From 87efbb63efccc46f4f583a5fa267b4f10e655806 Mon Sep 17 00:00:00 2001
From: Harisankar Sadasivan <hari.sadasivan@amd.com>
Date: Mon, 8 Apr 2024 22:29:36 +0000
Subject: [PATCH] experimental power analysis branch

---
 ...peration_instance.hpp.C89DB8EBC3EFFE7D.idx | Bin 0 -> 1268 bytes
 .../index/algorithm.hpp.3E7EDFC4AA4FB32B.idx  | Bin 0 -> 2408 bytes
 ...amd_address_space.hpp.FB92D54F0A7BAF6E.idx | Bin 0 -> 890 bytes
 ...buffer_addressing.hpp.7FB2D5AE48874EC4.idx | Bin 0 -> 17286 bytes
 .../amd_gemm_dpp.hpp.55564336A8EE078E.idx     | Bin 0 -> 1586 bytes
 .../amd_inline_asm.hpp.7777A491B073ADB0.idx   | Bin 0 -> 6854 bytes
 .../index/amd_lds.hpp.3739CBB0B218918C.idx    | Bin 0 -> 1316 bytes
 ...e_read_first_lane.hpp.F8A301BEB898DCEC.idx | Bin 0 -> 3546 bytes
 .../index/amd_wmma.hpp.143B2B78EF519046.idx   | Bin 0 -> 3650 bytes
 .../index/amd_xdlops.hpp.5A2FDE085CB4BC8A.idx | Bin 0 -> 12144 bytes
 .../index/array.hpp.EEBA77115A6888DA.idx      | Bin 0 -> 2010 bytes
 .../avg_pool3d_bwd.hpp.40079D84C914EB9A.idx   | Bin 0 -> 1646 bytes
 ...c_instance_common.hpp.EDEFB32170FF2FE7.idx | Bin 0 -> 1814 bytes
 ...vgpool3d_bwd_bf16.cpp.CC3B9B35F218D95E.idx | Bin 0 -> 2172 bytes
 ...pool3d_bwd_common.hpp.C26898608FBBAD5A.idx | Bin 0 -> 3936 bytes
 ...vgpool3d_bwd_fp16.cpp.11C3DA7A01E70C82.idx | Bin 0 -> 2178 bytes
 ...vgpool3d_bwd_fp32.cpp.4BE51B531517C414.idx | Bin 0 -> 2140 bytes
 .../batched_gemm.hpp.E9C31BFE1C18D556.idx     | Bin 0 -> 7204 bytes
 ...gemm_add_xdl_fp16.cpp.25B35CBDBE38B07B.idx | Bin 0 -> 13184 bytes
 ...add_relu_gemm_add.hpp.B9BDBBF21D219B22.idx | Bin 0 -> 2568 bytes
 ..._permute_xdl_fp16.cpp.6413F9674606CEBE.idx | Bin 0 -> 10658 bytes
 ...tmax_gemm_permute.hpp.3291370BDDEB76AF.idx | Bin 0 -> 3200 bytes
 ...batched_gemm_gemm.hpp.3359348295C6E9CF.idx | Bin 0 -> 2086 bytes
 ...emm_gemm_xdl_bf16.cpp.38FE56A993CA3FF4.idx | Bin 0 -> 3920 bytes
 ...emm_gemm_xdl_fp16.cpp.F5472E9B2342F0A2.idx | Bin 0 -> 3920 bytes
 ...emm_gemm_xdl_fp32.cpp.958D740430E8AB79.idx | Bin 0 -> 3840 bytes
 ...emm_gemm_xdl_int8.cpp.EF48A52F4EBD9A62.idx | Bin 0 -> 3798 bytes
 ..._permute_xdl_fp16.cpp.20C4D0C379DDEEBB.idx | Bin 0 -> 4830 bytes
 ...ched_gemm_multi_d.hpp.4E3016E76D827F13.idx | Bin 0 -> 398 bytes
 ..._gemm_reduce_fp16.cpp.795EB59AA922209F.idx | Bin 0 -> 1562 bytes
 ...m_reduce_xdl_fp16.cpp.6EA78CC685DE72B1.idx | Bin 0 -> 8444 bytes
 ..._permute_xdl_bf16.cpp.3B79E782257965A3.idx | Bin 0 -> 4800 bytes
 ..._permute_xdl_fp16.cpp.71938F78DFDFE1BE.idx | Bin 0 -> 4818 bytes
 ...max_gemm_xdl_bf16.cpp.74FE791FA1A9F5A4.idx | Bin 0 -> 4096 bytes
 ...max_gemm_xdl_fp16.cpp.B77888D4EF3A20F0.idx | Bin 0 -> 4112 bytes
 ...gemm_softmax_gemm.hpp.2D7D43E1651E0616.idx | Bin 0 -> 2018 bytes
 ...tmax_gemm_permute.hpp.A9BA319B35F534EC.idx | Bin 0 -> 2990 bytes
 ...hed_gemm_xdl_bf16.cpp.ED68A85BB00385BD.idx | Bin 0 -> 3390 bytes
 ...hed_gemm_xdl_fp16.cpp.A653F9336CDDD36D.idx | Bin 0 -> 3390 bytes
 ...hed_gemm_xdl_fp32.cpp.B7D138F072E276A3.idx | Bin 0 -> 3312 bytes
 ...hed_gemm_xdl_int8.cpp.AEAA0DF5DDD6BDB1.idx | Bin 0 -> 3270 bytes
 ...atchnorm_backward.hpp.7322FEA177E6E308.idx | Bin 0 -> 2940 bytes
 ...orm_backward_nhwc.cpp.7637AB6D79637171.idx | Bin 0 -> 11414 bytes
 ...chnorm_bwd_rank_4.cpp.15A9A6FE401A579D.idx | Bin 0 -> 5142 bytes
 .../batchnorm_common.hpp.7297BD2852C87A9D.idx | Bin 0 -> 1590 bytes
 ...batchnorm_forward.hpp.7C4BF3E98C6D2CE7.idx | Bin 0 -> 2766 bytes
 ...rd_inferring_nhwc.cpp.5CE047E1603B8333.idx | Bin 0 -> 7998 bytes
 ...ard_training_nhwc.cpp.6918CB89E3F27E24.idx | Bin 0 -> 11660 bytes
 ...ing_nhwc_obsolete.cpp.B9948F07339573BC.idx | Bin 0 -> 11668 bytes
 ...chnorm_fwd_rank_4.cpp.051838AC0BBFCDA1.idx | Bin 0 -> 5154 bytes
 .../batchnorm_infer.hpp.F7608824D3FE5589.idx  | Bin 0 -> 2752 bytes
 ...chnorm_infer_impl.hpp.0B7228D5ED433292.idx | Bin 0 -> 1686 bytes
 ...norm_infer_rank_4.cpp.16E993641F247D85.idx | Bin 0 -> 4966 bytes
 ...nt_wise_operation.hpp.D7064E30A6CF6778.idx | Bin 0 -> 9884 bytes
 ...lock_to_ctile_map.hpp.F32EBB9EC094E255.idx | Bin 0 -> 31166 bytes
 ...lockwise_gemm_dpp.hpp.5CB9342685AA4AD4.idx | Bin 0 -> 9864 bytes
 ...m_pipeline_xdlops.hpp.C000528561553163.idx | Bin 0 -> 13708 bytes
 ...ockwise_gemm_wmma.hpp.33F41B2FF8044BC6.idx | Bin 0 -> 14406 bytes
 ...e_gemm_xdl_traits.hpp.9E775B78211AACFC.idx | Bin 0 -> 2320 bytes
 ...kwise_gemm_xdlops.hpp.E5193AF5C1961F73.idx | Bin 0 -> 24308 bytes
 ...xdlops_skip_b_lds.hpp.BC4BEC2371CE0CF7.idx | Bin 0 -> 8568 bytes
 ...blockwise_softmax.hpp.D0089700D1C9998B.idx | Bin 0 -> 2884 bytes
 ...blockwise_welford.hpp.93790E23276AA4A9.idx | Bin 0 -> 3000 bytes
 ...ast_add_2d_amn_bn.cpp.7B44FEB5DE986168.idx | Bin 0 -> 3752 bytes
 ...st_add_3d_am_bmnk.cpp.D83172D8CBF829F1.idx | Bin 0 -> 3846 bytes
 ...tyle_pointer_cast.hpp.6866D538A80ECC76.idx | Bin 0 -> 392 bytes
 .../cgemm_xdl_bf16.cpp.D34D118D7BC516E4.idx   | Bin 0 -> 3280 bytes
 .../cgemm_xdl_common.hpp.FBF552CB8A2753F4.idx | Bin 0 -> 5352 bytes
 .../cgemm_xdl_fp16.cpp.93DC5CFE4CF25E39.idx   | Bin 0 -> 3360 bytes
 .../cgemm_xdl_fp32.cpp.D4109241584CC5F6.idx   | Bin 0 -> 3276 bytes
 .../cgemm_xdl_int8.cpp.7D0F341E5936B471.idx   | Bin 0 -> 3284 bytes
 .../index/check_err.hpp.8BCADBDE7B4A301B.idx  | Bin 0 -> 11746 bytes
 .../clangd/index/ck.hpp.5BA5ECB9E73F13E2.idx  | Bin 0 -> 2384 bytes
 ...luster_descriptor.hpp.B962ADBD27415A8A.idx | Bin 0 -> 766 bytes
 ...lumn_to_image_f32.cpp.B78257C7C44FE524.idx | Bin 0 -> 4954 bytes
 .../index/common.hpp.030C990214BE859F.idx     | Bin 0 -> 2272 bytes
 .../index/common.hpp.0CF2F8C8A70630AF.idx     | Bin 0 -> 482 bytes
 .../index/common.hpp.116380F0FD4C88BB.idx     | Bin 0 -> 4426 bytes
 .../index/common.hpp.30E7C524D1E7DCC8.idx     | Bin 0 -> 3458 bytes
 .../index/common.hpp.3324DE766DB5A26E.idx     | Bin 0 -> 8422 bytes
 .../index/common.hpp.725393A22E55A928.idx     | Bin 0 -> 3338 bytes
 .../index/common.hpp.7922E75661986F8F.idx     | Bin 0 -> 2048 bytes
 .../index/common.hpp.98AE17ADFF3CA289.idx     | Bin 0 -> 608 bytes
 .../index/common.hpp.A2FB78C7F0E24A0E.idx     | Bin 0 -> 3378 bytes
 .../index/common.hpp.AA5CB32F503E8500.idx     | Bin 0 -> 4182 bytes
 .../index/common.hpp.D8FBDC6F47AA6D80.idx     | Bin 0 -> 560 bytes
 .../index/common.hpp.E36EEFF638BE348F.idx     | Bin 0 -> 14368 bytes
 .../index/common.hpp.F52D770AC14BC38D.idx     | Bin 0 -> 1952 bytes
 .../index/common.hpp.FF4590F87A18C9AE.idx     | Bin 0 -> 562 bytes
 .../common_header.hpp.BCB8C81E7AAAD43F.idx    | Bin 0 -> 1002 bytes
 .../common_instances.hpp.C20932BA881E0448.idx | Bin 0 -> 2736 bytes
 ...er_element_picker.hpp.D0295372D9F2DFD0.idx | Bin 0 -> 3104 bytes
 .../container_helper.hpp.57BCB53456250D8A.idx | Bin 0 -> 5916 bytes
 ...traction_bilinear.hpp.3FE8920B21D2C083.idx | Bin 0 -> 8878 bytes
 ...bf16_compute_fp32.cpp.34FCAF706B5AEB99.idx | Bin 0 -> 3394 bytes
 ...fp16_compute_fp32.cpp.038CE2A80837D541.idx | Bin 0 -> 3392 bytes
 ...bilinear_xdl_fp32.cpp.37E0EC20BE4060B3.idx | Bin 0 -> 3378 bytes
 ...fp32_compute_bf16.cpp.4B6BEF9EB4DD9C01.idx | Bin 0 -> 3394 bytes
 ...fp32_compute_fp16.cpp.F92204694821E2EC.idx | Bin 0 -> 3392 bytes
 ...bilinear_xdl_fp64.cpp.D740E9866ECDD95C.idx | Bin 0 -> 3378 bytes
 ...fp64_compute_fp32.cpp.29832DEA4C5E2D0A.idx | Bin 0 -> 3392 bytes
 ...ulti_ABD_xdl_fp16.cpp.95E96326F13F98F9.idx | Bin 0 -> 9356 bytes
 ...contraction_scale.hpp.9A93BE54CAD13142.idx | Bin 0 -> 8836 bytes
 ...bf16_compute_fp32.cpp.2E396E91009DF0FA.idx | Bin 0 -> 3306 bytes
 ...fp16_compute_fp32.cpp.A7A07273739AC223.idx | Bin 0 -> 3306 bytes
 ...on_scale_xdl_fp32.cpp.1A2C9B556CFF860B.idx | Bin 0 -> 3292 bytes
 ...fp32_compute_bf16.cpp.301F203197216C44.idx | Bin 0 -> 3308 bytes
 ...fp32_compute_fp16.cpp.F7341E88A32312DC.idx | Bin 0 -> 3306 bytes
 ...on_scale_xdl_fp64.cpp.7417F385FB954651.idx | Bin 0 -> 3292 bytes
 ...fp64_compute_fp32.cpp.02A2D58B76014A4B.idx | Bin 0 -> 3306 bytes
 ...quantization_int8.cpp.D342A33945D6732C.idx | Bin 0 -> 2866 bytes
 ...quantization_int8.cpp.E5EA635D0BE03F7D.idx | Bin 0 -> 2788 bytes
 ...quantization_int8.cpp.DA805D9EE5DF962F.idx | Bin 0 -> 2770 bytes
 ...quantization_int8.cpp.7D138D1D290C5C53.idx | Bin 0 -> 2684 bytes
 ...antization_common.hpp.2C2B760A7CE8525E.idx | Bin 0 -> 3132 bytes
 ..._scaleadd_ab_bf16.cpp.8FB5CAA0FDE2EFE8.idx | Bin 0 -> 2030 bytes
 ..._scaleadd_ab_fp16.cpp.A862E22D755445CE.idx | Bin 0 -> 2030 bytes
 ..._scaleadd_ab_fp32.cpp.E8494EFCFB4F3B30.idx | Bin 0 -> 2008 bytes
 ..._scaleadd_ab_int8.cpp.97C188D55BA3C9C1.idx | Bin 0 -> 2054 bytes
 ..._tensor_rearrange.hpp.1BD4F049133037BC.idx | Bin 0 -> 14330 bytes
 ...nsor_rearrange_op.hpp.1DE16BC6AE42EE5B.idx | Bin 0 -> 846 bytes
 .../index/conv_util.cpp.2DB2D9F346007F63.idx  | Bin 0 -> 8702 bytes
 .../convnd_bwd_data.cpp.7C4E178AEE50DCF2.idx  | Bin 0 -> 4154 bytes
 ...d_bwd_data_common.hpp.FAAF76603FF7D70B.idx | Bin 0 -> 3788 bytes
 ...ear_residual_fp16.cpp.17D6FDBA1D601B9A.idx | Bin 0 -> 7106 bytes
 ...bwd_data_xdl_fp16.cpp.32F247EF8BEF9A27.idx | Bin 0 -> 3386 bytes
 .../index/convnd_fwd.cpp.63437EF3B0E7154B.idx | Bin 0 -> 4144 bytes
 ...v_multi_ab_common.hpp.F16D50D450CBF542.idx | Bin 0 -> 6384 bytes
 ...ctiv_unary_common.hpp.A3C6C6A8E8602CD2.idx | Bin 0 -> 5982 bytes
 ...convnd_fwd_common.hpp.1735C0647D85C158.idx | Bin 0 -> 4356 bytes
 ..._fwd_max_xdl_bf16.cpp.32DD457450CC173D.idx | Bin 0 -> 1820 bytes
 ..._fwd_max_xdl_fp16.cpp.57DCB3A8914F4BCB.idx | Bin 0 -> 1818 bytes
 ..._fwd_max_xdl_fp32.cpp.7C22839582612580.idx | Bin 0 -> 1808 bytes
 ..._fwd_max_xdl_int8.cpp.E84F6E18A0DCB8A3.idx | Bin 0 -> 1818 bytes
 ..._fwd_xdl_abs_fp16.cpp.B14E3C722763D4B8.idx | Bin 0 -> 1318 bytes
 ...nvnd_fwd_xdl_bf16.cpp.054A24EF5BFF4A1F.idx | Bin 0 -> 2454 bytes
 ...ear_residual_fp16.cpp.C4A66C697A8DEBDA.idx | Bin 0 -> 7154 bytes
 ..._clippedrelu_fp16.cpp.4759FB8DCFC58E7D.idx | Bin 0 -> 1324 bytes
 ..._fwd_xdl_elu_fp16.cpp.922AD9048F6922A8.idx | Bin 0 -> 1318 bytes
 ...nvnd_fwd_xdl_fp16.cpp.E044D0CD44CCFE7F.idx | Bin 0 -> 2476 bytes
 ...nvnd_fwd_xdl_fp32.cpp.911AF34195327474.idx | Bin 0 -> 2402 bytes
 ...nvnd_fwd_xdl_fp64.cpp.3D0123C922DFAFA2.idx | Bin 0 -> 2402 bytes
 ...nvnd_fwd_xdl_int8.cpp.1BF1AFB682499568.idx | Bin 0 -> 2490 bytes
 ...dl_leakyrelu_fp16.cpp.BA4E6713D7C74F09.idx | Bin 0 -> 1324 bytes
 ..._fwd_xdl_pow_fp16.cpp.0F662794CF8A0DB2.idx | Bin 0 -> 1318 bytes
 ...fwd_xdl_relu_fp16.cpp.8F7B959CCFC1E66F.idx | Bin 0 -> 1320 bytes
 ...bcasted_bias_fp16.cpp.54C5613FFA1E2460.idx | Bin 0 -> 8448 bytes
 ...caleadd_relu_fp16.cpp.303711D8C093AB3A.idx | Bin 0 -> 7736 bytes
 ..._xdl_sigmoid_fp16.cpp.7462E7274E7DC93D.idx | Bin 0 -> 1320 bytes
 ...xdl_softrelu_fp16.cpp.47B66AF94BC3AD6A.idx | Bin 0 -> 1322 bytes
 ...fwd_xdl_tanh_fp16.cpp.A1CCA35C668C7A32.idx | Bin 0 -> 1320 bytes
 ...ion_backward_data.hpp.F488158732BCE944.idx | Bin 0 -> 6146 bytes
 ...ta_specialization.hpp.E2F7DA4511290758.idx | Bin 0 -> 714 bytes
 ...ht_specialization.hpp.D1A61821B9657E4F.idx | Bin 0 -> 916 bytes
 ...nvolution_forward.hpp.8E3EE29C519B7A7E.idx | Bin 0 -> 2910 bytes
 ...rd_specialization.hpp.29B75FDB4D9A7ED7.idx | Bin 0 -> 948 bytes
 ...descriptor_helper.hpp.C7E0FAE3D228F16B.idx | Bin 0 -> 8190 bytes
 ...olution_parameter.cpp.AB24172BD3A32784.idx | Bin 0 -> 5098 bytes
 ...olution_parameter.hpp.98A4E91AB8FFEC38.idx | Bin 0 -> 3392 bytes
 .../index/copy.hpp.D47879DBFF83B386.idx       | Bin 0 -> 3316 bytes
 .../index/data_type.hpp.FA421D61CFEE8D86.idx  | Bin 0 -> 39350 bytes
 .../data_type_enum.hpp.9C984CB93A76A458.idx   | Bin 0 -> 748 bytes
 .../index/debug.hpp.3D8E2A10AC86C578.idx      | Bin 0 -> 1998 bytes
 ...hwc_bf16_instance.cpp.8DD0CEE151EEE78C.idx | Bin 0 -> 1328 bytes
 ...dhwc_f16_instance.cpp.E45DD97792659A1B.idx | Bin 0 -> 1326 bytes
 ...dhwc_f32_instance.cpp.0A50ACDD7655DC8C.idx | Bin 0 -> 1326 bytes
 ...hwc_bf16_instance.cpp.C65ED37BE0626399.idx | Bin 0 -> 1546 bytes
 ...dhwc_f16_instance.cpp.322ABBBE306F5168.idx | Bin 0 -> 1546 bytes
 ...dhwc_f32_instance.cpp.BE9138723EE82CA7.idx | Bin 0 -> 1538 bytes
 ...d_bwd_ndhwc_ndhwc.hpp.B8CFD83A7F04917D.idx | Bin 0 -> 12716 bytes
 ...evice_avgpool_bwd.hpp.65CD0007480F503D.idx | Bin 0 -> 972 bytes
 .../device_base.hpp.07667515082DA1A8.idx      | Bin 0 -> 2394 bytes
 ...action_multiple_d.hpp.C066C76CB17254D1.idx | Bin 0 -> 1326 bytes
 ...le_d_xdl_cshuffle.hpp.9FB9459677B68FA8.idx | Bin 0 -> 20364 bytes
 ...vice_batched_gemm.hpp.0D6DC4AA71BF6793.idx | Bin 0 -> 966 bytes
 ..._gno_gmo_instance.cpp.2CCE15A52A6097D9.idx | Bin 0 -> 8428 bytes
 ..._gon_gmo_instance.cpp.09BE891BDB3DB7E7.idx | Bin 0 -> 8972 bytes
 ..._f16_f16_instance.cpp.FE05EE1022E21C41.idx | Bin 0 -> 10890 bytes
 ..._gno_gmo_instance.cpp.5255F78DB2565F3C.idx | Bin 0 -> 10584 bytes
 ..._gno_gmo_instance.cpp.B4883A86B075066E.idx | Bin 0 -> 10588 bytes
 ...batched_gemm_gemm.hpp.FDFBA95BB6D583D3.idx | Bin 0 -> 960 bytes
 ...gemm_xdl_cshuffle.hpp.8AECE81C6A31D6CE.idx | Bin 0 -> 14622 bytes
 ..._gno_gmo_instance.cpp.24C81AB80A523E1D.idx | Bin 0 -> 8092 bytes
 ..._gon_gmo_instance.cpp.42B3A408DD25BE6F.idx | Bin 0 -> 8092 bytes
 ...ched_gemm_multi_d.hpp.65CBFD3787EC8604.idx | Bin 0 -> 1174 bytes
 ..._gemm_multi_d_xdl.hpp.5F31589668DB0CDD.idx | Bin 0 -> 15014 bytes
 ...d_gemm_multiple_d.hpp.1EC37BB5D09F3746.idx | Bin 0 -> 1386 bytes
 ...le_d_xdl_cshuffle.hpp.1D083EFB833A8BDD.idx | Bin 0 -> 18206 bytes
 ...duce_xdl_cshuffle.hpp.008517D2C19930C3.idx | Bin 0 -> 17582 bytes
 ..._gkn_gmn_instance.cpp.3CE3CBFFC33E63B5.idx | Bin 0 -> 9496 bytes
 ..._gnk_gmn_instance.cpp.C2D6CBC7BDD03D15.idx | Bin 0 -> 9496 bytes
 ..._gkn_gmn_instance.cpp.BE70F5327F5BF3ED.idx | Bin 0 -> 9496 bytes
 ..._gnk_gmn_instance.cpp.542AEDD871493BF1.idx | Bin 0 -> 8262 bytes
 ...gemm_softmax_gemm.hpp.2CF55436E05C96D5.idx | Bin 0 -> 1012 bytes
 ...tmax_gemm_permute.hpp.0F50C2899670001D.idx | Bin 0 -> 1570 bytes
 ...mute_xdl_cshuffle.hpp.72B4F2954B40EC19.idx | Bin 0 -> 20010 bytes
 ..._gno_gmo_instance.cpp.CA655EDB759D211C.idx | Bin 0 -> 9360 bytes
 ..._gno_gmo_instance.cpp.54BCAA9E5C4C7CA0.idx | Bin 0 -> 9372 bytes
 ...gemm_xdl_cshuffle.hpp.5F249F65D9D49000.idx | Bin 0 -> 13924 bytes
 ..._gno_gmo_instance.cpp.9E1D81891D4D775B.idx | Bin 0 -> 11160 bytes
 ..._batched_gemm_xdl.hpp.336993554588B91D.idx | Bin 0 -> 8590 bytes
 ..._gkn_gmn_instance.cpp.01B3A3CE0FB07AA4.idx | Bin 0 -> 4068 bytes
 ..._gnk_gmn_instance.cpp.6732FB26500DDC13.idx | Bin 0 -> 4068 bytes
 ..._gkn_gmn_instance.cpp.5C50F8886E987210.idx | Bin 0 -> 5068 bytes
 ..._gnk_gmn_instance.cpp.991CBBA160B6AE8A.idx | Bin 0 -> 5318 bytes
 ..._gkn_gmn_instance.cpp.909733D52FF49021.idx | Bin 0 -> 9704 bytes
 ..._gnk_gmn_instance.cpp.EF1B1B5023E691DD.idx | Bin 0 -> 9706 bytes
 ..._gkn_gmn_instance.cpp.6191F299E76FA25A.idx | Bin 0 -> 18620 bytes
 ..._gnk_gmn_instance.cpp.1378958A5E1CC674.idx | Bin 0 -> 14848 bytes
 ..._gkn_gmn_instance.cpp.155CED9F59B77387.idx | Bin 0 -> 4066 bytes
 ..._gnk_gmn_instance.cpp.021E4B84540D9AAD.idx | Bin 0 -> 4068 bytes
 ..._gkn_gmn_instance.cpp.2D0A22E72C5CD4C0.idx | Bin 0 -> 4066 bytes
 ..._gnk_gmn_instance.cpp.6662845BB6111F27.idx | Bin 0 -> 5316 bytes
 ..._gkn_gmn_instance.cpp.3B9F48B295268F74.idx | Bin 0 -> 7482 bytes
 ..._gnk_gmn_instance.cpp.0C31B57B423E19C5.idx | Bin 0 -> 7482 bytes
 ..._gkn_gmn_instance.cpp.8439FC102CF881D2.idx | Bin 0 -> 7482 bytes
 ..._gnk_gmn_instance.cpp.9E2B022D116E5337.idx | Bin 0 -> 5482 bytes
 ...atchnorm_backward.hpp.43E337CEA4624B30.idx | Bin 0 -> 1304 bytes
 ...ard_bf16_instance.cpp.CEBDED0AD1292F09.idx | Bin 0 -> 12744 bytes
 ...ward_f16_instance.cpp.D1D5284FB6757AF3.idx | Bin 0 -> 12744 bytes
 ...ward_f32_instance.cpp.CA2117C76C8B866B.idx | Bin 0 -> 12646 bytes
 ...ward_f64_instance.cpp.B99A50DCD60DF309.idx | Bin 0 -> 12648 bytes
 ...orm_backward_impl.hpp.8B916A92DFA8139E.idx | Bin 0 -> 17592 bytes
 ...batchnorm_forward.hpp.0E1593921379F2F3.idx | Bin 0 -> 1108 bytes
 ...ard_bf16_instance.cpp.F51389A9E014ED43.idx | Bin 0 -> 11382 bytes
 ...ward_f16_instance.cpp.5B31F4805B2AC1BA.idx | Bin 0 -> 11382 bytes
 ...ward_f32_instance.cpp.DAA81C77B8C2B284.idx | Bin 0 -> 11288 bytes
 ...ward_f64_instance.cpp.B473866A874EE221.idx | Bin 0 -> 11288 bytes
 ...norm_forward_impl.hpp.6D2CA9CBDDC84AC3.idx | Bin 0 -> 15958 bytes
 ...ard_impl_obsolete.hpp.7A0467126AEB6502.idx | Bin 0 -> 14182 bytes
 ...e_batchnorm_infer.hpp.4EDE573CB1DF1297.idx | Bin 0 -> 1108 bytes
 ...fer_bf16_instance.cpp.FEC3D08EDEA4CE6B.idx | Bin 0 -> 4188 bytes
 ...nfer_f16_instance.cpp.B01DECCDEE63A78A.idx | Bin 0 -> 4174 bytes
 ...nfer_f32_instance.cpp.889B58511EB200D7.idx | Bin 0 -> 4072 bytes
 ...nfer_f64_instance.cpp.0266490535848A23.idx | Bin 0 -> 3232 bytes
 .../device_cgemm.hpp.CD6CA3F05D9AAF4D.idx     | Bin 0 -> 988 bytes
 ...gemm_xdl_cshuffle.hpp.CB4931718165FBCB.idx | Bin 0 -> 10532 bytes
 ...ndhwc_3d_instance.cpp.8AF56F1E7AD14DC6.idx | Bin 0 -> 2168 bytes
 ...gnhwc_2d_instance.cpp.8EBF74F132E896AB.idx | Bin 0 -> 2166 bytes
 ..._gnwc_1d_instance.cpp.C6C85A0C333E012A.idx | Bin 0 -> 2166 bytes
 ...umn_to_image_impl.hpp.545E1241F8BEB14B.idx | Bin 0 -> 18496 bytes
 ...to_image_instance.hpp.FA1AC606D0CD0897.idx | Bin 0 -> 2682 bytes
 ...dhwgc_3d_instance.cpp.CBDECCEC6F806814.idx | Bin 0 -> 2168 bytes
 ...nhwgc_2d_instance.cpp.9A3328B90E30CC03.idx | Bin 0 -> 2166 bytes
 ..._nwgc_1d_instance.cpp.163F7E8C2077B0E0.idx | Bin 0 -> 2166 bytes
 ...f32_kknn_instance.cpp.79A78F49E52676FD.idx | Bin 0 -> 1794 bytes
 ...f32_knnn_instance.cpp.DB04A314807BEF1C.idx | Bin 0 -> 1794 bytes
 ...f32_mknn_instance.cpp.E624632F1E715961.idx | Bin 0 -> 1796 bytes
 ...f32_mnnn_instance.cpp.CACB611E1CE9F1B5.idx | Bin 0 -> 1796 bytes
 ...f32_kknn_instance.cpp.9E3DEE644D650E5A.idx | Bin 0 -> 1794 bytes
 ...f32_knnn_instance.cpp.384AC715F1A68557.idx | Bin 0 -> 1794 bytes
 ...f32_mknn_instance.cpp.66D5EE934903111A.idx | Bin 0 -> 1794 bytes
 ...f32_mnnn_instance.cpp.8194F48DCC1F513B.idx | Bin 0 -> 1794 bytes
 ...f16_kknn_instance.cpp.ED87AF8FD1B46167.idx | Bin 0 -> 1796 bytes
 ...f16_knnn_instance.cpp.9B28805BAED0F92D.idx | Bin 0 -> 1796 bytes
 ...f16_mknn_instance.cpp.7E6A2A9800DD0E7A.idx | Bin 0 -> 1796 bytes
 ...f16_mnnn_instance.cpp.E0C6A616E43DC9DD.idx | Bin 0 -> 1796 bytes
 ...f16_kknn_instance.cpp.01988F25803E8456.idx | Bin 0 -> 1794 bytes
 ...f16_knnn_instance.cpp.AFEB4718FB33138B.idx | Bin 0 -> 1794 bytes
 ...f16_mknn_instance.cpp.D913F33B640C756C.idx | Bin 0 -> 1794 bytes
 ...f16_mnnn_instance.cpp.EF41093E7954CFCE.idx | Bin 0 -> 1794 bytes
 ...f32_kknn_instance.cpp.CB877B16EA5C28D0.idx | Bin 0 -> 1778 bytes
 ...f32_knnn_instance.cpp.7AE9E872C7D0A549.idx | Bin 0 -> 1778 bytes
 ...f32_mknn_instance.cpp.E366D05CBA9CA229.idx | Bin 0 -> 1778 bytes
 ...f32_mnnn_instance.cpp.F88287BB0527CE6D.idx | Bin 0 -> 1778 bytes
 ...f32_kknn_instance.cpp.03F037ABB1A17197.idx | Bin 0 -> 1794 bytes
 ...f32_knnn_instance.cpp.237A2B88B920F202.idx | Bin 0 -> 1794 bytes
 ...f32_mknn_instance.cpp.177CE21E4E683B24.idx | Bin 0 -> 1794 bytes
 ...f32_mnnn_instance.cpp.FE8F7C3907743A26.idx | Bin 0 -> 1794 bytes
 ...f64_kknn_instance.cpp.1FB1A94EA79EE13A.idx | Bin 0 -> 1778 bytes
 ...f64_knnn_instance.cpp.6BFB81AFEEB1C1B7.idx | Bin 0 -> 1778 bytes
 ...f64_mknn_instance.cpp.D3FA995D6DF314A6.idx | Bin 0 -> 1778 bytes
 ...f64_mnnn_instance.cpp.D00F48772AF74F17.idx | Bin 0 -> 1778 bytes
 ...traction_instance.hpp.15AB9D8608FA40A0.idx | Bin 0 -> 19816 bytes
 ...tion_multiple_abd.hpp.D0EC339053CFF7D0.idx | Bin 0 -> 1606 bytes
 ..._abd_xdl_cshuffle.hpp.9AA6B0ABD9EE929D.idx | Bin 0 -> 15790 bytes
 ...action_multiple_d.hpp.CA3D08C20FA799BC.idx | Bin 0 -> 1294 bytes
 ...le_d_xdl_cshuffle.hpp.18531F18DA84F612.idx | Bin 0 -> 14086 bytes
 ..._f32_kkn_instance.cpp.634AB0569FA8301A.idx | Bin 0 -> 1792 bytes
 ..._f32_knn_instance.cpp.1229C32BCA5E8056.idx | Bin 0 -> 1792 bytes
 ..._f32_mkn_instance.cpp.B21239C1EC74587F.idx | Bin 0 -> 1792 bytes
 ..._f32_mnn_instance.cpp.92168E842E069568.idx | Bin 0 -> 1792 bytes
 ..._f32_kkn_instance.cpp.8EFB8D272907AC49.idx | Bin 0 -> 1790 bytes
 ..._f32_knn_instance.cpp.4BAF3FDF22005D6C.idx | Bin 0 -> 1790 bytes
 ..._f32_mkn_instance.cpp.F76E34EF68D8A65B.idx | Bin 0 -> 1790 bytes
 ..._f32_mnn_instance.cpp.1704AA4E0EFAB142.idx | Bin 0 -> 1790 bytes
 ...bf16_kkn_instance.cpp.29A1243F9B97E124.idx | Bin 0 -> 1792 bytes
 ...bf16_knn_instance.cpp.020B0D68AE892602.idx | Bin 0 -> 1792 bytes
 ...bf16_mkn_instance.cpp.A426959DB1641563.idx | Bin 0 -> 1792 bytes
 ...bf16_mnn_instance.cpp.1B1383BC1860443B.idx | Bin 0 -> 1792 bytes
 ..._f16_kkn_instance.cpp.C0ABEB6CD7C15E09.idx | Bin 0 -> 1790 bytes
 ..._f16_knn_instance.cpp.76E857B0424F98D1.idx | Bin 0 -> 1790 bytes
 ..._f16_mkn_instance.cpp.2C4EF55D06D154CA.idx | Bin 0 -> 1790 bytes
 ..._f16_mnn_instance.cpp.D9C825C0FBC7162C.idx | Bin 0 -> 1790 bytes
 ..._f32_kkn_instance.cpp.9ED7E55B00DD6F70.idx | Bin 0 -> 1772 bytes
 ..._f32_knn_instance.cpp.87288671AE2510FD.idx | Bin 0 -> 1772 bytes
 ..._f32_mkn_instance.cpp.AA8697A6E53E704C.idx | Bin 0 -> 1772 bytes
 ..._f32_mnn_instance.cpp.FC0AF91C98199AEA.idx | Bin 0 -> 1772 bytes
 ..._f32_kkn_instance.cpp.265462C1E86DC0F1.idx | Bin 0 -> 1790 bytes
 ..._f32_knn_instance.cpp.32EBBC25AA6C0CF0.idx | Bin 0 -> 1790 bytes
 ..._f32_mkn_instance.cpp.5C1840A2D370420C.idx | Bin 0 -> 1790 bytes
 ..._f32_mnn_instance.cpp.05BFD4E4455B0BDF.idx | Bin 0 -> 1790 bytes
 ..._f64_kkn_instance.cpp.BFA2076F1038BB65.idx | Bin 0 -> 1772 bytes
 ..._f64_knn_instance.cpp.16DF4E204049F20E.idx | Bin 0 -> 1772 bytes
 ..._f64_mkn_instance.cpp.25AD39BA7B2B9C77.idx | Bin 0 -> 1772 bytes
 ..._f64_mnn_instance.cpp.F46067E1C6499190.idx | Bin 0 -> 1772 bytes
 ...contraction_utils.hpp.39AAD30D5F9C497C.idx | Bin 0 -> 1434 bytes
 ...nwk_bf16_instance.cpp.F95715A59DC357E4.idx | Bin 0 -> 8300 bytes
 ..._nwk_f16_instance.cpp.11B1653CC880738D.idx | Bin 0 -> 8298 bytes
 ..._nwk_f32_instance.cpp.1FF1E6FFC39B05F6.idx | Bin 0 -> 8220 bytes
 ...nwk_int8_instance.cpp.27AC774D5173DB46.idx | Bin 0 -> 8178 bytes
 ...dl_nhwc_kyxc_nhwk.hpp.5A99B55004C02041.idx | Bin 0 -> 16046 bytes
 ...hwk_bf16_instance.cpp.38978BB017721061.idx | Bin 0 -> 14580 bytes
 ...nhwk_f16_instance.cpp.015682E6D487BC89.idx | Bin 0 -> 14606 bytes
 ...nhwk_f32_instance.cpp.24521662C376153B.idx | Bin 0 -> 14474 bytes
 ...hwk_int8_instance.cpp.050ACEE75FFD374F.idx | Bin 0 -> 14450 bytes
 ...dd_nhwc_kyxc_nhwk.hpp.095F638AA48F9495.idx | Bin 0 -> 19404 bytes
 ...on_nhwc_kyxc_nhwk.hpp.03A5D53DEE4E2511.idx | Bin 0 -> 18908 bytes
 ...nhwk_f16_instance.cpp.FF7059ED53F4F0F2.idx | Bin 0 -> 16992 bytes
 ...nhwk_f16_instance.cpp.2C9D98DCC57DC34E.idx | Bin 0 -> 18026 bytes
 ...le_nhwc_kyxc_nhwk.hpp.66D8D8BBD73054E8.idx | Bin 0 -> 17954 bytes
 ...nhwk_f16_instance.cpp.3FB08C96D8C7394B.idx | Bin 0 -> 17088 bytes
 ...dl_nhwc_kyxc_nhwk.hpp.FAAE94F7CEBB58BA.idx | Bin 0 -> 13892 bytes
 ...hwk_bf16_instance.cpp.9102AFA84B67CE1B.idx | Bin 0 -> 11566 bytes
 ...nhwk_f16_instance.cpp.656EAF4E8A7F404D.idx | Bin 0 -> 11564 bytes
 ...nhwk_f32_instance.cpp.793B97DB82C3A02D.idx | Bin 0 -> 11488 bytes
 ...hwk_int8_instance.cpp.A7E63495E5270E3E.idx | Bin 0 -> 11446 bytes
 ...ion_int8_instance.cpp.486913CE7740FA6B.idx | Bin 0 -> 3328 bytes
 ...ion_int8_instance.cpp.8577D499B9483812.idx | Bin 0 -> 3332 bytes
 ...xdl_int8_instance.hpp.78D2B94EA941EB43.idx | Bin 0 -> 4024 bytes
 ...ion_int8_instance.cpp.C31E228888CDC52C.idx | Bin 0 -> 2610 bytes
 ...ion_int8_instance.cpp.108EC38462E7A7B4.idx | Bin 0 -> 2596 bytes
 ...hwk_bf16_instance.cpp.7487AFBFF492DC5A.idx | Bin 0 -> 8312 bytes
 ...dhwk_f16_instance.cpp.B17F2553F57B25A6.idx | Bin 0 -> 8312 bytes
 ...dhwk_f32_instance.cpp.D3A7F488899285AF.idx | Bin 0 -> 8234 bytes
 ...hwk_int8_instance.cpp.31CAAB24531BB9D5.idx | Bin 0 -> 8192 bytes
 ...ice_conv_bwd_data.hpp.81B90E235167286C.idx | Bin 0 -> 1020 bytes
 .../device_conv_fwd.hpp.257583CFB6A8DC72.idx  | Bin 0 -> 998 bytes
 ...d_bias_activation.hpp.D4F53CD24AD4278A.idx | Bin 0 -> 1058 bytes
 ...as_activation_add.hpp.D947A059FE31F7BB.idx | Bin 0 -> 1038 bytes
 ..._tensor_rearrange.hpp.90268624B8539288.idx | Bin 0 -> 1280 bytes
 ...a_nwc_kxc_nwk_xdl.hpp.A70DE3BD61066477.idx | Bin 0 -> 30982 bytes
 ...evice_elementwise.hpp.00198FF49AAFC017.idx | Bin 0 -> 1140 bytes
 ...ementwise_2d_impl.hpp.30FD009D4EA9FEE5.idx | Bin 0 -> 7354 bytes
 ...ementwise_3d_impl.hpp.EA5D0D82E2445BC8.idx | Bin 0 -> 7890 bytes
 ..._elementwise_impl.hpp.54205731D4AB5FA8.idx | Bin 0 -> 7054 bytes
 ...ise_normalization.hpp.2F6431A98ECDBEBB.idx | Bin 0 -> 1160 bytes
 ...tion_f16_instance.cpp.B8CDDDF357D7D999.idx | Bin 0 -> 3896 bytes
 ...ormalization_impl.hpp.D9648470B11E62DC.idx | Bin 0 -> 11014 bytes
 ...elementwise_scale.hpp.A8B4C8DA3A99ADD2.idx | Bin 0 -> 1168 bytes
 ...ntwise_scale_impl.hpp.792CB7A86911D632.idx | Bin 0 -> 6994 bytes
 .../device_gemm.hpp.C11C88B2080F572F.idx      | Bin 0 -> 764 bytes
 ...mn_mn_mn_instance.cpp.236524D86BD9A065.idx | Bin 0 -> 23776 bytes
 ...mn_mn_mn_instance.cpp.3F5C595E84E47B20.idx | Bin 0 -> 23776 bytes
 ...mn_mn_mn_instance.cpp.5836A0C2B3A14787.idx | Bin 0 -> 23778 bytes
 ...mn_mn_mn_instance.cpp.E0256C2451B1EE93.idx | Bin 0 -> 20106 bytes
 ...kn_mn_mn_instance.cpp.EED2788EEB0B75F1.idx | Bin 0 -> 3550 bytes
 ...kn_mn_mn_instance.cpp.F71CF786157F3658.idx | Bin 0 -> 22874 bytes
 ...nk_mn_mn_instance.cpp.777F64009D887A71.idx | Bin 0 -> 22876 bytes
 ...kn_mn_mn_instance.cpp.9B8C2015B4635A97.idx | Bin 0 -> 22866 bytes
 ...nk_mn_mn_instance.cpp.7475A9472170E32A.idx | Bin 0 -> 19314 bytes
 ...kn_mn_mn_instance.cpp.5CFCD880070835B7.idx | Bin 0 -> 3550 bytes
 ...mn_mn_mn_instance.cpp.15161CADD4CF1397.idx | Bin 0 -> 13202 bytes
 ...mn_mn_mn_instance.cpp.C192E44A2A2B2438.idx | Bin 0 -> 13204 bytes
 ...mn_mn_mn_instance.cpp.0E4AE68E13A6F5B4.idx | Bin 0 -> 13202 bytes
 ...mn_mn_mn_instance.cpp.3B87280201C19511.idx | Bin 0 -> 19820 bytes
 ...mn_mn_mn_instance.cpp.E948F8CEDD1EC4BF.idx | Bin 0 -> 11192 bytes
 ...mn_mn_mn_instance.cpp.0BA2333B2B00FEDC.idx | Bin 0 -> 11192 bytes
 ...mn_mn_mn_instance.cpp.F5D89038BFC386C8.idx | Bin 0 -> 11192 bytes
 ...mn_mn_mn_instance.cpp.935CC5A9EFAF5617.idx | Bin 0 -> 9952 bytes
 ...kn_mn_mn_instance.cpp.81A12F337280E714.idx | Bin 0 -> 3546 bytes
 ...kn_mn_mn_instance.cpp.6BB8AF1DE23C88B0.idx | Bin 0 -> 3546 bytes
 ...kn_mn_mn_instance.cpp.6559BA68564A041C.idx | Bin 0 -> 3546 bytes
 ...kn_mn_mn_instance.cpp.5F5C907654F27409.idx | Bin 0 -> 3546 bytes
 ...kn_mn_mn_instance.cpp.8D291954B92003AF.idx | Bin 0 -> 3542 bytes
 ...kn_mn_mn_instance.cpp.9FCCE6CADB35B88C.idx | Bin 0 -> 3540 bytes
 ...km_kn_mn_instance.cpp.479D1366D0F3CE30.idx | Bin 0 -> 10366 bytes
 ...km_nk_mn_instance.cpp.D6BBA9767E8DA592.idx | Bin 0 -> 10366 bytes
 ...mk_kn_mn_instance.cpp.F658EA6412E57274.idx | Bin 0 -> 10366 bytes
 ...mk_nk_mn_instance.cpp.D219505C6A3BDF13.idx | Bin 0 -> 8992 bytes
 ...duce_xdl_cshuffle.hpp.2AAE1676CF63D1B6.idx | Bin 0 -> 16068 bytes
 ...kn_mn_mn_instance.cpp.4E5648F8B76EED28.idx | Bin 0 -> 7866 bytes
 ...nk_mn_mn_instance.cpp.51C1492BB7273708.idx | Bin 0 -> 7868 bytes
 ...kn_mn_mn_instance.cpp.89C496A745057432.idx | Bin 0 -> 7854 bytes
 ...nk_mn_mn_instance.cpp.DA4EFE2E10A56944.idx | Bin 0 -> 9184 bytes
 ...kn_mn_mn_instance.cpp.80BE80F117BA699A.idx | Bin 0 -> 13170 bytes
 ...nk_mn_mn_instance.cpp.E76894F842B8152E.idx | Bin 0 -> 13172 bytes
 ...kn_mn_mn_instance.cpp.741D3DF3032F48B6.idx | Bin 0 -> 13170 bytes
 ...nk_mn_mn_instance.cpp.E3C0C7076B1A5420.idx | Bin 0 -> 19788 bytes
 .../device_gemm_dpp.hpp.86A83F148A978F3B.idx  | Bin 0 -> 4414 bytes
 ...km_kn_mn_instance.cpp.AC3765D99685AA83.idx | Bin 0 -> 6654 bytes
 ...rregular_instance.cpp.BE2169842E629D5F.idx | Bin 0 -> 5340 bytes
 ...km_nk_mn_instance.cpp.C525CCAE82B88202.idx | Bin 0 -> 6654 bytes
 ...rregular_instance.cpp.E70AE989E444006E.idx | Bin 0 -> 5340 bytes
 ...mk_kn_mn_instance.cpp.F48BB29B1F2C3D07.idx | Bin 0 -> 6574 bytes
 ...rregular_instance.cpp.9E7DF32B7E3B06F9.idx | Bin 0 -> 5260 bytes
 ...mk_nk_mn_instance.cpp.8E876A3BCC68461F.idx | Bin 0 -> 6654 bytes
 ...rregular_instance.cpp.F358D66C9DE4EC29.idx | Bin 0 -> 5340 bytes
 ...km_kn_mn_instance.cpp.B5661BE447C6D950.idx | Bin 0 -> 22954 bytes
 ...km_nk_mn_instance.cpp.E716A8264A34B5D8.idx | Bin 0 -> 22954 bytes
 ...mk_kn_mn_instance.cpp.B8E2B7EF254C9C8D.idx | Bin 0 -> 22898 bytes
 ...mk_nk_mn_instance.cpp.4104F408858A9B41.idx | Bin 0 -> 19332 bytes
 ...gemm_multiple_abd.hpp.01CABF5AC7B8D71E.idx | Bin 0 -> 1366 bytes
 ..._abd_xdl_cshuffle.hpp.6B55B3AF2C69EA61.idx | Bin 0 -> 12112 bytes
 ...e_gemm_multiple_d.hpp.A369BE7D23825A77.idx | Bin 0 -> 1094 bytes
 ...tiple_d_layernorm.hpp.1996D4382B7902A4.idx | Bin 0 -> 1192 bytes
 ...norm_xdl_cshuffle.hpp.6897EB3658A7024F.idx | Bin 0 -> 21030 bytes
 ...iple_d_multiple_r.hpp.D51C7B3444D17F61.idx | Bin 0 -> 1416 bytes
 ...le_r_xdl_cshuffle.hpp.76B6E1492C6F98CC.idx | Bin 0 -> 11816 bytes
 ...e_d_wmma_cshuffle.hpp.5BE3B8B060B3F7DD.idx | Bin 0 -> 15240 bytes
 ...le_d_xdl_cshuffle.hpp.A234525D7676EE3C.idx | Bin 0 -> 16830 bytes
 ...e_lds_direct_load.hpp.D18AD937E7A47E2A.idx | Bin 0 -> 6142 bytes
 ...mn_mn_mn_instance.cpp.46B874AA40214074.idx | Bin 0 -> 7822 bytes
 ...mn_mn_mn_instance.cpp.0EE0667EFBB4A693.idx | Bin 0 -> 6832 bytes
 ...mn_mn_mn_instance.cpp.4B2A61CBEBB69C46.idx | Bin 0 -> 8358 bytes
 ...mn_mn_mn_instance.cpp.E45F1434A64DD3E5.idx | Bin 0 -> 7368 bytes
 ...i8_i8_i8_instance.hpp.35BB7F78B886DD68.idx | Bin 0 -> 20044 bytes
 ...km_kn_mn_instance.cpp.6F38C058A13FF1D3.idx | Bin 0 -> 1862 bytes
 ...km_nk_mn_instance.cpp.0B3E6146A79A7808.idx | Bin 0 -> 1862 bytes
 ...mk_kn_mn_instance.cpp.487F7DBAC716F5E6.idx | Bin 0 -> 1852 bytes
 ...mk_nk_mn_instance.cpp.0D311D4C9F7E0828.idx | Bin 0 -> 1862 bytes
 ...evice_gemm_reduce.hpp.536083621BE180B4.idx | Bin 0 -> 988 bytes
 ...duce_xdl_cshuffle.hpp.E3AE2B4768599A5C.idx | Bin 0 -> 14556 bytes
 ...km_kn_mn_instance.cpp.7921F720A3C738B2.idx | Bin 0 -> 9590 bytes
 ...km_nk_mn_instance.cpp.67A4E982DE011DE5.idx | Bin 0 -> 9590 bytes
 ...mk_kn_mn_instance.cpp.18986AD46051D542.idx | Bin 0 -> 9590 bytes
 ...mk_nk_mn_instance.cpp.93B77417ED344C2F.idx | Bin 0 -> 8356 bytes
 ...evice_gemm_splitk.hpp.CFD1B5B13FAD53CD.idx | Bin 0 -> 932 bytes
 ...vice_gemm_streamk.hpp.75A52E82DB5C165B.idx | Bin 0 -> 924 bytes
 .../device_gemm_wmma.hpp.32E8352ACBB1259F.idx | Bin 0 -> 13728 bytes
 ...km_kn_mn_instance.cpp.48B71A3DBBF3DF70.idx | Bin 0 -> 6924 bytes
 ...km_nk_mn_instance.cpp.9F95D8CF713B4DFB.idx | Bin 0 -> 6924 bytes
 ...mk_kn_mn_instance.cpp.4656D94005983107.idx | Bin 0 -> 6960 bytes
 ...mk_nk_mn_instance.cpp.92DF213C04CE83E0.idx | Bin 0 -> 6924 bytes
 .../device_gemm_xdl.hpp.8498EA503EEB08EE.idx  | Bin 0 -> 5390 bytes
 ...mk_nk_mn_instance.cpp.F73590DBD2CB1EF5.idx | Bin 0 -> 16360 bytes
 ...km_kn_mn_instance.cpp.544964C8084D9197.idx | Bin 0 -> 6928 bytes
 ...km_nk_mn_instance.cpp.6E89B9B9359CE514.idx | Bin 0 -> 6928 bytes
 ...mk_kn_mn_instance.cpp.2346E4E4EAE3DC7F.idx | Bin 0 -> 2116 bytes
 ...mk_nk_mn_instance.cpp.DCA2A91E3CF88135.idx | Bin 0 -> 2116 bytes
 ...km_kn_mn_instance.cpp.52C4510F756E0106.idx | Bin 0 -> 19630 bytes
 ...km_nk_mn_instance.cpp.11CF3D24C90FB35A.idx | Bin 0 -> 19630 bytes
 ...mk_kn_mn_instance.cpp.3180F054887BBB56.idx | Bin 0 -> 2114 bytes
 ...mk_nk_mn_instance.cpp.CA124F6888DE0DFE.idx | Bin 0 -> 2114 bytes
 ...km_kn_mn_instance.cpp.5B0B95461D582E5E.idx | Bin 0 -> 6846 bytes
 ...km_nk_mn_instance.cpp.144394AC7C436910.idx | Bin 0 -> 6846 bytes
 ...mk_kn_mn_instance.cpp.9C26807545ECBC7F.idx | Bin 0 -> 6846 bytes
 ...mk_nk_mn_instance.cpp.48B6F9C0E69BB63A.idx | Bin 0 -> 5954 bytes
 ...km_kn_mn_instance.cpp.FF4791DC4DC38C18.idx | Bin 0 -> 6970 bytes
 ...km_nk_mn_instance.cpp.6D261DF3AD15EA25.idx | Bin 0 -> 6970 bytes
 ..._default_instance.cpp.FA9B7FC9807074DA.idx | Bin 0 -> 1562 bytes
 ...kn_mn_v1_instance.hpp.E10E580D456628E1.idx | Bin 0 -> 6634 bytes
 ..._default_instance.cpp.030C33F297B655BC.idx | Bin 0 -> 1570 bytes
 ...nterwave_instance.hpp.C8A37285FABA85B1.idx | Bin 0 -> 6640 bytes
 ...e_padded_instance.cpp.1333260C2CA69DB3.idx | Bin 0 -> 1572 bytes
 ...1_padded_instance.cpp.B5F4680BE43A28CE.idx | Bin 0 -> 1564 bytes
 ..._default_instance.cpp.4093668F9B2A46D0.idx | Bin 0 -> 1564 bytes
 ...kn_mn_v2_instance.hpp.A50E70F0D93B6F8E.idx | Bin 0 -> 6634 bytes
 ...2_padded_instance.cpp.74A0C6504550F5D4.idx | Bin 0 -> 1564 bytes
 ...mk_nk_mn_instance.cpp.A977355E082D2CA2.idx | Bin 0 -> 6034 bytes
 ...km_kn_mn_instance.cpp.2559A4A420F11829.idx | Bin 0 -> 6900 bytes
 ...km_nk_mn_instance.cpp.513C881E10886D27.idx | Bin 0 -> 6900 bytes
 ...mk_kn_mn_instance.cpp.43610B5AFBA5C7DA.idx | Bin 0 -> 6900 bytes
 ...mk_nk_mn_instance.cpp.448FFB8EF559CCD0.idx | Bin 0 -> 6006 bytes
 ...mk_nk_mn_instance.cpp.6CE81023C2E130B0.idx | Bin 0 -> 5754 bytes
 ...km_kn_mn_instance.cpp.62E7711AB97BA061.idx | Bin 0 -> 2640 bytes
 ...km_nk_mn_instance.cpp.FC50F49FB1ACD418.idx | Bin 0 -> 2640 bytes
 ...mk_kn_mn_instance.cpp.29C6735FE1834089.idx | Bin 0 -> 2562 bytes
 ...mk_nk_mn_instance.cpp.2E316B49437298E2.idx | Bin 0 -> 2640 bytes
 ...gemm_xdl_cshuffle.hpp.54159ABC735BA0BB.idx | Bin 0 -> 5924 bytes
 ...e_lds_direct_load.hpp.C2FB5C5C6E8BC96E.idx | Bin 0 -> 5968 bytes
 ...m_xdl_cshuffle_v2.hpp.6D300E1E67F0C8FA.idx | Bin 0 -> 4660 bytes
 ...km_kn_mn_instance.cpp.9F68ED656E988D30.idx | Bin 0 -> 4286 bytes
 ...km_nk_mn_instance.cpp.FA79EAE16FB6F44E.idx | Bin 0 -> 4286 bytes
 ...mk_kn_mn_instance.cpp.F19AC2A6D8DD58E6.idx | Bin 0 -> 4286 bytes
 ...mk_nk_mn_instance.cpp.478DBC5276E71817.idx | Bin 0 -> 5610 bytes
 ...km_kn_mn_instance.cpp.1759B855C9E56265.idx | Bin 0 -> 3410 bytes
 ...km_nk_mn_instance.cpp.9CBE0B752A41740B.idx | Bin 0 -> 3410 bytes
 ...mk_kn_mn_instance.cpp.2A9A70463B0B1D69.idx | Bin 0 -> 3410 bytes
 ...mk_nk_mn_instance.cpp.E46BEEC813D33535.idx | Bin 0 -> 4736 bytes
 ...ayernorm_cshuffle.hpp.7536F92E23B638A6.idx | Bin 0 -> 12868 bytes
 ...mm_xdl_skip_b_lds.hpp.31130252724E07CF.idx | Bin 0 -> 9254 bytes
 ..._splitk_c_shuffle.hpp.D2F306B1868E426C.idx | Bin 0 -> 6604 bytes
 ...e_lds_direct_load.hpp.C1DEA5CF04C380FB.idx | Bin 0 -> 6736 bytes
 ...km_kn_mn_instance.cpp.5C4328570228C900.idx | Bin 0 -> 4672 bytes
 ...km_nk_mn_instance.cpp.8E11BCA1E8B40C54.idx | Bin 0 -> 4672 bytes
 ...mk_kn_mn_instance.cpp.02602D130B49C5E9.idx | Bin 0 -> 7486 bytes
 ...mk_nk_mn_instance.cpp.76573A416742936A.idx | Bin 0 -> 6586 bytes
 ...km_kn_mn_instance.cpp.80C5AB0FEC7DFC63.idx | Bin 0 -> 4724 bytes
 ...km_nk_mn_instance.cpp.073EE92A5EAA844A.idx | Bin 0 -> 4724 bytes
 ...kn_mn_v1_instance.cpp.98C08C6647949567.idx | Bin 0 -> 8894 bytes
 ...nterwave_instance.cpp.436428D0A84C0AFE.idx | Bin 0 -> 8794 bytes
 ...rregular_instance.cpp.5798AD2D53C823E4.idx | Bin 0 -> 8382 bytes
 ...rregular_instance.cpp.34C01D1CCA40C5F9.idx | Bin 0 -> 8376 bytes
 ...kn_mn_v2_instance.cpp.2C6EF29B32E3350B.idx | Bin 0 -> 7560 bytes
 ...rregular_instance.cpp.483FB0E67DC82A12.idx | Bin 0 -> 8376 bytes
 ...nk_mn_v1_instance.cpp.04904E13881A008D.idx | Bin 0 -> 7414 bytes
 ...nterwave_instance.cpp.BF7C969BEF21A812.idx | Bin 0 -> 7058 bytes
 ...rregular_instance.cpp.11676556038472A5.idx | Bin 0 -> 8382 bytes
 ...rregular_instance.cpp.E685A8514CD614D5.idx | Bin 0 -> 8376 bytes
 ...nk_mn_v2_instance.cpp.02D5EB3BF76A9EDC.idx | Bin 0 -> 6614 bytes
 ...rregular_instance.cpp.235E5C5B653C03E0.idx | Bin 0 -> 8376 bytes
 ...km_kn_mn_instance.cpp.37CD0D23BDED400B.idx | Bin 0 -> 4654 bytes
 ...km_nk_mn_instance.cpp.2924213EB866E058.idx | Bin 0 -> 4654 bytes
 ...rregular_instance.cpp.3FB0D5B09C446777.idx | Bin 0 -> 4058 bytes
 ...kn_mn_v1_instance.cpp.7D14F328651204D1.idx | Bin 0 -> 8982 bytes
 ...nterwave_instance.cpp.8A9B71EB3B2C8B82.idx | Bin 0 -> 8880 bytes
 ...kn_mn_v2_instance.cpp.CF381829133E9AFB.idx | Bin 0 -> 7646 bytes
 ...n_kpb128_instance.cpp.117C2517C948C007.idx | Bin 0 -> 4898 bytes
 ...nk_mn_v1_instance.cpp.C13174B8631F1EA5.idx | Bin 0 -> 7502 bytes
 ...nterwave_instance.cpp.8596E811350866B4.idx | Bin 0 -> 7144 bytes
 ...nk_mn_v2_instance.cpp.09DE6214A3D6FEFC.idx | Bin 0 -> 6700 bytes
 ...km_kn_mn_instance.cpp.515130D1DB38B8AB.idx | Bin 0 -> 4724 bytes
 ...km_nk_mn_instance.cpp.252C84373128A03D.idx | Bin 0 -> 4724 bytes
 ...mk_kn_mn_instance.cpp.6BA0BD5269CBABBA.idx | Bin 0 -> 4442 bytes
 ...mk_nk_mn_instance.cpp.CFD1881950858A56.idx | Bin 0 -> 5856 bytes
 ...km_kn_mn_instance.cpp.9CC79FF6A90215E2.idx | Bin 0 -> 4654 bytes
 ...km_nk_mn_instance.cpp.577D62F548B80AD3.idx | Bin 0 -> 4654 bytes
 ...kn_mn_v1_instance.cpp.FC0BE921C33562AA.idx | Bin 0 -> 8982 bytes
 ...nterwave_instance.cpp.5E4A00F1466A4D4D.idx | Bin 0 -> 8878 bytes
 ...kn_mn_v2_instance.cpp.64CDFD8CF1246F3A.idx | Bin 0 -> 7646 bytes
 ...mk_nk_mn_instance.cpp.38A1F46E05CA49C4.idx | Bin 0 -> 6148 bytes
 ...mk_nk_mn_instance.cpp.6DFDC664FAC2C818.idx | Bin 0 -> 7952 bytes
 ..._gemm_xdl_streamk.hpp.424C1309002226DA.idx | Bin 0 -> 5558 bytes
 ...mk_kn_mn_instance.cpp.16580B63D7D6888C.idx | Bin 0 -> 6992 bytes
 ...letmodel_cshuffle.hpp.6149EF88F8FEFF53.idx | Bin 0 -> 9082 bytes
 ...action_multiple_d.hpp.D8736023CD2927DC.idx | Bin 0 -> 1928 bytes
 ...le_d_xdl_cshuffle.hpp.A6618A549365EFCF.idx | Bin 0 -> 15300 bytes
 ...nwk_bf16_instance.cpp.A4DBB4F8B2C78E56.idx | Bin 0 -> 1656 bytes
 ...gnwk_f16_instance.cpp.315EC2A3A6D6E638.idx | Bin 0 -> 1642 bytes
 ...gnwk_f32_instance.cpp.AA168C235A47B299.idx | Bin 0 -> 1642 bytes
 ...nwk_bf16_instance.cpp.FA613A1A341E3811.idx | Bin 0 -> 1822 bytes
 ...gnwk_f16_instance.cpp.F48B669764788B12.idx | Bin 0 -> 1820 bytes
 ...gnwk_f32_instance.cpp.B296FF7D04FDFC90.idx | Bin 0 -> 1820 bytes
 ...nwk_int8_instance.cpp.DDED95C8C9D22159.idx | Bin 0 -> 1820 bytes
 ..._1x1s1p0_instance.cpp.15D71B57C7DA4CC3.idx | Bin 0 -> 1630 bytes
 ...nhwk_f16_instance.cpp.1D323BBA44E4F194.idx | Bin 0 -> 1624 bytes
 ..._1x1s1p0_instance.cpp.900C533B3DB605B7.idx | Bin 0 -> 1628 bytes
 ...gnhwk_i8_instance.cpp.5AFEEB9815EAB2BB.idx | Bin 0 -> 1622 bytes
 ..._1x1s1p0_instance.cpp.BC2C6EDB175EB210.idx | Bin 0 -> 1630 bytes
 ...hwgk_f16_instance.cpp.8A9EF4FB27AD09E7.idx | Bin 0 -> 1624 bytes
 ..._1x1s1p0_instance.cpp.DE2A54B24DD9CFDF.idx | Bin 0 -> 1628 bytes
 ...nhwgk_i8_instance.cpp.EBA14072309C61B4.idx | Bin 0 -> 1622 bytes
 ...hwk_bf16_instance.cpp.5B618BE41EFBBD23.idx | Bin 0 -> 1714 bytes
 ...nhwk_f16_instance.cpp.6161CF2711103FB3.idx | Bin 0 -> 1714 bytes
 ...nhwk_f32_instance.cpp.77ED46D094B0B411.idx | Bin 0 -> 1714 bytes
 ...wgk_bf16_instance.cpp.07FF698C2C4EA5B3.idx | Bin 0 -> 1714 bytes
 ...hwgk_f16_instance.cpp.1FAF079E735BA4BC.idx | Bin 0 -> 1714 bytes
 ...hwgk_f32_instance.cpp.FD07473382303B16.idx | Bin 0 -> 1714 bytes
 ...hwk_bf16_instance.cpp.B6F367ED6FE27214.idx | Bin 0 -> 1658 bytes
 ...nhwk_f16_instance.cpp.BE1B4F87BC348C49.idx | Bin 0 -> 1642 bytes
 ...nhwk_f32_instance.cpp.F02158140F1CDE9D.idx | Bin 0 -> 1642 bytes
 ...wgk_bf16_instance.cpp.9325EE22E541E919.idx | Bin 0 -> 1658 bytes
 ...hwgk_f16_instance.cpp.BBF8A6B4ABA52D27.idx | Bin 0 -> 1642 bytes
 ...hwgk_f32_instance.cpp.CE7231939A842BC9.idx | Bin 0 -> 1642 bytes
 ...16_1x1p0_instance.cpp.5A5CAFE59DAED728.idx | Bin 0 -> 1614 bytes
 ..._1x1s1p0_instance.cpp.B9B0155CE4F94655.idx | Bin 0 -> 1616 bytes
 ...nhwk_f16_instance.cpp.F1D87E46323F05EC.idx | Bin 0 -> 1610 bytes
 ...f16_oddc_instance.cpp.B674F131151301E0.idx | Bin 0 -> 1612 bytes
 ...i8_1x1p0_instance.cpp.E06437315FCA8871.idx | Bin 0 -> 1614 bytes
 ..._1x1s1p0_instance.cpp.159F0712016AF057.idx | Bin 0 -> 1616 bytes
 ...gnhwk_i8_instance.cpp.4C82308AB1780753.idx | Bin 0 -> 1608 bytes
 ..._i8_oddc_instance.cpp.872D44B53EF52F10.idx | Bin 0 -> 1612 bytes
 ...16_1x1p0_instance.cpp.83CDF2CC13C4A415.idx | Bin 0 -> 1614 bytes
 ..._1x1s1p0_instance.cpp.A4A9E7E93B446A25.idx | Bin 0 -> 1616 bytes
 ...hwgk_f16_instance.cpp.729E5DD280A9E536.idx | Bin 0 -> 1610 bytes
 ...f16_oddc_instance.cpp.649F61751C48D48D.idx | Bin 0 -> 1612 bytes
 ...i8_1x1p0_instance.cpp.8A0C33E6B1E03B1D.idx | Bin 0 -> 1614 bytes
 ..._1x1s1p0_instance.cpp.8C4CEC479CC69615.idx | Bin 0 -> 1616 bytes
 ...nhwgk_i8_instance.cpp.D16C834304C1756E.idx | Bin 0 -> 1608 bytes
 ..._i8_oddc_instance.cpp.1399C252DECC9B8A.idx | Bin 0 -> 1612 bytes
 ...hwk_bf16_instance.cpp.3060AD718DAF4F4F.idx | Bin 0 -> 1948 bytes
 ...nhwk_f16_instance.cpp.8A08D10AF6807EE2.idx | Bin 0 -> 1944 bytes
 ...nhwk_f32_instance.cpp.FE154FD36F572C54.idx | Bin 0 -> 1944 bytes
 ...wgk_bf16_instance.cpp.56FD421A51C8C5E1.idx | Bin 0 -> 1946 bytes
 ...hwgk_f16_instance.cpp.C29BF8AC2D59E669.idx | Bin 0 -> 1944 bytes
 ...hwgk_f32_instance.cpp.9341E131FE1393D1.idx | Bin 0 -> 1944 bytes
 ..._1x1s1p0_instance.cpp.11490B4C8E752971.idx | Bin 0 -> 1630 bytes
 ...dhwk_f16_instance.cpp.ED0E8CE7C6A61531.idx | Bin 0 -> 1624 bytes
 ..._1x1s1p0_instance.cpp.5ED435B77281FBCD.idx | Bin 0 -> 1630 bytes
 ...ndhwk_i8_instance.cpp.069AF0356549719D.idx | Bin 0 -> 1622 bytes
 ..._1x1s1p0_instance.cpp.511C2FDB97E265FC.idx | Bin 0 -> 1630 bytes
 ...hwgk_f16_instance.cpp.32DD26A1AFC28FC7.idx | Bin 0 -> 1624 bytes
 ..._1x1s1p0_instance.cpp.72C23D8E55C505F5.idx | Bin 0 -> 1630 bytes
 ...dhwgk_i8_instance.cpp.80E125D05CA137B9.idx | Bin 0 -> 1622 bytes
 ...wgk_bf16_instance.cpp.663BA57C2C8487A1.idx | Bin 0 -> 1790 bytes
 ...hwgk_f16_instance.cpp.FDD27F96EE937A50.idx | Bin 0 -> 1788 bytes
 ...hwgk_f32_instance.cpp.33C6E967995C7F50.idx | Bin 0 -> 1790 bytes
 ...hwk_bf16_instance.cpp.59DE1C78AB7D5F80.idx | Bin 0 -> 1716 bytes
 ...dhwk_f16_instance.cpp.F3F4F9DA4C0F8BDA.idx | Bin 0 -> 1714 bytes
 ...dhwk_f32_instance.cpp.A371B626E37E8E07.idx | Bin 0 -> 1714 bytes
 ...wgk_bf16_instance.cpp.AB40F2DD408B7367.idx | Bin 0 -> 1716 bytes
 ...hwgk_f16_instance.cpp.18F54267E4ABEC46.idx | Bin 0 -> 1714 bytes
 ...hwgk_f32_instance.cpp.192FAAC8A46AFF2C.idx | Bin 0 -> 1714 bytes
 ...p_bf8_f8_instance.cpp.F56F1203139F8DB9.idx | Bin 0 -> 1776 bytes
 ...wgk_bf16_instance.cpp.4C04B7436EBF8811.idx | Bin 0 -> 1730 bytes
 ...hwgk_f16_instance.cpp.D8BA94FD49F7A399.idx | Bin 0 -> 1730 bytes
 ...hwgk_f32_instance.cpp.C14D5D3E331E58A4.idx | Bin 0 -> 1730 bytes
 ..._1x1s1p0_instance.cpp.066E2D36BE87D012.idx | Bin 0 -> 1546 bytes
 ...dhwk_f16_instance.cpp.341F4486D5AB6E5A.idx | Bin 0 -> 1540 bytes
 ..._1x1s1p0_instance.cpp.5A4CE3D80E2AF506.idx | Bin 0 -> 1546 bytes
 ...ndhwk_i8_instance.cpp.403AE32B0F021E71.idx | Bin 0 -> 1538 bytes
 ..._1x1s1p0_instance.cpp.9C8F172FA791F403.idx | Bin 0 -> 1546 bytes
 ...hwgk_f16_instance.cpp.2C419A4DCD2F2168.idx | Bin 0 -> 1540 bytes
 ..._1x1s1p0_instance.cpp.217BA51B10EED1BA.idx | Bin 0 -> 1546 bytes
 ...dhwgk_i8_instance.cpp.CE8AF0B7DC8618CC.idx | Bin 0 -> 1538 bytes
 ...hwk_bf16_instance.cpp.8B802336DBA274D4.idx | Bin 0 -> 1662 bytes
 ...dhwk_f16_instance.cpp.17BBA75937C47055.idx | Bin 0 -> 1644 bytes
 ...dhwk_f32_instance.cpp.9C7E5EE7984A7775.idx | Bin 0 -> 1644 bytes
 ...wgk_bf16_instance.cpp.15DDB04E41B23392.idx | Bin 0 -> 1662 bytes
 ..._bf8_fp8_instance.cpp.C031B76FFFB0A20D.idx | Bin 0 -> 1702 bytes
 ...hwgk_f16_instance.cpp.C7145A60489EBDB5.idx | Bin 0 -> 1644 bytes
 ...hwgk_f32_instance.cpp.C5AE3E9229A8EFB4.idx | Bin 0 -> 1644 bytes
 ...16_1x1p0_instance.cpp.8F52AF019F02B944.idx | Bin 0 -> 1616 bytes
 ..._1x1s1p0_instance.cpp.3C3BEC3EC7808D5A.idx | Bin 0 -> 1618 bytes
 ...dhwk_f16_instance.cpp.2554570633260D87.idx | Bin 0 -> 1610 bytes
 ...f16_oddc_instance.cpp.D8F2815AC5D0D7D6.idx | Bin 0 -> 1614 bytes
 ...i8_1x1p0_instance.cpp.0FC1DC5FE52A0C60.idx | Bin 0 -> 1616 bytes
 ..._1x1s1p0_instance.cpp.0151E7B579E72F93.idx | Bin 0 -> 1618 bytes
 ...ndhwk_i8_instance.cpp.C058D2A1DEE10686.idx | Bin 0 -> 1610 bytes
 ..._i8_oddc_instance.cpp.278F3C311504F9EC.idx | Bin 0 -> 1614 bytes
 ...16_1x1p0_instance.cpp.8DE51E7D513808A9.idx | Bin 0 -> 1616 bytes
 ..._1x1s1p0_instance.cpp.B3BD1E3E3F6E4877.idx | Bin 0 -> 1618 bytes
 ...hwgk_f16_instance.cpp.3955AFFBF9D5A2EF.idx | Bin 0 -> 1610 bytes
 ...f16_oddc_instance.cpp.C97D071C6295D5F3.idx | Bin 0 -> 1614 bytes
 ...i8_1x1p0_instance.cpp.E621A4FA60DE216B.idx | Bin 0 -> 1616 bytes
 ..._1x1s1p0_instance.cpp.06F3665DC4884ADD.idx | Bin 0 -> 1618 bytes
 ...dhwgk_i8_instance.cpp.8A48A77FE8BA877A.idx | Bin 0 -> 1610 bytes
 ..._i8_oddc_instance.cpp.85809BA7EA818954.idx | Bin 0 -> 1614 bytes
 ...wgk_bf16_instance.cpp.FB2E0A9CD009454C.idx | Bin 0 -> 1914 bytes
 ...hwgk_f16_instance.cpp.21AF094B4B66A48D.idx | Bin 0 -> 1912 bytes
 ...hwgk_f32_instance.cpp.D78BE8BB81161A14.idx | Bin 0 -> 1914 bytes
 ...wgk_int8_instance.cpp.143C90F3DEDC2341.idx | Bin 0 -> 1912 bytes
 ...hwk_bf16_instance.cpp.426F56794F68AE44.idx | Bin 0 -> 1824 bytes
 ...dhwk_f16_instance.cpp.1C2FE2C88C10C9D6.idx | Bin 0 -> 1824 bytes
 ...dhwk_f32_instance.cpp.BB1A75845E7DE245.idx | Bin 0 -> 1824 bytes
 ...hwk_int8_instance.cpp.A1E84AD887D755DB.idx | Bin 0 -> 1824 bytes
 ...wgk_bf16_instance.cpp.43CB1A6BDABED53D.idx | Bin 0 -> 1824 bytes
 ...comp_fp8_instance.cpp.336E2C49F8565116.idx | Bin 0 -> 1856 bytes
 ...hwgk_f16_instance.cpp.117E24CA3DC21528.idx | Bin 0 -> 1824 bytes
 ...hwgk_f32_instance.cpp.59CFAC251237A8A2.idx | Bin 0 -> 1824 bytes
 ...wgk_int8_instance.cpp.B76F39C8DBBE2FD2.idx | Bin 0 -> 1824 bytes
 ...wgk_bf16_instance.cpp.9F991AFB3119C4B9.idx | Bin 0 -> 1842 bytes
 ...hwgk_f16_instance.cpp.A3477CD2DF481BB7.idx | Bin 0 -> 1842 bytes
 ...hwgk_f32_instance.cpp.C60AD9A6203D4F40.idx | Bin 0 -> 1842 bytes
 ...wgk_int8_instance.cpp.B74C847DC291EB3D.idx | Bin 0 -> 1842 bytes
 ...wgk_bf16_instance.cpp.91739D7EC5D8C4E5.idx | Bin 0 -> 1868 bytes
 ...hwgk_f16_instance.cpp.612D072BD9DA9C78.idx | Bin 0 -> 1868 bytes
 ...hwgk_f32_instance.cpp.91C96AAC7F5772DF.idx | Bin 0 -> 1868 bytes
 ...wgk_int8_instance.cpp.A06BD8EFCD977C2D.idx | Bin 0 -> 1868 bytes
 ...wgk_bf16_instance.cpp.0F35369ACDDEB6D2.idx | Bin 0 -> 1998 bytes
 ...hwgk_f16_instance.cpp.3DF5C2CC3145E799.idx | Bin 0 -> 2006 bytes
 ...hwgk_f32_instance.cpp.B3BE2F1E169224A4.idx | Bin 0 -> 1998 bytes
 ...wgk_int8_instance.cpp.ED09CF7C1628C3B6.idx | Bin 0 -> 2006 bytes
 ...d_data_multiple_d.hpp.CB0E4FDBA2354B83.idx | Bin 0 -> 1476 bytes
 ...e_d_wmma_cshuffle.hpp.62AEEAEB4FE47662.idx | Bin 0 -> 14888 bytes
 ...d_xdl_cshuffle_v1.hpp.DA7CD25895EF8940.idx | Bin 0 -> 16464 bytes
 ...wmma_f16_instance.hpp.FE4A66EFBF344CEC.idx | Bin 0 -> 9704 bytes
 ..._wmma_i8_instance.hpp.22BF4E121F8B1D63.idx | Bin 0 -> 9704 bytes
 ...bilinear_instance.hpp.8935210F01B214E2.idx | Bin 0 -> 16388 bytes
 ...data_xdl_instance.hpp.7A273F666FE77357.idx | Bin 0 -> 22174 bytes
 ...dl_scale_instance.hpp.A306BD940C6CC8E3.idx | Bin 0 -> 15542 bytes
 ...d_conv_bwd_weight.hpp.FB767EE60994A05E.idx | Bin 0 -> 1098 bytes
 ...ght_wmma_cshuffle.hpp.2CEBA1D92F4EDC1A.idx | Bin 0 -> 19286 bytes
 ...ght_wmma_instance.hpp.D8A887BE715EE7ED.idx | Bin 0 -> 10682 bytes
 ...ight_xdl_cshuffle.hpp.7B0144E3F93B6769.idx | Bin 0 -> 32782 bytes
 ...ight_xdl_instance.hpp.DF771DCE13775534.idx | Bin 0 -> 17894 bytes
 ..._fwd_multiple_abd.hpp.ED9472BD269F359D.idx | Bin 0 -> 2508 bytes
 ..._abd_xdl_cshuffle.hpp.52A5FB3B86490259.idx | Bin 0 -> 20290 bytes
 ...nv_fwd_multiple_d.hpp.79D64EC0672BD2D2.idx | Bin 0 -> 874 bytes
 ...iple_d_multiple_r.hpp.4DCB681EEBE1DA60.idx | Bin 0 -> 1706 bytes
 ...le_r_xdl_cshuffle.hpp.CF5C02CB8B4910C1.idx | Bin 0 -> 19906 bytes
 ...e_d_wmma_cshuffle.hpp.EC51DA2544106248.idx | Bin 0 -> 18826 bytes
 ...fwd_wmma_instance.hpp.377DCBF55D8BBCBB.idx | Bin 0 -> 12338 bytes
 ...bilinear_instance.hpp.4827161718A2E399.idx | Bin 0 -> 21842 bytes
 ..._fwd_xdl_instance.hpp.DE209B7BB2124A4A.idx | Bin 0 -> 25076 bytes
 ...dl_scale_instance.hpp.F388F6485863F7A4.idx | Bin 0 -> 20762 bytes
 ...leadd_ab_instance.hpp.043570DA9A94B804.idx | Bin 0 -> 7636 bytes
 ...add_relu_instance.hpp.8003C44E683CA082.idx | Bin 0 -> 6892 bytes
 ...rouped_conv_utils.hpp.82861A708A167591.idx | Bin 0 -> 4012 bytes
 ...vice_grouped_gemm.hpp.96241388E9D6D4CD.idx | Bin 0 -> 1524 bytes
 ...km_kn_mn_instance.cpp.B5F90340138DCBDD.idx | Bin 0 -> 11156 bytes
 ...km_nk_mn_instance.cpp.BC9F6A61DFACFE50.idx | Bin 0 -> 11156 bytes
 ...mk_kn_mn_instance.cpp.F7CCE132C2B5DC0C.idx | Bin 0 -> 10826 bytes
 ...mk_nk_mn_instance.cpp.9DF3794E443669AC.idx | Bin 0 -> 11156 bytes
 ...ped_gemm_fixed_nk.hpp.6F8F8F18341308F9.idx | Bin 0 -> 1692 bytes
 ...tmax_gemm_permute.hpp.3C2429E11C534BCE.idx | Bin 0 -> 2222 bytes
 ...mute_xdl_cshuffle.hpp.75AE4D40F988002C.idx | Bin 0 -> 15912 bytes
 ...ouped_gemm_splitk.hpp.6F08C0BD625D58A2.idx | Bin 0 -> 670 bytes
 ..._grouped_gemm_xdl.hpp.46AD692E57980C90.idx | Bin 0 -> 14358 bytes
 ...km_kn_mn_instance.cpp.37D37DC699ECCA33.idx | Bin 0 -> 11062 bytes
 ...km_nk_mn_instance.cpp.1D6C6BE3C31CA169.idx | Bin 0 -> 11062 bytes
 ...mk_kn_mn_instance.cpp.24FD07A2963099BE.idx | Bin 0 -> 11062 bytes
 ...mk_nk_mn_instance.cpp.32B3D11C83B5AEA6.idx | Bin 0 -> 11062 bytes
 ...gemm_xdl_fixed_nk.hpp.4C6A715EB09F77E1.idx | Bin 0 -> 17172 bytes
 ...mk_kn_mn_instance.cpp.47280D4E92EC5D1F.idx | Bin 0 -> 5776 bytes
 ...mk_nk_mn_instance.cpp.1DAF74A63A4FD2D6.idx | Bin 0 -> 6768 bytes
 ...mk_kn_mn_instance.cpp.CD8E0FEB6A298EEB.idx | Bin 0 -> 5944 bytes
 ...mk_nk_mn_instance.cpp.C65F333685492D7B.idx | Bin 0 -> 6938 bytes
 ...mk_kn_mn_instance.cpp.D0AB01FA41C26CA6.idx | Bin 0 -> 7272 bytes
 ...mk_nk_mn_instance.cpp.43793441438DD042.idx | Bin 0 -> 7270 bytes
 ...mk_kn_mn_instance.cpp.2F31492B9345223B.idx | Bin 0 -> 5692 bytes
 ...mk_nk_mn_instance.cpp.A7598EAA8CFFD492.idx | Bin 0 -> 6686 bytes
 ...mk_kn_mn_instance.cpp.FC712CF5F768A94B.idx | Bin 0 -> 5778 bytes
 ...mk_nk_mn_instance.cpp.6FC4AF50AF38A41A.idx | Bin 0 -> 6770 bytes
 ...mk_kn_mn_instance.cpp.B22CC0B7C42EE2B4.idx | Bin 0 -> 5774 bytes
 ...mk_nk_mn_instance.cpp.E11E98EDBDF90D0B.idx | Bin 0 -> 6766 bytes
 ...l_splitk_cshuffle.hpp.CE155CFA093515AD.idx | Bin 0 -> 12684 bytes
 ...mk_kn_mn_instance.cpp.5121BAECA46B0C23.idx | Bin 0 -> 4952 bytes
 ...rregular_instance.cpp.1597C09D10EC33F9.idx | Bin 0 -> 24842 bytes
 ...mk_nk_mn_instance.cpp.FBCAD14B8CE81D3F.idx | Bin 0 -> 6612 bytes
 ...rregular_instance.cpp.F03BAFFDB1120A94.idx | Bin 0 -> 8954 bytes
 ...rregular_instance.cpp.EADFE46F5C6C50DA.idx | Bin 0 -> 24924 bytes
 ...rregular_instance.cpp.5656416FBF150D8D.idx | Bin 0 -> 24924 bytes
 ...data_f32_instance.cpp.6CBCB642104CF767.idx | Bin 0 -> 1390 bytes
 ...beta_f32_instance.cpp.59A23F12AA5E3659.idx | Bin 0 -> 1394 bytes
 ..._fwd_f16_instance.cpp.E5229B139F1A9B82.idx | Bin 0 -> 1572 bytes
 ..._fwd_f32_instance.cpp.D9D1D3F585AA6CD7.idx | Bin 0 -> 1572 bytes
 ..._f32_f16_instance.cpp.4F6E8B2271D69F1C.idx | Bin 0 -> 1592 bytes
 ...wish_f16_instance.cpp.3151733ECD55CFAB.idx | Bin 0 -> 1574 bytes
 ...wish_f32_instance.cpp.7C71C8C8383AE2D3.idx | Bin 0 -> 1572 bytes
 ...ndhwc_3d_instance.cpp.B02146942667831D.idx | Bin 0 -> 2168 bytes
 ...gnhwc_2d_instance.cpp.BAF41643A25FDB03.idx | Bin 0 -> 2168 bytes
 ..._gnwc_1d_instance.cpp.3EF12405C93313FD.idx | Bin 0 -> 2166 bytes
 ...ge_to_column_impl.hpp.C46D2CA10F5E5642.idx | Bin 0 -> 10726 bytes
 ...o_column_instance.hpp.94D5335C0371893F.idx | Bin 0 -> 2682 bytes
 ...dhwgc_3d_instance.cpp.709987A54287EE7C.idx | Bin 0 -> 2168 bytes
 ...nhwgc_2d_instance.cpp.E16A024C295F734A.idx | Bin 0 -> 2168 bytes
 ..._nwgc_1d_instance.cpp.D56D4C0735B48980.idx | Bin 0 -> 2166 bytes
 ...data_f16_instance.cpp.7967A43F40615422.idx | Bin 0 -> 1384 bytes
 ...data_f32_instance.cpp.BCCFAE80AA603EB7.idx | Bin 0 -> 1392 bytes
 ...beta_f16_instance.cpp.756B8C951E2AA179.idx | Bin 0 -> 1388 bytes
 ...beta_f32_instance.cpp.4E9CDD5C09C49289.idx | Bin 0 -> 1396 bytes
 ..._fwd_f16_instance.cpp.572C171AE3FAAC3B.idx | Bin 0 -> 1574 bytes
 ..._fwd_f32_instance.cpp.C34D1A2542CE24BA.idx | Bin 0 -> 1574 bytes
 ..._fwd_f16_instance.cpp.4D368D0AFAADE23F.idx | Bin 0 -> 1572 bytes
 ..._fwd_f32_instance.cpp.82CA4512D672E7A2.idx | Bin 0 -> 1572 bytes
 ...hwc_bf16_instance.cpp.59F686652075C1F3.idx | Bin 0 -> 1838 bytes
 ...dhwc_f16_instance.cpp.075B2EC4AC466F2A.idx | Bin 0 -> 1838 bytes
 ...dhwc_f32_instance.cpp.2CC50FB6EE948F2B.idx | Bin 0 -> 1838 bytes
 ...vice_max_pool_bwd.hpp.67E96E50742F2EC4.idx | Bin 0 -> 844 bytes
 ...bwd_bf16_instance.cpp.142DD51C4CE34BA2.idx | Bin 0 -> 1344 bytes
 ..._bwd_f16_instance.cpp.DCBDF5FDAE579133.idx | Bin 0 -> 1344 bytes
 ..._bwd_f32_instance.cpp.EA3D33FB1230F655.idx | Bin 0 -> 1344 bytes
 ...max_pool_bwd_impl.hpp.85737AA1C70FEFBE.idx | Bin 0 -> 7006 bytes
 .../device_memory.cpp.A744015D851F9812.idx    | Bin 0 -> 2700 bytes
 .../device_memory.hpp.C2F2D98A33C55529.idx    | Bin 0 -> 1700 bytes
 ...e_multiple_reduce.hpp.188F02C49C64E773.idx | Bin 0 -> 1302 bytes
 ...reduce_multiblock.hpp.BAAA7F55A7B0C61A.idx | Bin 0 -> 12272 bytes
 ...reduce_threadwise.hpp.956BE9BDF24D2435.idx | Bin 0 -> 9686 bytes
 ...lization_bwd_data.hpp.1BAA9B467AEFC28F.idx | Bin 0 -> 1082 bytes
 ...ion_bwd_data_impl.hpp.A627060CAA69A19F.idx | Bin 0 -> 9770 bytes
 ...on_bwd_gamma_beta.hpp.78751175103676B4.idx | Bin 0 -> 1118 bytes
 ...d_gamma_beta_impl.hpp.A5BD375EAE722F5E.idx | Bin 0 -> 10178 bytes
 ...normalization_fwd.hpp.DBDBAF89813C337F.idx | Bin 0 -> 1086 bytes
 ...lization_fwd_impl.hpp.2930CBBD3A2BF4E9.idx | Bin 0 -> 10942 bytes
 ...n_fwd_splitk_impl.hpp.6FD53387232480CC.idx | Bin 0 -> 15510 bytes
 ...ormalize_instance.cpp.FD40702354DAF5AA.idx | Bin 0 -> 3074 bytes
 ..._instance_factory.hpp.115B69A1EE76F903.idx | Bin 0 -> 6784 bytes
 .../device_permute.hpp.47CCE84C97E4E9FC.idx   | Bin 0 -> 962 bytes
 ...vice_permute_impl.hpp.BEE7C34D1958CA2F.idx | Bin 0 -> 5400 bytes
 ...cale_1d_instances.cpp.7EB32958B0512D73.idx | Bin 0 -> 1674 bytes
 ...cale_2d_instances.cpp.2B80C1D2C455DED0.idx | Bin 0 -> 1674 bytes
 ...cale_3d_instances.cpp.07A2A609833FE4D0.idx | Bin 0 -> 1674 bytes
 ...cale_4d_instances.cpp.18341306CF0F5EFD.idx | Bin 0 -> 1674 bytes
 ...cale_5d_instances.cpp.81367B36B3C16096.idx | Bin 0 -> 1674 bytes
 ...cale_6d_instances.cpp.A19C7755BA5CCAF6.idx | Bin 0 -> 1674 bytes
 ...e_scale_instances.hpp.6002ED9F3C6C6BA1.idx | Bin 0 -> 2070 bytes
 ...l2d_fwd_nhwc_nhwc.hpp.8F42F3A58241D669.idx | Bin 0 -> 1690 bytes
 ...d_fwd_ndhwc_ndhwc.hpp.821043598893E6B4.idx | Bin 0 -> 10036 bytes
 .../device_pool_fwd.hpp.048422E56031C936.idx  | Bin 0 -> 1110 bytes
 .../device_prop.hpp.51F1A892B5410B39.idx      | Bin 0 -> 2020 bytes
 ...evice_put_element.hpp.50E2F2788BD7B2E7.idx | Bin 0 -> 788 bytes
 ..._put_element_impl.hpp.CF1E04C6FC237429.idx | Bin 0 -> 3456 bytes
 .../device_reduce.hpp.17BAD15A3D93B7FE.idx    | Bin 0 -> 1132 bytes
 ...ice_reduce_common.hpp.CAE83A095739383D.idx | Bin 0 -> 2026 bytes
 ...e_reduce_instance.hpp.15BDAE27A6CA281A.idx | Bin 0 -> 1438 bytes
 ...nstance_blockwise.hpp.C5FE82438250FD9A.idx | Bin 0 -> 2296 bytes
 ...e_b16_f32_b16_add.cpp.A1938EEEABE8B42A.idx | Bin 0 -> 1062 bytes
 ...e_b16_f32_b16_add.hpp.20070C9325736FAE.idx | Bin 0 -> 280 bytes
 ..._b16_f32_b16_amax.cpp.3CFF0C60C7399DEC.idx | Bin 0 -> 1062 bytes
 ..._b16_f32_b16_amax.hpp.8C9F3AF79C3259FF.idx | Bin 0 -> 282 bytes
 ...e_b16_f32_b16_avg.cpp.39D9A35BA78EDFCA.idx | Bin 0 -> 1064 bytes
 ...e_b16_f32_b16_avg.hpp.D3804A89BDDBD68A.idx | Bin 0 -> 280 bytes
 ...e_b16_f32_b16_max.cpp.26EBD8A22938C32C.idx | Bin 0 -> 1062 bytes
 ...e_b16_f32_b16_max.hpp.EFFF67751118917D.idx | Bin 0 -> 280 bytes
 ...e_b16_f32_b16_min.cpp.82123925B2E47DE5.idx | Bin 0 -> 1062 bytes
 ...e_b16_f32_b16_min.hpp.1AFBE8401A0BB727.idx | Bin 0 -> 280 bytes
 ...b16_f32_b16_norm2.cpp.4AE84C1155A0BD16.idx | Bin 0 -> 1064 bytes
 ...b16_f32_b16_norm2.hpp.E3EB62B3883D54D2.idx | Bin 0 -> 282 bytes
 ..._f16_f16_f16_amax.cpp.116108BDF1BABC19.idx | Bin 0 -> 1062 bytes
 ..._f16_f16_f16_amax.hpp.9716CDC40BFB2514.idx | Bin 0 -> 278 bytes
 ...e_f16_f16_f16_max.cpp.F4D8C3EC38AC4A06.idx | Bin 0 -> 1060 bytes
 ...e_f16_f16_f16_max.hpp.E4FC0C2AEABCC117.idx | Bin 0 -> 278 bytes
 ...e_f16_f16_f16_min.cpp.8D16D62A0EBE3E6E.idx | Bin 0 -> 1060 bytes
 ...e_f16_f16_f16_min.hpp.4404F6E1807E5BD4.idx | Bin 0 -> 278 bytes
 ...e_f16_f32_f16_add.cpp.6312AF648C08A121.idx | Bin 0 -> 1064 bytes
 ...e_f16_f32_f16_add.hpp.70C411CE7496D684.idx | Bin 0 -> 280 bytes
 ...e_f16_f32_f16_avg.cpp.BB89842592147883.idx | Bin 0 -> 1064 bytes
 ...e_f16_f32_f16_avg.hpp.08DE17CE7004DBDC.idx | Bin 0 -> 280 bytes
 ...f16_f32_f16_norm2.cpp.15C86B2B21F27130.idx | Bin 0 -> 1064 bytes
 ...f16_f32_f16_norm2.hpp.152CF392CC9366CD.idx | Bin 0 -> 282 bytes
 ...e_f32_f32_f32_add.cpp.07FADC8850641162.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_add.hpp.B81990FE8C5D8509.idx | Bin 0 -> 276 bytes
 ..._f32_f32_f32_amax.cpp.562409E747DBF8A8.idx | Bin 0 -> 1060 bytes
 ..._f32_f32_f32_amax.hpp.DF4FA5048878CC52.idx | Bin 0 -> 278 bytes
 ...e_f32_f32_f32_avg.cpp.F349F4E1672B2959.idx | Bin 0 -> 1062 bytes
 ...e_f32_f32_f32_avg.hpp.1B257007DC8DF36B.idx | Bin 0 -> 276 bytes
 ...e_f32_f32_f32_max.cpp.58C6C0B25D770B51.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_max.hpp.9E23A16BC92C20DB.idx | Bin 0 -> 278 bytes
 ...e_f32_f32_f32_min.cpp.0B95E75A805DE595.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_min.hpp.B0727F2EA98E1F00.idx | Bin 0 -> 276 bytes
 ...f32_f32_f32_norm2.cpp.FA19AE81B0ABEEED.idx | Bin 0 -> 1064 bytes
 ...f32_f32_f32_norm2.hpp.7E4B9D2EA0AA55ED.idx | Bin 0 -> 278 bytes
 ...e_f32_f64_f32_add.cpp.B429C5C7F2FA95D4.idx | Bin 0 -> 1062 bytes
 ...e_f32_f64_f32_add.hpp.5BB1B8BA3E60EF71.idx | Bin 0 -> 280 bytes
 ...e_f32_f64_f32_avg.cpp.CFF7599DE0E79FBB.idx | Bin 0 -> 1064 bytes
 ...e_f32_f64_f32_avg.hpp.46A8C8C2BE2BA19E.idx | Bin 0 -> 280 bytes
 ...f32_f64_f32_norm2.cpp.384AFFA95058603D.idx | Bin 0 -> 1066 bytes
 ...f32_f64_f32_norm2.hpp.54BEF21DD1A84A95.idx | Bin 0 -> 282 bytes
 ...e_f64_f64_f64_add.cpp.213C81B0FE4954D8.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_add.hpp.496DCAD6F6453198.idx | Bin 0 -> 278 bytes
 ..._f64_f64_f64_amax.cpp.13D835E23741A152.idx | Bin 0 -> 1062 bytes
 ..._f64_f64_f64_amax.hpp.2472648101515DFA.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_avg.cpp.11BCD7B99E3C6DB4.idx | Bin 0 -> 1062 bytes
 ...e_f64_f64_f64_avg.hpp.E5A6055D6ADB74AC.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_max.cpp.3646CB646CAA9902.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_max.hpp.58F762D7856C2DD8.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_min.cpp.1EAEE023D4C34351.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_min.hpp.94217680DDE41BC8.idx | Bin 0 -> 278 bytes
 ...f64_f64_f64_norm2.cpp.5276760A8102F515.idx | Bin 0 -> 1064 bytes
 ...f64_f64_f64_norm2.hpp.52EC3DCBFC6F1222.idx | Bin 0 -> 280 bytes
 ...ise_i8_i32_i8_add.cpp.BC429E8BCB688FC0.idx | Bin 0 -> 1062 bytes
 ...ise_i8_i32_i8_add.hpp.29384EA1BE892314.idx | Bin 0 -> 278 bytes
 ...ise_i8_i32_i8_avg.cpp.C4B64A596D6DF817.idx | Bin 0 -> 1062 bytes
 ...ise_i8_i32_i8_avg.hpp.615A6DAAC49A37A5.idx | Bin 0 -> 280 bytes
 ...ise_i8_i8_i8_amax.cpp.9FC33F8197771FF0.idx | Bin 0 -> 1060 bytes
 ...ise_i8_i8_i8_amax.hpp.CF7CEDDC624CFFAF.idx | Bin 0 -> 276 bytes
 ...wise_i8_i8_i8_max.cpp.20716445BE6CBC5D.idx | Bin 0 -> 1060 bytes
 ...wise_i8_i8_i8_max.hpp.A0110584C215CD9C.idx | Bin 0 -> 274 bytes
 ...wise_i8_i8_i8_min.cpp.D8BEF9441B9F4BC8.idx | Bin 0 -> 1060 bytes
 ...wise_i8_i8_i8_min.hpp.8F13163D01C8AE8F.idx | Bin 0 -> 274 bytes
 ...tance_impl_common.hpp.F2201C19BAE9FFE8.idx | Bin 0 -> 1688 bytes
 ...iblock_atomic_add.hpp.F65D86E6C84EDA60.idx | Bin 0 -> 2304 bytes
 ...d_b16_f32_f32_add.cpp.5E6D995E61ADA8BC.idx | Bin 0 -> 1074 bytes
 ...d_b16_f32_f32_add.hpp.510427928F98A02B.idx | Bin 0 -> 284 bytes
 ...d_b16_f32_f32_avg.cpp.54085188C8447423.idx | Bin 0 -> 1074 bytes
 ...d_b16_f32_f32_avg.hpp.0ADD6329411B78B1.idx | Bin 0 -> 286 bytes
 ...d_f16_f32_f32_add.cpp.4393525F2F4E8DA5.idx | Bin 0 -> 1074 bytes
 ...d_f16_f32_f32_add.hpp.21E4CBB422BF8D82.idx | Bin 0 -> 284 bytes
 ...d_f16_f32_f32_avg.cpp.4BD601E4D7DE370C.idx | Bin 0 -> 1074 bytes
 ...d_f16_f32_f32_avg.hpp.619ED16F95775BEA.idx | Bin 0 -> 286 bytes
 ...d_f32_f32_f32_add.cpp.873C6FA8487AF602.idx | Bin 0 -> 1070 bytes
 ...d_f32_f32_f32_add.hpp.DA5C10F93175C08B.idx | Bin 0 -> 280 bytes
 ...d_f32_f32_f32_avg.cpp.A62B76883CD2E814.idx | Bin 0 -> 1072 bytes
 ...d_f32_f32_f32_avg.hpp.82487B9F10609357.idx | Bin 0 -> 282 bytes
 ...d_f32_f64_f32_add.cpp.979C0C60ED1AF7B3.idx | Bin 0 -> 1074 bytes
 ...d_f32_f64_f32_add.hpp.16AAA28D16A62BE9.idx | Bin 0 -> 284 bytes
 ...d_f32_f64_f32_avg.cpp.3DCB7674F6E295A2.idx | Bin 0 -> 1074 bytes
 ...d_f32_f64_f32_avg.hpp.3C8AE5F8E636FD38.idx | Bin 0 -> 286 bytes
 ...d_f64_f64_f64_add.cpp.EC103046C3FCC222.idx | Bin 0 -> 1070 bytes
 ...d_f64_f64_f64_add.hpp.EC5F0E8B3A5B73CB.idx | Bin 0 -> 280 bytes
 ...d_f64_f64_f64_avg.cpp.F69FD80D45178C79.idx | Bin 0 -> 1072 bytes
 ...d_f64_f64_f64_avg.hpp.F6C8C49B5460F242.idx | Bin 0 -> 282 bytes
 ...stance_threadwise.hpp.C398C657987C23EC.idx | Bin 0 -> 1634 bytes
 ...e_b16_f32_b16_add.cpp.7EFDD1FE32B6F3E3.idx | Bin 0 -> 1064 bytes
 ...e_b16_f32_b16_add.hpp.2D7F2A668C85C670.idx | Bin 0 -> 280 bytes
 ..._b16_f32_b16_amax.cpp.DCF9F78C2573F20F.idx | Bin 0 -> 1064 bytes
 ..._b16_f32_b16_amax.hpp.E64E06F9E993015F.idx | Bin 0 -> 282 bytes
 ...e_b16_f32_b16_avg.cpp.71BA7AE5BF315EB4.idx | Bin 0 -> 1064 bytes
 ...e_b16_f32_b16_avg.hpp.5814F92CA6CF0198.idx | Bin 0 -> 280 bytes
 ...e_b16_f32_b16_max.cpp.D9E364199F88743A.idx | Bin 0 -> 1064 bytes
 ...e_b16_f32_b16_max.hpp.B8A4A04F978ABC79.idx | Bin 0 -> 280 bytes
 ...e_b16_f32_b16_min.cpp.B60965F1C97196EF.idx | Bin 0 -> 1064 bytes
 ...e_b16_f32_b16_min.hpp.97362179A977A149.idx | Bin 0 -> 280 bytes
 ...b16_f32_b16_norm2.cpp.C510348DB0E9EAF4.idx | Bin 0 -> 1066 bytes
 ...b16_f32_b16_norm2.hpp.B82C759E6266A9D7.idx | Bin 0 -> 284 bytes
 ..._f16_f16_f16_amax.cpp.6F6EC6F8F6066B26.idx | Bin 0 -> 1062 bytes
 ..._f16_f16_f16_amax.hpp.2B3A1DBDB134572E.idx | Bin 0 -> 278 bytes
 ...e_f16_f16_f16_max.cpp.501586DF17B9409B.idx | Bin 0 -> 1060 bytes
 ...e_f16_f16_f16_max.hpp.EA77360039638CEE.idx | Bin 0 -> 278 bytes
 ...e_f16_f16_f16_min.cpp.F85FE38D0DCF8C44.idx | Bin 0 -> 1060 bytes
 ...e_f16_f16_f16_min.hpp.D976B4CDF936A58A.idx | Bin 0 -> 278 bytes
 ...e_f16_f32_f16_add.cpp.630A4073E2B46FEA.idx | Bin 0 -> 1064 bytes
 ...e_f16_f32_f16_add.hpp.7BFEE24C60F74EC6.idx | Bin 0 -> 280 bytes
 ...e_f16_f32_f16_avg.cpp.BB154BAE452FBECD.idx | Bin 0 -> 1064 bytes
 ...e_f16_f32_f16_avg.hpp.442706BD47327B2E.idx | Bin 0 -> 280 bytes
 ...f16_f32_f16_norm2.cpp.01982D48C410CB47.idx | Bin 0 -> 1064 bytes
 ...f16_f32_f16_norm2.hpp.D90C9C6069808C78.idx | Bin 0 -> 284 bytes
 ...e_f32_f32_f32_add.cpp.30899F9F46343B0E.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_add.hpp.3C1359A00C377B18.idx | Bin 0 -> 276 bytes
 ..._f32_f32_f32_amax.cpp.F787647EC7B1DE3C.idx | Bin 0 -> 1062 bytes
 ..._f32_f32_f32_amax.hpp.2ABD3DA274EED7C0.idx | Bin 0 -> 278 bytes
 ...e_f32_f32_f32_avg.cpp.0D0748D60784A612.idx | Bin 0 -> 1062 bytes
 ...e_f32_f32_f32_avg.hpp.F51F72472293693C.idx | Bin 0 -> 278 bytes
 ...e_f32_f32_f32_max.cpp.35DDEC8F63798EAB.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_max.hpp.4E3093BB6320EF1D.idx | Bin 0 -> 278 bytes
 ...e_f32_f32_f32_min.cpp.1CE993853CF695D0.idx | Bin 0 -> 1060 bytes
 ...e_f32_f32_f32_min.hpp.448A5294C15B52AE.idx | Bin 0 -> 276 bytes
 ...f32_f32_f32_norm2.cpp.F71A95723F67A7AA.idx | Bin 0 -> 1064 bytes
 ...f32_f32_f32_norm2.hpp.1D44113CF201458F.idx | Bin 0 -> 278 bytes
 ...e_f32_f64_f32_add.cpp.DBE748926DFEDB8B.idx | Bin 0 -> 1062 bytes
 ...e_f32_f64_f32_add.hpp.55EEA55DB1346E7B.idx | Bin 0 -> 280 bytes
 ...e_f32_f64_f32_avg.cpp.E1314404E465BA33.idx | Bin 0 -> 1064 bytes
 ...e_f32_f64_f32_avg.hpp.D9B461C00E68A892.idx | Bin 0 -> 280 bytes
 ...f32_f64_f32_norm2.cpp.C0DCB6010E545A4E.idx | Bin 0 -> 1066 bytes
 ...f32_f64_f32_norm2.hpp.082457B5EE500DB0.idx | Bin 0 -> 284 bytes
 ...e_f64_f64_f64_add.cpp.FF7D2E3E29E0477B.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_add.hpp.774FB922ABF22CDD.idx | Bin 0 -> 278 bytes
 ..._f64_f64_f64_amax.cpp.095F5D6B26D2F29A.idx | Bin 0 -> 1064 bytes
 ..._f64_f64_f64_amax.hpp.5FCEC88A65C41974.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_avg.cpp.BB5B12454481FF04.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_avg.hpp.7B5AB5E79B93DFFB.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_max.cpp.BAF3E66A0EDF3C75.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_max.hpp.390E8CBCF4E86ABF.idx | Bin 0 -> 278 bytes
 ...e_f64_f64_f64_min.cpp.778D73C35496348B.idx | Bin 0 -> 1060 bytes
 ...e_f64_f64_f64_min.hpp.2C7338E3132300C6.idx | Bin 0 -> 278 bytes
 ...f64_f64_f64_norm2.cpp.59008E97BEF36435.idx | Bin 0 -> 1064 bytes
 ...f64_f64_f64_norm2.hpp.4329A8B0483B68A4.idx | Bin 0 -> 280 bytes
 ...ise_i8_i32_i8_add.cpp.509D27379633375F.idx | Bin 0 -> 1062 bytes
 ...ise_i8_i32_i8_add.hpp.7F575A4F4970CC92.idx | Bin 0 -> 278 bytes
 ...ise_i8_i32_i8_avg.cpp.BBCEE5A5D01A00A5.idx | Bin 0 -> 1062 bytes
 ...ise_i8_i32_i8_avg.hpp.FDD16FEA381DDE62.idx | Bin 0 -> 280 bytes
 ...ise_i8_i8_i8_amax.cpp.C6B6423A4D3C4156.idx | Bin 0 -> 1060 bytes
 ...ise_i8_i8_i8_amax.hpp.F8ED464DDA5BD732.idx | Bin 0 -> 276 bytes
 ...wise_i8_i8_i8_max.cpp.C1E70E7984041F16.idx | Bin 0 -> 1060 bytes
 ...wise_i8_i8_i8_max.hpp.1F0CCA2641B90428.idx | Bin 0 -> 274 bytes
 ...wise_i8_i8_i8_min.cpp.05E7FDBEF4E2FEA1.idx | Bin 0 -> 1058 bytes
 ...wise_i8_i8_i8_min.hpp.078E8B6FD22E7DA4.idx | Bin 0 -> 274 bytes
 ...reduce_multiblock.hpp.C0B69E5EA2AC7CE7.idx | Bin 0 -> 12110 bytes
 ...reduce_threadwise.hpp.71DBCBAA668DE14B.idx | Bin 0 -> 8606 bytes
 .../device_softmax.hpp.497B752134CE80EA.idx   | Bin 0 -> 928 bytes
 ...nce_rank3_reduce1.cpp.6DE6521393D18F5C.idx | Bin 0 -> 1570 bytes
 ...nce_rank3_reduce1.hpp.B94CF584D85EABE4.idx | Bin 0 -> 614 bytes
 ...nce_rank3_reduce2.cpp.6002EE43735BD2AD.idx | Bin 0 -> 1570 bytes
 ...nce_rank3_reduce2.hpp.453CC2F7EB0502D0.idx | Bin 0 -> 614 bytes
 ...nce_rank3_reduce3.cpp.62DA65F400167AF3.idx | Bin 0 -> 1568 bytes
 ...nce_rank3_reduce3.hpp.3F8E078976E83385.idx | Bin 0 -> 614 bytes
 ...nce_rank4_reduce1.cpp.22BB990A62BE50C5.idx | Bin 0 -> 1570 bytes
 ...nce_rank4_reduce1.hpp.7F7D38DA481813D7.idx | Bin 0 -> 614 bytes
 ...nce_rank4_reduce2.cpp.AC6F1FB653149557.idx | Bin 0 -> 1570 bytes
 ...nce_rank4_reduce2.hpp.44F668BC92D2FB11.idx | Bin 0 -> 616 bytes
 ...nce_rank4_reduce3.cpp.2BDC7BD673407442.idx | Bin 0 -> 1570 bytes
 ...nce_rank4_reduce3.hpp.1A5543D9F44A4D1E.idx | Bin 0 -> 616 bytes
 ...nce_rank4_reduce4.cpp.E94DE85158A7059C.idx | Bin 0 -> 1568 bytes
 ...nce_rank4_reduce4.hpp.DFA0CA7449260515.idx | Bin 0 -> 614 bytes
 ...f16_instance_type.hpp.D6CB9472D3A0B40A.idx | Bin 0 -> 1588 bytes
 ...nce_rank3_reduce1.cpp.7A9B6FF1C1DE5769.idx | Bin 0 -> 1560 bytes
 ...nce_rank3_reduce1.hpp.901763453949395E.idx | Bin 0 -> 602 bytes
 ...nce_rank3_reduce2.cpp.B1019C659E1C387F.idx | Bin 0 -> 1560 bytes
 ...nce_rank3_reduce2.hpp.E7E5CB6E00C1ACF7.idx | Bin 0 -> 602 bytes
 ...nce_rank3_reduce3.cpp.D8971A21BC00D0FF.idx | Bin 0 -> 1558 bytes
 ...nce_rank3_reduce3.hpp.5132DD2F1F4E6659.idx | Bin 0 -> 600 bytes
 ...nce_rank4_reduce1.cpp.70DAADE138E1FD1C.idx | Bin 0 -> 1560 bytes
 ...nce_rank4_reduce1.hpp.D21E6E14C4987978.idx | Bin 0 -> 602 bytes
 ...nce_rank4_reduce2.cpp.C8298BB69F4C257A.idx | Bin 0 -> 1560 bytes
 ...nce_rank4_reduce2.hpp.E0AE02E34E90B0FC.idx | Bin 0 -> 602 bytes
 ...nce_rank4_reduce3.cpp.25E3FD604E588BD0.idx | Bin 0 -> 1560 bytes
 ...nce_rank4_reduce3.hpp.33AFA8695A15063D.idx | Bin 0 -> 602 bytes
 ...nce_rank4_reduce4.cpp.86DA1C4D6D7DC8E7.idx | Bin 0 -> 1558 bytes
 ...nce_rank4_reduce4.hpp.35C1B364BF55F9FD.idx | Bin 0 -> 600 bytes
 ...f32_instance_type.hpp.7F3C1693216AB251.idx | Bin 0 -> 1584 bytes
 ...vice_softmax_impl.hpp.B9E8182EF8FC3442.idx | Bin 0 -> 8366 bytes
 ..._softmax_instance.hpp.E665F0AD026D2FE2.idx | Bin 0 -> 444 bytes
 ...forward_layernorm.hpp.469278C9F8399B0C.idx | Bin 0 -> 3838 bytes
 ...action_multiple_d.hpp.B9F979311F947532.idx | Bin 0 -> 1342 bytes
 ...le_d_xdl_cshuffle.hpp.43569A70DF974D2E.idx | Bin 0 -> 22282 bytes
 ...ranspose_instance.hpp.9CF832E322CA1530.idx | Bin 0 -> 1608 bytes
 ...pose_instances_3d.cpp.5DC4547776B7C3A0.idx | Bin 0 -> 1728 bytes
 .../index/dpp_gemm.hpp.CD386ABA025C1896.idx   | Bin 0 -> 19250 bytes
 ...ual_reduce_common.hpp.746AA107CF079C11.idx | Bin 0 -> 7180 bytes
 ...reduce_multiblock.cpp.C5E150B4E9B1E7E8.idx | Bin 0 -> 2770 bytes
 ...reduce_threadwise.cpp.7BA56606324C3914.idx | Bin 0 -> 2634 bytes
 .../dynamic_buffer.hpp.9B076D16FB6FDAA3.idx   | Bin 0 -> 10820 bytes
 ...nt_wise_operation.hpp.0715A40D7E0DC380.idx | Bin 0 -> 10376 bytes
 ...lementwise_add_1d.cpp.CD0512697217A13B.idx | Bin 0 -> 3570 bytes
 ...lementwise_add_4d.cpp.6D6ECABFC0DFD709.idx | Bin 0 -> 3870 bytes
 ...yernorm_blockwise.cpp.B1B3F8F277F6314A.idx | Bin 0 -> 6260 bytes
 ...ise_normalization.hpp.B86CF9747F7CBAA9.idx | Bin 0 -> 1404 bytes
 ...ementwise_permute.cpp.780590C9EF3E30CE.idx | Bin 0 -> 4106 bytes
 ...ntwise_permute_3d.cpp.2AC2BFBCE3237215.idx | Bin 0 -> 3762 bytes
 ...e_permute_4D_fp16.cpp.267AD0382A416744.idx | Bin 0 -> 3922 bytes
 ...ermute_4D_fp16_2d.cpp.73F9F8FFA71E623F.idx | Bin 0 -> 3800 bytes
 ...rmute_4D_fp16_col.cpp.427D8FEA072CC243.idx | Bin 0 -> 4752 bytes
 ...rmute_4D_fp16_row.cpp.70A9A9EF321166D7.idx | Bin 0 -> 4210 bytes
 ...rmute_4D_fp32_col.cpp.645D32DD67217BF3.idx | Bin 0 -> 4734 bytes
 ...rmute_4D_fp32_row.cpp.51034D29A216280F.idx | Bin 0 -> 4210 bytes
 .../index/enable_if.hpp.10016CB3379BC2AC.idx  | Bin 0 -> 356 bytes
 .../index/f8_utils.hpp.AB5D9F32110A912A.idx   | Bin 0 -> 3088 bytes
 .../index/fill.hpp.A302158AF84698F8.idx       | Bin 0 -> 4772 bytes
 .../index/functional.hpp.AB02FF3D06BB8256.idx | Bin 0 -> 2350 bytes
 .../functional2.hpp.4855B705916A8A54.idx      | Bin 0 -> 902 bytes
 .../functional3.hpp.730E80629C119B9D.idx      | Bin 0 -> 2212 bytes
 .../functional4.hpp.579A0D5EFC68EB03.idx      | Bin 0 -> 1644 bytes
 .../index/gemm.hpp.344713A6C512A32C.idx       | Bin 0 -> 5262 bytes
 .../index/gemm.hpp.C45C6ACD321C3642.idx       | Bin 0 -> 17426 bytes
 .../index/gemm_add.hpp.0A05FF4E11D910E0.idx   | Bin 0 -> 2284 bytes
 ..._add_add_fastgelu.hpp.C1858AFB6E6189D2.idx | Bin 0 -> 3062 bytes
 ...fastgelu_xdl_bf16.cpp.D697A303BBCD00E1.idx | Bin 0 -> 3228 bytes
 ...fastgelu_xdl_fp16.cpp.BE72E2A90F751444.idx | Bin 0 -> 3226 bytes
 ...fastgelu_xdl_fp32.cpp.1D688E059DA5C6D3.idx | Bin 0 -> 3220 bytes
 ...fastgelu_xdl_int8.cpp.7C822BDA45FD8B5C.idx | Bin 0 -> 3226 bytes
 ..._direct_load_fp32.cpp.DFA0EFAE6F444F2C.idx | Bin 0 -> 3256 bytes
 ...ansquare_xdl_fp16.cpp.784D28B0B387AEB3.idx | Bin 0 -> 9570 bytes
 ...ddsquare_xdl_int8.cpp.C81A82E4BF27D67D.idx | Bin 0 -> 7896 bytes
 ...gemm_add_fastgelu.hpp.7F025B27BB774C23.idx | Bin 0 -> 4170 bytes
 ...gemm_add_multiply.hpp.A076E88A74402B66.idx | Bin 0 -> 3096 bytes
 ...multiply_xdl_fp16.cpp.2190407D709E315C.idx | Bin 0 -> 3128 bytes
 .../gemm_add_relu.hpp.2FD450825E8EDD32.idx    | Bin 0 -> 2290 bytes
 ...elu_add_layernorm.hpp.2503A6C1813DEBB3.idx | Bin 0 -> 3418 bytes
 .../gemm_add_silu.hpp.8120675903700DB9.idx    | Bin 0 -> 2292 bytes
 .../index/gemm_bf16.cpp.B8DB90A206756A9E.idx  | Bin 0 -> 1820 bytes
 ...g1m2n3k1_xdl_fp16.cpp.ACF8DA35700D8D8F.idx | Bin 0 -> 10580 bytes
 ...g1m3n2k1_xdl_fp16.cpp.83BB18F6B66F2289.idx | Bin 0 -> 10502 bytes
 ...rm_xdl_naive_fp16.cpp.6BD99D80213AD3B0.idx | Bin 0 -> 12012 bytes
 ..._xdl_welford_fp16.cpp.AC32388E298FFA30.idx | Bin 0 -> 8538 bytes
 ...ias_relu_xdl_fp16.cpp.74D4B1F74A45E199.idx | Bin 0 -> 7158 bytes
 ...tmax_gemm_permute.cpp.A515B9EF146522D4.idx | Bin 0 -> 10794 bytes
 .../gemm_bilinear.hpp.092CAB87B366F08F.idx    | Bin 0 -> 4822 bytes
 ...bilinear_xdl_fp16.cpp.372A779C3E21DD0E.idx | Bin 0 -> 7866 bytes
 .../gemm_dpp_fp16.cpp.D49004015318A845.idx    | Bin 0 -> 2590 bytes
 ...m_f16_nn_instance.cpp.684811D08F35F1E0.idx | Bin 0 -> 3584 bytes
 ...m_f16_nn_instance.hpp.295DF91B6CFA2E94.idx | Bin 0 -> 1396 bytes
 ...m_f16_nt_instance.cpp.CF459381653536CF.idx | Bin 0 -> 3584 bytes
 ...m_f16_nt_instance.hpp.CF5C609546474B1F.idx | Bin 0 -> 1402 bytes
 ...m_f16_tn_instance.cpp.A27FD0EC622C1968.idx | Bin 0 -> 3584 bytes
 ...m_f16_tn_instance.hpp.C57CA23FF5A33CBF.idx | Bin 0 -> 1402 bytes
 ...m_f16_tt_instance.cpp.5D254C2DB3D84C02.idx | Bin 0 -> 3576 bytes
 ...m_f16_tt_instance.hpp.A9C0C844390863CE.idx | Bin 0 -> 1162 bytes
 .../gemm_fastgelu.hpp.BF1B6FA711DF924E.idx    | Bin 0 -> 2688 bytes
 .../index/gemm_fp16.cpp.3FD2C2DD621D7AD4.idx  | Bin 0 -> 1820 bytes
 .../index/gemm_fp32.cpp.6736B177964E2D4B.idx  | Bin 0 -> 1768 bytes
 .../index/gemm_int8.cpp.C4070EF85C2974E6.idx  | Bin 0 -> 1840 bytes
 ...rm_xdl_naive_fp16.cpp.B168DA32D72A32C6.idx | Bin 0 -> 10906 bytes
 ...gemm_max_xdl_bf16.cpp.40A7D230DC3CAD48.idx | Bin 0 -> 4450 bytes
 ...gemm_max_xdl_fp16.cpp.F50CF88BE688DE26.idx | Bin 0 -> 4450 bytes
 ...gemm_max_xdl_fp32.cpp.1AFB0F6127AC986E.idx | Bin 0 -> 4340 bytes
 ...gemm_max_xdl_int8.cpp.A29BA4DC035DE52E.idx | Bin 0 -> 4446 bytes
 ...ansquare_xdl_bf16.cpp.E1D95735FEB3AF62.idx | Bin 0 -> 5084 bytes
 ...ansquare_xdl_fp16.cpp.80F76169F90BD8DC.idx | Bin 0 -> 5084 bytes
 ...ansquare_xdl_fp32.cpp.DAB08B69CF521C93.idx | Bin 0 -> 5076 bytes
 ...ulti_ABD_xdl_fp16.cpp.5859AE2CEFBAACD5.idx | Bin 0 -> 10330 bytes
 ...gemm_multiply_add.hpp.6B6D21EE52049C9D.idx | Bin 0 -> 3304 bytes
 ...antization_common.hpp.A23B3AA4CB2F3E7A.idx | Bin 0 -> 1606 bytes
 .../gemm_reduce_fp16.cpp.EFAED30A7BA2F6AF.idx | Bin 0 -> 1554 bytes
 ...reduce_xdl_common.hpp.BF32975909DCA917.idx | Bin 0 -> 6844 bytes
 ...mm_specialization.hpp.F3D9282B71057FD0.idx | Bin 0 -> 2058 bytes
 .../gemm_splitk.hpp.749B3E67AC6B8B9C.idx      | Bin 0 -> 14040 bytes
 ...andalone_xdl_fp16.cpp.EE84FA2A4697A71B.idx | Bin 0 -> 6576 bytes
 .../gemm_streamk.hpp.B73183768811D4AE.idx     | Bin 0 -> 1450 bytes
 .../index/gemm_util.hpp.1BFB1C62397B2C8C.idx  | Bin 0 -> 7886 bytes
 ...t_f16_tn_instance.cpp.DC996A5B769C800C.idx | Bin 0 -> 4018 bytes
 ...t_f16_tn_instance.hpp.6DF980223599227D.idx | Bin 0 -> 806 bytes
 .../gemm_xdl_bf16.cpp.3DCFB3ED2888403A.idx    | Bin 0 -> 2632 bytes
 ...gemm_xdl_bf16_rtn.cpp.52661E2CDB7295F6.idx | Bin 0 -> 2658 bytes
 ...quantization_int8.cpp.FA91569EB26ACB71.idx | Bin 0 -> 6876 bytes
 .../gemm_xdl_fp16.cpp.FC8CAA78DACFED36.idx    | Bin 0 -> 3214 bytes
 ...gemm_xdl_fp16_fp8.cpp.44C0E7F4DAEEA470.idx | Bin 0 -> 2970 bytes
 .../gemm_xdl_fp16_v2.cpp.BBCC41C35D9D878E.idx | Bin 0 -> 2192 bytes
 .../gemm_xdl_fp64.cpp.326BD4F598F31DB0.idx    | Bin 0 -> 2468 bytes
 .../gemm_xdl_fp8.cpp.67C1EEC83062DA6B.idx     | Bin 0 -> 3028 bytes
 .../gemm_xdl_fp8_bf8.cpp.AEAA96B6834BEA56.idx | Bin 0 -> 3068 bytes
 .../gemm_xdl_int8.cpp.636EBC0B4897BF3A.idx    | Bin 0 -> 2654 bytes
 ...ingle_kernel_fp16.cpp.4306E603B691D828.idx | Bin 0 -> 8404 bytes
 ..._direct_load_fp16.cpp.7BFA821D80385E72.idx | Bin 0 -> 2806 bytes
 ..._direct_load_fp32.cpp.39A0D92907A2D21F.idx | Bin 0 -> 2726 bytes
 ...quantization_int8.cpp.A58C21992E5540BA.idx | Bin 0 -> 5932 bytes
 ...l_skip_b_lds_fp16.cpp.86DD71A6DF5D193B.idx | Bin 0 -> 5942 bytes
 .../gemm_xdl_streamk.cpp.2BC4B9D993728CB3.idx | Bin 0 -> 2616 bytes
 ..._xdl_wavelet_fp16.cpp.5100C57848816EFC.idx | Bin 0 -> 2728 bytes
 ...mory_space_atomic.hpp.C9D4AE0A6DC5899F.idx | Bin 0 -> 2868 bytes
 .../index/get_id.hpp.D8C96396D4A9E3B4.idx     | Bin 0 -> 1060 bytes
 .../index/get_shift.hpp.0A10E058B6B6AEDA.idx  | Bin 0 -> 520 bytes
 ...uction_multiblock.hpp.B962E3920F925D70.idx | Bin 0 -> 5630 bytes
 ...uction_threadwise.hpp.39A048A5A487AC02.idx | Bin 0 -> 3626 bytes
 ...uction_multiblock.hpp.6BD3600B2B4E50BE.idx | Bin 0 -> 9224 bytes
 ...uction_threadwise.hpp.324930ADA73F028A.idx | Bin 0 -> 5210 bytes
 ...m_xdl_cshuffle_v1.hpp.E109D2D6AEE43DDD.idx | Bin 0 -> 15548 bytes
 ...d_xdl_cshuffle_v1.hpp.FE93431287EC85FD.idx | Bin 0 -> 21792 bytes
 ...m_xdl_cshuffle_v1.hpp.A946F68C0A5B7C29.idx | Bin 0 -> 25522 bytes
 ...m_xdl_cshuffle_v1.hpp.EEF70E721362F483.idx | Bin 0 -> 22284 bytes
 ...blockwise_welford.hpp.8F24BDC37125B404.idx | Bin 0 -> 10148 bytes
 ...blockwise_welford.hpp.640DA951EF640017.idx | Bin 0 -> 8814 bytes
 ...se_elementwise_1d.hpp.AC0144CD925F6905.idx | Bin 0 -> 2948 bytes
 ...mentwise_1d_scale.hpp.CF7CE1061BA1DF52.idx | Bin 0 -> 3184 bytes
 ...se_elementwise_2d.hpp.DCACD2865A1A82CC.idx | Bin 0 -> 4026 bytes
 ...se_elementwise_3d.hpp.4B57404BC7344584.idx | Bin 0 -> 5488 bytes
 ..._welford_variance.hpp.61A857F748242277.idx | Bin 0 -> 9138 bytes
 ...e_xdl_cshuffle_v1.hpp.6AEBEC03A02FE520.idx | Bin 0 -> 15286 bytes
 ...gridwise_gemm_dpp.hpp.E915886DCA1D5AA3.idx | Bin 0 -> 14384 bytes
 ..._abd_xdl_cshuffle.hpp.8ED0C5830B90D078.idx | Bin 0 -> 17608 bytes
 ...le_r_xdl_cshuffle.hpp.E96B152CEA3EEC31.idx | Bin 0 -> 15352 bytes
 ...half_xdl_cshuffle.hpp.75601854417950D1.idx | Bin 0 -> 18098 bytes
 ...e_d_wmma_cshuffle.hpp.7EFDFD3F43E2C561.idx | Bin 0 -> 26262 bytes
 ...le_d_xdl_cshuffle.hpp.7B2A45106BD10E35.idx | Bin 0 -> 15078 bytes
 ...e_lds_direct_load.hpp.8A2E3D6921803C3C.idx | Bin 0 -> 19686 bytes
 ...l_splitk_cshuffle.hpp.79ABECDBF9253649.idx | Bin 0 -> 18190 bytes
 ...pipeline_selector.hpp.4F4E408D810A020A.idx | Bin 0 -> 1134 bytes
 ..._gemm_pipeline_v1.hpp.8B594F4BBFC9415D.idx | Bin 0 -> 5874 bytes
 ..._gemm_pipeline_v2.hpp.1C7FE339CB47DFD5.idx | Bin 0 -> 744 bytes
 ..._gemm_pipeline_v3.hpp.D2EF7B80D9AE44CD.idx | Bin 0 -> 566 bytes
 ...ne_v4_direct_load.hpp.7C3BB633588D635D.idx | Bin 0 -> 1944 bytes
 ...e_xdl_cshuffle_v1.hpp.76D93CEF02B94AEF.idx | Bin 0 -> 13164 bytes
 ...le_d_xdl_cshuffle.hpp.F83027A680F1BF3E.idx | Bin 0 -> 19936 bytes
 ...gemm_waveletmodel.hpp.C9F027882C9D4F2B.idx | Bin 0 -> 1354 bytes
 ...ridwise_gemm_wmma.hpp.4F8B2CDD1348D704.idx | Bin 0 -> 20164 bytes
 ...m_xdl_cshuffle_v1.hpp.E796A6FD7BCB2B56.idx | Bin 0 -> 20202 bytes
 ...m_xdl_cshuffle_v2.hpp.E0B3C0FB7BAA0FB3.idx | Bin 0 -> 14774 bytes
 ...rnorm_cshuffle_v1.hpp.FC9F72A612CD78AC.idx | Bin 0 -> 16418 bytes
 ...letmodel_cshuffle.hpp.647408B24FB9D296.idx | Bin 0 -> 12588 bytes
 ...xdlops_bwd_weight.hpp.65AB95ECB27BA1C2.idx | Bin 0 -> 22188 bytes
 ...ops_skip_b_lds_v1.hpp.8CF2D2B713EFE65F.idx | Bin 0 -> 12088 bytes
 ...k_lds_direct_load.hpp.76232B181B348ACC.idx | Bin 0 -> 20972 bytes
 ...mm_xdlops_streamk.hpp.3A8F1B585329CAFC.idx | Bin 0 -> 23270 bytes
 ..._gemm_xdlops_v2r3.hpp.66B1A28D309FB5D4.idx | Bin 0 -> 18192 bytes
 ...emm_xdlops_v2r4r2.hpp.FFFD55E6D49B5037.idx | Bin 0 -> 23224 bytes
 ..._gemm_xdlops_v3r1.hpp.F4E339DE2B052B4F.idx | Bin 0 -> 11174 bytes
 ..._gemm_xdlops_v3r2.hpp.A91ECFDCE099EED9.idx | Bin 0 -> 10860 bytes
 ..._gemm_xdlops_v3r3.hpp.E0F0A6CCA772EDE7.idx | Bin 0 -> 11116 bytes
 ...batchnorm_forward.hpp.4F73C2A5286FBB35.idx | Bin 0 -> 12750 bytes
 ...rm_backward_final.hpp.C7E55A256926B340.idx | Bin 0 -> 9462 bytes
 ...elford_first_half.hpp.25F962D583FA62A6.idx | Bin 0 -> 5070 bytes
 ...rd_final_obsolete.hpp.BEA9938289B2ECCF.idx | Bin 0 -> 10526 bytes
 ...reduce_first_half.hpp.03EA80728C2C5A38.idx | Bin 0 -> 9856 bytes
 ...lization_bwd_data.hpp.43B5DE0C3D25BB20.idx | Bin 0 -> 12016 bytes
 ...on_bwd_gamma_beta.hpp.8028572B7299488D.idx | Bin 0 -> 6944 bytes
 ...on_naive_variance.hpp.9BD4CEF549374198.idx | Bin 0 -> 12224 bytes
 ...lization_selector.hpp.88C73ECDE0BC86D2.idx | Bin 0 -> 1186 bytes
 ...zation_splitk_1st.hpp.6CD10C30DDA5197D.idx | Bin 0 -> 5352 bytes
 ...zation_splitk_2nd.hpp.FE37C91CA5D8EB00.idx | Bin 0 -> 9936 bytes
 ..._welford_variance.hpp.FE77CD9DBE7C3CF0.idx | Bin 0 -> 10872 bytes
 .../gridwise_permute.hpp.E425D88E25D439BC.idx | Bin 0 -> 7052 bytes
 ...se_put_element_1d.hpp.59E84F53787BFD31.idx | Bin 0 -> 2964 bytes
 ..._set_buffer_value.hpp.1407258F85A006CD.idx | Bin 0 -> 962 bytes
 ...iple_buffer_value.hpp.77B0D514858D3743.idx | Bin 0 -> 1274 bytes
 .../gridwise_softmax.hpp.5A728FA36C708E60.idx | Bin 0 -> 6866 bytes
 ...forward_layernorm.hpp.9F9F7C52CFA18397.idx | Bin 0 -> 8282 bytes
 ..._tensor_rearrange.hpp.73B250CAB9F4E6DB.idx | Bin 0 -> 3378 bytes
 ..._half_layernorm2d.hpp.52F6FAA1A972E598.idx | Bin 0 -> 8388 bytes
 ...ias_relu_xdl_fp16.cpp.2CC919AB5B044CF3.idx | Bin 0 -> 2688 bytes
 ...bwd_data_xdl_fp16.cpp.618FF0F87C438165.idx | Bin 0 -> 2630 bytes
 ...d_weight_xdl_bf16.cpp.3B512EE28C055424.idx | Bin 0 -> 2742 bytes
 ...d_weight_xdl_fp16.cpp.35241A5239C4258D.idx | Bin 0 -> 2742 bytes
 ...fp16_comp_bf8_fp8.cpp.BD38E20DB8B3998F.idx | Bin 0 -> 2932 bytes
 ...conv_fwd_xdl_bf16.cpp.2A5915DD526E33D4.idx | Bin 0 -> 3254 bytes
 ...conv_fwd_xdl_fp16.cpp.F0393DD4FF0F531D.idx | Bin 0 -> 3252 bytes
 ...conv_fwd_xdl_fp32.cpp.329BE83F917399A6.idx | Bin 0 -> 3174 bytes
 ...conv_fwd_xdl_int8.cpp.E847303767D0E187.idx | Bin 0 -> 3302 bytes
 ...relu_add_xdl_bf16.cpp.E2C08E4AF953E7F0.idx | Bin 0 -> 2116 bytes
 ...relu_add_xdl_fp16.cpp.9CA386C0ABEF35F6.idx | Bin 0 -> 2116 bytes
 ...relu_add_xdl_fp32.cpp.77C435750269A2BB.idx | Bin 0 -> 2106 bytes
 ...relu_add_xdl_int8.cpp.8BC00DE6D83E6AA0.idx | Bin 0 -> 2116 bytes
 ...conv_fwd_xdl_fp16.cpp.F004628471E9028F.idx | Bin 0 -> 1946 bytes
 ...ion_backward_data.hpp.04038859D7920D2D.idx | Bin 0 -> 13936 bytes
 ...n_backward_weight.hpp.CF4258D0B5FB5557.idx | Bin 0 -> 11580 bytes
 ...nvolution_forward.hpp.3DF8CF239601D4CC.idx | Bin 0 -> 21098 bytes
 .../grouped_gemm.hpp.76EFEC97F28430A5.idx     | Bin 0 -> 4884 bytes
 ..._permute_xdl_fp16.cpp.D04ED4238574C098.idx | Bin 0 -> 12730 bytes
 ...ped_gemm_fastgelu.hpp.81337D2196A77665.idx | Bin 0 -> 2640 bytes
 ...ped_gemm_fixed_nk.hpp.92D3931E2C242798.idx | Bin 0 -> 4728 bytes
 ..._permute_xdl_fp16.cpp.E01D0ADA6F0061BF.idx | Bin 0 -> 4814 bytes
 ..._permute_xdl_fp16.cpp.9FDD703AE6D2D15E.idx | Bin 0 -> 4820 bytes
 ...ped_gemm_xdl_bf16.cpp.60909DDCF9BB055C.idx | Bin 0 -> 3386 bytes
 ...ixed_nk_bias_fp16.cpp.3B6907AB640AF24B.idx | Bin 0 -> 11882 bytes
 ...xdl_fixed_nk_fp16.cpp.C49639C51E0DF485.idx | Bin 0 -> 10730 bytes
 ..._xdl_fixed_nk_fp8.cpp.33128DEF03F0DD92.idx | Bin 0 -> 10828 bytes
 ...ped_gemm_xdl_fp16.cpp.164525B4D87B1687.idx | Bin 0 -> 3386 bytes
 ...ped_gemm_xdl_fp32.cpp.F97F080B232E6F17.idx | Bin 0 -> 3384 bytes
 ...ped_gemm_xdl_int8.cpp.01E8A20D5ECCA353.idx | Bin 0 -> 3264 bytes
 ...m_xdl_splitk_fp16.cpp.5CCCCCFEB9008060.idx | Bin 0 -> 4056 bytes
 ...roupnorm_bwd_data.hpp.3714A9EE70464716.idx | Bin 0 -> 1234 bytes
 ...roupnorm_bwd_fp32.cpp.94FC313C790AEF66.idx | Bin 0 -> 6438 bytes
 ...rm_bwd_gamma_beta.hpp.A4BF65E232956D79.idx | Bin 0 -> 1244 bytes
 ..._sigmoid_mul_fp16.cpp.E1BDC41DAA5F65CC.idx | Bin 0 -> 2446 bytes
 ...m_fwd_splitk_fp16.cpp.5CDEA132226AE994.idx | Bin 0 -> 1954 bytes
 ...rm_fwd_swish_fp16.cpp.FDEC9D120BB249C9.idx | Bin 0 -> 1956 bytes
 .../hip_check_error.hpp.B6401497C46CBC5E.idx  | Bin 0 -> 744 bytes
 .../host_common_util.hpp.DD6CA1FC5D70D95B.idx | Bin 0 -> 2672 bytes
 .../host_tensor.cpp.BD60013DE8F91330.idx      | Bin 0 -> 2800 bytes
 .../host_tensor.hpp.A2A2489B4F3A4E06.idx      | Bin 0 -> 16564 bytes
 ..._tensor_generator.hpp.44F3F167E8AE0CAB.idx | Bin 0 -> 6738 bytes
 .../index/ignore.hpp.67D37926F4126F70.idx     | Bin 0 -> 400 bytes
 ...age_to_column_f32.cpp.E393FC8829D80520.idx | Bin 0 -> 4960 bytes
 .../inner_product.hpp.D15AD8CC8FA359B4.idx    | Bin 0 -> 7484 bytes
 ...nner_product_dpp8.hpp.5615826ED9AC4030.idx | Bin 0 -> 2696 bytes
 ...integral_constant.hpp.A156C83DCD53E7CE.idx | Bin 0 -> 1574 bytes
 .../clangd/index/io.hpp.27CD569BA11007AA.idx  | Bin 0 -> 1856 bytes
 .../is_detected.hpp.993129C9FD0FB446.idx      | Bin 0 -> 1390 bytes
 ...n_at_compile_time.hpp.C66FF603A6D2FF7E.idx | Bin 0 -> 1570 bytes
 .../index/iterator.hpp.DDB0074CE39B229F.idx   | Bin 0 -> 736 bytes
 .../kernel_launch.hpp.4FE4793F82580CE2.idx    | Bin 0 -> 1756 bytes
 .../kernel_utils.hpp.A2D72F83F06BC2A4.idx     | Bin 0 -> 210 bytes
 ...n_mn_add_instance.cpp.5CD53C62DE24ED15.idx | Bin 0 -> 1924 bytes
 ...eline_v1_instance.cpp.2F582983A57414AF.idx | Bin 0 -> 4048 bytes
 ...eline_v2_instance.cpp.FDC41D8D362237D7.idx | Bin 0 -> 4070 bytes
 ...e_v2_opt_instance.cpp.9D0F889F526341B0.idx | Bin 0 -> 1768 bytes
 ...eline_v1_instance.cpp.536F114F5BA9349C.idx | Bin 0 -> 4072 bytes
 ...eline_v1_instance.cpp.5937911B0257B674.idx | Bin 0 -> 1750 bytes
 ...eline_v2_instance.cpp.0318E39A03EE883F.idx | Bin 0 -> 1774 bytes
 ...eline_v1_instance.cpp.0A7ACF40DAB7236A.idx | Bin 0 -> 1776 bytes
 ...k_mn_add_instance.cpp.09C78F4C6D71CC59.idx | Bin 0 -> 1924 bytes
 ...eline_v1_instance.cpp.7011152E4CED8319.idx | Bin 0 -> 4048 bytes
 ...eline_v2_instance.cpp.4ADB4E587F42B7E9.idx | Bin 0 -> 4070 bytes
 ...e_v2_opt_instance.cpp.AA43A9C5EED3E40A.idx | Bin 0 -> 1768 bytes
 ...eline_v1_instance.cpp.37A32BC5170A6461.idx | Bin 0 -> 4072 bytes
 ...eline_v1_instance.cpp.322AFE9D596702FC.idx | Bin 0 -> 1750 bytes
 ...eline_v2_instance.cpp.6202EDA536A25A77.idx | Bin 0 -> 1774 bytes
 ...eline_v1_instance.cpp.667185EB457E88CA.idx | Bin 0 -> 1776 bytes
 ...ernorm2d_bwd_fp32.cpp.7066935C4C40FCB0.idx | Bin 0 -> 5730 bytes
 ...ernorm2d_fwd_fp16.cpp.6079EB647B2C4778.idx | Bin 0 -> 1970 bytes
 ...d_fwd_splitk_fp16.cpp.7A8CCC9ADABCC2A0.idx | Bin 0 -> 1974 bytes
 ...ernorm4d_fwd_fp16.cpp.FF678597E26B24C3.idx | Bin 0 -> 1968 bytes
 ...d_fwd_splitk_fp16.cpp.BF340D61393FB17F.idx | Bin 0 -> 1976 bytes
 ...ayernorm_bwd_data.hpp.D81DD3681C23DD68.idx | Bin 0 -> 1638 bytes
 ...rm_bwd_gamma_beta.hpp.8CC35A2FBD2934F9.idx | Bin 0 -> 1628 bytes
 .../index/layout.hpp.08D624D3F8E5BFE6.idx     | Bin 0 -> 7000 bytes
 .../layout_utils.hpp.9A8926719F7B5577.idx     | Bin 0 -> 7498 bytes
 .../index/literals.hpp.AC03AD34F59B947A.idx   | Bin 0 -> 514 bytes
 .../loop_scheduler.hpp.CEBBF3BFF5C82A80.idx   | Bin 0 -> 574 bytes
 .../magic_division.hpp.7DE3E22991AC58AE.idx   | Bin 0 -> 5530 bytes
 ...c_number_division.cpp.853FDB10989ECFC7.idx | Bin 0 -> 3322 bytes
 ...ng_specialization.hpp.2AE80C7347AB00C4.idx | Bin 0 -> 2250 bytes
 .../index/math.hpp.234B356D49034450.idx       | Bin 0 -> 5434 bytes
 .../index/math_v2.hpp.562B4AF9F5A86873.idx    | Bin 0 -> 5860 bytes
 .../matrix_padder.hpp.0539595BAF18EE86.idx    | Bin 0 -> 9614 bytes
 .../max_pool_bwd.hpp.4F15616E2BF6E596.idx     | Bin 0 -> 1614 bytes
 ...d_instance_common.hpp.F8135D216BA19663.idx | Bin 0 -> 828 bytes
 ...axpool2d_bwd_bf16.cpp.E133A720E87AC796.idx | Bin 0 -> 2018 bytes
 ...pool2d_bwd_common.hpp.4213849B15309F3C.idx | Bin 0 -> 4964 bytes
 ...axpool2d_bwd_fp16.cpp.AA5B517C73C2030C.idx | Bin 0 -> 2026 bytes
 ...axpool2d_bwd_fp32.cpp.3CE2CB1F186065DE.idx | Bin 0 -> 1962 bytes
 ...n_mn_add_instance.cpp.F2B6A1DC94E17C44.idx | Bin 0 -> 1924 bytes
 ...eline_v1_instance.cpp.B364344C594966A7.idx | Bin 0 -> 7000 bytes
 ...eline_v2_instance.cpp.A590A56B0084D88A.idx | Bin 0 -> 7022 bytes
 ...e_v2_opt_instance.cpp.7FA7EE3992571786.idx | Bin 0 -> 1758 bytes
 ...eline_v1_instance.cpp.E84C4EA0690C6426.idx | Bin 0 -> 7024 bytes
 ...eline_v1_instance.cpp.1D02066DCD48924F.idx | Bin 0 -> 1742 bytes
 ...eline_v2_instance.cpp.DFAE93983C514BC6.idx | Bin 0 -> 1764 bytes
 ...eline_v1_instance.cpp.905D27CAAE2E6FF3.idx | Bin 0 -> 1766 bytes
 ...k_mn_add_instance.cpp.0D22D3DFBDE4264A.idx | Bin 0 -> 1924 bytes
 ...eline_v1_instance.cpp.FC2B955D88E3500C.idx | Bin 0 -> 5692 bytes
 ...eline_v2_instance.cpp.AB0F8B7DC8CEF96B.idx | Bin 0 -> 5716 bytes
 ...e_v2_opt_instance.cpp.38A404E476FFFA7A.idx | Bin 0 -> 1768 bytes
 ...eline_v1_instance.cpp.1B80E7B3E985B173.idx | Bin 0 -> 5718 bytes
 ...eline_v1_instance.cpp.8AF96AFDCF49C20F.idx | Bin 0 -> 2408 bytes
 ...eline_v2_instance.cpp.AD25F4CBEC60DDF2.idx | Bin 0 -> 2432 bytes
 ...eline_v1_instance.cpp.EB9B1319190E914D.idx | Bin 0 -> 2434 bytes
 .../multi_index.hpp.46D763C8845699EA.idx      | Bin 0 -> 244 bytes
 ...i_index_transform.hpp.97E44055DBD0AE25.idx | Bin 0 -> 47022 bytes
 ..._transform_helper.hpp.7CC4F9B71E8A7CAA.idx | Bin 0 -> 2414 bytes
 ...a_instance_common.hpp.3AFBFA5B5298FD17.idx | Bin 0 -> 1978 bytes
 ...a_instance_common.hpp.84C16E4EE625B892.idx | Bin 0 -> 1982 bytes
 ...normalization_fwd.hpp.C0D3E8108CC38C5D.idx | Bin 0 -> 2588 bytes
 ...d_instance_common.hpp.38A39908E71BFB61.idx | Bin 0 -> 13126 bytes
 .../index/number.hpp.25A61C6490A403AE.idx     | Bin 0 -> 450 bytes
 .../index/numeric.hpp.E1739325ACADC629.idx    | Bin 0 -> 530 bytes
 ...ermute_1xHxW_fp16.cpp.5780304278E72678.idx | Bin 0 -> 1446 bytes
 ...ermute_HxWx4_fp16.cpp.FCA2351A6D6E3584.idx | Bin 0 -> 1482 bytes
 ...ermute_NxHxW_fp16.cpp.6FE6AEAF71058087.idx | Bin 0 -> 1446 bytes
 .../permute_scale.hpp.CFE8A4ABEDAA7993.idx    | Bin 0 -> 5034 bytes
 ...pool2d_fwd_common.hpp.3724B9CBDE79A373.idx | Bin 0 -> 3740 bytes
 .../pool2d_fwd_fp16.cpp.09424A28FF97D19C.idx  | Bin 0 -> 2890 bytes
 .../pool2d_fwd_fp32.cpp.1AF35A327BA2145F.idx  | Bin 0 -> 2852 bytes
 .../index/pool3d_fwd.hpp.32DAF50D80F54A30.idx | Bin 0 -> 3978 bytes
 ...pool3d_fwd_common.hpp.EE0EEEB8E82E75A0.idx | Bin 0 -> 4464 bytes
 .../pool3d_fwd_fp16.cpp.176B982A3BF061DE.idx  | Bin 0 -> 2502 bytes
 ...d_instance_common.hpp.4DC931E3BBF2D1BA.idx | Bin 0 -> 990 bytes
 ...le_avg_pool3d_bwd.cpp.4403F8C1766CA1E8.idx | Bin 0 -> 4198 bytes
 ...g_pool3d_bwd_impl.hpp.C8240F2385534214.idx | Bin 0 -> 5756 bytes
 ...file_batched_gemm.cpp.8A238D0EF3B6DB52.idx | Bin 0 -> 4950 bytes
 ...add_relu_gemm_add.cpp.CEB2ADB0FF2D0A07.idx | Bin 0 -> 3930 bytes
 ...elu_gemm_add_impl.hpp.CC900C93F3C48179.idx | Bin 0 -> 7496 bytes
 ...gemm_permute_impl.hpp.F57E6277DAC86E29.idx | Bin 0 -> 8382 bytes
 ...batched_gemm_gemm.cpp.B214B5448CC6B6C2.idx | Bin 0 -> 3482 bytes
 ...ed_gemm_gemm_impl.hpp.95E513E815A63274.idx | Bin 0 -> 6562 bytes
 ...batched_gemm_impl.hpp.65F140203E93B240.idx | Bin 0 -> 5272 bytes
 ...tched_gemm_reduce.cpp.EEA3A1DAF90E2787.idx | Bin 0 -> 2518 bytes
 ..._gemm_reduce_impl.hpp.5DA2586197AB1178.idx | Bin 0 -> 9058 bytes
 ...softmax_gemm_impl.hpp.03AF7B0A8D358951.idx | Bin 0 -> 7122 bytes
 ...gemm_permute_impl.hpp.14377DE4299D9A14.idx | Bin 0 -> 7574 bytes
 ...orm_backward_impl.hpp.30164435038F0711.idx | Bin 0 -> 7454 bytes
 ...ile_batchnorm_bwd.cpp.8CF98960A7C44AE9.idx | Bin 0 -> 5480 bytes
 ...norm_forward_impl.hpp.07F6CF6EEC2BDC7E.idx | Bin 0 -> 6872 bytes
 ...ile_batchnorm_fwd.cpp.493E9D388D70238E.idx | Bin 0 -> 5850 bytes
 ...e_batchnorm_infer.cpp.69EB815355931680.idx | Bin 0 -> 5364 bytes
 ...chnorm_infer_impl.hpp.B4DE6B04DAFD57EF.idx | Bin 0 -> 5968 bytes
 ...traction_bilinear.cpp.B7E7F8EC02159FA8.idx | Bin 0 -> 4384 bytes
 ..._contraction_impl.hpp.E5FF1E81EBC1E7D5.idx | Bin 0 -> 9162 bytes
 ...contraction_scale.cpp.C10C8FEB4E77E97A.idx | Bin 0 -> 4342 bytes
 ...contraction_utils.hpp.E0F0B9314CF06E36.idx | Bin 0 -> 2384 bytes
 ...ile_conv_bwd_data.cpp.E30D0E66DFF68BCC.idx | Bin 0 -> 3784 bytes
 ...onv_bwd_data_impl.hpp.0CBAC32BD7F22884.idx | Bin 0 -> 5652 bytes
 .../profile_conv_fwd.cpp.7612DDA826C4C111.idx | Bin 0 -> 3708 bytes
 ...onv_fwd_bias_relu.cpp.2AED2F5E417190C4.idx | Bin 0 -> 3942 bytes
 ...fwd_bias_relu_add.cpp.064478D11C1695BD.idx | Bin 0 -> 3946 bytes
 ...ias_relu_add_impl.hpp.660B8209627A1FBE.idx | Bin 0 -> 6692 bytes
 ...wd_bias_relu_impl.hpp.0A6D3A36AE3BBBB6.idx | Bin 0 -> 6212 bytes
 ...ile_conv_fwd_impl.hpp.DB77593E3FC3FC2B.idx | Bin 0 -> 5714 bytes
 ..._tensor_rearrange.cpp.50518BDD3D0D27F8.idx | Bin 0 -> 7144 bytes
 ...or_rearrange_impl.hpp.ECD3F7A8AA337DB3.idx | Bin 0 -> 5620 bytes
 ...se_layernorm_impl.hpp.228FBBB77B1C2212.idx | Bin 0 -> 7016 bytes
 .../profile_gemm.cpp.1660D1BAA0C70B61.idx     | Bin 0 -> 5226 bytes
 .../profile_gemm_add.cpp.B4A511D84B31E476.idx | Bin 0 -> 2530 bytes
 ..._add_add_fastgelu.cpp.374CFB86EAD905D5.idx | Bin 0 -> 2400 bytes
 ...add_fastgelu_impl.hpp.7165E9E08D31883A.idx | Bin 0 -> 5046 bytes
 ...gemm_add_fastgelu.cpp.CA0713C21D4F7C04.idx | Bin 0 -> 2630 bytes
 ...add_fastgelu_impl.hpp.23CA5590877F722E.idx | Bin 0 -> 4718 bytes
 ...ile_gemm_add_impl.hpp.A44323833CE8C5AD.idx | Bin 0 -> 4680 bytes
 ...gemm_add_multiply.cpp.136C377C7D5B0E54.idx | Bin 0 -> 2398 bytes
 ...add_multiply_impl.hpp.FBBEFEDC4A78D751.idx | Bin 0 -> 4998 bytes
 ...ile_gemm_add_relu.cpp.7B9D4CAF29075E5F.idx | Bin 0 -> 2534 bytes
 ...elu_add_layernorm.cpp.E30A7F678D4C6FAC.idx | Bin 0 -> 2402 bytes
 ...dd_layernorm_impl.hpp.5DE61D830E8EC4CC.idx | Bin 0 -> 5142 bytes
 ...emm_add_relu_impl.hpp.9153E795377986F0.idx | Bin 0 -> 4714 bytes
 ...ile_gemm_add_silu.cpp.67ABB3D17F5C92B4.idx | Bin 0 -> 2536 bytes
 ...emm_add_silu_impl.hpp.CC18A7C19DDCC841.idx | Bin 0 -> 4714 bytes
 ...m_bias_add_reduce.cpp.CF67B6B3D02F20A5.idx | Bin 0 -> 2636 bytes
 ...s_add_reduce_impl.hpp.CC4023F2075FFC72.idx | Bin 0 -> 9850 bytes
 ...ile_gemm_bilinear.cpp.2A04608A0019808D.idx | Bin 0 -> 2496 bytes
 ...emm_bilinear_impl.hpp.3D87C8E5D761FA48.idx | Bin 0 -> 4594 bytes
 ...ile_gemm_fastgelu.cpp.2EC621FE126AC880.idx | Bin 0 -> 2304 bytes
 ...emm_fastgelu_impl.hpp.090E8EEC37D7278D.idx | Bin 0 -> 4054 bytes
 ...profile_gemm_impl.hpp.DB85F2B55BD1CC04.idx | Bin 0 -> 5994 bytes
 ...gemm_multiply_add.cpp.0030E28460A11C1F.idx | Bin 0 -> 2432 bytes
 ...multiply_add_impl.hpp.A03C0417F818F4FD.idx | Bin 0 -> 5000 bytes
 ...ofile_gemm_reduce.cpp.0A54AE6B66160321.idx | Bin 0 -> 2496 bytes
 ..._gemm_reduce_impl.hpp.FE75615CA769411C.idx | Bin 0 -> 9070 bytes
 ...ofile_gemm_splitk.cpp.6BF00CBB81DE4660.idx | Bin 0 -> 4858 bytes
 ..._gemm_splitk_impl.hpp.1DC09EFBBC454E8F.idx | Bin 0 -> 6752 bytes
 ...file_gemm_streamk.cpp.46F4E02E0912E672.idx | Bin 0 -> 3594 bytes
 ...gemm_streamk_impl.hpp.73FA01C66CC8D3E3.idx | Bin 0 -> 6022 bytes
 ...ped_conv_bwd_data.cpp.FA67D9F6149A96BE.idx | Bin 0 -> 4052 bytes
 ...onv_bwd_data_impl.hpp.59E9CBB730B20A37.idx | Bin 0 -> 5452 bytes
 ...d_conv_bwd_weight.cpp.34186A38546F70D1.idx | Bin 0 -> 5272 bytes
 ...v_bwd_weight_impl.hpp.A8CF2351C95F34F5.idx | Bin 0 -> 6012 bytes
 ..._grouped_conv_fwd.cpp.8D8BCE0560F621C4.idx | Bin 0 -> 4522 bytes
 ...ped_conv_fwd_impl.hpp.90847DB78EF4B739.idx | Bin 0 -> 5934 bytes
 ...file_grouped_gemm.cpp.7EE7588374346C13.idx | Bin 0 -> 4578 bytes
 ...ped_gemm_fastgelu.cpp.AC58B27586C8FEAD.idx | Bin 0 -> 3858 bytes
 ...emm_fastgelu_impl.hpp.DE8318589739E396.idx | Bin 0 -> 6480 bytes
 ...ped_gemm_fixed_nk.cpp.12EAC200C0F64A96.idx | Bin 0 -> 4264 bytes
 ...emm_fixed_nk_impl.hpp.B583CC73B5188079.idx | Bin 0 -> 6542 bytes
 ...grouped_gemm_impl.hpp.0459F4415131A8AB.idx | Bin 0 -> 6162 bytes
 ...roupnorm_bwd_data.cpp.B5C8D8F20885019C.idx | Bin 0 -> 2880 bytes
 ...orm_bwd_data_impl.hpp.05BE2AB219AD2046.idx | Bin 0 -> 7216 bytes
 ...rm_bwd_gamma_beta.cpp.0D9708552C2EDA92.idx | Bin 0 -> 2888 bytes
 ...d_gamma_beta_impl.hpp.F912A6BA3766E9B6.idx | Bin 0 -> 7364 bytes
 ...ile_groupnorm_fwd.cpp.B71B17B33AF365C7.idx | Bin 0 -> 2984 bytes
 ...roupnorm_fwd_impl.hpp.DD036B9EC45A3789.idx | Bin 0 -> 7184 bytes
 ...ayernorm_bwd_data.cpp.EDB022306AF7F904.idx | Bin 0 -> 2966 bytes
 ...orm_bwd_data_impl.hpp.CF00FB7A122475B2.idx | Bin 0 -> 7288 bytes
 ...rm_bwd_gamma_beta.cpp.58ACE0ADA4571F63.idx | Bin 0 -> 2976 bytes
 ...d_gamma_beta_impl.hpp.672AE63374836A4E.idx | Bin 0 -> 7360 bytes
 ...ile_layernorm_fwd.cpp.FA9A3EA5A9ABAF4C.idx | Bin 0 -> 3112 bytes
 ...ayernorm_fwd_impl.hpp.BF1975AA887FB0A6.idx | Bin 0 -> 7462 bytes
 ...le_max_pool3d_bwd.cpp.7793240AA812E755.idx | Bin 0 -> 4238 bytes
 ...x_pool3d_bwd_impl.hpp.885D2D1066A38A2E.idx | Bin 0 -> 6066 bytes
 ...le_max_pool3d_fwd.cpp.11655872CD06FCBE.idx | Bin 0 -> 4654 bytes
 ...ile_permute_scale.cpp.A23434A32C62DC3B.idx | Bin 0 -> 3164 bytes
 ...ermute_scale_impl.hpp.9398D48C7353608C.idx | Bin 0 -> 5136 bytes
 ...e_pool3d_fwd_impl.hpp.4E7051C23FCACC74.idx | Bin 0 -> 5902 bytes
 .../profile_reduce.cpp.7CFAAB4D3A545696.idx   | Bin 0 -> 10466 bytes
 ...ofile_reduce_impl.hpp.17B63EE6A3A6D55C.idx | Bin 0 -> 8812 bytes
 .../profile_softmax.cpp.31A9AF77C4279825.idx  | Bin 0 -> 5040 bytes
 ...file_softmax_impl.hpp.28DC8A49A2D3559F.idx | Bin 0 -> 7920 bytes
 ...profile_transpose.cpp.3798E55CF7D973A7.idx | Bin 0 -> 3004 bytes
 ...le_transpose_impl.hpp.8CE368E3C650AFE1.idx | Bin 0 -> 4964 bytes
 .../index/profiler.cpp.A1998E185346D702.idx   | Bin 0 -> 1590 bytes
 ...peration_registry.hpp.C3738D3851333A5F.idx | Bin 0 -> 2204 bytes
 .../put_element_fp16.cpp.4CB3E91BA1C3425F.idx | Bin 0 -> 3146 bytes
 ...ization_operation.hpp.BD94143B871C7D8C.idx | Bin 0 -> 5448 bytes
 .../index/random_gen.hpp.B234A29FCA71839D.idx | Bin 0 -> 1046 bytes
 .../index/ranges.hpp.66B5D2199BD23CD4.idx     | Bin 0 -> 2112 bytes
 .../index/reduce.hpp.7DC5A811A95F14CB.idx     | Bin 0 -> 1156 bytes
 .../reduce_blockwise.cpp.E231BF6C564BCCFD.idx | Bin 0 -> 6846 bytes
 ...ce_blockwise_impl.hpp.B38B56E3F9BFC70D.idx | Bin 0 -> 4880 bytes
 ...lockwise_two_call.cpp.C62D3A32F608E374.idx | Bin 0 -> 7212 bytes
 ...ce_example_common.hpp.A3D8D5EB42288D30.idx | Bin 0 -> 992 bytes
 ...iblock_atomic_add.cpp.22DBB7B98E8ABAE4.idx | Bin 0 -> 5590 bytes
 ...k_atomic_add_impl.hpp.014F022BC601B7F9.idx | Bin 0 -> 3762 bytes
 .../reduce_no_index.cpp.3EDCFAA9D727E5B6.idx  | Bin 0 -> 4554 bytes
 ...reduce_with_index.cpp.F84EAB2F5A59285D.idx | Bin 0 -> 4564 bytes
 .../reduction_common.hpp.C00F2970D4BF7FA3.idx | Bin 0 -> 522 bytes
 .../reduction_enums.hpp.871AE7D90DC9E069.idx  | Bin 0 -> 1400 bytes
 ...ctions_accumulate.hpp.7D2C2CFC245627F0.idx | Bin 0 -> 1634 bytes
 ...nctions_blockwise.hpp.C958E7F57DF3365E.idx | Bin 0 -> 4528 bytes
 ...ctions_threadwise.hpp.7A06A2856650B185.idx | Bin 0 -> 2020 bytes
 ...eduction_operator.hpp.28F266F9F2F32A7B.idx | Bin 0 -> 7810 bytes
 ..._operator_mapping.hpp.8A11EE4B065A6CAE.idx | Bin 0 -> 5628 bytes
 ...rence_avgpool_bwd.hpp.EF9EF399DB15462F.idx | Bin 0 -> 6474 bytes
 ...ence_batched_gemm.hpp.AF38237B103D733E.idx | Bin 0 -> 7908 bytes
 ...atchnorm_backward.hpp.DB836B16A0EC29D6.idx | Bin 0 -> 10216 bytes
 ...batchnorm_forward.hpp.0421D3C13A4F2108.idx | Bin 0 -> 9380 bytes
 ...e_batchnorm_infer.hpp.B110A1DBC8883C05.idx | Bin 0 -> 7220 bytes
 .../reference_cgemm.hpp.9B0EE32433381722.idx  | Bin 0 -> 3734 bytes
 ...e_column_to_image.hpp.B1B86724AE64837E.idx | Bin 0 -> 9180 bytes
 ...rence_contraction.hpp.4808E834C3E32AEB.idx | Bin 0 -> 3162 bytes
 ...nce_conv_bwd_data.hpp.7B56299251FBE761.idx | Bin 0 -> 7830 bytes
 ...e_conv_bwd_weight.hpp.474E92061BE55B30.idx | Bin 0 -> 6822 bytes
 ...eference_conv_fwd.cpp.3703F6A90F557994.idx | Bin 0 -> 5830 bytes
 ...eference_conv_fwd.hpp.4266D54C9E6D337C.idx | Bin 0 -> 8090 bytes
 ...d_bias_activation.hpp.CC5C1A67067C17BD.idx | Bin 0 -> 4904 bytes
 ...as_activation_add.hpp.05EAEEBBFD920597.idx | Bin 0 -> 5038 bytes
 .../reference_gemm.hpp.0A74F0F3AFBAC0AB.idx   | Bin 0 -> 3096 bytes
 ...ce_gemm_layernorm.hpp.9AA4C354D9D24BEC.idx | Bin 0 -> 4930 bytes
 ...ference_groupnorm.hpp.368D3290BA70398F.idx | Bin 0 -> 4944 bytes
 ...nce_groupnorm_bwd.hpp.C24505C3C5BFB298.idx | Bin 0 -> 6272 bytes
 ...e_image_to_column.hpp.14F89CCD7031E508.idx | Bin 0 -> 9402 bytes
 ...ference_layernorm.hpp.B9C72CF73E810635.idx | Bin 0 -> 6506 bytes
 ...nce_layernorm_bwd.hpp.F0026FDE37931C67.idx | Bin 0 -> 6248 bytes
 ...rence_maxpool_bwd.hpp.AA4F2921ED8F7A67.idx | Bin 0 -> 2624 bytes
 ...eference_pool_fwd.hpp.3BB9B8E57191E80A.idx | Bin 0 -> 10426 bytes
 .../reference_reduce.hpp.0C04C3B5F65B2DFD.idx | Bin 0 -> 7802 bytes
 ...reference_softmax.hpp.4C7847F5A451A485.idx | Bin 0 -> 4100 bytes
 ...forward_layernorm.hpp.495E2A0ED186B796.idx | Bin 0 -> 5280 bytes
 ...ched_gemm_example.inc.341E18570766CD6C.idx | Bin 0 -> 4858 bytes
 ...gemm_gemm_example.inc.B0C303C33F007446.idx | Bin 0 -> 5004 bytes
 ...cale_softmax_gemm.inc.35534F29FB76E4C9.idx | Bin 0 -> 5692 bytes
 ...tmax_gemm_permute.inc.DF3E33E757137FA7.idx | Bin 0 -> 6114 bytes
 ..._bilinear_example.inc.427C55016011CF17.idx | Bin 0 -> 5596 bytes
 ...ion_scale_example.inc.619AD6EEB2674ADA.idx | Bin 0 -> 4884 bytes
 ...ntization_example.inc.453F2B26B4A84A7C.idx | Bin 0 -> 5400 bytes
 ...ntization_example.inc.41E3948246A8DC23.idx | Bin 0 -> 4970 bytes
 ...ntization_example.inc.4B57FDA755059955.idx | Bin 0 -> 4974 bytes
 ...ntization_example.inc.3D388FA1A8D19B72.idx | Bin 0 -> 4160 bytes
 ...vnd_activ_example.inc.F2FB9ED6F9BA4B8A.idx | Bin 0 -> 1426 bytes
 ...onvnd_fwd_example.inc.6C6FF11BF5A73D93.idx | Bin 0 -> 1368 bytes
 ...d_fwd_max_example.inc.4C55B2EACEAC1C07.idx | Bin 0 -> 6314 bytes
 ..._fastgelu_example.inc.E8AFCCD92C8D9967.idx | Bin 0 -> 3718 bytes
 ..._multiply_example.inc.3C87D5BB1B980F33.idx | Bin 0 -> 3652 bytes
 .../run_gemm_example.inc.4868A822F4677354.idx | Bin 0 -> 6828 bytes
 .../run_gemm_test.inc.B7894CAD83BCCCF5.idx    | Bin 0 -> 718 bytes
 ...bias_relu_example.inc.1D7F06778A26066D.idx | Bin 0 -> 5138 bytes
 ..._bwd_data_example.inc.11E97FA73102DBD5.idx | Bin 0 -> 4392 bytes
 ...wd_weight_example.inc.EE2E4EE0E60E4D81.idx | Bin 0 -> 3260 bytes
 ..._conv_fwd_example.inc.88B5CE7F22532219.idx | Bin 0 -> 6876 bytes
 ..._relu_add_example.inc.7B0A8630D499DF28.idx | Bin 0 -> 6284 bytes
 ..._conv_fwd_example.inc.C65A743D8AABBB7C.idx | Bin 0 -> 4326 bytes
 ...uped_gemm_example.inc.FE653C75BFA23618.idx | Bin 0 -> 6854 bytes
 ...tmax_gemm_permute.inc.0CAE0C43C07E96CE.idx | Bin 0 -> 7728 bytes
 ...pnorm_fwd_example.inc.85BD21F884D2FB4A.idx | Bin 0 -> 3752 bytes
 ...orm4d_fwd_example.inc.8870D57CA6553449.idx | Bin 0 -> 2926 bytes
 ...layernorm_example.inc.DA0460DDD4EF632F.idx | Bin 0 -> 2900 bytes
 ...te_bundle_example.inc.419BC377BBAE0EEA.idx | Bin 0 -> 2148 bytes
 ...e_element_example.inc.42CC4C3518B6F771.idx | Bin 0 -> 1764 bytes
 ...litK_gemm_example.inc.51FAC96E5EFEA0F0.idx | Bin 0 -> 4824 bytes
 .../index/sequence.hpp.25AC5091073FF3EB.idx   | Bin 0 -> 26038 bytes
 .../sequence_helper.hpp.2282D53015E164C0.idx  | Bin 0 -> 896 bytes
 .../index/softmax.hpp.2F97791F40A73FC9.idx    | Bin 0 -> 1412 bytes
 ...softmax_blockwise.cpp.3EE65033A8858975.idx | Bin 0 -> 6792 bytes
 ...ace_filling_curve.cpp.EB322F6A66137D22.idx | Bin 0 -> 5936 bytes
 .../index/span.hpp.D6316D1D55F0830C.idx       | Bin 0 -> 3328 bytes
 ...forward_layernorm.cpp.CF5214A930A81517.idx | Bin 0 -> 9782 bytes
 ...itK_gemm_xdl_bf16.cpp.AB98F65366A9F1A9.idx | Bin 0 -> 3188 bytes
 ...itK_gemm_xdl_fp16.cpp.DBCD688A2F933855.idx | Bin 0 -> 3098 bytes
 ...gemm_xdl_fp16_fp8.cpp.60034E8F6043ABD9.idx | Bin 0 -> 3328 bytes
 ...itK_gemm_xdl_fp32.cpp.A3B8F8ECA755E6B2.idx | Bin 0 -> 3098 bytes
 ...itK_gemm_xdl_int8.cpp.4600C416992C0160.idx | Bin 0 -> 3070 bytes
 ..._direct_load_fp16.cpp.2D633E2EB00CCD78.idx | Bin 0 -> 3078 bytes
 ..._permute_xdl_fp16.cpp.3597E35811A1130C.idx | Bin 0 -> 11556 bytes
 ..._permute_xdl_fp32.cpp.8A4C951A87984080.idx | Bin 0 -> 11558 bytes
 .../static_buffer.hpp.E059C0D43CFA4618.idx    | Bin 0 -> 5930 bytes
 .../static_tensor.hpp.55059299885EB585.idx    | Bin 0 -> 4350 bytes
 ...lly_indexed_array.hpp.8A80A7A4CC47B954.idx | Bin 0 -> 2362 bytes
 ...array_multi_index.hpp.E363636635FF265A.idx | Bin 0 -> 3500 bytes
 .../stream_config.hpp.015AF40CD82DD077.idx    | Bin 0 -> 618 bytes
 .../stream_utility.hpp.7DD00DAAC010212E.idx   | Bin 0 -> 528 bytes
 .../synchronization.hpp.94B924A949B60CF0.idx  | Bin 0 -> 466 bytes
 .../index/tensor.hpp.C84FDD70C53B91FD.idx     | Bin 0 -> 7706 bytes
 .../tensor_adaptor.hpp.1AABBD498E3DDABA.idx   | Bin 0 -> 7392 bytes
 ...tensor_descriptor.hpp.7E73ABD05058119F.idx | Bin 0 -> 11290 bytes
 ...descriptor_helper.hpp.B97BC6F42C53AF70.idx | Bin 0 -> 2568 bytes
 .../tensor_layout.hpp.83E496410F43D7E9.idx    | Bin 0 -> 12068 bytes
 .../tensor_partition.hpp.2F54ACF04D747753.idx | Bin 0 -> 5406 bytes
 ...ace_filling_curve.hpp.058953EFBD7A2B25.idx | Bin 0 -> 4256 bytes
 ...or_specialization.hpp.5BC1AF8D88F729F2.idx | Bin 0 -> 678 bytes
 .../tensor_utils.hpp.B92D33FF5A9EED5B.idx     | Bin 0 -> 4658 bytes
 ...st_avg_pool3d_bwd.cpp.6DD8FAFEF9A89C97.idx | Bin 0 -> 3470 bytes
 ...st_avg_pool3d_fwd.cpp.5274F49320881D42.idx | Bin 0 -> 3298 bytes
 ...test_batched_gemm.cpp.A022B37F45A32155.idx | Bin 0 -> 6944 bytes
 ...gemm_permute_bf16.cpp.52CCAAF6582C3CD0.idx | Bin 0 -> 15106 bytes
 ...gemm_permute_fp16.cpp.FB71D53BC1DE7E13.idx | Bin 0 -> 15102 bytes
 ...gemm_permute_util.hpp.38E1B00A8B9BAD43.idx | Bin 0 -> 8154 bytes
 ...ed_gemm_gemm_fp16.cpp.A1B1DD17614F30D3.idx | Bin 0 -> 13082 bytes
 ...ed_gemm_gemm_util.hpp.1FF4EC29A4661CFF.idx | Bin 0 -> 4388 bytes
 ...softmax_gemm_fp16.cpp.0592033B8F1CD807.idx | Bin 0 -> 14694 bytes
 ...gemm_permute_bf16.cpp.EDC4F40BC781A6DC.idx | Bin 0 -> 15052 bytes
 ...gemm_permute_fp16.cpp.0D1473C6C2BC0BFC.idx | Bin 0 -> 15060 bytes
 ...gemm_permute_util.hpp.DD433B78FEA57450.idx | Bin 0 -> 7610 bytes
 ...softmax_gemm_util.hpp.C46CFF8584087968.idx | Bin 0 -> 4566 bytes
 .../index/test_bf8.cpp.16F0A495F47E95CC.idx   | Bin 0 -> 13388 bytes
 ...lock_to_ctile_map.cpp.D0968ED6B32C18D9.idx | Bin 0 -> 13980 bytes
 .../test_contraction.cpp.6A4280349E6B5DE6.idx | Bin 0 -> 9750 bytes
 ...raction_interface.cpp.51D57FFCD74AC94A.idx | Bin 0 -> 11242 bytes
 ..._tensor_rearrange.cpp.69AF0D8F1794ACAD.idx | Bin 0 -> 5752 bytes
 ...arrange_interface.cpp.6D91D3DCA5B6D134.idx | Bin 0 -> 15844 bytes
 ...se_layernorm_fp16.cpp.6E7042446B0FE0D3.idx | Bin 0 -> 3444 bytes
 .../index/test_fp8.cpp.AF44E860BEF4215C.idx   | Bin 0 -> 13386 bytes
 .../test_gemm_add.hpp.EFC09082310C599D.idx    | Bin 0 -> 3774 bytes
 ...gemm_add_fastgelu.cpp.9C90F62B27922BE9.idx | Bin 0 -> 3332 bytes
 ...est_gemm_add_relu.cpp.6A11991389EEE204.idx | Bin 0 -> 3290 bytes
 ...dd_layernorm_fp16.cpp.E9706C96D2E415C1.idx | Bin 0 -> 5208 bytes
 ...est_gemm_add_silu.cpp.0C3DE18B907F5A14.idx | Bin 0 -> 3330 bytes
 .../test_gemm_splitk.cpp.921AD8C68523E94F.idx | Bin 0 -> 3696 bytes
 ...m_splitk_ut_cases.inc.CD038407CD6C080D.idx | Bin 0 -> 7386 bytes
 ..._gemm_splitk_util.hpp.D5E2B28B24FEB85E.idx | Bin 0 -> 2342 bytes
 ...d_convnd_bwd_data.cpp.BA5BCB136A7B5CF1.idx | Bin 0 -> 5266 bytes
 ...ata_interface_xdl.cpp.8747321B4790F80F.idx | Bin 0 -> 7392 bytes
 ...convnd_bwd_weight.cpp.1E0C24433031CA72.idx | Bin 0 -> 8652 bytes
 ...ght_interface_xdl.cpp.B93D201E166036D3.idx | Bin 0 -> 7454 bytes
 ...rouped_convnd_fwd.cpp.5B423DD33073B028.idx | Bin 0 -> 6280 bytes
 ...ulti_ab_interface.cpp.D44C4377C972C730.idx | Bin 0 -> 8800 bytes
 ...ace_compatibility.cpp.D8EE1DD81FCEA622.idx | Bin 0 -> 3526 bytes
 ...ed_gemm_interface.cpp.BFD580574FDC4028.idx | Bin 0 -> 11368 bytes
 ...ouped_gemm_splitk.cpp.1985BB1B09859227.idx | Bin 0 -> 4558 bytes
 ...ped_gemm_ut_cases.inc.B622DD65303F1D04.idx | Bin 0 -> 9582 bytes
 ...grouped_gemm_util.hpp.62D7D32AEF140979.idx | Bin 0 -> 7828 bytes
 ...orm_bwd_data_fp32.cpp.EF83656CE5FBCD24.idx | Bin 0 -> 3434 bytes
 ...d_gamma_beta_fp32.cpp.266D9654B5E1C67F.idx | Bin 0 -> 3444 bytes
 ...roupnorm_fwd_fp16.cpp.E01BDD088F0473FE.idx | Bin 0 -> 3460 bytes
 ...roupnorm_fwd_fp32.cpp.06336278490A0D06.idx | Bin 0 -> 3432 bytes
 ...m2d_bwd_data_fp32.cpp.65C39944BBD4A16D.idx | Bin 0 -> 3440 bytes
 ...d_gamma_beta_fp32.cpp.B3B6429EB0BBF1C4.idx | Bin 0 -> 3450 bytes
 ...ernorm2d_fwd_fp16.cpp.49CB42570B950D66.idx | Bin 0 -> 3436 bytes
 ...ernorm2d_fwd_fp32.cpp.57C9823195A1B889.idx | Bin 0 -> 3438 bytes
 ...ernorm4d_fwd_fp16.cpp.4CF1AE7F957B873D.idx | Bin 0 -> 3396 bytes
 ...st_max_pool3d_bwd.cpp.AF800737CBB45892.idx | Bin 0 -> 3476 bytes
 ...st_max_pool3d_fwd.cpp.BE8A2BFD19420616.idx | Bin 0 -> 3690 bytes
 ...est_permute_scale.cpp.0607B34D8846C932.idx | Bin 0 -> 5080 bytes
 ...t_pool_fwd_common.hpp.CD53AEA46B3FF1EB.idx | Bin 0 -> 1562 bytes
 ...softmax_interface.cpp.9F0A118BAF3CB91E.idx | Bin 0 -> 13524 bytes
 ...est_softmax_rank3.cpp.F74F60A3C98E1A4C.idx | Bin 0 -> 2138 bytes
 ...est_softmax_rank4.cpp.05FB4DA05EDB8025.idx | Bin 0 -> 2036 bytes
 ..._softmax_ut_cases.inc.A86C258A9B038466.idx | Bin 0 -> 3360 bytes
 ...test_softmax_util.hpp.A85F4CCFF59B9D2C.idx | Bin 0 -> 4926 bytes
 .../test_transpose.cpp.EA23A250A71A4476.idx   | Bin 0 -> 3372 bytes
 ...test_wrapper_copy.cpp.732CF6ED569867C1.idx | Bin 0 -> 5522 bytes
 ...test_wrapper_gemm.cpp.74B7113BB3F74220.idx | Bin 0 -> 13228 bytes
 ...st_wrapper_layout.cpp.6A8EFC4867A12799.idx | Bin 0 -> 25600 bytes
 ...wrapper_partition.cpp.EDFCA4C2C7218226.idx | Bin 0 -> 5914 bytes
 ...st_wrapper_tensor.cpp.A4AB6B2A22D4689E.idx | Bin 0 -> 14482 bytes
 .../thread_group.hpp.568EEF4CF01C2640.idx     | Bin 0 -> 682 bytes
 ...nsfer_direct_load.hpp.5F79EA1E492991C0.idx | Bin 0 -> 6644 bytes
 ...ice_transfer_v4r1.hpp.CC30E09EE6FAA0EA.idx | Bin 0 -> 3890 bytes
 ...ice_transfer_v6r1.hpp.F8F9EEB58869763C.idx | Bin 0 -> 2992 bytes
 ...e_transfer_v6r1r2.hpp.34BED592A25E7D4A.idx | Bin 0 -> 3768 bytes
 ...ice_transfer_v6r2.hpp.B8541658FFD968DD.idx | Bin 0 -> 3398 bytes
 ...ice_transfer_v6r3.hpp.F9C1F18AD045D6D7.idx | Bin 0 -> 3876 bytes
 ...slice_transfer_v7.hpp.85ADAEC27150419B.idx | Bin 0 -> 3830 bytes
 ...ice_transfer_v7r2.hpp.E55C206DD2AD1B19.idx | Bin 0 -> 4554 bytes
 ...or_slice_transfer.hpp.CE773F216A04AFAA.idx | Bin 0 -> 24630 bytes
 ...ice_transfer_v3r1.hpp.2D0C9495DC723ACC.idx | Bin 0 -> 13920 bytes
 ...ice_transfer_v4r1.hpp.42FD0DF31A3BFDC2.idx | Bin 0 -> 3270 bytes
 ...ice_transfer_v6r1.hpp.7AB8A56ED737D765.idx | Bin 0 -> 3766 bytes
 ...e_transfer_v6r1r2.hpp.AD9623F82BA99EFC.idx | Bin 0 -> 3784 bytes
 ...ice_transfer_v6r2.hpp.D28171A7A44ECC24.idx | Bin 0 -> 4606 bytes
 ...ice_transfer_v6r3.hpp.C25C8A66C4CF26F7.idx | Bin 0 -> 5430 bytes
 ...slice_transfer_v7.hpp.5C98C34294C833A3.idx | Bin 0 -> 5142 bytes
 ...ice_transfer_v7r2.hpp.891A886C76B49727.idx | Bin 0 -> 6980 bytes
 ...hreadwise_welford.hpp.CCCD2666E9AE4B4E.idx | Bin 0 -> 3816 bytes
 ...ntraction_to_gemm.hpp.51AB58D6126FEA84.idx | Bin 0 -> 6932 bytes
 ...d_data_to_gemm_v1.hpp.341C45E4BB62B32F.idx | Bin 0 -> 22458 bytes
 ..._conv_fwd_to_gemm.hpp.14B4756DD59EEE91.idx | Bin 0 -> 10592 bytes
 .../transpose_3d.hpp.F3212DEAAECABFBD.idx     | Bin 0 -> 1528 bytes
 ...transpose_vectors.hpp.7945D9ABB9DC55BD.idx | Bin 0 -> 4920 bytes
 .../index/tuple.hpp.3819E4E0F87690B8.idx      | Bin 0 -> 5602 bytes
 .../tuple_helper.hpp.73EC15D672B6492A.idx     | Bin 0 -> 4482 bytes
 .../index/type.hpp.3093CAE9B6EE9081.idx       | Bin 0 -> 1350 bytes
 .../type_convert.hpp.5842DA15997DD61D.idx     | Bin 0 -> 7482 bytes
 ...ype_convert_const.cpp.46FAD4095BF1793B.idx | Bin 0 -> 4342 bytes
 ...nt_wise_operation.hpp.DD0C457D046F7875.idx | Bin 0 -> 19700 bytes
 .../welford_helper.hpp.DC75228A25D5F78E.idx   | Bin 0 -> 1736 bytes
 .../index/wmma_gemm.hpp.260BC1EC1DC7109A.idx  | Bin 0 -> 15720 bytes
 ...workgroup_barrier.hpp.E88FB4AF7CFEB4E5.idx | Bin 0 -> 1398 bytes
 ...p_synchronization.hpp.EE3FBA63C405EAA9.idx | Bin 0 -> 1004 bytes
 .../xdlops_gemm.hpp.C8FD7C44D6227313.idx      | Bin 0 -> 39200 bytes
 .clang-format                                 |   0
 .clang-tidy                                   |   0
 .github/CODEOWNERS                            |   0
 .github/ISSUE_TEMPLATE/config.yml             |   0
 .github/ISSUE_TEMPLATE/issue_report.yml       |   0
 .github/dependabot.yml                        |   0
 .gitignore                                    |   0
 .pre-commit-config.yaml                       |   0
 .readthedocs.yaml                             |   0
 CHANGELOG.md                                  |   0
 CITATION.cff                                  |   0
 CMakeLists.txt                                |   0
 CONTRIBUTORS.md                               |   0
 Config.cmake.in                               |   0
 Dockerfile                                    |   0
 Jenkinsfile                                   |   0
 LICENSE                                       |   0
 README.md                                     |   0
 client_example/01_gemm/CMakeLists.txt         |   0
 client_example/01_gemm/gemm.cpp               |   0
 .../02_gemm_add_add_fastgelu/CMakeLists.txt   |   0
 .../gemm_add_add_fastgelu.cpp                 |   0
 .../gemm_add_add_fastgelu_generic.cpp         |   0
 .../gemm_add_fastgelu.cpp                     |   0
 .../gemm_add_fastgelu_generic.cpp             |   0
 .../gemm_fastgelu.cpp                         |   0
 .../gemm_fastgelu_generic.cpp                 |   0
 .../03_gemm_layernorm/CMakeLists.txt          |   0
 .../gemm_add_add_layernorm_naive.cpp          |   0
 .../gemm_add_relu_add_layernorm_welford.cpp   |   0
 client_example/04_contraction/CMakeLists.txt  |   0
 .../contraction_bilinear_fp32.cpp             |   0
 .../contraction_bilinear_fp64.cpp             |   0
 .../contraction_g1m2n3k1_add_xdl_fp16.cpp     |   0
 .../04_contraction/contraction_scale_fp32.cpp |   0
 .../04_contraction/contraction_scale_fp64.cpp |   0
 client_example/05_layernorm/CMakeLists.txt    |   0
 .../05_layernorm/layernorm2d_bwd_data.cpp     |   0
 .../layernorm2d_bwd_gamma_beta.cpp            |   0
 .../05_layernorm/layernorm2d_fwd.cpp          |   0
 .../05_layernorm/layernorm4d_fwd.cpp          |   0
 client_example/06_softmax/CMakeLists.txt      |   0
 client_example/06_softmax/softmax4d.cpp       |   0
 .../07_grouped_convnd_fwd/CMakeLists.txt      |   0
 .../grouped_conv1d_fwd.cpp                    |   0
 .../grouped_conv2d_fwd.cpp                    |   0
 .../08_fused_attention/CMakeLists.txt         |   0
 .../08_fused_attention/fused_attention.cpp    |   0
 .../fused_attention_bias.cpp                  |   0
 client_example/09_quantization/CMakeLists.txt |   0
 ..._fwd_bias_relu_perchannel_quantization.cpp |   0
 ...2d_fwd_bias_relu_perlayer_quantization.cpp |   0
 ..._fwd_bias_tanh_perchannel_quantization.cpp |   0
 ...2d_fwd_bias_tanh_perlayer_quantization.cpp |   0
 .../conv2d_fwd_perchannel_quantization.cpp    |   0
 .../conv2d_fwd_perlayer_quantization.cpp      |   0
 .../09_quantization/gemm_quantization.cpp     |   0
 .../10_grouped_convnd_bwd_data/CMakeLists.txt |   0
 .../grouped_conv2d_bwd_data.cpp               |   0
 .../grouped_conv3d_bwd_data.cpp               |   0
 ..._conv3d_bwd_data_input_fp16_comp_bf8f8.cpp |   0
 .../11_grouped_conv_bwd_weight/CMakeLists.txt |   0
 .../11_grouped_conv_bwd_weight/common.hpp     |   0
 .../grouped_conv1d_bwd_weight_fp16.cpp        |   0
 .../grouped_conv2d_bwd_weight_fp16.cpp        |   0
 .../grouped_conv3d_bwd_weight_fp16.cpp        |   0
 ...ed_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp |   0
 .../grouped_conv3d_bwd_weight_fp32.cpp        |   0
 .../CMakeLists.txt                            |   0
 .../elementwise_layernorm2d.cpp               |   0
 client_example/13_batchnorm/CMakeLists.txt    |   0
 .../13_batchnorm/batchnorm_bwd_nhwc.cpp       |   0
 .../13_batchnorm/batchnorm_fwd_nhwc.cpp       |   0
 .../13_batchnorm/batchnorm_infer_nhwc.cpp     |   0
 client_example/14_instance_id/CMakeLists.txt  |   0
 .../batchnorm_fwd_instance_id.cpp             |   0
 .../15_convnd_bwd_data/CMakeLists.txt         |   0
 client_example/15_convnd_bwd_data/common.hpp  |   0
 .../conv3d_bwd_data_fp16.cpp                  |   0
 .../conv3d_bwd_data_fp32.cpp                  |   0
 .../15_gemm_add_multiply/CMakeLists.txt       |   0
 .../gemm_add_multiply.cpp                     |   0
 client_example/15_reduce/CMakeLists.txt       |   0
 client_example/15_reduce/reduce_nhwc_c.cpp    |   0
 client_example/16_convnd_fwd/CMakeLists.txt   |   0
 client_example/16_convnd_fwd/common.hpp       |   0
 .../16_convnd_fwd/conv3d_fwd_fp16.cpp         |   0
 .../conv3d_fwd_fp16_comp_fp8.cpp              |   0
 .../16_convnd_fwd/conv3d_fwd_fp32.cpp         |   0
 .../17_grouped_gemm_fastgelu/CMakeLists.txt   |   0
 .../grouped_gemm_fastgelu.cpp                 |   0
 client_example/18_groupnorm/CMakeLists.txt    |   0
 .../18_groupnorm/groupnorm_bwd_data.cpp       |   0
 .../18_groupnorm/groupnorm_bwd_gamma_beta.cpp |   0
 .../18_groupnorm/groupnorm_swish_fwd.cpp      |   0
 client_example/19_pool/CMakeLists.txt         |   0
 client_example/19_pool/avg_pool3d_bwd.cpp     |   0
 client_example/19_pool/avg_pool3d_fwd.cpp     |   0
 client_example/19_pool/max_pool2d_bwd.cpp     |   0
 client_example/19_pool/max_pool2d_fwd.cpp     |   0
 client_example/20_splitk_gemm/CMakeLists.txt  |   0
 .../20_splitk_gemm/splitK_gemm_fp16_f8.cpp    |   0
 .../21_grouped_gemm_bias/CMakeLists.txt       |   0
 .../grouped_gemm_fixed_nk_bias_fp16.cpp       |   0
 client_example/22_grouped_gemm/CMakeLists.txt |   0
 .../grouped_gemm_fixed_nk_bf16.cpp            |   0
 .../grouped_gemm_fixed_nk_fp16.cpp            |   0
 .../grouped_gemm_fixed_nk_fp8.cpp             |   0
 .../grouped_gemm_fixed_nk_i8.cpp              |   0
 .../22_im2col_col2im/CMakeLists.txt           |   0
 .../22_im2col_col2im/column_to_image.cpp      |   0
 .../22_im2col_col2im/image_to_column.cpp      |   0
 .../23_elementwise_transpose/CMakeLists.txt   |   0
 .../elementwise_transpose_3d.cpp              |   0
 .../24_grouped_conv_activation/CMakeLists.txt |   0
 ...d_conv_bwd_data_bilinear_residual_fp16.cpp |   0
 .../grouped_conv_bwd_data_scale_fp16.cpp      |   0
 ...rouped_conv_fwd_bilinear_residual_fp16.cpp |   0
 .../grouped_conv_fwd_scale_fp16.cpp           |   0
 .../grouped_conv_fwd_scaleadd_ab.inc          |   0
 .../grouped_conv_fwd_scaleadd_ab_bf16.cpp     |   0
 .../grouped_conv_fwd_scaleadd_ab_fp16.cpp     |   0
 .../grouped_conv_fwd_scaleadd_ab_fp32.cpp     |   0
 .../grouped_conv_fwd_scaleadd_ab_int8.cpp     |   0
 ...rouped_conv_fwd_scaleadd_scaleadd_relu.inc |   0
 ...d_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp |   0
 ...d_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp |   0
 ...d_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp |   0
 ...d_conv_fwd_scaleadd_scaleadd_relu_int8.cpp |   0
 client_example/25_wrapper/CMakeLists.txt      |   0
 client_example/25_wrapper/README.md           |   0
 .../tensor_transform_using_wrapper.cpp        |   0
 .../25_wrapper/wrapper_basic_gemm.cpp         |   0
 client_example/25_wrapper/wrapper_img2col.cpp |   0
 .../25_wrapper/wrapper_optimized_gemm.cpp     |   0
 client_example/CMakeLists.txt                 |   0
 client_example/README.md                      |   0
 cmake/Analyzers.cmake                         |   0
 cmake/ClangTidy.cmake                         |   0
 cmake/CppCheck.cmake                          |   0
 cmake/DoxygenDoc.cmake                        |   0
 cmake/Embed.cmake                             |   0
 cmake/EnableCompilerWarnings.cmake            |   0
 cmake/TargetFlags.cmake                       |   0
 cmake/getopt.cmake                            |   0
 cmake/gtest.cmake                             |   0
 codegen/CMakeLists.txt                        |   0
 codegen/driver/main.cpp                       |   0
 .../ck/host/device_gemm_multiple_d.hpp        |   0
 .../host/device_gemm_multiple_d/operation.hpp |   0
 .../host/device_gemm_multiple_d/problem.hpp   |   0
 codegen/include/ck/host/headers.hpp           |   0
 codegen/include/ck/host/operation/gemm.hpp    |   0
 codegen/include/ck/host/stringutils.hpp       |   0
 codegen/include/ck/host/types.hpp             |   0
 codegen/include/ck/host/utils.hpp             |   0
 codegen/src/device_gemm_multiple_d.cpp        |   0
 ...gemm_multiple_d_operation_xdl_cshuffle.cpp |   0
 codegen/src/headers.cpp                       |   0
 codegen/src/types.cpp                         |   0
 codegen/src/utils.cpp                         |   0
 codegen/test/CMakeLists.txt                   |   0
 codegen/test/gemm_multiple_d.cpp              |   0
 codegen/test/include/test.hpp                 |   0
 codegen/test/rtc/CMakeLists.txt               |   0
 .../test/rtc/include/rtc/compile_kernel.hpp   |   0
 codegen/test/rtc/include/rtc/hip.hpp          |   0
 codegen/test/rtc/include/rtc/kernel.hpp       |   0
 codegen/test/rtc/include/rtc/manage_ptr.hpp   |   0
 codegen/test/rtc/include/rtc/tmp_dir.hpp      |   0
 codegen/test/rtc/src/compile_kernel.cpp       |   0
 codegen/test/rtc/src/hip.cpp                  |   0
 codegen/test/rtc/src/kernel.cpp               |   0
 codegen/test/rtc/src/tmp_dir.cpp              |   0
 dev-requirements.txt                          |   0
 docs/Contributors_Guide.rst                   |   0
 docs/conceptual/what-is-ck.rst                |   0
 docs/conf.py                                  |   0
 docs/data/ck_component.png                    | Bin
 docs/data/ck_layer.png                        | Bin
 docs/doxygen/Doxyfile                         |   0
 docs/index.rst                                |   0
 docs/install/dockerhub.rst                    |   0
 docs/license.rst                              |   0
 docs/reference/API_Reference_Guide.rst        |   0
 docs/reference/Supported_Primitives_Guide.rst |   0
 docs/reference/wrapper.rst                    |   0
 docs/refs.bib                                 |   0
 docs/sphinx/_toc.yml.in                       |   0
 docs/sphinx/requirements.in                   |   0
 docs/sphinx/requirements.txt                  |   0
 docs/tutorial/tutorial_hello_world.rst        |   0
 example/01_gemm/CMakeLists.txt                |   0
 example/01_gemm/README.md                     |   0
 example/01_gemm/common.hpp                    |   0
 example/01_gemm/gemm_dl_fp16.cpp              |   0
 example/01_gemm/gemm_dl_fp32.cpp              |   0
 example/01_gemm/gemm_dl_int4.cpp              |   0
 example/01_gemm/gemm_dl_int8.cpp              |   0
 example/01_gemm/gemm_dpp_fp16.cpp             |   0
 example/01_gemm/gemm_wmma_fp16.cpp            |   0
 example/01_gemm/gemm_xdl_bf16.cpp             |   0
 example/01_gemm/gemm_xdl_bf16_rtn.cpp         |   0
 example/01_gemm/gemm_xdl_fp16.cpp             |   0
 example/01_gemm/gemm_xdl_fp16_fp8.cpp         |   0
 example/01_gemm/gemm_xdl_fp16_v2.cpp          |   0
 example/01_gemm/gemm_xdl_fp64.cpp             |   0
 example/01_gemm/gemm_xdl_fp8.cpp              |   0
 example/01_gemm/gemm_xdl_fp8_bf8.cpp          |   0
 example/01_gemm/gemm_xdl_int4.cpp             |   0
 example/01_gemm/gemm_xdl_int8.cpp             |   0
 .../01_gemm/gemm_xdl_lds_direct_load_fp16.cpp |   0
 .../01_gemm/gemm_xdl_lds_direct_load_fp32.cpp |   0
 example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp  |   0
 example/01_gemm/gemm_xdl_streamk.cpp          |   0
 example/01_gemm/gemm_xdl_wavelet_fp16.cpp     |   0
 example/01_gemm/run_gemm_example.inc          |   0
 example/02_gemm_bilinear/CMakeLists.txt       |   0
 example/02_gemm_bilinear/README.md            |   0
 .../gemm_bilinear_wmma_fp16.cpp               |   0
 .../gemm_bilinear_wmma_int8.cpp               |   0
 .../gemm_bilinear_xdl_fp16.cpp                |   0
 example/03_gemm_bias_relu/CMakeLists.txt      |   0
 example/03_gemm_bias_relu/README.md           |   0
 .../gemm_bias_relu_xdl_fp16.cpp               |   0
 .../04_gemm_add_add_fastgelu/CMakeLists.txt   |   0
 example/04_gemm_add_add_fastgelu/README.md    |   0
 example/04_gemm_add_add_fastgelu/common.hpp   |   0
 .../gemm_add_add_fastgelu_xdl_bf16.cpp        |   0
 .../gemm_add_add_fastgelu_xdl_fp16.cpp        |   0
 .../gemm_add_add_fastgelu_xdl_fp32.cpp        |   0
 .../gemm_add_add_fastgelu_xdl_int4.cpp        |   0
 .../gemm_add_add_fastgelu_xdl_int8.cpp        |   0
 ..._add_fastgelu_xdl_lds_direct_load_fp32.cpp |   0
 .../run_gemm_add_add_fastgelu_example.inc     |   0
 example/09_convnd_fwd/CMakeLists.txt          |   0
 example/09_convnd_fwd/README.md               |   0
 example/09_convnd_fwd/convnd_fwd_common.hpp   |   0
 .../09_convnd_fwd/convnd_fwd_dl_common.hpp    |   0
 example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp  |   0
 example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp  |   0
 example/09_convnd_fwd/convnd_fwd_dl_int8.cpp  |   0
 example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp |   0
 example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp |   0
 example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp |   0
 example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp |   0
 example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp |   0
 .../run_convnd_fwd_dl_example.inc             |   0
 .../09_convnd_fwd/run_convnd_fwd_example.inc  |   0
 .../CMakeLists.txt                            |   0
 .../common.hpp                                |   0
 .../convnd_fwd_max_xdl_bf16.cpp               |   0
 .../convnd_fwd_max_xdl_fp16.cpp               |   0
 .../convnd_fwd_max_xdl_fp32.cpp               |   0
 .../convnd_fwd_max_xdl_int4.cpp               |   0
 .../convnd_fwd_max_xdl_int8.cpp               |   0
 .../run_convnd_fwd_max_example.inc            |   0
 example/12_reduce/CMakeLists.txt              |   0
 example/12_reduce/README.md                   |   0
 example/12_reduce/reduce_blockwise.cpp        |   0
 example/12_reduce/reduce_blockwise_impl.hpp   |   0
 .../12_reduce/reduce_blockwise_two_call.cpp   |   0
 example/12_reduce/reduce_example_common.hpp   |   0
 .../reduce_multiblock_atomic_add.cpp          |   0
 .../reduce_multiblock_atomic_add_impl.hpp     |   0
 example/13_pool2d_fwd/CMakeLists.txt          |   0
 example/13_pool2d_fwd/README.md               |   0
 example/13_pool2d_fwd/pool2d_fwd_common.hpp   |   0
 example/13_pool2d_fwd/pool2d_fwd_fp16.cpp     |   0
 example/13_pool2d_fwd/pool2d_fwd_fp32.cpp     |   0
 example/14_gemm_quantization/CMakeLists.txt   |   0
 .../gemm_dl_quantization_int8.cpp             |   0
 .../gemm_xdl_bias_relu_quantization_int8.cpp  |   0
 .../gemm_xdl_quantization_int8.cpp            |   0
 example/15_grouped_gemm/CMakeLists.txt        |   0
 example/15_grouped_gemm/README.md             |   0
 .../grouped_gemm_multiple_d_dl_fp16.cpp       |   0
 .../15_grouped_gemm/grouped_gemm_xdl_bf16.cpp |   0
 .../grouped_gemm_xdl_fixed_nk_bias_fp16.cpp   |   0
 .../grouped_gemm_xdl_fixed_nk_fp16.cpp        |   0
 .../grouped_gemm_xdl_fixed_nk_fp8.cpp         |   0
 .../15_grouped_gemm/grouped_gemm_xdl_fp16.cpp |   0
 .../15_grouped_gemm/grouped_gemm_xdl_fp32.cpp |   0
 .../15_grouped_gemm/grouped_gemm_xdl_int4.cpp |   0
 .../15_grouped_gemm/grouped_gemm_xdl_int8.cpp |   0
 .../grouped_gemm_xdl_splitk_fp16.cpp          |   0
 .../run_grouped_gemm_example.inc              |   0
 .../CMakeLists.txt                            |   0
 .../gemm_add_add_mean_meansquare_xdl_fp16.cpp |   0
 .../gemm_add_addsquare_xdl_int8.cpp           |   0
 .../gemm_max_xdl_bf16.cpp                     |   0
 .../gemm_max_xdl_fp16.cpp                     |   0
 .../gemm_max_xdl_fp32.cpp                     |   0
 .../gemm_max_xdl_int4.cpp                     |   0
 .../gemm_max_xdl_int8.cpp                     |   0
 .../gemm_mean_meansquare_xdl_bf16.cpp         |   0
 .../gemm_mean_meansquare_xdl_fp16.cpp         |   0
 .../gemm_mean_meansquare_xdl_fp32.cpp         |   0
 .../gemm_reduce_xdl_common.hpp                |   0
 example/17_convnd_bwd_data/CMakeLists.txt     |   0
 example/17_convnd_bwd_data/README.md          |   0
 .../convnd_bwd_data_common.hpp                |   0
 .../convnd_bwd_data_dl_fp16.cpp               |   0
 .../convnd_bwd_data_xdl_fp16.cpp              |   0
 example/18_batched_gemm_reduce/CMakeLists.txt |   0
 .../batched_gemm_reduce_xdl_fp16.cpp          |   0
 example/19_binary_elementwise/CMakeLists.txt  |   0
 .../broadcast_add_2d_amn_bn.cpp               |   0
 .../broadcast_add_3d_am_bmnk.cpp              |   0
 .../elementwise_add_1d.cpp                    |   0
 .../elementwise_add_4d.cpp                    |   0
 .../20_grouped_conv_bwd_weight/CMakeLists.txt |   0
 example/20_grouped_conv_bwd_weight/common.hpp |   0
 .../grouped_conv_bwd_weight_dl_fp16.cpp       |   0
 .../grouped_conv_bwd_weight_wmma_fp16.cpp     |   0
 .../grouped_conv_bwd_weight_xdl_bf16.cpp      |   0
 .../grouped_conv_bwd_weight_xdl_fp16.cpp      |   0
 ..._conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp |   0
 .../run_grouped_conv_bwd_weight_example.inc   |   0
 example/21_gemm_layernorm/CMakeLists.txt      |   0
 ...bias_relu_add_layernorm_xdl_naive_fp16.cpp |   0
 ...as_relu_add_layernorm_xdl_welford_fp16.cpp |   0
 .../gemm_layernorm_xdl_naive_fp16.cpp         |   0
 ...xdl_layernorm_naive_single_kernel_fp16.cpp |   0
 example/22_cgemm/CMakeLists.txt               |   0
 example/22_cgemm/cgemm_xdl_bf16.cpp           |   0
 example/22_cgemm/cgemm_xdl_common.hpp         |   0
 example/22_cgemm/cgemm_xdl_fp16.cpp           |   0
 example/22_cgemm/cgemm_xdl_fp32.cpp           |   0
 example/22_cgemm/cgemm_xdl_int4.cpp           |   0
 example/22_cgemm/cgemm_xdl_int8.cpp           |   0
 example/23_softmax/CMakeLists.txt             |   0
 example/23_softmax/README.md                  |   0
 example/23_softmax/softmax_blockwise.cpp      |   0
 example/24_batched_gemm/CMakeLists.txt        |   0
 .../24_batched_gemm/batched_gemm_xdl_bf16.cpp |   0
 .../24_batched_gemm/batched_gemm_xdl_fp16.cpp |   0
 .../24_batched_gemm/batched_gemm_xdl_fp32.cpp |   0
 .../24_batched_gemm/batched_gemm_xdl_int4.cpp |   0
 .../24_batched_gemm/batched_gemm_xdl_int8.cpp |   0
 .../run_batched_gemm_example.inc              |   0
 example/25_gemm_bias_e_permute/CMakeLists.txt |   0
 .../gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp |   0
 .../gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp |   0
 example/26_contraction/CMakeLists.txt         |   0
 example/26_contraction/README.md              |   0
 example/26_contraction/common_instances.hpp   |   0
 ...raction_bilinear_xdl_bf16_compute_fp32.cpp |   0
 ...raction_bilinear_xdl_fp16_compute_fp32.cpp |   0
 .../contraction_bilinear_xdl_fp32.cpp         |   0
 ...raction_bilinear_xdl_fp32_compute_bf16.cpp |   0
 ...raction_bilinear_xdl_fp32_compute_fp16.cpp |   0
 .../contraction_bilinear_xdl_fp64.cpp         |   0
 ...raction_bilinear_xdl_fp64_compute_fp32.cpp |   0
 ...ontraction_scale_xdl_bf16_compute_fp32.cpp |   0
 ...ontraction_scale_xdl_fp16_compute_fp32.cpp |   0
 .../contraction_scale_xdl_fp32.cpp            |   0
 ...ontraction_scale_xdl_fp32_compute_bf16.cpp |   0
 ...ontraction_scale_xdl_fp32_compute_fp16.cpp |   0
 .../contraction_scale_xdl_fp64.cpp            |   0
 ...ontraction_scale_xdl_fp64_compute_fp32.cpp |   0
 .../run_contraction_bilinear_example.inc      |   0
 .../run_contraction_scale_example.inc         |   0
 example/27_layernorm2d_fwd/CMakeLists.txt     |   0
 example/27_layernorm2d_fwd/common.hpp         |   0
 .../layernorm2d_fwd_fp16.cpp                  |   0
 .../layernorm2d_fwd_splitk_fp16.cpp           |   0
 .../run_layernorm_example.inc                 |   0
 .../CMakeLists.txt                            |   0
 .../grouped_gemm_bias_e_permute_xdl_fp16.cpp  |   0
 .../CMakeLists.txt                            |   0
 .../batched_gemm_bias_e_permute_wmma_fp16.cpp |   0
 .../batched_gemm_bias_e_permute_xdl_fp16.cpp  |   0
 .../CMakeLists.txt                            |   0
 .../30_grouped_conv_fwd_multiple_d/README.md  |   0
 .../30_grouped_conv_fwd_multiple_d/common.hpp |   0
 .../common_wmma.hpp                           |   0
 ...ouped_conv_fwd_bias_relu_add_wmma_fp16.cpp |   0
 ...ouped_conv_fwd_bias_relu_add_wmma_int8.cpp |   0
 ...rouped_conv_fwd_bias_relu_add_xdl_bf16.cpp |   0
 ...rouped_conv_fwd_bias_relu_add_xdl_fp16.cpp |   0
 ...rouped_conv_fwd_bias_relu_add_xdl_fp32.cpp |   0
 ...rouped_conv_fwd_bias_relu_add_xdl_int4.cpp |   0
 ...rouped_conv_fwd_bias_relu_add_xdl_int8.cpp |   0
 .../grouped_conv_fwd_xdl_fp16.cpp             |   0
 ...grouped_conv_fwd_bias_relu_add_example.inc |   0
 ...ed_conv_fwd_bias_relu_add_wmma_example.inc |   0
 .../run_grouped_conv_fwd_example.inc          |   0
 example/31_batched_gemm_gemm/CMakeLists.txt   |   0
 .../batched_gemm_gemm_xdl_bf16.cpp            |   0
 .../batched_gemm_gemm_xdl_fp16.cpp            |   0
 .../batched_gemm_gemm_xdl_fp32.cpp            |   0
 .../batched_gemm_gemm_xdl_int4.cpp            |   0
 .../batched_gemm_gemm_xdl_int8.cpp            |   0
 .../run_batched_gemm_gemm_example.inc         |   0
 .../CMakeLists.txt                            |   0
 ...e_scale_softmax_gemm_permute_wmma_fp16.cpp |   0
 ...le_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...m_scale_softmax_gemm_permute_wmma_fp16.cpp |   0
 ...mm_scale_softmax_gemm_permute_xdl_bf16.cpp |   0
 ...mm_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...tched_gemm_scale_softmax_gemm_xdl_bf16.cpp |   0
 ...tched_gemm_scale_softmax_gemm_xdl_fp16.cpp |   0
 .../cross_attention_forward_wmma_fp16.cpp     |   0
 ...le_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...mm_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...uped_query_attention_forward_wmma_fp16.cpp |   0
 ...ulti_query_attention_forward_wmma_fp16.cpp |   0
 .../run_batched_gemm_scale_softmax_gemm.inc   |   0
 ...atched_gemm_scale_softmax_gemm_permute.inc |   0
 ...d_gemm_scale_softmax_gemm_permute_wmma.inc |   0
 .../run_cross_attention_wmma.inc              |   0
 ...rouped_gemm_scale_softmax_gemm_permute.inc |   0
 ...n_grouped_query_attention_forward_wmma.inc |   0
 ...run_multi_query_attention_forward_wmma.inc |   0
 .../run_self_attention_wmma.inc               |   0
 .../self_attention_forward_wmma_fp16.cpp      |   0
 example/33_multiple_reduce/CMakeLists.txt     |   0
 example/33_multiple_reduce/README.md          |   0
 .../33_multiple_reduce/dual_reduce_common.hpp |   0
 .../dual_reduce_multiblock.cpp                |   0
 .../dual_reduce_threadwise.cpp                |   0
 example/34_batchnorm/CMakeLists.txt           |   0
 example/34_batchnorm/README.md                |   0
 .../34_batchnorm/batchnorm_backward_nhwc.cpp  |   0
 example/34_batchnorm/batchnorm_common.hpp     |   0
 .../batchnorm_forward_inferring_nhwc.cpp      |   0
 .../batchnorm_forward_training_nhwc.cpp       |   0
 ...tchnorm_forward_training_nhwc_obsolete.cpp |   0
 example/34_batchnorm/batchnorm_infer_impl.hpp |   0
 example/35_splitK_gemm/CMakeLists.txt         |   0
 .../run_splitK_gemm_example.inc               |   4 +
 .../35_splitK_gemm/splitK_gemm_xdl_bf16.cpp   |   0
 .../35_splitK_gemm/splitK_gemm_xdl_fp16.cpp   |   0
 .../splitK_gemm_xdl_fp16_fp8.cpp              |   0
 .../35_splitK_gemm/splitK_gemm_xdl_fp32.cpp   |   0
 .../35_splitK_gemm/splitK_gemm_xdl_int4.cpp   |   0
 .../35_splitK_gemm/splitK_gemm_xdl_int8.cpp   |   0
 .../splitK_gemm_xdl_lds_direct_load_fp16.cpp  |   0
 example/36_sparse_embedding/CMakeLists.txt    |   0
 .../sparse_embedding3_forward_layernorm.cpp   |   0
 .../CMakeLists.txt                            |   0
 ...ed_gemm_add_add_relu_gemm_add_xdl_fp16.cpp |   0
 .../CMakeLists.txt                            |   0
 .../common.hpp                                |   0
 ...ouped_conv_bwd_data_bias_relu_xdl_fp16.cpp |   0
 .../grouped_conv_bwd_data_wmma_fp16.cpp       |   0
 .../grouped_conv_bwd_data_xdl_fp16.cpp        |   0
 ...rouped_conv_bwd_data_bias_relu_example.inc |   0
 .../run_grouped_conv_bwd_data_example.inc     |   0
 example/39_permute/CMakeLists.txt             |   0
 example/39_permute/common.hpp                 |   0
 example/39_permute/permute_1xHxW_fp16.cpp     |   0
 example/39_permute/permute_HxWx4_fp16.cpp     |   0
 example/39_permute/permute_NxHxW_fp16.cpp     |   0
 .../39_permute/run_permute_bundle_example.inc |   0
 .../run_permute_element_example.inc           |   0
 .../40_conv2d_fwd_quantization/CMakeLists.txt |   0
 example/40_conv2d_fwd_quantization/common.hpp |   0
 ...bias_relu_perchannel_quantization_int8.cpp |   0
 ...l_bias_relu_perlayer_quantization_int8.cpp |   0
 ...bias_tanh_perchannel_quantization_int8.cpp |   0
 ...l_bias_tanh_perlayer_quantization_int8.cpp |   0
 ...2d_fwd_dl_perchannel_quantization_int8.cpp |   0
 ...nv2d_fwd_dl_perlayer_quantization_int8.cpp |   0
 ...bias_relu_perchannel_quantization_int8.cpp |   0
 ...l_bias_relu_perlayer_quantization_int8.cpp |   0
 ...d_fwd_xdl_perchannel_quantization_int8.cpp |   0
 ...v2d_fwd_xdl_perlayer_quantization_int8.cpp |   0
 ...d_bias_perchannel_quantization_example.inc |   0
 ...fwd_bias_perlayer_quantization_example.inc |   0
 ...2d_fwd_perchannel_quantization_example.inc |   0
 ...nv2d_fwd_perlayer_quantization_example.inc |   0
 .../41_grouped_conv_conv_fwd/CMakeLists.txt   |   0
 .../grouped_conv_conv_fwd_xdl_bf16.cpp        |   0
 .../grouped_conv_conv_fwd_xdl_fp16.cpp        |   0
 .../grouped_conv_conv_fwd_xdl_fp32.cpp        |   0
 .../grouped_conv_conv_fwd_xdl_int4.cpp        |   0
 .../grouped_conv_conv_fwd_xdl_int8.cpp        |   0
 .../run_grouped_conv_conv_fwd_example.inc     |   0
 example/42_groupnorm_fwd/CMakeLists.txt       |   0
 example/42_groupnorm_fwd/common.hpp           |   0
 .../groupnorm_fwd_sigmoid_mul_fp16.cpp        |   0
 .../groupnorm_fwd_splitk_fp16.cpp             |   0
 .../groupnorm_fwd_swish_fp16.cpp              |   0
 .../run_groupnorm_fwd_example.inc             |   0
 .../CMakeLists.txt                            |   0
 .../splitk_gemm_bias_e_permute_xdl_fp16.cpp   |   0
 .../splitk_gemm_bias_e_permute_xdl_fp32.cpp   |   0
 example/44_elementwise_permute/CMakeLists.txt |   0
 .../elementwise_permute.cpp                   |   0
 .../elementwise_permute_3d.cpp                |   0
 .../elementwise_permute_4D_fp16.cpp           |   0
 .../elementwise_permute_4D_fp16_2d.cpp        |   0
 .../elementwise_permute_4D_fp16_col.cpp       |   0
 .../elementwise_permute_4D_fp16_row.cpp       |   0
 .../elementwise_permute_4D_fp32_col.cpp       |   0
 .../elementwise_permute_4D_fp32_row.cpp       |   0
 .../CMakeLists.txt                            |   0
 .../elementwise_layernorm_blockwise.cpp       |   0
 example/46_gemm_add_multiply/CMakeLists.txt   |   0
 example/46_gemm_add_multiply/README.md        |   0
 example/46_gemm_add_multiply/common.hpp       |   0
 .../gemm_add_multiply_dl_fp16.cpp             |   0
 .../gemm_add_multiply_xdl_fp16.cpp            |   0
 .../run_gemm_add_multiply_example.inc         |   0
 .../CMakeLists.txt                            |   0
 .../gemm_bias_softmax_gemm_permute.cpp        |   0
 example/48_pool3d_fwd/CMakeLists.txt          |   0
 example/48_pool3d_fwd/pool3d_fwd_common.hpp   |   0
 example/48_pool3d_fwd/pool3d_fwd_fp16.cpp     |   0
 example/49_maxpool2d_bwd/CMakeLists.txt       |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp   |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_common.hpp |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp   |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp   |   0
 example/50_put_element/CMakeLists.txt         |   0
 example/50_put_element/put_element_fp16.cpp   |   0
 example/51_avgpool3d_bwd/CMakeLists.txt       |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp   |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_common.hpp |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp   |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp   |   0
 example/52_im2col_col2im/CMakeLists.txt       |   0
 .../52_im2col_col2im/column_to_image_f32.cpp  |   0
 example/52_im2col_col2im/common.hpp           |   0
 .../52_im2col_col2im/image_to_column_f32.cpp  |   0
 example/53_layernorm2d_bwd/CMakeLists.txt     |   0
 .../layernorm2d_bwd_fp32.cpp                  |   0
 example/54_groupnorm_bwd/CMakeLists.txt       |   0
 .../54_groupnorm_bwd/groupnorm_bwd_fp32.cpp   |   0
 example/60_gemm_multi_ABD/CMakeLists.txt      |   0
 .../gemm_multi_ABD_xdl_fp16.cpp               |   0
 .../61_contraction_multi_ABD/CMakeLists.txt   |   0
 .../contraction_multi_ABD_xdl_fp16.cpp        |   0
 example/62_convnd_activ/CMakeLists.txt        |   0
 example/62_convnd_activ/binary/CMakeLists.txt |   0
 ...nd_bwd_data_xdl_bilinear_residual_fp16.cpp |   0
 .../convnd_fwd_xdl_bilinear_residual_fp16.cpp |   0
 ...aleadd_scaleadd_relu_bcasted_bias_fp16.cpp |   0
 ...nd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp |   0
 .../62_convnd_activ/multi_AB/CMakeLists.txt   |   0
 .../conv_fwd_xdl_scaleadd_ab_bf16.cpp         |   0
 .../conv_fwd_xdl_scaleadd_ab_fp16.cpp         |   0
 .../conv_fwd_xdl_scaleadd_ab_fp32.cpp         |   0
 .../conv_fwd_xdl_scaleadd_ab_int8.cpp         |   0
 .../convnd_fwd_activ_multi_ab_common.hpp      |   0
 .../run_convnd_activ_example.inc              |   0
 example/62_convnd_activ/unary/CMakeLists.txt  |   0
 .../unary/convnd_fwd_activ_unary_common.hpp   |   0
 .../unary/convnd_fwd_xdl_abs_fp16.cpp         |   0
 .../unary/convnd_fwd_xdl_clippedrelu_fp16.cpp |   0
 .../unary/convnd_fwd_xdl_elu_fp16.cpp         |   0
 .../unary/convnd_fwd_xdl_leakyrelu_fp16.cpp   |   0
 .../unary/convnd_fwd_xdl_pow_fp16.cpp         |   0
 .../unary/convnd_fwd_xdl_relu_fp16.cpp        |   0
 .../unary/convnd_fwd_xdl_sigmoid_fp16.cpp     |   0
 .../unary/convnd_fwd_xdl_softrelu_fp16.cpp    |   0
 .../unary/convnd_fwd_xdl_tanh_fp16.cpp        |   0
 example/63_layernorm4d_fwd/CMakeLists.txt     |   0
 example/63_layernorm4d_fwd/common.hpp         |   0
 .../layernorm4d_fwd_fp16.cpp                  |   0
 .../layernorm4d_fwd_splitk_fp16.cpp           |   0
 .../run_layernorm4d_fwd_example.inc           |   0
 example/64_fpAintB_gemm/CMakeLists.txt        |   0
 example/64_fpAintB_gemm/common.hpp            |   0
 .../64_fpAintB_gemm/fp16int8_gemm_wmma.cpp    |   0
 example/64_fpAintB_gemm/run_gemm_example.inc  |   0
 example/CMakeLists.txt                        |   0
 include/ck/ck.hpp                             |   0
 include/ck/config.h.in                        |   0
 include/ck/host_utility/device_prop.hpp       |   0
 include/ck/host_utility/hip_check_error.hpp   |   0
 include/ck/host_utility/io.hpp                |   0
 include/ck/host_utility/kernel_launch.hpp     |   0
 include/ck/host_utility/stream_utility.hpp    |   0
 ...n3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp |   0
 include/ck/stream_config.hpp                  |   0
 include/ck/tensor/static_tensor.hpp           |   0
 .../tensor_description/cluster_descriptor.hpp |   0
 .../multi_index_transform.hpp                 |   0
 .../multi_index_transform_helper.hpp          |   0
 .../ck/tensor_description/tensor_adaptor.hpp  |   0
 .../tensor_description/tensor_descriptor.hpp  |   0
 .../tensor_descriptor_helper.hpp              |   0
 .../tensor_space_filling_curve.hpp            |   0
 .../gpu/block/blockwise_gemm_dl_v2r3.hpp      |   0
 .../gpu/block/blockwise_gemm_dlops_v2r2.hpp   |   0
 .../gpu/block/blockwise_gemm_dlops_v3.hpp     |   0
 .../gpu/block/blockwise_gemm_dpp.hpp          |   0
 .../block/blockwise_gemm_pipeline_xdlops.hpp  |   0
 .../gpu/block/blockwise_gemm_wmma.hpp         |   0
 .../gpu/block/blockwise_gemm_xdlops.hpp       |   0
 .../blockwise_gemm_xdlops_skip_b_lds.hpp      |   0
 .../gpu/block/blockwise_softmax.hpp           |   0
 .../blockwise_tensor_slice_transfer_v5r1.hpp  |   0
 .../gpu/block/blockwise_welford.hpp           |   0
 .../block/reduction_functions_blockwise.hpp   |   0
 ...roup_tensor_slice_transfer_direct_load.hpp |   0
 ...hread_group_tensor_slice_transfer_v4r1.hpp |   0
 ...oup_tensor_slice_transfer_v4r1_dequant.hpp |   0
 ...hread_group_tensor_slice_transfer_v6r1.hpp |   0
 ...ead_group_tensor_slice_transfer_v6r1r2.hpp |   0
 ...hread_group_tensor_slice_transfer_v6r2.hpp |   0
 ...hread_group_tensor_slice_transfer_v6r3.hpp |   0
 .../thread_group_tensor_slice_transfer_v7.hpp |   0
 ...hread_group_tensor_slice_transfer_v7r2.hpp |   0
 .../gpu/device/conv_tensor_rearrange_op.hpp   |   0
 ...nvolution_backward_data_specialization.hpp |   0
 ...olution_backward_weight_specialization.hpp |   0
 .../convolution_forward_specialization.hpp    |   0
 .../gpu/device/device_avgpool_bwd.hpp         |   0
 .../gpu/device/device_base.hpp                |   0
 .../device_batched_contraction_multiple_d.hpp |   0
 .../gpu/device/device_batched_gemm.hpp        |   0
 .../device/device_batched_gemm_e_permute.hpp  |   0
 .../gpu/device/device_batched_gemm_gemm.hpp   |   0
 .../device/device_batched_gemm_multi_d.hpp    |   0
 ...atched_gemm_multiple_d_gemm_multiple_d.hpp |   0
 .../device_batched_gemm_softmax_gemm.hpp      |   0
 ...vice_batched_gemm_softmax_gemm_permute.hpp |   0
 .../gpu/device/device_batchnorm_backward.hpp  |   0
 .../gpu/device/device_batchnorm_forward.hpp   |   0
 .../gpu/device/device_batchnorm_infer.hpp     |   0
 .../gpu/device/device_cgemm.hpp               |   0
 .../device_contraction_multiple_abd.hpp       |   0
 .../device/device_contraction_multiple_d.hpp  |   0
 .../gpu/device/device_conv_bwd_data.hpp       |   0
 .../gpu/device/device_conv_fwd.hpp            |   0
 .../device_conv_fwd_bias_activation.hpp       |   0
 .../device_conv_fwd_bias_activation_add.hpp   |   0
 .../device/device_conv_tensor_rearrange.hpp   |   0
 .../gpu/device/device_elementwise.hpp         |   0
 .../device_elementwise_normalization.hpp      |   0
 .../gpu/device/device_elementwise_scale.hpp   |   0
 .../gpu/device/device_gemm.hpp                |   0
 .../gpu/device/device_gemm_bias_e_permute.hpp |   0
 .../gpu/device/device_gemm_dequantB.hpp       |   0
 .../gpu/device/device_gemm_multiple_abd.hpp   |   0
 .../gpu/device/device_gemm_multiple_d.hpp     |   0
 .../device_gemm_multiple_d_layernorm.hpp      |   0
 .../device_gemm_multiple_d_multiple_r.hpp     |   0
 .../gpu/device/device_gemm_reduce.hpp         |   0
 .../gpu/device/device_gemm_splitk.hpp         |   0
 .../gpu/device/device_gemm_streamk.hpp        |   0
 .../device_grouped_contraction_multiple_d.hpp |   0
 ...evice_grouped_conv_bwd_data_multiple_d.hpp |   0
 .../device/device_grouped_conv_bwd_weight.hpp |   0
 .../gpu/device/device_grouped_conv_fwd.hpp    |   0
 .../device_grouped_conv_fwd_multiple_abd.hpp  |   0
 .../device_grouped_conv_fwd_multiple_d.hpp    |   0
 .../gpu/device/device_grouped_gemm.hpp        |   0
 .../device/device_grouped_gemm_fixed_nk.hpp   |   0
 ...vice_grouped_gemm_softmax_gemm_permute.hpp |   0
 .../gpu/device/device_grouped_gemm_splitk.hpp |   0
 .../gpu/device/device_max_pool_bwd.hpp        |   0
 .../gpu/device/device_multiple_reduce.hpp     |   0
 .../device/device_normalization_bwd_data.hpp  |   0
 .../device_normalization_bwd_gamma_beta.hpp   |   0
 .../gpu/device/device_normalization_fwd.hpp   |   0
 .../gpu/device/device_permute.hpp             |   0
 .../gpu/device/device_pool_fwd.hpp            |   0
 .../gpu/device/device_put_element.hpp         |   0
 .../gpu/device/device_reduce.hpp              |   0
 .../gpu/device/device_softmax.hpp             |   0
 .../device_splitk_contraction_multiple_d.hpp  |   0
 .../gpu/device/gemm_specialization.hpp        |   0
 .../impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp |   0
 ...d_contraction_multiple_d_wmma_cshuffle.hpp |   0
 ...ed_contraction_multiple_d_xdl_cshuffle.hpp |   0
 .../device_batched_gemm_e_permute_xdl.hpp     |   0
 .../device_batched_gemm_gemm_xdl_cshuffle.hpp |   0
 .../impl/device_batched_gemm_multi_d_xdl.hpp  |   0
 .../device_batched_gemm_multiple_d_dl.hpp     |   0
 ...ultiple_d_gemm_multiple_d_xdl_cshuffle.hpp |   0
 ...evice_batched_gemm_reduce_xdl_cshuffle.hpp |   0
 ...emm_softmax_gemm_permute_wmma_cshuffle.hpp |   0
 ...gemm_softmax_gemm_permute_xdl_cshuffle.hpp |   0
 ...batched_gemm_softmax_gemm_xdl_cshuffle.hpp |   0
 .../device/impl/device_batched_gemm_xdl.hpp   |   0
 .../impl/device_batchnorm_backward_impl.hpp   |   0
 .../impl/device_batchnorm_forward_impl.hpp    |   0
 ...device_batchnorm_forward_impl_obsolete.hpp |   0
 .../impl/device_cgemm_4gemm_xdl_cshuffle.hpp  |   0
 .../impl/device_column_to_image_impl.hpp      |   0
 ..._contraction_multiple_abd_xdl_cshuffle.hpp |   0
 ...ce_contraction_multiple_d_xdl_cshuffle.hpp |   0
 .../device/impl/device_contraction_utils.hpp  |   0
 ...rd_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |   0
 ...ice_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp |   0
 ...fle_bias_activation_add_nhwc_kyxc_nhwk.hpp |   0
 ...shuffle_bias_activation_nhwc_kyxc_nhwk.hpp |   0
 ...onv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |   0
 .../device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp  |   0
 ...ice_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp |   0
 ...evice_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp |   0
 .../device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp |   0
 ...device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp |   0
 .../impl/device_elementwise_2d_impl.hpp       |   0
 .../impl/device_elementwise_3d_impl.hpp       |   0
 .../device/impl/device_elementwise_impl.hpp   |   0
 .../device_elementwise_normalization_impl.hpp |   0
 .../impl/device_elementwise_scale_impl.hpp    |   0
 .../device/impl/device_fpAintB_gemm_wmma.hpp  |   0
 ...vice_gemm_bias_add_reduce_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_gemm_dl.hpp        |   0
 .../gpu/device/impl/device_gemm_dpp.hpp       |   0
 .../device_gemm_multiple_abd_xdl_cshuffle.hpp |   0
 .../device/impl/device_gemm_multiple_d_dl.hpp |   0
 ...gemm_multiple_d_layernorm_xdl_cshuffle.hpp |   0
 ...emm_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 .../device_gemm_multiple_d_wmma_cshuffle.hpp  |   0
 .../device_gemm_multiple_d_xdl_cshuffle.hpp   |   0
 ...ultiple_d_xdl_cshuffle_lds_direct_load.hpp |   0
 .../impl/device_gemm_reduce_xdl_cshuffle.hpp  |   0
 .../gpu/device/impl/device_gemm_wmma.hpp      |   0
 .../gpu/device/impl/device_gemm_xdl.hpp       |   0
 .../device/impl/device_gemm_xdl_cshuffle.hpp  |   0
 ...vice_gemm_xdl_cshuffle_lds_direct_load.hpp |   0
 .../impl/device_gemm_xdl_cshuffle_v2.hpp      |   0
 .../device_gemm_xdl_layernorm_cshuffle.hpp    |   0
 .../impl/device_gemm_xdl_skip_b_lds.hpp       |   0
 .../impl/device_gemm_xdl_splitk_c_shuffle.hpp |   0
 ...m_xdl_splitk_c_shuffle_lds_direct_load.hpp |   0
 .../device/impl/device_gemm_xdl_streamk.hpp   |   0
 .../device_gemm_xdl_waveletmodel_cshuffle.hpp |   0
 ...ed_contraction_multiple_d_xdl_cshuffle.hpp |   0
 ...conv_bwd_data_multiple_d_wmma_cshuffle.hpp |   0
 ...nv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp |   0
 .../device_grouped_conv_bwd_weight_dl.hpp     |   0
 ..._grouped_conv_bwd_weight_wmma_cshuffle.hpp |   0
 ...e_grouped_conv_bwd_weight_xdl_cshuffle.hpp |   0
 ..._conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp |   0
 ...ice_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp |   0
 ...ped_conv_fwd_multiple_abd_xdl_cshuffle.hpp |   0
 ...grouped_conv_fwd_multiple_d_multiple_r.hpp |   0
 ...fwd_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 ...uped_conv_fwd_multiple_d_wmma_cshuffle.hpp |   0
 ...ouped_conv_fwd_multiple_d_xdl_cshuffle.hpp |   0
 .../device/impl/device_grouped_conv_utils.hpp |   0
 .../device_grouped_gemm_multiple_d_dl.hpp     |   0
 ...gemm_softmax_gemm_permute_xdl_cshuffle.hpp |   0
 .../device/impl/device_grouped_gemm_xdl.hpp   |   0
 .../impl/device_grouped_gemm_xdl_fixed_nk.hpp |   0
 ...evice_grouped_gemm_xdl_splitk_cshuffle.hpp |   0
 ...e_grouped_query_attention_forward_wmma.hpp |   0
 .../impl/device_image_to_column_impl.hpp      |   0
 .../device/impl/device_max_pool_bwd_impl.hpp  |   0
 ...ice_multi_query_attention_forward_wmma.hpp |   0
 .../device_multiple_reduce_multiblock.hpp     |   0
 .../device_multiple_reduce_threadwise.hpp     |   0
 .../device_normalization_bwd_data_impl.hpp    |   0
 ...vice_normalization_bwd_gamma_beta_impl.hpp |   0
 .../impl/device_normalization_fwd_impl.hpp    |   0
 .../device_normalization_fwd_splitk_impl.hpp  |   0
 .../gpu/device/impl/device_permute_impl.hpp   |   0
 .../impl/device_pool2d_fwd_nhwc_nhwc.hpp      |   0
 .../impl/device_pool3d_fwd_ndhwc_ndhwc.hpp    |   0
 .../device/impl/device_put_element_impl.hpp   |   0
 .../gpu/device/impl/device_reduce_common.hpp  |   0
 .../device/impl/device_reduce_multiblock.hpp  |   0
 .../device/impl/device_reduce_threadwise.hpp  |   0
 .../gpu/device/impl/device_softmax_impl.hpp   |   0
 ...ce_sparse_embeddings_forward_layernorm.hpp |   0
 ...tk_contraction_multiple_d_xdl_cshuffle.hpp |   0
 .../gpu/device/masking_specialization.hpp     |   0
 .../gpu/device/matrix_padder.hpp              |   0
 .../gpu/device/reduction_operator_mapping.hpp |   0
 .../gpu/device/tensor_layout.hpp              |   0
 .../gpu/device/tensor_specialization.hpp      |   0
 .../gpu/device/welford_helper.hpp             |   0
 .../element/binary_element_wise_operation.hpp |   0
 .../gpu/element/element_wise_operation.hpp    |   0
 .../gpu/element/quantization_operation.hpp    |   0
 .../element/unary_element_wise_operation.hpp  |   0
 .../gridwise_multiblock_batchnorm_forward.hpp |   0
 ...e_second_half_batchnorm_backward_final.hpp |   0
 ...gridwise_multiblock_welford_first_half.hpp |   0
 ..._half_batchnorm_forward_final_obsolete.hpp |   0
 ...cond_half_multiblock_reduce_first_half.hpp |   0
 .../gpu/grid/block_to_ctile_map.hpp           |   0
 ...iple_d_welford_first_half_xdl_cshuffle.hpp |   0
 ...idwise_welford_second_half_layernorm2d.hpp |   0
 ...dwise_2d_multiple_reduction_multiblock.hpp |   0
 ...dwise_2d_multiple_reduction_threadwise.hpp |   0
 .../grid/gridwise_2d_reduction_multiblock.hpp |   0
 .../grid/gridwise_2d_reduction_threadwise.hpp |   0
 ...wise_batched_gemm_gemm_xdl_cshuffle_v1.hpp |   0
 ...iple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp |   0
 ...ultiple_d_softmax_gemm_xdl_cshuffle_v1.hpp |   0
 ...atched_gemm_softmax_gemm_wmma_cshuffle.hpp |   0
 ...ched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp |   0
 ...e_batchnorm_backward_blockwise_welford.hpp |   0
 ...se_batchnorm_forward_blockwise_welford.hpp |   0
 .../gpu/grid/gridwise_elementwise_1d.hpp      |   0
 .../grid/gridwise_elementwise_1d_scale.hpp    |   0
 .../gpu/grid/gridwise_elementwise_2d.hpp      |   0
 .../gpu/grid/gridwise_elementwise_3d.hpp      |   0
 ...elementwise_layernorm_welford_variance.hpp |   0
 .../gpu/grid/gridwise_fpAintB_gemm_wmma.hpp   |   0
 ...e_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp |   0
 .../gpu/grid/gridwise_gemm_dl_multiple_d.hpp  |   0
 .../gpu/grid/gridwise_gemm_dl_v1r3.hpp        |   0
 .../gpu/grid/gridwise_gemm_dpp.hpp            |   0
 ...ridwise_gemm_multiple_abd_xdl_cshuffle.hpp |   0
 ...emm_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 ...gridwise_gemm_multiple_d_wmma_cshuffle.hpp |   0
 .../gridwise_gemm_multiple_d_xdl_cshuffle.hpp |   0
 ...ultiple_d_xdl_cshuffle_lds_direct_load.hpp |   0
 ...se_gemm_multiple_d_xdl_splitk_cshuffle.hpp |   0
 .../grid/gridwise_gemm_pipeline_selector.hpp  |   0
 .../gpu/grid/gridwise_gemm_pipeline_v1.hpp    |   0
 .../gpu/grid/gridwise_gemm_pipeline_v2.hpp    |   0
 .../gpu/grid/gridwise_gemm_pipeline_v3.hpp    |   0
 .../gridwise_gemm_pipeline_v4_direct_load.hpp |   0
 .../gridwise_gemm_reduce_xdl_cshuffle_v1.hpp  |   0
 ...e_gemm_split_k_multiple_d_xdl_cshuffle.hpp |   0
 ...emm_split_k_multiple_d_xdl_cshuffle_v2.hpp |   0
 .../gpu/grid/gridwise_gemm_waveletmodel.hpp   |   0
 .../gpu/grid/gridwise_gemm_wmma.hpp           |   0
 .../grid/gridwise_gemm_xdl_cshuffle_v1.hpp    |   0
 .../grid/gridwise_gemm_xdl_cshuffle_v2.hpp    |   0
 ...ridwise_gemm_xdl_layernorm_cshuffle_v1.hpp |   0
 ...ridwise_gemm_xdl_waveletmodel_cshuffle.hpp |   0
 .../grid/gridwise_gemm_xdlops_bwd_weight.hpp  |   0
 .../gridwise_gemm_xdlops_skip_b_lds_v1.hpp    |   0
 ...ise_gemm_xdlops_splitk_lds_direct_load.hpp |   0
 .../gpu/grid/gridwise_gemm_xdlops_streamk.hpp |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r3.hpp    |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r4.hpp    |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp  | 853 +++++++++---------
 .../gpu/grid/gridwise_gemm_xdlops_v3r1.hpp    |   0
 .../gpu/grid/gridwise_gemm_xdlops_v3r2.hpp    |   0
 .../gpu/grid/gridwise_gemm_xdlops_v3r3.hpp    |   0
 .../gpu/grid/gridwise_permute.hpp             |   0
 .../gpu/grid/gridwise_put_element_1d.hpp      |   0
 .../gpu/grid/gridwise_set_buffer_value.hpp    |   0
 .../gridwise_set_multiple_buffer_value.hpp    |   0
 .../gpu/grid/gridwise_softmax.hpp             |   0
 ...se_sparse_embeddings_forward_layernorm.hpp |   0
 .../gpu/grid/gridwise_tensor_rearrange.hpp    |   0
 .../gridwise_normalization_bwd_data.hpp       |   0
 .../gridwise_normalization_bwd_gamma_beta.hpp |   0
 .../gridwise_normalization_naive_variance.hpp |   0
 .../gridwise_normalization_selector.hpp       |   0
 .../gridwise_normalization_splitk_1st.hpp     |   0
 .../gridwise_normalization_splitk_2nd.hpp     |   0
 ...ridwise_normalization_welford_variance.hpp |   0
 .../thread/reduction_functions_threadwise.hpp |   0
 .../gpu/thread/threadwise_contraction_dl.hpp  |   0
 .../gpu/thread/threadwise_gemm_dlops_v3.hpp   |   0
 .../thread/threadwise_tensor_slice_set.hpp    |   0
 .../threadwise_tensor_slice_transfer.hpp      |   0
 .../threadwise_tensor_slice_transfer_v3r1.hpp |   0
 ...ise_tensor_slice_transfer_v3r1_dequant.hpp |   0
 .../threadwise_tensor_slice_transfer_v4r1.hpp |   0
 .../threadwise_tensor_slice_transfer_v5r1.hpp |   0
 .../threadwise_tensor_slice_transfer_v6r1.hpp |   0
 ...hreadwise_tensor_slice_transfer_v6r1r2.hpp |   0
 .../threadwise_tensor_slice_transfer_v6r2.hpp |   0
 .../threadwise_tensor_slice_transfer_v6r3.hpp |   0
 .../threadwise_tensor_slice_transfer_v7.hpp   |   0
 .../threadwise_tensor_slice_transfer_v7r2.hpp |   0
 .../gpu/thread/threadwise_welford.hpp         |   0
 .../ck/tensor_operation/gpu/warp/dpp_gemm.hpp |   0
 .../tensor_operation/gpu/warp/wmma_gemm.hpp   |   0
 .../tensor_operation/gpu/warp/xdlops_gemm.hpp |   0
 .../transform_contraction_to_gemm.hpp         |   0
 ...ransform_contraction_to_gemm_arraybase.hpp |   0
 .../transform_conv_bwd_data_to_gemm_v1.hpp    |   0
 .../transform_conv_fwd_to_gemm.hpp            |   0
 include/ck/utility/amd_address_space.hpp      |   0
 include/ck/utility/amd_buffer_addressing.hpp  |   0
 include/ck/utility/amd_gemm_dpp.hpp           |   0
 include/ck/utility/amd_inline_asm.hpp         |   0
 include/ck/utility/amd_lds.hpp                |   0
 .../ck/utility/amd_wave_read_first_lane.hpp   |   0
 include/ck/utility/amd_wmma.hpp               |   0
 include/ck/utility/amd_xdlops.hpp             |   0
 include/ck/utility/array.hpp                  |   0
 include/ck/utility/array_multi_index.hpp      |   0
 include/ck/utility/c_style_pointer_cast.hpp   |   0
 include/ck/utility/common_header.hpp          |   0
 .../ck/utility/container_element_picker.hpp   |   0
 include/ck/utility/container_helper.hpp       |   0
 include/ck/utility/data_type.hpp              |   0
 include/ck/utility/debug.hpp                  |   0
 include/ck/utility/dynamic_buffer.hpp         |   0
 include/ck/utility/enable_if.hpp              |   0
 include/ck/utility/f8_utils.hpp               |   0
 include/ck/utility/functional.hpp             |   0
 include/ck/utility/functional2.hpp            |   0
 include/ck/utility/functional3.hpp            |   0
 include/ck/utility/functional4.hpp            |   0
 .../utility/generic_memory_space_atomic.hpp   |   0
 include/ck/utility/get_id.hpp                 |   0
 include/ck/utility/get_shift.hpp              |   0
 include/ck/utility/ignore.hpp                 |   0
 include/ck/utility/inner_product.hpp          |   0
 include/ck/utility/inner_product_dpp8.hpp     |   0
 include/ck/utility/integral_constant.hpp      |   0
 include/ck/utility/is_detected.hpp            |   0
 .../ck/utility/is_known_at_compile_time.hpp   |   0
 include/ck/utility/loop_scheduler.hpp         |   0
 include/ck/utility/magic_division.hpp         |   0
 include/ck/utility/math.hpp                   |   0
 include/ck/utility/math_v2.hpp                |   0
 include/ck/utility/multi_index.hpp            |   0
 include/ck/utility/number.hpp                 |   0
 include/ck/utility/random_gen.hpp             |   0
 include/ck/utility/reduction_common.hpp       |   0
 include/ck/utility/reduction_enums.hpp        |   0
 .../reduction_functions_accumulate.hpp        |   0
 include/ck/utility/reduction_operator.hpp     |   0
 include/ck/utility/sequence.hpp               |   0
 include/ck/utility/sequence_helper.hpp        |   0
 include/ck/utility/span.hpp                   |   0
 include/ck/utility/static_buffer.hpp          |   0
 .../ck/utility/statically_indexed_array.hpp   |   0
 .../statically_indexed_array_multi_index.hpp  |   0
 include/ck/utility/synchronization.hpp        |   0
 include/ck/utility/thread_group.hpp           |   0
 include/ck/utility/transpose_vectors.hpp      |   0
 include/ck/utility/tuple.hpp                  |   0
 include/ck/utility/tuple_helper.hpp           |   0
 include/ck/utility/type.hpp                   |   0
 include/ck/utility/type_convert.hpp           |   0
 include/ck/utility/workgroup_barrier.hpp      |   0
 .../ck/utility/workgroup_synchronization.hpp  |   0
 include/ck/version.h.in                       |   0
 include/ck/wrapper/layout.hpp                 |   0
 include/ck/wrapper/operations/copy.hpp        |   0
 include/ck/wrapper/operations/gemm.hpp        |   0
 include/ck/wrapper/tensor.hpp                 |   0
 .../traits/blockwise_gemm_xdl_traits.hpp      |   0
 include/ck/wrapper/utils/kernel_utils.hpp     |   0
 include/ck/wrapper/utils/layout_utils.hpp     |   0
 include/ck/wrapper/utils/tensor_partition.hpp |   0
 include/ck/wrapper/utils/tensor_utils.hpp     |   0
 library/CMakeLists.txt                        |   0
 .../cpu/reference_avgpool_bwd.hpp             |   0
 .../cpu/reference_batched_gemm.hpp            |   0
 .../cpu/reference_batchnorm_backward.hpp      |   0
 .../cpu/reference_batchnorm_forward.hpp       |   0
 .../cpu/reference_batchnorm_infer.hpp         |   0
 .../cpu/reference_cgemm.hpp                   |   0
 .../cpu/reference_column_to_image.hpp         |   0
 .../cpu/reference_contraction.hpp             |   0
 .../cpu/reference_conv_bwd_data.hpp           |   0
 .../cpu/reference_conv_bwd_weight.hpp         |   0
 .../cpu/reference_conv_fwd.hpp                |   0
 .../reference_conv_fwd_bias_activation.hpp    |   0
 ...reference_conv_fwd_bias_activation_add.hpp |   0
 .../cpu/reference_fpAintB_gemm.hpp            |   0
 .../cpu/reference_gemm.hpp                    |   0
 .../cpu/reference_gemm_layernorm.hpp          |   0
 .../cpu/reference_groupnorm.hpp               |   0
 .../cpu/reference_groupnorm_bwd.hpp           |   0
 .../cpu/reference_image_to_column.hpp         |   0
 .../cpu/reference_layernorm.hpp               |   0
 .../cpu/reference_layernorm_bwd.hpp           |   0
 .../cpu/reference_maxpool_bwd.hpp             |   0
 .../cpu/reference_pool_fwd.hpp                |   0
 .../cpu/reference_reduce.hpp                  |   0
 .../cpu/reference_softmax.hpp                 |   0
 ...ce_sparse_embedding3_forward_layernorm.hpp |   0
 .../gpu/naive_conv_fwd.hpp                    |   0
 .../add_device_operation_instance.hpp         |   0
 .../device_operation_instance_factory.hpp     |   0
 .../gpu/avg_pool3d_bwd.hpp                    |   0
 .../gpu/batched_gemm.hpp                      |   0
 .../gpu/batched_gemm_add_relu_gemm_add.hpp    |   0
 .../gpu/batched_gemm_bias_permute.hpp         |   0
 ...batched_gemm_bias_softmax_gemm_permute.hpp |   0
 .../gpu/batched_gemm_gemm.hpp                 |   0
 .../gpu/batched_gemm_multi_d.hpp              |   0
 .../gpu/batched_gemm_softmax_gemm.hpp         |   0
 .../gpu/batched_gemm_softmax_gemm_permute.hpp |   0
 .../gpu/batchnorm_backward.hpp                |   0
 .../gpu/batchnorm_forward.hpp                 |   0
 .../gpu/batchnorm_infer.hpp                   |   0
 .../device_contraction_instance.hpp           |   0
 .../gpu/contraction_bilinear.hpp              |   0
 .../gpu/contraction_scale.hpp                 |   0
 .../gpu/conv_tensor_rearrange.hpp             |   0
 .../device_column_to_image_instance.hpp       |   0
 .../device_image_to_column_instance.hpp       |   0
 .../gpu/convolution_backward_data.hpp         |   0
 .../gpu/convolution_forward.hpp               |   0
 .../gpu/device_elementwise_instance.hpp       |   0
 .../device_gemm_mean_squaremean_instance.hpp  |   0
 ...uffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp |   0
 ...fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp |   0
 ...uffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp |   0
 .../gpu/elementwise_normalization.hpp         |   0
 .../tensor_operation_instance/gpu/gemm.hpp    |   0
 .../gpu/gemm_add.hpp                          |   0
 .../gpu/gemm_add_add_fastgelu.hpp             |   0
 .../gpu/gemm_add_fastgelu.hpp                 |   0
 .../gpu/gemm_add_multiply.hpp                 |   0
 .../gpu/gemm_add_relu.hpp                     |   0
 .../gpu/gemm_add_relu_add_layernorm.hpp       |   0
 .../gpu/gemm_add_silu.hpp                     |   0
 .../gpu/gemm_bilinear.hpp                     |   0
 .../gpu/gemm_fastgelu.hpp                     |   0
 .../gpu/gemm_multiply_add.hpp                 |   0
 .../gpu/gemm_splitk.hpp                       |   0
 .../gpu/gemm_streamk.hpp                      |   0
 ...rouped_conv_bwd_data_wmma_f16_instance.hpp |   0
 ...grouped_conv_bwd_data_wmma_i8_instance.hpp |   0
 ...ed_conv_bwd_data_xdl_bilinear_instance.hpp |   0
 ...ice_grouped_conv_bwd_data_xdl_instance.hpp |   0
 ...ouped_conv_bwd_data_xdl_scale_instance.hpp |   0
 ...ce_grouped_conv_bwd_weight_dl_instance.hpp |   0
 ..._grouped_conv_bwd_weight_wmma_instance.hpp |   0
 ...e_grouped_conv_bwd_weight_xdl_instance.hpp |   0
 .../device_grouped_conv_fwd_dl_instance.hpp   |   0
 .../device_grouped_conv_fwd_wmma_instance.hpp |   0
 ...grouped_conv_fwd_xdl_bilinear_instance.hpp |   0
 .../device_grouped_conv_fwd_xdl_instance.hpp  |   0
 ...ce_grouped_conv_fwd_xdl_scale_instance.hpp |   0
 ...uped_conv_fwd_xdl_scaleadd_ab_instance.hpp |   0
 ...wd_xdl_scaleadd_scaleadd_relu_instance.hpp |   0
 .../gpu/grouped_convolution_backward_data.hpp |   0
 ...ped_convolution_backward_data_bilinear.hpp |   0
 ...rouped_convolution_backward_data_scale.hpp |   0
 .../grouped_convolution_backward_weight.hpp   |   0
 .../gpu/grouped_convolution_forward.hpp       |   0
 .../grouped_convolution_forward_bilinear.hpp  |   0
 .../gpu/grouped_convolution_forward_scale.hpp |   0
 ...rouped_convolution_forward_scaleadd_ab.hpp |   0
 ...olution_forward_scaleadd_scaleadd_relu.hpp |   0
 .../gpu/grouped_gemm.hpp                      |   0
 .../gpu/grouped_gemm_bias.hpp                 |   0
 .../gpu/grouped_gemm_fastgelu.hpp             |   0
 .../gpu/grouped_gemm_fixed_nk.hpp             |   0
 .../gpu/groupnorm_bwd_data.hpp                |   0
 .../gpu/groupnorm_bwd_gamma_beta.hpp          |   0
 .../gpu/layernorm_bwd_data.hpp                |   0
 .../gpu/layernorm_bwd_gamma_beta.hpp          |   0
 .../gpu/max_pool_bwd.hpp                      |   0
 .../gpu/normalization_fwd.hpp                 |   0
 .../gpu/normalization_fwd_swish.hpp           |   0
 .../gpu/permute_scale.hpp                     |   0
 .../device_permute_scale_instances.hpp        |   0
 .../gpu/pool3d_fwd.hpp                        |   0
 .../gpu/quantization/gemm_quantization.hpp    |   0
 ...n_bias_forward_perchannel_quantization.hpp |   0
 ...ion_bias_forward_perlayer_quantization.hpp |   0
 ...lution_forward_perchannel_quantization.hpp |   0
 ...volution_forward_perlayer_quantization.hpp |   0
 .../gpu/reduce/device_reduce_instance.hpp     |   0
 .../device_reduce_instance_blockwise.hpp      |   0
 ...uce_instance_blockwise_b16_f32_b16_add.hpp |   0
 ...ce_instance_blockwise_b16_f32_b16_amax.hpp |   0
 ...uce_instance_blockwise_b16_f32_b16_avg.hpp |   0
 ...uce_instance_blockwise_b16_f32_b16_max.hpp |   0
 ...uce_instance_blockwise_b16_f32_b16_min.hpp |   0
 ...e_instance_blockwise_b16_f32_b16_norm2.hpp |   0
 ...ce_instance_blockwise_f16_f16_f16_amax.hpp |   0
 ...uce_instance_blockwise_f16_f16_f16_max.hpp |   0
 ...uce_instance_blockwise_f16_f16_f16_min.hpp |   0
 ...uce_instance_blockwise_f16_f32_f16_add.hpp |   0
 ...uce_instance_blockwise_f16_f32_f16_avg.hpp |   0
 ...e_instance_blockwise_f16_f32_f16_norm2.hpp |   0
 ...uce_instance_blockwise_f32_f32_f32_add.hpp |   0
 ...ce_instance_blockwise_f32_f32_f32_amax.hpp |   0
 ...uce_instance_blockwise_f32_f32_f32_avg.hpp |   0
 ...uce_instance_blockwise_f32_f32_f32_max.hpp |   0
 ...uce_instance_blockwise_f32_f32_f32_min.hpp |   0
 ...e_instance_blockwise_f32_f32_f32_norm2.hpp |   0
 ...uce_instance_blockwise_f32_f64_f32_add.hpp |   0
 ...uce_instance_blockwise_f32_f64_f32_avg.hpp |   0
 ...e_instance_blockwise_f32_f64_f32_norm2.hpp |   0
 ...uce_instance_blockwise_f64_f64_f64_add.hpp |   0
 ...ce_instance_blockwise_f64_f64_f64_amax.hpp |   0
 ...uce_instance_blockwise_f64_f64_f64_avg.hpp |   0
 ...uce_instance_blockwise_f64_f64_f64_max.hpp |   0
 ...uce_instance_blockwise_f64_f64_f64_min.hpp |   0
 ...e_instance_blockwise_f64_f64_f64_norm2.hpp |   0
 ...educe_instance_blockwise_i8_i32_i8_add.hpp |   0
 ...educe_instance_blockwise_i8_i32_i8_avg.hpp |   0
 ...educe_instance_blockwise_i8_i8_i8_amax.hpp |   0
 ...reduce_instance_blockwise_i8_i8_i8_max.hpp |   0
 ...reduce_instance_blockwise_i8_i8_i8_min.hpp |   0
 .../device_reduce_instance_impl_common.hpp    |   0
 ..._reduce_instance_multiblock_atomic_add.hpp |   0
 ..._multiblock_atomic_add_b16_f32_f32_add.hpp |   0
 ..._multiblock_atomic_add_b16_f32_f32_avg.hpp |   0
 ..._multiblock_atomic_add_f16_f32_f32_add.hpp |   0
 ..._multiblock_atomic_add_f16_f32_f32_avg.hpp |   0
 ..._multiblock_atomic_add_f32_f32_f32_add.hpp |   0
 ..._multiblock_atomic_add_f32_f32_f32_avg.hpp |   0
 ..._multiblock_atomic_add_f32_f64_f32_add.hpp |   0
 ..._multiblock_atomic_add_f32_f64_f32_avg.hpp |   0
 ..._multiblock_atomic_add_f64_f64_f64_add.hpp |   0
 ..._multiblock_atomic_add_f64_f64_f64_avg.hpp |   0
 .../device_reduce_instance_threadwise.hpp     |   0
 ...ce_instance_threadwise_b16_f32_b16_add.hpp |   0
 ...e_instance_threadwise_b16_f32_b16_amax.hpp |   0
 ...ce_instance_threadwise_b16_f32_b16_avg.hpp |   0
 ...ce_instance_threadwise_b16_f32_b16_max.hpp |   0
 ...ce_instance_threadwise_b16_f32_b16_min.hpp |   0
 ..._instance_threadwise_b16_f32_b16_norm2.hpp |   0
 ...e_instance_threadwise_f16_f16_f16_amax.hpp |   0
 ...ce_instance_threadwise_f16_f16_f16_max.hpp |   0
 ...ce_instance_threadwise_f16_f16_f16_min.hpp |   0
 ...ce_instance_threadwise_f16_f32_f16_add.hpp |   0
 ...ce_instance_threadwise_f16_f32_f16_avg.hpp |   0
 ..._instance_threadwise_f16_f32_f16_norm2.hpp |   0
 ...ce_instance_threadwise_f32_f32_f32_add.hpp |   0
 ...e_instance_threadwise_f32_f32_f32_amax.hpp |   0
 ...ce_instance_threadwise_f32_f32_f32_avg.hpp |   0
 ...ce_instance_threadwise_f32_f32_f32_max.hpp |   0
 ...ce_instance_threadwise_f32_f32_f32_min.hpp |   0
 ..._instance_threadwise_f32_f32_f32_norm2.hpp |   0
 ...ce_instance_threadwise_f32_f64_f32_add.hpp |   0
 ...ce_instance_threadwise_f32_f64_f32_avg.hpp |   0
 ..._instance_threadwise_f32_f64_f32_norm2.hpp |   0
 ...ce_instance_threadwise_f64_f64_f64_add.hpp |   0
 ...e_instance_threadwise_f64_f64_f64_amax.hpp |   0
 ...ce_instance_threadwise_f64_f64_f64_avg.hpp |   0
 ...ce_instance_threadwise_f64_f64_f64_max.hpp |   0
 ...ce_instance_threadwise_f64_f64_f64_min.hpp |   0
 ..._instance_threadwise_f64_f64_f64_norm2.hpp |   0
 ...duce_instance_threadwise_i8_i32_i8_add.hpp |   0
 ...duce_instance_threadwise_i8_i32_i8_avg.hpp |   0
 ...duce_instance_threadwise_i8_i8_i8_amax.hpp |   0
 ...educe_instance_threadwise_i8_i8_i8_max.hpp |   0
 ...educe_instance_threadwise_i8_i8_i8_min.hpp |   0
 .../gpu/reduce/reduce.hpp                     |   0
 .../tensor_operation_instance/gpu/softmax.hpp |   0
 ...softmax_f16_f16_instance_rank3_reduce1.hpp |   0
 ...softmax_f16_f16_instance_rank3_reduce2.hpp |   0
 ...softmax_f16_f16_instance_rank3_reduce3.hpp |   0
 ...softmax_f16_f16_instance_rank4_reduce1.hpp |   0
 ...softmax_f16_f16_instance_rank4_reduce2.hpp |   0
 ...softmax_f16_f16_instance_rank4_reduce3.hpp |   0
 ...softmax_f16_f16_instance_rank4_reduce4.hpp |   0
 .../device_softmax_f16_f16_instance_type.hpp  |   0
 ...softmax_f32_f32_instance_rank3_reduce1.hpp |   0
 ...softmax_f32_f32_instance_rank3_reduce2.hpp |   0
 ...softmax_f32_f32_instance_rank3_reduce3.hpp |   0
 ...softmax_f32_f32_instance_rank4_reduce1.hpp |   0
 ...softmax_f32_f32_instance_rank4_reduce2.hpp |   0
 ...softmax_f32_f32_instance_rank4_reduce3.hpp |   0
 ...softmax_f32_f32_instance_rank4_reduce4.hpp |   0
 .../device_softmax_f32_f32_instance_type.hpp  |   0
 .../gpu/softmax/device_softmax_instance.hpp   |   0
 .../transpose/device_transpose_instance.hpp   |   0
 .../gpu/transpose_3d.hpp                      |   0
 .../include/ck/library/utility/algorithm.hpp  |   0
 .../include/ck/library/utility/check_err.hpp  |   0
 .../ck/library/utility/conv_common.hpp        |   0
 ...volution_host_tensor_descriptor_helper.hpp |   0
 .../library/utility/convolution_parameter.hpp |   0
 .../ck/library/utility/device_memory.hpp      |   0
 library/include/ck/library/utility/fill.hpp   |   0
 .../ck/library/utility/host_common_util.hpp   |   0
 .../include/ck/library/utility/host_gemm.hpp  |   0
 .../ck/library/utility/host_tensor.hpp        |   0
 .../library/utility/host_tensor_generator.hpp |  39 +
 .../include/ck/library/utility/iterator.hpp   |   0
 .../include/ck/library/utility/literals.hpp   |   0
 .../include/ck/library/utility/numeric.hpp    |   0
 library/include/ck/library/utility/ranges.hpp |   0
 .../gpu/CMakeLists.txt                        |   0
 .../gpu/avg_pool3d_bwd/CMakeLists.txt         |   0
 .../avg_pool3d_bwd_ndhwc_instance_common.hpp  |   0
 ...ice_avg_pool3d_bwd_ndhwc_bf16_instance.cpp |   0
 ...vice_avg_pool3d_bwd_ndhwc_f16_instance.cpp |   0
 ...vice_avg_pool3d_bwd_ndhwc_f32_instance.cpp |   0
 .../gpu/batched_gemm/CMakeLists.txt           |   0
 ...dl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp |   0
 ...dl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp |   0
 ...dl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp |   0
 ...dl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp |   0
 ...m_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp |   0
 ...m_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp |   0
 ...m_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp |   0
 ...m_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp |   0
 ...m_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp |   0
 ...m_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp |   0
 ...m_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp |   0
 ...m_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp |   0
 ...dl_int8_int8_int8_gkm_gkn_gmn_instance.cpp |   0
 ...dl_int8_int8_int8_gkm_gnk_gmn_instance.cpp |   0
 ...dl_int8_int8_int8_gmk_gkn_gmn_instance.cpp |   0
 ...dl_int8_int8_int8_gmk_gnk_gmn_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...6_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   0
 .../batched_gemm_bias_permute/CMakeLists.txt  |   0
 ...xdl_c_shuffle_f16_f16_f16_f16_instance.cpp |   0
 .../gpu/batched_gemm_gemm/CMakeLists.txt      |   0
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...6_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   0
 .../gpu/batched_gemm_multi_d/CMakeLists.txt   |   0
 ..._d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp |   0
 ...f16_f16_gkm_gkn_gmn_irregular_instance.cpp |   0
 ..._d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp |   0
 ...f16_f16_gkm_gnk_gmn_irregular_instance.cpp |   0
 ..._d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp |   0
 ...f16_f16_gmk_gkn_gmn_irregular_instance.cpp |   0
 ..._d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp |   0
 ...f16_f16_gmk_gnk_gmn_irregular_instance.cpp |   0
 ...lti_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp |   0
 ...8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp |   0
 ...lti_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp |   0
 ...8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp |   0
 ...lti_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp |   0
 ...8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp |   0
 ...lti_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp |   0
 ...8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp |   0
 .../gpu/batched_gemm_reduce/CMakeLists.txt    |   0
 ...6_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp |   0
 ...6_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp |   0
 ...6_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp |   0
 ...6_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp |   0
 .../batched_gemm_softmax_gemm/CMakeLists.txt  |   0
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...f16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...f16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...6_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 .../gpu/batchnorm/CMakeLists.txt              |   0
 ...evice_batchnorm_backward_bf16_instance.cpp |   0
 ...device_batchnorm_backward_f16_instance.cpp |   0
 ...device_batchnorm_backward_f32_instance.cpp |   0
 ...device_batchnorm_backward_f64_instance.cpp |   0
 ...device_batchnorm_forward_bf16_instance.cpp |   0
 .../device_batchnorm_forward_f16_instance.cpp |   0
 .../device_batchnorm_forward_f32_instance.cpp |   0
 .../device_batchnorm_forward_f64_instance.cpp |   0
 .../device_batchnorm_infer_bf16_instance.cpp  |   0
 .../device_batchnorm_infer_f16_instance.cpp   |   0
 .../device_batchnorm_infer_f32_instance.cpp   |   0
 .../device_batchnorm_infer_f64_instance.cpp   |   0
 .../gpu/column_to_image/CMakeLists.txt        |   0
 ...ice_column_to_image_gndhwc_3d_instance.cpp |   0
 ...vice_column_to_image_gnhwc_2d_instance.cpp |   0
 ...evice_column_to_image_gnwc_1d_instance.cpp |   0
 ...ice_column_to_image_ndhwgc_3d_instance.cpp |   0
 ...vice_column_to_image_nhwgc_2d_instance.cpp |   0
 ...evice_column_to_image_nwgc_1d_instance.cpp |   0
 .../gpu/contraction_bilinear/CMakeLists.txt   |   0
 ...16_bf16_bf16_compute_f32_kknn_instance.cpp |   0
 ...16_bf16_bf16_compute_f32_knnn_instance.cpp |   0
 ...16_bf16_bf16_compute_f32_mknn_instance.cpp |   0
 ...16_bf16_bf16_compute_f32_mnnn_instance.cpp |   0
 ..._f16_f16_f16_compute_f32_kknn_instance.cpp |   0
 ..._f16_f16_f16_compute_f32_knnn_instance.cpp |   0
 ..._f16_f16_f16_compute_f32_mknn_instance.cpp |   0
 ..._f16_f16_f16_compute_f32_mnnn_instance.cpp |   0
 ...f32_f32_f32_compute_bf16_kknn_instance.cpp |   0
 ...f32_f32_f32_compute_bf16_knnn_instance.cpp |   0
 ...f32_f32_f32_compute_bf16_mknn_instance.cpp |   0
 ...f32_f32_f32_compute_bf16_mnnn_instance.cpp |   0
 ..._f32_f32_f32_compute_f16_kknn_instance.cpp |   0
 ..._f32_f32_f32_compute_f16_knnn_instance.cpp |   0
 ..._f32_f32_f32_compute_f16_mknn_instance.cpp |   0
 ..._f32_f32_f32_compute_f16_mnnn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_f32_kknn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_f32_knnn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_f32_mknn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_f32_mnnn_instance.cpp |   0
 ..._f64_f64_f64_compute_f32_kknn_instance.cpp |   0
 ..._f64_f64_f64_compute_f32_knnn_instance.cpp |   0
 ..._f64_f64_f64_compute_f32_mknn_instance.cpp |   0
 ..._f64_f64_f64_compute_f32_mnnn_instance.cpp |   0
 ..._shuffle_f64_f64_f64_f64_kknn_instance.cpp |   0
 ..._shuffle_f64_f64_f64_f64_knnn_instance.cpp |   0
 ..._shuffle_f64_f64_f64_f64_mknn_instance.cpp |   0
 ..._shuffle_f64_f64_f64_f64_mnnn_instance.cpp |   0
 .../gpu/contraction_scale/CMakeLists.txt      |   0
 ...f16_bf16_bf16_compute_f32_kkn_instance.cpp |   0
 ...f16_bf16_bf16_compute_f32_knn_instance.cpp |   0
 ...f16_bf16_bf16_compute_f32_mkn_instance.cpp |   0
 ...f16_bf16_bf16_compute_f32_mnn_instance.cpp |   0
 ...e_f16_f16_f16_compute_f32_kkn_instance.cpp |   0
 ...e_f16_f16_f16_compute_f32_knn_instance.cpp |   0
 ...e_f16_f16_f16_compute_f32_mkn_instance.cpp |   0
 ...e_f16_f16_f16_compute_f32_mnn_instance.cpp |   0
 ..._f32_f32_f32_compute_bf16_kkn_instance.cpp |   0
 ..._f32_f32_f32_compute_bf16_knn_instance.cpp |   0
 ..._f32_f32_f32_compute_bf16_mkn_instance.cpp |   0
 ..._f32_f32_f32_compute_bf16_mnn_instance.cpp |   0
 ...e_f32_f32_f32_compute_f16_kkn_instance.cpp |   0
 ...e_f32_f32_f32_compute_f16_knn_instance.cpp |   0
 ...e_f32_f32_f32_compute_f16_mkn_instance.cpp |   0
 ...e_f32_f32_f32_compute_f16_mnn_instance.cpp |   0
 ...xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp |   0
 ...xdl_c_shuffle_f32_f32_f32_knn_instance.cpp |   0
 ...xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp |   0
 ...xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp |   0
 ...e_f64_f64_f64_compute_f32_kkn_instance.cpp |   0
 ...e_f64_f64_f64_compute_f32_knn_instance.cpp |   0
 ...e_f64_f64_f64_compute_f32_mkn_instance.cpp |   0
 ...e_f64_f64_f64_compute_f32_mnn_instance.cpp |   0
 ...xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp |   0
 ...xdl_c_shuffle_f64_f64_f64_knn_instance.cpp |   0
 ...xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp |   0
 ...xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp |   0
 .../gpu/conv1d_bwd_data/CMakeLists.txt        |   0
 ...bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp |   0
 ..._bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp |   0
 ..._bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp |   0
 ...bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp |   0
 .../gpu/conv2d_bwd_data/CMakeLists.txt        |   0
 ...wd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...wd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ...d_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 ..._data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |   0
 ...d_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...d_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ..._data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 .../gpu/conv2d_fwd/CMakeLists.txt             |   0
 ..._c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |   0
 ...2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ...d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 .../gpu/conv2d_fwd_bias_relu/CMakeLists.txt   |   0
 ..._bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 .../conv2d_fwd_bias_relu_add/CMakeLists.txt   |   0
 ...s_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 .../gpu/conv3d_bwd_data/CMakeLists.txt        |   0
 ...ta_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp |   0
 ...ata_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp |   0
 ...ata_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp |   0
 ...ta_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp |   0
 .../gpu/elementwise/CMakeLists.txt            |   0
 .../elementwise/device_normalize_instance.cpp |   0
 .../elementwise_normalization/CMakeLists.txt  |   0
 ...elementwise_normalization_f16_instance.cpp |   0
 .../gpu/gemm/CMakeLists.txt                   |   0
 ..._gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...16_f16_f16_km_kn_mn_irregular_instance.cpp |   0
 ..._gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...16_f16_f16_km_nk_mn_irregular_instance.cpp |   0
 ..._gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ..._gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 ..._gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ..._gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ..._gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ..._gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ...ice_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...l_i8_i8_i8_km_kn_mn_irregular_instance.cpp |   0
 ...ice_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...l_i8_i8_i8_km_nk_mn_irregular_instance.cpp |   0
 ...ice_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...l_i8_i8_i8_mk_kn_mn_irregular_instance.cpp |   0
 ...ice_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 ...l_i8_i8_i8_mk_nk_mn_irregular_instance.cpp |   0
 ...gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...16_f16_f16_km_kn_mn_irregular_instance.cpp |   0
 ...gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...16_f16_f16_km_nk_mn_irregular_instance.cpp |   0
 ...gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 ...emm_wmma_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...emm_wmma_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...emm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...emm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ..._2_stage_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...uffle_bf16_bf16_bf16_km_kn_mn_instance.cpp |   0
 ...uffle_bf16_bf16_bf16_km_nk_mn_instance.cpp |   0
 ...uffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp |   0
 ...uffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...c_shuffle_f16_f8_f16_mk_kn_mn_instance.cpp |   0
 ...c_shuffle_f16_f8_f16_mk_nk_mn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ..._shuffle_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ..._shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp |   0
 ..._shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp |   0
 ...8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp |   0
 ...mk_kn_mn_v1_interwave_default_instance.cpp |   0
 ..._mk_kn_mn_v1_interwave_padded_instance.cpp |   0
 ...p8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp |   0
 ...8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp |   0
 ...p8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp |   0
 ..._shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 ...ect_load_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...ect_load_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...ect_load_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...ect_load_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...ect_load_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 .../device_gemm_xdl_f16_f16_f16/common.hpp    |   0
 .../km_kn_mn_add_instance.cpp                 |   0
 .../km_kn_mn_default_pipeline_v1_instance.cpp |   0
 .../km_kn_mn_default_pipeline_v2_instance.cpp |   0
 ...kn_mn_default_pipeline_v2_opt_instance.cpp |   0
 ...m_kn_mn_interwave_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v2_instance.cpp |   0
 ...regular_interwave_pipeline_v1_instance.cpp |   0
 .../km_nk_mn_add_instance.cpp                 |   0
 .../km_nk_mn_default_pipeline_v1_instance.cpp |   0
 .../km_nk_mn_default_pipeline_v2_instance.cpp |   0
 ...nk_mn_default_pipeline_v2_opt_instance.cpp |   0
 ...m_nk_mn_interwave_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v2_instance.cpp |   0
 ...regular_interwave_pipeline_v1_instance.cpp |   0
 .../mk_kn_mn_add_instance.cpp                 |   0
 .../mk_kn_mn_default_pipeline_v1_instance.cpp |   0
 .../mk_kn_mn_default_pipeline_v2_instance.cpp |   0
 ...kn_mn_default_pipeline_v2_opt_instance.cpp |   0
 ...k_kn_mn_interwave_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v2_instance.cpp |   0
 ...regular_interwave_pipeline_v1_instance.cpp |   0
 .../mk_nk_mn_add_instance.cpp                 |   0
 .../mk_nk_mn_default_pipeline_v1_instance.cpp |   0
 .../mk_nk_mn_default_pipeline_v2_instance.cpp |   0
 ...nk_mn_default_pipeline_v2_opt_instance.cpp |   0
 ...k_nk_mn_interwave_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v1_instance.cpp |   0
 ...irregular_default_pipeline_v2_instance.cpp |   0
 ...regular_interwave_pipeline_v1_instance.cpp |   0
 ...gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ...gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp |   0
 ...gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp |   0
 ...gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp |   0
 ...gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_add/CMakeLists.txt               |   0
 ...bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp |   0
 ...le_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_add_fastgelu/CMakeLists.txt  |   0
 ...16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_fastgelu/CMakeLists.txt      |   0
 ...bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |   0
 ...le_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_multiply/CMakeLists.txt      |   0
 ...16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_relu/CMakeLists.txt          |   0
 ...bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp |   0
 ...le_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ..._layernorm_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ..._layernorm_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ..._layernorm_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ..._layernorm_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_silu/CMakeLists.txt          |   0
 ...bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp |   0
 ...le_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 .../gpu/gemm_bias_add_reduce/CMakeLists.txt   |   0
 ..._f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_bilinear/CMakeLists.txt          |   0
 ...uffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp |   0
 ...uffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp |   0
 ...uffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp |   0
 ...uffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |   0
 .../gpu/gemm_fastgelu/CMakeLists.txt          |   0
 ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_multiply_add/CMakeLists.txt      |   0
 ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 ...f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_reduce/CMakeLists.txt            |   0
 ..._f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   0
 ..._f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_splitk/CMakeLists.txt            |   0
 ...f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp |   0
 ...f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp |   0
 ...f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp |   0
 ...f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp |   0
 ...l_splitk_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...l_splitk_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...plitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp |   0
 ...f16_f16_mk_kn_mn_v1_interwave_instance.cpp |   0
 ..._kn_mn_v1_interwave_irregular_instance.cpp |   0
 ...f16_f16_mk_kn_mn_v1_irregular_instance.cpp |   0
 ...plitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp |   0
 ...f16_f16_mk_kn_mn_v2_irregular_instance.cpp |   0
 ...plitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp |   0
 ...f16_f16_mk_nk_mn_v1_interwave_instance.cpp |   0
 ..._nk_mn_v1_interwave_irregular_instance.cpp |   0
 ...f16_f16_mk_nk_mn_v1_irregular_instance.cpp |   0
 ...plitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp |   0
 ...f16_f16_mk_nk_mn_v2_irregular_instance.cpp |   0
 ...l_splitk_f16_fp8_f16_km_kn_mn_instance.cpp |   0
 ...l_splitk_f16_fp8_f16_km_nk_mn_instance.cpp |   0
 ...16_fp8_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...plitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp |   0
 ...fp8_f16_mk_kn_mn_v1_interwave_instance.cpp |   0
 ...plitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp |   0
 ...k_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp |   0
 ...plitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp |   0
 ...fp8_f16_mk_nk_mn_v1_interwave_instance.cpp |   0
 ...plitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp |   0
 ...l_splitk_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...l_splitk_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...l_splitk_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...l_splitk_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ...l_splitk_fp8_f16_f16_km_kn_mn_instance.cpp |   0
 ...l_splitk_fp8_f16_f16_km_nk_mn_instance.cpp |   0
 ...plitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp |   0
 ...f16_f16_mk_kn_mn_v1_interwave_instance.cpp |   0
 ...plitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp |   0
 ...l_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp |   0
 ...ect_load_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_streamk/CMakeLists.txt           |   0
 ..._streamk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 .../grouped_conv1d_bwd_weight/CMakeLists.txt  |   0
 ...weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ..._weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ..._weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 ...weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp |   0
 ..._weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp |   0
 ..._weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp |   0
 ...eight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ...weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ...weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 .../gpu/grouped_conv1d_fwd/CMakeLists.txt     |   0
 ...d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ...1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ...1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 ...d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp |   0
 .../grouped_conv2d_bwd_data/CMakeLists.txt    |   0
 ...gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp |   0
 ...ta_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ..._gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp |   0
 ...ata_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp |   0
 ...nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp |   0
 ...ta_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ..._nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp |   0
 ...ata_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp |   0
 ...ta_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...ata_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...ata_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...ta_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...ata_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...ata_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../grouped_conv2d_bwd_weight/CMakeLists.txt  |   0
 ...ght_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...ight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...ight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...ght_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...ight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...ight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 ...ht_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...ght_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...ght_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...ht_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...ght_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...ght_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv2d_fwd/CMakeLists.txt     |   0
 ..._fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ..._fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ..._fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ..._fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 ...a_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp |   0
 ...gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp |   0
 ...wd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...ma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp |   0
 ...ma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp |   0
 ..._gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp |   0
 ...fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp |   0
 ...mma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp |   0
 ...a_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp |   0
 ...nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp |   0
 ...wd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...ma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp |   0
 ...ma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp |   0
 ..._nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp |   0
 ...fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp |   0
 ...mma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp |   0
 ...wd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...wd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../grouped_conv3d_bwd_data/CMakeLists.txt    |   0
 ...hwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...dhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ...wgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...hwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ...xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ..._ndhwgk_input_f16_comp_bf8_f8_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...ear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...near_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...near_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...ale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...cale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...cale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 .../grouped_conv3d_bwd_weight/CMakeLists.txt  |   0
 ..._dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ...t_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...t_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ..._dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...t_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...t_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...hwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...dhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ...wgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...hwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ...xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...kzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv3d_fwd/CMakeLists.txt     |   0
 ...ndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp |   0
 ...hwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp |   0
 ...gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp |   0
 ...dhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ..._gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp |   0
 ...dhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp |   0
 ...wgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp |   0
 ...ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp |   0
 ...hwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ..._wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ..._ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp |   0
 ...xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ..._xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp |   0
 ...xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...gc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ..._xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...ear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...near_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...near_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...ear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../grouped_conv3d_fwd_scale/CMakeLists.txt   |   0
 ...ale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...cale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...cale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...ale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ..._ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...d_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...d_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ..._ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../CMakeLists.txt                            |   0
 ...elu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...elu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../gpu/grouped_gemm/CMakeLists.txt           |   0
 ...gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...l_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...l_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 ...f16_f8_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...f8_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 .../gpu/grouped_gemm_bias/CMakeLists.txt      |   0
 ..._nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ..._nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp |   0
 ..._nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/grouped_gemm_fastgelu/CMakeLists.txt  |   0
 ...gelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...gelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...gelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...gelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/grouped_gemm_fixed_nk/CMakeLists.txt  |   0
 ...ixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp |   0
 ...ixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp |   0
 ...fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp |   0
 ...fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp |   0
 ..._fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp |   0
 ..._fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/image_to_column/CMakeLists.txt        |   0
 ...ice_image_to_column_gndhwc_3d_instance.cpp |   0
 ...vice_image_to_column_gnhwc_2d_instance.cpp |   0
 ...evice_image_to_column_gnwc_1d_instance.cpp |   0
 ...ice_image_to_column_ndhwgc_3d_instance.cpp |   0
 ...vice_image_to_column_nhwgc_2d_instance.cpp |   0
 ...evice_image_to_column_nwgc_1d_instance.cpp |   0
 .../gpu/max_pool_bwd/CMakeLists.txt           |   0
 .../device_max_pool_bwd_bf16_instance.cpp     |   0
 .../device_max_pool_bwd_f16_instance.cpp      |   0
 .../device_max_pool_bwd_f32_instance.cpp      |   0
 .../max_pool_bwd_instance_common.hpp          |   0
 .../gpu/normalization_bwd_data/CMakeLists.txt |   0
 ...device_groupnorm_bwd_data_f32_instance.cpp |   0
 ...vice_layernorm2d_bwd_data_f16_instance.cpp |   0
 ...vice_layernorm2d_bwd_data_f32_instance.cpp |   0
 ...normalization_bwd_data_instance_common.hpp |   0
 .../CMakeLists.txt                            |   0
 ..._groupnorm_bwd_gamma_beta_f32_instance.cpp |   0
 ...ayernorm2d_bwd_gamma_beta_f16_instance.cpp |   0
 ...ayernorm2d_bwd_gamma_beta_f32_instance.cpp |   0
 ...ization_bwd_gamma_beta_instance_common.hpp |   0
 .../gpu/normalization_fwd/CMakeLists.txt      |   0
 .../device_groupnorm_fwd_f16_instance.cpp     |   0
 .../device_groupnorm_fwd_f32_instance.cpp     |   0
 ...orm_fwd_swish_f16_f32_f32_f16_instance.cpp |   0
 ...evice_groupnorm_fwd_swish_f16_instance.cpp |   0
 ...evice_groupnorm_fwd_swish_f32_instance.cpp |   0
 .../device_layernorm2d_fwd_f16_instance.cpp   |   0
 .../device_layernorm2d_fwd_f32_instance.cpp   |   0
 .../device_layernorm4d_fwd_f16_instance.cpp   |   0
 .../device_layernorm4d_fwd_f32_instance.cpp   |   0
 .../normalization_fwd_instance_common.hpp     |   0
 .../gpu/permute_scale/CMakeLists.txt          |   0
 .../device_permute_scale_1d_instances.cpp     |   0
 .../device_permute_scale_2d_instances.cpp     |   0
 .../device_permute_scale_3d_instances.cpp     |   0
 .../device_permute_scale_4d_instances.cpp     |   0
 .../device_permute_scale_5d_instances.cpp     |   0
 .../device_permute_scale_6d_instances.cpp     |   0
 .../gpu/pool3d_fwd/CMakeLists.txt             |   0
 ...ice_avg_pool3d_fwd_ndhwc_bf16_instance.cpp |   0
 ...vice_avg_pool3d_fwd_ndhwc_f16_instance.cpp |   0
 ...vice_avg_pool3d_fwd_ndhwc_f32_instance.cpp |   0
 ...ice_max_pool3d_fwd_ndhwc_bf16_instance.cpp |   0
 ...vice_max_pool3d_fwd_ndhwc_f16_instance.cpp |   0
 ...vice_max_pool3d_fwd_ndhwc_f32_instance.cpp |   0
 .../pool3d_fwd/pool_fwd_instance_common.hpp   |   0
 .../gpu/quantization/CMakeLists.txt           |   0
 .../conv2d_fwd/conv2d_quantization_common.hpp |   0
 ..._perchannel_quantization_int8_instance.cpp |   0
 ...as_perlayer_quantization_int8_instance.cpp |   0
 .../device_conv2d_dl_int8_instance.hpp        |   0
 ..._perchannel_quantization_int8_instance.cpp |   0
 ...dl_perlayer_quantization_int8_instance.cpp |   0
 ..._perchannel_quantization_int8_instance.cpp |   0
 ...as_perlayer_quantization_int8_instance.cpp |   0
 .../device_conv2d_xdl_int8_instance.hpp       |   0
 ..._perchannel_quantization_int8_instance.cpp |   0
 ...dl_perlayer_quantization_int8_instance.cpp |   0
 ...ization_dl_c_shuffle_i8_i8_i8_instance.hpp |   0
 ...l_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 ...zation_xdl_c_shuffle_i8_i8_i8_instance.hpp |   0
 ...l_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...l_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 .../gemm/gemm_quantization_common.hpp         |   0
 .../gpu/reduce/CMakeLists.txt                 |   0
 ...uce_instance_blockwise_b16_f32_b16_add.cpp |   0
 ...ce_instance_blockwise_b16_f32_b16_amax.cpp |   0
 ...uce_instance_blockwise_b16_f32_b16_avg.cpp |   0
 ...uce_instance_blockwise_b16_f32_b16_max.cpp |   0
 ...uce_instance_blockwise_b16_f32_b16_min.cpp |   0
 ...e_instance_blockwise_b16_f32_b16_norm2.cpp |   0
 ...ce_instance_blockwise_f16_f16_f16_amax.cpp |   0
 ...uce_instance_blockwise_f16_f16_f16_max.cpp |   0
 ...uce_instance_blockwise_f16_f16_f16_min.cpp |   0
 ...uce_instance_blockwise_f16_f32_f16_add.cpp |   0
 ...uce_instance_blockwise_f16_f32_f16_avg.cpp |   0
 ...e_instance_blockwise_f16_f32_f16_norm2.cpp |   0
 ...uce_instance_blockwise_f32_f32_f32_add.cpp |   0
 ...ce_instance_blockwise_f32_f32_f32_amax.cpp |   0
 ...uce_instance_blockwise_f32_f32_f32_avg.cpp |   0
 ...uce_instance_blockwise_f32_f32_f32_max.cpp |   0
 ...uce_instance_blockwise_f32_f32_f32_min.cpp |   0
 ...e_instance_blockwise_f32_f32_f32_norm2.cpp |   0
 ...uce_instance_blockwise_f32_f64_f32_add.cpp |   0
 ...uce_instance_blockwise_f32_f64_f32_avg.cpp |   0
 ...e_instance_blockwise_f32_f64_f32_norm2.cpp |   0
 ...uce_instance_blockwise_f64_f64_f64_add.cpp |   0
 ...ce_instance_blockwise_f64_f64_f64_amax.cpp |   0
 ...uce_instance_blockwise_f64_f64_f64_avg.cpp |   0
 ...uce_instance_blockwise_f64_f64_f64_max.cpp |   0
 ...uce_instance_blockwise_f64_f64_f64_min.cpp |   0
 ...e_instance_blockwise_f64_f64_f64_norm2.cpp |   0
 ...educe_instance_blockwise_i8_i32_i8_add.cpp |   0
 ...educe_instance_blockwise_i8_i32_i8_avg.cpp |   0
 ...educe_instance_blockwise_i8_i8_i8_amax.cpp |   0
 ...reduce_instance_blockwise_i8_i8_i8_max.cpp |   0
 ...reduce_instance_blockwise_i8_i8_i8_min.cpp |   0
 ..._multiblock_atomic_add_b16_f32_f32_add.cpp |   0
 ..._multiblock_atomic_add_b16_f32_f32_avg.cpp |   0
 ..._multiblock_atomic_add_f16_f32_f32_add.cpp |   0
 ..._multiblock_atomic_add_f16_f32_f32_avg.cpp |   0
 ..._multiblock_atomic_add_f32_f32_f32_add.cpp |   0
 ..._multiblock_atomic_add_f32_f32_f32_avg.cpp |   0
 ..._multiblock_atomic_add_f32_f64_f32_add.cpp |   0
 ..._multiblock_atomic_add_f32_f64_f32_avg.cpp |   0
 ..._multiblock_atomic_add_f64_f64_f64_add.cpp |   0
 ..._multiblock_atomic_add_f64_f64_f64_avg.cpp |   0
 ...ce_instance_threadwise_b16_f32_b16_add.cpp |   0
 ...e_instance_threadwise_b16_f32_b16_amax.cpp |   0
 ...ce_instance_threadwise_b16_f32_b16_avg.cpp |   0
 ...ce_instance_threadwise_b16_f32_b16_max.cpp |   0
 ...ce_instance_threadwise_b16_f32_b16_min.cpp |   0
 ..._instance_threadwise_b16_f32_b16_norm2.cpp |   0
 ...e_instance_threadwise_f16_f16_f16_amax.cpp |   0
 ...ce_instance_threadwise_f16_f16_f16_max.cpp |   0
 ...ce_instance_threadwise_f16_f16_f16_min.cpp |   0
 ...ce_instance_threadwise_f16_f32_f16_add.cpp |   0
 ...ce_instance_threadwise_f16_f32_f16_avg.cpp |   0
 ..._instance_threadwise_f16_f32_f16_norm2.cpp |   0
 ...ce_instance_threadwise_f32_f32_f32_add.cpp |   0
 ...e_instance_threadwise_f32_f32_f32_amax.cpp |   0
 ...ce_instance_threadwise_f32_f32_f32_avg.cpp |   0
 ...ce_instance_threadwise_f32_f32_f32_max.cpp |   0
 ...ce_instance_threadwise_f32_f32_f32_min.cpp |   0
 ..._instance_threadwise_f32_f32_f32_norm2.cpp |   0
 ...ce_instance_threadwise_f32_f64_f32_add.cpp |   0
 ...ce_instance_threadwise_f32_f64_f32_avg.cpp |   0
 ..._instance_threadwise_f32_f64_f32_norm2.cpp |   0
 ...ce_instance_threadwise_f64_f64_f64_add.cpp |   0
 ...e_instance_threadwise_f64_f64_f64_amax.cpp |   0
 ...ce_instance_threadwise_f64_f64_f64_avg.cpp |   0
 ...ce_instance_threadwise_f64_f64_f64_max.cpp |   0
 ...ce_instance_threadwise_f64_f64_f64_min.cpp |   0
 ..._instance_threadwise_f64_f64_f64_norm2.cpp |   0
 ...duce_instance_threadwise_i8_i32_i8_add.cpp |   0
 ...duce_instance_threadwise_i8_i32_i8_avg.cpp |   0
 ...duce_instance_threadwise_i8_i8_i8_amax.cpp |   0
 ...educe_instance_threadwise_i8_i8_i8_max.cpp |   0
 ...educe_instance_threadwise_i8_i8_i8_min.cpp |   0
 .../gpu/softmax/CMakeLists.txt                |   0
 ...softmax_f16_f16_instance_rank3_reduce1.cpp |   0
 ...softmax_f16_f16_instance_rank3_reduce2.cpp |   0
 ...softmax_f16_f16_instance_rank3_reduce3.cpp |   0
 ...softmax_f16_f16_instance_rank4_reduce1.cpp |   0
 ...softmax_f16_f16_instance_rank4_reduce2.cpp |   0
 ...softmax_f16_f16_instance_rank4_reduce3.cpp |   0
 ...softmax_f16_f16_instance_rank4_reduce4.cpp |   0
 ...softmax_f32_f32_instance_rank3_reduce1.cpp |   0
 ...softmax_f32_f32_instance_rank3_reduce2.cpp |   0
 ...softmax_f32_f32_instance_rank3_reduce3.cpp |   0
 ...softmax_f32_f32_instance_rank4_reduce1.cpp |   0
 ...softmax_f32_f32_instance_rank4_reduce2.cpp |   0
 ...softmax_f32_f32_instance_rank4_reduce3.cpp |   0
 ...softmax_f32_f32_instance_rank4_reduce4.cpp |   0
 .../gpu/transpose/CMakeLists.txt              |   0
 .../device_transpose_instances_3d.cpp         |   0
 library/src/utility/CMakeLists.txt            |   0
 library/src/utility/convolution_parameter.cpp |   0
 library/src/utility/device_memory.cpp         |   0
 library/src/utility/host_tensor.cpp           |   0
 profiler/CMakeLists.txt                       |   0
 profiler/README.md                            |   0
 profiler/include/profiler/data_type_enum.hpp  |   0
 .../profiler/profile_avg_pool3d_bwd_impl.hpp  |   0
 ...le_batched_gemm_add_relu_gemm_add_impl.hpp |   0
 ...ed_gemm_bias_softmax_gemm_permute_impl.hpp |   0
 .../profile_batched_gemm_gemm_impl.hpp        |   0
 .../profiler/profile_batched_gemm_impl.hpp    |   0
 .../profile_batched_gemm_reduce_impl.hpp      |   0
 ...profile_batched_gemm_softmax_gemm_impl.hpp |   0
 ...batched_gemm_softmax_gemm_permute_impl.hpp |   0
 .../profile_batchnorm_backward_impl.hpp       |   0
 .../profile_batchnorm_forward_impl.hpp        |   0
 .../profiler/profile_batchnorm_infer_impl.hpp |   0
 .../profiler/profile_contraction_impl.hpp     |   0
 .../profiler/profile_contraction_utils.hpp    |   0
 .../profiler/profile_conv_bwd_data_impl.hpp   |   0
 .../profile_conv_fwd_bias_relu_add_impl.hpp   |   0
 .../profile_conv_fwd_bias_relu_impl.hpp       |   0
 .../profiler/profile_conv_fwd_impl.hpp        |   0
 .../profile_conv_tensor_rearrange_impl.hpp    |   0
 .../profile_elementwise_layernorm_impl.hpp    |   0
 .../profile_gemm_add_add_fastgelu_impl.hpp    |   0
 .../profile_gemm_add_fastgelu_impl.hpp        |   0
 .../profiler/profile_gemm_add_impl.hpp        |   0
 .../profile_gemm_add_multiply_impl.hpp        |   0
 ...ofile_gemm_add_relu_add_layernorm_impl.hpp |   0
 .../profiler/profile_gemm_add_relu_impl.hpp   |   0
 .../profiler/profile_gemm_add_silu_impl.hpp   |   0
 .../profile_gemm_bias_add_reduce_impl.hpp     |   0
 .../profiler/profile_gemm_bilinear_impl.hpp   |   0
 .../profiler/profile_gemm_fastgelu_impl.hpp   |   0
 .../include/profiler/profile_gemm_impl.hpp    |   0
 .../profile_gemm_multiply_add_impl.hpp        |   0
 .../profiler/profile_gemm_reduce_impl.hpp     |   0
 .../profiler/profile_gemm_splitk_impl.hpp     |   0
 .../profiler/profile_gemm_streamk_impl.hpp    |   0
 .../profile_grouped_conv_bwd_data_impl.hpp    |   0
 .../profile_grouped_conv_bwd_weight_impl.hpp  |   0
 .../profile_grouped_conv_fwd_impl.hpp         |   0
 .../profile_grouped_gemm_fastgelu_impl.hpp    |   0
 .../profile_grouped_gemm_fixed_nk_impl.hpp    |   0
 .../profiler/profile_grouped_gemm_impl.hpp    |   0
 .../profile_groupnorm_bwd_data_impl.hpp       |   0
 .../profile_groupnorm_bwd_gamma_beta_impl.hpp |   0
 .../profiler/profile_groupnorm_fwd_impl.hpp   |   0
 .../profile_layernorm_bwd_data_impl.hpp       |   0
 .../profile_layernorm_bwd_gamma_beta_impl.hpp |   0
 .../profiler/profile_layernorm_fwd_impl.hpp   |   0
 .../profiler/profile_max_pool3d_bwd_impl.hpp  |   0
 .../profiler/profile_permute_scale_impl.hpp   |   0
 .../profiler/profile_pool3d_fwd_impl.hpp      |   0
 .../include/profiler/profile_reduce_impl.hpp  |   0
 .../include/profiler/profile_softmax_impl.hpp |   0
 .../profiler/profile_transpose_impl.hpp       |   0
 profiler/src/CMakeLists.txt                   |   0
 profiler/src/profile_avg_pool3d_bwd.cpp       |   0
 profiler/src/profile_batched_gemm.cpp         |   0
 ...profile_batched_gemm_add_relu_gemm_add.cpp |   0
 profiler/src/profile_batched_gemm_gemm.cpp    |   0
 profiler/src/profile_batched_gemm_multi_d.cpp |   0
 profiler/src/profile_batched_gemm_reduce.cpp  |   0
 profiler/src/profile_batchnorm_bwd.cpp        |   0
 profiler/src/profile_batchnorm_fwd.cpp        |   0
 profiler/src/profile_batchnorm_infer.cpp      |   0
 profiler/src/profile_contraction_bilinear.cpp |   0
 profiler/src/profile_contraction_scale.cpp    |   0
 profiler/src/profile_conv_bwd_data.cpp        |   0
 profiler/src/profile_conv_fwd.cpp             |   0
 profiler/src/profile_conv_fwd_bias_relu.cpp   |   0
 .../src/profile_conv_fwd_bias_relu_add.cpp    |   0
 .../src/profile_conv_tensor_rearrange.cpp     |   0
 profiler/src/profile_gemm.cpp                 |   0
 profiler/src/profile_gemm_add.cpp             |   0
 .../src/profile_gemm_add_add_fastgelu.cpp     |   0
 profiler/src/profile_gemm_add_fastgelu.cpp    |   0
 profiler/src/profile_gemm_add_multiply.cpp    |   0
 profiler/src/profile_gemm_add_relu.cpp        |   0
 .../profile_gemm_add_relu_add_layernorm.cpp   |   0
 profiler/src/profile_gemm_add_silu.cpp        |   0
 profiler/src/profile_gemm_bias_add_reduce.cpp |   0
 profiler/src/profile_gemm_bilinear.cpp        |   0
 profiler/src/profile_gemm_fastgelu.cpp        |   0
 profiler/src/profile_gemm_multiply_add.cpp    |   0
 profiler/src/profile_gemm_reduce.cpp          |   0
 profiler/src/profile_gemm_splitk.cpp          |   0
 profiler/src/profile_gemm_streamk.cpp         |   0
 .../src/profile_grouped_conv_bwd_data.cpp     |   0
 .../src/profile_grouped_conv_bwd_weight.cpp   |   0
 profiler/src/profile_grouped_conv_fwd.cpp     |   0
 profiler/src/profile_grouped_gemm.cpp         |   0
 .../src/profile_grouped_gemm_fastgelu.cpp     |   0
 .../src/profile_grouped_gemm_fixed_nk.cpp     |   0
 profiler/src/profile_groupnorm_bwd_data.cpp   |   0
 .../src/profile_groupnorm_bwd_gamma_beta.cpp  |   0
 profiler/src/profile_groupnorm_fwd.cpp        |   0
 profiler/src/profile_layernorm_bwd_data.cpp   |   0
 .../src/profile_layernorm_bwd_gamma_beta.cpp  |   0
 profiler/src/profile_layernorm_fwd.cpp        |   0
 profiler/src/profile_max_pool3d_bwd.cpp       |   0
 profiler/src/profile_max_pool3d_fwd.cpp       |   0
 profiler/src/profile_permute_scale.cpp        |   0
 profiler/src/profile_reduce.cpp               |   0
 profiler/src/profile_softmax.cpp              |   0
 profiler/src/profile_transpose.cpp            |   0
 profiler/src/profiler.cpp                     |   0
 profiler/src/profiler_operation_registry.hpp  |   0
 rbuild.ini                                    |   0
 requirements.txt                              |   0
 script/cmake-ck-dev.sh                        |   2 +-
 script/hip_fatbin_insert                      |   0
 script/process_perf_data.py                   |   0
 script/redis-cli.conf                         |   0
 script/test_convnd_fwd.sh                     |   0
 test/CMakeLists.txt                           |   0
 test/batched_gemm/CMakeLists.txt              |   0
 test/batched_gemm/test_batched_gemm.cpp       |   0
 test/batched_gemm_gemm/CMakeLists.txt         |   0
 .../test_batched_gemm_gemm_fp16.cpp           |   0
 .../test_batched_gemm_gemm_util.hpp           |   0
 test/batched_gemm_multi_d/CMakeLists.txt      |   0
 .../test_batched_gemm_multi_d_dl.cpp          |   0
 test/batched_gemm_reduce/CMakeLists.txt       |   0
 .../batched_gemm_reduce_fp16.cpp              |   0
 test/batched_gemm_softmax_gemm/CMakeLists.txt |   0
 .../test_batched_gemm_softmax_gemm_fp16.cpp   |   0
 .../test_batched_gemm_softmax_gemm_util.hpp   |   0
 .../CMakeLists.txt                            |   0
 ...ed_gemm_bias_softmax_gemm_permute_bf16.cpp |   0
 ...ed_gemm_bias_softmax_gemm_permute_fp16.cpp |   0
 ...ed_gemm_bias_softmax_gemm_permute_util.hpp |   0
 ...batched_gemm_softmax_gemm_permute_bf16.cpp |   0
 ...batched_gemm_softmax_gemm_permute_fp16.cpp |   0
 ...batched_gemm_softmax_gemm_permute_util.hpp |   0
 test/batchnorm/CMakeLists.txt                 |   0
 test/batchnorm/batchnorm_bwd_rank_4.cpp       |   0
 test/batchnorm/batchnorm_fwd_rank_4.cpp       |   0
 test/batchnorm/batchnorm_infer_rank_4.cpp     |   0
 .../block_swizzle_test/block_swizzle_test.cpp |   0
 test/block_swizzle_test/rebuild.sh            |   0
 test/block_swizzle_test/simple_args.h         |   0
 test/block_to_ctile_map/CMakeLists.txt        |   0
 .../test_block_to_ctile_map.cpp               |   0
 test/contraction/CMakeLists.txt               |   0
 test/contraction/test_contraction.cpp         |   0
 .../test_contraction_interface.cpp            |   0
 test/conv_tensor_rearrange/CMakeLists.txt     |   0
 .../test_conv_tensor_rearrange.cpp            |   0
 .../test_conv_tensor_rearrange_interface.cpp  |   0
 test/conv_util/CMakeLists.txt                 |   0
 test/conv_util/conv_util.cpp                  |   0
 test/convnd_bwd_data/CMakeLists.txt           |   0
 test/convnd_bwd_data/convnd_bwd_data.cpp      |   0
 test/convnd_fwd/CMakeLists.txt                |   0
 test/convnd_fwd/convnd_fwd.cpp                |   0
 test/data_type/CMakeLists.txt                 |   0
 test/data_type/test_bf8.cpp                   |   0
 test/data_type/test_fp8.cpp                   |   0
 test/data_type/test_int4.cpp                  |   0
 test/data_type/type_convert_const.cpp         |   0
 test/elementwise_normalization/CMakeLists.txt |   0
 .../test_elementwise_layernorm_fp16.cpp       |   0
 test/gemm/CMakeLists.txt                      |   0
 test/gemm/gemm_bf16.cpp                       |   0
 test/gemm/gemm_fp16.cpp                       |   0
 test/gemm/gemm_fp32.cpp                       |   0
 test/gemm/gemm_fp64.cpp                       |   0
 test/gemm/gemm_int8.cpp                       |   0
 test/gemm/gemm_standalone_xdl_fp16.cpp        |   0
 test/gemm/gemm_util.hpp                       |   0
 test/gemm/instance/gemm_f16_nn_instance.cpp   |   0
 test/gemm/instance/gemm_f16_nn_instance.hpp   |   0
 test/gemm/instance/gemm_f16_nt_instance.cpp   |   0
 test/gemm/instance/gemm_f16_nt_instance.hpp   |   0
 test/gemm/instance/gemm_f16_tn_instance.cpp   |   0
 test/gemm/instance/gemm_f16_tn_instance.hpp   |   0
 test/gemm/instance/gemm_f16_tt_instance.cpp   |   0
 test/gemm/instance/gemm_f16_tt_instance.hpp   |   0
 .../instance/gemm_wavelet_f16_tn_instance.cpp |   0
 .../instance/gemm_wavelet_f16_tn_instance.hpp |   0
 test/gemm/run_gemm_test.inc                   |   0
 test/gemm_add/CMakeLists.txt                  |   0
 test/gemm_add/test_gemm_add.hpp               |   0
 test/gemm_add/test_gemm_add_fastgelu.cpp      |   0
 test/gemm_add/test_gemm_add_relu.cpp          |   0
 test/gemm_add/test_gemm_add_silu.cpp          |   0
 test/gemm_layernorm/CMakeLists.txt            |   0
 .../test_gemm_add_relu_add_layernorm_fp16.cpp |   0
 test/gemm_reduce/CMakeLists.txt               |   0
 test/gemm_reduce/gemm_reduce_fp16.cpp         |   0
 test/gemm_split_k/CMakeLists.txt              |   0
 test/gemm_split_k/test_gemm_splitk.cpp        |   0
 .../test_gemm_splitk_ut_cases.inc             |   0
 test/gemm_split_k/test_gemm_splitk_util.hpp   |   0
 test/grouped_convnd_bwd_data/CMakeLists.txt   |   0
 .../test_grouped_convnd_bwd_data.cpp          |   0
 ...grouped_convnd_bwd_data_interface_wmma.cpp |   0
 ..._grouped_convnd_bwd_data_interface_xdl.cpp |   0
 test/grouped_convnd_bwd_weight/CMakeLists.txt |   0
 .../test_grouped_convnd_bwd_weight.cpp        |   0
 ...ouped_convnd_bwd_weight_interface_wmma.cpp |   0
 ...rouped_convnd_bwd_weight_interface_xdl.cpp |   0
 test/grouped_convnd_fwd/CMakeLists.txt        |   0
 .../test_grouped_convnd_fwd.cpp               |   0
 ..._grouped_convnd_fwd_multi_ab_interface.cpp |   0
 ...nd_fwd_multi_d_interface_compatibility.cpp |   0
 test/grouped_gemm/CMakeLists.txt              |   0
 .../test_grouped_gemm_interface.cpp           |   0
 .../grouped_gemm/test_grouped_gemm_splitk.cpp |   0
 .../test_grouped_gemm_ut_cases.inc            |   0
 test/grouped_gemm/test_grouped_gemm_util.hpp  |   0
 test/magic_number_division/CMakeLists.txt     |   0
 .../magic_number_division.cpp                 |   0
 test/normalization_bwd_data/CMakeLists.txt    |   0
 .../test_groupnorm_bwd_data_fp32.cpp          |   0
 .../test_layernorm2d_bwd_data_fp32.cpp        |   0
 .../CMakeLists.txt                            |   0
 .../test_groupnorm_bwd_gamma_beta_fp32.cpp    |   0
 .../test_layernorm2d_bwd_gamma_beta_fp32.cpp  |   0
 test/normalization_fwd/CMakeLists.txt         |   0
 .../test_groupnorm_fwd_fp16.cpp               |   0
 .../test_groupnorm_fwd_fp32.cpp               |   0
 .../test_layernorm2d_fwd_fp16.cpp             |   0
 .../test_layernorm2d_fwd_fp32.cpp             |   0
 .../test_layernorm4d_fwd_fp16.cpp             |   0
 test/permute_scale/CMakeLists.txt             |   0
 test/permute_scale/test_permute_scale.cpp     |   0
 test/pool/CMakeLists.txt                      |   0
 test/pool/test_avg_pool3d_bwd.cpp             |   0
 test/pool/test_avg_pool3d_fwd.cpp             |   0
 test/pool/test_max_pool3d_bwd.cpp             |   0
 test/pool/test_max_pool3d_fwd.cpp             |   0
 test/pool/test_pool_fwd_common.hpp            |   0
 test/reduce/CMakeLists.txt                    |   0
 test/reduce/reduce_no_index.cpp               |   0
 test/reduce/reduce_with_index.cpp             |   0
 test/reference_conv_fwd/CMakeLists.txt        |   0
 .../reference_conv_fwd/reference_conv_fwd.cpp |   0
 test/softmax/CMakeLists.txt                   |   0
 test/softmax/test_softmax_interface.cpp       |   0
 test/softmax/test_softmax_rank3.cpp           |   0
 test/softmax/test_softmax_rank4.cpp           |   0
 test/softmax/test_softmax_ut_cases.inc        |   0
 test/softmax/test_softmax_util.hpp            |   0
 test/space_filling_curve/CMakeLists.txt       |   0
 .../space_filling_curve.cpp                   |   0
 test/transpose/CMakeLists.txt                 |   0
 test/transpose/test_transpose.cpp             |   0
 test/wmma_op/CMakeLists.txt                   |   0
 test/wmma_op/wmma_op.cpp                      |   0
 test/wmma_op/wmma_op_util.hpp                 |   0
 test/wrapper/CMakeLists.txt                   |   0
 test/wrapper/test_wrapper_copy.cpp            |   0
 test/wrapper/test_wrapper_gemm.cpp            |   0
 test/wrapper/test_wrapper_layout.cpp          |   0
 test/wrapper/test_wrapper_partition.cpp       |   0
 test/wrapper/test_wrapper_tensor.cpp          |   0
 3737 files changed, 479 insertions(+), 419 deletions(-)
 create mode 100755 .cache/clangd/index/add_device_operation_instance.hpp.C89DB8EBC3EFFE7D.idx
 create mode 100755 .cache/clangd/index/algorithm.hpp.3E7EDFC4AA4FB32B.idx
 create mode 100755 .cache/clangd/index/amd_address_space.hpp.FB92D54F0A7BAF6E.idx
 create mode 100755 .cache/clangd/index/amd_buffer_addressing.hpp.7FB2D5AE48874EC4.idx
 create mode 100755 .cache/clangd/index/amd_gemm_dpp.hpp.55564336A8EE078E.idx
 create mode 100755 .cache/clangd/index/amd_inline_asm.hpp.7777A491B073ADB0.idx
 create mode 100755 .cache/clangd/index/amd_lds.hpp.3739CBB0B218918C.idx
 create mode 100755 .cache/clangd/index/amd_wave_read_first_lane.hpp.F8A301BEB898DCEC.idx
 create mode 100755 .cache/clangd/index/amd_wmma.hpp.143B2B78EF519046.idx
 create mode 100755 .cache/clangd/index/amd_xdlops.hpp.5A2FDE085CB4BC8A.idx
 create mode 100755 .cache/clangd/index/array.hpp.EEBA77115A6888DA.idx
 create mode 100755 .cache/clangd/index/avg_pool3d_bwd.hpp.40079D84C914EB9A.idx
 create mode 100755 .cache/clangd/index/avg_pool3d_bwd_ndhwc_instance_common.hpp.EDEFB32170FF2FE7.idx
 create mode 100755 .cache/clangd/index/avgpool3d_bwd_bf16.cpp.CC3B9B35F218D95E.idx
 create mode 100755 .cache/clangd/index/avgpool3d_bwd_common.hpp.C26898608FBBAD5A.idx
 create mode 100755 .cache/clangd/index/avgpool3d_bwd_fp16.cpp.11C3DA7A01E70C82.idx
 create mode 100755 .cache/clangd/index/avgpool3d_bwd_fp32.cpp.4BE51B531517C414.idx
 create mode 100755 .cache/clangd/index/batched_gemm.hpp.E9C31BFE1C18D556.idx
 create mode 100755 .cache/clangd/index/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp.25B35CBDBE38B07B.idx
 create mode 100755 .cache/clangd/index/batched_gemm_add_relu_gemm_add.hpp.B9BDBBF21D219B22.idx
 create mode 100755 .cache/clangd/index/batched_gemm_bias_e_permute_xdl_fp16.cpp.6413F9674606CEBE.idx
 create mode 100755 .cache/clangd/index/batched_gemm_bias_softmax_gemm_permute.hpp.3291370BDDEB76AF.idx
 create mode 100755 .cache/clangd/index/batched_gemm_gemm.hpp.3359348295C6E9CF.idx
 create mode 100755 .cache/clangd/index/batched_gemm_gemm_xdl_bf16.cpp.38FE56A993CA3FF4.idx
 create mode 100755 .cache/clangd/index/batched_gemm_gemm_xdl_fp16.cpp.F5472E9B2342F0A2.idx
 create mode 100755 .cache/clangd/index/batched_gemm_gemm_xdl_fp32.cpp.958D740430E8AB79.idx
 create mode 100755 .cache/clangd/index/batched_gemm_gemm_xdl_int8.cpp.EF48A52F4EBD9A62.idx
 create mode 100755 .cache/clangd/index/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.20C4D0C379DDEEBB.idx
 create mode 100755 .cache/clangd/index/batched_gemm_multi_d.hpp.4E3016E76D827F13.idx
 create mode 100755 .cache/clangd/index/batched_gemm_reduce_fp16.cpp.795EB59AA922209F.idx
 create mode 100755 .cache/clangd/index/batched_gemm_reduce_xdl_fp16.cpp.6EA78CC685DE72B1.idx
 create mode 100755 .cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp.3B79E782257965A3.idx
 create mode 100755 .cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.71938F78DFDFE1BE.idx
 create mode 100755 .cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp.74FE791FA1A9F5A4.idx
 create mode 100755 .cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp.B77888D4EF3A20F0.idx
 create mode 100755 .cache/clangd/index/batched_gemm_softmax_gemm.hpp.2D7D43E1651E0616.idx
 create mode 100755 .cache/clangd/index/batched_gemm_softmax_gemm_permute.hpp.A9BA319B35F534EC.idx
 create mode 100755 .cache/clangd/index/batched_gemm_xdl_bf16.cpp.ED68A85BB00385BD.idx
 create mode 100755 .cache/clangd/index/batched_gemm_xdl_fp16.cpp.A653F9336CDDD36D.idx
 create mode 100755 .cache/clangd/index/batched_gemm_xdl_fp32.cpp.B7D138F072E276A3.idx
 create mode 100755 .cache/clangd/index/batched_gemm_xdl_int8.cpp.AEAA0DF5DDD6BDB1.idx
 create mode 100755 .cache/clangd/index/batchnorm_backward.hpp.7322FEA177E6E308.idx
 create mode 100755 .cache/clangd/index/batchnorm_backward_nhwc.cpp.7637AB6D79637171.idx
 create mode 100755 .cache/clangd/index/batchnorm_bwd_rank_4.cpp.15A9A6FE401A579D.idx
 create mode 100755 .cache/clangd/index/batchnorm_common.hpp.7297BD2852C87A9D.idx
 create mode 100755 .cache/clangd/index/batchnorm_forward.hpp.7C4BF3E98C6D2CE7.idx
 create mode 100755 .cache/clangd/index/batchnorm_forward_inferring_nhwc.cpp.5CE047E1603B8333.idx
 create mode 100755 .cache/clangd/index/batchnorm_forward_training_nhwc.cpp.6918CB89E3F27E24.idx
 create mode 100755 .cache/clangd/index/batchnorm_forward_training_nhwc_obsolete.cpp.B9948F07339573BC.idx
 create mode 100755 .cache/clangd/index/batchnorm_fwd_rank_4.cpp.051838AC0BBFCDA1.idx
 create mode 100755 .cache/clangd/index/batchnorm_infer.hpp.F7608824D3FE5589.idx
 create mode 100755 .cache/clangd/index/batchnorm_infer_impl.hpp.0B7228D5ED433292.idx
 create mode 100755 .cache/clangd/index/batchnorm_infer_rank_4.cpp.16E993641F247D85.idx
 create mode 100755 .cache/clangd/index/binary_element_wise_operation.hpp.D7064E30A6CF6778.idx
 create mode 100755 .cache/clangd/index/block_to_ctile_map.hpp.F32EBB9EC094E255.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_dpp.hpp.5CB9342685AA4AD4.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_pipeline_xdlops.hpp.C000528561553163.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_wmma.hpp.33F41B2FF8044BC6.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_xdl_traits.hpp.9E775B78211AACFC.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_xdlops.hpp.E5193AF5C1961F73.idx
 create mode 100755 .cache/clangd/index/blockwise_gemm_xdlops_skip_b_lds.hpp.BC4BEC2371CE0CF7.idx
 create mode 100755 .cache/clangd/index/blockwise_softmax.hpp.D0089700D1C9998B.idx
 create mode 100755 .cache/clangd/index/blockwise_welford.hpp.93790E23276AA4A9.idx
 create mode 100755 .cache/clangd/index/broadcast_add_2d_amn_bn.cpp.7B44FEB5DE986168.idx
 create mode 100755 .cache/clangd/index/broadcast_add_3d_am_bmnk.cpp.D83172D8CBF829F1.idx
 create mode 100755 .cache/clangd/index/c_style_pointer_cast.hpp.6866D538A80ECC76.idx
 create mode 100755 .cache/clangd/index/cgemm_xdl_bf16.cpp.D34D118D7BC516E4.idx
 create mode 100755 .cache/clangd/index/cgemm_xdl_common.hpp.FBF552CB8A2753F4.idx
 create mode 100755 .cache/clangd/index/cgemm_xdl_fp16.cpp.93DC5CFE4CF25E39.idx
 create mode 100755 .cache/clangd/index/cgemm_xdl_fp32.cpp.D4109241584CC5F6.idx
 create mode 100755 .cache/clangd/index/cgemm_xdl_int8.cpp.7D0F341E5936B471.idx
 create mode 100755 .cache/clangd/index/check_err.hpp.8BCADBDE7B4A301B.idx
 create mode 100755 .cache/clangd/index/ck.hpp.5BA5ECB9E73F13E2.idx
 create mode 100755 .cache/clangd/index/cluster_descriptor.hpp.B962ADBD27415A8A.idx
 create mode 100755 .cache/clangd/index/column_to_image_f32.cpp.B78257C7C44FE524.idx
 create mode 100755 .cache/clangd/index/common.hpp.030C990214BE859F.idx
 create mode 100755 .cache/clangd/index/common.hpp.0CF2F8C8A70630AF.idx
 create mode 100755 .cache/clangd/index/common.hpp.116380F0FD4C88BB.idx
 create mode 100755 .cache/clangd/index/common.hpp.30E7C524D1E7DCC8.idx
 create mode 100755 .cache/clangd/index/common.hpp.3324DE766DB5A26E.idx
 create mode 100755 .cache/clangd/index/common.hpp.725393A22E55A928.idx
 create mode 100755 .cache/clangd/index/common.hpp.7922E75661986F8F.idx
 create mode 100755 .cache/clangd/index/common.hpp.98AE17ADFF3CA289.idx
 create mode 100755 .cache/clangd/index/common.hpp.A2FB78C7F0E24A0E.idx
 create mode 100755 .cache/clangd/index/common.hpp.AA5CB32F503E8500.idx
 create mode 100755 .cache/clangd/index/common.hpp.D8FBDC6F47AA6D80.idx
 create mode 100755 .cache/clangd/index/common.hpp.E36EEFF638BE348F.idx
 create mode 100755 .cache/clangd/index/common.hpp.F52D770AC14BC38D.idx
 create mode 100755 .cache/clangd/index/common.hpp.FF4590F87A18C9AE.idx
 create mode 100755 .cache/clangd/index/common_header.hpp.BCB8C81E7AAAD43F.idx
 create mode 100755 .cache/clangd/index/common_instances.hpp.C20932BA881E0448.idx
 create mode 100755 .cache/clangd/index/container_element_picker.hpp.D0295372D9F2DFD0.idx
 create mode 100755 .cache/clangd/index/container_helper.hpp.57BCB53456250D8A.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear.hpp.3FE8920B21D2C083.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_bf16_compute_fp32.cpp.34FCAF706B5AEB99.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp16_compute_fp32.cpp.038CE2A80837D541.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp32.cpp.37E0EC20BE4060B3.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp32_compute_bf16.cpp.4B6BEF9EB4DD9C01.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp32_compute_fp16.cpp.F92204694821E2EC.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp64.cpp.D740E9866ECDD95C.idx
 create mode 100755 .cache/clangd/index/contraction_bilinear_xdl_fp64_compute_fp32.cpp.29832DEA4C5E2D0A.idx
 create mode 100755 .cache/clangd/index/contraction_multi_ABD_xdl_fp16.cpp.95E96326F13F98F9.idx
 create mode 100755 .cache/clangd/index/contraction_scale.hpp.9A93BE54CAD13142.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_bf16_compute_fp32.cpp.2E396E91009DF0FA.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp16_compute_fp32.cpp.A7A07273739AC223.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp32.cpp.1A2C9B556CFF860B.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp32_compute_bf16.cpp.301F203197216C44.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp32_compute_fp16.cpp.F7341E88A32312DC.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp64.cpp.7417F385FB954651.idx
 create mode 100755 .cache/clangd/index/contraction_scale_xdl_fp64_compute_fp32.cpp.02A2D58B76014A4B.idx
 create mode 100755 .cache/clangd/index/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp.D342A33945D6732C.idx
 create mode 100755 .cache/clangd/index/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp.E5EA635D0BE03F7D.idx
 create mode 100755 .cache/clangd/index/conv2d_fwd_xdl_perchannel_quantization_int8.cpp.DA805D9EE5DF962F.idx
 create mode 100755 .cache/clangd/index/conv2d_fwd_xdl_perlayer_quantization_int8.cpp.7D138D1D290C5C53.idx
 create mode 100755 .cache/clangd/index/conv2d_quantization_common.hpp.2C2B760A7CE8525E.idx
 create mode 100755 .cache/clangd/index/conv_fwd_xdl_scaleadd_ab_bf16.cpp.8FB5CAA0FDE2EFE8.idx
 create mode 100755 .cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp16.cpp.A862E22D755445CE.idx
 create mode 100755 .cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp32.cpp.E8494EFCFB4F3B30.idx
 create mode 100755 .cache/clangd/index/conv_fwd_xdl_scaleadd_ab_int8.cpp.97C188D55BA3C9C1.idx
 create mode 100755 .cache/clangd/index/conv_tensor_rearrange.hpp.1BD4F049133037BC.idx
 create mode 100755 .cache/clangd/index/conv_tensor_rearrange_op.hpp.1DE16BC6AE42EE5B.idx
 create mode 100755 .cache/clangd/index/conv_util.cpp.2DB2D9F346007F63.idx
 create mode 100755 .cache/clangd/index/convnd_bwd_data.cpp.7C4E178AEE50DCF2.idx
 create mode 100755 .cache/clangd/index/convnd_bwd_data_common.hpp.FAAF76603FF7D70B.idx
 create mode 100755 .cache/clangd/index/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp.17D6FDBA1D601B9A.idx
 create mode 100755 .cache/clangd/index/convnd_bwd_data_xdl_fp16.cpp.32F247EF8BEF9A27.idx
 create mode 100755 .cache/clangd/index/convnd_fwd.cpp.63437EF3B0E7154B.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_activ_multi_ab_common.hpp.F16D50D450CBF542.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_activ_unary_common.hpp.A3C6C6A8E8602CD2.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_common.hpp.1735C0647D85C158.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_max_xdl_bf16.cpp.32DD457450CC173D.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_max_xdl_fp16.cpp.57DCB3A8914F4BCB.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_max_xdl_fp32.cpp.7C22839582612580.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_max_xdl_int8.cpp.E84F6E18A0DCB8A3.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_abs_fp16.cpp.B14E3C722763D4B8.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_bf16.cpp.054A24EF5BFF4A1F.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_bilinear_residual_fp16.cpp.C4A66C697A8DEBDA.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_clippedrelu_fp16.cpp.4759FB8DCFC58E7D.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_elu_fp16.cpp.922AD9048F6922A8.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_fp16.cpp.E044D0CD44CCFE7F.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_fp32.cpp.911AF34195327474.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_fp64.cpp.3D0123C922DFAFA2.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_int8.cpp.1BF1AFB682499568.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_leakyrelu_fp16.cpp.BA4E6713D7C74F09.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_pow_fp16.cpp.0F662794CF8A0DB2.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_relu_fp16.cpp.8F7B959CCFC1E66F.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp.54C5613FFA1E2460.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp.303711D8C093AB3A.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_sigmoid_fp16.cpp.7462E7274E7DC93D.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_softrelu_fp16.cpp.47B66AF94BC3AD6A.idx
 create mode 100755 .cache/clangd/index/convnd_fwd_xdl_tanh_fp16.cpp.A1CCA35C668C7A32.idx
 create mode 100755 .cache/clangd/index/convolution_backward_data.hpp.F488158732BCE944.idx
 create mode 100755 .cache/clangd/index/convolution_backward_data_specialization.hpp.E2F7DA4511290758.idx
 create mode 100755 .cache/clangd/index/convolution_backward_weight_specialization.hpp.D1A61821B9657E4F.idx
 create mode 100755 .cache/clangd/index/convolution_forward.hpp.8E3EE29C519B7A7E.idx
 create mode 100755 .cache/clangd/index/convolution_forward_specialization.hpp.29B75FDB4D9A7ED7.idx
 create mode 100755 .cache/clangd/index/convolution_host_tensor_descriptor_helper.hpp.C7E0FAE3D228F16B.idx
 create mode 100755 .cache/clangd/index/convolution_parameter.cpp.AB24172BD3A32784.idx
 create mode 100755 .cache/clangd/index/convolution_parameter.hpp.98A4E91AB8FFEC38.idx
 create mode 100755 .cache/clangd/index/copy.hpp.D47879DBFF83B386.idx
 create mode 100755 .cache/clangd/index/data_type.hpp.FA421D61CFEE8D86.idx
 create mode 100755 .cache/clangd/index/data_type_enum.hpp.9C984CB93A76A458.idx
 create mode 100755 .cache/clangd/index/debug.hpp.3D8E2A10AC86C578.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp.8DD0CEE151EEE78C.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp.E45DD97792659A1B.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp.0A50ACDD7655DC8C.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp.C65ED37BE0626399.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp.322ABBBE306F5168.idx
 create mode 100755 .cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp.BE9138723EE82CA7.idx
 create mode 100755 .cache/clangd/index/device_avgpool3d_bwd_ndhwc_ndhwc.hpp.B8CFD83A7F04917D.idx
 create mode 100755 .cache/clangd/index/device_avgpool_bwd.hpp.65CD0007480F503D.idx
 create mode 100755 .cache/clangd/index/device_base.hpp.07667515082DA1A8.idx
 create mode 100755 .cache/clangd/index/device_batched_contraction_multiple_d.hpp.C066C76CB17254D1.idx
 create mode 100755 .cache/clangd/index/device_batched_contraction_multiple_d_xdl_cshuffle.hpp.9FB9459677B68FA8.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm.hpp.0D6DC4AA71BF6793.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.2CCE15A52A6097D9.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.09BE891BDB3DB7E7.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp.FE05EE1022E21C41.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.5255F78DB2565F3C.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.B4883A86B075066E.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_gemm.hpp.FDFBA95BB6D583D3.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle.hpp.8AECE81C6A31D6CE.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.24C81AB80A523E1D.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.42B3A408DD25BE6F.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_multi_d.hpp.65CBFD3787EC8604.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_multi_d_xdl.hpp.5F31589668DB0CDD.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d.hpp.1EC37BB5D09F3746.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp.1D083EFB833A8BDD.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle.hpp.008517D2C19930C3.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp.3CE3CBFFC33E63B5.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp.C2D6CBC7BDD03D15.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp.BE70F5327F5BF3ED.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp.542AEDD871493BF1.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm.hpp.2CF55436E05C96D5.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_permute.hpp.0F50C2899670001D.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.72B4F2954B40EC19.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.CA655EDB759D211C.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.54BCAA9E5C4C7CA0.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp.5F249F65D9D49000.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.9E1D81891D4D775B.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl.hpp.336993554588B91D.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp.01B3A3CE0FB07AA4.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp.6732FB26500DDC13.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp.5C50F8886E987210.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp.991CBBA160B6AE8A.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp.909733D52FF49021.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp.EF1B1B5023E691DD.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp.6191F299E76FA25A.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp.1378958A5E1CC674.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp.155CED9F59B77387.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp.021E4B84540D9AAD.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp.2D0A22E72C5CD4C0.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp.6662845BB6111F27.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp.3B9F48B295268F74.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp.0C31B57B423E19C5.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp.8439FC102CF881D2.idx
 create mode 100755 .cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp.9E2B022D116E5337.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward.hpp.43E337CEA4624B30.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward_bf16_instance.cpp.CEBDED0AD1292F09.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward_f16_instance.cpp.D1D5284FB6757AF3.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward_f32_instance.cpp.CA2117C76C8B866B.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward_f64_instance.cpp.B99A50DCD60DF309.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_backward_impl.hpp.8B916A92DFA8139E.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward.hpp.0E1593921379F2F3.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_bf16_instance.cpp.F51389A9E014ED43.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_f16_instance.cpp.5B31F4805B2AC1BA.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_f32_instance.cpp.DAA81C77B8C2B284.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_f64_instance.cpp.B473866A874EE221.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_impl.hpp.6D2CA9CBDDC84AC3.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_forward_impl_obsolete.hpp.7A0467126AEB6502.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_infer.hpp.4EDE573CB1DF1297.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_infer_bf16_instance.cpp.FEC3D08EDEA4CE6B.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_infer_f16_instance.cpp.B01DECCDEE63A78A.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_infer_f32_instance.cpp.889B58511EB200D7.idx
 create mode 100755 .cache/clangd/index/device_batchnorm_infer_f64_instance.cpp.0266490535848A23.idx
 create mode 100755 .cache/clangd/index/device_cgemm.hpp.CD6CA3F05D9AAF4D.idx
 create mode 100755 .cache/clangd/index/device_cgemm_4gemm_xdl_cshuffle.hpp.CB4931718165FBCB.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_gndhwc_3d_instance.cpp.8AF56F1E7AD14DC6.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_gnhwc_2d_instance.cpp.8EBF74F132E896AB.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_gnwc_1d_instance.cpp.C6C85A0C333E012A.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_impl.hpp.545E1241F8BEB14B.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_instance.hpp.FA1AC606D0CD0897.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_ndhwgc_3d_instance.cpp.CBDECCEC6F806814.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_nhwgc_2d_instance.cpp.9A3328B90E30CC03.idx
 create mode 100755 .cache/clangd/index/device_column_to_image_nwgc_1d_instance.cpp.163F7E8C2077B0E0.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp.79A78F49E52676FD.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp.DB04A314807BEF1C.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp.E624632F1E715961.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp.CACB611E1CE9F1B5.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp.9E3DEE644D650E5A.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp.384AC715F1A68557.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp.66D5EE934903111A.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp.8194F48DCC1F513B.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp.ED87AF8FD1B46167.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp.9B28805BAED0F92D.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp.7E6A2A9800DD0E7A.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp.E0C6A616E43DC9DD.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp.01988F25803E8456.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp.AFEB4718FB33138B.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp.D913F33B640C756C.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp.EF41093E7954CFCE.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp.CB877B16EA5C28D0.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp.7AE9E872C7D0A549.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp.E366D05CBA9CA229.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp.F88287BB0527CE6D.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp.03F037ABB1A17197.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp.237A2B88B920F202.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp.177CE21E4E683B24.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp.FE8F7C3907743A26.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp.1FB1A94EA79EE13A.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp.6BFB81AFEEB1C1B7.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp.D3FA995D6DF314A6.idx
 create mode 100755 .cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp.D00F48772AF74F17.idx
 create mode 100755 .cache/clangd/index/device_contraction_instance.hpp.15AB9D8608FA40A0.idx
 create mode 100755 .cache/clangd/index/device_contraction_multiple_abd.hpp.D0EC339053CFF7D0.idx
 create mode 100755 .cache/clangd/index/device_contraction_multiple_abd_xdl_cshuffle.hpp.9AA6B0ABD9EE929D.idx
 create mode 100755 .cache/clangd/index/device_contraction_multiple_d.hpp.CA3D08C20FA799BC.idx
 create mode 100755 .cache/clangd/index/device_contraction_multiple_d_xdl_cshuffle.hpp.18531F18DA84F612.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp.634AB0569FA8301A.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp.1229C32BCA5E8056.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp.B21239C1EC74587F.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp.92168E842E069568.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp.8EFB8D272907AC49.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp.4BAF3FDF22005D6C.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp.F76E34EF68D8A65B.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp.1704AA4E0EFAB142.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp.29A1243F9B97E124.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp.020B0D68AE892602.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp.A426959DB1641563.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp.1B1383BC1860443B.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp.C0ABEB6CD7C15E09.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp.76E857B0424F98D1.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp.2C4EF55D06D154CA.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp.D9C825C0FBC7162C.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp.9ED7E55B00DD6F70.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp.87288671AE2510FD.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp.AA8697A6E53E704C.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp.FC0AF91C98199AEA.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp.265462C1E86DC0F1.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp.32EBBC25AA6C0CF0.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp.5C1840A2D370420C.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp.05BFD4E4455B0BDF.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp.BFA2076F1038BB65.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp.16DF4E204049F20E.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp.25AD39BA7B2B9C77.idx
 create mode 100755 .cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp.F46067E1C6499190.idx
 create mode 100755 .cache/clangd/index/device_contraction_utils.hpp.39AAD30D5F9C497C.idx
 create mode 100755 .cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp.F95715A59DC357E4.idx
 create mode 100755 .cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp.11B1653CC880738D.idx
 create mode 100755 .cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp.1FF1E6FFC39B05F6.idx
 create mode 100755 .cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp.27AC774D5173DB46.idx
 create mode 100755 .cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp.5A99B55004C02041.idx
 create mode 100755 .cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.38978BB017721061.idx
 create mode 100755 .cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.015682E6D487BC89.idx
 create mode 100755 .cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.24521662C376153B.idx
 create mode 100755 .cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.050ACEE75FFD374F.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp.095F638AA48F9495.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp.03A5D53DEE4E2511.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp.FF7059ED53F4F0F2.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp.2C9D98DCC57DC34E.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp.66D8D8BBD73054E8.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp.3FB08C96D8C7394B.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp.FAAE94F7CEBB58BA.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.9102AFA84B67CE1B.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.656EAF4E8A7F404D.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.793B97DB82C3A02D.idx
 create mode 100755 .cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.A7E63495E5270E3E.idx
 create mode 100755 .cache/clangd/index/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp.486913CE7740FA6B.idx
 create mode 100755 .cache/clangd/index/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp.8577D499B9483812.idx
 create mode 100755 .cache/clangd/index/device_conv2d_xdl_int8_instance.hpp.78D2B94EA941EB43.idx
 create mode 100755 .cache/clangd/index/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp.C31E228888CDC52C.idx
 create mode 100755 .cache/clangd/index/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp.108EC38462E7A7B4.idx
 create mode 100755 .cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp.7487AFBFF492DC5A.idx
 create mode 100755 .cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp.B17F2553F57B25A6.idx
 create mode 100755 .cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp.D3A7F488899285AF.idx
 create mode 100755 .cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp.31CAAB24531BB9D5.idx
 create mode 100755 .cache/clangd/index/device_conv_bwd_data.hpp.81B90E235167286C.idx
 create mode 100755 .cache/clangd/index/device_conv_fwd.hpp.257583CFB6A8DC72.idx
 create mode 100755 .cache/clangd/index/device_conv_fwd_bias_activation.hpp.D4F53CD24AD4278A.idx
 create mode 100755 .cache/clangd/index/device_conv_fwd_bias_activation_add.hpp.D947A059FE31F7BB.idx
 create mode 100755 .cache/clangd/index/device_conv_tensor_rearrange.hpp.90268624B8539288.idx
 create mode 100755 .cache/clangd/index/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp.A70DE3BD61066477.idx
 create mode 100755 .cache/clangd/index/device_elementwise.hpp.00198FF49AAFC017.idx
 create mode 100755 .cache/clangd/index/device_elementwise_2d_impl.hpp.30FD009D4EA9FEE5.idx
 create mode 100755 .cache/clangd/index/device_elementwise_3d_impl.hpp.EA5D0D82E2445BC8.idx
 create mode 100755 .cache/clangd/index/device_elementwise_impl.hpp.54205731D4AB5FA8.idx
 create mode 100755 .cache/clangd/index/device_elementwise_normalization.hpp.2F6431A98ECDBEBB.idx
 create mode 100755 .cache/clangd/index/device_elementwise_normalization_f16_instance.cpp.B8CDDDF357D7D999.idx
 create mode 100755 .cache/clangd/index/device_elementwise_normalization_impl.hpp.D9648470B11E62DC.idx
 create mode 100755 .cache/clangd/index/device_elementwise_scale.hpp.A8B4C8DA3A99ADD2.idx
 create mode 100755 .cache/clangd/index/device_elementwise_scale_impl.hpp.792CB7A86911D632.idx
 create mode 100755 .cache/clangd/index/device_gemm.hpp.C11C88B2080F572F.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.236524D86BD9A065.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.3F5C595E84E47B20.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.5836A0C2B3A14787.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.E0256C2451B1EE93.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.EED2788EEB0B75F1.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.F71CF786157F3658.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.777F64009D887A71.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.9B8C2015B4635A97.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.7475A9472170E32A.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5CFCD880070835B7.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.15161CADD4CF1397.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.C192E44A2A2B2438.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.0E4AE68E13A6F5B4.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.3B87280201C19511.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp.E948F8CEDD1EC4BF.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp.0BA2333B2B00FEDC.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp.F5D89038BFC386C8.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp.935CC5A9EFAF5617.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.81A12F337280E714.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.6BB8AF1DE23C88B0.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.6559BA68564A041C.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5F5C907654F27409.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.8D291954B92003AF.idx
 create mode 100755 .cache/clangd/index/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.9FCCE6CADB35B88C.idx
 create mode 100755 .cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.479D1366D0F3CE30.idx
 create mode 100755 .cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.D6BBA9767E8DA592.idx
 create mode 100755 .cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.F658EA6412E57274.idx
 create mode 100755 .cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.D219505C6A3BDF13.idx
 create mode 100755 .cache/clangd/index/device_gemm_bias_add_reduce_xdl_cshuffle.hpp.2AAE1676CF63D1B6.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp.4E5648F8B76EED28.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp.51C1492BB7273708.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp.89C496A745057432.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp.DA4EFE2E10A56944.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.80BE80F117BA699A.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.E76894F842B8152E.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.741D3DF3032F48B6.idx
 create mode 100755 .cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.E3C0C7076B1A5420.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp.hpp.86A83F148A978F3B.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp.AC3765D99685AA83.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp.BE2169842E629D5F.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp.C525CCAE82B88202.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp.E70AE989E444006E.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp.F48BB29B1F2C3D07.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.9E7DF32B7E3B06F9.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp.8E876A3BCC68461F.idx
 create mode 100755 .cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F358D66C9DE4EC29.idx
 create mode 100755 .cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.B5661BE447C6D950.idx
 create mode 100755 .cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.E716A8264A34B5D8.idx
 create mode 100755 .cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.B8E2B7EF254C9C8D.idx
 create mode 100755 .cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.4104F408858A9B41.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_abd.hpp.01CABF5AC7B8D71E.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_abd_xdl_cshuffle.hpp.6B55B3AF2C69EA61.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d.hpp.A369BE7D23825A77.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_layernorm.hpp.1996D4382B7902A4.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp.6897EB3658A7024F.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_multiple_r.hpp.D51C7B3444D17F61.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.76B6E1492C6F98CC.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_wmma_cshuffle.hpp.5BE3B8B060B3F7DD.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle.hpp.A234525D7676EE3C.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.D18AD937E7A47E2A.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.46B874AA40214074.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.0EE0667EFBB4A693.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp.4B2A61CBEBB69C46.idx
 create mode 100755 .cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp.E45F1434A64DD3E5.idx
 create mode 100755 .cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp.35BB7F78B886DD68.idx
 create mode 100755 .cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.6F38C058A13FF1D3.idx
 create mode 100755 .cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.0B3E6146A79A7808.idx
 create mode 100755 .cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.487F7DBAC716F5E6.idx
 create mode 100755 .cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.0D311D4C9F7E0828.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce.hpp.536083621BE180B4.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce_xdl_cshuffle.hpp.E3AE2B4768599A5C.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.7921F720A3C738B2.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.67A4E982DE011DE5.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.18986AD46051D542.idx
 create mode 100755 .cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.93B77417ED344C2F.idx
 create mode 100755 .cache/clangd/index/device_gemm_splitk.hpp.CFD1B5B13FAD53CD.idx
 create mode 100755 .cache/clangd/index/device_gemm_streamk.hpp.75A52E82DB5C165B.idx
 create mode 100755 .cache/clangd/index/device_gemm_wmma.hpp.32E8352ACBB1259F.idx
 create mode 100755 .cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp.48B71A3DBBF3DF70.idx
 create mode 100755 .cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp.9F95D8CF713B4DFB.idx
 create mode 100755 .cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp.4656D94005983107.idx
 create mode 100755 .cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp.92DF213C04CE83E0.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl.hpp.8498EA503EEB08EE.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp.F73590DBD2CB1EF5.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp.544964C8084D9197.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp.6E89B9B9359CE514.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp.2346E4E4EAE3DC7F.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp.DCA2A91E3CF88135.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.52C4510F756E0106.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.11CF3D24C90FB35A.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.3180F054887BBB56.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.CA124F6888DE0DFE.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp.5B0B95461D582E5E.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp.144394AC7C436910.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp.9C26807545ECBC7F.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp.48B6F9C0E69BB63A.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp.FF4791DC4DC38C18.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp.6D261DF3AD15EA25.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp.FA9B7FC9807074DA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp.E10E580D456628E1.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp.030C33F297B655BC.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp.C8A37285FABA85B1.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp.1333260C2CA69DB3.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp.B5F4680BE43A28CE.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp.4093668F9B2A46D0.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp.A50E70F0D93B6F8E.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp.74A0C6504550F5D4.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp.A977355E082D2CA2.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.2559A4A420F11829.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.513C881E10886D27.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.43610B5AFBA5C7DA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.448FFB8EF559CCD0.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6CE81023C2E130B0.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp.62E7711AB97BA061.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp.FC50F49FB1ACD418.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp.29C6735FE1834089.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp.2E316B49437298E2.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_cshuffle.hpp.54159ABC735BA0BB.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_cshuffle_lds_direct_load.hpp.C2FB5C5C6E8BC96E.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_cshuffle_v2.hpp.6D300E1E67F0C8FA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp.9F68ED656E988D30.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp.FA79EAE16FB6F44E.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp.F19AC2A6D8DD58E6.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp.478DBC5276E71817.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp.1759B855C9E56265.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp.9CBE0B752A41740B.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp.2A9A70463B0B1D69.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp.E46BEEC813D33535.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_layernorm_cshuffle.hpp.7536F92E23B638A6.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_skip_b_lds.hpp.31130252724E07CF.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_c_shuffle.hpp.D2F306B1868E426C.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp.C1DEA5CF04C380FB.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp.5C4328570228C900.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp.8E11BCA1E8B40C54.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp.02602D130B49C5E9.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp.76573A416742936A.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp.80C5AB0FEC7DFC63.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp.073EE92A5EAA844A.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp.98C08C6647949567.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.436428D0A84C0AFE.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp.5798AD2D53C823E4.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp.34C01D1CCA40C5F9.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp.2C6EF29B32E3350B.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp.483FB0E67DC82A12.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp.04904E13881A008D.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp.BF7C969BEF21A812.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp.11676556038472A5.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp.E685A8514CD614D5.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp.02D5EB3BF76A9EDC.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp.235E5C5B653C03E0.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp.37CD0D23BDED400B.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp.2924213EB866E058.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp.3FB0D5B09C446777.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp.7D14F328651204D1.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp.8A9B71EB3B2C8B82.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp.CF381829133E9AFB.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp.117C2517C948C007.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp.C13174B8631F1EA5.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp.8596E811350866B4.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp.09DE6214A3D6FEFC.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp.515130D1DB38B8AB.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp.252C84373128A03D.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp.6BA0BD5269CBABBA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp.CFD1881950858A56.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp.9CC79FF6A90215E2.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp.577D62F548B80AD3.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp.FC0BE921C33562AA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.5E4A00F1466A4D4D.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp.64CDFD8CF1246F3A.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp.38A1F46E05CA49C4.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6DFDC664FAC2C818.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_streamk.hpp.424C1309002226DA.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp.16580B63D7D6888C.idx
 create mode 100755 .cache/clangd/index/device_gemm_xdl_waveletmodel_cshuffle.hpp.6149EF88F8FEFF53.idx
 create mode 100755 .cache/clangd/index/device_grouped_contraction_multiple_d.hpp.D8736023CD2927DC.idx
 create mode 100755 .cache/clangd/index/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp.A6618A549365EFCF.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.A4DBB4F8B2C78E56.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.315EC2A3A6D6E638.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.AA168C235A47B299.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.FA613A1A341E3811.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.F48B669764788B12.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.B296FF7D04FDFC90.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp.DDED95C8C9D22159.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.15D71B57C7DA4CC3.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.1D323BBA44E4F194.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.900C533B3DB605B7.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.5AFEEB9815EAB2BB.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.BC2C6EDB175EB210.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.8A9EF4FB27AD09E7.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.DE2A54B24DD9CFDF.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.EBA14072309C61B4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.5B618BE41EFBBD23.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.6161CF2711103FB3.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.77ED46D094B0B411.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.07FF698C2C4EA5B3.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.1FAF079E735BA4BC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.FD07473382303B16.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.B6F367ED6FE27214.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.BE1B4F87BC348C49.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.F02158140F1CDE9D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.9325EE22E541E919.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.BBF8A6B4ABA52D27.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.CE7231939A842BC9.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp.5A5CAFE59DAED728.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.B9B0155CE4F94655.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.F1D87E46323F05EC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp.B674F131151301E0.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp.E06437315FCA8871.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.159F0712016AF057.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.4C82308AB1780753.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp.872D44B53EF52F10.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp.83CDF2CC13C4A415.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.A4A9E7E93B446A25.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.729E5DD280A9E536.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp.649F61751C48D48D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp.8A0C33E6B1E03B1D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.8C4CEC479CC69615.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.D16C834304C1756E.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp.1399C252DECC9B8A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.3060AD718DAF4F4F.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.8A08D10AF6807EE2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.FE154FD36F572C54.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.56FD421A51C8C5E1.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.C29BF8AC2D59E669.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.9341E131FE1393D1.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.11490B4C8E752971.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.ED0E8CE7C6A61531.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5ED435B77281FBCD.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.069AF0356549719D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.511C2FDB97E265FC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.32DD26A1AFC28FC7.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.72C23D8E55C505F5.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.80E125D05CA137B9.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.663BA57C2C8487A1.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.FDD27F96EE937A50.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.33C6E967995C7F50.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.59DE1C78AB7D5F80.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.F3F4F9DA4C0F8BDA.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.A371B626E37E8E07.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.AB40F2DD408B7367.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.18F54267E4ABEC46.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.192FAAC8A46AFF2C.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp.F56F1203139F8DB9.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.4C04B7436EBF8811.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.D8BA94FD49F7A399.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C14D5D3E331E58A4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.066E2D36BE87D012.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.341F4486D5AB6E5A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5A4CE3D80E2AF506.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.403AE32B0F021E71.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.9C8F172FA791F403.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.2C419A4DCD2F2168.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.217BA51B10EED1BA.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.CE8AF0B7DC8618CC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.8B802336DBA274D4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.17BBA75937C47055.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.9C7E5EE7984A7775.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.15DDB04E41B23392.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp.C031B76FFFB0A20D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.C7145A60489EBDB5.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C5AE3E9229A8EFB4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp.8F52AF019F02B944.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.3C3BEC3EC7808D5A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.2554570633260D87.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp.D8F2815AC5D0D7D6.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp.0FC1DC5FE52A0C60.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.0151E7B579E72F93.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.C058D2A1DEE10686.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp.278F3C311504F9EC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp.8DE51E7D513808A9.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.B3BD1E3E3F6E4877.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3955AFFBF9D5A2EF.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp.C97D071C6295D5F3.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp.E621A4FA60DE216B.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.06F3665DC4884ADD.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.8A48A77FE8BA877A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp.85809BA7EA818954.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.FB2E0A9CD009454C.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.21AF094B4B66A48D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.D78BE8BB81161A14.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.143C90F3DEDC2341.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.426F56794F68AE44.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.1C2FE2C88C10C9D6.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.BB1A75845E7DE245.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp.A1E84AD887D755DB.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.43CB1A6BDABED53D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp.336E2C49F8565116.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.117E24CA3DC21528.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.59CFAC251237A8A2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B76F39C8DBBE2FD2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.9F991AFB3119C4B9.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.A3477CD2DF481BB7.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C60AD9A6203D4F40.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B74C847DC291EB3D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.91739D7EC5D8C4E5.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.612D072BD9DA9C78.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.91C96AAC7F5772DF.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.A06BD8EFCD977C2D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.0F35369ACDDEB6D2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3DF5C2CC3145E799.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.B3BE2F1E169224A4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.ED09CF7C1628C3B6.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_multiple_d.hpp.CB0E4FDBA2354B83.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp.62AEEAEB4FE47662.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp.DA7CD25895EF8940.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_wmma_f16_instance.hpp.FE4A66EFBF344CEC.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_wmma_i8_instance.hpp.22BF4E121F8B1D63.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp.8935210F01B214E2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_xdl_instance.hpp.7A273F666FE77357.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_data_xdl_scale_instance.hpp.A306BD940C6CC8E3.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_weight.hpp.FB767EE60994A05E.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp.2CEBA1D92F4EDC1A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_weight_wmma_instance.hpp.D8A887BE715EE7ED.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp.7B0144E3F93B6769.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_bwd_weight_xdl_instance.hpp.DF771DCE13775534.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_abd.hpp.ED9472BD269F359D.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp.52A5FB3B86490259.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_d.hpp.79D64EC0672BD2D2.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r.hpp.4DCB681EEBE1DA60.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp.CF5C02CB8B4910C1.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp.EC51DA2544106248.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_wmma_instance.hpp.377DCBF55D8BBCBB.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_xdl_bilinear_instance.hpp.4827161718A2E399.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_xdl_instance.hpp.DE209B7BB2124A4A.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_xdl_scale_instance.hpp.F388F6485863F7A4.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp.043570DA9A94B804.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp.8003C44E683CA082.idx
 create mode 100755 .cache/clangd/index/device_grouped_conv_utils.hpp.82861A708A167591.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm.hpp.96241388E9D6D4CD.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp.B5F90340138DCBDD.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp.BC9F6A61DFACFE50.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.F7CCE132C2B5DC0C.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.9DF3794E443669AC.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_fixed_nk.hpp.6F8F8F18341308F9.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_softmax_gemm_permute.hpp.3C2429E11C534BCE.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.75AE4D40F988002C.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_splitk.hpp.6F08C0BD625D58A2.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl.hpp.46AD692E57980C90.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp.37D37DC699ECCA33.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp.1D6C6BE3C31CA169.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.24FD07A2963099BE.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.32B3D11C83B5AEA6.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk.hpp.4C6A715EB09F77E1.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp.47280D4E92EC5D1F.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp.1DAF74A63A4FD2D6.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp.CD8E0FEB6A298EEB.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp.C65F333685492D7B.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp.D0AB01FA41C26CA6.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp.43793441438DD042.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp.2F31492B9345223B.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp.A7598EAA8CFFD492.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp.FC712CF5F768A94B.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp.6FC4AF50AF38A41A.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp.B22CC0B7C42EE2B4.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp.E11E98EDBDF90D0B.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_cshuffle.hpp.CE155CFA093515AD.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp.5121BAECA46B0C23.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.1597C09D10EC33F9.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp.FBCAD14B8CE81D3F.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F03BAFFDB1120A94.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp.EADFE46F5C6C50DA.idx
 create mode 100755 .cache/clangd/index/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp.5656416FBF150D8D.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_bwd_data_f32_instance.cpp.6CBCB642104CF767.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_bwd_gamma_beta_f32_instance.cpp.59A23F12AA5E3659.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_fwd_f16_instance.cpp.E5229B139F1A9B82.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_fwd_f32_instance.cpp.D9D1D3F585AA6CD7.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp.4F6E8B2271D69F1C.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_fwd_swish_f16_instance.cpp.3151733ECD55CFAB.idx
 create mode 100755 .cache/clangd/index/device_groupnorm_fwd_swish_f32_instance.cpp.7C71C8C8383AE2D3.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_gndhwc_3d_instance.cpp.B02146942667831D.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_gnhwc_2d_instance.cpp.BAF41643A25FDB03.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_gnwc_1d_instance.cpp.3EF12405C93313FD.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_impl.hpp.C46D2CA10F5E5642.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_instance.hpp.94D5335C0371893F.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_ndhwgc_3d_instance.cpp.709987A54287EE7C.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_nhwgc_2d_instance.cpp.E16A024C295F734A.idx
 create mode 100755 .cache/clangd/index/device_image_to_column_nwgc_1d_instance.cpp.D56D4C0735B48980.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_bwd_data_f16_instance.cpp.7967A43F40615422.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_bwd_data_f32_instance.cpp.BCCFAE80AA603EB7.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp.756B8C951E2AA179.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp.4E9CDD5C09C49289.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_fwd_f16_instance.cpp.572C171AE3FAAC3B.idx
 create mode 100755 .cache/clangd/index/device_layernorm2d_fwd_f32_instance.cpp.C34D1A2542CE24BA.idx
 create mode 100755 .cache/clangd/index/device_layernorm4d_fwd_f16_instance.cpp.4D368D0AFAADE23F.idx
 create mode 100755 .cache/clangd/index/device_layernorm4d_fwd_f32_instance.cpp.82CA4512D672E7A2.idx
 create mode 100755 .cache/clangd/index/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp.59F686652075C1F3.idx
 create mode 100755 .cache/clangd/index/device_max_pool3d_fwd_ndhwc_f16_instance.cpp.075B2EC4AC466F2A.idx
 create mode 100755 .cache/clangd/index/device_max_pool3d_fwd_ndhwc_f32_instance.cpp.2CC50FB6EE948F2B.idx
 create mode 100755 .cache/clangd/index/device_max_pool_bwd.hpp.67E96E50742F2EC4.idx
 create mode 100755 .cache/clangd/index/device_max_pool_bwd_bf16_instance.cpp.142DD51C4CE34BA2.idx
 create mode 100755 .cache/clangd/index/device_max_pool_bwd_f16_instance.cpp.DCBDF5FDAE579133.idx
 create mode 100755 .cache/clangd/index/device_max_pool_bwd_f32_instance.cpp.EA3D33FB1230F655.idx
 create mode 100755 .cache/clangd/index/device_max_pool_bwd_impl.hpp.85737AA1C70FEFBE.idx
 create mode 100755 .cache/clangd/index/device_memory.cpp.A744015D851F9812.idx
 create mode 100755 .cache/clangd/index/device_memory.hpp.C2F2D98A33C55529.idx
 create mode 100755 .cache/clangd/index/device_multiple_reduce.hpp.188F02C49C64E773.idx
 create mode 100755 .cache/clangd/index/device_multiple_reduce_multiblock.hpp.BAAA7F55A7B0C61A.idx
 create mode 100755 .cache/clangd/index/device_multiple_reduce_threadwise.hpp.956BE9BDF24D2435.idx
 create mode 100755 .cache/clangd/index/device_normalization_bwd_data.hpp.1BAA9B467AEFC28F.idx
 create mode 100755 .cache/clangd/index/device_normalization_bwd_data_impl.hpp.A627060CAA69A19F.idx
 create mode 100755 .cache/clangd/index/device_normalization_bwd_gamma_beta.hpp.78751175103676B4.idx
 create mode 100755 .cache/clangd/index/device_normalization_bwd_gamma_beta_impl.hpp.A5BD375EAE722F5E.idx
 create mode 100755 .cache/clangd/index/device_normalization_fwd.hpp.DBDBAF89813C337F.idx
 create mode 100755 .cache/clangd/index/device_normalization_fwd_impl.hpp.2930CBBD3A2BF4E9.idx
 create mode 100755 .cache/clangd/index/device_normalization_fwd_splitk_impl.hpp.6FD53387232480CC.idx
 create mode 100755 .cache/clangd/index/device_normalize_instance.cpp.FD40702354DAF5AA.idx
 create mode 100755 .cache/clangd/index/device_operation_instance_factory.hpp.115B69A1EE76F903.idx
 create mode 100755 .cache/clangd/index/device_permute.hpp.47CCE84C97E4E9FC.idx
 create mode 100755 .cache/clangd/index/device_permute_impl.hpp.BEE7C34D1958CA2F.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_1d_instances.cpp.7EB32958B0512D73.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_2d_instances.cpp.2B80C1D2C455DED0.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_3d_instances.cpp.07A2A609833FE4D0.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_4d_instances.cpp.18341306CF0F5EFD.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_5d_instances.cpp.81367B36B3C16096.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_6d_instances.cpp.A19C7755BA5CCAF6.idx
 create mode 100755 .cache/clangd/index/device_permute_scale_instances.hpp.6002ED9F3C6C6BA1.idx
 create mode 100755 .cache/clangd/index/device_pool2d_fwd_nhwc_nhwc.hpp.8F42F3A58241D669.idx
 create mode 100755 .cache/clangd/index/device_pool3d_fwd_ndhwc_ndhwc.hpp.821043598893E6B4.idx
 create mode 100755 .cache/clangd/index/device_pool_fwd.hpp.048422E56031C936.idx
 create mode 100755 .cache/clangd/index/device_prop.hpp.51F1A892B5410B39.idx
 create mode 100755 .cache/clangd/index/device_put_element.hpp.50E2F2788BD7B2E7.idx
 create mode 100755 .cache/clangd/index/device_put_element_impl.hpp.CF1E04C6FC237429.idx
 create mode 100755 .cache/clangd/index/device_reduce.hpp.17BAD15A3D93B7FE.idx
 create mode 100755 .cache/clangd/index/device_reduce_common.hpp.CAE83A095739383D.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance.hpp.15BDAE27A6CA281A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise.hpp.C5FE82438250FD9A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.cpp.A1938EEEABE8B42A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.hpp.20070C9325736FAE.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp.3CFF0C60C7399DEC.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp.8C9F3AF79C3259FF.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp.39D9A35BA78EDFCA.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp.D3804A89BDDBD68A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.cpp.26EBD8A22938C32C.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.hpp.EFFF67751118917D.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.cpp.82123925B2E47DE5.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.hpp.1AFBE8401A0BB727.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp.4AE84C1155A0BD16.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp.E3EB62B3883D54D2.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp.116108BDF1BABC19.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp.9716CDC40BFB2514.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.cpp.F4D8C3EC38AC4A06.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.hpp.E4FC0C2AEABCC117.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.cpp.8D16D62A0EBE3E6E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.hpp.4404F6E1807E5BD4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.cpp.6312AF648C08A121.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.hpp.70C411CE7496D684.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp.BB89842592147883.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp.08DE17CE7004DBDC.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp.15C86B2B21F27130.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp.152CF392CC9366CD.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.cpp.07FADC8850641162.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.hpp.B81990FE8C5D8509.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp.562409E747DBF8A8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp.DF4FA5048878CC52.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp.F349F4E1672B2959.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp.1B257007DC8DF36B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.cpp.58C6C0B25D770B51.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.hpp.9E23A16BC92C20DB.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.cpp.0B95E75A805DE595.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.hpp.B0727F2EA98E1F00.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp.FA19AE81B0ABEEED.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp.7E4B9D2EA0AA55ED.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.cpp.B429C5C7F2FA95D4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.hpp.5BB1B8BA3E60EF71.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp.CFF7599DE0E79FBB.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp.46A8C8C2BE2BA19E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp.384AFFA95058603D.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp.54BEF21DD1A84A95.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.cpp.213C81B0FE4954D8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.hpp.496DCAD6F6453198.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp.13D835E23741A152.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp.2472648101515DFA.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp.11BCD7B99E3C6DB4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp.E5A6055D6ADB74AC.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.cpp.3646CB646CAA9902.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.hpp.58F762D7856C2DD8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.cpp.1EAEE023D4C34351.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.hpp.94217680DDE41BC8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp.5276760A8102F515.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp.52EC3DCBFC6F1222.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.cpp.BC429E8BCB688FC0.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.hpp.29384EA1BE892314.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp.C4B64A596D6DF817.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp.615A6DAAC49A37A5.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp.9FC33F8197771FF0.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp.CF7CEDDC624CFFAF.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.cpp.20716445BE6CBC5D.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.hpp.A0110584C215CD9C.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.cpp.D8BEF9441B9F4BC8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.hpp.8F13163D01C8AE8F.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_impl_common.hpp.F2201C19BAE9FFE8.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add.hpp.F65D86E6C84EDA60.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp.5E6D995E61ADA8BC.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp.510427928F98A02B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp.54085188C8447423.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp.0ADD6329411B78B1.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp.4393525F2F4E8DA5.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp.21E4CBB422BF8D82.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp.4BD601E4D7DE370C.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp.619ED16F95775BEA.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp.873C6FA8487AF602.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp.DA5C10F93175C08B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp.A62B76883CD2E814.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp.82487B9F10609357.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp.979C0C60ED1AF7B3.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp.16AAA28D16A62BE9.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp.3DCB7674F6E295A2.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp.3C8AE5F8E636FD38.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp.EC103046C3FCC222.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp.EC5F0E8B3A5B73CB.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp.F69FD80D45178C79.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp.F6C8C49B5460F242.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise.hpp.C398C657987C23EC.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.cpp.7EFDD1FE32B6F3E3.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.hpp.2D7F2A668C85C670.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp.DCF9F78C2573F20F.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp.E64E06F9E993015F.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp.71BA7AE5BF315EB4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp.5814F92CA6CF0198.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.cpp.D9E364199F88743A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.hpp.B8A4A04F978ABC79.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.cpp.B60965F1C97196EF.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.hpp.97362179A977A149.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp.C510348DB0E9EAF4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp.B82C759E6266A9D7.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp.6F6EC6F8F6066B26.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp.2B3A1DBDB134572E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.cpp.501586DF17B9409B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.hpp.EA77360039638CEE.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.cpp.F85FE38D0DCF8C44.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.hpp.D976B4CDF936A58A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.cpp.630A4073E2B46FEA.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.hpp.7BFEE24C60F74EC6.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp.BB154BAE452FBECD.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp.442706BD47327B2E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp.01982D48C410CB47.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp.D90C9C6069808C78.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.cpp.30899F9F46343B0E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.hpp.3C1359A00C377B18.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp.F787647EC7B1DE3C.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp.2ABD3DA274EED7C0.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp.0D0748D60784A612.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp.F51F72472293693C.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.cpp.35DDEC8F63798EAB.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.hpp.4E3093BB6320EF1D.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.cpp.1CE993853CF695D0.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.hpp.448A5294C15B52AE.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp.F71A95723F67A7AA.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp.1D44113CF201458F.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.cpp.DBE748926DFEDB8B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.hpp.55EEA55DB1346E7B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp.E1314404E465BA33.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp.D9B461C00E68A892.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp.C0DCB6010E545A4E.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp.082457B5EE500DB0.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.cpp.FF7D2E3E29E0477B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.hpp.774FB922ABF22CDD.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp.095F5D6B26D2F29A.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp.5FCEC88A65C41974.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp.BB5B12454481FF04.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp.7B5AB5E79B93DFFB.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.cpp.BAF3E66A0EDF3C75.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.hpp.390E8CBCF4E86ABF.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.cpp.778D73C35496348B.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.hpp.2C7338E3132300C6.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp.59008E97BEF36435.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp.4329A8B0483B68A4.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.cpp.509D27379633375F.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.hpp.7F575A4F4970CC92.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp.BBCEE5A5D01A00A5.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp.FDD16FEA381DDE62.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp.C6B6423A4D3C4156.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp.F8ED464DDA5BD732.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.cpp.C1E70E7984041F16.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.hpp.1F0CCA2641B90428.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.cpp.05E7FDBEF4E2FEA1.idx
 create mode 100755 .cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.hpp.078E8B6FD22E7DA4.idx
 create mode 100755 .cache/clangd/index/device_reduce_multiblock.hpp.C0B69E5EA2AC7CE7.idx
 create mode 100755 .cache/clangd/index/device_reduce_threadwise.hpp.71DBCBAA668DE14B.idx
 create mode 100755 .cache/clangd/index/device_softmax.hpp.497B752134CE80EA.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.cpp.6DE6521393D18F5C.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.hpp.B94CF584D85EABE4.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.cpp.6002EE43735BD2AD.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.hpp.453CC2F7EB0502D0.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.cpp.62DA65F400167AF3.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.hpp.3F8E078976E83385.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.cpp.22BB990A62BE50C5.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.hpp.7F7D38DA481813D7.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.cpp.AC6F1FB653149557.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.hpp.44F668BC92D2FB11.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.cpp.2BDC7BD673407442.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.hpp.1A5543D9F44A4D1E.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.cpp.E94DE85158A7059C.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.hpp.DFA0CA7449260515.idx
 create mode 100755 .cache/clangd/index/device_softmax_f16_f16_instance_type.hpp.D6CB9472D3A0B40A.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.cpp.7A9B6FF1C1DE5769.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.hpp.901763453949395E.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.cpp.B1019C659E1C387F.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.hpp.E7E5CB6E00C1ACF7.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.cpp.D8971A21BC00D0FF.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.hpp.5132DD2F1F4E6659.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.cpp.70DAADE138E1FD1C.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.hpp.D21E6E14C4987978.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.cpp.C8298BB69F4C257A.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.hpp.E0AE02E34E90B0FC.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.cpp.25E3FD604E588BD0.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.hpp.33AFA8695A15063D.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.cpp.86DA1C4D6D7DC8E7.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.hpp.35C1B364BF55F9FD.idx
 create mode 100755 .cache/clangd/index/device_softmax_f32_f32_instance_type.hpp.7F3C1693216AB251.idx
 create mode 100755 .cache/clangd/index/device_softmax_impl.hpp.B9E8182EF8FC3442.idx
 create mode 100755 .cache/clangd/index/device_softmax_instance.hpp.E665F0AD026D2FE2.idx
 create mode 100755 .cache/clangd/index/device_sparse_embeddings_forward_layernorm.hpp.469278C9F8399B0C.idx
 create mode 100755 .cache/clangd/index/device_splitk_contraction_multiple_d.hpp.B9F979311F947532.idx
 create mode 100755 .cache/clangd/index/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp.43569A70DF974D2E.idx
 create mode 100755 .cache/clangd/index/device_transpose_instance.hpp.9CF832E322CA1530.idx
 create mode 100755 .cache/clangd/index/device_transpose_instances_3d.cpp.5DC4547776B7C3A0.idx
 create mode 100755 .cache/clangd/index/dpp_gemm.hpp.CD386ABA025C1896.idx
 create mode 100755 .cache/clangd/index/dual_reduce_common.hpp.746AA107CF079C11.idx
 create mode 100755 .cache/clangd/index/dual_reduce_multiblock.cpp.C5E150B4E9B1E7E8.idx
 create mode 100755 .cache/clangd/index/dual_reduce_threadwise.cpp.7BA56606324C3914.idx
 create mode 100755 .cache/clangd/index/dynamic_buffer.hpp.9B076D16FB6FDAA3.idx
 create mode 100755 .cache/clangd/index/element_wise_operation.hpp.0715A40D7E0DC380.idx
 create mode 100755 .cache/clangd/index/elementwise_add_1d.cpp.CD0512697217A13B.idx
 create mode 100755 .cache/clangd/index/elementwise_add_4d.cpp.6D6ECABFC0DFD709.idx
 create mode 100755 .cache/clangd/index/elementwise_layernorm_blockwise.cpp.B1B3F8F277F6314A.idx
 create mode 100755 .cache/clangd/index/elementwise_normalization.hpp.B86CF9747F7CBAA9.idx
 create mode 100755 .cache/clangd/index/elementwise_permute.cpp.780590C9EF3E30CE.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_3d.cpp.2AC2BFBCE3237215.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp16.cpp.267AD0382A416744.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp16_2d.cpp.73F9F8FFA71E623F.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp16_col.cpp.427D8FEA072CC243.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp16_row.cpp.70A9A9EF321166D7.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp32_col.cpp.645D32DD67217BF3.idx
 create mode 100755 .cache/clangd/index/elementwise_permute_4D_fp32_row.cpp.51034D29A216280F.idx
 create mode 100755 .cache/clangd/index/enable_if.hpp.10016CB3379BC2AC.idx
 create mode 100755 .cache/clangd/index/f8_utils.hpp.AB5D9F32110A912A.idx
 create mode 100755 .cache/clangd/index/fill.hpp.A302158AF84698F8.idx
 create mode 100755 .cache/clangd/index/functional.hpp.AB02FF3D06BB8256.idx
 create mode 100755 .cache/clangd/index/functional2.hpp.4855B705916A8A54.idx
 create mode 100755 .cache/clangd/index/functional3.hpp.730E80629C119B9D.idx
 create mode 100755 .cache/clangd/index/functional4.hpp.579A0D5EFC68EB03.idx
 create mode 100755 .cache/clangd/index/gemm.hpp.344713A6C512A32C.idx
 create mode 100755 .cache/clangd/index/gemm.hpp.C45C6ACD321C3642.idx
 create mode 100755 .cache/clangd/index/gemm_add.hpp.0A05FF4E11D910E0.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu.hpp.C1858AFB6E6189D2.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu_xdl_bf16.cpp.D697A303BBCD00E1.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu_xdl_fp16.cpp.BE72E2A90F751444.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu_xdl_fp32.cpp.1D688E059DA5C6D3.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu_xdl_int8.cpp.7C822BDA45FD8B5C.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp.DFA0EFAE6F444F2C.idx
 create mode 100755 .cache/clangd/index/gemm_add_add_mean_meansquare_xdl_fp16.cpp.784D28B0B387AEB3.idx
 create mode 100755 .cache/clangd/index/gemm_add_addsquare_xdl_int8.cpp.C81A82E4BF27D67D.idx
 create mode 100755 .cache/clangd/index/gemm_add_fastgelu.hpp.7F025B27BB774C23.idx
 create mode 100755 .cache/clangd/index/gemm_add_multiply.hpp.A076E88A74402B66.idx
 create mode 100755 .cache/clangd/index/gemm_add_multiply_xdl_fp16.cpp.2190407D709E315C.idx
 create mode 100755 .cache/clangd/index/gemm_add_relu.hpp.2FD450825E8EDD32.idx
 create mode 100755 .cache/clangd/index/gemm_add_relu_add_layernorm.hpp.2503A6C1813DEBB3.idx
 create mode 100755 .cache/clangd/index/gemm_add_silu.hpp.8120675903700DB9.idx
 create mode 100755 .cache/clangd/index/gemm_bf16.cpp.B8DB90A206756A9E.idx
 create mode 100755 .cache/clangd/index/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp.ACF8DA35700D8D8F.idx
 create mode 100755 .cache/clangd/index/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp.83BB18F6B66F2289.idx
 create mode 100755 .cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp.6BD99D80213AD3B0.idx
 create mode 100755 .cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp.AC32388E298FFA30.idx
 create mode 100755 .cache/clangd/index/gemm_bias_relu_xdl_fp16.cpp.74D4B1F74A45E199.idx
 create mode 100755 .cache/clangd/index/gemm_bias_softmax_gemm_permute.cpp.A515B9EF146522D4.idx
 create mode 100755 .cache/clangd/index/gemm_bilinear.hpp.092CAB87B366F08F.idx
 create mode 100755 .cache/clangd/index/gemm_bilinear_xdl_fp16.cpp.372A779C3E21DD0E.idx
 create mode 100755 .cache/clangd/index/gemm_dpp_fp16.cpp.D49004015318A845.idx
 create mode 100755 .cache/clangd/index/gemm_f16_nn_instance.cpp.684811D08F35F1E0.idx
 create mode 100755 .cache/clangd/index/gemm_f16_nn_instance.hpp.295DF91B6CFA2E94.idx
 create mode 100755 .cache/clangd/index/gemm_f16_nt_instance.cpp.CF459381653536CF.idx
 create mode 100755 .cache/clangd/index/gemm_f16_nt_instance.hpp.CF5C609546474B1F.idx
 create mode 100755 .cache/clangd/index/gemm_f16_tn_instance.cpp.A27FD0EC622C1968.idx
 create mode 100755 .cache/clangd/index/gemm_f16_tn_instance.hpp.C57CA23FF5A33CBF.idx
 create mode 100755 .cache/clangd/index/gemm_f16_tt_instance.cpp.5D254C2DB3D84C02.idx
 create mode 100755 .cache/clangd/index/gemm_f16_tt_instance.hpp.A9C0C844390863CE.idx
 create mode 100755 .cache/clangd/index/gemm_fastgelu.hpp.BF1B6FA711DF924E.idx
 create mode 100755 .cache/clangd/index/gemm_fp16.cpp.3FD2C2DD621D7AD4.idx
 create mode 100755 .cache/clangd/index/gemm_fp32.cpp.6736B177964E2D4B.idx
 create mode 100755 .cache/clangd/index/gemm_int8.cpp.C4070EF85C2974E6.idx
 create mode 100755 .cache/clangd/index/gemm_layernorm_xdl_naive_fp16.cpp.B168DA32D72A32C6.idx
 create mode 100755 .cache/clangd/index/gemm_max_xdl_bf16.cpp.40A7D230DC3CAD48.idx
 create mode 100755 .cache/clangd/index/gemm_max_xdl_fp16.cpp.F50CF88BE688DE26.idx
 create mode 100755 .cache/clangd/index/gemm_max_xdl_fp32.cpp.1AFB0F6127AC986E.idx
 create mode 100755 .cache/clangd/index/gemm_max_xdl_int8.cpp.A29BA4DC035DE52E.idx
 create mode 100755 .cache/clangd/index/gemm_mean_meansquare_xdl_bf16.cpp.E1D95735FEB3AF62.idx
 create mode 100755 .cache/clangd/index/gemm_mean_meansquare_xdl_fp16.cpp.80F76169F90BD8DC.idx
 create mode 100755 .cache/clangd/index/gemm_mean_meansquare_xdl_fp32.cpp.DAB08B69CF521C93.idx
 create mode 100755 .cache/clangd/index/gemm_multi_ABD_xdl_fp16.cpp.5859AE2CEFBAACD5.idx
 create mode 100755 .cache/clangd/index/gemm_multiply_add.hpp.6B6D21EE52049C9D.idx
 create mode 100755 .cache/clangd/index/gemm_quantization_common.hpp.A23B3AA4CB2F3E7A.idx
 create mode 100755 .cache/clangd/index/gemm_reduce_fp16.cpp.EFAED30A7BA2F6AF.idx
 create mode 100755 .cache/clangd/index/gemm_reduce_xdl_common.hpp.BF32975909DCA917.idx
 create mode 100755 .cache/clangd/index/gemm_specialization.hpp.F3D9282B71057FD0.idx
 create mode 100755 .cache/clangd/index/gemm_splitk.hpp.749B3E67AC6B8B9C.idx
 create mode 100755 .cache/clangd/index/gemm_standalone_xdl_fp16.cpp.EE84FA2A4697A71B.idx
 create mode 100755 .cache/clangd/index/gemm_streamk.hpp.B73183768811D4AE.idx
 create mode 100755 .cache/clangd/index/gemm_util.hpp.1BFB1C62397B2C8C.idx
 create mode 100755 .cache/clangd/index/gemm_wavelet_f16_tn_instance.cpp.DC996A5B769C800C.idx
 create mode 100755 .cache/clangd/index/gemm_wavelet_f16_tn_instance.hpp.6DF980223599227D.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_bf16.cpp.3DCFB3ED2888403A.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_bf16_rtn.cpp.52661E2CDB7295F6.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_bias_relu_quantization_int8.cpp.FA91569EB26ACB71.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp16.cpp.FC8CAA78DACFED36.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp16_fp8.cpp.44C0E7F4DAEEA470.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp16_v2.cpp.BBCC41C35D9D878E.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp64.cpp.326BD4F598F31DB0.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp8.cpp.67C1EEC83062DA6B.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_fp8_bf8.cpp.AEAA96B6834BEA56.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_int8.cpp.636EBC0B4897BF3A.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp.4306E603B691D828.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_lds_direct_load_fp16.cpp.7BFA821D80385E72.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_lds_direct_load_fp32.cpp.39A0D92907A2D21F.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_quantization_int8.cpp.A58C21992E5540BA.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_skip_b_lds_fp16.cpp.86DD71A6DF5D193B.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_streamk.cpp.2BC4B9D993728CB3.idx
 create mode 100755 .cache/clangd/index/gemm_xdl_wavelet_fp16.cpp.5100C57848816EFC.idx
 create mode 100755 .cache/clangd/index/generic_memory_space_atomic.hpp.C9D4AE0A6DC5899F.idx
 create mode 100755 .cache/clangd/index/get_id.hpp.D8C96396D4A9E3B4.idx
 create mode 100755 .cache/clangd/index/get_shift.hpp.0A10E058B6B6AEDA.idx
 create mode 100755 .cache/clangd/index/gridwise_2d_multiple_reduction_multiblock.hpp.B962E3920F925D70.idx
 create mode 100755 .cache/clangd/index/gridwise_2d_multiple_reduction_threadwise.hpp.39A048A5A487AC02.idx
 create mode 100755 .cache/clangd/index/gridwise_2d_reduction_multiblock.hpp.6BD3600B2B4E50BE.idx
 create mode 100755 .cache/clangd/index/gridwise_2d_reduction_threadwise.hpp.324930ADA73F028A.idx
 create mode 100755 .cache/clangd/index/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp.E109D2D6AEE43DDD.idx
 create mode 100755 .cache/clangd/index/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp.FE93431287EC85FD.idx
 create mode 100755 .cache/clangd/index/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp.A946F68C0A5B7C29.idx
 create mode 100755 .cache/clangd/index/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp.EEF70E721362F483.idx
 create mode 100755 .cache/clangd/index/gridwise_batchnorm_backward_blockwise_welford.hpp.8F24BDC37125B404.idx
 create mode 100755 .cache/clangd/index/gridwise_batchnorm_forward_blockwise_welford.hpp.640DA951EF640017.idx
 create mode 100755 .cache/clangd/index/gridwise_elementwise_1d.hpp.AC0144CD925F6905.idx
 create mode 100755 .cache/clangd/index/gridwise_elementwise_1d_scale.hpp.CF7CE1061BA1DF52.idx
 create mode 100755 .cache/clangd/index/gridwise_elementwise_2d.hpp.DCACD2865A1A82CC.idx
 create mode 100755 .cache/clangd/index/gridwise_elementwise_3d.hpp.4B57404BC7344584.idx
 create mode 100755 .cache/clangd/index/gridwise_elementwise_layernorm_welford_variance.hpp.61A857F748242277.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp.6AEBEC03A02FE520.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_dpp.hpp.E915886DCA1D5AA3.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp.8ED0C5830B90D078.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.E96B152CEA3EEC31.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp.75601854417950D1.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_wmma_cshuffle.hpp.7EFDFD3F43E2C561.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle.hpp.7B2A45106BD10E35.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.8A2E3D6921803C3C.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp.79ABECDBF9253649.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_pipeline_selector.hpp.4F4E408D810A020A.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_pipeline_v1.hpp.8B594F4BBFC9415D.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_pipeline_v2.hpp.1C7FE339CB47DFD5.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_pipeline_v3.hpp.D2EF7B80D9AE44CD.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_pipeline_v4_direct_load.hpp.7C3BB633588D635D.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp.76D93CEF02B94AEF.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp.F83027A680F1BF3E.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_waveletmodel.hpp.C9F027882C9D4F2B.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_wmma.hpp.4F8B2CDD1348D704.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdl_cshuffle_v1.hpp.E796A6FD7BCB2B56.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdl_cshuffle_v2.hpp.E0B3C0FB7BAA0FB3.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp.FC9F72A612CD78AC.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp.647408B24FB9D296.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_bwd_weight.hpp.65AB95ECB27BA1C2.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_skip_b_lds_v1.hpp.8CF2D2B713EFE65F.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp.76232B181B348ACC.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_streamk.hpp.3A8F1B585329CAFC.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_v2r3.hpp.66B1A28D309FB5D4.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_v2r4r2.hpp.FFFD55E6D49B5037.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_v3r1.hpp.F4E339DE2B052B4F.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_v3r2.hpp.A91ECFDCE099EED9.idx
 create mode 100755 .cache/clangd/index/gridwise_gemm_xdlops_v3r3.hpp.E0F0A6CCA772EDE7.idx
 create mode 100755 .cache/clangd/index/gridwise_multiblock_batchnorm_forward.hpp.4F73C2A5286FBB35.idx
 create mode 100755 .cache/clangd/index/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp.C7E55A256926B340.idx
 create mode 100755 .cache/clangd/index/gridwise_multiblock_welford_first_half.hpp.25F962D583FA62A6.idx
 create mode 100755 .cache/clangd/index/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp.BEA9938289B2ECCF.idx
 create mode 100755 .cache/clangd/index/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp.03EA80728C2C5A38.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_bwd_data.hpp.43B5DE0C3D25BB20.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_bwd_gamma_beta.hpp.8028572B7299488D.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_naive_variance.hpp.9BD4CEF549374198.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_selector.hpp.88C73ECDE0BC86D2.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_splitk_1st.hpp.6CD10C30DDA5197D.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_splitk_2nd.hpp.FE37C91CA5D8EB00.idx
 create mode 100755 .cache/clangd/index/gridwise_normalization_welford_variance.hpp.FE77CD9DBE7C3CF0.idx
 create mode 100755 .cache/clangd/index/gridwise_permute.hpp.E425D88E25D439BC.idx
 create mode 100755 .cache/clangd/index/gridwise_put_element_1d.hpp.59E84F53787BFD31.idx
 create mode 100755 .cache/clangd/index/gridwise_set_buffer_value.hpp.1407258F85A006CD.idx
 create mode 100755 .cache/clangd/index/gridwise_set_multiple_buffer_value.hpp.77B0D514858D3743.idx
 create mode 100755 .cache/clangd/index/gridwise_softmax.hpp.5A728FA36C708E60.idx
 create mode 100755 .cache/clangd/index/gridwise_sparse_embeddings_forward_layernorm.hpp.9F9F7C52CFA18397.idx
 create mode 100755 .cache/clangd/index/gridwise_tensor_rearrange.hpp.73B250CAB9F4E6DB.idx
 create mode 100755 .cache/clangd/index/gridwise_welford_second_half_layernorm2d.hpp.52F6FAA1A972E598.idx
 create mode 100755 .cache/clangd/index/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp.2CC919AB5B044CF3.idx
 create mode 100755 .cache/clangd/index/grouped_conv_bwd_data_xdl_fp16.cpp.618FF0F87C438165.idx
 create mode 100755 .cache/clangd/index/grouped_conv_bwd_weight_xdl_bf16.cpp.3B512EE28C055424.idx
 create mode 100755 .cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16.cpp.35241A5239C4258D.idx
 create mode 100755 .cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp.BD38E20DB8B3998F.idx
 create mode 100755 .cache/clangd/index/grouped_conv_conv_fwd_xdl_bf16.cpp.2A5915DD526E33D4.idx
 create mode 100755 .cache/clangd/index/grouped_conv_conv_fwd_xdl_fp16.cpp.F0393DD4FF0F531D.idx
 create mode 100755 .cache/clangd/index/grouped_conv_conv_fwd_xdl_fp32.cpp.329BE83F917399A6.idx
 create mode 100755 .cache/clangd/index/grouped_conv_conv_fwd_xdl_int8.cpp.E847303767D0E187.idx
 create mode 100755 .cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp.E2C08E4AF953E7F0.idx
 create mode 100755 .cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp.9CA386C0ABEF35F6.idx
 create mode 100755 .cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp.77C435750269A2BB.idx
 create mode 100755 .cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp.8BC00DE6D83E6AA0.idx
 create mode 100755 .cache/clangd/index/grouped_conv_fwd_xdl_fp16.cpp.F004628471E9028F.idx
 create mode 100755 .cache/clangd/index/grouped_convolution_backward_data.hpp.04038859D7920D2D.idx
 create mode 100755 .cache/clangd/index/grouped_convolution_backward_weight.hpp.CF4258D0B5FB5557.idx
 create mode 100755 .cache/clangd/index/grouped_convolution_forward.hpp.3DF8CF239601D4CC.idx
 create mode 100755 .cache/clangd/index/grouped_gemm.hpp.76EFEC97F28430A5.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_bias_e_permute_xdl_fp16.cpp.D04ED4238574C098.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_fastgelu.hpp.81337D2196A77665.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_fixed_nk.hpp.92D3931E2C242798.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.E01D0ADA6F0061BF.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.9FDD703AE6D2D15E.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_bf16.cpp.60909DDCF9BB055C.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp.3B6907AB640AF24B.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp16.cpp.C49639C51E0DF485.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp8.cpp.33128DEF03F0DD92.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_fp16.cpp.164525B4D87B1687.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_fp32.cpp.F97F080B232E6F17.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_int8.cpp.01E8A20D5ECCA353.idx
 create mode 100755 .cache/clangd/index/grouped_gemm_xdl_splitk_fp16.cpp.5CCCCCFEB9008060.idx
 create mode 100755 .cache/clangd/index/groupnorm_bwd_data.hpp.3714A9EE70464716.idx
 create mode 100755 .cache/clangd/index/groupnorm_bwd_fp32.cpp.94FC313C790AEF66.idx
 create mode 100755 .cache/clangd/index/groupnorm_bwd_gamma_beta.hpp.A4BF65E232956D79.idx
 create mode 100755 .cache/clangd/index/groupnorm_fwd_sigmoid_mul_fp16.cpp.E1BDC41DAA5F65CC.idx
 create mode 100755 .cache/clangd/index/groupnorm_fwd_splitk_fp16.cpp.5CDEA132226AE994.idx
 create mode 100755 .cache/clangd/index/groupnorm_fwd_swish_fp16.cpp.FDEC9D120BB249C9.idx
 create mode 100755 .cache/clangd/index/hip_check_error.hpp.B6401497C46CBC5E.idx
 create mode 100755 .cache/clangd/index/host_common_util.hpp.DD6CA1FC5D70D95B.idx
 create mode 100755 .cache/clangd/index/host_tensor.cpp.BD60013DE8F91330.idx
 create mode 100755 .cache/clangd/index/host_tensor.hpp.A2A2489B4F3A4E06.idx
 create mode 100755 .cache/clangd/index/host_tensor_generator.hpp.44F3F167E8AE0CAB.idx
 create mode 100755 .cache/clangd/index/ignore.hpp.67D37926F4126F70.idx
 create mode 100755 .cache/clangd/index/image_to_column_f32.cpp.E393FC8829D80520.idx
 create mode 100755 .cache/clangd/index/inner_product.hpp.D15AD8CC8FA359B4.idx
 create mode 100755 .cache/clangd/index/inner_product_dpp8.hpp.5615826ED9AC4030.idx
 create mode 100755 .cache/clangd/index/integral_constant.hpp.A156C83DCD53E7CE.idx
 create mode 100755 .cache/clangd/index/io.hpp.27CD569BA11007AA.idx
 create mode 100755 .cache/clangd/index/is_detected.hpp.993129C9FD0FB446.idx
 create mode 100755 .cache/clangd/index/is_known_at_compile_time.hpp.C66FF603A6D2FF7E.idx
 create mode 100755 .cache/clangd/index/iterator.hpp.DDB0074CE39B229F.idx
 create mode 100755 .cache/clangd/index/kernel_launch.hpp.4FE4793F82580CE2.idx
 create mode 100755 .cache/clangd/index/kernel_utils.hpp.A2D72F83F06BC2A4.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_add_instance.cpp.5CD53C62DE24ED15.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_default_pipeline_v1_instance.cpp.2F582983A57414AF.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_default_pipeline_v2_instance.cpp.FDC41D8D362237D7.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_default_pipeline_v2_opt_instance.cpp.9D0F889F526341B0.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_interwave_pipeline_v1_instance.cpp.536F114F5BA9349C.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_irregular_default_pipeline_v1_instance.cpp.5937911B0257B674.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_irregular_default_pipeline_v2_instance.cpp.0318E39A03EE883F.idx
 create mode 100755 .cache/clangd/index/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.0A7ACF40DAB7236A.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_add_instance.cpp.09C78F4C6D71CC59.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_default_pipeline_v1_instance.cpp.7011152E4CED8319.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_default_pipeline_v2_instance.cpp.4ADB4E587F42B7E9.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_default_pipeline_v2_opt_instance.cpp.AA43A9C5EED3E40A.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_interwave_pipeline_v1_instance.cpp.37A32BC5170A6461.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_irregular_default_pipeline_v1_instance.cpp.322AFE9D596702FC.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_irregular_default_pipeline_v2_instance.cpp.6202EDA536A25A77.idx
 create mode 100755 .cache/clangd/index/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.667185EB457E88CA.idx
 create mode 100755 .cache/clangd/index/layernorm2d_bwd_fp32.cpp.7066935C4C40FCB0.idx
 create mode 100755 .cache/clangd/index/layernorm2d_fwd_fp16.cpp.6079EB647B2C4778.idx
 create mode 100755 .cache/clangd/index/layernorm2d_fwd_splitk_fp16.cpp.7A8CCC9ADABCC2A0.idx
 create mode 100755 .cache/clangd/index/layernorm4d_fwd_fp16.cpp.FF678597E26B24C3.idx
 create mode 100755 .cache/clangd/index/layernorm4d_fwd_splitk_fp16.cpp.BF340D61393FB17F.idx
 create mode 100755 .cache/clangd/index/layernorm_bwd_data.hpp.D81DD3681C23DD68.idx
 create mode 100755 .cache/clangd/index/layernorm_bwd_gamma_beta.hpp.8CC35A2FBD2934F9.idx
 create mode 100755 .cache/clangd/index/layout.hpp.08D624D3F8E5BFE6.idx
 create mode 100755 .cache/clangd/index/layout_utils.hpp.9A8926719F7B5577.idx
 create mode 100755 .cache/clangd/index/literals.hpp.AC03AD34F59B947A.idx
 create mode 100755 .cache/clangd/index/loop_scheduler.hpp.CEBBF3BFF5C82A80.idx
 create mode 100755 .cache/clangd/index/magic_division.hpp.7DE3E22991AC58AE.idx
 create mode 100755 .cache/clangd/index/magic_number_division.cpp.853FDB10989ECFC7.idx
 create mode 100755 .cache/clangd/index/masking_specialization.hpp.2AE80C7347AB00C4.idx
 create mode 100755 .cache/clangd/index/math.hpp.234B356D49034450.idx
 create mode 100755 .cache/clangd/index/math_v2.hpp.562B4AF9F5A86873.idx
 create mode 100755 .cache/clangd/index/matrix_padder.hpp.0539595BAF18EE86.idx
 create mode 100755 .cache/clangd/index/max_pool_bwd.hpp.4F15616E2BF6E596.idx
 create mode 100755 .cache/clangd/index/max_pool_bwd_instance_common.hpp.F8135D216BA19663.idx
 create mode 100755 .cache/clangd/index/maxpool2d_bwd_bf16.cpp.E133A720E87AC796.idx
 create mode 100755 .cache/clangd/index/maxpool2d_bwd_common.hpp.4213849B15309F3C.idx
 create mode 100755 .cache/clangd/index/maxpool2d_bwd_fp16.cpp.AA5B517C73C2030C.idx
 create mode 100755 .cache/clangd/index/maxpool2d_bwd_fp32.cpp.3CE2CB1F186065DE.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_add_instance.cpp.F2B6A1DC94E17C44.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_default_pipeline_v1_instance.cpp.B364344C594966A7.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_default_pipeline_v2_instance.cpp.A590A56B0084D88A.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_default_pipeline_v2_opt_instance.cpp.7FA7EE3992571786.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_interwave_pipeline_v1_instance.cpp.E84C4EA0690C6426.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp.1D02066DCD48924F.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp.DFAE93983C514BC6.idx
 create mode 100755 .cache/clangd/index/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.905D27CAAE2E6FF3.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_add_instance.cpp.0D22D3DFBDE4264A.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_default_pipeline_v1_instance.cpp.FC2B955D88E3500C.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_default_pipeline_v2_instance.cpp.AB0F8B7DC8CEF96B.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_default_pipeline_v2_opt_instance.cpp.38A404E476FFFA7A.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_interwave_pipeline_v1_instance.cpp.1B80E7B3E985B173.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp.8AF96AFDCF49C20F.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp.AD25F4CBEC60DDF2.idx
 create mode 100755 .cache/clangd/index/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.EB9B1319190E914D.idx
 create mode 100755 .cache/clangd/index/multi_index.hpp.46D763C8845699EA.idx
 create mode 100755 .cache/clangd/index/multi_index_transform.hpp.97E44055DBD0AE25.idx
 create mode 100755 .cache/clangd/index/multi_index_transform_helper.hpp.7CC4F9B71E8A7CAA.idx
 create mode 100755 .cache/clangd/index/normalization_bwd_data_instance_common.hpp.3AFBFA5B5298FD17.idx
 create mode 100755 .cache/clangd/index/normalization_bwd_gamma_beta_instance_common.hpp.84C16E4EE625B892.idx
 create mode 100755 .cache/clangd/index/normalization_fwd.hpp.C0D3E8108CC38C5D.idx
 create mode 100755 .cache/clangd/index/normalization_fwd_instance_common.hpp.38A39908E71BFB61.idx
 create mode 100755 .cache/clangd/index/number.hpp.25A61C6490A403AE.idx
 create mode 100755 .cache/clangd/index/numeric.hpp.E1739325ACADC629.idx
 create mode 100755 .cache/clangd/index/permute_1xHxW_fp16.cpp.5780304278E72678.idx
 create mode 100755 .cache/clangd/index/permute_HxWx4_fp16.cpp.FCA2351A6D6E3584.idx
 create mode 100755 .cache/clangd/index/permute_NxHxW_fp16.cpp.6FE6AEAF71058087.idx
 create mode 100755 .cache/clangd/index/permute_scale.hpp.CFE8A4ABEDAA7993.idx
 create mode 100755 .cache/clangd/index/pool2d_fwd_common.hpp.3724B9CBDE79A373.idx
 create mode 100755 .cache/clangd/index/pool2d_fwd_fp16.cpp.09424A28FF97D19C.idx
 create mode 100755 .cache/clangd/index/pool2d_fwd_fp32.cpp.1AF35A327BA2145F.idx
 create mode 100755 .cache/clangd/index/pool3d_fwd.hpp.32DAF50D80F54A30.idx
 create mode 100755 .cache/clangd/index/pool3d_fwd_common.hpp.EE0EEEB8E82E75A0.idx
 create mode 100755 .cache/clangd/index/pool3d_fwd_fp16.cpp.176B982A3BF061DE.idx
 create mode 100755 .cache/clangd/index/pool_fwd_instance_common.hpp.4DC931E3BBF2D1BA.idx
 create mode 100755 .cache/clangd/index/profile_avg_pool3d_bwd.cpp.4403F8C1766CA1E8.idx
 create mode 100755 .cache/clangd/index/profile_avg_pool3d_bwd_impl.hpp.C8240F2385534214.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm.cpp.8A238D0EF3B6DB52.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_add_relu_gemm_add.cpp.CEB2ADB0FF2D0A07.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_add_relu_gemm_add_impl.hpp.CC900C93F3C48179.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp.F57E6277DAC86E29.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_gemm.cpp.B214B5448CC6B6C2.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_gemm_impl.hpp.95E513E815A63274.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_impl.hpp.65F140203E93B240.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_reduce.cpp.EEA3A1DAF90E2787.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_reduce_impl.hpp.5DA2586197AB1178.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_softmax_gemm_impl.hpp.03AF7B0A8D358951.idx
 create mode 100755 .cache/clangd/index/profile_batched_gemm_softmax_gemm_permute_impl.hpp.14377DE4299D9A14.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_backward_impl.hpp.30164435038F0711.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_bwd.cpp.8CF98960A7C44AE9.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_forward_impl.hpp.07F6CF6EEC2BDC7E.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_fwd.cpp.493E9D388D70238E.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_infer.cpp.69EB815355931680.idx
 create mode 100755 .cache/clangd/index/profile_batchnorm_infer_impl.hpp.B4DE6B04DAFD57EF.idx
 create mode 100755 .cache/clangd/index/profile_contraction_bilinear.cpp.B7E7F8EC02159FA8.idx
 create mode 100755 .cache/clangd/index/profile_contraction_impl.hpp.E5FF1E81EBC1E7D5.idx
 create mode 100755 .cache/clangd/index/profile_contraction_scale.cpp.C10C8FEB4E77E97A.idx
 create mode 100755 .cache/clangd/index/profile_contraction_utils.hpp.E0F0B9314CF06E36.idx
 create mode 100755 .cache/clangd/index/profile_conv_bwd_data.cpp.E30D0E66DFF68BCC.idx
 create mode 100755 .cache/clangd/index/profile_conv_bwd_data_impl.hpp.0CBAC32BD7F22884.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd.cpp.7612DDA826C4C111.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd_bias_relu.cpp.2AED2F5E417190C4.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd_bias_relu_add.cpp.064478D11C1695BD.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd_bias_relu_add_impl.hpp.660B8209627A1FBE.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd_bias_relu_impl.hpp.0A6D3A36AE3BBBB6.idx
 create mode 100755 .cache/clangd/index/profile_conv_fwd_impl.hpp.DB77593E3FC3FC2B.idx
 create mode 100755 .cache/clangd/index/profile_conv_tensor_rearrange.cpp.50518BDD3D0D27F8.idx
 create mode 100755 .cache/clangd/index/profile_conv_tensor_rearrange_impl.hpp.ECD3F7A8AA337DB3.idx
 create mode 100755 .cache/clangd/index/profile_elementwise_layernorm_impl.hpp.228FBBB77B1C2212.idx
 create mode 100755 .cache/clangd/index/profile_gemm.cpp.1660D1BAA0C70B61.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add.cpp.B4A511D84B31E476.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_add_fastgelu.cpp.374CFB86EAD905D5.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_add_fastgelu_impl.hpp.7165E9E08D31883A.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_fastgelu.cpp.CA0713C21D4F7C04.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_fastgelu_impl.hpp.23CA5590877F722E.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_impl.hpp.A44323833CE8C5AD.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_multiply.cpp.136C377C7D5B0E54.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_multiply_impl.hpp.FBBEFEDC4A78D751.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_relu.cpp.7B9D4CAF29075E5F.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_relu_add_layernorm.cpp.E30A7F678D4C6FAC.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_relu_add_layernorm_impl.hpp.5DE61D830E8EC4CC.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_relu_impl.hpp.9153E795377986F0.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_silu.cpp.67ABB3D17F5C92B4.idx
 create mode 100755 .cache/clangd/index/profile_gemm_add_silu_impl.hpp.CC18A7C19DDCC841.idx
 create mode 100755 .cache/clangd/index/profile_gemm_bias_add_reduce.cpp.CF67B6B3D02F20A5.idx
 create mode 100755 .cache/clangd/index/profile_gemm_bias_add_reduce_impl.hpp.CC4023F2075FFC72.idx
 create mode 100755 .cache/clangd/index/profile_gemm_bilinear.cpp.2A04608A0019808D.idx
 create mode 100755 .cache/clangd/index/profile_gemm_bilinear_impl.hpp.3D87C8E5D761FA48.idx
 create mode 100755 .cache/clangd/index/profile_gemm_fastgelu.cpp.2EC621FE126AC880.idx
 create mode 100755 .cache/clangd/index/profile_gemm_fastgelu_impl.hpp.090E8EEC37D7278D.idx
 create mode 100755 .cache/clangd/index/profile_gemm_impl.hpp.DB85F2B55BD1CC04.idx
 create mode 100755 .cache/clangd/index/profile_gemm_multiply_add.cpp.0030E28460A11C1F.idx
 create mode 100755 .cache/clangd/index/profile_gemm_multiply_add_impl.hpp.A03C0417F818F4FD.idx
 create mode 100755 .cache/clangd/index/profile_gemm_reduce.cpp.0A54AE6B66160321.idx
 create mode 100755 .cache/clangd/index/profile_gemm_reduce_impl.hpp.FE75615CA769411C.idx
 create mode 100755 .cache/clangd/index/profile_gemm_splitk.cpp.6BF00CBB81DE4660.idx
 create mode 100755 .cache/clangd/index/profile_gemm_splitk_impl.hpp.1DC09EFBBC454E8F.idx
 create mode 100755 .cache/clangd/index/profile_gemm_streamk.cpp.46F4E02E0912E672.idx
 create mode 100755 .cache/clangd/index/profile_gemm_streamk_impl.hpp.73FA01C66CC8D3E3.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_bwd_data.cpp.FA67D9F6149A96BE.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_bwd_data_impl.hpp.59E9CBB730B20A37.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_bwd_weight.cpp.34186A38546F70D1.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_bwd_weight_impl.hpp.A8CF2351C95F34F5.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_fwd.cpp.8D8BCE0560F621C4.idx
 create mode 100755 .cache/clangd/index/profile_grouped_conv_fwd_impl.hpp.90847DB78EF4B739.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm.cpp.7EE7588374346C13.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm_fastgelu.cpp.AC58B27586C8FEAD.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm_fastgelu_impl.hpp.DE8318589739E396.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm_fixed_nk.cpp.12EAC200C0F64A96.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm_fixed_nk_impl.hpp.B583CC73B5188079.idx
 create mode 100755 .cache/clangd/index/profile_grouped_gemm_impl.hpp.0459F4415131A8AB.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_bwd_data.cpp.B5C8D8F20885019C.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_bwd_data_impl.hpp.05BE2AB219AD2046.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_bwd_gamma_beta.cpp.0D9708552C2EDA92.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_bwd_gamma_beta_impl.hpp.F912A6BA3766E9B6.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_fwd.cpp.B71B17B33AF365C7.idx
 create mode 100755 .cache/clangd/index/profile_groupnorm_fwd_impl.hpp.DD036B9EC45A3789.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_bwd_data.cpp.EDB022306AF7F904.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_bwd_data_impl.hpp.CF00FB7A122475B2.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_bwd_gamma_beta.cpp.58ACE0ADA4571F63.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_bwd_gamma_beta_impl.hpp.672AE63374836A4E.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_fwd.cpp.FA9A3EA5A9ABAF4C.idx
 create mode 100755 .cache/clangd/index/profile_layernorm_fwd_impl.hpp.BF1975AA887FB0A6.idx
 create mode 100755 .cache/clangd/index/profile_max_pool3d_bwd.cpp.7793240AA812E755.idx
 create mode 100755 .cache/clangd/index/profile_max_pool3d_bwd_impl.hpp.885D2D1066A38A2E.idx
 create mode 100755 .cache/clangd/index/profile_max_pool3d_fwd.cpp.11655872CD06FCBE.idx
 create mode 100755 .cache/clangd/index/profile_permute_scale.cpp.A23434A32C62DC3B.idx
 create mode 100755 .cache/clangd/index/profile_permute_scale_impl.hpp.9398D48C7353608C.idx
 create mode 100755 .cache/clangd/index/profile_pool3d_fwd_impl.hpp.4E7051C23FCACC74.idx
 create mode 100755 .cache/clangd/index/profile_reduce.cpp.7CFAAB4D3A545696.idx
 create mode 100755 .cache/clangd/index/profile_reduce_impl.hpp.17B63EE6A3A6D55C.idx
 create mode 100755 .cache/clangd/index/profile_softmax.cpp.31A9AF77C4279825.idx
 create mode 100755 .cache/clangd/index/profile_softmax_impl.hpp.28DC8A49A2D3559F.idx
 create mode 100755 .cache/clangd/index/profile_transpose.cpp.3798E55CF7D973A7.idx
 create mode 100755 .cache/clangd/index/profile_transpose_impl.hpp.8CE368E3C650AFE1.idx
 create mode 100755 .cache/clangd/index/profiler.cpp.A1998E185346D702.idx
 create mode 100755 .cache/clangd/index/profiler_operation_registry.hpp.C3738D3851333A5F.idx
 create mode 100755 .cache/clangd/index/put_element_fp16.cpp.4CB3E91BA1C3425F.idx
 create mode 100755 .cache/clangd/index/quantization_operation.hpp.BD94143B871C7D8C.idx
 create mode 100755 .cache/clangd/index/random_gen.hpp.B234A29FCA71839D.idx
 create mode 100755 .cache/clangd/index/ranges.hpp.66B5D2199BD23CD4.idx
 create mode 100755 .cache/clangd/index/reduce.hpp.7DC5A811A95F14CB.idx
 create mode 100755 .cache/clangd/index/reduce_blockwise.cpp.E231BF6C564BCCFD.idx
 create mode 100755 .cache/clangd/index/reduce_blockwise_impl.hpp.B38B56E3F9BFC70D.idx
 create mode 100755 .cache/clangd/index/reduce_blockwise_two_call.cpp.C62D3A32F608E374.idx
 create mode 100755 .cache/clangd/index/reduce_example_common.hpp.A3D8D5EB42288D30.idx
 create mode 100755 .cache/clangd/index/reduce_multiblock_atomic_add.cpp.22DBB7B98E8ABAE4.idx
 create mode 100755 .cache/clangd/index/reduce_multiblock_atomic_add_impl.hpp.014F022BC601B7F9.idx
 create mode 100755 .cache/clangd/index/reduce_no_index.cpp.3EDCFAA9D727E5B6.idx
 create mode 100755 .cache/clangd/index/reduce_with_index.cpp.F84EAB2F5A59285D.idx
 create mode 100755 .cache/clangd/index/reduction_common.hpp.C00F2970D4BF7FA3.idx
 create mode 100755 .cache/clangd/index/reduction_enums.hpp.871AE7D90DC9E069.idx
 create mode 100755 .cache/clangd/index/reduction_functions_accumulate.hpp.7D2C2CFC245627F0.idx
 create mode 100755 .cache/clangd/index/reduction_functions_blockwise.hpp.C958E7F57DF3365E.idx
 create mode 100755 .cache/clangd/index/reduction_functions_threadwise.hpp.7A06A2856650B185.idx
 create mode 100755 .cache/clangd/index/reduction_operator.hpp.28F266F9F2F32A7B.idx
 create mode 100755 .cache/clangd/index/reduction_operator_mapping.hpp.8A11EE4B065A6CAE.idx
 create mode 100755 .cache/clangd/index/reference_avgpool_bwd.hpp.EF9EF399DB15462F.idx
 create mode 100755 .cache/clangd/index/reference_batched_gemm.hpp.AF38237B103D733E.idx
 create mode 100755 .cache/clangd/index/reference_batchnorm_backward.hpp.DB836B16A0EC29D6.idx
 create mode 100755 .cache/clangd/index/reference_batchnorm_forward.hpp.0421D3C13A4F2108.idx
 create mode 100755 .cache/clangd/index/reference_batchnorm_infer.hpp.B110A1DBC8883C05.idx
 create mode 100755 .cache/clangd/index/reference_cgemm.hpp.9B0EE32433381722.idx
 create mode 100755 .cache/clangd/index/reference_column_to_image.hpp.B1B86724AE64837E.idx
 create mode 100755 .cache/clangd/index/reference_contraction.hpp.4808E834C3E32AEB.idx
 create mode 100755 .cache/clangd/index/reference_conv_bwd_data.hpp.7B56299251FBE761.idx
 create mode 100755 .cache/clangd/index/reference_conv_bwd_weight.hpp.474E92061BE55B30.idx
 create mode 100755 .cache/clangd/index/reference_conv_fwd.cpp.3703F6A90F557994.idx
 create mode 100755 .cache/clangd/index/reference_conv_fwd.hpp.4266D54C9E6D337C.idx
 create mode 100755 .cache/clangd/index/reference_conv_fwd_bias_activation.hpp.CC5C1A67067C17BD.idx
 create mode 100755 .cache/clangd/index/reference_conv_fwd_bias_activation_add.hpp.05EAEEBBFD920597.idx
 create mode 100755 .cache/clangd/index/reference_gemm.hpp.0A74F0F3AFBAC0AB.idx
 create mode 100755 .cache/clangd/index/reference_gemm_layernorm.hpp.9AA4C354D9D24BEC.idx
 create mode 100755 .cache/clangd/index/reference_groupnorm.hpp.368D3290BA70398F.idx
 create mode 100755 .cache/clangd/index/reference_groupnorm_bwd.hpp.C24505C3C5BFB298.idx
 create mode 100755 .cache/clangd/index/reference_image_to_column.hpp.14F89CCD7031E508.idx
 create mode 100755 .cache/clangd/index/reference_layernorm.hpp.B9C72CF73E810635.idx
 create mode 100755 .cache/clangd/index/reference_layernorm_bwd.hpp.F0026FDE37931C67.idx
 create mode 100755 .cache/clangd/index/reference_maxpool_bwd.hpp.AA4F2921ED8F7A67.idx
 create mode 100755 .cache/clangd/index/reference_pool_fwd.hpp.3BB9B8E57191E80A.idx
 create mode 100755 .cache/clangd/index/reference_reduce.hpp.0C04C3B5F65B2DFD.idx
 create mode 100755 .cache/clangd/index/reference_softmax.hpp.4C7847F5A451A485.idx
 create mode 100755 .cache/clangd/index/reference_sparse_embedding3_forward_layernorm.hpp.495E2A0ED186B796.idx
 create mode 100755 .cache/clangd/index/run_batched_gemm_example.inc.341E18570766CD6C.idx
 create mode 100755 .cache/clangd/index/run_batched_gemm_gemm_example.inc.B0C303C33F007446.idx
 create mode 100755 .cache/clangd/index/run_batched_gemm_scale_softmax_gemm.inc.35534F29FB76E4C9.idx
 create mode 100755 .cache/clangd/index/run_batched_gemm_scale_softmax_gemm_permute.inc.DF3E33E757137FA7.idx
 create mode 100755 .cache/clangd/index/run_contraction_bilinear_example.inc.427C55016011CF17.idx
 create mode 100755 .cache/clangd/index/run_contraction_scale_example.inc.619AD6EEB2674ADA.idx
 create mode 100755 .cache/clangd/index/run_conv2d_fwd_bias_perchannel_quantization_example.inc.453F2B26B4A84A7C.idx
 create mode 100755 .cache/clangd/index/run_conv2d_fwd_bias_perlayer_quantization_example.inc.41E3948246A8DC23.idx
 create mode 100755 .cache/clangd/index/run_conv2d_fwd_perchannel_quantization_example.inc.4B57FDA755059955.idx
 create mode 100755 .cache/clangd/index/run_conv2d_fwd_perlayer_quantization_example.inc.3D388FA1A8D19B72.idx
 create mode 100755 .cache/clangd/index/run_convnd_activ_example.inc.F2FB9ED6F9BA4B8A.idx
 create mode 100755 .cache/clangd/index/run_convnd_fwd_example.inc.6C6FF11BF5A73D93.idx
 create mode 100755 .cache/clangd/index/run_convnd_fwd_max_example.inc.4C55B2EACEAC1C07.idx
 create mode 100755 .cache/clangd/index/run_gemm_add_add_fastgelu_example.inc.E8AFCCD92C8D9967.idx
 create mode 100755 .cache/clangd/index/run_gemm_add_multiply_example.inc.3C87D5BB1B980F33.idx
 create mode 100755 .cache/clangd/index/run_gemm_example.inc.4868A822F4677354.idx
 create mode 100755 .cache/clangd/index/run_gemm_test.inc.B7894CAD83BCCCF5.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_bwd_data_bias_relu_example.inc.1D7F06778A26066D.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_bwd_data_example.inc.11E97FA73102DBD5.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_bwd_weight_example.inc.EE2E4EE0E60E4D81.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_conv_fwd_example.inc.88B5CE7F22532219.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_fwd_bias_relu_add_example.inc.7B0A8630D499DF28.idx
 create mode 100755 .cache/clangd/index/run_grouped_conv_fwd_example.inc.C65A743D8AABBB7C.idx
 create mode 100755 .cache/clangd/index/run_grouped_gemm_example.inc.FE653C75BFA23618.idx
 create mode 100755 .cache/clangd/index/run_grouped_gemm_scale_softmax_gemm_permute.inc.0CAE0C43C07E96CE.idx
 create mode 100755 .cache/clangd/index/run_groupnorm_fwd_example.inc.85BD21F884D2FB4A.idx
 create mode 100755 .cache/clangd/index/run_layernorm4d_fwd_example.inc.8870D57CA6553449.idx
 create mode 100755 .cache/clangd/index/run_layernorm_example.inc.DA0460DDD4EF632F.idx
 create mode 100755 .cache/clangd/index/run_permute_bundle_example.inc.419BC377BBAE0EEA.idx
 create mode 100755 .cache/clangd/index/run_permute_element_example.inc.42CC4C3518B6F771.idx
 create mode 100755 .cache/clangd/index/run_splitK_gemm_example.inc.51FAC96E5EFEA0F0.idx
 create mode 100755 .cache/clangd/index/sequence.hpp.25AC5091073FF3EB.idx
 create mode 100755 .cache/clangd/index/sequence_helper.hpp.2282D53015E164C0.idx
 create mode 100755 .cache/clangd/index/softmax.hpp.2F97791F40A73FC9.idx
 create mode 100755 .cache/clangd/index/softmax_blockwise.cpp.3EE65033A8858975.idx
 create mode 100755 .cache/clangd/index/space_filling_curve.cpp.EB322F6A66137D22.idx
 create mode 100755 .cache/clangd/index/span.hpp.D6316D1D55F0830C.idx
 create mode 100755 .cache/clangd/index/sparse_embedding3_forward_layernorm.cpp.CF5214A930A81517.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_bf16.cpp.AB98F65366A9F1A9.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_fp16.cpp.DBCD688A2F933855.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_fp16_fp8.cpp.60034E8F6043ABD9.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_fp32.cpp.A3B8F8ECA755E6B2.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_int8.cpp.4600C416992C0160.idx
 create mode 100755 .cache/clangd/index/splitK_gemm_xdl_lds_direct_load_fp16.cpp.2D633E2EB00CCD78.idx
 create mode 100755 .cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp16.cpp.3597E35811A1130C.idx
 create mode 100755 .cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp32.cpp.8A4C951A87984080.idx
 create mode 100755 .cache/clangd/index/static_buffer.hpp.E059C0D43CFA4618.idx
 create mode 100755 .cache/clangd/index/static_tensor.hpp.55059299885EB585.idx
 create mode 100755 .cache/clangd/index/statically_indexed_array.hpp.8A80A7A4CC47B954.idx
 create mode 100755 .cache/clangd/index/statically_indexed_array_multi_index.hpp.E363636635FF265A.idx
 create mode 100755 .cache/clangd/index/stream_config.hpp.015AF40CD82DD077.idx
 create mode 100755 .cache/clangd/index/stream_utility.hpp.7DD00DAAC010212E.idx
 create mode 100755 .cache/clangd/index/synchronization.hpp.94B924A949B60CF0.idx
 create mode 100755 .cache/clangd/index/tensor.hpp.C84FDD70C53B91FD.idx
 create mode 100755 .cache/clangd/index/tensor_adaptor.hpp.1AABBD498E3DDABA.idx
 create mode 100755 .cache/clangd/index/tensor_descriptor.hpp.7E73ABD05058119F.idx
 create mode 100755 .cache/clangd/index/tensor_descriptor_helper.hpp.B97BC6F42C53AF70.idx
 create mode 100755 .cache/clangd/index/tensor_layout.hpp.83E496410F43D7E9.idx
 create mode 100755 .cache/clangd/index/tensor_partition.hpp.2F54ACF04D747753.idx
 create mode 100755 .cache/clangd/index/tensor_space_filling_curve.hpp.058953EFBD7A2B25.idx
 create mode 100755 .cache/clangd/index/tensor_specialization.hpp.5BC1AF8D88F729F2.idx
 create mode 100755 .cache/clangd/index/tensor_utils.hpp.B92D33FF5A9EED5B.idx
 create mode 100755 .cache/clangd/index/test_avg_pool3d_bwd.cpp.6DD8FAFEF9A89C97.idx
 create mode 100755 .cache/clangd/index/test_avg_pool3d_fwd.cpp.5274F49320881D42.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm.cpp.A022B37F45A32155.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp.52CCAAF6582C3CD0.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp.FB71D53BC1DE7E13.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_util.hpp.38E1B00A8B9BAD43.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_gemm_fp16.cpp.A1B1DD17614F30D3.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_gemm_util.hpp.1FF4EC29A4661CFF.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_softmax_gemm_fp16.cpp.0592033B8F1CD807.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_softmax_gemm_permute_bf16.cpp.EDC4F40BC781A6DC.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_softmax_gemm_permute_fp16.cpp.0D1473C6C2BC0BFC.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_softmax_gemm_permute_util.hpp.DD433B78FEA57450.idx
 create mode 100755 .cache/clangd/index/test_batched_gemm_softmax_gemm_util.hpp.C46CFF8584087968.idx
 create mode 100755 .cache/clangd/index/test_bf8.cpp.16F0A495F47E95CC.idx
 create mode 100755 .cache/clangd/index/test_block_to_ctile_map.cpp.D0968ED6B32C18D9.idx
 create mode 100755 .cache/clangd/index/test_contraction.cpp.6A4280349E6B5DE6.idx
 create mode 100755 .cache/clangd/index/test_contraction_interface.cpp.51D57FFCD74AC94A.idx
 create mode 100755 .cache/clangd/index/test_conv_tensor_rearrange.cpp.69AF0D8F1794ACAD.idx
 create mode 100755 .cache/clangd/index/test_conv_tensor_rearrange_interface.cpp.6D91D3DCA5B6D134.idx
 create mode 100755 .cache/clangd/index/test_elementwise_layernorm_fp16.cpp.6E7042446B0FE0D3.idx
 create mode 100755 .cache/clangd/index/test_fp8.cpp.AF44E860BEF4215C.idx
 create mode 100755 .cache/clangd/index/test_gemm_add.hpp.EFC09082310C599D.idx
 create mode 100755 .cache/clangd/index/test_gemm_add_fastgelu.cpp.9C90F62B27922BE9.idx
 create mode 100755 .cache/clangd/index/test_gemm_add_relu.cpp.6A11991389EEE204.idx
 create mode 100755 .cache/clangd/index/test_gemm_add_relu_add_layernorm_fp16.cpp.E9706C96D2E415C1.idx
 create mode 100755 .cache/clangd/index/test_gemm_add_silu.cpp.0C3DE18B907F5A14.idx
 create mode 100755 .cache/clangd/index/test_gemm_splitk.cpp.921AD8C68523E94F.idx
 create mode 100755 .cache/clangd/index/test_gemm_splitk_ut_cases.inc.CD038407CD6C080D.idx
 create mode 100755 .cache/clangd/index/test_gemm_splitk_util.hpp.D5E2B28B24FEB85E.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_bwd_data.cpp.BA5BCB136A7B5CF1.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_bwd_data_interface_xdl.cpp.8747321B4790F80F.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_bwd_weight.cpp.1E0C24433031CA72.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_bwd_weight_interface_xdl.cpp.B93D201E166036D3.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_fwd.cpp.5B423DD33073B028.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_fwd_multi_ab_interface.cpp.D44C4377C972C730.idx
 create mode 100755 .cache/clangd/index/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp.D8EE1DD81FCEA622.idx
 create mode 100755 .cache/clangd/index/test_grouped_gemm_interface.cpp.BFD580574FDC4028.idx
 create mode 100755 .cache/clangd/index/test_grouped_gemm_splitk.cpp.1985BB1B09859227.idx
 create mode 100755 .cache/clangd/index/test_grouped_gemm_ut_cases.inc.B622DD65303F1D04.idx
 create mode 100755 .cache/clangd/index/test_grouped_gemm_util.hpp.62D7D32AEF140979.idx
 create mode 100755 .cache/clangd/index/test_groupnorm_bwd_data_fp32.cpp.EF83656CE5FBCD24.idx
 create mode 100755 .cache/clangd/index/test_groupnorm_bwd_gamma_beta_fp32.cpp.266D9654B5E1C67F.idx
 create mode 100755 .cache/clangd/index/test_groupnorm_fwd_fp16.cpp.E01BDD088F0473FE.idx
 create mode 100755 .cache/clangd/index/test_groupnorm_fwd_fp32.cpp.06336278490A0D06.idx
 create mode 100755 .cache/clangd/index/test_layernorm2d_bwd_data_fp32.cpp.65C39944BBD4A16D.idx
 create mode 100755 .cache/clangd/index/test_layernorm2d_bwd_gamma_beta_fp32.cpp.B3B6429EB0BBF1C4.idx
 create mode 100755 .cache/clangd/index/test_layernorm2d_fwd_fp16.cpp.49CB42570B950D66.idx
 create mode 100755 .cache/clangd/index/test_layernorm2d_fwd_fp32.cpp.57C9823195A1B889.idx
 create mode 100755 .cache/clangd/index/test_layernorm4d_fwd_fp16.cpp.4CF1AE7F957B873D.idx
 create mode 100755 .cache/clangd/index/test_max_pool3d_bwd.cpp.AF800737CBB45892.idx
 create mode 100755 .cache/clangd/index/test_max_pool3d_fwd.cpp.BE8A2BFD19420616.idx
 create mode 100755 .cache/clangd/index/test_permute_scale.cpp.0607B34D8846C932.idx
 create mode 100755 .cache/clangd/index/test_pool_fwd_common.hpp.CD53AEA46B3FF1EB.idx
 create mode 100755 .cache/clangd/index/test_softmax_interface.cpp.9F0A118BAF3CB91E.idx
 create mode 100755 .cache/clangd/index/test_softmax_rank3.cpp.F74F60A3C98E1A4C.idx
 create mode 100755 .cache/clangd/index/test_softmax_rank4.cpp.05FB4DA05EDB8025.idx
 create mode 100755 .cache/clangd/index/test_softmax_ut_cases.inc.A86C258A9B038466.idx
 create mode 100755 .cache/clangd/index/test_softmax_util.hpp.A85F4CCFF59B9D2C.idx
 create mode 100755 .cache/clangd/index/test_transpose.cpp.EA23A250A71A4476.idx
 create mode 100755 .cache/clangd/index/test_wrapper_copy.cpp.732CF6ED569867C1.idx
 create mode 100755 .cache/clangd/index/test_wrapper_gemm.cpp.74B7113BB3F74220.idx
 create mode 100755 .cache/clangd/index/test_wrapper_layout.cpp.6A8EFC4867A12799.idx
 create mode 100755 .cache/clangd/index/test_wrapper_partition.cpp.EDFCA4C2C7218226.idx
 create mode 100755 .cache/clangd/index/test_wrapper_tensor.cpp.A4AB6B2A22D4689E.idx
 create mode 100755 .cache/clangd/index/thread_group.hpp.568EEF4CF01C2640.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_direct_load.hpp.5F79EA1E492991C0.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v4r1.hpp.CC30E09EE6FAA0EA.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v6r1.hpp.F8F9EEB58869763C.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v6r1r2.hpp.34BED592A25E7D4A.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v6r2.hpp.B8541658FFD968DD.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v6r3.hpp.F9C1F18AD045D6D7.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v7.hpp.85ADAEC27150419B.idx
 create mode 100755 .cache/clangd/index/thread_group_tensor_slice_transfer_v7r2.hpp.E55C206DD2AD1B19.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer.hpp.CE773F216A04AFAA.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v3r1.hpp.2D0C9495DC723ACC.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v4r1.hpp.42FD0DF31A3BFDC2.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v6r1.hpp.7AB8A56ED737D765.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v6r1r2.hpp.AD9623F82BA99EFC.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v6r2.hpp.D28171A7A44ECC24.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v6r3.hpp.C25C8A66C4CF26F7.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v7.hpp.5C98C34294C833A3.idx
 create mode 100755 .cache/clangd/index/threadwise_tensor_slice_transfer_v7r2.hpp.891A886C76B49727.idx
 create mode 100755 .cache/clangd/index/threadwise_welford.hpp.CCCD2666E9AE4B4E.idx
 create mode 100755 .cache/clangd/index/transform_contraction_to_gemm.hpp.51AB58D6126FEA84.idx
 create mode 100755 .cache/clangd/index/transform_conv_bwd_data_to_gemm_v1.hpp.341C45E4BB62B32F.idx
 create mode 100755 .cache/clangd/index/transform_conv_fwd_to_gemm.hpp.14B4756DD59EEE91.idx
 create mode 100755 .cache/clangd/index/transpose_3d.hpp.F3212DEAAECABFBD.idx
 create mode 100755 .cache/clangd/index/transpose_vectors.hpp.7945D9ABB9DC55BD.idx
 create mode 100755 .cache/clangd/index/tuple.hpp.3819E4E0F87690B8.idx
 create mode 100755 .cache/clangd/index/tuple_helper.hpp.73EC15D672B6492A.idx
 create mode 100755 .cache/clangd/index/type.hpp.3093CAE9B6EE9081.idx
 create mode 100755 .cache/clangd/index/type_convert.hpp.5842DA15997DD61D.idx
 create mode 100755 .cache/clangd/index/type_convert_const.cpp.46FAD4095BF1793B.idx
 create mode 100755 .cache/clangd/index/unary_element_wise_operation.hpp.DD0C457D046F7875.idx
 create mode 100755 .cache/clangd/index/welford_helper.hpp.DC75228A25D5F78E.idx
 create mode 100755 .cache/clangd/index/wmma_gemm.hpp.260BC1EC1DC7109A.idx
 create mode 100755 .cache/clangd/index/workgroup_barrier.hpp.E88FB4AF7CFEB4E5.idx
 create mode 100755 .cache/clangd/index/workgroup_synchronization.hpp.EE3FBA63C405EAA9.idx
 create mode 100755 .cache/clangd/index/xdlops_gemm.hpp.C8FD7C44D6227313.idx
 mode change 100644 => 100755 .clang-format
 mode change 100644 => 100755 .clang-tidy
 mode change 100644 => 100755 .github/CODEOWNERS
 mode change 100644 => 100755 .github/ISSUE_TEMPLATE/config.yml
 mode change 100644 => 100755 .github/ISSUE_TEMPLATE/issue_report.yml
 mode change 100644 => 100755 .github/dependabot.yml
 mode change 100644 => 100755 .gitignore
 mode change 100644 => 100755 .pre-commit-config.yaml
 mode change 100644 => 100755 .readthedocs.yaml
 mode change 100644 => 100755 CHANGELOG.md
 mode change 100644 => 100755 CITATION.cff
 mode change 100644 => 100755 CMakeLists.txt
 mode change 100644 => 100755 CONTRIBUTORS.md
 mode change 100644 => 100755 Config.cmake.in
 mode change 100644 => 100755 Dockerfile
 mode change 100644 => 100755 Jenkinsfile
 mode change 100644 => 100755 LICENSE
 mode change 100644 => 100755 README.md
 mode change 100644 => 100755 client_example/01_gemm/CMakeLists.txt
 mode change 100644 => 100755 client_example/01_gemm/gemm.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
 mode change 100644 => 100755 client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
 mode change 100644 => 100755 client_example/03_gemm_layernorm/CMakeLists.txt
 mode change 100644 => 100755 client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
 mode change 100644 => 100755 client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
 mode change 100644 => 100755 client_example/04_contraction/CMakeLists.txt
 mode change 100644 => 100755 client_example/04_contraction/contraction_bilinear_fp32.cpp
 mode change 100644 => 100755 client_example/04_contraction/contraction_bilinear_fp64.cpp
 mode change 100644 => 100755 client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
 mode change 100644 => 100755 client_example/04_contraction/contraction_scale_fp32.cpp
 mode change 100644 => 100755 client_example/04_contraction/contraction_scale_fp64.cpp
 mode change 100644 => 100755 client_example/05_layernorm/CMakeLists.txt
 mode change 100644 => 100755 client_example/05_layernorm/layernorm2d_bwd_data.cpp
 mode change 100644 => 100755 client_example/05_layernorm/layernorm2d_bwd_gamma_beta.cpp
 mode change 100644 => 100755 client_example/05_layernorm/layernorm2d_fwd.cpp
 mode change 100644 => 100755 client_example/05_layernorm/layernorm4d_fwd.cpp
 mode change 100644 => 100755 client_example/06_softmax/CMakeLists.txt
 mode change 100644 => 100755 client_example/06_softmax/softmax4d.cpp
 mode change 100644 => 100755 client_example/07_grouped_convnd_fwd/CMakeLists.txt
 mode change 100644 => 100755 client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
 mode change 100644 => 100755 client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
 mode change 100644 => 100755 client_example/08_fused_attention/CMakeLists.txt
 mode change 100644 => 100755 client_example/08_fused_attention/fused_attention.cpp
 mode change 100644 => 100755 client_example/08_fused_attention/fused_attention_bias.cpp
 mode change 100644 => 100755 client_example/09_quantization/CMakeLists.txt
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
 mode change 100644 => 100755 client_example/09_quantization/gemm_quantization.cpp
 mode change 100644 => 100755 client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
 mode change 100644 => 100755 client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp
 mode change 100644 => 100755 client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/common.hpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp
 mode change 100644 => 100755 client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
 mode change 100644 => 100755 client_example/12_elementwise_normalization/CMakeLists.txt
 mode change 100644 => 100755 client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
 mode change 100644 => 100755 client_example/13_batchnorm/CMakeLists.txt
 mode change 100644 => 100755 client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
 mode change 100644 => 100755 client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
 mode change 100644 => 100755 client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
 mode change 100644 => 100755 client_example/14_instance_id/CMakeLists.txt
 mode change 100644 => 100755 client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
 mode change 100644 => 100755 client_example/15_convnd_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 client_example/15_convnd_bwd_data/common.hpp
 mode change 100644 => 100755 client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
 mode change 100644 => 100755 client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
 mode change 100644 => 100755 client_example/15_gemm_add_multiply/CMakeLists.txt
 mode change 100644 => 100755 client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
 mode change 100644 => 100755 client_example/15_reduce/CMakeLists.txt
 mode change 100644 => 100755 client_example/15_reduce/reduce_nhwc_c.cpp
 mode change 100644 => 100755 client_example/16_convnd_fwd/CMakeLists.txt
 mode change 100644 => 100755 client_example/16_convnd_fwd/common.hpp
 mode change 100644 => 100755 client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
 mode change 100644 => 100755 client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp
 mode change 100644 => 100755 client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
 mode change 100644 => 100755 client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
 mode change 100644 => 100755 client_example/18_groupnorm/CMakeLists.txt
 mode change 100644 => 100755 client_example/18_groupnorm/groupnorm_bwd_data.cpp
 mode change 100644 => 100755 client_example/18_groupnorm/groupnorm_bwd_gamma_beta.cpp
 mode change 100644 => 100755 client_example/18_groupnorm/groupnorm_swish_fwd.cpp
 mode change 100644 => 100755 client_example/19_pool/CMakeLists.txt
 mode change 100644 => 100755 client_example/19_pool/avg_pool3d_bwd.cpp
 mode change 100644 => 100755 client_example/19_pool/avg_pool3d_fwd.cpp
 mode change 100644 => 100755 client_example/19_pool/max_pool2d_bwd.cpp
 mode change 100644 => 100755 client_example/19_pool/max_pool2d_fwd.cpp
 mode change 100644 => 100755 client_example/20_splitk_gemm/CMakeLists.txt
 mode change 100644 => 100755 client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
 mode change 100644 => 100755 client_example/21_grouped_gemm_bias/CMakeLists.txt
 mode change 100644 => 100755 client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
 mode change 100644 => 100755 client_example/22_grouped_gemm/CMakeLists.txt
 mode change 100644 => 100755 client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp
 mode change 100644 => 100755 client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
 mode change 100644 => 100755 client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
 mode change 100644 => 100755 client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
 mode change 100644 => 100755 client_example/22_im2col_col2im/CMakeLists.txt
 mode change 100644 => 100755 client_example/22_im2col_col2im/column_to_image.cpp
 mode change 100644 => 100755 client_example/22_im2col_col2im/image_to_column.cpp
 mode change 100644 => 100755 client_example/23_elementwise_transpose/CMakeLists.txt
 mode change 100644 => 100755 client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/CMakeLists.txt
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab.inc
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu.inc
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp
 mode change 100644 => 100755 client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp
 mode change 100644 => 100755 client_example/25_wrapper/CMakeLists.txt
 mode change 100644 => 100755 client_example/25_wrapper/README.md
 mode change 100644 => 100755 client_example/25_wrapper/tensor_transform_using_wrapper.cpp
 mode change 100644 => 100755 client_example/25_wrapper/wrapper_basic_gemm.cpp
 mode change 100644 => 100755 client_example/25_wrapper/wrapper_img2col.cpp
 mode change 100644 => 100755 client_example/25_wrapper/wrapper_optimized_gemm.cpp
 mode change 100644 => 100755 client_example/CMakeLists.txt
 mode change 100644 => 100755 client_example/README.md
 mode change 100644 => 100755 cmake/Analyzers.cmake
 mode change 100644 => 100755 cmake/ClangTidy.cmake
 mode change 100644 => 100755 cmake/CppCheck.cmake
 mode change 100644 => 100755 cmake/DoxygenDoc.cmake
 mode change 100644 => 100755 cmake/Embed.cmake
 mode change 100644 => 100755 cmake/EnableCompilerWarnings.cmake
 mode change 100644 => 100755 cmake/TargetFlags.cmake
 mode change 100644 => 100755 cmake/getopt.cmake
 mode change 100644 => 100755 cmake/gtest.cmake
 mode change 100644 => 100755 codegen/CMakeLists.txt
 mode change 100644 => 100755 codegen/driver/main.cpp
 mode change 100644 => 100755 codegen/include/ck/host/device_gemm_multiple_d.hpp
 mode change 100644 => 100755 codegen/include/ck/host/device_gemm_multiple_d/operation.hpp
 mode change 100644 => 100755 codegen/include/ck/host/device_gemm_multiple_d/problem.hpp
 mode change 100644 => 100755 codegen/include/ck/host/headers.hpp
 mode change 100644 => 100755 codegen/include/ck/host/operation/gemm.hpp
 mode change 100644 => 100755 codegen/include/ck/host/stringutils.hpp
 mode change 100644 => 100755 codegen/include/ck/host/types.hpp
 mode change 100644 => 100755 codegen/include/ck/host/utils.hpp
 mode change 100644 => 100755 codegen/src/device_gemm_multiple_d.cpp
 mode change 100644 => 100755 codegen/src/device_gemm_multiple_d_operation_xdl_cshuffle.cpp
 mode change 100644 => 100755 codegen/src/headers.cpp
 mode change 100644 => 100755 codegen/src/types.cpp
 mode change 100644 => 100755 codegen/src/utils.cpp
 mode change 100644 => 100755 codegen/test/CMakeLists.txt
 mode change 100644 => 100755 codegen/test/gemm_multiple_d.cpp
 mode change 100644 => 100755 codegen/test/include/test.hpp
 mode change 100644 => 100755 codegen/test/rtc/CMakeLists.txt
 mode change 100644 => 100755 codegen/test/rtc/include/rtc/compile_kernel.hpp
 mode change 100644 => 100755 codegen/test/rtc/include/rtc/hip.hpp
 mode change 100644 => 100755 codegen/test/rtc/include/rtc/kernel.hpp
 mode change 100644 => 100755 codegen/test/rtc/include/rtc/manage_ptr.hpp
 mode change 100644 => 100755 codegen/test/rtc/include/rtc/tmp_dir.hpp
 mode change 100644 => 100755 codegen/test/rtc/src/compile_kernel.cpp
 mode change 100644 => 100755 codegen/test/rtc/src/hip.cpp
 mode change 100644 => 100755 codegen/test/rtc/src/kernel.cpp
 mode change 100644 => 100755 codegen/test/rtc/src/tmp_dir.cpp
 mode change 100644 => 100755 dev-requirements.txt
 mode change 100644 => 100755 docs/Contributors_Guide.rst
 mode change 100644 => 100755 docs/conceptual/what-is-ck.rst
 mode change 100644 => 100755 docs/conf.py
 mode change 100644 => 100755 docs/data/ck_component.png
 mode change 100644 => 100755 docs/data/ck_layer.png
 mode change 100644 => 100755 docs/doxygen/Doxyfile
 mode change 100644 => 100755 docs/index.rst
 mode change 100644 => 100755 docs/install/dockerhub.rst
 mode change 100644 => 100755 docs/license.rst
 mode change 100644 => 100755 docs/reference/API_Reference_Guide.rst
 mode change 100644 => 100755 docs/reference/Supported_Primitives_Guide.rst
 mode change 100644 => 100755 docs/reference/wrapper.rst
 mode change 100644 => 100755 docs/refs.bib
 mode change 100644 => 100755 docs/sphinx/_toc.yml.in
 mode change 100644 => 100755 docs/sphinx/requirements.in
 mode change 100644 => 100755 docs/sphinx/requirements.txt
 mode change 100644 => 100755 docs/tutorial/tutorial_hello_world.rst
 mode change 100644 => 100755 example/01_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/01_gemm/README.md
 mode change 100644 => 100755 example/01_gemm/common.hpp
 mode change 100644 => 100755 example/01_gemm/gemm_dl_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_dl_fp32.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_dl_int4.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_dl_int8.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_dpp_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_wmma_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_bf16_rtn.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp16_fp8.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp16_v2.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp64.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp8.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_fp8_bf8.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_int4.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_int8.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_lds_direct_load_fp32.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_streamk.cpp
 mode change 100644 => 100755 example/01_gemm/gemm_xdl_wavelet_fp16.cpp
 mode change 100644 => 100755 example/01_gemm/run_gemm_example.inc
 mode change 100644 => 100755 example/02_gemm_bilinear/CMakeLists.txt
 mode change 100644 => 100755 example/02_gemm_bilinear/README.md
 mode change 100644 => 100755 example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
 mode change 100644 => 100755 example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
 mode change 100644 => 100755 example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
 mode change 100644 => 100755 example/03_gemm_bias_relu/CMakeLists.txt
 mode change 100644 => 100755 example/03_gemm_bias_relu/README.md
 mode change 100644 => 100755 example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/README.md
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/common.hpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp
 mode change 100644 => 100755 example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
 mode change 100644 => 100755 example/09_convnd_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/09_convnd_fwd/README.md
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_common.hpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_dl_common.hpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
 mode change 100644 => 100755 example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
 mode change 100644 => 100755 example/09_convnd_fwd/run_convnd_fwd_example.inc
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
 mode change 100644 => 100755 example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
 mode change 100644 => 100755 example/12_reduce/CMakeLists.txt
 mode change 100644 => 100755 example/12_reduce/README.md
 mode change 100644 => 100755 example/12_reduce/reduce_blockwise.cpp
 mode change 100644 => 100755 example/12_reduce/reduce_blockwise_impl.hpp
 mode change 100644 => 100755 example/12_reduce/reduce_blockwise_two_call.cpp
 mode change 100644 => 100755 example/12_reduce/reduce_example_common.hpp
 mode change 100644 => 100755 example/12_reduce/reduce_multiblock_atomic_add.cpp
 mode change 100644 => 100755 example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
 mode change 100644 => 100755 example/13_pool2d_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/13_pool2d_fwd/README.md
 mode change 100644 => 100755 example/13_pool2d_fwd/pool2d_fwd_common.hpp
 mode change 100644 => 100755 example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
 mode change 100644 => 100755 example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
 mode change 100644 => 100755 example/14_gemm_quantization/CMakeLists.txt
 mode change 100644 => 100755 example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
 mode change 100644 => 100755 example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
 mode change 100644 => 100755 example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/15_grouped_gemm/README.md
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
 mode change 100644 => 100755 example/15_grouped_gemm/run_grouped_gemm_example.inc
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
 mode change 100644 => 100755 example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
 mode change 100644 => 100755 example/17_convnd_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 example/17_convnd_bwd_data/README.md
 mode change 100644 => 100755 example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
 mode change 100644 => 100755 example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
 mode change 100644 => 100755 example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
 mode change 100644 => 100755 example/18_batched_gemm_reduce/CMakeLists.txt
 mode change 100644 => 100755 example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
 mode change 100644 => 100755 example/19_binary_elementwise/CMakeLists.txt
 mode change 100644 => 100755 example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
 mode change 100644 => 100755 example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
 mode change 100644 => 100755 example/19_binary_elementwise/elementwise_add_1d.cpp
 mode change 100644 => 100755 example/19_binary_elementwise/elementwise_add_4d.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/common.hpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
 mode change 100644 => 100755 example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
 mode change 100644 => 100755 example/21_gemm_layernorm/CMakeLists.txt
 mode change 100644 => 100755 example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
 mode change 100644 => 100755 example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
 mode change 100644 => 100755 example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
 mode change 100644 => 100755 example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
 mode change 100644 => 100755 example/22_cgemm/CMakeLists.txt
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_common.hpp
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_fp32.cpp
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_int4.cpp
 mode change 100644 => 100755 example/22_cgemm/cgemm_xdl_int8.cpp
 mode change 100644 => 100755 example/23_softmax/CMakeLists.txt
 mode change 100644 => 100755 example/23_softmax/README.md
 mode change 100644 => 100755 example/23_softmax/softmax_blockwise.cpp
 mode change 100644 => 100755 example/24_batched_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
 mode change 100644 => 100755 example/24_batched_gemm/batched_gemm_xdl_int4.cpp
 mode change 100644 => 100755 example/24_batched_gemm/batched_gemm_xdl_int8.cpp
 mode change 100644 => 100755 example/24_batched_gemm/run_batched_gemm_example.inc
 mode change 100644 => 100755 example/25_gemm_bias_e_permute/CMakeLists.txt
 mode change 100644 => 100755 example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
 mode change 100644 => 100755 example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
 mode change 100644 => 100755 example/26_contraction/CMakeLists.txt
 mode change 100644 => 100755 example/26_contraction/README.md
 mode change 100644 => 100755 example/26_contraction/common_instances.hpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_bf16_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp16_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp32_compute_bf16.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp32_compute_fp16.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp64.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_bilinear_xdl_fp64_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_bf16_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp16_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp32_compute_bf16.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp32_compute_fp16.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp64.cpp
 mode change 100644 => 100755 example/26_contraction/contraction_scale_xdl_fp64_compute_fp32.cpp
 mode change 100644 => 100755 example/26_contraction/run_contraction_bilinear_example.inc
 mode change 100644 => 100755 example/26_contraction/run_contraction_scale_example.inc
 mode change 100644 => 100755 example/27_layernorm2d_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/27_layernorm2d_fwd/common.hpp
 mode change 100644 => 100755 example/27_layernorm2d_fwd/layernorm2d_fwd_fp16.cpp
 mode change 100644 => 100755 example/27_layernorm2d_fwd/layernorm2d_fwd_splitk_fp16.cpp
 mode change 100644 => 100755 example/27_layernorm2d_fwd/run_layernorm_example.inc
 mode change 100644 => 100755 example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
 mode change 100644 => 100755 example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/29_batched_gemm_bias_e_permute/CMakeLists.txt
 mode change 100644 => 100755 example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
 mode change 100644 => 100755 example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/README.md
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/common.hpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
 mode change 100644 => 100755 example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
 mode change 100644 => 100755 example/31_batched_gemm_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
 mode change 100644 => 100755 example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
 mode change 100644 => 100755 example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
 mode change 100644 => 100755 example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc
 mode change 100644 => 100755 example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp
 mode change 100644 => 100755 example/33_multiple_reduce/CMakeLists.txt
 mode change 100644 => 100755 example/33_multiple_reduce/README.md
 mode change 100644 => 100755 example/33_multiple_reduce/dual_reduce_common.hpp
 mode change 100644 => 100755 example/33_multiple_reduce/dual_reduce_multiblock.cpp
 mode change 100644 => 100755 example/33_multiple_reduce/dual_reduce_threadwise.cpp
 mode change 100644 => 100755 example/34_batchnorm/CMakeLists.txt
 mode change 100644 => 100755 example/34_batchnorm/README.md
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_backward_nhwc.cpp
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_common.hpp
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
 mode change 100644 => 100755 example/34_batchnorm/batchnorm_infer_impl.hpp
 mode change 100644 => 100755 example/35_splitK_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/35_splitK_gemm/run_splitK_gemm_example.inc
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_fp16_fp8.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
 mode change 100644 => 100755 example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp
 mode change 100644 => 100755 example/36_sparse_embedding/CMakeLists.txt
 mode change 100644 => 100755 example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
 mode change 100644 => 100755 example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
 mode change 100644 => 100755 example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/common.hpp
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
 mode change 100644 => 100755 example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
 mode change 100644 => 100755 example/39_permute/CMakeLists.txt
 mode change 100644 => 100755 example/39_permute/common.hpp
 mode change 100644 => 100755 example/39_permute/permute_1xHxW_fp16.cpp
 mode change 100644 => 100755 example/39_permute/permute_HxWx4_fp16.cpp
 mode change 100644 => 100755 example/39_permute/permute_NxHxW_fp16.cpp
 mode change 100644 => 100755 example/39_permute/run_permute_bundle_example.inc
 mode change 100644 => 100755 example/39_permute/run_permute_element_example.inc
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/CMakeLists.txt
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/common.hpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
 mode change 100644 => 100755 example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
 mode change 100644 => 100755 example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
 mode change 100644 => 100755 example/42_groupnorm_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/42_groupnorm_fwd/common.hpp
 mode change 100644 => 100755 example/42_groupnorm_fwd/groupnorm_fwd_sigmoid_mul_fp16.cpp
 mode change 100644 => 100755 example/42_groupnorm_fwd/groupnorm_fwd_splitk_fp16.cpp
 mode change 100644 => 100755 example/42_groupnorm_fwd/groupnorm_fwd_swish_fp16.cpp
 mode change 100644 => 100755 example/42_groupnorm_fwd/run_groupnorm_fwd_example.inc
 mode change 100644 => 100755 example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
 mode change 100644 => 100755 example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
 mode change 100644 => 100755 example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/CMakeLists.txt
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_3d.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp
 mode change 100644 => 100755 example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp
 mode change 100644 => 100755 example/45_elementwise_normalization/CMakeLists.txt
 mode change 100644 => 100755 example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
 mode change 100644 => 100755 example/46_gemm_add_multiply/CMakeLists.txt
 mode change 100644 => 100755 example/46_gemm_add_multiply/README.md
 mode change 100644 => 100755 example/46_gemm_add_multiply/common.hpp
 mode change 100644 => 100755 example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
 mode change 100644 => 100755 example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
 mode change 100644 => 100755 example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
 mode change 100644 => 100755 example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
 mode change 100644 => 100755 example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
 mode change 100644 => 100755 example/48_pool3d_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/48_pool3d_fwd/pool3d_fwd_common.hpp
 mode change 100644 => 100755 example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
 mode change 100644 => 100755 example/49_maxpool2d_bwd/CMakeLists.txt
 mode change 100644 => 100755 example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
 mode change 100644 => 100755 example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
 mode change 100644 => 100755 example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
 mode change 100644 => 100755 example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
 mode change 100644 => 100755 example/50_put_element/CMakeLists.txt
 mode change 100644 => 100755 example/50_put_element/put_element_fp16.cpp
 mode change 100644 => 100755 example/51_avgpool3d_bwd/CMakeLists.txt
 mode change 100644 => 100755 example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
 mode change 100644 => 100755 example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
 mode change 100644 => 100755 example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
 mode change 100644 => 100755 example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
 mode change 100644 => 100755 example/52_im2col_col2im/CMakeLists.txt
 mode change 100644 => 100755 example/52_im2col_col2im/column_to_image_f32.cpp
 mode change 100644 => 100755 example/52_im2col_col2im/common.hpp
 mode change 100644 => 100755 example/52_im2col_col2im/image_to_column_f32.cpp
 mode change 100644 => 100755 example/53_layernorm2d_bwd/CMakeLists.txt
 mode change 100644 => 100755 example/53_layernorm2d_bwd/layernorm2d_bwd_fp32.cpp
 mode change 100644 => 100755 example/54_groupnorm_bwd/CMakeLists.txt
 mode change 100644 => 100755 example/54_groupnorm_bwd/groupnorm_bwd_fp32.cpp
 mode change 100644 => 100755 example/60_gemm_multi_ABD/CMakeLists.txt
 mode change 100644 => 100755 example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
 mode change 100644 => 100755 example/61_contraction_multi_ABD/CMakeLists.txt
 mode change 100644 => 100755 example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/CMakeLists.txt
 mode change 100644 => 100755 example/62_convnd_activ/binary/CMakeLists.txt
 mode change 100644 => 100755 example/62_convnd_activ/binary/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/binary/convnd_fwd_xdl_bilinear_residual_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/CMakeLists.txt
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp
 mode change 100644 => 100755 example/62_convnd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp
 mode change 100644 => 100755 example/62_convnd_activ/run_convnd_activ_example.inc
 mode change 100644 => 100755 example/62_convnd_activ/unary/CMakeLists.txt
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_activ_unary_common.hpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_abs_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_clippedrelu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_elu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_leakyrelu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_pow_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_relu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_sigmoid_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_softrelu_fp16.cpp
 mode change 100644 => 100755 example/62_convnd_activ/unary/convnd_fwd_xdl_tanh_fp16.cpp
 mode change 100644 => 100755 example/63_layernorm4d_fwd/CMakeLists.txt
 mode change 100644 => 100755 example/63_layernorm4d_fwd/common.hpp
 mode change 100644 => 100755 example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp
 mode change 100644 => 100755 example/63_layernorm4d_fwd/layernorm4d_fwd_splitk_fp16.cpp
 mode change 100644 => 100755 example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc
 mode change 100644 => 100755 example/64_fpAintB_gemm/CMakeLists.txt
 mode change 100644 => 100755 example/64_fpAintB_gemm/common.hpp
 mode change 100644 => 100755 example/64_fpAintB_gemm/fp16int8_gemm_wmma.cpp
 mode change 100644 => 100755 example/64_fpAintB_gemm/run_gemm_example.inc
 mode change 100644 => 100755 example/CMakeLists.txt
 mode change 100644 => 100755 include/ck/ck.hpp
 mode change 100644 => 100755 include/ck/config.h.in
 mode change 100644 => 100755 include/ck/host_utility/device_prop.hpp
 mode change 100644 => 100755 include/ck/host_utility/hip_check_error.hpp
 mode change 100644 => 100755 include/ck/host_utility/io.hpp
 mode change 100644 => 100755 include/ck/host_utility/kernel_launch.hpp
 mode change 100644 => 100755 include/ck/host_utility/stream_utility.hpp
 mode change 100644 => 100755 include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
 mode change 100644 => 100755 include/ck/stream_config.hpp
 mode change 100644 => 100755 include/ck/tensor/static_tensor.hpp
 mode change 100644 => 100755 include/ck/tensor_description/cluster_descriptor.hpp
 mode change 100644 => 100755 include/ck/tensor_description/multi_index_transform.hpp
 mode change 100644 => 100755 include/ck/tensor_description/multi_index_transform_helper.hpp
 mode change 100644 => 100755 include/ck/tensor_description/tensor_adaptor.hpp
 mode change 100644 => 100755 include/ck/tensor_description/tensor_descriptor.hpp
 mode change 100644 => 100755 include/ck/tensor_description/tensor_descriptor_helper.hpp
 mode change 100644 => 100755 include/ck/tensor_description/tensor_space_filling_curve.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_dequant.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_base.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_cgemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_elementwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_elementwise_scale.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_dequantB.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_normalization_fwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_put_element.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_reduce.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_softmax.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_fpAintB_gemm_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_grouped_query_attention_forward_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_multi_query_attention_forward_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/masking_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/matrix_padder.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/tensor_layout.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/device/welford_helper.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/element/quantization_operation.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d_scale.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_elementwise_3d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_fpAintB_gemm_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_data.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_gamma_beta.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_dequant.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm_arraybase.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
 mode change 100644 => 100755 include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
 mode change 100644 => 100755 include/ck/utility/amd_address_space.hpp
 mode change 100644 => 100755 include/ck/utility/amd_buffer_addressing.hpp
 mode change 100644 => 100755 include/ck/utility/amd_gemm_dpp.hpp
 mode change 100644 => 100755 include/ck/utility/amd_inline_asm.hpp
 mode change 100644 => 100755 include/ck/utility/amd_lds.hpp
 mode change 100644 => 100755 include/ck/utility/amd_wave_read_first_lane.hpp
 mode change 100644 => 100755 include/ck/utility/amd_wmma.hpp
 mode change 100644 => 100755 include/ck/utility/amd_xdlops.hpp
 mode change 100644 => 100755 include/ck/utility/array.hpp
 mode change 100644 => 100755 include/ck/utility/array_multi_index.hpp
 mode change 100644 => 100755 include/ck/utility/c_style_pointer_cast.hpp
 mode change 100644 => 100755 include/ck/utility/common_header.hpp
 mode change 100644 => 100755 include/ck/utility/container_element_picker.hpp
 mode change 100644 => 100755 include/ck/utility/container_helper.hpp
 mode change 100644 => 100755 include/ck/utility/data_type.hpp
 mode change 100644 => 100755 include/ck/utility/debug.hpp
 mode change 100644 => 100755 include/ck/utility/dynamic_buffer.hpp
 mode change 100644 => 100755 include/ck/utility/enable_if.hpp
 mode change 100644 => 100755 include/ck/utility/f8_utils.hpp
 mode change 100644 => 100755 include/ck/utility/functional.hpp
 mode change 100644 => 100755 include/ck/utility/functional2.hpp
 mode change 100644 => 100755 include/ck/utility/functional3.hpp
 mode change 100644 => 100755 include/ck/utility/functional4.hpp
 mode change 100644 => 100755 include/ck/utility/generic_memory_space_atomic.hpp
 mode change 100644 => 100755 include/ck/utility/get_id.hpp
 mode change 100644 => 100755 include/ck/utility/get_shift.hpp
 mode change 100644 => 100755 include/ck/utility/ignore.hpp
 mode change 100644 => 100755 include/ck/utility/inner_product.hpp
 mode change 100644 => 100755 include/ck/utility/inner_product_dpp8.hpp
 mode change 100644 => 100755 include/ck/utility/integral_constant.hpp
 mode change 100644 => 100755 include/ck/utility/is_detected.hpp
 mode change 100644 => 100755 include/ck/utility/is_known_at_compile_time.hpp
 mode change 100644 => 100755 include/ck/utility/loop_scheduler.hpp
 mode change 100644 => 100755 include/ck/utility/magic_division.hpp
 mode change 100644 => 100755 include/ck/utility/math.hpp
 mode change 100644 => 100755 include/ck/utility/math_v2.hpp
 mode change 100644 => 100755 include/ck/utility/multi_index.hpp
 mode change 100644 => 100755 include/ck/utility/number.hpp
 mode change 100644 => 100755 include/ck/utility/random_gen.hpp
 mode change 100644 => 100755 include/ck/utility/reduction_common.hpp
 mode change 100644 => 100755 include/ck/utility/reduction_enums.hpp
 mode change 100644 => 100755 include/ck/utility/reduction_functions_accumulate.hpp
 mode change 100644 => 100755 include/ck/utility/reduction_operator.hpp
 mode change 100644 => 100755 include/ck/utility/sequence.hpp
 mode change 100644 => 100755 include/ck/utility/sequence_helper.hpp
 mode change 100644 => 100755 include/ck/utility/span.hpp
 mode change 100644 => 100755 include/ck/utility/static_buffer.hpp
 mode change 100644 => 100755 include/ck/utility/statically_indexed_array.hpp
 mode change 100644 => 100755 include/ck/utility/statically_indexed_array_multi_index.hpp
 mode change 100644 => 100755 include/ck/utility/synchronization.hpp
 mode change 100644 => 100755 include/ck/utility/thread_group.hpp
 mode change 100644 => 100755 include/ck/utility/transpose_vectors.hpp
 mode change 100644 => 100755 include/ck/utility/tuple.hpp
 mode change 100644 => 100755 include/ck/utility/tuple_helper.hpp
 mode change 100644 => 100755 include/ck/utility/type.hpp
 mode change 100644 => 100755 include/ck/utility/type_convert.hpp
 mode change 100644 => 100755 include/ck/utility/workgroup_barrier.hpp
 mode change 100644 => 100755 include/ck/utility/workgroup_synchronization.hpp
 mode change 100644 => 100755 include/ck/version.h.in
 mode change 100644 => 100755 include/ck/wrapper/layout.hpp
 mode change 100644 => 100755 include/ck/wrapper/operations/copy.hpp
 mode change 100644 => 100755 include/ck/wrapper/operations/gemm.hpp
 mode change 100644 => 100755 include/ck/wrapper/tensor.hpp
 mode change 100644 => 100755 include/ck/wrapper/traits/blockwise_gemm_xdl_traits.hpp
 mode change 100644 => 100755 include/ck/wrapper/utils/kernel_utils.hpp
 mode change 100644 => 100755 include/ck/wrapper/utils/layout_utils.hpp
 mode change 100644 => 100755 include/ck/wrapper/utils/tensor_partition.hpp
 mode change 100644 => 100755 include/ck/wrapper/utils/tensor_utils.hpp
 mode change 100644 => 100755 library/CMakeLists.txt
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_fpAintB_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
 mode change 100644 => 100755 library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_f16_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_i8_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_scale_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_bilinear_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scale_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_bilinear.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_scale.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bilinear.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scale.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_scaleadd_relu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_data.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/transpose/device_transpose_instance.hpp
 mode change 100644 => 100755 library/include/ck/library/tensor_operation_instance/gpu/transpose_3d.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/algorithm.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/check_err.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/conv_common.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/convolution_parameter.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/device_memory.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/fill.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/host_common_util.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/host_gemm.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/host_tensor.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/host_tensor_generator.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/iterator.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/literals.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/numeric.hpp
 mode change 100644 => 100755 library/include/ck/library/utility/ranges.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gndhwc_3d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnhwc_2d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnwc_1d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_ndhwgc_3d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwgc_2d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nwgc_1d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_silu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gndhwc_3d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnhwc_2d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnwc_1d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_ndhwgc_3d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwgc_2d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nwgc_1d_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_groupnorm_bwd_data_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_data/normalization_bwd_data_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_groupnorm_bwd_gamma_beta_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/normalization_bwd_gamma_beta_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_1d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_2d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_3d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_4d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_5d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_instances.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/transpose/CMakeLists.txt
 mode change 100644 => 100755 library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp
 mode change 100644 => 100755 library/src/utility/CMakeLists.txt
 mode change 100644 => 100755 library/src/utility/convolution_parameter.cpp
 mode change 100644 => 100755 library/src/utility/device_memory.cpp
 mode change 100644 => 100755 library/src/utility/host_tensor.cpp
 mode change 100644 => 100755 profiler/CMakeLists.txt
 mode change 100644 => 100755 profiler/README.md
 mode change 100644 => 100755 profiler/include/profiler/data_type_enum.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batchnorm_backward_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batchnorm_forward_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_batchnorm_infer_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_contraction_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_contraction_utils.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_conv_bwd_data_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_conv_fwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_relu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_add_silu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_bilinear_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_reduce_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_splitk_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_gemm_streamk_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_grouped_gemm_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_groupnorm_bwd_data_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_groupnorm_fwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_layernorm_bwd_data_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_layernorm_fwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_permute_scale_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_pool3d_fwd_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_reduce_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_softmax_impl.hpp
 mode change 100644 => 100755 profiler/include/profiler/profile_transpose_impl.hpp
 mode change 100644 => 100755 profiler/src/CMakeLists.txt
 mode change 100644 => 100755 profiler/src/profile_avg_pool3d_bwd.cpp
 mode change 100644 => 100755 profiler/src/profile_batched_gemm.cpp
 mode change 100644 => 100755 profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
 mode change 100644 => 100755 profiler/src/profile_batched_gemm_gemm.cpp
 mode change 100644 => 100755 profiler/src/profile_batched_gemm_multi_d.cpp
 mode change 100644 => 100755 profiler/src/profile_batched_gemm_reduce.cpp
 mode change 100644 => 100755 profiler/src/profile_batchnorm_bwd.cpp
 mode change 100644 => 100755 profiler/src/profile_batchnorm_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_batchnorm_infer.cpp
 mode change 100644 => 100755 profiler/src/profile_contraction_bilinear.cpp
 mode change 100644 => 100755 profiler/src/profile_contraction_scale.cpp
 mode change 100644 => 100755 profiler/src/profile_conv_bwd_data.cpp
 mode change 100644 => 100755 profiler/src/profile_conv_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_conv_fwd_bias_relu.cpp
 mode change 100644 => 100755 profiler/src/profile_conv_fwd_bias_relu_add.cpp
 mode change 100644 => 100755 profiler/src/profile_conv_tensor_rearrange.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_add_fastgelu.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_fastgelu.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_multiply.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_relu.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_relu_add_layernorm.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_add_silu.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_bias_add_reduce.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_bilinear.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_fastgelu.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_multiply_add.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_reduce.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_splitk.cpp
 mode change 100644 => 100755 profiler/src/profile_gemm_streamk.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_conv_bwd_data.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_conv_bwd_weight.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_conv_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_gemm.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_gemm_fastgelu.cpp
 mode change 100644 => 100755 profiler/src/profile_grouped_gemm_fixed_nk.cpp
 mode change 100644 => 100755 profiler/src/profile_groupnorm_bwd_data.cpp
 mode change 100644 => 100755 profiler/src/profile_groupnorm_bwd_gamma_beta.cpp
 mode change 100644 => 100755 profiler/src/profile_groupnorm_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_layernorm_bwd_data.cpp
 mode change 100644 => 100755 profiler/src/profile_layernorm_bwd_gamma_beta.cpp
 mode change 100644 => 100755 profiler/src/profile_layernorm_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_max_pool3d_bwd.cpp
 mode change 100644 => 100755 profiler/src/profile_max_pool3d_fwd.cpp
 mode change 100644 => 100755 profiler/src/profile_permute_scale.cpp
 mode change 100644 => 100755 profiler/src/profile_reduce.cpp
 mode change 100644 => 100755 profiler/src/profile_softmax.cpp
 mode change 100644 => 100755 profiler/src/profile_transpose.cpp
 mode change 100644 => 100755 profiler/src/profiler.cpp
 mode change 100644 => 100755 profiler/src/profiler_operation_registry.hpp
 mode change 100644 => 100755 rbuild.ini
 mode change 100644 => 100755 requirements.txt
 mode change 100644 => 100755 script/hip_fatbin_insert
 mode change 100644 => 100755 script/process_perf_data.py
 mode change 100644 => 100755 script/redis-cli.conf
 mode change 100644 => 100755 script/test_convnd_fwd.sh
 mode change 100644 => 100755 test/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm/test_batched_gemm.cpp
 mode change 100644 => 100755 test/batched_gemm_gemm/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
 mode change 100644 => 100755 test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
 mode change 100644 => 100755 test/batched_gemm_multi_d/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp
 mode change 100644 => 100755 test/batched_gemm_reduce/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
 mode change 100644 => 100755 test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
 mode change 100644 => 100755 test/batchnorm/CMakeLists.txt
 mode change 100644 => 100755 test/batchnorm/batchnorm_bwd_rank_4.cpp
 mode change 100644 => 100755 test/batchnorm/batchnorm_fwd_rank_4.cpp
 mode change 100644 => 100755 test/batchnorm/batchnorm_infer_rank_4.cpp
 mode change 100644 => 100755 test/block_swizzle_test/block_swizzle_test.cpp
 mode change 100644 => 100755 test/block_swizzle_test/rebuild.sh
 mode change 100644 => 100755 test/block_swizzle_test/simple_args.h
 mode change 100644 => 100755 test/block_to_ctile_map/CMakeLists.txt
 mode change 100644 => 100755 test/block_to_ctile_map/test_block_to_ctile_map.cpp
 mode change 100644 => 100755 test/contraction/CMakeLists.txt
 mode change 100644 => 100755 test/contraction/test_contraction.cpp
 mode change 100644 => 100755 test/contraction/test_contraction_interface.cpp
 mode change 100644 => 100755 test/conv_tensor_rearrange/CMakeLists.txt
 mode change 100644 => 100755 test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
 mode change 100644 => 100755 test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
 mode change 100644 => 100755 test/conv_util/CMakeLists.txt
 mode change 100644 => 100755 test/conv_util/conv_util.cpp
 mode change 100644 => 100755 test/convnd_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 test/convnd_bwd_data/convnd_bwd_data.cpp
 mode change 100644 => 100755 test/convnd_fwd/CMakeLists.txt
 mode change 100644 => 100755 test/convnd_fwd/convnd_fwd.cpp
 mode change 100644 => 100755 test/data_type/CMakeLists.txt
 mode change 100644 => 100755 test/data_type/test_bf8.cpp
 mode change 100644 => 100755 test/data_type/test_fp8.cpp
 mode change 100644 => 100755 test/data_type/test_int4.cpp
 mode change 100644 => 100755 test/data_type/type_convert_const.cpp
 mode change 100644 => 100755 test/elementwise_normalization/CMakeLists.txt
 mode change 100644 => 100755 test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
 mode change 100644 => 100755 test/gemm/CMakeLists.txt
 mode change 100644 => 100755 test/gemm/gemm_bf16.cpp
 mode change 100644 => 100755 test/gemm/gemm_fp16.cpp
 mode change 100644 => 100755 test/gemm/gemm_fp32.cpp
 mode change 100644 => 100755 test/gemm/gemm_fp64.cpp
 mode change 100644 => 100755 test/gemm/gemm_int8.cpp
 mode change 100644 => 100755 test/gemm/gemm_standalone_xdl_fp16.cpp
 mode change 100644 => 100755 test/gemm/gemm_util.hpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_nn_instance.cpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_nn_instance.hpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_nt_instance.cpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_nt_instance.hpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_tn_instance.cpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_tn_instance.hpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_tt_instance.cpp
 mode change 100644 => 100755 test/gemm/instance/gemm_f16_tt_instance.hpp
 mode change 100644 => 100755 test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
 mode change 100644 => 100755 test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
 mode change 100644 => 100755 test/gemm/run_gemm_test.inc
 mode change 100644 => 100755 test/gemm_add/CMakeLists.txt
 mode change 100644 => 100755 test/gemm_add/test_gemm_add.hpp
 mode change 100644 => 100755 test/gemm_add/test_gemm_add_fastgelu.cpp
 mode change 100644 => 100755 test/gemm_add/test_gemm_add_relu.cpp
 mode change 100644 => 100755 test/gemm_add/test_gemm_add_silu.cpp
 mode change 100644 => 100755 test/gemm_layernorm/CMakeLists.txt
 mode change 100644 => 100755 test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
 mode change 100644 => 100755 test/gemm_reduce/CMakeLists.txt
 mode change 100644 => 100755 test/gemm_reduce/gemm_reduce_fp16.cpp
 mode change 100644 => 100755 test/gemm_split_k/CMakeLists.txt
 mode change 100644 => 100755 test/gemm_split_k/test_gemm_splitk.cpp
 mode change 100644 => 100755 test/gemm_split_k/test_gemm_splitk_ut_cases.inc
 mode change 100644 => 100755 test/gemm_split_k/test_gemm_splitk_util.hpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_weight/CMakeLists.txt
 mode change 100644 => 100755 test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp
 mode change 100644 => 100755 test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp
 mode change 100644 => 100755 test/grouped_convnd_fwd/CMakeLists.txt
 mode change 100644 => 100755 test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
 mode change 100644 => 100755 test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_ab_interface.cpp
 mode change 100644 => 100755 test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp
 mode change 100644 => 100755 test/grouped_gemm/CMakeLists.txt
 mode change 100644 => 100755 test/grouped_gemm/test_grouped_gemm_interface.cpp
 mode change 100644 => 100755 test/grouped_gemm/test_grouped_gemm_splitk.cpp
 mode change 100644 => 100755 test/grouped_gemm/test_grouped_gemm_ut_cases.inc
 mode change 100644 => 100755 test/grouped_gemm/test_grouped_gemm_util.hpp
 mode change 100644 => 100755 test/magic_number_division/CMakeLists.txt
 mode change 100644 => 100755 test/magic_number_division/magic_number_division.cpp
 mode change 100644 => 100755 test/normalization_bwd_data/CMakeLists.txt
 mode change 100644 => 100755 test/normalization_bwd_data/test_groupnorm_bwd_data_fp32.cpp
 mode change 100644 => 100755 test/normalization_bwd_data/test_layernorm2d_bwd_data_fp32.cpp
 mode change 100644 => 100755 test/normalization_bwd_gamma_beta/CMakeLists.txt
 mode change 100644 => 100755 test/normalization_bwd_gamma_beta/test_groupnorm_bwd_gamma_beta_fp32.cpp
 mode change 100644 => 100755 test/normalization_bwd_gamma_beta/test_layernorm2d_bwd_gamma_beta_fp32.cpp
 mode change 100644 => 100755 test/normalization_fwd/CMakeLists.txt
 mode change 100644 => 100755 test/normalization_fwd/test_groupnorm_fwd_fp16.cpp
 mode change 100644 => 100755 test/normalization_fwd/test_groupnorm_fwd_fp32.cpp
 mode change 100644 => 100755 test/normalization_fwd/test_layernorm2d_fwd_fp16.cpp
 mode change 100644 => 100755 test/normalization_fwd/test_layernorm2d_fwd_fp32.cpp
 mode change 100644 => 100755 test/normalization_fwd/test_layernorm4d_fwd_fp16.cpp
 mode change 100644 => 100755 test/permute_scale/CMakeLists.txt
 mode change 100644 => 100755 test/permute_scale/test_permute_scale.cpp
 mode change 100644 => 100755 test/pool/CMakeLists.txt
 mode change 100644 => 100755 test/pool/test_avg_pool3d_bwd.cpp
 mode change 100644 => 100755 test/pool/test_avg_pool3d_fwd.cpp
 mode change 100644 => 100755 test/pool/test_max_pool3d_bwd.cpp
 mode change 100644 => 100755 test/pool/test_max_pool3d_fwd.cpp
 mode change 100644 => 100755 test/pool/test_pool_fwd_common.hpp
 mode change 100644 => 100755 test/reduce/CMakeLists.txt
 mode change 100644 => 100755 test/reduce/reduce_no_index.cpp
 mode change 100644 => 100755 test/reduce/reduce_with_index.cpp
 mode change 100644 => 100755 test/reference_conv_fwd/CMakeLists.txt
 mode change 100644 => 100755 test/reference_conv_fwd/reference_conv_fwd.cpp
 mode change 100644 => 100755 test/softmax/CMakeLists.txt
 mode change 100644 => 100755 test/softmax/test_softmax_interface.cpp
 mode change 100644 => 100755 test/softmax/test_softmax_rank3.cpp
 mode change 100644 => 100755 test/softmax/test_softmax_rank4.cpp
 mode change 100644 => 100755 test/softmax/test_softmax_ut_cases.inc
 mode change 100644 => 100755 test/softmax/test_softmax_util.hpp
 mode change 100644 => 100755 test/space_filling_curve/CMakeLists.txt
 mode change 100644 => 100755 test/space_filling_curve/space_filling_curve.cpp
 mode change 100644 => 100755 test/transpose/CMakeLists.txt
 mode change 100644 => 100755 test/transpose/test_transpose.cpp
 mode change 100644 => 100755 test/wmma_op/CMakeLists.txt
 mode change 100644 => 100755 test/wmma_op/wmma_op.cpp
 mode change 100644 => 100755 test/wmma_op/wmma_op_util.hpp
 mode change 100644 => 100755 test/wrapper/CMakeLists.txt
 mode change 100644 => 100755 test/wrapper/test_wrapper_copy.cpp
 mode change 100644 => 100755 test/wrapper/test_wrapper_gemm.cpp
 mode change 100644 => 100755 test/wrapper/test_wrapper_layout.cpp
 mode change 100644 => 100755 test/wrapper/test_wrapper_partition.cpp
 mode change 100644 => 100755 test/wrapper/test_wrapper_tensor.cpp

diff --git a/.cache/clangd/index/add_device_operation_instance.hpp.C89DB8EBC3EFFE7D.idx b/.cache/clangd/index/add_device_operation_instance.hpp.C89DB8EBC3EFFE7D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..50693bbf1c7652bd6a4b695f50269131572d869f
GIT binary patch
literal 1268
zcmWIYbaQ*d!oc91;#rZKT9OE4G6({3aY<38EF%NMS|$dDiaARs`Q{xm5NNHPqNBmY
zX{F@Ny)ogsuVqu}n}r*bY>v%S=?Zv%Z;Qs#l8f;txwH2@|DNV1w)teGs&TYtw24pV
z-?IEBYtt1<%fE)D-#L1A*^HnNy_9PQ6YAnbCKqe+^0qz_+#OpZ5opMJW2N-Rm@{?q
z8#U^l*@@qZO5Iq>#@^t$d2zwhLnp#!M0p*Xpc$(g6es8TlUHsn+rxvEQy2m+3NAc#
z>(!g6jw^{lRvGuxrysmv7<8?=f7LOO3(efe%*75~xv=4LU$uqF!9B+^;_rnU{=0f}
zckQK>|7RFXv*Z7=Pv}QcqWRr`6c14w$;~NB^CwPUEhHr9Jo&Qe*L8RJ*x8QNx*YYi
zKKH+1=N}QnZz@qmVS5;IT7OqF6j$aZIRHbe_t5VJYaUuFFfnnn0=YoI!OO)f#J~yR
z915{}883hRI1d9e8v};`tAH|?VBln8WZ+?8<X|jHO)FMqW?&F1RyS@c6f|b!;SkXh
z(Pf?R%R)dVbe>FL$ovnVKF&lFmQ5F&bNzgVAeyjJnSAW-#&zCk!s^F1h}8(h%A*Nu
zOgDUXVLQ(ZG-0hpf85#B{z;$-TTHDf<(P0N5lz_DOyT;Yofpwn2VVKQgf-&R0yNbj
zenRRmCfcw9g#|PP^kJbBn$Fp`W4ciYn(CMreH(6etx7@@&e-e9wn)&)3r#rJBsU{;
z_PuO0;i@lR`UDShoJA9Euc|U_PFmWDCfsqZ`uCR9CwxF*VQFD?m;<Ms^jLnmtxg9h
zEUh99clW}-yVlh2Jeh-L$P(w<vQCG7%|=tb#r_<l`brf=G}XIp6a}AjeCq`Y3v&ny
z!K~aXq}bp(LHatH>iw}TwH7BV*wKVxx)`~@=?j)B53dNw<8|X`0jfrJ_sJ{K5)qci
z8PN<mUDddCV$)V-G~tWaO2f4kj@1K&#pJ|vU`g}N&GZ#Z8kiTMseWp}`h}feyc;Mi
zU@Bk_bJ$z{!={NA+UL<!f0b2@(BXNq4=5}lEujuG`^UaztuL1U09lEg8yJdGa}q%r
ntGFn+SRcfIW-}0z9YjC?hwI8U{_jFwM=*17vNLeOgc%qBnk}y8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/algorithm.hpp.3E7EDFC4AA4FB32B.idx b/.cache/clangd/index/algorithm.hpp.3E7EDFC4AA4FB32B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94088da4eddbf26ced021a3dc65670ed81446ca9
GIT binary patch
literal 2408
zcmWIYbaPAKWMFVk@vO*AElFfyU|<jg;^LB`%zj1&26>=B#hj%<zW$F51it!zWL)LO
zC40+NM(0SQw7Lq*^An|aj;(sV@r~zdj*s^@3VWQ+yVx{ADb*(b-fVv}+v=I!)z71=
zrsS>tnIiP*|JEnz`rIEc1UbF)WcQTV;dw0kVBNWsC*(iANqHex!^UHE?^$~5Kj%V)
zd6Pn3D1B77Gdvf-_ustor2mt@btxzJYv_5-S+V-bd@rpTLe|U7oW5j;G--a3UA=YF
zl}g**5blYw&0DWzKV~a#P1JX;Ene}t>{hpk>&f37Z(rrNe0wWapSpzQxXW$1<+2*?
zkL3Fj`H~;AMoX?{WiZ)s^@2;}>5Q@mh3lPF=WdBL5fsU@b=$dR*R2;1wy<VTQ_O5A
zJhg2r+g8a~UCHbf%$e6Ecua3cKK?s{-PSQ9VeJ)G#ak<0Chcoz@VIcS;DUAhou&Lj
zY4^&1aCh3Q)#5!B@Tc`ub@E~-K5yUn_{L3{dkpr^iaocE|ID^&SM5CJztfdlHu2@O
zuiZBnuGHT3bI*C6`btOl^k4Udtz>V{FDg8JY;$7%rK<&}7>X-%lahh)BUiq_;W(qp
zGad#`4h9}xR$fjp!NAGF$iT(G$io<!reye4>wF{+gCJ0W1giumm|);!V`dOy2P$|I
zYvycyUt}K-gD_BmGOIElm|);!Wn$n4nZQt#npXUsje$XE|E4+lKW&N_d3YpwB!yrm
zi+<@=xIXPAI~xxVpFE!m+(kS>QbO_wo}{*<0Za=IkF2n)1cIlcqM`|R2alS(nhJ~u
z)6d8S_E%kv_ynEr=k{rw&zdc|qU9n@>@?iM$>mllb%hD&Pen;pgfnG1WqDyd(@D2n
zuZ2%^r-=*9y3R4^U8^hvx`0oVPZAagJUl`ILLxAp#b!m>*p&9)Ko{`K^DDuGd3c1S
zgk=#tDQzh|n4@`k<b>tK5Ihxg6)RXs^YEw{shPrfmU33wZHpWi(ZujjC1?4r?Fx5k
zVtDvs<E_b}YN0fV7+BB&lOrfA!IEQ4VeUVn2~u-_@uecFA_ogZV2}!^Ny2!kl^RPb
z4{i{~o2;_AqrP0Rb8V(c&=gk%ONh-{$B5UNHR8;*GnxN=;{c{TEj~j;LgLrp*GKS#
zRE2a9JYi*FEd)=}Owtw>Qan6TCQ{Z2o~)cKQjU_7kyD1rb>3c-x5|E@FWy8yVVm^B
z5Z`shGzp)H-Mxzfs;4mHb-`?FCH|R~Zrf<$f&-nNxs38n2WgUc55D%9^YP00jd-K~
zsN2OO7j6e6(LMTKIWB(t>g(itH1Rqt4+8TosD^{(+rQ-;r?>iR6yOaJVEvbq2&`%u
pii?tqtw9Vv5CH)KFa`qyFM`Q)x3()nSKQ?}ivTYlKLbBP1^@(g0~`PV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_address_space.hpp.FB92D54F0A7BAF6E.idx b/.cache/clangd/index/amd_address_space.hpp.FB92D54F0A7BAF6E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..608fe43dd2f44ba7cb12200346b5cf3fdfb3ffed
GIT binary patch
literal 890
zcmWIYbaN|WW?*nm@vO*AElFfyU|<jg;^LB`%nv|Xfr){kV$RwL2mKBk@U(r_ur6TW
zbBU>5y(^3Tg2R$WS(TlUi<fzZEMLI&@!gWOWeFbEzMGP?fByO3cl7Pq@5zC-j$b#0
z`|iwNa3}A6)>4PP(?a)VntlKH+)wAq%6&fR3BF4WnGIKW%*wL=_9$uojfu+-W^F8f
z*x+X)E%r)+-EjZj^Q~@v0?l7L**q9aKFqnGe&8xcMcJ=*+3>r?tF*hGTFf_!ua9tV
zowAA7_T;pS3b*|w4~L7n^yej=4r;e!JJziHBICo8C0}p#wk>&AWV6LE&N0L1+D#MB
zH)rqj$3(_6FV_@azW1zmj7}Znsp9>7#g(~9|A79MeE!TeRoJqhhk=iuK}eiW+!#zS
za56J8a5FFpF&c_(*)HRE>mxq{y8wd_KMOxMm|)-lY6b%9{EqacFFOiZ8CZe(d3ktc
zzyt#)SQe<?)2&c3at6~Teg;mUenA#Nb}#|cpLBYgQ{Gd3W_|`9ptLZHFb9}`N!Lof
z*r3QIugSx}3p8HIP$?WtFmSRkF|dK0_uAu!|BTF;-}o6gfcgbk1h~KiO#k~S-HFXc
zx8wL3xPa0^EJ7S$0w!IQnpUjN$iTpv&CUJb%nNsB9wBivaeuJefdr3`l9-Y{f~Uc!
zp$_BmZe6iogSB8UBTx~85;Qar;mODa4#y4iwD+g(nz#)Xia=pd5W+$YMgxUGK?hR}
z3p1cFD8OK<U||Im1_cyM7#2oAVNd|U4B3(P;EW5`sfSEJKQV}N!Q2gWE~k<(j0X!o
zpk7ee!Hk3j8&DV&W-wufqSTy3kk;a&<YHwI1C)iJ0B$A&FdeLaeVzZs?_4H9VFm!v
C-~#Xf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_buffer_addressing.hpp.7FB2D5AE48874EC4.idx b/.cache/clangd/index/amd_buffer_addressing.hpp.7FB2D5AE48874EC4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c614279cadec8251fed1671cc7d1798823b20abd
GIT binary patch
literal 17286
zcmb80cUTp<_wXl^%nWdYh@gU^+(MCNMZtoK4X|SGu4Q%Y1<R=G+E5fllvP0~qJRw*
z5gUqYL0uIa;<{E)5p`E=E4m7{UoMy2<bD6y=j-$Ao&At=lT31Qa?VL^pO7wH7IqYb
zjz5OXoicXTC?p7iE%|Tetgs1PWkIlZ7lgTqg@d|HyYFB9S<6#PSBMGr(YNlUPjU!c
zJoToJmHqkArDsRfLO#7~+5Q%rHhE{~&-dffM*ML(>`3&HJK-D0KYMb*Khe>4SliDP
zVTU8G+a0>l`rRulUkkfWFH0wPHpDrtTkN9MMWSNMITpq5yB!<-<jLuTkc-#Oi*c9F
zEtr$(x^c#{@x>=L99|u;)^6MGIU!Hw5B)<5<!Z%C`t}{ScK2kL;ZqZ*ieanLkHjBY
z6+5@l%76T$$M+bsJ$yim0j;7oKOeKBeC6Pbr_sAh8)qDee{k1I=Z#Vm`c+@nu+f3i
zrn&P5bZqo|;J|^6ef00zOmEp0O&XM@nNpZId}qrYF{^)_SpA2%PKovJ4VZcMWy2n2
zJ$w8+-M@>~x3EtA$GF`J+n|jd{_x8p@pHndbGJU96cgt3Up;fZr#P)fSVybP|Jq$_
zvp@LYU^QfK$exg<+Z_VR+l;$4^o`@#E%So5XlHEEPP>-e;_z#=OUw#)+aZ@e)t+)C
z->~QWm@DH?UY^>&S!D9r{bll`c?%<ZWp2OR?B|8+XYY!t`1}5=?Gt<UbeaLqckG~d
z%Cib_vTc@DG%c%rk4HQ2-|A{x(z8|Yr>K#&H@=!O^^jkjrf@~R+lMx}>$)HJAKP<g
z{XUt)qVl@`aksSVV7IgjO0(OE3pG)ZyTZ~YEEx_z3=aK$QsJi4ZMQ#eefQ_`H_P@N
z7N-=qu{rqikNeu9N5YiG?oZ#3fBtgf1>YUD;%3k3GpV%Jn%P#Nb+yg@b&H0<E$=;E
zd_4Z(Wbb|%bBbzX|NFBV26oPGHF~phBEN50&xAW5z59s0-+KK%?*74uo_U>}x~|$Y
zVCUKoDP?yul%9Kkc@g=`?7Ya~iL*v8I_+#<{Bv&i{m~UW{EjqsRa3XTjA&NxeokYX
z9?GG6CAaH^Ez&mX<vluS!o455kmu1yI}Sbn_nk6->D^rwzfLMRTr=XA*$Z!0?1|rf
zws*qmCty!m>qGrVrEeX(ZsP2;zWoC(xV;_!VH4g^l6)o6a&laz&dz!0!7i6}1r1#=
zv48yOA%XFy?I!gGZH7b~PYZDS>D!*OHxpx`UQUS$>R~f!eN6PemiEnW+-lQL^HUS6
z=8xT-jtzb5HfzzXCiCllUKVpv_a?;`IlfxkX7ZZbb&tJ{p6hSj>6q1n-flN7^Y+h<
z>$vv(sMk*RO}@<iIM1qR(Skh-|2p_(&%*M7PN%mGooqM2{d#ZRn;rT(Zh?x|xfN3<
zPFm7*fAUI)qN9#6(Ni6}cIlL-|6{M+&Wz6;22CEl-M-6^WwC2&9`-G`WqW<Y(@2-m
zF6)uK-<Yl0*TnLA_7~I6cz3wnG%df|h=#2$l?_%l59qi3=#f_LzSR$b@m;rGF7BT*
zX#JqjbXUjaEp&&LH%O~<7dn==-q_^H9*xBhBhI8poe7=l)_wHs0JodXdp%gUtjW*o
z?hLM*qpRNW!||PoKYCdvj99R4=i0Ly?Eclny_l=FaVgASyh?E$RCDmfz7vw)9*D4t
z{ncaq(^tKwr57EKjp=-&%{Z4ow(n|u_+)WN_p3FEcUtc{SfO3C=xWx^oyjK`HvL-i
zV#1`1eaDVGyB9ynD<kFg;+dz0{2iorzP|ZGhm4goN^b0{F#H%J-aX&+?fCWsV#k)x
z2=-1MQ+wv|gTKz*pXHl-2JU<iR@^wUW$av+pH{3ywbmqRXGcDLFm!kRmZ9SYU$bhv
z=1AXmA=h?ww04`n>QYI%W=!a?&f&%19PFDPXb?HC^OZ)`0_Vk8<~0$2{6X28Z3EA{
z*0NYSd}Z3Cq3(U2W~cVu-E_{TB$SveriM4T_4>q*qduOQm7ITTwnhFlRdVcebJ|o(
zhkmJZ9yeH4&A+4a%eL?gC(8wAoC}(~yb!iHAZ=93t?jL)fG?9X{bOs_OnHdm%}F6k
z{<ayv+~Lnz-n+^(eM<^%yx*nmI5FZ&%XnSQAA@^Li?;s~=<3<6U)>j9d|y1@8c?%s
z*uf<~tXgxVM&gGs_i0URrgxOykCv}`T*;H$l=sN(6O`|NeENdo2WM@PA6}f}us!Q~
zx+`kkcER`i$)O{UxE3$%ezfoT?6=E<vnShr{<_dJDL!x9$bXcE!Q*#7JF~C(j4wHN
z9>*5w>V{`m`}^hj)srr)dHeES=mMWpn=_isT;N%9(L3R-?~LCrbzXWDH5{?%2k*dV
zf979U;J0;N<R6hQKIh6gn~pznZdw~ppE+;J=y4W;02<FpxH|gD5-jWhGssRu{;3H-
zNB}y)(%nwj52gbZfEV|-Y2Unfi-m1wp==;z1HEABfU1O~p~JVSVGYk<VS`!dC*dbj
zk4tr737AOgsns~WUk>tVj)e%bNQ`J3Bl-!Jb(MberV*Lie|z~HxW9Bh7IqOCwN}kq
z4J$tc%UWVpHML)ktKVsQvrrqMLztlM7SV1|sd@<Hm5l#f)?eK^YwjW&EX0~yYAv+Z
zR(2*Fj3*6wH(YVfVPMp1ENnE3G#8rJH;N2d3VtLaJ}c#Z%J+|MSco=@tO3Fr<9m~r
z`gszOX*$g?V@It&>s1qH%@ouc?lpWWKWIn%rOb5w-9C4)u-;t7C}EUoR2GA=q=uI7
zTK{cW@RJ1=wwOgKfKUMpf@M9+s?lvzO&TyqQ|gL^Ze&!}$-+9>tLh<G`g)Jf00OD1
zlV?G!OK#q0ENnMdwE>U`aTYAyYW!fD%}$~H8y(x;GYkv8NL445>L(SSs)u0N(yOY$
zE|%imdaHMz#6ou>bwjRxL)KP31WUh0V@!1gH%%=(yI@j27IvB&OaVfQQGFWNS5?*Z
z(cjCym-McSh1F)E`NDkDApSC$^g9%lxqPY}uoeq(W{E%{&@^4kz!am%_tmfZtB3!>
z!g907647pn*jlixZ`F6fPei8QAJvyPJaLxj&Tb;3y4G|Jto#rxYl14>>A&p6+6Rs7
z60s0MB=(EKe$k=oAy_u^6!uk);~?)fpKL%u2o@sEjZGD%R`zGWLZg}vzUP!15W8%C
zHE|Ch^KR>I>s|Rl=Y6<q=LO=f+U>BAY_1_42<d>-;V2C^jo#!vgOurpSZ^$>FbhSB
zc9EjLV5zh2zsRVAQwE_qVdIb9#zI%p;X|@;NcO9G2$s!#swRHwgXevQWjq{=g;nO7
zW(l)QLtHsFVHa9liwXGN4GU|{62U?+Q4E*rgZda-nq8+$t2bT$_O2xckT=peYTPS7
z=o`%&qjB5aq41lH(3Y&aog&&P+E+cyB9H2Js<x!G6boC;Z5#l?0poh94Xd<a{+v1c
zdhE?;goU2Q86ba#j#UrAvZ+Uv)S^kZ3YuEh_=bg*=BnaFAzpM7EbGYq<{0N-X+ZIj
z<UJOXu@GgJSz?^%%6S@CX--yl-@Nq09*<_##E#^ps_UzJRDRGNul#Y~n3|7XcELh0
z=}a8-j)U#19)hKRV_|QlI;?7c_*?4H&r>C#HIdpb;_afo>LFMPWKxM8Rl(Yo4i)K^
zO`h8bfn<QvMU*aTsvfkt_>kb&NzDRtFtH)j+O7a(_DvHte;lc1OJ1+>b3pY!FfqbZ
zrUVEjMswh-sT!W`^#(q88S;@>(sgE`o<dKPSqYs$s>#}I`To*`<&SKH_M|tvM6`<-
zNz;S&=7^VWo%V8}4;DI+*D8~RGP!2e!??spnw0d|c4v!v?!A7(LX5eoenLM}Q@y8>
zrjF0Lmz^2*qzDU}%o2wc;jluK%V5*aRL#kuDZ7?58n+w^JxF6QsCo=?t$GNSEgDw6
z*SUhkQgK+@$5`k>q+%5zR;gL_Fm`E_sjh2&A$bFv?&)bIwjolMvZa0HhhW*N39<d8
zM|U?5ySUNDa{(rnh>qcDAe=U8hR%9~sg?&vuAXw#c6)?{c(YI;5DJY0UEivzLxmn*
z#hOtQc3~mWEc6lxFO3GSt}@J2Q{i~`_WoVp_QArhW}%_N(8{$nxN;Rg7L$6^uRi5I
z7B-tD{sh9G#%i>9sHvKgjg1akhTa{4g(S0384${h9dfd$>d=eLD<>Cg0_Ri{XA_0C
zsb*8N@`KLlOI^g#k1slsWu9iP;~EgI88yt+wyKUd!j^8eq%+U3kZKmX0E7!hA$_$f
zp?6A2qfc!DzGI;?8RT?SJso*gJ-7+`I=sD~&5-xKu@FS0E=tueN}g2@PAY8dxS0!m
znYRHftDKC?EncU_9*;RfJq%J{ax7?pxZL@1lYa`Iidc;Rnh0QLUP5Xaa47@cbkJbf
z&#haYeb&wetG0D*_0&FNwMN|<db&wtb%<t24fAxVr-AM?XkgYBHB!_>iZ#q;SzRTf
zRicLRCDG!NXk*?()E8pt3vnRhm!kbku`c5;V(1sXd<hI)0y+5z*gFCaV)BvDJ`!@@
z-wGYJLT9s)QuAQbJlN7~LDgbtUkvLqeg#8b!BEEUp#3|({3{Ik%9n?ukZ{DwN1~8O
zo)eHm0&+GlUG)IcA3z>v%}}EyU9{AIalGUZFF7+#lmZf^PK;BefE1p$OQGAPfvh}T
za!BXDzgG&_%gbj;0a-jBmO>Bn-_Mo;vib5Hsb7vXl+}My3OLD^pON~VVdZGj@u+#_
z&#n+%ivpS`U}yIK)C%BI0lcZFMqOPCOV#M%QeDOBebM2**pkiX6VdUB*oyIABK}MC
zVf;aK{J@u2h`53;|1RS1JTHTI88m<2@<>j`(>HyhC9J*{@oTXLtN4wG-|$=l{Yzk5
zrWh}v{|lZYq<{#iBONHKy>G9!iU0OY#_D4kJ(jWA8C1g+6s}<N&Y|v5&<+Kg&9-_(
z>2O5pYTgmlJjE(csbO|G)ys<GWkqM^C4c|4ZS$QY0IRn}hufl-zJYb)mGKidHn~q$
zb6s6scXLpn)^o1s%E(#k?8eAb=jm>C0ab5jZ&yYQwGG`Ey%x2vMK{_QR^N!)H$1O~
z+SQONkAm7Lp4UL_8lIz}Hk#)W*u4by;fK5X3!dM@?(cc7fZZ#2UW~dgM%?!!Q1=L)
zb5T$(>OrRqs~1twMV<>#Pyx?(P|zKoi%?Jz&m}0Rgy%#lI8o}!>Q9k^Q+VDh1@GlK
zOA5~7Ia><O=J}))e3IuvDY#JTY2G{3CsObeo?l79uXrw#g3EX=mx9ZA{w@W7XKa0S
zzJ2cwHW_3c&(qASVGcjk65vt-yr_d^<M2m^#y5*5L8N{~J-?z}^nH;UDfNt$IL=q<
z=PS-sYiSrumcJ=_Wd<M{;cV2DR)p1DgmZbmi10<83lJ{g`3}N&crHS?i02Z7OBhqt
zSF9>SWY&#X*&<ffnPWyZN4Ch3b>>h)y(3%Pk#%fUSI8C>vW~6lB*h|0(Xmy1MKN4a
z+R*o>@i(xlin^(P>~Q;l#h0B&n~KH}?ei7&r2}Cc(LRxqnd_fPOP_eZwhx)1x~=PW
zq3?lJ&(@w@=rYx}J9O-wf4<d!)<%PU;yf5*d^!v?z!xo5V_Z-FjMZ$>CYxDRYUIhc
zO|d-{^m)3F#1~FU#5iqN&~~Lg@hvO=H97V^qt}3KT8P!^sCpe5Vi3mz+1BA44^*QL
zbv)=3T2*ETr~`{tXGdpe_WBJS8#>e1r<nuNePhCa_6n;7qC)}OWayqmx^GGvP(eET
zWR^_n0$M<m1f*h9Zh#hG^%%4`2HmOMqVr8^Fl7m-VCO@V*2k@Fch~}}8KQlL=*%o>
z9JG()J(iuY-cHznc`*B-$A0KbTfig&XdZ*wGY?}Ma@>Yo7^kCF>8LZ+BVy^KRu`qt
z)N)V_COtG|C}<CZoWiY7#?^a6=I$TS;U7^)dl;1Nuzgrq%}%7?9c=gx`qQq^Oa>W8
zQv!olMa>5(FeWYNpL+zC&d)hGdooes9-1C@=73B+01O9!*SxAwxYufrD_@?ZJSJk1
z&>|Ar(T*6~w|EEbX!|q;LYgsVK4>#E?SU*+Q?`Q^kWCBWH^`R_O$Dl6f&3XqNYx`G
zGgGx9LuJg1&}wLRM60SS2ra<sE@;0Cdef<+DG|~?V<v=(4SF*c9Vr@8N>pJr#H%6a
zgG50b#q%17*YF$-aWv2Gq5pf{C#it`6+ACS{)-Xk!$cte2*!gxZ))-^C*1dcw&rxf
zlBH?P;;^Mz&#_*8wluvRz3bC49C9<qL*H+zJ23#QG_CBY{lO%t(I)^;b52z~FFKrO
zPR+3HXD(--j{Sk?!fDasG#^Oh!lt>*v8Tx+vO-NcBRbk-LxC2#umM$?ALb@b8~NLb
zCkU%SnxGoy91r=LE*&(W^<Xs_IwbSUG9Ed^Gsk{p=tGNbdpttPB+k^#tYJ<jsfEC$
z5O|pbWuk(bC1MS#lUQ9UqNSpSaRusBfx0mj{ECKtMWY!<NSz`it~^p28Y!8}f1Ko$
zR`_N8VKP3>jhwqtbBoo6jT)M1Y>&X-mK@kITEVKX#@Ch_c}#4Xb^~yu#?5G%v>Cil
zmkJ$HVSTF3n53MV?}x4FNMdyha@d0EGt09Lxo$&VlqWZxUKakr(w4ZyxyUaU1+bU8
zi2N?{oG7(Ol-jWJ6sbiD&xMj-p%g$(7$(k--xJ0(*d)${DHX`({#ObADm7shCrUVx
z=M)L2@Vr;TdwI^1a2C(m63*uNq=Zj0rrrWE7^bWv?Fdy5VhM~XM*8O&{#NPBhY05!
zkXiz}5f7LSEmB{=ZZCL#54#Z$n3Y$+Zo~s-ycl&O9x&qw)Qxz+jB`;J;sMivK*as+
zLOfu`1*i-0fEnLGU5E$FxCnJ29x&q))P;D!j1#3G;sG;Gk%EW^%y_R9L_A=|SyB-3
zfEi~?LBs=Qd{PP`9x&rVDTsK$R09yvfFR-lGkzrn5f7MgnG{4kV8-QA5b=N+f0u%Y
z2TYk-0b)Ij30^usMx#^fwAh?l{F(JuCd{qK7)4fFUrk>-bDEDNU<}#7i)LhKRv3tt
z*<flOX*SrxswxLeeL$@46GQijBWW{OjfI_KVSl#p_QBA7a3q~gtgb_y*P;G&p<y*z
z?i?-mr|t_@v*gZMa(}9hSk0F^=ga-+)=xKDqK&3hGkss$fN^K5Ofu5~tfnHAit5o8
ze^%xtmH4;w0!V!+YF~;5+I5<PCQH$jdS*4;l<-ZdDIIaF7D~8~=O+?=!t*N$zv8(}
z!eu;{OSqip?-Ksbm?pZ(JDT#_v?Jjj6(`Cp?mr{m%yk7_S3IcAT=BNEQ_Jg7b%;fu
zqM2e#H*c)|3G{!02Ih36x&=CH;jQ9k<gl5yf|*F4i5gJV#cC<il_Gb>ACbdHRG)FU
z<Pa{IdBvx5fA{PCaT)1f8gxiwI}r7m$);<}$y2Q&;ecaZE!%WG9X+*l)1?YTzGBML
z(?P)M4%u#p?9MdioNRYacBdLcHIt~SDRs}jxKH%oC-$WCZB$jCSlE%aL{u62?}I(5
zCK*-LXC3ND700NmKGAYV)^L{WlO=bg4I5R}CtvPJRdsdq<d=yb+)~MCtb-2gpxG&1
z-4!l-wf4Kev7}b3T7hf;S~$0GV_Te`vmZzP&i)*=(Y0y79FYKB07q?gZ8_?o>(GFD
zn^^6r>&Q_jT_=u0v?1=)nqal7wkt>7wB0!BrSIjzT;NcBC`Y~Zy*cWy@9)8E%>exX
zjt1%nax_dk%$=obhiiv(G(tOqqtW`&9yDCS>KOePj(*ht$kBNHcn_Mr!|DY61db-^
zCvr4RH?09%Hq&*}IhvuH!O?8PY)`r;Vs(yT4o9j%<>+U_&z>}#!0H0S0*-z${KC;v
zpj!$W&~=a1aG(q4c^S|x<9Q`8tOTAcHy#NLkvy*ghE+UA14A_Mq@gcX*8;;@p4S1x
zI-X;JHWs)u=PC|p<9OZ#w3~QN0{SH2!5rZ&K);3OWS~#xISuI3fCr7iNWcyB+j-sr
z^gDRo3v_#V|8^hH?c+HE=rVXd2n+{-C*3QsdI%T}@q8E<4)c5*7>*N1p5A*<bATa-
z=M%tig6CYI%jFLe&H~+8p3ec@Ii4>8!zJKpK8I2BfFY0P%fN7%=X_wu2cG7`9rXq<
z+~BzY7z%j41N3)*2Q%|`f&MPf_kjK$&kupiL*CJU1Y92R{1~`A=J^?LdB!qnSbYv$
zp7Z<{aQTbp*FgUoc$g2<)Hgu?hUd3H|CZ+uz~uw*qGvl;{TI0W%kxLz@{#8+z~u|@
zqPY;Peg!UHdHx1mzVW<FM9cWVWx0r!^Bf_f2%e)v6vZC|tr5{0o})z+&GQBkZ4fmy
zQHa%zBHGAvtcYTH&J-OpS)hp31ES*ro^!-jIbtB&sLqNV&Wa(7uZSJ4@O(?`a!c&N
z%8SG<MLa(fgP!plU#Zxtl&}Aj81#uR|0=fn%9k&Nt(HQrKPzCX6+CZ(ftz43+ekOV
zz|A}#fOQT)9oty5V4W<Ue}lfi@xj^eu=($>4U@kHn_uJkPw4k2pDQSWenmY01^xcw
z`8D)=&GRSd_lf6k(C-`1%aGqPmTSQ3D&)6{=k>^KJ@R3W!3N~Ef#<!*YcI>&62}60
z?c+I5^2%fW9#$_)UYB`}k-cN&X3Wu8FMF@&d6SGb$r@IkBv(t4Ycf78`yZCuF~{Vn
z?0=N!-(>WgtYPK3a<yE!CgYp3`%Sqab!4!5OLo7-^F0~elQpcoNUl~S*JS)bcKRS2
zSc3VZ?DUc6FEaWfYgqYWrP^YpCgV+t(<a5h64LRCQ#{Yf3QATqtUOhzma5cbd{ogM
zRovOJRJNkc=J}YSJ;w8KMSGm*97UVM^C`vml+v0eNU?fG@jb)yS;hA(&sP<ztBNg^
zr#D4NXl=SCLS0n4ffGmHbS;Fss)<Qn#S3z0`~A<^Eqd5NYBXIUp*1E;?l*i+D)WfT
znJlAZxfKl^Qb(q=XnW&pEfNkzLR}<up;jdI_ScV3r;N5D$)=H-k@n2tx&RCpfH&LB
z3Pg(n(UyLVB&?xt0Susvm%53hN5&f~v_HF_Tjiv+-r0lX)Lx1fFGVg<S`00US)!CU
zLlO>`YSWk_J!sai8yZf!Z%rCPeg%9c<__|^!*dbxE8@8X`IYcofdVR!`4kkZd!-h8
z`OHn0)FO-LY^g;y&nKl8CwYD)`MqM99jul~eq}tDOMc}%f0z8eGu~G|0Iz#G!$4-}
zq3H0C<*+mVJ@L+UM{v7pm?Ri&&Wp9#P+t>muJL?Zw7JdmL(%3T&;N)v|M0vH+N^_I
zed$n>4(*s7&xIbjunC<?OfDpN<ndepy$WD6R(=n9-Q&3!x)eiiR{jpUzJorDQ;~lv
zI}V_Cn8-RY-CUyU<dC~%|8-Zl{Y`Qg+n~cX=)`)Mh#V4;6Sc&LhK|UzegCb1jA1eK
zDTXa*Q-@3EAFjRX9a(<=aa2WZ5b1;Ih7?uU>~W8-bh65O+!U)Cq^U-i4~gJ}T2_o)
z1y_!I1z(O@39UG4FSO?<NC@JnhtPwgzCvG)h6qDA8YztAXq+&Pqbb4^j%EroIhrTT
z<7hDu76XnW0J*ZuaTFkT?l@i#$o*E1;{mx;#ql;kt`2j&8;~oS9A^R{ljoyAILh-W
zAe`d)cOd-E^Hm^R<@pv6Zt?s82oHFE0)!_#mja=b=XXGO$8$Ll%6VQw?jnm^|06{q
zlIIvvh~as&C~W3ALliQ2j)OuR<m$_SLI%$<NQgmPc{&o(dA=wK7bUKItt_mSIX);0
z2YJ3B3paRvCkyX*-lzy074G|o6yXrhCl%o&V;avBlWx2jNLL8mZpoib*8$B1Kif>G
z)-W%XEWUDt%TX<A%aOWV!pkKy=d{t;maP1D4O!>^uOR;a1W#PNHF#g^vZw#KoT*_T
zSwYhsMA|1}AY@c5J5rzJRM$xTtE<y<9kH4S>m@=xGcid}n*?1MZ-Kfk(2a32)F(rC
z##^C&E6>}Y^ET+h%2S{=g|9Ca>Qedo(x5(#|9%{*7l-upO|ZHNX*VHP#+#9DGjd}b
zkM!}#opA!vC-9tzoD=yo*CeD(;_KUjbX)lP_9M6b$cxSyRx^=TCTh(10P;S7xO0j_
z$mI}!YpfJ`mZC;XK3ejMmbmX<mE5mN-YmI#U2?rHc~YiM9WltJ`=N9OFBIIkeEG%u
z)Bm%NbgjX!U+PulekHC)drf;gmK#d}`V`>FvJ%%shifb<qe)9-ReAfAc7tpO<mxmn
z#Hs^wsKF9??#R6c&1hb4<F=qrFDDWU{>L8lNB&ann?c?-WEosmnqOAhQb&@wJW9&~
zr32+ZC*<$=>^=0I70Kwx_G@KV76HV|_2cD6j1S882W400;h&Q0pOPCfz9BchAqUdc
zfYk@G&jYzF<1*Q`Om0DUN34F6eZI+U8OJKFu}TX%S!AD9e6}fVDc^aY*k}BvRV)7I
zxuSP9$>E0SLN6QkJZO=}jz%Bso1btyFsm-v`5(#Xk=&YY$@DanSS8bKUs?mrPLKlA
z4PSFXkl*zH&sFDwv1)^CtmzG0veRp7*|Vv2)wtTzsl{q@=jIJq=C-x2wI|EY2I>Mm
znU)1>gFRT5wWq$P7j==b+E3rli&|H#4%H6zU>VU-`cYoY2_L5!XU{U5Kk0w+Vl-7Z
z)sy8iXBlRBvwY=z!+dX+lUxFHOMoZKJFWnR6~LQ@C79fhF{}aJEN{3GXg2~6mKRI}
z`b2VeM1NGIrUHE`@M0@y576xap6pyY3mCG1H_OW%1BPS3o8{fk0NokjNhh6L{{V)I
ze5i2^7_I?tmKVDY4A*(S4fMBx7t4bc0hc1+LsbB)Pl3x*;6sBXtiA&JSHO#E0#@Gx
zm-oPj<)J<Um(Re5<(a}o6fRn`1Mt-%S}j`BGhD2$7twmrn&p8KME3--F+Er!htQ(q
zeo@c-_xqy7eLi%0B3eA*`MGHEoafh~-D}Z_egBPU_lD;WqQwVZzFf2@=lPpx@r~zj
zXb}#%&}}ueTMfDTqo7?B&ugIF8lIz}T{O>eFfa~wWm{b)tdj{1)CtGxJ?MLng@%}%
z9Jech!L0lRZ1;la_psf2o-1Ix3Z56Ec8d}B{Rq@9g6Ch6+pnk*+Y0v}uRZ)7noE+`
zCCQg=H6(8*TWyr>nAJQkTOF6}n3cROTiurJm{ly7t;%IPW(5-!t3<_)T0KnsAnjqr
zgE~f7y`orMQS7MsrfZuRH{;I-sQtm}C9&2ec5?ghi=A8de^xdpfp|W``N$j!WA!7#
zA9+rfa5@W>$?=_pOQqWM{ph+N7SHr!1lr<b|LYw;?sWkFxiis~L25AlQi0a+^y%2~
z2L?ov3(RT?bVz~qm<}W%ha`6T`s~EJMoG8h7ZW?ZR6$GGEjwyv$y;i?HO=Tbv{mKL
z9oY7&)9SeE!Hu<z%~yb_RUpT=re8qNz8Tfd;vV#5YE}X*N_c(&Ene{a9$LKTxdK{L
z@VppVEJj>?5y&Ef@vG0nKJ`e5@T+tsEfZmNZY-d660~Q$1zK<6IT>0f^Sl*WZ{>L#
zwBE*Z3ban)ITc!`@|*^((|C?UmT{;$TMwI%^(JJ`cr&uz%yT@lj^{Z6StszEh^!NN
zPD0j6Ja0kPTX^1&?DnHN%oWT;wKMtS<O8Vo0iF*b+e65K$(N#9rKld`XsK4T#PL<h
z?y6LW>N`0lmu#;~4wPxiimW%&&tI5^6hMmtHp6dcdik&UV?zhxA)ZF~G{3^)B^=MM
zu=^6e&sG@i83{K`zo20)ELYHSr9L$k?-GxlYLuyS`Og*ien7o&an=P@M&#1JGr7o4
zjXxr{*qzBmcE+orGr7plI0`zGi|mZoKxcB1opCgDCKuT$|2t=JU~$g)ImDMtkn1PN
z9?S|R$u3E}FL_$7e_HlnRxww0$>n{?LfNNK_GebINOmcb8!`Sa`+S%E=}=&GvEs5=
zX~Z~1@kvqqnH5b{TvC-rl<91eF*g1Fh>h_s1?^JmQ^ByYlSdi7z?orVX4W?T&s^h&
T>7Tk{d5zDF(`>T9T;cx!Fd+vO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_gemm_dpp.hpp.55564336A8EE078E.idx b/.cache/clangd/index/amd_gemm_dpp.hpp.55564336A8EE078E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..49910bac90da4d573f8df2a1697494a007b4100d
GIT binary patch
literal 1586
zcmZ8g3rt&87(Tr%ZMnU4y`1(I3L^pfrchf}1_XnE#qk&@nIjXnF<!19Oj@GVp}4>X
z3L9Z#u!oyb#62P#ZeU<IfjK6fiz{X!$`H4OF=m27L?X`D>BR#HH@V-v=lu8o&;Nhl
zce7HGll6Q6k_uCtWkr>{cmMzz-HuATEuRCxH7)?o_dA=vDZloV#QL{xTjjeYqXH|P
z_~65>va%O@&gPVM3G(lM-#XiEt*^dVv*`p#i~7uQ`_@FCw|{O~zb^aUTV_RO@L*Lz
z#YW}c?IrsL49bhY&P*RqlJsZTd#;`v$_bbl%NygrE@&`0tHwST?rl3fkv+Y0r1XjJ
z!ui&MxmUrF0(YkT#@VSzZe6>X*m=grFL5O<*;`DhXXayFwoBDzWA1OK&Rs8VxiZ%n
zWwRapv955-<BuoX(to~tXI$>?wSUr;dF;_tyPJ1RsLBn@*l}t;R9O1NHr^YQxqIl3
z!jLzQ&b~bubabD4)~8@|r@4ILa5w47(x>Y>j=P$RgLuxR&7DgLUq4eb?tHA<*k=i1
z$R8A^lC2%n=Ovq#PB002|0t<_Xlq3n<8vXh)EFJhJWRf9_$R4SemcD8(9>i7FU&pv
z%j7^;f;jkYQ*LCy$&SUL=J<?-mo$zy%JQ9vX?ox{@s%45yJdu20-)b|KP?MDh}1xr
zI?cF$^h@<s5g}X$pqA7^7I4_Gy)O39dmvjx_@Hb#DW_S$;c%DOSfNw(iU=BIt4Ng}
z3pgBkApYS*_#VB8;G?jbR6`bUxQ2KAx&6aC9*GD63TsFWWC4e3t6QfO^Pj%VCw#@r
z%|fw2ECt>?G#21MPQzqjRlVG&MMO}jIfi6t7I6Lt&e6!B?^10Nf<)!@QoWo793Ffi
zv<{DirHTj<%9fEbngtvl`C?0L_)mkwp3}LI7SK`_aQLru^*fDj%>bY976TY62$iz{
zyuFC!=Ilkqjv4fh`R9-<@3<`Hoky~K<8l-<8_brFrXJo`S(0rOj7f}1!*ftDE+j4t
zudp(xu16MJ7c1gLvP!JhRV&qItOo@(N{t!QPeH9xYsL*x(4w(e@F}M0(-#8rhWBuI
z(4w-a@is0*kgTAqr$i_sLh(n@iEW9O{|sKYCO6iWXUNWcZHa@Xlg6m<*VgXRBW+Fb
zJp=_K#1Z~@n}SBMab=YocM^W&HB-K(B%TDd<8jeyJ4S_Zr5$W1!ke`bFQ+?w;Z9Z@
z1yunmEvBt*&KekVyHUZVAT6T<mTd<B1!Z9}6Lt&*nJ~tL-Pe%Ht(|T9HyWk*$@~Jb
z-V`+W8de&o;09*HdaNr>McR2>i?zcuqiv5{@ff|u_ijB)-n@?p)ARKI{|W%~#|6_<
z^@|4)VJ4x;7bheAC<}vExMft*<8XWuej>s=&R*rhv%*6!9!Tj`-n{ht8nSU)sHDgC
qIM-hE$}Uf)!(QMpdR}B6U@rnP?K+n+wNJ#AGBUv6eE?R@0^mO~)%>mi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_inline_asm.hpp.7777A491B073ADB0.idx b/.cache/clangd/index/amd_inline_asm.hpp.7777A491B073ADB0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0515b5691dfc453c2f3870d1aaa8fe3d3c46bbf0
GIT binary patch
literal 6854
zcmZA530M=?769P6$G9*Z3sfppg$a&TT;QtDqL7l*C9(Zp(Zr=vsjsbyeOeU-ZI$hP
zzSb^S75${FE(xHrh*|~3vV_G1(5GRsVpzmimAbdq`V^skiRmyW<%1diIg^u_J7?~_
zGxOfkXv&8%%*^F)ZCVqt;X@h2Fz&X$=nYXT{lT{87-myu(Vq6TSrgs1mfWAePIqMS
znv*rAdlyFDI}v7b>wmV#FKM;z(x>l#X?omwpxF1(+6(Jn?;1WPt@GON$)koZ+!o*w
zng8?e$@><b@maol)GMPD=lF-cPwG}ZeC&08^|y_>qf>mYpYZ%`xO(rDu-=5;otKVJ
zzrq&zUTkX!tA-%+#31trK_BTR-`=9zzwP129lhP({$Tv9@od*}ACF)4EtjJIak+8W
zra8q=GCQYlaqX}4c%nV#w?Mb?=gHl@3iH2x&mHhRXY@VSxpKU+H*Ca#XwDGLWk;J*
z&mUUunRI^Hh#b}S#>l&IpJ=L2Jjn2ln{itH`3v^NS(l6XaWAL)uN}X1T>I7J>mPqQ
zZ&m)LGdsuUj$hW^zUI-wuf7c$DtjUEZ05w<-}xmPda7qXsR}(6rwjQqsbs_#Q7v7c
z?{C}i@sLZ=gDP(@^<+;!zkxTeA6RmV{Uv?)*va!YWdC=1nrg5;#n#XjeyP{qySa2s
z`1?)9f>$T3H4N}zr+KXN>(rk1U9jgB&ovn-xwkjR|MV~<^6|z`lDx;zM?WqbTD$Sd
zuSeI01^qF-`p_M-;;n-_AKkfdX5*M!kK8ZK{^ilGeRD^f-?+bfREx*BxduP$4-r0V
z!gqaLJWUhYJNK%4vNk^a<gS8q2iL0eqRP+xZ(*Uqu={q>ja_Z(FAKHm%d-zn9XoE}
zf|7*vhnI_24qY9vx%Id1A)!YqM#qOI7Kcw8Ib?6=lB?HeUR}0idF8)$Ole-tL~mZR
zG@CGt|DM-ljO%76am+9uhA#lO0tonX#kjoe!i@ALJPAJa{z=_{zYlTj<n87iz}pv$
z%Zo7L`3`Tqwx_?TF*uiFUa@5fW`cd5FBq4X1`L0`!}P5AUr%^Z^QL15f5zX3w=WnM
zH<@FHnP0rRJ#&#d-0>Af#I1+~@QxMZ;@Zz~9Pj;jQ2$4*Fwn767jf&d?c`W7E}kzq
zj%91agtAdD`yO<Bm4z@C;=?;ujLT5g@pWfPG{Yv|dMD74X_4Ea0N%b}T-^IRvXl-!
z^3L$q?m3Puu5PXYynVsA3?_~Y7xc-MZ7a|2c4UZjiwxlH3&v&0Ku3lLpPd-expBcN
z#}1jqEz@?ejuqqLG01U7znLfBOg#4be8*0S;Ff4>e#eS&@!}jiMMbQL-Z_+E`bEVm
zf~s$-hH|#^QCrB{eWz;kJPUbOato2SkTH_45cw5sE0J5V?-F{K1opKre~-|6*!Ky&
zkNtqq2iQ-E{3#hDeZHH>yRn}U`7`VuBJaWOCGuYEoj`X&U|;L-NkEgZcLCjny&LFm
z?BhU>LttNV^WOpe4*LYq6WAw#o|NoT{C(k#2lIBY9G?OqDKJwykW^5n!Ys*YpiIL~
z2W2{T1}HPI^$?=R`3#^m;CL1&v#_&4nT?$T${ge^w?5LRwG{W~Y(+y8VXAas+kkGv
z{s!nb*xP|_$KC;ShvdN%ze-<~JgId6$L|5JJ>VlHP6n@J>{H-%3VfudJ`G-{u}i?K
z1p5qlosm5FKQSJ<8LovdaeOczJi4zV;e*EoPeKX|4v`c|BLn;HKYkruhqRuqM~b3R
zNYOMJ=`djr6ZgK`ls`-qhsg}dxrEKd@m!+F#qm7C=HYlAQRLxxK4J55JfA4?alC-A
z1vp+n6a`ZJrGE~;_1?a#UwO#*QV>cZKx|<dUj{-Mb~y;;*cBjDU>iX&Vw*rPVVglP
zW7mREi(Lmo9d<nk_1G2=EZ7YoG+;M^(1_gxLKAi~2+i155Ukj(Ahcq)fzXEC4njM2
z2M8UK#dag?pX1uGed8-4&g?1{M>Si(KAMS}#WJ&Pwhv_DW^n|wZMKhG;%0H&vTb&Z
zR$@Ri_wMB+&&*g`VTZ^=rifWNUMW{1spKl87<tSTsjy?^u}E?9IHY)aJW_%@0qHQ2
zA0|_9A<J{I^N2hTJD<q&u?vX20NY69MlwY_Q;x4B@=EL~BCo=(Ch}_R8X~WeJa^mJ
z^F2Qru6b~ODQxa9h0XEKh0XEKh0XEKh0XEKh0XEKh0XEKh0XEKh0XEKg>4&UoeG=d
zoeP`eoeP`eoeP`eoeP`eoeP`eoeP`eoeP`eoeP`eoeO(@YjK)!!53Cr>B$DkylsyH
z&ws)mM97Le7T<Z^zRFUVCEft`_EnXET5>t4%CRdzRe@~;l@Z$nN)yg!29+7dYe7|u
zT?eW<?0QhuV_QIJ!TB0M)qvxTplZZ!0#y@sGpL%et)R5xe665r#ql;!wPCk|svWxn
zR2`DV*~~!p&pD=eYujf9vJ@C8jfr}Y>9I>eRth7fhBkrBglz$t1=|WTtK<)~R|o0y
z$7;Pe-k<j$E%rP%jq{&`6zCr!DTEDimpYjcMaT?EN>+(tN`(@|RICccR0<V}EuxD8
z`)0~~7!5;um%fX%m@Y;Nr{PF3Y>d0qsl_N_W=M)<V^J(t5sPARY#fTkDdJEpo{dMb
zctt#lC9nx7mY_)Ji$#P^FWYkb;8Gv^6{IQfnsfzIfu>@o0ZqeB2bzwZ0W<?!4^)qB
z0BXR_0-A-L4Ky1&2WSp<DbP}QO{$<Wpk>(QK+CZ!fL34|ff}(*Kuy?Ypl0k^ptaa_
zK<lvUf!1SNfLgE{fHq(^0&T=@0@{S#473^B3e<|-3bYlw4QLy7JJ5FQ4xk;9*H2^f
zwk~FBd}Ou@JvRlk;)Ry+sh~{-o#ZsoreUXpHXS<yv>DiXn5)P644^gOcot~0u(Ls%
zjhzG99PCoimV!>a(Duolt_<c$E(cvXb_M7vu#KQIVw*r~!uiagGvjzI=xVX+Kv#!d
z54w763urAkUjyhGaJ&(8jo3|~Yr<{@T{E^7v{sz26?CmQ-Uhlh>~_$#V|ReALvq62
zIa|Zj+poKFwnuB4Ndl!=&t5|JV(%k#ANGDi_hTO*^Z>Szutwr8&3cSPVI(soR}!`o
z$191V6343uTZQ9QL{Wv~)r76a@oJ)|#_<}$*5G&zQPfCr@zLHk_&eThh=WsZ7yEs-
z;!x=GXICKw0>o11Y^5)xVyA(ShMf*VI(7yK8Q6Ld^w<Uv4A@y9WMOB6kd2)KLXPAk
zPv7B^Zmbt<&#nbP3t*a9^BjK!=n?EfpoQ2+fgZ&^2K1QZV~<N8zIyOY)gX><A)YOS
z7u$;CuMp2G*jD0c#ePaWpAudi_&L6tcy?nyBc9K&dx&Qbb}#Ylm3(?d=lNM|{WXr`
zT^UzaoS<>MH{*>ofti35!~`KNViqAqGLcA`gvlhxMT9BBzDSsh*j<F_!cGJx5zyxg
zfGLnHj{CL(v%eV=%S(JOX4`CkA0}?DKWfomyB4;@b}@?xEh1B;`=prAV(e3dp29v&
z=xOW{LQ5oH@Y~(v-Cfl?+WyYV_KV(WS;%+``I~fcEo7R7ydk-Z&@M7nx|lx^`V;mo
zLT_Q;CiFJ;9YXJ5r@*WfcpKkN>Qo4ooCfMN>~v74V`qRm16vQX^f;dZ)CL^S0(BO4
zHmI|)b3mPgT?*<_2$k-lGSHO49LeRNDaWn=O$D|QG)8O_s7*MZ88l`buLVslb{%Ny
zu<Jonk8J_91?Ou3O#_ZMf~FC>2{cXE&7f(<wu0J<^R<Ge7025^(}vv+ns)3C&~!)^
z=cN7E=kIaF!shrwA}b^#rE%prksX(Ot#{Ic-M_lEvi6w<?r!Zf4cy(zc<1g`#yfYn
zGTynnmGREqt&Df>Ze_f4cPrzayIUFW+}(;NU^_p@OK9l~O+;oQBc(gGj>zgHiv@1G
v;~g)i#Sb!35vxD6cWTj5%c2GQKfd;b<3}$3!!Z2js(?`|CiE@=Upe!Cw?ENd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_lds.hpp.3739CBB0B218918C.idx b/.cache/clangd/index/amd_lds.hpp.3739CBB0B218918C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02f003d38cefe7070e533444b389ea89a37d235f
GIT binary patch
literal 1316
zcmWIYbaRtoWngel@vO*AElFfyU|<jg;^LB`%s-3_49Y;ciaBd1+2`Lj5IAc3JNV86
zMa?D--&=f60t&9j7I+)837mU$Qt6C6ljlxTp8D@{nw@2j(mpE%JlS1te=qsthCB6Z
z53N&gp7lVIdEck=580;u-t8sjSpS+~Kil8pjOe0&A1b!qxgQ>_%~}(9=z48|=bzrA
z5C8QYO}k*Kd8x*2UEI=LDnI9`+AZMTnj7t(ct!1CzfQ-w6;Ezu7(FbAwV5M$?||i<
z$5*Y_Y5bIVSyAOzY*Wo%6l<fID$Vmn>CctT715TSx@XiIHcX2)+W5Rn=7!j@x_Pz_
zpZ+voAoxIo=bXa}59X?MoX+cClx?iwwzazQb5<tf$8Q%k4;i+%CnUP^<jHoP{BCw#
zj$LTFyuvZBIT@_sy{?zfJn-ajzjs_{S@PQ`Ddn?Pc`jA=lbv_Is&BT<tU6`Aoh7U(
z-Q86<q|BQCk+6R5c9yl9tqhlKnz??fUasHUC8jl-I;_{0*qzaR_~e#lJYQkl(YENK
z-JbPX=Er6SotD&3vof2vskXP%O@8C!cJ0}78Wzt}Kl8{U)c01=^f||6(i)vNxNcs#
z>PzPg=Lx4i*K3Q0rbd>91>dX`dGU1Ho44|Q47{u6PT!Za=H1%kQx{b^pPQhyH1mqW
ziX5}ll1Yi-i^NR(Dz7K6i@hIzWBy6~H&$nht8&H7u1oAnnW-N4huLD=gUQ8}xk(Pd
zv>>y5#ZHGgtIseo@vs58K)}h(#VyRh3E?=Zowu8Ku-%h~ftj6wQ$k!q6HG90GBL8S
zGB9#77Nw>YS1>U!Fc(!@TONL9&&tCosU>L$*T*TPC56P3){;iz$#KaEB2+2~DM_w5
zswR+QByg11pkr3nRKwehJe+DAYJ4z7{1R833@+cxW@6!#b&z$03A1o2dnkLtcoNq)
za{O7o-H4HeQ%PG%2WF&_{g$1dEG}@N30voPq%VEhQ3$k2SzXy2W(bUC<N|vb7EF#^
zr>)-IFP;lhEv&2tQ|0Mas2DkeX%k4;QrQM39PxXG;j`V%CU~7YXRh|6nnR`Lcol7Z
zUA0xb>(eSU$L~mcaK?q})I*@X5)2YtkT_#t*mL4)<XY>`PtjB#zjX2R=~rJ4;I-+z
zjap5Rzsxi=y;t?$+}<3TFd1l*0=I%B%!k*Xc7!dvtjP!zmgkdKfeGLEGAAfwMprvt
zn;u%6cmBRcf)UN8Pjhze*WaM1f-d~(Vb{~650`_27%Ak7QgaeP-YzamF4hOqOq>ZX
cfADiMy~t<g;NoWBh6M=&13YvX7&s8309SACRR910

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_wave_read_first_lane.hpp.F8A301BEB898DCEC.idx b/.cache/clangd/index/amd_wave_read_first_lane.hpp.F8A301BEB898DCEC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4337c456aa2a774a8ba5fbf4181be163dac067f1
GIT binary patch
literal 3546
zcmcgudr(tX8b9|ElMpTmxjd3X2#=7EAP5vFAP-ZBaSIbuB~)dEtwE{KLeXF?5ma1T
z+0shK*0!{hR@+(~R-MIBT)|o`M2kFptsPhQAy%QaFY1H3UAN4hd+F!R_8<QvGx_fM
z-S2$oeDCiT7tWh^ArK+Uio%B4$~8_ALP$ye^=n+!ODTl1$%_Whe#d9e(J?`zj-SqD
zOe}kO@1Cs_Bf}2GFDNzOVA24+sdV`8t*e92+4j~RlFV194mIA1KC%3)V#gZ=*_!yq
zgiKR>qib|5rNVFLtoil%m*uus*56X3eR1aT*o<+_Qq9Db4+q%xW8W&uE?53etq922
zv8&%<a|{)ieVL?AFZ#~eqKh^yyctxe?(E-b_<Z1f?}`HzY{5Bo`#-`?ro`*dja{nP
zs&5-wvn65KhCt7xB=SzezRS_qnv98i4}JEB`rjWFyS4fG7ju62XXIuSu=|y4*Uj*8
z-FETGu+7%Kt~Yl-**}`G_3bP7bEI9nw(f3ws%Ijn&e65#^|c+w;|E=SB|8eUpUAz~
zG;6folcucBJ?&oA)}ZNo&Aa?l+h$|pv5+;HcLyIYaz6i2L8>=UcF`RY@|a_V@#=WG
z%G>rs<(cKk6_e9-{e_7)C!VS4JfgmkOf6}wMfL95iYIB}@3UXA9phI%2%rri2*=7|
z<vc-B5(n~|6IO8{di%x+1_cBm><o9N@&x&dXrz-<n6i0UTihozKMJ7ZLJ>B|4Qif%
z{KB-#+h3h;FPG2??s!#fBu{|tKC<%XZ*u3>$mu{LZHzI-@dViG#V`EWx2(`hlnX<+
zR8}hI39vVQdZ}yuRm1CII$9~kF_IXTC&1qJ+R>5PUQa8LW)W7&RC1mG``zmUU)Rmv
z-pon+<HeH2EKh*F)6lc<4lOluYzfv#G%Qbm{jsDcZKUHJaR42oM7UgB9?TQuFPlN!
zMjW%>Q+v2GbETY<7voS#D9aO&|MT3I8H=c_q%c}REYAzh3+D;2Pn>F-(ZA>CGgE%A
z4yra`jv#+O3NhUAz2EhQXH8NKEUhB)iSdcaJOTNW$6Nz1&C#DCwg)4um1*TX0rsgI
zXRL4C|9)i<ttZl!2#cO4z`pL8-`3io>r~URgk2O|q~Qs$CyyOC{ZhrrGR`zUOAuxy
zEGyWKT$NAPxA-IE=W5KE-8W{Y8SG@8_(3*dXPPq`kZj+N!@2KHj*(3UjUg5?iovDI
zQVc1eP+u%a`Rl))rA4^G)DRE;5#eSkp_xhq9Q4bwMY~)J{^-YGbCNj~yv$&$)mjKv
zGuUpjCxVg6;mMvy7T2t%7+m#O)nYga3WJvM1p&Fn;1WYg{NAc(JFX^Aw5YtTuj<C#
zZlbz5*PQQjox{2Wu2Ww+ApP&2zobe;I3*}W36UcxEJY=t^pNyW9~7Rh5l}NFZKnJ|
zH4)xKF`KAB0dJuKwonXUo%vAr?W}XMX}d7t#504d@v9k}3qz`+K#!!bq1E$x4(3xL
zTsNoA46h;Xelot7biSW(ys~H8VFtSyw;nbzB(R0eloN2le=?eGyz%#Za)nl@#gN~S
zEMghocA(ub*py_-5|Aa%k^<pia8X2297LbNHoYxLKnoHUJPHWpMI;4C=MqwYD_3AL
zI4>s8Bt+g4Yzc>}9A#<UvtPWWo55hC+86=dj$~u9H4Qcq4V1P}xRb$knRPi}HG@~-
zRa!tx?zcZOc64+4Lu$a;opEH-<=B2M8iuN5kY2Iv<t^*tn@|5H8V!BdZmbiXA11qv
zfkqZ|gr~oZ@5q9+Cap;)ATu^c3bva~=2<={!<-`^tJykNFxfiaS}Y)Yuw5-Uz^=0!
z0YUjwGQRUe&xTNt6yn-Glp{0M|EGEqBBdHy5)HhTe|vA1EWGz3u~Vf~sbM(4lU%?}
z2<X~y5vhJ|GQc6s)QI@?0?&I1UpDnVpZSn@z$77OabrX{i$lL2g<zf#i@0GTY+9>b
zI}<|4;0@6mW&r|S$+_Hk5kxoy=RyF8eZE=S_3gSYRqO3PO<RjN^7P2iT-m@iF@qDO
zi4W8x)g#UGK^CdS2Ng++e2^14WndPAt5Nj>^hy60rTcU4j82<15CY<QZY&Ee?EjnA
zg9a|+hqF)FFlH7aL%IQDWqpc)v9pl=(B??=xET(a1e+<EAKP&5=NHZo4t{g@#6$A&
zY;Igr@e>hmO`Es=cGpu})!y3j5WTKtoweNCr>3S&FYq#PBlnF1ZoDy)Ra55m+wq%5
zdE3){Jh?MQKW*yWQxx__b%cg+9~V((y^t?<Y>x!p;aiOM=fA;(CQu$#J`<V>`5vor
za-~!6T3-JYx4<R&OOI+dc4LV;Toa*<j6&hu>#1KPED(wkmKtHH7M2mhQY$R~2cg2r
A=Kufz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_wmma.hpp.143B2B78EF519046.idx b/.cache/clangd/index/amd_wmma.hpp.143B2B78EF519046.idx
new file mode 100755
index 0000000000000000000000000000000000000000..145a9262f3b66c0bfccb8ed74264080319232490
GIT binary patch
literal 3650
zcmYjT4Ny~87JldX%zMyhUT{&a#!X8g0*hvwj?)DTb-^^$c7}pKSpOykfo)K<Cc388
zv0DU+Yh|^~uJTi=)pa+W!6I5z7GY5ABD7d&h1%`T8f+H}GEGIhgQDzxc{%1~Cil*F
z&$;K`dw#z2mM*c`7HS!0asCo_alzJHj$xPx`gd)0Ix2u+?r0dsU0eU&y3Hj|MwX8!
zoc9++Z^*LEY&<lt>R?Yn)L3IeR^j@DX;T;FP5gM#7+%q2etKtZYs*k_v8$nOQ|0*m
zD;=)(tv_4N72nd16gRkbcEvQNB_G;#U}NRV{LQ=DH@V+kRCBvi>hnP6u85qH9Wz-|
z%kuQj*Ni{km|Xt%@2mEg{(EoyXyxyAU#!2acYShmc*fUP*_Z6yhu5@g)}5W!Q?t3F
z-Ev~j+{68bea=t&x3|~7_QU<*wAbshEBlv}Wb2+WI9v8Le{}7Q1-EK<=hh7kkF9om
zdewaDouYl^E6)_(T=eJOdp~xJ{BYXww@VvFj@`T0ThW_zCh?s^PYo^>q8#bh(+7nY
zn@f!A-V<{xujMQaySO02+W|4JE_o`G981Z6&icyP(&sW~eSSt47x7_jcBbwxm&ej8
z_>0nxolib}A=x&bqq}r<GGX=Jq1DX~2S{^i>xus4O_tLY&!*eA7m22{!XuJ4u`w-n
z>{l(i_1PbK$lmgopIMVPI`6SLpERDh%dam_TI71YIPc~ZhM7|G0(<_;EBzW)FHR+2
z0sj>w1SdwIaic$O@O({cZ_oQ0_6d<EGx!;CfdR$5efH{za+h&B`<Rv?1E3!OLvVVe
zaz5(qZO^yv(6Dnv#U@GP)8d%0uqksH1_;EatTD8HoBj_C8zXASDDa~Y7o5<>+~u2&
zpBgITWE)e7mAA$P1{BBpVpHzUWG6N3Y&>EWzbZjKA~tYD!oR-x<E(alj)pa$LL#3S
z&xD0@rhtNFbWe9g<gZs~SR*Ppc_+Q3z{Up@e*5OB&zYa%PSb}@XNVW{UiwgF!dtA|
z8~tT!-sEzsGLs>G(EF*1GNH;Vf9hD-aQ_J>Rnak|4fJhPMVU|~_pNT`)g!|rRAmN3
z+<LbmFrYZU*km4if5#hCJdGiCz1<KPP`v%GFLrNScJQ!<eVm5tIPk|IJ~*L@`|NDe
zbmiSLYT*|QA$np63@F}t{q%y~@waI-^yrCR;Cn&k%|BedoBr135AMnCo=Vp8YZC$k
zTB!WZmXFQ5`aaUI(P*I(_)4JO4Cg2d(2!Tv?HZZmcs^3js$fn`=cfl0;sOfQ3m0#n
z(j7KW&WWJHFz~~&4U9<rvMt|UGrmM~;+OOmEx@<Xv{5F!$*EWm)O>cbNxlhB4gi0E
zswfkx^z6^jeLeoqW7HcF46*1fhQNU0FTRtCyd^7Is5p`#9?*Mez?2D9`tGb)@SVx|
z6IGEIlA+Hq1O^oQ_B14|U3F-q9FgEljskxaRLM7-m;8Ijg$j_9PeZox+u{NPTDTMG
zKiKnMVwE2nv~UjibFwe6g#H)ql$vhmjmY`U6Ekm)4-9C*Sy1RIrUi}l4&lKnZ>URb
zD`-B@#Bmq)-O3j<29X#AqZU&{Br!sa3dIVsDwHH7Y0-TmNfwe-XrZuBg|dV!ExJi0
z*+RAoEftokkV9~2F~vnvBowL824RB=m4Z+TTC5h4lz~vD#^oTCt8q05)u6?AilhdF
z8a1v3p;nC>KxhChmc2+$f^bre8$oDP<4zDdL5l@0lFJ}mR^u)Zy42VQf)BLFpCTCq
zVNi{4fN(>NCqS40Epm)V9)R#bjUR&WFocn|+3)~eW1K~Wyp=3s@Nk+A3UgzA@IlIE
z!va(t+ATA0a5EEq)aojz68h+FPT-<&8QGjJD^!lVk;AD#HiOH^-SjM(vvFBeC?#kW
z`5VHB%-g7htWEdIY>msJQa-`G7(#lk!q>===u0Y4SQ@X6L8JnOn~`cT+o(V$W~^IG
zSh_3|GSVZSLt~&2ap-;uMwel^(fu+VW9*QL>FX2*4((T3HST{ieQWfydf#C|Bo>p!
z9I8o6s>K#Uc9Y$VR~3mp)ou$RFPOYwMpuc{Q>k9Ch428F2EdF7Bhq|H9RM3*d@+51
z+zs%a*a&FUlm>tTq<<=qX&;x7^yz~YvPXf4p98@cV?5}23gM$-{dkVJ|NAX5oJiaz
zw;AI_QL5V(LVhs$!Hn@2Y4}t9U_*>ugT_ti6=+?=-VwaU*~06eS;M=0;UXa%q2ImA
zk4AyBs89x%p+a`fu0n3ktwJ8)JfOl};Jj+w23(sO4*)ly#(v=ZA>8%x7e{=)m)Fs>
zASN;&vy7q?Vhy1TQ^x!dIcKC~SVPDICJ)RHk){VyJYWssHZZlpe5@CdmP<+-SP^6D
z(J;x46`K@pr^`wwMTg_LR3NueBn<4JG+UI;iSZb^lsVaV$6FQ25u;?(;gdKLBgLpt
zl9Z&wt8gS)N>-sPDNBdPawJ>IRw0Mv&|!seq(~}Kp;C}aL5CT|(e5Lasc|(()u6+d
za<tn>HEP@dQUmC)zBt-lq?2mg2~sEM(2*SNCemd!_JQOB9o8R5yN5KW#uFe-fDSX7
zquoM!5W+}Y)Y0<S2Ko~}G-&-P9~dYQ@w+OR;TTC;A@a9INKHmq28^>{W3CKb&h@Ss
cI<sWi@TGF^f4&jfs9DUcM|#Yo-oY^c2UW^+I{*Lx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/amd_xdlops.hpp.5A2FDE085CB4BC8A.idx b/.cache/clangd/index/amd_xdlops.hpp.5A2FDE085CB4BC8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..44ede942d27bafa742f9f957a2aab379492d5f46
GIT binary patch
literal 12144
zcmd^FiCa`hw(nZ(+d9_`h(da}iqc@9)Fx((K;uGVR7TODW>80CG_GTON!+8Eh#J>0
z2qrE_N21~e8Z#Ic!7UOL#B7R!xEnMHAwEarF21^b@3AWLFFe1`Ti=g!?zyM7Q>T8X
z69>HV%H*yVi~G$1bKjaU^9^dTSX%S{GiJU!B_CNV&6-*)b2C2~R=aYsv(?G%_rL2`
zGgj@h^-jy#BkoRUw|LzW>*qTfMpiUFI+r`;AOGmm<olo-m5ytfsP<&!^1hzDv=$d1
zh1MTj)A7;o2TlfzPVII2R*Pxm>U^}J`J&K{aR=LOu2~&Ax6x-p-0gc2sW)S`RVRP$
zZk~A2O5g1J=M(F8EK4}O@9OZ_dCMpFe&xYdSNoSb4I9+vZrtqg3ETF2()>4n-0##i
zzv1bgqR&fbwOBP=nb!8`E|S0e)v5n8b@{gE>%L8&ld=8lgxS{*`kamVHM_~}-nA=k
zXW3Wu@4GL%?B6>R2KL=ow>EBFhmMXH*4;@vkl?!dv_-&-*GHtKIJ%BYag2DFL|?y=
zIriwCmBn>I1ATnwOx@>G)v4L3_7|6bzvx)s@^Q8+2Y33Pzc=91J6CPB*O$$DY3G{%
zoL;3(98}f){q(gTwaLvH(X_AUlVzDlau+V^^YOSIN48e{ynfxF+CvSsZH|9<qP7(}
z@$aguzPU#WyXTJlETnp2#{13pB)#7G%-_PRZY}BDHnVB>tckC$sUKNXkg_Ft@0F08
z10(h*fhE^>mL6#4x-oLO?@x6*lDpsqi>9Za>b>Up;da|PRF0i7@2zpQf3jH6_BQ+C
zxAywnivOZwwvs3-Nk_A>*!V2-f2d+x=ajo0kNV(d9V{#r*;A2bHlU)s?b*Jjes?`r
zdBMWUu)PdxW&<j=F2e@A8(Udp#c`^Q9YM+w)Xr=`rRkmT0*>}vV7KCzxi67QWRUj(
z;uce;2LC$n`6B)iKMTu2_8jhq(SS<JDLubmF!=ZcEAFXM7OsQ`c^@Ed)!Opp%*mP^
zRvgL2JBhNBILzjmO4zI`o8wlV=O5{4VM+ER&HDgx>rV$&4-eixor~=j*52M;^FBcA
zv$I{FHAN4)adDu9b+LERybllube(c1C*b{YR@{|$@G?>^qY$$JmFGu)n^1i><85xC
zxrJRr_G`S8MguB=zt3|wiHe-bRoYruF0tnl&1^u$5pefjpX@^Zq5iz1d$F<?2bs+?
zl|@fGUikU2L!WqKp4l>@EF&E)O`GyvZ9xo$R1Ny2_gnKeTCtP6n}L)J)WK}PN?6|o
z1z**7s?$GYTG@@2-8e{CVPBnII-{WH->ulKTG?=AcsuU{EOc)CFsakj*WR$=m-t7P
zQ)M{~HXBgsx*k^szqxdl75CsOkCF12`(ZSo64C2?SoG5BsaEU)KW<><29|o$t>)<E
zm0yqF!-uVng*`>~r+j3M2CV!gvOGR}(+6$z^<nmA8&S3qhv?1Ak#jbEH0sw1D;@wU
z%doNxw-+kWqkA0~lk)LP{6oGLmO|_)e141uKI9(VX&_nSQ_EFaT3Eb2Uh_Uc+&AiJ
z-z$UGP%Dm671l~=736(96YoB%Oz!>HQY-Gp$Kx1Pj`1Ng8c-SX%IeS`mK^S6#V>M|
za-x*;m2NbkGIZAcA#IN=c+HAq`5O5MDIcLAvw5a6r&(s+ScikV*~!APusw@!Yoh^`
z;iK0qZL?{Z_e188l@X;(9<mW#SAD!_?^jLq88%fuq{@dhSPa=H{~66!dIrwtAJQ!B
z7P8;sb7M5{p-C;*CM`JpCbtp=R<=@Qt8C@1;@<PGE(?#=hph<<R04y%53n%nP}?h!
z5sPE3xI0+LMoKo4*6O_b4IQ4_+dtllBS0liiIWsQC{#R0(57N5j^=~*6H$I5!Da)W
zNp{UX`)bnhDBhQFKK<C9!UxT0KxI)&uJOTq9*6qEXHH?|l)P(}eKqF7tcS{GEAGvg
zz&px2LEZ;gNUd8M|6YzGL=SZ4n08V+Nsgus4nw1C{`0K37kBhGr2NMBjM0FNHGOAp
z=r(OW|3nPfXs$H(+UVll|7*rK7XDmT(oqjr=0`3d<pK&36SsE#f}1~HI@6zb_IV3y
z!1e~-S)+l^Y-|(NcJ{HGp}Hrgm3pMqqYlD~=jy;&IlWHq(Ki)kE3vW?w>KNG@~N#l
zrR(uY%dNN{SUHT9!}5@A`XSXl=chY7=)J<%yay{D9`uX`tb8`{#QPtf8+V4g+seW!
ziM^6<O``#o>=qCI@@o0~iTXODY&%u9^UY;6&s3%~JALB9j_<jZAPcKP_A0(wjRsV9
zZ2q9j)&AN+eVv${t|Lkvk=A_fA!Tv?w3M;>BEaknR?cvDjRvfIzWyVhR#gvv<R5y@
z!d!Nj=6!(p3+LvEte_EFaUbsL4@mg|Im`xB^1A$faOjsU=Wq+b7B=2KUh_VliSJ!~
z!S|D+H@LW+g(V|<GVh7efXcq?%Wq_MtBBM`)$B<IRx;#)FX+6e$H0JBr|aj0sj?0!
z>*N_O*nfWbpT@j>M-Mip$_k{ckoR8E$yw!}e77i-`{J;$B4jV(z8DSoQXEnnA9v}2
zpPoaQR`QUNhuVpyba2ZzfBLa>pp#o^ZDAGEUO_dp0V_uf4>x7G9?VsIENm6Eui`33
z11evSNDfML-A~tdpZTHb%5?A6=<MAZUteFfdE(`of9Z#sjg3{tw(~yV{qmn$Z>?Lh
zbdnxw2-|^_9Xz%g4X9MMKDgqG($<f43+5*hl|-p0mD4gd46GmXj(((3_6t^i!5z&8
zY<!z}r1#QoPjh%r0xYba*z5UnHX2a*K7RL#L8EK#>Jg2weoDU}?*qiA^ZM?Xe<JIw
zew3JAlv1Tsj;gg!Ry8htawA_KbSuk8N<L5Aj0UV+`0JS4IRVav`sT#!C{~X0rED~y
zaw%iiHb2McO77~PEzH;Mt9c(FzI<e8K*Ok0|EnMIrXLTG@<3ifzhqjzKlf($|LAkl
zg#A<bXOQ;+78>R)$~n-GHBwI=O$*(X?ovc-_-(NL<lEc7;~uoIFb}eO`1&#$u<=`I
zNPWTdw!8HWW!m@_Dc>UR`L8F-4BdRK{gGDN=RW3ELM^NY+iQ5hGa9h+`!}_lv!*O7
z<SJnnmWS<mT*YWW<=)NdsP==u9^on>7FLSvrQ8>z0hI@LUoKuSZSo6TrM-pCw9nMM
z4-h{LZrongVbpJ2>~CSasC^e-#6|-umUkyioRJp9-8OT4+|11E_kJ3FXu|%qd=)Z%
zqqEO4ebmg%?;%GZ@6L~oJbJwsRoQABxf=I|Lr!IBI5JJfTX57C+-InnIjC$aj@l|P
zrC+7|`liiFRCWp5F5wn1o+`VHZI@+y1>3F&O#iX}qj?XmSNf@}r`9tBBBRQBIeYaG
zXrwkWq?MV%t8A2WR1bkBXcIyp9;$4jb7BvHCTWu-*<|NrNwx`Tn@|WWAC+YxXBO%q
za52(~Q3!-Dm6af83F;wm8Pdv72u!)kjv?nU)I;FoNINde%aOBOme(Pz4u!zQKxNmF
z^E&Dwa5C1CaR{WhDqDn|i*OIXa5dv|YTV0Uv%-apuL|QfCRT-k%|&6c6`xO4yZBAu
zlb?^H^KmRJaFrF{=mHs+<EV1n2R^B?3LI4-aMS$aJ#{w||J6!mPq6z59<Zs_b!p6S
z|KC*Bh~14c&Lr+kGGNz`sLf-JzkWw$9^&@MxP-V%$bg$iQsYjgR}WR$0pdO&<67db
zB?J0*xH5VE?&1q7J5Su_W&DV^ACUp?9`p%qx8?1rDtk=ak7b-e-5GR1()h$Kv0F~h
zQQ1c7-YDb!)V-e$crj~!vw(<ida10Kx{GC8L)|rWz{hX>ZAgNDTD8i~QTI84pIvmD
zc`vwP;PpE%47<uU5p@#@f$N0IE)w-334t7j4+>R1Gz5+_-bt#Sry+0xs>wf3`fNjW
zKA*Nc?8w8N#kA#NS00WNxC}eWaAz@XW!P1Q;{Z2zF3*_yNkna&%AN~<E&?LC%G$)V
zaSPNoysaejkMWme9l|?Az|~o09b-DW1#*TvBjDbwvWS=nw?I9^dOF4C?&azgE6~93
zfe~<zQrV!GL2iLYhK+QJjX%mYDpsHgVH2EU<4<%=j1_27*d$3d*)>^`%?_U(0YO=1
zb7JPW1)3W^SCY+(nJ38>qwvKjBG1gURJH`gEJ1F8(@=OCiU5*DW$7p;9k~VGgu*r<
zrwBJ$$d!d+1>TLqccTa(M^v^4#q2?Dfs0XCF>;D<RDxV3C|2My6jp|uA{-q<u45=x
z;NvLlxGXP6u5wv^7KNWh5!q%=tg>npQ;pmL*P*aF<P_oRI&xh{u>#*l;kQu)+yGT}
z2gTe$Zh@0=STc4(IMc(MYY~nWI2E&0{GvFJ(($l#JW}AzID9jX5C>8=9+r(q3j8Hz
zU*Z?VW+=q372>}Od>n@##}Tj@RCWRnJAp?E`~)XH!AWu`OKg;JCP~aBNpdJl^vJk`
z3@sre<WM&BfQ)O&&{{G=4rN2n%lHvVd_<DuP&V{286j{6P0XN4awr?Rk&X~}KOMTC
zj*vrHVlhn;xP}g`p(Er_mUxaP0d9WsJ-=6DTkdJ0vYC#VOl*mn(KBNON=A-k#KfXc
zM$ySAR^U|RNafj;$uz1w{fJIQu>yOL!-JSu^d1!LkvBmOa^xT;7JUwi&OxyP7a>Ox
zVq(!3q39xc6O<xHDPm&Lm!jxWvB<#=KSK?^5%T+Id}A{&&nN=9PCbf1zL2O3NeJAP
zRd$G|he!zUc>MGuY9k4OY@VNHR6RgLAjJdLk{{FtWhp*mVE7QkIPRvGxHO432(|pr
zI$;Sc2}Ic@&Ac2XbVX$giE|<8CVT5VM7qh|IvYth*<0rU+6}w~j+1+95S+qW1B15^
z0k~TRzX@*!9`k8Pn%ur#<>!s|T&Or<+Bn-p3e;9>E6Mzw{*tVN)*%$S&CeTW$4G%@
zYqLYe!kpur6DiPKZLTDn=bR_W79(vj3KdIs334t$k$_=%cyA0w7G@2oSUw&G4=cvQ
zz%U*TlaNJCO(GT)frjN43=S3qDv+_J2_R#+1%reIMWA501)YC|KdS~2YrYLcELSpE
zSfOTMa1`iRu4s_3LeV5;O-&$WxjzOO3yKiD`Cu6YESLlWmcL;TukZ%YuKM73=~fUx
zcEn8?6f0bIAQ|8SonM7}Fi@-9d7W1UGpa$bnvVB130AIT(5gbsq*+bHe!Kj$|NicC
zaXiOL#ab%vB#xbQ{7O0=BJh`3`x18&$4VjYUx*U{59(k4UUA;6Y245)<hX@8i}UXm
za@|640zXBLr>HXwp~^?l^%TVc2I7~yZqU8Lb(8)zU5Be3S2Va^p=e-J&jKeUA4r4x
z1)dG4Up^-W<qH;o@Z|!7=mi0Ef}1kPU0^AJ+~sc=q%OQMZOypQxGV8<yjKS82@aFb
zS*=mVnMBJZVe&bvd1PEdv=R~~pR?Kl8P^i6mW0XYtae_;kBIh&go$(YG0`5&ID={#
zG)$bU8>zNY#`~$ZpN5HZwU}zfGOnRo4Gn{HRlhQ5=LCkR&8JHzwqbl!gW)#kYDVDl
zjdLIFd4sQ_D&(j_okbQ_g<MrAPT*_EaSe5b7_9OT?7D{H00ULd=fj}L;d?-kbAe8b
z!=Hf`=iV8dI0#fA$4w7mKj_=l%Zvl{3zRrFY*67~0f=yJT4%xG&#FO(n`Tpf`u96S
z`_!lYzh5R$7vRIF*8^a*V4v|pG-?1KP@yJZ4kXkB`1^V-0Mwuo!2fJi06+j)0-sNV
z^#|56qt(jjpkLoxY*V?w(SXB2Q^!#axVOM5#F0Y6A&{#)jz^`C-U8PXM?DD_`Nk^h
zSVhCdwRabd-bMRB{-JWVH@brMfom|7AGqsA*#TUKmB+`#s5pRtvtr%~ggGP@C_Qis
zM#%vbArNv4MwtNw{A%v4K7a+dvhz_giVYB@q1eDJ7$pW_!7MqLCSc=o3r3*<xPK@#
zaOaEy16bI3qgnCR=<S2}mUw_|53sKo@Q2v;P{!+tZ9VZ71HOUSHpsY;*b0d+WH>4-
zBDNwKpCz`l#1}frZ-B&BE#rH{c8~aq3A|5i_hr16+SXEEF`?_IZJmtssV$%SiU}^D
zwgMSfQCk)Dg$Y;L8EQKtFw|N27#r0U7$L6cExN!g0kyyrV!hM?B#q9)I5&}QVtg(V
z=S9*@^w&e39@<U#f1Wze({A9;-n%9Lact|~i|>Th*tr__fEcKAr_MAPZ^02;a3mbK
zI(HheRmS=F#e5tEIfG8Wy;vaQavV{PBgNIU0!LH`3`>R&gi$LJ17Tnoh%b(@5jV$Y
z6>^Vri6fUVG4;75I+w%>TuB_2go&Z3B+->5R^WOPRZseesb58-R?$9UsCH4uF3QBz
zR!~O;WiWI=%<xsOm#(1b0y#JTvr(&p(!jpZro&gC{Oo7%n)iEv{pB_9_fW=9aQqXT
zAZHWtjWS+O{MHkHG3OhI-v$|HlK4!LAZHu#9vR;we)ovK==dWN|A-`r>|rhSTTA`L
zKxEMP44NRahXU$XK>fwQ6w~-(njo@=Gt}=4^@oAdvxWF`Gy$>)DBAI;U=-_Ms-aZJ
z1xAVPnZPX2ae+~o13}@$%aw7>Dz3CsnJ@Ki0c%8M?P+@nb)j7(<f1MK#nX5RCD9}a
zji=)!G?UJhP%@&)NW!UzrpnlZs7J;*h~~(+2+<-Lmm*p!<0?d}WPA<LYcjrt=q(vP
zMf9nRQ!q`zl7Cs4X301Y(>xiMVOl2R8cb_s+<<9=j8h0rA(H)ELUUzYNob{v>j|xw
z@hVDJQR)3%l<tyo1*H`N13AKXl~Hwt!r%V+EvK~TQnAHG*<);bjD28NQMQrTHWDA$
zC6pBtTQTv0@Icu)Vmn8CARO?GPHYc}59|ucHc;CJ>I1ugvLb3LqCT+dDXXToYU%^4
z9Lm3ZjT&WNSfj<Rpw>Swznj7X)*n6O#Qh&U<V5`+JmmZ44<4%j<)Qjt9`dvA4<7Qf
z?hhV<r+f^J(lgv7p^IE#RGQ(oee|iP>BjHwL~^pjsX0S|l;K3UGa^!;f!e@OxY2U5
z+c_vwpfseVp-{n<r6XrLiWGP^(srXz;IMRp*trKq3Var6XHh6*7&?LLtVWRn-$vSP
z6bg4YouGBzL6Lx=RL&<wFOkEPK&<A6jb0sxKdVNW+>9bXqj4ppI4;!8V!5dZ#d7`;
zy+{sfkZkhzjruqUR4A02?|;)N{c+&Kub%LWQ3`gX;LhR<PQk7e94BxVc4Xns;tbBh
zt}Gk}7?v)dN2AUTHy_{^xGQ>*T{xm62%1oB=iM->?4Sl)jlXSF)<M8aZr*}BSh&#d
z<!@ucR+jRPY{uGV914!9EF1UB#zO=?j<w@BR3rl@aK96H2w<oJ@-G^NdhtcGaBqGR
z2shqiy=o8q2^8#kSByeEynp3M#Le{Et!i6Qo_7X5z`-I~Jj8(yWxSpQt|!4FdTbzp
z8)RHav_j$#(WHoIMKV51w6nw^qDwW=s%3nS1l}XTFlCh8C)$1D5O^&OTuXytYf`q3
zYU`*&;C!m(Q-_E~1vIdL1`Aw8wJPd>Jx_Vw8F+>U1Fox^F!_tuRy^g&NFnhnBmvM7
zJ%RQsl5q)%FCl|vY9juCjL#Ckvm^jIswdcf)iSOn@wH^IOkKpEm+^h#cb^16$MtGw
z{9`g$BqQsn-#Quq9oMU!@f+!2k&NV1zkC`19aq_Y8o!?o7Rg8z^{b)*&~cU3(D)iU
z7?P12Une~M;m*ztnDUx`%S*Tw)ch&0`M10*Fr;yOP3k}IfEA@0zw<EHBmB;TFFgJC
z9iRlI0lrL)-*f<L3cu;#`o?cL;0^GCzhV4-11<`XUvq)+s|}!Vp*+9^#&0#CgboBU
y7wA9K0LlUejQ`pAWd>aL_%Ac2y`lGK#=GNZ@ZXhK_-{ad?E3XfljoncSpE;HK)vVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/array.hpp.EEBA77115A6888DA.idx b/.cache/clangd/index/array.hpp.EEBA77115A6888DA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..542e77153c40a38d2c47e4de696587057dd31cc2
GIT binary patch
literal 2010
zcmaJ?eN0nV6u*6aQ2P4XURv8nsnlA?_^h=QDo_ZR4Fz$s!5o{>ZG2{K>Rijl$H2+N
zp`!C6&IH66!<G>QH#5ox>fD^NDUJzICy^*~n4k=|u;DO4+<i|Dy!gj%(qDV-@7{CI
zJ?D2%v(lF=>60P!bU}LQ`oiK|8X@FQKAvLtTD=6JGzy{8hSt5^Z~U+jr;dlOsvg{&
z-|=$26AyPDd^b1j-11Sm{_4?v16|6gf0x~jJfC2EJGP`JD8|x%t2FycaYy3C_%(AE
ztsK2|PM2!`BJDs!!xWb@<kW16Zc4FKcFNSrVL3sjz|4832cbRJYkSIX6jiNAS<(I4
z;hLd;b|#OHq}^-Ty8VyD?F*`U-|SPZ{P?4xLwC;}Q(oKrUfYO!O;_EK&lHXI!~2rB
zm+W2T*1hAZgpHGj3=gZ?B|drXf#WY`GzOhy*7>$EKEF5X@@raxt_^n#Zpy<;^{4DP
z7YBAXkFS;$d2ScfycOK6&u-ayG*sjJY1ppAua{>;#N5m>whelful^PJ=hWrD73n7?
zsoUA-_j}6L=iMjSK}YWNjG6Z=RZs>M!ck0=N&w_XC^F_u<nFM>)ox^wjz!oJ?63-e
z<dg&{kOWKgFUUFzvO5A4lv;_fno(;7fc=51TbuSr41M5B`H}ssjAaEta$hO&PWGD~
zwt7FbCNIWR0NHP_85Riu`&Z>xje8y}@{;fftzuLv0f64H^~BHu&9@_%(i7SkZnO&k
zw8tFw<A!^#Y)o;4HiepO0sviFzh_JTg(*%&X$kGm_^Sl~`k2#k!Fy76MME)!*4nkP
z0s#H}ciT(+9{p58ZcZ|1Wvp5OfbL4V+}wT862RZw2S>;vSOI|Uj=C9L*V>YSDLHYj
zv+3*t0No#Q>3mJ4bAp^roZIZSSOI{(A@!N~sxRi1$Gf?3E)w$q^zhAXXHwekI3HKd
z73{KN9*|r{&k?HTE-dn_<sZs)^(Bl@b%<H%Z8{z1v+<Oi#add+@<QMTi*+=S!SrA;
zM{^>}KpRAskv57f6KxV%Hrgh#9LO;(HvR+`a?MyEJ`x?Tye~hAJYCqiw;d!$rg^=A
z!~%Xj?l<}Q*U=Ihhp0m^)E;ObMo=C&VX;cDGJ?FOYcq2*r^1QhbQUKmk}#wc^eUkQ
zoq~5ua$Cn(+$H9T!y}$;+%jI)UP(fw{pdN6I%~nCD~PH*Lspy3W($N0(U8@=(!n=~
zoT3DA!!(U8c1t|$g>MoUuX9)g5|=W9v|*Z5a=Gq#&>*Q~3YQ9+s^FxRb<(N(6pNiP
zPS<Q2oN3NXuxgh{t4DV~it{B33D+)!r$YQj>*7Uzqc{inWw9Z_uo&{kVq=Wa1y^IS
zE!?&cL=WCb(EO{L4cZ)R4uy-d*rK-tg5@14_i{VNnxDx3H${UNpJ6Wi#tT`g<skng
zOr=g_aY{~P8I%T*WmFnPmPu(6SvI9jWI5yx?d%CIxogHckvQLfXxsf=5)!nZMFT-Y
z2O;(5+XuY9b4z;9Ub_BV-v1T=7E_GSR%T@#rH?YgvxY*D1NaUDiTow;sfw)3xd)y|
zR`1n!aw5%L?~^EF^|2vPYN!fv!uKIK7|Kef`PKuz;hm6azU4q%$X&QDmrstzo$m?e
afAD|N(?bJ45W@4z)blE;f1XeIYtVn+Juue*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avg_pool3d_bwd.hpp.40079D84C914EB9A.idx b/.cache/clangd/index/avg_pool3d_bwd.hpp.40079D84C914EB9A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..45c5f80942202e01938b0a73b3df5985bdf7796c
GIT binary patch
literal 1646
zcmWIYbaP8%V_<Mj@vO*AElFfyU|<jg;^LB`%o;`phFn$#hKf0R0|T=U8AzP*|Crdr
zxobtMZA;J$Ve`Z@wmjb?E<FiuHa$6MOVF`@zgKPb+aj#KY5JGbUw6EJ*Oa%;ZugN&
zM-R3;+0yHthn(yeS@F?K#zD>UPWt8WwUhq^wK%T6C|~w|)#ua?U;b_Q=@NKl$G@ya
zn_aGFiBIFHn!Pye)1r;%FIZ3dA+z=VVY9#!&s!=2{GwO!Mm$>@*3z++PqDPBx$qVH
zna$hpI21~+oPXNGc5}?~eOEOso=Ui`5Rshyc0;6b&^hCyi6_D$i@tE|Ym!NFx^Rg#
z_2w0k*O_}nLOO3wx?FK-@8VA;@8VZD=C_~nZ0p+|T*zrTvtm!vt?!>*x8@$!iEd!O
z{cel3X~cH6*H;r9MMU$|jToB@=e=ZoFf+Nya%SS<=?R*_Iw$9r%y`{;vF|A71<MFl
z&S_ohD;dr_udvsiH%nxC;m5@N+l`aY#m+i$!)W$<z4Z%fg>N5Utt0bq;p9y(>&v&i
zTAtCePK)!%wkbZ|jr$mHOns?WT$!7602mkI&ptm}I&o4f4+9TBkP8IDe5`!>44kY?
z3?h7t!i=&xC+FRo?f*xJiA#V%SV2HR223z;Kvd7TBFY(Yo41~affuG<j8zP%pN$!!
zf6K)8ds@wZa_}&4!xRXy3K=kPvM@pw+-+X^Zg-^!HxmcY0y!=@Aus{6;COcJ{S$Y6
zFR?Lk^0Ny|u}i6f37B+IYFhC*W?;Hn-zinP?$u!y9$|hHesj3Dg!wJ`Z4f*)Lp2ka
zk9dUD0@Q+GJeV3LPH+J99{Rmt%|mMi9$|F>bt{-67|qB94g{Fn)L6Gok;#?~Vd4=M
zWfv8NgbV`%kFYMgE}Um}|Ag;}pMMW>09|Ou?+mjK$aCfQLhytggk9lbCM@hO?1SKm
z+lf2F16)|#Ron}~Q`S`0gE@~!SlL6_7sd<m6H<RM(S{WiG@ShWNLFzQ^CKC~Db9~D
zPKj4Z7-m3of|T28L*p`_0m3H276>iER>F1&o|c&wJfL`lwZgQbU<R}-uCMZ4ar8LQ
z>&l|aaxh^a&sf<8!Bdk|Q$pfdsyQKeT5?+INIZKj4+Kw7P7g^=&rZ)BCI|BmFj7IO
z5)rBT0{U>b&VFtt>&Ni!Brx^_*#+V8#UrfAt_kPCya_Z3lzb5;X$xpu!>ro^P0CDM
z;3N!79z4QIDoQ#qMTZ{u9(}Cc-vYFbmz@`G0gtdMyDFRq(+W&zpd=5o?)ZxR!IQnB
zKC%L%Qi0z9DJm6&4PZ{>5f)bvH-LGaM_7hKMgox}<wWIFVeWit!1{%qU%VTXGDP|1
zk&F`+mPay9R9qfmoRXxHGR!Z(woN~%XMK7zBM&4~Laq5L5Vq-V{*h-ubHz=>%@KOV
zEyZmRJUuf#OPEPK!g`^45inj+YEB}^tm2~NVr38mRBb?ku<fJ>`BpogWlVxX3_>tb
G1_l7K#8FEC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avg_pool3d_bwd_ndhwc_instance_common.hpp.EDEFB32170FF2FE7.idx b/.cache/clangd/index/avg_pool3d_bwd_ndhwc_instance_common.hpp.EDEFB32170FF2FE7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5640c2495f5ed2bfe7912ac911fa15fb6596a732
GIT binary patch
literal 1814
zcmYk6e{2(F7{{;OxZU-RqwUIEuUECKckc<;UGHv#z@q#Cj$mX47Jp%k+oUA3E{pAk
zMFk>6vBp(b5zJW7#TW<!Mi!$Y4hLg0F)$NCh$zAwnT9#VfH9&AzCH5po+f?ze(w8e
z?w<E~+chg|YlRYuTDERwrX|tVSWHor55LoGspebwv!sZkGTjGPAFmsV^HX;<w!FJx
z#OK}0J)YaSJJWl<Dx3f0(c60tJ=GI<^__iz71<mA{r2~oAD2$$h6CF-{b_zZIez1Z
zWj)D{&7)T<ZVh(!XZG$GtE*ZdS0+-s>R(tDY5(QQz;OSfUmrf7)E+rnadz;g^?Tpp
zq0O$EY+@GmUTgXO<*(=4yswP<zUd6-qu(84A6VTSOy+X2H_m-p&3|$Cg{SA9pqD3i
zU6Iy)ewu3D`dVLJSRZ-r;-apP&NRt$8r0H>$n`CmgF8BZK793q`sBbpU%A)uKb_pX
z<x*2u;4ghjKlZjVzW3YS$ul>PP^W4;UOd<E$3v_<o!;2e#JKQo^kv2NL)V5lx0|K}
zAMIo9La}&vD9R<c7G54%HS*8&an3yp3jsR7*~JlFdZMHHz{F=N=PtoQgpS1Q;t1Os
z#M)lN+Q7M8Sn$(+k6j$$NYB1aOLksHoVyqcUfRpr#SuoU_jk2drxni4VnL%d%`T2G
zR{#CQ+e-%PI5&fZGP+E&iz7^XAM3x=dUAqu7h$24E@kcF2&u&S^!xaD6s2CSSv)kc
z)Wr*4!8?~c+RQe2?Ct3|@$ZKR#%k{6g&+#fCwIrB1$Bo+B!W!TM39A02xK8G1TqOp
zAd@r+WMLErSy&6tWMoh-k3Asrh|grYTN?(x9LEj${r(^s2ALYR3nL>oiMcy7BA0m~
zilPh1cQD1GhC?!vK_<s!kSRz3nG#b#h7bZ7iXo7xNClZ1Q)ja6*B_J?kG>`1hJ1cs
zIhjoF<hG4Zj&<jXc%hsv7n~_nuoZ&ytd(q~K#mg`F5ZK}p(D+b7+)?NHFEg^H+)}x
zHoq3%H&7X{$aa$(VY_hdh<$C|iUB=+x@-<U6iiSd7B7g52q?sa07;AlD9nTbMVYAL
zv>`JxAcau?AqD|b85NMu=nCnN7vfADkii&$OvVIcG1d%`-;(IDe?z2W+^7+(0*x9X
zppX#)BpDK*un`6nwW13_qm~Rvu@peaLV#3D1*9W=zSF3#={g`ouW~4k;=qh+abU)c
zI4})l0MpP6U>b%2OcR;FG&K{LreOlpLKZMB%>t%nSTklS@qA<9A32?RHeFL#2yX_@
d&YL`Sf3DcW@v|vU;o6-SPZkt4+gWnX@_#ZnXW9S&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avgpool3d_bwd_bf16.cpp.CC3B9B35F218D95E.idx b/.cache/clangd/index/avgpool3d_bwd_bf16.cpp.CC3B9B35F218D95E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..11d356e4bb9ebc7ddd119bd433675c1b4145a128
GIT binary patch
literal 2172
zcmWIYbaN}=U|?`g@vO*AElFfyU|<jg;^LB`%(=`A3@Mxp3>9<shThJ*Z6R>={F`9I
zv^h&|mNnWudA4@rPKmoBl_G9WJd}JUMorrE`0p=SuOivqFP)pq7@eP;xhOMl=Y&%X
zMhxN8jxA*RvdM1!sptLfGY)>e8vpa`-=k08`qo*_o+oQz_vYs2qX)S+A3gc8@$s84
z2R8Z}?yvXh`~7<MY5n?7k51n=|M}?iVV(UEXU;EwbK<lAXV-l@_EyZXt=V^4|J{wx
zQXgKv%3^r7sEoJMYQdoe;fqSMKUqeZ%wEtuVO?hIfkOfNRPs-DEV}H*nY;9*$X>;N
zZzsBm%l%C|{-I#fL1C%G{&l+*r7K;-UJCJur8pVwEw}r-{zQJ#+T5u<&tL4U^k4d!
zcbP(NX^X?MCtFJcHZ09$sq4?VaE>qU=exVd`6sk|b!3;z@Lbp%)i+`Do~PXj$JI1B
zvSm|wyS(0<@mRd#tBRq~8V{*25|(;B3tsQwxKMR{$B(If^CFY4@D{ps?3;Y?=Z01b
z>vPL4wwtxw=!sqy*>}HVO%8j;+`A_0xD7tM6IiV4*gWs|#^Qh8_YQLHbz7>sc#`jz
z(CN1fxw1+)Vk$IDC0SQqlj?nTfO$@Ek6vDs)Zb~wk0+@(emL~@RtVdLsiiR&4qjgr
zw7ZleV3ApL`*Mf6qdrj^HO?CS5$l+{ea`Zxw%dm<1-UQ3%*ea-^PPGN>(&<vvRCh&
z{#_Tm{raq1H<y-Xb7y}sOEo`nFNx8+X_2<~rwgoWGPp|`TbC8Ot#0eeU%m6Y+mdCb
zHipf`zB^h%eU__=W}eZ$+OyW-e@TjJ|Afo7r)MznuYLb9&PIRH^{L0V`H9SZn)2&^
z?EaAaYNyEMb7v^cJ+$00kj>`D)1<n?=dUdJxb4Z0l2Z!|{PZMmy=r56I_deORLQ$9
zD(a>*o?g&v`q^0W`<c>tflL4He9`28{QsMf4=*=NoiWd^PpxKV(erb`$Lr2Nu&loA
z>vsM`m&cy@&vvYedVZ$!RQJ|jiM!qlcG_y|pL%)IWm~RVg+fdDrpE7+C!d_Cu|;%i
z-qJJc-W7Nn&%3`t^=$mZ2Zw6fs~&CAo_5`J*RhWNOs;j$zdijWX1=cb=2vI&@;tL`
zby1m~+-dt_R`28$tgP8IN%XJck@f3NufMA{S0`Ql>uqx>dwVCnpWi2UTuMLo^UhY2
zn7@m%tQY>$<2aky`?}e1!>YK`4}aRfkTAU5f8t)<Uxn+7CQZ`!N@sGNA@@0B_YL_w
z>;LSRjsN6tuuQf-nqzhkyIpp`?Kk$f-h@gp6j$aZ-2~=VasTRgwU&1p+)Vxq3?V#h
zJc3{XC<6xC8&(!fzHz*co5>F-&CA9s3MOFE2I5yF9=|J!<Y9;gN-MD_aexV!wDojt
zxu0FP_H#1@0;RdxxP`z3Ogb~`ySUH^Z#izJV4$=(o47QXfJr~Pwo|TLq>+W2DF7%f
z$R@}SCScMp?8`nB-*r6A%@hQb7GM()02452hN9H8Vs2nL1oI&?7mCk#LL?m}ePB-D
z2{EuVaDjOT=4&P{6yNfMNI6Ocz_h>=Gjd_|GtfGahhfg)36XG=@P+9GTBmJb4l^Od
zPe}d6L>pFCo)8sI6@Hj7Plzg~DiTkPQw@oy&Z&;X)8N#A^WrP!<R|Ri^O%Vzgj<GN
z85RjVAv`iX%COLYxdiA%P{1RcX&_>t33EYVh243XQ<wg-vV^D@su;n9Swd6|RgrjV
zhH6MWbwhO|o`#_YoY(Ee-14^nzXcc2pCT&KFyny!6j4Rusfnl|@zh1sk$4&+8b~}X
z4lTH^c|x=~w2^o^96CrmT@GC&o*suDoOepP;t^*BA1lz079190Fh{Y3SaMh*@vJzk
zBoK1e9M(uY8x9*unA~;lqng2;lcIouFJLTS0ZXJjA%ez&7O)ikF!BKJB;_tGMqoN%
zF%X97g-HYB1eDs5;v~o*8m0(lB2W=1%_AvtGYEhwdZpyPSHJj_7|;R<B?$*uWCQa=
zP9jK0aZz%yJBSehA|N0X#$aF!G1+72_LTSGY}UxIi11K`Ft`*$B!U@^U?%6L<fH?+
zUa3CJEUawo9Gu)dyaN1u9)d1TYJyxILgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q
m6%>_JRMj=Lbo31jjZMwvOr&&;oJGalJfoOgeI4Byy%_-hS28aE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avgpool3d_bwd_common.hpp.C26898608FBBAD5A.idx b/.cache/clangd/index/avgpool3d_bwd_common.hpp.C26898608FBBAD5A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..11302a846b4f26243cbeca462a7b53f3838a4de8
GIT binary patch
literal 3936
zcmbW42~bnl8h~%$0)bra4LezggheGpwq#f|0wMw?Y)UA>#FPm5ih!~yE_owjOF?-$
zV5KUn2%@7!v7mqz1V?dDQPC&V(uzVsL_pkdsn2(*JvaC14D;%p$<6uBIsf_BoO7~#
zX+VGsgP_H6OE)IRXQ*HZf{gG_dPZ78r54^WfS`>f=cLJrzqrs}hPl>nH}6gu8C@-=
zuJo>QX<kHoJ=b)^c<1H9g2Afjk*(!6jtyqrOm{bzyosi%v%@7N?AGUk=F4tQ3{h;c
zcYlv^^B-xPZTAwN*k}cE&gP%2>!`I!dQdm|Sifr78@=dborl}spDUZ)0>wrbG5*84
zt5e^0M(b|>=`8ot^D6$e-jtqsPcm>Rc9ZGO0a?lO?uA~89gVF@<{9%Vhh*%I2U{H0
zdc@X`8rwNoTeTHNysG}LQC_ztfnQ+SE@K@k-)m(gDPz+Q|MGA=zN$I@*GGe@+sQQ*
zmL*9;3)jDnIQ+y|Se`28M|HpLyUn{oJ5Re|{Pnl3m;9gip4s;;XLM`crpt-Kp-ImQ
z*Md0|L;n?84dQ#J`MM*W=Z>10TMxT?G`x<lX{tLQpbkZq4&~E)cJ-&cYtbG!b^N&F
zKUSwMm8CR}WZY1f>0uqacJIF_nJTO_^icP@)_QP?@AUS^&7y1`86O{DZSM6tU;1`n
zfnK9+*hbIq$FFT*<&<t<wIn~&zV++E4aM#(QOKpv-Z$fcF-L}~)5Y%FgTLA!Qd>65
zE_A4@l2S6Z-j4C7{yuv`weAnkK34wfTX8w*HSKpsT@t_hah`cVwtf6--u_ey!*6bs
zPDLL*0Qs_V`47IaHhD7ET6*M?qsaapV=YFaJ(Z?ewQu(EhpR^)WZxgzaWJU8t1kR$
zy0U0%ajt~N)6I42d*-|iE?-mGGnOR}x_}59ZSU?1+FP-ZMW1)9Xuj}F!kBNhw1~F0
zCwIu#!m{2wg)`J@-<ah$@?u+hL+<PMZuh+|UvtCCZoO+1wfiSX6AHR+mUIh7b7Rk!
zy|`ZRMQ3H_!`Xc`&lXHs*9_EV7*++Xh>C0L9XuBn@>GxcrhHTSrsP#C@j^X;t^cQF
zpW8tMGK3%$V_?pIz94;V3S_Ad)pKFmiem5Y5eUJDxn|sXf4(4nSPRlKfbK`1H!QT>
z!9ySmd|0F-GX3)f>Fer1R2l@O#jj4U$G@R3TNiZ+954!k7E+jn6k8xjnFtQ`ZSx<|
zg3%a3j1Y9gsLEC421whLcC`HJs(cth<u>v;pc_FIJcYA{V&=#AXh^j{<qJsn<dkeH
zmUw(dE%YW%vS4q}b=*d!zLE`O&9k=^#%7JnDKHxD9Zo2)XwNG8@wa{NU>KD%<-}~9
z=E`|M7Dg2ug*Ay>6ogaETNhKIN22)8UV!C3ky&iNd;cdforP}m%A;PruZ^H=jLjx8
zj0!LTn@A{{X_85pW^J%L<r_=mBHS2;JHrczz$lB$veOV>$yb2|VAR#!)e8`}Iqj>*
zS)v?m7-bu<4K*SH1A!r!1f!W~rV$8--C$FPflw*^#D0GF5eaqG3I0bBcMd!B$UN2Y
z_$Tw{b0#|zbybtNZaFHagZ1&?qY4tKQ8ht+>yFIh_%zL25CK1ZGoJK#b-50Las^z%
zRRpyW*pSFpU`rx9fgOqL1@;7mJ^byM<8bRm-2Uy9x!WlVfT7Y*<2ertL^`?%is@i9
z5IY22%v?-l2toas{v=w$TtcD%W&nu-nSlgJ`;_842ZDWOEPZ80uiL|5)lYZ?!>z-)
zz%ZN~!}%l%3=aa7ke0q7>{X4g7J|mWF@~TUK~=Ddprqy1cdaMkwi%VDZuGv+?(BYt
zD`#K~I&dGaJ&Z*nE1DGo#!?%84tdGhUOMBt8GB2kIwww^{-}(ye+W*FEWO!*8=erH
zAlDeaM!Y5(P)_}1+~5VV=x+pMr&#;zvEA=I8F&`ym<2EuCn2^#Lv$gXL`Fg*5*Z7P
zNn|24A(1!3o5(4G`Y?P*<je3S(IUno68SOwNVM2~F_BNaqU`-iw8VZ1i302cNEB!v
zNFp&JHaS%xd(W(FeTtwGL_(rqB$z}YNC=5Skx&waAz>tnphjRI`UoneN=Xz+jU*_)
z^+mq#o;kzQS@j6_1R{H<Vpe8j?0@4wt}XMo1c2`1g=Q8h%lSwCMjp%j?uj_dB!-{8
zFIPO~g<r6kYR659Gl^pYQw}CcVDYk?2e{$t+3LBVyK=(JVepad8@wE2ZDL7hTLFbv
zKp{@Is$ALr-`BUW@kNzZN+%!<>Pp+-!>Ns*w9^lQCUcWLz$64+%Uw&*35TIb`Q)<c
z-Brz069-qFqSvi5A2Y@e4Y|IY_yyx<tiFOoYJD|9#I|nm;ZE>DH?BO26XmW^F0+;q
zX+qFwZnPH|Yw*fE6ujyH@=-3E(2?soDQh6y*Z;jNG^5lTmt=z>pPia!yRM%8W-n_z
z*p^|IQV=}^<%RL2z@1icp3T^`ly>~KVHstq&`e?~m?{nNqIe1*f}l<cC*ovmyHhQI
zx>^!u+?}qZ9bNWzAUXZBeYvW|(rf&pv;U*uU8Vc?yu}{R|Lk?UY`5fQdq-f?h2SD5
zZe(~Ncq&97a|o(-QWIC1n?bv;>^yaaHKW5fhuYUQ``jMZfl;xm*bVF-Fe-7A2s9Mz
z9PFZ@5a$pQg}R0ka(6gqb$)iVzWPa#^w2IkMC28&z;n!$GNUyKbCNnmgM<M;C<Y(y
ztACa5ePu?!r(XeWMn^{1)JA%l_+7*U#K-uyU`YgZ!CXirqzPw%vC%!XFH0S-m*XN)
zaFhYij-WDFMxtmqnxL__idP}UIUdthXl6;e#4(vnuEr^z8IN?Ob7DGifzEg5yJjB{
zZ7jhT;6hwe(Ea}#ofe;@nvP3)T5Njabi-u&_2DrE4G5xZnr50NdgFMT595K?6RJ68
bWNcz;Mu+gf%ix**tNhnv3>pw*u4(=kBU)fB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avgpool3d_bwd_fp16.cpp.11C3DA7A01E70C82.idx b/.cache/clangd/index/avgpool3d_bwd_fp16.cpp.11C3DA7A01E70C82.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dc381a607236b05c0a3889377f367169a3e7ea7c
GIT binary patch
literal 2178
zcmWIYbaSiXU|?`g@vO*AElFfyU|<jg;^LB`%q7eW3@Mxp3>9<shThJ5Y$kB^{Fh9_
zv^go4?=sbSyxEa$BAz_kRj$?0auP@I60e}LJL}&EOnQ7|c5OzNZo-K<ma`5Uiw8+M
zus!HHWD(L37dClQ#jlzs$urC9pIk4W9&Ud9=hd2;!Y4nzoSLWKA2&n)ykEbJ_;dHZ
z>muE|+n?;19UEW!d#?EXe|yeNkAMIC=hGrX$<4~=_v=ji`RUoG#Jc}q&Gzp}Js*8w
z`pYfsVvCAQJGB-ZT427YF#3~al*#S|%@fw;#T=Nru*Ub^B*E2T(yCjgE(`tA_v5>}
z)ZR}o%Guj~Z|0G^C{g$6w&~3i+MB$@6AE0_R4!LPzx==Lk6zfVX(ubX=FR>dVG*Cq
zbt%l9BZVt2nq7O_mfi#JC-UC9c~95A=Eon-NgMQDoM7hF+!A=2v1r=%8o@U=R9F*N
zFJ9vi*tk|QL&S8Rid&ngXS1C0@m9tXU2#UK_gUieU%u!xb&C~tZ+Q_dIhVh>WAo#j
ztlAvsj;|Y1XFdHf_r#@b4cBhQzP-g@`<~J8me7?KJLTD*-+v;Tb0Iq<%1G;`OZ>K#
zvds&k+7Gn0J53IpkhS(=2A|C#%cCzIty_37-_lsmQ;0?X&N|OjhS;d%tZ!?1wMrw}
z8CbJ6msfB-F+a0JwEL3I;rER<s!msGt=MU+qgz^H^X_7R<@vwu-NhYzmjd7J+4H`<
zw0iZ*wXxCLw{CFF4=&sL##U^DLc~^AN4Z-Ks&5<3He4_{mk}EOvDP$xui(l&9)F+r
zcZAj@3SZ9jbd~z#WtyygAznA~r1F!J<3>vkIKQ^uBYeL5Md{0pD}`LMEoZ*_e?kB1
z&3OS~nd-|tmVek7K82CLZevW$XJgaM9lDlwbB!V-)vxosEj`Tjbk_4psgkj;Ei6wS
zNWE~R>*vg#>Ss~sS6r%7y7l1l$Nd)KHC8#A$DS{qEEvD+&Q8x)H`Ytqq`u#M@&#w4
zU}yhh8S~V~b1Z{@X085bdbhvhh_<%=tdB=M%zV`<6k5tRwe>IFe6cekQ(Nn<$4uY+
zLf_(f_cy4Yz5n3Bp+5~(Pc~_u&Q?Ba@W^>q)6GNsV(vdUx7Bd=-R&PAq$R&!-aBne
z%CpsLqizUHJg%-CaWUWN&yrnBCYG0l6sFd_b34yl@!ah2bN3Ub|NnL06L)!O{J7?w
z)T{M%fo5w1V#QmYEou9DQ6eHVH}(G?`4XNPFWsJ8t-GrbzG%uMeXn#TSBd$bGIrmP
zy_5bg-f#U6b&1P;`*mANk1@}`x+3=*|EiSGfa1#Bq?^DzEAC$%uh#NTgPX~pfgzNK
zjYkkn0A;{Hd&A0t$v2MIaWnY=rFq$SMZpA2+Ccn@#N&5Gkvt61KxriwB@QqFleV6&
zE%&qQ)_!iLK%g`?8@CXcfJtX&eHRxx;Vs9_6bzIWXA_qO6ENvV*LKR4i!`!uGX(&p
z1=$4o!30eDg?-tF;=7KgxtW51(gJJ(0$>6r%}|t@R?H17jbJ`x=0foqPpG7$qz}vq
zJfQ}b1}-q~z<kZbh2mSDP$@^L0GJk-Vn!~keg;|x@-WOfJfRYf6236KK<l&(%wZ;k
z_z9`Mm}tYw$`h)>slpEv<_T5hR7K*cajGHl)H&6Wcp97<a30JOpzA?lig1B}h=C@|
z@WKka^D?I{{bgkdRWVdCf(f&Psv4>y@ze~}ka+5b>PS2dLk&2u{e!EF<@p{dCZ14k
z8Ez$5#Pfvm$nYq^VxZfLx#eyDe+w?4KSfleVRi!jDWZzRQxj1`;;D<MBk?puG>~{&
z99nQ+^Mq=1Xe04-ICPMBx*WPlJUtFQIPa8n#UsuNK2~5rSa4W~!5qaBYRO@V#IxeC
zl0e8=b66wsY&dKrVRF~Gk7@>YPKp8sgn+St1uT*BgbEr9TEJ5D!^i`?la#x(7=h`4
z#XuOQ7bXpiI8bUwint(yXqY0Hi9ki5G>@dn%^(1#=#`TDUj5=zVn7Qdlq4KrDFT=$
zauPv0ii?tq-GMY?=!C|KGYd|<31N*4iwF;4h=7^Lzz~LDh9H=c2xd5fnVg%FlMdv1
zrTQ?lu(GjpaB}nT3h?uJ2)a0_337P|i3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`TH
mP*hS;RoB$g(Kj$OHZ_+sk<vAC78P^zjAC;2b#!O+W&i+DfFMc$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/avgpool3d_bwd_fp32.cpp.4BE51B531517C414.idx b/.cache/clangd/index/avgpool3d_bwd_fp32.cpp.4BE51B531517C414.idx
new file mode 100755
index 0000000000000000000000000000000000000000..07a426bfd972935c0809b6d60401578053a93216
GIT binary patch
literal 2140
zcmWIYbaM;gU|?`g@vO*AElFfyU|<jg;^LB`%q7eW3@Mxp3>9<shThJ5Y$kB^{Fh9_
zv^go4?=sbOOx}@gBAz_kRj$?0auP@I60e}LJL}&EOnQ7{c5OzNZo-K<ma`5Uiw8+M
zus!HHWD(L37dClQ#lM;+$(Z=R6Yt-N_sh@!^fITW@X3#og!5wV`e($_m%I0Me^x%O
zc=XRdL(6yT^OxPPdwz7geg3DT*V`lOcFyoW_U6QA{m-uZcI>U0V_UQLwEp`WpT#~r
zeYJ}H*`gxTPOSxp7F=IcDE-MY%H;Nf<_YWaVh$`_SmSzclHlnuY1J)LmxX@m`|;gf
zYVVg9+pF8!H}lN9AW`?}w&~3i+MB$@6AC=lR4!LPzx;pJAG5Gq(@s|O&71u_#3DYK
z>r$9IM+#S5G`sh<ExiZcPvpII^PaAK&5u8vlQ!tRIKj-Txh3#4W6`wjHKK2BsIVq3
zUcAO3uyL(qhKT7r6}L7~&t^I0<E@M(y5fvd@3X|?U%Z%U>J}^P-tr<^axTAh$L7a5
zS+zON9bY%3&U*S|?ukp=8m`^UeS3?+_Pw0pEukwfcFI>jzyCyb&Bg4HC?l<#F7ew|
z$~G^EYCq80?ld`YLe|=g8GJT}ERVi;v~A(Rd`n|JPa&4|ch-5PGQ>t5XMJ1CtF<-0
zoq;uLb8!XNlj~=eh<0DnIsCriM%C#`tra_Mb#zNhY~EcAusr{_y}P)B?^58~J$v4l
zmsYP{xi&U(`^F89`N3s--^hw>P>9&(>L_=sLG^8e*@g=y=Q2X$KgOE!-|Gl5Z&P1Z
zp2rcsq3`9EB}%-NOJ{8qeX)MpnkRxqyOJe?9xVJSd$04k>z7?$ZiI9$+B(PQ`~LUS
zU*-5~tllCVyd?O?741_E?E53m<o}pC%jC{9nfZQ_I=z#%+sbw&b`~8iKDnu9UFn=T
zDG!V?l8*lIIc8hDHa#TMUNHB=%OCM`xb0-Oi9G&%@nna7@Vh%IrEjkD%-j6^ZpxRQ
zH62He7xum0RCsRA%9@>l|7X2(cX@Q+RQIztB`UMkglz;4Sw|i|e(~lD7oDx5Q}dRb
zS@*8M)p*|h4a#TlKX`EHPeav{P1@70yX`vG(Vxk+?)kT`zr?n$>%IBaS^9gPSy^3F
zwkLPmyu2%S+B$6PA`f=e2|rT5s@l6}^WrY;&&B3dnFp-5OWN5xt^4sk{>wfkulbVC
zi?0XNf9>9s+4^dif{D86n*L)4R%I#we*S+!@&O-nCHCK?%`u%pUi&oWa*8C@oLc-g
zsk|WmN4@3TPyPnWWb30jUiYxutq#2XMt&Mks7G;SZqiL)o)!16j#q1Wr@_tS&%hAM
z!^R^BCV(<vpuJ&b!Q>mq>$sWxfYQ8dyrN(NCT$>oMdI<hqDUTwXrQzbixLNzfJs|V
z*OvR)b!$I2Qy@^9n~hrtOu(cwv%ZT9o$!|9W(o#Mi?fMKg9(`QqiZ|m%0(JkxS0Zg
z(t>P){9pnm{ldQNL-Ae5)7(r!KxqLs0Rb=plV&JNO)GxJgzzCV7mCk#LM0s~ePB-D
z2{o`ZaDjOT=4&P{6yNfMN;ygez_h>=Gjd_|GtfGahhfg)36*e^@P+9GTBmJb4l^Od
zPe}d6L>pFCo=_D|6@Hj7PpB%VDiTkPQw@oy&Z&;X)8N#A^I(<$T@MOVgbNHr3^ZYe
z7gpGvmpOImFDpx^ilK@TOqeB9)le0Qr)H>z#8Wp^N8)K1YQTBjUd%0T`~O>T0lgri
zA`LSR=mil~B%Ydx8WK-kL>-BzA)<lA)8f#Adx$4gn?oCkr^BIx#M9-_MdImk=)rlX
zq$?hAR`9U`{b<2qAqI04OQ<D>B@)ky!%6}nXU$=a#IxbBk%Y-z=RT?#+&L+Vi6>OR
zSil07{&+$KjRh@WiS=RR0p3Z<U0RI5C}uGbhUtY#1LFjgkdfjf$RHY~2xcNs5hyt$
zDRMIifGK*V<i1zG_>>sX0tqDv2UuhmrRF4pG!z#l7rO&##!%zg+426%<Q}m`hDC&j
zFhszNV_*nDFvAecNCYz+!A#Cg$w>!ty;6OcSy<WFIXJm_cm??RJOo{w)C9Ragv5nK
zBqS{?t!-@W>>cz(wXKxRG~}^~ib>1JDkv(csH$sf>F6668k?HSnMmmxIg5(9c}6k0
L`Z~HZdNTk3F(Db?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm.hpp.E9C31BFE1C18D556.idx b/.cache/clangd/index/batched_gemm.hpp.E9C31BFE1C18D556.idx
new file mode 100755
index 0000000000000000000000000000000000000000..19d0939a1d3331553c88b786c01bb059d052ad80
GIT binary patch
literal 7204
zcmY+I3tSUd_Q&U7BZ*Eo0URZ%WH6@2q=7J#02y(NN<dN}iy&I7AfUM7Tft|wTR*6+
zj~3Tf*;Q1m!dmg|r{JQlu0pMkKBQO|tN8C?T`LxC@vYdM{9$qrd_J1_e$FrV-gE9b
zXKrqqCo!>aUy6#K>{&E3uQ2BoilTy~e+7l}@=F6K>XDA37L|Ur>DS`TQN7MSO@EOy
zzRG<iz4}6Z=*Qn($lPRYJbdk5%F4exhF-Y(sr|ys$NfH7bSiUI!t{X8tByY&vaJ2L
z-EA+w+TL%>)iaA;-QRoOhKC_vzx(>o+z8X-gHs|e?q0C$zP{g@0R=&i-Wk`HKI7Sm
z#xwUHEq-2Pd3bZ=U-rHz`91mX-!;X3v9aQs;^E_>N!|?#+iTN9n=j3s)U#-A`^wwT
z?&U4!KdJoI<H^l1jR=jIpO*X96;I+^=Hag4xg$*(^vFS<WlwnDyZ_T|qm%aDseM=c
zEW2=T$n!siq$>*+hlYf{apse<_)(>+>gQgMT6wU1b!6pA^YWCG(&mdfJ#vPHY~Aii
zR_&dBdcoG)FQ(=sXL~~av?nAa_S9S5#tsh-D%!PK_0ge{VbTo$c%E2o@SOQ-&DY0o
zU%Q^3vb~Di^>N<9tOZ*p4O0fqNr_1=58A$b`w!WJ=KsdLTddxC;_{VcAL(z;DyjOq
z|Lxo)p`k5lo8j=61@U9@=Y2WbW?cKz{YSgUDMl?RIK0N0V~YE~<;#l(Mj8JtOieg=
z#M1gmwR_xl{jrDj)4%^Ew|K<+&sWSJyYG+1*WbOK(K6=Hk>JrWzmzoHJoQ1@{D94U
z8<Z1{FUYd}weCph?~jWP^x3xR-g58SpTjFQB$f^P!>sPH?O8Kt{5vXUTIg%B>y}Lm
zeZF+#PjM-28&<v-H2j|*P*2w^8GR^vO<i~6j?z`*>w3*g=+!f!Prvn)@qG=~x@vXc
zm#@~h>}1b|Irn_JGo!R@>6uYQt&y7fu0B&H-Y)p!hkHSTnks(_`uNL!<C84aXJ4sH
z-FWkUMeoM_->lDj`F3XM)RJ9&Vkg$>JnX&+3(Hn2ncM%GvtpM<_xhWgmN$Jg<>!_4
z|4@E<{j-~eflq(_Zd}h>!tYeufVKZASUfX#v6Q3T-8Z7I7li+lrXpjc)Rq3k;fmqI
zDOE@}YG^cw(67hBYtzSd8%0wCWr0CqNT5_&RfoV9Z}_pYS=&xiHd!!2F+v)FRd))W
zP58Y2%+Sj20sUk8i7Oy*1-#z*K`A@tj4N&vcP#qCK~rXVjDd=Q(ipwFcZ`v6X#HOw
z{yAhWO}#D)LKGp=Di}qFATi;^tg=nv!8GNN1qF%%Nzj+>5O`OYyqNmKvQsq0%YsZr
zrX=X~N{65*&24VFIPMlrMaTkCAxdlL6WA$u+ew{YetZp0MahC>MY1G#t!Iaz{N%KQ
z1122VK~n={L8u~Bnj#1~1QoYd+`f~3IfkaJvLIFwD~+J*(IKe*Vc+<SeHY7V$|MWI
z6k*Z`VXt-wzCX3OM%icbOPYE^7VrvQ8bSYuPQlalOXqC<ATT(<7)yzN1I^!{Z|8>`
z?8a^{);qqrJG5tjL6Q_hk75Y!{P0WW6-Go-D;LoeCr^{6NR!qPs_dBNzxiM2_pkmb
zf~G98AVLu#ts^*~L-1mM!{<FRU#p-gyDXThm@17BuIUg^^YW$?%<3&=Aba$q_#Ls~
zLurv=8H0bPMaIaO2<ePElRv;Bt7F6b!4g@PH6SvwCPGG?G0dOCqS0tH5wfrrgFp2|
zi_u~tWHZ`KelJCb&JpI1z35;a21G`OiIAWZ!u<Oz3P!<1$i=!0{xlO^Mwf|@+vqm=
zi&6CGJYoLA5<RTPfXL`E5%TK1VMTIgie97FL})ZWI=WHbN-=}aK$OjANBe6+%;)o?
z{aYqZ<EJ5-&QC{F$QMTYi%nd>FF;fV)-tet-<g86SOz1?;7!6iz`6t5%YiEg2eDTG
zSAlsCB=3Pg5_=_ZmDqj|xPzE$z*>XHuYnOYc>ZHxJ%;Tkfjf!6e+sx$m}?=q7SCS?
zTphNb2kt!P2Cz2Z@f%=71D^jfSTAFHBXEuQ`&WUxin$4roACTMfV+Y1H-Wo}xfQIf
zc>Gov(TeB42iAMo-UeJ7{{92t9$<a~$xrb7Pl0=i?Jt3QNm$O_0M-A$LOwkGyln4}
zapAOh7&M0=Mc&39&&ryLMn<PcN%bDbB$5hel9+Tv8B8`IU7RkFRKG4smyReymraO`
zV-rdBvq@|^q6{{hkTK4fNP2-W$(W8P!<bFT;<Ut*&R|Kfq#{bQWD>GDZSkaQ*b;21
zh|+ACgd9#sJn0&a1V<{OG)E>O!70R(9wH<Nsff~qOhPWFE1q-|SAr`QQJO20klX2w
zm-mYn-3jhgL}~6!LLR3lo^%^ef+rPGnkSQx*XfNXy~ms2O+}RE%_KCAA19K&^fv!C
zqVfEAM3ec+BI#0jd>*1H{1im9`Pm}rU~~96i01Nh5iQ}Dh@^wP!@q-QDZkWDK?hdr
z=I_6AhNeY}*<vM|Y>BjRi0qaaLN>F_N;cURY2y&tZ83x#W`~t*vLn*LA+kGS2nlAv
zN;X-D6gWh7Ax0)SVWk4}9nnYuUIWcFNR<n4fTHS?zSF}-2hif+z`;z%35=p>u@nMJ
z!O!x>N~8JCb7ZugpxH^v+17cwgC3n{I$0QXJ|-k9pjULTROd=Yi)vY^)1rD+vRYJ=
zN~2bGLkhdnqD7%q+O#NqN{1GONGWLLFwmk)>C&QHSGu(*pOqf1>>4e4m0m5$%k=Pc
zj^t%lcos);FF!n=Bl%SrUdWLgS^@4AkU;`j3`{W$kk47N2T~w?hgvy4w0IUYXGtLC
z07)|6(U!=51kI0R_vBe6neTv0WRF1eh{#f6H_YxGv2J~LT3ibQ*Mgm_ZXFC<hxtPo
z_#x(vFmfY|kylTPn_%Q7%q1|ggs>bTX+6I4G+EDk0h;#${7E<38*IA$m%)l4TI{9j
zrIwwgMXgG!CfT4@>D43~`m6e@Nj8{OX0;qkT8vahs!2B3RdzMW2B*rYCfP7VHAGFa
zAzl@)CfSgrN>Y<-NKvJzQ8uX3)hHWOS!&r=TFh7Ft5G(n3f1zy(Bc+&bqj>Z)i!Za
z>Fn$i$D6y+;w7+L0!}vgc;qEWBU}%bdScH@8Fpg%p^98JEso*GILS)J@?#Ob#lM9p
zhtF}6Rp#=!h$iuq5Y6CcIEe={`I(4j@v{&u<QF=LcZ>K%h!*pU5w(D&1)_*UE%0Uw
zNl99~3zoaYK6_-<*n`g}ESLJwsK8PFzC(+5An*?9L-N*13FiAfMiOivX!eovk@roK
z`F@oV*>2G6MzU;OVkKu^AT@Qy%tVkpWNeHbQ8XhE(wTJ;Bu8~NogGoMP9VgZ*$9%q
ztc|rJie?2uMzb-3<hIdfv?Gc(3No$O`R64myX(#lrp2wmYy~|D?^bYZ1(EPJV76g<
z8EDJEK<s7UD8u?4pxuG(<)AMI6VaE0P!7WhSAf0(+xLKePgfT9V0|U%E3y3m=nr&d
z;Q-bj1pPs5uLh<X`%?|k)i8vt=O{2ovAqVgHQ1jTaMWObkAe0Wwx0z3Nj%?45KiLt
zoC5tRY_A1<ZC4g*vAz!Ub=ZCm^yj*=a1QIwgZ@0WUjpV5_U97BTml#Iw;q^!Y;OQ<
z1NNr@91Ymt%b>lC?Tw&s#Pc<R(1_P_74%oJy$SS9U0G<t`Wv9Xf$hzpZ|=%MGuGb(
z{Y`9d0j34}(*n^gcz^E#a~IoNLEDP`X$40s_V*rW?_qlz=-cpoZ6LJa^*jLm18jc+
z`X^mkc!Kp$LH`un+d<#nm4$Y!e+l}R#Qy&MgucQz>vYnKpc*XIq!0S|&1&2ykAmeW
zv2XcrOqiEyJty_a;J{$iC%=WjZ%Ln&n~>B1eYZ}e0p0@5Ei$_NUX*0MyC)($4Vu%W
zPs)u-lKF0+i0lezu8`4AH-A{y{6_H#X>F!J6Ixpp1Xhu?$**Q<G~aa<8SMmUPLN6E
zRF}?@9XD6<;Zi<~1oY^<sFDr7x2<stSmkfZZ%Jt)--Q*K=s9SflV$%DrFami?z7f+
z8ew{|q|<8|Eh`6-7WIstl><SG{h9u(+?8q3%$Qkp8et+?bQ)potn4x^IvFP`N1Yaj
zFhf}Rl}3y4OgxKDBTN#DP9sbTE5FrfF`Y?gW$$P)i^*cidp@7ZXUV&~kSS!z`+5s#
zw?Kb6KJpr+`1tPYNPI4W<|5gMJ2~Ms5sqK%(vNpW$Appi=ybX;5+5C_GY~SeMgxhD
zMP~^k@v-S_Ve;IxXk%>#LINuoNPJv6R~U(pTjvfV@o}?m10gT#HK6$L>F9PVi_b#0
zTNC+-i01S2(VbN}B$wkmjy=HbA$J<IcmTKqT{-yx)}MsrlbCCPtHt_rz@6*L$>*@X
z5t17*Hv!j#_07OFcje?}tZ#$lHq1|edxG`tz_oYf<aVNeR#oz^$hZqffR=CMJAWt0
z=_DnJ@7|Uq%CDgLl@wCDX3XL1vukfj$0?WDWhEQuiga;^?5-F>ZnN7;w$UBw<`CK4
zF@!v3kChaJC(^?qvU_3(dCgucNd|AEmqTRt#>g};Z$^&K#e#X03MTp<hWP&L_!}fY
lP?VEA#*s&H_T;q}N`CQN33Lr{4ie%9Q-g`oMIHt6_&-w0kVpUk

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp.25B35CBDBE38B07B.idx b/.cache/clangd/index/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp.25B35CBDBE38B07B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..76c32d800f936f1761ed15e3b84c5aee96ee81fc
GIT binary patch
literal 13184
zcmai*2UJwY_s4m6*7d<Y%C3SGX;P$#SSbO-hN6g~u@Mu41!KXk&k~JN#Rf6a1O*gK
zez9PU{$jyc5L*z$f>=R7MT!JeFc$dF?mxp^&-tJ8_nZv-+3($X?Y^0N?@UIA1P3qc
zrKdM=TFA2K>5C&3dU|?l`ro3(3uo@7;B2g?w=C`Y_~?-pJ|+b_eP4vRdmGNL_0-tD
zyEg80P`m86<|msRk5rfVKFF+Vd1v+f*O0%4$>)yvx&D$dqc-(&uc<RcQC}<y>C)$E
zoaMGJ|8uqJRf=+&-#?axbAv19{l3gUxgdGX$o7rjF7F-rr+<Ji6clb=GqA&k9qo-D
zhR&#{AMf*ab-xWi$$z|Bc5Kq5VDVzZPkBART5T1OKd=5^MNywF{l<Ko8agYlOa5Y`
z$SBv}#`(v`KVFlxH*4XHc*E2JRrHHt19FtRgA(?S9%^HdH|^>Z)0)X=N3Y+S8FhE!
zi2%bZ4qG*)KA*LJI$yDVL)_3)F!Rhe8~e62>*p6tAMoo}C4RA2SB4(Dx_7eUkMl3a
zgwD_IJu%bqM+YC{A5A0fc)kDo$*ZpANn<x`HW{t=+co{-J;w(8bSBSoaM{JVaZ@dx
zq}&+YH_E4@(tOFDXXhOI+<xNsz-38Nf4lgilS@}FIy`uDSmhe`>|Rcb?w5><Kb`xi
zvV4Mb@zYJiC;#=*;_ByH<J=A`J?YtPZ{1Q$*N!P^mia^4x9IJ@+M!R(!t<m48ug2F
zw^e?Y15O?`BO^7!mJYVF`~P|_e#nZ4S-0kn4~lhvXpwNSLzjRFodTsnzYLoiG&_4&
znml!O$bk4c0x`?)5TWTDqW?=|M7;Z^3#%4xnYtond*L@7??mjF8T(Jy=QA$OEKi;l
zpH(_z!TX7$haVreV)vB<%ct`ATXUCea}F9eeajmAUw?Hyw*F?9@W}<AH~5uD_co65
zxuQxk9TRJ{Yj$JhhnWriXEi;Yny@AQ?66VN@Sz*74g0;FN5q&H?u|b4gMUddOZ5(S
zdl>rW!P5=nj|$g%*6WXWJLc_#UIwDs>Z6nL|E~Mt_QP!_9`EdVbj$Vw;ddIH-LDl-
z_-W|y^^;Eg6&bWR$l#~URlO5@Z~fZ8Tc>9|Jq|QnZvEhK@mA(dFFz}({j$_4&pcOM
z3RF(B%6aJ=+ih^LNyh4p<2zR9`G?sTPW9MRUAV@h`O1b>IagM^EgVzc?WYB!2R$d_
zI`&`J$JFxr!p7%ii@l`r1L9L%%vHg;u1j8@?{w}PpC8I<p8xk;gZHUhzp8@6ji+UM
z?>MrjgU{s|VFykR&<rWh8%$m=ZvXbn#L_!&PrT~&@`&Xh*_YC579^MX%zFJpRbZ}z
zLu9|`p6PF-%-dh|J^7#6KUehm`a$mx8#*r*E?rnTbAZzC=Wi@R6CVXGc$l)&Xt&M%
zxijyRW^;di<@l<|%<SFE7ROHXjxG*+_3={YL%%vbUH$U!n$?TX*giV8`<p|X{G;jy
zMlDW_I-mYn(cC{K@|2j{I)3@tEuY`AjhnNn(C+<6>7T%$9TPf@`ff+=F_(c~J$22#
z@&G0*89uIha@p9g%XTzG)-L)0`jcx5r`vuRRpanFqhn@ZY)jyUgl=;$E_3i$+^cjz
z*&fr^gCefztsedHx##NOP4>sb9vR>NY=LWf%7wZCuXBrg{5Pw<_BYqW&I_ZL*4Mh<
zD(++7vLwT6;<XvC9<3@@eH^;7>_6}BM@3#U?;95pG2hYeeEp%NV`GncH$S*)zSCrh
zWYw|vN2Th_T;FXMPW|!o1)nL4ww^s7cOm$8*>^rA(-gj@ZhMS>>U*|T{lU<^I{J~`
zqUF(3zG|nZXA$t~kn5gDdsHx6PfzwxdYI@MI;QFVIUB_IzBuxAy$VKC*;VPP(KSqV
zwG^jhb~gQ61rw?4sB|>ZHB9#G8PnLfu*ZBAjG}V1GFq)`nB4tZs#tS+JUw<em7SDM
z8ePL=uk`tH@uZw@RWOpuo=Q(6UBl#nFCI<#XR#((1(T@kqI5CQHB277sj*?!!4DHu
z5KUz(rIo&}VRDGy{vh8m&yT5KI+eYYUdFnH$zeY<MU5U}_u2&337_l9QxsDqUBmRz
zX@LVhgQEAUU<Unp4`mNyUBl$h%f<vA8QAWE_U9Ay<o=5OlCEKTM1z;6`kd;H62{Y?
zd$;#i>l!9cD$1Sv=fU7dDws-TTcxeOu3_@*Q#mh2SFPQmgfUc}+<vmDu3_?=i0*d{
zS9A|lK^T>1DQBs5O`Ckoar4Ftzg<#61eIqiXRCD$lV#^Y11}uA<)?x<RPL?pZKP|M
z9D8HmDff{lf>iJgmED!@Cc1{n>k3_u4|UmkUeZexsD4IJoDsx-|1g^*hg0u=_6Xdn
zf)UijLFr(kYnZ&Dqxzq({<)ZJs5g(wi_8|;>KZ0*8U5R#rm8=ts$edaJ1aZu>l!Af
z{WW>i`{gD_4fW<zd4bsiTV2EC^z>ZEf%*-mhI$LByu@sYt*&X4A3eX=5Sfvof+#9?
zRd!YD8YX9~ed;{bDEFuemQdM7X``=enEXrZyaxwcPtH)mG%7nQoz=RA$%hqhqVxM)
zq3_aSD%&gV)w+hse;ACay_vfHy$Zgda(88SBVEJfGX@n&$*GN_R4|#!Zc4XKx~5G&
z;p*x;^;)wEW>HyI%I3O;$rl&B8I(8TW}FJXrt%!+oDRB%$$yRf_Q$->$DjV!U;Pw*
zlB{i*ej`_@=$G}fLaUF{lgBB>NxFvV_g4QWtI#I$xDv+F?uWG>rq(q~ei*WA&`+5j
znJV~_%5#--)w+hsk1~|5!}3laS3x9|=PBoP&^1hcekm-fpnurc{~mmVVuU1X8>Uwu
zjGMITRq$OEd_{lWQ`yr<*D$$W^4tAERBlnhOe*(M_A=5nOnzJQ_<PT1Uw2W#QYu?0
zE!yiECbxujz5DruV@C}27EyVz*<xE=!{pY}XCr>^*X53(-U2EwG+St^YnZ%ndeox!
zE_!;ThidQ6{3#oaCHWD!J%a9S?}_{v+#X{t1&>nb!Syn5DT8jDpMb{`te1mJIo2z{
ztpfK~34JT!GyeM+5)?xQbB-lJv6$n?z&J9L>+xh@Jmv%vlz{Em5#MzrklQDbpd_v<
z3(nTmx*ax9N^&(gRYPBHSOZQqm}|kg7QDG$2hMevUxL$1Y+nyf^@x4SZav%f`W`6c
zGO#KGYu1rME(fb}&W1BAy=U(4e2yCC5KRv0$PEjLQ6Vv5Y<PUnS(kS6&$N?dUxja*
zw@Gr4A_!5aA{5bB#aKj<ipVxH2|H^aT^sA5kmPsZ^bVc5`~jRl;4A$I&L1(if>SHE
zH#un&-D~U6LXGy8s;j|~ov<X=fVu{AEof@NiR*Qssl)sd)Gx7pJ*ew3H-fqm^D9ul
z!u%REuW^5GK=TH36R4Z8eKV+=F~0-#JIwDv{T}lN(0suCeFV)%%q^g9!S=17Zslx#
zy~m2Hd-~rNB)N!~7m+UPMzENty}>83jwB}vVxpiq^Y2h9$t9pF0aNxfNxl!N`<N4n
zYa;2z^(^9?MZ7p)B10~baK@cO?&eoLH<(2|Ach3Tkk6SQ$+09j7IOj#PQbj5^jk+p
zF#FEOci-Q#yjr3cwyevtaqN8U43g)^n`^dF&(4L3xiE>@E9A=%ei_=1$AY?c2A>3T
zk{l<9aY9GdpCmsa0Z&K>>qe5xNkBQ~YGPbXteEa%nPIZ~PLJ6tg`5wa^1+q~3b_C}
z6=1#r<~P8e>o>vtCg$7F={B}6hEBzpOQ2H;xbxrNhfeqTAgGY5h;<cdb9{yTf>^)6
zTurR2IlF|v4ZOJiTnb2X9r36mXb}2`c>Keeg}eF&pM<%Ryhae$V8<x{RRMOC8t|@x
z{_K1u`4xD-!W>InW3eNp6X$g7NO`1N9zP99zDok`k|C@=w<Wa(O$L_dsXw%UQwszz
zfyLi~!KZQepFUT9)?d(<Uhhld@sjv3dpa}`k9V9|fKFc&ZFJ6NHA!A8h-<N<l#qZD
z>=^e*z&*^RB%l;?6)~>jjw8|MWBh_M3;d~nYNLLB*yZAK6Yia<&^;9olY61hUViP8
zd>ulsbEoeeaM7S-c|=Eg(YeGtm)P)&&Lv*Cq-|#M4oTZtly-R%HGECfuZcM~T%)jF
zqi7p%{Y);TA5WEA=-jp$#LZyD4U<7k#+(9T3g+)Y{2ucb5Vv6d5yT%k``I*Ho2`s&
zFjUa!!RQgRO;HLOJs3U4Tna{|n9IPZ4D%B(dV;wejLI=rfKdhJN-(O#oI~7lhzGk+
z`eGBK0zPGG*|~tYVEwZK(&mp^eL?|=U^+|T3=BR=-RO8Ih_U#6x(%w^_<ec_-Y@a{
z(*)j4m|GyQ1yB%^KwJ~>`?ZTW?;@WJ)NYqZZ+2ReTt))Q`1=)T<)8Ik`SjV;BZ`Sj
zG4bbyrNpI_bHBPfFAwg?SV}!&12Nw~?6~1ZV!jdcCStLPw2gq&gNVgu%*n(&8QZ52
z^Ayf3xfKmQNopn8!P3E+<yVsIXz6Ip#w<y8wsf{;DT*X_wd`u0{%^V?$u5>I*6e9W
zvb&|bHB0m)*~8MqnoXI(Iq@&&)JE-WFUcpt<0SOtS9l6MPGLR`9;Y$qfJYAIGvILs
zb1u=(CC0qJi$wn-XO^$iF+rQG^D$wiW~CJ$6LvuN9e~EO{m^GWpy;#+!ir!#JIJuq
zv&jXqwT5&?UMa0q^KMp2s}RKqQjCB&L68zSkMQ0y&#c?V6gv9a``8a;S>A|CH^T39
z`}<0J1)c4UV)(OF(AnN77IPdiisR9VLXIa!@t6~cQ3AG4Bu0stlZa6gXI8opKv%I~
z#e%S$NjF1B2JE7zd5RdG;=UAa_aa2)zilg>wy%TTbvz9hf?XlzBCspMd<*PuVZH-)
zcQD@tySte0f!#gK55VpL<|-0cMf$O)qn)+`Utq2#fz_PD5B_<$SQxT^UeS4CeV$(t
zE6z|~(N<=VuUM~i<_|dyx~E}Z$%a1J*jH{r*e&iWtfWJ4tG1lODmd&f7<^K}k>ojM
zbL?2w&FVQcZqOEU_z88_bT?&rza;O4etTg6yL?IB2mSV8-VZ_hVIbGDAt)R30qA!C
z+aH8}2QeRleup?u-n;j|`$zlEu$Sb=g6(6$fp=Ug*p^~06Ku;cR|>Y3n4b!^Pcc6e
zY@cDS64X@!ilbi$>KB-+1$8y%8bMuyxlXXG6VQa+AlNqG{u>3`M$E4S+gF(12<kVu
zza~N5gt=KzH)DP)sNZ6KCs@A2$NwN$d=T0)OiBJISbW6XB3QIwZWSzAF|PrOH9RAw
z;S^Y`#T)|`F_`1PCJx%t7fFr>n|RC#V3UA35o{7MuLGNPn3KRJ3G;feS<iX;f7gaO
z-Hi1V6mo=r#2{AtnYnP$(vfd}3!+1R3{k`oL+&rJL=lTQjws?V#}h?7<^-Zhz??`F
ziI|g!B8l_t(UU!IHcqbBzUN?a9y+lDv!RWKx7s2od#zagQ(^E)-BXg+3F11T10Mjk
z5UVZ3ne|4~I}&i8mpbPBTPLO6Xp5w*cS%kX#3bC!55(#R+|2_L@PKzS|I~3G&4jIA
z(yT1dKX4GMv?=7l{(}dxL1pQvF6{zKDtu{#5$F-<#f&6*u*YC8X2dG9N`p@-vXZ=7
zT#drwozQ(J4vP;#pF=n-z6YWAcvvjS4<PgbXZdtkO!YJC0rX6Ah-D7x%5P60F)Ji(
zqlDbr|H_)BdjsfGc@IwSp>5b+miwe~g`(&>4T3U=Z3bzJmnAuq*k)qhM{M_D&L+0m
zm=6=%!<df}+oPOg4unRJO)t7k9Wsp!Nh9O<siu=5>6kOfunaPh>pRJ?otSr#A-k}B
zCJD(T6S#dA8Ir|y78=p$OIrZvo{=DE5`@m&!HPjujKkJ?@UF*UYBP8@V{QfSR?JDn
zH3^5Umx;+`V#!V`ZufU9M(oMEY$VC4;FJoUOpxU5;Itj{4shCmISriBFsFl4I_3;;
z%D}u6oOWW~3+{UX#moD^eIMri;JzPoHn?YFJ^=0qFdqc>gP0G2`ytMWe-G~LTe#7(
zqa<6ISedclmzJnI_*w9G#m~jh6H$QQKtyBwCUWxE_?vSI93B|XseiBjzU+;oi#8$s
zM{*j`VT1*%M$;O!%Lq?I0V4(?8Z%-dr;(bG=A6QZhle8?AO0n$alOX*vH_Qt*h9vR
z<TRz%6kitH(bcGsDI*!N2TupW|6OrmF3C~>nqX<yS=iT-{GIq6s&nOoDj)l4HF#HJ
zKWqf=M$GTP`yJ*O;u?efG>tf?;q)qxnB;N4)n-uv#XP-Y3qCZ2(ysRKlS~%GWFBX4
z7)Ijil3RT`N%BQOyeOblAx{wVFkceHOPDVU;$_TN1n~;ytAcnH^EE-dhWRf+{0now
zAm(GfE{NAL7YJej=0ZU%#C$^#Z(uGG#3IZ$1@R{4TY`8C^KC)Ajroos-oacfh{c%q
z5!ZeAd1MpUY|Mv=>tW1CiR)3$8`t`7-@l+~mTvy_IuCuAyRoG>IzwvL;rN4dvUIX$
zwIfM(wRE*+B}Ykivvjj&Ygt>Hzy31Dxbq!t&&2&auW;?yU6P$RLhn8eLBjpR2eJ2q
zE$&fI(XQ<A18x$;P25x1A|Ex=uJCcQ?*;LDWcF|NONVjoN*_0i5o}`wH$IHt1l3J2
zV^>X!gy2&LC=P!EK5sC81fP$X<B3~5X}fEZe2JJ`;?*v;sHidcWHC{aJvE-DtQ026
zUK%e`mIO<34^0nKmgY!uPfbr#whAW6y)?Z{*`k*u_tx|_Ws!q+qVP=xRKM5`zS}YH
z0KXkLSV#lEG|cJXn~v=>z&8W)PVn7{`8>Ft2ef)sMZTyaU-4V{f_(7;b2a&*n)9xG
zGcMddzodxz$2`Az0c@QxbIx$Zu<(R(QCko5dc=ze+IpDRW6WhRs0@PH@w96}gUT^i
zfL8^!uLS=}o&#!^h{AHnDE|9G;#|n<TUqs{u9eje?`Q<Ffw*lTXjyL~aodP_6Y<zY
z(6ZiU;;|WXGI2}B_9?_I1#=PcE8?jyE8|et)h;t**VV)b*4%Y(fa(Tz-CFRe#jg1p
zd|qSz06rfu#}T(U?7DemSRNV6?#D0dCZs-H@@Hon>NSE>BVVJ^R=AyBVSWwHuW?lR
z2AtnuZUUz!Y~Kt{&7AlBvSj?J>EcrA2-}JIb{>UhFFW0CRCHD_6)cP_%-CY?A=hVP
zzi0^eAlizfWdZ5T&REMm3iw>5<zWRRl<BN`qcZqpZC;XRna#3iiG?K3Hk)nF@&ieZ
zHjB1rrFKc4Yc|)OEqF`vLlW?iCj*jPK>{i;KP3TAG1rr!^<)e?9U5B`xA*t~yeC87
zbNz5>pqy$x@*BEFSq7eEcup<{&vMLhBqWZsEeX*al7z%#P9nicJS09csibGtMq`br
zU9i@F1Sr{i4Em2Tmx6vN<|m;41ak%GS75FL{YuO^#4v}b`4txs{Q{o+vFZq|<!HA=
za8F+$u0To8cIdtxyVE}Cvk#A3g%DcElbjR5TbCUSdR}5r2ModSv49HJrGjHA<}$&t
z40EO6Sc&<m;P@2tGr{p0<|@IgN@xq&X-P{kdx5!HFssH~Bbe1-t`ltQgtnwhAvXw)
z4Y>bC!LbqZE5Y#<<~M@b8{A)$VAh1WSuktH{8lh~i}{^k`wk!fgJAPPXe&i1Xz9b|
zBjy&trUi4WVAG0u4cM&Vr3M8peb}tU90N8nnB%}94%(7XS{nt2c+3glkbpT691=0F
z1BZ2(lfWSf^LlVt&-p}DS^gk{YfrRm{)WDtxVHrv1|bSH3`I27a4e!o!^k#baSjbc
zv>SGKckzN4FIcc)n*F6_+KoHhER}Rf<;#?R_8vX0-QcVmI@Z}Z*>q!ZL+)43_J8)e
zs;xh#5yLcM!Coa+eIy2-?Ddi4^@6xwFl9%T<R6LEk9<;NsSqu6YPSHf0%w84%5^b;
zV~lkpg3lv7$v+04$C%3?v<%uNO}cChq2-t>z^4M+SHhr5K&v%5WLyquTO^X?LgHCS
z+EyA^LPXC`yAO$NJ-YFruStd6xd$y=G!V-Mg3EKn@+0Rv`pqLVjk+wNv1L?H)KHc~
z72BL{chB7}pWeneg=L(g?VV!>pvjeXV-tHi_s!0>tmtRALO1t&ZNRHX8;RFO%$rEh
zO?-V(yVlclGv;LCm5l9Eh*t{cA`(<Y2C>KbApYsyKP%cDr6;-##BCrkflWFz9M^8K
zVkb&t6G7a73szf+)mB`vsw4rGd=6q!7Iho#RxIAlMnT+&-DVrH+Q!|c%rM#Tf_tG4
zJ*_B@sNU?hve2DAE$yByb~9PDN{4mrhAn=OZ^duXu$~Ux)A6u=0Qwxj!}=Wvy~Bs~
z8ZkQ={0?j11Y0LtFIIK0nRRfiZ-)VS)H8CSLoR>YCAu%7!)4AjKOAgIcQ4$jlIR8o
zMLrm?J1fx*42lBGH$ZU%^G#6P#C#hRw=oxkq8M`tC`vHj2gQBNRYa^JXhHb}5no`g
zCSo;b7VXo!somek@8&AaDog&JXF&H1d^Zn5pM&^r-i6S+{BFJ}S+D>7yQyk=+rmA<
zz1V}4<Oq)lFZL3$fQ!aU+O2c^$eDteDd_WKX9;2!=G}t08}lAP+=F?qAnwJyPZ0NE
zJ|c)mFdr4fqnLje#NRQW62w!SKb|Xa-tNCF?f-QNw*8L|&Dy<x{Q6P^F@>j+Y#Sgo
z)9wM}W}5|ZGdI&)IDHP?;z<90k%dzioz;Gbp}S<+zi-(GJtAMQI1#?K{Lp%%I3Zq$
z6=H;C%U7&iwfeg?!dl^5y*21pE82dgw-$Z9jDNj~e_YK!#v|k9{Oefs^-BKrcl=`<
zvX~k@Z4TWBtT$|WC{eUiwioqPlA(#Q(NOiE0nTcJp_)z|I+|Pf_<rUe5ZEus)70J1
zp|6YW|MjhzRcC8E`!0@7U0vNgx_kBL)yLbWw`I4U15M3>LlOiSJf;5-Vfg<5j;7}m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_add_relu_gemm_add.hpp.B9BDBBF21D219B22.idx b/.cache/clangd/index/batched_gemm_add_relu_gemm_add.hpp.B9BDBBF21D219B22.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d037c1ee38cef3cc3f1eb910c6ee9901ae234982
GIT binary patch
literal 2568
zcmYjT3s4hR6kQ?#Le?aFmS2GI2Tj615P{SHic5$HLB19Q6)^}XBnGVns!%}$@i$d0
zO+m4(3L+KIDnd~irHWD%L==BO>o_vhDx&B}H^I%moynU$d+vSr?z#8fjU+rQ3^IhE
z&~@Qkv(x2ix)1~zVLzEXPZ~vnASMh!TWiMS(oPS1nog{Eebw)s@bF-E4pq|k%cR%P
z(AM{YYj20m?z%)so{TlmcHA2r&UF3lWROM6f$g^^dV71Mr{2tTA1rO9c&6z1gv|FJ
z<y01}6V^vxr8KP3Tbr4n+a+-7xnJ=^Lf#SIozIH5zEb3nCly1+qlxRgtl!K+IsKFE
z!5_{Y&wjz0*BR58vEMk+VuF1B;K{)8eZtI+p3vK;`A?<;o1U*^9`)|)%&B3QolR4n
zH<)dux_5le$oZC;ab(3hyMYiR``~DkJ*K0n=kh96#HNlF)@|Hhos;o(rOTN!!Y0yX
z!<x9KH-3x?&Dco!@a9r;W0YgN)w6*zX(fM8NTlO^>DkCSG&?Hu-S79_?)^J1DR4Tq
zZJ%NTV_%5G+oQ#)ZlG%H0!cM->?a*p@v;YNx^{5dgGXrxohf%enJG9o6s$WbtLuAb
z;HH|<um?Y{Dad=CtFxl;L&eB@ReiZjs&ML&>fwZSY6^95Q`h6R)XAi?)7+t2^2ffN
z%B-%hTBjaeWHj-aH1e)1$Hu${;~wqm^!V=&Hv`*-R`>Xk>Bed^FZ~Ke`4W;_zi(|o
ziv0rHV56gwj=|<Iv6Ey&!1?b}c8Yc!=cflzq2?2pA1{%p>iP<Ax*f7ym{mEP(ox^y
zwy3PBFRJHtFJ*hrhd;AJrzRer_;mADkcxe3aEjK}SLiwr-a1kix{FgT8-Hmt?qIX{
z5R*%h?kkVFk#NVLi+4nIf0@Xj*t<~?)A_1u;ycb^;en#fpF4!FC$A^&=5H>XN4JRl
zN470{%`>dO?5(eJE?&FU3qw>I2C)k<Bb%**D7qxbV!jTdQ$3~YThmW1fgu<-SV&%o
z8O+m}GiXqq&U@<68EQ;2q(jIi#wKgc0vXJ0cTG06q<eXP36-RW?M>68nP?V>qM@+>
zBf_N#OS!^NdL#;sjEM9_&YA`HA5rNo5OrT_rIPe9f4UysM6(e7({b$re!Ax@m1Ka$
zPc}?8)-1R`l$X9<#=~Oku`cWlt@F2|!H79(9-)14#8P4zOCT%M3gDO$Ccs&v)&R#w
z*%4ZG5Uzw9OCS%42Y@_LPry>aRe0i}!AL$gA4K8#VuCC1F5nSJ8L5oMk>O13<KhTj
z3=AWVsACL~z7y&MAQ#jHK>nzI43Q!q<pU@H4FFKQC_Z+sd=La938Dl5C5e&%Bo#<e
zyaF(iDaZs+mLN-ujGA^^Z8`JuDwePfY6GZfi`oK+C1C-2+o5&<XD_h_I0w`L;Ji>T
zKw)px8$do1AD|Y#s4u|rBs_rgL;bWIE-8tkIS`dsKGrro3$%lgZH#T3@G^MOefXu5
zTUfeSPuZC-!+l_cZO2|lNFCu25l+aNj3fu)l8LH2t8N#+nr{FjmSW3jVgsg_2_Tl3
z1t5E|J%G4kZZr|fL+k+{AF&UBcw(Lwt&B8weX&1?JP*s9VooOlvngykv0d)`Toe3d
zrFmgjXG>d#bzp=RKnunXcFFOgZ<hbseI2W;43XL5))?}WEhdm0kprCEPY!SjON9er
zso*KTB2d0%KEUPk@&T^Ew!oFJEC?tFC6Llq32@2)rIsuGcd7NFI~7(~(PV6y7vUlk
z%OnJnv*iFM7t6IAE)Uik>QP9vMiHY(+Zwf}GOpa~+VU5s7CnG2z{eQAzmB11O5(%s
z%4lWxD7&fzvu5Zs5WHD%jmd^{L-WgX7;#3OfwQm(T?BM-q9`#Q-vCB73O0)HZo{Kt
z!PRpXPaMw;i=v*dgs&^=3PcG+1A!>3M5}-(*@A3slyOh;TYb9u9(8LW<}9GK5H<_f
z4n}e<a~*LtUZ&Yy<oZtd=)s5uY5|;U4$9G<NEmTP-2tQsSAerJvOqf>VMOVz)J{(P
z?Ma534+Aa&mVr|)2hU)p?QoZS;LUSb0o2mq8em_KtTZ){<*k*iR^OPZm$}Orzrj9C
mH&hvw2d=eo_07!~#zrQl5Q8AKib9Yn@zk8SRsYfZa{Uk7f2HUE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_bias_e_permute_xdl_fp16.cpp.6413F9674606CEBE.idx b/.cache/clangd/index/batched_gemm_bias_e_permute_xdl_fp16.cpp.6413F9674606CEBE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..394f471dd928b29eb09df39f9c41d62022d70314
GIT binary patch
literal 10658
zcmai42{=`0+`sSN?RL4xwRUay+9Dxok*=krC`k)#v?%SSOr!nKNHV2GCCSh_+R&6X
z6_qlSAvB7VQbgONWRTGEz4zAZyglFde0iS!=lt&P-IxD%@*6#JWU@7nH)7uCH6im?
z1o3%1o)-SMe8sZhQVd>-dAv3IZcGVSYBX5$vD@9_k<-S6R$D}xtZ<H+8xXVqd0fq|
zbC;vnEE#=j$+ODUmp09_b^o_&rJB%uVNUI%y!peuhR2jFpPyt}-1Og-zZ;(%JRMzp
zEOAAS;rO8?ujci4b(wP|+{6CfDA?NiuuQA2p(Uot|HiVFMg}Jyto`p944(D#^A(@k
z!`usAKWo=C*c~zB)T@rWA6A8aYTvRwV`p>Enz|KLaO?be!#j5d=3aSpEa9rksHC8u
zUK|~tRguuzXX3EIHB73*iKE(S4l^p?K<MXotAt|d^2-}GzP_m1UbOqKb-VA)hPT&+
zNjf(MOc=VWG(fM??$XTON3G36O{9{>#Xs*~(6rNGfTo5`!PPSsBfOJ`O%QKOymx-d
z=jSDzt1X=GNu8hWRO1hH|4)2YFm|YXNgUt(<9_Cf1uUC#Vq9!w5n70X(2V{CaG+VX
zXt58Z?W=HXsK_1Cp?^f+TD#!{d%CnubL6J7MXZmG$%=&o-Y@%5TKHFaw87$4Rrd~@
z57UU*C5TxdejPeHLS<p$9;d?S`j}-FNw)v`Z!a~oyIPPRKO;)p;`y`zmQ6KU+sy<<
zt9q<hyCE*fJ7q*YdN`%Ox`*WEoZ-_0Z$!%*u39<8r!SPhtGSRmT5W}QiqWa=!{0DB
ztz&NfcWCjMl)Rv=9e14nbaQgC%-tnVSJ!(Jk$BOhN2a6E+(ssAZfX9@X&1wb&q=+e
zyl|5Hg-;o#QEn`YeN((`(ZI(Go8vrqGp!qb-@|raGU9^SxG_F{dQ*!H{ER=R=l-(9
z%+|<mrttpxQ@KVjLUs1_k)9|@O>xxsIcNN@rC}ASn}0sq=BaK($PjO#V?<71X_K2I
z?Xq3d+I<6dS>GBgY>kND)};R4GW+O}J+&V#%pA=o_x5$O$xD7vWoC24W&GmCrymaP
zKPHzQ^V)QLLp%yR=AYH<f5uIF*+JLFUK_$U8oN7(jUTv?<++RgNnWsvRXY#I7yC3{
zoBY`!^VRr%VfMNuCu$!=PMy4KPUn%*Wn&{uuE_QsNpc9bwk?fssy_xZ7W6s&(7&l^
zhI>QeYa4^*5wk^c+d8KtH#~^n(6YjD-LM&F)cgDrv)1I+ZE%urZ4FyCNoB9kQg11A
z3tzWjQ}q?qC4%XNwb|1QP4q7`t=@L)R)tMc`=_k-2)%26i=>Z4Er-iiA~P6#|J4UG
z|KyXgs~_BXQa-uIrkM>o=UdjzxxLj|G9~@Fb6{QOpJs>F78eKh)4rP0y>sQhLGw=j
zG$g(wWxD<4*7WI93k~g0JiPNDyLbHU$cyg#%F_mgI;2bv>z^{UXSQM4ty(v=wIAN`
z;~qC#sje>bcu-|ASoQEftAv?7Jk#Gz3eRqGC^R+;@bKCC3)<1|<v6o__4j9++-do8
z;MBtGva23`Ik`^`eH^cQLvPc{-+M2bg@S!6(iTk{8uIkr{l80oeRTcW_4NEnt2=+a
z(7s_~*@wdTrU=ckepyScItC4SU6nTL)&7=WG}+D14v$*Ua%f?NyN$sr>t2yddk^h1
z>%q9hW$IPg)wU09PHLPV+nRAMdU^1i!occdxuV;nKmYCBJjCv9)bh2}0nf~K)(Wqk
zc67*yy?WuvNS%*IV?x)49D1&|XwAEpquw3wv?u>Qa;^0%z0{Xychw9uFY`&RUs_c$
zbJ21(WX{TNJYKi74^yLl2~|<&jpgxJdx^bNDX`vC{&hDWz4W#7-BTj)`-R6Y7B7}4
z1);^qnaX(i)HY36_%n~qM58iMz_$k}BAI2%zSc57E&^ZNfuX=qRVljY)i-L7uB@A^
z4GVA!=aA+(Wc2MpiWpRT-E+CbdnbbFxP^g&frd&!Xybq{L!)2pNLJ@9!nB3V!a^ws
zZ5owad|y&)CWh5ZdF*`C`SwadXxo)58=tp~F%rRKT-#P)tF07-wr^uE=e+ky62W*(
zTMMi;m4eU?Hn!(eZvGK2fH`>d!J=Sor69Cx`t1Xovg6Z4FcH_531k{dL1@p||C)X}
z_0eAhftZ#EB&te5=uv9xcRusY-ywpjm^KlZNR)!m<JNiaO_FS~*MuP4cQ(?@MzU`Y
zQe^yYt(GU<JueF&1h+7z`xq;wAoSGoO*RL_tE<JZ1}{yJQBW_XAasy$S5h^axmye?
zF+I<Co~=?4dTt=|ruIt84owKfM<YZWqNNmsUNj@T@~~cPy$B}YzKsROT1r9aUsQS&
z3k89PMKB4|HUb+>r66?7ZR<?0-U+!3F9u=00Pzcu@Y{owiE|%`W|V)PBZ6_bh*Tie
zP>L=($5A6}^OJl5_~G>#CmN@t6olU8cRb_$%ZqbGFbmh#6X>Zb1)=vOAGh!Bdi9J5
zreWGe;3829LMI765Al6va!j2!8`E|YJ7c9F^nr&V2_5%#p3{c;xbIY?nTn*}9;C?O
zi*;r%np^O_n1Nd`6PRf#1))>VHfhCp_AS$fdANoB$Z$Uz^zA{4oU9q;x<2;I2~C)b
ziyTI(hjAxgAEd}>errhapuZbMumpEvATZEU3PPXt()O#VACWEse@vSSOf{5((7C;D
z#iy@!_*tE|7}LFEy)2Z1(1m*gjDvI%4obiepN|Q`37Sel=vy_1<Br;NCaUuSaP5AQ
zep00%^qm7s+pJ<kfBt%W_8{>dq{V)HkRp$+P2c{zhu<|Ztim14Hk;i?DF|I%H1EsK
z$9_UFEXVX*<GHp<LFmV$OD;Q==3NlMOx(Y@z+6Kq2>p8H<KF^1vW>*B0@t2pHmi?P
z5c<RG$8mk1&o&pqGF)3%pxa$32;HH%x#d9e^?Y?+Fs7|!Ru)P@=r3*dXPpjLhN<&@
z!nCE#(n2W+&098q!SXm6kB1t4#tqB$cm;em5}8FJb8@NpY!cE-Li%KtRQik34i?YZ
zCT7?#sP7jvu<L@bk<d31D7+r}uBUMnI7fjSS3Vw`<7vDboOjbW0h|+PoCpqy;KbE$
zg27GT%Hd{kX@+4OZUwhi@ZfM8jA#Qd4!6UocJSeFCyeQY2?P^+q@wnn9b#B{8Io5b
zVhb>?Me-M2Sa7<{f1SqSKYzH8g3@cR8*EOisW5B|SjB*xbj+}u!D=&&w}90a4iig=
z2c|HCMDsIjh<=DE*)|4mn{5&rNVZK>cDC@t-R$j%VYeV=3+hD*sxLGg9DG809gjB)
zw6j2$41>>BfVcuQ2__a6seQAnM5k+J8HWr^@VYHfOhhLJdXXB6iRf&m@fOh8LgQG_
ziKX#&(AiGo9iX#=!^D8&5h$!TIm41~h8xtrnQJmIhW!^T{snt79{rO3VTGyg<#-J$
zL8}ss2toD|x1ccUq+y12F>tXX`j=Su?rPszb-e#_xf^MLVRK=4E{rBEGi*Ls=0hJ2
z7l35}jf=pth{o06S<M|v;xholP8Uy?5k;q_^c>>8*%h<vAIx0p)BeQsx(K3+U`%>5
z-x@pnoo&e_#AiQ4-_P7i*bcmQ!#(I|Fs_jUvK%laHHZ^}&z8ai;Z{O^0?99S9dd?E
z0mBq-XNYscSNqN>!E3V%$@fv!4uRnzupqrL>@)Cu2IC1PM;SMx@L;%oJcq2#p}u6J
z@idX?MPx-zqyk%CMEwYM7?Q7cpPi<Q4`dTCO`uKMXV_-YYz7I3TS2!K3_08e`fc3$
zF>E_X+rgaUJHfP*o3^v&U;8tYR)5A<xETqXk(evkf`lzJZbiaY4m%q+6)hG7HL3C0
zYlyjq#KeW-W2Gwog1Rmp*2W-?<SrqfT@T{*G>!ss6pgc>cNW-?f$-UK5SP<@IrJ)r
z-Zbx04nt_{Uk=j<CVr!;+IRjU!*(E12kO!likt4V8HST;Sjm9X78-8{r|mS}0Zuz;
zyc^7SgDqFT1V)r_*MMQ4g6C5hM|jt}lkWaky3JQbaj-o<!$?Bl^@HJGU<8LFRh%MK
zoC$V&Q}Omx>VZ{w$C5xR2@D9qu*sm6Oyj+vwU@)gm!{eoyJh`Ytq(u6&b%q{z(jnt
z;*cy3^(TtfBm1q-CB|WcfX}ugNjvJgb9}s*k`5YoB1tEWKO)IT)TKlC>?b7og!+=9
z^Vx@B_7EHhCeFN?+ILPo!%7XLQsSyItjs_rCELreCI%)_;^s4~se!4KxMU1#ZeT7Y
zJ`BFnf$|_u-O1<67g1=9Je)#Xk$ijC<sW(Fx3^0g797M&Uks|nph3o~z-GnR1QJi1
zEdkXMjwi#yI!O_Fa60KBXgmZHu38;<*1=?gy_2r(DBaMYjyJ+gW@b*j+_8gqEtS}9
z+lhx^AU7C6{8EOscCz*$dW>OhoNPQuWH9#TKfV=q@BZfV*?17gbNYhM?gsI08Yh4_
zfyR44yobZ&mSY3<HA*5Hjy&-c6|oUF#l;4Tort|>*lMt-<{X-ds-9Ov=4&cUdNxwc
zMjE8~N#nGwUdFk0Q)Ac%Nc91tR=NtQR?+woQhh|@YNT3C;~Jz|L*rVcT1(?Pq*_Pg
zdZb!U;|y@ffUeueu&2S{G>0cW3iQ60?Bn)BUujBELcpJYJaooW%|kkQNJ6Gc0!M5g
z{l)$UoPAWpki<UDyy&!KMlXw=c+1DRj2lIMX0Vg*xCad<E)s)77WczwEQw}j)<yeg
zE?ByYuQ2LNF6fh!g$Dupd9)e33Hmo_TnPGw9G;bycHxAd(?Z<;8)WhZ_2e3Qi%j0q
zxDlB&a+t)ucy`mo(~P*;DMDZ}4WeJL^P&iSi4A4gM$l~pJ<`|w?cFRRwU-aaOE_ly
znCaxik>D8DRz$|!$*Tv?dd?XR9`W1L0sm|i<J!kS?-*wigO@E|HU7i#VR(_^fgca*
zq(27V4gPK#CjdWz#(RLjhr>(!j#imA@pFH87M8@%vGS<(UX2$eTpHe+OrK#Rq!C?+
zTwUx*C?a}pTJnQPp4O$481@e^{Da#mhJ6m6&pC%Abj{G)(xN+`gnWFHWDii6E6Hc8
zkgSTvkC5yUjjNHYn#MIqRzu@jB&(%y9g@}2xE{&sIUJh#VC;M24f}t%I;+Nb6x;ec
zPQV*w=Vs?ivIA@KYU{)JWhM9srh#=DbQuqZ%>e5R8lML1(;SYNSLsszB6^wf5=_j6
zncV*7gIhlB=@!7~0+`70MKHRE!|Q*&6rEdBo23}oM7wEZU{NQ1LMHDoE5Pj^hyKUu
z@GTYkr_wkLe9~Y#*M2(mPv`he?>wbWx3?MM_9dDU2{D4CI99n|Q#a%&Z1q-&ezW6?
zT*j~#JuIZeyJuL-9+pxPH8X6l9=)U_`@pbPJ*=c8;AYqoBrKtoT`3Zl()b@F{D;P6
zNLWVWyGVGK#^p#@PUCw>c#p;pknRK26|^w;2<leR_z}{5MB{3tTTSB{q+3JdTBKV`
z<2s~UN8@^=TTkN*aLeFaErl|7J5A&3AiU0*ZH3m9<<fU64}^I%p9j`?(2wj8!`=kp
zO`5+6);DRs5QK#^UkKKP93T7gZ0JPy*E{}Sy+GU_9K<QobmTB#M<0pWX@52bCmLw~
zG!~p<X&evc@!$KaJxk~?t`<CNIe(StYOIbF=|7@;8D*M^kPN%nW$_5k+Gik*45UXA
zcqB`RT?9pnkQ=$Nfw7cmaqOM_EDt3fSVFDM^<g<j{>4|~3evxVEQr-iXq!DVN<&(K
zwLv1%Pec}+2~0$;iO7q?IbfRugUKF|SQuLiMIw?L#8J>a%5C+YLa*a@H%9cq9bZJc
z7m*2x%*d_B^%a>)uD<*lk{5KTWin@6>}$r7Td)SOY=Hh`JcrU3{#Bm4vJCIJq1@1g
zBnTMR*2&hD#IlDPY$oLCwf%)ha{#msaCZEUC1d#Bz8jx*=d;%lbN%}$%;5`iB;6>Y
zFo#H>@Op4qPva;U7R4P8MJmN7i`#xhD#fCLiyKoKdsx^1+p_~Jy(F+q;=(mW6k?f7
z<Go<Hm&TPasFHJ_h*OIfN|Cpv*7H~Srmp2YGUB0Z_+K4+|JQaLiSbW*wlYVG&p|9$
z#d0YohTRTU+iAQ5tafmi<b<(S{hAOa8<hIa>yN+<X;rwZIMgE!84>gR=bFpiCWNGn
z#DuO`S3;7SC%bu7HhBd*;SIDOV(&|GZVcORNIzeaw>>$~dBTnTvlDTlVCP^@&cn$>
z&Y8%Q94_Lw;X!@PJ(EFYSU;aUy~(!^juy+oupFe^zLrD(axO)w;JwSau#*G__|7RZ
z*4*;G0M8ekbA;VGnEVE2lV(pp8M-=2fBYh)Pc$r^vw;eqXxMHVC%~`-8YhBtB8(*s
zotbfeK*}}^ZJfyXf>ggC%DIXJ)kw~HVsKhObv=!vKsAcSS)iT;TBKnHPgJ#><4LU7
zP3^n1JC2-sn)M-vjA1*GsFPNb31FT;E5TCeS;{HF3yx=h4=-!*!r7-QNaqUYv|k#T
zu;$dT7x(dgyg?Rk5S5L1i!9#KxDi=2a`@7X6^;S>dbcW_5vN=j%FQPqjPkiO7Q+^R
zQ2~vMz^I7Bmt!>!L~8ELQD%sSG{GpYp=NMzhA|v&1<zLK%1|+E8;owFcdQ-8w!>7e
zd?)yJa$X8?9kAk8e6`?Cklb7DOZ2p2wDJNv)BOZIpK!-8_ukQgywaTDA3A?|9&hI{
zt+~J96hjk?ZUS#IEIv+rjcukCQY-kjf*;4X!Q?iW!r^uZXy?*Ee6|y2c0wTG^HL}L
ztUgj0{l7HI-_ZemH_scJiwU_<?nq`xA{Tr$6<=qF#Vaye^Hg<>xEr2W2KZ!f?enp3
z;d7e9L=#{=r}%tBEn^dsw{<DH;uj}#8%po`;NI;g+6NGAQra^-CTi$JjqVKA^SalO
zKCy#}YvuX{QJ)MV!8MZeDHYesbv=!vz$=Q|7zXS4Nm<}emW9E3zE1_`l9YMvxV81{
zEu$Z99I>}}eZGEpB0BD#@#St0Y)y;tjAB#_V^q3A$I1(%My73>xEvG5LGL(p5ml0F
zpTpct!<4qgHi^smDr}2wGL83w?OqyJf=eZL3LpOVtTLyr+gaShdnA017&1%pF$}BY
zsp6?r%&qt?Mz)rG62pZ8!~s<3s`xra3X!A|Ry&HXV%)ZVM52$hqWOeGpJ==X%=dh+
zct(_R+fp02qDDSEPV~cNf7Y*COLq;ut$2!=NGlT=5mWN=lJ|uL`eCE-=(9jS3oOap
z6+5h30cHe~Z+E!h6!8>lEfk;e$TVNapL)OF_Vz&y?C<GK=t|-<tgoIgh5WSq^vFeF
z*jeIPx}<f6oh_cNOV*uXgM>k1lKEh;i_|j}SreZL2Y9Gw8jaJDEFJaa_;$o^M`~mS
zic~JYgT|eR-$~<-i2sqspAi2OjUR&WA#|m_iDeh6efOCX>w_C0zX?PmV0;hcl{9`1
z@)tBdqar`6LY4nZMV?3FJ1X+KG`_}{U+0tZBuc}ISMg1hJ6iII@@b?H!xn+Oh&%EO
zTLYdo+(G&r^vOr6hQjdq=r^(7G@^w_bSO~!?prFJlDpiCYdatM=EERvoeRLefb$s@
zecKn&_%V1s=K988W|sslz~lD#Q}k4~aotozl*E{DlbrP>?xBb{JS4xxBL|M;dU(VG
zTA)8ad4Xr!CPXwJajAGj(JWs!cX=l6O@)2(rQk&9in|Nde?uElG}?f|!q@&9u`Uv=
zM^T8k9{(o@@)t&4c)SSibshJ*kt)2F`#GF@#VvOIMN{v;cKynWq<;R5df!03&kdQk
z1pl8Kp7(qo6@E8CccH3?QP<QEdue$L@1>>YrLCu<t8XB8a&a9p)NNSb9=6Us9ebPn
tU%yI>q%u=8a|=r=Ynwjy{rV3aG}zd|&~Ctp9+HuxHzDwx;ywzE`5&4Nk>3CS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_bias_softmax_gemm_permute.hpp.3291370BDDEB76AF.idx b/.cache/clangd/index/batched_gemm_bias_softmax_gemm_permute.hpp.3291370BDDEB76AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..902d5acac4b8e4d9944197f674a2f7aa140b33e9
GIT binary patch
literal 3200
zcmYk82~-qk634p+5C+<vab#xb;gsWn(86#t&@h~XfD9m@PeDcyFdpD?D4?JaNDL1V
zv&y+d<!KBcVKG27hej|8Syu5jOT?_;fwyKoA{%s@z&Bsddo$g?uCM-IRdrQQFO5*C
z3Q2+pNslPZ%E(XEB?w|3`_0SGU8ms@M39gm3Y+gPi`z6J7hZ^YaK<a^azj~kNms0`
z$@<=%ZM6r7>vEhcKQiC;b;!_>sW%GUrafUjX{ExwW{-*ket9!mu-3lt{k`RaBx&mR
z#Jzn~#~o%|nQJxD*k>YX?)${IE;!$4sXF1YZHeQL+vIKq@`)YaKfE?Fain1Mdbi@-
zm1H9m<MFP}=jGehKk@K%jJbSnb$@GVs_~Pr-m0H8ul*@hXFTEgfvY*SsmZMaH!nAq
zI*@0M7BJnrGs(x#1bsV7r<@A}E6W>Y>h93f5ouk#n-h*Fq)}6EPj7F0HFRNe%ZQ7`
z>+BJM;C->^T|&hH!C3tX3*&5`!49|3qVHZ^{@^2O+`mgOaP`uIrRBrFWk)|v{Ov}h
zncH>qob^uUbH+p;2UXrxD~&ev{CfS;i59)7q5nv3sxowBVTZ{%jkR%Bl>_R`TWhU{
zrYaYeFTF6~M&DT|s`$4@&6?Xc&YgUeSwE<V_-ZUZWm$#g<i)W)lgICu*@qP@P9R-A
z9DQ)|*xBE&ef-_twFid!-}XHFV128u{*fzB@{{tjwoFMz?@LDORdKx`dLVKOxA^9j
z{P@EO2cvdgz$s;tw9MY`9#0<ZkSK~>zb<r<#_CIpK4DYbv7VdOM-DGL`sTxb^`%=0
zhZ_cM0yjSW<%@(zIvcP4SSI@HklT~kq}m?wAU}D<F#p?xZ~X2RZIHx$cK)AtG-^q4
zX6uUU{y(=i?PoOBriSnKFUVQ(!v0OG!R1F!&o8>MZ_D6eht!VoeIiNn7dOIH3wP#j
zaFpJ99e!$o(mqHseEXg0zW4Ec!7m(s*STbP<|2_-l$G{??Ki2n)A6d4@2ib?hVu!k
z3=GRaG<zTnSUdZ1emSA}A->a%Bm|<FD^$E*d=A0a1Hx<`P_V>o<+rzWTrY}vB&)F(
zgnPkk_JB0jWf`wh#~lAJNeGdKgPubc!PnQB(Kx@;{?}l$SL1VdMj}3yt)Ff8=MVC~
z^}-v*qRr2e#9ZVbpcha?@aOPm{9j-Do6hcolOZJW7qno$-u%r3-#~X}0g;=LnO8TR
z(lalMPy2?B>5`OPh~09nBNZDGyIi{u)eOpJh)1cBR3|7qF&j=Eyj33M8E-;TR<xBj
z7Fj?q;E)Y%!y#MRmP4M5XSlX3lo#X0As@zvLkdR0p(r}a8;7}sUcw<Yt>#b+9mAp3
z%<6FMt)x<!R1U3S)^I4DN$1d3)z-L~+fEQ9wN15+LnW#b4sBO$=g=-t?1E6#UnErx
zifS&`fTD)WdqA-V^Spml$~JdCJYz^wGFs+>3v5c8a>$G}<B&OR&LImg3kCLbVO*5B
z+;T?FAy>wgLvD;4hul@}5y($cXgE;vdV7Qc+d8ye5El1keB;m}l8U8cgYX^7U^3!x
z-J8Lo8T@fuwSY|v_+ah?hfc1&8*I9<9aWQOI6I?gY)GmWY-@o=LW|P>@H6L?^z(G7
zMsRKfIWp){N5TClcp^rT@%VoZb#?#$4tCu3vm&Wp5cWbbTIQjdjH~#(uWJrTi4DX~
zICDo6M?ai-ph2J$&OFg1(GTg9)F$gqVYrXVKwJhsD4)p2A8Ia?Kc36Zq=nqV6U(~?
zx`$!;2zi7jmQM*x2}ANIPB5H(@SwsZs73Y@pE7!SBxTGu7GULt{DlIntmG>NSUG{8
zAVA8~m7QA)^U)5-hb`Fj6ATyZ5C{)JAzJoVpJzGoSMsGil4=6WCU8aulIj544sb_Y
zzFMX_>gv8lKvH3Jm>dV8qE#FUr^7k4m|o1G2(JiV?6;Cx8Hj^VVv;zN%p`MY6|;&%
zDXNr6?AHb^Z8%7hY6Eo}EJZ<+R6EGq!Gp`{cCNhxB0E5h>=lQ0Y>jR`JHSqJJ&5bU
z0vV>+vmOF4H-NZ-Yi|Ir22f&q6NsC*_BN2Ufep5|flnI*VQvRmJJ;S0igpOb_70GB
zVEfMJVP=ZqT{5;^>cO!d&&@PP)k7S1YyhVQaOHAT1J~XJVNDQ)>{a@wTR!Q}EoBGD
zQf|rVaF0L_PKQUzBRL(e3DjsioTQ2(i(>EqRe^sM1fv1^`+YBSHSzQ(w%z>o{l$1$
zE9kd^mix!6&yG9QRItlL`XcRRlBxs!I=l=G4?9iM|HP=rM|%4bC;l+WW^=Hntv&E8
zc{ARjXf8-9nvV9s?U=@-h2pj<Ru!vpaJxaV8-h>{$cLS*>Hl`LD@k>MunQ`*uNimc
zdN1__(yHl^78_gQFdXMPN^uxUvC<Mrl2qc{L@5rV7TBrQHVsl>H=cfHz#H!e;eII9
zZan#QeV?=6P&be>?<TW}O!+vyJXK_;!#=!x<9#(af{geK4c@Z_d|N<;`hldn!M7W8
zZpOycX}!GMHF+BL$fLHd*qpQBLDkRtrZQuRSZYGZX7^(c+B-)O5`5JdpGxtm2|hK&
Gr~e1tcx*}l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_gemm.hpp.3359348295C6E9CF.idx b/.cache/clangd/index/batched_gemm_gemm.hpp.3359348295C6E9CF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ef6c3e6aa5e27bc4e6ac1b520c289862ff7c6db5
GIT binary patch
literal 2086
zcmYk74^Yxq9Ke5&K;lnDejtK~OhgS7t;rLT){xYU%+xu8gIti^d1<2J&1z+L)w<2f
zn*Fh?Wj4M3EnRt<Teg)eD_5><?MhvxUMpL5Zb>b?&4KRmd+&~S-_P&+zTfwK-|zQ5
zax&7=L^Cl=w=$!$xWKW(2g9%+_pjVhR_LT)*fn1at2}fkZ}sYd1dcQ7`nKZJtgO=Z
z0&z-1Mj;=6KkZw;)Qjz@D=nFy)o%T2@}2p^8bQ+bW<O43slyR>cw(3ELe<OU##=pm
z6xzzR5dqKiZ|o+1Nl#Mi9%}PoH9nkt?d8dus)r{Z)F17~o#Q|d=Kp%IbK4Kr-K+Wz
zERL$ZpM2u*W7DRIa_*tHj<o%?6N@6&H~-{EcgjMlX-%RwUGI;Z9xZQdPCTQ`>HpH{
z3eC3-O>~WZP~8}O(`4V!J+Eq6=<#Q->}}oW!ZZUd4Yyst6<xYo-x69<vUcyz3OnOs
zlli-mI_0giYe%GdJ+FSqaJtTJ+#ShW#@-RI?qK|~lc#n{(|)BprR@tgoNsFAeduU3
z<u01MH|D7OJSI0kN}f3Qmr!%p&I)jKNfb{XejGN~6YwIJ@w#(!XoWrB)wSlaRvXZN
zZm*+hgrOXF&97U0<BRAVige4$x*6AnHJ!D3tDUy5^yZ!ZjC7yh>0w9qH+5jAv_U7f
z%B*5*c6;D~+3~HnZKIJ&TH%cvO+~!IKE5jRPRq}env&X>vqxGFUsp%V)g6OpUUS=i
zu}KFyj`_d&lc_&gzp*qmZ`Tq7CdevE%XurM=D6_h1OdDHNygB#7i}Hgh4&VA?$a&r
zdfxY-;f;}@-%~oq>l^#@0am`Y*m}g>1+m3{OpbmuywQ(J_s58IW;%On!rTqx8M?X8
zq_QH*avWpg(_e@PT7&`fok78Z{U`*bQgJn4PUiOlI+cOb2o*ymotmJ%Q~&bA_3Vpd
zIOYrO2^!(H_n}VPzq;GHp=8l8k4>HF-an6(C!Cs~Jyuq*s(j4dvh?utJ9S5s#7vyv
z0FK)0mk0*I2=M?9AqkMEy(b`|KommeSaUqw949PsmIQ<rTJ$7YXIm3VvSr$qLSnia
z?r+vBb2&I61VSwtR2T?DNCZR(g@bT}#6XOY9LTj~z_B0}p*RqSkOC+WiU;usX;oU?
zbg?iDCsI@?2<cQhgchh4Ae5>~MJP+21z>mJ#8UZEgtFz?2<6Cg5XzP3dMT*o6G6q9
zr`<DgA_RmWmEi+ELINN_C=`Su6b+)0s>Og9gl2(R2+4rVOE7RMbE=(Tm*RvW(hv*F
zH#emE-BAta8y|O1i8)HLd=jxlNtQt!p-v*xna)j@LGN%4)pzPv?6^G9aC#_Nj1vW{
z0xR6*y(R4EC%ZMY8SY(tIf2M7zLG%b!5x^tfq*73TncmQGC{vvj#aJ8JbtAI#|b0X
zD1xiqq%^)ml8I|VbS7Y$PwLEka|A+(=6NJp_!dNGQCbk6R*^NDTxnHXQ%JIjY>3V_
z*QSHIZDXlIb>DV@JHbYo5xHPSy)hdu!U>byG#eqk$*a3Ovg68#yk-jpC)f-&7gpg>
z0&ScT$gB3Gl*<W6Qp%NtBPrz@B+&2E72>bs5}&_K#|bXrBK;K!BE4OM6H*{W$h^>u
zz5|(7@6E>v+Z>w`I)?$c`@r+M!Fu6DJ}ckb2XL{wsGesI*4Wj)|B@oP_niBPp4Sr|
zv8<qIg-0$gvz6z21me^`y%&xR%v;0YaspU^L2NA0^Lcu3pnzd4QVbwP&maE(gAD}?
J^ClF-{sVrK?z8{^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_gemm_xdl_bf16.cpp.38FE56A993CA3FF4.idx b/.cache/clangd/index/batched_gemm_gemm_xdl_bf16.cpp.38FE56A993CA3FF4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..69c184d54b789584d78be307d7e17ee2b4e0cbd1
GIT binary patch
literal 3920
zcmZu!30M=?7EXp?gv_u6ArcC*BVk7%5ZOewu$8bVASe)mBM=tDB8Xta>Z2%vQh`$H
z26j<FD%NV%B399274?bjLtGG~ic}tb3a;-aeU}N}_ws%GId}eh?m2VrIe$JPg8~Bh
z8Z?@(D5x+qH7^;Z(P&KYmzyU^TMrgd>NHwm>)DvN>`|`foiG2mRo2mX=#MO^$K9Q0
zGn}jIMfHJFi$Ja9n&_abf!j)|SJ+$*ig%VO>1y$wxbEc#R`^vkIj^o#y<l^_vperi
zX1|HVbI!DRWXC1fTOyB7?dwz$6DP?}O{Kp*dONv2r)ZO3VVbb?mA1;}p<C0}|2#E2
z=N87&c=es=L{s9=e13!BjDPnjw&%4e!>rwdwvi*kw@smXms-a=Xf5s*O}jJJwy52$
zNb4y0c*WJQ2))anF&cDY)oSdlzeXEdz777#{jd9-Uq8ON{LC}k;V%OpH*+e(_t>p0
zPLQ&eo7dR-{Ls~N=kAO3<@e?ip4x~{jtHmQqEc|yo<pCtxd}MOdW=JcgPyBil@{Eb
zs2VuGb?#*&n$&ZXS-Hok>1D(BHU|gCq}{!?Dv!7~XIfSGaqVlJkYK#ENQ6JAi5ri&
z-V+*?ti?|YOMW||FJ>OxtGusj=H#Q1*rDSND{4D-T`qpKb?`KCb}~^qkmujT-4fGu
zwf<7p^@vC%PSej@{<?VnC%x_sYnB`HB)qKVYu^ZVvINFo^j+K?U8ZIBc<plYi0SQh
z##=@=o1e(rQJ)-)7)e%gQ~65u&D^FNRlx6cPp(;NWIl?kPwLqQN%xsu=rZVwS(DOQ
zrX9Pqf@P(5S#4<SlJMP?G3Dv1iio3MPrO<^GGdk99`n1{`8S7eee={NcRq<_h~nBg
zv5~ybd@g^bkElH#II(g0Na{>wMfHK`yN+s`l*Imbyh20xksrnMv#0dm#!1KTt_rxb
z=6B7mw9IRry`vZ6o8#7u_xiOvHod+#6mq*5uYJ<HxJ4yucG##!g<E>MDSVn;bGy_w
z!aI~5XVn+e|1@W}`Ka||NJYEI_}xDpOe{>=uMc|Z^|dr#HHqOD`&*NJvx(~kt?FM4
z&I%p$A6(Y9-mGKwGT!gx;G1)zuGH-+6Gg2y>~XqZgHOcmx(}KEzF#{Svqxkd-5omm
z3*G*g?ZR?FMxjrsL90ep9>?Wc>Xjnr(;M16D}_5d7^|#@_YO`}Jr5l_$b8bLWc8%@
z;q;57-#)FaK059;H8)N7o|=8Q^7F}5i?jWQ8k>f@bp~>7BrKVG*7e5dc*BSKa_K)T
z>UipvWg|x;23?MrSO010HT0{~<k-fl8C{j3=l!-h6JJU)zOkuWN;_<ow^dVjYW4bE
zqitgDV9Fk$_W7vS&C1@DA@P5>2H*IGc=32^=g`X=?OQ9)`@AomT~u}9VY*ltyiwhw
zrBO#G|7hBde>MK}EIj<$d$d#Q*6m{-_V4~7<Mqs*q<!oYQ(uS=m9Kwn+E#Z=K=%pk
zzL#g#>9FbI?e8>ZBqlbEdxKva_<L_{e4sORvGn(vB?8?u?#DZa-R^hH&Uhpr&--fp
zO35Rocb2a%gq3uSIgjewM+Wd-n0fUb{B`zq(*4ihjEQT#Ydm%?a&LM6No9Hbh69IY
z3>_<rE`|h6(--U6=deDo9~3Rs34J<|eP7CH$SukgS}D<JtcD;Dw8^UsN8)HS(gw59
zkOR;m`G=#YBHXX9^8!aA0d0v{YRUo9_OFW0oO{V{#*vQ!?SwgL$^p_Y_U~K%+BV*Z
zBPoE+$MW^%0BO(od&0RqtvVc84QLL=(Ub$E0~yj?V-Xhty$;aUm^DidkWLuv%Q|-~
z;2w^w2ec_>swxLaZ%im2I=E!VV;qS9v=wH>lmn#2C(L%%UO0RSN1_1jj5#yq0O{@5
zIX!&K#$JL}2<U1$T1{8^-xu6co1=Biwa4!+j>Lj3hL|Bk4v?<*37{w3elNt4FhDQE
zmTAZV(hZS^J6}$GpMoPvfHuO6ROJBaHtCo}ih5r+j-&$G1T#^U1Ef2rLgoV7DiAeV
zGN5@Zp0OMt-G%-qbHJtlPaF{f+8i@y$^p_nwJN47=Cp3$hzQUnSc$$IApJc<FnP1N
z<_(TS1KJ+5SC<2%&z|CNcqvz3;YcE&-7$AvIY9bcq-wf-aSQ`T!U1iE*{R6^()~wz
zold{I;)^3|0qu&pYRUo918r3i?>f}8aYO)U7tCd`93VX$w9jwvew+O`vIfvxjLVb*
zq@NzENGO>J7{ZYlKs#U#>T-be)KGp-$ze4&9Ek_CBj%_k2T0F88>zP&Pcp%g4S+Vl
z3>L`&(vs8-xv>~EF-nXDlR}WbsIM9<4nb}cY>R`1hEoJwWFrJmBuF3BN1bvh5EQJ1
zBLgQ3Wh|H|f($?d6kLh~Me{BMDWwaf^tCW5Jc`a(Fm-NH-4|W$2OrQ0(w*lX55vP@
z1Ubqkk;-uux=JBnsi1T|fY(PJ2ev&*%|Hk;-#tGZ#v{lQ_mXg!1)O4pv0!=$(jWEL
zhQY&J8?srpwYxx0=5FS`uz%r*17F#EQ#si)?X_W>CKk44C!PIq6!;o44RzsI!ubaU
zlnp=S>xcSjz_<iiC@74l690rQ_=Fx0Q->M>N+J^o>WDxzP~nIzf-Un$6omK<SjTb&
zAQ|2&UNoFjY00UI-UknVkUd_*TL&LQl>s;Ox5U5@@FkQNY6?LvV=ps+%0iHqY)b>!
zTLj5ra}3}u5~K~=#sG>t^Z>`nga}3<$d%|yg_g}1<i}C9*euwruqqLh%(pYtE>IVl
za8buuv90uB-3ii~ZLJS)AL<#n4VjoxE_O^iZ73N8>A-Z*hC@t{j!Z{wII~9z&nOEr
z4+MY;8mJqv;H<(*f)0_XAa!mC8Zv)Hu$?f*0vRMoCzg{j95jM-W;q+fD<?=7mWwe|
zBZB0zxW-U?2-211Y7B)8HVe2znUcZ`V0VJVWC{u&6Qn2VNx=4l-@sR<qj1wX=SY#V
z|AA-VT}{kQY+>zSnZS`US%td^(gXFt;ZZOxP-vOBQk#O&V1+R*78EP!Mmb$jPU!}G
zYY;~!z|;|Zl+T25Zi=&`{%w7P3v#TcTUFBoVK2iff##6uGvytG2F>3HL6*@4Wz?j>
zI|m_Us!ch0p<eUVdZu!u4q4G96A4puQwJyxutR{aOwuV|Z`6CfEU;g2#sW<z$apkf
z8NL%iCZGult99r)g_RH$Dy&k_6or)t6)CLJ&@_crI-0Jq%0M#|RuWV)zmlYi!H+rk
z|4)*VdtPQ_%HZ$MCTP%T63R@Y%nZuRqRec{ESNWy$Y0Jxp9}~a|4}2qATKv9Jwu$8
zowG@jxsjGWzlpYyGV>@?Oqn^9nNOLyl$nw#5|dyVoEoZxDr1XORB=L0Q$wB4^jl@Y
zWbn0?EY{X#yYbvTJiWYE+UZ!YFy&er{=cWJkMs?Uj7`kUmvO9Y>>V7PTwIqr>)9^%
S)nNq$ZASbPd;;kqwEqH(6C!B<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_gemm_xdl_fp16.cpp.F5472E9B2342F0A2.idx b/.cache/clangd/index/batched_gemm_gemm_xdl_fp16.cpp.F5472E9B2342F0A2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f323c850f6bebe534ebecd6457c44ce4fe9fafbf
GIT binary patch
literal 3920
zcmZu!2~<<Z8cr^kE971R(HKc(-xKzoutV4p2^3{-*)I?fA*`|qE+DcLT9meyA|P}D
zwYC*h#G*Y_6e}oCr0!A$r0r8_T}stA$$P^M=RD5gyZ_8L|1AH^_Z<ii^z#c)rO|{b
zfyLRW`H3ivMq`0rNq$~NJZOY7X|&=4rz4|tW*pS6e?EV$x~H{!eoK|>jmFbryPD>d
zX8$T9|J8|g5rG%|E6Zz~&CUnL+Epp2uNG*}woi<+bXmGaA+3qkR+kpPHoIN1{#b2X
z=$n9^qOPZx?YLJST+%-~wQgzp!Q2m7xfO*zgF{YwLu~6arHq}Y7Uo~h7RCKjcED(F
z%~G^Y%-E461B20~S9;O6>f6&^s1A(Iad<g&52FvUwrjW8o|5yZBrhcRTl%L~mnVM?
z^gqE1nz<3~bqx91Ya-RKwIf`zNx68{+a*D3TufTg>77f9wXU{@p6uzx$F(9mc5T>`
z6fCmH>bG|ono1~g-|vin+CR`XG&^~tTG8msL&nUAI&t=e)kWc@)w7;18TrCT?+=$P
z1yv{7Ut{h1NT=<2%fJT{Up(7BeUF3YuAeX2h<5jd_p8$nDgNww78y+1SMsaGJe^M<
zynbH&-Mu@^f0;|AU${TiKj?V9Ky6#i9rm;0&;yn+tq1qLHr2O{XZ1APS=*D>-F#C&
zC{suA+qS0mm+PX0*T;g44E#Tu**D$Cm`_*XcT3cZTY^N@JEMYo3~aRS&viOhM0y8L
z&2e`dsQWN4^bm<UHrFtPU&e!se9ja$c5nEqF=3=CzRQ3&Zn7us2J=blzvg;%6nj>M
z;M|58tl-2qZIyW?UPHPEH;MU=%hwJJw8YG}I;;<I7<+s#KDn?)un&peoWa-^(U>|;
zZxsrK0$cM>YKtbya?c9G_<e1=2kFQ4#M&nf(J%ip{gEhQzUlLnnCa?^dHs0QcgKGc
z&*a|kA2Tyu)akrBx8e04i-wO+EB>_Dq3pZUqHK#<of4<wWLxM{Zp&nad9z18H=2Jg
zQmLTuL&F)<zF@oHQSZOMPW0K}6ZCmk8GE<Dw$(Qxpw!ot6fV90p6EB`XA>`ztTgVO
zXWJI7wRs-vb9`cHIOS++I%|B<DcYu9q1IqW=NH;<V}CuP+;zlcbn_Pm<};V*mX~Ui
zszl;quL{irs=M=f_LoyfOYBajwY%*~Y3yNm^KR{*nA`pPhKF6O*>ei~*`;$c_eYL-
z*k~O;Ff(|5M!|K<(^0MVKj*p?a;6z+H<Axa{{H^_>B_br-)^l;=p0RMeBIFzDUQfj
z5s%(?(fV9G%iWjDe)Jo0(^~Dw&BoY-D;^K8M?`D6(Hu6Jyxf{FNS1mvmL8k-X$q@r
z|7YVo`DlsF&pvcFKzPaKhmlwB_5VJ0sBG66uQwHos_3ri%&erK^hcVPFV@vL-~I0R
z&DI|lEJIZ)t2ggS%$#~Fx_rytep^s-N5=6#KFjLf@#ccRo%ZdHYDLY_!rRS5$M$rE
zxRt5zUo#zE@@4nku*q|g%6IN<?CZT;OyBj*sWl518y2%`Jwn$Q1=nP?#~Hhen({0s
zCne^++}YwLn~PP(Pq`bNXBJKjuX}N6TD_8O-6+xYt@J!QVrUz7^;J{tS*Hh|b-&!M
z)bPS_ZzZwyfWP90&VR3O7~|E_BqiBNd<7bf(-P>4wz*g1NED4mnqg+DG5|g#`{BK#
zlyp#I?Gqdc2eb)hq9y}KTfQhcHT*o_V;tECXj{xyO$Ly*w|w)-pOueVaU>bg1z3TG
z3?S_mdpqe>{^}+i*#u}F##564q<tHnJ<UA)PZW;44`^*nTUiE>_GeUm@G$I0kR1nT
zQ_Pej14xJK`h68=S*D^wO9ZrlBhZxrq~j*eZ5cl7cN<6I0j-beE6V`V>2ak~UF#Zt
z#gQ;T^D#b429VDB&Y-dWZ0}Dv@-Cq5FgunEAYFTf_kDm#>j*(B26PP_t)VOZ?*oT4
zd->7hCf_9CNC+5Wj2WxS0Maety+@wU4J6}80-$v;9c39ny1nY5Q8M$~F&s$+v>v9X
zECWa%MgPhkv%m0f97zJSA!f*u0i?gLSJHQWwR#*!QUG0!m21cV(gO_9Z&yFATf&hD
zKwDy#Oc_A>^dOHXNWS<2N8SUp3+AFO14s{tD`#4kMlx_D6wnr!g^COyeW7o}_T=kP
zA&x`?+7WY9lL4g1+INS&?qTNOhzQX3nEe_VK>Ai-htK|lW(RR(GoT$X2bK&V{rK?C
zxbj86DIAFev=wH>lmVpwm@3#>-mBt-BQb!s#;jFj0O^+tcbY98CFtQu8lW{X%~djh
zbY5zjBnqPxqrh096oM3@LKRpXf~*u(M#Dlw6@e6~gy4(>>4kbRsgy!dVGNWER2Iru
zp(uj%L;d7ZN<<~gDFj(X7gf<?;Hq#cI%9?EoT}8H96i)^k4}&-0+(30JQO3?qg091
z9=%a-xdkgk70U~_uXpXNe2^+e2(rMXAQY}gkmWArp>P*a#fpp->Lo~HuCWGG0YRE@
zO*CMc2$IL;X}}E;q#4&t1NszHJ;GR_c7pUpecA9VA&xCM9P|1Qz!43d424h*&=Nqd
z)DNg52BX2MFc%sF7$kKCYLF$%k`1K+?E&15)ElTlKB&*~OaxghDvqV9vYjs4PLG9k
zfffSlE%gs-3V+mJZi;Q9ZOc;>2K$VeMmdPUu>{_N2zZ>T^1+=W_wJvNc5fEM!|s|V
zz8Zon_3wagZ8dEUya(t_;8pzBnBdCrF2F@e?Fe=gq$lbr_pS>>1<}-*ibO?nt#VQM
z@|%Ya2F@q7FKRkIm#+cqN06pmQw?~h(A7XyrItpeSg<VE&{qi3ie<%yN+(EbmNgp+
zyRZ0^qA0u54;)#OsmXzYhWC##R@fjx+H!1lq0k7@j$@|_Z-pT3Irh5HjR?|#<Dd)8
zhaep}j=Ipu;C+I6N}UvL8om_pjHC_<y9v?_btB+ggI|y<^-(xz*e*Op(YJE}97E4g
z&m5KoRsc*XwN*HpAYD;c98Lwd1!^cYS87lY8YCy<Qc<a#Z|tCpc2K?nPYu?Q8Zb3Q
z02;u8>s-ysdH0!Vp99!q4V_;@_lNHaRtem@)Ss#BKr~S9U02gZ)sz_Eb%G_Ou1zJm
zqwdSqT2Sh1LY$9ERjqHRZv}-4FCXMeEuG5sKs}br0-plTSmEge8H>g$!ZQ(M92zIr
zibvz+T1jY<Tq_w(mTRS;DRQk0G()bHiDt^R#Hd)Vm51gnxAIc6z+Woh|G&Is$r<of
zAm=~o5C4kEO;ar>%$H<jinF%lY|YKfPN!wVL(*uuRG3GF8B{2y!Yx#oLxqJ@m`R1{
zRG3eNSyZ@{3Ja)ELWRlMDOn_F2BmILKozl7O3FB)qNd6WVEK3(u^0iX*R5fzbDab(
zu5Rugo)&9Oo%J0|wEtf(XT64|j;@}8p)rqdW@%+@YwxIK_l~)Za4p9#un6&u_420&
G)BXcIMF~m(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_gemm_xdl_fp32.cpp.958D740430E8AB79.idx b/.cache/clangd/index/batched_gemm_gemm_xdl_fp32.cpp.958D740430E8AB79.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e5ab2c6614a0b60135d5c57c4540bbafbecebae3
GIT binary patch
literal 3840
zcmZu!2~<<Z8cr^k3*=q`K@0)`41|4)0b+o#ggq=8BXt3+vfZ*HAqbQmD_R6p78SRO
zRDE@!%G2UXi}Z;ih+=&x6^l|W)z$^Ih(7d9(l;cW^KuSf{+VxP{#pL-%Rwp*33*GW
zPy&<01)0h6c$7k+Fu-4~T$UOK8jAo|u;=3PpR>mJ`nLyOr;2Wc56V0I%`<xLd+gf0
zEoob5g-xhIe08+AcSmW-#&av`8oo<Gw747@-81*G%^-u#VB17B#aD8#y*lbNQfk`1
zDQ@wHutWK+FZ<n%2OeIt>KZb7b@yRPS4K|hI?kC30)q?D3%)O+Ri1zP``h30S^p{C
zW3zeV`<1+yzHbu@Iajw#S+2ZSvyYtA>$p13w9cmb*)(7~Z{0KW&+S2TWl<6Tr0(YS
z|8!j(de%OC?2c5_jvN);Otx;?FU|c#r=aJ<<nw4*YV*h&lV<}~6Mc=BY};J?i&zy*
zo^`de=q&g2zwf;Gbp6Ow+!LqB@9rl|>|2_MGx=g?pPc<tR(jxeLBsHzIIUrEtLcJ|
z6OVpgZm{Y0?mEubXxFJ77dF&{G-&PCT_n=)*tYMtApc&6+ea9Zn9)8fr*cQKxu>K}
zGirlq{y~ooFN^X{1lN7aEcf2iF?aIvIeNhJ8@mI{2fln2oX-2Hy)2k_Yi|7T_^Nn^
zP?L*32foIn+|5(t8_Fs?lKPJ&1;~Ot7d7`E*;=<<!rt5cLql|>0q4gUo^{R`qt4~?
zv3%=Ovf6#|VF+6`pP#I)v8|au??%Wn9{b(NTu!ny55ehwn0SaQzUB;`FdKC|dULUe
z7O5x2jcdoSyt5~FmdOf57uYSIq}vad%<br?kNLfcZyLq#dosEz@O(EbqQdE@yJmz#
z4{yA@mwi)mJ#w?n!&UC9kG+sI(i{&lvdWKN8Qz$sBTAd5E$D7u@8@`f6>;TCOvF=!
zdHndxT|)YWrsm{i^XO=kz2E#&(i!B_@k4>RQ%m)Y4FkHZo7Oxyym`v-E4KL0A%n?{
zE>;oAZ)q)c3s>?wFP#gD31<$j*0oBol)huVY321bf8%r@VAwIn({%XbL}fz$c+F*_
z^$#e1hjyMly7lBQmhH9oC#}~z2Zs0$j3r&=(&yI`)dx!dlil?3P)%n&C$^>jl+e%g
zMSG1MQ~Pw4(XT^YQ)|L@NN>e$-xJ!J{`l_Fu*WG!<~Lp{F|N(Jb3NMEB*#~{Ro1y$
zXl_otw;qc>pUavoC@$}D_PyKPV7^yA>Kxle6!qmT`u?volX?wrEuQH`x)=7Hv@m>4
z)>gO0J9`%QJj(r3*v+dW?^KJ{j;#9`c|UF6aj17<oO+@q;N(Wl?&<AD56>9wj(R=s
z;P}6uH<prZi{l5Ey{k1m6^j+5bPlBz-IsP=`elBZ-?T>XgE$ZV0&Dlr%=&w)tG$L!
zJ^nL%^v&|aReoJZe&n7WEBi;%vt93m*2XC&$BS%Df8W1)f^jXh<M0pr=iF`cu3x@K
zAbawl`^=tA5swbmmxkV?ueL5I@;<qjKc;DT`Prfj$A#BB*dK%uJDtqh3uU|C`g`|Y
z<SzXx)w5sF<K4J~F8p%fMWb7%@2$xvrJkl^TP!cXj10YPabxY;uE<9D2dg98x&jM(
z-~9n&BmED*LWI4om147LUEgo*ZsitcCg2(r3bS4;M0fgD;>ZdLg>=Lm^^^cyNcpih
z(N1V#&3%C*Qb5~cw)#qdwCiNy`Od$>nsH<ipm`WiUkQ-*aQ(3Bb=l)497zDQ9Fwz@
z0BN7tk%TF^!4@3(1kiSvoxTzv9ZIX%`e@0IfQ|#S1LnX~0;J<^c4u|Ag^b|HDnMIe
zmO4s+bV^*&P^(exuQ;*<(Ds-;LkW=1IK`=}xpeq4jw}VV8|KDP0;D$$*q;uwYw98>
z`GBsXqE%Gw|9!zBTZPxO?)k5HtxH)0h6tDfwh|z{O%y_n`{_dhjzj_42D8yq0;KDu
zhrj)6ydx1uRs-4sv(Qljr1w=kvPq<Ox8q0>pxGE(M+uNVfxgM?^SJUHN8$m^!8i;h
zK>Bozww2eE!3`Wq1avW0%u)iRJ7|(W2Aiwj<4821T`(8A5+Hr?jJ>^QV(%o5d<tkU
z%xj(!Al)g|Npme)PQ#JKfOf{5b(H|=E62Lr&P`tp#F3SN=3{()B|y4wU)7T7L-cGM
zkpS8qbDyIGNZ%9h58Bb<)Pf_+0PTT!Fq8o4Cv9KEmAncW!ja{GcEw!jN`UmUp}e&v
zhjj%w5(8*1#?@5<q~AWhzs=?G8cQ5W2DBMwrlkZ(Q)J0xt{tY*ga&Pv76>u`4bX+R
zLy)DC(v|S0p-#XQg-GC%1nH0Z(^XUQBzZAV6i_=TZI)&TG8he3n^GVtn3+P5<y1*I
zH3nvdOHpaFbS0>m_u|CC)=?@!dU<-q!t_vXV2=u|srD#BMQRC4BqcKmd`*Q1${r@C
zBLpe;mM@0+2(s9_crn}s)G0!nrB8wkLW2xp@~DON*-Yn}tzb_a0Vfa^FcdMEt5C3N
zPc94>Ls+L!%HXCHI#vw|L<99;0)ot!<j1OZSSBfpg*${|2Pao3U9~_68ltv9p`>tT
zfprl<eU2;m60l!EfFK%*y`tpI7hR(d&MLZ>30A>wXcyoi{`Lx(0uBJ@Q#b_dCdhC!
zd?qur3QgK9iy%lp)K9J6<P!Nx)%n&*)~U6MCB-wf2(3knHp@~7(#F_^1#=Okt+6c&
zmMuZr8QZa7fe@s<u{{gkIzc)ZJFws*6QrZDBMbg&K$`+Dqwpxz@thgXhR`So(v{(A
z2&GGqTn5(=D)m^wc}+>?fe>(?W^^+qR46PDj5f>H2$IL-v7vAX(v9iHhHp!d?o4+!
zv=@T(V0y5jn-C<Q$!9}nf>i+?OJR+0)3DmWQz<MFb`zu!6%w$@;5V47Fh)42(_NaR
z8GPU=I0oCA?F_F4-T+uq;gfJQL3*R!I9v*D3*4c?EmecWs8~&wpG!Vh^MOjLq*CPr
z@YEoW!ckRAgrcDgm}f8}d+CJZF+SL171g1N8VZXR-V#uU!f;iy!_csq8YRdIs-%KC
zBPsZtAf>{FRg-*BpPAcws(ow=;&nnHdP{3dE)+X_d@xtx(W<$=sPD|Rz*4|zvrL*G
zW6&5)cqW33MPt=kacG=cD;|wkYbBrwYOO>xQLUANrl_@2(Nwio8k(lo%0Y8xTC(H}
z@be4)Uy~)~ey<2*3i$hz0RswUtt!k>g(<2qO%-OU!aP-&r3zD3;b*EaR~4qK!t9w)
zgVd0_eiY~>)auFea%Cy0Y3Z3+*=ut$KBLHI22m0-lRyWB5}q8Xfoft}+B!I)tFK27
zV*~}*Flb>0Msp1384EnUg+9K1{w{MJy{!0l7XPo8Y05I2&$i@P+uA!gxpEh{c`P({
VH*w|#&Si#(*CD~NqEKoC<-b<2_SOIZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_gemm_xdl_int8.cpp.EF48A52F4EBD9A62.idx b/.cache/clangd/index/batched_gemm_gemm_xdl_int8.cpp.EF48A52F4EBD9A62.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ec7787e93f504ac8f6ac40599580be52a750d036
GIT binary patch
literal 3798
zcmZu!2~<<p7EPXy2jo2hQHU~TfB=y}2pB<T0tsP|7*QFO=>f_RNI)nQ8KmD?REia+
zDzs4Ruu$7t6hXy;Se(!nKh_ZitN!Bnf!3in>Axhb^|KZ`=f3^!J@=k__FhCT6^pOX
zDHKt<v?P}-OhG9W3IqJ93Y9D3z#trOC40^+y`48CH2LM~n-xJf!uvfF*v3Dkjapp(
zDeU;grocMe^=_fx+C7fw><}e!d!)tJQF?ONBP-=_U({E&!Q9};U9-E;lpP~q#GQIS
zKb_IG>s@`MLNfGhg<#)LBl^`FXTJXT!{l>01*J;1SAdU6dyIQ$IZ=K4(C_0t^Q_*L
z?6lwb*XcMxeBYcjQ@h@=32t0stxaa}*u35+7*9dH=TL|3yz!f9psEK|DWf7zP<Ojs
zzIjC|Y2$|v-I50#M-ByDCwcYz<*LPoB|RU;Mq`vK8t%Nc_+x?Xs|%l>cR1=ASY}nl
z6!12<1hu(+|Ld(6%huePNPOxX`NQ4RSNr5?I8!Kf@hjLbx5^gn@Twb}k*Gf?ZL)O#
z`0DX9J;Mu6qx7fCtCI&K?gopSde2<OTG;Duhh0^1*<}67Ukn^w@`oABvt}pyhWrPr
zo6obi_U}4Bo>|lyx@8%&(r3>}v$5_|^x)BJyMt}6tbZMvEx38SB2;k0EM@RXQp#+J
z#TmcGuQ0jW#)&7NeXN|De)({EurkDZLBr*5w{F=X<LvD^TNhJ3-R?rXfLAa?-Qrq0
zRLuKMxp`ko7{WR3AS4Z@ZEFzDx+Y#C;QV<^Wk<@35S;$pVy?96E4%(y>-$cJu15#a
zB8}y^`Q{<4sO{L!3gv2XJLkaSZ2n-m*~ydJ<A1LgT1E-`p59Lqo$j)VsB%8!sTVQ3
zNARTb66d<?O5{fShe@8x55JIoPIGEz<W(NIIJh>?Flgm_ntNBnnog%{RuLC3#z#Cu
zm`9Gh+~rGeuWukpo0u4jy$A1=cZB$zJX>Ppd|<=1&!UZ*)+Ik^**IbPB}e+7f$3vw
zU2P-CaoT|`^Wp>@=TC*ihco+^(`{3^@;|NKHVOI~4mvjm4@Sg$TMmAFw<)#w$)@hi
zH4i8O%{$wE-g+#6dwlcVG2R*%kvK4VD81K>&fZRJXk7hHQ6I2Kyj;5dh|IA0QR>;B
z1&y)nn-<Hq-qUYc!$}H{QpG84HXq%YS<7i?I#?~E6_cK)GJF<IWL=oZkG?eV`r*rb
z#jQE!#<%wl{n#_)=bQJk*TeVs>r(x@t!90#J&7Z4vTMsex~=cbuSjm}6>WLHfB(Gf
zm_ohm-UpVhUuKV(?<g>}Uz*$E(mKFWWq<2j`v-M5-?TK=?bBr+g003<Ppr3TH1!*=
z>%ZFf+C6eGyp!=^)z1{;zOne+tMQkXj$1w&ZRu=r+;gzozPV<2Pg7jn!i6IZzqThC
zwtskd^Mylf{jBCy?^uibk{0dmv6~iES2UV#YyBp|KXu~%u<5Kq?(HJUs|_<G3;fzy
zU)UBLY>(+5bsT(f2vy`sez=j*UU~l3oo%^~yC;MfzRWbN`&uNd&6~l>lC;gtc>3D8
z#eAfsLU5_d;icie{O*B~i4&*pE$AL^T36tEb>Ulm&*w?5GkAMHuA9Bd_HO6~<8Q8&
zC{B#8rPRLsCRr}|h?rFJ!zijxa#Nop6pGn9#S)toYLgLVIfX(BnL>^RfCp(l%<WQN
zbf<qcj>H0*kMWH)0BOF3LFxf3vllob2Q&}knP>pgbH+Y7-SOYB1{_%cXgAEwL<5kX
zJLki$Hx<M6IFbTrC8o5}0HpmA?xao>POrt0#elZQ>`gQPX$h@r>*IwN0G$YEN6e9_
z0Z1oa@5<{qD!zjwNq}Z!Y(ot|IxDekplRmjM>w(&&<>abLj#b``Oa?3rt>Y`II;-P
z?wC761CU;KmH&O1WBoaTvI@{ORJ4X_@V_tk%C;aeHSy+$R2+!{uh?OB#u|Y1zN*Le
zY4on+IFbfv7REBv0Hj;dx4C_DFOK3!GN8E_m!Sbjf4|AVcK*clYdDe$=u)iIN&}ES
zNs~SAZ`kk`j>G`k8FQv<0MciE<nz7LE{);HGC+G_Ub8d+=?=N!$~k3AX*d!MXeZ3c
zNCS|*c=(*hsrS7i9Ek(8C+2CQ0Z8}l``5zv&2$Bh$N(+C1T!=M>EERLL%ulRd;mw5
z0NMldU}yl+Pmk6lmcJGc;K))yyI?MK4M6&jfuj8K79%eli3hYR=4zw?NRL0eyUlet
znTsRofVRf0^)&!#C7GdOVHC<VWtoq6*KhzrkPhY!Y}g6}>1giAhHV9ffzYN1f*`}t
zFk@IjD358hDY78QKs3+@zC@6p%07*QR|X{mrl|D-_ePNZs6SmhMI}?kLw!IQp|mOT
zAxIG_(wVYKwrX+;L9Ujqj)ytnLR8umfq7NUdeQoI(|syIdU<;#z~oSOV25haX?GZa
z2I!<Kla);-@VE4BtawOfgNiDBl+iFBL6-WIM#Eje%7Tij)v8@L7!5Xs$)o0NS1?^R
zZ3TPcdT~Wi2T;vmu3F04xy}q{Q`qigRM45KW6h~3LE4$yv7iD8l4s6i!MYQqy}3OL
zo{=E==6n{sM%W3U&}v7}W(h)rChwdei)2L!+9Q_B$`jxbp>2TDsXarxKnNP5vp}(|
zcyfVN5g~m}u|gR*urOE{1KX~u{KuMe_aC&WA1@Im!N<^ez)k++Ixq!1tR8KO%@Cv?
z>Zj9*N|`cFyY)g@q0XpORyx_G(4Igg)ZPTM5Tp~s$rKs{LAo$pOrc~6(v{(A3Ke>|
z<g}hFw^0oCZ%wymLS@4H#As9ej3C{ZZXDR>1nJIn=fGMLq<|^lKwBY552gnPx(Gpf
zGCeuak>D+W;;1bVZW^`+=pD5c!p8(TADvIYzJuRjuG$dcOC6r_biL5VXW$qt9?J!m
z1y%qosrE<sHbHu$-Z)$eZVS{<?ULG;B&bA3M<ucn9obe=WtG%PUBOd>JZdM^E)j}`
zGGLzm9L1tnPKSkHk2TcUHB<>~5?CeBjcTLS&K9F$oo=j<RZKQ8yiAZ#ZNAzmKB&)R
zt)3Yiu0`gzsx`{yvE86H;njn=YG>BY^+kOrivk+~r%kbDf{aCD_27vJG7gQ?8O5XV
zI-}+2a-C5!nyfQQK~r=_8EA&iC=<=p8D*haIwJ+Dm>f}*WDfX;1pEq9rl~sBLADzH
z`6ysOnMR84+%8Mt2e}%9pKAq~Su0lN<mM^zvok39lP^&+wc!eFn4=AIwPB$)%+iJ#
z+EA$tS8BsNZK%?Q1==tzH$6wak8m<_8mfos8yMn*k%=)qj1dxS&!B}(pE<*HmbsU=
zkFTG9K%lFc(|lW@BkTV?Wm;NUvpHNlo&$fj^BgyKkGb<~JS|)VA~U8~T7-lq1WBk7
Gl)nL(SLrVR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.20C4D0C379DDEEBB.idx b/.cache/clangd/index/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.20C4D0C379DDEEBB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0701564354c24c0a56e68ed0d6d155c43a112927
GIT binary patch
literal 4830
zcmZu#2Ut@{7tIS0AIXc<M56%=vh<Gh5}Figp-8nOD4IMpbWju#ML@+MAjN`!$|82w
zh7D2C6%nv3$|BZX6%iCgcSZk6bn@c=ZN86_JMYY$cjwNXbAb?FFRy#5B$CHs-*iD-
z@<MqMiKGI4l9Gi94PbCsnM6vj?udAZKebW5+c}!x_IqG=@<~@szFt2)wPdm{Gedpp
zvc)PdX8zL9HgtcCCpz4_SI}LjfNiSwxK(|+E37rLb%zNxF5}iIwdcJpk=d`_uiv@#
zLj0urm%C0D@@KHBHBpzL4!xdSg_4&eT{#1z&s>ije)8@g7wVx-Gp|S7*x~UZ?W+EM
z|6`9w-|<OqWvfqD$M+;&{`Y#vo7pxk>z;XEEFU?gVA|8YY?fbV($MqR=w3Y`%XUsv
z%V@)qGskvys>jiU8^RnK>u-c6tp7DB^U~Myw_(B;r(y)weQdqK9HTtF@a2}sbj#Dm
zJDymDxCHJzZhBtpdXu|Bhf`qLB~FG?UP|S)fbRB-P7UPP*j~&kfqdD)a$a;#!Ih0l
zyIqd?9PCY=wvh3>Pu{?xeaeo$UcN<r^=Cbn=XIfBdXs~-{p>CBN^6r;3mi3%bf)cT
zQjrh0ZxsA-p!OF#N`s=~suj)Xz0cP2ZwEG)ZA^NRJLG>gdSa)~t$>bGN8{@SZ=xF0
z3iT866NZYnHVvl<%bbOJ8Kw%n9D}s?4*P1S*VT`qPDVNn)mI|@4(Zrv4y>^F<H)ZE
zN~Wf7DE4&cd-EG7Q*{5js=cS_;FFB=8d;rG)4S0F?Il_3bt`hW%hOL%+}~hdZdQ0@
zx|+l%n=cKXAHDI*@g6m%abeQ~pM_&lec9XZJTNX=Z8^BfQSIFFgu@n{v_;3y=(itz
z@ZP>D=RxjL<0bezn<uZA4mF{5MtKDZc|mQH1})m~O>VEg6j)ro*^_9^%27RbtgHL<
z+Px|C{^D!X?CjH83~LJ>T&#F;t|OxVrpC%(FF&5j+2NYF?dIy)7Ux<nTIn`8njGb(
zvZ@S>m}~E~^cEE=yiz@}{f=ez2eo*$@I2LjYB2lb&2p7~X=V5NuYBY>`*vx>@0u8x
zIj%V|bn%dJ{9yTdlalhCi%&dT;Q4YS*yd@z-1FY*6SrChB3}tqX3V$QyM{mTwlQ#Q
z#a(?bn}KC>KJhccxAmm#;a4;sJV4p+96Z~%fa<rCQ90h{IP{^XVn%|25l8gcUg?WV
z_~l{u;Dhy5hAe8AmhaHX$?{pl1IqnH-s3wTr*<wq!ch9{$%vOjzskH3g@D_2*&6!?
zo84R+zD&&->|MMnVy`trkiWa9R(+HeW4c+1{pa?!3;!wx=IOUI`shUTnce=o<I|aY
ze}$|W>YsmhZ1&jsp?=%z>Pzmfe#v(>D~V}cs65qr*<j+};HC+kSJIleQwK~>U)x=l
zvhCW6B~Gv222vbLR)$Pdy0Gu%x0Z{KcGWj_%x{Vcef@aKyAZEG>ADQsm(1tC&wUuI
zUYQ)v9Ibrl;c)oXjKmu@g1K({s*+lA9$os?^Um23H{Lhnp4O9sv5C(d2FKUAzqx)g
zdr44tPwMsM##()a({r2Ko0K*Uvh<L^m~>n2H|z34PW|~_8g-(DrLA+1whw0vUrjwj
zxn0*3bM9zU_dmC1j*T1F#QXa{t9sZbm%9CtRZm~Wjbu}K)0g2HRQ`(fF~#(b`jxdq
z`)gIQ4a{@=riWc{zkF+w|Gqk<!@+rWPg)!2*R&2iz4*9ME#jPKQi6y}8hbiPk(87n
zSfmUTiz)JTlHcL7jz*3kk#Hl5k&*;}rb#{=s*`YkPW(zbNdRbfraMOhkT#ftE~;TO
zhH2zbK<iQTlqCRZv$2e`=Rf*X(8z&+wx(DsO90Z=W?w5u*SsjDk>dfKk5KXv_Kz<(
z#i6n*&@!rOKb^D;OyM%Q90@?$Ip)trpOaO$(#W#`txwTcmVlJLTAa5px_bkiG!M{@
zOh-)#KziopcO%OVeTk%z=Kz{R;V4P~(%#gZ;^)C_VDV@`8&V9J5`c7odseJmmN}0`
zUJB@K2xS|h`QuAE#iRJ&n@3Kh(8xhxiZR8QDFH}F&P&=NLN2VOlNJNoO2bM=0+3!h
zC+mLgq|MK0<Pbm`Pz+Qg0BQaSoh?PJ`>)c-VSu)w*r-SV(kp9^ehq8nmC;G_0qw+e
z;z$6}S$YRNcjUHp(MbycZKh$SBLPU~cN#SN=$Br?NI3}LcO&w<5%k9woU_f%3z>80
z>mnLC08G)P=qgD7(xQ<4hd=&(GL}Z33utYMwxR?eT_yC7qds`^8=bTe(3TpOIud|%
z9mkTXuS2V+k%IwkLNQ@V0MbX~KL~DGx4)&4;{dHg(NU2Aq??P-=?<S&Z_~(209}eu
zN<n3P|B_Dm;2p@G-t~Y^iUm_FG%R!^0O@n53=HgJZ;a8%F@W}@cutl8q}#?We=2#s
zwt!CJ1KN%0rY!+Tw;#M@b!Pm!2aOyFXnTsivIHP~bJvF8@qH5$X=EOtZ7H^D5`gp{
z%|RgwxaE5~X*r->nXcLrfOMbl9?$X`;~E+{9ME<YI~55)dZ<Q|`|!8Dljx*HfHv1K
z*O36EKfUhXX8K}*HjSJBXf2A?1PMS|7#E*Zg@Vyr8_jK(em{y}xU;=;0IVivI`!J>
zkJ-NB$J)-e(#JnkM2`BCdSG}U8mIzOiQ(l4uN;YjhXb{RO#MkyFnkU=M-4hbg#jyz
z<pqX`;r^(<Dx8Gjn-Jb6BpM0|)E#;1PkMvlp=hX#QxU=|BAhV10^wC4(J(x?7()F&
zO|q?ZLF0ozZv{E9b+F~a!-P@=wj<UlBGM={O1d2=SYSJ1wIU`3paC+Gl_9(`A`<5x
zG6q%_YZ&1aiAE~J9Jow%+PmgS+;TDo!gb6?f?)g@D166KBna*cs-6P%C$+<HU6w8z
z-f0ZiW9hMB@-bYWrO$@3W4IB^hz%PDR7nt=SR#qwLeP-OFt`9)Q6kf%s2HS7$4<us
z-d-r7AjRMH6M6^y5KgeDST*5e3=cwsq>F|!g6xieD|`eN6(Q3^h&fyi>MHOSiz?wg
z3!S9|LxSo$f%=oGVt6nbEIlp^FF<$&NDPt7LWEZ+I~E}pBE$*4SsaP42ni#`I}qLu
zBnIZ5@3Zv|ovZH;LY}ETGX%y0r5sd+SjUNg!_Y99ST-ZP&BVLm-2=N6t2^P8hoE_g
z)GgJ|^QKXx4G+B1k?R-?YYFN<@D__d;q8TbNprMucqZGvvuFzN<tTHQuySB?0LS*d
zF+lIa2^VYfTZL;u4H|10Yr|8-a1#v^ZK&vlv*gyl+4{{+0mIE0W^5P;hMP0Y*-*SN
z+=5}jZvLUs7;ee1WW$bu;Z_VQHaz6bitVxU6YtyrG44hTcOxb+I#|k}&c(e6#)#p5
zsGp1`<s-a&B#Jm-5#l03yom!WMR=uzX2GTgPD<R-h(!%p25eYT7;eZiWW$RA!v>ia
zcR9kv7xk598YTx!68Ab{5)b7`3;tmGSp}Y8uNT;=)<i8PRPrVTpZg;|3HI_B&h_9%
zKpnzxR}a?+*h?_n&BHAMwk-^I_i&GZ0t~Ms$dtH~5-+!AT63V5Vz>>{h65`V!|j-M
z9N394+@5LAflUt<J1CF;Zn<ES;(klK#|QP1W&+*`;4SXDa1x9h_=>wNe2n2SXpB1a
zx!Tuvepz|MC2_5XI?jW&4g&*+CT_)qZv+}4Bf3=xuZj>|DC@vU+@J}k+30Krd?AK!
zL3mq`xv+d-wn3EQwoN!iqtVhRVb;LP;^s|E3P;1G*A~Mo5nd&s9Pna*mBlTcaPmgI
zrMZI@L8JcE(=nWn@)e1g1*kw~grm63C<#rH86~61GNTkUMP`(VX3C7R&@7qJDs+|1
zC>KF<5$X8r!f}5#70DErYZP~ZOu0t6E`SLH<r?L-03JFh*C_V|uv|&PI6nBD5&YLh
z7@KroY@~?c`zI3`BvKY(CK2We!W0r_8eygr<{GIfhmWoo<qE&*(^Rt8tX{QZWs)#C
zEj=SMD<w6Nlq#J@T1l9z2{V~66A3eeFtZ7B6=B8-7W2X17Nl8m{&MmPlnJOJ4Wlb7
zP4rRmbk|j(`lwD)Q=iPTvv+iIcHz33GK?Ii+vsck*Hh+Hw#GD$w$3kl28PCF<`!1g
Ywwg1hm{@u+m|nhX5%SC^H*du6fA}s)Pyhe`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_multi_d.hpp.4E3016E76D827F13.idx b/.cache/clangd/index/batched_gemm_multi_d.hpp.4E3016E76D827F13.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e073f3b8fd794c64cf7adfdbd7bc808a770a93bc
GIT binary patch
literal 398
zcmWIYbaQKCWMFVk@vO*AElFfyU|<jg;^LB`%!fdFA~OR+#hks94rVbM^0<6Yukvvz
zeq(5#v}M~SvyHmSndWcqA8%~YeY>!IDP!Y-<Y&CiZQiH8urf^Kb`0TJk=c06T&P|2
z^D57S#hGg?(u%?-c}WI+t>k0*ck8)CwDdP=$<WloJ02H=!e`v#X`J@t<<2`D&y-os
z12xy3xjBhDe}Vnh9_{nr=G&YKS#c|ScXZILeQy+>n{b`_Wqa`T%XFdXtE8^2{d&J^
z_X6{2U*C4hss62Gxx=G!=*i!MGd53>wewdv!?yVp+aHU$za6gciF)xtmao<8eRBIj
zr+&9~Yu&~$VTR($+$5k!8H!TViXn7PB8V<7N-oX@F*ra31ONjA2G|-C92Y3=?C#@W
j=j32yW@2Gw00turFu@oM46FzyC~%>G6UJa*U`H?kp*MjN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_reduce_fp16.cpp.795EB59AA922209F.idx b/.cache/clangd/index/batched_gemm_reduce_fp16.cpp.795EB59AA922209F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..531addd0c345bebe6a709d6937296cf576f80747
GIT binary patch
literal 1562
zcmWIYbaNA8V_<Mj@vO*AElFfyU|<jg;^LB`%y4D~hBOWahKe~$Lvr(Pn+aT-e{=bx
zH5MoLlr`9^tlTS=Ewng$V*jB|8_$-XizZHyeSiJE#>wKx%a~s=wH>XNyZ7<kPG#c;
zo&%v%%QP7GW;`qX`sW>!aYg<2^o<oecK+R=kvIMFP0kPBTKC4x&xu&KUeCYb=2y-S
zMn_&fEM}6sr=Pd~&%ti}dfl$C%a5lPZjL*<oJZtn`_)|$^?QEqsrge9f9H<zC#jDQ
z-{mQURX*l<>g&+ynEUzK+Y{m$f$y8Sb`~rvW%Twv+O|ja?_;^jo0_`Csjcr+=02?7
zw>)gA`~Lr(9Ba#`UV756D7W6I!s)D&&Brolvrk=XLPRrkJlj*M-q-xfJNZ6&an9VH
z=PxRg=4n0WRZ~w5ZE@K3XlrS}rlmP7_uS1i(%B>SRDOP^yW~vOa$%OJNg0cUl%MLz
z|D3uxd9q5&+^$88o}y({JjyG-su&pQdS-l)xT)W><aGq=gHOHP|63QQpEJ~Qej~W(
zy-SImjK;dY)rPOGv<aN8NZIA^akfhKZN{kd@5f3R+jh9SX?rJJD0c39VE;rm=c4u$
zF9YoxF7f3nWs{xUS`Qp@Uh3iP5v6^lOV+<a&v4g{{LCY|;m0N|X?dW1W2+}yz}8Zo
z7Y$oAo)*e7g`C)I#59lTr>=X@q{%^Z+Ux~3exCMvO2Q7?IZ>}(lvr(8SUB%D|K{ut
zzDvI^EUtNHzH9HbE9dgo+>X4laQn(#if@_4Hz;ZBUZunzdteryi@3!_lXDqQ7yNKM
zD{iecGwkybo(Fve%(rK(^A^fnqn&vyY{7o(O-=46R@$A4VB%l<{-<4KoXh&G{@lq<
zCDXcJ?mu34^4_<WMV{(~9VOiJ+j@*199(~GHS7Q7YQ6p@zn4d|EiFqrd_^%|BQU2z
z&r)}7PuJ&eo*1^=91A=4$MbfcGuECu|9R|(m#i~mpVwID%=KTdE-qYmaA)!7D;w_{
zRh&Cl&7;g88K<P}uxZ7~8BOi0gufhqxKI3r>6#U{Ugjv>GCN(VQ}n*2F<w1<a`&DP
zv-}Wkm$LGH>8MX4?RF&}?;5PR@PD3RVPus>{rS)HYm531?5Hr+JSscSB<A+a<@3~^
zJlqg>UTuBrgXq^K3pg%&hut_@*TpxrBVI56=XB%cGWI1cYRPwvn@^?vXOgY`$<I(+
znVTdJ%x4DTS0o<4D~jY{2w`9dWM^UL024qN2q;QTD?ZK0z@YM@J7isctR)LipopS~
zBV1dctfDLuPfbw`iKnlq59h&jFmj<-U?unQR5n}UM<$j)D<LZpm@1Y)dm(!`ul<9o
zjOF<rDITEZZX!M~MLdC`ZlXv$F*h+Jo~)ZJ5>L)e4v8o4CXd8Zb5ld&sk^Bo@ig2t
zka+rT`bazjHv=S|p_?I`w=l<>#VfGw5y)3oRyHsP1AS#>59e)+xnZ=ZhhK<^C(w%5
zN)V=qC(xeP9?sim5|qN~y680{Pavlzry)$yIpyk(Cc9vMkX|P%H<&O`uaiBT_nKo(
zKa<sl2|&GEnp`F@MGQr$If<ayEiOtfRt7NwKm-^t21@=)$gPWiw3sO<fFTGb%)pSG
zo05|W<ocxhF|)9;v2$>8^Y9As2@3Fg3A?#y2y=OfNQ#O{Nn2Ul+SxleIvI%T+NfG+
uDPj|qkd>2HQdUt@*VNY0Gc+<bHMdkSlhHSE6_@bv4rX%qclKoTWdH!*hkAkl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_reduce_xdl_fp16.cpp.6EA78CC685DE72B1.idx b/.cache/clangd/index/batched_gemm_reduce_xdl_fp16.cpp.6EA78CC685DE72B1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c8f75891626f2bc507dd1743cbc654f2ace07971
GIT binary patch
literal 8444
zcmai(2UJv7x5wF+ab%dw(A$6v4jqOLLy_iHK#C2;D0UH$3y6iL#)1v`EU2JD1cbqk
zf(lpy5fyt7qb6UXCQ)PbduoDW5*3Z%otgK}oa0(=t;bsKb$<TOz4z3;_c@ak7ZdZR
zy^hX^xp74~X?e2+IyySW=-=YJMVTuRG)i=IimJc*Z0gv@!Dc;s?jG2drLP#)Vq%z+
zS;ZGmnszbu*I<{@<SRYzZh!sv`|zH9#TR<{*-?j0K79VBu=lu)v5B#&C6V80l3uj1
zHsgM7tU_*C-ucV38}mL7+<c|w(cR*QFIxJummHgQGAty*sJg~ZuXaLrZtt3hPv?ZL
zIDPSa*G~a6e)dZ5n->x(Q`v9nyt%9IHJ=wb^~V`k8nP-ngrDz8l+GF>p3ZnG5~|Fz
zQl1%9R=PPCPMUl@e@f=pCZaz5O-!7r_uTk(zuZ3XYW>ID2X4`GD)0WqZY<vr=I7Q|
z+L$m-5?hzF&G?$hR%PPmTh1Nd&njp#ayKkL>R;J$C2{xmb)z{^tl}zrCB9sc!G)}T
z5jCYg+2QHTv!8#6Up@P;{koz<_G#BeKXWPd*Ijl8jv4Ay;9%=LO5ggA6Z_)o{mxa5
z2p;5BHCyubuEn=@Yogp65`KS^cP!eb#uW-DXM`2r`$59J{APEOZBJ=_TBF~$Czj2g
z+Y~lCzH$1Yubd2W%SP3&s_32-vGmbcpD&J787s_f$8VXCzDRcamUN<imhUS^Q~L=f
zb37}}PWS1=TGkY_>!+OIZu&T$w|_nINTZQZzG$Cs-R2{6GVhfgb+R@+VQO`3*1xYh
zCI!UTS?|}a4$}Xa7+3mi_#G~$q)XVgPFcBfQ9*pQbA57_+w<kdr@r6*{JZ#Po8>m?
z6JIs?ws$GMUga3^^*N*5DUu-%J*Cp?Rfk$ji!5(sc-U6#xcINkwKlUnP88Hn<+i4D
za@`BkAC(NgbbM{=!6_-0$L5XR{Qc#}=d8EcOfoON;E}g~#oPY)7iZw|#yXK3XBnZi
zh!%(=r}i9p$HwuUJ5NXES;y^t>ajMoJ?_Fg{TC)TwyVr~|8|v!%6E8dm0k16xMwqR
z!t9<yr<_{{?z&y-9TzgBR<cF-&i7JxP06N@#@llKJn3znZJi>>dlxvy>N4C8E;+I3
zcxzb6!07ic^ZqoS(lI)%I=QO_nh$T7Q#Gz(N!$uqHM?b=+{|<2r6mDh&f4-*NyV1^
zCfsnv;dyzN3huSf>ll0`Jb3>3Yj)LtKHRI@d{;mH2gS*jkAYpw9DHR1&n`v9ntC>f
zCyOPI)BBv7zB;<AXGcE!!dHLrQ}%;<RH6B+FRwozJM-HkwhPY&eYMRt(R<&R%8vDm
z#@{`d`ryM4r54}aS`qML*_(rf-UW)9;|4+c$D&7k2<uCIdH(3nBMWb>xsaMUW98H$
z*~65jZ@W^<|GL2HestNlXL@q(#QU|a70s`LTvv$32aj(n%^!5bE%uM=cUqgwQts#O
zNxtAR&)n)>!QW;3Zsrs};6<`bNkH=LuO4qc_KkOc^cP>5CY>H~bh)_Ws4(Z}hKCP%
z3$uLroxjHZeIawpn76O)e6w<^XU!i6OPc3&NLL)IJ$W$Rwe9ZmV@=%D@7KTkBVhK4
zZ%ZbPJzy$37<(^+f8KDSF62K&!|&W)?a-AP7S`72%`W@$!ImrU8a6EK@Uj1Em{skM
zGg94-EHgB}Z0tSk$uL*58LYuSkI&g~6aDH{ozTFHvDaC=BxlYVT^$|0s}{G<q@Ux(
zOs0+w?<@3`Y7FvJnm=(-+58RL3@(b9G$cn0qvaZdWsB``VS>tt)naA>lI?}|h8lxq
z`<8VD3Olc}VrCwaeS|(X8iQr;w@cd3z8O^|X2u}dP3UH%F<1^ee6zOX%ko2FW*m}b
zLYbk)U^#m8>ps5e@5y3j3X-jaRw9kTa-8{tZFhq=n2VX|NR|ubW*UR#X}7xOpKXfy
zNz8nPWLu%FNMo>^KBvBG?Z1l+#LP@2`w9K5GzQBHbj`c<gfkk%%y=Zb3*C)02FuGn
zOPB2r+4NG(%s_IWFwjb4uv{WrAq^Xd%o8(nknAP&GSnEYJa^6ZbBj#V#msCZdkQ@b
zH3rKYZo0LNa@}#B)met*F@iCA8iVC+!((9DxBtu$Goz7gC$uxv7%Xp3I?(*)zfPr!
znG_^j3#~;OgXNu8GksE{&QB3DlaU-Q40qBPEY}u4aZZ(Voe(qWNVXB$h%^Svs+U`m
z|1A0`$3SN)l0z&)Tr~#E$A2GTdw$?)r<j?AWQ9;+s4-YRn<V1Bmwl!$X2v4fN$6yt
zF<8EO<h=ivAFquNGZT^QA@nfP7%bmvb3YQb<4v5HnTF({!l9NLgXMd|jXy0-n>I<z
zq$1g0=x?ntSpKne;w2Dz%oj6rk?bIJu+SJR|NPJ;wcEMsp_oZRva`_HKx44n`+fd`
z<p&Ic#Y{4iU4$+M8iVC`FMi(U^>mh8%w!<hLTE8qW3a5VC@p<)iKUK?uCMsep6)rD
z^jW?RT<gG7JJozWxYpCGf&dkS5WNBX8X$=9UhvyXa|;BuKp4?mA*hw+Hi&M6c%q*I
z<tZ3O_zWn|5Ek}yzWB}km_Ebuhr#hM1d@u4;MhnQdl!NJCtii+{UZG);629j7s2l$
z#E_R%GiKF{D`xD282wLt56iD%?AI{bP8noKgEAX!&w+q{50dv_grAqx^{;d8O)&wM
z?*XSC@Wp~rrg!#<+`6$yxB*T#z(*^bvCHw_G3us3-9Mwl;D7~|Zv>-8nw!9=iRNZ7
zY9?&zx#jwuJl7op0bdTvaxx<Vz5<jLH17iYT`&}XNWfQveKle1XodQpI9Qe+6*wva
z4-m`m1o2KV!!Kd^-5}mga~sIpz=P<2fWsf4z-+ls5fY(Xu>p<8BQSpi+8MCA;k)$O
z!QtQPvity;4}c6m!SbsZn^lYx;bO+7n6OMz98#*+{RqABF_=CEdtA|8zih!eE0dj!
zfZqsHHo`P42>3D>Uj`E~;}C*6u8tpg&!83L>K9H{kiV;cINstc?;q(ukx&=-bwMnC
zAsQ6MY&YYEdFYpm=KZ!RY&}{wk3sSnjIrP<tu<=+&NoXlcNXO^1Pki9v#6kXCs^zR
z?HdJX*e!OGVRjw=F5=?9&TU15u$J*%OIEB~+>{po))FVQjkYt&?F_X8${A%jVYmHh
z;Z-YJrlO`)f=wlH1OmPaY^rFk0h=1|CQYpcn_9x&XM?X+iWVwSQ`R${>lv*F!J!y!
zRCOFC!_4Jyg*Yg%e0*ShBw1}|L2(w8_*=n68|GK6a#x{u`*Oac@zxJDx=_<rRrn9u
zOUv|@NpNkJ|6K2Lik9mwr)Y)V3N3}&^<B>r&h8TlkXMw{k$o*dUQtp{vkD|Cnj1jU
zK=WRZ?4`LMB>kVVb3a-20=^%H^+PyrvYIFNLkeb`pA6RjBrjq4*uYq7Pga1qg5HCT
zplpOl{1TRL24yqNtsrkDi-+Z3gTrgGN5k$+yz{ze(|BDp^~~5RW*mMI8VyEX&uCqh
zfLAec72yb*FxARm(z4L>#Ms9K;8~3P{kz_#gS7={4Yq@EJ6PlQv3v&@cM#@A7o>zm
zSy!Ov!(+mEye4oKh6X{Mf{}$|C$$@ba}k!8No7Ouc(J^_)P4x|Xe`e;b8h%`h&(wT
ziUPPGOfenndb58^-!quySA$|T_>rcr0mT}c*Med#&Fet1j^<Y&dIg5$`ByOX6}S-Y
z15qE*wW*~3Cn+S$J2*P{;miS1pu-44l_0JpqsO9+Dz2ir2E;Wq*Mhi~=9A!hl58%P
zKMk&@X|89Y>X|X5$tosFML52yv#e*$Zw6@BIoLTkkY;&VdfDRq6FE`2p9|$g<^C?%
zZH^Awo@?Q`X)AhxTaX)v{R7Lp2f1@N{bhNNAP){FW-RZ+`HaR6d-UaR#((Ge>gpf?
z-vW{rkm9Ba_*Rg#(%c4;Hkwa@<Rl3d0@QWMX~H;jN9lk%Z6_;^v*e_Bt;{@h=wxG3
zNQME+r^(W^n}X%jW$6@U$TBF(l4VhpEz71TN0viTu5qq8?hnf^FkV2>LgR%LEw)^2
zLu%$(=24VynNLxnw9pRM#B+lBs6Km8eXIrU7~o!JE?QhL_T7P~K`a_$K?@k*=UFty
zf>xT_K+s0>Nf4Z*`7{Vl6V6JS>2>?{%$I1hmx56#SmGCFm(Ot-`6#{+Wdc=<Z52t<
zSmb{E>lw<gsThA1VZ3BefKeZps2(ll)@nP%@(F<nRD|gSMJHMG1qma&Jtiy1p;m>;
zLj&+>qp0iAuS*5DZlI-l9z^Fs8#!415{NF*d>KTSX}$`gtAz2f4UM|`%tp2G3UGI|
zjVxY%dhPj#kItx@p$rYj%~&lxu4jBaArz%FHH=*iqfL?od@W;FOL+Cs#GK@sTUXGm
zU4@XVFbWSY%U^?#Yc%&jWDiiA{3Z;$NvG+zVc2cLYjy<>>4p_6kuO)qDU)z_SUxB&
zNK1J0qfV+%o>V8xfveVbvgCgcp}P8?96njz%hk&lzk}sBG15&W#+1B@whg(tNs30s
z%E-zB#|GR#(EsG5y5WhQ`2GpSuc)?CPw9r6yD?sO?eAp+!N|=g>m^f8dXnBGiayi(
zjG`%eQz%N&OQC40-c*XF>rL0vrs9w<8W+CdQ4308OesvjE#17_x@7%~Yfn&8Sk1Us
zGuo6HM>G`f)JJPFdNF}9RLn02aXB6H4}<bB9pjro*+g>-$Xn=`-v<tTB<ABYJIaC8
z2X^w&q0*s4@VS@e<x=?&9I8+pjo^l3KY?fzmq5`3ZW5*)Z)Z(kZD{))8iQW&=p{!I
zTmdZ<^*sRjYByMQlWLXEe+W_y6t|-1--F$I(hGc7fL4(DqJTWlNjPKta?~<=&H+>1
zZ|>u|EZ<&03-}pWJ_D^kWclY{`J8Zlc7h;&{Q6&n>YUK4j-1@oIiXiQ%_<00k&~`~
zZ-C$i2*ndE;P*oCUYg&7*L(83>ZQ5cug@LcqrKV=@^(5$=m2>K%>$quAV*~N$!ORr
zCJKLu<%^kN#e}ifK%=X^C!t2?K5)Nlx8dFmldGQOpS49J(gco8<YcYR8y%YoA1P`d
zG%iOKgQmekVqt-=9GcypCQt1fuR!TyKS=w*7Pm*uzWv}&b9_G}Vm|gDqOi(p>^x0q
z32y;yo~aHk;jJ{cK|&kMzz?$gDF{0SV+o&uurq{@Py4}7waL&#sJ`S7)sdKkA_$1;
zX;y(qMRNm)8fe}NqP;ZtgQ%Y{zHLJ5Pkq%yM%ZhJmoxET8yV?F;=xYG>?mrA`p*wW
z0zS<#O?v<q@ac}}6lFMOP?Y7EMNzh6Hbps(ITYnu<WhOi0*eI{Eworj(PG=h+QYJd
z&$G>=DBm`pqC#1rHZu|M2bjnM%v3!10{$@Lc$nb`H#3gSgwIHx-CVzb*FnqcAR|3U
z#)0LJGSZ`j&juvT9Tcs4fnuzUlZ_|d+br+o=H!bn?pWT%&BYgAhN*q(h*C(v-^R7j
zidWz5k@+Zxq;i;otKn=CB?{_0K2nV<0dAXiW3YT;U?P<%oCC!<k}32!EZeX;VuGP&
zktyrQ{#GwCWj)O*;8gV8Qv*abz)0L07Ws#$y)?fE<$Ln{H}{oBj|**m)e(d-D`m8a
z5_ZcdP^j-BNjKsG<6`jPOMUK--OfxVUn^%~%W1A)j4BxItq{xaV2pOqT*bImF<Ni-
zZ@czEcRlR8QET=y(*0!ma8{4TUVZgRTEm?H?vi#;?^$&Ivoz9vsm9y6wvz~?J`B2c
z&^!QP13-DVRm|{JOf>08F*CfF@cng9zCZhT&<WJAm3k{#Tn!&fP_xumu2i#9IPD1&
z-@T$r>Wf!WsREKJ$hvtDIPI6>mM08a$_kl6j(0!Hd&)fJIGcH)cchF7J+>89vURfc
z!AA*}cXo3Qz#n?SHzkLda+gr0jbOTw_VHz4T1NOqdDEL3kDGh+SaiE5C<lGqFBaVp
z2`Xsb34)z8SAw9D<|+_W(YzZ3yJ@ZgK@H8dAgCq$+wlj}GUqwnLg~tG#$z|5z0t#a
z5&1y%tuz_CT@bblMv(SaLs&K8mo2R>35nCyS2?{fq?eq%a99!Pe{zG3qIf&Qbuc)N
zBkpFn8#MpOa6i#}K$knHOVvNE%blV5sxEhp=92>Mlz^)LSit?N<qseBcY6Q0^G`Hm
zN-t$B&LmOd>lLJ>kLP-vzX&U;{(q+x-N$F6UtZ8}^@~y$cd82u)aHNw@^C>%Ctq9C
z<s+{o`r3s$*BTZrDJ;m!ou8AvaM9wt`~^B$+DbYLNpTS=E+NHiQp_dALQ<ShiaDgX
zm=udhv49j)bLM8FUmkSg(-L(BgM@?iL}J#!$WSuMcqHd+tUt=c($vh#IyhulSa?Ka
zl$W_jsDsi~_J94ag^leH`=N42Cl@z&g}0Bde_)VZfYj4(gt<jb+)^fb((qVF(D`pV
CGnohg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp.3B79E782257965A3.idx b/.cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp.3B79E782257965A3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00857c098745021f4a84a6af19caf5b7efe1c202
GIT binary patch
literal 4800
zcmZuz2V7Iv7tRY19?6RgB?=}$fUx%-DJ$$D0S5wtVHN})7?52lO3R*AK#Cw@v2}rp
z0}(B?f>=Z>S~pTf9JLjkrLO<YtCxrWU-SF%<y^jV?>YD0bH9^dA5YKYN)(D9)+ak9
zE+a~sLZK*vzw``oVj~zZ6)2SK<_jS=@tbxES1vwT?mFgs=z3-$|60{{)f6QypIWcN
z4o{`1$nKMNnWr}v*X5UvdGs!*i}=xO$bNEqxKd5GNV}+UXU*65i$WtKww!&uyp=a8
ze!*LM2m5;M;Pte+7IoCgDLgP<cB#(cv2UdMXRjAd^priiAXs#3O;z!o18L9x+}M2K
zFI@M!$fnObs3xv;_u=GM8>Y>A`g%ubSM_EUr*o{<73%M=poZDkTk6(#+XRX%JGQNH
z{9Vs>k4@DDUM!<aReQTEH{HOhfm^IKqFxYO))SZ6H<p&yRO6+?+jtL`xp5E6FrjN+
zXYVTaORB8Ar}X4Th+D(@gU?u-U$0)WcQs$Td-~@>qqGCeZTEZ=0@KtJbSv<)Z_gA5
zs{3gF#&)QwQ5AReO>knZB73r~O(==)wiG>39%>W0X08)|Z=5!zI8x{rzilx6fK-5s
zZp{tVxVNh~`)b-3)2EA`+oYxLyXlg^32XoP_a?i}6NWJ+Zx(C&anLc(&aHA)-MLvM
zi&RUZS@#Yu+$fuLt}e7++^=~5c<fQkoXz3#ngxU3xP3Y7^Wumoc%om59q?kHGwI=^
zs<EKK-7PjS`SC-Vrf<q<+x~s=Ij6pL^rGubdW!lBBX;W4YUvkRJ)uqauj{tdrJvb+
zIof@!+an^*XrQ&ZRWa4u=!$3A>r>M=t2e2>aQjv_pr!H3r4#k9XzBYX7V;iQz4V~(
zHA+ACQsapKgZ+cH<M~I=Y~I6*_czR^9@X$u-^+9O;jrO9Sjd9+^d0scHHww>Bhvn1
zHz$}m&Dn;&zdSbBgQ`|04n&-M7N1f$V=!K15UuQVv_ff9>Du9rn737~k0br~>U%4O
zzx$n0WOH(;c}8Z6_iLS-T*0Nl5+CnB4KG)pH;dO<?VcWGnviBw|6x|;?Bh{hz{T8m
zYEB(Xddim^bXGd#Ta`B?I$NVSI`&QS$f$hH=RCcw_9JO~Ex!s$zEbl3rOfFAP4A^I
zuIn<$KDVl8c*-Mp{N9G`LVx{r%#g{~<B6St58TFlFE{IM_M5scz&?nKxH=b8wqtfr
z<z><xF`0`KFBHZ{ir@7;`fJ$icInt>>)-0LQ`#;qP>DD!<KJa<d%)2-{es_^xmMM)
zkc-;wZ|Hu@eyYF0Gm<Gb<Xc=BvD{Y|yJ@SM%xmlE$)+9|R;uM`xyQlZ@5SH#T6y($
z_E_byA?!|T@%bj<sv(Z3aPV%q>YJO6Eu~`(ci;S0Z=BZ3pOCLTSUV_6b$asD^w++#
zditZk6a@cM;CJKL?zqkGuk3mpWxC#|{Q$EqzpyoO!^y<Fviyaohc)^-WOl@H62>R>
zQUdx%ZK~Hxoj$Xl+BbedzWJZP+NlHk4`yjxje7cEQG)uR3yq)2$@Us|_RQ|+sl;qF
zdvlZzEpE0G-{^lcGuzpd@o2`=YSZH2zpXveXEo2LeW-l;i0)Eok@;>QYvsV~Nbl3E
zg;Nz(h5X}(6U_|5#8t*!wF~$@fnE2_P9+W<JyGDOa<^D>;2`fzUSwO;(+{pXUp?;(
z`id6q_(NoyTg)FRGtQh>D@&u@bPJ@ObAKyO&q;|^0QyARy|id}arh{c5=Eil<{Wcv
z0)T=MA6A_YU%Jt06N9=M&;~REc>*A<k#Q_f`r+%NOiDDMO*y981VCC}6^(9VD^D}1
z%K@!J(@`J*(x$I-dQZLb-o~IV1GE**N`U}KTbsUbdQmX7l|hXMbTL9JM%W*}pv!0B
zx1(n>l&TrjmB2-prmH{zq`l~cHTQ#l26Q-}^=bMn0wC=jQE@8i*7tufDan9#VYz4$
z0BOI^3;SP2KW}GJA_48da!?}x(gALHF;aPELIyPf&@~9H22uO?HRmFzc{AMAy@5du
z1TKa&Llyy$PFRz7wS95rB!e0ZXg!*qA_0(2?$)lV_-Xe5gSrCH)--EH0wA5+{$rlM
zyX!3`MFeOQj)^t_kj`fc%FJ6z4>2hzfOciMatVNR$wmDm-g;a6G0K++;9C&s76kqH
z1$`>gln&S)algi(h5#20ng*Q!NY}c0B5N+ck7iH<0L`QE<OzUuUGVPCR}YWHFsKoL
z=F+%w1VDNh*Mg;|&G?Q%4Fa?g&4@(+q`RbFrVLx3d&;230a}ZurAPpz4^Le^Up*bX
zlu20!XeX8vmjFn2uaMfNHW9UpNl5~<Gs{_%07xIHK(!oZl|~uV&jGy!p=|+`_3syS
z>3v`u^!NI;x=hMi;Nr+~<PreszT^7(_A%#QGpK6-EuaaO5CG|O`}!?Uyc-fQsG)$i
zquD7C0O{eDuY=xoEJ$Tgg@Cr9*(@Xg(zkqexHq&JwK1rx0BuXNRU`n?<7}6%Cl99g
zFex#BHshFS69DOFi)L1pyy*f}9Sdj+j)gV>kpBCR@miB9ktTzh2xv8$nk)g37RSY>
ze~Xg3K&5}82N>>+ddtJzv$E-fU%cArgJ3vclWz;p42p&n{S&dk@BlPG5zfW%dPG={
zgu=UkQUU@>bOh#u;hw1HTtFx&U}cG#kWPN6pAxxp2_h^(!k~UYnUSV{A}<&oj0Vq}
zRE`MC$w?UAhzJ{zFqj@(459y@1lhJbq47bp)!+<l9BiE7)k0+gnM!1d%+w$CpBHp1
zBHT&_aazjXS#UQliHd<Bj>SkIjEaHkblidj!fc?hfy0uh8o7E98ngtS1C%%5B+)j~
zDGUu;1f8JB0Vj#Vkxn6Khyrv9u&qmF8CTSRJ!;u$3E<m=3JA{RU-^Tx13u~=SX3f<
z@ZT7|3|%&t-g4zYyTc#K?|?;X5sg~J3@*1eyE#?Vck&(}G!!+KKox>&DNFxEB{5ux
z3KzmmF}xHJmLlQg!IdGxviaj$#Jm>a!&s8oKDEdSa@>Fj8<22#(8=D_mzmCb{va23
zP4{3J21+xyU5PG}8(fYqpO;GoBCH@|XZpDh8?Lqyg0&o+9fIJ=LmdZ)B@uCQEf3UV
zuJECn11E`;lTP7i_*_7!>EJ&8)pO`pF8gC!$xq@hL0uVg3^n2PV7L*-ND~Tixvx~w
zpVc4iWH8)B*@O+#!f;b%Q#O2^7;dI)#)b-n;pWQbZ1@ub!!498*ie4pwaI+iG%%c}
z!sEc!f#EtTIvn__F<e(gmjho8h8w6DaG;h{%AJc@n|k>?IJ_3bpan68n}th(IFc3w
zw~XPwsP8<nDMp0FNGMqrwTM$K;zeHj7DTv(6eHNX!2L=Z7rCgOiXIylB8Ka$=(FKd
zgvSVONYc#6DS>F<T+Q#x?v)Xy?DPb?P+OqJf?B*^#{24wccQ&ChC2zILSS>laA$#Y
z2<&AT?jmprfsY!)T?MWouq@z<1cxALm1J-$mK7K3ABJ1AthumsG2E7A%Z055!|hmh
zT-df?0$?wa)=Dn78eKhCZ7?t3Bx$guQz#lbCok|Ffs*|1{t9MGnkzZG2tkXGxghYt
zf%%dq3*828-8!|dA*5gOl$JY0!lwlD2iGR)xzHQKSD~xsiEa}jY$8P$DmIuT>BQuu
z73d0O7$3tc5n&|~0SgG88<-^N&*Y?)=*qd|F?<^$+(wECd^TXulCDio@<P4lUOId~
zz)8}(NhcrFXU+*04uk$l^TzNbG)a!kIT=l!H%dWM=8eRtc-|--O`kW)Kr`lza?qT4
zqg*t1-Y5^vn>X5opqr4n@jE-i7pHP~Qs9#K_IwZc>)?IVJAXDE$(0m)sFMiR8YuQq
zXAvwaQ0$>DqItz0>MDZ8P7%i?M@h~lT^y7Ci^SL<0pcTtH{dg6MK4>k6u#J)pO>~i
zU7V4Xos+vEGnJAHoxn`eOe0M(X=agTHfe4o%}ml<Pn!9pnL(PVq?tpSUyx=VX~v|)
zCW9ZP6yG?1DQOv+EGoyqm<sX>ycONubQI~{N{bgRTB2fS@5ukm$=St3+0a4DMo;~}
r{>xg*=4fy=wRyVw21cf4=9bpBYF4Vo76N6Kr%wUZBh1wc@uU0?C&@1h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.71938F78DFDFE1BE.idx b/.cache/clangd/index/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.71938F78DFDFE1BE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..77ff6c2630e7afbbefe14ff49fda93904efe8164
GIT binary patch
literal 4818
zcmZu!2Ut_d7S0V2FXSQwO%zDLV1NY7(o3Wl=_ND)5v7YtE=X@mQE7rgo>X0VD9W;y
zwE!*_cCmoEsEE3{mQ{3hRYb9@D=X-mL?<`CZ}WZp_sspz%$$4X%zr+Dy*)fGXpl&5
z3EsJ)#LQSl5{X0ue;JwSE58MU40RGIcl*Wg$W(0`_1~SYb`<aYHT!ORu`VYveZ!fa
z7`uswr4KH>(0Fi2zjxv5F}pi%jUk-+*rUbs#Bb^>?k$~7r)kjQ4*O{}Xe8yWsb4V`
zxRy~jcBL_)&}QyBmf>md!aUCV+=DxYHe358{kg)k^p{QZ``fCX-VIT|6TPMU9-j7k
ze9iVDofn4ZW2{eB1=Vs|JB}p0%bKw4E&9=)a+Etso5-_VRm|g8lb6^wm>V{9SsEWy
zzUqf>Pur>;rZxCjL#z~`D!ydQ4l%N93{x@O?lj##JL8w<Ngbmf?e#AE`AiMmRp}bk
zd=i@Aepui&r{PyE?E(FpH6|M_osJvl8@v=)EIyRxl{{Lh&^)?S+kCKoc*lLm;efe!
zD<;d;S{=pC`y4YlZ2fk>g8ry;$(Fr2H}7i%KUjR~O=)}WA)lUA!CZdYV_HCoZ&LGM
z#sMY&jV|+cTU~g+e__OYPm|gAmH!l-J$dNB`eD1b9+eNR)-3BbaN~cN!^w!I?Wy^l
z8zLBaa+e*PJg+KpXfWVCYaiP;Ny%5qbL&I#ZzoiqCM)wsnUoy~<OCnreM<i!tLcnB
z*=e&b#MSu56U{P?x?VwCNSUd6-Rkv-C_&%%>fJj6u^HR0L}^BLWs7@NyW+?})_vo>
zS1xbm`l$s|HESMHvNj%SDoW3DIl*oVTg|(*mUdKEzk7r#THLa)>3aY1vg~p}1JWEG
zNofe({Ad`dciY*vy*Xf8Ma~ubo1-rE8peBQ2a5UUM#^zzmsQg$W~b}d|70}G^gUOx
z#CMF$*uVc(i+#WedT(MPEi_c?+s?u0zL9)U$t%4GS9%+@RojEsf9i75RkL;FobIFg
zJJg~*9*;~Yxw2BtR}_|HRd%%tPc8pzew<$LKjgo5nGZF0F4)<5FuW_>y6e*;rL!-F
z1vMu5(>lxcH1+PSJ>;Z!cyqw<^D!4|X~S=hBwR9`S!*ER*4ti6+iTGro)W$3V|U5&
zaQ~uGL;Z-bJg;9n{SKS=J%4q$n%gkDyUh4?!OIgjIx^Z;vVAV?>bH9HoY!7t+Mr(b
zXJ0jC``_iIy{xiQ1zX3mqi<r1QjghLROdX%HPjL9-sLdLkNR2h_()dCWMt^F{F<|A
z!QbT@XgM%$iZptz9f;-G-nn1bnY|-+z5D2#!qOU@4O{dlo;36-y~7=T+Iy{AXfytP
zH!Jn_s{^c;yEK<M1#YwoTjs7-Jk82CNq(L>&a8OY_+W6Ll|M1pVf1C~p_0LvRLAEp
zO&**)%e`PaQWX48(ca&Vb|e~4@7;L)bjO*_>RaLHBkI)Ks;kkrL;e_>(x*JVZbE8l
zr2CI8++AbbbnBmQ4vgj;G%|fQKE?g+)7F(QS{w^zr1Rgcb18XjJ~oh`&wSo<V&L_a
z+t#K}+<T7G#dEgXrr+xSFgf`{Z`Y$&9+pNX?GxHnFJ3Py4V<#+xoTZZPJI5jMnoT&
zv_AH7h4N~Ta`(Po|0!3y5mr-YtC`Y8!C-k$-Awz+%M-Ejh0_zazyC0vQlVQlcd%Ud
z`*j=3P8P8%XJ7isgFZmh-T#bkWa&7;|Cn*+MpS##@RfUM!yhT6j66}CI#4Nw*xO#Q
z$+4754kwXtV~Vkw6o8INfBN^#(&u!o)=(v_2DCH7Sx*X(=4qjEZA|(El^g<SLyDog
z6d-N#ci!>7cV5j@asZ&MC|2rHfV7p#$CkH6kL#)ABtTanlnR9T<pEtBqVC2`W@>Dq
zlEZ)tm%>$-0;HX;>!43NPTo=_#RA%#WzLZTq+R;voqt3*R;x-{1!yOR6I%+9UR3k?
z)v5zi5mfS0K(i@qWhp?~Q>CQ#QP3H%_Yy!GQH&T;fV97BLA*kNsgO!e26P=lsY7(W
zJY+6zwI7BKc4bk?fxu+}WdTD9kWOA&aAVh;nm?%IU_kRIJem|_^o5JMvy>Z@R7nYd
zwqRLsqyXuZF3#raQyl|Tawwp!Db_S8K)T{GuiJ}T-;a?>5Wu%0imeFx<pF)_Ts)Aa
z*FMHk$^O8_fMTE~1xSm7JAQcg<Y+vV91Unaik`9*Al>9S=*`>k%utmS185VL2}cT$
zZfBb_xE$(kDme(yd<vf-1sUx`?X(IkxTs1>0kjLlMNbNlKB)MYXvpf^ODZ`L&>RYf
zCIv_z*&Nq%`tb27RgwtMt_)YU6d>JQjm}><sWD6?uK;vCLa7Ip_4xr^PX5T_3F0sO
zO(jPG7k7&LTq!`hZ>Ff%{N2vIs-!qTo3czfQh@Zief<{4rq8=k$q|4SPz35yfb`Ig
zvY_cbGgGN#A)sw2HnXGv>8s;8Io8?zR;r|BfVN}U=|}<6cf9}W-q^OFjY?h&Xj_Uc
zO$v~{SN)Ltdf|I#Rnl@m+cWIhQh@Zct-(K^?iS^#lHvhv#xmnb0n+csey{uHaf}|7
zyb{p56x|t8fOL9dQpPqE=t8c_;M#ZVyb%m{&~vbXdqTNTQ29zGFgySapuv=4cq1Zg
zL?YoSKoKFUd?g+jz64z|8%}~!0#=sj2#g2A{ZKy*=!D^wh_Dh_0+j;_jH1d{!h+!;
zXo%dTYD8E~Ov3PHMA(chf#Jc$5S9NEADdl{Xj0JR7LWs*g*Hy`P@yJ)NF{<qMCy<F
z%Wc|#2pfn^9JTFt7ClH@O~$|`b`?k<+!X_LX;+T~!f2qRfn-Vaj9A?p^;Uy<fWiix
zB&tR@MWT_jpc9lgC6%va4Z{sI4VduiW4NKFArmGa!?~JVCj3AQH`X*}!Xt-r2=<g{
zA+cvL8ax;7>2D)WW$>$OLCQF`95;B2p;m$ve-=tO+y9F~f<-0r2w!7(AQ~uJG=v^#
zd*sumdtgy9GEa<{!sVcV&QSSEI59j34U(k+!>bTs6%s`xd=nzvBtI4-W@5x1E-Kl?
zTa1Jf<0eGdghasuO8z3tAY_Rq6CwMcKC%N!@!E1t)rso|LSCe|C>W*(YBnexiC`0R
zL(x#VfNKz84H00rulta3gtZW?W#?oU1jVC#Vi8ktxmpXnW~;Lq@U)@sgH(Uke7MP`
z89%pGoJubS6}W)4Ko4df!}%<}9u(_MJ_;Mgw|ufy!f+G12@^^@hMUq&nNSlk+>CC<
zggS%a=5%u=EL#k>pj$9uj%t+8#V=32b^*lLiWs#bd>9=pKTw;JwgY3ta9`9{PDd&b
zVFeONoSYbO6eFI*$<!mldO~MlzXI8nG%I3Jo+gh8YX!rNG>w?BnBfP4!;~~IVu}aq
zAv;W%9N;8rV}z3s70QaV{YyUsPLgIuI7OjRGLhStdt6B<+UWuEp*vHT0kwF)lGlw_
zUMmHP81Cfe6b^4UhC91CheP$ma2Ge1aJUVIySllC!!`)-8Azw3RT7(9F|627Br)8Y
zVa<jmh~c&jTQ+Pp7%pH4*s!a?Wx=;dS}U>KQgo@TOyPY1h4$~A6%dlP3TMOW0gFo7
zDSVCLo~WlR)dP3#^sQ<P@0Zkih}|++>o72IvXWj)%ne7w<s^3-BHTtuF4Su<Nz#Fd
zNnvOh9o|R`-;4-1Bhj$N;Eo_lNna)=MW7Mta1zWKSXt7Y38%&AVp)DLyaf@q5V`?d
zC|Ft2vx!L?5p*LWn+C5HaFg_I!p#fylBEe24prr==8fSgXo@n762nEPNN$9qxZEfM
z&5#>qqM34|EHq1Ql#k}ijSA2LxltinC^uS<pz9IYc-q$ZZC{?g0ys+tfx{yB{U2ke
z{g^69zNFYAonqvQJ<>TwuGk}8V&sZF(lrKFB`G~IC03G`jP&@78xkW+0-t~R&>)ct
z2s48)(+D%2FmnhqmoSTDrUK48VbFDibgzrHu4rvx+M0~?%$(f3{DQ3PRMHxm8!4MG
z*AiwXVWtvh9$~H{%tFG97bT>C|1C&9iGB)-N|YI>G8I!*SDWcYb9XhMsd#D3nKgT^
zrmeuv-oeqy`5U_N!ui%*-G6(_&}OpcvGq9mhCHJMCZ=W<RyI18T6}XiI>W=e7$Gl;
Jbn!%d{|9c~E1>`Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp.74FE791FA1A9F5A4.idx b/.cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp.74FE791FA1A9F5A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7d77b5be82dd761f924a8cfd59ffb9a0729ee050
GIT binary patch
literal 4096
zcmZu!3p|u*7oYd_dNJ?oGNozV65}>*GnR3W+@?V?D0d~deVIf<E|F3qC0ea5X?;m;
zm$lutR$oP_O(>E2mMyjYlCo5rwrugm_sq0U@AUh9em`f<ng2P@d7kr}^M45S7Yg5L
zG8jHd{(0%i*()&ygTVs7S=rLmOkk)0EN@?T#QoLxowbI0o`-qe2{`>*PO<wy<pbS1
z2Vu2w6&kd|<cjuDQ|^k%w{CY$rQe{xnTc|dt7gy0uWk*R>=g%C2vU!c!4}QNmpb}B
z9JVumBb|tFpK+~S-aYh(v8$n~hsSTcaHTERM&BgOclOj%t=}72)S|iR#O}ZTXzn@p
zF2|`VqS<MC-s$M&z7|`DcsG_^Jim94w|e!7p-w?S5ms2+tsCDd4P)Fe2tKzfCVa2K
zz#hqQjU}#Ti$ip;_cI@>N2+i8(dy_dv$Nu5=n3bU$z1`*ZJNs-Ug_pGXvT8?D|~^T
zxjtva_xq*e0*kpn|EjB3)EuwrvD=lXYiW+WSax?{Q<+a_V6edv?f!ecf|my;wBIgt
zOU>r64Lu4cMM|e8dlz~Kiq1`hp4oYO<n_qOHXD<>FAmOi37S7#lh$n_j@BCF?As_F
zlvX{9CH69REew<}hYY*h=DDnkV$^3o48Ni`=BnCwm14RKhCSQ<Yw}3)OhwZ7R$XUr
zmt9#}>6laC>#u(15qs=5acQ$4`^Jw?%C)yL2c>n(8uV=%yj{!64<E&9dUr}XYkt|R
zr0?^l-7tU}*5zFN2J?O+TGh1^x3o{^H|%J7;@@}TyGHKrW3wYO<C`geGf{xC`hBGJ
zwC>Pis(9Edt>Hk(-Tk&s?6KgUo+C~U6=^npmwH}LG#&_F*Vw(}jzP1lZ-4Jg{`Ii^
zW0pZ}eO=?E>wc-o;f#I1|E_V{x?l%)?L${8Q%l?)>79SOGTOWP-Z}4Ena>{m$vktJ
z{pIPO`${(kNhBA|s)&r_l>Azi*){F)qrbN<_it&w>^IN7p`5=(Q=Q{|YW&WBq?Nxn
zh+f>8)l<f8lqCOxH&i;sWu0nk^^Fc-_pPW<6O?@Gv1p1SB`KEMDdQK)KkQQ<+ECP<
z>6+bs;HTSpC7NMPgMLQ_3`<^ij@2ifeAb~l@Q44!r+ZXHkI8MuIhz`f*&pS-Yh`~p
zvw1_t-j<Emr+lxCIk*~@pEe<5%A+{U$*GZShZM=tDBD1zjwjUxMTT7-wN|P@Sz*<k
zb3FHS?JCzA;vChg>bf(2`*cKY<^5fEj$Ci#xR1ViHCne|^ya#ns2}VuJk^VQBsy<v
z#XEFRQ^)Pm;n((GCgtoNqPQbm9Y5QZR?3CKw}*1-hkSP`FIlTKSTnza`7gb{|1DVd
z?XH;1ZKTia-tori*=yd!7WQY3B0q6*O|G4rNqRoh=kaIo_3%m4?^vzZR@^w)!r-53
z>#^9dUbLw6%Jzo?1K;#sEQ|^dw?67tDlm%Qez&SkcUpP5Y08B-&1#pXrp%z`OXgfk
zcz$X}ZlBkQ*X*9nlDPQZ@>Na|<MvBXm$$Fh6irKP)AJve3x^L?PhaN8o+=;DnYvYU
z54YaWpLlhvRXS8y_h7{{wE|(8w?^TwCkFN<dp{cPu~4ZOn!oE4w6<2b-<y7sw9T+|
z&*dAcllhysG}f(YI`~(H-4^7r-?6;SYCCo>**=r<D#eRwD|oNMZqFjK^3xMSlo$+l
zwZA*I$72(LMll$aHEyjb7l4`MKU`haga-XNuLv|0z!tcLmRvxX_jmq<)(KH9fi3~C
zJ#Mch7Z4WkruV+t@U(_N69Jry=jzJ^gcrm-NO+$;tBOFE0ho((wd4ZAi<rebheLh=
za2$ZGa4WW4Ksd;|AW^A+FDB4r0B=O_jmW%zKA=m~h78et6Mn5aV<m8LVY|$e3kb*F
zzPh^gxbOjiE(b6N=cvgAgsIqq{>HgGo)BmVfGu%LmRvx1)ftn@iYtfP2{a7A4!8qL
zE+D+MhkI6JQPV~;@&R0lV3mmKKOfMgB6HSpr?b8T1R4ok3~@uITtN71FCh|pcRGPU
zg8*!fn`_DigsVdjpO_eFP9)HH0MEzgtH}j~>x+lY5;d-#CeRcB8{@`maslC!*xU4*
zf{s54Gy%Y-xG766Abhq$mE-z;Rxg1j0k{OgOA!5kKA=l8Q~bQIcH0zzE(I<;oTnid
z5N<ig<+>zx{Y{`N0PKdleIXYRZVgpS;}t|O2{ahMwz#djTtK*^sm=cUhfW^?jRvq2
z?xZCb5WZPo8uH<LjSK=616Y6yX3GVH2mSZ^?ryMdAkc6CJK~Nkxq$HK@sil}e+l~u
zGy=eUoUb7l5PsReHgo-9b!P&N0k9ozr!E%|emC~;Yul&s#svBmfDLd16}f;gLz<kD
z)s54WrNsPfqDaaI^HGP@A*oVCT#7`&dc%1GL9$_ki;|QV=A}Ug<%#p6;c&qj#F(GW
z9!Uu?p+ZQ3xZqO=No_^MTajp(6)uG^KbuddMi14LkoQ#xNeNs9p0Ma}Y{4F7GfVH$
z8}n93xCIe!p;LJ1xgXf@I9Y;{RIXcYFw97TrgAGmg5gf!%%jZDW}T$`Fh6a0{vfC7
z47N?hPOvXiXHy^8%8(GiDL%3R+*#P1VCSD^w*L<Ni;?-o2p@VvegN5I<N*CiDi8}K
zVK&GTz)40Cv=fiT(}qK^B5z+te9Oflkiw8<_ywG7NFgfB&!~f>d@)~5n1H16#rZLG
zMb?SeDOkmbT`}SgcPu-Be=)L@w#yK484?5Q2T2Lsjf|1#%`U<gDI_cu7k)}1ovCYj
z9({xdw(jQQmH>B^6X<)>D$-F5zIO3+34x`8v<8a&k<(xbcx2_zNsOe-b<7PQm64Q%
zj)ehia+2cea1CHPk(9NLwE=vOcBpkEuE@CC1=3d{mX(MNObnL-d1RCb)05OfY@xzf
z<%)Bo=#myA9>vHax@aYcxP%@$$fTfXG8(0qwA8WGhgV2aRytPtus<QYf}xX9D;;9X
zveky%MpAqhUmLOwN!hXNv>~}P<y}w~rymf4^BHIuupz0yUIwR=aWcKGJ=@+0PC7|B
zupNxxRg;t>+tCP89!WW|os1xT!nOdn^^wEjg0OEvQ_GkeeoayfumztwAGQLBmhm@q
zYIO)rQuaGA2DWZ&YHR~b0Z$E<l+ii#CMkE!oq$WV4-U4b?T%=Zbyb+la@bXHzu*RB
zTu;XaU;zr0vJDY$qbUXQ5eSmK0CbQD6R}`Il3Fib9|t=DRtbd2UIjYDAM;ljk8(s@
zPR|s)KyWbGTR{hTU>=`pHKy8Bg}R=UaT3Rr!-u2+uN%b59uhj%6Z8C36nOar^Rt(P
zq++lbW%?koSge8-hs7ya30Q)Hm53!OSV>rtf|ZJ;Dp+Y)nt~<4Bnp-klYX+mtF{XK
z4F&#Rl_q9gk}0cYLjI#fEsmU8zq7=pQ!^)bZFZVuReI(cY1ZnDR7Mu`VKCBYbrr2<
z(&`#o&85{0T9wdhDy^=i)pS~w(rON^X47h7deSQJelr4+7b{`PxQeP8L8@zMYKU0A
z-exSOXx7}>+F$56ySTY8@bFw{JIBhE<7i>{|Gs7G=^M;9GBz<a=UQ6x_;&UJr+E&#
UHedS8VGI59P`_BOMMxmyf4|{H(*OVf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp.B77888D4EF3A20F0.idx b/.cache/clangd/index/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp.B77888D4EF3A20F0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..640850ea63abf72b111ae8baf0d5be75a715bc5c
GIT binary patch
literal 4112
zcmZu!2~-o;8qN&E0GVMGg3*XTfC3>BSp$YekWD}WsO+ea=zt)r7z9Bmn+l=`MYMGR
zD^_verSd9T1Rsj>JSb0G>WW$wT&g@3dJ}t@aL$`^^38we`~SQB_y6umSb(pux(<Wk
zEeVjNO0$<>3<g6N{Iaq$ld?f0Rhz+({dhGZD(#L?r*CI}bMf9f$?;cc?vZJ&j!xy9
z_jERShI7{mwD|!egNN_$-q@MJ+!OpD2k|z2#$DESc8;s5ndv1-gM*EE_uGf2J*Re-
z*_hUEDXe*ZmCAPv?l}JGi|5dAfBq?9q#)+U9US#*Jz7}Lg6=e<;s;MNT}S`NR$KI@
zx$EKUoF6J#?>0<U4R|YU^JhggmG{58yVCBQZ`a4Qwp#*Jx|5dZUvEEJ>vDAO&;4tI
zw=T_mR?;-EuUB}gc2mo5$(ypHG#bKopKS?~T^v3ArZyp|;qHeiD?+$$JNI1r?$E+r
zZW}$DO}RDv+2;f&`)&`#e}8wxYfuz)>R!TIBl|=HVtWZ|hj0^n<4KdiA2nWS98!5z
za-(W%;isYJT#ZXFg4OiGysn%u{I#^wcT3)O&3Pu~E4U(Ce<u5Y<bXwKFS^X=5B)Qe
zA94qtMPHvk84HVUsw@msi+Ma)iVpno>%lY5eJgcVm*2C@5Po;<K%ZovbgWo%h}VAN
z#w@9|-<0!irlok}*5C8rx2G-EjI!rNS_>OF&D~EI`1=L=YdjMSnN2bd*Gu4U)ju=L
z@H0F=+^QFS*0)DsR;1HEX}euXQAO-jlH&5Bw*8g(%)g`LPmIO6?U<}*Rio#BF=y@J
z7GIC>MuD3@U*fQF>P>@=gwd`M_HbFr!h=6kAAJVe>Tk|He4?K^wq&8&n<@XD_B0Zg
zL`m6^VJ<&-tIgM1SC+ZdHP!T<&%1u6_Ws()nN{g3DL#Gf*@5a|FC!}0*EnBdik?P#
z`QD6<HatS5R^57>@!()5pSNLGWA9+l*Vpgr`m%^h!wY)*)#E<h7OYoi6`a}{GR&^I
zS-{UI9JD)DTXtCIg=~xYBVJ439`jPY(S*smlQ#Dn?wek><XVtaH@NituDav9zxMQ4
z=9FJc%6FQqHI%r%p;evznP+-uFm;jLZ956}{=He0U8TyN0Ij*^8|`CX1U)sl)7QAu
zY3R}Z=RDc5U};U=@W|SF>71jr`y698d#E*kZxY80CPFifs%A7S(>8h4^h=4@;k4A@
zT*9pI5$e)NddQWL|2!N{PrGf)*1q%8qcfKuxrtT`cTEyKzY(Aof83z^$kaH)O}83v
z<Zb90`{g<r<Z#yZ$N{tWwUw1Kmhgr79Z#}ei7qa!XisvxxL2$Fv!3k4Z#^B}OM>eR
zZ%kXfsBDME;80SJea4rt{QE;MkmKuKH<4?`w)c(&e&Fofc<}x5<+!ZkUv&X_qv37s
z_jk!!1M}9myQHNg%%!fyZ>u*5ez|{l6%lFrA>v3`VbOx38!IlWT@ZEKtP1{p`WAhM
z)Vt@mRHa|H&8>d&eN5cq>+4c(MRXDKI4&QC3i1n_>z&;E8bfnUO7Dke*B{pVb-<{9
zUd&o%?a@~59qY<>so8};nYXZ~>E&i`|KvHV*UdhB`eW6rE|+<GH*1O3g$z;7ZEbHw
zzkGfqsq|Ps`7%#?e?fWK`Om9NOW(Z-EzP*}a@*>1W>rMYhOvgBQ9Ub*?Dx~_x!cOK
za#IsRRTvCbO@Ij7>RwEskqidK$NAa{0P3W8tfr_Z?B*B@5oj2ot#B(H1wh*NV{U8P
zU;YgQIuFnexPy)YAnj;7y6t`8(>el81hfp7aTEY)x0t&LBiVY@1iApw*0{Bf0wC>E
z^>#S<(3eO8T?A-j+*m^akoIF1ZF&-V2}H*N+6K2_DFD*JUU`Wsc>*zkN&#Jj;6;eh
zKQHJKd*kAYwnM&m2{aD4nBis`3V<{fn|G_pu<9{^h60+0^K=ye>68;@8!OwJI|wu!
z(2lsHt^y!k(PMSe-?Huk$yg2OY6PoB)c<)wm(8<%k=UD~2?QDpTr6=*Z3RHOCan4J
zUxTL;33M@_O>h$p1weX7(G!bAt&7J9bSa>@I9EdfkUoNaNbPp)d_$lKfVRLbbQJ*U
zla=aR*Aczz1S$b^DT0?GoPS=><up_LXK%y$&jdOjxCn58mI5Gs{*0BCOX8J}1R4)$
z5iXjn07$omX(Zd`MKB381kkp)t)>DX-Pv-%{@m9tZvu@1v=A5SC;-ykJIX@8?$t^o
zP%)sLaHmNMfb{i_8H4-EpXF#WmH^t7<!Yn=NDl<;np3~qW;cN@1hg~mtg8S>zc^G9
zyXKwmEdq@I^h|uFmI5IC`qs+yHO-nr0*wZ=9d4(o07!p&d2jQKr;AMqbQz$Ha3eJZ
zK$?*$U7B?jrz=Z^IiaFRY7RC>6K)<!l_TPEBnqYj)*A?t7aQ!9q`WXMEjlPyoEr^`
z3)UdUoKSlt<%jtxh2)F#$3sYJ9U@+bM8i>GR|s=LeL6R}t1k^5sYXc3$<@gN<^+~4
zn4`RA={b61-bw?OA>uN62=}R?#=?iv<tRzXX30X}h$J|3vr3T=I1^a&s>}(sPEwX^
zOAf3TlCol3ao|>wlr`I$11CjNd^Vp0RSjwc%ABALB;||w>ci~}cCJZd@hdlhWm*U=
zykX%$0RfxxO+Da3!wo`&$Hz9@2mZx~aWNu*o={J~XmT-u{v;*P73jlif~um%oS-Ho
z<%9WX!zq!}YVqnAdXw_Rc}lHfWM(lUf}_f35m1cGr`r{XxB`iRTMBgwq)RSR^sM|b
zf29G}h}Vn{kvX=#<<IDSwqS)K7f}M7RZifXZeFC57<_j1a1Dh;SU%*#ap|d?0zMn-
z7_(psp#XyY`=)+y<SLEM#6@X0uK@pQ#JU>cLr<7rkbJoqLVuE)jm=i77@1fWMQ=?p
z;$DpS(Tgud#HDmSK@A0qmkTJ}C6CSHz*LZw4cmqT&kEF4P)~AEr9*6WZS|q5k<?7x
znfg#jNXkyvP9N$<i>y^uoZ9FM)@P(;#DaPO&lOmmT#e~|9as(~u!u>@k>zLtQ$|wG
zEN2s_W+WwK2~D7I!t{de{ifcqLwL}@DU@qA{7h2rm^%qm3Ev=EuH4Y6%^^&p>eKiV
zOr2}a<-@Ijs|H=lg&ca5lpE$oz^*z52HKMABQAhg*KpT(csk&GK?dY1Psavgfl6wz
z0TFMY)dETl2$E|*9Tb2C=)wU>sz6*22TuYV56n@1AJ8E|SkU;g;Z}lv<+lRuG#8tz
zR1517@p`&|V3tA8^1Ff#^1wXCdw!|jQjNMEk&B3#xtRb;0X!2RR(_Gtv7VUc_|Cx-
zLNF)XBqSA!#j4V)jl<%UTJc!CQY#TlRBA~uiBe07NtIg3Sh7-UIksG>m4c-xwX(3R
z@fNt%QovOU{<CH#X8l`ktdPUECkq-3MmB9G(`E{7rqgByZOX<?6>8e!oOG85%SLrp
z$#PaEFHcEL&&bTmUXhl>NE`QJET_#R+RUNNRNBm>%~iCyk~R}lB`M%8XUvrbsbH$O
znz{x-YU*ff`RmT{ve0Gv>lsecpUf7z%o4e|dw9+;;JI?0ET{j!udJz@X~rg|X6D~n
fS=-nOX4*SC8#zqj+j$$Xd;@Y(pP1QxNFd{H*A7f)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_softmax_gemm.hpp.2D7D43E1651E0616.idx b/.cache/clangd/index/batched_gemm_softmax_gemm.hpp.2D7D43E1651E0616.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4e073eccda14be895be2c278436e00dfe5387e3e
GIT binary patch
literal 2018
zcmWIYbaT7K&cNWD;#rZKT9U}Zz`!5~#Kk2=naWHI41YKo7%JwR4Nc6xZ6NT~|73ga
znh8FI+fo&r7I<b#=9H|tn_f8a{I(9A%t>=hdj5UaEWER%bGqoQDGj;v=AYYF{{CFi
zmw#vH&#;;wVCDR7tL&PBl}FwS&HPwa&}e+HkY(QU88L~0cV1i(FWBDFJiqE*UhmeN
z?O_?EJ9~atPnUmiCpWKj=2L5Pa~TymE)Pd58NT+l<xZCTU5j_ycl*U{+n(C|Dmd&;
z!k13%d*6TPNPW;;DmV4=w9IC)Jr}wzrKnE5)I90@{)0}__uZ8$l@EwnT*!M)|G?q5
zhaMQ^aUYkqxpyFa%6Hj2Uv_=j@1ybf8dtzfo~Lf1oY$W*uIc`>%j}`hw6jqz;V-gJ
zxlT}7^Qpmi)uTu;uQb;aiXq2hQk`AzsZJ`KK0Q-w;?lzZ7iWqFPh0tPCG)j4-~P?l
zjC*i(&bdX0Yu7q!))qZT{LRO|oqz9Ji46e@+;6`u+m$F)zH>+8YtsUG=^g&Ocb>DY
zyMLfIbNeh?hfY?l^Un^o{*pg?D^M=N|HsFpCiR+$VnK=%g`A3F6;8^kJUG%LVe5X{
zp)tndA;<OBi+pN{O3O1uy_7_zr#7|tAO5Il{){zlOSHgAAJ0`QRsR3ZiE}(u<-5*o
zXY#Y??Gtmh7Yl!yb0)|~=f=@(O|x!=e%i40%C-0FH*C4Z5f^*$(vs!AO=k0M%t@J3
zzUSJ*_iwEW`y<xcYR<izQhQsvwYkMquI{0uqse8Z&VNkv7cKnCP+XatG#{8|+@~kr
zQaY{P%frCM1LOjMurRA|JOd{)BZDY6qcCH}9vinUEQRmcnAmw(g>BevRKWxT2Sjz{
z9)G8`=dFJ8FmS^3E3+!6GjOsnLG+)M*i^9l=d4BCOl-Ui!aiI+LSO=>|KT>N7U!_=
z7$GJOpmelAv<#SlNi!6srWJPr6F8gvjy~sw<_erV!XmaJ4sd4+i`j}H@x*P#k$5V$
zDh_Z53ai<wA@S60)sc8fI!XF4r|<|T>n0l@@jQ~fVLYy13+^?0n+6N;2+P>Ycp#i1
zYb%Syled*e;wjiFAn|l;bvzJ`(zVq^;_2J!Bk>Gu4Ul-Q_O4!#FlS)k5q7h8L*jYZ
zdm!;V?LCorAx0s#ut?w$4mAoz;)NT9Bk>}PBH%p1^_^0s>s}q^;1QOvl|TwpNn1%I
zo|LT=5>MJz8i}W2tAP~4nzouqJS|%-B%ZddHk_vu`y-J-OkpMqFrpPiY!D%=qM%}f
z2sc|LTRTK(_=x%#!u$x+!Ndtp$-ReuFIe-?T7gG6M<gc?7LOr*Lh3Ij+ORSMZQ&H*
zhw0%FR^e3PN7xd}8!HUc0?P=DT;L1<GdJVdRB7Y-(APls%Gt^}AxzS<)pJ6a<Zkco
zf^ct`QJ6W*y)ZRE_kwZ`%v>JfbdmHjnB^^t>#KZM96b(<j(E{{IhZhzXBuyZ;3bGA
zAju_|CD<e6k|dLqka#9Zwg_IbWHORmvS~8h%=QniGM49iq=4Rw*NR8V74hEjfd~T<
zv=a=Gc-{$cxkHb8k3QD!Z(-yS=4Iy<g@pmk_e@;i;s6%7Ji@Umu{to5U^FnuK_vrB
z(NhD~FYNr{-Jr-56_G~-xr(TYJS;eYCP~IB!%TRUsIW`d<gGDKi>R$AQX*BgRfQ)#
z9^oX9BzRW;wQc%AJ?qn(fer_idN79rOQf7c5U02(xi}a|GYL!B{4nxZu(yUyginB9
cP>6vau9iUn!Q?|QL6tQW2*DT(3?c|70HQ0>E&u=k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_softmax_gemm_permute.hpp.A9BA319B35F534EC.idx b/.cache/clangd/index/batched_gemm_softmax_gemm_permute.hpp.A9BA319B35F534EC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0374afc0bbbc96e6ef601ab988ae253bf09544d4
GIT binary patch
literal 2990
zcmYk74OA0X7RToXLP#760s$sKP=V73b0iT71SWh+AOs=^u?omHf)(l(s1gN%Bi2?x
zv<RVyBG!_Wy5i@qpKSrr)vmR!tsL!EtE^jety(K~56X52VR$bm$;>bJ=6^rlo%fb%
zHJYAT1QAo9tt>69$h9U2VmA9JuP9r+m`4!z><OZ>^T%a(*8CuMc+vHbt)(T06W_I|
zcIZD$AAgj4{M$z>WY2E8T<g=8-70xh)8!GC^zh^3S3Yv!9S!NypMLrHOpE8an%}}M
zd}DUIT4yZXsymhTsz`XgGR$MJGte#NswloM>T=i~qt?ssSC@mdniO~i+-p*t-WG6X
zm2$9hsRRZOjFsv3>~>yeGG5zoDnIQj<>n{fT{(IA2`Sg}|A<u&i$)UE|K6_WAJBf3
zd3NNovDfWngxCIZo9fhYuM4$%mg@}*ev<uUiqE*d;#jq)wyS77QFnItgWrbE>=~UH
z3Ey%r+~uOVc1*QttAUj6y52ULGhC8nDpuTvF_U7f;;4M?y_A}m?ZeN8PgIUp`DoMf
zPb!+DH$Tg0+2Wr*`-=|__4lPM*uOgAq0f%hO?&?Zfjh^po&VlwY3=#O+8EEzJtRAp
zSYNuL>gL`umG`M$sb@jbf29+KpVx->{r(=E`11U~Q|4CIfPY3~cFD<e$?Hyus$?(r
zcDE%w&Dl7S^<~U(Vesg;)wgdbe{_qgwZ4Bzy&mqgtWy2${hnBRufMk}vv)4aBs8|-
zbe^9_(q|*(1<%#Xqs8J`7d`XRCFfq*?4XwOdG8J_{+O3{Odc-2@aXP7<r}+pJAOa(
ze%39QINMKGI8}7(YT|3`8j4%4oVfHyrAyV@pRFv>9A0|D+jur~ne(eN(ML~TZI-6A
z9vC*%E2TGi_Z!E~-WaHRGGtY7;)SJR-7gcr<T%@i);Jc7{rQK9+VS6I9c%V#YT&x!
zKxVnIH17dBQ~q~$4Jg+B^9D&cII_sTC>dW?Ob7*5gvV^4Abp<DwP>L<fF$fVhj4y)
z1tGNIO*w3d*)^(D3%{Jj6FCu79f<0{ar#2oPDwW1E$)0~@P0rNZ*Uq;e5XGVLR;%8
zjh-z*aW|Fg1tek5Ii&JaO9&w!rX23RIasydlKNL0-YiEU^^V{j-`8JUfXDN#jt>wy
z|8wM#Yz{5p&;lOQ7tVz!D_m86N5CfZL@C|b|Md}TlJbr5eQPE+Dmf<kt(j`64v6T0
zR8E+rdLg10v1okX?ajKNR5y~EL(d6DB6I1v7<tp)82QjX7zHswaWmbgl#CLi5GDj8
z6{EsPN9%%7nu&BGMtWM0(Gq$IMmbDQ+)SUTTqYNzd?p{G0;T|?4Vn!pQ_DyYB(+hq
z5u<8NHAb5>n=onwRU<6srh}xKK-GkEGpL$z-T|r|h-deGD0%PD!$Dh;lF*U>)L?ho
z9U~9g10zq`6QckoAOba5&d4!RFba$oG7D$OrBZaU`o$(K8`zijjY5G#na~s@mrN%|
zp;ahk3R6&TJ3!U}YBV}s;N1lwh<icSi`)CayARp97(C(YC$VYZ<-UV=!{(4wH#l@d
zIVZCydpxE7)BeLYBqg>L`=dPl-2B2g4U&qqjr2!(8r%$FD9^R?*2bchtpjl#gmCMX
zrgwaEbSJ~-la#a2*&fNy7tXgw@)5!adnBJGOtVMwtzg#*PF%j67hBfk{(;Im2o49K
zhI1>v(CQvoWB3nS);xNi0>uqxf}>E}BsxjKDUwtklNW{JR%xnqsI2Xv+797dS(|et
zn$H!1>qwG{rDNqtLPKjXilgH&il^f-N@vm|QLqdq1EWkP6Qh;PN{q}9V1_`hPLeW%
z-V93-TR?7sML6p%$X@?R&4wh)&_#AEO&~Ub6giq8$OPet+d<rp+uK3e4iU(128kKG
zaSky<6k-cVEV$hQDhot&cCM2=(ex(IZOtUr3O=nsbB<e|#(Jo(HcHri)&zbgkaLDf
z)|ns$aXa|8g92w=JF+(jzUuy9psbc%GpSq}ghs+k?uAicWMC{BkVTP;Fw)AkL1?WL
z<OvvMMP|jK^~{dU#;7u(atRtGV}cQ*ZJ^!;(Of%7ssYpuh`B7;ai2Wu(71PkLnmzH
zYTf;_(o;`7zQiLbwLmRKv3fwz12fF!$Er=9i%9JVI2?iXoZ7%h)6j^2b3Lo(C~%zl
zH%YaDpbe>UD+Z$1hbX6ZU(SCnuy<{keB{N>UR(tv<!9&TiaHx1j_^XgG}swjx$Gt<
zfPK%57MB(qXX;4i0+0X2)W0C?>tjGpo#NAFPlWhC|H9~N#ZF9UYG^hJS(sXwje>MR
zXcuTWagyqT&_2ZMIbV`HDOg^XU!Kigd6cZks%@z8*JOda#Mwpc>PASW1G1NyCJ}@S
R`qdfTx}sY*bnA|8{|`*8@YetU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_xdl_bf16.cpp.ED68A85BB00385BD.idx b/.cache/clangd/index/batched_gemm_xdl_bf16.cpp.ED68A85BB00385BD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9792e0985063adfe8e7feb01826a2a0f78a436ac
GIT binary patch
literal 3390
zcmZuz3pkY98Xi9;W(G4D+Zf|MCd3dwjg;h`OF}}*<wul3sE7(FDRN6uc1KYurIKqR
zQm33FLOY^5iCj`oie1{Gvu4g-W<1aNpJ(3ZTmSohYt6UT_pTXtM>{)HDICr^*fE|P
z5*H|j!{I346B`#3?g<9=(l}gvanlx`s9{s`i+@b?riK=mO+<=V&b~s=wmDQiDP5PZ
znPF&q%_3>0Z{PCv9aS<eA8<!)I_*1?imk0#X<BK8D?E-$K3KutpHRE7w^eqf(n7(>
z&vS<dcbgZ+MY)RkXOHH;Y)@_f_~ENfbnF3*RbEC-hfb!>b;nOlua-WzHmmpLaGTbr
zuFFjZgr1qlai=>Y=eqLKA2(JXO3$lHR&5Iubq){QP;vxI<usHP_@uZLh3*_KD}8mU
zV0T2N+P36BO2#c5nB~v?#6>>$|62HcL#O_ob4^FnE7I;6UYg3NaC)X<Uvi}5SWv8+
zf$I&_wE0YcS#yqgc79Er;=MB+{dtdE+?_Tmmy;e3<ZwF6zKQjv?Clx1>$sov<6D7E
z(4Uv(B(JVB{KSk>FYO*_4L$5Zet4s7Oq`N)yTlhobcw|YY2wk3R!rA-)s#q%i@E!F
zDBbs1wrgZOM6$jufnIO+f~#)9IDh%nmDqPlFVAK+eDu%l*rVo5Z_;szF6(ZzKlOaw
zhEOAw#39Q7DW8>kMLEiwPI`n+liRvwbuYz^$LBe^rP}yAUDk@UX#H~R&8&;~4Zq>`
zXUtqOhkQ==ZpsrsdR&-hMIm|acB~fW&J|Zs?q}}ILmVB;0_&1xyoNdAIz|=_`gYZW
zT(>QXC%&-E`X16{ooY^Be7)shqF%WbyV@tbD#+_`FIrt!h|zKqo~*p3wd!=6f`oL&
zKjT7uO;rA~e>&_)dX;;jQC75ggw^S_UQTiDiIwBV#hlsI$)knq?Oyl}Xa@8QDBWxy
zZrNJ2)r-2x_E~)4UrpI(%EuHG*%z6fl!~cKfv3pMKO;;T!^Zfl^S<vEyhDO3N&afp
z?_EYS*@o=g(|MO>nogt@rg=86n=dC}p{cf>4uhqwFTEGc`s6FD$Ep)wH;li1EU(#j
zf-<M6J~?v0f2Fr^u-MF3^WTkUM30%ISN-f7ICvpl#kHr#ZNg&mO<eqDro3^gZ=2KC
z#Ib9!R%7nH>%EFu^OWh29J`yAUS#Z-8*_eZtx+rMK<B=x0LuaPb$WNnwl&IG-=9pK
z)DK8KGE^JaTQ_l`_mImyEN{po<!VL1j)P&ID^o9xOfk-EteGYm8_*hFAJ5%I%i4Ew
zw_Z^I_Yt?WHYwD*rRxZ)&d#&Hv*COmiSx^G;s^HCl3rz{npnTBzg6Yi_O0Bx(xjC8
zu3rxJEIS@_&e#;Cn%qbq-CT0<B<n^(BO!jCQa-ftmTtGpr`i6)%+O=bwxmJUvoNF0
ztPh+|Rvu-8`|jpRe{)(J^kZZ|o)%%NaBKeIi=5_n4|97})@HU8)r+wUX6VQF&T2f`
zHCXhix?%JB@Q<S%wS?iM$#-3M(pr0r&hDtokf|<xm0NVTj%qTd%-(3nnJP=nc-=XB
zx9Icp`MsPLRohI=*!~aIXVKAjoD1R>G_O6&j;H8YmaHGX^z_|*k%$S$_A=K6@yZsK
zT5JM0=o*5<>FwF`d345!PDD22aF`xmPnHkh_TfJU4>}*j8I^fU;rs!uub{8a2S~HO
zCp0#Hb0{Jru7K9T>yY^XX^U%J=hN@zUnL@&0Ih;ok>LZR?FlJ4?>9ZnCL&t^t%g@q
z-~*)ny4xa~EA3tr5pO`#@H9z2KsqGV&2Z*;`)4Au70}vvZD~G0I?OMzuk^R95hCIS
zXjQx_g%6O97NvGb;J23&5qCiA;&sV<fb_90of?Pb1+6699zdImnM?2i(wWh+m1Z?I
z{Y1nA&{RB?zz0a@TG@&CJzoeSA})Yt;2AP}fOMYwm7Cwj>o6j+9ncDR1xY?Y`h3cJ
zZA`kYiiiXPS`n`($p=W6)(xwVP0w5*BEEoD$E%b10O?y|e{ws}CtzCw08Pi!DSUu*
zQ@xH32m9A|BH{yRL%g9PA0XZAF1eGPxP?GOHUe4;uO-C?NDsVsluBB@znX|_12hZI
zlHvoT-#9*P<aD;x5)nT@Yv46x_^?R7WnmrKMQ@3SC!m>lrW79_Jv;d(*I+b2gNOtH
zS`II_j1Q2G2?>pD1HXo}t3y4q6HhZlNSL{vxivf~*vJTBNduEGv_J&CAOh=#4Gm<x
zZibTs*<}&*_M!|nHjweUT1b|J9$i$y{st;ua|@}=QERv&>~bLDwYZST9Swpa7qh%5
z1^vt%P!xR?-5MSuv;fHE^}isONny&r`FlJ4<BOvLn(hsQTq=dC2zwLy1LX3|A;?9Q
zQ7yPGbO{jgEFvVbMctq%BJI3H+llYzzygLGLr?e`S_tIw{3Dc`A%f;FZqi;Sn+`p1
zQ#X*Uzh2)H9ze>0`n1-!Lk+ywUi#bMYj~U@ge8VMv8?67u?I292nkbXsB1tEkT7kk
zwi2uqRxZBOjU<dgV=$m6Ntg;vg#k~FgsIWg81RTlm<COQ0lOGxfckhA7OIbpu7aBY
zy$nP=BMU^TG*xA|yU@|1ge6WUVOkU|d1yHj#-gy~p~GN>U|ycZ1@of8Xe3<xdVHg}
z8~36e*h4vKIR$7S*gax|CDtclI#eAs*x4jZm#V7<&ya-aQT5cIhv59+5O_BPT)f%D
zJy_iK;v`s3k*TN;tAYi9ugbeHpc+_;fKuR8ur46tT^m9oZB!qM;LZaP@A425MWYFe
zqQBc7h<I0skVp<yf}%-@>)D9;EuNofOqwC|5!@jO!jh{6eAQnfXfgt`!8iv&bA`tx
z2zo(yJc*zmg~vY;bY6IzDT-!^3eESIC_1+|#>Iq0fV&<1!;isY|IG__^5B2Z4jABY
zk%BNv5QYiDki}3GTO;i?m)shdMUIKx9UT%Dx+8p71UGVLR4^`hQH9$j2qOextRM^(
zgu4Y{xFF;T!WcmqEeJ7ga0IW04I$2=V&eE^5|Tuc6j?^vfnu{>n?i7q{Y_3@k+zDn
z+G5RG%XJ1+J#$Ud<*NVhtpZ({p{A~(#boK|u3#G)8KY(@CQAA%t*Hujj&X>smzBMU
G6YjsPM{|b&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_xdl_fp16.cpp.A653F9336CDDD36D.idx b/.cache/clangd/index/batched_gemm_xdl_fp16.cpp.A653F9336CDDD36D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a7a61505e784cbd681862ba1e9b30523edaee1d
GIT binary patch
literal 3390
zcmZuz3pkW%8y+7sm>JB_OonkbCZ<90DTnRkkaG+xO6%}3A*T!?Dy-1aS(0sOBsyXh
zZP=t;`4dr_(qWZ~PJfAX&}!eAe_u1M>!0hI`+DB@zTfk6KQlZRCnqZ<9L_$>B_$?2
zF+>iB!%@LkkeCqZ3mQ(!I9y6)`}%;mQ7g)wpI_X{;#bzZh?QCS=>b~N<=k{lxT@G-
zkGbPXyY#7{ZHsy~H>r5M#N{~jt?f&%w6|wx8)la;_Q_Gave;~U>V?_lPSvFvv!w^$
z&ir_Pt4(=goTpq+{`2B{Jy|`kU%q#U7wphq?zg0U*TJlrfs}WXw#qwK9JqCFw9D{K
ze{H)dad4_Tv8*q4roTA1`%?3++@h8YovsjB_s9^}s@+(Y^~IXffJ~3W{HW0y;e(RW
zt<h(7H)Z@;HE!q3s=FPCmw9{l+w4!SecWG;wCCiW%5F11{%+4Hx0~8)t9JM1g$le(
zJ<p(49fyLKbrjg-7oTrYZ#&d`x9E2dp4&QR9l3kBz`CzyT5c#ad2rmR_qX)V)1?lf
zA8OSU>sOh-Va4eQ2cCBFfApbTJySD=rxr9;Z9owdQe|p3KK}Q`ldb*ds}#rOcmX~d
zzxl{-c{(1h*xHq<*}CjbjGmp=(b|#|f+y+sD)wJ|9aPwxr0cHPZtM|XGjM5b$?a9H
z{3Y6Hk5>jO1uW$pE?{~e^x;oZx&~BDjtj<9id?+19D>|x4P)&(C-NSB^1z=79PPQu
zDx_Fbj+p$Kc`Z;gF~Z<fIc26%V2JTz@O0)~R#Xw<;!+dRlA+=^YK=EuV&}|tYQ7)i
zwO)PS1bf-gRZUg5^JT{#uHTu)satK<91z(Q>eqb>ZEh*Y7zL@<mNpnJFYBT!DDU|r
zk-x!G`>*Yj5xaAnc;!nD#N(q^m#y$~OXQ`U9k-~o{$!i+ynKz*oejeV!GpsZ=Xypj
zZ#=xwkLK-oGo}1Xd;X!iF}k|hF_tg&)Vt#$B^39s(N<cc7KHl0H$0j34-Y#_4$^IY
z=J9;LnYmeES<&&S_I=sq*}fgC{;nfq{47Ub=ljCWd;YV_h8CQ%A8Sr~cyWBBdx62w
zKI)8t-m9lOf|mMQgvm|4xB1CpN;c0jx9MyD@XqR7ZO_5;UXgxDaW-uyRrfMhpLxvs
z+%EX@VwIpiQjgPDJe!v~%@hvlaPk^=)fLTh@0X7huNbpUY0!DIP<eQF*FcIH^HC$-
zy}{@~S>zWhi>Z(2thc!@O*b*$LVLYf<$$YGLkY8E^nHEu(U2yVLCeM^30&cXr)4aw
zakFPb^OHN?8^5zS{nyx)ih_kb!T{Dfmz=h5HB(>pKL4zAuPVR&f$8gDQrfvQ7iPRW
zE(e`RS@}ZNvGF?NNCmN?&wkRx%e%YePTMAY+brSsJGOu6cx`g}>rPb*hg$udH);{h
z<tK|jR8ChJ-+p7e$whvKrPG+6uw^g#f<Ec)j?m4{L0g|(i&W8Z8gUQII^_CA$<a2w
z@X47sy~)E5OsXxD1^kV>w{|2=j0Ic?exZKjm-|=chrR!M&Hdppu|jZg8(YXs8En7G
z-<<yN`nH!lFEK~*_5Wx*Bz&n7vAos4*45@4_51O)ZMh{nUeyzQ{E2>C@>ES2kNpkV
zwUbR3q{f7vL~uAxQqtS!Q*N3h#21IdI0TNW7(n$Af6P9nUh0^3K1@P90c}h$ricO3
zb|?Ff=Kfk-PeQx_txeEY5d)+h54@j<I{Ve1glqydjX+Zr1EkjyGYg)1U(F{Wet^~`
z=+eai>A-=m*p9PK4@t;+Kr;voMKM4+Jj=^`YH!b560!l%Mg${eF+e&ZFl|V<@W4|N
z;st0Of(}&-kdBw7^(qj45RwocpiKxS6fr<LuiyB*^P<vDGA;?wHgYx!Vt{nvYA2b%
z+q0o0!~@V;1T7UYK)Q%`;@tFj3r0eI05qLIR}=%Jk7hnI!j!w3NJt2v)d}i~Vt}-;
zWmIo$a_R&L2>`SnL60H^NH@rRi0MVIfpZ4~T9cqj6$7N(TaAsavHyG|A^w0iCzz{?
z0n!~j#VE71^+Xb~4$y`KLnSdldia@(Qu?Cp%_Jlc&};%*NeqyF<Z|PZbzj#760#A{
z`UHIyG0f2;Y^>Ml@CXU<0W^!iQW68Cr>SA1FLM+9lyE_S=F+)(Vu19gSC0x!p9kxc
zkWfIY5!B?x0O^EqzMu>I8Zzo1530Vnu}6lC*>G&^;Y-0pMu_tgn2e#NGHA67ycU?y
zKqks&I607|qd9XbnAbohN@^*U4Qdakg4vBH&dX{trp3@=!VDo}+6-+boR5s@GIW_R
z@5z`xL!Sxn6!wBOMOKhnlZS@FH9eLU#nHL@3&7zumuuR?T!2;qV?{oZjAc<-3m*OZ
zD$p=MByx*HgfdY>xB#>b5Q#h^C2~Z)peQ=~s6yAik4L~R%&pCRVK?*?7%MWA)Yv^T
zXz|?TCc8QGa{R5lz-aCot}o1r%pI-Sog<Ggin{%{n_zeDJ4dE<|9%b7O{380@Y3P4
z%Mj-obD#X>>bxrn83-BE)6&z2Rw83YG$Rdo0q||WK1J4*+NT-193BRG7>Gm$mWXs1
zI!t&=(8;pIc@8FHhE&4^&}w9iO=T~D&VpM4^NOr3nHLR1W8vbbQ!e4XVvae1L#rvP
z(V;WobITFuS)Pm;(~NatW|J`!nu#ua3o^!`ade?M;QZj$M3n(9-eJiL!#f^(1(s81
zsdM3}-~qr^MTG*Yfmetq6;1`u1w^89Atf?Gxljai9Ee0kLrN5nrp}4}%~K!}l@2M9
z8ma+BuM|$_BQ^~p129+&bC~!rTfkURU6C4FD}y%9r5?OU1>(HY0#5Wz2F*a=BA{J>
zpoP-yDg>>TZofj%*V65;2>Q2ld%rAtKvrtLFS6*&TpO1V9t}!A_=lf>39gHRC=vYc
zvl#|BT&yIFlY|kHFkBLbNy06XFj^7{Bq4t;l*LRnLhvIp4=_rCU~7DM1b=hnmgtz+
zsJJlP);Sd}QWC~U!URbeFA1@juxQa4T*KXE<?sY~1w|5BiK3$HOm$deL?t?_E>v5f
z&RA}3YxmuXm8(o?92*0xMLPelm#)dw($&*9WU-A+7Mqzbu|St;TWWBZ+SBMxE{TZa
L_p8^+xZ(Z*TF-Qz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_xdl_fp32.cpp.B7D138F072E276A3.idx b/.cache/clangd/index/batched_gemm_xdl_fp32.cpp.B7D138F072E276A3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..85ac05a844349a423034c1d3b4cb3024ecd12d6d
GIT binary patch
literal 3312
zcmZuz3pkY98XiCMXUw<^4TdpBjJt;VT}Dc!Fr*lzl!_Tt801py>_RA;$|Xf1_e-iR
z5hdC+mDEI~6FEC&tLWHWN%uNy=Imw0^PG7e@AIwyeb<_At^a-3;Bs7D$4EGwOE4!v
z5E2)pfWzU);4d~VW}6QfELFkb5{f(4`bUk}sXaXLvM)8XxawtOiqoe*ZI21vn{HMv
z&Ns}oUVhRkX|jK(Y4?^U)%gG5GM4qY^&}O$xR|6Fr4?HGW)QAgvLYuUKD<h020MJR
z&|5h9WoKox^XnlRp_4A^>Z`<$y!_Ym+@~+TxJTFV-oI7Hs|tO7j_#y=OQKiz)pl(k
zJiaor-PFBLKYyEJ$9PA!{)yLpLkm5I<ad)$W$W^EvlPvPgcr|J{Xz+LA;;=l@(;%N
z`rSBNlp3evx3}UPqo9l%OHn9p{5e_T(PMdTyd<NmA??z<>bIE<o_FbPW$B{*y!qa9
zSDv=D>(mPD@62&LlwZ)QeW_nGkayjS>$!?nN9r2Pq4rcwDfFl8xIgA9x}5ZRs$d!K
zLyaM!ezEl{W|TqYy=OzAdwtceSyhkX$vJ1s*4rZH#Nxy>eDrn8iMHOBGQyYw*WXv?
z@|sx|&&EOsZNfy|Hv5MHhLc`t&Ec-tzmgsu%f9gUlN?bzV})+VT(9V=d+jcVA1w9=
zwV^K<b`Dhbx3(zCq4^y24V_RE-cvWPjvY(L<9Meo+w56m6zO#J-Tt9Z@%YmLBi(nI
zxoVFzj+<Xhxf!4vx7DzrP_6xJtP!6pP|&0e;KT9|4yP)}K3UaogpD_|adNkEZT`2w
zd#!fXI}`i<Yr5*5Ek`OI+wb1Rs$0_A;vd!&?AO(2+iYLRr{?(IoY!dNa72hIs$|}X
z3tf-VpCn8KrtfRw7TO$$CPb{%T;%5&$K7>iY<@BO(}Ls|g-cx@t{*fEyg#V(d-up6
z8;UmgQGAx)O(^`*k#)3gR8yN(!CXylc-s<mSZ&3(2uJGhe6;@i`oDgx3kj|voiJ=3
z_6p7JwPEER$*Z1h)gsss>XTBw_)u?%6te1arjDRn_@AEy_?}WJ>3bbNTCrE(H~UXP
z_JBUYpIQ5Bci{CL^4;6xaX0R&GVarEy>|}(nohsy%vrm%+CM)2(eRoAZ;@h{GOPPd
zNkb35gyTi@-(EVIAv$+dyJN_GN%nVs!Q|_KL4^~O4o|Lrf4sG|@yT2o$!)>#)6zRZ
z*|x)WEBZbbm0#X-=K??PN^FXMLtuDf;Ocp4>CfKk9bMHtLCUwHUVQaic{ugJ`3fHk
zD&POQKzKf;W?efg-PVAWXL4>uX+DYllk@Vk@SU@6(-Qq(+BObcnTkJo`uv2?qOQ%g
z4keSN6$8WBUS4>w{_;dhUEK&LeCQ0aseh01p42s&kzrl?=$1Av!REYyMJcXX^^QKh
z+f<Lud3>p_I<9fUd-82J`<k19hE2&OCOgt6l}!#7?|8dgq30Sip}uOfZ?kaJY-pe-
zqV~>dbq{|3mh)@0M>lBgDP@w59K7e$?v+%g_gqEvKybQ@J(>_T?^cS*W6y}V8~wi~
zHN0-Uz37K2kM`B!<d5Wx4$qL2RTKI?>ASBjaQ`ev(RmZD`BO2_ewBxD{CieFfpHKn
zHc`OKLU1^X`1m(3COvhD$Z8ypXO5byO90$Gk`L=+VteP5dl3;?320N)R80bqb~@Qx
zy6<9sJrVH%v>vLbDgj8lDW&8*_qle6i1-0oAJx~C0HnP#^<u9yeT!7aZ348Vrlo-d
zARTZ|7}<Ho^)V4y3utXrn;-#5hopL2PyW{ZhKQ^OG!td2NC47X19tUSY8`k+M7#k_
zN9kk<Kss8UB2q**RT2>{pv_P-H3>j^f3I1KyJ^7{5-twVYz4NW1R$Ne#8ocf!4Do0
z@d7jrrKw5)(s|t4->1e}`9x$RpeZPYAOT31raU+1s|cHjNFboKP%VN4AYIuyVlX-}
zSxZFx0nI=eY7&5SqrwM)$o3}K_hvxrpgLp;K)R#N%#6*y^^u6I12hX|X-fdoom@f~
zYu8#OBC-n5hNz*k1Ry>5oTHp%`b#qr2>`S)YOE{)NDp!Dw6lAJ=ZVM$KpUV2suGx{
zpPKMR#zjwwh%caxP$OjtK>E}8Q109pfpdsR5TG?sjad?aG%hA2G`0Zz+EMF=@2kJO
zlPO2yu`Spx@LfQwB1$tXO5$M!a#*<>+&Hu>kcs^YR|m2rEMr;)oeNZA+sddMF&DTL
z^e+&J%_}3~Vmv7Fvd@dsw93u_-Rd~#xWLx~%?;*?oh>uBMh-hW-3K%|kck~Gl^KzZ
zW)J;yp3v$*B=)&fgwZe~*atK`5Q*I`BU+AmLs3LpsiH9J<8ja|i_KaMAHy_&x#B3u
z%*~X;@~4~K;kis?vChsL%(h%=xf*UU<=3{fD^G_nh#&h|ZiJ8Fdy`X|k)5nrf0XaP
z8k3BWcnm$p9O!luk4a(bz@5T?friBCl4*#AIl!)A%uuBn@gnhPR2mJgK;r39^=R;Q
zlX!G0od%DJ#M7th)8Jbs@eHU2G-z=cQhB8r;Uw`4$%eCGiby<Tvhi#f8n_kEv^cdg
zO$TF<u<26??RalNg)2BL4HXSdm<ebb1*I8jCh<%urVMCl63>ic#(*!2#51RuGhiy<
z`rx$1p@(1IiE)GR%PYpg*J&}etl+NT20*LgI{>OdNR%)#Tng?Bh{QL6jK~<Xf+Bcc
zKqS5oWJJ+e;<V@=BLk86R*(^CU^-AVu6XJY;@Bt-k~UM@8YTmtkfPFz+XA%uQw~c;
z;A}9?L9kreaT$V@%Z|qp>~GoeHw61GJI<EJ4#>-__eCE2Iz0xrQv|rS!2jei{Mg&#
zVw)KL_X&js4kwV7k<xODv<#M(d}$dbEyJZ{jI<1yF6DVKXXalg80&9V+a41g6}%-R
zbZc06gkW1F9~V8X!fln75z=zIw2YFLd_i!8xJ8eU74iyrbe1B4NK#f)RdFXTTWU;J
za#z>Vn5|89U@vf5xX5|&T#ALGp`EGz|9h&bOVeW*%rRn`n3-F$tZnAo+S4%|t9dRI
RO;=7VvV6@FH#tw-e*urDQNsWL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batched_gemm_xdl_int8.cpp.AEAA0DF5DDD6BDB1.idx b/.cache/clangd/index/batched_gemm_xdl_int8.cpp.AEAA0DF5DDD6BDB1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c078cfe13a195de6a65ebd795d7698fc5ac0e7c2
GIT binary patch
literal 3270
zcmZuy2UL^E8Vx@XAPFrokw72>LMR%_uatmFiBbiD#V86qLjdV5tYCR6c8Ur{J)$lG
zK@b;diV6Z26m&%a^;v+;fro&|@(>#qR^KGgA@Q8|pObsOnR{pEo0;!Uf_*$a+Xxs8
zH_9g~DLR9%g27;P!B3cx9v1)_%d{|<tkbP)?j;X9>HMqveU<y70JP{mhLbhvz{)Eu
zFBH3p%yMj3#9z%Ssy$Q2to5O@;?!)J4Mw3K*5>5~sQakfg;o$qdqN`yUcPwUPxZT-
z`X<x4dFn%oSKN+)R%c=uy{q%FSao>(&ga6x*KfU3&qbzxr>V}g2YnyQq|I*AKl`|I
ze_G$EfaGQi@5gjeoOA2s(!+G4sYen+UkS3EfYr1u-D{amIixZAau+W~!y)=)Rio%o
z`Y*hW`l4MKTD;w57mbgV1PjS3ErRb~Gp#rmyJ|A^jjV}pZymj5USsXPm6lCGO>-Z(
zH90jtef}w8XZTl-*GmF;lE^m^sKAfn=)+!`R)W$JxG_$mK2n%kpO4tU?)#Sy{q{Mx
zU;BFzgR?jE`AIxH#rIktU@MIetoQB4n-~k+8{<_%C{FLx@q@XcqiJUn+(|te<qZ{!
zxBsriyW4x8n7aR+LuyKOzgw2`5`W9XW4Avy*i(fsuA3B1xSep=5`J>fTb#MyqAL4n
zSLBjSw|3mrZ`i;?8Jsd&W!ETUpV%zKcdTccDJoIPX*<5sXCQ0Oux{)01a4hZhp5!g
zgwbH`usTDzu*u<7k0;NLu}LuI-BI4BzR;--|AeV}<cN2BGKOB=A;bkoW)-Z<kG;}g
zQe#UQUR7hA<H$T7&t7q@cJ?DLaCdRS#V1ksst;8r#HQGGqQ(~sqBr=YF5IuFLn9s>
z@46Jbq-}ciCQdv0tgS$Qi1_^U!fn;ZUzL9h>)q%ie6tGkz;17Tb>k(k8|97m5u(oQ
zE9jMKtFqjZY1<E5=i3Bzvuz`m`<*l44f(ZOFiw1Wk%8XYNC^Jgc>C}AKklxiF7t|Q
zKh>poMi61bHuD=#c}R>T27g9Ah|&Ba4N4GWQd9F`$J^`8<!cR{EC0OzFjUR8La*tR
zUcD=+!B;Y4wwuiLbnCWxOy#=n^cGs&|9)X|{+3tYnBf}7A_+AvE$4~3RQC%*W}n5v
zUXFNCRUBi%Ly@Fl^GBNanStej`W@$vNZ9L74T_A$)?dADFiH8?zjyagtT}C<o|9kC
z8rT>2#bMF=nOaWfN;J#LHi4YT*3S3cS64u58-AO!shHnj#=R15m%$eI@)pONrQNT)
zaMJSWq>%4_H}7TdLQy4?+qNTU%a_2}7_P%B?$l+x%kySQ!taw?j&@H!vy`mbvMT7W
z<c^ApT@}MUEx3>;_2u>{Ym{meuJ3cbd49QL<1knI_z4C|wo5$YCcYE$TY2RrHUI2n
z$#}5ybN1&0BjWzxs5;%y8#N>sk1x3-uK<@USH%3^#FmgpwU&KD@8+c>m|uQ8%F{So
zKSAoZ%71qu&$XL~_3Ml+9$+o&Ih3vz^ZGci=d`e85q0E-=sz3>&hUg$+v=H-mcy>O
zdYwjx&byZIr#y<R#%H}kO6k+zg0{Qu?YMUGj;qr1d!i#gzs;(Ntyxcno0B3F5e#O*
zufI->zV|c0BOw?J-vVnvlmU2mWS`B<=4)*qy^HWj0H9e|mW~V{z2tm%@&0B}6&?u!
zv;o#YTLzH!!etkX1l>N2M|glX#2Qj$0O_#a_T;u2&p+|V8bDL96b%_bI(k>2?fcw&
zQ+Q+@piQx+S~7riY}nRk;`s+(;*mf=)37vM89+K!ncSt0jS%CJU_hH=&2?ljM<=RV
zKmGdjh9)K)&{h;Hh72Hmpxd(1+v3<w0!9dEM-@kP89=(w-BT&-$+t*6;tyzjtiHAk
zAYGh2!V+k;H{g+f0Gf;?Ysdi7;w!_9v3Kt)@kl73jj%>KGJtfQ%7>&b`@3M@5rEdi
z>gmb=(ydo5EjfZeX7I>bK(n!IJsCi{Em$MoX6qUp9$5uw29}{I14#Fc_-Jmk*m4Pv
zgaMj~WopU*(h{Es*EkQ`FXNH*fTm;V+A^4<2h9attfE0Y@(Z9%uqK)^fb^#cNul-V
z1`|B80nj8YNlgZjPLGZe7Jz>;<@BvCoc2*yB?6yDr5VGEq*e_*B91@E0o{%Z9J%lc
zV1**MpDL5UM~^9?rAqKPV5I_?v`pceK(-CtH>ZLH3sll-RZuyjTsRdhTOg9wtb!;Q
zjf5h9$0I2ec76fKR?kU~3oi}UG#D!_X!%$ZU6Xl|AIBH%@7|hSUlFN7;OkTM>9Cpz
zd;_Wh9i~X&8&VDFFc$)!PNmb~^+0QYRiwU<uY%H0Cd?3?D-cQDp&(j;20~Hdo?`X(
zyqQyA8*C0c1a?FBfU#2RD2&ZfLPc}Onfk*;Kqj@2f~-Obt)G)^@>|}uV68(SSj5W1
zDg=(o{^!*_HwTBBq}@ELaM%qmUJ3V;=knBUlpeU1z7rwvjSP)Upos{4Q?jW(JP>%4
zAR(zM6%w*Ro#5u6J(Y1k*^<C#=rZO({}A|0UFJM!7q}XjSL#=Vd85!|IB#{<HPyhR
zGEcA{k`{>q-2h8N1^1JW349irWem%fz&9tG8^dEK@GZy|#<23?{NO;Oc880%Ee?)S
zT~RgxmLr>y*>G2I10YqY`=J^b0f*|sso=hVNcsjSh*&5aieR+?k@O`{5T&A<=R`kh
z8i=Itfr5yH>O;|ldi7z%xlU>piW$WgngpJZI_{@;0;Kv~3Ehdn(V$&`poNO<5(F((
zY)>HQUyAKn1pQmFov(}@R92Ypi!%CkuALs8$d~3VOcx06OM_%7{O^l~0S1#O4`bzF
zj64*`!$^4;ClAx)VZ1y{l7~_9Fhw42l!pm(p)#M{+*G*nlR8ExH7zAs5EB)>F)ltK
zF*Yd@lRBru2$G@_rQ?00S1PNhV%5|&@B~dAZ7pxz<sK|uoHubkX`UX{iQ}?l=`vS0
zYw`kTQwIzB|Ldg~&<u?kCQLJPORI%8wswo`9gP<2vr#UY;^~u)tXS>trR0bCFP9NP
Ao&W#<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_backward.hpp.7322FEA177E6E308.idx b/.cache/clangd/index/batchnorm_backward.hpp.7322FEA177E6E308.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3169f340a348a4d8808214fc743bdf1013d72c1d
GIT binary patch
literal 2940
zcmYjT2~ZPP7~VvZ5Lo1Vu&g1ANl*+D0zoJyBDOXXa)?xu4FUy{AXqU6hj_G9!B*=%
z;|yb^YOSbvYZ<3FYDW|qXltuX)p}FUdbQS8MXT)s-F+`JWcJ(dfB*OW|NGwyc?P|{
zIuOI&aT@9>UDYLi7>0#1e^u4KiXc9QNyQje*L-YxrMp)ha%O6`D9e=d$IC-9`;53<
zd8b5aZRTa}^PTN!*6_<0Eyi(Au6t*m(-&Ili=*sk`k#EU`RRnae(v|L*6geq(_*}3
z+0b`h)$(GRtoIB)`kM$z*NwW`R#S3O%gHBgnq?go`)6)mePVy{qsp@B89#14KrKCX
zwmjl^e@$dt$sE}t>v8$6PkyWD7rtri?LE0DX~(sj-T8qRmh9SFmbmj$PfaA=za;O|
zbIOzJ^}{UjhKeI`U$!lN_{vXmW#2e`->q3c6j2X@o2kfjiEmH;FOSBD76;#m^6L*W
zy>45R5gC6qhdz34T3wJYbZL{zyQ-qwEBowh{6*KimglE8|1;6}Aj{X>v3<#d8{tO+
z4)L~)xm>s=ZsS5jeUkf0XMm|^<=v7G&pW#JXo`P5IPKAr1y>^{HGGt~tk>nq@9t?j
zS6{0u{CT^+y67m~u$Uk7`3qg<nvR9^4&H;b;$y-+*FP>yEp8A#{BZAzXZ8!1x@~2n
zeA^n=nm~3+;28R#B7g5s>a}b4?>*4C`|X54!|FMU^`(}^Tk|)_`T{3jbLCGwF{z8c
zu3gn_k^fn|P&YdLvDEpuYh%-VPd|01Kq378$fikxm|2J1{kw$E=Utq0VRO=^v)i-l
z*G;aNV$$Df?y|IA*}6hoSi0V~t?f?M^CQ}44KF4<uK#z_cv`tm_$_bIII0J$nqOJk
z$4tj~U(e;)#!GXB{J=0Vp%$s*2EQ<WhGRatymC{kc%?d!9}tcax&U3+;1^5NmA^L?
z)ZFOBvEgAu3y5$*cm^gG@-Z@$NATwDnCUz%8G9>?k2CA1giMhPezEJ9yV~1^PuOt~
z$AZ}PF@l)!m{`CYTEDT&X$ZJK(v4%mYycr37y;3+p@4QtE4{My-Xt6gVFP3W857{|
zHxz*RTxC_mm~sMMZn@TA9QU3-PJm2c;H(p5CTZYw2r7#*a<&LXrozBU5V_Xe61F0o
zD6keFsN7nPplWM1heG!+Cu{cIKk3H^j{s6A43a<!L`J1C5hgdQ%*iYoB{;lZ@<GgW
zK2Ai5qQFq2F$^cHA}fd3jpPLkZs4kpS6lY<X=B+9<Od9H92<ob28CfPD=^&>Hu~N=
zoj`yS5#oqo=qXwp9SmC|6-$F*Yvf{iFl>z>$Pmt^W7qS=gVVrT!HK+(yvb}(tu}Fk
zv##PTrl^skkz5%#F;_H~Lu@cp)WGbp?i-3p{qAH<Ez?nI1Vuqll$uI~$f0zo*bF$~
zQn^xC)c95X0`tz!u2`Hf0;80D&7c&c79tZcA)HBL(m~D)%m`;rGf#w^1y~R+SDVXK
zjT1ItLpWQC4bh_kjc~MvM)d5!j&SxgJEE5l@)52;TY%^}fCJ$i8V8co37iP$)Ho5&
z1zZT{N_S<!`pQ5V!j)^w5j{6>Bb+<MjnwA>9)$B~JcynbcoEK<=0)oB0UyFuYpXdu
zcC(pL8@M>&sLc+U{Vr>monR(t;PQY$MIl98rj2*gVl3F>Ze|`OA_+HvI8iDp<xoqr
zxGAZzvsK6(Wq&}jgN_rVKS{DO%uxDMB<l<(6#fbl9+YHRaw<H*I-ncP9#J-iu{m&s
zL7TN9wVch4o2xo+Cah^>Y=()#IAb`GBg)~>ZMEPpQE=3Xf%6O^C}=YpL{reF6iAV?
zBL{MdwZL2t(FQr&zn6!bPN=lsea*v%ejEO`|2o8a=&JkF9U0S%SRjWgAP&SMNC^@V
zlnT-jlmT=Qkt0YcbVtULas<VZ@d#3qi3my}QxKF&rXwhW)Imgzq!e&ts8Q4y1SL=^
z1f@|K2+~rS5E-+LMz%LNVM;I|w~Q&rWQLr1hM9)QqPA!tvMFuIeQ(RQAvx&;8p%oL
z&`3_Z%Fg{V;DmjK9Z|?v=OYRZgTn;%oC!`Or!&Wi<aAkGb{JWvF5|Mixh>u@wXUdr
zI5Wp16jC@KidcmlL2-(B1Su7X2uf0<AShLlj-U*M4kEYGjkLs_?M7PSN$?={Jvko4
zzE|Z%TH>AIMOxxh`?zA5{{`lh3`}p8&rvmH;DJcqWT6Y6hTjxKM}WvEEMnm0(1$xs
Q7#7LC!>}m$7!4o)1MijjlK=n!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_backward_nhwc.cpp.7637AB6D79637171.idx b/.cache/clangd/index/batchnorm_backward_nhwc.cpp.7637AB6D79637171.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b3e27787241414b2fc01c9d6e9e8ffcfd810c8d
GIT binary patch
literal 11414
zcmd6tcUV-%`p0+990}#X(m@cAD$*7Z5e-Gkt}N1&gB6fwjo3g%G|_0RAc~5bprQy0
zs1##EK?S^`;AMjmDF#8n77=NdXpHwPzs$~f{XNfp@*h5r!};uY&e>`2JMX+_r;0_Q
zR2v@8J64<+H#Z@Q=J9xH-2eE56$>YDV8ZZtiHFLjJy_D=tom3ou+Y;w@NUAmiH7Zq
ze;a#QoKohriw;ONt5DC|vrq8xy`kTVLoS}*&&wFLIC;$RAIHZ(N*Vq%CH}zE^u0e#
zF04*AX<g^CCuM!<d+)H;h(Px_$+r{Mt-7B0>4zh={D_0KL8*M*`6s_hIcYKPVyk+`
zn~v9=A6~hv`;=t+WKQ>q_7(0|bT=8_IN20-$79ZS84tfLQ9CfMt<>yY!peH3c{86c
zs;(Y+o|%_Ec1b{lul1=od*hITC1Jahcih-hnqxKl+=;tC7UwS-TjurY^-t+D6Q5sP
ztF(GlN^4Dih2UYqZPB4yi}v>W+2ty3A9(o3Kv-Jq&AW{zY8lU_lz#Z_k69MxF#(<5
z=Z~2G;vKcGIIY9&Y`M#WpX;xzHhS1kpAMPO*0bf@xw>4>PMa|IsjEhc`-_~Xq=$vr
zJy>LaJg+x9_J<jJ^Ro}AB}J)T))u$zTzx*7|6~D7sk=w-RQ>i;R)5qeuY_$UoVPmh
ziw(n0M5Ib4YF~JKcfyr3;rfY7R}1n|%p)f6O6vM?{od8<%*)c6Eq%3opC>z~$MDz;
z)!m6l^z)s9N}SI+2D|Pyz45I5!m;m{stupIhu1r4aYMG#vXB{De|S?WT2$t-^?Llm
ze1po6k=Hhw9|zUEEv*XCFKpZIm5SEzY(o=VyK{~u7`}cw|D@}LBL$19Hn+}Svo<gH
zO!%ye7uMO91=PJRT77n8e0#F9<aWRFt&OWTI`bARn!j*mW^cH2W~o}<5~~D%N!Z@7
zzEJ;*=H?5xSwCjjqt|AgRkZE&pFdX~1ibiq%~yt7p9MUc8)BhxMrhv>KS$6~cO=_t
z@2G>m7kjQwy5XYdktr$Y&Di1OKl$0o**A64Lnd!qb?FbUeDn9}MOGmhiOX+XiYeW0
z-c_l)=*cX8p7UQO-)stM_|)Z8`Lamp>~3JN)82ndNitj9s@q!B5a8JCJ2I=QleIDH
z+i)WP)6aPaQeyd|(xQd>3!-Yy>*bw@8h+WhKd}DKhPvC&ENAOxq@S3wBVf;c+jAD*
zeEU4v`~9DX!dq^c2h=@`dOF>8^;oZme<_#WDi#W6zV%mI{@~#LDLV?5p3_*cJWRV@
z*;Eu@uXSYnh|Kv-8S|f|f4AlF?mTn%<9+KL_U+%$UorL`|H<hE&HleuTUIk!dbI-9
z)2w&bQJ2Ow3ja=htnjmG%QwADCUv#!-R38W`aW?^>DwRpr)Te{GT!d|$Sf?(TlL+C
zn^*mGvV3EjW~L>2REE!K$b2f+j15sTc-JKg{ZY#=ihtQExU(aM5$Ieg`7`pvfer=t
z;jwbg4mV#I+1J@_?ul&cojpw0sGoj%pY{74g;$Li&R_NV<N)KnZ}{4Vo08b%ZK3<`
z>&!CSA2{Ytq1V7#m!esEHA!B&i?1kGe&P#_Bdg`MJiey=F*@+m*)2afE}k$Y=3g00
zGPjTY%W*<slgB5A*v0W{;-Z@XkEihqo4ILL^!Tr+03MIklGD-}EI1!C_)m}axH>s0
z^tcMOFp$TZDwt{x7D)E#xJju5&2vzpVmR4E-b7=tK=RDB|GpBhtFlIgisWP?1)~vz
z1(M^GJlFLpCqGr8rf{;cys_F~f#mp}(C;iXv$|ENshn(}V4yo#AUUbz>e?!ms9I%S
z2-os-sF@D+|NeujtXC1(=Ex<CRHf!~PZ%pdR%5V0@@7fP*6TriBb0fu+!Iu_RE-A<
zByW?9KJROpUe565a`FL4AAs_I|3OtU<r;rK7;CPiN-f~NV3firjllxRHO0!;w=JzL
zW_ZEe6XJ%&=?)f1e!BmN(*7FoQiY1<?$+1UHybRF{B9NFnNYK*T7{a&$-45onu7(B
z|Jw1SdCi(Xl~t%koNT6GW;9qJnYUu@ym&8j4(D9+tQ|9&4;eNEnxsH$S*zHc&}1jc
zsn8@9GkE16II-O4xgyP;0n8bojBcda0^nQ##$mn(Ebf6(nCrpldSHdQ0oXSH2h2^t
zzKP^!;MfetV7&#5ZXxe)1-7lg4(n~erVZF4h7L-3nRlX76d5)RSfl}>hkJn49w0y!
z8TKBqx<_(7u&c*k%&-k$LId!?^_zfe6Y#{`3??=MA?6lf*FxUk3Y=TP*H~`@PHkX3
zV!2$0?&vMTG@4-#1H;2)LyiKAqj*FaHU}^{zyRIIu$O`1Wn5pb=<SU4{BKSV`Kghs
zz4oHZsz{My*lJ)}4V-X~YJg=8F*FZ%?4CcMZqE&kx2CrqvRsB0d5DCV0yF~%E!aKS
z2Welej8Ba8`4A$<u;EJKYUpEzjkJ$+#uQ~AMNqVTG(j<5F}}EFtXHgz_`3JJ?0<du
z4UjH<j{-0ebwf4qnsHp$CR46|H-YI*vVT>;vI_T4wZt%POuDaxkxsqhVRCAZ0`;T7
z96ga?vw?Ot7>)TdP`pgmuK@fCV1o5ZpjnBHPudXvO;WD{npLD;4fxfhUIR31NWB*D
zYe`)KG$mL^8wREL*+#*23(6mcI>>2iZ~T5w%!cmSgCnL@0JKqib#ClPF?qFS4L8xf
zU{WveM*_q4fk}NNzXq1Cv6VCIJ7D<^vu;DN{GTT;Iu3aoKzpd7;^5wi29RN6<YP4O
zdOHbyPr?v1ehgax#RZVXd<lAAf`OQ?L7!_d2=i?yzD?Hu1^WF0L$Tfrh0QP+^CKvF
z1g9X@cYd`x!+h6||9F=$Qu%)TcKJJxxcT%k^0C4F5gUmK3N#8NC_*Q~4AnHgykbGu
z2G=b#&HCv17@&`7R;(u`C{Qnupa}H{EmYIQ<6KFf%P)#Uej%83mQ*P^MX>*Pj7L{K
zoaFp2;TGp*dVzi~7>RABA6WF`4bSFo?qvDsLYptF&Mqi+lH;k>$((QsXq^H&=m~<L
z<G*V!w?6%k+vM=%NP&Ljzm|USRZea<6dTvf(v)XdeN}yJ6y$K8e4NM)%^AZkvtA|(
z2pBd=J81-J62qRO=yQ}J=4+JvHA)3@IVE3Cas?${f!TRSEzj=$n&pEYMxhtSEm9Ao
z&_{AVQ0OQ54N!Q4*~ROC*U#^b`CoK#f?;p<H*!(EDl}`QZl#H$K#q*9OfgxjTN9d%
zx(z|L>b3;gsoN1GP!|wnuWnC}gSrDj1(bRLrHKZHW+jxm<a4%?P&TCQEum!L3(fvT
zss2T&Bi-ZCrP@<{ecpg!n<&dB$_)vgzKO%9+1M=L)?XfQ$s+@#bT}a$dZJj|YiX{n
zp-0Kkk4wCct~cpDa3qB5>ONq#54fYQq7x10tfj{p?5v~ZqX}ny0{WgHo%I<gK0`X+
zi_rTb>8!6ppR1&^z6Hg%$odlKCn2466BIU)&bkeX+OV@0m2U1{+%+$Au=NW203WqJ
zVBDUi+O{dXIpIBUe-Fem;pQ(>>TP>&aZC9ErE-DNM?#2#uq<w_YASa}h;E1>vQIWp
zeblow*I}Gn&{xpsV<TQ>woDe+b7sp=(n3#Se~a+D|C7JRkaxo~J{)v-a%Gd#pjnf_
z96GLYb1yxs;<=BKkJ3jY!<hi|&4aVi#>ucnP+SDVFkgn=m*F(b<<O@bPRCpW#WiI8
zyU_1049EI!Q1~01fw=>UI^ayiVXhe(S;L+!`NAnhBs4lc^N%Y1qFZwk4=s%>EA8N{
zw+afXpfh?D!`4DUE#}CG&{4;qmX96WW*mCSV0Hjl4B(h_;fnZGLGMmZRFEF1=zG{>
zGHgAd>q%|^bOXsvfNmnW8PLrnw*b0@<W@kplH3O9Hj*z;^aV<GGF?11r+H*AT|VU3
z;Nrc}qkJCut>%XM4A4G9hT8?eqyV3FQI14TaWFT6*JPx8Byq^eg8VF~izoJHsQEKA
zz<dL0-XOUiYSv?3p6D(aQGNF#q}e2wB-wG5X1kzk7j#39q1k>Y?1v)6%g;9jyfNH#
zc*u`-eE!|iJ=-F`<(!=FNMBjxi4shly-Tx8cx;1JgN?Ag=Rn^a(tb}v@oCaNFF@}L
za4H%UhAo3WW#rya1;tfl{aWZ(i|v_V8=<g~+zVQvs1@%83F(KYSHCK{@P)l6RzB)X
zqU-+6g*jb7p$ic4(MzE4lH_in&`t6ypzsPa+Psw%|0{jQu=&s@A5O+yD~7^i=#Tjd
z^tgi81H)Fpi4|}X*6%>!9kTwfQ1mMdz<LYxY=OR*JE2!66l3m&uHDcZ^J^%4jh~06
zh&zf(b9;FHG9f<`>Y>{<p7D<hJzQ7Cd8=cT$}x(_s$^08EXok=U<`Yl(mYO?V$P;)
zvMC~+lSA3%;IN2cb18l<d4C?Ikw+0n^L$D@pEAbJyF~e3qNXE8`A4YuS>BOh-5lJA
zNIAqP#2R&V%emgz&PqSmFP!|=9^r`bk`z5|HRM1+4jhYm&#<SV;55n4;i%`(7VAAw
z(1UgKIZ*tsT<7)&SxfRgjs<Wo>%YV3&%@?qKlsDTU7uXIOw3A^m9qGNVOOcFBB%@S
zyKq{7VP69NOOm?*znkP&fd2|}N?%F&D{rI6+(3MOpxOEC?D7!L7rXwaXH84qnH9Wt
zAN-&FNq-v|zL7too~w9+a=AhIp!s9iyVRJwlq_BNlQ`SdbK0<8&dq!T79W8ts+i%X
zrL%0RP3{-7cIdEm`?`vi<Gv8kA%a^P(yTIG+M$}EvLKCNXKKxqIjUoeC)0jYwzMm8
z#}G>^EA#}KwcZ4*Hv`Nz+ks6o!M35c)3Itd({45<!7d<3BkJ0ZxBnWe4&DwzOhuG>
z(dUPcKnWF$#(-u&QL3NtVI%96*67aPP2cM<>`7pC66ZEKs{)p}IM2gy6Krt`pK%$s
z9cZ-!J@jCP?EpF*fH*#O0^?3#f%Ru#^fSDpGVF6;`y4o8y@yiip)@e}Qv6=r5pKCt
zntjx8toKtk{S<L{9iZ$6C|UNJVc$|3Zz)||{~e|Cjw0Ik2c`W7j=&jqD>T>&39pd?
z9aEqzyDm-Jd#6BO{JfoT%ueV+vT!H#!<-5QsgTIDrb4$==!Nv`Z8Ls+yz=S@ZaLQi
z(^@jhlK@KzPFZFTXz$jV{NV&=$HC)+35OJE5ov`S62nGWL=hBi5lv8xON=Y78S4@&
zqud_fV-cJ3nmJq92Tb;Hx!}Lwl@yMC5<0tgiVY<_%^FH5S*}CMV<nVqT#CaQ9kWvS
z>Pif2rEH~!)*-`MD_av}qijQvt+FjacFJ}H36uo{*(=)<<e=<8PywY}K&hb?GhD-z
zB_w~O)IL%&S9fl`pkt)|iZPs#hfECdLgR7%Mbdel9ew#jK6<(&N>419__R^levJ@q
z)cz9XCZYWAXRWyzbW?L1A^TY6xn#?-MZABUP%2px7u;fcVj;IKD}ZSQxh^Y#WhGvh
zrMK#5p82}}@sJ<5^7}bsc(qp-4*7;w`gF%Kugf>faq+(}NrV<Q&6ZPE<&^Aonv`A3
zar`6Yz;Y^xtRGoUMPaU>>?$bP>$l3fLLGzE6DDyptm>jFMsF$^R?Sh(8&TEWx|*8b
z2d;myCEgzHv0-oidT9znWFV4lg4Ozo?Fxx1!{li;7y9Jl=#yp(p|B7~mo!@fJxZV#
z3UFxlI-GbNN0&5P4TaTY{W>VB!_g?sK7gJNpbx&k9eTAxA>!)t>F&8Dq0>2g2$>M#
zj_d&q1{YmRUjU<SO2agr9B75^mWF9Mr%Ap5buN%x26f6vu7WyMB-cWnT9O-~P9w>!
zP^XpT=aBy#n&9v6fjT`{uQgq}ZKHb#U+P`$t(*w&nhdOW$sz)VwMhdu=>*$`+eYEq
zc2Rb7F$oR;K?YIR-qT(dKQOGrWQQPJw}{d!B13BlHBCazLZiuW(S+_Nys1iDE!Is9
z8nJ)Ko0??nfe-8V{p&urN>o);HPI2E&UaT$^4S{wA>aP{hF69)S&zK)kKKFY<_2B4
zm3sF%^Zh(vZyY!@b-}?MTfZ>)M=lcYx2JW%kROvr5eaQBUTI20zM&|M%lRVduTAi&
zLZm8^ZQxvvM8j8hs^Fq==$lKXCkmms5T_>?&Z~Ntkg4tK(C0dtx~_)eYO;PE^s6J&
zqYt3)0Zwgm(LEHk<8;c4y#;R$TT3#z_vYWy-xBYHzN5b*Xak@(0D?CHdLzl30KJLi
z&4Aua@-{$kBY8WZx0B4hI7=ou1<)xZrvf^a<XwQ?g}Lug$u&WdvDuJcApM0e57e!#
zNZ`Day1n|>XtR2=MSRJu`t0jL6D42fqn{{o$5iR>lkkFao9O0`c0g`Fv<Y!R^v6pD
z4>h)U3l{|B0pmRK+?Nhb)1d=OM}6FOJsh4bUd2@$XEcuZnTE(nBqQ`AC@wuJ{RI?m
zK&*VMEJWabQUNux;b^?qorfCdNxlg+Zj#&pH5xFln7cSiIu!9MV&Zp6FJ?-M|NA!`
zUGR9b@nsmkoQ5w$@#QRhIRjrt;LAC(OUQ1&QloY-e)4psi10a~GiJ^Xn>BqJFI-lE
z7ZVq|cs_@IbNwMbOztar1tmsVm9OHf=Hq3drs%7#qoJug+-3aNu5Rug6YaFDCzv{0
v8vTEtYU__M7-?)`W<F}Pm5spOVa(WZhE966j^0|@B5@QY4Dy@=CiDIShcd}q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_bwd_rank_4.cpp.15A9A6FE401A579D.idx b/.cache/clangd/index/batchnorm_bwd_rank_4.cpp.15A9A6FE401A579D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e95dacdc0dd4574b6b07bb630a2ba6b4bf6b14f5
GIT binary patch
literal 5142
zcmb7H3pkYN7yo?8G&7%(%ama-gOD+q5Q8X}Fk?&_nwZHgTE0+&vPh|@c2jbX(yHAo
zt#r|4Yn3)x-L}i$c2R0mv?|?nk)ri~XX>4q+WmW;|2&WLe82B}=e+Ovz2}_w5b^x{
zC<X}P6UWO)Ubr+y6G0GN@RznU754<dA48Cgs>5PwNGI1|p!#NQ!2<7tNf$7kyuxyP
zHCA(>8IOC{aoxB4c(LH@-|s)f^`aVX$sOCzS-SQ9aqQ9UVoo{dt9);1k>kC%Wugf>
zW;(ttR`F4T@2))>>~l4nz24t_qt5JykCJ*@mJa+aTc6@|D`xuKl$2%0w0R}3`rppx
zW{;#_ig&pYveK=IOIvg5U2i_EdWFY{?U8rqzx*Rm7KwURZ-4XaEXmES?dS97XDq9B
zPWLJxTfA;(Rr94T*YgGX)g_L*aw5zxN@`w((sp-N4HVvu64n1s{5d;!UE57>JXK1)
z@mxRM+R5}j<wl}|hPA^n+lmP^+>F<|C;#Nzp7NcdA2+u8$m%MZT7HrEG8cwMf46k?
z!2@<<Y6Q_o{o7cHSx2(pKI<k-@aW4+=kD1#zhXxZsZq<yj3_NEbvzmDCAO>eK9p5s
zRohl%pdk~A$ku4jFzFCGay4`K*1An@S*tdCUDG6*&1dB<TdB=-^9`ww*!nJbpP)^P
zbh*mz@Ws4>Cnno#@Jh~gyTs0{>$y^2md~CEl*(HUBh2`lCuIB{K3c|D%j&GCExRts
z+`MtWchRMA!5Te@myF>5<kY%_KCEW1vl;ad$1iU^5|nNBBR{sk-D96_X0ylg%cRxj
zlBx8})VEi5rE+V!)=9_gT&9)s*B-w=yCi})vW_2`8!2l~75LeOO(6K6S$ij5=e_8N
zbWG}sD&r<j60_>t)dRYFrrMJhgi8&2!un?M5_i0L;aYy2lQ}Zjw|+&cMev=3(7BiA
z7gih`8ZWhNzG2%sn7w2AhdaeZHjedf?D~Y@j=iUwxUzA=GNIGl&;yiR(>kv{FnJnq
z)>g7+L&xr~Pxupk(hLsfr)8hXZfN87#9B8Bq^^WR1KSpQUz*b-3#^x{Ey>NBeA4$>
zU*65UX*<_26K0tv_sQ7vo>?ZJ$nX9uq4fJQK_$gMH&>(0qcKzO;GlntY4HauYyaB!
zhp@UCe{B~8CI$u-4PT}%Zo9p5gFCfyX8RTFjXf*Vw%MaUsVL3*_o~fWjd-yEb(<S2
zabT$HLl@&>DUWq6pt6eZs_S*T?fv6{{xIUndsPt|T!a4eZL{NVQBJFpnipk8mz+E8
zcPH+ct6xaS>mGxlp}>j1C%Z7G@G54k?Y!ND2fYj;1sPvr(`o)`^EfV%JuV#Qtjgut
z!w%KP;p~jO$kX*}pY)$}PM@Q}|Dmrt-Cg*{NPS_}{t?sAAxY<*7HQk#7u>RhD8Z2`
z*6UY0ypKJ>U&q&{G&Hl$+p=p~R~@+(5`EgIbhq{$Sy<DQ9Cvzm30h)Esh};n@*-h*
zN<(3nOI>5we#^%$b>otMc%=REt-+R@b}D(qYqz_tZsKt4>6~oW>mOcSNEwptF`P3r
zPc!HJ{j$18`m~lDYi<FpwAD@~l?DBXg@x_!<ZrFJWj%tL#Kvq)%S>Ky4n+{c7GwV{
z3E{S6RJ;JeO-Lpa6aw@a#m|!U@QP;hz*#aHfgpH*R)DTTK($vec5%1aQ%^D)2CQi$
znt?(<^{h=V`xosU4kM!yV2x>EM1_FrAj|yyw|`E_qM%EGHDeOPNg<$G_~-Meo#s8K
zw2>viIu3~=DFh;Kgy0uR`DNt7QQcxPIv=QYv~-9H@llB&@hf9exPJ@H$!I8WZcH+!
zDg;#L{WP9`b-lk81ziTz=2UYtg@EdfGQGVnhkZN9=v?5OqD3Jo1XLG`YWEF1{5h74
zN`YEiOPi<=P+j%SmFlY;_g}}OQNZ~m;v`HVpnAK*{h>{T*Q>P=8L*B(B1j4W1KZU<
z`%QTlJ&uen1lEQmLw$vS)_V`kTvOjv7fV26fVG~Mp0Pqe^}))D(*1#47cv?Stc^%U
z1_}YyM~-V)5T0jC$VdvXmLQVx3IW${9qUPrbnqcl(64|sRhMe05YYO9j!oU9y7n_<
zG#aSsB>F^!fa-puXYpBs(+-i5)PHmSHZ|o<P3QOBWHb_3o03dv3IVMLJI@5_7rvDe
z&{$yokMka0b*sFr`;~xz#sT#|SVK}5ewB8Ltjg#yBt<4ex-i^<=P(zZ42FB~Jd}vs
zFs_34;G3~1hO;NL15_VzCv%mEx$SPii)R+oG&B{AK_X#=prp?#1H&aY5=Z!v)z~AY
z^hA?T4Gb3$1XLKYJ=>i+cfoNwc+J9ChVZqMM=Pl#Q(^#KGw2L+SSBvE4cfY0=RLmI
zzznOa%N@ISE<rUFq`>6TR4@o@$SWX&CbqLn04}L6R8M%!uzmqo^5Ru-HH9}NL`8!u
z&sB+FO#>5oLBl;`IA4da2gv})f^P*;;Q2|tKcTb3z~b_RgC;O8z?{7HRf}6QtvPUU
z7=Q-)vG_k3hdgM#m>K$L;1d@x@j``q)(?xOj~xS;BQkX03II_V#wuh^G}nO*VD3uW
zX5GC{b3ZKrlTiFJDcF;Y9|f$+j|Of>6ZgS<CPD;T&^Xd(tp~%!I$||0#0KJtkPD!r
ziRo&{RLoRE7Gfo9;7c1ZOHIcXb0CVm<hAF!67s;PyNk-GS`vNYcsHP(qDwJ^EiF37
zeNOu0<D@SFjuG&prkIxujp`CMRdD$7j;cB`4v!NGS;263p1TqyS6+NkdyRe`EGMs7
zSj8zBo|oz8+dpV&;*4nwXLvT6ICC76sD|iF6E$Sd<ftJJrl%STWC|ck`>FkS@i);c
zK;(cx8@RD$y@Z8r)-D~RhM4Zou8hmTxxhA*5UMQI<!OhpB^z>P09$z&|6XOe%XZ#b
zaL=?0)%Yyd&m)J4r%6)WI4TZVG?oS2Duz4zItx|E%hyYZ)=p}iv!dYTUjz&{WEyhd
zb%EhDCXE9t8pDm4MjUt<V7M{Um;+A_!|6;q2ljFdH({D^;2ePArc6@~?9}UD%%Zzq
z-e?5Kz;v;Ru=Zgpfi#d`jj*6GT%;q?hdBz6rN{=N0%74`_}BX!M&-VUA$c7vcvd#*
zNAT{|BYk$~f|g^zS1U6<jjvW_d{e$!nelD-YK6sT^Oc2#;X)muTImW6glay8G@+VL
zp{Y>Kr_e&ER=PqPq1p;Uwoq*a5mjUeX#pe|iA>axnaErX*^1a|$X?_CQRU*XB`@>}
z4uOM_Hzarh+j^Z%gvhgV;B_z|Sed*qk2FY%IRvI0OyW<a37x=%0;ek<KH%5jZsf5=
z&=rPrd0Zu`8}C-0OO7=h_2BCJax=s4RTXH0qayMtFcd&JQK*JQM3EX26U9o@aMAOe
zDXEz`>MC%%VE6Lb2nL1W5#$I4>^>MCM~-8_UVk9r&*%A?wsGJxQSa`scz`GQ7zG_;
zxINEac^zW78_!LN4j(Mt5YyPP2Jo-P(NkV&a10oOes%$l)=G~0m1XRm*(>i=|GVFu
z7@6+yWa`%bQ3*MD=JbVHzbE=&xRHBO5h9<kAZqE}^jkM;a~{|@kHAwV4xAGJ1@c)?
zMS-)}IYN~RUh}*nAv(9N(Dnv#@Z1+SGwe3NlYDGcd78?bDu%WgKAksRi7t0r&z9cp
zOb2@kFblAR`EuWN%WA!|V)3X!cW()uv6#z{j!NGH{;h|u70ri6_4GJ>ez>>ufeoOi
z%L#)y^Kl%32+kk>#{t6@3kZ=<!YV=>d5&Q!LOgk%N(5I!Kc9s@PI_>k@1&s@jyg3w
z9X0ZIPl+>hLl3M2)KVGLDe#!zw;Z**;%C_K#a%t5TjZW}VCC~sUBc1=n^RA-kLe39
z#^I|~AG9uM-5=ExnB<dU96@G6P7oNl@&#NB)|=|nwA9$ND0x;Vz|=;NWaXm<zCJEu
z#S^<(xr9YYiHnot6Ocru5{XwmE>=AzsU8=p9%GZ^lEA;E$eiSmF`8qw2;+!kOk0nl
zJ6?wu%$psgAK*7xpA=+3qfRg~p6)SYrkA&m@6?G?JSVxanE&?H(8Sct+``hz#@3E)
f@8Ia<;_A-e(mB)oCmLEi2MW_r{=8WN4H5ERriOu4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_common.hpp.7297BD2852C87A9D.idx b/.cache/clangd/index/batchnorm_common.hpp.7297BD2852C87A9D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..74aff7b3254da03f5d5a3c41f7b7ad5d3ed191ad
GIT binary patch
literal 1590
zcmWIYbaT^VV_<Mj@vO*AElFfyU|<jg;^LB`Of^OZhOJBt3>9;hPCA&y<jB+ZJn&mW
zcavF0b<`2(mq)wV;}beIUzD9QmuKdhw_!Ek*Re+D_^+MdeB%87)|UHc*pFwvTr8H|
z^ZRm=g~d6q<A&K2w`zHd9L&4-`f<`#v+Pw-?)v*8MZOok{G6HcO#D#N5BJX})<j70
z{F^(iP;$G<ezw$1jZ}GysLfv+y%<*PR9s;kHsdd&(EdpgJrcX#7+d5CrQc!MR6WUq
z@dfW}5e~WeZ`iJ0-5AdpKld#!@9}o4rT_IlTx{K@R8XC9m3_Tz$cl`u-X%w$E>SRj
zGiSC&@rfmqU-Gm>E<F3`?9SP?-&yy5`*~FQ4_~p`MQM)b%4@i)&YB1CJ(+U!jmO%4
zc9H(T+^g;f3X~4UN%x&Oe`k(ysL#5`$#x7CUY~yzSLP<o2ZowKrf|gSPpVP^OuQ@%
zY=Y{7`e1^A11JXs-8cB6W<8wM#>2qFz`&-<peqC>7&w_28Tc3&*%&V@4}O-~H1{zZ
z6Ejc&4?B-Im|)-pYX>U0(zwjt{@mHGVoc0H3;0<0M8E{hf_EAMi_a`Mkjcly3Y6C1
z)({30FlmOO)U@IkK%*Jo`}jE=ba~0h!lvY}6aaHJ+lKEE_59gu(1dxKZD;?P7go;5
z!zLpkBLh>-yLH8W4c3CaKw&`^L7<NyK&+{8QgXw)XpkWaN{TSm;)fchZN3MLJ{C3;
z2@@$;P)J=qTFCIqCKXLs=h6yqj->TEXokQHV`Kt{D=fI2@>l4}+3!SmcgUSMuWPe>
z>_LVg`7*BLhS3Y3#63(vVFM*Yn3*hW$~wxrFkW_0yxyfx@$Z><*d(MSlwiU@o~?un
zjF%I%^p?U$mV;;(<V;RdcX{Xq3NL9PX*rnc!mT%bzsWVvMN?feb#C~n_ZzmMd9~v3
z!zWcN7tb>Du;~cvNWj92hfP;bR~f-G=QQVo@!Gb0OnGwH%LcE<V2T;JAZZU8Yd!nF
z_Bc#de-3hmgpwr8&Al%+6$U%L?nm>=G`VEnBb|lM(Ne^0TT$_XAe|L>9Wc-7zKvX<
zH!sj8<b*Kq^}!39Wlj~N8M)Q6e6df`EFLs>?MQoY#)a$DL!fFwUO{nK+S|p)ap73R
z6JIp54_;ep%y4dBCdd$^gmAiOW#I3}xy5Lz&oD`|?+82f0L{uX0xs8o^S`k|Qw>Xl
zOk9wx3Uvw(IA6jfFM9N!?%s4-8qEQ)+zgBhP*#Siei&Esz`|)FH=1f#<_4w+K0Q7&
zm}VX}C1)jXScGwbGd#?sqSTy3kb{bgl8fCz3@#7>0qiga1H5<uig3e4nb^*V^T#Y=
RxiFcPlbeH!9a;#1<pDQ33*`U+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_forward.hpp.7C4BF3E98C6D2CE7.idx b/.cache/clangd/index/batchnorm_forward.hpp.7C4BF3E98C6D2CE7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4ec87b29dc8ebe816b573c32e812eb6dcb73b75a
GIT binary patch
literal 2766
zcmYk72~ZPf6vvln2!TaHV9Cb{L@-_nfhfdcxD+AD#srXT#50Kzs!525sEx<ews>2#
zI<{3ot5pV-iWbGIo~XsMo?|;+BQn;rUR6=FwHxT}_w7tJ`^$UZ_kZtwZ(j;1t@hVI
z3>#WRc|FC`#&^XqER^|_Pn+rt;9=O=01We1AEHgmuc$=l^G^GT6urFfhjbfZS-dU0
z%q~$Xy7qBDGIp<9>?x_W%zyiAbFlNr$)hjjk7%e~GjaB9@9mV^k0qJ2uO3)^Sv|G(
z+@#X`1s_hWa8A^k#|AA6o07KQ+`Gzi_fpZdrMx566DvM19{9xt`(G0r^kE}!E_hq<
z!u#Y%`}Ef78x_Z69Mb*2*IW;)URytmAVPC<Yqzea>dvk?aO&!pH5FOWBQ=Z4TPxo#
zDOh+}etea-OKxAvd2o52Gw8zJektCYF8sr@#&^s9zW%N$_T=?-0XrMpghM{(Us?1r
z<0M{lBVcKp=DBomJ?~?}g&5UugK~2Q#>_5`YiqZEpk8m<-A&dU`Q*{=*BKk?dORKy
zwd-Z?lEc4z6<>ce;!MfW6X(K{OHTQh9Da~v9P8b3WzCR}FOx3qEL(7;@iXgAS7OcF
zLoep!+!%eHI53~YiymB@x3|S}C-c}W<;Wxa?z`>BBc6DIBGd=&eKUTOXNf#Md&w%%
z$}KqMi1k-o%dQJ1{l%BJ6W>kw)>%A$ecYIhD^`nRXRdA5zug8(#oKp?9*o2rt`6$=
zx<{?1Wp2TQcTcw^-bg<>>+BI#tz+`hlB!7hVp;EnPwNfOr_EQM88PS0*E4c7I~HAD
zwt8OtppbP1TXQ6D+E4Vor=M&(|E3|cpMSq=-WQ6*)rT6zKeu&nI4w*npXqV5Fn!Sl
ztL|EAT)7s<L}CV+3&io`QZQjaSBwnx1Acj?kg|u?W)Y5cV;#cyVT^-6uhYRsmu0ja
z<n70?5Y_?k0plR((&_NSN?TkvVRIFZ1+xy({OCcLknh*&aA*CYyZa)Nn*4b|Vj)Nr
zB=+jKVN9#~uGT+l=Qo)rZVuuFhG9Sxs1bMESo7m#VdLg8uWpKYIAhKhWk+?~STi=Y
z_|x(+0+=aJ*pXpn)B6oSwWRjIL8+*8D7#5u`-a;U!_?*g91s*ibIAjeB55vDAT>!x
za%lk>C8N1I0;9!fXSc&ap`{QZhsA-A$Kv5AbjM<Hdi|s0K4Ars0EGz^k%W+jNJl7>
z$bzJo>y=PguhSbKQ6?%Ml0j)uvDBNtXH}Ftn&#m^B9IWBCC4xv6bK49VnvwR`3@h}
zE?1Px)sD^;_j#})jnBtHxG+3~U4w&2VPpubU5qd$1lF#nuxALYoi<n-#s<c&U^IRG
z1T{yAMsb=mQk83pyv`IR`dC3M7cUMb2_|t=VJEe9{gY-ewUsKQiBMeHU#dnZO*$B%
zEU6ZfRc=jS1H*w;Z#6=)nQX%$*_HPGEX~_Aca~|($>t~==m~uc`!hoW^r?_2f<kOm
zGL;T(2Eu^Y3~ECLv>8*4*^tbH8Syj^Fe3^yK_fOgnMM>Wgaxr#)D}dcFtrd-uo70p
z)0$*WgX%WIhS+S$HblWr*b$q3pgj|=bEG;D1t;M|Je>oaNDeN-h1gulE=0jixDlIM
z?aqJ-o>UK~uz>f_&ocN?r9bnQ0YVO2lmw)lZ8)Y&fK(38UIt`x*rKX9RWf`@G=!!P
zJD;lV4~)*mo}OlUF-j1{bsrAK3C3~6Zp(DFZwSJ!E)o@SVK%i^o@`YtpU3PZ7KpjE
zIM4}n9I;c=FR)`;vOBTS@B%xgB<qC(txOxwR^^_G|4I-dS<1YsL@*&i>JUsykU9iQ
zC2YVrkOj*munyWtZBKSP_L~XCH-A~p_C4PpOGrJt%Mb6I=l_@OAJK@0Tcs!d3}?jE
zL<$rqvI!j|Qb8ubP)HS-gpis{K}bWUBa}&IA(Tz(AW1)wYG6WAy>u8PDxOMY6TyKY
z-(ZGhmYb2$YSx*N(MreD$U53YBZJnWv~YtK2bO#bvZpQHhODy>x6`cNzxRY2^1Rml
zI8$?lECJSBrjjKgq?V;1q>-f~lqt(XC|jn3<di#+WSlxDl8h_fg#_g?xsYVsN;i^>
zJKxPE!~DmVjQ5>x`P2#JI^P4=r5G0O<IXFBeZVjYd;~4G&hz#e3P18A!-)t9770!N
E0Se>KJ^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_forward_inferring_nhwc.cpp.5CE047E1603B8333.idx b/.cache/clangd/index/batchnorm_forward_inferring_nhwc.cpp.5CE047E1603B8333.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5880cc9acd2cd31118dfba8ebffe1af0363e91f9
GIT binary patch
literal 7998
zcmcJTdpwlcAIIk$L+HVn8RHg`au>N}(6~g&HT6hTq*#R7bQ{%HTB)R)M7z~?r=p@w
zzicF>n;F+?!bA$)@3-CUvi%;jzvG<WZ-4*tdO366&*wbn`JQvW=X<`-GiB0*39cqA
z*7*2I;<SXtacmZgrOo^=TAZFdm4SU4ES7jr^|Z7wv$32mubW{OiE5LVudf@@9bj)g
zq~$~C$}_`OEWML+r>mo@>h7smJ)IhvYmN)lr{sT-iog3Mqx_owT>ZHxEpvG}YW5|+
zo}RwioumJG<)1@peiPQ5yYN7F$D13*Jw0a29p!R|pNx!NjtSlJyWZB2-~|^ft~NX>
z*<?~P_f1cejp*}IkNb~)ebTN+HP!?tu3Y-VUq7UOeqrPFO}%l^TKmGn>e)X>o-C?T
z_uaU^Fr#N$Z)Mq~)U}$Gi<k7Q-DUV-=-Yx|(SnNb%CoI!a$@#>OAk5pDCUob#!D~4
zufEJ0^T)N9Kd08^l(s56%?ULh*}CG<&1I8T@ods0ex-*dzImzsC}n&8ncLfy9hGIW
za-n6K*=5mb9}oGjee1RMiB7Xir*`y|W#^ZLOMH55XNo2-9X_eNtTlXEaJXr-)2{Ga
z;=B>Qr2z)_{j*~9vQGMQr}xZy@t|55tu*(Ri;tP#z52&1+j2NY2akOaHTaeeDLhu#
z!`pTw?WeNF`Sb3m$mE@qa?jSyyC1&Q=$%EzwSr422hL>CYxx&arwUcPD>T8z#C7I}
zrq@>pCX8PgwuN(O)9!4Q3;yu|lVF{3*{mI{7Z)cRaXa1_9Jn>#2-;-1&d4tIUZoZk
z4f0i}8l5`XjJjjfH@{^?AAc;q8#KPp^J!D)mZ-(HMz^;NtqFVkac208^JViB+5^SA
z4?c@qkl@3wdzco!EisKVDor_!)qJb*(s5CbJXO@8eQZAEUX*w;z_!h?=w@D?bKTB#
z73=0$yQqzV^xuAU%vw|QVQ-bvR{!_Ey+3#?Z=-C)u9xRa3ZEv0`A)mR3qCwme8;R_
z&!A>k$f?)+a-8oLG=~-EHf-K8bi>(m?K7_{f8v(qo?PPTaebF$I?twU{g_+Q3gan9
zzD9Ma?A!9~#=<Q(<0LCrdb;b~cl7UToN{HAq|JRrd&q<o;FoV=^sds|xS9U2+V#Vp
z?O#@1*$x(GKIk}j*SO(gT1B2&yK~f*v)8M4xO`r+*xCL1qmG9==O31i&ikZst9a70
zyM<xG;^Owc$nyPa%g3eqyk7m<!`Jnj&dKePhn)qXx5_i>7E1FstX;PG$gper7nAck
z({_gQ&eZ;kwd<zm>Kguq&6h>rZ?-%1v@bL}YNteb)wdl7oOkSbbn?Ke(B6cb592L6
zwdMA)PrT3DoqAI>b}4VIorIU3e^T%ldk4E{uh{;?0(SA{j>qS2th3sCr~LT6wymWB
zqun~*Z2oE)J#S=rQOBPRhM$km+~B#{Br)Sb_4@M{GP;_0_BV<|o$uN^>mp(TH)gyF
z-hch-`g6NyJ%89#{UTP={<XxE*R`p^=DmwU+}+!!99lG*UaKnw1%_Q)p`^8JCqHfh
zJ&r3fy6jPMg1d5^E_;4vna4W)hfzf@(hN`CD9~=&^!~87$F=%nl-Z%)uX4-yie}em
zYIpfLGX1WsJl~6DL-md0ZMwV)S|!UYSA86NH|v!$Yf(npoZW!M;vKv6dXMHiu_jCm
zWzpuU<~&8gL?6YUbF5Rt_nci@I53rwt<|h~iUP?IZU=(lly`+1tRxnT7U&2}6$O%K
zeQE`vg{R+X!ZhY?3pESAqCoQZ<yrTWZ?!zvgyD=lN_CXBqCoP7Caa17%bm3pE1Bv1
z0%Tu+s{i>xO%6Riui7y;Xs;%m!t`LKW~QeokX#&RlxO|2M@NG-mFb}h@~fcXe|}Jt
zlGfj!)oI7+b6^V7gQuz|Pf;NGQc%H?q#t=n6l*fm!xH5s28sg7Z~aQS<$pNc83IEX
z*-*t$S5YAO&si}ZdxLGaX~OTA&Q0}9M=A;=f0<lm7dzYMmL{CX$i}M1d_{p|R(isB
zi&}LVjIHr$bh6R{6nz$sI16o%Rp3={sCTX36BRaH53K9KfWW3Dz*<7G6pWSvdvqh4
zZUA--U^M1t;M@#cFt-A?Rv<*IG-l|rebPC(suaByjM@sUkwDSgfXz1GfVm!wt_Mz-
zC1A9KWGQfyf^k@H0Im(d19LO*YX*}r-vq8Vv95A-%*#31{@ax)dJnMPg9VD-52*dX
z2;X=V7$3zWrsxu2T|%}m15_DquTu8=^wnA^rGq||DvdSIo?DlxD^YX-FfJf1ya+5W
z;yI`2T3}HN+|Y<rx^lK9zUMlyDLRlJXn-7-qC>nw{4vpdn$RLek$y-UvSg8Xw%_Ne
zOb78p<F(OcicTDrXoD$fR1!hSW0S`rjT2g9n)duhGf2@@z_N-owHjDdV^cYm#%WGF
z1DZ7Z$EaR_XSsjX1bcuL8WTnD0|xtm9p<Az{V3VK1PDsV_NRcsDYE@Jz&{5p@cos5
zR|(89R{?$%saFGDHK|_&{EMVs3wX7pUI+Mfq<#hPu3#NS&mroHsENV>MJK5y@zAtU
zv=mxNq4PjIqg$X=3&}EQB_ml5t>h$kL#uAgx^+3ZBVJ`>FzfgLI6eSl@o0O1V-Lx_
zAfOk7V7(6n^kLS!Q?B}@_+O5L?(>GOua@l|we_ce1Tf=e`@OG$kq;F^;XD9*Y-$hS
z^pM;OIK3qI0Zt!g6U|KFMpfy3rlTN}AX_~3lT9WQ6k!rUP=ap4NYv8wxAdgvYdo{r
zY&ysw$Ov6#(~}J*6BJ<(K~MrWK?k)Q;dQ3+ty{DDpikz=-pXt0&N1{qCeuFb&{M%e
zo6>&>&g1IT+cKj0%;cDIOa<svVj{=bkb%uj(PBF>5w|meAQKEj>w)qBC3VI5j~#s`
zf2J{-0*aokGus4_yYZX)6qUF)s%+YpYs*JHFl1_LhRKd=M`#XQ2ZEfr&IGw|T?i6#
zg#?vCPATLfYkdO5`qOM}lbD&Pg6>t&2PFc&3y#_udsPm8l=t1gdc>Q;U#2n^<N@nE
zAVQ{1s9N_j_4#-E6oOhF&_Y7U*t`V-yWDMzuo3!hgaf;Fs+xawS^{SaBSh;(4@@>4
z%C&gB-BX!4M@-dBweX^f&BeB;Q;N<6TA4rxF?}~CQ1)}M*PuHoN(>o)>CXyLN<-0p
zWBdr0i#CZKIBQU{$i&kAOcA?SvTE{w#Zq@E^e1Dfdke`j=q@8!4&CJ>cSHAX%(3D<
z3uaeW^)Qp$0qr_)kWC2jsm?4fn!}uR`M@rp%wTRqtJ}~XS&*E*Xld9V#oj6ulY_DA
zfd-muipjy)5|X8WEhV`Dup3Bj2JB{%TLHTj^E~y^KKJR}>ljmWp=B=gLJvxr5<9x}
zZR{&X$OnRae0ru%Inpuw4ZC*GR~serl)55;$4i%}nn>gsDMM5EC^}GdicyL=LG~$b
z1o@-{5;QesCZ-!u>ju<A(@)WjklzT6FyDmyn<TeEejCZ1kl%@Up;**3tiJUtWHVdE
zEfeQ7Z2B>DdklrB<AoLNAs>x@+&k!#x#)OnRo;f#KQPM~Hau+LfD&K6-@BaM(!d1q
zPN3EaG|}=?bQe(TBKaOryGL?2Q0vCL_C#=6#NL)`jKh8c=AUrVPtl)&?PoGY{sL^i
z0C)VLzkuUkzz6d;;Q9^tV_pNz*T8|7i1n}ihY6c<4VWz?fp!w;gr<_B`=C`H-a_a)
zQ2(FTyG?;gSueJJb7Q7L4m9OBj^$<we=b@0ah}5EUGuTcg$dWjfEQ;)jPL|#JOO&h
zxC5y{?7;Bs2Sb?i$I{jspFj3od*U2%;yMxJ9N`>^Te@-G2u(OvD8$+c$T@+}pM$S-
zta_iCeWED(Hqf~Z43HHREd!=9oIg_Z9bj??jKq2yuxkU3nA?F%JGlh=q04?qEXe`r
zaR7Sb_6MQ+K^z(=`Uvzn0w-X-2>KVnK+MHZSPTbFCyK6sh81u)($V`h6O;R2uJOi>
z<U|e}^!@oQUkc68GEz)_%WolB2Kh3Q<&ZBYxf}AkF(2MA{inVq7ltu@QwJ>T@Rp?L
zE5PCke$6=iUa(Ck=yMUXB_f59#IcfSm`GTcWSB%yvR(2R6iiTBzzpy2j|2FelENq{
zoJ|fXs)b}3jFOQohf#8pyJ1u}$xoo$6X=EdJNi1{K+KQFIv6+210(W)C7N(#8B|w%
zhQJek6>43D2I&6snQe{-bG5ibDB4!TRvY;tMcZlE5#*rZK#;SBGeIsIE(8fRganm9
z?GiYUG@My&-!t4W-3hSiXzyrWWMaj$<rTU)Z;uapKs+~Lr}(7zqqhF%^P&al`IQ>&
z{;E$;+c0zJr|mav;IihErtI(Mv6##M{jsESe%j<tv!Z0it<C|<bEI2U0*gxQR#n$;
zN1gC^*Ei?`tCbybVl`e!9`pwx?_&JC{}UGWatY)~pe6Duif)FyW|D71-ffaQAg_bu
zdysdJ<UXj?2S=j*F8jTVr#{rKV-_wG_-BHFFm$=)l&;Y-_le9>2;2mdusynty5kYu
z(Dz!i<M`_9%-tbIAp?1Dy`lJlnpjg=nazZqiBcGdp2lXv&cqgyWpJX5WI3EDC%GF=
z?8aPQ8!0+k84<}iPPBWp2>DZ^*@_KoMN_r<x0-{kGu~?U2KL67oD7@@a-QWJhc!0?
zH$oG73CCmY1k^hL2VS$9Jk3{52^&^0=pNCuzVP$v{3|kMumVj1AEoOp0b3=TPD>02
z{rNKE^UCcu!=?Y&rT5*M#{XWb2L|1O`!S2V-F&wX9rOo1+i~*aUYn+Uj00r>b{0@a
zyOE-qow1JOY{1SYc_UzNBsmALb4bnw>|ByJ1NLUj=<s09lKwB#XkRg>Dpak5`e@(0
z-BVd(UuHUT&?n;EsTYMUD=sZ&)?Cj)&jZ<m-dt7H6(7fu)y&rkPRKx1XL4a)7;%bA
zVUQF?q9OEW96>E4%V3a<WH}6yliUr1x-tLxLd{E?F3)BxtAb`#a12Te&=-2<HMalr
zJhsix+b<YJ2qs?GMi0EaeBE$ymeT3+rA%xSnTT9*qY#sj0ZLCujq4x(qV&0oB>k1?
z{l(vZ@u<RL#o+2RT%C!l(YP9gt21zQ4z4O5_&3fSxT-|IYP<8xjaQ}|PTcG{vGH@K
zMMck;6*Dt(dIT$WpcN|;S0iwB7Ou|5)wyZ$sfi2*CIl<7l~snQstu(yI9i$k+J3&~
z+UfyZT^?Ug-%aS@IZou|JzB@c-OSa}<o{h27!ETUZaQM*C<`lVTYHBwPA+4Ooek_9
S$Lk0tOiF<MVLlT<5bNI$+o){-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_forward_training_nhwc.cpp.6918CB89E3F27E24.idx b/.cache/clangd/index/batchnorm_forward_training_nhwc.cpp.6918CB89E3F27E24.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d1da7863f309bd0c965e476ae5b670c50b1e6562
GIT binary patch
literal 11660
zcmc(lcT^Nf*N3ONN*%Bn(hvkCDiV|=${<LZ8DvnwKr3P(C~M?J0YP+k4M>tCD6oqO
z6;WATOn@kgtN}#@brk`XT|mSDDqzB}ehu&TRPB4dKldCshn{}sce}c}Zq==-+cQ(<
z;jv{bk0)L%TN@e@8H9K|o+|qv5gEQLorNI+9&hcg^K)i|>DtqEdmCmM2Hagdcb<W=
zw@y)I?&W8VZ8NrwYT6yRJ92wuRCz#!Pjvgl)wY3CvZ+lzM$&F|Sq9y@YxS!ech?;$
zqYIsnRiFDXB(p2My&@_3<@1L2p{pl2IyD%VuRp%`)(64Or<21z@JCmz-;!W%Syb`e
zZ||y-PlZ(kcXfUkYV$38ociy*ZO;w)?%$u?mE7C@{6WL591ENCRGZn=Lp?p4R?oe6
zVc-6=nKzcsmcD!-+!A^BKl_v7x~IK7FyBvYielQz6VFoa$E~i8%H3o5?R(h@O}F@4
z-j{BvwwB)Y|I!jEwhWSb@64Q6W}w?M0qJ)~OM|vAZ!Ih@n*u7&js0QiVolqck$Ts&
z{Tyqx>WAfsYBG|qp4)V^sXXRe+LFNYxAo#m^AAq(NFDP!uzSz(s%*RT72ADpjn6uH
z>VVF)P=itG^|?E;+)S=r$$0K4+EiL;@BG7@$<Y^onkwFwXclSSt5Q1Vk)U+h&A;|k
z-Agfvusk%`z~aH`=j9LWE+l_{`&_GEYofiafu3VafY2%6dSLzp#rbi^XeQ^94Yhn~
z$|@%-rMb6G6{VgYym|fi^KE)hFI$%H_=vW=zVYLvlVPWH&m7s6ARgjxQQUNF(UXIA
zQB(Lrv!i*}mu|G(s9KV8Z&1Se<xK`o8?%kqtre|W)CpT^ebQpOom}qkQ9PIa<-PB8
z$I{3jJ3sml&Iz->*wnp$Op?{vlld3VnOZd3p5^B`>^}H=Y*lH~_p6L%{O-T_rtAFe
z^ShbqqEU{^%ICZv7M1Z=$2GJ3ZT9E0yA(YeZ}+T!<Ch;;Xz+02FVBWs{=Ol)>+V<Q
zZN6(tkVdaYlOR!ha!XFnjiFzj8l>J=>+a3!ihH+Qp-$62z_VDgkjfsPDmyeJW&PEM
zzG9ahhTmNn`sZ?W?_|sPFJFAphL#KFXe3l^k5$-eW)z)k{C;R_bVp29%BztoGge2A
z(Nz1<b>{eQbMMdc+WOi8C>sntees8BHG`U86z3i;2^n;9KY!u)FYVXwJjwIZo^!8&
z*}6OA`3le4x^+5*U%w_OcYN`(To!MaKJRnO-f7-{R&9t~+VZnxxv}9A+XuoD?V6%B
z8zdvkWd$Lb`)e&b2YI*3j$Bx5v;4q`kJ%gF?KA!HC;h$4!d8xo^KS{8`ORu;_>DJi
zAG70a7y5?IS?WJb8duqQ^_}{O>PHWdTXy(}h99pqN*YJRB(<Hd+oifvwrP4xc&}=@
zwOL@}3H{#IEZ&;ye%mL+oz7A=xqr*z;f(P|4qhDIT^)7x^zy@9K|R$a%k{Tkn^x_Z
zvCQSSN86lTYBCRBYn%G`RD)K>*~2@2^Zhe4|FZHQ`?q|)7k%XQ*$wN~$5&8J!!|!E
zcWvUS=LYvs8b2hbT1`K`|H8198joC<8JjQ2zHaP2GxzoBDF3bFpVId>omn48ho%pU
z{{3L>lfo^ABbD_$+kMyHx;@L>X6p2e@rVCfS<jDkXiEzU_Z1p{3SO-G_2t(5ofCg}
zK6}HH>%z9Upxk-ML3gVc_57rMbj6*U&#H@Vdl!gX7ilk%#5&Zx6({?(exvm5KI=w3
z)LP4nhzSkWr+7StnH46-uG?uS@qBqarV1)mLE+yYZ1n2?6TWZ_s&UJR6;P|ad5ndE
zg?|5mldX1qD)JirLR&yBVP!*Y!;$?5PPSKCQHknnUkj+2tZd3RRqa1;vO{9e%h(~w
zN0fNe+52CE!fVj*?+?zz<&SgP;tgVffC^(vFjp`a_8&NTx@d4{YJ%Y(0%|EMoAJ%m
z`wyHPBu;4GZ>t#~ph8*MM8QPA|G>$?9cxYZo+&U<;w@wEuc@JF*ni;Uu)#e$Mx^Z1
z5>T_)tBn<mwfhg89Ek#|E{dPks8B0ec`|>pdjElwH#9XxPcKz25m2jGd6dGa;r$0r
zPOLR6_B73^q<PEPdsjnL4f%h6a3<x`)1RB#{yu34HH|gV9jvR-f8gY1$El$j=asuv
zsQ<7xuvD;A?>}&Iclys){WIRh38)pUtjE_A_8&O;lS#o>-jbPKG;ao5cq|wc3$*|K
z;7oYoA>Ty=j$$#O!mVbUnerf-$%96D&_aGOQwU88p%vz1(Bv4&<<O*@<O|T`0%q`f
zs<&&U<T)QPFM;_>ATJm(9l*SU<W6AO3FJ3I%quYV6|m%PjF>Jkz6(si{1!O91(Ps;
z0HP1T0rMx|^a<<GX6TW<H-l3YY32YJbpV)ihBT88Eb@Uh=9j?oB^Zaf16X#D+zCXT
zxL}&C=!93m5!-hIyKY<&nt2D@-+{?k{|F|21TL6A1NYBZAAH#6RdBpCgLTCzpmGWr
zV#5-kQbKYCFsdLcb{VK#CiUw;<vQlU$3OUOP+oC#z+Ex8rs8FX`Ef-=Ge?2ZQSu3D
zfl)2V&A_yoY|0iex&?^178%^Mb=#7+YSwH)jF8dchAYkZx%f#j1qcHOE!a6&!f8X+
zM63;ze4aIgX2OPssd8V_Ot^LU1WXau5d=kAM-sHgdJREQ?opojx@h-kImsij;{U{7
zn)w4t{(w{D;evSvrO)70%x|H~Tj+uLD|Gz|Cu2@Ru1SdCOyrh{Jh8q9N%tV4yhF(2
z5SoVdQ^@TUX<v@q%aIK0*O1dSB*uIPN$=qDRJ<z;Lpx%PDVli<OdpfpY67F1u(wpo
z^g_pFdDg1*xm|D+cTAc&35-sH@tj*}rWlwMlLb`(K?P}l2^d`>?XLloYovWG5Y&?P
zcYxpy$qhi*07m2cH3IcUV1oHE5I!dLCZOI#>dipdOzJH_y@k}<fUu3!p91x#SXZCD
ztRmx;Y!#rH9hAuq%89!%Hx&<2?4N?U*@0%F`O$i~TRnj451<+LM+a2zAbAr~--HM~
z8>wfL{0mb51+&)Wttm#YV`A7rltS62Q1)DHXeO1iO(i*v@=T+qVLhGlOvkMK-~|86
zp;Frck1x6tUPtd7mHzX;KAGu7=ebsakx%=ZN+pF7;-*NYR8mP!qg2vJPN!7TG3yI9
zIVACKHnRoI)}K8FyLO)bJc0uC0|{ECwQ3}1X?Qk#X~#yVc!ZeQI<rS`Un6Fo&OCww
zbpi=mrM60gvovx!UDhqStvKM$8u_HGO3`ki@4xOWE1v@|G*??3{r3z0dqu3I*uN6S
zqts8uZ#dV!G~=)CuaD^=^n3_?@tA!AWltc3`3)4mfj*dfp`;g1#~g<wafskFBuzsz
zu%3%#xriw50P;G3W@EhwNsCDPGsxo%n#t*QlP)XP$M_5RG-Ie@sL8GQ*@@KNZxVip
zcwi9^Y;dQ~1e%#(IOgjVx=tx#zDM!zQ3A|$6u*w-2NeGSW^OA%EB0?8aJz={!-=Um
zW@wBMS1D#hj0r(jF(QK8V!Q~N6*C{xeW-jNYIDy|GcTa<1ss9-8x(#cc?S~iK!kn-
z36EfwZ*vs?XLCa{O;FMV2}iX<X*-lRJk9h#mmYGm{{mgVKw@%EM6QYCWSxQBGLSc}
z=x!w4jU7od1<0cS&BFRg<aUy@KZo4UA)=#IA*U*GvcHX_w{dwB^}4E83=Zm2MvS$Z
zwUFD8uw-a$jLAmLhS2QP><AL6i3oB~b0Ek`&50lvH5Y=!YGQ&UY7&AjQEHcPG(}81
zrPlsWHgBiKa2>Bt7q?SV%vn$~3u<%B?LgU%*SGz|9WRO>rH;EyBlK*9M9a29St~iD
zUqkV0NO<NWlzb$I^cE!9LbiM=lBSYF`e!8j84)33Kl0j74(VT!^jFfp6nT{5A?@b5
zcG&DOW0tauiYn+@1t)V0PWQ0GWA$9h27cM^p0~lMD`(FvwuS}3yZ|_J?e9?@_iAOw
zcYFH<g#w_=3DfK|!!*aHY-5EC$~l9Q$%U)8XWSqA<{CiE4N7o>(&Yp$Tfn-qFI9kD
zxrV=naAgDZY#?3P0%a|vD_=qJE7Fx8pyUJT%2*_cC0&_<q$#9penPUJNLTJdUi(N_
z9z)V&q<snUD8a6rr64^Y8lsZMK6Z$9h<xGa>#a8WS%%Xf$OdTzLuDMK<H5*yazC31
zlrw<_*Cf7=7JAj?PIVdZnLc3cuCT!K<*n>SAqQFIpb0og<sqv)%v?@~bz)ya2Rkv6
zA1QAQcDDgNAH#*Xkf%`g6b4}KhT?8G7xQN*`3(IrCm=}zB6u5;ZbS31z6;5AA)>s4
z$m<{q#QF&&Jwe)^MIL9-d`{;wPf+n687G?YvGF0M)DZm;3ogWjOmjQG=|n*=yDBUM
zHidv#iPr#|8j?37i_OS}yBEzQBa39r%fcg~d_Ep>RiN3V3wjBZxV}KMNf*>XawkBY
zB)<aa70F!yb&>oQptmG{0O$kBp8)zq@(l{zpyc;oF*CsOZg;?IwhI&jokDU;Rs(cu
zFt40>u+g9kRStNk<uZh9U+hZ{;=UNH5=<=b>Y-;noP+(*3}wyG4|69JcapRIdnkEN
z2B6JIvYDK<laVx;3?Mm3mV=1$@{m^^BKr1GBt1&n7bA~i9B@{zb*>#=efKLwOePRz
z;+-;LE+ePQXcAXZ#N0yCTS%U3SY6yO?Y&;??g4j0MB&}?%*4PS*pTRN;4j|>aU+LP
z>`%;M@7#sTccBjVSZgmeKUs(F++qh-0;Q0Emz*?{NGT+eoJ1)kk-U{s*ot}m?T>KL
zI-O>=8r_s>Hzm(La;adpoAl*@aYZ8d5xQJ0Y3331d<2OQ(*|X2P#!X9rVEO@AhBxy
z1SOxyZWWIt@rdYV=}4MRc8i@zwiAaSHnWMm^3ejW8EED>k{&1R%aBJI?naxBO$}vs
z->G85MlmHQrbghfQ9=nzC}Yf}lzJ&;g!wEbI7`}}qwLO6Vyu@_j^&gW<_nbF1yZk|
zL=_Z~BE3a<-lFDk_6cvK!yRf<bXfn@L#ui?9zRhtv}z`~5{<4zW3gV1tg4aRp8Fms
z{!b!4*$ZyzO#de8?;*+wlqPrUz8<2SNOBUToJ8_gN_i{hluZt~zpQ@0yg$NCEF{ys
zHDF>5=By8a3pOhczt0*rP@)D(g75f(8ux;d=ZE*LoPxY&#JAAwqTAHk951@Zsf{C+
z#P(|T1c?~Y9Bk>R=16GHYR&|?tGN>-wii3#>m*`{6w?(-?aDtF+1~BcbgtTn$$^?V
zcu~FowT9V~7sj8|Xyz->_zHA5Lz?*u82*K`VKnm%=zjwvu^vlJily8!Z=s}H$j*}h
z1qo0c>xod92!~-#g6c`QB5Ya<s--|JtfxXzDwGdhnn{B$Y3xAbUY9W2plBPZXF#V6
z=!W&ZP<Jmhz<db093rnj3@0CkQ?XtK9m=45wM#Q+;pDSe&rkIGxoypb;q0>R4luey
zX208jX&VmKH1iaUev0#{`END1X-xTCz$UMQ9fFB9MYu`0IgU9ICJ_WhnnV(`#$*ja
zQKBd(d|k9CT26=Gc<x&md!&(#dj-I#0GM(QS7i2-nb$RA4Ar;C(rc&Wt<uNi+i~)+
z=-|WpZGug|vL&2`#;2kD^~#C)^|t#{l+|dOu~xEH<+>Zq*eKZ$WT#|DkVr{Hkb{x~
zK~73e1i2`=5F}O-6C_cR5Oj%Bx<sjR4@R@~S869Y3#w*8`KtQ#2CG&B-SBbj4kW}i
z#GPxl;+N}+wYGK_4)}z5#v^O(udeO&ef3ih(d6n^rqm$izc^{Zy1_@)N1m-J6Fjet
zU$va~ujk6j{u4U=v2nq&z9<Jw8_3Sw2u3&J&dkl2>`dPG8kBoMNi(-0|2EX(0(bef
z`*V*u_OuOn`&QN+9LKA<ux!99a!K-PG5xMGZgAh)PMQfkxE_p{I?B9`B3`)GQBHMu
z$<n8L*HJ#CePA6Gg!utw`GAtYkiB%LNNYs&#L28HHAR{->>({%Eip&ehq-LrUbvxe
zW9~P?Prg~J)?M48uvRcg5it*;<RN}}g_tK$`UJ{fRw3pMba?~aIe#Oj7rORB`O7E7
z#39!>{4xqLX~-=NNw9q`lIG%<S%^7+JPzRe7GjE!TM=o02DzU>QttZd$~n%5%a}Rr
z_z0O8;>-=38sqheo1JGV_XR0yYdajI#_5dH!(^{xPmpMVC<tqgI*x?qtm90OyN)|S
zVi&O(UniL&@xgS3(z!zEaW3i$QnTABf6O^hF9*ukWwlNw-_P_Jo;~25v^F8<^M)N)
z>ey!06le;$z40B-^c%@1Z|Dwq{p$424QjA3xcRTgzh+Ub_`^|cFyK$fbs;t#(f2zN
zF5J@065$g0Fr}Fl8Y|?x+UA9ke~R2Qh7Ne$UvB^P{celey-GBb1yB|s-Y{eXludFD
zKshAm0+dVgPJnijyc?k1B<}%e56O7|<&nG(pnW9gLpmR-;qiA6(g#T{fpiJUrI0Sg
z%>96ebxPl_cd%1@Tz#f;v4@RBV?yL_HM)0|RahN2968`l=_%^WxwHPaNOnkTTWdRV
z6F_gLf{Uu}iFo$wDri;(?YT5G*GBAHv;Y1SiMyh2Pp4H6&AAXvGtE$|ndDBW)k*Su
zsP&%Y%}8rABCbzHTFE5mAgvsd^N>~^$w!gaQId<1Rx!zyNVyUj;rdl0t!k`)O}w}O
zo;(r7&WkdA8Sz^XKYc$rg@>#R>U-db@I?_f`%dQe9sc__hdbf%Lhxw-KAnwE7vR%*
z_;fKoosUl!;ZuM4DP)rPOGXCQn>DBg2Q3N=Sv=c+-u#6D3+B#Q;KvJ=oACnisUJRF
zh)??;rym!EE?&8WJ@X2g3elj!L--0qX(bh9fv2j(-9%N<Q%y@<s69;N;OI2T*~QgT
z!(yWG1XKP0?^{jX;Uf$TjYf_dZDu~k%Gze!czZoNow2rJ4NVW(ca+q}Z8Dg``#<7W
BVCVn<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_forward_training_nhwc_obsolete.cpp.B9948F07339573BC.idx b/.cache/clangd/index/batchnorm_forward_training_nhwc_obsolete.cpp.B9948F07339573BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..32a4b286c2ddcba2c441427afdbe0788e885d65f
GIT binary patch
literal 11668
zcmc(ldpK3u|HseTYbCTD9490ZN+F??BX{Q<A*I{K=;D^?f-b0-W~Of{kxM9<bW@^|
z?x~bArVA-#niP#13X^LtrR((Dd}p7v=KK8qoaf<r?6Y3y{n=};y*{7y*=wzR=DWGL
zB%1Pg;$?18VZo6Bh{xlpvi}j0;VUy(u;lZ2Q3q~#hk1^+R(W^0&@=twT+tb6qd4o;
z&NGg`g`CwjEj?6U&{)tPy2iQl$%~iuN=D$|LXq*2e7m43+up`~Th>*I`h!>A_-(Nq
zbC_wW{{-*7-;v;OyQ!k@^P?Qm;W3BWRiE2Et!myh{M+8_E1S0MerflMSk>FgZ%RY+
zt)>r8UYvjbu_foH%{LFmzy93Xu7dRJmsZ^Q{HX7K`@7zPpO$Xz5x#tR=g!YpjaR*x
zJ8$O6V9QfS--akU&sb?yecSm%?%R2jC&iB+HuJJi<lWX^M8P(*tFC_iyuW5r488y0
zwr2)+i`_!H6CNoQZX0PG=DxZqCiKBGmG8@5jNCpuFX>cI^W<TuYtd`5LGSpN9*wD?
z@ddr#Z#lRw{-t)|ilWhFtrh$GyQ1W3{OYpb?7b9u??h_Tn1w%XtPPr#Sz6$h*PQ8I
zTRSf?cX~>iLg?)B^08{2cT!*5h_+lQv9hmE$hPtK$kqEj&M4COgUXdjKEp$_E*7^~
zOv^4maR0SxS^4_KGrFcHb?NKA>J-(88k6X_(&YEb!>YWeXr4@>yaP<ckDJye{zpsA
zK)Y&|*sHfa#Nc$v9Xopu5Bo4*t=Ig$CF4h_d`)hD)-Jc&xTWFYmKyid8{0IajgsmG
zlvL5gb&6IN?Z3^`V0}r+oQl!ghCd9KC@T+No&SU5=MVQRJP)-TuUl}WKlQ~S#Z@hP
zE6<McyZ^_9^|0O4tUT@gQIlDUMF|&<nr0{PG9P_2!!GLkwu<2s1k+AS@6xxY-5c+h
z^V#>pZkOv9s?H_!ZJaqrZ+p?}oJ7O1E5%h^B|)h%Cu2g}CLN71&;8ojnDgm}u_s&e
z74&;dK2IMTs2%TDd3JL5{;if?vwp8Sn`oBrJ?q@%s~?>6#`g%$<*)RZVdMY%)aSWc
zJ$kf7?-HK({<U6_7Oz@1hcEtP{nb2??ICTCxgM&87q!p7&@2wBI<+iv!j`z5*+J)1
z-oUiH4*~N$6^uH)DIG`k6=2is%3%$y1%e+{Z{$eS9$(IP{7zVMD)#FP>r8Foh?hI-
zF1erYozyHB_-%-t(&NAW>f5tq@%`}?DP8MENIviA+*I(UT{+-Dlws=Ivd8-#G{u_6
zCWP1>m>qS-<Ni45;i$S12kwR}_+eVmZ!SUIrl9ZA#2;@a96lKpYPH?Ban1a%W-;N5
zx}Dzdjh*keFwA>}?<nc^E&E=6j5q(bvohaXBHkNbUHjvmI!WD_==kOv_J>qM-L}kC
zjyNsYWp3nOcV5@JIi0uup3m;-+b^Xn8$PZuX`46gWX{hId&<AN`+3!g?ttF%YpZm3
zmyM~gNnPn6u1~Rd@Z5c(tjVMPQLV<?YbW;nBHs{p?6&fYqYpm)cIDEh+TtR+Qa^Br
zc3Qo3?F*jTj1|3<>JJXdW-8~7-n0uH_WkC(qbPd!N6@=&`3tv>OTrL({0X?OUvZSL
z!PL&nvksRr3Tm?s%}=R3`7ZOw{j7?DTXcPv?zsR1IyTcL)?&fD#6J(FN+U0Yq^)l4
zOJDRQz0_xqMr`#O)7bNqs-6bTId?|7DLbUtdFwv@>6CzXs!Fl;k32<zNt^Y*f7tGK
zWS{XkUPN?Qpf1JZ30oZlDxFfd@~L%mc#MgHiSFQm&0vFnW_v!Kb)V3x$)}dHvYw{i
zxWNM_TPcN<pr=(`d}=-`Pn4giI(Xn@hZom1#hb-^KDCCGjTMZAg9lD_dNxYm;ZZG8
z;;rQI81<3rdV>c}_6+iBo-^Hjm=bR;E0;lG8PxmxgEN^ck_$`Pq4$DMtza);ByXfP
zc;Ms!@s5_GQ`G(VR2VB8Dj4bx9yoc8T<@MSiAOZ})B;vEP%zLOJaBR(@+<vW{AZ;K
z70Svp<!7o59yodP^XD7qUQxcxr`EFac!lxWg9lEIt1>EZotR!i^H#B~FNdfc%K!br
znH>AMd6}MX+bSjALe``dDwRUv-yfVw(cE3HCr)`}H=LTynv9kkJ#z5C$qhChVIyxS
z_oz^-*)~j8n5;H<;N+fN`|kRtzTM8JLReWxUPm~1;N*{nxnFq8=eyIqd2Hod!LY4B
z^Y0JNgclzCZG``L7X6ByDyJGL52KkZsGkK*WG6Fu&@d00VLlBFPm^2(4U0&=2@P*z
z23;O=97`mx<q^{cjN5>$V#Krq<93oez~l}fyAWbJfoUh0%v~5U-C$Zbn2xy**!BTC
z%>6*r53DhN1hyZs4lPEUJp3RqQITe{!T4-o%o)<mF<^2Gm}71OliR>l%<W)uJINhD
z)PXCe*@jN<1UA^d2UzythS1Dg;QSWM#QJ+M<2`V|{0TUJ!n)iEi_XAUX)5cAi$LWf
z(8Gq8fy!l)i-CSI*|6I{<u<9`11k40%bn}@*{mFrKjf~Et1ND7H$JC`XeJ-%=aY9(
z1@x;(ZU7S-$gXSz6B>bt>k+x<J5!eT3C!7w7$Kv@jcA(jaqy91@)P<ITA+QPgwuwv
zkBIV@d|EJ^X4Z^Yqso0vGvVgp(=kPuM-UWg9!b!8^YsL6aNgjG&)ev{QAR3riuJ<U
zw;E71Qx7KAlQ;DoOn8oON~KUIY-+k|mC8Wx@bj_%XyyXYzW}CjcTY0~z_5U<s2K2z
zN&8zs{}yRq1`Nwc`zpY%BJHaIznbJ)Agl!w@b&6|S{*RNTn~ixr2ZVJJty@BAZ#G@
zMxfS6>P<k{MC#2ztr_cTUMq`JJKfd-n%P4c?xAeC3v&;(;fjL~GcJkJ%trZ*I=J_r
zK(!~(NS3IXcBs})@)o4F1rhpQq_&sjQ%LO;W{ulB6ZN~IquE$Xq^2ZNR$OanCW)Go
zL~=6anoP~c`Yy_K7iP^T=jA^izcOXWhvDeyT^sk0-?i^w&-pqVvm8soxaPsGQc0wQ
zxGRz<l_ZjrDV1cBcTp<4FzfQSSjWphXkaVy()F5zUAstk5kdaC{sgVnSUZlh)VmhG
zqWwR%u?R6<T3%zguMx9IYY{>ITK)vB6|5b}S?W7nD(n$GR2*_=jeAvCs%W`%;lJ)I
zGtcZ+^>rrs|9-`P&n=Y{2j`S{l=`UnXmkBbGrnrRx|p6p*Jp4c9<#5Y+bhUmegnmC
zpeN=JQ1SuJ#k?I!wj+X*ku(|2!+Iui%S1$d*~mQ`d13u5lAa~)uOgSLXg;S~+TB)s
z8tp5Trx`sJJ#}vO%sy|deC+T|!~&C8V1a`^4XCF9ZOr#5bdOTR{2L|z8^y<5L&?{W
z{DhK!f|;9?(~5(Waon%r^5Mi2EP6z=5H~4$Otc|EX3-*ooTA+cS`fV$(_^Uo7;19&
zPcyAh*b2vB{tAU(N#28mdk~?YM8cDpXXtd7hsXtVD<j5SU@qjQw=C(I8(^{!SP+_}
zz>*-5KtzzWz?vXifh|D}0tbS`0x>}nfrOx2l;9RlI*4hZ1TFt$;}&WX=br&x+(Jn)
zr$hC0sL3(6Dq{UJuq49%*&x3`4fkvvbghF6a8JI3ZZFBm?}FkkI0x(Rq2xUo`P-0W
z8zM$R5|Sp7k-rbQ?L)*!I*Qznl97J~Nzah>SCGpU9QjVJQKP&jO<KV|drP5XDV)hI
zX1GNiRQ$(cj%GY9Jc%I?tQ%~?-K+DO6Q(*2g+o8@bk5qW-<@%20ee%qz&ID!b1ri!
z+TI!3{_WvGK_M3?bHZ$^v^DCci7Bj*O4+AUZZhHS!+DQQ-;@Ew+^6{WsnMJuTh%K5
z&svsdo<qrVNaUgxC~bkV+(I+G(4iN4V2^)>j-TOd%yGyu4iTJ+oKg`HAqSE4AkJ0n
zJtCJ}L`2vH<aB|wzmA-*lMz&kY)g?FzTQJ5eTeJh*5Yhf4=luSSg)5~PlR<Xbgd=B
zx)HiHl40Em#hqkW_d`iP8P;2oWGmUIL?lfl!}KTQ_7fS_N09pwGOSM{>1opbGIF_$
z!+L>&^hQ{)N;2!uV9j9JTyfzX!Gu3kZHGZNuPf*&<GdaV#>J8=#WbLt21ateUikY`
z_nJ%(hasOF_(dI9<A0;*CA+rBKxP?eI(A|fGRwlu?U=Am9M~?wPK=a~l=TL?K!L9H
za4D{&8M-w?Kg>N)+ylSC{0T}vL0`-}kYooUI0Z>l&?2lKKyC*RQC|*n&q4lJKaZs6
zN&9Qa<r-Sd>A|y|ZfrT98_3QI^MFMjAZFr~z@n1m7-SNIEVvVCCIOiwU|ty>vBC5G
zaYqH3-J3vdK#6lK&F)R0c9J^)>L9rjpiYvz0qQ2X51>Ah`vK}F`6ECdNxn~^`;_eZ
zA@ltv|Jvi%#riW3XyuVhvPz&;i8*wBPTkmUR5Iid$nCYUfi<wvh65{5C6JijJ%z4M
zp*QwN19WSEKA1b8xPwfI@1W!znN4DlB!*0B2}qhiW`hjmmVt=+vXFZgA_6rZN%Kkj
z0_0MFv&Xt9`zr17U%x=aqybSHUMVByHnP2q?6`&^rUFSTkPFA_3TkJ+)7g4($lVZ;
z_iIsFoc|AOe)ApcD_aI}Phv`OaElXr=U1rwE7am{BkESetM5>C1v{{IPzpQnl#^!S
zD1|tZ<0*xBl6O)HJ27v1_#Os*r`5nVqlcQ<L&-MjxJ^|yOa^vTaYG{HBSv$rq?zBL
z>+g`r7fsNu3ChMl&2&R?Hza25AED$U8CJ1K5{rm1+l8dN$gtRt-1g&q!fpT}_hV=Y
z*9|mt4oT0E_Jzo$5QkCBX^$}GV09^*HVP<y0W}7vjmwnqGG&1I3Z-_1(#L#_;$I`}
zuTz%SDKXZID4Qb69rI1f@+PSlQ=(#u*h#FQTq`JV&VI)mskU`hq897Fr_k&voQChH
z0h%?CT!JQ)AXBWDBeQZOv**4CivP2vo!}0)b?o{o8jKL-9h5qE>46AQjw3mqQjRBi
zC#Ad-bK(~3%v0;$ts2a5GxEsY&`L0)5_5XL|B@JG?Z>QPEhVa@B>0N0)YMi=wvBou
zbQW@-7u!g)tBHx`#&|U`RWOy9>stw|2of<OZ)|BJupu;ifjvRa0%wB6R$^;>o<uB>
zV!A^K?)-E0FsFr@%QYJ@8Bjd~uNIDWjWl}IYVc7&Ghe{SFF=bkq?x~f-d}h(jAp(9
z-LGIA*0)l2TPbJE+bHQaGI(}C{tl>y^*AVugQGCVL$!F^5O!M&3KF3P){~$p3Ce~p
z%_KvIWOkr&ZxxsnC`uvqRA`$Dov?lwjy?>>Vm=NXj+5t~fHO}(53CnL>q01-lhe#K
zIP)6TkHz`yYg&I(o1LRo1N~}pleh^?Y{JQ!W}3l-X1t%utz6iBi2uHW;a+~w%nK-a
z0cUa1PBVW(>7V2bybn6`L0Oh(Jp>)Uz?s-S9y!J%g42*w8gj+@AtXJ7h#CEH<Z_&x
z!Cyp97fJgf<Xl9~z{`+r84~0As*$uB*LSQ>J!RyqPr2;ob)a=1vG@u%3^&HfG{P`~
zph&|=g4P?ZCuoCcgDpO9qiCayPP}nFvUKaoIyRx_0{vVtk@MPFqh@AN_q<8ez^YiM
zg_5~@fTy+KUD2~S&z`36x13>XxC9L@LD}n-^NXKOIhv>}plQZj$y}8iM>J!hWI>Rn
zk|jYRB@sc^O4bC~D%ld`pyWW1SV>HfL`g!>ElTMYrOMqH&9+~uh2(Uonhs@i|4W<A
zUXC3dK9yYn20I2jbKO?Z_FaL-&Yrv>p9x%bNsl_?*z$CseQLq#T>A=@YNhh67fe_;
zc&d8Jw&4o-ud8C$uHyacn}CI@!{*i-<gOe@-(X@bIZErmggQJ*xv2m<MHzSvipS_f
zDE|;@aATmT?D03JZF-xAynjn-a;EYsZ>}8jhTM{LE~DR;Y?m9Dv`W)}3m3tNsiBN(
zDB^`{4P{${=VAl;oEplLwD+%}0x&<JCO@HMFJy03pVb(%amGy6mFgmOH|!yeDH>vq
z?u~Nz&+fd<1N$OAV|-*w%JR`sjS5lxVTy=(1|`pM4n@o>D18NGITJB&pu-#J%=sHJ
zAE4t0D9eF}*^V5y<D7_?WaN~LB-lO^Ni%UyMNBqw$;LSqF=vs}S<?P0a=wbB-1+4t
z-u5R78E<xc1kVVz=Y~zC!KSzv`vuAaNy^;Z5+|vtT2pl}S!r1jBw8X0z?zMg4WZd<
z*%Rcf<xG&+K`h4SNoGkrG2NlG?oc|Miw2UER}1BfIRolsK-q+|%GU7v`JUQ)hdfBD
zc4U0oyys30+pX$+bs@KjSnay&e!_+OqldhGH9FUa)tZcb@UQ2;22qu`KVK6J`5kg0
z#O~V;{Emc6w=}a{xLh_&X(nW3h-|suur%@yk#p*ZA#Z!z?lbQWnp7QDqM3An(gE@6
zWG_H_NzMQ$gXBzrGD+SK(0-B+0(6k%LjWBjISZgHl8*p%gydt8J_ZGN{N+G8hvdtU
zzD)8JNMFIs{eXvc%D}I8uv0u8Jw3SG!)BsM!LrviJqHSl&CcnK8*-=gp6$q}-t<c(
zJES$uHEp;Dz=sY62i1k=W7)4up;0Nc;-=qRAF<Q$|Naz-J7Zw6s__&Wb19f+8lXl4
z$sJImgXDKm;~mK{NFxRj=O-YI1d=n5Mh3}QNF$5ne58?2askpPAh`r7mmq!IzH+2d
zj`c5bKQDn7&Tn9!i*CAZ#BV`-bbVwL9vm7laKjPdK@krI3f=>U|Nec$6?nX0T=c_5
zFI-%Li;Hm47Z(@HijaAkadWjrZ<~>7U_g+6@G>voMT?jEE&0ZKi4QMOX2uJ`#bvnY
zkBdIIxD*!$Z)XsL!j^?DXG`wE9uN(a8!oRff>u&d=DVs&oDEeKT?HCyLd{VkYa3fT
zdk4qKBTZ%)OrNOx|9z`IT6@e`J^gXxCm0z|GBdZBI?YPQQp<FTc%-_E+qaa|(`hD{
G#rr>F=4#*o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_fwd_rank_4.cpp.051838AC0BBFCDA1.idx b/.cache/clangd/index/batchnorm_fwd_rank_4.cpp.051838AC0BBFCDA1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..373223b49cd01455fe8c55b32c66a3dac4149fb5
GIT binary patch
literal 5154
zcmb7H2Ut_t624r4nB-DJXbB~Bsi7!Mf|Sst2T=$ZAYkCq1O!oVEhw&tV6UvOK3N-}
zD7dI7yNVQ1tSD;%fe*#C;`*#8?<9JX;Pdf)kMCo?``<Y;XJ*dKe{RD4eSEal5X39d
zKR-QbQM@dIAS&QLXHj;F7XY&{1j(;xj7;a;b5ed=c`8KbqHH!jsaQ$YZ%@ifth)Vy
z=>~(N=|PE|rk%tmM_YXJ=IdVzroC8Tw=Qq)r)QtciF_h|#??lu7A1JQfc@qDmsv*7
zUJp`QPCM)UbZ+oM*m(Zhtk<QwL3!4tCcbN3ONahc9CB`b|A=LmVSTii=l<~3zq_cF
z_aBtK|BzUCn^5#m`4F+Pkkz91pfC2He^M9DrxZ3fjmQo)G*GLnHFmwi_ukI)C?gxb
zxnWTm!0oxc{*VdFP~bOy+Q%h*Yj=3k{tUajUdxvU)&3UTG)z;*1v%B7>5PnW)wlW%
zs-r8OQg^MO><SzXTt6*E&;N1F`j(vRu26H1_n@uQqRU<xvBvv0+nF;%T-yJdH$$#E
zFs{kB%_7z1WZ}CPpNZ@DHU_xTBRM94q1ndA<xMKhxEr_HwuE|YxBOR(!&12|VbA?a
zkSw2YvMKszEFhVdxRNn;saS#TcD}vJd{wbh-KIZXTGSaX-kd|RyFY~12DZtNY$_}p
zej(}n<+^iC%1-8OON)E?w~4>2)?)PmP7!$5m{0fdQwswU13b+=NOK+PgDq|aEGaKJ
z7PF~)p5-dlXpb!4vfd(vuHBJ^MH@FA8Bs1ce=_&75@$_)$D;Nl?Ct4S`%c7Et&7Bb
z_ui-bnHO<L8BLA)4&D0QrE?C_V%Ar`c;|Y^rq+vpY)FQ@{o=*z+41k#p1dP@{8@G7
z_I!<EtJU_l*Q{4HUS40(+t8VlT{F6`s<CSHmGjAFX4H<c+=R*pQzLq`V|#2hj25Kn
z4*Fdzto1B?GG+KlWpSzZo~WLEHTyjHXCnLBR(dTwyl|hjVXMo=wzSzh8XF85UCBh{
ztZ+S-a@FF24$~ov!uF-Lx1HV$867`#GF$%o-OL%(?Bc#P+~_&x6*@+NAB=IH@nw@5
zMPp2cn%l79LSx{m6Ah!*b$f<9O2X{eclF*VE}i%8;G?LTK|NwY;M;o@^euniG^iTi
zx!BIC#>#>H+MRH}=G5mE+@BB4y3WLUY>H;i?cbY!uA_;vppD=bb;rM?Ea;bQ)K6y@
z-*WXkaA|+_;p)Kd>Si^M`|V%w{-Ix0f4Xq0N!?yMUUgY&NehnjjMWYA4Kf&+deiaz
zY&Vuat8s7t?@u4@Uj2d7wBgj5HY48tYYhYUFZ|6+=tl|~I$o@png2_{(vL2cv?%L*
zwPhjmtrJbhQ~B>+Uh$6$i~G&jpzbf(^eP7*cUIPcl#k1EIi}Cf6%3!domV+%#m~);
z4taX)fn6Q*=2=y%69=wTKJS=bcyDJ?wBI`8^u_~!d|aD6hjv%4{%y>cJf~LEHGv5p
zahc6B1D)iyM$ZTRNYZG>=7W8R%I@f!v^S4$<t?~pF-D*~JUFXn^s2)6=&+NbjZM48
z&qKRZwvCq!hVYtIsLS+=PL-7Lm5l32_aDV<nTG0G^`;xe9=>Ecj$V#W7)5fHq$iA^
z2r^|L+p+bd{u(lpjUad`lBz5ch>&x@A9?%}A(j?nO-6a~2(ClY(GZERN(8}HTumwZ
zcMG44#sak-Nl!~8pxVe{<eKccw;g0O1~`wBk5Ul{sJ1A3Q+?<|pEDWB`O)4sN7CN6
z4EwPakcB{<i{z3-0*2<9g(&YEG@qu3E&zLLsA%Yl1a$7b_35HNayA+eP(HBEKr%=o
z0j=2@+U0*7TQ){O6Mp2Jz2WsJUN<(6jLrjUOdcbO1aux0mmY?DI(I0e^TFOGL=#LT
zpn6_Rb^rU+(KIs31=dRPN>fDws^=@79`9;W*-S>Gfm&H!nJ5xaomr|-ez4ujhKxo4
z=X4UCDiTn=IHAUYm@#ue5zPZ?L#m;kNI>;kTFy>JLIRnLCIK~#M5BlVRF|^)4^tlc
zKUPEqK&`H#t|<~?^+MG;$42j7GCBu@rYNsS5(%i@P?v0errtG45zPhG3}c3aNI-S@
z-vx2|)}6UPK;wWlMV>+v38>y2e&Fcak>5eLL;<yeyaG`qpt>2k{!u4$_8Br737qSa
zbk#)ys#~cpr@goSRHuMs0qg%d(1F&O2Y>a!50KGBV69EkricWz&Q4m8bBHYI_bKSt
zKL4hd6*+`k$ch>b_8**qDI`(v!z3}>j%^nXI|0MlY_<e_n+sqf914kv1KS}&V&cp8
z6(h!;e!o{Q45!P;iUNak2}cMeO&A>vPc=)mhl@<)PHv^A>O{z3cq}273Uh|@^?&4Q
zIZzZ;JDueygNjN3r^Lp}l{zP(nGQCv#4I)82CzckR>Rj0Zmp!=CDCeNHG|GDfKB6Q
zUaO$;@RHki8M>R?DzG2=H4~le3_SeuwIVGolYvWhH+65=#gGF)orMG-sk1xBJx)^R
za89@wL0&-7uY53}0UBBRd#j$m)P#A1MMB<ziGA&hJi!&}7u`A+$Psd+3EY6;TqUk5
zJU4*UxEfMOm&=eshFlYfW?wRXd>9Rx2T~N$Fx(SnAwv>T3`~(T+<Ik6#FMw*iV2T8
zB3RXQWYbK!DWLR3h7v42KvV`zf((g<N{~k)U#Eq#$SF;r{=&0``^n-FScC>dkkLRC
zLOO%NK_(-RCNLO=3zP&>B@n0yq)H&r63}7NV5P1=svrVGfw&+*XC|<KE9ZB6)a>0t
zemiNHy!?vHxU<bepoWSnikhY0*epKKpJ#2b<<q3LNkJZGSJwZQ0Bn^MlqfJ`$i|?>
zgj5XckKwj#+fY~@4EJVxi&0iS>ni=yjnDG3IAc0v1}wWQZZOq=D23<-I#S5SfGLIC
z3|LZVwm~37IfrjFZ~i%M83@zT$kGf(ko%aB)Mo0~J86n}u2#De^HZFFZ4x0#%oDJo
zQJD!tKam1pyU3D6*NvnOmRgpEFbWKJ<+#R5P%tN0jFt>)R*-u3TYN_?1v!maX327Q
z5V(-je@ww^jE{#cD|`J&9>cX6+Dypx7*1o*nDEBKa5{s|gy)6fIt(2q<ZcYtW#}^D
z4TIr&3_T|Nc!BodgoW3oBr_(P87#@gi|r*wYhSVH&R0vdK_M{}Oan3jq$S`<c(Y18
zu{f+~iKj45m>9tp1tcQ8ZQ)>HIA4iRfxQn9m9Hg*^!Y|o$b@eW5$qChT?lVyNd#sL
zGbW6nEO_JlsKrO^ChZky3UU!%<dVQL$Qk;OL@_*<oU0Gna+7}ao6x`)+a~oZ{s<sV
zxcmW6YH_7#fosMUk3NR;lz38f!c*f((FsqBCuPUe<w?;A&yXiYCp<Hrlutg5uM5it
z3T(-@foSKQn)a9598*B23R@B;yDLL(%PZBg25{&?YJ?Tp^LU0%F!G`kypAWti-+o<
zvdi`&GGBAjVjOu|v?OYvqD&T~Pvk1X06@Y-o)iK<*QJnvC=jFCE38gk(mBSYyVNK5
zy<9JAo(S#=X-<ci9T<D*G$$tAcw%cw{PEsZAUg_^B5tTApKbYdvmOiS=6_Eu95>KH
z!ml~lQ5bH|whx2Nh~YkLA2B*RKGW`*?e5V@^J!_n&~?D)83n~eo5O<u%_wzOpC8(A
zamb>4H3&SK5G_6y$S>c{7TR_eZ#2I{eBb%qb$WHkG=$rKFAq40(u)$$T;IFo=w~-p
z9|Zx}nAmtrQt>qL6eCDdAba80s3dzEwvCubFx->vDMo|NJ65W8L`F`U{m_mG_Y5a}
z?xdw34!Aqwyz^Ycq@JGLIoq@7NLB53eg1(P(beQo`+U+3K0GJ7t3=cngUYCzs5?OB
z`=qkTHLd=qmy_C#-KzL3-z`5l=^`XMDJ|X}oDVD<0_RZt0`d;*QOVMrY<><;c($4F
zfDt4^@|YrdOp`n&iyvihU(F+-q;2PO390F6nHk9`NV-^wq)Hz7>4|B=EyB_{Q)K1j
z2~&w=OhHvqMOn!|)ITJM;^#AsLJCsTqN;1tX1cl0^6>QXw$-p^8M|08e)LvTM_13l
n(8$Eh+|tU%&fdY%*;U_(&YbS6p=ml}cJLxJU@kjQCLH+>;hiVp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_infer.hpp.F7608824D3FE5589.idx b/.cache/clangd/index/batchnorm_infer.hpp.F7608824D3FE5589.idx
new file mode 100755
index 0000000000000000000000000000000000000000..37d2d3492da6dcc11d1e2cb691f817b47bf530fe
GIT binary patch
literal 2752
zcmYk73s4hR6o!}77=kxEmRzzz0tS*GF$oAtL!beah9s2)NpO7BfEAS_SQL~&@zGJI
zRvpB*gRcRl+R+L=D%M(SMbI)n0qdw8YJFR!TA#Iw+68*|-kr(LemQ5)`Oi7`-p#R?
zOyB!a)W`x$`SdBXCiSBzDv14+&MI+>cofw<fTGH`{F<9pbuBLNoV~Sn!fa(utk2AS
z+5Ux}wAAjsRKNRWf$03cvCMCWWMjU2u&_Zs?@Jx8t!A${sd{7U*@|bi8O5%*o}7E7
z)@i>SYwWyJTH9R3tA5gWCVJgmr)|Pc(fGed46&AO5w2YKd+g)K3ft&^qi@Xk^L;k6
ztkZS;s=N8^m>&$*ikVX@%db!8OYfeDkDst)=kpznj#W3K{V%0eUQE1MwqeSXhu05}
zt~99Qtp_7-ws^0!U;pUk@S-ls?%EFjx`37QMPYAf(fc>XOR5`&J=qa&_Y|eif8KC_
zI^4z=AMaeebm@^3rf#3~`TUB0ubR?tUhWdk@OteX|M<=Q>A-~=)wq!LXH%=somBsv
zmK3mjazu!1bLQi=2Dv1C+rb<|*rI!r;r4|6U!5uPJ#(!{?<wPzycSG*bGZFkck*Ad
zg|=yJPgd5?kNN28lXCvir3pEW?Mrp4{+dy3JF*_FIcl2s#UXoh`|<p}aZwlQJTASl
z)(|CMO^-M=cX>hAAVKuh=KH&{E3yaOv+0{&Tq+cdTvoJbOs!(X%yGBPuN0papFY*S
zI4u|6ZWGsabv!sVW5PSb+WPiYo$Ajg>RL4|7Z+`P96xqsu4?m~ycMbQPMYq_aBXed
zZeE|&d8?)}x@=T<lKsca?X~tzG5g0?<miu{o4xjB)5lH8;S1FA;$a<;Rqv>$S9DjY
z(mB&VyTPswSis9FD_^nzPysZH?8Suf!-i2p!3R`mkPqV%<-7aoMgM670TqZGhVX~5
z4!-^R91MR<zqSp)V44SlDJCN@L)v?zeEY)H9UJaFSMqf$phW2XD1Ov%O33%=dw*AV
zN6XF)*_#37j~s&e!EA#5ygrA9ZdKvk6Iq4*c>`!6qZjCvy*HHL+5!eld=Z;6fak~N
zH~Jar-WxJgB~zxBng#4$Xjm3HyZ*sRU%)^bLM|-><xoM83dUdxp+g8yv)Zgh<pRc>
zX|`gr=qy7CvROtWLM`SAds|!C^rq(98_9zJGcIskDsn<?W}SN4i=B!Ci#aD1zj2+^
zD|7My6DkOWeGR54z~l&WI6@J9{Cekrx@WCEUUhzPmk6-s0~kb<2to*w0a*|d447~b
zPLL8PgSc@rDxe}r3$#I~2Egcmj-VvpBoWqX@HG&W3Q~j61Obx<(l~P3=;-^aQ}}$q
zh=t++ELbX(24IN_p&|f_(g-yHScXYt3P$ZHOV(Bwl)Lq8@6-Y{*Db&l3JN(Ymp4cX
z&li~bvpJ1oJCV~UwG%mw3OkY0sIlXm`V4(8F4mY}%s}NW&JQ*18JbYWE<h!%48Z}E
zkxCUogOwUgDq2N&y3{TmDhMzxvuhNlJY8NqLH0Zcq9r>P&dvI!<+PYxv(Oxl{$?rO
zoQ%l=EriXIXyMi!V64ze*c{M7QoMtR;)G7Z=1g=FQF37}VRJzjNeQkLtR)ZT5w^S!
z^U|;_ALbJ_H*}K}?<TZ(pog$|5<P?#FZ6P@jZN~ax=o9!*s)7Ak`z4l2u(x^9(%MV
zIt5QARuh|o$DTYmS&zq_u1QZpW8Zu)QD&nawXzE@QV_{4HNZ?1OymfKWw(FN>4>*~
zL0|z_I?}~D^_-B{$rfniDnt97-TpmCC30%IRoQZDNYxTn1udYtynx9RWO9Tg^Za^`
zP!tL<CWT3Z25={i|6C9tS<$mEp*RAG0*n+&BXF}6P!WNu0*nS~B5<=z15HsV-0Q}T
zCv?eY4*Gz;Q~m!Yz6~66ptHF7ZV-E1Nnj|>0;Q11DWhc&=aflgp#(`~gdUk(rXXx8
zSq!FdI-H22WRzi8xl*PagvlCbO+d;4<4`+@?|~!JL3|IKIw$cxaN3;Q_W&@tak)g4
z{J4CM-rW(_*vnl<1^`A&Yawo#R-z3hNU9~;sg-LLgiWQ5!K9;gL=?B$O$c^px(UG^
zore(Yv3Ur=-Z(EOv1Cfoq@E#_mP{_q?0Mjp4@HT4xbF%<4-_TAkIb6X_xJN3X3pjh
M6o)_wH4vNr16T!to&W#<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_infer_impl.hpp.0B7228D5ED433292.idx b/.cache/clangd/index/batchnorm_infer_impl.hpp.0B7228D5ED433292.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f1e9e4e8bef742be1b767d0ea730a9d580132700
GIT binary patch
literal 1686
zcmWIYbaU%tV_<Mj@vO*AElFfyU|<jg;^LB`%mgL|26iq6hKe~`Lr&-2HW0Y7{KEt$
zPsQmUmGuuA1Sn5u%H{J6+@jvN$V@D;GsN@M#)K#Loo#bEi(+?d<a*)o_6<+@bQ!D5
zo6lajT*);@rJ?Sgu<yRv`@_ZO=DP8IIL&m7_ruH}?pY=mgjEeA9)vQVlMECqyPI&o
z;e))(y9YOa-h6h9yQO1d8s`u12{Ml+mA(!<IMvOoeV+Rgm7h}z-!5v`nJMhMm|bzQ
zJBKmXcb%DvJ7ZT(s@;BmnfLi*o0hHlA#$AY=Pp?N=c-w_xXXBtP2ZO#s@w~o@$^4>
zB9>ml!p_TAD)pi0&(RsL3!{8{ehX|rHFwU6h>$GKxVah;+jBGHuFujw==8<#>-Eqd
z%Y?&R%SB6WH!=mMwce>*c%n%{U`f}}rtoPy-#lqH3ZIt2(OtA^%EN^s!Me(Gu04~k
zyxsEW>!F$Ns;{{I{+4qz>RgViZt%88x%S%*$!ecFw_C5|qBZXi_DA>kl&4hH-a4N8
z{Qrc%e9ONo-45>5EuX8E`a5g&w&D=;fU1?!Ix$a_Zrk6LIc>>)aBcb>*(vcBjOMPt
zd+eJ_R(wBNnPyla@z_J9LD|)5(G9NUk64%ZTXYs56<Ma_*{ZZ`(XXuJh{-CwjQ#3$
z0X>TXj_#hC&(Zl{{vwmV{?b!;f1Ft|XA|dRlXFf7N;3D&YcVNMU7{Z}!NvICN2Rj`
z9NKP-;$N?2%O`m>ycS=!;JlsD*`#H~G5gex{wq~=`nIRB>AvKQ+gCTQId^7d^em;T
z5pETZFCAu!hc(LaxJPg;EL@hd_+z4pxxAT#s#oI`b&WL*vpOxG?JT>;u<qBqh~moJ
zBt>9)o1S{ykn!y{J01pR1_m}s7D;|E!NAGH$iT<K$i~P}l$utY4wPgrs<yT~{LG$-
zhs}q@ha0YuEs!M;&eM9hwCv*%cRofQHWv;TKA;6)U}AfO>qcdi2vBbrOBfeIZ<tz`
z4vc3d_wiIVTjEC+7Pf4mY!R3+3tOE~9Wqa=4z2|#=UnFsQ|Y#+^<B=JwD~~GBRM1a
zVZvTB_xGP`e!~m2+>pmm5GKsSX3Ap<=T#hj_@s*E;#qDUHferoxOF^iGW;?~JXwBO
zB%U0<91>5SUml64z^{PBQ{-1f;wkYfA@OWEZTS$6vg5Qv;@NZB!+G6a%q?&G|66bY
z9VIEFfY2f<DT~CDlaxc^$xF&3@f0K#ka&uciby;qNhKtnt+1^W!Uj8GJ0zaHusxjD
zU3c?PtkYR-V3->?8HB^Mbl>2Mn)Ps28&KFr(ncF5JoS4*Y!koCRiJROd9gE07)CR4
zK@uM{)hx{MX7LJad&I=TmLrq{4>4e*`_%=&^sK4a%^Hxq$PE~~mOPelMLcX)JXUbt
zdd+8{(H{eKfF{++)xo_9<eAo)!%W&3bHiv+55Ex5q-@@7c*@{mtK+SM^Uf((cQn}r
z^8*#-Nae`E^a2%y)kVR07nTP<OKqC_7^ujM!%PAu%)^$>kq+lwX<TM+f9~v8U|1jv
z1M^!>A}9cgi;|18K@1KM0Rfya1_J}Iet-dP7z0>sz?n>JYyS$#h4hFrbMWx8b8&NW
Qurlz%Wf@ozOdbRi0Fkw6g8%>k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/batchnorm_infer_rank_4.cpp.16E993641F247D85.idx b/.cache/clangd/index/batchnorm_infer_rank_4.cpp.16E993641F247D85.idx
new file mode 100755
index 0000000000000000000000000000000000000000..981c6c9695a9371f890c554b2b3e64e908bb917f
GIT binary patch
literal 4966
zcmb7H3p|ur8-Kkqnwi%yn9MNDjB5=s<SvxUpoyVz39VA-B@@vWwbF{!MM_0CTd4gi
zDoRS}sv`MNDZ1DsNwuL>pVW5w&P<(|sol@-_xb%i-v4>dbIx=A&*dDm{P_GREd=q5
z@k<lOE{#@35QGYTQ<tX1+XE1YAxK(rO=zS@&q4d4d;5XBv-CN;c$dc!iqkYNdiuui
zwLL!a@@PXuym`F1cy70_?)++Ey!Aj{W5?-?e&(+mig`u6xYu(c3#@x$(j%UzvnXEm
zmaXouoBycp*)DKn{%6y2`q17R`<n9W@sy>tmmLjW^7eS-RSp^UzJHuDI$%WUJ|8vQ
zG|;kkuYbdV44&g>+dZ$EwpZ8PY#E@r*P;DBS;GGQU*{DUoh&=?#!xG&rTxC;wI$Kn
zQ7igH@fG)1Sl_v_@Co+3Gje}B|G7m7|Apy3*OzZk-A;Is9eG}<beVnFuC57>pH7T^
z>(OgCSQ=G8bS!8JZauMM%Sge})-$6nW~Z><Y`l#PvOF8R`a41lbtmmZU++&X-WV$i
zysk`ewTTLEKh{^bob|Ap9=_<r<!HOudS#O^!$psV3y!JyF<z5tF-<ntz4wLj7<cmb
zkJngeT$5~gv9&V$<fX~H{d4bC{k%{03O(4n>eVGXo7w%HhB3K`+#10YR(WBw<Fuae
z%{eDUrA8}Nt9fc#4ZC7v_1c_!Ldyy{p~T!7;+@@1;r2QK>t_1-n)vptH|>beiI_AL
zL@hGE#n|8-=wVJDyjHNXDmuJ<ebQ!|K`!Bn!7;J^Q0}4n*<DK;E4+)uZ4ayGl&lNI
zeD-uPeYu{2r0CijwrwlBEidGtp2vpY2H(0IvO40qpsH7iTy*v7&6J+E_i{cRyuVTD
zv|DT55A>57tyjG2unrrml=<zZZB54-He_{VH5fg;J6s?0dtas1$U1&WdB?Nz{L^gf
z1+O)HxDPwISJN^|rhdG$xd5M15wR`A9#8DN{$OhI_yAhK?bz(uX18{TpC2z;W*ozC
zRWb~Ck)ab*&ttxe`DxwI?8-a3xf#V?_IE;C*%6H`|MF<PQ8!*tF(apaP4=V<UIP``
z_p<MlW*e21<)2&<*u2ffUvKE~87G7NbQ973kJE=ra(~>Eu&06hsi-P9=0Z!8^&z*(
z==)9SX6^O&Jw(QzQ!lpcdZ1pLvxB6qb~*5z?vq3BMy?c2cSte&cb@u+<43n@yY6;2
zHy>Fvba46RfI>e(h+t>2UpgnNc7uM$`D>mgAAU()WfEswZ}Gl^k$Fjl`(A6vb+^|e
zYFu*m+xMw@AAYE#9`)VPo%3L{PCRosgygT+s^3L1dSh48NPRcq;+pNFW@c7?wFP}y
zK0VF4eICnnGZR02KJ8Shx~-wB@$G=I|L4lk7umlruk7l4Fz9kN_AsSw#vxwoIn8s&
z4>V?W=Pgnd3yb18CiW-p>_Puvl?lZ7smsnO@0xow_G;U#8(96@pM-3mHl0pV;i_@l
zu95|1E5>=@3I0h{dqqEe+<G!>RQT7!pq~?aGxa_RUo08Fr=rnpZSSwt!pxg*Xyu({
zWt(EgXO?Y7LW49bE{R+1b{anF%NpX}$|?V8*r9M<oKtGLcz!vGAf$us&p6HVu2ay|
zc?ixjVcALr=r7VAWAkS>l+TY`BcqWBf_oFZsZs&eQ>j&t(W*_JWMm0Yhaq7aQh`W%
z6!`JpIQ(bA(GRo9=xm_I2pCZ+z9<ode{Xeyr@+OFj3ob)wST+$;(#FSUNX919)i<J
zbS<fX)=`u*pPOr`TgYexuvRCi6Qu&G#oPMd<OZ@2tD@0B9gRejqyoB@%=DdMjdmSU
zMdt%`1QJ1#3aDOXW4iJ}md!LW8U^;&BkAc#1zdc!M(x^<DRmhd=np_`Of<%%0;<<U
zJn?@Ve(5k74Fzf@i8(<kpgJ#E^Qc3O*IhCi0`{g5C?u(X>MgSlRE|75wSbI<1GO4K
zjVKjRT@uE7$mtHOprA{^;v5=>B^6M8hA{suSLa+R8Jz>v`XqgZR6uoI61AkZ;^}8K
zBnd43w{u_UtXf!ke}x4ZjRDpS5<^QWp!MzD$LE(u2JKKo76a?Qk=k__DNVI*zd=D`
zfi;avqe}&}9{yPAP)%LesEW=7)|v!OeW`%zPwA{j8-{EW$!IXJ)+OoEqynl_V&hVs
z$pC5X@>p1t)UAX{9e{C$iHwq@2@Lnba8r&c55^0#4)-Wsn^lb}ePjFWN|f{!U0Ei1
zE?@%*KQL1`ySsf~4I0@gDNA!4CNHcIu&<;%ptl$v#tpNBlT1}kZeuRe4_3l(KUF^(
zoMbiGg%&dZEE5cKF)ker+qzyM-A#|s2E!~S%MhN-fqO)a`sljrcMiZB9EX0Xy|1e%
zOPz<|R+gR@hR@yf_R1-}rAg>mlcxKoF+wdYaNq^af#wEd3u{`1^i3^exR1JzCL{wO
zZ66&4WZ=V5Kt?{M5J86i5AOxnP5R4G;cVcTq~k!=V7nMclC}wX0$A`Zgt7#{?fLdH
zB<j1;k{<kg<SV0)$>0D@vu6cURmK7Si7a)<KR`5=o*Z$A9CcWXp~H!RZshn{z^EjW
za64su8a7P_BG|E1N#ArT3>T@36sQwvixj96F+@yA7MN)uQXo=fB$5#cbQ4A9aAs8N
z^uv3$l1Ik$XKq??(!6tLdjXLYDuoW)QuIu}xt4}o{~FU4mL2+~c3=>`<QXip#g;3P
z&fCe%?<xsa6tI;PnA}!^elL0zS248q^pZ5UHs<wPA61lbmL1C;(yolN$FqqFh{@Jh
zKvry?0&-=$L6lm4>+F`b^D@9$xyIayFeE?rtHw5%Iouudtjk;`?~F-{cLcUUszI{q
zU71>gC9ThN2euN&|GCuQMFO%?a*N2b(u!{-f?i;_6W>XO;HVNwdZ|DYc&UJuBsU5i
z#&8~=Hw)TexHI2bhH^}f2W8|BKP6x|okQor3c+v&hrxrb9>aAxx;)r?F<g(M$AeyD
zIFrNV!JRQ&pQF!%7fkLeZ>CdQo-SYoqhdO+FTgAU{XufG!Qf)JP+h16(-R<?kfDIs
zLSqGFD&#_xAFz2ee0inKm;owaN4yD~IaoO6B-yn}`sO+Xt%~8JP+>k3ePmUH;R1Dm
zLS+iH1qziZU<efK1O@_y$`lw06e?3NQJ}D<P)}$8j|Ygb5L!XBbFs?SSDN`J0Hu<?
z0lC}VZ?7MK3^svbe^q~3{=wd(Op-nT<O#uLaE;^%Ve4fZAV+RqZZdRG!?|b;c>#S)
z+A987lOEnvoUaUsBKlCEtAHA!KmiGfLIoruie%_Wi(8Wc={#r5U5?=ohc+m4V?ZZX
z?3gvx+tx=PzncY4u4$ub506;OFHAc+tzS~L|GV;Fo`QOjyn>)DhFkNkL*#_G@LgmG
zmhjhp1^GYsd8*Az>s^12d4da#mzxjpUuvP)H%$$bUc7K6SaZN9$=gcq(`5c+*{y`(
z4txh0s_)%+wby)8HlTp7%9jPd>6gviPU2|OcUM<iubD8i`@u3W+nePrtErp8`_>L5
zTd(}?%>>^AUx%TW@Tv!1O5SjCFKq*D!{p4l1-i))?AqXAC5ecv!+(PwJ8l12^IB-=
zm^}7w4|ZSd$c`K{wtX}1KXck~{=}FoLBGK<ByUZ4IuIXngdy}0#D^Sh2ye5|5X5z6
zXuT2$qQxW&Z~0m6C)vx;yU5_?TH7N7V|w^*fap#QKK2U;1dUCzgp~e8EpSOZwrXfh
z+Yh&kKN4CAkH?$@8#Xwd<f#k~i{X}h%OJVqj(kTMN{LO3o&soqRfRw!_!ogZ$Oa%y
zNNUQ0RPbs?lyKsMuJSYa2_35OixR|%aSM@!kXZytlz%43KjY+|V)^GH`DeWRbAdP}
z5&UO@1c^oClvN0-<B4QUO_M^^Q1=V;6Z&iU@+WDL{Iwah3A%byUEQZm_we+x*0FRm
yaW?1t(<oivfMsZ8Y&wx^G1<yyimii_3)_*&v-8oRo7oEjQqUP;-ZPbEA^!z6iZmGj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/binary_element_wise_operation.hpp.D7064E30A6CF6778.idx b/.cache/clangd/index/binary_element_wise_operation.hpp.D7064E30A6CF6778.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ffabcaf6f933f2b12b750f4ed7a08268cd53635f
GIT binary patch
literal 9884
zcma)?e_T~nw!rV&XNyGVa24fn@p-x2K<*U<q(IRColz7qMe)tZ)YM=~sag4>%p6V4
zCyFJ-D5ND<bO!LDpcRackWPwJBKcb@e;1~rkwNA&f6VN2&t2{{{`KzX<5_&~UU#3p
z*Z#HE+GdR$K74hQ#WHmA$hj}(zA}+pES9d~-|SarP4x$h<xRz6nY*s`FHJA)4eye6
zt9x9x-?d&V+<E!2r$2KnYV2247#|*%f9K}nV=ry4zWm6;VN1R|U~gWU61{K#yuE7z
zC(rt1_oT5uZTT*H$#=hgb#+$i)W60|%-MQo==QQ}p1QiW+=UA^AJcCJX~)|agk=p`
zYwIz1V~$exQB!`$u>L#4*HtVU`FPX7(D%QH8rPEF5w$73X4O+ie+WHybV}yHd!tIf
z+){Y<`*X`*4y^fhb7rT19m-3xy!}#?Eh75uU-{XR-u^XHiZ?~15BPJ)%zlwk0mbT+
zKBIrnXT7=m+3Qyx+cK};l(!~!-F<Q4=eG~vermwGFK=BxKE6*(>9#w+)Xd#^;%?T=
ziQAw5aA(9<or({Pnl}6O7oT10Z?RZg^a)8<+Rk)gp<ONdFTj5R)h8^X*!YAzx%bi*
z=a=_{N|5=DUZ}VVl{i!geR^Ku|1RItJ&^V2Zqeri%<1BNU~!LY*&mO~a%7mV$YVT@
zsk$jprRNvr@p||iw@|TK^b3q%5GtlX6=&?LBdd=7@TpJ<wCI^UQ}sUX7ms&dNdMrC
zQa{Eti~bGk{0$583ClwkOJ|D(ba>>gl{we@H19C$=+FDBy79m|BDMza=KAEXMRAZt
z_bA+>s6K%zy;eN=QA`VfP!Z<xFzyizH3h2l`S!7_hq6X|CRBngdJ*s<VNp|{N{pX<
z)(4+l|EEytV$mBJZxpqd0#)L_x_6)|^7_MOE$e}=7qyrIRT368tavGxZxbrrEcy}P
zM}&$g?yDSF_+skNH@65Ct8eJeFxMG&f9Mh`_WgOzn3pG-byNed7Il~c>lpBtnB()F
zNeB=s!tmXAH`V*V;(>E_J`=g?twL`N-goGg;HrfCuSgD^bLx-orOiS?n5>xbV$oSs
zU>!qir<~l{QP&p89u_^h3ISK4i%+1+xDgr6x4-=9DYF$RJVn)w2NsWC(6%|?Tr?BK
zVgfBz_+l}1ra+bd4&U>_80$9`eyo$2Aq7h30wvHVj2R*(yO<%rOL{$_c-K3|Too;8
z0Nx<%VG6A0={3h&I)$bf9TtOE&v?D)k10?kC-!gIrPZB2HyfMHvsK-AVDaSkPivoq
zbAD!HI~eZ}HJAcbrp_9^_g~9uzVTyRH1<d6{38VVgwa^>5sjU;V*4|RHOqZ#x0>;4
zQI9FGo&_UE?bB{eyCGgBrgV+MYlK})fhuoIF3+>QI>a}n%NZ{hwU`1`7EWxrTl3@e
zexepJr56ETB&=czR9W=(rc)oyoBWti5r%65-Xv5^fhzxVJ)-jH=c&J%DlOn@fjHe1
zsFK&M;F5FI3yaMcEd#zx)M5%$S^jSQyw|gm4daQ;W*VQSdLLN)R@&;y1!=$hV7{at
zc)h5>6sWTD(RZJF+?w)@&K?pq<k)h;y$>udocq+Wg3H}5iyFjykK=Kw_kqQ$+Wv3m
zYfD~n2QpQ3^cn<Q1FKJ<%6o^8uRZw3xN@@>1;7hL8%%*JC4K?x=goe}utlJEI4!!>
zWxc;<N-Lk=bL`Vs5Bo6?!>@K$1H2D8^re>*tR>rmEM_C!+^y=y1FNr0+7s|%r>|b{
zV*%p#uXldk>V3%HZy6Mrz2y9-_2Tzpy|gmkDhAyYSjU!|Cwo46Hfg?45$m>q@dBY@
z3RL;<`Oc@yM&Gdt6*0Ri6kZ`_mnl%Cu5_xpIsR&}P!YzS<eC)cePHqDS8qP|+S~5A
zqBz)RhxxAg_w8`t+ZXiYU+awdAlBL;#t(@$m;!4!xWIpC<NBjhgo+rCTHv+9*rq_0
zqgDJ?-?%3#go+r^9=wO@ePHqN+D#FOD@rc-F+X9fGU!|;M${D2U?+a)_ulO#8}|9J
zhlJ8Iou9FKA5!7#S<g*+68;pSGe6Oe3g8u@9j3q+oeh|?bkBcHyeQfs%vGfDB4I95
zph`>3ja@$%b~WtL#b=Kl%(X*q*cZDlcq3GIUpKpZ8u)4P8dG2$7cUe&pEzyj470mV
z?o@T-fyI}0KRk8g)0^HF#fD=fK6*R2+U17*W7xXKW)~d&NhpXRo5^RY-Ut4^-GA~$
z>vwyM76@ym@if)@z~Z0Zef4fx_g(|M)$8|w-xCcq1*+V%>fe9;=(MqZOcA!MWu0qT
zH=mI9w9LwVZgx>mi$$y5wA<Zj#wUrYZsk_hY_i3o>Onk+P$&;26wbp5xwwl^ERQ9W
z$P)=A^JGH9_%K3a_!vSH_yj^ZJcrOEK8es&K9$f6K7-JFKA+G+zK~EJ@H`;A68K8W
ztAVek>;dkfTnW6Aaux6@$~C}iDAxh6qr4aRUdsD{@27ki_+iSYfS;n=0=$KCJMebO
zH-O)ud>8m#%1aqv%4q)?UrD)$@gmA4jF(U@W4w%V72{QucQC$#avkG!l=m^dkMcpr
z4^lqE_!-LQ89z_?3gcHOw=>>Od6B{wDWrde3NNI*TH&iHmngi1a=F6GDSxc+k0o1e
z$9~)I=-C&5>$z#U!*L%*e;=67%STTGy?g+lcaQ#(?c|64cPh3&YE<OPbz?>7^mkJA
zoc=k<c&Ms+Zg_4KDyjN(b$SpU*{VKUogIYROMo{>z<m{i1ktWMZ-jvD!q7kzm;r&)
z=jmfWAPS6V0FO*8h5(=Y|HYP3Yd5c~?Dfdtnt!w>9SV4$9~cveYK&L_tHD@6d~F5-
zXby}3i~_Iwj}1k?Ulf?$J|4p8?Tdm&ud9#yANu&B!1V0#7wFlG0@I(zBPzk?&x=7a
z{dnwc!lCq*%SACgG*ut$92|q*vZ{}8j)=h>Th-H@>4e5P$Hkx(#n^<+j>fT3_4mN}
z9>iduRDBIN*HAtN_G8c+8>;HZ!G4_bS#X|(7(A_1y_DHYS#LC+sxMI-OO!t7wyXNz
z702I|KIpDuSRxEy1|w(y3`RtO5sBbvVGtthVg?_mkHLq~H$x5_EetsVj6*u10tVsv
zt?JuA-3CF}1XbS->UIc1KSR}bfpZtQ(9slu8#o)mg^`1*H-oboT<8$1`U$X~fCy=-
zb6`IQ5z;i5z<CK==p6_60w3(bh+dUFj;>m_ovZp<#ec1$q8Fj+8x{YJii$H;)vFZ$
zDn&)7Ox1TP{yP;FZHjY3n8Jvh&~!Ki;%hTr!YPFlUMwUtSi*&bJr%z&BO`2T_U$$C
z-`6ckcwo!HfJx|^VG_>&Kh0hDa?U69-}M!-+<=r+Y@qNbtZBNOv}v|92chV+ctR;@
z!z5+IW+Y0=v}WpZdS}`)9fTgw{H>(%vEvgZ&9u*qlnY_zh?!#~6@j$~bh#{wz*Ynf
z%F#s-PdTLshDm-OV&8{E$w$C?1avt!j)3h5I4C=hfSYpk5s0Upd<2G4&O8FYmwXbU
zPg4EUU_A}GoO`FiaT=m1yH7(L<x!_0OR|SqJxrI&*uxwi7Dd_ZVR4j4d03X@4XpPD
z){pY24YYi*VqL80I46w|-mzGTqU>I*#8J*ztc;VKuY~3+;gV|<YmK7g5->tvM~xCi
z*<GW=QO>AQ#!0SMLhF@qWQ^;@Og7?qTtXPni?7Wn9+yz==-+&@_`~vNMZl45%?`$&
z3ku1$6LM$A5*m~}MACY&u7_YeDU1|Q$a=6-cCUw6%7fO!5Xtpmt%qQ_S=B>GJ=iHb
z>Y*29S3Sf~PO67g${F=AR`M<g-31QGmmu^KIB+vF5(%N@%z-%ugHy}dFv*)(;3n2x
zay7G7vtZ0@80n0VYG$YGsAjz=yQ*0X<)mtsN;#vNjg`EO1#V;Ak?}MTdz_hL!95NY
z#E_Z^71W%*>RH#Y3mLE3R6U|=gdJloRgYGqZ5WlSdb}EM!$g8`iMr(WlrB+jOzuEJ
zqjD!mnjSViTDnWSK-~p4Y1~~9y$kwF-VMEXL!#tHP#eL9`?GN}^=pJ=$<3fPgH4`H
z&G2|LjF5Z@)JtHKgLes{FF}9Fm!bD%NR*t<?D?!Wx^AkT&r<VQhU7xlvyizlk)Y~@
zETxc*l)Q%7*RbArMyvWArru#TTu!RKRZ+JpHayT?ESuX95;go=alEv#A#E%jlNh)>
z#bBAKEj+q#T8RQPp@pM}2`y1zrm@iTLR*LeGkJw)8+KY0m?<kPz?79JFcVc+)w1hp
zJqI3qyT@1&eh+Y_$e9um77;1QZFL9BdE*Xo+a*PZMmr?MS>uA`oQeyHvlAK^H&{}9
zXuLyGe{26>^m0YQJEXszkh^~@p+WtJNSbD!79r>2wBgg9kW>KH0tm)Lmyz8HDF8cV
zcLBsw9#jBBBsYMy0fObsY=Dpkuv2z4KrhO!28f}Y)BvfJGa6v5<VFZ>1PA4mMi?r&
z39L;JjH7L2D#Mx}l5$iN^rf8M1Y;yOLufNNB<C?}9t)N;KaYjvF*{{P9_vNfmB(Tz
zC*`qJ${BfVtmNe^a5?KPIiE%6vwoCE<+E{;3z)Tl1<RFIz`_bxB;}|A)|Ya60UIN^
zko7KPag;L(*;vWN%wEhQ<jO5(!;9Gyk`FQKAr_43J0oWoa){X}yAQEg%7YHEA(C5}
zwN<p=m!&jPlOe6lPTA4QdQo<@vKY!qtt^%D$X1pq`5Ftn#=1*xW24&GILhPO*dHW!
zFlz@3mOE1i3+rH!l%qOWU&<po*l5XjSm+(*knB;c9wk`1?j9w?qu41sJW4OhE{_sJ
zImx4>QXc71G9{NPfu%}!$s3i3jY@CH6^gY&3C5k(NZN)~D3O$-DwMvIM^q@IC2v(i
zw<-=~+!w?iXe8t@rpJRxd~IgnanHnLfjD-J^gHUWy!_G3@xQtHj>wAse@{%@<-{wD
zOg(yJn56eccS9>6okn*(wt<62cRjX)Lq>PI!MPivu<6Dbn$`#za_4A<xMoO@EB^$<
zoqz=Cai4=e=b)eTm@k9#GDP9=5#YNjKuh3`DB5pc5a8*tuB6Ym^Y^?u;DJ~Q?+rwt
zYup!Le^Cm98kYs=#o*Z?YBz2NPz&E7w2i9))UF&oiPyF7TgADEG;M2@aEvs$h%{{*
zm2jLsTtu3-DkU6;pNmM-wo|&vxNO90%_{};hjITfdh5MU!1EYu5(Vb<0iNKsSLfA+
ztY{ah%>PI}?Rf6|fnfprFBmyt%^E^-!deiaP%V^DxE4;xrMU>jYO#b8wM0V6S~8(w
z+Au<6v@wJxXcGwKXgP!?X_E*|)us}fq0JyPUz<;8p|+4v9%y+$cqM2nDX#`?HDwQI
z9?F%VRZ^}3t%`CDXf>4UK&zv?7qq>U_k*^d@?p>pQ$7XSDatLNwNP#ct)21>&~8w^
z3))@EOPRKm(f%`SCFLTf6;Uo>S_$Pcrj=2yVp<jD9ZcInxsGXdl=m@hALWBgJ4pEq
z)6P&n&$RQDuQ2Tj<#wjEQ(mNKixkqoLPaa2yjsy#Q!Y`o63XR@R!;e2Mf+GX#u;Ln
z88>mb@78~Q{0~)+cU*j6k$raLa8pz3&;NX23&JZqu>~1db-0<}*)8feFX%AKb1*q-
zepY7B;s<KS9ard^w{^HB;%%KMFs|xw(>ngp&CCsfkIYuMev$RO$RaT@%=KGr$SszR
zhbz~YDLt1dk(dO)dRc()QV;h#yq-e^G>1?$?)Rh`J)^$*d0p0&H7kUx{~u{}JYU3?
zYTo%_rMQQR-k3LixU-$t-qH5-IhZ8!dKa1hMb-uTql)OJZyQU%T~rnQ>wAkOVE@F0
zxZ=M|>4N>ckaTe9%`G-~z`l5sD2B<rLB#REqe-;Lyf?(AVoV|mjQc_~Jl+?Iip+aL
z+_W((CiIO<LNp&H;KbL)<seqmzV08p<F+{?A800C7Yc3jk`Qah<_e3NH-tDI_xe>0
zo)tI2<*(`)&J2>y9`77a(%J8W^L?7IJ_+`dG(o+A_1wVRG!vD&fn`X}SJZsPhDH|u
xhd@#56&r?{;&yWSMB~_*J!{hJ2;-0Yi`HVX8Ek&^4{onn`q`D1@yr%t`8R1VMQs28

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/block_to_ctile_map.hpp.F32EBB9EC094E255.idx b/.cache/clangd/index/block_to_ctile_map.hpp.F32EBB9EC094E255.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cfee4b6bad95913549e9ce27bcb1e6be5fb43b77
GIT binary patch
literal 31166
zcmeHwcUV<N^FQ~TJ#9S~u7x7PrRW7wkRl2Sb|ZF;rdV!l(Zr}R#S~+r0ed0F-V2r(
zK?N(=#R_&*6l3p3MPrR58vV`XoMG4Zzu)t|&-=&sd3~PT@#D@tchBzZ?Ck8!q*JRF
zEwaCJb8FVO)zDA-ebt+|xw(~x|Au_^<tKi!o7)&qH@Be+Hg)Ow+4;J*+d~>k0b7GC
z#nr#KG_qF9`Fa!ES^s&1J*vnfSM;Ay+cBV8zoM|eFMeCCU#o1#`42i*9pCQpgKvKv
z`fXx`g;&m}d;Ku>pW9vgj=6D`mdrk})b`!T+3CH)KAPA0uyyCwya`^>@6}G6Iq~R4
zkNJE03_8}qZFSnbka=EV2cBPuil8;$88Pzw;tuIiB|EO$yVcoVrT!PqZ10y_vT4q#
z1w*4hO5X5yQJZJ?R+rxy78t&GRN&-~n^u{W^nA{M$#pukZnddO{Us}=`CPw4^ZEqo
zzu#^A?kAH@UK`gl?|Qo%UHT>}!Pn#GoSZkm>CJUdrhhxHd}Lz&ru2d@(w?vU^~XiM
zueP1<(JpArr?cew!6SElxTxlBTCU@FbK7mn{9y3IY1dCCt2bv(uGgc6HDT)aV;*&G
zlW}Ox;}?A{9_ZU6w<Ioe$Mfh{wqiY@Yn9};Cg)l6{|b3=!{^U~azg##Pi{V*mvTOQ
z=L_=H&yPNSS-a}wG2=^~6j??*=~o!C*}H2yx0XGR7ktv_h1>HXi-OJ$tSX5;M~ZW&
z^_%74y`bwaKb@)+e6-SxM->BR2KjtYe0}K5Q*EPa+@z;U7N6`gD$jfN+D3^zPfbX?
zwfJIo)7<!>?VgPOtaagl-yH5czOMD?B@x=KT0dt-*PG*?eYK<Wh*uelvU0B6I~sB<
zWbevRL2kdjyL)&2$$!`b4?CAmi=O*KLbv@L_x|Hvsas~C|I5EaZ|3xEJE=~;urI%_
zGvLGNbbF7vgEF0UYIZnEbJq;7+kdTWec~25V|>b>KdxoJR)*&N`BBFHOZRtqU3uL*
z;7WR@3Wsh9eM0wb`EW;@dh>q_eLA(c@i#+jrmXzXTiN`f*nN73H3vT~_io(xf9GGH
zw&j%NK*!P3Jul7uq-y-5<82;27JA<Ns_o!{gjIn}PRtr{<L;%o?H4WjVD80T=PLWh
zB_w(z9<K4r^EEp@-L~xhnCwR%)O|27yXt<a`p*YTu8nM4>&m^yAGDqH!M$O>Pi{Nw
z(~Vb_(<94vKfgJ(Z-coxo{t-x*!h0*U4LEdo7aDR&Bs$tH#&LYKyrgNjav4}`S|A{
zpM2{hMb-=b`oquG1RS{3u*(VmnYwSn;mbA4y|uf_xwiGkB{utT<o5eNb`0y&`~BY+
z9q%8uCLr!r6ORjn!^Ybq3$8nc`7Oz-`OfW_jFHwYZjCn`>)*Cu-_gR1Wm=Pfk81CY
z_-fwyg5l*pdtP{X>*%zFeMUCS=$EwYRPBEzHt5i^T~d#I6$TS&<Jz{JKJ@#e?I-;o
z4qvlJd#7Wl<?kW=C$+05y(f<x-u+Iu2K_gM_2}5O>5u7YRaczZv_CZS&NH8acdJyL
z=C?M@u_C^FxsFA9y}Gq?9RIwvcV^F8HSV82cW?6UD#@?6`YhO7Q~4qy+|qA?SW>0V
zi!0d!6S{1lyst|}kZn(HjXgg+*WAa{i$7U@+lV0vTlzgY-}AHY|Jj%M>-h3nO^T<*
zRjt=#(1_vpU*B)Ae^~!+p`8QXdj83yPYab}v*z13K4^B`=UTI2cICHT{?+2c2mSa#
z@sQooUw?M={QIF9H(#xNtnK<B>gU!k557Gi!28{jalend=X34#$uYZ<E5~@Y`d*#i
z-;q??$C>o@jEvNz^d-@C`jqz^pmj|eQsey=bH=By?l`vT-_z<9)tfyve{T4P_I2O&
zy*fC3<MA1HCj=czn^msUvyEx<Q+%gSozW)$=97)pN2C@Gn7GyYPF8qQ%B{jSttQ5N
z`Ag8bi0G_p>q%mKrDdt-p3V%~`((eovc{n0r>gfJ)PG6irhPpklcv6RsKqZQtoa?5
zJ%8xEdtRlvYi}jErT+BGoJGG@zdEbqf&7gNj^@t#ywa4p@0}J;H56`twd)f3eqi5C
z0qZ&qT-4o1%(?MhpHT%J&%gE2%lsqF+82ub(w`>w9`MiRb<u%sdY@{&>`c=Kn>TlO
zIrcZ(3a>Ovuefidu%Et8xv+5gt9SPVjPUQ=_g0?;>vlT!236?%OTBqY#g`U(u;+xq
z$A<;9b`NdP+$+=L3%^AtQ<Hl=zTaf^%T?X(_PC#sG4@_+<GJ~*s%*-R$-3a%a^@@P
zbaBF%Lq}HS)O<d##Xp~(Y?P7SqFRAXy>h^-^YPJBFIy_s`i#u;o;D^d|LxDGdA_rB
z{otUc9SVH6-K+bzN3L~h!jSJi?Q_x_Hj6pyV*7Zf=eY|p9o(G1N&0V+%_MFrG{YY;
z=vDDAvDS_5!ti!(&Qo%QQ*w|=%2*!9+)uvJppQ+c4o^vy0#c<YlOPM(xnHR?;BHS~
zY31h3QAm#BX%b|S*3KQcDspf)FQE$jP_bk!mTH>>S(KR7NBk~a-%>%S2`p>ns%zys
zCMjcS8{+wq?*6`);NRZOxrO*_Aq`D}EHr7?*fqZ$$?z1yfn`1MT~GW>f-D}EhT9i4
z2}<x19G%^qS)xyt7-y0)mM@fY>2d4a+=V)&Ls#K9$qR<gMUcfdPLFvqt5GJ5E_C;V
zte=o=CP9|+aRD`E)|hd`SE%~7G2YIRf;v(NMF!91o1DpWoS%9MksaNfKE6JFr5|O@
zCjzUd%sT!`7UEmFIbX=i3pv0fZd!S*k(-;vY4K{-e|(q1&pLYw-?w&i-cszhlqi$n
zbG&|8UGYepS5M0cjiIY2WqMNfHc1)Fil-%!;`QO~LRWaoV?}$c_?iS+yu(`wd%}-z
zb{AsXx;dwb`ZUqqBxNjnJC!7jm8XIryyNDauLR78z8C~qeBSS|ZSkMKUiK0yyPhKW
zOcNTJ1X+A5ou)zK=l||5d<}j1L!o~t<xPSt_F9qO(pCLORuHPg6y}Q|`7nhBDPvjL
z+GFqM^==unoGtof1B*eBB`{}crym9;Zt@ZWV3x;7K4YXfla#UKtIzVnXZ&On`axd`
zHJ?H)(j;XpznA=3eb}oNHlYbff1cu>r-YjXS%PNj?!T3HtoIaZ!4D-%zR8lGNsuM@
z#)RZYLuNZ{!T|W8YZ|$x`I!V+ss=^WeeY<|U6_DQZq5>mUSe@K39>k9Zdi8j%*!=i
zf+zH4qvEqssbi8dmhYO^U)MA9SDWw|{Lp9`G@2%u1X)66hAgXDV||EC2!$WYw1j3_
z5=??D)&IFR-rsU7$xEmQGyPijc`Zko1X*hKzH$Fd!h7XmJ%ey}Bptm<Kd`yh?tOkf
zkL$I9C04qADx8qLVEwoVvP8b)cXmkjkP5Qk3-W(mQm#u8CUMi;6{%GIqiin@zWQ_7
z9a!)nP(KO2KM8&&!RN%h^{nSE|FidOLKqBxy5yHGg_)#`<>Zy0wr%W~Aqze&%Q+_r
z$|RwdNsuM>xYzo)=+B=>ZVjO~g^E<D_?yH{D_!2O8e@OISgA|nF>xT6Fm!V*A#*Kp
zCc)>_eLkmo(#yU3ZNi5z0Xdpaj^;2)8OvYMzWbb&-gXy~K#4si6`qn{lORj{!e`yS
zuJT7k5Gfd4Z(nb}(hqD-*q44iXLj*vPay)Huz~n)05xb3WJ&yLx3DkPW1q1=R*62V
zV1XE<jAcpOBI(OUwP4_2mS-w{GnFutAWK94A&Y*0HFU1K5C_9QUDT(8Rxn5z%h2{u
z5_@0q02WxR=|oQ_?j}K&X3=t|FBdM)1Pui<bWWn@ByW?Hu}tky^vx&29?&fyFMmno
zFUiv+Wh^TK`$Rnqh;tXdfRTNmSRW|CCP9|w((IU|KNtK4;{reQn)tqkaWP04%jB?O
zzg#cyH#F1~!Dk9+D1#tNi)jzs?mvF~4J>&W{5qj^>Xv?BbDJ5vFRgA}BhOQ)?OMx<
z{hAVL5@c!j{ugJyxiNK@P51`nWR!)Bvgjs3mX7<UPVP1M6PqDQOD!QwL6i)FEbsUC
zPdj@lxIf4Vj7fo9p+F8c39@wC_}OaOYcp)$VK6^Cg$g@`0Fxj~_nN1CAFa=84Q~VU
zvxN9BA>k$|W7)sKedL{1qoBLMvR)+XMcpLG;=EGy?R3YM{jl6&WVcIXJ8YQ-L6#l?
z(-JP%`(lrmP!(Q!n#F&bCDA0v(raJ7S9y<$zA!)VMz<v2Tc9ZoQpR#;>fqU#UvIVv
zgJ5*iXiyrhXA)%TBYGEg?3cOE*d9)ZAtykz4T3Cv^DlOox4LtnEZAW+za+{_5@HgU
zZV4&XE&c2xQeS;Ed^9XF=*?}(_crv#Ao!e*Gt_Ffe%cpcL4ha?sxT<1^rNi#)V&q^
zz6}>);9&S?3uLySn*><~G+Gt%^}(7C!H|Ga`P-uZZLygISq8Rg`FQTwTVH^k0ArgX
z`KCZW3{u82BkG;Ju)n{9%>ia;CDB*HA~Q%C%geMLU;lCemTFU2%}W%0iQ;1tWEmt!
zbWc1TalM={0$Bc$=|6I1la#SI_Rl-|dEzfHL(tu!zC-;=Kd|}JwX+U>e5G||SqO%4
zeQi-*TS83Yrg^9hO2>6@<l2de-xr@SRL>gGXAR7cLGU@BJ-a*a&a~n4K&`{ieCzwI
zU+D)nf1a}a=<s)iVCV%5yO*z*U+D)ne-W9O+TWi(0g;CHn5_6uhCSIJ$TH-w&h)$2
zQ`*5M2LpCd4815Om;_nAn%}ZPwAY@)HsKSHw9A^$Wv!Y?%2-xS{ps-MSDShYb%Eu&
zV!sYc-5|*F&5h88Umw`NP8O=QG_Am8qOwd3agE1&rQ`9<)24rTUebOyRL*?CXFjML
zgWyxX?eexyz=WbgP=BD%b4frhi82YY{Lr`m_4HRBk3dR*WfSq;1iH;2$TGabI|;82
zX<(rJ0NuT<Rl2QJHAxxEXOq52|7_9-PvI+IxvJQ&Ds@eQETcZHmL5BB>sA<D5b8;y
z|0FToB*-%Q%T2>x^{WYQ<XY+@f=4td{U~eBe#;}R*7mOOHZXm`0l`tFA7#ylPYkSk
z{qg7U10e5dl7E^MZW3e}yDPPJ<;ANX%Yqlk*=0q!47%PRZdwHj0}Z?Jr0rV|#8#i)
z3zimibcU$U5N#&G=lt}%+^I2xQoG1PC3wy{MOmj*Gl`pKQ!S;>nOuJTclCn;j~a&J
zD9L9O7>WkL=ltxJ^rBg?f3myq9Sqw&#d=S%n*>>Awi_OMQ+?VHBo2DBPYB&7B$xzQ
zX4M`R>elDMHJi{Io>D-43TV7ZkY&LyEq8ohwVJ0*_!L-vr9r>ax+X!Eg`u@C(K-XR
z!)^+rJCF|aF8wHLe!A@4Q*ULzb_t~JTp=ne#1NAp%i^cLlbXyM`IfOgrC3M`Y)=M3
zmbB>)>Z~t+qNOpz!-9u3ak@URdFi+B{rKu<dk1$R8ier)(VvitCMjddOPkgF`b~Gx
z(l8!bLP(YX=de<OEXzD=RcJCdFA0`8uuK+wCxdZkkTRC#nOW}52Y`-_g7LVa*l)lJ
zG)NiCk%4U{Z;ZbKdmt?JBC$e|7+?})S*~v?d3$YTc;SKYlvy-r7L7AW8B6!?n@tX@
z^09G*y(?Mo!V%UW$dZ2Q*Ktekj2~?i`omKeYe9>(+9pAkjEo~5AK&Wy3Pc+g#v@DU
zBTHSAl(9U$P%`Y(+)QJ$-6)ccu-O`<jOEV?fnBe6z5!2x(Y;0dZ^4#r5M;@$KIPBU
zcRc?zbV|BF(uJB%gOsss8hif6^+Qfr9`KaMM1M?dCP9|9k$>Dz8x#Kuc0gd+NPIVf
zPBBOs%b=x!+Ya1*4WkQZj$?}7F)$Jhf-IZ9ukq>ZO&_j>?!rDgMfRT}hnu8~<z7_8
znmq~Tac#Hgvm3SqgCNUh?@r&PW|SZADI~z?-c)McR2rECS+<=B_1^Zh_ERu5VRZkp
z=zqZx$sowGebwq(%j;IV=OqNe;_wajO)UMu<{eL+hkU{oHgOk*!H?WktasrcY!GDG
zx%#`P8y~fbbr-$`mivnJz7k>*WZ7l&8gQWg=O4IsLEZVa@^wh*M_KccJHMojY4M@E
z&;_3GNYNfC-X=ko+!Ldhj<qzH1N$Eg*-gp!CT#fzL6*HApX(Xc(b~i&d<0KfqxrAV
zYM2CB_VrHuE#&$~FFo8Uf_6KjkTVK3iR;WV^xd~WOXeMGu1`q+=v|mVnB!vNS4_f8
zg3rlsG45w;3GoKi1-%(3hmMmIOj5=&<M@NB9V-#TK3^pVt&(G$1}S4%F=zSE`|)?&
zh4-Pu#fr692`~w=9Qo+phI;qsO}7bsfn}#gc53BKf-D6WCgk6$yJVvw*fRw(69n5J
zWh~E5(+MqnG7a7IL+}qxoURXSKAGyC*`r<Y5u4Blek7avWW!=N2(lDf2A`Pi@kbqa
zVc3D@Nul$k1e27pJjxp3JbSvAVcN|V$Xqb(41z30kx%sP!+!+-P6Q~PbkR3m^fO5r
z%cW<1D#V7?m4!f<p^8eypwbVQ4VqZG_nd2)aQR6s=Of*%1~3%Km5by^lelT3;MTka
zOyT*&_D6bsIc|bY_#ECSl?J8KCMLljy69e0YCpJQMR%bYm_%ucJxz%=Ng2!0;G~s}
z;<|z@3om_Ms&HNkG6}L=a!*$hE6l9#Iw!;VR#fJSwM^osS!Kh#ySn}Q{`itYP+%WJ
zZ&qkQE3^oc;B&6k_iP+aPE3UHh2D%)LdPi}P^F}dWy*(-XP&Bk0$5<7MutWvlzx;o
zuL)O|wX1&tq!vVSuIx8g4l_v^%l1E?*IzpU7Dy-v_F)owm?W44S*}li=krHDz9^Oj
zKX{+DL|IEBOakVa0@rRZx;Ik4e!EGVGf}V>UADF8yAr0+Ao!d+C2f;4lcOeixYc*f
zl0u3Ut4Ul7ZAj@tyVqsthSQ(yIqxof3FGom(H<(5O@hxUev!HJ>kX4UVcUeaIwAX<
z0GFRZkmde{rZcvGNzcPufxg|X*mo;2CP9{mvAfcyPwe`cv9oR#LN>$BY7k_3Jh5*{
z+LAjmxL@F%GKii5y5AsWEYFpy8!oO~1S1T}bfv{_r6tTH$nwNzeCK;3J{<)nCv^9B
z$@h2AQU*bmr%ID&-^_pKK5Q~DjRW)nw$cx5es=oZA9lU-(-;_AX!i8=^eg?q=I1q*
z*SYaxaa&)ZavRflUMQ*y#Yki*Id(AO!_Rw;gYSdPz9PP_V1qIUvivpva5eQG=R7aL
z8-~3^Atj)841z2#!rpIUc`@RF;aJWgJ~`l6HVCr3^cbBQ{OxjJaapuuE&gNS@M;ib
z`DgvY)O(-x*k=<4!w*fRK~rgMla#R(J!+TnQRYQs_wft%OLV$Eu=#c0<uP^c*?)i`
zhtX*0*D$Q~1Dn6>_t6mF?l2CfXJ68t<0U#?az{_5?wl&cO_dTv(`m0eXGqm&NZ|~x
zm27LJO6U=z=6wL3hVecC434J~eJWK(zrF69BzsJf?YI$8^OgWzboc`nYTgat6A9V#
zQ*mEa2K(`mtUi+6QJ>+Z0er)_F+fWkf6^kD7X<hlctKFwIp3#JIC4_+ngAJfaBf#U
zDc3>2w_;u!z!Qx-16=iZSwN_HMSu;ua}Q~|hqT92m+m}DsvaedYo<ogoySSL<D>&(
zy87YK#Cfl74btJL<at%`A8wwabnyFlUQ_(hC8WdQ(DS<D|CM=e)WKBsyrK9nGp$%1
zT=$;06o2&U>tK73yNZrxi4JE9a!=9mxS_-TNs1Lc-`rYtFht0GMMoP&2mcp&py<s^
z*O(51A6TMP$2Fll|55_~;_=@~;NLucp#;9*@k=G}C6E760{`LhYejvn*x2(YsOkij
z!xL3?B9AAj>Lecjq^dvhc(STa=JC&}`ZJGLtLkbVuTj-CJYK8nYt>4uKkHO|9gnkB
zJ)6gyRedv$x2m?Sst^1A4%N1U$GcS9E*|Hqwp<?XQ*HZrd_=V!;qepI_JqgDnx3q2
z{Y}yI6do_ud=_hgc+$|FOEh(fW@C7T=CeWz#4T8NrfX`tW@9)<^T^S>(Fv`CaaCcb
z=8JZh3y18~YVkO6r`DX|qndtHtHkgzO+UutLd~O4^A0rcO?0>{i7M3M89uE=o#ycw
zE$R%9&uZ#f&Bpl8Y3ey1U)DS>Yu@OZ(ZL_%xU5Amd`<JeraACB1=gPyeoc$zapP-R
zONMW2)oyFy=orz#8xnV0YsBMbx3xA5-_g`NnvLPRntGSV4>a`wj~{C4LmvOBsekhL
zZ%zH1$1gSYC67l_pV2hXVcy#4a8;8ynl@*6EUi421~EL2s^h4Q;Z*9AN(0eA)SZ)Q
z=wuqn@KowEl?LM01se#Bok|lKo<)6T(Lg*O>u@NJpG6xpJeR6-sg2=zRGr7;`Ba_H
z<234%Mgx19PFx*4xAoI#Q-+t&`b&8HEA{!62BOyk<e94J)W+~Cs;=U3CRH<eyq2nK
zd7Mo>vZ*&70d%nY<FjdFhPP043$-!4gX%kIC5CgUp3CF?RNv3z15`i2<2<V8@%Rwc
z5AnEwdK6G^bo}Vx?WtEln=pKmswb(9;ZsyS#p5$nKSL|A_Onz!%j1hwzsTcDRKLXI
zD^$P2ur+G@ya6+#&)apUXMm>ztux)}72xGSr>pMt4)Av14V3Qm3Gi{?RkrT@HsD(a
zdT({-(14+iL8i-ChkLnzp9BY9*?}+$0h0v>dU<p(ssmPv4z#ynvc-UO(SgS(xN;x?
z>xl!mZ5?j30yYo_UUq?(gam9P4s=@R;Qb8PL>%ZC)xj_fctsp&GwN{j6Y!ci@KQ*J
zJK2C_$${6k;PR3JQX~hu-*q^E2i%k#cs7U4S_-%&Iq=c~l!p{>TXNvFfexl{!0(a+
ztz_MKK@PnjN1zF$J1@$i7kPYDj=U<@VeQxC$ZI^lD@Wet@jW^69*>LV$YLH(R2&nP
znt0f<Zd$pky7c*uPP)^V`09o|?dGOC{fHljf=MukB1t5N>X14dYDgM#r~~Q1p<!eg
zhkhVGaA*V>!J%XeNw#o!q=k&+@fZsk!{ezIGL^@3Eo3f_ms-eD9<Q{Jl|0V0kW3yw
zvXDnSo+6Mb0@t5u0-46+83LKX<5>ck#p5{wnZx7x0-4X_l>%AG<1B$>@p!X9HuHF=
zKz8zYpFsBUc#=paiClkIiDVUz*N9{dkF!LQ#p87%S;ylIBH6&>Y>{O1c(+J)^Eh85
z`8+-$k`p{G5=jw{FN)+MkCza#gmC@OAtZ;#TL{_0<6J^=d3>0V!#uu4$Soch6H?6M
zQ4$#?albc4B4c=*Dv?wk&y&bJ9;ZnpjmPN{N$2r4iEQKXU5VV~aj`^-d0Zlq5+09}
z$vBzo-xQfl;qhFV%;oV~nXKh;k?d9^n*m;~)%{u~ule?I3K^$x&!4Q2$vmE^keNK5
zuaNmX-l&j`JkC}~HjncZlE>p?3OUB(#Tr?xalgMpBP)2kMk8x@Jercxl;cmOB$dZg
zDVfUSS(MD;aT+CQJpPrEUm3REZ`&)oD(wXa)QwujjhZ*QRWVEqPK`#4*h2H(#KWhs
zY1FAp10~`-F2o%d;_+gKI4_YJmq<-K6%yxF5_grvGrUfVS*O*;Um(uuw83=R5REMl
zGwQ7LSnoMrs?3h{)1<g*ZyxLG%#iBgsatohm4eqwA^fr46|07aaEw)h2S1YNBgq|4
z&afj&aj&F=W@c~&?09nZ6giyXG})FWSGs08+o&1gh9_-|aI;v<$TqzGMRNd7_pY$E
zBGb$1X-2dO)Qn?WyQt=_ep{zzyQI_=Lxc=?{*|d2%ZB%Ty7QqH`%r7a4Dcsf=o76T
z!_T$Q=R7XaLQ8m@Oq(au_mCfb1kfWRFb-c$2M-51C6`B+h?f~5Cs8vlj`e7irjOG6
z*%5IaZMcqhU?%Ee+W0Vio0*VTX!|R)Cz@;+R|js5MFL$U=y)5fJM)E@d?6Oai<;4P
zF=niiw;62*V{F9O;oxD69Z0SbV29uEDO#t#{52=05^;uwgoTbaZ(#vNgw<l`xKQJ`
zP!oTaIQJ9BeiDMd5ErhopVVadB&mLqgfS^UNg_^?S_~JGI)$Vz!>37|(>%UJVlR<8
ztbK+QGD8YQ;U?h2ikKnQ;&JU6QY^zMa!86CiZV+;h9gqsT0E|uBFFN$eu`Ws^Taux
z29Kvz@yO)DiQ{Plh9^_|WEy}+CJav$sTq{W1mT6MzEC}wAl%c`dzuG|8U`(bs2QP(
zs8B%EM2}NqAcn2j&Co@dKO<lf8BrfVgAuKW3mlKjAPPo|BK9fJ3{HdwBQg=aY-r;`
zgAtF2j|y(Iam(ZQ0{TB0`8Q+)-3u_fMkpuF2gY#1VjpLr<19M5mEmk-iCJ%n#n?+~
zMt0&&B1U$?)Ei-)IFTr(AXP>%C;pTUyNkzk8iu<IJ{kICgmmJgF_se^{iYJG@%8PJ
zZ9$8{cyVTEff-s==0&=!y>nT6pLt!D(ZFT2Dm%rWq4mztL{xFQ^Ac@xi8e=!p`kDe
zc85LgAJZvLO9Mo4uh5+Xs}8J#u?!J?Unckq!}o~9IYD+zkVDZEO`MnI;LCC~JTwvK
z?{etxau{NaGliFUEYZjO-uGIHd;R9#S0_cThk62B6LN5x9K!E?uIN<UAdrF{FhWz=
zUS>oEAOrYlOXJ-U<M9`~&PKQ@s%PD~N~pd{sDW4ax^t_LuvKV)*EzcLfKc^-;6TkB
zeQQx>m0tTlQ;2i3#j)8Ef>Z1=Ox`*!v_MAUJSo&XDTFh8TXftOLof!4I15Q+A&Fut
z_$+apB_Rx71FMFFBSt#|M9GMzMdcCG^2S8Da@V@<#JR~5yvb6P^(e>U$l-f<SO`8W
zRAoInCPW+)YB795sCk0#@eMKfhFF#L_?GCn#lOKx5`2<W#U2vpX%cywf1@JeDB|Df
z9Em<hVi^B9^42-hg5j&A=2iB_#F;9GrOGu>>Tt@e)Ql2llW&BIB7^=fTQICj!F{?=
z2lr^*yt-!mGvTd;&UniW4mZJPmeA;9Gmy{~Jy>bBkihU9!Do)psL;HPc3nZc&lT!0
zY{ZeV=NmB~3}*;5Lx@J1D~oc5ouo9xnZ4LWCH|t)k*SxPiswxw0;4j(lvHcHQk$~f
zFH5VDr8Pxk7bBxVPZ?p+sD5<ka*M}uOCZL(qH7w2%LtCfU#x3JMHAh*orG*Bb@5)u
zbsH0NoHWJB11B1ZDPWkIvD0|F6YuEb*4v}!*I>Y(BjM*r1b&e?Cn_-$mD=bqA<hG|
z@d4TdO*i5^LTetO;S67;HLuce)NAogXE)vOcM#kg|Cbje`r=^BT%q8oLhF^^w4rnK
z+UIb#*`n6iqDJ84)T=$V)1|$uo<YPyH}7sSsE>8$4!Oz>xh~$cp<@kHoDn;YBd9y~
zlD2#K`>_Jju7KT#VQe{UDMmy&S}=)bL^*WS2q#CI730WDt>E&Dgo?}xJ}AT-<gH*M
zx*P?#p&4BcU(6B9=ZF>2dIHzK7;{aGMU?@%sx<T*yMi(2Adp7DIVv$7VjjqOsr*GV
z90E54=(7=aj*1*F)hbal${ZzCcV1C#S5yZMx9)tV+Me+^S*w+-HDm25TCEfwFVX5Q
z(b}+Eouyi<rCJY$_h^niT77m4vsVk*t0gd;r`5{SnlYTOCFW~yGkj2U9OS>BLE|%M
zJ9cDQO<S&}A27U?#&4zVQ10QDh}PXn+c3O~+ICR~!~1FUe%gZJlQiNaZHnsz3|<;>
zieYrX!j5Rf;<KsvS!(;U)E;ep-I*nI$dca14M=zHlG^T)+GFr}^G;Qk|L~{U9OODx
z)>CCa)Ym$!Y<;5Qhbz0e8RQQF`li0d(0-U~S6n}?q@+^|66SY%Ye<mp?4oz^Vhg5g
zt*-G5b*tU20YlyO?i{BRoan5rr{0rC;XUiJwqANKuC2G;n?uDGk77$@+*WnxSkYsw
zSQ)oX-MN5xEFhIp=0OmN#}!f;6&zR)lI?)xi{qs`^Q4$ODUsoPDJGxC2PNA<p8t?!
zJH+E-lI_@=IO>>GpFQumWIN8c7f7}O9`BHCJ7ix}I3QSZ%uYFx;azgfE*|H~wp^Zn
zw`|+Z<NdO2|C=~!zg(X^?|^JOz_;hgwmcs1RXp}8l`+N!ZUm@DGOdh9xTF^&I?hbq
zyBfBjy^8N%C73bpRYLYEwRjw}SE<kNR;|`ntuAhEtt;DljJsbk1~l_M&2~@o!e78Z
z0ayw~;t10z$!b`#8pCwP11<i6)&#!@p&(QrPd)KGfEgwr;YNB1?zgzlg5EQ-NwDEx
zH)6~#0nPeQq7Nls<~3R)2d$B#^UaVyYUZ4vUhMK>PV}v9qhkN#4l+z;fyHTLvfw6w
zDJ;-n<gcKT?`|fpz$h45E7;&noL_ON^!wod`c{og14ebc6g6IIiidh?X3H?$cTTE$
zPO5|Z)Ccp5?<d)wEsfQk)$G-x(OsiE9d-wYLIXl$SX;B8X0aGlq&qv=J4G`!)7jpc
zL+=K>%h>_*E%y1AXtdcttyt^}c)Z?XUvG(KYG{MSzJbT<1p7K6dZ`)V4ToF7zMjYV
zf<0e|W@_u8U_Z#?qeAFWp+0LrB-#&&(da?ZorgvHVIE%<?N`NUruMFh_G>)eL+pD<
zG`g%@=8JtVkB<}kaT3kcWC5`k@OZLhpDabQZRlso{xgrKN&!=)7}h>b3Yf;@OerW+
zie=kYmK2o5<DHUyrxcB{;o8>hyLh}?3fL{hu=YJtz#bmIknAs{XtY#x=S#`{lE<&4
zfLBrsYkw^Tyyo#DIcSj_i$`7ExmXTb%;Qb6eUlu`)M&PB&*t%FIbgFK!`ipV0b6*y
zRk3eXqA})6cWzVc+jyL;R!DxsK2j5t)ecP8rKlk(YCVQus8KIe&VIV9h2GWbGhO&V
z3wof%BF4ls&;mx*8LGS<>jyq;fB3*vuosfG8p&FjfeemttwsvNc*P59*~msiUFRC=
zwlAgjY-ktD9Tv-Pv!TqE+h)t{acIy*0uM4$*O-3Kpn3+a#%!+DRA0^GtyJGitD#NR
zdtAYa!3htCL)SN`5gXJfHaO?Bh;v#L`UKId2>RYgfkPGeuWm<FJYek``Ecy$XXLsd
zgZ?kxN4)O_?Q0~*;ox9$9Oyc`r@b2bd30yAN3{K#8Q)7w6Xwv+!gDQ5m60unDjw71
zRF~(>XPm{4dCs<4YHzi~F=e#XlCag%h{s7=Eo~UyZVBIRiDGz%C42{ui!IHHElJo;
z&15{38H@&iw>47ma99SKDR`wP5t<|F>?CqkjJe96M2y5doQQ#DVjg^Pl%kGODxpWt
zWwpjmR2rkz3cFir)*c?3FkKJ2V&v+THJA|>FqN(hJ?uI-8!SmPEv;}pxOV(*%Pmec
z2?p)?HEjOEXWjqf2V=Yzk~HU+%9kNbuw1DT7>5AHe`&fPYkh4Adu?fnUxqa)&=esI
zF=i2hs2I6}IG%(5I&e6)SwW0dcdk}zuU6}_rID@1XR8SeCu{E|Yh4*m(YmE*A22*p
zYcW!5#qcPt)hMkk!^gCQV_HLYu9`<9=g}Az4YiCqm(iZM=w0WgT03bpu3@<0p^5pl
z5yM5aO%ZL+&PkVPm&>#}VhpPR<>AVR<fpazOiK-ZS`(IA8sN0T^`NEdT8jfE8sm6C
zI$W8Pxb(o1XzOUtY=_QmJGVz`0TU`g(u_1plr-FOp~1+Z#7RKkFf<sMlc?x0^c61B
zj6_Ln03ED$PF9_4bYG}RFVv0<-_;u4)sj%FyDVOx@htWY6D<YJX(U)OK{AruP|(n|
z0D|+Ta}4bR=(CYwiT)0ZAc6)Xp%NRgFVJ9QQ{u~D05E7>BZm?jT<W^@2DJ~q6(&={
z6O2qsRCt&|3B7dXPvT_b`~bfzZxSbOw3#vq|7_$-qE^D+g^n2Mk!)0q7<gou)LI;U
z|L527eRZdQ1OMhsSqFOu+woU*XPA4KCqvELn|U((vSWje%~4`>=KwOG97g}cLDkY=
ztfe_q+G8y($67k@xbs*`H-@KM8ceq|XCgh<VxMbqFpG4qC4R1@5s#bAwX|k<sl~q3
z;$TZ4$D-y~Ja7r<;Gk1C^LV>O-Ol437Ig=YcUsh)JT9@25=%L}&w)5<3;D}Z4inIH
z*fhxBmU6gj=^$y1=L<<FdSFotjTi8Eq0o3CkJE)t=|Wd@%Yx%WuxAMlW>IGe@mWG6
z9yiMpS~I*^ux}O|?CRizpq>ysm_>b3P*3vsjG&(3agm@F@%XHup5^g0kxUcI;pH{B
zwM8;rEQfA9NIVe946z($;6j|G7@sXRVq0Ri*eqLY&Exjj;=2s*5##rWjW(HSfuJcx
zd%oykCxCo0J|CWE7J_lLH_M0TmEjYj{e<X1;|D_ZMD>>Ff$9oE^+feHkMD`<JsuZ}
zYB7)Ri|TzIuOwt8DTjMLgenq}PRgM-5^l^0$spx0%?7S=NTaQ!1#?+%BaOE4IG1?m
z5<6-j9WL43_Y+Ts3yFIn@nk2IBBB-%4~EYY^(>Ds5cL9&FB0`4k1rAR5|1Af@|cv%
zG*d!#$kioJNI6X7h4m}NkChtXVHJ!pDSjM}Qzh?I$<9tdGbHyJk|)ENl6$7)$)AAS
zvv|B+YOq~u&Q3VnrPkY}w;BFTBEL!HF!KfS9HiL$QUY^HJdk1^@c0iY_75JHNWmpi
z7&FsTWcL)=li@|O`y$zsneB^Z_r*M3C0AZ02cpYEcdnD`u9F*}Sq~>%x$b%%Z;<P5
z;PDBWoRG_5B#tW-lwOd_<1M8t6qH`%@m1OVs_e<yugUJ$czjoOzsuu$vim(A7t8L&
zJWf`kla+X8S<Y9y=PP!MA%Q#sg)C9Zp(zT^bj2%2@n<Kt93?16sm|l@9Hlm5T<2h&
zFA{tf34Tn`6$-Trg*c{o3WbD1p%IUh3WYWdpAo{(2vH0d3E@RN9xFB*D<)w(HS_s#
z*PL{>`I-J7efjeL{0$`D|HHm+R4rip`kMg)SiAu2h&PXMWjRFDEL4D&03=nYw##Y|
znkbMArrMtJc!?IgM61oNBbI8lmuk%!&egngwd%|)&eMGJw3-a(Ya#hs9K#vZkwNS7
zN8dWDX$yw8QpZ+Wmzl|h)TfY!p;<ij<nIqtwtb)ZA3wIBkbzNUlxsjK!(WC5qeue|
z-RuvS%qeVbod8F<zIxvvW(=K@^iz;cYZl4Torh%okQ{{SV}7I0a@DJ2x4~h)i+7iB
z_O%02l><<3&-~i_0!zWv;E|5f`#jGrYE_(}1q*9l`(e-5asP3TD^3A8yvNgO<7qu~
zFH^JT1irWqSECh=(CTcCuF{HEX>}YajN<@tbrqyQvsHHvB7@4Kp8(GeaI0=u>`bQL
zyhhaj<;3U04cCYt;E@o_rP8u~_&^tq&l2Cqm`+gNqR%=p4&B3`zQs!G#o7#S5Pdd?
zaT#VQ23P$B_l;r%!*2#=>1L?{_IrCpx>pQC^9Ac^z#25FXrMKZRWzW%RXqd$8Mj>+
z_U*)TJMm+7!+sLJpVVe{!ATN%k_FRIv&IIVPnM3D^z#w-xVQe}O9Kr}m`|f*2QCd)
z(2B=aArLK3Ozwcgrct<qF|HDlRtc?`iMv(kwpDPJebG#bgkLbqc`#8+Q5&SFEl~KE
zndLoTF&ec!a2D_w4mA&qiXQ9<Sya!W!6@Az4pd*q<6K%jmo{X=e3;q}(-4Nw(&}ev
zLlkbfW~0@wFudF>6#}#RW++hl`zIUHoPiM#jMUY;YXn>8c~X@;-rt)qRmtb^PPxiX
z-ru@QuCj|^RJ<_8M$r-WqUl<2x)#d<w~lHxj%tahd9fM^44qMm1O*TU9p1pGJc47I
zVOAc2Zy5DPP|{Wp-6nPXbZHA8-Pzr{djNa2?vC!<i*}Fgp2*sq-cF9w8RD$PwZ%Id
zGt|qw7suJ#yElgpN>vU@RZ&Um&Vy3eL4NQKO7RD!#;pC2ROJxQe@Lovh<BzRk`fQ`
zrs#30%5k2*K&n!}<6OB)u3Qy$x$exBYvl6I?_9ZFuH1yR@0P3V=J|KaHFop-yXC~)
za&yLiK(2Ct=g*U?<S~rv2d3)H&>pnq!6w^I=yu}Hl*l<!`y7d5hlg_{;T&ni<D_$>
z4Z{~m_yrbB1;#1~zsTdOq}f%Hg!~X8U0O>9MI-aVg1u?6xdUb0`F^AKd$LLG(X>Y&
zhW1PPekq(y(@{x3%H!>_zFiJy!?I7-_c5Hc;miDz1tAMyZ``3&+M(FdG{tN%kXWN|
z4R&j-S-1uojB+(-Qen9oXfSHkpg$YcISjQ?ng&<JrhCISc~5+t@xP$ie<NrXBgsIE
z8O42=7JH?7zEVR_;lpmOd1gVxi&+3?^G8>XG^dBIwt&1gtFNo_ap@n(YqR=sDAF3q
zp*q$&9O_{0z@cH*VI2Cw`U8hXSVwRu*<ww$aCoG}I+Dj@E$UcH1=gc!7IhkrGcDFk
zp8t`>`iRF<1nU%md;T=RI*rFO1nUeQ&l0S&csxh2&f)QV!8)JED+TLH9`6&Z`*=J_
zv`!Ma-(MwKSMhj_XkEkOEYX_9<8`8S9gjDN)(t$)7OmMl-Yr^p^Z24@y~yJw#JYrV
z{mCKL93F2W)-60fOst1_e2Z9b@pzPE9VKzk8zWiA@OYkNoyX%e$(qLFZIX2xk4q$L
z36ICg)^RfT{3)_^3XfOGWR>jB<fBN|isXt6zm~19dH!*Vb)3RIZ?a;Y%;SxUbt8|n
z6>Bz+cPnJK!u?*JV$I{*7i-qV8u$Danso(_*J#!?JkHTbj>bK2hemerxKJa7Jie@v
z%RIiSk*hqut&!V29!;&IDc7G=YE9+wRBD~d<5|=?i^plyn#SW_sr6SLXH$|*x&9YW
zQot~JbKx{M&QgAyr6O}rZ?ME{;G^}8B4cPiVUaP|C|niB@U*i%|8&<i2~TeRZ=3ix
zwh1hS;!e%FY;2o2p;kJ<yD5rPTan5oW}R2{^J-<bjht7Tp68<sE~*|ERX^7LL~Zg!
zZO`yiwc}HjyB#c1LrT=ztbM6gb*UDG);RcZHNX8@7#iDf8%Z6jX$(7cY@i++s2{@@
zY1Bp908bTI!v${pj3RUFMPI7wOa4XIX+G<;XxtB7nT$Rg8AhpuW3}tPl^v@2Bd1<k
zA`Mw}=WzGo5d~%)HT{hnDBby^`eP_wPr<F7B{<pA1dq6IOKYi=VyVsWNK5cYOB0Mo
zbCo%EA7zPP*eIgJe)-K14Y*0L&>5C6yb0Q6*3N_LL!)pW8<mloK2q~#(~(6JvS=$d
zyt%Y>F73h};@Y2uqoG-I6SMl^!A6}uwEA@C3#si3KExwccBHZl6mWUU9TxE+ADiU1
zn^?$4?(2keudjuD_5bM;+7mh=I&Nd@U;g7q_r35xz<FLL^P&QAeF}M~s92W_vO^W}
zNKtW!A!9}%j};YDA7N_MDk*9J8{v^!wUHWEOlXu=Z4{4Z(}uHYQ#^;Ge1ctVR4c@(
z!!m_%)HiArVip3NG?Y#kl`hDLr3&E*Mx8>m<?~LSIDGi-)7<~K0}KO*;TMgvhbR?z
z695fHmBX?IvlJ7gt{No{vGY*-Nwr;9gW2$vsJ0Rwzf`Ne<f9c1YyO9|NTxrIYjMZ5
zR)~>84~#;I_(Mm|t?u5cSH?K-CjOz2KNJm97D1BLYA@6<?7{KGkRLjA@SF9&{pjD=
zk1$gQ_IjfVB|5lJdqRV&`Xnki)ITt7M$Ji_Hmo@b4MxF9yu!y^9(a&ZUlM--3!}h^
z$tW&~=TY6cPfXe;w!#CeE7YsoDbb17p=gl67hNSM*^a<lD{y{-(samr_UVF7{pU;^
z_a9$jc*O;;ZB)0!-eAHlXa%F9B@UD;v$4^1+7mMyv9u-l7e*@TNTmw9<(i;)Oi+R`
zehR7^DIQamV2oygWGcmDwi1k4DY|o^;;~Q(##=Fnr6{eFiNP+I5dxZgyc9TIie%TJ
z3nY&Pk{zdznq@O_2C-}=#6OJ}l<`6>d`GCOSNi{&K#L!eqe^RPvi5nRJWq_k2SD6_
zNEe9Fh_A*UZ}xn(kL7<W_}^H;=p6+^%P3aL1Z|xbyH0D28<(qkmd9jT1tk;GDQNTo
z+5nd{${C2VQOgts2V)(f!Bxl<*A(71!#^7ZOmPoEPZ~5BWlM3NLkbJmC|1e@#VA*a
z43P0v`oBj)e8cCL@F}-EYJupk(R|iue(asrX%*M8vKRka1_O&R!QdDrPk9|veV<r^
z*D(pF#0IEi;1*G=dRcT}$i>a)FV}yV(dz@~!EajS-?RV>P=y+ZTICyB047a?jMGXB
zs4togxA(Vyee!8Kl=*=Yk{*v_8)^(Jx(b6~R9t0d!z;D<E43Z&b1sKb^DM0$zB3v;
zp42SA$|gKTRa5wKe5=*q)qK)Lwi=qP#<2${Yi}iM?=qaCwMyXw3P);nMrv=d_EB2=
zD6J*K8Jb6i7Jw3{JCA8Kj`4|EziH}knhzVQ%UZL`T4xNda}Aw(L-S!nw~Tr$qX8`7
za2aj0jCN;uC-vV6xmaceWLE%U&3u}`@EKbD3{-eCZDQS7MC%vP)(l^wQI}|AhA-2k
z%k+K3SiTnK_=!ZHNab04Mv9!6A~(Z$3s)f#S8ZDqc+8Ulfj27JqAwP2gJ2Ud3fSV_
z_pkR%7-<9E?i|r4NAzQ%M0dp6cf>f9L*02tOt>R9;&IX)u?@rb#PEA!6vM@0crlL`
zl4c7@61G#bY%Z>g#|yUXid-91?SCry-%!bq&Dz5t4@Ti(w4R>W?i7!j&>{B!e0&EM
zje`*|3M%7L!H`P0L$p)7JB!l{jSEd+k(puD!@?PAZfjl%<1}e$0cPBTu*4tecUSdg
z{AWlqYFmv925#1&+1vO4NF-|8sVp@3&*29~c)i%|@qd0s{v6wCbG1Y3pZ(WU#;^!B
zR9P~LU!(Q+{6*Gk@4KJ=1nRb@ch8#aD(#q5<(L%7uG03)Rrbq~cy)&L4`IrUO4(?p
zf_)>koy@1(W=b70`E=WzQrn#@-4?eoIGDT{ON{ljq2W!l1Rbc*@TS`bZFvxPqi{BE
z$td6O4gNTSvIzXHGTA7$U-lc+I~a^<%n>O23l5nZ{$<FQ9}em4GP`@ZK8!!z@Gov|
mJ=kAAV1ISh-ZuaA#J{+~sZ9Fa|M+q8ll(!C?%jI!a{GVEwFQX)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_dpp.hpp.5CB9342685AA4AD4.idx b/.cache/clangd/index/blockwise_gemm_dpp.hpp.5CB9342685AA4AD4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c1bc1a5d6e95f18003edb6520dc6ba73b5c77f7
GIT binary patch
literal 9864
zcmb_hdt8)7_un(FitEEgbXnxK0<r=N3p`xhb-Auw1k@KVh?l6WNR7Oqcmpl9%<>hb
z6cUs)r6feK!a&7W!#gNw7-|Vp=4&ESq#~I5n_W20yuW|?^!encX87*+JoC(R=A1KU
z&P<Grj(%Guk&J#VcG=vN#WST6iDVG{f6?N!In(<{B#Q?~B+Is!Pf6aCuO0GW(W-Zq
z-DSy9efAu_J}=;lfY+u>7!ncFHs<kd+lm1;FT4IQHm{_+Im7Rr&6Y(oidI7Az2B=k
z9o`hS)t614b3oDWcE=~~p9cLmC#avWGIPbJ2D{??2ASKOLh*onXHnwUL$;Qk>3dFn
zc})D}oY#L$o9<Z<5&ccvq#JWWBeMDxT)Y`R?n3!y{m7bygUd1`S0@QJyFY$A{>q%o
z4NE^7kQ^)X*=ad>YW<P6SMFT86!I42b%=B0<{d~HuW0zd`da9}A6ed3tsqaPHrjXV
zub+7G$JZq#2WFWMel7DIyNc`EKeTV`R9&kdJ9*f+;=CGRutSl?Y5v|17yS3ySLZs?
zuIVP++hP+`x_6#$;;RXF_oSX!J~pI##!G|q?>x@fxas$g2U-66XltYW=ik2^lJ@LQ
zystuBTUWcp=icBCo{TCiJz9I~VZz#*^7{9F8`QS))UFLjhDKew?{)vQ`_~V5r{72m
zT@#!9sBV(>g~z|AJZh_t^EtKaE%VvIhUOg=JHp?4y7H&2`eVu)4+j5hLuvcjCtGLV
z|8sh~ZE&pP?*DqmWDi_DX~nKK>51I%(%B~}q}%`9d9&qe$^Cb~u)Mjt<jF(p%?sXH
z=WRFYeuZ~#LgLh{-@h$gHmGdRZz*d!&*iR?pV^sr_Ufhs?J=sXD*ab?&lRcH_ItJ8
z=zgyK`mAW#@bmaK$45I)jz4qo$~oPTE-%^CfLB`w2Ob%>aivXvuh89^c@xiC-+kqF
z(uQlE=X2(kT-mxI`Q)-CA1%$le0e~Ao6UW*N6N0!^&{sz`ziXRluyo{8&UlBzSm!S
zux{U`m-=@${Nh|4^KN9D!)TWmk5!uwikb1lvSmY+%dehN#Csq8{1=<{yui$hV}2~}
zJdk};r}%7S{kzx395x#;D|TR+FzF{lid@q;RBhEhzF*Ey-@p9%uAj4C3y$v|+i&c_
zuRi=^+S#ss_qAj5+GaQO-_Z4A@~YYTmhFOve$R*A1rdwt{zyw)ooipP(DlV5)7v8a
zHOrU#e!Hybi#Jl1O(g2@c=r(J@4h^HEPi0!PnBy6UazWJCUtZk`~9_*xs4C%ig(|6
zdiRed=|AgY4yP73&3VQBUefn>o5J5(l$ep{5%ocz^M%LT9;#<tS(Kb{dg-H&WPy+L
zE_ZjiCB}W}aPoo6QSg6nLdK%yb6;88S0d@RWcK1i$0x1mM`Rj_sO+x{Fg20^eF#V-
zAc95L)%okszPepb9DF3A)=}$fY9waVP3gZuFC0xA+SlroHL;*{ihar`Q-kyn_wvk%
z7e_pn6JJUP*af(l8l-Ix|FZN}|3@9x#EH@!py_~dy&Z}S3wf03q;nf#O_WsRGDt6j
zy0=3Smn6^Y-%Ks4mlF>v@(k>rfnRTjB9R%l&Wxz(thOfRRAjpS^e9t<bo`5}CyyBM
zeyTOGq;xrG$|140Ly-wX!^ZEPP#R=Stf@!=XbNCTZ-*j@f9BQMwxkZWAp@z%YOq}m
zQN10COif<<`-)v9*W{!h74dWQb2c?dzc@TiKT+yfV?jJAeT7(EAx^y=ip)+|CzMJ0
zd}cvLP?1Jr)ksBpIux0+u}(Wb#o1p@%&3U3qp!26K{~bm^T<XEEnO2gO1nF{yO|oK
zGdky$pBrnIBq!dKK1S@05%=B>MKZ%fgWqT$JX=m|sECW5i-V~_dX?wTB`Y>B**1_2
zr?g6@8ft2gURzhJ-g>jjRZd(et#j15ni{0DQVtEgH0Relax#F@qNC_+YLNcu{zaz@
z`w7qF#GcZgj-D!0gY=f{KMt8mMsAQ3J4y#R28}Q^NN@dW@mp;z4F%R@7^P2u^aObH
zb|{iFBK^!;W%79gNq;JGp2*G<d2fdzpLM<d=Yu{!FP9Sq6>+t5bucwZ7k~Xj#}aiy
ztDLAQeT3K@p&IPzP-Oo!i)q;h2E8dKaw_6%=j>o=kUo$SeBitIZOL-tPU#9_S3yVU
z=}_cg=K>r3TZRN{;z&hqf#w!<ot_Rw4iz4h)@{Cf!Gd^Ek#Mtcxv4?==+g(s4lVil
zvK4Wq^cpf`4N>%VD00lHAnVA<-_{HyQYunSWYts`J>7GW`q5@v2j=dyCWEQS7SL>g
z$-NzlRA;GgE;z1guqH#ONThwFo2hwDKiE6XX8!xNa$-yA;da9vObyaa+KkfQBWL_-
zO{^%r7c_feQg4SMHzov;a_{CqYa&pQ1p9<2Q-k!~LEn4`XYziPlfIPparALEHAr`+
zmW-OE)6&&+rnK5otui%8KRWxXb=jP4tL0=UrM(=zRHg>$?yY++Z`#&kAtx$IKLxv|
zH2Cy%DDrH)qSX3$a<UZ}L`4kl2Cb<<T9TIX`XX;f7K@3@6obnO)c;92lzmX>jUmrN
z_2Ix_!BP9Gq=I-B{LjL8G-^To0Rn%31PqyicpEHlgD0AbAl?Vd`{0RDS9bO0q;IEY
zgn=NA^BNb9%Vh4mKJSgKem{&5L}#V58=9&hx+q=T(Den;Rq5)6jv$D`mBZbtdlym=
zQ}ihjxHC``lQNE>d{E?r<FQ_|5yX7(%m-hF3&Fk+oH5)AVj*Y>A%Njx&=*4lT7@9)
z1N}Z8mw>*6$NNFQpU0)3FXeF==*xIq3HnNiz#Tyl4}ty=j}L?XFprOb{s@n&Kwrh<
zqo6;^<5mc7g=n@Wt&q?PQy6XoWgEEhSZo6=k0aY4j^R!S?}TVvQ$g&6gie^ka2F`M
zz>UXZ7if7L*#&V7?<S7Bi7N&Ns(<2G$m23HzKl#}d?guONn$Z#31Sr)T}5Is1`6VF
zGWa+dio1#ttBw;d#OCFD|Kr|wL6uAp4N`;9lT=6~Mw^$e1`cll=@uRrfV6<edqKLF
z$K@a`=W#Vit9jf2(gq&i0_iOtcYw5mVa$SLGE?@0t|*AHBV$9*aRhOJVZj(?=_Oz&
zficWd4};+_jA52l1EDn#$-K7)Ce%O@!%Yy{1d+^}n_xl{Br)6ux;6+yOR%(UF!+Ag
zluD%HZZd5*nT|^&rAs)ikYTG+f7XAv=_XLOcXxLW#B>1X4Kh;(ho6WU+(4NrbHkqn
zF(@b~6n9%e40Q@s;eI2Clbt7fU{Dss7oA`9zywSXQ(aO$@m+$L10Ffx$Gm?hc<cl}
z=Iup9S41Mh?#{{*tZ&8$;(jt@KXJfMGg_1GIEg@OLMx^bv?rNl(|ra3&tMX!infgh
z%>MlR8?ld6^mh05z!)zTlfsh1apqESX4uSdoH^>9>Zd0Q#j%HdvoijB@0b#*&uIVX
zP_zb@u&*mSwDq#T@d}(>A6akuim#x0BCa`P6#MQvVyGk2Fjw&wH<YioHXBB#pmI>D
zF#+>w(vF#Xujzdvi0g>MI-+6EuO|-c8TPwXu;2dWlk@vZ#h`$oKwMU-SOfYR2u72X
zinXAx<#7|}n;;lfBNdxL-^{SqkZbs=+u8CT9#-eSZemk)@dNs`*FkX|M123KCns`d
zh02c8@2w`b)kKT=i$D`N+w(+=`qIx|KdbQ9^$Yt5qTWfbVTQYo_^cygyqbL0GaPIY
za^j`v;mi9AVjfuLfynH34_NL2k=boA2*p74-;=utVg(2l;KO{sl9*KzH6Jz2s)(9J
z&0|D;jEqOS4jVPB+Bf}L*<UVX<V5>DnGek(scObWjGKVAh&D|9!g%&!W_5)quaH<g
zc0_%h`cdwtuyU%F0y3n4D4E}VLWX?8<09f*M7)rX`EI*De|pp+e;N>lSQ~sTMP9K%
z48;j?92YOdb12C&X(&3LAkMU$ITR;N!>X{rhJ}m;ngv>h777bFE=@?|P`Z%LBTYK@
z+!A34$1N3>a%hFHf<u`?CWls8t{TeTo)1Dk*f29F1fdXYcr)-VhEcp3_*TFuW(GAN
z)PN28mq7hZQv+ItYe7@X<MSY#=jAVeaDm6yLAVY!=nsO}1ezw$GTaQBW**-F;RY|?
z0zwOq?|^UzY;cPa#8wDwWm}FwGm{~0;KcZQAl&2S+d*jO@k0<Ef(?dgW3m$131JL(
z!H_O+Vz?WGZeIQ|2#<N3MU+{@gZXPVQD*Zvm$>E<Z^q{l*E}BYCZl(gaf~k{qYHVw
zkBr>MZ$Fj9tC9rqZtGP=f|%PLCuYZqHM%XXGu<+JY6|FhXdsm6sU#qQB?KV<FX_j^
zfmCKH7@)jyMUq-!23vu1M0bvau)tSCbTtf5ef+WCt;1^r{xV5;Avk~Vr+shF`-ej}
zMj9H(a)?z9aYEJ0ddnR6VXxF2G|c<Or;jm0`|1Pu2pzGGMB!tCxSm9;XBev^bo`A(
z*huV9P}W-`!SS(1LL;KFK!QITEi$GP;?V+AKe2YZ*Yb7|9j_IvS|NZbx)ZEAA>g@S
zZ&8JgXRN5O@nU0Q|Ai3{?ShWBkJ#=bDvY19-U16ghNmoBnW?aXKMSH~lxGZ9Km<`8
zrH;YVs~}EMPw{3EG}$xRi$&0v)h~N9G-t#d4;I7bdd~G?=wr}*40;yP@*yN2!gw5!
z52G3W6m*}09u3-fTvHUnP##AV!e|~x6+#Td6`-$x5cCv5JPm%Q!JkFIE8u?x0$BvB
zBqJ+{j(K1u(N~fX9*0$u2!^Z3$SR^^9$H28RV0MRVO1moF=kG58I5%-rub-=Xg^#=
zEMU>YfU$bTgkr;Dqfkf?(+%m|VPHQP_Vb5<BVahf9|me6w3a^%)WU>X{xHxCq0RhZ
zpcy7K^M`>h&~>p$x*@L5J8heuX)S4vFL!sq9G@DH+|z+GAi3Is8IYIU%K<ZCLG+RP
zIAG;O5Pjvo4(Kz2=qLAcz=$M>qFi*qsvSKS$b%d(ffPioT<d^`CqdN7bq=UKnk`#I
z%bB{CTP&9|Gt98a;7~DfDkg5|E%+`P3ytMLrvFYz>Vz5Sl7iSkbPXf~lRiPbOmvqS
zMm5lE+gL_q!)R4nA2#PKu*m`?24?&+GnuKb$R0`xND5`%H#1;nDD$@+#Ayd{$514Q
zl_ai`Oy+Ggu8K@X8>Bu!;!l%A7Q-)+_=_YFzXN@PZXP`aOq>J?P~+_>TOuLQ;3~aB
zEHV0S-=Y2~^V{^1f4K`_eW$O?RN!Rec}IJXVW#32?q}f4AzU=z5((l}@W0CZw<zgS
zsO+0mZ!?<IDHeFM^-fo$bEpOsHQ<dOD$rwxqL#-^plAYb78II6(aht!V09N%ELJ}P
zt4E+h!$3v%6-0V8dY^3)g1C*iZX<qdr3%ULLLwrT^;U729_opqo+O~+_bofzH(~Cl
z(GG$br-)NB+m2Vnb7-Mrp^|-hnj(!uOB72uv{bQ_L+3$p9+XTg=RtKIGz?z=#RZ<f
z0A3d$fblm#aRZc0Yd1i312jAiya6E$w}7IBmu~@83oqXS`W6_?<nMvv9w?dC?}5iX
z@ME|g6zx3U4ytz0FuofU-Jrx(62xxs=mtNAAA{mC&p!s$W6&`EG|`<V5zKAQ5ZxIb
zUnIJVBmx&p5HAtkC5Exi=_oUm;@KK|5BF9vTlWd~F>qlq+}FUu;v~;WUMvPAsuQ{W
z_$BpA+zvc@#B6Q{e$(?!ZU<fwup)?=M`l1KhjKxe%kR#+K(~wEeNTbkDSr2C1ph|n
zeS(-pG+D$K?b(Rwe*ruW$|72Zvx#Rm@#3*2n}0r=h}rz}*(4~Ne?FIpxjdgsf^z?t
zAC*gj+4y-xlgE#rN5nk-`8*Po$3LG(M&+^R_y4%PdS}3~RGKu72^uqn`QQBD`7bb)
zE)HJ&9F_j|#5vx4%^UQ<te5H?G0()<Ohd$bM0k%65u1r}^Zy<phVSM>M9(!1GoMq$
zaEc_LRhGA|C|7*Y@x|Y$3ZaA0u-<dO!^VCJ0Z(BPUO`m6exo*0cCLkP`UZ1@g4I$Z
zrI8j4B}fxEG+jEKLn+b}4y^|B)xeFj1<bebxB$!xc)S<P_wu+L%*%OP4d&H6ZUFNJ
z9^V4<TRiRn^9~-bA<{ME`TTILs9ubBTTCm4NW%oYb2%LBJm}{7ha2f-$p&J!fmkzJ
z+d_mb#D?J=#A*k>(7VWxUF?8z<jm!1CHg;p`5Oy|L(`4A=XMRJh{ix?-gB9Tgg{Rz
z(ney5#~4hr=?FbHYbc1OO;aG_eH#8Oh!dSBs#yqn!TAL>OG9Qk&2mRCz`Ghcvhl_Z
z2RMD{VolAR`^Ww=GpGwzWlsE;L(bn&$CdlrC(i!(gZKV&Y%G`1x%FK4;q36ksNRg1
zeE73bYsy9vi&|^Z){XAmXA6I0ZWz(&$j19YHgYS(v_d?(V0+*PG0q83X_ZF|a0~Fl
z+W2q5VFy=P4o)+jI_)<yqcNUxl{v%(Eumw3<$rw2oQMC7nLS(=TQU6Hmk<8ogE(gW
zH2F2&c(M=HsdPRpMc)iIn_2SP`M7kS;?CTS^z&mCv1&9z+#IR$jMt-VtRT-Ie@vJ$
zNYNMe+>SD_jYPST1fh-u@gy;vB-7AI(37b(7_Uz8F?wtI=1gNyTa@<dBE9jC0Athh
f&maGg5V6cd`AO4(lmpU0KmP!U-*ZM1$o~94&cHK=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_pipeline_xdlops.hpp.C000528561553163.idx b/.cache/clangd/index/blockwise_gemm_pipeline_xdlops.hpp.C000528561553163.idx
new file mode 100755
index 0000000000000000000000000000000000000000..10b29ea0ed436ae2e6fb3d266d37e569c6993da2
GIT binary patch
literal 13708
zcmc(GX;@T8_BVC<c3;F|5p8I&8x&;|SsM^&5HGTc(WqGrny7Iw?iwW~IvS0r#7K-W
zDxeTi6BW^5lpqpEMUX{BMKp%R1s9Ycn1pd52=QOtbWxReKD^I7`7r%pRsZ^&s$1Jx
z>pEdnboAN)6O)k%qn6H}@xi--iHS)c@-Oj&#j}%TCMIjlO-z<%=D$99VTO<D>8s~T
z91d=-KXxo<$hyc!dxXIJ&2<H@S{yIkaHLQAl)%>0mji{CwTG_Mjp}pe;fb2t!QwY5
zmZudDx0c_xOSnAKsqRKq*8S(Bel0ot?iw*XcENOoW5Cj;hI@*sLGAk$S?!HoVdv7I
zyud&8vtFwf6#QzoYWBgjyo~zHxN60#^>tkh+h1EYrX_dHsn-ho&a>;4J1YA7Lk+24
zRwc(+y!uhz4EZL_4*T{kC0`|bUzr;9rMKOkAm@{ZUzryC`&{!68{BdNX8m?<gP-4^
zsMk}fgcYv8)#jK^XgPc8t*8@6uF4(<=^89j7X9oHSRXjPWWk@&H|st>k$<GNY|hBi
zrw*50Po5k#$?e9cHZQ7jkJMM>?*God{K%W7=!Wl+j9cjE`){pjfq>e_*%ssa|B#XQ
z&g91f$0tY|?K~@gEt=Q6?tn$-dxcXRyOwl(o;CKLa|gb)aN9qoB<vaR+GwvO_e(cc
z&Yovix$w+r|9Iy@>jguO+WN0P@49OvBu<Gnz3}>(lv6A5wUNKzNwai0zW27-Jozqn
z<wp~ze)M$7#M9p#yj}Z7-}cDfi=(Qac)jQ3G2k1!cenP9U7HY*Fyf78x3!MthfBSO
zrUYD?V?N`Lildtz+5bE=b=1(rxAx7dtV-Y2=X9TwHucwQul#&viP@Q-r&TW=RyuTu
zUFx?soy*5Mx@Eol*~2=U`8QTM5AWLF{tJGmPjJB1IdO}3jM`kH6h1b6VTA{%NBFFr
zCY(_fzdQ3;|HUJp_LqVJ3xDe6F>zASm4|hsSB*aU@zKehu@8RNcDUT1uq-~M;M?<;
zo_4sz#lG#gq-mMsMRC^f$ba{#_R0yAQqB)?tJ-MRufcr#6!&m-K>WmKyKj77*AO$d
zF!E>T7>`%tm;dXHh?E@(p1mh+vm5k}Z&S9~u4x@TYs!<E$v%gcxSXA{F!sfl76UZJ
z(RZq!td4!hxyoyd!=(B*{$;ts<n4*y-OS0Ktel_rsO{FIpUzzTqi~t}<68BWmqBmF
zKW`2RdLHWC$Nl%Op#1o*O;cQUUc0$I_&smA=f7gVZLTky8$E60;<4#5%g&xpUwWl%
z%UavGZO3iao1I$sy4JI%eDH-^3mrFaR%g89aI{+c`ovemTtc&h>w9HgJo_^7W!%1w
z%bB~zdZ;#Di<p|yyynlWn5D(b9=x$-$M3giAD%EFx-n&Jocz<si_aG%&RgF9giXEo
zHU9#Iw94eOH!6w~m(8EPQZz9!54mu#%s==}Z}1H_k(!WM6S6W^CZ<Gc$;aZ}4FA&3
zf)AD8N?4X^mbw}WO18F~6Q3TETx$)U!%d_h`yfX{LCJk>UO)WJ3N>jxjI=&2+MgEP
zj1^_E{~_<ho?ixcDxnwY!AIlcV<;%uY2OLQ)81=WD8Zfduw1izkfEUD!7f9;FCHj$
zSwnw9-V0uPA>LR~7Uz_zmmRlm>`{Pqn29t)oZ(_9OnM1G=7%KY`ek*8^>jNy3C^UC
z+4i&5hJv<t_geO5>`xPetYH8l??C-`AZKGmS$v<4PrmCrJ4Xqjgyn*0e}PQ1yP_=G
z2>&wgoqvu|LO(Li08PM9LqW+w)8-95&^f712^OS>L`~vALqW+pCntF|`q;!8RD@iQ
zOzV-mvFc&*xqd1mzFDILEqTCT&0uFkLCFy*x0~kwx^{&%c#|Hs$eg#x{EQW4iCG!{
zVtHy<wl&ximQ%>|6mm6IlqI%g;U6Ego?K}Sc7&w}yozA5v7#*FAIx}X@5Zo|O0Xj1
z@YHzv8VX8&%_c!PbN(2a5?n|RbL{7+4Mh)mz3HA0s~2mm!G@5Jp#Dcth_Rw9lea#f
zt$eyAyf+LaEGx}c+8GK;e%r$IMZ&XTxeDk*wtIj$z{gOS^pc_O?VhqNGpi#)rM3Va
zX`hNrQ<2hG(LSc6`PLnKJ3CGZ0fgnO%;~Hw*jV+jRDJOBdGu1(-XJF|mqfEmqS9DV
zmZ@?7%r6}9#|A4<6P9%3mX3mq6=j(==Bp(S*Zw}*8io*-yCB>JUt>jC-hJ+~>chBk
zla!z!>*=C#(HaU$PH0%9JNLbFnFR!s9`1=&_e2L{MOkLoU3+;ZxLIZaf%=DtR(A<Y
zcSTv|+@H6pV(JISZJ;k<d4z;VXppg@EKB}5Fky4z9nl(`2+I?2egXl;in6S56OKlC
zsLhmMMuzI4@$fYiJ>&-ioqu;pFSiB<(!&Pu+5nS`6=hk~f9#wuKUjW72_j*6U-SMz
zLqW;Od+sF#UTX1Bf<GZI&@6B^6qLNaVgIegV^oQ~!H1CNna#5^6qLNFZc|R2t>Z8y
zm=f}0&EkQEf|An;bYsLZT}et9OvtnBXQ>SZC1=FnyzROEjb2JHC*(z%MFR~*5BcWc
zL9R_<A1PrdA+OM^7-T3Yd6&zuNi$8)&9nwbLheLfooJG=>S3AxvT2W0KE@h62+I`^
zu7J0(qAUl#viZ#0{@3$L@FNR6NHfUUP*C#Wr*nV)wxPIJZ|F^Ws1eO-$eHY}D9h1R
zpN@<FhtIc4up~YBY5aT)1tsSP{b?T+`T7nk@FG22m-V|Yvo%(frC9rPa_O$hX-Wtp
zETuB1QX(7ORS(PZMXO8VhkR}gO2V=a_1{N?xVxe(w}v;B&$Vk@s02*L;ihrZ8j2qB
zsvoVg-fyy3LWuq$qJ1e*sNEH1xx4g~z4g}ILTk_vmNf86gSU+pWx20eb2+7O<YjBH
zCoFm3l}C(2cSTv6zn}U_b@kyyCAblmWtwG!4Fx6tn)Z?Wyv?)^ln_nGOYE088wyH(
zF1cUI+WtucIgex(P8ug?LqW-%sXIbmeQi70u0f>rO#7K?LqW-lXS|naGQ`A05Q;ap
zB$z4?me$JfTA4K!HN}a50ijX3Cjm>>L3<s>?lK+`EHyzu6O5zk!>soHYsGIZ4<|i*
zgNA;CMo^n&?z4W!oNd1Keps66IMbO92urgZXF1aYkEPj;vz@6G#L^tcInGo~V`2c@
zP9ZO98}!)!6bj+-h*Kz<#}iJWw-~NRL+X(aHDH)nt-yK|%HxQ76wBk+>(L~JlfgC_
zG*o`Dlng_X!H40kV7nDGD~!4YOIu;cR`6kXH}u~P1L<kO(r)nD4gL&ggDx8+>PKMd
z0O$_zI0tk&JU$4zgFHS2x<fqv4s_q~_!#JpfkclYmI^>uz~kefJI>=m&=vCdN6`Jq
z;}f7e!Q-=_I|~vu%vdS~T`7;tKv%}&a?q9YxB_$)JU$1yb385*`xc3I)RDtdkvODC
z^x<(>kvNj!Vi6aM);!h}i-UMPqF9V(xJ>L@CfeOH3Ots|#35y(50Asj#E}eFh`2(u
z=CP(i9K_=h6=F2Qm15sY(T;j(SgI6VD#f7`Qx8~XX7qchw}^@P8~g|jp_UO#Eog8H
z!{+(fr`_c5pESe7qzmZ?>z!r&2TRRJXy$Pz5;}Rj0fY^}@uz{1#^b#p?B#JD2zfj%
z0-=b<RUlOH_znnn7^V(9nW5f$XY*S&WSKuze^~kgS$~1llxnHE7U6UC)?Aq&wS#Xv
z_)(Q4NFCtU0a}K;z`qLuD7Gr=tomxxeMB_T0=ESL)L_#v0ht-X1L~vEQy|0ALm7T3
zv#0jQ%orrl@nGqS7;!}$H^u06seV3U+PS~LPY|T(=F=@Gn;<2aCvYeih3BFX#YS@^
zNQEf85RG`v=ws5zgRG1`?4X0fQk-R+6CE)o3!rHSA3DJS1s{*T?iuqPnP4W;W}@L#
z)jEa$P}mk!W%jp*M5CSFW`-z-&I?OfXjm2sqtC)p1&XLZu}sUhAzd4i*kR3qh#ZJz
zTJ1PQ9EVt@g~}kJ3}TtqsDa@%5XtOI4UDgWi3~r2;g2Aa+8ZJqF#Zutq?kU3Y;Si6
zMh8x7Ff+tl-|zG|J>$pC$MnLbcQGlM+8$FgL*zw&#?s=D#i7(~!%|{MVkixiu+)g+
z8qq7vP$Ypi2?Cg5_!6{VLI5)aUxW5*2w;l;0camU0G$F|TDh4qwqt|HMh@A?jh-Sb
z<spYW<VH^s(TK>Q2)VJnxQQHYA~!lPPrsD>)z+p1$bp%sp65Y*RL_J?od!NGMuOCa
zg4$3B-8=7FL9z26-};Og{RcAj1DO|<31akR>L!MLAMDKSKdp43sUR)&Um8H=T##x&
zR|BDR?*yq9bhSKw1iD8MN+&KzEud@R@uxDkPi5}(zzfm_nfnIW5Qe{yIe#H@p_pn?
zQ?et%fJVWqcSYbUf2`kMMog)EzyUdjTjlCj`4C3WR;aTT9IjBPD|oz4RId{``aw~B
zkYO`pgiSY!M%YB0_2D%W=T|IWabV*75`tbr3>|%t^K_GUEvGIa+wg@f^b1*74<RBg
zBClhh<^Onf>2W8=QlAu{bH&mc6ukz;(O4Ku>rnJM9^V#&Z;N3xBEV9E7~H^cur_(Z
zt&)RJ$)Rip+h*{iry<0u`0M=bAN@&AUx;al1zWaA)5sna(sed<=5Vt!gz1u~x+Dfs
zwTwx$p{fys=wqo25WUcyv!LVG8Iu)6FX%HBY_EJxeXXco)-7B=bMG(f7ZK~Q898i5
zUNmyoONRFYSvW5l-c1aLT7;dQ8a;TKNR|aG*MT2X3R}Q(3-~dmunR1Afge)}*?_ab
zo9Z(x9Rqv}yqVrC0bBy!OmAKgJuZl2sSJdN3@GtgQuqD;S9WPOMldtxVCceN>A4&~
zm)o(a9aZ3?3Jz})@g}hcM^5Pf?TTlnkz^ykLF8{d)nidV%-xc<DLh|^rJn@#Pl78G
zyX%7bI*;!P?so-$M&B-1Z<llYJLKveJl-dF-zVqzZ^~6S<%3xJTXNMc9yiOKo8>-?
zo~}@(D>(k`3e|QV?^QVORdD>r6{_Pry-=Yl<Z-FOxs>OBs!%=U>CY6ZXFN_4os&ea
z{|r%;A#&r%6jhl#&K8}s89&wY<V<&`$mrI^T+P&FOo~eM(pzhsg>2|n=<(1j6v*%{
zgm>}u-3agIaW1mVMRv^M6e7z)WXCK{1;Q1`hK@w<4R}?cAcoH&e2%A|NBBICFCu*L
zuh{D%-~JN9mw5VRgfH{B8sTalUqkpBkFO)k>&T9c=RUH$kL;*6(8q8%8ElwTZUwv*
zY?xH;2D}?=c&YTsh7evVeU3p0FO@ze5W=Lg25=4Ozb9I*0k0YeVz?H(YI$4-xQ^$q
z2VBqNX28v0Lv^Iy`|)}NK@7KmR|}6@0k`t}j{!gCaib94D2!qfaX=n^K+fS-xvo_n
z#pu}zUABV56$)JikJpL1bs|SUDC!RKxJWcD;$yNiqUjkPmx-ojd`wj?nwImpQZ%g;
zt*Mry>r5n|I|EE5V63Y0?1J8b?L=Wux14TAf5uXRWdesjH2u(;CRS*+m>6w+s+bvq
zQI?|yG3(O-Q5_IZZ8J?y5TmG1AhYK*qKS<to}IHbpk2eegquOTnRf|yf_5i!38{P$
zqtu;vW*xVnz!vldb<y7HTkbn8SLIC9{AtnSv=~F<7!rMp9;IRojXvLck?#B8#2Wwq
zI*g)rojAn$1UDN<JK%QcPZvTToM_u2g5eIp9X!1Q{5xPIqjv%B;^|!w*acCH{t{eY
zLLkL-JIIRY6Y0!-U#VUhMaTGdXhz>X@}>p<V^4`X2V|_>IeoTh=_oQCji>2&n%pOh
z-I;yHSS3bOiSg8%p*3Ob&hInEyP|Mcw4{l3+A|S|?kqoLq#A+{x^w%KK$H69*y}U;
zbe*vjZ5i#%j(rzsx>&3|{lc+`1&!V^gFoc68Tr#G>-{0}7sZ^0CK(A9dN92sAm_Ce
z4Q@q#^dOrVZwlzJ=!6W6X01rY7`yKaC?owH!T%G<%`4vTp7Hh5$^YR`(<4P@+<kXI
zyTa11g4M6W05<qMxmBLL2h&Rh(olb+z?Lf3GWIVTLel}h^eNPokcIEQXP{4`7Yu};
zzg?g|(>w)vi2hcA{;ap$F<IzUX1UwYm^P*eXfQ{b>2DiYvrOcaiG2BHHf_k6u3Ac*
zY;EhNmqCtLa#A{}X%-nvgOr2RG{=vn!OFpE>L_E$Md_lZ#|BGoN;fslp<~HI>7l0U
zge6a<r<&?jEcqyX)YPNKlAqE~P457(6rc=H)3AvoY?KSsbRn?xzVdxFT`epvQZ7={
z8xRtfD;KM|uw0p_rf~|EmMWL3X)KDRWy)o0mSkG4T&|`m4lJ!uu25$fuTAvX6pKux
zWJ`7hSzJL%=0jXX7FT(EPo}yjb7ZF`N5DBkUxx1sF82jb&W6fevgI6Zkvp}>-5Gtq
z!fC&P!{rL6avrZ0oz{vRJx6rPVVJH4F-iK1C3X^$P;?R>m#sz7YkB;W82pnMMpX?-
z^ozmQ8Qxso^xNta%X9xjoKoK_emLCDNv6QkOPT3QnGKux4}#SXf&;_tf>k?@i{-dj
z-jiv=(kBZ1iGss775FB@^pc>znc+5!O)MAr=b}L7qvj*meB@4TC?>f!|3VbVa0zlP
zLGElnuOZKC$eZB?<l2DTneA>wVT~x9;nkpB&D-*12up@==0hd}aVtkMya}|Mcw4^}
z!nQ&<RSf!j0`fPK;Y`qGGG9P{Phh(n?3fR_8^}*1c^tJHVi+y~T?vFy-v~?B!1o$x
zna|k>+D4Xc!qN^Qe1{;hZORbBGk9Dix)h0S%(EyKCl`zFFkB|Ol!<QC(;zW`IJrW6
zhhi$LM9=B3D(SXSaU=Of{S_tsS+D29vrrV1zKba0;$QQNk!>iBDljY^gouNDez6cD
z3i<qEIYgB6`NdioUd!hfYhipXpI>Z&;VpcAu?5Dru>4}?*B*CQ6fYU}U;9z?IFM!N
zzH_BBru#w&-8Ze2uxDk;rb>%pQU4)v2PRezH!GOyNhGcU*;XJ=c62Wy+lxG|0oxkz
zq>sQP0cTsw<40io2t28CKyDVmwuND80EpY6zu{#DPiv~RqPqLg(i8G=^Y=ve!_qEf
zy9+t8t=o-kck}ohvOR|!*=Wxr+w(lWglsSI_%gD+%;P$+tpi6kn>ui*122Z_!M2{K
z*Tc|y@MrW^ux$lLHrrNkX$3DH2ed*M!;iuCG0*=PTpsiMk3shsMl$|tF}PZkm<GKn
z24Cgz4KetJDAAM@mTro{HyO^oky*0Kzi2LTs&=Blo#;)rkmV?}9KFpJt`db-GJN>A
zu7RV?lje}j%Vs2OMjGmk(I=7p>%M!ZM+*~gOSmUmQg@8*CGj5g7x8SPM_Z2m%l$)S
zJ4~dW4OJNA;C5fn)86wR{WIToLtD<@x_>zSUPEQ1+4(!<SfwNLbY!c)(b9j2zCywd
zWYLZEcpnn>@%RW5j_~*-5>E2C4heNUZbs(KeE*%uypzWpz<dL6&rbvMG#>8-^SwOI
z1M@r{7lC;ZkE_7EipO`r{0@(wfbayk`E8L2TVxzgmkH@S&XWmwJiaayt}}chbYP$R
z>z{4(CYf59=^A#YPwcH=`iWpo_Yq601-x3|?Du-XYQ4bO<70x=F&-BP{R(({eo9cC
z5}a6n)q+(u&wo|uca`VACaA9Q{568QMsQ*6I|Qo^zWocK-wVEdl3bM}=bo1;w@T&x
zi8Oh?G&$#gq{~(5Jb$KKohj#@cUF$i%DMTM%B@QI{wn4DD*68E<hYLKua{fZ^ZYmD
z{ciC54RTcj|NJBcPEv5sTcfaA!~7jAZBX>vpy0;8O@X)Z{HY47RGvRg(JzhX&rql`
z6x{mcD{wyF{+Pn*7~j4?(XW7Se?p-;!MDGyz_<DK4GOCUzWqZ*zlVJLHU)0u`Two3
z`Zv%2MA7dF&)=?4we!zQ7je4Ct<QGRYP-mt#~ot79U`|rJ4Mw_v1h*js`x4XJKet<
z|CSmmc3tT^?1#9$|6zpb%SS}FcE55&wI7!1#E3dEo+@AJmlECD{o)a0Y(nZLq@x%A
z`m4ymM=Yc(ufB5OLi3ZN|Jv2K9O|-S!mH}kzculiCz%swrEmE3zkWK61<8E%FH<pF
z5@Q+jmxMx07raAl33V2U?fXTBf8np+yWh29i&ig2)QjWj>fQ;uJo5P=75ZCqrLHrv
zWZiFX(WlY8E+KTktVIblkhC&0yre~c)~i&TPXrB9VaEiAV}dio)q+Dck2?g14j!k<
z9a7~S|5>^HS@}TLzD{mm$Kxc0eUgHsZ&TQB<8eNbHy+<s*x%-Ho5H@0$LXSdy2$lc
zB>EPKLA<Z;TP6n4?V`C;m6_pXG`0*&1D8rH^!^-0e~!k}EFB4Wki%BwPQ4T?9YRir
zkS7%s@?sKlDnOnrU@JwMQvO=|F4Ej(*V<V66Z!m!bkxJf(&ymwIp|nGwH<u6gO2Vp
zmJWmEVX$XO{F7jL671<_W9b4|UI2R*K-~n(n_$oU$lt*7H?U`Z-(6w&T|SAvPd<Df
zpG0q#hcxp^^u3CZy?oNVR1s3j<0LU8iBFnmiy_%Ot`vhS#Ymb2Z_6w^?fw10!GB8x
zo+XbuKKT66r~l*BDNR5T<)MFvktvVB;6RDxhe9nwZE0@u`HRDcY@6qAB;Icpvduys
z)NQ7(81^+Yyz9sg=ldb=hjRI(kVT<1M#WMNa?L^R>?r1<vAJj*!-Z&UAsWYyW-)Rt
z=5q<v$g`TyB~&AAH40#!`!%G!#^c+__4Z$L4L*&?kL4OZ0qrMzuHg&Ne!=GwGC-Tb
z=L)WZ?^Qn5zYp5`e60UM7XCuU#rmDH@J@yo&zSeF-mgwv{Ex&K{g()R(fx^{U&w!{
eVqZtFuaP}pLF&KfTlI_I+KT0oQ4u3YnEW4H>6A<W

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_wmma.hpp.33F41B2FF8044BC6.idx b/.cache/clangd/index/blockwise_gemm_wmma.hpp.33F41B2FF8044BC6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a980c11856836e9014b0828eeee43b3fe6a15713
GIT binary patch
literal 14406
zcmeI3dt8)7_rQ0~JQ^+!NazB>qFe=$g#{kimE{u1B}Bt3p^28Mm8KT<`dX&u6+#0=
z1<K27MJ8S#P!}<c1&Yc`givIGsAN{!MOLBQq~GjinCY+IUwz)sr_bxBp5eRa>^z6L
zoH=J^n3|T75~r2P#?Mb%xn#i$PY5!Z>_PJX@)vTRxePMdYPn3da?gq4>}9*v!#Cfb
zmeX=>(at-Iw;W3I?(}K%-~IK|XO4?|{4%#DL`fG8-S4~9gmC%SGmdV2XialT^M-Ll
zPDQ`ele#nD*SswWhU%5eQ~xw{b=SAFJ~#bV&K18+_5Uv3KJVGK)&h68yGL$%e4G-$
z>`v9EYaFVpXG_KVm((8Vi!cpp_-bo^?Uy%H_tTGd9e?4^vR&V=i8>tp>;25PQbtAf
z*KEUMf_^{#{J*b9yc;~lzwWKW{!ci!9sAh0^7Hyu#rl1zO5N1P&mP37w;o(v_Tt=A
zKcufp{OP2AUb!|TyMCK?WYWy0!QIRDB#CSMF3J-g&9ZA)urIANaN(vv#i`r#RIm8$
z2=3}c8%};AxOaQZ?OJ+FC3pM^FBy6y``;(taXzr^z)O9rub+K!tf{1G?E?-km7KKy
z8r^#3U#~=YdAt&N(!F%h_3)hAC+Cme_>sTwpsc9)LD^H*Op-e<`th~cmGyBi?zrFM
zKIzTUHap>h;)Q1)YdE&5N49dsqbbw2?;5A~yjI(CW5$F*Z>=aDbGq))+HI9--wgfb
z+hun$zYYosEJ=4N__g1(cFZpx9-r&gFD6|bB40h+amvd4hKS+&e0Mzh)`&}n`9V(~
zKECev-)29lx;ggi-H+~CnqJ_tI7l^hub?`+r+&~y$9uZOzGo&Zb*kU4zm~LQ$g)SC
zI}|eN-SMYy?>ef!t<9Vpvb4C#HF@vD6(OUR4;yj9_w5z)PtQ%cesSi^?uAl#&1G@r
zoOtysx3d1MbBp+F-_UXDkLLt04_r8RVcLkTG0l%XHD=}FG-sb5Hk3T|m{aL<4J(Q!
zgnl+Pa@w73^&U>j($p8vpA}B8>{b}wiB5VoHbd%OmFC}i<7R7gb64df#T&d1oDB+%
zHQqb?{?8*Sr?;*={KyxnE;GbwS2jJG<Fn$ciucOr>ekG^|KpLbbH^<1^n1H%*7seq
z$hx>3`^S88GL}bu*i}At=H!$udsj?|C}=;V`({)B&GkR#d_O8L_uS&;FD{ukbFfSX
zYj0`Je^=$N0M`(iB)LmU%R^>IW&-*5fLprv#(NImC}0Q?r+TJ(TOL#_rf<8qVfBL(
z?ZJnLtB|}3IoXN<m3`aoiaG;Z6yQNrwC-A^<w3=c@sp%2yH}<<!%!m5Lmqi3+E!2%
z=SA2fcHGai6(A=n$)3r+mIoCNs~jCwxM}j+&fq}A3p^G?TOL&GIV0^zc4yQU1^5zi
zyl1?><w3<GU7Y_jqA&S~Gq@0OmPeMq<w3>14t*;fKlwRU0RcpO5W645fwqFG1O@+D
z9GGKYX%FMVWl|4z?7_~qf~rIfUAW@;|J?Xa0fR{`DV{0bmIoClrWSjRy0xg^8HN+_
zMI>BAkJ$>UlA2!GbM{(u=wKKW0@63w{u?~fRt%_oF@AdLD>L^ygEOh69|`?vmaU*F
zljVWW>X$t-aWL40So>G{3ETgKM^c5Uap_IEBgQ007(?pVE<|h>5^M!k$$BFr$-{KF
zNCDAACCxp}*Ycp^Z2V<y{`Cu;3V48slRT4rEsp_lUHaaQ^>KHd;XxuU1fdXS*$S$%
zuq)@&-`|N0a|Q)bX+%OJX$@;ZRdQrmKR&+d=l?i^BT-r5u_D^?pyF2sKXqm2gV#ux
zD~WiV`#7cLLB*R_Cp}*?{lf*$Fp`MBN5c1Lp{<}QdGp8pYgbg8+T7fNbROHEC(Ugw
zsLGBB9Sau@Iq{phEd}XUZ2v1!u@+S2E!6q!vQ;mq*n>BDQH(rhgyliS?>_wgspZR)
zoXNlmlS%6Y$900Ut)ME0*Or{SFd_RF7Z6De)yS<Hjj<I}rF!rVZ^M*#$^3F7Dv|Dy
zO3Q<aPY%DT|84)rYspL0GU+qy_!$|U)`F_keKKOpH?KZ3Ujgo<293K$X?akw;pyk>
z7tOf&v?Giq;^}tN6_y7TH{{pZkGKdX1^5&3e(>54I$J?i&dqu9^WG&HBNY%#R6fG)
zACXC9EvU+82R=Wu>gk3gXK*4aWyqrp#n=j}(wr!p<x*GwxB^BHl_>WprR71zpBFe5
zpZl?Qu>$;vINCjWl;uIit&b!t9cR2;rvOhPp6EHz*Ycp^j;C)Gte*A&8MdQ{c#`KN
zU(17vyX{|IH}i{`&nO^}h~EUSH_0z)EvU+k&;zd*WsmyRJRU&$6Wjku#)GvOP`O$Y
z+q!#lsRDvXEqd&(C+%P@sLD?rt!<f2Dvc9(5tT9+UIwnVf~wpO`e1wZyv$FXU>H%!
z49wJ89#q`xw&vL4)rG&44iA+{o!GIHbhx#kD!)x0Wc+ZW@(=Uqwz+{%vHhpa4fOB7
zxmhN!UN{)+RWd33fp9m=LpB5>qK_oh-xQ<Sb3AUgBa9|CY!f235kFxq=u2cd3l=Wl
zOzNjAj7n}<S!ms&v`Domlsa!w+KK|UqA_$i5T!jxwFiar9I*#QGTw(G_VIiW1sp_U
z=r|Ik3KU#{LKxpfVK-3(o%^Dshsk=FO55Aa)Af+abCw=v@jP1(a~Rh`KrM{nIkXnS
zc#f!rNS-ydFpg(kEyOZ5z;pxuegk9~`1c!Nwt;`Y0UkHNJof!Y7~2Rco+BFh_cwy3
zk$-<9=o<O=H$r?PB(U!rVX~2LA0tdR^7a`a%gEbjgxN;kzFr9E<=dwhLVNl4>4k`1
z-o9SY^z!!gg07dh&jgcAygeqEZsP4RL6(WP#{{!YygmI8(9he`525|MJ^c{T&)d@v
znttA%e$e$Zdj`KB=@pY)aR!OfFG&0a1<-*-111-_B~W6c-2ib0NM!BX2AVd|F>VKK
zJH#-)1e!~{+z8P|h-30x9FdDPjPtRSkF|`q;)t!hyavbD;PE$XE=ZJ4W6f!-V|)f{
z&hT7^HFZ22u*Sf1J=WCo+=t`(a3T$cM5zzY>BI9F-@~eVIE=9gt4uuS3ZA)wH(e7&
zDOb?u3NeiH1u<W6qAb7k?X-rhRq>?#p9p^9AsWTecofTRQ7CKwY7|tB)O0}=rDG`Q
z7{|6K70E5pDgBy*I`54|yTZPCNR%WWDUiA{QHu1545TZsC`I{11=8h7lr%n?K)R-i
zlGaBXDB5ZlrExyv0%>?6O3^;ifppyxNz|*$LkY~o<)PF(G>LH;iY-IqnMW%_X=Nyb
zaRm}9kTcI=6-dMQ0&=~8JZX3(N*7T01*B!%fm}O~Ck;kLsRM;~AT8rg<l4#0FC*8>
zJa-}2E}qwcZXG0~+QL#%S`WJQJZ}Kq2A(&9ZX?f|K(~qK&7j-Na}nr@AR*fpQHs(I
z(Cy%PC+K$aybE-@crFHAG0!ESE8$rW0eaq}8ept}_Za6O>>O}2uoc2udA<%w*I@#i
zkwyqG^7WY@z{K-rtlEqtn0^6P74W<tKeQjGFnKvvm*Xg!p@`BUtUko^5v)GKa|Kpc
z@LY-2l{~+P)$j3Kh1FF&S7UWG&&RO(7|+MC`Z&)uSY3mon1?xu)hBsAh1I8cK8@9<
zc|L>HXL$Yqt3Tkm4y)^UHej`Z=X$KJ=eYr^8+dNQ!<z628k~w!6CT}!gDE?lI9L}r
z=(#fwh*GAIDO%$snM@?%w@{5ZZa_i<&n-x3;rRv<Zt#2`3HNy}0HFW|>Z8-0I3jB@
zOkEMpf*s|S?3aElN(M+aKsp;1Z4lQ6NsQa!p?2VgjS&)zkjmt_I64<U#5f<v<YR83
zs=+BWcp~e@Q&@8f$1*;H<IdnD#&tNZj^{p{+=tVd<L|@M`tVG~xq@G=5JVk24G4(`
zH>dLK*Q`RyDx{+h-rNl_)#x$m&)wRq?EZMSCfq@gI?<R;6hz&YnIk(<6ywWi%w=AF
z8AV>^<y~k@7ccKZkzKs}3i7^!d{})~k@r=euOaViJnKQNXKq6v{YCy82n+?u2x=oQ
zH-Xy3*dzS*-Xu4<J%}WTS9IbDG`<%}s;F2eOrQ-%_c=(znKwCD!)2#tXAG>5QZ1TP
zi)PSfqn{6wTXsFzm!9!^CWLl{D81zOQV7ktMJXSR%twKApo-D~H1Yrnr0a}V&Fb_o
zRjEhG$~-G#RvdkUPyF%P9_{Dye_RJq|7#~o<sg=WJDW)-L3t8_7@q^}IesSI2jzWU
zUVx<nHuL1ReJ0i^7je`@oJ2EvQM!s1SFtZ05OhWn&tct!qB|6-@GU3SIoo-W9*I%L
zjHSC0s`-EDJYh9)Ut~~dYf&^E6rq>34=?%o@~dRX9R%e;2w}g$Ay6J-9QNJ53XjJ>
zTIM21JxJSw#?c^{3~Hn`@~j7~o<-9F$+qJ4FyM5|lOo2zqH8mcH!vqgwqPLgJVa4L
z%p7x`*KdZXX14YUQX53J@%n8L)5hz!LsUDjZv?H8x6cUiM&3RXXidD_1o0+bo{PhB
zaX7Ot7i)9zK(;54&^$b!t>uDLh{Fnb{X(oQ<n@bicoEYNfB4({!6Rly*$GlJ#5F@a
zdr=ESw(!4bJH)l~@=FkTiI?Z&=zKhob_mi|9JZA=qzFeB@$#KGY$ub`jZBi=n0GPR
z(9}b;p3hFsgXTP+y)=Wina^%oK-0p@O%QG3zh8(W3i<4%2unqLcC!;l?BwO;IKCW@
zr*n~H|5#Ijb&M;qrjq9>tf}I;8f&U~K87{Ncy7gUtvHcQy;eM@70+Y*C02cj!x*<?
zRXfk!*s~jZ)2S*--B{a=V;Fyj#qY2aWo=?f;<0<)L;p9iq`!A!Mc@Ch_^CK~J#lr-
zSZv0@tOt!)G~!_DRCUYpo_gnpyyvkfWsJ-Sp?jic4;gxp%{LN+L+ByKF^;hx&Q0-q
zQ4poGXy{oK&a$m_Ag+T@mK_y=SOlRgoB0Tj{RlroqtW<iuSWI1zJBlj^~p3YA*+D(
zAc4-&q^TZdFZbFdnKR~uC#KMKIr;eF9VLbFC&=>E1j;4|Xicl0^=8COUs3uPlppil
z0r4F$fyp}|zLV#5*liv5W%BjdZ9UH=SXY8mn7kD0N*U7=4q~|Vc!L_A+PX)#dsg!D
zFj2}Dvt6j+go3l97|jvqaI(4LT#l9wTk6T8nVexc9IYL;)|16H8_=)~h~pw879kfF
z>lC4oBBW)!1Bp9$`A#J6<hcZiC4c3R68`(8NG#>$uOsnwp7$Ve56}CMxR2-kXy|@4
zg4uHv4LyoR&{)iz<BJ!N3mbJENbEo^bdZpk5{aEW-$qWikvm<-MY67s?Lm=@jcBZq
z=R3&h4zGU~Io;)X9jMnqG(XDKMG(zKxgL~y7)_TmQK|)HEzbt<GVmc%BX~9Pd>#VN
z!@#;Ll660@jpr*6e+AOmm@|UX$k*Qs%3hvLpfvH^56XU?bFp78<^rNZ>{rP1Zmipl
zliB;qv41%Z<<}wqN}g-5e+>?0$<rC^e}?BKY}bUHsTZK_Nal$3e2>nNNt&W%E~>5D
zh<mt(gRfx?T}eb~vk<&l(9kS_o&X9%Y$t!T29cg!*%Domo?Q&^cEMr0;6+aeX-PcW
zapZR#h12L!ls*9E2Yesh2DjU+4du4;MEc^aVIPM*Ug73XJnJPi^b#7&24^|8E60AU
z@oKPL4fdmrMGqoLLTo;Tr1go?)R9vIn1f#Iw^+sI{>y$ZtJrKWM4p9cG@Ct#k>_DF
zn$4ajBsL)hn>}A3@e8D22KJ$$eaMZ?jyxEe2X3^NMCo-H`Z~C=(S8k-*TA2R_Py9`
zFZO2KgvT}EMB3Z4V|EXFYtZ*g{}%=P%~1e7vnC_Sd|=H+k{%ND{8q_%h&#`+5)w1O
zL(P0bXMzM1uV09x3;C^;A{<l15<HSH;*@eckqz)8SaSr&GOomNl{krU6^^Uoxs|vY
zoX%pFRy?g0&t%+<{km}w^#ybs67OL>c&ESUymR`*rPo8TIk-`0NNnyLgoJ~T%*N*-
zNI1lp&I?kX^?07vM<*D_EywcoWAmPS%rUlSL-)bSj)nP*o~fn8ro_`sLzGfuQ{(BW
z%KFRguXVPpm`7e|Z)ZP@ZpzUaNIte6=hM3AIX)3sPw%N<%LKgfr@}wAqA0~H;+1ss
zM3j;gNlH2%L@8O3tfU)nqLiXYQF3-EQkB#$QJScjsHA7iqBKb{NlA~C$v%kVECmZx
zb|c5#NWq-WUgWr!=PnrC1uC|Lyn%h*z+)(H`7!Csrx)h@@n25`3${3i?|XZhUJ#@X
znAQPV)Izf32GcruF2NZkcsi4p;*3(pben?ok@fln?IXHTLB@yq?gagqw5QLkFxH;6
z$fXwf)8&UQbfh-xjS6}$U;gODIC;%t6<M%~kZTcAu^uQvt|dJ8py53zhz=@|Y-S8M
zGNx}rV7rMy4Lh>({XvtLyp`fCN^@N2jA9nfb)C!64&=H6jbd-#iClN`T#8&vd43(a
zzRvR<<Z%c2GVAW3pgSm>@m=I`mzUo~>bpqG<gHlMiX)kmXv3;Dp1ZNwjfYdGAxht3
z@mrqzu&NJ7(!@xVe#5HYc+M5XTwyrv2Xd?>h<S`FI`%Z~mkf)EubCY+JCilib`-lE
zWwT}}MzO_=D}TN3o$j!H5y@UMgbYvWqUaIaFuCQ@3+)L}TBTYQO5Knsy@mo_<L?&i
zL#loJ-GXuyQ2y7u15+x{G!_S3KtUH!F!guls5;~VQuCaA0j2Vsb^)a`?m$5uD44|!
z9Vny&sd-NBK&d>Zb)a;{w@}zE{_etItUipx*}8BPtB>+ri`BI_oCZpw^gdR<&-ld6
zmrl6m_Z<4iZN<_tP3FG!h7)!FCl_8hpCmtf)q$MlIAprgWkZy*ge*r!3xow6tq@jl
zl!t^o#BmuC%6P6u4%LXOxdAye@Z5qNT6n&J9B%M@A35CTxd0prfUB<zgfbX-|NAlC
z4}P21`x?pG$(F!-;KJ7EY8YM(BN?9tx6|Om_)}1R%A#>mx&`7b{`)_H(@*>+_Rlc<
zXTH9T*li;osPFVAU(7tJGhYA4jU{@3PhuGJjX366zCxK_p)5K#sUsq4=9_U$Ei)-|
z3caaCugH;87VG^urgj39C)hr;DBT0)J)T=|R12O$2ev3(!HO$bO`S8%6o@mu0$x`j
zgs!oov>PjT;~>g%+jTqoV(Qw+OtIdzqk`s`(JxQUP!|8=`jcMI5agCygY@?`-xHft
zfgYy~B1)g2%ui4j^WE(zvz_%fy{$pI(tO2`RrA*vh-|F<B?c0uzs5kK^w$`OMviuJ
z%dJFKZMG&mmaUORD6t4V%+|<Sq^m^<G$1kW*64ar0$b{PP(}}$!FW9+u7`(Ne!2q^
zcfi9mFQ5S{F~WQSk~1PcJDC|#f)Y#qYD7{kn!t=OBAt;p!iX}Ayb+rqaT9MuF(ei<
zBd%(j$M+p^L;tt|NUxBRkz>6$%G$RMyz2PFkRDL>@J`?^j=GDdu!w7;px7v=X@o@s
zTr!uf*G*}~bp0ob#9<H*^G>}1yc*a!o+!0~wv|t0?}G9!zlF0IOPg8ZDz{x)r7set
zUL4hnlbD5@1jQ!7ms&^{AG5cy+-Rj=qz;k@%=cL7$31&$>q6i2_WQ?;0i)9ow>G?*
zA^q2rH@bk3hPPe~rn~RGtDDxinCgfA_buF<cGB-`_1*t%e*MjxUwvO1Hu(*CQXrYt
z1iJ}NbXT>{RC&bp$0e_kkF#B~$5QuCua1-bW9yZ1I%h@t=Me+@#nw|C7F}->9JUEw
z^yWDI9toVx_siegH@0rUsqK36y&m%)IZ!cGw`?=Zk0Up>&Kw7~<9s0dJGlMMSZ=$H
zPAke;@az+2TbAcMxjdDWAk9{kEj@F<qjr!sN)r(&i=vjwWRdL02=*iSQ$hMqLw}IT
F{tt>vMza6_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_xdl_traits.hpp.9E775B78211AACFC.idx b/.cache/clangd/index/blockwise_gemm_xdl_traits.hpp.9E775B78211AACFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eab3d4715848a236cdcf81bfe9c9d37e16f737f2
GIT binary patch
literal 2320
zcmYk7dr%Ws6vmeoHY7$-B*G>U$>JliWKEoyz!tMYx`}RRjRxshtw?!@RIN<}sZ3i(
zw5Sy;MMbQ&t+ZA;KJmdHRILiJ#ZhZPP$JY9qPCWb)LM&9?{4QNcV}{De&3#Z_uR8*
zzs#C$w}(qH%$75~cuDT^tWXTYhJ&AXxo6SM5Da?}hGE4!_ots;y3;VM`}<5|nYx8L
zWZK|T*3P~Xdwk!@X)WprU(MMbHaI)ppU_a=*IS4w>NGPqQ*-K)-|2dsoxJhLDt2yD
zzQiJ{zI>=sc6{DA?yM=ro+~ZeGBMhGDShvuHTTBXHXLM^R<ws&l2ZEh&*W+OH4hpM
zx~WG_<qR5@(#%J8__huD@NwRYyiZ4Wesi54@J%UH?q9d}yk57tVWCHvu3u1cw2r!c
zbd_&X=Hb8R_qF`d-u36r(tFQJFNEF7kv<uowux{JPv0K3WpxEveSOdF`Pk|`4!!M%
zTwmm!nvZW(9jJI5AGfgjuQo&F*<F#@Uu<=z`8(ct`QnvpC!%*`bGUP5-=2t`WktOa
zSMNs6EiWG5RPbm)<L77Cp~acyWpDQs{Bk}its`dbj8PlOiMHn_+Amg}ymHEKufN|p
z!!WA+^%q&){_NjJ52P?p=(et8yP;~?KO<8Oez&3BZMft%G`x2DIumy3K`Z?<w2P~k
zbQaz3f4_Sp<GcS^GkI)7WmMKx?Y;@KvYz&v7Wu|KjQRHG!p?cy?%Y+nTNn3c)M*qo
zQ$J`;YgnJ~Qti9vb@VjpEqa9XpxaY<=0NWJ*}ZYCC55m4cHueOeEZzYSpUq}71K&O
z*(#TM3+8=q$y?pvO*0+g5g8ee5*7G}XpE$k)L1cK;b0oT#rB-*>FHvQ68Mmj7&%!o
z*&+rE%nq2B5HiI5AmZ1!f%YE>JQO6vN@CSwz@&q4uscZbwlC0>Fkiny;4+YqBuTP}
zLD&IYAff2l$tTiJ>YE8%0urW3rdY%fOsH(VRqyXO@sR?L1SQ5P<1}J`B_5FfyZxuW
z)0qSw2G;WuK2Z$8^@p2|ZSwRNEF*9^NYF~O7BL73!^8@X?6R~y-2CPZ0+)h>1WAHL
z48a{5!_T(tuNguSH~|t25`#qy!Gxyt@7BlrxJe3pC@67^a*SFGu*A0GEz<`SPwNzT
z6u7NXYBXYiw>!UFBCD;~4OT0_?RaJUSTVrcHyd;A<gdMUm%zip)9n&_q8Nftf0&OK
zZF8rB`UiuAR7q;07=(lf@qm~oH_w{^KJp=kbvp{S8vS}XsZ;AH*cm`t9gUDl#~{S&
zID~keD@b~^o`UU@lX|V5M#!XR5MuQlLcHD;B%|6$!5+&=qt-|xWHK@cu|^Id-slRF
zS#74^yvj+l)=VR0GBXIVW)30V><SX2W+*tba+1+9G(skZL5O8Igm}gkB&*s=!NHc3
zR;`sr$Yf;@VyzrPyww#Xo7zUfmn$c2S{sd!$;KeW+Bk%Gn=42TwS$5)BqtqO2aS-)
z!63vsID~kID@abYlY*a`oOEiPG(sjPgAnWF5aOLKNYHFTWTMGH?|{t*vfwbwSLE@f
z-m)@yTP71N1<GJp37rZVK|-%WMu{-0kdYwFD!4m%nFOOkUJ+qcA+08CDx_tELxmij
zaH`-D;CfK2;AOBVxDk*A2ZJ)SE+7lO1y2Mw1G3;(FdLo;$bw5j86FPEf;T}Ko)XA{
zBS9Hj5s(ERf{%n21Z2T~pbV`C$b#=c8J-=;g5SWsVS|7yxD1p%xr?)^ftkbAfq#vF
z0J|0tV7~$aY)L?X%?Jpv4FLf*ARxf10|KlU^Lnzq(V#+{RMF<-t)-X8gv4Mm!XME8
F*MFgNeZK$z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_xdlops.hpp.E5193AF5C1961F73.idx b/.cache/clangd/index/blockwise_gemm_xdlops.hpp.E5193AF5C1961F73.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ad152908d55b66aa5cf5550604844032d060ce3b
GIT binary patch
literal 24308
zcmdUXcT`lz`!{!P8C)+#WLa1#QM$rX77$q4F5tC9V>fDwNi=GVF;SD4m_!6Mh#Ey~
zvG;<2q97=Oq7qSziY*2sDt2s$U8BAOyN@$D`SW|;bMo_;oqT+rJ9qAsr_D3dyGy4|
zw?DA5N*vl{!sub+KN760tXh%(#*H64a;Ru!mF8$=HDT$UcQ1@N7h!*9`)gI4V<H}!
zKWxLOHsi}b_x)W~K5I2`)CKE7t4Hs1@K2xTw&lj_GS6XM=H(QO%qshwM-|WS;(KxD
z-IAl+9=>QdZ2a_2nbr{oAOAtsk$1-*59sEQ+5S)CUH`5(+t28n5%)puTbY}@`c&Tf
zWK2q_{`R9*m)iFBzZBYP?lhY~k5|&!_**Mq`Kw00wap88|5o^3eme3?o9Eqi?QeZt
zvpko6<#l4t`pzx99#kGTZ49r;oBLP&wXtW6y@!rWsBSk{H?qc<*2mGhZ0P0y@0MQ{
zbMb9U6MJX&4Jk9(2laXT%ge_0i{Co*YJGZuIv}~r)=pzeKN@g-e{tWDP9v_`j;)P(
zab<a%58u6d>i4~OAGG-X@uo?6k6T>zT7P4-SGx8WIcCH@<B*2!DvzUWZ+5Hv)B1zH
zslB>4Y#w&0Vy(}_kuKrGwj53=eG(RKZ1eZm(_e)Db!UF+$oDUviQx=Ue{G3sf62J+
z$8N2=_O)HnuO{N}eS_xhP1RgD@O6jNYkaB;@sRs5OCRsvG-|<!PbP-ldDi`YRsZ%M
z6o%g4wJiMuL#r>+()o&xCa>*7y6+E2bqKr{_hHQIXOq?s>h{}%_-U;oUnlGkHS;^C
zHSSHGTHkY@b>!)xdxcDuI(N#3*xS!0)URF<$6r=HU$Ij9uKSSjYg-9Ur-q!F8`e42
zXcO0A?-+!g0%!RQS!6Ii>ATQBZ_>c*s~>;TYJZQU>tDuy{<E^6DByAD;9o+|2K4^!
z{)$hp8vmRWcy9HU+cATKCvMo@Z&X9|tNz2+a@}XI^7R_qHfhSW3PW}A+*@z2*U#TJ
zr~k!`D6s#XL&K)du?uWddHr<7gNe%NV@JO%>*3$xV7osJiF0nuxHU6i{MzbF<Mg#X
zo|z`t+)KOozSr1&t8RR|HL~_$-K>@aZ2u@eI5EHKe6fdJoxb+q;G*B3Rz>Kw)m9%o
zEIZ#gvH!VUV{$SAYYJ{0&3=3K_xCr7RT=&r*5%CE7HX^Rb#LSMo`u7<-nqbg9Q?p;
z*Jsw7v+8a7-fOe@!<@MM=a<`TE9h3%zrN7HBg-#^pSt?zZ9gpBroOx|DRa86zH8j`
zj354fzj46Aqv8E7H%49iS^i;Gsr%zqBj3ASQXF&f>C66ilp9V;yERLc$WX<p<%BQa
zQ|f2;+TnTflTRILqq3IkzgpJi!mskuZGK0#w74F2dQ6LJx+iz$oxU>R)1Fc8e;)s+
zrRVa2Uww9E&3f0=)bzhiab?r?eP{1Iy5_?Zr*nsWo%qqJM}<44ue`1DT~Yc=kYlVe
zZt?=JA+7rK9N9V7f9B9dbN1Y5HND+R@lf6S^}-#$I>)1>U#6`4w#xyues%2JzApP0
z37c-r-0-yi@{7&qrhj|&*v}m&)E(RM#r>ZzeK=r3#H-8ojmMTupY(FcjPxWS@Me!K
z3-=wGCt9uDn$mNGo%^8$TPv4NaJe_)$u9Mf;l)38{N+-BM=$4w4L1UpZr1-4**$S6
zH`8lK+wj3vvEL=^>*ZCAzJGr>Dl@cwe{MkU?iJ@3m*4-jOTov34_ANu;nnHx3x*%t
z`FpYcL;ZInT0Ic1J%3RC>zWm5?vriHTh99AWxkb9>@OeHKDjS_IV|T(t0isnzL>Q<
zG1PnM)X|40E-G0x>SXfy@x40QWSmRhQE=CG-|x?M?u_~L#qJ9go&iNa+UqVoTmEgy
z`NH&(-v)N+>~-#&*7<AveU>IoJT!E`ACu?o`k`Uk`lX%T|1Hh0<;ad_$E&Y8JYVxP
zY`+ln)|cxCUmw5B>z$%sI*;gjr2S^ETZ<3Q|M2T?-kp@a?^~Ze9y>pOHO21rk$pJ4
z#b~+X_f=_^UV0{;%?!JF?bpRQD@NqLee-7FA6@e=POy0vJJ5aVm{`r1QJR_KI<&g(
zk@4E+^0@Cte_Ucqgl^rA#=$jy2h~WQY^CIjZuz38rLeLlWb%i*H~enQlE7|i^fqDH
zENVB4-j)Jb#M`N>9v3vsRijA4vQ~6kOIXYWu&734<{cUI{Syf}n5>iqJYK*jS_-Qc
zSO_EUQ8=e}>IbYDn$y}AMJHM*2L$^Ag1e=FdYp=ro|cwA&2&JXNmfeFHa)$Ye}LR(
zfc5(Ap@Cb~D4w+BPpthD`&kNLanH#Mh#q)Lc0g`~rAE@#NCPbeu=uRr`{-fs!Ui>J
zPgwFqw>;6;QUHtZm#+IeE#KTiLQdp+mWX(X=w>ObTC_l1^Y^qryROx!h3WNbWJBuN
zD7tMF+gb{!r^Bb0Q=hL(U8F{_gk>piw-iTNiZ?8;$A(ADsC}tMM)Juwk@hC?vlPG*
z@b|;9Ejyr(?2!klrBw7Q6(cPLumryK^3aG3SDhBUooF6dWe;z^hi?N6x(z+k3#NM*
zwI~4?ih28D-VGRnIz`U==HBAJ$y_1rNOnv1Z2kdqh-dEcCHYg69Z(P<Kj(z!T!^Ir
zmath3yJpNicZYO6VX5MDRop;J0W1+&gWD`vIx)-vwIwY1oGzamWGR3pQZv1~>^Nkt
z8hH_x0BwM4^AC`t<CFof*S&o~jRHttY!%(Mk|}L2fW`3Dlevdiy?e$9X$Z?~$$7S<
zvlPG*GwH#xt6qIKJD@g%B}KAMk^C(Mu$cS{uD6)@**Z0{CoG#pw@qYp%>}R|4Y;fB
zH(-*gC9)+f^;lJp)s_NSlE>frX7as;V`>ydzGf}bu0<i10$7yf)g4NO4%5gSAuQ{V
zb{%0c7r@fFkKKs<e|7ns^a**<GEuusbhQ+~(&cxX))%H_`Dl?1jBg=tU&uSYVR-PX
z&2L-7FFPQ2QpXWVcZ4hqa{(;f0@r<de3Zvl2c#n{S%h?pP<KlKEN_oY>N>%pBtnaf
z@S2^x{Z8H;7`k`sbg)NV^)F<6NF7OTNuJF=Kz`>zpZT66C;2)cSJH~>lJ2@R*iryX
zAAZCd-z1m$WLc3KV%%cdHvhaKJK60#H{<i=M2Zk{3U8mnyITri>6an(T-5*BZ1NJq
za+cGb<px^{V0kxqYGKiz@$L@jEz*uMNmnMlXDNVXP<8jyyXFOj60t{W(Yxurn}6Pr
zZ|zUsRQHiWqzECK+)TdBKR|x}lJ%kEIkvskNG9ZHx9GObKR_NL_TS09-Q7ctIuLTW
zTewg250HlsQ@(m`yJ3q1(h~A`>+!zLKR_ORWvt8MpHqWvQ34^qZT+^o`3J~f)%DLl
z@6$1hJWW3CvS@vo2sm>AEaQS57Pc10O&|-Au;fd+e2FZ?<^ov03zvTF^VN(K_9&c)
z>k8hng4b9IV42j#_njwh6&uvZo7A#g)GjCU(p&(`<ok{TuBZ>+P$NIm)Q)Z)eVTuO
zoF<@yz3*LhwncG-JejwjOh(gOykXgz(mlj3!Kg-dgvH=yXxsb)<SDHdciolV@?c98
zLE5ldRIL_WECsMk)owE+KKnRFjeH2p3Q@a)G}BxFOZxYNw)NP)+tUfD$xGI9&TF}D
zmf{UdX}3p<mJNMQWCmf$<g7C}e@g)@)2C!Rj!D@11!*dI(FW0N1Cbf#0$64oob*xJ
zrfrWT<VZA2d)z*{`Nyh-b(5n+wB#>WE)14mO&dsrJ$X$JzaEjzKk)pVY>nZaeKX0o
z7|5t>611BHKT82Di)X*nX+UB{580|UdBs`5?yS(+QdqT+c&nJE*DS%Cww&8IXp0s(
z!$RK0+wbCopq3>w|L|(-`209|5&4$WSbLi2T66J+<>(r_;p?o%wnVK+Th{Tab-b;m
z0G6dLhpepfol1$uCM>yvHkUNlTmZ}JtFLM<eKK#N18PTJv_{gcAyd&@0LxmZiOD@~
zeKx!$YDxO!Hd5V2YD)nu>&`#+_HjsDMWz8^2~mZ(H2(m3gZ>IXe9plDEou*+TgKa$
z@mgTWerH2r?NujxVh54u<F)bL%|AfiJf_Xv?O(^=B!&rT<Q}oj9%2KTi#IH5d;Q?_
zTV4av2INzB394O!+EM^Z&h-mFMZdWHBk2wDlB-C2mFRkN0W4ci%H3{kPkgOKQP9u=
z-oAkM1%|EB8&ZyBCygR~L0WPVt1c3yV=msXJXv||yfSYXkv)WEz38@{Oha=4EZdIc
z?mj<tpTPmSkd_Q@Gd#Na2guv6$30to|H~m{@Co@6(q1ABH5b5A;Qo2u!KD+4Y2!~m
zx14t?=i6HfU@7Fsx1G^zOItMxCACDjMff!T0C~s99+}h1V<(ZdPim+T>??>pXfA+d
zXP%Z{yz0qt2NXzZn2B^ViAFLPz_P2yIOB%_11rhmCM-Iw&b9dm$R#JPd|L3uv2?RB
zuuO7ZMvMV-0W788&KPsPF8vgdrlh+{MC%e!Z7G1|U_c8o=%Xo<WvdXfr0lHhe4Br)
zTBt<xayop?W$U8nZFV~#XVQxI+PoLt`~%NdmAN=>AN<83HS#5|SSe~(5(#83fTjB5
zlHYx+`gYSIZy5Me-oBJ~0EQFBX?0&`_U%u)i`4N5+dsnYmf{V|2O)z-JUaUunS!Jh
zkCEyzStsV=4a<)AI$ad5x)SwCYI%t5ACg+k1+bhPx~lB9M^>l<3MRGO<b<1Ckfi{Y
zQyca_pL26zk`~!RV~cqEB5G}1a>#1bdV4M)A57{9(gwLU|G@K?O8=bqO;TzEQNM&-
z#_7s9qJEo;H!Lr=kNJ7=z<M%6Nkh+wZO(|%mI7Gnif0CMH-|U2L>&psEJ-y>vbPk#
za(SK2D+j}&Nn~6IOO@EBiqv8*faPl1{&D?2_3x)evG6TBc>5i6#a;X1Ex+CTkOX20
zlE%im#kOt!f#+|2+G}_H-M~MI0Z(3W9&67NL1`}Du*@De<eSm^tB5KkEGw|%3aqgd
zz;avTx%#UsAAU~kYr--e3)69kr2v*Y_1e%03pF2Tkry;{H*de2&X9Wx=M^2POYKZV
z16fd81=Us}fy@Qe(O_d*8CI(yy&6nj^HdU^N<o(54a<v7#SV$j<Fv>HUQ@!`mr%`c
zKfY%{?%${5iA_uDSc2OvAx@3CfI1%f@A+e3-Pq;CFd$tPs10;&{sHontMciPSzX>I
z{XxhZk>f_Bu@t~!HFnsi<D$9}ooMl4TdFLcEBI!zazKnfASQ!omX$`))F^fa>qk}=
zNJ$H%uHX*H$_dGILh9VbVhGF1Magtg>I@#EtlX7Mccso?qRL7Nj!VHw&}B`yR|+0L
zaTbot!bvc4vXX^+W#IvU!LxQyS^R0*@H{ahPc(ubDl0oh_no3Y$RSy&7TZ;efiQPv
z<u9?_Ut-`T%ZiYddeN<3^rIj8v*h-(<OfcTtgMrQ*GUFYtg^CQ3f?Xmz~+;cy;AUA
z$pAL6telbT&PZC?F;^tJE0UIWOrvDiC~0ZOOu=?ju$FejLTtAXYvIFXr2<D(U}M*z
zNxZfCud%Xn7&{)u9ycwkKvrsTR4q<|g#+;gUS)|Y(2hDFl0kllj_?Ah`vR#SjHpx+
z{#HzKUMZFp$Ci#hZ-lgRm($<ndVm}w1Bb#>P)~4SW#t!S_X}zd%0*TdBfG_@J-9!z
z@>0^jlzPzbn~lR~<DT$+5bQuIONe9Z@(z23?aJICE4Kx^+d?-mz*K);>9g<MiSbBQ
z1_lgFfZ&F#REhzWVgh`YO>p|+&*q1m`%G3^MYW2jpWzVYz#ykYr!JJ{)XQlAMb7P<
z{TLMOoWP*&&V48f@Cfh*7h6^WJp#R9t;&kdL+1_ZR91pKg1kYD%gT7a@jCEI2nrh?
z&7g$wofy=6{JRuoa_uv@;B<@aEE9t{B9n_@u_=@5%;LV8+#re<bK1q+TNE$hv`biA
z!f8r4H;609N(tA#gbSv)l511RwFP}cx}MWja$yu7<qStTBZMww<rrr;#^U3g;W&${
zIYTvzPjH44EI!E@PO|tf&hQs!1dmi!&U1$IEUx7YwJg5C87{E+B4@bB;!B+25{olY
zLMG}&r*bChnTh&SoP|8HkT;7%vrr_9ld@1(ikG2;WvCNf3d>N>WvD;JE0D(u<jvyH
z6)2L$Nh?rSiVq>TL&y^ZzN}Osw<;F@iDLdlCQ7eEiFK$8*q^d;4JBSfUBGZAq7SvM
zM=l*KL9-_8Q;z~De#8ep;`QLW%F0td@F|NI3JD8^PL!T4w9Xb>z^js#Y(bYTgi)L)
zw9XS;z>SfWJVBQygi*Xq^jIc(Lx4sm-mk}U(Hk6NnYfuAD@1SbKxJa|d8`z@!5)-}
zTkdyR)KO7>Sq!@@Mzc8Kve=2my)TRJQk){SPmzMD=ueR%QluCbn^L6CEbf~k4Wc+#
z(&S2RR8{0k?Q^AIic2I-iR4CAM~T$FL<*+3QfgBvwWX@0QqomQVH6*e496rRRV~LQ
z!*Lc@ONMF|pO6eESbS14oMiDS$#9Cr=Ox2=$w*aItz@WW@de3nfyEak!$lTfk_?wv
zTqhaoSbR$|+>(q`_1%^Xw^@8gGTdSDUCD5l#rGt`Jr*}eh6Wbr;OHEjKviQQjw>Xp
z@r|n7gWdLEPm0U2TRDq=$1%TS6Q!TQiDz&ZaNcF51}E0wE)Y0xQoXHfu?wi)ChSv-
z0|46;ZrtN1ep%`yE4Eg)mMsD+D_Q9wkZ^+Kk4yq1!di~O`J9l?;xbMsV{sKHRI&Ih
zC!A&R4Nkbh;%A)jjKwpNFcUHFTZDu~EG`v=Qjww0mW0_7gO^FdG8V6qgf%SAmxO#4
zmq|hyi;qac5f;}-LJf<rOTu*)KaqqdEKbKlI)2k0JLglT;FGnZ1VK531|C9#K>7(p
zIu5L2@nb&bF&_)kN>HBhG0!Nr|LyPNb7x=Sh<%~=(}zRAl8le&w^j@Xy9)MQ2wOF8
zU(LHieJacD3m7zDpopsGT^9&d$;t=rANYZXD=Q=2NBTkhKvuqS|HhB1`z+2qi}M3@
z?fgf+_1_zcLR$z*q+g^z!~z9nXwuMRs8CSGB#lXi3I%16XtPMP1shOM7K=8ES$tBA
zKPe`FaV#jcVtlQb0LHPP<lw{{x`G6y5R<>1VTHg-5M^aOme*q~5M-qY`xfDFip#NA
zISvJ^vTQU#!(`<SHr>HJVG{sE5T~+iHv!{YML%{s6R6~p?(ePd9S5B(E9IiTT#N%N
zuw8t?&c?{&s{c@5frCN}bMr14d=l{}M2GF7E5#eJ(?<Lj^cDQOpt5Y5Q486;9Rm!q
zB^1Nr@MTgrh}ysw8)<=ge~mI;$5GdDEJVX(<pz$rK{0H+kq6DYZuG$d>|222psB=W
z$EMqOAPgf|5@fVaA?H)b2k*B(*LLyP)hf%T9IOi1lq2J9-jRbhz*ZdB!m<Si%PTN+
zM&V=!>$gbNA+8}EAYv(#XtnEVu>%CM1Bd=?a6>=tQwd5o7n#jPLu^`5Ztxv%@KLaQ
z1!cMrI9&*$I78572*E7wlOgn@c!r>x!O|BAx<&uQ@rwi_<zFQ9TqN|Sc!3nTK#B(U
zOZU!DC3N$*Gq|RPwY|Z60we4-k{&ScF+vZ(9wYJPn>HAsD`ll$e7|J6Fw!_v8rPW$
zm2564o9hbQK%^LFs^&UVoBaZ3y1;d&Huilk?mm}9-M{->ulw8pil?BsDJY4$gHuqi
zDQE!2D^S!56c09RP;F%A(RXX7OR|!KTyu~P#5FMnk!v1{5Am*tcpY>DF?M;^Di*K7
zu4}N4PQ7*5bsfdf05YS@`>W7DA=j4YwHZ<|rlp{m!c5_VEIUtv(!d!SxES~_LAlQv
z?z1=z8PZS;)GH`ckYNgo%Xw2dp9qm$fvobdJTV-Wy`a3746mga@Ph>s05POs2G7NY
zxi|*=Vu1ue3|SPz0Fm!8ZwABnfOkd+<}F|#fMe>RYTomOf0mU<uSg%z9<mah5N(3}
zHCZtv7))uFJtA2d;yc72G>fbZ?K89=MPvGm=?8j7RwnpP@TYSuoeNCof>^9i=fWw@
z;Pe?Rp3nKt=lns@5xbvjx195$xRBEqa^di$va*lU@8iO$-+7A*xy9*W=w#(@PX9L-
z4nrj?1xQzbg6UK%K%oUFjKvWJD3anL)V>G>uvk~bK3{}Fi`eIjP(%^?d?^YoW$C3T
zqV%8i=u#9z>)(rXd)fN;qR_qU^LtUmUiSIDD0(k_KGOSRtB>qH{F;c4>4IUp5Cdx_
z(zMXD{drrv|6@}<N<Xc4?f%V=h-gnmnp6}@2Qm|DGI1!Z5lCMk640Eq09vCdn%VJy
z7z^eAF)u{Zoqt9%4GSbA2wx%vB~vvU&AcF)F0j$eJCf-R8_j$z#l4o8Xy$9F*K3K1
zX3oWNbJ=L-T-<9ejb=u_Jg{JRoT{8i!Q06BHj0EjIhkxQI6p>_CoFppGKn}lZxtdT
z7%h`MFy~!D<YvqE7zw*#=ZiR!YVLaMT#q9`0UO4o4_|jP{VQu(QQQ?BY;`vY-=G{6
zPiJ5r3d&>gAwK939}g3$N%#g;v3Q5*zC+Z}AnO_&v<Ans!Z>Ih#n8_jvZRi{K*iX`
z9{8}6*LM!&!#KNLTqs@R2RXZgTqs@RM>+W@7X)HZR?czqIWCCmsXAm`hdQva4(n^E
z1C4duL>+FT9$=BkWSh6cCZPw7P2>t4a)ll=7Lg;_<w&8_ci$!1?UF*NuYORnJ1B*+
zzPWr%3ZlOGc}YGm1%V+;Vk?qdF9lKG_YAf^gFCRkrF9MN02VcgZQ%~*aS!k%A;v(Y
zyg9)JdN4lvt^L7Y*KGa&u22(FTT3?{ulDDNt<CGp`3TT-klRD7LvuzCG)-3S<EZ;M
z7A7OvQ@~M;6esm=v-XF_)=6ZY4>S(!MNO7U(O4<=g2fAulL)anD~Q(r2n&y}9qi>n
zdJt)dIX8%^yzR(iJBpyw{Vp=yMFU~gB^Qj!S~xqtuqBZj$ngO305g>Y-;rYli;tnU
z$4~&JSEIJoEN<YP8>sC@dVqJn&*GQ7_e=hbQ7J30dGFUO-h~}^VGkPYFTsu_EUv(9
zD{uhNA=`<3u{p<ye(`SPxf?~(FMf(lPtib-E6T~Gh9!fN$9*9yPS#EiG@|Vk>13kF
zS>xQ6A{T2HhEr>;Wl%r4pA#LxegXX=DC#fwXJ`ZD0Sp>yH`Il(DeT6$P@Cc>`6nl8
zGyJ6cDUzZ|@+5|q!pSKdgVQ)U?Vnhe#xYM%;p8bSeJUqUWpO$ur?WVNlQUR6le3%2
zxzN@u=Ij=8F4SHr;p7s|iORA{POjvfSXmZ)l#6C%S@1b7n#!^UPHy0ws6Erb=^D65
zitlr}`z&tc<VKeN0Vh9TaVnBikrP#;X-JobA}OANbW>P770FXs{%J^_#^O0ho`ald
zoI4YRXQFtDvyfvJa;JD6lIOAf^N~EC#Y>UA6gh#9K>Sz~zKr?|B<_nGS0HyvUy0<E
zEdMGbuVQfl@+d(5bn+J=k0KVABF|D3Na=f#=Ux^cLWze^cS^59iB&8<g*u)>Mk)pB
zP(U4uV0BNxH55U0Pd&1(M-Ei?oZvlA@PSmjp5{GIv-lkEd5*=kyk{+oAM;+1Sy$*8
z@AZtuse&O@NT&5?3)b0!1J!kTf_0wYKy}>#Q9d9#fxvCDo&pYt5foR5as^BOO_YCQ
z@k!C{r04=sBoYS|?P^6Ah$1yvU-BK%iN@LQh`KvsB*k|{`7TSpC(8F&+$hS8|HQgR
z_W1{*{D7rD6y=93ej>_GSo}<spRxFbX!k;Nq3xM2*-e*RK<COzt|aG5PSh4Ck>nD|
ziP{2{l3Xb{v9>_)F)5n01%l5@(X1^HTrWjaTi~@Mzm}Z9H6&U=(!G`<DNez<6wKhs
zSe}d-{!}cdvN#jVnb--;WLcSub#rkf#aURF#o~Ebp2zaf$MSp@=U~qq97wIZE!cAl
ziwm)DAr7JRV(eSY;ypNV5AIIs<v6jN#g({YCA-Ca1_zwM5v*+#P=h0=ZB&b`Yq0~^
zMzCMxrn02*QUhVV7`<MM*=kAQl1cht^ae2oG7HHDrRaN1^rsW|q}b!6*puQ~u}7`g
zlTPI8qVILlpH7q~V&D@oh{gIRVmQUmMEx@sPmz45Nd9zUWk|spQYgh6B>e^{oKCbt
zNna?1(}}iE((jYP=|p=jg*=z^bfV0_`WZN!4s;F<&B0-`<8yFC4vwUF3)XGH!7L8l
z!t!sy5nKMrAH4;~Q2s(3QOME@adaU|FUFz8Y<<NzqL}3`#?i(9<c}++8#s`L&9_j@
z>DLhYBH?8rsz`{XZc(uiRZKCgJYp@Ild@q(fC?eDq&Xv-8cgd@*gDh)EGkINCQ&zY
zN;bS2QnHE9*_4k>pYPYPUo5qGelq>^kC>Ywn={gw+i}ixoQ=8Fa;91~=GMT)HLx+a
z2Ci2F8*@uTacOMKEe-WbV`FYvC@PEU;~vgO4t&uv;;A4j3DyY?ken%#%>lbbq6_p5
z$=nd_7PB}<w9XM7AoT{s73oxSsyRFeIp!ol-IQQXyCf40$V7vw8Snr_JwUNwb<4^_
z6!nneK9*c`;tp&_@^+*J{YzY)=EQWU80;j{5OX>@5MW&pf;kBtTKMjZMIqNp(!&1d
zO(NJ8A_jL;4m-S<#9Wnr@6gfD=DKHd{b0Ny#AvN*PHLwQ4UHO_$QZv-V-l(HyC1pl
zM}E)+vQme-)}g+viP7~M>Pt<G2Gp$q^`m|J9Cdq+`hmUwUd)o#tHm2%{2z-1L0pn7
z*QTU@+H;vm$wb{?W<mZx8I`8wf6A8RlN1CCmS}O#DVOtrQTaHuOG)dG%8H3UTO$P4
z2w{+D^l{Y=(-&8R_@<aiuoA|`OpG}wnGQ)FGUl=P5N|xhGgem>Z>(bR4l#I#7{<m-
zjB9W*6{hR3aUI3rU69G#bYuZOYSan`@4J^$Hu1!BYwgsLdS{u)J`;sOmo(YX_RCNh
z2)xhtZSY^R?9t$V#Rx|a8Nb&@`-h5j!+hRzKHm>Ers2Cu-#49!fcKN#>!cpNAg}%_
z!Rpsd!7q-JHCrRN*9aY;6G^;7@T?bt0fR~;Ut>Ny0ZS8f8zGoaOaK9_S3)o!mjDS2
z=^$-XmeUdNgshwt!%m73RDIWozBQshRo}H@SgjaA@eR@U2CL4WiGj~pbv{+{o%&C8
z9h)J=Q+2&j(r;wdb*>bfE5%dwohuo0r9_H%NctVDIxmr8OQd+{0+Mef8B3%@iuX(U
z{j9pLlr)u+3y3+Ae<c|!r9>7dS4s-S^^&1piUq+fD=(yw7p%E36YFQP=E51=@eGcl
zcGnqfID=zZ9DfEIDXzgCYj6~`%WANp2FJ2Ez6Kis!=3|~QB6lz=!`P8H|ar0!WkAa
zqjn?v-N+r*o62&M1s;Q43Q}9sF&6merioAfc7urz8-YYNn2)p2*-|dHD;Gnkm$N`}
zS|E8ttP-+}h}1TpZlMou6r(qaF|eIVLWyGZCNT!$V6u`Sb<L38rqg7Tq~F9&lOjo9
z#7=_)lKucY%(Jk579AE4SR{CAK1u_h3h@TALd*wffB^O_A(+q0K+|V<8+*>K8T!k=
zVt6yU@>A~2ejVvSvL<a?si{<PuytUNUafZp6(}p=>Tp*Oy|NOaj&OxD3bN8s-O&}s
zo9r>Eqg_FklRYN2!4*bNR$|mKu3$mPN~}886;i}y#iTa5!o-l3Bz2N2mFCInWLMY%
zCCLEfa2VCUKXbM}b83*FvND~sozCKu$o(Yprv7_5uP*0ZA)qNMlZAGZ1%Hr0O(`Ji
zwW2GefRJPdQN3PtW!(z(2GJF|T_)S>>W!i+J3-W&L{~aN@<er>=n7kdB*-tSw~DT?
zogtIlY<0fq3d!0{$qlyUqMFXs$D-|HQ4KPc<mHLBPg#5ryC1~f5G90ZK{8RAj!4n;
zfqzaXSy{;2EaVeF&XUY$zSTNDX^G`v63Ltv?EVr=Akt?x?kvBT=yCO5JJz5a$n<ME
zaRsl3@B>K}In2w4c`fbVQ@rOXK7`^&eApx21enB?n@?uZhHV#2+l8KB{b!!I(J;AW
zV&1=W+k#K`U5|5V$CK&)i{Sc;&>m`pParKcpXh=<frDJc>^C3af_H+<BaL5+<+W_4
zOCI*j`=_((wTI1gF(31SB}kGKvFS4I2|G5>yEY_*+kEB=-bi*Wc~t>#OK~BuDrE5v
zUbTb8MZBtr#l^g;n8iDJ)lL@g;#Ip?T*9kLSiGB8?PhT)uPSA68Luj1@g82ahhm5>
zksdq3%SYJZKF@oeXNUVKANG_TZu5~g`r)O5sZ{6*dtOVIb*TGc-`5@gXJdZ#Pcvp8
zvF(`ruXzIsn@kV$(K#4AFt3TuF`t`*xlOX-c_D?j10M#`QHa)QIz30(dbjHxLgiF8
z=akKPgMf!)b*P16K2Aq_pq%em&Kp5MZ5nVQPE|B2&`wq~tu_8sTK}Z^DUeFNG)|Mo
zF}Q)#G;o1z6j^hh#c4>Bh635Bux1L2=OX*L$cI{(%aQ$Z<O7PfDaxn0BL-5b)F^5i
zS^Qekyp{rKup$L(QZR$(V$EC}NP`zySd&FDxUH>J&1VT|mzp9>Mye#bx_9*l(^XbJ
zZufB@^{_^_8y!eJtch+DnOwU?oZBLrYe!sAE_gd<pdtLdT<~6!Ep9n#B`c>nyVIPO
zhVU<Qc9%IVxTdo5h_iddX=!JtBfE5@rJZ{W8IGY&R0UNdLp6)fAjdPv1B{=h)a<Bw
zlmsDUPz6M!o6k+c43(ANMfcxH8nor8maLo=-Oq}C)aJb^x?g3FyQE6)sr0xDEHW~~
z&4(`O9IOz83c(J}8RkqV5qpka-6>L5Tw1%dp>xp1+a-`9ZEGz<)3(#Lr?mc>{vLEK
z4A2Z<&?L<y4?2@mI86%2;3=GD3X7+5nyD;q<TQ<(2krbu&Zm*nQT%|@JYeY$xPS*-
z7^P1|nyJWx4%SrUGZpDr96lArQalZ5rm_6fkk2%he;P7OLy44s9@5N19&{zmL;mwn
z2*vY}W<E=wk9_7M9i^{Cnw7|duDq4Ve<cc`coovDV(F`p&nl#&^ad2wfQ%r`$Sx&{
zy3gY0DC#*fLhyk24JhgbiwlIP0>KFSS5^vzs6rN(2~lN&5$2Mt>=B~&u((3hREQoh
ziDl(CQS%#%?~0ncq6d{E_e9M-7C#U*4_N$A)I4PIWUQHtJ*cFajD03!9mT0wlgiRl
zaX=~#qx5-LGY@-ENi+}p%)>eshtI>Y6wk++`7Hl@>@%O`pN|dmaU$ivf}^frBbWfP
zaur8iW$|4cbr&09<6Bnl;i!8QL(qq4)-6chLIY)_qmkzx6hfs&74oV=p@3DE<Iz-h
zK0~Hws3#0`-d{`iuMFEeisTaXjp*B-uFW4~eteIv##AmQm1FRJF=D^ipN{5jG3K`T
z9v!{=V$6Mt;Y>Esa^|zxRKjc%OxuKkFs-&UylmUeCiOGo2`Yl(MuqwgQP?56K^z5o
zkw`4_!EGuEA7bGlNe{3b(gky!Xm#`PZD0hIKx$|@yG?s%GVd{&j{}bsVsE6|N&0av
z+wD7$$BuvQA9@{PyWM=)n^t@eneL&Uu>L@;t5lXF-?T@L^6ifDZ*o#eUM}x{gO30+
zmZ)4_dx!U>c)AcfUFb@2h7gh=#8I3rgl7vTu*91V;DqN1CNR~A$tiRz5Z;9mkd-o_
zTbb}KEM#!+h*i^c{+xbKvQKgl^~G|zxLi&FtETYQ_l25_#_a!4DmLh_rl=KfU&H3`
zn@`Nc02F_E@ko;D@Fh=jf}@3PYY2^z7?@2D4MjbLp0*UdC%ng?;lgkRjT6Q(XbvaL
z;TXJ@vsueA_2zRn`7AEuY|2<%#o1J`_$+60mc=(Xn;R^C#@Re$@l0eh6EW{wglrbE
zcq0-vBBp&6yimb2xKy+$6`AMvh{7I`p;w7Q6^qY^!WkCVi9#KVUyH(P7SEP!W=l-{
z%Osm+EM6notYLA!WRuV0GRdZl#YZHYBP_0wY-(71U9!2(;wO^L6Begqn{>>yX9X5k
z;5Xk3>Y0cU^Q8)OeXK=cYf&G_Pl8jSB(l_W!vbZC>=fCV22ElTVoXe6Bq7#B10yTB
zb}QKg*&<F~#3sld;PeOB1ld=d{uND#Es1H{>Pq_KS;T+HN7nhs0Tcwu!b9>d<V5j)
zWWS%4bO({+K^l{jm7~1%QQm<Hm}9(rjCZ2=EH9sB@f}{i!{WQV{aw~IyvIA<qw7&t
z9`Mc&_;$3ujo5l4+4+5QqH7bDH(@7=w_*EjtaDz99ZRu0(BUKm38k5DtpF9UyYlqF
zT*HH_{~;j-`i)E{^DP?CR+$`;NGcV3!I_CYhtE&^EI#`(iE?b`h3&jQh@ySIKZSHG
z=(_G-{w<i`L=!gMz5&6HgD0x1s_Gx^{hu9c@VH3Rn=bF56Fi0YNa5qaNk05^_o=7x
zw*&qajf4Z8#Je$H>p?#~LmH4F4WyyfY)PNZhE|IueK8wat&sE;G_-mw#^;CL@3=1b
z4-GTiR3f%e%r}&P?rAzb>ys_`QDvPiw9gg-D9#gn@&rGsyz_+ic|rhS$ciMpzL}C&
zrW6W+ZIWIuwO=kp!eSs=A3GM{cCbiPmU~d3)4@6<qiMeVgpTKAUYN|=fq?{uR`VNG
znUXX6M%8l3nZ9u|w%<&brOI+=3cL{<F)}~R7pA});ieRl5MaI`g?3iC5LGV3f~7)^
z2nbR8C_X#&4`=CX?U{eYD$hP#+Iz&JjGzDKIbZ8Hv(J&f{v-c2nc>tw*&T1XxdlE&
zR#tKCR<Sa?n9~=tGQ5J*S5O&#De_q2lN@L6KQwhA7(#~0eA^6s8w4B3j4@v|0|P57
zb0p&&X$VXVxO0XWl;#U(DC2A?Xtop%j7<*)&7l~=8{|Rr{WJ8zS(584a_rc0n~h8o
z&W*F_8{vK$;(eI!rJ*xuw_w^W41^g3x|R%E(<L?VX-(Tp1D2t|)R%sVqF$m{SR}IY
z3Pru57;=<|sac7gE3pp*<yDqjZJ^<>`bcM*ue1RIc#VW$zQ_g$aDxqrRhw_FflmSR
zopitXz8WBq#6CedAlShfR&aPp)#l4<pz4O0_#HplwH`|x!R@Hcc2?o<MjpFSJ8B?Q
z^B&c#g0A5`YFNA(du*n%rAgL!7GNLv+6K$@IOH45*Wy4^;aVIrz~;+vplPzQ8GCNV
z(NF`)jKHQkJP_Q##-%&=1Z{2W^&eU%a3_dX)qHOcol4=!;R>^XmK;ISoXnSek;7JD
zLnA<s%E~C;QU2hj%F39qF%cAf=ldOV^eTf3%wUgRE#Q0?ut%#_aP3yGQfCLJ-@!_q
z5-zrci>GnV63$q{B~rYf)9<HJsp$$3O(o|7qM!*IE4f4#Cs%R`#pgJ~IWCsQNpEu@
zx7njxFFE~7YN<Bmy>zTYQQ&(wVM85?WpR8RGE#gEb-ae6AVku1<%QuIie+*9HDm<*
zaG#COSK51X{w1{^PwP_bb$HvO|6>U=urA4pHeUn<lkLfiEjgO{(X)sg>Zj?~frfSt
zh?)ap2MFH0Dm_=%H{woz(jWCgXuV*7oicKR)aM^H-8>H!TER-S8aMWnaY~cT*%=H^
pE30nwf4jW-U!?4xu%zRZ%eNB6uH8C!>15UE4autMZYs;4{|DV|nWO*!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_gemm_xdlops_skip_b_lds.hpp.BC4BEC2371CE0CF7.idx b/.cache/clangd/index/blockwise_gemm_xdlops_skip_b_lds.hpp.BC4BEC2371CE0CF7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7e053e3f298d8be4a892c11d7e3023ca206335e4
GIT binary patch
literal 8568
zcmcgxc~q3gvZtC+acq_#!zj$KDBCnJAj=3012Q1u5^=d+T+ka6#RXABj2iVO8gQ2=
zDv@A#aYY3cRMZ65fW}0<iHe9Z!59&VM&mPXH)?#%K-v9H&hgK9uzP;<tFOQA>Z<DM
znjw7?6Ei)ntm4P@ot-g0d#u9B%IXdNe@1rJq*cVqYKo1O)$9$GYepsC2zJPQG4zk)
z-M>%$rh(<uI%f<UzovJ&<KdM*jO(|{>)mlxfur}ooUiq5yx{EYIeYr;8F{ASJ%-5G
zIS&(8_&&9IFZk<(4Grb7PoHFGK4|{v;qG1oXK0?ER$RKVblu`-S(;HV$}WAiz_oqx
zk9Qr1jemdY-M&Ayzvq?JJ?O#a>;Jyrf5wLU%B_#e{;WK|x%kgr=dBM;nD<*o_N@P2
z+*y3}YRT$R%Io3INi)5lWqjdpe|1wzZEcmutqu;!9iBb6TN`mNX>M)QfA$?8*r7w&
zNL|$0k>(fc7Mb54oxFWPw&SyfxtVdNj-@zG3Q?WRTj7~`=0x`c4F$RlV~ekO7y3DE
znOgi_T4UPT(f69%o@B+kH%<6q@k-xohi)IXJ~p;tQ;f%RQg8R2-D&;t_NmroaLH=t
zi2;+G_k`}O@OjvO@7z;CJGT9v|JI%4KW`*|^enXYV$-krs*y?S<C@Kl=H|G@xaRnX
zocpxf%ZYz%oic33-ZA>Y@8+ajj#_fe;pkS^4Zil{?&W{przxh;erfKQ4D&+YKOd(A
zF1pEjn?L#aY0^6zcR3D!KWfddGks$V{%!C#%)D8&;^ge~hTQu6?NN2%thTarM{VWY
zBTZZA+Jox`4PP7MwXNXu?^hLdx)(n=J7ZK%{bAk25<?&77n7ccjV}B6(-9flHm}w;
zJS;!^z4w{S#)$v)YiuradUN~Rt|bnm*7k}#_wF*Cuk##Fr_tkVbp7%Qt7mW6)j9j;
z+ViK^Em&V0UHDUS{*|)EXOo6G{g%?l$#2K5tbdhh&m43)({#Nd+unZIGd8|mfJgt@
zKE<E?VZ8m3DPBAD=BE&-OKHlT-}uIXe>e~PF}veWpHAKterA2cx{axs!wdhs9TQb~
zIyUXz?~iB9$$0w+H&W~48tb9fRbGz7!Ox0CxJGze8Y^zDE&tlP`i`(p8f$D%+WT6u
zt{u8YTN<m51UL%}Y>Ix?<xl;}Nkvrrd8TE4mIj}<`^ux^|2~^FM@1Bz_R@NJS{kHP
znq`VxQ=SH?hzF-Vv>x7;2I($o>ppC^yJDY;=sE4{;_Gc`koG=%e)Hk3F(*{Sk<%I%
zjmFX--SbSZ_a5hlf9^muoW282cR<(Hp-AGvfO#XYO|DT9PcHHZTpmG>whl!G79Ul7
zHvWULj-(?ONp?++vNT8!Ni8~HJt+RFBcYrw27NJ%XzNg9bk%R=)7;O`>_n`&$UdUn
zM^tScij0|bEak%Ri4D%A0~eXqW!9UP2I;Z$(gU7Y&DyLY?KthN_4c$hNKf;K-P#?d
zsZ_*?({3(q8cTz8X5p>4=7L1M9r5OLm`#|<(jYzko3H;!-}?1r74hQqPU5nY__cK?
zk`+HBX+3?J<w)$gNDk<8U}Rf|B6G@eQZFTLe5oScxyUx+vW<_@+M&q&{qEbMZtUpb
zK-{^=Q*e3;>b4F=J{+?m@@>0*a~+8b7dZ!tb9_HqI~2+7l&$<^LE%(K(v^!`1N}94
zv#mps#kr>On|5`oRS_*03DgF5w=}Qlt)@i@vkx71B%L^23i?tQ)z+cN@~_LjIM>nU
zGY4YJMGSt12up+X&S&mNCk;$&bR<enr`n|YS{kI!_NsoFA98Y#ia2w+o3@*er9t}K
zsI!V4$(8S^NM}xKU9=iY^NOC7Z<{-Je!YtLaJrnhlygnCcCSRXx7+R*ILXU_xN(vD
z;B+5YTZbaG{)bom^yShI>_|5*a)dY>A#QCQiqvN&T+m-!{Hr6e;UWpH2~m~?=^y7#
zUl)`=Z?%f3Ijz&`d@T*qzpNd4IqxggbQO7n)2=SA8cT!pqkDyV!xK%Iige+0fHt7J
zr9t|InpD)yyYaOnvE%e=(65Hnw(gb4m8%8X+B_baT)D_$P#lJ!whl$Cvc{*)*ws};
zF@@3+y)bG~mgtt~hwG$lu-jli^k>S(xQ+3{ox~@I(h}dwi~Wy{|6=t4YH>p(+gF8q
zMEBl(#G0}b5PAXzzG#~wWfvj*A`C+FrR+M`Uk8S^ib0Z(-Vz~kJjzN*OeyJ$9;m$1
zIP}Zmb7O(BxPZ7`7@BQ^mwYmLU5~3il)1XQdZXeg)3|HAF$Pkmb=P{MX;J3p?&ghg
zfwEM;R6Q=0qp;K{3H3@%l+cjW;R5A>IuEqy{*>hb{|gql2waQ66E~BxBG4B>n7|vs
zxDnzoMpCv3jGJV<8H}4{yakL~WV{uOTV=crjN4?q2aJ0lZgHDKQ&tYfav4{Eu|mds
z!MIn(`@pzQ#{0pzU&f8ls}T}$8!2mqL5=WFft$d+3A|+-(gYDQPH2LD0<R+4Rirxx
zLdx=qHebfYB)XWG1z$qqOGsagtu6M_vy_<7UMMRgJ<DXgi}c(j;|kKbf^<FC7Evgx
zAU+i&0I_Z5_G5nSGLLkm%%Ct(%Yn}uZuE*-63PKZ4oG-4C|1k37!<`aE(JxYjH^IV
zCF8T8I4k38ptvS5W(?d{TG9pKD~-^n5&EMp?Yq>PgD*CugMzIhBUX`-s5DF|td*8j
zLOgHkV@gJsqHKQnd=s9!m}WRAE$M}LMAuVi#xO`3i(v8S$&`h-h3N2PrL14Ket{y+
z4C^*5P{f&cwC}u%Gn6fY;6-2*F0=-M*ML#D%yzKf4lcrF4uJgua1kbc8thMli-<f|
z!Tu_^h{#hw3<V^vaZz%(-$46E+$D)q5phTDw3t9-1xY{?=yJKj=6T7%5TalM5(f0f
zB`8=*LP~F3LburcpFfEBPWf6P3`CukmLyClaX{Dr1CC4CP~V|C9G9|`uoQzpV|~Z!
zMC{25%nA}{DTFSCaG|z5Fy%q4jN|ekUf==<Er4*L)*?_BL01{a6+yg=dlx|;fp>#(
zH<*Olk3o-P5GvIE9fW=d;X>^tq-P0<6rNf_j3vY*<Jb}shd8JqA}Qlu!%`mUpM&N(
z=y88~G!^f5ee2}34hp6V*BQ|G6pR^IPc%LSs|I5=MB{7}tOkrVGOhz-9Yo{o6znn>
zFUxo(30z6^Xe5{@gR(6>#?6Z`e{A0Oho}ADI-AJQMMD~@OCNFFTmtna2*G7Uy{IXe
z5TiW7iG8HYJ`#b6wQ<^#30r?zGM!Ks<QAkCUiln?o<l6^Hrmcq^RGm&IUOil3HB=?
zL|DrwVE+k(h~3`^bRz@_C)^EmHv|c5C?Pf_#2<IFB`({P5`Phw%ZYzE8Hi368|6_Q
zJoEeVzwCCx5Z5j99@`}FiO0po4HnVq9<X~5h5OQbfBKq_KZ>p7dB<|%w4Au(#!<F{
zIIWOz0r4y#0m$R^MnTzK;D2#AqO$FhgBwlR0^+`a7*JJ|RT6V08G#z^b8&;QXk_mh
z0hF1j$yrR)6ksw46hmVqE|$hhXn_5It|APi*r#+Aj-N);oQ12W>C++vnm{K=+(bH2
zLg_SJM*4K=xhZst#AVP731!ku31!hN31!=7cNK5X1DXfU!cdBU7J;*DD8U;cN;Z_>
z-4G=Vr5b28mw%NfR)fA8A_T4heT|Gy0zE0qp8|SH#+QIz0%!CH%IZL02N42a2K{9j
z*8{DW<*xv}BI6rCZ-6r%sx1jlcq7CL+yqWd;3n`bptofCW}wY7UP#;*5}h!kT;iT9
z<5eVn6?s$e`6NDH#)YJ3p?rFkkbn|mkR2_clo*7gRS=sB;)sqmu*=z+shtc@xuf~|
z`G(-Rj)unLTg&|&jUIQ6f8Ba-M`HN1^*x%ZJ`DRyZjl;Y)Ol08=Ba;cJVYPl2~lh9
z0H;a$eNJ(N*yB|X(-|QD@799F1AaGZsT_!>8zH3;Mv1-sf<%5nO!zv=sz_v&z;B;9
zV4iw2$hsxS1_y=7K{oC=^u{es{~t&1d)GhM!MC6h92y}E$H&VE&&ygW6k;07A#}NX
zvag5G_2T5t^r+Y!s(9lAKHgnW-vu3pFq{BaOG`;ZjAuwN495FA8Va9uKJm&YAvjHC
zTZIE3Ll5V+&|2O=0!^<SWiNof02dsAvK-=`Ljna}MZ8wYZ)mN4;6RieAm#&P5b777
z<8N#&g`glhGbdUrA4s4k`HEZ19!O9g&BGqSPz?82|4=F|bq^tCiZ%5X(K_9no+KjO
z{DAo(XhD>10rM6}LJy&AFPQg2l89?H5K{vQavU3611SPuhM3EcfJ>n~0U3N5QV`=f
ze2J~)7o6hqe%3j6KYAHqOPPntLxZ_IW!+TWG+0)l%uD5^!FvK_-YRbm8aZV;l}>|c
z6VKUIOoM3?WdW)H4HhOS3sePaP*s#g*+r>@Zl>8yQwiNnx0^1ZmBei&@y0Wjvh~Dm
zy})RCd~GdtBV5^rPyK(M_vOr<e~BqOQZB?O52gq5tRzjH7ARhtu1=RwHK?m0P$;+t
z)HO1$19cq)qJmo#uD&eeo8WL0e9(mOEqszK)hnE2+3c!zgEBTHI#U*-j&T?2h*if*
zXo7lzJB~ovMD;`oO;JygP=-1~LMK6e65PcaPJ+)#&<lJD)Td<r6a<{&)r41vXFaIv
z!CkDe9(?LSFXQlfFbRAG)K_HrE8ufQmcIhVD-bW_Z-M$2xTB6Jy9K&i&_m#6P&dnb
zGx#)vUhoG=<UtaL4o=x25_w4A-9K!o{w(ZxI?v*w4blG;${im)K2@l5a`fa^XwUr@
zp8ahXPUgWcND-vPkXL!<y-M}62RmPTA3K!x)7b>&H`g8b<=)ny7MJ@7MTDI|35o;>
zB`cC8G)gf_LbGgVNu%U|Z4O9yHQ26}aWU8y%eWM5OJ!UIwpB7d3$|xvd<|@`$@mT^
z?tnC#`=Gcl<ENl_DsWY_=NpYn9<Sty?@O?G362<TDO*D565=fIGUBj|NKrkHIOP#H
z<PV>|G-8+W_nWU(^;Z|ppE+~;;qkxtHbyKSv0JLZ=w>`f`Rl?kNMX&cozhYU7DC>=
zyv-u6ErZZy@-g`_gnlfJNy^Sc=y`D=#f*p_e1)V-A@RX*rEDuD<7>{fm6CZRXssh7
z@oluvydlY&g4eG6!q7(@&shDE%MJ7as2@N#91e35%gw{)E;!?{-6Cl70Gv^<<>o;b
zhz(G-6{&F=cxB*eLraAkzit_(`vWkGVU7^@BP0gjTmNvwkcsP;E&j`vpw`<fE!Au>
z-hjvf@%S=cn21b?$Fz~MWyEb6@k5)Stc3I{A^*VCfU<+6|3NYg9shd7xA9NPy1;8S
zhgDoBrKM~x=Hb=L%Ph{eB;O>RnAL!x1BMBdGBjnFnB!Z%Z%M_I%)m^kc#;{K87?AB
zR%n)lmO$td*{;?>=sMZ1jzW*4a^d4!2>n*r8D$HJejy1Kx>-m<7LqU-8y1oXfpZDV
zB>^(l=gQCLl8{{a`CMYi70=&~jy>>i`_5V1;1&?q1;SAuZ20_GP`RhqYn6^?e<!8o
z#|WWxQ%{pg1mrYxnw%JQpRW^%(Pl7jmJ_22FjvTlQ8mO=%ZX7n46c?FqdJJGlM|yl
z7+fb3Bg&c}vPoRP9xv*9$m^TEkN(yR$CGD!%G9?rR&pJi)Fyw_%YQGuA9nsh#Zmrs
zRJW)I{2K4VJb<+RjDuRmbDI;bKi(k0f3=x1w&etxk@e0DBku{D-DTX3rEe9#qX_g5
V&{_H8@nmHcAU^8E$3XG%{{XnFIjsNy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_softmax.hpp.D0089700D1C9998B.idx b/.cache/clangd/index/blockwise_softmax.hpp.D0089700D1C9998B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f4dfa8928920852b8741a6566836e70939f87fdd
GIT binary patch
literal 2884
zcmZ8i2~bma8h^ZSgyaFjBf&rd<bZG(0mBhR5h+K2K}DmHu#s{k5G+IzT!URvL5mJ;
zm1=2Sv0bQ&ZLNw2c457+<H=eV#95&(VzG*qqPqyxTHSwM=gZ5GnSArU@Atm@c+1jL
zQ|GfW?5(17OPQ`R-x0$wPxLoenhYZj7?$RYVV0)Tsx0F`B+)nbd{Ui%tzoo1bE)@$
zhG2G#?;4+dD^|d63|bUWcB19Y#*g-=?$}*uZuoLi`MjYeC3w6z-|5cyncKSG!uZQt
z*9>0NdxaV=B*m=_f9~N@-%+BUQ0F9Uu6y*&#iF{|%4hvOVKukYqB|4fxXqf)L+sbz
z4rFI%A2VDTXd66B-27r=qS}9I()Y%emKDQCBBSmpEI}zd9wli-Kis|gtHw~j<XQ3?
z<tpilmb8+shfX>3F4dP2-w2fpn};W!jBW^4ZS(uZ{K<b>$DfSq8#C0mhVpASJjp%u
zp>NOb!7Iau*=Kuqw)JM89$P;=k$ij4XT9dl{^@fsriCq@5<4u{l*r8SV^^y?xW{t}
zI2qlyN<Tfr2&@c9Xg>O{;O3&&A9o+S_M?|ujkNyWdY<=OuPL9t{<zs+21>6h{cDei
z3+Ec|-qim-hszt%D4U`#``#0-;j8X{$Ckh3$XuLEKW{BO_{;l$UaB0HX|{dj$|-J7
zTG`(Hg{j|EcB{rm({|eRh;r|HF%N(2JpU}Suv@phUuEjD2vkPJxlbytjh{2Vy8Ehn
zlk{^*phdWC@WgoV^XQJx|583%)UjTpeLD4nc*;^5p}w&;uYbbzLF9jZb~#jemOe;Y
zfBZL(nkJPTGgp-r^rDVCyme=T)8O9aGjPHSBh><RAVn}HnhNw|^G0J1G}hH}a3+ES
zeFOOv0odotwSSKN{CzDOcSmp#E69%`0Egt%3<X~AC~?M}JTcPW#ovb_m@5Ofx+WcF
zO>7FkyJ^?o*|-PN4`GD}C<6L%U5t>NB6krRpMl_5R;+*`052=)3m)E4>dMBQ5uC-!
z;!_0RcMo2zyL=$$A8gza!KtiNK1BdN{P_2_&Zqr-9NZPbLSG@DA^>05^J%p4Qj3Ix
zvk@#3h$IvN_#%sO%}lPk?~eN+xXPo7PZ5B7a{EP3$6Q~!<Gu*Ccv$!p0l2?zpZnzB
znw!|T3xZWFl{ZBIzP~ePY;T)e69;D@SS%1rC<5>+x1YBCbl}PNY}^gO$*g1%MF7T3
zI=%Th+GOah1B>b@bOU<IU{SpUF+qybMCp(=K?ci%W1zook=y@T`f=0%gCGN?fikc_
zkV2_Y28AR@y}w>E5h;DV@ltEC+v&sSL!8#0bs$K!qnh|*oG=V5I<crpp`Fj-&yvA7
z5ab;G96N3vf1V6xfglt4i8c=OQHs{_I{sPsd0j_CduV#jN%^U|AmrQ_V@!cUgnv~%
zvg7Q^_mQ}GVR15u!7a|BrVnfqq?W1mgO?(R-B~x+eEk~f8lnwxAcnAzu5~_X-3n%f
zO;-_AfaQDSt#fcB4T}v+KsVmFz4+kB_6lTDB~eAgP|SB<@LgW2qzZ|Hq(K=_v9#W%
z_<fobGmcur4E6#ss(Pk=1ymhY0)=DU2Ra-xVfF{Vs$T8MnPmkH$WnwPA|BEw$OV!G
z@sLo#jrOGShY<{GaIva+5L{95u?s6-92!QtA)z7Bv~Fx@Y&7V;)72WfZ~u$jY57+4
zojdd4ixXNWf|N;QQd&16IU<eb<Py0ZmZ8kZgiR*MTxo6?q+!ab;5}VC-&(%hm|R<v
z*X<FUhz*}?Mq>LWgAG&}>nWfslSky)DpQ}q*gCrNjRIvS;)wY4GKe{1K5S(}&*NuX
zI^R7yExX3W?R(Pi^sSrbw^85qD1j$pAd_zsZ~H><1X)Q`+PM9m0pS@2Z4`PbH6V2%
z#7dA^0a**7r!Cd>JG<TEQl|NZ^+SnSj~JbpI8B@mF@LMS-<{05K7@i(yQzI(8$0I<
zJ%_hXeBexw8m7jJ4sJDbwH;T)EVAQtgpNzUS<WoC!%R$*jf3jBu&B#P2c}9_EroeP
zj=35EZEiJpwH;T)Ewbavx#f18iEFZPaE%?&z7s|sZLUIIkw|l9h1oVHEb3lEP+ijM
zsJK%Z2dC8*&N!ObTq4)bu2!zKvzw>Lv)M)GVW~k2K3UY+2QQE`y1dq#4`PC}I9lAn
z7eQ7zR=Gn)gE4=4u;t`K97PZ>iI>8Lp<PX0kVbQhB#Z2@Y-M&PObx1t)M(oom<Kd0
z*3SwZfzDSagrG-X`j?LMtQ4bRQTeFE5cU7Ey%LU&k9;s~z+cEYYVM57`=^P2HMeHH
z?)&#7YDpL~%#U90NM@uRr(i1l=rj_T2{sPCa;R9=PaGYpGEupT-jF4UOIAUy=*#y`
zzIC?Erb6>_>w~wI{PD95J%;f>hG8@5M=t%yu|48seZia2sEo}DXRZ%#rZ)%sKdi6n
AxBvhE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/blockwise_welford.hpp.93790E23276AA4A9.idx b/.cache/clangd/index/blockwise_welford.hpp.93790E23276AA4A9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7a53b2246938b6167305bef4319b4f1986ba58b
GIT binary patch
literal 3000
zcmb_ec~FyA5dZQeNWgrALkvm41c>2EgkYkv2^a!I!=V8Oi-JWXC<PHEDu@y<iUOnP
z;K3l!3MxuD#S;V+Q50xF1f}BzLX~3cAmW7-`$*>z#y^;0hHoaD{J!0{Z}<0h_r+f>
zla)?K5cg1dTvXVWH3S4fjKD{=CFaMsI0W$oFXFO_{Jo=1c!tm3jif9Xlqgc#bx#BY
z*y$v@(=G28m=`6xR7T#83{tZa3%**EP^4%<XG>$ac;w}mmjSD?#o2<*ukLnA`uoNE
z)AQfSsPgA=iS(%3Lq8<s8`lVt4><~Ti=D?nY-mM?ZgtNUo!HOl#x9r4L#{jIBcCfO
zh7M-l3Q19Y%J1+#?4G@#u&}*9`GvGDamc7jYTa95?-gX)>+sjZ_!Uo2U;g)Yb9CCz
zjS=$dygqN|_6VPzU)y5jZvA^o9LhFkDo0orx4xsOG8uSTX;hG-bTM8a$og${WWuQi
zuV2vN!Dl_yM<qqEn?Al(Py3@BwU(FKO`EOfbhNEgp}=0%dnAk`Ag%nv$!iS4Su<FP
z;ts$1z?K<k-7GBl`craMQYw{{m=vwZzrGI}y2oYJvQ6rTJJ!>>{RzoN<<n38zCFCk
zVz*+D>hvZrsqyRCa{54iP>GU-&j~eNR+)D<=lRgup^tNIci36|q*`FgdwT1pRbRIt
zqaYxYb|)}eSD7G5c8kg=4;nDX`c^6S=!t~%O=prz?$KL)sHf6d?1vl*t<Z5%=dsiV
z8v88%Nqyn2N1~7MD<2-I&uzO_*I>D1+uui{^OuLnojuP4JSjV5MXrii+m+33Cew@5
zNZ$E0S46cvO8FR6fRDT1?yA+iy~gM?eFU>1*yw2lLIgVS1he<1#0e<tO;I9HGiVHs
zMnJVnv9amYGj*9Mssq&KG;>>xfa<NRH$<<UF}<cJ0jMo#7PcAz)w|PZ=c1M;u#8b%
zpmwG^&(a8}F3t79$<m82OhfT{2u9JRP&5M3Lvf&#aTu;U{Dj_D#yK5RlnmBe(X6-{
z0oNljVe3>$AQ8;zKxV`--eUp@W129jFeiYx%(+_RB6g9&kVu%T*mVq%&Dx|+51vKh
z37DOYoh{^yfN?2YGcEEB^j-1l%B8y3%o}R6M`gs}{435n1dO_nDuKBWFpe3As$Ifm
zaGB$f4R^eA9@iP_tQKCpHXv-#pCWhM&|#4Unox^^=*+3Jw9(nmwzq+M@?Pc?(^ub)
z21fQP?StTJ0v0|e+z}EbnA7znkmQtsBVpD=YX&?L;cO+ch8(gLFy_E^7;{O$<Wjji
z<Om>{d=W%K=?Q81kR^G-?ZgHybs;j=cCa18k>L*y&%>X7dglkC2Y4+Kh&*9%iymZ&
zPOkJ+;lWW8na0qXge~!0vJ~!1!jxjA3@-7)3HK&lD%}M<2_1xd_!=gU>TB8q5=O!T
z4FgS~QQ(&h-pk5;{7IAv42htx=_{>Y2~EO*eo2_Gl`nr>yi}zF)CG%An&j&AZk&*{
z4e3f7k1nrRl!V!H?1eBHfZXhTA=*%zZ+9~Leb6M147NXDWwiIT(KBgGTbLRg0(gB*
z^N%y)Go{c7uq_a}!=&IhPG2(+;F44?eBwW8UxlELB}54sT9GjMOu2=Y9$!OWiWVt}
z$}!7>y%#nwb9<dOArD7dd;9lP?kJuF569QlzH+5n;elTnQ8rT>SO_VE22obF!;77@
zTjx)rI&!4X#k2gpF&yX=j(iSGDp#GF)?_f>a}p6k3IU^|gS0fqp(d>2$mL6~<1bNC
z1`4KVtXy5!+2o<$&;@v46W9j-Um*xW!kh`tW2nM_*m5VZ$M>70D#eyYJqJD<1dX?t
zxY!u7cKvSqiXzdwKc^_dnwai4!Nc(i&@LP@#|YLwnB)drLx^f^)V%o@z0OR6-}=`3
zjg5VSHzskojWOr_zH~>azTu%+%--q0)cp7qCEW6TgxRNU*$glO`9!|4_DBRo!8kOJ
zIByJf3}*R<XQb?(L@r&zX7`VUX1FN|8r~vcu8iL5uz106Cy5xaG6u4))i{<~vZiRy
z4@vT?tcDBkCmf`o=5D^VxcH7_!t(jtzKef<psafIGLngcVZ1S68`g}{uZmf#vL5|G
Z8WsOPW6c5ntT5F9O}aO|+fb4L@*fnR2OR(a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/broadcast_add_2d_amn_bn.cpp.7B44FEB5DE986168.idx b/.cache/clangd/index/broadcast_add_2d_amn_bn.cpp.7B44FEB5DE986168.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c07668af67e09d7c8090d681f503b763f176add4
GIT binary patch
literal 3752
zcmaKv30P8D8^=Kr5w41Xh-heTxum(2E4ig8DsEhq05v1T4Kp(%8;!KR%*?SSE7xgk
zthAc?%51D`vVBq0Cmm<nCmTz%X)#lqtKV}j^gQ45@#yjV-v2r8J<I#P7h^eWc8wke
zvv@g2A(!SR<1iSE0rI;lHz)0}CI<5z34>7_{FNWG;&*p)Z*?DUWy-<jOWe%$BQu*j
zDt{Y{eO^{`y*%g#i-mhmJ)0VzZQU0+&~&NmuyW2n!`HtIxA`#Ll+Y!^O)CR8yghZ^
zen-PCO4;T2e?|GW9uTiTI6Cxs>}lxe)FxIEQ~tyJdy%D7qn9Du_c_tl=yO(j97<R}
zRZOe?S~>cn;fq)GJfnC?Z1z<0$>+C9%eQ=<Y!LajSl>AdnPp|=Un$}@vE#IAgYIYa
zp3}{DI_CU)jb?dMmYCcU&=$C&^TdT7wu}65nqk}1iJ>UQp><ErYU_WJ8MBCz!SL3M
z(`9j~6;7A7nsP(yKHY8jx^P^6?6hO~Tk0aik(1eNv{$?T;~nVFyWj79ozvy6@n*5f
z>0f>owyj(9nHjtEk6KIJ1xGfP2=#1aI`8_DLoDnGCbOklL{ewv<&yS}dTQBi;$eft
zAa5glea1Fi>V=~T1Fcu)Zg=^c@hWI*TSe#R6K*ki-*b<bN(m>K=DhjWnv3SiN$uXP
z#}iCyKCL@Jmm8!%`I)Kx!sxwTZ+k}VX{)1FF=MB=FX$SLs|i6YDx;+6w=*<$_hn0;
z!}Np!ch5mbDf{~$m5=xLFWiQ|e|U9JYIDZM*e$CG(FA!Ou@pO5$w~_jNF!buflhwd
z`wougt+hQ3caEMryZPqYMnCbv#ksqi(i^M?niXH-VnnT{vLwlt7r1>Nl1DNFsD!m5
z>R_#I-EW6#b*ZW8vSj|fv0ZnJd@OoPYInV=FrhjKuH}kmaTX_TbK9da^+#RBgTTl&
ziJPfM(h6iFMGe94&%dwxdC!jCSl2(k)+RQVFCs3nw#q5XdLl84!9N(gn&VQx!oJJd
z@zr<drMpe$#xL)`s@t;L;qP$btwG_nqveno&pmeC&#S(&rMH})w*G4W<L(a^J!QE*
znKwFK)|VUT@)`&DjSp8Cv!u-jmXBH4S}cDSUX>C4&!+mzL7&*#6Zd>C4wby9vMsX+
zihOf3%FyV{i<a3pe!4;bp#4u|@7tll%-r@t&!f4PQumgs?Y7KVm$_k5k9Dv9iaVAa
zS-{LYH}QzP_3B7%so4GMnZ2GqLE!?DNnOp|8(rP*4JOg#ha)>DS#gOvo-Biira_O!
zDWi~AcG*>}_%EZkZ~i20l$4}}tq}M{mCT(Kwd)IBTOYk}<<F6cGR&?&2dLL0d)V^a
z0oNt9;+6TbM|4@u*^d6tZ%icoUFT2ACSJ}Noh7&LJ0R)Wk{zMIc9&FduZZo+co}ht
z@QlX@b;|2rGcvrdJ7e7C6sNN7`M5SKHLR+m;817HEpMBTzdX*Z><Rq+wOMX~1#Z<^
z`LZ!2-)TwaYljRh8b}(E7z|9q({yGUM49=}_HbCZAAE~Q8c~Q$#Z&cX8X!9xJ)l1u
z&RGIsk`dV-=TDevfE--%ZZy4VLO{}pMPxFbtTWR9nd7kUscEv!8IlGUktuizd8PsK
znzVky6E|M7A<R-lw$ZXNoN0i(C$^#a<KS5dNrQ*T26zLVnFh#R?yljrkG2kwG-42$
zj;CA9G(gUgrmQ+Zz+g08NIzHKUba&cf`#5fKXgN|$X(>4BC(g)7ZC3Ao4r5H3v0zf
zaE5h;Coq#+m~yv@(0dvX98Zj&4a^WMAP7tVf#cR9esjzaY~pF+1C|WInVifBFa?5@
zSmY;O#T&5p8?cUm!4=_&-&_p@GXfdmDwli`p9-cxu-HaCUqyxbg>)4uv7t(A3}E^l
zm;4(?0v;f{Eif*y1ZFe64PM&_x1S*9Y)ZB{Fq;!gvp3Oa?Zx8YXz%Dnz&z)CPU>qV
zYcpcb*2^{kW?M!TxxU9G1BZioMm$qs#=)==%uo@Z#zzkj2NyC5t$+spVu|0xorm|v
zQ)L#bVvxtd3t@LK%9}ZQGXtMIs)2(QdI~ZyBV?nn1?2apXSnG|RUUGUiH3=4*O){}
zLRS%VjP#T`vOXT!eV#?0HJA(HU6c4toI@~$LZN|ZhhQp&N&~mXzOZRH^L0wK76fyg
zI17L$g#0+cD&jhE)ikajS6z$d#(|OGAtd5AsRY5sfyNQ4U@(p}7O03v;u!)55Db%G
zHHvqR_f}~FCjlDoNW$0%omHC1O5~&>CDvStwO8i_N-UZ}Vf!zZ-YtHuiyX2s%Q!%_
z6b6fd61U@iOpoK+>zX)Ns3&9sUpQE#CsK<Qv7Q*!(sEYiMU6H1PiHeus0*e*um~qo
zBQZ{lQhLJn3E9n@wa8O=9z1k+IGFFw_W%x%sPilI2LuX}2JxGMG@YmfB^J#j#g!+|
z^KqSb@DQA=ldK13L$E|gqDE33sT!r|q^OZhN2W%pI;kiX9gmRn4|QHccAD&!9IA4@
z%wri_MG`Lws%?22Z0dDyCj$vuGMP*R!MY<%bK{%Rub$cv983t-SIuP;*lH9)2vH*`
zMT#aU1gB6^)JR5=sZlB=6{Q`8eeDT>cN~xv7+V;7fNKYp9LeJ8G7d^~#iykSMf#>a
z(@Ex@7YG(s(<p0yaCvnnVx~FM+(1RG%?!td#cY0x<h)|O!Uc@38!@zd{<q~ga!!d>
ziFUvY!Aa&xXx<{F&AtGwhd*u=?0>Pk&2+kuGviS+QpimOXw^NYXw{0A*Rhd7OcpZ;
zYz2bNSY{|4CmlUg{Ap|wS!%9J?qcxpX1l-mqS2Un(@SM2&`@s9*^<705&Rw*ZAY`S
z1yOe5k?)%Q^r+P7nD!U)fD<I9S)wQviT%W}fKG}Zy4LN~pM`i*=qS)oKgX_6H2J@L
zh&;C#>sgG2!T6ioa>hS&KUj*4&oazHOW|G1bz6&kdHU0CJQV2L5bPBR|LNB%xGo4z
zr>CPus+ZZS)#qS-8<|^-H7UlTCA*(?aVp=>Jb(H^>=T{95|9hANm8p91<nPiuc7oN
zz859%Dst9g`eEoR2!f;aqfr_vZ~A!1;a)Wo#ylL4sPYTrU^U|7_-a&0C`2=AI(?0m
zSTog3CDu`ibyMd6B{l@`__lYuhMoGOkkiaF&$9sLNpAIEZm+NdS!|YT)<R%LVt9U@
zE1=0scWo+Jii7`mN0=OGCi1lk`JXIDvZ{Exvt+vY?<Wp)Fc_h#D^hjEXjc;!KI2^a
zn6My?n3%wij}t5rio_T`s$&GIZoI0SsOl!Dx)S;FOc{b<(g;nQ7G9g6LxOb4`g)-T
zL4h^~#L!tr6e``=)60jo$k)%`$<V>u(%shVe_3fJrnAlGSXf!lwX>f$-`T}&p$BsT
V!_jrIA&t#R(g=<Y2*HM9{sk+odyxPD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/broadcast_add_3d_am_bmnk.cpp.D83172D8CBF829F1.idx b/.cache/clangd/index/broadcast_add_3d_am_bmnk.cpp.D83172D8CBF829F1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eb1b2b8fe8e8c1577adcff6f019e66ea2a160db9
GIT binary patch
literal 3846
zcmZvf2~<<Z8h`_YH5Uj5LJ0dF_Dv9)5SEaHK-g~*5D-~H6cfWHi&C`+6s+3AqQyQa
zh;5-25ua<N#UpMgDn^kOtYTHbQ&|M1MT__@yf^p8b6(D2@_+Ns^3T7_gm@l@v#f{1
zv9fq2GI3!#9*4sjAb$mg`J2bpaJU)*4p-9ji$ErS=tmrG8EG%iZpupXv9fHsvSGh(
zjpDn0?$fYu?E1<S;WrQb^Rl>M+&`~JHS#Ofs#`s_v!yeY*UL|qd!-4k?a3f!6~u|Y
zoaDxKzdROn^-WR!+dOA?$+J^xSN}1dm*#b+pY;!5#x^xu&^BCX9D6LeI=%2_{<Y=l
z#id_iSLUPJSF*e=QtCsi^<P&emNor8IKzBYeM5J1uXr(c+Zo?mrO({v9m>l)L_aTV
zfmF>JqlRiD<;j*u*3u49%%~fRMrmEmYn%H-O}pOnjt=0`PyL-+9N(?!S=D>`VO!+s
z(DuTsP{XDSGV7J~hZCOC_@k-KmDdeBO7suvljp6c3o{I&*jpx(cy&cJ3T9#4j&p*6
z;@LZ6r}&*gg{|ps&bhpSOV7<inDi-L*wv@&9Y;JXjS91Kk9O<QtZS2wtR-40e)6Y$
z%H?g8y&7(JEM5#Zt3Mm=n4#0mXI^QpX*;^9li85OFWP<Um!@8uq@Dwxz22Jnsqx70
z?Bk5ntZeJpz`t&KP#DP}cUw=}9kVkCeKejARdgPxT{GyspL4Qli!SH>B{h0wW7+6w
z{}<+Lj?;~b-+a!<3#g4<&FvLix4F^Iq+RcKzMI{yMe*wFuNQ_zkoI@`eCzJNH^D1%
z+&br{wr6w0R8z~2<*BI#`@wm^^qGo(1dn?KLwDSo)97cbp5zxMkJ6JXbdp8{MjFHH
zjy%Z=8-v@^P_)dacVOZRb8Xp{j?A++ef#Q;$M?Dg@K`AU5<B)%56kg~%AaR`fvjWi
zS3Q1Q-_F*#pm*1Aggr@XT;7%D@nAqD#LuKA<$3aAw;1-!S<3Nhxwp=EbK{{x=9ssT
z>hV?9=v=mazx{%}!h7`3!aUeBt=rG{c}r8=?K0#39}CYvNrm3NYiZ8EVBR(H?7l)#
zUF9U0u@Oz3D-R@}_*d4V))lR{izyAGiZX*up6xr1+b&C5#uly(1UuCd>syPA93yrc
zeYb6(R{F3!Vb7~4--$0Ch~{vc?!4xlizn9)Yzfb8ITjH2<N57%U0XNS_~g7fuSWi$
z6H>b5zT9v8>-l?Y+IRF-_!WDp_{P+xoij---}_|e<dUPq#KG#zK9|Ppb{?JBmDkXs
zW?3Q6qqmhlDoR<{{Os$ero)4oq@)jrUa66H9TqAwgQk}ZR5fR}I{Z6f8_%@a&x}94
zCwl*<nPwxcDLUKlk5MGFqwJTLC-$GD#>g%npBiMX@xAanRkQKER!)hVePHS1Fut^m
z&$`jR6Cd<;d3>>MXOpy!<N3$ZZq2C9O%o5xt5OvSWnXJJ{@N1~n|QZtQFAKBk&i1V
zm1Qg<Rj>BIN%6KdS`Ohh;&8Byx{dj2gJ@Pi<TR_&mj*WN1l4FnP1Yc5uQsT<j>o_^
zGRq07F^Jkw!%%;<0cua<A<DCv{3HmMj;Q_cep;&yP_y^X&*gM2B@$HQ5j9bRsJ+?%
zHP7X%r)KF6X9%iMh?=B9BCa+-T^HYe;;$FyG6|{)h}uBIKzp?T>U)0Pkw?a=e<!HM
zA!>>S#b&hu>U?o_!5b~){9c3~T5e_RQ-@%-ADfO^40D(q7Dn7aZYV}VbD<4JB6AT+
zz@4f(AGr~N<yvwpFa`g1`q1~T5j`ppyxxAjKj<d@WFhl87(S>9!J%5Arl1>wSrE$-
z5D196&PRcOU{ilnIyf@~C-ai|U<$H%iaZUYT(?|LK;-`gP($aVz(FuIlp2XG7fc8?
z1XCcGOXGTBB-58sFp|g0lK>fC*}KnXwzO0Yf&~r&7tjsCLb@;*BN1JM66Lnn_8T1`
zLx^jnrc!$_$fCG4__~VS#L7xUfs!dgqELxML=j4sdY=W<Xx;h)St*T_hWg(6LjI<y
z-Ha*~Je;PNW&k#D`d%D(c-N_9q?>8Xq=Ihnt)lZ$;(lT=cXNwI`kX2rmgq@{U=%_Q
z5*nbO=?61i9p4lo;pCBcO5uz!jzGi7;=CShI6tRQ!^1(wK^9;x9u6i4n`0!S${YZN
zbrPN$&h8Q+2NT<fQ5S(?Qq%dUh#;6mB9TEUL9ijokPL#(`MGjNI+fjm<i5SHJsspO
z1Uu_GlQ42Ia52Wn)yUNpBf1eCJsH@9pz~3kLa<4w2_MTPlW3DfjF<$b5jYtH2N8ml
zDAYX^Ee@m#yjcuvB-@VdjuD5(@y19_lPeQL&H=iJ!xz82vu8>diHQlrBm^^q%AleY
zb?t841GiWGNKDvzY%?$y4|DW5XddBVt{xX9IC_oai&};{vaY?8y@!&2wCVa_fz9~^
zMKO!*!AKIb@oZht4Z$2dM~S$2E=qYN%t5o8_ueCENw=ljfhi@I|9HL~uehxN!F+AL
z9_WVPXzgeviq(!)qB!k1C5qRMSE2;%1eD5;^JRi&#dX9?{yIMDItY%UM@1-YMUO`5
z^H*%MfWduKBqfPNA{nIO7ZGX|(_en@NAioI#X$W4!Lg)Rw8S7djufXv@uYYqN+2bm
z^u^ARzSPj$F32t>HYV#p@W4BO)ac4105150e{W1J*Ef5xQYL<4^zv@~QPMYbX`2G+
zCVP^7z;${=8i|jHt9pt!TH+x=Z@0EtBd6zo+P*`4!L#C_wFbcvv<NrM7gBer51?<W
z6AzDX>or@s#mz#M2C=MI=(?RHJ?gQtHV!h#f?*K`jsn4!3`>-b6ONtP^Y`L=#MDf$
zO!P52;WQygo{Mu`F_kJo)AL0B=Q)Q1;Wx-=8kt4|DcC*6EIv$$*|d^$40a5#5=pu+
zjtC<ziz@_lGUYe#wtf0*kd-9b5;Paja}Jht1yBBl>|5Yp0E5w&!)CIeg-ex4wK&^3
z+k<`aunWz_2_simS2sZYWXH~v&Q~&!(+h2dN-h%Ficq>)lmBkv{?JCmid3Ui^j5uN
zyQ8|CnV`R7+&T&RUMnX0<QIa5CK2Np4h%<N9B?f19IyNW0J#CdNtC3uV9>BdkNSv<
z)h%Sn0^<U-rbo#am$wI5ZC{DK1J4<ZMq=+Cp>(}S*Gkb#>ScY`U(GIqEt(L_#4~lU
zDM9!kB?`s|E0Ih~hSnbPR9nkY>k;ymqen!DCr1PLu4ewgjQeN|a$vEQ82$b6KC0zK
z;qdxvh-*{5Q++WD#2#YQ=zOsh`8|gG$Csa3uxCYBzta5oBLo6Y4Gx=}TFBM+T+-Q)
zDi8@1lh!Av;8IXME*TRAn3#l#iI^zD#0{93iiw%BEU5&+2r*v`udbn~rA>fziTZls
z24SHN20Gzuj7f$R6aN4@Baj&s>~7@piLD>a^8fXcP0h@$tZnS<9i5zAJv_a9eAij5
VrMh{ujK~~bnkqXsgsT#X`w!^moy-6L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/c_style_pointer_cast.hpp.6866D538A80ECC76.idx b/.cache/clangd/index/c_style_pointer_cast.hpp.6866D538A80ECC76.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cec437ee3ecf5ce1a18e0f869f82b50386273ca3
GIT binary patch
literal 392
zcmWIYbaQK9WMFVk@vO*AElFfyU|<jg;^LB`%w<422`E=FXYGWIe9Z;|&Yv}TWD{Pl
zlvO@rZZJ*L>#(9<i^?OX+Z-R?x0b!UF!AcPd8yr>w#20PFwT`TTs?d1DgV=<*S71d
zxE|H<>G<kT>ti~03Az>fh=qQ5qU3ifqwf$;P*<Oh7sHFc9Tzs|nO<gYs?XDWyT#Xs
z=UD&2&HM6O7}eXwPDomB=SXX?oV0YtgZSe+`5zwW%DpYO;9bJIy4&nO*<`LYtuL<3
zO;QB9*3JLT^$Fs0s(Bb#SQ*&(#QC(r1Oq1%BLg!7BO4<_QEFOo3{WNS))o6TSPS+t
z^04tS@Nob|zyM|dBNtfXhI!iiQ+G|=wj=Gq85gco4?(Iq`GjGrFYkA<4i1^H1gM&Y
rgGCu8T$GxV2x1o(B^QGQ*<msa3~aX<7H2tJUzx$g!Oj4;mVp5PjVyl%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cgemm_xdl_bf16.cpp.D34D118D7BC516E4.idx b/.cache/clangd/index/cgemm_xdl_bf16.cpp.D34D118D7BC516E4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa65d663ae6f3a44a5e1507da1805bf4062c95ed
GIT binary patch
literal 3280
zcmZve2~-nV7KV$E3Nfh=2q8d%iD4&*$QB|Xizu?XfGkCV2$3})NI*7eabZ9~5g4^y
zum`c*9aj)hK?gw;5qCfZM4%f7ZA4p+j*3%c<|V~BQ|IKK_uc=#dfR=s66EXc-Ht$z
z*IM7~l-P`DA_PI&*iW93p7>B5f($7Tlzr%W$cuGDLanD=e<j++9Xk4E-DX$OvTHJj
zUj0N?4zq38$0(p&;QQr5)IjyzR*tz%;8mEo-M;A10(p&H0dhmPaQ$)U`ssNcF<;*h
z?HdkHl@DyJlJxZthh3a#Eq^^k&8ZhXpD#XC#kHJwzl>=Uc(2*bthIaW;HlJKawgt&
z9o#sP=H6jm@@3>_-JGe|rgpu{mA8Apw?w*Kk#>r2?0b_zzqe#0i4o#a=y`U@kZ@Rj
zDPe=B20yZ|p}^nk%Z_WN#(U2Q{SvGKPmRw0V_?&x(O;J~yM#Bmcf6XlbI@tL_Kx@J
z@{mSotG^w8d%w#sO|lPH!xP13(VT^P<0sOZ+A@5^v-!I%vr}fzMIQ+X<Hnpz7xU`A
zy&rtM)aO>C*HU(ZPxa1-$=w&<ekf|$wIoPO=zN#L)a^OYUwWji@>csfqr?Z&VR^Mi
zWQT7-+tuf-@$yTtyH<AJGGFx9nbP}3YOhB{2O@YKS}}cB1FjoJmCsF~CRFu3t)-p2
zoaa$mPjt+!z3us>>e^Pv+?{^@-lip^9l|{|=BviEL=!6#>?lFCj_gj2vNTQG#s=B(
zZt<?t;0p3hdQ$gMhk|)$hPM5gNfTs-7eR26y@f0MK~{fMxAIXy3*966`iMyPiG=&x
zUV$@r{NaiPU7w0d<i(ngj&F#V{3=LI)vc?u8WXaFqbWPqEzi3_U=cg&E!w>}=Z>l+
zsOtp2D-C^T@3!`1>n@*6X@hj~E^UzV%r?;9FNha**T8@CRMH*4b^K#4R{7+Ymzr8!
zK3pqay^niPus^|C<8eOyQ&830S+RRu&u@1qT%PA`kT#@O@`ctMT>kGD6<x<HxBj|5
zz4gHj&R&bKm-_-bMk8)_R3Z{m;OQPW&d`A?&+JcLYG$UqPyA+cA6{N$-THZBhTZFq
zzguPP7fgytFW%-m8KjgOl#3R2#`<P{FY7eqX!W@a2r6w#CR%-jfo2V#|LKCZoUOU|
zaC>-?m7vdOe_*ZcYpLMLFe|HeXYb;>@qJIDpVmB%@634ca&Sl7%XiWLsNmVYVm5gu
z4PTH&$b27nA)P73?SU1hgN@@u@bX8}kw|f9ulo8^zX@}%Twim$nL1%#F*$H>nRMjs
z@!aN#sR83*zoSuGKVFd~Y>F(okXZOHJArdqHNTg(OydWAwd3u+vtL;v_ibE3Q!?^B
zfW6kDFSx~NVAP|?`bXQxhiBbw^;i8g+VM@-{LUfGw0-4zCXLmn7h9S}y}wV~G2d;(
zHuL<|w8K^3({o$P)&#grnKPal&{IQe>U#%Txd-`>JSQb4Lk)sh6~3;-pWU`l2|*Bq
z(n)l>!eDn_@wlGsI$YFPd4ft<fytUAO}fGWIYE-wU8A>im`YfV$yy{WZG{1HahGMC
zkHwyw2o#OUendaA!T`D4(_2mQWGaSA2*hL>iAGl#AfF?CO1bCQK29ZsVKR%v(pDHC
zHwDc~vdIghPzlR0nL%RE6b8t5Qd>^#=x^Xq3E`N`Cb10_2FTB!pB&6NR{T4a5Q52E
z64yXsfZYGeTzb!J|0^nCB_`{TbZ80#<YDUfVBTPE1q}+tWPO%CUtxg!XLn{={xO<}
zN(jbeCW%Q?7$B#|#>opw5JWgJPRdE{F;_!St}{0rObHwbOqp>M2+FbK*nwpss0nPM
z3A!Mt8El559CnT|*nDU1#9bLFdV)wqQyHmD;3lGJj5HjrXROCjIwM_4AP0z)8E22p
zhs!Zn&Bquv#?2s`ovoZfpi@OE%er(owoaCa<ps_)>colqQ$Y*j5d;<U#L8#{mGC4u
zisVJ&D4G|IqZnQcj-)&(j$*ZAb-`j0G)_AXNAcS6I7-w@<ftN(^pbES(~~I)WIKT}
zBg+xg6gJfYN3p?w@ja_>;R<ZA`it~eflS$)->~iGtHCSN?oj<m;0D90QD)qX5<m5A
z@5+9NNC>);ypj&05HyG!gri_`FpfgVAvg*phvF!V9HykwG)Apc-Qpf~1f?-)`rsBK
zs3ucW9~?G<(wTI9kcuD%OM&TY3KAJh0UeH^Dg}&SBU}n(v9pxNQ4w_P?BwWWDYkqq
zrj{Pq$?9Z(qTkBDy;dWla)w-)akVS>m7W%DkFjVDk;4E3fW)CtW)uP}$bsVyf-v&r
z1mI{H2UU?dY_2>h1hs%IaAXNv;wYb;Z>rk&pV&X)XcK#rlFmzd?2eSu8Q5G?X{p+v
zOM~~G>{^dMdJrNiHW6C@HxZSXNN^Ns5{aW|lV}{pn8e^nY9hr^EGw3+dS@If4oC5<
zcpN1fCeBesCK)E-NM<NgQj<f_T6oEkF#>`zEE%?-x*{mkl4%PH8t|%9X4F~)<-t4z
z+yDg4W#{t2m|*X*OEX=BRjIRCjlEfIJ_x&SoL9WXEtrmNqns`00{3&5I-@9P%N@*}
zg=85h-T$lJ*h{BNx9X(@YU~0vYY_ILaHP9w2)=;z&LC$XzzvE!wu<S>uIfDpo`YLO
z9y?FD3UFVs0jI0Is*3;?D3fN4e7cO_d~UkV=~(JGfMLMqVZ%*-3RHdhFdsMEM)pSK
zaG%He%jf;R#z+G}g|LvM+F2*q2}dGWgrkM<LL9llZa8v>-Erg%dn*Y_k4?tD&e(tb
zbg3NsTtSd{+WhyCfCdCbD4Xgi|G2NwewV%{q^OA1krFW!snnp=s%Dg`DNR|M9FL)2
zY=AluCTWmoQ4yLJUDHQ<@gg&AiVs6ihpBHMTHxZk(9PXrp01U%vCx9||GX@Y;p{nl
lBNJ0|OTk>5`L_0sPFx4JwVjtP%iDJiVaY1brE31re*l!2$Ws6S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cgemm_xdl_common.hpp.FBF552CB8A2753F4.idx b/.cache/clangd/index/cgemm_xdl_common.hpp.FBF552CB8A2753F4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a1eaa350aa4714de7b30be95c3a1294b0d8a608
GIT binary patch
literal 5352
zcmds*X;>3U8ppdr6G+k>0u4ltWumBAK!^e!XdGh@1Oy@|IY*G2pe&GZiR&t$j<cfZ
zsyxW4sOxy3BcsluK}A=66gD^t3f>3fs^|dDj<~DtY=@aj>eKp(FZT1Kp8Dlq$6M9!
zTUA}L;h~`q9Wg98CA=srH7^mzFwCC*>GO12DFeec*kV{wZR_GCvOAL<{~^0#wN2t3
z5Vd=|ZCqnw4bN>GbD*;S?lk!a``&1+<imh|TS0!twY@u{BPRX3wqKvzK-t`yfAn6J
z@p07^N9UgPg=s;AcUNUl*f@cqWkRL6rMH>#XbsgQbqR_!<Fveg)^`=g9y~7}Tqo>n
zTP3?xCYrrf<c^2E>mc5D%%$45AO175zVCF1BEGC}<EivJ2Q^gz>?_lGkE2}|jBgt5
z{@M79A1NvOqh^}8AT{R6n$=y)R+59enVoJASCR37=jI<$v8PY%&fXn5HNRInVd~Tk
zmm`KI#}#p_qsIJva@7x&JzK)lj$t(ij=y<-I6i47Cbse|5y?vX)Bf<q@&eBrtil_Q
z6T>+tjT_GDY{Cn+++6t9b@e07o`*xeV+VT*-hFwQd@?KX!t1C1Iuo?y&Ryo_+voOD
zCwJavJ?Lri`>9;a3QaiuVOR0Wn4ICk#T8`_rzpFbfgU#0rLm#5sS(&!BK)gF{q{rY
z*Q&zKFStN7M)Y0XGUY*-r;@i-FNs<Eyz6(w!B9h3P{Bp%l}=UtFNbDq+s^kA%+z<<
zuv`w6*be9F&j0nZz1q31`;OgS%Z~bqu`jN6n)}h~2H70rrRrzb2Tqr;+h=?`q0!}<
z3$w}Y?*9AL?KKNeiF1xdx@_zZe!seI(6R4od)aVaS^tHoJ)5rjUU#XEEzZ0*pgmV#
z;JJ54^VgG$eTKbUTWo?hyVlHnKHE{wxAm-T{hM!Y+DnhvX5No~YAo>^|6jcueinS*
z&i}i9MOIQT?NK3{UJqp+d9UO!m>5P<Oo}y<U@TgrpYj}sBhydJ>)|l0Y1xVC#2!hY
z{Fm7MN8ddAE}6ryp=C#=BWomq@{K)dcaEi`?%^=#=16CzGkYX~^6gnC#wB2-R|w37
zmc6aL9Yzw2g)=Y?fjzapZ>%DSTO0=M?IfSc=Zqw<_S^pZTm7Cc9mip?Y1y7>Z!?l$
zEGsZ#x-K<M|DH!@6VA>x26)*q2=WyWyaHb4Zpqic`!xuF3<e^jC2jMbGjTFooIM-<
zjgz_JT!i#uJwo~7e1r<cg$S(>uP~F<6wZN$j-+yeAmx5?bDspM@KYe9^iv|F@>3zC
z7OUq%qX<$X)*zH4$_cb+I*|^cJW(D(1)>6kie?u{E%l0L7n_Os&9HK{-Lk`@G8;_e
z+5VT_->`9lOcW;0uvm~JOhPDGn2b=0FvU!~7LTle9buiebOTi4M5F;K387>v8KD#^
z1)+2*9ia>=10gM?MJSWXG!y^)n#ynQSAF2(q+B2$2Ro)IP(IB<3V{OA6oCpvQwo%b
zrVLafno6KTG*zGq(bNJpqN%0od6xb(0u7>Rq#8tP1>9D^vy^QG!d6QZxZb#!i@gx1
zeli1cn>J8bGofPVN8|l7Zb{HrwjGZEK{f%Z2?*d63D^}m>&7@#GH^0P8WIBk#>w!R
z;pTuIdrFt_a*cRB!0FoDCSVVXjqhu@%npoS{L%hJvOO=HbA1kfwBSsSp{vW7i9>AK
z%eZpnO)9twgp^z*LMpBbAvITxkcO*4$N;zo)ISVhq5=FK4hJU<V3q-x$8BPz0jQxa
z&HmcgV@}KHtP^l|0_<ORrICdl!$J+~eq8=Ct-*#MB|M2UG=-+I5+OqV67zyVkjDY*
zIIy#*$ASBCU=A9BJONlIQ2itzPXY>-Uwk({;%0T^oKaa06N+lHm$n+7(4Oez;N$`~
zU9C=EAbY!CLJu<=&nBTM1euHHBBaOl2<7AX2o>Um2(7?ZAhZ%+X{Ked34YgKCk)UQ
z1Um-1!LF9a?z=4*!rMk=)VU{mquzLw)s0HyrA?7padqEcrsu92$f`j+91=m+09g&n
zwLn&javhM>p<EAS^(Z$0Sp&+uK=dxK1h)U>V2FEAXSpjqlavEB272(u2&=WPDt?*4
zBuH<*w+AdukUo4LgeLJPA>_;VMaYluhmeRbG81$dI!sJQwBMoxD=!JEK9u_=)AUkf
zgB7q(cxa<bn@()7^p4@BOG@qByXn%+z^55Zhvn;~55yiZ3&>AqhiDvd^6Gox`5*dS
zYy$q9EF&ez&A@*%$`!!B0_82he+$Z8K-vZ7!<iz;ZXoTp@ZJr|oj>HCai@<t>B4km
zEy@sPAfy#)5y}*1n(1)SY3qeqjiK~AvFF-5!z1reucvXEp_oZlrp4IR0NkMgCZ1xj
z9BEA)VE~a(KlVkBe`C3w105A|mfRfu^eL33KuF0_BBWxe5K^<$2x(XvgbaXX0A_CC
zKJQ$glZkzFe>=ssy>?027xl?F*qOlC6zh=2epBGh50=5%Z@alK05hJ^j>>}9t~(I(
z?n##ZlbOU0-n?7y>|W0NWCmpPT)Cn2^F?;1;1kOw&@1TR6Dtr>ij@ed#43c;5$b45
zJxzoLp;jPjwM72+VHNEgo7>$+#n<3>J-vgNKI-8Pa?|6qb+3=0<D&r&hd|dief7io
zkK*$l&X8{9jEc?hHi7Qbbd>=6?9V(B7eq-f(l&?iLwulR1R2f`H&g$%BX8<_IveP(
z%_uyIZ7CFuM<XP|We6?A7a|mc#~`!_UxZK_ux$feORE=v?F9?NMV~g@bjbk?|8UMY
zv07YI`+p7Br7ll2RnY5_^%bU!d9E>xGV!kuebWdZCz!!7)7{Ch4};ttX_S6<FQJq(
khVRU^vv;5zop=}@>KMkgWbG{3F_x_9_VTyK8D=o-e_=H++yDRo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cgemm_xdl_fp16.cpp.93DC5CFE4CF25E39.idx b/.cache/clangd/index/cgemm_xdl_fp16.cpp.93DC5CFE4CF25E39.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f14e6179f8a7122c479978f7fe32a75eda878a4
GIT binary patch
literal 3360
zcmZvd3p`Y5AIFa|Gc<F?G$SU*xQ#IkA>`U1<(_&OmzXnh8xm4GB$wLKR@#t|Y(<d@
zk$T&eOY*KRDsQsIrYKRlw2>~mZCjh(GxNUB%<1!a=kuBW=X{^v^PK1Z|3Clde}sPC
z-V13Ev_j&SxgkD1h6q6r1N*0?OV>Y9g&<1`1Z7q<hYhC=JI(*KZ#vi}p+a(bhVbqA
z%2s;^mhq}$pIn{~GiKjF$wBMSAN78V9bB!S(Qg&)Tj5+Wc35|uXsB{N;<RP?&y{A~
zSr<R0cWJgB{#-WV7I1n?%Zqe3b|v5K*v)t25BL1__P4hIHtlY~746{yfzS1;y)(ZL
zuM=(klzr6dxc=k|N!voYU-9WE?SH3dIc6PmcY9o0nscv%wDoYWD7uS+ax4Vpj~ZwV
z)7p_ew85Nn0?mt!jguCIFOW))-mvC-C4A#_%Qt%OXIlo6##@ICTm@@QUSDZGZ8~Fh
zmnf<(HdpOkQ@3qF<&RdKj#KZe19?SPtenh8RwhK=swuWL-{DhvcPKo(_+$Vyl-F&M
zENcGndaU1P?=_^xjuv9pHqSr(fPMONdiNpEgQG3xdkNfxj{B9m-<PfEQte{yp4hK;
z;hkP$vOgi@&LIc(qp|%?<WNq;htVWP<2zT+yFp7%xs-S1U07=SUYhA;>uA3zuDA22
z=Dho}jJrLC2OkBw?(dnMT9ncnEH?M3Gc3N;v44C@y{sq4Um&40@A@J6dNb#$UBz<P
zVcKH)_dT4NgPzM@e>W6nL#vX;@vF}^Jr(K?j;!b|d!?n}+D^+7miY2djAa`R+zG5;
z`kT{(^0o9d&;i!{pwBVsw65jsU%gjOcJ>UTUT1rU9-Pg!+`{u7y!Xnc&u&*}W&Bjt
zd%+F;I#)*F%<aMyyODM8BtAc1xHn*PR+_(+c65HN39mliVxoXsw0$s<*nA+gu-uB6
zBpMv0Wr>Kj(i;t&_BtA=zoY1nlEUc&`J7{(R!<jn{#3xM>h3#~k7jpsHn*5Oyutsj
z>|*Syz&_nDabvyp_lLqxQJ&=!QwX7EH{G3UjyxP<n3w7PICRNrt|r;LKSt`%p)ZD~
zY)|zR#$BKDjV|w)<z3Y?^}cCk9b6n2F}KcfQX~9bz!qU`ggu+)8ER6nudhPO<ZriD
zM0~!?wlT15{;lVG)aKY!`?hB}Xr6>~&BM{P$t6AGm)qIBu2Ih_+8_TJa?g%@@QR}^
zEyebeC{Qg)@S&rtkm>X(AcxhhwmP!ZbL@2Tv`OLpw6T;*)-6rpY<H*2<713ccb!kW
zBMPoY_f}O2=CUXAn3wxE|7jB&+xxWTO8Y?AgOMkT&L#1V7F?JRL{%s9>T|dAs<vmH
zIZ=0M^PuM7jmD1`-+wWs4J@>7nyXALyYcMMap%9L|Ful=ON01H`3sBHJDv2#c*dsO
zBYewnE8^o1VL9;Pj+1T8)or;Nd7m?!PZ<~oM11t+SZf}cIQhJX`vuz2TD==e%i0iI
zsscgUC4K_p4{kfD1R(^WG%}4YH&`c-{{(0Hj^(wN)>8?Am<*F)y4(QShmw0>Jm^{x
zl@N-_bTVCAZh)L9&Kay>6+WjD{(;FFWDSPg0C{g8zuDKUtP6%>G1-skr!F@@F81(N
z5f9A8QVFXtnM$V8<p#*-iGOUkZGUr?N(jefEwUCvZh+h=oR?&svxY(?tj1&pnL(2q
zAm7?}wPD{#t1gugfyp{#9k$#6d1&a&SXS-c2`V8Bll91YI&uT#kx|pQ+q}c0R6-CY
zGs#Sv+yHrWFyuwwfU_SBT7$_fZI+ST0QuG6=1tpbX--r^FeYn~HED7KWJnsHkXA>A
zAVU2tIV<IXsS1qhx#)!hpTLn3DPJ8Gj3Ru*7Q_Xk#w246umnakkxXNdZx{VHJ;~&l
zH$)<u%1m7VhKcA#=0+TCVs66GH_UGo1gZk_KvojKam)jK#8l}4hs42oum#zo@W89^
z?VF^gzr$5wl&V8zgOi3)SO;c<x5Fq+hsFlcf>I+;zA7;oH6fX3fjCz=mTc6vEI5E=
zOv_1Y1t^iI`ud9vLg#fbjEW3I3NK(3H9&DBHW1?|+8`Q7F$OU>iZzJEk%S@90!f6?
zct$*q5*P_MN@OMKDrc@|t;bOkD@j4%;$XqbZjCa*ToQLa2sUG-=WUB{hd?Y?)^b(|
zD7xHjt-HEL$J%7WVXR0n47OFFe09n4{Z)6(6n$|b!)Ty-ARWwt(IE9890}EhI0{h@
z!BMDsD2~F^!xRM06{{23^;L#}Xd-N7ofwgfaCMS_WGHf5BCI|4eySyo0HYc@8f=gj
z7^Umb*`Ok-Q&tlFgZ4aCA)->IR8gc&fkvesX5IZ*9uOkJ1lxg{pisVQ7ucjFh&z~s
zkp~ihqt!@+l6WMZ;t*ielw^t{Gm;sOwjx_ilqubYY{SvF$hQjmr}%;Gi2^zk^EH*0
z$^c7Ry~{G2mQ6l@iKvJxG6Taz6y>5g5_82kisnY+D25w@qgZY%jwIR=METBm?RXp|
zXeZz(k)3F$oVlL89!E*+Bn5RY5lTp2CthOb#o#k+Ko5pdO}?fLXnSB(mGV_bhEaWz
zJ`9Q*MzfGCBd{mPJ$4^tty)<~J5-PzDvQCiJH|PCce(}Bu{25%sXiDUTH4Hsf_C1*
zhBM#{9mVkfdOY@0S%X);bUU)$5=<K2JwDj^jC288o2H%ygJJLiz+#Yff8|;Ok^wG;
zY$RI|1GtFTcCs&matWWrSCq_4^~NHC%XwJ?WAd3xz&0Rt*mkndg>tPS$q=_)4w9qT
z?!(KGVylU0jv9>GlkCaLygHB^aO6mG#F2m`z>zD-6-RC)HynA9yc8sjPl=JalqQW!
z!@g(`B$k=~{vyExq>8%pFRK_2%p502i=x&>iXo9ggQAp=Yn6}D%E!13l9Y8A`o{;T
z5=mq=^?6j7HlMEH%kW&zV^DmVEX@TAb)1}+3S8aXmsx39xEMQ_8T`LdZC$pWp%I5`
iV#+sPWWCtN&fZaf31VrxLQC7*FNWY1;^CvR68aCurr-?#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cgemm_xdl_fp32.cpp.D4109241584CC5F6.idx b/.cache/clangd/index/cgemm_xdl_fp32.cpp.D4109241584CC5F6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2f716bd3dbee1370d92231ddaa647a3af58776ba
GIT binary patch
literal 3276
zcmZve30M=?7RQH>L=2OakYEB~OCVt@JBR`W3j#q_2{-{UK!~hitAL<_MNzDRS_BoX
z2tF1CMIWeCX{}3(8~PMc#DzkEiio)N`LOm4`QA;2?|buoIp6&5|K2nAo^$3*A_D#W
zufPzrEGe)cL!KwYLl8tqf4O-%DFfONWJrRbg1XMA$C;1aY5jdKBOQ`o7Mz*F6~69t
zab=qyG^@m|*v0>u*&-0df0)b{iq}nz*01=X#3XyH%1M#A5?2L1w7IuW{5*ttwy<k9
z@2XyH$mbvhwPEIojZxpM8@r-X@us-2L)LSF7c0!%%C5-O+fhE=Tu^)aTD<qY-zJ}(
z$@}=uJ#_hvs~O)jQ~o3G2(HO*82>nOZ-Ym-<t^z?HNB&dO^i&Ov}JrxQKzi>$x-q7
z;q}wiil6sYzb`)K5ZB<;{HRreILgb~1wA95W<ygubB3luS-g{a1;fa_gX?P6;@a*6
zRF{rf-}kDxy^8%KH*YpfxOcgXZKyvjIMY5lQ{`w?vi#89!I+rcjls}h`8De_O6tt;
z<)%v7c9(KH=dk)w3*lYg2URze6e%No^RreHQdR8?ayFSSp%AEeG4sF@-b9_!M;ggE
zlGF9i;kA_fD-O+!Kgy`>tV<rxe3(Xlk*(rI#X@OKN6lsHJ?C3WxoIDcwRsx8JWOkY
zeS3}iM;bq|4{agmSBM2V`uv=lfeu64`iR2N?d8Gs`)Ijs=OlTraZR*u$g?Er#rj``
zg59*@hOVe!?dGAFpb4MoY*A?1wCPAtKB=7dWUGxRu#tH;%ZQe$5QZ!iRF^kjt@tpz
z;6D37_+P`xI+uJ6`~Afm<A0xOJNh6w<K=^&_ASwCSaRo;eV<cV)FJuAkq=9Mp)`Ba
zcTU~hxism*jj>0&O612{5-Hbh-?*gv*6pipbZxJ7+Oza?WCQ)auf#<%NNY}s`^IfF
zg<f~v=w!v(qpr;0x1RlC*VD`FDkioEUlA2Awht=dcuVb)-|g~CRf;`~n_$Ucg^Q1o
zGA6s@A3b~oskRzgEm))1?CbNP*P#FAgI!@QL$=2+?iktRl;mt=q4-NS-MYL+6!pqg
z^0E@Yu6H)maVR^XEZsi9Z<qYz<zzeZsj)+^8bi*!Sbs+^s@mSlD_8Q}>!RuI{Hv1d
zr+lM)`3)lKr^T$)ykGw8qVxuAYsf4{IwwY(_valb>bbV~>mR0@8`-dMi?Zs)qofM1
zAnVQkBU@%lBlq&dYr8gFXbfC$SJ(E|tk5^L|Dpf*W2c-?mWe;wr*!<;_J{s=UGd6K
zF}w2*Wd7|Ld}7rnmzbJC5_Q1k#_<Ea77Zt6#;uvRo;$zp!@oPUTHo(U`(u{D;nOd4
zyVIRNfA1WYc%i9V*&W3Tm2SND_UcJ*$BC8glQ-V!Tl&<6WbNJU#JCm4vSDwwD>&{+
zoS?1^&RA<#*Kp<;F>$?t?E1gF>4p36ytWWTTYRL}4jl}KatkvO^0Xj`wI^^XzSe8A
zE-nIs5Gs*MRT;Dcs4kAeVteNMi(AQ%1eFa~2Bs<l<fW~Bhs(S7w&>!*P?<ucP*n!V
zYh#NB8rVBWb#bAnOe50hDg)$geWK0)Vf7^#lA&@CK1fGpfV|tsUn}<h=LB8c3RET&
z$yAjA@+tiLjGHdMPV3^LQJF<#(NzY>JrN62zbJ|(>Ec3AnL%WbRR+kt%FD+qo}T0C
z;$l#lL*y8$43Gy0TZapqwvFlHqEMMf<SkMeAU}O>nRwG;|8rejI4bKA^~fp%<W~b<
zXKiRAyX)d2QJG0(l2r!CIr8M(Vj=|LTBeDG=?^WnV1(zviviCBDIt*Nr2<B{BCaFY
zH;nKJd<s|sBNhY;3>9(;%|UcQ3UQ=)se=(Kf)x!o@Zil{Pb11^@OVVYP%^<<Jd(x8
z!caCN8$&sa95sPTz?0@x0J>=&*HW`-bAmZ$n+2Q#^)^?y?NPF9zpX-p!E$FU1KEvl
zY3Vv1u~-4ah}2lBo(&_h#<3WRGmgWM%vgq@1mgq@B^oDUNKTjQgTul|GCdhX3c3PA
zDeM%kW@aio6+>z4G&O;mN5ef=@F3?fA|MFV;a(^H>ZWy!YZ!W117CxcAVsAc&Xrwy
zK74*|SYi+dhQVXCNb}y09YNY<Q<Z<a6JaD=C!7jq!AOKo1coAYA~6)D6NMp(js!!|
zI?-wZFCUuExi?R946Yv6P?JwHf*B^C`J8-pZ1+Soo%}FyHW3daWG0yn{J{u?NnwL(
zfe|W`$_51wBQz$B4FXu7F2)CiZ~aXRkK{6P)rHm`X1do$c<mk<BQBoH09yg|LL$xU
z0_Y@Xt~Z#3k`Ff+Lm}K%8nPr<suKhwLV^%OB7z7*8#o&Tns9IAY{bwe&L%Z=#6EO9
z`Yn}#?o~-v(!r8*{?!HTUyeV7@raZ!6@p<r63dUpP#ixFLo&V$Lkau@3?=dtF(hZn
zIU1kItYi!+SPBfK7^awLW~LgZVkpfpO-(({5lIBUqmwuoVTc$GptZsXQ^a%t%?^xe
zljb#A7%?Uo!=TT=$a>CtQ}9d>d-PqI>%f{KE7jtZYT1Knx6O;TZT5<!qG8PC<nqAq
z;F3<WcxZDkI-C#ZFH#TxuXUqN=lZtBX^9r6M9UUTd%WfMfu0e<X>@g-P96+~L1#x#
zG1t~Lt4#<dm{SySiqxl=)HzX!^Efrvx%5PO&cFkR3+id^3!w32N-)KED&`ccJ^g#~
z>2ABRHD(K7#FgMm)CA3q;D#Z0f;)yh37#18B6wlQo8XNhe}cc7a^&ggM;HB{pOcu2
zexx8sI%iauHnObsbHqt=BTr)ds?~9^QYaQk5VT5jyIOM_r@4*S+$Lrur7KVjk_T(!
z3B(0D3w2>KjY<iif8}dICj~IrdQ5{w?u(Z!_4M-o(oWyn!`xM9{QpK-Ttg!hQ!~E6
hQe<WGg}sB5iyP0GW9zs~pXDF82IsfZXStRb`Y%>+ybS;V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cgemm_xdl_int8.cpp.7D0F341E5936B471.idx b/.cache/clangd/index/cgemm_xdl_int8.cpp.7D0F341E5936B471.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87eb49378d1c077505be5ddea65494d16cac883f
GIT binary patch
literal 3284
zcmZve2~-nV7KSS!6+==XB!mD7CbDlLqU<8OVH1a4C_-bxCSl)XM{z|Obx=^TZCp{I
z5fIS<ZS~+bKqHE%AS$9D?T)B}ah%aEGUp}5Ig@j8Q{VgFa^Ji6-Ab^pP}qk+kY}84
zerkM<m;^x(6Z>Z8WF<XNfgoKf1mzz*7xpM^+)n-e?axWp32!@+&b#r8pV%JwxhkQ`
zyUf^|C9Vkd?cG;evaKzy<yYtShn5R`4mwWs)$(c}L&)xr_7|)UjT_N*?_6T0NhkKW
z-hDW}w%FDGpP#<p+VXs6Wu29kUEjW;O|%=gK6qrNNv35i&0T{JWskJb;&0h|7ack@
z`D$%v$D2L+0aq67ef6>Q^OE(2r9R$_;{^<vQ?2d>*dk?n(+5#ji(OM&RF0XwXzJ3(
zmBVdqbuBh0l5@zmQTP4)PgLGHcb#$Yhs^hM_|A*}?YTObOKcG(7Y@E>1svAoHY~YQ
zyk*3D!12xVyiI8WYWzpjw;PW=zm)Kx#+e^1gg=~Git=a{yQ@{*zTL@>zvXV}Qu6O&
zT2(~$OAi%IzqE}k%Q$-Hf8MtmF^y!WsYjvhIxV@$LdlE34{ohqM>@yL`oD78ZH)-#
ziT=EulNsK+o6M_L$;PD2rcd*|27+wrFI0DJYs=Lsa{OJd_7`r)nS_{;k!xM=RR?+(
zR6h<{zN+`iK;&@7$&S#Cf$sWw1@r$HV42r^sm||aOI59hDw0>sv&x`a54X0o-se?R
z2D4PV@}yBWY&P`nsNc-b(cJg3eTKiJIp<`?%YoGH-MhW~nDimSk#$?2R|f_cl%q{+
zudH_~5zoIbNZYEu@On^c`^1-k(NPO&>VunU^pwt}))!RXseN+xjJO}URZV(3`|XAi
zje+6*$K9ScFQ(MJqeO%saGE^)SU={P@%5zbzS8juM}wa67+bEpWn5TnJ-zI}1KWJh
zI!;%^<LSn7W2tAC&C`%eaR$pKj+WPTdmX$wq;iaOYWnO8iEt|7_{{m^pE7&b)Sr&f
zxHr93Q!Dwk|IzxehO1A)h<0V>?6^afLAO%gQEGahoEf?3@ZH{1I4I6~Aed<`%&852
z@m_MH)PK#s;<v`v=9>t+$(uqaBKH1x*jG<@@cZDe_sr_Qoi}MT7l#&4Hkm#f8qJ-m
zI9)B?(zw}27{ZtNFTD4%ZYp*D?kaWfC$e_wKu^%pS2m2|9oetfJq_};(hj6P)w_ST
zhi&gSHbHZzjjNnA_C2~d^;^F0p3lJ!QGc%ZSb4AU%#QwlBu>ckZk9^!N~WR)Elg8g
zpUTeY3v8ZzFWMh*R+ikl=$t0MIeG7D+dToo$no2&O%HS>jR{m=#uSFJb1$Yvt;z3g
zw0hY#AX|UeZO3%o>&vgg3!GEGGH+{L`QCjvq_f#-(EcDP)2S;!*zaTG8QChLTo^k&
zqkY)*u=Tkkof$j(OLZz=i8q`1|L~fK_qg^;wWxMNV0Y?ZW_CfUB!>V&oZY_8r2Q@%
zX~bX%LK$QRLvFD9FaJ0<-mcqrVNU~%7>LPgWHpA|0C}ybXy};MjtLs^Z<wr3R%gl$
zkPAePSHFId4AG%*OxEUT>&Xp}%Ws>V^D(XILLf0F`;q)8as%WlcOgM^@4JLX48UYM
zna+?KAh(kKmD*?9Jw+p~#$*ne!;~8!Uk+B4S{1FP(uk`tnMG#N<p#(->DPX%813NE
zh!L2~C3AJ<2FN2Ljn4{BmQT`%VVKM(^L69~$fIK>v3<t1V>DtACToy2=yC((_d~fE
zB`4_)G-3!Qv&n3_+yFT%J|TMx8G?unQ{;k_VG{y^@}2k*U`-Gbm^vpE2+A|#EdhQb
zs3B~q28JN0F>H*Z0&al;I9(_HbZ;_QJWV2@>8x}%XeOZ<tPC7wvNCa$#mZ6;$N>^{
zPV6y%_&gJ(KL)S?&VzhzzQTh5``zgr%gVoF`{X!qJVD@M8XC_37VMaaAZU~zO3@oZ
zMFJ6yq6N`75(~sQk_aR?iWS7-D4rRw2{wzM3Csi>B{CCnl%$o!Q}&c<NpX~{m8>9;
z?L_LFEJsiy*hn3OVx7NNpGAaSAhua;5A9HJ1;8&hcJ>~CR6tO3*jzETtfXUe*Vwbp
z*|Fi;(O?YljX<4qKYsR8+5B$T4+k=W22p|-pcjG$Q-X06LJ7f97$pox;goP3t){G2
z5V(9;&}VO+(qEniPgfaqeOMnC^gM2!A^=BH{7HU6Tb~n1Xf`Wbkxb2jdVAbWZw_Kx
z<dJwRFcZiLDs|2!#yZ*XTtO#{+<7Z;w2Bw0Boo*~agYdV3Y+4{3^v123Ae;Zd0M61
zQXH-4u2;|*(eRR^l?)c<YdSrh35IkCYx0}j{v1X~Xp~`;DQG64B0~|5q79>QBsLV|
zNMb0#QLJGsj^a7-T;-ezoCF*tauRWrq?@!r*;A@3#Zj_uvVtz#1joTEj=muxD9emx
z4Jspovd!4mpjLra73!QiiJ$^lfPmyi&_ZsZ9#|8cJ(kn61yq?MWdv>+VKL~|Z%|ae
z(Itd|9b+~(n-7`?k0m@h6BmC>1wrX-x;98S1XW|JX@i=BpbR!c8&t%RW#{x`pp8A)
zOnFG24wwmehxsvEsg-_gAaFMjEJ3$NTV4!Z9*5ho(K(bH1T=$Mj%|FlkSj+ofEVC4
zUc@a@Y<v<P?ca|eL)c8}Y;_$_7v4~sb`hOgXX~7XnT8FR7@RP+!0f9)xqu$5hg+bS
zTdY_>_GYwd(c~Kac?fC;+mXQr5!4>G$B_f<fTQK`avZt9E;w?9U2!CYg$jbQ;#08K
zFZO>wD>fT@T|rRPtoh$Z1RfwLQu+9k@-bTRsDi$#U%S|PxyKASCi16fQ4|!T(4g4V
zxRgW;{o+@ukYMsWiYg7Et25Mmm|h;nOsWq{OM|Vg<KVc=dAW<L+agU1Cj&cE!T)RJ
q@O0-d(9<_GGBGn>XtmgSsjWTVhHJURQ<Ed~T|-<E>h4YOhyDVcd)K7^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/check_err.hpp.8BCADBDE7B4A301B.idx b/.cache/clangd/index/check_err.hpp.8BCADBDE7B4A301B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8eba5263d1990f206a69ee271daa4ca211de9e68
GIT binary patch
literal 11746
zcmeI2i(8XN*1&VdpoaGimq^6CU?75G<RWsBLITzciGWIJ;6t?5Dh4lAsM=kD)~c+R
zpjc^Dz*|eSTNOfG5olGYid(x?LA$QT%RY7yZ3S=Dg>Ki&^1YF7IMY91{T`l2o(D2N
zW-{-bGv~~lGxNTlU@%+_mPp3WNyuGfes`8sB9XX@zf0d;GQSUzNG#40N$$>zTW7s}
zE!1u8Gt=Mxx<>l(zJzO`6G~kFu>7+98Cd@^clv~qvu!&SiCs_TL}l+de^9x1meZ<l
z67nBJ-|382y%~FU^)DfAPyYCZ%ff%Wo$r09a$TP1+J(moieo?gxaHyfQSW7Ddh%PG
z-TQAmbETY*esU^gg>Syc-Mtso?h&I_zc$_3voy2hasDFB{UaZI68NZ2wo@A!_s1P4
zZvHsPTDWW1jr(Oc`VQNWvYR{aC9@P>@1L^y;8=@)_(17T^*;~&dC0z0r$e{49_zQ|
zwDu@B?45gOPWMv|z18>ps#~Y#Pv3jIbJ&OTQ_=$5M!fHP<sH@4#jVMoU6p;le#>Vw
zoB}T03!L91DDS_Sw9Fdv%|Dx#zoTAQKgi|dU)R4fP;bv4cm0|qVp~&mesaCbuS0$@
z|0mW(ZGU(FzS~RV9*-&Ld@c0r)aH);Cnf|8*;l<u3a8GwzBh5INpOyp3U4QlyO?bD
zdeT%Gvwg<H4+ehu#<km?aYuK(nd~{T-<{O&-b+n$8uV55VdMDtZ8bCGZ`_%<IboX1
zHH$24pW^24|2-#8b9L|EM@%>VV!ozM&-`WG#pPXZR9$}Sc+A+={!3DS8M*u3whhZt
zY@FHe$pVkP4{ud9XIwl6Su?xl){j4O>XQ(4=DM;ou9;_T=^iU8s~>&uxA)6$uMd9m
zc=bGI{fey*mOPsK*75<JQ9-i9jgqB#i)LREZ9+2h*x2T9tt&5a>MP+=Bq_4ze<ZR#
zAQ2?Mf!F8<%XSRiw5_|r5s(}K8Q1+Gkr9v#bTsn6dH7rBh&NNa8*L+!ZA8X(e?0HW
z)6r;T@}BmAivw-ljqZZvuGmxe$MZ&m9E}dUw4Gb~&H_nyBOi&6jC1^v$fO`qIT~D2
zA9cDC=zF%i!D^7K7JKRbc;1VbqtOy`*3tnZP&#maF4f7clRgg)$lTEX=-)T-As%KY
zp8F1@-vQSnBzGF5r+>>MPQxgyZvvMlP$D;jOEbf-{NuupB^_6IDW^-*rBh*~oL;LB
zM9NOfPN#~Oayi<ZK%{a=D~EI{rIfRR)(U~hb_lXV2pxcwv%{-)7>8^jX%>=BbuHxz
ziMEggBG;0%T9QsDOv=>}Z5;`uET1bx&9C%6E(#E!2=Krm>ja%QQidYK1BaX~WP2kO
zfuaaJaKPmtl!G^NBPbfd0|#se!4BTY9iZp{4;-+C2o~avY$XaS@t^~ia<xRLCEk?X
zc5Yhz;Oc6B;JJ}uBO~dMl^q*=?Imsfc%BQ>hK107^IW(#oFQ|pc^s95=dxn67%GL>
zQW!_&;JGr0En~R}Ml``lDhJOs!-!^<yCAj;#!)$VZY7CbNf@?~SQ{Bf<>0w05?h7r
z*|@Ug(DkxsgT#r^1;kPTL`n!qV#pMbhBRC;oSDK5T}CVwPReB@WF#?U%1A?+rkKV|
zZxQH<AeIU)<%%Go2$EPf6+s&EMo?^I4WJQpjSx!}Bjp+)p%Ic;HZ?*T@<~vfWDTSP
zbR7^&6({97AfW@2ST=P)8geHnI#~m<5}lRAQpJiMMG~weiDi?Oq#^Gnirr{ngX{<U
zKi&IfnCKB<sxSeS8?Fjx$Q*1AK?P<7XE9U?!KDy_3Mzx(GM1Y_XaX-(Kr;x<EO$Y0
z7lfdID@pK5!my15+eiosP(^~PkcA81ES9wE%`YwI1zQfCP;v9=-Y@C)<CQtnch}b~
zjp7=04t&nR5Zcmb>PP>&?SJcJAD$Cj1W#&fJU7{8vM04JapeG)mEcM3X~>wG!;fA$
z@7!w*lp3z6r|yt}>S_471v`}cR)61X7j5l-)|{vv`prCXqECa@Y0%P%4m`f#%wkp1
zXmO&e!KoVDXa&!;fm0i}QD^7565>=s+^CL%ET13y;c07yz;gle0C%(yol>VlikHVT
zbq2YCA*0;LP=-9i9W5(cnXN*~mFF^bdGb7l^5yvq6@k16+^I$KTsbJqL4{la@(Q-T
z66BRE+dyt(xe?@z;EqbMgVGKv<Q9;(u=O`UeuL#!khikj0rC!TN2OYb(n3_o#YA3A
zn0hObTZubWIxof;%37jAZXogotPe?^V+^ZYwxO?-8>$$}?6~QQbY>r24~q5dzIYrI
z$Ju@E5hxz9`_`vK@hR>@bg-hAbjQrpOCmqM8h<JMBV(`ilE^P5rR)2<e&rCrC(IA6
zINT?kA#=Dn3T-nhJd2@H2rq>ww6`({FJrk0e4D@@?WY-hn_2FH@Ggj=Dt5$X;VTKl
zHWF?lQ8@oqB)keaM*Z;g+kIy}3>Bj>Re%>2K+FLWb(4_vf}RPX^&$FDrcSSqLyA|#
zGgf3!7#K1tj4TBjgYkPAsthlxA}N=hm^}$8N62A<^qi2KP^M0w6Ni+m$YrcHPm#w^
zz9OHc!2DqRUJ<B@z>7|TR9plS%V83-6$C4r2v{M+3ZZPh-U@NZ6`-hK{iPBVl`Pvp
zVf!ry+8`Le-w3KkHjJ`^#twee2ppNyC_99+9B&6B%O*P;m$rbSh4t4Ppt!+uD=1oj
z%Ym&BjQw?hssp^Jkx4lVNwknj$c02GWD}i25>iM)*?N5;i9;?Xx?;k_#a5!S5-)0`
zj>IUjmP|sfBSIY}N{-Agq>hBL_4+yzhulDP4H!4atgG?%X|5U9Pc-W=T?86kc35@<
z8bLXPl|uwhf`x=xNCZ1*zhW{Jxt4_0k_h&pehp+OWkda%`wQ=8ecG4jLgk?Y=tFrf
zQ=U11K9uLSf_y6spbzD_Gax?$1E}VC?it9R!2tSDo~t183NnB`<n_Pp>C=8RZ|chc
z`2`2aCRu#LBD33v;_Paanx1u&=S)gd&pO6)YeBgd*A<?t2W36_HP77#<$d%kp4&l`
zJJ1i(_k`DGDu<=@+U;>_P{)El*9+NV5a})V_N6{BGk0gsTMhNCgy)hYllABfHSlT;
zE@Nho+FbST9)B*1?<Mn6&~NkHV3)y6f-yxlMS)!;%aaEptpeF9P@s1efV=<(B3FZF
zHTa^p)xe+{&>**gXB+sU_qBs(JIf`+vxN9kZ{)dB;#rEkaP#bd2`!1s`tzJwWgd#d
z%2H)9R0^t67>a``163KzU7+fMo^{8u!>d-}I>d7}qO#$-<2Zd$RUzlhSWtL-t&2^J
z*JqEK-Lu^B+*26!6nd67p4(4G?Z@S8x%T(Lx}T13cjmbWb%X|$AL$<%fHYS<SA%lR
z^Pd-hv<=kTK!b8^2mkF5fV_&RR}l>=w}AK;kO1T_i24hnLFFDG{s%|^<#j*Ddq<xu
z5=0{!JbJJmhg1NO1)!skDSCAH#DnoZxBjaE?#A2Ii%nziZ2Etl=*_LigH+WQ_B_wn
zB2GgTHRBy?)?F7CT?i2?4uR((3?4rj_k&)Yk|ck*d%WP?<5P>?_2GiYOZ)MhPOIyQ
zf_W}mo81$YI#N+>IYx!N7<FkaBp4l}m<Yv0ja*B#wIrB22+uVTp@FC=TNfrt6I0gS
z5D#1QN&3muVWgaKjBzYdPEyWf>cdhle@y;ZBr7CYVKQ|sDQAN*HW-V18HCI1d3Y;~
zX@#-W4jl*jNrhxG9k7%uCTYd&fqxxIsw0!>fTf}UX$^SbzxRvj&pw!crL2!5{+(in
z*RW$d%&@~u<QCAhV7$ElPYWe)edaVrbbu`|aSI**SAK1dOeZ%lxr!=@(nq5b^alMz
zq<BqyPckRoc|_$zqe^lNITMj`HMxFBR*14fG^)o61}jWNt^iF1_#qdPs6rBr>M10K
zLNXD#m?RdnQFk4Qsw2^;t~z3<BNLGuNMZv<(YDYPG0|jjqNs(D!pPxNyaU=F)22U6
z36hG2AQ+h}FJH)KvMw74HZ}ul1)&wEx+XxDR@`K56Z>w5G3_v(4!ZVEey!*FM-{#H
zgzqcE*4!^%{pqr}e`x$POrmS@Gck$hTJT%TCV=(eR}XP$koUpwJ}?Q;4&t{16Barb
z;sJ@{#*iM59C7J?{?yc)lV7gSFL-^f|I5XUD&0SC^;)YvUYM{5d<stV+RWkGw(To-
z-5&kDxGEWf3_aJJQZ6qjkGY<zgrG|HI_d@l-M|~>GfVcTJZW9D^rip2VE;MmHo^YH
z^=`~d#lK+1H~YP3DTp(<DtK<VFkDUd6`s?Z^gXY1@SM?T?74X7xoN^QH4R#LF2|J9
z^MVJ@<s0*RuG)ESBM2KoO#>EQe8IzH#TP!rOcabZjIMa@BnT%#O?N?_y9`4v!!YDl
zFt*}#I?r{2&<RXHTS!cW7$5Rtyf~wn%%ZUc&+R6{ZVYgF@mSDQhw-8#0G`o6X3>D?
zoPOB*uP1w#^x9RODqai|y5Ab2fr35j{PKmlD<(Pei5Q&$r#(9++kjJB4l(6mpuITq
zi5LqppgN05U@?1jP)lNJiGixmq0YbtygInN#y+U;?}SVL)y(=D)IN^x+xdf1JOlFb
z>Iqs!iuPjU^wxThMyUt;4rX$yDPB{CqYkD-Phm8Y?2{aXb*sQ@6%0rHtb*uOtiB4w
zEfs>WehZA+f<d3-K2%!`qtU#p!B`Ec$cG^K5X2zYz=#@5(;fGx+BO)C^U(&zHb_No
z2W>lBUqZAcWHioC2{D$CROC{kEyenlSI&+9mu(RATJyhk%RKMk{}gtKNk%CdQ%WXK
zN!!Qnx!2yg`FgL7cJ6)se%Je#m;K>+f+}o58+@p_w*7F>8b9$AE;1sLIpm5=h)h9)
zpBFKYIYph9FfRp-dOJjH2R$0?c1YL`DaZvRqJZep$O=e80ZBnVKq3whJykW&9Uut@
zNDAdgI}aZl`6rEkuQjS)v!u5qx)o;}cS9N-ljdh@vYAxA95m%@s%{|~3-O~`a-3u+
ziV5@PUoFwp5<jXX$4Q2w0WZY=eNQ&VJ@w#P$C<1!n^~Qng76fE({3GSvci5coOUBw
zVqQ4Q@#5>!C7DYPISMY0Kixk{TIhaZ`u~W>|H`rGip73d<nsI~ThcyCuXa(&UEEyV
h6$3o_$(;N15(T#Lz~TTb_Q#?Oi*8sH-`b`>{|i~2JeB|e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/ck.hpp.5BA5ECB9E73F13E2.idx b/.cache/clangd/index/ck.hpp.5BA5ECB9E73F13E2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2fa7b44ac04ab1410f344f5b3f71575442143e1d
GIT binary patch
literal 2384
zcmYk82~d+q6vw~!5y+B*gv%&K@Ip{PnPM#-14=+eE=5{LY8_FdB2tu~f>LmbLcK*%
zDus$I76C!Qf>Dr5Q3qmGIvNi=sz4p85k)KIX!=63Np~iBllS}YzPGz?-y}S6?%es7
z3?o||m>4Thh~h8|V~rn0LVPq27-j~`Fp0T&`$J-`cnZ=77boA}{I>bN)k!<v>(LqK
zWRt7S)nBZ#PQB^$)%kL5ZG!Wu>@%K+b{2fRJFcRpKQilRX@k;pL*nzZj%?V=SNGP)
zN_sDkZK->2zceWDUbJBC!rY9-OM1p76+P*S4ODO(@>_Da59_yExb28{yMAgDH$O0H
z{Q0oXqq_qf=KH^2mDLe<{Ic>?*ws51<_Vow1y45%4w#&D)vQNj=GWq(mR-HE)9Xn<
z?eDk$7Q3&jYB}vJb4xf@@9!J=%CXvSP?EU6i{CM5-k!X&!BU}(m^Dl!`PnbEdCMEg
zttD}BN?Dbqa36P_vozXaR_>!|)~`3HL(`(WZEuKLZB;&sq}Y`icp@QTpGj^6e7{J<
z@Pv%G8#vv-H5i8A$>CU7b35W%$$qsxJIsm^_rZuha5NYykv~ekyY?H`OdFPik^ztn
zK!CweN%XizyU{|cLOyFRV8rRHMLKI^Fbv)p4iH1)%=OKwV^pI=j2ULr1DqbrMsMgk
zY1!VRRFIM?V$4zU7C3J~XfRarT?gSms&H?&U@b9JFK~Oo)nFJ)7Pv#9_AW5n`TC&N
zo6p*zXcud-ixnD-E+5+=QORE(3pO=nOGJzX273pbcNk1>Xs~lXy4A@qWmd5)e`psB
zBYp+ES0FI#is|qFrEW^wvG?W-!9W@iXu#QEOdDt7<!cn777Sw+zkTNPE1kX|z+~)q
zu~M3+`Q^0^-S@IefhQhGM!FbcF&3K8YlT(vVsw`SDF;sqxJ9=$M=EXyBX|Jh0gR=<
zGKSB8rg}LKftZNLP`9x1be*gD#>*IZ05}7fzhMlEs`Nd6?&Mt5rvXU=cM2p^4u1%l
zI#7YYjkt}ZZig<FdmhX#K=)>{*{F8*Sx34BYgXX;>8yP^E2RbD(UJ1J>Vz~8^(8<`
zz@IK83-5D2+SH>(&<LavJSYg5N#^92dMXgy0&>f=wFSOWTh>Lr)gZ_Qk_}FDMv&8&
z!WhTkOazf6QbOId_*tHz<mp`r#9COcg(zCr^a-F)U;nhkbi&D6@S?&9Kh+4CtMWJo
z+5=<{jH96<+wGQK^s<gXkOL$KyeU|b=p%0^xL1gv7D%lLtW5E+e3;d4jWKPBt(58o
zVtcN=@vv9NNDCeh9I3?FzQlJt714Cit$&O(UwY^W^iPnMZw%WJ?5Zw*iJ%imCyb^`
z#jZ%GpWvY8A*caT12ZW2+PeR`&ysB2^;ZI^gsBuHwC=Om_B6u|K|YXtQ%P=BHrFO!
zT-1vop2Qn3XX~HiA5WaMEezdBBxw>|hV}w<>pK9wob7MCnp|ez)KzvXkXG=eGrqkP
z?zX4xLJET6WVrE+(kAxR^pHoL=*|L?WvV0EN)YJ#2#r8b3xU3S&{Lw15rMvAP;l@;
zlDClg^)&(skr=mjXzbpuGX?FsjHE<rEQS;2E_Xa?xk>0Q1X2h-bO#03mKPT|&HopH
z5{MF}P(T|I0)6|Tsg@}=o{!_n{SnjysW-h&<(g+Ie_rz+L+@EQ0LG%PYks_ay>Iag
zoKX#=+H@S(hl_JtD#|4YGJs^jC>p$ZLfFOi!7ANDj3&{>g;XIyM7=b0C%O*+Ibf>w
z-&T9=|3u#5jo>bjyD*I|bgMD;%31N}{~!<$!7%D>i!cu#p80VGx>Z0_5NZUa+kQ~b
zsThf%2uP9XreA3KqzgXquOVmx(q!6oXU)4QdFq@U2#SFeo0>`I;D(o5cW2&2PzIz7
zX3`j4{Z+x9Tj%K-XeN+MaHT*av`cqssncC)8;~|rm(e&K_U>Njv=%`eiF2g-zE3AN
zaDF=5iteXCo|^RZsyIAD+owgN>$DRGz>mIGT{E%Lg3JGZ%e1hJ{vPBM3o2}?be-}!
zkmsQLXEQXkCoUXFej3Cg$ODpR+FZPRLzGUTh+m}`swLwThOyDl)qRW0L`#-UJT_^Q
LnaIYP5t+b$iF=!D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/cluster_descriptor.hpp.B962ADBD27415A8A.idx b/.cache/clangd/index/cluster_descriptor.hpp.B962ADBD27415A8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e78adc52af9d1dac6c45faedb6cef36e108f7c64
GIT binary patch
literal 766
zcmWIYbaVU0#K7R3;#rZKT9U}Zz`!5~#Kk2=nNNZA14aghiaA>+IC`@f^0Yo*{zbuv
zySsF~YwKQaxj&3c%jR7+$_={RW2E=5JU#qqv)osqL)9n0d+hr6W`ps(m7Tj-dxd6M
z`OBZZz3g>;`L?Y1)r$qoOC3UMR$cd+##AJE>Z*pK^P=s%{>CQ`_bpg``mOk>8NYvY
z)ol*lEuGxg-EMmB?4h~ydtS;VH(iWkKX+l_&VyTjtl{3jAj(}&SJ&*VrFh;u#e;q;
zT9zpPG~vHccZn%0=552XY#pJQUtXCW+<xq<(v(m3k@GKc{XRS=GO25mOuNx!-h$VU
z{olM%yD|TJ>HCSQo7)(QD|3?+fxb1>dw3^;$6y8z0|yHOD~~dd6_{Y)WMX7sW?*Dx
zWGG5aE1n1p6SlA)=Qb>1lV;>$6;c**f@$P6=$Mr?)$le@Sky(-3og8M#eNOeg1tat
z9tIu`piN*P(3ZSn|F2teOgyaI+T3tMcv$(A_$*<1VKgHb*ex68Y41<nHF2B!<vG@~
zqg;!bd06E*<b+_VfIN9QO(dR!9NZP5Ssg|NeFu)Bxw+emx#eyDe+!_n0H1&;Oz)bC
z-K+t*i`;<1(i+kRFyVD~<{VyNav&QhEXysc4HMpx_TY>Q*Qtj<VIEE%VVDnhzSgbw
zm9d`=@?kW0JWMsvhe3Q{Fy6_(vaG@3Yixmfk;C$;{+rvILlY(gg+*jV;5L2#y?%}O
ztK43ou(-6iKFknc%;Y43tSv4|E>;FHSXYVcc=+@RmnsuG8v`3i1OnJ$3<d@O=4}n$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/column_to_image_f32.cpp.B78257C7C44FE524.idx b/.cache/clangd/index/column_to_image_f32.cpp.B78257C7C44FE524.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa0f8ec93224f22f4cafbe04c4cee6c8434dabf4
GIT binary patch
literal 4954
zcmZvf30M=?7Jz3U!~}*wNCFW;f`%OehJ6bG14tsf42vyKWf8F~4G3iutVLX)Ziq;w
zSW#Nk3Mlv#TSbd23P=$JvF?h+4Mi5)r@o8ros4~t?_=`)=bU@yo^$Ux_fCX7ARvN;
zAhJYxZd%fcB@_ffnBYHaMP^ErI)c2KfgrgxM<R#Phun1g-?J0WhJ2G|Fe<w@aPW{k
zHF3d%aBhXLY>xfHe-#cky|{9C=Q+8+WxzY|UQLI4jS-Hpk;cX&GWuhG$4{>f^(4of
z>U!L?ys$WXrDtQK#oJ4)JAGOO4${Fn)w^^;R$VG%-KjtG`kh#P()Iq6j77OGpS@5z
z7w-t(f8foEf#mP@p9~r9|IfsT(xKqkfyU?Dys{||!|XGysWrO~(=Ejf*@MTm-)M?P
z;=*>byIbw<%w2F~c80vFY;UV=CC5XazxMr~k|PUqU;mW6W-t#8ZE$YSJ37lK&p&Fv
zp_5(U+SL3Ph3^_KzKGxR)WK-zaa&&KkuhpZ#bN8YCUIjoJYt&bQz^IRxmYH9#uQGC
zzF|E$H4$P#Yc=f+Fk@_MDSQyxy)k&Ju8Z{ss%ig@*S-M}Pvnae7jC`sSbTVq+h#i@
z^SklxV^gw?|K=q`mJSI0cKNOF=2>-bY!Q!KE!FM1d%^3{OMLVEoyfr!&-)RpLME$w
zJN3A)Z*EE+UG$1^?J-ATwu6&ex9Izu8)ZHLYfi0MyKu|)&7U0%v&r$g>wfm}y~cX=
zbazKPIfGWTI&SABWO<6QL2JjtZLji-)vh+YW{$EiG{ws0426>WF7V6)(XiF&&kNjI
zZ}t|oC&=!5O$zMnO*}3{XK#Ad)-*B^wV)y{_UNHz-;AuBxPf_gi-Vg3vxE5&r_7Ak
zv)!(EbuUYbewpCB=;^w|g<Y(^FO7cT{Frc3)E2n*=gh7F&DWu`ZE{O$ZXDo6+}S-|
zP$clOuZy0)DXqe6>%+u}+rvc{PduahSA@o|FVxy~e4XZ<(@n#D;`XcxM)CZKo}-xl
zjrE)b=Z`J<GX8AV#U&|8k-S>Y9nED49e>!KlVn|Vw5b$qx0)(^Yj3YOX`B-~VV|EY
zc~_k|T-ugU%x)IlnzYy>%JhpY;zf31C4C&JdeXa=n@uIT2SUz8^@_W{ZX6l&4Q)*e
zf91FB7F}IXFnppVb9d@sL)5zX6>Dw28AaCaSw6m<R`n?Qo+P?_;_{xIOAD{*yS7HH
z=5}4lTrTtDZ-`%>t{jfA{eFTf^nO%bT|brZcJouM`sh=~%9<P8Z%Hq`I$HJYz4Sac
zxj=h<*0|PJo%*-Y7k4{@vhvd65fnkFE@>UtW4JLGvH(GFwg#K?;Q{LJ!=LrD5rcjD
zC%9T_VL(lz(ik5eQ0?ct?%n;Qy?Zbu5~#PLl&z@I-#^eqKB#@xZbx#vmRcw<!KfJT
z!vm_<^;jJVw5&XhAu&K*k5cMU+TTCWq;A7o?@zpq^R?6>fC-gKWqx=-b+6BUs`-#b
zgO*x2P-{>%bU!?xdTMmvT^GLexRzQ7P}8Y&&W8t7XC^Jp@?;<gdi}(XH*d!$Pz+Bu
zO}Bt)Q5S3N+jTj<oPyy}zElXs817@?W2+*$liXcJ3U7rB5~X@9fAp^TV&Jz1W!0cU
zm@;Ufruk9AF}x1N>QEkG*4_T(sdcj^81FCicY>p6Pfv$+T-EFY;wqF@h3Z3l%y8+D
zOG(8|It3Tg#B^v+!5wH0ByyxVlE{hXL?UOJGl^ViE+lfLxsu3@=0+lSnma+dZTvLH
z%D^5qFl`<G*>D;d9-tdQWRcq?&cE2;dldvxh8mQiv!EHDGuT*Rw)q4~!7EWwCF%{M
zGt2wBHmq+l8I1T#0hiJIR%FR_WaCz#*nqM(pa#&DrD3h^o__IX>J;45z;h-PQ*bY?
z7cnggF5yZ{pw6lTwAiT?7r@+kPCU}O3Y-K4!NQ<uepC<`&J;17;hx8EVUREq`o-{^
zIXP~SY}b9?e)~W7_p4)gvUV~PiZQ$dWtN~E71yHLwI~y2!1nj?n>(WVmVrPG1O`OL
zTsCwgcGvQhKy2?|?*d1A@LzU*MqTofM!}^FsV)>#a36*biIVk_*(#lqp(N>)=1NPF
zPO9`(Q3<MFf^t<{i|W;)Y&ccvv;OCe4YfI7PC<G>hS1D6B>IPQjBsNhUX7ZpMjfF-
z=5_CC@Zggcpb&G!Hc$-RfkOGK!N4`gK%cO#M}1)n@n=6{?q;dygYh&sngp&;z|U)*
zr#@Y}3y3r589LDIJmz@7z*yu`APzSO7eg^zI(5yD+7vKy2oHot7#_)vw1yOX?n-#S
zX!L>_1(!4AMo>(_6$}LrQuv@K@ca)C)xqny0xq#`c=q#0y|3SF1AC3lV-qoB_%@WX
z4Q0XD<9>C^6N+x=gU!MYV&l;1bQN$aQ9d;LXSu}PFNry_G-E69Zlr5O%mu^wx_p9`
zJ#U`t^yly6z{ri`W~TC8EH93Ql#-d16FT*S4=|Eb<l0b-;R=d^AZ4z153lprxBpYN
zxgB>O<WqXP=ooHBH)E(qo72rnWI?wektN-dL{@Yw5?Rx&Nn}H}A(1WJmP8`Dh@gTc
zx00GPjs}7ts!^S4)JV0HcA+}ER9uj_!S!~EdlfL`@p!}r#c)TCBhmGX60dt+9WKe7
zHj3bVstQUFAt2~W`q8FOF3ZQjvI}Pl=R>2_2ZPe^Z+p*!u~Dv30Z@$L(QeW6Abk~4
z<vsRrMiww~<+zgeMv=T|5j0ve!zkwB34<u`f?WGt4=9E;folHaFkL@Xm%d-RI~drL
z8%3Eyz2w=Z_iox4ZpbhsbswWyMnw5zxG}?+q~p)xlXPY*Gor>Z+?-`jA`6xUi7Z)`
zB(h>zk;s~5O(GkX4T)@7wj>g<L<DW5`_DJX5`G1u%E5AozP_pYhQ<99pQAvGNw7#*
z+gr-Mt_v&JHeHM498^v&j$5t6>j&a=hjee4%I}_9hjkcC{sP3ItWe^WyHn>=6yc@!
zsGzl)O3k4)Xsse2dljYVrI1^zLK&pem7@BkAGc4>T2umy8D4v4Xnu6v0Vj-%83IGW
za4#({edrRyC0Y^^DNGf_R6zH0&+t=O<^<$~L#lhhv3f8f`~%!_g7`ti`8<6ri^@<3
zXolhCsC_v~TtcQ<T8<J<_jE0J(k{$?1Tuy5aN-WNH`Z0Z^WrQYAfBr?*BE-+yDp~s
zVfINL5C@nBI79J|?M(50gYfr2JZt7G5fnFBJ;FbE8Ey-jtkO_PG%rA6B~hTwx4plp
z=7uf<6(jeM2g9+TOZh3HAsza3z@e^OM-TK;DP2l5FL0%!`;bUUR}yrj|FF1zdCy)j
zr~EnjUeHq;<M*D@&zB=WT#g!+qeSA`7FN)!Oz`zUk;TX&BK~P#Y|E3Wx9`E51v~*U
zYYcbbI1p_G!wcnwi{UFy;%>LYh0Y<sQxq?XsJ8Y)zje#DAZvh_?n0NrJMJ01EA<C|
z`}rn_-<{)5>`4q?h00f<3!&2s`W8PPvOW_J%$%4`#DRt3&U(&7W;=~?@6mG^8X6S5
zP*O-7eH479WF?7;Bt;}DmK2j{6)IVU5`5WtRV*?V+6LyFq?<(Ks#{pFZngIU9V%E3
zTj?A)N3a~uKE!Pc!&3}X_$r+;LK&r^Qj}kc5=D>UwWzEXod@gairFRc>64lY@Lqj^
zK5_EG&jy<2N1qFzyj+>Ss`B#G@+5UUe`kNG>b(Ij0X`~{3*{meWrSuBhbL%C9yvsn
z^c}xH@u--Jfb6?)T&z?>R?1h#LAs&e{w`l?k`Ii6OoJ?;c)B7T%TQOyL)&S&lj*vb
zKy2xu>BIux4evqr{uwJ{fM`sZHmYcN5+1Q0U=vCNBr@Zf5hV-yH_x0z7CZ|QS@JAN
zWX-cCkqys=M7BIz5{Y;s66FYU>|h-Kytqt%ELqf`4sZ=e?C4*ZyyDk@bZD?OxNz+-
zoXVzhVWWIf(Kzv~P0ub+UY-<BP3Rl6NQxJUBoqljQ%`Pfv3s<5Hpm|{iMbRe3f2=i
zJ*Phy;Wgpysc!RcymD)RHMun*nMtWjrfsq^6S54zuR8vDm@K6&<FO2lLZaYk1W8Cs
zOic!;ND5M?Xi#U+=vtVzt_~xR={MJ$sTrtez+!V}y19FLNxY>#b{rcIp^K&9|4H?Y
oc*Z8PW}BK>SXtZJ&k;L1yYih3MGi8Ket<k$%|F<89y%ZSFLUO&^8f$<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.030C990214BE859F.idx b/.cache/clangd/index/common.hpp.030C990214BE859F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d6e43323fee7cd70d84e74c82306d1c4f74a2ab6
GIT binary patch
literal 2272
zcmYjQ3pA8z82)CAF*E)dM*5i>V}?PlTbNjilw7iAs2PSfDWOog6kTk#&_=PkZ5yRU
z+eK1GmrbSHb}TJP6xwQ8R<V(-UES7xGxg0s|2gxX_j%{}-uoLK6cF%`27v$apsbYm
zjHMI+02}_&Gt!dc$N-qq0mv%5qDWcQ>uTH;_Oc+gzPdco>$Y+Awpzif5c~Qkd7{N5
z;tl?#<twTuHz)jPm2|tpF~l#Y;d7t1^z~<x*JX`ulHA=NT&-#>TWHWGKEm63qJP=h
zjI9GPkGykJ4XKmj3Hz?j1fLj#QRR_@IVBHPST#MS`^kU165PzpEid}8Ima`uz)i8-
zwv*Zx^xBNEV}s>ecEeD_v&wc}R%!L6UtRW(^qm?gecf?-a8NZhlNr(3bL0HQ3y1S!
zXOFgBZGX1u){EM9pSg=<QR_~p<s^qE4(I#S1=02;t#!T18|IWVCitZoyY%_>rsWm>
z%>ILGT|^Ia`-7i#$5i{>^{0ehn^^UoZ|sGP1w%vT)!&D_Zq4=|3Fx~+>Gr!}|2+Pd
zsS~X=dkTL{;!rt#op-*c_3_JR3yTg6KaFcz8Ryq3%&y}f+Ea*6yjdR=(Zep{t6UH8
zk^&1d7JgAhOM2VwJmC^rG})a0Wp5XFFoQcuPi(Niyvu3H`MlxGwF7^gyVO!Ye(-J1
zmOaYDiwf`V+dQ7DZ*aAyHf<ND%c{XOy3BenevqL!N{I-3>fSWlE$LqUy2hxOjV2wV
zPLX&_x0ki!&fXctslIL2RrRY6?d*KCEk<$9qa^c2<j{FiNZzEc;HH8ZKU&suA<%1$
zmD4|W&dGa=mzBCt{I|Ee4?8n<e@r#pF=t67Ht;I>Rs=2Yg6F9-vv<}et{Hz8gS{Nt
zvc;mMP8yzrr+4e`Fj-LXh-G{=^Hq6AjdI=xmr&Otg&<{R4=1RRd*J7}jl!Ef`<)-u
zR7AVyW|Elin49%uAE#%h#O0F!pp9IY-Mr7}z<?b9xEJQd)d)b>BLNnhgp(F!z8GwO
z^=1#1Yyb&11~#}x0Jx#>gM`><rRmCsEQKK%ns5<S#Lx(WkQZL%SW_IffeEOPV2NAW
zX@rI_#Rz&R4itwpG?-Rmz#bB&W7D}B0qA5ha4>;{swgSm{EL+@i)>2+xE;-Iibeo<
zLF%uL@;c>?d()^P8H4Y6R32L+hzx#G?WXS+F8YQ=o(kXF(rwK(0=@6uN6Y#*)wqR4
zo(BK6r`wxq1p51Tc^jyD7v;|chS0shTHvA)8uzBgFumu_^!37Vi!wtp0}{kkF+(GW
z_!GmAzjYpr6&aG5kiemG3^W4J_0@+DJ*362NFUZCra@wSoPwr20GJ6}Y0^plH4+7v
zvZThSUuMU4<!(9=Q$)gWsYp5pnH{$uZG7^+{{%F<(cPHH>}Bz?c@=f(OFati#&Xj-
zq2TT;cOB%x^3Wo$pc7whDLWkvJw^LOhoW5xm6oo?or+4JnKzT?kJ^tmMUU}uLp~*s
z;ody&Y3MD62eX5P2ra%hGh6J`gF{a~jy~?lyyW=t^C!bS6AUmsOcJK^Rwj|@Ah|@Y
zgA@{l4pK^#Iw(>Sse@FXskEWc_QH;<SNBuS0PQR^s}~lT3&!U0J===7FkcZ{WQHCW
z2J3D6t9WuI8N&s9ffy|ehO10e98?sV1DWx0yS43!rN!ciwEK{aFUR)z-!OnsB%F`Z
zs5myh{r1vrXJI;yG)E3f=SXr0C0L%{r{0Ts()lO?Iu|$y^{t?}V7Sak#zm1ZTxq1#
zqMED*eOby60WdCy#lg|SqS3;*>a#*|F9z(*s`dTzp#cRCa|m-mE-ARoL8gP`4sso&
za8T$VrGrujMLI<4AeD#83;9H+3u{r|Em2FDQJ4iHVYtFbp+&9YjoZJT6=4Kxk%mg;
zs4M8op(pi;6Q1Noa%~?lJklsqi_jop`RcPr<)eFtX7#ZnGin+%t4|!6-?M4nO-^Qg
z>X8NsWoV=zEk1dvnv<TkEWJgon5)IdPlOZzED2RRcL2-?tAJ4XgeoCap;jg1vn=YI
rDO*dP7>liKBxdFuE{`v;uoPN}fQ6O?tO=Dzs9Zv|B2*hf6%*?J1?l54

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.0CF2F8C8A70630AF.idx b/.cache/clangd/index/common.hpp.0CF2F8C8A70630AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3125fa18e50ad19e7480ba82b54c3a150d5073a3
GIT binary patch
literal 482
zcmWIYbaT7K$iU#7;#rZKT9U}Zz`!5~#Kk2=nQTDud8`Z!6?67Z4$Na#<Z&rp6cqaR
zRqBMpeodmyD-$Q@`Nwa&s5SA|^=Td_6mlvSZ`@<O;lWu;anplGy@W&Ub}tg2vD{}<
z$U40Yt*A`>59$$Z7BvA{Oy7UKX@2(k^b`K+C-<M9B3&PpeD?L?B|kN%ZThX+Qaquu
ziF^6Je=ire#a!JZb|duJ#xr?`W-@P^b74`j^vPxg#qGJ@#n<ha`lnTAkNKvjaR<v%
zEe=dz{=@IPFnib1>uu}a$}CCF)43z&UKH|J>~ex%2ZuWU=8JRkW}n(4^=k38Z8vQe
zvOT`^%g|}pbN(xVCo~rFrpK|LP*>TSYVi8p?c)>w2v*LW_VA3HmHR96XWW@`T1$#6
zbCZC9!cdf&Rt%wY5<zrvQF8HY5Q7IqKmahnV1O6KU|?WIFqsfcb_A0R!Q?_PIpItu
nMuR&c8ZK|IuHa!}W?^IJ;^yJy<X~k0MHm!tz!(e++z2KBYY~{!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.116380F0FD4C88BB.idx b/.cache/clangd/index/common.hpp.116380F0FD4C88BB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ccefb22335e3d55c02b46b748c11bb6c781f9bcd
GIT binary patch
literal 4426
zcmai1c|a3a6W<LXkg!0w4+9D;0SR(Q02h#;awu{X5aduUMWA4{2`CV-N^3<RqD2uA
zDk^xOq833Z%BQXM!YX19kA4W%DzsWl)mp98zFinL{r>s#$1uN{_ud>kZ{`s_FF1I%
z0fx<4I<H`PMqU~h!!QIfMV>s{QUk;G5g1lbe<p5u<Xt!6ACZ@};&d0dl^E3R;UCxQ
z-Q4@pP<K+RBK+{{@Q#=v;oFk!Cw9B;jc#w9yDELx_3Mh&9oNOXs*-&U9uuE`ut9gk
ztTuJjlRMa4@twnw$45r*L~Kg_e0XN9pQ%mu!NR?!FZ<G-8|GKX)p{paT^4lIHf38h
z3$pEN_YJRp<G=EHexJO5lW~^mgB4lJ<{26NsyIGitoictWq#;BozEL^$%{2b&j`-9
zM`pDjzgk^$;H`;O&41slaM;;;vO4_ey|b&@^&O<$9n~h@cYW<nR8PBRRekgYXUZn2
zg<o6ko6|K#Lj_wGok_m7;r`WigO;DHcv;+`{c>k+@1>&ow_k+YJ$7ARwT&?Fjqi$A
z#-9}4zhs%|o3J-&_`5L4X!(Or7ASHpE)R4lX7MT_jwE_|r60c?86*E|?X&kL7cKt1
zzqHy{bn36!4_D49i>fO=v%x%P)W%Wx?Kc<K>)Pd*7!2R|_P3uG3$|=J7O}Ccd}&Si
zxt`<3r$76ovbl6?bYi^j32ndovmy?*MGjbaFZ#NaSP=gtci@!fE$(SNXmCf){<W*F
zCA24mUAPcwakIJH@T=VBVH=6&0^yyywl6wAXx<upVEUXUe$891@h|VV|8n;G!NyKW
zHZNwQ>HNHvH<optZ2s`@ANgJTA5$@5AMFuUnG~;GU*R76v~<qzgCE~aO@EddIsI&O
z#^h!8A)o$ryG2uV=}U7z^UA+dy!a=wsUMuLtb2SwzNgaTO6+@QD%bnFRXGNPnHPMJ
zWliji&2LTD+@}&>?a?VSX>@xyIcvkA)=6brj~0n?bnm>ry{=Q3U{#x*@ydVG#-CDK
zG&;{^FK+5?>=f=dc-oAOoVeNCe`jvb!1i0aKG$$IzR1<bi+|9vIZvc`Ds3B&DyPh?
ztki0KxvuP4!LN8)sOn_bEsfnpqsN9iE7SD14!>F7hb>4n@2~JTDc@PY>UHL$OtYSi
z85eF>KhkRpj~6G}swxZSE=lv)^NM@)*0X5`%Tpho^_@3VIhf-meDuETn^zUXo2wo3
z+j{0@i!O4zG&A4L2>u(pabZ`MqHuY7IJzPDR8@3V;FF(u91*5NQd*RVA(%G0Md+uA
z4VILZ_MW2E1nIBkFJcI&`D-J^)PCo4v|2z)wWK13fI95yrQ9z!D_&`EY{!nnNOys|
zF2)~2vH8C~yz%_iRvyO)!^lXjNK=M@I&Qg7yY-s5E&`j0)X7A$B||X2mIh`$1&d$Q
zeDBYkkz5^)1ybAcY_%BzyBB{|o@I1y&MqB}Ia1s4?Dz};b!yT=<9E9X;&eDBNNvTl
z5-<eR%kl>VLEbZ@Ivj&Z7&%oi)qo)oSl;NHoTeWI8af<9q;?WG88QUa%6A+OwWOz*
z5ttu3yaLB6aO?3u7@x<%%qC-%D-Bz{+XHXuaD-@yh$j*-1ngJMwVU77WxSWb{E<42
zh_hq}#^-XeSvFYB>nncNW>l>XM<3aD6gV0%1nf5@c$;3iY!jfz@ioTCXzgethJgBT
zsD0_GnbKk%juqN15l9Re0_yg<;x>!tai1YIvK`I~*JTLC*JX;LGqIY#eM;;`J&qkZ
zJc=J>&Jfsrdc#S}*G+%)qST<>p}bH%hQRLbX#Tsd%7p^tIZ_Am0(BVz>aPk@#AW*e
zZxdJmQg6dGx8b(qe=xo#PRD+?M!VqUd%_+)jttq4;l~&=1nl<@I4rp;-Z6l_96BL{
z7oy7$Ffh>ARhQb?*~sHqBDGjcY|Idhug4VS&o2DcQ+~Iu*Ws8VwKdOLz!3N>FH=K1
zo^CDDqw9U_N4jdcir{cep0P}khz^2V#Sx6%rt!tZzYyFlw1~b~<7>vR86db}Xc2wA
zKur7#>1JAbXCD%s#kpjXUXl=s#~`*ZYh@F03@1pLhl~PyW(PK{D{fg-iNr~^NwZ<K
z_3Flei#YiZiK}p76>bbw>@?WCYUF%~9)}=z;M^U!fOVoC=hm}W@=LNXIwW8XI)RFz
z;vosnW&8ef@6G)dB#!6D>w-P!;TJ#D#@)<9Vig{&!smiLY5u}%Z<lB8LJDcHG!!!6
zV?A;uhnF^@!6oDM;wOUSk_mbVYA8`JaRT|wYx!Vp{m~?J(wdoT!l9uARzDw#k{k6S
zvCKj?8S)Zj{KR0lkwy^`Q@&IHxDQQCmp|6L#-m5hc}Xd74xw4nKs6L5jZ{MmrHL$}
zyeJ9|N2X>|L24+1ic&*~)M6INjAdrvEAre%MyeqfnY$YDmH7i&5Mk&rw8c*oMWw>|
zDqJ7Z4BZ!17TtQ;kD4|q%?VNke{sQB?;T&YaIZ)E5)*MODv3H=ScjWT+*`5%7dEgM
znl#!^cW5vY|5<zQ%-YbW$SJ9>R0g?$o{d)14I5T3?we6)H|>rovX<(c>Ips(sHJ*U
z;4%m(rJ#OA>Y3B`(V-#(kvVkOl!FI54n<R$S_CN<%O@fuNQGFThVsODYN$|LsD_Hf
zMQTVXR;r;A_mT;ZV5lJf>4^gV!!SW&dXhj3TJe+pX!rZ24RjKsq_L2M_Y3^5i>~yI
zqRXtptyH)J#0NtM9YaqXI0n>6JT_^-ta53&_q}XtJ37A#x2?h{Sh?Z1AQ5%F(h@1G
z46Mw+>xxi~kDgS$oq^I^Vpw9qrdfp>tMEy%8}f?m(PIm&!L>%^M2{;F6Xb`u_J_DB
zoCcE%?WV^T1hIRG;}@A;9?KYC1FG<Fu(msQJ~uS#;~(kP7L*hMR`$N|`68mSJdY_r
zejTo_UV!`t7B^A;*~*b&cXTk7LM?&Y)D+wDc*TQ1AN>FQ<D*nVz>HA1G^l#?6_yYb
zK^Ey3O`ztxPt&KH4~l7%AygQcgsF=nr$;VC4$~Hi=}`-LI~BaGpf%vpEiEo7m&jeg
z5|>m+6ly3>lBb3WC539JNK&MRloF*HD)B1ufiq#CqY&w7&W0%0ms`Ukf-KZ8oIqVF
z{ZH2lH+M7Xv#G-+Z2If)fI2)BFuVp(bo5gIqJvie64Orr5W_Ac9vl6jxMb>Z__)*g
zbY!mqw?Y3b!AcqjG_dx*{`h-dw{mGb<Ct3=E>&OX*1+OxcFM9+{}_Grk&?obqM)vN
z?R(tMbqhj~ScQA4aBrx?oAdedTEFxMNFfu;oT0nIV+7@neu=QT%hS)BC<ua7>MJJ@
zJWx=O^y>sdgnHl##vUZ&a?~3frJm3mP(?XXo={Q96`srpTzOROdh+UO*EZzg7F=fw
zJ{eqxHw`jKKWM<<W2bD_+})WbbVLENAUNh(-=N@bRkSt|OBYE~z%smiH2GsMo(cB?
zEdpKOa}s2TMTiRoNRXkHp{`H?1Q~7-?gG_5diI*Q&aa?egCNUst#Y=z{IBx_s)VWA
z@R{%(<r%qY9>_SPYvQRjA%H+81P}tIkS|df;b^lD<N(9GSaurAI<Txg%Zgc6!m>^*
zOR=ms%T8xm7nYsDvUEJ-zp21r9Ma!{Zxhr%y2I3Gx`))$%X<c8FLrQrl1y`UadmT_
Zin&g##oSrek!77(){SL7Sysxj{{ToKTX+Bf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.30E7C524D1E7DCC8.idx b/.cache/clangd/index/common.hpp.30E7C524D1E7DCC8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b586e6cdbff4463124a4c74ea37afd065f383c3
GIT binary patch
literal 3458
zcmYjT30M=?7M=_kvO`z`1cFQ=VF@6FeXWWF0u3kvMg<p^Dg^`~$qPar2r5{xpg^%e
z^;s8CK&iN3QP5f!(5Dt(U8`1GMXFR$OKXezZU*m6=KDC`JtzM;_uR9RAeYJ3vJoUA
zL$1xuRIM;Z5QGc<3RMMJy(ofc=m?^1J(--HeaFl4$N2He)rXzBnh!UukkLnt80nIX
z?%Fw9@^ZK*>oWQljh5a&V&R-MJbO`7dHVU0rFRQeDvuU>htNM3WTkzl^W0uMc>0>7
z)wDkc$GwF~)vayEY8{7Mx{e?H#O>9uJ*y7pz7|^b_?~TNWPPS>+`BC)EdPqSMd^KG
zvNT*+sa2(%%&6GgG`WG&b1!kSsw7uDJR`RLw5WPer|rKTIdDHZOcNS&;M@Gx5U-So
zg1i*lf$=xP8?D|kGnRIuBTrA&N<E{N?`!&L$tkLbOk-0#=y{8J^x<*guD*9p+xzRX
zJC68O-I9Ke$+BMF==RmI1r;c_%X#ycyMmY0-8<FL<hHfz$j$!qTQ^try`sz(+_mfu
zTr`@q+CfYEh?#fsgk$=l=46XzWk*%C>Bjq&X&c*UV&#=iqbqg}>!t0x%N~q%&tH`k
z`_aR+<idf(5RK#2xHo-*@<-yoI%pGfs|H)5A|s_2-Ch*9okd$>rn&Ap+|6Ib)t;;9
zZ13^Yo?|b3<!EG;9Afv^TuRHjfd9CJ7K~Y5J#Bk&&2^6S`ozfY%gVb20Tua9am<l-
zjdS@shgj{~9*pl`<o3lEE*#$QKCx4mQNL%)#arDkz8f+%>#5mybkoFOWqc*oqEFJ1
zS*9K;-qzTv{e7f`5q#l{WvC4$C#rsy;^~Xzg=6x#o{^!zwyzQI^q=4M7x@Qwbz>UO
z472r3oB&C3*__z$c~Sp-x~}(Mmm^Dk*th?Fv;DaLiMN5`FDkN{i>j}E8?BpNh{Xx}
z&C&1c-<7S+7I{=GyYbG~=D}=l@i%)-4%}ZL^la#i@hxh9`?8#IRuT7)b=%6erl&hI
z&n+4|e&!ytq2s5uwoJwDE!-w|p1<g(OK$}2EAPo>`u9bXLs@qkN9*cc)FXELIxk(A
z{#dvESYc6a+9;@B+OyO1&is#YjfHq42)3NHoR1TPZbU)&_7sd_Nl8qw73|ScQ5FEW
znz-_Dg3#$y&;|f7JvsEC@8H5dCThz@Fn5MK4<`r>{0zRx%`eA4Jus<cqUJ!~nrcnQ
z2?F{Cd_e!>=NEF$)D^xlLZ{;*fbwE^u|EKi1@goOO=?6u73F|mq$W}!P7pe6+LV5h
z`uE2_ynfR_Ma=*p)+CmP6Y_$ZeDebVN9NvSq6`4wQF#oUkbv4~QEXqI<t`Q?1pp;W
z$;SyoM^VIAK*1<m6WR_=JUN!mL@fZohH68{32C9(ZldvgvD+awVhR9GJSR7t-~b0Z
zU}5*70N&6bHj9ILaS<$z7H5GI7^FQ}8||LWe3HvVc|c!Im7C*)d}O!rAGw$OE<9(V
z)&L-)inur-J?M>+RS#8&2AHT707OtDxHus#9A8{dZ*ol9#6cZ_2MW5v1}D&iOQAW_
zdns4DK^B3&j4Ct7390`>>DJ#$QV-{V3qYDB42cy^p#GgNXxcwln_c9fKEQlDJ>C*0
z@aLavPVB&#)!9tc5}4;x`81r6{y!2FgvTc5&1RxZ01!|GbexcYvHg-WPeK$EAVlDP
z4~7R1C-8#5p5EKy`FOdIiP`{tjEb4y1fiSgCpHD6AO)G96y63k3oD+&n5y&-MaN#r
z56J4UKSeRdQaPz+uqL=k!>jUlv$`oPEL<7B6jmeb3Y0O`4IhlKFdwRqEd*J&f4;6{
ze@YF|O%<ljfhJ+Y09}2zKvOJC7A;GJd5}C^YL*Z+>)${(Oc|C6O}ba@9k@UF%YGnU
zOkd1~CVkZlFTc-SxdVs;m4PV`9qceZlucbRZe)xt<}4=OWQ;B0EHU6xI4J}dEZ<*R
z)_QOWuvA6~E2Ai2xWbpZx4Zs$s|5TY7K-Jt;^)@Celown;vNwDEB%T1L|Tq<cUl;&
z1!6y?AK^Y61|atON$??RNX(Z5wU;ZNv5i=muhcgHx(^;B4K(0FrJ)1|rv`;SH7!1v
zQUq+x<9XE7fqO-Pq6pYF(4sM8YIqRoiV#Eu!qx!^aF7MT1h6o<gFFZ}CLA&#XZi^P
z9cN)_c4=-fM9M=)u7+)kc@8f3Rr(THNF5F=5_#O_0dcx}x*tR@)3)X<KY8L8Am(v-
zHgHNU+qds%M?%m_6Bects|edHOf6O$a2m13fGZXk8*nA!5(7>r)){c6-lc>G=`V*B
zYKvt)zyn(oThN$OU*039xBY9%zXIY=Whn8C{J`gH7F<(VgCzMl_;|pOvakS$00S<>
zA%x&6UqspmU8&&%OV(U#Ync7&7^C%LweJIgxYWGV76t@H02GIQNJDpEjf2qZM>7<&
zuniR21_}>q!x;^<^#dBhV4VW7ek?=r&YW1|nE4wY>bpHMH45GV$FvaxAI~uBAo!+g
zKD>#Am6(<g9IOy<fqrho3-;W3o+h<<AO>#?QU(zX(B_5ik~Xqt0dxW_0gS4tuT$GM
zZC>uzGJwW77!M9JSiS(QANde{*qNSG_urkP0L?e$Tfo`cD{Isqm^E_K#28aaR5PKw
z#+X{7HsCZ8jR99IDK_9rBqau%PNFm5O8rZNp-tE_z%l)*fsVl)1Hz|YFl6{trYZ-h
z%fgCHiwW**{BK>4Iu7N4x7eH56Ez6C5L}?&Meu^lqT;IZ@CBv-ovKKk57P#B6DSe=
zVj?GML{vlq^z~XSy<l5t+OI%2SQ$)As_SaS-H#{Pr+`=-D^7s*0{0nsihh}qPtlla
zh}^R<ovDuCU><<a`UMAlhMNtD^=pj~KgwM7P4cGq;9h~5zzU+^ngdVK?>O=)@@RP?
ztTk94z@&cDL6ZfUIV<$JfJ(&wPDo)vdSQ<q<_@t4GLt-dl1EqaID<S2$)lJ&Iul0}
v`^-7u`r5zCLpU=%JcXhePR?!;cUKoNLUbI0hzL01L>^tpql7$qkjMW6vD`i3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.3324DE766DB5A26E.idx b/.cache/clangd/index/common.hpp.3324DE766DB5A26E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b10444bfb8d103dbc3bb728f2e8c2efc34e34584
GIT binary patch
literal 8422
zcmb7Jdt6jS``>3+cVRiODoTev;BwJ5JX{u(n{KJQ0;@(T1fpVILQ#X)UNJ-Mr=Tw&
zYT^|uH<1dj+^i%J0}*d%sR*wk7ZZ$1@scE=Qu>=+9cJD?<MVO$eCN5$%roCJ^E}7M
ziD6+kTqP1i{KS-``OD@?B@)R1{4aUg(nUk8BoY-Tk))Wuk9qH{X6>Nvw`%2qNeAXM
zxsM8UIFyoPbTEv#^7@RT<lV8|lS}KbxIODFGR*!q@!b;l(<@Bx<tBfWbXKEYe!r=_
zM04tmO<l(~N2crGmTv6n-hiICga-~M;#O8YH+~j)<<zF$$d>QE{=DK%pAq-JX`MCU
zjNO$VFD)7G`gwqH<c|3@X_quBKN&llD_-+3^S{4dopF7`;g~-KLt%o$ntv@#-<j`u
zGw1YO{}%O<FQz3#_PAtkJG>(JO&7k%tzuNs=ARxf%Q@QbnN|FBZ`3;f<|>y2?=6Qu
z&R>_mvExW~z?{=4EU;nb^?ko9N-yg)!(X}Qv2^j#Qdw1!OK<-#M|DKS?iri@=lHbs
zSGVtIsI(e?^~~JAg+)F^Q@h)bHM`9o*SW85n$T|d$(+BPq;2_=ckGJ2)#dQ%hqj@$
zCs#c!`^By0?HgC;hdzz3`|luC?BVt2K3qBNHNTTH6U#pR;O2_1wNv0-&%Gbsn0&Et
zvDZ7XW7{f?Ioj{K5~6y_b$?teIC{5d&t20wdF+{==XO1BHALO>%sOD4d!0Y-Jn3Po
zUH)2A<c#Td-ztvZ9KZSCHx9pV5{mYXmyIY`9{7723PelBO)cwa^4!~M6*Di!Wl+I%
zb<M=;Wrbgyaup*DrhdIUlYIwo3p;R3G9oPH@8Mr`Z_Ud;e7mM`%8?&@gx9~?l^(w6
zBP)ktn_HU?Pjari-m8}FR4SIwd{v{5R%R-0oM`mg_gVL`h}ZrazexM<XYmeBbEbC$
zoS$T`>#2{HC!ee7ujAi%nlU_WPfpvD*u>iKrk;3HWYMI~`2~-1KA6yN-l-G4Nu31+
zzb)v^ux`F;Ds;DNdU`wO=!h2o1yR4=+YxJ{`SD6XM5lZAVDI$Cjb(+h^{ejLcV7*g
zA+P#*^|g@djq^6Gy0pJ|p>N%GdC~0gY0}%ye=J?QEN)(1vQy34VbQ1ma&d{wEAw2`
z`~LGe&lg@l<`cX-<^KDHnH}3auAEz6uZbyj{VY4xZm6;#c27Umlzs1Blr7$EbxEdM
zWIyJxwg~-IY%@!%vzND*^lpvb>s+5Vbn2ZaU%9?3+IHF2*RIS?UXe7f7r!OecdMt=
zG$`shi9sR}7juhMmPTUJ4@umHfe3@(?MNq={1h4Lj}?66J}OHiv9Y(tFB?`+$LNYj
z)lRFhMXm!RqMyRg(b7oFV*EGKI%eInmDASQNqB58I*LOqjl>4aUiHBCZ#+1cShFSh
zi52p+7+|enMX;Sk0e3`Ae0#_8S%$Y{Xdn>>$^+dcHg<B1v3>Lh_kXE-^kjz&4Z^0;
z@@Pj(^U@wG#Kz_S^!UApiMB|<3a&C&YfD2FW}V)!(4pFpYl~D^!A<67V`-?uoSD-G
zEv`t3u|<wpL68XwOG6drI!&B&L%S9qIpeTG4mjt4-^(9~jT9tKL-9P844e7oorQ-#
zHm^h>j+c*jw=^VB{~q?%hN8OtF`Pt?eOv(N3*i6qhx%A{xv*vMsI*Q_G9D`xBj;k|
z|MG__ZF+xTQP9x|jkZXMl~gj7!qU*`W*a>wS5yr8l9Nop3iG&mDoZ1=v9gkcY4B9H
zx36lDolLVfPt`^oBOl{#X~<N+njYkMzSeD=J$ikxM2xkLb+9x<oWCeTxa<2{k}Yz<
z?xSQ;11$~RvJz=`(hmXW9@wHGSV1Gx46rnGf@R@h8_%!vsIo=QSRq0dF~HJL59Kqq
z+w5|U&ay`%@C>HeOyeyLnZXabM0FWlF2S6{_ETh223i_wU$<u4osVXHi#g$pNvl<8
z2U{9qUsv*W`O(PBXYC{o*#1?=SBF>{%LIqJ;{Y|kbA0#46X9N*WGq%lL(XYP`|^kG
zLvxOGO3%OgU$94^c#_j?rVp|-WRN>PHXpin4KdlGSFjIfnX|&u&?wwHB-Z>kCgM*V
zM?6V?g}<YvA@+|uZ*2AXWsbWoQe*p}vY~QIBeCIRFW!v5VwVMYZE^Y5yo<JCxIEn5
z(hz&e()kOLlkr<A9afY5e*^y$017eAIu3Wd_>+G6$I9Q^G;^e!_z5_E0-8RSi$8+n
zkD%#ethf$2u0xu>pom{1$FGrwkd>S<*uVLJA^z<WR6&alJD&ilm|&mKpOBIqI8w!n
zgXX0=C+cK6Cqhh4)XVfv<c;D)qs-_;mix+q&1oMUirWO7nBbnECw_USp78I_KN^6$
zJtufhCR?nwPYcPi{-F)8@*F77fiqp@Gf+MQXS&J)q%1(rbd_yL*@m3SD#<yBS7|;3
z$uiuIEnRpwRhJ=^V*iSBU`PB*#b=;+26n{1R4hP>0%S-0OT{*%XhU|yzt{SGRX_c8
z>kyuQyiL3vS(exB?v_tt>KEE@q6vnZz?1mr#QnhU2NlKN0skF~Gtux&<VnWk#4Sj>
z1^H9_InsX4;!-rc6nT>QaN-HXpFk>#FC+dk#Xgx~Z*2BF=!)lSeBBsEc8A;tI85ft
zfRJJIkca2rlg)i-B^BfC<NL@U73bT}XV3!s1$}7r#6xRRO~o_u02|=74G=+iX3(`a
zL|pIuroTBPp(dC~Ly!wXF8EP=5cq@OL-8Tt53#rqghE!o5cGu*OZ7`YC}HL0z?ZZ3
z$ACY^;tJp^SX>E0C2L;^`byS+6{xEqfR0xULN%*j4f<+UzXpUFDxWYWzEWPBcGkQF
zKE@!j1?0w6*u1!D=^LifxzIOEH(WP@K>@m8iu5YI3w`_auj)kx1?fjqWK<bl$iR3S
zo<=VQg&1`NMFeeI;^@6;3!bYcSYsep#MJBF#<Tl7tcIBPGSCFUG!}b6*aI#U=R#yI
zyiIW)h<V^m@d4luu=Fngp@6k70^cGUJ5DTt$P(7y5fG2C{>p$aWBq*(!uPDdDu}6K
z_vR${o@D*kKx7T;{}hO)SpRhpSO>x6c;Li(h^mKZiW|Vt0HG9L1>q{VP}~B13#ce=
zgYY((NbxOD-vT#^JHXJv`uhchUs!*4fxpZ8`vcT}u=d>$-p%@d4C=?M{0SJIu>PNd
z@Ra(G^xKyoIAX)kc>nd@`T#;8lKx3;L`&M&$JGdzh+vW?aN?99Q$on4&xuh(qC!X#
z5^W#S)OG9ro7k_;OE;2Gm}HGe@giI7^P4CnjvxmK$sF-nWlkH(Sw-?i+-*)4`?{|L
zuki40a_q;6NnS|-#0iHJ!HGE#Nr`v?CS8Ci8u&D1OhZ%Yxm%1T6{9F}uD<j0uM;KQ
z2Nie$p6;G{a&OH}F)f+%{qbA4TQxw%lbmn%!GooTBSRA8oVZ%Ex^E?%xJI*vL8+ig
z1qP>sCY{9@pvhoyCTKEQoCSVa(6>aAALD?SlVcJPvSQr*A|)mQvbi|;<}{debN_=#
ze6aP4WSFRtz+l0P1egex^G0L;$PEk3TRmDgj%@Yv$rFAMV*{t)ZgH$Qm0Uop;x`2S
zv?yekBPV8}&`cCZH((pmZbRV|Z%5kgEY3yRTo&&|+Py5!L)ttRm!i;86i4RGiRDNr
zNB$I_MEptQOK}b2Ygl{+31?XSGstiT&7k_VNT_AyR}g=NwXa8fJ&PL<-@xK)NVvw@
zUqgm#tp7%&ZbX4}ye1?xvHDHO(8TJuAfbiIGk*_t4!N{Rh2t*_5MCi<**MASKlieq
z2jlK67?=f4<b~Pz<JLjjop$|%1$wn!L&!oHj<?-h(vVFjc@XB3xsV~<oR|&P+2BZ=
z5;9;lb2&rRSd=$u@aONGx^BcH>(sho#3nCsvUJkajCOOP8k!JJMkG~`mBLc}kP`*J
zYRWY#GA<`(fHDL6Fgcd+D$IounMeMnJKxQ-pKEVkvrcayYu^9y>Zrl@+~NdItb^fo
z;7MKz^KBs1)6<g^n}Kg;Pv@H;++^jQz<09phafy;<(Vij69rSwXCWaAxlp_nX}7Y;
z;7&AhCkmkQJxIHUm4Aswe#y#9QD7+wrW`6mLK$+QxEw{5qiBjNkXV7dDXvC*HBBx#
z@iY=nv-W3^?^)zW^=na7E$i<*63?^#>JVSY`n!sRtE|69G_4U$r~aCeZ!_z^1x2;6
z{#%jQO8pn4RP>vY^hFrn&!P519Z9K1vKqYm=Bk(Mep%Sgl%jFJHp->qYH_tM;f_>X
zBd%dkDu}7T;B*kvS)2iwY>Ye;#7q`vL0}e)Cf>>AiRWT2u*qCVWsO;FE~_c4S1VT!
zA+{VISeXuexZ>dSZ#p0U{T!dAbr4zyaV$?URrM4%fY1Q`6t@E33ceKI0R9GxZ-RJ}
zwQmD;8w67Qb`aWGf9+sshZ$7f0YV2WzX$w1*8hFr@3Z&;@DEu05X6V<_+6mxV#n`-
z@Gf|Vj`s+JN9=fy!0?EjPdBK$S^FLkdRYA)F!ZqcPe6D=<yCJUl$9PCc^vb~s5TBK
z$#czw`|(_Nb-GT9v*ej3h$C08RK(>@F8EW7%bkPZOEE5Y4zU=QJB6%1E_VuHI@QPJ
zP6;c=<xV+kkIS87EXL(d1&eXHQ_0%na;K8@kIS7Z2&CiTa;KWr$K_5ntB=c_8di?U
zR|m7men`#as|N$cn0yT|l44B0R?t$6$#;Xrn0z-`drZDI2%`F!eC@11OulxQN#&S)
z9jqLa?;h(PlkYx@G5H>_7?bZIJ3b~~7dt*CUl+`#<6-hWV#mYed&JHMldqe#$K>l_
z^)dN+Sba>sC#)PNN||UDIrpU^PL#GGBgHsT+RkE}DCM#kCrW!+j1#3i7UM*z6wM;%
zt`z4={^e*S#W+zqiL?~sM5%_wI8i#o>f=P|44O&xaiUbq%5kD}g|)|tQay`tqSU}*
zoG4vm?Qx=XjrETcrA8D)$HR$I6RVFCr6yJ%CrT|;PHGr@D4V}pl0*5T$Ex*fUyoAa
z*FGU4VJgW$E)TolS6!)?WGRk$6I+btf<6}}kk^RAm9UV<7N`4w-^c0~fw~CXsD26P
zOIZ6OARJ-sOMx$C?W-WH3XIhLB&bia{%Sy9!}>b~!YS5Y9q@H*deQ`Z6N~Qve}~1H
zh|feSLOY!8A#FCBewdKf#NtxKm$KJ=IWm-^P_j)N&NYMz_WG|vd=;yI7OBr7H>zKY
z47IHNc_f@??Jptz5^LXx!W+><YTu01&8)u`WN2aiwIZRF`XisKF)hsBs>!!SQVZg4
z^EYb377`WQ{o>1XpWx*OX?(Yvzedx$Jr$IxG)Kh+5hyb$zIiKk+|~gl&6s^+lsJnN
zP~`hH9>)CfnhespxF|ZrL4TU##ND9W4IY#f`Jl{a@ljA7W$|TDUIq_xmT}^BP+n*8
zZBX83aRyRmAP=%!GYOR0EH)vfiN!~e@+k74n^1|Al`Ou9lou%`pWpHN%-`Ng&`B<b
zyUpL+$+F2+j0+6&mv++4{SW&4#6`_#x)c-sGMNMN91v)1{%Nc(fcydoG*<sKc4<hS
zh6EbBe;UhTBriq+jpaX!E&jiU#JMAJkVsS|q~z)AAk|l22Z^vGdFi{!x6A?dCQ+71
zUZan}^f7=wYU!gFeH=v}&9iy=7f2o?5`X&WOCQJ3M;(3i>3c-tRQ-y?eznoR4ty<S
i^cbC&w~w#ZZ{(=Jpx^+1Nl2fb#E(7((Z|vB@&5rkeKh?5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.725393A22E55A928.idx b/.cache/clangd/index/common.hpp.725393A22E55A928.idx
new file mode 100755
index 0000000000000000000000000000000000000000..de3a34a8e9bfe381c3279530ba33253e2274d56f
GIT binary patch
literal 3338
zcmYjS2~<;88h$)NNXR1#5W?nVO^}3j3I-gGqFK}g2!<q8OxRIci8=uVhC0<2DYUjC
z6*|X-r8+7FDKH~~TU+QbY8|QCBRFmyYN<20QH#qIPv^eZ-n^G{@}2*?_rLG{_rL%5
zU*ig`7Wc$3Rq=`{dr3tB3&SuG{FhfaUQ^;2Hs_9ERr}AF?S>oi!l#Dw?sZu6-j+}H
zhSjS5mzNeUoj$v!XKP0H|1#Dcv8T;d?*F=HM?;)K<8VB^^`iRW8<KZ|=B3U$N{UAN
zgkuNpf7{viUQ^K)M^oYxOL^<!_I;)CMYEji9dFut8!gASHr|O|Vp^F}^WeToe7&Re
z`zHx0IYpKIF%y|jLklB+#_s>wn%MX5>ek8T-r|H)ys?k&P9B?TvaYDk`_IwSJ?HNY
zC?m!*KdfwjG+vz<Fm~g~qM`MzkM2eeMSgb9Rq$^^Lyuh7FaGgqs{Q2y-{SO9QFCoq
zbI`H7yZ$<R=)~TejSYHH!y8Kv6_uwYncq_OHFjQU%+~4GZEM-C2jA@f;|Z+zro-kg
zKDuf2)aTxd_qpVJ?bC9z%CjQA*v)TbaL>i=58D5rytGCC=od-u>JA0hbn(jbh<46`
z<l;-N1D#igUo{O|x$SoT7jI`-yZ&))&aGRqZ|Zp0<f0cLN&Jbyma-1K%(|!{Wp=>Q
zcerWUrNRdV66-P9xxU+@d0oQ8+w$Fd-*kwh#@`=``fqFDsvqP!;VbO=nyG5L-`dZU
zyPsV&pA^2j?B~ttovQ{qCvT@*Rm|LbTsQh+dEcQO!XpW`JuZ{4eQLNow7hFu;NZrW
zD+3a*tLu%!74EF&re)u*b+7kvjdpuJ=*^a|nR6>WePy?&yQy>o?_YrxV<$)cIjVCE
zq&>VEDVaaMLnYR$_LvhkuZunq(OPuUR=Yd;Wu<Gwh3;upF7}h-ek)d9Z7;-KF^ni^
z>#y`VxRsCldt)G881F{`#(^~ke-i14rEEjpmwY?`PA3QxBov_O;5y$z`@yj~HtvOC
zAksb3ivo;GEp`Den)iR4dA(<D6%Y4-4q@ytE(NrMu7=~1e@%6bkNd$5;)HR26rc^#
z+mx!{jF~VV?grNfvx7wx(Cf{cA5LBV>BSy44kHGM?uot>&<?Atf4TE~{Y)8;R1t_^
zM{p>h9juo}Hk|Fz-s0g<N+5_G#HE0Cu>JkrQtvNRZ9JTS4pO$1O9AbWt5My(rA|7-
z$0hI#vBFph1?UW0v;oU4M{zwL7s6?|P%fgtIbFMC+}6_@!pHe=S|L=3C_vLYk_uIM
zZwC$&SUlW46-(t)fN^mbEHDVBc;_as9_4ZM1P^CJhfsE?2L*I^O~xaiKD~FMh==o_
zLpVE}O9AcBy8U}u5r3qghx<T>Kz1OP0@~rwYNhXm^Py>cJdTS2B}eH^0Sd8~HEX|^
zbnX!khn)!kHV{!jukYsz+y2&qeZ$9n;SpnmF%k;U`k|Z;xgUj?8~FG<`1u9g1%4Eu
z>ER@4=i{kfB_9{S)ltGI5d~=aMhmCv*#_ZPd^`cJPUa^2P=KbNPv71hJCzSBNDPDX
zXZyQTfN@!{`Z1V4N6FgqFJbCj9W}pC8JSw_LIA0!)Cb)<5(w@=2?JF;0nB3aJajoo
zCeVr02<>DMTY``i8ih`zQYeW)AV6wl>hF+Kbp7Ga?#zAn35+7gGe?Am7lN4!>)8b}
zEEdS|%JD(NEU?CF4TIUdYz&s`mFr|kh%O$#3h=kxk^Epxr1nsV1TcwB&eI_g!nLFV
zQQ#TMrj5QaI~!rBAQ`Ao6(B{zPm*#(-td@9Q&*zEkt*RQNl79nq(wMP3K0#XyMV)_
z{LnDc9o&f&9UXkOINO=^ysG^h^Us|A4n`R&3iU&3$?NPK=rX3P;}bwFQ_E3Y1klJd
z45pQ78B8bBF_>PaXRvfxI)i0`%rN?!24G;YERe-uMyWB1_RN-MGnh$gVlcDR%wQI&
z#mP!CmYmt%ceoIqF~T##8zmPt9+)4p>(Kczhz-7I&n>^TALvB(9@;5OoaJ;v&hP@r
z4ume?oyVG>l(C%waNjUdm;}|>dpg(J2fKbr<`5u?6D2_UC4ihG7oeaBpx`J3Xm<j{
zaAE|gPY@uM6DvU1OMo~|oB&lS0pdCF0;Ea;BybW0NPg&YVVGneL}8)^2#3k;hlWwx
zgTrLcL&F_qdX{cw-6Scs<eY`*=qzCMvO49&0yd>Bg`Ugz%6HD8Zt3blbxEWLSXu&2
z09rx-y+H3|-BqXDvg{vgVIuwc{=TSZAxXgq$*zhbL}`LoM)p$r%8X*8v+f7A%~d^V
zKTPmhK&?>6p?z6EqtGyzR-t7uokGW8dWD|B(iQ0pmKmNINq^H2ZeXyi@GJ&1%8jwK
zXSO_>!Ax=!gPG-K2D8X5PKI6xm{Rh2pi`<Ar~&#o0dxYLlbua87Q1Oao`xw6lm^D2
z%SOcrb4R`;bnY_6na*;*6kb#R)}m|y3pU}j95uaxHNfiJfB?3|HWfXWACT{yLvIp{
zmwb@uc(nqp^U?{BE=YH>X-@Ypd|@xyO@pLC@u(9%Gfk+M9Y63291cznR-smmDg;&r
z`J|!BaFmo4kaoGFsQfF^h$Z3I-+W|XSQKrB(q=GiN@-I@o6)oxLz^K^69<3tvAk~H
pbTCaAD_2Cv1c!!%g-1wbk@KUl2&We&r_FHM45Q6R+Emcy{{hyK#(4k$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.7922E75661986F8F.idx b/.cache/clangd/index/common.hpp.7922E75661986F8F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6f9762ba2354e7418aba806494cd9193af57a24e
GIT binary patch
literal 2048
zcmYjR4NOy46n;GEgU2h<e@ffO`)SGuO-F-36<KtVA3NeSz#4Q+R;6=n0}MsQWvH8O
zMnpw6MwpHX&N(y01vT>zr@}&v(+z{-a5!Ot1ST2Azvzs+@0sI$O?q;_)35h@=bU>^
zt~V>|Q5ZqYEB02E1}gob1VL!gH&|Iw5(puP0_3RL-I>?>T9?~!asNMOS9BR3Xo?Me
z=jTqWtd3gzVgF~}dFNj)edE71yV_T^>x*m~lWmiovu-!vG7Y{}T}vm1YZE?Y@9#bS
zWOr-o`1{jp8r)wwZ#NBVbUphP-2C>vn#rRXho?d>mX-gm|6*iX+hRxc)}momPF_n=
zUw_Y3NZeoUSs9;)c9t2htr$vQ;=Wz~=Xk<}-d#6Z9mBU0YV1pscZH1{**CBuu=L$4
zn{uy(Pw!oB=x(Hry`6jNr^SX->K&SnIoiU>5&G3aPt`r9Ysq}x{C8pFhG0+bXg6i;
zROK`+>sWox-=4FVebc)~RiM%_zXTe{(agi`!R43cCMSO~BX6iPdUwj#d29L}tMW5T
ziZ_>BKl^e+blshb`YR{yuDJZesk(FhGXu@N?x|-(4a)*e2iM<y*wT_X5G9nYvkVN5
zlsvEv&b;ufF6yE6#>z*{m9I44E8-JRo_|bw`oGyi1lN`>JBX@HamTjLYw_?5snHWw
z17)CPL(r%r1QB8liFmx{#>SkKJcf)y1Pf)+$_5i6tDDs=CpJ16G7=F4O3=s#6SRl6
zSXQ-<cQGV|2wF-@%LWt7Be_q-`}=Dm2^tYt6Kj`^Ofcw>h8Ww#m)pHLM;TI$2qGnF
zWP@**^JMDI_oHhMF{Bm|%#=AyHUvElr6^@_WO&^H`sbUO3>kq4JjH8dgKxOBDzo1}
z|FMuErz3)cazx37phL;$H;nuF6RzYtGZ<2Z2u8|C%Ld;tHR*_7pSpg3BoU4XHj_<|
z4bI|Gc5!3rnV5ctWDtR)IE`%Z4No<`j=}QoQHBgd1RbTLWrGPsMc|#FG#}74IzsN&
zwsTMRv<=PX)mF~V#X%f|;yEWqd_3=jI;gEetRO(^5V8w#3KTCmF%n}%0Xm?zit(Zo
z5WJ1Z2>DGYWTbQ1I56ks@)gL(`7z?t`D`qT_wxA)<m3Gq3F$&M7A1Ivd<F6eevHI)
zF&iWz!Co<6fqbGL5d0$~LVmFzAKoj<N_wo|M&)#IZmgV3;!+gI!=+)wyLdNN&L{CH
z3gqF_FcMsX8+4#HBnc@B<Pp*^5?!Jjl%uSY#1sYch-rXenow5KD1m(V8M&na0yj(!
za!X?a?ggIua+9mJE<|m$%(TpdM(4k-dueQEx(c}+(GC`dW_3=+ug-$hMWotl%dq)i
z64X{UgY|))kE@SuI(u{IIQp-xz*YpL&HrZFX8l%cL2j)^Yl7mpRh?8VDs9a|Zu4}r
z31mXykX&hqLGJ0~++tN`>jYvO90nJR2k1j=X`+FBweieYf^PP<P}EXA$12*U=UIq=
znxPk1Mfvq23pE^ZtL}wGMQ=sn9Cn8b!i-)#(7ErU1q+dzEnxkSIBX78guF381qvR^
z35wvcynuuAc0LXx(Jsb8@QLuY9c0>%QfrJ3qZ@9AdZJvV-2+kJe&m)m4!B{}Ah)z<
zz+Dmez%T9nV8!xao^&ut=K04A2ZAu-W6X;qX+7Q1Gj_e{Y9nQ|n4--wgyqF=L^MAB
E4*)ZvWdHyG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.98AE17ADFF3CA289.idx b/.cache/clangd/index/common.hpp.98AE17ADFF3CA289.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d274b8774d3d01495c14e57a951eced240434e9
GIT binary patch
literal 608
zcmWIYbaRVfVqkDi@vO*AElFfyU|<jg;^LB`Om{{GhO=x83>9;ZPC1`<Sb^u*ZzJn%
z*H+CfyVI@Ubti$<z-USCw=c!B{>?6%Ic?iy3x(Ui>i$jmnB>iPpYcGk%ZH<8+b-}K
zy$|g1HREaJvgn)HeV4^V*6R>M#Rfz6mhD?JTYt}qSakE<?2RTxQ9tL1*H2mY#Oc1m
z4!`qrTeOxPO0rmd!)wlhv$iHXl-nJ$o|Kp`=9PZ&Il|<ZVe+^8vo$OO-p;$)op5=#
za<x}>>=w)5toJPMwO7u6$kwv5KeexFL2ra)p62rj55!m}+r8nx?B^SQdAeg=fO7in
zbBkZ|blJ=(mRjC1F}1?Ldun>`Gkt-1-lmtU43pnT?#N7<=We9C?B(g&<f3&7a#7Ex
zO|=wRbNad|M{(=E6QRFT|1M-VdG3`v#n`z~C{*hDgkRhtuVge$H8LLaaNgzDc%sL(
zJykbMH0I|~@hP1XG8u|1bCZC<%21S=Rt%wY5<zrvQF8Hh5Q7&)Kmag0V1OIOU|`@t
zFu4#+P<%lE6XP^fwHYom?l1@n2=em@F|jeTuyS#5a`W)AGYG*{L1Gum1I8^3u)-J&
Z44eoiGlI#EVDclFJP4)$g2{(q0sxT#upR&a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.A2FB78C7F0E24A0E.idx b/.cache/clangd/index/common.hpp.A2FB78C7F0E24A0E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..99c952d679ad1acf08affec1b0f3d2f0769891dc
GIT binary patch
literal 3378
zcmYjS3s6&68ou1bkdW)mgQo!wk0d-26W--H8V#@{C}08!K?K4}RIn9X#X(zb5k=iv
zt-@+iDhf)M0^)lWcL3L_b?vrw5M`~-g6Nhig>hA8x^>UJ?a95F$v6LZ&i~)@-~asI
ze@%weRMlJz)0Y}*tIF2fxfq5?;J;>l^(r%tVIT7`thS{;XL;smxZA7Dn;iXCmyY(u
z=K4>z4%Bxnt$29($%Ej=oJXaJ{afEmwbZZIEV<HkcK?x@pGrP>bZLZ)AF2EC>8*}E
zJ4{Kh`}$8V(yBn;ae2$3^HD2j_3xF;Ek6MQMql@EJl5TPuw(YQPx6riXJYSod~y3#
zin>*Fb3l5z`OAS<eIiNSednSb<89M9OR8U52V2^ITCkvS|LzM>&ErO6<;J9rsn6F`
zZ@ITHI5y+cjTfiB+LUql;Kr4l>PIho#+_#k@9%R4W38%-f9<+)_}*l~+W$Q-FHKGj
z4G~z6TzN5m;rsUQrVcER-0*DQjjxgu-r4@Si?Vi17<!<g#grG%IlC)==2XYs%k#L`
zHTQO(3Ut3y>lr_DM!*#w9&GXM*>;q)39g50%SMGu^o^@3&U>xQR875YUTizIMv%Rv
zcZ)O&_g+zSM_(~?vc0FXzEi3jjcap0cG0r;%T3N*F5i#*r{_@gxeOQc*NV^5|B>AF
zW7|J_Io5>FCWO|LRWnXUOL|v@-d~cEtC>}lUs;{kKIdT4x+ahJj}xn7$fWth)6Hfl
z7k<&{&2O)%!iNuD>U`0(Irq9sC3fWnR318+*u&c++^wm}(ROIhT`n0czTdHH`O|Bn
znP+bruFodxQ_=%IcD*_(`EK0@x{qoaha<WkUnROvFFV`#=M(kC9Pai3=f~w7{!6C`
zeDedoJY@N=b1E$MTX%M3SFBBc;CHm;XiLvK!~4(J9)q6z6Zy^BQ`z&{x{S36<|m5j
zq{owi#ILQ3=NU}pHoqBMw_b4`tJzdljB_xIu($SYbUVI<#OJwUAX*&lMFR#~j`~So
z-d;0iIW839b1)3}^Zeatzy$QJ6L9iW4p}PKHe4fd8H|q+$9U3!;{6&timQ%))4;=}
zFkbDfmeRn)r~GMVdUf|NxkB6pI>>o)0S%0Uv5xOlG^`&caW8m4lsL+Z26RBqA74y-
z`*QXG4~H3pIOjMI8kqRp{Pst0)=XClai}m5!VBTkz&PZ8JGAy{cj`kS?hGA#dA<S~
z7>5;q-&Nr{pl=o81at`Hg$ih3912qO;}4T#`bpdqrWh%X^rQi$xFJ=RW;unYleide
zYsFd#4UX;lMfX;8_xh7K3Ae+<VG<h9_V$=!ec?ymg9H`<PuF2Ok_JqGJ7L~F@CNq0
zC+&_smpmfGdC(z{7wAF*liq&Q$(~owE|dsyA#_mi6apF;hojq`1eK6OeL~y~I{5H>
z1T-)XCv34Ew{8aLNIYDCfkb|yD-9^cUhaRZZpYl165?=b0Y9Ffga#(Qj}*6_ZNl!7
zxCea22yuib4JiIn-lu|Q|C~k=_l1v75F~iffVKx?l$}3M^v0672u5qf8VL<(d$fsP
z`+BYTYZ8xy(Mf_NHyY6Po5>&cX(x&Tg}6IRPR5ft(|`%Mu=_Dszv{B`nrpChoa(y7
z_@Rk~P6P-gLS50VBZ1&9v@lRZ6Ts$f^F^0~Wa8xVn>=WBU>FhyjHQ&p#OAr@`Jh;|
z3Y{o@{2q!CKo_iwL{0%a>)swRHYs3o{)zqubl&`(9e;cD%h(rixnORA1bHWb>04`f
z_GvB`6i5r)&@vYkN())cF152*v9#F1kSgK(P?AJBAuYmGDIucF36Sfa>!`)j*!}N%
zX!d*pbMi~{OGV9rgbHIRm7+(&*IcYyhB8EAh38UAMNUYfaGBC4T1Hm{mnl`EWh6&<
z5~W2ZORKxpQT4*wmUoN#uZ+Pg10{i8NJNF5o#+2zieE_*AT20Oh4LbRA;`dD#vmh$
zWdvogSY}Wri)966v6u;%0-0x;ftkg!K{kt7l$KD&)2g(xSgtac#cWC&i{&Zv9IR~V
zqW*2CJBr~uh7ciAR9rNVV11}z#MH;^ZuUTDZuosFK__Y?F;3a;*$yY<3@?BhOy~l3
z{G8$w|IHp5JXbD}d!lCAZRFHH-}^R>PXG;HBSNw!fR?Wnp`-{9#t##r(+Lp4j}W2p
zL4ZhpqzGLv0iyU(BGi%uh~`I&kXQ*2!;cXm`Ju~&X;R}6rHSSuT&9L1T1L|lE>oir
zEg!2)=Ngwa{F~BLTtNbQbuK897CCgo1@>5bJhN9SEp_anA<A)~hbPhlY%P&NhC&FC
zDav%PuG-6UvZ~rrVIgIt%ma-wBq^97HEdCas7&z6s1eItnZ@1W==+PQ2Wz`^Pe({D
zNK>bUp>w&wpf<3WQEg<g40Q&JWvVk-EK8lmVkWsMn0cmIZf3D;c{Yn#R2D7cX;oQS
zELWAwVm6hH#qv~n4u(D)u%y&igDGj6C=H;86CgvB;b2#zO{H^E+9qL1eUv^LblIpG
zVeP073{yLkyUEe+cjR?DKU`oIapAPp6(lnU6v>Jl2N1x%(5`3pN@b;vJ@l!9`BL90
zCSRk-=(uzOWQnpIY?9x#7f<M=rm3&eHylku)C#Z#s81Sdf$FkKJ7w2Ym(<*&j4%p*
z`-LI{!&HnJ$e4bN8Nir9j2X<B8i$F4#y>3nm%{tu50WtTe2vyG&_6(~PzD8ugsL#D
Y!wvIiOeJHg8B@WSa>fi{%=wJ@e~ly6V*mgE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.AA5CB32F503E8500.idx b/.cache/clangd/index/common.hpp.AA5CB32F503E8500.idx
new file mode 100755
index 0000000000000000000000000000000000000000..42c7798c5b052c63c99329d7e6982d16fe0870a3
GIT binary patch
literal 4182
zcmYjTdt6i39lz%qLIT&^5D0JrhF1uLh=qz0K_m)hqSFeP0RlqA@Yuk9G{^@lPzM=M
zU@LP*9@hG{Q>kTW9hRwUt&?h3>!bBGb<-+3M9XK^y3W>}d-dd;KlpLJ`Mu8lo!{?!
zPF9A+vPj7=X)7{ntL)V_3Bxcx=+9BTy5c9mFe9D}Q~S<Q>&@4$#VMa`-Qd~HxYmw6
z@e9)}9kcCg{_^zDouo^j?i)K;`CgUz*`uM2OR8+I8D|gN`o{;4h7O*cK4e;_y?b?c
z*X)`nGf!J1(`wdy=sGsSgAwV2lG|bb8~EzS;Wt;BzslITd0YOy4UfAXG&0h>faKxd
zeImUw)ViSN!oWWJqR%}l(@!^??8>ik^c}JLCOla9IDf|3D;s|-{r<tR!w2tQ-nP{8
zB($WikNw(y@}m`{Kgbo&4G+z|yP5Ngo^KC7o;p}M)iN;Up9eOdx<1q9*rl`1D)VSl
zOOAdy|M2}k{bqWn;|t@VzwVFrhTpZVD@aqG&$y62>ByYX`1#@L1u_NK^4mi<nCH{|
zB6F|3ky~A4eDpjZGui%cc5vzWr=i;<Q<f|5g;nmqv-^X!!1(_9KUcKhn_W@y_`=uE
z<b7NIJ}e!*7!|U(Y5jtsyuLNgqh~zhvs;3;XqFA@=I&^!_BwFxKcQ)k*5lb8xn|})
z^I!H`Gv8n57}}e1r+>fIx@{9X=U&gqhc>52`}X@i?VStH`l<hRaqPxOd}iw4;mkKi
zcJbrgwdqZ>#utxg?Hx_O^t!V;C3t*&ckQGT+J|)sdp`bb&+-`A4-&~(_0?tNOE#Q~
z-@?uvy*z75lzH&#`TU)^o7HPp@9QsKCjIEGjj>M7z9Qq=^*INR$1Cpt<XBs^eEI~2
z32F9UUbXXVvy`34Ffh<FP$ddRjyi>500dEoN6tO@rT&nV<&i;zX9N@lF))6t8{WM0
z!6}Y0A_KFB*;f>doCS=30E!T&I@xay>^h}jWl9D{DWiNu!N|lEjA>638~tw06$PtS
zFfddZ8XyX+UO3ouImp^@P{Dd5?^tDQkSMU)y5`~CbC3VKQOZhD_BhWte^HR_<vVVc
zjqk2pp<sQHzd>mT5C!%xICo;@u>+Rt-mC}mPnIUjMM3<R%}x8}diu1Z3f2#8FGLv<
zAPVfi{8C_F@6QUq;aK78S9z@R6$Mt?+ZykDd&Bjnf|aA}k;+IPQDAj7JLz^xR${Ay
z^+IaBQtu-Qtgf+W=KbQ2>;j}lxr3BJKB91|owF|$9_SB6g+=OMWw4JZu=*YA9^0oT
zdl@P$3T{*y14Mz<@BaFZzC>|iz?(&^0F$Iia#4^k-}*>$`{xUvqRsgs{}^RVfGDv4
zuDnG5)1QTz6>N+r1E<QS`icUp`z7C1eLijYkvGes@F~(1A5oC-0}AEVLz|e3=wQ(9
zwMuP(D6s#LrGJojg<9WIuz@K2tMXR^MS<0yOpDz6?cM%FR9NJ#QEGfdfz=nR+FRb?
z|GNI-SS%hEUnmHyzP3qL`~6Dg83j8T1y7JC_=y6mpN`)6{gk`K;ohtY?cF3bd5VIO
zOVGp!L$TUk>QH-;>6zg5VrpJ=G{ZrQpC!b-2RJWm-eipMJRuihm|l7ymsNhN^zLqg
zj%Y^`MtH)2*Ne$|=`*=?&+(hjZtr6`XrE~J!3j9%0ux=p2Yci3W4*+&!}&Sb3e>H@
zADiJZL}tPaVqV~&O>GOtv6Hmpr++22jmss_uCn{#pCzzVRZ5|9RXK$!R239*0hJ5*
z;j9w438*&FxErXtY1{);Jv8nGs$PN@CT@7$KenS2a4^M`l7^cOlLq=2ficK8cLLZ6
zbT~UE4~f?c(r_Q2InJDltIolUu#7nTsT{11u1>@OFu9;S0<(}jZ9v@yg0UGU8FYXG
z!{9T=)PlSPULhqfRu{WVUbJ`b;r&@N${+_b^_ekbA6fb=3T5lFDU_qnp-`?qmqK~^
zJPPG&^TSEZ0&M|>3bln4D$*526U$;<F@<b88-?ohbuswUIasf+r%;2wfkKTy-v}t&
z1oTZbb^^VV#?3(648lq57NBjRaVyZa(zp%i+CVh1ZwI<|8g~F)2aP+y<W3NW?eTwf
zYyz?4W5b+|j!htRa_G%K-Ao1wb2~B<h@F_V0Cfv7!yJ#y1cE0`P@pdG$Mvpx$@Iub
zpm}0cs4jFHwa=g69DDn@X(9(TG7XOzl7m6AARedTV6ZHh$MHC*m1%j*#2gHfh48p5
zIT$Jn<uNUAFj5xD<4@<HUZ&^q7jiI47R6(p;h;fg;PGT|Fh&-`<Ew;&MwyYv7ZV3#
zWwAV-Rs1XZM&Yg^_38lX4l+yl+CqmS+*$Zga0cWp+*HK79jM!h_x8#KlKGj<w*&$+
z*^_X6C9q6YMv+u`V)+cBtH`OyBgh3zF5u>73G4!@E<#8W*bNf9!3^9x64(QBdO#k*
zy&$0%Ovf1Cq8?u2?THzQgIRo*1}EiUE}u)GdcIzR18}ej@J+yt`)Y?g7FX@CAVxGP
zO#b*P!_0`vCfvoiZ20I=*@PRIG**$i$W1|m7FX>7^T=f=ftkk4SMaVRFw2-lp=@I|
zg>sBJ6v{Q`QYg=uN1^<P{3sH$AfkXmg%O1mDl!zs63b#kF@<af8-?nOb+6zGNnpLP
zo<a@A1`0I-V<Vt&6EHT>*a?hI8aIQ8W)MYUw}6Ni8n=RoRvNbfLmP-C_U*vXPU8+>
z=%8^Yi0=f`uswbSq4p4-K%_k~`AjV?1qXBZ917L(by^&NgN=Z11a5pRA!~)l+#RE+
zJ(41mV(~O#vOs+xyroDV<g4@D1oK72y4J>7*}NP*IL!8RlD<q+Mx`%LEl(r5ilT}V
zf?UAt0_kLPT|m=CHpijiN$mz{MBf97dO!)my<lc9Fk_74qNWyKcQ3mdzwS`cgr^-T
zS~j0eO;H}7N1+D3!97D9bOPQ<W@=Q{_coi<FT666Ba`Fsb&h8Xr5E0ac&3=u_DY)o
z9IHzlXV8BZZolVu)iGssiBcaOWzYl#Ptt~lg-7TjLl~Xgi-{!5V6xPZWhhyOlVuoL
b8ptw;EJMgrOO{b&If*PI$TFHN^<?>ffIvXZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.D8FBDC6F47AA6D80.idx b/.cache/clangd/index/common.hpp.D8FBDC6F47AA6D80.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ace4d33f1125488c55ec161e80d23e57eb2f9133
GIT binary patch
literal 560
zcmWIYbaT^SVqkDi@vO*AElFfyU|<jg;^LB`OnpWMh99g93>9;ZPQK`O*nr3NyRl84
zSgqruiSnJBu4J!KD}1t~=hIh<oE+ypfmTa2R^`0DQ$0DYctOK?h6k2PKa#4WUx*sX
z2ln`yooVH==$qMnm)%6!>kz|_v}8VxyOTM~u1R^nsVcVkmTOUT>*pNv{Vl;oiLV6j
zta#n;AR3%#I43Yi)z9DwXH|`AyMxrzKe<r?(l7oNuXt*>_V;@=FTRlDN8^~@9z9j_
z`sHes3WtMBUfIW3l~(4)zpGYN4pu#D)%^K|TeFh(>pACbl2T@dU;7lZ%F<Tga@O|=
z*0In3PC3)C@~`FU>{rXI7S>FD7~^Qeb1qlASR-P4zbr>_>&g?N-!JTa!Sr@#P_?VR
zLck)gw|ldeI4sq3d}+ybYX`%`WgSlMoOZ}cZ_p}uZZ4m4+x8|yab<22F!&gXQqzhd
zbWS3OE-p$g-Vb7Mfd~lThA|izI1o%=WWWF$jKRPFiWewghcOr!_z_G21QQr#P{72v
t=>2~+$>WZZ{CxbpJOWJ2ENtu?oLt<j49qY!3=F&oCMSZ)hhXv`m;gPtrNICI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.E36EEFF638BE348F.idx b/.cache/clangd/index/common.hpp.E36EEFF638BE348F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..30311b4d79f8bb7dff5e0aca9e82cb388b8da589
GIT binary patch
literal 14368
zcmdrzc~BHb_tQPgF3S!#EH^C3jeuMMxdhQc6i^XGFyeh{@Ib{Fqo9I<%Auf~Drf@U
zCf*tlyrV%-LlhB3BWk=)RLqxO&+a~EP5wxI^-I;4k}A9BP4`Z}e*Ny(#ze5Ju^&Z+
zjgLr~F(G-3fTAc}@^4P^>?w~3u-lNLQc8XpK4ah=Kb<!N&rRtm2=DJG+mdy}aAAV=
zbG6OW=8lSXI3JbXWV5&1b*ERuK6ID`xJNbb>yzBGDktT`1+P1W&qtg+*O)!#t%LrY
zXKV5o9V^vy$lB6yy7P@jbHK#6HLi<eXBUK+ra#5@k^<|42eN%{2E3nTGF3lgbNe6I
zr)-Il>C1gPTP|I@F=Wq;tPSB!;>2Uewhu~v6jz+q5VP9e*|g%j%!F;LJw`kqnzBCY
zY`?ofh65db&>FDj=_bv*lW!+qpE+i9|Ghs~>gwJszBQ-i?u!78p?XEvPPTt@aermE
zjO&YjdiyRhynSlE{dYs|C!P9Av?CzYC&kM*zwS%p>VvtuH?K#tKmFpMoxj@B$<MCa
z`aAb}^tsq(7&dqJ&C|=T?C{j-uJ*L}?)?qgH{V^>chu2+H0av8R<l7Dq&FsJxGZ&j
zaHYR<spsYa=f+KCJ^HLXRr+DriWv*f<(UQr_k3M_;Hag(w6?ds$JN%$&pTU|hGuPA
zG0A+$^&eB69mBKz_m3R#c%jd!40Yks;MT|H9fiIv4|?nS-p}Y<_%?ganY)#b&ucv|
z(tEh&r-u!nv1?{GyH{Oay{WA9wbsK|C&aUgCQVxSdr56c*4i1PiiM#kpUUR#O{w|n
z^}NX=zHMGBRa^Zo*mct3n;R-_sINQVFk@lc^s0lYQME^6T?WS`4H;GtWpu1EH+^*T
z>f$vkN3!Q`JWFpGIWq5P@w68?-`-r%-?Fgj^76cQUxk{!EL%51t?D4-Y5hfzL-$0_
z;<vMr!@Rk+gHw__XJ$P=RN8W6&*~c6QUB?`YwsP)KPK7=Po%^I=j{A)%GK%pLTw{|
zO7(Z1+AYvIZ%l^w*$s7r$5aZ9mP&3`Xio`Pv~cAf=Mf{eJk>lC)V6l)#mTo;ci&u?
zx7%gk(WnON^7P}|Yty#6&AWT_dhS{Mfifm-wDk~mY0;_3mlJZ%84X$|F4*DyOQ`MT
z$f@axjgto9JSs81$WG_;-=n<-P5I(@T8vY@QT&_1F>d9y*|`r)7WZ0t_K-#Kj~Ar=
z>uoknTA*`6Z-CP+zhy6eFjd?C-9>NB4e>hm8)nsB$t;L^U+1$Ux7WGUmS@X~hmOvA
ztMlWsiw6GJx9sYZFD|&S_4dTvufHpuj2gYvwgz8|jyH7L60<Y-n$*)Oc6Lx?_bs38
zt330^MZMCcjtABkL=_f>xTY;I&|I}{rO?YbQP{t~{;vr;4=?`V1HBl%>3e9>63>ht
zx1x1VTux>DnobxS7`>#8{Z0=<1GG0g`hK7t`;?{KZbZzypmNsZi&v=ogi}8=(sYv}
zhqcND7oBRKz5j6g!^EA1XI51^7PQ%H|K*`!;nFEj${yQ~)QfAl`y}MhH(r)=)N)6D
z6W{w`%9BT-nLC5$SG8KZJ&S7nZ0&`pf@10Du-IEU-3y#1=#~{We{~~hWWxT97Pn1D
zb#pG%{QTIHDc)IM2X=q?!ZGIHwi6>=9itsity$swtp4)VhO*Nw@e4Ftg7@U6^*{f4
zTW0i|)*rGG15Y;h&F*#BtnaKB&G|cW1Gcnn{r&2+kP+jaq$hu$ztQE*%b~KH#ci?I
zo+eLfXWz(<++<hJ{PlpwsTj8Y<m}|QMcUWqN6oQ{coX*M%=)nK#vwPH3_6T%J`9dN
zy0Ey;VzJmep?qV%n8~X;8;uq`xicvH{nd>xew=g3vX1>edE>1F7w;!`Lms`7{eg-c
zs?Fvsm@$@CrzovQR^!Sg{Z0w-<Uoo&heYR)g|ef>4n{;fA{Oa;9@??}SZggqjie|x
zl~zlob(I|@)>WtGbCV1jswdPvojz0(Pat2KE1l~sZz!>|F-7^438pLFPg^oQ|F$Nc
zLMAjJX%li&c9htWEX7T-zR`K(qK8JF5RVQtXRDE@n&6}8D1t8c*X@^}-mYfx7D7Bd
z(1P8AM0=2-vV$oOW%{i*zr0f*#3KR~KVn%CYe<Q0iHhft_1OA({&vyaaSYX$?5G~A
z)ni>{M~O8BREQs$<ZX7qc8va3V<DbO5H3QZBBZVCD6tdS7562dmM7ceM$gaI#N)`M
zQPNS}<PE=V|3e*bth1kBhFVOPs=<~u*g@G*Vm&c6m-|Y9{)C3r5kD^%;>18xc9Ce2
zOy0nLLVSOJ{p8)#k(zi0S>ync9zgcWjwfB;)i0h;-SuI(1a~J3nG4Ou@`e)2DL@t(
zxG5p$>djg6B-qD`Vr?~SwdD;gGJMjc0nT;j&oI<y<oEZm+C6Nb>?pCekebIKnxOl+
z<=%=Rrp9=7kdm_F4aZx+Hxjju{Biz}Zk`0Yk#ATEEyePN$6<`!{nCVzjaRwzrP#wr
zbeNp4qT~5(g65NH3FllEY2s;QQYDgBBAK$|={mV?#ys5j+&+v06o^o)uZFLVaw%%s
z*wJmtYLm8bEMUV%h$0N-4Inymg5inJM!!5H#1qL9E0Jg=SwhiKVmqngZ0DvXhX$@U
zd83IxCzGNrqFv>Uf{l6-YRr?aUs7x}`<N7a5{f!$IvL0tSR|#~ZqLEB?Z(D<Rw%``
zAwe5rlpRb-{m(7G?2M8i30{JNDYlnJFD-ck<4bcdm<=(JToU0$1chrzcnujSJH^R!
zhq7#JVcggoYYuV}L9yGBXgiSzMaL60{i|<Y-@NzRd?EHAlg5a~=*b)YXfi@p-K#tQ
zb+jfPLnfR?($hpf6&=rHS!)J5QIo$|t%c{19lL3`Y0Dcxch%{L6>Ru`(VBP?`Nn!A
zU5`AJ9VNCWWaVT$XT(6W&bn{jam-G!DWVjayy5Ad7xV4zH}?;YlVDfEMHWH}vAp3?
z&2KWGs2Np-3^kKX$i_@IHdS_%SWAuilA}s#hi1<mdczlRs?D6ON1}SdYKjif-848=
z^7C@R5G|Zc4(Kt`eT<Bi9ZcCHcs=8S&)KIE+=DFTBlOXgH#{zT-&YI2o03}1P)TG$
zA!Z7(xw50gIwEQzhs%EB9o-|Zov)SRugOBKNV650DLYu`fOd_aSN|)Qg?K!n%4E@G
z6M4f^<*<3AP1Wv{5fU6mCWHyYOyvzF?na2gEmCJPY(edg-cvb!WyNkpqK!mfDLPo>
zWd7~aB?I@|)55;w_}s<rn(_vQPd|GfaCvvA0Ymj6duhgM%~(g-QDTOsdU0y;r`YP@
zX|6UELfkvhnhg>KnaCUX#-#;GYu<leh`4>2urow66b%eFP0tqXI$ZWvAu<jE2MNnH
zFx=XC=j$8~*Rw+GNidL!WLoluC--g9H?ecGuFcfMqX-q!ku;rfn4;r3?Dm?n<`2im
z*l^Nq#;!r4HAtrHV4*t}tbbF8+BzW~8)(T66AjaoH!%F*e&xbk?X**d*qso{#@fa~
z-oWs4!-YAum)-6$)E5NJPQ-K~6J<w<C74R#&g1#crFPwR?-g={QtC>7k-rIlU~eC`
z=DM|ywvk|O@)a+kmq^}FVwt|;MBk~+e{E(put0)c2o`o4c3ScVuy~iKKjyr}vah*4
zQ0!pQU_E)m8>ipRC0AGd_M{&}eMJ^{iI|r}dnh_etbwQn9GcYZ2@~gBawIJDG%`A*
zw$B4C0UPHS=cGWHq6BP`Ns=iKWn$+{>;e=JuvwVN!kRpsi$im9ZxB%dTY^JNaBmP3
z0lObd_G3mB*58i}c=#+<Ka0g+HU;cEto{y*!8&oU?mMi<!|$=)d#tYto4&_ps&KdW
z*cM>gLsGTz!q`=qVOP)&E9mZB&JDYgc3erj0xaCqF?_N1v>l%`HAIzK^#}L-+K<?!
zS6KTMHUu)MKN~qkvT^rO`GEN=EQ0~b{K2}^`O3Z`hMg&yX$;cNu&GFrij2XLVc0w*
z$wS5<i40qTBo)XQbQHtxMUuV97@RqVtwE9+WDNcv!(K#^i^v$1IQXfwhSEC)^Ur`v
zS28F=TuP576b{~|?zX59@4`*SWFzg!q#c=qQP!*Zt|aMQ;B0k<jWdq31pa2&Byo~9
z4=oliR)MmSeKsQQmEw<KHzC7K$c%^gAe%jih!TufBj;*F3<tpV$hn?}fz^-(c+<*>
z;7gMM&XI<_8otVs{ZjBiq`|Ovk;z?T343AKd&uOTD!dG9FT(~Plnk4T!*g*Y2qnX=
z#oBAJ0S}kp@DdyeLddYYvG#6WC>iz?Havw*K`0sa64t(i4Z7TDwiRo)Vgs-r#u*+R
z=j~RkBSg=~*7?{07BW8SvDjh#WX=!z_%z_S)HIYK0qmJ!KOor$<i~F%jh3a+en2~h
zEu~FMX-^(5qfN_r*ev?USpQAqHp&@UmBTF~uFnrtUtTL8P~~vT`I|z{Alr7LA+BJZ
zE2_J^g>`Q6yR`B>oV3<<&{*<?3@pyzbs#V?IX*?`0-FJDCj*MO1qSS8C39YO{y1C6
zu)_m~$M6gj?;k&ihZego4&$LE$Ylu%<>6(t&oX)dPl)BT&vG7yn1~k3V<A|XQMkp*
zuP|(MkLY0VdmS%MEjFxN*7&dIH|Msk^PZm!*CHoThXr+51Bl{seCqjGCaXRpbbW=G
zSIDT#A`-S@J|H8Wt+KIiHgC2Vwg@vt*oc>idpPVKjs`WzuxT`tMjHXO-3E>i_uM`|
zOI^U0qu_EB0vHI`ZCHOBHiQoe*zd8{_gF_2?)E*l<KYf$)PYT4ApzTgT|2Nl4{xCL
zH_(Q#Fj<ebE2kX*29YEzrU+{R5r}4q233TzJO?O!MCD9|gj(Gykqbm&K6%Z3o8ofA
znSY<wxq{8FU~AYi^CvCDu(z<;Ei41A1fe?7F<hL^?|-JrOnX&BW|E5RRSlU*9<o<8
zWF{5J9t@e^_tq98$7gh^G3+vIwhUYH9JLmkt;Lq0`Z%3xwi{c*L329P>=L$A#onwH
zTLQK{E&W~6%R|fm@ehL%Cmf>4MgWJvmdOB@fdKkI@<ClgnR0*uPDW%!xCg(ZA{<tP
zhw$)y?0X-F1M4#E103`KNAYkP-9L>^0M;eOf;P>koq4!`HZ9=c-WnCtEJL-AsR`Ij
ztec7Tfz%uvoQXqtI15{3VJjYfjAI_-SXfxVzQ;!Ic^T`SROxEj>+q-T>h*b2tmD>}
z+e&og3T(ClTLX0<Z$yrrOBw+;u<OPTOYVJR1u=uU*eDm<@{C-9jY_aB95Z`t`IG67
zCT`SZSQ}Ft8AMJDYincc2stNWW+Ea6LQ;m98ELdRm>JF&uxUq*yf?EP8!pFYfFSV=
zu}%$^0^@RUa1HLm!^d&(aaH&vHaMy3t<+=hdhE|j+9q0e6R*1&wu&~aqRoKsqI~m`
zq^_%S2&H`de8M0D5UtT^la@Rn?k@y{Jqk7C`7{V&8P-nE&Im#{hIKM?l0m|VVci67
zVo0VC9~N8JV|yOK2JGB`-2sk`)^oU@8>mK(VmTHs$J#t)s%UW)tqqvRwj}m%DvIp&
zNyjWhZNjWF56W)*<BBn^!fXFHL;IosdUpr(7%}CFG%d(1!ycq{4$@Y@4-9*p);Z3@
zkiA8S<+8Sr5*#TzRUb<)6%x+=*ei#0E>R6!E*F+bOup)QKXS}5EMT+Iz--jzp>n~w
z^c^+=1R>i?5LM)OdDEfH=_^0UkVK=$X%r4Q$ni0|y$5&hE20_pKS=vONF|b4hqc$K
zMlKatyMm8Y#=mvjvCCBdH?ozFeRA+-L>kIWGLOn$WU?1^1K*s`3~3xk2H?aqYy;A0
zKn4(rGVFDvcO98SW|s3TWv`Gk?^$N!0oiycsLN@RgjZLB`!4;**%g#2IUFv%4f~lM
zv$a|FTu}Q->i|gK6UI^`?P2!}r1`oSYd!K_kD?$kJ!}89s5h2NOFwDqfMLZN@_GeN
zn>O674SRuy3C4&p5LdbYyET9Q{j}b7+3$$8`5Ei~jE%qyk=g~;Yr|6b{rnG)_fA@G
zcaO+gf+WEJGOA$45mweZKvGgpwEg;}pz$Z7Uu}(S&B1IkY>-uuE$9}8?IrDH24xZA
zHkd`o_^~#m*M^K>1%|zYH18li9<IS^HCTgp3~I1f4ff;V<JjxCDtr=)Phu??|7!9u
z_o`!!cmARWQ!Pt8H{KbiPX=CL)m#DKNl3_*ZXh)dI~HgtYc>24D02TWfni@G*=yv+
zD~}GO-+`<FhRP1%b45Le|Bxbqz?0?xayo!~!C6nY+BV|n_j|kj*9sJD0^uF5M#Qg|
z)H^8(R!iUaeEhE``Un240RVkY@K;ol`0w4v%zZvB%f$vd53nOxj*Mr0&3j&4WlAV+
zu466*9~C?%!kf+-&f37i8uB6(Fz~9>o<-Ydy_rffd!s%Z6#!u$!_Mt9Hv&>?3@IAK
zr=vuuCorT;6~7)Of}-G};2xF82kZtzf}fBBs4ozhvsi;pqoJ@vF7G&~2@Qor)~cG<
zt3uW~+#@8)K?8GCZDba9$->>ihl2$P!4>r~o<xmEqY>GG;aT<KaP8ADtGoa66a&=n
z$f0t@J04#yY6KSH=<A)1qJX|R2h7@Hqjt^t$IT!V3yBDFl|op5?V2M2PdySv|9G&N
zuV=Ys<ETY>gfl<RZRYzp(iK}QmJzlaY&zH$iZcv5*?O`IRQDJ*#WBSbD98}&<+u`g
zf;l4*H8Ltf){v`Vh}2tbM-HF{7~)3R9Y*fpS1}|RZg&p3Lv|c|4wB(e6h(m=3|ofP
z%diw;HimtS#gBPMjbSrsai(fWlu6rU(mlYDEL>?heD8zpm-R@p+%3+B|Lk!0;pDb}
z;*VkDUE@h53!o&cBzqOeKS?E5xmdQ?fgf9fESDfV9?nJq*{C-U=b(TbRX7&~<f_7(
zP>)T>m!H2Gxot)&SeBz6<*MW5sDC*c%+Iet(G_SYq;wdPoDJNAR5Go#D5@64@*ZI$
zifTl$AehAgHrj<f*#rMLYZnI@8qJL9Uir6b7BC?)peS&IOhSQMKyqU!uoZ<tA)3qO
zS=~kUyy<L5{_Q9fU=3v%91K;(YyZ0FRhh|z>+d4XyGS1<ZW+4mQPQt(fBoyguEx+G
z&a@PhfsbwGj+CQngJzVoWWY<yOA19$2(!r^xGFmw4sa|{Hj3Ih3~(%L^B8-A#bDTN
zwE8yM7+4l|O5Bi-ZSMY_^D8397BPRTnL%dB(p)(nXv)Zo+B_O_UFY@!s{%E8@)NE=
z&;O(WGYweAdr~c!Y2h=I)o<e`RehVblaM3}>t$gn7>DC4BW570wAz1tNLt&p)r+dH
zp7Kfi)u12<iz-s)aK3P@f^$n`&bGW;!ria{w*=f~@@vJ-3SglN*i3Y2*pIjVmLb8t
zBwygJUBK!;rT-$VUO0X@5%PBA(~f$AY?8zl)~>?(AVUoM2x~v$Yvl}^O>1Y<`XDlm
z{j$$zeRtmMlRm`8rnzx0veSztTuIuc0`-JbW!PHeT#Niv)x2{f@&o0{`RZP;Q7HK4
z4BLV1JCHXI=VGT^><dPiAvYhKO0X|a{e0RzpAG_QGi(9vUZ4t>(!Qm1Uw*ud_ATS#
z#`e0TBfYNN6Ok(vw8MVd5hil5*M8bt74Eg44&mWy+M!x?el_h?tvbJ&?p6Kq`~lT;
zB>WyqBZTd^>pZ}AaFd6;=5Fu+n_pf2i?ML7Y19AQO>M5YHf@Aa$4^A5ew^@bxmI$@
z=J5U5M5#u2Mub5gMevc^Cip0(6@28x2|jY-1b=ek1b=ek1b=ek1b=ek1b=ekZnW;(
zeB0@G_kY}Tg%cu$9PVNjs2}E|<0E-$m0SRqnKrJ+R=nYAz}5}ek%ybHQ8Qn(V%QdJ
z-GUuK*KsKW?=~C&a9d*XEx+Eml22O0K*&QPI>p5_kQh$GXw)J*&@fO_Bod^!qXrWp
zl|lx%n`!X6eYrHkKx(e{7mkLS3t5r7ss;y5@+{7_Js{8WTWH1O-(Ih6{nz^v5A7=Q
z#xD^K`+wWBhiiAeDRsHHk(lLF+BlWA1dYbQQK|G0fS+7%J)hC~;wd5Q$64d2t+)Ga
zsNe1XU#C&Pfg>lPxE;t#TMn(B!`JWOk|4R3!Cel7qb5l)V&VB8h1rMsfphut-pacZ
zW#|yW_Tx)iug1O1?3vWhnG7tYt(Vg7u)nvL1R2BT_*MO7f8cx*M&qs@!g^4&A{8y}
z(jm`JPQFe7;N38+yTDx?TpxxwHL_-G3zE)h0_PU&#=~tms11h#4B;{Hl(~zG{JuGc
z2Ruju$9|LtByc({E*M^kzm@ryeS-xiyHH+kRDNLggz00f5k<kJH3}rJ>jh-}yIw#M
zyz2#+|E?F{FLb>C``GmY<Qfc3*Z;EUdI14`*9(Z~yIz1*>Use=g02^k6zF;ZS%t0_
z5WsZ30KL-n0#1RNGke^ePL68*f%X)Yz&{S-AG?Y@6g7~48{YMZ*&8|G$!?;^I)jA4
zF+*a94vQNeAKovb|A5G-=z)W%$gY{x0RAzKe~jlJyFzq|>dHM*R6l;gaQ-ote~jiI
SyRK4ER0RKaDE~NwfBZkXve7O8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.F52D770AC14BC38D.idx b/.cache/clangd/index/common.hpp.F52D770AC14BC38D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2df7dfc43c5703611a3f93af80cfa62319a7682d
GIT binary patch
literal 1952
zcmYjQ4NwzD6yAhLemD{iW6&dnkU+pf2t*?hELFh>2`C9oM9>ygOcixn0x=9|p=zZH
zMT12<Vr|tC!CD;^D%xVJK?^t_#2-WyREi8@D<Jqg+Dqu>ax?el``&*0_U+sEH1P_>
zJ}Lk)tKv=TGjlR1000#K4LR9sFfst+GyqJMjk>#Qy~90jRt?3ZntMN9Wo;<lwf#O{
z60nQqaoYG>R7yvP*_aj4*Avy%Wyzhs=e%*=irN{MXX_P9zSumTb0Vy-*>Wqot>#Qs
z>YY~0;D}lLXjQ%Xpe}uT2|GF|B|Yu;I_t)f2CcE0oxfbO<LO}yR~EhRv$t&dU_*&@
zbesHfUBG^A)SNSe5p3?5e!%Nt-4V^9@R_GN>d!B3+xT!?-*TW+*!6yivS9PM>c+ig
z1HSp??egKE{@6t+xow+TH?gl(X4h^JV;#cSWk1R$C5ButGHZ-|^NdY1MGu1VUo+JE
z)4L-x%oPQTo38l`1X#eLOU*Cu7ZohYdR`k)ReyUxb*$>y(I;;%Zo86x?|R``)rR-9
zpqMzeY|NwTZ023o=^|7d*B3kMk=fF_rPZ?Zdrr>qfs{+E6}nUH*#f%v!5LZDh+hHo
zal2(K^xs?T3=&EXe6`cFMfpj0zV+F~zlRc=cK7YEx{Rh`w>6gTSHjUw=8z_)<ZAx|
zYuOXsU`>=VFYT$TxwKQyGxnZ)9(W_%=gO|K*7`eE-kip<-Sx(``H`Y~^Mm%UE-QZ>
zvNA9&>!h?No<4FjS1eRl#O#oCo*aD9?f+(7v->Ewu#nUuEAEkhw0*iSxsk32dk2cc
zObx3~1l5?{85WlIYWKeUe6+OxYOLp`zJ*uLj@K?e@z=nw<$T2@MFI8BR$ZteZ~Y1(
z2>{yApGlW*Fsumh$6H5+$S}JBbQcnEb0?uBZSFt)?ayEBaV5Lr0zQ?`u^T|g+YA6)
zP!=atUAVwLjDP@KAV$QP-2gh73<8<B;K<SlPIHsrTqapa1*niFoNPA$?@ilRpIT1e
z%h!IxBzxe?0=i(T-N5C$6*SX}4IV8_vIt-Hr~9+*1};BPw}B&Xmd3c-5<<EDT#4P-
z(_<m{>4Ty4jr<W_r5hPP59PV?7<NPC|MJQG-J$*I0yi=P7htX!)ouWNl5Lvs2p~Ii
zwL$4Z)QZHIsF6K=0NnLh#G2#WKS&f*%9MJ*dUzfCwy2<HSqTY2r2^@E=nlv${UOQv
zLXQUtqy<j<KL!XXK6<g|?V}o8Eul-?p*oT`+??f_G3-J?B}|DUHU$l1hB>itX1Iez
z#@BpVSa~uHkEHuhmk1vizM~6c;{E<V7(pYvBR+s41YLkG;6t{&L!KuTcX9AY@({TU
zx<9S1t~;)ot*0WWQlxa+s}iZ4Sb`|QiK#_uC#Df;oLI6b*@<bxwT`29DqXfdEqgB~
zBPf^46~f0xP_3tS8q5dQip-d3Jjcz_XlWcM`e~htqEvtV(Dq0RqzmC5*qeA&Y>f$v
zgrF*x3WIwQRKwCZ*w3a5lakgSQs7m=m>37P&pE|+reEeh?xa#sWw25L_fk+*u*!)g
z1SdE#b+ForX@WIQEIByYiD_qPBVkN9F?iCpaUqgcvXne1LQpkJ?O?6K{Nk^olUaDu
z(nP5mULOt@9?3RYM5F{(g5$ynn#@XeF!%;||JtSv_AeYW+-;jN=!Vt7-L}bs?(EEU
z88&RlUTL^(18kd__$EUD0FQuP1oS506asPx=t)4wd;u_(@CqD|jGm&#xm9cx8&F@q
ipTLtndCJsjm^asl=fweBhYauuh!N0-fPMt@CE$OjcZJIU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common.hpp.FF4590F87A18C9AE.idx b/.cache/clangd/index/common.hpp.FF4590F87A18C9AE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..60dfc17692991ba702f109434de26b4ec008c442
GIT binary patch
literal 562
zcmWIYbaT^UVqkDi@vO*AElFfyU|<jg;^LB`OhZNnh99g93>9;ZPQIIWSb?YYx6$sL
zwSP0jci-t&@Vb-0YGAY^_S+ZZ?CWK>?rtg(a+;W1`&@=&wJd85Bg1*YgW|VuI#r*j
zaWFKyX&|cgz#>hrO!@+k1}F1_*}awuHukd;)Yoo2vi#f|zF6tLIc9dxuj)UX5+h;F
zV*cWna=_7$2%dD2=uY7;l`VYVp9pv?p0m%FJBfL3J>OQFWu^b`x}0o$WpLYj!P<zK
z>uhVMvXpDg2>ILYy<AWK?dMwkl$13~BGn7z?j{7JB+k7&)4fOM@}a1C-loUdlcL^=
z6&-)QXYQnzhFk5^3w`f~s87i^nxoy(es5=KDj)0XU;h|eb{>j&(zLs};wtN{;-zI>
z^BfnrOubco)uYjCZu6yLku8O6O1*-ucV-sYUP}&raBRIDLxH*NZHD5?+$3NCG8Cnz
z6+`HpL=atElw7<Y#NYxE5Wo#%Ffec+n84_O0X7(efdLdVP{0mjFfi~Vm;wkUFw&ra
viE+{U|7w!Q9V7Yq_<4B*n3!4E*f}`4xLFyPVQLr{co9ra1d|WJ<Uue2XiuWP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common_header.hpp.BCB8C81E7AAAD43F.idx b/.cache/clangd/index/common_header.hpp.BCB8C81E7AAAD43F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..08aa5b65fc57e3fed4370c7d62298bb9e5778631
GIT binary patch
literal 1002
zcmWIYbaQ*e%)sEB;#rZKT9U}Zz`!5~#Kk2=nX?%g7y|eh7%Jv$off$6jslPC{Vu7M
zZ^O-#D-U?CVRYVja?hTt(|vA<Uh$~8b$rs!n0Lodb=K*g-Zb6k^g4;apAlJ=fy^R4
z%QxvQd=e!h-}%7PS3_4v@vDb-Mc?P#^<N+Vcf4z{aW3~W4vA#t&<kCSY9HilW7i+6
z<~ePa$mX4-=x|o}fxZogRAS52*-|P~B&&b*tPJ_ew!Lk+v2nvKORcQLjTdDD&Q`me
z+{y7K`Q?>6H<K)`Txt+X6>2{9(6aE1-^yK&m8F{IUXAXYD0laF>#<;i;vcVB&J|25
zXIy{o&Q@;&y~XKLKYp8h>d$;K|J;+_W5SZRPE7H(e(%q5;naoh-zv;~=K}66F?nx(
zqUfeuuIjf3Z`Vvu&F2;7sSk24e6DS?bWh41#TZ^@o?|zwT_v|G&N{*VdcsocJ7xJP
z{>PM;x%?_o^FGgC9vu4XuIHV{5A5H}(Jwr4GwrXDxpc7Ad+#L^XREtt+p4}<%Td4Z
z=j&VF%De7Hy}T-YH=tySpp3{GllL;&d)4wy_suL`Vxc%A_QdHe{8`)6|1-|?uhwTM
zuFOpWMioO*YFe=#koK03=FOJ#naIe)X~b?M29g4S;}QaMPH);M4iwg7(UXJ-f^_91
zg4o4H$;A(VG0mV3lVM;`K`>bnOfCdd8o^{kFl7-;H3SoGBLjmXLR1dH<Uue+5KMUl
zlO4eXCPx_1fH4>tWDrcaw;|~mDj|eWAdFy2A(*NNCMYbRKorJcVBkeCl@Ux}&Vd0?
z&Vd5XS3kMmo;)cRtHH?5t;i+ABOoXw%r7P(Eh{Ils-~_X#UZI8uFT8KDk7?^#H7H-
i!p5Kk)62jhfMD_?m`n(!ID*N6U`imE3J4}QoCyHS8SKCS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/common_instances.hpp.C20932BA881E0448.idx b/.cache/clangd/index/common_instances.hpp.C20932BA881E0448.idx
new file mode 100755
index 0000000000000000000000000000000000000000..90c5ba47e8b95ce14a3938495846428f1a1cb3eb
GIT binary patch
literal 2736
zcmY+D3s4hR6oz;2t})35FjRC!P}DjB9F1BB!8SfB(E2PARJ6VkG%D(VRunApfrf`D
zsEY|;wH=9xNm}Wsh=$Qr87mg;sBKk7tXf2%SPDMtNY4hYyU9%M?)mn=_ndqGb7QAO
zM<16H#N_2u(vlN5ETsv8P{KcTLyE46B#4$C1d&$a+&^nw@klo6b;#FO62g26Px{T6
z%}ug~P8}DLer7>?LHx-*Su=_pO_|?JE??%GW-Obc+@9X@%rwL|qn~r9?4C+CHzL3g
z^|Il4SjE)ePfb6&z?uJR<Itu-&AX;=xl%sp=B4qgW8%;oQ|z?N3bQ0*&BgQ7vUNQZ
z0y`{G^<n2KgXs;go^a-}$!q5qZk+h)V)4AG3-7gF8QGD)U$*Z^%Z~9!ucT)&W7^8f
z<}syzI@)s|*wXX1On;eGGph1Mb7K49Yd<y}EBaJ(tFpsoTNgd|Nd4=w3lkPBZfvgf
zYRDbw9h>%0F`&aanhTZG44s#;*Sz(0VDRzqbJ5M)hBsdveqtMUdwtESBZ;?92h^_^
zmAyUd%hNa7dbhRvw~w!&LdN>XuU}L*I_`j9Xvc^C=T@^rtSKGw6~%RywRLCzF4hnI
z?#`C#7p==C8OsMB+m$naQE1=4ez7FoY5Bs(yuEYP$_w=Ddz&0lvAOxzkDZ9tFwBUc
zi|n(hva}T|XL@D2D&NZdj^^a7_{l4&_2Xc#t-U9{hzaS%ddICI!*jCxCl9-te>0y+
zw0z#Y#QeN|&ic!drBqV-%l|&xF=F^byQA&rA@|2cT-wm`cK4jtF^8XR-5>dzD`mx@
z$B*|hu4_;0>q;8O7Df-eo4Pr9*<Em)+}8fEUb%S?gL)DKJ6JL}$UP7;8hrp?S#q?q
z;$UtLJBdMFAXqM09^@W)K_9nZ($nCiowhk;3{rr=NlKh#VE021MQ%*Ku;oNk5QAhO
z3YUZjxd%ZIwddcqRgSmu4DtqnOd^xH2Lj~qgE|&@S9aD^C<__X3j_*@LgpR>LEd^_
z$JkSme=rD|n3YT9a`zwz3S-URzi2#@z#tzGFcL=Q9t1)0txX3yiKk&+gf9q`eUt;-
zgImzo4{9*KbL52-eDn!};1027QeuX!c0GihTz$7R?uqg6T?R=&lrBjRbPv1$t{SXz
zqkP|?H=UcZ7$gNj0VyefRk|L+Dy;{r2g>)331AQdqM4GJLGD2iylTE*Frs0}U<UOD
zfse$er+Xk|6u8Q#Nl8pf&43cH!=${+S@(ew1ojm$Z)!Isu$zFpti&I%gYbd34`b{b
zVBR#0v0H$7t1w=iR<dr1^ZX-H!OE3#7F+h!w8(42;=^J9D<>#tVIOw?6BQ6~J<lJo
zOeyP&Q{h>#Ht#ZC8`}$*_ZDNkm`vVfW;a)`s%lbIOXAZ}u$pR8Q%mA)Dp+kbsjVfk
z4+>USP3mgjUk3%NC#y+4SxaKS6)XpY1A+}!ur?qzAh>Z9tP6+>2tH^01*-0Dr0xQ3
zY3F*xB!3$XoA>ed>4WXIkMEmlx^>wsg$+nmYEk)8japQ^RI3)1F4d{A-+bXxJ*gI#
zE#*kHxM-=3RO7<<lBF(EjSI%}pkThY@iF7+fV*3{`^$qz`iTX~AMWwqYLyGI9k@8q
z^UaNy!P|_GcXe|&hs~BCRVfm-tFj}NLmWGis!ALiC{;Ek(i<t2NgP`!l~o*bw5o^}
z^qLZ+DMiA8YV1hk5XVlWsS?KqN|Q~A^hQc!62}%wV-?37ttp}fy|x5tOA$U5=sr8r
zI>fOPX{*GsfzoDEBE6B)n#8e%(ptqaM{A2{L9Z)8x>6)Gh0cz24sq;6x+-yOpmf=k
zNN=QcCUI<`bXIZ9(Yhj9(Cf((q$f)eZXdo4?LZtJ#0jLzgBU11nN5jj8G)EQhy{q%
zgK)H-ETV;39AxJpJML2cuG)b(JctuWl?O3U9Ap>IG6FGq5DO5i2jOTAvJ10pklhB^
zg)3!)>^8_QCQcw#9>hS|AiG#&1Y+_a79ds+!qGO!F3fU4b{AyF7nr}$b|4N9;sjFV
zK@5}&vWsUKftWmq1&GyyaI_1u<5?+*YnJj?I5j09Rn5Qn^F!CKKYk&Ie(zrq8+<5n
OVR7B_2CAQbf8u|<|Fc{G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/container_element_picker.hpp.D0295372D9F2DFD0.idx b/.cache/clangd/index/container_element_picker.hpp.D0295372D9F2DFD0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..26f22696207015f7f54b4e10e5a11ed53d17cf2f
GIT binary patch
literal 3104
zcmZ8i4OCO-8UDV@1(NXHkiZv0A|#R!fkq>d9DV|m6cC1W!{ykqH1w>g0cAFxnsCa`
zLA$xM{vfkaTaG`oYsV=b=f5&+EpxUi)!Co5$8M|6R&6Jzt=PHEt<}zb_r~|!J97^2
z<@w%w-|v0D_xYaJ-`v=kOAx9LH}^!^yFxrdNYDOaUD3`TV1#z2Ak?#CufL_`&FS>o
zmm{AXm>JyTnELMLS%-VR>Nqs8&Mmw@wqXOuoj>3H@uTyXRbTD0JXNsq-?t}H-U}Z&
z_R83YSKJqhdX8-WWBQ4T<+bzlfu)~qUHI6I!Y#k*Pg(u;>ggvEl@nKA%Z~2dF|o1d
z^>6)Szq~pX8H=fJKkabzsMOo)?_MfdwrOR=cQDI*;GwDDb^YIps|y|fsh->PQuDO<
z`5y&e|L3`kcZT=eco+wZZBhNJboS*@?o`1;A9#wt^N6Qt@mcGOpY#2do^Sqoc}uip
zE;@L2*~yA6r$4)K@n>CmZEE!v-JuU1&y6llja?bc{Z?DPWc`_s)J;cSgHL@qug<ba
zQ>b76hu!Fnis74enjfe8jfwj0iS4<Y_6=4oKDXO{F(cn|xUBlv)QMk5On15W(`NqT
zqwl}6vbXnxq1O7UMJ?~{GS%E&Vcqrh;AHla&gOriSa0O<%N#<fbN)W^<klr82{P*u
zjbnQpJMTS^I+a5@geg|-y1h@eqU2TvXURg;ukmXogH)_V?3eezndh&ME)QurV>Y7m
z)$=nY1MccgL(A7pzLmzlW^SEEr<Dx2eLr>mX8fC`ku)xYWw0=Pp+hoAtx8uTare5f
zhkK?49VCm(WZzp|);W@a@0-!mFHrSl6G8dx`vkTpaMisBQu8=jLCZ6F3+}wLd`}KX
zSOo!Fz$Y1q5nS<$oeySDoYHZ5%pGw>$|VEtM_XoIag3kT=(uSt17DGEwq(HF^VW;i
zrtKO>GJ_#(A7U9KA4r{+lFZ=epVmxPzPW!X+k(aLxIE>Off&!EeD&CwA5IP@6Ku7$
z`lx8Yy?Ot&M_$N^SlJP?1Y?;ogJi({@{H9_oG>pst>ZMz9d?CFB?Inb`vSLiK6m=}
zIh>a5ZL}FHB?Io`(W_H!cls6*WMU`SZSQtUM(+Q~x9vOnHtlU<9bn-@uFxFG!1t%F
z2a5HkE<4+ug?AP^ACL^V|J-=geLh<Dm!#X~vXxTNfcwI}GkcxGCC$m?3^qe06%DvQ
zK9WDO;PAz?G)`d2Y16g&l9B0grS{bR$1goyspE3l?%rZ=sbt`L6m9Q_txcDi&T;qn
zePE~ac>8OK-|}<-y9aO?Tm@h|)<`S|ehQ^)xBcXK_XQQ9=6Z88oGGC;x2*zdB-H72
zy5T?w^;CIkq4R`#%e+;PE}?DpZOy<-Lc{KG1+<aSNL8d3NGXi&I$oUm<hY(t18-m|
znfwuI<c$hs<*f>2<82D$<eduS;XMlE<-H2z<9!O$%C{;|fDb58hz}`Hm=7yZgpVjt
zH|oBp6bKP^_$ZDQIDt_@iH9&6l5y^z4yx{qjLpG3ZOm-c!A%6o;n^uP8NVktqbRe;
zAUC5lv(y01&`6>Lbh6bRdG%)Hdz$;=&I;eD&gWLXk;2ou>2)^Pj;H;({bsl)JY9!#
z*I_dd1_X`~MpQT00*I5UkpvCsMyOx!pA94<(lv)yiKI$0+KMx_;u6po5W;kgB!I9F
zp@rr`i(K?9^DGM#O=z9Ht_aE?v>~U#ETg`HJ_i&<=o(zG20NfAkX2SxQd1#+pe*M8
zFIiwKTNd>bYJknyvZ$MoGnB-lCxsHCuT+n}Y_P8%XAWM)7XjJf4zVzzbV3+VO|~qm
zC2Tq{KP7&B<K4<MLQPpFlgtitg?X-wf>}WmRKu=gMcZ5%^=bPI%A?V)!3H>*4W0j8
za@PC1r5sZNVz$Tu+en)QLV#kYu%12~2*NU<otd2mxab;5#UX82WwnbE4$Gro^o<(V
z-`ss4RgZRja;`yh_Od7og24zU#V&+kR3NKhRUn&SQy{0{R3MMwQ6R72RUn_>Q=nF%
zRe=IRK!HL+NP)saSb-u!M1i{1-Pv;1ajcGG1x{dfLWzg4dPv5{Y6|q9uDvzLPPakT
zkOeFwbQQl!EB9>x^8;85eS=0ZwumzjNCdC5!642)U?NXfxmT5e6GkAXkwzc%hR_ak
zhXoo3w~jSXoPw~yyEQZV{VjPf+}A+3vFvSe9>Uu&4_TJt9E9ad(FGGTt`*D^N#`<`
zLoJNAuC{Vu7rd<2NJA1zAylW=6@g4Kz9`}{z9`Ou&~8_E0~AeYucg-`qd0cOaRVGD
zq0eH=v)Ci!%{XH-E&%0#NyiEir!wRgZGR#pwmlYI9J7o67(|o&(QPvhrIm9p)o_L!
Jlmn{>{SRu9f`9-3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/container_helper.hpp.57BCB53456250D8A.idx b/.cache/clangd/index/container_helper.hpp.57BCB53456250D8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3b36d59535731580858f001a6be62269b81ddcb3
GIT binary patch
literal 5916
zcmb7{2~-tF)_|+-dmzngKtM!UWNBFie9De2Hjf<xsDO$wh>8;zCjKBZWL%CA7e*Wt
zK#eSsphgLZVu(Z~3NA#{CywGqhzg0Qfbn-kgZ_*<^=RsKpX2$@e}2xPtG}-MZr!T7
z_tvdh(Gd}2%n?#8iO$JLU!5jL2pN<Atks#z_R0{l9*s~=`LU!AvVK;MetDuPbju7!
zb;2+Eqm3q9n$h{9`J0kxzvheF6NisS@8T}5^(ub%(M6x{+&kB<?vg)wx+bOeUT}F?
z&EWk19`z`v(<b{`1wMIi{pxEshrSwwPna<LV$0s!{RRm$@-{_frJmXl9%?T$eK5JM
zr26Fet=sqBS-ZM?$DM}@y<X`2crz|1uHC(=;o174nI^Lz$A6a<e>`!s<s-i=)Bd0S
zmauzZ=`A1i+|PXce~OD+o!P0k?z~COR&($41&MZ1#To08bN}{gu2YKR!PG|{c`wIT
zZU6etmG1FpCnB%QHg=6hf6qUZ`R_%A=O@BPjVl~xcR%Qjj0;TMu_dST+SPckuwUj3
z#>ocnRV?w>xo2ZII+`<bd_GOzGh@AHr8#rd+jM^Z?Ut0erX7V<x7@eSsNPrJd_6oP
z?&k70d9xzpOwKK{8miO(tNQta&fSZv8;d{5+0*F*S1c+fHk|5OoR}S$Tf1+~CwsqX
z@Se5CdBu^;*5eD?z_fB(T*bHj)qC#MG9KBdJpvv)*!peD`O7;x*4$pvUl_5Y=jN!=
zW1DN9U3vK7zGYQvw-Y{_fB8qohV=6XDu4NG@Jq|9hdy8U!^(Fb)Y()${lRp-xml_2
z*WZou(KQGz{<Hb`jv7@aez$RsiNQYy7j#|jdb6)}`E0k$&7A|Q3w*zd$>`e^{;%4Y
z3AdAuI_s4DT(jK#mjC7x7r(fiHL0@>&of*ppE+Sk!Jj5n=ItwV*Gsy4TDGe``OC|@
zs~4P}e^Rz}dClS8M<0jv1StyKLQZyD9F*^J`b+t>_TP%P6#1Pu`C>_ak<R1l841B#
z%vRyxl~Ahp9i+<ru5Yqixb;?8TfK*IS?^yKUK)~(aw_uA{UI}N(<6VKvi}*oaGw78
zzwaKJZRPgl&bFHSaVLz{{g%?~ba{93-fdg@29mNZUdQ<Sb-BFu^@k~KmHiXks(vU;
zDy$0i<i-Tv&VIC#dHP}3=$dlN`eS)rdhNBgXASO5T?+9ouM}Oz3BKvf8H2HZwr3nN
z`wwitU3TGL<NXW%$USRf-`gKDuFHJdQbVucX{TC!heE$MU!Jlk8fNq+M>}pbc<^r7
zpMo8|FU79fb?~|V^~v|<PMnRe+vV%+_^7sQ-j27{6U!%+*qySxtFFD&z51L(TH(2Q
zb^l&_zcKtIchf&|#jdJfH!PkK>G@MqZD?<}YRu$=c5jyiw4R!p<(hFgtmBq`yWDEv
zKvDWTyWq0dGnaqsdV5uY8CN-Abn)~mKI_wrMSDo3>L$NxtJ|+@$)W%+#K&{-p~DZv
zSjo{;gn4Y~`}o%P1)G<#$d3qAW3C!|jx>mIFo+%&vT57y{4D7ALoD(pLQTLmf#*nr
z809!Ylf$Rx9wFu-tt<*8LJu)|h;2t2#5frVLV+91eO@T*PqWCE2w8L1zQYg1IFA+t
zR9`yJDQb9hiAAwQpbt<Vj2USVGajQEp2X-GG5<BE!6vvx5E=k9K!k(_G1fYQP<-r|
z{R_=tw_tQTaP7osp+U?<Rv0efRPyhRiQRrI@*%?+bH+Zy55(B&2?9yW^QW%b-b{wl
zETAiLwHbaO#$I0#STty!8u8Q7YGK+pFuFmeEi{PX6oSxVm--54MO6ulrV%YETuSKh
z12GmL2z<i%p0{^6MYO~afg#|AAb6xfjHOHv$~-Hp%6{EtFKF>b-Zs4EffzeR5Xc#8
zX1zV|*+XGA9LM?anuoZ$l`nr}a}pLs2^u1~$WUJMK#YsAprL9(&73ml?ccK~nAmX!
zqcdbSLL<(m_Gr|vbt=z%76lQZ0^kZDYNSETWb>$DpZW26M)R7WX}*G{8bCE*%L@%+
z+|8y83!R<*@wvB+*>%DmwE)*b_DE<DV>3#y>Eat}o!N5^1qoXa$OVS-ng?QBO$32U
z78jgfuS=TBqDW#-H*nqHJJKL#lBpo{TST&~L78YH45bE6O@<O0#JG(SgnG_xeI51b
ztW7NPCz^^dSA_jW8pJpn3Nz~)8-Bp?w{&x1$14C;keLY$Vm!tQnx4dLn#ax$Av><I
zsROtUVw2DyW`eFDl$pLXE6_-cQqUic&bmF5n$k}eHn99+#bQ_LI|j=9(&am+blS4K
zxvn{@^;}+Qt8~=5DX+3sIcisoPj*iB(B=TX7<(3Df9)g*PcOz1QjROei6WO``%>&I
zWzSOS?@O_&6i18wwb-i`2WV3YUyH+QajeLPvDaZ~|6v?{SlWLCdmWMXAHm^Ar2P%p
zx&d>dz6LzG0eee1qydMEtj1PqY_Iic{tDK;f?1L8W8M3h)kZMMR*?_!2iio&KgOPq
zvA@XMz;zoaweHXFgs`0uBl3PQ-4Eln{=pvytK(p=jRgJ#=$-&p<hx*T7i_hz!uNnh
z57=tmna`7%=E=s>++c95s>xYr)hP>>H%7*Kq#hgoV|i<2Eg=qZ5^_N<67oV`5(-3t
z5{g8T5{gIh5=ucS5-P%|2urvcqiQKPW7I6=c8uDk+=Ee%lnVeAfJEOOKzpQI4X9em
zO@NxD+ybaY%4$GrDYpY^mvRT74k>p7>XvdZpk66I2lQOZLx6@vHX3&$RM~!S1(x%r
z*tQhg(Rdj1Ds8K~U%@IF%PSR1GciP4folbKZSLdmg5oZiQ8yU(vgF9C*8Nw=&-|?X
zl-i`v$6CcIwF#V0b4&}MslnpI+NvF{5AqC%X9O9CimAa=Zpw({>gYO2w8Gobdy=U9
z40bw$-Nev1hvnxmBk~RGbOXDIp>hk$Z(&B{76@p85OGc|FslXTh};3e9iXCfWXX|k
z)&~~a)1Gi>A4F27A&m7$Y7v?eX(b}lAE`fR2g{!U#Tl@q0vfhC1CuG!ASN1x(uDrd
z&tVZtGfK@S+qV-!cY=zBGc8p}Qr6TdRG#I(0OK#ffr_cd=k|`8GJh^+NxD!x#;$ZE
zIu|mcP{vRV;nVc^uWOZA!17CVmfDDM+>dqov6ds-%cqs4glFlpyp5HOBQ=NR9c&z2
zv?o2UQmD+te7_%@_rp{%16>8~D!6OUK9XS+Jzz#NOiW4B=WU-oF(7_p9c--<hv0&^
zP;rRB=)gE}h{)*3IMEHaf&MlyqKm<jbI_>*+(ZLvz^Mk@#FTdg<VS!JSq%Yd2oVib
z!%Q_Ki`)&t-Jqfdv!u#18vqM!0Z3Ri08x}_OV(T{2kDyAVvuwQB_Qpf(Lp*i)gS##
z!$ht>T<+1&EbpuEbrqu~MiJvGu7&ipo;$RA%EQLbyxz+S3r2JISABcnz9ziG{h~!(
zvEi}NqOP>Kv{^K~X_OHgg~F1WLj_2OrlzEySDroSQufu$`R}<jJxPeXP-4<wP%B7>
zrlO>usSA+LnmUqxUTZE}S0a1+2XWw5@N9(uG0d-m=T!)x+gA`H+xYwE@BGMIE5Wdm
zggh;s_*yWm1vAkPa%E$4W!6vMDstn;e#sYJfBY=Ban0fPX0WqDb}xCuj^lEc_b~Br
zqT9ssyb13l`e>p_qLWzmq?)8U(WNT?YSOQ3&u06+w=Z<*$y_xhF`et)^#z}+4SmDk
zvt;je$F}mP6;sI~TmZ!dFcBA_4HRu)LKlJNP@-5Uq^asdIgL6A8G9qfvb<}AYn(V2
z4`&ahNZt|NaddxJK1dZ5E6$U5=9MBvsiI=(G*~`f6(1|EL4qnlLW$0aO0h3lk!&iV
zG09dUrKnP3>DsaUeARpjr8=i7#lEy@X)2K(VyB0AGR=i7DSZ<<V6K?|yJ2QG%oF3f
z4;=cyo#sqUoI4JH2aRWz_-n`zL{bN4`8P274UD5q^A9-|gzE@hW0oX!qn+SI1z3I`
zDE5JpI^yxAUFtCXOWm59iZM17bFnpJEg_EKB;>-lNXU!vl29NMD4|FuQbO@eyo6Gi
z6bTh!rU*;88Z*^WZpKWrl-n`WF6AD~^hmh?m;#XK+XKuVDOUqiE#)R)nxxzUOpBD&
zz^J9%4oth0JAmnsayKyDQtkz&SIW<Uc`oH4V1`6)474+TkpIU&$k}HmH#4M3nuHpb
zU%-Ymq*(rEEdLoB(gex!^&qbYLouy91^H7j6y0hN<bz;H-HOVQ&=hW_)NN=ilMdlF
zN;_I2uATp9XTqZYGxc4#e(mz*z8B5!<rx~T#8ox7SgM%i|M#<mMhUS-xZKho-JJZS
zvp#8IF{wS?f#Mx#uL>P=&lUWUYWWT6*eEmDC^M#kayNI%OzgXI3sEK4lWS`YnybiX
z;Yv(LptUTqR=5q*4wf&$MkRP66=3;e*ytFZNGrF;_w&zA|GwAf|5^OfxFm)MH)=W_
z%^ajdxJA>B%=DFMny}5vT%47vd6}+x2!AU2f>5aV8ZN$uiLVjjYl!&DA7$$5ji;W;
O9u=wz36BVyj{XA#mZfe0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear.hpp.3FE8920B21D2C083.idx b/.cache/clangd/index/contraction_bilinear.hpp.3FE8920B21D2C083.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8ad808c80040e804fc5fbfdeb4b14a29e17ef2d7
GIT binary patch
literal 8878
zcmYjW33yEB_rB-*X2fuHh#6~Q##*Fe3zs?V1TzM;^+)__r;5_5{A-P+6<eEx2(eW}
zi;xlt6@=I&Dq{JCR8qStVqa<x!vE$O-<&VcL*{+%JKz1jbKdjbb5DA8@pzK!D2i)%
zmxymiO!&f}C`u*$-}njRM(<~ea=x0PL?jmU9Ta}SZh6)-H@kIOuNME^dAy#bU;mX8
z6Ot!QyM=XLCVLM34Lz!MfBo=}?-JYC(o-trcgrfvuO?1gkU7g8KBUB%b0oQ1-pqdM
zI~||Ys5sBR^B2!XspJ3mcmFP%8f_bxd!bj~!DXw({qw_?;Lo+2*^|0I=&G%`UEDr4
zGjDriTl%WyLBYEWDX;gfoY$sH{N<X#vFqb!7LTtoWlhHLQ(+ahR9-#tbV}j&lovNo
zrA+vw<KW*u4f=Ur_VMmJlKbp?m$&`L<*&E*Xy%F-dZTUImqV|&ZMwJ3`pUyCS*;I5
z|9zwS{i4WiqYjU(HejJ=^xAT#wzDH`28}zkYxbXmR>V#D;ljg$*geVrrd@d*lDzD4
z)0oDU$0S(u<NqCXaYBQV%Qx?vUWc`7_d$Mqr?CNR(vHr0GqmnsJC4oE&q^5MUodWZ
z)56$-8bdyxb@|k>a~+==$NV{>PD1j!!o|PGE*zeoReMUe;oU|yFJ4+MD=PQc?J#9U
zw?U83Z|r+As@tV9D^9Mf7rXdSyX%$OJ^s=i|2k@2|Lr9aFOn+TCm04#zh!vjKXbr_
zk`Se9PRY-wqAPw}Cbad0I&<dN-c{kypt55-G&9e-Rm*+X{_54~Gl8k+<K6{aF;;Z2
z(A|?^PLw_BLHq3=tbLe~tmY*>=sm5a_1afG?>SFRy42nA018(n_!$dA{K`*#^uzys
zw>t0ApgRK&T|WBL^@C5Z-o8}%d7B@)`F~ilr2N6MYr5TtY~kwt^~(V3ok}5V;$!As
zEz@T}r|ORHHs`hLS~03?gJYJ8D>EMl2bD9Izkg&_tr-6T8#o?)UlcUH%T&Xgf}dBH
z?{j`(M9}NcE*v?zWJdcc)7NesyuEtJpgr9RFYG$~{fs%??slI%!LzjUPx)u6RO-`r
zUCXBj&mL@CUp;$h%KYuUsx0+{D?@TV-)z^GxdvZ~e)q%7SMTCSZH}n%2~28xJ7N5k
zZ-)ixk}gxVD79Yh=~b#yqnR#l{YT4U#w<2MF`1a+XbKvbGE&=nswG!Zl?H<1Fgp4v
zCe`RARB4))klCSkfT~m$gwxD;nvGUWWx-3SWtiA+g5B6yRcZ;!S}?AKk&3At_fqP5
z=AJ2k^V^}S5-2E3!MIeWG&w!@&Hm>Qt|}i1%6>5J2YoLe_<2`pnV<Dds|M3ssft}t
z!i-`1Y04YCgpbn$D$iK|T~T>fvALop@^XpX|NSqetTzIkx<m|1n^*g-sx%R690%iZ
z`8J;>HTc|rL+v<KsUj#ZnenB3<gWF4^@&|DsahFUshL@e<9>0x{`+6zMm-vZ?<^Vl
z%MYqjN31f|I9A$huP5VjubzAZsuCawQDBUcW779l?~6qr9WPdu+Jf?h@e65<FJC4V
z)bAdCP*p5~aG4n|%Uc~8b8YID4cWn}VilCT%y?I(jJh6#xw!L~sx%UmcVK)6`i&}-
z@kVyil`n^!$*FlpRT>M*d}f?qDkVXQIZ>_bi{YyVs!Dx9X<}?5Noi3-HGJ^)X)&r&
zPY~=zyF^%?6?WuE^JC>z#U==yjGd(8CJg!Y-}*OaI8>#YAZ%d94YHuqQg(m(wsA>y
zRS6c9vtT?c=XcB1dvg*Tdz!0CO+lGvoF=JthjM=Q*m~<eSCzVgP}x{n+Gs{x&-FL=
zgbYxX>VmL?8F$G0$Qc>&(VZISb;UOnl*eFvEXVZZZ^h+@H@km9RjLZgN@iRsuX3m6
z#3cn`O`54nkf3Y@<5u}JkF{p=vJ<Wj_GkXO<~Q()8+gF`UqXs6#&)>)@>O1vs(dI`
zxdFx-@?Mm2BSwy|Ta!G4ncjOvY}V`5saMgiYO@TMS%wMXe!@9)VyR;gEt6X^`6xjv
z_iFX!*<D-pZlh{7v8KltS}gzrB-9WadVKoO8et;|wZc{&pSZNv*jhpzu!BeJMAcmA
zl2A|V>G4TQ`wTym&_EpM@!5$s2nR`MG>-Q8Do^_wzn0J>oaFIUwl*0jOK3LWZ1DKh
zrOg4HBjbgD3&G<ndu<WmA{nm&Tm>GVRcgNg{vzWPz!dQK+*aEJxJkyDfSKU&Nk-cZ
zxLd|Y0FQvj=d)T4V2+Ic1pE^`K7-WG0G^TY-++IE$5&w5Rlut<z7KdGJU+2$4*(y?
z_!ZzQ@c0Ttdky$n#y>IqiFtfxq(w1|lJOFTOPI&!MA}k@OJ$tIFo}75Ri&+CxK75u
zGyI)-d>*20Ww=$w`xx$H9$#&1`x)++@iB(Sn8)`})s8bfF5`0y&oPhhfvKHmcwWXu
z42zh@XW80KhBsyWgy9qB@%gp(l;KkuM{<nh9^cbRo5^t|;YwSgYfRjHC%dAm^}=3G
z68;bz;v}h$!7)w}_Y|DsBsu@e@K@%f0Ip`ZTE=S`u9a~r!&K&^fTl4_lkpaYTV$Nc
zFq1hcz`GglmhoPOdu4ot;SuJffafsGk#R1=T*4yox*>Tx6tbj+2FpT2gs>!0K|0af
zrw}oZTk`lsK@_E^6TRIE5l?c<Ng_5dW-JN#CNPYtn!lgF)!S((imGkmew(-ti_+8A
z^R_m$-W6`SLZ&Iwr4vgV9HJfMmV?}9!J^$^<=!_rqAgLi6kacde<ThfYFb~<+w9Pa
zSGnaX-T2cL2jc!W`(6e8AWM1KrMv<iWepEm!`q5m3Ki-rdV3>UahbugjDr2?&Z$p^
z7zdWrQZ<X+;voO8ZLjSh#}Bp#JILeFjt+A5&i2j@l5-z>ABTvKstvObbC8_F?colR
z^K|=k2f1-H*rUNg?z;@^%fKO0tZGSMPXdR~hN`85Jslk6rU$@&030HPs&)eGC%_>h
zt7_-KehwVujzwTE0tdO_6R<x42f1G)vqv%qxm_%?$1(@GTOzY3G6xmoMrPm098_!>
z%$~s<R7}~-p3NLoEcwiy&m5vCRP740Uttb%pL@)Hk2%O~ikZEbIYeGny_<E%@xhe;
zL~c#wO-25N(xBY?wpZl8yA~f_rOEt-231?m8!qQ9X@#G8!=GiG#KV&K5V4}Ft>a<q
z2#Y-G3M=iK#rmptm|G504Tyy5G;jAz#N*s@oN6?zhyRDS7JY0C)P-oWm?a^aYMUh?
znu5)e5KU;7glOt)mV{{PW0r(y8fKP+XbLxr1E|_`({!^eL{l`F#lBT-8JL!VSr(!x
z3Cyw(P3d5kg=jheW?6`)6JVBwXgUXGS%{`0Fv~(TJpr>UL{lU)%R)58GP5j1QzA3V
zLNsk;W?6`)3}%*vXv$`0S%{{5W|oC$y28w|5KZ@(Sr($Hn3+k4DZD}ouPm%g#8}UF
z>8lWBJBwSgNESmP62ph?+j~uyVGeJf!#j)bt6H=nINH!k#=WBr{Rzhyf@A2q*r$F!
z@0%6fG2UQ_C!-Sw(23snD<Up2Se6h`L|P|$-?E5!k6Z518NSAd?^hpo7wBjBi{W3a
zBgODK!|SZ0*qf^T%kW>;QEWlg-ZFg4I*J&68<e?xK((+Q0jd^=fi9AMV{Ggq>1(LD
zNcwJcyGZ)~!T-2O`h#(>i=_Vze&Z5pQMCw+aFO)q0?q{&i9a4N9$X~;Ujcsw7l}U=
zFcn-R{yl(uz(wLe3V0M;B>n=x0&tP|{{j35TqOR7fDge%;(r7923#cmXok_uB~qdn
zBg19PMdDx2a6NO8_|qAtGZ%^f0K)^!MdClf@C0*__%ATLz+9BLTMTb87bWf)!)MGz
ziJQf77I*m)H~F(Koj3QlRaLDzR(Dh48ejuAC9Wm5bW`HmVLLY^t~+*jQ{wt#e>WxW
zEBwk$iJO2E+?2SPfHT2OiCX};0Nj+gm4GY3O^Mq8xB=WEag(sgis2EX+vt77@8I`4
z_*x)UE8u<w)DnqE>&4)mCr~jw;Fbqe46}#U@N8<{a+2QO{s{OZgwoa)11^S8+S+Qs
z)euTsO9M=UP_Z?A2Mh~XDCMM(VWEuAGCV8ee;EG5LMdn07+#a{4Td*l{E*>87D_pN
z#PE@fA2WO`<2MZ7uuze6RV!gwBI9=q-w~dZG-GP_4S(cRR<%@aP35)7xl(yZD*uFV
z8n>p&*SGS(t-JwU-^=Ul<w1mVc%2*>7jkQ%{QW}Ss*pR0{};FZC11b5124$y7x6kp
z@_Ns>^_k3n#@jyQ9caDh-1=O;{)z{_lGlqiSfdTnd7}-EXoHKsA7ik_$k!Jb0v8y3
z=l$tX$AGqZ^K0mKx|g@#%RA8*pYh;lJVeI5p7FkfpY!17bX}C1uG!K77b$ipx9p@|
zTGXdbD;;wYZ3nmPAetynomM*RBHDg#*-!cxmZQ_WBQGMp<(9WZTzRu~U=QU%u3mbl
znBOTTm0py=FUsJ<iQms?-T&^Zjb&ABC*V$Slgl0kJPdAf+0%fh!A&lE8SpZ=$z|^X
z-UT<g>`TCx;3k)y%Wy7plgq|4jAw39`|IZf9!r0@YM-8%H4N7<H`(oGhMSpNoL$xS
zFx<o3WVc5d9%XJ}w_-zjjJz{HiqSQ0xkfP;-k}pqCx1lC<(6D(8HL~Kw9@$>(YA5R
zHd;+=N2htGfJBTmSR!d;mUS)mk8443bM(|CFuw#jHBY(UQ*uQSIekU%(2!QV!!36x
z*&<Rp(K|XM;u~&xL&Sm|8|t=-S#?#v^$Ol_1%E#WEPY2^&%@U9fuaq+(aiYN&s=q0
zX`5`d*hC0at+utcO-NeRf~~<eQ3tAqR<zM0r?a)QO{iMc`dIte=#ewbI?P6ooN#Nn
zjUGAEt<!BHGkTi_)@ZPit1koVGO&@WCxJByY~<?cU`+=bY4`wG4}gs{d;+W|z(yKA
z2i9|76OvK2BCr;LjWqlOtWUruGNEdb%o@pTq~Ta*jb%2WVO2|H)<k9#2~o9;%({`;
zNV*x!n!#*BE2@^wtl7*aq^oNA%$m<^^i;aStXG(go;vrK^&Yd)Q>B<$i<ym{8Y#Sb
z3a=|XOT<AJhIe30!nnjOmq@(g3;K(kcXlk^>_jB=6}-b@THzMA+#+ec%2<9V7>{p*
z(s#zvKZJ-kzLJXnO}`X=xF|%`s$x|qNvRrElTZz;A)#7WOG5Rqo|8mZAL~o#L;O%e
zK^P>V=Gfdx@@#=EBou-n5^95OB;-V=lY|_Kp%QA3?IqL^J4&c4c9kY>-LRX4dSDL;
z^~Ju@G_D`^lh6PhAfXXBLYmZ##E}vjg`*@i4#!DTyzw|*LKAVKgl6CjCzV(vU?fO*
z7T_!y&j*|jPAb7zz*rf_0mjLA1>g#BQi&!2CdfDuFj2<IfXU#b68;VFHyLjP+$iHm
zfRDf_gsC@5fF<Ch5{+RP!<<y2%NZ_bPAbu4hRO2$avQ^K@_h0j!-Mkt@g&2OG+$J;
ziwrL^Cza@JhPRoMO7uCy=gdhZI-BEc?i3~Z_J(P7&xpa9`diozz#ZTeHL7YEfEhC0
z1-MJbdja>#I14ZfoaAlUfY~zs1Mm+S=K|)+I1exnoaBY408hy{A246Wg@A=Jz6f{`
zoaCLC058e-3g8tPUjw`*<J*9@<wV{Ayd&d#fcIql9PqiE=of%5WLykbEaTY>XUjsE
z!*Gs_e`NThj2ALoC<|o~!$mS)%y2Pb{cmMsz9==h@#DT6Kcw_So6`ThKU$@f9-XG1
lYChSoZJQP$EnBr#Lc}FSX+a-b(Z|;Gu`PY{4bBwh{{c9(*2Vw;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_bf16_compute_fp32.cpp.34FCAF706B5AEB99.idx b/.cache/clangd/index/contraction_bilinear_xdl_bf16_compute_fp32.cpp.34FCAF706B5AEB99.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b0d0e5f371266fbdb5da77b0b49cd4b16521c1c
GIT binary patch
literal 3394
zcmZuy2~-p38XX3SkeL7pBoH9VDr-Ox$j>4nBtXI@P*Fe-9Tr823!s2lsVL8thk_u8
z^7K)qpa`fGAGolIN)?M6ODj*G0@mH?LyH#QBz-@Vp3|I@d*;mj=YHRO%biGBK!7I$
zLBvV2Y<Y6#@)-z%Fv)*LW_oIyI)ZTM2$FrcJ}PGIPeAwn7u%lQC8t{A+jah4RKI`N
zd00P3-)2=3bL`mYqr|-5wfVp8e0(rT-Ehu$<)b&1!666>`Ks)8BU(Nhl^3(-@`J`1
z0bN(qyk~oVHe8`ovgS4VaWbgN!4aRcr|l=~XjY!ug^kWfgn`R^jD(-3D%~&V6`s9v
zt48J?9r7t2J@evI=6?PbyvllkaK+HFfNO47K0IoS2)Wa16JmdrKGyA7f0HKfaV_z(
zj_WO~D@^=XcUr-s>X3J)N&D$HlWpFpj|;a%?*7@O%_}FnHTYqHO!g(IK0J5$n4?6i
zc+@6dzuwLL^y+h730fETcHUoNHxitBig>&4$IflynA=N@ugO@$>O;pic0Q$zC0yG4
ziN1eu>PHRKjT4gFv81#=b989A*qrmwz4INN?hv!Dz7k!%=WtG>!8NP1H|}^0*o7Xo
z^pzL7Og+_Z>F&o0GCOmO&kOGJ7V7i-tGkcmeJ3Y+2Y;AiTtBYeey-muIe%NvYneqs
zK=_Tw`C$()uVkF6V4u=GeD1I$<)B$vwbi%VGY(|iX!N4QX^oP6?WIp`C2K#Q?%@pR
zbB|lH^5+fSb!M2B4Tt*~pGxsxKkPNrO7B-BI&<Z~6PdmwNSBct{&AVxm6&h-`KxYo
zbpM)`{?^{;I^p|Q9mYGWTLXU?5#+Qjx_fl-aGGmc`90^>zfFk7qj&AwaYll5G+E8k
z?_l;1_JtUP_UH_4PJR?|uh{d1n&8@^#)<W@MDg%>qEDLsae?p6+Je1<xvAo#Cz^AA
zukw~(U<d#3%&t}QQ2tOu%_es4kon`N3%laFZ8xVHTKN_xIG#wHEb?D<xh&HsjQiyF
z8+NVs1+x{K{lAu-`!>Au!&=r{t-F`ZzdYHyDka}aZk8rw^!2=b{h-E`x2a;|YcBzB
zvOIEhvWdu>(-`VAVb#UA2q+(!aH(6Bbk6Bc?vePclB$<y<GRybD&<(kXtE%-&NKT)
z@8qlRyPkW>YGskD&*uw0vyN1{8hF-AZsxccI^U{F6wOTB)x@iMmel?ma|?JGM(gVT
z7*gKE4qNH)^H%wU_NIk9vI6cnUw^aMpYQngntX83WVAnZ$Lfx^wN<Oexf`AvKYY91
zClBju_~z{Dva=VDdyNQuUX6(gw{J=th-vF9DUST_;zL_vt!?~M_4vGj+V4jG@%5Fq
zLoYL&T0{wJYwlkX-<k+)NK5#9pyBC<JXY42&7~JTDJOcynnMq6kT!A4x(>;=ca>b9
zN-?@BaXwb^1>sv(EJHGK<U|^ZAi9&kTnWhPn_?gd2tqJ<OiLw@Tdw?ZA4u<eok!ka
zXf#Q4X<Q>EK-zx8hSy{7gN-pXlB5l21}r5^(~Y(3N(QfP!O$p@)}iUJl>lj{4>@g@
zCZrV@8b;D=8e3NhkoIVgK2Q+P;9+PiNxRY9jFbRr@48!6J6cO>F*KZ{wP;#8N`SN^
z!}z4-&cTHYWEn}bcq~gLKsu1Vb>C=2cM*mzBx#I>ag+e*m2tTP)dq!6FmwS)(`j_3
z5+J=!oqa=-wzL{UBS@M<<LD{@(mQY2HcM?vJ8)zLNk`3y(o_PZ_l(!+A3L-81%}3u
zv^CAzNC}YMFA7kLyE{c-Xedc*(lm9H0O_(rHBK|VFE(N5Vv>%aMR1e=>FSopX5;VP
z*I;NINi%2+T_r&JjFc7m@QJt%Ll==WPQ!JS0O^YrcRxyPikJ*!DM{<`^emMC=@Hq@
zHjlop?=du<q_t_<I!b`_^H}ZD?1t9k7`lX{#Wb;z5+I$PoRYz(A&A=6cmMe}GsjE~
zC%8PW0L~a3MGg9&Mvy!NxI%acC#-?B5LOT;+<==94hl|)fmjHWfKQ<GKaDI-L~tYc
zu!wMg$(V}ig_+`n6|j;(1)P`%=1HiSA|R4bF&BUZ5*X9|{{v6TDW;v0GY95@l-yh}
zH;|I^0X~70TnGpWgmQ2}NIAuJP;wT)!jF=32F`wzoDc~8D7jD&>IdaqE}k#lv)fQa
z)?LgK3*c>dSdIRd`$Yy<EEg4=2Y4bXxIM5JQNcZchY0H7L?8$h!6<MaNjb$bQgTMX
zNTs_rz(%FJ?!aB8yAmLo?yh8mynWNjXQt%4So5p}Fax;OWE90tQ&DDv*-|QLKHy8?
zBRJs&yrfjpG9Z(}q=PCh7xis>!62hBcnov6TcLvDwH*H5eJD<_J=hZ1CY<1Sa3s(h
zpoNgBDExy;Wfqtvrc!YL4q_@5Pv9x0Qjr3w7^VVcNjZhzP;!>QQe_rgfQ!m3_yS**
zSqKAR)3X3Aifp>VpP<~HAocApi@pGLoM5t<CU7Njg2iT;z!u_!9$U`@S_Mp=R8m+O
zRNA|abz0CY8PvcDO|B*nni@`MakY5RyKw^JVm#P1oWQv_4>}Z1XmhoB&~f1N<m)RO
z5k|oY6JVk;jds9JWf}nhD$^JM0;Z?2II^Z;;%Q4F*+Dv&ZUTFS6FOWS6S&3DB*-^Y
z*er~B)Ff7)FVFg&9EvErs32HLI2>e#3fG0Fal#assx-(EII1+r8+fZUC<p{iHwc!2
zv@eBAQ*sOoh69WRdqBoicr|s%7#OQ8l`XJUSt<b#s4SH~@Sk3)d(Gcjxa~gjh<q~x
z4+9Z2N7w^0io)`#C}zM+rS?w1Nu~Dl!F-k42ZP}0+QW$^%c}T9K)F|?J+tXWt|nw~
zH=dgSZUK^>oVHwXCL=vD!%K0Iq<}wu$?yU}5~nZKiFO{lyxFz1VMbE&@<c@W^dD@Z
zJSmO*BqKq|A?h<UXfrjnFr1;Qqb+4h{N^y}QkH?9z9AQQczMtF5&Al^?L95rZD#+~
rU5>E{&(zG^(#qP_&cSJ}^E@}fELS7Gi<r#`kfoshUx@<Mf|36LXIYT{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp16_compute_fp32.cpp.038CE2A80837D541.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp16_compute_fp32.cpp.038CE2A80837D541.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e5dc6c876f8f7948904cf422dbbddcd57895a4c5
GIT binary patch
literal 3392
zcmZuy3se(l8XblZA(;?@B$$xMoA7=~;OC(j637eCp!h&gn1EUf$O3{Cx9X}z1q4O0
zRA6;22gTwOsz(8JRU(SIg0ccDiU?>I#I9PcqV6Q^&!p!x=j3M2+{yjE{NMd<;uKL)
zfPo;ARK@y?wA>U5f*>s7os*mW#d%`{;m{CdeRX5}(zX4-qUW<)`_76V&L&-A?ka2C
z-RIY5RbaJxO)6{T(D0yU<41GnU$+nKPc@$5GJ0U}<pEVRVu>uU>N?5)b~tTg{8-z=
zla#0{+Vn8nu*n&#m=&25=)3pIT2F7xrT9WW{b2q^qn13sns8b29A5b3rvrf(H<q5c
zcBftuxFq^R5_)><L+);`ud!O;yzo`M$x+uu*WM1ER7ZDrxkjtB*&|n78ap`|?Ew{q
z!o(Y;4J8`Yj$U44RF(IZ@&fmO{q|aB_2+jAx5~=ij~!RKlTe>%<$K|m+xoiA*Ll;*
z(|2@5-c8U9oflttTlcSHu465APg<?o-8oD%uQ0IJWW0H^vm&C<F7#gPjE9#dZ|+~a
z!rt;2Gs$j#%IzlWmUwQ=;#xtV6K<aA5$0pB{hLYe`HoYkuU84z?Mn2&8CH|i*hNvI
z3wvCL%3fs2_^tEh!s@e|3pEeLk$2Sr%ETLwTh^Q?X&rl1p7<PzwrO4L0#;Af2E?tV
zT=Tp?MEO+QbLpF<m$NS`ZYomq*m!V4gc8Vh6h$2Dv>!?Eb$j-BTW!Sa_6d*lzr^fT
zD;?~?mKb(7g4Z#en=ZRoj@Gy4%=fitZi+0*lhjIMzo*<iIc*zr(Sqt=S9hkmm_43n
z)96uJ_g*qFvt=`A5Gmibb@Vs?gWB4&@>t5vA9mdpH8Df9w5@5bv?dGwvOzJ6&ng}7
z%pFLbEWfpwGc(@#sD_tSoX)VZt#@`gTIw-U;#BmRa<4k!`u0AR4JvAEd?PC8Tko-`
zGCnPr6=1U9MEzI8&FQ;;_8*XXKe1{HD6u@)?!)Q!`>1tjUZD!Ht;Z_6UW$&Ix1@Bu
z`|kao)rvz~CmuN^TOQe<_|(C@-$_yRB|i9?S6w>v=)$}&LSDO%59DWty!NZR@kVp)
zQ*`qF4P&ZrLZeE2TgTHD-oJjtySuO^DZiq2{7l{U;ncwXVSkT?%%+9$FXjJym{GJi
zeotJ*_dQuA8*^WT1U-*?)?~#V?MS{FR=CSVP~zO4x4vPI(Wva*U{fd;lMT&EY3~7t
zPGwE|u{ZRYS}?$k(K=mO;oN-gs^i%dP0n6fy&#jKtzXknv+7>s+D%&RuajQjB6|Dy
z(%WlP`7LWV{O8sdp4poR$Is-|2u<pKIuY=(DR%zOE|-Lm;|t<i!$0iI7xj-W&)m{B
z5Zn0IqT2DY<-Kk1+*Wi?e0F-#sov0AdMkR{tuGH8Zz#uSc@~NKX+^?{_bYi?so$ZB
z&q&9ns&8`&GVpX1LAVd)J)!52mvnRqg5X>#m!}6}$Mr9db?YWZUaM^AXdFRXQ>`ua
zFh!p{npN>Y`!yYnCuk;>$<_m;ecl#aX!}K8MMq-^noVU}=mFBfmzM0^mc$Uy(FB4P
zQAIpGKsv19PHoZoilcONK0%vN&6s+C^qg0Fm&>14`7w|c1kDn#ob&)`8Etd<u=;8l
z9bHJ!bSj;r2S~3@Ed2Lj>(ZxmbRI#|s5F)yAf08*?lh$?KTJo}1kIsxEc5{B?RVUo
z<*xfKW5_Ckj;F+%>H*TlqYYMvPHz~aqe}@|NEPz*0O{T0D5J!CpKv-FL(ryFQ>Gpu
zUHM(T&$O`CQ*?AOL93~1jvgR=`0SAV=&RTDbTpBm8B~Ua9w6OV5xmTN>+?tkl1$K+
z0!t@7K>D=YGVbwHNdp~SM9>%&W9k9Yoy7smxFhoC3}iV$PZvyg(gUOi6t^w}-@fuA
z9Ze!=bE-L050L(Eg89DnC(hT=(PacJp-OmqfOK}+${a5$f*5Un_2dz&z}^VMxdLtw
zv>^OMX@9UFhO4=1FL)@>kS4T07!kvrfwL4UVE8OBOA4oe;bI_`LbGA`JTOlRJN5Y8
zz=WL6y|7aZcK{ADQqC9n%1F66V2+HGiw4m$C<hHs$m!fp$~gkZ2vW`u_(hO%;UGMM
zl#2l|5m3&*_1wPVoioJ5K}iIXAb1--Y()FheIbhLmW3=X00J>t+yi)s$>PBvSPb<r
zTn1!fSOxAEA*Wj|QjQC_;iQ}p2*XJ^5fFuwauOg3hjP+&8JE5~_LCj)TZ95(5d2DH
z>HmISb9{1%35J_;P5E#&Fx-r5#)k=j;dCyY4>uaaF)qf3M#gY+t~nnX6Fx&Uty_7r
zX<J|`C!6*HUUJw0hKGVsIoY%VDCDqdW!1&9+grvMMAHlb!vU^dOwgXS9Iujc6vNrU
zY$=>NhI4{BQWzaDVTe|A@j|vT6U>y5t#|@Y3E4^r2$7Jj$bnn}TY<8KoGy+?IVa#`
zU@d>(Z(yytV6K6+VnOVbwP1n~9@51Zlq*&mZyhiD42&_H#b)uL2Qb`{ZOMnxg5lHI
z)A=weVDp5MF6p2WoFvh&F5%#94Ce#Bfg9a{yMY@400wT10#Q?L+!I%S;+JP<HN=tA
zxHLXITnuM&nS8j#FeHdpbg>C*9^@xjd1d5(B&Hbe9<PKy9Zr+zP?w_cX$-dmb_OTv
z4ZICbGz^3poTw5gr%n`}4k4#YT~dw*cn1683S15LB@hG}?28mgr}hQ<n@B@lN|SO7
zM}{YC9nOHLsmp5eAsb+0P}2>#8Pp5{K?XG=LF81;`^`rjMLTN-iG#8Zwid&zh08@$
z(M3L4#U9ujjM4}A7>sf@m~Ajh6;Mr$5;~fglI|}8%DvdvZc8f*;1k6~0#Oj$0wg;v
zJw^9SPPQf|RCkf8gWq0h@B%?JQ<uhgV46kC-=nkJD5+^F8btr?-&jpXYC7?cjVRNi
zjVUJ7X{Kg$jA6kvm$Rf1E-aed(t5hp3@!)`4VyhDe6Ba!Bg8S#)%MTsa%}hlJ9`Hw
hXQ7+Br;o4SEK$(R0G^k>gw2UktVAOhh-F49<o^J9hG_r*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp32.cpp.37E0EC20BE4060B3.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp32.cpp.37E0EC20BE4060B3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e7d422e801b86f94a40f1f12d9f3e50ef43b32e7
GIT binary patch
literal 3378
zcmZuy4LDS38$O06X3k(Rqcdjm|8F9~J5m}mnei9KiVEd0qWm>d@>5A7DOR*8wb{0i
zv~5zhKmBd3c1wNqza<o|<SW(M{c_<uGvD_*v)8qAU31UOxzD+u=Y8M%xku*j=jXsc
z5K)4EcB*`RoECy0OyZrnUXk3Yi6DJC2$Egf9I>qYxrhFvDV<x-n$l!f^<qtT|2SHZ
zP?&FIw<dvkTv}aRKKVu0`sv4<-P9LzR_(7%-{IE7HL~2Qy=d^KZYRye?72~c<IbK=
z+eU}%Yn*nJ_t<pwe?GTGBX6^HNTe~f>*2(U8TYreA+Z<t^ABsDo)t2C%cTFoZ+&b!
z-i|8{pph-t_!|Bnv(5!9?b*@U_2$n_uC<1W;L)skYm8dj?pxm~l4C8nNn}p3O<ZY5
zA<JXn!pXDp#+wO+esyKx&IxS#&&N7H@CJLf$V$6iGKajK-dkqJ$^PZD?3)JeNT}Rm
zW{0fGqvTeS;UoTmLdWgPXU4rp-QQb3KAzB95qiB~aZOijvee(;j>glvGrmpKfo~V)
z{|7CI${(T>(A<vKzG@C|q;1*|wYLO&awy}{Ry1=!K5TR2j@jxx-I$^S&0K%)T);@9
zb|@CJW}~7ww{J+ibNzL%{-3&T&5o$z$-_$nv6GI^E8_OOFsWSorog^g_t+Sl(|*X0
zRut+gni#s;{vhdzt7Vi8NEYn;?PPwEB(TBgk*wF;*v?J6l^sO$%C4(7I<%pfwjw$#
z@vF*3+V@t)6uEAVatxl4db(7;?EHDQ!sa#8CFwy*$KGE`s?0b84H+-(?zk0JJ`;8O
zO48F!n)zqzCPh;kwL3B*)>idhX?)&#Jf}4*r+(G#GTKx3v*ljDDz<kI3OFD3HIFuY
z*jKV+U82KUhs`Gk9O=b!>tvjhQJkW=!Z<C^PIfS|0rO{1eQW@`E7+~yc+B7|4<BCp
z@$5RklZ>X)SJxc_j>lfyEN@Mz7_|^v827TT%qz1va;-D6MQ5VergdeYmuVXIz3daC
zanB`ld7dcK|Lpg6MJB(Ck61QD3nx3UKZ0^a&w};~-!<zzZlBWc8VMc#YpiZ#*U@tO
ztTcVEv0Sb3q5PGuzq{evM`L{IPEq}`4o)_;45#(A<X4v7Uy@jvZ?pH(h4$gm6Q%fp
zT~Cr?ucu!ge;Hg$V;O(f=grwu6W>`^d}HTL7mZ&XIEu~NOfSZ)cHxBF+b($isw!>U
z0)wE*L(GWr$F;d$oC&+64c3Qel{Hwr%=*~&ymx6v^inrD@95k-P45MKcKWrSSs9z}
zDq>eW3%}HneE!3msH%_=uk^=@tHOnoCH;D9*DaGB69?WNqrH7ADgNuBQGSYd|HD57
z$s@U+wc0%s#&d4xiQ1>ao7^P5)sJi*usobsUpkdnc9#!Mbu5W`7xea*^_|~U>aZp>
zSJf1?A(@*}@e~SzU@xSP=C>gq=#(e~!7(bvQ2{aEs+VI{*873afhKf{jG&FF#s(^w
zrcXC!?0s^3Go2Da(0WuomI@&4{O6{NSH`3@bjlKfW>H!CDuA@$YE)HG41-6fEGKAp
zsyjyokoG>&dARt(-bOkll%RE}I(jOAw1ier@h0qMIi2zqLDQ*pwhAB}_2tFG(Z0}r
z1`<oqhCD+n6+k*EI`{EW<K3_6ln{cZQE5ySKsrN{^_w;|_9&eaM$l|3TVDl`F6p#y
zmD*KbLy^@49ibJWtpZ5z8a!cCcP8&0owAIeZK<{#6+rrHp`S){_asiI1QWD2Ra;L5
zkgobVFVu7FdLRQ?MbHL311l9kdf$P1=b7G@f230)2|A1###RBOkDl*0AN=&Wo=%A-
zXa<#`uL4M)ks8Qey%wFIQ<f4mN=5Zl0O>mh&Jo$RL01^aN`hwcm{ux)^h^J{7X|lj
zoTF1>2wIn_tEU1;4=>lP&i>)TF*;=hL5rv&jtU^HkSAtZQ4vI=;8R}@bCbCSieo&?
z6Iu{{QfPm(Ac}`!VSKo1Xvi6~zZnt5t%0=|DxkO<a1+B*Kye`uilNz1JOqS@VW*D&
zH!vZmbT8}_#Vvq^gp_juE)r61Ay_CO<$^$v1j<3f6LLzolX8~8(wCH*3+DQgaz4Pv
zmy`<z!M;$=^>SPFuF_dTVxvSnkte(iZ`Ppw<Gc{XmBT_7=K-FOEba&#g=BF75D1|j
zic5e*2&;&*Qm^JTy)`3>+wyEZ;UjSDh$_mNBdeGKQz==65BO5}2#U`K^QB}Jf8Z~L
zRRU_-%kLGwV-QssJcb2yZ?I=&I-6fsK|yhrfF*|eisEbmTMUO2jyTbZa=yt{On`|G
z*@``|_aR&H1fD)*D~rKmAJ__v1VT=k2c+C=Fk3{*IRPgTDdz>eM5LS)NJUT%c1g%7
z^MjPL0#<5Xa0RYvURVScsd-@uSTgMe7&*kDEAs}*?F!J`K2kmfG*O(%Vshd9qPPLe
zfD4lZ#SK}8T)17Zc|u7UO;8D@649?RlHg?&=K`*peH?&;ntcEOYWDF1e$)1;l-2(*
zHgrCICW>og+FYoP;yRcP7e+gZ(=ZwrZaj+9F*+9-48>6l<-)u|aa~N83-brX^)Njy
z9A3CgqH|?hlAUt^N6iOzz)sBv9>7D*2Vx+e_5q9<Vt<tB3gwP)mmBd@H%<^vjc|wv
zfad`nOB7dzExa4W&48KOG0y>W)Q;I3c&i<A00@{qX1HD=1eM`T$}ub%POug{Wum4s
zqRE>ufT=Cm7TBsS*d4g5Em#CZ(+hss+F<EkdZ?FJnX$lF2qPA5Hc>^H^kfxtV6J8e
zXW*=6hy`GQnjr!~;Itv&ToBu;{AYl2|E|7eN-LkoC5pTA+&$qeDdZ_}%3CrO@tO09
zPfeVa)mPv1<^Bq-1bJLMk^rAa5IOlZp8TqM`hP#;Qxj5%|7j#Z9;B%?gE~`NhmJDz
z^>n37v9Ar2CN(fNG@6A0!F=xp3w;*NVL5tPdf1u%(^a+!muF^fVP$P=@8IO@GS|)B
YbM`zA-&Mq7`}rqQ7KaNZ8iB}v0skn1LI3~&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp32_compute_bf16.cpp.4B6BEF9EB4DD9C01.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp32_compute_bf16.cpp.4B6BEF9EB4DD9C01.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cc3d4dd80d4a9c3f4e093c50644aed961100d888
GIT binary patch
literal 3394
zcmZuy3s@6Z8XbZL$V><X5)35rmbWMvasx;V2_!rfG>Q~K22fOdl!t&&A6RuOC@7$S
zFXSOY3knOXlm$gs*{X%=DrgkJ6(3k1Ac`Wo-AUTZq~F(kABXvJX3n{D|8q~m1N{9x
z=?Ef;56D=bkQzG;K@daYFC{g3O`A4?uxSXAakO5seB*s!(EFRi)x5$V&aJq}*q>j2
zKsi@woM~*oHs0`*+=XF2`D=ROPcrTW8?%dZgag=~fhe3(fRt1$+WqWez5V2MeDc+9
zyE&#lryfrXKI#JBl%!^vT<lx+Hva3_Im;jPM<eZ%=wM6k@Mj@q?B>Tu9|oEyzN|Ys
zd#g{#txjC|hFNfGpe8qUG;YK@^OG=0$KAKVp@Xq!AAT)xv85BQr?qh1(j#<b(Yl7s
zp`NpX6=q%TyM}4?{#TkT6x~&9$Dxlq#>_>v@vB<<k9m9d+m2i+uGrSpuqaD;vO7U*
z@9NW0#`S!UhIP$eF@f!QmwLZ)d>9m)bmg&SZMDvxy7G39f&(@!I!%&SRX6F*9oZA(
zEUU&ZC(&tZZ?ukwH#=4iU;F#z-O?MK>#Iy`hayJGBL%`<_Vf=`-C_0{T5D1o6lDtn
zwkyl7eXR6)z?8W<RT*{PxA<q{>+*`{_GN+5*9NoiM02w4U9fDut##gG%d*nJ^3kkY
z_O*tN8)|3Tjl?uldmjWI4-!X;t<GIHxPFsaSCvwfVQ*8%#T#sw9n_D0?jYUxx#23i
z&zMtf!_0BNdwVY3;;?d&uh|bzGq)<ebfzDEfkdPx_6-F*l?F2DS&Kfd<X<V3_&u>)
z+gQ=Rp)#$yamlvl{W}E)vb6kL?Hk{2u0Jh4(|j-S?xyeKQt7XDc8^jlQNJDRoX0`l
z_BqxSjmKKMkK~y*O<cPar*L{vLVaKn)2ZxQ^IgE&$9sj=KRps@wUo2nlHXL*ca-(M
zexLQWd35P)mowL&TP4y~T{!-t_)S7?RMF_HWW$<~>ewQLzqPSs|FDQ%`z7+$1V|`w
zuwGi_IkG!1m-XKEW_Z)Vj%|h81Iov3C+t=)aW`^ald@*g8JKllNbgDXe0MPIjhBG;
zsU-Z#r$&6+tg|6L<F=Pw?EL?|A<IF*cT*l7Ztreri(NfAudwmxgH!jTs{Ou6KjJUw
zj4NaHN6T)fT{<05!mr2?d8QvTI~4G9;Bxc;=Ko;qS;Q}$5wS^oYCJi*Uw>;J+E&=P
zX`YYKs_b1xrfB-gBON<u-g$KFW@@n8XGM6flXtS2hleaPRrx5AmR(n)%na@SM9s)P
z+B1+d<}~=S>`Btb;r$=><+o*?_HWvf)VI8?yD%sGy-D1OUOk21nnD*|pF{1#6W?5E
zD|?mVdQKd(v8J~}(lahQlN6IZYw5d5(RW@c#xa}Y7B7w)Q#K{=W_Vmqj`+{M)C{+K
z1Hq3=mPEZ661f(Yu0&EY*W*bP1mX0{d%eyhBPb;TL2wS0W2y$?l&e2ZX=!hsz6&x#
zDd7ZdLN#HkVTwLmyQ%PQ*H)CGAZP}a!BPXHT|Z>Db&ShPQA#L5v#2ZsH9%T;F`{VK
z3OWy^L=rTg$~RR5q`m8Uj_>)guok5(B4|CT9zzX~meF<;JPG?(K1x|a&?psUs{ztM
z1s?vTV;f%6k(C6^<S}j30O`c2tiDQ<eM2Z^AwkoqG($B&dXqNmhAuU_5~YL@G@Hsc
zPy?j@*5lA3w=eF*kktgOn5NKG1Eh1G)ft~`%6^GbmJ_rc)y`B6kUk*x*NVFRKOCim
z5VS5;m!SqoAKD+j>)GV;D|959ppAG&Hforn%W7P8yxSX5$})luqlU570O`tek1d~#
zy{kbfQ3Oq=(hbxA=_WZdd~irohf<ajG)Bc3YJl{^fS=oh-IxD?QdSVOK2@Kg21x%J
zsb8FN=ErK3@)bc#s1j2(Kzel3FDD#>w{M{%F$8VMGqh0yq>~d?rMOTLL~F;`e+LXR
zEwwP5!{Z2G1mS-Q?N1TJ@Gwr8%b`Es8U}JY?N33(a9dz2g$fw%4&0@13K%X1Vkrz8
zhA#vQrSL_Z{s=H3r%EqeDTZ4CD;X&_2h5R?az4OEM#=?)U>THyfhXitX(#2ZfweCw
zHy6zHCFMjw<V(thfDm6O=hl9{ICt+1F|kn+o<sn@hBs@`{_?yK=BmR&n)3ipOqx3Z
zCoyR*1VS;?!*CgpiJ^-$ZT-b9-@mXR%<Xt~0{93#c7%)S%#kkUz+6tcxBwS9d<4V2
zfR~(f2>=0d=n`0ZA-{X)OFH2~=h3ZTdg19MTvSJwbTI{{BGSbk*o#OP58xppU8F!N
zf-WI~gB#f{c?A><X9-zSI3^5d3)xb5)?p<Oo~k||J!gWM64KKdI7>)RPv9vbJ>@_y
zfu2y7kW=*sDQ5$0G_v3Z+%&SV04&hRLMRBG$^xt$V(6;AfpWQl+TXs){|#tkxFO4s
z3r`@1Gg(Y7+(isGVi|E^ufW9<N~&stN_m@@4T2WwU1Aap4uc0z1}rP$HB?On@5XR0
z;A*7G5jbk33IL#yDu3WVm8yf`HD|`3pNpenxGqPR3)L}PkE6$hjgR3p4vh=vf#E0z
z<-*)yIL5)au*NW4pQF!(tp(qZSgWcw$+emRGmX?Z00)iK2!KE%HGaTvDmAckh>cR!
zER-wfMjE@UPya$}l)_OF2nPeBOPH%#7v7EG7QjMd-m}4Mjd^<mZ;g2eg21VH!#5<_
zP}S0;9Nn7k4888OoVMogJ@$xLiiyxf49gMj2eA}Yg_F-(0!xjVxdK;>nav0DHD(qB
zf~IB$6HV-cs`8=S>*A~CwETHo!ko|J3*ae$5hTo2zX0%VqU!-iV+eM@PGbmsz}Fap
z1W2ZafFvg*#i~S7lH*dmR2T6o`2CXrFAyYd>QWnLbU75CJ=Jt(T6{um9HM^ucQ$T)
zd=l}CjRYnHYfqa_)zQ^MF}eXmUv4P%on=UqGfj+)XK;Yf%X_|$Xu)ijlc%+Zz4>2#
q%QoZkEG(^TZ0#Hzon7b5b>|Ca&NFp!ld#zS0jnr}i^VdnAmsm&JDKtT

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp32_compute_fp16.cpp.F92204694821E2EC.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp32_compute_fp16.cpp.F92204694821E2EC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9f074231c657a90fffc04ffaf639848291fbce3e
GIT binary patch
literal 3392
zcmZuy2~-p38XbZSkeLv+fC-80OMtLx;13`W5+Lk?fPzS665Ik-HgUo2S)hU}vg1=)
z6%bKyuNDj9QheH1eBk<UK?Frm6qN#M)i+7sPttRmb8^p|x%1!en{T-j6%-Hvs0bpI
z2CYv`&P<wxAc!vU&&W())uxUhh7<%@zqchO?wk8S=gt>S*RJxTXP0);ic4B{_qp^L
zY%o~4TB_S9=F&_)zR0${7Rb6_Y24||@5l1{6L3-~QscEKZ$)Q|?S~^dQ)78H&P?yd
z!S{~`u7b@qnVXC{?=E^PO-pi)d+;z3Y5#~mK2tdI2cPt=XL?|&|MdG$vb~O5yu*4g
z;(f36N*ntR6=uGXjC*bPzz@-w>(|QZpye0g*C`9l?O8X@l^e&_a{DS|t&`y%)}e8X
ztJRd7y!fI5?v7|id3;mpfJXP$KkCauejWQfG%(`0-Kf>zQ+d|TzMI-v))@<Oc3$`I
ziI)tv3C>UD-a2CY^5=#p?L%EvUUSOgH?D0;?~|#Al+{oAu4~#iuWkD;)lU5#ewMK&
zZN5KT+9h-HKjONl)yp$Qd<A<DyD`!?k56mPDo;v|TpKZp<vyv}<S`uIpj9!e=_sej
zr|mL3ZQ`WJW8i1%CF2*`b_K-;I8Pg}?&`716V;!L%k0_PWCy1jw6vKQY>LX5`-iGV
z(@0<ZWy>zN4AmOc(=F4_Q%<nB`=dMpcjhM?xNiC^KHX~Mac-U8MAvKkl(ewjk-_F>
zU`H7JCX(JjwK{##t}gjddq!!-VeM`H`Rjyr(tvu)p5q$1wAh7v85PlU4j9_^jk(1y
znjP|c{-gQp)zhM76|XjzpOyJjH@5s1d%0=}`=D&4NoMh|=C7s0E6qU1r5)zLoO!?Y
z#nEfiX8%0q;8j(5pm<d$_bo3aXKz(ay<2O|+%nc@{RQZ1GYgqhs^+cJKi;`Jv}&*=
zAuf^L<NqIn(Du5~%G)RX9Shd$|McC$>0amHZA_o9L$6Ml6yDmx9m(12y4cR*PQAx?
zo|TboG~_^ol(YGZwOYCFQ}k^_chhp~#l6dWC*#(-6>oiA*OU32{^C*g8jlH=hU>pe
ze)$}_tUBjqNS<f-uF{8-QO&ER(bXgE-L2=7R(#ClNcca6W!DD38-3+gwB{dydTs5+
zP0im%mOINtT3a(;dAO%=c=l2&k?ZWV8{bO=8j{=-?7C*D=@s2AU@V+`xo0}G<_#}=
zrNe{msUNgAM(1Y-+;q7%zDU4r9sgJAy?d;uJ*)DKIpc@-tRCiPy)b_KCe3>*+I?)>
zDOuI2zFzigW0UN63yP)h@3o#<x6g*(FzVFh-J*ZJtk-&pN9)1`?LHq1vU&H16W8qM
zcog2UIlpeKB=Nr;)1gO4rnc&3KeO$4+11#1>Dig&BX-Uyc~6fwE*Tk@M)ijd)>I^n
z4*6W0U0ai}Ar(&{AqcfJEhzPScL^0)h9I~eThCGf#EvU|jECYop68Hpl(d+j8Ds`i
z0g$%O%6k25BE%RaMG>?S*+@?TGxYJpYs>Fl-GY*02%1Kw(G>t`r>PC+J0``oC@Gwv
z>14W&0wB%rTwIa6l*&d)@dV8y^Oy>Nv{&<uy8O2C!zgJ1L2Hq<XbOOIAZ2^$)5t3&
zC@GqtQ8H?%07$P)*mU=xQPB`eiXdnTnWC!zNUv3=|67xscn~E;60{-NP)7lf{_Y0n
zjM%pFB8IFW=$KhCnhJn);c&CT{*yT`QBoX1+mLOT3V`%(L4aDqt<N}03L|JuvL;Od
zkgnQu$VtPi{RB!{M9`7sNJ9lc`rz3?v*C9Whfq=iK~u?89R)!8q*yQN@sO|?CB+gn
zM#g9gfb^rF-_G;9FaL;=mJ+l!S(~N+NWX~Ju3UettpO!1A!s34$W#ENU)`o2jfj1&
zPel?5TA!_NsQ^e%>rk>MYh5ua@-;#0vUM#L0O|DP<r!Quf~akO_oQEUgP9tJGuRAw
zIAd@WzoPtQ1TmZmm_G0!47UZgKCr+T?grd^;Gkf*2#9>(E5H*Z%3nqn!y_4yTv%f`
zz(h>>^ukOr+zMEUpaO=^1#?Bpm;xXWDPu-}2oa2F|Nns}<mA(?lq2jgP$}mOoCB3|
z-oQIhDHjSt1ECyT5JFDA9ZES1VBx2fa{(@XN;x0k<ENAh17Utp&b9qqWnsZ=0a14$
zTj&mN!-v%<f4g5qaQSj6gR=o!pbTyg>;=l;e83k#Jq!;7fdUvsl$F}K@yMtt5!{At
z;|`Db7rlM6y7iC6#BCab#)1We+e-wOudXt<2`~{WgL45_43#k46L^Z1!Gl1M7zPin
zy-?D<<0X}dLS<9U;R=PhmwjW%-Bn7$a5|qZg5}0=L%yL1S^#tkA{Dt!C{vjO<_MLk
zH~<HsG8GTtAylR!24W#h1<Dd~a$8Z#SprLyp1J~8m7e+nUzMJQgYcQ2LMI}cC$}Le
zR~W4R?O4edppN0XbX^u)Obpkf>#^XXVYohBp9OasCQm5I9SbT!vm)}9I~BZ*;Vi&X
z83Q|Dr!ocr0F^Na00A>&P!@IQ*yQuG5;BHsGBjCG9mBO4S}bVk7*1hOSg`OIjxtad
zoCplZ7#It>9ENK%v{}&9;8qi_Ah%7J>kq7W18!>eC!+f?b}_-QWN-?IH<P<4d>X?|
zfvHL-9D$=sC%k}{N+*Ir@JuIQN`&9YT~{ebwV*n{ShvqKS?~(>4G^gq@r?w~e&7ob
zsmLu^dDaY=sTA1>IH?qQKA5jk<PZ=tQ)D=6L>=T74&`1|c9~F0=CO$2JT}iA&L<oI
zBDmbq;nNsy42)HLmjgH|zUvO$ReaYU_|Nz*oHasD{smCVF#tnF&IZ`1$ngMAMNSBW
zGjd3Ja!Qh1BqLpt;VHk6%HhvnGQ2<#$;_oXKDZ|7)-s?zK1-UMBtaCX|70bp(iGy0
zjRYr$s?Yk0tf8reVpJWPwpdr>XRS*S>lx`A%w_<-r`LRMA74khy@!RHt;yfrH8f_i
mP0h?Lt!y}U4o=Q4b9wG_<}tahLb_o<&~lRhLP4Nf2=YHm3XjSF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp64.cpp.D740E9866ECDD95C.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp64.cpp.D740E9866ECDD95C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1d5b75057c7b30e4f14a46a502a1ef18172b339d
GIT binary patch
literal 3378
zcmZuy3piA1A3u$DGIP#gXhy?~TywvKSB7z!W-y4bl!$T|YNLEoblLifkW?}eLP>I2
zAtAM&O1q_tiauRcOE(qUDm`{nn>{n%_d2uBv-3QD&+-0de*eq+`~OdfueY}a13^S_
zzUfKQ)R>tFf-s4{^{FYV+f)$b84W?w%T9(ZTQ}gWac|1}O0Mqx0h_^D-QD*j-yaP)
z;IreHH#;V`|HIg+3)`}b68Za1T|O}bla1d@eswqlpQ$n9wct>x^de?U9oahda-Wx-
zp4&Xx<*}gw<4wgIKQFyFto<Q%gVZkS$@c$HCOa($tM@#7W{&K6|6*u(MeC<eu>}q1
zZELP21Km`bq@d|^Sj(i-)>4}bw`D_R`H{IzRlz5VFYmIMZ)0_%m^ILG$JLSc?IOK~
z=K1%f)z<zd^K>2Q(z4A4r^C%`zRz0cJT)N;Ii6J_8j0C7x646R@b%g`@yk`&7Fv8s
z(Xm#cqxa$FwSC;>2K_FoTM}wF-9GDNf@WuGs##qxoA49)6$GhWRV!pg+_jq;8@iF8
z`t!c*MvYp#*z2^XwXMhUwX?zQ;vG+D-RM=jglMO=_6=CdD@nP*p&mP*;l!%`id>|y
zuX<xbT8HVrz=IOg28RmQq)}h5hrv$O9YtN|T4<d3hn<FJm*^GWc^mZIy2^=x?~Vtl
znmF|DXMdTsH_O^#r*x;RNOra7QT9g1OurawmR`yZ`;4G9UB{1REZywaw>qu1xr-(B
zHlbw>QtpMBspPcX3GLPSA-t$<%Nm1(g%wB!`(}|~cb<fHEGCdXy!V*E!K}_rY@OOV
zaI&qvH)~S;Pd#VGP`zGo?&tcoXG-7R^?p;nljmR&xAed6Z>(-GLyioux)|AhMRqCD
zQ0GcXT|?s^b)~je+@iwcyX(4mEsVsd4(m!uKy+oi@v<)iRnHt3MsPbF&3k)GtG-Gq
zoNnc`#e^}Jr+Z!F-fv$)RDLy`X?S2yGJ0f<H?wze_=ks?Uq!OgCX;gYGgwI7r>s!Z
zr}_5hE6*B7gdDuYcxTwl3~kh(dXcyJM5jU0ydM3DdzmA@{B}9jwcq94gi9K4vN&Xz
zS{b`t+`Rwkgvn2CM&5I_kDB)1z4-dUiK&s%_QQYRM~)k|%1YvnTr+N(>{M&2H#>1Q
zXYVVMt#2Qvy?Yi~+o8c4?o00BZlSwz1C6t2;Y<4R#~r@(zKh7{OHVQAc}T5E)f(Ns
z^20BAwT~{mdfeCd+DCZW@_o-rG3$L}VfU|UXLdesDK#INvv<gH@cq9|r88{S{(C&K
zXuPW0t^06U8oFYi*QW;!<MB~qKI_G8=WS=$hfKo!98(i(jy9f8%UKc7_xXxL^pgtN
zmGh#-W04}KrnH^~A3v8zL^;!&TAg+`HgDrU-|#ZFVO`hifQ^^%H7R+Eez&qH3F^C0
zczIV3vVKz%o=8Cu&J)Q!*A8S1rGz60&Y^O26hO?k;$yjC!*8!X`s<>U5Q5gGYHKQB
znr^O3mObd+hEl=^TAixSQUIiFf8X@u`3Xr0N(m%r7L}!;07&yMh9BG&#o(co2!aMw
zprZgt3+iuIWS^7Op_Cwkrcr6?3V^f^eMkPw;Gg%Slu&|3sVG|kkdAM9w~gu_c9nsw
zBxp^Zrm+Gby=wWE`&HU|U!au51Wl*XnF@e(vI?t5l^R`zQi2JZO=W8+0Ma?P&D$hq
zg<TkuK+s_`!&DUj={>{svnyM+j-r%h1Z_$+)lmSX4+y<yEbsq{qm%%GR;8+{D*)0(
zr8Ty*1f8c)%2I+3rUtVW0O_hT&kcshKh~g><pj;3GBgwb=@yA*$lwc6JxW<Z&=?g{
zR{*5D)j9Lp*Jz9}kZ6L|;%ONx0Mbu=Z~n-?^;0`ai6Uq<s+zh2ApM*c{x~>Uzl?#b
zAZR9!X{-Q9k3^^yrZ=6dMk$d5EuxBa6aeWIY5aO)D(N{2{cm2w@L*1`HLM;w?;HBx
z9Eag1z(fdLi{TEyK?vK8;X)u3LRVq<Vz5{UQ?mTOQwce_O<_tHZU_v;q?{eF6O(f8
zz+FtrEdq<gP!2kqkdxb*lrsWG9;DoSFyDie^8g+mq+9?9@PKmgbO!w&CxD0|w?7$$
z2YBve6iZ;~PDbGazB|;za3A2~4x<R)n8jx{EFoULxQHimgvW4@h$!+YBBRU&bG^tY
z*1+0}jN%Ghy~rrOz}E}P`ITJQe=B2@K}2Ej7>3ZU0SgM&v90s+DHzV;v&6807|!Oi
z#c<rB)y_@b=wKGi!f;iNDi;n4hSNASF5CwgPUq0MaDXu!<)B<x1BPQ9j0<M~!__!y
zT<B>GSLdj6p%LJIAgY!x2eN7%pd%owHUnk?vT7&bBp|C61F-;B4bvgy<jaJV(*t^*
zq?{G7@+9S4fQu(7Cjk;qC<hxt$jO%vDQ66fmHcH7?3MiG2|Shj6$k>S{RQ_G;V=2J
zf^vKOR5E|vKjo}~;Y=2j3#S{yHCdWm*dPqoVrg;V-hjyyO7gu1mEcVg`O5bhJjQS?
z;3_%Q0$3<H)fqS|In^6@PdgPZ0V0Zg|3SH<+=$uMNoju))(W!-^Mh4EixI))>kxh$
z!}Wo_QnPG;jZ(7&K%mqtKj1grteUTldFBTk1Vj|T1CH>5ps9$hCEu!K6owJQ3Z?+t
zM5H3$t>l+<fv%FP%z?R*s}_I-O0Mz(Uem6+{xv3axTN9{kqU>$afCkVZOJd5oJ!k8
z9Gmh?9pRnaZL2i`x#dH|u{K{@2sbG7KarDsp_4fo00X7Fu?4nDcjE@!l<vkK_)p&r
zbO|9Re*utk9Kca3(G-{}l?VV(Dp3SP(<MU36IU&NVL-VQX=050d3{RkdRO^FoE#J%
z72GX5$9|V~x%SLBX-q64g#?1cl8=gC|L@<}q_{-lJ&pKD7pctrhB`}?hGGm2bu|f7
zyl@_qF45H1nmvc(%y$*IxqEopuq<7SoXqC_(<xh*%hNY7G&V6cx3IFco9_S?=sD_G
T+lyFiZ{K)|*SA8S8UDy$078zk

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_bilinear_xdl_fp64_compute_fp32.cpp.29832DEA4C5E2D0A.idx b/.cache/clangd/index/contraction_bilinear_xdl_fp64_compute_fp32.cpp.29832DEA4C5E2D0A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..251b6218bed576b74207e66458dab5ed4a57d3b7
GIT binary patch
literal 3392
zcmZuy4LDS38$OPjm^ni;V=&Al|GyK_i(w3t83rk)M2T|H(nmjR8{hWTk|aM75s9r;
zL?vvcw3Sr)`m_C@O|@moU{~$7U6Ou$XXe}2nZ2%cU2{9$+uYCdKF|F;M;_qk=T1iu
z-?)IZ<oGXRW*`W{AU>&IrX;p$B8WZ}LDCMM3|q0`iTmu{DcjCm)~VL0cHKRNC-+YJ
zRGJl{(gKGLz518^vD-iCSia90*dM29Xr-v^|GhFOA2CC!6!Pz}^8V{<s$O4y)~wZ8
zQy14_DEeG(QTTBE6m?=Uu)47kt3AH!gZ-iB7trVn<@{>RCc`Cjwoe6DKEG^rb5xOX
z01N+Ky#E`E_o7J4<0e-XBX1N!PM7xiqZ1+{kh@~lFms@zC4WV8f&E3_5~u5fSE$CR
zQOo;Umf<bpnvl+V1LwS*zdZP2BI8dJZ(HfQa=}3_2NScj82JP575k5I$3o)GX5EzU
zbIs{YVD<73?6A*TIV<Kh#(r)2s5Y*@D5N7pTJbPCQ5L|wrSbT1%fjQ7q1BhRexPPV
z{AZY&L3OD;_~K-secPK~nyeTuO`iQ}8l0g!w|j%LOT}H>dY8ds&d~6|<kG&S%tOv?
z-LD39-d(%Dn>muTxiht{z^lQ=Bez!Xk#j`&+3@V>uQtdUOJ~)d_Sn}Kneee1$ZfKU
zISn^B9}Nn!`7W99nI&Z<|Gm?<E9G=nU<}`KPD+OJj-r2EIdW)6_`j+56So{|y3UsR
z*=lFKKzqYvnmK14?7K%f9#K~Hu6JQqmJp@z5Af5OoV`rRRW7^BSh{I#h&(-K<*|g;
zCke@i={+WsUK@=%j`i9+`(U-9HFvbfZ=@=d=VB*$bWb$uaGN3jR*}#idA75rBht*M
zucH3M&58MyryRItrAPK8w)0x(Nm1oxMMqp4i(HBwJ`aR$T4%~%=a{V3eY&>y(O}|0
zeuVLI_LZICdW$Yr|61D9BJRxF$U3jJbmD<i?l(r>tM6L9HOcMX#vj>Q?kqbxWGv^U
z``((!o-9rX2&KB6%k%LMIiLBl#HFEPJo|Ko(~h4aS)G@k=Oub{d%PO=*utAEm%pSO
z#<%-56^h0!uX>sL89sl<YFM{s+5VA>UF$FGb{r3|TPUaq7_V6Lrr~)|rt#~dtX&^;
zPBqxnTJ0_vvix{Iefa71`jAbgc3n07<$`f8)#IQH8*4i8ovFfl;*TFZjq^K8Xfa)`
z?nj%O{LkL>91i9E%nA1QuUS3w%(-i<*3j70;u{Ly`k6Q4!P4IOqrGnJ4<F)x=&yQ;
zC|Vs9XO4P3a}$i|U*hJysgF{OwdJG)_N?CYF|jJ7Y@h4>2hCNPqn{p!|8V@V=v~S{
z&)8S%#ysC9ZGU@f{CeU0>}^xln+*1J+-?m1&ORynNm|r<!^#*z<ITkp!pd&9hwgrs
zEW1hF(aDM<sp-jh5{e+4XR=<;bI9*BGy*|z4uxZ+0%FEhKXbpPdEIr)b*3ZH1kK{H
zEK~q#`^}qQ54{aCrlE3zHlP?VRWMCA)o(2M>Bcr18b;8%6kWCoAno`r{lcYjSp^MU
zOweo!d$tN7EohI}w=;^)qoFGa8c=|d3Lq_P__2CdTS+|)4IyZ4ingu_Ani}hD0&%s
zt&oN;BWN0hrmq4>C#?GVQH?>tAPrqY&{PVQp#n&6)MVe$qD0rw&`^Tbr|8dC0i?IK
z?r8@D)Nkp?8iHo>m=-F4bk2{qXJj^|S1{xYf)1MzrlkT%=P4TW4!3OmjfSouXlshK
zkqRKaSL~;;s{1oeLxTxgi=w5g0!Wt~sB@epyx2@b!wEW+5~{BPNY}Ium@58!TSr4z
z5j34bpREE&x5$|C7lXbHG;}#ZV-!qR1(5zZ;O+%M*VVH$G>V{gC_1_-fb_2`bxPAt
zwH>9Qkp%5a@ikHbq+b;;4Ld2ReM(1G6EuUzuuuV{Q{vaA@+qX}Q0m{jhT)-{P(HjK
zblyzr-yDbG#=uwzU5nwiz*Y$BiQ#U*O$cp+;Zh(K!i=F)2|1-rVM-Wo2`t5=oC|Of
zlX7Ap7L#&Iz!LGa++WTn<dn81<;;MYgp_jvP7+d11Vj>2ZV^}{fpTyX2sz~#kaFh0
z+?$k}59WK5a^Aq(o0JO%!QM~~o@!A4abk!l%5fp1@BmLlMzII>A~K2q2t-g1!~KE3
z2u2Z)*sRSuvAk9T!+m+a3*a#vJR*v6>c}W2z(h($;RC*ujN%DArDT)<5Fmweffbhv
zyLS9WC!)}KbTer0V7KB8`uzMN6vNp9wgeUs!}SIF5;(olI&D+8&oPQ;Vz?GZiwg$}
z!?iiuT)00loW`MXVb&OqaWF2N5)9Yj=y2gsz@0(dOt~<~o6P}pe8`(Q00$rPW*)%9
zhrF2#$b8_<V3vfOa>0;t7QjNS>CV7et?53%N3H3L!Q$zr!<|IPDHjqcX9SGY2F3>1
zs11xOa8(-^36M+=4BTOahm;Ep%H;)WW{nk2xocuLgU#T=iN|mzo5_VKU^t7-;=&1r
zwIGy~yACSBZV>q@cN{#%a4z7gxzP^TskzY|xU0F*5BN>Hu~=SrYJ9jgmM{dBL*>F=
zVYn_wmuvsmxWeT^q@rA$Fy<ldN<Dt^mQRGo!tBBV;nktnhzyl`6#h4c&joYU8Z-~g
zQ)`eA2-O-C2m+@YRQI_t-*&Hykca|!U;#7*^ct}ZmFt*{LN}*7z!YwMj#;LtsO}@a
z%i(brK(pOzDJq|w+Om^4w&qzcfR*k!bIcs%R`nCd1_A>y+<nl;#C?>Tos47(Ow}sx
z2prWa?ghNmDjo!arYjD8OcYc3CV+C#5=0c`8vq^?i35(B53PZ<nhya0YCiM@zSBNL
zQsR?hlp?7qv8kTQ2i0H72Ly@x|1TxRlaH~}kDB<FW`@75=*H(6aq%&+NNjRk67fn#
z0^=8H&X`G=rKL^7=(BZoWDLndD+X1@G+^l&a@+-;LNAfG&pfuhhq<eb$v>Uy8*_Pc
lP0cJUt!?cb9G&L70Jk{{jQGyJY<<6gwWxHd*k2<E`5*kCmvaCB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_multi_ABD_xdl_fp16.cpp.95E96326F13F98F9.idx b/.cache/clangd/index/contraction_multi_ABD_xdl_fp16.cpp.95E96326F13F98F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f3d542513cada78f7c26179cd00e24ba7f4cf784
GIT binary patch
literal 9356
zcmZvi2Ut``*T*^Iby?WEupphK?kb1~h%^yJRF<Y9#S#-^*K34mlmKdq3N{2pM1&-Y
zK>-!k5=BC>0HWw?OYBnZN$ib4Ecs^ld1q(*p4aDbyubb5J2Q96IcLr_HbSLJSL*4F
znjMjvI45Zar>Cc9g8wc_T0B1)gJ7hmm-<uF#KeUyezFHe)(3vxVO~~vUDT|9{D|tu
zlS1r0MM#{>dar3dxi6!>f1jP3(0AO5#51k1rm0r(Z2R)3-6LK5=G}6S%P-3t8uQux
zn_YLUU6Q;%G?e-`{&u~;?fHMw6isVgUM76yGvi3=_uuWZ`b^$`drd({Ygp6y*@6Fj
zFwpOKR+{kQN%E;R8+s-<m}h>w-SftsN6)`Ye*5GH>He5kxwl`Xwpu-inBl8j{jmSv
zi?`(^_sagN@#=-BB5`@tqn=p!bmib3xpVqX_#s>ozolWowd3EF<VKxt+xKeG+wGV1
z_nCdTn;`AiDIViCcyG69-p1l3*P5Q$w0fBRelO;RXl>K-=8DTf9=R&fUm<e>{PI0w
za{G<{N&i&T0-MUzzn|XKw^@00L$=T1^t>M(Ll^ua8Z~@MLQHzzKUsY7j{dXjV%0yI
zMe488_bGEUO9|=>?665(t-SR6Cy9;qb_pi4a%>i@Hr{7iA9&n<Msh(`{%P+wAyrPZ
zJlD>1ToM=aWyQ_;pDw-C{lq=eX2%ahE5gd&!G+&_Q*z5<cTCvPJoWRY#fMiMer><L
ztdQICdvoQz$U6;+p`}&{H?~~*OFDN=-slmpH{HtLByw9aYttpwq?bNVPPJdojH;Xz
z@VxWd9}%VENN2vvKmDxrxxpni|4F+vGxY8IQ-4Vw9r%#6C?u`qtMY48GT&xaZOxcc
zxaFYjDF4>Nf~mchFMJi$tES=?Uak(hIAh`NAe8mw+}K-Q$L75{lc`q|e`8SA;$Gi`
zUUzz$A0GSTq<ef{<L@U<=Z@rTsod(<J7LPCmIcoP#N!@hZw*a~u^v<UA^A<@f}3C&
znwjo#$KxA8W|EOrH1I*(yN#h=L`Qw*kYYX6ZNNm~*4@%5kBK+7ilZV1HeR*<IscNk
z*&hzuW}HvBf9Z$Flr8b?O;3V<4PTT}5cKD?*B6$h7R<cAZ)Vd6$E2vwQ*!I`70%(J
zZ`Vx7S~s(>;8b9wpM30x%9i6x3hH*xX^B5r@z{7-v3ta-^SJ>7eBK}Z`&df8dr5@h
zCx#(w=Jm1f{<h-rfevYz;mi%K`LkU|mra~gA13Se8e1gZx^;F^f$v`y8(l7kHTX`e
z=c_h#Jbb-)q+L?Q)!^oR^UiedJ8^B&q7z;!oAOI%Hh;-2=-+wdCtHVr<Uxh@V_r0D
zyM62(H>G2XWAoRUllHdFSh2J{ewp)|tzFC8zFXLgqSwD|xasEEY<GK;T$LYgtgwFY
zO`-DNu-ntDYmS=sIoL35e|NOJsr+K#+$dq+PNiqLQeD33UUR^!@TK2+U3pL>o8Wx(
zN%xS(gTs2f)qk2Do4-G9=Bl|xkB=uPh9$+doA&&%t&hp$2}aM4t(!Mw`I-pF(hKj_
zZSNDZXH~rSW7C)wlLvh>cU!=?PgA6~YTDM%GQYMq|8|k(`uo#o4vMz#^RHL-OBbVA
z&wsnKE7i2**tHC$B`S)axKw=j`)tc?(l_ScKkxcakNA6oTioX_{^*g~+aL8W%pG3W
zWIj<+<noWyW6DHv`-9O-=7%3jQ@?%V|I#uyWl+|VrHM1sdg<v&tGf>>O6`7^>dnHY
z$i>b@sV%U5(*BXnxV`>P#o#=t-sf13u!ykN7DP7h+1@FBeTEmIX;==@4>HykM3xsh
z-5Qu)x6cB_#Oev|oV!e05P87YUq5^BI?7gpCg7*TL}8}dLMI>CvuN{;)89$ZWGpL1
zN^@;NWY52s)*kB$FO;BYEQg9hW!i$sL3=N5Ut6<zj|9bG*+JxBtSyKfn)UL<7yI5#
zmY~nDY$dWX&=y1vGhCT-Kj!2{37UjuCy|qdwjlC2U(fvPUro-5Pz;vc^xgVs3nEXx
z_P*WxN%cY*S~guz2<C!$Z9!z=v*p)!T4vpopz-)=N0Fn6wjlDNUgqb-qR)3qP&}4Z
zB9%;A5P99j{*~dbTk3herC3hkQpDPV$m<m~ZwuUqZ;_y}Shf?{$+QKLGuM}&|HAX`
zMXBBlEW6sdDzycXbM^CX?s%5`n^bQmmiyWDQ)&yH99y=(>yN5g5;PUd)*@>IZ9(Mx
zAId!SgN`1Qpeb065=B{P3nCXpR8+NPPwm3Pf|rB0i?^q?AaY^3Y4YCd@1u>-Tr39|
z2FSDpkq>Zh63_WIKa-%@xN=*Ot%<fE@-N}0v7Pru?UkTUu<R^yHr5tIK70Jdot(Ry
z#oD&DKMI~Mo}NU$GIZnHxlgW`8KHT&?IDIkjI{-k+au1`4!O{HM1nrWl`BLFV{Jj?
zhhyUtLMDyhEJ5S2Y%j8xX$vBEKWg9P@nE{M1kJ&+g~&o*TM$`q@tnC!oNV><dJU57
zDLOYZ)07wXfZHDM)J?Up7u@!;Tn?_~Fp%mM;99}58hq3+gz6_C@B{=?u7|*SmK(so
z0fMOB2>y*MpM=np5J~lH<eH6`_H&SX4jM%DT;!FDd@1K4pFG6$=MeHa#BwDHs6@l)
z{Z%Na3WZRvMkA_GC}Gis<BuM@9Wdm0;Slsa1b#%|g~QPIFw2#otOTYbM?iUmvb1vd
zao669j`rdOPtKDkGr|kroHs+mx#2n@29q=V$WZdaSZ=I@Yye*90?#hCA3Kn02O3NI
zNz5uT{K%a0!enkTPb!gB+b0g*5`Iyvc}tT+pd?Rg-qPeS%atIj1O?TPfb0m%)nHQ1
z-d_VIH7wVHtd_mM4rFzdP0y5^KXWzuN8p94;BXZNkcP~Y{M9~}&U08OLas%0f{0(h
z@27DJbX-HtL+yy+wmLs(*_quTSMbw~$h;9bkZOtBz)x#@1~Imra0zTKK|gYvgVD+%
z>&548<LCc`K7WEE5u9ok3y)U>tpN@TMt_0{5s1U-W%!Y!;e|=uBpDfY{{J;I;J<DN
z_eSG`=)066RVkW4>h^pPq|BY@^dH=tDF3MGq&D)hxJiwRqEAoYCK!>Mc;N!rT%Z%;
zGwh_{jipn}@XTi;J{y^mMtC6y$#Red<y>Tzi)<+8A*(!OPx%ny53$czB3UK-d=)aQ
zVxO-@R@L<RKv~1K%EFWncz^r^;-A2XR=f?w+gRQX;_Z|J?O&W-Ae!;QfWxNDXb;H9
zgB&(xMtfN<2cvS9E5N9NWi=S7Sx!TyX~>M$mxkQZkUPu4Y1DLZLOL={r}vYvz|ioc
zaDW%a8;!T5^RWvgyFk|nFI0kGC2KE_fZq|8Yr(IU<vQ@IW4QsG8|c2_h4<j~p0&?A
zk!mN3CNB|mIsWp?3z;z@PRK{@`P9&I!d5hJEAk*WazX(bSio{E3amv#sa}Ty>sUUD
zJdPrrV+nq8_36Id+mo>&Jq4zxfLTbV!SpoC&0x|D)}&%yI0GhUD37jrI_6j2Ck@Yc
zF<2IZZZC7fPO#j`@-DF01$Lw&PACD35|&HBq7?d5y$meM=+JRO26D<kK2*;{wwcJ8
zvKm>aDU&!)WcX2h#|w7Wc5)Ic@q&Z3gPbf4UU0T{mXmmb7ZlbCISBxGL20d&6O*8-
zTm3ZQ@!TT3J>G)!E%YJ1(6HlMYN~kQ9Z26%orIUziD+U=Y87I+STni`GmvoxHH*Bk
z3#oRYiKLOp!s8n+q&=46)}zJIQu^3<@py*f#Bn+r>z}*OVqj(tHd3|$wxdW$HTKk{
zn5zSyH}~d*6JT@#Y)KQGP!C4+EH{8r16>xJ&<IA2ET06Ulaz_+#@0_0B-8g~?K&@n
znTI(M+Z;d6)a8#&LA`Mhv=R(ff-&g}FRTKCRV=52K|0H;!C*DZYr$YG%j>{k9m^SD
zkiqf>FxbFy78qo)oDBxqEVqDF3oy2}6|7oWZUd_}mfOLqo#hU&>R`DOtU6g%BUd%u
z+q|Ggqt!?kgiSizF|+Z<lJ|Hn%aL<AH4^0XfspppL42{smMHU4UoxIk@1{?#oVzd=
zFP~Q+dj)bL@WLCAzX4au-5~F#JZ;aOBjvF}=HUnP!6Bc<3B0fs9JaDt01gF|$+-uw
z2F;;|u7(&c#+WYO5~M0YF{JhRi<cxvcbARm&EvJnod78jc)T{bdX^i2Yhbw%xJH&w
z0(X*f!seMSqi;v1VB6;_^Ocjs38s<0<0ys5LgZvbc_CzA$Phy0oQ5~C=8#5TXCgO|
zaqSnu<|1|Nys#UocBAp+Ext+}eo=n*_xFYzUhZ})X$;Td<!-l%<#e!1XL&W)t!8;G
z*sTS|5w62lmE{bu%V2o}*ll1r3+%F3&IY?|%E?iq&kdODHy$syh~S7A62Yc6-oCet
z`|Z2|FI)rZHDE%O8z8;G@=cK5WVr>TEiAW!w3X#HkhZbh4$^j(J3!jOawkYTDStQn
z`p_#6(-&)(cl;~(oQ$|8HXqT&9%gzVqz9sizDhkNal*FWPU8i44F+9<5k%mH>oDj#
z%QwL52J7W-g4a!!Z-MzOI`wOxhsuXs%(THnWhJw+Af5>yHvGJU=z{Z(@Q$W5m>+D)
z(4fKolzg3hT_^?j3AQFn6XzRTgFPsPIEA<nBCi96|1)o3VE9qK0Q(lBXe+uoi$PKh
z=A;hI$;AH<um?5puV(oKC{EBtstJ7hyr<rt7c!A?CSt;hVx%gjyJJISuk<Gy{|&I^
zg?{G!<Rm-63ohm^a&i*m1y^%dIZ142+=kgo>asL$!)z7H>0p-5@@g<!&GK3>Tg&n~
z90;<U0cII2Zve9mEN6jP7R%XSmd$btC|bac4q+=OT3K!bMH|cQplD~g0~8%BcY>mm
z<(I&}q@jk!x;wq1)>~uU`8TZo1_r(XT|mVP-N1LVdN&O0W_2|htVV&f|7sMircnu*
z2W%BJ*(GWf$8qDB^KlP&^?*OE<Y%P%8R^c)nY;4a7Jl<*3*xZTQ>c;d%%|aCH5y4C
zBr5`kcba4sZEXxUhH32;jC#fH_%7(z#jeU8aOnYERING9s)}fDa{rt4!mp%BdC^Af
zt?lI`_rnX0){b)G;&{Qy+DT5rye)rEpSsFeeihr!r!e3teQ&a_@Yzz6Eu%GG2J_3H
zbJb*%V~?)MnvtlA6KYYvTI50ogCx+f$#Zvi_a?_^?9uxBdXhQ}wMjKnMUhSM;nS@X
zdgGq$M}79wU~|V~xABdZe>Y&^9GIK~E7Hi%2~k{R%$hs+Y_bQu_JB7LG$%W+y)2i5
zM>%*=y#hQcSXM)znkJbvCpEit+JoX}=Kb&erT7Qi;ZvYIMUMo$a2k}SS#AdBX3!0T
z=4|48hUG@&*GLaRyVgxEy#MtvId=Tj(6^e+($qlT8kTE8Sxd7;yif<qI?5HPwY|qD
z7OU_WSr}PZkTl=^{tqU6`Xa`?7bm14`!uA>OK8|P4e3_BhR3B*Ye&Lyyv{V~O}ZLX
z<|;=LZSX=FQkBv8NHcL^`%x?z<^vtWQwpu31#P@E;snGIK@%^HsAstWhBbgL62Vy*
z7}m)0Nr*a0{pP{XZg>}G8k>qV*-3*vG?3!)#2f5oxf~42S*`$s3YOJipk_G@Nz;%n
zOG8dxI84-}ap^F`8O1S2^hR)Qr02q_{HoIdt=`k|TpUDZ2azRt?O#<}QumE`cmvNy
zJ}C34A;f2CP;O<p0F(ukf1TTTZnWXq+juv5N<A%UH@v0Z3=NkK*U_&{Nq(RF<oZS%
za|FI(LxC=kcYzb>6fg9ET@UC^t+iRk+gHePj%gD<9@!{>-k5{Da}bl6%|(8>h>1z^
zP+%S!P47R10uQlViH29AaH?0KkyU66<!Yp=rc1CcyqQ}%=i4Maj8DPtDb1-hj($3u
ze^b9&%;A(&U^d-N98O6M&q2D}GAHDsk-2CTX@L{+kSY&_Q9guJhghyek(EdnesMw-
z8dpVgA)HW+VylsE88_d*?&Wdj(K+11E8uX2I=-{Esdo%gje6rAwt!a)jngzdvITTI
zgcDl9tCiJTVPq?-w}Dq1ojgux2d{RPJHV@h<xcSGq)a|F0BAok5L0_Wv3%XC;c>>=
z{S&Z<8a~bb3E0bWIrx<W<6$enuYzSYj8ub)j3|!{XWw+BOPP>w85%b*)O`fw2K;=7
zQa7Oa$U*vc&Aztx%>j;@L2W`dO-PqYzGZRz!?Iw<Wm+Tbc8c|pr@`$s%gx~0Opgm1
zBkX#H<wg|LNEdF)>igG^-R=DgZm|f&MZlA{A(<9^!;e0(@IpKnughcbIBO&+VH0$v
zASq?J3?yYNp9JTV)LLlb(6R49mrCipyJEy9lLIZdHxY3WQ_11xK}*iiusJ6T@y`#z
z^dWV~#OdR(L-VzV+J<PGXeDv#JWhQ{cCx|pVemi92FEqvU&C@eDC%ig@@VtEu5AM@
z7UA<kF>u9TNQOvLF}IWDUBK;PxdgZpmP>&vWw{KvGL|!tI0G@s-Ap9TWLb^GYRbgX
zi48ycj>QYJx!Jwx+t1<VFf^B&%g_RD0YeE~0z-*hB11`Bl8*3~vIP9u693<s#j}<i
z$N%;cPUNRr##~(Z+BiAo>!kT#EJ#?m=&QwvUoO%6icE^0-eP(=pI$DYmtWG$h4eCo
zUVcF@6Y1p=digcIOrV#G=w&j!Orn>w5@#pO!x%YdOfRmtNMCFq;iWQTqi~baBm0>c
zhMQXUF|)D`7&0_y*zn*H9_9lA6@IP`|La!^Te+R1le4m~OMf?aPp?6Pef{lyYzBId
RGPh7gq@vJ?Az=`u_kX#SS_l9D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale.hpp.9A93BE54CAD13142.idx b/.cache/clangd/index/contraction_scale.hpp.9A93BE54CAD13142.idx
new file mode 100755
index 0000000000000000000000000000000000000000..261b1f924460c135aece7b8736b44d3ab653df02
GIT binary patch
literal 8836
zcmYjV33yD``@Y{fGmVL>W0|Rv#Ei7o5LDvEeC=dVgQ~5KDoQNXqDX&Yt^6yCkf_-A
zXsIY-7sL{pk_sVOTWm!v4NXI+gbGssn`fLkU!Es2ulwF}?>XoF-tWx&Q8%|cCQwm2
zeco-#*bx(k8WcsTrT<1x7(Z$UQ<U7giZW$Y!SX&)=UUf1R@Ud4lGH6|{ix!EzI}%U
zb*eVjzH{=)=&MIZ`}_S-ys>6sKvwd_J^iZ(KKW$buj5-S%ZmK3?!aY@w@1G4ca43g
z+OoxO{@A!=YTt|p*3M@)=FJJN_v78QA-;8T-k;jvIx{c+x5mYzwmW|der<U8w>Qpi
zIX!X2dk2eyE<AJeTYKQt<Zol%y0m|gdaiiFm1##m82fd%jsHztXn=tmqK3bx6`VHz
z5%JrjF3p=1*BTbktncpHbGw|{(W9uv%b<cO=hr(9eBV&{_)?FChTfMdZ|qN=Yi{a3
z6+ce-{q@~hPkJph^>>XNnV#6O<41{u23Ok^>N~?{{<h}M;V&NT?$>MEi$BXoJ*ZUu
zVh$}VJ#_8ELF+?HKObGZ_x@Nmpm)#i38l+3{Rj4{Q#8}PHu~_05gRw8_FOa|zSW~|
z&l;YdPnr@ps7=4TE~5{2YZ7*{_tu7ie*es$)U9T(J}cICp8f6T#=x$33u4Bd(#j4$
zJ?t7?JH$4iuyb<EnfPPDw~yYbwSQ{Aadl6x?r`+FU)US}`*Ov<ua$l{CgJHhpTYey
zvW`!CtIP1x;Iiw-+bmDZ^Y2kKGSKH@*v4%g_Nbd8C!fkqd*1e5gRE9xx42b$ptQRG
z!dI6HmwsUgUX(Lv%jCp|+XKwnv{?fl^Z(nqGhjsIu*CME5swR>%$j$}jo*(ydGp$W
zJ^3wN&j*cM^24t9j*2Td%(y?d-i0xyb^FqMZ`Q6GbSmWQ>K#VL3@^!AvuT}iNTX((
z=CF(%$5KB#p~OwH8Lpo2Js;{b@8HaX>5WeOV4ohn^_QSNE0$NRyb?Dn^y<C|d7tg@
zODy=iVVrh<;nnQ}woiNRyYk{!_wtfgb=`Vx<aZsr@`%^RnTq`y`u6BCeLib61D1S}
zI;PdRw$qepDIb^qT;t_?uF8HLGAu2+f6HS#r+gnhb?mSpUBsO4=?d>Ix=@?dZlh=$
zc+Cym|J5I*nyzpC=d~^tN1oENpQ_Yr?fD;V4>RszgA|jADUMd4!Ryxxe}Av&{cft#
zP*8>%hw6mt1`pwlm5D<>Ua<csRcR>*C&73U`Y5K?eLR#mJDKv>H}fh~rKzAi0^=jl
zN2&3ehY~n7>?ilw8g*6W4MBOzj8EAR#Z(PEl-8Z*r-jF+<*7;&K`8=b5gn!7*2GE1
z*)8g*N@GEB7#;eXs7B8iVG$ih4#-*8Tvhx9Aqk90Fjz79au3D*Y*1R^F1J2mo1k<t
zb|Dk~;9XPvwERO)ROKx}sAH@{9@5M2?$@<?zqLzM-V}t5%(#)RYOlwkLze#<AFnDF
zLAlP1*XeY7*ErcSy2*$zRe4)b>_$6z%%Ipkc}x5I2B=D~AY20DCA#oK&8-_f$#~vN
zRRRTNoN*kP%jXxDl{DVldZnra2*MIDE}`o&sebOCiwtx0H4PG!onYKaACp{u^VSCc
zln7O^3rd7Bf;?ufskGDnD>L<FZzc#Y!T6GH-QtYp-D3myhE``ab+x8)-&9`v)gPhO
z)tgiARR1A6LseRcQGN#F&-6{wtNqf)9C_$#RjDs1N0{*lUHV<wAxpOWl(axqtb$U)
zj3spG_suw&d*amdm#R`nP|h&p8A>^H*f}&PykM28G!T@vU|dTV?r@{J*2@FC=o{Ev
zQ06e>oT_&bsy{cl)u*#_hR3K%BSC3tY)SU`%9$|Ne7_?ls^TXI(~Q%IT8rld?`~Du
z%KaLv)!MAh<vw$HqgQ{#vQ_-=er4dIH}<JYT`@)~Gp5otQO1w>BD$KN^bGKQ<rLoW
zUpWOcy>f_@BhS}txrugFd&w;?`556gzR%WNPv{={ahR(4p`Y7ZESf+16Vw!&y1fe5
znqf17LNV0sT^X$%wj-z`c658IT<e6L2<nBs+}`TddSh>bKEu!4UYTlxa1cSGaFp9?
zbJ}PeP0%Eq<n~sPHW?=qlmM6jZtoUpGXQ5$JRfjAxV@I6EdX3V@%Mn=gWFqz+7Ey~
zP`n9n6S%!f(>4Qcrg%T#esFv1Q#$~7fZ|-hTyT4<Mmq+0jN)^E=fLgtZ|ywbd5UiU
z-T=4PX0)4tHz~dkcpuzeTh>YeODTQ^_zc`$^VgmOKBssV!&%Jj^#(18VG_lQ87^jS
zZy9Mz7%rhWjbR#dd*!XIW4Mmu9SnCcx7UreoeXzUe2C#8=JxhL?PrERQ+$Hq3Fh`%
ziguFWNs2Esyv*EQv(T<Eyh8CEhIjPm39HxcGQ3N18N)K>_WF@l&aj-~SdOvW?d^<O
z9LG3`g<|!JtZK-z9uITNVOfv0wkOx0xb5$pnyU5@e&m$(5s8sbSrKD!j8lxLYEyBl
zQ>uP7!_~|w*Le-YH59L9xR&BA47V_+T=%UEw^F>F;dYAmGu+Rdavu&bJV5b5h6gFm
zWthvHa$k-yJVtRI!#s(FzV&tToJ-_N-QboRe28!^Ayl2{Ih{yiF}D=UB^5iO6Fp}X
zNnB#EEHOlh5gQvb7yEw|5W!Tfwoh%V=Y*vws<w&yY~o%lDo7vCbD)vq-Q||M^1;IR
zbYj)9M$-0i%RV^~EzMc&i#`?e^v5peO_%eQ^1<tP#5&$z^jcv}`gopGkR0&@x18W(
z#YBW^bYj(6NYV~*%OUxG!t8Zg)rm;bGPor}zTXFXzbT6}epcB))hu?4L)@im4ebpb
zvMB`HgB`LBpdB4z{;Jm1-qj&qO4a(>`#PkQhuMcYq?DuVQ4T5P7<-IE8g??+lffZY
zLDiOmeJMDkA*X>o4II*dGr*n!4r#dAV9y4JG}wHw=YvBCPSq}h{W3VDf!+c89dL;4
zR<$y)mw`j*PSs+WJ(fA7A*L{U3Uf#UT*d6Gm_rs~I<u!Uhb**AX3t~}Sx7m|p2Hln
zP|h;@S>}+2Rmkjx`u$#my2tGIm_r&;1+!N$hmgCfpDW&<%Lhv}rE+U3Zza^!Lt7a2
zM$0+#4f;azrb~HSNnFO8E~9uYk66ni#f_?##v{@s78BBC<T*D>AyslqrPP~vbDiip
zJxk&OgJpp%L?J4jSapV$wB6jYTM9FxXYD3`EqK=$pbOJvF%w~$8k)rzsupYtHWOi*
z&`gAB>S`v!H1#zTVVZ`Si7-u3W+F^ejF}44lniDnOw&>@Q(>CYz)Xc{$^bJJrYRfD
zRG6lGFjHZgE`ym0({u;SRG6kRFjHZgVwsr=)0D!@RG6k!%uI!8N@r%VkE)i*Oqt9~
zg=xxRW-3h6S!SlfG!^Ri6JeU}F*6mWse+lMFw=RBbY4fem6(^l+nz5(a=&kJ%PqOz
zkyBPh4LNl1wyxM*-XWKF6(guxk|8+B5K8gKNrnLuCmVv3<#n--`V*_ZU&-g?a7&KV
ztJpxD=J}Q-iTT`;Fa368?h?PY)BBeAt6BgCxTFx9V{?}jqK2AF3h{k>-zA0E2m83B
z5C`L6mlWbx_?1fvaSBdxi3zA$B48r8q!bqdE(Dj9;%dOv;1V-YwJm^Kz$K-a1(*dc
zDaGS}$H66~SO8c6E-A%ZfVaRUwoKI?06qYhl;R7(7vPdoOlFwOTvCcl87^fmDaD@{
z{={5TiWv+um`h4An_)I{Nh#(t%x5ks#XlJS!CX>`e=+<^zuzmx#|$4cmy}{W$9V1%
zQWW#nci8hKS?=&vZn-L@Hwjy=_<YK!Fuh0b0G}P;Jq)PYdG2#w9u5+Q)OT-g({_@6
zewzt66T;=zECO5v;c{!%0Iq>>xiwn>w?eqwngWIeEL=|Q0>cXwUt)NP;#&-Fv2Z!{
z+YE10T*R=5;s*>LuyDCF4;enB_z}ZL6u)5jf`yBvQ?*Kll@z~Z_)_8-Y16;yvGM4!
zI;ysrTQ~CtvKTk>kj?x(iMMd;7J7XL57@yQ%j;P@FpCFCoXZ1qDK6mF0(yS|4=vzM
z$-ls@7wGjXJm3l)zlaAG(ecW;wVd+HdHZtSQI7YRTOZTw6+EDVj+bPxCK<?llMId|
zLnnEEvcZ~6ucsISQVibt&U)C{zy0rX>gz6-#XDs2F7n259$d~tDE_FNe=6}~9{gBd
z7Y3q>yDHw1($40VZ0T~}`TX)@(vk78y0jPbYKwUdc@9{^L)P&2qVJ0$*2nXtL2|?r
zZYhx?iW`{kt3*ho5~`NY-%jT(#2csHZ5Pl}Db3R#eVX~4W~8cS8GL3Lyjb{={%B8x
zBp-d9TdvDG6~>?wJwcKr#u+Se25;^mWUmuFagro1GFTSLvxulao#+XbBypa>GEWjk
z;p#+Bv?PhYam#PATL|;hiJpK-67O@%eYuWDZl@f*9W-x-zC0_L&q}&HW!$GsnvQ5H
z`iNC|lpHUUTQcS12rJTQRjHJu9p{$g^8Jch8_Rvnb-$}RR#`1J+07bS8`@+y3$_N^
z#Di50t!R_otgE%FO?I=s*1k3&MpYYT9cGi=EXo>Xlie)F8e<c0scOk!O$M7Z%%xym
z3N~q&X<$tQn^<>M%K&Qz*o3F6S~gg-!6vk$YWZNz2b=6=m%(}&Y_gZ#0qY&G$zE0l
z)-tfkUKY!&vCJmQK-E&1HHF#4cB<McW?jW>vX`YZYdW*ZUY5zMnan18Sq`)2Fq@F9
zs-0!lv&<$0r)q`FTBzUe?Pd3v^&Ybc<*Hf*vsN&hG~{$%FP*<7ic1uoo(+0(MzP=J
zJO0TS7<Z<bep;!A^+?Fq7#ovVuPwGEL7od;B)aQ?JxEwL00)qG?o0fV1acE_0*T<_
z0OM%nHWzR%jo4BFQ)#5O5pW}o(DnlErIFdMfWOj+><r);8i`#8yiOyq62KA~c|8Su
zN+YgBhKV%NTF7uAjj(=X_#=(1wlUmBBdRQhSu~P5&hR*mpe{1JNF%2|8U9Hlrhgdz
zLnEdC82-n+k&;lBUca71SnQ~(J?EC^(m0-FE;}5ICwD>BcfhKMSy*~y%}+9}Mg4jw
zMAhnIU8gMVH}Oq^>SKL^8ejv0-p03`(gYe|BZ8V>6M}*;h@dvu#wqRM9ejtN5DXzG
z48sU=qSGmjB^<*E>VO>x>WrNU>W<w>diEiHNKjAgNzkYGDM{A)V}F7M;y{8%;0Ths
zeSu#PG!jP=G#<y3L@pYm37Uu#37U@6ozn7R0b_ySc))my=K#(Dr?kWrz!Zw-0nVd%
z1>g#BO3PdcxRT;kfU78854awj(o#17ZlE|FFrDIufDgebEwd7^5}eX9XEU75oYFFv
zF<i!+(lXaGTu-yeT?}{8tnn9yztAl46vI<8D^#_s46ib$w9I0L#rpkT%Y4G{33E!z
zOyHQnox(EzD>AL=HD&OA-3j*q?g6Ll9hrcc6z>DvNAW?xgA^YDJOWPHYjOZ{C_V~!
zl;S+VJc@q@{2iRKH=PDNP4QX4vlL$dyg>0)z^mYtz3dv`HHr%X3n{)0c$?y4z+$?Q
ze*^wa@jbwM6h8rcLO1$fz<(*O0IZ-mfnfp_!VHEpD4xl1CdKm^&Zk0Iz;FS@ix@7F
zc>IVlL#sS4di?O{$f^%vRsVQ?v}4-Dj)rN)kv{Fi-U(?N+D-`(mlWk4`7u;}Y$rdq
Lmmj?WnWFqZ4I%Ni

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_bf16_compute_fp32.cpp.2E396E91009DF0FA.idx b/.cache/clangd/index/contraction_scale_xdl_bf16_compute_fp32.cpp.2E396E91009DF0FA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..316fef4806799cfefdc62663b93d1ae3aee8e811
GIT binary patch
literal 3306
zcmZuy3pkW%8~zN&9KLawW@MNNkuVNnn%7|r1~bMGiqNX<bBdw^i4Gh3vpOLpiu^Pk
z{<eflwH38GZkuYW4W0C(rdI8*@{`Knelyqq`ev_dU6*G*?>+B*Ki~7*&ttA!D(y!R
zBrsW?mzlB}(?bx%1pm!mot1uC7eU6U2$EOV6cfMVv8VCFuMXF?RUSLD@RCt!SyTBZ
z|Jvzgr0^1_PSGjRhBwW9jWaA;C+?=|n%Svp|2<r*2t%02Jl4GyZq197-1v-(Pg?cJ
z6RHf)x4z$5q(&&$s7N2bNT+qVqS6n=pXm)B@^mibxYc(AE%r7OPkgBr`sJf1$H#i@
zhq|sm73#!~=7{dR>|5um>>ZOotToP3yld-gW)GYa(C*)Bt@CP~!#^2N>((`Rn_8Wv
zbR6iBB%YQWjXJd5D!HE4m2%@9Pt}({cUw`)n%wuZ)`vW=lFPp)H$|?0)Ob5sr@}e%
zoc)4j842BKxgD%^oT8>P(HFWq<F58wHx?NbpJ-?owphv2jFO8VQI*ST@;{Bd@CZu$
z95|H{UE}dqe?_$Pe$|p<#@3pDJf_@H+-O}KAY7hx%qi=oe4kDAE%yjjf>KaefmF@D
z`BCZfgeh}n-J_~9pFfMsjf`5&Dq&W;dp7UoC%Jy+S15P|>B`)jdIoKMl~P(sL(Rc;
zb(ivgHIMV)$GLc>du_et66hog456y*-$hb(d0k1M+v)Qz+?}`O*~f)NjN=aOp>kWj
zB`rr1&HbI)8aF?TyZZjisa!*v-Jnj(tGma4uDr2zmzDWw!|+c_w|N)@zp!4~y8rE(
zrksPV(fO}Ln}D$_r|e#P?Ar4&4;o@0We75AesepwcQim1SJ*-cE;QYgq!8I}vf1Cd
zZ_n+a9hN7?uXQSau+0f$EZ^C*`0c={E4H&&U9Br09R6p)-g$IR>)5@`{^85UKKZVF
zl*iw_J0@kdi9rA4w-NunId&rW+An!ie%vze;{uP+0`rSCWyK?w+aGLje!Zd2Juc*O
z>;doL@AU`xzbp37C}MW3uP%@`yq_^Xb4j$v6sPoT?0ou4VQId~?cnRXznq%rIk)WB
z>$@DgOW2=13Nzn#xN??r+&V`(Dx+d-l;7>y*xj26UiXT#>&hpD)xG=EWF9g4W${tI
zPtN4U%y7sFS;~kVmJK!LRJ&}Rqlj8nKSI8BvO2EoS!CjyqWG#2PruHH!%u&1@H4BY
zPC4Ca(zV$qvTI|NMa=Wu{<J*@zA6vg-kq2HTXSaVd!&2Qz%lb>Do)!=_V5=``?W*e
zJ%isrwSDyCz{1M|%grD7|4Ws+PcS5Kuy@Ws_`0|%!Qo~_TF=?}^0tXr{Eh8L#iOYf
zBPAW_JH8olzc|+MGAd$x+s3amQ<jwW4`gPfIv%rpusXhoHB|0-tXJwO+Vf_`p+*I3
ziubrtGpn>Bdu?W7I0-?h`wE6GJVfTxNpT2**^%uynt;!@=7&KEVx^8c&SxNtaM+k<
zY^w<nwkTOx`n-Iu4+HrYhfR1UwweH8eooGRUXCa%=%l$gY)Up|YQiMka%5#?&-IOT
zQVb3ok`39K0Ac6n#|39&$NK4{SR6JZn{hM&!meX$&v%T5?4gt9;4q8KGS&nLi+;FM
zUwp3e2%Qv(!&EZWNE09|qZRIa5p|`EPKw51I+<>&2@norMwB0Lc-%uL#pAGmEZ}GY
zgwqn%cPpos4AMyvI7}ncOf&()D|OkoDC9*-Iw=Z=*<`k{CP28Te&W{FMFsa6$U+=u
z@|d=o0O2in9NI$ctA0X}R2+`hi>7D-gv$e@ItdT{mq;gt<1mFxG13GG@7{aZ)j-tV
zN+-?7;Yf0%sU|>Jd1k;`^<m^No%9V3qh!=r6CiwO)eW!Q_CHeTqy;!EAxk)#0O4oy
zuJc~EFQ2B95^$J7W*BJ#gpsV2)a+0C2%=N?p})^$t+ft{nQ_g8u#?biI<&tvnLh0=
z=ECm+Gcml2Vs^kz3>i>N00d&#B@~kYi5TV%O^MT~9SM^|F%IBp(ai)iwdiJp*;;gg
zAaIfnni!{3JD8xe1-5}icJ9DEkjTy-_y-c%%>i=)VRrogn;xfAJD#Ak2G&x7&K0;y
z2|8ckD<$X@Kp};6@CJ$Ym*GIMNKT{^^a30eTvI(KgywWGJpeMG*eozBfY1~JaR8wi
z2EqcMrhEI@s_om%0`TrixDp{e21+p{HEO&JKQ{F;6~ns0LBz$?Q%Hz&0oR`pcLa|9
zgt#a0^e4onK<W=wf^sr1tvmX=6|Q2(wG+ZCaGr4$^)M4E)4((tq2dIbWbg`#c>`}5
zp%MZ@WKad3;!af$Kcou>6)Q}gckCpg7|V+l1nY`oY%g{YoO-w%@Kn^B0p3M19^grc
zRQP}|AyV-IUJ@vZVlp6;z*HbDPN&`^1f31A(Q2L>aMNm@AMn#^UN{J!Y#v-bIGuX)
z5On6iTx&Y)fxXsr2!T**I)XsZ<aEGgg%?V_sUY2UgG-HDI<pG!E}=YhI@~(jLv=U5
zE&J-Ji(<xXV=nyff?_6Y6E2)v6l1cPT$tyMmCWOyEvQEi#V8yK4{kpcqjIP`=q(hZ
zbLc$i0Tg3!7(BSE;0^qR)yooQh+-DNLd$;+z(LD@001rj1%u#8|5eOA+&uc~OcHJZ
z8i&S(%|<apjv*HgHyk{?D(VdkHTQL}#tYUpU&5;r?HDbCwS(2cGgL2R_-z!k1eRJI
zb_UK`9rgh}S{;@H`DBNoad2_<0*7?49=M8nal>O2vjSFHb#MVLT6GWskyaf-LFi;1
z{;0ZUNh=d@@pln$1w!cEtdtB)O`V;Ul<k9m>SC=m(i?*b!dH4p$%_+_B&drZ$;4wK
z@wk|HOv+5oSc3l<ni8g~r%yJZQ0XYc*vK%%BuHXsLJMI|ox(EXczXGWe8qnLE^NNH
wjnLlm?@moExI8QC>9#XwIygGH&T{jZ4W<dqo!tZ3rc!wdDR^FhOh<wI36frjJ^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp16_compute_fp32.cpp.A7A07273739AC223.idx b/.cache/clangd/index/contraction_scale_xdl_fp16_compute_fp32.cpp.A7A07273739AC223.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8de9268ae7d65b7bcf896919b590c6333d97ccba
GIT binary patch
literal 3306
zcmZuy3piBk8eRrtX4V>)X+}mPgb72!G`~S&h8bgAL#cFGZe7rY<kk+IY^9_~33b%e
z)&)`Q_EAx~?sgYj9imRU>vD7wkuJwxGw1wf_Ves{p81|N|7*SP_kWkyT$x1j3_}oc
zoGc?bem$;(Aczt1O<kX|vRNBJ#;6FAQB@bYV9i4}!>)0=YlWq?XBS^I*jZe+`-^ww
zv|>_Fkwd#@ljp{NpXxhgW8V0^BSG7Q+h6(gL!~?rVIbkm&IV5T%lPyKNo|iBb;#fQ
zliUVo|1c#XM6sct^!b})>TM@XGF0#ttsctII-llzu+?{&yNTELZ<T^MIoO%8(H`4(
zx3B+8ptWEm&GYW7zAPt2&#0`c(lA9n*xYxD^}LBsyVu!RC2S1lo$;x3zWw43bzh3Y
z{`mt@%+I2eVLxrRh&xET9e-octiLa3Zef1>hV+j!v;Cix$z<bkb)nh4hwk`kl{kjB
z*e*;=itb*K-pb5k=hvN$IN#kKb@iF$p?tl9)72LQ4HmK$25|+wR7GNW&X?hruD&r}
z#gi!!<*oy|Ya%4~%9a<<x0V0xA?2?8M&mLc!RnM+hm=>c16KQPxrFpbEBJXONZG8L
zpA{aD7*Z!@C$&HM$>XT>(6IH)BE~)!w^J3oSf{VN61iFKN=5oj9lhqhQVFf7y8LKX
z)y16OO`}|SQ8V0D3b);wA$E|8{i*%7gQ1i?!Yk1zSJ&)(NBGvqpBLxT>+M_uWY#*%
z8;-}AdOI{9+R_zu_2ajubbT84g;vAsj{09qZ*1FRVLDR%;nJ!?S3SR%ma7^M4{WGQ
zJK7kL^V)MWFqEbhcV3vEc`ovP_59u>ep2~Aom(nKeEOsE8Yq5w#+zg1p0=B<4mTdy
zf9Kr}^V4J3+7&0P(*o(MchxN$c;0lyde*wDRl8q&_<L?eILdAu?cCxWlsNijc4lt|
zZ|~m7`1MA7-7`x=KHVHW?RV|Bj7k648vc2qYe24PTX}K8u=)1;8y(+lta6F+zdZkl
z`-dNOpYz_z58LE3TC?}%%Bnxwj7?u2;X27-Wh#Dd<yyhcoPOt{Z#sT!`u?CL@%QU{
z?7NFtUtRN(KejrtSF@elhg(a-BCQlZ?BCSglMLSUc%@eD{w~<JAgo!j|HNh~W9s`w
z=TDUEakp;%k@s2gsiQ?cSyXfMm|IAe_NyLwZNeH`=aKf>fUGL>r#1K9$90%vcehX%
zJ>8nnZgwo8?Pihv2ak14ft}-FbC=6KPrvOJNHcR2FRhzb)I79f;Seoj)4?UN_Y#v*
zF0W<`JQ}Qjx(?7+S=>)O6I_v2!Q8XAt#kNtKs_^F#!reqd8BvKFstoxq<Quk`WHzb
z|C(dg{=V#`L&s`1`n8Rho_3WE4hjUWtS*C9Ew!S;PyU>)BrDFrz!7k-!={jW;G3Sb
zc;`(dH8VLTh=d^21G(?ccOmmoQWS#VTr!uf3dDY^J~nN&{a1Suf2AW!3E0TY$XXR3
zY@6;(=@oJt=*VIMW|%RoRRO}hw6y=c8kSE%NplI<m~6~Yg$cOf_}bD3*EgZ0NCMU;
z>$6k=!j4bs`M=B`eTI_e6R-){gsln?b{fq**E-_AA0-77Fq6zQR0Rlop16Ckpr!OU
zN(v=lDw%4a3J{jk@^-xpyHbpjA_y2IqsFQL;Xp>n?jv>&AE2ZK1k5M%*{T5H717z<
zipfPUP*Mm1)5tU<Re<nXZPqOcd5Hogg%L1|%raC32yeY>*X(awb_qig2slC~f}#o#
z-t8mNioXBf7?c!5z!WmYKouaox1z>L&+|efN}5N&q2y3wRe-SK>~qWhq2U^ow1|K)
zGG?d>5dLZ14Pp9)_X#LzApwiXBDN|(__6HvIpLklKcl2*0;ZGc2C4wz!JT{SFNcMD
z(~%_vY-na^tqKrEQsNU*zvv=}R^HIFKBG)aEeto|m<V7e;jn4Z{&L84X@43n3}*wj
z7kn4Pr-SKUkO9MIfmvR#OBgN&VlQ|#IFtmPaw1`J7;XwoHRx=Ctp=R{2sG$?f$szz
z9Abh_Il*c=YhW!_XXgT3#Omz4fwx$lT`&k1!|ZtfA9{jLIq_;bOJFHc(>VbriJERU
zm@QG$$$?w~>EIhA+MkvK!$aAj4scZ9st}sWJyB~;1JisU1BTB8Gkw&WUck#otr-Xc
zeW0exg<s0H7n=AG-4$^}0(cA*UX+Bev0~!bSZM48>joE*5La%YTATwo-fD4sVDGIK
zcLQ$TYH<mWctaK6wB(CfC*N8SDqId%0AKMd`Z&0%{#%p|hEv!S4%{IOr?ROWm;#2Q
zY?K36AH(TvItNZXJSEapx`8^)sbH#9ou&hDkU~uicL(lLb(;RbUkcNNr^Ki$-2>7E
zfs!@GjyradFq|o5`ohLwI7`U#g|-3Rg-AtdE%03oHv?uObt*i-6RA@X0-*?sVz?AY
zMKBdeOVBB;NKI!2tTejs44gH(KL^aw=zb6gn&>`sFXCoO>w<LK^)4RT+Mbe2^aV4+
zra>Fp9-zHtpm^L(8^aA*h8$=h7;eNe;=und7|viZIB;)a?}&Vr<_85Zd<vMNF)(((
zPGevI02%}12mB@mreto-sgc)bV~LW}*fb6-E{5x~^*L}G;ieFmQaUQsJkZJ<%gs7<
zk+@WZeS{QV9bS&eP-(F6uQA*lm}^wj5jbj8)B|{ER8$6J6BUJxBE*%34C&xx5-Lg~
zhQ}Cg0W37?FaylcsDmf))Tl!M2$-nD`?721v|>JoxC@`d7r<G89U)Ydey&zA0VW!5
z$pu^ux8wu9hFgk&Xu>U1;*)SCM`}uJs)zC+RvG^HiGvRa5;O6rjR&U`jOIV8TB;Kp
zw=4!(Hi00q$#F@`iJt-Sf!aE{WIYNM#ps3x`u;|~BCZk5pD}q7(}e9N^zfYRHOG4f
zi|1}7ur>dyQ{yQdGYiXU);81a>>ZqDI=jvSQ~9QjE@GCkL>5o-3-^&~$&o(*+X9Mj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp32.cpp.1A2C9B556CFF860B.idx b/.cache/clangd/index/contraction_scale_xdl_fp32.cpp.1A2C9B556CFF860B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..694b3a5cf3c8f26ae5d265ceb19e3fbc25a4f057
GIT binary patch
literal 3292
zcmZux3se(l8XX24AQ?!)l0YPphlo7oVa(4nkOTtZ{Rk*XEyGI@SSUypbVaJDX9W>Z
zi=wV2$ioE{MXV2~Vnw!!T2~RYiWP)bSFsA>o1NtNGwC_aIb0@lGxvV;{rCH#<uchV
z96_Y1@-^9Ms|XzgL0IHp{;Ir;W7-IENgqMh>}!pU&lwgn`={f}*z=puE<MlKUU8&y
zvgE$wOWPYGy;h->_6+lnll+92ta}4>8OX(vbqn{s9X(isSR*#qcM)jK<ABwBa(?)w
zO$YlpmgWA;=O1S25z4Q}&^I4sbNZZd*{dz@=?7n}(K=V)dZ5cc$<tKy@k6!Hw-`V1
zc52Y}Y2W4hLaq4M1>V<3hl-q)gH!VUT4r9zA18;7vWMGU<~5Y2eB*IC+~I_{*0t}}
zHTk!BF>yU@rQKPjamrv5v3+H5Uo7*@_E+P+%_H4zXT1t*e5_ymw6m`G_=2dlcfaGf
z>0aVAB^tLb5+2EH_xduZv#h&+>B<M0f6w~QBg^JPx?AiUJB7zA<SQAeTkgu0nKi}l
zCLS;JPkt{o(NoygysiTFO1qIkLTW|JHT%6#3(e;!hGj`<l?x+QmL$4f@!p?bepAOD
zRrK%dF~9%oZug>itNpnTu;=9ky6m*T!JRLov9T*Qsr%B31M-@+U9u;R2rh`9P1Q$5
z&)s=7+Bf{+&<cZ=J-b4K!V|)+PF!K$`<ZEfC||Lrcb>ggTYr_ET-SMGXB}9-nAn*4
zQ(3T){snH<;U(CPknE;|i{e}@!`F^kERW_R+YF!W^_Gk4t$#kITgr%wIe-SQWJGP_
zXP?S!btpUZNjk<lSF$E{%Jx#{;d`wOSm(`jC2_H3`{3<%hUewHO#!|6?M&X%M`9L_
zRX%*Ry~$yH)2JD}BJI$M<dwM_vu0(Q=Y%*$S1)ei$_=M!O&;YtD%U!P(FE;ZA04W0
zEI*Z(5g+SsB93u;KY7c!?OP|eWj_XuSUu>zvo{JAym`CMyMB70el$})Hte*q)NNhB
zVeOnHE?sq}trLQG-&{Rz<f5<WvX~wzD@y-nUh9sIT5S37<hXakDz6E*Y~POO2@|e|
zt`($QULTQEdu}{r^HN!vOXqm%t=qjTh8m`y{I8>SBI{os*Lv4GD0VMqOgSx!l?O{J
z6E23lX`Q@rCd4GVv9!fKqDXsmFeED}r^%(GIBuTLwTc)2%~<XB8)NhG?oF+^{&&vv
z2hx5xC@uJ8(&1D+QEV(YPOE5puh{x9;w5sYYPe-~<`}o{7xs$}-km)ScLqi_-?zSd
zI;G{;uYkd0zunjSJTv+;s$yOi*Z;m{^0;HO%W+~jVOGHlrSwByqDWX{@UF*C(Bo8;
zHe`0R@L)_=-OzON%{|IDOtacOw5OL{zlgKERMgMN?5H%1uzBCRs<x_1zW;-^<(Z-<
z`GwiZVJL#=@89tBTt5;=N0%T7VT0Ll)j&?R`r}YgFg`jFVn#=!Nt%OkjMOkgA8XF7
z9=N=oj>eL-A!f){1EifE9ua)Mc<K=yT};xZm?>8ckanIbJlpj;Xb&9?Cuw8Mn5hOx
z3vWp88AYBH86aPfv=QIPS`Co)K79SamNV7ObTpEr^)Y>h8Xz4=E7|cl>c<K?svv1P
zM(3yj(xFBXmGyS_2Iy!!NeeImR}GL}k+}Ab(xiNZjz*9)4WqHt0O?$9_EkM>nUao1
zku)1)Gt~g;l8mWXM~mA$1LRARX7O3pYJl|CI<MF9-LX>!$TE^<@|o6Zn4z!RoeZ+w
zbpc1xNm`+!&{G4XE5$Oc#9N<|>1Y^9>tT8fH9&gz-bQC#@6I+l8b{KRSR_XckXD`=
zwj6sk(MU&^k~EItOf^8d;j3Pc)t%4N>1YB;OE3vn4Um2y?>p;p?P3QVO(bap%z&W=
zNF#Y^>G^-mLJ+NzSC58Rg_c@4VahWV!W_V`Y0>^NWV2|03KvdT18XUK7AM?*n-nVG
zgdgye!b8M~a1btqwL5$cB`K#$BrFamEP<trl5+;mGD^+|_{b=^5D+4Ra_|Y1_NVaT
zL?k!T5q29!ne<fUnDVp&R${1t6E46-OnHicNKAQ#f>1H^bnE<n*S3F}ipl;-coHGJ
z4JU$hQH=xTG8@bmQ7*Q?Rz$f7flx%b_yd0tlmkkUUV7Bq3ett+!4bh(hl54BsHTc?
z;Q^i><zf%){U{fA;O<Ac$bifby7(7lpD+6N6ARMChG!#$kHGmPT~xzKxy%7`0x1_q
z;1~!W!3j^`8A!PVfuKO>0`HR1QVlbd3j<X-9H+7!C{7rA82iIx!U?tq+aFFb9Cor4
z)uh8`ae@!{5~>sj;2@z&@c<qYXo?eoAW#BJfwH8WYBNxBbHQAVEVu$!jV$;AUyUq;
zfv}k@z-2<psWu5E#|2!CL!S@kYaIF_ut?+3r9e7!=y3Uva;nWk$(aIE4LKWNqah~%
z0u4C{kWg};mlfG})uw`S+jP%=x3xQO19|8;AGd;AXIrrLre_t??%FuPWHWiN7Mx(Q
zSv>fnIAO##;=!!Lbw}2#+I7$XC(MAEMjGvaokkh~02*lw00A>;tcq?t`uew1DHu-Z
zarJmm9Vhg;`aGC?oS<=OJQxd{pmXUwm{*)I;2QAYBE$(pt|1R@9Jme0W~de~taX2v
z@!JhWN6(YZP}nO1w|;IJYzp~isyz&UjT7d;Tq7q=z)2$~UcgHuCvqU4$q9@fX|CGa
zP!3i~x~TRwyp0nUz(S*c3%~-6{&@p$js69L;F<nC-_>JIs}S(WcM<ReLO22B=FcKj
z+WC1Y`CjCIZQ|jx3CX()W^p<xsY%I53N%KLRO&LBx=f-jQ?gUDmXlwD(?Yd%W?{N|
z`gGiY$uJCJ`AckAv>+pszZsiy-95a#eMG)~3)l{xbA`6%e|48*#^YO9T3OGVZ)fl5
c?Bcp`5tt*G?c^q9b7b-~G$2MCs1<_zACCNd6951J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp32_compute_bf16.cpp.301F203197216C44.idx b/.cache/clangd/index/contraction_scale_xdl_fp32_compute_bf16.cpp.301F203197216C44.idx
new file mode 100755
index 0000000000000000000000000000000000000000..29c90e89e36cf388827f22d5177b9ff0f36125ae
GIT binary patch
literal 3308
zcmZuy3pkW%8~zMU%=ZmrGLvBtG0ulH=5-jSp&6$Ur9_F(q0;gzl4$=<tCT}TN2%4%
zPgG=6*&L#+P+?t_Z2qXIO>MM)Qrc+0nQLF)>~*c{n%n2Sz4!gS?{h!TEb|qMM=%8O
ziuK)=7`FviK@dcf_@rz}j=!pmAX91xvaRI8im;>sM@HXQ^V``)=dVVzY8=eJaOktg
zNuzv<^q|!Zt*RH#Vs=ieTTktOn<da~wR0N6vWC_~Qwosd!q9zcQAg)LiEEioZf9yw
ze1GX#{l|&K<F0v+PoYoW>!$Os+i7<hjM?t$9-i~A@7A`@%-y3=pC;ZXJ^u1#(Y=>L
z>&wmqgR}va*k2y~E<CyOO>^$LA=9uVQHS-dM?0>s%)Ni(@NxD*wU;B#7w%CLZ}XSC
zm`2=@hH={2v`eh;Jc|gGDsKyiP|v3>-PgaC7ovB68ymqV3tBCgXU}wA^L`tj72d(s
zT3wnPT;JE<P}4j1t>FiI+UlgLZM8qvndD@ZAJykpmrOox-D6zkTA;VlG_MAq8NGk3
zhjHMoM#O@kGf|KI2hZkKSoWyvGJ|&w`1GKR<CjWx0L3z=Bz+k=k{$MPwB(z^ydY-j
z4;B2GSL#=9_TFWVX1*6}N;>ay*33D(;_xwh$4Eiw?X{Dk#>bqI`{HrGdE5s(mg<-|
zh&rtv4y(-M^tBd;b!Lu~zc99M#Ok>8Gs~PM2eQ`uhsj8>K2f;kooK0H#&8--&`{A~
zW^#y8(pHfjkNo&xA2U3LHSc+SntC$tq0+@yKi6~?X?goG>Dz-oMLFCuFD?H+<MOKB
zt&M%xZmg<y-=g29yT9z3WOVRc(-Vu0KU*dWbOiTaYez``C{3}FtjZKh(6%}geXTal
z-u~`@9OvPzgqFH@t5#+^S5QI6%8Qem!=np}8=|{?l0Pl==&IVA`@ik+qSA`X+ow+p
z6Pu3s->o*frdqt`m)gpV1I-DCem5HQwt24A(wECDz1{Qcu;FyOkwc`vvq2@szVnV%
zrQU?Meaik^xWPQkZ>ZV7)i`Ks&zGXb)hE=B){7qhOQ$4ncqly}-g&}4(f!I;_{5_B
z-bszQH4_kZvT59Z-)eELUGsQs$it2ekrhL!TL+{WS9-YXW3RI;?w-r~!|JoD{-q^@
zWA)osn3$*fY^2*h%kT4gV-w+Xiv7cL;VAw1dfU3<<})|?(MuMin|EY>AJ|yrxzkQK
zkC(7_VX<0TY4NlLsBRs<*=z6nRpNQ-!n=;@$N}yPr5cm=Fo(YG#T&aUpH8J`<VkDm
zI_ZtR^(pO*``VxI4=abcfj8x0mA--TQ<hTq^CNM-VTE+vM;2$V82tNIZQ8^3!G!Nx
zdO_aioV71M4n7jTHjnHb|62arz2I+~^11@{e)z7~=3GIvD>KVsz$eSJXu6BP#kaFO
za7MW<^I##8lAahXr67pfsa<cH`jC|<Wfg+pJSvYP2cpB}KW1CMVLzLvv73&pC1{2o
z!&nZGwoFYOe>vf=i&B;mG@Ht1%3+qiShcCBzkLTvSwYb1RCSgdAZ`7$hTjnS@i|Hf
zC1`D`Hb)MSw*8pi*fQyJ0;L2Jv=&v1AqPkct9wpoT`Q_WDM19SMpe_01EeLi-32d#
zZ|0+v5Q0XjC|eGY4qz@lRBk@dk5a-2nos3(<N)avKZOlSU2S=EB$A*t^)!v;0O<{D
zwm&MHpErV1mJ&3LO4F1Bq&F$E?x|AOmZ6kjf@V=!3^_o0e~<ZPAG0HE7_yF_LsUXk
z<pAkJo?@jn4`!lKij<&Lsj3=sfb`MgO53@@<~o$JlAwdAL2Nldy6oz(;jix}DpAU6
zg2t#ALk^Ix*xcci);zWjrGyi-h$`a90n)t}TdYb~fA~sAA_$tP$267$qz8Sw8lAd-
zx`I;H5Hy`i*N_9Gk>t2_DWB&ch|=!&&xbVA4V5rlo2xB=g@$2MqWxpY=Ft8YE@Cgx
zc7uB{oCkPrPyxgFfbRy2h~Xk2a)U>Mp(NyFiG(kQ;l{w&i<Gkm_Fkl%2k`JB<pM#V
z*R0%MAtvNx2`1$XfuWd`vjw(dQqC2)ib*+t;4g-9a0i9<x3FV)5GTkA))q#e@Ra4A
z^fUrSo=^e9?SP#p>FEaCJW0<05a0<t?VB5p<Ya4m66FwaMFO}C%G^}f1yAJ@%WNmM
z8@xBzScJK(t4MP$;Chhe7Qn)TG<O7!9;CS#h&`Z-cWPp5#<|xE2p1lgCxBaEe-bXT
z<|JM8fxd)vu>w{SxCO(VfwP2k@c}*(=mJ-XXvvxx%1J?C65BerfP&##PFmjZUNM~I
z#PWu13Y(pH6<Mdly%^2`9EJRu0#k+jIRFQR{CNSd+5ExbK*-4^11YBm^hD&zEP<tn
zJed=462UWIxCBT<@MQ2U2|3v$A>|f=g$h|%1Qsb|!5z3OWI+m~vsr+{hqx%&<biTI
zb6d~szmdF)s1l}!8Nn%%<ENbYHvg-mGKMo)3@-dSz;I2LCKuithBH}AE*u@Xo0v7=
zvUfkBq|Q;-v;3=l;8Y`yAscDX7sGXdu0pEJfw@Ae000W9S^}2LrmApRW$omvt1)vh
zT$Q8Bh3Xit#!=(K1Y<ajL*v4FU^vP_xiEJaPUq0MaG=4)BW^`DQQ_O2YSEh7l~LPD
zoF&8}L;~*)-VSjV*?@)rjo}8sK%peoz*?asF2F^hB)-6Rwj{99gt=@WLpgY%go|tt
z!(|ND0Xhn?Gy`S|u@nG-LM*+3_iQZTQ3*NO;3nl3fCUN#umLs-1rP$ELIM1M-)sTK
zj@&k&<@343)$qA|0nBc4TmmkWPDzeQaUp&y<86g{r?-~!8dPFpqoR=*=!+n+<YP4X
z7)3tDB*rGJC;s${3s6>>L!GOthGKMvhPscYw}_`n^I^`Pr=`ttbaD~8y19GUuq>Sy
v3d{`tX_c+Z)mva_WNcz;Zee9>w`j2g(C6z|+k3ItV&6E*lI5NfC4b~k<dlxU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp32_compute_fp16.cpp.F7341E88A32312DC.idx b/.cache/clangd/index/contraction_scale_xdl_fp32_compute_fp16.cpp.F7341E88A32312DC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..80d5bb01172df881771dcd77bc8a3d802724d6a0
GIT binary patch
literal 3306
zcmZux2{>En8a@sUk`uxtsSt{)J(gO2v4jYT*ox9^bUA|3R%_^{=F(Ps8C7d_Vf2nN
zqpFsf(nVX8UOHV=r&?NuPRs40bm6viGbg$Cm(26bd7gaF$^Sm*eZT+v-tP+&3x$s{
z1QA4wcP7Ma#Z?dl(IS4yTa)6?DkI2@I)d!{zBXdzmH`*q{aK67oTAg`R<~*9=G7K_
z@G3RVLxT=lx6`Xf|B$4A*5J%!jb}RRwb^lpvCQGfC^R2A>c0G+>)UG0n{RykaWvP=
z?wMy*NBpbj19tz=ZE~C6_7uF0OT+C-JO|ezWnGj<wFllzEkf^JdHww9aO2F3<Y;&6
zZDITt-o#mTzq|EMQe0>6r?PIfEctC{U;fwG^`Fyfb<VUOx<>ADm>4OEh|O>wEvYFh
z%2i*wtn<zO_O1JuRYV;yjOeUnT1HRjPQLQ29%)ay=$3NS!{qh#;uGn=*o3A&Ebori
z$hHl+V7_YOrpUgy9WD90So>?wg<t9$>@VyRgav47L<FZMtm7B<8eG$8DiZcJOiw+(
zu|DhSyG$F}NxLTx1MZy6`N`_GhMu;)^0|UF^nAQKW5Zzdik0-iB3D8A`O)Lu7liTX
zA|=a>{!j8I5(C(mtqX!(wq(5SY|h9&Q4!~#Vv#W!lN|FhCfzrwS~)Rcs?PBWfBb!A
zXxPF-gJB-Qk6PoXr!N${7wv7_>$CZY(cucCaQ~;~Z}J1eSyv+i>P+-6^(@UAnqIx1
zT4%A?Uu>eXzM(ow*UP4<eBb>Q|334wafe2d%{`@tXFYYl7TwG`WTN}-hu2p&WIJRO
zUKjL6`3{V)8Y4@pzx1T2Y%$DzVen&W{oCy|C;9b_590b0Yb0Bp#xuI#uyL)oY>rCT
zk)0tbA|Huca@jqWXY}O3{awF=_1Zg~=eO)_%Dbyw)O(eYEG+CuM>p^}(IZ0vU!hL5
zzmL5iiGNZXxoQpLj&BeBW}3%6tAQKL3+9i5MvR01*m%oaX~U;)c?Q3wTZc_HRZf0$
zeO-CZih>tydd$evCvCqwT=?VE?-x}9(PG<cp4^I(W9!7F;(?ee4_v;KiG+8{kDUJ8
z-@Nfo*CETkgN%<3dlKHZ*s?Zr?Y*}Ti>pr++-V7H9ea6JZ~NILWsjm;Z-tjE*L)v8
zrR84OGtuZ1u_(uN-^zvl(-r^iX_io`R~lIv6%OrA{i0p-K+iJgvRzH#ahhoQd7~R^
zO0~lQZ(r_m;#i{Usps`c735FBY);WsI=%M+Iq&p`rdy!~k5-YtkC2SJSKYsa9{*QW
zU;mVD>0m>tkK2w*8+Q%{3Ap*A$!CH~wv~(=I^3%Bv7zwDq8PFBrpTHf9;Qt(T8APQ
zq@JOE5Dq(ba&{dXPF*{B(&6V5!?WTC$10|G>Xnv|$5e|sHEj(B-*uF1uIwoK^-HX7
zbLB{w*{qVtEcbeHN<vf+iXiG`yT>lwN0w306$pZxl1y2$K+L!7vD&un)qkb}^eAW;
z0W(QVZCRLu8>$nF`n%F7XaoUkkTe*w0AbFfI>+<N-#@0H%L!PAq{EU02;06-x!Cf~
z?-&IQCSW>=PLl-)yPvv!JhQo|nu3N9usTUyQx+gBBJar`4ZWI2LBk1{LZUEb0m6aW
zOA0D22Kp)JN&<EyIkIE{!og#-nC#3v4i#BLz%+fDi7Y@kE;6;RV*bGq3c7@V$t1Ft
zEI>F>nQ>E%w5EcBh7vG?#GuIngtKm2H2IkqU%`-A0uEORSCa(@7x0Bjk-cA{C}<D?
ztC7?+WdXv6OR8*D-7o)4L6;G52q}aq3lJ_DC1*Yh%qgWJYYA9OU&}-mAY5_oiP6)^
zsVWNkH34HJj3x^Z{^^@*+#Q!EVkzh<0`?|(vt$9nr_$)hacU2jimWDJZGCMMS%C16
z_|`>kcgI-@8cD!Z5>-<cAdDo%#3p~3hagIOCLa%Lr5Gt;xDH##8TtU4O^N)sCYwk8
zi@7k|1egfmw-~+{EEYfp4EF+F0@z~=4+g;kcr~m4o06cDIua&_;YPqnD5tXpwn90b
z2k;Qe=>kB25YoXfQ1V~Qi{T-x5Np^gXk|iE>SMX4F)-#s1`M|Yc6_-e5AgVM%|H;y
zhnkBopD#X;t-~kk=gszZhR2|SrxqJJlSdpgxlA4$PB<upxO7V7;%vb7l8ajcOE0;&
z3vlt0iwl9!3##~ROK96w^TLo&F=d-N!%ut<zJ0Ty?$Zk5Sc9d(hE;}RNQg^kP%dr&
z3`BBqYhW#cOc?G8Tt#wmKj0^V;_#HPh;(KlT@WbT!sO)SqZm%-(tTjLF`U6=_`qp|
zGfku-9ccJ1;X$Colc!<|OnLHD9DySbiW1cZ-aMEJ+yMlgbOFfe^nt#&oX!eZdCTdz
zfa@)%69JJoq=Q)!bkc<(r&|aXDp<rG*eh7X6L>0EBnSk}Sp@DFqEONW1L+Q^wv}hK
zC+#M>gz00(a2p)(SKc?CH|wH|;WP$~4J(D=S_~~V+$b2X&Cq7U6#%ysk*{<`K>-Zc
z19}RsvH%tet^xokxXKs!&bg{Etg8OqvvZPp7_P=rV?%ZfS7)iSp@T7;%p$X4;W3=T
zqOhUCFr3Pwvf&1TgGXFnx*=h9@G3;o(#;5uF`NZh3KlQ}W(pQ?0!|7R5CFlP1<G3J
zGrM=yw-MD0w+t7-n!~CRD$)fjS6KiSD72CTI0~(F18xef6a(>GE1^S(yGR!@q=Sth
zRHTa-9%HyJ&{epLIWSkai!*RmxQh?)nY#<D96={t+;TcYV5rbf8(^c*Pj}$1&`*Eh
zKiAKR;?4!+JV!QB8b`LHGxU5?%qCn)otz{|c9XtHq~U*0G<-pjsJT~VT(8%mR&~OI
zry_}77lo{wLl8+q^rrR1U;mguWtDj(RW)@AMx|+L_-Xlgn`)8$wC8_C*I~JE-P}ER
zo?bQ#E7yh2<_rGrl&QzoH#9OfS!8BmX>Ds~@8ARs9CbO11q`N89E18U<%^U8kUs$o
CMVD#-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp64.cpp.7417F385FB954651.idx b/.cache/clangd/index/contraction_scale_xdl_fp64.cpp.7417F385FB954651.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0201b38d42d9b4fb10ada3866cf0a2ee3f3752ef
GIT binary patch
literal 3292
zcmZux3piBk8eSI0n6+GHGQ%)*lWQn`gJ#UQj2Yun*pf<Rxr{dHLP~|rlS-xI($Ov*
zPCK`x#IAi*Dx@~u?2@eu-P<LlQtisNan{T^zcu@L_B@aGS=KxAzW?|A@ArktrP3Z8
zL44!n>B;fy2`vOcnB-sT`jl@QG!f*oE`p^0>-fUOYx>*`dnasv4l=#d=lEEsqodb9
z^H6Y+ET>w^A$IkAc;DEPwPRn#CH~pFAe%_J!0pON4GBV6$ljM%-{Ys*ej5+0Y3U!<
zqK(z8cOUZ1GOI4A812%2KT;wrU15^{!eFXX_WV&aeJDnl`H$B1b29!MF;31JH>v*k
z?qulwg9XiBv?^cS85Wj(_u8{S@yPnyH9nmVm+xe?MC30f0^+tk<8Dm3VRzE|V9JeO
zyW}0Yw)W4y`VeP*>Ovd+3hiRUFt&3>`$*Hj!&8Ve56XnsEm-MK0&a;He}BYtC?vsI
z=d$sUD5K-9{Q9I*f))Pf^Dp);O?()*dexO@(`&1AcG%UN=Qm7~CmO`<xG%3rip==>
z{x=s%%&4z1J$(PVp{kYP(r-FXwr`hQZ%wW=){fi%@sOz-H-e-6rLrr;X7!nx)cVj;
zFTd=orQKgE1pO?TQ()zkuD)sK8s8o|{Kn>B*z&87Gw$W)pL}z|-!ToT-}A7<%KgY#
zIS5G1EnYg<q5G%K@f_Et;)Jlmh{Dbx=Sv$vZon%Z+ahJB-PXvYD<zd%jW-9}GHa=8
zXmgF1&f1y#80RmXt=U<{55H+zm%gv;bC-8l)jWh_e&2OdF}J8#(r#fSiVWVQ2npMg
zx8%*hSmyDxZH+zO4V-E-%UhII5Y}IJ_CaFeP5WfNDZhVk|FYoGL#aYpWUf$#wGGd*
zFluAoa_@>tNV{zCYexK&#$!goL3*nTPeczrOWx_1_+gjp%qveGY5a7E>74SeN}jX5
z_wDanKAakkaCA7aYhcF807XmrA0<QayQ7Tevd-i+y{}UI6Vb~OO*MEW=GTra=^jbH
z7X3qhWYJ)&87J!4O2?i1imN_7Xx0isOPo5z9u?&WqU8tWeerGg(#HhJ(k5T0dzsB?
zcMccWrq!DXKFiR)=gBLa`%_|a-^5y1)k~$L{IL}YO%EDQW)5muk0w`lpBp^2^zc!Z
z)VJ*zb5=&8=c%`MeG(%&1M(`}9+!(dXrH$;Vn-7ec29Xq7ng6a-Z~zgZWZ*EAzQE4
zpR?=_Y+m@wiReCO`Gm}SX5D*}<^|j6<CX?%9kjcSHf|`teSMk6f(5seR&CxBy!vkN
z;WEL`rWPA-hv>(R^c>r(D6{gIHDMk5uJ_uHl@~j=?@40!w0+SZzP4`DqiFtsndU8%
zrh$^oq;oHF6fu4+k7~OQRD3lwJ6Oti-BILF*V<72dwc2X%J#iiM9vSYE1!p1PoN94
zHjk%nN{$Ie5k$8v`}MhABmzSt5rnX!S#ecBPPXb}pO*IS#iu|s3=Jb`4voW7!6e;K
zyLRv0jw}paNYeT=eYOf9?eMrBoL=<hDTXd0X%m_WR|SxE`m(9{!e{>j7&@P%jc7)O
zDuA@`=r83v&hD+n&=8W=rRf@|0Mar>PT_B%mkKa6oTM=t#!&&JgIEiSDsB7jV(4O$
z1~kA`0i+dzG5vviwTJbP<s@xrZfL0jNUx0Aa;L(0&vOi2K++5vgQ)^YuhnE<r_+{K
zU}z{wvuSKY6+k-g7u%owZA#j3L_yNwTH$mRK)T3Vsu9&Q9)qF5Bu%H$4O9T>eWf){
zI>Pgf7#cy+A+!*V3Lst4^lbXekDqEVbSX*WG~7@Hkgi_W>9O(rTLp$LA!#3)4_5_{
zu8(;0bacFQs~)m~q?zVSOBF!+q5MX(M_2oQFf@v!^=NtqDuDFc*zWId$5$BWA<Iab
zWzMox0i=<Xctz^4HkB$3#^0t%oAH+v;e;iy^nzi*31{H!1#^iLUck!>UM5b=2lKsP
ziT3|bA1SAd9y|q3Ob63_DLE(L<V(qU0#9E`E)WFzLOJ*gl<}98;zS5HWH!7Mm|4<O
z8Cc452ACm&3OF$b%n?zZA|Mh`o<Sf;1U=m<M07>yNPz}Ua6C97IBqcdq>D25lnW2=
zJSi7DVCPA>xC3`j%0&vKp3uenm5n0zSVT4H;=}WCfv@2NkuJ*dp<FD0g@kgM4Q5Ly
zmw8~GgmUo*{t_q$-;$TE95*Nz>{h&n<B(s7;)IchkpwmnC)gfr2^^&5`e!Guon{u(
zaDvXI^WcQwgf3T?2fKk23@(EQM-wM7F2;j3;DjDmj|bz66Z%|z9?S=vb+S(7pu>aV
zgb6SaP<2`XD*;s}06+jMBHIRh1h7tc08&o57$`Y&U@oTQ?18<QlJfu_VoFX1WMU`>
zTR_Sw7YQXd6U<ag)LbxEEm2}1R!dYc2%bz7+&$ziC>IZu+pW|3L*B)dY%&74IX(j}
zk=+5BxkCjL?wUAZ$TsA`PT>TT&E&ydffFn?iw9>P-VJ$P<?4b4IAI3N)Us&{Y}K;q
z2He!L=?DBKvk3=_Jg;)qLAkOEMkCpqPqdQf4Yv!I!K&c7Npt1qgWtx9sbH#FvmAhf
zTC)T|pw=uokWV%X-T`T@+>}tRX1qS%w#Y?Dngbqif%gXcL7FQUC;T=}Z~<2>N7lev
zEk~}vRV_!pz;`l7SI3F)mj}x4lP)Ga6Bn4An?Dxr|6^j~HuAL<&&mZx2u7PUSFUBs
zd>WXh)^<nWsMdBN5URC300d079o{A>r(ENdoGCC>tHlP`sMW#;e6?C6Kr&eix%pQS
zN|Dr**i?b?Ayx_h`^3QqM3YD+K7AOZ7cAF`jf;*!VkQ|RntF^)j$4&L{u>Y<q^YG%
z)1m8PxSpYbzCTmqW5s0nvy7)0nQ+}b1VT@d*vpY^KW`@AX6oO)<(Tozr%j(>Im_DC
fZno2$xh}457QocO*_X|c%HvVLZ@gt1fyjRWqVtgF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/contraction_scale_xdl_fp64_compute_fp32.cpp.02A2D58B76014A4B.idx b/.cache/clangd/index/contraction_scale_xdl_fp64_compute_fp32.cpp.02A2D58B76014A4B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5542395726607e7b1e986b33614cfc65f99403d8
GIT binary patch
literal 3306
zcmZuy2~<<p7JUQ^koOo8NhBo9h>SAiB7`ZJgn6b>&_eVvDToC@KZBx(;}<N5fS`cC
zfS`zT#hE(%|NV-ffMv33D+*Wz99SS~p)Xng^3t`MwQ_j5JK6i3d-gd*`1|<0zz{?f
z@1M0kVI!`BAPAj!XKYMQx~PmG)9MJ4Rds6FiXuI}_So!vriW^qx%K;-X(^9(itdR&
z_BsB;Y5i|4n{(6No_X&0l&kD+cEQy6?dPWGnctHaA?kVeG7>6~#O%(Pp(l2kHrtoZ
z@;M=!L!a;N?;dH*P};K9v}kAnwQhRw&b&uStB}-~!bUrKh34*1!E8hOL&G~MGr3Fd
z=X~5SRg-7)#kFS5$J85#I^TXVdlRg7<=Cfz@!owKR#q^&Z#&$oif%0H|5tlOV)Af(
z;1x^j!9F40B0L#>tiY(u@7R!)+LpW1Q%5D~c;(Y-_Xh^^vl0S21uG66_ZSUBD|f`A
zlEwRa^f<jv1J!o*QXl+{Epyz-xh4L6@#C%?qOxu&C#u-&qUw3kXnm*C13~WR4{Vcj
zfBb-|Sm^|9zQ-x+-1w>=7n#t~IIpvnC>botpwH@r1H91W?a>Zx?$sIlyHo-wlAhB4
z8b1B4+$Cp)Np;#Y>e&8Gs!VU+&e9JN)Zw<F`+aRQ`(8$<t{xbPNy^_bFz#Ph?_*JU
zET`(`wobh$o^@q$)sI!;#6!mAb;go_VT<?0ff7pW>4o2vXi7&c#i?IU-=_BJu#cHB
z@_Bs{2Mwb`17X5-&9TC&$&RHR^{xm@f50ibB;U`!7T2N~HaED;E0q}3vaEBqty&j@
z+x%?ky8a#Y#qqK5t5a5~m-gvR1RUAFTjPSOPh_{}2)B*SGJe=8OiTHvs&|?gwD7uX
zJ3bj7Nwsk=FRT6j#4U~U%;wyx@``%%(-piD?#%qKjAUceDd}PD<4L2^pB9C3QeSnQ
zZ`!*!zhU*{n-kJlyKm#K?{z*<;IiwCQ=Q83_aCRT{ASPjwfB~E+_l`X%W;c$lX7a5
zeQRy2*~)<OuFP?Uy}G2;X!d1ca$^3%Q^hSusA~p4We!*d?~yi4jrPP8cs59`eAnD}
zR6XV3RPMdBw!QQQt?I%iJ@#nZ#I=s^uK7QltQ;(Sdr6p({d3ZqPTzNaw+BT9k;_UV
z_SE;hW6Z4eXc_C;^*G7cy5r!pyzEKF$$pMi;SKwfB?iO3pI-my#-32U!^3Yccd4yR
zQ*G)$lU_&r8p5?bII&IVaX+o7`SZp0aD!i0rdKVC8+EVixlXA$+SuDSp?Bn$vq!{k
znb+54{SqwV7r)B55OR3a;g_Z5KXpyq36PeTF6=+M;Z%O|m(IwqIgz<<+n0=Ra=Ude
zjU;Rx&uy=(`M~^RCiH#>_dj=^^gWP7368%$PqXYX-4JzQ|HdrG{^voWL#08#zZ9%>
z_`V0p$X*{COhFL!>by7CdywTQB?>`sbE-L89*Fstzf7x(ObV|{3p9`z0%qtl%;W*W
z<`}<V2*pP<kW~as*QcAw1B9(NZ5kh%2;`uY2m)qNSqyoYgU_BwJJ@%3D@s{Lz?xJ|
zraVB{_Se(ka^%ztloCn6x>Q}ZJV4lPD*IaNq+caU2_aw|s*biiK-j(MVa=Ya2T!1s
zFalPms%yyugnen-i(iHRy9lL72pFZJEO~%%5F_+Jz2%cWl(K?=0Tr<20m5seH}}>p
z*#8oxgc2}~N~6mIgwvFn57emAI+PMlz)UJrTOJ_1_o3xQKZ~+k7?Mc95*3M>JV5xs
zQXi%0$1|}gC76KKsA^jB0O9h(jdrT;H_oAy<pdl?4P(gzglqNaCYN`Qg=ipB0@l{o
zHj@Vk*IgPk9{xDdh*G{KV2p}s%L9ZPHvEsDd1EXQrK}`iAyvqh2M9m&Z@<RxxYL4C
zq6t`os-Yzh5Ju7y5;H!}BVDCL`&&27qy5E27_JBO++e#g+yYp*L0@9{60pP#wgbb(
zK<oySgzh8gWY&XOU^pAF73da$MGACIz)68l1VnRm(5(cW%&H`v88GuAvvUNFUSxJ&
zz{`uwE(C;l!R)O6?{0!lW^Izr7#NF4Iy+z|BI!JUhlr#L1c4$*2Y*4K{bdR;Jd7P?
z13dv}g3y!=1gU8POgte2hT8*sPg2tpczTkWK_JKzYC6|>swIX`7b#&li_h|e7X(L(
zP?60Qslo+Z4^qVnSb2~tF2KcuRPg~m9#CcJYm3C4&C44I6(Lv1gU4_}2^HCRk}3wk
zKuoIG02?u>;tE{Fq>3N#6GJ+9N?1}h&X6wHxg?ckTUboNa2>vm7+xTTGx<z0oMWlx
z)!DYo^b!>eS7WPj;ecYeI$NC!Zv(?oHp+!rW4H!egA3;c?hB$8*|LD*7_JZW1!OJO
zz*<1o!Uuc-%o@Xefv*780`ns1WXpx5GX<s!ZFT?-3T+kufkK;uLGWCg;XWd+Mz)L~
zU4iP&ANJl(&m;PR>0>5va})$9?;I_fby3D}ZKgIC?i&oJGwED7k{HflGPv-+2fQ#M
zU)ef?0vOH#90kKz0!szMI0I(|!*~PlIm47hG@h9pxfDnAjK-#M;pJkuCR>vW#}1AJ
zQIu>)Le1({o$0)sGdGE%Bvuk%SURkN$WXR2;kPkd7w9Ui5_4d#uu1>`3adm2gmbF|
z-W5S7Td^daAuv>EkS(xPXpkFlQ)rMs@SkfC>^o75Y$Zdw#+lQFmIrw5L@j^|c+e~N
zXK=}IWz7I_tjpEqL3=$qUtIBdHggwoY|b_3L7zd35~<4eIpoH0BVeS^o5f(ULT}uG
zyFzaQK)_sY(i2i}8AnEXT!tI*R~dh&clqhQu6O1taq+8TkvJ%TAo1j5Ecv*ae2iNk
zpR$Jd7?2R8tTK<Ps-})&8roW#esr<WoKEv&ESRsO%XZ<rxqEmDycRR9T}_u*82;TU
pi^J77GBz<=xX9AV#?Ib>=j3bv^lTkPOqP#-0>%5ArM^mm$e&+hi>Uwr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp.D342A33945D6732C.idx b/.cache/clangd/index/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp.D342A33945D6732C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3ef24c5eb874664d3a80bbc64239ed98b85af519
GIT binary patch
literal 2866
zcmZvd3p|wR7sp@2yxwMBX@;4`)QA}(j0s62mvI|Ija%-;tA=DFg_PRe?&Q{Gx@jYQ
zh&Ey6QmKa8F3PsGDay8@NOaMPZFQmlGxPsE==0h6eCB7)^F8Nz&vTx0UIAWWu^AOX
z+@ifUCd8yH*FX@22ENHDNpa8B5yXmuARF`Rf)jkZ7fieRHLQNyuI|SNM*O^g3qCI*
zmyfWdZ@gBt3c?y|W44Uu9yoY-;J8mnrXVv`|K7+^I}<&bp6p!NLGslrgS};y^9dgV
zncacL?EcY_Pfuzu950BRX=>SdyXMB#^!Dd(hPFL_{zWwS^v0Jl;hcfQ(np<%af73A
z;THO{R^HfX)o5?*;s1SsAnV5Y!J#+L1}j;e`+GNBd35`vW^bCzwl`N(Ewb``TjhO5
zlt{6uKY7ji)s6=#*~a!e%ml@O+#~5{O#j9{ck>Tfa4tM5$jGS4QEUwuKH0&Ro?x&I
z%}Y<Xzc|pBzrJ7L`Y)Ru_sS@Kd+~EcU|XC-+;Z!RLs?ea&1$((9I5z+RhMf^Lxy|{
z8<thOm)qF?QN<+Z8GJ;-+&l;LmeT#_HXhz)w>XOIueC<w+*F~)c5)JVDyjXWTJ3)M
zmGIa)t;B(~ZZmg0lrWtx+}g4JN2-5{Ztmd>>me6y&(g@Ljr+=_(NX1Tt@AgveeN*y
zW=FGUw#rU@$-G<5STMyk$yJ!y^(wE~YFbOnW^acxMrKz7DveFO(lr<CV#RMvSog<_
ztegNjcGj^jVF^30Jhx=$hl|Xek1G_h3`t$kg5kd49xn>lI5YhT$<W?y9i{ZNgF8QH
zc(~T{2N|ceFe1MuDvaIgv*_iR{Cp?F_NS7wtFN_nEzYVurMTe`c&j-ys4sNbuba91
z`0mnFl=mqovF76%Pf>^2tGvj1acFgr<bhm1@A{!6vAk&(P3si1x_ZcZr%3wR{xZ)u
zyDR$bi3EA}D$~3H#hD|0(HU2crO@joOGa93npw6#*5&#3AI+Ng*Wu{G(wsxh?~PrS
zJ}uX1IN4wn?){x!rH}PWo3`}yF9&SOj^^(1GwcY`S1jt@vhMhUlstQ9$5{nAebyz-
z5hYopN6$)A|B|N9Fc9Yz7V4fZEA;TM@EGKk`j55${_Jb&mrlze$5HXmMF+d68D>m=
z%S$p|pBC%q_rlWiazai1C({jwSAS^gi%*dsY3a$07v8&>-znPqGVmWO?WEQ)r`mSg
zv@LG4H&k8M)5_?4Z?T8xwoB-gv;X(mzFKvSzjADnK0M^!TeH?!_|$$zzqj<y&&S%s
z@8oKW67oBW+BzEQ(`pKio&I&6k?w#Sp13Le-Bxl0eow2u;@BJ7!bsnlA6O1C0@F^T
z@OSa%&5wz+F~f=zb+*N&qmSM8ygk3SyQiqV$=>IyICiaFQR3ILIq9A0`GTS0p!1bg
zDdD~)Yh;!eD}OUCuqn^Jn0R$hOq2hEx0VlXiP}x%pBpkSy}!>|$v$~|Ol=mgf|Z<>
zutJ-JAezL@(|)O8c~nw3f)G>CDQuO1cY*4o8_^#WEUDT~B`pJ36V;@t1Y-W@Rn?ZV
zrWI6@FTfaz(NzNR-02Tl-2+KsI3fc$P$Lji3B;b1ZTnvMU*Ah5`2$QvsZ5nXylUCz
zN5#`~d#EHSz!a21QwhZ1SA3UWz+B!+B`pD%i*k)r0`b0sT|3R6p7GE^0s&?-+4?Gh
zxF~DXX{i%u6O|MKFplChl|Wpg*h6dRv0Y3hMF6aW>d;jJah1kD33qI7y`z%C0H&cd
zno1x(=W)%jc*9<y77_sPbmnw@l|X!7REExeV)8SU6a+AeqSI6Y@v}q1@Sg1wS1QQ|
zU|m#~t`dj`D(!ji{wxWml0pGypbWZ7ARc-BbiZY96qia`3a~b+t*H`-$6iQN9Tpm2
zrILaHrlWM4N+3p(Vpb*-;NO9I(&mN2YrW1SoWMAk5xi)epl~Qg@Md81kSO0Y4NmY4
z`19cXz!m~d%Km{*!U;3XOlREbT{00`78s$16AU3E0G^Sh_phg-d6)*4S?;Vwu&yv0
zu%)u+Cbs0$`TB56*nDb~|7<!CtL(UmSPRU81!KLBHtxN<<Bb+Ly$NnYgUjGxIDg!d
z+Q}1{{G5FtIhv5>4tpZxBKL1S!Ry^>8bq)`umwybvgf52g0H`%JPMcP3Xc*abx~|P
zTe?CWCr}Q`gROxRnjB3YEFDhZ9GnMR7FJ)K^36Qp1eHTIg)@Z{=9u|-c~d@#YEN>1
zLBYd$T%Hd+PwFD~JLaMD{lGHYne7ipCe@Eic6!_`1k1WYT}RlDFbWh@Ic>0V*{J+f
z2-a{1c!y|&*unF{%|M-%!w0=_!UnUM2DgGo0#3>ioN%(lEXR|Fvk07&gE-+d7n`dK
zcPig_iR_n9BnDY%Pt|6^DS(#(YNedgiC8*=ZUAQiCv+G(2JkWPX29c=lRFV&jaiQ;
z4(|y#Dd%^>X%04r4s+`2-YVKpt9qg=?-F(dtO%?p$VoZe6S0gX3<7pOP6YFVg|L`#
z6@(}Uej>yQvl=e}2$|GxumTw6oB78HXUusD%n&C;m}rvag1Jnx#F%)J<%xMtvLu*f
zlI4r}j$5#*pft+Y%S5gMOfY_@;5`$+T#1sCWXS^M50$0-fgq9p{{vK=U^O{?s_yVm
zs7Ef3LZZe2B1?#lj|HZ8%wly7GOCG9q2gN8w5LjF?yj?G6bXI0jxI}YfpDRNqm#49
zl3{Mon{PVnzgC$uW^xP+xkkn&bIdFRbFFM_?eyod`POa>rr2u*$-~dZQ*9CQKj8e}
AkpKVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp.E5EA635D0BE03F7D.idx b/.cache/clangd/index/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp.E5EA635D0BE03F7D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d3bbf74155243e18db0965768f4dba63d537fdc7
GIT binary patch
literal 2788
zcmZvc3p7;w9>>Rwwryt9>@j9+6%!`PTc$D$%6OadJ|x>KNk!d3Ii*|Zbe&3+p%M{}
z%AqK_wWO5eNOe@yk;m;JDsiI}rBmJi%w2!9)|s{DyY}z@`TzF+|9fp;cNdplI)&o2
z);&2cIx#|-LZRT`pO7ex?WR#EEWnZr>H>S>dzLXDeq&qbEI2OVj;7~4I#VNwv|jkm
z=)lcG;-h*2WXfRYHlx;m^dUi&saGEL&utsMQ$TbRQAE^5le$^0jX!T+lY89UFuUgX
z(5|t;Fa583tgq!4l5@4po(#wAPp4gI?s`dUYW|e^EH2~I7eA}sCtm0N+1uXN^4vNs
zW!KSxoaU4ikG*!^gkk;rH|Lz{YMQ7|k-Fwh>^|4oDYAH37nj|sV~||eR~1@%`}*$3
z+XpUJOE!ww+GBiSdOp`Dwlx<Qdf46gmb|Kpt`c9LHM%9QCgxtg{a|AHf_wB#e}0E|
z@p$Y|LDJxkpW^3ih_0yZQg-bV9r;s`T_E#G{qX_6Dl>Z4e`Tr%+D^W`mH*hpe&?jL
zl~vO=*mz^LEIf@xTdNlO=~h=ehd;JaQ~ZZP9`gg!C348ykrz~Q?<p_4NY`=B8MHF#
zV$C!D-1v+rZ?(h!nr*f(yzt3#*}UkE(@mnWlf+rChPhvP!N+bk23?5T^u98y-nG{1
zruLSymd76%h0(@a=w+-u^zO7@BM(RM9vF?B8Ljj`&G)PeEY`Hqc`A#nj_BzNUn(PV
z4e-cwcP>hz!z--v=-oZ5v!&O@PvjQAZ+m*LG{nb%|FcX_J6>haD{Nq4u+uwcUD>O_
z*sRHJVS7<%!G-W(?T2k16JLG)Jt59q7dzZtEm)RwekgXtVmAN0ULIe_5&2)Dl^B+s
zDZGC3YG3_^t2TIbtRTs2^BR4Vu0>_B?#|lpJOrhMEk$@{G3R{6eycJ^$&m1h{;K^?
z{RfJx{4&-X9ecCm+>zmN_2mx+7tjxtk6w<ES#J1CYFX4C!-`EdZFP_k#<iNBX<M%Z
zD@D5Rbg!-5;ICD*_RhLWH9_;K@Uml>2ffeq9(5gQ{}@(~kgx9dwqxvmYle=WZ+rp!
zvu)Jz;Y&xbrH%)(i;DW)AKtYx@3i_xbQpbm{bF>N%GZ(XzjAl#Cy_>@m!0N~aSHdE
zHt*VLXZRm8*OIu!D?@)P)KvStF>#6T*-uti0;`x^;*2@ysLZ-3qd2herPJ|HpJ8tQ
zq<Y@6()**fd7mu(-DFXIT3y4FHRZo|UW#!(sT0EPQadYa42jVfmA7u+wK(MV<+n<w
zQ!j*U%74?_d*x%>AO2C=vpqh{-^;bFnrKWrvFB9&jliVwqoUI7HZDV7Y7W!KX1~-e
z+13%XxS}Vkq5o3(<T4kt{LY}3S39kos_`$EbZ^Ie3~OvYe5xpJNb0%mgScXev`fJK
zA@Oc?+pkwUga_v}?LDzMx174~YQ~L;hHiJ)oIF9X*Z7>`qLUd3KgLClDp4rx(7}K}
zx9T(oHH1PTRS*?iA>bRJ_?doQf8H#rC6Ym11u%-BEQLU9!F{0h{EajiqeKGitL%#^
z1Y%cuYW9H7Z&?hg55RPU&QS=&>(}h)EahhQGpG`PRS{KOArPljZ73+@L_B0r{Q%}6
zJOhP5oPKD9h4o3-M`{#*fC&zvqY#L5YBVBror!m96kmWj9FC4cATHVW#XiJ--gXA{
z2Y?v}16K&dzcd`gul5VP7}PZYvk(?bArMz9e~7yyxH-b01_R7Qn7Be9zVBFpEa)*j
z&7cMVj35Y8ArL>ywOQSt=C*=CT?w!TqQO!K#Ba_D^+&pY31m=%0A?d>mO>!@IQ00C
z*^6*J26Z*S>WI3ELLmM+AlYhX%fHT`1_F#DIIa+grP1pWgy1tzlkTv!X?o#6#YmJu
zd9cnHsY<BwU}01k*Y<3_C|In7k!%~b54;%c87lp|UcpEs)JOwXO^cXtZMgLw2fkb}
z*Atcl9t3>l-2|_Nk;W`z9q0?&N{Rlr?F4+~y)^B+2wkKFeLap{%W6v>Rs(_YF+L8L
z!2y`@lqFeucr(ZJKsLA>(}pRA<@CR6|A5yeOIK1R19$-@@ET<QOLq#ER4)%jkGlef
z5-nNYkXln7NyA8lKy+czF;az4(S<p}NCv^sh1~`dOQV0+=@>~T=mv0FFmfTfa4L_9
zla94zdgl;u_<X(jD`7laJ;is7gO+%MWx|2*fy1!XTW^;AqvtkYS;Iy{1Un6SfrQHE
z2xbm?frQHE2QFizC2GlpJ}_LcNj^o>n=DX^sZB6Iut`2k)0<3Clc`YPP=QVIiJIPI
zhMKYAah1vSv%KR<TtEcs>gpWW?XW~3bMkSU_Qf@E9_)FHWNET^@EBMSaGZPur(FzD
zLl!))`@}<St!M48!Eu4gfdW|ar_Xmg9>S}8<mvSzLSdp{KEW;IV>|sywlAB6<AIR@
zb^&hiO>h;s$VYkF#T+%C3L3b~=r<TMO8;)MF;awzRN=)j(jK*+VL6}<Gb}MGo?$to
z&ND0*)MbVxK_ycb%p=H{{KGMQS5wq<s*2L+O%ZaEAdN~emp_Ed<^R7>_&}kNE$a8j
zn>v(IXrbX@l&~p4iHcjhDF&Dx(Oxv=S%?a%%D~i^v(?>j@rngF-HpZ7U~3V}Y;5gB
z_709_n#Mx?B}Te`x5}9_cOGxPo&n!*;UW`r3(LiVr8?HyrdCdx92fTps<ZcUS0zu%
F{{V$D(f$Ac

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv2d_fwd_xdl_perchannel_quantization_int8.cpp.DA805D9EE5DF962F.idx b/.cache/clangd/index/conv2d_fwd_xdl_perchannel_quantization_int8.cpp.DA805D9EE5DF962F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d9cfe305dd51ff28302055445d0eb853f57aff51
GIT binary patch
literal 2770
zcmZvc2~<;88pj_AJRV6{UJ}A)7?!ZFY7xS|B`hHXAy6XlKp`SRv<=G<#HncA0TC5+
z1O_8TYh6HMt)f_^P_+zd0hOgLgWxDA=!kovH<@z*&zYQ)|H-|-@80{}@BY8UM)3JJ
zC<qdg5wT7t&0DUCAP5e=xq0%eJ~D#5)j^PTHD}`o*9>|Y^vy7pm7Rx#M%OkE3s&ds
zHU4_1FC|^h!AE%`9R(kz*pc#@pHiIZ7H;%Rz07r^{FSBB(*84ImL~bvvEPmlZQz!a
z3Oi?5J%hLV&Sib=;o}z&n|!e{bJr2wmX7`>)*bIB!(R?`e0pQ)u*R<Ehl#t&cWpJS
zT_0N|3zv1di6WLVQ=jZw@ly8My|wkS1K+dZJ!3N8_r=CSBB`wKdC<n{Ll>Kq^0bU&
zM8t=hq(2h%owjtC)&}iAtq@yPZ~ei3lKp1loLD~cvme`cs8%|T?^C!vzq5*T@)jmp
zq~#>`-_b3po;J%99j<DOyz1)Zct1U>c54QwjMjC;F-zv0w1*>1Q|gyh-^p7&{j{f=
zTDAQwKg{$_J;LBjSZ4p{no(y}D6?y=^q~hMF<5-btfVYNuOzJ`xny%-L7TR$fz#9}
zk+rw@4B@|8cvNoDcYT#s`yhnL<{cSn?kc|Z=brDY+t+fO7$(Nqb;`23{J-mOuS5@`
zVN+iA-zZvg$kCfBYGR_TesXM-zy4PFx{iJ`az|v-zHFVy%13y#C2Ce@IzF}4>c~D$
z@c5R2B@Vh#37xUQJ}Knj1$XYZ&>IGv78K=~z2a!?sSIs@USd#t*h{NNY_6zsEGil}
z-1cMnvzy)9oD-|~XZwCXe|hA&h*!2!Hh!nc#cQu(Y}Ld<eVa4!&VoGWqGMm_hl)bu
zd^Y1Lop1Y#&sx{7Vdh629kM#nZ`+U+5%Fv|l52f^c=z_Laa2iv(bJIn#(bCil}CAB
zosr+_z0UiB8TB!sSD-u-@-W|fM{<Kh@wl<H)bw(4*U{HP-m|rWGorZ8gd49sg4W;9
z(JB^hh*0k9U6fpx(Y5k~j>mhuOz~nf_Tk_Qzxftbcg_?a=*c_a7{JzxM)MBFj=08k
zOn2XNUH?G1)yIN=Kwq>|+;>x9e@^iZ@3EiBc{2I0@|g$H>EO$Z<Etvn4c#+7txgJ>
zx?nuncwPRiZb)*fKXr4<m58=V+tbqa!X^JMCTW<*o(%K7Cd@9pn=x#bI9%HJ^MsD!
zTJHW=(+O4NLuJ(-Q_8EEk?Hd0oQh1wP<Z*>ZvF+ky>AK@4emVuQsbCsOLp&tKRi8e
z{+PQs{($#C73X(7l8DC0P4!vd9J?-y&wP9D)iU|ID`^M)6vb7mHZ-|7_{=<hzpeMn
z>})juzs=h&p4{PIzGK&@&70JyA`!Xh$^${;H=lCbx_bt5QkFkemh<|4Nmk0L1l~te
zjpFa-x?HKMYxG&xShz91n){;llx&QbzCWwN_M=wwqbesPw@|iXQUgJ>Q#-$pDUckX
zk|YR1U?^s)7SR9d4>RRyd_s8BCMqcjU~N<zR|~}MM!m+vFXR#$vI5{(%~(t=5QkAp
zDn^$6vYSeZ0hoePSZaZIRr30~M~uoIQ%ORAbx<8#Ef5!<TwPPkTHZ$`#R6=BT5!|?
zv0~?+{z?AqA}T2XU@A(*)dKN%m&)<B#~x8sQZm4FlulO*#7&x0vM!Hn6I4<nzy_!R
zt`>+d-jHMvHn)%KB1-{gu~_D6f%tadanyFu?iiIM1{g(A1GPZ>JAM86QKVo%7l{Md
zgk@r`7Kk71=PrA^DLj};iUyd0GU#f7_{Ax2>xqHy5~!pv0cN61x>_K9H+FxI>u8!4
zm9z|CJycIyEf7zS2n&3DY_3pA@c`o}j;jS?xpZZ&Gx&=q%h&sG+eZU5XoL=1hXbDi
z+JQv*%my^ViSA?z9|W2L1gRVWx26&Hm^}jq8JoNfyL96<4h}F1H4?yrLdR%OK68q<
zpq~AYz(&4?CLy*EJHsewAaIh(J#!~HVvfe}q{xGpcXw|d*Co*i8=4Icr$Np%r`eIv
zvvMuVrL+R1W56{Ch0jkseiDJu@>^9|S{VhyETzJrOC`;XE66kgWuw;cxoLzpTiY65
zpGHvGRBLEMSV1!7GZWJY3Y%gN??EFRFo)R^<$VZLHdviT!QqxxmeDZr0zqh(Q{s{s
zFl`cGvJ{%BAjV3||K2bcOf$F)KWHMj3KUe;BCtd-El@dCYrtt5;fy)YCIEW`1gRQi
zF31&gWx|b5tZUVZktz7#Ts=KK7PKF%Ah@rpd*;^ShPVYZA&sCL(k<XI=e8{#%i5sZ
z1a@L_nbFYau;?HQRin-A<c_%;z%1YsfFM=7%>~(Ec61m7O9g^dO*a>`5L-CA)5GD7
zfqU?#K~;I<Oj6()SWj>tRY%UPWr~<dFjX2Mb`f)-JK-!?qUzGQC2pA8Z27>Fd4_%~
z;Cr?&K<dHa@tX0A;7+ip80E8GrV)OapANhojqu0(=ZylefO(@(EOg$8hw<i(_!xiQ
zNQen%NARIQ#Z-S4b60S|TxM+|m*y;2kz9FtuAAydno9onNr68QBz5*DnIKr`)s!XQ
zUL~ibr6TFFjGRmWBc)MfO)XR#)1lII4fOTG@z7vfoDxnqVla(OytqETe*OW0u7*zD
x)=TXF_Wx#CW((LBmR1}aI|oN+H}{2$Jv`07Fm+iJV#wl0EGO|qL17vK<Sz)}xSId~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv2d_fwd_xdl_perlayer_quantization_int8.cpp.7D138D1D290C5C53.idx b/.cache/clangd/index/conv2d_fwd_xdl_perlayer_quantization_int8.cpp.7D138D1D290C5C53.idx
new file mode 100755
index 0000000000000000000000000000000000000000..002f9434925a98b9900fdd4beb6b901ac771a468
GIT binary patch
literal 2684
zcmZuz2~bm47k(t+<&iv=1c(U;qyd6~uvEhmwy+yk0ZSD0u?iG~iUAa>e=9m7vZ*Kq
z74R2CYXKKptE{55#ab#LVxTC7vIs?Cv?{iu{4bfgEHnR`$v5xZ@0@$jJ?A@b66)*Y
zlR<$X&uzY0sqq;xBnX0N_%A&}n$$&vpaC@q$|}DeHu%MmgHHc~j^G9NgfIE-!BJD&
z3r7mM-iE5DTZ)nNkjAVj*<5IG=}nacI@{WAkH+kHqkajkm=>Gw5za-@Tbg+JFG^j1
zyPNfThwC2Z@K<NXZuH&mXQh3J6=CHDWmZ<MXH_<pCbs*p`u^}ml5>IC4be>6%h8?1
z)@|Jv|4I9_nZ-W#EGjX7-^<i@f{L$ibswI~ZSmP>P#6*GV^NzesZ^U7KmBMI(JC?L
zQblQ{w08USV5oZ=A<?Zj{-j@U(#<@=aLmG2{-&ixjZb<z<~hdQ0lfL2`}6|Zb(x<~
zs|M&NH*9%U=fAK$tvT3t<AvYt9QecOF%=2FiuG8z<S_Xkn;U+r5M;}RHjxp1cjo&p
zTFXq-win#j3OyLj6%U6zl9!nM!nncCeR{Xug7e?#kjyVg4>cU9VyZk($mqGh`+Q*;
zDb%Sp)#2tr%Vac^81YWd(R}oe%c0&FK}6;35zVINwQr>FJhvrrVtzSY@ni3aN3+L6
z9zeZaF)l@N>$Gp`PMru47x6nS)8hliCrL?DF5F5T`bjoVR?It+ljxkoOd06!8G(&E
zPMWxnk$%kdBOb50R@TF)J*CxoeeQg8=nECH;bT1)-|Q-rdue+)UWqnA?t*W~H8pt}
zu;sMH^wp1Je-)GbAAD=KX88U3;Yf>Qfkm8sPD<H2UX$c)<==HsLxHzSkyp3hfOc<n
z&%^s&xpt%PQ`#J8ZZe9~1y%cY$u@%rKb$wo+*wmTa=)V+$%&(e4`xkyv|q?nf12Cr
zoz{?1ceyzKbDOe>h0WHz&kbxMYtHm|pT6@<65@09W&Y182X0*wKiPidZPLBFq2X;F
z$yd(sSy3@jReaOysQrh6fr5wUJQgO=?FrSM67g=&>fK56SFCtJhj~--2HEULTwY1D
zukL#%<DuvOv1X;P%n4r(c5e(5<+lvv&L;*`&)1&nuy^{nqx3cDDWyav*BLq|>+=*!
zrkgsa9jgbN9)@53HL5i~@~l<Oxsp0zcTV}y@aPSf9(Wp#HC-#04Ng|pN9V1R&q-e`
zd~G%{pGw(lr|%*e_?@C5^_i4kgDVSTEQODS`Y#U!Om<{7(bkO;J1%$4xJ^$y?pmFa
zm{7%j<LtBc=*a%jf5rT?`)AEFwnzr!-lDJ4P+4c3DtE-lPjo&vVoP7`Uq@TVEGu5^
zblfVq_;}jUQRFi3b!(H~T;|1n-hG|2rpJF`^*Ieip9!9LnGW~W|J{&U?-OYpG-}M>
z_}lmKXKZJNCQ_0Qoj8%i4ScIVEyz2Qz9%)dR|SG}u&FSK`04>FAp(LhGEAl^0iSy1
zW3%cZb96=;jzTdw9Y_j9lmO`Ul<Z^UAwM0Z5`uAB4OU|)0np(#jp7!|wgh!31gCWw
zx`s*s^ygdm%Boitj!_ALIE}yvO$mVBSD##7!HDUn5<+oWAJ#Wf0-#UjoIACzI;>wE
z+KAIk2GdXpfUe5>;JU@t;7ck&g3}b3LQ?{uYg&tFZDY0pRKgaVrolA25&(Ub^d_~-
z_U<f|uo<T{VNIG60Nv~M9n2jv`<6=h45wij)=~nXM@obdV+X_@R6-C=(_uPY34ort
z>}WjuM~#F^*o4#Cur^%@fPObUeB5dxnnNXo<Fq=gPF4b-KaL0Pa&|JkO(lfkv=*#I
zQv#qNX?#NZNBnC|l<svBc1(y=Pz+%qY*1{_Z~|pTqfyKlG1dkpXR_wjx88eA!($m<
z41dr^pavDn3KiH^Ki|_?x6Kas4k3lu057Nc##on@3Ld_!Ez}MKw+5QQ6Ho{S<c(sc
zh$#~!5NKb#gJD^443E_mYI=cofJE?Eg_@RPEf5PW5F0u6%ojp4n-$4(7#<)dKG3~2
z``U$AB8tH**a)-`#mFqO5vU4^QCJispdC;%B4vdzQA~}cW&-YkVkU^mV&yXab^GBw
z=L(D;zM8W-2;{WO->b`Fvt2O0%oH(0fOY|m7gV8daG!i~`?=gZ()|P!W20;u@Io;z
z%3ZX;9(XMj;s>!P#z**zwE#E4ofIRm<Yb1JE#?en0e4ai!jcmY;pu?G>#~|vgHx+~
z@EfRWs55{_L1##m6$64|TDn?nAXOBj>C)KX7!;%H(%E1}Kp*hK6?3$7I0xa-!QoGD
z*Y32*>B&_gVZrQRK2Vr^bidni+SMUNRW~v*P!X^--ZsVTEp4eCq#Xfdh+@Ic!D6rz
zSj9sW^SBgZj+if&01sJqIH&+ZSuvg{=72b;fy7ZvhzOTij)>zjON5A)S#F5iGRqxt
zUuKCB@uCH)idRfA{Y$y7LDnqxR2rXxf3flZZqm4PYsE{H0+bICLtOSd)s^|{5w}J~
zLR-N)1Vt{r#-(mcNyJUR_y8hF6(%EUR8(C{Q$tMi^5D`aV)`m=9VW{`=;Z9;Dsr>Z
zwQw}HTVwS9RvG%M4cMP>Oiayq=6q|LPuJPn8w&I+*LvzQe0-w`-W%N4tN26z1F`m}
A?*IS*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv2d_quantization_common.hpp.2C2B760A7CE8525E.idx b/.cache/clangd/index/conv2d_quantization_common.hpp.2C2B760A7CE8525E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..611e56f23c185cf5b804578cbc210ee144a87fd4
GIT binary patch
literal 3132
zcmYjT3sh5A7JWS6QSuTYKtc$FBp4w4@-*R}01omI&;cSiSSpB$ABtx9h$v9MU0N+F
zI6o^bh`Q7nL`I5Kg(^DM@jK#79bG8Y(S_O;DKm6L#c|HNo#PwUV&}25?>Xn*d(M7X
z(^68x#S}GjNm}*F;>sciilU_OTTxlQ;<g<{u?$63H}&RZ?@&&6y^%TgU{Ext%XYOa
zujjAsM}2aazw-EMN#*9dx;c^0?%sMcn6zf~f`eh>>s(Llu`M+Y)ta~58=QQytae|q
zm91S`Z2NZoV8cMhqn)*`o9e_R=TbzXns~#{5(2(+8*OTLs7sWc_KeJv)PFo!AJV`6
zSwipoD*3U8aqZe8jTVoS_d1I^M3*OCx10><JYJYmwyZ2-*5VGYy$iELL(7fzUXR<J
zUM}f3cYZQ-^>>bk9Fm83H3uHJ5I;5Zo7cuQ0j1@QO6~nA=KEo@tRJ@NCI<I?mVBw=
zrZSh!?nq6ot}->byog-oUt;qy^?N!iRN9T^%ViVOW^Y%>+$@ppbfDGNaA@*_ZLbGv
zl(`1uVcqqaQDIki&R=1zb&j3CW#97W|G75oT6OGBU%xWp#BW5V^cAZ_2TwWK6+Al<
zA&X^{e;drmJQdV7)?xkg$tSHlb4uDb&A8HC)_eU}-NqfKMK|ju<NCzemac2JFMYQ6
z?2{4MtkTw&<Aq)SzPL7G|77!zSw$O9mwwsS==&G#8|(bbzqaqTnU=R0+;7h|e>kyR
z<^RP$>YO77$A*5DxxJQcx%x6wo$^b~ozrghcdI({4j;)a>5V$Pb=b4-pK+-bYgaDn
zwWlci@K4Y6{Bh(*3FYZeaZ#!$JpsxYIzX?p<LpE)TLw30M9gJrB?!D2FDU^cpePVX
z|K6P7`raG0gknJuBahJ$Ac6ot5V&oc*Iz^b;V4V1K;X~#dk_#J$j)0s19u;8V(E84
z5X1y22>hSm>p%79Kc&{Ov=azSjLD0DAW)w_JF)A=Krl;tgFwkBMFa%Fw9<*z?n9?Q
z;0OXeqjx933$$bx=AMTkB@SWFC6ovRaq>7F0TO}iBJPD_-e`?m2TQwvz?bor5)e`(
zyc!!hGQ9RUOZ$SropE;}KsnPeMEpU`>dKk7)$gP063Q6_X1UpqfRG~1R=uI|<?#nB
zZ4Uw?WAq@v3p6Cc&pp>P9oRRJ$I>1kP%~;50zwUot1{N@@2(ACX&DF%jKP%vFPKIm
ztXz||IR5dMF)S?sfrinz5D+4)%D?&aLh|iB63Pt((eh|t0z?ov6?Refj@h<s-K_zZ
zR)fHs@fHyfB2>OqHMB)mRI;=J1Tsd(5)cI2uBt{?NtVEMPXU3Pk+TE@K|@jDc&DlP
zB1?OLz?1Q02?&DE9WV6L<$ZfungKyD6YN7k5PZ>kQ$qbPyN{)PK)^AajDR4xU9&j+
z%)zxHmUaL^I1{cSAP64qPE!4=u4*PryMjQ)sHPAQ1mhiJC-OJ;2eGsm1X4ySB_If1
zbY^~V<)x^IrQJcGWwh=D1i|F?{*hE$_-dAxgFwgVTnO-j>EyEBT>8Fw{Zm^4OG`nZ
zU=+>-ga}l5@zM%+@MoN2I}w@Q`tz?^+8&rS_T@F^m<z)Rw3t(C)PdMOVos~k24a0q
z?Aa~0lMyBm>0|X)oD?w^lo0d_98GiqM-iEROd#fTS-O1O2n3j`FAxG4tcC^Hkz&r6
zYg~vOj-CPa@`i!o&?$iCZ31bJ(Tc*n-{&lV3XNuC4h9i(rWvOBvj2GlRtjr*2f?-Y
zHPF0oAg%Cvp4xk5TxuugVv}Msupaag*d}ixs1|enS^ha#9h^*F(0iCq%yCw39yT1*
z^scD@lj2c8a^7+<9C{7VywSX+Cr>Sk>-jVYwy87gGKEw1H~VK|a<m~>%X<*6Me_lg
zcOInCa)9Rj25CG&pm~=;I(_fiT{kv=FL4xeDy2$|BOvB{mA-0B46US8tA$q5C^c$y
z6=*D_R*j7{C#umV5BqpGoUvE5*Gz07Tnfo~2gBsKvDsz4wr7t(tq4~{Vec0_dYndy
zSD&$mi}H(2#ETMh7C%cOZcog``^6{Xv0}?X0^R{J0Tu}~?|Vq2;Q`IN9nxrX_CMJh
z(qb+oIV27JN6bZ}M5Lpapz%Q@-t}-vRYuaaki5t&*jRKzbT;Oz%JOrJzw<B#Xl1n0
zf}Mxv2tjyf#2`DI&h6iJzI+3yr3$GE&1Xlt{g#nkZ=!*Yl12HT4WNNS5Z*m8$R1I_
zyJ4)g5Y!?1kR;s2C(}PaTD5t42hg%`SqyHY>3yjt|6=|epuMBLEqEZUsy^3y_bMlW
zwj^1y@d^$Nj4m=N9s~d_50}T_HrlGsIA*PEO##|{in|=!6jMWiynADT-CIrHt*Dc9
zzyug9hHSh@7%>Z%jU$Giu$DJ=Tzl*BU{A;Wu5`#0pA?^m&G7Zsqe~7%p7{#s2z^8{
z-W7gA5Z?ST$WU+EUCFl#KLWL9xMvKW<2MgC%xI8y3;-QHH+lhHAr1%x;a?CK1TP0@
z{<VNKjs?*C%K&K{bD;TG0Mgk14kF_DNW)}mR8uiAOs2+nsxX<p{=Rzr3*HT-MevUY
zRZ+gUVlMv>%)?JVA^1R1p~7R(+eexk^6u&yZ~x)9Oi*Z;Ataazd;1+_5FY;vCKwQX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_bf16.cpp.8FB5CAA0FDE2EFE8.idx b/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_bf16.cpp.8FB5CAA0FDE2EFE8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cea088cc6dcecc1ff8228f94a1b33fc5d39c9d1a
GIT binary patch
literal 2030
zcmZvb2~ZPf6vsCamL$7@1j3mF6$Jw*;=v&rBZwG|5V@B?H6qj?BBuh1T8*Njpk|~h
z#ei5Z?kG@{!y=_9w}q<JqGk{k#VSGZKtng_d^*l_XZD}{zTew@?|bi`KyNRvCwdUH
zJlZ=wJ|=ao4g^7T{7y+tj@>;4g6@(aD7~OH<o+h63vFPM>=oBqk<aLrRvhTPYZ9op
zBeGL=S>zF!+;Pz1^GCEs7Gv)&wIF|tO6pd27F<a@MTmxKOR86ppLud4L+am!ooG%~
zt)ExV%;9xqk2dh?pa+drPdg4GV?%H6BjUsqLIbL(+s6NOAk|r|+GVgtop0RH^KLAv
zIue~g8Le%tm#`fc+|F0LhbCttyqZ>K7WlbE9lcX~x?!Ait5opd-kd8$>hs31rKYv9
z<!ymM3Y+!(r!m2zC+#G*b>sEPw{lUQ{(zdLRJ%!6_UBKkUS|0{GBUqt*Ne%<Z>Dx%
zeN#XE7lWGlrO_k12bJ8=+fFA6js*~w3}^TJ?s4<S`GohD=9Oi`;?XBv6KO~GunQb$
z67OQN56d>rb!fr*4jb>+5yqZWncbm}W?q3W_+qh0*89w{s*ophPBtpqT&qypvR-2A
zR#6rE3>>n>r+5L8?6AHIzRXNXu5F*gz68$-Kb?iQN0>}*YEsXk(5xc&wlP=f<dMr&
z?&nw8HJ6Ev2JFryv2>l<w{2PyoL#-AKW?u`F{s;P4Kv5*u#;Nyl6+;;(&78gHH7|l
zY?EYP4QJ@X2sik`8*<{~oAItP*=!#Lnr7vZn)TvX@9T6-lrOpVP|;MI`_w?nB$@mX
z#4Rd#_fymMgbTOIat18!7B9GX)qniJ7)C#)D7oSOrmonC&P4<-4tZ^1^EQvQ<}4GG
zW~_5d^sUPh@GGsIhOw~c<`HsrZnRG2k&u0yHpBGTfnB`)PnRcd7^*4{+Apa)(ef%|
zwMx9$BjCIr%c}8un|}b@@cz%_d4tzt`E8Zf6X(O=v$auO{oXFeIJX;)&-N|#+N%tK
zou$tGtU~E%=f>1dpP{UbHxAs%W<hslOPNY7cvaFiHAw!qqhRD6y5+AIP0`iMLp?t{
zia&a8L-a1Is%<UO;8w|}!WK10EblG8*B&)%vr>11!-=_hE5hIEac|Xzch}1n+grRZ
zPaCLj_l;4BYCNLP*ut5v4SSE$X3bz0DsRrIi0$rpCG^SODZD{dD}I)D$z_7rGdkN7
zomHj@*X3z_0VRoRGnYNMym77kABRrAAkl*f=la7+^3IuWKV>!rD^xtyr=60Im7SfE
z_U+>ATxI`oUPNE%&K=37T}YAZ>XFgO?n5rNajQ;_*wdd`jizM8M}-g|i1FQUNGQK*
zH$n`BAdCc)=vu-5tJ+Hvx$Vvw)11c$(I3Y!4AZp&aDuzNujBjemk^==#}t@C(+a?e
zQnQS>YPAR<2I80o(`>W?@DE!oIluLslq19d9P7fm`dR^ae~)bqpOe>&LXkK&*D)t)
z1>oZB&r2hgvbQ3{5FC?XGF>YGpKZ*gH$8UpLx^EG)`#^CwE}RLN37`$n4E?Xg*ZlF
zgr*gMZ!If<ZSK#yfDnUltOx7Sv;wetNU&weV(Vsv7>r{oOr>iDU?@4}`xFo2*X|N1
z-*naI<h8@y(h({V#VBSJHs~UXA!Z00RJJC)AVJ*NFi1c#T{GROpjludM9MdlKry<I
zE&>e)U&MnnLjiT6n4!=x90Y-1yg1EhK+xur?(wbF3s2#B&F#&(U<x2Vnexq$P)v`Z
z#{%s_F%pBs0-i!KGK0(lbHdMRNwEQy;VTZs@#_k)L~Ovx@jh@6$FD1zkQN4t1?AvH
zYa9WJ&UGteo7C0y;YHJIY5pJ@<i?{l1_9B)8+dy(#sJ$WW-K(00PlkQc#y^(UxSQ<
zMqhde4vr_#m;?mfe(898LVCItA24+h)g9Oc#q<~HyMydt7f+yZ5O^2G7(&LENhZf^
zT&p1|$?H;PXdc2f^4}!_56~2>+{dR(l8~cA3>QJ`;-fdN$C*!z-xM8PSdT<TP%6zp
zpHKJnu%T1<hNeagGjrEPi<fYh@|HOm+q=zh=1l#+T@xmYJ#D(xOzT;;cC+U?&RgK*
ZV!6;_&iv)ZCSKkV#1*UDR}y@oe*r*-&b9ym

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp16.cpp.A862E22D755445CE.idx b/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp16.cpp.A862E22D755445CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..234c2aa9309c32376462e4c05c113e9f79aed6c0
GIT binary patch
literal 2030
zcmZvb3pA8z7{@;!J~Q)uGtFe&nUXcpAhj;ZxJ+ft2$SoK!OATZB}@|CuyjGK*hxvJ
z95$lT%1+TmQc}5eBW=`Ds<kC4q|AOZd)_*ov)?)MpYMHt&pYq){GVrh`CRS`4G3Bu
z#owF|of57FK@bi9B&Q_BWvW6@F9Cu!m$n8xUT<JeeIh4v*R`HHMDNNvwY#fV+gD>-
zNM2izrLood=Wfz%TXXt~qgk9Red4ExjUCT!m!3~Nq7nr)Se$bszFc7&BAuQP9&An-
zi&@lYkneJ5&$}iU8PtFA%<!$f==<j%9;2%hlU16qI$1jJ%99kknX#R-!e$O>-Rhip
zA5kBIO=I-d8?9$W+w`4ozjYbVE)V8ja?dFB6i56bY^ZD+GwVK?(BErx9wxue5<2TN
ztgpFITp6*vV}FOxC$5BLl2SBIlJ}&qJ(u*7GEA80(PJNf_*uH4Ckj5=#tX>2t`&FG
zbGT;Z<=~{;Rp!5YUw2-@=#PvG&A-Xl)4FoNYAW8Mrq`DV4VVlT+ge)mH>Jrbg=fm|
zsl#Nx4QF@3*&ILT<c9b{wab1S|G8DyPBdB<>bt|L_X@MQxjVmPxc9Yr?YRD^FXy+!
z`j~vqgKG?DnOUY}X2nKHh*zyu!sxL#zLlYrfn6bo39&T}_wBaC4wmWFVx0Vf;f;vg
zI3|3r_c8y>pl0R;iRa);wFm9#>)Azn>a*{!`&nE!sNQLc7*NO(*|T<;DWA&DN;uF5
ztxd47uW!{a8q4^|@bLenwq$Im!I00_t`Oxo(u|k3NIXuYu`MjU>IXIig$3PspW94U
zLtj-fd^)DR1}@p(kF~r01{>)r@egH9UrVzf2G&{J<c!NIHEB+$*WY>E1|t{A`__C{
zNqJhNIMH3UGl6&7)VAs*>qDNUV7vR`OkrDfWYH<LpTwy{`~I4umWmzYtLgLftzJx~
zmr2=u1*`ncHFJ5VN9zKLR2FVKA98zggTC*?xhVN~`|66?UTeu+GK+4|zw+0LfZD60
z3settG#3<fHg8^Iw{%TKQVV;Fv`qe#IouQeFil>W`e1OMxnTW4=ErnodPEYoxX*-j
zJY%$h=VyD?mDYRl1*bZ~W~MWIJKb6k>+>e6PI_S!w$-^U4Rm`nuHn>e6V%aI=*U_-
zS)2N#v0d0cRJ=C3jq8U^CT5@}#@<{z=avJ(e|D8h2aXoe<z@Q~K1cGM4m+71z8gsV
z^fZ2GVSl#M+sbnPXICe8G{ro+%6<}hu%xoSX~JvMoEFU)-s`EUjH;;T+lQv)UE^UT
zCEZ2SZFI|)s(Y+IDurJ@4_O`3aqSKrc_;74vo}BKQ!>IzJ#J#2oR$y~paMbk&{qM0
zyfc|790)-o0z#lE1^=%qpV*M}o^qZ1Ar$t;F@hj8r2xEv!&+%Qx%Dgxd*PUbkf=%l
zI5A5%ZC(A07=?XtOhu?nr2xER6T|FUXGs|f``}m|QKu*c;9Z^O=Xhqt%@`DdV?8xJ
zf>HoJzUPZ`u(Qz?6b`^K5h2o)0&vB}0@|e^TTc`g;+TR^W-A5YJIms9+7MzY3JY+I
zA}CcU0C&4qBh1J1PoS_Ljx`Vss!{-+dG57|?P%JJ!u~iWBV?LV0EUvH*Ca24zqzYI
z`p#94lFR!#vPQ@-CL-yQjDU-n2-QW6KxV*tnDm|Hn208ziGj1=C?2G6A0&W@W(#Hq
zfgtdWXQAi;2$EEHeB4rRb3_Fb>9O>jfZd=gMACQNz(g8!4MX4yCL+)YhG2Lw5s^+b
z1P#H9Vvv|1Gkj%6vi`P087LE!h!<>zT7O$n1ymtOHh0v`(4jL=AJ0x*SDgW}0$Jj<
zC?*2bQn0kfNLwbmheuP*sovm6KuUPDVn9GNm<`-D#bAJKOr#~y3I=CET6mCRJiY~K
z3N*i33;e(@Pz(tOdiamEbUN!;E8bwT1DONH1rt#mC>(G**u^hUOb|GWiRc3Q*9Ip=
z$A>FOa#CcniQ*+lA^$yM@B*odOo+L6y22N^!9ijuG9fBH24^18o~ml<hz5a(Vr1$p
z3Xis88Iwlh&DPPR>*_6WaAZ3<ySQ3uv6h<Hna%mXU2OwHqq)ZOelVSHzQAIk^`akb
Y?HM-umW!8bX><9(u-ht*yUI%FU%nv45&!@I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp32.cpp.E8494EFCFB4F3B30.idx b/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_fp32.cpp.E8494EFCFB4F3B30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b8061eca8d81e2b274eadfd0280729be046256f7
GIT binary patch
literal 2008
zcmZvb3p7+~6vvM-9cIp%X~ugDQW%pVgbIy9<k^(29D_>Ug*22$N{b#MDS9E@D~Yjg
zDby9>mQX5VJg!HDN>7zY3Z;8y?)tJ?>z=jdKj-^?`&)Z||Gkfgv!i3190b`1I>$u>
z#rP8;2%_V^=omp*vMdDk$U{)v*+$Ptk<6vECqGDz!#-UX_G!hq?{7MF%<ALwEvs6$
zl6KrPGOivUHtW1>=tENP3>B-UM^i&yJ~%LpYRg^;5*oRz)GxhN{Wz)WHp+TFkp8A2
zYCN8QB)I!KGWead`+9AaiBEsOFZ8_z8QhsV*$_a`Vs(hSGZVi~HCJMLy&t4K71&e-
z|G~C-c)D@y)5jgxzX~>?BaBJ<XxQ?|!|@w?bH`c=SCw6l4~QKio0rVpIO*F{eO1dV
zH;rOxZ?(yxDCqK$wsq~yZv!8MT`dLHXLSrC&Eo7XO&SkezDKk<e!P^J!D+Y2Y4tw6
z_pc#pfk$xukv&n%Ym3{f)w8n~8|>5f8|%fet3_^5-RdUJHa=Fn`^!_VuK(&{gs7es
znyjE1L@7>u8s4i*+FP7lveM)5x>$c*i=#yNq8=%ODmy)SwDNc!lu>bwB6=DsQco#8
z%uS41M;!ik_}Sa4T)to5p#75~w;GK)#T!fLZCG?k#CTg1*No3dR@W3gM`s*bZ>!|n
z35gZS?lHq>n#82ilV2Xkv6aI}!fx`lEzQU|S9FPn`rD|6lRI3qwBuwRv98H<_F<8s
z=@#r4lkd1EO2$c1L-!&qv$70buA;FE>|zcMpIG}T4&&#({=D=4REyQQ72_k?m=o`F
zs-S;lYND{fvBGp$$mT<>7vA!F^;-j+ucQSzQ_nYB<fp#&4&{fKm%sli%qxo|mNFkQ
zJCdz$Q^M^NjD*g9e!`IO_$jk2dad0C&B7SHkL}euNlKzagPnc;>x*dODAVSFPqjnh
zUHi|w<Sbj2aDaE|JQ_E8|3zX;Qpd6CGQH_yFPIb2cu<%fi52vG+*odSX{$2q#ppqE
z*X~H9Hknn3O*to-L*kn8Z_}}s_0)>ivXKNY?#!97AvkPHeee}g|Aby4w|qg%^iqf1
z(&KZ|{{CPw*F(8Myn;Giu`k{(U7(+Sv-5yz$&Jl*l?E*@CobyevhK%w2|^qw_2ZGd
zy(V39YCSS&y-~?B^D?*lv+vxy>lhN3n90l?9h=V1dN!gcNJ&~G45@e(sMu#<J1;-u
zTO+gLRYTwzgR{(e$G#Zm<-MwjHtRZZDpnkru6Axz-@{n+%2ki^!$|bw(8V98a~SS7
zGTp?B&syCSY)j~>|NXJUCo;P_ojWs-Gs+6GFv#8`8)mLEKRP~Qi>C|(Dfzwe+~9O2
z86j+dAWR;Xr%MI@R;5p<@9yqG)r=v8um;C44AZ3o@OE2{tI5|r<p{wY$7GmHlM295
zDQfZCDrfiz!2`!Mn8uO{!25P-vG0g-&LM<#IGzK~p-KhdRMEm~PVAGnP{<d@3?f5b
zDgd8P`)=c7qnUsZJaJ5dNpz_ITy*m&y>7_D4Iy~pm<m%Bqyq3myD-(eFew%x@NkU4
z2u&&gi&vJytVjAK2w^>r<zP9QQ~;iN;l69Rwcag+@EeXPFoiA^fFVIpa5RT7>#hv>
z7gt?!>N8tY-cblBMph$h0vAyXQA0FAWx#p@`4`JkjLxI;fwSN!9wc!e6o6t1JOytM
z1pe_VBs~B@J4;(95-QE}Wl)U4VQ_)nperQuFWo>fIVCv_;0ub$E6Hns;XyHy5=jFz
z1izFPnFT7tS8TG$><ZCBSm28Ih1rP7?25*td4Xa<Ie5{M)`OytTK%c1QdQM~7foA8
zTLYp&Zai8t2_PB_0d9|E0>CzkDf5(lz*&$V50cEmY>*;P@u!F2;&=kdG=QLg-kY4B
zPAO=_J4vym*n+u0F{&li7Gwvzcmm0&fU_v3#8dj&NqlC){Us<`5D={+dGVIWe-9tL
zK(g3;1x{LRIS(Uv^Pzx<!0-^9xdge%66e5j@+1VM(B@K|=&S5lbh49zs-lt_W0|G(
za<0vam5Y=)Rys@A^ZsvFg{h&bJ%53&p8i4uLt_(DGmE8K=ITa^?Uhv=oqY%nYi(D{
HxI+H|eW}8G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_int8.cpp.97C188D55BA3C9C1.idx b/.cache/clangd/index/conv_fwd_xdl_scaleadd_ab_int8.cpp.97C188D55BA3C9C1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5fe42018566b8b8e939061a2b39421b22b7168f8
GIT binary patch
literal 2054
zcmZvb2~ZPB6o!X^Oh{%D5(vm)Kt#h4IgKdDVFV(V22o&uhzE%Wmu?W2fQXv4P?4~r
zptynuhk{2T;4a9aEF!Exko7=C1-w8@SY13(U?-_+Evt5_=I`nLUiW+b-kV@Qfnbsh
zL0)3Nlz7S3wOSAaF^M;Et8~*YO$fTD4M8c#T0{C0Oy)BlO;H4~Z5Lj#Zptt0y?IYJ
zn7m!onOhyPSS0PpvYQykF6BzNe7Ujh_`08zgPq6fwiIfJq2DWNd?>@-?s!POPS}xV
z@z*rRbVG%QFstYF<qVC7ja3^uve4@1U0=`>8EF~~bjOU8z^g4{*YWWKI;*FS7<AZY
zg*%Fr<X()`dirw}zTMPAX|Q@UYtY@obEZdj8{4rvwl63u?a=TY&gPMP(N5(3WA1UZ
zujX!ac{zS?ZpoOFjcuaj7e(&lfvIHSdH>9W3`J$_Ujvj&FRpB>H`n{*EvTUK@=Ma4
z3ajlV652<+&Ue}v)eoek2=kI?981YcNO<{4^~Ip7(m<$b$#rh5E8dXyagd%{DZfR6
zZT)!sy%rMM>1~5?_h(VIpWStaAG~|HS&MHBn}P)f=o+(;4`WOI79_wmN@r)c%wl#d
zR$>0(O|uX4+?eIQvrl7;lQzk9x*R1Bp1&z<&)d7>`qA|0;!L~If}R@;-f)WPhBx+7
z1Lt9u|Li0Bai=0%d5Yo_0$F!<rvHKbUsKNqag!Sc>aMM6X)3iysC7#4)77$H(^uxa
zBf_YIHd1L&>UO*_lbYwDp>V6CEeX!`vCbb{=DkO`rBK({R3{)nud!(qUme*tmN{vw
zg}y8|4N`tt)xVv0d*j{f3ADJLf|aW*zg&}@)4qX+Je!d!>e;)eXm?qXz@lDp{PO2l
zf)izT8XVcRvqdeOlHS~-ThDRd@3$3Z2%L6>wO4o@il#Kwj9xXnHfFxylJwqPYj2*Y
zbMT{^y{wXb^2`bj{Xo><iqlq_QD0+Rj{n=wD|y<RJ8_FE501RJ{#A*IKP&ceKz?aR
z@t*@Wn!oW)VieuYDeIkG*OyA=ZmEZhGNzn|de?@2ejoo&MVr}*qK7k<zV5}6@Aox{
zYrI0eWmD+~%3{PjXI7@Q$XB**d6&~NIcH7L%~KC<tg}oSB*oYrvCkt3hb~w?aPE+W
zNjLatUU`?08YU|bj5-^-CvHLWj-1Rt99U1p;^Mh(O>OhKQa(@k`Mv!}scZH;e7pL7
zVihI+(1{$SVDd@!{js%9f%?j9o%Xc+uCB31bp<j09rYWleX#CNdgX_Q#!C0)P2~?w
zdEJ`o=-glSkX$xwxU4X$ZeHVfzi{-FjKpfoa-Zxkc*~K{Y!0StCc?A$iL&^0AsP_G
zTJ<s{)VFdM3Wq`vu8n9j)k6HM>Sv=Ut+#ku_A?Y-MqmU%m}&udGoKsa@F~3tg_jbT
ziclG90eFktP!?M=8G*vV1ZE%%j#>cTk!)&pt-GKIg_jeUgplZJ0eDZh^(9}cyk-m%
z5!gt}NLwucpV~L!5$VBBMd1(vQxFPMEdZA`Dws{r<^`f~7=h^sT~{ptcP-vDtsSBK
zjKV?!qX^1S3&1^#DiBVe<rx%SL0~dMW~c?=$>&Rx-4@PnM&XqNrXe(@S^$QmlJ$w<
zFa&8z(-yk4yztb(aD;`JfNX-KYfyhkwtvBu9Hsm<4aRV)A(ai1i{Ypt$_4>Tmhu~S
zUMyQj97$r4j6n{;(J=Lg0%15)$czAW04EYgs#<|?F<e)u8xD-XNCc;<7%)nzP<}|Q
z;S~~ojkrecpg6!kh5AG7Fr3UHPX}4VaBY_MbZ~17r?4o~LHUTgno>C+GGb;$b@)C*
zO;HZGIgtk|)ZzP#A!LMsU_m%U(5glOK`UI(u?=c#ZxKN=tQpIIHSkSXtNI35gO(xE
zqv{v1jN$r1{YbDE_$Q21UGv>YPpJ1Thv4Fb16A(;qq`#xhrh^Awh{%VxzPBai!hw-
zLgxeTV3ly7YAmo9!&ySsw@MNn7PnS~5~b0J7R0ED_h;J|UVSik7!Hqs!oe;GiunE+
z9WRdCKurB5ftp$*gse?LF&aaM?#uLE%wbY}b*JgE42>4JEOc}C@LXi4&vmtMwle#_
sRRfdh>>1`WXU(>>wwYt^;5c{Qd{dsWt&^9&fxs^k_6g!I(FlP41IX;m1^@s6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_tensor_rearrange.hpp.1BD4F049133037BC.idx b/.cache/clangd/index/conv_tensor_rearrange.hpp.1BD4F049133037BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21c4fd727b825a84dce57232129632b5121d56f8
GIT binary patch
literal 14330
zcmZ9T3s_Xu_Q22D$H~Tq#%v})p&2I8&@f5B8RSqfyoUkNEE&PT#PWrON+~@+P(ei{
zB*9Pv9Z*Ekl541>7^L_JB_yQ4tEiD;KFYt>NBTeKa?d$y>-Y7+Ui@b5$69->wfF2M
zM1+Rk9?LN1=@B2#O<s`Hk71bS`2XfFNS!&HZye$O_HjvFQNp~hMmsI7{j>U5qwmfM
zW7bTs9HF^fmfK<7V4JzJ)pBXcS6`DEKW<%7{?A=lwspY3InxeJSf}poKDzvaRf~2u
z>mGGPAOG~|xj}0t`i?yG-iJ4bzm#(1+?<GiWrn@r-L)&cWWxWgO#JbKgTux=n6=>X
z&As0Z?(91E%bf>Hmc?nRd)7Q%-SkO-?$ZGm*LAnQ^v5BKW$ezaRogNKmlitd-|PQE
zM)e<C7qd5O-YdSVZFu<O<>fJd{kb}K7#HB*$zItRG3JKDuz5#cVcyX%8aJc(uA+Ef
zTAA0H0>g##^(h}*H!>eF`LnYEjbn8CeVsDK>~3H9l6C*d+{Cf_(_Avf75(OuIPj-8
zQh&(+W7f%p{w+l(W<Gbv{Wpi|to|3&`o%lF2hO|l!!MCpH=lap(DP9R0nZF=&-?G+
zhJ9`woHKUqL!X%gyYzqk(%5@TIq^j6Gk!^be6se<wguBp?$3`{AH6De?0`ir>M@t1
zwuf$aoZgy$%f&<E?vohxdw8j4QNU9#haFcOs~>&2W6O*$;$miO**bYnYrWUtVU2^f
zJoRjF$y3@FnoGBz)%X;gT>RRm;f02>XIKB>zkS`aJ|AS<-a2)(OVq?_?bT4{v}E52
zm5qU(`YojsJe60Oz=m0q-EO*+Omi4N_W16R-L(@&+@A@>KNU{c9@GEnm)~Xwe6qgn
zWcosvsn1?%$sT=@9XxLJJ4;+DhWux3$O--ZjhBeaTWdFe9QN%Jr&qm>?nnsQ6}rax
z^@r~}_jnaIdwT7Au3}orGqH(#{t3<cc-qM!4Gk|A{Ih=6)z?SN{^&bn=C~bG=ahdx
zY3`cgjr%GTBR)5r`7Czz)Wm)_M`xxVx#l);N5BvF63=$&byY=!guiw<&3S!h?}J^&
zZ+E@zTU)X}@XW6XM;~1|HNWEf2M?#cbKt?};n8D)U)#7Xa><am5VxXjaodNon%kix
zmL6Pl@X*c~=9gO)w})lF(PqB&`%Z23j;t9u!EH@}hU<3C%L#9%H!(~1=G)w^Z|SLC
zkZ^Xu!pyz7k*6;Q{(bzw+d;*5DyNJIZn?a5is6R+qq?i>RO62g`6I4+!;<gsR9Z?$
ztsVQ};yS-SU;TFK!r9Bhr*Hhb`fZ1h$BWN;IDfs?w`Jc{pMG`MCan)j(9bQKTX(x;
z=fcIWPr19$<&V`BV~d9W>yrA$-J|~JAFqD?<OjjVr0(i>4_t2Fdbxe?<@SB|{KwuL
zl-&^c^xXG@PY-Jhh|cni>=}|YYW0WTS+X>Xyd&RVvDtj_`#~=h3{>oKoDmS%HT%*1
zZz2Xcu&$fK_C~)pG&*_Is(mSEGtUkmeq%sJ*ZFI<8B9iX$C2CHdt9bE`Hwl7cBA%A
z>AQ~t9BS^(Soy#(D&eNHwxay;p|6!;bF$htjlHsNna729)0a5eU6MV<4SYR1<uAwM
zzZwQ<*DU^_|G=1E+rE72`#ZWHMs9Rw<}aE%?R<ZRQGEK%iFGTt=Bk+GunG7;!vC{F
zpdxTAqkL-^vkf)^*Z~wPF074IHy>0ni)HE@D9!=TO&KY1N@Ml-!EOb=shG7$-B86)
zZ$@c+TE;#Y`N*)^Yp;s=SXMV#F`1{v_LnH_k}1LWehxgXVwNFgoPy)~d~>LT4c<}q
z_a{Mqk5tU(va*w)I0?Rt(&{R4Zj1hXDYwT|r(z0_y6vFYE=*$TC<z=By6QmM@t1b0
zm<<Tb21PdbFiL-w1o|!eC2MC(#Oo?%f!y=^iuZXc%u%BJ8#kz4U3+POiunqu%K=4>
zpf2E92@F)^ZvDzQ`k0EDE32EUm}_K|v2HU&Qv@IRvFO&u^k@~6E>oSL=oDs=;3;W~
znI19W+Sm$j74wnIy{UMU@6$&iQExf6+)=MkTzQ5}-weFvZ^MAwFi8B#Jl*FJZ@Fw5
z8atf(RmH4BQ`iZLox&O?j+WF-){fztj{Vy56iMC$Np`pf{jY($_)}Cj<;a|S5zFg}
zR7?p{w+a-ifcJnm21@GUKdUzyUz~7G#e68wAzl&B&ouD4_rw)Rg2C#88y7sUViw9&
zlp=~Bniouys26*0CjA{VOvU7*PD?>i3j6}wM$D4HjN#1%PrIE_F}VnQS@AO8YtSI6
z*O_<De75Xsdp{NPCBoK$Vx6!iQ7+OXX78}=9x;68D=H=%frmkHSeWhWY6<-C^OD>n
zmlj`8F=_HV#w*7278}D#R9a=)BF7<lTU5*nr0x~PD}1lP8VSp6dan5!@3Z$+%ohl=
zgTgNON&G7kxMW`3rJZE*1{JeN?)5Av&I-#C=OHaiR_?zqOz+L$RLne?Tc}vbYn$XQ
zX<L5o_@v*$r~j>DR-taa72f<L!kr{6Cu-24n&i|sRLn|*jaH22ClUIBgsoJ*p|2Qq
zrdh>&Cii&^6vu=mwhWQDHCA_B<;>0$74xagHGrZ)m}`XdY|*sq-aYu*b31}>sF)(u
zZ3!q!gsnJUFDWea%B%eBsI$9@$wJ^kP#hFM&Orj#N8I@AOz6B{RLl|teg}&01aRCl
z61eA`BXLWo`b4RiJOnx^ocNi%M&1`kxaW?sFlNE=msCtK!ZJXSAxznCz#I|$=GtbT
zw2sKfDkcMgr$KRA=r~3vb-b^iLul7$x&{@KgTN3)2(NGQ%M$iO$%+N@-uQQ<idlm&
zy+Y4x^L<Lfj`usG8ZonSy^2|lFc*ajk9qf#u^mZ6o0h$GPsJ=nm`P#ceQJV1ie)F}
zhRyDG%=Vs&$wb&*Q0x_EYf`=|&i2fe@T*?O?%h-|pU4Kj0E!C&_s*+QzrQSZd?sq%
z{yY^^h!m~`#ah7=A_q&J&~svl)0QOW|5VI;S>bfWbY5MER#NxJklA&k9CKY&%zC7*
z2oyy^ztJyBV6Rj2NBh_I@;T@OS=}PVBA$ACxI|q~ba8MrPTr+rQe^5w#fSXRCcY$5
z4-ag(6!XuDjVdNprV<s2ye{+e67_gxO5_ng(}_W3=6?hCe_-H$z{L)IKgCGPq$bap
zZyqV<d7|vg?s-QZd|h>6oL*&zN}{PGp`t!9(+gl<f=B3OR&$w^P%+sHsF30jxQ1xf
z5VNRL`M-28fI^l>;0ji=f|bDMs&fY~+;#N_Cw^G2X4g>3+xTX<s|7VZgFVBfNN5Mo
z2+t_g3<wDbm#POl1Vjc{Q8U&Y8!E+GJH&>^T2M17I4MjrfgO?}lA=&EH6%4$GNT<*
zBU7!YxfIMxAyld#?64HVmx2We3L&@<!lc|`heC)bgeW8^hmdjzm%_Il$|15GtVmD`
z=2{4qB99$vA-onWNYDhqO%NtUPdhY0L=!|IK?j6%K)7gZez`j!vIDH5U{KGrUe0ju
zi`V$SIP0BFQuN{tp?5VQL%8YPOj7KygS+0{B>DyaSr5I3DM`$yyqWZ#CS)cry_ZSK
z1a|Os^7WGPhaLR%ex~DM9N~?q4=^Dk2I+%LQt@hsFsCptDbm@&s<)b?vVb?SKGuXx
zJV`&vB$Y>Yh||ZJL>J}9oS;uINreaR0Qw{oa)4xgvPlXac9`Wf%S$Q)?U15RF-c_{
z?;iS86LOC<eVR#1qjpGlO81fihaECOp9v-@0SL~cUkWDVJh`CH1(Q@y+9410d0>*t
zDLdqYJ|9d{6cpS`UkE1TUd5m<29p%n?0j5v+5}!wX0$^Y=*z&AB$i8p1M17cgdDIE
z^p#+e>S;SvgHttlNmZg9YCvBDCMmoNE~>8u6LQfy(AR-U%CmN;2Yo%5q+-bqji7G?
zla$&8C)PKC2|4j8(4PX6RR7zd1)N&IOG=G)Xajv4n4}U%aC3bJm_#?{t=$FsE-*<&
zwH>;_sT;hcgky(ZF!aK>zCxH`_>bHM?QIw(R?ebN3c*b{o1oyfi)ePqkyRY00QMbE
z5PFK$oMJ=7fcfI(j*@uq-pK>_-pYV012bg~t{lviIk-wNQ|9oghH=OoZNRmGnOX*}
z1I*Mia9v=gmcgeR#-U}nNGumgtT;k`V4bX`la&Nw_rmP@&Qxf2SBU0{ybz*C3uxb|
z3ZX}c<_HNEyAijRfC>jI1a4wAn^*}HH-!N99jy@hJ<)tG>lF8=fC{H81ny!ryVx+X
zb8$5Uu<w9{&<m{Qf}9F1WYoxEqq2`V^XuXq;~bCUrPbAj<E7QjhU4W#_e2~o$9p8;
zcp2?E1;@*1uPHcQM)}6!cp2+A701h%fOye2`B4SM1jXZcX$`ZXc$o<?nGjFK9_v!D
ziI(C&p9|Jpu!+me3-VxM9wbt+XM8>+P_ZYv5T;PEC%PD>P_ZX!6U0%mC$<cxQqCSz
z4)K(;$5cYR=xzKss=-<fHstIN!S^9ZC50I01tSWlUu436h~^*JE#8bCG`uI@U*W~;
zbXGd6u?dx~YHUKKn;M%?>8{2mRC=hf36-8|Y(k}%8kx{X>7&LbRQjp036%kAY(iy_
z8ksOu8LCDmv?{G?obr^hYMk<vlhoo;@Dq$v#;I}2Qzoc!%2Ot(amrIBt5M3Eshp`6
z_a(0?MVX=&w-nz@Ri>(O%2TGPQOaAaT&$*2o-z~El-ntnf|_zWWiF^Gw^QbUnsPg3
zKBy_TQx<}nayw-)sFB-k1m#9hQ*Ng$12yG#%5qRsZl|mSHFCRMpxgy&<aRZntN}HZ
z@|3lprc$1=4%Af2Q`Uo;N_omgP*W*S*#v4T<ta~rT1+MUdY=L18BkLxPuT`)D%LAI
zKuyJZWf!Q$=*H`~2+E6~MzQ`rINXQl#W*NtIx;}K7ZR^|_|M*lq4(hx3H{$ik+9qP
zu7}Wm`-^D)B2tbMbpWrm@0v(NMdJ!e`YwgUzwrH?Aes~M$wG8m0Tu3r5V(QWY>)%=
zJ0F+KORPKCEu1WJh<}c(f*0fxE>{i&yr6{mmr%q;!fm98-Nb)4MeHHm9*Srn{tXn-
zNVrCdXea*d6mgDl=P06=`1ex8b;4bzh%DAWOFlpGL(gWpY&k9Rg7vKbdWtAwxgv_#
z!TRr@h$@z=qKHGR{~?Mv%yNe*;%C<XXNqWLxmJq!o%R2nB6?V^M<&G3C0J6pyF->-
z#%h+ywUM|>1+ee-4xu-Q<_3`xxEO{7RJgxG;AW!PELSAS*1)>VJ>}Cz^Xuhqc8@^e
z+}+|Hi<-XvzCkEp`v&_)qGlKuW=0`7EF{c=nsMg12o!YVEOD`@Im>@m5DKfaf@ejd
zW;&N{Mu9UuB;A6Vd0@_i2yx%=lgI;09>gNSCh*?`K`7*Hg5XUMi3HWaRf8D?v}y>c
z1`86@gSj3eP`IiGOFhIQK@0e|KoAN>EfCxSkx0-DTsN3esOg4~Zm=Lh9x>*T05PBl
zb0M}ol86MQ#8^rZrNmZB5jDhELlHH^Rznd@#Mnd;O~lqj5uL=?NfDjI)=3e!i18Lh
z+#<GH6p_Okb67l=9M+b@;<=Qt#uADsVQnQ8v70sSrik6FZ8t?Uu*L?8Xkcv(6w%Ha
z+bN=*wY5`3FKg_jh+fv#D-&Wi6?RhJ6*t;TZA8;1$92&i1hnsx8=*h4njhsLAl4@W
zD%^7;Fq73}%E4j5eb@Xd-}wuK^H4r;`4B|qd#(_I#1ZmDF>u8Yg!26+@Yw`@$k!Wz
zYlI*w&vH!=MCDoT6a=9>+X6l<;D<c<Ct~@Dj7M(S#9EqINg&pI!rX+*c{H~TM6*F&
zFR{oFz`pBwgkB|@t8$nUle&QR-Oyu{)m)X&E?J3#Lw9(MNfYv&v)MTeRkO|!&Q{d)
z2=)jU10@f8M0!M{reBC(1gc2=EPgSlX*FBJ#K6se)*4~8qGm#HLb#Zq`M)MaCPbrV
zN=QlssvA=*DKV&-$)9i`3{{4i5RnO1B*=&0d<aK=kPngh5RC+75K;yaV!q%<PzIJV
zh(UrHFxNmBs@ZBFq6Vx;&<Mee5RN>h5h5EQ8VTAUqzxiO)9}5tfu#*%M1ffR3N{nI
zEkPEuvl_cxp=59Qe6e*~<Ixme#Y50Ogcp%-J%U#sfe{HZh-(HJMhUnK*;9BO*~E}d
zMj|4cn6pU~666p=4n^b=LoOxABZj;u1m-+S#}~x=3&J6#>xk<*YGCUKw~iWE5it}|
zN{fiOh*DZY3?&q?kr*~of>L5AeL`R^rF4`N?{aEj+lcEnYGB(4w~ZQD6){v%N~?&u
zic-3p7<N;{9%9%-32KO;<_UqhhSIU0c<-kMc96Ipqy~17a0jV@9VUjul+we*e3(+&
zKnx8O(MSx9l%R<inw}7tn<yPW5$~U<ft@C<r>TLRCfsRiV6DW^N-1q6=2l8+J2A9V
z#5rO(M+rKKq4NoWxs%dyk$7LE26l<KUZMteiEx*wf%OnW52dt+n0qLty~NN<5!Z>~
zIwiP847Z*Tm~T-!?vlZG$xF!j?~zyU5xh+w5Z4FPNFNaH0X5Rc#PFEv<uPCIQ@v!d
zhAfua$gClo66COkoF@e492Rf&Jk~pp#oKfZ>$--;@na3kt&xKfA3q9MLjk3<fHfCT
zO4qZ7^%POW8j2`E32P{MLSQbTbd<8*rPRQ-u&!IEfo);AE!4m&SVIM+w1PEPP)c{O
zh8+}9#Tu$8!EV;D`w4-0H>IP7^{$}?wvTn)M-6Nr%k85Ec7Qb;pp+h9%?Bu@hgicQ
zia5*~4pV{#*3j^Tz}!ITXkxvasDYheT~AO0JHc`%sDU-JhGt4>Giz?9l>W>bex`_4
z*3e1`+F3*U69RKPrK6Me?xY5Go^?G>4eUJ2ou>xY%^JEXrQNK#n^O8aYxtcadRRjb
zCFo@hy-x_ty|NDRbSC6q;TZzTzgviAi+q?~elx_y)Vyl2P%&i_V>VUUXA@gCRoNF2
zV-ZCZ5nB;OR1sqpMN|=66-68-#={hGnAi?eL@P12Qba4UwNgY6G4@bI53%)7#A7}n
zQN&|ndrT1ptg(Qq{R>!Ifn59Z>rufPD=4CZwN+5W0oHhcA`Y;&0~FEB8k;GinYA@j
zL^o^frigCV)-4m_wW(l9;iU+&<d;PArCd>qWxN3PJr+Ue8CG*fzDyA7I05Z@FM`lE
zR?{ZCt9X16(7qQV2t7|U=VezCulNMC@68B8YgtXLe6kU*2L-h6)d)iCSWTUbZgoxk
zB=qBA{vjD(rE?xZs7m+bJcGmo2me<u&MQb<ExzgN<Lf7G1HPHSB?O^5Jc&yR!k4{V
zauB-go#ivj4_!liK`dWThn_svk|!T}#AAgp8{ugSn#)R}St*-Htm6f+?{N!4j}pyM
z85NIQ0^0Yy1)&|RrbAvK@vcQcg$FJO+($I~WS#rXMo;gzmRu8-(wTD(6vxdsT{+i4
zeAmyp1>(DY&OH#_23a|4AUg2Ga<PHpnU+^IiJKIN@7%e#Ky=^p2tpr0Bvm|w-XoIW
z(4Nkp%=yex;Tt7uj5Qv2V~w@pZmg4RxSNS_iKv@|@d*jo+R;f<u(hL;r(kPG&5A?T
zPKiyKiq2%IF{$zRYCk3|UOe6KgG{%k+t8I%4aC$yJUU6$g0&WGqQCPWr~_*q*u=fU
z3+iEFJtT_v2fUyW#y3KOxJP+G6GS(`6mfs^f>RKE3Z_uU^r#ky6Mcp!+90+Kri!OA
zUeEzC9sFIf_@Gtz1~{e*;?Z?@H(0yDhOVUUl5uxQxM(WT9Rx2E-bW!XTS7ET<S@`g
zj=hquEMG2oStf9q5GcAmFIWoPQV68J>ELo9kou;B%Y#64hgSn!4Fpn4%GE+3wWM4f
z1X4@N)k7d!(*4A;pTvk$;m1+KT54EHAbP$qH{nqin%gp>Stid-yaW(H;aL{~za^S)
z<vUn$xB}SsunVC_Sj`bR7>g?|pnY$<5c(&p`BO&4Q@()qz3xKjuSD~!Tw`Azt@yjY
z+HvK7zt__`>+oLJy6VIk^Go5Tb<>H)<(uwWcOBa69$F6_4!K%S9S*r#FC7lK4!({k
z<oap-bmEHfngg@}I{dw!Hb{rQ*9&t9a}*;P4_mcX9p3!fSRLN{+DSUR`L%I6y!o{W
zI=uO{NjkjwwaGfX`5k6C;?1v3(c#UnP1WJeuT9h8&F_%zh&F#FXfr`adAoKg=qPX3
z=7Ns$c5NQ$khkZ9HXn4Nq4|X<1Z^ScsCzzbG3cm!K8H=<C?;zDBW0j110A~XDF<yi
z=)@~Go~Q(ECFs!idez`i4UXv6y9Ts1pcCDg|41!pYe7ftI&B^3s9mS62OZjVji7A=
z9Tmv5O`xL!nf4Uus6gh>0*+K5)3$+*3S`<2&{2U*+XXr*kU4aNqqwd3q2CAR`yd@y
zf4|ikzqxSnK_Nc*aeij(p<I9&dngxV#vbYuHV%1c3YTKWp2DS?v8QlpX6z|G>En>6
zloLxi^{S+lwUo-QN<{Y*_POvn6YcVRR+BFW1knowu<v~)LR*NYMNa-=?*b}3(L~@k
ztmYe8r#J}#6kcf}a67BnE|-S)|Fyn;)VLGh0>6-bUuBA~g~Z(^s1TlTA{B+Krcmzd
zpMh_E(>br<ig5aHGrNVNIOG=YW<gD_V6QL~X}lu5qEIs^Bq$tNDkw6@ikg$mlR{Cn
zm=r$Af||*}$zkGx^TSAvNRC3yw2-uLWXH6~G%IT6f;ksL(SFW_@LaGUK`{guLzrkM
zUVkw}6hjmeR6<B4grmJ!36YgxMS?mo*Fh-SX>|}@2Nona1;M8v4DF*+5OE5kke~}f
zx*#0wk}ins0;?zxQ={;;bl+QA@r!EyE$y5n!LajF-<_Yr|0MiV`o@CGr#MS%{|f$<
p+}C`(*}{J(xfPhw|NXBt0$z@OFGIg?LBD^Ee&;_f{i~3F|9>pH6Al0X

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_tensor_rearrange_op.hpp.1DE16BC6AE42EE5B.idx b/.cache/clangd/index/conv_tensor_rearrange_op.hpp.1DE16BC6AE42EE5B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..769092f8be4a961b52cb19a091baee4a4872dd7a
GIT binary patch
literal 846
zcmWIYbaQiKW?*nm@vO*AElFfyU|<jg;^LB`Ofg0V22&;mhKf0@6Axw`HV|<6s@XPU
zg2Rf8YT;uuo?Y=~&V84poS!ThsjYHTb;h4+S@or*2Y+m0Kbvt+=k%s6)gm9?TyW|=
zc`o)-=q=k$0dGYOm;L~CZ?5USlgiFJP2@GKY;1_$^j0oc;eX!)+Z)R7Vm9qr6ZuVD
zH~ugGEtSoZ*=CxXbZ0KxR$aCBVMor4>Rq-*s+S(EebIQ-y|7`y)IG(YkCuJZUdgqm
zN3Mit?}00x9f@D|MoF?S`eN_lnHMHkI;*tbD*mpp-?x>#T<0oXO+RTl++2U9$UXjr
zYJ|b_1<c_g)?5V!8vZYIrPi%A^6iY7`@?psOjXA$9s6Bv({E&5JuQ5r{gh?dUb)@1
z%JWlmYFxhY8l5_Os<<*Y=^-$b^nH}=zu461urV-mFz|4&b0~lb22P+H0|O%uqc@YD
zwwhkc2{tBHpaLFt9uY8sP!K+^sfbHM<dP^8FDC<!u%fUkm|)-lo4}Aev2m*YrCoD)
z82EttCHW-9zyt#)6C(pZ8_<G%ia%-%x|s5qn7G*?T6oyF*p#3aFznxW_nh-YhE_Hv
z7NCAXc0my^fw2F^*`VEj_buNIw4aNChfk4D6->bFXDCWdD}D|%80J$JF0c<_*7ET1
z$nluNe96NjC?{wRvy+EM(ooV1<^z}-MlNs&zzl%-A1Dm+JWOxQLKC5TF<SzefQGP2
z^1uuM^5i5{VLTYk!~_l$n2{_z((KZlFiDt+Knp<Lhbe;j8z>C&G|Y8C|K=nrA^Dd9
eCIbrqc#swsB^Pr8RWk8(oYwg*wdvF^1_l6#nh}ct

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/conv_util.cpp.2DB2D9F346007F63.idx b/.cache/clangd/index/conv_util.cpp.2DB2D9F346007F63.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aac25c6623f36ecbc3adcc848ca592022ac0f4cd
GIT binary patch
literal 8702
zcmds+c~leE7suy?VG)9$2`P{Th(-iONZ69dsw}dFO#(q7LMBx9)gQ>B0$LHQKoO9t
zsI6M}{<LB#?$uiLXGO7i6tyb0Rx616Mzy~Qo|k#j+H?FD&SB1c=FXeUyYJrnzVA&`
zP(Z*l8w{IH1Qlhf3zBdQ!^V*R`2~5Yf9YUYIvvA`*0jkL*;mA-7uQ|>D*24~RLT`>
z{>k6mE^z`HtZQjupN{{Y8Th|r%LeZiPE_Rfv5byirhRgz<z{Jq-buQ(&iXwr<BjTM
z-L^3=AF0Dra&q4kE|_0Z{GYOxbzvt0XD#n!KOs8Wl0=@Dt|xQ2Swc6PU0GIRZ$0aM
z)c0_bBX7X*@!jY9gBC~+N-dnnbskO#dwlol-DAJr9}HVid~A(!)S<QOo>xqYjIR9R
z*`B&YN57EHJ9(*_6vh2ZzSt~q{8;a?L|SwEwPiA^&2I6&)5mFr=dLGQxYN>*^%5+;
z7`fHB!}a(zU%l|I9DZw2Nu}EZ!^{mKPI;f4u<erM-CbCoWs#P--R}VYBlq#l&(evg
zF1h}LxxtCEDlDRV^lp^RJviaYH{D-_E4`*a+7aE;>r=C2ukZH8s>!QMzYGmc25%Jl
zG1JR8El<&no^vna)RA6VW4%o8;^Rp+{%2`9FN~MOic6EMDzoZbOLHQ1^A?!q7I4D8
z{5JK$21QG#{=`3jCDd~#ge7Mu$=xdU63xc`CI@pIggtzhuG~!*&sg0tV0EP`=Kc`%
z!^VP`(0TD2+<HoO8LYGn16vK^Eq|(>mAAq(B5q%M@2t@+`Hh?k_l&gY1wOG&;tQT%
zH~Q~REHT;i>x6hAr?|i<pi6h^rpw0e2Ga_fZcSc)vtxJq)}L?XIzDxkWYNXa#Xfa<
zfj_GEdkdb#cCTnUl_dH0NybfubTzGioH>!$WbjQ~(;e$##ZBv~)-sucW9E6=g){n2
ztXLG5-oCx~w)1Jp<Z03<JLPuM=vS*Ay$JW)-u}~!!in4Gw(m}`9q7H`pVN`IXD8#h
z>5RSiy3Dx!5!<AS<f?+Gb;Y%!8^<DLUr%|t`j0ltTRYat_-E)_B{h17lO}YpDQUhY
z>NtOsaMbB<o;^50=k$7GS9bEuUE}<Jykgid^LsMo)74*2KN@&So?I&KVLxpDJ*{=4
z{hP+4A^WcUdgkTE;z-^5-VbEUs<((%o&1aYVBVR2y{o<(v(Bi_-M(hM`j30vGukpI
zw1+Jj*L`H&mzC#RSEWu$yJp|G_WDVQ?8Ty&qivM;a@B5;0V}Vm+7GUJv^Tx<*^w)q
ziZ`ncl!b7tvg_u@KanbhSARHnBB%30^Zo(LOtUoaLnj-)&x#+|x9@O6+iUJlZXeNq
zqdPvfd|zi#M`-Pm7YSuoj0Oj`{L<ZnZNB{{yZdUXm40xc^t(0jeFKWfwn%3CIYa%H
z`jG0I<DTvNW$LN>M}lpAmTro0-8TQM$;#`QOKW5kHb)Hh<gfK29(yl6w&?OtRq-=}
z9}0p_s1`l+3O6>qwrIiA=e4h$-;8QX?Jd+jeOt$rUYA(3UuUwDXR)X?<?`X`Rb|Fl
z{`}13^8mx}v)!*$M?ZeRqQxmNVhT1zZ|Hz&>IL~XW1f$G<?Ie;ecEgcBMPxXy`cj%
zCw7#d6h04np-+n=&3|iMvHroc-B$-Y^=Wd_Y^rBEcIfc-(HK@+SMxNwS}*QBzg>SQ
z<RU!@W3g!QvoInGi_#l9z~5fCf5&q7QjX_)T5oKaYgEu`;{Bf1?Mpa}uF=9)Q(8<c
zMleP(m_rBnYVA9vP7fD;v00xMPrmuze*5UnIgP;$3uU8eiDYX#D?9t418n_$<+6|+
zMy(H7wAl&&;I%%b{b)nZla%+o+I@${$D7R0AAV2kS1m4YKuk~Ad(2p#Iwk*@C5Gvw
zAGw%1DBcBVLqi5f5)L?dSu<X61V@Qgk1Y~c7wO<bmj^=Sj1%*dZs9|HpQ&tnGD#=s
zHa_&EZw_p@m}b&jGTJUCem34FCZD8n5Of*aZJ=<qpti+?PB^oj$8S7WQ|(nWXh<g{
zYzac%Y;S}D*Z~NIF~T@hyI6MY_#<;x&4^R~_8&Tt$WBCvU=s+Xv(pjEW8`tDcBSmn
z@ufNoKRV<0-C{aX0oWA);bnlm4COk&u0y#Qu$w97nO)x1AgDVCaKembW)5S?eslVF
ztpz=?!#o~vV9EWAzLa%1onTlptl`}xIWU|N@?cCuD1;FS2@W0Y?MT8XdpjD?DBI#1
zzq5Jtk`EZb=)swZldt$a9To~RWn*pc)O+Zf(+NvAOAq)Kbb{^1Mu_9aL5S<dMTqCd
zL&(a_3L()bkqI@pE<6`x;H5k%LO%9BPSl-&yg<Yz<I5&eHu)&I2{qJl_Hj-WDS1jH
zDAe|9#FoL!Ky3N^{E1Y%l2IilRKS$+%8+oW1ooA{iLzD$`)ZWyfqgy7Ex@=1FyQ?#
zooL{l#CbQ3<c~5Uyo^%YO>n{)ch-6gCnUH8A#dCpp#VGpp-?;&p;$at>m!^<#1j!B
zZ~~!pJRPB2JQtx-yj1HIoJ=FU0wBB$;LA|119%<E%>ZwvIH5~Cf1E=Pi$*8RY|Wgg
zK(jNm<G~UO3ltgrntDYAKU%#K&-(6YZ7_N^f}NEV8l;@UZx|fGT}<zD#m#8CKEvu!
zmN`((>0DFs0hKMc$g7icy6ZLREEfwG?OcKrQVS_UJ^~+Y5y6Q-i$KIC<I1dIfx(G5
zL7cWU;DpjbshwGHLM>1uwhW65#Fo#^w}vGJC(10!IIsxdL?sYZf{7GY13@*)^*~UM
zatknM0Y>otdB5wcPuRL&8o}{U=;$+V&%$}N!{*cZdw2e^w%*834<{^zmM*Yt-~?O9
zMu;QiAjB1N5#kAX2w4fOv{aJcM$fKZ==}k~21kS}=Nf?wC**f!7nH)d;X4wTPE66C
z!hki7PDJTPF`$s56NUPP4EVUE9oZhw?XtUd$R<b~B%6@h3syz43Hb()YJznZ_t*X<
z)=cESmNTHE$cfySjS$C|gAmu3ixAJ3hme)86+$AWhz&D@PPo{)Ad|J!PKuC^tB(hD
zXP{jmVv{+@Tqv8IDQClcp%ZbgaUK*Y?UcyWt#(x-whX%r#Fp=n??SaJVV1Dr7}1F`
zyE0_rt^}@?z=N_@1J`Pl>w#-M$}PaW1z5rRSIFvKC<?z8jo=)FYqViCx!yYdNkr)V
z#u40excnr0jxKLnvUD3Bj9sx;v%Dd78O$|uc|+<liZr^6B8@JiNTbUr(&#dZG`fr;
zjV_}|qsu7L=rW2lx{RW?x{Tttx{Tttx{Tttx(xE_wYD3Uz4H8SSkK{tdsrxZ5opW0
zFk4(iIz%*86h-B-m@k%4<i+<w$e-_zPzXN+A^B8!6qUUR`~(R_D!vM#G=3UFIs6=i
zN~V@X!91lC<$zxfBotQxeih2KfM1Jp6W}*d4EGX-IRti&Ec~b#eRpBK#Z=l<X{0hA
z!I`q7@b2=jIz`DJFlX-k$?Fu)@Q}?2_R4=9*aiu8T`NDpE4%0ZGEjrt6C>CwaP#P2
z#ZuchE31?CT&e(c!qnDu64XxZ0gXkwe15+)cK9BZi@S>qCL^iw-K7eOd<DMI6a~5m
z&Y~!s8;+2s%EP5OPACM52v~PW#XnW4q$ov@5<^kO)QlvG3b+LbY2+B}GO7Q)%jB>I
z;lyGfSPa4_t^tA?prH6OAovV~Q~WtF_#DJRhHB2#aA@Na`iM>laiMnAPbb`QcZ57~
zPlWt%KZJtuV1#1u811T`PRzz<Bb1CMBb17#B9x71BUFqRYj+3e#3Fz%qE`QOVky9v
zqPzm&D^P9(cq7G77LywRsN{XSLkW!sojY%2g^Y0;R;dsE<ss@5e*SO-Cl-vrf8A7q
zn<`_-9Tx*L28<!Yq1|gLghY-aDa;Z&A!dnfDDq-?A>_~UM<|39f{@%%E`@JJClXi*
zHWaB?DumKlX$a-8au6zUERn*P*UaOra$rMo6<}4NTnkvWC^rFC6UA^}o{SE3rFJ4v
zecW`cT=)T%`jmzDqK0h?PttKh%oLkJ_uzyV(+eSgrawX<%n*bUm<eXkAvmF8st`(J
zrXiHW%t5FeFw21%bOBCO0cI7-wSZZRauZ-SQGC5JtbA>MC})_6>y<Q@k3A`ukKhFP
zEx6~7WYwX)Be;w3!_u(o2R|qc3w!vD=U?UP3zIc{`OCNc`~kTEKTr+(xj%ni^IGq$
zUhfAyiTsw8k>p1%OJHfzevi~%fLkZp3vecY8baGPKTnmPu36%wz;zCWsi@N=>U56w
z6cF3jY;p1XRR5KMI^|>KoTOwGrqo)oWa{)|>Qqggsxpa;Y!V|gqjhjyJ$-{wMs(w`
zqsN$<1VsdehcRXZ*f9*lm=<O%OSXrnw~z0%>3+hGrg#aYlWhL!syUZu#UD3*!bJPY
h4vr#c7k9~2Yca>kHSi;ITelF|Tw3r<|4<Nx{T~t{LDm2O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_bwd_data.cpp.7C4E178AEE50DCF2.idx b/.cache/clangd/index/convnd_bwd_data.cpp.7C4E178AEE50DCF2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f8664073ed0b589c7effba1fe7e7f3aaae9ca6aa
GIT binary patch
literal 4154
zcmbVO2{e>%7k_6AGE2>9>|-~WiAaNCEFlwOijlIEu}qmHB3p_|8%mZU@fDSnQvXDM
zlsZK#MVjJk`HCTBPf-r}fA5T*iB8A)`kdpL_ul8d&%O6|fA>DGhclCDf`=f7VCT3n
z?$!Vl1VIG&KWb|v@4Ey9k+BdIS5WEY9{!4s9X(lHvTMt;FWY=~+0c09lD@ABbBhk|
zGkAIp?>+1s+0xgl&Skjuj>d%5Ande;wQV!CC}iXT$Tf#jhca$R`)qanpRM&;RR!DK
z&bBHD!k-0Oshk)ks-Jvt)v<|lO;PnLu|&`ME<1H>#6<s+QHfyiB=6JrL0)-PZ5~-6
zZEP%q=se85=_LPlVb0b3R~I@vQfWpAf8A{PWB2z`7A+*rZ+sZ-mxp_~hj;cGtL9jK
zxk-94I$O@?P*1<w7N6zh!1L9ml4|dgs!FoXb7RKm?t66Hm0FRIn(vN^b$x30Ix{}y
z<ui0;hLV2HsYttu7i~k=nb$}5^5gLXwZ-Yx4F?UqBeHz9MVvG736-EYdn9Nu_$=i&
zrDOZ&owTt(aB)iyuh9N>MedE{gJ%uZEmUOPcfN9J!r+fx@rkA(7o-)$@0N_bYw)?V
z+p?Lz6}{wR?mgS)cI&$zxGA0OH_EEbU-WA{N-w8y{iV7aHa8eAne@)6HleWX-Uu3_
zajq)0hHm|Zl3x#GDHjFjP%BqYE(zs4z8GvDX>XCz_&)j^`@WA$_c^T$XHTR(L5H11
zN}zWQ9D3{3_R+N1b%Pvo_p*)by`K?zp$>X)dZ@bfiT$MDgU8RY*K7@Gt=Ey1BvY{0
zCANo|c2Nq?)VPNf@#_N8YVNGrh^1R2!ZrLzhjK=D#1tB>y8c3EAZJ0CnX*}9Pew@o
z?YU6~4o<xW4w%Xyi-Dal-5ZO#^5py52j`Ba?7d&^!CzGtnY>9dRGmpJSe+Y4x5*++
zvQNL~uV<PWtTp4eUAvuc71CbTXm}xY$(7}2UW~V-Ft0s`BgAN>8lR61TS{n-2u>S2
za?_!<CbVFpPRG6!>rbdiQ}MXBntz>`lU4ksdl|v$@8Xxagky84{>{vKmuOlgd9h}H
zydOj5QvS2-V@&7nPeZ{Gw8GvKrSWQ|-&Ui(J`FJ6V4vO7@tXb5uB%p04n35U`jk80
z#xH6=!JJs6Qt8xk$O-ki@3E&*=9ezS<x>5#_YRe{lP4?`ap@hNJyjWU-A_AH%OaTC
z-OAU_47$lBq-^1wTJE#ZAvQ5=x3^hTqHk}AvsLyz?hV52p6d5~ebozPT88Y<-;xwn
zZ`}*|!q@N2X#VT37RUBH%QV5;=LEyH#CfR>vAwn>366h<xQwhkvb4o@F1mDN*LZDB
zZp?(+a=9%}EZ&qeExs6Aoz-1@GjC{Z+r(kPq8)*E=x%h}Btucn+Abx_^F{1RrETQQ
zql1M*(Zxfmlc89nI6lo~IO#KIQOTY<|0A6xYWKrO6~EVOTiL#8yYcQj^>nCL#|FMr
z(s6z7n&?wW2<AH<Zzw80EHE1ZL8wq{dhvE>4i4G?K^#ZO5iMFE;k1DNbQXMQM%^5J
zfI)h}+cYT}L9|S5LJ%!5g-4_3Sz{m`yiJGbc+mo^{rEMQ6VI88q>;YAJAAD>GJ8Xg
zv4J$w2i|5t476wg*8j?zB>Ei{$YYSJyde%Ajh7WIz;_{Md@--^IRyi4{GIOxjm13@
zWvwG~k!xXVf&sx?v;b>?#=M4o9!*9#X#MYe3$h|>x<mSKc;sr>nkYlWiWXo!P;~Ku
zdRJwWG;$4Wjg!JDh!(ItB>SQB63eVb8tDgbdq5s&(E<`1afwy;{8(m*gF=3{4MlR-
zN4d)UtPKJ=Rg)jbV+PW4yuJqCMPNCaEKPS1Dt35#o&D=O`{B2RPy%r0#jU=ak(aFZ
ziyH6sEogGU?=*afvZgWv9D*Oeta8|jDpaen)I5M%GykpA(15ln0wt<1C_7*W5<P<z
zSPor>ZX|wbz-g>$Gf@*hQi-fY1xFegGUQeT-XICpSS3XO9k4<nr<n9Zj}qksiiQ&5
z0jP@-feS!2r<UOe?;ctI5~}&;1$)?0%`&h%ZI(70L}8ieRPHn(K>mOj5Y|?gpRia^
z6fW_<`(^Id3KRT$1O=ZYEI6QmOWv)+Pm=p^5?GF^zN#Vc3ztq`bDBifCyS-!kI3Dn
z$Q5hgP=y@@kYRV#KaThOIQk<oKz~n`tlo!AD@pimf(&6EpuHDw%?JMzyY9ehGK(xG
ziP!L^l{Oe&6%r*v9fX787-5YRfCNxL&g9(`>RlZ^l2JVoiGW}5OG3JVD5n66x%jAY
zFtjgzZ3{hK+wi5lFgQ&D6$H2LiQUB$*_hE;qwtR_2<Oz^5Wu!(G&2U^T7avw|3a<o
z@mXyn=!09>&$cSSsFR2!ZIBq2qe0Y|Cg~7$#8Py(H20pKdE2b#g1nH389g?$;}Qa}
zA!fI)_Wr_zgs4hvcxsaEA8;zUcBrZc&ya%MP3P21hV7|{N#=3glwh3nSbAb+?<JSJ
z??@kh53BLYcnTP%`-dD!7EK40VYMmE)DobY&dT`a9edei*6cGie7>x79XLAcOL?0m
zvxj8UD_{pgP61v3|0uL*h=T!&<>=$}6@S);f~QQAsCcSa0(}3!2g5nMWll0CNO9Ka
zFZ3qId$$*)qp%zrm1YFsfx)7hPLs^2X452ds<~Jydu;Pi4SiE{Rxg(Z>NZ6@=rH+Z
z>Rbu=&tMvg8Vq;%^meSgdL%WVylpp(D9!+944|uG9*w)Sq`U18_*DCkl}X>C<DX^?
z?$-Ufh7xAqe7JX}H#T<09fX~_(0)|69#OtCzcuA69HA;pRXl7S?#rOI%6xnH%ftdx
zKmskgkwHFKj=Zuw`6m^Wl$3OU)FjYb;rsSg4D5iWPSXKpdA=-2=HOxcK3J{A(h}cR
zt(dzMc1o5HtX5~Ki#wvjB6pYU123;xvj<lUT$<_GgFp~y(RTodS&-{#yP2-x`v*1X
za(Vf|;@HiKS)%|$7H)juCz*I4&yadNFv@OAc|~xVx-4CB_8>O6(}W);paJ1(Vi+Va
zwc%=F97*6@xSALb61cL!FFe_WQwJO{b<Kj21g_g3Lk_$&=Ihx|MJwz713<<JnT$<W
z5(odlWP)GtUFwNc#~+wPayJDmg`)(`1Ob@9F9g64D*#qv1pp>0GAPPlm@yZm4M8G#
zN(}y~4gC4IPW*`A<gm0hPoP4rr2~QjxxwrFp&;=l6!7ygC@gqWI4rt{u_aKFQqpr|
zFj$=ITmpWc^GfFxZu6a(l=)~kqLQ4vvWk_p?GihChoy@Y7TS<47ij)YtEj4`uBEM`
kt4}pBG%{ICGq+%9&`HLo%M=v#%$Bo4kuIK&t_Tn4zw>1SHvj+t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_bwd_data_common.hpp.FAAF76603FF7D70B.idx b/.cache/clangd/index/convnd_bwd_data_common.hpp.FAAF76603FF7D70B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d712d61aa89e99b5dc070e81da832d0cfafe77df
GIT binary patch
literal 3788
zcmcJSc~}$I7Qkm<7)+RiMRq902eKwWAPIsX37a7-%A&F_A_yw78sKSFpehXtf(uJQ
zl(JT>sB}TC3-w7wL=>wcN(C{piGU0Ge17!Ob|&*a|Kbnk`;wgBUCuf8+%xCi_<MSI
zT%$vf>w3@3#Mty`1cD${@JUNgO=y;bpe`x|Wmf#Mx_3*jHTJgt{c#HiMy2$PL(GtC
z2|VhxEK1>>si~Tt(itPy;>tMpnJ=5auuFZU+j~|$*>YfUmg&@Q5_;c?s%eAbL$|vF
zH^S5R+8!)y+@ap#b;P+~;FrBABP9y)b&N}Q7fO3rmU@Hn`~QdZ<Q6>W@)_RN<yrb_
z6P?<8*-Uq+xumJE@~zFgGq1V4Y*Fg_b?IAKmRi+um+p4j-s+zmh<eo@;J`I5+<jal
zD9glt%e^6EuPN1SPh%Qu>#5u0+&1|>QnAwLvJavqMi!^E)9MOFrH3P=mZxZ|%kPi<
zQ5I|7cbeARF!K|8#fj3S>mNe9Iv-M&Zi(&Rzkj;6&h|UAdZpn8{xO<^0T#0BT7$jE
zlf43yTtVF%+V2&X-r8HwOQ^=(hZJ5)%75&?(bI0F9}wDJ_u@mlt$xVOMVZxDWYJU8
z9fB{j*(woHG|mIBi9mt#uU^uozAKK$PUq*A)NmzsZs7BrBbo6KR!%Px#XtO1mbvFi
z)y=xsJHxFj+YSUeR4p#LSeW_m-Bsr258Bs!b)dtPQ5KV9A6(SRyy9OMd@22|>XyK0
z{coh<C6vdus|WeTwjmh-N)F)*%#(B0$L#;v-(%-*l7Yw@HPS4zT@&`pXUvL{q}ivS
zUSWQRq&Uyxz!yi(EfFV{Y(J5me))&asF!DHX_oazdSZhF8_p&xYEJCCvw@p(GJ$5-
zR-^Z<K+;&)@W!jv_GudR{&>0K^!ll@eXLs(L1HJBm5K-B4#c?jH5EvGkR2~7PMx2;
z)wnNwywqaoWiR`V&QM6ZUPax+wYc@yKOArCZOwM}zJ5*jndk7B!}Nv8CCd@^_(lFv
zmd(8k^}e5{w8pJUI+G4k?c%c=S^h)*{j7Gwi%#FR4or&%16zXEdp&U;@r$j=+adqH
zTz)DoD{<X&ki&&OCc)(j8&M3>f*{lgG1B<!3!=!&K^h43!eZ)->5p$)Fh~=0t1GB$
z{PhJ<U>MR?fKp@Q(w>9thaxWvI&ICAVGKP4>mGs)akcO(Z(sgZw_$k%6<P_c@n!_|
zWP94kNW>NqB+piAtm;^|55-W9A;%OSgrQspcZrO63?7Mu%Y=McpQoLtql`p$B7)GS
z>7bp;n@-NbAGGuG=$LsERSrWP8IEk3UMGeVL9`9Jb_pl+ufPbZL)THk!=~bUd2!-c
zt|6G~#&9D>W2igBogkGKgGAfI!rNfMj+o;jk^#(#M0|`-A^|2KksIblB6rN4AgzXi
z?Csx$?E^cb^DiA2+%c`zT}RIKm3~4Il%vhj#ccvI<5<XutHaeJkujIp4g}@t@JM|;
zV;<3`+x?#uXv8sgUX<(cUS)i<q_FI>R5KcEX`^x?0zd0u4!?a|x~4lGbObYlE%EhO
zt^9UGGsWWMXX70VCfhfm(UVFDs!7)*_7Xw0=vpMwrfZXkL+7aAnh}&s=aPs==aDEu
zEkRv2QcM?<`ozX!B8mtqUM6-TjSLcp;N+6jz%p=dnM)eGIu6|P=@~r-mK(+lbHO!v
zT--grc`U99WM~p4Nf~d(&}2$7L4G|fVO#xkc`!GW8A?Pj@NtCbez~{v9A2jIyGyeS
zS&uY<CKHwk4>uP--sHmhIQ$lP5zUMy-c}5iz)lj_2OpM@nkHWPzMe0Sp+ZDR!<#YG
z6Y(UG2oVvq*}wJ?V-mSK2Om>@=<#`HkXt#2DLJRS6IWN>X#<X}5w2M)b8M|>ZLo}j
zL_q}Q`j;$!{ai5(EIEo9MeGNLehoW)4F}+p@L!nvcU=axcJ?_u9TlFVtp7WRV<J0|
z$daQje>fWWdC}(O%{w+`zUo)4=~iHKmD(yD+_U4Gy%3S#o&h;f`rLI^;1vczGw2ya
zYFDo?(d-UenhEAcFe6;?xz(>U%mzmc{|n;g&G07Gb03BeiHc#1V%Qp=RFfA}_I&Fl
zeGoZ5gHM!E3>7d01l8}~5uovMaTqWd!VDo&_Jm-Aw%yT@RiJqrXPXdLdD=|6I&{*H
z3AS9UE7r%G0kOrlBoc|eaXJ(B(58ByvKpw`niNeMzHD%4Q8+~QfkTVJB@vIpBT)h^
zK^0erp<;@d=xd=5-`<nF5dv=h|9N_zzM$)R(4fz}Z19D~;B%wzKg|HE;mmMildkNI
zs*N2o>Yo>zSMS%_LY<cq=J2vzQ|qcd_2lAXu>b3r>xkUP(0te_ANIu~dgCiEMc;dB
zvniKkl4ON9-%M@v`}iy|ZC+N>q2l)7qw<);yeRK9+?JQKJV03i9OXVbex!_JbGZ2O
zK*m{eiMter@+^5I%FxUp&iXsT!khVLZ|cn<^4;g3y_4=lArSdUW+dstl?OZJ!G5?8
z_YE#6^jI)DKFi+zi1epRU5b?EU=m*?sDHD!&%di51dSg2{#~WT?J5LAtq?0JZW%+Z
z5o;2$5jKfz5F3I%UFkF~cg(B+mB&=yl&CEjYNl^SB6EFng1~KebM&kOX{j-3=Vpc3
zKY{-|D!2eaG?}<SCMwEAC7Gx$6Tu%${87UN2%^Zu*?aas4~7d6q)do%sO2V!g;GiE
Y41J-xnyTUgiV|H}g@#d~g~S->Pt=Nxz5oCK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp.17D6FDBA1D601B9A.idx b/.cache/clangd/index/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp.17D6FDBA1D601B9A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..362e15c797c29c6b1ce8b495364ff82aca874297
GIT binary patch
literal 7106
zcmc&&c{o*D8$atfj&q!CJceUTW*N#n7ebj5QYyAtN+NtwsEi?Q=qfUn>lPxep{Piu
z$y82*qnlfo+>~47HmKCMkMCRd?&*7;?~ned=jrwS&Ts9#*ZaO}z3;o0yMw)by9|l6
zGRPq=JUGUmLL!kwu>a99QCq7qFegMJ#g%k#*z7uLBGzBdIhiNJuVvmOD&)PTCb&YH
zH67aa)*;|tV^b@8aQ4kdQwI&PX3>TQ_j|*CeVFy0WM?;K>uwOcM(<8Nde493?};1{
zDRax&gXKQ*a=s<gqchW85>I9-=VoJ^9w)S)J#S|+^oOE_k#5@shtwW+(`dPG<NMu-
z*G{ge@1MOB6J<h8nu$29QI5(!D)F6r-+!nh<+beS!<!BRg<}P^8fTK%Dn7FtIj6E)
z=k&7Lij_J(kIpHCq`xn@pmKO%Q~0|2I)l0s8=iHd>}N%ids`YKTaGqv_D4Dv`y9^k
z%JXdZv+uo;+Wcl@_Mh_#ommB~!Nb4X9U6Ni(f>eEV{5GvBZ&WWU7-GCa}E*}l(MT;
zt;e;zBO!$Db+XycUqHjZ@(54Q#&Zc3lMO1}m)%!w)U`_s?envD8+FK{wJL5+l@!_I
z@7FG2?UxhVa`4^wGeK5mJzaY7hxPV<T{)J<eqzuvGSVdcqAR%O>3d4h*3GJ?n`(z!
zlHTkxmaWuo+!Qw!xbl<TNMqX6CUxzuf+2q2Rt-6xicJUhUuv>B?55SdiBWm{T=#X6
z%7eA()b{S1!uga8ONH~QX-N}<W|ZcWhBnoGW|Z}`@cR_kq)+4{&17`}*51C1y0a!H
z0}KawLNzK!cMb>?%-mREKD6ad<J}rL@xA3&|8D!Gb=QtZgTGZx8I{%_xH!)1CS2TN
zXI<7}{d03jGqCb7Wz>GiI8!;hTG9OxYn`*JcYa}$o%5soWP=jTZsfexzOz9pjyERy
zZ$(nX>R!E<i_<#spi0uEz`0rCz{FnTS@%aHhjTq&Q<&9^h<ID6l}{wuH>El}^3xxR
zeyHvl>NPRt3(G0)OD(C?0X@$)UB6PA-Ii_Gn4HL3+3)?L%X4_qc=%NHoY33om;>4~
z)1O1Pr1ppvnLA5;R0@pcyx5`F&a|?f?b31G*pkJNI=-AOakTGX&gb)O2^D=uvNT0_
zJ3Z0|VjP>MWVhU}5l@%wd+(=QtlHyd&=&h!Navxq_U1Po(FyxoldEreu+BMXET7dG
zkNIhjP}sXqynBToOJ0_4n;dJm`CEF}Z-3g0^q^gi4DE^9SnZcrY%f|^I*Jqy3G>!W
znI?@^Ii|m~Q++$~@@kos<Gp?N{1O+hug~E;t*$?_^XUn#rdp#BM&QM~Yazm&{W~1~
zVEA7Yxf~+AM7VRs%%?AX@Ae^=&%PB&5=vIGIDY1h;f0R2l2QE{ROmy}@qcca4fw2_
z>p6PVrMR|!O*_qut*Yl5ORZ-wJ_4TJ_i--m=d-$#@u*=u*tgRs&Um73=f@X!_to??
zBZ)_+y5I9W7O-I7n~DEW*^zj^GohL%+;-&8-C<XqX!n&}<Cd`Xua*1dUuyPWaqEV8
zlC`O{+=jMgRWoMm-<W^Q6TfMCUF+k>UU`GfPi?%fOY8soPkUGPO-()8C%!)~{j9hx
z>($HX(8*)l{@U<r@`d&G75&VmH-A0CdCkdTd7_S)x(|x1@^WHzW6llb_`GY~_t5g1
z8lP*xWOBK&*KMmb(=&Th=m(32!e*(Atrp`fYs1ec{Z8H5c0Bw@S?__^__sohlfyEk
z=-uG~dSnubQ867Mf<7r^kljfnREMggJ}<DGHvdn|XTo#+>V{+nc{4^!Q>8`b1*DDM
zbqA?L$cHh=8!=j$sw^}wAiat|Dfg+{)k6%jD@MywWhLeXq(i;qA5==_j4{YA7%fAU
z5t$c|j!a?g4r~6ri9vS5XkV(Y>b!t-#(+l4YW0FHgcOC*nPf^PnfBj5aL9?Qx#gRe
z%k5#1H(*2LsB$9n0@76%e-h~!GjwK<y)jyiswO%wAl>Lwp-?Bdwoibx7NeI)ERmlV
zkZxLkmUqyRMP-nEFj|SKBs4D|eT&DhuxcaU972l5=rG<e@p%F1J61JR)rV@Q7-SEO
zmZC}t&kIPutT$77HC(lxLH5FERjR7!ynyu7;8fR>RVjf2r1cohl3*>K7m%L$U=$Jj
z&~Xcc?1|BeR7Jsg0qOTIM)GvV{S+DGO&F~}Rp6T!kpBGCWtWAyN*9B?4x`nn>LT+3
z(ow-7(QZr(`F2|O>uU%B1pNY}zW^m@F?i*?#;VJ9f<i%afo?9)g9HW51G;%coDY`f
zgJn3+28wL_MLfEU>hj)zY$}4rE5xh7VF;R_kU&6*N{Q+?mkW$@fd!89fN>rX=L7S6
zU<G+<!#l5iLSd)Bd#=>}&KHwQ8u)n-GzBQ60A2V61Wg4BsYILxl+%DA<lzP+^MAGR
z5Y$)0*9mR`;Wqhj{esm4m?u&Ik^&aNr@3!OQi(VX2&dsLM9>U?WDxmGfMgPJ77)(D
z$BTFlhDQE^G+;qh)>Aft`x~YKY?9m@0G9%S8qtiH&;p6%r2N4BFE;}U$_9>X-~=Zr
zd3pblrjf!q8U>A)ikE>uV@NfAA#xNnK`Md3C8#D4xI|W>H2z$oMxrK;hJoNP5GG>P
zVX*M5G`o8z>A6Ml5Ba!^(W}^El|x0}>=F6c3ajftc^o4ufqo@0hSOgpm}Hts>wSU|
zy@1&Z6d<v<Jt}l^zePIc-CiKr3q&BHJl1xd-vC|r%{R0LPdU*7&S-rTkEO1a&n0$h
zp8n<u7;Y@g++2=_S(m0O0>f-3KYc5V>Oakg#ZNGv;P(W<1OiGFOe7!<5abYJ!~v=t
zJVq3h1I#(V5}L%tUL3FqVw;Fl+VU&fzJHh69{%T<TNHByvs_V6(FEp6`wM9=w@-$Y
zlM!?f&<BAK{4j#v1@yZ(h6N3C5VxA)4l<#cu%Lq+i#iYWr%Vc9Z_`Kg7yO2xrikeR
zaxu=2kkHP~#RjRGs@lSQkDzL%YPPU2t@|ByynoDj#dl{XpFxw|O4`E`nB6<HcNjtQ
z(3l~;XU_f?$lM((&&JaN1+#m@UEl==!b*+V%`MZo-NrQIh5aA;vcbmS<HQ@xq&gs0
z2bAH{2>KdWz6LH3Z;guH<@)}l6_zV=0VNj*Kmx%E1SOA%^8qCv$J^bjh83nMUEl2w
zm{$S+S1rZP7LLnX`IOS{et%3s>4sac8d|^EdjG2Ff4isIpcD&4CRsC+Yze!VxGPV_
z_E7Hh2_tABpceu`+$TkVUPQ#l0R0#d7Xx}R5tjgZ2@#h9dMOc?0eTq`mjik^5mx|u
z1rZ+y^y5U_2!tAe7(TxaAkYCsiI~*^<Z#>x1UiX)Ct!6F`4J#6g7e9B&f%V=cdk<q
zv=LY~0%!b1tw6FBsN(o6kUUGoZ9t+8EHnlw?njr;JmrhVmg8w4b{Z_fADjl3)4&7b
z^ycZHu`7-i|92N0n6i^K9mtl(rY5B>r2|_y1eI2o)`2AwL1ootbzq`EP&suu9oUD$
zU}MYfKm97)NCzExQvW{m#Y7N6S%NHa*xn)NF40|5aGu<i!7Ky|br3;|0JDgQj{)W}
zA}$8ZVj?a9%n~9l1<X<+E(6RmA}$Baaw4t(%nBkt4w%P@xC1D501bS;oj|b@$A_~=
zmboud<F5Usic39Ux;$K_VkM>ad<OejQ&Dpj?79oy`S~OZ$=<_$z6~^Q;}x>lQd~09
zUF+2U4Of}X9~Sa%P69uq^23m08IybNfPqF(Lz>}2#ys&#Tx0CH(g&=q`xz+wjIUh;
zEdUAyM0^CO93d{@IiNX*ch1}vT;~hW$6fygSbhOE5dV_lRrK^<9kN))%>W7+#BMYb
zC}a|G7EsQ@yHi++Ff+L4BisxNnuQpQV6F8QmT}qhwOFVg1NFyvw|zS9Jn!1@6ZROP
zAgCY#&mCtp#yvMryJ_)o>%24v$iiuIu`UNUgNq$GcnWCr9@0CJEhLI<z<7E*p=>13
z69_1go=89(K<D5k1THVkAnu6{H)uI+xiB;c9__F}+;bg1NROtsP{G>-CkE0Zw~;X8
zzCQo73({_1^BQycA?xmZj!@Zt;oJ@^(p5mVikKcx0-2M<B~=Y%s&RbjfL~ScGvy~g
zWY%5xtMp0z9b3P<@Lm<YT~+t!!tYqpuLolFxOoUV11x928o1PMNT{7|Tyix4^V6?D
z_E(~xP63%y#0P7DObw22raJMhxhoong`EwUY_M=-fp5l=JNL8?7fnx8wcE$h@aP}n
z*Ir@U2Ootd&3?$o`<3D|l9zi3ad)du{K72&c6`@KAfQCu#D&8!f(`?nVf=;wmRKJC
zuWm=MR$A0l7VcIAHKUn{LUd>Kp`L@k^~ikpQer#NqY5no#;`yu=_#4OoWO1$eLKF7
zO2SrHEfA~44>1Tj1uUn48(atXT23h$W|eYJa_Xk)*6>#4zWjM!j;2EIH)-Pu8uQr1
zCXeym6Mu2I{%xt|Kshgh#!_PW;d~G@juJ;eyD7T~=qEt=2@voIz)ndEtp_}wU<r|X
z7X{Zgc1hvv>jj%>*tKATxHnSxpk=gW3ndR8xv+?G@1^iT`l9;MFop<fKr>j##;|PA
z_`kZQLQrR#GaY}#h2}y)YiMf-$b;rVK<jAh2*{J>NkHB-?*;Tpgl2I=Tk=mVRGJQ&
zYhme_yVS3G%p$G?OC1ri5zBD@M#@GK&`#N%3kU`ua}D=C4Y%5oX1U-RQdIDEe=Zdr
z6&O9nRpt>_A(6KK=MN@%P$wbx6!9SsCuaDLh)BP{paB1nE!)DjZw(C&^(6%^3?%vC
zY5=Zo!PO94-G-|XxVjZrBXQLiSA%dh6jy_BH4ImG;A;4S8W<k5ojW7PU}qi*FO`o*
zXCMN?LV~MBY;9CU_*aWcGR0U@CZ^^VmdjUI>4<BZDH*9R`d`0F$jHhmELK!jQPWtW
ewRD-TzM-+afwZ>XN^uE$hn-}*HP)+u6Y1Ym=N_N{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_bwd_data_xdl_fp16.cpp.32F247EF8BEF9A27.idx b/.cache/clangd/index/convnd_bwd_data_xdl_fp16.cpp.32F247EF8BEF9A27.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1dcb1319b1d62ac03144fa0acdd7a57f078b6124
GIT binary patch
literal 3386
zcmZvd2|QHm8^_N(#ti0I2HD3r_AHfcFeX%tC3}NT)+|wR-4I&-B`Re}lC6}R_!mOU
zrO0~!D&5<qv{ScKN{jMuPs#uA`Ml2Y`TXbenP=wv{?2*c=Xsy!eP2H}q0n3!L6$|i
zrNu`lucjji!X~~+$%$)o#So-J0zuNs+x_DchIn#!rsntMmRyTxe$9MZa(iW=N__TK
zMXzU9wLJ>d+`dJ3cDO9xr+z~HG%}>T_luzIY0|=j=ILlxQT2wCNc*YFuM?;KD`ipY
zIo12(OW{!ehaV|s8X*Wb^j^;GOS><<nkZp~Cnte6^i11!kFLqMZQsXIL$`kC_Xmb)
zDt(J;(vVf{{owtVf5(JiQ?Oub%U=emxj%BhS{_O1@e?)Z`-+QHJ_cC+;$rWT^IR@j
zZs2wvYo*Gc*3V2;eSaS`JXUX6Uupfd^Wn|WKaM+;>j;8l#&=#GFHF4~rPJ7J6q3=M
z;o<u4FMTJz6#u{i1{(^`<R3l!MZ^z_OdJSsk%)2+4PuH14GJIsD=IUkIdkM<D%ZhL
zJ|(%;@i%N2Z{7VBeeVk#Z0vY$OYXfiH1-iite;qvt>!JSb2GbYNq)iPQt#H{93(Hb
zDW`upwqNad!`{V9Ki22=Whq>`zk%1(@jKSj>R|qN;#R%H_#nsEZ;qwJr!RP0W;Q5)
zYqW(MGrsV>O!KIfk!e{`TxHn#8czFftFpRwW%(8Q?q64bBQ~Y&vF)YNC&!Lyo-bD|
z(Ur{LhTp9+PY(^cB0hGSQN(XtmM-693y#vhnYs*QJFhJs_I8SE9}&#E^+2<@n9tWp
zi91>T;^%)x8I_sI{`rO#-)nzySif$)^_?-3H`j*_h8TSB5jT;{ZZ^5PO#N3+gGy-=
z%RGx!xBY*J{#LfrzMVI`_hlDnn9*$Zuvk=O-zc(ZrcK^IP$XFD#XbIU{`mC?o$ZSK
zIisHjBUe<*)sIB&dzaLC>&EAuXWb1p>`7)-@SRo1AD7gPlvmWaP2N2)dDArW51W?N
zud39g4yqqrUL2iV|MaVoSjiFv<9+?z7h)PdT{9ZHQYmd8GIb;G@zCnK5os0|)W)wI
zw6pf&Y^mtB-!fE^`?)vU6@B!grY>OQlGlio&d=BB+qv$+_r8UewZ-i|z9a6C$9Sy7
zyC~^SnFeNH<@EqjYk~P;SKFs!`85SM?H*juIKeNMX%&oU6y<!*caaI^MNg!hSTot%
z7h9*NHy)~2y11C<^sMD-@n^hepijDb_14(4Sv>ovZT1Rb=L`n=6&w3SS0}bBJaIS`
z*UImcseL|W^Tt{IQCjenokF6{z@y9G9$4L)1U;>5ckxq%v3G7YhJAYYHmc+P@@)wR
zsv_SD9PzxtHIG2Fn3_@2ruYbdfFOz?FZ}~tPiL@b0SJOiFeKPBhS;?;A1@cP`axcn
zw>08S$O=jdx-$mKvB6vJ9#_tL#-e!=vJ^v#J!7D}r+<Evt6@<$iUbp~9o<f1#z46;
zd(uA0p0k-nTS3T@3`zEkfpT^GK6b}5D-RYekdWmVa`H0<$~PQq8Tvy8Cs{OKLS{0U
zax(_XLp5I9J1x6CSu`I)R%R$;GX~0U8~8f!2CD;DG!Y@Q8Ep9(1Ld!;|0!7TJY1Va
z^Cx6!hBR}=K>6p0XNtg1ubV~lBV<{IEPKX48A*&@la$Fs5PI2%&Ev!R;bJHrE+4Kv
zyB_#Xpu7_pkbDv-pCp;>KM?zC6?z(=_$k0W1@z#9uv+3$bE<^mYIHSOm@$ecW637)
z6%;Q8SSg@z8^GE~)|_|0p-^vsJD}svT<7`lHT@GtmK8RuCW*+`bJlyqVHD2?oP0px
zbAWS>WaHPNa(*t3+la{b0PY^33tzO{7}z%zA5%dHk=#ffILvpe-JV|F7(@)?`Zx~`
zI~9H$^-WaxhZwE}+*)!0oE68|#}vdi6T)u5-A(%Uv_6oaWSYC5h*#ED)(I8}#pP_}
zoM0!RINO%(1WONln<h1<r%_yst|bo_({JMigCbw^1BBlMPJ#~%5XB1trw~y1Jm8!s
zdF9!r>x)a>-w|;}a3f~Z4H>Z6q+>d$L<~F6bJm33!fI-o>-=nEm?-YW^_qPe#l5-S
z6!PKvP)KAZvVsm!+~3TfLIGw06bdv8oF%wpiKM5uE9_Vl*Pv_6c2L5I-ekvyj3Y#t
zQQWB6IX9%)_iOw)Fv+0fLXL1YdV=(X`Y;rN777<p$XaMaAqSx&g<OQLkakF(Zw$KT
z_JLUF7)}hO!Eyj6hiotuZv>o1l6SPEj|Nw}l@p;VFIA3!Wy>0O*0Al*RU^DHRhXLa
ztUz%I6$wqai%?urMN$)<LNE{@HRq7ZaTnV;mOo`fOb{p^s0yFT_cXu{iLYZ26C}ze
zYQbR?F9x#3fWoanww2_<qk4|22CqI4-W0VJU0@5s7>Ip6eGHOSQKzfVRt3d3VjC&@
zxC~%r<UU66c7U~$411m-HRsesaZS3W99$+8577#txJ<=TP03Ua1MD!Ma3{bzNv<Ax
z7h$e4tWTtp1-Mz{Ry>t<QQRlKQb-8WGSW)$D7_#oN^5X@agRyIy>z_{;WTvITi2UH
zKDs^>67fWQ@=1T5KZOE#0Tc@41<n#IH?f@4_X4?`s&rKuGWR4b$rw6E@e+WQ0196O
z*hP|WE|*L!6+}EElBxsTI`Tw<?&wl;?iUo_gl!U#-j4w62%vB`z`99(Vf_A#Ve_XJ
z!n*_AK@q+NPb#8@(>D~E2ba#J)G$;llv2YqER9mLa)6bSHAC?(fOU}!Yr~Y9bG@Os
z5#2}}W`^R%bYlvc&`l_0MmM97Io+H>7IceQ`oWeE^qML^AyQOyQ}ltI4-aZ0#p!F5
zOp!z9P&z+YE0|K66fA|3@F9R5BKsc2I{?-}GI1Bjt)6z9lo*+$L;Q;2lU883C*IJR
zwsIA+63&PqtH{5R@lkOx1iD9ih|$FvObJOADlI1~<H~kgs?U~kl~-0!R8h6DwG%A1
zcW_)V&zP^nGvxliQ6+T^juxh^t7kCZ$YkLnGYczgO-nUX^JVjtgl<7JXRjqLU^((X
DFnQ<y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd.cpp.63437EF3B0E7154B.idx b/.cache/clangd/index/convnd_fwd.cpp.63437EF3B0E7154B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9351ce43d998642cebf4b8463675b5c266f9ce0f
GIT binary patch
literal 4144
zcmbVO3pkWp7yf1pGBY$&$jmXWGmKACjB%@^T$(T$w@77NrZ|!^lS`_TM5X9L3X|&Z
zs*;3?r1O)Q+&Ys~y15jR+fhd))%gs2QcsWP@1N(fpS||l>sxER>s@PeV>vj`2@td>
zh!q{i6Zm5w2qGeXd_hFWI~fSNiG!eMVUbrDzuAH?bgX#k(MnWAG1GMO;wgpBSvy33
zi5&OsZ@MbyGsrU8@nzWe@tJK|EArNxS8R*a>^-^V&@M;^N<DLTHmY`TbJ}TR%~{@<
zsP+cY%hrye8fx*;p**Xk_l|>t=Z|L}ohnbMe86+u;5;YR8&g^F?B&oK$K{(p^AFY@
z(>N=t+sy1v)nU!A>ky69)CCNG5FC5@G_}L5<$m;2sKCNp_D*r&cFW=vrvlro*>#Og
ztQmawv^)8y%YzcUP7ZI}6YayI1>_W;$7FN=C^>!5sitQLpD51sfD2=js{Anx&K#ZW
zkl489``DuW2K1|TueJy0=C?*AN50|b6lj%3J5k;W$-bAl)S`DSuBQu|<kG3E-ngL?
zk>W0gx)(P;4QOm%Q^awUp<&P5d}$QKpu}+3F1ACh)1B8~%e@Of5KozD6w#reTCGx0
zn|d_V)}5oet=xF;SN2ZhmnRR)9!z+9we7k0wRpN!(wY4$NkZdh-3sWVrm!~6$3&%(
zQt!O%&%(c+lzS+V*YBg}5V5Os++tsu7s@r&F43dvr=?b}^-8cRx|kqHfzdg3G%aOY
z6PoC9ynD~{4`I7=SGiZGU1%YwpK~D%%9z|wG<nvwMDO<2UiDLZGtYT?3xXfsrpj8<
zM2a_LHuKCX4UQi#IS_iH@$=UBlBeFk<FA)s!nIbY7p2#K);^Kvc|%O?OE(N-l9{#L
zgy8A_RpQUI_x1`&QiZEbo}v!v9$QpnYLxl1e_Cu|MIZ<6KJQR$dv$x!R9%(92Z?H8
z$AcBO;6`C{ZFT19<V5n5rf7=1QR$%!9#4ZF8RxOd&x%u{N@}QJ*%9^rDgK*|CZ`DI
zXWffBw=6a}b;}VuZ@6L^rgtQf)^6CnGQvr;aQH8Kj5sW3$R4A&UEC=k<hxN^cp+K)
zBbArcC%)<EHaWu)`#inXv+`Y6$=Tmi7OH&R)k3rC?QjluTPcj`wppKEl`$uyT<c5c
zju&~w_jiRR-Ai6heeEz};#Bp8`?#G)s%*^3H!I!tBDUhWeQBCiJx>(FC|hGF6z?t0
z)U}*_s#_4fNv|X@+i2nRt(T^*YHT0zcpYrp)V}F`+KA1;OL_ZVQ7;bv19$nM<iBE+
z-gDO*pU7U5u%|qfPpSI4%>TE`ANyXN&2HMF@;)h&Gehg8^#O+GO<_`08^<POHzo8L
zDMq+~eeEPiY2G{eM#DP~Q;XlGeq~-iz0_IltmstSy!h6Z*Nf9mS!LVux;8HI656(2
zaBC$Eh2o9FJ)#^2y-xbp((er%)`{VFw)pj{m*tzzo!UEBaB*u%Yn;jlqM7&@37a@;
zXIJ=xZc`3FCM+No1wk^zi>=4ySl1Mx)eywBhODuY1sa_Z$d9~<^$#=S?_qhgFQT@G
z?6Hyss3~n>&EgAhFUX^p|IC_d*mVz6HdKv6dm(Cu9D^uX#y$-}aMERVIqW(hkM>5?
zigJq6B@0j+FQ`l2^yPZ85_+W%#3jxonoAa-ws9%s!UY>Q5YWpIwX%XTUa|o78jR)u
z!8$Tj5nA)J$hTy;*1_94JaK4G#F~I5Op`3YTBxrTds~hql1HyVtT_+|D_KC^*HnX)
zRAk%aanSFG+5)m5NEV>J+kW_QOhFV52Zj7h-M({4H%5@>pa=#3?Bh@bZ#Cap;ahD`
z=&_oN2Z!t1lkW9Cs1k+eYO}RD00*TqEAgLE=p-~s5-P|5{hdVkl=s`dx$sHER)ekK
z25do@(de=Aj(fsLq$*qW1!cJce^6}X(kM>=R2YfWmOLd=*jLf7o%LezK%cG|3L`-T
zAOV<w#(u{DJl8_Y!ccmtZ}G>5b;{aEa3mTD2EhTEV6flO1JAX_S*wAd@LYSGy&A{^
z&*k7aY5)RbBWo4UfKv6*@8U^nAaKAh8LWg?zzv?ON>-%-=ZFc3L<N#LoU2G|xjXNN
z?9A60?r`VdQ$QT(vUT0TZ5zGDSkv|s;WA?ofmRzs74Q=ysSlt8FhyY{WCk%U$+)$z
zKE6j$2G6C?DMsKHLIAq<1c^qcNhPPdw9fPBrQS$Pqx=L(yT7_!7vu4IXgmWzWDoW<
z{jLLZWs%!N1>#fyjc2aXtb{H%w{Zpnu)RXOyZo}_%8}bNHciSvufesh^Kl{9(HIHN
zAW96^7;iiSNGrt_F}g<0s-vE?i7@cAs^0cub}DXYQd_@hVL10@X**)eWH7A(sr;Q;
zezSB^KTT>2@J3P@9b*8G^%ngb-0ur!AjqmKtLuU&@mwustqBrUnJSewcFOba=$k*9
z)LksU2p^vGhs`h49*iV}=cDro)E>{J6X<H+0y7{OOpss#ER}W?M#hKmTu2B!`fPnE
z4HIZrI9pN%JCI`nnP33i>>RLHw`|x;Mvl!GW>x^q6!!iPK2dollV*27Yxz4@=e|Re
zUOJ*s%~q|b|MCNQGS<na4opiaMEs8O0u%x?D>7Y1$1OM>YY#Bnj#^&W)1S~;JYGc5
z?*B=Dk>$Em$MlEea|EmOcwC^`q)}dY6CdN#DonxPxeS<L3}8S&fz2jJOqe-AGKbBj
zQf}RRkp{L*dr~j?!O#2mVBjF6?It!uQ9)Xg|CLAE5iAtxiZemUL04kX-}UbQK1QxP
z9E!fQp!3l$u<z#Axh6exGXI`5jayX}HD?`qrX$@l@d-eN?}U~@hVO&|-bP*7mi2J=
zdn82)n<AYR;88(dt5I7}$+6}Hk|=RM{7~V8=!ani-i!Ymo`!b)rM@qoL?SUUG#OM-
z^9PQB3VZhv5|LvawvP01e2A+waNm7!IdZJY)|9qYo8{rSY1Lj{lL7{wMI_Bpz`y_<
zdqIH&hCf2ziO97d_OMxJq2>DvMSQiSxy$lMS>dEHb-UrnE_NX&kbLymdeSuC+-36#
zyKnT2M*;eA;Lycfo6>~LAC5ozdemHJ<o5nTaG!_=tk>pm?f(Un5hkto%W%{X5-M4n
z%mhPlq*3@;u3oO~7rqq1TkCI!3>DB~(*JCw2cR0#1AqW%QmHLJB9QNgd<(@Tbp*&h
z$qyTl_YVx<1+DUfg23E@pwS!SzW}g-pjF?tyl>mUZ(ILw+rY4(wc&`!33Hdh$jZr2
zQNZC9rzsH$Q(3O8r7qJKI~YvIx+s%WRLQ^1ooBPacHttsIWuO>r&$?l|LjzaqM@my
mOVy*pGmVT*Oc~~u)>;<ov&|f5sOdAE*rDhp9`?>CH|SqqxxpC#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_activ_multi_ab_common.hpp.F16D50D450CBF542.idx b/.cache/clangd/index/convnd_fwd_activ_multi_ab_common.hpp.F16D50D450CBF542.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2ad3b26124e25d5d25cddb68438be04f0c45c0dc
GIT binary patch
literal 6384
zcmc(kX;c))7J&N>!!QHQ3@|MS48yP)l|?oY5m15S0;oYF8i_^`6jVf66k!q(#D}7?
zL}XDAa795ykhozq>T^eM0fam=69I7vN}}NM5?>9br|KPl<VViIK21>{x2wDA-dneB
z71JkAn9wztqI|<9Cq#wEE#OiVrH_7M<6<I%I25%>Kv4-h>w@P-w>mO6rzzLf9k0Gt
z>s?aLnGj+T^|Ca#H^TUxkKKQ5+Ths1_uq`Nu$UYYd1vq>efLLS{T$mJzkFN!()Vwc
zTxq^g>dIN=+Pg^EF=SX{{uswV-{@<WNk8t$%b53mO@m@#{g}J^Hxw2W3T$4TEd4g+
zS<Kh1ZOaEmIQE>+y&Jzc<U;9%!;85OTI#y$9@V7P2G%yz{xJ2$__L<+<mjB}re#vI
z7h&3_SCtRiZmb*H@aE;jePxnq-L;{Me8M*P#)#(RDP4Xq>Tt`rDj%-DAjmn_Gpr};
z)QNH*srYV1lf1?J=0y3HvHF2Wi|1%bzYj^x8aie3@`-71=osg0^|#m04E-X=_s`)A
z-@8rw_514B?JF`gU;Sw}eM4Q|ZtFRZ3hef#SXfveH<@ITKY#r4^y-io+l=@9Ceu&`
zk1=iWfw`%pO8C$6S7~2)l^j~XMSojQx@ml4qDOI-_T1&R8@i2a&m~k=zRCCKJMzoz
z6}bYLd`t6M|CetauUrccTPyJkEx%WIxK}ji%lI$nerZ(82@uD9#&lTR+#aC0*fDtJ
zLAl?wa;<ld54?XI%!}~Z8WbF(EV$JsTyo`tfv4i-?NP<<iA%cP+D&cO(3)~JU%W54
z>3aSX@zWQ~>%x@2{Y{bctXz*<G({Qeoyc?e#aT1WYwT{(WzVxuA6^}KD6*-f^47G9
zdry9Sv#m=z(8>OYYh}j8{b85hgx3UXrsvJl_@#1-wa(eAe=Kl3pCP;K%N&UH<83_b
z>G+MG_IHsla?G<!UvDmHbTRh*&9vGgJA{f|6BTL%6y?@;Da<BXx>SdwLs5*BCl&P<
z6mmG#HQz%==so`8P+dwBi7gE+rTs;}SYJqOTC?s!L~5~~E~SOUHikCR{-R%OAfmQ<
zRG1l_IB`dZ!$V?wp1o~<K?%6XrzxuBd;v8n=Ty0_>UqY<(8!{{=x;67qpE|y*SzgN
zn5V<hMPf6anW(?$7o*nQhdr#Gx9$zr;b<eVj3+bbFIe&KZ{4E9+x!>naQH|p;Yo!3
z1uK62^g*^m=X^^Y4voYnJQJ<{f)Z#TrJyS&d{OKg235ZPVSK&IFcHv9D#s+1V|A&<
zu&Qz9LR)wqICb24zsFi!W-E-?3XZIR%VdIUCXB?K1x{Juj5!<JvSAeF9B|4Z<t<>`
z0w!2~AB^u~)^O4(-*qE&GoNM>rHShP&`gpn$%bu2GjbdGa7<glbt{a<oC)5U;EOp6
zMrDB~=4=?B4HGcufOiglo_Dlw_FB=>iUA)xul|Rp&xRk>=F&_Gn5ST+BkO>5Dpoqp
ztOwb8aKa-<1DiB(V_EN{G|DA+(k-OK6wpZlDJxLDNG}x#wXO$xJ(y#88tA8iHRg2C
zNhjNH1ig)9`%OS^BHL$xeg<x@Kl66PiaqpE<X=)}sXIFz4O7<TEKKeW;c%H2@M*!S
z=Q8)f_dfWs-#7X(d3*JP{6F|yCUI~g;WUZXi9^_jxJ;69k|{w$l7?cE8_Nlq+*WRn
zWl9h#L677iN>H~QY}_}o?%GRz)OjmEs{l4wnB2Lvl6vA9{z>{|>h^r~4x7w^k0z{i
zb;-EPVBLuXdtme)@M1@0E?hTyBmagH2^v7$05Vo!aVjR_S+eH_;39#r0rXjcb^FQ7
z+OG4y_~;vUw+{sHLyy?~t=UFQ+x^6FrESGWZ(tpOqx~Tc_Xi`2q8S&yi#{7qJ)}L?
zmhu+#p!t_5OeD;cBupYmE|e3b0HK03r2>X1uqnBW0>&t?QdB%!0dv?Pk9|AyWZFt$
z4(jkDaD9Xw&?owUy>#r!fsX}#-@WV#iV;6RJ(4&}+}U_P;Y?chvS*8SacHI)1kE5~
zM?f>JAZW!rWmHbIq5bA8)Lu4P=Ep84U;H%Bzo{^A?10C%xwqZdSlPE4qW(nMN2(*$
zyto$UuD}InxLh<I`(|)p$D@i2vsz#_=2n>1N^%?cwBg7xul><vijF^x-fv}T<-<-^
z_BV1k)$;zAzafY7;(O__4j0jV)W`MR<5x(jhXLJ0`0E3&KA6IGA~Gg+)wI|9$7#}N
zRpf33T~<J&Rgs%Xau#s2NX`asHpw}_&B45E`rcMq5BL0ln`Pzk`%`+QsYL_+;kewQ
zS=XK)t^M0r68Cmo%^u4mJY-l0ONXhf%~l2Y-LReQ_Br~q%~+ctHdRZwplna%Hr^D`
zOaTZAK!`0=2!cYAcYt6A$vZ)?ljI^06p>sEf?|?)fnXQOB_JpvxfBGYB$t7pjN~d1
zRFPZ_f@+eFgWx#H9U$(2!FW8KAnwGRQW+REv#9Yha?L7mt-=$7W{$z=V=xIftcH=*
zIKt6P4S3YRWGt_Pk#$&}Iz7+(_YSRCG&w2(-9*~e3tqi2gKfB>rYEfZg1_fKp31B(
z(VSJ~+<4A<^S#w`Hbd)NY5I>b$*5b-w6l8cpc!}CT}>JO8tZ?}eDAJHGe$zAL2O1&
zGb;^N8ngYTnHYl@f(k%ffb&tBDFksL$vZ&2gXEnc-br#1h>J)r25~XTyFk2)<Ps2<
zkX#DlQj*I+T!xwTf6)H0+4xI4xqb8Jridi>&~&&5w)b%Uy~Ase(TeGI2mY~DihX|e
zlm-Rpev-6@wH$H@)kX@t#0dTfF{_fY=Yxi|Ke2j?roc9kZX>hQTo{@Q>co#`wu5Xt
z5Zj477@h|sut%)p*stTL_cVLbXBTwDpOv7zJRQu_$-I0cSZ^fr@=YMyMCRogV3UFK
z@`{Ahnx99NO+Z#M5t<mXyR;+2I%kIT1leh*(h?H|NZ5|3*hv8{B>O8MkmahcTAa%^
ziwsc55(SCGYLp~MB1kTf6Qlrv0<S#mHWX=AwFkwTb>cgz=g~>w<69e6ETK@p|9_i!
zO2}{c_w4p5pTsFnZxIzI>t&tQpqYJOx{tKy0Wdv4@<EUs1oejD+`s1U4ezp6ev)3F
zZ`$iZ)t`+V@D=x>LBrn4|D5>=Ma(0BZo>UvgV$@A%6iom!_O+JhF=UtdfgAE`^g8(
z!L*#@Lm)YX^?G%EfY#Kz1_{Vbn!&OeFBR8g%7flLjEemv3csP}*ZfZO%;S@MyxuA)
zJ;i&5NHs(45*=9;qvH&RBz1g2#~Chi^@$11D8WIAPfQvoWnRxXcHMNq9p+AeAST~4
zv>ln#%-PJHjTCoIJ?!mV%cGDVRRFrx`(>JW1zxWph#h2;^t9%ETiK0|qK4vO^0;jM
zqislFGs9*syWT73#VOo<vULW0pijR$+*@ROYZp2$iRH#>v;C);IBpz4@!WWVR&iGm
zw3@q`pagCLK^<V+K^~EGf^jEic4vkHu<Dc#2VhUWC&Mmn?Dh{mq&n-v4>|H3)hQ9X
zoba_j9QOV76wUba{RQlsXeNLkKu{n*kf0!b5JA)V(+QfvpFz+Z{v0*^ajt$yp=ZKQ
z6fl-ZmJm+9RI-$yWs+rTLMM>R7N|NH8?!L>hN>b{6(24$wn9-x>MDofto}`+w_9*g
z-@?+$dXUjzLlaY}xy(#rETIOeAEQih)dW}7StLbS;lCQ=sv)jg<0|^61$#-@3PqXY
LswJ+fvl8lmb#~eN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_activ_unary_common.hpp.A3C6C6A8E8602CD2.idx b/.cache/clangd/index/convnd_fwd_activ_unary_common.hpp.A3C6C6A8E8602CD2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3274c2601cd68fd56671cca3550c8606a606b1e6
GIT binary patch
literal 5982
zcmc&&X;f3!7QQFMkdR9lf+ztlQy4=)MnRdRJe9eOMG@)*MS&s&HA%GsLIo6=5iN)^
zR8g!5wG}OjiqaY?4#f!ssfak?^XXHUKJ=Z$p4{8zFMgQ?d!6s@opbit^G&$_{Q1!{
zFzmCH{^Hbx%w=>8!?fVPC^KVKxC(|<s$-bArZYM=?dmL@i=n@tt^O`@){`4q%&5Gz
zoBfkM<7pRHJ*shy-dK>bFK*|{;Ul6&H9uF##v9{nL?=b>Q!J&M)?5!A_wcU{8qpni
zePvLiLH$90b;0*FG4o$<wD&ru&tF%$nNRyG(A}bHgQ&P^>{$N>)oWp$Uq@Zfd$K>o
zchh|X`oeE&cO1C7{p^nu!96R!c=dCC(X!K4HP#)+cwgFH4P0egsxz@=f$I9FWd-3L
zA=Bft5+mlvl?GH1Zw@b+bWJQ<WZRW!bEtdrLboFja~Avdm%lUKd!b|d#=B~=WB$J`
zE#-Ut=ZpG+e2q@*{<m>A4X(S^RUR4k;WhOv-JjSIHOFhEc3*3^j{VNl%@L9F53lD5
zv)vDNv>#V53R%Z%K2x=KOx)~y@aip_>VTcPk)bhBe@vWeP9MoX5P7XMEqSF|tIfAr
z8{)&(yv;Q9KCvy1OJDiqT&$SKR4HU!$*MI})v(U=OyO=}ss8oltaR~DMY)~gqHjCJ
zwc^g#X;#<wye~;PF{1B0pgzox+5Fz+D6iWt+@(4$Gb82Joa$vSvacjwxqLG?I%uE4
z;_T$XV2<;FA71(&G~JT9LVVd$v%dSN&eojz!8hzh9oy=&s>8Y*PW`AgJFM$BtCj)p
zKaE3k@-0*w4r<k&Fp6~@=w}@tE#fvMoNstu+86!o%vSq~c<tIZL2<Ol3koJ9yn~Aj
zWr441{i5D)GTNV;`uus<9jVHKi1Q}%m<a*c1>C=28N1@3c~TI6L~s_b4;Z~r!b*Ed
z%ZWA3;EZQxdEQSbWF5WOTH<&3bKx$#WCy!*9P8XN`*x4_*IR?a;~TTa2RFVeZ}@N6
zo<;F{R?p1uFn#2*JF3R^kuZ=u?U!B`Q0`xt7O}m#IcAbx&_5@>Ewk>~!`j>bd#l1;
z{M!msk?ZB%w(}T83(x8}ADi4hJEPuiU-a$X<)M<-D|X+UZo1(I=k}d9L<YM+?TSm%
z;q+T7^sgq=R^6)cT4L&^G9i7JE2;R!;znmr%ldX~?~B!aM{<U|;`kma57x^T2UW&4
zIypDI$XmN(7kFFm?X%c=t6D?sa390ywyL|Ki|g8SVkEnx#aZ(ib4HuygiDdvna!IF
zPe*eTgB#g5-Fq79l5gW?B1!6UO&W&j#65_LUf7z$rg1Th&}ZmtDG0`b*@t0(0Isjj
ztu#*~lG!wvkc2tITth)ffdy*WCHKO*i`g_L6fk5M>M95+5M5s3?02TzNF8HC0imvt
zP!NTH!Ay8WqR&wKKypzHj@d(jq-jah6$E2VqhY!{D3G+|o9hiTibmNqRVZM<FwjyE
z@`#K-hmLOfBK)wr{1IWJYeOiALO_28++o9^)it%c%lg?gH7GEPF^jJt7>fa`i~O4P
zg_B;fUWQUOO%n<b3_?pm$RnCMw`-jobqQkAbfAD0!%AC0NP+g~U#49Nn8{$%rb7Wf
zgRh|=q`+mL0}P7+%l&K`4h47&o~DA30?*&LrX>sne8r}5pnw^}jH4i=!0SgtrFLUW
zP1!VUC}6@cQB@F(MF$^l@$OQY+nruXzdEJ?1@v_FOcX@1gP|V0=DT~r>pbR~pJvlw
zMI|g5mRbrz9+8odD0&J_fPHyt-bE*C4S*BlfHw}zkdwgEpb($aeh*aWL_V<32eS}^
zP80x#0^msELf}wHVigB=b6xN9Z48_c6Jq6lI3Y2Xn4>S@gw$MWL!x}(oDbYdTmWVl
z01pxu0<S{w8N!U#S23G3l3S*%8{^`sM~`h<)o3`81&p%DOW{N|FwO=J=mgCUBGsvK
zVIN&C$j$-;V#o#YY+yv<T)@ty^78<mNAk6zt|YB##9N_#%^l6%P?n)|q^W&KL+A+k
z2PfQ^Zd??^GtOjHwGUNKGU<ePrkID0gorP;Ady5*qEA5<5-So(^`sP@R3H_Syk5ZR
z1)5ZB(F>H{ovHU~LFa`RT2OI;pCAxr9`EY8bnKx=!Y9#5-}UMG@6C&MeiW?4`aM29
zfIkKqyMX&H;E8T!#L0FqVD^qcMi=0A0b|54Ii8X9ILqTZK!*%Y7tlfs{%A*!nsa>Q
zM`zfFHayf#7qooDKt6@Ol$T9~BNe`MWx8r1U(O+(T}Wmud!`D##}QMkF5yThNXn5?
zkPL8SR1e93g^ctNosfaKGV-o+93umZ(M`PHM0Mn)bIPE?mw@vn;D;E#Y5RqQcgNJz
z$FKawcvuE^2<~X+Xy%6M*!*AfCQ=?J*1@bg4_N1c2092QdH}13#HfWpOUe5OX{kBP
zIWv)^f@hbd=?b@%z}+m}E&Wg>!wD;QD?e17BM0LHuI&o)o^l01O$w?Y?@*))^O*BA
zkqU8F+$3hgTe|R`$--n;be;Io`~Dcd_5^$^FctWslt65t_Mx5qy(emB@FaOBBTsT?
zx+_yWX`<EF`R#-M!pk-SokqY%o8iO|@Eiic2(QWztqXm<&qoD^<(Qri)DZ)R<(OVT
z#f5-gNaB?6rd!6(=%=S_sq~{agI^MxYp480WEXgGd50!1?!)v^nmF@~&rgk}hZ)e8
z4yFzv$nxs~{dxs4uAjrlLT_Oh@|O7Q$f%Tl`40`8C<d%zz#(s60$3$fTnbpFR9psF
zWmH@aSmjh)0az7OTnSi}R9ppERaCqKuy#;!HDFa!aSMPYViu`i8xXYp6T7y7xg@_G
z2->OqcHr7h@{wI&Y{`ciWXqgGL8(!-mwR9|901M-$j*op2Z75$GSzY75O6s};?3bZ
zyq@1v6~TCG06GnT(t>XT&)XmZDYoU`|1xmz+s1<pkY~lSLo*Fd=v(RAq3(qfhE|4l
zD64UTup;cxfVEY%Yj6IUj4XJP3+|$<gE%o4pQ}V$15|RK6u);<#|b@--gGq9$i2)h
z2Fi|y6D5FKLdB(kTS~=cfLlhz<$zmG#T9^CLB*ATTS>)LfLleyI{<eF6;}gpH5Io3
zpEj@n84D-cfloV$D_<>*%h53Chf&-O1l?dZ65Q=O>_LPDs(+I1i97@IQa<;hSi%Vp
zriZd1)jyqXJ^Ik>4UDX90RB<tw<2I%1Sk{S7!Zz;V_#!_%+7mjPZ~gf=K`Z#YDnh+
z<2-V-LM0rkA)l^EHN2Q!Iw;;-#3!bGo?1U2TF``JqKgLN!`5R_vCqP6VNH?&o(vcu
z!^*LP3^-A7fD8m7d}PTD$NFs=+Ay5OEHR~cN>~yKlCq=}Bm*oNSxC{`5A~K0{iNQ`
zOlM_4cW@r`eYa*c2KD~uIDMSMymUThH29Nv{zQ?+?kvrclhCO(K);6S#N9xDHx=&z
z277>Vxa``pv?*bnKk`YmK7Fmp3A=c5)s!1hxAvu`L$`k&grV33begEw8wQ@kAOuCl
zZ@QK(M{K&6L%Y-h{aUJE9ni0%;=RCNFKL(Ga|2aF{?ryjH=hTV=Yc)iQI^ph_V#Y7
z=#v<%SJUtMTaCwaKFQ%1&A!Rc@ruxpXPB*+(f}oP66L!heC>0RDytz*^a8tHvV@=}
z12ati)=XyDZ02ljR2eQWES7C;lj%=6Dy{~yGO9h6k3ygE9r<o3ov$9h`*Lg(1B1a*
z4|M9u@fs%vf#)CyL$19>oS1e~VBGsrYTvn=<x{HFIsh5M2w@zWSw`YB@3{GvvZt)*
zqw*hL)(ZOSpe03gks3N5Co<`o6ttGUmV(yN*C`R2%b^|RKQ54VbYeOw2QXApnQ9+?
zTliZFCxV&5EM#n)2w{d$P$)B$f+Crb6cojbqM&GIGzG;m<CJK!>!QUj4{;4lx)g&H
zN}!|~q*Bn=245=?nhW5C<iB;07jk4eDmxc?QGi17-#bVlPo}3*2+K%FSte(RGFFIs
z<qI1*`S7ts3k*}9fni4EQ*-5#M)W#ar3;+H8?^+sLOUZ9Q)9llm9>qfg_#9rr4+-2
g<Wf1jVVE)b)Q()5kxNr@X+bWn$)zp16eySf2Ur-G?EnA(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_common.hpp.1735C0647D85C158.idx b/.cache/clangd/index/convnd_fwd_common.hpp.1735C0647D85C158.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9dd712f57a7f1df28311dbffc9fff44dc9a6a6ac
GIT binary patch
literal 4356
zcmc(jc~}!k7QnkC&@?m&5YCVdQ2~L7awSnYglLlCCO{OqL=+(m2y)DzEQ(wjJkU`*
zK|D}VQDGGK16dGZ6-7Z;2NdxD5yb;#TxC|Bahq|vQ~UWBe^K9;{_0n+>ecJ_>Qz-b
z$V(#mtcxJ?mU(3+#ilQzAqb*J{L<1>6?KykgsY7pnWg`fhX>rZH|Pkssy2^XvS(+O
z^GJ2Z?aXyv^Lh?H?K!qt{%DztjbVRNrx&Z=pz?f%*n7u7Z{zmPM>X5?S_W=Ed}4q2
z-Sj6#&1xTpUhi%%FU<6Bz0G})#g1L0@X*7CHnW%k!!2>IXAikH<%~ugD|>nRty}k^
zh&Lw@?AoI}JsA&qcSj@tWBVa{S6MP{dF-wD<gY(`PHJAU<K28&X1)8}tcGh=VH<ZS
zB66Pw`?mxn$QPP$)_)nbS=hT2bMLNlVC*yKw0t}}R9$CUi%gPXUACaraKH90d7)#w
z?SOf!{+{858i^-Ht)49_c^B!jmT^ttq9L^tM%_;@=ef3M1vfmVpM4zvlfTf)E@yq!
zs<zb|ws#1uhAwpnrZwDL+4WyrURX)JU3_EpiZ)JERe!%_LfE3>h6wRx?XJhtZJxQw
zZOL034>h!zC};dES^d&5bNImGK7Lwn^0GDa!n@HkR};3L#)zMfV(TB9H}pK8yQ6dP
zt?{LM>_37WlRn>%QA`zFt=N3eVrxr9zum!m7QL~KPbWF%2cEV((5ok5&)jxPyLz|v
zl$~2#=GQv+Y*WS`m%R0;(979!CPjQaCbK@eGD~NHP3_0sX7A4EM#yCL9_gIZUqaft
zee^<+@RM3MeAk7ZYU!HR=BRt#r{R~dbDiEXHwzN;rvKKee_@kLyr0q0=*gRm*W6<2
zxMd8l7G2UBt}ITluUkB2%JQDn^7g+yxygRc&a1s}89S$th2;0U;@Lh<S)8oWJk8Rk
zrRMM5(KXL@-#`ko7bRTiby(2)`qSh-)5?UfF8kV%w!wJG&^AWa3AMG(?8zr<r3Ggr
zUo#GO_KBLMWpN;Gqn2xW?7B8P?@Cc&UdamC!8w;|{*j-x%fyU6SaQB5{cNjI->Y-=
z_80dHw^!ULwQK?h9!xogq^(MdaVJ*tw5o%in#Buhal`~cP)l`7(?1^wrZx$osUy#L
zzt&ive_w|q#)Q~V-O%vQ2Z8~Bm}(%YvCGqN9YW3GCeIEh8wSA9qkwf3OeNKVH@^NJ
z&GU!UXsCFm*q&^rp%Mp)ql%;sQi{|Zw2zlxkJ+SwqkImZ(uAV|uE3HUf}=u9p{<I<
z&f?iBk_aWPDw4WO=aHmQJsP@}@qN`e?9u3Gdilz>T5A%Hx|zFCuYjZOTz3ajuTC6G
zs5r^K2xusa&eA2_rd!8Jaw_?*lSYiycH@w}iLto5xr#isJt3J07f~|f$}Q{uej;KM
z;aJubmWb4Ad?{z=*@wG-*QB9*6F!U7M34pFN<{)wff+<r0&5isO@)w5$P?PBWJQ2p
z1ax7xDgxA4*6n|-L|(YYPRQhDf;T_zvn<~8%}V3^c);Tduoxr+3uiBMBaJtwubZvZ
zyxvDNXEU?8Wb@?u)c999u7yOrvgz4+WV6LU-4!k8sE}{8lO2H8*YqHk{8*V8aQKDM
zN?z4B{pWB-u6?huWJHUG8q<xb^*}>S=q3=E(oG@a)A^9o1#|&KLb?zlC7>(e04af$
zQWYQ?sst`dK>1*dLzN(me2KXq<#h!s=|2%xZvoC*z<g35O0Kly%&i+wc1Riuh7(^c
zFDJ4iA7J@_LF&TM0)Q2$IHWhqyKArS+;L=H<iBUFnrhc$MA&KVU_FP_96wwwa{e^*
zo|tlp!4eCy8AnTjTPg4(S&^FdL%?sxJk)ShOcU#n%{VHdNg$Haq!c9v9qHkY(3;0#
z<I385ejhk_W#br<+0r+#ns)0314n%^U%JXLKg<szf6O1E04xBaKr9d<876}$2n&KJ
z7z>6d1Pg&kj>#cX@D%ncTatOnFbR@fRI(Ia3M5Mrq)@WlQ@%;^vRhXOpT{}JQM)dV
z#`EK;RfeMp`~-^ff_AyTdy1wJ(Tp~ThPL#5z-=E0CTHE~t(O0~_{%Jua2nH>+VODI
zkLd@IKhqzg0A>J0fy_XNWK0=ELCheCf|<b(g)l=Pk~8HHC3_@GNfU82#Uq8HlF{hM
zTn76WV)5nzUM>)l-S&$fxpHJaIuo*e9!&J1dCLm>$IdA+OG=(6hu#?447`~7k%*fY
z#|t`hy}91dnOgzuDgbp*S8NE~^E9K8O+?Pq+!I<6BCd!!7|+-Y%EL!w(~1000+tf6
zNjHqKgAzEw?5za8B-cdTbv(3*p|6gk#+b1VX(@3j!1z?c5|;u@0Fe+ALZk$k5>Q-6
zAL=YjNkE8z{Qn)Lp00WGa_E5HgaqWmX2$*;-K~v;=NFkRf_xor77kG)JCZurFRqV1
z68qeuZ$j*C?mFUxbTlf)vEQ}o-#l{aK|?RGPGSsVpqDBGZe_rq40xN-w5l50))>MQ
zVP;`aw+OQch@#k0lqas{`=I`J^fQT&=7VYZU?!=nJ@vTk)03pM331VN-TB?07_a6`
z$mAVUvo_?q2Q!GYVfnFSP})rALn$hlA)xY@xJx?;sqG0zi@=N`RhHlSuDE^U&+TU8
z2>9*#Cm#maY9K_w7aJ_5?u|IQ7r5;OGBQYAoD1src--sXBxGMy`r}!mZQA3Ql6+6%
zem3?e<?HPXL0)b@{h^H4Sx(%)?P>N}<jgqgKy!e|k>&`I6U~XD-!FDp?RCv8B^IR(
z*Tz|;z?N$Z5s%BGC^a^5$=Lj9sY}zE#wz+_kFSM6RuBYJRW()BvHwzk6`ZUf2=N7i
zEYzzZ2=VQKEJkDnnS@@sLys&t7qp7OG%z&M#I$f7T|IpUT^rG&Iw5pbb?kG-U&Vl|
JAP7@c{U3g+v#$UE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_max_xdl_bf16.cpp.32DD457450CC173D.idx b/.cache/clangd/index/convnd_fwd_max_xdl_bf16.cpp.32DD457450CC173D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9f7e7ec21062f420950c299a6de77d190c816bc5
GIT binary patch
literal 1820
zcmZvbc~BE)7{-@CHYC{?f*i79B!o*LpivMJBbERm;DsU!iUKhpG(in=IHIT#P`Qc(
zuM!nd5iQ__aj+C9P)~vpv|2FrLN2d(U|J8l$;|ghXZ&XN+1=;&&ij4)yt5IZ!NFD}
z2nt*mnk5x)jyHoKh=E^en`Ox&QwSO~hoG#A=Ey!-UobK-zRp3|c-BR!k$Z#{ojPB)
zsY1sl^%Fa%a55h@3uG<7VSAgfP``-_^tl~k6Z^+;HxLKYbCD-G!T{G1&4s_VkBDGZ
zSxdq2SN)4ym#!S9<=N(yG<<mY^l7VP)5pHt_^fU5S;~a%;|k=Hs^{koIp^QL8oh-1
z^KS3cCGgO5m2SLl+s>!A^vPpyv!V*$Jb1^xHh5#*n?r+*Qn!Zr!Vf7*RYqWafc)Q1
z_hZz)OZP>!wOHw_6YsgxIcehE#rw4u<70w4VgBm@#giw~<w?Cqj%7EoCSs24NFo|`
zmshKfKB)as=2f{e@vuU=F<hG_j+IXn*jcxh)nh(3!utwx>)`b)ahKMEm#zQ6usj!g
zCzlwrfBh~?<t@~6i;X1OnY_Jv*Xe~3xyv$>of?bH;EXU!bP8PjUfC1OBwSMa_}KSo
z4$Oq_S%l~xSDD}cM_zFCSt^=dE2l4t;=YpUCp^E)4-j>QiK_FrY0E;~1MeIc@fQD)
zb0Zq7UhY;N*AQp@i0k6#7f|%>D2ZO~$YFKGRIaqodNODur+IPGGuo9pfu)1*tUjM2
zlis@+b4Yu?O5yr&>*=M#^0rb})r)W;*R69#e;cD}R)5Xzyy5uudo^uwZ$37r-(G+;
zurNVXdeT7C-(9alP6S`%W{BeC>JzWW&e3@#)(h><?z88Yg!%baB^^&zYOa6np!Z~_
zJ1*@>(H?5v6nJG<x<KVjs~vUkD)e45&gbrmx_&11Ks|9yjOzWH){7PuF@BO!g(I4p
z_=%@;&TL{^rJw1Tc9)Z_W#8Ok+Y)oT=1aI!`r5GKB`QQH?c5hj>R(!g>DF*P>QL+8
zuZRSeo}RnFRiu_|ep!CA=-p6k)3^4~iVgYEFXW?ZTCPo9o|vkb^qDP}j$QV$la?};
z@rG>oMzRZ1*admXD<qz3{+i2fLkDktN!j`3i&bhtU2jk{EpNW&a<qDtVxBvyZ_d0o
zI(U-1r7nhbLC1l!g@q-#zBlKDB+K&MkLXUmCGk7EDZ1<9bbfbYS%Y^&{VCDdsvN7D
zkm2(=mYH=*_E!I(6@??+{yF(=K1JTmM+WuVopaO!TjMusR!2M<7}7P&e)jPOy29l%
zeSDQi0+hBzn&4vsL9EFY3cKD#zfuV+AP7TX#KI`}-)LMcS2Jxp9Lc2;R^pfj(=3bv
zu>YCvqxnr`wN%0q9K$e7HwwTZl$}2fEbAzt5|-iE95%Nx3c%}EZ~e2{w&*#Puo%ZA
zn8Yv&z<au<H45Cy+Yo3Kju~bQbE5!Ux-i&eb?>(XDq#_hi7=6F6o9K$&sc-+Kh#hO
z5jZBpWDBDJtZAJ7OF;QpD3uV2V=7FwH44Dm(5|a<)$Nz5gylG<z!bVs0G5dp)9Uc_
zV9`fOXhLPX34$@)8A0I0AXEb7e_<jR&57m-A_aF7%6B8hZ9|@5M-U}$8*&7;@Ni3s
zTK<xTAArWBu|ZFeEM9Gh6;y-s;kF@8U>A^DEBE{^+krPBGs$d_T;WE``Z<mK2c`%{
zM2Sq0H-f<^%mkB1FbYaxg1A5pzJXyNU<1$*w+({;Hkbi!8wLPu&<wW?!2^55lbq@B
zj#*)bU?h~p1fPgt=BPOn#Dic|l*$B+5ez{QCYS~!6HDR^-O^;DG`8U(${@yNiYfO+
zuFbfno=b>|hN6KDK_cloNfQ1l6o;Fd5n+-!nTpUX=o17C{tqq;ionu#qSYij-?{Vr
z=kpc>cv`#rv3=a8e$Ta;Y(K?;<>=(>I&Hd#*NmC7yyq}yqa1FawM}qn3?XRg!Vr@%
F=s!E=h_L_w

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_max_xdl_fp16.cpp.57DCB3A8914F4BCB.idx b/.cache/clangd/index/convnd_fwd_max_xdl_fp16.cpp.57DCB3A8914F4BCB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..983ca1342792db4c637acefb777b2f8d25e98ea5
GIT binary patch
literal 1818
zcmZvbdpJ~i9LA5qnc<v?kqm=DLvE2<Qbg{?Wi`Z1$Te~+w~@qdx@;v(6v-I25{WJ(
z><*iTT4{5OSdFxjl~!`^(c0QG^E|&l_IWzbbKY~#`#I0=`#bOZ93Q%iivk{k9QV_Y
z9E@a#NkI@q!p;<Sa!j}+1Wgej=m_suKVh=a3VrdF9Xe8aOE{aWq1HKe^rd#WMVU=(
zsKPy~Dz_qX$meC`<$}(k7V~b$VC^8npAf##EDL>jsj^8=qO2OuNnWhaQ5ogfk|+Kd
z?=DVCGM`@Ff@%c6yEyXrc-PGPyNKi0&+T(&kC(patIQ^JZhQSDa$Go^WSWtAS+Qe;
z-524=$mI5%8z~+icso9SM|rG%lF9L|=pQrfHVpI)PA`1-!XROW8+06A9y0t13EVa`
z5#(F1eZXyMvHJeABGL}`qt0u8(;B(fY@zSzGqq?rBkqTVX|88zcJ3aXXx_yYdTI}U
zDY!s*+P%ct$5LRL-dmZvwMsK`59H0?Kf=1)P~{2Py8dPuw-Pl!`>IneKkLH@<IR2f
zPU{P!_YEa>HH7|dK%a}-;3R#)J(urAyP2eIQWtok@102^FQLqGLH<foQ`7imj%BWC
zHP@;gPj;^B(Fht<PxGzyP|vIuUWh5~@YW=6!cpRy8uO?t>E6`asC^!1)k@`9enXwZ
zjok)<V1WVO*^_-gD|T|H9M#=9z=7N7aNBCGz1Ij<Z6Y=v?kJmgQ_!?B7jDXv7;I-z
z-;E7jDbyXw&|^%W6cp$-zwvZ8(|c?(Dj;1mAH8vkI~^9C-5;Frak(h<ncdP|EtVTI
z^}r<GuKNS?x=Vw_;e(-hJ=bSf>f~(C>j-Qv>1;nsC9@NqdUp-%s3`a+Tcc|Ajnfr=
zf2(Q61UFHpHvf<Z#Yf2Gs>GE)a5X!@d{VRTQj>IWK=b0qhYj+)0IT?#!i|(eQJ-w*
zb<+4WMZcOy`p=Cr$7oNp$Zb5Y&NWZ%#t;vVcQg9e%X2y{>wXU7gm-0h)!nc0Q)Zjj
zsaRZnv&PHa-q2sqG}!i}B5nHfr@JW+)T0VM-LZbP?wDuEFZSUx-@~OdsUc=BHEGvc
zhm90!U6dvRl(r63{}IH?ZjER3X%w#x;8s`lM-^3X=G>cJw!Ri~kJqf1=07->HdQmw
zJ@xc$QS7NFr<;}M^i6FQ-{`$otL#@dj%hD$*O1@PX84Hq)lOc$H1C0bE1%05N%)tc
zrcv2BOCN1K%$b>2WSG?)TUaB)fx|;=gYpSs(b~@9gdYXYDfHCRijKDDxx>1(jFK6<
z5aZ3<l=Oq)mJ$%8rOGQ*8T4vJaC;z#g~F)3Sg?Old}!%Pjdj3&W(dw7!$g=UFBX7p
zYI@7h@Gss(aNZb(VOUNq0K3W_&!1#GI*;HO7$(33C9wb;9h5O#LC$-N;J(8!9>$Zz
z0`S>h{YE#vA^{2oVwfaFB8Uax0(%#UpusQU2+j+`(y+9gSOBhQp3s_FT)Khad@w8n
z%gBob;QmH~Tf2%@=m^da!w8I!#RBjcy}M<5Pe&7i^Tn_%EGs7#fRiJmQf_1E!4b>x
z^zf@45-5vgNOA@{2BG3){}(38B5D(jL8Ra+A^XipF<X=;m<Xc8Y*CKDb}wl=-*a+?
zh;4vKCDOo}p@B@r*p?&yk|;}>B25KRp)8mJQ$cAcOO_%_1z!rvzy^q#00Y1$VYa9N
zVCUc;UOn5MoQ2Jmp~}!e*kBO$wx|&BHo%xIDgxM`1I!i`0Bo=$%oYU??DxhpX_^aE
zo)pT$Q}9%<UzA0l5U3y?l!Z_bDp(k0p%jz~+JKTH<HJO&r6fnB&_p*(k%$jT*8Rkz
zHAH+i3&-3G?FBXjMI788e*k-QkMxw3l7{gF83ZND%dK-GIXP@1$+{_!*DETkSZ%kq
z*=cLH%Xov~4w|K&#<yH0RW)@@t&Q3`y7~s2x0slkZ{0?<pct7sZcuWe2jHA{+q+75
GK>q>%$c~!;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_max_xdl_fp32.cpp.7C22839582612580.idx b/.cache/clangd/index/convnd_fwd_max_xdl_fp32.cpp.7C22839582612580.idx
new file mode 100755
index 0000000000000000000000000000000000000000..be88175ed34b966bc0e34815c8b634981fc8a9cb
GIT binary patch
literal 1808
zcmZvb3p7-D9LKMVJFh#&h>^ia@*2XJyoHj-kXM@;uann^LhTGSQfjlIh`h2S<&ivA
zc6%X&RdgzkQF-K%6IQ~u=|ySVJ9Eze?AddgbME)v`F{TA|NqTrW_EC#o$)vbat!AP
zV|Z~PG7tn&u}cugj|!27pjkWw35#2ONBE=G=(BHe!6P0;M(m`6YRw<Ublz6XP5Vga
zp1aR-dd+`3P;VUh$bF1gneZU}=Tnf}K?i;52T-)%zI28D(>3<SLmJJ$boBh)78+;K
z_>VR7T6b$LyEAp~XW@}U!M3f!#E8Z_wrpE3R#W2kl;#Vf51;xBkA5>}R#HAqg_>M^
zpBSxo#yB^M_0?La_QHOQmK<<<wf8O0a?DC6fw%vmL6$<R|DR`rvJLt#PY7oOA@f~F
zildmcX>{%=E&qIz=~fqxaoX{Oie2o~Si$3i$Fn=u?CbHe3(6%JW@p#PoMjK#RrmT8
zOt|wec<Zyh9$H)Kys5r<txK^YgE+}4FjOV62lEK+vk!#4o?gR=mX&2Gh2LWO3uWm!
zktY@E9?^{w6r%mK2}xH@RypiA?b8so!7Nh-IpnTDS4T46)je@$z?*sp&D91(MJ&X3
zjmz?68Gdl3HNAN%c3=FpWJ-L1=^}qwc6w}^Ls(DcZBeJu8&P?_W|fGg;W9<x)6TzA
zbkiO-;5u1kozA3r*R{-cI99v5yWZGUrGzV4X}TZWSy?(5+sI~cU>V(r(F)dyBGpII
z<0JIECA~z7HnY{ja%uW}F^7X(S(`Fyvo$W|OP)7-GrE!We15Uz?)A)-)LM5uZNPWj
zzv(Yy!58&vkzq!V3b&M9-K@UQ+84$#N#}9o%DSx<)kW<7o4t|Ku`j<w+1XI5VmBRl
z+}pT5u-&v`G$-Zh$NY${qO0GuR5}d8pP9Rl!p3!@pBpCq@SZ*D=Du!Jvc-*K8Y?*g
z<J2=u-<OJWG`mH4UVHi;PhG7fot)?O+TB+J5O&P{lR=4NCXT;PwuW>4pE2Jzs|=$>
z-ajUJb`_7j8aBwEpA|L7eXTbV+*OY#`DA1>E_Zx;VN+7*!ZBoifw#xhUrVd3WoV;9
ztut-Ljb`?wYB;bsttHy0-@agF2ws`f4^-Y?oqoMZI7;=%ooR7ux>1&L_+q-J#U=gR
z-8Izps$;?J-S?Iq(vpuJJD1YXc$u|_*ygfxwOT%F^}sr7bJe6;VZ538^%u*@me#J6
z@$$x)Ma7X$R$2JYOG~RtNWjow9J6F<K7W;6P=Bh!N#GfOp}ezgAY(|k)~9em*QdxZ
z2NER2gqll1kfv&Jp2|~?781M@g19Jx%1Z+KFG^mTx-t`;h_5LL-i2W@LY9{Vz_vAg
zC8>9Et4XjIh7kmzNCIFN;!o#heC}nDU>^+Q5j;&207nKU4wWmPohHGa7{(zusw4nT
z>(jsGs+Zq^LID`2%24r=062S#vsB>Ie?m#H2Zm)4S&Ae8E^nC9oPGEHDhb|!VFE&s
zmju9~Tg+b_@;`D&urG#52uWEI08emwnyq>}Z<F93Fib><6iEQg=S2uAFndVYhiFb{
zX{Qv*r5aG3z{Ef*O#GirluO<~HUdh)U5fZ!rC2TY2~Gqmv0CgA)NX}sS-r^%WNZL3
zgRBMi1ZJ_-Vy$2`m=CMPIzgR|YcEavgMR|sguozZ0dqhwwpwf#tOgjX#a2NLgkrVW
zCa8f(tQK1Y^=l)-;dM)lVjRkqrOPtFQ79LoBMk7XLb*6PjsZRk%Ei<14B!Ff66r(+
z*d67P=p+Uh0p+4}lmWg3J}){%e5`;UCeXs}(p*_j`NYBu_zAq*584eTf*`;D9>Ze7
zqxWKO-FVxjWn>W?o<Ks$@)S8&>b9*qRHCbb@*2gpDwb9@wwvwj9gLI=thLPbG``o;
pRMphiX|CU(t*g&8G&W(e%`6z3=o?KPm1xc!f7r=;i;I*y^e<4tdTjsz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_max_xdl_int8.cpp.E84F6E18A0DCB8A3.idx b/.cache/clangd/index/convnd_fwd_max_xdl_int8.cpp.E84F6E18A0DCB8A3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4f6fa22878dbb44c49345c236cabee1b29d14047
GIT binary patch
literal 1818
zcmZvbdpHz(9L8ston?0xoBOh~?w6>uE+?!K+H8(o3hPoVmayb{%F&a9iljs+MZyzx
z9+yK2iwa39<l4n?DR(NQqSJ{pyXX1+ah|7no_S~H{rqO;JMa693yaCzgohxT0M@ZE
zZVX2hf*>-wN5@13b3`C$Mht?E<u|)}9T_%9zHBHeXotRt#Xq*WqU9AaWIAj9sn=J-
zTu#k?2Y<50C*kJ92IEF(!N!Mo$o8Z9&L`av9rDV3T`U??Jm2&5bD3dagvCE~3L}!?
z_s*wXja;067V+I$)v-q+y-a?w2XVg?bo!F6)~aJ?M1lqX>w4a)#W804r)7>eQhM=3
z25I*)mu^t%gLj-r>1gze=*aakyN9G$bZAuZ9O9;&i`}0#RqIFLC<97W#|?QS><lA^
z_N07TK~HfX-JLnBGbg)Tl`#-9{XFcvqKYB?7_;JQXIUscQIqbHb3KyzT6@rJp)Z~l
zF0IF%W_jQ{CM_=w4PH*lcE&K~vbg@7M_1^WHD$%B@>hB7jYkVBX_#TSz~egYjFwI>
zzx+*>c1x~n*0{S~DnTPhA@(ZMWx~$aw@#y|movV7%sgz8W;MYJYok#s6#bn{B2P*h
zrui0(HP%~vD0Q1LUvsL{>^Dp)wCZ*cH6G%HC|SPysAB1<m_(tjL>jMe4&t2IS9Bn$
zImtC<<?`%d%;1xF6-IX2os-6)r#vIa^9sEXQywMc{b`GdObx3Dzq<|;k(&+GjOPUv
zyL2;ZR-KCb;*-@K!#ne`1Dw)0_08(GLG_N}CjD*CH<dJxb}#5JTu!LxpY)|29IVJ$
zSU0Zo$)t3(zRXFKHNVIoZW;Z?$Q8ZK&Yj%+TYmp2*;2$|A)Wa_>t5mXS?R=i=E|dY
z>H4j>EjE0EV7oA%;K1^@zqe+Qo2=c!Zr?Fj>f=`=5U-sYANY&2|AtI&gwgP7=GKNm
zYSB#>&3&2SZ&oCn;ik2z82Z~FwOijkqmJ{r+xp$MysYgY8Oabw?N8rut+2j-HL>fu
zxa`bKz_Yw_=bK#`PN+27zu1_k&kS>>;}&cB-!FA!=PhVI+7b9?_n}qI5)-=GllT#}
z>W~2E$^d7hb7@`eXC6P5!+np<yKiuz@LFqZbuQm`w9=wDxF)|&=cwDjQc-m8+_Vw*
z<6aZ%fMhRZEn7iVEb_t;iM}NXdE;sOH>;^>W3q2EH^gl_0^5$rR#leTeDO+<+Jm2~
zj+d#cC{T-MFtgL=3mFN?Ekm^?1!*&DziKA*OnGr4dh_>>OwTS=8Bcuf3^}0nO_lM$
zpazPL3-dF@KoCtSKU;CYp^1n+06}a7Mx=y-{)@tmrY$<!2D{RUST_`tV3L$j0A~Er
zb1j8`@fH#5h+-IqC4~a;9zx=UDd&e7M65H4#b7Zxp#Xf?JN`w9{JBXYb{~rIFrF+F
zfKz*PYwdJ$S`f$!#bi;km{0)D+{MIr53KtUu?{H4!8l2w09;ZxL7Q3mR7S+QpjaFh
zml6uVy|sGfRyki-M64@{i7-)KC;*SLx*E;8+wKyv`%z4Q36eqqIEot>U4f<t`+w%K
z{EFHz2%D@=-VG)Op<)UD3lm|JG)RUZQgFl&ei$ig3-SbS1W}^4AV*;D&AXq`o&1i3
z4nR^Nse@;-y>E=Ig)Z~32pdPiDTAmGHcWw)L1_ou(+*3w*c9RrHlBi~f;JGg7)6W<
z-iNRW6ap2*fv|}bA{AtZun`JE1s@CUM<*0i2_^)YpthhwV4uZ5C`xUO`W+oyTt!?R
zv;`(bR|^^is{uxBL6g7+gP^vcL12SvQCrX)u%Rd}k0ZE<j`ELI7o4~Y<iE`WoIoOM
zbyoT-CD%h6Sa%P|KP-S3gtEQdUqnQ4FkVcYh>)ZtCG5zy)>>qO-6nZy83o1dW)=*~
z9XqWIW%bR~O?6a%<jN^2Q#aGpG&Hq!^|ox?_LGsx&nm_g1G<f@9FygV-R-n%55^w)
E7iB_?fdBvi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_abs_fp16.cpp.B14E3C722763D4B8.idx b/.cache/clangd/index/convnd_fwd_xdl_abs_fp16.cpp.B14E3C722763D4B8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..338128dad4dea23272da777baf09e94cb04ccc25
GIT binary patch
literal 1318
zcmWIYbaRtqWngel@vO*AElFfyU|<jg;^LB`%ot_{hHwrBhKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*VC9~a{DLTDBjS$SzLkn1Vc_}
zX1CCV&Z<4@zP(>>(7gNQ{G)FYi+*Q@-8t93nESz9;p%<6ZWzSJ?%&JcW-fl<vOssc
zK4-;_*>~&y9_U_Qf41xD^QU`etPy*x$n5&DaHCe(e6{#zr_X;s`efe3_k!}vFZV?}
z*_x}d-q!0&*6i(N?&}<ye?6FbzBM=M^~Nh9IVU79RZ32sw$sST-BbTV{m$iKOP$Lt
z+#}-BKXgQWo?dRA!nFQ`M`+`*Nn8Av6<=HWf7%_zE2q{>-pJoK`+G#>^@O4zH`WhJ
zwBw|?RijU_Jg80CvuRoL(b?wj?>h>WmF#j@pxUXXaWVMXhl%$sqRl*;?ypqHnh=tB
zrB=dot#V^yrl-^5!w2g2spmdXa(JDSe7OE$#(AmctPb0nf<Ttne>xw%v6$@^d+UMo
zW%-$DmNvpW%>*w*6ldgoxp36hI7+gKP4@d^o`25g4vEcmTFJ7x$+zV7j9VAEv!;BH
zsaPSB$+_|x*Vd%t-1CBY%w(co6&{{h=s8K@2ivw>E#?(lH^yAx%MLvC`BwFcfS1A#
zPE5XCq@o%*B}wv~OUnzM=b8&2e%r9+%EgBpl52BizstLs3+OCaUFGugPI~F@xZuS%
zC3YKca83^@oBKxi`9>v;<j{%DFA9Ctm8!&M9h2;RqgA)+*<{(qSz(`#lsxDwV7@&e
z&s!^Ke)uZM=z0J8Qds&YgdW$Z<+9oG?u7i&&q@|s%WhoLbTOH8=41W4={C3e0<OGN
z;hUKEFnPHKv)rGotusEFo35<ay=i~fNMoz0^fJF)3tcuo+?f&?WvbP)-e`m5mW9uL
z$n;dFElv*8GC#Mt{N=CyyrO-wX><AW*}|FYe0CN;zOwPY@r`u->XRA$YdSt%yscDv
z^#qe^!uO0HYk%H1te4V#zc5R^d+!z2Y5P{(dYGYb%j|TeT~WF#^M2lS?BSIwed1P!
zb>vj_OV9dj@jLj(BlpG~5%$(837f5NX*tTz>HQPqTlxC%T@}q33Hb@_D$&ObE>C%u
zdoxXR{xA1=`@=+IO#B-b-M4t4{&(y7s<4mmjwr`0`JH-pL0-J=(RlM&vuByE@cE|S
zyZ7tk;>z5lSYTc&XqKqV-tsSso5_!XA&8xgT@6eCWx$~0yyjL9jj0ED7(#&3Tr6B1
zU;-xnZk?6#`&&!HxS9Nc(yVN(LSO<W%}|t@Ry+g9gW1B!g<^W?s&n6}wwTOi<O$-^
z=CXxZR<ky1&Rns?t3Y9PO?G#fW|%b45RgM)mRwU`AJE~dw+g75)soc_<|G)+#D(G{
zpphQzAuuD0QgaeP<`fqt7b^p4=Acb`uNh`4hW}v-3}6U=S;4>%h+rn?rsSjpxn8M0
z%q*;I>>QljJiG$@d>(=>PHKW&9zx>6A`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*Z
jR8-Y9wRH3i42@0A<xHe>jhscr+&qJsTzwtg8NC?*#1I%s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_bf16.cpp.054A24EF5BFF4A1F.idx b/.cache/clangd/index/convnd_fwd_xdl_bf16.cpp.054A24EF5BFF4A1F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..26863ef07b85ccac67a6fbd97fce266a2aae2e5f
GIT binary patch
literal 2454
zcmZvc3piBy8^=#$8pq5T3}eO+6(iRn_ZT!7cP7_d2DwJIK}brQ*#A>%qeY(W(MZE)
zDI!bOHTiE_B(X@gyIdk_sgR2QnfdQ;*yq{vJm>qI_k7;>_j`Zu`<~<N;^>$w1wr;v
zE@`~zlyEErK_v7}PD$F=CjmiaC`;ov_&rW~Wrhr`&UBo)F!(g(cNxz980)9MUGzwu
zuMjxj9X+q+7kYGRn0jt($6UFLN27sUradG{&8M+&VfVbb`k73xoZ<16<*b<PZC}+(
ztli_=d*?5FN=zwV`r>%8rq`^;|HqqrZAbr@(~bHRnO!b(+l!wby;*j9@lIg-@(uoE
z!BB)1-Qi+JVAcH4aM!@mPqQ{9PYNciU)-DJN$Leee-l+WnV;Seo%2+1uj%FTi^7=B
zhdCcx`_3oy(FUusUERvh^m&a+&g_h@^n3X$lW(PGV4OA)H1K2Ua4e>}&)F|)Vq;th
zd$P)HmC<ZpaV}6ZeDfyLcTM%zCc2zy3LRzo`*?a}y0-lB1;Jc*5r<*(ZSx9pD&U)x
z@=TnDx$P;{x)XNoMmOlC*qc7KzFV%3c`e4%&7?QVVBN??Ox3%+i<feIv9bJWUfO|^
z>7$CiIZ-da`(7<z@~mvM)5zd-FmCg-O;pkNUlLsnr#f;>PyXpBQ~X@Mw9vMqsq7g4
zT!!G6+HuZ-N~|D@{q~_@Lz$+p?z_srJaq=;ysH(4Pa_W!T@4>4#?}ricb{?1hi~aN
z@VsfP>YU42sm&+c4hN;nX!tKzB);zFbrs~ZT}i}4Si6)H^XEK$(zAo6Unf2iwAbY>
z%arNDls~+94dlg>&27Vjv1Y%&MP`0A)^5pNAiZy8YefW>e)sy5(plCAwn>h$dEM2e
z>qQBv*eYcY`2p=icg<^wQ}sWoeLQ@*C~s`;LAGNb95{4j%IA{cAmds7&kl)kvuQpW
z)P7H++0{XVk#6ax3U7K+9_MB99W4o^FG&sO`^`-&Z!)@UijB1D6_m4bTLmv7?2}#U
z>#z7J?uiOBj<s`G>7sq?adfKijqfQvG_`d0nqsl^FMI!c_15j<`aaW7b{w2(L>5Nw
z#2Y<Hj`+{asO7E@QzL^z^LsuY8743p@7o3sdr3P<^qbJi0>gSwD;-k_mevmtAXk0W
z*=D4OZix)n;+D2aW&VKT=d#aVtWWx0%D|h{j?6|n*q6_Z72dEh)2h>sFYI8(sD}i*
zeBNICZme^uV_!fWv+7Oi^Ul^s`aSNEO1Gm67?ElH1It*ey$>ovw*-}3%Fq8)e7z*}
zNlI`P&G-H)b<|_^Sez{N{+srN3%C#6-<oh%Sw`2xr@q$dHxg45@Lcn}t{d;&z4*t9
zYg!CrPt}{mHQ(`7vItG-B0MMlkx@B6LbwR4`Wb3fcM2FyLxu<L>%Osn7Q?FHCteZw
z>(_YtSTbZRO#<&1cCFy77n5R>5Az}lHb4+LWXf;1b8Qv@697S6JPuD13Hs%WE+>5E
z5940Hoh4v)p%}p-6p@H=qtx?k21F%7o+zeJD5@d>I5s$axQv=JM!<NWn1CaYL;~>9
z>NtKeC47j0@j~$?+$Oq60M34(-QcWMcng8TP|U${V37b^ditvk*GA<C0po{aB92HB
z3BZ@za!KuDrtSnxAd2O2@^T^pxSw?ex9zcJ1p(uOVkw-ItVqPTv=cMUKzfOgH;QQ#
znyN?uetXSQ{oS9JcM~wXQB1~><wOGT(wk>_`V(Pl1dKn5WpT2SA^~`1+#}VBsd0;d
z@kOy5PL3oJfT5)5y~z&fKNs`{ga1cQ5H202%Y))6(iWZCx<8OMAY3Y&>gM!!M?mAy
zSYg+|+6Z?WyiF0r0%z1o*t<0+Jy=f`IJw7E4wGdcDx({4bT}-~6%X_L1PbHK4-yDh
z&O(k2&V+E~E#%pt_z0J5L1u$$g^o?SK*+%sVXA6WJFxqnG=4%@Tk9)yB`FmtI_M0-
z#jD`ypauw+s6wQJo`b<c6Bka^TH+e8hPW~*%PjN{RD(meQ&Cglfb*ofv3qm_%skQ8
zw4F4s{=Yj7@CN+94k7wlIDufN?<G3~*_}zpF$i}nvXuma5w0epDYihiXqLk91hEKL
z3)T{62}*@J35RseNe9*;gLSLYS~hy}N*&RhWu;{)V7@@9&{*LFuf?jv>T)0!91wL9
z&heTP17?Vu{NmMdRvxMLv9KsUv=9&jI)}yzhkPwo7uFRg2eN_j|Cn-wYXMtGfh-WN
zC2YCQV#3UI77J#rvut78b(S4$x6X2bUBnicIW#-rN3@omHmohKFZwm@4;P~3q{w8Z
z@F7Gf|G$F30|bSNA0@a(nFGo?3->tKppejDC|FFONM6+b7}Ru)c9+0z#7V+Z1cWFn
zBkfFL+ifG^o#m+VWJTHz3#OH|%}$oSg07{ynbziidrMK?q_SnJ8eKzETSw1eyP>hE
YxvGg0!^mEN;^?vm<G`_X+TaHL4;as1fdBvi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_bilinear_residual_fp16.cpp.C4A66C697A8DEBDA.idx b/.cache/clangd/index/convnd_fwd_xdl_bilinear_residual_fp16.cpp.C4A66C697A8DEBDA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f0d49e1faf60f0cd9a8147566c3e528ff6e93dd4
GIT binary patch
literal 7154
zcmc&&d032H8=o`Am}#EY$+S$`r&ZCSy(A$m5|u<fBbBtgtx~C^#j8Rpp`8>-DTQor
zn<d$j5+?L|WveU^dC}{8X5R12^SFH1_5I-wb6wq?-}5`qbIyJ4``q_^&a%Y|7Y?&Y
zr1}1f<H7=B)=)?!k|g#Q9TT;FHwL`KNu;=vODk5ns+mfc&Fwv0U_{Y+Sa)$sbA#K;
zZ7feK@1043m$N#{c#3wPN4~@d+9~f3W@I?J`IV1s)Ia5Fc-ipsYByO;X{N8|?*WnD
z0v7p?jt_|^$HvDbMqbX*e3GWu@m%vuSEv)$aA4q}s={cvWLeQZ-$&{lPhZ|2%ZZ+u
zyt;hs{=`xG+#(n8PUVPyHds7q%05`!b!Q^{>kGElEQ6M^lJV{9ea{rV)1h%vTduL2
z)w?sM2TqT<QgxIqbMb)$x~tj$UHs7<S5Fok@)#GHyc;5KJ<(Bb!aDZRFjjYayJ=-e
zqoi)5qs;Yf=7qaXJ+drMesiVXuPUt2WTNL_Tyw_zlMar&;Va}qa>&R%1<NqCG==;7
zOJ8-p>peWHd?mSGJIu{%zT>^>r)^OzeI;$n(8b3a`rJI7-Bi4d_PO=N<*BCDImka=
zknStH@7Mw<uO~jQ{z@}HJMD6HNb16EEd@P)Hw3C!*e@S>Wz9IQxcZRq&TiX1rvf@G
z&eiy^y}~zMFyKiY%~)h@a<TP_Wvu3TmZMVy=TyhuFr7|Pb-#-8SnVx;X{R&&9YI>Z
zjY;~&MBce&+q@0YUYv4ItubVo{677<`F5MU)pql0=W{b|1-iVIzB1UJ=Sivp7wYAX
ze_%!GcpkJ$*`s5jIPqYqdCupiv?SIHr$r|ZcD!ji|A%{Dzkcm1?^Du`4!uaY7vd{1
z^?3X$>7kWu`^6y(*(BqWO${~OpDc=9)m888nAbUek&`*~T=vTomFl9ARV8a=UUcuN
zwP%<=Tq$Q~+xvOv;FeC+1fPT+_b&SoN#}`kNA?bXEK9G=)7<B9`0c4hyA6z!4T{Si
zTTWkpTJJ#l61ifwrVO_(vgyn&#yE#7=Q9#_I^SULaShJ~@wL^3A=alIzg##kYBS?o
z+N8qa3w!>K+kC>PBOvSQj6+$Idyan@-C!#BkfB`hmAm(FjrXOD?Mur;vadZu*^}#{
z|F+fLlQpP&;kku}SKQ8slffB}7Z};db{*=98h3hKT6@3#;2%9p40o+Kn`>y9_s-ou
zGQVciJC~V3sh;8?-Cv)tR<4^X9yrvPKO%A?IOvJo)w`lyqq|?9sEmH&y*ec`AwFTZ
zT^UbeD(^LCOV={%hVCQBEaxx(!c7eBMykqMgBw~Tbnjjsedj)ZX=UK`W4uqw+dnLn
zFP`_+;v8pEIsL;zTJgMU+u`Ky_uuNSemlN-kf%wpq3wJd*wTOZ?Sp4gTe_<gtY0kJ
zdb?@ZYNO}}S07)Kuw!R!y5^9?LwrBF^j_+#YP$TPX8GRi`0`b9pIg3?xKp}gO5d0L
zQyt#-VHP@U%g$|%dnA8ADs@g5v#es=`R%%W>a)Y{jd-7D?lk%9#6=H%=a=VqdQ9F-
z_FoXyZlKZ9tW<fU(1f#lh5p42MFuU4>)G4G&#R`DY*rRSNEL@Zy|($iclGjDUx!PR
ztgpW6jQ@*sKgRo1VytUi!QN$lsWBQ${h#<)4eQ4;Z=}6KV{^v3+Vl4X{rOy!WxRAv
zdd(oW!|%bkcT2lnk(K!&%N?yGS?@;gpyH~|mtz@alcJkG6-GtJhxxgYNhE2XmmZ!D
z4XF(BY7z;RqsmDN1QxvnKa+_|{+w{d5C+*Bqt&Qt;sOEbMIu{vJ$1X7!yvn2G@Ht1
z2?V4Ux1FA}e5KcfkfJd<bZY2yfq-<->P>eK%Ip|okX<lZo+>XX5Ri^YlZ_8K@sBry
z?1s@?Dpx}wAe~qjTvEzfbB94*hS9UBv(*Ix(%ZVVPdaE7v?HV_jLsxeGRd_6{)1gA
zvL<a-*~~~_kiD=j3RDG2fq-<?A3G&mhUPdk$UYdYLDi5F2uL@1Hc!3nC`)CKeKA^<
zswyrJkiOAB`-o`D{#a4cQjFGN=_m>Wq<d^@s2T&B)eN#bM$1xVB?JP}_g`o@wuLMY
z7A38~Xc?A_ia<d6^)U;zH@qrO26-h$Yf-hN1On1;X|zjoIsE~mq~#c`$I??22uQ#G
zY!V(Y;JA)K_QYrvs*0FEK>E|jqg?%GTvY~n6-Fyjm8J;<r2l#95^H72X=ji<Fq%W<
zND2g`Nl^iT(dkSKqOJ388fl9I1pNl&zX3Jq4d9LQ65DqBmlO(`3ufhl*^r>1dB7kK
z7~(h|7~~UiA5iVXKQzUFaisjZ-*zg3CMqThzmK3v%1Ipf5d=-<B<tWP7Z~RPa~$UZ
z(>!2><9uL~59UKmZJ6-dEFMz(({rWXY#kZZX%L|xXgW|z$CXCV44|9=4B&_ex)pG?
z0uvl(0p%>5mpG#sW>nzNjVYZD80kO}@0bCQ44{PLtw3@sP{(m5U}O^eX8|OO*gqRc
zX5;-OJ#Gg@R3Z(S{nQNA%wQyd0fQ{^U7$cvGnyF_?nzeamb|qouU7#I+6SEbfD4?2
zOh?L|rbqiGX%sY3Hc=iv#*liV29A>ClGp^Kk)(yAWVvJlm#mwthjRl!Yye0QvBm%p
zzAGy?vFMMh<C2(0y7szGaEHqGUWz0QTLt`*Ik8RNY%J#F6#lSF6gTd-?F8!2FrpHe
zR|0FOm6F&N^Gw?H#~5(~FmC{5NT{5R3K~tZ+KMgu4Ip*{NJ2t==uD@Gv9J3NW6Zo=
z?L_mdpZSBCX1N|5W<~1O{_qC$f+-^3<pBhpBQ=K&t&gB4G!sdPEELDChEUgxO~Z^V
zmPjy8l2{S}C5t5!5D$p)h=$|=4IXYt3d#ePJbX#`*oz0ep~7=Ndz{IL6wAY={}^aJ
z24;}33$NBwoO|Zyx3hiz_zSlv<|Awb6+;y>7*Q9V&v+d=8dy$7&|X0A1>$gE1ib_3
zcW?|7C$@|DSrZ0N1ht@9$ikI_F9VV9(xzi!YtGHIQBiD=ImjI0IwPnhVktx}rg`Bk
zz3scOP8@TNJq+3is$s5S4>Ozl17FAfBImh39o>9xo5ic?@mSb$gzJS1W0!BtJ;N8y
zYtEDTVL-7Dk2Mr*Iab^OUQk2dq34T9$cA&+`?{*Sc5r9H0s&hkeszGak~z&>7;A!F
zH`p0}dHDg;^cavnh6gYNeFv=H;jw3ZRCKKCr^B{W5G?Oda)Bru9l`PrC69>n0VN;D
zq06dx%43xFpI#<d(O?GUmoreG$dmV7J}Rb_{`B`_3QAX8ebaF2hy5J$+4e-GYCRQO
zUjtQxCD6e()ZF}-aKgA0dz?wu%_LjHqAAwN{<^N0$x@8ao2%yr>xsAvkB1T{JvSr}
zv=Gn>ff&AmMSxyJ#KnMKOvELCUP8pBfL=<(Wq@8r#N~iqPQ?2GeLoQ&0Q3VyTmk46
zMBE6(8-X-Fzg8gH3Z#fwwiV34aT^eABl2xPwvEUS0?|R7hYN@~Ccp56hD1=VHg^%U
za%!D(m`7<(CzkRXflVXshzQyYSj~V?qMQaYr-3m4A?R5ka~8*G%l6v5e>yE1^YA(#
zT?Yu&%owm91IwWrTThJn4_$Dy`oH___qPia%)7VCV^fpWlGTUBD1ypq$?3xkkDzQV
zwmyth2&$l^pbtxC=qQ*c|65waCCGN1lKv{^o0%Ab%81EKhjlE1Zj{<62j|ImA7&vC
zrhf!20?Z;JE(XkEA}#^U5+W`I%u*sQ1I#iaE(gqVBHj;}`-%7fU>+di3c#!&;#Q#A
z3bgV0wgJ^P92ZP*eNx30?qJ)a2k7<yQ`oWCdb&)+vYFNY*8pX9gI0=ePNKi0=72`T
zk~sg+15Jk|5Gl(U(6tfNie@DYvK4QpYYz>pea0N704NsVH8_Iq0m^%bzPuM`?gca9
ziXiAD(3`|-_{wcwg-<uMC}6&v36wI4b#OM|WaD)&%u(1D;~%~7N!rkCq~RoM<IYW8
z8dk9oGlY_u5(}PO>b0MFtQvEhIfY*s$n$^#oDClv@PH8!JMw@t#K%|PGpyJyE`_O?
zNKYgrh$MOv0VUIu35W;iJUm&zV<9$g{+SS;w=vCFn4Qmvz3ko^5lq77{r%bKtQhUq
zWv?NZU*Z?%c8HgyNbG9C667JkK14KXHDFg0aSc$Y!2?v=@7$_@7wV6H$*kM^tBgoD
zTh{+{_Pi){tE%qd`3IP5)&uE!qPp*a^?R@cddU@*X6<pEi+<QT9R}>f#EwS*`v?&q
z1qw&;b?QoYnzp1*Dh@MyA7J(Y;W_47)DgEYPr{;q$pT)Nu)q67eAMEX3=T^@Y`gI9
zNbv1qOS6@L+x({5rfsP<%f<QIa%SReXb0?|ub(8mEx-=?M#;iRjGzNRe*lj=FvVfx
z@Go2NaZG8ZQgECehn#EKjn~*e?Ml5)^r(GSenXg{RSi|mU`*&e`{d)Z%~TTRVYNWI
z7C+D+=v!d@7Pvvp?wqVvo0C<_KjUecYt4h#J^hO3rry_8zWzhlfN2)fj(=B#YiCF^
z6vpsBpQPC4N;V8&Z|g1B^MOU(W8auR&Fpd+Kdtmg{;`jxy4~ed5j2hxCjxar(0EEb
z0VPlp2q*_oasUAj0_GqP779>tOfUW=7p@oAf$I+IET{<9iGS6FcbZ3=Cv+irLc==o
zFTC(hMl>T~CzwBIBH!J3A*c(@g$|trL6^{$5Rfa)m4Mu7?gZpP^B^EknkNDI(0qhw
zvhAiukyTs?7EvM;A_xv1sSrs(Q3_E)1W$CBLi{^8Tp??kwb0~I0ikR7Omx)R=zIJ|
zC|~^hTMru~(ptPJ=>9L{;!oG%%|N^vj5ovaW`M9c1x;D=%5<Z1ai3J!Z{ZPZ*82Me
zt_uzcT^|$>#3lI&`;j8>ra#^c!kZy@^EbR1CTy+^^AF|E(J{bz3S}yF8ja3CL?y(<
z93<`SG$chFq-2=VvT|nTmR8m_b8Ypf>sqLpXes@#r!2O@3}qEnb&jUC&dgb}4UFcP
YDjLh{8P1>1TDUlpykN<^MZk&lADZp?{{R30

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_clippedrelu_fp16.cpp.4759FB8DCFC58E7D.idx b/.cache/clangd/index/convnd_fwd_xdl_clippedrelu_fp16.cpp.4759FB8DCFC58E7D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bbb1713a154f91b4d7591f998a68fdfa3ba112a7
GIT binary patch
literal 1324
zcmWIYbaPW-Wngel@vO*AElFfyU|<jg;^LB`%tU4ehH?%DhKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(^Kt07_Bb@eGVJ>Ddbhwzt!c)duRd~j+OqGqsAe>3;5iUF
zwa}?iA+rDa)qj7do_LV=E8Vbq#+|xZqT4?|e8E)l{m8uadS++BujlJGTs+JA!Q{xJ
zhu%zb@6xu%+kbfUI(~D~*Gfx$wbe%}PB;YC#H{JMy8PtypPxS08tzPgQhnt3%Pq=c
zm2aiaWb04)P&IpdUV8|K_L+m(^S!TcIoEwDVWZ=wj88^ci;HrXc5=l(_+R9|*juQ?
z!aX7`{ewo&L;d%87uez!hVEig_A;@X>C$(pzPp5D>$cY^Z<q@|%W1p!didRX#UST<
zH2d{}6z|&%d-&I0KUcP)=JT_^)*grY{KPr7tZ3nl`Zi^A%@gx;eZC<N%B@aFiG(T5
zvg<iCEuibb7N<;>&I5J()MKA0Ie4!te7L`0M|!WL`3X6@fGGzj+6T^&O~3rYTe>FW
zm-1zu_H?D|W=$K;$w`UZmK2`vxhS?!@W@|_=Kl(oH@Gf2Wu9D<5`4jGS=pB+skgS`
za$3D^j;pqcxov*TdR~)FoKN@Lj;CjKsCWw0Fz+^-%J}7$m0m{sR;QJpZ&j~oF;j22
zpqy(tDNM&}Gmn*^;-sfGq9%9brcJZ5k||rWUbgUmVhQs>$6c%3r<|`1+y36v^j2@F
zHMg~+S?cwZc7<^v9I0!xR663;d~|U+?{wLMd-bla`Kx}WxF1+%YGWw7!+B$4uE%y&
z(Us@5ulDHn*Sl|-<bL91@)SE(xhSho?jQGXEV-_7%c`|w+GNXr=l=$|&ki_YHTQ_(
zCEc1gBB=+O<JXEA|68W^b-n8E<<V?Q3yTh4@id>}lvkk_IqhOE-{aLhF>JXjEbZ7I
z&)ey7M$7lHO_kU6`jW|?-)?waQ(?SA>O=CWxzfw~{?0pSSAEmFt@>b>f&5<GpzyU?
zEj(}JboZP7w|>UI_^EBk+mMan%O<#O|C1pwD<b4=fysp2*-vfMEiG>z$ZvjqFf=Y}
z_Vry;jb`8L<~>)r=I_fHkJ1@;R_rf&aUivP$5kQw3Ee(^H=jPtn>0m#4S!R+dh{^^
z;~CEiZl;N@{-Jz6ezldzTj2#S%=vz-tbe<CUg;w1j~wDJ?^#-}xE=2Q@j83M?75~+
zx8KBThB4PM6j$aZ#R79&L9;|<_LhH9+)RE93_<K{>}p^FC<6u^=QX!_XiPoG!w>?L
z=3?RE0246jck8T_-``pq#?9mplxAgP6#^45X@;WIwBi{+p1E9S!7kPp>lk^0SS?vy
zVA^1oGjgF=TDt1ox2i2BbAhV4w7G0yZm3zCHD|6^;#Ht9yC%CkOfyUxXb8xCFuP%P
zFma(c38>eDJp`t=C^aV$q@lPdxmX!UGY8GBSu}U+-33KVfdLEwFtZpK0ujvQ+?1Sj
zAlEC^hna<yjh%y&n}=6`pU*?k#Ys(&%R@+9SVTh7!qVEt*3RBRPgL7V*-S$oo2Zzy
ojI4s9l8UOjrk0MrfuXUfxtxiVu935-n44!XldG?zJEJ!P08+OO!2kdN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_elu_fp16.cpp.922AD9048F6922A8.idx b/.cache/clangd/index/convnd_fwd_xdl_elu_fp16.cpp.922AD9048F6922A8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ec61376da5fb72c1f3d225aedbe9dd8e3114d2b9
GIT binary patch
literal 1318
zcmWIYbaRtqWngel@vO*AElFfyU|<jg;^LB`%ot_{hHwrBhKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*VC9~a{DLTDBjS$SzLkn1Vc_}
zX1CCV&Z<4@zP(>>(7gNQ{G)FYi+*Q@-8t93nESz9;p%<6ZWzSJ?%&JcW-fl<vOssc
zK4-;_*>~&y9_U_Qf41xD^QU`etPy*x$n5&DaHCe(e6{#zr_X;s`efe3_k!}vFZV?}
z*_x}d-q!0&*6i(N?&}<ye?6FbzBM=M^~Nh9IVU79RZ32sw$sST-BbTV{m$iKOP$Lt
z+#}-BKXgQWo?dRA!nFQ`M`+`*Nn8Av6<=HWf7%_zE2q{>-pJoK`+G#>^@O4zH`WhJ
zwBw|?RijU_Jg80CvuRoL(b?wj?>h>WmF#j@pxUXXaWVMXhl%$sqRl*;?ypqHnh=tB
zrB=dot#V^yrl-^5!w2g2spmdXa(JDSe7OE$#(AmctPb0nf<Ttne>xw%v6$@^d+UMo
zW%-$DmNvpW%>*w*6ldgoxp36hI7+gKP4@d^o`25g4vEcmTFJ7x$+zV7j9VAEv!;BH
zsaPSB$+_|x*Vd%t-1CBY%w(co6&{{h=s8K@2ivw>E#?(lH^yAx%MLvC`BwFcfS1A#
zPE5XCq@o%*B}wv~OUnzM=b8&2e%r9+%EgBpl52BizstLs3+OCaUFGugPI~F@xZuS%
zC3YKca83^@oBKxi`9>v;<j{%DFA9Ctm8!&M9h2;RqgA)+*<{(qSz(`#lsxDwV7@&e
z&s!^Ke)uZM=z0J8Qds&YgdW$Z<+9oG?u7i&&q@|s%WhoLbTOH8=41W4={C3e0<OGN
z;hUKEFnPHKv)rGotusEFo35<ay=i~fNMoz0^fJF)3tcuo+?f&?WvbP)-e`m5mW9uL
z$n;dFElv*8GC#Mt{N=CyyrO-wX><AW*}|FYe0CN;zOwPY@r`u->XRA$YdSt%yscDv
z^#qe^!uO0HYk%H1te4V#zc5R^d+!z2mk}Xv3rqxZXFs)(-?@2F<9)X4&8zo>sI3px
zb}`#~ocC<yoWGhi9~V2steG!swBe>~?$ibR=iKVnE!*?;VcsMWy&nD(&XaVLXJmSP
z-j=hu%m1(P`S{gbdb89WF6PU8nEdZny6x(U@<)Ptm;P=_3)sGX-lO%>XMN8~g{b`&
ze|+_ua&cvDQY<j96*Nm!W^efy#m(f$z!1dF#;yh?fHGjvab9z)hsM-{JPaW~X)YEn
z4ln_eez(p_`Teb>VcbmqKxtMsRv|C}lV&JNO)H)O<YkD>`!l2K@?}PzAXZCO8<;kj
z<&0b?mX@wM_pNG+$y}gnE^RJbm>X)=X3d!^mUtB?%&y7q4$};i1{wl#AIxr;9ZXy(
zP6F!nU=M-mElSNv1ZgNPN-kCg(#%009*NG~`(9@eQ(yo?0L&}~hCl=}IX5LI9mw@c
z^<idVWn<^y<mTZO;OFxYba7G><nj;_7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-
lsHCE*uBoM?Z(wL_YA$CYrEBCYD(2=H%;f6p=+5ZP004=F8kPV6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_fp16.cpp.E044D0CD44CCFE7F.idx b/.cache/clangd/index/convnd_fwd_xdl_fp16.cpp.E044D0CD44CCFE7F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4986ae43f4974898209c5bc9843cb262c56594b2
GIT binary patch
literal 2476
zcmZvc2~bn#7RPS_gq!3>2*j`kgs>$sQ7Jo!0Rah{fDn)iD4QZeEdh}ou|B1;L=e%k
zNXyc}1=(bTQV^Gi3KgLUSXrI|B9013c`jAq-K6sY&b&L5GvB$tbH4AK@BDAP1x`*?
zL<rg*E=Y`th!2s4Ac%_n;^IXI`eY!e5@U&_&Aua|Ni$?{`E_Sb$)mCOE(JmK{;gwr
z7u*vTYJL{nn>eq#J2>^_uu9>S#oJ#L+^=s^%G?ggs}#ubaeFVjpuEf!FR|a?>_U#p
z=GKpurQ>^oGRlgEo1V=$y#1JKgO*00=k{h;dgbsuXorG+pDS;iNH=a@N_Wkk6Ay^q
zz6lWqbEhYDSltUfsb0K};H?j|I>yfW9jQuE_#lcKz{Lqi2)h|AQ+6@uo6tpl^T4_9
zNp6q3{_H%b-ea2leKGgU9|bICe8C6YCovk;yf`{1*3#-FCr$?poJ%l|)UEAv_RXAj
zeWa5;j+QTLEPnFmG<DzNB}a|NMWo$s$~R`)^Mm@I`1==Bxye2`X<&5Fs`<m>%Y|>k
zLu>ngAZWULTcNZyobs>Kh-iTWG30VZja}%!&&-ZF$7{)nWnF2-DO&e=i$A&h$bL6&
zX4cVin-QCH$*vLZH>B1o$Y(WHMynrYZ_jeqHhw7joH^C4_q3}xEo|4(-Z^T2_Xd;E
zUEA`yZ}g#$zpHEUIp%4n`~Gs_NWXHF5A2gzLyID|4xKvac*4-4S~X_W`t^TL7*`gA
z+t0hi1o32Q&Nle;dAw*Tyn7h-jICB6iE&>}*suguwElVd4F-pdb`mMgCu^SPtvkt3
zk9u`+dGoJFBXzU~g|4v~H6tnR?)ze14|Nu$oN0O$sa@a6IUaF|y?7eG=NZm9Giy3_
z+9_fEHlM#9<Yna!m~1H%Wv5a~&hK>Pd6)ZLRTJQr-6CakPy5!Mvz1d1_|F74DZrsl
zc%jXz&G|&Zf#tjD_l587IhHZ=vr|ds)=qMLPyMn#7Iw4>zQhRLeJs4`jW!s}bZQg@
zUOB+~p-7l}YO)C*jpGeXm6v)wj9xhI^8D^bxpMLH%(4ELn>l8`sg_b^jH`Qd<L;(&
z_WGs1Hn`0&bh_8e>lyWB_C#@(o9m9W*H1Q%3W9{+TByS>NKQ?izpbZywzAMJZKyWB
zBJ{|baLFQyKeLX^PUqFdI)!Pjo9%h8>7BX6>g%MNi%ls%c2)X4`FqBbraWqtd41gA
z*Otb2>Bb}7t;GC+dfN5JCT}`glN!oOE_T_RQn;Gv{MK>jE9ShvO=WUsWZH4I%k+`1
z^yJvTVjK2(wk>bEQ?m5s(FWTehtuAeQ)bS7iY#5q3=e&<@F^ng03+cL32nU-K9cl{
zk;e_5-<aw(hX8fTSQuNZxhLU%b|2-BjEf8R$rp%PJ}h?aL<hHZkQK*0s1f?-etuo)
ze0M*`Z>WY<RwU0Tu6Wi>%a9o+i++kriU~b~gCJVaOJAXL{V_7$AA(R4fkc%W?8}!v
zJ4l(QpL^XAlkqz-j1UmI)BqeAm@-_ck~>AlyJMJ4AXB9V;MCfv(lUC;AQ|t4VNHUj
zp40%G^^n!<Y;^V}0`0}HhpY!IH2_zf`eKXPGLp%7Ukp<S6spt!T-}~W?U*ukBja~t
zSec-#BsBov>V&Ud(GM}DK;9TuqpN924Z!{UO9X=v!x}Pv7lw%hqN3CQJbTrOIrpSm
zNX82>Oe4^gqz2&kH>^w4x4g@yK%N+;)9G4L1Mr7e<N3Vly}D$)ABGhPit<te@X~Yl
z1RHDpn`FEXhLs3PRH*?N5=HEb3&j3&!EErPuV#Xvda#}{xLZ}V_dD9V7pXV|RdG;p
z1(lh!TUHJhbJDQaG)tNz=m*Rb+fp)2a3q2nzy_*dOSk<OhiQsK8d$Ce+k+3r<ZfPY
zh|bM9jYm)=OC<-;1%fJDDm#E_1W&yXK*-@XNjbVI_MjXf6zp7yN><O+hxJ!liT|5_
ztxC6!faNlD86KdV1XqXOIsRsz*lV?IYF=Ql0OLqsiR$FMqo(D<9CcX)B{GOiFlPiM
zF-S~sqX>#H5GJ?`1f?)2OdvPVF1D{klAzRV`OcrSZi$X!-RmGaRB$?i8X|@(9@qnG
zQ=(Upi=alZ(Mp@(s<0r5YFC5UFq;PUy^`1_=NVJsgf*?WPLU221)9flC6Zpv<-%M=
zkPCXof+WIT4PwGfB@py<@+d!_T0bJW)m>^qU}GQ@ELWoX)m#qDSt$-o6$_GRe>G?$
zym6%pP%=#VO8f|F30o3D6$oktTdnb|Ve2&>ALg&|>|nb!o;_^8#uLDT6%Pm#t4{K5
ztyaf^St~;jMMQ^4NSr7v&RVhvl9+#<0I-0d;FYBe%AD|TlV_3bWdnkO1EIhb0)@qd
zNAJg6mk2i*Svi6{Oe7-|MTK?FR0n$lD#=+%MVY3mW?^Y<v(<JRpSPZ4#WXY0{(o=j
q8k&p^I=XuLhAj5RO`ErTV`{EtqRutmzMk$R*n@ZUu-k!ih5iRMA7Rk|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_fp32.cpp.911AF34195327474.idx b/.cache/clangd/index/convnd_fwd_xdl_fp32.cpp.911AF34195327474.idx
new file mode 100755
index 0000000000000000000000000000000000000000..78aa926495a40ef8d84f4d9c64005c65a073925b
GIT binary patch
literal 2402
zcmZvc3piAHAIFbj#xZkdni*plLhiT0(vTt7m@&BxN*LreMVZ=M3NI~cCE0S>)>}&{
z*LpF8rJbeN)MIH~+SW$f+Ks3bZ5va)<vla+^T$5Vp68kG^Z%XC?|0_+|Np+{1o1eW
za1{vJ7SBr;2~%T~APAzMzm(ME#1TaZszzCQMaRy`AT2hza)mcwn=AZ2<W+QwUxe08
zHSbprMn29Lto9o=S@q11*V}ErJ=L;7??E{uO*d`%>8YB$);t@*&Ix?~@$xq_*CwJ6
zuio3^pOu$MM(>yNGVGk$u0dIuMgg;2t-9Cu&TtD~w;A;BS8tQtY3$MpXzuxS<d4bk
zYbrkLzW7!3+xb6_(}Zs`?5$^abI#m*FYcU~`-*i?;i6dV6n%UDBqTn@*PlA?BvIIn
zH~7mcdTPx6M_lU&zxk%tNLC5o;xy-3&6M`iLg!Cii+u|uotFWGCBybQ>lX)ZdWPo)
zhT2%ttG3a*J6%@Br;fC8+J^e<Gxt3T7G@h<qi7Z}bD9%GOhNH+W7`R3<0`w#@T1aC
zU!4@jHjS3#v-y3VUpmJVKhKIu;wkm_;i?<Y<!v|oRT4P0)tjOcvZ-THzcg~r{p$vJ
zC}Z*$={)J1TQ;$Z@s!H}w&A)<WFhwfE1lU!{@Nw}!i-dLpqA&WvH$QFez|E!HxzIb
z8z~pw9pZl^Ji6AgXiD4eVRW+3I`{xLprM!Cu-ND^QaBS{J>s3ub*R-4J!H?l$+xX8
zjrW-UeB#>(T)t1-m8!}gRccAk$)QI41vMRv)GourtEM<X>P#-%@@0K2vExL;%;^gi
zsR<2D^&i__Z>C!gay#w^J=A@+6cSt(#*o+~8GJYJAgz5ll&sd`b=7>4iETaZT#TXI
zeatL+X5J%y@*i!AW_@C;;|%UrVN;xo<#q4usOx(Su~L)V5|#&DV4aO;wfFQ3{%kMM
zDUa%Mlz8WtCa#<~dL?9OQ+%aC$*C;-MMn;{x>x+9O(i>}?A>JP{A;r~f@r3&%K6|o
z<YccDUOUTP!}<ia4V%(xyfR&sxu?uz$E6$7X40(x4i7t2lG>(C_8$8^$13>QL*dI`
zE@CI?3zC6oivr5x?_TF5`*wthehPS>d5_@A9BdjB&iK2MYErQEVN;KO$p2vDppCr6
z5XGasu%eocr%ZKHW}6T8<Lg+rV-|X{2KChDZeWN>hd%J{6(>5R*EVe+#oOfDZSQ(z
zKiYq`<H+NuoV&gSdHJ<nxpj%lcZb?u4EeIA1HU`cpItbd8Lpdi;t!QY#^pB(S542|
za9DZKfA>!RdG{9|Z;-|BmIcS8@li`1zp<)zz3e{pPPueYZFr&lO>Nppw^`Ox>MqeN
zJE-Dlq?N=b;bc2Y^V>V0oT&&+9T*%Tj)`}?ox=_*EnlSt#y#vX>U0z~3q{+Q?~%E)
zWjfm3$wK1!vBZp#OJ<t6by89<H~HTw8KT%S1qh-=KHVAO-I7PZghG%2kHb^sg1-6k
zi;K@Ko(a5lhJXn`F@i(rasfCY;$P#{>R(I~Fn%Z|;0P4C0G!pducDG3bDw|-MDa%4
zMnky(Trh0b;ca@h7lEQt?62ex%LU-eg&$l5F4~y{%uW;&aYTw-0IuymP3f7o^Ce)y
zP^^kmrOE~1PS@w8l?ne&BE&~Aole)03&5jpS8>LZCJh8kFp5=hDrC6;{PTfpbxuf&
z9uW#cF@w&~kqf{tn;i`nCTc?nm=F}xa5So10G2*~RATicTAzT~g<>*}tSlFRKhF3a
z-s)u3OTg?vF%?Ip$OT|1S-2;Kihgs!Y%uu0%>)q`!iK8gY&96m-0mO#q9`B&bx(C4
zux&6nv{p7X@N7h23>#~JTHn1_#%biAwa{9BbALB555Fy?2kDkY#fpf4%BFgPb`gOp
zTh$XBJ0hU5X`bL9qNZng5RxX5b)m2B0lEPGK)aBcW33A#*l2YJsqfs{YxKo9v{qYR
z+aEM{*vE6oBGjIb-e$Nm0>MN8R>1$Gm(bfXBY}T9rF`>JL0@tX1`+5XdK6HM2uu)@
z)d(03?Na74P>TpmVbj$v!M@QXnb+2m%wcmHcy4351j`p)=Ab>3Nn|?k0XPn{R%XPt
zS|-dSgIeH7G)ZR1wIl=BfC`epDWFL*Q?4ah!j`K~ntGh$Rzhi+l$~5KBNEhrS)sKu
z$F9{{z!s}J04>A#zgZR$uwk|e=mHTq!j9`AC)jCS<OaK~i`-%Nb&&_`u`c4lywwP}
z4DC+#Myz#b2Ai!8Az7FdBO@uvaVbu+MWjsre}#huq$r?^vwW|QWtS_3M@B_JQL6-s
zSX;)4;*<8G;dY^~q7oLT466_jB6$PJo8swVOu>6o)m3R43<tK;R%aJiH!C#@M+19P
u-G91E*V?GPNl)L<$i&Q?Y5j@KX1gsqEKN(>ZEAE5Z#Txv-<_-A1N{#!Mo>rq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_fp64.cpp.3D0123C922DFAFA2.idx b/.cache/clangd/index/convnd_fwd_xdl_fp64.cpp.3D0123C922DFAFA2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..97c2ae2bcbfb6203d1fe200b18c664c98bb44fde
GIT binary patch
literal 2402
zcmZvc3pAAZ8^>Qm<29Gej4@*f>wcN&vaaLO7z|=2;}$bWMWuAqu-0W1yWOU+vZQ~9
z!cfU=cf;B&e=S;~n6}tbavL4X+BTIx7XNo<&kuXfzUR#M{GRvod*1i?J<s=@03SA+
zuLeQx(LTw?V-okPKoCSkehG=f*nVXQk{~R(s3CYrIBth~wmQ@1ckSt`L{oKc++mkj
z9oIQY?{1g+bd6m#42u*^4N~(b9R9qi&iQ^DDbpQNrRHs7Dn(W87FcJd@%Gi-dG(C8
zElRZH<(Q)NN>rS0|I6H{cx+B5<(|Xy?GfRB3Cpx1ubU6wyXxW`)7hRZoLkPJyw*39
zjSY$TIh4E82?6=C!slh-B3&2Go7b~FQkU3a<<^(ONo`rr8r?!vtyBDUCTA$=-T1+!
z5|l>cNxQcd4}{snC5L*->!TmS#m1-FR>rUJtz71pB<B}AmHNVuEk2n=bILWfD=qs)
zVn(zyZF=tXm#6fP#e{_Y>bd7pT5WZFvRwe%6TVI7>tKEJ^2N-hew7%`Ey@koi1c%Q
z)wT?2hI=<BsO)@oX>MJXmDt=ME|_fVU7h@-MMwNtGIBXe%u47pdVW@~UVIUHJoMlJ
z_46j99G<N=#bD96?87L;OUma+xH0gzQ|^;tIlFSh|D*CU^lp3{rhmc3kWV{vF73CU
zc?;B3tj?@v;ckMZ_r$3Jp;Ia=8<UAqFJM_24DwIbbiGx%MV4fzoo^ZTKNOVb2S--c
z^u$07JFj3~Ol>KA*R)y5vi7~(=#G-txn?Hfd_(OOsAB>zvnVb7NN9K|otxW?Z}{-^
zv{odmKKHAmRTWLUp8no))AX)g(y^$XmhR`D)vJe&X#TV_STof-v+gZa&h(gbNHA)?
zsDEeSkXOfp*lSn0KX|sD<MiG#p;q?M2eZmEPMg|%!S1><;ClmwuZjQjinCSOUV2|d
zafGpZcm3+Y|3=f(CdwXP?(%d!+ovl*9eG>T8OM74vV5h{cuDjys<I|jBXI=pYdv67
z=;2$%?wr1hic_*4n7C2I>x+ALX1A;^4CAW5I%*>MWq-)bU&0f1B|Z6!&x@Tt3Y411
z4%*XZ9d@vil7c5|wZlRL3w7x&(`R)HW!&Gs6Sntco~Vto!S)KIZFOH>x=#~WwjY(9
zF=w-~QYQBNknsJyC_7Ml{KScUt#5KM@AqV)8=KOrW4mnhTAmDeoiEcGXRG}<XD96}
zXiy)Y3VG~zq3B#iQ~S->`A5B%e(&|OndB8UAGkht>H2`z$hXPt4=Cv(9Gu_#N7nND
zvB~iRk4lnO_g%1GUaE8x?W|jR5wG1(pO-o;D|dHXlg;-mK4HuZlN*KN?#4MkUeT$D
z+Iu^Wi0f^%@BJL4=h!-vM=jMUUk#9bO?B3Az8w(6?@Sj6s{$wa19lzX^<_jljuuGv
z)>iWZ8CL2H8_)Ivixu>|R9Kyma{R#8N)SYjm<kT@{^koTDineQ7&L~c5ai2OT-2cs
zx@7OilXxfq!88g@Payz%VKQ@NypFH1s67bAp>Y(20DL4oZBRnZp1`6w2*#qZM1=q>
zsySLzOxgboi{c@83wn#OLIA$dXVKtoUT_}=MIo50!i5zAaM{IWSAnZ;Di#%tU_2U6
zR0zNoEjh&230prbDh$CIXbqA=0Iu$Shu!|8aVj42M=*s#(NhS(Qsy1B>5$oNEGiJe
zYG^frLI9qvWf=YOq9O!~3PCU#O(rP>;QvhzU$&l%GQ^_xBA9?Cs44{D6&WYV*~#QS
z78QhG5}HI*2*8jq=1>9&`OgKjL1F$e6P&;pHr4>m0~10v$;Sg8f)iN5Rs^s~i#G4o
z^6))T2`8X>P<_GS!T69|`RvwnO<_|lkn4B2a*#|I&_Qy!mRu&73dcS#o??}itBezn
z97!JFOgMptqlO1)K2AV(Bzu5vMNYi&fpDKS%d0S?vOpEUB}f%=i>y~+0-LNIGjWkw
zUv5~3Msjrxb-AFpBwvqSdZ?X0a+|h`#skd&SPAp5jzVtB?F8O*Nwux?!cW2s6i#4(
zGa!OwoWKlcwiW@qAyvw~26Ax%bJ%>XN^ovuliYFZn=D~VGI(xva`Ptt<7I55X2NCy
z1(*f21Id-!aXoh{ymf7mz?aA-xhdB-8No&*@Pd)?4CZCxH$(EK1kxfvQ@|LIT)AV{
zbLlXBtvFB(jQNLUaRNu!Q4PERCt$#g4UrS<v>{@`%ngwn?6x6d!K@9D5A3rR0Rth`
z$-jv8>MUT3wf+iY;`Yl)g79F1lY9{&m%lGQSU^zZ+EQ82ruIp<-n37w3O^z;9133}
z(81%;afgv`cZ{F1$|kfbtcJzm3F@1@i5@IdBF39U)gWup92}jTU0ioDtu^TkBRg~b
uf4fW3*`oW2fuXU9nT6%nZQFNzYHP1&qiwa*U6aE0*@yDvx_K%2LjMOCx>u(F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_int8.cpp.1BF1AFB682499568.idx b/.cache/clangd/index/convnd_fwd_xdl_int8.cpp.1BF1AFB682499568.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6ca16ad7bbb4033e92b7397c9b3fe02f4847dbda
GIT binary patch
literal 2490
zcmZvc3piBi8^DjjIA)G*hA|opxf_?US(nC$aT_Do6qO;l#k$3e6{0OI#jF|>UD$uF
zD{{%wY7yG@U&&Jcls3C<n<iPU`z9p+GxI#3vCp&fJoBFK{eJIzzW4SWe-AdhRS|+*
zB0bU)qEf=-AP6F&e|`$@^C4LXszO;>S$n`JZ_EaHy!N)&>!Q`qsfL?wJl*GOn-^VN
zZXX+?-1=!P=bT#T^560X44<wEmq&jITwf4+`*I>>{Gg6{OP^Tb{>N&9*!d+6If=jO
z&bZmrwQNUkHoOz9kUwub7n#};X31O<v653WE~ia~XRMu7U9=r*nEt-&{l{(e=-;&r
zMWXnO<bnLw50>TUzH2V^Ek1eZut-FWdoFu3ICQah_i&TbPK7H`m(Rugrk+J>v>IhO
z8IQEl@80O+<vp(?&bzYlt^N81+AV7xH5E5Xk9IE3B>DG#XO-L2)#;>>o)frBo;~>R
zE3=I0%#Y0v4n=sU7Up;_{Mt3%zr;SM*Gkg4d?hV`I<QYy<M?8Q(yRF9!2dX8Ih^#m
z`kCDWmLTXEC-zvGNZ9-}RZypMl~%8A6=->(ui(E=s)`E33i>$~dZRfqDR}~@@o<5_
zVk+rJO+D+oWS!;e%K0#b<kT{6my?!XJw<fp&F*M(=<%4(%YU<>%fDHjBD#<(3cp#o
zw&!Ec^S&J!MHfj``EC<xwC^Lf4Ew>sX?1WxZg5;t?$a4)oKz(kB`b7`20HXF5W4(3
zZAxWg!l-I(Bklf^t9EjA`WMP?MX-9peXB2!wkq|=hR>DE4?jpcaz1G4&1~|S+-K=i
zj%DS}DZ(?c--<dMhyUuXHSBClOb9dY{`r@`+Yav)s@}2>P|e^t(UTOadtK+Zc{Z%%
zPuDLUbiZ9|ZNIgfSHM$jIPdF8|DSMA6U9?z%_~OrjIe6(z7VsiaeC&oy+!4>Ii(5Z
zA^P|4rL1N3q!=AJTR&L%)XgDhXnPeVZKmOul!4!Wt^VVB<Vt4$fyA;P%eFBk{*@B@
zQm^}dF0(J|pgk;&xOYQ?9nti3w)GOeFnj3E+S}bT6D|1UOglUE*{J^V&b!o=#$(>I
z*DfKpMU;wd%fYXkf<gl0R-3cBC*9tbirANm8oqS!PvFAe$p29u`(otby*HeJd-)fG
zYVtj65+l$0-r$cM=H;nqjkWMYZ;lJ`3%<wMO&v$Crr$-Ty62wUAqOO%F!fKH40$Oq
zjhjAX7`MG$XNOkFOY)`KQQLNl>${jX=iO&3%(eV8AI@iNvL_F5EC|huem~v&K)>y4
zk?24lJpJdpHT-PtmyXt_eBXT$deU3(sK`M*I&Di%Q2vLbjL;v*E4FTC&-7D9(w#J;
z_m&EdB6hmeg52q?{dz{@S#u@}Cl*5MGiS1+vzY<;k;aLgMa*K$k=9h55*O@;ewwd&
z4=<`9_t8Y!Oa)Ed<NAAzNA~;<h@YMiUMvGaq>zb#Ku$|G9<vvMxHv41C=uw*m;5yU
zB$Ws$)AdS_ABxFjvX(>uOy`}T#r78Clpuc;Q^*u8i2&FgcU&midHW0=<BMVhi;yJ(
z;F#beBUNezukaXe6yvdYqC^0k*%(_^P7ZsF$LvJ0CRS5lA^<)$L~G|5pX)=Q11R>9
z^MNG-;L7|}2d)D(1CI$nu@Y8^C=meHbe|^Pd}Zy0$LvF~3RXo~A^?8Ks>d3R8r9)3
zyHKo%RU}9Rz*9|jdehHp0`ZtY6qB$dWr+az{p9mP`s)L_c+4IY6R-pYi2(Sc$UD`Z
zY0!tq>_)LNR+%Uf07JZ}gZx<Zn=8vZ!espT+EE7K;wktopj@D37~DTvg>dy@eHBo&
z&6Gv=?*0WL>Z|6e<_Y=;o~W<5nczr-YX+MUfG_AOY9wx{)W{q*UpE51MUBLbl^Pks
zhMU2gyrQp+kO)sS&?tONd|04m-nK=D$mU-QWf872L)jJ71L3MLR9r!45H5*9as`#%
z|7zR=LcZt}=cB9U4DyKbcIrNUtu|Z^;VM!TwL#kuE{=lJ28BYn2nEpw0V7-`ijp=M
z4^UJz4e<a<(=dPy){B|4%xbIFZN{R9Q+266AiGpg*8$VLHh$<f#gVcTlnr1R+`rB%
zx-A}7aFRfwx8l?--Z2cq)j@QKz!~8hAx7&Oh!%}fJjuWp;Tprn>rsNjp+@3)mKvGB
zCM59Qt7)C`ehHOqG-kpl1TvUZP&m|AJnT|mJy=f}_`VoB#wsMXjEeK!MF|1<f)=5^
z;+r7#HHA&r&jkm;xPM#)gv)>#ir@f*YX{qH(3miDgT{hc8#E`_X@lkrJ8#fDV2^bT
z3>O-m_@+stqrtTG(($6=!^DKoi{LZSr!04FAbZg!@7l0jP{{sZC<M4cP>^)FU%CvI
zE+Y~m<D*e^N0gVWoIF+mR>UJpgiW7th_23tL>xz1O@*|XvW>yC-|pbZqN|$P>Dd@-
t|JyEELzB8iM_1p#h-PAD@#)sjtZlWd)XgnjRLN|QP>h?8le>&3^nV7KZ&?5U

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_leakyrelu_fp16.cpp.BA4E6713D7C74F09.idx b/.cache/clangd/index/convnd_fwd_xdl_leakyrelu_fp16.cpp.BA4E6713D7C74F09.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9f7c2c4e4880beb6b958abb44be9b8a044b60eee
GIT binary patch
literal 1324
zcmWIYbaPW-Wngel@vO*AElFfyU|<jg;^LB`%tU4ehI|eNhKf0BLvo96oAF%x|MFnh
z>@(Tr&-we3x9>CExISB7zEp%W(k6Hb+oZai>*X^}7C-)SOK-Zq(8mS_`3H<M8B-Y6
zt?CixVmVp0XWh5=>s{E&FWYzDJpAO_<uyCz^1F*4*scDprgUTCzFl?Ij6HdJ4a+sf
z_~V5>6qxV+`>Rnr{{Jk|)#p!F8b<0Kb7t#$^f)PWjh*klGt=k4?>=d(e80G6j-T~$
zrS-ARm2X>x;_^RVdpm(EvDLETZSl)l*D3?0bj&=&eB<15L)NV|34F-?C|)kJ#IuR-
zInU|WJaVmuf1cjuy>&o)i&uDpOXo55$tMDT<xk{Wc&%)1PpyL3`;A3cIz*PN<$G{3
zWY0D}uWeg+5B$DR^)`q7w0Zu${mm|KBuW|`y1i~JNPS;4t^beOjg6gLfBe`Zn-6)p
zTwWe{L&!_;#@YZWN2P)}@8{|j2`^B+zT-o@)g9yG3#6ZL$ZMQ>aAJN;jBNPj7v9n}
z8NZY->$In<TsLdla86E2+_t3hyxd}4C*_W>7R~<^EN^gKa>_gzcXHW+QlFwPEmCi7
z&E>Ru-5g(S6>}?o&^q6nM^7f|P2uUOh2BDnKiJB0vzT9OExmJrFS|A9`PT0NoVk7s
z3*B!OsjZzcB~7ZNsKsO2>yQf@Y<zSztz^pL;;obZCYCTCblkPdeaiXTYun#<O1(X{
zYi{dYh1siKXUwmO3*kszqovXjx8~!Mj(MHS3dO=oW$jBppX_V!&8|qYDriq)zC9t&
zJ1ZzY{FP+1-2c7|)&2>WkL%QO+H5KLB7gF;(uBCl8Ly`(Md>~JVgGt*pV`6*rB<E6
zOQL_=Fg(@3zCTJg?T@e0*Y&E$mq)WLEi5{`PB>qq_0ESFBi+T)vd6+DcktW_D6aE6
zHfyKHo6uRuD!)uw_y2|3^V^Bne||_SFnw@*%2a9fqjl#K=I?!@dU$Vw+l>BuVxFO^
zxfGe(-B0JA{eSM~;hakTQ0ve)Q#E;3mnoQRTyg7RhQe*L)0JLDMVlJ$vt4gqz2}PA
z`cmywv%TWdb3O<BzN~n(|8n8uy1EAkQp<N-6|$eu{ing#^1;KrNf!ER_?wQaMISRT
zo}pH7GfmX^hw}OO)%qfDg%`Xq=le0S{_W;@nHT0(v~-s|o+EqZ?ds`2R$tnDCQ^CJ
z*?+9hgcyPtiYs%IVu3lXpjo0ad&|ElZYDnlh9Gt}b~P{olmUZ|^O{>dG^QTpVF&?A
zbFpx7fC-rNyLDE|?{6&)<7V;)O0%-D3V{ijG(%BpTJa1ZPq!~+SxEA&eT+Ootd^`!
zFl{i)8M#m_EnRi)Th$hmxj@xi+FZ6UH`J`nnlo1{@hVW5U6b7%rWqy;Gz8>6nB6ct
zn7B}!1k~%n9s<)_l$w(W(okHKT&xUY1b_%IU=AvLz3NBSW~*wZzyO9om@orFa&Aga
zI*{v?>ch;!%Er#Y$<4znz|ZF)=;EX%$mJm<E-WG;X<=z?V{2#cpeL$rrEI1lk4;oe
pT1HkuQAtHrT~kX(-@wq=)LhO)O4rC)RLspYn90@G(Vfwo0RX_{6GZ?3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_pow_fp16.cpp.0F662794CF8A0DB2.idx b/.cache/clangd/index/convnd_fwd_xdl_pow_fp16.cpp.0F662794CF8A0DB2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bd633b7c76a7bbcd9e5e93cb6b4b756eca5db3cd
GIT binary patch
literal 1318
zcmWIYbaRtqWngel@vO*AElFfyU|<jg;^LB`%ot_{hHwrBhKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*VC9~a{DLT*nFdTv$z8D35J}|
z%x<9zomG3*eS5#&pn3Pp`A6R*7X8i)yK}C6G53SJ!qxkB-7tuc-M^Q?&0PFIaM#hp
z(^+ipoXy)`_u%OD{b@&EeJ+ag3G4nip>ffV9TB3d{U_^xe){}(QqlPn-(}>NUp}|-
zME31~xauh{uA1e)?Tcw-`+87&KKJb{*OFJRyfMLe$!8-i-Qv`izDaQc|37)G^=f}t
zkQNb_{-Go4^Yn7_6sGkjJVG0fP1@qWtoYi}|I_X$UOBa9@<#r?+211~uP1N?xv_p&
zq8%sAtr~rb<w0%Ao=wY|kIpuKf8SB4jAxg_0@Y46jf=b2JXGG#xh-cB%l=@;D+((Q
zt*DmrT&vvJcz05ZTYKZ<n)h2Kb~CKq+{^p_;>CHUY*z)`e>_;kaqZ{i?#+*Lq)IcJ
z`F-nCW<CAj^W>I`L-N^$8!tK^ul|;i${|&9r{DMoza@8c3D;7gH=;R<qK{wciwju1
z=lLP$r5@fMTf<g}vd8_fJeqKL;ZiyCIg-;=IGgelZZ9)xh>||d`gViZQk(UA+qtIN
zG4|Q))HT*Txk%)6fx3Igxq^8@QhEH|-m<cM*6Y`=ef;n63$~`#lF)vQ`M=lX-OtW4
zkxnjWD|_K{YTe}8PnBypPDO-1QK*W!d{WSs+xKygTbZbR=;xDt4!)}^ldKBdcQodD
zY*!Usd0zW!kFI~cyHS(-iB-zp^O^Y9TL0>QR4F(o$~xz@iy*Ih@t^s>Rr|Loi<DY*
zYF^T<c_Wf~pgDf6nDM`5YG2o@9$y~Kwsc$4;dPVqGy-!f^dhH)^>RI4&6C5HyJF`2
zlRMJQtiE|YJy!WeFZTaM-*ej!ulfA&?1QWW{hH-@bF98qv;BUPsxSF=qG+I9RrHc|
zQjrTdWaN1FoBp5ud3jDHf2eijo27a))~>kqFhk*%+38BVqI6g0{k-eg!z)+%#H|kN
z$f@d=p7q({w{Y!8WhK39=lLXWyy@HKb%8xy**-k@-me04PcHFe>`xR=PBT1Xva~XK
zTjbH@|0X_PAId6zR(Qb+bG{#*^=~)N3*BS=kwg4tU8M1f+u{BnulGJ%md2|!`7gJX
z#JA+)%G{(_U|uU|mZ;3$@-K>;$&Y~{h@FjH4NL%Kz@X#2=2j1lsRwx&LV(g-EL<F5
z0w(=#ot5(YTT8>Znf!s$tZb}8U;-x1P?VZhJOjv+UY>t)Gpj%&BTo>kC954w8_aS>
zE)+{kSDpJ-wZ&vEP&Jn}mo3Z<HEXlx%oR(#3KVA7WOs*YhDier0l5!mH_Q$uE)*vL
z^?I;}!1NZS<|KkN6c;5ID+6ifprRjkpZJVff|vpW7y@8sF)#!on8~>*Iq5*ISE>&)
z3o9Et2PZcVuK+)vhoFm-njn{lkhrjjgrtR~wT-Qvy@Q^pww1D(hCDV=F=-iD1w|zl
iRdr1*9eo2sV^ecE6DeIIXHhXX&tN83Uq^RFZw3GlK@AT8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_relu_fp16.cpp.8F7B959CCFC1E66F.idx b/.cache/clangd/index/convnd_fwd_xdl_relu_fp16.cpp.8F7B959CCFC1E66F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5abe29475beebcdb9151af66ffaf3d1b72ff6918
GIT binary patch
literal 1320
zcmWIYbaPW+Wngel@vO*AElFfyU|<jg;^LB`%vfdyhG-53hKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*He|rRezk4&L7k}7uUd}z`P-J
zsjioU`m2w-ZtdTHi09q1da=#zC*LlQEVP#I(`zXAfAjNI(&3uYzu%Z7?#3|qJL|~p
zQ+e<(uk80%X8nEtOtr)2Pyd{>DXOQPN95@7qpKonXVuKqx4$nwx!P@?d(9j_>!gX-
zw=LZBP2=L~+<UiVcQWz5Y7Vy(-<G-dcu??0#dMG7sUcC%PI1Vp?mh7T>Aa|^e0Lwp
zBv#HnsGM~+{@&dQoO_g%Ls_M~FWJpJ7dic3bWwYN_GZ6hb{2W}l2c=wPkKs<9+<qY
zB3EZhj;3(K_lX}*&EgZwzqhBJ-Sfu9mrRY*RAw!3@sB+0UiUcXrV3}>@}?DzL9Kz`
z6IHhMv9K*u<?Q8S{{3^_mWkaAYd81u{%>7uo6WOI@!N-ku7Ybmt9NaFoFi45+05@-
zpEB#|2cIXmTpW_mF5Gz0S^WE%j8qP(k~{szKlm-Vqf5A!3Z3CHySO$v%Uyp(;JxC9
z3xk%ZE{W2<(j}{3;b(ZGAY<im-t=Rto=p#=Z){u1v*6mXB1h{ruA1kw-!F7p>ffO3
zpLcrVl$5}()E&Zti;U+iZ)v-I_)?I2`{BN?TTAcMOWbwXFk#&*!KcOZufDDJ?M+L5
zbM64&T+f@OoBZb-?^@6k>dLq1xar)E7m*oK=_Z?2|5!Q8zl<d->a6hPM(aboInKLu
zS9;XOtxU_Q{AE2!(8eiDcJUX@4;OYR*NdF*esJYhV)Sx>iv>o<|L=_dlws`@xO}pO
z%dXb`xr=xzKCHf)bnO0>CI7aj{4Y7Z;7VA|46B!ILPy)4PfE?ay24=JsRL;j+Gag2
zHh4F)w=Z(#9gF^ZOJCdXI{C9~MycH%$r#=rhELD+1|P3GpD=&#8`Z_Oo0Lwu=XYNT
zRpu6Hyu0|}_Gk6S|7332)4po1&bJ`(z7m~PWfoE$+s+orO}G<z@j!m_>w}?jT9dD@
z5>=Xg?{V8RoAduf_Wwv!(Yx;7FL~ol-!`uc?CHw;BZBY!E-?305<kZN#PZ}c!!ssJ
zCq{3JJo^0K#OLcnImFKjFL+_j_an6a?dEx+d#yimh`;QLG+uE#-2bC^)!Ev!QXy);
z#pRBFf6Y)_nVS>~%xwkD5|!Co{zY*!`7tmAv9qzOfeD}t7<8Q1-0Gn*^&k&J2vC}f
zg^L4Bz@*=;vr>M4YiSralRr?Jm5o&hOu(cWic-^xX8?IHTNt@eOfOw^?pxIslevsM
zL0sBgwlK@e0+d+oZalUG3bR_W+QKZUS(`Oyu2|w#pfI~8yF1KGm@c5%Aos!ShUsGB
zLU9sMuLpYwOm9(YP9jJ{aZz%yGLU8t`n2Ei!r%L;Q<(w-7y@8sF)#!on8~>*Iq5*I
zSE>&)3o9Et2PZcVuK+)vhoFm-njn{lkhrjjgrtR~wT-Qvy@Q^pww1D(hCDV=F=-iD
m1w|zlRdr1*9eo2sV^ecE6DeIIXHhXX&tN83Uq^RFZw3I>4;hO9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp.54C5613FFA1E2460.idx b/.cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp.54C5613FFA1E2460.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0d5098787953eb9e790ba318c37e89f68d9a6186
GIT binary patch
literal 8448
zcmdU!dt6N0`^V?mP18)Xo2h2H*OW?1O1hcuMd&6))N#q~<|L9Xi25GMC7o_`JEDZs
z;Cz)rN+=R?Tq4mq=v+E-DV+$%gkR2g&z$dSZ!f=pzOR>m^m?^g@A<6Rdp&DC>silQ
z&Gnr*bHs?jm>%t$8XuDq!DKKPYUFQnO49N=67KRDjMVZQp<g9DbX4uG?C;84P+rx)
zYLjbUQg<vbs%VCYvs>R+VP4hMw*2?cT<%0n2z+;|ileNp{^a-T?M@dHFNa;;Wu+Xy
zHS)rb1E(g5&i$F4TTwaMI(b!KV6F8?e{_0%+sM1zz1P<Y%}#6{HMrc_%X`?A-GAd>
zkv-*0%Rij0x^(jl*ZqB>OYeje{lmMo%Qsebj0_I+eUtS5;qJ=Cf+r5G(i0~FZEjdD
znjXG>|CaWcUw_sRTj$<dQQ-bv!>VenWR-O2$mn>YcJgYckAcS7h`%4Kiz|2O?Ow!f
zRMJUPUD#DrzrtqI0z#-6kmD3v>GP!8|M~b+YYyKk&Q8Vi3#CC}p*i&&71m=ft$N4O
z8`;wY+Wnh`V`sOQE^SGFe^tG1Z5J=gVc!;xdD>2gzL!5P^EWkr9I73CNhmZP_%%24
z@+RS?i^^4+fz10)()Q;Gncd<=59+2U6>OP8sMNP9X~ooa@Qs}fDh^&=vioh2)Owy(
zmX?vfpv<PbdWD|bv%yv4byeyXim$epzSR<6E4{Ek$lP~gw(w}Kg@M~X-R48*Pheh_
zrIC5lOyYZXcBySvdB)n%vcD+UN8RYnmiRqF);7s8;jR3i*ph}7&%YWM#K)CHJ%y9@
z8wdH=^cb-_b({}HI!LniNpkbsEAE7L2Y<NNIgVQ@SrFgUC!Ns~?jL8*jSAgXv&F7I
z#bD6u@eEig%E9z=!i#O#2PW+Mp?qubl8aff?p^Glsk<j0j5_K$>1W%jRlZ&WH%lUC
z7H4^wX#0+jK6di9;ihl9PrlT2waRR}oqf>H?7;rE`seX2A^t9^$NlzR@gBI4#S6ZC
z_(jgW*O_}Qd`EvOd~?V;Dmng9rnkZ>N3A5UHNJKB-P-w5+hum$TfQAEnsT))FTYCp
zmqbys|DPLEcFkM2%hu2^|IN(TQzX|74&{!P9%~<rUuv5w+`@kO{lE4ej}vOy{{C|5
zqW4NJ>*ok9FBz6R4p*M76!KuJg2CIx--bL(&%f~MfR15abXT;~-GHI$>U*`{Z|)3t
zpH&q4vrk!t&wA_3Q*(6OC2paS@ljVgj$XpdLf(FJzU{Yx%0E7c22J)<{4zh}H(f{1
zn^$_4DA>AWjrTj1;&^+iU1CDVoVj%eFDrSKyxV;tTH^SJ|KVrdx9+UWopWRBwmW8x
zDmCv@d&*5Lz708dFCN#6S(tD#`{~{1$Cb|Ksn<06dB65k+$xp2{_Sz6eItHp#;e@c
zv`cEgoJ#oX?$Yk2#kUednhjnHruWZn$bY~(ont$B&#og4(MhJ~M|Xd>WVHV4IEBSs
z?K<=3ZC|ey-+xAZlu0xPG|qUv^TWO)_j+%HK8-l?QgqWrA#9pWo&K#ErPewhup-ld
zMZ-IC_!ra^;_zpSpWb%H=a0626Rnuky8O!j#d-)Gk3=b(*H-hV?^-@m#M9ILkgB*l
z@<X;bHeJ<W3txI~U9!r_HP85O{xWT-e3JbgF}HDWWaZ?)Cr*}Bz3qzQZ_1M{*RR`O
zk>%L1@t%idX`rp|E7iB_vlLB3>@)5@w~2@iTVwuP#JC$Z<;$j5d8lqK`Dw!WKZg!8
zmv+AJv--X9m&(ri;=9=F(xHY!Qm;2tlbzfBlGn#auErP);gbIO3w-J_x!4j0gV1H^
zs>us7vdMo87F7O`TVJ+Rg%M23R+?5Kd4Xg{>80rL38EETY!NA&vP}8%0?E#&*8SQ0
zu+@f(1(LEE%S=^XAlcQ_>xSl|o!K~J4JjumCTPeDB+pXbwBu3m#R4uCM9TUseNB0R
zWM4b?1C|S)?pI-ikg|=YjYwV~Id*Z{Pv2|h^>MM;q^!r%Q<E1+UcFg+{fhbz3%S@_
zQjTIpnaK+zr`N=lS7=7u<zn+l*^XsrDld?{{iel9AM=t{oRLDx5{xOq*#G^5njH9c
z)P0e=!3Hk&71_jqWuPW6kSuA=S8M5W2;gGjq-?>mP?r}-KCz%t@wT5fi;FEKWfPVO
zUtS=&)AQdfvj-CnbFulPti#e#l@~}JI_hll{JvxX7h6cmmMlwkd4c3t?;I0i9{7FD
z#X?Ehm}Sh97f2p)KfP0N?Jjo}Mi?n;Yif^^7i03XUw2wRCXBh*Vp29@87avNB!74`
zd#$U>_*O2qfRxQy=4$c+$w@KGlKr(wT6)f_!_I;aII#!Q-GiARXMpDcfu612{fbQD
z2uwZ#Hb`I+)i9+RY$>h*^BS<A_$XK&g-I0G!jxL7y}}gk50%#=b6Gf12%<t5dm~O1
zfl(2RM_<H=Vi;cx6Dd9d_D8^p;%b;$4XzZ|f_p7YN6e~|hHd1psQ&C0usY5?8=P3D
z%*2UGFs!781}FA`(LR`hdVmxA!D2r+Q+xm{4p6MxB#O5y@wq8~5VsOURKtCM?*l`M
zCBT)?^@D&Pq;wQLFy)V;22QxJT?A+iX=ZLJjLH~V1WaNWJchv&JzeWUM(K%Pca7pq
zaz?ZZK^F-!HZ24*ii<#}2=wS|Rs`l07lTeQT`va9Vz8q05)hSu5uNQzz>(ro7*|Sp
z1d}L*sioA6Gl>!4jesiEegw=$sP}8@NM|))eXT}1v$eOiKMI<<U6&FW1FkV&ynT9(
z&)4l4Ur_eRVnvpc^Do;Q&?3i)>FnvMX#Pd(_Ib90>0|QSt>DxOZfF<{d7GTF+1GnX
zVE_aJ)chE2o9=5|^Jo<*7~2|)QM?*=Iw^KA!*v_!NCUtd05$X=)4ry5W&7}uPx?2p
zSF(2a;w&C1lmW90lu-{S-9GHkj;x*ZiBC?R^ZkHkqGk0bk3w6Wg7QaO98QSU#d^q8
zoN!`0sUdO}4PRZsiWug|{Fzrs_n#u*716mEc*S%s0bU86OMzEPXDRTcAG4X1@<%2i
z1s5qb&@v8_!a_7&XkR9$O|~hcK_P815Kln8M~y-Gqg~nK-TbDliKI0+F$(6RAV%$`
zB^<UAc{F|gus8GU)}fVy%PKLPcmbRjz(;SyiI>26Nij+oNV8<w0%evP+f5sph0+GH
zkt}DR8hNw5$6VI0>|8;|=0O$mm10~xHlaA-j=PW1?5TwbnpU|x$VTQ)=H6)I!wCx~
z3vaZ)hxCN|b?*-F`0Sg(eCOcA=7ga>avsd=%pB2_F07AR_gia#4cVzo(8vTGYOpqg
z#%79_+!n7lvAVBGem}`7X)F|mr~GXDG$7)v0+SqUhZpo1`4U23LKu1;lXwN<S5%A)
z?|b6Qz}K81pJ#1sJq<+$l!lSREz7~E;dW)aj_nY!Lv_>a-}nDTcKR9!uYoCg2~Ol=
zZu!`3#LJVC*UlL^<f(|0$pP~Ss2~9+lLKZoooj$uL+7KwJWA(UVAfK+a<1gQ(J-_1
zvriMsGs#sXOEQuDi6`&Pek0mk@!8*(Qc@9m^?6;xCxI>Hoo9WOaV?9SAA4i_K(rjz
z{JZ1LhI;!7@?!~RDZ$)OYP;6o`?_VAV-P9Wc-RD^Ws}<0+qZ$)c|#p1HgY!dsPWGL
zP6p7p05}D7ZU9aLog0DENas_)IYs9t;55<sG;mJS`3!K*(D^KI&eFLVIL&l!0Zt2@
z&jIHgojZWv0Yb{2`(SyWo?+bw$NS(y={;cCL)Ux2v4_&p@RDvXJ3OFi$B8J5s9DG-
zGHU|j=T~&Lliu6`4jtf4J-q{hJAk10HrU^$GEAK40*5Yo`@7(9m*UNHzjJ@}NGX|I
zPgg;Bm7dDxVQzWYJoMmg^~2G9ZGNu*^J>p}F+<<!X09%oLF$<6Sfld@PUxEJTBEEG
zC-lwrtx;6Q2?KKjYji9^Zbq()e~wRRh;sZCw)|P}x0rM=ZJxFUI*j1N2K5cJV3x*9
z(?C|>L^>~>MjHhi$5Kh0$N)hGm0IFN0|*-E+z5h3I-df;DLOZSpoz|>L2#PRXFza<
z&Sybzmd?!}Xr^-u2wLcT4g}}ud>=gT!z{|S9`Nj;xI{`U$>i(bC70z8SPsEd)NqfR
zhSuu2lMesSQ@MBAZ?2+Xf0ZvjOGU25RQ~9sO>Wxi?gnT<l4R&Uks=Sa$JjP>;JJoH
z-+;+G1)MkrqGQx~04I)v(Q*0-u?({+!^V!HRolbL9<6E7Cs%_63?=lia1e|R(hp{b
zz~T^f5W|VXU~!mYlr@lHLv|&A)&x#?vAu+-)77bG6ocXq%p_gFh-aiZ^564#t;OT{
zi-v<Ikq#gQT`A}z&yevHDcI53PYMBuk1hV$_CPLQU4Ag;6jGa_>|o3(qH{5Dis@Ve
zoDw>h0;iPDQs77_Mwc3-b+Q``%DSoSsbg7HL+ZOK`;66dNb4+}EN7#uj3!?H&26jW
z805zv-xr+bv2R@t>zn<>yYHDi{@x7L9WCT;kmiu)fo6vcZVu@*Y6HDCx+h)&y-Rd%
z2mN;H<a|Dhki<MS?fv3Mr?p#R$LMHT{@D|zHt(Tm{IAxn7?VizN}Gm;ok^s7rH|1?
z^;?pfU(WWB8|rlsUZ;<D0p?bK1){HA(VS3yY~sa8GK5?Ny^HjQmqG6`ov(oY6>6e>
z+~Thkct<^z^zRp7{DRuDuO<Bu{QhBl@)tjY*HyiLde0wp{^AYkFiG2<={}DykC&4Q
z>CwIjtc$1{N7+T9T`{#q<3tI#mQeR!IC;(FQA$tiM!<T6`aa64$apEc;-%uHGuv4m
z?FOAb1=n(pU(@^S3!}@QleEh<Qg*;Ju{9B+;N5ll;hV=BSqyS1{0PDysY?f($i>`p
zv0$_mx<zLc?^zmM|0H7n{xHLHr&`?uQm~n4vjm+idc#wG7f;*C{p^PHj;yMA?>S$W
ze4nYUsTf`4JS#f-ro!@Or4ml$05b=aQO8L`fSF6@JYeS0IUkt$bS?mX0nifA0uUE~
zJN3LBAl^YS+S15rm0itK(;DRy6@tzN=+295B)g!e8ac8Z$F3<*a!fXoUDH#I9M}$H
zSq;h$*vcQ>)8j-CJBWk42q%Ks!8Dr7o=c-pb|{U)*kLsKiv1Oh2sSZBqvt!!cDtsQ
zlUY%+e)8BoK2D_Qr_gAP{u&x>hY8!MV-&hmAq|xMCxJ51lkGWXAR{SeWrR#hPKrwI
zlU0&r#XmpssKQ{xQPpUw8aY-~AWEL93k-h^d?H9%m9RQ7GCC&e>(~`5<KvbuTOPrP
z9c#$=nyRj#s;jB$GO8LyRTHVIyubfo4E5=9s=AV@CQ#K?W7Vkm=#{c=_{9V$Fcn!!
zYz`M!QRVY|)V!yesVV!YYYBwfI$|dmS2uSLPiqZJXA?(r!~gG7O+9@BqjAQj<0n{5
foHW^Hik-t$k-e^!?Q{*znZC)`j5%Jjz@PCS!iSH;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp.303711D8C093AB3A.idx b/.cache/clangd/index/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp.303711D8C093AB3A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c21fab0457bb98c8180dae2580fba37a752acf31
GIT binary patch
literal 7736
zcmc&&cT`hJ7th2P6Y?NHAhc{inuGvKl`gu7lmMb2Sm05mNJj;1Nk9=+DGCVKK%|JD
zC<2SUqpl4VcM%W;S7614d=^|?^?S+g<UNnyIs2b-j@S9U-@KPQckaD&@67XXbDPE`
zk$#Ny$ViA<9!4RNNJ`khwB@O>MHmQVl1LdlF9t45(zR23F{8y-@6O~a&6fNjG3A%4
z8)8F^q1zTKR~*?;7^1B>d^hRX3Gw<!?X4yY>Efr!zXfi5uB)f7*UsM1EKnTISh0Xz
z*k2Ij=B6gRJ$B>eUvYw)Q$C*KK6+}o%HrPq^6NeY<LcRWoGzuY?(F!`epDm>rZD99
z-~YH@-m$LF=W())(^I|d&e1y>yMpCMryQAaaO9<?-68sEbV+MgY{)|S%%O_V5VKdk
z2haK~Enl>HBrS18g3}Iu6(gDI?_4-9Z;9r?YEk+#;b$L%KMIb$={Wz~!;M=$?_GYc
zrr@lq`ea6RH>bq6CatTs-{9zqeKkDRm^z0&x~IvfrPV`IBX?53KyQ%69MA18bD5jV
zx?1Bu4P85ONFOY-d7k1fpH-ogIi;Bz?A#N-g);LmPilw~n7C=brC!y_vkP4glMa?g
zfbl<lV)P8fzjnlFwG9s_nNu4;`nv&LX4jwjw7nN%TKnqrn8(v9R@t2-E!^&2et2nP
z!t0u!PZS%)ieiSFxDLZpled{v6bsx7<`w9tMG1x;lVV4l40o$2m+0{BZsp}^jSs4a
zH-!cNczL?IMx8#;_3du+iPdSgVzYYx8)w*__}qzC?yr?Ifti@Mt;S`pa@f&)9oOh-
zJ~OShP1;lY)7uY?zXo=NeY`SILh<VKOE~&?eD;G7ew-B}A~5Gbxh3y7`>osXLJ-N_
zMCRn}E^ShBEp(ahV5!%;IjQ+ifu87~>Po@2C17z;lzY8_LBl5g5W-^$=S<@){^jZ=
zPc^bl<FY~dj3tyQUYDO3SL9N&W7nm>QI>x$8Z*wE#Trp@Zq2pG%15u)Mf~M6`$?i#
zU!{*m1^=3wQSX3Z+9kE_GMbl%>4hFYw|KWFbN8&Z%CT;_y~p1@ecPPT&N1zs!x{F1
zigDl7bDT~J4yQ&pgbO1F%~$t(QwtHLna;L#4g1~ZUdmo_x~KO{_Z5fgAo_bFl{Y(i
z$lPaa1^=4a_qAVV1V`JodAu;WJ2?KpW7Me9t2nTE*ikFhIb3|dzB*({%S1ubg^!n}
zuFcjKjA(_uiC>u#7<R<CJIsZVZv8Tz;&tU=XHV|*(BB_y&f&do7!!x}pQ@X-Ft72Q
za%&>ueWUB$ubWkC>drK#ti9M}lk;q;zTcMG#OsL>-Nuhr)#tm`uPbvB^%*~0RlnY~
z-al0N^YvGs8+tz<PFK5ONqw%e`d#GdyMg1q!)wHqA#c{*@<nO`f3uHU7K0u?c_6}U
zyV(^3@(av07uV%ATxqy;FQ6_l`jySOz9YeZ($Csg#JwLb`c##0=f>1v52izDVa9_U
zRmxfN*T8Alk<zNE(^A@7SxIpzIvjI%gA4Z2`&)f-Jhi)-heT72eDXHF&5ms`u3O)E
z^?j3vrB71d70vqTi^S)53Uw9Me6GkzPOiKk+f~)LSiAdfao@V#Qx;`;)JNTnepM2f
zdBsklXh9~eM!31~g!~PcmY+Y~8?#%NT9U$vxpi}riM#6tR>>z3J^M~w@m~|zodX^z
zM~mMvNgK-qma)yRZJkctY7vsmIOV6)G84l6$XJ3>9-LRuomDMQnoA;~#%yD*RA7Yk
zPbGLHV1av67K6M5qcx}+N>Tx7oAHa021(py4Duq3)}!h%r2^7(<itfoe(i+}vM)w!
zQMK4o0qK|}tMBer%O7Nry)jyos;MLukWS84&s^5}X%U0$htZ+bP(!JJ^qR)Fox9j!
zzcI-3FnStwn!Z#(I`<~;XLqC0a|kIFqw~p>d@}8y|KOCOi#yLgN7Hu7lLE0RYHYR1
zQUU4epFTP-cK%@%gB*m>94bdiDj;2ZszB-VptXQO4#sFhs-dz}K)PkYk%^t2>Qn|f
z1fzASx=g8n^lg{@RKt5y_A$u*7_ClKSCk4Fed<!-sAjF4JZU~go3c&0QUU1~HqYMm
zH5+X~NNE@yKQUfaDrEFQd%e-#+64^qLX0+|8YxQ!q+h+aNs7AX8O<Os!05@;$qG^d
z>32^bY%v=S)n$+uW3)C^dy-T@`qPm2N=FBSa}06-MjKEKl%xXEq|~UTX&tH<{KvD%
za$k59g1XRL6yfHP&@(eTEO|_TpxI>2Y_c9)85kA#xSVqxnM6U0!PH_f4H6Wz1eliq
z3mlgMUMVof@ixHQM#N>nyo@-%2k7<y*$+|Bdw_cnXv6m>m@^tGuZM4@BB+onlubiW
z5zr9<12_#q#lS#}{|thuX;lpDaJ~dgF9D7?E(6YG;75q5P2)jA=CT9de0|i8Gf$rx
zH_1^DbR*E-h#M9`bAe7SFozl<Xdd9@0ec+(1b9E;Sn(J)!Lro-ru0LMjev_!%mqj;
zZczly2aJ3oUjUE-oL36yj7hFRnlK;Gv(U4JIR&O3%zq?#2l^L++R|)U(0|xj;;j*)
z7mEP}?E!)w;0={fJ1^SS@}T@9je-i*g_`g&hV+GoI1&L35zr!HLlH2-u^4EGiM$x_
z#lRHj2Y|u=P$Xi*0sMV+jqy3BuD({nR>pVbd%=j%ymK*y^u#ghyXvN8Gn&4`ps?br
z$m6D0xm*VN!x+&3>>7X*T%ERp*e;KD{UJtl0#+x`fyCtFsWHz)jyc#SbOMD=pacp1
z!DE-@tU~<1`VEuLeO|QiLndEs+f?7ZPuMB^17E!Xy@De5#R-F;*2>me&>JxHgS8=!
zY-l!05ZQCbt}dg7jbQ;NR1gxjOav4}#4Qv91u+r#1BL!Cv0*>H#T2w3IP~L2lHj0z
zun3x9#`}O{8&VXuVhb7qMnk|B60?%_nQ~_w`}PUIwfOa8zf{%(OryyblWpN);C6Du
zi}+_tE6E7j4d~r~2^B@qUO?}~F+9|<xRjjecwE}k?A4*QJuA)>c4R-3$G&EbSj#p6
zK^+hW8S<XKHHmGyxd@wNU}xY84;BP9v@>*t$E$x|h-cS!!HjR7D1$p~Gxbb+*_fS@
zOp|P&IYO3qTRawoog_;#4z~wSV-(CzfxRF|HgWI|4-&HC1ol2(mp=;zKP-AM^GPZm
z+<bO4J6T$bdC@e>>chwz%rY%Nr3L82k0R(>;Pe(xbFryuD}CS9yG)dvKa^r1k54QC
zloBE?1(Z@E-UcY!h`0<;%5WS%ueMibjB@Uqn;lktK<<lbFXiVR@3-9SUEf@v<)ym<
zuZ}jgeU++~zjtY^(LF@P_San3+y{F8iv1mLSG8L0!XD?7`T1lgSctCla=p$EvYCq!
zrZY_aZvRu^Wn6eTxQcT7h9ZKlp|4TE_is0#?*;^13FwtXTm|S=M7#&k_YiS4pjQ*|
zUO?YV#5I6kL&UX!UQ5JvfL=$$^?+Va#BG4t22^l;&jbE>Vl{Cd*qjFrINuKV?L@vE
z*tFw3G#?fgl2RTIi%?$Z9Ox5Sje>-LUAHe|-rfdm+Q4l5>uq3e8$fY<99SL4t1SdQ
z0jy5|SDZfutWV)MdtRCItD#A0SiCg@m1bgD_Xaq<0rTODIjv)ngBLs<|26tIzMRdm
zySZ5ttB%x-)XiYIkDwYx8fLJxK~OCtEi;%<5tL)ZF@s$PbQ>%V{%&5NNj7;-SpU56
zv#kPxswt?c!X^em)0NXTpq>c28?bf*S>1!6m4H=A#8rS*MZ|jmYY!1u16DN=?***A
zL|g+{HAGwsShYl42UvAPTn||FM0_5&oCkAoecORcJC5P1Fn5sjWpM2XYC*G*Wy#X<
z(BLd4=Qk{(x&XfmOow0F;iRgTJkMm`zwXlRS^bV&KQH<3T8Y6O$4VPXlZmg@LD}I4
z7#G;dsqJ8lBS)H}%nsF~s=UD`dhaJ7Xa(R_;JpokZU;Kszi3*>red<JcdE$^su)^v
znu9rMKG4o5MrQ%gDIoR&g@9Lx*Zv4v1b9U_hKDt_U6Mu^UxG8uSp_P6AmhZuxe3*7
z*mh|vXtQCPd5AY0uz1YRWP+r0((DHuxQh~O-VZE^*s~u9AU?dL+oF0iQyJ4<NEZ@L
zJ`tdc@RLl^snNwm+z;seIEIZnrn{s|$91=&S;<P3wv6{RR-$BIOm`zYBkvdgG}y-!
zMsDUL$C0qdU$(Eu6==6E1r2(CSI0V$&)h9iEIN((#eaa-e~5n22(%iB_z>V60<ym7
z?8eaAsK@#bzpM7=x@s*+9j9Zz`HA1Ie5<zc{>eV<FlYrTtwb}81E+D|10C=e_LKvM
zjoZU9`yK#V2Z$370<D8Y+yppHxP7mz_nPGMt8xbBHa&pVgEyGhQXBj}JW5FWu7<y^
z=z8}9^O^m3RiZ<0b#9jPe5NG0O@y-BAIUY0nFznlkp#D;7|;9&IsnWDa2}q!Sa3`3
z*zn+<PMfX_6Vq+?!fTtFuW5aAPlGierk&)957%xcZKmvs?2Z>bwcIg$5OWq?3td~7
zuI?Ov^mbTCC1GBA7^ob^J5mIF4V+#BKj@`h+!GUf_&V3W%25N4L@ryDn(kqQ>3q{*
zsLaEV<pZ{}wlKcAHJ%n9e7lQ(vvLxGW>7NZp#BJ&Ny#LjRg_f(w3@P-fYvbAsN(Of
zv0dYgqqVkcWe7gSiYv+GDKtDx>)3usE=}QM1Pyf$^@l|$Y?!f0lB-jEk|oVjwrTKi
zqse`7cZ#6iG;cceI|TKi`4EsV&6j}uY5oKhKnozC1+)bO6if@2p^s-f47WRG?8I_X
zGAEf3Hz}MH0!rni$`I_ku?0zPbMXZ^(VS#8Mru@in1o47jY#X34B{o?@4r|$Adw>Q
zp>+D+Fcg0pjSrXN!#I4Hgb$<eVK_camJKJMFTLslqF)@?uAGpVlpGcj8NM_+ZdrV6
zOjJxLDN;6%6oU_!;lo6Hm>?TQBt*tbbo7i8OrT7pPNLBnh`b_G!ClF9mZ6fIyRsTf
zMP0+z&cV^id4`LbD&Jnu#z_1BJ!NZge$bh$t8XxcXKXTcnz^O*bgq@Asl|_~Y&VY-
M@@$`(bAT7=e+h|0&Hw-a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_sigmoid_fp16.cpp.7462E7274E7DC93D.idx b/.cache/clangd/index/convnd_fwd_xdl_sigmoid_fp16.cpp.7462E7274E7DC93D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..92cf32bb29a413143a68c62008085379738aa01b
GIT binary patch
literal 1320
zcmWIYbaPW+Wngel@vO*AElFfyU|<jg;^LB`%s6HShI9@FhKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*HfQKRezk4&L^!;7ngKkOIYm{
zJ*C0#%Hz_l_4TcicYOcrr14L_?Z5d^nVn25!~1zBK3_fB`t#MVx2y)cb~4P{6=753
zdEoHfH{V{f+tvKd4qIcV|NLmmRtbJdjjsOg(2c)Lej3Hs?bBEL-c#fMvC!(RNcWy?
zyid2fcgE#@zV_~fXhwkio=-bB_-5Nqo7Kf@>1?IH_>xh$+Y61Ctl#vvUs+Z1f-jx_
z^lKiu!0dmY-nRKX5SCo2bs=KXH_wuDyT06?WZe=qvsCEYLBq=asinMX%Bdj?{<|J+
z-WafHi4Vg*_hZ`GRuOwDKmU_ma-`2qoMVg1MHbbnsircYq?1o8d48}Kb>W?))w8Qy
zd1a}J!G@?w7hV`_nEQUJS&{I9(rp_b+Lv#6F4LUVp<7cB$nyG+=h3T$roFqS9Bg04
zKQqnJ#=9t2X~D*47jC>*___L8Mk<F?$(iGlf7<7?>RPo1abDvRySO%a(PIA<f%l3Z
zFAQ3ux+G5fN*AyGmzbmt4=)7uORwvRo+Q+ipKyDbSwoccY2LRR#FpC3-`mbL)sC^x
zW~Xj-(4@&hbK0ymol4qXPq|=VJ0~is#Ny4CecSqewqN9B;m!>0*En8#ZTtI)QfH6d
znA<38HF?$ROLi}9!dOz*XcY<Au6wzpW1g_@;~uv%S^Lt@Cwm=yS63!k6}azc%=NfE
zY1WnNQ>%_mTE1Ud@}P3c)r;QoO#ExDU-h5-tZ1>d?8dc|t|l>OKGr|;mP>WlcqP&q
zyF|C<jnSzF_We=1X@7iGzOGljy<E3BD0*Y!Rn2^jz?=#_OWn0nqQ}A|Z}8j-*j!(B
zY|hOW5!zFwpI_X)<=^7jbG9E|^ZDV~2U!RDHI2>vdj8HkXjgsH+pYRZkB0nSU7^)(
zVj4`d-4ExV{r~N!bABa%sCDF<rFt{kk7-T2{V+q}mf7h_x1w}c=KZ|u*uyJV`oygc
z>&U6<m!9?6<9Be!BlmMVGVQs|Hr$lWow|VkoLk+lWi@{v=1tlZD<MCDT`l^U!R0Aw
z`8U%<-Ty0}k6*2<H%r~&V!X_U&;M?1oL9QPyP~CgOZGLMSKqFB|1rHZ^T8U&)J^q_
z_DO$s7gy#c#R79%L9;|<_LhH9+)RE93_<K{>}p^FC<6u^=QX!_XiPoG!w>?L=3?RE
z0246jck8T_-``pq#?9mplxAgP6#^45X@;WIwBi{+9?TX-E)>&CSDpJ-wZ&vEBTo>Q
zHkU2TvYNG7bLNUAUIhxXYqGn;G{dBUhJYLbv*h}e3)2s0*GB<Wvs$t`z?=l5nYd7#
z1T@lvJp^WCQEE;i$eiM$<YHwY%^dXn=JzL$U3f1u1qLt#z^q_k2t+WGb5nBCfn2Xt
zA7&O-Hg*n9ZXR9%em)OD7bi7AE)OAbVG#*Q3rlMoTRVFPJyC5dWit(VY@%Y)GO`Ma
kN-C=Anp!&g28PC_=5i)dx<<~TVs4(nOs>9;?u_0H06we|-T(jq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_softrelu_fp16.cpp.47B66AF94BC3AD6A.idx b/.cache/clangd/index/convnd_fwd_xdl_softrelu_fp16.cpp.47B66AF94BC3AD6A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dc4161aaedefcae9ea81c864742531522a3c5ad6
GIT binary patch
literal 1322
zcmWIYbaPW;Wngel@vO*AElFfyU|<jg;^LB`%y?!7hHMT7hKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*HfQKRezk4&L^!;7ngKkOIYm{
zJ*C0#%Hz_l_4TcicYOcrr14L_?Z5d^nVn25!~1zBK3_fB`t#MVx2y)cb~4P{6=753
zdEoHfH{V{f+tvKd4qIcV|NLmmRtbJdjjsOg(2c)Lej3Hs?bBEL-c#fMvC!(RNcWy?
zyid2fcgE#@zV_~fXhwkio=-bB_-5Nqo7Kf@>1?IH_>xh$+Y61Ctl#vvUs+Z1f-jx_
z^lKiu!0dmY-nRKX5SCo2bs=KXH_wuDyT06?WZe=qvsCEYLBq=asinMX%Bdj?{<|J+
z-WafHi4Vg*_hZ`GRuOwDKmU_ma-`2qoMVg1MHbbnsircYq?1o8d48}Kb>W?))w8Qy
zd1a}J!G@?w7hV`_nEQUJS&{I9(rp_b+Lv#6F4LUVp<7cB$nyG+=h3T$roFqS9Bg04
zKQqnJ#=9t2X~D*47jC>*___L8Mk<F?$(iGlf7<7?>RPo1abDvRySO%a(PIA<f%l3Z
zFAQ3ux+G5fN*AyGmzbmt4=)7uORwvRo+Q+ipKyDbSwoccY2LRR#FpC3-`mbL)sC^x
zW~Xj-(4@&hbK0ymol4qXPq|=VJ0~is#Ny4CecSqewqN9B;m!>0*En8#ZTtI)QfH6d
znA<38HF?$ROLi}9!dOz*XcY<Au6wzpW1g_@;~uv%S^Lt@Cwm=yS63!k6}azc%=NfE
zY1WnNQ>%_mTE1Ud@}P3c)r;QoO#ExDU-h5-tZ1>d?8dc|t|l>OKGr|;mP>WlcqP&q
zyF|C<jnSzF_We=1X@7iGzOGljy<E3BD0*Y!Rn2^jz?=#_OWn0nqQ}A|Z}8j-*j!(B
zY|hOW5!zFwpI_X)<=^7jbG9E|^ZDV~2U!RDHI2>vdj8HkXjgsH+pYRZkB0nSU7^)(
zVj4`d-4ExV{r~N!bABa%sCDF<rFwhTHYv+?uafO!Ta>$bPP<BZn8}CR52^~LPCxZz
zYgX%tW#4x!v^>AX{-{RXi{9JcE*CJ*zI!o@^Vfr%C)cx2eV2cyrune-BV%RSrc@qd
z-^@SLa>EM$*nj%7b=IRrl}sXkk23E6uwM4HBxCFm<(MU<o6`ceuaA4A|LL4eI<MB`
zzubI}{@rFMuFOq}1?IPcW{Jw|E&rmpnfw?Sg4o&E)xZQ$1`ImRYi{+>n0k<hAp|JR
z#lpn_CScO<)>$dPzqK@so5>$2&C13q1SVk83`MDF#WR3BUqz`{|E&S<8F_+OEm<95
z+F+J5a-mpSy6W7wsx2mSfvUN*xolx>s9BpeXRcV{RiH4tCc8UKGfWz22*`afyJ2=P
zaiKT~sMmu%1g5trH760Ip|~iySQ$t&2Yrt?ay!fO^ctqX0EPgVSquz;2xf9_N=`bE
z>y_%m%)-jX&cVse!z;kg=OO6gq$bGaAtWv=A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%
nRzXooMO9r>OGn?p(Ad;m&O}Pr$XQg(%`=$E)z{IT(VGDP7RVK{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convnd_fwd_xdl_tanh_fp16.cpp.A1CCA35C668C7A32.idx b/.cache/clangd/index/convnd_fwd_xdl_tanh_fp16.cpp.A1CCA35C668C7A32.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ab8d38683ff6c6a70161c6360a0d273ce53086a9
GIT binary patch
literal 1320
zcmWIYbaPW+Wngel@vO*AElFfyU|<jg;^LB`%vfdyhG-53hKf0BLv9z{HWPUIKP7BS
zQn;_xV)n>sLcYn~lhq!(pA(Acl4xY;?$-PJOV4Fh*He|rRezk4&L7k}7uUd}z`P-J
zsjioU`m2w-ZtdTHi09q1da=#zC*LlQEVP#I(`zXAfAjNI(&3uYzu%Z7?#3|qJL|~p
zQ+e<(uk80%X8nEtOtr)2Pyd{>DXOQPN95@7qpKonXVuKqx4$nwx!P@?d(9j_>!gX-
zw=LZBP2=L~+<UiVcQWz5Y7Vy(-<G-dcu??0#dMG7sUcC%PI1Vp?mh7T>Aa|^e0Lwp
zBv#HnsGM~+{@&dQoO_g%Ls_M~FWJpJ7dic3bWwYN_GZ6hb{2W}l2c=wPkKs<9+<qY
zB3EZhj;3(K_lX}*&EgZwzqhBJ-Sfu9mrRY*RAw!3@sB+0UiUcXrV3}>@}?DzL9Kz`
z6IHhMv9K*u<?Q8S{{3^_mWkaAYd81u{%>7uo6WOI@!N-ku7Ybmt9NaFoFi45+05@-
zpEB#|2cIXmTpW_mF5Gz0S^WE%j8qP(k~{szKlm-Vqf5A!3Z3CHySO$v%Uyp(;JxC9
z3xk%ZE{W2<(j}{3;b(ZGAY<im-t=Rto=p#=Z){u1v*6mXB1h{ruA1kw-!F7p>ffO3
zpLcrVl$5}()E&Zti;U+iZ)v-I_)?I2`{BN?TTAcMOWbwXFk#&*!KcOZufDDJ?M+L5
zbM64&T+f@OoBZb-?^@6k>dLq1xar)E7m*oK=_Z?2|5!Q8zl<d->a6hPM(aboInKLu
zS9;XOtxU_Q{AE2!(8eiDcJUX@4;OYR*NdF*esJYhV)Sx>iv>o<|L=_dlws`@xO}pO
z%dXb`xr=xzKCHf)bnO0>CI7aj{4Y7Z;7VA|46B!ILPy)4PfE?ay24=JsRL;j+Gag2
zHh4F)w=Z(#9gF^ZOJCdXI{C9~MycH%$r#=rhELD+1|P3GpD=&#8`Z_Oo0Lwu=XYNT
zRpu6Hyu0|}_Gk6S|7332)4po1&bJ_OM&H<ww*@8wxwD_zDD2$4sPR7A_2$)kLe$oW
zYP*>2J<fZ!a^7Ffx{rrFV%E)<Gv07hHh1a*{&Q}18<*|*{xENni(b$3wN0B+d5nEC
zpG?aQ`?$0I$<J)jL$04$G=B9m*IkYOy(ZoDy`Rk?*Ih~1cwT+G8vLg<t~h=(+tic)
zSZ{3nzMY}CGB+s}nA-}PB`UMG{EOmd@?&5KVrOGl0~0_QFz7h1xz$5s>Omfc5TG;{
z3l|5NfJwhwXQll9*3vL;CV!wbD;uj2n1D$$6s4vW&j9jZwlH#`m|nW-+_$PNCUY5i
zg1EG~Y+;twtj(G;S1j==P?%kl-5sVGCJi(M<Pex8bFu{$qK`)V0#&nGvf9F&1f!X_
zP@Dub(t|w&W@J%nP9n&h;-chYWgyKQ^d&_}>=@tt2TXwh3;{4J7#IQ(%;emZoOB@9
zE7gaYg_Vt+gOi(wSAd_-L(s)ZO_0k&NL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`n6!+n
mf})a&s=B6@j=q7Rv8lP7iIlF9v#6MxXE2kiucJGoHv<3<%^3#(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_backward_data.hpp.F488158732BCE944.idx b/.cache/clangd/index/convolution_backward_data.hpp.F488158732BCE944.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5dbc82176418fed94741f95d7a27f40c8b462e1
GIT binary patch
literal 6146
zcmY+I30PCt636cxjWK#P5sfw|CJ+h)fsn%`$Pz)N3bH6FSkbxx?v_@qRbQ>5R;6k|
z#a0C^AQcr{fP!_2vM8WdsIBmPRx5o~amUu>RjBq|ikCCwgD?5rGnvf!pEGmkp0S}L
zMl|>1xZ&}kOBW<8p5B?`xL)G_qQwj6lo5{G-IwE*W?cI6^8433`>cEFa{TJPh}FxI
z4()O~`_9n!@**GQMBgvzfAOkELCC}Jmt@YrcCt3K>kq2xE=l7*w`sh(@-MrVV|I<t
zH{Kk0W!<pB31>s<!Y&px1jem%*t>kV`S#i#IXm{$-c|L#{xtUSrB|7^r>yMn_|1ba
z|MlswyDiq0WwAG3j;f4X*%Ef<=EqY%%<9p1?Y+RnjM!;EzMgQ*DfGnd=aberPTrAH
zJ%7mPb@4x)IUc_Abnck1di=aEE!WVPRw2ai8M~wM)lbEa75sydmEQBO4f{ED*TTy&
zd)96Gc607bqh(`JUY9Q_7L1dhEt_%xvPZ1>%jeSXH3ycDvHjrumdM=YgLgf<5LKIV
zZMN+P-~S$&eQnI#=MgF4Iw5@afz(Yq%GZT~eZ(By^_&yQ752(ABTDRh#y;Nu=cYzl
za^t4go_>jO+xi^|&OVVo(rrP(@`431#@i~lY4*LlT>7rWWi<6G$-Mb-;g!U_3DdJq
z#PYw$BJX_o_>&n<$~p3JpBEMkKihNH$jh^5>djHlLOMCl>>M#Lc}3(8cgy^Tf1<tp
zy8QiFfz#&}9J&x0b<g?7W!cXM^**H1=G@r+{%1=*uDp@GW8eL$2Qp%u-#ZjI{`8>d
zbC=owi{#ckUg40~{ntRB>(=~HFDlAw;vUvq{^Dr4y;peg&0_1;cXJ)<-IJ}x<P|Q-
zyDb;3uA1kIWheFFPG0%@)58yMdwI(7s(Tmh(o!8)`jvgB)0KVZ4UuihFnxBkv~2RC
z9vgW5f!|%G%`Q)y1`Vq_n-W%qKYU)Fl5IO={PzPk?R_i$1<Z4r68mT88y@#pKMVJt
zxpJd#mj{cia~Foxg4U3K=RxOX`$GyR|M~KR#QMzj^SAHGaM9$%zerH!s*i8<*{#!e
zN>f&@{dRL&(#(C2K3cW)$5rp<MISNxH@+Yx9`Vx`?rHiv>0xcn&QX!O&$u)_3gZ?n
zTQK7la9mey;-JkAy}sr-ovXOP#h2hNb05MfRGm2=0|?+6y`^%qZG8jJIZ6gIWHW|x
ziauT24CdyoDoPJKcZBC$B!e=Ll|eA4c#Cf{Sztat@!7>;Gk8ubnK;XwEu2Exsm)-y
z?fM~OZ`JPLIVZ^=7i76GoKy7e)@HIkZH#d2^{rPtXD6BTk@XP+?E!B81iJlyPlLY+
z>2#0h21*9@GJEkydXhE|KV3-aY8&il<v9n*AVd}t#3^jLv>DU~*DoA7^!6<qGGGuV
zJOI@L81Uwe>)Iwx!xsT_ZgpQXkLT<qAO13bQH@?Q_kW`JwdSvh<#$fHbtkr>j|S-0
z05)&lBp;2Y<GCvgn-o0fEcy5rWZ$+<wqN(QNbY`8(bVGlWLX!|-&G@A0r?ep`^}r=
z<$j}ZBEx2>n&<SAmuOkEIGf(`HXl!y{3m5`(DPnA_qJrP3uL=m*F~-H_$OFyVZzKs
zSvG7J5XC>Mzq!06GyhkA9WQJos*NO65~f>o<}Yi)yvI56!T>tJ(!L~u9koL=kPbwo
z_t%eTUmHOmq92DST0f1E!QU{VeX9tD5W_e`(S~V^ys5XPedh{hYDVNkeGm<&!!7NL
zCIr)9MDNgd5RIW@EbUuFh@w%5#?rBf-lOkX+T|Cf&?$(f(y55%(z%xQ1QO=ad5GrI
z`G~TCW`m`D&4j%`_hQZgnuGZW&?8`J&l}+=(4&|OfEHjr1@shH+P9uy1!~278t7@v
z=YgIFOM7k$7l2;Cd=cnH#=VXt+kSNTeuaV;^lo}1TT;DAFCa4OeHlg22yZ5MB#lHg
znvO=44>TXV*<=p^J%sr%(8G+KWF=_>=Gl#x^MaGyN!^yX9LEde<>L_*fxHM1{{iwp
zFmEODtpttlB>J7i#Jrs$@-uAQd8_F}vV2y&j28lFpqXvCa2jrAvz$mLn%UNwNoShb
z_DCfmsU%#=HR-yTlCAZYrDQQt6*DvQ&FYBim#@mi-^~D;0X}RC9RPX&e3)t_Kuf@f
zZGj4)72v~`_!J2|MS`W@56F7`_r>jJ;zY;&Xg_b3>v{+M+iVdH4u-c`UV2hbZ<d!$
zFu4h)N)CfwZ>-<qvgM<myr9;pO~_$ay#_h#YS1u;9y$*ba~PltFfnCAb)hDvY>Y0(
z#FUNG#hD}pc_AHi>0n~Y><8U`FfnC{L01eWrc62L%E82xxdOpgAet?G1E?FoNh<Xb
zK3fy}x^GJD!VA~I<vMsV!yDjo1M@H7@(boh2y28G$q_Hyg|NGr??Ko-#^WPxobPT4
zko6E1RH&6qK?j9{k}2q-a8V)!6&^~Ypdvtt6jX#Nk%EdCB~nlkr^E^>(m{z8RO|;O
zR!~t4O01xw9F$DKN6_OD^p>{5)TJ4T)2yd{6U#*s@k%0o(f}{4CgPtz=1Hr`490oH
zD-YY}6Coel=aWhK*uI8%)nNNtBGh90S~97Y*~d>de&^(WtU=uLL+B8c#bGoIW%0Z8
zU6jQMG@(NlyQUHkDd2>Nq1L2=eKTx$fl@k{Erb{KLL^WD(JVS^Fq3{ZosDP?or9<n
zXeA70GFJhu!dwls8uJEX+(5jg;CNvJiP}IWFis|($;8aKkQfWGeIbb|#P&tRvk2SQ
z5n~;;uOm@)*uI{4)-!u4Zn4C*b&RaYT_vikWV95Bw3|i4*1nOMy#Un<2$9SZ!%8L`
zZJC@TCZe0%4VCWc^67|5L0*ar{dyu_&k8v&6cBj<<|{;gg|W2f#aP-}R2IzzqPjpL
zrD)bnvmKFV@>(Kpat+m(S+?3zTSWb7e?&nv2<^I|bSR=>bQq$qfqo5UmXVo2GcjiY
z&0@TE@5<#7`_EOm^Fk90Y63xuvXw_R!5GF*K>Y*;GJXo0r{I9u_bCix{0ubD@OU$5
znme*zGYn_;Eud+^<4Hu5M3Dd0M6<dh2d!o;XzQNUY{26ih~EZ+{3jDlG9FJMgHi~Z
zZz@rz5;VV5;+0AQSbW=vW*h$gHsZUDp!uc|O&T80BATp@?3ab($tId?Jf251d6@Hw
zCch&G<>UDl5=|i<FC>12*nbhx6yfnxWY8%*Un@~t@qDdBuo4TKZz<7~;_sId-%>o^
zGNLKN<5fgc)sg+Ga6HvSQ;o-Kh^7W}Ez#6=<e*wS-#VhH!{c?tuMYdKCz^UTE~S;2
zj;)V_EE(4l)mr9u{gZbF_||T+6_c?UjLm>*U^7HD1FC^7;MoGG24)ds7OrVoBq|Hn
zv~1#;&1xDiR1sqpwyz>lRoK3ocvdre>9#G#-}<!5;(rCISFE(9H=6AtxyR?l(q<#K
zLB-5h?u%##$amn<wu;DCvC_p02Z{V3=1W9=iE)1Ds!!ej2s$a={rb|rxaG6c>!d(=
z!C+_5NfntFjMUhn<qK?r_GGatOYWb07&cLKXrUHVyhhPc=yp1RPC&QQcp8sxrzs>b
zg{{8z(};Dw|Gj9J=x{gC-MHI52=pNCc8>u)hP&OfK+ocC*GgQi#3-df`EQ%g{Wd6N
zt*Ea~u0x8Bl#fKT59Ir>zAr)k5_2YzXEHr`p_0fe8Go<rF<pJ`zemLA-=c3}eVz5L
zNMC1zE7R9RO-SD-;Q9pIrKqLeP24=KCmC$><PcR336spEcj)%!G($RGxDmf8{r>G(
zP7G)d&^=hrJfL}4&hLPJhvhs6^jt?dL#^!IC$*SjV6Dd^EU?2wb(lT8NR2P)_U4F0
z3XB&Xf#V}^lOn(EDSIYY_S@Ln`Kr}gDd~B^LG7Sr%DJdrv`je<wTG59Tmk9;trQk7
zgsMZel6t%lqmI#{&Q~3$W$jlwsMA5q?z8(ry&tqv7r+a}pe_b2?tIndpp{yi)`@5y
zfnJ)3WG4#KcILyRsUWHfrl6#RxMpp~Ka!yor|6;eyhpms3ttk|mu$_YfJL*mQy*q{
znW!$a-m1Cet1=J&i$%ao2TX4sUtXzV3U4Hyy(5-PgPXy~bTya^0wS})mr*Q@^_C*!
zg-LW0qRDhJqT@i1gSVsyFPs2+0`p0rCmD+;Rr9B}2DWHn+@h(iAA}(lLkm-vPK=hD
ly^SVM>cx4tUT%A}Kc;h>QM$%)p6nxKAKM?-IL^$j{vX;p-eCX$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_backward_data_specialization.hpp.E2F7DA4511290758.idx b/.cache/clangd/index/convolution_backward_data_specialization.hpp.E2F7DA4511290758.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b8e8c6271e93d2c8675bbbbced4d14de8f74a8e
GIT binary patch
literal 714
zcmWIYbaOkz#K7R3;#rZKT9U}Zz`!5~#Kk2=nL>;V45~o6iaAp+9Q0x~6kvN`<F(NB
zR(Iig%M(f6?t<SK54}mb#XDO~aq``DUGMX5tqO2lGhN2$e~bS92)%n+tp|(et!<WX
zmfpnsk8y*&gOT)!3ocIF=k^^cG}`^}`_zbk`N^v5yt;UE6AL13qNZzx248)W!W()%
z;T_8=*S{gfW>2q3Enm6n)v~-s-7G!J{k!L{e;z+`*)hIJ=Qwkooo2KU7bx<08kMj<
zMD4udO99QfGG3BWx|4$Hd)EI7txV09)znNC|E6Ni7JOlG&qD{#X7=xgrm42o^6;!x
zz8n1M$&#-x*qnu`e_mE7FPxC>7WpUa#ka`|eG4=8HQXsP?2>)D_*Z#e@v5LHtEO08
z*Ro?+_UH7^;>z5l24D!42Ye4bw|ZL|4+9rF1E&CkfI669;ACQC;AUgwWSl%t)j?AI
z)n`@)W}pI29!^6r!N3XB%)kItaP;o%q~~l$!uT0jIT$#3Sa>AB1Oo?H0mC0b4x1CY
z;*a<lSb);pEZp2+0w!IQnpXS<s4|$-b#HE}Kp!IyCqECr5X@OHnvo0aEKilP?uomc
z>tK#z;Q~7bCd|VrpeGOv)5F6l!YQJM;E8gI!g)K~rdF(Tcz%PKg;T^q#1p27g;Ug5
zG#SAY6A?3l@nHUB;sW~*<^&#2Q3FvIm?D^Ofr>z0MJN(65HW`-DoV{s1O-lUQF1Xi
SkY?iC{Xnnr%e^x<7#IMi3IMzS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_backward_weight_specialization.hpp.D1A61821B9657E4F.idx b/.cache/clangd/index/convolution_backward_weight_specialization.hpp.D1A61821B9657E4F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eca7cb0f26e6c2a8ddf2214505738d8e3f7d0443
GIT binary patch
literal 916
zcmWIYbaU%rW?*nm@vO*AElFfyU|<jg;^LB`Old|21{Wp<hKe~;C;0mxHsCq>IZ!Rv
z>2~XdY`ztXI+PW@9xyDDUEe!PHE_~8)+6QW(LWa#HWY45*`NOU`SEM3uRAT+Y^bZ1
z{{0+}r*$>+jedbE!Y{XZZ&kmN@P%7u%Td>w+Z(>tmz%xu(cisPmq%@*b*@lrl9rrf
zY;AcdSJXwDnt)YF8evasB`2Q>^_sJ6k;dyyjVl6=<!pQVbyHMb)hi)u4kOkJn)A34
zdYnD<o){jE>u@pKr7(ASfs3%y#W-cokL@+jR{uN~CM`7Q^SQSj<$@u{?iIgx;}X)D
z@qC4{`}WUzPRSnMM06zA>CSQ#JU;LLk_T($fACBa-*1ufnNjJua%Yda#`L-0*DNhj
z_|Lo3j_p)Yg43CYo5~xOU(aDEuFOr^2MVv{>rX|!sQSv!z{bVE&BMYY0VWtYfO0@!
z7pJpA^XknSRt9EH25wFsPGc~^zzLRR<YtU7om#B-{}UfS13OSZ9}6D`n1Jb@Eii9&
zOYGerJPbTQ{Xz^vnqY!~lZlammjkH(6yqW{Lp8&5{0yu>{k$x^f?xus|5~Z<U7tU^
z2lyFSfYRJ7+&};%VA4gYX~lWKn1=a+i3{v{n5%iX#f`+>U~YhkvT=d^;Hgs9J#m+F
z9S^sVflxS15f8T*r<ehPC(bF3#FONdMB+(tO2K)-oUVIwQw91MdAJ361chJ*z=8+p
zFi_aQoWR2^Wh7+=Qv{<Kxxj$~bJz~IsTJ!Sp5I_);TCfe^M<Kn;g;Z+P(|=0JtX0-
zXW^EzmvTkONsCAu!FaHc208^4%m}AQ8cEv26v4t9s0b9!2t{H>VwNyPMX5Q7p!hE?
WN-pLG(oEb38%3vFHd0q(U;qFW<UDEs

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_forward.hpp.8E3EE29C519B7A7E.idx b/.cache/clangd/index/convolution_forward.hpp.8E3EE29C519B7A7E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0882c97343f46feaea6eb49da45d751de553e366
GIT binary patch
literal 2910
zcmYk73se(V8pkJr5CRhuz%a}WnD7=uE`dORfZHNV!lMKU5sRpiNGL7AC=e=ytx~P%
zO3|vj?po}s6lo8ufOXZn<s3vP<=CQq+Qas-_`t4@T9<lyeD4JI&Yhf-$^3G^``!Qd
zzu(+3=NSy8{sfU-nYXdl-caF95JUj;t#7D%S;HrYPG)Rl=V!}mos$~rMA6iF-Lb$=
zBOTQzqzXS<_QA=#-mmWLwRidttan^&T+;Q*vp?i1#h1us&s(D|r2n|{?!E8Vj=7&r
zAK!V*Ut?-bxBGN==01((9XUw1hhHZR`A7HKHNX7p*ac^5SL3ZKzx5TzSx=VV*t>3%
z`)Ss(!H)L)>Cc~0=coGbcNE-uuX4?^cba;J%t!lrDDN+3QhWA$4R=3U^4<Bnt=m_7
zE$?XCkXdNjJXO6dbl23@YiTn+j@ALMm;V%=zvXa;YR879!$TY86ZXJeO^-`%m@FsD
zOCB7oS~OnrBhfd~JF@JGrQ`js%YWVWMc3+Uzq@~gCVaPN=}%nKwGFrq^vnyHH~95H
z->-xH9S4s4rykCkNxG{#lW$)PyepTb4*&P{2RzY!-iKoEiV;Izh@UcPMbe2+)|G{;
zbHBY4^mMzpakxtOaTxe!Mj}!*m5n8KHa{9KePDZBJ@(`YeY?Bw!hbG*{?(<B(7#NK
znNDZjNd2i<e6~n@_M-Ug)843-)%SM$5BVO>9o-yJ;Ppz=A5X4$J=|h`t886B$w29!
zC+cDnhH|Cb3u5*V((-Fpv){ch(oMe6F|E(){h>=dE&Is5t-a>$g^GXle0ut4$>i;l
zJ9RHqoqK890Y^|)_he4pri$MD=T7ck@A~IEVa3Jo9=Vqv)g5lB%KJ<w+}HTY@2e+X
z9E*#NKV!OVG{m(9*4%7yKACAb^j&cF%FOF+&i;pc-S1y%{Y_#mQZ?5%)!O(xf{<+e
z@bt*|xPv6-1u^Z!e1Jlr$RtDp9uXGE1H6TWqhBu`JS-py2@0SEbOs^v;m-xA+cVw!
zyv}SPi2xLkBuHX1_<GL;=sJ(z{kAUYZzLf`0WyIs2WQADg(C+Z?;Hx`lk*9%O1dg^
z_Com139?xh>*7~gg8lh^%&^(d95{PH!(SfiF6@q7dy6EbD1E#jo)IYYnp^*y<Rk9D
zs6}HweE%R3FbECNvlleK&R$jDE+h!QhlhUHQjk(2ApxbQa#+=XGE!y^Stu(exn7>b
zIt1iKxtT*2xfN3^j7?|F1`FXr4#mMZ4naN4VT%IL2+bU_Kr1Fa)TgudgAAC#p-h;`
z5?Yrpn%&$thaVmK`&5RK1g}b8H6r%|`gQ=<@tfy;nEF)|#5GbHO3R@XDh-odC5O01
zxkj$#P>MVaQH<c!zR;R53!elkp-MKlu>?VaN?|2Sv3nBF?hw9I=}{0(#c@`NmdBw?
zB#4BONvI+cz-SnU9nnEu5_ZIvYkL8ehvqXLc={K6;FLNIC<skuf;>HpgI<%q#u|lm
zF+rX_#zE_)>zATT=$Gd=+smj23sd(vgVpW3nG#JFlNI|=WGUj%Qp-{fl^M&-=pi7%
z3gZe6l^e@hvX}y6e|jrZz$6bA1xt`l5&#j9p!!Lm6e%UxKeb3L!TxDQS_v|N1Xi&%
z2o+-8*y&s|_|a7_5;SNV^yqIAxRPBN94dF&5FyQczu5;CdBEh2vc{vJ%9ROcVlqzu
z+q3BkDiVv9E98+_bOemxZn^?SV$mrug}dpCY!D^47iJB%4ji>H$x~r!GOjEQrg118
zrgO-h;7&maNZ>AZFUQoPY0)8a7Mxn%|7c|^)6<E<M7E1b&@60bY0Da5=uOl%GPXs*
zC{78{Fq%U#FovbAU7MN<-y0oc4x$QGMKW5B1U8v%0f)4<OiXs#9>Jk>`w~o5bQNdJ
zK|8plcVy$pTDq1q)<8FK$VI!j^t>w%&vVmm&X`;8=HzUlTiCI-UvfgzFSY~54M&25
z<FcdJ!Ew1(Q_H&i_6=Q_iTHV(ae0Ao0qZLX>V@?z4N6Yk9KBso=4m3D3ga{pEf2#@
z6b{2#mzk4;hGDpgl3_BZi7K9}U`rUj`sV0W)vm3Kg89Pv?3yH~5!SGD$$NhP=Alh(
zLPmivCF3aYmCLXKrBbQJ&mQ!JGOR$W)T(iZ8Y7Ge_;@hO%n|5uLNl22$Md4Y=TEh?
znmx(T!^WIHo;Mx}x~>uYDwI$=Jx7ErSb=2%8BlO!%9IK$Q!P^~(4r*J%CrhB(@I+-
zQ6D@!wf`K<xbOjwoO{9kU;Yzm>51>1Bh$>LLluNL6VyO0hf-h~qB^^y!ZW+RZf*VS
rK`2L75JWJ(rSL7g5d;x}|E2M5D82<rlG1Yl^adylmCLB$5SsWOIdAnD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_forward_specialization.hpp.29B75FDB4D9A7ED7.idx b/.cache/clangd/index/convolution_forward_specialization.hpp.29B75FDB4D9A7ED7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f2c450731f828dcfa918f4fcfa3fcfacd04d3fbd
GIT binary patch
literal 948
zcmWIYbaPw7%)sEB;#rZKT9U}Zz`!5~#Kk2=nFfpu3{gxB3>9;FCtUPmHWYCEtl1j2
zk-zkg-7`lq?b?t;dxa$@w@A-Qb27RecC_BSbn8;35RS$(duLjfA36JC@1!=TX9s%@
zO{+QHcWBNfKlR%#CVti5m+>DxSo%+*dv^Wo<BukJ+B^y>pLg_5<*|^rtHpEQ`>k)A
zy!nRs{IzpAmEFGv&%HS<T66RL@X)(gxaZCd+&Sa$t2ztme-*QCZL!kyyH#i>zU6z0
zNw8LG;<X^7HXqNxBVSt8mL6SetnrlfXV9@r2`_ED7&0=OFQnyizZX+_YoIB+rT(m5
z^FrSxYsBTQN(DRc>q*>?WZB$y|8b(I^$X5Z$CVqRzOiV{SG=>$P)<?9Xo*<ZF`pSH
z(w}&mwz?d-@U%p#%CuI4z3ZBO*qRCak^gyaewnr@T66!~`KPn*1Ql22ChY@;pM>Oh
z_r0gS2=X(qa4~T6vhZ?)2?h?J91z4kdG5);Q@Wd#ftizmo12GQ2TU+<f@K-G87Eks
z++ZHIZ7)9qJ5av>ivS0hfazZ(dr;ucosvKN46HzDJ{CSfFaeX^*gAJb$=|t+JPf=*
z<AsHVO~C{MClezB9|zF*OV97^*>HEO7e50VP(MElzXX_o=`Tu6E6xMPKFk+PTwvG3
zT+PERt}Si^vm}_)b#HE}Kp!Iyw=j>e5KI`Rn2ihUHBXhY?uomc>wt=EgkxZefF^N@
zYa)0OoDxVpNlr;5o)o7PoCgaYpu<361G9^VTS{9>A7&gZkbsInA%svQp)FwnQv{<K
zxxhgLbLI}WsTJ!Sp5I_);TE?RcY&#5;g)ce2te>8JtX50JSht)I|NT!L>eAOu<!;t
v1r*o_r$}l`+QYOIrRF4p;=i~kxmX-XGjSh!@KyJxdW<h44+9T4Wq=6)7rIt@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_host_tensor_descriptor_helper.hpp.C7E0FAE3D228F16B.idx b/.cache/clangd/index/convolution_host_tensor_descriptor_helper.hpp.C7E0FAE3D228F16B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0bd93c1507d1e0d545cc35d7521cbda5456923e6
GIT binary patch
literal 8190
zcmZA630M?YwgBKh$4&$-X$t6ISIc66HqqADJU?6p`Wc~A+R$YvFo}CKMom;k#U<vU
zBH#j!5OGkH009gb#VCRcK?YP@5D{@7aUsfpqXd2EQ=bb{n3VU-@L|L6UFzJW?!8rn
zhmIWie_k3*@E4&==gx@!T&vM&c;(;1=moPIL8B>uTccT;aX4bq*KQWK?`N(Mvy&p6
z<E*QTxLs4XhbMHqnE&T(4c|MahK|^Hb-wv!+n1(mt(QuQcJCbM)!2~n?QcWl-Dd8)
zcG>LY)T4OxjcvMtR|ecTBh5SzwPlYf<Xz!-X+iW<pFPv&d{>@T5j-aImxoK9<ScFY
zM{e3S>z_&nc#5}$i$BCHzF#Xn`64QHu&Gb-FRz@wDxLqh>-~ksUuO=g{Qm<z<3AZN
zeu_gtzYk73&-f;C_gLRq%X1{pfpvfHcJe{-v!7lp%F24YgkQ&Rvj4U0y)>Io?xsuo
z#qnKJoG&B}yO!HLb-3=sxBAYxH)o~sFJ&J_jg@D_|K=NhF0VPVwSgC^PB|TL3JRIg
zqj2IVzulWEo^Q?k^_4NGa9-fk&nG<b6c&ch&z?SZ+>+Z4Wp+z@$-xHgnXfO*uI?Ce
zB6``tR}0Vd|66K!!hDZm$(|pL_W16Tkr#rJ(<by!dPne0|IpLr-P4bTIrp&TtiSi`
zveRR5q{}_OWiDU)hL<=>PUm+!p7i_0w{3uX^X0RF<8Ey~?mV*Q;;D2^*#D0Hc2@qR
zHS4DI`mk(N<i)OQr<lu^Uk%QXDy@q``pr9d>u|R~DfRuK$v5W<H?Ez?m~(r!JvF*~
z?t6ESPcu)MHuL<jg0_>l%br5hq}b7#h2PAbwo@5#Q-0q;SwqIZ<TL`O;Y&eN3jTlm
z&^S71-5t9F4+HCV_I$r?=vqW4PAQ{&@w#q0XN}|U2w!xt&tE_G-^ghUN*$X~vl;*E
zj~^Pxu3g@FT}Q0rfuASTgtl{<o=T><nz{b3KQxYSf&O)d-1@JYdhQ?R#%bJoXm}Tc
z%dpoU8ppR$W9+6xp;<8F%Z2T|G#c!Gxh(ADtnJA0k|fD*3i8iEdJeMMByD>8vImW(
zsXaJ84#MJK(wh+Y6bMV9&VaBC>H-KWpsoXZ9ZXg$=J-aiH&VBPy_NbI*q>1+pgjRe
z`9ic8;^a5|#+RbKl=?W@k5kv9y`H)IZ%Jw6o_tZ&h2wkLdWNy~SAwk)!dRH+K{^lN
zs-2(zY)lU9o2zu2hvY%JZBgi=Ftu;^$og4-&g1qdnQW3RlGR;~QaMJdA-}VH{G1h=
zTAGz?L!==jTM<e{ESr~8?9g=%HH{7&Kize@H>+)#`!Zkkk{rJqTz7*v^9gW20lsS7
z9DfzuuTm$WYZ7|1^gMLWLtiyEj=zGgSI}Fv_v+x#cMF^5DY1^1#*<jH!JZ8f>ScSE
zE|_(HW#DQDEkE69I<Lm6<(IiE)2o@ZN~})1foDDeE+;^*#-deXb-7BNgicAwv-CW4
z$wR#wn^uX{=?e0y)qK$HkDIV;Pm6H%gP<4hAH6Xu;A!+pf#X*}&?*?AM$kFAgW{-@
zAt;$T1%gtjw?WW0>I?|Vpw5S&eCh%SDxf|K*0T_-M#}MZV6CIR1=d^CjbLr0z7N*>
z)U9A`rG5(5r_|5D`iwdrt?@{%pMcf`>Ro8vg~4i%a(p3L3#kvG^$>L_T1%;`&{{=(
z9IeNx&!P1kbv;_^nfu+F>>NI7_;O{og-9V0>NMi`!zdlb2sJEohTDLzex4klWHU=<
zax?BhX%EV8vfX{$uw~r&S;`HxTwWf@20aPnBp}`l@@DEBkaMUDP%dEiCC8VbTtZ!m
zawT;Q$~DaWZ=b0gV7i^5MB7*DOQKy3_SF!shBau>C$$}OXJ#qcf^0!{*7i1#+StN8
z=)u^kOTBM~XgPiZST?{QwLHh)1j|hr#OBsLu-v0Qh?avmNbQ}$^ZLch3+o(}h0s^>
zC4GGqY)3(2)m#MWB1EVj=6vM)lJq67x^nz9hiOhM%w6EH3!GRqh2T&~y$&7Lp_5t-
z$0wpgBK0M7xP(q>bR2&f9WFBm2mX2??CQOJ%2=)f(<%sL7m5Q@9Cb37lBrX`ltR4?
zOxvh4z?4Cq52k$T0x%UYhonAg`FMfr4P}@{OQXqxxC!i=*n%j!weml>c3Gxm8zGG#
zy_*X5RFKu^#g1bQ9v<QKN;bP>Cvzhc?3rwCj4Vs|ZO;8Kvj%g#3-992Iz37sWmN0p
z_!x7HKam(ShH1IDoM?r(f+!Kpi9kvwf|v*-Jqg4l>h&P5r%nTN8ZDm&Vj4}~4B}?$
zEg){8&H-}{EuRBo4o%MmF_$_I#60RU(3gRcO|5D$SA##bSPf&CPlI@xx(37=>IN`3
z(DDr+Hqi9zAYP|#26Hpb-wa|iO>Y6Qh59atcd6UK+)m54gV;{fAA$IY`Z0))sXM^j
zLCbf5*g?}*qPP-C`(sgzrA|kEIvUxnmyhOr^rsf{aSU?-iUrg=P~1UXjOJomz8J+~
znqGop3H5#y_fuD(xq_CjK(T_RSE5)+eFViL)HP_Xq2+5(tfA>=P&`9ji()PF$5;E0
zK9S$l#gXGz$ScS;u^#00bX(X0@)qhmkn^Z(K(2vEHtnv1e4Y9($akq9gZ!8}7Ufuc
zQ(o<U<#hA=Q-XTBiLkr$Y56YjEnZn1-g4<q76->D9a{imykksEiQt_G#A)E2Mx6uR
zIn-s~SOz-QmTK^>1`~4wcsEctgLgA^J9xKKcYt>XbvinxqfTuj$LFJWKAM<|(Yu(s
z0=+A!YtXxfdED$zTc%{j9#NK6gUvt=`I%_T#4xqmPm4<Gg0|S&lx%^vK(gg-1p7v|
z`cA4J@{LdbzrUmGlyhx!->lu7a>}%ovYnYz#%0WAc1oG*lrSc|D(-8g_I^@7vbNs_
z={74iy(y*7dUH~{l5K)Cfov!*L3+ui!(4l`^`}P}eoD51(m=AItO5HPwxLArJJVx%
z)0eB2SocEry)cvw`3CI10f(|7yN2Db;ZU_NqMH0(md{`Hy;8r`W+i=5hEf?jQ_Wv$
zJ=?3g{-r_78+Ar!b+T#s-n!mIeRO?@2IvM54blxF8m1dYG+Z~FXryi=(MP(Eh{AQ@
zM4#$DC7P_8OcY~`@l<<3%O`>{5s1^km`0rg#vJNuFjj-7+G|?A0gMgQ&0uV%ZU<vK
zbq5$bsPoa7kDh9;Yx!a{7E@QCv4Xk=jWx{Qevvr%=Io)Hm1EE@H0{D5wW7|Gx~Y)*
z5Sk8Am!heZx(ZEI)W^|uocbJ^&QaH+sh(NwP-Qjhd>x^#Me(^SmW35mofyOM?s|7m
zb$Tez+WMYE2EBpEL-6onIXwkWA}_&<$Vc!Y@)dlE`~*LuzCvH3enLMYvtT9~C=4XB
z2o@rL!JjBV2p}3P3?>>P3?T~C2a@&$>4S)@dMnWgVFW1`ECdsU2q8qHh0#QzLMV|<
zP~Oq0uYh>FU?-9UiD;}amMB7qAc_<siN*=zh{g-!i6#gWh$adXiK6w<Z*DJ+U!-3|
zv{=8GXf+6{!AG4*9KQyHHPmZCSWCSTgpJgxAf!@n0$~$%CJ33-Ss-LlXM>PUT?zV1
zcB^vy5zrr@J_`Dy)aOAs4?gPB!0{JAxIldogp1U-LAXtQ2ZTG+O&~N;w}H?`{Q!gq
z)DJ;;Nc|jy=hQDictQOVgqPHrsLy0e3&&@nK8rdV_1V;WP}qY$Y{l7&!d~hk6pE<J
zP$;85jKX2+aumv`Poi*=`V<PMsH;(^W{xjj95;JLm_gaF{3JiJA!eg38znW&^?7O6
zqyFdawJusd$|I`3I@`2-8F-XIe|7Y<d^&ohV}G^SR!i}Y#c`I-S9uXq1gSj>>{)EP
z&A75g8epEDu4D_40?6jQ7VK-;(_p4s-lC$V+mn=RW2Ld=C~y*`lWcQUZ((h>*Kcpt
zH^&^m51jXbMU9=~Q_wjDEo$r>e;u8#qeYE<`?a=d{yiH_%GH9UU=sTSkRGrn`TV~u
zm-Vf^Co1RM(;%N_M-5Kd<m3kG7LZ%0AA$Ucx&!16>K!QWU@t&8em~0lsgIz1gt`{x
zTIPeF<+SS;f9I~edyWM|EO@boZvevv>aAedO8p}kex$AiLoIl*j=Bkko7DHfaF6;4
z7@knap&<@Q`JHIkiC$_)bNoRx9Hc&mhGW!qXsBa8SQQ=cS;l*hl^%~(h7$r<!?uED
zEA@|H`H{L7EVU588uA1zPpIS25{D#xCt7x501NaOT8>fIp{0)bP{#1xTk83~%8ZGW
zBFT(-0n!ULV@j_@`K@Scu`AgIOM}U5Zvv@_jsGt@KHGQi>C2989Df!}XCa7%T?eK*
z>RVvCMcoLdM(X=ux=-B-rdH~wV0udZ3{20c<Ixn4<oXF{N?@)Gzcnx2xiwm8kH6$k
zn!OjLz3d==42o86yS(m+lFi-bPM#*BL!&=Zhwzl)zgMRIwBlqBEgxfu5!C<gJa8Hk
zfjA8eY1BDj$f2$VLp2C$mvtUE4Gq-IU}&ap2SYn`2N*i2^U;uxf;tmA51fW#>IyVe
zP}iWLhPlq~w577Vo6VWym-0(p)n+O0n&cSrXq*UgB7O8t138U47vx;(D&VWYRc$24
zSA$&5CXn*-50r0`-VAaxO>YOeoq8q8E7_wm$LFB(?W#7q^KnwnXA2Ccbb?$=U4e21
z^%<1UFxT^GYyPo1dblzHERu!HlRF^YVYkQmF^&ti1Wvo7WDAu-$?fqFq=zipi=(DH
zu1|LEd_?n<JjrHo1Z+pZ&N}4+NEg_1;^ppD@5}2pj#p~;k$gxHi$fRN)p5CbIP_|d
ze@<PeWEv-pBbiR2bc)UB#;5ZNpAU}|lx!2Fi6qi$l&V>~n^QkHP_yKNeM+Q7;7|m;
z)XQ*uGCCw<FBWS9Iy7J}wZs1z-dzX_nf-%OznA1inwW*QEVQfHURKnawgxWEc$?$r
zXy&-F+W+%vhd^@(+}PFr^Xh4+NkccaT4n$Gw`i2Vea`v3^RKN77ffIHr_NtoIzN8@
ea#nxQXmspXe(0<6sn3&A+_gGq7gvodOZgYtT#Tpy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_parameter.cpp.AB24172BD3A32784.idx b/.cache/clangd/index/convolution_parameter.cpp.AB24172BD3A32784.idx
new file mode 100755
index 0000000000000000000000000000000000000000..386438875cd57d626260c52417e57c52caf2ec7b
GIT binary patch
literal 5098
zcmZvg30PCd7J%<S2!vergb*Mkm=yxD1VT`vvPB?*Y$~!`Q5Fdh5kW<16<a`!q7~~_
zMcD!sTeVL?MYOK9bys-S1+}<Wt##MF=khMrn;GBN`+X4pIWy;;IcLt96Ov;!nj20S
z7MUAcq0cLyOv5nDf&3{epP^Hc;Af3t6&t=EuP?cxw7)Pm;B0o<m9vKRK1DV8F*h#j
zxA(ZjFZqmBV5qlBY5j?Fj5zmx|H8S8oum)Cva7<MJQ`4kUE?iZHt*VB%DCzo=C`*5
zIgKkUx=&kmeYWNB>(_#;f}_3vHLm5-*tFPFzp;CdcOR6GI`OApOYM<vpV*x<YkLbH
zuk7MQ?ydK&KQi1?@5nkK;^u5fa$0?10r*kcbCK6ymOcA%K}njvWXZVqJvUD*vVT*)
zQ5&y_u#qn}eD-46a?P#7uV0nVm@{MUh}g01Er)%NUa@>y-qo2s?8^KdMpf9w<gixN
z-Zze{HGE0?q>)pezL;Bodc>so-j$awowG6oy)9R2r{0xLYMZdh@BB(;!A7$Yb;o?C
zJFn@!QL58vzEAt9bN_9L(Qv$zt<3N4=ror7(rXc7I8pJa=H!&7wa>ZpYu`TeHLi6l
zyx>;5<?W*WM6ZA;Gmox1ymZT$TT$mj*C-eLRN%B<JMF?&R)^K0<6P07l3&}J55~v_
z7Tpd!mZ=L|Se-s3?bs@IdS~Iw|Ge0{OQgEsqSwWoQgm#{=2ctYxU=zx+Mg9uN@{Y%
zzf>gbdsq-ixSjsid_$=uX}0aGF7wXO_A&Z$<5eDS(!tSEZ(C;Hj;RT5S0Yusgp|el
zb!ii3X4Iz7TE4;OYiX*kbL-Ast+%fVl4S|rg<mbNdpdSn(8Oe6fy5<|?y&3B!t!$4
zafi}r8*hDaVrg~S%`R(~X#=X}-f3CS61GhbtQM)F9&Mh<J2Wb)Dah>d{lahZ*6-;Q
zwMV?PKI|K!^<Ru9Fm_L#dtgq>^wUdIW9%}STwYyz-|EYP_qQHbsp1cf)9<|aDye^C
zT*m~av4p+HsHy4goc3z1F;rT%>2AM$iSA(Kfee-F@pT>F)m89MS<2nF6)raDR_)p_
z^yl+6R~*0SYmgto>u=Am*y<7&v!vEI+{|zxsN!ar<n*FLXS;D0G2nk{_R4MFtvEOQ
zTzX_|AaB}0{DY*Z#Mgg4Ebd?GwA6E4))|_1cTw)p;0)~kK55x;=Wi2K(r9u2IU=e&
zs&K*-PU4WV%WjS9{73(=OA@r;@Ru&Lqu5XGW$zA&(nh~q7IAG}-Ot{0iIEQ~mjwIR
z+AWy=;Qh(xw9)R?(OQ1srV9h=fo1DGw>*$()JHTs3~vJa`Hzm4HQh+=TN-_n+5KC}
z*W3Qk)vDfIPqCR<{wgz?zh`}1))oIfPq)5&)crZ{%pPy|<P-G`H&5hr9XQ)l5nlD;
z$%`kotyy1gpE4$;E4@y0;I5Uq_8cqm)Q|gIe!TZ<CVyeOGTOCfhaCu-TDQN=rtei+
z<<>{ycMcPu-_y<eZcmwe_OjjL(l;k3c;z$G($~EF%V}=fkcnq<ug?9!ta96bN*BDb
zt9-ZD+h@1d{Bp<PMdR14&b;tCwQEsg{fb)q`q%G1@AKI`)j|Dc^uG=I#ZmLYpCzkv
z%jW2(43Kq%^X;+)pLJv|#j#207?Fr4`ebF|nLZf9+#JhI16ja>eQE1T@mKqt%~CTk
zg2`mEve-1X4Q53yCZA!k9rLR;Y|Fs0DdhUeZj<F%gf-!cF=zlwoj|T{9xZk5%~Vy`
zn~f*ehl|38X9e&Af~fTe7G@js?uiC)Y%+!s4ZyYmIA`VB<~m~x3o|TdBB|k*YOQ5L
zpChSZ9H~LW(eSer_6iqD!@JZ9JFLBJ<=+N|F~e9{EE>yXAT}c}zpU2=!)WWC&3!x|
z$;NTQneOaN?F%P(9=xHDgptA&NTLW)G9>XRaUvv1ganaE!=z)NOs1B_L-OPM`9abK
zX``q;;Y55;JVK2?+z5gxr%=HG)*m9%v;CgyB|r5W*WRP0*!{3PI*n)s(q`aA&8HD9
zK+*zakXwPM6-Xc(fy9W)+kl`A41w}?;Mop@6jL^vTmK#S^Qt*Q-^<Q8p{A?tD8JK)
zI>4+0Hq=Sr+b#%QY;|$~;Dp>k?m_AB?4D8Zuqvz;&<Hs$x1)xAet*%wc6@JvIZn9H
zU96~KoQNG7>kqxu2qrdy49c!y9#4)J(I-E#p%I&a<tD(O1T>-%ST=sj%0{3<<*|)G
z3%LncHlg#IfNK+Qhw^4%*^J6tfMpBHt-!JsWh1aOqTB{7+fZ%?mhGRisvQiY-beY6
z4ESIwQ2~z(yU`ba>A?;fPPo(E-Jz320+AAu(ncwS)B@x!Ae_>y>8W{C{4jqLsX}9+
zVM5Qwy2lQI6z3Tyg*3)<jERz}4j$kStmq-*mui)2O&!go-ZL@3O?y6;9445At!em@
zZ_Kgw+va3gUFoi9SXHsZ;0v07cN6fTY)oN9@6+emFEhi5M2kcwr3)vL8Oe4Or6xE^
z@2v>4u)zs?fxQIo*FoTbkfXp6At!+oLM#CbA+~^xP`E`nlPXy_QNSp$qhf*+I))CR
zLPjA%MT{bZiW$WS=^1(xO=wYc=CCE>$USOVS}t@&JS*NrnN2~C*D`k|k&_yLMFYSo
z4-K+M0}xZ3^)VxvIKjl24Cssxb=s7!VOSW0N>`ltP$Eq-G?mH3c2K**nc>{<A<!8b
zT*HR;PIj2=3SCnN-0OfCas!BHfN7R>vpTIazoeerPn=zx8&opJZj6a0ZQk6qB{_7e
zHBRL4bA(V=E<YEcJboTR`TTr@O1(;b;ksF}Sw4^|C6(Tgs(`EtAkrEzqy|W!+yI0I
zAf>qA(e{X-|Nj1hOdbJUfLt9?fz%&5UFoF^h13SbZLpM=R*X2~e*FAfvSzFO)d7@`
zNm8jp5sFYpA{3|AB9yF7K`33FK~Y)9`JVbk*`L#KLZ}jkK^sINqT!Il!D7TdNr(iI
zNdu(8aHcF!W~wWdxrUJIy3jS`=6Q4;4^9fU3g$zKWJV%qMeC{!pu#*a_t)WlcdKU%
z9{B`q8Z}H_eF^a<`Z9UVh-bGg%gDFz>3pQbNdyv;DLA2sR%F0;8iCvhOWcw;^XfmA
zzgKyX$-rd_;53|zBV@z1LCBVCi;x}H4k3HFeIV50Aa_8>QSOM4liUd*mYjtUTh2yE
z$Pt>ZM4S+DL<osFVuU0d2|`ki6d@T$hL9ib=LGFkI4Zc1!tKL3Fokn$bJ*0O#tE{`
zLMYER521Y9e1r;E1#WPij-^AWkX49K5vvHHVpcIidY0Zqt2hgt!{@Q)k@rTjeX<i(
z-6>Pa3-Cj$__0j2*Ud{||KXWQUME~S7rjmtvBB^_H33c&;8MkLw`Ntv)=_saTGNOe
zQI3qVpGM@0auLcC<sp<W%15YlXlWo^H_LZc0HjKJr5~g!;9CU{X$_Fp06CN!fXo03
z&fd)vcRYCg_Pqs8xY6B^oDvFzAuxCfp<)=Mc-MGSPT^$d=hF-#sl8I4!0clWJ=mKM
zK1UqS8o$5V(@*A=C*6|+waEms07%M_N~E)G2a<LWLg_sA`!`h$_mt->2D>XkBib%D
zg78LY#GpL25sar6(ugMD*aTQ8i<^LyVk&TCus?M5AA^0|t>Olw!q&o)_K9<7pY%R7
z!h`PNN^OAlNztHvQl)91Y@9}*h3c&MM8LuQ)^X&y;?Ozhx$=sYQ|961wls|(hBEuy
zokJzvBm3@>-KS!h!c;>^%1~rNi-P&VrdmQHltQHl%9No>Q(#G1qB04}LPQ~ANTnX7
ze0Yjx@n(5Is&uP#hg1c4RREFJ08R~XfpP<IF#s;bz1=ro-kDFw$a8@@5+}q0aS-)-
zoCvWGF`11Mt-z}ls3||(xgNEw=J4%WGEV5lKqCrx1q$jtG(yMIAvsjYEA)r5B3=<9
zE9MoOu4Eda=jlzCId##ICk5?Z$nSQ=<kv>>zhN_S%7lY|WDfrN@Bo6tIPLz;=Jy-S
zGPAHOY8b<E^tr`fk(gKqlC2olOx(uS-ocU0a!PQHi;+9qC5+~{yKs3P>Jg*DBO;@s
z6|VjxWy5_%|1|0*@EjtRNWHv$hWZ5r1uIp<!i1sxz>rv1xBr_ysN<g|j~Sb7sZG(u
H)06%Oxpj-@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/convolution_parameter.hpp.98A4E91AB8FFEC38.idx b/.cache/clangd/index/convolution_parameter.hpp.98A4E91AB8FFEC38.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ae7630697725e62560cc58e42c83dffe80828536
GIT binary patch
literal 3392
zcmaJ@dsGu=7N1POFgRq!X%a}jgg{70FcC<DyuwlzAxK!D5hzq61qm)LQZ3-CqHR}P
z6>HsMeX!yYpIEy^-J)BcMMT>@?x|K-6~z`xp}yB$Yo)7;`(^Z+>_6rl?wR}D@80je
ze)lpZ(`JhgAc#q>%!O6uwWU4;L5#zHHMP~Vd`W^>>`M>}n-AtYDz8Sf7jw-qs4`RE
zczo5%zY!YCcAy_h*QS10FfHv+eeKtr{mG-8DUD|yTyEJG;<|OTd%?pkwHkH!TFY*$
z;mC=h^=~ekOSVVncF*{Ce6DPuvh9<q)MRt7-y6=%tG<S|^X-d{Xjf@dPHzh6Uth8M
zVeD=BUs|1(i_2o)OD{Qfnf!5UXGUS5eC}i2>RE3!cC-hbJ~KV))b>kH!ipPx^%?K>
zy;a@XKVLL{=@&t_QXl_v<?!gr(d(t{Q-@7oeA4!Z@Ty;S)n0S$o73D^qv_pwENV#6
z`cbyCq_6t*<Q3<V4L-m9dT73N^R^|r-)~SZztJ{Yci{8*5!2bCG2-G1ze&rY&)5YQ
z!|FQU?!Qua;|<pjJ7>G{JI&RSacf4lhL5%c7}knr?N45)s3S5m*Qc_V8XR_Xb^H|{
zN9hK1^Y)xtZcFsh7UB9s^I})7chr4a*Gm6mOU#m8dz^zKm*XE!)$AxN9=R4VAZXA`
zyj}74tA9$~+aGx|e^39KvKfqF-%v=yyp$!ADkJutocfZ)d2YC8+SqnxwCT>Z$1+j+
z=B~l?VM|bqea*%jbi?tFEJ3mMyLpof^>NC$!-Mf3OYhbG^};|`Ld&!r@1M#=zuR8M
z?06z<Inb$+-#d^k`fBBWPUm&4ns+!jF0O0$&+30>{wtT$oVpYdRg<;umCmsHbBUTo
zRWleNLHIP?fBW7*c!`({mk}tG4GrN5;fFg0{{r`X@Lt_s=le9Fkv=^^PTFZAD4~!r
zasng>miD%-)BW+dj1<Y9U6A@n13=$&<3Zo1*SagjB#X_5v0)rfK%AJl_nmi}_f4aT
z2ryslU#uV~9|0j6hs{q)`o1T-_ts}(Qit^oDuab5pzr9NF!{^YA)%NYgT*>l$MFQj
z(=Ws?3Y&06Hii^Q3B-vwmL~+ody37MFZeEC(^n_%u_TT4?W|qJ6VR`?yFWGN(TxXU
zG7O6&Rgn{T0^&J~ziBEya;|_S#aNuerl@!V;>E6pxbJ2qY!s6QERIk`#P9^f%cg8h
zzCR?Y5tBh!Y*LvlJOT0c485$|8NZMv10b<xl{p5)JGaiyEc$bO2TP8{+Oce`k|!W;
zU9&V#dMDH=CK)W&vRaNOAU^u>>z~~k`R@}k8H&Yvl|F_iAU?MLx}rDgSC2_DQ1Yy(
ztUoJ$mN%kffl%1B>dliZNnz~*wjhuvpxv>eWLNoh?Ep>apDGi`C3eXO3QwQ$FX60r
zap!k^dhJ-BnABqZC{<J>PeA`~^;x2v9u7zD50P6~OAv_9>&}`_p76JeNk6QuVKp32
zK-_&~@aN%o#t>psj>Sfm(aaMNkLE9Ah}PZfy!ut*BYEDhM0NShnhhckE`<I(9-xzG
zloX#d8In&UTY@bU+JFh_NsaFQwvzNgh2sm2kR{Az7TSRWKnW-wNRVlO5k7Xq2cC;i
zf4)3NRx%=>k;Cdphjs;2VdvbCg4;AolqLpxoXa+5PXHA_8h*fyFnAa6Mf&r&@>ohz
zCrt!Pwqwih&b~c!D@7v<WnrK`G>WBS8Ms6vYq-_q;Zmp+1~NmV^zd|(hqF_51{|Z&
zs~N9mdpJjy!{Olyr~(F32VlohyOHkkHBphM^gNuVNK?X7c~tuGs&e)NoIj_pb1Wo3
zKZny_|1il{Oe2L{p@!ts2+5Hbr<5zbIF(%G#RbcQy*Ra8?ZxRSJp%}%k%2NWkY*Z1
zP!S9a1{xVDBLfXfBNJs}U@Fikl8R(t0MRIliedmaG%{0W1_ngo4$HwZX?TGU_oh54
z*pmpG+!hRppiz!BCk66a^k|cL_~^2j7j~+1TMYNDv)w_`$mQ#zpQ9SNf?Yb`pa{PH
z{A7i;qLCX^Af*4Vr)gA76jRWT*m$Zd>$$nX`l5p~=QQFstrBa9Czf<cdI-b<Q3(9`
zVGXZ9uJEyYQo}K5;bZrNhU4nCz@nm)vG;Lj<QQ`%f=wt5?lJekhTHhkuXBpv5{+_S
z$$bsDrLIfXo*(P0vCm0UlAIo&ZF#oo@H&(gd*Pn+V0X1$Sbt~jllUN<B2FszG@^`?
zd2xYUpcfa!1$l9DPVU7iv<jog3eqAkPN`LTaVo9Kiwo8UdvR*5+KUVI3uWNn1R81l
zGz^Sj8io0VF))p2r1jG>uz6^t^V2baSsLjjdgKKPi2-?mLK1->^B5EoBk}@;#DpMw
z7!;C71ck<+kVGLbP)N)OK(X1XVVV+bhj4jwt+~litL81n@9R3}#l5ic!3G|c?u`qz
zg2Lhwy4NinL%g_S+-nw&X_O{S3xx91C_6g)W$*&hfQPPo^?FonYFjL*z|_Ofxi>6a
z0;C88PnYVm4GUSox83^`Zhv{<o3m$c{?zfpcwwI7c-?yzbfL&T{``K0W4vV8L=KXH
z!(`Jq@EM00o^60Ip@!He_ZkKRmw#%WQn9IF#S49cHI04po+Q=fvrF9(*Ho9)<hehB
n-Q?*H?QYthkX5xLYYd*#Aqpj-00}`TJ)dgNXNc!h<@x*{-c)8a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/copy.hpp.D47879DBFF83B386.idx b/.cache/clangd/index/copy.hpp.D47879DBFF83B386.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2101734dc29eb0af12376b4655d9f02581222f76
GIT binary patch
literal 3316
zcmbuCdt4J&7RToVkbpxVaRO`wn}|k0FhN0H2`KNUNGuOgKondh0!lz5U5d!dDhepr
z0IpS(7CuTJ&`?$ilA>TKuYibxC|Ja83oeVVRk7{>9cKR7Kla0XlDYG}ch2ved*<A^
ziP+o2V;n^g*Jy89Tuf4wI)Wff{3}gL+<SqBAhC-PM0TztI9xI)z+OnxZl;=+r|YSz
z2Ac|9F6Ql8x?-@m$+XAcV!5x&>4})~_o=7HZ@jxTQ2m7oe^XL&@>Ba)t_KgQJ3k2?
zrZrv+-?L50VuvY@?5)YRf1@9mo?s*UXwx>$>@fZh-EGQBzNPqqcFs>}*7h{{i%)_k
zgoWR2oLqVIQjqM}LpytIuG*rirmewA6@orR%-vS`K2hDXrFoL12-6SLbmQ-tycv)M
zW(kbO#5H|q2iBOK8QPwa{dQl`OM1w%!uXNap3n=&o2}n{Yj}0p_PE^IWrwF;8NZeU
zhFfKKrYejkuC8ZY{J8eV&zjuXWuErdhtp5RpBb6fjgA;R!YWAVzrZYs)*a>WoD23=
zP0F79voO#*?v{5BME(|gO8+@+`v7DF{4#!JL&LWIRauuM;UCM(+0U<dX_(cQ=2xuM
zpD5OTFFDb8+4X&h&g;hGb&K~T>^hk8H|g1$`~cl8KVO^fxU*;^8o6%V{O^wX>4Dev
zZ`Gbpm^!8|+KHx#L~U!9Ykt^nt{>|el|pklFdg(C?u(!P_(YPHe`oM#&eLi4M@~Ys
zr{BE3`7Te?_QYCdFl4yNH+KL1uF*p=io#p22H*JF{lm64@X?m3mPxOBJ5o5mO(p+*
zKi#@zvLyOc^~CR$$NI%BQ@mYQgb9{5QC~e1`i^Dg%a^$LgmcY|AK(5o+i<ANvZ5?;
zBGbe##KY0yv(T&E=7nyiJjtodDzDbs;q**4uk%IAC>j%*Nf%edC}RhQFI(?Uvr5)G
z9nY$^w(&md{drb-pTcat+hw}j59Mu^F1cg9+j=&9I<ov_>BHqkJAR)!Tj2O>w2V7(
zCk?@lkKI3Ste%&uaB&Wf_BDBvqnrD#tUto*`@nPhW*aVPbp3(&?-H`5hvRli@Om`~
zP2N8CcytIwbWj8{(KEqjPly(Mk(&Nuz`)U~K{569tLnKM3g9mVV2sT*h?a&r%-&Rq
zwcK5|-t%h|(Z*G30I31SpADj=rAeE;2}z9EE%nqw5REP852y^Hc<Lx7uoBqJhkz-;
zN(naeVpua+HB-42ELuT8jOz%FmBbcV4K7A8o*{1;DYMFO6@>(w1;OOlTH#t3lEgM*
zN0Q=g;wdDD)pD?!pFSoBM>%XHSqYn!5KIJv;=%eggP7!-AiN1K1nVd&uA6G?ZvYfe
z9#Q~wJomGQ!H=&D)GYtCS_4l8U0_5;W5u&_AV~>yB^VKEGv4UURK5Yc8{j~U6LvK1
z*{7LwXtz&>G$#keN<qC8=)@#UQ|&&_YZQdwK_wU_*b!EiYfmWBS#%EG-3msnu$qu1
z>~Pt+3iw~bp1tQBOT*ly_!D+Ba^n%gDCS}0L7@RK9DwD7I*Q!~!}}!r6kpHl%ox|f
zMeI;JCMjZ%+Eb_<=<T3GigW<IgJgf((+T>Px%rDwOlTprB}F`io}MIa4cHn+($1ir
zJ4osV({5NtIv4`GA#f$wKfu=Dmb*Mq4aG$6A}`V`!Y9I?Bo%B_fg52Im|vyqvG~l;
z0#|@|cK~P4Bv-sZnc%o@ZT%Q!Gi4Bj2CX524zDT6Ju<bf3-7iyw4JYx$OlF1fAR2*
zUZB&svSL?TxrfODWsYVj9@o+;r*IiFjv1x4%sJ*1x(1wUz$4O+V)ekOCpqV@x+3Ry
zKI$lnIj|iJi3FgSBioTe-fV9Q`LKN`<jeM@kRRKRLY=_u1dMdv1<WogKLqAODh~s5
zn93u-9HH_fU_K)GgpXS0MBaNlO%$_aShC6BMO=|7NkzaY0yeRNQLGpk#Z)c<MhTT0
zfzb$TGH4YTt6&AeCvR#Vl}szU@nyB2x&K4b%>m{C3iW_a4=g2&P^=epdZ~OHbZ%3*
z4|MuSK6Qe%_WY%UTADf*$&a)moTy_mRK_5vSkhhDu))W9fpv5tEdNV}_eU#eD7KQd
z(wI<3u{vPYfiW=!q6O_gw>RPqcvV)fu{o~H15bzu6=9^k3+h54SJah4Dxj-CkI+Q1
zyFkB7<$FNCM{-r^(g=@Bx`*)f*arA#exv+qd@QE_!FN0HSCEChg%jz!5bO)ViG1b_
zVBY{vguj|n|CP@1L)`z<tZw&kyPI~~2ySMAm}nFG*v!7BRRBv!t_GwU=Ev)WT;}Mp
zmnZS)vw)rjEFw(Ro#c?!oR&1ael55be4-o0Liiyfl0x~R6iP%BKOo1XsFXs5z%2wm
zb=KSl;1jd%<Bt#41noG3hc8CO6tj3Vo<ee<%R!$I`IYbV;NIB!vIQ2|MANgFE>21x
zo&yE&6ktIF_4e%-*ZKF)KF5O+nTuS>nMLp;C|imOcq(uvjPT!6`=VyjElu1djhwld
wPyXD0k(n<1V9zx(XWjWr8s>HKZNH?g-Aj|r)MYKvV<39-g9wAXWRaKu1&(eC=l}o!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/data_type.hpp.FA421D61CFEE8D86.idx b/.cache/clangd/index/data_type.hpp.FA421D61CFEE8D86.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd885df2c0913a1ce47a716ab2b79efc99c45da5
GIT binary patch
literal 39350
zcmZ8q34Be*_s^ZX%}czeJTG{%h~#Avf*=VEMU|?hXshU^i&m?;=%&?0tdS7=5_^Ln
zh{TdcMCzxis;F9G4MIW|B9e&M|8w5F^D_7E^Ks87-}&CTbLY--=FFV)hW6{zCq3E2
z<M|KzO&B-&>vt6o506Id|5sm+|M(U5NAG7mJSMEz`ugRu*J7T@uYO{C@7xUKMF006
z&N4Z^xc9~2vu2-~!<QBeu{ylJxw*pe+p)HfFMro5Fnmz?+qui946pwF`ts)e^5>ZL
z+^-oI6+L5bn*)QY_B>poPF(*(a^=gHC+}{ZzrCONo!a(4bP6orvE{pl+jB-lynS$6
zvv=Nlbz9v9ZTGd0ThH2b>iKaWEKBOMy+i4OH<P!&)b^{LK6Cz>GNEJ3L3LF>(EF>h
z7k4e((elfd-_Gw{xp`v47pD&D6ZBrl*6Ok^yZ>@FXw%1E23O7grMg}D_MZ<qCJlVs
zbMxS7rYWZa#y@wgWa`6)U!*+uQ}g@H{lB|=@2&hZrss}zZT89PtsTR@ZS&8${R0cf
z{&96{#m{lC)Gg1G)je&aYKKgV)N6PDxcGcVyOej|dG&vb-}`&<_=)3M{4*uupO?0F
z-@Gz1=ckANY<zRab7LK|ixbZ;Y;b*HVea{?<UcLSpy_SnSIl~&N5sPsM=D>LyWxeN
zUcKg|cpva{w9Oc5G7k($dw$u;?28vH)6ev3wtmXfCi~y3-m>$+?)Ns|Z@Z)1WBG#2
zd!?V9xV&h0+asaP+da7T%h}U%;@wjT)%WstbbPv`-PV^r8q)Q2_&l!%mj<2tujYfU
znRzdr+O>P!@%Ouw_B`X^mG$!0(^+%J<{j>R`lGC)r{8`xqhj`!^z?1%GkhBSG_UOH
zfYvV`A2H%??XEo&FMZqKx24rx|JteNHyqM8B{cDm^evsYdA;U+?bl0puEgGPe5`%6
zyXDMoM}LWbaB`-$VRy@ZF){ro^jkJ&%;J*CL3ggS0m|*<xM#Jbr+r84-SBkgjvK@O
zm)oM|X7j5RpVN%#TUQSqJGfKFAEsTfelTEm=Rapp4D550ez7TKUufO%@z?fu^&kK7
zsK4eOhzg2(Qu|WY{;rMw$<oxWdfo78UcQ4%CT(dJxn{03y+P;)lLCg!dA-Zdl1B6U
zzgRSPOr!2S5<(|TJw=lnsdMJ~52>VkpLuVHN829`6g9SX`JcYU^uqeERMVi)PkQxk
z5Hf3$_wDXC+qY`*)-c<GqnEl@uDlyKB~9-?H|zSBj+fuKlRbIE*6-fjy<^Dk6)$f%
zx#i;DAE*8~a?IUM1<(IEbbZP*_j`tSdu!32pEjvMgWm0Ci(H-d-l8*OyS&wH(AbS+
z%@51pd-+z?1ILBiPyfn3PuE`QbEIL|u&nVb6Ve*%FFk*EvTwY*4HRaNTlCwx8B(V&
z5B#|M%&>`hXUC51_D<QZMXSPVA|*$|o2j1_{1F+JUo)i12ZK#hUYh#uxk0O^kGe8s
z!qj<v?=%XYx+0}>Q1{^{4!-o`#&vhL1o~fFv2e^0zeWYJ+T{4_7mj<ZiEteMW`F9S
z?)wgvz2Enf%L^WyU3C85`tIMR*6#hj@%L?$|Jio>cfa=&0;V1QerW8ZuU1X|<>s-=
z$36-x^ceK@s>ThktQce|&_3_%UlF;!%cB9o->=U~F8ufEzJY&t9J1xBzn`1$)h>F)
z+$S+NuB0WjZThI^xZk#q_pYnnWN!21w-Hy~_(k89KDylFtMA9X=hKc2G9k@F-g&d(
z5eu0qz3-7D%}~N-DD6;sP{x3N0~<75?Em9GULG@~_ZaJ;Y<Va*Mu{w``5P(SK5pOM
zMy^Wz87tZ3muw0_>1M^RTyo+`^&Sh^B8_6KQ<8Q{@<Qom{rlX$m}$dCYNSmc_Mj~$
z)?TbA-K6}SCYzh)4ziGC(#I^Td=i>Z+M;x`9+drktMjuBZKPK}M%qDxcF+ivNFp75
ze*ffJLyH&rOX|Z|H+Ai%Zb9i`a@K;BBS~bMe_@9^1;t)uuakI1*RBAOOCbIsd3XR*
z?vix7O~(u0dXXbeW`C`}Jx7A8hsjqxf~+sEch|;OmV0?Dkv?D#o=z;&343EMv3SbZ
zoLS%BzkbAv{O-)FP}d6Ksay)hqkD$FGH(efXiS3ovfQ@P@U65XO3aP!xj3gmM3d28
z<g7EdbGmj;H|I!P3b~oG_M{iz9-8zFX*GuB_K1W&BCP7T#N14O?abO$c<BQ#kM+{q
zjCDq{oYB}DaEUC>^gbVr&0hMqg?t0;EI|oNfOf_ulDz)>dO)Tvb)bd(0;K6m*mNLq
ziKGT`r62v8d^XFATymCvGSwziRu^0fr8gh_p~;Y54b#2IAt$j^*Gl0vxfF<&TXS-g
z8YSsAa@5&ymzw;RLc`?}pKGant65Cl!zYc21KLq8ZJ$e{P$H?}KJ#bk@1OtNi$p>*
znoG60ls%G5cqHp{&^Q{6-CuG3=Jiqw*(klih-q?Yn#`ImmzZ0lgo5sCrp&8m^$#^$
ztwmI8olqjF@#q&CzPoks78~ga<=RPuc0&8+(oGuEqV&yGm*c(39cd(EWyq~E;CWmk
ztI6op3F_lJ@7hQxbl{aHzm=u{ly1_*q<QJb4|dncZK)6IxT%t!DzWd)C9-UJ-(UQ#
zy*<Z^6iZ`SUTr+vux|ucH{;eQv*X@tKLP_<*T2N`U&1(CBCF}Ft2e&=G`%OQ0vLIA
zE1|m;_I6w%skybqe_KW$4EG{g&M!TSYO|o)xfH&%{ezyvdwue_*qc;IZ?fz%rB;~|
zs|GGHJNu6F-IMBaXEY`uP@~zjeKvF=E|Jut%c@?bL#m3Q*+Y%4mMyDcOym+-J|COb
zI-WSvY@|>cz*tjF&8M1tQ6kItug15Ak7-%%Me>~GFVeLly<v{TrBHs~qmGdU=7+bv
z$Vn&hk*+<`pJ6@HDMey|_xm4xIXTLUbb@v`k81OvLbw!oEjJ!tIOp~7kG(uT8_swI
zs--|}f)a}t@W#1<zvfM-w2=(wOH45}pJMXKk+?)ws}JHLA0#(!?@cNt$g4taRiU!F
z<<iYsFg5i-{@5fN=>c`Siw5nY9Z(`EFmhJqn|EG0Y$KV@4!O+ax6BlrBXNl=$GN71
zdUc+VsgWY7kCr37tiR0aAFjv<?)P2xXP&)3gjs~W)qImaA7&9Q-K5_}>^Zoh>kk$(
zOZtcvsZt56gi(h}WQDYE-l47K#QPSKERANYIV5xr%r#shD|E2A=ON|iQ#Mi|y~<ca
zeTRnftDEsv^#9Cd!*0P;3gi7WVwncxJ(q4)T;`i!wk!zoBC+uK?&{iIXb)TppD#4^
z(xdnLF52NmzK7oNK=ynfvo6G?JF{!ADXX_XKgo;qh9;O~@=P*$qeN18VDqFI7m7N0
zlN#r=d{%087J3GkZq{d>Z$0+tnq(oHopn{EgjOkx#ig6I;{BpqGp~-ryz-@1`H&Zv
z$ZC^YJ?`x53sSwvLuXr#(xRH@NO1KqH79Jsg>kZN*vxrBaUt8hNV$`=l4>g{Yw%nO
zH4*W~7yj?<IlS6JzIcbdWSQ2wOk?eeOMGOzG2=Q7Z@MSDF$sQ=kqT%;0W>}?k<|W8
z;$!iQS)-9~sQWckUqe~r;nKromt93`|81{PQta_D-lW<&?^SB8DmC_&Tw-n=!h6mi
zG^0&tXo4)aM^t}ASrg<^$gN}CEB%V*{5_9-NoW>RiRV-p47tS2I=wVe?RMky=PhL1
z2u8Z5HoOMCl}k72XZ^3%rnl}w8yvz|V|>T3HV9W_M73SEp};ne)qJruh%uH)p35Zm
zFfQG!PBDikd|A|*^;D>=lhpSlZH*F1(M^6{fBweFC}$sJ@AZ}b6>H3Jbu(5!BiVA&
z2C^3eb?>m`c^IC>C6Z!v`Fe1dCJ($k<~bG2r?Ta#%(COs&H8Qpy_diE^Qnc*bt<dV
z6w5RP`i@f~tIK;?d&mFL;CnBR@gK4WURNyF6_yp3Na`B${AZ&-i+I^a0-;YNnEVo8
zrsL90I<kH4D;MG;Y~+?Skg@vM`}p%KGM>G&WX<CA&%-U`Yna71%VC>kHgs@_q*$8U
zzi{Kmb{e_tY{R|v-nJYGt{x^I>9R`^?CuzNvR@=Q>zeCR_0Uv3g3YQ<iCOh}Z-)6;
zXSp1dXLz|<MXOcV9L%Mgb=f@n<K`>#EhI^Tc2cZ{6|1aRTq5g*9qFqw6SMqaB!axI
zDVA$cn_RkCJI5|qGUvHn&e`6n1v6?Ry-{Q6q_<sV@9Q(goVUc6O-mP?Z#b1|Qz;v-
zxD<M6-_Iv0xoeANS;%)#uo5+_1PaC_KC*9Gv-NL|xApcS|2cDesA~^llH*dy?ZqW;
z{Pnr#zfAK9gO51d<Tu;Y3MJ<DVy>;U=Z=wWps~Z0ze)1k1Pzc&H|zQjb5C|TG`BHn
z3wrZL8om*xd@hmHukp6-xi4O;gc<{V?6hV%4Vo^OZq`=P=6dVTzqgP%PStLs9J&!y
zEiRGu(!a%ryJTey(8ztMkDep7)>}8@S7h|R`Ss+LBR-C?ke}aW4?7`+o`6A=OC${l
z{A1apmmZCAejZjwR}}pUd_*qYtgRMl>b_1C>Hvm>k32tOH2_y+41BTthdn#K`P7Sa
zfnlmr*D7Is!6j6AG#`%#&VS?9r=CCM+Q{=TeP_|2EXqD0mzdehiD5?`T}vr+7Rr;2
zE3&=}R3k2tHMn0%i&h^u`~g-zBUxUnWXmdGap`8A+d4ef-gKCSOp{<>FV|X^!@$lZ
zvWAWve`&m0ahs_z@L?`!mJ1pi7`b$l&ZhX6uMa<KA<LcJ;<_4k9a<lk$a*d1bm2$p
ziauVX1JuL<sx5%k1(!nOd+jhCwlFL8PYapw7UOM|Jhw__l$hDDHk}4%7xjC?i`;ZJ
z<z-Y`1|5S-A-7?lHHnzj@?0h?L7*+qF!{}ZB?y=9?0$}!pq|-z&N<pP&5?f4e_#~`
zS7f~2r113D{XE%9eV$VZyP#SwsBHA((#=}@$D_CnjYdFwfY~OVc*etQ!=;<GqHevk
zV)(ZfGT%AeCMscxpcrw9tl@z@(ms5C^GQ$+Ut)Q!H#J`m$|09-R?ns>yS|v1@8z+^
zS+0w!<sy`eOE>GggTH76+pfUWHH_u;t?##N>Vm7AaV=m#^Rlc?7P8c-fZS2S?m(Hi
zbhB=cR<bUQEN5dWw5-!|=xI3uC6Y$0`n#`gJ$Vb}c~*<XR4;~ko=Z2WD!q3wo%qy?
zbcNPXqiZ#wh;b==zmYLvlddUK-n5Xl&SqMwg_UZoTwG#qBTw!LYxU}9OszZbYz~vC
zHVGORmyq4Rt`YT3&-9t`-dVY3k8MuoLCNbNj9OeeAADST0W_qy8eAML6-HlzDFM{(
zo2qtGWm5u|_~f@HefT7Q&h4qdf>Cgrq-_JM3zu%zZwX^MZ3(G{0S+o|rEFOVU6M;=
zy)$9O=cD%KUh^W|p-xPiiM4yU3iUB+@MZez=CHwD<e;<sCAwAuYi}-rxZm}r?+r_<
zS+Kdw9%q|wnIjF;hp{KZ6`%Ls${|lXS(Yi#T0w<4DtR7-y5JI7qkShAkNvA?Ae0?g
z|45d9U|Gr~vOa7vqUVf|g}OJn>r^z7lvYWgz;fwkb!^ac>h>Ogu%$kzK9_04W!ed)
zoAlx~y|;JvLko#_e)1x%brFnATq5hE28U83`dWU3COwqp^^@;UtVzQa86S_n75+?q
zqCdPIFn;v>k-Z#T-Hh!!I*{%I`$HbDF-D@~n<%kdxJ1?`t;#Qb`<mZi=cvnasFU?N
z_;_5pSwViIp8Yj-GgK2amouv6jLKMCx>-{*mp}9IuRXmy);rhcM~USqtWvl{*4RZE
zKKrG6<JrOj%mu4Wo~vPD!6lMDZFFx-2emI70{?clgj>3H3#Jw>g(mZ9<QLUb?wT60
zP7B|!LJF;rB2Z#(<9;6BerHweK1PCOoFzqMN$ksTiKNfW7pAVc-RW(pQJ4-cX`Yuf
z_DC*~^~I|l$FyCu*^_1UAtP;6LpQ4Id0e_lGvA1=IP=Y07P7%<Vmw1a&%m;eOJsdH
z<YN4;kFNa=QyG{tju6Wcs4gzutkKK99Z1@Tc$0fhg{w|!RR?X4OE;_X@y??gTZLN4
zm!MLn%9d34id-UTe6urutPOuF!9rFzb&O4N=qAuHxOB6+XMQxJSLxp{89+Up(=6vS
zHt2HcW__|x8g%o|RbHecbmxUsTL@zkm%{h?$~S)bga_&GdXZ;=cSqOmz<1<Q<h^_|
z_xMYX(!9tG=LcU-wdJs6<5D1g6LO(hGSTf866YMtZ!2N9VJznoi}y|Vg!jKWH@gUI
z0WdSpH2KX0TL72!tOwU;{XA|QeDY5jE7CKPl?|@QnBc$v+x;yL>}P8~&|Gq8`y9~l
zxO9_#`CU6rwso+P=iph>OwFf(wT4S1eeY#D^Ifkuj#@~f(-e?ILX)6PaOq~f=osl+
zdi^4dxUe>#rD(H2Rp$~}6ZP*`7R5Xl?oH~PE$2LGbsqBK(#>l0$E;>AEFB4X!F)ZJ
zXmbh6i%Vqvc;ssMxOaZpXd|tm@QEhBL{kt-B>lW{(6r$%%r5dGhn-)iOxMa_VCGWz
zI=?vfMU2>0SPC@?b3uyinF6*2E-|}bdVL+xy+z>~FA@WdqFUFgVSeON$Zb+hdaC!2
z4T>#fowFaEB<7QZH3}{<x47h|Wve=wJ9v3abvD?Gn&l$saa_7t2e%}4O_<pdN;a51
zbd={PRxY?AV@f0Mb4SPDylNqfonL4cG0%dhap`6?+p}r*qjM%SSXHX6D#2jIrJMEI
z%l_Y;y3dw)UxEI4Lu-8lMtLsXq+xacJGNB14m}M9lN*ZV1}r+cMAr1dm;UM@FLF-R
zFie~!+FAIFT)J7a{>T{c)`zJUlHycT%ha$kP*b@?){M#*e7gKQ%kD+~aehZn&9gy{
z1XrQfXFNId+N6>fHd)B$u-qzFTbHY#J3A%j6z}u(k<;-R6Iq*wQGBtf`C@4ET)Ihz
zO<^O(_Db|3zd0W_lWH?5dmxv>BWGUQ(LLI;;Q}_QLQ7pjB9;&~Ky!(?C3Nr`^vnG0
z_hCYNov}vvj$qRnT-}VC9??6_jepEcu&^FWq2VcD)aDXNiJe#9TyZyetrsb97HS&R
zrh#_NrBJBEQ62#UhF3K9A|+1ZQmQQlB9}nC?$Sn*+6U2;3Fo@n$Tg?saf!)q35+&e
z;)9buzrF8+4V#%!hb`WiNua&no-N+sij28gQu?Si3zED@w)2UzsWux7uv`j7n|I09
zD9QfL7%!6UBo^pe0rVg)MPgN>=qcxaWwuOa?PL?eYDr%W`VE)(+=bsi=j{=E=07j8
z%b8s~)#BkhaS7RNcfIN2PtWDN@a%($7P8W*z7%U=#h|friTN#Q<#&7Fydw*}Jk~oK
zcClh9hV><v$Xa^+($9zA*FI$?7Z~cE$d)Hyz2Oo`$+P;!{X2L{M=!F^S+rtZD~31a
zQh3v4&m=q>aPx;LUSzkEm_W4zcn2;8;&R8(k#FZuWCOwX(CMx!o>vt%5^#x+T|RqD
z>Q_fYN*D<=hK-WvMo=EPbd%P|=~edNJFL~h3oK9~7AUOMa*3oBW&hZUzR9}pMYcN2
zby3$Y!V7Q-<vQT1+m%C>9Q$<QiydHT!G^C)s%OH|f=kS9)xYmNcw_6ybrv#Rf(8D6
zlKDT04N_bpYjyXMd7u9hbAxpRP(?4$hznq{<Pu41=-#6>6E+mF`iFV`tY$e2^E{Vs
zQkReWeXwbkkA);SpLbFUJqZQ{E|K+XtJhm*UwGm)nKNB3&ZNhI$(&0hrHmXNGc$20
zQyI=UYr;$OV&4a@!q-{rcz0Cfrid6&il9;-$(Bc;$#98TtsS}XMDz$-8Jo0V^jTr@
zTLF_6mu}Lil&J3iJ@&@kt8}?lI;cBbx>;u)-zy&Xo~IW%;Vf01uGPW!;Sx%A%=LZN
zR(sDZyW8=m7b$Zxlc|;r;{lh#JFNRKd{yJU;oB|bTi6<?R>G=5L*Wvio7(l6a?_kc
zJ}@7F(SL~N5Z1up>ShdU>p6Hq$`qKt;RD^&w3`|m__=hm{vK!<JyGuC?BUEp__qEw
zt1h@AW8=`D)`yRqV})@OX1;fQ-(}+_T-}V3FCXiXa(1tUOmfbAvq|V|&?mTbv!;H&
z^6RG~%h{G8G^!KS_XK6e1}>4bsn@nikE6mOSf7CTe}^2gLuP$~OE)R8XVSru!<I7=
ztR||}h-#IQxO9_lc<x#E?1q8Pl}V!<=@8KmfqKj(veMe-op{c6d>t4PpeZagHD3sZ
z1TNjIzF*H@7%}#OYkt{ALbt(|G?#AD6VsM6&rL4%Cbdp|X+CK+A8LV1WNq#n?UgVv
z`~^6#018vQ$uAykE?l}<PZl;ikvflQdq2P&TBNovf;p5+ByDXz{IjvmGnulq)7hXB
zsg?-CJ(og*+IoLjgSo|D26~ZmPT~}*O@Wrjr9ezy_eM?Mx0>~3?+L2HWfE~2=20&3
zv6*xI$NOhy^=D%nkPZ{i!(hhY(oOoS&4}CYugUfzQJ|vDr`mid8kfQocZQf=n|t(E
z_6|FoMQflnV75EBiljH*&S+Hg<b)T=cak3H+5_lgT#CdAkEY&zDDQ(#0j8X9JilQj
zgDaNnxBuwJrZ0c0akh#8r)fDy^4I;{t^K#NcXnKQ`K=7tJB0?Z$kcohXeL}@HrZYJ
zHy++<(%%-c$vKZ**TSxAtdDYutekzTj;G}Po&<vt^uT=0k`Ja^E|Ik-eC@a6P7KG^
zoT7!MXsmH@=_W1v^Y7^B?<-(RgmLY$Y<Udh8kcU?THjs%ue{sLi*$xsUqrP<u&Uxx
zsHfb&-=;(E{8Vcp6JfcxK@QshI|y82W_e+>cZYHRvi;EKz{Iw})O-U}1eb2o$lJ%3
zw_QGmH3sO4`Ko8W%AUw2l71iB_pX{@JBU4FA!)S`dIp!s`rq8ipZ9w;s0p)wfcBM2
z!&70>;nGd2_4HaeVboao!tlV$s^?{RAeYG6KcI77x5GbJEM&EFkyoKuDnNnY(#@Lj
zi6!OS)UjZdfR_9ZvHSxg1eeJAqv@oQb#d1`Vfcf6tL2jKa<GGN>1J(t|3CACS6R&$
zJC)1fp2JzG;Ob^PU1?f=yrzp6$#pj3Yr1v~8Zno`$N4jV-fNjJb?uDD1$IcScEFU$
zC1!W9!;D#*1|{XO*%t=NHL_(5Xg^%KN%J>d(?7cYBkRv#-MAn{TmZubmq<GFo^StK
z&093JkSWf#d{t|B71lpoBJ1z+rSClY!Ruw*c&yM`Rlvq0m&iK2?f9>Sg~Q+SB7Ztx
zvRv27VZF|!@RI-haOd1tqcWalCViL;PEp@elvOI1nA@??`Hy@b?0@V<E;@6Iqgoua
zXD)@@{tdi#bYc%HJIYWjy~47aB6&^$eTz%X?%y_7&uyF#R%anUg7G|Ea->UaZ;?wR
z9S=A>G@$%YTju~_VkJ`bQ(&s%5?TL!5zy|hmnXB0-^tFE<tY++3R*Om$U5=K4YTL6
z*v~8^)w!s;p@!XnMHQEB)~wZ^CcJw6H7NYctX!XYe#VLpS2yF(pK~9_ThD_X8+6VR
z#Zm$bJuZ=TvRBQ|In!6X2dfB}S*Dx(ro$?NOE;^3PN$OY-}^b|dp|bsNP0YI^jsqA
z)VNz8F75DRb1#qe&Te*2wVZ=m;L^=nzaVOH@&4{?{|Y{0A&n@6!%1AaNjn~R|F)<7
zSZ9SYh3$l<pMXJ-OE>HEIj;r&*S)>SY3HYYtZR>9`;ANCQ=e}5d`{2BrthJ)z$}>{
zc_zT8<`T2Z|M+<7!!~bJIoC>z6{H8TwGv!m2l0VRBg<d=*0X^d&ShB0Z0A=zO+rt@
zSL71&Iuo|{+Kd+(u|o>ndOKHXH5V*sT)J5=O#P<h!p^2J*2BPbR<)dk8spN<+W0A%
z`E(IG_OZ~}B1+Y;QfLueBI~@v{P*k8^S`!`pJ4TWO>2D(R{vbON$r2@T)Fo5ConES
zo1ZHAPK9xaOJrSay{_c-*;l@R>VkPAPVtO`<qDT>){HKVKfND0yn)9RXYF?h?9wqu
zf~$wAh2jz6JQY#UH~QahtrHG-kyB36eO<c`0~nXWCo0%oNF4u5n8kKTVFq1g@>>Nv
zq+H@7FRe)1Gw1i3t<a`nZalA9&cpJ7OJrSc{l}jhZ=8>|kYAkx`CTRKF6hEsx>+-t
zZq6Pt<QUsmhX*b<`7MWibuQhc`}10idTsrDFY-TU@vrLIRT!GM6pDZOku`Y0op`p+
zW=GA~`%NbLWSADX#Ow+mcep*j`k4!8L%T>?U4*v7C9;Z(M~}%IHk0k-w1)}VteKet
z4p-rM*9U#`bnYuJ|6w7Eoa1ka7FGhf1eci8^^y<gk2(3lYIdL$`r3cg_dn2{xkS>9
z+nHy(<xiXlUmFIBGR3nDzBZT0D$`P5vE6;?Bi3!fD49pw=g}CHNV?_unb)ia2Y!Y2
z3|SQ^mLix!xpcE^%ZA>_{BsmM@Ds)g_6%l^gR7gdvFEWb)<?if80M)JvgZmg=yB;L
zmG$rYp7rQ1FVYRFsY=(X;L~s^R8_^zH@;e*{~BApE^s!78YQd-ngf@Z+wHTS?Y{6`
zy$u#V@XRf;wnb+3&Ly%czY33;eQIGZ=Vvyvb%CTGf`O1rWZjwmWrsoQ=j?C}muyn}
zRR5H5;EIguSv&W>wtB?R>;<76pVB-}f!Uu+Hz~PCr~OBk>1-kfdqAEPng`nqTq5c2
zxK8tT*U@GclIGm{oT-G)gn@=jWYvB<qUY5;W4?8k2+v0J(0j165k69pYh=6ke2W+J
zcFySxS}Z7Usix+spuBO3S=}4{jy!bh;0)Xwc%Za;01xES%{n+>_lvJ*QFcBPrj=PH
zzgf_`xJ1(ZH4y=Ow#}T(UJzQ#V<qA-*fO|8(u30272D-){h7`JL*#kYavrv0xkS>#
zrpp=(kXm1YaSf)!&wW2<qZ(Y@jKG|80q2@8VMcXO6xY%4bzr9B5=oEx51IQo<M=jb
zH8GpQXni#MaBy`qe(t<&(U{MdIvH$sh?Dd<XaHQgS-D@f{Bc=CoQ14%ew276G#=`R
zOE;_Bva!+gZxuKj6F<M6BQ?{TxgYlU(+S;f_PW2?i<CM?pA}SF0s9kN3Qc_c=nucT
zIhuV}*BN<<Er<B|c-(5kbMYEW;b-D8h0U%r6#p4YAl}S?MM~@<rN?rYsfhur6vrwh
zOaS&Mv3ryrc;uBa4k(TTN*Droz#a<157*-XN6B@lUS}!%Kt1*;mPx)8nJ-0$yAI_s
z;H(sLR_cZ){uodwSqr6R2VE*E1BxYUu>q)&tTj?I+@fQQI>}mB4@j=#`U-pgT9dxk
zWX)I={`=^hd0C9g0p346JQ#zY<`)<oK!)Tx)(@}AkMwVH84x7bQGUWw_&I*Gw6Vuz
z)4|zvB%a;H1O1Q*Kg^HM!xQ|F!cXmE3ZB@96ga2v#$9LhDZ}}>e4LorW8+Dqcw&3q
zb)td+vk09<8sHWL1LhF_IV2DdkTM{dyq-+nKub9T))B`#5+(qd<n>JQMw-jk!5F#3
zkxRl5kjGDaw#y&ap98Dt57r;-Saa;pAH#o)#2sb!XPq8er?*?{vWzj{z8-pCZ-)v5
z10Lw15A=3BUF&)VJk&!U>h16l76TsXp^x--c=VV7kM+>Udb?9DJrR8qpw_rQ0@fQk
zBR~rGM8LC^=z#z!-0J}EL6-wa;eH1A@zKu!Ive*bIOWo3S&^=5{@1lYv;{EWp4Q@?
z=8p&H88Ab4%+SLGAWd(Pru*ZuJI2V-9XWa!0-jqckH#*0a^1{QrkiZjO}?nrLuMx1
z43jS&+F?Mv$rf+&MT-&xW}0j>O}@B!$$$itEy3iA=MEW=XtE`md~wr`0kcfDStj3m
zt}&hgvrV?yCSN?|#(*SK>m*ZTsOvm01D2a?%T2zQT}l@NR+wxnOulGTWxz_4ZKcT<
z5B4))mC3fs<co?Sx@AChb6*S`!O#~2Qn&*Krl11`q;RJT{CMW}@ykA0)Zv1}p3|jM
zm#!CG%B*?U#@|ELFZM<O?$HIkiagk#nUazzH4qx`4oTTz0J0<{%K+p_N}d78my~=1
za8^>z8h{Iu!U`okr9x6F48VM%%qJpaAyF0@fF(p(VgR-gWt#ywPn7cp;382j8i310
zxoiL)5#^BqNSBp#Su96}tYjE~9kQ~+0GyVU(*__(QIZrfkGYC6*8nU~lm!N0x1#Je
z0CkE|X8;~3$^!%NSWzAufQ_oMQ57Fsp(+&ypi)&T4M4T3R2zU4O-a!Vd1y+70jSiJ
zN&`@>Db)rbRaa7VF^?3gq)-u9N0oI3AeAbq24EvqHX4AfRM~0(vZ<170CK33V*qlg
zl4}6+sFG&@E>Puy0VtqKfdRNomCFX8kSc`+po%J02B3y2H3p!TDzyfnjw*Ha0J=#*
zcjm59*qzZY3R1W)6zUu33k50M5eg?_bbx{s?)8MZq01Aba5pDBCyIkIq;MZ69NE#g
z2_}5*(1g!nLq`rBoI^*t&vm#woFsx?O_)6nevmzPX~NfN;&Nq@SqgVw!ooU7rdG^1
z2q_HURzC#L!3lB&w<bK$YeTmthI3aYoWNKX=Sj?Y5{rjE7;uq9TqKdWf@Q!}J@Tp^
zb=kE-W<Z%9S*Azfu?+@P>7A<dm|L!O8Ur5aoge63gc*D??Kqi6<1oz_3A9}T?TGuT
z446+l%%`0*T{{B|SVlW9qh0X0s$1b?@4!8+aO_7<D|l<}U4=D<u2oR~+?@(*3RN<=
zhx<_BM?@bgn9^4}{ipDou_C3({%LX`wi*Ux$vv{<-ncErfP=E*pd2OuXXPGe<=(?x
zr#~5^M0S+OVF;iz78Ha#V&QHlI$%Kxce=vYLbV4{xR;eX#pPlJDcr9LQ_!gj^hNGb
zg{Dkaq|0if%WBjAxs2xwxTex;ss&v;7*MYIm#YpuyTO1uwP&5$cav+YjR8wE#}X}E
z05)hnH)wrv`+_kJYL0_iI0ERv1x4jfTlnJWv;|*{J8EH3Te%#ypzCojE&OUM`&{};
zE**uV0J>{I_S{7aI~clXK?*o#VGUtAWX`>@Fok9Aa?f(j*1<0KEXabpWZ{{2)Tkka
zyJ4YPj&4|x0xnosj8K>J6|^_*aD{iU$0iZ~Boc^+yBLs6Vv|V^K`UQF9BclcnjV`;
zdZ1R$ZpkK&?0Pjl)aAtm#o(@6_?GCm1qI>GTG+u@Ces!F=}I8#Ukq5J1T0b<!a8x8
z(sG#+BmjGqfIW%>yDhtApVD%l5`+M1j8G8n{)OKk7lH8oxvLj09K&6%UT_b0@4|Y=
zl_cE5UApidbm@Y7z>y1I7rnRu=bl^cln)*~>q~yxFDcBiW^kXOU0p6W=*|Uqa33z5
z63~YWnhtl|!h_rlsG;p@Xe0u-+JH>B`xZVHeYfBhxW5)=kN#S44|meSdphi{Nq>3R
zuPxiM>P?quy5xtu7z{We(G!v%x@<7uq(o0jersH3E*LPI(AmUqqicVj0ZD`=5kH~5
z%^`FS@xxZl7^et5Mf`Asg8`=rJx%=3R?dKYLi33qE|?i`hR`#_?+e$dRR(O3=?2*k
zcNrP5QKlPZzkV(!6$WgQ=_c6^Pn0qsO{QtGpD^%lmg#2M4}F~&W4c18D}F)?ouSYf
zil1QXh*xO5;`hj9v0=B&ROn2_5AA9UNKj~k;)m-G1|%vpQSp1-<#fP++X}s{_~9v0
z22?7vQt`t9fdO|EdPnga?(zv>K$SwP6hCab45(IUwc>}zI2cf)&>F=Lt(OeAtI)fO
zA6mB=P^8i#)o-(F+n)i&DlJz1dbymc7;s&s*HyovE_V_J+)(Ka)erZW7*L|p64ejA
zofuH6(o)s$jBC$<0c9#JQ~jp9&iXN+T&3lzpD^lQ)95wL4>uqfqe!Dgn%{@6W9JMg
z)@ZTjhX=VBa9yL<H9uh$d_$u*G`}gXBaDnuqR|q~5AECxDAj1G=7;VJ3@FoRndXQ6
zj{)TxE!X@8yZms_PZ35g?wlwLp|gqqY!WE6@njN^OdP@xx`MP^L4pJzlLTZEhcJZh
zBrSK6AOvt)gzm`Q9C3c?<Z^R_6mV`tD<e8L!uY@)8t1uA&N1Lu$@5pqf<TnZ!x5}j
z+`SP;4))rY=?j<X&<|an0E|&aeamQTfl*8Q*3wszfu4}?T=0LyVS{Dmbcb|JXVETq
zNT@&V1BrttBc71_Pe_5nC{iE=6i5!iFn3vMd07e)fJ!N#QgR4}xhkn;l@x>kI#z-z
z!ksCFi6%~I6sLIOz7%`@G=)x6%rP!EO9mt={)tME&?lEDJ(nnbgqbH@ailA)1z@kz
zbFb0|XC9WvDaCP0X^jA84xf$tZDN7CxcoMuX1KE^PB<)+YQ?`=2@?8ooa%^ETcfR;
zF&3$gMFwDt>e!;T##8=`u}^jEs|U~$0|nu3pExJt90V!g@`?VNH~<39Jv>ocMFymR
zUnhQhv?T$~9XWBDK}XI;EQLF73NKryMAj+MU%K1{SjitKF%Oh(xJST%8LD-L+Dv%+
zMAe#T0OqLHIR;>XYF%IemZ;Vx24J~rU2XtYtJc*9V7qGFt~L{@G)uK+8GszNi(mlq
zRBN6A*sogm8-RnV^`HSbtXdD(132zN9dZX&907Vgi2AW;!G)eGtB@ro+Y*y6p5A4^
zQj=||$ro3!3`jQFl1;v7LTA7-lWm#F7j;(#tTwe?ZHh#v7Icq==Ez-Qg)da8_*W`{
z!c6p3iG8Z{5I)FE)iF~I6M!@|Hcjmz)ZI?iu~Q90;02eTZDV%O0o-e`39)iHH!obn
z2m8}`L`wM6>hI_J3+!pZje>n{xwZn?pKTho@xhfU`?G!X_CW$UvQebZH{slKf~5|#
z4imF-1@x$Ylxsh(9&qi$)&s5+ef5CL<+C1ec_Y;WE>G2Zz;(Q%9&jB`Z~^FR4DEsY
z7h_f7bPFlmqZnHX8nYpVdlBP3Xk3L9?mvv90QwDsZGig<<9e3WRHhV@DaGQFmjOGZ
zh#gWSo)>!2<xLE^au;G$MbU*AS`~L2#u0&KQl$A8X@SB7U#msdYEft&VT`Am<EbVZ
zu{Y_FoAf9gXBcC<?$};$#6~sGSrDfmG7ed2u7DKqJ;n;b$p`QgPM2expRoCRu@wH%
z@E-P9?!$x>2Jnqb0Vvj66zjow2%Ir)>GoT?1Kq<Ia93|}R}aQT69XRWK9BVf+-G3G
zG}>Ys4aQ^t446%u&!$0m*q#B4sC^N2>~S5wW57!4vyz730%U;85gOhO+@DctK=)_B
zxz97c8#+A0w9dVpg=KIaeJziU!cBavTBso2?lBWq4t`V_DGcC|cLd;?-r|}b{D<po
z8e^2}_Hx~U=eQV9t+%MwgVE)N0S|Sbhk6LMZU#)DEvC?5JY>p%MA|%&2I28k1}vcV
z1=N8~&J0*ieU{UZuUt-B=#&lZlRIJyCEO(YZ;}Irl|q*6$TApU_REg_2H>phI9qRk
zVUH`49Yys3YL!s$+@l-21A25rLAV<?rm#%PH2*TqA?T=eT6CS(P3R~~bjK1sTmUxf
z(VO*dLPt5ws#Xt20F`+t2=@oau>-3LQn(X1epU1UhZOGqjUxow(IADpdgDE~WdteQ
zw;RWsVJ_cpXlOq$k8Zq!y)OLe8vQ)b!P{9cPUmhcC?l4t{-tV=u*$fn_PnR|5e!9h
zG{+pRwE(1QJyW$lXeeTN<Y|sPtu+GfeN9#>?j0_u`}0Z6d=i`HI&97ET}UDpl1Myb
z%76mBV}TxxJ}wL>)FTV^DBLn+K&c*Csz;#>ngMt8PIvSebT(wbeZBL2y^FA88Am(D
z(P%u8$QUzeyP33Oyvs$20rO~wd9<^y%Rt9~WZF5IcEMx+sL#Sj;Lhx*gQ7D#d<5>m
zjyu~7NH;Z4X9qi6TiwH59_{cUxEH(dA<9($GBrq8-PNg%I#sOq1)5`l;X`cD92@HE
zot1gN=Gb2k40pN4Lz%%L9+w-q_ywFh!Q+^VR!>Ob9`M+eaSa3Oo?}!$Mw<wYjRV-+
z7=t^+3k;Wsyo)lz<slD`;coFr#YH8ga6fqL!srJNDctuR??JzJ(2KdRJEpK2DV40H
zlDDw<c_djMN#3~DVT?3lO(WjI@^U}1?kC>3nq-W7#Cnf-<LP<^ER?MaWpCVeX25aT
zdR+DvtjblgwMzEJ>ST<aigl;rEv$LYE7tRhH!g7)W36gkt9lFC-!avCO!Y=(k1<wi
z)|HyK&=HSl)+3rX_B+P-Rk!}CdkaI@DcyQX_r?(m{r;`2rns-a(Ecja$O<(Yjk_$f
zN;RfZ?KZ{b@yLM3s`atjOn82rW{uNCV47x~W&mbt)|m!iwq~7e0Oo1dc?MvSW?f_e
zk~M3x0obBhw`k3T*Ur$a83tgdX5DE3vNdbA0m#*?xdvdLX5D814rtZ`2H=opJyZ{1
zbOe>vC69=JBRQ+7Oer!`iZ<waF*~GgsOvFCj>OI;HbK{P2JDv_?w6WWyH1+D<%+5R
zBL$D7fE|%hH_HAS<v`Ta7_du@-6i)BnqQvm$TQ5TXXMy3au1<BUXdMF>gQAp&tPJY
zgQyHb2l~qTm3UU<oA7TUce?yE*pn;ukV-uQRUro4(L?U&5yDVhrH54M5yI9}wH{Kf
zM+h28jUG~?N8p&i^0=#q+|?scsl%`k(D8XF2rM0jfq)bU{@})4!5_dEg$H~<eUxRM
zFU90bv1pHGz*#BctQ3hxZ3bM_J6_bI(Q}UhSM<m$dK8`kVnB%=S)xZ_!5C1fcdFE5
zgleeMJJ;!5gfldgXvax38V50UOFV5CPdlRTB?IQt4s&T|+zVyEQrdYb?Sg)&7^?!T
zGCV$o@KQ&mMn@$Z2EAbS{wLA@qz33N$bgfQ|4AuOuq_uzu?12OVPL!_Ij%`z0#GT%
zR!TjDk+N2D)JkCpj2_r?Z(L^TdvKsAPO`>H&2a)?z!WKbiqruQ_A_7-4VgsS3hy6B
zL*i&#VMd-zLnhO<!VcsV8Zw2p6%^B{G-N7mi;Fpy=`<QLjkZPI9^-U?fu6_az$qH-
zZ7@aiz#KT8F`$;VtEHV0z?BrVDIS#rCm@W<0VxoR1C=j4IR!XG;Be!v2poX(s2e9;
zhxf5qp!D;#Ew)_tPR5cmfFD%Hdl*opw<ywsar==0H+B0>-GQgq7*M0PsL_Lk9`#7~
zd8CKnz9eHzr7fn?V4*5!(dM&g5N_2n#zJaeNFC!{fgTvJg8HnWA$St<6IV<S7@&9@
z5Me4QR{e|BK(xZJd+(~Tchw%k%<)upJXOO4V4W7bPU|6z>FJsyT?<3t6IXB%X#G5(
zh_FNcFKP5IvEe2<d)!GvPm%^0KY#&ei2oT9C~R?DB|WZ^-oliAhdAz#FadZ<dORh)
zg*iS+b|lGR2%yOazAg`5g8P%WI)xM-umry@_8ds5aE2<uE{yX4r0}>TIGJEv5-=(7
zSR~jU*=tp3F%?=Y`t~rOQj4h6B5|h*RR_qGhbTck4{HKacvupwP7F%|-<5|V!4#H#
zKJm{dfx_0&RpPj6sI^<famxTaC61@WU280l1lf^LUyW$6gg$Gu2Yl&@X#z#$flP1)
z#Kk0}@Zcrx6j$&PC_E2XB2-q1=3k-(3ELR=w4wL3*9B#1j_#PFw-$gjeQ28gI%-d>
z@cVVge!VpU*g2peJoE{EX`KHcg$Fxvr??i)kiw&zV9M7!;?LyH+&LHy!llyJQt2D$
zQit1Fa1Rf3g5MF(3PK7GZh}igG|@o{gfzkM19<cWa3045Z4(n*aZDhE2QR_AaK8~!
zc)$`|(0%U;SkjQC@E|3^IF+PCCMnT4DzmqmtHjJzx}if316C;36-qNfDPE&k*BF3x
zigld<*r-@H8h|Z|b&CPWP^=jSV83GBuQU@h?SqQ-paD3nSPvV3V~X{d0XU&pPZ)rF
z#hPCa;OQJ_#ykuQ_L_;VFf5>1@z5(+ckKP=lgRlb+OQaoSxCAGi{Vwox{5Rt7AGmh
znqmM_i8a*#WD#o?X(lXRa)>p@0OS#CUOj-PN#X1A_%Ha~@c<E|@K`U{R^5SKSPBpB
zf?bfkRk|FLF2}BQ`QI}jLypLhBOAL8TQT68-tn3qZ7^GPEZ3tAW~<KCdKW=Se5gk~
z)VmtYR-LBM7(qKor0o-FCxh84asiD(TMT>bavHUqc15!lW)5%1W6$8*;lu$cJlG6=
zO$;^z(-03ZBMdN?v__Y-rZ`};=Uvt4Rn3A9nhYq{{L3{58n75pr@dOIjTA<l<+@|}
z|Fc28x>+BIBM!UenC>`MZ-a_+#kGN=j$+YlQeE)?@F+RpJc122J6z8>k1;r-*WkpB
zTfT4)4^@LNH^mjI2F6t$o(2aXR-hWizeWiXcE{t?o^fg)VIiBWI+Fk2He1gvY9E}Q
zSRMyd$AS87HjH@#`Vo(FgZ0m-=Sk#w5{(^>0T)TkMbb?;4tIlCZ;)ofR9;T3<p$t3
zvEDWS)x=tD0H(^;sd6)%?AcS|Wox_vm?c|h8GyO6b#6U?b{wco9)$<18KdyPA{=7x
zV5?&f`PBNUs6>Aj{#l0;F85NFTA~M+=xu~mN2wlMs<#pBU}bu6ncfDwExV;$4=&f+
z2>t4&9(+@8BdChE^x#{18`N3YEt%AoNm~f2d^QcsrtR=R9%E$F7qjUg0mz|YIkX+l
zER2yuU(BI{1R$4&<<fR|_KPucY5!b0L;&(=SRQSMW^u;IqXYBkD*`az)MUQNy%e2M
zccApuA@`;}!>I9+HC}S3Fd#t+Pmm(f*OUQEY3Neg4wq64NT#95v>h7v7_f|nE~D+x
z2+M%wG;}#_hl2$JR?yHDv>ocd3|L7+SJHO4O~8QvsO>-MgRXuII6-YEsLw`M<O~L!
zq_&gP2hR{N;1so;qCV(O&43bWD>39zN^PZvJj$r8%#cSpwUyWBfu}*B;qjnN*zhoD
z6QuA+O{lV>85B}@bS7++IP*XXkH>`L2*zWA6o|rvi}3_k6ehrV=q238!gU&?@Ni4`
zl*E_hski1lje&?TLv2HS@GKttbA)Y#4?c?h`L69<AB<?i{v2Z)<AaAf*q@)<KKDTf
zH1_AWwr_pV2Fm{Y$@Y^EF5=jqQzhF}$>+Sw50?Rnk}Xm4`Pj8=W59CBwp{YT2m%a9
zk!&gT05&IhA08D7-v{@UA%#bQ!kR`k3{oKW6W)Q_QGoM!Pnd#+a7f_+p4=&}Gb&(`
zT`K8ICAR^D6_o?njTwW-e8PJf!xaR|ML|OZJcb8;Lh9m}l)rYp)3GDyrCaIiTj@LK
zql2f2;2s_V3MUK<0R<BV5Bnrc7*&dYl@cf%xSFK)n56a=CX9KiW1e9ne6!kPv)Wsj
zFmhE#Zv93$#*%`^@c>eS8k(T^Cn!O}p|K^3V~OGTNt)tFGXQ%P$KLwmC+xkCD~{v!
z0B%S+-`g3H3hN4wOG63|J%uUDU2&(Nh&<SoP$Rb$|JzC+s`>1BPnCeDibJT8X==-9
zYLEb=sR3!KL#UAqwPl7HgaAgcg0{iKR^ft@<-SR3v`MnzW;Fw{rGRY7fs0TE?2`ia
zNe(ndqJj=@#lu@+^`emsQg}!!d|zBufXRi&u)-9Uy(_4dtLdSY3Nqn=tZ-S3fvg}C
z9=!_9Gz_Ss?W*d}ZDKeps9zq#3fm%k-VM$Fh8BeS4Fm3JQTMcG1-&a#cO?G5rC(H<
z{;Z&P<>`*R`lTNl(xA*d7#5Zpm1ih3kANk-xGN-<>%}o77Gwf}u#kvRu;5*I<SVRi
zR`htqKVAtGPDw6S9E%Mep(%<Z#Q@|gj@){WP?pCb#c`+}K+i8I29KeIqZR7#peq_T
zqtUVn_wZO*?!vp`WI+m#jfE+A5(`pzG%V}_7!3<jAO;qmy-sz-z_Kx%$Gs9hYK0bA
zp+y^vA2F3$H#B~**L<v5A8XBouN<da<8%?2rdy{OfSI~=rU96(TW1@9dAfC;0a&D4
z7a4$L-I{CwR_fN31|UPXX6Vg?#<f$o?lb_|x;5JX<m%R31F%oG?lS-fbn5{Fa7ecv
zG5|+(>k$KRT(=&t2XO5GEf*q);df&-=?p05YPC51!pDLzVt5-K;RKwA4ntJ|6(AUl
zc)T#|#w_as+N*#Lz!TbxQA&MFX^6n6p}lM9AY|Y!2Rw$y6~m<`S`#6K2NOfhXH&+f
z9;4Q&AHWorLWic%*HBBrg#|1%_(3G$t&J;rb}x?(f%h=rs@~$N9&GTuv6tx%K?ASS
zTU0R>TlBo~d7y_FJa1Y|ron>UJApP&pg}m5u&2zY_W9Id@VxO^Mnh0p#UN<#-aPOb
z_FpstKnjm`hMxp&4$!lCgfmP*e^j`KM>WF{4UacN3J+z5^^1#cNa5kjuoiKB4k<id
z8K&U+3{rT6GOX8(+T1Ef>Wj}Q?A5dBbJ?`7VXg684()4LTJ*}LFBw)6z4Pb*TsUC3
zGZ;&Gj5B-;%l#DbKSctC!}5hBfCWFnqA^A>X<1By1mFn?ctRY)!f>+Oa<UwR0DdHR
zD;_Nk-xour!Lknmq#@28vQ)Azm72P`7aYi2CWSAP+Tp2W2E<cGJPkt~jR7;MV<rtl
zy`2FG)R92LghrD{9f>pyzYJr{qK;WKOtA3JrjFS(3^h;2I7V&9sJH;VK*KK3wm1tj
z#s%8#0_`aP1vIRHw#B~57zMOj0qrROmuc8#+7@?U7~?YSahdiOfI=EpNZX!rMI&O2
zLfWg4_7Q-!rY37m&9H*7??4^%7;;#bxG@PS5H=1EJYd*3i5--LqmanQXmP;j5#q4t
z;CXmR;Zfmm=E?r>%8}llhYOv4$)+w=EEOwvrew}Ee8(M<d4~bWlFV5KAWt&q8GwAr
zoNoZmO6Ic$;DThnU;rv4bA<tzPt5a)*q#;=^Fjl#gqW8YfNjLQ%>bMy=JN*NA~9by
z0GEmRvH^HR%#RE}x@=CD#d2iG<_rU{LpJX)0H<a1X#<d?n3EJSkGYC@t^rt}m=_p;
z-HLg)0jN{Vbq3&pVt!x%9xLX@24JIV-l&R?tx(Ms2B1<kR~mq7)m&`=QZ#dlX2?S`
zR~Udw&0J{!sx@=90Z7%&sk)d)3N@!t5m-mf>kL3DHK!VYjnur+0Boh^tp*^QnzIc+
z4vY3^0CK50*8t>EbDjaXK+P8nKmj!u7=X*veAxgLQgfjJsG{a715iWFH3p!TnrjU}
z9W~e00~oRqMtdHn5xXtxp2eDfu@;DnRtD5+18cP*f_^wfcT6$Fd)TB8+@uc?6tyhf
zkyU?G7I!CMl;8m$aYo$Z3it>G;Xxh+6@I4TKT`=5{F)XkJr*myahb(RmZ~^X|DOuq
zBUkAy>|q>H97pO^cs#u79LStuBXNjBFD^*o0U}Xd!T^z=?(wLQcub7}8K&kL?Bu)a
z+*+P1Or*0~oiQTuJy}_cRR1D15S0T4+*Jeast(k~81PVS`A`iKfOT5HI?aK##~5i^
z%QP(r0o?xp<%9>BMCAl`#NcgsWJxqGp>r(klk#YiLJ=#}m<l!4aI`6+QjJ7YIL4cV
zTzQO1T$t>4#h8Q_=W!(k?cp@>KTQG!&$>bqTS$5c%0ek|lp1zJpODxmqz5Vs?8T?c
zj_LKgp=fG?Pr?IKqT3MzQY7CLDHMS}ToEgw%n+v%&2PA=0yvLNi7f#SA3?1e-O@1_
zCG2YO2$a}HSP4_)Mk%s4Zul@DRi>%38MkK{uvzxsEC->d3j=<W9lyz~QBh~WVcBum
z0GyK@=j7JHaip8F<7Pd8hfv@pc(hAgUE^j7r0@ur*uilj2(QPZT4D;Czu_Jp$r7ss
z?XIve;jt@)Wk(_HTS$kZ76SKD-*Or%FzRT(I{GRy@F)vBhKIPs$KWv-NZ~OpAGv~8
zp*JX`@IaPmroq+=?U{QwU~a5Fjb7Ib;6Fcr_n<-xx$=;f*lPdF|K|1ZmIr^4**$&Q
z_34IgOYG17UHiW%yMBI{{W+lP0QaBhqzL)(;FtI$+?{|F2ziO|)=s!WUP6oFVJ@*p
zv8>CKMrDdO2D@WGg+eP7GxkCT+*SPVDnWvO<z&?{*|3AOQgy5}09#eZ*7_YJmd7F0
zai|`^I6P2L9#IolHFzio`W%m<DO5BEP{qTrnlKvhP?}f|?0Gj-{~Kx`Ruuzk)gHBK
zZ(-D%p*dz~;+dEAT95TwZyfd5Em@i)tNzRjHY_L^56X#cj{$YGT^)@^0Gm3r7Vi2Y
zyo9kgu*Vp^m(f=Os)onw#CvfA1yXo`POK~p&<TehAuK1}!7>NO5Ueovr}GC6Tz|7a
zqa0CD=<&_|e74E64h;B%11UU>hw#K!izf!4Ya35PopFWjgfGHFbqZhPy6S&j4a8IW
z>}j>Cqt<ZfXo}{T^8dbwV}1PxV7F}79NX&wJS7iB<S{?-ap-ajDLl|8eiGbZfD|6w
z6FWaTQG)G(2lNzlgd`;<Nr^Qao`{&ML}FF2q7~^Ki}Yy2F|dv|^=M&luSV}&qjxbJ
z1B-g3cQqUX>ok?d7><FppG7+fi^PRAav_bvlP>JBD`?aT+7-9K@Z-VTLBLUL2zc}!
za2{dQo#F~F3Mo9MD6Y9MrYL;O^UippScmMfrJ8@K7KF#c8BnLSsMA^s%G(0nvA|%$
z-K@9RthW@Dw}ZOlV7&<!hZiUa4@in*6voDa6dr{Xi-?vUNa2x3@wHGxfitJv#SK$f
zS&bGJ1{k#(1{CNm3iM#y>tR5l-l9+sHaI=lOLd2^*?UKCaYqlvlSb^8`?}A4Jp^l>
z0dcfN91X^Ki2*Zd^O-bAuo}#x_IcD{aC-1brXjd5iks@t@r<^wqT5Az_rC4)1z6Nf
z&zQXN05L{P1uF*xN);+$wPal_i9^G$Quwb@d%?;vmj=(J;lc^nc{F$)4Ht?xp9asT
z;b>%G&t5=-7tnCwjK@M6ypV<q?R60iUPQxDL1woUQ(G|=cZ;iNSQU*Bj2cz6dlh|N
z0BUGh4UG_t8a1?g4Sik!YH3(4jS!3)wX|m~?IQqnG^~zB;F!+JQAeMzqb~};7E_Zg
zCb#(o52`^^<FRD1H{b>{Xe~T|EKUO~*9)}Q1v&uB4);>u5`#IliuSIegOGvKFbqOG
z&Mfu<D54Tmr^MnJJqA2bA|5D_f>Uv^-mzGZ#tkvXxTSZzrAG_r3GV8h@9JIfFgIg7
z)}tQlU2)c7z%<%v8jZo-VFt{m?Pt?Yf?;YAja)>dgm1W#My;e>aqAszeNa_A94_uY
zVHjK(Qh4ZF;qC4y{&$o>bcSKiouoP@8MYl~tB%<QV6*DjT)*weZpl_1+4TTA(ZVzg
zQFL*dWW=cMQN6^8ukEuPaO}M0im?lY<bibs^?QTtzd;VfRUFG`mmIK5b_mLKj@&Xw
z4ibPfa=;ncAt=`u<(3!aAOvpZ{rL39v)@et7w8oFTnZhC#&8C#qtC6Q10T9Vv@#%-
z_DZFLaoxp$G}<JMwnA@J25hChx6+}wCybVE_!c|>Fb)B?9==&OaQXCT7O9VEo4);n
z1Z^{0^3Ilg4I7Z&`y^l7fMhomNZtjKFD?w6K!*Y;MgT5Lw#$-_09=#o*Q5XesFb`b
zC0{&+%krp{I#fzA0#GH{sw5u)sFm!sQh)#?6Ypf=i^n2Z9?7IbGKnz&1Cq%r0<ePE
zRuCTnSVQb<NPqyWBldL$Ad`4!5??gmv!`T|4w)pz01U_^uL!_SV%tf41R$H(vq^ve
z<Pv+X0VpKig~V4_uNIOHg(OA*t`hrI5?}y2UnQ{uP)uya#76*1iM`Zt%PnHRWdQCF
z`yB)Dgm^z8zQU^b3F+{J#0bDsVt+~k3_$0nBvt?>%eKj~j{r=U?bBuPmIT?JU;vV2
zdy)azC429ZeTAO1OYX2sjuC(?*`6f_7=X@Ma;yOSCfk2AFmhyDj_e~a@??9Sfw5n<
z?>7JkW&1$`a9Fk<HUMX2?=!M5_Gs2V&d43k$T0$NR<@s&0}MdtvvRBeoRjV642+Ai
z?V{`>Fs{h<D+Wf9Y%ekZC9=K50Nj-AHx0ld#e0$BE10$xDIFFmF#@nyu`gBv3_$0_
zN~{1ZQS3_$j3r9c5~Z8KSf<#PDLw+QO0lmp+>)Z$Qw%_=Vox;yX^K6~0HiDSbOW$Q
z@!q5O3P!j+N{2m4i~!^+_FN^v0Cdh(Vg+EYV&7|E>{X)nD%}LeKE<|A@ezOniv57$
zmP3mDkO4TN*pC>1<BI*b0XU`DPZ@xxiuY5+S2$4eRO#?ki4lNFs(q3gU;sK#Qey=m
zPPNAw7;$P;oZ3xbOjhla4UB23ZJO#MFlMUunFhve)jrz*%v0_248S7QzQ_P1tM+6A
zuu`?JGyrL;cbe*phhtcONmDzdsWAeuS+#Fg0}Mdt&1$RwY*Fo742&&m)E2dyz}TwV
zw;C83sx3qH5g0pF`%VKRTeW8!fLzs{YXJ7C_I(E6fNDQr01m14Lk8fk>U~%B6)X>T
z)ed*n7y+nN?X_xv0q9(-#tOha)qc;wxTi+lQ@aU_I@MlhVAQEmbq2;m)%H;J5pH>^
z+MgP3nWEXJXks~LX!aQfAW^d?8h|;PeU1TGpxGA~fF+uJi2+!rd9Txa1v|?+t;0Gk
zMgZ1p_VrqT0qDG5ixq%W&7Nvtq-s&AS~r2QL9=f#Fg9pW8w`vz&6cM52)Cqb_H@H7
z+co=k1CXWJvkX9vX3sGId73@X0PNT7`whTB&3@1T)M^cCHLI|6uGK<nwFm>yt5)kL
z0Qa<K?rBX8K=?f^(g5_mrwtN-I_;S{!!31Mn>wwdfzhu{8!P}1HQPhYM*yB`_NQ8a
z08G*CQ*;rSq1$H|fJEJ%XaMHu_BjS%fo@-50G8<XB?e%*ZeMNyHt7vE=~ls%ut^Wu
zq(>NlUYqoO0+6OZlcqN{0O4tRqygxgrVkQ;&H6K&4YzF8+icc58W{aH>w^U#UALv{
zJ_4{^w{O=21RzVdXBmJT-JW9r@^pKi0obqG_ZxtNy8WO5IIP<b8-QcF{a8JK&H%6o
z;6Ltx%P#zJ4>L>QKj<N>b<*X?bU7Liov~3mLypOiy9tgXxw17^ZYC&G`(*1r18_jL
z9xwohWa}XVa74BqF#wlj>m|9Fp!Zyrtyc}eb=i8|0F=qrG6PT{TPx}T+^qm(5D$Qa
ziy-z=#_*R6Fvieiz(u{qMLihLRx;p<-r|ZLY;Zufm*@_`6`)dYQK<(T9FTqL^bkCz
z$MTp&TTG(Cf&*?mZ5~g9gfk^`seLYW2<Dfi)MqIT!DG&7+Xr(r{DO+`O3oitaT$5>
z2UTD-%YQ`$?G~&M*EIiYS|A!r8BnV=t<_ox9sHrz@}U+a0GsrtoAee!gHG35rt3in
z;1bcf_ILg+3+|xc{wk#K-(W%Ws{1!s*l`f}=@sEk`ULig5I034`gakQsRJSgh<`$K
zTEWx}F6j6t{(=is7XQ^2+)HJ^HdFI$^_#0`O#vgDF#^)P?*cN8w|sa27R+fhB#pMm
zWj6-=hj-(@<bvB7b*A+P8=v~`Pq=3t9lDObhKqSry5JuE!!Fn<@rPZUhF9mWx}aH`
zWm2O0m#9G)CyN31RL4C-Sdv7|k!T1@lBzjU>!XmcTXHl<PCbAp^r0Z|!!X#eu%LkR
zUxUHTT@1MdwF$ogBYZIDPr|rr9S!S1<39r<5M93ugH%_PQFs>rEf{PZk6gb6180RG
xiW+8t7H_zN|LhB<j352!yO8Zy<KO3@685m4@xg!0BKFMe1J5VlKP&B#{trc066gQ`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/data_type_enum.hpp.9C984CB93A76A458.idx b/.cache/clangd/index/data_type_enum.hpp.9C984CB93A76A458.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2592dc47562aed3a56924d64355f35fcf138f7b6
GIT binary patch
literal 748
zcmWIYbaQ*c#K7R3;#rZKT9U}Zz`!5~#Kk2=nKOX&At0`p<J)_h>yQJF>t|Os`Hm9J
z(31796}M^1TKTqBM*RNnw>yv{y-<HA`?>JtQ?|LU(7c%Zt!AQ))7qa;C9d)1J+a#1
zee1p0ys#R+4^tdVgfiNSLbpDuU*a;iG)6ILy>hbr%B30iUz9#d@HlbyqFYYM7pav$
z&Kz*Ov5<dP4cnGWJMMha&p5r$(=}m-gxTKjA6TbYx`-85<|bJ%GB9v3IbP-S*p<l7
zz{bPC%gw^g2__ggfbIr@0ApU>j8gZ7{0yu>X-*bS4ln_e4l|v5ENQNK94iA8Hv=y_
z54!}IVBiF6X5?keI5q$F{oaQm{0y8x{Q@ik++YHxf5SZO{i(YqZewEN0_$L4;AQ0k
z+62+ha4<sdmCCDyfA|?#fV#O@xH!QCO!tl1>P-)K`H1o}Fao9dSonCr1WfwXd)dYw
z*((<O4D3K@9u^)BFaeYPSe&J-@MZB_eg<ZsG(QVJ7np!a|9!n{U7qCX)BFq^KxtkU
zUM?^JlP*e4D?Sf&Gc5cVxxnED3lLbi0fj+<1`~$SKw(f|!MqO(C7>`Uh+wK<-UkYU
z{0<X_c^fDU@-s{r=3SsL$gePAm^XpKAV0!{Vcr7@gZu^)W++O{Nd!ejaZz$HH;BQz
Ok4ftKLdEaz85jUiK+=x@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/debug.hpp.3D8E2A10AC86C578.idx b/.cache/clangd/index/debug.hpp.3D8E2A10AC86C578.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c074be25ada10deac695b36d421c011a1625a9d
GIT binary patch
literal 1998
zcmYjQ3s4hh5Wa-K<#9QZc#z{6j41?!MleUjk%&B^2~{AXMNt$G910XjXi5hK)FK+G
zl&T;FTd}n;Bev?4R!5vDX)7uYR+|xOeOIO8gNllRQ~O`wC&^6q?)&$4_xr!y)oV4H
z-%tdJTc<56%qh)eAqXM_e`Bd3?>PfOC~#2La#+83$F*4k+khnt=PlNWg39Fy{7w3w
zPI*Q~^9KCSc8|q=vFoG8H={aJBqh5P{ro`P@4e+oBU^j@GgLn+|3v$(3omT2`YnIp
zI<HHuZ=JazlTq46>QdXc%<nt*D6o<-s(-Mq`BGd_in030{HHx8OI7~q8uR6{P+3Gu
zQO%*@j)tvht9!3^tJP%r=6O#}O-a(hr!L#~lohYKF;kJQ^LaPl)3aE4z}#w)23<;W
zdARdUsjzU>CbW1TE>+jKwcgouw$Iv+Cb!oPEnPeE%FC?YkM6)*q81dkm~CjuFL`Rr
z65m*|n*aIH>V(j6s*5tX-aPJ>y=B&yi33}HPBYJaa%Ovoo*<WyDL1(RnN<DKlIwL5
z>UBYPV|T4ZFN&8P3LU@qn!#9e{Fo@OfsvY4is}l-hJU^7wZ2-JbHV(5#)^i3e%p(o
z>_7C!Oc#lK)gcM_l_t8cq48B{#oYq+*OgvAuWe7Wo3+|H*+6)x_{Kjcx=&YkNt$mJ
z42&w3%}FOK+XBlQ8%_=u7mOQDtx5{5e5=gRJcvkryYHo`qKJ(~L>TH<#*aLkS5({G
z^>5dm3}0jZl0mX=>6P-^BbVj*Z(?i*uE>q8i37{_HWuD+ot`i_!Z6v=62Eha5zW+c
zdnD)LHf}V2Qkc~UdXLy{4v)V$gSaudxEtYtdWh@>VFMTVd7a<z<WbCDMiG<kju51b
z47MBS4%BabG@f?QpUIrYp9mm0T#f+YOn_L|^}B=Jnv^`ICx{kv#kk$T==4|Rd(+Q(
z>gn39ganm{>;}5Cp4@u2D6sBZ+C7<^hw?-S2Xz6EK`^TP&Xjdty=DRAga{#(NrUYM
z#&2%&>)tziTL`EQ+|#+!al3);A57;1PF(M(0MP=35XywXb_3l<ILT(>;b#eSbryl6
zxX5mx`$SJfRMH3XF_0WY3%CN@ZlJsWsP%sO5LL-!im*xUC>M2y!rXsczf-h2GZjNT
z_z0m8D17Y(;V>B-F%Mh~G34YLU()xS><^P`?+X+ho9NO+YJ}g}ZjrkdpK`$nAxk(}
z5J514WRXrt%92hY*jNU4qM?&zBx&`dgR*;8CPv^gJZv&Q$Rv?TrqEnL_rA`ikN-nK
zxj8~L6cQu6+1@xzj1kf~QZ;N5Mntot@u_PQaaf$lfwY*`2_>=<r@ovQ;_mO$Q4<Sj
zve+yy$OjG<P=%f<C=y^G^Oi#f3&$&x_f%OsfGX59%EwWT3ZUdpD1}-9QSygjz8Q<w
zRM8<}loB2T6h*~2p#<tZh|+`B=ayW(o5uhe&6cPkLpa-jI6c^~Z*Znr9D<%@cns7A
z0|sM3CJ^F48d5U@mMeZ=w>Hd}dkipCk*WZA4A48OFeenJN`wgZ08pWK1Tur`0W_p{
z10I9E>7~nIOwDkDZ&Y8cH<$+xv4@|76R9NULZNr)Tt0nV_jZmOMkr<}RIn~agyZ2J
z4n(mio&!a)BAw6zY{4{#FJ2HYh6qj<U{CKTTub-_$Fy`%$m6-UdgHn0Z?nK?@`Yk2
zzI@UNN%>MIq(Bs0_$AzHpbEXzunb6-Dolf1AU9B)-d(utLpkc&8N0(ZbZ<B+r-@ck
zDm9D&z7*9;h(>tqzuUHn$HDaAv+;Py!jMytNxvHn*+w4yLS&8YKkRd*HU&Zc2e{XF
A6951J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp.8DD0CEE151EEE78C.idx b/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp.8DD0CEE151EEE78C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8fcd998f419c14c9b48a00f1f15d9f1c1b72347c
GIT binary patch
literal 1328
zcmWIYbaT^SWngel@vO*AElFfyU|<jg;^LB`%xGo?hTR+t3>9<sh8!%qZNYQ(e@Viz
z*=x36?cwjt&bPg-GB3~CvPeT{lEpF)UC)~z>uMdnyQIx+RodfJm=qe?*&8Gs*dCm^
zWU<hpBJ%v|PahAro_O&1k^H_ryZ8Ov^Td4h%QrqB|IOTUYffxLuHJu(;4|yaiJW&{
z^M0S&(Ywc=p54Cx<%hTTRqxmAjQ{xhY?YlV>+|{7r1$--`CD50|LL^&`;(rj9$tQQ
zk@=d)w^Avk0$wb8lh@u}{QMACF^84G%URc010|kBe-Y!=w#pT`95&1Eq4b~o?=njs
zw%X58x)Y!FL7``%{@uI_ZdMCJcd;mQnJ9MNv!4IY`b&F6cC_k4!>X)%&M#|}Cr#dF
z<>U~2>86#2$0akSyUx;EHt&tvSN;8=txAIXa%0{tD_VM^_|(!hpYosUa|sdnW_3zR
zWZEKQXNlIe?wU$7MZJ{wbUYDfTl~6%C1Y=PObu6ie~aNOwv);(_fNj4k?;+X4ZZxL
zSkft{ZA#{}$Nmc1w;MN{lasp6`r`3C9k=Kfw~ANCJ%8F)v0Fdf8W1%pbdpBR)um-T
z4N|g9g8WO}7Aq94ZN2!UB5k?Q;_h}a_Pt9_bxdO6s0-L_c9CJ%raq3hrE3FTmGCio
zEj+v8gT#gR8k53Uyo7n|rIMcOUR}9h;WvRTS1!KXvc2tA<tzI$>lCa!R##mKx;MX~
zdgaQwccZp%-ynW(*{iA>i*-AM7QQ$oWLx4GVj;E5!Qrw6_v)%+?q>7vcHG#~rnW5n
zU0_b4_~mUXliGF(rK_<1S?{c!*)qq@P10+@!msYX1#M&(U7sqRJJ+dXT9@TN|Gz=*
zHy2MRJ>{Z!NzCp`htYzA`sS~Xe%!qzKkMp;w4f<lvI8X4x2H&5bXI&aZQ0}ll6PNN
z>_2fJ^}?}Pn*uyu&3SDTIrUrdHP-oZjW<OX%Wf2%9p|g3_GL-X|7R=n>r+3>*!$L2
z`u?;Bs@~S`ZOV1ORiD_(8z|7YCd*o_Of-M4@$@e7>aVX#@_mnpyxZdPFXr^xRo^Fx
zKP~z8#AagdY{OY*&*tVlIDhi#lP1wyo1(S{rFBPdJSwW|{=@aea*HEBUvKh@amzm8
zZ+G%(IG^I=+cx&g70y?`w~>GUv@%TS)|`#^-M)5T?c=j~yKM91SrPy49*8q%zViN!
zHM^wTYlh;=+$4EmUfVM9{hn6ypBy|4{s9a@T&!FMU;-!u0Y#~4#jk-(v->A}PyGCQ
zkclUV+kx8^t}Te$o!bY-b8LUZ@bgQ_Ek>3gUOnDWm~e=nkot>>Hmr<5y`0?qFyX3*
zUeh(twrc=|kySS*NV%;xG%f?#WWsF$Qw`)<aofRoFq)AI?A+c%zZa}|XsvK4#O`Ih
z{Pp7?BUyNrV5*L<*dILEE9xW20Sep(FyW^LtY6sq#k)bmqTKQ@VTPjAoJ5fCi;I$r
z#ep<qP{XX<8#$TMY#0L>0zonmker*6lL2IUr};9ou(GjpaB}nT3h?uJ3c5P0337Ql
z35g4fNJv^(THDy#**oZoYFjCrX~<&}6_b{cRZvt?QB~K}($P0CG&VJtGm+9YauF4C
O_X=im^K<lI^Z@`P@I24}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp.E45DD97792659A1B.idx b/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp.E45DD97792659A1B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c0cba2ba8a5d6ea55c39460ed82103f63ed98f0b
GIT binary patch
literal 1326
zcmWIYbaPW<Wngel@vO*AElFfyU|<jg;^LB`%qV6ChV2{-3>9<shWHlWw&1z?|K))m
z^T_<wlJY{=?^SK{w7FYW^dvw<rO;<$)TB+1|9)lF)y%u|Wr9rA#13|*iXTaA6Bu`F
zGV#+8h`aWARn4CQMWY`(ezfbqUzcyE|MP5?l{ND{`^n+6H*>E(`&2&tz?)Y*74MR2
zzne0a%TG7I|8Lph?e)`ZK0mEHy59Wt&nXP&?AMyteE#|C^{aoUb@%U6eWH4F`O!t;
zYa-uDrIZSIvFuG=dwcQoLtMoiRt7I;U1JTDcoO|Zj91$#SLAZoEWd}+Z|>)1mON~=
zpJQ_;KJ9}-&qDpXc^BNQ7KZL(QRXsH?7U|^|DW}j_K56g)rW>vS@)b@)+kS!xXsGR
zA^6fwD-DlJW=waTrMGOh)w#d-_Xjzljf*czGU;k1E}qtWDw6AG^G0Ps(U!e^o7h^H
zzBu#1E%9|=mq%9XB@TV1o2`y7t+bf7+%wg)+j-dikwj^8ir}L6DJ6D2TBrIpWq#So
zqiE)Q!sPTv^+{9TIz*)N_r7NOQs*zVSl3C}=IdjYfBp9k`c>#onYLuA3RisTN?GQG
zZtV`O?LmtJCuFT{y_iz5Y`Mtd?shTuJ3(n7lUz9J8e(UkV%W8*kLy<Q+JILje2iWT
z&vI1oyeNO-sjalMvrS%f<L7CgLM{YUcSOB<QF1HW`K`@Y`Lu8W*(IU2i$uOJ-?jJ3
zm2-J(Z%5zY&JQlLeH$q5(iu>a+9|h7K}$w7R$)=*oYv6VqsO!Cb_;II<n{I8e;1gO
zD1Ld{5+&ZsrL!iAeAv&oc8OrouNSEvjcwQDzq)@a>4*!D+frI|G1BYV5Bsm)vZ?(}
z*EEH8dPM)YV0@~9ZU3D$;>X`l+P5<F@NCbxOY0mH=iHf;mBQaTA$pddQhLea$A6TV
zr#nh#cRA^;EME6{N$9Qfk>WK~ENS6AWk<qts!VszxG>r4@9gFK{!TrZ{B7I!oSNvS
z={j%se7>{w#(Tx<W-gAbDWSIwH?FRDJ!fago*S#f+pDt`mK7{l{q$U8`>K21aenFk
z+k|g!lU$sbeD5UNzd12}dJzkCSC?w1ZrdRm9dj|SCBCBOle21l=!MSH8Js^)szm;I
zRCnTu+4=LI&wsvWasGI9<-Dm!Ow-?NzxwlNj63r@vCB6-eNF$}JrZZiaOKS#>+T=2
z$;FkqN%Fv)CjRX6v!xR!wem3d2QUP2v2y8y37`xF6s4vWzXmb|*LO;lu6uQui6@BL
zgxeghEr{Ea+Xlvi(TrSRLwXPWUa;n&wSw9G6TT;Y{yhlNYsc*jQw8L?a(lsej_q$4
zets#r#mExGtH&Dx6AtkcQhzbghLsVhnv<I!CVVKw?q$6E_2WQcUKU;@nDFrx`-3NY
zMSTRRR^T>(2|o<JdF#K`0Y8vxB(t9yuzq3Z7w-nC7Uh<Qsb(lj%}E6LzPKp4SRBL%
z1QB4s7*xRZ|8=*k_&>%#hUDCooD86_cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQu
zh=intrL~Q%oxOvesJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcduY3
KH$O)YMjrrZc|rF8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp.0A50ACDD7655DC8C.idx b/.cache/clangd/index/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp.0A50ACDD7655DC8C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8c27027b2d1f229ef1e6239cb6367c2d503f759a
GIT binary patch
literal 1326
zcmWIYbaPW<Wngel@vO*AElFfyU|<jg;^LB`%qV6ChV2{-3>9<shWHlWw&1z?|K)+6
z*E!p+_V9OR=iA;^nU`m6S)?H}$zqv@uIJ5<|9-LRYUbVfGC`(lVh1}@#g8Pm35+{7
znfPf4#9jNms^(9HqS22nKW6K{Uzfj6=jYk1s#3O^3bplS>EdCtZ~iuZ@ahv!#k-{1
z@4AfT^QW2L|G(()_xfoypTE`}U2p#S=M;u>_Up}SK0p0+`_;eKy8HL3K2bgT{OF?a
zHIZ+nQc4B9SoS8by}kJPA+BN$D}$G_uCWG6Jc<4y#;a|WD{?t(mfu6^H}~^0OCGk`
z&#}34-l#(0*oEnN<{8R!FKER!39`<b;PQU1|9{zE&TF==o%BJ%*7W_tl6`_FJ)&h5
z7ihlBkrh$Nlx}>tkoVS2xoPk3)&61Vym9dbPovn>4S}Z}pX#*!Ie24Y2iKuE_sHf$
zK_zJgi#L4La8=prl-VLa@y($HC9<N7QSYV1=ihm#T-dYAaTCYIaw98wH_=r0n<l?v
z+5~1Rrp$W!L-^#WGKDpp+mDqpR_#~sy(qR&aNe&%rvJ|G6V&&Jd3j&*5^7z)>q=i^
zz+&eGhn!bl)bQAP?NEl1&E?>(i$|Tine#M_wLFzr><_H-O=XCUbZ33LF-)V>s-0o!
zg)|l$wlCJ7mWWQg<Z`&b>qez_rB=oYTNmBZ605gc7naTY#UDOhKz2!}?IMxy%XjU)
za^+m!+S}1LxbuU{Y~KcoyL1MWq;|^fQqYnSja68bIj1$W_UQ2}yWN5tGkJY|_}>NQ
zB#K|&wnT}ya_OvzA|LMatz9Bm^y@{cM`PPH`LFJuN;=}g<F=F*U5xa4_QU?Gw`^*^
z(=|<@ogUFYE*PI`VB3FZjrj5RllH9)Jv`ep?$SEP#5s2+Wu@@9PKci6r<7i@`0*bl
z=IM^o*<DV0D~s2CUJ`oie580y70Wt3$+sPAZ+s~#GF+%W^_RJS?Qi4e<8N}`-~6$S
zML+UZ<@24{8}B=>%k605I<ab7a#HxmYnH~rdv2@_Z?Dc)SXQuH_0w~W?W^v2$DO-p
zf737Jc3MwcZ>75Qk7T3SF&mts!(NA-$|=;|wqwy<j{P5gJZbm(w`zg#^ov5DCVOuB
zarEzmlezQeKcD}6Z{fWD@1Jb6JG0N-$Y1@rduJb;jqb9|lV@f9+1<G}i($phoBtc9
z$wn7f<|fGlbJ~n6qMQ-8dFy!?`~w(*xLCRL!30nS0*X@8ieCenX7^9{p7{CqAQMjz
zw;i`LTw4&gE4LSn=h*&+;pdl<TZ}A0yn4JLFyRnCA@vs%ZCDwBdO5lIVZtz)kqc~P
z@1frd);zRUI22;{GG6}raiD5m7G5Qos^cs62T%5j`Uo;af!hEkeBN}`tWwirHIQl~
zE1w#$eqrYq?*^$B<(7x3{wol+>2Cg!XCPmgaGS$~fjmoY8yJtFC^aV$<on{H<YI9U
zBM?M@0b`K$qz~FopHi6_0~wNYQ*ttZ!rp1V%q*;I>>QljJiG$@e4c`?&T4{Oo=!sI
z!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#KHMMl~4GfJ<&E-s_bd6j@#oWDu
LncVywJs5oeuPQE}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp.C65ED37BE0626399.idx b/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp.C65ED37BE0626399.idx
new file mode 100755
index 0000000000000000000000000000000000000000..114a030838de07a7899c382ddedcbf5e495d1eb2
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%y?!7hNT<~3>9<EhMX?CZN+o-e@Vho
z-&eP*j&&c|8ht)j$p4n#G?}gzl|Id&>HOCF;^Qw&dUWLEvdEHM6FC|zEGi6{PcY;}
zW}Q|oP=6H}YFA_RSh8l`|I=dT)5GtJy~(niH%G(HZu0ckletfS{rOw?=#xh-A7zqC
zKPzkI+3ByVt6O;Z{Cn<g`*z1}I{h`vOpP@;K0JHdzp|gnmDS~E_11fSp7ro`%{|Gm
z%3SF)+2>WiXU*PT*4|fm`0G9G`P{d+T>I>#e0zcMiqA?~y2Yt2eM0vi{4eug%XR$j
zzm9_Hc?WH;&)QeA+fn#Mi`yz86a9eSDpu*Szw$4>bzbW;wWa;RipuXxOL_hMPOV^&
zEA>oIUZ~(Xlc9vaH~d;j*2i~um-kQTke%aaa$(UEHI0k2*DRdiKS#Gy<KX)c0aFj1
z$}6@#N3Jb=(0E3u^;sYLr_XyYx@j?OIcH{ObLb&+A<rtuPah7t3a<Iw+_`CSj#Q~u
zGrzCA(b=L3$4y3pJ2s}>zwu(>VcV!JMjX85SNa!!sxJ}Sc1Kfhs={i8iMzFym3?WE
zdTJ{w=hYhJv})_4sELm|=LYeZy@`JFwasYpBo(I*F1fcu*aD_*47uQUy-2e^{r$ox
zUi@6`a(Aaqn4%KYm9gWXz#`)*%UjxRIbR8KKmPdKtL)b+{u}IUNKjn+%F(+p{>s~}
z%iJ~_mZc{gpX+(EbPN0J=c-PEr-fu)&TGr?Tok@&a_-KFb&s4)=ilYHv1OLpvfy_U
z@)M*lZCj%BwQ}jKiBCT6?{K@o5n1Z+GI+tlQug1CHnL5@Qy=HfRamqzh3|jz{wI!Q
zCtT+$3A)@m^t&&R&E&_^DpUI!@h`q>U)}d{^4!hVx-H3hS);8;OiqSX=bW%!uFacy
za@cZLXmU9QJ<lmgnR@E%rx&qRENbGGZ#!0g`grt3j*Gw6*ZgU}zWKB9-@JWq^TXVa
zeor!=uT9=|G>?C#kzIjSLfqDg3~TMCueo#bKHt}A3OkRi6*}=_Tavr3@uqv6VN)KT
zT>RK$OYXGGQ=XpdVzcWx=bm1<&?jznSVvCXS@RuBmVRJ<BD}+$XVvHTt}-{@@TH1W
z++LUB7JVXa(f&uTj5p5o^Y1p8TXyd9%YuBXvtEyE&VJs%;YR(WNGrQ+QSbZ~)5Wto
z0@^pnKbtIC!BAY8n`8&fZ>B9e-+umdZQ)@E2xJK2Vdc>W6F?a-Sgd`@^i1K5Z@dit
zKxr;9E(I_FlP*e4D_+9Lz@WJzcHXyX8yFdRg7_r(RAGj~Xhtrup}mKGFIe-?TEXo8
z3EvYx{~lxls<!2Gf~f-XT=+a;JjeDo3_rh=++t)25)u|NfeDBB38}xBXu}G$iI<a?
zA12(KAmz5&(722RXowD<0ZbUkGvYHt@Ptf-ERlHDLiR9T>s*fuow>KB0L?|V4(5Mm
zF0j90J^-5Q%@+#O3)JE*1m`V&bM4`xuZ$f)leo;etY9YWT)3jvYTA8ukf$y9tdaCO
z3c0{|heGUL#>-zn4%92iBB%rtermw_g`Ho#8z{`j%O?&KewlrxF=R%Z4ag>SE>oCI
ze+9xe-OWGp3@9w*CIk=AqSTy3P`nivB^Qf>7{MR{3>bqdU36|1>~MR<7|f8Io05|O
z6!uQ@WoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!
oEhDR-sHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL!sO=X=)vd%0ErZr&j0`b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp.322ABBBE306F5168.idx b/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp.322ABBBE306F5168.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f3ecd0d73d45cd2ba6096c95b34df5cde3887b0
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%y?!7hJ_pq3>9<EhIkj>HWPUE|KysC
ztLtVJxUol1l=4lEeysLbnT=(dikrfLM~abuzeaW}Nj<&uNV}iDN}2fYx7+no7$h6k
zYUS=^dFNuhe)V#HahvAz-TQBRdiU$u>q&RlEZb~+^mzEEtm3l7tvmm`^)cLS6}+b{
z;(o1n*Y5MDXTPt%^ziq6?`@Sk;y1oNTV<!pnjC*kdfU&czqzIV3r?@g_xx;pc>2*r
z=4(D5+cv{|p7*<}dhg%J$v$p--5hQwzU}Rr=S$tTxuyp^PYsEBc8Wt*b>HFtf9FL_
z?c05r^Wi_EX7_cbwJ*!M)h=-MhKOdwE&AqJa&Fg``wOjG+-6#F`X3DZET{cCYTleF
zfs8h<y3QSQ@|<ME_(D!<{hF6oA02M@msb?Dwwjx{!1<(aK$1bEvtqnOv`}Ev{nZLm
z6CysYV3+8e>wJ)TrYhH2S^g)_*%$YQu)LU)n_JO#SpAViNcWS6&D}~7&)tPn-EVGr
z6~iWfOEG2E(+_Paxk?4eXBTe0=-gkuDI--N>fH)?^{4w^=)Bnx9HZqJ=BQL&>ho5D
zW9zAEEt?Rott}zhN4F{+H@03Xkz-L}@q#ziP1UpILF=uul{^cwl2<O+yY6w&ymR-Q
zk4}{n<+I%-tElO@MC4+@Vnvs94S!DEt?et8_W3Q}d)53};Qz!j)`Lz_R~Juvv~T6@
z^RuMRUc0fk(Y9PI_4;MG+<D;~scW=K1Z>wAnz;D8T(&4y{@vAowWd@#;dbV%8F4$^
z_qE1)+@92B@?>h3+oa?5N>Yx>DOUwkJsR7t$-Z^}RKnqRwXfW_siSR@<v;(wliY7A
ziu-oC2xM~0SG!4Gco07O>!TmBm&~_jRTy)szLlDCt4nvL+qZzCjZ0su6h&>-I-_s2
zL9%4wLJbzrxtlLe4An9}zqtIR;EdR(Ta=fbTOK~So6~mXmHm~kZp%yBq`t2{e{B0v
z>xs+fZtuIv_uDq*)Q<xz8!GR*GHv-4v?)u!-ttOy(-Wa`FNMduXU*Mm=0uu)gy~${
z-@Y!_Q!RNT->u1b`2FP5CrbKLPfB0qoT!`kQ!IRoaGgs{<F_8i)PGAGglF3>=wW)l
zd+V7D_M6ckw$HN@zP-p|y+h!aeXfT0w7b?Q9dn)c|G(DXcp>lO;k8LGYd54_5)jPF
z|HeEaQ1}Bwab<3j9WcMOL|kN)6H|Q3!w?Y25X8gEqX#B{GGK5!U($N_RDL#I27jP5
zml&4<n1D$arKS}xVFc#f^_^0s>s}pZ;R)i?;nRm38pLPFXM*4fnF?7T@vMYwVLZ(h
zvGcx7+rY>OG(dt+6(($U|Ag;}pMMWB0fnvk9ALsgo)e!tjOW<?hT-Rzl3R={K|;bp
zaO*<+gw$V5v|$Ar$;-*h4^s`J8M(k7?LG8+!J3EG3axWJE_CMJngUdfY{=p_*B&nV
z%Gd!E<}&B9f(6IUg)3^UrrlQuxyy{t5@se)uf32HjCUx+?q$6E_2WRjf-HhcFkzS(
z%v|80ggF3ck|$p<OcBr|Pa!z(sR8R3c7E}0ph<kZeBv-gFSD;ShRlew0Xa^c%M@nt
zUxBbqck_=t0}2bd2*HD?C^aV$6mP{v$;ILzMlgr~1ID0&XB)EC&9`V{3}#5qP07gs
z3VWydGPAI<v2$>8^Y9Ar^LYxoI;#nCc{&M+3yVldT3A}!*xK1U=!t4uDVu4?V-ppV
omXTFZR8mn@*VNL{H!w6dHJ3Ay(lv4s6?69rVRG|x^kDP>0CVPR3jhEB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp.BE9138723EE82CA7.idx b/.cache/clangd/index/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp.BE9138723EE82CA7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b6704811cc3f17ccd9942704529e1c99de6d375
GIT binary patch
literal 1538
zcmWIYbaVT~%D~{9;#rZKT9U}Zz`!5~#Kk2=neogF3=26J7%JwR4e>6%Z6@&S|H(BO
zuIpwOxUol1l=4lEeysLbnT=(dikrfLM~abuzs7bfNj<&uNV}iDN}2fYx7+no7$h6k
zYUS=^dFNuhe)aP6;x^6t-S#&>z5Dg-_N2ROmW8Kx`LBO+_1T+4S;oIBmhISO6TGJ_
z;(o1v*KYmOv*qhAJUkxny{&Rb{KnU3tL#)+ljE;TZ~IyHx3=_u!RdARo}Y~mUq8CY
ze9h-$+h&;0^L}?#@BJG&*~e|Ko5StIx4m8Se5u<u*YtqrsUcC%PI1Vp?mPVd@4Tp~
zeY+2HKCGK|Q2F|-eI>hHg<rI|t>QA#5B#lSl^*-+{RLU4#XfUemOofg**~>Z*Wd5d
z3I@4S*YxCtDxNbLO89%juYJk-`0(&^{|OzkbI+MvSolOu!*Ir$g%kAW=yqxxd><;n
z>#?Td3Ukkqa|<6do)Kz&*2n(obMr+vEvAxl+steZKNK$HS>^ia!$DWUHJ=xEZd#lZ
zRjSv_|5m{0Y*B^tCNseu8=qac^<v?1+sG|O9J=LK`X_&qFA*!dqp3GlVYR}<-B!!W
zcv_;K+KS3)wMIFu+WIJJ;^UcfgLuqjtYp5l87-cq;`G5OcXtR|z}Af+7vioLYWk<Y
zU--zYpQ~N&Zr_9{DnVTtJ1z<=GM=)$W!f$0D?#qdFW-A5{d&d!f>O?dPEl7EPkXd)
z<?Zvcq|RQuvA5B-TrKtbWx3pW;T)-Jv`PeQ*B6?&_`6)TC|3U6)ql07R5{^x=BycU
zJKguS#(LbI)MfHyYL?ri<NKAR9F<e93Z{BAwq28b>;9>P!|!Texo=ZP+a$|>{(mRA
z-&7R$?Q#*w<d&~?lf3XCeD>EzKVmPLZ_TPO=2U$vHRV>9?o79D0Yw{^zEmlS+NyO%
z-)Mtm$-;#iES_^WUz`}KWqy8f`AfkayUt{~`^@#%R~Hwq3R+qJ`O5bD#ucaTeV^Bx
z-(9ZcZ=El@S@v7?i76i&gBd^X>SoFK>a!_Jzuxjnb<-1}axaC)yJyYaa^^&ueuU}o
zA8*VBW2ZkgN!hu1QR{uR%_m*sv?iOc5>|@7^GSDIhWek5A53p0n@|1mW>hz;c93Mf
zSDrodBHyO%5Atnm)wa17#x*W_QQMugCtM`bwWr(W|Nqcm`&Fj(t^al8((i<`OB9u^
z+<Oz>YAXDKp|~<P$qty`jGyek&;FIekB1>3kRgbNl}8Uu0A;`+WSX=6H~V#-ybS(8
zX)ZA?1uy}VE=o-+p3TU>pt&M;-nVHR7#Vqj_$2sLVTQtJMlP_Sy@!4;So6?Y!R-DC
z-xELo9%KTlw&rtysRHtx_}pPU$M!c2KfjdRVq^&t5*9Ln35WOzslS+L!wR&Cmy?$t
zE(~=iGZ)x>Fq?ozdh!Ls6!8QJc?!XKt#dstbmrcg0yGKPg2ivHJzVsau>&Z~WzJ;<
z3xJ&qSJYZfyRQy%vl*Wyl9Bd8PB7k~5WAQ0^4E_8^$M~GD#3)G8nAw0=NIn=3iI*u
ziNl0nW?yLxnGt6LvPqrG6lT+3fv`<?^N&1Z1*+EJ(}$@B@(lUl!2oo#sgMOiPRL5g
z7K!H~1P@1sqSTy3P<#~^B^Qf>7{MR{3>br^YVcNAd)}JI7|f8Io05|O6!uQ@WoBVz
zW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*
guBoM?Z(wL_YA$CYrEBCOD(3DL!sO=X=)vd%05n~LA^-pY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avgpool3d_bwd_ndhwc_ndhwc.hpp.B8CFD83A7F04917D.idx b/.cache/clangd/index/device_avgpool3d_bwd_ndhwc_ndhwc.hpp.B8CFD83A7F04917D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a94d1d3e3f7a5d45a2d41c5e32d46c8c4ecc69c2
GIT binary patch
literal 12716
zcmai)2UJu?_rRGFM0l<%ZNUYXmG%gIX@V$%NQ(tQjSZuTCYpRE)`&(8h**Irf}lV^
zQADH|fy4?%>@n6vEU`B<sDWs#G5<k!oSW}|&L>BA_IK{x`=;KxbLYiHM?`G!HZbTr
zK00yogc;+gfq_96_J8_}X_KB}0|VLEz#uVy&(JG_E_x~d^7*=7w_F-GBI5iBqZLJ`
zio(2V*0*>KSpD6e{be7G`u<F1-xpVA4LqHd^`EbjhJIT7yl&US2X(u)H%)mlKcvN^
z!NMgmYyP|QVuE*|hN{q-c_DL?(uZ~k8Xjt&kbC-1#mk^|XB@gLYtfvG+!S}_Zurac
z?*eR_Tb+K$-@4MR;Li6m=Ukk+vuW0x#zVd<7qp(eHBU3B@mAkM6Uv=d8*20a*|)iI
z=Jy}XIOlrC?p}OT^2`DGp}%Z-ukq#ZOpjN?(=R{${-?`LAw_%cYl_YCvfp{X+s3(u
zQ=O}uM}9VY@4!3K2OIDFQ2cyp+`8lCV_jyMf3R=TqPR}uT|b*SX2(^N83*UZetf+m
z*1Btb*o}F<!@3N1Td=)Jem=)<Nd)O3t)Kq%^0F^R#g4siFurrJZ{pr3!y3Cq4vxQb
zDDa~7mP-|DtGvVR_SDXw<(h9dYd@`NTBsQ45{t}+=+4A{e(KzSn2%~6=hu!Lef86%
zfUsd<>wkLev!*p=#`$s!!<+kKhfh?0V_{?-yn06CNZIO>37>8A81VjP*=bQ?sKbK9
zTxIixX?>E`eRka@ZD`JjjNUoMsY~ZHz0*3W!9D!bY=`o-37Id+s~1ZOZR(nz90^a4
z+S{;jpY`Fi5yQ)Hd_{1VWur!i8pj7WD&DKq70#NhIICE{VCteBH`nQ8!{b}K+Mb@Y
z*uN-Y%nu(=UQ^%D+hMBBot;;wJ}|N$c<jUCJAJ~<SDHAF?tksulNrB7XRLERFu~0{
z+wq2XLD+^rw?7Iqu3z%m?5*Ej*nKR0_UF&L9?Z;IcKbwD(Hig1F5}$RozWLNmFx@5
z+4}j3yzdgO_VAhDv*A|xVK<+x4NpHm?!RJ)c|_%e-ea`^!`uqbJl~NV<o1bUr-O!u
z^$V;>z<tBafLsTYMCX4#shiSo$TvHt*mSlYQ+t0)$c3>F&5P2%ncVG@tOv7m({IfG
z;%lo>74t?1?TUH&VWM%Nl^61qV=V?1P3hEk)TE(Kk!5ac3$3hw8?)istgq1eE0>?_
zOr1V^@`q_`F(|fLDQ=t`^}+(1TN}tOS{Jw19|lHjjj%uG+jk#te6Xt60-G?gTUR%m
z*B@N$zM-bd&?vsr5O-qY3Sv@0O!Wm<3H)Funs{f{Z4>+sQ>n+s^;o4ZxJqwBb0bu{
zu&)XBU@Awj@eypTFSts$RY}w4E<3kdU}L7@+SS$O^~W2rb8!d9@v+vXSi!_siSjD3
z*B4wRVtvN&G_5vAfi+zX<T4_a5w*VHDp9xky-@dzQYx?)Qz;@AMZ`{DaFytUs)>P9
zb<GNF$5b{Er%j}<zThhTr@NXQJ~BAn1b1L{sm8|DtTpWgR~f$G*eky%x><=$naV;+
z7gCkJ;3^YpHh*F^(=<$hz1f2biA^Dq^#xa%{9NI;;8KXU5}PqArBqu=ee?xa`P`<Q
zBnBtWP~dLN$_Ap{K>YLtSDEHkA}?|7@~a89WL6Gh<Ac~jUvQO~^v8^gy9dN6aCc^<
zm}rZMzrNrqiFUWoq$NLXFu_L5N*y+?!(H_SS4oPSJF#lR*pn7m$y7YF9$v3MxOhS2
z_x@9s6ls-Mr8bbOh@y&g(HC4LWt8RON}s3>mYA@5T*5||uu@-el|_S=R%HA!^cw}X
zVJfA>zLfOR7hGlW{7SElKG(ApSj$w_5$!tGn)ZULeEX!A`{n4cbr!fYtE#8g)9dvI
z7Z)zLb~|Uk(QyU#WEM(@R6<nxf~%DJZf&i{i*wprbDk*9v(~g1T&321`S?2Lm_QS3
z$R2VC8y~`*^aWSh@B6I8e){@(O5BmDR8VCFwbd6~<ydy`nca=U5*65%)nz@=t|z|w
zf~y=?_Bi|ZzH12<*n(Nvk8Ss3Z+*d4enAH_u6{CXf+_C8RCHA5`1*s3@4m_!-^qGe
zcSB5=_#!sBh|Tl`SGm_ca*yZqqn3u4S{cap9qb)ne;9PcNHQ=$GWzT4%w&6i%Vg#2
zCHjhVd_`>J_JZ4(HsRyxb}j}6cqR5Ozjh(aP?9$x^G!(R)sy7S$b2*8Ey#Qe<flme
z6nP5$=Sck=@=K(CDVP+mSN%EruP`df6Fnw+i=R#Mm;~rkk52(T!XA&XH=m1!yZr2e
z`&I5SmgHRIlZyrlBgIH6M$Uq_BjtAFD)<ZvJcEV^zKDD-!r$LRx_fAVkUvJ!V`y&)
zmX=^(zZkn0<0zrO1xs6mypz|GwV!5tU&NB^s&RGX4poxfG;WUE&GX&B*!cA>zz0>5
zTd-RT4&aTJ<W{U{#qSEfPPEsFTe#j0OY#b;T|wQrpOfU()L}LCgxr5MeP3`6b;_aM
zkO$?^k%DunQ?8I(HGXIGa?=kwOy%E*>NjG=mm8InsC^Q3;FC?GNDf?TbnF^ybQV=*
zQ3qjh4OOiXtlc?nVsldPB4#m*nr2Z8u1MuI)O3yD?jbchT6>%{e(R&{uf`YrFkpjS
zM@im=G`o<G7?|Bivm5drq}c<x4r%Hj*CS0m<h@9<7jha_r(rPZGq5@X@(Qe8A()SK
z7vtAkxftUau`}XCC1=LY1oUO>mw@Jk&gm;O=Z4M&G(U7cpkx%2jQR`RWR#GMJ`kLO
zVp5<y1tp|Fc`AxYh4NICkP79cXhbO*CG3@<5oM6e(TH-$Ye?)G636GQ`}mhVZSlN)
zY%rUMTNCl%6PU{Pi0eJ#&Y8zK#8@BQh+ZqBc4gF=@5Pc_MqSFN7iW(_<0ExjW-Vt=
zcuWRACh^=Nm7kD-Pe?3hUDVBGhSnqeUn>SaCP7@#rs)5K1aJ=YSrm7<VasDllGh^D
zTBH+={}!pfg<OKnON6tN<O*b70eL6V>_kE0_dg=dkC2;@su|k*1F8Oi`~<0=Ab;`u
zS4jN|@-nPmhJkz~R<9Jy14C-84+lk`9><}_@jxCHN^Ao+FC|*Ow@Y#z=~YK!c*GFY
zt>x%vhT~d{sl1I!+o*zXn{BMyMtucWQ>hxttEp=>b%*lEY8oxLhDtS1zlOTjK>ZpT
zSp)TJsZ<N~YiXZa8piGM24Q1;bSqk1OM2Ck{=7xpId?XG9pwtOIhJ#*M1Rk<oC|2a
z<$OSS$RiJV3Ee!T%R|0`^N~kBl;<N|K9m<Aj{+z!K)M1b--Uv9p<rQe4+`1?xgG`8
zL(V1fxnw9`c#^!2jMzs;y=kz1<3PsxsGT=C!nb8r_u?)ENMfN;k3+aGkD7Qm&pi7>
ztaylRxoKm4P|stQ$TRsp^G5cYu9M_(F5`d~{?O$^Kx>h-7Fmk+6e6h*auJe>Aa6v<
zji{^8--KK?L7!ZTTuLEtMlPEnmm!xj$Xk%h7Rcqur5ti4QdYv}ZAHqhkar^Uov53r
z&pu?n5AsoDeiZVrNct7F=PYtO3w=W)a%_Zr4oT;r{&^&whkOZ1mmpt8(q+h3kaPv|
z4dikIw(ll#xe2)$NzG8d1xYQCTanZX`3X`!f&K9mc|C=`>>2WU2KhPidJg#o@_GUJ
zCGvU+`ETU)H{^ei@*nuTWUNfaU_LCv%4Lu<u`(0#TC7|Pc_WrKVle+ouw#kXvsoaF
zrA<&?iltJ><ybDqU@xe^as}i{ELTF_iaT$`)_lJ3lP_CMcd_Cwwic_Y{TR$w6F>hl
z#av>QOZ<dl`(c<X@>4L|G273+d_^<wjYI12Kz?GBnD-B<#{>CxFd!iF3rqL36>P;H
z#61q;zWjhXU{HYSaKGHaOmG^HK8-)%?le|$IH=={hx3@AmN?XsVD44oJNDZ%d4e*7
z2`XsE3fh@F;voY~otst$zhf`SI&WP7?*&Qr@%91a>+K83&)W}>zqdc20Pg@mJ-mAW
z3iJ*H6yzNQsHb;NKv%Hk6>P(MRFbb^%d3#DVascfuVc&WkZ)ki8<1~e%bSpIVar>P
zZ)3~bkndp2JCIv&mlmuN^;t(cts`dq{4B9!W~cSUjGsUymJ;YxOw4$;Kw^o8P8*0B
z&q1&`Oge2MW_${@ZR?#%iJ92e7g5_q)SgG)lDwGOE{42>+Ae{-l-e$Zyo}l|gS?#D
zE{B{(ZPOs9Q`>aN8PqmI@UU^&Bg)rZ3}UlqJ2l%*E&1#jZg$9fY^9YWTLa^yao})z
z14%c;VU)Ln?P={99=-w+);8adR6btcS(2yNO#yq*RJ*BwrrAvcG{boYh}32}&jj?P
z^Ot~<kwY>9vnmBSq(DwZ4ylmSkYgI^A=W@Ta!iMufgCd+CzI~UV%<w}3hAB#IhAxz
z6`b%x;D1ItIoPu!_C;)a5xZ~)A+dCb)p_E;k8To62H3wMp8V*>3{&$IYR9{A+~SiZ
zpZEM@9+Bi^WSWdDM2%CBX$s_2WST0NXVTbkHDN^)?#d^$u|A*1Ys6!AHel_^G(KQ_
z1~UuyvEn}NE-bXC)VPHyJswVtKQlwc6zi~C9qtL8oLfEa$(<bcX)N|_&$#hNN%Bb?
zdJ^~NzKwZD(yN|CahE2^$B54{(wp=2veUaWmyDgqPMMk1E>j$6+c+YV#_;Zw<W<ye
z6_l@{{Z`RvAzw|quBP1}hpna&f^(=%ju;L}&Y}Hs;O}#(O|Fo0FTlE@87rExz38*{
z1S0P<-bJh{+S7)jD}KhIKjXooD>f1HO~gU?k3FQ<9x{M;1<x<C9j7h1D9mJ{KAC8c
zm@W;(vqAJA|2Z4=_M{^}>+uRRL3^r^*O)sj*0eRmY7NnGO*U^{CmgvV_moWW9#-6g
zebAn7<ciB<$in;I{L|w-Np`VzvEx}7N%pYzu;cBLWKU~PJD!?kJ6Es{9JGCceE|6e
z`vUR{_5<V}><=g)H~>(O;2wYig98Bt1qT7@8Qc?4FQZ<aMF;mb>J2EwC<IWbA{4CL
zFhv-ka78$veu{p8A`}sTq7+epq7~7AViYle`YZYa8lV^eXrN*spjbsLpuvj4fZ`N!
zfQBfB02-<o3TT*O7@*;b;ebXcMgSVA7zt>!Vl<!xMFOBPiZOszBE?GK-Pk%uicH8^
zNRb6O8!56O@5dJVu?<fOOY#-0y@F*vhufkK?N!Lvu=X0{>sWgo@(rxL0r@7@-h_M$
zYi~imjkUKS-@)2DkXx`#i&!d>{0u9f!6@VfHh+QDqP~A)tG}^E@H%3(j%fJWYl~8>
z))NhkQml%JhWn?sD8*_6(QqFx$<;(zO;p@1w4GGV_Yt+=24dMjYz1E@rq_uzPtQp5
zBFg@OROe#K0vW(dsJsO7QYtTnyo}1rATOu#a>!{^PJ^6I<#fmyRL+3Bf|{<N*4&RT
zzZZUX;>XiozBSL3*1F@zhkZ4--kQ0}SZ=EOw)l(hB`U)VB9&nVk;*WGNM)Eoq%zDP
zQW<6tDckH}2C<D{29e7D&LFli%pg+v-x<U<h8e`{YwzyuQ@7zB3-YR{RTXvML0<Ng
zu?Y)0+x*JnmMY?2MFPcuSCRLs$oqmX5UUF?j=4bmFOV=E*Q`Dgwdu^F`EG0z$fM?Y
zRKrK2Emde<Mm0R_D0o(tdQP?6mIXMMkm?dVd|yJnE}@~q!Iop6a{L}=z9q1<QrlIP
z*!Z$hpKP)5@mniq{{dD!5NA9|{tGMr5=+Whf0xB~rwx5C+J^0Wp7+)k@Fuezv;C@z
zZ>8mlyWWYNygp)xBo`u8A=2=wx3N<p@__7Dh<ZT|FN6u;LKI&J6S_r6RRrx7A*Ui}
zuL$`SL3>3gya?JWLh(h=UKKL0LN?<0Rmi>yUfNV4T@}m&RH4u+6b<!btI%-4TakJz
zvKRavQhx_|8&YqBT#eM#kZX{-2J$|n*#|Fy8jz*|@_wY*5BUJn9DsZfX%0d@gfxdB
zA4ZzPkdGkE5y(f8<|yQ!kme`I$B^b2<e!n|XUNBq<~ZaNNOJ=6Nu)Um`4rNef_x6C
z&cX3Mhji!Qc$`P7^H6>sIh}{|;WAQPhVskE=`xgGL8>cIeg!#Qf$|olYJu_=<ka$S
zxnB#MZ!IXi1<tn?6yE~pTPsqvLVK;qsTJC5MSiW&UMmW3h4xxed@Hn<gw2yMn7>Kb
zJ_&<M$0V#v!r-)+ghP`sn7>IlHVK3In~c@T7|h=VSiJyp3iBI~7h?57$f;PJ3ON(2
zGa;|S>Q#`lusREJHdbdtUX9hOA?ILq4&+>{&V{@NtJgqYi`8o(=V5go<b15ohg^Ww
z1(3hR>Te+zVs#<pBCIZgybi0^LEeN_n=tTyrC3!8xdQuC;GV)?RAQe>!4*D1u?Gsq
z<}g=WL2N3BGxxA8nL^D=#b#17^Kf2b|Glf2v!CrgJ5<>F%Ka-q^W5jXA)Z-h2eI~-
z5yVDv3WuJ;;o=}xMvTgc2~WXG@?O$wFX_W~#M<BI)T)-Xma)j;f8Wl2@4qegt5eNU
zh_cPstV(29b|3zAL`Ba(+hULZZJ~bJjR&J&&WU9f4iM`DB1~%te=uHp{9~#Em3wJ>
zb?5D;a&K*KKq1-?K%pL?o<cLsBMeZud-$7Oipu@m`vHn@j{p?q8TIDgh04*M(STw+
zV*vG+`@4we4Uh)_8YmA06swDUlMABqVBKIqak@A_Lwtw$37bQGhXNYrI}Ffp-Ei=`
z5xNn8M(Rca8tp&&O+Jar3H}Ly#`uo`^kLM8Z+r=rKZ^PY(D<nFfF?vu0Q7Ox$AE64
z$eZZRiwso0g(7c3zKtSpL%xF|??7%skxhadN@G9o6k)XRt>Fs~{aE<Y{;CaM00-Xv
z<)h#C*QT+>v>dCK<2OSn$>~^~4*46b{zmYj()##SF(cXCg>3I>AHwY&y1O>1pvnJ~
zVcP|z<2LHdpG@UyYEey9kbSDDpWqs5Q3K^Q)Tai@YpHWBb>aF))rr?TCMw>c>^}Qn
z2?#3xjupRSEk8BU|4P!(|4MWneY*RXry<R5tOw?p&avPVOp@oC&IL5zbp9LS2QJp_
z|9+OS{`DjtUh<PD3t!vcP2$}o$&EO)5f9~l=R~m6+_*tDtKaGsPMw|kzN_8Aui2yP
zux%Z7<vFIdy9V2O?8?({r*7~2VrR%Nce+q^iRktayNmvKi2FRmp^%3>#Bc5|sr(4{
zc?9K;@X$win9zTWEgy?hE4%&2{*T2AG?azhj(=e%q5ljAJi|Q%zra0T;9$Xj<B-2`
zAHm5aJel-^JTjS#5}ZQ9Q=mMBj7)*@R1%&F<*8(3s*v*mX1TPsm%+q{RAa|#9KkPl
z*!(9p>7*Bb>@PbeXC!4$N@DZ<2(~$b`|zRr_32MHy8K@0&)(?DL#jOF!XF^X`AC%y
zxd5pOAm2x-`;Z?X)dR>6k?Nt~Mw0_B@k=vDvY~ex?KDnQ+<CO~8*03@WLU`=7sQk@
zk$ol#6t&4l_Su5Zr8Zg_&g^>jt!LmXext7^SKngqy2&SS^a+u_z1HJI-zVi(=&f`1
z#!hQ*w-)CVwv+w)KKrKCMz<BWJA1Io9wI)6NC4kvCHXM%IV|{=*E^3obXF~AiYc^X
z3N;f4%~IO2l$!B_=Itxt8L17o7O_XQV8<5h%7^cc+l~q!+b!!^D-IBw14PFk#g}Ra
zWBpryqF?`=8f`l)TU8M|{zHlF->$zAAAT}q*(LJsRiYEwk|Y|CL_<W<KabkwQJGhr
zKZm*GwpR+H9ll(?a%bQ9t@%6ZJ-2Ed-rB2wD(^O@EFI+gdI0~|TQz^pJ2>C+#gCoe
zuBd;Fk=3aE<wah_CoihXRre;RGZ%RlDbFGYF{v7nvJr9#R+iv5@~1vi8=@)?_W6G?
z%!?N#^`5`2e#D03G!8wDNAlG)ZNh)Xg|nBl!^=&sn?L;eeZ;$$U$z+V0psza0gn^;
zj%2{AJ$>3o(^J~s^mc6fX#ZocVDdjF{>8w+ReW_8Ume8Pcg0tC@zq{@brWCZRUWy$
W1D<T#Xzb?h==83Ov#Y&>!T$lt@WQ|V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_avgpool_bwd.hpp.65CD0007480F503D.idx b/.cache/clangd/index/device_avgpool_bwd.hpp.65CD0007480F503D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5732d6b8eaae7c71439e2c1cd70237eeb4078cb9
GIT binary patch
literal 972
zcmWIYbaOkx%)sEB;#rZKT9U}Zz`!5~#Kk2=nd*!T4D*;67%JxM?eon(<RH-YIk;$p
z`P&7#`rRB}dQSE&-{g*+HHqAGGqTKlqtf5^v5Q=lo$^~-Tffho`P}!};=>^t!56~i
zE?xIAeJ|GR^w-wUy?rJ}wKtoz*XHeJDH0!)*=(zKBxD;q^dH`uey3Qv>XS)?RimrX
z{=kGe_e$cArN6!U;p2ybFEex}^;oH9-kd6(w)d9kN$zV~N>Xo6;FJzioA7$`)|Wff
zzDu8s>To!=`Si_6-Lf}10>4~u`N16T{rJ%GODYTt9<?p_6tN&&HsapI(-#-qtA6TO
zlzj7XSg_n`iAxK5uH-r@Wbe4<m>F#$<saV{o4MHV{=2T<r*8e9n0N8=fn#P3s~1cu
z5=xTxPc)j_n9O&fs(fFadGt~CiKQycyPvSXUH|1TLvdwpQamu+)CCW7aJ!Uc2r#j-
z1GzxJp{S{7&A<WS9AB|Nc(Pa2M>ZyAHU<tMb|G;v!NAD`3~L5P4#poAm!{=iY0(p4
zVu9%wWf2vJ>1QZPO)K6DR133*krQk_%-j$^A@vs%ZCE*ZIK*Yem0_mwa7f5XAn_z+
zC6RbivQkJqX<2C`o{X#v5>HlE7KtY(D~H6Bmz9U}Vz*o7#Hgt#GxBhV$cSjboRQkc
zs@fn_>;x1RQWVmKIl9}6x#eyDe+!VKHN^E{s)3HykU-)|YDgmSq%@?Ec+wiuNIV%0
z86=*phAa|KPD2ifC$Aw7=dCri3OG79VLH%tiZY7oun>m1l93A>Rh}wk-4k~?*F800
z{ld;K-pvH`qM(R0!qJL?iqbH>Fi~JEfnpJ66fhQZ5|w}v4hs+lm<)_&c%|gNSHJj_
jSaDHuu`(#_a6C+`R9~@eeKZpnCj%!)1Om8V3<d@Om9Sw<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_base.hpp.07667515082DA1A8.idx b/.cache/clangd/index/device_base.hpp.07667515082DA1A8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4bba0f4cd960cb41e6daaba2cdf95d318440aae0
GIT binary patch
literal 2394
zcmY*b4^&fS6n}d>8H~5>?c3NN7);nu$6$_016qqV9g`GGEKN>$Q2C<>iHs2Zm(<Zb
zQ6wHvDN!^k^N5_n(lQMd|DobB4<6AY8e(=1NAq7$TG{u$aT&dH_S^2g@80j;``zC?
z%S=y8%M7Nd$+_tj#d&4(8H%E0_~R`rU2GIk)K!|IDz>#`9-VRA!k)iSqE*^w_q8Xt
zo7YZ2rzT80Y%2`+lwVz}9lV=6a3#yoVb~a*^4*T!ww5oIY4>Z+?sMz9U!417dtv|a
zXDMSpow_&Da-=xGT<y_!J3=e=z0fdNce}Q6;!ux$S)*`a!gAg2&6}1-ah|vtS*?w8
z{tEij)pe?Q$v|B7Z|mN;<m|1z@Qto>%wShv>)i*NR8=2)4hB7ayP1jXEW8_bj%wPa
z+=*tXmpPcLW0%~@XbT;m^Oe|;-`zCh<;d`oUR!om&XSF%7iATnY0a+8UUy%*?Z$z%
z)-ucVRqTrT%MH6{ZfFm!yWAB$Q?D*a^5*OcYP{d@PvMcN(#%)ZUwn4VvUNjE!Qt1P
zOE;zGE=4Wn^5geT)?Cksdm6mcD*B1?Ru(U~hr1<<r}y5zW^{2hBc+g4Z8Z|0f^b*y
zqkK5{P`7K(IVDX?D5PbyQUYL}vNxnp@%uY-I64@cbx0RV0L<xz^6J_jCa<AsF2olP
zu_BhE@D1|Mw3WIeRp*_{I64TQ9WRSF5CHGl&Xa$<*IP7+qea;4kU0zlz&v+oqO0ZK
zv0GKN8lP=o4PpXd&OVS*^yWB&K9Cl$6p{!fastC=3y>i9vwI28y5@CqbSU<ZkVR++
zfPY@in3%Q5>k6a;u>W}BcsT*Tf4(~JlYIBk`y3sJ(=)5hF$BOrzyGawY|(vDN?IVM
z5XW#F0WhyQF|eZjan<)6EyUh-xm`~H%pY9*_TZvYaifY>VQ&Si5EB4%?U_g8y7F)9
zeZggnj6*ztdCQQi&QrO%8JC8Gn~*7t0GRizj-?NUoon)4YB&=vMLdAHWzzGno@mV3
z5=hgya*a@f2>2^+lLVbE>sQn(=>Qx)mWkyEfPdS(g6OTcOvmu)ICuPn5g0L74SY0x
zrG24<R${YKq*M|B^Vxx(9V<`QT2!<K2Y0h>F#$057#pPBkBXceJqDXCYD)|OFkhdt
zs-s`Hf4wg?BV*(c4`43M%lH0_mj;`B^U|-IANL3tWK-M1fCQ4~PP8^}K5&g!vo;B+
z;aeag+<?gdZeTS}3}{>A`mevZk$h2ugCvG0MuKzfFQ1n`=<rAcEOG`pB|ZjFRHCCp
zbfq#&fz@`_j$kZ+SU5J%Du@k?!D^l^kjboI_u=9^t|ov*W|>(9lgT2B%%XyhvB)a3
zsvsR87mma;3nBrj1T-NNoC{<M5fKsruURtEU}VV*j*?U|Jq#*~5Q8|F3@k+>vtLq(
z6&k;vPOF;$^|L766rTiT;~vG@ZO|hYIbs|MPzcZ&XU=mOGKa#1B0^t(=KNTIktOpx
zN>at@VaQozVoV$`3QN%@vtM$H-5T(Nw&Rxa+h9aJya=$G-vUr)!DH8gn@u?v@WtDV
zHY?l<6o=pDmjVQVeSoXumjMKU_TsmB{)4)HC@1?}skWP8kuE{!@HbQE)MfakR8wja
zw2wup_EZl@P#`{z-yV>~_8*7Jvr1Og;hdA@$x6t%>Cm3nLMHEcg4GtcB^?G1)(lRL
zUoUVY@QrVT-!1+d8T}IbKT2Uqsf_}@u_%IxkivaqDJmk)FB!xJH5>tVh%?}~6qEsj
zhpXea6LhU@crno;vXDT{BA3(^4xKreR$sC2r8`$>2APx0DRAcuvLsm?P&w={d;q_|
z;9M}{boj*uYS>v=J$fkboNO*RYDntDT}>1xj_jzvFSq6F&#0#Oxl!0g_*|M-G@suo
k-qMBMiTn=*555PLh4d$WXu2^aKp~e=a!~vJIt@$z0b-)GP5=M^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_contraction_multiple_d.hpp.C066C76CB17254D1.idx b/.cache/clangd/index/device_batched_contraction_multiple_d.hpp.C066C76CB17254D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..31dcb6393c81cffd0271fef0ff2372cc2b51ae2b
GIT binary patch
literal 1326
zcmWIYbaPW<Wngel@vO*AElFfyU|<jg;^LB`%%6-53>$%R6?67Zj?TVqAaeZg9sh;x
zeA6Df`Id>t9_;g#ICf9&mswljlg%zirnRUn*G|1(yJ_yj)0bKeB_x)7{yo3@oaK$(
zJ8NPD^tgV^i=J{d^vLS3{PMAOdg>;*FHduxjrx`4G<n)SwQr7%3q%(vESh%Da<%Ba
zFHROh_r5u$xSiS^*Xe%fx8uYeEnF>opR?I%Ubs@E8MXZA*9DI*$93P!7Cox|H*fFk
zlN)MGW?mLa37+ud#rK0e5fx2eA4?qg!D4OMlK9E^mvHa>DnS{G<_S^Tnv+ypuTC-b
zikd!cwUCnj(rS(`+wz+1z4dfnOsf@P-^u=$+thgnbK;V3yO?=5P2_&EY-v_E2N&D-
zf`a~awIaGTCiBnBPTl&p;7yO{uZ9f0tDh6v;uiQ8bEo$m@7v5IRsED-_>3M?@}2J4
zyADp@DtvA3gT+@QR~lb>DyYH-gt9+gIUOqWc-C`9@}|7gBI8gs%UKhC=sVkNQ$Mn8
zbE&MnxrfKaya^MGKizcQ&2ztklO=i2WYGt67qXiEX_&f)eQvPhY4*dsX`6yBUGU}H
z8NI(jsQBQiT?}WwRYb|>hxYDPm+7x@<~>vH^>=+%?qSQ<*5_6ypRP2lY7Y^6Eql+X
zEOn_4?~T<*g|anu*Y&=8Z5Q44?y%qEOA)+Vezl9RCtUk$WY2gl<Y6;Iab<2&9WW&*
zf4i(B*fnVe8xu1J1Gg%>ss@-~;ACQCXJBCDW_0)IG<=hNJxPFxl@q8Q2)L~@t*jY1
zAe`ibtB<{uebX<>#Ky_MtuCvs4klpI-+i3)A6P#s7hqz6>DFS=5{F3_rKS}tGa+2T
z$OX3DQ>CnX;x6YpnB9z=U^`)M4e=9Fe=*U9m4k;{PgzeBW(N<qzOp_N&p_D#iD#&6
zh{Q9}GcrNwGuAUk;+ZI$AgMG}Hih$Iw_E1KsHrG3@^EX)XllR=fQ11I7ufGGXYz3C
z80c8R6!CBy*&BHxc*gd|a9(O3t12+2oq*=58mZdB45&E#@JSWR#k0&nlVo+25PFT2
zjo_B^a2qQd!+G6a%q?&G|66eKaI1@|%fMU)3@9x<LxchPTKY&l11$q2o}rc@63@ug
z$OfU$*wh$_XQE|-q|#K&6wX^~Y!z^HZo+h+hpc3*)M4@X)PVI1JHL206EOS)HKh^e
zS_xW7!xX_pfsqbMDzLy{C`!#qR01YCxWI{)NWIP!--EEwgUb{bB^T?1k|!4=jRHv~
Z?%5|qBkyh~5o6}z<znE4OEB;tm;g3b@aq5o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_contraction_multiple_d_xdl_cshuffle.hpp.9FB9459677B68FA8.idx b/.cache/clangd/index/device_batched_contraction_multiple_d_xdl_cshuffle.hpp.9FB9459677B68FA8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..543307e8fa5074db8382e594f18aea9c04031483
GIT binary patch
literal 20364
zcmd6OXH->3_cwRW%sE#$UM{^iDI!%tI#Cc%#NMvGca4cHMx#-rD1x98Yypg-0gb)U
z>_!E_7z9P6iVzE;h^QbczJvFUGta00hi7FyFKg+(zcu^Jnc1^v?>&3=89A_j|EOVV
zYC*FGu3R*G`E;nJrly7eEnB{H{vD~B+HyNJwUsFs#;yLo!dv@(O^5Ea8JhD9zBBck
zUt+u(HI_funB;fGCwt<w5%M#^6}h2lmLpf4dHUm}eW77@3s+>$55J~8cf+5x{_8SZ
zJ{+%U59<5+DW^Yg_XeH5DL*%*O}*6B;ciI(Zw?RNY4rAdztp*Rvx1rpOJbsQ<L!dC
znv9#h`A0N!(u_&t{3c~x`)TTlK1WwPo1;^t{W<8$Z?Cj7)=fB(K6_}5@At+<Tl+t%
ze)7g4y^B_#xTL$6-LfV(+;Us&wrHoH#qD{9$7VFN+q}9o$8=U`SpADv8~Uf%>g^k0
z)qHIHj7g!BdJh_$9-Y71Z$WXy`ycvwjm|!|-E!iKG|ikM?`bbz6ny@}FtQ}&WOn+%
z8>2<X)Pl|3_CKE8f7|H&tEO&?-E3@9Tv5D!-OxUcug?`P@p+%NwWCBMyk+Xr4Ckzu
zlV^_~-@8COKQQ!~SMQt$wn;00i%8P=bxrymMWeU*U&ES#N%G%Ax0L&bh3Ec0cGCcz
zv&U;Pzu%rcoKIRdD0}gYA@BO$)pJ_^dQ{@v7m1ELq0Rc_!_HwH=R?+B4GFoJ{ko#|
z*OLL?w>36Diwuo<sk>sk#k+qD*W_HYE1PRDV9~J;uYaz%Xg762&V`ztPcmL44!P*x
zl$Gz*Ry^tbn!Rg$T;7jtz29v|`up+YyJ`*@ef5b)?d)@}Dn?c~7){(%)NSUY^X1JW
z`-WX#Gv)5gF;m|@zJ6oNP|poZ2hG{kbhETaa-w;EpQbJ0TD^>)&mV0!tj(jvdSBh@
zDYZivjyiO9&?Ue0f2Nw=Z$6r2dNXc_{f~}E|6Jf#H|fA6t#`V~i>;H|PQIyMvhQSG
z)tjdICmy}p6urdZ>Dka@o+d-D-<vad+VZ;_rf071d7;^A?!a&7Oxw}$x!an+kMoYX
z?K#=@A#d!1wq@h5MQ`fe-nA}j>9)3G0ln;2YNRLk${x0K(ZDHUVHfc2mWdx*WBVRG
zxOYds@bj*LO=r)$tQc?H#a(NAz|09&5r?|X)K3mPRBz%g9$omFbFNAJHvjFsMSt%;
zd1%k7&{f*26T`k6wnX~d`?h=O_fBgaJ-TRNnf{fzebVk_W^}hNzB7EmpFiB_TDzdZ
zxmYu(=l-c?t@X3|zj?NCvSE|glW&&lndNo6;#7B;JKplETG_szCpbtC$V{d=gog*&
zZ@YiHNiU!}M9*o8t<&MCRR;I2WQCY`6*=WE8gY1D#SFcGif%)bdwczJc+p>ywS#W_
zc*oqnD&&P_ZnRhRpQA44)Iau}<x=J|&Lwn}?}V_11FKTu*#$i^4UDBb^>#@02cJ0b
z?V}oSK5x<K+VRg5+<Kn=u1fvpyZq6X{gY?8A9#^w*K=XUGo5LTODZP#pV7>naA<MV
zM%k5no;T<0{o|X_>kmdJTzGiKrBphoaQ5e*@Vmav<Lt{rv^O59G<#dDHm%=95K}H4
z=Df^9{m{szw?@n!JT)!o<)3>pJKFcUdbDQ@9k<%@<K^ld(-)dd`8C})aN*_Gb%+1l
zdMZv9dDCS<<hEZ+M9cUF{^`7N8w2-m^A`=`>z4%d%6d31W|zrQC-2@SYki%13!i2M
z?E299Y<SPsDv7+Ue7<yJ$i#;^?qQjSHt#EuOXD)(o!3heJS#0dGS+U>vpF%~LSvuX
zQPuBV{;EFIx+$dS&E(yoalPeh#Ugo-bzb`6h0TY{dOV7sJFU6c&Qu&Y(D2<*$K3Q`
z!#$km|2m>q!EuMNb;atj6;p-|a(w3UEP3AZGAwBSA6!1%{s+@#F8lpf4AOa$c)Zwq
z&ZfOn%>O_+zwckLqF=b%PXj|;m+!whudd+6@73cDJN~n;|BmUS2hA!i=(T*A^Ybdh
ztn<0MWhP5n7OxyqIkCa&yZl!!Zh`g7-V91vvTptD80n*cjL^oIfTRGk_alOTNq7=z
zk=;0P_wbXIc24nk9CJg~{WY}qT)^kJq+QJ}?F%K((VNukmiwK<dCO1r9^JZZ*R%h)
zs9<D5X#Eeng2&wT3mjy(U`33X$K!^e+^6;Ru=jJFQ`fci3;Ud38|bor%3!TCv2urk
z=ao%+AImr9%ED!<xONRL*FJv6qg6Shh8E3^GEDh(@9Lf+T(iF43D=du=2gX^?-mza
z8G6O{FWGNJUDsa?ZoOEx+jZ-OknbNP?_Za;Hp6{iMB5dU1Kj@a^QL>%v?cv0e>A?i
zu4HhZsqH839<N#RhJSl$e~tWS`O?^V{pMaj)2D35hlQSMW7kJsTpkrQwO8HZBj*Fg
ztk_$>{o6SgeYecj=Dfcv-M9sPbUpX`kDtOXb?lVY>pe(xN<8^<o{r6^tdy$bQ>?v)
zXlHZYZR^xre?Oh|@S*=-r)&Z{JZJPQ)tYPdV&35*yXjVw;Dn6kw51z$gG?-DL^sw~
zov<0r89IrE2GiuGcFV5%{`JHxynQd4q3)^W*l0e>z{alPOj3DA{%^~DHP)Teu}>?u
z6*q6+cVh4@hrp$?O6P9BVCAwsYKX(4+CNgy)~DXixYZ<|_xH|O+D{jN{Vg-b-d=hj
zI&Ay0u4ZGrKM%Fuc+h9l59vOhAn`EoKhkbvn?n1$OA{Qf&NzAO>C~NXysqwa^GR&K
zpV{G&a3uQZ&!fk0nLau7P1%x-KMplYl)PBseXCE>@$mF36^ozFu6t|wVSdnu2lsZr
zf7=+_I&=HnzC(xk9?CMiH$5Y2RMyt05z(aqKG5BLmE+lYpQIDMsU7I8h~1dE_W6mf
z2cDd`|NO<?+P5oGRz`jJ@xrdE5~ySUe89_=J<Ge)&FZ>a>=Qe$sKweV^qcZ!-%syu
z;2qt&_4>?VITm?c;)bnQ^0Bf(&qT0^udH%6+q*2k<TO}%AS`0r>!kx~&AM7H{(JR|
zDV7%&r8HRX9(Qz1{Io24ZPNyaE7clhC(*R{Rr3qZUp}<fP{_+s>l=`<IHy6^b(&H5
z#%t>?8%<fW$-n4)xc{lq=POsa8Z3F>a9Z=dhivY%cgyxKEZ37pZw>Pc`(xXzlTU5W
zRhRxyBfS9(-M(*+oN{&7)8GAjEdMmK=Y~t>uhu+`>|$H1dGO(c$d_pYW}O>V5R&&X
zYJ*uz(C#(vCAWL^jhy}@;r*VLbt5Cx%+;2yS~TO8x|$jneamO``=uK-fL1>>#dA*m
zIcKOMYB=!e{Nv6`8fVriKFC05keZ?l8I&P=6(KBwQ^m(!dckQj;Eq{dA^9t0sv?9%
zs&%e=rnBc>ZO|RRD2|iIaXnOouxMTjJ2^V|mO>5;gVhvC&^-wTsK{5ASJNHuAD%o}
z2EM^BdW;MnqaG?kSmbJwHP@fIrD*_d{GujKy@@kY5yGNB-T2(1fFHlt26mX`4ky3E
z*{cX)G0eGs^Hq6Rhz#_{FKS2K+fk5;d}Zl~__XyOpGX-Pido(u`5QcGB_S-vCYnFp
z6;U4<7=~HiBKcd)q9k8gY7U3};Qt|60|+5%if~>Q&g-iPVKKFz`TYFEzFu-*faMYc
z9b=%6ihN~peVx=Y%I`K0EbxpvkVXg6RuRHtws5$O)1@&ZWS|!wOARuuL2fESSj_iC
zL{?VEJd%Oln57BHoA7Xzgs@oD^mRQp_)WA1(8Mo#$*I5Oj8%lNSY2+j@~b~|L>pLQ
zmLgtW#9ONfVd<7QYJ~pzqcb#s9J7=`^)l$GB7~)T*aL||p*}(mT(Dej^0J$}mWmJ-
z`$y^R6OyYdWncgvZUr)`Ku#(`SRCCiT@Rf8`+Q|D5xgt{_o5_(#cBBdp6NN>e`x@O
z$5IT{i=ml{5Ehreo#r1jin}BO1M!O<A@LFNQxU?_b9zOy(V|r~dcYL-QXw&}keI3n
zVR4IUavT-0ZIBF1$GzM^Mt4v*6(KC1o8}nvD>h}zz)Z|if^<so0w@V#@fmk@OyTLQ
z2o0ctU$hUZ?}J7vLRbPmuGmwS>g}Wf)G^CLsQwTdst92TOdHeg|1tYdIdH}=x{kVD
z$NHiqgeBN-7=Pts+z&D^1G5w(=VIigB7`NR`JPFQpJ}ZOcw&|YWZr-rRfMqg?{z2r
zfoxlb9&o~ADU+C$NxG{DVHx!8jBkA2uE>{xL0G-_K*K%IK}875#L-vNauPDH>w&4b
zml}z24emur2+NGm9p`sPFW2RPh*_FBjb?0Ll!UO%T7BxL@1{?gCj&$9aO+U-Iy6H?
z2+RDS=tpiv8)Vv`2kxavkQWIiD)N=({49qa{aZA&fh}gK;^bAFgNhKA?+R-_Wdtvu
zD+6KpMJ>p_1$n3lVOf&=a@G9>bH?cd5l^~>leTbnDneLRMwa)BJhv-c4)icf2j||w
z1*iyNS+#Y{__xRMp6LPe&ap_0tFYE731RuUX3^HN@lk){O@ez#g+{5+Nks@t#7p0<
zH?FEx$$$?Y?rS7}jkQim2+O8eL!)WE>aJ=4DSpvCsD2M`5+xxlaZYdD|30yJy$pEc
zUhX6P`^Zj32ung}sbg~W0B;%az$|S@-i9nxgs|*b$u+!=QD1?T0kc#h!%F0!B7|kv
z_p+6PPuG~rfdd}PJHhIm;I1Ns<=5|yN)7J2gvx*~?xh*YoAHd4gs`MO7Ck@gyJsu|
zUYO+}GJS~6p^^}meZli8j5KwQ%7Hb0(Om)F6%15_u%z!?Q@{MIPPjI(!Yp+Xc^y_v
zCHczoc<VQt=EslG09`Ok1E=1=8K?+h$tVo!GBo_`Vj1wmYm^M-$<RVY2+N@ZfpM<k
zwP)o(AG1V4_eks;m4vVy*;i+n@nd?8s5T`)O<^y!cl+|6S{De^X7!OMBuAnD#mrmn
zvo!#ZNAiqQe}+e*B&3g%{A+Xd%BD{;&=X7L88Ux{ceIiama_>jHlJD<Wv&NIa4+Q&
z<8rKXN<vu9FN)M3bx8J^2LfiPhZ^-zM@0zBjgZf$+KP%}WS}p8(Q{<^9J#6pVabmg
z@3LvukSjds8i+lBM7SyOP!Y8*oWxO04bSOTiN*54eZB9?zc`yiLYG5=sX|FeKX*PY
zTyk*FkSq<LgGW;V)hnQfiV&82Lqe(^H3VX#YK&P9K=lJqUq!yMtgQM=^Oqtu8F0k|
zZbkA|th-7=Sng*I=znPQsAbxq8}8+}0FMiD6(KBT-9EMYRQ)tu2K@0dra*ZLUTGyE
zEOiDwlGh0vF3Z4R{Gta){{h}dN<vuv;VL(H)z7on0J``^jhuQTUPvV&ERQYJ++RJM
zSSSP2a4$v3s0ce7CHcxy(R#pb`2t^UV2^t#0rC>iT}24X6OYy|31^cv^?`tUc?_kG
zp<G1>%hO}V`#w)ochduAxR*+aaV1s<CHcyd9hN!A@TY-tV1!vd3i6MFnTimWH=5S*
z7b7CpY5*B#DV3;~O0-pku)Oa#YoXThEsZkZhR0Hh^h>c^l!UN+RJ&)Jy=SSY4Qy~P
z$v~crXQU*AW$ElW%l7!HsR3sV?dEM4u1Q2iDRe1?BZ-A00Wqhp3WJFnii&b5mP2C_
zyNEajmX;xX5*`!Ei;7QB_6b@L;fad<NWLFgQTzvT`~&$@d<GewK|LrgKz;>iF!9mq
zs$l!;t~%#o=^t|J0u&Wnf$LTfMvRcC$Oo?ZAdHAXRD^S`;au2b)k2GkYh15u++Y%l
zlc=2&6<d&O3o@o7$U$NbG9*}C70Z(mkPe`_DuSmns^hHcIN$jyA1UI%+pdZ0OJZ_S
z5s6GAkq7CxYtNWp7i{lU;e>@Nu_%|=65}G`*vX<oVoSm*5v!C%rNowKn~1%fMU}*s
zI6zU+APH-b45y-aDY1DeabU66OUXABzmnLzV%uLyyk4>G>p}2(R<scyID*B|z%d&1
zqWx_EjvH8<2Ex+7aJm%fz$P6yu-Gddd_(a;U~`adKM1@Ivh6t_IEP)@To9bg;tRm>
z0_a8iy9gXFvbY?Cm4o3_6{>+vHE>|DS2g&C;)lTIA=~~Ccs*p>n?Z0hU{vch2!74t
zx4`i&=tcW`2OQtAIFSoW<c1SX5*0f*=N+6oi+y)+0Tl1zoOiM9yExxnZ2L(r_#_ub
z`A>1dr&xTN3qH-_OwJ{f>rMHyIF~FIXLByuEH3B7a!#9^97V-rPJGO16GiWO_TVLZ
zw77Fmc5u=TPOdyUtEpj)htf#M;20>4VQ~_aCQ+<uc(-5AlIn$0s2Jfn!i)HHs2Jrp
z%Acaqexn(r@KZ3TK@!*?=|?mZD$+n;8t6wf11icvU^(bVRt72(xxhrOACUw(giF*_
zM{g2XYpRak*jmIQSuC<7n^;sFL5@ezHw3Gz4&=mIkj;;+h4M5`t;IX8`#a8)#KIzu
z`lWlI0XY!hO^>9N$biIz@(nW&UTQBl?;<K%Bn~Z-Kr%;B(JIkvl{is+7&sgTfn?62
z;yBPd4xA{i0}gc{kjOw(JOO%7fD^^ZoI^4fNMs@^e&h6h<D4i?;~dgxJJAZDuF4|N
z89wLSpL2tVQz2Oefb|{8b|42TGi8c_$jtEmN$JOoEH@r1;$c8M3?w-KsMrB}?|>d;
znV@1DKWrO6oE&taB8m4&;(aN;#T(q>4aspDDh><QhXorFHyNsu7<jprsS7d!BTfIC
z0qK{oV+R@F6A(&Nk7P8k7W{%_zpy^zDiW`<K0}%3ApIGu@*LQ&Dw7<hs+5I@&*O#>
zyd@d2sJO*D-s0WKh)uknmY#{*Fdwh_dZ4!+*wR&w0D2KD&H;Kkz?QCZF3`(maWl|s
z2DWsSUjx0@EI!HUo#brkDxc!?PO<njr+1oSvLsk^%A5?@DP(3?;gy*fst!d$_aecC
zoH9j4n_$@{bSEd8PTZqS=t(e1*<f{vLb51iL)Ar@zahdUi5t8s2}qW}u1Yo%v)NTq
zrgO-H7P9p|o20jD@hZpiCb7IdSp|~k!E;b1dFUM8a^2r@UUbu@LTM^AB<D=*xskL2
z8It6T#k202T~<8oi{D_QW`i@1%0H{zrSlrY{WnM;ezATmZ$w^9vPW2|%G42gaksu_
za$b8EssHO+f&_S2Q<a%1G6GRii=?%vH%SNKz>YWH&xhTtxD_|T>SdnwyAmg@*u-6U
zss)m+1ritXAW>yJCuJgwXb2G!Ubk>03uj}*ov8OtG=MI+GQCCC&7t=970uTM$gsb8
zERjBz$cS9A4F%FYfWf6eTFT-CPMW}dZ6{*KLsO>3$k0ey3`;?o6C;xmanev)gG`8W
zii&-_`98iEk%IHAkA7z0=c^)AJmtGR<!#7{Ld7$F@-u!a#jp5ouh{lxesVM0-p+Sx
zXWKvUlRvQS5kj{Jf$2X|m>elE{l^P-@q#^l-xeWgix5I_ieQl<SW&!32-qX^BbY2R
zUYr;ti=hrqRGdQcQ>ZUp9A!$7EDi};vFMbEKq|Utyyr9Co4i3(eBeDlQ0x|3k9Wd&
zPpsn4c;9EdA7O-wX5P1%#UFUz4=j!pd?N*hKS59=2xIB<slu>SVK~7g6^RYYJ1Bbx
z4d_&qc}X%Ak4Dc2-`;#R9ILb)w9_M6M)H*S`8FtPW1m-ME6MZTwlmbGYfk+ELL8uv
zX^<$0@$L*u#-srSG1i@7$(VAWAjY{fEE$u?DTpy9X;188l{rr;(+nuhfCl6U?`z&%
zu#lhLtPK@;NIwr5l4BH9<Rkrj78fA<0^~s33z2>y+rC~Zu9wQ5s!p#^5g`>LSR5%e
zh?E-9_9&@A6pN#!2GK0uAT`*);*C<HjZzc(e2mm6hQ*ttMw?h1D>aH`ahy~qPO3+r
z-z?SH%;I>dL%h_9wr`P&Tj=;nmKbZPGDS=!`WQ(cBYhI^iHdaIJe_wU8$588-Y}<g
zJ9F{k7bA;e<U}&Xq9TQNOyRx1GOCiy*tRHB%S0>419)e(LRl;Pde9UVdy#xEbzj)_
ze91e1eQ>17fi>@>!T8k)uzvzn5Facmc0$jcus;clf^Y6<7~O8|i5V~RwwHNNV(0J~
zn76$~ad2Tk%(cMY{#e2LcJJGpggGR$ji29)WSfx*-5|;|HyIJhZyR7|@g=!UrC7|l
z7IQu%HWU>loNEb-%Q>BL&WyHKbFS5#4>?g_Gt0Twu=p{j^O!TE?Xl1{7P?XiY=*|0
zp$)}PVUMTKoeW%5yoBB_VLytSpm!6CQ;=~A4KYN;1!R2zIaB_d$oM9*A=?GV;K;NE
z*^yi~7L=gBS+FD~MpT>>dYly8Nm5KyWC`9`LO+VL1@CMYmkK>g1z(bM#lgSOw^A5L
zah1@wisAt!DSpWlLYC>`*a*m@fHAQwq9Pi|qgk8?<cTc)1;~G4aT1UxvG_cYp9jXo
zGGRdg`2`ji0C@q63xT|l#kYX`7K`5i`5RzNESso!3*>KE9Kp#WIHo<4lSi^RijzmN
z_$((s%NY~PC@L~Jc_xc*aPk{0zRAgNvbcbg7f?)2TX-|JbFy~Ml16qrq0UZdOZJqy
z>QaV;xkH9|8RUoT8iBXYacF%U+7S~xs!Pa)MYH7zxZ$JF<)ffUOwj0|+TEXS^G7x!
z&XK#<Nqp#<)l1y#S)2^qlYtLigB0MN!r~I(UIKimOz#5syDX05JmNS%Dx=Mu$7U86
zbLPdI4c&YtoOub0t2y&(&W3Kb8qU0i#SPH30d}LC<2f{a&f-RB+Q{M;(DVh1U&8J$
zA+x!fVD}~#zk;T(SpH^c+RWn9g6V0Y8{JH21k*DtJ}a1>WpSopn#tlUp?j9#OrOsd
zx@WUEM=;G{`Ev!+To#uJ)@4Eux;Y;R)(==*E?AedxI(b5U~#43QYrMN&sPa9RV=O+
ztgBi68o|1T;?cWoUkCNDFZ^;Q^jr@Ds3JuG&j=Rh0M8r{Kouz$c;>RW8F)5>0IEo@
zf#+)$pX5AGvU#RcoaZSPpXNMIQ>>W2ef*Jq6~0*Y6Qw$dQhlQO<8<<RO*>|2E5!*R
z;2H@A(GkZ3c|2fp7w3TMIWUNh_&ShZXLA=Vz_kSoq9gtU<e%8w#Tl;G8Eyzw{mY#E
zGR<9#`>?Cm<8zU{|7Qs%a+fQpt8Q?q$J_{XHUdj3nH@l92a7KOolC%yO6E4uxy|BM
zpwkL0sbn^AIvcpJcygA{x0CE_tnrzt9-7s|9>j->_*7<ileZzqG!b7QSsxNyNsa`&
zZb7$1uq4_#W36+S@}zn#yvVlD){q3n#4h3^v+~}VPE2{fOBk@BMbdU;N3ubp;&)#6
zcixS}P2`>$TjWhhwu!wdIELiMP#<D;)K$0FWFqseuj*QMm1W38ML+d^`XrbU6+yZ|
zmJ|i62QzIUx*<$kfA#)M+W_4GOj`$M*TK1wYiWFBg%;t^k>W^b7YW_SO(pij&>|W-
zQXC5nVxct+P~xFuJoKVC3ECw=H{xz^iVRxpfQ|%{q)BID;LE){5ic>)IFMD|#*+pd
z@^Bt6xnU%iffGo`bUpGU;`{b$Mvvsww-f*C!wHFj@PZ{GSt5-VMa3l~zl6-l8`V`+
z0A#@ydp9i^RkGXwYf>uem5KtKR&H9pdylmq@xgeq|EW7f+PFbk5kQn`<)5|nKfybH
zY2v7y(<$Z*S>L8p!s2R9r<yY)V(Sbvb!u1~3$<dQK3$imQ2!}3A#>;qBu!sJ<~pkh
znl`bx1$Avf+C(UwfuR0L!G!vnEWtEOu%b9yFwJIhsbE$rbR)a1GZ3_{6znOk60EBz
zCRae6%m1YSf-Zl8Ffu_HO(W7&VNj|tgkX}f!mIY4lfCDxh|oI^7ZwT7jYvRURWCtA
z9g#cIF}5;mEnYP{S6o4+`Xgl&xW8>kwvC=*M8!EIo@0}6%Hj&51z)co)m0@GbOsNR
z^dT}KwF;sljW<uDsTNXlfp8)L%057Ix<(sO$VN1V;zP*s5b~oqAKB+4ZxYO_s|q#9
zh~mPys25v4dk2ULXKiOAYC8I;_tBy#Ks|s#L-dB2(zY^*UYW#{9FIiBLy5&hi5)Ti
zo!I`N#FfRp9!k6^u9NuIu^Ik)iEllNlYwtCo8eCZz9}p&0lp<{hW{?`z02Y_&OeR|
zqF!<{=f9c7#hh(1=frLz+Y%O6bGFr-6T69QYgpXEb#LKZs7k%zy1!v@2N&4Eg_7;u
zc})};2}7x|jDfx}H0>|q%b36<7)p$xh|5Ey$xxG2j^Lv=l%_yUQeS}YL!os8{CdA2
zDxO2@=PYi7){QKF0j*!K_$73C3BO*{;)`79(!}Cd(E1h2-wdsrS$qM>FQBh^CHy>+
zHX%)ts21^!vh)?wBv;TPuFjD*BTZ7cBjWla>1(7(_EhI-Q>O(PQ={?*N#7t%a(#gh
zRY>|4X_D&+oLE58R-}1PRkeZhR7l!}G)V;#-r~G=0?j((t8`v_D~nGH)~5vrYNXBx
z)@N9JR<J(H;!MFhlf_wrOP1h4pU)OtvRRxXSm&_(xq@{ri^~LuGNBhWdJhDL2P`fZ
z9LiZ-Avjd9xKi+_6#VJ)Rf0zqi>n2PYL>r7aHyd;t|7Q|-khZ$|Mg5HSrBYAw<Fnh
z>MC(Y2U+h#Zp37%s|v-4iQ4jDm3LRQGcP)gu)du$F~Zx5pT1guZuPPM+NUQjy6}1S
zoyp(e{vS%D4<)+9G$s9|6|>R#$zj}3%*%`UuNjZ+wGNj9E*@yW4IP4EhhRfu|Lv{Q
zCP!$R-ot@s1n(Td`_N=m1V1Q(A4_ooZ&<(vq6NHX0pE`VrsN_N3o8Q2BG@a~ooL`r
zG?5BRSvg09wc}=JM#+Y?PB@)aFX+|_7E~|V1l=~lg4{R}io*+KVI6TrWMkr2Z$+}L
ztaHmnVlM04ly!FG)hVftkJeoM!Rvo^-bkGtwPDSuw>R;!R6>_ZIFr<ar7pO+=+f8-
z3*7JkT0DSb$i)wFw^&*mkZc3ZiQ!D?mpVZ*w>|x8{LM!U>Wb%<hivmucj8Ywu}>cI
zqc|Vg=Ckej$S0p|FF@V}$d~dLpy35*B*lfuwvc_k5cw3+=gD&7WmA?Zl4VQ#_F`CK
zv))oX$NRkTeZD&he$v*)ZMh}$3i+@1$$_8>vwm0GlmAcEBF8!$rzs00$rwdkgJ|=N
z^CRVn_;>_O+iB)9T~$DdA5>OP(g*KD@jf&n@e`gSlHNypBrGD%5$kL#Cu`-*s5cIW
zp5ZWrcwco@u_X!ch!@7<RF+thh9fJ_q2Y_t`s2Jq5->>u)?`wh*dYnHQG5%S+yd4g
zRTXLYQX4qj0&WyXaVAlmH9f0FaSl<O8^r~jNdad~<C6l;p@4H^u}=Z5Kp=4!z8>i;
z{5-M8@zU(vZJDBCH*d3>A4*P^M8sH>%DPV~%DYg!3ysLFAFhf+(o&>Hitdi0(EH7y
z^L?-h`z$g4EMX>^1<bS9i8TR>Cg4sda-6d`P78OAy88~lo;qzOR+kQB+JU;0`$l{d
zCFs7RmWWiMV#YWmi=$^MJoGPRsbmwLoK)tPwpmjf4=oRA<sm(?X>d6l(#mIX0kSTj
zn@Cg?BCSH&eyT1k+wI1Fcf40>p-U|sKyFA#Q5u$-vd)!CErIWwzz0(Moy2=2@m>U<
zeg4x~gV@)9{OecNnVzS1t-jM3if_|91f34yYuYJ$&cj>5>IJn}I31i;2WLY3WM{ys
z6$u#}1GQpUoCLL!D9(BAyn2Xf-<@A-Z$`Gzi@eU-n<1@|5mfUJgOP{n#a?Ia&5+|@
z1l9aHFtUzb^mW$W40!@ZP=!h6MkdpXzs}m5A-{1WsOBHw`X8W`T)8r7(T#ESL;mMB
z_?*9$R(;I7czhTe4u{Lgj>PvOP~6NOr&Exag7k^2#>pon-bDITYc~th&4M;D)aR#m
z8CHK|)M>1h2cXdb=tr{bME`LfO<7@0wNhEAOBisXhPMvq9Z3a_hzlzmig-8j1`(G8
zc^ww|(sb7)LGzMeKx5Z>!LptnxX<U`8-LWV<)43DtFLxn+so@r*zf<fkNPwC<t*{-
zm1qH1N1>ics234F#6f4z?JUkl&e_O?w&x({92Vyy=Uf(_L(b<|d=)ugrTEW-!@uj=
zUmEp)I>&48=8QSM^zvT3%J~wTd`Vxjn50e~_YjX{@iach<;h=)=E*AO72hc=Y<P3|
ze->xuJ^M90^{M9_T=x+v6(i|IGgL%L#V8g>OT}mwZ;*-`SiDgxZe($cRE%NqCaJiI
z#j#Q`mc?;WF^<KXrQ&85$4kX{7AHu>1d561u`gAAbTH6nMclUBptaa3O7u!4L8Mr{
zz&OwT{m*0EuuW{2n76Y|@sq^-6N^s)^HZ#oJPXXvvbX`5Hvq=_zX0YhSe(I`XRzM?
z56=7#79Zuzk5XK?^w5Y8b&Hn$&ut<xDL9Z)ev?6Dd+W*3D~-2H27iesdViMqQumPs
zdS?M&>MojqM-%X)?%_D+ah%@d-L_sG6VYds<`?fDn1@2B9nD99`7AC#;|kCu+FpnP
z3u*h`{&s6d4l~{HuUoaEu?q~_R-|?wJN%*l*Fvm2r9X^yw7jtnUl2X#OrCQ+NX=+x
z7TfG4=SZ{IB~V%d^_fr&N=u<Wje^UdPZ=CQPBWdy9%&`iC&!-7=tf!#^+|NonR@P)
zih7dNGvUM0j`CX>vS#?w7D?-nJGq{~4v9BEzz-zB<vrKNiqm7J{enM$XprbMNUZ3Z
zrU9KaU`5xo9O#q-E4rqMoK7NVMP2+oPV*jH8d1w>*3!}k;sLeQRp02)5tcy966nN=
z#<CPTQPDhrS`VNZD;mp6=tQR6DH_XK=tM;mjkKfDf238lH?lYuS;ew75(kmyK^m&!
z*aztzXOHQpk?v_07bC4=WJc;?a47>XkK#>9{3t4Z<#m7MEs2CX3vA3%_#PBr<TWqy
zCY1j^FW%=3iDxI-FMI$>Lb4?4JUWvr+B<2vp}OizCgMzp>&N?6`56<DFv&0C9@{us
z8+)XRhMv(dgdCmJRbMrclT+E7l#%mx#cuqcxu;t1zk^zj7)t*2{Y}2#!rPz&$~xGj
zd@LFmi^h}S0>@p*@i3c|KY<KSAbZ+gi0lj5B)#%0FtSMXhQB*)eXQw%LxaP-%VFM)
zs=*Q7<p{;(b1{4|sr*QctS0`728vbCh}HklS~!O4{{tj_fb__e{^^O{2$}|Q_>Cn(
zml8pT<Q+N>Z(ZIAIwXBWzCOc@Z=vihTOX+W293;tm;j`%`tpnp?LKF9pYx^(_(smA
zk@F*S!RJtD`k5vPIzO`NwI78N;XQH6IqGh9I1O9bL%hKu-kAu8FksdG&oG&MT7;#f
z{CtfLC5fMw#Lpl-;9NED*-Wc8@li_fj1<0Bb&|9neuMH`H_}mKSoC;tV&C2W-^B~B
zQ*ti3p0Kj{*Y`$?{8?k<reVGDi~ijLR}!LWs;j;tBx`_kM4Z8W&VuSy9cNI-S<s7|
zcBtJBnNQ(9LhX+%{tUH0vv@tyUeBgOBawC_i}$1M`;jYAyUxRyQ8;f!4`cgy-F>_%
zHR)-*ZW@dC^Sb+4e1O+Iz~XdXH=W{lKi^oZ|KW<}zb@g{566!fG%Sk8%A5#H5`ioA
z_KCnR5e%ca0GJd2SL*!>fL{R^MsWmZ62Z9=Q-E{ZoL>YtjN%)d$qmkxMhAJENgj*y
zIg@;f+fG<z8Aj)L;GJIrT}t3c5)F`C7CxZAhqCvqVU9*2(d=#MLF9OlHOzU)K94oT
z%Fj-T5&Q5V{gP|tjyL#?$D#B%v?2itnIaCQlwYAzEo|Uy8aO|a+Qfx&(DWm<W=m&(
zJ3SbuYRDG_YUIJ-BN~r%o7Rn$kShUF)S^ZPLlSIiB(b(^>C9yrooSzuB=@7HHje%r
zPk#=lKPS+i6Tkihip^aQ_xma1r*2&*O`bS%)VR@O#*dvae1zJVuPka~>CX}L=VbbG
NH2pb={v1Vr{y%?+DQExy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm.hpp.0D6DC4AA71BF6793.idx b/.cache/clangd/index/device_batched_gemm.hpp.0D6DC4AA71BF6793.idx
new file mode 100755
index 0000000000000000000000000000000000000000..742f35e896ad2ede665fb8be300be6779482bea7
GIT binary patch
literal 966
zcmWIYbaUIs%)sEB;#rZKT9U}Zz`!5~#Kk2=nLdmR47Y)D6?69X-_1T`AaeBg@*f-a
zzSP_GUZF`onaAwe;d|ab?{8f&F;~fSP5J+OmsR=|K6$oghZFDbRQvgv*hGJy883Nd
z&w+D4X0BiHMgH!&n^hYQPiBaJ_rc^qc;m^Q>VIMTr5;*sxG9{+wAl8c*@mishy~xg
zUw*i{;Az~Xyta2tx8vL2HFsQF7-(Yi=H6SGYkViZbo`mZ`#tK$me#G?xNH4htexw2
zkw;sz;LzR^29-w^S*R7~FBYHo_{OrQ(+|&C@@+?=|E@(cZ^b(281~<q*qb@~V8kU0
zqcSHuu65@3mO8O9NQ!dIe9T$CiOJb}i{bT{wR1`oRv0|_;glHj-*qw5%2q95mWhWG
z<(4dK@qO}LRDZ|e?lUv~du4MS%e-5lKedN_wxgEId11}=eb?n1R6hk+?e?~m_j$#<
zRqbSD_nsfkuC0eA_u4UTG|wn4uFOqp0EVqj=-1w<qUy^9nAkXgTp-}m*3`CUfN~5@
znF+@+iI}o6F|#vpiL;AKfC&apCSZIpFmf^G{k*n#->(<D1ejQ1`lVQ;#9{h>&px-|
z@u%Lm+)OM${pM`uQeXn1zbG}WxEyE}%pHuJVEbYA!e}Ngu)UrtW!)2ZIoI)Und_R{
z!3>YxZkZFKrlQQq!zC#rsR46RY9Fg=gHW*(P*_|^Tn}axOd4nt$fYphZZGDRxBdSu
zxOli^Wn`6Ms(85MWaN-|@-p&BJOvpAB%Y#-A`(wYMhS_hETfFXQ;|_Y;;G80BJtE@
z)Zo0e##RAG=O#=CdO=%8TOAfiPYqbVu=9&|GXp&=C@GC_tG1xFG)xgNjFilEU||HK
z8M(j_01HTlqSTy3P#%K?76U9i7>bLMi}gVXf(JxE04IzAOeU(C#!X5EGL_8SJe&;N
G2nhgFo^J2}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.2CCE15A52A6097D9.idx b/.cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.2CCE15A52A6097D9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a15625dc53d6809026cf4c28be8076ea2e70949c
GIT binary patch
literal 8428
zcmZvhc~n$)nuhDUOUb5y;D!ro>-Fx}dtDGzTo5b-OmHDaF=(Qq9+?nD0@2Kbplz7Z
zL`vhyqLyn*jKR)gaKL4nIBgav+eRb<5i}x-TEVmi)O6yM%qe*P7|!vYdhhT39-pF?
zyD%p)A)$Gw%@(&TF)wR*&Qi{1vxR#9t;<=v>Je+REf2EU@^)9JWzB7v5OUe|pU>>s
z=9iuR;egnwj?(v)?fY{c-by`m;*%ov(TM0_zx`77*>zjNod+vhe@OmuWk+$Y>-CE<
zL2uB>Nl~-KkQa6xiT~Fp*QSm>+5gGADWz`rl$6izx87U!%+J3UP1@6z^`85kULE11
z&TshlBf^U3$`W5$Z$>})$0+U0#=Xn#e|TqA?s(UpUWvuaBQx)>`MvdS?rX2Fdw<Vj
zXUoW|*Ykd$F)x${<PX@?d)o0l&*fWVmTvwZZt%&R`JOiy^sm}=ON%(O@{82;MXMq<
zAD{5#@s0%J)V&XT1lG+A`e~E-;Lw#m^$z8WSvQxQV^SOkI^0K^`aDiiPiW5U$jN=$
z^jFXHZr%Q6%=D={YU*PaY(JF6)o&TUGIQdIjU7LIwZU<&K5j!qNvi+A6F*FMJHMN`
zWpGi+&|~X&juflnU;1)I_~jvo-{O>=33;Kd>EFLux_`t}&*B?Dx(79eoi9-ig$;Y0
zkZ?0DI5j6c<#5MxzfbGSlTQzF6{~C5%9x6e(!=(B{_7k4vqI%-9}Nlq_V}&A$2QeX
zUzJ!gBK6^c#2erDKAPjVGw<th5xi8luzyOAj_d%l{NzWkd|Vbj;P8mhH)pPX?#S3T
z^kWA%<(4lLfBADSTDw2x_`UvrJN~sZZN{vyo=xnhKkxc+-m|IsZ!J7}EBpKX`*Zf*
z4jy$rWbmPn>sNo%_W7yP^Dg~W|76jk1%WTm%^B<&5;%P3g0cB;5BXqdz}*5N<l-fF
z<x&0m>*HN_-&#=n$Cy9OElpLI_NmEAOsxOwyK58X{MXrmUpN<j-Fo@k@!aZ#O+RY+
z3uhb;G;Uk(uT=S(-(xQ&o%n3#u<x3Sk`FD)t21Amy3KVV^Ll&q{!F<xcZKuaoVq1T
zcc=Hh^!b*}+mmk39-f(V_<Ad`=l8umGJn{PdC_%w)8B1Ba>_rpXTgq_X1{u_@5a1I
z6Z-$narp6`Kb8OXm#ceE|I6NAJ|CI)P0gNHW7mz#?HFA%e#SKa=J+f3Ht$=u|J#6L
zN4J*i_g8+f=Yi|W6ZL%hwOL73XHF{>&j~yBA3F1gfJgV5zbPx(_59MBx{LLmAA*-H
z9y9y5?@sUN^Uv4`5rgh+{_@nN6Fb*zG^5|^pRB#Mc+{$gj`HxN-G84S)RdTgbBFx<
zWwYm?ffK*_`-Tm(>Nd<?cm2x;rP8e2>zD72NxIy$o!#->#g%7o9ZbFX%Gldw+Y;{f
zOyqXW%}>~U_8;-?8_w85=fdXmJw2liX8lvoYy9JY)jK_-S6|DxI=^P$Yo7)y5f93R
z4Hs^lU;mH~FMBU^VBNXMpUXOWD33hzpTvGJ<Vx|)TNjqxKkM51`bN(iYW}d1zhz_#
zOEz5_wk^3|LsXsq@!7$9Yqo?5t<~{Mtq%^~jZFUf-^y(3^0QtXPd1x9^5olb`=f>i
zGV>SMGQ#X({+%1|cSPq;z~g<5g-PS)1~T)!%LD8KLOVCU<<R3r5$j8De-+5g@h*qh
zL;O28zU85>v`o4E%&eY)%xv#+ggqjxbK_ecHt4)sey*xI$o8Cfc|1RUS?9*LJmuG)
zJtZyq#{-!(?{YtTzaE_%-}3x%{_NpfvHpR~^WNnl_8~8JZhXtPmkur0#(KsCGO6C>
zaC^9a=f<~u_uMyAE?l}gDv)``yFAc7FsyUqTmCI1?XxQ{pKA$Z=6aVy?V<ji8{hKU
z<tx^eMRfo6CI@u>*0xpty6@Pvi9>@jgg7C=C+JP-LI$KvVKt;&As<o!37e=pi*X`w
z#B>ojVx|ZjF;@hRSU@6jq&NvUQn~~jDN_QDlq&&8Dj*3ta-0ktIb8;hoGAlG&Xs{9
z7m$n`B~Ag3lCA(p$y9)&<SM{X3P?eY8m9tBO;>@VW~#tZb5-D|1*9TJi_?IkrE9>^
zGBx06xf*b^0@9G9$LYY)({<qJnL2RvTpc)i0qMvw;tb#z=>~9&OanMZt^pjQfDGiA
zaVBugbQ3scrU@J~*94APKqhjmI14ydx&<67(*ll_YXQe9APYHBj;N%rk9|h;oap)8
zDaH{4YZK>8gtbYINnLBc@1&6do!>?9p=%}K0epCbb`lDh?hljTWQ0AC#a0rrFSe72
zebLE?*q2&K!oJi_680r0BVk`|B^mp2JIUCWos5iqrIi%yEA6CUUvV-D_SIHWv9Gq1
zihb3|sMyz9NyEO@P8#+#C!=9sZzUc3dOPXZ*PV=xeWR5O>>KT5VBc^u2KLQXGO=&A
zlZk!P$(Y!;TFJt`)lL@nEhl4PKh6<1xBKli!!ZL=f+L|T`Ho_V0i8!LctaJDuod4>
z#U$*+TuQ>Hn9E2g$J{_dBj#ojZbBA|NJPF^Od|5dQWB9bmXU~jv4KS7i_Ih=Un(LA
z`BE`S$d^h<LcUZ+67r=6l8`SolZ1S^h-Bo;#UvwNE+rZHav90Umm5e%zT8YQ@|7Y|
zkgpVzf_$Zv6yz&qq#$2uAO-nKGbzYdi%3PjT1+bP)lyQCua=RDe6@j8<g3l3B3~;a
z4f$FzX~@?~NkhI?MjG<92GWqPHIs&Xy@+(=>&2uaUoRyc`Fa`Y$k!W4N50-nI`WMo
zGLUZ+lYxArlnmq>Wn>`VXdnamMl%`6H;c$bzFABr^375*k#ClfiF~txOyrx*WFp@x
zA`AIeF<HpBO36aLRYn%_tp>7?Z#9#J{K<~VT}P#$j95o3q$zWz!0~IUV=Amon==j8
zraPv0tqnSHu-=(}J1Ee5(&#5l>OOXfej+5vPl6=-$&eI31(NEgLel&+NV=a6$?!8E
znSLfD%g^dczLWZ%0i7rF@NHSh2;1@STF8i)r9wu+EEh5|W~Go(Fsp@(ididUG|YM-
zqhmG-83VIf$e5U|LdL=z?}+bu6?}&Y3h2B9@jWRsc+7+?ZR;Z`C?jfU)J$K?cV+T&
zK<5>SukC%!+^lec?S9pSCRVtLxs4U>VXokWQylcGIl+y&krSFQ*YLt69{Mf3a1(O}
zFFeHTu?s)hyZfS>712*@Vny^5+gK6(#0pMCKe3t<(NApTMD!DDcoF@?7G6X@v4a=U
zPxRPD^po7Ignm*JE1{p%#!BcXRd5peN!6T$eo`YRp`TR4OXw%H@Dln-9lV5olE*Hg
zpX_F3^pl%d8U5roRz^R$f|JotuI6O)lN&i1{p1>6MnAcQm(fq|;AQlaJ$4!W6gR7&
zpVGuC=%=)?3i>G(oPvHzHK(AT(#R?3r_}HY`YA2Ef__Q|ub`jeu`B4Ox>*(d)FxI%
zKedfj(NC@5RP<A;ITiiXMovXPwT4&GPi^5<^iw-{75!9?T}405&1&eUHL)7{X>F{A
zep&^mp`TXGY3QdlavJ(+HN1v?S_`kCpVq-^=%;z?8v5yOR!2X*iPh0hZ)0`z(<?X~
z{q$;1M?bxh)6q|_;dS)WTX-G)^bTG}Kiy;3(a&(R2KpIItbu+;8*8ASQNbDLXH;_r
z`WcO!fqq5}Z=j#i!W-ykbnpiH86LZVex{o>(a&sRP4qL{SQGus3eH47vzjx}&urvO
z^fPOC6aCB<-b6pMgE!I7^w>@Gv)rtOepVA}p`X>pTIgq0a2EPm)trTXRwHMjpH;(K
z=x4R?7W!EoyoG+2$8L4|P1|zU_UJ)QJ#iUHeo4!GN5fg~pMA$&c(><h@9{59Sk!&A
z5weB#kV;7S1TKDrS`z9Zi)kXvirFH}iX|lCtXNAT&Pr(#%u3l3%t|FB;jC0k63)tL
zGR(@^GR(>)B;%}HOES(XX$s6L*$T`mC8Xf2QcDWXs%a|Bs@W>cswJf2tXfMd&T45I
z%xc*h%xWd1;jC6m8qVrzI?U?XI?U=Nq~ok!OFGUPX$H(1*#^uSC1l{NQA-BSnrSA?
zn%O4Ink8i7tXWGY&RS^}%v#wN%vvR6;jC3l7S2W|N2mJUszDhOlP5x&k~{^{wB%`B
zDJ9DJa$jxhcJG<rfS3XCzI#sdX7?U-t9TOs^n}nqD%i*My~~+^&U?_^CxrMpS+zFd
zt0!B&sQ1oB#6(Q<okwrE7e6Ta!qy1y?vRj>zP_6g`;KN-Ui!=Wi{9M>V+Kz4-LL-m
zPGSD@*LJ_{-5u#4*~|AIAIq!YpUpau;N9J?SHHf#6Vx+f=Pcvn58UzY4jUP!`EK@C
z&Wml$h^#yB-JQrK4);Ba8>9ILJ%aoFhj({WpQuRRZr`1B?+O3Y+vwd-PDl}+^$7%J
zh$$i@DMf-Lr^t|$6a|u+qC(PAG)Q`i4#`L{Aekv9BrC;&G{G^U>yqk@-?4X?m_I*#
zCBsLCkM=zq-!Ihrte)QehiBztgkyLa<6=b2l8ccr%PvO7thg8jv+810%$keQFzYTx
z$85M51GDL3Ow5*xu`thc%<Q_1@%`CictGc$EAUB?Z~(q+g-R0s8*>#27a)r+60t8<
zl8Ak=ibU*7E|Rb>Rg#2#sfr}*%Px|!FISR`eYuKc>?<x(u&-2-f_<fm6zr=mQn9aA
zl8Sw`id5`tF4C~CRg#8%t%@}4>n_r<uUC?eeZ7iw>>DmJuy0h7fqkQj4D6dOGO=$~
zl8Jq@icIWVF0!z1Rg#5$tBS1dzHRODHA|oVEo<G{m)6aF`e4!1&Hw$R`ySYAi@P5#
z?0U#%*apYl58fFWY+tf)@uKt>Y%g@J*_L!ad?{<$nw8#|v?7J&czX{&|G=Q0A;G;8
zL*wIw(1670z59iQ_a87Jdg7$Xu~Vk%eUz9#j29#SzrB424jMdk*zge}M@h13j5f!N
YiyAUELbGP{={s}QQYPWqX%0&MUvi(4TmS$7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.09BE891BDB3DB7E7.idx b/.cache/clangd/index/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.09BE891BDB3DB7E7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..264f59322027da78d6adfbc63be19c673e9f898d
GIT binary patch
literal 8972
zcmZvh3sh9~zQ_6P$B2Lmf-iiarQ7}9+r51tDxwGu%4ojG6hlizU3w!$g0y=h^cXiS
z8O@cYj?W{R;pt&`E%6@m9(#~NiqnP6KukyFglK4V$~$uI!1u0oS&RRg@BV$i-}Qmn
zvmdi(rX(kGj|T-ME>6i^wIq8H7Zel};s0Hey?S{o8x*u8EGQ^<*EeaaX4S@qU-JCN
z^xa#6v(n$`6QAHL*{N*Zm;K<zjDyGDFI0aWJbKXYzm$G>H7Nhqy=A`dQ|p)AFUs+}
zR23KYGMyMZYML1S^p3+x|N5{l!93CX;hV0KbLYmpKD_H|T0H&EABAIg|Get0b8mFL
zA2sysx_>_`EPb*x<@vS4Mn7yGs+FzZv-s{ix0dIO^z81MQnVx{^X|$&e7AF6d}+<j
z-3#154f*kE?)PLreKI7k&xY<3j^#F9x?wNc_}^Ur6WMbbUzyvxV#f_Fx@_6UGt%cT
zkKTAJ_F?O`WaG1@cRGc7Cx`ug!?1e?FZZZ%Dj!d|zVHm=a_+x>?(mf!tuFPr=FW;4
z*W+jXg|hCxt)IkAOxSjzCT{N5gR8ijO(U0Ojyb;m{@*`e=R8xBxGuVQM(4iAzn^%{
z{q^Kc{R@i+9$mYAh***I>?cd3E)6*JDyM8u&W-S;fA>nszQGBN3$E3l>-R(C*<$5j
z<e=8%<m-vuW@JaX4&7f8{6Wph)KmRDMe0hnJnqwX(<AqO^xMn5S4GHm?+)np)v+7>
zk8bczT%J-qc*cYMDc8R1ek411d+rw_qIt<XueYny{j88-Cr`Zl{ClNQeGUzdcxCd6
zCl3#QSwDJUL(a*0;xB*cN~`y|jy3hJJobe<ZPJv;E?3wO?(C?a{ltvCSLYqMk@ela
zec5|%b{l#&y#K-XYF2#t^GBbZntkza`up?e&kcQUR(Ai!0ilmgo;y76^#N}W47r^z
zgjZcWSAIml`qD_x?N{ek|0(Xzvr1;Di+WsGm6B5Pw>MWO&-~BReLr^3`@(nWtC8G_
zc~|PSym^z3g&H@loy%1I!XNP$ryl=s@}RF93R4fx&-D&_CSi-`eCE}b(fcyx>YSzS
zH?zG97wt;#e(|GC8@EoqIqk8`>_b<5#Gc*z`jEUq+h&jU=1zRG<?v^n<GbW<dv@9j
zXL_#B9UI%b(s`(L_n%Mx{?{M(ocfnNzkD<#_sa{rUx;5bBImw&VdSI<og0!aH*MUz
zc;8ncM~`eiso!1p_U?P0%MaDF>2*`4R+ODmK7CTyw(nrs;~~E`HGElGyyK}wzNsaz
zebxAVx5W$WX}^DcYFm$g#K%VWYufn9XBUreU%7tR=(l>OYA-Gry8MCjWYpAM-_8lU
zl9F|OoBYS6VO{$59rO9O>()*2uA8>z>L>R~q$xR9FWt7MUb?cC-S&0WveP#X%((vi
z@SCMul5clO;dac*OWt+5IqBRrcYJ|+Uc=cgjYAKt`iGwT!=L)B*xqQasLS|q&V{`%
ze$Y*czIRescmCShwGa5H(zhb|de6k%DZSrG`L%J*!}zxcTrRqP<NU(Ar#+irTHpAx
znm1_3?-^Ob!VPtUwxsr|9p%;EJKcZJg-wxy@0+AW*4qbe$E1GqZ>4MURy{M4f`S||
zCtgq7H)>!gGiPp4Mx-OMbKBxSMzsBewC??(VCsljq0DUmxR0YxMB5S=M;t4RUR!eW
z^H64{e;n=z@7%To#si=KY5dLUQ@VsQ)BNLTM|5P{5*UyF?M`FyPkF~enKb{nm!nsw
zwk0s$`3pCE*SDb?!h)Xkk4N#N7Pl>daoyq9uFr~YtPW+K@{b2N20YWY1jaWP4Lqq0
zZ?r?18UAsUBdT-T5*Xh;^X2&S7k?ZY%1rl<`#SnYwk?72@8M}5UViS(Poc~#|2V=C
z(Yb92j8`vNy2jhD<G(i|r0u`9rTn+ON9)E649gG_h2#LBU!@BfpmKyfQ28Wmpbjl2
zijWl3MM#P{A|%Cp5-}+yN|2P&B}hs+5+tR3k}xSJ%8-=PWk||7G9=}Ek};_yDv(st
z6-X*M3M7?$QZT6|s*qIERY<BiDkRl>QZcC|YLL{@HAreX8YH!R(lDtf>X6jabx7(t
zIwbXc(lKcy8jv*74M-X}1|*GqGB9Z-nvgWpO-PzKCM3;#GBIf-T9CBTEl65979_2F
zvM^~U+K{x<ZAjWVHYDwQvN0L!jGfy4xyy)~88@dxjh#6b)Hvt3c4y!yA%?Ua9pGc&
zBjFzG+(HWp`Ao;oF1Q(CH)zpEBKpM^645WZ84>-Gk0kU<EhM2|ax)V8Wgp4tms?0i
zzwBmY^eaA6(66+Rf_}x#DCk#xq@rJKAr<|qn^DoP`A9>*)<PQkH8-Q7U-yxYe!Ycs
z^y_X$N5A1C1N}w|8R$3MjDddBM<)8s7BbOqx)~GwmX9p-TP<Xv-*Ph+`fVTC=(k(Q
zM!)T5Z1hiZPMX!R1133>KuvZ|ZdZZx(DabD^AGHRg(Pgo9k7Um?Wjvg_yBb&2`5q4
zlJEoS1`@7=77IzlezAx|>=#Q&#D1}qMC=!9NyL7!fkf<=3Q59#sfZ-(mr6*&eyNls
z?3Zdu!hWfNB<z<9NydJ;h-B=SOGw6kxs+t=mupGJez}2U>{kj&!G5KP6zo?@NWp%k
zloae&YDvL<rGXUeR|`qSezk~H>{m-j#eTJvRP0x4NyUD(fmH0*3Q5C$t%x-2*Gfpk
zeyx-=?AK~Z!+x!SH0;+4NymP@h;;1NOGw9ly_9t9*K0|~e!YQo>^BO@z<#5M4D2^b
z$iRN1lnm@QYRSNUqk#<UHw($cezS;7>^Dou#D24sOzbyn$;5uMflTbT3dzELtB5S@
zw@S#ueyfx$?6+#k!hWlPEbO-n$;N)Wh-~b)OUTB4yOeC~w`<A9e!GEe?2mWGx1S}$
zGR8Z{gG!j00O!*Q&IvG^ICCP*5}k?dvwp`9)VT9*hK2brD}#lx9p_RpSOg^nOQ7Un
z8I%&NfKr21P+G7CN)OgS8Nmi9GuQ-W1zVu(V7pxfZUVZ7v|aqevAKW|w&KaZfDutk
z1&oARE?{KTN&%ywRtp#vwN}7rsPzIyM{N`^25Pf_F;QCujD^}RU~JUM&gAy@XjnHp
zcGJ&mV<;>mc3|w}z&R^$hm#l5c4q^Bdlf5u%XYkZLOm;7McvE_cTtyd!f6itm7Gw8
zx{ed-QG0peJ0AQ#Ubu<6l^1?P-Q*DNI6C}d6)Pg2SkH>cCpNPp@`+`fh<sutCnBF%
z$BD=%dU+A~L?16ApV-Qa$R{>AMC6mISPA*0dR9U{shO3KPb%Xi<dZ5n3HhWtPC`D(
z%S*^7`FIKWq*h)+KB>teA)j2u%E%|zvoi9@&8&=kav3KhpIphw$S2ovGV;k@UPeCI
z$IHkkxAHRb$xRL!`IIVFK|ZCPRgh0<W)<X9$~Xo2luAxPKBbORkWcaQ3i2sFUO_&k
zl~<5YX>usYr&h5l@~QQ#ihOD_t0JFT#;M4sR&pxxsdb!+e5#jMkx%vUD)OnVyo!8k
zlS4&5t%}u<PpfA&<kOm24f(V(PD4JelGBh+tK&4})4aTfe43BfkWXvnHRRKp92)ZJ
zRjiJDdOfQnpWe*s$fuWaI`Zk2oQ`~Y9j7Co?&Wpl(|x>-e0nRdBcI;n(2>umVh!Xo
z>RAK%jAqtAKBJ5?kk6>(4CFKFI0N|%FK-~9;o}YDGg^5A`HUuqfqZ5aYa*Xn&zi_*
zHnS%3nPr@bd}bwQBA;2unaF2)c@z0eA8#U`*~**9XEr%Z<g=<+3;C>i)<Qn3nYED5
zD&s8Vvnn|Y`K&t5LO#pOTgYemcnkTgR^CEBtI1&@pIyb;$Y<BHHuBlctc`qj8D}G(
zUCG(VXV-By^4VVAMn2oe+sJ3P@;36>O%A(*Z^EYALBHP1t|2ZXH8^!~;A}l*=jk`j
z72NJJ+<$46Cd}_R8wy#%T2RF#ybt$)LNy6Bpv5#1vSOA9S+STz%!<_{Vpd9%AS-1_
zkd=x_!mLzH5@zKz8M1Pg3|YCDWX#IdBx6=dQy{BkDUel)Nx`gAO$uh!G!?RHmI_(5
zm{iQF)udupOVc2$WoeMrib=z)R!tga^)wx_dX^4Zy_j^&>eZxU)<`oTYh)RaHHyi=
ztWiw{X3aDcvSyYES+kf-%$n6?V%AEtAZuk=khO}*!mL$I7G~`<8?ttm4OzRGY|Pr#
zWMeiiHEu>=HxJ7gn>rR$LTUo2iK!FYm1~syxt`kQt^TXpK5>1L0$=mZp4NTnjiRyq
zql?|nvE2f_%U36V{&3UBHU7=fanTb3*ZrHClKPE4zd72!IXpbPXW*-vK#D@zzSR$0
z8t*-lS$^@aYpeY3zHxoW1-_<#?^Z$HQ+2yu_iv8r9Md)MVLY1qDgVT({mK5#y}I`5
z8Mxdo8$NR}pS1s$e{<xJNG<S;qkML}FC)f#*1vfQH|4PaL7<EYX?uVWAPM_F$A)Vr
zf3H)wUa$GHu{~mA0@=V*hQ1+fPZv6}f=hTJQ0zaPWC|-fl;{#+CT5B-lUx$aq)Z8B
zvP*`UoGHUhaVaoUG8LGqE)`~KrV2C7rNK<g)L^E&beQRxI?N200W%}hfSKtsVP<Cf
zA7=gUvzo9B%VojL%Cul+yKI=*nKsNuJ4d(Q0*7UcagJ$Mw}Owp!NmRL(M}pQB+3lD
zBY{1~f7|)!=?=Uz9!5Bdx1AnFL@jw33AOBDWYmg>QBbQMMn$c87!9@VVRY1nhcQr_
z9>zp%c^C_|?O|-xPN%c|wlnaUYDh@i<0<?jN!SnPB%z#y|3Y0s!g<i5heY&?<s_nC
ztRNBnl7}SpOXVb?U#cJp{j!H-^vmTWqhGEd8U2cf6!a_Qq@Z7^AO-!Zhg9^d<)ort
ztsoWsnuj#>YvrV&U#lPu{kn&A^y}rMqhGHe9sP!f4D=i2WT4-uAOroThfMUF<z%AY
ztRNHpmWM3#TjgY-->M)B{kDf}^xNfRqu;I|yTc!}ddbQ~kDhg|S^eyqS&v@Kf3*Db
zCq3{YC}=^)%X#fD*^GC4Ty#8`u(!ssaNdIX=}!ke-98Ii*zxk&Rf|_H^UJAAT`b2t
zIt6zQ4eJu#t!qj|QlbzMk}|P-ugIw0ePTzC89Oe1e1hIXiF<sc81w&a?b)|q|AB)Z
j8$4vFB&&uw%pNgn!0>3zn$)A`<SC1o<R>OLDfNE<j_|R~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp.FE05EE1022E21C41.idx b/.cache/clangd/index/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp.FE05EE1022E21C41.idx
new file mode 100755
index 0000000000000000000000000000000000000000..622ee5e30fd21bcccc55b23f45e936fb4ac01857
GIT binary patch
literal 10890
zcmZvidz2G}-G=u!6PB=C)>T;qR*+p070qojlbM*y%p@}zghfGaD;iiZDii{EsoB=_
zh+ru}BtpHQ#Hz(w6buF>R%;AIk9QL^=%=8mpHYj_2I~c_Qopc0f%lJ(a~$@`v+q70
zhRHIUOrNCdx1QhB6u4?qY-!|%UQ<(3Q(OJN>u<Pj(U>}H8`9Ji+w$hD1JMKGDf^3u
zFH^XwnfKq7$80P|R|#wPZGELnxwC7nO<Uae<R`y5Htfc=r#<o#^VXl{<nMk_s=l%L
z@!z*?Y+JBSoog9%)5Goy*R5$C=;<13&F%c*+xLHXPr7K{luPtmj~3ru_vy8T1${S&
zr*7%kU@FeMB3v22=i8Tm_~Ga!-yb>l%eU^XMB5K9>FaZAwD1qsdzSq0?cwN(um6!q
zUHMS;n*8VM)(v}R>s>om&U#?-Klk4!{F1$6Md90bCth;fj?rKCJ+X2}eq_nH;o`}U
z-Cmlucg*KU{o7{#^O*zxbRO)8R3G@o`q>AsTs7!>lm6JP*FLc7yG4h#+;}M0oE>_X
zBXYUNwC2_4tlJ;&b8o%9<;7K&Bk#}I_{qha<>h0~NN+#%;{Sg5Tx)Go=euJ@O)}Vx
z_dghH?OAosA3AnFW194nK44ngpaC16o!#}&inC(!oLfr=zq;k~n_EBlopRRtT~80)
z@WtOqYM#kHaAc3xw*6~%^+yl6o*tE6+j`?N$NuYLO2&ET@`Z2i-o0<r)Y`CpuQhG^
zOX23jtNr<&%x6<qY%@K~+;i|Q`y+pxFk@oPeq`&V500Po<g6<udc?i^UYcLny)qmv
zaYH6`1|Rs{(oK8i8H+Axergkc#cLn#I=c057q>D4#~d8RpSO1F>2EJ*PkEU8`=e*`
ze;T;$k=v&}dgd1szqI{at=?;zJpJ+A)1UgWGwPqb*g0{b>$gju;a+|2j_+>i3pc&G
zaM~?h`@TN4`K(`^ecx-FN58M#{JYO~ok?Hl3tzfn@6aQ|d)AaMJkNN3-g#F)^RFW}
znYwna-oE(hZ!GioJ=?zQ)lUM;AG_hN+cqYyEp7fT_xe3c-W&GJpEr)5ap$u~atnXo
zcEL{z`ug^TKaX{9`{J5&7EgL!zi#PEEf1~qzWSFry*GABI~EV0+Ui)qy}i0+$AcHY
zv1R?RY1<B-xBjC?#x1{Y%9EiflOBC})2cyt%y^@+=(Pj$M)tk(-Tl`TmV7m}RQ~#1
z`vLjAHB&~86`PZHcW;$mp1<;{@n!z?;YWuY-QUJNS@Zn*w(j$LcbIn0i_ZUc92Z)`
zM_Q+Idv3mM{F>7fhhE+F^ZaP%edqjf&ST}CX`B2LLYvt8CI%+%H{Y2`RfJny=j~cI
zF?Z_#Ve`*7ox0%z`}$>S--2BmzP|hNUk&=?4<Ejy2X^P5IJUa^i6>sTV1>H%&<igf
zT0UDE{rdUOy?N;?`s$t8%e37+|FdTH(kaIb%W03lzhuk7_qTJ>{X-92z3Xsy&*}F*
zI`WZU=N{iS=i_D5dl&sRyiYu1>B^hldS`dL=TA?jTo?W{di(iXv`-$pdDMSKg`VBJ
zctK`e-`Iau^4ni_Rz5i6qqoOgYb?9><F|y@3enepalr=P2dA!5zWH+V(XA8bJ^b^4
z`+xrMl{0@@G+~)GdEAlv3yWJ<PHA1yexQ^)Hs+7tzx=WEpK}*5mEZh*)^Az&j{`5=
zdC&RR_bpv8lbV`Z{{7%bt0wc84Q6KbH1&>Y9@Ex;s6UVN|7uIG9d+XqUlay2)9ZRf
z^N2zHhlcL33YU-Acum{Trm1z^+1}aFe`sj_&AYPa-n;y^p-nUEdR+Usv-=MXeUW`#
z^H0N#@q?L5>UwPR*g^e=hW7vX)j$6HSzmfEGrO)knmY#e9~%0R)n~qz`9AghVCJ&A
z9@#vywg1r2M|XU%(mi@paximwT~BD9u%Q3Y(EpyCJ?*RYG5?UJIdwf{z?4Y;p`nkR
zGAr}pvKK!a%*?3kGn&sB)PHE`pMwuQHS>sE8rn3iuDja1I{FU{-E>{#>gxxbeezG&
z)zbf`x_<AE4{rbXqO*qdT9RbBpBn#<bxx7x&%o9sS<zclWJPaFk`28rMK<*IB-zp1
zQ)EZ)NRk7+BSjAMT#`8ST#7jK&Lla}J5%ID&nJmT&!>n-FC<AoFQiC7?@E#jy(>j7
z^kR}k^kRxc^iq-}^iql>^zJ0N(YsURM(;_I2fZgn9`xQMdC_}Q<VEjGk`KKvMLzWY
zB>B<%Q{+c4CrL&xr$|PxBuPQ9q)0)pCP_uFrbtB}NKycOAVmT6!6XII2U8S8uO&%C
zucb&suO~@Iuct^yA4*aPeJDjC^hS~l^hSyd^fT3&-6t;L-dXCb6V!OdTH4b8Om*^-
z8`&{(Y~wkje!+z;y>R(OEY||XEWZXykYznynqey}tr06MtuZSstqHQ?(iXPC(iXA7
z(iXG9(v~0_F707EEbS3HEbTEnEbR%h<I)j!z|s+Mz|s+Oz|xT*2QIlV2TLx(!IFz{
zu;dcN;nEp)!qORW!qORY!qS-_CocIg4@*A6!;+8ju;dfO<5CC<uoNN!EQOc=OCdo5
zE?r?4EL{;7EL|}dEL{n5;Zh8XuoNRAEX9}zOEEzrE~T&pODQ73Qi@5iloBN2(j9ig
z(j9Tb(j9Zd(w!hTE<IrnEIknqEIly~EIkSG;L;oR!qOY@!qOY_!qS@{FD`vyA1r+l
zA1r+_A1r+d^5N1S_QTR2@x#&|^TX1gAU`hUunbE%BEwRS$*`0YB;!&EE3i}|3M`eF
z0!t-93NF>K3QIMj!cvW?uv8PI;xZ5pz%mdCz%mdEz%r1a04{^!AS{ECAS{EiAS{Ck
z3gS`=Yp~QJ8Z5P#21_kL8ZPy)4of|v!%~mwu+$Tz<1!Qu!7>yH!7>yJ!7`Mf5H5|d
z0ZSudz|x2rurv~6;BuNeO*{EbV!ApVXofoD1U3E@%eVCZ>-FTf3`>zL`w6}kRLJrv
zu(e25^wtVl(c6k-LvO2)4ZXcccJ%fN+0i?S<UsGJkOMteBn~}SAr8H>NKW+53OUj9
zMdH!(72?qgMH0{p6%x?9isVA?s*nr4SR@g>SRoO;R3r(#R3QnyyGU;I?h3ildy3>i
z@2QXny|+kS^xg`2(ff+zL+`7Q552!ge)Rqd`O(WolF`c*lF=(gQqU_EQqZeKQqijw
zQqc#B6hI%SPyl_fNI~?$3I)+?Mbgk~71Ge_Mbgph71Gg%iWEX0s!#~MQ6vMsQ6b}`
z@A~eY3)&JJ@1-HV-Dh<xjVrNnQ#iV%|7H-bwk%m*!mBGwR%Bb2Y{>R3*^wPtav*bA
z;*gzLaw79t;*o_c3CON2xsb&yiO5oxBxHA%+{m6Rd62zX@*?}P<U{sn$&V~&Nk&$(
zq#&zVQjr5$3Lpow6hzjtq#^5B(vd@13LzU=GEQ>iX1$}O|0W$yTZSx8<7vx~71@>{
z8?rq^c4S9}9LQXTIAmvroXC8Jcw`|%0<tSZE@Ux7BC?br3E7<?H?k)~9%OHZyvV)`
z`H=k?@*~R`l981RDadMuROCR00?5G(1(CH3X~=qpbmUNmLdZsj4CLAB>=WPV8b4CW
zQcM5OQs68W7|V+ce5)%nmi@?8#_}n0nzig?;q&vXWe;+Rwd_OAnJh1uPJX_%z*w<o
zEi+c^S*wf{d)72-#hx|KTCry>u~zI^b0#bHYz4-KJzJTvVb4}&Y}m7<SsV6jdDezK
zTZy${&z3XUuxBqYcI?^9j2(OSDr3i<J<Zy&XV0^C?Ac4K9eehi$&NiofpK8ZQDz+2
zb5t1z_8e)}fjvi_bzsj?Vjb9X<V+6iaRr9M9#>{K>~U3w!ycDrIqY$Hmct%bVma(_
zITMFHXMu5I&sk=i*mG7HC-$6a)`>l5o^@i+Sz?{obLLD=?C}ML#~xp1c<k|2hQ}VC
zW_j%Kd6vf>Ut)Re@i`NZJ)yt|*b~Z(fIXqg2-p+Stbjcs&kEQRO00lAA!ib>=PEER
z?77N}3wy3A<HDXR&APDX%Cj!)xk{`Hd#;?xg*~yrh}aX$jEFt4%81w#)2xU+G0%$F
z6HBa!Juznzu_qN6342nRk+3IK83}t*nw798<yi@PQi+wYC*@2M_S^-=jXignabwS2
zW!%_vr&%}l+<DfGJ$H$9W6zy4xv}RdFdpoA%8Unlo+{(Po+r(Eu;<CM9_)EatOt9Z
zoXLYdZ-Mb*&s%1^*z;ByFZR4?){8xFp7mnSTVlP~^X5!m?D-0e4|~2c<HMe>%J{J7
zOS3-g`SPp}d%hCu!=5i^@?p<kVEowgml;3y{8h$}J%5_@W6z&w{n+!DSU>jsIg=lI
za)FVtCzlx+dvcYLu_vcl8GCY`m9Zz6SQ&eA&Lm?`DKHB5lrp1WPpL8r_LMZMU{A@j
z3igx|t6)#bnH21)1xCf5T4q%2sZ~bBo|<M=?5TNH#hzMXRqUxblZw4SfeBzQP-X(y
z3sji^_5x`(fW1JT4PY-&VguL<<V*qV1q)0Nd%-dj#9pw<1hE%Pvq9_y^K1}%!4ez9
zUNC10VoxhD8uqj@qhU|0G8*=@G^=4x%d;Bxv=XafPs^D!?CAwY$DUqhbnNL>M#r9>
zW_9f8c~-}sUSf6Z={b{*y-<M(VJ}o>Lf8vcnGp6uX*PttP@WB8FH~Yf*bC)MA?z6i
z#=xFYW(@2ZRmQ-ck!B6-8F|*go>5{A>=`+eak8hZIo$NaH#dAhrrx>cxmPuAg)iRx
z!iKk29UfZ0Lmt^aa&+U>iEh`%W$hQ<o~%DSMjR7pyhJ-=>Zud{ey3-^@%!mP-K~u;
z^2^D=Hzv*zF0DU2P8&D7aeKYyh}Iz$Zy!~Ectrb%j>gL(4?cFy-twLs_thUBD~`RW
zaaY$!*YEC*zmf{~1#!!7@#ZdWMYP3ji1xT0(Ghnba&ZpP8FwP`aUM~K3y7|`3sH=V
zh*DfabjRI@p123m8}}mm;yy%w+>a>7Wke;eAgXZ{F%S<R2IE0QEv_NzaUC%f4<Q<H
z1MxETvdN9}{rH!w`g%Uza!36Po;&E=VU1g=?Xl+uT)OmOz5Z~=u#Wb|o!*WC#(Q5G
zizH@9Z>QGT)A&*6klrq>tEcfY_TCx3L-WSI_i}yPW9sQ_lwMigV_ADu_nuMpm&?o+
zp-~#rYc*SeY-Ssf-E0SPm>ocznFDf~oj|;q2NKKzkjv}>63rrzWR`&3W;c+>>;dwc
zy+A&*56Ex!1IcC?NHHrws#ygJm;*pTa}Y=~Ye2eL2MU=(K!({kL2rD%W6$BY-kDNA
z7oBrEdmAV0Z=c*E-`TeFK>guyJ>wQOuJTW91HKv9I$~Y@;qKwxV;g6#ajKY>{x?0G
zyqxjbr}4@4p9LI$|3l-G8gG55|9tNF`yLt(*Z&SgEh`#dxc+w_Y6Y@IZ9w*@9mo-N
z0C7<c$QgA4@lhT~hzdZis0&Dpia=6S0&+*)K%S@v$Q$(n`Jz4`f7A~oM`a)-ssO1`
z6(|r500pB#AT6o^=}{dh6b%6xQR4*tr}x;Dju!e={j7G)?OM>dOdG$~&&=_+oWPl>
zk>y)_madT%*;XSPvb{!jWJir0$Xty$WM_?>$b5}>WT8d^va3ceWU)pfvQ#4p*<B+y
zvZqEKWN(eU$i5o+ko`6CBg-|Ck(C-L$ZCyL<UoxA$iW%~k+m9W$a;-*<WP-5$VQEf
zlU#o{(30NcZw<Qsx`o%zJO06y#}EJgbw%TYrl#<T582)o-v~kOqtv{C3+9HexN_e7
zrnx6%Q~#IvFU~*t=z^0UE?j!ml56WY`RZ=gG@yB)dC=e?Lr-ZPHmOYuoYU4aNj-JM
zX{VntQj{+8czu4^GJK-DOE`Dz|MzJ7=#Deb8rOOD_zCBoe}UCz=Xlo`=co%E!Qt(p
Nix)EbrHVmQ{ts>y^kV=3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.5255F78DB2565F3C.idx b/.cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.5255F78DB2565F3C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..86617bc719968eaa23eff3b9499e3acc6e54211f
GIT binary patch
literal 10584
zcmZvi33L=yqK5n4s%o0hfdm9320;)34Rq~ms)}9Nm#{NRXh1ZwNh6!VA&_xn>{dfi
zl#Dh{Ga6!CI@mZNB2FSPIw)vVqM~93MMcfPK~Z2{6Hk%<o%3=IH~H(<_w%KyQ+-pZ
zX=BHXnKIbt8@Xugx@C)3FJOH>UuNoGWc8}sGgE<2_xaZCY??V|a(^S^M>jFK|3P;4
zn(MlD$!G{RlYaSi3*T8bYwf(&!{7Hmc=5L(zZ^Ks@9w_k;hE#iH5=D7%~~?B|F!+=
zwm<vIgp3pGHvXyf)6m^H{g<3}PoLgh^1)BvoI4rbAa^ai=hgW+H+Jap&Xx<0eEIFl
zYu8mBKmOg&j#Y;aP5yIXeSPfH(}n9JwGETE_dJ+)Ma#M5*2S-%zh_l)@9VF<J9+5T
z|788V%=u|kr@8aetItm0LmyYaeAP-rns930nQikvpZ#7McP?!w8NGE->C;D6cIWne
zb@9^`*U!;Utchh+brDy+x9!b=fBPt(R=p9J)3rWrinFXNG4w*!fm0)fOe|GD{gPk*
z%d>Yq@$C2`4a0Ju>-zCqkJ7bA*v;R(vF`SX15Z5VY&<ui>&s`4+%zI9<6ZXcyBo6P
zb1kpb4+$LC#GjuC44N~&Hj6yO9HMX1z2PXg@0+{d9}}dv?`%?YpW8BIRd5<}?B@fc
zudY1J6d#LgbECi1b@=?r345|{dAM@IweQYOe4uVPa&qISZSpWV@_6jVufF&)XO1**
z$HUoEqw`KwJbvVGw<AaH4(1LXFlhL+$^m&@L(!G1g->3(<LSq)IujiEXyB27;ZL98
zGv@T#N=_|G?{@soQ}*B;)3>}))FC-$;<UblACN!u6@MH2ZSehXo$0%0+u@VdBR4MV
zed4L?-jC*wEohqg;F|e$r}jKECLyj_*l@2=vhVxbx;(yQz{CBv574{XyYqd!Zrt_!
z&K;RI*WGuUV`UE0(rzg4wtUkI4dye;#}5DQ!Z_idW%upxcV%?IqQ$$PT}!rpv}fJ;
zt9u@JyYc1nN!RZhzNGn!gM|;Q`Ds(DIcG%o=R5smjp=pZ&Cp{#PoG^MeQ#UEJ-6=K
z{^1?ZOxgZoM%E`Kn_Iij9QbJ7w0BnSdo1(GN3Q$uHf~1v#M0IGr=5)6er9NT(f^Em
z&<<a3+B<)kxbm6s#c1BKiwzCx-Jj-Ob?er(JInu;9GXmDz47t!Uo-#o9jSPrI{(U?
zjP=~yZ;ovro>-PVaM}!yIc9!*?10++rR~}OS<<1?+vj)x{qze{j9-+Q*Ss>g@#NuO
zr=FZ~%Z@X9_P*L>>2>>V7a#0CvFzVZoOg`Je>glMYwkPup03>3{ZQy=T8~fua&gAY
z7rSnFKX&foP4ll)pPW}Z<fXUw?>drrJ@&zdqZ79e=v-RUd~|lx(T6@?lsr0QN!q01
zh98@Ermy{Z-Nn2UPn_$1X!F{$Z~V9B@Gt6Hvl2)5PFj*UbbhJ2^T~g`(q-ZFaOuAW
zZTxzEX5*}1cYHYe$o{I^vLAnSMa{7HtA4!d%TxNYue!W+{hc3P{H1f@6MyN__v3{R
zo?05*A6{|h8~$kTQSMXEPi{Uj;J2%nUR=?C<H}^&56TPr!Trw6^?^nIh$W8Qdv?J2
zjFvCXPHr8U`Sq_|$Gj*Nl12aZ#aAC`@I}@y3zra|&);WnbYxuX+lkJY=L>i9ck9?T
zq^=!pA8EfmeQe8kC9e~mo|^9E@0Hm$c+;IvJTd=*wqjtqZ+2?BPmezR+XioX=$UWl
z@6W3egTA?`>4n_FJKF|t+Wo!t)N?1+zu1Y+OilOn_w3L%c++`@-g<jS*)4_XzP!|Q
z?;gGTwhi9&qNW9#XU+a|=T7ubsp&!fLCf0)Z+c69)|lObw5AiCl$y@)XQZ_a-gLu%
z3o^IO+c-MiH!U@t-6Q*&w!xcj-g98ax96{1-igjiO=tPDI=2np^s!cr9G}>`KIog1
znjXoGT;Dc$)88%V`;Q^=DX9~klA6B4e?`Z(!JEF2G4qwrSJa*8M5m^vGyRzz+Xipi
zw`%dNkr(=<;!o<;j~saCWf~067&&9s<^MFC-#=gTl6L}0THA?x=*{1suKddC>z%l<
zX>BLW+Z+ssYDuVpcn2?)8%d}Ma|;Rmgt?W3&O_#F36DJANO<J=7Q!RXw-O$Cp_T~9
z3ynlTUT7f#@<J;SkQZx-h`iWHMC8R5A|fxg5)pZ+mPp7;jYL9TY9SKxQY(>=murcP
zyxd4+<mDD3BQLiS8F{6aD99^~L_uC@Aqw(JD^ZYFYl(`y+DKI7)fS>6ueK5ud99Xc
z$ZL&6Ltbkk8uD5z(U8|`iH^M9NOa`&7NR4sw-OzBqm~%R8;!(3-e@5P@<uB$kT+|I
ziM-iJOytcLVj^$05)*l=mRQJJjl@FUY9SW#Rx7cPw`+-wyxmA_<n0z>BX7478+oUe
zILJGV#6jL^ArA6RD{+u_Yl(}z+elpG-4@~^@3s;b`6)S5E?+x>;i)-OA?4-dL7JX3
z9n#9&l{a3#o`>(uy%SO-Hv(yO?&{0QdxIFB*7n8#m!T*PJxSp*R7OJ;m@8@MHOw_M
z)Bu@}QXYA}jPl6ym6S)Guc18hLX--~3uRP5UZ|u3@<I(2kQbv=L|!bTBJyG-6_FQf
zsEE82r4sT|8I_QiDyf9LR6`}?<tUYrm&>S(yj)3T<mDPFBd<iMg1k~j737slsvxh_
zPz8B4N>$|5GO8l4R#FvtwT7z5Yf-8pua!{^d99La$ZIuJLtc+k9eKTs>d5PrR7YN~
zp*r$Llp4qzWz<04sH6t+Mh!KPH>1=<-Ylag@@6G9kvD6oiM$o17V=gZwUD<esfD~%
zLoMX(D7BHd%czaKT}f@^?HXz$??kDCyi-OU<ef_DAn(*r2YEM2UF6*|>LTw}QWtr*
zhPv&1k9BDWH=ey`ESp-{<jffEt!0AY{9yhy?dh)EyQa2Zb2gFSCUT9Jz0aw2&!vT>
zcV&_1|2DPw*#j@7)=Z-Vqi<;MT72bV%VWHSXFLU8T6TJHYBf~Ng!V9SS*mA3Z((j?
zLLXs{v!Puq?4Mvm)tKwpP(9`nF0_q<{mZ$~-!Uh-&|b_H{?KlJ`~G}2!=oQx&+zES
zH!(c=@o|<%KR&_o=*QQwJo@n^9FKl{Ime?PpX7M-<173;`U%yHfPO+fBcPwq#0cmo
z#90CTgaj*~pHRmN=qHqL0{RK%oPd5pk`vHRsPGHuCss2e`ib?7h<;)dBch)eXGQcA
z6Re1SVjU}@pIE|)=qHwQBKnC*PDDSk!Y`tqRLw}}C)G0&`bkZUgnm+-mC#R0uoC)7
zb*zMbQVA!apH$9C=qDvP3H_uBzl45rH6x>+T+hhpCpR%N`pI!tMn5^h%IGK8u`>F}
zC7g_YayciXpPb}m^ph+6GWsdijDmhjJ)@wX(!?m}r^HzW{gec&pr2C5D(I(_a0>b<
z<(z_kN|IC1PpR-L=%-dQD*CDQjEa706QiP^8fR7XQxmL;erg@7qMusAspzMcb1M3&
zNlrySwZgBWpH|Ii=%>{)8v1EXjD~(%oYl}zORyUHX?3iIep(5qp`TXHY3QdVISu`^
z3crSadNre?pI*=C=%+U^I{N8xR!2WQ!RqLz*ReYK=_Q<wetJ2lqo1DSboA3J{5tv>
z)r^6DMm=MopV7n^=x4-P1O1EyYoMP|#~SEolyC<68ReXTenyfr(9fvw8|Y_NGbZ|(
z^^A#rW)owgpBZOO^fMEziGF4sYoec7!kOr2mUAZhnMuw>KeNJbqMudGSm<ZfGZy+;
zO^k(pR-Coa&q}Zs`dM|Xg??5EXQ7`}&ROVZB{>WItO~z{es(ouqn};R*yv|BF*f?y
zan?pZJHguMXV<Yd`q?F%jed4HXQQ8;<ZSe_EBrS4In|7Veoj5(pr6ykIOylZSqJ@`
z1nZ!mQ^z{!=ag^``Z?vCgMLnubI{MJ@H^<|Rx>X8x%G^Ter^-vqMsXQUG#Gktc!kb
z9qXc>Tf({M=azFW`ngHYML)N~@3#95zwZa%xihPqDI1;}n7h(jJhI^pfekCY#pAdg
z`(uZ;{E*%?wOWjk&^Ej`79%`nAw~quVvLBGr5KSg%P}HjR$@fKtj36nS&I=3vmPTl
zW+O%n%w~+3n5`JGFxxR=V|HT1!R*F}i+K`1N%0O+FuZ)&^3mQwny5XxqQ{WGK1_n)
zUV5*Q-rY#=mM)jpb}JXw-J5CXukh==y_wpWoz2w2>~5xRd!9b2>%eax$>mau{!txA
zb!k^x)NxT4Pig+Toh$Nd_kEpG%E-*P(z`$Gt*x1~w%f<;O85#0{-0OC(wkx*rn`eC
zWTajJfspPs1jBrQha?09NMb;QBn2c$azKWp1QbYWK!v0QG)Q_thhzi{NM^u<WCbip
zcEE<@1RO|iz`dN-j34#x4YSoLspFbGB73@b&pa6JJ)-w??_3sb+*^F$%dcKb9oOuc
z**AMTdB^vTKkmD?AJ<SpXnFe~3q?X}Ar*!;!Ov5un1t@fTuMSaFegc<1~Om3Lp2}a
zp_(t`v6?R?JXZ6igvV+=NqDRl3IwPYA_7zkg#uO!#YDhrp_B+%EhLG6)nb7N)nY`1
zYOzqnYO$DzSS^+k5v#={5wTh-kf2(MNKh>mN?0uw6A7!OQX*lslq3>X%LOu2%Mlr>
z<w6;&<zgaZwOmSMtd^5R#%iTNfodh9K($h+V6{?A6s%TCiGtNik|<cM7N}6IMpUR)
z3stOEi;0TWYAI2%T1^ratF;0Rs<ns))mov3)mkypuv#l68dhsbqG7dOphLAD(V<!|
z)UjGGCOTH@r9{VSJxO$|HVO==HX;U88-)f|8^y%HYNM1GSZySUfz@V#3DsuAgle<U
z#A>scm{@I=5)-S<Br&ntDzKp1idax>6<Szr6%z}qtx{rPwUs0mR@((ORNE07s_jA>
ztL<W9W3^pMY^=7E#Kvl;z=3Kf;y|@i=wP)|OdPCsN{NHjPLeoS?H0ID?M7Uvb_-ps
zc8iIN)ov+qvD!@%7puqRjN{wiQOD<uhcqE)0;GvK6Cq7fCz<VBL2kh?@0~apz9aVz
zNK12<LRyx)4ASy3%g1@#)(&sP?|*aErb{cYI0=>Gl|N2+%tD+9n8i2|F-vhGVV2`W
z#;nANf?16d6|)v68fHCCbj(Ja7?{mCF)>?lVqv!9#K!EziG$gV6BqMjelo1QRt#Hl
zdF8eH`z`C|e6=&0dOIH6ad2nvy|r>(4L4=kt}&_2Jv;a8;oSj!fxGGB)|+lOQk(mX
z>NDH>bIt4Wv_;(LUEilRcN@@ci1#<<H>W!%=lA*eU}`-uo1NXS{p_w|*Y)$x(p#T>
z_Ikl%vks)#-ZOhI@RkYQx*$ER?GNqk>jA!q^6g8xU|1-k0%oy@ikPJ$Dq)t3sEk=D
zq6%iUh^m;iBC285i>Qv-D53^tvxu6Qts-i*=Z{YDkNv(SUZ47Iv-GTC-g5)q=Wc0j
z&keMHx1nYdI*u>gX2N3@nu&l}Y$hURshLQa<z^ydR+@=|S#2gNX04fMnDu6&V>X(J
zf!S;(CT6RdSeWf*Vq<oiiG$f~CNAb_In(6!*UhTjRinJujkh&*PyW&a72YBHH_x5v
z(52_rOKbb?qxyJz1;dNe7x(m7?}3XHd+BKlkNrMS8KdS)mwpSmV!#!=x0m<)MP^#t
za~Iy;sjEsA3B8Obr;6~Hg(@Oo7ORMeS*juuX1R*Un3XD`U{<S$idm~78fLwU=$MTv
zVqi9_h>6*%A{J)5irAQ)D&k;vtBBj4FX%U%d42r=CUvCm%DZd6cWi^<O{D)O67urJ
z1s5h{r;!b*gW6|qp9S6}BlXnE(gl|uVTr6-7|FZz#VwbHKR)ukFMLe6=J>wy#gY5p
z_Ak6S|Cad+d^ca7^UZJnvaRZmxUhZe!ext=-j<5vZ=J}noWDb$W2azxMwiZGGe?iS
zE;DWHh^{@mU2$bE-LP_P#~n7b`weD)bx@!GZ)=aMvU~Tvrr)&#1`fVHC&UYqqFpU#
V4G~9m?{VX}MRd&M;Wv>9{|gJ31&#m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.B4883A86B075066E.idx b/.cache/clangd/index/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.B4883A86B075066E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a0adfc213a0231c7b592b547a37f3d106deae15b
GIT binary patch
literal 10588
zcmZvi3v?9a-Ny5u%esUh;bIUBgo}ts=*}>i`wTO~Rm%MiDj<=IL~br}i||RniZL}9
zp^~<Xb&;4Z;zqKP=Br8)B_b6eVxy-hMrpOE0X~#s#n5)y?+kgqbH1Fz%Rc+g?>WES
z$<DmlnQ7z3jF~YwEp6nYaT``Feq<q+mX_8f`ESi5;boc0(4$jY+J;E;thtj1=$RLs
z#N+`-xjF0Z=-N5+*`BYlqw1eL`rIe+3pHn7{CI5XwSU}uVcUR3vxWrvy|U}=_{sRy
zZ(2_Tj|C4dj^%cHT%Z5DSzmtj9=-Q&<K>f^)@@$*ug%Z>?&Q%y=GK#;E8jPsDEjg2
zxz%G=hu;{mA$In=tHV3QLrebd96o#Y=$YIFpMG}y`$XfHU*F1a{MY$#S?KWZ&YWDe
zvg*6UkX^0p;fa;KSC4z><)V+C@9@=^Pbup<{cPzIu~&9KAAb3^Z|6q)F4!u)@L-pU
zt}}YY_ij-?ee9=+13Mo5^vbxm-=8|BPx-WemQ>2+TZRZDo*Q+L*){pzzaJR??;)!{
znKFKQ#rt<?6FwWcY|oOfpUoSV*?IqwLkIWO@5w9rTo>lw^I`Fim6I0?-#O~Qo39P%
z*D1JU>)-4}oqn@v-CxdTahD$IxnUx=D|i3=r?<RAOKBn9Cmv-!``&(VK%gk^M8R!+
zni~RoAV#bH`Fa1&AN@!pE;ijZbo~!UZaVYZ35T-hmmQdJ`|&x6W~K1d*PBP|%5(G9
zl<yeyMa$>8bHzcwekOZr>Ahd=FF*A`w^OGc&9DX!ywjU@VBqwwp$ng%%YSv@CvHeR
z*Yk<nyYxSKq-Na0zSYCB*we>)-Eri@(~nLY77G-yomc+iH~ZXaU6<8T@PnS0r)NI>
zwO26s#lJqh{pxAsxTEWLG(7v~{jYRCfB30S4(}V2$XmDK-QVat-}v^y&gDx7KGXlX
zfm&Dl<@?k2j@=uJ{JP7$BikQ5taWkK^dT>Gix&+&cGtd@<GgRK{ak!|&-S|hx0ViE
zwD{$U^;GiFp$+5v^k_Ktr-Ls|x@)hu^z^?gHx;kDxaE>Dcf>8RjycAdo(*q>dhGu2
z%K9hj2G6{9eALj`$9G3_7xd42brAp8@dqZ|Ub*kNFD~~TpbVJ!`w!ChA36S)bD6iC
z+H>mMyk5+pqun1pGUJg=mT~JBcQ>}}dFhR7Wn~ku)NE9aXI<1L+N<WRyzSgoVfQ;X
zJwf|sc3ajU%<21G`fDfN9J+1EZ&z;n``RPFd}D3=&2200x__Ylx6sX5&;Iyk<@Zld
z(SML<^{XCybWrH<ujYUCt8oiT9{TCsMMr*n$BT2`|MT4Vv5POh_yG;w7d^CGJNeS=
z<sH^6A6~cn`20RQ_KWZ5eEHV9OQyeEeYxrR2Ua~cx}wMCpI-XM-rMf%7#aA(p~S)c
ziF;o^aiQ8g(LK_k?93lxy)Jxysjc$3{QdZBx_GYW-|M$sT=ryW(9Sc}kqO3!JCChi
zIqcZ*$0rv5N{(KBdR6aTp-a;KgIgb)_rc03r?bv)bzi?{)s{c(O#H&B-+v@_wdM5-
z8w10)+_z(J%^|VvsW<ohP`_w>=Cw_mhCj6X<8zO``lB=PVng$~CqI7eY<A1)1K(x7
z`_Igyf9!NVaYd}J>;3#+xA%YN(%h0hOSNwkkEKO>9h<dg<BI&9l$I9AsVyBDRr_{i
zX5O2Y-!0IsLuyDKIZ{9AS6}#S+jx0;M`lKHx@Vwgm(<`-cd07wwKo3E+Z~xH$?42M
zW{1?^Pdh(dI{*8xHdb|HW+kV41bW<*8vN;*8GqUTiz(frozkW!r+a7h?wcC?>BR?`
z!~5T>nAIt5PI5XYD`!Az@TZrp=ss`K<Uf3$k#<jVdNIFvMQZS;AO2Ha;L<UBXQ#Aj
z$?5E@?0%`ipZ?QixjL`suBSRO_a&$A4BT0e8vN<ghZ<&{YrA!2M`m_%x>umr&8fkk
z{$^p{KMu`1FLh)lC#Smyx_3wo{`8gRXW7yhu8zz|o12^-&5vH28vN;NnX{_TtUB^l
zM`mhrx=Wx-ht%Ltr-c_US@T-|WW1>KDQkG|Ad`_ld*tldH(K-a2jr`M@^>Jmr*_=a
zZ~gSbfiKLy{*D`$p4wsl(;4}pS_-{Q{>xA1MhZ1yZlTaOm|H3I17x9=1muNA5|9^K
zNI+g_B>{P{mPF*mMiP-1TS!D+Y$XwSsg@+<rACsFms&_dUTP%?dAXKk<mE<^k(XOY
zMqX|u8F{6a6y%jgQjk|#NI_m{B?Wo4mQ>``MpBViTS!G-Z6y_Xt(G+8wMNpA*IGzJ
zUTY-{dA*i&<n>0<k=I*DM_z9w9eJac4CIYQGLSb~$Uxp`B?Ec0mQ3W$Mlz8%TgXJ-
zY$X$UtClR}twyqtw_3<T-fATadApWu<n2bXk+)mOM&52E8+oUe9ORuwa*%gg$U)v|
zB?o!8mR#iBMsksNTgXM;Z6z0Zua-RIy+-no_gct9-fJZf`DwY+ZX7!@@~7uchcq*H
zCZt)pvmk}7@YoxN-264x8c1ucwUE|X>ux0f1o2RM>cjwtp@Pr~I0%J9YatbeHbW|+
z(Bp6b3YAf47v=<oY9I>*0yGO@0h)zE0h)y(60lh)BLSO*1PRzI7KqR+hDB%=3q@!a
zi%7&~v5Z7)784|5vs55KvlNz~St^vESt=q4o24?6uvtowgw1k+49#*_hGw}?hGw~l
zWNenpNXBM4K{7Tg1qw7PVFj9%LIs+YB2utfDI*1&l>{l+tQM%ytcF!+Rtr^VR*Oi*
zX0?n|Y*rJbVzXADL9-Utpjj)_pjj&-4V$$x(y&=ekcQ2Afey`jSchi4P={u{h;(e$
z%Sgv&JwZA)8wCb58({;QjY0#OjUqCz*(f6en~ek+*lZS<&}@cHXf_K?Xf}(;#AdUM
zOl&q2WMZ>bU_rALwxHQ6w4m84A`6?XGP1DQN|1%kc7YAecG!kyyU>PayNGOTw#&%I
zW;;PPHai6lG&^Ajnw>%inw=tYu-Pdi2b-M)IoRwLxX|o|U1)X-U1)ZT$i-&2j9hGX
z6XasESKvXj7xtjpEA*h*D<Th@y)yE!*-Maz%@cAbg#0CDM*hUyiI67cPJ%Q!cQT|Y
z$`rkwTW&3P+D{rQtQC+}S}P%~vQ|M_J!bVdpUc{iesuGd{~E_7SHZcnz4r3_jQoD>
zX}z_6s*hykZ>5Z_)X&d(3$9LmyrSXN<U)FEaO{wFXz8s>O@AZ*kCkY0C4K!*6*yN_
zv7uL4I3v}wp$5!NZ0G~b5iYccgYDy7s0wo(7plix%7=FFuzi#d?ZsTphZ2}80-=`z
z?b{1gtbl$(Ju9G}(8LPpCqy^_{e(Ctpr2633Fs%3@&ft^QC>hlp_&)aPpAk8=qFaO
zBKnE-tcZSM6Dy*h7~w?p6XTqSeqtRbqMumGi|8jtc@h1@YF<P?u_7R%pH#(4=qJ^)
z68cF^tb~42gp<%uigObBNp+lreo`qfp`R4xCG?Z3c?tcbihzWEauqA1pIpz%=qER^
zGWy98PDVdD&dKN}*KsoX$)&uEesYwT(NC`CW%QFP0y6q3Rjh)3N<FKfpVGuC=%++D
z1^tvbr=Xuw$0_Kil=2GtDN$ZQKc$*i&`+rdDCnnFu`2qh^{k41Y7?uXpBmv*^i$)U
zihgPxr=p))%B$$7MtK$e)M{QuKeZyDqMug9YUrocvl{woO{|7~T7=WkPm6OJ`e}8X
zhJIQpuc4n7<u&xvs(B6lw2FX+etH$Fqn}>S>gcC8u{!$c5l%-xJ<jRqr`K^h`st;-
zj(&QS*U?X}=5_SbD*`(D8C9%-envfOpr6si8t7+4I0OBRIA@@rQO6nRXO!{=`WaE)
zKtH3JH_*?h2pH&RR<S1fnf0uRer6MEqMsSzO!PD3oQZyB9cQASS<0K}XGVDw{mg3K
zL_f14V4|N@#aifR)w34*Sxu~kepZCD(9eo<7W!FroP~Z?DQ}^l73D4Tv#NOu{j7?B
zg?@GwYoniC&)Vo`H?cPQ*%8i0KReFZ=x5h)Hu~A6yp4W#l(*5(uI6p@vnv8N`Z-ms
zgMLmu>!6?0#5(BbL^ucioH*y8pHs&<=;xI34*EG!-a$X7ns?C8sR%ge=T@;U`nmP2
zi+*ks>!P0<;av1{<D83rZXM^MpIgei=;uax7yaC7-bFvRBH*H*SH*hh=hd?w`gu*P
zhkjm!^U%+Wa~}G6b)1KOUMcUPpBLpl^z*8D5B<D~fY<KlZEsEc@$w_587}|6;C;*e
z#Uq!$DY$95zj*xluj_W4*w)&qYw`=#5(+(o-$#^?fLSad5wlc65@xxCWXwtlDVWs~
zQZZ{Kq+!-eNXKlHkb&7OArrGzLKbGbglx=C2|1YE5^^znCFEhAB21C}Hz^~3wYz$>
z|0YdTpIwzT^q0?2Mt)DN=Scr6>KT)|4mww6@yXTsm=0q)x3fz+Ea~jC{<Q)=J#~!$
z7U$O}w4G^R3_I7z!R%fm7qfSby!Py0WAM^b*A}3*m<c@xwV`q*RDrpY3B887h6%k5
zStw=%<b`rZKwhY11muMpMnGOHW<=!0az;d6tYk#w#TrILUMglJ<fU>(LSCw5B;=(U
zMnYaLW@O~$az;j8u4H87<r+ptUMXf2<dt$pL0+k36y%i}MnPUJW>n<Waz;g7tz=Z>
z)fz@cUMpra<h61}Ltd+7G~~4!MnhgNW_0BBaz;m9uVi%O^%_P;-Y8}a<c)I1K;Eci
z4CIX(#z5XIW=!PGa>hj7tYl2&%^Jo;-YRA+<gIeXLf)!mEaa^k#zNjMW^Cl`a>hp9
zu4HWF?Ha~L-YI4r<ehTHLEfok9ORuE#zEdKW?bama>hm8tz=x}-5SP4-YaH2<h^pn
zL*A=oJmkF^#%t&OqfYXxfa@0){lmiYyLbG#lXK^U%;Zr%7}DAgIzmu@BnCxDQc!{<
z2W3b~P=TZdRY+P;gQN#_NJh|rWCl%0R?vcE2W?1B(1GLzT}WQgyOC~OJWWqsH1%J9
zM!wKSLi^D(BVTMI5wp}r5@xxLWXwt%DVWtZQZZ|7q+!<ENXKlnk%8H4BNMaLMiyqf
zjqLWkZv3d@L*^(`l4qLi5!o~RSDBIDdqnRU{u!odbM4~|pMUY+$+J}Utn7LIM%({q
zbhdu-m%WltXJ%$*`4=`yZmN$y`Br$#^%H-DLQy>NM@YadMo7dgMM%OdM@Yu3L`cD`
zMo7i1MM%S}M@YwPM99EwM##i$MaaTzN65zPM99JHM##nNMaaWERhSAV{)gR%Z=CpF
z{&w5OxnD#|lkeT&4ufy@KQ{+9yv|Qqv3E@J=^i)t$nw8{dxgLI<4boh)00o<jLMng
zpP~=vO<Tl|-urFx>23qN4fU^kzcpj{*9&q!KAPO8_i*?0Z(ltWa)tf<&+$8-etNiI
z_w0rw+j~~;h5oMgMcDMz#sBtSb6n_U{GWy1N9TpzKW&RPCYyR`y<Go30ssGQ>8bk!
z+M7bn6gq>y6wM@H7Mn@LEH#saS#Bm7v(iioX0@4A%vv*PnDu7TF&oWfU^bh{#B4Q_
zh1qT<8?)0)4raHRT+CiGd6;M9&d6)uCy!W<jPm!%mjZwJrp`S|t}pv;8I#k#(b7&!
zd-xmqH(DeQ#Mke#@ZZ;Og9mmn#;ey?lHCV(7yPaKj^x(YZ@TcezAFE$f!kYkD!KJA
zeprrwpM`%LMwj%|?HB&m$q#ajLI-f4#Yn&`#z@30#Yn;|$4JJk#7M!c#z@7i#Yn@f
z$4JL)#K^#G#>m8M#mK^J$H>O)#K^(y#>mC&#mH;V*R&y*4{sdECilzw>FXExd*b?c
z>I1e?$j|>;cx^&<Iu$10)SP>A7W$vk@Z#kQufM@-!Vj#Oe*M7%*N6Z8neRVHOS`}Q
z;k@>T3)&y1N>Y){A2#NJw{D-}-5hve-u?3zEKGY~#iHd8Cg;X4naFZ{;HF@Qjv1XY
zJKsF6%jl7JbV(mKqHB+C-EZxw>8525cif-dGQ=35+?n(LJ)3n~cJIFZ`rkfq(BQjr
dLxL#DYM;DbLx+vJC2Q=@7cpZdd3V!<{{SFMDwO~L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_gemm.hpp.FDFBA95BB6D583D3.idx b/.cache/clangd/index/device_batched_gemm_gemm.hpp.FDFBA95BB6D583D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c48179452a27ea1d886f4c181a58728d64ca3048
GIT binary patch
literal 960
zcmWIYbaUIm%)sEB;#rZKT9U}Zz`!5~#Kk2=nFWjt3@3pC6?67Zyz76+K;-E6<sTaI
zAEf@cD<C9sY(~bHgWs1<d3`PCb?mLY7J>ie$3<QyoM`SmX}9UVt?hE#&u(|Oo`0j|
ze$@Dl&%7U9T)!{AF1dbKb{?Nh-NS<W52stpS^PLvIKS}p{Z;F$3f}%QS2NFZN-lbx
zV!~GLcdE8aQfv1Vw)mnecf5kOo{--$@px*UqSGIx675g@r{^4N3cGqk`tARwxaSLM
z#ecGESSZ++n(w)_;678c*tMwK!+%>IM9l3yJ4awkn8mlW2(QNAJlCFX_A{H#zP0_#
zqHd*S#vSLW)x9gNf7KF}uVS4ohTW$o@MUJ7ys>Co%9|rQOm{|o)eBBx5D0D)kly#e
zZQ6v(LZv4}{lnd)1g1Pq`C)T(!>#^9B1?oaH#!J#zDakQBK|@2_g}8(cOF#EdH#QD
zn7mQXD&cGWZYEE~N_Zv;)Ot^m+EQE3G3$g=;upPFITma6R-{epxq4!=eZ&+Eoy_;b
z^Ud~V7FXsb#RJ1%PT>AF^MI3I1ejRafm|TqG}Sb<X5fHutR%V}UfPK4W@BPzW8jo#
zmzDt&44h2BIAdVsWQ?eNz`I@Z?Lh%17MOlH7CCX4eukpdwBkIVT9`eIoM7`|=E7)3
zF0i?tDrMahcRAO=Oa-b2nFdo8yWKJ;MomSTk%v=OMpgr+Dz%SQwLz%Z2`DVBCT#>$
z-R;HP^0xoK1uqY$l8llvOf?UuvWzkkPen!riKi-~io{csQA6UX%cvvqG-Nc8c$zYr
zNIWeWEhL_{j5ZQaM@9#Urz@k2#M6_}gY(uJTLm1Qn=l>dD^nR$byxsCHDLY1&M)50
z1PlW~S!sl?Oa)D)VTy`Ua}q(h3l=gAurOgLE=n%e2gNPt>DOixPE{Ux#mvRc0go=A
O1UE<)0=Qrd1_l5kv|q9S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle.hpp.8AECE81C6A31D6CE.idx b/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle.hpp.8AECE81C6A31D6CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2dd29da73b90b93e8cf6f6e6b8fb70895533927
GIT binary patch
literal 14622
zcmc&*d0b8D|G(#X&dGH<Zns7It!PuJv|Ul8qdh5UEHfDU42BuRFqX-dr7<(Mq{vo?
zvL&I3M2bqbc9kN9vP~+<%<sv)ALo4k`n|s2`D46ZeC~Um&w19*{yY;N=;x<9L?jwH
zGjQ4bS&`Fdkw~OT{w|JOG{>@+Nc6phNVIHk>4f>;S1V*sn?6cbHAkHf_@>Ztxzi=l
z`Sc~;Q=EED_S$;D@`i5z&B!-7Fy^n0F^)Tb)v~mkZ#dO^Uht14@|zofe6sKEhw)#x
z?pgNBw*hEwamh{d;Om*gUX+%Pv791$qrYjs+5U*Pma@<_e`@zH*t)V%=R?kKN8F?E
zxA7wftv}*^T*vfVS&`g-Z1u>g7pkl>Gl!Q_zpwRLX?Q1WW&8A>`i>dBK0nu(F__~|
zt$pHCwNU%Q?WOu=KUDr&wLe+?Yu^WNHoaPP(=}GB@%MFm^e(kg>MxvohnH5m9nFkz
zuij;ruy)fmt#hyUEY5s;S!c`P@P!LQmLDlPJafXu(!NVm-0$~29KP*#?#Vq{m&`9-
z8a}t_o23r%<Hn9V^f73rUFDkdQK8mN7dyieR@<l*etgmu{-?&pl0EwCnDQ@{&Y!hp
znE1f*^T&)v=XS=vO|U<DeR5aezyV%I>;1l);pt`6pq^Sb#H7PHaj{L!9lMv|iPa%h
zE#n?;qk<mmN7*?WRs8lq^Oyb0ynlW8z5hw$25pD&i3v|S*Xec#4xHOL_vXfQw;8La
z{qXX}&z|DZ4vXXk@ADQo=xr|yPDL}FTV_3cv}k%}rNM>05esW`o=+NmqcuQ#_l`jI
z+MHPzpUk=4$UT1RTz)mZVzu+0ULR8vzB4&*7~kxP0&9w!A}c62&7m39>`?uEze@b(
z<K+8ugG;w!sn+{1#EV}iQJ25#6`HVnkD{?^_T-sClM<^s&vxFEmOWcNLo#%so0jd1
z8A-ckrWcxy=tca5%3f^pTIF|r<(O0E&Tr06@jdtT_4~s+Ua!6~uk%2;{mP0aEj8<V
zp0n!L-x%F0+qkRX(Dml`bj-F5HEJl;p>OnxF_#8^TxwiqvnTT0A8%&v+<t6K;=qYA
z%e4*xqfRkbE=}$|#B7P4Pnq<1z|HV_N3E}Aj-G|uYtJ-Szj=}BzHoxRddA1HuYMV_
zHCfw#f~M6n^O_XHWc7<`1=kiN^*P|5?wK=J|LEF+{#G5ce;E=pH9p2ECSv}M$6M>a
zNi*u)_V)W<ePY~4Uz?qJV^DOm<?acYaXPv;WEt`1qe@fb@6K(^T3Ztl^fKE0_x-b$
zd@(B3Afe1KE-iWD0IxzcDlLEJH<?Dq(^fCq`<3(u!&H}9X73C3=QafQj;J4B6dv5$
zF8|ll=1HPnzT1EK;^DFM#ruxXODAYazR=WN{3@Vd;muvk>@(c8jMKl`yS%GiE&Ei*
z>=om(?bm5#_4`s`G0Rk~Pl$81*(Af68HwtL=GEx#-Cz??-|W$5pORGkbFZ8@>lNet
ztuww^(jK<!QcI%k7=xoZfny`fbe>HP4_S~?*v4qzo$_7hxHj*uNLOmWMA0pa)DDjy
z4kzAu*5<Nq*US?erf_KKm-F(aqtd$PjErBgXISBbyqpoMxSKU4sT&TSOk{ilHKLck
zvzZ^Tc-Z7Y5idf{Mf^2-%TMj*tyXCk869Sm4=j!EYLjoV&KezYYRci>E=vufAAi``
z74!7;$DuL9?9$A}-;TYsIdu8ErL2Q*#r&wvdS6VNe75`RhhCE>(~)%p51+g0Q5t$O
z`M|t1*R`WPw(RX6X(Z2m^P(pEYGOu~UB#L+>S6xY2RvI|QaX!%eD`%%%NHHHSFecZ
zmSkm>IX2zNyybZGZsz4vPiwNyIv)Ld@w~1@vptS4-{A7#P2;F5C)WPFjy-pHNS(#A
zmuA{C3UWMdMc>@GcSh&=)9=Q&&fh#H@U(H4))ITpX2!fZ9~Ts-WtL5kTU~Kw!s}Oc
zEhRtMkNm0gM$+-NcB4q{)v%4T|4K<Px%cenfv4oD4(rgOz6UiTtQHJUZ<|=V=n>=T
zcy)1AL{f}+uy_j79Jy)q{K%o)(fVOohmH5$IAO6xU+;l^rPo%kTUvD?<*n~0HcZlX
z&~RR!bx?hge^bVzRK=kCwu4_c#xFnUd`gjd`qX)su|Li=+T8C)$6Y7hMSbJ;G$YCT
zz?B`ctJfsvGqwkxS=v4|N$yyAqTlz)ll(KHj(2Z!eD~e8f;~0nt0L{wl8xGuJa@m&
zsvLD^P;7+L6884&1-f6QS4!W#9I#4P&o&@2Z{m1cuR@V;Wb);z(Sv_k&=`ALZQa#a
zCi3*w&B4=Ot(KiB**QZ?>u;y_2Jg8g)6d3cIm)9Sm@Yf`*xEtv{^LDoUH5GV>o@0n
z5BD-&sQdB!)<aE8U!B=s?>2tMev5<YtBz1x(iX?y5eueFzTT-W-H`cU_n5Xn%wrc0
zsqhIr^UfBPE&sM5#Blese06r;*1q$+y8p;IHbwS6?ZgcYXUl!=p*Qv7MvZu4^ES%7
z;gaLo>J7I`Rv+r#^Emm(!1it9#2(9TbhNvK`0kTw>6{L;(oDI3tH0Bu>Z>;v*7mM6
z&5QUW-5=kM)3<OswesW8FAILgDStk>d(Ecz>Kxxm%LYhKj??{JR~nEpLw{GFFWM&z
z?>k=6@WiD(W`D1HS=~~XUtaEMI`p?FCuhV+(c<Ovr)!Fd+kC&xu6WhUd?|7qEaF<Q
ztOXmZh^VKT{Gn#Z0^k1lEME;7DMVZwR&T>vD)O17qqQ!z^WH%%)Ss|4P_hQfUPXX~
zzL*-B+clw1id@M+jhJo3LsbM=Bp$yFtC{lkLUp7zM8vgYX*<?c5ny5WW|g)e?N%*h
zLI%pCWO-yDB>@(h?{(2FpP~mcq~$K+vXM(R8m=P1qWvyp{Kdas$)soynafk`_7wZ8
z2(aiBzb^ghwoiu)8IytD(Y@c%#wr3VhW?o)+M&;<NzrK1OC8p(BU`5=z%s~t#Q^V%
z$3IGu3t@STbsu9Z6#*8f6@v>NZLiFbq5*`Z37a%wI~4&I*9B`rR_5qm)kRXWqO(YR
z7FnnWu(+?$e|<Xn@FO)Oa}#lCRIfD3Mny!uF{MFHk)wv>6$GD^O*zpQ>60~OQ-;};
zv5G)Hp0Xji9WkAAWJs5=++}2U8M%r8%dn<#mjfK9c*&3@VJT$#6*4X=0xaC=l2`3p
z52dRkF=44hl1e035n%ECG9+rkE$gwmNR2Fe9~SS!G8F+9KdY%b+D<17?28OYFPAC9
z%VfGr0xSWpYm6UXc(FqZnG%*RTGmBQNlAbu(7kAa`3L7SQWQ)EdWa1k;{GZEEI}HB
zDy%m8j*+6?WPk4=**!8{B>|SvDf(g36JJhNM|xzSS|q7Ox+(%JlfT?-^KF-Tn>u0$
zO9hftkY1DoSZ1xTHVax486!pZWG>Cvq?w$xk^sxRP&MCjI?_st90*H2Hm=9cDgrEv
zqL=J23H>Wo3t5nX;uu*RW27R$ve?DC$Y<iAH&SFvdU=M;o?%B70hSf&@1|F5(%;h;
z1(9B^QifM4YZU>O4bQr!v?q^Fks)iMk2$n7ht^aPVA*IEVYEkQ+5{OgBfTUr{Sz2B
z6#<sb*1LnpzBe{eN1B8si;`qfS}Fo8JA)TmO?gm|qJ{bqmMA2PLJledEW7)=M6n)U
zE>zB?6H7a>CZ{C862H~zsQtE`RWj6<9Ago3DI$umB*3!w+FkF(9WRecksaBW=h*Z)
zwpS5gNft*s-koMfv7$htNp@np0iXU6^`?<1%!z2x@z>T5G$(eRlA;jOM-4WpAp)f&
z(8q~-%VBb(%|$Y#LuQhTTyn`wlmu8#UNAkafBelPb<~S2xe7_Dkco<XW-0ZFeKU8s
zhZ>R+meWZ6G_qC^U^$~(cx$fd-z&5}S*c=LR!p07N&+mG=HUHRxo-wbkqw#dBW(PL
z2$YflOTM4}iBU(sA1gyvgylZ&bDtP^CHc$}TfK0M#aB*J<V;vzV3QX_a+L&F?kG%8
z7FPfDjTEtD(WOXMO5|TjfTeWlGqW8XkJO}S6zQcF8`feA6#<sAb9uYpS8g~UMe3xN
z`$%@5^r9rdQtmeN&l~T*|5}P9gyjK}Js?}BB*0R4&GG{kuPv6Mp=6*|Y}tyvR0LR_
zaF3?X+Q%&Jiww!`UZD)H5bL5O!1A={scqE#Z-{VflY#P)OFr^a5nyTJj_H)eI%cV(
zKBSjwB&kMvDgrDo(?h;0-~8o0EhHx_`Ltm^?WQ8Y@~UN%Q{=TXy`(6VO!qaGzs61~
z0xWOlUF-gKEIUVv1`?JAEN{TBD)O0SYjw-@UVFA{A#*ZNB9bK%c~KHz>019PQh}|1
zS3_!qB>}4^U~Lrv7SW<vvlm|+K>WTcGC(YM4NI?KLonGAReS)QL3}_<R8avqBbHbt
z>o~>%3>M23GiJq%HF%I7Y*Wk(0!*o*2rN;>4tz(I$hH#sgI{5}3)t}j9tAUk*aEZM
zA6WVa_Jjs90+toy;eaKohy$jsmgQ#V>3cGpd{|9sR8zJPU$9&!rO`>*!jxDp7HPyH
zTd-6Rs*tHF;}qBg@*&pm#AaZx$mY>@*Ju~8r7!^bP#L!HYpg)F704gl2+L(+$4q{W
zeFklcpS#PcnoKpDGR>we!JD(hCz@WSEWur}M2bzXP?jLzeZDaqJm-XQ4n=bf4BNm+
z!N>Q>ICf<K^FuaGdPL%EB>NQYk$+fXtHilNoR7r$LR^HzMLgEff4~j8UptScxzQe@
zJ;8#~T&PE=0EKym4ddI!c#Qen1|b_IQ3YxzRi+ODwVp*|X*M>4gJiivY*&bf0hXx3
zH#k@b-$;+j*bTN>FE=VSy*BQnnC43Hz*6i0#;*tamEr&%mtp%d>?p+kWjK(><=DPl
z*j|nU%5f0iUV$wuu(c2mslXn9^~=wQyN_l%lTC}FL!;<0kQka<PY+p759RSSTJIXI
z4-!Ok9kh7|Z2?O((D1xD;z&*&#d0Huj0}JshUk>2?|CdekM;PiQpTvTRfhNT7K)yG
z&idDP#8kRw$Yya-Jy}RMHtNO$U<XW}S)cE{^pO{7uohWMKur5xTJ&v8lxH*%>~5^v
zjSYamzt_2oZ{6=m{_)n@&+PV)qgA=2qegl}Bh7)orMai{h^Io_L~~8F58vKQk8Y;J
z07D+2#&TI$nuSf^FnZ)VFdI+g@hxn33w!Xm1nZaJ0X(k211hjPNRdR9O~7G7HbGAP
zGL~M(M)09Tl~a6Xe42RDXlmZ~nC2=eqe{vc3=_?@Q%3ESF<3sD+lh>KA~{$-=YS_$
z)C?vnKJAITkv4%f_UwuLsSr2OCQY;{-`-5ioB4LgYKV#|QyNfuw-!a^cQ;>$k>j?Z
zZTiA-yK}p_^sasy|MiSP^^ruC5d~T-joEf-<Z5ylRNpLWFhr<BT%yBkubp^DtW6E2
zsG-Kdp0HdOrRbu@z@D(&9;CMiIm4cSDQ1ZM`jl0{kSte5Ih9c^d`Ev$PJdD^d`H`m
z^ETuP9kE3B-PY5CL6V2+$BcO9sgp&vB8_rMqXMAgp(bx`%~PB9MzhCk>6T)B5D1zp
z!@6ZcT#j|ig}4IiRtRw!!<O+*f#%8?wp@rS7=sGN2%H8H7sjAch^rWbDj}|B4623r
zA*20}>C1m#!)VtCaV=v|E99?ZOzIdjp8pYJ@<@p58J&7Ymv3)i+!~m{fPoUJQf9BP
zW1oWzi7Huz1_&xh_@c~Ip@HR|Bk^-&1i2qe@-T5dHi8t5<#x~}J7_28Nv_vdGpk4W
z+>Cu5(Pge6x|c!QWbhW8<u1}T7kTVcIAYUvZ&y!ps+=|F3Kknwl}ufk4}-3L_UsZ>
z`peI-kP0ZIf_U+|LwVhyym?$sIh6~=ubJ{{ro4IaYo!8OsURM=QBG|_@r*`Z(a4(@
z&lnUCgMxUx9XW3oif<AMNJ2qxRFNplr68{q<PF&W{$7uSFMSsqv)nDp^%gappJ*ZF
zS}4RNlxv9)-=SRZ2=PP8?jhyIfB%pQc}R`taV=$6D{QZ$?CONLiE?eChVu(=rd*qa
zxP@|U5#m<LwN;4UQ+DquH~#zgRLFa3JdZz6b{~Z8-IQIo5N|-P8_;lmBcqXPv=DDW
zu3LmS2D!!vaU5dfkS_my9CC<5?mUi1Y`m~N0kH`}d>FYNM#JG?SS|^<CJFH|<a$hq
zQ;=&4kD=~DL{(XB0f~fa3(<jeEKL_i9m>)R$TZYkh^;-3r00<lxC0WtBDX7ugP2gF
zs>FbI2nijmSKu2rhA2-nGH*s!pq@b0C91*<%&Eu6_RGfhpkFMPk2Ugn*T{0WvCeH_
zZ21IhKEZOxDp(R*YCgqs2u4|M3$3$-HUqt9xm~n;7j47iSXv${#Obs~I&B0-o#k3-
zjaJ$S9OKyDzS;9<$$lj(y36#w%V>ZNix{nC-Mr0{Aq}r_pVwF$^2Bk|c1=t@SUr?9
zG&0(ayfvJlo$oaDq`sxNN5S-N2{pfj(kl_-CQ7eKsOc?~UW*WKKzbXH3olSxklq#{
zK8*AZBNtwPjv>8cJf83&&gp5^I#&`U{BJ42B+mrxLrn96|JvI_{zS~VviQb}!80U&
zCN$@du=o)+f&vN4ZKqAP^X7bV=8*3uSy`Bq^e6$#6R-`4E=y`)`Z=@(1mrA9G0fl5
zju4O&enuyOF@xIJbmzb<?P-@Ukkd+~TvDl_Fr(?K?R!`4e4<Gj2G9faA<&ui6XW;w
zD$}0$r<Ins(zYNdP(CEebp=bW@DeOh6%k>EbIfyf%=$==%2>`<;;Rb|hUNTp{LFad
zFYy<&1?U6_+5#nkg0>)?AVFKOBv{ZkN@tXyEgM;7BWF<F9%Z(^f?Rl<i>z{yGZe^J
z?mDu*j$C-0kF4^MGgMw!?k2LniClPGgsh5?GnD#>>LTmg$OSOCW^!g1u=E19;B~$L
z+ZAAM9+%-kW!MkGCW)%X3Ol%v7VSwH^x@8@GqXsgoZ!qBE*oBAeCysvl62%@X&&wm
zJwlO#m_%hs7QTR5EIG9tEX~1k_(Gzp$ikRVktI>2vK$LNu-pU6@&V<{d$%`~<r~VG
z_cp&FtKWpmnT9mdkTEFd%CRn9rKg9uk;ODH8V!ss7?`z}!tJ&^%v?<xY#ePk5Ww|^
zfBq@tSYAHKRthMC0;)eha4}_4Oj+}|hGuK%z9671S4XpTLR?R?^+NoZW*_qyG>Gh(
zvbqda6i^+~psXpwRzRpOk*G?@(9lyZvCI@2`3u<cf)HQAT9^1DR!<^knuV?5Fj!K@
zF=)sAps+(iek|+4`n=KJPU~!^jd>%#gVxz0#5-x7oji^S!s}aq``bN`<vg@J40xsU
zmUwIOXryGM0OeBsa;X6z2c!%`+2&GCLhO=D4HaU~Txukb^C*WrVUUwYdFD|g`P0gy
zg7Sn}O#|i7APh<xD9;9JBrjGCR8WI3>xn{+QAol2geWvD3i$|ea1;vXaUvR!h#Wx-
zSyGHtB%)zL?30Lsd3+XGo<#%sNFW_qrVH^oWO+`AGmvG55N9K=Y~&|2HeR{NPiS05
z<RgEfG4U!weqcOU5^0F<Aq@ykSW<Tu-$xoyk|t`1#1D`L)Sp>W-4>T34JdrG+(E2)
z5bN_IbqH%765_*H^RN&nVa+5VPR5$aLVN^k9ueZBSo5e5AH$l*gg6CjrU>zIta)6B
zPhiaxLY#^<Q-wGUYo-Zt3l_Iv4JZWnL?Jq_u@OWeEcZ9o`kODiu-sp?{4d%9f<w|y
z%l{VQC`KN|2-?>$@-;%dmXWU&;&qICoe-~Q<m-ia10&xc#L<jAT8K9?@{K~giIHy-
z;?0bFvk-4#<XePzKcm>s3<G1tat9d20U<ufC=LqoAx3dXhz~Q0!$O?ID3XLYnNcJQ
z@exLGM2L?vilah&j8PmD;uJ=aBE-iT#c?4%!6;7fIOe%e*|)P7b^qha&2Pu%4Y<6`
z@D%ZAPq66|9L7h=yJ)*z^dvBnzf~@G?;|?b+T&kzZS29n{C-<+ZbGnRIQh@D$e<Pt
zgU}r;2Puv#F9CV$UWII{PypPTvD_tWd<hQ)RohvtxoM+)(+SdJ2`ww3&A=n=s<+J<
zaXINZX}Hhm-)Ahq_3i4II(e;zaVd!uI<b8x9>N>?P8{5c$Mg6qt$$S*?Oml6S7{D>
zSM1GzqxYj%4<ws$meD!OnDXLM%;*#|rcii*4v9}w-dlpR1N%c7l(&_T>g?TT*RV2o
z$>9Gvh=%J<GPkQ(dKDYPeo0i9o)CH@a&?|2V}kqq>owq{k=1f8w~J(}|IcoM3nfLp
zo;zB;q_>&2*i8F?1g4C?H(ZiA&z_{24=Ke%Y83Ab-&2bB)F|HB#UZ^oWDg<;x4dK`
z=dkpgFp*W>^}^PHcOj>zyy1n@OI?=5gv~$X#}X;06t%+QMJ?r9ONH=LsiPEi!uC2U
zppFXT+doi>55nTZ2g>&Y6~a%wn^JTO+q<cNZYqp#k4Jj($P|{wa`DJ29y#zh0qG?O
z+Y^vY0&;@(G?$@adHbgBCR@>ljoYv_2rjvVVRQ<42?y;Ysh9FXn^)f#Nc;jBgPxN(
z6pI_MF<ff0+)mnLC+!TU23trvQeJ!W9sOsim}E#;+=?yWzL({8(>lB9LC_JH0dhFX
zTXE><%%lqEL%(Qf5%n*{nx(ieBm$&ji#5xHxEyPi3vmV3tl)9Ri=Q&|w!A$~bo8HH
z<D6HVjazZIC4eLr_ZjW`d`^F1_M;*n$t@B|lgly9Y-9vxwC9>mGgpZ7k!HRS7a`3e
z9$%7*@7$R1B;;RLa>%&JiZZb@Q<$aYW4nA|mR5=#OR+B;jzslv0d((=g3^g;9)C9e
z<ITV98a|WpJ@kUqO$sRY0%`=WdySN$QK(z%kYb&%5O)wM4)TRDa&ti2)zB^=cs&`B
zU7Zlu({}Yj{Ft_T%;PIToBOvG|By>kuSP1UkqYIN_&F8+Tv%aThl1CkFi=vK+l0n$
zLKAp=5Ct7Xp?q+C1ce_#;{jhg7Pg>|pLqTMYiMEbi3%y7guv$KtDc0A9&)iXS6Gcu
zJ`DNHc<*ju;q$hf|JjY@KZ}pr*Q~fJp}7i%tq@*LR5EO(5LYp5l@M1mY_$+SWY~v7
zT*I(6LR`zRwL)CSuysQGh+!WIaXrJ<^SE$P%9sz2=P&-(HLOVHQ~p7tW$3>?k=y1g
zHmw~#R^yYM9bJmUL7aN*Y-kxC%j0q!R*oZhT!FnSc-1fQv|1S+Y8?BIRUNW55)oWP
z(u=~v^KImI8*vbhNmP%Qz!}{cKUcqV={{2uN{*mMNWlujn<k<L8CaUZr%$8|{po=d
zr~#yrWZ+CB%@huN9l2dc97GaC(LOzUf`K95Bp+Tt(hJC#PmpgSx0`4rB(xIMD=7G|
zqHS;Zw{crG{?Ce6vvWq(X2VI}ktOF+;#^7>N)6y|$)c}g>2>~oTB3So1&gkCFB#d9
zs!#pnV|${XJn;9Y8V+Q{y|lw#+8M%WmfJ@=?Bg*!dm(|L@--K@{GPkjsq5+K@L}U1
zrmdHTUJU6mf$5#VX!C)4A=A5%(T2bswwG8y<r^^A-=={V(_G{y4w3DNr}g4#d*~Wo
zg^@f)`6P^=LkkkOpuT({^cagD<G%1d;CTlfuGoLLH<{B@Z2A<t@^=ogv|TJc0s@;C
z;?&>J(9^#Yy(*@L6;lDc!+K76Jg5A4zp)8<Y(jp#UpRt1jvznq1Fat>b%rEF5B}tE
z3>qmj{=?^#$#cq*$Lo;6I%LLov<aDPLY6!}hzt%QGw7J*jv$jG$PzF-gX`JFPjBF0
zPoSJY*8HCpRf*~e9LN$xi)76gvGgLgglvXXiLu>H%<;G!4=Tt0U`xnzvQO{jK6j+L
zo+QVkd>#i=A+NxYZ8h?S^Xn-R+GODo(DB>77q2*;?q~6jrTI_UvzB$6Lwl}A8tH%X
zh@ySfjn%pzu4?@Aj-mrESll}M?!Ufoflcj2O0}Qf6Y`t+pSRdO6;|_hY{yqvi)q_p
z+L;fPI_SY2v=@JkvW?N$#^~|oq_d3KS!MvJ>4y(Tt~fr7Z6n&(jKs~TA1D=h!T8-Y
zA7Y>3^@Iqf<ddr2p!4eI@f0NQaK|HpRY0!u`G0y?ylBSaqdm5MI2e49$dmusGmP>N
z;v)W<f?rpEf4+}@fZuR0pAp_8eLaSGiadZrB=X~bj^KZK^FN32KS%OE1Nfi5{7--W
Jrx*Y8{{Z=lISBv&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.24C81AB80A523E1D.idx b/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.24C81AB80A523E1D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bf1a54953606e1317d75bd513e71f598351a7e67
GIT binary patch
literal 8092
zcmZvhdw3INqQ>*hOiFJ;3#FyB*a9;5D>H^R+-ymK3gOlW2&oke3RIQ_s(?UQHgJRm
zq${_B5aqVYrajP=u$9F`3oKQ^hJ^=NP$(RcyDN&K;$=7VJZaxQPM@dWGw*zFfBnrg
znSPlWS2}d)hCG{Xh_7_PtZ7wK7@N(O9{ZkGHFw76w9R%V#b#Tu?c?#YM&}f!-FxHU
z=rc!|g1u{xc1YY^G9iDbXR){N-tE0z_l2KZ)a6vg($AOlOq+BgIph=$-P+Rh`uEu{
z6WzA?+OZ`2;p5-D^TAhxZzm-@cly2kcWc+5j$S)=<N3v*BMbh${PmCLJ5F5f@KgUa
zyN>iYR+gSMp<&gW7hbHs|6Ok-Qe7Cm+VAZC+rH{K1K6RPmi+C%pSv678us7Xdi?g4
zR|{I`LGHJD{I+q#gr1szOt0%*_g(0|gss!JZl^cz4VI^+_|J6?mmJAj>-a;Z<lc41
zeyyVV)b$lv)%!R8cFFsfMs*u9`+WJ1wV~qLs*XeDlgnK%eYN=h-jlPZ{(06PSN?Qe
zSUCUS-HpRv4xLG-UfcG`Cyl@7XZ1cgC{)z{)4zOiXnx+sA${uxhK?-jzOjDn@rUPb
z@(gm`Ug@$=`JY{K(G<1&>D6N=F51wmbKRKkmrHif+tHzXNYU@kANuI;9p)?;?HF3#
zGm!X?etn&J2`k_2JzETq7_#z-eSOc&UHq(XN>17*=iFbsW($0#ObwOnS^n1I=eM++
z>_0JS--=VOjw#)Jr)2%9f8=;>r15uZDy`&kbqDm<I{xd@m8n&8%AScfT*+SXNH(+X
zpL<75v6W98M!sBdV9d-z9m(7sAHRR6+sS>yUmfq;z?whXjvkx*O<7UNjh?5+4Q=|o
z<3oSCyJ2a8bNTiB12d2J2}l3;UU;|T#TggQAD*;e>$|s3JY1hYcX`SBYu9%q?R=k1
zr?OxB=*fv^N*1rYy1qwx?fEf(uR8JQANtfyy}2o4_r-k~=1%I@)7v+Cibpk!dns$r
z^d+ksj-BrHQu%_hn@e5Ke7NP_p^rDuoUyKWQ)!n~!^;<(ee&q<k_W$0G4j*z3&Ul>
z`&&xKjoa6FxZvO;)9WA3ELQ*Q{IFn`)MMTE-3k(3kV{im?LKkvSeIaE@KyJn8Hu?)
z=H5)7eCO-S(`UPSZ~SKdwr}ob`4`=Ld!ggRrb~TB5BzHJ)tR^N>!%&AB66kV!krb{
zpWg9Na$WG=cc;5o>{u;%XRJDB<{z5CuHUiy_q|Gf{NehCWuD`8^*82(UhMSv(6{$y
zzjgS^#SfEz&F!(~-Veo-pSn5eaYr~~#I~_{EAnnFNh`^|Z~tU%?pEiKQ)isJI{rNT
zaPxh3#I1#A@6EgU!>6YXy~&?jJ9)ph=7)2iPC9aBZr^k6v%XWpmXZH8Ykkt9$>C*R
zY*?0bO?!9xLGP*Ao^fwhXw%RB`>&7pU%K$t)IDSWdSvH{rT+15|CRU7T$ruD*|g@L
z@34`fEzK2y-SeORWx><|!5^<(pFHvI;OvS1tS>(k7xwIPA==EW(J#JLdUfievH?f`
zweFX)`*RLIav|eN>aMSk{b~;$8|JdjTR3Z~gV=1;$W?W_q}$&<WGkC!tH{mF&2JrI
zSA*8CguHiN8M32aehT$OY`RNAm!#GqKAn7j`$hkVN6S(uPi#6nAv?Wwh)<{QTh;xA
zy*H1fP-9}#X$fgbtwVfz@K0ZVx##M_eJRxV*mPDxR)^LhKJE3+extZ-+VK?X$=GyW
zLf#XtLwx$?l>EOb)|X}qH99t(k&uzpI>e{H8*z93qRx>!skRBR=>mJfgw`QG{WF*I
z(zG$hGas^zi%sWb=JaeG;?uvRjeq;X?4wsxsIjr>^n~=J)*(JUciOY_s<YeA9-Z8J
z*47^WX?xw}{`sjD<_E+)MdC-0T48-ato9r~t7>xV*=+wYR%gxjA7WkXFi388Onc&V
z4w{^cgC^(ZpvmhzG<g>fP2SBzQ_uxy3N8Vff?I&5sEg1PT_Q9^w+Ky1m!K)RBxp)*
z37WDlLsNFi(3IUWG!<Qers7hdskjwrs=5kI)ulpHb*s?SbPbxCOM|B7)}X2DIy7~c
z4o%&yL(|X=Xc{g9nugnerm36IG+ib%O}7b6OShnDxh!Z}ZVQ@)j>5vWH$g>zM}J5I
z90S@?{5?D}x%IsZFGfFMUxC*!TSM5bm;;2}gE>gp1DG2Kdl)k3Cmiy)8p0uu3lI)@
zT##_c;~EHuJl;=u<nc9xM;;#_Jo5M;;gQET5FUAgp9shkYKVY5AwUG=2|*$tPiP<l
z@<cxoktfy=5qV;Oh{zLzL`0t0Kt$w8ej*`Hsv#2cqyUkSCk2UwJgI?5$dmm<MxI<l
zWaP;KA|p=@5*c}N1Cf!Z_=$o%rG_ZTQvyUmo)RPq@{|UmAW!uZ6?tk6QIV$xh>AQl
zNL1vh4Matr<|i8Rv>KuzPYVzYd0LQY$kQ5#hCJO*bmZwZL`R+;AUg8&AkmShHxM0p
zhMyS7Gir!|JR?91<QYL?AkSzZ2J%clF_CB15EFT3fSAZLgTzFh*+5L>S$<+6&#EC7
z@~i-{kY@#ng*>Z)Sja1K6cx2!po$&EkcK#hKyt;_ZtcwyM@d_XuP{a=x2`PU;^ko{
zwEwqcy=(;}ANw4nO7=I9s@Y|D&GK-7;k+DRI3EWXu95=`SIr@Y_waz>y*yxe9}gJ5
zk_QZ5%_Bze2!Ij10$>E602rZC0E|#AAV&0vfDyeSU__q?7_m|Wj94uqM)F92k-QRM
zB%cHrsZs)rR4pM!_Q-&dy)s~Ap9~neQU;7%Eh9$pD1cGC3SbnU0vM%I0gO_uAV&45
zfKk0FU{s$97`0Lbj9RTCM)PQZ(YzX9G@k|-tx^MwR;?jM_vnDpy*glYpAHzkQU{D)
zts}<p7=SUn24D=I0T`pw0E|&>Ajb5VfHA!$U`(G07_-s@j9G0W#`0KzvAh;wET07!
ztI`6DRc#?=pkrX$8YZ=3kYf;}!H&U@9&<d_mI{`2yHo$CU6(qhR<Mbz*#5@n5;;hG
zA`eMO6d;L-A|xqMf+Q!(kd#CPlA5SO(h@aDdZG@=NHidsi6$f~(P~T1ZmS==F(Bg*
zNv$Z%FMKq9iHv`)T1%2!KTyF{F+#ELQSArKMk)3r<`#;*gt?Yx>uBhQX*PnnnP#Jy
z0}T5*1O0l2{R`$shCPfqWM>c9+xuLE;?R$aQXKkmEfj}-TrJI^9~Y)M^y8Xo4*j?Q
z!=WEn&v59+H8LFfaUna0etd-D(T|T(Jo@o16pwy<EzP4JAEtTq<C|$7{rCXGqaR<-
z@aV@kGCcb6Av=$LLWB~~Pl!?i`Ux$RfPO+PEufzerUmp9nrQ+3ga9L;pHR;T=qEHX
z0{RIdyMTUTgc8wDj8Y={i7k|feqt>xqMsP1Mf4M!X%YRz03)KGSkH*)CpI!7`iUXC
zh<;LplF(0zQWE+}EtG_QQY|f^pA@Dg^pl!t3H_u1BcY#E&q(MeH8K+VNg=z0esYA8
z(NB(2GWy9al#G6IEiI#;9HwRTlbdN7{p0{6qn}*Q$ml0GGBWzfA-jxzN`z9-Pl-|r
z`YA1xf__RZt)QP0rWN#4nrQ|7lmMfkpHj~#=%+L?3i>G_yMlgdgi_H@jZ!N5sV$U>
zerhePqMsV3RrFJvX%+p{0HdOxTF<EHr#3Pw`l%tiihf#z($G(fQX2YcEtH0SS}m=i
zpBAPy^wXMY4gItLqoJQx&uHkUH8L9dX(79YetLw`(NB+3I{N7?l#YISEv=)U9;S8l
z)0=4>{qz8%qn}>S=;)_6GCKO{A-j%#MualZ&xle6`WY>hfqq6UZJ?hKrVaEnnrQ?5
zi~wVxpHa^k=w~!C2KpHxyMcaYgfh|3j8Z20nJtuwer7FgqMsS2P4qLHX%qd-0Ar$`
zS<jg0XErh>`k5iSiGEguve3_pQWpAIEtG|RRxNFzpB1Jp^s}023;nDBW1*i_&sgYZ
zH8K|ZSs}aC?pO5kZQGCER-GqIMQLK`g!ppD-S5@enaY})*eWQyAbUvsewJrcXYTqR
z2iPAhj*<#H#oO`429?};H@SUngvb|1xK2nD9b>4}ik#vcPyA+aYDI2wt|xxWeA$wB
zmN#Cyd@%M4a>nPB$4f8YDakF2uIV1zotBoC8UKV@d&|G@$;;bbiS5oS&zl<8s9SKz
zK4#X=p|Rar9kVjyOSgzM&Sx*#c`LTNOP?-E{D$#i&!8I>d1sqqyN5Budc}|F*k_+b
zW^SI)7~7pYK6gr7@C)7kZHG=-t7E$hyB6lf+wmLcY00fO%-hcZ>tqMUy;3VUCkKgl
z@{k0l07-O;kR+!BNp{MR6sH17b*hjwrv^!P>W~bl0m*clkSwPK$?0&mt@Upu)~%zQ
z^$%Vc8GSPJ_;JV2Lu_^Y;ITmb_~U=Shmu>L1jKj8J|{F0w*B6H?8Gz?4m00Gc+5f*
z5ipBQM8qsL5ec*0L}biL6Hzd$O+>}4H4zQ7-b8fFMiVjGbNoq0r{vZr7;sos?0ooZ
zS8<T|Djt$hB|s9ZL`YJV1WB%vAt_Y~B(+L~q*ZB<^eP>aQDs0ft4v5%mDQH!PJ4dJ
zgC{cc<~}p8?7_z52gCn;mB%-1wn^=q-uBI>+BWHm>zO0=E)CrNO~TWY$|rlDvOV24
zXM1Lr@A>JmIO171&Daw<BqpV#K9ts}V`+NH5H>xzw77FtmyB-N1%>?w3>-B0F}17Y
x?Cs$5{=dDMIk`RZd-i&y4=0GS7JD}EXupT8?n+-**UVv$Poaj6DjH5k{s$JDsx1Hj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.42B3A408DD25BE6F.idx b/.cache/clangd/index/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp.42B3A408DD25BE6F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..16d2c7f352d7026ddc8e252df1c1f26461a0dd3c
GIT binary patch
literal 8092
zcmZvheOOdw!p7%0=L`-XDkO>s25Qdv{_>ng7{3Arlx!4bQd5YesZ?}Lq{LEM8-3GN
zlg+Q!6cTmI1zx2bms}bObfr=mS1)t56x}tyE6Uc&w)P!)ci{fx<+^yTbD#Un@A}Q0
znKRFMW@>Iu&ZZQLCCi&zzTla%*^I?vi3|NNE-RgTg|=AEL|ZK7J5EenFgYnb_U>DU
zCZ9RNr0rXOq?c_^_Vm<Uo@E8;cW>|Ozd!KY%l%Fjz4DLcLt<y%i1Iu6rd!+Es((m)
zjcAsw*N#S7TaSJB{>R^pyB!(v-02Sw+^O4mI(Y5ejrq&`hs*!Hviigl$MLJZejc@M
z_u;`u^WzexH?CRq!iyF6{y9u;u1F7F9eMV^ZEwY*(e|7z%m4h}&)o^KjR$URKX&`d
z8)+T%Sob@FpKqBseTeGI8+yI}{tJVa+v~OMJL#?aY71kdeditsWFJmg@AzGb=-%CF
zy;f9l>iVjLiUXUUU;fdhNdvPMo-eFf@6W6&>ysm$T<Lo0n`QU*om^P_#|6J%{quEx
z>5@ZtHcxoXe<qH4bH`_&wY<tL7<O{3KV#JA?|#{|B;{h(i25=9!z%`DZkTfH!O|_B
zvCiA8UDnzEvs=oTtqeN7cFK&GHw}HDK5x+F>^+NXdKG46{Puj)r+?|SsC=>`r*KG>
z?XM$8I8!24e=uyJ5SW;?`tkiE&Xg{D);l{X_Oo;DFW<COeIXb7v-hriXW9I19VbW4
zh}^&G)EjxZdpff>p89K&=SD2oSy^I6O|3tuz1ipMOIM1^7Ue%3Y`l`V>Y+qt!{7Ez
znr$hZF`j%?elTxdQy)@VbK;}UfhYG*cw?G(lim22<;c-l-{ogy-xzXwYEIif`t<(8
zolUQ#Iagj!Jvi^!@IdhIcLRGIFV4MizIA5#_783yf3P97bY=F&Yu9TccYQ?WP>FAT
z`oxSg*~?a6-8eX|?tI=~%8oz$yW#c4H@C#^xwt>x*hT$%dgo?O=A_1{FD2}qvwUsi
z(bGd;DlDI}HP`j@C)@5eo!B~W?uN`Qx&78mC@eqw#F5`djeD!;vCn@<59HV0+m<_Z
z>i(A2v_lWgX?Sozrt(MUCuzII!5e-Um=^JZlpDQf&+$V?`_=l#U3GWPwIvTOy%{&F
z^V`dF7P^LQ{%*;R@9rk}UcUR@QpfQvmxfOs^UboW^KReMPCHx~<VyC1&Q&{~s`)gk
zzV_}vPY)`pSt}OIU31PzZJKW1ShMHVq1iwEc>R-n&$0T3?<-&3RoeH_ocH!6zSDZ;
z;wMqRCJ$bB_s7gxPu`sQs3Q<RamSRDRVlZY$7UzqvwpTddAsxQsWZ;qef~9}wf&xb
z;;p4;?=HUi<L9TE-saA&pLIZ8_v5+GXCA&%I^vxBtoIbZ?Xmw_urc!GS%DQ_Zdws}
zP5ofbp@LHjJyYK<Qs<og_n#ge^~%y)#e1jx>F};qulT09eOEp_b77(OcH6qYy>D;M
z>1Z#i+Oy=TU&@O|*Zy?v`m7mu#wE`1C4BXTuyn}q3&D0~op$k^+^fYe=Z`+}^@d;a
z?=5P5=tBIJnBCtV{nc7~Y`kmn(gnp1VzES|y#IPu&B)Yf>hT$tqJ9zmBD;pr)u8Jy
z>fX+azKIX#M^m0qofwfA*ENLexczGey|C}*;b<x^RL4fdMs^M1dR(XfdDn^NInmS;
zp*kfZ<?*heyZ-sxul8PDx<8tl7OE2>5_)wF;X2ECd$=&a@W~j<^iZ8<O`F~|gzHzM
zo9fTIz8Tcpk{_y*2PCI<4dJ@x<I1nMf4{J}w`FRmP8yIjq-zM*H)p5*SvJ2iqN&ND
zIzA#kvTF#}zr;>^@4~_(SEH#Zp*k)iF0yL~*QL)qySO5;=haJz>U!1IxBk4d{_?2Q
zm?FDoH^R?8l-%~bp2TV_bgY|&j?*~kI5!6!ukq0FZXP;86QC2^0(7D#LMOUK=p;>o
zPI61o$(jtE?3STZGzB`vtw5)0Ds-w_g-+8n=rp$mov!K7>24i5Lo=W=+y-=}W<qDW
zP3Y1c>FM3?qoUD{(cLNh&VDSa>m3cJ&PVL8lb+LMuO#;En5&3=FXmceKZv=J*jpj9
zKEfi8tt2e+*eb#zkF6yv^4LbgB9HSC4tZQ9;gH8w5e|7=E#Z*IH4+YaypQn6<0}b|
zJidzX$m44Xk37DS@W>N<L_nTUNd)8xRYX9ZP)h{l35`TRp6DYY^2ACaB2TO$BJ#vq
zA|g+0BqH)8ACZtJRT2q#QWcSqC)E-Oc~T>hkSF_yj6AuL$jFnch>SeBmdMDH8;OiO
z#YYt6DV0P)o>D~=<SDg8L7vh`6y&Ksq9RYNBr5XMDxxA!ttBe*)JCErPxBECd0HjW
zkf&7<4S8BE(U7M#5)FB}kLbwLD~XOgy^83_(`$*2JiU?V$TNJzK%P-a4CEPA#6X@=
zOAO>0jl@8n=_4lc%t~S+&#WRQ^2}OdBF}6jCh{^I85up7q)bO9q%21kBv)w4(sN|C
zBfC3=zjr4_b$#E$#mi%#4wtS1dl4kB{W(Y__TNCNu&=;tmWKrlTfhQ_^|F9rOIX0L
z6)a*n4+j{ofCCKY<p9H#aDd?|IK=QC9x!|X4;bFd1BNf*0mE1Dh!H#jV1xnzFoIVA
zj8Gx~MyL=FBYH%@hy@~GM6U=Ku|x!nSRo=t@<@P@3M9ZtUI{Q#i3Av_LPCt}kpUwY
z$bgZ(GGOEq88C8%j2OkE07fZL0Hb&nz$hgOV3Z04F{(!ej9Q=qM)j(IQA<?7s1+(=
zG>--ttv~~e=G6eBm1uy`Dm29C9vv`xfeskms{=+a(E+1Z=!h{q24IW=12Bfy0E|&$
z0LG{=5Mz2wz?cOlU`($G7_-C#j9FnKW{hJ@_Z%jsXslx_q;ZaMkREY7(w)*)4D4+9
z!|qFcLX%OOUFiAXvo;nIXX7C8HXf2-6CjB;5t3w+Ajvivl44UJsWugoX44?)HXV{-
zGa#8ZvpYElu6^Xj==dfQQ<R>X{&4sb8Gce!RM%Zb&qXoUmOH)sp#R)+tS3=jHyb@i
z*_$c*hg296dYyul{UqiN%6<uR9c{0t;rIY;Z^qnC+k=>^8T(rd9N)m$-^JX**jq9C
zt@eY~p5xhOibX#*NU`Y0c2F$(v2`?yer$kd(T{DXS@dJ885aH628KmHwuNEQkM&zw
z^y8W-4*j?w#i1Y9L2>BE)zKXKaRHh`KdznT(2uKTIP~Kh7!Lio7KTGV&Tr+=k8h@U
z^y7mRkA8dy#iJizNAu{%2WTGs_;#8{Kfapb(T{Inc=Y327#{t2zm-Qnp_vlUPY6;1
z`UxGBfPO+9Eufzepat|3+Gzp(gla}WKcRsU&`)S#1oRX9RssFQW=cdqF-VE%Cw5RG
z`iXV4h<;*#7ST^^r$zJ=s~Hjf#0Ex0Ke2@o(NFYSMf8)JDGB|gASI!n)Imw;C)Lps
z`bhy=LO-dUme5bCW+e2J8W;)vq!va(Kgn;E&`)lrWb~7Rl#G6I2PLDQTu003CkJR5
z{p5C9MnAclk<m|XU}W@@TNoMrWWQBLKc$&c&`$|c3i>G>l!AUr9j&0B5}+0IQ`%_-
z{gi4(K|iH|QP59mVHEUJ{8k11)MiRWKQ&0H=%;p2D*CB)w2FRefL75@ZKqZAQ>z&j
z{nQ3VML)HLQPEHJTUGSankfzav>>ISpVmQX=%>}u8v1DgT0=jroz~D#t7bIx(;65J
z{j?TFLqE-L)zD9GrgZevgOrYbdIzPWpI%4n=%)v09sTrnT1P*<n$giuZ(wxv(_0uF
z{dB)oM?a&PGSJTmQU>}N9h8B7MjdUSpAn!9^fTIN1O1F@#y~%#ficj}XkiTWGyGNq
z{mf>{L_ae~ndoPBP$v4Bb+n0oW`H))&uphn^fRj&6aCBv#za4}g)!03^jpmyzl_&z
zTYmb!>^xzLaJmzo_Jn3T_ZLvZvmN)yHx>?%*HwmQP>E@YS>d%X&!h*~>%Sjuy+5&v
zOz#_h@QV{&$Hf_rJSwIrDKp6vURjMPO3qC7gqLAgEPsDx%caYQLXS<FmQ)x%C|uJ~
zU8}QUftVuWV`7|wTTb(1V)o?MI<u40gX;!`XtA-e1H$*aVQTNFuEpFQn!SzKdmbHA
z#I_L@GuK8q%zPW+F$--(z$~^A5wp}rB+POfkufW6M8T}K5f!u6Ml{TN8_|1m-7VkJ
zCob=JJ@jl-3R8;1bH4iWCTre;T{)rM34Ia<jQh{jwb`8NwPx?S724fzct1J3g5T;H
zd!s1jY+GpecxL?2@X0&+#TU)<woY#e?M|MSJUe_ozSOMW_v)LlHncmve|k#z=<u?5
zXhHk_nmD|`PWza!S7>hRvXAdctdoU~b+OQKP7XTG#X-kAdFXf-51rr?pc7mIbfQy)
zPIQUTNlpnm$t6K2J7wr(mkgcaRG?E_3UsPdg-&&;&}mK$I?bg)r#p4%be9gD;WVH#
zTn2Qe(}d1+nb0{M&hDGKn4(dRQQhg5t$qXLY`FiKjvpSch0kR8myEtqUB6(!87;Ff
zfpb{KLgLCeNPHO&NhlK_iDe=rsZ4?-m&uTnG6j-arb5!nG)Q`x4#_ApAem)mcPf2m
z{_Oj|$Sf{>dU5{!jam1H|NAQpZ&)lddo~MtHlOU?q>JQ_-mM;Ab2~5MshNed3ZAq)
z)!kX1Uf`WSCln_>>!ulNL@!%pbWHEqzI}4zva{@QQMs89B=n0Pn3$G6YV??~;~r7^
zi_T#VF6IB*J0K}}aO#ku4-IE|K~h6Mc04@tL35Bi!qtDk_(x|`Ig>IbkjMTPC5;1z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_multi_d.hpp.65CBFD3787EC8604.idx b/.cache/clangd/index/device_batched_gemm_multi_d.hpp.65CBFD3787EC8604.idx
new file mode 100755
index 0000000000000000000000000000000000000000..81cd5060a5fa67cba6695cf079162b8c73e37305
GIT binary patch
literal 1174
zcmWIYbaU%tVPJ4h@vO*AElC7283cj2xTGjEkCB1F6(~?KXYa(=>_Y|;t-pi!6qN7s
z`7Y~nVQbI9J7o`RJ74ZLGl^HpbWQnRyLEQ<E3G&lwpO$Ap65<hOZoh|Em$VaJBj1P
z*?A}3MBgP}c%6TocV6|6O$Qf#{SkUtnLobznPOo|<l3x<$2RA4Zm)Cj-uYQ6>d;<s
zp+W(LyR#xIt|oLOdtICDCLZ_bV{Du>f8v#$`BFZ7($Zh}PJcSFPo3X&{|SHVeWC2x
zRv+J5MWhSw`^owD#FxtoEt3Lg2)i45$T@bn9+c&7W<8mhb}6a+iLc?y16SU7PSaxF
z?5^L+F<oJ;5#w3IvajiDzWtL>-)etM{fYAR73cOE^GPU9(PY%vtfDw=s>f>688bW%
znNN#!nR?;WB+>RAyXLU;%~-O>gW;l#?6O56;xDv+{1#o`(fZlv^?$F`$9pbkor`$G
z);l?D*CoE0JWHqW%?-NKB;9xHgvQf=MBm9}X7j@NSD7qZW9)E5UdL;}x?jEX-)!&8
zEw0Q>ssn~T_doMVt9`@0*qE5v8Mq|bCFQ^b11A$ON*EZq82u8ePaN~_wH07u<p7ER
z0hfuUi8TY1lU$@5WX}3!x&RXkOj?0OK^!K%Q<yWv=VXGHC=(k81DCX{v^to8Nf)K2
z6{|BLT*1f%w%t>utb5`v=Q<~@#o2O>{CgN#xMY=ORbg(0If;=I>=c-A>~_nX7&R4T
zMjkGC8F>wu%V0E6HOOT!)v0}~strQLPC(U?8j>b3;fli#pH#72Jj=|(rKGF`H-v{v
zOIZtvr>Cq3=XHBAx4iBDZ^6sMB`qo~1M@Numx_dnJj`|;E>#ItB%Ydt8WK-kLLG^x
zA)$f9)0EIe;%Vt>nIO#6me59$(~;0Y;^|80BJuQe^^o-GOX$PpCim^Kes37Zjuyyk
zjjaNX&P|vO3?CC26Lnaa!+giW1&&BqumMBMOvwS^FfB7JB%Yp`9-Mdg-j;=rG!N*Y
zS@+a{^$R<{csCO;@&x6j5hj@knn=SeXDCX|Nd%P{u*icaUxwnM<YIkL>f{6w5WpoS
X6?6ReofGEF+&r8NJTPHkaziiy6ik0;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_multi_d_xdl.hpp.5F31589668DB0CDD.idx b/.cache/clangd/index/device_batched_gemm_multi_d_xdl.hpp.5F31589668DB0CDD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cf9895f7830edcfec836354f11d942e3e0ece2c8
GIT binary patch
literal 15014
zcmc&)XIK=;)}F5FhGA%j8Nx8+oJUX?kfex+VjDnE5p%+fu9(v*2reiJq6C$o7%+<>
z1{4t#1W^<a45%wACIrJOh>Cv2&eV3_U*GfG{c(AA_sqL>x~i*AKGnfP1`W#c<njDw
z3|SEozGN!N<MG7k-^EKJ=h&<9c==X5-inN}agzg|drEFE_K7uVWAfexJ#>a46Sv&0
zb=;f|3iDRKhsn*>3wB3n4nFJGH0jpzKc{;QX_e2Za&~br+jM(f_Tn`^hHkhwAkf9Q
z*5g~|cKHFrTlv@0%!Z0~?{hF#X$iXHpnrWs$-=N`OTUa{+aKF=#M^!*tAuWU^Lalt
za|YucUt_hoZdmQ=%J(l9`!+wGpq^6|$B%lE9Pq+GT#z$+5b1BHuBBGJbBy>#N8Xk}
z5liBynXmj>J7DvNoz@pr?tA(EnqpX^tC=f~xVK_0uVB5(!_JyhEp?H8C38PN98el(
z@LWE7Uzl~xF7s^5I8E`OlTU}$S_auonG&-v?$YUDo3uXjtvyoo0<Ww;b^Oar?V6y?
z{SNfhwkFn=X!b4`C>|<tx5?bo;vooM;5X_)gpO*Hw$(86ZMS~<tv#F*<n?CamT|U%
z=RumI<x@kaY4=`NzuSAGI5YXoX+zp^$ctIxZA-50pYgFEQYAil1{EgI(mqmT8r_g^
zQj!sFXFlYP{f4_wuFni*f{t6=ce4`q8Qsq}FuFFiZrF;Y`X>uNct!oWD0*^q+yUUz
zv0`$^gPK!M`18|q*2VV<JCq?`n);TGa@MN#&T(pzB@F1CYyKv}&2-7=Pv?V99(b-Z
zyE-f(`nP_$8P>J(jFZnU@0E|<8Rs{8Svc+8|6Dtx7ZD^Iy=cc`-)Wm9dAB8kLmyVA
zGl$woYM(fCdU|I1^RsU^44Yjf(qET3+ka@_9*a8{f}=JFEnB6|da;#)aca9RYu%Z}
zC%(&8wyCbTEF2|Sq$#`CZ&Fw1w`VQIp*B@=e+NO4UQ;YnlrqnImUo}^S$oc<EG5sR
zE{Z9CSQhTj%aQDEjV=rfnIlZdUX<eJE+gRH#*2doYGv;q&}Uch+~&?u*Vq$_H2PVp
zeJebBw|&u*&*dQn$HnpHXEzqk0IT%NNAU`nkmiU%&AUqZhnJjrQrWEfxNzCMvv(}U
z2d!Oo@3*++@+lY7wK9%GKU}+Y>kidFoNveMI`P|`vGeAo+7tUmj(MCO<ok_Ib_h--
z=XxF4969}4-iqR%pF)?LP7c|1=rmpY_RZ(xTYm36dd@6S7P-#PSbg*FXZHrmJ`Jw=
znjY|p?{MqujGfD+uTnR@_d2$@N6V+FJ8w1J_xy2t*?eK{>bbG6BHy(Jn(O=7969r6
zNx{L~ecMjw*T}-HUPXG(N%b-~WZrlGU5(LpuBtkg7M~UcCH!oZjXAb^^@vy6_CHN4
z%@eyq?DiXcBMM7<RSzxyoE3RXp0IgKSlY)6Gak7oJegA|eX^v-;fssLC#~9g)u()N
zj|=b5J5<gxc;gezcQ4*POKaurhs#sK3vLFfeytc{K85Gw*c3pfi47;;D6XAhmIgEy
zj`*mZ);rr|zxRv&6YqE#2QB#Qxa+g=d5x|iv3W6lhCE8{>hq{LA@S<{jGVfem8#hT
zW3E<j*|hDwTSa2*v5K_uZ5=8u<vA0Y=9yYMC2v366qf1vFuO|p;*&(TtS#GU|D}|w
zpVzI6c1>tM{^ZH_s=sDdKAbQ~UUBcm=|`VtZLiw4y=uHmr(1#SU~+x^hizJ0ep^2D
z>)EjLuWoAD+-G7B?uxIf+q&>CjaR+0BdVv@Pg?rQpv);{XzAtjBYn#5WHwiaFIu>u
z^>|}x&DRwb@}Em8diC>}FKK)=zI8qwbi~=~*u3yjo6Zj|y*#-*N93IJ?#jdax;6Wk
zEZ%eCvF(=?(_5k=&4Ly=y>a|yBM#BD?sM>g-q7N;9<`yGSKQ2wo9`>0PQM;dGA=%N
zSl)nrW%XwUub8Qpt#PBBH^NlbP&TD@%KCA^*^3wYFK<zG<{b_*-#lKXk$XZcD!H4~
zd+26|>WX{qy1ODOJ0d$bj`91w|G6dK{D(EJ{bQK@DbW7m%7&%~x#K5fuAI`C88`Dp
zaIn2j%IekgaxO{BXM5K@lAaC!$$UzT4w`Vc{IS79Y0bIwC1sgeHbqT(PJcAJrx@;@
zMSZ=J|H=JIbCylrTEkxbR;AmoNo*SVwDDBOJRQq(BXvy1?*RJa_g6{|kk`&0*55^K
z=s8!1@Om;YZo6mI(xSP8^Z)qsPHV&&_~dNOh$$DRrX?-&v;ThG`SkF@2i7IqPScH>
z4NnxR7)3ALIX>9!<GSbre;j^bdSv9)*uaJp22pjwodVJ9<YTH?ANo&*9<5tf!Ta9B
zj+^;i<Q=RRsLq&ea%g_?u|Kx2+-ToFa{6?em`9SMRAQUjTHX0)yp09Y#sJ@qs$c2{
zN=jlUowz*K@2<+W5tHw2%{MgoQGD=$W6bQvKQAX|PQ3J`t>E11ufyjqJoM&hu8JtV
z<E7-s_sAup5A*hIFr92rzrM)rVUNEAjjp_tOD^g@)JgvCJJD>VG{eEdovuikYN*mA
z_-G=0A{gu=Jex9St=ANX?Ppbw7_D|4oi{sV&GR0e<{=0EnrlC$W`Xah>*1FN8Bbg7
z8aZZqlgy@5&@bb+qBYO`kFLI0e7gN+=$nR)OLem3Ch65Xw^uGdeD2NWt}l@>ACGrz
z3~oHVXQh909zFi}plZR^^F<l%4h2gKZqI0X(6N5`+P4#z7qw(9f823m`>eA$k1k3y
z!ge`3&IwA5V8$#N+m{ZK4PPGf$-mzMdBmuZmu~5tOVoWA+}paxXY&Cq?@KjI8$VhM
zC!as*J3nMqad`Z!K=XGG7Dv4<=|5&-8l4YTy9}-})^O=-GLxyYwi$PHl`QX-MpobD
z^N9#oX@HT9?5_uI()8s6H<TE+Z+*EYd%?(pj<5%7&pess<#f)u*85c6tF4m{C3KvP
zO6WVz^~2efEw}dm*f1hDe@5?C8KoA-&-St#AF_4GfSgeArmYX5)(n-*X_GpamR?Id
zy?)sbo!NqTm&1Q<ku1~*sSf_K@x$uf8V?flb`_T=1f+hgdjG=o{qAQ)4ogO)T>Bch
zBsNVRowIIoV)NeDvRawb+8UzM$2)!b)l;5&?qAiO6nHhQIr%~B(nF`}7Pk|Vw*HFg
zb@SJmr#r}!K)a>YF2C|VS#%waKe*1$bV|4zrS<Zc)|#ot^NT~Y%VzNQrd$2CBEWT@
zEXC;ZJK36F8)62VOIu6|i-fa=EpsmEbM|M0K}n$M3bO>m5kr0mwam{YX~=TxP1UpN
zc#}TfjV*ds!gP(E&}-7ir!hsdBq~qV5%+FgqZhA?n5Kd<CZ121>b#}#1tOsD&0~%M
zmt&y6lJNLwxcje!@~t1VTPOlDPaboQGCN1PC<$f(!&aIN%{FL}08KPY2H<A^YbE)I
zW$cSjVb*_n3PFD_9@7fxR_LW9m<3MQdQo%iq%k7k)R)JUk-{=ktR$F)+-@>(=7b@7
zLZFTox&lO3fRU157PYx_*8)R}M#U^qq%aE2q9B+>)Ui?3(l4x^5cr~Zc?Bh}ptX`<
z7URai#@bmML)C#4Z8{wY(t)m$U=|ZCb(JyW4hn?82hH*l>b``wN`hIe3|<MACheFe
z0!C<-B1%+587T>7v2!}_v}tL0ga{a*9i0LWr+|l&V3uAgA2u&vG&ooYG|)m#P}Bq^
zN`hIOJue?v`uItz3a~&6bwTwmsHr5F#pQCi_UM<Q6GG4n&C&slJD`J-U>4Wr;eiWG
zjp9`RAI%a6)#IR+l3<p8VvnWIG>dbDK!{%VDiB>o0#Fdl;=ktFw(T3{zSIF`XqGBM
zzltza63j9vW=><*UinTDFhsKy0+&Lhbqa!67@z8^$-RChsRLCsO9~LAAOR@IKP>z5
z8Y-s+%+dj-h~);Me*-O~Aed#?fr&8zlY%USpdZ@wM`-gA`Y8!!88PxqyWvC8cp-2_
zEX`2TjFdq^FiXh7*U=Gb79)kAH)44U4c|g%CBZDCY|oe9@6v8l0h;Jtb^-NWz)ng2
zVd<b>`D*zF(!7x#JjRdjXY%)dcxn_RLi-R5G1_wOrd4rA7N`RiG)FrSw4<|C5Io2D
zWd-Y)Lf!A0zyYzuQ=)jvKuIvm<T0T=2f@fr3E-ob-30lYph!tD%j~Q}qQvOG8r6Xq
z%~B2o<v>qKFw30x2G3RVedh~-Gn(ZEl)ONvq9B-M?g>xtyN-+dh=2{6WeY`ap>&l5
zv&@gl;$Pj|v0DXbp)Fq^R4)+bO2Shk0rD+q+Y3H_`QB^HRU37ni{|+S1iyf*lHhq3
zdd~=OI{Sx22;9)ye1^K8p`DWa!*XxgbZ;|@;X)usEG^Km1<6Q3Fw0^SLAa6e5gQ#~
zf)=VI^ed6RC<ta*R`YnzjVVv3Y64rt@|+MoCv21ivn=n*9o01|sZt2WpoN}5ttZe>
zNifSQ`^He7SMXjTppkl)08t4Nq=I0Ub$&fxgvtdUG=UXbsD=>LpoJ6!v&65N_-ej%
zbDa>lqlG%5X(x1263mhi6kv2eqQOWA1|XI;C~t$)lmxS+4n6Vk;a;to>Oc)G)CmNg
z=v@>9vt*fAF7FtxW2Fv^5laOSQ~-&RV3s{@-VawCzfh<N><~*F5XAvoCBZEF=7f(|
zzqR<A5cEKsE(Nku;He~-W&ee#ekohR%hiDnn&k%&`~YrBf?0Cm%?zEny5}{4HDY;6
zik^~&N`hGqExuSiecJFMAqYX6ZiIS`&|FC{OJ2>m*)K!4AhREhSRTWkk6|Ar!7Qhu
zPk#=Xn%N=*qY%p@sP_oXq9B;%bYq5^XiKP(Ca_2E(oTrlk&F}svz+(PPPF+@a7`WX
zQONoQg};z%RuIfmXqWz=DNTQz2somJ?!X>*kZV>D%yRu>=i9&!$*$@^gcgc{f*2@O
z63kK&;)Qw_CcSh(PsDPaFuIQ3ML{r2l}k;Tr%{WF4lqY7HwpclNWB#Vv)nDWh}yPw
z%xz6z+08<VZju&Cf?4WKs&#Us>}ho%K`d!NkOrhmf?1yTT6FNsm?P7Lzy<B-D>VL!
zB8h@vmN&ioMy;+f@KOibXqGZUP)2Af31(>-<H+2Q+92;8j997({c5x$1;H$BXZ@Rp
z?x{nkD?lu%K#+=JfP!F_FZt7dMjF1BiT>VEBdOMi^hH51%eTQ@EBq&a93%pINHKGP
zOD+mF3W8Z8!)GoYVuRd~G8-ZU${Yy0b({mC0Yw6YJsaq$uK96W(KUi*z5&N?AQ;;l
zPKnSHihKx91(b;pq?sd7cm(#v4x47mp{N}CU@TCkM>s0u^a#zU$c=E&$4L>IaSssg
z0S0(&nmG(y4}$^t1O&=d3BL+XmC%fe1PRZm{<BZ=(cz&C+SLU@c7gE2+wIXSZr9wD
z-i;KE9H&VoVTFAvjRLR94Z@01#$*~re3L4|3MT_J(nOP+gcVj&nn?z9GSI}1rN=D&
zK68#5ULi>421VbXggABTQE;f(min!Tq>;?{xj@vNk#_&lNJRW&Tzm@fPjPV};1{x3
zLwa{WuevAmNRpZFHowm(Wi%$41zrohS+vk=AqOq;TJ#UXS)&Tg9Du?D&=?DfW{$%l
z$6*+YuR@2b(2K=4q4XxSWpM+vZGi6BGYFKqCl<7B>H5_CC+S^$lBtDOwa^ALc4LoP
z=*8kXXjR7@uY(?S-0^zoUJpH4{(2Zx4})3U0IeFh&o@Ai2KIUBy;J-jx#yhGk!>LZ
zw~#?tVMr#4^hhE-SzJNtR*+Jxq9pT}w0KNfV(Zk?@UHBC@Jbodr!bGO0DRi|b!Cfq
zZ@t6+=NrY@J}FS<_*ma*CJNF~P!H!QG!qL0V&Q1)v1uea{wt`1b;Cp$<q^-n*@QM4
zPa4FNw)li_v4E6kKNRkVCaf}CgrbYkjK!tUp%i+vxB^NmpdE|rp;tW|hGi&FmK<1l
zOy1a>Ra^Gd2fcw0&qtuJ6&}y5Fmm?0_1+tieZ`ZKcv26)n7PlH3!mKY2>!K<vB4r|
zq9}Cm8R7~D4JgVR?0#@v188P16z+xEc#i^Qb%XbW-4YUyqO`$Nqf@4A;weX*79)2>
znO>l5u)FTYb{D8#7~>iUt#|?oPe229MMXh`SHx6Rpsat`6%`pZW}s2Dcl!nV;Y0xW
zInw?n>5L;SjVz(hW6}p(0?m9TH9nKNIL|>_r%d-!wm6w~edqRIe&vTCBtBo#R|kvF
z)A-|!c`8#s_ULv2l3FOm;wPCpD5>M(dT3h@?b-1LC~4r1mr`^o>t0FbDn(!A;%k)N
zHOc@>lVr*$y)rH?r}WCX_&TL`or^0dtqMwqeZG>?s^sD-imqbU!}TPJUHgD=A2)Vg
z0<M?909+~xlvO5H9^aKZfwGHfg-Dc-p~+(?!#N+#q?3N>WKi`M%^Sa@zmFPNnfi9r
zHw!n!c!{*SM7m+kr<p6H)fE;Gtm+?k%~$S?j()7oSUGl5ID!FzvWR6>A&H!lL{7u=
z;WtHkt|(>ks}8Qq@Y+4je=$un4TQXb=!cVknz={F?{RS>A#dd3M}+(l7r!RtuZe!R
zN}-uIg!~N`w-WMJE^Z^_ZCw1Fkbfuo;l!3^eh~5>T)ZC0*8|RY6p%-;7}w-bpsd8%
zqxcHwuRsr1urw+a`7fXjt|D<(fXqozpR*6%BLeOb40|Gv2%krUFN@z0K5qzLcE@c*
zKpVlZ_y-a2gJ9SlMFF2E;EOS?|B>*ILE$lI&l<d<8G!A0cwish3jg$xNceTYq7GQ&
zq6oJKP{y+l3iolN{$c2U7>>p>(o6|-DB(u^I%r$Rjq-}70$wph&A%ujT(lLfSVXB6
zQ5x7}aJK=y^Fb&)2rcmn0%gwu&w%v{xkW|8fj!m#OgU-P^5B1j*7!7#(;;PhNd}t;
zR&}J}-ED~>mdiG8EA;!FW>QJ*RMG*nVd*0AE4mWwdT$~BEi}MyPBWRLTP8UKKR9M;
zO4zZzjh<+$)s$8>CBa)At98BiloL`b6crl?-3G#im3t$h+sMVQ3EkI(3oG|lLbsKR
zzZ1IO2^TDPnpqEY*MooX*zf7RUtZiG|F4rj9AS`gE4nUt?-S2?%%5OwX@T<7IA|0H
z?XYxcl%-3{NK2d((Wor3Xe6E3NRU8jCs3x?^J6Ka75+2CO#8j3+QS_$#OU<;ll`SQ
zB;Y0v5}KlegWm|H-hf{Pbg>4{u_)Cx?IFw+(#&|lcnKCZ%}mgqV9KJ2f{C25N!pV*
zW0M7wIb&0_r*Ot{fpso$#wA8~$zgLCxUl#bus#Nyab<-9FR(cQTv(hBbn}4)P7-LQ
z0JszY9~Kt^>q6j+ohx$Uz~&Ng!5DiOWS%)tn8P*CT<D+6HP0*1;R@G0YoTo|*E|*7
zEi67P79<WuYYQJQrpO8u4K3DO$_Sk@q8F}xyHf+bUr>%y1Dr#m2Ng{&JQIytHvBf|
zg1b4Wj3eFlkW;bw;O-ZSk;kF%I5fqoLNnK(!!@>o6exRQcn=GeJu&1-4ng4|t|w7+
z#xUb=?(Y|c)Gg6PD{KW$TS32_CK|_nxb>VyGiPA$Gq8WNX_(iqQ_|B&Jg|2Gi*fmf
zlu*$~WA~Iqx+al*aB#$ii#D(5s<Dizz%mth;ab1j&D!mQ{qY`gyAADO4;1c!x_CgK
zY`Wn+thn0rVij3)O9L4qkjDWJTw0<e5y<1YcoUFs;^HhI%L2XGsyQEs^MQm_!BWtp
z6lmfG7tJI<NdlWY(##HMxPzO)r$fVZE`CPZJ|lawDc@Vt_AM8;k+yAIoJd(DvMC|W
zq)`@WT)dmI*v(?xZA8M`3xs>QDbxwzdIAi<wkA;aAaUGS6XN1ipXcF<9Lo(#;|67q
z-7@yVNLY%NBrB{GV3`8Ea3aucbapwcu%Zit9X35W_dA-=*U5JPc2zO^)~y<TsO&nD
z;ax)SE@6%rpqU4R@dLsJW8CsYjCoL)2aQ-^EBc>U*qa#`)0+6peGttIwjSJ*^<trd
zP%(?b1YsOBTW7W*J9eGWxlS13z8cNkCQNP<*0?LxjcsoePF&pkHsQ|V8p5`QaKdj+
zqmy#0A^LLhz#3vGi=Pv=&j}}N2Q)GRx93D(E*|)t7|P;y!nU1oVgp(`;nq&{<>G<u
z#84K;0na$#%T;>McrM-qJU4N12C&Hhj;zvW0-H=O&H`>(pf5X~4Q#Tx<GH{$7X)zq
zyzem(fOU;VwVvlG;L93OA@D5(0oZS&<Qwp>01e!`LfI4GmjDf{%rt5p>y`pz*66MR
z{#BrX>jmVfpm-~kVg;d@L?}+=;v^_e;^J*kyp4;Kp*Wd~w?pxEF5UsHcfg+P{3%eJ
z!W~bA;#4k9gW@zUPKVa%+~;>f@lNh|6XZ8R4O{~wr$TBbuw^;Qph=?_q$TS#nn|N(
zE`CWGz2xFoq|qxbeoY#^=HfS`(Hky)OFF(KUD^3tNTU|+_&d_*9T&HfMy*`jMmo2V
z9_;h)Nu&4N@okiJ8)bsSDC+l7(qt~)PD!_O@eWG5gNsurX$lvoQqoi|PNRCJQF3<v
zbV{1e9p6bwcXIJAO1g`ScT+ugbDz(kq#5iuj>X8S9)-fA+<wdrD7pdrVZ{(A?_02E
zN_?==y$A11YqtxO`1ZsuFgc>=S9s0DzUY77Ciu4rZR|*P+!DwA(eu?&G*C^7s!3Da
zIY?=;ztaEm{<mnLn35J#me@O|bWE8TtzlS?GR`QnXB6pyV{A7b8byY(xSW)hb2HF#
z(yg2vfHP9uszz>JaUsOI`4-4;fd)8#pqVVvEsM?PQ!4{<>o)!_LvDI6rM;IuweEh4
z_65oW_ggYD9iFWzUE1frPP;S1V>Ye{`BI4vVLfzQ52xcKDsygSL~&@e2^#1ECS71O
zzFWdsucknGg~e{F1DJLI8Lncw8}_E(fDE5P_eA0@V1y^?UeR<tl;IV-TYsjpP{y|Y
za-n7}H{m=CH4k&~acFj&>jADnjVs)=;X2g5&Q0d3pmr4(KZfGR(1`78Y#>D&NJCb-
zsiZ8GbjMj@ci+P`o%CgKA*oSF8nFCzq(&WSfb}bDz@Og6BZv0*pEqA|$U-@!;$o4N
zSQ>Ci14HqZRL=a%O!7yaNF?(*(x{HK$6LWc0vT%#5as|wHfj|B-vSVbqmV#(A&E1f
z{eF!`14_T`{(lu2D+MwY#kD2-j_)A<9hBo+9h%7|-LlDPSPIxbqB25pnTa37x0q<@
zOeoCc=Cz6|P3%yPyLbkbWll*$j9t*M3)<jrD2=Y7w9m1|h7|>!LN*j;b7ReUC^`@I
zS(o+r?;BCP!l?=OopTa2G?7E8h2mPMgPVuwoT0dmi|e69J$qk(s&*)DV8`*DDtgr{
zD9qySNpZ7^SHxKhg*pj;U#sF%p_wFLmIU0fP9UWQmg&F?H|V>sUhEFA2`4HCq46c?
zinn(9Idk5*e4h(akyz*$3kPFsEofO)pu6eQk$>IhKI?WmeRa*d0Mtk?rnHLLUIs4n
zkcucSfLS?3k<uvAlr_W((xHNM!EX5c%%?X83T`z0eZwWr1qL`~bl-4^k8$xSAU?&#
zg+N@$;zA++PWjju!~f?Njr9j9z2c%9%l)FanfOIq`xdm-dE`7Hn~a|U^cil&l@IBB
zD8W)iWdo$|LJ6Mt(qy&37v&-ONXivJuL5{sr+=xUKJ2*HU(Nqo%9rKAGbsO@7Gz6z
z3HQ51e>RVQNw~e_wp2R_w@$W&TFk_ox8C}#6z%XOf&K|(6Z#e+xP=I1aVIgjlNg3g
zfo6UZfj@~+7+*aUw4lcze$4;tny$6a9DOqK${utW=LyU6gb%i*>$NpiRc}9E`LBcT
z^*6iLXTEl;Q6ZUY6n%|t+IL?H&}Cd)PSNFDe4V1Nb8!VlS8#D9MOU&I8!0j!#m5hL
z3+T24@;^ff?tmcwPRepfPwZWB^`<RQeha}q7zg~~z;M=&!~?&0F5U$EHgRz>&`JhI
zEPod8$pS;zK#~uv@_`c$B&aU}^v?iG7MB9kQeclGQFo8PE&;l*8oUEK?SLLEPKQqE
zT$~T}^VtSXx5n$=g_c<3Y33Q}^^8OfEyd?CH1n41^Op2waU1E?#vM<j+!Cq&Y<iqV
zxutROZpv*pi}AHQvRK6zBlsYzA`g%J{ybvw|J<kC6b-!hUE`hS|NKR&Emp@x_Y2YZ
zJ6ihILVs39b<nqti|gUodN_d{Z-BlH?D%bO>oviFhN=JBS=|YpEB(4GvpW}_LCz53
z@KMvSG}OVge4Sc%Q_plOaAcc$X}};2*x|g2W;y`B14y`;G5;Hou$l1>;PC?t#z|0j
zX3Sp?CEU!I9}6YejJunL=4W6pZqv}bfW<g%>sGeEKS9FEhF=Kz0mX+$cmO#^K>q~#
z>>Ip<{FhLIZ`JFa-Zkcrp0XXCKmyQ6VD0$+=ZxUl>6`xeuYF*X`)$9DV^ZQ@3kiQT
zh5Y$RC_Ks50RrWxQ#eMxkRFgFz0`0-jJc#^F6qqX4Tng_Lo9CYMSWJ)(A|s}6DYL=
zN{co2bClXSN()zEuVfdFyBHnKL8o(o)IC7<#53YfwE%rl^!F!P?0F{v-6UYby2Bi3
zk^^m6_q7k2?Bn7C(BuFYUxM0~pdrhjNy;)wIriJ#1*I&Di?c~tHW%+9WqVlM!atS>
zJoA!JIDSC%eLw`T;kbqHY9R)(QS~R`^^+KcqbRPj(G{8E(=v94PYI2ugdsc8Cqm;B
zVaQGt3p8SZA)cu9`-HE<cW><TcU;$dNtm*XErf9kVa4K3La&oB#SeDJb>p9e6~?&F
zjdWb`B^&!j-vRv{=&@GU4EfDa2VX#b%D7PMlxJr7uQd^0Kq2!}d?<(C;A@`QIce;L
z?tX9waO?oV*dgF1A9_OZ-5jo;zJEVh?DQ!0GkT&9@aupXPQ-YT;q#{sM4kwLNyfvA
z3_Pa(O#A)1Py(mF_yUcG4FDHBJe)u;j-0+Yx7(`?!mf=6m_-v}!hY&Zx9<@cG-A-;
zVF5#i4)^mP$Q$+#3(t@JIh_4Dg8ey!{W+BV>CgV`-hlEi2>;;m0@$B{?9ajM&;J8s
CXg$9G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d.hpp.1EC37BB5D09F3746.idx b/.cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d.hpp.1EC37BB5D09F3746.idx
new file mode 100755
index 0000000000000000000000000000000000000000..35fe9c96b67e3a3e58751c7ec4abd8c1c2699e83
GIT binary patch
literal 1386
zcmWIYbaP8$Wngel@vO*AElFfyU|<jg;^LB`%oU6b46lK56?66m#(Ez%5V`Arvw7dj
z$Y<q}%<0`6b+$j^lFDKae{!AV*Y@N7+Pi7_Ii*bsBD3ZczuWu#*%vqK^(LI&+_N-a
z#Pj6cxHRQq?)v?ot#Uqa9^;p`k*#o%S;*A&AvwgM@Svx2%8myj)?Seb{a3f9o!hXy
z!?HidRYW!++0a!v-_ZQXx1?7B)@M{SE<ALt*s=XVgn--nlpn5j6Be@X^4TokqFyoa
zed_YS{do^}m}ec`cRgRVZ^8Ku8+*2;ZqYq`Ym?fIj4bw!M?Grgo|)54UQa4JE>?L{
z>4)B`6HmU$&7D-J-`e`XAhA*0U@z0sJa5UCsT16$-_xCaxOMhad0n}l3yd+nv9DWa
zHs5wAU+Tr)zhH0V+?T)3JoglOm7HgGcydY<zyBoNFZ#A&wewnAS_Fb5SeJarIkhK3
zAUp2BQy=4<Uk(Z9O);Hk>v=-^xs;0FtkXqYE$2^$i>IV{Z!+5S@2UTp!;Xn>>wQ;W
zKbL-W)y}B*v$nsRsF=)L$U3udpM1|mxkGlcLg_vVOUrv3?>=1e%i}2L<7Vb%|1Kmk
z+Jt!SXHJe*<}R+xO_~dgH!;5Qht=oH`30C*IDuRs;8tZ(6=&dpa18m4E!y9g%Cj*s
zb1-nrv&-v&2?kCkU|cgWax>cB66ZHrr=BXx#Ky_MtthLg0VZJj>-Nl^%_;CAT!4uc
zX1tT8lQm3w_qt`-T1M82qD<^S-Ab}b8ejq@%}|t@R=ki2;YvnMu<Ky1g3*jzU{`sn
zlyy(s<y;4I6AKsEEihF)-0BAE)-VtAa2pvIA@M8>Ea1G@?Up$)YAVW%JlrZWDjF~?
zseP=f4MN3EKw)_cc^8<u6^9=_sbaZ!mJMiuthy3RH4nF@tR@oANY)66XCiBY#IumK
zK;l`+TETg+00jCR6mT#*yS<oO-uC~u5a!`l6jhXg`Iv`WNmNM&7OFhl+A`Y82>0m7
z=pgZQWpt5vdNO)QJbf8`B%Xnc0TRzp#t?~Tq;8~#u)$cy7)j1V-2_R_RK^rZ&P>J(
ziDxcjj>NN2w?NWoDPxHwXQgh1BxfyS4VPPMY!z^HZo+h6>^R9dslx&e78t<907_r5
z00hR4fhIh{dALmsOpth123Bz1Qv=p7?EK>0Ou)n-s3MK<tdpRVG|U!YO3g_Gm4mP}
q!T_^~p|~iySRa&|xIhF1@WL3t?AHBz`SLqG37eRCc)1vO5E20A;=^$O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp.1D083EFB833A8BDD.idx b/.cache/clangd/index/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp.1D083EFB833A8BDD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5163a6be3f5b749f7f72b1e3919524b4cb08274c
GIT binary patch
literal 18206
zcmcJ0cUY9i_CGssTbGw*7wNDRWu*zyQ9(dO6bsm5U1N*UB=#DOQBbN_0TB=tY$&22
zqDV)KH7cTjs8Iw9f}&Be#FpRS?%|#9U%w}R-1|KD9zJ(IJMWY;XU;iu9331KWH?w$
zYv2#Tzs{SrbULl2r6tGzU9xoXPwT{5T2(!@w0=#ioiKI5T`yVl8^1^kuXIJ;#6%}s
z*HdWG+@(Q3drw$5FSPrd*yE8;Ep7Vz;@w`nMK3sZzum-Z@yJm!xTN9UHsAY-m*O=a
z-&yu{^vkq4FF$_lx$cK0i|xwNCN_M1=+Ry{+|s)}a@VkmK)o=%d)w*so-Xsk4)k2;
zoHHj}dv$2E!D8#2;Qr1^ld3VUa~y|_bG6Hv6lp#vlO6iw>8KT-&Ym4OWbEeHyafhJ
z9v?k&eeC@&ZZ`7Rxbk&rtA__i#aYFg+%25DVZPU;8MoqQ={`I#H#)@Y!npRgE|s^v
zf(JY^{PZ$d|Mo9I%f_}wjk@6QHfofu#9;CF;}0yp-_kSY_t#wv>Ym74l?5g%ts)D`
z9sk~(Qr9P^@X7S1e)pNROCIcH>t`$o9=FJ)ee@u2N#9f7wFeE#ST|rq-KbF`maF!S
zf2Z?9e(;IN)F)mZc5`mN+x+$Kfl<2(-jA7Cq_hiYY<Si6M9qYP$f(e{^UC`k`(fwL
zIqyw2zv1HVy>wsq<=Nj`jch&zG)*~PRb9JjQ(gOxwx65$gtE(a%Le=X=yxgk!Rax7
zJHMH;Fu`}nY1{b5WltJn#@pQP7Gr$m?%7q%jk4KGKMjxSGQ5~JT$$0u<CEX)kSe9$
zhkss=**4v9Ojo77g?CF@>e7L03tRu3e#0rDXnAJ(nf=a*pHDUHKX&u%&u)i(Ug)_b
zk9~gk@$Fg1QWKRX1$}DWezKVOGJi+mB(3TlJ!dcYwKb*C=<NQxvzOntpVRxqVW-F2
z*O(2Tu6v~A%E6Ew0aYm@{^@r{S>wLbtN-MN3!~rdUE6!xqT+e;`(?Cr8}qjHVRGKD
zZS{2qF|E(&c^@p5RiRb0IJeeXryF*^ay5D{emDH^c%Ev0P_Kt`KIWb@9qjaEeu9;A
zpx)RlBa@b6Q=+yzpLN)g5&36s5hFc1Iel^e?RBcHf8BgxeW>QVXM*X0ktdHnxV$dk
z_l)PP0}j@A0`iw%Ki~hE>z(WWcFkWj`RzmV<s<eQ-uh^>{<E&bvJY>@UYtE?^e$zI
z?i{VA{3%gAMy;~DXx{wj)I~qPzYLZJHC4GDtqruF6!}lD-cui1l}$gh+oom5iUnDc
znqf^P*Gq1Hw8+@_XzY&h6YML!?u<IQS6^Xue7iDjVpkpgH*G3(VSP>i;2Cbe3^+8F
zo3-r4+}j5tE)G5THor&n$)}sU?AvE!@Yp%QWpK*LdBao-V~@q0@ZFPt!1<i-`P{ri
z_8zgv?_HUhl<_A2K<bjY7Y>{YHUCFx;(y0(Z~4f?+m*65m%LMX50WJ3+l@9z(z@(6
zE>BvRJal&8cQ+mf=nvHIW4C;yN093m|Nh4vvsO*-e-llAkER^2xH+>yZ~jI7!FMH_
zpUQJ{lK+*{)#G!!gof%HhG&M3FV0OJ+w=VM>tcI5W{_{y#z-nod)ee44IU|@Ps_e&
zzc1O{^M{pu;N%T^^DmBk9p*4%>iu<lCgmt+k9(F8*sXSfjhorr6KA~c=3T8TiMBgo
z8uQvQrpP>jol`fhTfz4}HB*ca4coVT=DJbKmqlba2HSt?du-qOLBE$?KW{c~^}fv+
zQ3E%;FAw$j^^05e)*A0Fo61_!8<Jac*ISx?_;Wn>)AUyH<?jD%I+I)G^<CDLWnZ0X
zhl_8%m$R2IZ5mS$bI0w{kI&{NeN4_bEl;TZ98#eOS!+GAsU^a1NxCXJV^st7!xqzN
zhxRSd+LiqzB;IW3%TH!brHZxwJMTYpOnI78XIbB{B5vuV^5V>wGh0u6-Rk%K1k2@<
z$_rL_|GjKQ+oaUBv1_LE`FgYOTlXV#r@tua_e$4#9@Sdyb9n#O;uW5+>d!kCwT!MA
zc6!9;srS#6?mC@!W%=#wNiB)3Qzrga-gNWG{KrGSuFoAO`}4{3!*f@DINA5viUm>c
zYi8H|J^$ksw|CWN#|C6n*7)q1c%kUYkMR$`<ULT_u-Cqy(Qon5-vU<jY_X^@^8U^I
zFS++*#b4Um+ipEuXJ)4JV(qg*rE8iU53YK<yTI#g_5AS*O+Ws1Cwg~nV;R$9YJBW3
zOQ~ad_cE6DpZ2<2%IS<N2M2DCHPAiQH)C<+c(3^ACG*7Ft(OHa<7;2V=S_VX`~LK_
zqaIIBY|vWr?9RvAMR{eJ{YQVTSkUvIhCQ|Z&qmyyUhdgnN7{hw=NxMB3V1!<t4C<~
z;H-awYkMBkoi5*>UpkJB8Q1%1>UeY2)0q;}9odPcw&x5QCjB$w&z?V7{XTVAgZ5;L
zi+Z!>JGFIv8Kz1O+paTCmTjJEy1tgqR^_IytL-ykN?5IO!nx2(-7i%*l>6-eDc<0Z
zoO`pj{~L9-qF0-F#CJi(=S)iT=gjAR`8>qbyQ@6t_kAdWU0Japa)Z|`rMGJFG480A
zVOyb>`J7q$;rUB+2H!axPF0%TD46i>iOlKwZ+gvLK1F^0bZa1=c(;paiNmd@@>PM;
zj!06w76ol_?0-mNns6*=_PP<R8SBR#98j^dpQtS)FrfINUE?AzM+fS<b@_`Q-OCrZ
z-+nIsoMyF3d7!e-)ZpKzeqD9IzVL3+o#OdA^dD2UY@Yhg>BrW=0mbn;`)BU3MXj?Y
zXZ3FTIOsum-?q6Prfmr^8{hSLx5&bG(~U>H;;qde{=LNF(Y_m!;LQa!<%cKE(A&JO
zUR&4CWaXBzZoRID_gy=3wS{Z(K96BHEZ_h1@kXo3#sRBiO2252sMvL(Ib~zfmXM)&
zIlT?;By6>)`t#a^!RP1fu=rpdd*V*sg8&~>nO&1SeA2_5+;pp<XS<G%^sBCTyV&sb
z)akMEgSVTz4s~^URes)NQP1q5g}2gjhMU%gPJHA0=f|gA%C4h9`C|^+#r_m9?OuKG
zq>XuB<}|PL*UvAl3VuZ`4VoCQzi;4f&!Vy=Pul;DnIRg{-Q-O5+L7Cq=p2yGeSvDV
zHi$w)H@wZVyc&|eW>P@<nm&j8rhl2(V4S`Z*_97k$AtzzSie^E;F@&Fl9lsjIP%yf
zc{zCX_Khny$xxr+TB<Xg^%>4xBebxi((xbe&5lLCyItyKfPTP~Pa@M#q8=In6v@?z
zad%&57s-%ssFrFQE8WJ*H3BHQcbCmtv3A97DVj4vOSOYB-ND#sL?@-u!^m&diy|rN
zhbeiCWgg?C5kQeM=dVn*6m6BF?=U5UG0I@<Gy*6lulE%`4@0-4Xbh&LGsfwRgGK<w
z{N%nn@he+-78woGQq_pK8j+7iXmz1P`dV5dmB=C_%YMtM)0d=ZDDLMJV||Ko)d=Xv
zVvo*%)n+4O^^gZ%nM@?hM0Oeh6ibU|kqZ;1_)C#I1kZ_*<xv(I0Tio&E(71c^X#gJ
z+;J~ONLGY<Q463bO|z)8xBkeIA|2ezF-mp}_o5a+u|ImW$B1leH!0#V<tQaPiYaOV
z6qiX4SFT>_F-D3;;u#e&$|A;JBY@&+pD>_tjq@b~)B}&@Z;|QWB6E!ZipP*MA^zuL
zS4h!d+)EDQmBR#U1W@`EG@l=3X4Qk&nuXWbN$lkE?SHho&`4|U5WKcNa>_cVYGkMs
zjmCZKW{h{^JyHwk!{=#Y(5+YJo@yg89?5!IvYxik2%rSj<X0*jGJ~{H7fd-tNlsD5
z8Ud7{-wo~eA>L4FfE+NTU1Zvh=c5*#l(j1c7Pnh)QuGTROA4c(!dPeoP=>MdfA*Hd
zW5pbZDJL1rlUOm;0w}|uZh9r3Idq{E4ZxILjL|Nvm}&u(kldKXlIT-QwUIR*ODQEO
zr3^KqlTxC(F+8+ix*l@Elx$X(&34xapp2SJWx7W{Ezw4Xm{Led3Mmtf=%gGg%dHa)
zUZg&am5j8Kk*m}KD5_<N66J4u`>?1d?xl>;E@SjG0w~k$_=XYn{;UDA!<5$|)7N<E
z)B-3oBZ`-_m|cHrh`M5(`;5e&k&Q+G<%gj@+YDWb^9_&<?&XEZ^abukEr2p>jNZ>K
zPrSkmkPoJ`iWIFPJB<L!kG;bOY#XbMk9{Cs-9wDx5aXc{K$*My(3%VK#);a<1ou)z
zNs6#Cs0C2w^=VzUAY_1#0qTY+&qb!s@mSOXC`<ImoGW;B|DF{2;}4}VMrn8()dDCh
z%&*_PFy-E3Df$Uhk{IVCrjJH+QVNRK@Zt4?^pGe1&;ulUfb2D*lTsYAuJwsxfDBn-
z$^}k#fm3J%P<{=|zF|0U`5h^e;?>QiWV!f6s|8S2N63dhcU|BpMYHi(wlk*Nu~AhE
zpoFg({Gs=wAJ-Wmd;Fm{BGWf`(rN*ebq}no>KhjglA>|Amu!a5W_&aPC=vYQelKDT
zE=W-@rsOl$`Pdk$1yCZ-IG)($xOkQ9o9<SzU8-1fm0AEL>L1DcGQ(c8q{t8VvY%1x
z$J?kDK-r?&aBWIb)qN?_#gopV>~pBz8Ud8}PtxBn-01d)6b-_?>|u=d;5||cpd@~J
zAwQJf>Y<Ia@uW{vlG9ij)B-4}n+`j@9cnQ{h7|Zi8I*el<);xqNppFBs%c*--s!RU
zLs<-;#rSGOCuMGIYjkF>5*c#EAByDYNX|$jfU@hi7YoCNbhDA7QMi{(hR<YrYXngC
zm{ymJ&AHL6jU>32C|VMQZJk;GWpDfMhHY=s>ZC}9hkKl|KaST$Er7D$C1uH5tErD=
z$P!aBDaTC8TO&FtgMPA}n7KYr8*%uau3)4UjG;yVB~$O*wA`baQ}mEG9?LZ(yM}Kg
zwE)WD(x1iiBAzeTL(Z6z%$g>%o*Ds^to_Tz7tIo5N^eXF=ValWnMMHR$oTS-J^@Y>
z4A4wWc`q`3kKGEj=%n0BY%O_i&@4qkc<BlmSs`Pt5kSe__sU&i(W9>nS>Q=0QSM2U
zzeWHhx6F3yqXUv72FMYA=&i`~Ej|Hi0hHq@zwFPSy`V`Ov6ym_lAOf)q830Y$gA31
zb!~hEi#SY)qqXDct{MT9;*q0!mh4`-T8aYja8EFnC-60)7C<SlV6C&4c0bOduDF+v
zNc$tQ)d-+eNSh8>>$k5pK;1E=O=Q}JhpQHylxM$cx1Tv{q=&pZC`i_Z?TcCf<;>yY
z^ikb<H^`6~UY9h=Jq@pmS^%YTSb%nO@7jw7NQrxSDKdSDKcp5wIs1JN>oWICTlJ6&
zrqr;q8hobJ0x0KK-Z43uR&rkt>0wGbD@(@|wE#-(GT+tLkV#iV#Nl<>M2k1kGK~Pr
zy-U7Rj<oJ5lA`{Ya)41Bz?MNRfYPA-u=vD|%}@1^8~#uNk|kipREti^@5k*=`D!&A
zpuU*WB2u*An?x;u^59^4Qrv`31u|rc*Cmc}kHg1aEr9aq&@$th7gOJBBLmz^0VOHG
zH>_F!<?%k}pN7sabdn)++{;$VeJk!oEjlTumd_h<dtQV#VsI~Ev?L6lX|(`Ko5Ule
z@WN6&=^yc=6B*M)Y`xV2D6f=W?GI-~;9iVzFU=xJvq(=Pfb#lC$mHCsGZq>kTTFQ+
zGJS=&Q7wS-#)3*b8~j~>6#a}pl*~9M<10%ofbwyIoiC2qAL*f99m7SkFub~I0Tivp
zvwmDM*%e0wnwUW((L@U1PeY_YNi^{RH1MiRNcR#l16P+<-9x(fkQo@8E^lUJyjNBY
zMZD@TV|kbv4W^e@J!BjoGT*O0yyBt3wU7`9H1MZ2QHI8m(Q%1M{dv3VT`7-mbj#-=
zC0zDAK0B6eA|)Iv9(x~_FGNakOn9vFmM=w0I1M~@cP;-GDZvxw@mgBG5-DMsdF%sO
zz7{FLJ>v0YYWYT_gnO39?xkhBNC}=Nk4=r`Tagm15U*-Qd@Ishq45!T)qBLhM|u^S
zGtH}ZPysurATVq^HsOOasS#k^bxs~Cw__HTvOHE>aU3Q47Ng<+;jsdWw~{!C5+{*3
zjS{DkID-;rkT{bPX9`%?xMqOM&HKNwv}&T=M0@b*Y1LG_sRWwgJENb_Hq&<|foA#6
zBG7E#*_{ZYLA>GFv^1MG1{)#K#DrjObu}>|CRWhW3ff#CssltILS)GBs(f0SPg??!
zSC!Ml%4wB=!x_hL##g{mjBylW4_Km!C!u4*`FH>2+(+)iBlsW^e-IhK#l@>Wi^QKr
z2B7uEw~mWHXBN0owCWl&>KZe8l*SLIRf()oB5MrFORK8c?$xXfcuXd`zLx?HR-6;@
zs$qSG4T05!K$hlJCur#j+6qR<t3J}Qk90QyZ)S`)GY)_yns66pV5YNG*1u|A7Vf{4
zGAX4(ArgV(i6olnH`r!*+_s{DYj|v944Os3V16JL#%u5gE&YSG5I&_2jNwxd5n~IW
z4u|0ryy_{kev0}CYX6e9cu9LfN0!p?KC$8}jd+Zg8MDib9V{LM&=jwVrlir7IRqWN
zY9Hmfj~WP$oJ12=!xd}UVtcB~@(2F7M|&-M2^<qJL3o7f7+V<OX-avT>Ic3fugaqx
z^5_9DLhGW%bDoF!Mu})VLc`0935>9te}3WX-q$7ndEX(b$5UHRN!OE8i>Ex}sey3K
zNHnoNObudv{KIvWbRGG_O_b**Y9QQl5={mGeh3)=ypQK;>3P}?CWjZ}TRH&t(N)*I
zcl66;mU!`MMLM-24_GBIi`W3GvjfloX#qT@a#~tWTSM1)jEv(t#t*PWlPrKwc|P~L
zJ@?YzBk`vkX-5OtCpZOIH#-stf@an*HZ{y}xDfhUyt@1gJN=cER$bxvE1YzNCa|Ja
zS2_MFiLY@c*Elnwy^6D_;;aOGowK-3;%d&Kn#4Ca{TrO2z`x1q-z0Gjr(Z+jTF$(d
zvk>@ooOvCIZ*f+)xNbswJ!e%<;@h0nZ2?2d1Y3jSwDdS>4PMZ)7o;`Vz!+~Jt${kR
z0_zQl6?_2HSrynp9y?s(1B^NNV!Y}BWAT7-gQX6<l+-%*V|OpCJTa^?hV_C2z^gX0
z%8ddJ{4*ftkNzIMct?lY4)uVYgmeup(Ijbv4_~3gS13a;OS~$Q<|AoS7%7hvDdJs>
zA>>#F-%Rt_Icd-meN03ik;o667hV;GJfcXv0eNg7aUAl9BXKhFNJf6(TJowC<dH(+
zUC3h>iFYH9-6YON9@)qbtRAmAiad^z_&D-7PU1Y|kw@Zk<WY|Nz$W244tZ3N_z&dq
z2Z_%ik8>oxi9Bv1KX9XXRSoi}A#pwOs3-Al<Z)ZT;G^Kftj>wSjttf0#PC)hqov2l
z>sg&317(8D7$%;jrDw_OS)CsPBJ3Sj-&|UnOJ2`~bWkBZ9?S!;`j>Y6m+ritC7LW6
zbTlH=+vwt;<WX3;E>qnvQ$0XyLi`qrMoC96#tqT5b2L2$qG}+ROElRwm?f_Yr>(;2
zo{u%II<JbLts>~2g5E{bR?&1%=!jPp(0T>J<;SZEX}v-cm(Y48Brd1BmDA3`)V|WX
zUuiP|M>6`63~{f=GWxM3-o)r{B5@KUPhu31f#R_bFHdF^;NJ79gN*zjqkxP<hchid
z%qRqB`UoRG!YCl$!(&S<&tVi`y7A@C$o^o=g>zlY=+`n<U;ud4Ek<#Ru@`VXqo^lw
zBco^}@k2)Oki?G|#Um0wW)zP}9L4HJv1Y>eFR;27ShEmKRum*wS9NL>HZr6<CTr|Q
z_AA!HWX3X?aS~Rfo^h;a`oM@Gd5UeHI&%s=@Tw|GTtyjyuJG7}5~njp;Df=C@TWET
zR88D5rputxd9!4jvDRJSx?JFN!JdzszI|eL#@)WS;V-W1U&6&Oq3Z?LX~&F}k`C)-
z6p7pg>$L$HZ6I+nGD=47f_2)3jCPSY8yRIIcftA`M@Gj<T#k&&k-K1B{y;{5koYDt
zx{2Hc>rs!4>IFREQ?l!m;`JW?lK=n`(D*V?XP<?g2<L3WIcLZqVZ*{%o#kvH=<LAu
zXE_(ZQ&4zXgJ<C|y!Qtg%Y#gap#6=EV<R&O_I^r1pSh|YHr=rWilY^Av^`i99*2^~
z+gKZL&#^AB-K*GM;GSc%!0AVDR%x1CA4Gmw7u4x+82Svi;Nq^+%TD7hzbSIRDeBwV
zu*Rv&oy13S+~7bv7=w!go-FQ2ojV6N4PsN=piY%TN1ds1i6%h~4fw7`fe%oq;9fVP
zz(x``p}-~*x1qo`6e>iWFHqnM62C-&FG>6!b$yR41b!IRCyWY&DPX6BGCxk)LYfk<
zDQ&-v_7F^68m&yDT?D+BR_-P7YufZRtrXhNu&!rVUpVr-s-E?%X9FN1h4TWe>qFL8
zz%8t63yEVnmsrjR_Ka60axRG^PUT!u1q=@iSc{5jX|XU6iRPIBl;x-HmHJjX(o7>>
zH9|7N5PTb66`~(vC7@7ADA6`je<aa1N-~OQ8?8T@Xd5FLL$r<6A4{~2myS0LZd>FW
z@XOCjdDR5t2{r<nBAr6CeQ*3d(H2Mbh@;#fB;Zw>DBDexyMVV+J+@MA5Nq(Nc*-`O
zau;wC)gy^=6Y>!$lx+&-F5on(M;he@2}fSFo3h<axeGXh>XAXYL73JNw%P8Z+y#7?
zvOP?>Lw*C^aXR9&Z;v>FieIP1*GVrigAr#ih7kBe!UrE3^_vd-aAEJ4i^kksYQ*uX
zC#csG<oD%r)9E*Xe*JOk2K9P_{G$6E@$b9rNgj@RX}4y&zj6Ng9^V`5wd0tTc6&<V
zH?-Ru61UTC?Ib?HIv!yAcAGdwGNLfVZxOAPa@gh5^>6mT=BCIMMg*@+*c9c^(j3wm
zDxiZ3NN4CP?f8{+hSYCSppD=O;GpXWEj>b;2<PV+Eqf-o8WPP5Rp$>i*(V%bsh_7{
zXxPW0JPN20kOYP646myCkqU_LI)y(SNl7Eg7<4z~xtkgY3LsIb->ks6{d%_X(JH$7
zs(FXa@_c~$k~Yh;k;F~Nvx&rQ$g>Ug6>QK8<oSZcFOla<5`RSwUr{f?MukxhVHAOL
zD4iTi51yPlY?#h*N>8w1TWG^A!WwqiG{bEqPNfY~Nt{L-rja<EHcTh+UfOUkiR;*I
zb*!ym!|K^?^(1a&yET&dA=~XCiJRGO%_MGNyS0!wno~q`-2@vL%PC?>9M37@Nu0<j
z5=orQDUwN?$|+I>43Cf4W@gdSEHX$?KS6@BL24B1kow^fbZA}CXvYnI6|BJr#Icv-
z0Js;yKgFt5MoY`+eu7p<FtP~7A29X<zdcsMhv1bGkEoKCRtlcDMDteJN!+(~)rdpq
zE{JjF1(~cz-2@*j23f=)TLEXXicGc#ymIiWB37q}H5BfpVpgY^#HFmXlr<9C%UEd{
zi7Q#1N`W7#?R8=5sO#By1un3P3#<eD5S$OZX6iRoSTlG-#dADGOHYwEmHH(W<`_4W
zifG&NZ|~W>YM|ag6X9+OmITWMG)yv#KrJGJ7Lj>_<{bs+sznB^B6Dy%Ix?W%A4Gma
z2K0j{;Dcx|iGx3gMhf_|$osR%4>F`38PkBzqQN8%{wx|P;4ow#hFpZz4@2HzWV{}R
z28E&FLVE(TPe3k$hn#@C6Ug{H0S!t(!-e)tWS@y#1dllrd1sRGcqST@iG~a9MaaGg
zxd<L~5%Mk~<L@Fgs0a-g+OHw|Ysf|LsIMXKYeKxu<A>Bi*U)gG{Q>fMfCi8j(5I2a
zO~|K-#BIo@4GkbIpw9~uzeGMSN&FcpKck+41^j}PUr78FxqL<5Li@i+`7hZXM!AGh
zM0+@;45x_x;wb+(Dv0zY{I^m;r0)=rL=7f=g@80_FkB5hesK5Cpn?QnAd~8yN%?~>
zfb%Jo_$Z|d`A(dTpv1>0T}Yhb>yzr5LzxN|JeLyZQo3+Aby)PS$0<|6qUTZKJW3aG
z%DgIpmL~{?3ww#QJdwoPY58^%C(-gG67QhpJ4l>N%aci*Ld#P~oJyOg(%r#+@v5D)
zd?(qSMw_RR?Yn6CF0y?$E#FPzJ+yoeiPLHGbn^EZv^;}s-%FeCCENGW@_l4`AtNqi
zbdPA>-FekjM(?VS>E=~;Sj8RIM!2!=vWmMT{)<)oMdEv`;vR|bvx@s9ZeSG+B!0ju
z9+0?^wQFR13EyjC6-{LOL)Pvg+5U)CJR;j4vx>(ge!?oAkhqz(YbJmHlvO+>+gn(>
z7P7sSRkV`rah!1+XDQsU8#&{RB;LdsZzAz#&UiD4w{XT=NW7IZ-b&(aobfgi$8)yv
zoU`z~1kN~tY)|BD6Up}Nobh(DJ&7|;BJmE+cn67-Ioo9N_bHrl3fZ2@*`|{1J2~T>
zLOYy8d~rum(g;c+7?u>uGldESLxA4^zE$>sPq^jw%HBF!`E4DJfT5!kI08`VgZ<<@
zTAD}t$t8493F#+?F^*xvo1R2d4+M8%;#GM}v{UmD-2X9Fc8s-xYZe}Ra2i5g9t77c
zHr15)8X0ZG(0mMSD&zw9FycLoAq;VQqeDeN+5T4ig-e|ACC&yM=k4#OO<AXFUW?~Y
z$v9OqeIS+8frnKx;{_bg8po4qh<Mg3o*e*biKI(IGH*t$>4`7?y`27DVbwb_BKl`J
zOUQ_THN~!jx~vIS8S;R5drE0(Dc$)bgA@3SY$Fo@c1fbCa)RXpeZxJfOPpYp(snsE
zt*%_=&HgXmh#`@Khg(ETi|D>EJpBH^$YRLdR##QQaAA*d2&pcjg5mNwm=_;n%puz6
zRZWaVlkma;_XPf86eW!!_4hF4d6)`>i%EhnmW~Q6SXuc0@dByqt~whA-c<~!{J8V~
zyJ`D1m05VqnzAfSg0VwQj4K%{K}`}_$3)g0)B=17MuMG-j$$ogvTu>~ThvEz7K$ik
z5!DX@Y+iMeHa|(GXg<*vpM)1#sPe*dQkQr^$B>!B4eG)!XaKhlA0~Aj7c}syEl76@
zG86o^y-0U2G86o=Q%Ls|G6TO0?jO8r=V<9U+7xykr)|F#fx%9LJ;6g!mwv&zA9e3L
z^4zXzJ8%zWjCmPj3;9MKm#pZY5e^HKW#AsvC1EgTC<()pQ5S+i1F!m%68}kBhIKT*
zjy4ru5z-iO8e<4S{>f*nP8w}^o%NrehfjIsCI43ad<cGUzQlFCBxI0b_MAjhfF=x0
zU0ekO{GP;guV5TOxp1)s>yXX5f%W5YA%b@`>kq*Ou61E`U$I8QeY=;l+RNF&`9J;R
zgDZiO%Z=Y^Pvmiw85ru0+7tO!5+_meBoe1l@-z}>Q1T2CXHxP^0T=sP%OAzO+k}fb
z#?j-XLbRfyRRnrM^G|3)P!3#~L-Vgmyq^{CXZ0X<gYrDA6K81Y8S<@4U8D!gR&ud+
z;!&SBkN@+Im3fT*ff=;-8BXo5N4?jh072oCkXI5Jm}DWZEHdyYMP8+3U{Q^{s>#5i
zi0xU#dI%d=!uBd*y#-v#I+e1X0xoBR%GnSBSF$}Th2NJCjqTQUbzvorf0NL#Bs5Yu
zi96AlooIr9v(VryG9t=Bp*d(QoK$?1pkbwGq!0mBqA`_df`F^h;A%4BsY9W4Xe?MQ
zIIsAWsf!^6RsO#zaeN0b;v&WdDj0AskI{d^xIlKlv%E{9sg)EuibBCrXqd1SamYUo
z^%rmo@=rnig&FTgLw2K~0zQg@kD_71jPsCx9_laP3glmb`U^8YhlZR(Lj_!ef@{z)
zm@%)qjr?z;{sOLMe5#p#VBf)9<IJ48qEnbiA>~m>jes{S2%@m%QrC6DT0$ioZctZq
zLJx4Paf7;sv$NsQ$oV=!;)wt6Ed@sj-%{!lPw0VHm9RP`tcjp-<*ZFP>jL<~jp{%D
zY;CXjpEwE^T2jMyJ@u;Qap4umUn6TisyM!i#Me3gI*F?}zM8~0IQ|BSZ*u%i64!8i
z4T)<xzLvyw9A8J`TO5Cj#Pu9sPvYAge_O!U?&cSVl}{_jPGmUk98QmcJEh9E$Li6c
z=1KqAB3&Q<v+;-JyE^K4hR{Q#V6Us|mW_9`I=`0hNQqlLLC$1K-0BT-7E<B`lz9Q=
z0O?@-<V~3uQVyVm9Rn4I(}pn6jwf_+1Z_w@p^KwwLvYDDQtaKEX;(7E-u)?w-_YIP
zkhq=h-cI5JtoZ@f0aEO>y`JAa@!hlpthtYn?jxiSQYo*H?kl8#;{=xk*8Tr`N7lEe
zjh>UdA^Lx!nNUiDhj*Hmo+guZ>f&a&a=_Z)gQYHM77o@aN_mRv2gNfT<r5CMLRA{1
z>U1TVLTBNJt*C1&vJe*JE9&zV5ifQ*lz9$i3o#q^(P;ZEq&t{OD^tlAx^!BZPU0uD
z=@Vf_9Tp(4gbjrg&#UTK&pPtOu90<ZBwy&7S=VL~M{_RG!iyY!hvHn~Nu12NBnudF
zPIyDr<=bHX8hcj_e4lMB`p;+Ok@10^Tc7AU<IQ@;I6q_Dz$D|6V#fKofZ;IU%~F?z
z3u}Ir5?>{&2zE2#-HZ{`ZasEk+BsdLIQ)eOu1f^h6&$M$ue8e<t}BFakU+zeQ&*M4
z<lvm)PpfOmg-`Ei#QPa@NDuR>2F9X62yL2s79Mq19NCKvK?`Hl!Z^X?T7s8V_sur4
zz+*bXSRP?ULG~S=D8}&-GX-wgXX5NFsBi98{EQZZ`o*9j!VS6;`RqhNpu;#ThkSBS
zkZ?m*BA-eWB;1H~$fphk2{&Ld+q0N-X3AKnGPXBp5|qi~5gw+chlO;NL{lsehp6q7
z>cfbgQQqIY8j~bsC5&h%ve=1~0?tAvS;$HlVGgp$K}rFaB9l^NC5*TdSyUpWfUA*7
zHL?<BP=_q)kP`6A;^xV&Mx#3_<&qf7B*qEe={hQC9B&ErC{SOkBheJ~3)}e=89hao
z!s&gBjNT$kFlQYPhgLa4{bq+*u*xNI9&MFJ;*+%ANg<HJRqwR^OTh)^Ri9}6Pb98j
zOez>_A#!M7^cxs+P==1$K>Y_KZesMC1pGR!u)J5UwatG%givOWHDCR27+^(SzYp$b
zX17uwD{utu96^tPD+X$PaL|5~mL4VFK;P4{_riM{j$^<5%Lh=GcL(*}p-tUK<6X*O
zEOQu7p$_UX<M^2A50eE&?>Kbd{$T`+35U>>@+8#{V#SVKamW#j^QTV-%X|Hm)Q-n|
znG#>7tf9Q<3!TO1>7Ckz8?MsyRayrwX8dm~=1vd92nuku;9G&&+D-Z|*NzuK;1BgI
zS}-4od9@(cU9x!Qk}VxQPweboYqId?B;n5~!k<$+|3s>gzi9TS-ZN)vPoFVu;-v2<
fPnkM?g4Xw)1g#0epEHC%#|wW>7yg_k{Q3U?iAi3t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle.hpp.008517D2C19930C3.idx b/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle.hpp.008517D2C19930C3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4420018cd8986a90a01c8cbc7815c3a318d51b9f
GIT binary patch
literal 17582
zcmds<XINFo_VCZ1*#(Z<!=Xr#CP))dkaiGpBSjSKSd-{g6QfD4-CT`{i6RQ3MnSQm
z7&R7R14W652mz6%MiG$S5k*kJilDw@&Y8^Q)Bkzy`|aN6-YkAuYi3Vfv!<LI77`qM
zB1k3+oFDSt;xE6Q$I4_fUG%?|->&$2hai&~+Q?+z9lSO3vmI5bdJk5v<1D{lpSrB^
z()}NzC$q^JBdT&2xFkj#yr%0hsw%ti=*Y0zMO)s#-X1jhs7>MC`KD%B$;bCEjci_a
zE;#f3oBMOm%+a&B*EesM&WlmU-afT>-<dRT)FGedds}>0<_{cQ(698xz^g6W+((c9
zdHEBEQ$vS1-teqjWVfxfeDC(^wFyJ^>Lzv8P3!ghF2=2Ps;l?c))tqh-jAJmCto&e
zYp-Ev-yJVkCcX*Gwk<ufxo@H8Bvs+18LDqes>2Fq@XN~B>g&>C?rrrK5$4af+Ep7j
zZC|=^ZNZ}Z<sB{F>HF4Xe=_FRXJc>wv1k5(r#I%AJ^M4Lsps2W=a*R8&RKD*Y~}bF
z53D2YZNKWT{mMfp(_ou_!Q~&C-4^!0ksS8f!h?IZRr!1|)y?jA80d3)&!uC}D>tnB
z*N?8vTPs)Cc3%kiU=ozoZCPVDeD&>`xd}c?R0XOxlZ$5#PV2qA=OCZrFfG1F$M;{4
zE?MuY(qA#U{m?`~(4J^BX-2O<KdXP*yuH=4+?L7D#ldsi%CmoJ9M@+}za`>|lAjN{
z_n)k7ud?RXm4#=9T?tRxzE=8|mTo^_30_t$9Q-JuVVK46nx#d9%DWDg7fBhuV=gZ)
zT0MJ(#b?{i>S`Ak-92pjD5d4A^6=N?#TIM6{btdyvlU&n&rYRZ>@Mm2=FOnZ{Ms^5
zw>sj_s^6}pHr?>7X&Dz5>6_;A(>J;!u6(ul$dc5^zWs}rbh{SZe(ZMJuB~g@`uo@a
zIFU2)e)PfUW^Tysl}7&=vC#gf5x-22Hj95zb@o)~mZy_5*3T(&&39Y7XZe9?y_vCx
zzjs?az4M^&zYUi!o##9J*pNe(VTLY>-?r?{EAgFEQ8&4(dj3I^jxA1Q`Dsf`T?Y?+
zIjhl&ao-v{Z)aVx|C4qvtx=sdXHS&6-n?!X;`vKhuI2HcW=<G1wPO1Hm?FhO%RsHA
z@#S7tf66nb{AYB|n^)K83syTb>psN&&WY{E&o<T1(Tbh*wTau$`T+%UuWw8@%+B8w
zu~y~vzHEK69L)B=*DV{ca_5S<5ucAY4c^*bzA@UPEOmyfEa~*;p_i@nW6KS*>ofkf
z)o5^UV%87l#wmvH&tyG)@#P{#Xm)w*jMR|j{l4w~y;PR4=Ce&1VLSf(V&xH={KOZJ
zHZH3FVXo<+Y>y7ttUjX`F`o}ztFy>`>dc4b`B^t7U-_i@X5WK%d~<Vm*p|O4ymb3e
zzq7GdW}II<V>y_!ZQg>!`ON&CUxi%XX{@?<YsS3PPuu6E6vw#sWiNz()ob#u^57re
zUmo_;u((gQhAC#BI~e8l{7(NVd$ukSXDUv6es<}*X}wo^UUZ*HZym9AY4W8#=Z+r9
zZQDG&u0R1--3}db^^4H9hJ@#nbBEafx@GG3=YKsm{>tKA0}9P^pNt*)vZP;`)`W=<
z56^0xF<uoFotE+4$)Jort0&Y{tb3Z}@#(oC$JYJyXkJ}&<M-XV?nU!HH#e;GJn4OI
z+4@bdZrRNq^I>Y(%*3HPruwX!V7++XYQD4i{L*--vdF?~_2nM#C5_9b_ZEKgYhcdu
zDZx4mM>HHf>mPXi%a^Ze2YvRWx#8gDhgt{XDn3lP^j+@G`bVd$J5sCcSEuBy3h4;D
z_*r=Ro}^!vum1G$%q?Fn{!{zux=KH*7th{T9FAH&j?35+<?(<?->^_G`OG`_^`o9V
zeLr}`y1J#YbMAKDAGM?XWziJN%9g!Reit)c7C(Es>eYXo)~=j5=gHgrCD%OuJS4qe
z*nHkQK7U$#jo3BQsAcS*e>6bJ>PgkbS6-(xX7^jl?tT?@=}GJlLk3&!;mw<8ew+Kd
zRa3XS&IRYi+d>b2tM}O*`JKQG=Z0NBb$8c@i3zRypRaB4T^Kkr_4BI%ONNeA**{#c
zeo5p*yV@=GL+<_@e=@}S*aQEs5178xj&B|?+~2t6_Pa5*^Q|3X9&VR9Lz=pro<Ep%
z`o@oI4CXD$J=wRj4Y(~S+WN;HlbR*vYgX-ko1s_J=T@F~RK}Xf?d<57LFXNd7HZ9N
zI=Wdf&Nlpac!lYgH+TF#>y&e}bKq|;KJ#sH=6V{vi<}H!L~!S)hI%h}vElyImRi&P
zFCAOteo^1|)jcmyv4BqXqu#7B-|&mel;7@;{K8|Jta`z!KL@|h*%>)m-u-&;h9lPx
zDoS;<^LhsU9DBf9yX%C(21Cyn)z7jtlR)--r{VedMS~ZvxAAd|mi2z^Tjo1`*@)~1
zw>!P_fBez6Hb=5q@~@ESm(?@Ees?mzTC6u}>ikNd6#GxJOSkvDd;jU8@7A6DYUH%b
z9rlg;#|C`SKO=F{OZPQL9)hbfgB!kWQA4*DuYX;kKgMx<*s|D3nJ;cBLaax1txB*-
zI$N6@5_E6h**RZ_XW#UXwadCQerHVG@<A6K=vKdZd~W93N$YkW%eAdY|1|uA!QiUZ
zGd<>vFMYqdTX8z_fX#*0+53{0|2bX?tWUO%S{Kz}aMj_%ug~gcxfEURo-zH|@QLG6
zBW6@?{CSFOagca!TtuYP)}YrdS290u|2Z;k)be)=T-(Q5ed{;vLh;LA8VsZ#>o>*7
z<_@V{8zG<el|yl!q$B%bz;G+w`K^cBqjyhyvMq2@diu+|@wv)1C%5jbHa63$+~PX&
zgDlH<oR)0mn#G?#LF=+?#XURgN73DV0PiJ}VmL(%XQ&Y}Eu=>Ofr<MKa<+c{Mgg2W
zWKtAo5yd%a1UAu<+S<2R?VO|l+MY5goK=Ri<{E)b`s`l^me15}RsdI&%X`-FJ!_{C
z*radd;e(;q!si$Q9+~{XYW=|~H3FL`4tH;|d-AEC9<XqiNg2E{gE!L%Y@+l#wxoe^
zbWj34RF-z=-VP;=_}k=bmqizr9-gQNtWhqT1mz~dR3os7(Q)TtlGjh0^?)fdIl?QC
z@B=jho0wKk2s3cqzpM{1L?%(JJc>2c2y9|u;Z_6|4szgu0-038K9$g3BmOq&u$)w`
z{j)_sV1!JPnEpwOnMPm}t7~fy_2=#c_5n&{a+Z;wW%M)xn^+fg*BtRKdC?EpA(N}T
z$yMH3Be020!GrC5nPr#~=%aS&gzlZFUDN`b*q%yk`*F>(BmIC0%H=%M|2!&-T3{3B
zq8YX`<}CU`2^`RPiQ^S<ysk!I6PGup9y{_BtCe6N%B5JaDi+)|0-FqdRimwV8Zll0
z#-g5%=Io+5UyZ;fBSuVnaOI1r(+Z$MCL1~1jhvrG{B4qQD(=;hqni{U7@53bP2RB9
z8i7p$#;^L+cce)i4|LpQ(rHFG%{Xa<OdB$`G8rQ=!A~OAl!p~YDu6S}=N)VIj&;)r
zoX@x|tuwEF`pXtQVDqtkq4Fwp)Cg=cW#HLd`Qq$_Jm`;1wsU>9b9x$qP3H7#9DDmr
z3!?y`=tB{l7{LwE2yC+8i&de8N4zKX0UXMu6v|7Xxkg}<Mb|1b8j?eQ)dOb8<N)9Q
z0Pm?0*ktLe#+Vz{!-gt=h(6TADtpk-R10jf;ziD|M}xl_)&~eEmog|XLmyHLY_jse
zlcSe@i$+~A4&}0iv){r=8i7qVt*tT{=oP^z|C%OAymk_rk7|KUBER{<JuF@p(g)b1
z50yiCIqa_y*ksEBo8{h_#Zwd@1byfOYx040)(C78>w3r4W91un9`GoaOtw!ZtFIB*
zByM<5UHZW4FBL!w4e<4>ay@cFYJp95k1jphR(I4~4-7!L!~<nK7_1T4WM5mk>gb2?
zi%QV%V{rv#xnQml*ku2_bG@!N52Y%A9DQg5tK5J-q!!pDVc(Qn@@scr=mB$-OA4b*
zVFqgiHn}t{%f_<FyFdX3p-M+`7LmxUss%Q=?7sHpm78ZTDnK9fp$JwPfj*=b*yM`#
zova$h_qYO#Lb<$VO<toppcdF9Z~VgPS6X&{rT~MH$tKQo6E|8Tuu1-dGr!;3_BHGW
zj8Sounf}RW;;RKVDfupOn8Uo=j(tF1WO5(M??WSvz$T@-n{skIcYn|WR>&kAD#MYt
zQwwZTcF%s-jMPOdl)wPh=rweIjhv8LV3Vrc?Uq3;!y^=6G%9X4tKZF<X#_TDDqL_Z
zIn35g0UVJ@FKgC|CahXuljZ_Nc6w0yR0SA=Oy064Z&?S8z$PsPd94xW%1(>2Amkey
zwH)34`iD%L1+wu&kY8x2H0$@lH0n1cutfEV7T9RPNF#6_Pg5@1j~zGrsM;Ow<P<wO
zT}dsl$;%q!Y0I_aLlnRTm1Hw#u^IKVT40mcTi$Ui{q+JBU??&P=frT%M<cLFcl4wK
zKDSQ4F$4<KaMe(&8ro<CHt9{vo);JNgF*>(kx46bZ$)jZ7T9FPmkU;!d7w32v+6^+
zX06B0OGJe?-pLKX^%AAMywhIZ9lL!T3jk4SViirSDPFVXnuy@<d7ADyz<1QS45cXD
zWt4XrCp;rWshm-k)BK20KBBps8Pv@L5&QK(xgJpIV}LS-=F32N893poElQa{nMv~<
zpu9tKJs4CEf{1-9P`1+g*FgE2=6I+d4~O8vElS6r{xO=*LH%<y7eM_2I0Ua4I5dKy
z)BqI?&=mKiC_RCmPhbF!SLB+23C9fbrNL!I@r9dF!Y;<AiwPtxk_0@HfJAC=8v2}u
zfw+V?(&AB5vx;)o0{4+9)v*qBtT*8XHl%_5gm4?{)W!}c{FF6*$~s_{YvMKB$9-Kl
z$A7bXXq8+<0gGi4V{4&_WklpVERz{q>{dh+L|C3@Y;k-cN^hXw8#oGk15v8t#VTG8
zM|Ws&@nRFrZM@h<^GQKGDd^z|ioP$1Ni?SlVk%)=0Y<I~C;zU&rNm4JZdnE^qKTr_
z3YC8ayyzbyiWs%pY3_ttoix9OTCWM?lA|xEgHjxL=ops-xZbP%H_J-nUB?f_K8KaU
ze8WZ%n&LZ!qN%=9|0X;<P|oUD7LO05CWu7}bp(q)Y3MZXcHV^aw`kzhGd}f902w%k
zq0eC$fCq{w<v`CII06q<9P^?a)d4Tg(fIy(t=^NryD3;Hg`1edO~UShl~TE3shkht
zy}Z$0-Wa<JR=UjFU*;Wfn7~R6yio&hjGK#<9`Z&HY3|^~4!$qZKjp=zG<Wi1C(UOC
zqqBlB9tf;-Q82nlbDAKg34Mt^T@cf0&Je^5!X`St1tU*n<uan=Kg@p&?jIcK>mx7s
zSD23*2F+>Ks(~Gh2bd^5WOW`AuZZS2Yw&~}K=fU#K^M*IIo<V~IbHxnX#=ObfiuT}
zqKMXM-A$Z1UcyD`6sJ7Jnc?;pr8G__jWZ>Dl{2_XyYD=dm>Wd&`J8z^&4rwKA<eCv
zK`WiUi!<+{_0Ku;=QO|M%wN)+$m=BXrg$XbvC7FcyB^YT>Lm;th$ytz+Q`}9s9%%{
zID-Ptl@$C1XYhhB_B24Q*(;HZQ#sF6PK6U%D#C|FYu#)>J(|RsCUFjbD<0XWYOkvI
zLkc^Yog7bmt1BzMYG3ag$*{=YFo`q4_IMK}mut3O<cnpjZ5cfV?z6V{X)b3i%UOFe
z1}a$V3f7Tu3}+I<S>ZKVlyW(PT)K~Ma0WMMe#RL*qq&na=p=0IcQLKY`;Pp7ezM{{
zZ(q4)r%QSyn|I0Py|EV&rJKCVO`31<F1Ki|;`OU|Gh*Mw>o?Kd#_PAyyj^hIF1TZ_
zB1-!N$9**K7aaG~d{WRqDVP!aBtbuk=2SsHl`!6^BR{O(pW|7A*H1L+)w^<1<5X^L
zD)%`atabw$?@RILC!oO{#hOI1q1b~tHXL8od404(3$+O&-emOfxEG~nMr@|}B_qD1
zc`FdN0?IxGh$%D|0kMeYY9LnA+y%rg!gz&1byOcB;0Sc^{NCZ_;Kyt<0Wa{z7x?~o
z?ZC$hD24h=0aqR$CZL$Kg;lh$Hh5;qHAf0~wfAiIs#sL;Y!Ye}2iBn<uKF<d$gtX+
z(=SmNQ^}|*nJ}_m_AsVBj2GbzK(zr-`kg?vljc8w>JJb`0<%V7+6cS|KLV;pw7wgt
zx@kTE{ZGJIWPMJCrpeF?vyVk*;lKHLom#9^BZxJE0`FB>sa6ncX|5AY>I74wuNO?}
zX>Jfq8fb15Od4r!67-vdek6U1AhwYFeHA0F&k43!gBtQ0w7UjFa0G$fH|mONR#8od
zj_MN}+!Z4<M>oi@m{k<BMx@Ww=Q!99cQP9C>eHJLn^jw5uLS-iA~)RcEHe!I<v{HV
z7h@NCX!ESp!3T8kfp|8v(o;U*Db1aHKqt*<LO_}jh-W7&r3(S+G-n6_8H9t1MsB|D
z@8OFoKhl1r2aXZ&ArcBD@3V^gWE6^0E$dLrO1Ori^oVtO#QI~FYmS<5jYr=*IQ+n;
zqgIX*rOS-RWkw>yFOBg?qd9}|$e{Te<8h7VV#c|c9=;`va|z9*jB_c?6^wJmKY2n0
zGn0J3igB)@_0^1XHO*~|bK5_8LK~gFopElb^&O0J2hBZ<vWMwUzSqkrdue{hDBsZ>
z4xGaQH9f<@gm6HWHv%|E(E5$Qc_YoSz&ZAxJRz1&zY93;qV;jWIgaLEfX6TNKI|~?
zI85^~;Bk!R<G|xM&1t|njdl;|z&V}ftHAjx%~`-X>z_O!3(O?$B^x+r)A}6XoI~?n
z;C%O=JmD^#{vL3?N9#*~a|z9jK-mcTlkYVFWfRTKK-o<5L*V?7PX7>0cu0S*4LG;a
z`p3ZeG0ks)^P7M2gg12hx4`)=t?vQOJv2u`k4PwC_led*=n+NpR_L*n=55ep8_frx
z%K<oyq)&h@2{a#qE{AA70hK3Ue`0?UDo@g!2$hL6Cqrd2&F7)=Jk1xN@&e7*VCXeC
z6+2r|DugbDa2Vkt=u$-UP3UryFt$U7KI#K(>|4<}0UT5g1Mx|%C{?h|6>Ko}P~#`6
zjPghAn~XeO96K<MwZqZG6zx$Ni@#KEM~ZSmyIj!03kF`yS-Iu}oP27NV82Nif;X?C
zR3Z#05$y5e{*ecj2(Fl?&)faUnIly`=nD@8{fB}f{=!WCyrFZ?8QY>mj?0YEWyYO&
zjSR*pgXUj=(J#QAjM`(s=ornB&?pkRlU~~jjkXe=*|TrxgY*cG|9S&IJHEYRwDRRq
zL6ou?^K8bIlr)Di&!PD@VE!9$B_%xt%umrA1I=TgD=BF#G>@e@3l7MF?zo?4|2}Nd
zEIXTlXhq6o%`@3Sc!fdT$Q#G;HaG@DTO)p867PZ|Em7Jm7;F{>;1v<?M$v&n6|1OX
zUGcyWrA}7a$+}^dYc4bJhd%#luy$o!vo7jAd)D3<2V4uLn_s9^R_UO{KM(lifdHz;
z(B~96pum?QX#1$X2_eI15399@HN^u_M5n)6y{sud82-42vdHCz;&@n;o^ck>2!A~=
z+h9On#YuEJ87L1l#A~=H1sMblAT&xoiqZue1XH@v^3jxTjKLU6H&!vWzrT68@QKwS
zTTvQkJkEwtE3|8cZuqPg1q0Cj5p*Zq4(-~Z8$R{<ct~jf1iBOMgm#_K4Q~-d={dB2
z4&4dAhIX%^8#!6)h4#JB9rL1HpW6pDgsw(oqK4J0q0a_uS-o1C>sY-yn(JA;dcyb;
z3Pt1U8z|g0`05FDZYirMWev#SP+vgd!GSNJQ0J<zop285%$e2NOLy)&R_h&WiX$vh
zy1`l8Af1UuA!pG^7++hV5<FlP4_I4L0`(mhF2Sm?K{uUddQCDErF_OJpY|CAj8y^6
z^^9FT<4tC617p`fb0=fhNplxt*F|#@uu7tRPBO4cra2GT<pFOp5AuOsKF#I8uAJry
zU{^u&9%!|P_$<_d&}uKuhoRkJ=uJGw5omXW=2Ot_6wRlh-D#S0pj8g>ETVKBT3x3(
z7h2`gd;?nDApBiUL-Y6S-CGJ#%HplFczYb1BbUfqXVYB3TNlt=$XgfET*O-!(R`D)
zzDe^f-uf2J+XVA%f;IMJqO@Hw-%j&x!F)H(dj#`6H18G6_tLyiFyBY>e!+Y{;SE>9
zoVHZ|{u6R|4lWJ>*x_x+O3`i@|Ml{ZC-S^$Iq#&K<#Zx6YeWn>q2*2MX>R3BTWLNf
zm>v_HHn<kXn$N9&BuZxl(=#-u2&O57H~zZr`|*jnd1&aTGA5~vH5sw#j72(Qj~Tmq
zG=0^#wRrO37ytzc>Z@9;5T*UB)_&HQxZDq{)(6%Y-`!$~JWB<us9>$|tV4BV9qQ>@
zg)Vkn7dr>*<(ivbJS(?`!i^m}K6rbHl8c^;3Gv2$az9-{{&IhcQkWqrj3=(0h^|_F
zQ|KF$6ecu<nN0NYU`RahB=apE_{M`knnUBkWWwRlAso7p`4bMk!r=&-gTmoh!dIcw
zRp^FSVze8Ew$;#?T>aEQ+Zvi{p=~YAb<nns=2qy}3WMoY!mk|$)2o7CCk&=n0l(KU
z7<+vYU4v__XLaxmxrnZsw8B{(d^d%Toms67tPZ}|7SScMRs^eq@1N0f%IfB_#<<In
z2WE9|(45EW=Fyzb>gLm2!0HyzT*&Gc(tL;2y+d;iYgxnEk?+;AmbEn3v6gi-*Rz)O
zH21JtJ**C1i$v)Hr+0yz<)VuZ-YSQ8Aa3Y7Z*`sKT;3{|<{P}#4Vv?Kt2~<Xd8>Sy
z3wY-O-ks!sn>WAB+YsI%nC%b-61Nm1n8nZ>E11R7yi+jSN%JnjY!}UOf>|8Ry9N8*
z!Vr>wykHzJSYrP9;f3W>{2%Z5|BeE7RIKstE4$e7(cc+5495NrPh~-_d7?p_i279*
zHW1PE^S~I+iJW~O;T@0gZutHHT^#Z*=Xn)&S18=!b)NG^q#-kf0hxjWt{5KRAFsZk
zqMp8NeFQxp!2mpi<eG;e_+&NimhR@L!L6rJMSkU#zw!fc39t)5{!sl~gjA&r4l0Ae
zIE?vt?BjBu9f!y4?gq!Kk=Ks3p&Tm(<4VB>?_74jo;!Pkj#(*c{GHt3o!l@Ir0(R#
z?&M|=-pw2Dro;5zyw`3%0Eg{+3dWqg7x|Mja`u^mK_=<mkDDli62TI0qV`&TFRwpy
z)El+R2F`K=7lZ?P6!&mW`Q$neH81D%oG>mjS|!vEVQ}TKqd++}LPaCp&d;Iea~Oa<
zkzDgG2DkIU1gFNe+5b`f*So0%y<d*zoLF=jHS1}{^E5LG7YX}z)KTi^Hduk;P&lX*
z2II|yD3!4;Wu&7Lvhyptwscmax~%6c*K?D|Wn3QTl*i2?{n5oabrDYJPTEv%wAH*X
zI>TgC#mq!p1W~%jsP57Hm{C2ZxrdQ@nCZkm4XDz<L_F3+=?YL?p}82SifL{HQX`m7
z><>Vn12Bv%KL?@DL7GoM(-U+!m<*+4IGxxRLZ3nyhL@|u-;P-JpA%o_qc%?D%@cV?
zJQ<D!R+$H6zdi7Oj=0}u6if2o|2-I8GgL6D3T8a1Y8#_!qd6R?!U3g^1*%w@vw$iK
zj3-sS3siS$eh5?#X?_D#ZwR0HE}fgS_-L?%<^-cEVf;uzN|@jhW&+_-Mpa7dOPSzO
zT3^Mes%U)`6I?~>s~J@_t*>T6s_BrUol&*Z`gSI`olf7us5)qU2NTjkr|)G<dzpcx
zyuFNLFXKk|9b@{A*1uyM-_iOApo#!g{Ubnd1fc4_5vVrO`i&rDBcMX2T|l*q*6#wr
zyXf?BKov*p<3LCpojx6?(t#i8$aD~#4ki%33RG8V{Z$ZrmDXniRW_~92Ep01J_o3B
zXnhU{$)Q8Idq8!M*53oc_vrK`KvhEPOF&2ooxTZ}Hi3bpyiLHd3Ahn%2Byulz8N?+
z)A}}`YNPXS1Ho<d_a6h*V_N?hggmCd{}!m;()za`_${5j2dH{zeGdrfq0=Wop9FHW
zB1#D`Bmqt)KHw1aIYjFZ!H`3={v<R#2?rAUlhE-bbR(PyO%rK-B6Lio_2;4Kd0KxS
zI-aNX7oh0{T7LmLUZC|w(5DCnU>}9<FQLy(!e=5qYqtGx_Z;%#n>fo&oCkR_QOG$J
zawCZsf5|z$Bz)H0XHss$+`Y)BZRgCkbN2Y`5_xsOpqQ+p_`N1FY-AOU<dKJ5^J)_Z
zHJ2WLf5~V|*Qx*c3;0y83;TY!+dc+e)m94nmE>?1?;B8OG(klZ9Uk;T&t4dSHzw%7
z@UKU$I6Sz#p#El%ys!bqMy*h{6`JCR^5c6f-FBKgp>8M5uc7X1!uVM%>a4r0;x2u^
zsD2!aJ1gDSO1F9QtDn(Y7s`ez@c!~@e(5J?hd*!mpYOwK9%1vjQAZ!6u1aORQ<;%u
zK}kSYUG#0$M(DMX+>T_1Mq78>T9%EvDglHh(9y|J5Ox&IAf9m}4BZGP5>F8g!=m8~
z%$bWTPcIu;ehg)l&v@q3_m>sGr-I&<AAvqc$U8^0D}|o9P{J-DCvnn}zQI}>|Ig^;
z`l|)gl2&9LMty#Yu{p)~;ZPC#4AdA6tfGN_r}db1dd!Z%!999%_t!IOJc{y)?-muc
zy~z5n!`i$@@#_;Fdfi30*L8weN3Pmgsa_E4X>Jh22AUfMv61E`L2RPASrD5E<0qMD
zYo>nRjq8q{k}+e_m{8LH*BHNRj6dPS!0#~dC*67+j5!WM2}i*(Q7{yDt0--Qe%qiw
z=0APy)`m?q+w(s^qwdW3#`w*u1WPo!Cb5$g*n{Es^JssfemqYaejlr~kL^be*50#P
z@7aF%882=?lu-Sao+P{v2i=Dwan2u~`wl8+N8&+R>e617GHvcIbm~~jJCyQXWY=87
zOEvsB!p;21W_|+UBZ9*b!3*!IMd>#I{dXMUWMO2oFab06=4hL!e#(#CdHIWjVPEar
z67^qC3;bIEG^OvcihHyhSN|{oPr&-7^&?|-ldF&uDPTMc=r}qVcqY?v(O&4e7y98=
zMMqiC^9J<8PNc#6&%oDbjG6!WVT!NHP>YqZiZc4WN)78!L%&yf$T~fw->ay9X@Q%*
z+4#6~%mbYvD5E=^(;d!*L|S(_r@J)wa85lm_i|3Xgz--{P-mC0iW1fw4`jLKha32d
zEw0=PLB|L^NJAHu+M5M^JP<#IY}&<wK0f%st2vrFAAjCKo(`pRHtF01QqTvS%>%+X
zJ2alvKl;Ek?!m3!GLEkuKM_^6f$?de518|TPd>d=?SY<q$YS~N!o>4BS*lu{)6Tk^
zpFWCeaGNu@%{h{ybaMvXgdc02*#Uf#W6+9}&WuQB#t?UW6bwHKf{FW$hQp&_Fdo+(
zJ+t17KM<)xMNI%E319#*JPIt10$aixp~*%#04EfsXlM})Z81MlS>9YR_waR8)K12y
zlYUKJ4t&bV<t?rPYBlu_U~nMabucZ{CE3d6e|}2gU;3b?YGxJ9^bSG&dmlW2@d*>^
z)k;=T$@=1vE=nD&vV+V$<nsUeF%V9OFAz{0|MUHHPtS==m+Cz)(4rH^YQ?cucs0QX
zo+yR-r$E>@zhh5{7xYpSKBnN<IIb^Vs8)QrXde1M@{k3c>C<TNblnUFpEu(71TtJ2
zd=@9e^8@dtWVq{BuJ~f*j*l7p;M`@h;pEQ|<WFz%=P>f8ANezo{OSAmpFkSd95%b{
iO?r5rfIxqr;eNg&MtTpEjr`j{_V*o}Oy*Di?f(PS&yqm^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp.3CE3CBFFC33E63B5.idx b/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp.3CE3CBFFC33E63B5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c279fc209ce70aad4beb16a5a0844674ae62e156
GIT binary patch
literal 9496
zcmZvhd013s!^U~e8F2&@k-!a<RbF@J?E5+DAg^O(xHT><jtiRDqA8}nq>-s*O6F1~
zX_hUB<*TnIwpc=oW>z>dWTsK5m=&*CMq2P4=KGFw|M7WUKG*!t{c~=w=WNfMGiKz7
z5uLN6qJ~c&xoGZ;ho`!tqM{P)zw;lSH#^;i*W;q17H$1}!sOAJ;w{&<T{?NpLwrS<
zSEACVzC0+4nxL&sxp90=(mMU}sA1PHFVVVE%O?!{a{SI)1CMU|*Dn`ub?<z9%nKJ!
z&l_2pcz*UX=2H2)$M0VA=ofuL1sBit`TEa?7wv7owD`cbLDxFn+BI+Gk`Inv>Gj_!
z)jwSyL_%Gkz1p<-U~{p0G+Deld*OlS$~q-{y7#Spxd)mTE)>hoO_@<}=c;!{EKOZE
zIq<P>GU=~xKC_G8^>g%;k2{7YzojJ}DBpc?RL8Glwvx{`<<CC4Awlsis`~rwbHV#A
z9sK5(tmEt4?_HeqgZSc`AGoG0f2QASTLurk@t*}J4*m8{^xGYGGP$QU*UF>se6f4i
zqC>B&i3ygRcxdi`yS~5udFx6_x3?;0PrY*K<mNB9?zK}+51-X>{R`W@c@s<eZJTo7
zV7rljk7$=arh~lhP-5mw3)2=^lOFr>%#~$-EX_N;YUuX;`!;r5_x*QdZd^I_Zhgs9
z8UBk)V%{76K)?NmUo*y)9UJ#q{fcK+tlrRb&Vh|jzj$C?(V7X-vp!0_Fs*$HIZ?wG
zJvl17H0j{^H&^vdEgQP=*IxAK`Ryjv+%o5*{YRc!q8;j1SXcXHeY9RxnpM)6SXbLO
zZd}sG0c&!_g)zU>o;|s>|F?7}?I7<P)8*+gv!1xTx^vl3Gkf<Ceb%I7jStrCeg8k5
zu8tnPcJ>}_|9vGp&USrh$-p<hs#x=Fw?p+OvYICCTXN!RN+WYYmU`%ex99B08$MQ^
zpZU$IKfYSJFvA)@r{cNfGg0?^wcwxMu751}=f=PjYP^}zZF%<e@tHg4lc#F=rPH)Q
zkBwdN!-D6%-jWmC;dhoEo!PYPDQ5Ptgg0WFCO_r==*Smc_dRq~|FCI8`Ta9@YzWSI
z_>JBE^iz{ZZ=O=INBK(H-7o3(c=!ItYhUsX{6`nI;N{^H`V}49Q8e4VHHG~tVeFpz
zE8pF_EHmk9(UxDY^?WJ5VeZw6Z3EAC*tGcUq2sk*UYq*rn)np@(<ha^*`M6I{CLdk
z<#Cq>$25nOb330N>F-az?mzBY+JZhySI<~+?AQ9EzsO4Ai-Ga=)AsH--M7c8Lw({t
zdsgn>@A~EMW<B`9rek}?oElboCF!0IZ+YivYuSaosn5O9T76`B_Ox9i7azU--(3gi
zU&syJzwg!ERhtIvx!U9_n3li(g%RhD-^#sm;rY<}C2wxu68d)W$`jKTJv1+2>*3(S
zqYK)f`S{5*bsa0S-!=pH_1da@ou^)2?SFd2jTxRE`xlRRx#i3(uW|Q+wU6}jT>E&l
zlGy*RIA6tUf5e_%XPu1S{Yk}@U?^?Nhrut+p}UT}^!w;JH*%gk`OuaJ=En7YWpMP)
ze?RO0A)j44=~V4|c^{rN-+txwl`V_^(l71a{8OI)s=4L;_i{e(JZNINtL6Nj)5Et{
z4Nv(@;07Q5r1|f8i`MpjV(SN+dn?`tCp>EscF4xmsQHh~4a8GXQPI5*t{wi8m=;S-
zz9*_6DLSb`c(AV};Xm%mXYSqJQQ8z2b+;XR+j%F22PclXzWr>;D1K}#HO`KEM)ynz
z4^AB0=Rn7b`b$}{)V+4xFS_5I;lYU$-df#b!J9u;#Zq_JaYA%Lhw$LU-CqgR{pcUx
zE0(&`j#Hyk<HLg!n>Vif{Q8AQ-ioEB*m3vh?(M^a6W?)c<FEl^U&x82{C1oXozXo!
zIC0?f;H+DI-}huJHPMa}qZ5<EgA+gd$;;on*;9YSQseD7JvzNxcyQu(S1hh*zK`G2
zF=~t*=OpK(hX*G<b@%$!Uw`wQFP0i>$GxL_cM1<T<J!mc_TrCvEOnP1r$wj5g$F18
zacbJ3f!xn*EH&DWlcSS6ga;@7{bEMmr}3YykEJHraY}Sbe0Xr;zy3ZlWB&E3sj<`q
zJI;#EN(c{5{CDRGd%u04`a&#~Z^vDtyL1Q-P8>CF#?1M(sgdI%JtllSJbmQG_KJGd
z{-1g@kLm1Jo4lU92r(K18RKOjV>K2s*2_Z1X&hvnmxGMgc*uA!51F6|kO^J^GEoyD
z6TKp2k|seWc_qkXO@>VN%8)6V0-54fAX7CJGS#a>rfC{vnpcBN*L28quMU}^8IT!X
z12R)HAv3)uWR_+@W_hifnR7mqnDBY*>{jPI_QZtGUuUzue@lty1rpiAo-*Qj9l4x%
z_5(Aegh8JvBMkaXIbqOeO9_iUTSi#)*>b|7&y^AmeXfjf=yT<SL!U1tJo<bY;nC;I
z36H)|N(A(UG9sWaloJ7cv6P7Di)BPaUo0mg`cf&8(3i@HguYZxB=qG{BBL*t5gC2C
zoXF@ar9?qrDI*H{N;y%`S4)YCzFJ09^wn~rqOX+_4SlVQXy|L@L_=RMB|7?g8PU<#
z%ZZM@QA!N-jWS}OZ<G@QeY2F9=$mE4MBgkYCi+$>vCy~5h=smYPOOOU9QRaA_`Vf6
zEk?RWPIOKs=PqWSJ8ieHh~)8kZgalDK9ll2<AD60$q{0F446zl1195Vz+`<am~1`^
zChKRx<a`{MTs{XT=jXuWeLR?aJ`X1E=fM<w0+>R+0H)v<z!ZHVm}0&Nrsx;JlzbAH
zQoaPH<d?vdeKMGGz6_@9m%&tg3Ybd10;b|uz*K!Im}<TXrs`M0)O;G4TD}IR=GVZ~
zeL9$Wz7D4D*TFP=2AD>^0jA+Mz%+d(m}b5Srs+4qw0strR=x$M<+pB{&gv>5CcLVG
z!?u!mj^HJ|k}$|@C1H`dO2Q%Ym4rtYDv5wBRuU0esw5J!TuEePrIIMfY9&#TwMwEP
z>y<=DHY$mMY*rEz*{UQ~gr!BR?%%Mi-wtB0$*i<dr~YrL<}5as+c(a(r6x>Qi1O4?
zk#pJ8NO>+HH&dQJk*jD=H4T1{_MAd)pgrf2Yh0d_F7R7jo@MPK%!DWg`<X_H!G5Nh
zVz8g7q8aRGf;5BuOaskeKU3pku%BslG1$+BC>HzKMvBFLwwYqFpRJ-<>}P{Ci~Vc^
z&0;@W<6^O&ZFRBO&xI%s`?*Gn!+x%r;;^5qqB-p6f;5NyTm#KvKUd@8u%BynaoEp?
zC?5OyMvBLNzM0~&pRb~M?B|0tkNtcD&0{}b<KnTOZ*}q5FN7!o`-MhIz<!~b60l#W
zq6O?1g0z7BLIW*ezfj{6uwQ6(3D_@&C=vU`MoPqfv6&LFU#y}<>=%Qyi2Y&%En>e|
z;}Wr7Y;}p)FNG)x`=v%o!hWfllCWQ@q9yE?g0zJFQUfhvzf|LruwQC*N!TxkC>i_Z
zMoPwhxtWr&U#_BM?3aVIjQw&0En~l2<C3voZgt7nuY@QC`;|sY!G5KgQm|jCq802{
zg0zDDN&~H6zf$8;uwQ9)DcG-uC>8tFMoPtgwV6_}U#+55>{o-div4N>tzy4g<5IC-
zZFQ;GuZ1WL`?W?&!+x!q(y(8vqBZQ-g0zPHS_7?NzgFYYuwQF+Y1praC>{ItMoPzi
zy_wRnU$3Hd?AL>|j{SNAtz*Al<I=HTZ*}R|Z-giV`;A7*z<#5dGO*vMq7Cdfg0zAC
zMgwhNzft2du-|BP8Q5=zC=>h5M#{u~vzaon->jld>^Fn7iT!2+ZDPM!<1(?|Y;~E~
zZ-poe`>jUG!hWlnvasK(qAl#Vg0zMGRs(Hezg6S1u-|HRS&{ysD}Rdm^S6gj6ITKL
zCKz_x!bzIdeM4t^-S74cfV;Ju0b<=O5a;HAcsCCuxCJ25Edohy2}pL!K#E%dQr#+$
z=GK68w+>{u4ItBP0$FbBCi$K}yzWHFPaR!x1&OJN8O~F+xPsJfsj1EqGp-;#Ej`<L
zq!w3@k(QC|Jf6EvT>n7wz%`{*TtQE%=WypiL|j2{tM@49IhrHeYt^;~d9e7ZBc98+
zR;?opGFwMjWUh{I$b22)k%c-UAd7WGM3(A^ge=z)8Cj_#3bI;9RAjA=XvlgU(UFZh
zVj!D!#6-60h!x>+w|B{Dd`4?$uT^t9<i<NcNS&R~B_{kx6K1Q}vlb_-m_cNVSwyaw
zL*$EjM4?zf6pKYfsaQgki)BQmSV2^aRYa{=L)43PM5EY1G>c6{tJp%!&B+y=Nm)3`
zd%VA2xWk^SoMAcRoCmne|8?;355KFgY}-5`=PqY+$!00_I4OSAJ}R>&W(AyQs&NI`
z6SHs9iqi+SzP9YmH`=DZb8e>7wDP~XDe@O<df1y2QWG+r#{;X*yJNrEJD|bdoYgH$
zcD~X%E9^tN?I|<ttZL$^#l2Nc7-Y7Zu*h6B;gI=i!XpdSL_ij+iHIy!6A4+aCNi>8
zO%!CcnyAQHHPMjuYN8_>)x<zHtBHwhRTC@1yDpbJGO>Q^T6?zocj({AxqwzIdZ*nT
zb9aoeH+S#UJ=ytHw&x_BR&a^_(cYXZ=jzU$aQ3D>w{4H;VQyQ9XEppLV_FD<%(f5~
znQI{&GT%aYWTAx!$YKi-k);+QA<Hd9Mpjygf~>X>6<KQ`8nWI(bY!E27|3P|F_Enn
zVnz6UZp`#{xjQb~vz3&cG|+hibYz@y>7L%9PwdTkuDlFq5}j+XD<=F7BQlBUebNUx
zN3VTN?~^gexlG$MBtvq@`3Y@5L~DE10bejAU2wkO96#mz<@?T4c6mx#N~ZIU!+8~A
z@4>eBAd!FTVLZHZP{b853<Jb6ED*<VKs>_(35)<FG9r+~NI)_p11XFGq%tay#%Mq~
zqXQX?0c0{Jki}Rx>9V_G9i{9kwPzzaJ9&sR3(l<2G2u5pk=aP?m6{W|*{1eQWgNP3
z<JH51&usdw?LuZ`-RWF)><gKZeUqH6_AIx(FN%Dxr-^v3<29&>Fvx5ZVUf8e!Xfib
zghv*dh=43M5fNExA`-IPL}X;8i73cw6H$@1CZZwhO+-gFnuvjHHW3rqY9dyIomXXS
zO!yU9WLgq46Z<$fS$o~nH*r9OJOR&)$bso8^gIAm<XIdcCcuEn6f$5kMGTm1fCZB+
zWWi*MSTMN&2PRj@fyotdVDbSTOumo@lP}`I6aoU6LZJYrP$YmU21GE$LJ>@{NCZ;~
zNMK5Z5|~nv1g0F2!ITSSFy$f{OeLUzsT3+;Dn$yIYCr{3EmXl&i&QYRfCi>ksDY^!
zX<+IB9ZbDY2U9Q7!88H}m`0%ircq>oX$DL%%|a7Qv&aO~3RqxTg%+4rk#*CYH=}TB
z+pFIB^QO(er|m<aZ3zFN3;x`8Y1v2D*g5TIO`S14Fl}bk3@3|<njZNW{(}F-X_2in
zBOhl)K2DoEy|4{O&Ag3vwTo`=?hqT-v2%Q<kqNoO`y|AS9M+|4;w{~}OR}ozhG`8-
z>aS!9|LFbyZB6c(lA4yDk(Hg(x8H!lLp&@m_TqXBWV}hqdA`}yh}(z$lid10s#ow<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp.C2D6CBC7BDD03D15.idx b/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp.C2D6CBC7BDD03D15.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9ee5aec10f6b9ed47e7f4296de854b627d3cbc05
GIT binary patch
literal 9496
zcmZvhdt4Od!pC`LRz$!Hl9<L_kz*>hbHC513*>4lUU1c{<Eo)2EK?J;6KPs%nUeQR
z%3H_UG)>E-PNpthG_?+`E}CU^P*E%NqZVnwJM8<8`}^bN^WpjIcYa^?2XlFLcxL#(
zf`Se?F);%s4V*J=@~jE2n3$L(`+MfB8B?=ucr!63X3oar#p8x{6Yu)%^-JFzbrWw{
zw`(z36Sn4cr;4@Zske_U>$F0@GN}KpD^F>isRhM-z8>|)B>yw7KmF^)!EtXa8vg3V
zGcyKOC7+-AviVQ>heZ!9d*-Vi;gXBzdYt}a)|@@9|5?88^}OHO4engH_^E?OuBBgn
zu=?j)c_iF<)%AvT`y0#EBPrtbsk8U3uWXm}#hzVz^Y=B*o-J0MdvJ2e=q3Ly_-ERR
zasJP|<47-k-PyPKw||L!@Uynzal5qSeSw`92emyNzmXhYTQv2^>LkTGC-~O~=R*Iu
zwEx>*yB}K-_wmKC|4kk``tQ^`2k&1q;`q8vPoGgr#@`yh@Z0Byzmh#I{n=pMcNb5M
z(#8)jXmz>Y#r?wCyeGB~-ZZD(veBpB+ueBM(w8M05+BX`CF{M|BM%JQvUNwlHAC~p
zoqcv&`POIu_ST|}h2JfGd6?Qk%k0x{%CB<{K6RkmwCc>S&zwH`eBl#U=L@l;$oE$d
z<tOd9u0QwvW^v7#b0abWLna5#Up@TAVb9T*_-~IE3>{ac7WAUu`>Eaacj@Msl1m%Z
zPv7YHMf}+rqpMdhKe5R>cg2G3RaBp;FTDTL*dIUD)5j+YrH|H6U-;UD`&ah6(Wdmu
zU9o}qiSA8(if8sLd$eKdi8a5H*6C-v=JZ@z(31<^^ej2C>GhSx>(jYDlV?e#V;&y6
z_>b+mpL4zqW8d#r)bQf|kJpB#6o2+u;&;<aR+jBNtzVqmZdRA;si`l$*7@t?|7_j-
z*2b!rr&ZscVyH=*0(|k{gvaLZs#mr>Hn892KfUf?<>HK_Wd}R%3A|iWH9!AK*8}eK
z&WGMQeC@-**<1RRY&bXmz@)up2lHRyHZKW1_xsNLMgAWzzIb~7+#y$A8GiEk;MFVV
zy_ENUr(WCUUAeUK#ZJY!UAK>V<M1W2dF|KA>Ni{E`!AHv`oepD$e0^>Z5OPX_`L3q
z7qs-$#Od4oO+!X}U*7G;f`3nY?dIN3)kDL-8n^Cr*E7bk)z=nGj2Tl@xUcT+-W6lW
zs5KM*$|ucZFMjl%yUn9HnNK};G=1BNMSb7hF>H*Xoql>nyH_87H?ZgMjy+q>^z6Fi
z!ybuWu9AD*d+W*%Q>GtWdvy2kll?2Mb$a;FT^}83uDnn<Vf|ap)rS}4OniIbydy*2
z?cBHMLVozMz5m)7T-$s1^#*Ur#G;k27MwdanA>*YmGFUu?`+x-{%+pl6BFk=F(YZ?
zC!yI#o@{;gvlq_Rwynzf!1VtkeWP-^P`$F$_fo;_$?mT2&nwv4baslzc<9OHbJN|w
zeYQ?X?)7J)cgObM6V9x#zG=Vn^Bvbh;mi$(LSLKx-afqf=FmrP=dS<ei4BiWOU&5T
zH};KpSNZ;1#4aCu^3=zLhrTyI*!Je)rg?wrmv*lExzKmr+;HII+~Xbc#$>sg&hI`m
zU{i2F>X!o7_mj^X|C%vpdB*b_53b8lJkyI;nS?dhq|KZ=&EKAiiHXhFzkI-EF*AW0
z_i#)}r`S$yB7;4bM84v0ZTf!UAbvywHPVjniM=N&GB|P4uBBa{eCKj7ff{DVNwG<7
zB7+lm+2*gk>>HJyK#jKJwAi%vk->@0+t+@1^TOO+3Dkpj+$FY4>&W25_h*fGZdYpS
z)oo*j+i`A6ZdPP);$cVE^zS|5)!YQiXUEyG*<B)o6OZiIetOCyOI9RMf4Ad%WA7au
z8JyUEJT&F5n|ohKpvKs7a%^%+WN_kDpKtvsjy?H%0yWBxvtqOEjtow`Vf7ykH+Azr
zi7^k^v8R=1Y-DialMk(2divY--UMod9cRR5w2KUP;#1G+t;NfF0yW-_Gh;IoBZCuP
zo{;%rAMO`6ff{PZDX}SSB7+m(yqI11Mf)#TCQxJTI5jr4ePnRrKmR&BdFHL)gaoSC
zj=RTpPl^mq{8xwKJ>NZEeIbD=vg3}i9os|(Cytph`H`8Y(xT@@Qhelmc<J!%O*?*6
z?cdbxaThxKHnD{9FhHz_1>!s$5bxoE1djkDdPE?}BLT@C8A$OcK&nRt(mWcF?$Ln^
zj{#(QOd!i+-67|EcgIK0bNI6>i2GF%{lndr#Qi36fVkfWW-17SK2u2;^qBx*&}S<M
zi#}UPSoGNdVbSL*2!}pbNjUVm0O8Q*D+rH1UrBiM`2gY37b=K=zEDX7^o0Ns&=)I+
zh`v}!MD)b~5z&__h=jgWNhI{80Fls_D~OD~TuEf~<p7b<S1O2tzEVjP^pyZn&{r#n
zioRM&RP@yVQPI~bh=#saNi_7e0MXFbD~OK1UP*NH^#IY)H!6sMzEMdG^o;;9&^IfH
ziN0A$O!UnFG10dwh=smYNi6iO0I{OJGtb+XMav`)mxb5;09+JB?vX$~_qZrAUIt92
zhyj!FF<`P@7EHE?1(Wr$U~*m#Os<Fnlk;(4@?IWHzK92t_wir~UI9#@NB~pt31Es|
z5lpd21XJ{hU`k#IOsPl$Q}Rh*%3c{vxkv_6_Q_x>UIk30NC8vvDPXEz6->2A1yl8@
zU}|0sOsz-*Q}bzH>Rug8y+{XB_vv67UIR>{$N<yu8DN@T6HK$n1k?1HU|L=aOsmKO
z)ACt&Oy~Z?{t#=q`GC{5intHsY+XedWVVX1$XpfSkohXYBMVhTKo+Zrh%8kR30bZp
zGO|)d6lArEsK{Cs(UA2jq9Yqs#6UKyh>2`f5i82loF$K~o`3HaV&8XInS)OL-|dKV
zJGOwtNAAO5cZDf;Efu|%-Sw3F5^^Ku{sTEkyQ^vNL$v!OavkkHk6hz&f8zqb+2x+!
zD#}clVz8g7rx@&K8Yu?*nIO$zKNF%E>}Tp|2K$*B7lZvwvx~ugHcYYD&(>2c_Op!?
zi~VenX0e|Q(Jc0}bu^3pY>kV>ezw`gVm}wAIPB-@DGvL&MvB9JE=Y6O&xL3X`?)%r
z!+x&D#bH0!?BcMW4^uq$^Ys*u{d^<EV?Q6HdF<yyG>`p!9nE7uU*qDjpKo^Y*e`@B
z0sDn|O2B@hkrJ?92+{)f3n5y-exZ&QuwSTg3D_?*y9DeP!<2~qVm&2dzt~8L*e?cY
z5&OjuEn>e|M~m1m*0@CM7n@xo_Df+(!hWfqlCWQDq$KQ@g0zJFQiztYU#g=e?3Zd>
z681~YE(!bPFePKZTu;f^FE>&$_RB$9#(p_O%h)g1(K7bSH7*(Z<z|<R{YscpuwSXC
z6zo?TDFyqLAgy4(5~3CCSL$d5`;{7(g8fRfOTm6MOsUwf)>A6>tBsV3{c4a_v0n|*
zD)y^&w2J*|jZ4LTwb`X&zZRx6?APik4g0l5O2d9FNNd=yg=h`?wK`hEeyzr(VZYYw
z(y(6-Q#$tR^^}hNdLyM{zaFG@?AJrIj{SNatz*Al<I=HTZ+7X}Z-glW`;B_az<#5V
zGO*tW(gyY$A=<!xqmDMP->7jJ*l#qu4D2_<l!^UjJ!N9Q*+`k#Zw6@-`^^w-V!v5O
zo7ivGxJ>Lfn_VXMTVcw=eyg6cu-|H=EbO;}w1xdvh_<lbs-rFJw`yD#_FK&^E85?0
z@y{`T+?aKSxJm}b4IJaFee{1|_E^p8vbux4CJ^WD4WHI=3=kW~0&#I15Ff__32_3D
z7$*WraT1UmCj%*Q3XmG70%>s?kRGQ48F2=X8D|1nan>Dj)}idX$d*MYxGicT?xk?s
z!!!{FnQbC0GS@^nWWI^;$U+klki{k<B1=s~LYA9|jI1;f1zBw(Dzer@G-SPr=*UJB
zF_6tBVj^2j#EP=_l}}cjSom{WS7J$WT5`6tpq5yYc6VBubLX8{l9id2<E*A7mSksU
z=Qs;@4~Q!tPwBI)f=Vp8N4jT#v-FZ!l3`^Ga+c5>+1~jrOMI~NYl-^`-on=s2AQoT
zEHYP1IAp$-@W?_f5s<}NA|gw*L_(HpiHxk&5(QbUB`UI3OEhG?mgvYvEisVIT4Eww
zwZw|DbG1?NkyXv;Zow;b<dBZJ^)G9!?EC!uHu>$H3)eYP9pfWQp3$Fmm%Eqa9xZ1O
z*>V<<E9VgTavo7A7ZAmA5m73a5an_iQ7Kmt)p8Y4E7uV9avjkqHxSKo6VWQS5c6~M
zMQ5kX9^^UJ%O?!8_f>BH+>y>Tzu@Wpiw^zpV^z!M;@t7h=7sB|v}2_F8T-`jKBl|h
zS#wP+$r+P#hn_pLZ{zm)@4VHr`#a=!bD9=k%}<rTTGrLxoRpT-%~@Jmay~BM+daMO
z?9JWp?k+oT>0CPY`QNe%4F{{5xKH8HswNCFTTNJGu9|Sjd^O>bg=!)oi`7I#ma2(_
zELRg5S*a!pvRX}4WUZQL$a*!=k&S9%Ae+_1M7FAl6=i2<+W)GytiC(D<?SmA=Z^Vt
z<8pf^_G;6sopY7!nDbGqVbiu0*qghw>yqO9qr3YXI<w>wec9feFX!vddmiA1PimFF
z<%+$zQ%<Ko&J&WuBaKTBXM{huHy64JvmM)+s9f=pHGbzeu_wMBS$WRUvnReD*?G=f
zX48Gyl0(kF-}a^4@+bh_;Jz%ud4qHOgFh|Ud!Dk(Q!`V$InMx`M+I5&kw*m4pLH{C
z-udh8&u7hTMu}k<$QX@*jAdBJSdE2@V>rk-jf0G5c*uB-hfH7u$OKJ*Ok_mJL`{TD
zVkF2UO@d5jWXNPqhD>1;$P`V1Ol4HaR856UV>HM#O@mBlbjWl~hs<CM$PCSZ%w$Z+
zOwEMMVl2ok&AO9aiQBP)Qg&C^M>{1a<v!;qJ4c<4k38{-9__UBwA|<?Zd%VY#-ZD{
z|Mf}Vvukg(OjJhq(ax-5PgF+E9dfqXb4<(gqUig&8;JWB&M^&yL1r5Wi_A3;4w-Kt
zJhIS01Z1&+h{#d{k&xvEA|opeL_t;?h>EN=5Di&xAUd+qKn!HFftbiv1F@p)JSt=3
zBag_UyCu0>au4T2)V}%YncO={ZohkS^ip$|x*rEBbI*$s<7dERN*OSjG6qc6&w|O8
zvS6}hESQ|11CuM|z~ss}FnK=@CSS^f$(Qk93Vs1hp;Q1<C=<XG{UVrRsR*W6CW0yX
zB`~E@2~4R>0#o+OV9KR3m~xp6rs7w?R7w>vl`;iP)vto7ma1T?Wh$7OUjtJs)xgxs
zG%$6)4yInJgQ=J4U>be{Orz8Q(<n2*H2o%+W~m9LS!RN1`7JQ5QVUG0%(`RFm|Qxc
z<x%g<853td-15TTGDN=Ul0LoGXT~IcHKO&D36m%JCq5E0*~wyJCPiOH-r#?6Vsz^x
z(U()AFDFi$RN8`r9(jOvwTf*W*CrvcZHM;l1}5bX=#dmZuz$zS$#>n|MUqubH%u$9
zQ!k~P@VAWrZ)?guscD&6+1+z;d*0i-?|p8T7t^_}eHc%tltS-Rs$fXJ`^n({0g|F~
A?EnA(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp.BE70F5327F5BF3ED.idx b/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp.BE70F5327F5BF3ED.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0b8a348b57f07ec9b11e5f036a7f10a9c1416a5a
GIT binary patch
literal 9496
zcmZvhd2|!y!iW3Lgc90PC}oGPP%5it_I(COLBb*hfwEr$SkRydXyFqpi(Fh-3Rh5(
zAcBfS6fKC<>s28E1cmFZl@>t}EfR`YJcw4%_D=JCLw<jJoWpza%=<LY>0~nVc4o}5
zAw$~cB_s@<HEi*`*^6el5)u;9?EecFEts2Q!y73H35&O$96$M<9%B37UcUV02@mm=
z^!OzqXU5I}J*n~9y6(3=U(;c|er5Qen^zvwI#J8V_y1ztYdvQ^@$%z8UmD?l?a47O
zUHW>#u!{7Hb2pev<?o)Hw&sa{_o^<wbiUV_KNc-6Z@IMez{>-EOC8Z^=<3HlIsQxL
zwJDLGZVn*To&IsX{`rFqrRwpH;`O<Y9(W;`n)Ye=-hBlJ8XkRA44$7dyLiG=9}ZcX
zv3|1usBbdqt3QA4P5#ZRiBpcYuAaPCOFs~L>(cPnXOgy(lbZ|Y9)B)P@huMj_5S&)
z`!63n`*Y9F*Sr69Y0?jOZhQL^*OcWO`s~_r=fGS0A362GuTxf@{ri{~bLVBQ2-kjd
z>GU{l>X;!d>IPmqC~O|^@U9Wt7pJb7aC-kc4ZmLgtawYx0|Tz+>`y#Ca`cX!dj@X0
zXTaohD_$+#x#Eu9Pi`Ih&C?r3t8KOH{sZUyy!eyHKI$<qlKsWkXHNWm=)>2R35ny#
zx7Q98r0uz`uln{Cansl5$7Y2_%?@3>cKFl7-V<y2vnPh!Gr2+?(wE-<ed_gh>86C@
z%UjeVueJL$>D+<|k>}Q(+U{Gje)+Bns{h<)-ur&ig(G_A)D$7`Ky?1fZ8PrPIPlk2
zfh+A3LrF6|+xw4S*t_I``njhz{Y+YBp6i;|`{^OQx$q6|Q>V7Sym9;snOy(bi=@EB
z>62Ffu`B-=SF~l)djkvWpFQ}m%~f;8AAK<8oB72XOWr!8Us{s7sLS>4-PdmG^hNrA
zcD}NEYsH3nky~>NHEnx{A3q}b!DV}+%Bv3!8(8<J&l3)=&PrSJNxSmUhRTX%1y{O$
z<jL%GX!qe?J{Xa^V_@-?O9Q5tpHaWtF=ESTZ{6Wrr*j9rd+qv>z_~w1T@zmmz4YQ8
zc{jRFy}$4HyKkKSGOttLJ)^i$Z~9h!Ne}A($|H9aCO;GT{7kfL(}`etbZS>B|5(an
zk@kVbn$-EbyN~=CeEq|}hD7NlXSFNE&z7usZQJ_;_VN24f1&pe^s$8U9qR(eJw@63
zs8FBq&1b2`_fD0}o!RcR&-yKClYV!w)pY;FSIU<zF29wTA)MbaY4(v5-!0rVC%bOT
zUH$I)anBpky(@PfB9r@sedCe`9-XkX#n_*EHI;11S~cVK#`NTto7cS6>vEoYWZH+`
z7duV;=gBPxQYV*xa_Zy0!L<d$9_;!!%l@ZqOYyW@sUNJke_F|v3}yX=BgYHoH;y@*
zSDi7xTi*pU#oEp18lT#|WPaVOtgCzS{e0J}&mWmnRo7+h&4oMD{#g8#{Ns{HO}FWF
zBlcWl?|66Hw|CAlsgdUX%za14PG31EaqQqG*IP-ePn&ODefe*rj?Mk(J<rmzk2*hh
zbd-PX(38^r<0k)}+2XOLa~pSL-n{b5@2|c;tvqYS$)-6sf7`|lJ~(St+lJ@E_qSP|
z+UDVYXTSLH&&(fw|LlTX+Iayry)@eP(8Q6AE>_u;k+5*dJbxQ1At5pA;JU%Dh}p^1
z<mm~;9TGdViVgO;B=#@q=Jsz_4(G=vQ}^1_-4eT{#Rg|OZST`vA9=ejoJ@_jr_&PC
zTEzxuy8X0^4eHh#<tYi%>}hWc@1)q^Om}(JUsG2!E;E^$U{7ZxX0(Y7&a`>!m#c65
zxMXiKHN~FplGvqXY;dMWpV&00-`JP(lc^$mIyW)5OKfna{U@vDw7;?MnPh6BJ)NGI
z-Z3^f)AM&%S{;`Db*Od17<)RuV}4F-aHjupZ0GlG_N(8Msd4snPGU~y*x*cmHErY5
zXU@LhOQy!!(^-jGsj=bq^yw9POR-K*rl#7{*@@XHvB8<XtoD}bS8dEqrv7G6_etzC
zAvQSEbu+R*=+9kclc{^`>5hpVTg3)v`o^W)p`W(-Y-2Jt$)4_>*u70`aHjwK>+tM_
zH^VcMsqyx7&%~Z-vB8=CtL^ymZyt*Lm`oMg)9n)5wTcbSbi#t!_boh~5kD?GNwMQ$
z?crP7_gqlzU&O<Bc;^?eN%NRMUeC}tF$@DHqcLDIUIt8-VZmfI7EIR5g2^!)n4HFe
z$$2?2d4>m**LW~_FAt``2w)1D0H)v-z!Vt~Oi>fT6ulyt5+i{rX%d){R{~RJWH4n-
z22=LRU@D9PrlKieDqaOll~KV|H5E+NtAeR98km}<fvI^lFm*-;Q`dAbb*~Pl!5Cl~
zngOQaHNZ3(6HHSx!8E-lm=<G!X=xUime;y%I_E3<ux&nB!C_lQJTH;>!RZMS&l|`g
z;&~64DI*N}Opq|>Ga<sD&z2DueKtr~^w|($(dWtthdvi19Qs^{aOm@8gh!tb5*~d%
zM0oUtG9sWa1c`vY5F!HlVi^(97lTAZUkniueW{E{=u1H&p)ZAqguYxxWc1}Ak<ph!
zL`GjJBMSOTkSORYA)=tKmJt<wHAqzS)euq9*UE^7z7`}J`dWx+=<8)fM_&&T9eq7S
zbo7lfVxVsXiGjWmA_n?q88Oi}gTzGN3=tE3tBhFaTR~!>Z-s~z_nqss{k3SmNW<5{
z=NSoKibBu5Kt-O(abkQ7m`ou9CR4<K$@*9@*+Ldfwul9j^KoEug&de%5eFvk<H6($
zc`*4R9!$X}fGHFTU<yS7n4(VvQ!Et06pKVKC7%SQR49Qd6-i*qJ{e58PzF;jlEGAb
z3Ybcv0;W==fT{XaFx5g8OtnY_Q}bzHYK0n@T9F2(?$g243w1E{A{|V_XMkxG8ekel
z2AHPL1k)@u!8D6ZFfE@2rd4QxX%$(wO=tZ`l49#d=dg9wjcG}-^`f)cUX4@`&tW`?
zR}co7tspEiS3x*rzJl<`LIn|!#R?)KOBF;ymMe&itW*#MS*;){vQ|MfWW9pu$VLS*
zkj)BWB3l*2inFx%sRy51)@KK?SDRM$@UQ-F5$Y_Xm)m=>c@Yh}tD5rEQ1P$0CrWuP
zBR5c<Kaj(;Cqjc?MSH$NuBAN}kt<!EFJ0g_xjf5S#F?q480<4qiorh9Krz^7!Zd?@
zrix~;&(zWk_L)i-gMFsS#bBSUrdaH=QHsSr+d#3{XTvm$eYT2bvCr1hEcV$-7mIzi
z$;D!ytEM>Yb5V-JKG#5T*yqADhkdS!=CIGz(j4}=N*9NHuF1t=pRcBP?DJ8I$3EXc
z@!03XG>?70isrG;*U~)p`AQd$eZI-XV_&GI1ndh@O2EF*Knd6v!nA;Wp^6r;FVxZk
z_JvB9fPJCKC178yrbO(EQA)(V*g%Qc7sIrOeX)ubu`kxrBKE~fmxz6_$t7Z6s-`6D
zOHoR~zSKZT*q6ezgng-smas3?(h~NiN|%IvsmUc_U#_NP?8{L~#=hJ@$=H{}w2XbZ
zik7i2*U~cf<w}=~eYwdcV_&JJ6znTeO2NL;Kq=T)!nA^YrHWRtuhh~C_LWMPf_<gQ
zrC?vJrc~^!QA)+W+CZt;SHrZ5eYJ{Kv9H$BD)!Y%mx_J0$)#dntEM#UYf(zWzScl#
z*w@0ehJCGy*08VD(i--)N|%Oxt;wZfU$3Tg?CViV$G+Y`>Dbr9w2pnfiq^5O*U~!n
z^-7nHeZ9%0W8bKz4D1_G%D}$SKpEIK!nA>Xqlz}LZ`9HT_KixHfqkRNWnkZ|rcCUc
zQOd-=*+7}tH^a1veY1)-v2WJWCicxrmx+C|$z@{Ss-`ULTT#lwzSTfk*tf#8g?+1v
zwy<y2(iZluN|%LwtI1`>`vX`1l<>!|i@qkV;$iM#6P>lzpplPG(yT7ewY8U1ZcjhB
zTDut_*3ANOZVrfd^FV@I0219IkmQztWVZ~YxD_DPtpaIo4M=zEK!)1@GTkPS<+g5<
z@5PVTpIZ4-YgbBfdPaJ#bH6a9IHPk$hO@*>DbC5x$#d=orWEI9=jJ&#9!H8BAL`hD
zO&OI^+)e5>*tzwPQk-RF4R`JkI<mc1ZN5bbi?15uxq@re8p0s6HH1axY6yqS*AO0A
zs38KfSVKf)sfI|%at)D@l^UWTt2IPL)@q1`tk)17*{C4~vROk+WUGc)alUs{yZq<|
zt%bc-EofEH#yR+$520OB>}D<Oty0fA+*zdzB3sHLa-|$1U&<p2r2?W@Dk4gy5~5rx
zBPyi|qFSmVYNZ;YUaBJ+r3Ru|Y9d;t7Ggnuf#~cM=T^0S2sPiQhMm$#JWs>=oM|Ks
zGTTU4WUi5L$b2K=k%dMgAd8JeM3x$fge*4_8Chv03bNWrRAjA@Xvlga(UFZtVj!E1
z#6-3li52HZhkHNoTO^FO525@)`S&_|Y5C&^pFH&4g^K3Q<MXFFn^!(BWqeLbpRm8I
zJty|`J9o)bit{Gs-KJGvAK1EU+1tCDkIJ?MJ)93<^|gZT^1s(~wKu0_r1fxaraX1g
zoqV>uU#-2lXXl=>^OnxZ!v46LpBuo(6(OF}co0VjgUm(<i_Ap`hs;L^k1RxpfGkFc
zh%7~jge*shjI2b6f~-b}imXM5hO9@3j%-ATfow*IiEKrP73VjvtXwkj!q#>6-s;<`
zZ>n=$*|YecEk@7VF~r{7CACXO=a1EQzNE8@FVl7Q<^s7ucRq=axG}R@6zsTSZ|;!S
zp}+I+>hQhB<>^_~$L!5RT|;vn+c^ijl44J1;-5rLubctS$7G+=d*u#rPSZBsl`A>q
z{0U7Z#h%~58{Cy6IB#&CpYr|keHSS^-aWf}59jfW^9$NP+x&zl{<C<obsqKDf0{Qx
z=W(Pf?mg=%<()G7Ej#9Qyvwnix2BU~PkbEPzG7u&<j1dA8ND+Yhi=_^{o^~&ZT_|S
zus3>6aE?*?us8B<le5)6Vw;~A#oyOcPdqpAh^;3KGFwkrWUiiY$b3EFk%f99AdB@x
zM3(A_ge=z+8Cj_(3bI;HRAjB5Xvlg!(UFaMVj!FK#6-60i4|w(Q5l;QdqfuBE$Kbd
zdpTDsd-2day<eO>e$Q;@qqkR30nbA~C7%C@6XR#VWC9GBObG)f>u14a11y+q2@59Y
z=fLCw9GF}Q2PW_5!Q=xxn0yHjrr;OA6aoU6LWuyT=oi5h10tAWi3q0Tm%x+)5|~nn
z1g7kl!IT3sm~x2>rs7w?R00Z^N{Ird>Q}*311gwmi3+CX*TB>Q8kky%2Bz-U!PEmf
zn0kp0rr|fhGy(>gMu`EY={LbN116Yei3z6Vx4^Ul7MNCvb=zDpJ20d9QSZV9GZ#*8
ze&KH(V*lvk-W%CVht{_4+j7o~*|YpJ?@O5N%q1kuiocA#!T;jS_}2U4FXzNx&YU+Z
z(2T?H8%eubB(`+7N=|9rwoU4=w1UCC(vpS^YS$^fedjKctZKSpS_3-tReA__Wc`0z
rJ9g`yk)4y<GcUh)pMH1V<zaa-lk3`_@pkAq)HjzJGHT%6WW;{~d~hz6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp.542AEDD871493BF1.idx b/.cache/clangd/index/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp.542AEDD871493BF1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a3f796fdabe954eae49d2520e155e536448c38b
GIT binary patch
literal 8262
zcmZvhd015C;>UT<8OM=DLIg!Yc1duXv#+8s!eulKx5lN+QA7({+%WZ$M&_<$ZmE=6
zRw6D}bHQ$zLK$XiI67#qg^F5n%`(!0-(i0DINv`m&%^u7=X^iR_jzaDIp@sFkesZn
zK5+p7nG<qmP05=+&Jqw15N7;Nn?7}Nya68v2L#O8^zDezgA>K}_usvB=D3|W3lr}K
z#E*NwXA(6+TM_;6#IlH$`ket85AMv-I#BaR^f*0y>p=Ib@BZ_bn*(iIUmvpZ=7p&_
z72($>znMN)zWVwz%U(Uz)ti6wO4sv$PM@`>_1vQU@Akakc3_9>g>$~BzT4$LW2%07
z(35yOymhZ(-ND8pwK`I~H+km%^`&jYzTUHIZ~y&`GiQpWSH|S!Keg!JS#x7nj&^_L
z98J>nbr-ku+kdu>`6|RadY2Zyzih|N0U_rDH<54G=1#6&9i}*ERsMb8is!jo2fzO%
z>BLIgXE#UvD894v3(J`KZ>D~<p;y0$pUgOU=(kU;2ST<o{V!;ih1H*)+Od7sp%0b?
zdWuiJJf-{7KMZ=i@7?H*yUHhzyL;=*x>H=#xiJ?qCxxuqxY?0CvN-kKG5Zg;%K1C1
zRql{B^2$TuiSNyfo7H#JYo{;Xo%hGw><f$fZQi$cO~}e0u97LiWz@f~6)#P2-JBEn
zS>}tW`woB9XISa+Vc%R^@aBTWt2;lxf6elD_D?NbHo`jT%a|METQ`xDHGJWc0m&s1
z2e0p3)Gem8-<n^$(1WM78dcN&`7ie!T{=fQ)UlxU-05ppy|N^!xIVo0T({t15ns9c
zjxN73<d<`o&umJ&M7Pxr^3EacmJgZq#+}8FmG(<d-qBm1H0pT$OSOAG|D^4`!Gl*!
z-p%cMu6WDk4lmE?vF*F^WtTc0x^^<DVbtC^C+|hqGdE<Zliq7{%9iZRq4KoE?-%{?
z-Q1Z8eTP3^zCQ9|z}W9*{NvK9*F1l&alfI4rYCfqpFClB;?`+o={bJxc&+DaLl^uw
z<86nd_#}7u)4A0X8|E!#CTE0g3u+j>)b{1kQyuocd`~~pu)6GpiCb2Co}a#Lhb#W<
z=)vp8l<!u)lXj#=3<|aFd;Q#djvkM<WAoq79Fbahd`sbE+oovtr?8>Buid@+?7YN?
zdxaZ*z2Eu0(7Gx2%HQpAxy{<ymk*sdcl!Rg50`~T(_b%9_9P#BcK(ULkIRDZ^a^bB
zDp$5H&vB)Z^J&BG$Ia+EcX8f=<G)_3{zX;_-q|uLZ^4m|E_Ca(=up?-Z{CvAQXky8
zI_ae^)*jzI<ZMRC-H5SA+J9QzTzVsW-1=?JRY&J1kKdj%yL!;SI`qoD(ck;R-Vb+F
zuI;}2UV}4#eD11^SyxUB<UY9Zw)gYmotrm!FU?+fa{R29r-p4h?3r0TqxHqFmRzh2
zsYpJM?tZSzCgps#dS|g~dDg=`d#8P~v)*sIILXoHnHeiy>0-bC)jB0S?XO^G`A2^Q
zU0B)oOz4hd<##>axD7`<r_=jwKl<M9gP(txvi{7=8(y3e9Q#2p>(+n0<@zy~T`}tH
zxzDnXTuwjm!N&`mX8)z%+Oh7ZY}dW?4WECO^6g_iN5)&4uJ68(xw$ek`Wu1kb@*80
z-&1F;h<#(z7wcjb$4evLN+)d2nwWrTuS{`=QUL+h*n=xF-xK44sL^8s@*}JfZG3}q
zF7f>YKG=M@cmO{%h#F?Zovoe2e1jQ>?ONPv#?IT7LDZ8*9A*t`;~UI4>H~M}ZP)NF
zLDW-59Ak|M^$lj+b@+2-rwjLd5=1>~#HrTQr+kAMr$4;=^T#({*%d^MG2$p|RBPX0
z#!nt!lhJ+X#*`q+WyA^Agec!&#_n%DliL5jcS#U6(ul*Y;gP<<jNdx;{tq_x>>okY
za3hYl#&`4$W?b5({rT$tAvqxdLyR~jG9}(OnDN<XRxLjN{d#8*HPnb>t+8!=!=w1z
zYkF((wjM-1ZNzcbxM1I4#<$1C9qPgT%mz_|jX2U8*~T}R@$WYivcC@fW>pY1%7~+_
z(V@P<jQ{%kXx_93mE(e_5k{P3O$zf3X8iYKBlcW+vFb(;m21T9tnJ$P1~YCN_}t^=
zV{=aj2RviM{aW=K<r~a6U~1mPX_hGey67C}TMx^RKHOY>O*JNBXY9QB#u?<WXZwj^
z7|0k00~yP(kg*OHGLGRO;~X4hJi|lAJ9x+hMu1Fk2#|@42$|>*A(I#hGRYx9CNnZ*
zvO|VUVHC&|hXR?(sF0}+6*7&{Ak!QgWICfmraSaUnfcj!Jka-<gSViB*f$dY%Wf|v
z_K%Uvh<zV0Q$iT@nNq@_&y*1ceYS+K=(DASMV~DrEc#pt;n3$w35PycMmY5O62ha;
zml7U*zKrnb3nfHAUnnI4`a&5I&=*UHh`v}#MD)cnBBC#q5D9&$lt}1HWkf<>E+I1d
zaw(C~m&=HZzEVOI^p#Sgps$n>1%0)IsOYPuL`7dMBP#k@3DMBkN{NQPRz@`R^%A0^
zua^=XeZ7q6e&0OD8_Tfe1P{y5X@3G1VXl1`kjp;WPmGfRlgVYkWLylGtdj+k&1J!4
zT`ZWKlLM2><-p`z9GJY52b0g`!Q@>$n1WLPQ^*y-6kGzBqEiG@%oV{DT_Tv0Qvy@U
zmB5r-5}2}622;+J!IWJxn2J*YQ^{4pR9p&}s#671%~iovT`HKGQv*}W)xgwT8ko9M
z2UE}0!PH&)Bh$S4w+-}N^<m*u5c^R)KUEM0nXMo!GFL%3WWIv%$U+4Xki`ljB1;uS
zLY6CtjI2};1zD{iDza8VG-SPk=zcbrITh%;d+|>*%a%3LyxdB&7QL`~Ug{QN+zIrJ
z8*ui&Hwfk}&g4Mft&RV0?Ow`WOPNn?TxaVk`z_=~%Kj&EC2g;w!S~Siv&eO{{W@}u
z#eT*DezV0sua%z}FU4R#Q%^D2&oojD_A`|<gZ+$$X0V^BqZ#aHYAg))GtCwT`&lo=
zVn16?vDnWxQY`kfl{Aa}tcPZ?pRJ==>}P8%EcUa_78d(CFU4U$S5I-+&oxpU_H&gq
zhy9#~=CGfuqdDy7YAhV~bIleG`*|<LV?SR{@z~EdQatwal{An2yoctopRc2N?B{DN
zJofX=79RTrFC}2VP)`ZiFEmmD_6wD?fc=7p7O-EaqXq02YAgcw3(Xb*`$aD$V!v2V
ziP$eTQX=+?m9&WcqK6i-U#z1=>=$b+BKC{T77_a;FC}5WR8L9RFEvsU_Dhwtg#D6-
zmat!{qb2N@YAh1=OU)Ju`(-aBW4~Na$=EM9QZn|-m9&igvWJ$jU#_EN?3ZgSGWN^O
z78&~$FQs6=Qco$^uQXB$_A8aNg8hnzR<K{GqZRB|YAg!&E6o-K`&BQcV!v8Xso1YJ
zQY!YVm9&ces)tswU#+86>{n|nD)y_*78Uz7FQs9>R!?czuQgH{_G^{2hW(m{*05iz
zqc!Z;YAhP|Yt0r7`*km+W4~Tc>DaF~QabkQm9&oix`)=WU$3Ke?AL27I`-?$7Tw?9
zZ{be?fBrW80<q-dMJ(fqnWMD6QL96Y>ygdg9j+EO28gw>K%9*O;%z*TU=x5un+PP?
zBp}%)11UBINVTaznoR@JZTcf}zI}M*$>N_vEW!EVG2sd320?IsOvjiQ^JXPDKRzx#
z+1won&QFL-NH%vGo)A~P7};Z42^E~*S?Zi=em@J&kL?>fz}ylrW#g`=WlsVw>9xdu
z2k&}n34_em5*C@OB^)weOL%0VmI%mVEfJBWS|TCKwM0f%YKelZ))Ez2t0fw;UQ2X8
z4;$1jrT$H=m2pFp(Iz9*yw5RT)X+fRZVSx5$i4z6U&J7?MJys$#3AxUJfct}Ac{pI
zqEsXy%0)7wQlucNMJl3Jq#^1>I$}mjhG@=n<^acuG?(zCG5wT`lwsy}(fofNeErDP
zYZWbrN2EM$9xh%d#hf5TuNn&{X=IYy+`$RXPac{4h!$Mfzv-iSJGZsGqK{=HnoSG;
z(?421wXBnII4mYC(fl^H=(;WF`#s(3jKfJClVtNL&4p{MzLq^Lc>Sx0{T#jlRfIui
zs|btCRS^!EuOd9MP(=h}v5JVuQWcSq<tid0D^)~6R;!4LtW^;WS+63xpUoH6m}bk~
z989x`*cZcnG1EjCWVVU0$XpZQkohLUBMVJLKo*;bh%7Y`30ZC;GP2S{6lArDsK{Cq
z(UA2fqWgLKo#IzUUfZ<7c$3rGq_s6qxaG4xZS~}oEm_9lsJ2m&=9h@wXXv>6Tl8(?
zaE6?rn+xc3ZpegI{kPmP4o4(M^e{Jkj}Gf|Yiz9dm~q%)aU__wc^b0>`gSz^uWWqR
z_@3tKGER$K6MC9wI)i#ANG6$IZH%u&EgPTk1ij-0^9io-V}6*w_c~>iN5@4cnmeB6
z#;EbOw(N@f|JJVB`<Sy2&Sz8xh*eo2PUV1jl?M`30Z3FuAW4;gWK{-IR0T*?RUl2(
zfOJ)VM0afED=B4ni805><jCIU+|0SsfxcZ`{~Tkw#H9GILNVQ97?U18{P1wEi)(*t
zIeqj=ddmE)8>f#x$&bi9YP@zWd&2(bwKov^1N`JS5C)lTAS^Q1KsaQ+f$+#e0}+tL
z1|lL$4Makg8;FdoG!O+@Z6GSL)<865y@BX{Hn+D^1AW`t{%M9MhIciuKE?&UTX=Ur
z+1>U$|MIaH*k1%Hw9obv<7U8Q3K%e%LIzCM&4S4muwb%<ESQ{|1CuM@z~l-!FnKo*
zCSSmV$rtiq3T^>Rp+EprC=|dH-6EJ`fe5BpD1s@uB`~D|2~4R_0#kO&V9EtDm~x>E
zrs7t>R0<R@l|ltf)vbc57N}sVg({euTLV)o(7@CRH86Fz4yIn9gQ*wlk4*FAM*@BS
z3^5ng)VzXmE&mvqHg){8u`LhWEraie&OdfQ9{>Ky4og~38kaZ0J$_<9o|y#%Oz=PS
zJ;DFuc>mFf{)dzN564fLP|$({CO$!1T3K7$+5`oMJQmtEC#-*F*Ra5xjCLKu+joqT
zWL4Aqr1$L^k)|XHkH`M6qmiAXW8&fyl9E%prFQSt+s^W07p_we#t{*j?VL<y4eIv~
HGVuQZh93|_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm.hpp.2CF55436E05C96D5.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm.hpp.2CF55436E05C96D5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02b70043f1aa046371ace09e12187b3b835207c0
GIT binary patch
literal 1012
zcmWIYbaQ*d%)sEB;#rZKT9U}Zz`!5~#Kk2=nUfhA7^GPl7%JxM4UF|ZY#{R1|7G(&
z-{j}z6AqfWvbZ|d>XqNz%9~s4Jjt*9$Njff&x5Y-Vt3ToV)^{boY}?A-E+Tp?B40d
z%G!1~dPeX|yQ+uhs_S|C`E34NIIPTXAN4qy-|pKT{euUCOILLY-`{_5&M!{hXJ<}l
z2t1EF723zD^<7K%Ube{tm6cH<vOgx?HQFzrSSR>P)b*~>`i+I#Uv>C*eEPtbQx?A~
zJ$9qNTa#DO9|;z5`|>N?=GPuOUt}}6=dkbl#SdA`i^U4=G;J-rUvOXP#jD+hXT%({
zkMG^E`L1I@r?F-Mi{BE9y~h?+ED;YsaQl(*j%J>AHBIqp{3=tOgOkm(+$3j-su-P<
z)$&$b7OL_5@~xSlb-&~@huH5wA!C`Wvy>?^E0HVg+>B-6CWaa(o`*%Lt+<eqI4L=%
zcd0>oQ0JaGO2U%ACC>`QZc6=iU3*@`(Z3<D|9h=|FQtB!H+lD-+}eyDhg*h|H@S4!
zhOa%p=yvz7GXj=jQJeejyzO7r9y&+mHILIr_HIcoceNiZrmA;Z8Hy`&lj4DKB(kjX
z<J`734gn@sb|4oBIL$Q8tQj~U9G%!7i40;2GufD!*%&xw*kzQ#1Oq1%FkTrLIT_2_
zO888DUN{IavB31pv&f6X^fMHtrWNM_)xzvy<OG`!GZ#iPa)HhDR4MD8xXZaNcDrRx
zjGBrvBM+yXjGP9{lGHv{)drzrC!ny5u8bv2GfWz2Hpp(6aJLt8%iI3{7Q8&1$}-Bz
zFjYL9Dl#faJXIN0B%Yd#8WK-kMjeT#A)|rB)0ELf;%UifA@Q_jw2^o^GCD{+T^U^@
zo}P>z5>H=7AI@8AY!z^HZo+h+XU$~H)L{Ym)PVI1JHL206EJ85<)jh5G7~hDhAApa
z%}G=OMi4A07>fSSUh^hug+43<;WEWV$;J9Wl}wzvTW0(gYCaRf%*D+Ck3OIbHxNSs
J7mUHc003G1Rn-6h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute.hpp.0F50C2899670001D.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute.hpp.0F50C2899670001D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2e0801fdacf6a40e99de0a25b92fea8403ea1942
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%$tl13^_o#iaC2H#rxkj5IOoe@DoFR
zY93e5StpMzt{G0b6K_<f%<)cA{il3e)8I&Xe8GyW*=eq`T0}Izy_fkp@Ac~^e;%kF
z_7V9o<#x%<PnM6buV4J-L8{Qj*2TvKW38%$E=s%mEAJ}$V!iD=7t6y5apyW!_wSx7
zd~VyJ5*ZD98&&Qp^P^M6IBwq1c79oLs)Lg^GU0@sl-j(9s~GtX*%>iwv(Ac8bA501
zEtvnd-<}1Je_q}2@yh$s8_(`O()oS8GQGml=d4TR@&=<z8|?jjY=2cvoa3&aFvaa~
z|3M=*bHj+<Oa0B?Zj>FfJ5zSxCZE`OlPQ}NGq2mS&ip3PmL|Ss!~34b%HB(RC-3rj
zwjoVq*^x_JUD*%*_}NAH&AT^k`+~}B#mEU#dZjn_h3=T@6LF#MuTXYm^SuRmKC`21
z;&*3zXY_7lO<Z#Bshae$R_&gTazBNC?lP8(V3tq}@#I@!v_zq8?Zip_3yPIIKIhDE
z3@K9O-I=fOK8QO*KzK_gx5E~<LZ*;aVqX>S)&I0L{IE9Y)#_C{r+wa(ZDL+~`LEm8
z+<Hfmd;6!{Vd2-%(3$GG%Sm;qd)TRC4YsG$JiIdlv%hS+@G#|8?TX7YOv0ES{x3Wt
zwPLE(Pi`a8&~k?2%G{*6z?hca;lJ6xBcoW9iH(ziTS-<)157Y*0Of!{g@5znzrSu;
z2{5s6LPWS#wN!0k(kAi00$Y|I+`z`f%)!8|z^<SLCKx!GfQg5Jk(+U)lH4LS^{7b#
zOsp{d4w??uF#UJpWo^x(nVUtK*nzey%PMPt379lPQEFN-4>Q7*EL>pM!K~%s)>6>Y
zhPjc4+eF{Q3c)kgH-+<HqKuqi_rnZ;(TrSRcYCUobx+*oTo>Xer2b-}4J!{1w~nxm
zBuq08x2~`*5>HQ94~eHQtdGPq5H>*K844RB@r;Cxka)(z#z;I9ITICxeWr4zNIWw+
zGbEmcoCTa0yWKJ;MomSTk%wDVMpXl5Kx!YWYJ*U*6Hr*eM8N^(+ls>vpH#72Jj(_Q
zKVdBim}(wwZDDOBo{6vt63<lF6p3dhY=*?M5VnBxy1kfN-uC~u;N#&|5>=9cr7s?C
zWl?1rSbPJ+RZ2$@VX>~1E)q{qN)L&rFQt#fGmtVs;u%UABJqr*jF5Q7QpQL;6IBx(
zgng!}rbs+9RWl@>g{lRd2TQ)dWCKdJumD_ZY!z^HZo+h6xH`x<sKZpjd<RUWpri~_
z4GdQWZKPCcrf&vMmpt4S`WA5BQv=p7?EK>0Ou%RqRFy`!*Fn%h8fHsTYEGgOFn7U>
zXUOsBzrt&;;RB0pxJ+?Ta<Mf~B@;K>$@f#Ig>$TC;pOAvVc>z}WDv=PVDcfDya*-$
DoNGQ-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.72B4F2954B40EC19.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.72B4F2954B40EC19.idx
new file mode 100755
index 0000000000000000000000000000000000000000..06b1588d7d1ec2371471223d265cebc332628cc9
GIT binary patch
literal 20010
zcmeHuXINB6+dn&J&e?T2EM=(+EQp9wMT&r^*s)_<yT+avODxeOMvWE3E?5y`Ku}Q-
z=}qhc0t$#Fc2Ph;1VscD%YSh9=FIba_`e^L>w3SuURQGOkNKUMGwp6OXY}BK18qiV
zY4w{u_}7JVex8A~w6yfef6IPeIxj{-emiPu{ThCI!or`PcQd&2qQ&TAF|$IudBMky
zT?YOnD!5mP;t&7T?*5wY>-U*XKJ(bR<X)|D%afmMw-^i_m3VUznxgl2x%=g`vckp6
z$$np7cM5CS?SG_el2`8QeBJo-I<;0`Eaq2b_U@#}v@ebcJT3kCSVS7TvfKCH87Ow|
zy1&)&QXbPNz4)8+V#AA@Y<q7hPu%BZBvRZCIW_87{w~{oVGE`doStXuGG+YUfGDH-
z=G@%ewI`)xs(-a&6gM}oZgO9FsCbX7-;VS<!@7^o+Bf@TMODqLtjGo3Z=JR?TN<2M
znHByb=zPG5m!TKFP8zg8w4~njy44SRuErcoKH<4*;^i369B0LeaR(2m_N4?q`~8D*
z{4e&aqY58AaZh}n=5~5&NxPZNi|hYzX20J%eYJmS-#qrkyX2EkBMoB5UYyf0@0WGA
zW`yo`Q00!O(VG&h7gY0p86GP>vH^Ef$zJE)?j$SP^J@H=pC9_2)v`Sw+G$a6@Xyny
z{v2_5*#4n2(mTuE?F#AqdsOx1UeQ-3bGh3}8s*zpEI3kj^~a5OCxkESUgP=eZj^a6
zEyjOLFImd=tbnPK$zgLsmM$y|to!NKoF&rg;Q1r&d)7}i*kw54M$qEV^R|9b<#>Gl
zn3wdS+l~!OFLdhC^1g8NrMEA~Zu#v+va{ux3Qk*m@Xuz&!{mwq54=O2=llIsbuu-4
z`+d*hso@dEt2a1b-TnQnMMM1)I{vYDuH47@+^qvi^Ujy(%)c==j?t+MKRme4>PvsV
zusk_%!?a|bs>G@}^G3~ft)93#`q_t2gLbR-T8|&7Q!zZNA^k(5!{j?I^QTId>`i%d
z?9*t2m-nI)*xQT4^pi_UKZMN|``w>cv~%N{mgz<v$4>1Qb#A|3%x@>1v}1-e4b_vU
zF3uk^WYhjR9XsB2{w2YxYgy;j(yd9omS6p?DDhe3c-4}7X(x8cS1*ck49mDb%X<IZ
z5z$9;qNhaqoESd%bg4u1t$>E1buYRP%9gH{P0(4f<26_I&h)38Uwd9Oe}ClJ_1KBe
zE0T8X*X^}YdvMs*sTOwr11&>JeD^)>_1JyFww1DVHpfOib@rc+8eTjU512R}*Zwgm
z)oxAlvl-^O*V1(Bw@kXR{phxKPm=z);vH>P_|(n!BCenAJnH81d5by@zZDm3d;3lJ
z?`g-L4{-?n?NyC+!@z^K;roV~t$u8GV5w<t@QKFsb0wc&E$sI-_SR2Z@dL}>lDpPS
ziJiGk+uZ4|o#}cnldo)kWplavvj$!?>+Pw_({_YhmIS>H?w>GmQM{S5X`CW;(+vGp
zHlsFe8FMfvqG{dpn#EmoJiSvr7N$E#wKv_A7TGXVD^gZjknQTUux#QgXK#_IPMM@^
z@wTH~Wu3g&f3@E+H!{`bRm8^l>&+fHE{s;toZbN&Uj&Q4KNQqu#q(KL`oEbvDW=22
zh5NPN7)N=}u-{?zCUpo7eVae^rIVJ6G%N4-*X-c3>~~26tHi0&*yYoYGLDw%mVXa_
z>@s6VyMEl_c^7QAxE#&f_+#>vKPPr5Z`d)seW&`yjc1*GI|udj^dDNTTPsWKdFH{q
z6<Z6dLn35lAqCy5e|A2-`DY!~442_IH;x-VsAHjGQE^aXr-l8ituw}tp1IstYORV7
z+1Kt?OnmLsF4ozf1FuCzF5j4MXZtew=4P2?oWY_}kJ9_=7TP2)U3_-!>UZXMW=6Pt
zbogzsuJ+-`*o9$pe^gBxbTnqj$M;+9Y6cH~5R|NlIlN<qeafL2UDjwt-^_@WB{z<R
z?Ov>SyzTq|&qD=G9S$-rw%LWpU8ILyv|n8PlAU$@ao4rd%o>rqXX5DU_j5$?9frPJ
z=raH9*F(+8TffhD^f)zT&gT(t3WiR-eXNW9l!~eL(I?Ix*%3Lq>#C_<`x5Rv-EQ^x
z@Z4=oi^upL5gUd?%hq_G*<Ltu*s3tO^m^%nkv}w5th-&<({#G$F9CJq565?Td8%9K
z{`%Wl_17C~`ki&lm1}1!vrdipWBuKBJ(;7YV&CjfpYm!!oKs!)ob=W0vbfoEoNqqr
z`ir9A=iP3<>J;a@y${;B#PYX*I|laJi)Q~-wZh^%iTm&KzCM`tx8MC6oxk?}dZyz1
zSK~jLd{6WlSU)^+(vHWIL&og!4cQ$TaAehfW8c4~^*Q>?-#FyO*>*~^N)t3dWE`;E
z=V`Zu)AqCN*Mywa@!WiC_x+(Ut35-y`&1V?>%DUNke#w<;2oXBMLBndHyT>~@Z+P|
z&AGP0V+Ll|a6ctJ`3t>Xdh|n|-Nto;r1s{ckEGa%GmF<ptk}}m;n;59Ih9VX{<;tD
zGtwo;lg+j~Y?^y;*XGh83lolh^4g@|DPn;6-dDb^%A_?t2O7u!^2{T-eAN!;SzpGK
zS*{JTke+{bd3?B0uY)s&Ub{Bn?ygFYRpP3M3pNYgmY?1i!@GOhoeY)!+-SJb+0(gv
zbw!+(*1Yd`dK9h=&%c;)?p)!gYfG1f)-AQM7~E;G?pq(nI$gtpP78ZAZ#ffI6tMX_
z#p!t02+98O2QnS~&RtqMXHSKDTF;(;8`MWVb9MOp!^))<oiB_E+z?`;I`>+##@9*~
z)Z!g=dG){^As+%R+kP1~W|)n^s%xJX9QeuQMxOB$ixuW2)2sX!EpDtT?s*|^TUm$q
z=Zntq8TN}qa;}|k8r+b5{KZU*U9-yT&X=N!)hp-p+Gu_|!h7}Q=+}R8PbWS3<!b)K
zpsex<(?ZfuIXB<68UA$J?@sl$55Iph-Eq~zO=HhrJh|oo&W>Aqr2W!I+x+Xy7o1+z
z&j7u9Gx^tk_S$*--{(zRmfd5%k=1V*(f51y;0gn#I|O8-dyVhgW#6g$RomL`$AdF&
zuNV8R-*L!srm;!O#}O@h*L3#uJ+LYuc+K-AmjbKTq%V7PWP*j*Y>r>V8`tf#y7ibl
zujgIAT-hhxK9&9M46|?dHOFwGPrrl99x(m0EFCvC=)YJ{-jr82y5@t+j9UhN{imP$
zXdm&zIOnt<RX-it8&Y&Tl-<k6U+Ld-R?n;FzW=Pd()D1f*0SFg&OD+`To<!kv1;Xl
zUNSU%ke2Ee=5JwdjnE<vP3u48ntA@{+k%HOG=fm7arbIGRwICd>$mZ5mYYs6K-2uS
zR9Q%og}P`2P{elKw!~ScC2`1#P+nv0*VsTKfFhZ`<a5~91p~AZODJcM^bE4p2%u;e
zo2>d!qSx5~bs&`EoZfNHLL-2}htAOJ5$xupjm!z91xZ_wu|@zz-&fy3iRubv$cyyo
z1-5*FT{HqH1`BK*GmlT|D?^@yat&Kv!yX#ZMybg)s7UwC<dB#&bdJ$J$H+ATD9TV?
z?DBKck22&=YN^1M71&uLfYScMxUgA9t1^v{C8?!Bq-YRXXarC?42Vd#+jxZ6Mm(Y9
zAZZTiky^A-cK5TfQIrK6AZtP?V)csH4jKUz+w+S`>?6mMap^)R??lpfBDqEY#g6sq
ziWhttCPTBx>t4YwSFn#pv{9A@4KKboBvlW&4AD|uVr7?DeT@K0ryYvzJvZ1S>mggx
z&=bz)3FoO1KykcK)g?pmM?V?rJy=WCgv^?dlSTl=xncLdMWY`rQa99yWsO*0r4~SO
z8FZxN@z#!Z+Q^VJR4<a&iwrcPjdDWv@yL!trFzJLQ1*!1?-9Fe1W>vt9{(6#Wn5r@
zIugoGgm)qvjR1=4mb-&;B?bLu$c?=2TV(ncIcNk>Jm>ch318j)jS+GpwLBGBJ{5J=
zh&D=4+;{skb-v2bbTYb^u;LQ7)d-+?Ne*;P`Yy>{hPo0;F1E<UZW;j;@9bRhcZpxf
z+U-FoFR|WBY@rcA>2XW%P~A$41=>hP-X#f1laQ%K0HxO~%b9;4S^i#z29R2+uw@l#
zNG*WUCw8ifSLudf2B<Ts<p9z<fb2B_D1EE;jmbUn{T%~jM<{_vFAzCv1W@{=?bg-X
zT;EO)Ig^Isu}(bM_G$r?{<F4C8Ib7u!2sElS{`tE4>)s;0E+)yNqT~~^DP<rj?{7i
zn_a**8Ud6+zQgMud^g-fhI$gpUF>-m`)dSH0_qAbHU8KbE<@c2<vKRIj-50DC}ZOH
zJ2y{0u$0%DI#5gHB60El_CKw57-`MvFH(uddQ7g~Z~en<QU_TI?~vI$vU1e|>X^8B
zV|IrZ6YKPl386eh?hjEfjR4BDPSKr8mv<UufNV&Yzlij{h@3S7D0AOE9(sHC=8M`$
zpHMQ8Gy|Dy1W<msJ=A_fk^Lwmq$HHLBE?&xI@AIv3kL1HrQ_Fm7>DGfKi5R!Ya%a=
z&}zqs^t7}{pB8@HGqi8#y8+rrLh4CG(nO@x2&iYtQ~PyB7h;a-AtTbKyU6`6>62Oj
zWzD_c$M##XaT14gNiBIuI}dfz2%!AogZ@5RUXv|D9%KR@VXH?(9jXOTf)78cd3)H=
z+Zc&kXPS{*VDvNsC|jP#4f}nDa=jihCht;)+{=)kMgV26)1}G!v)8^iK%L0DL?OK>
z<fsuqIWP)sv<*M7Ko2PhB}}XrCbrZFphQ|M*_ypFcdQKgk#~8B+dsrzH3BF{BMWjP
z$A{h4MmnUHG$c(!3XK3twD!GE?cEOK=zUYokzBh-VzShtjdFE;=Q)pVCmSIPQp+2W
z;tiQdwE)Vo%{@oFnSHoIhP=tUJiy8a*ij>Za$Fuidi3AZhH4`zsU-nP6Nvd#3!ud8
zxf0cq{O*Aa^&_=B#_b=I9;rne<!Zb2?elK9Xrp$dmQ*B7B|TCLpu~S2+?<m(`lAf_
zkXrJvMIKQeY5|lpnV}<2vmc#|kuI5|d?d+7wi*GHWcd%T2hHl-pp6VjL+MDGj?6Ry
zC}~{>hK>I)I8=uEkcRTHNj~nZ5kSdMOxX10Qpdlvkv6F%1xZs#EouRj>`5;tNseZ&
zB-)$QQh-ehu!BYb<-E^!Rf%#D(YxIVr4XAI5{g;?<?5mlQ$F858zMt92<0L+yGT@!
zS^y<S`TRuF`M=J~&}>2}$Cl-|yG8&d*Tm7T<1>>q85%~$r5P!jk&8yOQOx#As>-^r
z)<zuZQ8JPylOCxBP;MyVR=qVJZfS%Z3FV2%@(D3EY5|nL7Z;5*y0<KdSXlBdH?Y|a
zqU+S6jS{Mww5sw6p$sLITHLV~`)LGF9vgHTacYLvbOY3$G<1zsTw^^o0w`rU(+0lh
zHdo6~FH%bxHYy`lO)Y>@7hyEy?TPxn+DJ@%h6XHaz{VQUMyXjZUfbckwG8zowcNz~
zP3)@?K&c;jFZAin`g%QNN`|fyxmOYuq!w+I7>{+Iy8S@(%P>+)5mpotzfLWH@>)``
zK5hA%Lwd-J)KZJwYl-qu3!pSsO7Gi$|9rL#`I1^Hv1KLEQfdK|cb2sd8-06w%8)Ce
z+{eoMWH+h>P@3)6{Smoq1tLp|G}MSp8;Q@L7HyRCQ~KT-*+nKpGf6F%vHUVI<!S*G
zt)+A3E=%$zUbx0jXLwZuBWqx~gQDbBF<37K_W&%_c<%-p?^LV4n{xNXf&|2?E~4HS
z(OB@+c~u7ElEL%^e;M3-(!zeMw;ww|jlAk0);lO*smA$-7BIfFy>_;QBM<Jt{0?j&
zJh&6{J88TN^ScBLZ-S(nSOKgRPrO@?95fglU-oe?ul1ph4x|qy*s27Ns?#hNUKPSR
zhOkp74f|{6l5<<?AjrT{O=uz1_<-y5fO7#~gIDbnTkjKha?$vQt=N8_*cC8@BBUa9
z93oT{#7Ke|a~L(AxMz}FY!03quiD8fcd}jKjUZqVNi|_gTSb~MMNg^;PTC4oZ<+RQ
z8E0V>4`bcK*j&I-*f0uP3-}~9JW1m#SpN#P1ZTc=%=AmJr7&itSiclo!iwUFm!y9e
zTY>|{s{&cWK-L<1z^j5;%V5@Cz#*(<2#rHo%TOBcVJ-L2IEB?sVa>sz=T%{xZWw3Y
zK@+m*xNZ*nac|cL?RaAHtlo$^Mr*7gS?MCH22n?dka*(TSiKc>1h<D*Z9x16WB|?-
zuiA?Et;hhpQeL$e@q3X0IQYCO0`U>Z032Ih6^HmZWB|?ruX=@gzd{2+Md*C5@S1nZ
z;sS?Nr^Ngzu?$=Wtcn$z#EQ)X949u76DtIKT5Nin#_?j4c$$Ak%%2hJ(+S)jG{NU;
zjfhvJGxBuC8WxvM%8B!x*~Qr`nIj~*i-^0X{>Kx4Qu2_-Wk^y+<4Po{q;V~h)CySF
z^j2Tj(#P{zteVhyf<1V>ST)&qat{Gb@ts1UsXeCl6w0RgPHQXEYY9!M8#`bX=|Bil
zgdlGiatO4^#$AtP>#-FqTB#=71|0(7Hqjx^8QF8DBM^Dj5$t#b_XRA~1mSHDYSf%q
zstLV?ZKQS{fk;*p?$Cf8!5Hx*Ja#Bz-C!B=swmDYit__^jaOabbT4rxuz`8iN6z{q
z*BPQ)<6ot@tDDvgCU2aLY_pLMXtP#5Ykv{-5^x4%lEFB^q?+DMkbF9x=7zEA6CU{q
zkAjVXRcl$FwQP3*Uu8|MvZi*LkQ<XX>=4D-!W)|D_PEyj*oCWP)Q9*C83a?K{Bi%q
zkHt}O<j*b0tOX5+h9Mb2CgTeu`@$%NF;!<KV5qEQ>wJPGg=M6M_t^YB?hJ2aRoyAI
z-HNAvq`*PTK?<v}edf|1YS;D%6k)<Qevi$74?DM=RFlyVCg%g=`+?~RZ@{Y(akoV5
z4bdX6O2OSyXq=9{(y@<Fo`t=#X#9lj`-JT;Oy*P8|0z3Iz-8>9GIj`H8^7d?2Jd3&
zKb`}yP>fWQkP&)Vh9za#5ZohTU|7R_f>wn$A$CEXt$_z2O=Bq4Bxj(2SFJ<3>yW9S
z?$#rIJu(t-FwzZ1rh@w0g7_`SNWdXTHw2l&-sDxGhz~_Z0{#=3{E4gu^>`AQoJ7`w
zdQ3tNNyr_t3p{a39g>lIuEt*_bA}vJkUMzfJc$|{QjvR%CZOO+T;!03+#zIYRd$ke
zMjw_Zuc~0ID;RrWO;j-kRSaeKY8ZnW8oy+0UNVjX{}p5NO2A<M$Oe7P$R0Ch0&y?a
z+bbAUsV2n*%OCa|nHqI|3%m;0`lLXe)q<&kbQWpqCX(Gm?IHZ(RaHn{h3vq0m1>e)
z5Gy;+{^V<kR$r20)pJ(=IV*={fmKax?<TgdfScLg%`|?`_I)obc&z%sj`+Zi6!@Dt
z%gvk>jXgJWUNj!DnWOl@oMkZ0-^qFG<hlVLb~Nd_ItK@1(w2jhYBFyyCcNq?WAKzw
z2&4XtF?dGf=ZxKR#ziQvXFAji)5oihW4Gg&@^c%oSp#k_=<g=1*Mv>s%~~BV!=0>!
z;HK<i4R_IaH*2_Cz%Bu^{oPLevK3*~E^h2DZX9$KtM+gndxWloOGvt|&L2V#VdId!
zug)4m0nuj2ArTFNyh^K0bA8T?fUZJ1kt~-YBr8JJ!g9HX<o8fV$hJr|Nk&*MZiC9V
zaAp&`5wqRIS~jst;hmdV%VrwCXO-_+YWhF0mLKTyU7W`*&J&*Z`0nGVjnNm+5XGJ^
z(#aROLRTQIsV&tcGogU=No4R`WGUbr#x{rP0(~T2C*xJZwApQ5Rmymk(m0N@j^i8!
z^P0}Nr*pjooXNRo(zt@NuHYPD5Av$locn98mw+2M_XYubYLczQ*ga%q4;fQ<F|w(@
zWnN*od)9Wlx8U0Ak)-QR*vSZN8u;^Mj-D{GCyXsT&#Mk&$HTY}V5ufQ3-v*MmNXoQ
zWr4VD14uPlS|D~;zR91@&3L20sswDDfK9=;VO26VPNs1hHcq2)CN|Eb@ewh9L~tju
z>Mt?>7mXvtW)Wg@!99%>n?=(2sMzc%jibb7Q8bPg%cI3c!t=+(@?$hUE;c((^G}GC
zC&bnQKSr#Kq47zv;YqQvP<~46c}nah;8-yqOTT}t*e_PxM<|aI^Ko=}oVa(KxUW!t
zTFjrO%TJ5@oEG;J%Hzd+JY60y_KT<MKO^=%Bkl=&FnDC2t8?kXn%}}mwlKzm1<S{h
zd@@&>v=t=T$x_&eWgBq^Va=#h?XVQU<dPQDd3K?NbmWkZ`a!O&RVns3j#W?**mm-u
zI{Pj>n1CD-kU!)EdDSz<;u+%u55ig_N===Nhga!8+B|AilcYZhIqsou_s~GlMZ}7t
zZjaGG$Zit_fV`d~f5^tRnmVs4<PWBfSJk85^=KqSQxJKR7S$<!c-8*b_SKC2*r6L4
zr8L$)P4IttRTgWXCE)%Ay*FR!<I#ic)WMwwdw{>|?-Mezp-yWj`SUxG@jKCQ@Sxz(
zfK+^+k=<uZpz~7A(E;=sB44uC-!ihd3>8{OV8;mD=B7$D2MX}u@BlB9ya5MBk){Kg
z4uOmvtb_pPxZtrV*QSwz^~hvBasej|jvzQ*^^%djWZJ_!ktiG62McK&G8ot+4g14H
z5pM~brDG=nXJE4o0ZTOp8SrYOw&~206|Za}Z*WGedq(UGt9eYj0oe=Z=<Of{Vd8dS
zV%@d^&Cv(R#<ZS$47Oddb$4dJHN0vQYrBaZ3w9L_K*(T~AXy3R>pVyD=g1Bg1o2<L
zosNL<8au+EQ`PPstOp4{L_S|celRAy>Z{1-D~$t@PayJx#m%dNkWUbecOsXa$OG&z
zuiAxNcG37CQXWLMu&YQ%kur?N(MTDMYz4>m7*ZaiaV}$$%UB9ilgF6k(fAr;a*f9M
zj7dI?3mB6E8W%Dqg*3j-m|Umv4aVdKjf)tYBBry@elcTHOyd&9ri8|&j7=$x%dt*5
zmP0z4IM!IF0?Q%(<4H18rxME{&&iWmT;~OrL$FSEGS;cWa!_A9$@iI7V<jXAdDUr7
zews5AOm#dbkEih&PJV{Q37kBE#)+Ifk;X}!Jc-81oIIJvDV#io#;Ke#m9rJvPvewn
zG*0J~=`_ydl$kWH<m8o{nczCS;N&l8T*b+&Xk5+7t7%-r$!lo*l9RuraV;mWrSU6H
z{))zRoU)Fy722=ol=U=z%_(2gxPenP2zcC#y%S?29&{%=I7uu|5)A2tmIGZYvo?7A
zOIaPz^h9H;4>4gUPac@PC&0h=Me+iv$TJlMK!`U5dzbavIbjYd2t_8Ls0(b-scAk7
zrr6nbz&y#)Se|3-K~0e?3u}6bwS}0ACnsYaBDpS*;~?G_XZVD(hOq*jK@6Vy)D)};
zQFF-Q4DyG>ZEN^p@l-H))35dI?qEK@GkLXjNU;uiK}|E=24~4<oJ%JKd7_?qqW)0h
z%r(yK?(ZqrCk2k!(G(m*IF}`vVfCRb@Ojk@<ah)15S;mDBxy!waO6XreJm-$X7w@Z
zLkdV^a>UF3c62LLl!Ut_(aDcg+%1*H8Q3d>PJW!lUT0~%mYu$qoe7cAybhNPt#xF_
zb$Qh==`drko4jhc;c#mKjgXF@$^r}nsIrmLkyP0z!%<Y(Xz6IGY>eR;s%)HWoTBsK
zy!kgBPDJvm@uuT#1#}nL-9>JKd+`w2Jw$G>=y-CXYgdNcz?u-Hh3qPko8Vs5BD-4T
z2JQt69<g5PQ*Ob2)r*YkMFSvEg`kYgpZcU5o+K8SkpwdqV8eNmX_Q>U7Lc0cRl8Z`
zZq^M9;ZFg5zxyTM=`e9GP9mF=$XBo@X~;7TjetbQV(<DTV@iKEY26jrEfn`zSyAGT
zo@6^OUbv$9{OM(tq#y-(rl0^woh-px2V-1YiodN%mpqX#yyKEzdlgya-~3E=_(w+e
zQHYlDf2H(DE)mPt;<gkY{;&Ki{x6Rdrh)8r^{GDW^%Vs}J|E99I7D70jI#{m93W&|
z@$#;Az>Jtaq~JV~pGTb`-}&q1vhtO<<hlttH$w7*$dtCj@-P}lBY8A36~^rtk{_e-
zRmS8hV<FhV9L6Mv#zl-t5six(lVTc|FeW85E@ezgX<UtUs<DA!*J3$&EN3d%l{iiw
zN8?mZo=W31PM${NbWWa5<4jJTN#k-(Ue1|<E+clClULBVj+58XxSo^O)A%(fe@){C
zPTnBkRb$-!ZlCe-Br|eKtb0oA0G=JpC0X+7YZ<VdNVNa$WF81SIsTQ@V%xtpC&T0f
zg*DvEy26r&`x|7*H#4$kVJM}VTO6R)*RJw9yDk2%Jt6*yZ2m+(!Wu|Io=IpJ#8hh&
zsxCV<ernae$;dMqjeu<8`pKI{wD))(OA2z3Sq^f5OzZm0(T-alrmZ0bPF<Y(f?5s=
zPttS$GT)1|cm-Qs!OoEWYdyYl3}$_x53riZO8?IT0?;(1Kz*kKMrG>|X5*`!U%dzL
zsy+sN%mj@&NIFPgKtrTMDD*^R^h9I<2ld1niHx3#EFhiI>H_!vBJu|#)anBF{z~IO
z)H@LQ(=KrDAR6yPo;y(wu&=Eyu;(rsA4K*Ck*hH5VaPs=#?i<=8oAP2&;A&V?;^ju
zbny2O`8}kAzcSRjj1K-PkzXYmNNeZbwa6c|GdU$?l)3b&P9CGoqwzIHd5y;Tj543b
z1&p$Q#)XWskjB>;<#ih0V3ap#T*NpRF&;4aysDUSE~aq_<6J`HQpUNI#^u<s9GlZF
zwP6J|r(J5pN^DNM)P^sxIqgy#R$+76rS4FT9R$_$5=&lUT}U30qZ@3{A{?afsz<Es
z5o-j>mjvUi>@kg>u(BsKe#*+8(zuM3mC^V#r#Q{E7gSw5r--NV8BTGA#tEDvfyRlP
zB9X>PoFa+F$($mY#wna4g~q9zeJbZFw4cV=r_nf_vrngSCTE{X<4R6Z$+Z_$=?hNr
zg2q*xqKd}VoT8e>HJqY`#xFU=OB&a5idq`K;uNoFT*ukhajru9^_+b@jbC&2uW8)C
z**6GyYfb+<Kg?bF>0eLeZfnT6Sl`d8;(yF)@4WY$w~kiwt5!8??Bon;bkBAD%|XsD
zPLV&~X7z5f*0AD0QxL~neVtCQAsNUa0}TMbnB;vJ=Xz!kY_+{*4i|c#i)uaG4i}q-
zi*3Ow?QNbub-k{|Z4whSV&_Ke1O93&9@>b<3-~f?dYO(9F0<V(vwb0E*mrHv@zTIG
zPDJ&8<P1M@R>G|86C3UmTR{eYzbx~5@jds})JHH<1S4-yukcDl^{X!eLg&H{J3d==
zX}Q<G6jA-z_el@ssG&NhttnsE1ndUcoK`$G0gn@KA~sH>Q@n||dm`>3lqX^1B)U8a
zcTb|rld)?uU4Jqjn@rcAf{jz?=TmU^6#Ds8Y@AA$r{eCZba@)?nugtl_esO!((nWU
zr(@%E`uTM1n~r-5<r&yGgD%g&z8Q3RChnR^*Pn^UWzzL$VdE_N`7G?4ML&NQ8=s}i
z&tl)RLOI;*Bul1=ku}jr;p%Ii;K+avPh2PU<xW@`yy`k~ypDW@Xs-!Lnvf|R1(IkG
zOKxCOxQ7?j?}26COCJyYkBbx{U$Pd}mqejOxE4bW)74i)p`i7k!6y*=!j9v~Vj7u%
zN5fzf*9wnJrg0h`orcE><(YVNCXLs!o@-ezVgEg1`#)j_f+}b=K7AhxXOD0-m2~YL
zBYP*D(2`8=w|lBEn2?MngQ>ojD)eA8BiT$}s>;KXJZugZO^JKMDt8H&s^A(cY4JTH
zdrwcF`Z6oD2#E>OqWTJ}&|)ESETpIJ9g@65rjXqv(~Bk7u_@$hW4lNmSQ|y>6J-@E
zR>TU27330;*gQ_`4Jw2r$HeBRX&f(B#0&h`U$fYe3nK@%t~`;XNu&qaYT{ChB+UYj
zbMHR#>f!183Gp?yc#S*3i7QWzAr1EkS}k7wlc@j1te=TiyNu+Qkt+mZFnT0YTgJ%B
z=p(E@vE!fg5tjP$E{qziX|e?WGd@U|^0!;$HeCZU3JF+00UHTRC>iS~(>M+5r_neQ
z>t_lW&NGNLQD6KO=4C4**~*v*LHKd3cU-tPl=5<AipiFSlmGbQMryaj1HYHl4kC9}
z!o~7%A&CZuH{<}Mf{|4)ZC93B4+jRnWhM#R>@V#27oD+>#-`CEiLbfAP6qGWg=HAk
z^tn$9`%Cl7zFp?hzl+Sl=C)qu(tk+fGNfNd<4UAoN#k0iUn^k98jwY<zW6K*W&v_6
zK;2<JiD5;OMq~n4Q+aa5SyG5i;P!K-Oj4XPp?vuNnJ3l%u6fGp(O$oNbJKRR_f=RW
z6JpE|%!goOSYc=LZcmK!eOvL5Hv`;|C#O;`7}*Q@04xaG2mPmMaZkVlgngfg%@VPr
zfRnIUl7OX}d-kvgV5N!4nF1rLrW^XeC_gZ+&^2DQ8QX8B8&1aV$#lah*epe8h-9L%
zS*n1gn%n!(@a6t>v-x)yH}S2BvtAz=KWGT6T9{rfH2%c&`b6W;%;3+=5P|=N>H39n
z6Yy82>sK1D#a-88s(c;px{k){anJR*m+<@s+;ansH)7w7f>yz*P1tvnfG?jIxkzWA
zWFuJ`|9@u}zE>c`2aN0iefp{XY5|4@V%pXk|0Aa;)qJ}EL`Y;4ds2#IrN~CmEA>cT
zk2*oLFV%d}00|Vx=@12_{*FOVQ2EF)pH@(Bk>o8hfeSJ`Sw4~iYy#^u_hv~!K~3X@
zf5{!stqxmzxT;$TxqTHO<|F7E0g+-plEz2H{81W5iTNlRM~nGr8Xpt$$7p<9%pa%m
z2{C_y#xY_(hQ=qw{7C^9ER7k}@_ga4|5MEW%ZeHDd2(_C@Y`26!b(1f+z+B&bi&O&
zjK&d29)XlXc{Fm5M!g`c=2gd#`!O2FA$c593guTB+pF~Tsv^d#h`vr$%y<>kI0P#~
zFm;W(8ar2GZz18doipFg*$F;<EN2}{U&l)2+*9f6S81Gk8jZ_2>vH-!RvqVFCtSZG
zSpd$xUcf~?>{g8qu-NyHO;z#uA5Gt{IBZ3{U4QH^10Sm7&hqh&);WKWKZoKDp|~rY
zvD~`vbL?Yae1B4~1(|F?T_EPa-KF+Hl4?4+1FH&0y5Y!Dn3@wv_XM(pp@#^L*sNEK
z>=k`wBM_ShV(JFt7Hq#oP}RID6MJORS2(h;S(b2RgIAr!W@iN~)qI}?`u6WlcTKtF
ziaY;fE(g-}q$@#K7KGbw4oEehbG6mj7(V)k16zXrCH)Cshmklz{b?BNhBEJ>e$BC_
zqJKQwaC@5c>;ohFAl#84Ir?v3nnBOXO{1N5RqDDB;vv@M5bGucui>mqxPU8M*+#Lh
z3HefloGiC{!pVhX(H?QTJz_Z|i(WWo#JO9ZiX;o?AZv1vb%q+@0xF4J)St2mfmtIq
zZN%0<<W*N$$E&P6<f-5T9AZD;FtRuF?f0Wt?<j8bNu-)D-=K4lRVU-D{@x7?D;y=0
z8tWNZJ!1o21^E;P+i#^WE~&qQgBrn7lN#0EzCn!*;Th+<oM~_Sk7w>JKL5x5_m#cK
zIN!xqcd-XtvLfLt>$r#Q4d*d$^{Q((1P8QUTKISGv_dYPyup8-g_2SH_DvrQ27Js%
zh>eV_k<MSJKj{Nv)9Rcx#w{0h|MgdefQ-yzF_IP27lI!n`D4@(PQ}UX>u(<t!gMuj
zz91y8{)mx1qMbbTcZBfjE$vs_bMABeA|aMyt5Q5d`1&i9bqr-E!F05=96R6T;l4()
z>;oA|Ak!Wq@z3}qpJ<SJm=u(-xP;XK)%)4ovGLq*9T`%PjeN6_Kja_Cw~GsA^d~g9
zU91KAp?|FQGsF5n#*1KU!2J*{cwfkVXu;-Mwsh9A<E>4Of_AmECI~;r3qJ$eej?SH
z(SyhBezxJM&g3bRMvWdbcHH<069YzSO=_cPO%#5P6@HErevT1-jud`Q5q?e<evTG?
F{yzqz)%5@X

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.CA655EDB759D211C.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp.CA655EDB759D211C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b2277470981902e94fa848bc8c5cb3616d03187d
GIT binary patch
literal 9360
zcmZvh3s_Wj-pBc$%Q##NO#w9#OUYDh&i%?s@s@;&f_cI9Hed>Z8e@dp<YlFFv1PKg
zl481$B7rUFSj1+=m|?w@M!l~s=JsIMEu+n{Epx}-1HA{n`#i7D(~0l==JS0%oH?9x
z{&R9NXU^O;HY#dbL1uZ$>P@SdsHmvH_J6)jrG=~=;s-`WmDl{w!ll`tYC}&y-Fa}&
zx|lRG_xbpF=Y}7p7qsm<_2%L1>l=?3lpK6K^5cCqu@AJZU6fpLH0Q~4Uxv<HIW5m|
zXSxq>{!7}B3Vr$9-Ttee?@Ilw^j~X?*YwwZzAf96_-SZ+*AKfRUR`Tm{M5*JX3ar5
zoRzRhji0~KXTRlALek`Kzx?X^Z_D;P<!P$RAG-gmmd4e8{r8q@%`d<D%G@(e^S@90
zu0&b4ePUkTz@wLEbDj^69=T!TcyV_3vCDgxE?aU&V84o~Au~dG6EY55Xh_`H7I}Tc
zglheZvik<MjEmk}aL0uq+3Ps6Xl2DafAkK&C8g!`lBBI4c9lK5w{F*yB^&o0J9XZ8
z_mkAZg3+TERISU;e9$bU!oHQqj@Bjokp1*=|IK?A`wB<jnBcza51%C$41B_0*7->y
zbLocR<(bT9hZY7$kVom$)CuZkZxws&OY^<6!<Zbsv7|g_z|zl#?ROo!u%mHv^XZ=k
z{eI`iYxVO(QZ}s-&X$(XIKs)R+Mfu%^y0V|KDKK2liKRn{og#Wbw{Y+-I{`_rPKBG
z)f2zoS5@`!V;|>tW!{<*eSQ9<>PO1&-uB$?@Po_P;Q43P40~kOuJj=vO?Yton<o~>
zt)1AoZ_w2je;T_c=SOP5sPiM_%<cQfzQ55LRJ7*3><p=H!1HfBRCQ|7*I5_2UqAb1
zFO%hMRC119ov?mG>Wj+6eD%TNePYKekCg>BKSsaxb^3tYi_XOSWBa)=9g9!DvrGBp
zerZLG`<aOkBzzE~wk<43_O_fHeb1IR_FXvr`U-hiRcgWN=MMPD?z3|Fv{6anw@w^x
zcxXb=j?>a>SN(sy>z~^$33*c!4#%b}of#cIapv(wp)d4PYi8f6iu1F+kslfV_9FL>
z-(<G`9Md@Dt?|#V-Zv`s<jwnIQu8t%-KQ@#r$4YMH}Afroo$`t-x$TS$xl~R&D3)E
ztsm@bnSA|`IoCb8CQ^K<>Vv=iaQ47Ue|RZAx6r?G*p$2l^1H)avD5RmoqfJ#^Lxb~
zt^e-WU-&UQPrrC?^Bne2>dwn6;_H?^oAYf}+LW=6eEw<Z!gKF69_22rySwD+T@RJt
znL2cLwehxm<^D~AQClq?-J9@b<`0QA%R&ih%_OpQ+~xf#xsi<Z4;2fc2}@S4+dAXG
z<I8?(|M5t0T-M0c_yeJP=HGm8*`Lpk`>AI9tWi%$KPG)w{7Yi~yZ??ce+^b&efr#e
z8%)1k9T!_F6i&*%Inw&p*eh;9f9!*b_vv5P&pUqNO#R7ok$YymzE66!vwho3#}~Cd
zxu>%&W#yfpXx5>^+qV5~=lGXG_xzl2-;T-KI@WuBxpLlrHm#}Sr<FfU9`Vtu2j9JX
zYhhTw7CQW2(?)-qU-P9R{qt`Jf4)}c-=0!A?9{z!Morhl+1E(plsDdv@@+2h{+UEY
zxl&%Pn)a+ZI+nU~X_R+}Ye;nO#y(&4{)_qLnU8m5$+@xA0((5%HGFXI#u<x$-*jnS
zQs%OOQ91T_Wa7vhdpFK_^3_Y-^<Qjmilr9X<0MzofZmNW&bui(X^k+-6Bo6_9%rx_
zn|e2W<M#90zP}^?yIAUX_V^apErq=sXS}xk(Hq3AcV3L87TMzwt`V`l8)v*@<%pS$
zyjT`X&9%paT!UhIH_o{30r~b$<=z2tQF-=w6}zgqcjJt|UN!o!lca7jmYQdehq{JF
z_imi=cP&RQE{NRj9~hNukCPLVNB3@=ab(cKBY)p;@{3q%zC9l78XVobamK&CX)Rc?
zsCMtbsKxd;B{5}8@5UKNm9D<a_kFS*|EY}H6Mp+J73a;Jmb<8LcX^kOS*|+C`2a~w
z?+5p%6IY)(bio|$d~g{ty&stKa-7%GN<1e?-v8}eJGT>02j(u~xrn)&cz%Y=wGs|_
zuAOkmb6tc(p6ezY@_Z}dk>}e9k38Q+c;xwR!Xq!V5&?Okoe0PaT|_`$=q3X4Vk;4m
z7u$)5yx2uV<i&0xA}_TP33;iVNXScFL_%KbCKB>;E0K|x+lh?4+(l&M<!&M)ue1^c
zd8M5w$SYk$L0;)53i4_zQIS{MiHf}1MO5U~ZlWTuwGs_^t(|DdYh6S`Uh5_r@_H-L
zk=NUaj=bJQbmaAJq9bp#5(9aoofya)UBp1%=q3j8W-Bp~H`|GcyxB!e<jrnkB5$=4
z3wf)ZSjbyl#6sTcCKmE@)8_X5?Z$bt)3PDWPn!=ZCoKok`t<d;_nox78`3vGx;y=D
zNTuneeJQa#=G6WF+LFoGr;xPV4CkZ~=UpDR{Ko#|OZUz1FW(Zk<whs<o#A78&+Pr@
z5iUZwT~z;hgpUv&vk)NyW-&rU%u<9%nB@qOF)I<GU{)hU#jHh$hFOmg9kUT324*ut
z%>FE`_+@^@fpD{ZYMthuHnD$C<PXa?o!?!3?l$*b^REO*oOgsa!g4M+&Q<MC_9?gL
zn%2KhJk7M{Fx|gTJYm{%0&@rLc^`9-@jS)A`<oa~Gv+qN6UH21JrA?+{#w@a9Oe+~
zIf}W?<vHl;e?QkubLhu~X%79k4w^$hF351`$2Bn=`f+Uxhkjgu<<O6-WjXZYLM(@V
zT%C(UKfamf(T@+)Jo@n+G>?9Kkm1pfZ(?}#<J%Y>{rCXOqaR<(^61BhSRVcOIv0<A
zLNhI(pAe=6^b<O00sVv^BcPwq#0cmov@rtu2?16>KcSWt&`$`l0{RJcE&=_-W?Do)
zF-(i-Cw9;x`iVhCL_e{K5z$X<V?^{51FVRCVl6A8pBQ3A^b_k`BKk?qw1j?An3m8_
z>Yyd`lY)$deo_-7p`X;oNa!a8SPA{4T2?|oDa1<XC)K$m^pl%u8U5riEu){@LCfeT
z2N@at<R(T&Ke>&O(N7MrGWyB2tc-qgh?UV#u5-!gr!><F`YB;rK|iH~R?trgG79=B
zO^kwmN*kk~pAujd^iyhC1^tu|tDv7!=Tgv5ZKhT9Q^T~1ergA;qMsUMRP<Au7#020
zHbzB1HNdLqr`ED6`l%sSML)I9rJ|qKOl#<;g=r1_v<_NBKP|{;=%+O?8v1E%jD~(%
zfYs1Xt7SFx(?YC<ep;PNLqENl*3nN7(>nU;9kh;qdXUl4Pj6y$^wZlI9sTqGtD~P@
z%j)Q-hgco`^g5T0envBGpq~+@4fHcQXaoI>AY-7P(Zm?&XS6W}`WXS%KtH3FHPFup
zu?G4XbuI(_%x2m|KQm05=x27&Ci<B{#za4}i80a7Y-3FHGXt!Ner7FeqMsRJP4qMC
zTqgQi&9sGnR+zTX&+4Ep^s|DDg??5OW1*kb##rcQ1y~FHtXkGWKP$vq=x5crtbV_Y
zo!>-VxxDEd#dsIH7q4^fi5PFWyL_E<Pc%n;WJBVlhaR@?iH2*#r#Z{a1#_<-_m4g4
ztbIQ<HF|2i!^U~N(cXB6bzZgan|c=e&V7?}e?rIfuJ4_f<Nm`U?msM6l-F!n`RcJx
zZHqyJ2Mu%9`pzBN;F#X80{WXgoy7AoE@C?ghneprJZ7Pj2$;oAB4U<0iG*41Br;~D
zlPH+gPNHJgI*Eo^?<6{Aqmvkz%}!!swmON0d0yH)-r2ix-i_%SGn{>1JS*_u11H}y
z`#O82@AhcUzK-*9ZVnRf<{=4g0g~tzAxUltlI)ftDQ*Rl>Q*6XZVi&|)*%^g1Cr@B
zAz5y#FO_9Ye`n%iWuASMjhs4ifwMM=^CnMCUf_Huv;F>;E5cnDTJ2+F<ie55oHseg
zNcL~v1HrLT<XPXpTRc9`y^#E#ZE)ClDu`zn<`D6;K<0`#sN#GaRB?U|tGEh#`GjbQ
za9G6`@leJ4c&Ot2JXY})_R0&Y_z>Z-N+=SbO7IC#CHMub5-RLP8&nA)B4CwRBtn(w
z6QN4<i&!OA*y}y05<^79Dyc|<D#<57mE@PON~*Aziclqmh=f&gkqlL`PlhVlFJqNl
zVXrcwN)8bjtCS)IsuZ6BRf=E1Dy70+phA@rA_`WiMJiOOJ{77|zlv3Ag}t_gDm6q@
ztkQ}!sM35IRB3(<tF#JxSqxQLh-g@)7wJ%?`*f(%{W@0Z750i6s`L=ivC1eipvv$W
zP-XZHtTHO>#W++MA!1;aS!6<$=`*3q^qW{^R@m!!s4_#u#44-Ef-1{rL6zmVu*#|+
z7FJmyVqw*+v{_vLSv4~)6Ve@NcR<QY%Yw8veeD$I{2S*jOfQ5~lwJgBUHZDdw0l6?
zv%5}|ZX0c%sRP8b7k)=ika%h#a{<EPoC^{T=X`+hIOl_e$GH$70?vgX5pXUBh=_AB
zNJN}V0V3gC3K9wDa)8J<mxDycxe_1>&XpihaIOZ3igPtcRGe!8qTyT%5)J2ifao~a
zgG9%<5g-Q6jUX{_ZU%^nb2CUxoLd26;oJ%m3+J=B*>eAZGl!c4sd!59jQ)e8WJ*b2
za$dD>k9wX*I9I5~i#s+ixlmJOU(d!zkH5~jI2|f~g`HPYJJWtS>AIvu=S!&<*jXQ5
znpLRVFQ-gTS?oM&I4<QBurq2e+Aj}D9Wu#zXmDb|ZC|ZO`S6r|L0!l!ys6*0j48Xx
z@o7JQ{&>;ji^4XWyfAr{bDjO1V=oVgPuksc%a<@UrJv0on4jdZ&T|y|cj|eF;`~k*
z-G>5{(cE(vIW%=B=e*0Qw(kUco~bzRw*Mfm*%w~;veCL`|K;{&=W$DXOz&eBz#53>
z1^l~dARK1Cf$*4x1|nb<8;FQmY9JD3xq---l?I|<RvU<lS!*B~X1#&vn2iQvU^W|w
ziP>r(R)3CY6E7d%oJ!jtfG^j#!rA3<-YsLcc${=?Rb=+a81nm`;~_6)m2*5e|JP%B
zA3eb?tE4=S_w8AG&sS5PI?VNy=S9pdl;<R5u9D)A=c*|Vd9I$~kmp(`4tc(k;*sa8
zDIR&gp5l?`TPPlRp^_4i7pf@%d7+*XkQZ7g0eP{K5|J0HDG_<Go)VE4TPP8EsgjbA
zm#QfVd8wX~ke6C033<7al989IDH(aWo|2K5TPPWMrIJ#RSE?xmd8M9GkXKqL1$niS
zQju4yDHVCOo>GxlTPPKIt&-A^*QzNEd99w(kk?u$4SBti(vjDzDIIyep3;%mTPPiQ
zqmnX^H>xQEd83{(kT+T=19`KOGLbi{DHD0Ko-&a)TPPEGtCF&ix2h=%d8?kXkhfYW
ztDi4jUA(I2F|e;R-<R9-!}6ZZZ~s<0KSV_>>-&-RF5JGwitcphxbl~+T)tve)QY}Y
zRBzR9F~9%S{E~v=wRX(9YYxq@t^w}o*tmg%;;+jbJY(9;gJUwMUY|5%=&<3MZl+te
zO`dXV!bD??a!bnp_iEw|Ba=tpc+=R_apNbXc{pB_)lt%jNy7Am#M|#Epk`)g%p$Y@
E7l-7Wa{vGU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.54BCAA9E5C4C7CA0.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.54BCAA9E5C4C7CA0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..be8b2c55690f9f26a1fe60b536a152127857a53a
GIT binary patch
literal 9372
zcmZvh3s@9&zsL7Kmu&@9Fhd~~Q&dcBXB_4}<7jxz8(#C0DTuetOD*!IyjfC4`fd{T
zinXO!NSLv;DB&rC6iho^u(HTf@{*?EwO)NY2TMKUo@d~Dp6B>H4t)2wpD!Qw^5;K0
zlSYpiu_@W*O3xm>a_RgPnT*TjinIUctne?U?eIs8%e8V#<Fpyek1Mg?Z~Jk_vkSIl
zSG}8WBp3dacszN;%e`Kr#uX;l8}A!WT|as?V`PVZ2jiC|&Lv4_pE~=cykk?`%W+wO
z2WRy4<kV;b0}VUq%^L&1pKLjM;nH6hldWwXCj5;_+ID%zwXO8GE#I6jJau;E$lo40
zc<I6e?cRH!@A;F1&-{G+a`Lv5k4(5$)pBXYx@V2sgI+&>Exho_Z!f<7#b1{?^$C+X
zht9;m`pxIQH;NZNyXwQz>e<hI5zO>md}~M3g=2YpGK#!)2^lNm-s&_Z@x9VD=9R0@
zO#bZodCUHM`}B>J-v2eW|6d2cO&q)bp04*rm;OL)nlS5pXw2;aj~~w%J-NK0uQK+;
z>ddVRnqN*E6x;Eg{nfj&^Yz74UC!K^_sj3O)-h@D>g4s)a~6+$&=>n9^O^piiD@lo
zcOM++`T6Le-~QoAnLcG}BB`biQ}xsbS%qxPdHw4VG1Shjjgs;9bM5^xljsw_)eY|v
zx@N9D@uocU%9ZsU_X<IuTGMmJYolJ>wV-0@^|Ho4*QG38v*;X`^iuuRQ>iofe!-${
z6ANdZ`~1}--^3j`^7j~{e{#yuNufTIJ9)31`ZcREb5?pk`PYPk<hWiFKG`!qv*+%#
zc>37qzP_KFIrjHS+|Imwy5q9vUn?6jsng;QdU(D`xIQ^{L-SBy|Lq4d12;a`zO+=e
zZ{5q^y|bb7<&W1Le{A%I1&c2oD-v`1t~z?;3jbA9&%pG~r$${5t?jX{$NF;rf)Ve>
zUESVtX_#yFfY7VcK5*|zt|;2uhuwK?z`>X&d)!(-SJ_-tykYZUHg)FFQ|(th!t@*S
z>4p1uu5)i)d@Jc%pVw;XLlxQWrj|WWvpiuw6TR=%uMRa_WuBS*&fehqnUlgBpWpq|
z`i!qiQoGK{%I&50sgBukSHiA@o$1d{%jmP<@Q4@QF1oU5#ML9Qidz0d@A9$(XX@%)
zBkw<VHSL$n;TM<ezIkqjHLK=K+P`aZKYaJi%Lg7wOzP?`+MHV!T~J4DeC3VYD@)_j
zj=z7e5$?Sp`{ajzxcBT_Q+&4Q_W8P}dQUmL`gwNUuz#HXtZMnrS5_@Pa`fk8gNwTU
zw)b#K{Gi;;kNxZXsYQ<u>2l=jhrZvk@XVM!Yy1DP_4cYB9pB*(T;2cl+^StuXXH(p
zuxq=u{Yc=w#oy+gDBP%yFZ$^7AE!_Ial`C||MBC%g;Bdd>Z{*mjo0@bnl9}7q~MwK
zbGKgDbK;15PT`1;{w*g~PaC=YR@0(QH<$HdH-EkEh<P$KWzNqFGctGH+FSb5%Cx0F
z+?X(H-(#m5RxCO7Waa%&_^%vIxjN*%1D%f_>-zM;6Nf)~bocSV(?+jtX`hX`KJbUX
zCiGo#b-wWSfSLY6FE?$KCwt|(?Z*qwO}r6%-0JYv@9A-8{ySy(LSf7KANQSETQkFz
zvubJ9KZ(obPWqrQ{Y|-NG&OaWE9-9e-R&X=d$oxCi@Nc~iRZ?MlcT9A_Bg?v5EnT(
z<GAqp#N78T9Ehee>~X9+wq4}lj0c?DRL?JO`aZ@r(;m-b=PimHoU!%i@0Z^`w<;V>
zO|!>c++FU99GvmQ+n?NsyD%`Ny=#&^?iSzezR1BDFFI6wukg&&i_z4>_Bhd<7#%q{
z;|-@5ZTZ*K-pw(t8TL4xO@BOcaK>+3QL@7?o?O-5HQgR3#V7TS9GvmF^P5YSd3r91
zrXI1!Dejc{k%Ke7klAzZz_iOkG&R8<cXoGf7dbfNKZku-w68FAXnWUWd)z&~d(X(h
z8Q+YZw)>~$pPY-PCfegTcU-&3!5O>!^B3gY?rz7cQjd*wU+kh{vZkg_eYo{E%9`7I
zuIwb|2_#XGCvHRipKpXt>ph((ZfI2G33G0a$?{ec?|$O^es-=W-UiG~#CsldGx1)B
z%vBQ(d9I#t$a77EL!N6U9P)fM;gRR-36DJAM0n)+X2K&cR1*Ptp`HlH3r$2oUT7u)
z@?td+kr(TUh`iWDMC8S0A|fwU6A5{#o=C_`O+-RoY9<o$ay5~Wm+Og)yxc@&<mF}}
zBd=5w1$m{OD99^KL_uC@CJORuHBpgQ>xqiI+C)_3)n=k1uT>Kbd99vk$ZJhRLtbkp
z8uEHI(UI5diH^M9M0Di!W}+i+R1*Vvqn;SZ8%@MO-e@KU@@6$LkvHp!iM-iFOytdG
zVj^!<6AO8(o><6RO~gXpY9<!)8L1hq@7<WJiK!DIO-`K*X-eu8NRJzjk8E8}vz{=X
zfRtn8Kw4p}Xid%nJs~QxFvBw7^Dc!&!0*k4l;>RyDW7=P!UEu}B;G2>oR5Ph&d)&;
zm&ZX9mrpos;wlM;O}vkXCf?6O6Q9RJ6Q56bY~m{kk4=J4fF{8&K$DOsK$DP91Z)y2
ziGWR_PlP7XFG7=;Cqk2$Peg1ID~X6rl23vr$uB{ZlqW%xlusmVk}8RWO|nmhCfP4T
zlbk0*lblavY?3R9j7^G9fhNVTK$DWEK$DVB6l_u|iGoe4PlYDcuR@cWr$UpOPgHDD
zD~XCtnool!&96a|mZw3JmQOTn(kh9DO}bBqCf%<?lb)wTlb%m>Y|<-<j!lNofF{Fl
zK$DSYK$DSA3~VwgiGfX~&x9t^Z$guqXF`*iPfTnwD~X9smd}DF%WpxGm1jYdl}{{e
zvMPy%O%J6$<aM?aF<E0$$3PmJIu_En)Nzn}hR<qSk{&Z2gS5n00%@tSv^B-AjQVu-
zA5V{F>|IUj)S+!Zcy7$x``S{jkvp-id<}_NL+*33^VQz-+}Tt*t6JKe8xz;QRaa^6
zmxg<W4`}O}-*vw3F!J`no&^(IULrABiAtjB>>{1L=mxv<&Q8?nv^Pj0+WRgI>tQYJ
z{TOot?LCY+$ar64;QkeiH-x!{@z!Dvu-+G0xPJ-j-G({LdMh!Py1m=oZTIIwG>3j%
zEzO}H*FbaV#|0S<{kRH-LqD#D;n0r@upIhvB`k-2T$tt1k1KU^=*Nd>9{u=Qnnypr
zf#%VV4>COZ@f8e@etZqXqaPn&dGzB;SRVcOFw3JKU+U)3PYBTh`U$nPfPO*)Eufze
zWCZjRDi{I%gc?RbKOw*h=qHr00{RJIRzN?Y)GeT&7@|e=6KiP^{lo@ZL_aaei0CI)
zFe3ViHH?UUVt^IVPb^_Y^b^CZh<;+JTSPx8L`&!=)zT9BNe#4weo~N;&`+vhB=nPN
z7zzEP04t%NRKiN=CxuxF{iIU2gnn{}meEhHrDgP!8)zB*<RBxXpIpJn=qJ}OGWy8@
zRz^R$gq6`x4zn`)$)#=?{ge=`pr2ApE9j>*&<gq~K}JD8rGin=PpM%P^iu+?f__Q~
ztDv6}W)<{PO5F<jsUccLKed)t(NArlRrFJXjEa701*4*$TEnR5rv_LR{nQdxML#vn
zs_3Vdx>fYkLbQf{S}m=ipVmNY=%)o44gItVMngZXhSAVZ3$PmcX(g<Nep;B-&`&FM
zYv`wkXdV6ZT3Saxy@A%zPY*IW`so#nj(&O#qobc5V0HA<OIRKK^f0TVpI++L(a#9c
z2KpJbw1Iv`18tz65o8SXGb$JZ{frvMKtChE8t7+~um<`WVb(xDqttDnpBbV}^fPN|
z6aCBv+C)Dy$e8G7Rxl>|nKg`werA9*(a$VlP4qLvtciYRsoO+9D@0r9XVuac`dJOM
zg??6$vCz+|U@Y{rY8VUstN?4FpH;$I=x2pl3;nE8x7Fr1^o2{Vmg_5yQB2lM&&-9+
zo`}iH^W-gb_C(|4mzT#6e6h&h6D24K>CR_ItslTeMLu|I+Y)gt_K6(Twgur^?9(yG
zLW_M;1zBvdPmLf;E%pfxWVyvYO@XYm*e4s1)fW4d0kYO&Z~P(aE%x>uve9C1!XcY2
z_Ex$zPZ{5-->(}D*4`!#Yd5T;^N_@3Ww*=j=sYCOzA!E-@~L9m@9u3R-eY+B+ekRf
zd?Vp83ynm;EH)Amv(!i=%yJ`<F)NKk!K^kC6|>ezG|YM<(J>p1#K3Ge5)-r4NG!~g
zQYZ0k4~gIK4|N`rC8J*YuDz8qwu8MV_IQ=Hr=Rn1ka!OdN$?1eM2`qb@<@<mj|@ri
zD3DZ-3Q6;5kaUj@$?zDEOpgi4@>s3uZ)1jiIbgb!VZR8vJ=krE^HPt=>i%H&DbDA&
z`Kv!zTX*X8hxYrT+q7=8ojbko+wgAcS1%^oH^;`t#yg*}1n%1Z=DK?Snx6LK6CmCq
zc#(L6#9IQH3lI+HT##@$=L3YtIUgiE&V>LGa4rOifO9cGM4XF3BH~;M5DDi}kVrU}
z14PET93(Q%l>kw2t^|pKb2UIzoU1{i;#><54d+^rXgJpcM8~-vBs$KG05NcG1c`xj
zGeAt7n?YjY+zJp2=T?wdI3LfA7u((h6SxVGmJL}p+<8mHWGx@Eyfrz5eF}N!dV{lO
zIF}mKsK_M-=Vtp=Qb2kC*}5#*xtQ{nVlJb+yD(Q#-u;ld0*XVPE2cQ)xiX4Fo~xob
z<oN=MN1iXHc;xvqibtNWqIl$m0!lz$D5eDDg)&M&UZ|o3<i!F?L|!bWMC8RXN<?0)
zqD17S0!l(&DyAglr7}uFUaF!b<mCcNMqVzaWaQ;CN=9C;qGaTi0!l$%DW(+Ul`={}
zUa6uK<kbR7MP4nYROHn%N=06+qEzIy0!l+(E2cE$wK7UWUaO)s<n;ndM_w<cbma9i
zN=II=qIBeq0?I(%D5ebLjWWtW-l(Ds<jn%gMBXf>Oytcn%0%9*qD<ti0?I<(DyA&t
ztuo3&-m0RkHooHGbE{^Y-co3<;{Dt8@8JA73$6T!%~)DG!oImnhc5BX=R)tYqmKSQ
zYO!kHoHQ(Hy7Qj?IBilkJG}Iwee>OY?jGn|jH{nAxOq;}(NFEaF4LLmy_|(1CTk_L
zvX`@X9BOL%*!RlAbvE05TK7!nkIpX~d;PACT>^KuC-)3XYGdcOpWnq{9a}moa$T}*
zO~FlG%>UWsIoB-h)#=U!O9#2=3B5#Vm3Q`jo%?j=oO|66AE;wW!$<6!2eE^aoNJSg
zr2TX6T-J0-?6vELeU$*#ChLa%nPY2qu6TBgid^Xgtekjv;vdFx!eQpi36EJQCjw@%
zoQRmEaw1`t%ZZFxDJKeMwVbG!wQ{0i*2{^G*(fIlX0x1_n5}YRwdI@2fa@Qx>O<R4
zW$xtMInI+8leLEQUPHW2zLj}%Y_}+qXFt|SGm|o%*Oq-9cuD4+2R_F?FK4p-mCpL*
zyJP)+nKo{Nd*1B1kIu<-J?cC!F4x?)Z)dlCJE!g2NY($wd2P4OTbjLOksZe@7)LX#
z`z}wr=$Q7g9XpJU8=l@bE^73Hox0rJxod)=>V`Sk8uGwB1GL^!O49##Yy7?4y7#=V
kS8|_z{qIlpa=akQJ<<{f4jOh({K$v0sS)Fcjv{0K2YTp!U;qFB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp.5F249F65D9D49000.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp.5F249F65D9D49000.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9eb0989e92b0f0869f7d1a8273bd6472d4a96f06
GIT binary patch
literal 13924
zcmbtaXIK=;)}F5J#&KwdoEd_GAVDMzAPCri0t#YS!y0j2!z%8sd5!2QBA5^Z=Bx<H
zsu%zX3MN!QKm-I6K}0|lL69I{;moo7{`#JK*XOx==Dqb&;nX>&PMzwp{=<h)l?nvj
zv;9|w{<<`T5(osU_<u{5E}nn4k3ir#Kp<FoxOh_I@8weUlG-lKHDOBg`rOzt%gulL
z#jwVuN1~tied}I+bTnmCa9zb;!_P8M+OU3oknJA7^vPvc1XhtgX`L@VtxR)!dAH-+
zA<u^MYa-|Qp~yKi9;ipv93xwDjpmyDl^8ujuk6{os@)k+9;gR?)r_tdPZ(G=wdj|@
zgIbiByffD~-);}K_b4)sb=`6M`e>)yxh5|g*ZhW-YOiUY@8$iqSiNPy$-{m-H|qwp
z`MCuwQPC;fp<iFSaLau0)T^iR&gI;8R$b;^xjOTl!D&C;>RInM4_bRi?P8U2$h23B
zQ%@DTE;KEx-H~kVwd9y<n|@Q<;ZG-KURX46^nuBH4)6OdFnrqJ#K3!#Z~I-iWHqaQ
z%#>(}c178>BhdrS<U}r=zI&nX&B5Z2<;DBAt(hM?W7oB+Z*MzpeckIm`>4+RK8aTv
z9_GZau?d?~cjALf-@7}SZ<6QkOl!ZqYtQ-<Z>AekN2eXD)SOUjuzgkEe3#(1mhhW<
z<}R+1rYNm2_ByYkm%Mb2Qm5iF)2fNncbVjuN9%8C-;|x>5Y4PF*qEtyWuxuRHLfr0
z-`1iTQ$N*>S$yT;%AfP{cXlooXH5M?)3~Z=dY03S9gWfyg&8{@x%?9nIAh0%#s>pt
z?6_~!t}sJu#L?LqhxIe2&UH`MD$ig3*4MKx)!B9meW>E=u#xW<tq%%vJ02)LQW}aR
zhDQ$97+%w#I%MEK<I`P!HrX)4)sEb`;Po=Mm1EaGW@4=NUkf;p9QS<Xksf<deRbNL
zjdjV9>XwDt2dpOeRJ^*=oT=&_w4&ngdWBP-KhN3u>A>nm%^S_kt~{UD_0FN++|R$A
zir%eVEIrkI=4i`barvF;McU~{>mGUhb9KU%Ia{?8<NmPxY5MO!y`}Gsa!H?Zpghz#
zH$idpo9TO|PS4!+ME}5^Ro9b08*lR;Uq3_3I%1Z8t<RGwIlnL0@2fv{MbPn-xD|Hi
ztWei|(@En$?;f5%+UG)q_Vl@RcP6#e-gln=s`65>pQiphw{5PP$8SelWKm5aw#8{n
zO>XEldpj2nI`QPy?1G86UG(ncU2{?}?H4;|_TzrVf?wP3mHL$HtvlV~yk%g&9phqj
z6Nc*jqIFv}K`mzLNxOAp)^1ty>rX@Xj9G3~8W>moO{34pq<7hiv)^xsx_+=!>kivK
zKfxpL$T|PKMcU&hb~l)M?KtaLdiJrPD|+qbueQFy=^=~GMB7?q9}O9Sh(3QQY#(0{
z_wZ(5Gwo(AKJ#9CNx56IqN}d@?;E1}=h&6M361`%zoz@D=O%p;O{d+qeUNqISzEE{
zA7SxD$AT^|IPJTA?6%dZGu3k!84uE%x3?f@#KiC+FCx#*d$z(e>-QYj>~H<U?rqXL
zzGFjZ*pLy|c28sGo1GWV3-mv-bej2r>`9wnhHY^#VeGW7XFR-bQ)KY>@nysH;<rwc
zB(y}Wy606qderZerkFhr`ss7ff~e;k&M%!D`rvK!{IG-JW<!II6_mB9FZdK(GVipn
zzp}DKf6Mb-xuL5+oOmdleY$zZ(lPm$&rLgB?B94QNIULD;M#loHRd<p$4&1aD0$}c
z*l|x&rqjKrHyWL9Rc{H^o2Fn<$2#Y2NWDDzuau6s^!*(!r;ZG4vQF4@H?wm7d%bZ>
zTU#eQE=`N+Zdf;JV(8V0sf9a+`)`?_9=xvS+MBAXLsh@r7f6g;cFb%rcxpG>Lh1f}
z#m(*>lQf2`-mbCvrD67_EJLIIBK;1r%4XfQBeoQNc<JpO^3Bt2kj6fVu7*Zph>wMJ
z_sHyFCNDg+(yr+(x_9u?+QU<i7gvu@c=OwZr@J4=>pq*fq<pGg0)2gH`GrdM){bk%
z%jVTbdaM)&W}ZLv<>^9mvN?9mtd&uhR+%iW4xGLA?5;xxu2kzE^$0z&*7j=C>_WdU
zN^SeT2`4=ow_yI1M-B^gJua<&c=1f>oiTTkPcM3tex{x7Sm*4zrS@FPpCetK4p`ZI
zK{MrCQb5;U?Olua>aG3rQCjfU{rh7c?JtSG;+Ar)w9gjvrOON(r!;<DV(@dO;k<}h
z#J4c5W66t}hxue5c9>`5R(WgIE8i)R)Ly?UG1X;flb&=RKp}z8RxGU9e0fs1n{7*V
zQK$OvTfS9ob?W<S&yZI~joT+C7Oe0zj<y(bBIEc^Z<XJ@SlnH;I6bZNRnB(z;G?mQ
zUcuqN%s3h5ZuWLfjdexljh0za7w`3>XP;U=>ygFIx+^C>2wwPH3wzkL#rN9JGao&k
zIcc<!b&=kwQ5$})pS(=*bz9N-(<j1`XTE=VcARtI>ADu-%@uPpP7l~?WP3g-;agg@
z$6p7{EXyrjuP^_jyXegyD`yfCL+a119}UA*I|Ez?n0z|6$vk|kyZUU!Xe9D}ublqm
z(=YqQM^^9BP`j@(eAl?_$D<7X)JpgCDHxfMUs`JZ=R1Sx^Zu?VSu{QAR{2kARsSSw
zdITFKcK^D<8qH7MS=qJw>9o+KNm-qy8-26_>nb%zJ^11%>auEhPe-qBX=zv8oufM0
zK7HdD|J=<Jn_APOOS**9Urg+K^Kk9C^`E=V+eRKe+cKirqAhJ+*UZda{_(R3i<u^~
zozAE3_g(xzIXk^7EoSg=Hl{hxLx1tT>cLuRPg@(y48AFEs2yfgy@y)wz4?Jt_1>#D
zCR+C&W$ko#Rzj?IQGTUW6Js5eG0U^ZZB(Is#V_F(`@Btg)<cvXY>^0-{1rOOQ78~7
zF0#8-wZ37b8Zvbju$O3+OSHaR2yk=nKcx?`eO4qa*(*kaTm<Y#(%>U$FBd@3HB*1?
zsEXFBpaDY!Y$PR)q*UbsC`JwAg%KOp%oQU?+)Fd5)l3eM3!s=;SNboClXxm31>8#n
zC61u9<N_#`)6d5*v#-0Yh9r2puPD7&l#^Tl#rn_Nl}g7_CX3N%JkSSH{R3$r7eKLF
z^nUr_a-H{@NC)?lOX%ei266!uY3{x$Kh@Gg1*Cu}mr2FTq`F)H#kp9(9^0>RMT`dH
zfm%tYR&s<~0LAady}{oQyHkwpFr|mo>>*9%0w{qa%B~v6j2)ziXgtt6BzlJ=asiZ}
zwgn+`gEP9tNaQMDtC31IGL#FTjJJE0XmjxR0Wq@11AQVTpGZf!0Lp~yb@xgigs6#;
zC8o5K#_go7TmWTqPv(L_cU1EfQ6D^)_ek^}4<r*nnK{Z4eVLoNL=72WN)4q_L-m&n
zpv<2alM%h))*dxvf+@R@$}TiWE`YKCUD~yzdE;A|y_8boQfx0W0hC1#i<b8+UuL6#
zlyNV&km4<5E*C&qd|u<|)!7-V6p;{9>X4`ot4=0<P<jqsX!tqKRspF<1?)|t&rM>m
zTnPG-gt<UKu*53Q>*|SpI)XKk9_}Zf(96f>AQRBf7RUbN^5^a!RlYBKJKc9XZNSO|
zP`0jCvUwczC{T=ouznjz-3GFsTmU7K4D6h_Z!9(k2TbW84Le9Hx%ff3eJCy}eyF($
z(!hINh#U%$hg<+9YP->(L+Qd$F$%=JG?IFa*r;RzDBISKURKUHo>D;ictzoqVmPHH
z7eLuw5_%@tqzw<G)vG*OrH;nR`!0a8b6IM4&hg+xF=FssYLH3|)~`$eC3<LF%<T2g
zPKuE&o=YdG(uti-CV;X#`KPr3*I!G;$O==wl15)i8@T{VOu>}9bV==aF;c_>RUwrs
zJdjKPWe-C|tW{Z=Ax0LsmoKE@7pyv&0Lnq1wXYoJ?NL@i#&|_ZbiX9pNiKkL%&}%x
z_Jgs9G?6Z*<Pmy#c&B9oC>Nf4XouQPZWJRYJWvbCw2(vO0w^~uPM=t_sVrO#8RA}w
z36)~PMlOC(yz_?G_-u(5BOBaH8)?{vrz;ac$vk>C-aIjMg9=i|i!MP9CCF1QfO5;J
zGc--*i<cVej|WObDv8(!WCAF+Ew>dkWjU#8A!Xc477}J5Q@H@jgLe@T+OPXxR79${
zmq&!?5uqj*KzV3e+@4*`j8#KMn6jHz*^SLxCV-Nszgg_zl^&sqMq)|<p;thd%LP!L
z<}O{WdR?_j1({(=6>U;QJIDo4o=Lvl>n>d3Aw~nSUYbaQCVZC2#1G0pcUIK)bvvnr
z6fh-;6ef`>asiZ5ab>|J$1m%|XcVT@le+cT)no!FWsHhOPDb<vMWl`A(u_pSNJ}n&
z^7824$$2eNMk+`XtF8h$RA5(=37}MNim<)4|Knsuq=b8EM50FAi%bBeTDpqvnVoc0
zjNI^Cx=6Dw(o-&gQe#`6B#PbomkLtD1LY!zTs)9W0HwB}$7X_};!#B;#sk$OQ9aU@
z3!v0(s<UYC`>c-`1>=D>Qj(37om>E=afkkcl{#-vsG)wnizZZ>v0=)@56ZJuyA8uG
z;Skgp&!rNnRAN_?37~vbw9!@QHd1E<0XSw_3$0zg|3}c5LV_{&I9@LPb<PqGGl4+i
zrjxO|Z(=&hu=Rv;Jz)us!LSiXIRaV2d>HmDQa+0;;befMgAmD6jm7dXtu#2|;BR~F
z%5jgmgi$VG1}8Oxz0@d=Farl>aC|n(C(IypE6vq&o}XZFi=fyJn(3g$5Z;uo#bsL1
zf232mM<mQeDzcnLAYgD*6&CVw2@;m@aRm}qa2T>6JhePhoJX2IliMl7z9Ow&k;4Fs
z<hc=K5Xw4F*}+9G7f=)%>=x_}+NIbrZew_8tozuZT-!LeaX;ErHJ5u<)ht|%7@S=g
zwUUmIcrfe*%IX5;HDd0Q_*ui8g*qjdg<a=v*yEA*QKR=!!H||w>_N))AT@-;<&<_g
zr31$>#qOm0@1#xPl+#sq&+$5Y>n_2tzOKH0FfI6Vylt`+4~`il@dnqe!pCnge>bJS
zo3a3=QJwkqzRO>G;sy(Wg$Q&HxggE31*EuuGz8r-Y!T^SL{8@L8`A0x>Bix=q|RH?
z0<cJ)L4vMLQWwu_2y+j|X0V&m+D+-gMEZGNO=)&{BKp^=fgF_-$um>X0fTj-6G<7v
zhzwg!S(Q@`pkv5fvGQbTE7-jCwDu5zmhl0sqclP(jTrc&L7vFseSb`fACubLs%2R%
ztlDOFmzxBw&BBZ8ro6f-7LqiI{YH6x<KrHR?V)_Q_Aojqj1C42-bRb$*)U9nVc!yV
zZwUv^!T&~fe<KIZ!A~Ij6UY%9oWTa?bdYieW!Swvny5eX?(ZbU))D%3gaLF!vEfKR
z92r1I6nh#OoJJDp$lk9yQbBjJ^!w5z-IO6rwRdTfZ+zTC8TL>{TzeQT38R1Tof@QN
z3v)gOV=Y@!mYQIRpu#>Pd2-LKSa$LN0f#iIe*|R(`_JH{*RqDPhw&KfLe4vBPjDfI
zJx?p2r?uf^W!Q(b@k81Ik`(Y^98QWzaS>?@o6E43q*Wz36tGBsrGSaJu>0Ah4in}7
zT9vL2;bUKBC$!=b8wqJ6F$VSkTP~7DqA{=sc<M;|JhF#97-IP8;R1z_Pf8T~fo48%
zv6W(*Xr_seKhnA%X?=*j6#I$R{lv%3v~DvWf2MUm^KlEU-a>0~f8R>0xAO59TK5Z|
z-$om@(Z(FVoi=Re<FB;FS6Yi}@1UJJXlKBzdn)r}lq3l|*k{t{GZ_rIAHyb4RteNp
z2&_Ii``(Z5=r6^@Txwu07f5mZq6QXl*yoW~#GPS|?%2wF%zYdoaE>TC>~`p9-z5eN
z`<!rmPIz#tFDD$!`S=y#_==BT6OOO>xPh>0Ae^|rZy*92h>0BjKv;d?w>J@1O?=!&
zIJOZU+*Y>}j_rKhML2fxaW~=E&Bt4j)mG%h{ry%HxD`$0a1^qN;<s-@R@?Y^FLK<A
zJixRVb{}%w$Hxbe<3T<?gd7j?aU5dekQVp%acEE+a^Y}1V&eJj35ZGH<0RymgghXz
zX4tF9@hTrDBgbSuPC<?-9ESTmHiU9gTuw?jFO^;H!As#*Pl@Drc`mMWBVjkvhy0Ym
z*{`sP)Q7~FVdE&nILZ!8FL1Q8_5<J8G1#KIQ2#D80R9k)0@$4&An^la05<`KEkjOa
z$Q$k#B6+C*F4W`t`lg5es<H$3SV#A*qm{uOpoV~Z%qPY9q!}!ZVPBJ0uSpiLNM2ch
z#f}eBGk+KDPU9XoQc4>sbx2hwgzTMs;Y9fm+|W&{cXOUNA$8F<<uZkA+|Y=08_@tr
zw4q3W75t18KjSM{R<VFFC#m1Jo0+I%ChT=7?dSZX=@p^<ijUg}?KXmURCN*BU3|P3
zY41f2oCzI7+6VbK327%G2hL=Yk#;hNCw0c!y}kLjBR--2xAYE*F4$CLr54UqdXTUO
z8NiM+Sna}3qyg*`!^TsF@svH7%2fA5>L=-Ui~j4AVcHGXg;NKZ_QyH$Wm0mP90*Eg
za7{+1lrn`JhQTF|{@ZC=?vP8OHIis!IMm=U#5s<vG6YMRW$%AeJtQL;?{_NUkV*{s
z(Xh^@@AEy?s<>ewHBbkR<@x=yG>nzR=TsPWglL2o#8HOz)9^FqkiW>E*A}1=z-t>R
z8p&%Lr7?=v7AOkjwFPMe@!E3HfLvq`S2pYl$UGl8aJUc+C`9&=@-hv>K1JqFkpqWI
z(0~$T56;)?jOJy?fx{JOKn1dgyC=iGMCLD%17NU1oL<TbXP_ts8;9k4<OkNmuun*X
zC*%<DN!TK+dRZL}@=|bnY+kY=8rP$&S^xwbU`SyXX$luwhP_B>T%-nrdHqQV4xAs{
z`Q-Z^nx+zV(9xeOJ@O467Ix#LyMh!~ko{mzP@ls?%c^ql7YzFi3BU1!;73yUk<^Df
z7S2E@!}DAagug=~dF{^Mba_;Cs`oku*D*975gt(DSYBVEI6CC~Fg(X5<l2OSAg;hY
z5%(Y~>_L~knd;1g)GWwU*Nk)UEWYIGgnQ_ql{;ulh>q*7hUCpV6+Z<xG$Y++<O6QG
zamwZ(6UV9xIAnc5<{yv?jM+N}^Ck{&OdV?#S#e_>?!nU5k_B7ebZY&oz_`2j6|lDv
zx}`)v?l6Br7``CPIowJyt&}Es8N;?wOdB76rI@dL+(|K=90sq&CL*g`f)1c^iC6HD
z6hGu&G-PE=hzC%=#KclkT#8RE`5jm!uVVfn%H0DGB`aZqkwpcNn;LiaxcD=yyPCT$
z=gQt9Z&eQYihOyfkT5AEET8~|FYbh8Az{bI4u!-JK6Wo8yg6J<3@Ya5K*fZ6G2zXr
zwU`Jf=BLBW#Gqz^m*X@O?#+ZZm*X@O0nG$2UEYOkcOfabE`vRH=q}{L$0K*4u^diB
z7KzA~J5UmlG!YHuW1mDclEYcZEDKq4=X*9X%jV-d$m|Xu-$iD3`8XGO=Az;J5awBk
zhVw&~R|y)y4^f^KXgGu<d?7)?N~GK<uN*PBGApb?%22draDE}IM#_+o;F}Q=)*xl5
z9y9C>QuPL@1Gd7j$)svBAE%J2DSVtts;2VsO;YtHAE%M3X?&bcs;2XC2C16C$C;#R
zCLiA-Rd4a}ZBq3%A7_!OS$v#Ls%G<XCn@YCm7(<AoAhXGr1T-_Vc5Nt+Fq{ofOA}0
zvX3_9Zi4%1$$mbLr6sX^e1Mi5;Nyd|<RBj(q9up;_%JOw%*RJ)$q_z2N=uIN@iAI*
zjE|4glH+`Qf|i`%<CC=HBp=_Pr8nrIaOmUxq@~GxoI*=e_&Al8rt<MkT6&X@(`acL
zAE(pObUx0Yr5SvjNlP>N_!cd_#mBd4>1{sFqNQ1UoJ~u!ISkJ!IH)}$#gDj+7Ret~
zAgD#X_jxvN&f>0rU5kLD8J~-?rx^&u3@(S7U!n&=eH6PAZCpc}Lt(ua53HdF1BSB+
zhgsRn4Ri#)hg10+B+fzlT<~~_oL(XpoEs-<-=B3LKJP61%SB0WwXye%(z3LHGYV`S
z>!^$rmyvql%p&<?59kOUdoZzt6qk?&98vc214M`(c!u|p_&$GzPm$A8#6oxy$sdAX
zhVT%CM|et#pYqQj*>ez#uqRhFVv9{}0$xfjrBX{7!@0TljpZ$`jMMLNLp`liPn&`o
z_qNTPwo%!j80W!}luab%3Q0jP9vMkZ<Zv0KQ^wDW%P46X#X=IcFURj(QTRG*JjaK$
z#zWeO^OYJ}qlPwu4TS$6Z=meG3O4ZY5v$5IS<9UN=Uc^(Lb*u(z6Cwwn=9d0P6Tkr
z-)qA2H8G6CwS-+Q|NhiYc(xP6xZ|&z@arZ5IJ^njZQ_>#_aV=HXc%|g9YTJGPymNd
zBKwp4QsGtPcNGP2$6X5YOhLl{pY|@7uvs04{$E|5=`^UbXkBHFw{<fu*-Tr)exILM
z=pnkXzy_OI10ii7MsZ8silkf7C~hfnNIMSMz*H}+yh#U#o*d5L%#M(L;Fn=O5WXLX
zK(3=ELfXV{ZzBAfh+wWg3Q407uj~?qe4|hxH}Ezj-NtX<hWxgnV6HtLX~!cYm>9nK
zp#kw|5Qh_xb^^aW0S!z*cF+!WEv$h;Qd~&7gID6Sn^dVLJpha3FP-2+7afKK-#s#O
zKURJ?We`r8L+HXu0IgBQ8S+2szY{)_Z!X2k|A1USpb3x!U#nkrO*^tV9vkw%dy}{>
zy%M{+purE{GU{pddajlLO2$_v*;6cM>F<!`J2V1nF}Qd~+UAm@V6ivkZ?t%E#iY2H
zA4y~{wV;LMIj<l4h#tQA{)VKQi}WEZ_P!yh7V>ckQZ3=*3Zz=W;j{<ElP|h`uKCx>
z$Z#AxoAy0fkE<W0gi9&m1xmnAM1-`JU$Bco(inaL?J|;H<_cPvPg%864&W=jcM7XE
zKK@EsedXg$%BqvYnE_k-HRk@2g)c9yL_jMsnlsrBVr&P&s|?1VkufM3Op;+^(YRPN
ziNlvsz-2TV+@4{tqp{b~M8J@p;z%ibDu?p`2zdKtFXP<y*CWe%Gz^|rd*7M{7LZ=B
z{ct4W9{=+qoE;auNNKom)BhWB?iUt3di1{K*8d!&@7Em+JMvCiAfnihH1m;PN&iGM
zpZK_$W}5l<GtGSF;})7};p0}CY31WDH1maz+i0eZkK1XcosYlL%vTP>{kpG6{^0<p
z<XqA$m$U(;_CAi9<?-<oQtb(sFZ6DLQ2}Yr@gpeR2+D+eR*9oD;wS?sI`mcoG|uyJ
zJf#uOVMwa+CdxinfQiBFPeCOANC6sf$Ro_E346|CHXySN$euIYV`#uJe*G^UDW@X?
zurpj1A%z!7O}PH&yAN11cC^91f9=yxCjO?=zWj(04xuc?iXp~8s(@E0`=SCC0u^H%
z)N@I3F8@LLC294N{~#^<zyiV<=pOg@pKo07Sp_!y98#RauLH<Fu7Dm&TMmz%7aO_d
ze>Qwj9)llnzkiwmi|Tz1Gb!TR&r4GIB^R1;225(y@L$V5ld7Ld37l}fPvoktqy(PV
z7<L<_vW+s}GPmQD#&OD+v#$h7l0XgQ@C8b8fsY?i%8w|0?(a9#%A09@h?g(+{PKLO
z-qeM7(+UY;A)y6!2MIP-vh0%*h*xi1^1a(G=n((fUf_hqflu~D3a7&!QrJUkz+(l&
zCQ=%S+?uO(&RFkyt2_vAa}qTuiL&Pwca<7+mBUa8#Kh;M_&NXCUG|*|*aMUn@M2}3
zwZLL)2GcEll(nOJhobu?(dv-!_T~?Lt7vtw?po`Viw=^6lUR?JDecRY4a^CmH(sUe
zs~B#TVWcpO^Fv&gBZbYRCVX1}*Ks`be^w<$@{eX<^7X<CJJFC!yYT7$f*AUO@Z%yw
z2jSL14Cf+4EOLuQ!?`2tI&!;?hC^fk9-gY~n;UMbb;z;~jevS+Z;-Uj;}Q|5tYA;R
zhs5{zccFabl#f`@uSou-4yd=Wb836wp>XH#r>|}+VGL8i#Y@7lgD~T84APB3#@y6n
zkzp({<M3sqdl?zSQhHBc!|TWlFjVGn`Ly?YAg=s+j50k&jR)^(KAe(id&$J~U)%O)
zmAZz%w~p?8ROzPv&9ncPwfEL(b>3F~_Z@}%C!0RJ()wRNu5@;u&9tr9*Mcp(3kkcB
z3B<m|zb*{%!BHN*{}6z=_*_t|)M#2ik$~$Yyw3>0WFajVKsvr;@h?lx^`00Wz5kK@
z^MxM*#7;!V%qL9?NmcUmAK~HY>ov^VXSn-Nf%^}NU<CJb828hQ`#GHZ**k0bKX3RU
S5Devh`g1@1xSyWf&;J7<B9D9k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.9E1D81891D4D775B.idx b/.cache/clangd/index/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp.9E1D81891D4D775B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd80163482feef2e98357fb5cef9232639fed717
GIT binary patch
literal 11160
zcmZvi33wFs-N*BvnT?wRxj~R{X>{klm{}WGBs-V5+$@J|xe^G<a_9m|h^W-&RiZ%E
zYTmY9K%|(0l3*J!yz$Cn(rSX%grZ1`M^KJ>U|YeqbjkY+`M%Hd@;uD5AO64ZHy>uR
zvpcghW1{Q+jL*ulmP}l~a_QQ|OjcG_-^|}NYg2bH8Td9oD{KAJ@6J9m_4rMF+HYR-
zT<vZ9U%%_zt@l1qx$?b%bIHIF*S#E>xaX$3+wT7Rd$0ZbaQVJ{4@`OCl_Iux`#;Z?
zw_l$#fm*bGQbO7^cKWbW$1eVH_|(6wxOmTo`_`o{{%u_3nJwGyzGlR8NA_Q6`>$jF
zTy+0Kcdb3})_U%fPfknPOKojaU)#7V-SE-*wfC)w#HMx)e!H}!=F93w%8M>-NQE1l
z4o;dXZ9d=c>!n5K9>|;hr`+349{g;9c-zW3U(8r_;KcgRf3a`Y%E-doC&kug8;LQu
z#&$h=WJ`1S>HB|}xOdRj#&>7G`L|-~A3siRc>k}tlb;_r^hNE+y_-Jt-v7~=PbxOA
z>Yuyx+gtWLxbw4*mvzn>H++U0_<FZK@7lb)9Y4>R@WYYyODDfQXvFw?F9zS+_xB@L
zU$y+_$C<(nEkkEq_+npSzrydI%-xZ-dC2OuvRiU)+Se2I+1n3(x~c8L@K<^T$ns_U
zwZofF{&GcrZM5j*gGqZ|Tkk&Q-ii3~>g*A<3zqKr{OfmGj*oeLi}9=9o5PC^CT#w}
zX?J-S{-nLObN{~2rrh-+v;8Nx{Po)#cI^!Bo^fL4Uq;Q`UD!OAe`#RB@i$u2i{9FG
z_}SGhdhWzA!Qpi)3*O4&!cS2jezRQq{O)s+B|rXh)5#+{6OUhC*cx)m4v^p${j*>H
zwCa6pSNhSdZx@`-I}^He=i1;K(G?v}#Wp;bKKDY6xogbbOBbb|*?eMMS?~WoHRR15
zt=ifn>!1BJYs|CxTRwkh%h}Q6?enc$ADFfG#gd|~-13}1^l2>2e(@jt;YUN;E{(aJ
zdGcqo_Wt}tfSY&plfB=(5}fk+)}zjV_pV9p+PM0BVdBG4gPjlT(ItaV+%u?fLgjPs
z%Dc{;fBIQ<*1z7lo~=8&X#3t!&d@y%1qK(+x^c*_XT9*|Hw!3pQ_+B$N_$4vicf=2
z{=Q)U;l`Votyd5E+rzELzQ6O8@Q%lBy)brb`49Z2b>IH3+<a%ySIi@k)$^Ap&t6n!
z9lmM`SyKM-zQC^+ZJ)GePwfXE3GsVB?sxN^3k7Sn&9Opd<2virr-}wPJs-U;eeC-$
zTA!J-^q0R{pP2X2RWDpO{b<h4pDb;w%Ki7Sf=B=Ur|6K`yXQT;WJ7E0>FM`hxYT)R
z<m1}+d3y_vEm^*<-_396r$1=F{_LLBhYvh^&!>y0JaO`+JAc>u$Y+ONdi<l$X5YCb
zXZx42-_1H+{lTkWE&B8~+==2>LN`9~@x@<EDq8wP>%NuR#QV0~cj-jxnYRvY-?nJp
z`1c>1esJ<N6`%e&yMN*Qoj>{g#Z!TEhhJU!o|L`k{QtP`O?hX~zkAIaIQfqwk}v&v
zQGDASO?8vMAF^@mrH#|SdVRyj72{7gocf|-?)jUa`S8Ix=f0NaO3c3w%{`~SzW%wx
z2Ya1c_T8wb?mzq7`{`i+?}ikf>i5fUT3=h{HPo*uUs>`5$;!$eabT;pJv=gxnl&%0
zq<?n*obFBLXwm&A_u}&>9*h|?^Qd@cc}Vt<zTF#tx$n!Hhpv0+^jmq<w9Im!>^?c&
z8-MwR!iEcPO<eU@9+k)}kIEi3zkA~^-~7Wj=XQ6Mznn+S&MXhk9(+~z#$TQ={OsCW
zHt!I7WzEbi7Yr&G*}d_Xz3Rmuj{d`qTk^AR%PhwNv4!0mfBE*qPhP9vKkG~$H7Bz?
zG<#@X_r_mddpi2jqHXcrdDM?H%VV?0-qF4Bm&*@LjqYtP`XN7SZf1FDVCk~%jlcZc
zH6=Iv_RZB^udErF<za({jq2X`%cmEQd{rECPS2yJW|jwJ56J1>_{;z5GrRud)y-Xb
z)bz}9-|W6Q-5Y;7E48$2%|Ea2d2X!S?&oC7n?F3i=j7Ot`6Z!nsL0>98H$Cb0L4Rd
zdWa3PFv-SPm}KKDOmblkCb<{~lU$sGNj}WOBp>5pl8^H+DTD=>6k-BQ3UL7@#jpsI
zVoZcdF)qTS6qaC8ib*gj#U+@O!!k_DF&QT1xD1m@Sb<3;rof~US71^Nt1zj?RG3ub
zDoko&4JNgi29sJ`gGoKC!=xV5VN#FlFlmGhm^5MrOd4?mCe5%3lV;3>Ni%N3WHcNt
z@?Q-3CDB-P3Q#;cr-!Vt1(R0Hf=MfG!K59wVbYG-Flonam~_GpOgb?KCY`tglWy3B
zNjK)gq#Ji((hGYq>BT&l^x_^&PKr#LaOE&kG9@wvXli6CP+6pG!j)q}$?cKbf$oUh
zafSTj<!!m$$4Pj}R*}$?r00bhswSZw$aN%i2)T)bT7lUr!eXASCM@RJI>KU}Z6Yk@
zxhle8o~tGt=D9k;VV-Lu9On5d!egGVCOqc(I>KX~Zz4SAg(@OoUZ^Gl=7l;UU|wh<
z0_MdkB4S>wCL-p=IwE3TY$77&r79v}UaBS%=A}9!VP0w?66WP9B4b{zCNk#bIwE6U
zZXz<~l`5iOUa2Mu=9N04U|wk=3g*=+qGDdHCMxFDI-+7;Z6Ye>wJM@vUaKY==CwMa
zVP0z@8s_yXqGMjKCOYQzI-+A<Zz4M8jVfYb-l!%9=8ZaHVBTmV2IkEwVq)H`CMM?1
zI$~npY$7J+qg5pOB<VSrMXO142XY;W9zt#+(N<uqiddMps)>bptBzQhx0;BBdAo|(
zn76Bmjd{C{*qFDQh>dxtia3~es)>Vnr;a$7cbbTUdAEwVn0Kp*i+Q(>xR`gFh>Llz
zig=j!s)>hrua0<__nL@@`FJFLOV6QpR%8~?oX8xYxska*t0Jpz={fMOj;sc{D{>c5
zDw4WF{^!z_x!s>f;V_&CEr5e?X{Zb+8CnID3ataGB%z1#aGYSFhb?8HhfT83!=_m1
zVJiuXJzRo=9<G#w9xlm250~PghpQwU_V5WFdiYWvdiW#{J$#CX9=?+B*drta=n+Z<
z=n;|v^av>ddW1?MV2_v(p+_tgp+`)L&?BZq=n*T4h&@t5f*z?<f*vU;L64M@phv1C
z686Xm8G7VW8G7WT3_Ws6h90?+$k?MK6zEY(73fit3iK!`1$vZ9qF|4jP@zXHRiQ^s
zs?ejRROnGFiHbd1LW3TyRD&KZsX>pH(x6AHBpUYU2_1U$QXP8qqz*lLN{1f3lIYlD
zBn;>=N)6~Sk_Pk`DFb?pN@8G-nJ}TpEH$CWOq$SRrcCHDD~X9c(L{6s9P~@0WkAX3
zDxg$!9Z)5SK8%O|gatiTsRcb&(t;i<WkHWsNi6KK6E^hNr8e}~NgI0Xlnp(0C9$!`
zNjT8slseGkBpv8+QV#SumBhgwH{n8$Tk1lOn{=VaO}WtHRuUI`yo3imUa1E?Uebde
zFXcgxS4lkVnHHG_pSq_<rUT80%mA7hnF+K!vi!<t@cfcwBnh-4vI1yjWaSm|kMDGD
z_jQS$w~6_|`3rmAt?milv(SGJ66SxuIBU?zUr_lalSWP&*E3W$vTU3`6y7;LST_B_
zHj-a5^roQ^|5Aj1y`z=nc3<e|`FWu<6*^4$I+;`YPAYU1xr+*&LawJn`)QbOrbB7u
z4m#9{T*HK3U|_z12^~OgVM49Qje*diK+imzrdaf2J1G|Z*e;4iKenD`(T{DWS@dH&
zXcqn08iqwbwt->Mk8NRC^kW+XEc$V2ibFrHlj6{i>!LXH<LYS+{kUeDLqD#A=FpF;
zVL0^T8W;}!xE6*(Kdv#rp&y^7c=Y2tDIWd!E{aD#zMkgMk8h@V^y52d9{u<lhDSfX
zf#K1QZ((@!;~N7!`Uz=DKtG|A63|cRq6G94>S+P}gl1YmKcRyb&`+pg1oRUc7y<o+
z7Dhlnp)nw!pO~gZ^b<QN5&gt2N<=@go)*zhY^Fu@6FX=T{lpqZL_e{C5z$X<VMO#3
z8v`QxNoh(#KdF<F&`;{3B=nQ&X$k$LW?Di&se_i#PpV-g^phGG3H_uNMnXTSF(9Fz
zoTg;-lRGIH{p2o6MnAcpmeEgcre*Y#J7^jG<QhgsKe>UC(NAt+Wb~6812Xz4X-Yvq
zrIS+7PwAo*^i%3-1^tv}T0uXhgI3T_sbLiKQyLfr{gf6)K|iH2prD_crd0G(J1G_Y
z)GkUzKee7#(NArrRrFIkXchg`8b(DwwSiI5Pi<jT^ivxHD*9<@N<%-blhV*n>!LLD
z)9Psr{j_FULqDy9*3eI@VKns98W;`zv=&A~KdmvKp`V_nboA3ZDINXvE=or~y`I+5
zPj9Am^wT?N9sTqgMn^xrfzi=VZ((%w(;EXi`Wb1;KtH3CGSJWHq73vi>S+W0jAq(E
zKcj;-(9fu04D>S^7z6!`7REq7qcLEhpP8mi^fNmt6aCCC%0xf2o;K0XY^F{0GdpM#
z{mdH1L_f2EG11R#VNCQh8v`c#MblLDFa;l<JE`bV<Sr_D3b~$+?x$hCnU1EBJLqU9
zat#xGfr0r3CVBw5g^9KzHwL1I0zLCqnzGQ(>ZC06v$`k?{j7T0LO-jSw$RV&pe^*X
zY8VUstOmwHKdXhY(9dcNSm<Y`DI5LlPRd3<yNj~X&#tF!^s}338~yAK+D1RShOyDl
zZeVQmvs)M&{p`kojebs=a?sD|q#X2fx+n+zoO;?pKc|^?(9h|h9rSZ*7zh2F2F5`@
zr-gCQ&uI)e=;x*>7yaB$%0)l7i*nJ=t*2e|bDL=w{oD@PML)NOana9hU|jTbTNoGp
z+{S>5eqNgL(9i3nJoNLrC=dO-dfG!jubKAH&+DK)^z&*M5B<Cb#zQ}^h4IkOYYccj
ze*PD&Il0|mu=bpm812M323Nhzc48t&+e!2oo+Y;v3)yZbHnP)B9AvkhxX4~R@p`!U
zp|e>RzFYet!<69p7*n!1xH##bcrhh-cE*%c1S^vMX_!BgIo-Pa^|1f{njC(7^&s)l
zpJmR@hTJs7^1rf)Pw98#7mtk#Tt2hQnbg}4=a-b_l=b#2{ZqNYx!vC;`<0n9wptQ;
z8Fm$0OIT#CmT<^?E#Z-cS|T8ewM0agYKep_*Af|7sU-@sT1!-9t(IuWdM(kBjap(L
zo3+G5j@FXs%lK5+5)0X`B{s5COB`gkmbl1XE%A_NM`lNRcKMx=JEQ(CPhI-ep0o3I
z%L6hyIT%vmq%_C^aX}7<5Ar}lPyiBxB9Ii6faIVIqy!ZpHK+n<K@CU`>Oe-&05XFn
zP&61-;f*P10og$t$O$??ZqNntg5DMK+cH-%F5kz{^E-*Jll=el%D6jbzr)Wpr)7TE
zhee0Q{o5JxOA4X|asP1OKLMGS&*d8+{N7C)5B&6vznnandHD>RJ#4<8edudv*pJ#D
z9h&*)K7INO@~^=E=FEfTx1D@?Yvw17&KZ4m&$Bl!XWZ4k`JVNSfoUsuxS8zWs|OGA
zuao%Iy>h$nuJ8pizXxq3gvTMajj+gE8{v@oHo_wdZA3s8+lYuPwGjzfZX+_X(nb_y
zwT-CAS{u=j^){mSaMPF>O9IZ0GnxNw|HA&Fe+SN+@$p|S9C7q*Cch-XBu4kV7#1=M
zfz~nWM*Gi9en~k~euWMnKi;&wdd?e}-h$Z$i~XnZb0zTIRlNs4a{0J6aMFmL-m-yZ
zBm5@+j+M-D?egs^u&awhcF&W>7PCNHF$csK^FTte03;TRKvJ;;Bp1s-O0fc@7OOy7
zu?D0U>p(`a0b~}NK+)o;?f<IfmsrIXkX>v8ImHf;TkHaP#U9Y)$Yi+uxGb{l%0<O9
z!9ClkqJ5X&UIr8nQ2agX-}EyuxBF%v*uxc}`(O`Ous~b|2gFzKKthE8Bvyz(QiTL0
zSI9t0g#x5js6bkU2BcT$Kt_cDWLB6!(TeDOu!k!wAiKf_aw;4kx55STD!eO{TDoHK
z<r|{bq)OM!y!@c_^5%d4Ebt#>Wi9G?IKSuN!k&lSCEZAuyn1TLt-HRm-pwwZzi7e2
z#aX2*m#nxwvlc6xOf!M(tAaUs`Mvt|zItL`$GV|!?!;)n!TkqZGvuZpjU5*mfAa)s
yAYXKy85;5b{p+A>hZT$*HTwF(8(B_}6;(IF!^aF2HGAM7@75)hJEiy*GU-3E1b5p2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl.hpp.336993554588B91D.idx b/.cache/clangd/index/device_batched_gemm_xdl.hpp.336993554588B91D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b4ddce7eb640e15f8df6c21126a6655cc2fcefc2
GIT binary patch
literal 8590
zcmb7IcT`kI7vGt;;IcfHUFl_|ND)}6vNWX#5{gDNHey9Zjfw>^YJz}BL_j3KBE1Vp
zizr1ALD7f-EFc(E)Tr125rTpZzro${z4QI!3FjpD{k(T)?%X~z-_zawfDVJ<y4G`d
zfPau5lfhu{(0@CE0)KiS&0x$?WiWPU)Gt}OwOgq8cV(!Qmc2H(Yg_ZH(gAp)!`~%M
z_DET`D`Vf0!<(;7^!#0vx%qkMKRQMZqTrv}c1pqT9@h>zL}t6W^8=mck%2>zdGo_{
z_gwnnBGXXUaDBADci}^_#qw}ysjWr$lpLuK8h;&33zEKl`*Le*KIe=@g7J?DC#LxO
zKWIAEJH2bwA3KUXnxt-~rKnl!y)pioxpt;C*GSbyb(yU~z?$r!=40MXM~}Lnh~%W6
zlzM#!e&WZkjL?1U?CtBk%e!;wv{vmMjq;pyi~Z(dD`GBl+E?6jh%ECqNYe4vHhRK5
zw(inoX{+;AviqD3qZNneEpXS0cL*vzzaToF80fw|CycS9WaOl)>h%XY7H+q6jF{O5
zZ#sE)*JZ4#<?ku)3YfG2JiNeNK4&bv^Om5`OL~1i$7o42Sv{-6Pv8AFk7xS_`6dq%
zTK+aae7VY_G+bR(W{umPE&FwR4qW5uH_23P(h&)MOPiZ?OfPqOZ&_kc``k|@WqEv!
z;ePKcStY@4udJPm+GjU^Je@HznB23qIWe+~JKr#7pZXa`2iqXoMk(Umax2H`l)jgO
zZnd-pjs<7KCJ6nctl}IK8#`cCsaK|oTF|nFw8vf|y)}C{?9+pC$#>40F^<m4TU*oj
zx0?I;E&Q^b8Ll%bFP&(#p1N=L_Dp`lzVDS1rAznvRM;(XYCLS$cgOKqRppvnqu#Gb
z!eI`l?dcVR!$Q^_9bI2@$5Yo-RA+cftt2b;R<2V%AF;mDKkdrrkYJB{|9;$Hpr1<A
z_{Yzl`Vw66G+odic6ZIvrI+<i)xU7fIK%BZ0u9Qt7ezlc^5_}y`q4;M>s8|HoEtaa
z@Ilu6;^lF6ljF{o7yji?J|k{)mG{>2z|)D@cOEs|YmZR=ZOLj+<JkT~WRhR%(W-|A
ztB!>AwQpRP+&Nm%d#A$B;qmvg8w2j@zxroJeb@4N5pZd1#iizbv27{MCIL}OBPq`}
zw;fyfTDdp;SCvV2%*Eb~R&|=)&ec|@=H_Lu{MZq9HJ`bqH89XiuKr?svgw||>kmB5
zH?j`Q4d3qi+IU-1p<H{f^Kf%}c6x}ypRB@vq<P^lcI``ylM>ysR?gcL{K0nO4zB?s
z*&*ruWtWJ1KYp^!ygV2ky1eVI?(hoR4J@;*&UR8}n!<?1a~_r+c>U<K(cXKax1ajb
zV>g6mg$(7L(m3&V-mW>NPL;4Av~5;>#K@f1BfGg_L60V${Xw*8ZC#z3&`)T(?PAz_
zX7I*!_O-U(n<m=dbNTFBXvkh)w5{CI=l+(8)>n%PZvBw6Gv_XIkLzcvhF1~x{smnj
z)hj;;3#1mF4%zxNV8y@*yPJjnl`B4_@78sj*Bjb0?0IBwuzB=~U$(d1&tL!Wm+H44
zU-=fMpKkca|CChM!|j#*BR%$g16}tI^lq!Tyw2v2i@BF~b~e3fiAh(Q^s7hsg0>wd
zBV8W^LWOrjxtCJL{C7?5j!-+f>uRd$zblIyjviV$D#g+{7i?Q`Hp<K`R!b(})b`-K
z!X$k|%c70%pP0YjlUP=_HnC`b?ckBxu*{3TbE|XD#ikB6U0Gzn*&95oAt|J8t;wLU
zs7utH(QwD{lUrr^yRpK>kFU?k)<^`IHC3*SHa^e)J#=ACLq}e1efm$Y46Yo1Uj6eW
z$NutkwR0b8I@a70UN>lZrq=JgNlUFJ?RDygy|Ijup9WkvB@U-$z3sTS;@;JY@{m8g
zyw|GOumjJ>U0do}w6k?7r>;6Xci}#*`rdez`{pvEZ{)LGY6rSjebiZXQmI?YU*=uq
ziqFzJLu_|X+BC_%)bFul>ZjvQ8X0;IH*~7TCT{o~uR8?x4{W<T(yo@0SiO%QtDe7o
z&04ivXC8M<Sr!-+qM138|F28XH5Dkoy1llc{HS7!bIppYwPS}2_gyh)W`CD%^LCVT
ztG|7gR*bMQ(tCQ+@u?RYW1VKz)-Du1*rVStZ*@46|M0fGSI?}qy&J<eZJu*tb+7H)
zXDzm4(YrNP{j;I!$0nP!f#ZEP_T{;if8XD$E0<Qs7&Y6z*lph7qNdYby;t32>Qf8V
zZf<T*TrFdqo1q++_9APi_Ow~g6+$@!7Fo&*rR#$ZAGxEqds59VrTX_CW=X%ly){;U
zuj>>;s@!%`-nqG9>RmTurF6TqCcd0-z-yh)s;F~Y|J>OzTzIl1rmAG&-i6QqSQO&C
zVAuY9{i9b_Xxa|SA8hgR-MqA--updxxGDAA#!pT?w_9yS)-?axQ`6@#Yem2N0%NJm
z2MT;k{eGJ^!@74)$8bSom%in-*0G$T2=jWcQp1%_a_*yDu4)$p4zC1HLbv5^lDGD(
zYCQ{e5@hBG+(ylPXB-{#(|P@DrOVY@zwFo(u)2A?>Ai5NIAmoSD+_r#48d{0J`R*6
zguy~=^hX4r^jA67_lpu#MIwiR`eC3eAy|aj6m-Jh;J_&^R7N6=P^J-TN(dH_88s>o
zGag#Rh4Pabf_z|=51b_gizoy-gqPRO-pzqlx(vZvp!OE%NC+0u@^Jo@+0oBXf{JLP
zLqPoy(3B7?qFs0Lb^{q9qYS0dDh~kb0pLmq7SYi;lwEw|f|D|oMY<#aRst}P5G*on
zna`)8vyHAwPyuZe1Jq;CMq+|RX6tPbp5ORBL>BUqNH<`21D=Fn5ufnRwd{VE=7~@P
ziS$5j57dwlEV63u$l09%OE_F8hjd8=R;kD?VuD3hU+DGj*rt&q4M}Z=APh_h0~!*-
zkYPgPA|L=8cg5AL-}zRC3)Rt9PoeHpXe}Xloh^SFJjoi6-Nc1lBoa$-V+p>5V3BR=
zFPgM+ceik$J~G~WVE7))l#s6?-H|bCLe&GgP!p{ZN$5lpRuY0m4!vJ?@Pe~LImxg@
z3QuEAGyL*DhKv*egjz`1_&t4@oPYvjE}W#t5EMgBG2}@IUL$4nW8z<OTL@{Wgti$2
zvSVPPgkX_O^<@E3vZ>cOa2ir}1gMRmco7pUQuu>UV_8IoEf;c-b+dp~7P79GV3CsH
zVXI)%hA~+vjdTfvoG_>?Az0-2@crK34mBHbpf%cP6ex@$2Z#w4IqBE(=!VMe1o4T&
zAtxN2NKCLu<-DLx*?BKaCPFoIq9({~LP;VfSfuWq?BuncaDoz?h(w}+dNhhUF~K4i
z?^=Aiu9(~=4P}wY??Co<U?d?}<f>Nw-y5_1jX98w4qXbkrGPIXSfp(sx1w`mv@Qou
zKqq<!xbJ|HgkX_7`IqihIn`-!pbT221aM2xDq?~~9z<22ITJcPej-#sx>P{*3KVl<
zf<*?7q}z8od_2Q}EVN27;1;7*#N?|;mO-noO=zJslt!!E2eS8pf`nj^H%D(by`2#C
zVB(khQUkd)aI!#5um~g2f8CBYMU)JZipF9~>KINR5+x_U8yMldB?VEC8wEA7>qtQy
zoE!(uC@g~VBB(*(WayO)7g0DHdS%mbJ(RD98dzCeHvwBxEwLiFS|WzHRAL6Mk%%EK
zkl5L>udk}*tyY-@NkKdz7f*0;Dw6^cAt$1+ocY1DjY$^WOcI5SRy5GZ5k;cnr4<A8
zaa|)()zmr!^l?EYk&gO@feBVxZoRtAPo<jYAyZI9kVOOs@5mGs6J#-k<*JXJn<BNX
zo*+@xVx<D^c*_|7BT=5QvgkM;u<|J^ui81&yyN~Rf+_gk`Fj^!Gnj(8&U0zV$JysA
z;(1#QW?TN~1IArx(hV%JD{<#Q_IU+3uYfXc0&GeD7?8x6YS#(Y$D&h~=vZY?r3|X#
zO2iaggp)5qJ?tAb`RUEhh37A!t?cdWJ+O8f9~@173G7uy|C~hlCJ{@qS+qF&?4np#
z9wJ5z)QEvoaFk#JBcF=f6?V4P3&R?jpC34(Uzjt@*<zzH7+7;8FYaAfb=<TNL)@`&
zaOj)~+|U;`{QzL1t(9Y-8s1vhsj7C!wvGMG!y1jQioUbG_z6;_9`fs<F%Dt#7NNol
zsA!ECVL(3&2(j`OPqY5eSaoSD+SAU)&K8>oM;Fpj+$*uXwXd~`)pzqhqJ<(LKLQ$J
z9j*I>w>LMx9(SoF(^3gr!FFcy%*yvV^8dHK{iMN`P11gYJf<LpAXBI)WC~IVGL??g
z2(>gq1G|SQNGH_N={SQ>%b?>-LOqkvq~x=RiCKgqg|i8jY(kB~IfP0M9p@4nxdfl$
z=MfrtbevC+`E>n`668@jE+F&^2t(@kg#=kh=NAzu5lt!nF@iis@g0>{-_RC?zCl^&
z&T`kr{vZXOEKeHp*73H(JhxWogMYeAn~p4RXJ|JK|1(Z0CR;MgU>B08CiCwDJsgju
zpdWDhff5c_5*1|Q8fbxIXU<$3mA_`C%tPx}fyq_C02gLF6rmsyPegc!A7p0L2l#Uj
zqeH|JGO>g_KJ9qb(~l=iC@dljMZ{E$7x^XmmgRQaqpivag)%}JGnRZvv3Odyf7*Yj
zF-!AhuBOfWUIVRl9H<=!ZdmPQ)6*65q@Hj2zmQyh%5L*A1HH+p;@$^S?*lu^KE+VA
z80z5)L!x3oxe=P-3PYmlOz|ny#)BIk1d&<9;~&oNRgVs?m;OilKH4GFGSmfIYrORH
zM1(dz>p1oh5sN9DM0g}oc(tWxoq}J(X>_mzLNkGwj0IPROqIDKddx!%3#J8n_Y(i%
zeNZ5XM@~<@pJU?BxQ3uw0M)|aR6ID5f@C-?8G7K9CI#7WS~i7$ntVY~TaHsil7i{%
z>B>0JNWl!n8MN4RWINJ$PKr)6o-^B-#&c10q483IK`OAs)*}V!z%U(HQ8)`2WC2S&
z2$F&vV3-4}D4Y)r@_{98y`-Q37#094jIno+SHv?lHVd{MVu&Yc%)lcw>P16<GX#uq
zg0LlnG!Bq$l0h1+AfBS}3fq~C)ME1wZ72`o30^!g3GX%TIk^NvosKON2q%nx_IA5s
zvP3vuOzQxv4yfbU4TrZJKiC^6F_~xu5-oy%F^(+~2uli!i0@DwQ8<bC4#g41;b-QX
zMBhCXf_AVpvz&=zCB}=0=#TqsYfK7k*|t2&#rAA_8cG0$3BVLLGE$HT3=`>C1Pn!V
zoC;i0fd}1{E?K~X?mCxz;DKF5qPdb)4CL|1Nup_$RRZMkAWNbdm{kho@pw!MdH}Bn
zsAB(+g2#aOn2w(S-V-{03V2WH_!;0mqvPj*_neMj0Nx8aehGLl>9`m0dg-_i@cQWZ
z9bml!@_1G!1<jD(4E69mf)unsehVG9LVhb9UxEB9bbJ-^uhQ`~$iGI%*CGEp9k)S#
z8y&YpemfoCfczVD`~nJJz!^AYNWn`ed`ZW>P}obyeNfm($Nf;)PsgvI@D&{oK;Zx#
zzlOrsbUX-!gLFIug+mm^QAx1B+(xN#mJAh>p%uM+DP~hRwkKom)KFi4l+U5Ap|fz3
zj+a|s5j6nD-QHs>$EfNb&kSYDsW>xm^+!oAzUJb4t>`N6VC2+irAYoX<erAwxNjys
zG(PX#aP%c&#1X1-gdWblq>)w2!{s$^qCOA~O~au*rExf%9S)aJxCE+}&^u2F6qdl5
zSeGNs9z`7oLImhUBT#V!YEd4GA{3(tEnJzfuBZiwFW}VSd!S(tG{vnPU8M-^D8dl`
za2y*&5#|_YWSTq*zOd8gKlEXIxkpM40L}n?q!(ZDaTz`8+RdMN;eGlyPlM|Qa)tN>
z05@3N$WUO2umAYm<AzoZhuV~3%b-aaw8Dlg+g&H+6OiYQT-6C!o#<{OxqFsd+523|
zT$PM;8UvbRz>r!ol2DAKER0_kkl<^;c}+j1_<aGN603le7r!dt6>&8{4Dnk6_EB~3
zFV!m1LnYsw{%eFMQ+{pl^FV`59H9_LO;4xSJ-Fh=Zh83SYR5|j8aSxOuXemFI?e~Y
zd<x@BC(=p$c!6~~``1n1a)*DPd{d|MO^2oaDtdSAKg_d@qpUYmEG=r#+C_wR5ut<g
zX8gu#P)yC%7mxXDk#lE7d^2Ehe4w}xzq8<Dpg#QN%@7NasI#-)0|VUp|9rb{;qk!p
zIVdhdOhY{J&o19+Z*6@!eEvUV!Q~feVVTc`?QEtXjUdygfq(o$L#ESl20>=faV9}#
z(s33+X3=prL1xo&4ngM7aV|mT(s3R^=FxEhK^9OLH&x`YB;b(*7E)cl09Y4*nYc@{
zB`;CfcGssH1kazVdE}cRf!iQ*+aTZ!0tMW+*^-AVoGNXLH>!^QoT-JZ@5*%L;Gn%V
zoH2hxN_6CZu>NgPZcoMS+z_N&JTWz%aHK*?M3{*PcZ~6f@#XX;Nx+smeGahB0W)#!
zVM|`n@agdj8p;&$yBUs^2R&iV@w|#|bci~jSx4QT9@<`a9Vt@<-?Z0b)dIoqPvuQf
zyGVm3Y4mrkH0YcL-Ec`E1$odUkIv77&UqC7iMiD91bLMhWca60<|$O5+C?NG6G<rG
zD*sGSTW-ZK%|n8zP$LyuP&?#8ja&-j3GIuyI7vW*GFLjVP6spb$cLKdmls29F6?4y
zw&Z~jn~VIv$CB|oidh)6!KIx<&pS33p%Zoa>V)#0Pz6`d@lL2c0;k}k4Q14xGpo?i
z`=*s~*G3K$zei$?-X8Egt=Dq$&Hr5&KYS=WXLk3<F!E~^U{wJf+%m@MyB{qO_Af$=
z1i~eO@WQbb=)c*|4FxfNi(%j$+{$DhDGr1#{`#7Zuc8cWCftS?xLNE7T(cu?JRzpw
ztY9#Vs85rxpHR@UXlwq-jCI|z<`!mCjHVi!m>L)|zK%5vhT&IPh6VL$Kz*80pT^Xu
KIrV8qef|$i4Y~RN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp.01B3A3CE0FB07AA4.idx b/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp.01B3A3CE0FB07AA4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5399de0cb8fc7406235beab0b408115456933611
GIT binary patch
literal 4068
zcmZveeOOG}9>-^I8q-u#&7@L>_TFp1Xlha!#nF&D8cC)oQlUvYNv1;R9xrk8cD%;T
z5l=*@yi*jelk*sK2sy-*mz&p{gcD9)&b`~*XE(n;?mW+|S>N^j%=h{2z4qE`>B7Up
zn(alR5Pf)VMoP}~o+6RRf&XUbWTjo}Arb}Kh(x)?=O&rvv<~U>dqZcM@66)8uMcCv
zWMXZ4&K8$@UhkNwCu2M#HgVxyw~ll~9iL&mGQyw5n+69K6bEK+bz0i<!m~3~Q}G3R
zie+6N>Rrb_YU@0bGk@{CuYcQ?*icaF^vm<q&t31HT*zB!_|elPVV7iAMO^cU(Z96c
zz4yK$IPL8pMz4ei53e;=Je@I^9QpH)@m`HjKUNfUwR|17XkN(XyGt`0(|UQur47>W
zd%k(`shOu*O2(dfyy*GX)VfKh-VbjqI()<RwpV)Lwg}hV{Z?n5{?CUUQ&QH~wv4>k
z5{hq_n!f1POX;{0x!b|48^c4|Tkiz6{j0;R#k=sT%y7bP{pOEF@v(1n!(HVUO0TD7
zc<E0@Bhk)~)~x|M<MKXrKJ^<tqkW%~^=;;iwf5D(^vh3fri_g*e5B@9Ig-!L-dr5u
zm|n2F*K$gF&BH#Yp<|>es!z%Ms^gEtJ?5-Vj*{Q9Q%oh6&pTgmdDy!$-TuXIcHiHz
z?v#7tMz4q-&2c9xX1DHcT&y`cSR!9gxFP?IBFFhoGwZ!|K+KJ8wR={-Y;r0u{3y%4
zEdRmU_s*L1(7xm6X8AT9tuXkgpSZ3h9hVcw4kf#2<d<}sVUw@z8(w*?pm<P=jr*dU
z^jmw$O}fz+-)m+S4-BSa3j6PI_6u^Kf57KZ>6eQy9G9<bDnC^C>ijBGdU(cfSDSuX
zX14$>x_k4rZ$tKq$}8D9H*yoF+57KWbt&sv*nEeM(ymvdFO=Vj@!dH$KhLVIPc0=*
z>xf<F+CKcRmJ)}UQIk$h`QROOv&K2|N~7n{i9-|rF(EYh^7aoOH~QRr?U6MhZ0ugo
z`Hz#0490!QJ5yBE;QC#1`3Ju(Cx1=MkBC3!c_7u^weO(vNYDKf_tib|!ggppw8uwr
zY<87lh4v|`eCyqN*-fgnzpqul#NOA%{yKKamZIul$0n4yBlYX-w~<Gi{<yQR>e6BR
z=#;!iGR>shcJ)K8Y*MC7oA@;~In(bVd3fuE&bV!J_C=n^8yd7oR<Y(hQkLf*Uv{8v
z#e{a>)KIH`+Ix#ry`ulVwcOb=Eon+)zw>p0@7FH*{ni#vTetY2?d-v)g0@Gil?TTu
zw5dDxM7(ZCU7Mop_jlara60C;Vf@9Q{tGx`>tD^e#b@iak(=E7K3wXna2oW<R;#W5
zEZ-Sq+7|iTJ$>i9fHhSM<NxeY$^GyuNq(hQ(Zl4QJ7zmQNbY#j5O&hE_Q#gf-&=ig
z{2JVp_on&9y{gty{mMR_=;Zz6rH=#cWc}`M{~{}Wc>T!BixMaQEvs`zMWe31E-UQD
zh_V-E%=nXE-D1V*^&z{}18nf<NumUqSY~Z*c*p#aekys8KQ<uR2G{Yno7m04+z58?
z=eO+#o-I6LgU9f;i`b=?xe@HpV;hU!FL5&1iYD^5x0QFOxe@IAX>MT^zJBv<@JQZv
z5IacCjbK-1^mu)4pjB@hJeIdT#2%x}jbQ(K`hXgxf4iRz9>v>ovE16+2=>=Lla9A!
z{`$-YpTOIWVn=IpBiLCfGqc^Kmg{$tny>q(y3Zw*kB1GgO(3+yIH3Yw>4=FIMQTal
zkvbB1l$HV>rK5m{v<P@ehk!?GY2eX18hDJB0Uo1cfX8ZC;ITRuc$}649;f4gr&6dq
zyXPlCMXG2)33Ki)HP1D;C9%W|%S|BkL?%!MBG018SQ3z=o&+RmAOT6mQh=oN6d)-B
z1xOT&0EzSnkjQ`lNypNFr1dl)X#)*NCYAvtqh|og7#KjZu`D22Jqt+Izygws<p9a)
zIY4p-4v?y0sxaZE+a{>SsG@<IsG24y)zu+RiMT8P;~|O4RTvf)pRd%s=)gSHU_>ou
znGK>IBYuP0gb~e9m*K>I9PrgRaT;nPPFx0>tiec_Pu623%qLA43G>M^oP_ygHBQ2O
z(uk8VpQ^zqm`~MX6wIeg7zOjGGMs|>R5eb)e9DMZFdx-m2<D@D48eS4!Vt_yWjKQQ
zs2WEw9~p53^XVFlhWT_oM#FsCgwZgcF2iY<PgmnK%%_bw4fB~AjDh(~J;uO%#)L62
zpDDu`n9o$>49sVYI0N(98jOYcY(2)peAa}qFrO{MS(wjO<1EZ)jW`SQxf+av`CL85
z!F<kyaWJ1N!#SAGRpT7Y=ZrXK$sbwpO4QYvb06!O;7|A`39Ci;#iZsvNLVf7)SEM%
zmFqV0Ys+n@TZr)Pi-~lkUu+s~#Si;f``8PAPRcFLoObSPD?jYu;NT)`U^D94#x~CF
zS0dpz3JKwESxKaX1d5VSKp_bNik8qoF%kwSR>A_sNw{t$6u`f#zipqwMe$`Q2*OeZ
zfd~#LCE;sPq=Ez_iAX?_N)nKif&wIkC_qw53Xn*F0ErL+BvK+k(h3@oG@=1XD``M7
z3I>o2VgSh~89=fM7LY7r0m&*^KynHWkR0Lw$tgKNs#uj$cnJ8nu#d_Qr~p-<peow(
z7fya$yq<qRJhUF;h1)Fr;!^Vw#xhUv39ZbnwTjBv6UP77rMHW-@OnS*uhUzF?rGzP
zWt2=Q9N6k&M!ZZ?+%@vUem(uVxALbdtIl5beb@5CK5`$0P?vC6l1j~oBufE=h6oX=
zvrQm1Bv6!w0t#smP_%{yiqSAYu^JXAPQwAEa#MK-MYTyPH{d}>zpXtvC07YG3MV&f
zsrkePYE=_KpjI^r6s4wsLTUsQt)_uu)C^FpngxnebKNQ{W%l%Mr^W27<m~8goyp&t
z|9z$i9g!&B(v9ox_DEpv&MFN{oF)?|#>G#bGF3FS+Y=>Py2%;(*|T^#cILPqJ*~vO
zB-S>zc75!7hdYFZAO~sqD964|G8fmOfy0J>HzHW$Eguv_3}F=iZ`8S;yT^cmo+L%H
eoR^Qv*DpZb-``C+c(mO4du<XPHh$z7EaJb_PKjgy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp.6732FB26500DDC13.idx b/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp.6732FB26500DDC13.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eab90d28fcc767e11770b12255f5197576421348
GIT binary patch
literal 4068
zcmZvec~lfv7RI|uX`2RQX|_fLt*TcGHngk)GB$!RHXzb~Eh6C3A|b2-B7!kd%%H}&
zf}<mP5bUVnu5^NXFb5QOF)k<>9Yo_GA_j3#ai;8?lKaQ>IXvon-;a-1x8ALOvFh;f
zhn6By=mK?7R%*e#;Uba9n*Yf!$V<OAOe9j7i9|(}XS5HpJN&GEz22AZxv+B2%R`uo
zN;#ELu*LDd;x!xdWGWH0$xS_Q>u^uZvH8uHCi!r<#4DhzG9bS~_Ra9~Pa6zL_>$^M
z+kyVe^68Jd`VJQqmMk9pYHP~%vKraX&(huxynb?i+0sSJ3CHAJl3jHP4<}9ex%=+@
zH`i6^y?-<-l7IW>)z-Ss`CioIe=MJ^XzhGkS2oZ-IIVbb=;piM<g})bP$Z;}TcCfo
zxukw!eS1}8!{g#-6={vy`ZqzX+Ya53-&SOlZ;g`g9<?Utr~mfvNK9RSs(tc>_Aq=y
zQbuv>bLq4yJE#45H-bXDJMIK@{j0~R-J`t4cF}Q*^_$;rn?19)NG-QJUvn)zOR?ZY
z94*=z+EL-VGhx}gzE1Bc^Skvj)7xx=X~c`M85f`2OpTmf{%C^RY8&e5>6<0KHW_6r
zN32BBtD`LouJ=rS5MxzUXgKy*Jvw`hF2?Sbg=-SIa`D-+ixXe}5WBbJ)$X<(>*}3T
zHY%cqJxn-Wm)WtqwIuk2m&9&K`G(S0t_AjY9&#QPW8!aYJypHt`8`=}`CHqfi*{d{
zdfr)^5#}&GH_!9lk-A072~XtTQZ_5eqX%`4!FElt!4qd))dw9oQ&u^y-ORbTAmi2^
zYE$f#3vYtIs2r<8@#QYn_TGWch5M8TYd&1~-Dc&t_i7K8zc{=4L54c(O3S_PS6D2e
zi|^ii>3KbW)qzX-1viRP=34sfTHTcQG`!Hdr)J>Al=HQB;yrifmM$~tvO0yt={+;o
z$-9F-wpUrlhiK~)`#oZAp0v-o)Jpiz@K5=>CQNs6d;i;w%KI-z=V`(t_Yj4Tb<Hg1
z_F-AWHbam6jIOrdd&`MmQc9y{A0_ssS;`&8)kYJ0XXqQBD6k#Dqa%)mxN-T0MXMq@
z=>xqUBUju+H7=eeqf$m*6Z^z+UvJs=qsrzUowFltFta!M$h|-A=nYMWEaOs_J+ckf
z-nO{xZ(^33ICsWinl8ut0(Ge3d|$%WY<=|cW&VN1c6Dy<SKQodOOLJC*R@L1?U@#4
za?a92oTiBTTScusk*-Q?9d))b;LUem|9WeSTSQ~Ye)CMP`oQgR6WsStbB##bQ62TN
zn;zH{W4X8IMvv^M(}w960$rB4HFtb`SX6oXazykdC-43yOIO*rVlmQp4z7*fnH9M=
ze|>%QfK%9m)9xQIt(|ZG;k8eOUFq?>y2N1Gb9Bk;G_tKb>|FN48hrVhkFJ@Min9AM
zAM|RDg-G7nBzAkAe^n`7YtbFGeC^GgXLtL?Hx9qLfY^DXT>a6+af`*&{H0m*|HF5;
z*!8FNp}QxHF~j4uqGXv^W@;RG$M})HtNN`p(l^cwkL7K-SZ-|`1l!dy!^S82t3fk-
zCT}~7okNU+V5{EucJF(-^spHo&)W`ShY`j>uuJDUh1Yp{7n<SGylo}6k{SoW)}Q<G
z2%U6%x4CErZ@ZhghZzUK{`b5wC*6Izz0L3#-nJFnni>be9<<UPYtQ-RsTr=}ZELZ$
zsc{hOywrvHPST<Kmr9NI{e9#6ssoQFjxkRrG(@~`d*;cch6IYzP(UFK0*cnqKrtEy
zC|1J)#c4R8l$4SY?pUb0v(#8^uqZ7tZ)j13j>rZomna!hq?QCEsUrbN=8}M<v=ksI
z9R)}#mjWcxB0wS?0wl^sfTXoFAZZ;9NII7WB%@^j$><nBGPw*OSuG1lR>uO8&1C_}
zX*oc0Iu4LrE(b{EL}j>8bn|58RAn4cGn8`$CDav{8u$0m1Mw!k=LpdF7C%`EZtIKj
zNfM=8m=>DAQ)=A&Lw`#&V8lsGh~{@(3r1Xqx(y>9K&{7#A8^1MaN<X(n{c8TXtDt#
zVLsV{kuaZZ!$_D<>Twe0lLnlG`D7DL!hEU$qhLPOf>AJ^YQrd)Pw8<A=2Hfog85Vv
zPQiTCfFYQVS}+9jQ5%L}KGNd|<|6}+U_NTX5zMC>FdF95Ef@{+={Agp`LrIVVLolZ
zX_!wp;WW%=8ZZXtGc6bc^O-h`f%%LcXJ9^Kz!{j&G~o=)XB#jU=Cdsr3-j4FjD`8E
z9%o@bYrt8U&o<#K%;y>~4(4+$7zgvYHjIP$oF3<3K4-u=n9nug+))1HvKOL(zJfMv
zcrt9FX%jZ**t@UW$TyMPUmhyF3FD(}nBO-AnefvdrXH3;Z_O#H%$a-UbO%3eWo_js
ze4)&5?22s79aSaan@U1B548*_A%UVK6i`TlfTATdP>h5Dij}ZHaT4yc5*EO}xIcZJ
zfnIkZh@oW=h~R)y5aWjw=|TdML?j?d1qn#Xg#sjnC_qvQ3XsSJ0TLkuNTfi3q+Mu0
z(uf8mt)KzPxG;cZ5Ccd?!2ptVVFAe^7Lcrh1tjOf0g^);AUOpGNF}Fq7d8R^x>YK@
zf$~)b2uk>kNsVU*;eph3l`ft2xN<$eZ=55Xrwi3TP;|;9CaXG}pLQ7OU@yG6&-%nJ
zFbS*f;-_Uu<}MtW8sjHDpW}MBnV<F=?nMays&l+a-v~>G@AzpCI}cZ3S~#xAq{d^4
zuo8ZUs)!I_0lZQZQA3JUk-#I>B=9H|1w2Yk0S~DV@Q@k-k5<vZqt!I<7!?CNM$G_^
zRk6Th)hzHh6$d;{%>hs8q#P|2fqzu4%FmVRk{sB82Ojy<>$b>tyig(G<Yp>0p4dR8
z{D=Th8$S{#%8vpH`5~ZaKN={;j{%DHV}auQxX&stHFMsl(_(&}E<f(m#jH=mpPxkG
zLL{0!beZt^a#(VeU)y%BXO={qk}!K#Vv;E7vnNU!y3}PY$oztrkqf5{8*U;VAu%;G
zx3IDtskROarLCpv5E}=Xt)tvOU}Df;CaHow?8XI>eyr>No3$V1JbKJng5n#|O`#m`
b>Fqng#mC9rYl@xyUnAz=;nOEi#iITPtUF<`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp.5C50F8886E987210.idx b/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp.5C50F8886E987210.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e43f5a3ec057d679362524e9ae18fea1ce94644c
GIT binary patch
literal 5068
zcmZvfdsI}_9>?b#c?<*cev1RmdA|}E8AXr@5Mlt4QJAYBqK=q^APGJxuPeI|sQ7+V
z)J1X(0y8Z}$480Un4_hJsSK%kMbMB;FM(*@?t#0G{QkJJ7JHrV{(bnYIcx3xo4u#&
zl9SK)$YhBNbSsMs%I7)AWHK-DudIA&;rF9tGOepjwzBqwp`+xE#`C9(eT5MVYxn;4
zIikhq94ji{7TC@|rZYQJROwq--QbT09%dezfA(D5cm@shj;^eYF5BU=*5PEA#XK8b
zQD5sj*nd7~+AlqQ2g;XMEPMXe&N&w=Kk;epe(l-d<Ia<Bu6kpQDlqS3`NvH&e~z2d
z+Iy}2kBi#ENB=&{=iR*ZeOuGr`H^`12WxWqwz~sOm4i2)PhGt%aqG3UOWF#Z`I&{>
z0#o<aik5{fH|o+Xw^w)Xc<q#-<&W65-Jf3yy22N2-<ck?XH3<SBY*XOm{U-5>_+_Q
z8%gM<*+r{=yswyA=NH_#^iphM@13jBJ^!lMSQ3z0P<^D`>D6-;7xp);Jid2>;un4Q
zR~r9FJL9wVhdqyMA7t=xZ#{h+q&nFDW0k|}+kbW-`Zq@{czo}WNl5*DjNcR&@3Ncs
zP6SuAB({!f_5Md!jq}m&fx2Q^o*9$d{`+3{!)GEI6!+it_}r^?(yheKv8!e@woQJL
z**n#m<329Me;-oP+@5WIbm{)&u$-WGs$*<t<DYm;axMDJcUP(>yXbu0<um0ET(WHy
zir7v?%R7NZ-?r4RrzTng7hnJB$gVMuHzatwv~Ib6HNve<VPxt)YVMk@ym~Y7haJB;
zx=%hDRXBZX3R8c3b=deHHviL`P4&>P4;t$pb7j|6{eR|_yR;@O{>hNKJpW8@%U;dG
zf(O=<@5cC?=vg$dyE{7JYG?GMMX}?)a1Nd}#TZ?N9eZ!!@|(+5bF+&++nsSF{_D4w
zy?4WV-B$ygb;EMyUnQ61>`P5KLPbEsqP+EoI}5#P*^mS08|ouE)1Sq4rkN`$XK%mk
z-S}SKlj{+wYZco3m-qFSC12eA+4BpyaPV4Ko1t0$YkQ*g+v?A@eOH*>7#^K=lFY2!
z+Ppc{R$CmhXS&6u==l1?x#i~?UQKU(ch|Zvk&RilTX<LRCI|Dmj;>kS+~Yoc|M6vw
z`w#Y<&I$^h=(yQ$+3CoP)`a){XJ-G^e!b&DcvTS*K6$4px3P^6t*KEawWO}!v1cH@
z{go@rPBrxJ{<8k<Th0ON)H|8esoAbuz6#xce{}xs+Na-s^ZI?AA94FcmXqfFoR2Ov
zez^0>J`dBEQ}yJ)@@;>nG*q^&xHnKWbK}`h=Z)PIv*XSie=mv6xLH7anzAvd;!fqJ
zD!=DWgz4_{>Wq(y(>^V$X~`H2PU<)={D~O+qWTY1caFO7Fyi|aW|xNtS3G_Vv-c)_
zQ}XjC=$fkkgf5O-S<<(-<B|SQg8Zp>PH)7?-)o(!NB5?$slL3V`&ysz9hVQzO_{WZ
z+4iSvV9n?hS=p-M`A@{%?G$>XCUMV%FjsVjL6+y^<l{0th?U`w;%VK@b!kyET+!)b
zImju<Yj}{#(GTVfR`)(U;fnrEEQdISBn%HyS^Ml!Z$sCr1FmS6SPpOsa2_6{^18Xf
z$xRXJ<*sOkSoU=CR16PN`EbYn-0b}qOm4Dlu^izTku*F=<zMH89Tmp+s$J1cvFz*Q
z>oPn@<>#J;LpPS3>2gK&V%f{d%Vl_w%1a9tmZ9z=?@ud+-+SY!XLZfDW5V3>R9aPn
z^m<~_snS8xtFlHCro{k>=`cWIdJK@b76&A*!vTrwaX=DU0+57`03@L&07+^|K$1EV
zkffdjB&DSQN$Ds+QhExIw3Y@Wt)l@+>uEqTS_Y7ejsYa2X8_4+SwONn7Lcr-1th2C
z0Lke%KyrExki3=$B(LKE$?JJQ3R(e>f=&RWpceqC4pxtSvGVfNp=uRKxSEoZv@^to
zGqf$h!ZE1kjVun8QB?v`sj6ZmVFnD4m=OacR*C@<H{gK8jW{6jQXG(kfdC|7Bmha2
z5`ZKPBp^v62}rV(1SDmk07)4sKvJa?AZY^)NZLpPk}jnI$ru<wGDZfFOeq6M*1!Ui
zHL`$YOIbj21`d#%kpm=G$^nu$@POovJRtc}9*}}T0Hk0P04bCTfK<n*lcmMsmZwfp
z&j2Y~Jy%NMHjU35)HezdA-Ot8nwFLYsu*seBkO4bHX&bvq=u?f3^&pdNM%7(M-j;(
z_A48rIuElQQFXv<LRI@vz?)IkVVJF`>MY2Z1;L;nvmqGtV|E0Ce$0eo(2to>4EixE
zia|eaL2&5DZ3qtixE;ZvA2*>m^y6j}hko3O;?Pf65CZxM8$v)oVMhq)Crl^-{e&4M
zpr5d!1oV>@goJ+5hLF%t+7S}^NfSy!KWRou=qIfx3H_7>p`f3#Ar$mec7%d{%7jwT
zPnl5)`Y9_)K|gIlXy~VH2o3$T9igG0HlZ~1(`J;0e%gxC(9c*92KpHr!azS`M;Pd5
zOeh2Wj2UI1pRuA0^s^R(g?`qCu+Y!i5f=Jc6UstAYere<XRRm;{hS5ipr5lL9Q1Q`
zgoA$0gmTc&nNbe<IV;LRKW{;J=;v(+5B<Cy;h~>5p*-~SW|W71-iq?jFIW%(`UM*z
zK)+x|1n3t`r~v(f85N*ku%g0<KfdyTY_PA~jyUAOZZ&nn)+PSJhRtH93z`^|DE(q)
zWq4Efw!}J$(~&Nb9#U`1Us=0k?up}f#A#12&p_#LIR8{nT3hLuIybSY$yFgEEen%l
zAmMTxBtlMrM9N8!C^-cZEvG?Z<P1oxoCS%Kb0G0@9wb37yhzd)7Y_zQM_SOMl`7T9
z*H;lL1|(h;K9Vpc21ty+0EzJ!AaNxQNSwd{iSsxh2_*qYf*=4%@B|=9B?(B9AOT78
zBp@jz1xSjZ07>x_AZaBHNSdGlN%J%y86^WqhF}26@C+bXB@0NFU;)YUEFd{02S|?K
z0Lk$jAbBMZNS@#U$@4rQ1*HH;fe-*G@B$#!j9QQux%ituPOSzhN*yhwrk-`HX5Fr>
z5jQ|cYREKcIW(_4=9pPrpDa!XxCi)4M*!)+5XJDVNLprJkDtE4F{!>sq<sh<LAu_Z
z%8I+6AA0SqIIVHeyg1U{4R`Ew_6S%nPDlDhhDuK9Y9+dd?p2bzN8Y>%o%F@fe)c7N
z!zD_(X^mD4->@WF+^-r{wDcBkd6)(R3D@8t5gGy{QbU47X(*6r4Gj{bVL)OvEJ&P&
z1Buu0APE}bMOs?0c;3*huxzQZY{t;TtfAq5pB(8yCd(apocZE$RGw$dmA>s;^1pMM
zGc$Kq&TQH27d6?Okw;_kg2ju(n6_~0C<jL;XSs{3+h|V@cb!*KBI%{jC3pw;_yz_|
zjE;$&9H)KdWj}5brlCXs->m<bkg;Jesc><#vHZC32zAs1<@jJBa*Ch-tEu^D^0fFA
GB>lhBe*caD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp.991CBBA160B6AE8A.idx b/.cache/clangd/index/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp.991CBBA160B6AE8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e003ae8b04f2338a47a17dfb28bf0e3e0b7c9220
GIT binary patch
literal 5318
zcmZvf2~<;89*6Uuf?<;-EP@d9?aO^nkq|(wK~WmOg#e~1f=fh7QHekmm2Ogp+C{6P
zV=H=8Ol!rewVF1ybu9Kojbl-16{ECuaH)exwJyb_Gci3+{&UW74wv`4zmJdHlbg$X
zxtSRmhr?B>^f{T;^NXuy^;M}<p`P!`s*2K6eN-w#fJ#-}@QwN6f)0Jik1cmg6XrJT
zxU~l}P(_Ezs@4s-fbMWPSB%8?H9}_3*}dI4`)40JF-phdZNpOQ8d59QYhLer<eI}i
z9bde?A-w1Q$>>SHbl%-twP@MPPhQzj)Kb@|`SyCrqn<lgj?^q!_$o1=aEp3N)09i2
z#(vv%{=&T$L+S0`kD<b!em>RObanPHYV?P%7NFLv51Q(F+Mi5V`f~c(^RJh;mi9wa
zO2s+0>uZ-CoO`f+YqsO^((CI>n#~9Ajcnbx=XCTBsH}d&_~>l|SC)VIpZlAp6|Xwf
zKKf|;IDGZ=vZZHl_)pjx5%WpK>5=JO9p_Rz|J5DSF4w!l7k(bJYVCuK1uxvL&Ww&Y
z()fMpd^G1k9;4cv-myMu^OTy0cdsUno!w>A`2D~+{IYHiE<1kZOmTKW{V&PFTVd3-
zug@$?3M;Ey-fubWe=07xs-=7M#hj3>i|qR^XT~j9X~~H=8x%jCT>kRmy5l46e3bj~
zvfsA3H@$l>w&;B{zR#s8pEu3x*w(r%?Z7Z~#NztZwZFwzMV`CF%j*Z_pWbk2`^p<_
znqBn|!mE!*yyTZ~?(MR1QIo6{32pnD77j_i61{>7TTXuVsbxS~gflm7#MD!^k-NXC
zYY^K5VwYBxo!vpL$sK$2UfPQdgAH_ky>@$K;_%o-JBNJQ`1t6rVar#v?fSI-=Ha(4
zmSxWWr>pI~H-Z*3OV6LVmC#c8=I#@fRi~?qW(MoFyydL8ma!<byRqly*dx2n<tJ>m
z*4FrRh8&`OOS@lqH@a)&pY2;i^T(JEPP;GXocSuU{6s5}GC8H_@1}8<;~(6A@cxhs
zx8f>H8QD9CMVBqdI81n4<Jf3#iGI$q>we<81K$<ZjxYF(*jW-B9VPCXNPIlm)_etF
zo6_R4_Ky+xO8de$v#v6`Z_EAPI72sT6MP01_5a>im&?DlZsQk*ur{WAQ^}Kgw<qpv
z`~95F?%WfcS6uT;c$)c#ppz*+0maj1PJU8iDNj5~?OA{1?vxD+Y!g4PNg2L0{6|3w
ze5I*&{~J3y-!ye4l#KKFN3iT$g7W^hepe(>I%8Vvz{Aa{_uhH!$Fu8%tmb8(1kM|F
zaQFv#$<iki;<HLNZ6ANDi|JXD6Z~=a>2A$uF{>vX9j;w09P9Y=Qgy@EC$lE5iAlWg
zjEdKY4+FEZPCkm*JpAH@iPvMxHs48l+rFgWk3PGFmu}98IMHw8&z29n=V^YjbYE%7
zIB@ZuSKGgQ-sf@H6GL0gZ<l_(VDD(0vm)d!bHHtR{c>th_(1mukHZ^({(kR`qiT(A
z-O8#lc}!2#fXzWhRppZTv;XjXx_#rnT$R2pc~AhJXI2$zd^LXF!P9uZ{2y-psWv+)
zF96T=)X~1tq258M2m3F+R5rDbDNr@pQxEeQHqJXJHL)f4;p)=D9Rc_Yo;ubyc8qsW
zYQv-3T|2KW*&Bf8d+I3PsD9o-scUD(WHcotE(*XWdg>705P$EW)c>9}=qpLrl^B5M
zc<ON9a6j*$)K5ap``gREyB2_(JawpVsGoOG>WbpIm5=-Pu0PJ-yYBazA8p-zdBmW=
zLc&0dQC`0%nTY(}L>frYktPy!lz{>rWuibw8)(qcCK_~%fdL(3VnD|lSkSR17Id6}
z1084LK*t++(D5c7bb>(uonR6`CmKZ1i6#+r$bdkHObB$6K?0p*l0YXLWYEbb8FacB
zUEH&+RH&nLtdf-dC;EH$9e8`p#H`*ogs>0`K(Z3cdJ}0T0ZCd&K$2Dxkd&DMBxRuh
zNm(gC(q<Zvw1oyFZKVOpm>EDa76y=vl>sDcW&z1sSU|E?7Lc5o10-kR0LfW7K=NiD
zki3NlByZ&bDVPO73Kjv7f>i*dXchq}T0}sKRuPcMi~xx&2$0B%04bRzKuQ(~kdjpb
zq->S}DO+Sf%2pYWx)HhzWg7(+>WsQPkS6P9DoOc?`FqbUWs4~r2>17%S;}-^VG5Zt
zL;0PdF1@A*4^P58e?YB^R;HDMZIZwDK<oWo!hsQAVT#&wthg}ZB+PD%xCpZiCqBXf
zZ^wx*V0PlfF_1|IMnXU7!bs>R-53e|qzxybpS0s7^pj4Ugnr6_QP5AhFbeu9H%37}
zWy2}xr|dWd{ge}@pr3YNH1yLhjD~*NjnU9g+i)8CX**6sKkdY6=w}=l1O1E(W1ydL
zV+`~&Hk^Te#*Q=4&p2@g`dJ6YLO<)mSm<Zn7z_QZ4QHXBwc{-Gvre3ae$Ih$(9gLr
z4*EGa#z8-4!#U{Z>^KMgoD=7upLbw9^z$x^hko9T@zBrPa31=3JI+Ht@5FiN7aW)X
z{elY<pkHuf0`v<uT!4PTjtkH)IB@~`MF%EAzv#k5=oj6X2>qfB7olIY<0ABnPF#e3
z<iHU0BNv9CAGt9E{m6zR=tp)OK|gZh2>K-lCPBaC!X)UI+?WLYk`0%jU$Wy8^h-`$
zf_~Y7$<QymFd6z~Hzq^BY{O;fm+iO={jw96d;O#9ZmN3jR=Kgh|8q!ANM2hWDZTr?
z=a7s}iB4DkrRPr!V}D&U(#JEM=$9C*oSieO8_H*X^L2-3IwUk?fN~Q!ySX#F)jDu%
zkmt};6S2KVBB>@pqSO>fw3-HqQ8OU1Y8E6;&4I+Ld5{FP0FtN{K|*Q-lBAYElGXCF
zq%6a8AM^A^7W_xh5=8G}2%6wQLd4MCL~2Puk~9fO5|MzUv=ksIngS$+C_vI$8jv(i
z1CmBGAQ>$KNQPzr$sh)htd<2NOS6Dv5erC8%K?(3IY4rV10=8I0m;)mAbG?CQqT&3
z6lejE0ulfzYDGYbv<OHMiGV~}1V}_9Kq7<yDQP7@O0)z>2}yvIwK5=OS_Y(yWI*b8
zouqs?JhvG`bcrA(=~9){)LFY^>g9%2o)1oJR_r8Y<Ls_J<dZXhdxmE^s()0Za%*x}
zmpjL2-1bh-w1(D5%41-2{-_%>;?E!ROegnECKUJ8p+0x}1xLN(nNEyIj8~?WN3dXj
z@1qyk)kY#)c>#fiq>%)PGEyMXMj9l>$biHeS&%p*2NG}OK@yAtNTN{$2^kScl2HOl
zHp(FB;&snf+^OEZ8Xvyz>7f-)hYwX&sytQp_xC<cf)(qDRIp+_2@<8JK%(_DNQ|BV
ziPf_pae59UUeALh=mn5Oy$BN0BakG$1d^<mpQVc8d9$8A4_8)LD)XLROno~1^)*eo
zP^k)fFQ+`a>{Ce1eYALmdta$<(UgLz)26GYKkHOQy_c5xbLPG1iP>`}^y%y4+fVHm
z5Ev8^+&?pPTsjl#pE)KhN)tXHIwf_)$md2G(&PwnIH~92|KDunz}UD!g9*xWq6lcn
c(1gUKWUVen8a6f}^7*V8c*dmBMr{0l0p8x=761SM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp.909733D52FF49021.idx b/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp.909733D52FF49021.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6c8c17ba4e2938975c8324daeb347a9c6596a12a
GIT binary patch
literal 9704
zcmZvhd3Y0b;)gr)g%a8-<tm3@fMm{^WF|8cbds6La5Y(~5{@Wn%Mnlzk#H%9$BGB*
zg^I3<A_8j=5EKMrqQ7N56F02jfp&?4%Hh_cD=Vyupl;~%#D4zx^?4Zb=JWE>BblVh
z*yyNHUoaIF!>30V&zZ4sYNv{diteR93l`3w^$$`};j6BwSiJhL@lS6!5~%%n->F&h
zHLIWd_luO}Zhmd{!bkfaxafQS{iBf~jgN@Y(;vR{-IymXp7(OMMfX*t8}|;rcJI~4
zIt`+>P2Avl;`L{Ec1ilo{9MnU^k3?qIoRBE`?IN*lV7$*PMm({w%s9Jn%unF*>%(t
z4a=So>t5*=OP*PD<1Z&KpLF)DJo<?rZ#sGEy6lrR(T6vDUbkoA>0jEW-geXNk2W8C
z_^9WLPZt*H>i9o<O#Cu?w0X{zGhS&Odqd{X(qpT(&6r_Kp4HgWT60}`^PeAY?6+?A
zz1(NkFB{_Z`Gdc%{qo;K?r3;XGrnGUFZ+JggX8<n+&%2uBcEJy{P1^`2hCoEqw{uD
zuUPrxlM}D_esQ#4@7Mpbcjlanw#<CKu|i*-tDik-`%kC7-gE7r-hOIyLywJP=J$N8
zb<pg0kM7v>c$$A>?u6RTk$)`S8q|8!FJIPW8B_J%z-|lo-L!VA;23ksRmBs}oqyBY
z^2Q!t-_z~Ix;OOChaVldWZdQ>#!q8zWD}m{SGX69BSSy=X7p7{Xa4uX+O0Li{yy;7
zkwIs>JahMXH{CayarEAK%==P8*N3IPUqg3WHOw<OH)-D1*2CQ|^HyH)L+=~s4__0D
z3>mzbO;4yH8$Y|UYG(Zlna7sz`tjj;sik)ef9QtcPd(-yd*95=g^pdy(dX}4`Qh;W
z(?%_StM0#dFM90uzD>in)n2vhk(CFpn76ck)2@jZyrykCws*vLlZwVyOYYgWc}LB_
z{cHa9PqK?!zkPM~kvVUlnYsJo?%TGg6B_G38#4JV_U#+)wcF0#C;#P{{NHA)%l`20
z=24^nMfcu1`@}sV&$4mZBNMt$*niLJS+{jNz4E<7gZqr!IYOM)$4o!aW9^&!9>0Iq
z$<-ey8#X_^>N)0z{#$nTp1thmAGRHycIn5_qdUb3Pp)RB^xQbijz^dodp}t*X47YP
zeSGMh!P2YK_q*6@-`exWfj^ECkIn6I@`nccwWB+ie6p|ZyH58TbDOTKzy6FgA=l+H
zYWk81J9<AqWnJTfi}w6>{Wq%8cz9IT#%K3@_{{!UfdR&?Cl^I0Kl;@%`=jvq32$yZ
z@#K#6cYWTa&z(wo>^;3V_ZYJ3&<SntrpwPhUAJcEm=m9moI2&vO&>>ZvG#6y@bUf6
zcDsAWqGtDxH$HtfvT?~r<IekO#FqS#BU4u7x9<K=#u|Dd_x#iE@{vUDvRhi@?Hl&3
zT|52%CV#N~`Ac{0T(dprn!Il2&%5rhUf;aF`PT`z-tx!O=YMp|+~X%MT-xxPZ~ejM
zQ^%WcjlcTR2fy9<Uc-v})?BB^RgYKObIliw%D!tps@cA7)f<;nKYx4nDf0ozR-Ww3
zsaLMNouU(6A`6zxnf6oZzFq0etQfv-=%8wPT)d*GXJyZ>?T6CR{!#Vw+Cz7a*2Y!S
zV@q?t%6{G359PW0OZWC)^kUmv)%2LsTw7V&wf#_@gAbiF)-BGdH5C&|^Qg|FhP5Bc
zvvuY_-){VB$xGGr_|n{`vQL-xLwSyUalrNZ-YFxh=_^X}fXV?@`=LCyO&#>ANBvf*
zrY|qey()WkZ9kOfUuxqoADsL4SJiZ^G<UD;-nIQup6AcFW`T21$Jeh^wSV0YzIkTt
zmM?>YYMNY8S7Z4zC{fHczJuIRH*D^h8#X4&z{bQF*w`ow8yjO`<DwjFT#SQ_kMgkb
zF&;J{D!?Yh1lYu=2%8uaVe>}4uz6!%*nCkRY`&NeHYqB>CdDM!<fsgr9Ft-5NByw*
zV}96_r~;c3Q(#l0Dr{;@g-wfUuxT+3Ha)7trpI*Hg3%ys!B`NsP&5QvC>Davh#IgN
zF#|R;YQkp5OxUcb1)CMKU<*gXu!Un`*zBkcn;o-Zi#Q`ge!qV;McfgtOy&D*V^#Zo
z6>idT*VK+1u`A)a0VwHO+ClEP8%%e?4W>Kk29t?1U@{2?OeV>I$;MeQ*#rwFn`FV{
z;vAS<f&-IFa$xdt9!x&LgUKg(Fon1PrjQW86p{j%Vq64MOo(8LNfAtM+zX~R;RVy1
z^n&S&`@r-id|>*LJ}{-Y1g4abz?6~_m~vbOQ%=ZW%1Ieaf7}nIKj8<{pY(&N#1$}=
zgaW3LRKQf@Dwt|Q1yfC`U}|v<Of8{-sU<Zq^|%hEp3uS6lRB8eco58BA_!(M83Z#F
z4}lp<guo0XLtq+l156`ffN3NRFwM9LrkOCoG?ONnR@?&9N?2f8Nej$yJPc+y5e74y
z41;OMZ7}VG4W^y6!HirS8D2gq)HFp#M#cb*jWm_1{3u>JVVrvihZBa=HKgOD;9^|@
zkjEu;klX17)6KfUbbH)jGEN3ehGoEHJPequlLeDySuj}-3nu5}z~opCOwPlB$vb&4
zd6oy0_wZl}P614T6~Gic0+^yx1XE;1Fh!3Drq}5O)606n^m@Et`kX#6eXI{mpT`HL
z<dndaSP4wYBY`P9WiVw{22=LPVEUbYF#W6_Ouxqurs7n<R9FQ}#iM|!I#n=LRs~b_
zs9<VN4NQ&Iz|=e%n7UI3Q)hKBb&n2a&=~|X$Ogd-dV*kvoFOnnYzWMdCj_S9G{7`i
z15CqXfN45SFiqA3)AX2NT22d0i?zVCJQkQ?XBf;d8wNA%34>`nZ7^-t2GjP~U`B+<
zMdg!1>4f2p$UsUYP^JscnDsZ$KVIjkY0_JX-bOl31i@AkL=Lr*5VFxq3}myFn8;Qu
zv5>>9B#dmg61#(?LZD|e-AkiNpRFU(ul)Iy|E8+if8Qvd(8|AWJW{&;+;1D@>q}4X
zvy|&qs^hD06)4vp<Razz6gfq^HqhXwY1b>rIog#6c4sL!_T2@_jeU2Ka%10}qTSec
zr)f9#-8tHgeI`pW*k=k9gMFq*G1zBPG=qI6O*7bMax{Z|HcPSCXA2aIeYQxk*k@BT
zi+wgtv)E^IG>d&MOL5re3KWNZu1Im%=TbC>eJ)LN*ynOIhkZUv@!01J6pwwrNb%U`
zQ#6l#K27u3=W{fVeIZK;*cS?vfPJAz3D_4>w19mfO$*o;a<qVbF-wWq7Yme#eX&T1
z*cVf@h<!0li`W-)w1|CgmhxiXTcEty_ZBHH_Pr_Ei+yjJ_F~_gqrKSoWho!_eFe&g
zeP5CCVc(acec1P<X&?4|IogMPDN9M%mkN}GeW^%E*q2hYgncPZOW2okw1j;*OUc-m
z3zUp~xk$;_ms7NieK}3b*q3v(jD3HW@?+m$p#0eP7b!pX{VCdyeSezvW8a^n{n%Hs
zl!ASwKq=T)ij;zVB}FUPSJJeCeI-XL*jKZZihZ>}sn}PGl!|>dMXT6X)3l0xHAk!1
z*Rqs`eXT%g*w>1bhJ7tXYuMM)w1#~xM{C&Evy_f~y+G;M*Nc>neLY3%*w@pvj(t5x
z>(~!wsUY@)1uBUBV37)9KbWF}*bk=ZAohbfI*9#HmI`4%RG>oG4;85p_CqN;g#A#O
z4q-o(qeIv?vXp^+qd*zhH;R;jeIrF1*f-L&fqf%K8`w9ql!<+_K$+M#i<F6dGew)&
zH`BC<eKSX!*tfEjg?+0)S=hIVl!bjOMO)an(zJzrD@R+{4`-<`_QM4#jQwzt3S&Q<
zqQlq^r|B^E!#O&PeLG9p*tZLmjeWaF+1R&Jw2gf`P21SFbF|&jx0ioYar)H4BGsv>
z(baf$`PtL(u|G3^e&^(Gr6*TII}H`eubf^wbmiPS&jSyauI_hnzv1Oqo61jy=&JTt
zX*-@<VRHOsJ-F6KLY+(hwxMf7xAOlq<?x@bt^9n@p}wVOVlBk=6+EAGw-7fn(?S?z
zwuP|BTnpil`4++>3oS%I7F&pj>^(rdt?>W*4iF!*bbv_6@&O_v`wtL5vT}eZ$m#*2
zB5MbThOD;`9Y3F73kf2JT1W`lXdwo&*+NWYtA$v|;T94`wp)nZ!PETF*LqIfovAJT
zAo9eu7rqhhJaHp4dBPyGdBP%ddBP#{dBP(Lc_JW-c_Jd~d7>i+^CXBI%99YXktYVS
znI|T)l_wT*I8VaJcAnTBJnhZnqd!Vsu=d<7(%~9VzH!$yxgBmG#=!uw4i<=Wa6r6+
z2NE0tkmwMBybdpr&*1}-91@W1kb(RTKak>3fK-PHq&YMo-Jt^o9YLUwBLrkP3?S2C
z0$C0VDC`IW*$(@6n%Ajf)h6l8`{%x~8RB{eZ($kYMrJaEL1r_AMdmVuL*_GtM;0<f
zKo&DZMD}Kg7ulB~K4d9FBxE^5WMqGa_>q+iQIORPQIWL_(UA2F(UF505=0JVNC?@;
z5ChrF5EI$T5DPh+Az@@YL+lP-eq!W+OZGj`zjP|Dt*xyqzXtf=k=2)bZl5^g+!>&i
zxH^8-F5UE7i5r<|B@8m#N?2sBm2k*>E8&raRw5vatwcojwh`|)aQpDJ5g)SDMkHjp
zjmXIUHsVKC+K7Uzwh<LsYa?0*zi|A{B@@3~y`uE}8PG7`_qSoTEPk!?m^tf5m9Fk{
zexJJXOA0S8NbZ<hHDg5S>iSOgeEG3*w>oxu=aK7=m#*%~^z@Y9Y<+XQ`SoOHOTKjV
z=uV@5e;enR)cI7GZhh`AUER>T!C8KEEWgp&zpDKW&W=Y`m*pA;XC})H#8?avYq3C_
z#R2ga4<uLukZ6fOUds#QvwT33B>~Bn4CJ@`K#HXRsg??)SsIXT=|Dj%2o$nHK!#-i
znU)D;Sr$;(3Io}e4HW4g8Cd?NA9rkdh}O5AdzRCS=_Qqa2IY5^`&G5Sn+!jTfa?<Y
zSp?ibOn?Dm11u01;DGo54<rNxATb~Uc>`V`U%&?>1tcIjAOraWejp{F0I2~LNDF8{
zdO!yX27*ALKnTbP7(iyg1hN7aP&g0<vIF+-^y^<gJv=_}ZM^gXqc=~y@8ek@PkhKy
zo=C`Yp2*1lJn<tdd7>bzd7>g~d7^c2#rzr9Pd#^UUNAqgVBEP26VDymKS<Ndu4IM(
z<|(&THeY$w#7S3IOe#+m6<2j!yt3os)g2d`J1!>XOuv3+2}fUZDe2U?vWuf@b<KIT
z-Oi78A32=uUKJhIqfgIXefwQ}NibxXR@mFSUUvm}=l^$Y-30>%4jMef&2WNvkx!Bp
ZZRmw+e~*7e@4ALROr}R&W=E*T{{bdReZ&9&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp.EF1B1B5023E691DD.idx b/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp.EF1B1B5023E691DD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ff38a02f7736442c10eac6526e71a26679961c64
GIT binary patch
literal 9706
zcmZvhdw3IdzK1*WgA&>*<yI~N0g}1aBr}<rpp(o@281R{Rl+3-N)ZAIA`)%_i^_@@
z)*Gwnswg0`hMRyO5EI?ycuhRv2wu=GQSivkT6ATFRT0z^`aH3pKTe;AA@6)&-cFP0
zFT-SXbi|0GOhv`8snJEVrp=$+p`xOqYw6Ft`EzHyPbw;W)fE+sR{taZ+3m*ywVxa~
zJ43#C^^5=7PFe1z*JsXuqW7W8zvn+V9vR&DxEMY6(aYbBdishvuXI`PP(`|N|Dap;
z-*}?KKx)VM4W6gpSi`f6)92^ry8opAQorVK)5N9Er<%`v*%CQ@?%lihhIDCC(`sku
z5l=TPds?h}wM#5{e!(5ToVj+wg$we?r+-{{=IpK6XKJF4ZTO;Y-~4mGv`)Tz;nF9X
z4nKC>^S__XFVfZVe|8)HW%hW}tm~(}+A{j~%#kH0R_&NJ&6qT!vALz@*7TOYKGoRg
zg_#d>pIg6dh}Y*2|GMtWe+|B;VY_C0J^w-WgQ`cz_L;tS=(oo{z4Fx2?<x<QJqpL?
z?5bX|^2cY!U-$i@XrG>M{B8gAS(k5}zOk`FU!JR<Ibr8dXTRQe%b(wQc4R}hO{3;^
zf3juZ%=eD(+V@nNe{=S@+K!R`S+p&vb*o>#tkW{4>izy*<{wzNZkym3b>$7k(=T4S
z@Ev(mx33@Qvc2w2{flA8`!61|<(To)s5{t%XZdyRd1J_sPrn&?!;<O$8Bn{eX6S$S
zKXGi}g-&bkzhvP<BN<1}-6y;+Cv<&i>ie~HmsLYOgK`t*Y->5%^%`&GWk2-1WA3oE
zvB=;-i`ewI8nWs0>#L^Mzm$1$`JNvio0D2{&#=GTKJ3{i-J>6xo*CfSqa5FO-^!1M
z9h@>^`P+5>xqrcvOM6cox})}nJ&&(EeBGQS^_%yMzwC8w$BF${eK(<KoLv0CjxD=t
z`X5~T)j!D|ZvD>H*~e$SbAI~XE4uF3s*Y={|9tSIx7l}Yf6#8d@R0nsHTnOVsV@7&
zds{|~{EF_mZRY6*LY`$~vd6}C9e417)idtya&G1OM+Ws8zWXY1N-s10NVj!w9eC>D
z8E001sBGBs+^QFuANp?H-E-!$KmM@e=#;BJi5}lAj(cV`b5r+CL+yBknYRDa6{9wP
ze%~iY-W?>pHua#3z2)tFZyx&72=T=1ZfAaIpkF_}d-10S>b~pnurYh$t@XE^m&WBf
zT|-S>JZ@LdjW@l}IPda(zuoqYsx%%Q(Yf*YeIKnkI3v){xa-V<RVQy=bMl1!ad_;w
zw>F)AX4m@rzUb8JUL`&Hfu38s4PJHRw6=fqwHKbNTf2MI=`V&)zUhh0pG5Dp_HTal
zse{jVxqsJ!CijnbJa-|oY4OKnF8S%It@&feZd#Gww)fu|Ysdg@<8$xvkwosAJDcU5
z8xE{nH}&t6KHRzS>OH&H?##I+y)gafJ@;5|Y+2v*>$tn_{L{HhKfZJJsnY|NH2mgU
zf4J%FsiwQ)uf6=?Z}-06u;QV$w<>bgQ`Pot^ChFQ_u7wZcD}Ic&1<QjzrFCR`G{mI
z&-CWh>sKzN=((LD74sI)n(|ZWx4Y7rSuyN|Ap@)FG4YCt-7C9yZrhY5ZC_PCuRC(@
zNNr3tJ-RgRQ`x6$+onA3`tpN)7i@2RyP6(V8rN3Vc5d60$JY6Of4k}A;+L!Gv88dZ
z%3ht?Hs$f?S1$ZLdi42*YWlj;xL;*It8G&re>EdFVEcn7JT(>LO5+h7M+|M-l*g@;
z2fpS}zg4Q~YfIxEl|4GQZOY?cYU8gQp8d|rYC2XLcdhK&xouM(&z*MjJm<jnpTBQa
z+t2;zTj$qp{W3VPW}++R8e2ZVnu+e18;FT9Kx~W!;$j>SALD_9m;fZkL?Ca>3*?LW
zfTWlNB*$bRf6Nc0#1tSkrUGd(4M>maK*3lLC=?3;88HLMjF~`I%mNC>!a#P+28y^N
zT=^9(pH+=jZD$ia>2cTO_9xwyaNQ1+bS-Hociau8JK+Y?opgiA#2GM|1Oq0MWWZ$O
zESPM91(Qv(U~+K|OfJEJ$t5{3`8W?IpWwmdlRTJ0TmVx@2w)0H0ZcJ2f+;3MFvX+@
zrZ?^d)0^;u=}mgU^u>K(`Vu}ceMujfQd|O4N=RTzNeN6jE`up2WH9BV45mNs2h*SM
zgXvHD!BpZ3m`XwcQ%Ndds&N%eHKBs3CRH%CxCW+{(7@D^8kl-q2UAbzVCqR7%wRkS
zW-t*1Gnfp58H$I%3?)KfhLRyLjkp1(kubnCk_MP&+yv81m|&Vo6HF^^foUZyFs-Bo
zW;h-OGn@#68BT`5wBt6IcESeJPTF8bu80gPFA6miBf}%3fJR3qmZ|(cQ(7=CzRAFX
z;dBjdUld%dO91k?q;_&U-C(*|H<)gZ8%)N@fXT27n2d)3lXbFSvMdWG>tVs<oE(@O
z%Yn&xI52r94<^s@VDcUwOu;FDDX;>Vf=2*Tbc$e#tO%y)5yA92y<mD-FPL7B7fhei
z2d0npf$8)3z?7U4m=Y_2DS0F?Wv2|L%*tTO9vMu((+{Sf^@Hj6_`y`13YZG3fT?&C
zFjc1trpl^dsvZ?g&8dN@u^O0~M*~xL>R{@u4yNwW!3;WsU<TPBm_bhv%#bq#W{3@e
z8S;d{G@J&Q25W$6cnmO2rwOLXnqZn96HLo#foZW8n3l%^Gwck58D_&^hCN|0ZKn;U
z&Dvnv9vjSv5V^d(D3lfqZ$t)CB7rhpcHXT2<J?nqj+%*j3(;Fi`$7<GAwlF&3ke|`
zEyO@JTZoBlwGazA+(N?0b_=oFSt<m&H_<&bs&urDNT2fYm9LB+r>fd6iD5;{Qm)si
z_QQ1*DAzvZBIWuFIYqlR(BP+O*Q>}m+LZ@(XDK)K-37{xeRq*^W8a;k-Pm`hX*c%W
zIogeVCQC8cX9^U9eWpk;*k@8SgMB7VGuUTxG=qIMOR?Bz3lxiewn(wqXHztbeKt+A
z*k^Mzi+wIjaoFby6o-ATNO9QbQZ$EsE=_aT=W;ZMeLhR^*yjrrkA1#K@!026G>?5g
zP4n32b2N{AAxjC^7YdYseW6GR*cVc?fPEoN3)mNOw19mvONrPQ3zUd`u}F#77gMx|
zeKAdo*cWrOh<$IC@?zgxpuE`k7AY_Gy(!v@eQ%ofV&9viz1a6<DIfNI1<HqgUy<@*
z-<P6&*!QJrANGAY+J}89OG((53Y3I>sYprKmr}HZeJM>#*q3s&gnc<n$=H_*l#G43
zNXgijQ?!hIIZeyhmvgj?eSennW8Yt({Mh#wDL?l8DcX;Hf137V-=CxX*jKWYf_<ex
zDcDzvl!ARFMJw1>(zJqoB}XgRSF@CgeYHTT*jI~`ihVUjtJqi5w2FN-N2}P^vXq8>
ztw3qm*NT*eeJw?6*w@muhJ7tZYuMMbl#YG9K<U`mi<FLiJw@x-*VD9)eLY9(*bio@
zAoha=Dv14HkqTlzn4*K&52ooL_JcV(i2YEO3SmD~phDOW6{!&RLn%6h{ZN_?VLz0k
zL)bU6l!1MtKpEIKij;wUBSjn7H`26$eIrL3*f+D3iG8y`nb<drl!<*aMVr_+)3k|w
zGe?`)x3ZLleXBrO*td$5g?%eUTiCbKw1s^uM_bqrXQ?pu!v!jg{cw>AV?Ugt!`Kg}
z=`i-gIXaAeJ4@Nvw+obweY;56*tb)(jeR>!+t{~rwB6pfmw!`n?(F;`)nQ_ztMSJ2
z<<hXxKQn)Q_snmltEeFzh6v@GcvlZuIlIpD$YZ75eXi&;tbALkd@)2<wcSr_zqZ2U
z*lW6REsuvfmi}!+=Y}rj?{w4AKi^XM#lR!IOLqX9iR&a>IJ%pO8<}Y)3^LnHSY)o5
zaL9Z!;gN-AA|Q*+L`3!;BHkAG{=P%Rhb$c;60&@V$jJUf#E+~TA_}s4h^WZgA)+Dc
z%|yrJ2{w};a;TYvkd0<yAe+s^M7ElVg&b}sVPw0R*zG*UAAP<1<h_~N(hHF%uKn;t
zxbwt~%;X7!%;pJ;%;gD(%;yP@EaZuREar)btmlc29L$p-awtzi$VQ$R$Y!3H$X1?M
z$l*K*BingmxAT;@PL2FHdD*&)XQac`uY7Ws?nyh`K#YR{VjV0H=iq>N2M;7T1R&8N
z0(l)?AfLkrBsnA?*&ze@9eyCip#Z546-aYvK)OQ*3Oa&7Ax8+va2P<Q!vwM%7Ess`
z2C^OY?=+`F#j4HH`428Wu^Hle7tgQ^aU(Ms!XUF5!Xk4S!Xfh+!XpbAA|Q(yA|iV;
z#Eb095FfIXAri8jAu_T*L;T1}hA7BthN#F|hG@uohUmz_3<)BKG9-j-WQc)mW{8Pw
zWr&3w&X6#&ogsERFF!rJ-<1a*>04TgYin!k$~XHSJ+}H<&(iVJF0KGA#MS<P>e8v-
zLfptq3t^Di7Q!NPErdhnTL_OVv=9MVY#}1Dx0QImf%C)HN_@ytE0K`pRw5(&TZtc8
zX(bA>+DcSpt(9o){L-m=7mxpP^@`H-)32dlV|fMGy6E+eqh_riQQF<>(q47tTMFCf
zC3nrPns!xbcYTL?zWiRfR~<dI<M8#TO1rx=-96>c{oWdDem%+AoG<Mj*<s}1^4s7E
zspHvBU3xuS+TGBz!CBs2{=0P7;^OCV<#*QdM{!D3+edKa-KE#fat$pXZi%9<-w(d@
z8d`3cxub5FF%|<eCd$B!wOE+3Q5I&L#leh=axmj99%g)$hnZjrFcYEz%tT9snHUve
z=C!;q^G3Ze^I1Na`Jz6UNtOgNDJsECwq%&eQ5j}_%MUYu)DJVoQedV;6_}}(3NtmT
z!c4O?m}yZBX1b-rOpod?3tB;#1*1Wjg{%<FLeUV+49kF-5j9|DS|-fQs0lO6vS4OK
zEtrL^FwDZyFwAVrhM66;VHW8d>0dslOE>78k>7s-PC2&zh1PdmytM4W^pMJDc==Q1
zK2>d>Cc~K?a9!Dcrn>`fASS>7u>lr{3vfVufCmx+0+1LGfxH1PkT2i^k^&Nt9FT$h
z0Y8uuP=M5c3Zw-zAU&W11p`5#P#^?k1PmZEU;<eI3n&~21K9!lclz}opB)_=_%>d;
zee2B=?+3W_=7|qk$`c7$&J!8gpC^7~B~KJ&HBVGzEl;#|u9!RRw#gTl(0Owc^Tu3!
zF#h7E?TbwOIl1xFxrHO=RyJLK!}tj|R!k_5Dk^She|UZS!yDTlHnl%Y%$j=J^b(G|
z`D)UkV`V2t=jxhEYP(z-?K*rI+qEh>v|F$4J$m=K;>uviFs-n+XT9zU@Xr5lZ{216
m`VSm5*v)W)c)3rK6>Z1>wXes2RnNMHKTM)WTw_P5#{UDs<zmnP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp.6191F299E76FA25A.idx b/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp.6191F299E76FA25A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..28f819c0fa917cd0c536d699661d0ca793a670a6
GIT binary patch
literal 18620
zcmZwPd6XOFf$rh{zRx!D67cRAV?<*t;y`IvX}8d>(q7mu?Q2N_2Dc;IFc%NpFbM<1
zB!p#5LKw(piE`m4ESUr$K}f<T-6#{oVG}KfWys)&CWH%tJ7J%;&k_CpF+K;K{&hX)
zwGl13TCL(-ArSaQXllw^6E2ylUsPE%H8r(r{-6C9?YrOwHZ`?(`P9@Uw>~Nz*>l9Q
z@`b10zvs(__`kmL;y)@^5BFR<vwBPHhTu0;!PS*(p1<*s=F^WhhowJT-t^w)p9-G-
zr|S*p-4~|UPrY>ZbKi@+7GHap<H%*(|G4zBji3K>e9!&&{W|gNTNhm3JmYWg?{?j$
z+<fz4&Ee}7UGcjukL_Rg^CMS(rSqx2``L##t#@4>efGmoUp78^%hQ4Xx%x+MJ~ZpN
z?ebs$!_BX}{OXTV<}(g_>2Lh05B=sl>)Jl)5_LCqmNaX6_|T^N9x5JC?apML-f;4B
zTWwd|!u?%q>yE%U`})%VO#J$u7argH(90Kg9tpj6!-cVgt#IekUwF3cT5^jf`Elps
z!w>(hc*l{eJU?*k;U5#<JJs-3_*BidyXYm)Z@l-~D;K|gXz4kBdghXP?5y=l*VjK=
zbVGXon<2U`yyu&%?f>^={ekA}v+KA1?ZIVtP95BM{|E7FpB&ur+BeDvJ^TCY<-gjW
zd2#KcyKm)YE;;w=O|O4>=fR7fI3e}z@@?;&_TG1Yx$T8_Prd9b|3f8nS8q6D{|;t;
zad#wl>G>c0W!1eaJip!ihOGJI%EOOeV}Gi(`k$7K_I-b-+idr=-uodv=?5Nex|iL1
z@A&mqarKf-?R9_okE<MaUGMWWPiya5!S4P;dfEAB{iJi~;|Krq`n@gtH@w&F@!xXj
zjB~F!zjLDELBoIi?|-`SS?}}Jz{j7i|KQ4t4_*G{yyt<Hxd(5!@#PJBFXireFmv)F
z`~yc1``*cpHXl83)dP3`e8uMHfB2`D$b)Cz`LkQQH_SZw$obpts~-5NaaZ)L{Fc(=
zUwv}VLHFpxb#lOd-RCc!zv!mxcU}9v+m`SA#eI)$dhZvC_#59%`FDz&_Wk(4JAd8!
zFOPcF*PMTF=acHoc5mEz&Xxb`_viibsP&HXqA%=e-1xnhE_~n_>FlL*FFo<iZ*Sdo
z<_*vNpPiM9zW0x@+m|iQOy6|Knw!55_HB_a-Ej49Ub|7e?&e!R&|G?+^W=LD{aXCD
zBd;#qedBXSGv?ec*PQ+Te|hme_df9U6FWC5EsL*zW7qH3-}=of^}qkw*`L4k(gls1
zpS)@H&8H5Yz4rMFUQ{|hed7)N@4tW5E0@Nc;;yB4JuV-<X7y)th4&A?<k=|Q9sXAB
zvM0W8X>EV^{_G{sdLFxc&6Pil^`tZ3{n*mtZ`pDA_O*v!JaEGwj$HiSmOGBd5~kPI
z{Optyue-u}O8cX$p7_fjnm2!P&%sl^WZQaT^tQPEF6~Jh?eAXr+b3`Q`lH``_QSvT
z_kZ?xbN7Dm9n+~FF7rS4?w39}QCb(f;?-R**>8Tea?96dHosDRXY4w3{zGd%{ne3^
zuK9fHN0&cv;OWY9;!Gp{*uLrUkqaK|Pkp-o77uUS@yWJxl!xzsv#(fVeCps{&t2O-
zJn4jkr|p>9e_*D%flN&;QFgxTy~BLka=JsBDy~_wX6f;p`N#k9U&}ta{go>srX9=a
zx$}oNF4?&1_|4?-s{0R~eDQsAPb{a&`NJ!htXz8hW^&l|@w;!|ee}Tn%jwSf!yA@t
zIN|us<nY&j@{aJ-84t}Zr|J2_swFDd@teuv!Tu*U|NN`lKUq#w^M^Mr*%Ul}GdcX?
zZ*F;5v7z?aim6@mhc8@wq38I`<nUbOw1+w4+lJ*dF@JdNlC?{Z-%JjFx>EYrm%sMp
z(dAT}KfG$ms-?$oCWoi?)f@YN>sh#eZYRr*-!a!d{_*WU{dfCmD~i*uY0u=3IA6ka
z6iS@lxlo*o^GyC-R}?vci$aOq&V^FD)F9P@8l+lOgVeY*AT@#pq(;<$)Vj1FwSpF;
zR@8#jxpW|Pf)1ok)PdBy^dR+u9;9B>gEY7dAPs^6q(L-*G`fr+je-%RQ8a=yxlABU
zf(fKaG=bz@JV;*PLGmIG((E#WGz(^sX3-4N;<A9W2o{hQ(E`%yvVyb<R*+WF3ex7X
zfwT!WkT%f<((bZ@v<r5ScF_*f;c|d<2o8`A(E-xwa)NXUPLNK~3DWIygLDgSkZ#cp
z(&O@g^avi19?=8R>+*v13SN+2(F@Y&@`3aTK9D}q2h#8IgY*l2kbcn*GT;h;3<v>`
z0WknF=n8@i3PF%TF$gl`3V{p>A&?<41X6GbAO%4HDTo5duqzBQEQCRZ#W2W-D*`ei
zL_kKw2*{`_3Nk80K}N+W$e1ezGA6`8#>5!NxGN4aF2q5`#W=`>D*-YgBtRy_1jwW-
z2{I`pK_<l{NYN#N6a^8aD2gB@mjqH0B#@FQflRqlAX7pLWJ*kdOuNz`(?S|#T1<n?
zxH2F!LIz|;%z(_gvLLfU7Gze;g3P&cAag<vWKPV1%)9a+^FkhEUd)3mxC$T(LIGq!
zEPyP!iXe+Z5oA#;f-Jd8AWK3CWJxT6EW64e%R(7sSuBIBxGEqkLIq?+tbnY#svxUE
z6=YScf|Ok{NLi3U%AyRi=Bk0L2{n*4u?Dj4s)MWxb&z$j4zl5DfNTg2kPWc`a^~cj
z&0m}g6=#$)r$e1FbLONbr*g;6WR9K4!I_LSU0FC=nU<&bKsBc?T_{fCDhp>ToXqWk
zYI2t@lv+}QRLg3R>ZTf`M$&-P$QqEErUs-|(t^~=T9DeN7Nkznfz-)5kh-Q0q+Zg4
z)XRF1`lcSFK{9|e$Oe#xrU9f;GJ-V9Mv%s)5u`~nfi%e`kfx>yBrowGd6@^vH+hg|
z$qdpgn?ahJW{?)i0@5N|Kw6p>kXFeG(kfd)TANmoHpvFkCfh*Tnl_Mj$qv#k+d<l!
zc90Iq0n#BmKsuTZkWR@7(kVMZI-5?AZpjVOExSRwn{JRE$pg|Odq8@c9*|zi3(_ll
zL3*2BkUq%=(kJ^s`kFqFe#sBgFZ)6In|_c1DF8Ac2S5gz0gyo{2r?)KK?a*akRd4q
zG9-sUhMFOef+T<xWC5hm6hMZhFvzeR1{rRKK}Mtq$cP*P8EHm9Mx`jos2l|uZAL-H
zq!`GU90M6^#z4lUILNph2N`e1K_;XG$b_5#nP?_JCZ#0Eq?`nqY$id9k_b|iMUY}s
z1Sv@pNJ*AJN=*r5N=kuD$tjSjW(s6lN`p+xX^`n=8e~SwfXv7lkeOx%WLC<8%*t7i
z*=81GPRfDI$vKd@W)5Ut%7e_yd64;L9%Ml(fGo%bkcDOeWKk-DEXqZY#byy?Nh*OX
z$t94bW(j0jDuXP`Wsv1&8DvGOfUL+Bkd<ZyWL2tytjbl8)n*l>EXg2cSq3RLWso(g
z2C^pCK-QWykaejJvM$#_)|+*Z4XFXLAvZuanhlUMmKpElY-L4p#y^vUI(McpsmZHO
z$IfJqz2*dGGRo=G7tU6uRnyy`xM}S|aZ2v=g|ih-#chM)IPF5Im1>Y`l^Ud)Q-jnf
zH6S%A4M+{A0jX7LL26Z6kXlX)Qm52`)TwkJb({{QUa1GESLs3OIXy^&(g4z+GJrI2
z29QRj5u{ON1Zm`qAWcdWNR!G0(!`lS@=6{gui`=S91qg0G=ns&%plF28KgyN0clZL
zKw3BpNUPEc(yFq8v~pIEHl+=uO=SaV<7^=9N;^oq$_~=b*+Dv#4v-F&1Ehm<fOIOI
zAe|~FNGIn6=~lWyx>atFZq5zTqx69Es5~G&oCl;==>_Rkc|m$PFG!!#2hykVf%I`c
zkbb2fq+jI+>F4|)1Ihr%fGPkozy&}Cl|hg}RS;y53xW(OLm)${5XcZ00x2j3kb+78
zDR2VFurdrXtO|n+b77DXWdvkI6#*IHA|Rv6D9ETP3Np$?LB^CZkTF#ZWQ>b}j4R_H
z<El8wI2Q+*P$oboR0)s?E&(#BOoB|Rk|2{@5~QdUL5eC7q{xXNC8Y#XQb`~sP6C-y
zra-1tDUc~H1v0HngG{T^Ak$nLWJZ|*nNejxX1EN<tTGEStIC4Ra#@f$We#Lcl>?dM
zav<}{JjlE%4>Hf?K^Bw+kOfr%WPvMyEGmm2i>e~XB3A@iQkFoLR3(rlt^~5IEQ2ho
z${@>J8DvFS0a;O1KvuX4$f~jmvZ|_rta4S5vQh>qt7MQeCxfghYanZ?8ps+~16fzr
zLDp4ukaeyOvY~8%Y^WL_8(ahA%(j`cCub}3XEN#;9h6~)pVX$0x1aS7``%i=YDG~R
zGii=3oOYzfEQOsOvov;Q%re;7G0S4-#w>@OAG17mVay8H#W5>lm&UAwT^_SCc4f>e
z*wrzsV#{MDW7o#4hFu@CI(B2s8Vju*@N0^6t%=Ovdy1KjllyD(1H~>~cKj!aaOT$~
z(+`t{^SkK*nf@L2h)f^BZqex<Q}As%{SbDKPWPd?F5w;~3;%hz0pWg!JtEu@>=xyI
zOu@G)_YiiEa(!rZm#A@HJs@h_SC5Dq_th<`#(j00s&QZ4qiWpObcqJ{H3Oo-ea(ny
za9`7+8r;{ksRs8oJ*vTdZI@_qUppXL+}DnX7WcI+s>OY6n`&`i+oM|C*L8^w_jLoJ
z!+qU|=x|@xqB`8ywW$vGbv>%XeSMecabG_mdfeBKh#vR#Evm<TeVgiWU*Ds8+&6TI
z0rw39V!(aFh!}9+(4q$1H?*k%_YFO2z<pzv7;)b?AV%Cbj))QWjV)@#ePf#%ao^aZ
zM%*`bi3#^j17gB`(}<XG-_)We+&8tU3HMDsYQlZKOL*Mp2ZYCcenfcO=UbG=eZEb3
z+~<3g$9;2`m~r1cAZFY*kBAxf%`IxieRG?dao^mdX56=Qi3Rs917g8_%ZONT-_oKM
z+_$u;1@|pIYQcSLmsoM%Iv`fuw~mMv_pL2z#eHj=T5;dnqgLFvb%_o4Z3AM%ecOoG
zaNpLVHr%(hsSWpSJ!->!dzaX8-##FA+_#U29rx`mYR7$ho7!>T-lKNhcXWvZ_Z<V`
zz<tMvIB?(5q7K}5w5bF49X;y6eP@?Aao;&0PTY5nh!gjnE$YO5XPY{4-`S&1+;?}0
z8~5D<;>LaVh`4dz-J)*Xcekk<_uW0}#(hthcyQk{ARgTJjED#KJuT|NeNUTuaNpCT
z9^ChKi5K_11LDPf?}&JD-`k>I-1oMr7x%qA>cxFum-uksHy}RT_l<}T_kAtu!+l?y
z`f%UZqdwgCcZnbO{R861egBB~ao^vfe%$xBsUP?KJ?h8(K$iq?KQJHx+z*UM0QUne
z8o>QPn+9+{(4zs|4|Yiq_k#lx#Qori1aUvuqCwmbwrLRegFPC={ZN;La6dF4A>0p*
zNC@{sEgHi8P@9HuKh&ck+!wk;z<psr1l$)!M8JKaMFrd!+El=Op+^PW4|hoz_rn7c
z#{KY!gmFLIqG8+*w`myn!#x_t{YaNYa6d925!{cANCfvIEgHf7NSj7*KhmQS+>dri
z6!)V862<-Kh(vKe+M-e1kG5$P_oF=;#r;^9#Be`0ATivJjYtglV=WrP{aBmEa6i_g
zG2D-LNgVg%0}{vm_=v=DKi;Bo+>f_u9QWfr8pr)amn3jMF(3)tPmD+c_Y*Ce!2Lv<
zCU8H|qY2zkc1aTVlLL~({p5%waX;CjN!(AiX%hF7J(|RQu}ehU7Y9VdeQ`uY+!tF^
z#C@?%McfyARK$I$OC;Qv21LSrX+$L4ms(W9eW^_)+?RS(!u?d2q;NkqASv8WjYtaj
zQ!Sdp{ZyN#a6i?fDcnzYNgDUl1Cqx5^oXQ!Ki#5f+)uY@8u!yZn#TQ1mt=51Gawn<
z&x}Y0_cJY;!Tn5|W^g~#qZ!=Kc1afZvjdXF{p^TjaX;IlS=`UIX%_dhJ(|V+T$ki<
zKQ|yb+|P|j4)=2{n#28Eo91vo*P}Vy&v!{4_wxgi$Nl_><Z(aWqIulUw`m^t^F5l!
z{X&-%aKA7h1>7%;NCEc?En2|+LYo$FztE!v+%I-X5%-G&QpEk@h!k<Z*rG+;FScnB
z_lrGR#Qjp2lyJW^ASK){jYtXiOD$T${ZgBjaKF@}CEPD}Ng4Oc15(EQ@`#jizucl_
z+%LCj8TZRQTE_iKmsD`SG9VS)uZ&0q_bV-0!Tm~`R&c-4qZQn*c1acYs{>NS{pyHR
zalhK4Rot((X%+XYJzB+mxl3f+mj^_~eR)J=+?QKa#(lX>W!#s0RL1>Um(*~-HXt?J
zuZ>6z_iHU$!~I&D)^NYpqcz;GcS#-h>jP59{rZU1alhW8b=<GFX&v|LJzB^8Mwc{j
zzcC;U+;5CX1NR#(+Q9urn>KL2(W8xpefQOWp8D|pi$-)&aqpVFJ0~yKEh_HcvOhm5
z@3|jsz3HjHem?)2-i}2(woTq@A2Z+hwe{S0uAl$(M$1O;<ZWs@<Ezgc-{4q0|LNUJ
zcb_yFDZl=$^On4N+AGHSx7-ie^ij6(5*{~X+);Q1RXt>CY|W5qu(d;`#nug(4qHED
zdThgx8L*8*X2doPnF*U8G9KGJWM*v3kXf*;LuSRc4VewwK4f-m$B;R&okQltcF!{R
z814_xEc0M{XPFn<H_Lq3{#oY74$QIuc5s#ju|u;gge}Z60Xsa)!q|~n7Qv3rvM6?J
zmc_8+vn-CCm}Lp<<Sa{Ki?d9`mWE8i`zJMIDeUx+rLi+Zmch;rSr$7tWI62okma!p
zLsq~p4p|YqG-M_0@{pCWD??Vnt`1ohTOKkQyEbGs?D~+^u^U6ySm>%=cw|lGSDh{M
z52QYuK8z2fKI0C<7mm8m)YzIn(_m}+OpC4SGaa_R&-B=aJ~Ln&`^<=K>N68I-)B6w
zxzEhlmOis!Tl>t4ZR;}|w!P2n*p5DPU_1NFi7oY+gq`ZM6n47L(%6|k%V1~wEQ_7%
zvmAE5&+^!XJ}Y1s`>cpv>a!AdxzEbjl|HLrSNp7rE%%v>UF)+NcD>K)*o{7GEOhnp
zw<6CqH{E{h8CWrW3OxTRxKkE>;GtHip)?8&lvbgI(kXOMdW9a!pfErg6-FqN!UV-D
zcqp^N3}sPRpsWfjlucoSvMcOR4uu2Csc=HM6>cbx!UN@1c%gg>ACzC=hYBbHP(eiy
zDx?TO2?_x!tO!Fz6cMPXA_^5##GvAeI8;KBfJ!QoP@+PFk`xkDN|Ay}E7DLIMFuLX
z$U@~5IjFoM4^>bUpo)qjR7p{SDl5uR6-5QAs;ENA3K^=Vs6o{gb*P4-@kQ-jH1&gf
zv>!in><hfZrk}!R#t!42!e_<~Q)6p7OoOfMFfF#O!*tmC4%1^BI?RA=>@Xv?sl!az
ze24Ma<_<GsTRO~wZS62Cwync#*!B*yV>>#`f$i)tC$_u8+}NHD^I&^B%!}>oFdw$R
z!~EER4hvuhJ1mGD>aY;D&|w00xWmHOkq(PsM>{Nv9qX_dcD%#l*oh8HU?)2)i7j@R
zh%I%Pgq`ZJ6n479(%6{}%V1|aEQ_7%upD;2!}8dL4l7_6JFJLZ>aY@axx>oXl@6<5
zS39hVEq9oVUF)zKcD=*u*o_WrEcE0rgZUrUAG^4B;fMA3UBT7=&wq;b>F=C8f1bZ`
z<;wMw_jX?U$6FKJ<(c}i(~~iqp2G*>m~jj56j)JIkC_@<GiDlW?U-q?bz`Q(){mJU
z+c0JZY~z?2u}x!U!sf?}$2O0d8QU^u7HsR7S+Q+nX2Z6RnH}3PW)5uUm^rcCbIko`
zcvAArF%PzPj(M?tbIgbBpJRUPz#I!;2j^H2J2b~a*uoqWu)}jKj2)R{5$xz3i(<#-
zSPVNp$Ku$DIhMdq&aou6ILE|={>fWc9LW6pt>2|9iklLf;3{lYQ1!(%R6o7sk;TcG
zI|K7ybQ@0GuzvEklKb{IfBv;)b>IA_w=dfMmC1LgzcQX%TkOB{t@%%{*}8_CTr&Uo
z&h2lOl*9h~@6=q>gsZOhE!y`*O-_&I9|Om(U|)C);P=UsZ*J$`&wT7U_6d4JyZHSR
zPTKJ8`QLZ<y4}i&nq1jFua8~UKGAO}etI3XJ#g&%=(StdYA0%Pp}T3>@eAB1dj4U>
zPg^H{y7{ZpdCod{yv{$Qcr}!U*Fb4`EtHPeLFsutlz}%u8F?d=i8n#<JP&2&%}^HJ
z0%hf`P&VEMW#{cs4&DLf<egA%-VNp9Jy2fW3+3Z|P=4MI72pF<K|Tl-;zLjZFF=L)
zFjRz(Kt=f|RE&>7#rZf?f=@st`6QIci%=3TL8bT<RGLpiW%vwKmd`@v_#9N8&qEdX
z0#uPNLY4RuRGBYBRrm^2m9Ij{ybM+2YfyE*4%OfrU({!h9(jEy|F$&$jyiXrxu3!3
z#Xj?3d;83b?dvliw!hE(*nvI^U<dmwh#l&)5Vp`~0(Q91!q|~Mi(p6lEQ%fLvlw=~
z&*IpLK1*OH`z(nq_L;cQ^Z!SS-IZg{X#4lc`*$4sAbafQ_%Bxc@n6rfoU<a~UQ){B
zv$?`lZgOa9D!=eyX5qua!iS}W59OKK?(^p}(ulD|i<g|BSh{@0msXy1qOi*ERjpbk
zcvf#%v-ZmyEmph3x!rY+dfi#JE&NxM|J`TTZ#rf3Y15~lvGvStUp-r+)fr6YQ;jEc
Rdf&SB!El)d;_eWM{ulDJ_iz9J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp.1378958A5E1CC674.idx b/.cache/clangd/index/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp.1378958A5E1CC674.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0141c037b8b451ac6757dbef814ff51463cc8df7
GIT binary patch
literal 14848
zcmZvj3zQS(p~v_8zGVr^io7YT7@`nFI(cW30Ldhi*90`l<S}^w3u{2=y?_zCSA0}&
zTP?+^h_w$CwP=G<uu3a2sjVKVCcUwsrPjDXZM7A{B3d7KQ50|3b7KDYoXa`p<hTDH
zzs;_*na%DH3k3qF9bH}CRiVx6S6?&O-PP4Kq4nRUYc_uCudJ)<>aks2n|J<RS-PpQ
zVEm^K9*?h)rHyvZylD31>@mxjAMO58aNAw47JD)uZC|DP^3<Bu|8Ps6M>lHCD}V9j
zFW$K7aV?R$*8SvLFV4x&y-j{#z3agbyPvuJtI*u1QuV*hJ$~v*!9MZE>x&N_-Zt*C
zcc;A@{O)t_Z2x=twxN^W$*<qEW7$_<EdJqxg7)|qXE)qCaA?iLt}#bGIlAtF!_Tg8
zUiQKJR~^~gfA{ZtZtUN-bY-P?=|%Gw{q*&{Q}(_$V|neo`qIZ2O)Bo&_wLv`j!(Sb
zb7<;;BX`fc_!{oRTh2W4w^#n{t<ur=t*<Yi{jWbB*gf}t-JiR&yS87-4o!RX+pjNJ
zv;2$YyN-Te+_7i_|D62Mh2|5BFVxR_m~MV$`eS$8I_l&-W8Ck)ym|HI^QWqXo6dG`
zPj5Od(J6~Je0So)Kh7{+mmfMZb<XJny$^M5o&NZ#@Ey<BcK!W3E4PZ9s_c!A7ry-J
z<nBjza_cw8w)DUKjl|Y#{xCZA!<F}cFzcf`e>d;1A6|IlCD&?m+?HvxH^rIl)pbjj
z-mv!6`x74P6aRbWJ30Lqw;n!t+rk$L6L0BlZ2W26ICG&`_~-$8{tl7TRI*n;)_i+H
zSl82EobvuZzrX0=yL@7PRx#Pf9(_ICyLSHG(miJm{Lfui7Z!fUyKRI2-g{=pZd+R#
zr#fK%=Oee=eZ>396@fEfPCa$&f82BPH<pP{j9+?S``vF$yZQ!h&w<R0r}-z29rk^Y
zZOlA&-S?k(^f!GoUwPo&*T{jn5AEMs-oF0%r`Ng`PIzLUC3)HWWf!kFc<J*Swt5<$
zP00loZu{!hwcS6yD|yFH?jM_Y@>kFGfApj({LY;ze?soxxcjp=-d)1quw~igKTq8A
zgXO1RGcEqFUg5r%Du*_SFa7M7Z~SRCx4*t}*JJ64uk7D4YTA-thA(;~x#wW>o8&(C
zl-^&9m(JT%{Lfn!zZ2N^sczhNo}2UVg+|AT6`vly-V&R4^TR)T>e)%J&3@$2Z#EP+
z&rQE~-HF)J!J}LDm|XeZSugIrm$Q!=yZYW-=Is9cfBBPP>-;$e=RJk~J%8CJZ2tV%
zJElYDef!1V?|$IXU*00WWnAMQbMq<P^=sDc>YJaK<1;<;){>pE2T!NCMc>->_h+|?
zFKKgoe{}k$ow5EydjxgwniId1-aonEv+dE3cRw=YUt81Twp5?HVBFGg|6%aj4fe~<
zzT&&H=l+SKcZcWhkbk!2_WMs4S6vZ$aKLl?#-|_pN7{Po)$1!Ww%vH^L&s0|+%$Uc
zZM&cQ;L`VI{^<k5p{I)DuG)X?C+E*;ob)~W!KdGQ`{u1@)$88e^x8Fv`*y^iSf;xE
z+U!ZA`>UP1?>g_$<ePh6-tzIke{ML~efiZ_?0k9c;<Ie@eY4_So32}b#Wd2@)uS%m
z>D^_UHI~Meu7OEClg5nPw7&jF{`7wS^Ebb@#2Oz<W3735&-4i+H|=@C<6CF^$FGL}
zFqTGJ^Z1_eV@7V;vvB6alaC&|?(wlS(VD0AOdCCN)1G%9yQMiC|100um2Ax`My(J>
zZrby~<xhSY`}jy^EKRp&ZI4zMxoOWw0z?0luQ~1=OH-}6zo$Pqa?_rN2WS13vz#=K
zrIFS=xo7g2k(>7X<#^@SZ(Q~Kv9VNc%@cYij2XFU&s`f=U%6?VrStiT_l|sC?l^en
z=lkATIIC|!BWT3-Luiq#Npy%4II;cc`vy3fOLRyl=wQ>yI@t7r9yYzKhs_`uU^B=D
z*o=Y^Hlu8W%_Nv$Gsz~{%z_y<vuuXVB3NLv$QIbFf)zHaY=w;%c-VNEhs`F~V6(|K
z*zAHGHoI(x%^^5obI1<ZoPrZJr|g8yCAeU7$u8IgL4Zw=1=!qz8#cG>hRq{*VDrcx
z*hE2uO_W90yn+`tuk3}*C-`9V$v)Wpf*&@&?1wEN1Yiru0oZ~<5VoKkgiR78*d$qk
zEhL0s3&|na!a^9fupEXhB1B+|$Pw70LKL>B9ED95WY}a`hAk$<V2jBy*y2JQwzwRJ
zEg>XeOUMb>l0p);q@09J5fs=IS%EDjq+m<QDcI6N8n(2YhAksxV9UrE*s?+vwyd0m
zZE(im%&(sy2L@*k&TUisIr+xkk>_G~N>Vg~ohKnpPO||hueqT^oWc!uo|3p6w*e^6
z-OwSOq61Tx(}AhW>%i12dNB1lJ(&8u9!!H`0Mn2&fN97Zz%(jGFpW7Qn8v&jOp{^)
z)08uTY08_xG%IE>%{eoe=DZn9i(&!OlCywm$y>m*DpoM9IV+gfycJAd;lbo{JeYi*
z2h*n5z_jITVA}FFFzt#ROnc4_raf;5)1f%PbmSahI`R%Mor)7oXU++xGw%e`rMSR!
z<y>I8@-8q1MF3OC31A9&0Zg~z2GgB$gXzw@!SpB|Fg-aBn4Y``Oi>ZR6mue&VqOH(
zt9Zfm=Dc8f^IkB0iVsX*&IhJ1?*r4X_`&q&{9yX?elP<{0L(xx0A?T`05hlr!3^et
zU<UI+FeOC-Q_4wTN_h#)kP-qjlna3w%7?%VD`7ChxiFaFd>G7#5&<)ki+~x)N5G6K
zQ81&qD45ZF6iiu>!IX0{m~vhQGp5AAjOAir#_};R<4PRNcrFfRJRb)$p(Mae<Pu;e
z@(D1LN)pUuE(vBbp9E7;6fl*X0;ZBzz)UGAFjKh{n5ldU%(RjQGo4F=na-!d%qST!
zGr0_ynS2J!tda#Yo6CZk&1b<Jv=4gQC#}ALLH}SBC^opfP3>0^trOb0mk@A5Q)@2j
zoU}Ar%{(AZW9Sg4<}T`-v^Xs{4~XLo9nz_FVCu9wFm;>`OubqUre3QDQ_tzaG^h<=
z8ngy54V(c?quL0jQELR#$Qi*jsZC&-v?ef3oC!>`+6<;yYX;NInZdNEEnr%-7BDTG
z1x%~j3Z_+S1=Gq|!Q|CEn7oz;ljnFaZE72sHmwaz8)pO4uC{||*V@6fb9OKtY6qAO
ztpiL4=K#~Gc7o~DI>B^uPB2|+7nm-s3rrX10#i^6U<z6ROo0=?bgSK9y0vaF-JBat
zkJ<yKN9zI8!+F3I)gqXpRs>VzL@>Q-FPL7f7fdhb1=FYYf$7uw!1QrGF#T#jn0~Dv
zOh4xbGoTKD8PEp63~&K3gX$odL2VGsAQuEvQcGY;S_w>vlfVqALtuuqAuvN+2+XiL
z3}#px1~bfs!HlRQU`DhNFe6+9%&0mFW>gymGs;E5l+`krvQ`FD=43Eq>KK?YZ4As9
z7Xve{j)NK3#=(qpaWE6=1eghJ0?Y)L05hphf|=AN!Ax>VFcq}|rlM8AR5%69lsW}w
zN}B>R#ihVZtJ7enwP`TZTpG-bIs;}#n*lS!Wx&j;vtVYmSunF)7R<qUgWqhQv|1-L
z-JlW3Jjl1H|BP$?^&3x2J+E&d++^Wl);aw|nk<4GZL%n`++;FxtjS`?@g|ESCz>pQ
zoNTfrveIM<a;nKv$mu3aBWIc{gPd)$Y=;dse$oJ)Y$dIaUNtzq{qeSc&i5hd9r*zt
zoWaXP^IOt6PityK^CEJCXpSNmsAe|>zeqJtAy=rT3e1%W_gm8W^l~-Ay@=c(+)?BL
z<#to>i<El`xk9-ru&zvW*w@vF4*R+W(P3X#pgQd9id2VvU4`nfuP+ll_VqQQ$G*Nn
z^w`%Is2=<JBGqGGU!i*J8_L9heM5~Huy1G(1NIFCYQVmsNDbIGRHy;_#xgNt-&i9?
z>>C@zh<#&$8nJIIQX}?_6>7x3sZ31RH`Ry<`=$mlVc%4sChVJv)P#Lgg_^K$E)z5M
z%{5}izPUln*f$rb8T;lUHDljgp=Rt`%EW?wOO05tZ)p$<_ALc!!M>$PE!ekIs0I7h
zGO=RcS|e8MTN}iReQSYQv2QI>EB37wYQ;WZCOr1}8sV|eHwce?zCd~G^F_*IpRZ6J
z`?fN%Vc%9GHtgFP#D;xaf!eTdD^eTwZ53+6zP(KB*tge+9sBkMv18v}pmyxri`0&N
zdxhGu?<f-o_8m3iz`mnF9N2dhr~~_sB6VQjQK1g(JIln0eP@k0vF~gUC-$8M>cqaY
zNS)YsR;UyEt}=09-&G?n?7JGog?(3ny0GsmQWy4J73#viP$mNQg&GmCFEogNeW5@F
z><dLIU|*<E0sHPUabw?IBW~=w8^n!$cY(UG?=Dg|_T3fg#=fUaJlOZthzI+g2Jv9u
zQ=lH~dy3S9eNTmYurHR0h<&j}MC^+VB4S@GP!aoLk&4(CD^$e3w@kd)_tuCP``!le
zV&7Y!UhI2|)Qf#@g?h2?D-$2~eKq34zOO-i*!LBv5Bt6%^<m#vp+4;U%fyd;e~tLD
z?{5%4_WcFw$G*Qv{n+<cs2}@*G6`TmP$L2C2O1=R{Xl^RupcPW0QLhF8o++AOoG@C
z)<_Wh!3GIpKUkna><5c9i2Y!N2C*-diG+QrMkMS@4I*J*Do_dgQjtp7mnu}keyB`B
z*bmi62>YQ131L4}pdsvsiZq1%P=$uDA1;$H_QN$2#(ua#!q^WNXc+t9A`N3dT%lp?
zN6I9E{YZ^OupeoV2=*fd8o_>~NF&&fRA>bI(K3l*KUyPE>_;0Uiv4JTMzJ3)(kS+$
z6&l69TqZL1<r<N(FE@yceYrqo?8`+eV_&XN8T+v^iD5rhBQfm98YG7OSb@f{A1l%r
z_G1+q!+yL>;@FSZNF4j|28m-oUZ8R8$BQ(M{dk4Ov7acD1ojg(lE8kVK@!+c6lenb
zi6Tv4KT)9x>?g}4iTz}aB(a}tkR<k#1)9WuvPhHIPgZCW`%0N8*jH*q!M@TU3ig!(
zRj{uVse*l_LKW<%$|QyTRE?ytpK6d4_EQC#!hWhqQ`k>cXbSu3GD%}UT_b7iryC@V
z{d9q*v7au|H1^XKn#O*nOfuNd)JO*VnFh&VKU1I?>}QHJgZ)f}X0V?vlPvbLHIl`C
zwn4Jk&lYGF``IGRVn17<*-qcH<-M-YKDnkry9ZWJTAgUWT<#uNbMcyGZSuxGpY!7v
zKK`opntDn1l6mbbP%gLKebrR%_Pbhdo^GG+ZU53Y5uP}=xqZ>7)|*$3S$TdtvhwX8
ztm*mNtT!)dU5Zj?nq#c<5}K<s?il<8LRV)xWPP3KkqvcbKsMHy5!qB{CS-G+nUO7Z
zW<j>rnH8C@GalJiXEtPeo!OBcb>={J4l!pFJ|5Q)b0G^uOh9%IF*mYjh<T93AtoYw
zhnN@HH^h9%{vqZ^4h*pXa&U+Rk)<IfA%})o2svD5Vf^?abrwO6)>#x;t}_`qR%bEf
zc%8+O6LpqAPS#lxS*bGxIaOyV<aC{-ku!CcLC)4$w!>GLLQhW`e5TajIuNT&a~KcA
zD&r2r|2SQh>5%nRrbjkZnE}~YWkzIEm6?#uRc1!ERG9_YT4h#bzRGxHTb0?6?Nw$+
zc2t=IIb3C7<Vcl8kfT)=MV6~fMvhfk3^`t9apXjmC6JR<mPA&nOhHanSqeE_WohJ0
zm1U5#RhI4W6$ei&`Afe4=jWb5RGJIm`9sBB(D{LxPNf6VtMotyl>x}8G6I=YCLpuQ
z3}jJRfUGJj5U=8aY$_X&U1bMys2o5}l@rLNasde{0m!X#19?;)AW<a(c~xE@pUMa1
zSNVYgssK<>6$Fx05>QAL0t&0bKoM00D5{DA$toErriuZ@RdJw%Dgl&KC4m%`0+dpv
zfYPcoP)3yj%Br$o)79NwJN6jPynOD!DKX6p_!VDb+za>>Ut&6BeTnIj4JBqkHkOzX
z*;HaCWOIp`ku4==LAI8d6`3zF9@$o6He`E=*^wP3=0J9qm=oDmVlHH%!~|q_iMf$I
zCFVgEOH4%emY5gWS7JV7e~I~#10@zf4whIDSt>CJIaFdH<Zy|Fks~D*L5`MK6j?4Y
z897#BG30oO#gP*wmOxIHSQ1$&F$FnQVkzWwiKUS<C6+<XmRPpKTmI$0!12=UGg{vo
zjUPXLYWu2!JN~gV!rh!%eeR6ZWSU_-KAVi|{BomjK-XkCWPOwAkqu2|KsGj+5!uva
zCS-GynUO6`W<j<#nH8CDG9KC1WHw}bli86SP3Ay$4m0O_@b%^zW-er5m<h=4Vdh5m
z3^Nb1ILt(3?=bTs`-Yhh*+0zu$bn%NKn@PGAhI;fB;?RA3w5~ty>084O6RT(Z~szh
z?}_hSmw9XFowRSDKhh6ZU2B8d%i1TZeVd;i6<z;Opmi)v8#irg`xjKd+LZszRlTcy
ztv4^|UT{hKd+lc|u~nn|51nYedD5IoT>F~%g9+F173zAm^`2IBuezxHc=`se?7s4A
zYQIO?J92e;`%(7&pYQS8XVaE+?n3pBgf9!;F^u|T^!d|%*m@Kzr>s=B-(CAc_Dg$5
zE@1EcU5%t!(hj$-GMBg|?Z?tOk|iCGUeW^@Bm<C9G6I<-6OdUl16d>skX5n*@e&VY
zlWahC$qwX@96(OV3FMMoK!PLyxg|G{NAduPk_hCLyg)w52jrLhKmjQL6qJHMk|Y6z
zq!3V83Ij!?2vAgt0?CpL6q8~=aVZXzkP<*iDG8)V3Q$T)0i~rhP)5oCWu+|8pnCA@
zqxOVq-!{77iF2<mC(oH|Xdh$ke-$yWcjR9~z_G?_9B{1hoCA(EUI(P-^*{#R0A%Ei
zKqlS<WaiC47TyA6<*h(G&jZ<b8<3s1137pHkdt=;xp)_lzzaZb-VNm8JwPHa0(p5a
zkdOBP`FTH3fDZr#`5=(QOF$t$1Qh1OKoLFy6y>8pGA{$g_!v-}j{_z61W=Ms0x7%#
zl;TrBX+8~<;WI#4KKnJD{r%Cm6Z}b~b^U|0%A7CbDXGd_$U>C~$nGk0BYUdMgDh5=
zi0rL0FS4)7e8~PP^CJhUEPx!WvLLclWfF3z%0eC9xO&~-xs%4Gjk!(nb9b`mZbtsF
z0mn^m?k|*f+}5)qvn;!GdDqhR)YY}Db2rnuyS#IEMdxmA{i=0qTe#%P%USoRp3$l?
zWBbk<fBv}81ix21p*JK>oHl9lH>TSi3l}+Ef?GFb{(_77OVt10Tc`G4FmsmXqS<ri
f&b#!RdV|qywOweL!I^whrUn<Uq=B$UBA5LyLYOVP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp.155CED9F59B77387.idx b/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp.155CED9F59B77387.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d8138e15dd309456bc58999fb46f9e7841b8f9cf
GIT binary patch
literal 4066
zcmZuzc~nzZ9?pA^5CTddEP`MlZ%0B{qk@_Wf^5njMHZKcM@29!7L}>f;#5R&-|7)u
z8^#3$2Srm6E86Hp!744|#H|)Ya2XebR<U-Tne)u=kI6Z?=YIEF-hKDJ-@7j*Hagnm
zz~e<Oh+UDNTQbj@$KyG&|KgIO#n-KPJcS*Px1##IIzO>{ytre(KJoZbMtiqFxBAP1
zvdnv^%eD2Jx8@}}9p6;bRI}@T!-c{e+Ir~HPD#pM)4=WE8!ffApP_S`5;Fb23QtZL
zc=qa=ZQ?IapEZ=ME?znK#l9T5!OiO8W4GZSmT2dm3=8)X`ghor@Alm_?zd#^!1C3D
z1BtWWzYk2DS$bx}o1)Xzn_{=rKG=1<WblvIs|(9YYjSe7_sD;_v$PA1SKoHZc&P8m
z$<N3w?Mo>*d%x__p2pnVNptcN#`W1RX+E)YSAzS&yvk{JRUc~Ac~|azIPh@D?<<=V
zp_BScD#;&38`InuwMIYh?wauQ*B7q$COaE@3I|4c5-)t3PV~Im7UcWow3?fyS+hE4
z?BSJ++&HuA)cjZdox5*{D?@h+oQ@<FNp|#k=Uwe-y0)vCzW5?d%#XRgqCPy-X>3Kg
zO*tWI_i`w?QF@@B5hhK@G`;+Gq_k3#MBEx7&61ZdZ>_i*{<=2p(xl&$J|{KuikWzE
zs@1rzXYrY3iw4Jt>+K`{>Gi1F`@K!g*YeW!@r2Oj)Fb}}O;}ij?)5&@VNZnI$1uCF
zzVBBa<+f(Tuqa&bbA!gWiYH3$_^w6L3jFw|cBW?;FPX}&p1pPc`-9WY5YywFn5d8}
zhrgNJU#CXI_WLXeJltKl;ayvpQI%C4*(lCDy>;8YnT2IzkDmVbsB^)M-5pb3%r;Hx
zUAeaLM3cSOPx~I<wmMB6I$5pXntyqCQR{fe#=4O7gt2#h9(=F7Tu?dX$@}$zhik4}
z$O|ctyLv1-{xRxOpZ9X@M0t6tzB}D9{in6ni&s0m-PV5J$1UpARNwh-lTTGS9ca6u
zzLqgi-5FeaZ10|Li6M`=(=MaSR}D4x&X4>#w&#>@`ZxRVxsoFhQ`9j;Zb#Rqq@#Ct
z{CxihA8ldBk!cAXhR27uJ8n!fJ_zZ3zS(-r_U_&cRc4FCZ>VMY(fY>cm(tv&<M~@2
zrMGugmwz49e$wyHrjy;*11d9#fG_;xrMekSO`U&V*2?tf-D;V*$fNRD+$fhZ#p)d|
zhx?Y6v~?L9Bhn-Aro7gR{K3k<cD8MI6@55IS1k={3afl^g@5|y`ga4Z+UAZ8Jstnq
zCbbP+wyF-<zpi~jkbUohTTk{Tyer-wThvlnRUR_$+0{}n$>?`gqBz5_%ZN$&$8>{L
zp<~{PqB<K!yhqe7gRdK&bX#37{ie4pv*Negqu*ax_j%dPhO)to{?_O}9hA}uzN?P4
zU#4gITc3YhH#BPQU*du<uU#lm@%zqnj(r!<X5Bt_e)W%YV?Xc@9kKB(UYS4tceb_#
z(zBZ)503M;LsQkfY>7Z(Yi?M^{1d%9aDPpFXsR8W!qV;ncSm!>(T)w39?MTWxnzeX
zv9wqqwly~#tr~vy{7CQ020Jv3rQHN>Hs*$-mx=dhuaoatY=_QdX-|Qt%G_}Dnv|)`
z{p%UM_PlhKmh<Hi=7yu6%=7+E9`Zcc4ozffXMwY=x#8$P#Oi<CTXwnE4ozlhM}ecQ
zx#8%d+=ay+BFp_tMCSY6*fxBi?qRsMeYPw{7SBzPMai->3&LVBz+uT4;P4n6aCkBf
zI3k7s9Fa@_j*KAzM<$biqhct)QOOkG=olJsbTSP%CWZkVlgt24Ayvpet#GyiS5O?{
z_S{os-fOTVYS}!?BFHqd0)Vu#G7G}g7*Jvw3@9-z29&rO2TELn10}A-fs#-YKuKr_
zpd_>eP?BmAC`k<ol%$pfN=i)uC8eQ&lG0K@Nvmm~q%|~9(pnlQ88rixjD`V9M#})D
zV!R@fTXg$uMU)~5pcF+mhqx+3MCR(^HYC8Pl;oh!p$O|0D%?3WS9F0QbFp!1_8HbA
zvhNVfCdrJ5>>9)-M0N*a9V)9ufxa1)oq^bZ%B}#6=@AU(V@3pn`Irg8U_PcpF_@1v
zqZrJ`3@8TkaXo^=eB6lOFdsJ|ILya&C=T=SW)z3{xB<msKA}elm`@lH0_GDYgn;>k
z4kchd(Tox>pD>^V%qR5-3G+!KLc)B~gpe?w)S)EIC!0|c=931Lg!z;np<q5`L@1a~
znGg!*Q#zD_`BXDX!F<YqQZS#^BQ(sXjR+0%X%j-jd|HRnFrRKlX_!wNP#WeldW3=b
zj1ggAK4U@{n9t}?2Ie!(C<F5u1Ik$Pr&K)Sy?s++LaeicvEVGOTBeWNw#-#twS_Gr
z_wnwL+`BJrh7;Adbt0b~_P6zS;MDWmp2l}-M;{Qf)g+X8T51CmVgTVn93VnS07ME&
zfG8ma5G|wuVuZ{m#9av$nU6?IQNw%TX34%MKAx7GsN!Xm=Y$(o9_+n{#bQ_PZF2nS
znw1$3t2eR3p3^)NxGU7HILA-QKNQUlyNz^n<=!T(At?*^QHP$g!xBOw=RO<S(kA!M
zk^bunI~-&k^yxr(6u^ID<KVV|9rkzemvR&1j%<<0d{lE2WEWE<i{K_;pN*+7fN&KK
z5TPOfB2^?nl!^j~R?z@4Dh42hhr)}S)Kg*IW;Cqv;}hUaI0tZZ<h~zpk@>p;%vCCr
zS>`GuWeh-aSpYyv*#ry1q!{pFBnEt#90NW~i2)xj#eokeap1${IPl?09QX(+0el2W
z03RVIfR9iTz(-0+;3G*A_((Yke58^DK1xafA4O8YN69JRqm&fz(NY@tXp#m#T22EW
zt)zjEkut!?kPPrKat8PqB?Ej4M&ZZpJ$rKcD*^!uRwy}Cl)H4^$FIfWB297X$4<t_
z=6}E0T!+WYv~*`#x>=vPR@oy9v6gMG@@fS+vobSgXYq1AaXd}_f~AXC7{743l{H^r
zBeb=%A0c)a8S5AoNji#RBb?kM&ZFGNPY9nlX|n1wf0wagn39(MzfsrGo?hNQGMu0o
fUq3}aU~uTTF(DrEps6md)8gi!(KDvRAPN5kG4if%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp.021E4B84540D9AAD.idx b/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp.021E4B84540D9AAD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..28ab0b9f44f9d0547c54c87efc9651fb4df52aa8
GIT binary patch
literal 4068
zcmZuzc~nzZ9?pA^5CTddEP`MlZ%0CqB?_piAjqbGLKIm<B90Z%uvk>4PK&J;#eJ(s
zw5Sbr0l{I>xL`$#P86)tGENk&1rfJ#L2L!od1mIB-yf56xZLl4%en7&m-l8yM@HVZ
z<MF~5L@&?FE}C!2<MHg-zrvz|PunbbJcSL9x4h!4IxoI^g1G%Vef-hGjP~{t-KsB^
zl%#f{PFL1#+L9CRaCGC|`n@~v)Sk=Fpe<it*ddv@$Jl=};M$og>rc?k`nXizFGCaK
z`k%bGVjch6<0rL6s|r^Pe7-M3-s)m;{(;NTPm8sg$3sFrgud-orMtX$j{hT3+rMnp
zK!5z);o))dvx`rxe^t;}u`zmc)!m&(iw54lT$Nu^yf-6bTaWy=Te+QRtoo)y$~}Eg
zMqWyGao@}(C-0Qp-(8oTJ!xJ}-1t7*#ZAX{?2L2WpHnvNw(4D#I_FZ?yUKg7eZI6h
z7BuN$Q5pHOXhV|gqUOk_-JQXYe}CrOHQCY7lixqeop|O|f2`-lR)6m=rtQ6MoHM6m
z#%^BG$PKeQ8?s(J?AUcpTo$xT;BY9RK(f8hGv{(o{gs_f^!aB=Vt&-s<u#!}4rAAr
zT9p!_HV?a^YsHl{j4&ZM)%g7Tk<v0v0&!!6G)-Q*ta<I_(3e$77bg9g@ENI*ubqt-
z&axQa`6M>AWYNGFagA-*KRoVtdk$Oe{aRkUE|w5FHQe{D)r5qE>0a(b?RJOBy;|qy
z*Yy3)qg+;x7!rl*y{^&t7V$*ME$=l*(h?tj!;a)M!v$l><&!tg{<weI31WJT0}~OL
zX7`oR_2sOv=!ag5#~tj>Uq9FqVo;@3gx85v8@Fujnw?)V_Hg6BMx73*>u#U=Y_4%q
z?}{~b$Lei8e%bfnrbQ!l;CO|8OWwtyMa>iJ>#75j<Hp|hy8EN@;*zo{kA~NcJGl4K
zxtze#n9D~ZV;`VSH961MOq7?-(sw7@C;zgh;?q@jZ??AG@p6f1nChM7GP$ALp|a(g
z`btWFMMpr@kv+S=CtkZ%H#&_jUHQ7MH!J+t=$;1e<Zt)knUX_cQ`AvJc6;Z>gu}PD
z|9a;qFKvGNp=ojLtq%@vv)_<pxEt8}bd%+nZQZ>os?;+QpVw!W9j>W+dLhYGI)T6W
zesWu9Md{c6ZO47y*B|e`>Q|OZ_<imhE7i@YukZL_X*1JtJNL$!iHqFIj>L>|8dIp=
z{(Pt}x2UDlP#2aQhS%pbpXU#h{jH;Ao3rTMX}Ubuzdoew*(HAC^>u^&&Dy5+^*!zX
z-YT^YTDr0t`R<#x1^%|Z3vN8x6E|46ExO=LS$S#T{3n--JtU(C%SAD*Lrx<m<sH!t
zln0G@BZ}y#9Wjikod#akKI*o(TKsKqN$T1^ZjK&4_swS|*K11#QXV!(zPD3KgS}TC
zX}d_z@wGhrruy}$%)iD2TwJqIpyKzP=omZb*J9b0nN{&~X7oG$ffH8VyuuZES%0#p
zTOd8TF?|1cPaAZWnwKsSNUTi_%b0$Gsdt0A3O$=`dC4p-=gY%P4M&RxEAOn14Vq<x
z&SYs<fvdf#;b{BXGPh;N9$m0O6Ifa-5L=rXj#dr5czURJMXe2*#L_MT7b{c4(U0bP
zo{|SX4X{DuS=v$HXl-gZ`mI>~_pYTEdu`A}mbMqzTbr7Xbl&@~M(lelw?SvKw7bAv
zWokIOAbVk<o5=k9B9ZC2H?#~@R^JQtv`v>K%97sxFT*w+OT+-e6LEluL;@f(kpzfJ
zqyVB5X@HnS1|S8lptx#t{}GX?+T5hLYP*X}Rpy4-xvFLJ&2yD$WJ>_j%1X=#Q)57h
zX)vI~v=~t0Y8)tW4GxsJ76(c~O#mgKA%K$55<p3+NuVS(Bv6uC5-2G(1(cMA0!m6t
z0VS=bfs)qHKuK$9pk&kxP%;_@C>bpSl!^(8aBdB3(-jel1b}8L(mBNKZ+4wNY;JBr
z{0vG-2I?4uuwJ3Ul~Z%ueVoX&%{eu@e0oH73NbH|%z((QKx{;0w;<M`vMLnln^4&a
zh+9$FC4ezKg28;ufM75mGa?ww$8;zL^RXrrgZWr1iotwbkKiyLHy}97$BhUM^Kl)D
z!+g96#bG|)isCSz&?5xQCkzMy^9dtDz<ff75-^`=LJ62pw4wyeC-n#k^GO3j!hF(*
zkT9Rrp(M;Fn@|$wldUKT^C>+-!F<YqP%xh|A{5N0bSMS$sV0<y`BW=P!F*be&@i7i
zAT-RUjR+0%X&p+#e7XsxVLsi8(lDRVBMi)E3<v}B86(2Ld`5>dFrR5c8JN$sqKr9z
z%GxKqH?N9}h-G>J7LdlBmg(cSE_IfdZ)OjX>jc+u?(R#P;Xw6mnaF2{eXV`%ICWOb
z<Jb=E=t?1bnuIcU^Vz_J7(loX2Z#_70FgoxAWBF9L<?zv7$Nf!t$iNh9(=9bjh&BJ
zEOzE@grkpFuSmI9v5_5ipXMIN&8K?#X?{Z9fk<}PWu%KUcOx_h&RoEcIPjPqmJkv-
z_fBp}n*4B{^q-g5VSh`18TYLB{rIn}>|EBf!@f?wQf`=gT}niz*Ce=JRI)H`0&IDr
zWU*$1sW8A{Q5fKG6%IH&3I`maA^=B35r89AB;d#>5^$7?0vr`Z0ghJDfTN>mz%eQY
za7+{fIE9<SgDV1i<4P4DEA?1dy$KDe`>=-`2}eJ!Lfn@FE;4;FfJ#YaGIMQYq>KSb
zF7pFODGN3uOo{;?Mq<E+$uZ!=lo;^gQXKej5(hq9jsqXA#DR~H62M201n?1Z0{93e
z0eqyC1U`}^fsd4vz(*=c;G?7z@KGcMe3YC5K1xXeA1$SUk0xp0qvbU4(MlTl7%2mM
z49Nf=BWHk*Q8K`%U=%)JC-^GH0TiH6a)?_wRAl-lG;aWS0TpED&j0YWSXiJbob{oT
z@}c?fZ!Xv2@lws*Ip%KK$F4;>Y4?{?F+T)P6=cjwO_`g<%lOFgG<geh7qKvQ;dBd2
zzQ9UoZDTt^Y&SC6J|dj77e$9TxJVpFxlRZUoj7T->Jwk5u_2g}mj1s{=h5ySo?bGX
hpcro-h2OY<pz&h@-Q@mLot&q|%tIq*Oo>9`{sVAre8~U+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp.2D0A22E72C5CD4C0.idx b/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp.2D0A22E72C5CD4C0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cd5c1f4cafe6f13a45ff0fffd09ddd6f015d6024
GIT binary patch
literal 4066
zcmZvedt6L;AIIk$U8X6zmy(8a`JHphG$p31CZeLGl3AsROjMgLTd`Dj*JU?Utjn{H
zwRtdh5=D53*gQ*23?339%5~k!ZF9-uWl!_Gj=q09^Lowi_51w3@6YFTZr^i$@e{+t
z+pR^Su(XK_v*#41nutUqTmD~An4fuWfJh{_6p0pAHmYt-x#w^5UuD6Rnw?zU%^Y1w
z=%L2lIZIZ}Ii<OOZT-nrnuqgt=-#&7^3P-)-rO-*?4;h5ygBEm48wmET?=meDe892
zhF0R+5BVqTdjsElZoZYAvUtmmFUnq?T5`Lm?BneDd9(Hf292>vu9S*Xc1`cR8FBo4
z+q)|ll$kGknuFCH9bT8tUdazcV>f*>W6z~4Z_mCz^|WJt(W1%g4t<yV??meniJ4hx
zx`*qE56-RV*c#Pvr>N^6=_jsMybf;NsCs1e<xdAEMOBqw+<lUKCjEA5&AG1*bz~hL
z6Lot@*5oOH>g^T>#*xzoZA!0Rd^0dzT~gPQ9kw&t)^E-6xIf+VyS{XnbKCKSmjiZJ
zd|$r6V6wF!<G1})&Sg`!94q-Dvi4}<uq9@+-`JHy(+2)rvf%n9C%U&nqBU6=ko!8O
z;-r4t?Ip8Tjw@&$(9sigvEIQXqjGol!uZlPCw;12AOB#z$MJZeGW7n?MN?~<#`MNM
zkJcv*s*7~0GR>*IezpGf#jXxA$z}PbAVYIXuXUhh)-#8HPO|aHIIC{a7j}!U7>Xsq
z_az6H56(JuP+Mm1-!ORIHO1leL;fruXJ>KTcVG7G!J&@&_|W3SCf&`t#?ng9F=Dr(
z!mLZbQfuSKHw_HU*gRZGU*6!Zb@B~#e_Jizz5PSesQ?qpeba)t>x&%^9a`7E$#AV=
z$-u8sQCq9K_l1JxwdV^8FBYZT79IH43Vr^A@He(n?Yq0<e?QWa;M175?6;-!S9I7d
zJJ92)FJIvPDW^s3F}x;F7UW!=p)URUex_}uhg+_pT08oF^xMvdQ45MolQ*>3?OdVm
zz34mXJBf0pEWM*3=EBBZ9p@<7zP4elssxjP-NWx`wNI%0q$?iLQ|$IyjM_$d{<>~t
z`izL`gd9WC^Sd^O{nIa#RkQB(w47NH?(s*iYtNeqe9f7!dgg7Eby_$TcOJQ{xb`-6
zSGjd1etK2a9G4xlwnbakc72-H?e80XH_0M;%h^j2J6k_=zj|C(U2~w?s`2fnaBZet
z`1j3~j-x+>1P<A}H`+zDD||)Tg7coqv0HlmWWkx+ZvUt3#A)PPv%Yt6a_o@Z)yn5l
z={K~|&$`jOwXr=_&8{z2R-^Szvtm%=u`_A@em_&qzntbG?e(iaT=j`-$xVnK*;)7K
zV(Xryq6^a>uFZ`t{Mde6I%`^-{L`+Gkl#NHYAwmS@xr~>{L{?wN0&WnYy7LWaOU<;
z=D)0Qa~yCbXZD6#o2AD}RK-@_Av%}YVSdZD;ID2xcKkRvb?IDhpO}oj)F)iK-`t~M
zQTFU!zPHV!hgXMf^BZP~PgRN3_Gb1L#(}RGf0Dnpb}ow=J=GGA=gTf;F1E%&DBISp
zbe+HFNs}cW%a?7;Y%Gj}P?m=gcWUmre6$ix<I6Hr**N1Ml$GyaKCOPRsLm2k;LFZt
z&gRBJC~L(rJ-2QR&$h(B;LC1iZc5`Il%J#yJ0ctP)YlT9!j~P)94w53Q2uD6+TW3T
z@_{8D$Cqu*Y%Pp~P|lw-x4>1>fBtR~<9Sz{c)zvwPS7wbHK8QN2}i)INFoZTI3l56
zkxCMfWF!elGL8f!rKA8!MN)vI;wV7UN*a)KBn?P9js_%BB0!=@1V|Kz0LdsBKr)dG
zAelG@kgSpgBpb;Bl8s{l$tgKNa*-S$xi}7xa##7#&(DKeE|n8NQF2C5!V|~ekiN$a
z+z=I!+JDgr4Uq#>9#Pz{NEHc4QbPig%p(CwsVG2F8VZn99tB8RMFW!7(14`#Xh0$r
z0wmHPK%zVZNJhl~lF=}LWbzn5vMLsktcC?7o5uo@Q*nUgG#ntgJPwd@e|ea2(XG_-
z3G!H=;^k^V2|bQUjJ++~5TVR3_cgY#>}^okr{E5wF+Nf(cM&L|)qNz!78hu~u^KSq
z2-bh^2m?l(g}NOhZa}TWi7FiM^*C_|YCTRg15GwyB+Mrb7zy*qc8rAiqz)%xK3R{G
zFrU=pB+RE8Fbd{V28@FFR69n&d`gE?FrTW&DVR^`aSG<s4HymcX#+;Xe7YT@VLq+H
zX_!yf<21~t^*9alQ3HlxJ~Chk=A(8D!F;5{5zI&RID+{|k0Y4RG++$OXABqv^O<&x
zf%%LMXJ9^4k25fz(c=uvXB#jU=CcNjh52kd#=?A7hqExBt;boI&+2g&=5q}g2lF`t
z#=(599phj=r^7jz&(-4`%;)qt*PlPOv|IG<Rbe}3qV^?ylZ9@X?6)r0Nml+N-$X9{
zE@8qeF(Jl|d9*gzlwbC;@Uj-@*(bW9TJwf%74zLBCfxdagA|iMQDO=xT1*3l#0V%x
z%mBrTS)e#E_gR&`nBW$0q1=@}9vd4QC*egQ8~`pcey{s`9iC!M*Rn;4cPdx&UbhIh
zXu(sv@Th5Qwl<t!b{^#HB)m3`jfzh*ouKXFm+fhLnXqX&kr48DrnIe@Umj^P^79Af
z4{y^~=GM+-{IZv$msGe;!iIr=WA*J8gyZTG*RH_>_x4?E2in0~poIMcB{A+B1e(9b
zQi2c;msg13fRYj3Kq-iTenm=2;3E+Ue58y7K2kvfA0?%Lk3tmiQ8EhnC<O(4w3G%u
z8qvT<%V^-E6*TZ6DFQx(5bz-x0zRZbz{f}#;A0R2e2k0%K1RU+A1h^nk3}r-u`(9;
zSOp7woRkAT4spQ8$vEKS6ddr$Ir#|Tk>Ed4UUDCxeB}y3<<FUy+P7OQ$k!B1?VBX_
z4gdQk2@{cMM*nnr|1|mYbbva+BkLPyX0C%-%JdnDNy(y=&oz-IJ8fPDFQev89$;c>
zW-hj{v>Iq*J!qosgfL_)nK;hQ+1_EWi+?~+@R$%~sF$N>AgN%b|8Lc4h}+O%!wHIJ
gIFAu>Zy(>$e(s}OWh2KsIz>#HiHFCGjl`n=2If0FNdN!<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp.6662845BB6111F27.idx b/.cache/clangd/index/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp.6662845BB6111F27.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f50ada9877801be3408953cf6b517979d9918ae
GIT binary patch
literal 5316
zcmZvf3se(V9>z15R|qI?5d|Ze_v20xLO@Vcv0xA?L?Ma@RU<vNh(^GwrFC1}YVldM
zwe=LOwT<f=Tr@VVwH0l2qhKqoYg4tgDB^2<6<Z&5H)hYsf6v*R!#(re@B4mdE}7h!
z8JC-rb0Au!8Z|w4epN;7RDYF972*A>sjYtLgr7?FOo&P~zwt{;RY7NZc>71rf*sp<
z+u2$6g>THND?Nu|jxSxcy0Rd0$I1;Y8`hm~`f5%&=YQ|;dr{*)^xQp@a`M2I;Ahd4
zmi$ur%}jIt-RrlH2N(Qu?Rrz~!kPs=uWu~Ju2{cAmt!A(^RjKqo{Y={wcH+5|Guy;
z?RT^7?%ahvcMB#xdNib9!o2;<eyiTsxH9*hEkCc@QQPz9orQDi=4~i1UwaY%^3%-I
zD9>^xvgCsEVtG|b#k}ruvpzdtcWHfdMaA$JD)ZC2Ltl36es5iV+{cv-qt6;2Z?RM!
zJ@<Iig?rMQLAz6jU#V?izSXQKikopT=SJu0q1S%B+3(zlX!pfAcl*cFHxpZSU%b60
zS$JdghVMKRCw7coud404V#4XY({5ksc>iR0L+bm1ky{I^qu%WvTzTwb%kgy%?$FJm
z@Bq_^`MWYxBgG~4LG`qzEg`J-<h)J0cy-~>QqQeV`)V7kh4iUD+R3<n?!hI;GVg3D
zIz0T3!dDn8UNV8I8Sj^N`g&ez-He_=;k!b!{*iF0bMT{}4R7IjOY>-T%-&0KlQknF
z%YJ7g3R|Cr6J3+$?CSni#l|k`^H7uNOgza^tHXyy{Uj_Vie^axd*3Ub>^|(NJNDVB
zFTehH^nUvJF_C<>esb7fJ#lx&XXRc=e0j(xopY8wXw7gNCpV624lmuedd;~BbLzxx
z`+n^IMM`sL`*SxZd4_i_Slqn3B{bptjhD~(?PE9ZX>_ixI`VMF!Ssmco%-T@@oeJH
zUmK3hY8ZL-(b6HGY&iN=rM`a5vF$l|mr=~F%3F(v;rj8;&f<vT?-w_|v@q=cnzr+a
zvDtf{6Q;$E*t<M(Q|n2~@shiZ9VuJ3f4Kfr`rd$@`(g&xzk08^YuczEaxd-`ia*#$
zO^Mo?HPT|DE80)5EZlbX-5<_>lW3dMzIAkdyX*2NYa><^xqsGo-B{&6Xl-X#iLvxR
zlyvXF+-<v>ZyYX))20W!bE&xPbYuNn$!&Y2f4A)EJdxB;N+-Q8=V|R@TUt8)HoJV*
zsk7hiA2wq^!}c-#V+Pe&-o5p(duDCxX?Js0aTe85dGJs`Ps3k2TG#f|JpO`PJ~O!`
zqv7V!fPLRBeQ@`n&C$N>V*9_=XoFK{zq*t7=<T-Y$)R1-PhI^m|3S^#-0A}j%j@-1
zuOFM25H;|@a?Kdm!<at9tG3&FmZuK7ugUIc>hp-O#Pr;0y4vY?V%`T`b)`#wKQr*r
zS8u;k_gzz6Psx>oIsXnbXom`~Zf`rnO_cq=yub5a|0%DHNjb9k#Xw^~_x=v?K~k%K
z+mvaI-%iPW9I*LNkWjOrYT6&(-*%w(vz4PhP8%G8##>ZnQGrpxzQODGKAH!c&M(SK
z9Up?mdF{BsxCq~%*bz+)1Lp3&dN>3XdhPJQ@L=Dd*v5yqZ*1*a&=i7-ymoA0Y>;nI
z>_wB(Rqi(rEDco^d+p?a<SgHy*sHYHf6QC_+UyWC!E46{#v6TuVqcv)_;akkkrILm
zymoY8bg*ww>_5XTe?K?-NLL6ld+mt8h+yBK*wqy;);#LlyZ!`?Z`~_eA8y)tA#-qO
zna-%oQqI7u%(|jpMH)%qk!BKjl#v1+Wu|~f8)@LtW*T^mkpUiKW`M^US>Ul|7I>VI
z10HAQfX5qo;PGZ2c!E&?o?sS$CmKcIiDnUal2HPlWR`%3jTm^?jDaT`W#GwX8F=~u
z`h=$!Lz$k^vx-u#KT+em?!eb$(M|3BLUdN$ETC+<x?V+ENRTA0BuJ7r5+o@L1(KAN
z0!hk7fh29AL6Wx8AW7S3kYp?jNHSIiBpDk6lB|UVN!H4OBx_?qlCy9i$yqs&<ZK*B
z@)jN>c`FZ+yp0D*!6Ja9U==`8un8b3T11c(ts+Q@HW4Hxiv*IARRT%LCV?cjV35RC
z43gM}K~lEJASqjAkd$pQNb1w|qm;`iv`n9^F9d3wzD!Zd9!6+<dzW&FCAkez<tRFp
z@OIVuI3=xYX$2bJR;HxAJA{+aeNOaVb~-nqI}Wvn(ES9p9qG0pkar;6eyCkYcNA#S
zNsy3Fx(O2UNe@9nK50iJ<dY6WLO$t2B;-?0f`WX?O;C_ec?b&fDLbMdpK>4y@+lXh
zAfI*;H00B6f`)wBL(q^<+Yt@<v;)zQPrDEe`HYibAfIs)4CFH&f`NR-ju^;i9EgE@
z#)TNjXPpEK`K+5@A)oaSEabCx#6mvnKrH05F2q7U=Oj4D=iCGb`J9K~AfK}%4)Qq%
z;vk=MArA6+C&5EL?<RQ2=RE`u`Me$Rkk3015Ba<c@sKY#2?6p2Hz7d2;2{La7wkxY
ze8GVP$QN8lfPB$Oh>$P32@&!|4<SOnXh$OCiw-10zUV?C<V#LMf_%wMNRTgi2nq5f
zJCYz@av%xvB^Qz)A3F&Q`PfZh$j2T6Lq4`64EflBFyvzw!jLaJ2^sQbHz7m5>>*^x
zm+eS~eA$6y$d_G6?&XhMa$R-*w^|S3UzSd$PgZur=hN2A?uVDZ<J~0V(&I)cck80D
zk!<(sVFBLhl;D&wB|WY6T3&~3;HEy_-BPWK@7?=IH3<}@rhuZ=G*FD10g6?#Kyhjg
zC|=D2C8!0UM70Q%q?UlfY7CUDmY*tR8QvSklV??B>tAv!JAUZN<paEn4G$0Rr`%4I
z)gg`VsZ~jb{?Gq!$F)TZN-i|6^mgM%$LA}Zo%6p4D6HC?<DHJ}8{1F0=N#0Jn;ww8
z`I>h+ijKm{qkC)7h$}B>|9R9qo#vmWQ%>_zQowIPVX@1+(<w12T4h>!Y!1`-9+!L1
zOJ~yMDJKwGMw&>VC=&$~ZK8o<Obk%0i3N%?aX|4V9w@;i04174pd^z76gFX?WRnb(
zUaNn);%n-it58PsldCbBj!sfms=NuLYkY42V8vRUu6M;cM#lq%bxA-ObVGX;sU<;=
zWJu5>F$sF4fdoBDOMxE6P@qR)3iK!g1$wlW20fagL662X=+Oom^cXDzdJMyW9)lUs
zV+;)Fv04`NScU~X7PFwo8d%Wdv>fPh3<r7~=0J}#aG=L)dC=n-9`tz3gC1|-K~K;M
zpeHZ_=m}T=J;5M=o~RW;Ph>>U6R`+-qCo^bNh^V##7LkgVF~mkg9Lh5i$M=F81yj4
zpoa|@^kl6JdNLz}o{VMClMOQH>3O}R{M@}yud;p!P$_zYqN*!qPJQw|TvKhW8UN&<
z<jL^g&m`qQr7G<`oY;Ff`RSowS<~&M;e+z#Y68nAmX=JKtSW!%sjOAgXU_1-yceJM
z^A89NQU`~G_6ZN`n;VfmiiyzVW<|zEMfZ<OADTIA_z2^(a*UWk8aVC$o9#C+K4Ea8
ij-pv!ke*2zl9HM>NIw84KNr()^q3b=&e)MABL6?QxE<^O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp.3B9F48B295268F74.idx b/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp.3B9F48B295268F74.idx
new file mode 100755
index 0000000000000000000000000000000000000000..35c931eb6297a5753378d0385ed591b5c8e944c7
GIT binary patch
literal 7482
zcmZvg33L<Jp@uaVu!Ze{O^l7hA{J?8G@E9h0b_$OhFCn7AOpfI4h98ipnzBcfwXA@
zA!*};0$B<P5Z2hxQ<{KDWrY@Kn#PV&2qXj&u>?pUga>JPuVs5w_?@x;ch1u}zMgOX
zU!Tso(rB*k*otAp)SfO^*^G*%bEAu<wRgE(dFj7}i(<2nwR5?~WV>8T-*|ugq|q1b
z{Qus3YS*zk|6U>d!s>>jJ@>V%%<0#w-^tR7Jp-0}zN9Mt$Dh8u?zvYF4;1#S8T(=B
z=D*!rtGY7I{rxNZt-(ttPR+k{{A9mF=VG@fELmK<X~iF#=y%VbqR!v_V3|4SJ$d8i
z9-iD!Ms;=v$|s*K^&Y-<>9dR9Zr*q6)&*|k*|W)aez<$P=8=`coh2oI`{wM2o!@P|
zvvJOfvh@d^nRk44=H!2TI_|H{z1O>+`rwHJ!!~#1Kc+ALX~*s#7ykM1V+}*+9lKG!
zt@i8Fq2ug|%=>rx*G^q}-YV|+B$Lx?m9p>CYgu!*_pI8nP_Fqoy?xt1KiT!qtb6pU
zrwgv!$S?6-SDrlB<%Nn#J^O6=X!g9G!sYFA+s9X%vt|~*bNBC;uJ-pHsoL2t>%`kD
zYpZ_JXJ*BvUhi(}$GM;A8SR+0@Yw6W?Y3(FaDU%JSv&Hk4*fcN(orM7{oC*Fp1h)9
z_2?PXPju}v{P}~wK6dM9*~OmA!dH%HD<gBq?W}s~cP|Qa+8rLySU35hE$bS~UTko`
za%26=J%8}r>@@!H6L0PFR(|^EzqVFf`QqWqE4l2AHBabSWB2TrH+1VfKc>Xrt6MPQ
zz{R3f-lF!)XH3{7KO^ng_RNxfRqXyRSG)djbSJYk=kf$^1+RwI6&=j1xgg%E+p_iM
zn%VwsN2aZQ?$c#WgWh;^$f7k>`**B8GhxAUzOHsk)q#e+1!w5%1KwROJl|v5v5Wnm
zn)M5IZ{~)<&+UvAj9;90ecP?ChyA(UtW<g~sIBR^`}Mf5!|baQep&P;`JdF5S1N~$
zIxx0D8guM+!MxgCBX7oUY?anM|ChDewdJn`KHF51HIW|a?YsE=^qG0>xVtNUS~B+}
zUlp!z(AI6=KK0pE3t|;x=9LI5Cy!mccEPX-1E%&+=I4ADSzFMh;x`LxnabHeJGAlK
z@m#z3(usZz=Z+jdbnKa7+@<S%FJ0@LG5Y(b|FLJYcxc3?r_S8^__HtmYue7a&;Nq_
z^;hxp*}+3YUg;VccYorkJ16F^;zi}lUD44^xx)vZo%G9S`JjzmKO6PM;LnW<mCL^>
zo_Of`v+EA!RnKZ3F=fO@e|fEMmig*CPmeym`1EhBnisSmZ@sbe_RL%BfAjKbYyO<?
zayEv~maaczlz;S}sSWoYJ8|Mv=8R1GwQa|yy)k^-sEe~sob6uw#p-<*^LONZd;M;g
z*dMmf95Z%PsdV}1sfJeve7p2&<=c-oq`o@5dCR!)uKV?aOM~|cw3-VW{+0K1=Ka#O
zXExouI_eC6?&RhB(~?Z*wP#K**|a{fW^HNhC+%O92kz{6p{98I=IJ-T{9$FUQLk_B
zGjf>raPjxmj~m@ayB02+TVql#SBB@km&)F<dS{by5m$9?MsCNJA${$)JhE<YIrHoY
zBb-et({cBV?s+YP6ZhUbZ_N7n@9fVeqtkIwM$wO21}9b<J}UXshu80AlQHSIFrzTL
zWpLs_cYe6`_BYFZmrcf}<NS>L%$C84qn}4c4svxV$|jGd<8B$<vReix-t^>@ho5Tt
za$7bTm5vKC3bI=UCq5NhBucJTOLJTk((%9!1It?mCq8>{#Muwt>{OmjCZ^+F8NG(I
z3{L#rwBB)1zouoAa5~P*$m`fLIB_ic_(HMJ<!TpOUflo4H-o60>Z07DLT5UhNfonN
zCbRbNoh^IL7x&Jo_J+J8omWm%*gL+Bd?6oXzOWB6Iz&T8hiS-|5Ca(#W*}ojEM#n$
zh0GuFL*@_rA>%?EWL%hoj1TdU@nIe^AtXR1gaydNkO-L=79o>D5@b?Xf-DdUKo$rG
zAd^EfWO7)BObIEFDPaXNHKanOhE>S4kOr9+)*#bEI%IlShs+2WkQreEGBadCW`<43
ztdIqn6}BL=LpEe~*oLfBEj3y{<kf@JL9OImv}IW>7b|@HBHrn3A2@H+J0B?KeY%Z&
z5g#;tQ6DsYF&{MP2n|g-N<)*5(a>Zf3^bW215GBzK$DHI&}5@5G}#ynO@G7>O@Gu6
zO@GV}O)kPglZ$fD<YF8&`3Mh9KFULrkMYnHA_6ppr~pkNCO}h+h|m<HA~eOA2u&#>
zK~suK(3D~lGy{<UGy~B9Gy|~!H06j4O*tw<Q;x~dR3ZvAm8b$uC8j`Aji}I6qbfAj
zm<mlTqCr!OYS7eT8Z`BY4oy9(LsO6G&@>_jG>xbMO(SMN(~OwVG@~Xo&6o*ID`G*@
zidxXLViq**hz(6UYD3eG+0ZOg%SzfV(xK{5pmMdmm7F_AK~~ER6fRlaTMU<_?gOHA
z8i>&uAXaCA{JI~A(>Wks=Ya%W01|Z(NYW*sfF1ynbs0#}6(ChtfiztM(sdok&<!9{
zH-RkO0<v|xmE<P#p{XR_NZ*A^dY359d^@X;^qf9edcd60ULNL@_5sl}4aCq45KFT_
ze%cSj(Hsy@^FRVE0Ex5+B+(L3fDQo3v<#%s3Xn>xKpL$9>9h`H&<2o6n?M$A0ok<O
zO3vEsXIU+4GMKys<$a%On?7%n@_vGxqP%C2>xj3WKtE2rjmS;J+YIbWP(Jkgl9Uhq
zz7*v{zpsw?(C>>AANqYw#D{)5LDA@^lN619Iz`dwr|SrfemYKQ^wUj*Mn99F81yqq
zia|e<q8Riub%a4b6DJJ%nI^)ZpG{CK`q?DKqMuDsEc)3x!lIvz6Bhk!6JgQsPf&jJ
z`;(L({r(i?N58+0_|fl=6F>U>O~j9WE<thV=aLkMelA6E=;!JPhkh<jIP`N(ghM}{
zpm_B2Ns325pQ3p5^L2zrKOZMN`uQfpqhCl+0{Vp{C7@qOQ3CpfIwGK7h!X+*LK6|t
zFD57v{bG_5(J!Vb5&dEv5z#NkiHLr&iHPWz5|o5~DM?A_mr|63eyNT~=$GO|Lci2R
zB=iRoQ~>>fBo#n^AVmexAE+Y%^atW3fc`)e37}t2P%`@EBqgI?PEj)Y<vJpxUyc(Q
z{c;nL(XS*Z1^r5rQqZrYC<XmWJyCYU_m|R0lmv1!Q9cG%6O@X6HA$)HS5uUVezl&c
z=vNzwihi}3sOZ-cl!ksSNonZUQj~^%t)6J;*BXh2eyy2k=+_gJj($B!>FC!}l#YJA
zp6KY;8;Op7y_x9fHxiVAej`a4=r>Z7fqtW&80a?|iGhBjnHcCd6O@U5GfA1~H&c{}
zezTsK=r<dQiGH)0nCQ0>l!bmPNm=N(Qj~>$tDacsw;G9seyf>S=(iJ;jea{x+32@Z
zl#PD7p4jNO8;Om6yP4Q+{^08GUAJy7I!(2&ex%<c)y`6Cq_uWlf%xK@^a`uH-M!5D
zfmu1K3v>C`13Nrecy%o4?6hkRUmbCLe)lb1(_0*FZ;!Tnzt8OhqTMtQ<7R+ZHw)x<
z`++z&2gJL1Ai*sFiEa@{a!Wt~cK}Fs%Rq`-0aD#6kmlBabhi#<xD6oFZ30<t3&?ic
ztrQ&@eE;Xa-&vE9Q(cf-P~hxq<y05u78W>**qrKax!nq!EwP;H!Rlb!neUwHpc(`k
zq7G>#=Naim@Pi#RSjUN872X-q2SkfB5F;``tjGfSML!THazMPu0|}x4B#I)CBuYR5
zF#sfsGLRxFK&q$$X`%+Ci#m`Y8bGFK0$HL3WQ%qyId4UJ@*eCj!sL0pUYI<O4~X{A
zK#Ye0Vm&O7-{S}3JRA`3;eiB?03>=uAju;E1v~*D*&_of9tBACs6d)W1JXS@kl`_a
zOpgg<c`P8?W4F@otIsZ*c>ax-(q9g}LcK;i)3s-5V}~(ww+>66&hM08;QY=wsE?h|
zp?vGr^yyq*uITKb9<D6CQtdhUVfys&_QU%+pLpj#mDRGT>%1|}Ymls#&0Obn&j0x?
zyCiqGnc2DUS7~KQ*AkDTa&~+Rvs!j~+n(zUdWSmVoN8at2Sf*HASTEFu|XEdAM^up
zK@Nxy@<2jR01|^DkQ9`F0>J=~9F&2SpaP@@RUj>>0qH><$OsxhX3zw(f)<b+w1G<0
zlGgpltL{B7lL31l+%58ad9pLJu2^)=v<Exr3uDt4h94YEdNBO&W3qGLYFC{TZtiop
zlfEJ2v8pMPCQo%uY0X?!ZHH6a4j*efoIZEPoLOlc@%YGg?K@;-x;tj){3y3`r;5Dt
zGCD7-Vn~<5hw{4>+5L+LmXr<}EOq4v^fOt{|97^adyk&I`}Fnt7(XWjWJNQqUi!nL
TI<#xS@X$0eY;5om>OcMsNHE)Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp.0C31B57B423E19C5.idx b/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp.0C31B57B423E19C5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..446385c8b746e88adfe0a4d2a2874e176bc500cf
GIT binary patch
literal 7482
zcmZvgd3aN0qQ-MRpoF$Si(M#%hC*`IoFyk`6IviFg{D%L07eiiLI52R3=Fb4Gh9VP
zWf_Fwj)Ux53XX_^*ru{Lg0ut~lu_XZM+If8s4%kWoe90qY2QC?o~Pe)etF;g`aIvs
zIp@navb?NJ?dEU{m{dMzT4dIQ)((dwJM}knR`l^xtsIWw84kyswRK~X!@jYyZuXry
zY-fy{(RJLC#rsZ<o0T!jD-`woY@K^dx875Fi**P7_elAGiV3a`l|xp)`selE-R&81
zWFEZm-qZaT{;(&`{<Qtg!T)vW!i_htUY=cY`**8{PX6HSR!44U%rpA#SN^iTt4ldi
z*|D=c`r$A8xQ{emIDP)R^*4UHesr{P;J{xZ-`xJ!hbuZi^mF69b%_JZU6Y@=X?}6*
z1#3F><+e{c?EiM_&J}9{Z1#<pK6`3upZmuA_t29ID;_*8y}0PmlS|HR`s(Gf!VmYZ
z-ZtV^hub$Ry43w}c>3%v-&b}nNuPLe%h}vBL#9t%ed5rIZC+b3<l5`QKF+-})Vo*g
z^kYtG?ls>t?{_L6GOk<khGUnKy~0^rr?swGY)qZfd*3fVUAS_;`{>w>tvY<V<N2jy
zpDLbm?EJdjJv-$-zjxW5){EYodVI{%!)G?<{PEF6r<`(+V~fO#U4rD(bEBm`J3nB4
z@?B0Z?=kSpk1j5_Sp3qmO|cur?Tg2Zdo{M^8-d9<`-j~pp5i}ieeuTkS)0ZFo|4lO
zkN(oO?zz=JjAg}ks|Ti6X|b;37W3VRrzpC#EKt8Uuk6)-QkBJJ`KJEV?LYO{yvn`!
z(dCVoSEZd?^xIV@mpz!Vb@sL5e$qRa3W{3Y$$k9POKVS^f8pA~&*xM;&F|Pza(-;}
z#TlP(+&F99<&IOfXLZ{$V%mF?ZjL^8j(=j}nt!|@?ppLNS97AVD5J;Vur_yD(QDn)
zuP&rlY>vIYX!Q6acb|NJfA(9|?SK2g(T6vO5A)j|FE2lHWOcM`)WHLV@2njA(WNmB
zM(ZcWe);tq*Kc&Hlj1uXwrqMU=YhNbbGl;hJA;cGCM_AYg<Ej7<gtEB*MI!x_Qtt`
zrtqg{Pg(gV?KiuXN~zuDy2XoDR9vo%p12r1`f|^-Wr1!Ld;0US=wE+)o^O=r+_zvX
znX-ND<iF)DnA6MZax%E}&iE}gzaIMMlDg^Z|5fXmbL^Ajz1P3jrS|^g@3bE=oXj0z
zE+}Rn+Odc1yxg~N{`!Ky&e-aid^dizykO;v-_0NV)~-E=rMEYo;R-*hnY?!1^7)K#
z>)9Eq^>xOyxv}4HU)&lw@X(Qw`#vt8Us-#-%=>iNfShyMzjv+YuaEe5+OnaqPnh}I
z*~O=7x3OQ;jeFbm$_u^j{^9$B=c8J$<JYHt8tc0B&iu{UcUP=lyGQ%_m)EQMOKWZv
zt{Sy{xzH<fU*lHEI+vFF#HvQ^-5;3nY3<U1Csqto4t)5^=lkj3{q5YL1Is79JLT7T
zD{pqWrQiAeoa@J4-g0walJUq}whY_sr+Q{CfBb{}pG%j%>-y{Ll8r9rfm18#?FDai
z`&phmu=B8)bEiFKP!30$YyYwV8_n(+WOUe3m6ev&u4zbJ`%SO(I~&eDSD}Y8$f#6Y
zm{yqGG}v+X-P4D!nz3hZ1{s!$3(^YOHw|{I?mJd`@aVN)Gsy5%oS&AT(KOg`-@89H
z?)Y}@n;B$GD$YsEY1=f|apjUVzf3**?7>XOgQ>WrP0669!Hy&U2oLS+=v0tFMx^4-
zX`M5g20LE&%=j+z>d$V@AeE^&FD);lX|UtNYkrQktvoy-gFKXqi_(hvHw|`tc|!LZ
zQET*PkWeblPRnlBG}v)8GI^$$?{Kt=&hLHy2jBLkGOG%*3i9pguqQP?y=gL+9=^L_
z*H^u}XI8mG?lJa%&aCo;JV11a24X@C5F28FxDW@#hj<_%Bmj9sULY|f0!bkWNDj$B
zN=O0ng?vD2NConT{6Jbr1JXk}kP$L~%#aCWg)E@HYG2(x`SurZKzh>`52iBgp4c*#
z?udH^P}Kcw3wgpGXnG<ZXnLX^XwqRCnskJQCLN`r$%GkbG7$!vOq78p8)l)&Mp$UF
zQ5Kq9n1d!4;h@PyIcV}>9-4fFhbAB8p(%s~XbKSlnnF~7rZ?<`rZ?h+rZ?(^rWh8X
zDMmzSict}oQdokf6p^4QMI~s;VHuioM24mum7%GG6=*6E1)54!fu=9)gQhRygQhR)
zgQgl*p{YhxXsS^an*OjKn*N9%n*OLCnp#+crWVnlsYNwt>R}z4dPIk&9@U{~gbipK
z5d)e=)PSZLHlb-oOlX=>6Pi}of~FO*plL-dXbw;Zl-ieNX4OD-AkZLnP&3(g!MyaQ
zhXGu&n!7h#mYN5M)@UF`V}MwV1>!Ugh}U=^K@)(yniojaL?B6%fMiVuQZxm~r}=<X
zO$G96ejrWLfOJg<GBg9o)Jz~tvzke%H|~3c<mjm<N@@2}pFQ99lQcD__a37#r?gvu
zIi)>7G))6BGy}xaED%R?Ks?O@3A6y@rM*BREdoij1SHckkU}d!KH3MQ(khUj_5*3O
z2BgzEkU<+jCT#*)wAD=ZI^*~0P3sGoyg22qqgtlVouJ$&kdu`A9CD1fs|oaLh`ScK
zp12!;J#or|eoun(px=|EJm~kthzI?i8sb5}r=EDwPsb@5{d9t&(N8BS8vS&P(CDXY
z2#tQap3vxL;uM2^CP6XiXOa|yekMj3^fNVtK|fPZ81%DoibX%0pjh;?Ns2{38zU_G
z*&4#4pRFe>`nfp8p`S}o9QwH=#i5^z5f1%a4dKww)e{c=e4OIZ&nGAz{d|(*(a*;S
zkAA*}@aX6336FjuP6_B25|n^`AxR177h*&}zfeO2^b7SwK)*LmdC~7pP+s(Vlav?z
z-Wc(s-&;ex==atWFZ#teC8A$UP$K%pBqgF>j1dw2Vhs_|FV+(g{ZgEg&@Uw@3H?%%
zlF%>3h=hKrhDhj_>WPGYIZnywmlKqXemP0W=$B(eM!#G`Wc17RL`J_7rxf%n2}(h~
zlB5*$D>0&=U#TGq`jvX3px+m#eCYQjC?EQLNy>+QUp4XVg!`ATmiXex4a9dGSdCLE
z`qc!bqF+r?D*DxGqM~1|B`W&W2BM<hAE*52_a`Vn`u$1DkA8nO@uS~gOZ@2fHxNJi
zwK%1rUrSIL`n4pbp<k;e8v3<bqM=`FAR7AhIHjXsPf$Ag^(3XEU#})Q`t@3(qhD_z
zI{J+`WuV_kPzL&qBxRuAs3r#bjap)$-)JBP`pr0HqTftVCi=}JWuo7#CMNpLT4JK#
zY#=83tvF?&-%3yx`mH2oq2H<|7W%DPVxiw^AXbY%u=uLu_KjI*sMb|OdJd_wmr_H`
zmDBUYS5~A}ScO*M0Q(1~9aHH|yA1XsVN_)&=Heeq+T2@@wJYstw;nrurQ(wrg&T5H
zzv-Rst}ROfkJAH0J82-s$pEoV7Kn3lK)jO&5}X2%*Xackog$Frlz?QX45T;}AfM9*
zq&ig~ztaz-IW-{NsRJ2K1ITomK$g>LrpVAh-NLuFJ(iYPm6w&5XYb8qR^?~q=h<u9
z%&N{=o%8H%oy@9!YCp@K@64)z8UX6A_HQQp6REZFy<H&t%%#qZ=vLv(h#nwXq=6Wb
z0b)fKh!Z&=UgUuUQ2_FaULa8vfh17^l0_Lv5fvbx=mSzk70568fizJA(nTG}5Dg$x
zG=VJ9Y9{+sq$cm)9vMuY%k75Ab9sPh7Y)R?7$DZg0&y-5h<EWof=dAMy1YQ5O9YZ!
z5|HeYffSbl<a7CeRF?|mclm)dmj<M}bRffJ0GTcm$Z}cDwDZbyb07L@?XuL(p(t21
z%$}}Yb86cRpSGzi^>$8&oILyI$NSpINo@vgx{`W3%abMAn}ml)^|@5#`s`@x?ZK@F
z_prb5wv41VZA`YD7`&HVP}<yR+cE#;)Mra`OI`M7?cGZ;y=m97<+JX9d!Q{&T^~XB
z&=&FpJdk;U9?0kb4H+GzA!7mzWK58Oj191mu|XCxF2F&?1v$w001p`-<RKFR0%Sr^
zfXo~4Lgo#6Ark{4WMWW+ObSSlNkIuRIUqwO2W7~VfC8BkR3P&Oe31EqKFHL73Yi*I
zA@c|Pkokjt$h3e4nHJO_(*rtWdQgYV2pEtVK?5=~U_xdFO~|Z(1(_AJAS+c%n|En5
ztNN&Yn(2yj*K*Qp_q``<wkKP$zm1M)<jD#5cF<==C(aDrI~aFw_`lb~_JN~S)gy%u
zZutAB->gh~bnN(X4?p4<-<&zdwj4gva`<S=;lyc^o}8M(ipfJ;wQiHv*4Zv2vwc>_
z4&~W{2GH5*<^4P5-<Q+5z`DP8Nok+H{iIx>S5Jd={eSP~6?W~`y|{<l!*IM;R(yWl
XEYiA&>cHH*!NCcnY-C^v^}v4sXkqP%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp.8439FC102CF881D2.idx b/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp.8439FC102CF881D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fb5dbff092b9148742e6fdabe20a68fff520e588
GIT binary patch
literal 7482
zcmZvgeRND`;>YJcBqlFNhy=l;GWT`vy>suKd*@yxVY~!mz0{~qX@oMxZz&xqx=XvN
zUb?N9Zn1mzXp5+}<E5;ZS~H;=-J>-f1j{bJ;kTtU#j39AkCxp@_Du5qV{=ZPGoSl>
z=5tP-n>%^(j2&KCsr9tk22C5jaCWF>N?V)FmY@2bR}-Gu(8gvPlVh_j+*mh$;^-?@
z!M%6?_I|^M{BAMu+KL0GdhTm8F1NH->G_J`JNnK)H-EyRBQLFaYx#!yGI7URq58!2
zc}KS9{)#$SW{%?WCT%Dh^62jO!<2&?AFZ!Su4XTP^oP%9Eu25VdinWx6CHbfag9CJ
z?$y0#+SHH8KmT#Z*XI86MdSW`w&zQa4j#Yu!-BoD%8nfF+Uc?Q;i{8+@BeV;=*@pT
zVu~;4eI@<h#hMG$`B+Vt?=(C8%A`9Nb~a3}&)%EWcWz_kLdG8DcT&SdeNn@V+Jgs9
z>)Dt8^skpj)fP?q|HP>JWY*H5Q*w9SJv6Ym@zwE-J>Cdz%4+<k?w>QOH(&mA?8V+y
z?`(PhYWcMjPxe-4ZZB=C_j!Bg(v0N3RReCH%Dgpe=!WXUi^^9_3&v{(Ea|>}_v3Nt
zs`%mmPMu#fBxB0XTG2N}?!LE6?e2Tr5NB|+UjO>AMN@JPEq}$ff5H0KPk#CBYqQ=s
zH1JD%gQ@j({9<LSbdIoi;flQNJCBcx-&z0Wve^TI)6MliCpF;(D;o+ocY5c$J*AJ9
zUiy=Np|ozptNugT)jHpE;!>d}$><ecRq6Zn{=&-6SEx~PWszw-*SMjqcCB;i<TW?H
zDabsv?1h3;tH<ZOH~)HGNzsM7ZP!o_f9;$(q;jTq^oIenVs!s5<JR@kq)?4@^VFEW
zqwlZS8(7(4?DhlWPX8!8H;3N8XT*(jh3D!oUaguwe@(`lrQ+|)hISm8mvL`E=g&@!
zJ5e8zzg(<W$5saavRv3!y6bUqN|(ULTT5m?^sBWKKRHso?>GL>8pj_$mmy5J{N>I&
zH7h3-O*$Lu)DR4BOD?!S_ExfXZRqsi-eZd@<`=4)>Q7!je`bE&7URAD>E{{l|HcS*
zWq;fD`*TMJenvmRcAH-F-jzS+w>h?HH9hmOH`nz!zi{CCdquIViuOzP21b=XlT%UF
zvq$?GTgN_<$6k6g`P)w~nI|9otyApkhu7ae`ShwP$FmofbKI|D9~2+XD-g#$wd9HK
z-(6ofy8V*}Pj}k8c*4D{xzB(4&${<M*znYzefw^|+J0J<XVm^<b#Ha~slw{k<Ib`J
zv9tTv&0557nLL%hxcKz0GtbrKkJ$Kl=Vj(!`p4h@V$s>dOQr^<|8?Vodz<g>;{Ee)
zpS#(vcH4=Hl77k`d#vhkrL6l4y@%F5O3wJSVk`YVx5|3?`fu09>4%y0SB+O*>G1aK
zpMN-S^~a-@bq`%kNIzvwFkX1@QTEl&XXmJ;-d(Rb8WwzeyR`Av@FV}YT1Q@5^~iB<
z(RU?{J7-@X{&^1Hc;@S$ePTa4_4kT=V9<j;l}CDx{`y$fTl&_|XH-sI@i6Dv{#{4g
z=Dj%k882nCWjYS79<<pk$stb!ZPj_1d0EXv>e+Ap%6_ov($Wz|Acu@gO&4bt=Qj`O
z>5|=Z#;lwBhy6KZbZWX=X1DguLwZ^}@Ok+s$8JB&A!Ab0MVUo8%|m*+|BrWWeQ@o?
zk8;TP)O0~+LA&N5Jw0*H@`6RY(vU-*N=^65>@~1?NKc2(1xNO`b?TNw9#2ho&Fq@f
zJfx@JSy0ve7soHw=8#dT>B7vyoaP}t{mqn;Ly~^WlS2Zj>HN(6tmYv-{nM_q`?ITx
zx8>TNOih<%lnrSf($nG4v-6}Po2^awm;L%2z1E+~t?riBtth=5=|vrs-MpBq>VMp{
z<L~`Sa;u#_=iu~@l_I}$Bv8OPzLjVn4JPfU!K4E;n2e7BlkqcPG64om*2jX$`dKj9
z01GDP<G|$n9GF~y1C#ghVDf$*Og_MaDfk311-}5M5D>t0`CMSS{4OwE0T-B}PXtr+
zi(rZY5lqP^fhqYVFr|P5rtFi!l>IW8azF;t?Q?_a_PfD!2i#yPJ_St0uYjop6fjkv
z3a08;!Bhh(n3_)mQ}b(JY5@&QkIw_9$L|5t6Yzkk`*bjMzYeAz(7`l(2AGE50MiH<
zV0wLCFui^+nBIUFOw(tAY5Gkt&43A}<+H%F{1%v2zyh;eD>qsm=IRQq0;s>%zlG8t
zbg4(S=~E6K*`RZ3>tpE*Ip+d}or_zE4$@%KAsS3NOoPb;88Dd;111w@z+{6gm~4mz
zlMS<AazPGEF2sS!g*h<!AP*)V;=$y@JeWdI08<DFU<zRYOjpnarYqzE(-n4sDF#I_
z#gGW57#6{lf)bcgNCHy|OJK@D8B94OgDHn)Fx^2nnC_4pOn2A~rV>=ZR6+`vN>~9?
z4XR+OAr(wDtb(ZpH88c12BsF)!1M$?V0uCxFg;-pn0inLQxEB2>R}yBBWQqWgbXl^
zumPqw=mpao@`C9Nd%-k=CYWZ(1k((gU|K;7Oe<u8X@xB?2Wf-KTOZlM+F+m|+K?7X
z|Ftd5Zr*gjBda_6!DFe@K#a}+u{sOH=^PNR^FV?w0J(G*kf@75k}d(sx(wvj-9U=2
z0I9kPr0E)vNB02fx(;OM29Q_x0-3rAWa(B5DaXB?Cy@dpwR0&iDOc0$o!+sf*0gE6
z25ZVVU9hGM4a6`E5X-PY9K!+e3=bqQ0+5Sw0f~$VBry_@%*a4)#toz}3XsaEKpLX~
zc^D6n&geh}V*q&>FObQYKo(=QP<lW6a(44R78Wl?IqRs_<#Q$|=LzH_<-CL(A<pdt
z{3vn8kmJPp1uz|>XzbGoipD;jq-gBZ5kh01juIOCbez!GXJQnCeI`LM*k_UygMB7K
z80<4q!eF0?69)TijAF6RCMXvBY?5NJ&qfG~eKty1?6YygVxNmq9QL^c#bKXIQXKZV
z2;s2LMG1#}E>1Y?^D&CYKA)g??DI*A$37n+Jofo0;jz!h36FgtMhVy#5|n^_AxR0?
z7a~N!z7Qn>_JueRu<wddF6_G!lneW=B;~@sD?(h@cSVT{`>r@~VPA|<BKE}uC1PJp
zQX=-n2obR_Mu~`hF-}D6OEF5qzLcOO>`O^X!oCzC685Dik+3htiG+PQM#<Qh6O@d7
zIZ4Ubmm@^Rz8obo_T@N{vG0yiZtS}glpFi*B<04wJ3`#pcSngE`|db#V_%6;3ig!*
zrC?u4QVRB!2vM-FM2UiZB~BFVt1(K&zM7y^?5jyi#l9LLD)!YVQL(SaiHdzKMrqjB
z5|oC0ElFwE*CIs2z7{1K_O&?Cu<wac9_)J(ln48sB;~=rCqg{f_e6;Y`<^)QU|)|>
zI`;JhrDI=DQabka2+^^xM~RMoJx+A&8!^hjzLB5|>>EkSz`hY72KJ38F|cpMiGh7@
zjPhdNo1nbd_a-SX_Pr6}#lAO6yx8}~i5L53j54urCMXm8W|A_oZ$^lTeKSf-?3;07
zV&95U7WS<KWnte+QWo~D2(hqlMTv!dD^9G|zHh}%+x@$#^XInJ!%Bx$r?*lg&EL!^
zl-^jI+F=!2#e>rSV2&HriT(P`vW%vUS5|q)w4MHAPqLd2xmx!Qn0%)G#)va>i#K&i
zogLYoJ<{*Vt)}fX5MyV6SUU^E**PHI&I1W{0mx-{0f}}INU}>nvRwvp+ucBlT>(<<
zDv)N^fIM~&kZ#w347&m3wR?d~y9s33triN6^wlldv+bG8-0H%-!ou{4U~Y9$UQuCs
zr<+^dHLq)7`Y<uKdVn^-O5c>+YM<r<8mJ9yq4YaaJLRUc$JQI>l$;ve8HomBBnF6;
zSRhW~fOv@q5+ni0CAok^Nd%H42}qV?Ah+ZOQX~aPl~f>2(ttdY2S}H6AVV^Mypk8l
zluRH?vRWv8D^iQsbTSKz=Wse<@f<V|<6wYT2MfeGI3V7^0|^cR$mMVWi4GA+a!5e3
zLk4m?+(3#$0a6_*kmk^UJPr?#?$Ch@hXLeuc!5lZ31m5}7TSJe>5EVOedFrXCrB@U
zuhHq{+OaU2F=qCb%GBk84h4nje?|`LW2a>d*>WRwIgieh(nr(v<0=}f9p{gwE)Q)x
zw0HV9{wpQpZo7^}uct1TcPV$IFQ-q-sqE%c^7Q?>Vc)Ta^xfU`5tmQrD`_kJu9V$$
za-Ozr;n1upO(*E{!c*r3nqEw78vgV3bozy@O?97-_fP-i&(HVHoIIgw;?t9CRV{P2
z39T<DwZ5F(`f}>*X|raeaKy7C+qBKdY-i8P$!(w4vBU8EA%mFw?BN4D6?HD?+Rf_I
zudKYH{{Xp*t8b~7bNsifg~dI3mOR$mNwd5lx)s%9n7#Dwk~X+Y;ZXk+QaRQ)jQZKX
E0gn7V>i_@%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp.9E2B022D116E5337.idx b/.cache/clangd/index/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp.9E2B022D116E5337.idx
new file mode 100755
index 0000000000000000000000000000000000000000..34c8066366ba5dcbfd74f39a69d356366a8cb4f2
GIT binary patch
literal 5482
zcmZvfc~nzp9)@!-7($ZDQVgI%Q18A2VG#t33zj&5C8z<F2%4%@ss^>L9b5NSszqyC
zwL%3^BRzFouqDwJY-=@*75Ak<t<q|#qqfepxQr7r=fdxg$vJ%7=e}>Alkf6zZ@z+@
z?Ci2W3dQiLISb2e^GX913PqUjr(&LcMr(jVF-D_MEZlx{+=TpFsS*FKzOuV@FTI~L
zuUpf2CgxB;p>|O0pi7xK4QUH5E+}sL@{J9fR&Qw;!ZmCxIFVWX!?UUch4RMlKc;>%
zbYW3RMCZ9ngPL#HACF%!KYhp2FHaE%Z(c!eKK*7<a@Ob4w(9=+(9?M#Dskk*>zP=~
zz4i;Yf2lsy*?EiEcKy2Nz{97H%U)W^Jz0?P!_U`GgdB7}**0tG@Xbe-&ptOJXyPAd
z$A0II+pJpr&CDa&)xqrd#GS8p?Rz!y>z2ulBWAZg7+G8S(^b<sYH85(Cxa_X7T!!r
z4}P80#;%G#boQRQye_7ASA|seCQ(=W=jq*ls-F>CuIlbSh{(Y2$G?6od|l3jm;p5>
zXUvY_?$nhBHmymXJ}v#g)8E@W24lyI>jTv1_pGceer>?Coc7p*wSySd%otm+x}x>N
zPoq{H9z_pqR__Wc8S#^5!kMIqz&%IzO<by5lRtIJ`99&J)*gGWwe!sI+cAsGcaJBo
zw3LsnFMjv4_1vt0mNe(yi7(b{at>eLsCxgw<_$3q^^bawYnl1UA*}H1%YW_6xqG>9
z;oVT`!N!>ebwR^nX=_x-9D97z=X>XlK5{#H6&4-1Wa{|c(sH4pcKL!s#nj<zYZPCc
zsV5g|?~KQC*aXw2=wm@;xA@MzH9H?|oI%$fFI}_x?4nay+dm#QZ)5S{T~%%4=PqIQ
zR!%BD(zsvOM%+(3xP)8VzqIxC;KkG5qV@-E9lE;St{XQ$?0#+MPuX8Pk_+Qw=2n&k
z@B6R`?=_=i{M*qVqyO>Nyk9sh??^$TFsAjfZg%DF+(%6hb_$!;ep{7zZ^;MZg&i5{
zA|e+XIREC9X<-4()1_}@lwV?t&5p*zO?7o8D^|_5=Zu-1!L6KFFu!VU_V~1t{_%6P
zzgntv;W-~wRFZ`={@%Rp#<|ec^!D?E8gCpw*W9{1n`ys4u>D?$GXJ+_e>7C{&7*fL
zZtML1!sUOJ)|an+3;ph$rkfgL^RV~(SjIjty7J`woK-9ze{Hud|5WIxA=f9o{pv@3
zw!Cp6@AA-#Nw*4@{Fq+Ue1FBJ<}mAY_vlHZPk#HsKy~t#1IzNS&cFI;O4+)^S9d<B
ze>|;o^G6%5rp%f3t9G0DdgkV~q>(58>D~Hl^7->uf~E!$AJn#%ZXZ>fcYFHz>-{P(
zuQ_x(VprHN_n(H_zo?rwreH^=aOcdG#w}^TEbJ)U^Kzs2$Cm1vvF6>+9YZsX&vc1p
zx3>Njwk+s*W>wpcM;&=>?2Suzo|k&akgB$;3wCUFZLG?yJRP`R8d4v8t1P{)ddj0~
z4_C(KeONajH#@O!`ft`(@}lw;6^qKtk`aYMssH@l;h&_$Y0$A2g*8+e8tfl@|NH)z
z`f*L$iqT1C4O-}n`ziZ{`3E_U+dq5E<~aurYtVdO9IcFg!9U1xLgUGduTI>5ra{N}
z;z(tr#y`k$){}?#_WZo)GYvY<7e^=~g8YLVXODdPTFsWYaIIpzFE;ctjPwt3Y`bX5
z%~FI%YtWZ{ag;Jj;~(UB^~J{<xsxlyHE5A9j#b7E^AB>o<Ml~>7oWOTt3mU8u}-Pe
z_y;-uwKT4YH{46qpk`kjrVI=A4{~g`y;3nLM4<?<FG(MK{O2r0YmE+#j+8Hld{JZ6
z{)<`J@}#EWW_p~~iW#vH@(1&g85`F_xDf{#ZpJ}I7zvOOW&&iSkpvlOCP79SDUeZS
z3S_jA1{rOpLB<#vkTGTkWUP?|8Ea-i#u+(~ab^x=ypabPZ{|TJ7zL0CW&vcPQ3RQ2
z7C|N%C6GyG31mqLNlD#H$eNOn(oOQBX4U@14E{V8Y)a1`1hZjt0NJr+J%n3u(8O&x
zXySGpGzkj<nuLu2O~OuqCTSr-leCebN!m%!q%0I@QZ@=SDLVz4w1oyu+D3yWZKpw#
zu`r;?*ci}c><nnK78W#F8w;AOodr$K!ht4d<3N+MbD+svc+li+JZSQE9yA4u0GfhL
z08POzfTm~>K~uDepefo#(3C6^Xi7E-G$p$Pn&}DY89j??NWu_683`HPB;N^otNk}b
zu*3{lI#@;q91vk30Fed~5M`hM(FPh2V_*QW1{M%!-~jOk9*|%V0Eq?>kYtd$NjjDM
zVhI|N<h#oy#U;hd^Ox`LzBzW?+QA$Nj0JNfa6kk>03rzz5JgacXo3dB5DXxeU;%Li
z2Z$$lKms8E5(yEIL`dBvKa29!{$~)Fj0?ezB0bZ=JP39gvKPVHAUjZO9}4<T6mvm#
zquBR=aTkI^Kkh+r=*PVX4*j?T#i1W}qB!*9ZWM=p!i5mfPk0al`Ux*WKtJI?3Fs%B
zC;|P18zrEhbRi`4lOBYGe$tDO&`&y068cFeN<u&BMoH+WTnGjIln0@ppYkFU^ivL$
zf_}=0QqWJiQ40EL7eYfn?Llbhr@aUb{j>w6p`Ui5H1yMMl!kuBg)q?1cn}8q885;>
zKjT0d=x3ZL1O1E}WuTvRAuRN>9)yK{){C&v&pJ>R`dKH+LO<(9S?K3n2nYR~2jQTf
z^CBGda}Jb)e$I(<(9gM14*Gc)!b3mrL3rrry$BEeyaVN-pLe1>^z&|%hkn6@2+%Kh
z5CQrHFCsv{;6MfF7o4a7{el}6pkH($BJ_(MM1+3Pi-^!KI#3b%MJFmkzvxCq=$BlG
z1pSf+k)U7lA`<jV4pf4E$%#tPFS${v$8TKoo1*j4ysJo{b?Bg>R{7R0|3=jQuiWz8
zFgK-YwvJ!F(f1(em)dW*{5`5LFPyyd-jH5h&xhco5c$nzEghrJ&FNS3g72HI3hUo<
zOT<+;AVNg|B2^?HN<{&pRWu+*#Q<VeEFezB0pe9WAVDPn5>+A~NhNiYE!TMT&41LF
zDYaHzs7@zu7_`>N&`6#9?9*DKLZfu@Rzzz}O-N0Z=dZO6P8bX*Eg`L&<ahX<on1Q@
z`2zc{6ptl<E5+l02%Z2$@+2UNrvTAB4T#|xKrGJ!;&=`a&+~u;UH~NWA|Q#Ex=H@L
zsM^29>X{5&Qu6il&A4l~)l<Us7$zIER$Pw*BJ>0xQcnV+^b{akPXl7~3?Np|0^;->
zAYRV{67&KfQ7;0L^intN>sYa<=;ro!eJeiJ6q_%9{Dy_jUSrC4X8TS@^p4QU-zkn6
z3a0iNxwFG}IusA(<=tURVdh<{{?ZBG>8!x4f%39xmwG)43W<EjcRHz0l3qS7??^+`
z{+%eeekLqe{s6w~XTkvyCIS#?A^}k*3J`6g0Wl^95Nl!qaV8EBZ{h(7CIOIW5&=mj
z2~cuEa`%3$L)EYWP21mfTM5I%BzdL^yKPo!*N(fwKBdCkbugi8_}^=ye4q%hp563q
z(c!<8Mas#=lO{|oQB3O26vaJ<B|V3edk&|RPn|X0hofJ~4G8R|3{nMav@e8)^v($z
zIh+Vn=L`#vd@&*_I(2aRkc`Z%p+X-vZBR0$|9@w7{rbnm4H$^wB+YQ56rY%s5^Lzo
TCyeN$8)Yg*vkQzbA%FWX!-tG;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward.hpp.43E337CEA4624B30.idx b/.cache/clangd/index/device_batchnorm_backward.hpp.43E337CEA4624B30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..deb2d76cdfa77318a3c46f9480fb50e8d3bddfef
GIT binary patch
literal 1304
zcmWIYbaNA6Wngel@vO*AElFfyU|<jg;^LB`%nn8dh9aOq#hkqpyuFzXMOvRPKjT!s
zcH>uBuZ2tXVtzQd-r?OoGh6-46URxJvPb^s<!qh2g;l&uz@up2lX=ztMmMTo2VT#Z
zqF_)vuc6t!zWV)=>ioZAK5@!LA`{e<j_XPqiY=ThW^1s*plP|I>+U7Nd}dWe*X(w>
zP4G`ps*(Fv61mXfrP!TEYvUeqpGxgpKhyiY<23J;<vXtBWf`&^?~+bEV*AYWFsJzY
z!zmkNebw(y^wKc1KBCa=(3`~Joue*rq4%Nk>x_4PQhj&smGN2os98Mcxt_A;M7c-#
zIhE@Hk4~I<<-gqe=+eVe%T%AQa`~)VDZReop~75qj|k6|&Y4qBPH|)I)Ar2Vw`IW>
z_C0qy|9?89y`*J=6Q@dRtBlTk;k9BJMiT-|r$rx{a^cjWEr-sSt8Y|_*|h51#0{6$
zH#@zy*VuB?a^}lfYi9kwwMjGLkH5M6)GvWQg|)>KKAnipboX&g>nWWZG}ZO`+pnF0
zrzYN-buD|p{kQH4wqp*TyZ<w_8u@-=)MlHPUtF1+G#?l*ft=CPch?qO<z`~xVBoQ5
zvzG=F44h2B$YNmRVO%<gbGm-Z(UojW%s>T_?2?jT0-@l7*K23NYueKVnAkXhT7ZDZ
zNYlugfdj&Mdeobv=;y9j0VY<Mv>c0^I86FQ<<15BG#EdMGO=?q@JP!_>wpQEG(%Bp
zTCqJ7!kvs<VE1{dlyy(s<y;rL-7+UeO+}fJheuXMRs&{PY9Fg=gHW*(P*_q~QXl4r
zio*||RIyw<%gMtdrz{6o&BG(FERV!fP*y<VDJm->@syO6ka)_<%1AsFWfdfzs<J8)
zPfb}3&g=GKZh71P-va0iX;EnzSP%dmr733s^8(OOn(|0I1x*Dco}#8A5>HuI*#x0a
zMOOuhr>d)p#8cB%gY#ga!o&rRQCN`j@Yw6wJHXP#T4Sq#qjM9c14GhC#z-9&*)W=s
z6CBwvqn;YDeqrYq?`8%@hoG!9!c9hkM$#}nKns-Zbzv63)BsHar3;uzus~tu0;dO<
zFb|KinX)TP4-b!unF<n5)l3zMr)H)G=K<4EPNFd|MZ!`B1I$Di&0zm<$GO#?J~_h$
zW_d1Ab_`?v$WUCATx<<W&3up~4I;VVOeUVW+eF_8-aE;`!pq0S&A<znVc<qE0qGdj
Ay#N3J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward_bf16_instance.cpp.CEBDED0AD1292F09.idx b/.cache/clangd/index/device_batchnorm_backward_bf16_instance.cpp.CEBDED0AD1292F09.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a5a01a45fb4db913fb55693086641696fadab140
GIT binary patch
literal 12744
zcmZwN33Ss{qQ~+65%5w9NI|gc?RLA|F1Jl@+w`{S?V1)*v`w2X>E5Pm+H`NaSJ5J(
zJVnr>0}d{X$EVJ?z!{vuUB>MVDk^SIpBsY9sHnqJae0-QOn>j4$2lbV<r_YPoHqUQ
zZ_74Vt&gYp_!P8m@box`nil)`_^h7)JvcPb`KNPyd>--j@$u|FU@vvQZT36;OTYzb
zuh;f2zivzV#@;9H>$(54mx>E+JXv*B!uL^`onO38ozHCev^OK+?b*lIZ8){&`|Ee!
z;1lJO{^Eh-{s-3_9@?WH`|}6O4*mA0T}Qrj_v$0BynVRuKW&Gm!iD{j7296>yW4Wu
z^Q!I86;<zQ>*1fyeE+u-y7SccUyl9hhi~sMaob<`<hr$;KQujjVI_5ku3z=rFOL2u
z>*(ZjUp1w7uh?jR!@v6dhd-+CsdkQ@EbBe^?!@t%?{_*aFUsEq4J}Q3p=%$rSx?q{
zu;k~PCc1WIesMbHiY=clzV+a-Z|{7E-I5zG%Z~3In|rJH<;>4MnAp}!eVxZjJuA0g
z{ogwsm0zE{^BUjxN>A)hUuW8}b6M`0mpsmk?^)-5_t3=hl>7g3_g$CyvW&c~o1Y6*
zU;oFOJFfEoanq90*{65E{)12C0|T+1u!jn7d-%i4e*E!?j_wG9w{HyE-uK-_DN8FN
z#=d*6YxTX~?AiL@9j*&syXlE7U-z|se*7!`?18&43+g_-?>@`^gE8CeOK<yV`=|Y7
zKW{to<c(Q}_xZP;A2smD-?wD0G<H2d5fM6YWljD^HU0NJ`#}3$Yd5br<GX6S;O5mV
zCCqGln;-S*>PcAr)~7t_c<0|NAJu-c>HU}Pc-w!+j<q{)Pp)kz2U2(5b;XY73vN4n
z^~x>xy|(^X=}%4Bmp6uQ@492(V8f}?`))th*7Et0$G2y#d8qrTFI=~tAk}+LoZNeQ
z^4dMOonG7FYVU1*!Sl?=+d_;NF0?(f|M`%wSGWpx-&k@q{149tYzY13?w3hd_&tF;
z3w&SQ|N93ro_Hm~d|>1dk0|Vkz4_y*f4|Aqx}|pQsi@1ZvVHfeb*E3-^rJOLZaK00
z)(6VBWZiJ-2lwXv?!Gs!yL(%xDJlKcJ!^N{_xZkZ&tli2irABDFTeBoH{ZIxf@e4n
zjGpqp;NLFz>s8ecz2DK7wl{v)ry17{pNYKe#77UFx%bbGLu(R_92&m+zWet)RUSai
ztDikq@mJHP*3uKZ{%h*!lJ?}S4wh`;M@udk-g(hOljfZ96@?Y-8g)%({Dt35^;r%!
z-)y?~(p~3m_@?ugcTTNH>fk?izZ$vn)Q^_-m97W(K2i1Z)0TBdj(rw!Ti!*FZGZ3J
zn|I#v?GJ3{B~Q1RJ{dIKyx)AohCjMZ8y;O<`$}Wh@l$?hf|A(9_jdhwWOCqx&L`iv
zl4eAwe;c&6d?!8k(^rlTmHcLnO<I4k{h{OMU8Q^f;z=3)a)0>G2am_>sXcdJ#T%*5
zopZ^DyG}<QeB_t&RzGH&nu>a{=7$4+Z!i4jp5Dc`{QQyrYRiYm2FH4uAD#b9_Wfb+
zyPG$r*_JOVwfQ*wm-??hdof?m{@UCaXuT&Tb!7RXP4jux(pA1^FT8oH^Wps~Eq6WO
z=TkPHuP~NeVW@vzc^|T?HSEnJmsGD`zNljUA)A(N$~=4F&3C?c*}J#<Upc*eQTcpc
zw{)HP?1eY~be{decic~ZynNB6^ZC4`=Pf&X;mrfiw!zATg@46j-?PvD+9%KKz4!gB
zFh9qfkvnH>{MXZTY&9<0YE&-lZN|mhj4JQW+w!^3_ume^wP^AD1J`Udio9X5W1F#N
zn_>0ti!*k(*Ck$ao#E&3H~SZOU&QiF{xK)7&t5Wr_m#%7D~(m&$2WL(yBi-l^x=G1
zwR+VB-dELo{gm;Kmv!NJ2he~>?@Qz72#^8hg%wBxp#{o7XhAdxT96Ea7EFVo1<PP)
zAv6S9hzx-iN<*QA%1~&OQfO3CXkj!AT9^!j7EZ&Vh0Abg5i|l?gp7a|Nh6^}%1CHY
zGzwajjDp4}gT^F-7EPm}MayVtF*F8RjEsR6OJkwM%2;S|G!9yvjDyB0hsGs`7Ej}$
z#mjhT2{Zv(f=qyxNE4wY%0y^MGznUgOoAq;fF`7XmQ0hOCCg-JDKrIIicEoKq9$l2
zX@ZtYQ=z5GRA`b)Xi`dOX*3O5noNV1PSc^K%XDZNGy__O%z$R5W@u(<hL%Y)p=HWU
zXo@OmN-AhsGz(gm%z~Cpv!P|nY-kp0fo72wXgM?oT8_+trm2RerG}PEbD`zRTxfYT
z4_cnggO*S8q2<ebXa%$YT7fKpR!9q>70N<rMYIT7kt~8{rB-NGX@yoyi=h?EVrV6_
z1X_tKfmTXOBfVdF{2ZmS)V#3DXc@FJSq9BUZP0Ae2CbZyLo1i%&?;yJv<g`P%}(vm
z?9vXcl2$^il$Fq`Xce?7Sp}_{Rzs_n)zE5a4YV3r1I<Al&>YeMt(Mk8tCh9T>S!Ib
zI#~y;p4LOFm-WyZXalqc*#NDPHbQHZjnJBC6SOAT1g)7iLu;1J&{}8<v=-R{t(CSy
zYn83goYV=;DV@;TXdARP*#@nhwnJ-|?a(@C2eb~^0j-mELhF>B(7I?Bv@Y2Nt($g3
z>z3WnT+{{4C0)>ZXb-d=*#oVY_Co8Gz0lm$4b3gx(E4Z}v_9Dft)KQo>zDn|2Iv5^
z0XYC|kPbo{l!MTQ=n%9aIRtH(4nrH3!_Y?P2(%G70&SFzLK~H%&^*)w%_BX~#^@Nd
zF*ycpoQ^{qm*dbT=mfM0IRR~wPC}cMlhCH<6tpQh1#Oy6Lz|Y<&}Qfiv>7=AZPuW(
z2Gm(Y&Kl6?jG#G#dJkcLJ(f<MGo*KWujdb4|8d;!J$UuL(Jt~md!X;V+wcDzLf^9C
z&d5Cbl2eWIhqojk6b^4mAgmw~1S^;X!wMlGutG^FEJ`S>FcJnUoP@)QAQ7-4NhGW&
z5(SG91}mCG!-^p>uwqFptT++}ixUnjp2WjSAPKM%Ng}Kyk_1Z-0V|m#!%873uuQ}R
zE0v_el0?EvBWbYGNjj_yk^#$1%&;;^CM-o1tSpiRE1P7)vJeZb9FhY|6Adeu<ig4$
zd9d<HKCA*#0IQG`!YU#~u&l%itC$qSDj_AXN=a#`_oLd+QAWyO*@z8RIVp!#K`LO`
zi5*rYsf1NUs$f-<YFIU-29|?3VAYaZSaqZhRz0bQ)j%3xHIhbHO{57{Giip^LRw(8
zl2%ww;)K;k+F-Slc32&x16C*Lgw;j5V0DvjST5p%)kAt<^^#s#ZsLa3NBUs(lYUqO
zWB}G68H6=NhF}eoVOS$%1lA}Sh2<e0SYu=i);Jl5H9;m|O_E7iQ)CL(G?|7qLuO#j
z8f4b^*Y>`5&cEaSdIDwP9ru5q#Mrlc*-he#!*|c0gqSnPoDt}K3%}>H&12Q?@7})1
z&#~Tj{VFf#=eWf8l2zUh{3pwEK53w@KQsTJtBl2084=z;?dQ1KSbX)uKAybf+jCc~
z-*)yS$oW?;d<F5I%lY^m-?Jxi7M^DS3y59#TLXANA}kSLf~5j1uyjEF!U|-8pak+j
zPy$6DD1j;vlt3K_N)QVIC5Q)s5+s5^2~t6z1nD49f>|&q!8{n0U=a*TunGnxSO<d=
z!a_g^;US=eh!9XhR0t>`Is}wZ779uz4+SMugn|;PLO}`Dp`cJkL7|+2LInkdDhdkK
z6qGO)21*zY10_s^ffA;|Knc@fpoFt<P{MgQDB&U;lyDUeO1KUOC4xnO62T)ti4YN>
zM5qW*B6I{Okt`CFNFE7Fq=*D1QbmFisUty&Vo{((@hDKDL=-4dDhiY+9R&(w3>3y0
zC`>R=m|~zX%|MA}(V#^0Xi%a>G$_$38kA@q4N44)0VRgVfD$8OK#5T?pv34HP-0mu
zD6u>ilvoi9O00?nC056R635~|iQ{pg#ECdi;#3?caXJnZ&NwKXb5OY8pm4=O;hKXI
z&*DLe=kcJ#i+E7tRXiy1Iv$h+mH<iuPXHxBB!H5j5<p4N37{mhL{JiWA}EO>5tKxg
z2uh+(1SN?jfs(|NKuHowpd_gzP?B^KD1r%41Q(zPAwUsIfFiU2C7C6IlFXAqNfybV
zB&%dll65jDDJ%t)6rKV~ibw$^MWuj}qEkRIF%u{zZUV(5OrV&Q2^5nyfs)EnK}qGQ
zprndaP*PPYD5*LX6v-qgl1osekf2B<L6KU5lE%_NN#kjtq=__8(o`BKX*vy*be0ZE
zI!^~BU8IAOuF^qC*Xf{SunbT#cm^mLA_J5Rl>thI&H%;C%%GUL85FZHgJM=@P|Vs4
zN+!z$C6i}@k|{Dl$yAx3Wa>;%6jPumu0T;jfufWGMQH^}7Rv%9i)Vq7C9*)tQdyv6
z=`2vPSvDxyJR6j3kqt_=$_6D{XM<v47EmnQ0*XagK(QzbC>Cu2C5PpJlEZUA$q_lA
z<ft4_a&!(TnrTop*Pv*jLD5QsqO}Gkm*s+z%X2}=6}h0~s$5WVbuK7*EDw}Co(D>v
z$O9!$<$;o?^FYaG`Jm+Ud{FX5J}CJrAC!EZ4@v<m0HuHzfKng|Kq*iKpcLo=PzqTg
zD22QbltNJmN}(zQrBD}wQpActDdI(-6p12Gic}FOMY;$SE3<-P<yKIv!U~F2SwXRC
zD=5XR7?fgO3`(&m2BlaPgHo)EK`CJ+pp@_uP)bAzC?%={loDM6N+~Oi_5MF>{<|5x
zG!d34HNjG)7FfD8e_@reGEmBR87O6<43sie21=PO1I5N{pxC$#6q~StVpBFyY}y7&
zIV%UHoR@=AF3LeESLL9T>vB*kSOq8*yaJR8Q2|PYssN=zSAb$?c2Mlx4vJmaL9r`4
zD0XcJrIJ;GQpqbpsT7r<RH{l)Ds?3&Rjdk>DqaOjm8b%xN>zbUrK>=xX4Rln^J-A4
zMKvhZsv4APT@6YNs{y5k*ML$ZYCx$`HK5e!8c-a}0g8h=Kye5MC=TTS#i1Rb)UsMo
zYI!XvwW1c3T2%{5t*!;7j@5xu$Lm0;6Lp}}sX9>VbR8)5tR9qlUJpvWs0XE9)q_&6
z>p^K?4WKme22dJA11Jru0h9*a07@fk1f`KTg3>4&L1|Qtpfu`6P?}g1C{4TxlqS&x
zN|R~=rAaq|(#)DcY39wKG>c|XnpHC>&AJ(s7S;kv3vU6XMYMp@qFO*{(Ji2~vQ|)9
zc`GQbq7{@@)e1_hZUx23oS-<l6BMU#g5p$8P@LKcN*ikfrH!|N(k9wKX;W>WwCOfb
z+F3g&?Yte7cF_(>yJ`oeUAKeM!8$<c;2ofJhz?LXR0k*>x&xF>)(J`{?*ye&bb`{U
zIzj2wouG8FE>OC77bsn#3zRO^1xlCh0;QXEgVN2rLFpFVpmeKlP`Y(DC@$s##l>Bq
zxP%K7mvVvP(k@VXSPv*Yya$vX(F014>H(!k_khyNdO_*sy`c1pUQl{fFDSjb7Zf*h
zgW~3HP~5@|id(rsacegyeXI|ZKHdjPpXdXnPxXP)r~5$ZXZ@h`^L|kJML#J0svneo
z-4DtD8vtd14}dZt20$551E37(0Z<0nASi=;5R^eN2+E)u1Z7YUf-=O0KpEmgpbUv2
zP=?eHC_{P(lwmdu$}k@WWmpV@GOUI{8P>y~jIa?<M)(LQBVq)U5j6tJh#moDl#PNi
z%11#N6{Dbxs!>oz^(ZJF<^jdSJ)n4m2NaL;fa1{}P{!C8C}VsKlrb>|%9t7hWlWEO
zGS0?98Rz4mjEiwl#??3|<9ZyF2{r-B1fKw9LQH@%p(a3?&=a6cvPn=T`6MWlViJ@|
zH3`b3o&;rzO@T7Sr$Ct!Q=m+#DNv^L6e!bd8kA{14a&5b24z}JgEFnBL78DQpv>?Y
zP-esoC^KpXlo>q(%B;a=4RB@+K5Kw9Ylv9`tXV_N8lcS@de#7M&Iq3~q87e2HD^T6
z8F9Gd=ZqxW>2pRV?yNZ@$Gd$7oUW!{f7W<#pk=Ul{`wrp2Q@*PmmRCzxx`*oSzY5}
z_x>cZk5ARY&B}$FH48VZ7jCxnw7NRx=f<{+&sn@=>A7Uta=#Vlty;NZbx}d|YTpg{
z{_EGQy&%AxnU!tH(YZ<I$E(QnSo;6nyDlg=BrH54ip9k71d*JQnwAkN0~1Y!=dUZi
Ouy&F4lDrb*qW=M7FGg4Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward_f16_instance.cpp.D1D5284FB6757AF3.idx b/.cache/clangd/index/device_batchnorm_backward_f16_instance.cpp.D1D5284FB6757AF3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd4300cc770c9a257fa3143b663cee2a08223740
GIT binary patch
literal 12744
zcmZwN33S`VfyZ(GO_EL$r#T=2pJU78V|<K{vFx$zv1Q9AfdtF8Y+1hVOFm@DwtREM
z5D2tk$reHgDFMO}S|EfLmZMD|l$Jufuv{gSKxvzBAI<hy%9?al|90Q{y-51$D}N%r
z=SY7`u~(Lr{UjqO$ktgo>F)Bjtq2MV+PL&*%<JiU?7X0$N7n=eP3}ADs2cu_3OW0A
z=tbt&Qhv1ZXSQ48Tq6VFA6mZr^=0hMEvBsnmD~3I?N0|kRDCBO_{qcT$l&e^3U|}W
z+|}0L)+3j=#N~NamwfTzXJ4({{?73aZ}-kk`@Zt;Z%$i$Z~wa|rT?Kb?)t;x>ofes
z!Ts;%KJ)q6f~29s7k>Ek+?~(9HvfmyXMS<g^khToA)Ecq(}&(qYMOoO@s_t{AFKS~
zP{kMe_@Orrn^T-^TRV6B;l}BsU3*S`f9qqXrasxzbL?OD{AuQajFSiCJ#RejXvplk
z>IZTAPt8Ad$T9HT$-iGUGI7gK&U<a|{zv|;<ow<pW%q?-wH!QK<$dw-Kfis)Z>}!B
zC4bF#ep>ZeROSa4N8NT)m63mD*Q+Ov?PXoTzbeRTcD{Vyv$y7a^4p@2!=q>V4irCs
z>1FS>KD>Ct%)5h?C9CHSU6E*QKQS=7u)99@g8c6~@A%bQpIPVr-uj4nchH_AcdaW|
zGk0EeWzz2anUSO9)xW%0@<vtmU1xsrj?J$!&w3^|b`JHwb7Jpo-@y9RkDtF|Pf_Ay
z2dDPk^!33b55B)`^xhTDl{?>h-<R@S%mKT9`!lw?FWY)eWAYC+=bruWruGNhT22)W
zC(1k6sukN`dGy+>g`Zy^SL|wf_OAQ~p1gWj@<|>s<#iR?Z@j{O!QuxY-ydK6{HDw7
z{=`S>e|_q~qklbSxvy|n`G%Bds@#uwKVJLxQ+2PrIk@X&=<B0B?c?s(Z~NsJA#F`P
zp_Y5}OApCc_6J}8(F@lPhJNz;%%icb&b~7@J07z=b-DYRmp4^fez!R{_WAeb&0$5U
zowZkbDlXsqs`+gEWW$O>p6Bj26~D50W8d3{rk*bw|5@qA?WSwgj`i-6)t9~3bN~B?
z{V$#l-Rk?XTpjmf*x{p_!!v!qKEC(5b-sj^hidOV_NB?V)w+3&Wk>YAo=3+!_x$`)
z=SNLXtT+){CfhyP5$nV%?_Kt<?tW-!KH<}w_f6dHo49V`if3>6kvO{f(XpG8-a9WL
z*xCFagTLIFa^{K84{f)+7ZN{u|BAcIOT$(lyy3>a><D_>sypu6RJQKKzRht@L^gg~
zyxtJ%xaZ<)I>(Ra{P=+X)3ncS*znA!+xKNZSovRXrS>K{vWEV1#}QvM%i3}JJ4F#W
zw|Z}CF0g&E^Yu&OXA9#0WlMMF>9d!{KmWlg^OHxbqi(l8{c+X5AHMIzM?1Hc<zISH
z-WNwsJhM1=<rQZ>`@lMD-}`oNb^Eoe{<vqV^o!ef1)EPK-#dM1-6zkbK7IOkmc$2+
z?XQfQ3`;xG^3#vLn!e`Gx1D+J_@D2*y`^EkJZt`s;=p6upE9mXJXH}i=6AP0we&A}
z&0i0^edShj_1fU7>LAyqRhu@RJ6S5{zBc}FXPEB?XO6E8-oBLAuUfzQ+({tUyFd9}
zqwk%ULxSu}`6^@ORYvaroCR*u(-rf?&O@bZ)&|!u-DKOUZR*@fAph#gr(e49x9|OY
zZE(#}-m+>-&bgC7{%V8c;giG9{C;ilWlMR(stv2podmL{t9z_2W%*I8Saa^)Uw`ax
z2M(Xgiw<!u8rv2P=RfbJ>sn*$wT5GP?=mjgWz+?B;L#p8)|`8o%T6c_H3jZ!h$~cv
zW-qTW8U`&)hCvIb;n2cmIJ5{F0WCsCK#Qc2&?03dv?v+{ElNf~qm)9Ul0u87(a@r0
zG_)8R11(0zKr>MjG?O$zi>0y9Vr48eMj13F8MHVW2Q5y<L5rvH(Bfr0v;>*}EkPzg
zOQea=5@jMZPB}C#IkY621T9G>K})8|(2`{`v=o{GEk&k4OQosMQe`SMK?O7+1++Ap
z1}#mdK})CU(9&f(v<#X7EkkBN%cPmmGG!(-NhLHXCA2J>1uaWvK{Hb`G_y2A%cj}T
zvSl{39GU|yN9I6NR6$cxLCdAN&~jxiGz+yrvq%fHJemhBPv$|(r}@zGWj-`bH8d?X
zv;tZHtw0t)E2M?c3S}WQE44ziN-MM?S_G{~7C|ef#n6gnF*F;sL9<C4v=UkZtwfeU
zE2X8-N@XduGFk?$OqM|_r{&PfWjVA8S^=#>RzRzwRi?lzL5QnLR%I_QJGDczOFOh`
zS`DpQRzs_yHPC8g4YXQX3$0exLUT|DG>3FRtD|+$>SP_XdRh;yUe-f%QYSR0bV6&O
z4bU2712h+PL32qLv_{$ptx+~YYoblinq(8SX4(v`SvEs!p)Jr_WDB%b+6t{zwnA&8
zZP40e8?<)X4y|3bL+hX&&^lxXv`*Ryty6YF>!MxIx?~r$ZrTm4TXsY1p*_%gWDm4n
z+6%2$_Co8UebD-3AGChj53OJJLmQw2&<5lHv_U!uZBPzEb5l1ow{$}rqC?Pz<PfxB
zIt*=C4nrHEBhW_V2((c;3T;%5Li11$G>`N^8>3^;#^e|@FZDw6N-s1Y^+EGVAGC2g
z4sBeHLz|!z&?e*rv`IP%ZBkA`^HV=Gzw|?!qEpbO<P@}NIt^`FPD7iaGtg$_476D~
z3vE`;LYt#=(B|YEw0SxYZC=hpTQKN?0d>KU3kLK>BYDvffn(Kgj%^DUjpD#Ay86SQ
zzkKO^GkC?)LGJn28YO|S!nMmdf0t1j*q7$qIJ_lw*LRH&*Cw^e7Wf6AX4|HOKkm+7
zxpegoWAzSWec<+i^QB|re;nWkj*0`%;K6Io9nlA_4*5TS-_@U%M&!P<D{Ki~zhV8B
zz%lmT$~$9=94kL-T{@g5p;2%+O~PP>lW<rOBm!0>iG&qJqF_-%VMUW@STQ69mWh~P
z#gbT9j4)VnBo0<QiHDUy5@02gL|B|~SV<%aRx(M3l|oWrrIJ)wf(Te?Bn?(NNr#m|
zGGJwrOjwdgSXm?smYJAgWs_`JIV1;`A_`V6$%SPh7Fc;C4^}?Ohoy;zRX_@06_P?&
zR$_%!M2cV)lVVslVuMveN??_eQdniA3|2WQhgCr;U{#T-sK5_<h|5mwu&PNltQt}S
ztCrNlau5fsI#LI#p47u~5+|$%(g4dvT(BBRBdjLU1gn`e!)hTduv$qgtTxgHtDUsN
z>L4AkI!Pz2F46_7n{>nKAw95qNiVEE(g&-b^urn;1F#0kAS^d=!x|z(u!hMntPwH-
zYm|(_@(>TKF){|rOT4gr#0P7fjKi8B6R;-9BrHGi!<r&fu%^j0tQj%`YnIHynj>?t
z=E*#)1%oUY|J--&Tm8MXtDe1g=^1v>xM<Nx3Or@s|G|}h=c#=+1c$gTTyx?2<>%P#
zYqqZs{6rtCDLCCiUwdxpMmvoaI}KCdyF*;p87r<^-hppldcObWeVXOx`~UrV^M3lu
zFTb_^!t2i6n{|uEg#j<{Ue52&TXXJBoMj-Cg(d{_rI!O9ngUCNX24RRd9ZY-b$Nxc
zFi^sH7${*P43sbx21=L?10|e=gA&feK?xV(poFV%P{MUMC=o0Iln5RHN`#02B|=4j
z5}_kNiDZ$WMDj>bB1I%9kt!0DNF51E6pI2SibsJGC89uyQc<8p=_pVrqo7buL7{?z
zLKOvtY6?m;iv}f{M}rbAqCtsP(V#@@Xi#EU3@9-?29y{P14@jF0VPJqfMQ}MP)yte
zib<G2F)0%$CT#*Gmc@b+%VR-_6|tbis#s8Bbu1{1F;Ez1pfJHeVTysmGy^4$#eovX
z<3NcMaiGMhI8fqr94PTD9+Y?<4@$g<2PIy`gA%XfK}lc<pd|1FP!dD}C<!V7lmwjs
zN+L@HC6OnBk|+{ENmPlTB<e&^IOCvj&OzaVgTfUDg=-E<5=#Oli6?=QB$7Z$Qc0jB
z=_F8+Su!ZeJQ<W^kqk<*N(LoaCxep0Qb0-JDWIf?6i`xB3MeT$1(Z~l3Q8(Z1tnFa
zf|9CIK}prApa>>F5nO;GgaAb-0gBK9lr)wGN*YfCB~7G(lBUu?Nz-Ydq_cET(s?>4
z=^`DJbd?TDx=sfrgJpn{!81U~5E-Cks0>gtbOtDyEEAMWo(W2($OI)*WrC8aGeMC|
zf+D#DMG6UuR1y@aB`8@e3zRIL1xl940wqgjfs&=OKru5jC}wU3#VpLAn3Wk6vo?d0
z&9Xts=GmZRi)>J`RW>NuIvbQ6mIF!-&jBSz<baZ+azM$^IiM(}Kv7(QqJ#oPDFuqs
z3Y1)y3ra4}1tnMHf|9FpLCMv*pjem%6brY2Vi6WlEXo3kMO#40V|k$D@jOuSL>?%4
zDi4%Aod-%j%LgT&=Yx_j@<GX0`Jm+Md{8vgplGf^(L#fwl?Fv?4N3tk0HuHzfKng|
zKq*iKpcLo=PzqTgD22QbltNJmN}(zQrBD}wVr5oPtlSEURail>Dk~^fZ3U%>6@gO3
zi$Ez7MW7U^B2bES5h%s17?fgO3`(&m2BlaPgHo)EL9sC#C^l{b#U^Z^*pv+vo3?>c
z!b(6X;U%Dyh!RjrR0$|0x&)L`RticfF9oGkl!8*KN<k^rrJ$6tGEmBR87O6<43sie
z21=PO1ErjmgHq1RK`9sIpp>g}P|9^VC>5*%lnPz}N`<HZr9xGJQlTqAsbW<LfnV2_
zK0V=8DX>IU1}s&T2TNC3mzSN{L9uf?D0X27#jfn2*tH#$YE}(OHLnJxT2zBlt*Sw(
z*43cYuo_Tmcnv5uq6U;2RRc<mt^uW%)q+yXYeA_MwV>3hT2N|rEhrA=0L8%_pg4pB
z6o+zv;?NFI>R25pb-WIgI#CBoovH(+PS=4_&+0*`=k=h}i+WJ%RXr&6x*ilKbAsaJ
zPEee}35ruWL2+s)C=ILulm^}aN`q(sr9m}-(x4kaaWNMtF75)wC0wAmlnWG>c7f8!
z8bN8~ji5A&Mo=16BPfl!5tJs@1WFTc0;NebfzqU!KxxuVpfs~)P?~u&D9xf7lxEco
zO0#YTrG>SC(!yInX%Q`;w5S$PT67C2t*jN4R^AFqt7rwKRkecBs#`&6V{M?c@itJ}
zL>nk=stuGj-3CfKYX_yBw}a9y+Cgbo?Vz;lc2GK42Phr91C$QY0ZNDJ0Hs5BfYQl2
zLFwe3pmd5(P&!p7D4n_!lrGi<N*C_}rAu^y(xtjU>C#=GbhB<yx_LJ!-J%<mZq*G+
zx9$d|hxLHc!+SvK5j~*vs2)&ybPp)KtQVAC-U~{v=mn)$^@7r?dqL@AeW3L5K2Z8Z
zA1HmQ50pOL2TDKd2c@6)gVHbhLFrfhp!Dm0PzKllC<A-|lmRgS%77XGWk3&rGROu&
z8RUbY42nTe2Gt-agL)7YH*<sH=5A2j!VQXBxj}JjHz-4F2$Uf{1j>*Y0%b@Ifik3r
zKpAGkpbYb2P=>`YD8p(Plwmy#$_N_)WrUA_G9pGm8BrskjOY<iM%gGRqkI&UQ85b2
zs2T-jRF8t<VIELC+yjb7ctG(e4=5h(0cDJhfilL&Kp7Kbpp2<8P{#BaC|>3T#ml{*
zc!d`fukwQ8)m~71%m<2(`#|vtA1FTM1I4F(pp3I|P{#Q<DC1%rlyNl<%D5f}Wr9tB
zGQlT6nGh48OsEM^CiDa-lWY={Nj?e6q?iO{QcZ#~sV714Ge0PP?gzy${Gj-i9~8g#
zgEGaYK$+rGpiGG=P^Q!rC{ub0lxa2%$~2z_Wm-&wGOeaTnby;w%&-|yX7~&!Ghzmm
z88rjSjGh5ymd%1P%V$BE6|<nss##EG^(-iJYz~w;J_pL2m;+@_&4Dtf=Rlcf^PtT0
zc~IuXJSg*O9+Y`K56XhU77TC}48CB1v|xw@1FQu@Ef}CJ7<$0~Z_!9!G-Titr*A$F
zGcOtnw`I}LxJwp|a@>`R#wCFr<mno0`{w@|k9j)A%9qa1bM1LyZLj&^-M?SysIPN2
z1UbH`|M;p8TnGxPTfW$^e9^glvBTXt*t;~hc3*nlij}LrMOLp3S+`;Rw<|Z6+Tu5^
zsVv@f;rSO_6uR_gQNAuHw2IBime}k>`u|<KB|IWBI>xm0&i~RI(rM`#(wq~O6_%P=
SvUyAS#f`yb+lnfTOa2@E2ssD<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward_f32_instance.cpp.CA2117C76C8B866B.idx b/.cache/clangd/index/device_batchnorm_backward_f32_instance.cpp.CA2117C76C8B866B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f574f4832e5a9e2b0b3b455deba21effdbb0530c
GIT binary patch
literal 12646
zcmZwN2XtFig2r*nB$*q>ncx&`x#-&Rb-A*AU9QV@xh}Ho;1n!3*_L~+a+7<5AwHWj
zl%Z_GunB|`S_(Tr7+A_u!T<>wmM{d!Bn%k>LpBK|1lYhbtjUhZx96<R@$>uZUccYd
zJJNgSSg*9Gu&_SS%PYUNXx!B{+_b>U%gb;6$2~mM_2d;^UXS^Bd5v%1?<nd2v-17s
zM`OVLwMYN#ed?aWx5hoSf9Yt*cb(D8T}y6>xSK^C{^7g(&Z?pldtZ2LX>{*xD|2q8
zMY|T~dN)44u15SSvt-@*v!`EPSp4Bf7d{^z9h>>8ZR-bb#6LQ@<=Zz;U3_e;@%iBA
z&+&bQ`W5%4S5D-YAD`KNV)}a<{mb8fyX(-S-+pyr^65KjZ#A#aiQjtq{GqR3{_nq^
zaZS#wz4LQpcJYIIn_l#(F~?+WJHC7A?z30b9bEcu=A&6Fp4_wNZ!tHX4&0G<*s=e}
ztvq}<?n~bvzta1i>7R|Rp8LMZ`{vr7$s?ziuiqJ;T-vttqR;-D?{y!2<=aIk#L9!G
zO2&O|y7SvRsw>ZbdjAbRAC`1naK-Mqx-`;jNB&3uoU+&X?8H|m^??I{58k!Na{KWG
z5#L1iM4#IHc!XcXf9>_T*K4!!(XUN~%^%yp$Zg+jb-#MW#f#6qb=2lv{>QDU!5@CH
zdC2z>+xahEe8*h%x6Bn4tKYf(k@yYEKF)64{_z*)*N^|{g&R|zIdatGJbqt6%m(|=
z>K~Nl$NY|_mv{WNbn#<8%bz}U=DnL{th>(NwQ1wJd$wNrw4*j==<NFsA834e-%HPZ
zU~jV(ts~d`>B%k0*B!hu!dg@J%C4;0sb$8>e=AcLK24pwpGEA7Uiz-ldd~M9e$V~i
z*t>ss-?nXg-|;_M^5e<w1ut4_-zi)>*zm<S&pmLo-F5cp3tKZ+?CyE{v);Qp$nK0o
z|B-iY;-($<d~<tea7SP3q48Hgd3=T8(3-aoJaI7Yd`hqN_|~hBJiqTXW0C2{y>F4g
zu-$>T&n!-UV&Cu5UVnS_fMx7ii`nLS=#Ecj&Ti{%-B`Oaae3Z>xanTG=DI~WHR6H8
z?!q_Ud*p{5MbU%rHtyP6b=$*-Ope+;T^T%m@0wMS{)^vR@QB^M^7{M#`2Mk5DlBPD
z`^Uaq{#0P%F~8d$KlIjz&AFA9ultvO|4pTL@#_=cR*(IDJgVyW#F^5Xre`)joRmFa
z{>{e~haKBryZ7~&Ycq}*6Yn^1<MKO9E4$bYk!v1!|BG`kc7;A-X-aN<qq^j!XI>l_
zc+qk2U#%5Z_IAzle@H*K;o8k(*M<JmwdP|lM7ob{ir9NB|77nD|HZDx?^?WGJ)9A;
z>*bECv;EF|`P7@Qj30mLhF4EK{iT}F``h<DdGMpf7tWX>&OO|H{F>UNzjnUy-fwfd
z*6lgnZ@0(KhCEcfM*6#o*1dK9*(19<|9)Hb&OeNYjC}sY$sYzzr4{VkaKv)<j#IyS
z@JsXetB!wg?JxcQz3hi^zv53W?)rMH^{b7a8ID~%b**95hc~+?T+MgSe~Nv++xfwU
zwJBvwyzQl4PQOKdewP;WBbOfjSD*iVg-_)067RD4;Z=)PrCnNhhJP(;qjGk3ytKr-
ze13S{qIK%h!ZZ9YueKKX{AXCduUE<Z@b!j;*BjFR@yj#ihbtYAo#;RC$rA74`6<g5
zEn9qP;TiTCYHN1~#9rRd1wNPd@5YzD-?`^xW~i@o&T!3~q55aZxyi70lfiMhw;0xK
zF;sfm^Lo5}F1?P+o7rSovB_ZbOv~GN*6ZSh;bYzl=C_9f3!GaFE4CO4JQLTa-P(Uu
z+zr1m_&WWRU%uyPmS3|x>eE}Y7S2Dt!LWFPVVP&X=YY-oKOL>h{{P+VokjPD=Q$RB
z+34%^C;leSUi&%&NC1?P7@-15AXE?uf(j<VP$47)%0x_1lu)Qp5(*VY!l2B=3>8ko
zp%`IM5hMaCl0-s9ktnEW5)H)(hq4e0R1ArMiY2j7aU>2(5CIiW;-L~q0#qVNgi0bw
zP?AWfWReV(LQ<eoNh(wtNrO^EL8X&)s0@+;l}R$8vPc$`CK@W6WJBeU9H?B93zbLm
zpsd6Sl~3}aY{UjtKnkD=Ng-4*DK>fD)xJ(Uu|t)R5~xyA3ROnRpvp-(R0XMkau5ep
zC8>m}B2`e;q#DXeoKQ8S2CA0SLe-HvsCrTl)j%4c8c8Em6KR5KCe2VSqy?&#v_iF!
zHmG*e4%I<ApgKt>R2S)j>L%S#J){S!m-IrphzqKZ^g;ELey9O505wPkp@zs1l$*Gr
zhRHD02pNGIC8JPdWDIJYj6+S338+aj2{lEgpr*+*)C`${nkBP8m*>1|@cGAe$8*Nr
zwSDon__<deo<Bp+83N`E7SGJxXEsb!pWJ?vx36<m&MMpGQ}248^~*eO^q0!BPdCta
zU!9+{(Xe2n!R+~KU*~4Sg3Xuv<AjA5ezDB|*O!>3a|VCUM9+2ap@lw|E_j~(nm^%F
zfAeM4pZP~YasOB-;hzX4{ZpZoe<qan&%Im$GysGE763v34*(%R1b`4A13(B+0U!kE
z01%AS2!fFrK`?S72u5K9!6=O&7?lwOqc(yNNCQC#WPu<A@<0#*MIZ=)G7yA76$nD0
z4g?{H27wU7f<Or3K_CQ)AP|CN5C}mk2!tRV1VS(k1|gUQgAmMvK?oMXAOy=`5Q0@O
z2*ElSgb*46LI?{1A%us35F$cA2$3Nmgs2b@LUafSCTaq~#7rQVxCsQ4Fo9r_CJ;=@
z1cFJMK%kU@Kp6#ratZ<!6a*?M2vkuJsHPx<(ohgWSttmhJQRdb5eh=63<V)ng@O>O
zLqQ0mVIYLDFc89c7zkk^41_Ql211w$10hU@fncU)5X{UBf|;8^FbgvXW@!e&tjr*o
zwHbtP8V*7@3kM;bhl3C<!a)d^;UI*oa1g?EI0%d~5Ex@1FwQ_=f`Py!1A!?90@DnH
z2pR!G1d9M6f=7T5AtFGCkP#q6s0a`ubOZ>IG!leJ770Qmj|3r7M1l}0BSDB%ksw6s
zND!iE6bMl)3WO*g1wxdF0wGF9fe@vlK#0;&AVkw>5TaQ$2+=$mglG{BLbQwqAzDR)
z5Ury@;FN>F83%!L4gwb(1THxUTyYS%<{((81q2JTfMDSk5G=w1f<;<DuqX=%7Ht6`
zhQ@#p!(u>);V~e@h!_xJWDE!~Dh7lY9RorvjRhf=#exvaV?l@&u^`0CSP)`WEC{hW
z7KAt&2SOZ+10jybfe<I+K!}rZAjGLS5aM(k2!aX_1QQ?#E<g}MfFPs*K_~%&&;o>b
z8V^D|iw7Z|$Ab_r;z5X)@gT&jco5=sJO~Li0fYpW073#!03ksnfRG>)KuAytASCDn
z5E5x32#G8aghZYQLZV0nAyFoRkf;(tNYsfSB+(=gl2{T5NjwRJB#{I{l1u_2NhN`h
zq?169RDvLx1VM5Mf)o-2DJ2L}Nf4x#ASBad5RzFk2+2Gdgk+HnLb6N-Az3AZkgStI
zNTDepq_7kaQg{joDIx`g6qy1-ib???MW=v}N>f2dWvL*f@>CE~MJfoXG8KeWl?p<t
zP6Z*2rh$;g(m+V#X&|JDG!W8c8VG4B4TLnE27;mr1jQ5xiYpM5P#`F&Ku}77ptJ%Z
zou-43&eB0h=jkA%i*yjuWjYAyDjkG$oen|<%>W^TWq^>uGeF1?86afH3=lF@1_&8C
z1B6VP2|^~z1R;}Wf{-aPLCBPuAY`gc5HfWp2w5}>ge;Z?LKe>gAxmU|kR`J~$WmD#
zWa%srG}Rzzra{nLgP?^5K}!vSRvHAYH3-=>8-#3@4MH~01|eHygODw=LC99wAY|)o
z5OQb^2stbVgdCm&LXOA*AxGwbkfU-y$k90<<kDOaa#=12xjYwyT#*YxuFM4?SLK3`
zt8+ofqj@0Yu{;p+cpeCOA`gT-nFm6i$^#)!=Ye3QRuHVr3WAkeL9hxd2v%tY!K$nv
zShW>|e3}nJKFbFopXY;+FY-aim-!&%t9%ghbv_6-Y6HQ>Y#`XU4FsF8fnbw15Nyf@
zf=$~%D4+!(6tDsi3U~ns1)>0i0$Bh;fhqu@Ko@{eNDDzIWQ8CU@<I>_MIi`<vJiwq
zRR}_%E(D>N7Mnf)$MSU+v*IWyUK|T0iW8w^aVnH5&V<s%xtGgM?I75h9RxeKgJ2hS
z5bV+pf?e4`uxmRACA0*D5>^622`>SmM3jI~B1=FhQ6(Uh=n@c0X(<S$tQ3S&UJ62~
zC<UQZmV!{KN<k>qr682iG7!pG83<*(41_XK211!E1EEZnfl#K)Kq#l>Ae6Im5XyNu
z2<4(2gmPI9Lb)mjp<I`PP(dp|s9+T!RPYKADntbc6|w?^3RM9@g{}a>K^-7Cm;(d{
zcYxp!4iFsD0fIw0KyYXW2$i%Fgi2NkLM5*Rp;A<WP$?@xs8p38RO(6)s%RAmRjdkx
zDqaObm8b%tN>+hTrK&)v(p4Z-(`pc^Sv3gNyc&dRQ4K=1tOlW4RfABit3hy5CkRgF
z1i{IjAUK5+1gCU@;8ac!oZ1OO4Xpv8hSh*j!)rjO5j7yx$Qlr8R1F9<x(0+=S_?ug
zs|BH!*Md+hYC))#wII~0S`cb=EeLhA4um>Z2SOdM1EEgTflw#wK&Vr7Ak^tP5b9|?
z2=%NUgnC{NLcOR5p<dR5P_ODisMqx%G|&bR8dw7e4ZHz_2GIaQgKPkyK{bHTpc_DF
zq>UgnvPKXZc_Rpoq7j5f*$6_TY6PKCH-gYan?Pt{O&~P!CJ>rL69`SR34|uq1VWQ;
z0->2UgV4;HL1^a9AT*0+5SnE(2+gV)gl63gLJMsHp@p@8(860lXb~+Sw8$0^T2u=N
zExHASR@w?eD{BRzmA8V>Dq2Bkm8~GOs#XwMbt?#Mv<-wd)&@cwZv&xCw1Lnj+dybj
zZ6LJiHW1orI|%Kp9fWq?4nn(V2cccIgV3(pL1@?QAau|U5IR@~2pzlwgbvXGLWk@C
zp+j|m(4jj(=%k$>bh1tmI(a7uouU(jPT2`Ur|JZuQ+I;UMY}-gVqG9~@h%X$L>CBM
zvI~SR)dfPA?gF8kc7xE(x<TmX-5_*}ZV<X<HwfLT8-#A%4MGp?0ilQWfY8HxK<E)Y
zAoR!{5PDP(2tB$7gkIVULNDtDp_li9&?|aD=#{-7^r~JEdUY=dF6sio#atk`xC;cA
zaDm{GE)ZPG1%gYvK<J}=AoQ_55c+r@2z{atgg)5^LZ9jbp-=aL&`<k8=x6;P^z(iY
z`b9qo{jwi~e$@{`zwQTNfDV8#zy?4V-~%8Ghyf4=<Nyc*Y5;@*JpjTW9Ry*J4T3Pp
z2SFGVgCGpbK@bMjAP9qc5QHH*1i}y-0%3>`fiNV7Kp2ukAPlJ?5Qg*+2yW^I!Oh$t
zxVakyw{U~tmTnN-$_;{ByFnPH!ypW^VGxG-FbKn97=&Rt48pJ)24PqagD^rzKp0^o
zAdK)45JtoZ2qSU?gb_6Y!iXLLVU&)7Fv><j80DiNjEYeZM&&37qiPg{Q9TO67##y)
zjE#XX#>YSy6JsEZ$uSVd)EEe3dJKedIu61(8wX*WkApBS#z7dD;~<QyaS+D!I0zGT
z0)z=R0m1~I0AWH*fG{B^K$uVyAWY~95GLs)2$O6Qgh@UL!lal4VNy<lFsUX%nADRX
zOwlP2rq~n+Q+x`9DKQ1Yl$-)#N=<<<rKdocrqduyvuO~f`7{XAVj6^LISs<Jng(H7
zPlGT+XF!-?Ga$_H84zZ~3<xuF280<k1Hz1+0b!QTf-uWwL73&UAk2zc5N72p2(xMy
zgjqcc!ki&y&LBMZIsJ5dRl=M>VyDa*6n4g(L1X948LZg0Im2q~qB+AlPY<>AHvM$7
zuzRS*z51t>ik}w$`>6D+czHQ4uU21Pt@?TO3g@EN?%wv%U&i(>tf+KUReQC#T6;U^
zhu5~Rzhc3{MZX}6m-sHda@jA7{0j0T{d|h7%l%iZ3^2}r!;q!3b8^M1n2hk$X!`#>
sdsR?yNNAXO{$_*u@15fl5@kx7DLF7M$+qgM)oW_J3$Mw$+OY0_03X&<<NyEw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward_f64_instance.cpp.B99A50DCD60DF309.idx b/.cache/clangd/index/device_batchnorm_backward_f64_instance.cpp.B99A50DCD60DF309.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dadb38ffcb2b204684c6619e9880983a41d7c434
GIT binary patch
literal 12648
zcmZwN2~Zq$qQ`M2hPFe<Mq(0<p=oFudYWmPX&QQZXqs~%Vh&(n2AJW7;Xa1}X1Fwl
z&1Iq+*Q;D(VrkSolQl1H+|BdY^@vB*tVR=GJo2K?xN5TdG>N;qQS&ghv;1mnsiLR9
z^Z!qO{zWnKuYz`zm6kS3VPVDXWr2RLzhz-qSXk8D_lV!u{o8B8!VX4+g$3?E<f<I}
zK#lzCmDrd=?+L#Rf3^5mDW0KT<N2IRZ&$GIuVhx|m94w)@;48iSIa)y|A&K1$i5xR
z^LNl?wnYUCnx9<b5!Y*#Yc8Ds;)n3^GcR7+<sTcLz8Jjw^cxwg^dnDqUYI)dfu*VO
zhZ9xXHtI)4HXr$@xcbBCdq0}`;f7nExN_|H*?TSrC!VceBbM!WWY@~q{oj4}_m-I}
zXO(B4OL_Bek-Hvw=j4IrpQrCw@xbc8oY?3byHz~9@!Zau2Nro;CokUhvVZ%r-5rm+
zyI;FA>c8#|Y`pT_1MdCt<4-5Q@Pl)~?e)DA$N#cy?d}zEdo$mf`uY2--nO0lWOBnm
zbcFYUDK~u8k;ymqnS*{uY4}%<$!_Cf^Bs3CD!BYsz`OSGm4m0Bx#?Qzz{`6dT_2HR
zE!^_#YjHUjUOIKh=4Icz!YglnY1?}fOOyZ8X9^@fX@BtW|E~Z3`_o;$j3KRawee>|
z-`XX@$ppVW(-XDt;?6AxAL_gD#68b%^4-*a{_I!Gs~^7mhwa%<A3vGk{_vp^^M(rF
zs&CY^E2}zNUflVKW6{BgWrtrsS9{xZ!JZ2bY~Hly!Mm?J?5a2W&Y$9sG{5-k7oI&`
z(P1x(B-g$B)Yhz<k8V#X@-!UTll$vuH*GV2#K%wgjYW<-ZgDK%crNn6v5gn*tZ)QP
zzo~uu_18~+x%k{;KU(674qUhD(A&okX9O-cR`x!rjP*9ZzH`f|LoMUVALUu+PVeWt
z|C;;Cv5Luvtfyq_LGLqd7f;=Gvv<$lHuu?;-^D3O>sGVfyJ|k&#@>GAMAyPYzi=)&
znit;Ve(%43b*4M&{-e|rTggh}jgJKBb1BACNzo;_TM|E6p7&H#O2PJB>Cx|eW_zep
zIPv2#_D;3;z;V~U*WN8zY8FqvyL0*du00XIymevU0;lQo<+nV1^wj%1oN3mUL*rjB
zi&+x;=H}N<ogEsIcAM||(z<2zvbO%wPoKHG?`h97*9*s<8Qr`8z|I$|V`(+&)eHa8
zVP4l>`KP<`r;o>WTDBNdEV}9V4Kbs4u6gq9ug+v$=e*0(WBF{K>El12=^1+7b@V`y
zvxt4i^W4j}H#gk0W&GyEjW>ProAqlyyR_eZ{4ss`v6I0aZ!LTzpBVaZ-#aZ6-$cH~
zwdx6fI~;Sn=iaY>_klHmdt%ecOD9~x-1RdbJ@l~6x&4DJt4rgz$7&aU`SMHm&1|@(
z=j(F?(~kR|>#Azqy75nUxl1nY+O{D3<#UfsKDy-pHl_dmi%)Y*`(L=NEFln=aj@z6
zKmRbY`P-dezOo^3{Er#?f1i9`zJKXkZ*cz;)*XN5VAx2ozjfE#`#AjO<>_DDrll`l
zP~`}7uOL=fuO8+`u6{-QefQ}NtFx;XFQ}Lsjv}I>t{y_ehvWCwZrl5LVPsh4-0&@i
z@LLSFe>{Yy9F08P^t*&3%;E*rb5qt3Yt*ZU(D09E7ws+&G#*~OpnPt4DY10X)kA1F
z%;)VGi4o@4xgg@|`nJDtdH24LwZusGX2bQH4ffDNLhD@^adn;Z(`F6pW({?rX+OPe
z^=5->zPB3IY&Fz`x^UC^u<tJU-(RrsU#`E<z1492Rzpc>;#%t+gDcaw{lXCGj#8tF
zLwxnRW!&dGa>M7I-e6d?!LT$m|Gu(^lM7wpUpGg(qlxJF&`&?o9Ye%G#S*blaYP)H
zkuXBV6Y)?KK|v)D2~adaLnRW4P)S4*6hkmj$wV>~OR!KWL<&?YkqX5T9F&PLLGc6+
zl}4mNnF%wLKnPIjL^_m6h)@|s29$-cKuLrIl}Thm$%G7*MPxx`6WLG-p+H#)E0jv8
zP&UE_l|$q}X@mxqOXNc7gbtNQ<U!>V`A`K!0aPJT2vtNBK@}6lP<Fx&RYH_Nl@g^;
z<z#t6=*=7Ht{^L*D#=PH2kC&SBCDXP$!aJk>4b8TE~px^2C9~<g{mX#pxmSz%0qgf
z>dAVj2C@OFk!*x&BAcL^$!4e)vIVM@Y=vqg+o0OXb|^3Dh3X(XpgPG;s4lV#s+;VF
z>LGieddXg>KC%z0pX`SkAP1lZ$w8<gatLae9ES3dKBy6L1j<kPp+?D3s4;R3YMdO0
z3XlP)AQ^<3ASa+E$w{awatdmioQ9eqXZ~4x?q76I#_W;3bDyxYhL~AHTIdt@@pBu3
zbsyh<+k!~<iu@J!`A^uj5o?!*exE~=4H5s?twI~c3)OjFG|}(9GWY6DhJ~99OlW4L
zdy8S=miZ1{X>LFIX-^BiKJs7ohqIROOV=)qzV+%dm&_WXL)Q!Kdyj=jT;1^IFAz;d
zC(f@rnu$(<;-b@_gy;+?DLM;EiMBy$(RuR~L&bm)!^D6P!^MCQBgB9ZBgKFaqr`v^
zqs4#_OT~f^%fx~Z%f*5aE5w2jE5(8ktHgp3tHpv4N5z2<$HajU$HjpVC&YmeC&hsf
zr^JB}r^SI_q>LaK86yZr&Ip20FoIx|j35{lBM3&#2tqs+4?;W>4?;W_4??^U4??^Y
z4??^W4??^a4+2F|AW#ei0>x1vPyz)4B~c(y3IzhCQ6MBx2_Pge2_PhJ2_Pg02_Pg$
z2_PgW2_PhB2_VoE4Fb*3AkZ8Q0xi%W&=L&-t<WIQ8Vy1sl?Xy2lL$g0mk2_lkO)Ge
zln6qik_bYgmIy);l>|Z(lLSH%mjptRkOV@KlmtSOk_1AMmIMMrF(5Du0|LV_ATR<0
z0wXaXFbV?#qcI>PQ^_DCGsz$%bIBki3&|iPOUWQ4E6E@vYsnz66bk~&upqD;3j!;!
zAg~e(0;{kfuo??O3Y7vv3X=jt3YP*xijV?Aij)FEijo3Cik1RGDwPUCDw7IADwhgE
zs*nmos+0;ss*(yqs+I}@M{yu<3<m<oaUgI42LdN?AaDu?0;h2xm?#qnCdLGUi8Fy<
z5=<bNBohcG#RP&$Gl9TUJP16)gTQk<2)w|9z)L&`yuyRPYdi>PR2m3rOd1GjTp9>z
zLK+BZQW^+pN*V}hS{evu$_#>;F@s>{%pjNrGYDqM41!rPgJ9OoAPAHIg1`tM2%G?d
zAP67`k^q9B2p|ZW075#I4njJU4njJY4nn$+4nn$=4nn$;4nn$?4uVLDAc%|zg2;&=
zh=K@$D2X75iU@+Ji6CT986adZ86aeE86ac`86adx86adR86ae686a3F3kVj*0)mCJ
zfM5|UAXp>|2o}Wxf<?1{AW;$s5+i{iaS{lUAb}uB5(ttafgouT2$@tS2$@VK2$@_a
z2$@1A2$@nQ2$@PI2$@<Y2r?yuATu%uGADx|3o;0@B!eI;G6=FJgOEjKfsn;yfsn;z
zfsiF+fsiF-fsmzSfsmzTfsjpQgOJT+gOJT-gODv`gODv{gOIIcgOIIdgP>3f2nwTs
zpl}KZilBg?ND2swqJW@i3J6xp3WAlff?(yWAXo(}2v*4of>p7CVAZT3sFVtV%BUcy
zoC<;}s35443WBPrAgG!Of{n6)U}J0`*f<*qHo*phO|pStQ*0pEG#dyxR1OF^Ob!S+
zTn-31LJkNyQVs|?N)8A)S`G*rrGcO^8VDMvfuIQ*2%4mUpeY&%nx=t}OXY%)%jANP
z%jJTQE98QZE9HWatK@=^tL1{AQ#uGbql2JxItaR;gP=<~2)d$!pldn^c~l+<c}yM%
zd0ZX{c|sltc~Tw-c}gA#d0HL_`BXj#`Aj|t`CL8-`9eMj`BFXz`AR+r`C2{*1ylhD
z1xx`51zZ6L1wsJ`1yTVB1xf)31zG_Jg;XI3g-js`g<K&Bg+d_+g;F61g-Rg^g<2s9
zMN|<8MNAP0MO+aGMM4n>MN$z6MM@C}MOqOE#Z)l}#Y`~>#auB6#X>O%#ZoZ{#Y!;<
z#ab~4cFGQdow0*p=j<TZ1v?0K$qs^Dv4ddO>>!j-B_NbAB_Nb=B_NatB_NbYB_Nb2
zB_Nb&B_Nbir680tr681Yr680Fr680_r680lr681Qr682k<xJ?WSCQ^=ww#0F%LOR0
zT!NCz6)3e_gVM_j=Bt9P0HK1d0HK1f0HH#x0HH#z0HH#y0HH#!0HKnu1fi0x1fi0z
z1ff!_1ff!{1ff!`1ff!|1i?W&Kya`Q5FES%1c&GV!67?9aHtLt9J&L9D!K}UDz*xQ
zD!vMYDzOTLD!B@TDzyrPD!mGXYPuSPYPK4LYQ7qTYOxxGYPlMOYPA}KYP}i+C+!5m
z$vQ!B@=g$(q7wwC>;%E7Ize#iP7qwQ3j`PI0>Q<*KyZmJ5L~hg1efXp!KJ%EsG(~>
zs9|eBsNriss1a*GsF7<xs8MS`sL^XcsHJN`sAX$GsO4)xs1<8LsFiC$s8wr0sMTvh
zsH5vZsAKCusN?HEs1xfzsFUkJs8j1esMG5}aMNxO+^ib}H}3|)ExJK)%We?dsv88i
z?gqg_dqD889uPdd2LzAk0l_1CK=7y@5InjEgnGIjgnG6fgnGUngnF?agnGFignG3e
zgnGRmga*0+ga)<&ga*C=ga)wzga)|*ga)+%ga*9<ghsj%ghsXzghsv*ghsIughsg$
zghsUyghss)geJNPgeJBLgeJZTgeI{GgeJKOgeJ8KgeJWSgl4)Kgl4uGgl4`Ogl4fB
zgl4%Jgl4rFgl4@NgciC5gci01gciO9gch*{gci94gch|0gciL8gjTv0gjTi{gjT*4
zgjTT?gjTr~gjTf`gjT&3gf_Yjgf_Mfgf_kngf_7agf_Vigf_Jegf_hmgm$_egm$(a
zgm%6igm$qVgm$?dgm$$Zgm%3h1TXCc!OMC<@bX>|yrLHbuj~cEt9n83>Ru2!=nfD%
z*bWdn_zn;{#10TT<PH!z)D93j^bQa@=}r(j*-j8T`A!fz#ZC}9<xUVf)lLvP^-d7F
z=q?bt*e(#d_%0B-#4ZrJ<Sr1p)GiRZ^ezy(>246Z*=`WJ`EC%p#cmL~<!%tV)ou{F
z^==S)=pGPy*d7pi_#O~?#2yfO<Q@=u)E*Fe^d1m;>0S_e*<KKO`Cbru#a<A4<z5ha
z)m{*K^<EJA=spno*gg>Y_&yN&#6A%E<USDk)IJdU^ga;!>3$IU*?tiE`F;@k#eNX_
z<$e(Q)qW8A^?ncr=m8K0*Z~j*_yG_G!~qZn<N*){)Bz9%^Z^hC=|K<%*+CEn`9Tl{
z#X%4T<v|bz)j<#j^+6DZ=phh>*dY*x_#qI6#32xd<RK7-)FBXt^dS(2>0uCt*<lcd
z`C$--#bFSJ<zWzp)nO2Z^<fZvv=0Oy>jS~Z`#|uCJ`jAe4+Nj;1Hq^JKp3G%Kp0_1
zKp5diKo}86Kp2rnKp0U+Kp4?SK=9Lk5d5qk1V8Tw!7utj@XLM>{Hh-WzwQTNlpY0P
zlpO_Olph6QR2&6iR2~IkR2>CjR38Olj2;7Fj2#1Ej2{DGOdJDYOdbPaOdSJZOdkVb
zoE`^ZoE-;YoF4~aTpR~sTpkBuTpb5tTptG^KnFkwumKPPd;o-i7yuz42S5m@0T2Ru
z0E8eN1R=-<K?w3e5Q1V5grFP*A*cpH2<kx)Cg=$eCfErOCin>uCd3I4CgceaCe#TK
zCiDpqCh18KCfP|4CizJaCdEk*Cgn*GCe=w0CiO`WrsydUrr0SEruZokro<@_rsOFQ
zrqn4Art~Qgrs-)ArrBu_ruk_Qrp0LxrsZi6rqyW>ruAtMX6P9ZX4n}JX80KpX2cl~
zX5<+VX4DxFX7m{lW)0?9gBZHc>8IPXEVBj~JA2lkV&}{nbnN_DLlL%p*02h@Y}T+Q
z)IM)t%TG58kNDa~R?R(L<KA=cwB7c=zVqSE8dq&ynDd{v5{K2ypStEx>*h~u=TF=E
z+xxoaa&^brYZitR*OH4CM=rT;>CeidN{UmXBFc)EMPI)>CRVj+xq4oHfw;n)ldPEN
tfA{Q4V?33ZG<P?{+&{L342vXZTNAS4(lhNVR<63yv!HZc;SGj0{|`}zqp1J@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_backward_impl.hpp.8B916A92DFA8139E.idx b/.cache/clangd/index/device_batchnorm_backward_impl.hpp.8B916A92DFA8139E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7caed0647c0dbcbf1131e80e8d7417373c0858bb
GIT binary patch
literal 17592
zcmeHvc~DfxxBqnCMHnyR$RfxhI}8_az(EucQB(wR12v*z+=5S|=!?rYQR5OtB*6tj
z6cN;jD6$9$nux@(i70{_?j(Rja91K2H`HI_+@`xKRloN~URCm{LX~swXU=E(_U%4>
z`t<3(h6V)$<PDd}hJ6zBMP$UXS*%PZ>xup?UAAO?4+O?3WwI|)YCeiwT<590HtzRL
zPo{a-Jb4(EI>dMJSMi0eP1(VIJv~NDi?aXLD&xk~Bk${NpZh<sH@SFV=vSvapL+Fq
z`eeoM>Fq6N6IaX@gS`6Qd-R`;dOv@4wb;18aYnB`VY^qIDUfBXJhOD$Pqt~@PPSg$
z(U|f_N#fAqk1tx3pSb(1s%W9AWpCbt(%kZtH5cX&E%b69;}_WN{VumXmo@zg7F2IL
z@A3XX!AW(jy{FwO^~BG{)P8%>$obRs75N4>d*2?of3RAwu6e!r^^Q#5czV=<V~cP4
zE$`R%Ulw_VXALXvFJ6(MH>v)&6?Z>dzO6EL*5=6bI?rF8n-vA#WtK1g>Oh<0sLkW2
zPtWx|F)59I`*B0gr2f&yKQH}Wh>ULiX@f<r0r>Ja{}Yw*>jz&;c`@hYmio;VtK5?>
zM$XH2*!<O0X2G~avto_}Fw3ki)L-(A__*3X<n`4wbM@Q=|Mntp-IcXrU-mIOx9DiD
z!IgQ}*~`XLW<CvRdCbh>hRg`<V-o9hW7)RN>vuV@?Vc;s6$k9?re5+<Wh{8}?7N)q
zNh1@lgeiUua9@*`RlL!Ak$q0qwrd?@%`?8awsP7ir|<U7H~wAj6|gAm<2Q?<7krs^
zxNT(Vz`VwVW2!IS>HU7gXDQcrg)N=3tLCcB^pG)z3FB+eOqoByXRT_>^PkOE6dmll
zFLKP<mSLOfF1+<UR(D|b`ENXX9x@6&{mraJ>t5YEaVNyMe`_6IvmmzU?IVkela8L-
zmN@FW%g_I-k_mMMPFHoZUE}NfW#1lG)x)7}{M}JogJ!F)`&(OQltt_AvCG&$IOrX-
z=c6+N7B9$n-y+*GcIMg#ucJciRLj;)4f)h(>hSx8&c42zbi4!RHy@FQtW^(HoWJsU
zdT2rTe;Ym?S{!6vnWUPRGCyT@&6C+qV<%b*P1zo`?magenm)C14veoIG&ki=#Hcw-
zqITc<{&HSQl9-b<JXH79QitYYn{sD;_uID#kG@ivgc@0Xa`Vci9R&poKbzToP5f`V
zS?xW~&$WtuV=$!GnmaDv6|GX964|v;1>-y#x~ux{_bO><D`{CVzc{3=(W7re+b3_r
zPVcEtEqZLx^sM!V9bnAYjTenmo~EZA`El`fOUL^=PM=)8wlHn&PrsCH-TLNjOw#8;
zV_P3ASu!)>4<@3vJ||~t^4Nr(_xzUn?+T0f{j0cie(vNkHMOs7KJwpr{7O}~j`>HT
zicQ@3&j(Z|R=n5Qm#p)HdePQ1trowQt^G8e`C@+Ax^*_!#=59JFZZ7)iz&IbYV8t}
zQ*}v^k-M&q%wJJ^%kp{h^>^My7pm5Lx@zjG(1N3zibfQH4znX3y{`m*(V=_VKC;sO
z_|V{;s<*f1nz<wo>3dkWAlCfGhK}fKWBmU6S37@Ozd+$`%+jk{rp*9m@6Be)mad4L
z{id5t#{3W@k67sVzB}w?Arp&%ycn2BLZ*ZK2m0gtO~18pQ$les$n=qk6+ou~C?(NF
znOfpwC$F|qKwm`3g{HaCQWBV=8ySA&;j3M9ji4^NQIJKD%ZCrF))SL|zoheBnm*J+
z>dip884QvHrYM&D{&vpiGhQiR5V}bfR8&E0Nnnae@UaWF1s}~(z)^@&2^E#lN)ni2
zvHYlMY5cxEN@!py6Pr0jGuKlRn9`^F`m#?eGG8m;7*t9cv`K?5lE4((Jwx1Azv;b3
z0j&`w9onQLidJBX{rz8S*Q;k;Q9wIHDS&+npqnHxrGIsZZ|eP1Aqr@WC`C{xf+mu{
zlmX7h!ybxXexrZ`5al>DIu7k6fhhy4pWEF%>9IfohapNCG%tg$lE9SVra!9BIE<Gm
z;7~-#fxUB}g(SKtcPmSG-yHa}0uDlyqtNgu>@5jQ8TG}x`M+G5hA3W$@)Hz(LLHzL
zm@<ZQTy(!p-(CSt5hVk9XTXV)z?5;HPWO~gUOrd>Z4jjt8ka&xNnpxk+4WoZ<g1Q%
zhla=)N`SlsSV*Fa@}+lpuWe)f6)+H$at<oaL0d_5QQGx(?Ef)3yF2WON+|{MQeY+t
zObJ_+pAq@f{%{2}L8W9t?<_b@5|}ctLit;6OmbfZ9Em73P*H<AKr1li)5~wdb}L&8
zv?g_wQ5<D@idumwi|l?`lpS7hM;|IsT~;br`hNJp>eaxs>h$!}tqSOeZgK=#9)Zr1
z=%P5BS_$gyOL*7~m5|SL&u8={fhik)7~OQVG7(j^H!39?dS}C6Nnpyxmg2L==b0>4
zKr=+igx;AjND`P5)5qS^EqS1~0uDr!Q_%Dj>?;XO`P#oEx!Sbv_e#j4x?F<YFF`X&
zU`m|lnfWnO>@-TKgD8(U-AA0cB)TXep8dzIY=}}oe{`clXjBMoB+*5w)SoA-ei@^H
z!H9AI+8jXbq7|6(?SzW9XZq7ceW*k?3RVW&eE7iXT|sHKLqFnv=OJq@6V=_+79T!j
zx)31O21Jlz*4D6~G^|bm?a?htpm7N@N3FnRe4jrt(&j+%Tm>A0D0$E#51D{gU`qNq
zv-insta}(ieN>A8ivX7oA6@EO1ugq!7t9sV3YAa}&C8*OBrql0FZRx*jH_c6&=XNo
zp<XH)Tw2jZS-IiLh=jMn3h0d}KR}BgP(8E)Q}X7Fc{?ZT#x?~Uj41igG9UQ@t-zF`
zeMfSK@6~fszyL%!4;AO3og^@&B*FZ8O78Mu3fLb}PQo51p|K<|<qS-WD6LBvser=~
z<qTAuK^{dbFy+!uf%Oivk4Gq>fZF5=Z+(T=NCH!?F5hzbL5c8`hYX^WgYM<PKoXd8
z^K8C*bC3U^hqgsGItXnJB2TLoU6l6|ZqL01e^f#am2#6;-sCMLfhlzfqlQ&>*qv0s
z(WsPSXj}{(B!MZ7gZ~r1&B)7M3A>|GvUoO&H<kpZ+z!bsTjE}ptbij>DW{?0G&Gk4
zrnH0)EZ8@!*S8AjgDA(K#W8dvt-zGuuUyo7J2y5|0UZ$KduaGQDn%<Wr9(erTI#`_
zs4hmRF0Y~PYednCF3Ozb9y9)k&NPB@L<zJAbouas)o(+;T)HIvaSwgi1F2I1n+ohC
zfhn>j5p$QW9e`qhC98bA%3J(cL3|8@AHz^Qz6EhVGj=~S30FrDvsk+<))_lILA=Kf
zy~hS(jDrb5E`<<y^a-e~b&6px+zkR6VLBzS7an{9ngDc4VJ|#42;x=7>nbx67c7X6
z7{5o%7=mM2pIA1C;9acmE;g9p<E+VX)`8$k)})fg>70K$H=gLrIKwi|24fuZ0J#+C
zbQ!7?=s-dI6)1iM2AC<A!ksQ+QGcsJBXUmif;h`yR$uI;1o1QBGe?3p0AT}g#Iu7S
z?gYY4;E01ELCgkb*}xNbrXU^%!g1h;F^;L^@(-~TY3&8}3p~h8-T>b>U^K3&Aa($k
z4&X!ZM(Dl~4k0)iT17)O!CRo$78pq5@mt^wg4eM_*0JMo+X-S4=a<Bd#TY*fBbVYi
zQq59ERmx1nL_w@zR22m4xy9`Ibh~?<t{{dC4H-$EXTs146q;^1-G1!Av-6lR(+EMF
zF=9qALGyg)4I{ss?>nDD2Z7%~FakHFARYpKhiIG+eDc9itjED8vIuSDi95Q7$rXRM
z+WG~$tQbBvd>p<vE7rmAIyfB<C01;J;SDryhT+XLz7NCi)A%u*_!v$izi)*TTWQ<|
zC$`bJhIOc69dTnSqvsqdetoMCBZy(nVg9&Uf;h!_3WaikRW2BS9}<Vp=pnVSGk!>u
z>qm4t(tqq{()n1bI_N~IRR;&xK|g{UpsIn^H^9LSw7wa-G(%UyZ-)NOa2Ub&q04<*
ze;@kar}dAa%VS#q82Ue^^{vpm75b3!+MstEjW;pQn-~}D(O7X4GjtQ<PjDQgj$_<t
zJR*(>AUKgxC(`;vW<(;b-_59Z)B4@ah~2b)Kcn7H>-RGw_S5<dMx8<H?{UWWI1{{T
z!4mb5cImQO3JdX?1^Js6Ahdv9WO%lML9HNw-~?!!09^^*4sEy7I1$<=5-gWiKG+m&
z{Eikr8+cLvuPKOsuPNXqC>r69fbfX!@kFRjgkHFZ<<i0w_aI)FqB-m~P}~Mqgs5Gm
zVq)KMpN#C6^?9rjE4BmocA&w;PCT|9j3@XR*Z&ze00;f7*v5IbaRUi%=RDhK{DK?#
zf(s)0mt5dWZWO_<xWHF5e#eRLI2WRS$Bli*jVJg$=lq^?C3rnQXgxoe;0?U*2Hub0
zDBd@U#xeYW7~YHMV|m|LTECT_yp^9q^zr=kczy=KJ9yt6G(U--o<!@v=k31d9WdX$
zWlxUT%!<W;6>HcbHSAFAwfhC$`G)U3#S={?FF4N^bc?>^JYUlI73cYi#_M_K^*q%=
z8+hjpG>+n(qiDR9_utA7!!=^Xc-}vr#z}lY5<e2_J$z5eZ>*SVh>Yzd)Hw+ic!?mO
zbsu{P%CW;mksH*x3KbaR6h}8f+z6E$p*gl0LEH?Ln`s;eC&s~<M4te?6CmZcwnLxo
za3s+uLZ3u}<x=7V8!t|rpc+2~Mo)n~`C%M1iGwPP<x)BY|8SuFY}u?HGnTWg7#SQn
z8ix<8xX62v4?)L3@G&qN&zzkd=~KfF!W}u#KheM7wRMj_^)OD>pux8j3Ok_%sfIRl
z(^ZXIDO}_eV@n{2tw7ld`e34$COWjfJnK19Yy_H(U@VRvz5D!LwSb@XyOI^_d7++H
z+?3Wuo!GOUA4qTmFEr5l2Hv@Wcct|s8~7lC8+oCT<~Q=rjWoZJAK6IrZ}Y-!ntz-3
zxXpW#@|t*|iPksqZcThYT0fzQpG4zTP5f#a?*~o%0YHCKg1#y1`eu;A_tcC4yA`ON
zR|~7vxQzrc4hV5Tjf2u5i+!1$r!Rl~r&h)pGxYmkfZ`XRkDFgE<;`$E4R5r3-!}wq
zK_#qX4cD=L*aU`42{lwg6HqkKO9gFq4VN&oHpM4->WHO=f;d|}+Y6hRAbz6$ghCPO
z2nsDxE$K&?%T&uKv|P2ELaW_Ydy~hH0;(v`kKkyail*@<pxQ*^IN%lsym51(*dC~o
zKtF<$fhw8CDL|D%;~XI5fL?f*2x2Y}a%r3ggghD-0#zaCNBAdz>I998fU1aKoRLIP
zPYhJXKx&Y0fyym3PJn?4a60ZqLEH|5w!@DJPJ}^;1k0toChjAg*F>Fp11N3)%6n*&
zoS2ALu!y(=3Ok5@lS?^HOdRjw_)**oV=rVk=NaXBW+1kk3A!UrMn)*NBE?N!_a@&1
z8~G&e_4tRQe??11aW7}Mm+P7}KAN?CO7^}wZ=`7E4Vw90_!l^jsw0=Or_|C5Sd{|J
zWGO`|pnS1T88pM*NkGXdopNZFF6FI|4`uwGGNHJRf>_0xRMClmL!AF1ZW7U#bB5(y
z*J@<4l&wW!t2R}OTLD)9)li$DB{e(+bxy(FI2$R5*BGyB%qVh)Cyd_{hFZZFvL=PB
z6Vab#P0rFdgY(beCJ_B8&hQjxk8zkmrTfer6MNKelWZsXk)f3dEHZ%?Hb|T>MzmW%
zaSPaD1CmQ=WBg!YZ+5!36m9VMhb;@k2PGS%a`zVfX<nNCqu0W42RmCdlFETaIokV^
zQgfXXyv-%p7f<K{%0pSahE8}zglvv8Y3A(k$^=EDob4k{jUxv^%;XI-c`KY=n|&jo
zF?IY2H8R%YoZE43Chh^8Xh$}pO|Rn`%(pEww9-?gqc%Sdb&f+rJhclbTdY$E4dbLl
zGA;>~pv}JH#nHlq*^a|+kNO<l^BFLHMtW^wS-Mm5?&s6~(p-<7DN1m3CiaO17c#0s
zh8hzm8P!RGzc|tN$Ac!R?@`w6E>PYDPS}(L@jg)Ar}0Ced`RQRK>3*9FUoJ<S;=0i
z=_822eS<Z`bVv6cO`%lamJ0e~Lq@4R;I^N}6~Md#Sd&Ve2Ii+}TnWr8X?zBlpP}(t
zV1AayRlvN8#^-?fIT}|3^J*H`0P`9er$N0msE^mR0!sVq9f10A(vE<Da=>~Ap*~Iz
z3uuv}mk#yusz^Y)PI?(oA3JVjhEOjP>f;R%l*)v9Sx_GjXq0M$df8AP@1Y<)Z=1!d
zu-PI#Z<|fy!@TWb8t3q~IW*4YZF6ay$J^%7_y})%gvR;2Z9a_)c-sPk@eUCRXB&W`
z0hp4GY6o`hzzt)$v~Pq*4;}+hF6|Z(`@F_mUE>{bnjSwn8fJAsQ3otZiQ26sTq15l
zM7#+UHvtvGYB!WH5&LdLOo2iQq!yamZ6!>^z6A|j?Pe0LW9%p<>hZSs1FZ!yz$m~R
z@3IJDpi!WChm-*q#2}*}bDRrBgTN!WD^3^0(H^5IlnOjkfiFD>JonSM0@zfLktK+y
zfz4?eR|1<#8lM3+XJ}jn9I617TR8_D&e6CUI8@WP1~}BvI1L)4p{-MC1G>}b4Gut4
z+UN}qLQ~r44bq`0ZS)2i(3CcMgG^|Od#uyw4YHsqZS)4&&=hyQAg*V0)-yebZ?9&R
z)vPggteu0xDT{Zb2Zd8MjSurqhiRO{JLS+gk9W=EJ;?8m@UBN_oX@-F)3|_lEg(4d
z!SHMI<}T^@$1RUHKPxPKI?S^9-%N@4zndx9axee02}2)l`_s(5HGfR=;a_{pkRq2g
z%w>mQ8^v)4@+I0`XY5OGl)!Z+1RS3N5A5Fr@eOc!1G@5MotX-&jU)pqh+Cl37U+rj
za%qzqdoMgXk!!vOgnPgU69q8=suSQqjOEhSHU44z<$JAP6|KtoOVJ0O?m78~eQ*4<
zXvYmW6$78HEjd&dWdgLpb?ICHDYw%&5h@c2-hNvw8S-=51Ju5Cyh$Bzhu0L_U(B4o
zzK8i$6exB;+YYEE&Y}ZqI-noHM;X_nj2G@00R?5jN14e4w{Rvcbl}*+dA4x=IFQ_N
zHZZ+5dW{R3J&Jk5V&0N?jcdH&HQo|WT=<bu3v}-92ig56HvYu0H7Gh{l!Xkn2Y8ZE
zo}_Upqb#Ly1*5DW81GG@`+o<8@934B_ILp9#+2R8O<$HRAMg(g^>@F@-FU+|)&h0Z
z63voeY$t-aRI`*qQ9u&~f^iJnxxm*Xfnc)0-vu<gXj}+1g&-Kup!jG6YNUFgs3%Ei
zxpY7RH&W`F*lpE{0VDpBt)~sEv+yr_o%}zV#DNsuM0Hd}2C`nYquL8w2#y~t<sVLG
z;IGjZSwPq%YKv^xlu&}g2ZxU#XvXLn(+ElheyLzAZj8>D+HXIN4}yS$U=q<E0s)6;
zoC|z%!6>4y0_rL-n2gbLKz)wJ)j(ZM;~Jo@q46`$>KVQCYU8ZiI5+GP1+ks8Zm01J
z&iVz7UvkzjY5a<_ensOL-X?~3C*{TRs#v}&?Jpqb@4A&AMD+2zYdnp2@Twg&KZ$ow
z;)f9aUcT2}-k#t*-ZPIMK?eR2-t!2J^LfvF8W-@M1q9>l1u~aiP}l|SaKp-_vp3i{
z@!1<hYzIO+y(6MMfzw6&qKF$GnHtbj5F@-J{K=S#2#N?HXqjdi6>~20Up9*9mTQ(%
zy5;`MDP1(sL<1i(7NdcGG#EwjCZO3w>o)=aO|(84Xp(^s8N12AKN*Z7I0a}@XnhLs
zPoedBz&H<B5_3EOG$()$!9_q*L~yor90|4B9iX^FkA3Z-B;0C;HQwVZcF#;i#C*m)
zpK-)%5&<0tGQ2`Oc}|BFuz9tvb5c*@43oIQc&@<5n$Q&62o#Ng+C^vwe(hik=~?Yz
zCw%K;4_6*D{^nWczq^%x^KRu2p7}dhU3wIV5>Iu!K^@6a7Cv-cHOAa85p`|Yw6Gb(
zw@&$Z%4~vi!T4M-5u0L>?Ca3l)Hg1FId@Vt$hzm9m_ItR_j`Gry}Um5J@}Ly3L~{g
z<gi8kbmr=m!#-`z|60`FY*9a7n$_eHHfSyCx<u9_k?g${FZu0Bc*n}|-C6M#G`j^Y
zaC!mlgF&-;8sCAIcc3-V--DL-X#5CTKB93ev}~nu8?<br@gCM<4{M3N7Asz0dtYG9
z2!77?eoo^`u6HHZr7ux<7yROO;1l#jPeIS8zz};iyxEEzXdD#cV3$HJoyx?<S?0LL
zeZbLC2mVr&ReosvBE0~eD0H{~f8bqt`O>*_586xqVS#Vu{ax#J-}AhL`oErcs^^{Y
z(gmHt;*A=3V;XB3`2GYp@<xrczLD27()!!H^=;mk@SFHPO}rJ2eVcec8qa9rXVEwo
z?NV;(+NEU0CjKz`Cb#Pwe6AMx9qnma+_mTG#swAklHEkcafeafVY=qV3xgcK3>{~_
z<6npLH#;P}9z|WJJ?xCv7gyDf>Ix>!Oh7BK5Fx~r#DWuq2^0!fhW8;8-VEgo3LON(
zL12ofzs|#o!XbjOjdVwW?T0he_|E7^<%idGFhLM&p`sS{!XB>^Yih~yI6-^|b>2ZG
zS;N1Bj_>F-{!zx{C}V{^8p?QZ&i6PCUb7406Hf7j>xF00PON#t`4C*pTNRV6_VwnN
zj}F#4f$oABsUN8#Ewc{juLCL)`6dAU1fU|3>~^5PoyNz2{xP5;?GwZFj$x?OL_A{=
z&uGYF?PM%=(s&PJv4_TKj71uaGZ~9af^TfRZo)4&z4mXq?|(VnccZYZTk5>)j>yTy
zuqH9AH#W2z?J1%2c7C(zA2!wCL_7*!QKswxSps#Y*VNnTGkDrp7VtpCQv&i8+V>Uk
zh{Y%5(d7Fr6uyOKI0isR>OZ`-fJ?xOVpM|mT?Jgi?SU7Dy~s8J|IP=0^L&8c7C}9v
zeN_Z^-5rwyE^#eA`sxVcKGtL(YmTj45K~!`R2pAq4KK4ML|@Aq*3$STYxt7JuUNxZ
zG(O8|&T<3D@2faX6~WDJTzj`3#<9rQGI`xh-T)g_=fYg~3U7d?H2hWzBK`&xzX4-n
zmD+b(@T1*zDLkyP$Vx>EvpuZU9@dlG`T}cyf$c}|bJqMhjVn3xO0G+fUy?z0d<KMP
zBo>xSugc&%K1(SqR_9sU{cG-jv$?k^??2k`^|;OxFaNBwtL@M9UMzMGE=H5e&#d9k
zYyh^wKUSY!)9X!XkAGZ^|A-DcZdaB6{7<KJUcAZ8HhvJf1<g0Jy=RXl@zNacITXqP
z#yP+lM?)_)pOpvZoF4LjZas{{HRM3G@4n$F4)6STlYe;U4S$V-NbkrX(kSbI_c{<n
z2a(<hAczE!$AI@S5QLklb8YP$2feXt>%{(XFa%@FL)%Q+*XhV@OMp`e2*RiBQ1TKg
zlVI103_(nW%48bvhRWRpqxbL@&O-lnhirIu_a?)qYbOI7PUAyaGVJW|zJ?6n8>d!f
zIHkID$(*I>op0^Cbbe_6`r-?j%(3ediUsk5`T3Kbm*{Cc`Z+nftKIs#x;Uz2>Mn}R
cja<5u%YNiiMJ`>*r3blmB9|I+=}a#FANi7IfB*mh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward.hpp.0E1593921379F2F3.idx b/.cache/clangd/index/device_batchnorm_forward.hpp.0E1593921379F2F3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2141d90fc1678cebdda1297d7d1d1ee86cff003c
GIT binary patch
literal 1108
zcmWIYbaV4zVPJ4h@vO*AElC7283cj2xTGkvfRTYg9w<;TXYYhq@52TnZNG!77M5S$
z@Kg5K0Yg^(IYzEuR?S&@+;q!HPS0JDDgW>8TIqZD0Jkegiw*zD>U%DK=U&(QKGl(z
z>+x~HiXH3yPy6lpd@K7%u(yoKqnM6qRykMZHZVnhS~1a?bKTQE>GIwM@&~qk@iaK?
z+Q_vv-lN?-E!BVL70rDrcJms}g}->i-Ro`2qQA*7()8dK5sOJ#mL0!lZ?*oI_PBvH
zmHE~y$qyTn8uoAHR*pVuRd{2iW$s$z4mXSGO#9C~Rl3`hx3_nl;_+f@?faASJhS?<
z%a%T$viPry((^lu%Tk1mG7frQnzMYhy5)bze6{Q0^HLk<7*3n$G|kfE;7Lp4Ee^9A
z+7=aEGITsOp*geB)3)bWlc#Q%k*nILs<-LK{v33^ubyrkyy@!)`Di27_Wv`(>Zh*T
zuT`nGn_>30MJFrLjLn3T&I^n3{f*mxTxVkX(#^hC^713}!>l)mt%|=BzT!mKMd_{=
z4Q|Dixk(Mc@aOV6$MrPZlwW{}l>^8H0&XcSDO(08r;X_)W2nQcUN$CXb_Q+<b_q!^
z!NAD`j57vCZpO98MB0L7Lh`toSbz$w*{r3(1VX{d-tdquydppg*kBguYwBCWEGSA%
zD?SD^3+4_+PO$wjdt<j-=ESI}C^PbKOUX!Sz*MF7v8px*6*~ciC6p!fVX7+*KYUWf
za`7w&54W_iv;<5w54Vi43=&UPSQd#VCoG4=lNXjp;wcC#An_E16_I#K!b)%+%mGYX
z;NbC8DeIoN%ejt++gi`s0Tx8vUd%0T`~O?816?O6t$^e@Nf{)btfVXwPhL_UiKif`
zfW%XjR7Bz_Nh-m4YmKb}j?PV(4h$cC8GUtF9Kc+{$OVohm@{G0z_15J4a_r74OqXh
z^NV*g1H)cWN*du_eL;O`m>!@Pm92GQ1{bB~Bq{+D1}uIUU}nK+2H~*G{S)7L78DmH
o7h8jp8!w1}08SW#fq@6X<bpGqxU2glL@edL@Uigla&j>M04`o%bpQYW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_bf16_instance.cpp.F51389A9E014ED43.idx b/.cache/clangd/index/device_batchnorm_forward_bf16_instance.cpp.F51389A9E014ED43.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7789255f3753ca5a2022997185e1d3b07717c4e
GIT binary patch
literal 11382
zcmZwNd03lOqQ`MwTIwmK(xPo;J#2>r4ur#YfN&r{*t!dZKnVN3FG<)}YXxm-r>$Ps
zsvUJ|ua4u|I_}!BF1I?>scXmT^m28^b$qP$KGt!(v}5Eq&vSX6hWz-xpCtT2-aJW5
z))y4?#Cmv`o7N9^H@oVWdU$xOUijy9Ioh9C;^A?`+rwjc|6yx!@1N6r&i&+nMasLe
zk9)mhzKz-YIs?Bje*5MgW?UP-F>C#nTYvcO?l02XPY*o*$V$?+_j1!-x;}k*wrAbh
zD!bvTjN(o2en0=iiaVa4ec$xyKhAx2{;YbY#^>eCV^6hy)A076AF8PS=i8-og_*~k
z*M9uvgyn<Dn`@_jEZX_tPk%i5_DxUaj~=evGJ8+a$*srBo;*K3@Z9fza9m$=de0|g
zGg}|Ebyhrb^u|}bPq6>(I3Dwq`6kC7zWb1$*dE6}A32zC`aselo0Z4il9>Ch--)kH
zAC=z!TL0swun(%EzJC7jLEV>zH<woadT-1nuP42gwd0SrbsyPXKU}GQbobH5KFiay
zA3V9QP}gkn_?7YA!w>J-d(S41Q?oDe&dS5XXFvHd<JzlIwq&lf=$}}=A!yS(Z|vQ=
zKkhpH3xBt)@+x|5<N;sd$7i=)lDP5mTQfplIYH-+MEM<X9C_>vnf=w}jaRaLv%bGu
z`-1VKTl5+Kb<i^5Kd$#Je5vNr_-$3U1^oWumfBON%sa;(X=`4!;j2R-x{rVQTWr@8
zrByLG)|~L6&y6?Eue$O-pVs|%SotBA_q#Rk?g$zi+>r3zLl1s^AoT|SV!tKZbRG8m
zfb0i$v0QugvEQf%o|)axPe%uhy6m|n`?i%_KJ&THuLozox#7x^5&dHoZ=QMR@cE<0
z`%H6%t72a)?taYm)rt?FEr07&*W79U_xf8K2D{(8{q?g!VfUX63V!R32Omy)>kiLd
zU%s}h%m451jUS1qv$cPAv-Jt{vsZOrd(v;c@qgB&N4z>el@gE>*Hm`3qv)zz-$^+a
zHC(y$-gkDF=Wjo>lAU~d>v#KheX(a}_#Jtx`n~C?YyA$qbfWyY?d<-ONjHAh*Xp-n
zf66P5tqDvVc+>pSv6&~Y-*)moOa18H>CFecy>8FlV@bDXJhI{AgTFmzsPUX%eYbx?
zrwm$YSmydo$&Uvg?wQhkee?dII|qh#4sAPj%j1T_YmPW?j+tMg^L+ow`)n^=7yH@M
z-`u;^IPVkHfB({hh4~wjOU$n(AD^AS`((DI?C7BvZhrKx?T=r5`~|P$H`HHq-}a(g
z<4=Au)7zJ0K2y6br2pDa?o4ykUGKH)*p5%We75gM&AR2g>d(ha&$nE^`%lOJT9mc=
zwpZ`j`byKEj&6!N{p_oe-}kp|c7A!o%epsA=k)CrW%K7=WV+o^bGLtA-PmA$E+@z9
z`+);l(VLx*MJME6_U-5ccYU{Y*;N7a|MAu{Cuh#h-}sTs5qyKS{zRm&+I7?W5B}xc
zKQr_GvE{q-BCGwcH@)-G>euO$E84!gzW0{2GjDh}N4guHUHD7(K7ZiDt2d^UuJ9}_
z^|1Rb^ILuKVxe68nw~uuKDXO)(+bbBg}i0imb8l(zvMH6$9HWq{AGowWg)L!wl?+R
zg*#te7f^72Lek(0&n*ktcbTvE#S3@7b6;)vrORKw&Bvo;A#c~bwrlA>e{&!5<0|VT
zr+Z)gWQFII3x}*)wrcss3wL%jw>Z6Yi@(KE?~9Lq*U=vi{Ps*nsE>U{+dQLH{BlS3
z9oohnnssr{X`AM>a`(Q>+Wy?$uh#d~`Plu5f5`ODznzagfCRt_B!RGkND!=G5)3PZ
zgutSN!U`p!u);_ftZ)(zD}qG8iX@S+7-6uYNE9p`(ZSLaJ*;RF4J(Gkz~Y3%iY2kI
z;z%4U12MpgC-JZZ5wH?S0<1)m2rG#s!Ad5{uu@10EJ-A+RFVoSjikX!C+V<^#0V>c
zWWZ8H!OA3=u(C)NEE6%o$|l*ca!3xWT#^gROw6$INFJ<wk`JqZ6u>GZg|Lc95v*cT
z9OC}t^s$$a5?H0A6qbcpV3m<FSXN?%RZhxbRgel;He!QSNh)F4i5*rIse)Bas$tcT
z8d$ZY7FHdpgH=!JVKtBjSdF9+RugH0)l8aUwU8EAt)vxJ8)<{pPTFC0kPcX#q!U&b
z>4Mcwx?%N@9$3Ak7gitXgVj&^VL6BcmXkPPxrhta02zQaNCsgIks(;aWEj>68G$uQ
zMq!PSF<9ed9M%MxfHg@bVNH=KSkoGr)_&PJvtM|8|E=qk=hB7e+{<=odG6=frS>`P
zvN<i^y*H=s>s=c+w@34_`=$Ar-RC3A7C+tJ_A0N1A77&_zee+QA74AXzqjt0r{@=f
z?`q#G+@~1#--4(2#S_%x8Q|aV&Ha?VaM*>j)#73Q{vh4Add-r5ov{{=^QZn1?i*P+
zOECXvSi(OUmh?9+t^gVUE`SAq3lIU|0%QQVKpF@xkOhJZ6oKFZWgxg98U!wg1%V3^
zLEwU95V&9(3@(@jg9{eH;DTi^xDXlwE`)`E3lSmULSzUyN+~$XC^#x8I4UW)P#Ov@
zl!bx|6`|llWhl5X8U`+mg@FqbVc^1K7`Sj64lbO9g9{ho;KF4%xCj~nE`mjXix3gu
zB4h-(NE!()l0|}x6p`Q}Wh6L8892rmI3^f4CK<RW8U-$jMS+VFQQ)Fv6gVB#fzvS^
zIGxae(@7mTJ=KHLGd(!H(1X)UJ-BEZ4KA8RgNqi?;G$(TxELA(E{4T`ixDy4Vq^?B
zPB}QvI5;jiI4(K3SQ-m1mc@dL6|vxAWh}Tj8V4?p#es_xap2-)95@3tfHN=yID;^N
zGe`rtcp48bp2dTU7xCcYWjr`R1vtS3I3WZ$AqBVungA|=C4fs13E&cB0=Pt)2riK&
zf=d*M;1XpbxFnhcE{P?9OA<-ol4KINWSR^vnI(fu7Rlg}Wiq%FngT9`rGQHjDd19M
z3OGq6ILRb9DI_>4CAd_Y3NDqUf=d;t;8JBOxHOsuE{&yuOA~3}(qtOAbeaw>ouz|I
z7wO>AWjZ(`HG(rTBRHclf-_1ZxD1*BE`w!&%McmhGGqogMHM*36gVXmI3*RhOqvNU
zlVyU-6q(>MWhS^RnguS4Wr52QS>Uo{7B~|%fip1^IFm4eGf5M;Y?=)&n`MK`7TMsk
zWj44RngcF}<$%i(IpA_+4!B&J3oe)Cg3A@T;BsXyI5RbaGcz+dvoM1*OEb7Ung=eA
z<$=o+dEoM79=Lp(4=$hOgUc8B;PPcYxB^-Lu7DMQD-Z?X3S<GeLRtu}kQIU}6oudl
zWg)mCS_H0$6@e=fMc|5L5x8Pn9N~Vg>tiow#nG@taWX7fY+PI=v;<rUD*;y`O2C!K
z5^$xo6kI7Q1y?Fc!IjEVa29F-XJHm_7GVKrkrr@ev<zGsD+5<1%D|P$GH_OE1!rYe
za8_XjXO&iP<+L1JIV%TOF3Q1`%W`lPv;tfOs{mIaD!^693UD@R17~A4a5iBBXOlK?
zm9!FEC94EiDJsEL%1Ur{Y6oX$c5rrK2WOXda8<MlTotPVS0$>zRmm!F)wCL1HLC_!
zEvms)%W7~nv<6%as{vOdYQWXV8gR9=7F;c>1y?I-!PUxIaCNi}Tpg<eS10Pg)yX<=
z^|T&bJ*x*-FY3Y7%X)APv;kZLYXH|E8o)Kk25^nE5nLl{1lK4U!8OW8a80xcToY>o
z*Cd+2HOVG$&9oU@GiwIdESkYJ%Vuybv;|xXYXR3HTEMl)7I3Y!6<jN81=lKC!L`a(
zaBZ{=TpMcx*CyJ)waGSc?X(?SJ8K8mF51Dh%XV-bv;$lR>j2jwI>2?v4se~c6I>_j
z1lK7#!F9?`a9y+uTo>yC*Co2Zb;&Mp-LxBAH|qx1ExN&V%WiNzv<F-d>jBpzdcgI_
z9&o+17hEsv1=lNj!S%{saDB87Tp#NL*C+bG^~pYP{j?ukKkEnAFZ#jt%YJYU>Hz0p
z4sZ_P0Oybna8Bw3=VVTBPT>USlumFi>H_CtE^sd40_T!0a07Gz+yEN@Hy{SU4afm-
zgLDwwAR7cXC<eg|%0X~LbO_uK8v-{ZhQJNUA#lTV7~C)$1~)8*!41n{a3gdC+z1;1
zHzG#BjmQyjqjVJ9C>sSgDn`ML%29A*bPU`W8v{2c#=woqF>vE_9Nah?2RAOp!Hvst
za1(R_+yt8dHz6j#O~?swlXMc?B%1^`DJH>9%1LljbPC)Qn*ui_roc_fDR9#ooz_53
zYiwErJFSUn4fM1or#0|18lTbP-EYcX_@FIeMoYo1X0&YFxieaUd(U`&s+DDbwrk-{
z+!^hP87;;g+}}0tpFaq6zw>(k=U0!|&g|dq>0@8#z0P;>jn%E*TYcRxbB|iGKCPwi
zzPxbIb=uPFv~c&$`Pg@AOLs2rkA>H_m-w#RdGY>N&S>l02f9CR{KLWnxbSJC`|cOs
z>gE21#e4Nv{-(thzyqKKr~qhzJP=x-3WOHKgP;YeAZWon7+SCjh8Ds@poORqXp~cE
zR8eT5JQP}}3WXNN!=Qz!FlgaC99p;vhZeyjphc(%XpuY;TBM4E#yEq<6oVGUqo75p
zC}=vagQinDXnL-PrdN7s(L5Sjw2Fon!(*Vus2FIRb7)*~Xt6vNTC9qN7RTeD#i=-G
z25x|6PzGr6JRVxSiiak+fF_iHmcSFBC8z{wi98WnqDq98#FL;UsU&F0JQ-TDN`{ug
zQ=p}&6lju5Xi`aNsXP^0s!D~H#?zpssWfQmJRMrPN{43TMrcN5gqFcGpk=5GXo@Rn
zN-1cWJQG@`%7m81v!G?EENCWff@V@CXxTg)TDHoDmcw(P<)|EJxjYwIuF8dG=4NPS
zWrmi=^PuIaJZSkmA6mZ3hgQG~pcSYBXob8GTA?b0R>X^-6{#X<#k|<yex6=<<3bgi
z7FP)`fmWhQpq27cXr-zYnuS}SS(F7@883rYrplmMxfPmKS)rBla%knM99jjhfL5U@
zpxL+$noZfDRq{$`m8ue&o!g<=l^t3YuYy*ks-RW#YG~D}8d?pnfmWkxpw;qPXtk;q
zS{<*0R;TKq)$@92^{O6P18;!Vpc<ex@<wQlsu5ZfZ-UmOnxHlFW@yc-8CnZ(f!3m0
zptbT=XsxOhS{rYJ)~4E^wexmp?W!GG2k(H^p*o;-@=j=-suNll?}FB)x}bIQZfM=A
z8(I(Vf!3pXp!M=zXuYZzS|9I&)~EWQ_49sc{i+|DgFB!(lmnWRJE1w16Pk;=pt+O_
z+5jJbHlPNe4e~)~gK7}k5Fdgzq=ujk^I>SiY8cuGAAvTaMxc%IQD~!T6xtXcgEpqd
zppEl!Xya-e+613~HlZe<P4Y=-lWG#$6rX}NrKX@wYkXRRI<2W`4f>3hI-?oguT?L+
zb}pLHHoCWmqq(d8!vBnOIvSmY7p|0FxcJwX)qTZdi9OwM-SAzP<{$F1R+QT+Jt}@l
z9=64+m5Wy!yPLY&7Uss5%}bVgExUv)U*WTImG7nNSLd6fR(r3{^;>t@<yZKprDv$j
zEK{~&4L3$eJ^lZFx;8K<I5aFga^Z90nAmuckd%@dk{l41n73wa;f5;Df~`44+NS>l
DNnqBY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_f16_instance.cpp.5B31F4805B2AC1BA.idx b/.cache/clangd/index/device_batchnorm_forward_f16_instance.cpp.5B31F4805B2AC1BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f191e56fa2764321f2d283b2dd9c64c573dc6dc0
GIT binary patch
literal 11382
zcmZwNd301&g2(Z^2zUw6unA<nS#FkFNp6;#RNd@LRTfauY?V}1_I;^rm93Ig7C{3F
zsL<#%DhwdXD1wSIjyBy&(+=&1h>8oi4vL`T)`-%y%>+(Y=QndE=TOv-@B5@uf28uL
zm*VC5`CTy{9>&Jy&d#R(x>+6`9t)=b^z}R1p1;P!<4G?M59dP%%thTNmG_0qzRQx|
ztL&P0?;8D$U5EB|KK1pn0^|K(RjiG<$Wq$Aeea(A|Fvu2y}H-*!B6EC|IIYczF;16
zC*SDne16uqri7|9_k1Uup5L`x-u0K0d%nKsFXu-R0)AfH{jaaAE9*X7KX9?&mV}_6
z7QY#C|8qYV^<Eku|Mg<G)%EwGZTm7miaoyPT}LX*dtkg`-|=^k?78&vN4p<9e{L$N
zM!J6chV}5K8~)v7c8cfAZRQUGU)b!w@8{~_2)o7drrdCL&eoha_FLOt*}UL|`<+vl
zFK)9o-Tk{?=YIBW;%(RNOU`H;x-;UK*M`!R)@?g36!jl{;g^qho?Mf=Wngj8>W4qi
zIDO*2^i=+^r}LeZy&LXW&zd}sWhPZy-hTY$2MynSoa_BW@42?^xv$)O>*uxmr|x!r
zZeMPiJHC5WRCfK*j<Ly&6)B6;@3rhacIt=h@xRwTo4nEEzGoh}u0XkVEnBVInC|L1
zKu%mdYWko^|H!%5J~a-j#0w7Rg3}+q^U{XoOUF+HSx@~YFM9Q@j)I>Qzg*B#_uAIa
zi|6k5TKK%}y{7}#Os@DMV}Z}}xv?RRz5ii8{r)p0cO9A<yW`s5?|6Ofre*Wfi%#l}
zd$f1PEco!8PJGe!-x;SX&#wIJ*uy6mZrZeX^UnCn7IHvt-hKO~SB*PfzU%rmd*5Ap
zqUhJU^i{PXceg+MNMFtO7Y^(^(cJLO(U<N{Tl7rF;cxBR&JfGiGhc1L;M%Zt$A!hM
z_Li>3*PKVbd?rwPZAHnEXI~Bc@jAQlq5BJuhdlXb-{nD<_q;{gL-zP>HhR7D?7pWB
zhu#iV2L=!7LQS1fTfQ87{9o*iYbqE2z*em-`Q=#tW{>3YKUj}GaOR<HPnWJq`^~MN
zKaun3-uLf)Y+X>I(7&^F@k8cEyxv|n%kEhg_0{54yI%d^!;NJ+L(_qw9~LgVXW1KT
zEzf+`+LN$7`o4394Fi{=Z##1O;H4*iXFIqk=IFtJ$M!z8^>C>#RSW)nqU?>tm5oJb
z?#pr=_ic&4(-t0|p)20G(Qn107kgaieb+v|V%bki`^=laKl7=h=TKr(#SUG_rVp<B
z<kMCEdFaL4({kq@`efq$mm+>>cy9ZainpHnXz9@tUx)7Sy{YEqo>TVYMI#rzUyrPp
zTh71o+ncPLU)goiu!Vi%z3=Rvtv~)U>*9mghi`bS{opp|=|8=7`_uBp)|>2K8KVAe
zNtVw$yLYv%(}z9UI#~T=*|UE?8hmTR!Syc;efZS)sUO#^+_(C;?wc*AZ`=KY-zDEu
zA77qY`^jU~^l;?w|91SZ(K|0@Jm+z<>&Mwf^+~Qr-*9LBAEv*Pz5crW<JC7Nm(24l
zD)F!_oU?Gj)tl*Z^<%-4clr(NJ9l`V=PlFOXO7R@t2gd^%gsK$XRD9Q_x32B&bMo`
zZ`ag6{&F8O{6O}qJU_?l^E^wZ4_Pv2iT>)1JO4W0y#J%_KYcmR^VaD+f6n~5S8v?;
za&-3Y2iI)dFwe7WI<K6w(s1?0ogGcheX}EHp2aM$t6%+wKVI7Y#Aj*2-nJ=i`IJ`r
z&o9$<m$u?At!!qm*KS;|ncds{Z11?n>*~{-`8>YVH`M)1yluYBHxiccO@JkR^)t(l
z`hoLfe&GCsA2>hh2hN}RgY#$p;QWO@IDhF6E`SDr3t$1@0z?3~02u%-kOqPaWP#uU
zMIg9983-<j27wD=LEwT!5V#;21ddV)jxq|43JQ)&3NDxig9~QC;DSXkxL_F!E`)}F
z3t=JPLPQ9-5E%k4l!k%}Wuf3gMJTvX8450phJg!XVc^0<7`QMQ298k%jxh#~2?mZy
z1}>b2g9~Tj;KD^XxNsQ`E`mmYi(nDpB18nZ2pIt`l174yWRc(^MI^XL83`_mMuCfB
zQQ)FP6u2lE1&&h=jx!F93l5G;4o*jP;B-s}PA7EWbW#T{nnr_*X3^lHMKri*84WIm
z#(;}qG2miE47eB>11^@vf{SIb;9^BAxL6qrPEY|(Fab^o0ZvE(E{?{5i(_%%;zS&{
zI2i{np2mZVXYt_TMLf8884oUjCV)#|3E&b$0=NX304|Xxf=gtH;1We5xI~!<PErX@
zG6_x!2~J80E{P_AOJYgjl0*`?B$)&*nI?lvX35}^MKZW#nG8-(_2BeO4^A)i;Pg@t
z&Oi;|49oz|APnFP(g03T1x_&qP6-80Nd+#2rhrRfDd18>3b+)R0xp%Nf=gwo;8H~@
zxKx=6E{&#vOJix^(nK1#G?@l2ou-3JXX)V5MLM{2nGP<4W`N6J8Q?NR2Dl8F0WOnf
zg3Dx?;4(!fxJ;P|E{kS?%VJsJvP2fRESUu^n`VQ{X4&AfMK-u>nGG(7=77s#IpA_c
z4!9hd11^{5g3D#O;BrMSxLlbF&Pa{mjLZnmD2(8Y(g@B(P2fz-1kNN(;7rm4E|2Dc
z%VT-q@<blEJedbBpXP(hXZhgrMLxKEnGddj7Jw^Y1>g!q0k{HL0IrZ0f-7W&;0i?{
zxI$S7u80<ey3d8ZZAGjo5|$`RfF+CcGpm>ugDYmm;EF{txMEohu7s9=D`6$zN<;~`
z5?KPSl$L@kWu@RsMJc#aSqiR<mVql{W#Gz08Mrc82F^^);LOYn&MeH}%+d_5oR)(t
zXXW6^MLD=~Sq`p(R)DKu72qmF1-J@X0nS1#;4I7n&LS+}EYbqbO0D3m%nHsbtl+HD
z3eHAt;B3qW&L(W&Y|;j<l2(GNWR>75MJ2dOSqZL+R)MQxRp6>b6}T!{1+JP_gR5rM
z;HpJ6xN2Dqu7=itt6??ZYD5jV8d(Fbmezu+WwqdHMJ>2mSqrX?)`6>Ib>Qkm9k@DK
z2d<vhgR5ut;Oa#^xO!O+u7NgyYhVrF8bkxQ2H60vkv4*BWR2h&MI*RI*$A$QHi2tm
zP2ieD6SyYX1g@DjgKK8Z;F?7<xMtZ5u7$RMYhf+mT0{%D7TE%>m9~OwWv$>^MJu>g
z*$S?Wwt;J7ZQ$BO8@M*v2Ckj9gKKB);MzqyxOUkNu7h@f>tG$=Iz$J!4%q?DPVM0A
z%nr^j?BMLu4z81Sg6m|R;5tPoxK7y#u8Veo>tbEtx<nVaF4+aHn|6ciX5HYrMK`!^
z*$u9T_JHeQJ>Yso54axL1Fo0$g6n0y;Ce+bxL(-{&OsgE9LxdEAspZw(gCiI_JQkT
zec<{;AGki*2d<y?gX?Gg;QB>BxPI9WZh#Je8(;(A2E+ik0XYC}kPd<yWP{)a#UQvr
zIS6iu4uKnDL*RzQ5V#>Z1kOpF;GE0}&MBPWoYDzym=1#*X2aly#W1*GISg)uj({6s
zBj85F2)GeB0&bL!f*WO{;6}wLxKTL@&P83|T+9W|C0yWK(gkjej)5CvW8lWb7`QPx
z25y{=gBxe#;Ks!`xN$iSZh}sLn_v^*Cd34|2{{37Qlpa^$VrV&YG5ZdF{y!`)a0ZF
zeoBj;(uDi{(kt(;#!qPycj}aufjetT%XRNX=C&6%e!IM<cKR(;;u{2SlM+8z{=^?v
z011E<NCII6ksw%<P*}kv7*+@gffY(ZVTF+}Sd1`O;UpYZ1c`tZNg`oIktkT4a9BE`
zgB4ApVa1RbSg|A)mLLLF9EpP!PvT)EkOWwXBoUS*5>^sPf|X2?Vd;q;mVp>xDWYJd
zkQ7*{Bo$T~NrRP6(qUzg3|N^Y6IK?<f|X6OVdan<Sh*w@mXR1?nTQEi9?651Px4_E
zkOEkRq!3mSDGGxBMv-DzC8PvaDJg|jM#^BBi5XToDTh@-DqvZN1(uaqVcCccRwb!~
zRYj^`Rg-F1HKYbsEvbc7N9thJlX_SUqybhVX@u28nqW1PW>_tx1y(C*h1EvdV6~HW
zSRJGTmYvvPb&^h4U8D<EH|d7eLwaEKl3rL2;(*mh`e5~wepmx!0M;NGgf&ElU^$5s
z)-V}{H9|&UjgnDVF5-eUM#f-`lW|xRWCGTtMkcj??wmE>cw9W+|Dors>1$ex?$V6z
z*VtLM_1dEKn#sLyHf-u%61)Ch&D*w6Ei}5X_qns?^^fay?rUA8D;Gw7wK0A6^znCS
zbMMf6+()`!tv$W2Uchr7@BM!sf6DW$R+fEk!?d5$mQ86ow_o#pUO>v5>;0yK&wQUH
z?rUv(zFs|S`Rt*)J-uy9y_Wj8AH~~ti`Oka?(e`qmS&!<q3<1;KIl$u)}30Y`;&Ux
z)@ifW&FnAYXP>{;XX(1DkAB^hw$y!~`#R&{=_h>UD&x$T@5_B-W**g7`KHY*Kkf(3
zPx(Rf=l;<Al|Qrq9sn&s1waerfzSd~AhaMJ1T9DfL8F{Pql!Wc=E2Z{RWP&=9s(^y
zg+L4Cq0mBAD6}vh1}#j5L1UajV~Rlw=i$)8RXDT=9sw;vML>(>k<cPlB(x|V1uaTN
zLF1f5<BCJmaUC?B(m{*n(a@q*G_)8V11(0yK#S$E&|+0AG{FTlp#-!z9tSN>#X*bb
z@zCN`JhTL!04+f!KuhF_&=OT5G|442sU)-{o&+sPB|%H($<UHjGBiEcL(?lgGy^w4
zGbjTz#T7KA6tonc0xd<QKuhJR&{9<@v^1UuEls6COXum((p5UN44wfkLuEkA<eAVi
zRVK77o&_ySWkJj4+0e38Hnbd`11(48K+EO1&~jBSG$S`cGb$rA6E{IKDHF6jo(C;Y
z<w48m`OxxJKC}W}0Ifh3Kr7^h&<a%{v?5*<<9?l<ersA4rOm8jUJR{R6+<iGCD2M#
z3A9pP3awO?LM!8C(8^R9G&46tGb=N+a$XLtT$Mws;1$p+R0T8(w?MNf3p6XYLbEC>
zG#j@;vnd<2N?r-AQdL5$;#JV9R28&pUJb2URYR-cHPC8Q4YXQb3$0ewLaXC-(CSnj
zw0d3-tzOkbYv2vg8dL+cM&1amQ8hwq;!V(+R1>sj-VCi-HA8FREznw23$#|=3awSO
zLTlq~(ArcRw07POtzET4>);*GI#dTVJGVo#D?7AK-U+Q!bwcanUC_Ez7qo8P4Xs;s
zL+jx^(0Wu4v|io|tylFzb8rVVhjKvc<9*QjR3Ef{-Vd!`^+Oxr1JDN40JK3q2yIXe
zLL1^k(1z3yG$(gLb1EmaVLl9PSPerP;UmyS)CjauJ_>DAjY4yA7c`f0K^x;^(8kml
zv~fNTZCs5*o8S}BCe#G9NsUixP$xAtsX?F8lBYD~K3BbR=4_hM3f$ZMdOCgPedRac
zW=>uI_nE$jqsd-(<u~K{91VR1SMHQwxw-mrjqQsUV;;!ae=2^qxuV=+^)S1OheyTC
zUCYc}>&)GT&PIFd^t`e8=4)ono^vgkJJ0*N`99Y#Uyx^vSm3ohcj3}Si<kLM|Kv{3
z$jr(XH$<m~>7(fX_peL*0|J9XLZ^Q}Y5L48E<Qmf8-kMjViQd_EGbw~>6yPWr%=1;
Fe*h8Cz|a5y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_f32_instance.cpp.DAA81C77B8C2B284.idx b/.cache/clangd/index/device_batchnorm_forward_f32_instance.cpp.DAA81C77B8C2B284.idx
new file mode 100755
index 0000000000000000000000000000000000000000..52a0fe98471217ea0a8e48398411b440c8f956e6
GIT binary patch
literal 11288
zcmZwN33Ss{qQ~(D3SOWJ6>In2w7qG2yWB3f%Wc}GZITuX#cpZZrhB@iTe_u7y81@?
zEJYZs58it!Dxw1pipSk?p2FZ$;SKVfqks&8%DBtmuDFcjE1b#n_vTE_=|BH(?mwU2
z_9Xv%PW~w`DlA+R@8xA`Epj{ChMK*-yu5rDeq2L?ozE=w@_J~cmzR6jQ`WM+H<iz~
zKL-Xpwc(LxmK`<SrF-U`Q|YINKe|7B!;=Z-f1kX4SL&q;+aG_IZG5BoSo-cqH<th9
zhH17l%4@~H9CtaN_5ReH$-mrp$v^uSt6x9#`XA4Ib9`UTxIFZnzU;Y)s8|0Z`l=6(
zoqu!fOD8`ve-UCIxU|M^l(uIy{&mmx#GhYz?fnmb92<YWu2>Y!@1_37hQ9y!hvw;@
zzfXVupzin^K6Kyf|NCguuMXU{^PUZ#p1kwq$R_drEf)?{-Ls<h<f*TUTYhVFls@rr
znJuMlrkCAw{-%TbtarTIJCXMr8uj_nr}jqQ{m37@e=tT?UUR48!Kv%+enK34<D;kV
zv+UjcLPh_H=Z>8H?Ur@XZ5FTX`h9&Z;gLmmcnw{%t?JZJ*B2ig|8aE9+U~n=Of9|t
zy{MSSWQ{3&`>uq!n8SURHM$#K8M}Wp>Dunuo3BdQaNV6*VMky6Z0pnE*X|y9`tjz~
zvn4xPO8su1UHypfVg1{8J*eNmw|rm9ccmANFV|d?czb;D(zDf19O!xXzW>gtelmsl
zjh@)P>C4=q>rb8KVyE^Y?fif3nY(yeXW^gCxsUpFt(fnA>8Xe#YqNF+^qfBRo7aQ)
zZ;hE-ad`gBsrlul|N5A@W8Xv1?wIZSEd0aumsZY(oxAh49S=Nk-LKA%n>u#gr`K6d
zgso2)`TEPppZMW5{U^C~+qBi2A3D>;KV9|K!BZz*et7G-fYbdB`*8i4=Z}9C65jn~
zNNDOk{-=Y3O6Fc3_@HoP>!lt;mhYqc(>50GJF9Ixtbfrt$1Z-dW!L<BU%b71+iu^-
z_ZR$eXP)VTZQ`!%DerwNk6cJGrrxrp^X=K`|FMo7EeMSCHJ^Rqh`8Uo<MXfHx1TXM
zzTSG^%(**et{uHP#re{f;;C<M8+dHb#}B;s<bKP4*0H9w51NX~$nAPZ@H*?S&(5U1
z_0ooSx?b#9SA1Bk+jOg@{?{iC75(|mdsk%ZK0N36>D-_0ZW{RTzy{~Gb76ab|K0EA
z@~kJWKBndi{;le#{s*KzX|7wBbS|!Z&;0WBd%yqu;#&^wC;H~p##d^}4%1EN-?_Q%
z_m7sJG_u!g55ACjVAD-2o|)C&d?@$y@1k61W_1Z?Odt1D>{xNJ@rV6~-fiFc%h%3J
z?Y)~%x+mAa=stVj)+0YX`-S?x-F~syx#LLWdwqJ#lY91mqyO^g+`Y*yj=$cS{MepD
z`9)LTjqR>je`fxpSo@Z5KlsK_aN)D!C223;`ru9b5-$80c_?OZ?G9`6F}6B;Th-}T
z-?`^6udexG;^c)mcj~dk1BW8+dfT=14^Kz`y!7B_Z_RnR#+>$h7v9G!f7t!jrVVKo
ztCp0Od)a)K`}$tlER0;a6?Oh%yIX9UUA4rrFkHO6IOEF3Gu&tyFkE_W$+JFQWedYM
zYnI)tQGdJh9CD=k)z+%k!UdsX;Sj&&e(6^>p5dR?SRX&v_u>bumXs_EuU@`-#g&a`
z*lVz@-E}l-ah~2Qugq`T3qS6D_<WYu$2OxWn$cMPJ`vky&HBxn%EjKI*|0@p_4Hb6
z=b`yei~5^<Z2rVQ%riM3TL1}w3M7G0K_mz&m;^(GkPxU)5(*VY!k{RjP+Fpe3Mb)E
z5hMaCl0-r=!k~0S2Ngx4prT1MR1ArM;)FxRl31uX5(gDe;-L~q0+b*EDv=~YC6Oek
zWReV(LQ<e4kx;256)KISL8X&)s0@+;rHF#cB$-fpqKC>NSy0&|8!Ct7KpBVuDwpI!
z<&iw7e3B1kBt|F`F+rJ$8LEI3Koycgs1i~V=6U=1*h)z$R2eCQDktSo6{G^nLM%{~
zq!P+XtWZ^?3aXk^L)DNPC>yas)skAMI#LH!PwJr>NCQ+OX@qJbO;F9G8Ol!VP%Wed
zs+F`twUIWccG3>zAP%Sw(gD>;I-$Bq7gRUthUy_bP)_26>LtBUeWVYnpY%fwkO8Pc
zG6>}&E~p_g1T{>Cp+?9E)F>H+auYYy7#V{aC*x2PWCCiEOhQeODX3{O{rAeB{j=Bi
zUk{yG;=ORjUAtLh@|<D4ZCf;Jw`k0s_Pm6bth};vES|Xk|6v<5ZtuH3Ve4%gADgf8
zHF=gL%Z=AYe>9)HY~kTeniZQgt39meUbFIV>x~D#Wy!~y%DhjuEj(jJ6ELHR^&I)o
zg-v5M=Xd>biI2_C;AdJq^>19cakXcbFIWs8H_|tbEF3he@t)O0c%JQJo6~sDEw<<B
z3xC$-HO=$eeg5IU@qW^>udiC|f9sXmTs@=l_jsOl@5sV}c6rrX1pKLgghyXkrI>#-
z6!(vZ68_0h(mx$a`DZOw01W^$fCYdVzyrVx5CLEY$N(?{Q~;QPG!V={76@h_4+Jw%
z1cDhT1HlYbfnWyFATWbi5ST$c2+SZ61ZI#70y9Vjff-DL!3<`>U<UJGFoQ)fn87j_
z%wQD^W(W-dGlYeJ8Nx%r3=tt<hR6^wLsSTup)?fCP!<YiC=UfQRD^;VDnr2xRiR*p
z(J(N>SQwaLJPgb*5e8<M3<EPvg@H*a1(PxgCgl`NDkzv#QZT8aU}~urOfA!bspVQQ
zwL%M~R%*f2DlM4dG#t!u77k`O4+k?`go7C_!@&$!;b2D42rwg91eg&#0?Y^z0cM1Z
z05d{GfEh_6!Hi^)U`FytFe61In2|CP%t#dpCZh~Y#u%84GccK8U^2<TWQu{QqdG8k
zOb4cp>%i0r9hf?)15>ATU`Ej>Fr!!$m{B|m%qS5BW|WKqGfG8)8BL?XjAqebM)PPe
zqeV2B(J~s$XcY}+42=OZhQ)vx!(+gV5iwxK$QUqVR1BD$axgjLU~<mE<bs3AB?pr$
z4rVNk1v8e#f*H$W!HgBLV8+T=Fk@9Lm~k`?%s3VYW*m<LGfu>T87Jewj8k!7#?yE(
z<5@hI@jM>Pco7d~yo?7kUd4l%Koh`BU<qI*@B}auL;{!zG6Bp4l>nxo0!+aKn1Ty1
zg%Ds0DZms;fSE`W!AxX{U?%cJFcU>0n29nG%tVz4W)e*TGl?aEnZ%R8OcF_8Cdni)
zlT;Fz$ut?vWR?tOGEW9GStNs*ER(@ZR>@$d&=fFJSPGaaJO#`YkpgClOaU`RrGP1^
z1XD5zrsNV#DI}OuN-(98V5ZVkFjHA7n5jG!%v6yIW~xjDGgYO6nMTvVOk-(artvf|
z(?lAWX)+DWG?fNsI!y;Nouz}B&eOq67wKT8%XBc)RXUg%Gy}{GmH}o4&j2$+WPq6=
zGr-JH8DJ`^z*J0uskj1D2?eH-3QVOGn3*&a%uJREW+u-BGgD-OnJF{D%v70R>Zu+~
zJ=24!=Xx;pLJy{1>cP}2J(yWE3(PE*1!fk{0y9fwfte+<z|2xvU}n>7Ftb@UnAtoV
z%xsYjX12@*Gh1bYnL~5H%waiT=I|Uab3_i9IWh;#9F+s6ff~RxFawwdZUECD3}70h
z0ZfB3fSF5k!OUg3VCM2%FmpvNn7J|+%v_ZVW**H0GmqtgnaA_M%oBNF=E*!T^Hd&~
z`7|HQe3lPpKF<d;U*v<CFZ037SNUKXsS!*gGlFU4Mlg-S2&PdQ!89r(m?mlh)5J_*
znz#u}lQ4m4k|r=s$^@pFn!z+PGni&>2GcCeV49^FOtUhBSwIWGEMNs-7VrWv3q%2!
z1+oCl0#yKJAuR;6kQIVi$P2+N6op_G%0e&;RUw!qv?Rjwae|Mngq1`?@sfBbQIZTL
zOVXiKN!DVO(o!%>St*#MycEn*Q3__MECsVvm4aDD%fKvSWnh-^GBC?T8JK0V49qfB
z24*=e2eX`&gIUhY!7LZ$V3x~rFw0drm=&}F%nDWkW(BVRvqDsWSs^RHtWXtTTBrp~
z3$uV};TAA0!UCp6TEMg@3z(I(63j|g31%g)1hZ09f>|jm!K_r3U|OjaOe?d3Y2{Wh
zt-=bXRa(KcDl3>(v<l2BRt07iuL845RDoF~tH7*MRbW=rYA~x=HJH`B8q8`@4Q92h
z2D4gKgIPmsz^q|4VAk*&Fl$5&m^HEn%o<e#rj6Rbv@si)Hf{scCTw8Zqzz1)vVmDk
zYr(8#wP4orS}<!xEts{k7R*{z3uYay1GA3Rfmz4vz^oH>VAjbxFzZwunDw+C%z9Q2
zW<9S5vtHDLSug9stXK75HqZtz8(0IF4ZH!&2GIa!gKPk^K{bHcNE^XyWQ|}p@<uQl
zMI)GvvJuQi)d*%2Z3452HG$d0o4{-mO<*?3CNP^+6PV4k8O&zZ3}!QL2D4c-gV`*b
z!E9E|VA`o2OgpoKY3Ft@?ZOVGUE0C4D?6Aiv<1u-)&gb=ZvnGKw1C+nTfl5lEnv3N
zRxn#xE10dk70g!A3TCTp1+!JPg4ssfz-(h}V7Bo#Fxx~Mm~FBR%r?~qW;<;Mvz@hr
z+0NU+Y!~fdw##-f+f_T54(b5Y!5m;ZxC2awaDeHM4lo_c0cHp70JDR2fZ4%2!0Zqm
zV0Oq3FgsKSn4Po}%udz`W+(3ivr}|}*(p20>{Oj#cF`^{yI2>PUAzm-F3|;Mm+S(w
zOLc+SO}oMDX5C<R^KLM^MK_q;vK!29)eUA3?E$lg^?=#Kd%)}wJz)089x!`U513Bs
z1k=f!U^=-IOs8;y>6A_|oyrMjFYN`hm-T|#%X`7>6}@2g%3d&gRWF!*v=7Wa)(2)E
z?*p?>^nuwY`@rl|ePH&}elYu4KbZZzAIyHy4`#pY2eV)GgE>G4z#L!$U=HvBFbBi{
zm;-VE%mFn3<{%vebC3;!ImidW92A3K4$46=2h|{$F6si)#av*zxC=~|aDnNPE-+on
z1?CVP0&|EBfjPv7z#I}oU=GP4Fo)C-n8S1!%waYR<}e=yb65<6IV^|499F|%j?fV>
zN7x9MBYXtR5itVhh#UcPM2&zsN=LyQWust@@=-8H#VDAgaum!_H43Jiy1{faH<)hj
z2GcFvV7jFnOt*4_IY!689Ajf(j`1-t$HW+zV{#14F*OF}I2{LboQ;Dy&d0$V7vo@#
z%W*Kr)i{_FbOOu?HUZ`Yp8#`0On^BdC%~Lg6JSo#NiZkbB$$(Y63j_43Ff4n1aneN
zf;mN}z?@=JU{3KVFsH;6m{W2J%qcYm<}{rKbDB+qInAfRoEFnyPRnU9r`0r=Gn%*=
zjqqHjzx*wNq#2FGPMgsvZ2gQT2RnC0W5hPkXx4dpu&t;0@|W9OgDtLgmv^j}H&<>;
zZT_o{{b~5><9^F3tE|;EUX`8^FR!Y_-Ri~Nn#J7~XKPQ#LT+f^xYT>u@~g;-RX$g*
zS$$2BZ-FVwcV&_BTK~1z1q3d9+L@hW$juXear(&g82Zm1y*@Z3R2v?#aDi{(icw-x
bvP{beOASg$G5cM=Zhh^N!W;8%&=mh4c2D%7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_f64_instance.cpp.B473866A874EE221.idx b/.cache/clangd/index/device_batchnorm_forward_f64_instance.cpp.B473866A874EE221.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b45020cf829dc5eabe4376915454ff3040f980a6
GIT binary patch
literal 11288
zcmZwN2~^uvqQ`N9v|ebnrCQm!5H8CN;X=3&F8dOoua*`9AwYnzhCKno7S^$XT}PeP
zeXDhAUFz6cx2kh|uG6Q_U0>%_tyWvLwbq&Hs2w|<p?wqkn{!^y;m`k@d->!a&iViE
zIr*n@WohYBy^oKneWl0g7;f?N@$m_q{}~!~cRziRkIy3kK0cm1pR}6$KT?Z-{6|Re
zlLs1Hi+;D>xXSh7qt4ymytBe|^VhW-^ru;7_sIj-KmG+9IneTkF=}^7<(c9cc9i+}
zU-jmYbD!VIVnf~0^6!Ml_q(0{xPR|+-`sKi-XA(<GLQ2^c{hCZ#-VkmPaVI0TWi`0
zzRjdR8F}yLC;xup%kMvZ;^#Bp=^JOBy?f_@H(oum_44k&wd^Uog+4~V{ONAj``^^R
zH}&S9Te6&&EVBM-S>54%N9@j8$J1X|x}G^Se)P884oB|WX*+s{7pA|_^FVA>>DRW;
zL(YCR*>h9o$<Y;azdYg>x$l2|xc`&b+Y1uYDvxd-Uw3fDyP4m7J~`_O3UFA>o`CB%
z{d8-6&8e^N|82l$=C7VGUXi->M*o6e-tjnAKXOI?q30(T8Fs(+(9X31u^ELM|L|X-
zS*Kq4->n;#ov|)3U-inJ2PT%p>~SY}qV^Wwv+s+wXU-hz>eXs=ovXqw9r(G}u&`P?
z{_|%&fxEuHYvbeh^<DYlEibNj|EB%K(LZbV-}KPhu-?%J9?gB?nH8I@3-38{-SI)o
z+3VkWX<hcK4=ih6&fHHw+?M&H=BC#twUO>?Y(+<GgOBdt)4B8V*-M@a_-Vp)TVSjq
zN|`^>z30>8&<7qot`kQGzs)_;aBR)tckcUW>DH~6-*|6ILnnDs-nesa>+7a_UfpE3
zKV@yIUGdlT?=HRi8ujqCH@`Lg;ztKI+3t99PyI#pC(Hi(_m@1kcYIq+`^dKYW)FV9
ze01Hxs_lF3+P5X?i-<ek9`(~Fu8w$L_w@b)yZr}k+LgaNukB>xXC5n!dGMNDJrzIl
ziua#=?T}yj=FdKO{Vxx1So-8W?YIB5$Nu=12)Fr_8;Ewz;kGcd?*~7hb{m}A<iu;0
z-+otCv~lmk+n#D&y+cgwf2Ou-ta8A1ldeoG+wE_;XL}Xds_X1sVSVW1xepJ$x~e&?
zJuRl=Rk40{v!?MO%L^-aebnTi6Z_RCou7R2`CZL}Up~6ZdCj-sJ6`|!KQ|UU^>EM4
zgWAnmkDWPu!&_>0<;MwYj}>lNcI^ihsc)`$dG_rBOZ7_+zHsZ_d*|{mu9;nQGk-at
z+J8$~|LM@L(k?w4`?2O~a`x=Y3yrtEe@oBG*&i!Z{bQTvzD%C}@YBJ)Yt|q8P)6Oo
z;jf;L<6CoF=GVS(ed@V$O77mV^QqUr@IRFiab>9aL(g5FBd@%D>mK<+N7=4p83_+w
zn!n_Oo%eTdG{)@ca@ilNdgAO`Q59{^Z+Ld>;O?0}o!Y$S_p|TmPTYEA?asdh{1Wo<
zq2H|A@aaQ#`r-}tneM+YfA`F`u&Z|Dp7xE~aA=p$(73bp?)mfhBA+|``pdT7zSzfH
z>0`TEv*2ot`q!`d(esahzimG>yDGhEk#EKPl)#07f#)~g;ZtR&Pwc-eJbIC@Wq!D9
zVOhrcjd%FWJ7Zmb@weWx$hT^Kc=^KR#`7ER@Y$pO+si%neT#g{=ZBXpT;hLz;~n;K
zJ35B;$6Yv1O~Aj7?%J1r*}m&=c9i$%`~uD&mG|kzwmHqpIgRDl_h;LnxpIT1`a*Bg
ztlFfpdiyeK_X}H2t{iAyYzrbmk=~<TYzrpAP$47)DwKpmg^@6*a1ss`K_Z|cNhA~{
z6e@~DK}C~jC@s-K#gG^%Mi^8qiG_+IaZvFj9x8z(KykvMbVLW0ND`s+L=TljlAr_;
zP{|}2Dutv#8HfQYm83#RBB9bq8dN$-hcXf)R0hd_Qba*zl1!*9k_DAbvY~QF4pc75
zg~}s&Q28Vus(=(g6_P@zB2olpA||L}QVdl>N}x(fDO5QrkMy2&i)|I80?JIxP?e+-
zs)|%WS%?Lynp8tsi504b)IimeTBtfw2W2BRsCrTl)j%4c8c8FRo!Ft8NE1{uX@+Vc
zEl{nb6{?N2LA8^1C<k#sb&w9IPSOe0MY^E6NjFpv>4EAcy-<Cm56VfLP%h$v>L>kB
z17rYdkPJe(i5qH&3_%T(VW<%@0yRoTp~lD<l!thr#>qI;1et)EB$H56WD07UOhe6(
znP01L{RyAbKMWuA^_#!qUbaDF@?K&6Y@0NfZPFBbdv(Uv{wtDx`#a5ITc8Rwd9UM^
zHOu0^-jcIm{^?nb|Ey+-cm6uxW18xM<JZm)%xQw>G&=9VBj3!9*B!p|I^V^%<$23Z
z7cSka16D8bo+vL_@{Tpp1Am;Kv|i)4UZeH?^<vvbjo-!#eKcjk4;L>9+I;?HE}7E=
zc_(@$-+=S$+l7VZ-`|$^<>#lJTkI}O`}hBbcisEr{JM2+)w?h+hz4oB$oyi&g5sfg
zkRD0|8K7j45lRJRU#MUj3}!G31~ZrkgBdJ>!3>tcU<RvTFhghvm?10#%n%*|W{3y@
zGem}f8KOeK45gu9hO$sFLwP8ep&}H_P#Fqls0sx$jD~?3#=^i1<6&Tii7+t3WEhxX
zDh$kU8V+VS3kNfthl3d|!odue;b4ZVa4;ii1eg&l0?Y^=0cM1V05d{HfEl48z>K7k
zU`Db?Fe7;+n2{n9%t#ptW~7P)lTr#MWfV-xDVS7HFsY<qQboaxqETQ*u_!R3codjX
zA_~kX83ksPiUKp5MuQp6qQQ*j(O^c4XfUH?G?>vU8cZ$Kf~jR%FtuC@rdDXd)JiRw
zTBQXuhQ@#y!(zaU;W1#wh!`+qWDJ-wDh5nO8JLVQFd1iHGQq%Pl7Yz-12dM!f*H$V
z!HngxV8)7AFk@vbn6WAr%s3hcW*mzHGmgiB87JbvjFWL-#;G_k<7qsY@hl$9cpeXC
zyod)gUdDqNuj0W>pb20mumms@cmkLSA_2?<nE+;jN&u5n4kl+DOwKu&TyQYC<Y02e
z!PHS5m^!8dQ^$2+>VytVoz#J;Q#vpcX(E`3ED_8^o(N{5NCY!cCW4u$62a6{J(zl?
z2UE}WVCsb)Ouf{DsaJY1lV}o{Nh}G>B%TCjl1KtGNhX1rq>{iCRDda%08?-QrVs*5
zAqAL12{4msGMLFM8O&s!3}&)O1~XYEgPE+7!AzkkV5YDXFjII6m?<I!%oLddW{OGy
z(?AVi8khl012=$a5C$*}(g3DG8Nf`XsbHqER4`L{DwwGv70gtb3TCQG1yfQ9reqRK
z$t9RlNHC?8U`i#yOrvRFrm-|I(|8(~X(A2GG?@lwno0vRou-4C&eFk5=jmXki*zv4
zWjdJYDjiHCHG*kmMlg-s2&PdO!8A%Em_}s;GlOP;nZYu^%-|VdW{3<hGh_ys87c!z
zMHQHeDKHgRU@D=&R8oPdlmau8W`dc?GQrH`nP6s$OfWNLCYYHj6U;1{1!fk@0yB$e
zfte+;z|4|aU}mW-FtceknAt2F%xs<wX12%%Gh1eZnXR(H%%M48=CB+vb9fGzIU)zl
z9GL@Vj>-Wum*#?*%W}cY<+)(yid-;rWiFVxDi_Q=ng?bc%L6ly=Yg3g^1#fKd0^(L
zJTUWVKA8C|AIyB74`#l|2Qy#hgPE`L!7QK!U>2|fFbjAAm<6H$%mP^eW`QaIvyc{o
zS;z{(EaZh?7K%bJ3uPggg{lzDB3cAy5i0_-h!=rbB#OW+l0{$^sUk2<)C8u9nZPu0
z6PPAp0@EZ-V49Q(%wk#$W-%)UvzQlySuBddESAM!7OP?~OK1t0C9DL@5?%ski6{ZH
zM3#VAqDsIlrKMn&vQjWhc`2Bsq7=+hSqf&UDh0EgmTSGA6)d)uv+{T-Uap4{<pwBO
zZiG_h*%zvUR)AT-D!{DZ6<}6~3NR~V1(+480!%YCgK1`FFwNWyrdgQ5G)ps>W@QGm
zl2(FQ$tuCD<dtAnib^mmWhI!EsuIj9S_Nhms{*r%SAkh2s=%z0RbW=BDlje70;Yvo
zz_f4+m=<9H(;_WkT9gIMYFZ6uHLC`*npcBaEvmt+mepWZt7<T<)C#7RS;4e&E0|Vc
z1=A|6U|N+G%o<t)W(}(Wvxe7zStDw|tdTWf)~FgVYiTW*wX7D*T3!ogt*8aFR@Q=9
zt7^flqjg}`u{tp8cpaE^q7KYDSqEmFssqzTZD87%4NM!ifoT&qFm2KXrcK$vtf%#0
z*0Xvr>v=tx^`aikdRY%<y{ZSZfi{5Iz#718;0<6lhz2klWCNHDssYSK+6ZPNYXq~A
zH-gzH8o_LojbJvaMlkKv4yK*i!L)Nbn08?Y(=P2`+Lay5CfWpM6Kevqi8q1SB$~i%
zl1*SXsU|R+X)~D3tQpK^-VA25Xa=)cHiOx$n!#+LEnv2=7BE|Q3z#jU1<V%N0%nV9
z0kf61g4xPi!EEKNV77`@Fk59Sn60W6%r@EvW*ch*vyHcb*(Tb+Y?EzZwy8ES+i5$P
z?W`TlcHRzVyJ!cqUABYSuG+zLPzRU}<^a>d9bh_y15Af>fay>UFgs`mm>sMG%nsfG
zW{2nivqN@(*`YeX?4+GwcCt<|J9#IVouU)WPT2`&r|JZ=i*|w8#k#=k;$2{Ni7qg^
zWEYrSste3++6`tm>jtx%cZ1n2y20$0-C%aBZZLaj512iy2h1Mc17?rt0kcQ;fZ3yZ
z!0e^HVD_?JFnf6~n7yJG%wE|GX0Pf6vyb+H*~j|8?Bji4_K7|)`(z)OeX0*kCv}49
zWKJ-h+zF;rIKgyECzwv<1k*)bV7izKOc!^7=@Kq5UD5@nOS!=8r~P2|vwkr9c|Vx_
zq94qD*$-yF>IZXx4uCnp2EZKP17Hq_0Wb&T0GI=60L(!;2<9Lg1apuNf;lJ#!5ox>
zU=FH5Fx}J*rklCJbaOYDZs7*gE!|+cl^e_<It1nr8v=8P4}m!(hQJ(>LtqZ6AuxyO
zFqp$^7|dZl4Cb&H26I>rgE_2*!5pC@V2-d6Fh}?Zm?L5Y%n>;P=7<^rbCiyPIm$-C
z9Oa{6j*3w*N98D(qiPh)F**k37#jm~jE{jiCdR-VlVf0xsWC7;)B~o6dBF5=511a|
z0n;NrV0x4X%yBvn<~SP%bDWQZIWES*9GBx@j;nDnC+Gy26Kn#^2|fYlgqQ$xLQa4=
zp(enbq?2GyvPm!}`6QT=ViL?rISJ;Zngnx-PJubarof!yQ(#VsDKMwx6qr+L3e0Ib
z4dyhP26LKEgE=jx!JL-UU{0%PFlXotm@{k!%o#od=8TvDb4Jd9IiqI4oYN%EX@vK#
z|G6(4q|9j~cKVz~VQ0;0a<TL0G)36ObD9<2c02l7&V9LU$lW%y;@pn)+~)rt)!rQ+
zpPFBHFS0%QN&Tf&$N#W)L3NF_w$7*a*CC&}3%hO3_P(z9T-CAqBEJO-FDCwr7GJV-
z$)zg;OH6Tr0V|7^1zmP|aLD|pojJLA`2}KmVpfbXf&PEbUJ(`^5f!bSzri<u$0#|)
bAk#A<(?XL{i<e)q;>vp8(lv$up(*<(xI`cb

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_impl.hpp.6D2CA9CBDDC84AC3.idx b/.cache/clangd/index/device_batchnorm_forward_impl.hpp.6D2CA9CBDDC84AC3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6491ff3a2142ed430fb58c81dea917bc3add75e8
GIT binary patch
literal 15958
zcmeHOX;f6lwx;R~LVGy2G$6A~dYA;HQ4mlj0Ra_)^AyB6PB`n;L~%j|5ecFgQ3*I9
zhzc476cn5_Mif+HaF#e5yb3BRudq*1_40mX-L>w$?@wGw@ALKkx=z)uY42S{RA_K;
z|6W>Jf#X6KOdCHZhSk#2YLEV(J!jT8JOWkiw6qpv{5<5=bW;yyZA4|@hq^g7kDOWX
zd{0+)P-sk1{-@rNB{p5GoI}@-JU{u$l2_{&-TPW!zu&a%xR^DwC&aH(Pb|uQTu~VN
zAg8B;(tTCsi?zcSCgql04*sRIADnRme9hnW3H7>?tF=$>!jDX;c$drGwfwlwO}#xw
ze=KIorp<i6GEgkHNME}DLiVd&eRs@=c>XvcVXMdS*uK#TAD8`V+hjd=G51Z%iM5Yz
zPy9Hhc;olUi#-lF9UVW#Ikk4AQ`n5(#vPr}V9->!!)=@UiPv7&{H{eDb2Xgv$AX##
zPNfg(_ii71dKd&tYJa&>amW4Li!f=vdTgxa!6=WAMYm?!om5&V3<DUG$-gi1T)X|4
z_TiJk+h%B=&MjWkt9EOz;EwUd-ln>pZWJ7@x?V7E#r1+C-LAVeKCHf7keyff)rf&H
zbE6_AW~YWJzDkN;_RXGcak_KWtEcF9UG2407<hGMK=iWXxobyvoSZfHLaE8{yohIa
z$Ii(=vNCs}`|_Tzr%!Hrzy9*<)7jw%@*^B`H{UF-?ilrIaKgrrwRK}$=4WOX4KoM}
zon|%7&3Uy;Q?~~@Ulo+?^4_y`NowEGImh>1sOdiFjqmw@n?;c;M>%(XBJ|DLT%Ij_
zTWWZ{$0#x5+xH{4I=q^B=Uzj;YWb`{uYqa@r-m-Y4!x^>nVGY>oz=xj)q5j;Zf<b)
zTp!fCG4qLGuK_n0y?tHV)nqZDTSL|vmX=f%MCJz{)ho>%y7*LW!OL+6-n{>)7&E0%
zt6Tp0>0XB|g_n*71qb>jJaK=Uw7q{b=(R7oqnK;jT%qHTc65P#-o}jHnZkpp^^=zC
zZ8<YD<Gk7ZOIJ3S{FI#L6W4iQ<YM<rr`B9bbD#6`($N;cXM**;AjQmriaujQV~i&@
z7;PH3`fAjt4EwiNo~U>A`QE?4Ci8Gr^1&b-RsXgr<JukaFB@M!*kDl3(}aDaZm{!`
zrtat!xwY5H>q2#X?WZ+?+tXJx#M&<_nq|Lj^ebkD&$KYNyV|~W@sXtmo<-ch{oBb$
zUFXm1Fu|;Ni%0Bt$qOt_HXL&MIKS%YglXw#&UVZG_MgoUCTPXptV;E+?>)`vwbiU?
zwr$!bUoN%s8h3x#qP#Jm{l+~Fukt;&uHWh0`Gdm8kKKCumhC^L&ot>adV8{TaO2f8
zD=V$;MWnX<nAm>of}jVd(wkT5@49gJLX;@vSq#x`oUOHHmRVbElajIVC;W>o&er|3
zSG`8RUZu0x#KZfjOXJHsDZQ(M2JfyO9qH())X5NPA6Ket7nm4DZXV+K{%&~FhN89I
ztB<T3a6YTzSfu^o0~;3o_V8H7u0C^49K3k+U|W~rW?HxSC7-V^yklDy8=T#A+0fhF
zY)jXtHy4=ethi&h2Tpu)`jc*O=_}9K|Clx=wvCn+4B0Uv`oXgoBJ9>tOF9CDBhXqF
zTH0vUK>zsXm)>1FWO_z~{)n;#>TZEnvcMF@w0BGFLUi1ekhjp1&ce25p{Xn|MR&j9
zCX4!&_eB_hN-2k8Ikb>PE9Lk`SI?I}8i>#oQL>;{791!GOcAHYPgs{}UL(RFMA;6l
zw?ju+U`o5gCeJ1N{JSFTjVLFfcoJ1tBQQl3I`xn2NUn_`WDunkXqN(|ELtgpzbQ+3
z89IZ9thtut+{W4b%ZHW@1jr;n0>-a(^}AkL_)ZVDLru;AYzDBE1unyM-2#__k&}k<
zunnT5Fl|#9Jy~Fi&6LSMhT4sIBSK?TlO51&2kavYOtBSWMhvqEi*5@IP$@+~Q3Q-+
zfhp34&WB=u4H+ZCP*loZXucOZ%K}sEE60t4Pe<m8P>m>s(4Y`@mIbD`HFN!Rb2i%;
zKm~eHfO&w^mk+G=pS-f2?jM)GY713JT?`b(sKFY6DWRL+2UM!(4icdeddW{v_z4=z
z0#hO`HP0|C?dKrEUWl>}8t;R4vcQzcdvm5_7x|is&=XO1L*w1hTo#xzaOkJR#*Qh=
zm9Q<UO9N-!z;%;FE9KyuPm#kWb`xPwR7yTH&WA3tz?4D1t4Gb=F#J~$dLYUkXubz_
zk_DzjTkh@Z*|;}egu#eX3_BD<Nfxb?K8uEr``TxK2z?OcAT&ORUZfG25;Ik)KjGyR
zMhW#SQ7N2Q%(a(AE9L&gC4GJ0#_*6qEjk9;9s?>_V9J!qi<`D|K7yu)33|~^=(Q7u
z%K}rDzT4kTs=nS=3H4Da524OOXeJ9xS^l9qeWka{Zw62YQ3A~aoxXfv_38x?zl|$$
z^ATYPddW6uy$w3a0#j061QuB}q_`Tu_NW9ub3dmqA6T8*X7Z$l2b~J_por9Slyls_
ze6*_fGn;C=bsnp+MJgkvGVLXez?8Jc>FdtUFxD5L6{^E2D4v3bvcQz|i|pPp%@!||
zP#eAEB&TzdGm}Lt<z>N=z{2e|A`C>OoPgp9G>kL?Q?@^8^gE#Sp+JOv5G4~@XQI_a
zBQPavw~hEs-l&^;P#3)@9w_61S{9g+o!SpR?%Tsngnp=$W6<Up>?R9L+0oPgvw_Ef
z4<d9!l%ufSQD`I!OgVb*_2XXGzdtR)4v2Cb+8l@OvcQz`;78Y|Ymd(qVHl#6zz!v7
zFlq#*oH=!7)M(D&pL$S<svD#Xvi|ac)t5HvsOGL*y-9@DsDyN=pN__dMqtY2HE+)K
zNqN^#gzkt^3T;ZErz|k#m-u&e<I6pti_j2JHbec*XwT3HOu1XOd7{nE6C*^}15t9J
zDi?N;1*Y7KyYbcT3vI55&>K+>K;r|*_!@yJ4;Gl+c=Gl4A4TYeD2JfIA!s8DOnDUY
z!}Gag)Ez|VizpdTHv_Hq8qrGWH?QgUdBGt{$RV5T<dr*lb6K=f_8jnD8ngVW5(<ct
z$y;UeYFS{);|t4TZtHfe7h!kwq8w<JgX*FYn9|VKv49yrc!>zj5G5OWWy4TeV2akP
z@e^k6=}PupPLP&C<uceA?@5BR3MyC8I0=R(!DymShM~y>E9BieMqWUeS&O`TRMh#R
zf@@%C4UEKlvmmWtLRT=6c#ROGbk;VVbxf5fq##{lZ7<RI4r_adV7{cArB6`)abA!{
zs75&BktRq}geek1-wEGIcnJ`s?}6|=knkiDq+P&t7wC@nPC+UFLIIF4#-0Wgq~}0*
z4lHq_6|%p98;RQl6tb70lDBJJ*VR8vPV0J0N07q(!$U~5`uX>x(2$NpI$4LG`604R
zlY<})3mO(q(3qZM0?F^j_8d#0OyHXdf^c;ODGT^!(Kruy=K+7L$1V@5rN-aE)z_6B
z9n{WBAXWl9(oT)ngNfMPVFc+O5bgm3OcbOf=#m6IF;>V<5dL91qpAScYqzKHtTed)
z;0WA0R;q-<E8z%iKUS)S!>egr1EXr-0HUvjp0%(C!3m6i0uzA85caiD3pE}Ve!nt)
z?4Gg@zgnP*M>$3X;P(sC0LK9o$^n)+zysS0HwKla@yT#$xOW-F7bh&#c3xsiao}P>
z%4GC283!_gcQE=p2sXNLQ2XQVBW~?mnr&YR9Y`}Op<gBRC%78gSJV1x=vPhaYoL7%
zt*?Q8HMG7K#?-<wq`ceE;Wl(6_%*bD4IQxASm`zNdky^wZi4nrw7v=YHPQMNjQt9R
z`uz&VZv{jBK9TWAWIRcEiA+!;6HM@NuETN82>T(};G9DCSh0Zx<j}bmgGlV_2~rAF
zra)KRGlG-~m8mpNgUU35vBwJ)vY$%|&f+bzc<GV6&{{;QE||q6EdyvEGiFIlCtOu)
z-$O;OJ+CPK?lb9s_Zf~9(2TkU#B0Ee?vuW$a5%;aIex&Cr!{_1$Wa67lVYe{3=Of5
zD<EG&y963yHv_HCj7I?zM5fm{#^W4~Z!v+ln6HTbBjfXt2_ra<?U~1h6I{%i6|>ZI
z{mlA*W+RAx6Q{q4v&I-lHE5N*3B;SABWa-~#KA4Zu?;d!9WbZ^)LxnZjT4{)`JpE8
z!9R447#HG}J3mpyN-u!T3(yG@TX4t=5K8b%pneIu(Rk=f&}ws5I?HuC%Xwh8l$FkL
z-OkbYA{TOz3nltXT*xIFS8@)OoD<Pka{Vf~2!g9PpDL~w!PT5kHH~XHj~cEA(cj{_
z-{L$8zReA}%?%^?E;sltH<aK9T#pBwH^Ft>$U5$8f}eA?&$-SRcTHIBv3`8$a5R0_
zajxsQXxwLA>$c^ajy@d=S*aFA)DrK5l@get3Cu7&&)xbwOyG@&b#KE;7rE{iX;WR|
zx?iGk73Ww*o2Hs`tfuj8F5os7h^xj*ce#MOG_K<U>IlYBZd=8d=$1@|iCz;!$-ckP
zYhfsHW0HVZ5(p)3%zogtAB5s%RFD#2mjvj8M@I|pod5?ATnoKwVJO}X1Z119YT;;t
z6PUmRCJNg|kP;d6Uw4e%d!AC<{Ks&8G`^2Q?W0h{;e>#^ANDL%kb*8k?Tb*v*t63Z
zt(bPvb2(O;7(THtZXGL4@|xsL(0&lUAN0koY3aGJS{RNCY3aEE3CuuT8_$4E0fir|
z+Wlo0d;XfiO?%m)BN`;X1Lg0)0^6g9IzH-lMebXqSfXCiADh|BqW;1ZKBivDN_%->
zFE2il9rqU8eJ}4xa4s+8()wK9F_-T`>qBz+P=fP#A&=(g@s4>kKaUT|qxt)IVIR%k
z$9wJLeMouvypT`p^Lf{N-i_7|$ma*qcuqb)kH*_UKK~=2zbQiBl(c^1<MXC%f$CEP
zvQa(Lp`PiCXP94XPbU4$)iwXes(8;<D&(vT8QRC7_AzLU<5U5KmD<OlHI6t16y9l<
zLTeoP3DRN4<1oFh|HOFwMB`c}u$CExYa~dwnZVmLZf1O%nZZOK$NI#v6h6k99b>x_
z{c+apIE|ZG|7Las(Z_NAaU6xyIsJ6bndmoj`kM*%Z(Q<=!~0>gkcHkbMsFBzOzd4)
zy!+61B>_lL#pqWt0j-L=9h;rMf=Nj61Ec=~<BJs`*D}1f3=f)ZC`cn+M)tr{Mv%UC
z`I<t}F3}X4YB$x5FsIv1r%<e2EQRI^^9}Ln7o_<f^Zf~0Y`55rpg3R`2PnK0*e#`T
z0uT~_Az7djfJXxGCpZ<@r2;pC(|}zXjW+_jjWpf^ggw9zFA0K_1B4tJ?*+nM8W#e)
zLf}UDhk)H78XpFBhY7}cLKI3i17$NX$18&%#Y1I0q;LWpoB+oXeG>Faf*}MaL$73l
z6>_!_Z}2$Vh+K$Yf%q$VkwVTnVj|8tB4RZVtAQSVgF?<cV&VXupySiVE7zf+vXj@@
z$+yE(7^fsr@S@2_;^{X~PDG;B`xX#y0b5+OAiV_sFToIu6>@qKzY3=(QJt;>@j5Uh
zL`|*|69>nn4a@uSmKUniL0)x`H^g-ssyf?sbiPq1R*)iWB7Di9oM$u7mkh2oz-A2!
zWaUhkAngF=J3tTInM0e>yVjK~cmEHEk|TY#sIs`n)Bkc?8Fi$`)RCQSZP4I32F#BE
zdu)c5-O&0hbjBN(fHI%v51~EYxKMDz85MK3I0_O_!qDa<=Yl;40d<^y180eQ68pu-
z{%fJI7PflaXo3Bb2*xjtmNUSJSPR5j(!&Zl4~&WU&1j5hlE64#!UW`$Xi~p;@SwH%
z%jpCr2-0~do`;5{jpw2IJoF~`G1Ptxm88v&q4XHK5S+>wr81UyfkS%%=Xjh`<C+Ol
z1t(T;hQuZnoVtQIhR8MHEfaY=Y@_k^vwF^pS^gRgy<A2=muX$pzp*LTx6~DPBbyXL
z?Lw%Jopb>u@U#y>eVp6FS#2~us(@HUynI1QgMHKBC^G6bDQ|3wsi|Y6-uJ@hqc=PO
z=1+h<CJNG1VE&ZgspY%vw`_eo;@|28Ehu(Avdbu|8CCl(P~IhP6Quh<d7s7)f$|~2
z3o7nCT*y}bY$`}W&OvH2Cxe}XDU<<RGe9>y2L)*>aNSDdQeajJtVol~fLR%hPXMzM
zG(HK;PSUs>n3dDG0+?0M_!KZZMdQ=J>@<zHLEUXo51$qYDE*|n9qQqH70S#&-At&5
z6YMBc26eNb9!`7-C@rd+4fSxw6s<W>cL&tNK`vTcq3%wohX;osHFBy(&Ik`XL3+oj
z-qH9yr+QE051i@)jhi@C6OBJ|s*g1O#Hl{f_%o;aOfa54XyR%PRq*`8rzr}>ms1q{
zwIGe}G~S)`?OV|6E$Byjs{uGQ0B?eufw~z45F7`s;z(AdB_-^=3WgE=Dj2m2#t{4i
z<M#tY1=97bZ#~-|Px5&91Pt<f+Tz|gOU^xTC2{AWOz|Bc+#&l*i@)ykI|wG_CqbJe
z*oEL^Xp>B^LOvkEGc#d)_Vp_jaVuN)Bi863+nNzb3}aTjNN)DD6eK?bKeH71AW4t{
z3<At>U?xa`27zYSuxOUJ2erm3f)wl?OrZ?WJp=TlXG!<1G%f|!rDW;}QW>x=qwxt~
zeS*g2z`C5q6~MZJ#;1VwDH@*!)~9K_4XU<56KvR)S)tkvP3T#n%7iBLtWae^6M9yt
zvY`n*D^xq62|X)RJD~}g74=Z4hblZe1?dA6KG1kMBP?eqeF7sS(72H^ZX_FC%d9ef
zN8|UL@p~FKamGzF{>T}Br12-t_!EskbH<+uPJG(?`lJc7KK|EZ{<W`*PQD84_~`FE
zkMytQd6FI%%pBnVEcq|@*p#b}|5#eGDF5HO$Kv(DK=I{(8LwODNFg|C5Ot&wI3S9k
z!Mz6eBWPIPVIv620KOTZKfPS|Zl!T12+jn9h&~GhXVEwZc;|pHqCd-7o~8ZAbDZTl
z8einBE^@BKm%7ARU7~RfXI;Z}CHh;O)h$|oo9l9$^C9}XT$j5ve!$s2;M7E4$91Wr
z_0KrNXIv+&#|}SA&bOS(6LV%VPMJ(3-Yu|1d!6Q39y_UN_Z&|6mF{?omi67d(Qe)r
zhZ$*&qem@kXI70o@;Yc!M|+!fP+bRo$pV$ibV+4;;8&pI1SUL{8A5OgXH-Ia=p|hD
z5-tFHr}$z3s+#6j0N!ynp6|2!TKr-sG)@zF{Y1VaaTKz6{VcvCK6k*jKqY7{4PYyz
z%eMuPLUTs|7oD-y@&32vb3OjU;m_9h`&ZuVkZ6u<I8{A0oVX6t)YB-mSiLwL_am|%
zP^W@$61Z&u>J2n51nNQ%j>jwBc96^Z00<B0!^C9hl1v{aYVJT_6a1Lrae7dn*Zon)
zH?sPTtRL>Lti_4Rmqd@Cf9tHr%|XLdbKL@W5poh4rya!2!TBuI>Fc4e9=7HeQ4saz
z5(ch6PM0EL0u&Np>nW;2zJt+947_C?Q2t@dUk;<Z1tnbMw5;Ivf)wo)9YA_3IyCw#
zf~KpdQ_jltfazgG7psn?bg==klr9da<A68m=Qt1$2f_$m3e-z!{ZbIHl-8#KbsF%-
zhD3oB2uK5A1aAcDjkJCv2-ryL_ks?4K}S4pP{an*hk!T1hk^Pq!TZ!+k*Bwg-h`}{
z!kDEn5?(M+6u|3elF66<(Gpm{kJCcqsGilYXM5psRB+40VCJbaLy;nh5t0~FykHz!
z;_|@Kbe0;hC<|&|&$M1vIQ(qk;SLF}_x^oM_@`sSpSl;MFRFar2U%=4ui8y=K{(rs
zdg*&8d{1(03i*~1o&-hDx-9Bv=DP_k6Hz0gqRH+#;OhZn2+9GGIba|j6D3+}qON5%
zIsMDwaEXsq`-iI=T4IxDoc1$L53kV2&Q%ZF=l#z|fA{d8dbqSQ=Dzz-pC!m*Ijm6*
z>w{-e*{laIM}Axwsm)4Ppy?H8j*~*HR0U0|XnYfz-lTCYG_9raa@Kq~+YuMcN@c8h
z8I3Qq=9g)l!5L?8rkH<1^eX(#b>}!#!#dEu4(Q{+v0Pf*)#Fgu&VTEEfsd=vz+MN1
zb#$6eb2knT?20{+({+QjSN!Gbju&ZG@#XRy*?)?FYY_;*2kI?%aFj{V8n;f6lA$u0
z;ImgNFJ655=EQ&4b2$6thq$eeyH~btIVQ8;%R6GXjFobEgIvA?jn%n)H-htcgFITF
z$E)*b{XX7yAK#hq^LdMW-jc>W^LbwyN9Xe~G+u!Ywc=Y3wOA>i-;KV>Y5fLYNJM?|
z1PD*aP*KR&5^?9@!zo1E0EG>(HT8w=3Vpewh>5@WbpAFf!ffN;>&Bnf4SZ`8SrEDK
zPk{*@w+i|0CbnR;%kx`>gGQ$!S2SD*H^H+(kopPzC^T3(*n(^a!<54)lnI1PV1hS?
zmV0tS7Qr_jtscDVnmy1Kq=|YH?Z^uj>MgV*UQQCwO9FPpYe@!r$u!;%^!5Wg@{%{u
z`VDNo10_hGVCPTJ2fGsJXo<0jXL?|KbL9;qey+*&zjFfrVou;@QF)uJNjD_4x}RqB
zPm`4Y%{LiQlhPAb{fE~(e6<*j;~F5=&_{Ked&YPi-}C$>urbdF{Jj$VX(hlnuaO0R
z1LAM=HCxU7YivP$MuRfMb-=C;bjM4EAiV`nZ-E!V&A_S|xDdPwI;^52O3meMT<}Aq
zolfiP+Bu^YAf7dfXU(wdfKuJ8(JC63vqt4Ku3(KSXk5t}Rnqt>Yjl;y+c@<$&V!V{
zol|co_>n93rcJvJiO3WUoK6F$B8h`cUMG`R;lu&HXO4(3f$);HndY`RZW`YGP$EHd
z)f~GIxFn>|TrkIq-<^u~sm*h<&?)V5)^a(05u%K>ETi#d*77oqGdQyh&J8yXUr|SO
zyaB`;B*0h57u9hcUuP7Rxg4;v{d;EoX=eOW`Q-WXH4*Or^$aym`J?#hDiE)diwZ3_
zaQmjf5qM@P<Uc6DUD@#6nZ<@pC)@q^YaAbX)b>Kb+rK=oZ*1D1*Wu|j6xd56y+-yY
zyZtDyQ54z(I_v?CcsGBio>|c+|BT=Ndn_|+{M4A<=r8_y=e51B|7yt)fOC=f8iE$?
z34F??h37TSi)i6|$n05TXBW2oMWubq$CrOl;$z2dI&#tEchR-0OJ_TKN2e~%4mO=^
vwQMm(OUs!&I+8~l^5{e!JCjEj^4Ntub|Q}s<PrTv3jSw@AGEYw$>aY3VeT1t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_forward_impl_obsolete.hpp.7A0467126AEB6502.idx b/.cache/clangd/index/device_batchnorm_forward_impl_obsolete.hpp.7A0467126AEB6502.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3c5031a79ed0dbab67a58be681e0b4b8b1b534ae
GIT binary patch
literal 14182
zcmeHOcT`o!)@SCPBX~U=uXGTQb|`{^Tqy!dQxHMLUa+BJj7o|nF?JI%NKp`@20Iq)
zqSz3mM#(caL~Lllf`Vd0Vk62oc+Y0$Ti+jVWo3PDt*rG}%e&9-?%$m=ZO`7bXB-?B
z5;D$9ODlMC*xcBd+0mSqmR1MyZ`SOYpTAIOX<ZYwwB~NSFg*6lN>A0bfiHi27Ufg%
ztkT|qUp-GZ>HJgGXQLWcY~7jNXT43^)kp3dH9hwo?0PjZ5sftybS6I8oO{0{xAauv
z?>(D*`qvLyx|NT3_VDVJjtky5&1k59;-dAYXLXF$#qy)tr<IpG?=+uR{X}b=+q8K%
zHZ19~G$Cs5{1=ttx+-r2tNNv;Z}<DX4=?WBru{Di4(_?}UHPE{bwfstiQMpT%Z{!G
zt(}Lg&n|9?7;iV->D_~Yir-d89e?P4r^_$H_yLuI@x}a!<2!dQAJxv~%bB{}n+n4x
z1!}Kacc8Sw%Ws_Sto`2M*42l8G;Q;CgHb_LUE`uJ>bvY1{`mE@FXxMfCdI);K~n>Q
zpDt)$`ZC*io9B&R^~G&-=XjN6F22%KduQagpQRms;B$QCEuU+fFZ;Ycc-H6P(!5il
zuh-1+tV`(Zx9Mv6%?CSH2Myh~ZGzc8hY6GJB;6a`Cn;lqmF-WB+b*x0@Z8z<MrnHO
z?U}omi%-gZ=62fjB(`p^<%Q@~d6l;o`Atp?dReQt|4I3|Y4*pT)w%BZ*169<=a6=X
z*ZE#9E8gF4=Z@nk$4n2!yEf_G|LOV6w-Mi7_HMTRxp46Os7=|^_tq^=*_qXDTff}A
z^q5(0d!{H4xi`EUGy2i&9y1$iX3g1qIy3mk+@^8DIo-O*3%j#!oTxozQ07*;b4sse
zF58D?_R?+6ZK@79Xmd)t|A`l=rz}&ZCV%&<kwxE!e#)aS?(AH<bEZeJ!`<A3n$X)l
zJvRw!t+!MK&WN?#dd=gGw1SU*F=kInT+Cp*u!kMD*frxV*Gj+r(IdI}8F#Foo8_A5
zF-pDI$WdYDr8zEE7j-Y}j&2(HFU8a#z9zIJV&l1>u;A<FI?LNtchh%rRmYcGgqVf*
z`^`JM^yI{VHESO~8ndT#!1$tN<%4&;v$b$@nUpsyX5p%bKGR2pE&a)^<DT|IKO6W}
z_mY(hUoE&7X>&bfX4Z>eW*qHg&CN0#>{oPCyL#@^qz8o^KBj61dGy~f@ZQ;JU%g9p
z$UQO2{MMMGwl{4zhi0Yjvgx?%Wohb={CCY?t5f==emB#hF*o|xM`2w3%@e&#BmAbX
zsyN?iMZ~Lf4FyJ}yB<nS%C+at`m8$f`TN`VdQB<pwdLZ(E<2OMw+9z*nPU9b*!Jv;
zxOm?oFFOrh@AHFmxQF{FbynLqX(_#VG35KRAGaEw?C06;U&Ytv*r(bovcI4Cb$R*1
z@4jkYX}7n^sQ+l)HpjkqURrqPSKX?mb=S%Zm)y!LzPW1BlQlUBo`0O!JmgG9Y2HxN
zV_9qF`#s3J<<NI_;hB<?XB1UaPU1PbA3tt-v~a_p+s2BP^Fr${e>KB+#iS>5>{Q2R
z4_%rq-1~@<9wqIW^<C`52yHDb9J(WF{QW0SMeN>5OFD*yW7tL(S~_HXAb)M@&%Z4{
z+VQN2dlAadSoddaEeoJ1W8W@z3~lSK!fh<Iq!Qe&1b37LP;?I&t+RX?cTdFrq?AG|
z7Gg_Tv{Fv5b@P6<%}~T031utx*@}nC0x04a%crc|YH?G<L4>ji+ib$lvH(i^V`lXZ
zP6w++97reySS%pb)d-;Ihs}7gJ(AZp!U{skLppg#C5u+d@XrfUpM`zZ26Gl#l8d&B
z#itLgwiuBvK@u`;&^73Ge%@O>+@3T!9dYT%P8Lu`$CYzkkBpf*vJKWIln=Pw2dpOx
zpx92Iwk^yts!_zIq$U~ICj<AD1yJmS=%`Va{l>S$hNP4eNO=O8$O0%1YwVAHc`J0H
zh(kyzyRpS?+*KApaVncU89yGgPsDD7au^#P#`dxRihDEPUw6;XwuV?qUKC&v;PUAM
z)Pd9D+Uvf!IH(=gC+d?(d6G0(BY+b2^Shuj^_*cMHX$!LiiM-tR2D!Pc)s~7qx}Az
zMeIu``>^Rg>?jMMMBbe}eS407M-h7v%1&&$6I;jvC__hnNPN>N^#>JhN9vNm+a&NF
zvS_6oZu}5A>N5`!`;b!hW7GZERTe-Q_DDT;)|%0`MBJTFc43QMSds-$##`;}=KW^x
zauEj;%1LZ|5<AGEmC|?q=*i>!hKRTap&Y`dhscXG0w~clR0dO?O;@O}o|Tq#krOX+
z9c0l;x%b)P2>(B#+h7H0(J9pK6w;RkP^M2?)U?6=1ThaY@}lk7XFCp;1yGj0J>=m~
zaV<iH4M-`KxNRjimjzIMcwezT&e!#}A#O`3ffj)-pFThxKX>5m$vN(RBJNFIvJu;C
z#4fS`O6t?#9P8JqZicu6DWR7|FPBdrpia}CHud#=DO(SVL_JeAv-_uyR`nspx;q~B
zlQezNjKyZ$LDC4ItpAPMJF(gFsS4|m5-Pa16`Z*&S}D(tEDp}zVk_bRQq6oU=94+1
z5kT2e_a-1q>-`ZC_aT(c*k&_XDl`Hp8JV`?=ljRr(8IdqMaz(C8B)svDBIKe<JyQG
z?jr6<N;!pXPhk&P0A)wdz>kJrS?@*cN+`#1`{UR|7C<?9x1qMzwbf@tY)mM**ftk;
zmjzIYLu#&lp%b%J#G!<egN<{@Ow$OUoc-nO*a^JhD?O|t)eTk!+kE-}_4&1J_2<OJ
zuM@E~DPb)(SW5<9BY<*oMdPo1Q{VO%aW_KA!?t<YTNXgMwEXSEm}2jG5gQT8I&83x
zI0TIVN_D}{pV{s#j1sXYq3pr>d$6%AfO2=q^+B2Ev@eU;n@|p7(}Se*H3BI2=bB%y
z8@KvT5qlBJ5o~w_+sXncHKFPCb0(^dMeIi?=~y?NEae)}O6mVi)1z-fLRFY2eUihg
za(D|_v{H6w`7VwA;fe|igmRR(KFX_Q0hHQv-$(zh+v%l<)#OE)*gTWeMI(UnI>Pyg
zB4+qv5t|drHte$vhsgpcGh?RAYU@mWDlbS&uxbgm2Y(|-aaa|{;zS&lh{scXG7d|o
zSSfqrXnAo^%$)DrqvWB#LXawOSOty*zbi<06<&80z0zdk5~N0jS0lx3ay_j4`W!sn
zMvzA7kLn5&NsvAlK6jvKzA)bb%vX?7kdT5Lz(xdV2kN*3d4k^)q%0(4AqT(^C162%
zjD*L?3L2%9V+Cjg)EOz|2ti+VwYqNc>C=AdR@qjN!UDrWX|=)w!x=QJ(=f><{A_xp
zc2j3T8qsG&I7JhBP6(#Ii|!fCpv}mCGwK7?6{M}me=CdkBHz6z5cCkPkXmYD6{x<h
z9HfwTUP9s}<Vf48iCBOLp$ZwSDkM}PLm&!LB6dy0-hh>I00Td4Z*n!rt^D`t948GO
zFmxcaj*~9n(HC$O^dBdc;n8I*uE2vU@DR{L97I~6iGHAb)$&QZ3*O(dBo&KvjtqkG
z1!<7;AO>Y3t4!nt{Q;FFrD@_NC=EtWq5Kpeh1q?Rn0hLBksu{043ZR`>5NQK7^F~a
za{aK*pP47z+qE>y=>qNy)o#H77jPiOW!R~V)tBLbGFD%Kohn#;1rDfS^^dXBW84{B
zBPTt^0grJY#jmi_D^~vs2fSkSWt?#tX97MPI)qos(H?YwK!PE+lPD76Btcq*RjaTY
z^qe52VpS@O)37RyVu%*8QjQU6!N+;4<Ge$SypGxUAI)j-zO4MmbsYq|Wa5<}u?(5B
zT@#UtM*~*MAu!A_2!Y!w<-nJA%}K0t5*tC7CJ>LQlY@=GnF><6!Yf_Tht9$Sir@o^
zK@=A$dK4)_DZZpIxumeCcr_Qenj1**KF(ktX9E}lZ?YI%L*g~miMBu!g+mJ<{3cC*
zfD9iX#wEmK(|FvO{!kN}!w<U-oE+-D@7qLuPI`iDpCAc{EjaWE3Zu9lsq2vki$~U@
z);{H=O3tH_^Mc@tlWub!w^>}pg;sH4RA0@7R<pQ<>s-URP<;*8zlIw~aV_Up%k`qT
zj`OQy@l($0Dc6JQpL3qiId6&^xDgH9D2f}o;f>r#ir;WO-f+GYzvsrh=f+XIgtuG5
z+XL>F5bu=|6E+$NQU>Ri!HtJ8bMN~gp^fP%Pi;=B;ykO^9;oI#t65yjIoGl+uH&5R
zSlqw`HE_Z3Mowzvf*M);o(p<UG3>9jm7n%ibPmV(#Dr0=HP>fu7<Iyl$R`nnQ73!=
z`5ZuDum}lKJnj;Y{b0~raPN3L1aSABzbJ2fH`;(qyW?2rI2IvR6^QTUO0W`ESwSkr
zI;B_y>@7{yif%uCj)Ie7!eb(!cAPZDXNoUH2T=F{6alsIzO|8$eP-N=ba@R@)gVi#
zMi2G!!M~U6`-3PJsuvD`*7;bzJU6{f^h*^d<?}*5FV@JxTnqNh=e;Q|;DrKKU%)#T
z@LgDaXaOHaaUm}hviw5cxsc@-@}Y$+zlax#Sbh=jQ^fnx@``z(nAI2aZpFMis~=L#
z4`cD{V*VQzZ$ZWUHpG6DL%zvv{U+eco{IHnudMh#o!9D5l=5zyPMlL%=M=VqEs8*P
z6gs)s1~wZ4iR^Xqunh$Lf^<j`d`K}C41q*&ir_33mnlrj6z)`iMPYJ<#c5n%8W%<N
zYq-EQEZ)x<?B}{t{Q=J40L8ttPh}omm>Wb|xJzNMOA!Rb-qoF&yAHz1M6q09uw3B}
ziqP_nz8gmOnPntMqg_Y$fO#iKV_e5DXsqj422FFE=1!S29A_{n)-jerbA&lYFc$@B
zj@O((iWWF7aHnXI<01wvc3jM$cqGIlBWj`X$SWQNQk;q$Q;|ExX~;2+#cPn`8W!(D
z!Y*V4HYG@zNXTUIZY1nx@nPh67`apa5#)G;#o5R)n_@`Oke$E>r22p?z(xgW8CEUB
z435Xc<MAY_PsBcnIF#aK?2}BfQcm=M<AOvFag8^Sc!R!3DW`jY2<aX|EJtEF(t|fB
z<)jY~hqUc;IyOeNl1!5vzHJWQ9*haHK_oWSq=LXghRQi0vi#jd;!R`+MGI0r3am#X
z04wFp5WEU9L!?euk$4puQKBYI1jOOdKaJYIt<s0o=?t%bhBtybjnpsgHsPR&#0gS=
z+y4G^Le93G?N2AxkI42%<WDC@2C~RNJzy|LHm&dWF!zV<e{&}=_C${vW9;p0$(Wx)
z7N?LC)V{?F*p%R|u<8pWS7cF%ogg+ANRrCrB4-D&FWGi-wiTQ!cszmhw?P7L1p@$K
z80oeIEF|F80FNwPpHgS=;_-6Mj1Vi4SV;#)DQC`r2yZ4st4W(dlnjJ+N;yjgV@_7R
zPx)LRkb%MC8Eiz`cm}J_U|))>uuc_L(Kc6Mhbrt!aWgh)##XQ<k=T`UF5}ctGeNq}
ziPt$J+9%gJ^>rGj2+|hbY76fOee}6)v4NGYm`R#`80#Fy1`rwwB>ACp1RFqJ1k!zE
z@?JsW6&hO#QW}m(!(-{3)Z_x8#WT_-I`qEV?_2VQN66w4asr|t)gp^pif0sOI&IkW
zN7P?(@^ephJ+aFqqnT9eccl8AmL^EmNL9_^dq{PU;<+VvAI#&*E_4*6;I6@HY7xD=
z_GVBza!W@ZU;%=(0l95paUL?yL)NrOr;+(-7Uv`Ld=?iV^8ywZBJ)BP7a{W^78fJ)
zViuPm^AZ+s#JU@?9;8VHvUk_rg!N#vFOUqQ?q;k9$w!iAz`9$p9%Ky!k}}fWhV>wu
zM1oVSn}PLUYb20_uI_fM2ZK*?1-yPDZvwN2Y^!<wBo-(0`pGO_$?LCV@hV<_6^m1N
z{S+3b^7^SPUd`*TrWmY%*q<hw3}yytWTo;``WU_zq)F1G?sRNlp<b_0e>z%?$fXha
zQv3m_KcFCrKO*aoh}km7Vc$61kLu&_;5Zyj@p47Las{(Ji{t#`xB+0R%W)kV7Vx+w
zuyAp3=>e65!6W&(J4m=gSNE2<)vpGH(DD<pZ6fYMaWb||rdTOw#=-IuVzytqT(TsN
zyzi{S;H)A9Dw5bwvFz!pW^YSD>Sfr=JXOx13sQh#fH~|71S!xk&>T9JSV{Nb)=*oJ
zdUx;5pmgM!j(W0I;<<svdB`S@nw}t?MmDEeoR4hsSzL&03Rzr)Y>HT1jBJWoT!L&$
zSiBMIZ^UNMu`O1izX_YMR-wNco3U1*zZILYR-wNQo3U1*pMlL-tI*$$&8St>VWAG|
z!|W8K1}rqNIFUC^q<*QzYD|+@oXndhvv?J6x{AdqylD!HQ+d-=7O&<_S5ut$xc9ZG
zQ)d48AKy|YHRKdM5AIa+-|T?&*Rli2wMV`l68L1*zx9!+S887@&7FVnFZoEYZ$ss$
zlQCF#NWL{Bau}0u4IVU@qM^Np_NQn>#E2-0(vg2U8o-*n{{|LsMj@NgFsk2*LbkFv
z6ZvMMepFw{Syi%O+-=V4HjAq`>nhHThBVcjbv28hayC!7ZdCuAvwqI%8@MhFoFCOU
za$Oo({D!l8!>OtMJ=f(ut8eCvnmGyd5H6F<Xv=XT?am~HOOhfI7Fm$QS82`@Ay)k9
zZs)>){2k<2LdxY$a(O!l(tmn0VeI$q%_~S?d<Wa!VS~OqSbYck(*>p(cWK5w;1vRS
zLwGYDLGdNd<PsYwU*bG3aX}CyuKicv%<|=nT*!3W!W(SiJ5hgroHsbmcY@q5G?}zP
zbM6U^T`wPa5{2fN6MW9bP0sfg6wmScn|p*!?+(S?Fix}}T`^5PEu4Dr>FVhWTA*GK
z4#Pnb+DM&>!s+g1HBzr;@nNJsjKX2s!Rk*U$Gb?l%N~d%W7lN%Ktyv)3r*a%(d*2x
zz6}G&_$6`%iCh45d&Z)~Rp&*oK7YwLp}J%?X-<5hi^++X!X<@96OdOSgSHY2D{*T&
zi)=YQodZMtAgM-(@mPq*ttUN7`53sB7+h%)RQ!I!znupA=jQT}u^Ay91Zk|#*dRI@
zW5dP{qG*PC1``v^2%6E4>SEQgj4n1PmeDOzFY=|MxF~2*KZ+Ks7c;uWL5mq(8d9er
zU+6%x4?#g`s2{~^ka`WPUxR|yu=?G|csJ?<^MP!5kopMnr8pa@vnf8H_K7^RX~H_v
zTOY9b2kZd$ERfS%gQL`F4*qF{Y~C%=BJ(4TGl=6_k3o+;nRm=Mp<(Yo?xf_n=aKdE
z%j)|gj;vVvxiqZ+xjoV+DOgCM=^CYcodBj<&XX?l`<wf(BgPgP6*-=|l0oAJO{6Fj
zMP{O*Fao(+D+ZTmG`alcy<M)Kb%zJ>uUi7`W=^M>(}Q&^@7Ib^2Yg@E{NqL-zbyJ*
z_mO^!NsICo2KfqaSaS+y-hVdc&v}t;IO#m@cph6oB7u`G<BpeEd=+=R`oFNk^s1sG
z6vRpQ6{h!DoW+@Dajp77(I@-1+s?_P1`kk&2gm?+A;k`hx_KS#xAQNpdT=5}CQcF-
zlGxOY=8_3aoRZy<U+9MDEc>@>6f8BI^3yF8x{97aZYNL>oY}RUcB>Mx4b)zclCdh8
z;?gT+zy0>OvG8yDuhP2oC7WtJ%g98_=biF-XYlcyRKOb+@Ww1w7w{ew7xIRMtiF&}
z7qa>y-mZwZr~G2xvY5AGanEAjpT*;g`Dhj|Bbol?t(ksKD&{lEH+x#Yftx#|&mSS-
z5glr!e18W92U7WjxEc$qacdrgoXUT?%mc(re)bCn4>Vu<kFz_teneVWi-cNa2J=QK
zUqXTwR=CzzW)GW?Mgo>FA<PVBqacL~;S3t88fr<GyAi4p4BCu@&BzQEzm^*z!d8l}
z*TfIsRqlke1t~@^#*w~YuHIZn>Ms+KULta&J}?>SCA0Vd(mQ}0=}YRdO+EIef#o0A
z{txU2_(t4ylQwhAuKjmI_ZK&GH*$)#Gp1g5Agkp%g~2+7CoHo!8aED}x<28@zq!7^
zoi{QsuOsm~drG0X00;B(uJ@(jH~US{KhE;t${uOqEhOGzFK}rt?m-J7K~2(U50K*n
z<OwD!NUxB~E967*2W0&Lxl$a5JIAqczUI~+6#T$sr%TGi_FaidUspI^SGd9$k|QLA
z^9>e1R5(9m@gs%vBNiXw)JHfk`ul87olS9#8{eqi-Z+soHGyxN!0XfP*ipXiQC=Um
zV{ieH5bKdp&-RJt8X{C3R(rBN)7(J>cL*gBh34iVC?2`w98g>A%OJ5<g~FhM-i#u>
zsxWxK;w;WQi)%fWg}aKRP8CS3pxY>=d}9&n)UYw9z%|R-?jJW3U#aTqe^@cF`+q#0
z2Kh%aCYO<TnVwm;+~tZ`g`;35E9IM?(ABROo?T?rRMh@IUgI}ShxQvkj!h;t9_=%F
z0CiGPK2Z$Xg^YJ0XK+Gq)nAwNJ$N?Y|2>4&ni(@Ax;Od%6214gyJzrx@dyM75jc|6
zf*yoay%yL9?EbW1*EehCq*)tU{y&OC%ZKKVT(*Tj8+g#ta-xrR^wEVrcBYTE^sy^_
nbZmXZlFjHZlD>%S`9bK?)y2-<!Kt&eqm8YWeJew&3w`__x{|HA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_infer.hpp.4EDE573CB1DF1297.idx b/.cache/clangd/index/device_batchnorm_infer.hpp.4EDE573CB1DF1297.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dfc042df11c3761c919a776fd1e036519c4a055f
GIT binary patch
literal 1108
zcmWIYbaV4zVPJ4h@vO*AElC7283cj2xTGkvfRTYg87NROXYT}mZ)QV*qu&?5am-(P
z@Rxl@!b*1jus}gu-Q==2ZN5f|C&Rjb-1Utv+?;!`S&64*PrO?Gy_DtA?|0jr(UH0^
z!@f`9K<%E|FQ3+acjTUDm9U=mxl_r7y6$%@8gC9Xi}9)c;Iq!TDcV%YaQ&eBNgZ)6
z(^((1Y~yaT+CIB9<wsb{A2FZ36a4lqEnM6#;9{z_Jbz;jOW%@z7N`BLCx8F+^+@Nm
z7T+s==Un`i)joOcy0S@l&Gh9nZ{0p-k$#h7dC&v%ohJG(%(~1EriKSR-Z|T(c)Ro@
zwFg1FO@dcFo~-YAWe&T^yckC@Sr*T>&qW_KpW8q1y|?kDtJagAja??2PC0YN`ORD!
zUcxF7V61N|yhOLMA}r!aA&-_?Z-8&7&y|g3TG{^}EWUd4<`J_T%YWAcZJLk&FX=vQ
zR2bPhCx3OakC{RE#!i914ohs>{=I9kTy#L{ebnN&2Mxr!dFC|V)c<9gRkZnyFavAj
zz2eH;qy}L4=YM@8yQ$*pHvuMA4j>l@xTUnDY#E@OV#ek5=Z~d4<z`}GXW+JAvycT7
z44h2BIAdVsW}FtkVdH0=y}@iu%s>Sa>=M#o0-@m0vC8$IR{s(bU}A$=prff{4YQyq
zHLds<(5%?)mN_wMD$0yJ+)^@98Zc+2_OYrq2o*a4g(Xxa3}Lpxq!~HEu7wH1T+7G>
zcBQ9ES@*<U&UF=sA3mvKxp<a?hg(`$S^}n-hg(Ki28kytEQ`dG6P82b$qUOP@f3s=
zka&v1iby;qVI?@P+l#s7ZU27@cA&+Q(h5ixOUfYeWF=*hc=D3+NIV5e1tgxLq#_be
zNm2>Ug9QN-7dZT3UgzPqFtBieh09uFtAL|(6Q%=0Q%6Qe9Tqh(8W@_OsDT;v)PVI1
zJHL20GcZa7rKAzA*Adi_hUo!XplYEHv!Ez7Cs7HQFksQf05c0lGh}hCd1@8$ptZOt
nx!4+%+;~9*1aQI_3=BL7CKrOq{pRINwb`!U*jRXYIk^}Buy3bN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_infer_bf16_instance.cpp.FEC3D08EDEA4CE6B.idx b/.cache/clangd/index/device_batchnorm_infer_bf16_instance.cpp.FEC3D08EDEA4CE6B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a15dc51107ee8182145faad40bf3fd9955b36395
GIT binary patch
literal 4188
zcmZve2~bnl8h`@=#sDRN2;6<Y5(oheiviI<n-s*jf(TUG2r6i#0u?Z}Ds*ZQ_o{Wl
zWw4I&q$&vZ*(W}q+9I~ls;C7N1qHP(55-;47wKI2JM%KbnfskD-;kSg@5!AQYcNb!
z%49KVvH6zt+)o^2GMT&WFDExUqhf$ec0?hQ<(JkZTC!S0l^5UmHLmdfZI0`+n2wfx
zl|t~3L!J*Ef1{p>UyH;xKlq||)xu&=-cc9k=u*6B)zixnoeq@U!GitjZ;x*q+SU51
z{iNf#Yu7st<Q7;z>n+$k?ZD6H?Ydt#e^KL$C0PO&qq6h$vV2qJv0h|p>*`<bWGz0|
z(R%fl+@8*P8>U$<_5|ByyvV4H4cl?x*#7&+E`K*A?8~FVv+AVk!Vv$t#TDJD*$19f
zR1M7a3ri(yAIAStRkeK<avs+eZAq^=x6Y5vRX=d)?Fp$lgXPtEp1hK7|1@>Z!gIGh
z#_se&%CmoujJbFH;)pvP*46iY6zPZO_b4&1|3>|=|Hgz^U$5i48Z#^bY1NYi*|wPL
z8$#3pZM84EiLx0h&&if5_%MC3Db02BqO!ReB_6NzjuV!gD!ugMuC+-c21l+*D*OHN
zgV76*dA4j>C);pnx9jTMXQQK>5}1r<*D~EJdTPHe+ci5b<Z_W)`mAA3vh&+-Tx>eN
zGNkU@MIyQFoA>>T#%CKJhcA6ocIe<u)BJLWSVzl|){B+cwwT=yw|lr9iB{8H$@9Pe
za6Yw1S>#%$h`2mHZT-4i%E+Y_uk*6ZIRSk2PpZJqjC&(*q~0bzy><3pacpdM+*)+M
z_m?fc%O^$0cG8#3AJkn>{L}f)-q%6fMjuofYxMQ^OODniHnrScG14de(^=-+?^|ns
z?ZN1MJs94!J*(2YX*;&^-mfd?4sV;)(_FkHeRuhc<nASl2EJ78;f3-N|E?7ah&SbD
z<(c$8WMiO1X7Z)e=*<lG(iYt`ueD{NgX4RmE{qAaE=wJJ)?>mczoS_r<CY+!u2y(A
z<i!4#^pE?E+{W-ZYcl*7aDUb=TTz04CXU#XP=4YSW4Cp*m!c|q{p1lT5$obQyDFXp
zPP0EhXqRtrO6c|#fri}O6MMTV=Pl;C*Ocb%c$`;~SNPNB9|CK<4(DuAT_3Q=p~{kQ
zXvL(ey*+2%SZ=$Z@g36!#gleABmT|I=lL)84)aLZ5R|Fusu<e-^lay?!t^r-qZ6XJ
zhP;gVC-eSUwzm7ggawi7Z;+wkRrSja9ieSIE5kgH<BuZFtlIjKuAs5qvvO<0R~efc
zl9xAEj0)SzkLp-AE?Y70=RY>@V1)&bP_Om%J4D2Ap{aUhp<m{!sQSIZjgR#O@$75=
zhP=Zs@>+a8k9=s@cxPv|!YCA6%Q*UZ+orIl4X^zg|1<sT(^q1y?l>)O>z(*wNZYG*
z&P%@<_1x{@s`k{J0*iU8?cTLpGM1fv^VAGy`^l4JDc(-r{(XbZ^t}q=3vZR9_yuSC
z2{!HN<muKoNV@1_#IP@T_yT8plT9m~l#0GV(%;pF%xTEpUGE~BXw&|V{vmyXq<fW#
zRadf3bvfIQw`q4L_d$Jwq-ELZGjo*A{h!fZ(f8R_o_M*lqCJ9gNeK!Jnk4-X{j$5V
z=Z|t*>>U0!9k==4s~d~%x)Lra2!XKD>Dow-82Sk%P~cEK3LHjYz+rj}IGn(N!}U0D
z1VI2t=n3FRf&`A#lfY2~1stWPfTIZ-I9g8w#}Eu~jGh6GC0O8CJqsL1aKLeT4mh6R
zf#dZ&Z~`F!C+G#>L_!2k)QiBW5j82TnoEj$q&ixn$tDl+*4ju%JC~GzK>=>k^z{7F
ztdDDtUbjsv-IYGlMVC(>@Sfe6X0T1G<!WE)-O?p%yB6PGul-jGVnU{X6*r-PFcSs{
zH{pN?69I@ck$@-@1&B7$fEW`4h&8c*I1>kmH}QZ3lK@CGiGb8`>UUSX^gWFH>!t%I
z${|DhPZZH2;ed2V6d)rqrk_v_1s$qIL5J#4(4j^YbQp($4%1?w!*m$vFe3&!oWntf
zYjM!wIvjMk5eFT?5uhWq1n3AI0Xo7+fR5xy(2-gabfk_19cd&%M{yMBC@lp#N=Jc?
zGE$(UIU01dmIfWIqd`X-Y0xno13E^_fR52Ipks^-=va;g9jj$Q$Ld(nu|^hj9LIr<
z({iBWbR6h7BL_O3<3Y!3dC>7X9(25s2c5tPpcAwL=mebrI>9J_PUJ+;iCPhKqD};z
zXcR%GrqrCYJJ|M$Ahi}ys9FaoMs1X+<csDLA4NC#zi0cckRv#_P2?ybOpXD<<v1Wh
zP5>h1Bp^yo0ixwJAV$srV&yC#PR;@1<vbukE&vkc;ye2G1l=6CrEs-vuk;M{G)Q}-
zbgp)az8_|33)y$4s9<V)=_=d1eH1=!(y3SE|I2Z_rObfYz7aE$DNWnRiWJ%PBWgy0
zM6D>4m>B~Svtm%<W*kV|ibF}52_Ok80VQcBfh4UYl$4nQlCn}z(q<Y++Db#om>D1$
zD+48KW`Sg_ER>v?1Cq0HQ1WITNZ!grDVPNy1*-t1XcmDKtzy4CF*)k?bn;T2ZPkJt
zf*5ISJA)h_4RrG^u}uef1-$#^l^(uGMc+f0wDh*cR3YmAwL~<?uzo^SC~&9-1rAeT
zz+oB;I9!DThih=)2o(Vwp&@`HRU~kvh6IjMQNU3e3OHIt14nCU;20GH9HU`?V^u70
ztcC@SQ*pp?8V)#K#RJD{c;Ez;0Gyx^fD=_BaH2*8PA#bYq&2kdTK?)liL%q@e)9Gq
zIVXEY&ZM^oiEoF$UQ?w5*?^RU;#y{To#RWVgsDkW5|d>K@0@H(|6x-9VRHZB3`^SF
z*)|+AbL;>IN2h`ELC!9&O1Jl7-9LyC+!e9W9^RgVeSE`5L`05^8Z|o5%U{Qa4kiEJ
o><|>gDVpJU(N7hi2?`z_5=Lkd^|1H7h8W{g?G5AfAKJzL52#Dqng9R*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_infer_f16_instance.cpp.B01DECCDEE63A78A.idx b/.cache/clangd/index/device_batchnorm_infer_f16_instance.cpp.B01DECCDEE63A78A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8eb86213b12319c9890628a003246ab6682e6db3
GIT binary patch
literal 4174
zcmZvec~Dc=9)JU~ga8u6h}?a@21pbm5Emjy1QV$nf(X2HDxji96k-Ezs5s-}0V=lc
zTL+$1Dir}~rwZ0_eYRCASbd616`|m+sC7eM%*>VFA1^aGbHDTDBe^-;lRL!}7Z<0L
z$zn52`PMAk7j80{%-`|Mv*l**?<14dD`c|#9SzCWoYn~C<qzufrGXdb`o4<2b8UaE
z5O#RL>p_$5o@Els5!2PD%X`)?E)L+`M?|{&6fa%-_w|wOZj{Txf&-e($94|tXnlA4
zg!{xl+wRoZ3YIMDDcEYMIo;y&<o(s<4MA9nRa=P4TF%PyC)6J8K~}b|`}KZKVe_5V
zTff@6+vjhzSTA>nxn#e|ZZbvgt~q+((b4Pwo))>fQFwJI<<P3|ka@+`PttR1URKw6
z+SHNhWYe>R|JK#*nuWCBBZ{_VH8gKfvo_6BpPugUhF>uInSc{FvRwa8pS!rZtKVne
zsF14M-$%thY`Z-C{+%W39t~7v{jjiGiK!kNemrn@vMES)tnz%eH8k^3q9EHD+qN-W
z6MD1h?GvIRqpVrBQo%>Y6i>+X-MX}5UUo^pcQNjhSDf5&<#1(r%JBZ9)~8hbe*J0m
z;-dlAwr!AYtl#6i&h{$W;E}{+zxs2Ie|2}$KPxI{o5Qac`DM)-`Z72F_T9^uj+KR<
zX}(OP-u%iKQZy+y{)K+!hl=`xe@$3e<z{lX{?vN87TX!S=h?1)em{-T&>g7@zaO`d
z+N&(`J);<TeNyJv8#<MvR$5gpvN>}@`9r6Iwe8stqwc165eqsmJ}fqwa?R!Ffxy+*
zg8r2_#?($<nKAZETk<{c`}^Jx*%^IM8Q&0d_EAY=Q}U&24@#p3>KDwKVf((d>9=l-
z-rtSkmv-gU2432Ql|B5eY@Y7stnRDDE3)=fWu!h?vDEXeaxX7bm4tMZeo1_&x+tGR
z??*Oi-R7iT`49S6w*QW6BP^=&iirLR-G)ELM=V*B-v461$tTr~IZ@^n$mm<uf#>o}
z7gD}^be`L!pSwOg<V)^e)0)x}e33YOZ&KCqcZ|#SXqBRF%-2(gr;XfTZttjmskOMa
zcvS|4rA6#2)yCQOOzG*UonOd3S--=+`-Q#4zUtK0AGHmtAM&;YxAocUR%cDBFHNl5
z*Zs=}YnKn2aL3}6K)Q^G$G_@m>#96GZfer;eMh#`SFFAjzq!(Tllp`r$*Vkc{;R&z
zhWWe?uUosS`EBL5aS_GjVb!@4)<#1{1~+V3m;E^&dBpat^t3o5a=UNhknIgV_0tRI
zM~q*7OuhSV{~M*Xx}S{+4>}`*2bGMTILzle^(@pa@lyA)O_MXQuA*mK;@(Z$_ITcl
z$A`k+<lij|y%LkHTAT3fSWEuX#;}}yAw7vZiVFS8kE}U5x^&{3H+!CI8egmIF6Wj#
zBW<3yhRE^?tTVQMynbtnkwpjYo$z*@m?%qAdMLe|gTpu<1qrJ<t5Ez8Z`a8V9pDk*
z=Nu%x_g%}WZ#PZa?jxJx&>`+2!<>VpZ*?xdzo7Q>3Er+#9Xik>FxWXrx<{E@cO&Oy
zhqvn_hxYgI_i_%BmgQ#6%yaeb{ftox=VvQB{`Q;d+aoETG=xJ$={+2zLm~j_k&%Fm
zNPI7$911#Ahk_2(qo6~LDCjT_10ANrK!@ou&|yXlbU25D4%gwJ!}U1ma3c;nf+Ijj
z=m^jedIEHWkpLaZk)R`WB<M&z2|Ch9f{x-S&`~-Hbd;U~9c83IM{_jjXdMkYT2F(H
zHqxMDI0kf#jsYE`XF$go8PKsD3p!TEf{xX*pks|J=s1o89jD_!$LTrHaYhbwJja8M
z*YTj^^*rc!BM&-(6F?{E1kee30d#^<0G-H*pc8c>=tR8;I?*VCPD5$fPnTqxX0T=`
zpkbOwK+&35iDK3~kiC9xyXNASHds4&iZpHB`t!Q;MGt%l$CX7OZ12^B7?9XrLJ1T&
z)PMqq5g2fo0Rs*vaNuwQ4je%cz!3%lIFcZNBMl^Q6hQ$;87Sarf(DK@(7-VS0~}*u
zfMW?3IM%=d#}OQGoPh(5CwSm^0}q@)2*3#j0XUHmffEfPa2iBIO3UMurWv6bEz#5o
z{lw1lQSL52X`x=Be$sS4T@yI_d}f?uS|is4Nz>A<u2eX8bFiVBkwmbxW)u)+#sJ}F
z91vk90Fh=A5M`zS(PkPDV`c!cW)=`<<^b_#9*|%b0EuQ1kS0zuQ998zPsU=yFO(vV
z^H(aB1Ee3jlI2&A|6|O#kiL$8Z8?I2GnS)(FgXSYm*ap4IRS{2lYl5W1&Ef@fEYOg
zh?TQ|I5`K1m-B!Gxd2F%i=Sxoar&xu+p2Yra}0<Gh?CY<`U+eW&YxsyjrZR#DwuwI
z$6Cj$2Py{oNvB?&f5d%~wIU94d?OZQrZnv!J5tz7s09TMwWGjc77RGdjsb^TaNuw|
z4jf@2fFtY#aHNF<j<l1&Q5Fg~%1!}CTWH{DI}IFTVSr=o3~;Q41&+0|z;PB1IL^)i
z$6I*dcsmcAU=e^5>;iD2MFdW?i=W)_sfMo4$t!0Z>onMHFe9y|v=pui=be<as*V?d
zox6RA`*Tmfz!FCrstWzIV@S7Kq{4Zd)!QO!BvkstjuqD;VZDT^QQ%N53LK`!fWx#H
zaJU)=4%gzq5o!WBLQ4Qgs!8BTEeRZ@rhucg6mYbf29DO!z%gnDI7Z6=$EsQ2SS<@2
zr{;j;v>b4}ng@>8^1umd0XRV`04J(N;6$wmoQBs7k`~&rVFzn866I#i`{Ltmab9jl
zUgF1t<d1{%(I+h`1C5}Q!#zEcrl(9xPL)lUluVY?dzjLDIIZ_Ewf8W?nmKQ_1IN$&
ztdE<!ho{`j+s9Yw*Vp7fHdgRgn8x%A4Cp^FNI!h!s3^neXss$_1RF7k{C~3pPz<MN
j#&OrA4h|hWBup0`Nen|YLyf8d@#b{bxCt@iToV2V%&qI)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_infer_f32_instance.cpp.889B58511EB200D7.idx b/.cache/clangd/index/device_batchnorm_infer_f32_instance.cpp.889B58511EB200D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ef2737ba86e60c324aad3d1bcdff993db8ef6ccc
GIT binary patch
literal 4072
zcmZve2~bmK9>4<v#sG3iFz>$6M<gf#j0PiO3>rZcxzSiuyb5Z$LafyqtB9zycyv^#
ztcYDy>efZE>d=a?wHCpuXc4g<3zc2ndRtLpBQufz&TMA#&HH`7$?xUm``(+xPBR#8
zdq^ac7f#F1OwUPkl1L<8ws&?;*3$Ap5=o*=BFQg39^byaUE|r>=d-|X>)3Ym(d4@~
zQWxxPxPRh&Mfi5EX!Mk2mT&L9?HQU|I81OpwcJ%v(&hGhQ(3blNRsMvBWuXb!mLvw
zd28QeH(gV2^Y68Z*Y|b3dVcp!mgLOGet`+=2Nkt8wmYmGM@m}6UA6PPSH4bL^Y(dQ
z&!0UlVe>wzOZ{i!;m&8JpN93FoYQ$YCUKsxdEbgBhp`CL?A4_gQyNQlo`^_TrTQ*W
zo6tVtR4Mk2hsolX5ue&}9eo-we*K>6ipwP;m$M_^`nG#lN;P3F;LdeSad|oC!&R5N
zWzpq<!Ljo4SLpGW{n^(lUk>eHc3j9PdE|+Q{89I}Z|?py6_Ni?p}OCml(hXolw^Ft
zKa?M)B%QH5?IjNUfL1s-h8UMyy~{43S+5#?E!pJ##^gN9+F*_B_+{!5LPkyvn6>0T
z`Ly;oG4lW${xBo`97pJ8ZkZH6)mL|Nq&2rLaaP`r=*o>jf6mNRT^Us{eA4nLr{zb+
z#f7d~Vkt(e-69%^=l7(&1<m!_w2hVC2jqNKZDAELH+<)luVfw@A_AD!l#i;<uPp!C
zXT5s?s=YaLQCx|SXV|*TvI~-B3r5Ln8kea$AGd!z^5o?Sw=Y$+du=cA*|C50oTJiT
zkPEXTrab)1rP#=pmelIN>WbcToeJxQnnUi!HS?}MJbuvd_mI35AGDNBYdfCya(s*H
zXSt(fIi3@K`{3xNuN<TAHEqpM-b#GhR`_Xr_0st%D~ngVp8ouVAg}sa*}8=<?lS~i
z4d)MSten1X<F$|;zx*EQj7<FK&w4L+;mVofw<TL2d|nh}VK>)oc$uk3F5Mn}xbujm
zxqEj?Ysc<}l05foH#P@U_1sUp*tqN9cI%Zxv1zMzt-Dx0RIo6{f0b_CRulW3Sahp$
z`t7WyC6P@V%ZeixbEjgf)?T?<J5il~e#YsRp3{xmvzMDwzida_Cp7JRyl%Eb+re!a
z7pszjEv<boa;zSHAARlARV*p1DEz6YsJ7Ubn{^$j=$_xaS=OtVyQTTp6eQtL%BGCG
zbG0YS(;prEBG)JG7L^-vtj(HtCv<Mv(vd+o&b-l{*tR!Qx5jWs{cTy)#_kxufO88k
z@7o)7jFO&T?3^B&_Ml-c<9Fg2wlnVRz7eQjYt6A}tGMT7UHxCjR6jLYnuhgiJIr-G
z=3AOi!k!d`{a9Wj8|*v3`tf4T;J!zNrwi)7A05;6N9Bt|{Zs9Woq6ilE^X2?cd{*+
z^H<yMSNTH6=)C;T*0@P##Y&PzXEC^cu$lfBF{x|+&5}J{u8y;9TJ9qE?;q^i5@o$#
zg%Pc;jx%lA%f-v1f3WMfo(ab~Ry1_EI>y+vhl|IM{=u%Z(idf4wLK^7-@!rFf9|dI
zugj0z(bD#FJIeacX`gmW4p0Te+CMX?;DO}Tvz!|aPX9A)-2HuV^ZEzwq+7CrRB-l`
zjdY611B8$WaEJ~84kc0GP#p>!Mq<EWIt)0R#DT+gIB*0>07vKu;7F1Lj?|IBQ6vQ%
zrK5nONg6m>M+3)@3~-E&0gfeE;8-0C97l4%aXJn-p5%e!bv$qaDF7$v1mKjYlCrOw
zTe5Pza-yB$W_XXd|3#RygIluy5PuK*bUt%*_>$&@2HUh!szmJ5_C4n*>)&ew+d^Sf
z%mzzqL;#^i6cA>_0O3X)5Md+$kwy{_WuySnMj8-fWB{>777%CT0P#j1kYE%5DGkaP
z`-K**X{(<baZx$cwriyd0$iFD0fb6XK$sK*giCQigp>e8N=ZPJlmbLcX+VsW0mMpK
zK%A5V#7lWVf>d}<U)D2isy)_X+hTf!dKv7Cc=+GzEy;IEx7ntBWxgKvYdn&F+IeQ?
zegkg1bdzF{ecDE5#o7TvObBp@837J8p}?VL6gbR;0f(6};BXTT9B#&eBTNKvgqZ-2
zG?BoOW)e8cL;**cDd1=m4IFKzfn!VzaEzG&jy18sv1S%H&cp%7nK|Hi6Av73=7AGT
z0&s#^c<<`tblvl)>p$AoDZnX!wQteR(az5XdkinKP5aCJ-{0Q$=YFcJ|9KB?QC=}(
zVEGiHViX{iVhkX)A{3BDp#`K@L=F(bBcO?h2xuZI1T+yf0-A^h0Zl}YfF{bLpoxko
zXrd|<G*LAQny3Z^O;nG9CdOl+iHR6!Vk!(YF*OF7m<9t)Opk#k&f}noi#TZFDjYO%
zH4d7%1_w=CkAo(`6QD_m1ZWZ}0yGIV0h)w{08K(qfF{Y4ph=1(Xp$-tG)XlHnxuvV
zO;S&SCdE^rNr@C_QYs2GDK!O}l!gLLN>70%&C{Ssi!^A`DjGCtH4U1yh6YVqPlG1I
zGoZ<c3}`Yc1~eHp1DcG60Zm5FfF{ecpvj6XXtF96G+8wZnyiKeO;*o>CdYH2$%!0j
zaw-lqIW-5GoQ4BUPS1fR&-0+ki#%xZDjqa>H4mD+h6hbv&x5AG3!o{80%!^<0W<}*
z0GfhE08K$JfTogGa`tDM?cau=90|x@82~6y83ZU;ITld3a<ZMW(lgTD{m;$LT96(2
z?jZi%(El2gEOXC4+b1S1bB;?)h)<Fv+KogKH*lCRaF{f3xFB<3#u6J&Ulcva$=PMF
zbcm~)yQjy{X<kz%`+3QxMR*S%=A%F~<FsKDbQ8l>a^-j~bR_lvW_>Z7pc&TobBg|>
d0!NQghm0kI(SRVm+&9vg;%JyLX{tlie*x#u{K5bL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_batchnorm_infer_f64_instance.cpp.0266490535848A23.idx b/.cache/clangd/index/device_batchnorm_infer_f64_instance.cpp.0266490535848A23.idx
new file mode 100755
index 0000000000000000000000000000000000000000..400f2f459d4b718482510a32d16aec90836af633
GIT binary patch
literal 3232
zcmZvd4>VNy9>9mtU}8u9%>9353^Qf?88o6nCf1bvD-4>Fe-W`pJ3O@&&nkIYHvK1Q
z?dp#$yE&E3?xT&TZIykz{vfX?l|nsg>y>?zcP`&|&b#N#@7&MtbM9w;=XdYi2@CZ1
zKgeOR{1O6l(i5}e^jItwSNl<CXQfonVX-2OSgf3*jca;0_ewwN`NKTE4ZGEgj{7~h
zTi+10Av<O$w7RZr`N0K^mv6uL?7hjB-Schrn>HJ>&2m3iKU!tWGTb3FO3X3oE6)1S
zfxNeFTu-)WN^#&q<gH@|o=m2C5v;MQ=aENRcJ@4aS*qhy$=NY%KGJ)Zd-46b%Tuq)
zEC;4kUmtPQoSsPizN*yl<G%-w1U0?cI=N-b>xY8oChH-$mT#`sSe#L&h!ScCJv3Q2
zQq?|3vVQI93T*%L$<)D|ubujj_tZsMes%3o_p(C##!c^a;T_kr^1Ph;)(SFjRMaf4
z{5B*n`gmuePH}QwY+v-Iw0EOj<Tl}?`)2ZT?Sp=+wg@j#pu<w>YW@1goh6@!<yEWe
zDqWijuO3-EHsG6mCKRPZ;yW7;)XYbM1jp>LyBkbK>1KQD_WZd$T|-eZ=6zpE1CC#O
zSTzxA=eo0;)#z67qBQ$=AK9Oj)+y4Ag7LN{mAos|wE`b)d2>pF<lXrxGwbx6%=C!{
zmloe&^2+T`az&d@l6K?k&KtbyQ&g?+4fBoJe8GDE33Xk0XXu0DU({voy0kwwFbw~j
zP@eIwq0Qu$Ia14Cq8GeAyxnz2U#g#Bdc)!2H!06?^Vn15xyLt0cLzQi+Q(m<<~ZJx
zS}Z9lzUKGs)#APF_1zV1VejLX=w#4)Qwz%bt#b3~>wL$~du)_uo{rD1?`^*RLf6`O
z`A*CH!I)?IEoCPfDlc{Rmb|{R_0!h)H4C2}wDh6}TJE){><)bTpxEO?<(2KaslE;g
z30HnA^a<L8<qmo|T&{@uWyB>+FQ~<JAiViyQbySxm*@q>(!$sE0fr%iR!1glavh%#
zD|-45{cM&z&n5Baz4GSQUwqPVVw69;@61=n*74Vy)olxm*i|k{&bLhwKR>FUJZS4w
z>2#ReoH*F=c!NAr)RRGuyh_$J?Z|zt+<rm2FzRS-e0X@gXVJ;-PPKyW{m0J4qvgvh
zBj0DNNWQ(^W3=|K6WnTbw{wbfrB`Z+e63wc(;lUbD95u2Uwdw4YQ(kZT?wCD(Oj&F
z%RF1Q#eCmA%FWHCqj)Op!i%RRL1j0GY6rR+W5<5$kQg1<)t~0h+3vmxN&0@OFKz$3
z9e!e={dY)w@>Jo$+g%cou)sUmm2*m%i0OUVJ+|%h&^Ua!Ap4;IJI&#-w8??<&Tn&u
ztdu4y(Ua2aMbY<fWSUHkm;F%mqqkG=+p*e-$ee$9KB`EPyfGD>zx;<fH+{oa?eEow
zed^^+lVpU$3Jqf^`TBg<=|O9zk9afZw+k}<X<@7zuGJO>7NY5a(YY)0AD_gC9%J1Q
zt>zkVO{WJ&PdzlO4$e_sGS&^&YEuJK!|8$1S&2#NlDV^M&@r0+?!UDDyZX$1PujQW
z8SBkN_FWx4rM=ibj9IDT&`+hE1<&T0bCf)iXTz*eOER9{EFmNU93n%2LrD}kRE7eF
zkr;583<C})ao}(n4je%ez!5S6IFclRBV{CT6iER`$td7xk_L{J(ZC5%0mW<yM=4k$
z@MdWBDl@ysWnTI^9Hp(Httm6jY>lqb^etp|ne#uJn`3>}YGY@DzxG-ITYxav&fI6C
z>H9l-El<G<XTlsMqCfzl3KS5gzyRS291x)(0Feq35T&31(Fz)nz+Vu|EOF0gcUspj
z`*oqA_Ca9t2(TzN0tjWJfG{=&2xsGf2sQzTWRrj>HU)@g(=&9S)%uS3aDJ(F1G(;8
ze`e!Oj}+u8?;qW(owhWxG-XzICg%tJkn|dVGwq^Pyd-8?OB&wxSwd6@aEJy04ppJR
zp&Ar8Ooai5X)xe$6%HJ(!GR-G1aO3g0FG3Vz>yjfI7&qUM`<YFXcY|{t)XXJ>uTAP
zb=0kP?cUhy*)L#j=tC#{iMgg$d$iNG7Pd3bI`d9hYBc>W0gob|XE(cLJRxrpATiGw
zkc8(BNXqjBB<BUp62eD76A>byiHH%<L?j4kB2ol25jg^yC?5q)REUBmDn>yQm7t)B
zN>R{6<tS)kd<--(AqJY57z0gAf`KL`#Xu92W1xxianQttIB4Qx95itW4w|?W2Tfd#
zgC@ZzK$8#>ph<`c&?F=TXcAHaGzmEYnk1hDO;SjLCMhOCla!F4NlHo3B;_P%QhW+D
zDIo=#l$ZidN<x7qC8a=<l2f2b^J&neg*0f=Vj47Q2@RUGlm<;&PJ^a^7T7Sq7wsF{
zR$vcEBya%aC~yJfDp(B2N8ra$R$^M*hyQ1FR=hgk!^xTt!+(!R=7hysJ##uoc`8g*
zW%2uPvi|DGh&55Hh#AF-nmvtAPe|LS#T7|GbM*8L=CTcqIrBa;H3{S{_Y-i90+*Rt
z%{S*EQV&nBr7~|HvBknAHtzP+|C_bMaDuj8z_+y%iWWIKIZNCYldh<Pi`>F8KoPC$
LzsmPxot6Iq3deBY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_cgemm.hpp.CD6CA3F05D9AAF4D.idx b/.cache/clangd/index/device_cgemm.hpp.CD6CA3F05D9AAF4D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2464a976a0a640aacf49077baede3981b217b5cd
GIT binary patch
literal 988
zcmWIYbaT7H%)sEB;#rZKT9U}Zz`!5~#Kk2=nID0)FB1bp#hktUzTSr&1lqnYe{vzY
zFd*Ms^NQBd=7<ju40FuapWHP6O<qve$NkcampW$c{c3Xf%we1N8*Zl5pZ%<NfU`U4
zV?x5iPhqMy>@E|hmdjsG+O*%&sl484yL(2Z(5%fjE3a>{STFh3U+UQ#xw(@f)+M~%
z5&!%|*opOauS8mobXT!g3rTfHY>!XB7Pqs<;*4DEbJItODyGjR^?pYDYVP=)WpVDO
z{p-NCd9}f8b2_B*9estSdaPRfV0*{Xf<td6d8zE>wd?G4yt%x<V96YlT-Hgm>SnLt
z)2%&S=e1LiGu)SB`o@#H>T@sDGTF?%Ys^qwnVU2p=yTcS&ly=|owf@wv2p^rK)|ge
zp`^jU0pSGf{`6*@Tl9H0CT0!>ZeDg?Auz$f$;1d$z{t&*cc8OO@OVNxHxmm`fi9b_
z2$(=9Xuq6uV)fMcBLYlpFbnv#_-$bptUH{2b5q&E!vaj~Flj|iMQfNeLs4p4@p+(K
zFn2O?g53vmm)oA!cR6p;<}>neD+wxz!&JefnYh5t^;9YAp18}oj)z;1MUNlm<k;<&
zIWcM~%0RvRGW;4aho<(isx}A}I{}4xWq37UM!|G3a)CVvGZLl?=q`}AV5+;lm|NcV
z|F_`f;TCfe^MtA5;TCriN8(91Ng(kgog|TXQchAxJZUFsB%ZvJJQ7byQ^^44a~^JG
zO=TpWilzz@PgPSDiKnKihQw3XREP7{8e0V%otrQn7zT<mit4aPg3&<Ff+7j#*{24q
zU)cG@yP1IjDabF4@T{VsqBKko54Wz2t{N;9f$^G?s02(xFf$mKR=#Je%(<Zs3x2pv
daZz%yI8Y@McWj%zgZtH-Rg7E=T#)1mA^}MAUFZM+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_cgemm_4gemm_xdl_cshuffle.hpp.CB4931718165FBCB.idx b/.cache/clangd/index/device_cgemm_4gemm_xdl_cshuffle.hpp.CB4931718165FBCB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f72ee91a34add4bd2fb28b52d7742f549abe44e0
GIT binary patch
literal 10532
zcmdT~c{r8Z_dk2TulembIEIsBiX5*Y!^w~|ry}7hrAtXSjT%thOI<}#NmAWtpi!kv
zr6{COT@|;vTirCF^1T{Nw~6%K9NW9&d4AvL{`u?CyWfws_I~$Xd#$zCTAPpCxN%~Z
z#~U-p?U%*D%Y$ehkEe$Ihc6FXZ~y@hO&;%;)RJkBmOOS6-al4<@UG<NIh8j|X>eW`
zyUoe=_fd}>V-C*WbkaoG<fW<At7jY3^S$@hdiqe&Ay32Kl~-08SDfo;TvWF)MQ6Ot
z<yWoz3%AC$549Q5o@deGA>C@x5+L<DW2j>G{BS^U>29r~@~MlW7nZuptu704a?Owr
zTo|Q0;?WW+;8N}x_u&dom$!Wy<rKBic3sRZGvh6BKTiK@nQh+^mHak+#gf?xip<rU
zjIGL!LkW87g`Y%<EyI;2*RTAeQA?$6Dcw5d#bGzMJ#xag4t4YRsXI#-ZkaA`{`|bM
zzrrB*=LP(L<kHra=^oR3|Eiv&x=87w^}F99rc9i+=WYDgC;d#C`@QphSy1raG~LT%
z>T~0DYR_hWsFyNNB9fLxKhjHdc@b_?q8FoSwy@x4V%R>r1^+Zs;kJ{bUjG`U8&`KA
z=!SjR{mWV(Eyl&<$th{J<n1pUWwX}fT|l8(hp+0IGS_v1OCnsB?=$Ncwqa<>nkX$M
zM|kv)|GfEU=8ZYl??;6`{He|C^;GMM+Tj-yXC~%98eU!5G=0SB%#_*|{iyNQl~#cn
zGjA*`IX}&BMRM%<`ID_igdH9bn1A(SN!_)R*#Y_Sx)o=xkMy1yv26dEdrMm9+%rG;
zv|ya7`Fq{kBDS46ef;yD;HP=Nx#za6c=$@MAnr`Vo46AVO2uRSygHPhvNcr8+SRI6
z>&j=89ho!jXmi~e!^L+OKQ?)PUdu1AE-XORPdfk7iBrpKziod~TfKZ+{XZF%>k1lX
z204{3Nch>{w7;Li+?4dcTUvEzN2dCFZQv_qI4F-_UT=T<%}*;UgP$IZ^PXzUcb~KO
zPc~6++P6u04?S0<8uF`#y4Ut=KDC$L<(nDJX0B4UG4wagQBnx_Q~AS#VGGx<Of+70
zIXL7;KRpGDK@+Yeo{mecEIhF3^|M_DBlh^-(pOsLe@p*K+mize`!89Nu;B2f-+zm6
z`+H8+o~!LD#_wP_J@xnVN`o5r8M=6_(DcYVCKh;${rMv<|E_HrZxu3o<FXT{kK6t_
zEQB`RmXdSfhk{y7%a)`56E5|yJa|ehn)ujn+^V8WResHr)GK~ddXt=Q{dwAFxa{@d
z1h*vnpZe_%IUpWxtmn95+ZN607k9=7X0e;>a?TtXI@iQgA!Mg0{_<ZRV=l(!cq$k_
z@ROP=3>w_wHh=<NOXSWiP#tr)t*Sa}a$<|}>T}PQuC(so^lD_px;d#<+ZI*kt?oa&
z^7XCQ)Hm6?I!c*RMS1JgWh&P+4!aZ|x?@ou?K(hcGN|B|XJi;q9|Z@@wDjd|6m9wT
zq4rbBB?}$eW>eBPPo?HTMW5GS9quscw=^5g#!y4m)?cN`#<L?rw;T9`PQALLCA4U{
z<a+W^aQ?&I7J9myjza#07`G3rlh3M#u4wc09+xQh=$pnGyTA_f)B7gSk5q0w3(;)x
z(n>P_`NLJ8n!l|qT|1UX_|IcL9RB(GOtbwav4gkQ@Semu-SYFRI_t82xYFl_=%1Pv
z%<)g~2^yH{d#WREUh~cEI@`9JKRfe6cXXr6Xy!oVs_lZv`|h(x6}j(=PMhBB+VRiv
zS1nUoM%gn99<K40e-afQv3NEu$K%1{@8u&x9!lh)qCSsxQ*ayD_24N%fV>%C0jykm
zwD8&HWEI%YfX7BCM`(3Dc#4RO{>dNO;I+bQ$8rG_8}ZmWz^?;p-3?~Z*+1_=$M!8d
z`EalykF5q8)j+Si!7OGfk-JCNXIJo{6tO%38c%>hcY|3Rr2!H5f*b1i&<?Tu2}FMa
z)9wbdj2}E8ci-upnX<VQQT!rGjg>W+WzvGc+>KHCR_ag_t>ytGctDA}n;w=2*?%2Z
zXKtuKdBl<dlrw;4cY|5HpYjHMHBB1NhxTYLkAUbAnv1N#EdHaK+#lXZOI3&3Xp~?<
zuwmDuha5D<IkzDqT^;sEWIuu5fUXB7&#;Jkf3)JKX_}Ca$e#T@#a$0f4xH)dc|Yis
zzB*J#<O)hqK^b>9JuJVt=NAt@wMY}HA{MXyUgE9?CI=0RaM|OWdQi|c!W~NK4w8ec
z!7R)6CDUW3$0hNh16t%mAbN-vDQkLIikF7`>Y!)Dhqj320T4YvEV2f(guk66m|_@q
zT@%Woc}(d)#k}i*$&u4WjsI+&@mc^?5cv=-I7I7pH<%^n*QWTKljFtePy?})QT@v(
z`|bv_Y;rjGoaaAdm<rTI^T-DLY_w&v2D9w8IsL`;`pv!SP#KMqPiyDX7TpbIIp%nr
zdb9HLBuywlEZ+US#a$0f&i?tfSE|O8C`~9tWS{;%CS4CqK3Tu!p|t$;T|U%9b9f4b
zPk~W)gIUfzilnNyk5JQsG-AmGin&10-C&mdnRCk{O?~3!A&owg2~;wHMt6f*Dh&?4
z$yxvKggVqgEFUSsM@qN5!7Nq!#R7vy*9DqT0kQb@_ciZ&^pGby&6+Vy?mQnlp?UlT
zH2wmH-3?~>^SROCvJhhz6{v(p@KE+p?|NYJGw-0alb4Dk1yF>@skBilE$wbF%PTXp
z<Nr|z&E!Kz^pVFv<1zY(tidb|86iV2HNrCjsD(!PLJjzWc2?G4mS!i9#;px63pF8y
zSSI(MY~J<2<PW@|M{d7g0OX+>BG&+w8lc|YV3u#w0)-2mw0;mk2C<aTf)d)WyTL4B
z!E?jy^*LV%mAibPo5{+}+lt0SF|3EChZIW$`$N8RmmkFblR>-WkqP|p1~N$hjIx0h
zR!xSz0_?7U;aCS4wgCtlfEfEvhHV0-O~8)BEkL&gNI0AZRnwp@hxfyg`=JMi8{o(W
z65pUyZ&13}RWR&Xdf-{w7TW>Co}*RI(YhqIKSz(k7>^5-yS;gjsEh3lU0icYAj8gN
zW=j0K9T~$e6)$zfUaz0c#ytyn+CElhSWg4bf!Lifte1h;z`$<T#IW85-UG4KF|3c3
zj}#jj!%nuEEXBLeuv4t2NU<wnSYIn&Db{g@UFf}VD*iCTF7jSPps?{_lQ6B{TaQ@<
z#>%r8hCK$Q$Dku#Bg3A8(o-ZZfx;4~!_lun={1sm7fSDv_!$&FBl%mPw1uR9h0?Di
z-b4vEQN;K=Dd|p%n9qJnx}U_ElrWRy=ig5`yX0kP6X@JTy>qk`-ZWL6d!udcK3xQ8
z_70HT0cKd+)Wapxm{2t->Tv~FTmgeHLA(5;Hf8nR0ujQr!z{3<@R5x+`4Zq?0y>yq
zx%=S8YOeF(lw$j_vvw5Cwt!(RU^tc;&6ZHwC6o?6lWF!IWqgk^!3U3MYNlw&?y%sl
zHH<k&PsVEy|1f2LV{_B)K2z13r<VHQ=6tojgtV(q98&dmAi`N%dX{#_LmBX+hip=;
zdVvUUf$m#ih9_rK6gL0OT9-H^r?)`sEf8VCfDva3o1Jbc|Lf%pSZOe$I67xypUFvn
zKdH4^FvFk^0f#(D#CHI5hZM0a{=p(POR;@*>nyXY&j{v++Q-YAcIfmW*iX0DmX7yn
zM}${E?-l1na14dMSpxVaK$qKASwMxq=`?UQFGw|T1rkueqySHB<W4(RuTV@X59;$p
zXZ@z@i{yiv6lwMt%^aio_*|#i44TOx@o`#ooYv*&Cuq?L5@*t)OcH0&qAU_;(;C^d
zCf7fgW^%dlhozgIaM5*ZMl-xY>D{2rvHx>*Fnmx`c)PE##+{FS7xHHAek7B8T98j`
zVeLUOqLoT$ReXS8O-3G37E@x6IR0+x(7ivphHEiwpjn_D78b+KGMh!9AhRF>g)*U9
z*xHcoF<}G>XTk|o3WcRmA8RPXmO)_|iOZp|oWz%*@G^;Ops<F-wNO|~;yNg-Bk?OJ
zd`02}N|-<q-`_$Bw~#oI5+;&3i4rD}IENDEkT{nT=92g%B|J&uJW80yVfWn^x87Ly
zrk{Xe+km7ESaO1B2a<LYe+81SB<=u`4id*gNh~D#$3aOPi8nyW1`=<Cl8q$Z3Z1vY
z5#0BapmP$5w?XG^B;F34x05&pnx;StuKzA*x{JiSq3LcC?}4U!I6TSAQTwXvt{;$Q
z#{uIwV2T3~oYEk@l;txxB$?Em$e>6?mc?KKgW^)fr$7UTtev8@Y6DK()Y^b^8*t@t
zJFsde>9No%783li&^Z<o{Bh7Kj-+peBe%kd+<05zq^;18!%1*t5=l>imMP?UnF5ET
zz_A?vE@-)nqk9)OzsVWre-nwuy`Q@dUM9}ckY$$wektkgWVsslc0b(vCtc33`9L4H
zHd)EFWO1+{5fKj2N{47w>~^Mjs!Qs(xyYfRgDr#YI76OmIkyM-&QnS}mtr@i&#+U2
zc23Va_}Cc{@@S1bS_=>Br%_=$D@SJ_A}kRval*s;we7NfRlII+pJN>}(8^uOC0;m#
z_S+}}*kT1>*y}+5I(O(Z>`kD5lf&3!p}4m*6Xk|kp}N8de<Pqi@#LfU6SMoQ1_yVj
zcW0u?_5M-%qdgva#@Qhu0j4I#bZ4CGm<h&U*)zzQ>J(8X*t??BkTSkQS>rf?K?#BI
z3uTCN1?<m}w#)Kf{5>31Qp#P4F8+1r+Q(i8<M2t>xpwE9zy)Lc9W<US{ly*Ummqys
z8|~g_y$jk6-kBZ*EZY11&}2Vp;SJEFfx{u?$Jo7T|M>ShL>CWUJl}UO3etMVVmRSO
z178IEi=@8F(rzrXW!-5vV!QzO7f41~@{JjB#EfLpnR(--F>Ev(9Sward)f>~Z{{#o
zM<f$jMvgm1wzej!l_~zwXX-ddM?GW-I_?p3xG?XmkDnr%Y9cgAghRM*CPP6o9EkCT
z$VmHjvyW`*Gc-=1kb227db~J>JrDTjfhOL`&F<9tw_CqC4P;n9JHH`V^AH;AH<3gB
ze1A<21xtgSv4%73T<Kf_%`=*3#?j^*%_mR+P%Gd<HWYXQwX-BX2h`4yxDconlDG({
z6_L0Ys1=j=3Q)fSM0jZo+X$FOpn-*j95|)-lrq7tpJAU-de2DwoYH$v;un<O3li5*
zdNm}jrSxh^98Eh!)5EaQF>DO&5JTd1w8J_QucsZ>lQ@=kh$V3x?GQ)ec-kSJ#QC&k
zJ}u=!uYB4upB_%)(fPC+hf8Rq650%hX2=d`qYEUyNZVbchj8>#+OCwuWwc!xiOXrD
za;|^kW_7!LhnE&8(X9Uv|1nryXm%}>tcA9C8)#%lHqme-iAlSn*%)XOL(=1+B%bWQ
z8QN?n>EGKS%_cyb1dhJ#PJLr^@v59&4&}E$)HZIgdYy?(=NcurMj2usx$}j2>BtM|
ze<MOVt&>ihV8h<|an_6&Rq=h~jq89_9U!c-4os{A)9{KIwwTf>CcSep<xotG#@;&l
z+Js|w<Dx8)c_+}q1loYpn0>TxA8mjOMR@0^&TI$$48y9e++Bpj4ufGY(~`^dP`q`i
zX%;oB%U3w|Is4KKj;@S8S45*vXM!IyK@jH`vVnCraKMfu?Q?GIL+#CaND6_{z#dO<
z;KaQ0du1zbqhEgnqdtP3=;GklqO3KA*Fq5C`&{`@)aI>Kd`H(_;!hv*SZ{Rsm%aZ}
z2Ds9JoMdNl5wG}!Q@QeDxrd$S!6z{46L7=A&-$fU&U^8paSAAtkj@&4wUA+hq(KA<
z*A6$}+*CA_L_=#l1TtJGi6QZ3DA`Qn1Sm=1aJJOh`%2oZ6ttEaAg%!toZ~R4x+hHK
z_7m4>P~_2B3B_w+zOxsmm%T9T_xg|HPoxyrfZ!U?!%BgYHz2sdVdd^ZD&FEVBW?i6
zaTB9HTZ9KdB9@g`aSz-oDkE*_EUI!d{+_YZB6RA2At%8S%A$mlVvT1|*kN8l*<l-F
zP+i0E9yJ2T5e!>Tsn%24oNG>?4HIZHY#8|sk@?#3%_shKg`vP9Z`Z2ZZzdq^OQ$u`
zxuEg<)%(-4hqk=@*CUpw8?ALs(tXfJiWdg^%#m|F)PM*Tz@dUGT3(#7{#U1~P6~Z`
zTnwvy7tkK*&Cu)-C_2KGDQPwrigHO@1Vu$8u7IKn65ocR+a#`rqG}R1K~WQl+n}h8
z#Br1;j_N6s((DdOw1dQHlqikF8I&l4!=-%Z)73Ud=Aace0ktL|#ClNn9<-ky>o);e
z#{V7ll}n;*9nZKQ>E$IZWq2>?H%@W=|MU`<Kg{*b4J++TM#kum8G`k+;?C{L%D-Dn
z`}o0%|Lj?tR_}0Io@S5J%yF(1LbE4m<^+i|X(p4zSu~SH;%u79CUFkU<Z$@fJ%1g)
zAr+v{=Hm+qWOTCI2t4P?$mslSG5N-Qdf?hEic32$Coq4fw<6C=bf|oH{#A9~TOb*h
z0e%@7_sT9Yu$A9*F<tHBB~Jd=CUM{T&EwO)vv!~_EYw@rQw~K{WxYi`2;T&t*>_%y
zaI?P-gx7%>`=HKDLwJ+J4_<DZdh{;aDZc=-fZ^DkKWt6)nZGN3L!XWPYpS8qIs^aC
z%3*{kVk}Pgs~^qFNMB1G>LXNq)S_*WUC7{xy>PlY=3|x)_}9{?(>Y?9^h$M*PTzug
z29#!S33%TEdLESKk@VgL^$SpVfh?H!F0fyR((7dZ-Uau2P<oG~_b$Lchr;LN_`M7A
ztx(!Z_U~Pw?|{+{lHR*uA5RJ6xq>{y_AcNjQ_^IzfZw~Ie}Ix6AnD&1_!%~f5@vDZ
zzqFBSRZ!L5jMVZArSyf;z^6&)v8<FzYv8@dmu`r$9PrCY3z1#DVGDU}S(q)=%RGec
zLXJT35lGxS<w9{TiHo4Ph{P38TtVX7P<)%j)lghb;wC6=B5@lOw~;uG630=*_jgd@
z9VAYp#AzhXpu`y*ZcxlZb-9yC$Q-J`&?+v`$09=JAiGk;=I~bal6{bqMQ63>`}po{
zTfnDDd*d9D@G5|)0vK|ZTm|&2z(9-}FKQ*2FCRA&F}49g8!*L(1jEKcK`iXS`2Grc
ziq5M`JcG{so6H*+_WL}bIkoVj^+_X>e{HZhT0>Ep>@Jl1;zvsIks5?uHN)<s4foL!
ztT&w){n%jqu`M@C`|rhDkc>jLLwh{mCu!)V2|2%y{C|JUf4Xoqw*J_sHGOoTt?lSV
z>+0lIM2G{5ali<>U7U{~b&}m7W6|=$f<uCQ(O<%h%~GimzF#v9;7}2lC3%<?7r}Vg
z3ULn8^9vk*^YBjc!oz+FKi&C<HYS}9**^(>;qk1vS99*wlzSb(y$<4D&A3<No>#~g
rCS>nxRe7?9Z*4QkWT3gk)Xd7#Vt_HvvWJIf!@XK?uh!h_K<@RwJK?Aw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_gndhwc_3d_instance.cpp.8AF56F1E7AD14DC6.idx b/.cache/clangd/index/device_column_to_image_gndhwc_3d_instance.cpp.8AF56F1E7AD14DC6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9acab2618ad9eff6d3b3ff50964e9f90cdd60089
GIT binary patch
literal 2168
zcmZux3se(V8Xm|*42DEfUO^&|OhP0O1_{p)3WS6N2$ldtMKnkW$^r%h3RGH_MQUk9
zMXgn61K6z6MHG~x)<r$6sQ7>?s|bpM5m6e@Dn23wcEX+^>^VE<%s=z}|GnS+{_oBm
zxl}C9T?_!ibZKFZGJib?0Dv`m=jE$2OK<=%LjZunAKDXh@&*EyzMXyUL1JF{yJlDz
zK71`@AKbq6)h$N!$)>Oa#T-J%^oKR;afHmL+|~lc&bLLaFV4};BtZ7+zg!35CvUXw
z&L5{VgCVgcJsS%8{A)KDZWafGm>t)2y&mZ*H>f9PKnH~e=#hC$oEHyG==4wTtMcE!
zz2c$x@@(yB-}TOPNvNhI+2i%d=<g?9PkEg-yzz8>#0ZfcrL{=1?q6-)8Ur7_o0;WL
z$;&wSFla=+spf7FvhmOf$noc1I@?pOJVOfC)R&w;g5Xy^Snl~?H>EH&x%c}`{<jN*
zq>KbYwXOUdo1VP$LRQ5Z1t(rKe2d}xqQ3QBbAxv0t067-dA5fQv_e-pA=T~orQLYc
zK9!62ZvOE$!Mw@lMq%f+YmRC|ME_nk{!Hboe=c&3w#!$Eui7y;Xg0c@&A1x;{H&(6
z^c%o);PQvEK!e2l^E;B;xyy0y>;^T5_3Rq)kJpBZd(Z7WrB{Hf4*@qmJ(ZK4`&++Y
zWipjuB^<kA{b%&ya$jNR_Q#jh6e{A{_7k}wt~+kHK0&_Z>tB!<jy!mHg-A8s_MKo8
z+s9o>T&%g+R;dIMsnPiH<^zPy!F&s5{O9Dld`e%slwcVZ=WxcYI8<0<xv@cDldON1
zd#~kqn`C;&AI!ofn-dGzhA-kyzwq@v_+85~+_JA&8GYsLqfQ*AO}z&*Jonh%ZS(+3
z-Yy@X!;SCHSNCix>~F1R9l`nUZAu;2wOyjWDeA1Qdh(<;%rTacey2OC*4)Z-;{Do}
z`7ayX({pGtP}rsaa`!ZaPi5-1Oh&LC+lAy?GegT1T>IX^0`a4J|CqT6aW?nW{dVBQ
z*8`DcypB^HYZ==pHgt*1{YvllJt>cok7kFam3K%Toi0bF79YC3>!3Sr9f`YWQY)Wj
zwWuzzULO46oo4)z-HzZ&zRLS_lm0E7bh${3k7@MVgQq#}+g8)aY8!u$q`#?=yY_u^
z{ZFu)!x*#biF{lq7|N;9@>DC5A71n-yER1j+%L_#qt?Wz`lGrt1rM@V<gXg^syu!(
zAt!BScaSL1$=PP`LT_NCa&o50eQc~^kwV#@x$oq$Q&q~TZL1ILPw4iLmAv%w9BIT&
ze#j1%>X`h7pWn@H;0<c4w>(tTcZ#&*#aq^uWN2sKP;PZVCAd!x*M<XmMLB83=y?po
zS7>Vj;HyMnRXl)L;jIGZ0$Q27$|{D2)AfH95P>+f%)m4J=K@yNJM<L%x<@}l1mtL$
zi|6{y1*|-t>SSdf@ab_P@CjOWz&rTQ1+1)AX5<}3{{_}3Djc^QAN!X%5uu8xlKEK(
zHHs?35an5#UO+eO1&IjBg%mii8=<;T1F=Cw1ag4_=S3s11MH1ybT}0sI8d!RWJW~T
zT(%q=#KO1jcgmtAplLS?H(N{_#&2dh=SOA7RNb**zDw8+Jz$v39{@pw!llGxqd<hl
zrO7cB;QuTDB0MfH4pR}aZ}M53`lE-7(KZAmAx1z%C`pvHCPWobg%}+Xp&_&s6M_Vg
z5L1u#g0NFe2=;~f*iuA<m&jX#p@dk9+q0@*ywN`biM(Jb0+qagoB)wcbY286GDfiW
zL(lB@ekp%Ljat~cQWn-;+Hxcf-8@(T3$f8C62Z|JQbo5VHBY6KqH}#LeCAEgeN|si
z3;1^vnqKHpS*oMB(c5P$N=Xutq$a7PXwoVSVIt8DGJclW25ooUm380mA{V7xxFt2C
zezaI%q6<x1TUgHvxX?1#Evy2{ji$tr)?wV}ceskW(D7B!LXTx|GS*l}LdbK4bKez|
zfC5tJCIT8rqnli1gKRn`80{70(dP+tm7O}y?(L{BCkv-}VIwep(t>RNtx=dQ$=Sq9
zwWT_ncp+QJ*~AN5!S)z0_C)P59!ty~Y$}>I9!D%ay%ae)sJiqNbsQ2w5)&m91<5c3
zsFm64jprjz{aGHv_>yQ0|Gg5h7XXL+Ud69#H{TOso|pi9mXn^n5v@gM$Z#Njk;P)d
z5~9^oGD&JJ60)o<rQsC&k8B*M0fGErLCD9UTw6{M!;kI!|FU+=oM<ktbcpG;!rj9Q
b_V)4hU(R#%Tp3|&7a5ggE{+Y8n8o}D*`Q9a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_gnhwc_2d_instance.cpp.8EBF74F132E896AB.idx b/.cache/clangd/index/device_column_to_image_gnhwc_2d_instance.cpp.8EBF74F132E896AB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4d6db89ae97000e1d76d5f0a4a8f2adb8306e341
GIT binary patch
literal 2166
zcmZux3pA8z82&Mn%Z?c{h<})93??!&m>HL8l%Zi5a{b*m$*nXgmvj-UNNuax5tZ56
zSfp~RlytF*jmjR2wcCkWiHenNvr?$+KhF7#J!k)O{`dUP`+o29J>T<v-ya|n3b$)O
zkXx*1w<IoY69$4HBKoGLrEJMkgP`}C5VZSrbFd`!Iag<T{zjNyVbRM|LvF3E(lPT*
zbs5K{<gh}C_LY&kSkKw<Tdm7TDJ9&g7e{+sq|^IDVj=h|rq+0#YR(mT(Y@ggd)eKp
zvHHwi8T<c<YfpaJ)<fL!PVhZb%J})D!7GnDqmvrDx5VsxZ=;fNHr`<{yR7Z~V0`Ax
zyR5GIlSdzX3w5n!p2(beF+1?_^SDjRkLDvDKfM~eYPP&)@>+bbju)mOzF;tMq*ZfH
zPaNI*sDMDdF0W@hOTzvjSD)J1+vs(+H2dZ`j+0iVb7)h`{b#PB-WzX5B^iuWoUy}(
z%94x<50Bb@s^hrad)Q>BDJ5RmL&)n0r<o6nb4w+z39nXKK4{WN+i)xT_IUKnP+Nz!
zXO5)!m`>tV5`+9W@DEvM!o%;~YfZO_i}dMHhku`rFWT<sO+12xhv!fWzZ?|}1`tbK
zH_kOVRGF-5Pu%K5UKwLu-f^j%&U<){+hLa*`kwLY>W&cC6Lr|3+xsr=H!5w|7hSWH
zyfvp!s`epYNJz>N(#?(()3jC(^c{04?ay78m*1LLZTefkMv~{-!&Q3a(hUWF?er+|
z7X_^Q8#nD^rFN=!k5ytmM{uFd2}94l%!AUqn=HJA{(gsA?@)*lX+>2B<o#}CmB;Y;
z6eBBXU$$YCVCq{W{b=D}Ncx<AV8r#5tl?<0?#kCUFJ(mP<Q=+gqgBwo%>PeH&oV7`
zLBo!V)gS8Q%<b+`4ks+msX2t#L{olktf0QzQz@6}O}(kzRVcjwU2U-omK^?Ma^TsR
zPEQeCZdH;<F>m<9pYoQ(^$Cy%x+R_wAG+5Usc2_ijuc-Xg!czl&pI5hsfarGWbX~?
zRs;Tspf1sQm1ELYUCRXc15O?)Gq4k#`NxJ-jq?pPV)0eWy@kBzUfkQpr?p2#m7jeA
z2D7@J+43g%1I;!n=H9$FvN>VKzWW7V`1`4ZJ-!-yxAEF=PikOQ<C)p7n+J3Q_ev^W
zda#&L*JK|fCyLdtY_Ss#w|94vv&fOSuF&nq)nc<aZnM5MziL7wIqZqi4;NcH$L}Ou
z&Hq>*G#TUAKc3f8nR1~0-VBfP^KbUp8)T_;z<x!BMf>p6=C_X4=L<Kte;zsuAFt+5
zO$V*%e9qyWKR+>_z<&D8P0pM8?uXRgFMo8pFdttr|8}<fVr;hRGxgKise2?bndohF
zKJ@5dZLer84hjf_5C)dPSqNxl;fc(;>*%;5;4u!`fR?qf+71f=lt(*_49hn`(>TZ<
zEt_G@xC;T4Ki-m$Z20u`864z?mJP9n+=T$jDRJ?sRp_69Sbv0^UiW94Dh`3&VUNWw
z1Qx+!fJmb;uldxOLJSVUo8p}og(0vh>;wkk5Sl5?X;BrzGGy6;6&f}KI`zwwOH^<O
zok0%(gDkL(MaD^#AT({MZmAEnfi4xT1wRT}=Spjac&4uTuc4W#w0;Z*Au&jSKm`V&
zFem{)1?E2!U=TKg?GO0Y>rD{Lbk+81pv&On1%LpDkU~h2N(A#^0ieSn6oe9?L^M84
z0Mw&vLFf@m#Nx1=!BjYe9n1~~D9D#&IVxL&RrsfgCoGvt6DBMn*Q61YbWBYq0T6%=
z)Dguw0FFe>{xW}(8+icrZ>b}3KCaC<m(RtZd&*kF5&#kuxwE_gO7?CHxilA<gZA31
z+b?>*W3sG_;@DS<rk7fGFq$l^wzc1dQsUk5o<I+ZyzssN<*q!IG{C*XKzo;N$#<`~
z6kq7C2WXn8PF$R+<?_=mHyM<tFvSla1xV0y=nF4(+?Tf0V=*fXte`c>Wo)C#qYjjS
zgdtIt1QZN~s&ti(p;MKvvN3GxBEfLMy<ux<E-K7e-FQ*h3l?@tonUZOAxs}{qU443
zVG|`UO`m3><Yj5I3;-|4qPwNImSDF)E1FhZN09!cgIsx<{L@R+TADk}12m&ZL=yu9
zrNnL9q_`icDVtNx6d%Egu<)oMv|<g)iKKL<YH$z~1kw<+SrWS~5ry9IK57{3GIb5j
zFL2sA1iXm2-pzujB@&PfboC5jM<?DIzRTKm41H^7GcMia|FX-Cj47rosWfxTRaUE+
bEL(ey!wNRp#?D=TxrbMXs?gWfQ$_qQ?}l54

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_gnwc_1d_instance.cpp.C6C85A0C333E012A.idx b/.cache/clangd/index/device_column_to_image_gnwc_1d_instance.cpp.C6C85A0C333E012A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cdba2909702006c3e7e9e2c20ae610ffcf2aaecf
GIT binary patch
literal 2166
zcmZvc4_H#?9><Rg7yl3lNqRtrA}Z<v1ylTk&5<Y?WQvEyFvB#RlTty0va4lXtJ^=Z
znpt^uvYgZGkL0wKZJKM<Mm5duG*@1=SedT*r!D8aN1x}2_j&Glp7X-*{k_lU_xpa|
z_dN+Bp>V4$00`1VT2;D6i2?wC0KaNY-bcG^0N__U0MMSfo}^NfgB{!F$~Ty9vhphi
zOT^Vma{t=OL#wki`Y}m~<Z2Vce|9Q47xQmkRj;IvT{^8Zj(<hB`3_KQAm8%Lahfvt
z(jv~h#zshrGMf7vUf$yt@M%XR#dp1bE7avF+rx{BtPimy#dTgQzD!x3Gm~2|`zkoJ
zaAxzjCsW3rjb^n~VC^=IcBkGx`L7G@rt$X27aJ(=w?`LNO2Sgh9yP7ky=oiYcwwSf
z<;2<A_S246>cPrh-A0!#$o@P>N4Qef*3z-2{^#ns0&%Uwa4TQl8g}n#6-AJ8b%04R
z8yb5&T{bd@%J-Z7XPZb{1?N>K_Ldg+T$}k(*;%e{2`!eLGc@PNA2X<ct96Z!Oa&LU
zZ^k-34PO_;dfOUf^C(8MPsZ&LkM!tI^-qIC_md62y1?sTy<EBn^XX^Ae|X>x&v*V9
z<!fy;lbvzTdThE}G_^6Em_u~l<j#YaYqQC_d`s}+an28U+7)T(9yu>ZTXG)=GTwaW
zq9O&l*3+VYW$(lFx~}>}?&udY{6SaGk%(&H&UO1r6Z|~fAIjCWqSt%toJt(}p{dr-
zzW<!z<EBM@6?DQ%lTIZWJLRmA*Q5O7iqrh!JAR6C&y_w*vGQq^En($|Pfg8qd}e)Z
zr}Ea@KKCh_(Am+|kTg&;{AR;I`=g#rS+DBdY+m2-WasIsiEGmMo0p}w4<=$?9WeXc
z4>Uc#)AXO2tcW&(w3}R6fqTqs-;*EHuRqxH{6v;eFDnwJqYoeJ4)fb(A82eH+Wt+q
zdv(*XFY`Hsm7dKr{OP!{G^wKJ_VQCPA8MaO9k*Kc_cq1a?N7j6i4AXq%Nq^JB^|s=
z^eh_RNzs@Y>cgw0P`xt1R%~B<xid#}R!ul?g6yrEtaxYlgDChMG!$wBc$E9)v9qTQ
z?X#8N2&P-!WY4u|opgP=((kW)(Jwf69)HHZLZ)Ikcm<tA-_n%#P4~)1A6Dhn_sW?C
z*3Eaeq9}Jt#@;ikPIend%xi3JU`?0|U5#qvOK5-gj!s#Y<nCY!N9(^Ps_1a{?z@=|
zKdsJm8p+YbPs|C2rtF`Or_@!mX#>s;uYWfN4CQ7h!+(B4Iohm0$TLmQ{-oUPdI4My
z#~;;eO}MNVzt4f<;Ijoky!A6S?l+dPP3C)tCT}FW%Jrk$1Z}6P0kuw*S_t1pyBxp2
z=tlyKSRf%0fLLf&;KBfB7M^u!qJ)>hb%9u5J)HJHd#qX*khIc@jK6kPvm6VogVW2<
z%YqgLB%NDy=0L~rjzKIC52szwE~^#>B%PPOMO_Df36LJF>=$Yu?wF}=91L?KTsT0^
zZ+g*bjfFq~C|ra=U<??CL_~AyD4+g^9)*Q)Zn%)eh9J-l3_%uQA%+_xWKj#mc4LPi
zMO~T`L;5PWowCA0OcpZ%S>%OoIqv!~RSHjgW4uX7%aNs{z#o&EJ`@TfvWSUDITS=?
zQ4^4IsQ)3MATEo$9w~?LVXgCSB3h*qqPMv$96c|?4u?m$uyQ*Pk8oibb__g1z(Q<0
zHXh-AaQVu8ensrKEezn}A}s)-jQF7?0{LL11xJOba!bVEGa@Z;Cd8CmA~v5LiIl=Z
z+$8P>BuajJZ|;*I#U9wGrH&vaVV(2G0@0Z+j=aA6(z#T4PuT*t2!X&+9DC8DZE-EK
zi!+MdFp`7eEP9{wTTKm>*K;18Ug{_!0IoED!jMWOo5;e30B!+X2`9isBmKaU2q#9O
zz3xXo?hpE!1tXVk;OXZ<R0%M48=fX$2#XEexHMoAlmYwZH)$PCfk5C*5`#;96f>4O
zm&o3P$n20t{E|X@bPEO$QAD~0Kt)mMhz|H1qL_4x^;{H}z6cmP+%@EPbT8bPE5>!P
zu@QgtBnI#N&-unkIGTkEl0ce;%OEjm7A~8_rXk!}N9gjvwwupjvls$~(6WdT!-zv7
zATK>zIe!t<d7IT9@UIQ@a6q!fIJf;5>m(_VgxnDTkXrt3R;6WU!eQJNu?-6SH;k>F
z19q7s9w#D13A_mQqW6i;%bi?6UPxFtKVtQJERtWSM-Y?t|7KTET&ZsEbcUz5kMFx|
c{{Zf);CBMaD>-XOE23j%)<Q{SjMduz0=VWo761SM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_impl.hpp.545E1241F8BEB14B.idx b/.cache/clangd/index/device_column_to_image_impl.hpp.545E1241F8BEB14B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..940df9a9afd1d13838d4ab44ded568abb54fc218
GIT binary patch
literal 18496
zcmeI42UJzZ*7%ts(d*%I<<gr-FT$Z0sUjePB4BTc4NGi6QPI3u2!e<Rih`jiB9;Jx
zC`ebaU<8d_Y|&tgqA?nKOTGc_IOF@)`hV-KyqEmnT5&DU+@E{jbI+OCv-ixN86yV-
z1-S<+DfOE?Aa+(%%ml2Yq|}4_ofk8A`f8-4<f)>h6q{2&e9WAyuIhK^E$z~Cm5<O^
zUNELoGJV;}>tjt$cBweC-BvkHGd1dvx>edxJKHxm76v!%$*GzVQLrS<@~gz3etY7-
z#&Bt$@3v`vaoEdS@A&7V=ZV#Cyz)|g4etK7Px+)xUizxc$E$8+HTFMqZc}heO3&z1
zQ&$GGmg==Gy7f!Q(g*L}ZyQ*a)s)(8>BZI^TI09&zIkBbtK*u#9cyf!IemHZmd|rO
zc`J#zpK15~qU-m=x|dW1qB(hw8<+T)_xC(&czT@9%QH*%c>lxm6sXU<sob{bs-K14
zy`}T6OuIH<$Bv)szMJ<h$7cN8=<}g#3=iae^6Yr}l0(n#b@{4K-n>cYqYb++9{5{%
z)R5&bOcP>Wy&P!me#b?}(b2i^Tj%HI8!VOvI2aGixmzAznBRSp&6^2cc~hTR^l^N(
zSm)C@CztI%_Qb#^^Of*>V4yrSA?tXYpq&0_*_zwS_8<Al)+hQ`srYi->I0XX4YY3k
zxMKB>A)0BCIp1vjv7k^x$H`H7>BXV*cK?*zZGlp#?QXx{gICl?_tmZOc|N>t-RfQH
zEoU2qa;FzAU(BesYsJONf#wI&j)QOX*L_!DzYv!kJYcgIhMLd5a=xO>FrX!2$o{xo
zU#+0e1*3^c8>clj>2_-fyF4J-;&A;h`>(wA-ge}+-R|xuj;FVzm}knP``PI%(rSF!
zI;=R&(IIro=evU~w>}+v#PROy5l24T`Qqh+T}8jX*SzuuqwR-`m5RJakL;}<tA8oy
z?Zo5ZJ4-Jazu466+x+>JDxL*#%UtfMbgzptO!^TnDtqPL<JaEJ8QbG#-mm&G?9-@<
z>Y!OA)iv|aM)eYI-HiM+@Xh2%t->B@8aC%+Ho1IieSePA^d+BOeBN)&po!aKw<g~~
z9@E|$kG?qJ=STjHkIra}iSu9OReWqs%&1Xrqi&AN*qLU2>3PYrAun(1-Zva?8eLX%
zcewrXM2&ea>aoIisS#6Ew-hY@xp$-CwPW+EUKW_m7PC(r-Y<U<AFEkrbah_WZ7zdU
zj9q#a&#z2WvR?D$O_ic**_Wec?;3bi?^^1!5jTf#K6!YUm2<_xxv=P!>bQ$%qAp)|
zda-ZySuOt`JTm5gmK2&iqs-^wL|Ia+*}3BJ+HH47uFqQW_?FKJ>wS{8O`$0bYhxoO
zNi==Zj_v|B%~6T&u0Av7>^}Us@;4pT<C>?szWA+q)8`4%1ur#q*I(I^SbICaUN!A^
zr2`k+Lgy`-HPKC3NeTEw7kPV!E$R+cjg-V?P_hhaDGH^or0<eHprwA}>-W!J$e^yV
zlGu!-&8UZ>pekKMyUm<v3I=LGPjVM~U3-g;3l*#1-8-)J`c~8K5EJnVC|Lou6$Mq%
z(l;o%I^?N&ci5e%Btc0Mv``dO#i|Cr-kvh#t{PM)DiiTUy^aeN+wB^+eexF(=E_h;
z#K%$h<48?WP!-4T94r#oFR+$E-(E`MVPt(6c_<31;$Gl7KSYcvkij10F57Xp?O0z?
zP?bLac#)ZvJb8%}nv%QJA^SQMq$sFLz*X6ohpyaD*MNHDF21_Hh8-6w?ytH&_!FP$
zE}b>{6iJ_w8tp8oN?2im&h|-Pf35*l$Q}CV_R;USP;t2N^xBx~gEcjv7ZC^Q23mAn
zsCa^L4<T&Agg7bmA&*#z3@VYaqWGY4eN6ZxllSK|pp@LfQ`b|!<3hy~SA}JL?bkO|
z0~!*sx30Hg$AyX~X*u4B4_BTmg?-5n96@GBkc*=Dpz{1{ug635CTT!(a)*Ap{Tw?k
zR6P5%$NJHJTFa%-o7|xRnKmFlML|^-j(_mTvT%`u6uJ_X8suDq1}X}wlKRaRi(AwE
zDwW|UJ(WbOPpk|(E=t`XXg6*VtRMF3>A?fRKdZrR<S9brA@&^?`u%V6DwovWExI6s
znne5%_j-t(6a`hu`f;6g_SA59DKsNbQH$(rQGlYLDmjhUD@!Wp+|+;u<SstCK876^
zD&Bf>!h-`Q^NHbV5OEWhHDNtPK~+j0s>)B+IG1Wbf1=V?x36W#g^DXQP99vk?NyNm
z>_f!Py3Q6I7b>m_H9GG5vs<AQb|sIv7wPUr`ig?8>>uWSBf`WbObx1$I}~AA5wYve
zf~p*L+Y`QAC*+h2N{Gr6EPH|t6a`f|va4<LDZ?lYO^C=3ZG>GnLYbnVDvd{7-`#oi
z%TH2hLR1=&RU-;i6jbGee6#eaiTP?N^dKtxk^O$uS5bUWaalG>>FMC=&TjYyb$dhJ
zsI#Cdr=t^#4{w;{D~0;xE)__(f~a&BROOPBP5KVEC%-5|MBXR^cF%xvML|_A+blHS
zHM)O;8q^{xdjZ}HbQA?uxwmD(;^30*r=`%1{Ln#UauE3_3aauzW2}o=%gaV7bSEnH
z$i5!=DGI9c$WsCy1l1-;p&wDHLfTcxP*G5o#}8jF9MVhANCtI?N+D7yBp$4@penz<
zQ8BZ9kTgRIyAYLfq+3p2y0f4v&%aC>yTqr<91SRtH}cT+(C@fV@vGiPCk}kxB~u3F
zM0^tClUPepP?fg<ea_tK-9p^8Hc^R#k~re7I}56!G&gF>ynfy+etaTPL=D;_1u+5k
zPk=+I?-#^GC?rB1#!1jW2@a-?Um#Vgeh~GdwJnI}kjFU`M5B)&-asKY&<M&ON{J#C
zQUBk?@yU)azL|OiVDTDoz6M-r6D!xc7<mPj)FVNR18#93k_Hq(OapFdJZ}YVTN%^n
z1|=O)jlJP}r1c(|o>17Tiv7y0&o?++{X`HwdU}}Ab`ZoqJ^PqZe=dlgJw44PD4ee#
zdiV0SpuScReLQ@8sWTBo-(J2JG?EBnUz@%T)J>>rpZB*vcYTHm7LNkwqrio}BNi)Q
zL<Nkb9Tkh!a6~m6$+#YlsOLEYnPebSCeJ`287P!-HX4zQxbJ79&$7{YCeJ}5a(H<T
z`YeZ+=b{n0ygUyD<e@<J{e0w?kNQ$pyEpNnxAE(BKoG4}tj%dvQ=>r=#jt@21Th`@
zr^9e&MJ>p^g;|ZJ)8QwlWp_4``;|efGU!Twh<<_Gv2(D&?syrwUq-{IeF<V7*2u$-
zw3W3EEz8(@edBwK#S)}jg7j&xZD+3%<iofW>6Y^HQsiBVe5qVp=Vkx;re{6KqXmQo
z44b3~m)b35UEl4fbkxS;Z0T%OYNJ?;nGiFP(SivJI9lkv(1*!lC&Y5HMc#`z*`l$F
z#xc!1AoLCx$QzGS1$3r$)NVW;)vyoadgxKla|W`>K)snf6PaZ4#*~TtGf_WYJ~0zb
zVw{cavXRI*2ifKDoQv#odA^Cp+(ct((_rxh(t3fk>2L~*X;>`{%V|?#F&%5BV=cz(
zvDSL5&3Gp^*@;c*m{0fI(P*VMuc$xVv_=X21{m+UGn+K6lb=%%^&iH*<<-BspO*aH
zv>I6abDNf$Jn8Q369?AWmqVlF(2dT51o1J{e+-?dMfP4j&v_{hSwI>x30fz?zSM2l
zPJZWU0KYl}v6z6v6L2K6sYL9bhyy6w4W8WJv1nm(S1hJr=M?Np1?|Ran}$Wk>DVD1
zJ27659oJ)L#+g_v6YKD7lZkD49+!zfW1NLuv#>j58XmDkF)3no(1b0TupJ$x3gRPd
z^axwfp>aD~J;L^sMJE@BuOct~NW$oP=)WG0r*$HbVP4lpWKKtbf_MqJUqZuacXhbC
zsYGx5(K!f<E0O0)<V9;4i%G~ciRV?wa}{I9;M=Q|bw;@=x7VR@3O1qd)Lw^rX;`0e
zI_{Z{jTo=TM(c4e#+f)Y6AxsZg-x@t8D%;@go3ygX>Uaq^a&-3St1=!x{Pwsa_S%X
zj@ZRXWN;E$(06oQ^}Uka#YNFXa1``73N)y1bvK)+G(q*V`NR%dz@QcoMg>^>0oeTj
zM8*}+t^#sSu^Kv8Ll^oeSgeQ6^*pB_*A(Q&<f+ItmFKm{buD9e|4jeAZ+fZz%dJ4Y
zB#<cP!_<I0oDw3hA1-=M3}OY)TLD~YXCnpz^w#tIEztXxaUa7cO*51yJR$A+0t+v&
zl+Hk~_!0{*d47d;UST~Ze~oos^ZW+uyy5vB7T&S%(-}Ky*IiJ$3w{{xNEGvS+OB?a
za^-t3@~yfEq?Mdn*w|4Zt)%$_(4nm)h$+Z8g-m-D6Cy!OMaHQ-uSLdd8Tad95^+CV
z!-&{LK-U068tenopWTa?E5A<u+0D-_h)(n9pieI8NE^_K7Q`6en0~BwV@Ah(%4nh6
zLQh7EhA$e;Xc-v13=Cl$4<h5iaK>xEpfw<ZaXJ{A4#qRi17aR<V!RU!-pSvu0$NtU
z590uV3_?RI;9yoc$Dr3S=*M^^idcz;F-}3jDQF<$3>1)of*EI`F`4Ky#@T38HX27c
zQ2pl~TjXdlsYO>m*Fai}fxjQQW?&F`mi)7WpF;qR!a;Q_9?g6>Wiv6?Vx&=w^yq+0
z5U(MPYe<jwWkKA8du(FE2Rh6k9iTH;LtSeBb2*;bqXXxu3F0zfy9~IoidqhAm-8GC
zWbr_k$>V{2Jm|wX0mu?~c_NS{^1KGf*6^GPWT`x_1G05Ir-Pp9z>3{}1CVXt<rzSh
z!E-K<<?@^lWcfUA1F~%&v;8*y`3r%pke3$$SrKD836e=Vl3C2IeO-OssnH8!v{&?p
z5vw51cbf0QUSMJ5!l8^7J1yp9-vj;cfg$UnXMp}0o_9jmozR_HjUZM*%_^wPxDIO8
z@w^d<8<7K($KilD97KbKAbyPlzsCI;$K#NA7F>uo#K9|Y2$j>6T^C75s*b(<N?^Pa
zi0q{@fpI3!#lX0jXA*VZF;6t4TVVaHC}I@?$<=gwf>o)L92%x(ac`|x7u-IxRyl~|
zY-ZhnR2q<+X5a)8+*A%DIgNe-N#v;<L2{bo8$MyfsNyYG`;cGw9m{{mn)DY&%8xmW
zFVQw5J!u(`Ed#yTJEQ>16cE5T2UzAX9=WIDwD-o4U&tL@{9FQPhaLHLqr=_WubqYe
z|7W%T?^*1q8>Me+?-|UHU@_V=+KW~Lp`p>k7|nN_|G}qXabd*54;cw8E_PhZ$(Dn$
z%K<lUO9jEHU;w*QK8VO?`3g+tx}jz8gFnT@jO~aqQ%fG}l`YTja&Pv(JOmh*_3-EZ
zvgi3EJPuVGYDRObBrci9c+jU2#0BOHIEpoo<>(I3yaTLgA01ceF>|cBsR^;s3}}!6
zt!N+?h>O*JicG1`708&wWIGn=AVd&PVD%H&h-S0K-HIPsF=fs=Vu~|NW^kQpw)$)%
z_GE1^unk7gdieDAPXm;M`L)Ch19}Gxq(PY`&&g=vTPXb&%Bj9YkwT{rNwYW-Nk{6O
z^=O|kpKzAOi5?a`f;|Fh$K~KN#-trnK`3L=j`@Idai@U(Dc*^l270G~0j)riWy1p#
za2RzmB+G_<60sj;np)~E=}6zRM~wB2<wg>*qhq;|#3mqa;-md`FnBv3?dzdkJ#?hC
zB9H-TL_Or9c>{EA;Qj0>6tRkr_G?k_T0Yw6qEWdlnp3w!s=okf7a&XeLlVUT1Fe4A
zUL-BozPiBP?GX?j0aIE<QR)@3hV$QSS0=*%6$ec^3=l*|6-SPoRGc_+QE}0vZi$R4
zR9rc7Q*rx1GY{>s%Fca0<}V-C&)S`FeCm%OI;7STV2=c7Pg7>?c~kWp$cc_>sMjYw
zB>`zCpbyy|K`ccErO1O0ha`&C7+T%)3eVQ9TRHwK(p!pAzhcy%DlP~OIAJr=HJk{7
zG=eN>oTIBZ<cW%)v<RBBcI;frp*N28{Jd#tuM-C3`|IGib!<Wzd${%XVtnSLyC6=s
zn9N0&D2phLrdUkjXsX3jj;2{m<7m3YbdD;3Sp~4BW+I6DfZ0BtD}h-h&sD&zisx!z
zR?Tw_FstFY7MRuYTnEhRc&-O#^*rwfX8U<Q2viONHTHZBK&64_LqO#a&*y>4d7fK=
zN-NK~P(K&;qOCwiw@`l*&zqtCW}fq)ejd+Tp#Bz~^Pzq|&s(AXR-Ox>egV(hp#C<V
zo1j4xG-l7!3=Nujz77qp^IU`EHAtJ9l|a_u<h4kfF3$;MDkrZ)+H|f*+!B)4BW=1g
zBajs)`F^BL7oP-iH@4i3ZRx@oNf2YpGM@Kf%RM~r#g=<{F2|PTJXc`L3ZD02%Y8go
zV#`XNtFUDi&u6j8S!_uo8d?9qCg*rQk4?_=d;yzW;Q1mpxyW-1HfiDc5;nQS^JQ#u
zndd9m<O<^@lUqE`{}ex$_=%Z)XL12*&VV`GO!5H?d%%_xlCI{@F7(|ph;-$T(;vxZ
zeZy0)OkVSEpI9e{p!lCte|L`|&c`{(lR6(l-a!Ez1qBUa6w*CJjSXHx+(LX9g}Q}u
zG|+7zM}yo3aTMkj#?fH6!5oFVg>w`sjnrVjFhn|pqv6uw9F3HY<Y=sPEJqW~C)%=c
z!X)!a98I>J?960Qwox2Sv7N%vRNJW>O|zZG(RADC9L<o<&|vqRVLyYL*Upm8;$*XX
z&9-Ek(b8y679)+}Xn}MAN3qgaj`jl6y}*X~yK-P!&T|E@sQ`{lz7N>!<GB*pRPtN}
zY^r#!1~%0^*8rOuo@;?kEzfnprjF-&U{lZYeqgho=Yv3g5a_b!YXI^Fo(}=}A)b!_
z@d%$`IuGRMdHDq(TmW)*zbinv!gDK-xAOWofpC+T-v+{Mp0l7{7Vl59p<Xu6IZ!W$
z=Uix<3$2)c-UN*|@w^!tZ{|4<8t3u61sZSRIUgG5^Sl)rZ{@iF8W-@q4H|Fbxg4sM
zLml=!`=HuBo@=054bM%`qzPIx{bp#=%=2|<a-HWJQ2z!rVft;*unlq<zBU-y21haZ
zN@Tc_W%$Sn0WwVDc@;8T#q(-pxSHoR$Z!qMX~-~*=MBhk1J7B=FpKBCNNq3Dp?g0B
zvJF739O=*<0wf8C)GCk;-6kQBoijQ$$cT@hI<?4%kDoeq$cT@hI`znikDogGkr9iZ
ztw^&K8PIVhNv&dwUD%Gr(-Le^!gDFMDCK!Kw%d(en0^_yE8}?&w%fz=UTn9Q=W=XU
z&T|E}tKfMbw%f;ZCAO>NxeD7=@q7v!pTbrw{+`Chr+GeujnD9W7F(Ug_DugAwmQf2
zd2Dr_=L^{C0?!w*)kU6LuvH7sm$20(o-bpo%RFDfR#zA&w*_3BHf8R+|M()w>R+oH
ze(Bfq&R@QwwI*53ansg0`_!@MZRu@Ir@C0=r{7rQbBiPc!{rtUaZ_{rQKpgPW4KHs
z$;WV+Ml60jiO1r{lXxtCJc)1Lh+=w;g$gWwoQZ7D*9<I!Vbry>=WF=PBNji-JmP=K
zJYw<V%p?A%%wzi+C7&dqGi8#<Zr?FCz;r?K+M0e#)+VRqhkv;q{jb5>+Ozi_FR%Tw
z<S(D(uA5c=d(@e6-t9BZu0iT_Ku3pGBoWZQ!H>S4Al`&VH=#GJ)(x%J2mSUJw2^N=
z#oAA?Da~VUcsYJtoT~l>l7(tUw#~?c`Lt#f+RXMG3F0%P{fy5_JwvX~kUz~*WgHLQ
zaV}v=Z!+b*hSjg(p3FHs!s?H3Pr8OmOVR?Jdk*O{XE~IAyWnew-Q*b^{Tu^nvVuBk
za^qqsEoM<#qS%Q@=h8X3HaGuqXujKjeF~MP&r~HHTPbP%kmRE80xog2VDthmaka>8
z5w`@k1~{%^i;ZNb5a^o@0@*lqGY~iPX|5e$@D4uBRSoT``7~EGjHu?*Tt`Tb3w~I{
zZjb*FNqm|s6$PjAX|7D<m5I1TupBfhho!wX`QPv|j0jcv!<$d?mn5g2kh%r_<-L_|
zC?H1Ox#yKOVnM&F2L6ZMWdEPtq&T*=OZcq(Ac5qnfL;PHVU3pnY!iSB<3ymB$jcLf
zbs}(N@(n<5125kI>^A^+CeHwR8N56LSZDCxF9dppyu1)N6#~xKihy1bFE0YtMeO^#
zoce?x%^jaf-uO7uKaMQu(xE{5zxqSw80ZGSu9A-ZzqG~#@jKx09q7*{1IK~Kah{uj
zTQdk`@{c2qAZDZBY&4v@LP5+y!8ttVqTpP{Wn(Wn=C4*&BhOocG)j;T&Goe>2{lR?
z(;O*rz3sb_+4H_fmhX`ZO+!~aSX`mC>es@*cRu~;JD>Ks?%DX|nTNq-JM2>|f67+x
z=){4v@pdTP4rSChB#N!$G^W%@yPj?wd3(?w_Dpr(T^O~?^XZ+xt>90$0y@hgzPNMK
zKkcapW?d=z!tYW+`;f{i$dL}I4$QsvXzaVi;iSzQpg{xdNsR-GM@SBrO_Q;B44NH-
zA9hM%@fI?;g?iE<)Pb@ezVEkIX$85zU0=H}`m={DmN>YThiv=HgA^SDl8Wg(@j##O
z@XqkrDnVW2|IZ*rYlhsR^FRdcH*_By$)J`&p$vXl$P&aoP}sv*qBtKxzj*BIndauU
zX9xf52XV*lXD!RU=X!=%(JL&xVjHrs_!<kZd47Y1H$1<^!du2o)BRp2VBc5&YfpdE
ztGe^Ad>{7yx963c=YAXU_SUR<f7t1rkOj27Rejj!kKagfYK9nODU_D7E-6u*ouTGK
zgDx@1&J#1#q`x=$a%G&)5Y>*T9#DdU`KTUPiUv?V<z@bl$iey<e>>_w-BHtY9ciPT
zP`Z<k%AIGMXe*z&IDdqVQR7nLU6#YH%b_OQ!yE_o;-D2h#Y6TXLA@2wiY^G#8;~Jp
z=ZPq49>PEQulCunb~7@fu3RA7?QLElCpy~|h<Ta3K{nLh$P;6?o!FP{(Wt?yHCT(r
zp7u$x(KWWhbZJG4w(@-4i+_0{>C&^D$Y~o_CH~j@TRQwFyWBfZvazO(gHCZUnAyZ@
zsQ(&%*xPp9a`liCL#?L&+ui7m(#bk)V|U6(_}>Gq_OSCet?sA#y)4!Sf4i7J-Nm$i
z9~1XzxThXzlXIxoIpj_|_-*ZNz0>Zh+7aJ#2f5uro>U-+cahs&#&_(yJo`jdE14*s
zz}-$@Im>fB#N8fZIn8hV)Vq3@lR;5F`KepT=oWINccgm}Ny=>n(q4f+oVaM;Tw+m-
z{Anm^pMV;aAa}|V#mP(R#pv=xdrX7U5@<+=V-m&rOR7lEUlPSaC@qBQOtJGYCRL=r
zM8aq1*-I+;HK+Q3xVx9>-zNK~n=D<_B8|5LN_X&4vGedNZKJ2M`wzcP8tnWZUj=z~
zNX0x__m{MPyT_!*M#(ne_OrXpgY#<~<fN#xk1NfQw$Cs4^@Da_8?grtp%(t)tHvdo
zZyQwq<CXFvVL+8>)1K%5^2XAaZ%d1{+GeGY!Av~Ri|2PcuLFAPc-{u|wlSuThjiA?
zqs+9vUO7b{4lZf*`@<7h8Zybbsm>$N^cMs%8Tu#lM^77&dIK`18|(z}0&>5=j-tMO
zyZwN8E#n#Sh3laIIyi=&n<8haXHEzp|Ir3L{HjEMBA~eYP4!=vjsSGXMfb}pQG2I}
zMJ1Y0oHuvUyuI!JYll_)rQ;72>*THeqi-6Wm0DTbI$QQ82e|AUZR{<ST&RkYk}dn#
jn|*X-A1&BNXZF#KeY9jBZP-V9_R)cTbYdT^*~k9@f2U(m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_instance.hpp.FA1AC606D0CD0897.idx b/.cache/clangd/index/device_column_to_image_instance.hpp.FA1AC606D0CD0897.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ae0eb0a2aa5427090feb265b2935ff2ba0a9660b
GIT binary patch
literal 2682
zcmYk8eQXnD7{G6(bnn^@)|=U<($a5+Qy8^h*B%t(a@@5OOF~lmUA(&i8{i7KVMrV-
zSwKJoiDslG3{+GGIQ$}m85M#-3?l>g2ZkUH7eowdf)F+`uMKbSx#XVQ?|%37KF|Ac
z;bqOu9cqGj@zrIW?fj;+k{}43v}ZSUtg)03#7{E_qBA!fzPR@6f{Lq09=@~sHd(q<
zXXy>AyGJ>CN!zW^vjf%pa$`MHZ;bnPt*czsiWq|+3dzjNhp+cs+?C}ukDi(rJ@+_Q
ze)&N)|L&aqONaI1;rOK$;};I~x211<_S)vP{Ng=Zm;A6gSAHw|-;}p+aQyv=whOD)
z80+H|msYrM=YBtP@Yueq2hR6=wQaD{{K1tk2aTS7?RsWlY=7UW%fG!ac7wSx_4>Z8
zcQ=l)m8EZv)^2agbidO6-RaqCMdKaor)aeHp7FS2=dSIR@*mlFweqAl*E>2{UvpnX
z^~tlJlrBHN?$p@G+%3K2rG?w3l+zcE{M?Xg9qM|1`QqsZhlsb9<|g&c_Yxth>W{lC
zbMv|vD8Hy58allG!(_#b^=pZ2S9|7RK658_)@?q1=d4DdDkDfzMV1u;Q6@QQtNDC|
z=h}e{eHw*U(g9VVp%92Nr6R8vPyX}Q#}i#AGzui?Dpgf!ArNIVq(Trx{L#qtd(PN)
zjY2KyT2<}*Lcn)#{I|dV8TZv=8b!IJ!>aJSLcn@Qs;>Fy!bMv&ikXtusdVZ>z@OGL
zckjAioZmjDQD`LHsA_B|1bp}7ij^m?Z#XxhQItx$QdL=22>9*}UdaAhDz{&Hc)sy_
zzMnieaFa4<NeAr+;8#e*Ia*|L(oW!<oD(<~?E=olxqx%iZs6RU8#oW`0nWpDfb-H`
z;Jlm{IEto#qc{pUAMFFq$N7p5*R@1ls3&eVuBTpIs3ShDB@LpX6<1SB8bu=@lV}2D
z7R;i2rA4p+vhr3zb~fGy$j;jVIYdV*$hYVO<Pu$g+=5$_uk;8WKwjP}$d2MEKtA3F
zh!L4qoD(g{iYy>b;6&L4`Jf<Ehz|jZh>=#@skLNOi~@>@F+g!4F3MLXgan`@pA=-5
z;!}Xqd>T-O&lCygRjO!ypyARbJ98US!)L}zw4_d_tHEQ64+Vz12S3V?S~5aK0`lu3
zOr%AoC=~@R%0z*SQ8D0ROboa<6$dWP#DPmt3E&b;0=OiV1TM)WflE;-;8IKqxHOdp
zF3qHY%TO8MGEAoEa66@x<tHDUGMu9)^9?>N4FbO*fYU4yYiN<l=r;mqWR1X?{3hT`
ztO+=?-wd3YH3MhyTY$5$7T~OYD{xlU3Y^Vv1J1_UfV2DUz}Z=Q(cyt2Wg|bC;%4{$
z{nFE(k==Dt2O@Q3HC`{??z?6iU$f-vIZ`5!K@SpvjDSqY1jvlcfGo%Y$cn6hY{&-4
zj_iOO$f3s{t0kSt3CM+9fZWIp$b&q9yvPfPLKGk$@&RHHqnAfKi&#J$;s6Cv5Kst(
z07Xzlk2{<+H=rn>7>WUkqd1@hN&rftB%l;Z0ZOAZpbW|s>FA$3x*|7odzD(Uc2TWc
zE>>exqf1U@pef*z8ya31cH$2n=sZ1hdHbPesS{Mqs;a`vL~)H(7uHtXWwaziGXeQY
z49B#{#L_HqEXM-J(Hw9b#{n0lgTMv3AaEf%1YC#<6<r74md^i*WII-8sr<q0d?;)R
fvfTg8{66Q%edVl***c;E3xb#>ALf58pKSjFF>l#E

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_ndhwgc_3d_instance.cpp.CBDECCEC6F806814.idx b/.cache/clangd/index/device_column_to_image_ndhwgc_3d_instance.cpp.CBDECCEC6F806814.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2651b4fc8fbd7d5b2758bdfa60baf63438ddc3d2
GIT binary patch
literal 2168
zcmZux3s4hR6kW(h42DEfK!QXd*@Va^1POmaftZj0!4hDph(;*^DW4#qLZM}-qScC4
zMMbDVELB=WK`H83)S;r{2Z~e?6a@pKG$2*{L<)36=Mg&7o!L8k-o1P7x#zuInM5SY
zm=8gM1W8VsB6}SSL69kSXJuz57vLahiU2`5-#104W%YP1d^`QZiNw7AqpDvJ)PFN>
zKeuVqs}4r!ncBdE`HqC<$q%d6;RwmkI1QWSyWZwDyu3&^5JRaeKfeWY2i{aU*nAvU
z2m6N?v?p!u@~lYG-pTj!H#ntgeLdJ(q|F?kf-U7Ls9owbdP&qbs@6PvsLX!<_PUe&
zs|yuFUAI~i!~v>;Sf|&6Lw}xrJ>har`^MSs3BzA{f>tk1d3d8?QyBNe{p1u+N>*a&
zW1m4;dii}Hbp7Geh}BP>bhfihah?>Usw}v49K|ntw8Z(*9!k#I*v>=go_BM6B#bCR
znYrvDn;yICa>|ZXa>qzve+R?;Wo5&Iy6UQ3ullMuFH)V%kfrLvQHi?5op$?4(?kZ|
zweHAWf?=)M?VOgpn^u|H<=y+(`18fDew}9*YLTrJ-LPOLsn*+FNW9_q;)1H7a3|#4
zbM3<xZ>`w%i+kd`8B1{QEP7SPH0*NGk(+(_ofmhV)yUzMhoRe_o=r>5_`RFIES5?z
z5e#29{X6tnk-MN}>(i^56e?=h_yf5*qAjAoGD^1Kn_ti{N10#$QlWCP@c=)a?dBjM
z&R1P&ELK3#)KL6L-9f?zKVKte<QL?l*_5sX3Bfod!t%U*et;m?czreBELQV8<3atY
zM)BnKznD1-Hbie`Yrl**_tM?H^g#V$+~RGl#IB;IA!|pbS)~&*DC6Y5Ji3=LZ?~J<
zv6}an%G%R&x*N(^$8nzfYS)ga8?Vye<hGRU9T=zxv<hb=+-nP|Ff?%<eZP7vd#u_a
zA&n-51+AK|_DoWIsZ908@#U<i7XI0$%z!O&&Z5rV&7vm{exAC6IBw`V`uoAt-}D5N
z@oLAiaO3bAk+xN6=uvpTYoI7hHk29=U(_tIvc49)HvjP5-K7q+H6+fw@haIgt6q7T
zHCFolJJrY&i|u~JzDn0~wVJowm}|Krd{~XgcX*oB{=D)UR^!N{7|k7(%&u$at-s;6
z;%ZOx_TZ;S`F&~SRXpX=*vD5~wsiE-ol7Js_cB#s%I=W1Wd5TR7I|B>Mw#VXM@WmG
z+T$biwze_rz1-;?tQeoFbr>GrF;A}OPTqg!<k`K7iM*8uOQPDGqy=Lx&Vx0$@eiqf
z5;fDe`lol(Nxa^wvW<`Bl`X=mk^GHo3KFZP-%vW5kpf(VdqohGm75lykA08$o&Dd1
zklTMHLZ3uJs3qRgb0%P!nQM9TwbH+i<b@NV2rSLOGdyPkNPn$dIZ$i(X&Dh(fu&9G
zCSEfEq&uF8hh1-I$B2*&OLOoXkC_0{nTo`$6WG51XfQCI>0R7Bar9)rudV6n9abAp
z4gY3HM5#im7*t^>genDy@;qL{r)&4YM3iJp@}3ohQf;Z;pooYfwutxaEGXBK>k714
zo{jYGDN`OcAfjvzTLy}3w+STl+QIbs7%!hB(DPDaD692|%BKqSU1%sRPLB{iA^>z4
zIm(XHBd$Bw7bqa2ylCDkfLQp(66+LN6sE`C$le@y6}VjkX~(mBU>K!vD3L$`jM6wX
z8R!ZBX9O7Kad;7c7xczV)@>s&IVzlD|9r0>UgsZyL<YQ=$pk7H@M8WE5HesOqAmm%
zGT>ePp?&(0N8Hz#-nqgQ(!i#|`s4A~=5hI40jS21kQ)k+GPE(KZX&J_d)Lj#ZFbg-
zSCy4CuRm(B<+&38ZP*05lLjW(verse<Lb5v>qHWgB%lgIp`;Z6ZMEB-^3dZ72cw*8
zSox3#EgTwc#g<KtOlJjLuJ3IV?1hSSrbLj|020iU71X&KyaJiKrBZGz7_2$U|ApM9
z>pDh2fhlx70S%_n^{%pEHeK&359ZNl3Dm_cYR5fWF=5t5*0aJu8nZz6H316-x%hD-
z@}h5bdd*2T`gT-vs*S!KVvgA8+i^{}ivZu`Li9|p^6E3pLqv#(^+iYsA_WM_RHUxc
zoyV-q&$1Z0hiF~+_lg1!5Dty|Gc2iSpyW5h=qTv3w1m|4SS~bCii7d<jOG&-5KR`6
zNfJ|`fMsed38F0e$jp-J<?ZXo_y0J6WA5m~@L=2g->Su8YnrVc9bwupb#QXwy1Kc0
XF5y`@FI#SI5gZa@C<+f08-)D>?3HFU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_nhwgc_2d_instance.cpp.9A3328B90E30CC03.idx b/.cache/clangd/index/device_column_to_image_nhwgc_2d_instance.cpp.9A3328B90E30CC03.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94dad8f64f2bd8220f718e6b7b33e2de244d1c70
GIT binary patch
literal 2166
zcmZux3piV67(Pf!ED}j6CuB5&l!_pUOBy3AgizPxzN)&l8d{Wgb(xlFXPVLUBE9MA
z(u&ezs$EUTN-ukiPPb>&RQ1C2vYKkqvJ=nqi#^ZI^W;Byzwi6s?|r}XpA#e!3RATp
z$Rl2~Tau8u83REO3H@bcrf<#DfS|WH2-<zJJyeqMh^srZcr`+=T=w+%s7I%}bi#ad
zL-t`QC8AuS)AYC@-g|EH`_5J5^a}3ulahW{=}c}|JOqDXt36qyUeF|$-5l$2DEqN?
zqA@2cJNKW2?zE>}{iGdl1gCAK%%6`LJoCCSKBcvLYuwJaHfq^t5*<hK&vv~XNz9pj
zlh@mLtmO7LQ12(UM{{PM%niL?n6&Bm!Thn;PtPV=%+}OTeU})j>w{@dEFDRH+=-jl
z6UW~By_871B5!27NFx5ED30&E*XnboGXGi)$5}haCA_WU)<gGj-%ZzIQVb@lPT6C_
z%TkQW4~*M=XyCZsywhfnlaem%Ar|#S(#^-jg_RQbq-Sd_Z?|b>Zu~y>`ef|vXjhN5
zcY)-{A>HH_GL!O0@SkOOlkS}MylA>zEYoMi9QbV}QI;CuOFD=|Mi$V@KN}Z~1d%G;
zH_f*>)|#yAPTuB6SsQ14uIJJ@2JcP{x5vIP{4Mj>^*v$kM;oxC*Z0=t8dWy$jji8F
z*;X(h)p%DdB&HMy8D<BM(6!eO4IFZ<94vGzD(*~HnBE`MO7VVupjPjkbYtnuon94z
zq9C`w@iTr_8prSLu}UuH2rhIvV;F^(dC;!DHVa>2V8EB1H>jkjOj+$d`Jl(y>O+KL
zs*#m+Am1=XF#Qd(tE7A+Y}b5XaMYFbys=obzUt@KE@emS7JYf$M!U3cRp4LL{#DxS
z(&im?igyii+f>gO$D<ZC8jg|mvD9B%t7vcbRLjfsre9smDi_{5aq);NmJ<13YUtsF
zZoiBnx2i~{nm2#oPy0#|1_a3M)fL`R@A^KKsp;fhjuu}TfpdcubB>4Wt77&)`1~qu
zn*sl^pds01om0xT)s{){JG?x+%)nlB>K_}j0`DJg#Nw-;^A++Q`EairpVS!_RWJAj
zjpX${wBx<t54GE<nfvlymCXyY_ueY~C@?@R;*Xa2A6vg0>(2<TZ9O&jRr}ED;Ljyh
zPrX>#_iJSDqhB1+Y}#rs9P948OUa`|<9owXjTK_E1a7;&HNW<SR$9aZqi^dv?oQrF
zYAJr-7%~;-G&ot*QJuc8`{pc<^YeWN>{W_XI^^(iwng{YaQkZ~>+|JXx)(;zz=sw5
z>6wrXcOP+h=g+@bOkxjzeU0;~vF{G;-e*5JUsz16T6{g%R~MhJ{;=_6e#Rb2Tn_p+
zQWW1#=LlRD@K9hdgqUH?xJwI~S-MJaZl<-&{$KD=0Gc+$8giEwke2$Ho78O`PQ^nT
z(X<X$$8l)^>88muoUB2gMLZOQrkPkKXK4ZH^n}EWTJ%dmYCK5U)$muBIv#;NVXx($
z2rPod0FlSzp7Uw5<rq9dFeNxIcSc}S*clYz5xObedD#=fGGy5S5r)md&V%x_3N<{!
zU^0S05xA>qFBvcGoUg1O<(XPwTV=*6)DX1VQqxi&NOmuN4$V$y4q`9}nMn=?DliC@
zNeu!jF#l<QLD)=oAW#wFPqrK{+kjQtz!8Xm0CkRoiGTpLfuj?FGCX37vn2w8n7LmT
zk8z{-p~6-wNbqxSFSuM2ht80-fh7RdDDq_a0F>t28g^+ux&Ynlpy{w|YRA;svs9;n
zi)eVI)cvs(p~B7~3#BA@61-JX2|fgWfC|?hN*UtbV4{0hjsQ48r>Z<OV5(pj8de?_
z5SH7uT*%sQcS#HF#3u+q73wgI9IZky9~J;&v>if?QXx8@E>I1OfiR*}h{a*KfDxb;
zL)no4op{Ej`a>^oqv}bTq~(EtG%8bhXo1X*%fr1MWl)jw*#!_{RtWQlS4#1xuWT-6
zMF76e5Z8%KCcpQf1Y``ErXrwXs5F%q28KaXsb^!@v}J;^(wk$}(n3_Pv8M5|-X|>V
zv?kG@L@7+4V4~uM^<fhgFI}H*qT*%gunYk22VJCTIPIsW=%c24(!EqgbP-()5R{&<
zeY5gK$VlIkVW#{Os$7;X4df*`cVUZ*n~i#C2owUs5VS=Szdad+zKMPs80;!dE!;<V
z9bF<pMEcajf}|}HkPTMr8NyD^ybXNUPu!UL)-GmThROd$*BBX7P1n-s=9cTM*4wh|
a95{|2vne+Ap89LNe8SX){_fst;(q~1YE|F>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_column_to_image_nwgc_1d_instance.cpp.163F7E8C2077B0E0.idx b/.cache/clangd/index/device_column_to_image_nwgc_1d_instance.cpp.163F7E8C2077B0E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a90d1c9068025c9492535a13286ac31ddbc3270e
GIT binary patch
literal 2166
zcmZux4Omk59zH4zKL~^*Js_coidvvxiXYe<ilRZLcvuWGEvGX{6LqjGE$dp{ti&v{
z=G93#)9gd$w3cm|*UF7%I=$0eb<N_I>6#yHIUjfQd5*Z}xjYYl&hO*A@B91s6Dtx5
zcUl5~V7o}IOvy<?0RTXNzp9+<Pxe><z^_&SpgwmqUa2Dc+qTb_Z8qFt<yJf`5?3dY
zhc;9mU7wz#9hVeIt~W8f=cW^OVgAFb>X!_%i)S?Yi5ql_yMSUdxyv=nZd&I_3pw`&
z8zL!4Z60cP^?;klryY|NcDw(UugOko4=yM$KgyC6*14_vI&o#z?5@1I*ZzU|v)lfC
zDslYzSbAFp)@timZ}PoU|GM06m}q}=rGfH6dqjSvBq+IbplOrlb=$KomnZv`cATAU
zKY#XG^|Z2Iv&EqkvcAO85WXsHYq=d(|4a3Jp19WLSu0=O8uZ{v6-AJEeV9ox>Kgl8
z9kwt>$_^R5=bA`61(%fN2Z{^(zL|ZP)KR8w2`rRd&^71AoX{zO>vfF-hP*54x8ohI
zx^MHM+%5G{*%ZCeBlX^>OG?BS+9&>jhsZiljqgpcUM}5_`QkkBA1-*^i`@^yJ<at-
zvOVs3pGBubPHkid<|v&vwfo4m+6?j@&mz2dg7ZVRde!z6m#kM~ExURIsc-K(C`rDK
z^|Xjz*#~j`j_Xd#ZM`C&ec0J|ETmevd*i|4ST7gnpX91q(VGKxc11Qr&~)qh@4sYt
zIH^%L{L0O8Qm7<-hnzL~W{iJQah6|r-%C;Ey4Hg!PI^*hNmx7TQByM$lU85bk<|5W
z(0Q6Bw6}HC#Shm!d%JnKeV{K*)~{TX!5ci8U_Vne`HeK@&NZoJ&t%l=!$yxEeGQN9
zH~n`uJ*15w?Il-M;2tsC_vc0qX^-^1C{Gt^Wd*_%bjgX{Ag?{vzWUaYPrvJRu5LQ<
zbuNdn*0p(-KNCH^U8?B2xAIKnN9xDnC(TyuZBuOc^f9<6uHl`3S)(qY=r->vJ)Oq4
zQ#7Updhlu~RJT;H6+0MHX3r6Qt|A;RC%bE=Dpv3735WNfAz$soqx@JFHFs9mK3Dmj
zV5a45#(ay~PBW+}{{E|rLxKyJ@aL>6WJ-pOo8KvPS5DdBOuuaGr>gAwemOJGoO`1l
zMY&%z{=Qy$s#ia1470d}HDFRTHOeiophFp-b;!~s-A@xaYVWY{f|A~S-Dx&IuTQfZ
z&B}?HoEMHvTfdk{tgB|zhV2{P{I2&I*_E0U{L5p?@n-E2o?(*qC#AddC2%tse_X3J
z;L>0IJ`alhKhOK&otM7xkiL{{Fg`dsbt}P9t{uw^lb@*uR2pS+K71Q@G{~=bx_`0;
z3&h3&5DU%nT|D5(;-}oF<l^2g&=m`8g2OIomvxH=5<V@O*xMt_9KZrGaM%Iuux{}{
z!rjV<9jLqVR4lL&4zED3@LN2PaCXYasyg_efN=RRIk)L$hdCAk1)y*V6api`Xv88K
zlgIe<KeQ+;gmc0LEI~t{6BvLbVIhVSBVb81#CBo_AQd_^#{~>lW}Y#_LQEDj7D;kL
zx14m`L6yR8cZ@p;(H2@f2K+IVGl)V#L>4g)p+iAb7Bv>3L;Vd21#wy2O$c51{GCTs
zYV1QxST7$JiU43CqKx>F$pZOcC<2Fts1Q|dvKV|uC?XRU4Kd{=i_K?;B2-w28_(U0
zSW*em-B=ooUTBmR4!?}b3dAo1w_@Oz!CA5K2tdNS2lJ=>6#L<hEH6Dp9C~x#)eFh+
zl-L5c2+4-6X!cUoh}X<|ag9oD&sH}+4RXqiZjoJ?RqTTyIT+4Tle2!SsiE@vF2U~Q
ztt;9}7B={B^I#^N02g7Z04KtU5$k~Su^mHx-?HF$myZM~2^O%ht%v}mZrHuBrHK3N
z3qPs^7{3RD5HN%#UAL|d8w91m!G#uX#3>L;cs<47^7)7v%O?`Y-inleTN?7eMA|?X
z3_wH?=_UXwib_Ycz#~O5>4;#s1t>0k31Fn8bHwZT0a%zL#&Jm)0tA<_aGQ`aM!nHf
z7`**I7xa>FG$aewO9E*oItGbBGtse0Y#Ks0XA4~$&b;#+?l?oh5So%0k&I}>0<u#w
zk{0fVDtnvC1^%^w5_>0B_uvYX&Eus&JaTmaKx%s2rre&92Ak0zi!D&-_b`@LHrN%m
zc$|n3E^s4Qi#{OQue5Ujc>zJe{E+qUvq)ZnE`Chf|MRY*I8vRQ=?qtQ56?AhZy)YD
X|JA<awVW{0s)#6=xlj@sX|~}n7eGMd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp.79A78F49E52676FD.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp.79A78F49E52676FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b75711cca43cc1a5911045e402e6b70872c0fcdb
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7&Leo7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`%DTO5*+Pr6C-$?r&eJ(0vU1{-xo@t&*N7}!RKYq+=O}}-(tE}C^QxyB
zH}D(~Z7p<a{IJO;TC`lg&xK7sz5c_`uRo7Ix7;7m<Impo-Tzzn(+3BVDs3b={QTLP
zmTT#)->-J?>+zpQx8KkF`1ZVIeBI8-51*ety*YzrPW`T!PfwqJUf(@iKm5<1E|vBM
z|ILx?F`F)ho;am2g`?_l)^^wPt)kEDUL07mI!oLoZ^HVE(Iz_iTLOKeGo@Pfcl`e}
z+wFPWkKcRSW<Tc1Yc>3`<g(~35A7{p;t4K&YD?CBKezll+uvH9t+7kb#(tTy=k}qW
zUQ4`hz2ac>b<BRPuzc<|!51>O7R-Hn<kQ{V{O=V7v-NpZ0u~FIW$kR9qdK`?&-&J)
zrVUpEr!CkZ(0g-bU_|v4C80F|QV(61R420UYHMQHT0XU#|J8>+3%6LsyBrtIjjZIW
zI`$~%YuVm$?znm`W!2LkJx@-VDXiK2eQGvSN!9m`i((6f=6%{I{;&7mLAMIsDbrfK
zZgK6;3bHC|(UN6imE-DlbL`qG=63lgqo1Ce`1|ZTCv_{`6g^uWOuTVblV?HrwLk~!
zNZl#7);l*iWnAXAVf!L|X^H5>OD>1MJLK$9wHFPUBDZLok(FF|hQ{3Ge-6LobYWi^
z8m)P=R&2YoXYQL@yVka@{a{vk{o=cg3S2FFG`YT9VA*&`E`v$X_o>dBS4#KXKcDRE
z2+duQWc5(&sFayusW;c9bL&NnZ?XPa?Wn!7WlrAKH!%wqebv1;X}Rl{s;928YNEZW
z>9@Yy#jm<sEwagHnuqJFL$~_^*+PCC{Zw?#?)ev=mAB6NT$)tHTlDzJtOfIIpE#{s
zY1KLB;?%x$cgY<*uOcRi26@gb?o|)TtUosE!DRsp+r%r%Wk-*xFJ3Iz8*=J?<f+^F
zl6O+|XQ%(o;f&jK<@34X@Z{_5F}DR+StqQ0)n;<TWe01~Tp^V|{L-~^EteH6R{i8G
zv0m+t-;b(8@25PR)IKSzhv)gxA0<CL_RQ+oW3h}od-rymNu_fQwIZI)UGw1j#g8vK
zp5EFNbUi3-v-ZYgtY5=_=uKwNow8@XQ{@vQDds<GeN&5U=kjefn_HH){PMpWCA0Qg
zPHviHkou<eeuCQdFVa(|yt%F|#kOSaJw1l@N%0rwTrtfI%FmSNU3T*3{-0s&cjPO3
zpB|h)@yG)4ztzt6tIj=t;h%0Sq2IpgOMKC9`}4c*Z7q8;HAhvo_S0NWiA0-)iZ7CW
zf2`zquRodlm*F9m+L{xGd(@J@@fSE|ohz=)O|k>#Z!!I~vvxWK%kVJx2QY*yvnpqT
z37`xZq~)@({xgm4<Yo#4N{h3Jr-2EWbWv(r@q0!F2CiQV?lpUx2D9*l%PY(4!F2J2
zD<~@<@wBbA-QX66>saf+c`#8%F0l2zhkh?u^UzwsvHcCh&o3pn7+J!NIE*r2s=T&W
zeU4|1`OC-?uA-n~1rrYO6H<RM(S{Wy%&Ec;6Nb5li3{v@nAtoiMrK)8)Gj#Ua+QfE
zTt-<&7p8|NT+>?96=qVd_Ri_i5+Qy-MT*Lb+6YCu*1FCxMa>CPZmSKA%h-9sC6p!9
z5sD<0C6Rd2%F;+YS!G!yp1QTVJ<K?sa1Cn>B%YSF7Myn|#O`Ih{Pp8N4;!!;D8aNm
zHDLY1&M)2#3I|aYd6@8<16ngV*2YQ!g(X-dlwc0~D-gEnZvK&H%sk;z%2H}DVV-a~
zWjQ3Cp0%DG%p``Q)SN_68YnJGE)E7UzyyfFzz_zZKx7bt83JcAhf60dz5ZrljUQV?
zP;f|CC_^Y*jv)fUOwLWo$pCV_(|nm(SlQS)IJtRv1^D?q1znxh1i3t&gv5nKBqS{?
zt!-@W>>cz(wXKxRG~}^~ib>1JDkv(csH$sf>F6668k?HSnMmmxxrmCndqpz2`8j$p
G`Tzh{8pAmN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp.DB04A314807BEF1C.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp.DB04A314807BEF1C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac82ae08bb5af14c7c8919bc188a486b139a0abf
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7&Leo7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`%DTO5*+Pr6C-$?r&eJ(0vU1{-xo@t&*N7}!RKYq+=O}}-(tE}C^QxyB
zH}D(~Z7p<a{IJO;TC`lg&xK7sz5c_`uRo7Ix7;7m<Impo-Tzzn(+3BVDs3b={QTLP
zmTT#)->-J?>+zpQx8KkF`1ZVIeBI8-51*ety*YzrPW`T!PfwqJUf(@iKm5<1E|vBM
z|ILx?F`F)ho;am2g`?_l)^^wPt)kEDUL07mI!oLoZ^HVE(Iz_iTLOKeGo@Pfcl`e}
z+wFPWkKcRSW<Tc1Yc>3`<g(~35A7{p;t4K&YD?CBKezll+uvH9t+7kb#(tTy=k}qW
zUQ4`hz2ac>b<BRPuzc<|!51>O7R-Hn<kQ{V{O=V7v-NpZ0u~FIW$kR9qdK`?&-&J)
zrVUpEr!CkZ(0g-bU_|v4C80F|QV(61R420UYHMQHT0XU#|J8>+3%6LsyBrtIjjZIW
zI`$~%YuVm$?znm`W!2LkJx@-VDXiK2eQGvSN!9m`i((6f=6%{I{;&7mLAMIsDbrfK
zZgK6;3bHC|(UN6imE-DlbL`qG=63lgqo1Ce`1|ZTCv_{`6g^uWOuTVblV?HrwLk~!
zNZl#7);l*iWnAXAVf!L|X^H5>OD>1MJLK$9wHFPUBDZLok(FF|hQ{3Ge-6LobYWi^
z8m)P=R&2YoXYQL@yVka@{a{vk{o=cg3S2FFG`YT9VA*&`E`v$X_o>dBS4#KXKcDRE
z2+duQWc5(&sFayusW;c9bL&NnZ?XPa?Wn!7WlrAKH!%wqebv1;X}Rl{s;928YNEZW
z>9@Yy#jm<sEwagHnuqJFL$~_^*+PCC{Zw?#?)ev=mAB6NT$)tHTlDzJtOfIIpE#{s
zY1KLB;?%x$cgY<*uOcRi26@gb?o|)TtUosE!DRsp+r%r%Wk-*xFJ3Iz8*=J?<f+^F
zl6O+|XQ%(o;f&jK<@34X@Z{_5F}DR+StqQ0)n;<TWe01~Tp^V|{L-~^EteH6R{i8G
zv0m+t-;b(8@25PR)IKSzhv)gxA0<CL_RQ+oW3i0;+B@C;l-Tqhu4$F(+CNH5Y^;)Y
zZd-F@>y(>syw(&t{uSMK`iY~o>igvZayxsDG1O-V&y4Y3et3=V^0_x(Zm~C;o0flW
z$pRH0$?AvZGftZRbyHRS{k5x&F?02H?gI;-gqx_B`p&u%eapQ$v*=s?zN?Ji-Q$ug
zKcuS|zHt42??V5o&z8R~-{d)VdZEtW>39B}-@LATt!$-gw5R9%nsg>N3I7a^T{r4#
z<PQBl{iLa`$G}rw{*%GtlQaG@Fa4i&x41Gl$qtymg=(KN1kaFa<7NsBU<emy6HfyZ
zKp8Md{;~L~s&D=Q9tMA)v@)x5CYXRp7p0~Zzh`7%;QF=TUbDApFbhw(yt2F=%ut?i
z1!V;!p0>5N8{EWj9cvvp52k~O3v7Mwq2CMEJhWEe2{+;}%7AHcY=6V>^GnGsMivxd
zukBTz<5^?=GV+A0D5zM$REPKpslS+L!wM4SRN;pS!)Qh>uuot%Wm#9$E;!<Hm5C=@
zMp;G|<{qAKO>0e8n4VnioztTwLi~V=6qOaV5sGxJb)8{~niHhlRvQ|ZvGasWC`+g#
z6iF&eBJre^rIC2D%Cbm2b!&Bdm~lMe8rB*}JS}T2IB$E;j&mg(Q`3Q7lwgrif?0Ja
z#O`Ih{Pp8NVFMNeC74N14OqXh^NV+b0z_0r9wz)(AZ*j!{3FkpdBUZXrPN@C@Px}L
z%OUafto7_*S{RB_a}q&mptvZxI2gnL6CefyLl}etkwFM%2!a_3XEKKurhJ<d?76>>
zEg~p5BrKF60xrjpoSTx90pxn8`7*Pxvaxe;a`W&C@bh^Jx;m=~a(Oxli3^KJNLpB0
z+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g(Kj$OHZ_+sk<vAC5fyXyiez&0bM#>J
F0RULK+!z1=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp.E624632F1E715961.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp.E624632F1E715961.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c2aa82d390b3b8f61d417be73f909e3e3cc13b1a
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQNE7&Leo7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!mO6DG^Y@x;36Z=_Q=jj|0Svhga+&9<XYeW_<s$iX^bCkhZ>Am9ndDT;m
z8+Z<gwiY@ye%Lf;ZP#vocctcj@%j&^UvK_=e)^vpJ3Cp9d-gZaPj6=a{Pbgrf`zRl
z$GlY$G4;MpuluJzzhAff=yrSkpPx>joikroe6t~^#s4i4pPoM7y#DiS{qR3;x>VX5
z{Kds(HzX`wJz=V&hTxaBt9jk?xV6vJUTE|TzpB@9M<IUk_Kb*o8H;CZyJW%@U-18_
zdGEQsAHG-fnje?E!<GEPbGi1GiD4O2^$v8%_<BUYpX>jg_t)=;tleI-cfZiA+}`?m
zs>k##R|T17HC?;r=x3d)e8F;yqxJ31r@PDL?>Q@7i<9zL;HH{$rHI|qTTMQ;Y>O*v
z;_AgZj){s=n}Zi`_@?2i60y+aP`AhTLwqH?tc+RjwZ!GGJd`c$+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$*H*x5$W%>ud}@P@=kEEu9K?G(`5ZW(tDeGKSXKha!%bM
zT6c9x$s3MPD;6#r5vkrLk!;<bW!=nkVtVxMUEeV|`cse7B#wj18&?NQIILf@xS?!Q
zl*ZQhc81mq%fu^qUYIZO)K*&B+4jEiMum5s)(Q<<7v0nnn|Buito(nrUl#7*3tGJ`
zXwz?<y!J`AZfq@$=8k@t`)S?cyUC6s9F@T$FBWnpwc1`_QJQrsGUBT1p8jW(g#}mN
z3OHJFNVm&0C+XERktK8EHPg0m{|Ijm3*xlAn{{)igX`<)J*xiQFTS4Y-sPhu<$Z3;
z``UdWyT5Cs%+#6Kb)|K?>|&l3AG)7DT2vc%$#`j&xv{7BE7KFble3$uzD?lXx$Kon
zkz2Ucxjv%;$txRGw3kdWdM@R^;!@qQSr0A?SlA|BQ7$`rOnvcU!QPNl_jR7W-OlrF
z)AX~O>$b7zN4~0jzH{})*Usy5JDQkIgq1o=<tW`@+UdtB^sl{l-#MA!9T$auD)xj6
z=c(7&J}f^~aMIb+)UEAvQjJxON}R7toQ!|P)@{|Nl_Rf9E^bP_cap14R<^${Vx6vb
zscvdnq3Sh@<^Q<ur+!+{JL!9{Mt_W3a>M?uD}B~2znpk&+2wO@N^Z%|_S<Zp9uy$7
ztY_~B=`&Ae)h(Vh>2H<m;f5_&Z?irK_@q5+@-8*sSGsQ(A2hN2ZGQh%!*AvF8*P3V
zPm=hus6PLN`_-Q^e_!UXJ$@Q6?Voqv|L2j{cVFw<GimLTCI0)28I^m~O;}=c?Ctm;
z{!K0Nvv-@Zq@TZX#>JE~{}`vpXWd{ZuFOrc1Lp9wuE$mDR~**kVek)N2v=rR&IA)c
z88E2QvE-FAJSoe~6bO_SXA@5Y6ENwb)U@LFj0_B1zZTqU_BIV>;R%;lme+&n;t5w!
zRzTuuTWh<)EezMO)`9aJ+utz!{8Dm@ktN)S!zcr$$ZLDm=Xlnbzl=QLDheuAFyRnC
zA@vs%ZCF9VoGSb<VHnNG1@=Slq2CMEJhWEGvaYCIaKz;*6HmB|vWzamahleet}s2h
z+B>I5ON96V6)7q!Y9kcsTI)K)6v0H9xWJ(Rvw#Q1&CLl?ZmSKA%h-9sC6p!9VS0GN
zC6y(Sc+$$!NIY3(StOpiwYojbPM&ZLYYilxmbDg~w_EnH$gPQXoxpIAV3AOQS#>DH
z?q$6E_2WQc0~P}%m`P6!Sii9Ii+6)UO;kl5Cj3_*Y}4KRBhQ$5!ljg@)L@42gv%+*
zA@TIA_3U6;7>ZJJ5<zL8xG1?e7{mY*AO-_N7=!|mK?r6DoXH%{xWm!ld)!oJwuqqM
zkg!mOP`DgJ1cI5Io05|O<a($1GPAI<v2$>8^Y9Ar^LYxoI;#nCc{&M+3yVldT3A}!
z*xK1U=!t4uDVu4?V-ppVmXTFZR8mn@*VNL{H!w6dHJ3Ay(lv4s6?6BBWODO!^kDP>
E0ARSq#{d8T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp.CACB611E1CE9F1B5.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp.CACB611E1CE9F1B5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f266e053ca5584ed9fa9fc69ee52b7b5bf7dfd4f
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQNE7&Leo7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!mO6DG^Y@x;36Z=_Q=jj|0Svhga+&9<XYeW_<s$iX^bCkhZ>Am9ndDT;m
z8+Z<gwiY@ye%Lf;ZP#vocctcj@%j&^UvK_=e)^vpJ3Cp9d-gZaPj6=a{Pbgrf`zRl
z$GlY$G4;MpuluJzzhAff=yrSkpPx>joikroe6t~^#s4i4pPoM7y#DiS{qR3;x>VX5
z{Kds(HzX`wJz=V&hTxaBt9jk?xV6vJUTE|TzpB@9M<IUk_Kb*o8H;CZyJW%@U-18_
zdGEQsAHG-fnje?E!<GEPbGi1GiD4O2^$v8%_<BUYpX>jg_t)=;tleI-cfZiA+}`?m
zs>k##R|T17HC?;r=x3d)e8F;yqxJ31r@PDL?>Q@7i<9zL;HH{$rHI|qTTMQ;Y>O*v
z;_AgZj){s=n}Zi`_@?2i60y+aP`AhTLwqH?tc+RjwZ!GGJd`c$+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$*H*x5$W%>ud}@P@=kEEu9K?G(`5ZW(tDeGKSXKha!%bM
zT6c9x$s3MPD;6#r5vkrLk!;<bW!=nkVtVxMUEeV|`cse7B#wj18&?NQIILf@xS?!Q
zl*ZQhc81mq%fu^qUYIZO)K*&B+4jEiMum5s)(Q<<7v0nnn|Buito(nrUl#7*3tGJ`
zXwz?<y!J`AZfq@$=8k@t`)S?cyUC6s9F@T$FBWnpwc1`_QJQrsGUBT1p8jW(g#}mN
z3OHJFNVm&0C+XERktK8EHPg0m{|Ijm3*xlAn{{)igX`<)J*xiQFTS4Y-sPhu<$Z3;
z``UdWyT5Cs%+#6Kb)|K?>|&l3AG)7DT2vc%$#`j&xv{7BE7KFble3$uzD?lXx$Kon
zkz2Ucxjv%;$txRGw3kdWdM@R^;!@qQSr0A?SlA|BQ7$`rOnvcU!QPNl_jR7W-OlrF
z)AX~O>$b7zN4~0jzH{})*Usy5JDQkIgq1o=<tW`@+UdtB^sl{l-#MA!9T$auD)xj6
z=c(7&J}f^~aMIb+)UEAvQjJxON}R7toQ!`(>f3vtJg$`<Tahrcd{a}r`|-uf)3Q%{
z#d^<-jp;N!w{riX@|iXni=X@zIi)=PqQryzT&>BkUzW_+`f`i0_1tgAmtB5ycBaUb
z7ENGevKjAOc7Mf^l>KoR4HR;}mN8mP`P21m$vV%;Ro=fd41CZ1JG<TL!M_vPI{g27
zRM_^uxG()Hu++}I-fH&23d<==|4+L8zw+AGbz7C!r)X=woc!MB0LNoaUPtlU{pX!4
z>SxZm+%Ks1^0BjAnu+A*dWMsCvz{{)SLP<!0du(5VzqFg6^i{l4E_NO;mWManP37a
z0|xt>_ShNkVhG`83Is}vvx%pH37B+IYFhDoMg|71UkmOvdz%Kc@Px}N%j>~(@q{ZV
zD<JW-t+n0Y7KZCs>%e)A?Qa-<ekr-d$P#YEVUz(=<nEEQtcBx-3?omt1dD_c%v_i>
zBNy1Qy@!4;So6?Y!E1Zf=Xlnbzd+S03My7GRUv*t>MthRu!4j+Rrq1TS=JS`3y!#4
zW#S2!QI^q#navZfX|3rBQ<ST{b9%Hyh#ydqqOzhkLXobut}{$gbApuHYD42PcAzVj
zCDaj$B$XwRc+$$!NIY3(StOpiwYojbIG%6~YYilxmbDg~2Q!<A3mgwHpYx!E+My7;
zm+|t~j{`kyz+#{TGxw<h>lb!@@orE!h^olLg#QYJZMvI(<QX$hxRkP#8q5Zsa5-f;
zB%YqNo*hgJLs4o@A}9?M7bO=5gBTGY0s_E%2r~>O#=sDSV1^)=p>QU1_@RR#3;Z^G
zNo9)&3JwViWk}9V$;kl9d#Cv_v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI
z+SxnkiE3LZn`y{n6BUz|kyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN7m5a`SWaVDte1
D*}}c3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp.9E3DEE644D650E5A.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp.9E3DEE644D650E5A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c09d48887d1f4ffa36e2fed0b47d3915987aa186
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{qg1G{YO@Is@N|(ti5}?)uht7hFW3I=B{~g{o=<L9Zzp<
z3c4PYwmEy_A<M5}Kf=ZMZ)p_f`;|X2%Buf2EzM}>EWh@dv;3?lU(T%0u{wKiimPMe
ze4}rN^7STb|LXPf`tVwmxA|gN_VYgrTf$AmOMPcuiN59DoEh{jf8RyM@9uHQl^@Q#
zCtc|Nx3f_`^l9-I`*dRo_4Z9)jCcNxPmg{dX8p)(+muOm@6WOH7|gqHp?Krp3pOp^
zrk`Z}E0H+K&g4m~+hl{^(&DPgY7E7dxk+}w{C)gbQppdOh<V&hfdLHR;%wqcU;-!u
z24|Q5<W4bodz**BA1JNNs+<ZYVA4gYX~pju85p>JEx6b0Z5qtN6E3eTuLm=fCtN{U
z0g0z=t?dRkF<i%52hI~*-zinP?$u#-o^T0e2{nWPlFE`uJZWWVB%Z9YED}%MTHOv|
zxQ4X`5>Lxo3(j+Ff5Y(eOUW%pmT)5uqZF7HukBTz<5^?=GV+A0D5zM$ghTv<)L%@r
zVFd|us_?^vbG3I)kCq7WW8w){R94hRxL4O&*BPd0X^ZTsO+V)D1uBwJmeD~d(zMoe
zfhjr^V)rs${`zsC1qLhzN-$xVVkRzdeDogry<p8lYXu$@H^US&a)Dz7rs$~w>lb!@
z@otctMOEZss@@#Xn#r*?RthLA!6Kmqv*E8m*rvPrN1ieBgi9$)AqA+MvK$gm&sq;2
zgbYQgIf<Y&P+XK;91LQB2@r#UIs8$XVZ-SO-)6H#1O<nLg))SJ#lb`nf*FEfh9Z~|
z2xf9_N=^ol>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3w
vY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)J$<5EvgV6^7fJW~Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp.384AC715F1A68557.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp.384AC715F1A68557.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f6f9f8744f09e442171a2c072c694f0f81ac1d40
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{qg1G{YO@Is@N|(eC^%ZzLePXp025t>e@d_OKhx?c5Yj9
zW$ToiZ=%)|`0W+_H}z<<nb)1!mt}YM+?v0C=}ez>%PuD_TXy-}lagEVv;8)krv?T*
zSl+Yu!|c;fy6P5B68cl+dN^Ur)mxVJ8V^O!y2egc`#SC0#e`W?{?5)XIq-Ml^$58?
zpC?LuSyZ3*z&-S*%-{1jd5)c2sPoq+@Bj6i*LGj++c8OZ>5}uijTx1D)J?v`<k*|=
zFZ`QY<Y@0UV+lWR<@AdxXa2GIWG9<46j$aZ*#Yyno=rR3#o~V&+)RN14B_Hz;z?iv
zC<6ww;^%)<eO_<H!{85;R%TUB1rspoqSUnF_lyh-T)!6FYxXt`X5k5!SC-d<8Ojr`
zpsawz)3(-jgPRzxW32<{39j#yDqZ*LFgs7UgtCMh!T?ESNhF@MvNRG;R#_H_r*5ro
zhcH~jS_6rvWvvD0!OUjj0{gW0(C-Cn9$G8#gd1@frN9huY=6V>^GnGsMivxdukBTz
z<5^?=GV+A0D5zM$REPKpslS+L!wM4SRN;pS=W6eq9xV~#$HWt^sH~_BvzRAb*IL&Z
zrU*tea)Cn@=A5N1vZprvn7bFKS4LSz2ccKfTGIulXM4|%b0r*8(}DI%ut+Gu6dek&
zdl@f({WwtAfW<%wruwM?>lb!@@otbSMOEZs!hZ$AHr>rX@{E}$TuNCADM00v<&b!K
z)_U+DWGG6_Nd%>V;-ci@U=RaLfEWzS;gLV{4xJBsXu}o}6dV#3$`A$?2NOXEW(a~A
zieN?{n8~>*IT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQuh=intrL~Q%oxOve
ysJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdtk$H$O)YMjrrzjoqLC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp.66D5EE934903111A.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp.66D5EE934903111A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d83bd12f790f31b8ae686478970c869c73997e80
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{qg1G{YO@Is@Tu_kd^v9%ze{3W4EKH_DmN2V`U{LyD=s@
z>{ZmMoHE@oi{<yY{-1i}D6RVL?909wx3}l-U-D54zx>i5_;O}?k=5IN-{m)FZ_?2C
z5PU4|$Jx{(*ZqN>oOOE_B^unydOOE{%7?CJi`IEguJZn!Vc>h}-&u332melF>+t{k
z>A_QV(cb*S;#D<$|DNZtJ$@1}txj$G|JOOM?`FBjoSYVP>G|E64T8snrK<F{$<J!P
z@Xu)H0(oVhOYLoYPG<<E)iYadPBtj6%uTWb=5O8yVc+$3n{DT23JhQf7iSYs0uw+P
zFxb}bzx4TTlV~0Wf1tE7t8yxsfJqmnrWL<uWMJU>wcuW}w`nj7Pq@6YydKO@o^S<a
z1tgxfwYD4F#Bd#J9XJoBgNX}leea>)3)VceR^SOY;xI~qX%SrCDOI}e)nRs?a0z7z
zHJBcra7krJB%ZXgG!jo%Sr&<>Zmn(yGma-*!&(E0r)8}L=Q*~&Vfgu_<Q5|fiVM89
zSAC9Wjrq&S6Rx75Vg)lP#7{{5#Y7uckT9nTKTJ4Rd*}3Mi4Z>~o^VBFMQwz8b**)s
zVTzWv$e!BtWA0v{A{k{F9fTrHYfTrJA{foc1rB~#@a&d-EOKk2T_@092^I+@nBGGn
zb}!@QuO9~r8?YEC!8AWLVEw|*FWwDurKpNLO!%)r*rvPrN1ieBgi9$)A%&HkvK$gm
z&sq;2It)drIf<Y&P+XK;91LQB2@r#UIlP{AYt-LYyvNxhf`UWBLK(uq;$R{O!3;q#
zLlMjf1T#4|B_{*O^-l9;W?^Mx=iub#;T7QL^AvP-Rukm%bP^I57Lky&u(Y<ZwX=87
z6V<j-Hq(&DCMqT^BdegOq@t>>simWDU}$V=E@vX8Yvdv-=I#~A<mTt-!RP}3aQx#5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp.8194F48DCC1F513B.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp.8194F48DCC1F513B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c83594015d663dae4e086940cd84805fff0e894a
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{qg1G{YO@Is@Tu_aO&3Hs}t9}?g`vz^ge}ke_x-wyUz7#
zQ+G`>njPCMdT!<RL;E*XTyX2GF5aa)J@EJE?Ix2?UM;aYbEV|g%sGC)lP_m}Tc*Rs
z^=HaP@w(!fF&Fbgf|}yjXGrvHi?V*suexW^&K1Cjxp&D<%4hPX`p><5JM2G6oo=qT
z6Kan8a-R3k3a$Bz|69&JXkj^J>HL##|6ARz+I~wgeACh?Un<Mh5?VgCw9V^&+x|>(
z%YGld7soj!e_5#b-e^|K&HMR3lkFLbD|3_VfccwE{Mj0nlNaamF!%>Bge$Wur-BKf
z3>d_GzP@6j@Ri%#Oo2dYaW?TJFaeV;N=++%&xkOEkqfN5_t5VJYaUuFaQ#|vui4u)
zn1v@?URhobW*ASng0cb<Pup7C4epL`9cvvpPjG#wROz}`huL|;C6p!95C%vpOCs^4
zm8Fq*vdXeZJaubzJA~mH)*47WEo&_}&$0at!_O}zw-{N%jW~=_U|QTgl9sh_+>l`e
zx`0JO3FZQ??Ny)SS!4bJg;f+(tYE4`{DjnBOtfJI33ICO!-QeFn7F{92MZn^6q|Cj
zcTSI%2=Qa$30G8B)Q0Kd3D>pOb%vR=v_<yRrXO?n0u{+9%jh5!X<BQ#z!V({v3nUW
zfBiVn6$UH@N-*K42CQG$`Ng|IjuTaphY9}`2-|cw|Hw0Do^UB;DWs5-Q<g*G=~?T+
z!-}COH760228xT4i-SQ7FacsPFoZ!UAjus5Si41e$AwB$wuqqMkg!mOAh;Yu2!a`k
zU`8OA$+;;x89=UgnlCd8D;qlpCpQnT06(9npsTZ*AeX0;khrjjgrtR~wT-Qvy@Q^p
xww1D(hCDV=F=-iD1w|zlRdr1*9eo2sV^ecE6DeII7f~^HuSh00KSvKn9{^fW?U4Wg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp.ED87AF8FD1B46167.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp.ED87AF8FD1B46167.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3e8f17b950f0c45138ee0e1a9d2efbc76b7b979e
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nU|Ou7$kWZ7%JwR4a@a^Y$0%N{!ahZ
zD>EPNxa(EJFzw!==mocKi^}UBpJx&#W1ViT`)2+A!b<@hdnb!cyX>j@@l5i0o6nCp
zB^}rvbRDu-=ui<kZ%x;3es-ni{^#`{x?f-3d|v!dj-job<~{p~;-@dR9sN{b79=Cz
zE8_1P6Ibus^}1g?{eIo@qucHEe|$Q9wrReezW=f%5B^+!I!WJteqHRHnw^i9hcCIw
zUSnndzR^2umsF%=i;&~`$j4?cc5XOyhh26?N$KOJ6TF|e_PQSa6efLi%hY9|Kbk)4
z+h>-%Yu$f8a);%a2VoM<{<mTm_ik}oUBc-0B*RH)Z@Jyx>o49XuY9Y;S$;A2V;=YO
z?1_AttCd+ySl6y?o}-ehe8FyuqxJ38Pj{Eg-*Z;FCM%V>z)>~lN)flEx1PLgnaZN3
z4HpBaE!ZHi_U5d>8x>QOgw_N|J#<}CoXEXv+9ZbF@~z$c&pz~7xWy{Z<+!-r$V#5A
zW6|P_EZb>{N2*>MeJ!eSD@u(PxOVgRt*wkzyRDC8bO&(zf4!l(e{uN+u{iFfstYH%
zJPgm(IzBC6t@DB-&I>PUc<j7(D8tI)kmivWk7}Ev_bxr5q1q{66SP}xDdU$@R$dwF
zvs;38mvRJf<g&FMZ>aN~wnZsuR>lv;j<wrGmN(5h<$NW`b@^pE-ld;I>MiV=UMR#~
z_4BTq|1>u~FgiMN`^F7|_m;h=y2UK!(l}weYopvPg)j?-T!lp|=eCB{9&69ByD7La
zkEhgU{hb9X5@avudV1bEHnqqt=)nG{Et8^7ygYZxj3q1T{zKWyD3|r&aVDuo7o)hA
zRlk3=)avHq38|l4)Gq1nc`KRf&>OF#e}4CKR@>l}+l(i#S+bjLXW=HLj0FCj6GThZ
zCmQd%_VI}V^K?gH?@vx@D~r~BZP9w`A0b^M#dO@g=Uqqa#+MhLBseKgeN{f`)wgK2
z@|&lh-86`9nyz!V^7+o}4g1?;ZVRxgPFOo@n$8WE9n3{@xm5n}Ti4FDURJPH^;50H
z`t&>JDoh`jpDH-%>}l)P*M0cr<o!oh_MZ5hRCQ+csdsn$QZA?T?3!Asp8cb=$igyV
z=e9LhdZ*m{Idw&W<8<vm+P&=AAv?D}w=R_AW%#$O&M5Ni^T$bNpYJ{Yv*w1`{PcY9
z#S8A7m#ltRKK*1@-Q-DGf2y1x=ak6acHh7Lndw=#yAxEuPWyK;Vb-+6v-3+1+?{s4
zLhdi~Bnh5X_4|(W|N3P4YyM`*-YK=Azh>U~xBYW&$7<<f)o4%8`xWU-ZZpnheAs!T
zHdn6b_vtr1bv*{2@zPxSi!*2Z<z2^;@~OBoH^~l|!-ao&$?9A%Sjx>57{CxN&L*A=
zCV(<vu>RqjmRVN24S5*+fzryX%4uK%CS8=8R{Wlkfr0DSf_u&0rok*c;quDzdN4zI
z!WEPika*hG+HP<Y!*#56;5?WPCN8k`y@!4;So6?YfhXLE!zdM|#j*Vj!_O}zw-{Ma
zguS*`eU4|1`OC-?uA-n~1ydd3C#3#jq75rZm{WxxCY-Cib9%Hyh#wPAxT3P6Hq2t4
za9wL%XSkyBfMmPp`z(QqWRzue5Q;RdHC<qeniHhlRvQ|Z0Tsz9%c&z2=~?U9!xX`2
zMlNu`!8~>-#O`Ih{Pp8N#~H8~D8W=cHDLY1&M)2#a<iz4JWTk_0j-%FYh$H=!V)YJ
zN-&%L3WROCn}6gP2T)j9LJcO&6E3MNiNupqmO|o5D@!BsWR+!+c<R>bb}&16!ZoZl
zka$|wT5ujiQEE;iC=C=BB^L*S7+?a#U|<M?P#`i0!3;q#L*Y#3a4R{}XOZqz%h@7=
zf<wYW86x0v49U4EIT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQuh=intrL~Q%
zoxOvesJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdtk$H$O)YMjrqx
Cq24F}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp.9B28805BAED0F92D.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp.9B28805BAED0F92D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c53eaf299b85d9a8d915b28eb18c05af486a5e36
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nU|Ou7$kWZ7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!m%F4ZL*+Pr6C-$?r&eJ(0vU1{-xnHio*N7}!RKYyU<S>K#qr1-cY`-50
zHfrEGAlh2!)c9f3oV8uM`Q4S8`^D=&{QP?J=W|Q@_#S`uuJ8U&juvf5+-Nh8M`XD=
z^U+}8>DS{YKl=6f&!5}xxBU3_e2)J9J32pZekw9kV|_kfZ+gwopKo42dUjgWez($D
z#Rr#<y0XV?x)gfil)@B_s>4~^UC+0QKC^pqV9Dw%ahJRa>n}!|=;Ut+^oh=tYSp*+
z|7o(@^SB?s_qNV{%#+t@_+`mm(OVwcTfD>*T>8{hE>=IkT;2T7ZrZJNlM2LbMc;2L
z*eCRI!Zs@>hu{mhtWJ1dGGlt{EWKs(+^Buk-yhnlB)whO$d(my$W6C=!saKRE+)Dw
zX>ml$rgCy|Uz{;lvURJwrqV&KrGjxTR^nv~vNbpZs>65wNZ@<lk$gq;Ql~(z^5V}6
z+cm7ZmR)j|bXrrU7p#9gT_t=QW8&Ps>tlo$Y~JnOvyEfwhmiigpRB9cr5|n$h!WEB
z6sfzqr05Mts1=Ksjfhll)0Ax8o@LF<bDVng?p@z;Il8MSXcEUk<*loOr5x6;S=>^#
zD@tQ)d^<zyg=OItJTJ^sJhhdUcDB85yiws@sI_8+jf-w*iOstU0apHB+b;`u@LdYc
z-TLHv<UKyotsApnZxh>g=(b3#`|jgS8Umk|X<Tp?KEhSKfYZff>*fusPFKu7t8T3n
zwsm3O%PG+!**A`qL<dbOD_lBfqR5B!Y-^W@746!(iMw&uwe1z&^TaN`o+@5C)5m05
zQ|A5ORV%~3X{5}wnb=vvop0?X8Sq&A^y%(zmdn$%uI8mpnp$#Q!Q5{h=c}8}dQWEU
zntVbs_N9gOsROA2$0zNX;JIo}w#`DX-;&u){>~4+v|N<ksBLB^DchT|B<OG2%Dnxl
z1=IKD%=@}owc_NZx#jbAOYN>cv6VM4fa&I{yc12nEzt_sS2`T}zv1%roylD0QA+kL
zAF~d<ZCS7W$4^qPQe)-u6NQB~m7k_tm<m?L_4xn1b!+eKiECc>1a3EapTfGoug~3G
z=lZm%yQUfK-XWA7bKxo1f3L>}r9$7m{n;1e#@w)f<$RxY%P%KxTz2`~^O9Tnv;8)2
zUm6(jZh6n%54%r4nN_!VQr4d;=fgQ$WN*9HuXrYU)-86j>ep%iE+)*H_IGxE>4Cq~
zu2;zYVV)%MWmWyYBmKWVS^k>8S+Zw}bLg*`cm8evytaR}Y_V#zr|12ObS5_m|BMei
zZ`9?=75zT_WKLa=foHrlm;U0+8Gm^nXQVO~SLP<!0dqLhbmx~TS{nPfnF0eC!o}Id
zlfeW~1`L8e7oB`~_75WugFjGOnN>LrOu(dzQqzjxGa^i3;sWdLJ@k9Qnupd3JmE$h
zMyW6}xPC3T*X(T?%)%2cuPm<zvzaGcL0JKbr){n62Ghb5u4AnO=Q*~&Vfgu_<Q5|f
zin(6ft3JoG#{6aE30F~2v4ZJ^(TrSRC&Szn;wPm3VxkQzNHwPlKTK7w_Ri_i5+Qy}
zJmHGUirNUr>00YL!xfbWB-=gTX9-j!qb#F?P^4+C=>k*KoFL`4+R(TRs7Ov(P9330
z&sxtOrf7T5j&mg(Q`3R2kYJHef+;!_V)rs${`zsCumOvK5=`||1J*C>{NmjpH;bys
z!-W3|gl)Q;f8-el&=6$_HJBkh;gZUdNIWTJDI}h>vNRG;R#_H_r*5ro2Q!=}T*F!e
ziKk_)1?MporRF4p(m-)ha&a(-0VY5U2Ig=f^VxwM1wPzt5kbKrVWA9RU~w=JgkXjs
zn4t(}1cI5Io05|O<a($1GPAI<v2$>8^Y9Ar^LYxoI;#nCc{&M+3yVldT3A}!*xK1U
z=!t4uDVu4?V-ppVmXTFZR8mn@*VNL{H!w6dHJ3Ay(lv4s6?6BBWODO!^kDP>00YhG
A$N&HU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp.7E6A2A9800DD0E7A.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp.7E6A2A9800DD0E7A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24f610b9917cea210297ac533ad1d645db5ca971
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nU|Ou7$kWZ7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!m%F4ZL*+Pr6C-$?r&eJ(0vU1{-xnHio*N7}!RKYyU<S>K#qr1-cY`-50
zHfrEGAlh2!)c9f3oV8uM`Q4S8`^D=&{QP?J=W|Q@_#S`uuJ8U&juvf5+-Nh8M`XD=
z^U+}8>DS{YKl=6f&!5}xxBU3_e2)J9J32pZekw9kV|_kfZ+gwopKo42dUjgWez($D
z#Rr#<y0XV?x)gfil)@B_s>4~^UC+0QKC^pqV9Dw%ahJRa>n}!|=;Ut+^oh=tYSp*+
z|7o(@^SB?s_qNV{%#+t@_+`mm(OVwcTfD>*T>8{hE>=IkT;2T7ZrZJNlM2LbMc;2L
z*eCRI!Zs@>hu{mhtWJ1dGGlt{EWKs(+^Buk-yhnlB)whO$d(my$W6C=!saKRE+)Dw
zX>ml$rgCy|Uz{;lvURJwrqV&KrGjxTR^nv~vNbpZs>65wNZ@<lk$gq;Ql~(z^5V}6
z+cm7ZmR)j|bXrrU7p#9gT_t=QW8&Ps>tlo$Y~JnOvyEfwhmiigpRB9cr5|n$h!WEB
z6sfzqr05Mts1=Ksjfhll)0Ax8o@LF<bDVng?p@z;Il8MSXcEUk<*loOr5x6;S=>^#
zD@tQ)d^<zyg=OItJTJ^sJhhdUcDB85yiws@sI_8+jf-w*iOstU0apHB+b;`u@LdYc
z-TLHv<UKyotsApnZxh>g=(b3#`|jgS8Umk|X<Tp?KEhSKfYZff>*fusPFKu7t8T3n
zwsm3O%PG+!**A`qL<dbOD_lBfqR5B!Y-^W@746!(iMw&uwe1z&^TaN`o+@5C)5m05
zQ|A5ORV%~3X{5}wnb=vvop0?X8Sq&A^y%(zmdn$%uI8mpnp$#Q!Q5{h=c}8}dQWEU
zntVbs_N9gOsROA2$0zNX;JIo}w#`DX-;&u){>~4+v|N<ksBLB^DchT|B<OG2%Dnxl
z1=IKD%=@}owc_NZx#jbAOYN>cv6VM4fa&I{yc12nEzt_sS2`T}zv1%roylD0QA+kL
zAF~d<ZCS7W$4^qPQe)-u6NQB~m7k_tm<m?L_4rqA-B!I_Ir6&X;=I&*C%NimW&8Ug
z*6C`O>ZX<zs$Sc%;I!z!smGd4SKYb&S$1a+8$*4t{mdBu<%g4em(M-_a?5?QxoP>{
zix=GSm#ltRKK-QWUpLjOKfboL-ORAw*8MNwtoE$lU25L1bpI|sXrlYu{Qj$k-@5BR
z+WcTwF?<pFzot|E>(k;d_UXnFn(eE;81MX>|2$eg-1?c<wn>xreLTm~V=(W+gW`>U
zZ`nNgHvOdKUx~y?`^-e*+%6mZmcF<w<xg>CZjv1^hbz0aUg3V8o4~{1AHWc<%&MFQ
zCV(<v@VKO7#rriT*SMJifzsk^;>lnFCS8=8R{Wlkfr0DSf_u&0rok*c;quDzdN5r)
z;R?zMNIY$8Z8x}u;X2kja2`yQkqc~n@1frd);zRUaBP3W@bgQ_Ek>4bBMzffm@2RB
zRiEQoWBxMogsUj1Siyut{DjnBOtfJI33ICO!-R9ScTSI%2=Qa$30G8B)J8Z?*IL&Z
zuBbd9+3xv1OQ0edWf>iWB28;e7nq{v1Sz-GhQ?(;MRLk=>Ig-8)_V3ZMZ0Ami`<%M
z*9mlm1dD_cOwpkbyO;6u*N+2*4Ok46V5(u#OkCiAg?XF@#lKGtSii9Ii+6)OCaNM2
z)B9H-Y}4KRBhNU1HYrP}!HnVwmsFNS;z=n>A@QV@rIC2D%Cbm2b!&AynBhF(8rB*}
zJS}T2IFF$yH760228xT4i-UnQb2#7o_S=Rk6Q8q11O<nLg))Fdp&$&#U|<MBFhdZ`
zPy{mq!A#Cg$;kk6z0-V|Sy<WFIXJm_cm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&
zM76Dy%{1h(iHb?f$SNo*si>-JYU$`37#f?J%b7^&8o7vyxqC%2x%oMIF!}%hY_H`N

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp.E0C6A616E43DC9DD.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp.E0C6A616E43DC9DD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..594327118c05270ed4b8a28a1b2a3722061d57e2
GIT binary patch
literal 1796
zcmWIYbaVT|#=zj5;#rZKT9U}Zz`!5~#Kk2=nU|Ou7$kWZ7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!m%F4ZL*+Pr6C-$?r&eJ(0vU1{-xnHio*N7}!RKYyU<S>K#qr1-cY`-50
zHfrEGAlh2!)c9f3oV8uM`Q4S8`^D=&{QP?J=W|Q@_#S`uuJ8U&juvf5+-Nh8M`XD=
z^U+}8>DS{YKl=6f&!5}xxBU3_e2)J9J32pZekw9kV|_kfZ+gwopKo42dUjgWez($D
z#Rr#<y0XV?x)gfil)@B_s>4~^UC+0QKC^pqV9Dw%ahJRa>n}!|=;Ut+^oh=tYSp*+
z|7o(@^SB?s_qNV{%#+t@_+`mm(OVwcTfD>*T>8{hE>=IkT;2T7ZrZJNlM2LbMc;2L
z*eCRI!Zs@>hu{mhtWJ1dGGlt{EWKs(+^Buk-yhnlB)whO$d(my$W6C=!saKRE+)Dw
zX>ml$rgCy|Uz{;lvURJwrqV&KrGjxTR^nv~vNbpZs>65wNZ@<lk$gq;Ql~(z^5V}6
z+cm7ZmR)j|bXrrU7p#9gT_t=QW8&Ps>tlo$Y~JnOvyEfwhmiigpRB9cr5|n$h!WEB
z6sfzqr05Mts1=Ksjfhll)0Ax8o@LF<bDVng?p@z;Il8MSXcEUk<*loOr5x6;S=>^#
zD@tQ)d^<zyg=OItJTJ^sJhhdUcDB85yiws@sI_8+jf-w*iOstU0apHB+b;`u@LdYc
z-TLHv<UKyotsApnZxh>g=(b3#`|jgS8Umk|X<Tp?KEhSKfYZff>*fusPFKu7t8T3n
zwsm3O%PG+!**A`qL<dbOD_lBfqR5B!Y-^W@746!(iMw&uwe1z&^TaN`o+@5C)5m05
zQ|A5ORV%~3X{5}wnb=vvop0?X8Sq&A^y%(zmdn$%uI8mpnp$#Q!Q5{h=c}8}dQWEU
zntVbs_N9gOsROA2$0zNX;JIo}w#`DX-;&u){>~4+v|N<ksBLB^DchT|B<OG2%Dnxl
z1=IKD%=@}owc_NZx#jbAOYN>cv6VM4fa&I{yc12nEzt_sS2`T}zv1%roylD0QA+kL
zAF~d<ZCS7W$4^qPQe)-u6NQB~m7k_tm<m?L_4rq&zP<O><67yl711-xH#Nn(A789I
zE&H@rtoO{=m`+o@7e%f0OFt&`YJLA!<F2=u@xc8n{%Y4RzuXXUIrDk3)!Xg9%WvMk
zq@nRW_*mSJyQxLK`vW~q>-H{8G`rRJc9DI^=dQHH>rPIp^8TM;;CuSt+3i*j>P~0d
z@c(1<<f*!9AOFa`_NUC>^EY|iPA$~>>y!8Y_Rnj_ulDVnw06mo{5{5u%022PKVowB
zoAK}bn_A>&?>1w}dfwLQ7fsImV`G<0<u0zwO|k>#a5<LJ$PxusEglB{0ETd7R^>D>
z0h9rQIA=+<S+l<G=4J{6N{h3JCxZ!?bWv(r@q0!F2CiQV?lpUx2D9*l%PY(4!F2J2
zD<~@<@wBbA-QX66>saf+c`#8%F0l2zhkh?u^UzwsvHcCh&o3pn7+J!NIE+$Zs@y%2
zmbGx)kYVHrmtc`lf?4Oaz3Ov3Ys_Duu!@3;6-;%ApOE^Ci8ibtVNMl(m@rHi6BpP^
zFe`abY|7Q%IXzk;#E*$5Tv1t38>WXRT-RFH8D>&$c|fw=^L>^;MKa1VItWFY)|xIb
zMa>CPZmSKA%YcgHl;zYBiuA1Y>|u%yh1k7}m%n}-=u`t110|U7Qv=p7?EK>0ARmaT
z$isyH3WROCn}6gP2QV0vCDaf`N-9et@uZZcka*I{(nvg6WmzPiy0y9;%y6D?4QmY~
zo|d&1oX1d<nv)1h1I0zj#lav3m;f;tn8SDbHvF|xJg&<Y5fmH}7RnF?76%hS2xbU^
z8H!*=AehOyDLENHu6LR*GYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dt
zo~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X*>CO1Dv4@MsVb+EKp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp.01988F25803E8456.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp.01988F25803E8456.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8fc2ae409e11a6eff8767e25a8ce52fc7bca97a0
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;=~{!m%F4ZL*+Pr6C-$@W&eL%cSvhga+%MPPYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk>)?ksd_{IF@p+RwZB-IbdA#p^%({Ce}}b4&a99)I?x@BZJqpFTK{RB0p0;pflR
zv|LMX{eHEBUyuL%x&40T$G7J#<Lh=te%So<X^s!mIr~_>pPxS8yngiTbnUt_*GbL?
zmUnma?MP`|Jwen_L-0%6)x7R`+}dYqFEo0FU)AfF>KM0Zd&Z4b7u^hVmu8CWd+_gc
zzSQ1Nhwi^Oy(5!epxv`jeeSvoZcz(Et=N^rGMsKzy|1atzi|J!)9un*cU=8X?mjzr
zz00m8Yx@K)JlV31FKo+YGp4!D(pxsqjo4TH{h_T&(%YqtELkCk+;qz)Y)<)fG1Waf
zL|~iMDJGF;iwvzzxw89%T$)6^l=pO$=)ZBe9w4yb+q$BUj<WXzkFU^PBCYt#&F#4}
z|HSeqzDwJyIX2#!>9bX4{=}HuY%QhV_7(Xy_`R1iyrmR!DA?}Xq4zIjb6T!DWqPhy
zvi8GkwWZ1KD=sMpJW#xzA+jVQJJBRo=E0Msl8^h2&Al@zU1X9Ahg@gu+(Qgixw5QV
zCSP@0`PYi&M621}1IZ5i6@8-wHGNHL_>UZm{uJ!!`cg4VbMfQC_O?Tnr|jp<b10b*
zwRP&GyXl42Q?4AFb}d&or*-?u-F;>KtP_>j%(8Lk_Z7%~)2JqJG34B#ReFzVP3F((
zh`8FOwM;lKaLR_xmC->H%IcQRxTyCao;S)<^-0O<)aOj1Yx6%&v(a62JzM{k(#i}S
z%gg(|{|tJ2Z}9}zO)gqN;`hpWj20X^KYhCR+hV`-Q&!zQqoTX=dP3?P9nM!b4||=M
zv}*DR$=H_`)~61n1{`l%Gr{ZBoNSwgUcV)?oBW+0d}#^nzhNppzo*a5BuKM<^OWuP
zjdq-lGdut5rqG_0mCxrsUw3RB|ITfSTwDs#*LgEHbQW;0De;={!9M3p>9ZM!dZ(S>
zPuw%B!203SmiwBARry)3N=TmTSNH!Xv|MdcP0gv(m)@=IONmYI>6%)puKlC5#KtN~
zJ2!G+w&vz;?Z8J(RoZ{FyV-N6>@Z($T_m~HzTVq-rq1%qiOZH<KKG>Lmi%nL&E}~S
zgC8vK+52Jk=@ifXfsUMYdlw}d+{$`8$9~F(u4jwZc}}kK{+(gqd+OiW?N$%|oygYV
z|M$~_r|P1;`G>`;YWn^?&tZG~Bw$*d+V=mib6(%ga*sK9E$GtoyE7XEj|odv>1~sr
z)qdfhk>zrEWuHs!ZF^2<2&L6CudGYvD6Y&+vIFMt>Kln#jk7i!=4J{EU<emy6HfvY
zKp8NYekkPFF~d{bJPiIoX=PUBR4@UPE=o-+e$U9j!1Zgvy=HIIU>2Tmd1ZM$n4vu3
z3d#ydJZ)=jH@J!6I@UUHp5XdUsnT_?4m0tD%PGsLAq>#7*0Y0YacqCX@bgQ_Ek>4b
zBMze!n6TIOs?YJPF@G6(!c`PhtYE?+enRRmCfcxqggI6CVZyoEJEuoWg!ln%QdCyd
zM%bilt?LX|R34CQ_k5ovP?3zXj1EGPrnROEOc6|!i3=QLy@!4;So6?Yfd|ELFvW~q
z;DCcEIuv5}GG6}raiHT2SPYb4s-7CKeqrYq?*_S9R7D;p{N{kxOpdj&Qb1t|76~Po
z#eW6DHr>rX@{9v0tSo^PXp+j3NIWTJDI}h>vNRG;R#_H_r*5qd4{V-r4QmY~o|d&1
zoX1d<nv)1h1I0zj#lav3m;f;t7{VYFhzvq7L*Pv2@U!+DQ@-3b*u@qR6dV#3$`A^d
zV~9X7lXFvYGJss~G+$;GRyKAHPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fTdj~yH
xZ7XFn4S8&$V$w3Q3W`c9s_L3rI{F5N#-`?SCQ`aaE}~-YUXe_0evTfDJ^-hs?6Uv>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp.AFEB4718FB33138B.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp.AFEB4718FB33138B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5841a9e42596a101a931e2131c25f2c7cfab99c7
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{rKeM{YO@Is>sj(G<9qBYURl5l8cj4@15kTla=l7i&&?t
zU8<XEwp&blPH=u}ed@;ry+YrgmALCI{`NdSQ*H9<mnAb+zT9GLIrrP~WtZRhMv931
z&;-WiGozi0?yp$FvOn&kfkf`svUBoY6|TjB*Oz$iU0Q20LoN0HGilj^`X{E-*#Fn4
zu<d;@U;0O2sGa-&VzcJLA`Nf*$+!QPZu^?IRe4>Cx8}>w?|cq$Jm%!J6~E1Y);Xho
zX3T|tLA968&i7JHI5+QSymus7ytpzq$qtymjV~R_HdU)J<zes-U<g-cRZayHKp8MN
zvfgM$g$9EMH&Y-`TAWQh2~5DGi&E2y-!n2WaQ#|vui4u)n1v@?URhobri&+BL0JKb
zr){n62DdO=$65!@6I|aZRl4rgVJ4n%Ib}IDgaLZidUh}^Fq)AI?8e?hzZa}|XszJb
z{)XY_my%nIEa65RMkz2=UfZia$Fs)#W#kE0QBbjh35WOzslS+L!wM4SRN;pS=W6eq
z9xV~#2Xug<vZ6M^0lL<@&TvKL0m*jH_gMlJ$tcU{AQWj@Yr4P`ZSUD}u7qQ1I?w_M
z76~PoqC+8eFXQE}9|sB>uox)8RKuj1xWEAm^Ait>SDqTMeqrYq?*_S9R7D=9_pd<M
zrn~t^o^b$eQkFmpG)ZMiB%YMA6cSHbSsIBat1OGeQ@2)!2R2W*hP4I~Ps>^h&SNM_
z%}E5Mf#Ra%;$RRX0z^Onm=9rw!NeFCf)LCQ1Tz%QWDXaevG`-#^W`7dB7%ZL!a^C6
zb5n9Mfb!mHzRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6R?21?
s^4LVhq-A6k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_BAML$96cC)06RFm761SM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp.D913F33B640C756C.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp.D913F33B640C756C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6749483006057b330068d73b6ea73e6f0793ed29
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{rKeM{YO@Is>sj(l$H8D%ze{3W4EKH_DmN2V`U{LyD=s@
z>{ZmMoHE@oi{<yY{-1i}D6RVL?909wx3}l-U-D54zx>i5_;O}?k=5IN-{m)FZ_?2C
z5PU4|$Jx{(*ZqN>oOOE_B^unydOOE{%7?CJi`IEguJZn!Vc>h}-&u332melF>+t{k
z>A_QV(cb*S;#D<$|DNZtJ$@1}txj$G|JOOM?`FBjoSYVP>G|E64T8snrK<F{$<J!P
z@Xu)H0(oVhOYLoYPG<<E)ia;vNfs)u%uTWb=I@9X!FQhNoc_tf;2*#cuFR^O3MPOu
zV6e}KV^zTMmiOFDfk0_-Ht{4d0h2CDO)GxS$iTq$Yr(x{Z_{8Fo^W|(c|Djeo^S<a
z1tgxfwYD4F!f+jH9XL;LeWz6Ex>tvpc*5nB<<t-c=vnL8!L&HGzhU_KrQ{YPOSln-
zQ3_1hYkSq_c-EM|j6C5g3My7G;SfI|^%oOuSV6*^D*P~E7|qB9_Gs^+-wW0}v{uN~
z-Z?#5BE%2q07Yd*ZG;1Kt#zGYdUDGHlI@=Fvji%VQI^p`DAKgnbb%?_E&EvH)<nBb
zpuG|-5=t;dFq(-A9D*==c~HD^D8%k%y!`d!K)nVm21+m^pBk`!Vdod`26;tPMII*n
zS0HTD-TWiZIDj@OOCSZBq_QLuPfA$|i6^Zrjl`2xmPO*JTdTtZn<reuS_6rvWvvD0
zF%+fdB!bdFaZz${Fo*#rKnw<kFbD-AgAmLR1Tz%QWDehTg3E#B&h~J&h@jw*uuz5w
zxEw=rZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$r
wrEI1lk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEglF7}_(Sy+k0N!Kovj6}9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp.EF41093E7954CFCE.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp.EF41093E7954CFCE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..387a7c6dedd9e90e42cadcc23e7bb7d8eccd1640
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;>6(L`O5Pr}Y@x;36Z=_Q=jj|0Svhga+b`GOYeW_<s$iaFa+twg>Am9ndDWpt
z4Lk=#TML~UKWwsz7Q4;wuGHMW+5X4Vvp<hM_pURPpU=nge*T;1-3J?gKB_Phm@|)u
z#XmIWZjEo#?EdNE_v@A)-EOb{^U>+IbLQ)>J9kXw!Jo@loAmAG*U!CQxAWQa>nfHk
zHM7oIJ?NUc&MAy-q05Eru-&{>dTR{6JF=_AhVE7<YOZO$w<xi4ac|Nsuge-W2memL
zZ({ZR(EazOcjVFw)O!}H&s}%HEox!tE(YZy6UE%U<#v0ozj&X#@~u(swu`}^-f=zG
zo~W0(TA9U!b?w^bIc2xp78GYRmu(Y1UH;x~PrJ*Sy(SYJdcAI}c*<8a&Bt!nn~ZL*
zLvHSoO^2LJ(w4g=z76d1*x;1eB0lk5tLzIYE~YE@LiOxc9M&!D+2yoLa8Z6riQN~0
zO80vqRXf@RR?j^da{8m>$)&jt5$W%>ud}@P@=kEEu9K?G(`5ZWefJJVRp?Hc*5Y-G
zYky9VRauLcEEB67SFf96*H$sN%SRdg^xVYXXWu!gTj{3g+45lGjjNhG3#MNSbg+)p
zopNivbAwaHWo{d`FWHxth)%rZa`?ML&K^~J(U2)}i>4V_$(3hl%w7KH@Jmh?_LZU0
znm233wmW;~zB#q4x3%|!S>^SM?=~uMwd~R4`f`C~;}N+GCPCk)I%{4j-E;qZva=&J
zcSVxbL$;$_*9=R&xh|bsFJgR)_0MWY?UgNa^0vN-S+MA<?!8INUB6Vlbd6OL?Nv>`
z_1!Lh)!k~5O+M2+Twfj9-51Cf^5f{EtZR1qFJ}gAojr4j>Q|{J$4-hm+E$-vjR`Cj
zvRte!d%n-8K=R5)740RHjGjyRuekKD$MoQG#e%9sE8O2q>h*VXSC(2iW&h?W+wU9Q
zNYkI4{x^p+Zqt>|=ZeFVueZnC7GPzap!>Sb<c7-*)}pyWDu4K;Yv)=nD_E=g$ys8(
z+8w_iS%<ctDmdxvY3kP2{rKeM{YO@Is>sj(bn4dLs}t9}?g`vz^ge}ke_x-wyUz7#
zQ+G`>njPCMdT!<RL;E*XTyX2GF5aa)J@EJE?Ix2?UM;aYbEV|g%sGC)lP_m}Tc*Rs
z^=HaP@w(!fF&Fbgf|}yjXGrvHi?V*suexW^&K1Cjxp&D<%4hPX`p><5JM2G6oo=qT
z6Kan8a-R3k3a$Bz|69&JXkj^J>HL##|6ARz+I~wgeACh?Un<Mh5?VgCw9V^&+x|>(
z%YGld7soj!e_5#b-e^|K&HGG=$%+icmAOfF!2Io}nZNyS+Ppd*2LAwtaAj8IR4@UQ
z0fWmN`)qZteVED36bO_SXA@5X6ENwb)U@LFj0_B1zZTqU_BIV>;R%;lme+&n;t5w!
zRzTuuTWh<)EezMO)`9Z`*LO;lu6uQui6>l6SxyaMfS$FU9ZZX3`x}OzUrKH<vV<FP
z7^T33-93_)wQ$^!VdM#yV3AOQ>Gj%P^*Nq3<}XlKMM1?1rW!^wa)CYCd+7IqH4m*7
zLi~i(Ure-N1*zs#;fJZp)!sQhS|Y>`=q^QNMQxbbJmI?5y3TM#<pIfd&-YmZ70D>e
z=pYnnT5Gz%6dek&dl@f({W#D90~P}%m@rH+6BjsOVb0+}@!?Yg)-UY*;@u#}iK@uM
z^!^nH+jKYo$TJS0$CV|J0!>m`5{W0JEQQ3AR+dKM$tuet@zkx=;epK)u3@c##M83Y
zg7X-PQgaePX`r|$xi}ca023ev19NzLT$N2r$9D_1h@jw*uuz6DusE0qLNG%R%uoa~
z0>MnqP07gsa=p`hnORub*f}`4d3Xi*`8)+(oz(=nJe`EZg+(MJEiA2VZ0+nF^hC9-
wl+85cv5AUF%g8DyDygWdYijA}8yFg!n#-9;=^D9+in)75GP(IVdNBF`08hx`4FCWD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp.CB877B16EA5C28D0.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp.CB877B16EA5C28D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..acb6a31ef33a766f44b7a2d4ca376003cf335902
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x)3pYBm6dy?vV|6BPwZ#$ou}g@vU1{-xo@t&*N7}!l)^e|$zcX}rT2>O=T(Oq
zHSipGy0g%!@x!JWxu0(HyDK&Ki`RendG+Sc=a%;IJ^t)X-}}FHUwv>OsnkZ2!_S|s
z>A2Rk@c79OUOoQv=l1(8KfXPmvwmNU{*RkkmeP|OKKq{*jjy%)y=%w6qLb(KJU_{N
zc=>1%^RzVuqMMXBRGaQ?y7l%#aboLEZYzVAt8TGwG&|w_MT|Gv{FcVDwI;oX#P8Jq
z^j&;C?#J%tw%3n&@>&ePEV(Rt%R_sM*YkuUQobJ1@8|lz*Zs9SB4fAb?AtFiE6sO)
zp6W4u%T+<9S#z#kJ2<QCmfM2jjOMa!!l%pM+wEz0S+mz<f<v#@jTKMjil+J4?Rb;X
z&2`AnJ+k3Yj7i#hx5T%BT^<{pGF!wazH7C8A;raX<zA@Xz14@KA4!z9mMFX2Kk=gG
zh2p2aJuANy@GFK@PhL6gvG646+YKAe-3wjEdEuq4f?ITpSH-F0jz7$+*rOkA4Tuub
z%GCa|%=6_9;Z-FZA{8f0rC2quMM?OIvs>(xh~2ZgP=EUq$yOD?CibN8<pvG+BHS6@
zrfdy(wSkYZXrZtE2Z;;$o|D5|yo7o0GjIGj?N3NRU{y!dsTUP@7A`E9_lbYGx`6DG
zu-v67-#zcl^Ug~4eqAP3cIft#Soht>n@$K`^PRW2U9E%nn}cvihE{rFsQ5>_S^a)4
zYqFe0FLRb_c-=U5MOV|KdVkQF4Dk=^+14%*D%y2*Q+Z?8we1z&^TaN`?iDYc7-BMQ
zPUijgnw4em15YSLDv7S-&bM-t40!l_^;6g1b1oY%&9XQ4^nPU;A+7G+RP}9x*rp|~
zREpfft<Los6-Zv$sGz-MlF@T1{}q?&j?H>-S-`?J@rrWU;bZEH7Yp`=oVu^`^zC+@
zyv@_kZm!$Lq96IH^7+oy8(%xG%k5}lI$>7oER~~lhbeomkjfu^>DsxL%L-PjetK>Y
z=euL>hoVFIQy)%hZ_4W7d47Dddwpsq|1`Hv>x|u4Pwkm3`p3#jPIlwgZEG%UopST&
zQjJ2#zasmxf5u6xzFr<6v$N;coc&v0`Aoa~(t!JNW_p^{+kW5WH)ltm(_9gJEba$y
z>ZT?4S1fq4H}0YVL+;m7KKtosx}MEh=Q;V;%x@PHX4U+i{l4VD;fdEH<o*~>l=!l!
zKK}&!)$byI&)?)ZcDOjQ$~*7>^2n>ZFZHcy%3ZR=e~U4ra*w*nk(eBNvF{K6q!#(v
zyUkeA&ssU<V#?`%jMYzk5*UgrbCc|VIeTf_{bk<A6<D~L0s|Pr#o5HezyweR47M`d
zYR*$x63)Zm50qABRgMG`FzKSywBl!s3=CYq7Tjz0HVtOs371!v*Mk|#6Rx1FfW*_b
z)^>xN7_MWj1LrxmzhU_KrQ{YPOSln-Q3Onp*Y>K<@vJd_8F|806jZEW!XbV_>MthR
zu!4j+Rrq1Tx!OCYM@xkGG4X^eDl2LuT%l{N>kL;^9*}JJe4iyyk&Lp84nmQpwWbS9
z5lobc3mh7~hkh?u^Uzv>2gPwP#f)6wFo7vL6k_)>UjF)VpyLc!43uE1o*J-zVdod`
z2Dw>OMII*n=782rj<vB;Kw$|M2_=}te+9xe-OWGpj0-5NETIMy<_VWnmPF!7DN7;o
zq?M(Sc(Tf}NIW@ZIV7ICwYnY51w7#z)*47WEo&_#o}RTHoL7{ZlL(6W;-ci@U=RaL
zfEWzS;WNw5haAn?xSA~@C^#f6lpzc(4km&S%n$@K6v2!@Fq3mrax#Ei?=)X#7FITP
z4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlAvI>eyDyr(5
eS~~g$hQ_Amawbx`MlPaa?p~2hZhnp)j6MKu*wg<2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp.7AE9E872C7D0A549.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp.7AE9E872C7D0A549.idx
new file mode 100755
index 0000000000000000000000000000000000000000..091633c603024fe8031a0986e0c5da80ad5ad883
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x)3pYBmApM{*+Pr6C-$?r&eJ(0vU1{-xo@t&*N7}!RKYrH$zcX}rT2>O=T(Oq
zHSin|Z7p<a{IF@x+OFOF?n=%5pY4Af{d)7~bMCr*67%_3-p_v{e)?eJ(N7gd0(0i^
zu=t0@#MS#Yz3vx(e!p({(e3v7KR%s4+caNSKRsEf;NQ!llhfy)-=~+iFXrQB?MX6B
zc4wc<7M%17SJZ9}P<p|;`mUI*c-V~J3z#RbTNNv?({Z0${>2S@F1i`!F3l9#_u${@
ze5twB58r>6%Ih=UA$sh>WWVqX<+T^IVjBc4XHD34&wBnn=`Y_mUMZVuR(>(~Qy%wo
z?TLDstCd+ySl6y?o>O+qZ9#EHbJ;fG)8+5&_O!dK*=sVvq1Wriil=-<(|qi9y~*h2
zIuzy}*>uRsByG7{;@iM3j}1<lE#ed3waUJb;$pgTFI3NN#bN755~Zyr$}aa$yr_Ah
z_^EHt$}a`{iec81S5A9uJW2a@!-jMB!q;(LczIXZExN_4;?(iDALdo;(hs)=L<wnS
zYX4c~`SOO~suB*73Jnt}R?TZs629W>7CR+k_pC2e-~L3hRYkDLJt=&-K|@@mJLB7w
ztpQp4`53tt`s#m>xRCELIn2dNnD;*O#*fqfgaiaubwr(dQE_MC!jgHP_?N2-$Sw)X
zU7GSe@(!Qs){WV(w~1{#bo)uH`|jgS8Umk|X<Tp?KEhSKfYZff>gElrx+~<*s;eu7
zZC%*)vL#w1`^J$gx|$x=`-9G8h<{kmwswh7(XOkTiW|GGZLjd2CwB4mRPoZ8Atux2
zWZwT>v$E`c;0dEhCDE1K`POcd0T0DbpRWA9XF<Bx)x5Mx)2?JskXFCW^7W0Pc#-c_
z&nG?WO6JI>J}_F5>>1{vntIN3Ucl17J*EejD;888TH*ettJmMnU0G`7l>L#XZs$wh
zN!6d7{x^p+Zqt>|=ZeFVueZnC7GPzau=Z7($qknstVMH$RQ~Ww*Uq(ER<Kz0le5Hn
zwL5-4st&!M@^Di7q^uqu@#B-#>$9`JPjlO}&e)Ch)Sk(rf2^$JWH-h{hrNnAl~bpB
zZO4K-uKQd6Ea;u|HCUrB#x3{U{iss4>z7{|a9_?$PqTX4@4Niw>}WlW6~V{ie(<Ii
z`R)(&JXyVWQ6j^wSFdI4rkv?|HfNpZ<X<ztT}+r&^LKWB>4Co!uSdxJF`g*#Wl??p
z3HPhtW&WPO$#d-RLY=*;+y7r)^C~veJ?vyy&?WU-GaCet2}>pEZIkbAFZgF<xm;e^
z=hE>en^PH^QvWloKjf3oP+XatWCzUIHpVwEE;$|Zn42jufFWF*O*{-t0A;{nq6SZX
z+?MJt9tMA)v@)x5B$$9n7p0~ZKVxKI;QF=TUbDApFbhw(yt2F=%ut?i1!V;!p0>5N
z8{EWj9cvvp52k~O3v7Mwq2CMEJhWEe2{+;}ihyZxY=6V>^GnGsMivxdukBTz<5^?=
zGV+A0D5zM$REPKpslS+L!wM4SRN;pS=W6eq9xV~#$HWt^sH~_BvzRAb*IL&ZuBbd9
z+3xv1OQ0edWf>iWB28;e7nmX#&Bz4~6PTN~_v|=V!Z9@+Xs-l|gc3~ep%A;5@$%P?
z1BDG(43uD+pBk`!Vdod`2DwsHMII*nS0HTD-TWiZxPXQzOQ^vN;R%;imPF!7DN7;o
zq?M(Sc(Tf}NIW@ZIV7ICwYnY5DxPo+YYilxmbDfVPtRHp&MQjINd(1waZz${Fo*#r
zKnw<kFbD-CnZwU-U7`OjGCz+kA}BZ{ER-P#F2@jpV1^=?5eQ~-Zc0uDkn5f1%gn;c
z#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$rrEI1lk4;oeT1HkuQAtHr
gT~kX(-@wq=)LhO)O4rCmRLtEglF7}_(Sy+k0Mdoc#sB~S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp.E366D05CBA9CA229.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp.E366D05CBA9CA229.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f37c7ad45668d8aa3aaacc45e18b4651e1033ec
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x)3pYBmApM{*+Pr6C-$?r&eJ(0vU1{-xo@t&*N7}!RKYrH$zcX}rT2>O=T(Oq
zHSin|Z7p<a{IF@x+OFOF?n=%5pY4Af{d)7~bMCr*67%_3-p_v{e)?eJ(N7gd0(0i^
zu=t0@#MS#Yz3vx(e!p({(e3v7KR%s4+caNSKRsEf;NQ!llhfy)-=~+iFXrQB?MX6B
zc4wc<7M%17SJZ9}P<p|;`mUI*c-V~J3z#RbTNNv?({Z0${>2S@F1i`!F3l9#_u${@
ze5twB58r>6%Ih=UA$sh>WWVqX<+T^IVjBc4XHD34&wBnn=`Y_mUMZVuR(>(~Qy%wo
z?TLDstCd+ySl6y?o>O+qZ9#EHbJ;fG)8+5&_O!dK*=sVvq1Wriil=-<(|qi9y~*h2
zIuzy}*>uRsByG7{;@iM3j}1<lE#ed3waUJb;$pgTFI3NN#bN755~Zyr$}aa$yr_Ah
z_^EHt$}a`{iec81S5A9uJW2a@!-jMB!q;(LczIXZExN_4;?(iDALdo;(hs)=L<wnS
zYX4c~`SOO~suB*73Jnt}R?TZs629W>7CR+k_pC2e-~L3hRYkDLJt=&-K|@@mJLB7w
ztpQp4`53tt`s#m>xRCELIn2dNnD;*O#*fqfgaiaubwr(dQE_MC!jgHP_?N2-$Sw)X
zU7GSe@(!Qs){WV(w~1{#bo)uH`|jgS8Umk|X<Tp?KEhSKfYZff>gElrx+~<*s;eu7
zZC%*)vL#w1`^J$gx|$x=`-9G8h<{kmwswh7(XOkTiW|GGZLjd2CwB4mRPoZ8Atux2
zWZwT>v$E`c;0dEhCDE1K`POcd0T0DbpRWA9XF<Bx)x5Mx)2?JskXFCW^7W0Pc#-c_
z&nG?WO6JI>J}_F5>>1{vntIN3Ucl17J*EejD;888TH*ettJmMnU0G`7l>L#XZs$wh
zN!6d7{x^p+Zqt>|=ZeFVueZnC7GPzau=Z7($qknstVMH$RQ~Ww*Uq(ER<Kz0le5Hn
zwL5-4st&!M@^Di7q^uqu@#B-#>#yE=r`wwno8H4UrBYq{M`?+TRnpFFYp!gaa`TVZ
znnK6FBKxv`I!dd)ULGK`v**^F{aau8tXqCLQEb`eb7x9!$<OxNY`#4vAYggV-Vf5J
zp3JISJZaMRD%ZmdTdrQWtkIY$de$;_vf5YUw~G&&{QPBp|5d|p<@FnFexxfIzHt42
zZ$kg7_m;orZ<g$7caE&`&ilVS^6KtOeQPGIU9!Y~i!q~ekGjc`m>hdC{)c~3i~Q`}
zW-RGvt(<Z(<@7(sc}+f948@hXNp`@T9c}%qrXy-<I5$&Z07JMqn|K(Q0Lp+t*BJ?p
z?<sR*co_VF(#ovLkzfKQU6h(u{EU%-f$P_Td(Gaa!7M!C^2+jhFhhC56_gc_c-q$5
zZg3OBb*y#ZJjeDo3_rh=++t)2H{vjgfGP6YUiCSiHRdlPPq>PLiWN*a#7{{5#Y7uc
zkT9nTKTH^AEfW{m550$eFIe-?T7d_}$XxB6)1xIq{Fr#c6_pjWVS0GNb**)sVJ794
z2PE4)-)9L_B%>^&gHWVtt?2?&1fv<bz+nQjZnx}Xky{h(I)RRpV3AOQ={*!;_cC7o
z`f;GJ0gHhWO!HF%)-UY*;@u#xh^olLg#QYJZMvI(<QW&x5M>E9m?1pjlFE`uJSk-<
zB%ZXgG!jo%Sr&;Wr!0rWQ@2*PgIUEBu3@c##M83YLgMLJ>%n<NsX2+Dm@h6$E)E7U
zzyyfFzz_zZfFyIcp`}zg%khJ2*dl_0L&8ECg5YutAqZwDf*FBeCg-N)WB|F|X}-)X
ztZeKYoZLLT0{ncQg09YLf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_J
hRMj=Lbo31jjZMwvOr&&;Ttvm(y&{?1{2V<PeE=!C$&vs7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp.F88287BB0527CE6D.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp.F88287BB0527CE6D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bf87b5d9096e21a392eaf0b7d24e08c594b79342
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x)3pYBmApM{*+Pr6C-$?r&eJ(0vU1{-xo@t&*N7}!RKYrH$zcX}rT2>O=T(Oq
zHSin|Z7p<a{IF@x+OFOF?n=%5pY4Af{d)7~bMCr*67%_3-p_v{e)?eJ(N7gd0(0i^
zu=t0@#MS#Yz3vx(e!p({(e3v7KR%s4+caNSKRsEf;NQ!llhfy)-=~+iFXrQB?MX6B
zc4wc<7M%17SJZ9}P<p|;`mUI*c-V~J3z#RbTNNv?({Z0${>2S@F1i`!F3l9#_u${@
ze5twB58r>6%Ih=UA$sh>WWVqX<+T^IVjBc4XHD34&wBnn=`Y_mUMZVuR(>(~Qy%wo
z?TLDstCd+ySl6y?o>O+qZ9#EHbJ;fG)8+5&_O!dK*=sVvq1Wriil=-<(|qi9y~*h2
zIuzy}*>uRsByG7{;@iM3j}1<lE#ed3waUJb;$pgTFI3NN#bN755~Zyr$}aa$yr_Ah
z_^EHt$}a`{iec81S5A9uJW2a@!-jMB!q;(LczIXZExN_4;?(iDALdo;(hs)=L<wnS
zYX4c~`SOO~suB*73Jnt}R?TZs629W>7CR+k_pC2e-~L3hRYkDLJt=&-K|@@mJLB7w
ztpQp4`53tt`s#m>xRCELIn2dNnD;*O#*fqfgaiaubwr(dQE_MC!jgHP_?N2-$Sw)X
zU7GSe@(!Qs){WV(w~1{#bo)uH`|jgS8Umk|X<Tp?KEhSKfYZff>gElrx+~<*s;eu7
zZC%*)vL#w1`^J$gx|$x=`-9G8h<{kmwswh7(XOkTiW|GGZLjd2CwB4mRPoZ8Atux2
zWZwT>v$E`c;0dEhCDE1K`POcd0T0DbpRWA9XF<Bx)x5Mx)2?JskXFCW^7W0Pc#-c_
z&nG?WO6JI>J}_F5>>1{vntIN3Ucl17J*EejD;888TH*ettJmMnU0G`7l>L#XZs$wh
zN!6d7{x^p+Zqt>|=ZeFVueZnC7GPzau=Z7($qknstVMH$RQ~Ww*Uq(ER<Kz0le5Hn
zwL5-4st&!M@^Di7q^uqu@#B-#>!)t5_EwI(F1eT~_1;OYI$7ENzKC_Y+NHXwX8VP*
zV*>13-$&J5a69=+Q^Z|wvDx$dwN{f~zbu*2`f`i0@!W66mtB5yR##l)ie|F@pEjeN
z%kHmO@?>w^MFWQ1ucdtQUTLnye%F_H{`L8tVc@&(-`VX}5B{CV*5Us*(}Sn#;(qfd
zi(ggu{d=Cn_V{7Iw0o0o|IZBjx^9c|>J)9wm%?v-4sbl?<TVt(-G9{CqJHL_%l(3C
zFOwbRQcZ4ds%N;A;8V#^T$!6>2h7==+txU`Nh}lNVek)N2v=rRjsz1x88A5V<gQH>
zPf9;GQy@@UoJ~9oOu(dzQqzi`F(OQ1<O1vNJ@k9Qnupd3T)!6FYxXt`X5k5!SC-d<
z8O9T?psawz)3(-jgS#VK$65!@b8LUZ@bgQ_Ek>4bBMzeom?C$Nq-8A}H)I%j!X;QF
zlwjt1ZLj(q&l>X=D6FEOVg*wj;wPm3VxkQzNSITFA10iuy>oiBM2H^~Pq?D8qBhJB
zo^V}jU1zwW@_=N!=ld*yie!{!bP$R(tu<X>iVlU?y^NQ?ejI3l0gHhWOc<t^i3=QN
zut4EK@yb&J)-UY*;@u#}iK@uM^!^nH+jKYo$TKdW50xd<U`FwTODanu@uZZcka*I{
z(nvg6WmzPioU$AePu*JG4rUckxQ4X`5>Lxo3yG&^tq127rRF4pV!pU2xi}ca023ev
z149^u0+B&*CUf|cf8YIQPcUF(iwFu12@7Qify*(3BA5{fW^!&yP6m+co#xBT!pg?Z
z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%RzXooMO9r>
eOGn?p(Ad;m&O}Pr$VF7l-7AvG&Ck(;(FXvE-?(`I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp.03F037ABB1A17197.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp.03F037ABB1A17197.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1a849d12526814e5e41f14bd747fafdba3a4b5e0
GIT binary patch
literal 1794
zcmZuw3pCSv9RH7bZIefhd2TWO8nKpF#u%CAnHgINc}9#fHFY0IE|MCCo1Bok;<}YA
za;G|0_ekBmlaS<EiaN@B#J%I5$vx-(&iQ|T-_Pg!`Fwwm^ZV1N6v~o30Jua@lUb2T
zp-=z-)FdA}DKVx)767&(03f;ac>pV6&|0;x>{Zq1pZ#skeNH57e6xJ4dSz_hMX1*m
zGvT)CJYLAsB!^dP>)Q<RDhW+B%v<1cI_~4Noa9HQhiogHk`I(x%zVn+n%C0$0sVDi
zesp>U`LI;E)uCk>Tz=KeIc)8yHwXQz{>n)8xyiv>)MLwECZ-3c_>1%7$g-5C%U^QD
zb7C&EH0AdzGve`_TlE49o3EokVOvvudSc**>V<_*&S7|pr=Q~*$sHNUD9xn9-Gznl
zuu$%;&OpmJ-`L8Z8@hX8A;<m}+2jT%zvU<>bE>q4;NkD7RUZ7jfat>ZmMpmvZo@FK
zY&dpSz>F4nUS`emnwN*fv}+y%6N(To@7T%1WFpJ3a}S?P96ToyaL~gp27QMnsO8Q^
zalMHudpwNMK_~gHa-o%#W=^5zN9!4tx2H{9vCIlfVa~n<p?;E;90WJzu8Hn_TVzWK
zGI*`n^^#zRw##ZZOmKhw#wANM98n`$JU`amhpDX>biaTcysz*$I%hx3M{KM3u#$6I
zBN%?&s3N~XXfesV;7~!!bV|!SRO!nN`z@m@sV=i1r|XZPjPII^I!v(Hs`lyvTn1#e
z+YT0O`s-O*<_SeEoMo5Rn);(x3EzNL61})3>u%#%)tUTmmtGnr^O1&f#tbh7RYp$Y
z9=w~E|N6*rVo~qF&Dud`zhXiGXXmFaa054H@2Ly=6PTiY$O+8V>e38<)?-05X^&c6
zEyCZ^bf}!!V)sitl{~Oq7)BAKro4zYD&+R$7Tt%_`43ujrukXH7WZ$QyLPo<%yWh`
z=3({SMQ4pR_3C{=$!`cxQ(=x}AwHMKkq&+}^ScZ9*MieCjwibAZ`IduzInqgyR_^2
zg>g{p9-@ZHMbfC3*mpsCOiKI2NRKM2Y}bUEAmUMdOIs=hCz`SseQHEByn6QTfKA|;
zVzasCnIZrF`Ry;Q_mUwI&uWa7t<~aFGrX15S2J@TwCnoa>r)r+I$vxU@AkdbXkg^M
zv5InS!$;!9k8L5NbhrGn?ze>wOuZ*HPf<wI8SkRclOsvTgc<{PoEfyJ9-TU&)?C0M
zLqlc2<amNNXM}JxWwj`Z&{X2Nbg9JiU_k+u6Gvzp=DB*OmEq@w3FZZ+OUf|!&y&cx
z^e%er<kz~h#g>{UTuXWb2i($QXeSdMbU{?yg2S5@I^4oKvsata(EcL!tmBa$9kBJz
z?Cz$OkRyTP0lP<G4ZM8PR6F!MUbm~0T_^6eKRaLCWwe3{%_muo1yKs{%Q|0j!@~gf
zDOMOs@_#?I!hQ);=9AzuzWV`&ro3j@dXQAs$3ndG(Be`@B?9n~lnvz!BiDnpoERC!
zo`L`XY^i)W-NVWcieR8YG(p-F!N7o+O~{mJO4;bhFe936AadqMKs=i?0D&?Hii8NM
z$i4COOo-f}6$pZX#bE8EG~HVre4l48w?->rRire_w5d1{Mfa9LFt&r+ExzqyCvm5X
zRK$BZ$Af4s0TM3?iZYgpN`FYPJ~+8lE%EAtx_Bv#VC*37*emr4>7)-K=$EY$FCL0F
zkkV4I3~c@Ga)ry~wQq?S#D9(=!pqGnsA%S_?HP$hOTE$=0$aaWsdq&+;LhuW$M0kj
z3~f*wCtU%-&<FK5A#<WRN!npmq*U&c9NeTJIT(oAB+>%4HX$8QXA{x`^)?}p2yV1P
zFieOhn-GqO+dx2KWL)Ul8(=4fv;EeR^{|!zK)}D#TBLnT>EBZSjZ~H~Xw~ih5ueSL
zXVU!W0sa7UqXvYtBI2SY&?CxA7AhyNpr`~#D66V$p{lvM7^}gk&gz<5x1qFb_YmzJ
zNRCeCNK-pQYZLVUu^KwMdYJ77SR-%;&ddV8Yqyn+zNPj~!d|3?oBIJ7iZ^*5#Pc6Z
CPyP7-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp.237A2B88B920F202.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp.237A2B88B920F202.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f92a5e304d63fe677f95353cc386cd07bb41e5b
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;=bD3^O5Pr}Y@x;36Z=_w=jj|0Svhe^{IA#LQ`T4n?qT$umDr#>@%x1G^Lw=<
z9oQaxx?{1>p(1jQ{^z&t!V?c3e<}ax)wf5VK8x<3!+YHJ;BWQcH}{k$nAyvIWqG`L
z-h-80uWzsSEcpHL)u-F<xBU3_Ty%c@&e<Qke}4Trfn`qpu6dt-KJQ+?dAGj#pGR#f
z#~b`FXL84EJ~Z`2m%<c|s<T?#9nZIlriqs%T$(zSJ1KF(`U}w}If2=WXKcH4g)6?`
z|5NkcxQd76?7Zd6CGT)2zwlh{y=7up#?*5MJ6LB;*mlo){=MukyCX7o`^?V$GG)*2
zJ3oV#c;9-(!RYIk{W`(XI#>CE<rYWl+g(q0m&@OCR=O7R%5#B}YR;7+eoJpP`Pi~8
zuB?fR7wb4A>Sb*XU%cUyhO0`%LX$(?9^Vhim6)<JYQ6Uom%s9GnuS}e(p`>=*Nv>?
ztvdE7=X>3~@7(e9UdpPcKYE^Ax~8yZ^X^Mm8Kd^HAIazr;PU@z!LfgF`3A8#?xm^=
zCv7{V|9VQF?ZTjThsO4p#eox&*0x?O+1Nh0`{L2q*6N+A=QUJ11#E<NYqc_dxn#wa
zp}o2#sJ)b90(Y)e<MD<%-&tFff@Wp>;Op4B{mJsCNtc|j1i7^z7UOOGJf&V@zGKOR
zsI3zx-AymFo_6Kdx;3{WZgAvJF4uj_$F1ZZaiy|P?v=u{8;qq2i&oBQ4b45WJHyUW
zaN`=DP#^6(3nC6$Ez9$Cy!Cr(lH2Zv`l}hLQzu;Rowk=dWXryT(x0|;#Od!{(Y0jJ
zwkJNH?Vfm>R`)xJofKN>5nFNJ=u`uj{rq_U+0R97gI9esp1fsATytw-k!!$lIWebo
zE37)_TwL1sTwQVp&#H(CqCuWBi?^zWWY(X1_29CA#C*e7g1b`=Z$5ak!A<Mwdz+_k
zxASbfdHPwTd34isov(X7-wD0(P4T*!iz8#o=~ZgHW`cPQt$v(B|Jr-^os$XPaZTu_
zVb65sJog&YN71P}o-AA-rF{73k2fddABDJ2ntv%Vn?H<w(<&pkqo?*vR{djTAv<S7
z>$Wu)rcSx}#cfT2+g{OsQ;#;AdEJ?PS$0Ryt@-<x&h%NQHo0z@+T?3bQZC8P_Ss~f
zIx+Zxde7buvrj+is%vf%`cvilH^JoUEvtH+hpJ~?V<)P8o%Zcw!mKHGXXlq3I6U!s
z#Jj(rCrXqos?U4i9{N+}@A;cwj-A}7^VcWu|Mi>Kc3$n-(WJX{$@$&JjLJOjCSPK5
z>`eF<{!K0Mw0E1agrB!^`bCv9|Jc+XB^NRjSLP<!0rU6qv<jV9QnwO#82kem!j)N-
zQ^5pK1`KX>ByPJr<w`s^Qy@@UoJ~9lOu(dzQqzjxGcqu6{aSFZ+1oUjg(qBIUET<$
zizi$`T>*)wZLRGFw=i7CS_jT^Y=6V>^GnGsMwW0R4x<#9BCqXLpW|6${xb4}t0<^g
z!GuHngw$V5v|$AabE@#egrhEe5Pg1r_cEZc1dD_k%oVxXJEuoWg!nP>ge$5m>cfP2
z!gZ~6ond<C)Yh9tz5cZls7OX##sHy6(^}I7rsz<J-OG6S>&Jl>7_b;9!GvKnBNsUA
zdJp|xu;!t)0?Z61E^wH^gn3YW_|$;)3p>AfH^?iZD)KO+{tAR`x|@IG88c6~l)97_
z%qX64IdwTCo}RUy9Zbu=r<@D?E8=#t1AVV9fuu-MT@r~WtuBqklU0{R;;CD!!;Rw!
z*Ra+=;%QlH!FdctsX2+DG*DcWTpSEyfC&(TfjRuE(YGU0maTrt77-L25*Er31{Mbs
zK?r6Df*FcnMj)8UxhXjrK(2S1FEa}(8#@OlHxI7>KcA<dtFxLQm#34ExUh(Xq=lum
zjjf%%gPy3im9m+JJT_4=X&G4sMI{wgbxkcDeFH;dQ*${JDP1EMQ89O~NG3NwM-N6H
E0HcHbHvj+t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp.177CE21E4E683B24.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp.177CE21E4E683B24.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1a2c7da533fd7c39c3b7269ef7e28448e1c4d5e7
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=nHQKD7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;=bD3^O5Pr}Y@x;36Z=_w=jj|0Svhe^{IA#LQ`T4n?qT$umDr#>@%x1G^Lw=<
z9oQaxx?{1>p(1jQ{^z&t!V?c3e<}ax)wf5VK8x<3!+YHJ;BWQcH}{k$nAyvIWqG`L
z-h-80uWzsSEcpHL)u-F<xBU3_Ty%c@&e<Qke}4Trfn`qpu6dt-KJQ+?dAGj#pGR#f
z#~b`FXL84EJ~Z`2m%<c|s<T?#9nZIlriqs%T$(zSJ1KF(`U}w}If2=WXKcH4g)6?`
z|5NkcxQd76?7Zd6CGT)2zwlh{y=7up#?*5MJ6LB;*mlo){=MukyCX7o`^?V$GG)*2
zJ3oV#c;9-(!RYIk{W`(XI#>CE<rYWl+g(q0m&@OCR=O7R%5#B}YR;7+eoJpP`Pi~8
zuB?fR7wb4A>Sb*XU%cUyhO0`%LX$(?9^Vhim6)<JYQ6Uom%s9GnuS}e(p`>=*Nv>?
ztvdE7=X>3~@7(e9UdpPcKYE^Ax~8yZ^X^Mm8Kd^HAIazr;PU@z!LfgF`3A8#?xm^=
zCv7{V|9VQF?ZTjThsO4p#eox&*0x?O+1Nh0`{L2q*6N+A=QUJ11#E<NYqc_dxn#wa
zp}o2#sJ)b90(Y)e<MD<%-&tFff@Wp>;Op4B{mJsCNtc|j1i7^z7UOOGJf&V@zGKOR
zsI3zx-AymFo_6Kdx;3{WZgAvJF4uj_$F1ZZaiy|P?v=u{8;qq2i&oBQ4b45WJHyUW
zaN`=DP#^6(3nC6$Ez9$Cy!Cr(lH2Zv`l}hLQzu;Rowk=dWXryT(x0|;#Od!{(Y0jJ
zwkJNH?Vfm>R`)xJofKN>5nFNJ=u`uj{rq_U+0R97gI9esp1fsATytw-k!!$lIWebo
zE37)_TwL1sTwQVp&#H(CqCuWBi?^zWWY(X1_29CA#C*e7g1b`=Z$5ak!A<Mwdz+_k
zxASbfdHPwTd34isov(X7-wD0(P4T*!iz8#o=~ZgHW`cPQt$v(B|Jr-^os$XPaZTu_
zVb65sJog&YN71P}o-AA-rF{73k2fddABDJ2n%`Hka})b&=EzXV#Yw66P72k@&f)XB
z;S?SA!s}E{m2#NH;(J{GPd##!R(*H&C0~r&+w=D?`KX0Y_OuV4?3rC;^s>)a?dI%F
zCp13@AB+2OHnqrgzoR2(-QGoY2D7qq=h{#EFzMN%bsm$eynkmH_?~)q*4*mB!xPy$
z-~ave;IX=BZ~kHNs+zuk&vU9CKZ%%Dr?&n7>zvnjvfN@0P7Av9{O-&K!8T>7Dm^p#
zS?w488SPpkuk3TFy=~9w4574oW>Kx=oZ`ydBs*aKj<>XqZ3^&h<z@;DU<emy6HfvY
zKp8N|c%p56(t7(T9tMA)v@)x5Dwu#t7p0~Zzh`7%;QF=TUbDApFbhw(yt=#*%ut?i
z1$6}^p0>5N8{EWj9cvvp&$0at!_O}zw-{N%jW~=_V2Zr9SAC9Wjrq&S6Rx75Vg(Zp
z@e@*iG0}z<B+RM84-<x2%ftotL+_#A3)VceR^UM~5~i4u3+xk^qFn8r)1xIq{Fr#c
z71b5>VMg(U>ssqN!}QFltv8E${c9yqk&L>G0YZ_cwWbS9(V-B#m+|t~j{_ZNz+#{T
z6MkyI`h}feyc^_ZQ5AWZ@OL@pzq2%txB!JESR~Y7M*bBD+jKYo$TMc1a4B^uEtoJ*
zxSYBi5>L-s&kknNzo(oF{43&ivI9+0mq1b^sV<4clUA2T;>oJZBJtF%)#1kRglkx9
zAn~-UwctF4qSTy3P#P#MN-hot(#+u<_2<|gCtN(w77-L25*Eq;5`}^=7=wW!2*C_N
zFhdc{2m~`ZHzg+n$n{S1WoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z
z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL$>iqe=)vd%006n!
Al>h($

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp.FE8F7C3907743A26.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp.FE8F7C3907743A26.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8fa891ddf496febc262bf46e2d1a1263a2838f24
GIT binary patch
literal 1794
zcmWIYbaVT~#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg}q7({p&7%JwR4a?2HZ6<JS{>|l&
z&fGY;=bD3^O5Pr}Y@x;36Z=_w=jj|0Svhe^{IA#LQ`T4n?qT$umDr#>@%x1G^Lw=<
z9oQaxx?{1>p(1jQ{^z&t!V?c3e<}ax)wf5VK8x<3!+YHJ;BWQcH}{k$nAyvIWqG`L
z-h-80uWzsSEcpHL)u-F<xBU3_Ty%c@&e<Qke}4Trfn`qpu6dt-KJQ+?dAGj#pGR#f
z#~b`FXL84EJ~Z`2m%<c|s<T?#9nZIlriqs%T$(zSJ1KF(`U}w}If2=WXKcH4g)6?`
z|5NkcxQd76?7Zd6CGT)2zwlh{y=7up#?*5MJ6LB;*mlo){=MukyCX7o`^?V$GG)*2
zJ3oV#c;9-(!RYIk{W`(XI#>CE<rYWl+g(q0m&@OCR=O7R%5#B}YR;7+eoJpP`Pi~8
zuB?fR7wb4A>Sb*XU%cUyhO0`%LX$(?9^Vhim6)<JYQ6Uom%s9GnuS}e(p`>=*Nv>?
ztvdE7=X>3~@7(e9UdpPcKYE^Ax~8yZ^X^Mm8Kd^HAIazr;PU@z!LfgF`3A8#?xm^=
zCv7{V|9VQF?ZTjThsO4p#eox&*0x?O+1Nh0`{L2q*6N+A=QUJ11#E<NYqc_dxn#wa
zp}o2#sJ)b90(Y)e<MD<%-&tFff@Wp>;Op4B{mJsCNtc|j1i7^z7UOOGJf&V@zGKOR
zsI3zx-AymFo_6Kdx;3{WZgAvJF4uj_$F1ZZaiy|P?v=u{8;qq2i&oBQ4b45WJHyUW
zaN`=DP#^6(3nC6$Ez9$Cy!Cr(lH2Zv`l}hLQzu;Rowk=dWXryT(x0|;#Od!{(Y0jJ
zwkJNH?Vfm>R`)xJofKN>5nFNJ=u`uj{rq_U+0R97gI9esp1fsATytw-k!!$lIWebo
zE37)_TwL1sTwQVp&#H(CqCuWBi?^zWWY(X1_29CA#C*e7g1b`=Z$5ak!A<Mwdz+_k
zxASbfdHPwTd34isov(X7-wD0(P4T*!iz8#o=~ZgHW`cPQt$v(B|Jr-^os$XPaZTu_
zVb65sJog&YN71P}o-AA-rF{73k2fddABDJ2n%`HE`S#!|hp^CND<Wo=Z)%BWU*0Ud
z?ZPzCU6YJv-|G-Pw_^LD{hKN-xOG++?_!=F`1|vAlgTHqrd&FICFRnrIX=6RCwqQd
zrc=cAr)8sfUGdDAi}?WoP4Vkj$oDLZihka&x@XDG7221U+`D8a<uiFx_2*u`3X4xt
zr}x&|2{rqDInVoNh1UGV|1D?Vv#^}DbpFY=|E+FYZJ#9=zG3N<FO}tL2`vvh+U9ki
zZGWb?Wq*&}%i|oAzbsUIZ#1jr=6$b%<Z6cE%G@M7VE!(<pB#VXxbzhs2LAwtaAj8I
zR4@UQ0fR-H4-S0d+o8?P6bO_SXA@5X6ENwb)U@LFj0_B1zZTqU_BIV>;R%;lmp6jx
z;t5w!S3u%vTWh<)EezMO)`9aJ+utz!{8Dm@ktN)S!zcx&$ZLDm=Xlnbzl=QLDheuA
zFyRnCA@vs%ZCF9VoGSb<;au&V)1xIq{Fr#c71b5>5w6g+)^&y{g3*jzV2}15`n_Py
zLu-Y$a~~!#YWr;h8Y#gdp$5}Cr?%cK>h-UcKqF<;Wei}#JmH$wnl3P-V4_T1;J|?y
z$%EpRLm_rA<K?d(2U=jjVxR=m`_zE-3p>AfH^?iZD)KPlzXD;K?&cqP#>^8gr7oof
zvw<gEPF)U(r)RBa2h;NJDdz(JinyKZKp&_}ASsemmqg-8t4kyCWYuMnc<R>baN~Hw
zHLNv|cv{w4a2`WZYEB|34HOq87YBnFU;@NoU=ClIx;oz1#oCE2A}BZ{ER-P(EDk1u
z5X=w+GZev$KroYYQ*ttZT<<hrW)@a9b`DN%9$o={K2JecXEi}CPbVR9VG#*Q3rlMo
zTRVFPJyC5dWit(VY@%Y)GO`MaN-C=Anp!&g28PC_=5i)dx<)RdV(wm%Om2RT9*jN!
DGeh@7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp.1FB1A94EA79EE13A.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp.1FB1A94EA79EE13A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21d5975b5b2f14967e2d6964ef18fedf01b0f56e
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow@OG)3pYBm6dzhvV|6BPwZ#$ou}g@vU1{-xo@t&*N7}!l)^e|$zcX}rT2>O=T(Oq
zHSipGy0g%!@x!JW`JZm{yDK&KKd=9A^y<r-&spuG`~2COzW0ASef7bCqEZ`44!>p0
z2a`p`!{aAEc=h<vo7?ZV{P^~~Y5l$!{hu#iCE2MlKA*1_R`d1em)D<uomREqt8%jV
z!1B+Z`F13<uAU(3s3G`e+SR=FdEDA(VlOm$hF{g|*y<R!XnV$uT^HR9bC+g{?0fL<
zbiUNyZ-?&7n#9Pb7by2ERG+)<f?L$WP%C!jYZ*?rs@~UB<zKvi)akbGtvhc1CwHHm
zyWVBjlC^yT7oKd{#&>PYWizI^&eB^p&yCnu{r#b>O48e<jVxIqhun0(PuQID>0qjR
zbcjHi)hQ;CXNwH0O}VoBgIt<Ky_ENKl<2>4ydEI1;M=;Qk8!g1gpRM!ULvjd%gyb%
zGylZ$C%#MDt2s8_n(ebyX8wdbH`rQAfeL*a{N6Jf-ckxV6l}+LX#Yj)8(iyKFHH*Y
zjDBeD>veoyz*6T02b|Yk)bQAN?NG)pivuT*zIYVdTfI~DyoO4rfQ`Uzt5(J@$E>(A
z)K|9#wU=^C;LfdTJl;^}J!^|n(5#Ffavi<fpDb_cy5xK%$nE%JKHk>PQ|c{dI+aX_
z+B$X8UG>6hy_LPXYi~#1;LM*~F8fxFTgg4bys}SjmV((0W?zLxE9bO^#vb{dVP_?{
zaTRZ<kNTa38i~@u+m|T3{p)48IQGDPsV$SdPP_~|eUDWu>i#3$$|#rh=5beCgDyrC
zEvuFPwDi_qcg60M&X6T~dtOSUI<(GTuXpbD^QL*4p}%K(Mg^@uc*`PE=*s5ytrJv3
z)h8P7x>Q&<rSbHFKGsOZsZ%VkS2#`mR(y?Re%pbQQv&;M@Jjo;yDLkroU*_2)a`u9
zJ8Al})Bomh#%;Rt`CM^$^7ZzZ+XAet6LMd*ncQ&M!FtU~)#Hb}*)OYNpNC=IPtGUq
znN?u@@M+6^t;4GPtXCx@&-pLw|99$=oK9~_Y<dsZluC8&AEhNWR!P~pkqfgmH-GjH
zd^G8oX3h0ad(FJA`7g96lH6)ppM7<vPWt9Vv9!(S&g5K^pM7?d`Sv-%0qH$^KS-ZS
z@!TKic(Qu$qC|#UuU^a4>CF^9YZ*IP?W^(I#RpA({xZM+vY}mh{YIM~=}Lw#T>syj
zz`yFf=CApiC41VRZ~8Lr&OiUntIC(kMzY@a@SLBK&g3THpV6`NMxD;P!#__y>8a~6
z@RaBNq_Ozqw7<+Zt9_0z6j$aZ*#UEQVC934-oDD`xS0Y27{bNb#KXV@PzDTMNtaH!
z8lr2!!{85;R%TU>1QRgnqSUnFXN(LCT)!6FYxXt`X5k5!SC=<}8Ojr`pss+#)3(-j
zgPRzxW32<{IkvxH`1z&e79&fz5r<I(Op(|2s?YJPF@G6(!c`PhtYE@0^O?B7j_p14
zd%>EA)(Si*dPDq#)L%@rVFl^sRN;r|&DGvHJz65fkBKK-QC(3VW)x4juC=Z+Owp9<
zU#Boi{$c>?m0*!jgDIL*TW=Qi`qxUJUKw>61DG&RxTdwH3rz2!5WAQ0^4E_8^%}4k
zD8YoE8nAw0=NIn=IY(4Q9wrQ<8M(j#3-ijqr<@D?E8=!?0Ue+&p@pzXQe6^>C#5ch
z#FJK+M&ik;%OdgQ)a8(P>elLZFq?V8HLNv|cv{w4NIX4jJvgr@H75}i^TkEU#lav3
zm;f;t7{VYFkYo;j%bY#u@6FhcY!N}hAz`5mL2x;S5Ck(6!HhsKlXFvYGJss~G+$;G
zRyKAHPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fTdj~yHZ7XFn4S8&$V$w3Q3W`c9
hs_L3rI{F5N#-`?SCQ`aaE}~-YUXe_0evTfDJ^-M9)vEvi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp.6BFB81AFEEB1C1B7.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp.6BFB81AFEEB1C1B7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a14b620c85634329659e61b84691e11171dc8904
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x&$R}7mApM{*+Pr6C-$@W&eJ(0vU1{-_;0Vvr>wCE+{EO2DX~F$;`a&V=l5z!
zI<P(XbjM<$Lq()d_|JETJ5M}#{D!}-vbgT!&q?uT+8@t*@OSawnv2#3@pJmBnm$JQ
ze~37G_3U+(9X~57KfW!Gs`>qSlK#Fs=YIZtlx3yD_<a64v6_!R|GfV5?X;==ezlXu
z2bSMl!@4fwp;wBlfEUZ&#HqIzJwL>?ncvFb<*8Gw8%<7Jf1xY2ZOzs|pXki2R{b6S
zKh1X2-}~YB;<nqDdET`cepzx^^_GYB7O&?CM}&MmqTkQ;f6x1?R%c7>(zAM1UiV_(
z)o5N)%`I(dP+PvWbi<Ol+XP?8+*&X<_voj)yZPTM3TEqlRS8%qWR_LgFh_MVzn=B2
zMNJ#728t~(5Z`LKA~2$IijvTp0I7$rOR66-?3&fYFtvPX_xD!??lOz@Cg!!g$d;VT
zFY9t|qWRKy?-h@HN-(-wRO3{Xnk{hcX6?&Y3~TQxCv9=nXkA_-(^|i{e1l*d_fpk`
zleQgN|7uDf^TMEZhsO4p#eox&*0x?O+1Nh0`{L2sR`#8$=QUJ11#AR%Yqc_dIc3F_
zp}o2#=y)l|1nyj`#^Vij&a<{C1<lI%!Pl{M`;+BOlP)=5336*cEXLdVc}l&+e8-Xr
zQClZYx~pDTt+#S(Z1nc68$|D^y^p#jE9TO-Va;b5+bfQ1HZZ?(a0#~LUVW>h{9;Xs
zQc|Sksu^KBTsOA)`rMt=RPsUlXipjA|5X>gv=o<1>#~cl$f#{Lf08M<H?A^J)YCQh
z<jiNaXQpR;lWWzPq#87F*N1hfQy4|-_U)KudfvNgS%_?!TBhe-Ht)wzW-aKm)oG0h
zE)}v|>@B-oU2+G{s)z}qL7p><x2lI^+NYL&$mHlbFY&b_cH_gF4^kYIr+zZ`t^BRq
z9Q$ok<+ZcZ9;kZTzPBkC-Bx>ID{o)`!_8HzJez%6q!U&KH?{mvT)f`+xJcd>mwz(H
zP77{V-Y55Qu2IaVfPmhK4{K_^y~xjfd-0M(Sm?184l~O)wZyY8Zx-HmWt!-&Nk+3{
zyG_rnoPVfX_kQ-`l)oZf%+mvBf8HK8_vF`<OXm-zT$(l0XIJuM&u`DB6?466*(hG;
zJTvBUen7ydxb-XK8<zcwYM<|&=2`3)zQpsd&+iNa-+k}SZnt{y@kF-H_kS}zc&sk&
zH-FOns=Djn^PH;34|nS9Ro(vo@|ssMnQmbR!-6iU-<sJV*rqI%q-Q4I{k!0w(XIvZ
z%08ElH`$!Z*p&L8fj7hFGDC4?Zjv1^XX{>U{j={IV?Gaqe*i<cGOKbVm;lOv!HjV3
z^ae*8XKp575QU4giHCs+m~>HUTJbYR1_rKQ3+^?0n+CJ+gv+bT8^Lt(ge#~kAn~-V
zwcX$rhU-}Cz<DrHMlP`Ry@!4;So6?Y!Lj`f!_O}zw-{N%jW~=VV5+>fSAC9Wjrq&S
z6Rx75Vg(Zp@e@*iG0}z<B+RM84-<~M@ImzX`Q6Ka!V)YJYB0y;YVVvLEfM0!#1pQl
zuBZ<a<_Xue)^&#Ig^4n8f&C407Y~Ya=G4}kMZNyD5@?c)x{Lu#4^OzJwWbTqq(dQg
zFXQE}9|yX^fW<%wCj8WZ^$R<{csIz+qAKz*;eSs#7x-7i?c@TwSzSU4VWgzGBoa?b
zT?&aOtuBqklU0{R;>oGYA@S6$)$L$b@q}wwYasEothJDMde(YyUQudJA}HpIi;|0j
zK@2bfVlXg-K`0OzgkXlinattaN=~TkO%j>Q77-L25*Er33YTMuKroYYQ*ttZT<<hr
zW)@a9b`DN%9$o={K2JecXEi}CPbVR9VG#*Q3rlMoTRVFPJyC5dWit(VY@%Y)GO`Ma
jN-C=Anp!&g28PC_=5i)dx<)RdV(wm%Om2RT9*jN!lJnNi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp.D3FA995D6DF314A6.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp.D3FA995D6DF314A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..07640e52fe382b4696693fb68f7fc9b283a55810
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x&$R}7mApM{*+Pr6C-$@W&eJ(0vU1{-_;0Vvr>wCE+{EO2DX~F$;`a&V=l5z!
zI<P(XbjM<$Lq()d_|JETJ5M}#{D!}-vbgT!&q?uT+8@t*@OSawnv2#3@pJmBnm$JQ
ze~37G_3U+(9X~57KfW!Gs`>qSlK#Fs=YIZtlx3yD_<a64v6_!R|GfV5?X;==ezlXu
z2bSMl!@4fwp;wBlfEUZ&#HqIzJwL>?ncvFb<*8Gw8%<7Jf1xY2ZOzs|pXki2R{b6S
zKh1X2-}~YB;<nqDdET`cepzx^^_GYB7O&?CM}&MmqTkQ;f6x1?R%c7>(zAM1UiV_(
z)o5N)%`I(dP+PvWbi<Ol+XP?8+*&X<_voj)yZPTM3TEqlRS8%qWR_LgFh_MVzn=B2
zMNJ#728t~(5Z`LKA~2$IijvTp0I7$rOR66-?3&fYFtvPX_xD!??lOz@Cg!!g$d;VT
zFY9t|qWRKy?-h@HN-(-wRO3{Xnk{hcX6?&Y3~TQxCv9=nXkA_-(^|i{e1l*d_fpk`
zleQgN|7uDf^TMEZhsO4p#eox&*0x?O+1Nh0`{L2sR`#8$=QUJ11#AR%Yqc_dIc3F_
zp}o2#=y)l|1nyj`#^Vij&a<{C1<lI%!Pl{M`;+BOlP)=5336*cEXLdVc}l&+e8-Xr
zQClZYx~pDTt+#S(Z1nc68$|D^y^p#jE9TO-Va;b5+bfQ1HZZ?(a0#~LUVW>h{9;Xs
zQc|Sksu^KBTsOA)`rMt=RPsUlXipjA|5X>gv=o<1>#~cl$f#{Lf08M<H?A^J)YCQh
z<jiNaXQpR;lWWzPq#87F*N1hfQy4|-_U)KudfvNgS%_?!TBhe-Ht)wzW-aKm)oG0h
zE)}v|>@B-oU2+G{s)z}qL7p><x2lI^+NYL&$mHlbFY&b_cH_gF4^kYIr+zZ`t^BRq
z9Q$ok<+ZcZ9;kZTzPBkC-Bx>ID{o)`!_8HzJez%6q!U&KH?{mvT)f`+xJcd>mwz(H
zP77{V-Y55Qu2IaVfPmhK4{K_^y;vV}d-2l7HLH388;#zlaPH^xV^_~vsjFS0nrgON
zEPGC{ee3(Eeb?Pi{?Zg-*IR7%Jb$g#<ku&c%x^up#MF3Z$#J#GZ_esI7rDZjtpBIY
zXy>x~3l==t8y9HL;QMtepZ)YRlb+34=P~)$%x@PHX4TxConLz3@x<#9@BSK3lqgwL
zpMQe=>UWX9=Wl*FcDOjQ$~*7>^2n<@FZHZxTDxS4{}y9LWgd5vBQZI4V&5PBNiA`-
zcbl=KpS5zz#gx<k82?E6oMR}i%uTWb=4^M*ilnvOivM^R`~w)mm06V|!30nS3|_h^
zC!P0r#>34745DyxHt{ep0h2CDO)GxJ$iTq$Yr(x{Z_{8Fo^W|}c_WxEo^S<q1tgxf
zwYD4F!f+jH9XQXi{SCv<FD17aS;CDtj3Qu)U^F8a*j2rUelJ+_&|1N3d)4Q7)|kJH
zJmD$|DpoL6A$~&YFDBZsf`mC$_+i4i+B>I5ON96_@q{a?E9xU0r)#b23{y0xw%#o2
z^{<scMKbC#1_(u()|xIbMTbJ{UdGE`KMu6OfW<%wCj8WZ^$R<{csIyiQ5AWZFw6`l
zE^wH^9M6N|!|!s;e`jePaRKU;V3AOR>G}7RbAf+F+)gf_u)2g6OqeHJQe6^>C#5ch
z#FJK+M&ik;%OdgQ)a8(P>elLZFlX|FYglU_@wBY9ka&95dT?G*YEB|3=8KDxi-SQ7
zFacsPFoZ!U5E+DEhQOK3;o9eRf0FujFp4cAC^#f6lpz!@#}I*FCg-N)WB|F|X}-)X
ztZeKYoZLLT0{ncQg09YLf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_J
hRMj=Lbo31jjZMwvOr&&;Ttvm(y&{?1{2V<PeE_-_)V}}#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp.D00F48772AF74F17.idx b/.cache/clangd/index/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp.D00F48772AF74F17.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2e806c0d53de569bc332767a4ddb8d1cccd02bff
GIT binary patch
literal 1778
zcmWIYbaQ*f#=zj5;#rZKT9U}Zz`!5~#Kk2=nJ1YU7;bPgFjUMr8<v}Y+f3lv{F}=k
zow;#x&$R}7mApM{*+Pr6C-$@W&eJ(0vU1{-_;0Vvr>wCE+{EO2DX~F$;`a&V=l5z!
zI<P(XbjM<$Lq()d_|JETJ5M}#{D!}-vbgT!&q?uT+8@t*@OSawnv2#3@pJmBnm$JQ
ze~37G_3U+(9X~57KfW!Gs`>qSlK#Fs=YIZtlx3yD_<a64v6_!R|GfV5?X;==ezlXu
z2bSMl!@4fwp;wBlfEUZ&#HqIzJwL>?ncvFb<*8Gw8%<7Jf1xY2ZOzs|pXki2R{b6S
zKh1X2-}~YB;<nqDdET`cepzx^^_GYB7O&?CM}&MmqTkQ;f6x1?R%c7>(zAM1UiV_(
z)o5N)%`I(dP+PvWbi<Ol+XP?8+*&X<_voj)yZPTM3TEqlRS8%qWR_LgFh_MVzn=B2
zMNJ#728t~(5Z`LKA~2$IijvTp0I7$rOR66-?3&fYFtvPX_xD!??lOz@Cg!!g$d;VT
zFY9t|qWRKy?-h@HN-(-wRO3{Xnk{hcX6?&Y3~TQxCv9=nXkA_-(^|i{e1l*d_fpk`
zleQgN|7uDf^TMEZhsO4p#eox&*0x?O+1Nh0`{L2sR`#8$=QUJ11#AR%Yqc_dIc3F_
zp}o2#=y)l|1nyj`#^Vij&a<{C1<lI%!Pl{M`;+BOlP)=5336*cEXLdVc}l&+e8-Xr
zQClZYx~pDTt+#S(Z1nc68$|D^y^p#jE9TO-Va;b5+bfQ1HZZ?(a0#~LUVW>h{9;Xs
zQc|Sksu^KBTsOA)`rMt=RPsUlXipjA|5X>gv=o<1>#~cl$f#{Lf08M<H?A^J)YCQh
z<jiNaXQpR;lWWzPq#87F*N1hfQy4|-_U)KudfvNgS%_?!TBhe-Ht)wzW-aKm)oG0h
zE)}v|>@B-oU2+G{s)z}qL7p><x2lI^+NYL&$mHlbFY&b_cH_gF4^kYIr+zZ`t^BRq
z9Q$ok<+ZcZ9;kZTzPBkC-Bx>ID{o)`!_8HzJez%6q!U&KH?{mvT)f`+xJcd>mwz(H
zP77{V-Y55Qu2IaVfPmhK4{K_^y;#5HTSlhBYOUmu1!t@y53O&$e2_Cb>oixa=ghO~
zj!Heh;{R~>wD(&BH`Q?+Z9Ki=?9b@c=bls<y-ZIu$~^0%c6Z}RmFnWtmaL_RZgB5k
z=(FynxrRpN`tU3K4#Bl+51&`vyu@<3_N66t%W6$#sKtMO7A^as;)&_B>i<3}Y;#}S
zm;U7V)z<ZYvDw~_A7ZB6n{@ksW?1Pu6Xn$j+L|wg-})TjI4sC(C_cOYXst!PWbB20
zLA963j&i9cH#gNY7_j-=Whk!9O|k>#>=j(eh37XWKIdivhC#SEn|K(Q0Lp;D+Eu%}
z{%QS{<YDj!N-MJ}M}i5MbWv(r@iRsS2CiQV?lpUx2D9*l%d5*9!3^aIS5Q|#;%QrJ
zyTMHi*Rj@t^BmjXF#P;da*L5A+=#;{0;b4od)4Q7)|kJHJmD$|DpoMz5I-UH7ZYt*
zLBgCW{4n8M?VZ!3B|`j|c)}Ic74;FW(6!cehAC=0_hAyFw%;b8UI`WnHJG9~we@CE
zuYaus>XlKKF@Op4glk%By1?|pM47n2fz*5G_kuMKtrd7s+zeC9$OR56n4&`=b}!@Q
zuOA26YrtZl1XK0Yfb|PIzj!yuE21j$FyVhsIT!d>#O>q)x>8+23uZA-xTLxy5>HB9
z3W+DJE{(*KRhLEL$*IdB@zkx=?O;~%glkx9An~-UwUBsv)_QPWQEE;iDCUccl8b{u
z3@`y=FffEcC=eNhV1^)=p>QU1__L!+4=l|1sLvJ=6dV#3$`ApUV@S?T$;kk6z0-V|
zSy<WFIXJm_cm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f$SNo*
jsi>-JYU$`37#f?J%b7^&8o7vyxqC%2x%oMIF!}%hk)GzZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_instance.hpp.15AB9D8608FA40A0.idx b/.cache/clangd/index/device_contraction_instance.hpp.15AB9D8608FA40A0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3fd0c85c74e220c80b25ac27af2c87f7b684a0f2
GIT binary patch
literal 19816
zcmY-133y!fc`k6B{~5$F5?~Vp#$b#ENwYwJTKkHmgcQ;wkj+}GeTk%oaxaj^q=~g!
znoSz11I<keg~79GM*CPA3bcJf0z<}HEj=ym4M|E9YVA^*l9EfF`${mK@5K*CeqVi>
z-<(-AV;!CUmmj$Q{@O1jC6zWlaORPwC+bW|NlAzB|Km?Q)~czbr0*stC7pTZm9H%R
z`QO}g_*Y*!6MgGno_ye?Z<c=L=_c2u^pW6GFISGfsHHSiPPhKe*WQ2kXp8j=OFQGO
z*WLErxt&*Df2ZU>pWMCY3oSpm<4+s^m*bKD_p|)!&)n7a?W;_I+x{?__SfJ1zpFcU
z{B8cl@6_LP@EdPk^Ps-^vhVLlZQpt6R|mJZ{_<>-YJTfv>wVwde(<Lcu4IOOw|LU|
znE$?i&TIa~_uJn5(pxuO^XBEevH$wy)mMD_nR`Bc<t-ok)^*2w|Glod@YY=WYv<ql
zYW_V1U%&n0kG5`l=GCm<9lbaI_n!^F{LI|HUisFtwKUxL%z@i-AN*9)x94lmU;gO#
zzcKaex5j?`&g}ayznA>A-~MLx^^7<E_+z{O@xvoip9qJ3lUDJ$KYQ?*o{u?yveNa}
zn{~xU0ux_--qQ5t?2mU>-d{KL=^sA$XIDIRXzuLHH<$mkHS*C1{_fZVe_V31?#@Hc
zKJ<6*{_an@zi2!5Xol<Ro|(;-Cd<1o)jVx}_<@XS&wGzQedze@x8DEW?_Zg*mv$Ze
zWaAb7d+xZs_h)Gz-MX(K&G`9E%?H2rb5Hue_v)W6PWgrTz)#mw58m~&-%dZYdi?Ow
zYYzWv`2J_V_tZ-V>OW%NE_$Nr>3{o=vW|?Ii|^~%ufBWJ>&s94!ykU0{cr2{f8>?l
zJiFm?z54xMUif-J&Fi7f3!PbkJD0Bb=_d*<eEBi`zn;h~t$5JZcK$D~eYE-W-#+pB
z!wo0j`J?ar_n-f5(yJp4_N2$3e5C$>o|JUed(Xes{rQZqq^N_}C#4-tK6+?>lXL)&
zmZT&tO*=GZIr^>9)mKy00SsK5eC>h#O=94#fqyD}{|`U;&eciRV&KFLC+^ta><`@V
z5nMvi?{_x*zkYHoMIFMx@#N!I?r)L~=-LOw=l-PL+hhFmEh*|&3_PbLKd0UC;fJIH
zCOlmCu1Cef_Vw?~zvN9(M=<oZmi)H%@ee;FhFV^G`sS1EUt3F2*J0?E<XckqH;F6T
zmUrfR&;R%{D=8`&1J@^Ce{g@37<lAUFMhxB$XCCZqOQci&B-@kv%e7o-+ri}@SVTU
zNl~d7xFPw5Yxg&af$rL)_XqDi`4=hb3Je@fK6ussCNa=ExBXYzr@lCuqK;zVS(W^(
z`uK+*5_{7Bqu)Ot`eoj2O5JgI@196XO1t{1tFQa;Qj(tfr{6xM-Sl2kin<Y(@tT(W
z8g7jbJ|r&VneVk{R;;}jNKw~d=*Hw5uif7yu8F7T*iR>a_v|Yv>M9HzPCk6?{w6Un
zV~@-|VLbhf6m<&*9!`Gvj{Qx7c<I>NhyP`$(2}B#Vc<;inLGA3iGlTh{NxXek*cF9
z>Shewm3-G7`<ukTe~#}=R(39aB1IV(IFfwi(EcWI9p1fFfB%P{`0-~_)J+)ZP{|#*
zxjy)ixRmz~etq(1kN)`ADJlg+hmsE+*xw{B<qy9)@cfs~RKJ*_uExN1$=4m&-y{Z-
z9&2iT{O6xbJR`KL_MZvyJ0ku(P*?3g{^NCO+TmLdpNQX(@#n&0SM5Is5<@3r+Q~Py
z_>Ga8mLAj6WzL9c88T<av`m?^Vp^8W*)c6!=A4+8BXe#{%au7Vrsc_;AJg(>E{JIb
zG8e|QLYa$VT9M4fF|AnUl9*N^b7@Q~mHAXmJN2d(e>kP4mBq9&nag8Zxy%(YtwQF?
zm{uwC>6ms}=Bk)hC3AI5tCqPYrq#$?8`ElKu8V1PGS|nndYKzyT7%4uF|ASNrkK_w
zb8}2<mboRSwaDBW(^_S2i)n2#pN(l}-_+vIuGF;lnAR?HhZfU1G}&~b>7=F$O&2xY
zXu7HCLDNIcIW*^}=|$5^O&^**YRqWN)L771sIj84Qe#77qsETLPE9|Werg<O9MlY=
z8Kh<i%@8#nG#)hZ7n{^HFB&g3J~TdR{Am2t1kePi8AdZqO%P3xnh=^0H6v(7s2N2w
zO3fIWF>1!qj8ij#W`deYG?Ua!p_!s4j3!LYG@5B@X3)$~6G0Q9<|3MlXyUI&scExl
zW~rG&Ge^xlnt5s#&@521h-Q(RB{WOaTtah+nq@T0)U2Rcq2@B0%V^^7C{oi_(X3Ll
zhGvbLbu{bLY@peoW)saOHCt%5sM$udP0bFP9cp&b>{8RK@ElZcYVjAt)U-Y{ebkuI
zn5nU#u~1`0W2MH1#zu`Djh&i)H2u^Jpc$aXfyP12c{JzI#NY3vrn%9$sqvukP~%18
zrN)QGM~xqipPB%g05!vChN%gn2~rb66B13<`9h=l;s17MscE-l-%=2NHyM8e`+t(J
z+JEONUe<l;q+cao)=x&%$rojgs*`WX?9@*_rHgsDe$pdzSU-6|W|!&Y3#P<;x=*D`
zFFm5trI#L6>C#Jg>gm!;ckAiWOAqVm(o1)l(xsQ-QyJ3Bh^P$dWkgkm^fH`!hV(Mr
zdWQ5e!g_}EGF+w%>1Fy<rt~r+DpPuyQI#pZOsAeHy-c^BDZR|Fo+-Udmnl<vSw58|
zy{w4Jl3rF+Wl1l~sb@(q%dKZgFDtBPNiWM~%938TPi0FlJEF3smmO8v(#v-0+0x5)
z>)F!F4(r*{%XXQvrI+JVInv9Cs2u6#L{*OTa-4dO^m5#Kj`VWEdXDsRT&5i9<@!{v
z^l~FAS9-Zol`Fkmr=BalT(_Ppz1*;#E4^HoDOY-VK9wiEyokz^US3q?NiWZ-=SeTm
zt>;NEFRbTDFVAJllU}}0<x4L=qVlDeA65C%%XjMe(#v=2`O?b|>-o~lcbW2~SKw0x
z(kqCl0_hb*Re|&hoO*%u3fy{u^a{dyf%FPorUK~|`c$Fx3L~medWBI{D7`|bUMRgn
zw_Yf{!mwT_y+W6%P<ll^RV2Nlh$@m^QB)O4ugIwvNw3JQ7fG)utQSeI$Ym;$Ua?OV
zORqShiltW^RmIXPcIw5_D|YL}(kl+@#nLNwnTn-X;!`EkD~YHQ>6JuPiS$aGdWrN(
z+<J-hO2T@H^h#W&66uxtRH^hzBdSz-rBPKXy;7%MD!o#-UMjuPuwE*?QkSVzdZ&Eq
zlwT$OS$8U;PQ55|RGoT5W~YAYDP7FF^-~_1!}_TUGP_KtUN9x*%Y3R#dSww+CcUz#
zDwAHBQ!kTVnOiTDURhW#lU|w2R3^Q0pDLGLc|?^<uRN;CrC09M%cWQD*2|?=9@fjH
zSMD;EORvJGDx_BtQ5Djwh^h+dRXFtu=~cM(3h7mZ^$O`#xJ(t&tMsW#=~YHlrSvMJ
zs#1EDPQ6lkm2SOKdX-_lQhJpxQ>FAy`_yUaosOu}(mNehr=@q=sh^hKX}5k_dZ)ws
zY3ZGInNCZu%BQNNR~1oJ(yNN9D(O`@^(yIAx%DdPRfY8`=~cN*Rnn{WscPv}M^v@+
zs-vn}deu(7T6)!Py;^$JVZB;<)h<)D^lE&nMtU_7RU^HcsH%}(jZ?3YUX5F?kzP$$
zuaRDj%Ty!1TA!+wUTs9xO0PDmYNc1})N7?z>(*<fR~y!ArB~}R)k?3<r|P6v7g2T6
ztBa~S>D4*)I_cH9^*ZU*h4nh=)wxV{(yRBWdg;|iRK4`-qpDtd^-jHBdi8F-UV8Om
zy<U3tE>peq8holjdJPfPAiajDYLH%oQ*V%7gIjNqUPD-KkY0n!)F8b^pK6p|V?;Gd
zuQ93`rPt`x8>QFi)*GeQ7}gu5*XS}eO0UVMnxxkhQBBfoimE2*H97Sr={33aCh0YW
z^(N^xxlB#cYxb#T=`}}Gv-Fyys#$u?PQ6)r&2GI}dd*?IS$fSbQ?vA1e5yryEfLis
zy_TqIkzR{aZ;@V$TW^tGOIUA_UW?1rBE43hYL#AVM72t<HL6;r*Xq<;rPu1#Tcy_;
z)?205>N2%Tug$00q}LWvZPIItsy6AhIrTQ_wYl{+>9vLRHtDswOl{IT>r-d_D)D2_
zvk`UnMVX`O>>DyW^|MduV&1Kv^~fC7&t8z(WjgzUDKX#fQ|;1gkEnL(wMSLE^xB<z
zyY$-Kdb{-6!+N{)+Fho0>2=@-86BDwJG2NIX?AE)G*a!*oO*{Q-44x-M#>#p7>%?$
zG?%GElX@rgPSQJ}caq)-y_57#=$)i@LhmHK6M84<ozOc;?}FY%dKdIA(z~E{k=_Np
zi}Wt&U8HwG?;^bmdKc;4(7Q?RhTctjH}r1OyP<cJ-VME*^ls?gq<2H_CcOuG59vM7
zdr0qr-a~p1^d8cCp!bm81HFgz9_T%!pM!pm^mEY9k$w*PInvKTKS%mG=;ug32mKuB
z=b)b>y%%~f>Alc<N$-W;OL{N#UebG^_mbWVy_fV}=)I)(LGL5I4|*TzebD<z?}Oe)
zdLQ&Y()*zIk=_Trk90G1GwEjNX41{j&7_;5n@KlAH<NCLZYJFf-AuX#x`lKLbPMSg
z=oZo~&@H4}pj$|{K(~-?fo>t)3f)S&6}pvlD|9RAR_IpJt<bHcTcKM?w?em)Zi8+k
z-3Hx8x(&LGbQ^RV={D##(rwUfq}!m|NVh|`lWvD@C*2O+PP!esopd{NJLz`lcGB(8
z?WFfZ?<c(<dOzv?(ECa6hu%+mKlFam`=R%f-VeQ>bO&??=?>@)(jCwpq&uKHNOwSY
zknVu)Al(7oLHZ!{LDC1I50X9zeUS7)=!2vWLLVf35c(kLgU|;_AA&wa`VjOX(ubfA
zkv;@{i1Z=oL!=KuA0mAS`Vi?J=pN{aL!2Jy9_X@(qIrWFr|y9+XWVE!)P&JopvGnL
zKu>&?7rK{pFLW>IUhKP<bT9VZOS%{P?j_xeefN^?#lCw<_d)lO?t|_l-3Q%Ax(~XK
zbRTpd=|1Q_(tXf<r2C=!N%uqdlkSJ^C*2R-Pr4tvpL9QTKk0txe$oTb1EdF_2S^V<
z50D;!9w0pcJwSQ@dVur*^Z@C@(1%GMhCWRCF!W*4hoKLXJ`8=B^kL}3qz^+MCOrr}
zNO}-@kn|w*An8HqLDGZJgQN$c2T2b?50V~&9wI#iJw$p4dWiH8^bqMG=poWW&_kq$
zpod5wfj&a|2=o!sN1%_8J_3D&^bzPIq>n%!A$<h;2<fBHM@b)rK1%v1^ik4Bp^uV2
z3VoFHQRt(jk3t_MeGK{->0{8xNFRefM*0}^G1AANkC8qGeT?)m=wqahLmwx79Qruv
z<Iu-RABR3p`Z)A)(#N5XlRgf8ob(Ck6QoZ-pCEk#`UL3{&?iWrfIdO`1oR2gC!kM|
zJ_&u2^hxNGq)$ShBz+S4B<YjTCrO`#K1uo{^hwgEpihxL1$~P2Dd<zAPeGp|eG2*%
z=~K|BNS}f}MS2)|nDj98FzI3FVba6U!=#6yhe;1Z50f5-9wvPn`ZVd&(5FeChCWUD
zH1uiGr=d@iJ`H`E^l9kRq|ZQ~A$<n=4Cyn_XGoucK12Ep^cm7;pwEy#1AT_{2=oZ)
z5$F-pBhVwHN1#VYk3f%*9)TVqJpw&K`bFp$p(lRL=_2%t&}9=v^9D6e{UUTZ<3{75
zCXD6+H7?Ud=!vhIg+5FAEc98@XR+_Iq|ajCXGx#MzR!|Ai+!IZeHQyZOZpu2Inw8#
z&yhX{eU9`w=yRmcL7yXi4*DGFbI|8VpNBqA`aJY`(&wSilRgi9p7eR>^Q6y1pC^4D
z`aJ0i&=*KwfWAQb0`vva7oabYz5soJ^abb(q%S~UAbk<~BI%3J7fD})zDW8a^hMGa
zp)ZoY2z`<CMd*v9FF{`-eF^#!=}XX;NMC}!MEVl+CDNCmFOj|keTno-&@YjG3Hl|{
zFG0UV`X%U>NWTRA66u$qUn2by^h=~KLtiF+8TvBm%g~oeUxvO+`ZDxo(wCtxlfDdn
zne-LtE2OVLUm<-3`U>eQ&{s%bfxbfe3iK7ySD>$uei`~@=!svCybS#^blF7Fyg`jq
zzYJZ@xY2m138T3{jmvZydg7~Ap|6s@3VoIIRqXpJ>8sfHRnk|n@2jM*V&7LuU&X$!
zlD-Cgjr29>YoxD1Un6}D`WoqL(AP*`gT6-k8uT^N*P*YIz7Bnz^mXX#q_0C?Cw(3I
zI_c}s*GXT8zE1iE^bOKCpl^`A0eyq?4d@%BZ$RH5eFORi=^M~DNZ*9MN%|)AP0}}^
zZ<4+VeUtP}=$oW(Lf<5P6Z$6UThO;i--5nH`WEyp(zl>*k-i0ei}Wq%TcmG6-y(e*
z`Znp?(6>q7hQ3YuHuP=Mx1n#7z72hw^lj+dr0+oAA$<q>4(U74cSzrXzC-#B^c~W7
zpzn~r1AT||UFf@{??T@teHZ#J>ATQ(N#BLOOZqPKUD9`<?~>lDaBM>PRpQsGdlimN
zs28chu?h7CH8?h*p3>!4;n;-oP=jL=>H;+{n0nQDQ{t=op!bp92fdH<K80fwr1vQt
zn;^YU;n)P}eG11WNbgfGnEFUJD;%33-3;AKx*58ebTf1_>1OC=(#_D#q?@6eNw+|^
zkZyr)A>9JqLb?UIg>(ya3+Wc<7Sb)yEu>qaTS>P<w~}s!ZYA9c-AcL@x|MV*bSvpr
z=vLBg&~2pKpxa2dLAQ}^gKi_;2Hi%w4Z4kV8+04#cIbA}?a=L{+o9V@w?ns+Zij9s
z-45MOx*fWm^nU35r1wMbC%qqfKk5C@`$_MI-cNcz^nTL&q4$$M0DXY;0q6sy4?rIv
zeE|9Z=>yOQNFRVcK>7gm0n#1N9i%&;J4kmxcaZLY?jYR(-9fqox`T8FbO-6@p`VAI
zI343W^z+bV6Gig|HBS9JbUEWj<Dn*u<^nY?(|KHj#8<hYyGeILca!eMzPm|xW8dAR
zyRq+X(%sm1H|cKWGPz0jK=+XDf$kyQ1KmTq2fBxJ4|EUd9_Sv@J<vU*d!c(t_d@rQ
z?uG6p-3#4Ix)-{abT4!->0anw(s5h_f8_m+pQNCX_qz|ek8~e&AL+RN(8&88KMO-6
z?{^<`AL;&mTk3x3e$xHW{iOS$<EJg+5#@*OC*2R-Pr4tvpY#Cq0PTAKdVur*^Z@Ar
z=mF9L&;z6gpa)0~Ko5{U41JjNVd%r84?`a&eHi*M>BG>6NgswjO!_eNVbX)pgQN$c
z2T2b?50V~)9wa>oJxF>GdXV%W^dRXW=poWW&_kq$pod5gK@X80f*vA01U*E02zp59
z@ne!WgZqQU1mY)bJ@|uvaLjeDUv0u!+eZ!`i=VZfns)t>>yO2k21rf2;m8fg9{KPz
z|J1Y_kKA}HekT1Fvi|DP8}70X;smjy*+)y`izmeY`3mrZg$d$d&6#H&t$pR?w=sY4
z(7_||r2*n6mSKKxje_`e{2HL%TbUqU$Co5Py|*60|I{B|L)3e#5X9@h`21%seyS|<
zA&@6Nd*W35jPy8-)7STwB#2LA*KfT(Ee^-OUE?A@Se76@*0=rHTXN>@M{%Lo9=i6%
z__^@5EC11D{?TI{sE-^uax8u-fBYwxxR?)?B#2*3{AR=__EsZ^*YW!S_1+o;@jAHH
z{bczspBXSx)9$+Su6q+_!99HE!}li62{?1-nR^qz(f6Eo=X2V<@kc-^uE0I7X&Lcr
zl#2W3p0~A(_#Kvtg#_;DP#KB$<j<;mo>dv~8zOY(r7IqK<c0gOgU7Eqenb33z>&fm
z+kd$H6W{+%O*?6vygPBt(v5VHGK>t7GL1};vWzT|vW;w!a*Q02a*bS(@{Bx@@{N3v
z3XB4g3XMXMii{$Wij88CN{kYbN{v#HP8p}}PTbaIMwv+EM!858MukX~Mx{unjng7k
z8C4=x8`UD!7&Ri*8nq(T8FeDn8}%YJ7!4vd8jT_~8BHQJ8_goM7%d{T8m%I=8Eqn+
zHO}6hc$~Ex?ILw(Mu#SHr)G4@+@%>^GIwi6x6D17(IfLY%{V7>uV(bh+@~3RGMhES
zEVD&3EHYa)!z!~)Gi);3HN!4*zh?By?9dE{%!8USDD#kJ49V=#43Bnq;#ug`46n>S
z&G5<W*9^bR0nG@=JggbRG6yvyD04_NLNbqN#)!<LnlUQ#m}ZR0JgynzGEZp6gv^tg
zF)8zuW=zQ(){L;s)0#0Y^NeQ9$Q;p(h|Cu?<DzzV;=9zWX3WYwrx|lH&uhlK%nO>a
zAoHSTEXus38A~!>(u_+oFKfoK%qyC)BJ*X<xUAis_^!FC8LKj{X~vq&>zc7H^M+<@
z$h@f;n=)@{#+J<6nz1eOj%Mu0ysH_zGWRN@SKXcXF5ah%KAFwRFw1OFhDBzpGORM&
zlwp(Et_-`({mST<c|aKhGCP#vkomka&a1oQuPS>d!W(XxJ<9ON>{W(WW}h;AGW(U`
zmpPz}fXu_n7?wGxjG)XRWrPwr{{0=Vzk5sk#9x2o-|SHDE$kDoS1$N(&#*rGOS~rD
zoPG0s@pmV`c)R1ip+n>Uiuy$PiH7+5j`(XME@W@<pZJBu-zuQqTk<Df$L~SZAFS{b
z|3rL!AH03pTihpJ<GtFgH=KySSBtOlgQ2|zeiB1B-g@JS_<Ikm^YhKRz4eqHf2`s0
ryuH8Nwzt{;Po@3QbCxfS+){2e-G2PG6G<oHpGZnNp7`tSiNF4TTNMaR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_multiple_abd.hpp.D0EC339053CFF7D0.idx b/.cache/clangd/index/device_contraction_multiple_abd.hpp.D0EC339053CFF7D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7d14160a2b20adec1f5b4dc6832d13fd2c2ca6d
GIT binary patch
literal 1606
zcmWIYbaS&~V_<Mj@vO*AElFfyU|<jg;^LB`%y*0o4C<^53>9<s2F2#jHV}EbbN?a!
zxm$cp+SdHuw98B2f{ANN^^0gzMfP`1q8oH?FX>(N_q)`jPX38y{*wh_<`=)S{T}!D
zPK@2pmX9)W!Oho7MK@~we7xxN@}nHF+h6<^+?91PJBd%oVl4x!)uOh0H<M4-yqHmN
zqUOzvj1x7lW>h%Tyt{csHRQlq=F1<Pn-;Zk-s{@;Mdo~0&H7;7#T7y2ukZAHd$j1H
z*RcbqT!p(|G(HrO@NN9aVrE&g|CjcknV$J`Ilk#MFMZ&4aZTf-T_!%$1X?D3*7XxA
zYJ9U|lUc`t1H1ig#8vloKl^MMb+e+hr+(}Ec_v@pU7r^Bzu818?CsLX%qROD<X(5&
zlHop@zBFvBOT3_OYmM}o3)WjQ7Zs#^yzwZK?`&n}+daoN`Ob=EiMlRo!u{f4$EQ1L
zq9yz%L>fvq-<u?tBkt!^F0dl^?wg;>6Yex*ZG2y&U$)?%)3<)kcMaJW+Y-gymK)p_
zOuBVpTj?$>r%&Hjaa6X}y>!d{lpn@)P(pkWQ%uIJ;IrQ%y4V9WpKh_qSXQW}rD<|^
z%7k0{ET_!)xMof9ft!5&Yd?lv`|+ag*AnSfiE$o3>+fvJySBxOclIS;yB7`~&7qCk
zBg;}-G)!k`*EdcTV7new+Ltf3uy(`K6+vGf*f5`t-{o3dnVYl+82h^CB-oF0g}fGE
zV&w*Mfq++?MO~bM1Hv)8AsVlLr)IV&6FWBpuac~iI+%b-Tc<v;IyAS^Nq~tBrrTW8
z+!`jGVt#58+rnuHY)s5t47>{L3d&%Dfs=_5nEDub8MB0bzg5UzGgXv{18BUmtg<?o
zfa!m=RyzMH&lf{cCQhKVimZw{n1D$arKS~&G9&!J$O(2k%vPA&S-8M%h6(fVsvD?V
z!Tim`t8K6CiQwtj>%e(1-vP}9`3h!0h@X)9i-|U@96-H#+9pVP^>mPUx_Y`uJUu-<
zB%Z#WJ`&GB&j5*MsBEZ-Fw;oc2+oV$ZkZFKrlQQq!>cBvrU6p~a~C5QICMQ#%DN}+
za;{75V^wVsDs}>@R?t<jgsFlBIWSZ};SCe!;ngtEKnfLIdtG=q@bK!{>%n;yhaWzv
zV!3#h6Bv-P>Pj$I01c4UK;mi2Y9jHpm9^pS<>A#))<NRwD(fQg^py3Gc>2ovNIU~&
z130hSi@D`(|9=Z^9$qC;B^g*`^YAK*D$Bq;$-}E6sv-kRM!@(o)wV%6SjSWciKlC-
zi^S72)kEUxo9ZL+3``A>c!pYrh6po_w2a`qwZ>KfN9QI?2gZQ8jJY~2IXyLC{ld;K
z-pvGzMnN@cgkEz&b7`0&n45vo1WNaaXwo#$M2aSTdwqDs@bDVg8^C!CMX5Q7O2AwU
x)5@U#%tmEWnBr+z!hp*Z7bO=f164Bd8it21ol(*~fr*cYfd^6ofJi<B69BBBE++s0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_multiple_abd_xdl_cshuffle.hpp.9AA6B0ABD9EE929D.idx b/.cache/clangd/index/device_contraction_multiple_abd_xdl_cshuffle.hpp.9AA6B0ABD9EE929D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b91b2180af9c1daa7724576aed60c2f5c3069fd1
GIT binary patch
literal 15790
zcmcI~XIvG>_cwQDc12h&S2{=&q={S=M2ZECh~iqoo@mr)EJ00dv5SCIOF*%LAYw<F
z2nbRvM8Ieiv4N-{BDNSoEV2EE+<V-4UggE_^W<YZ_v?IjcJ|DfGv}O{0DnKfL;W>0
z2G8<ev0(Pn8AL-vqZ9pq$<oF1%6ScqwbmLMD-y~kO<q*%skpghJ!e%gFfc!MYww7x
zZ%emq3VKkZb8}GMz_E&)3%5g-ZAiVFI^@fTEeiu8Ts$tK4K8cl2FoLstl9GB@w23&
z{VyiJw$$#eJN5kXIVIlBZGW!on0T={Ic(GquV*=0qNkm9rgZX%w9fn6OuT0tm(|TH
zK4?T>r$ob)OJ&|?3YQ$Q@#)bRUeG+G%WlVzX%lM}HjLT1s-f`mlf%KTHbM8roK<m!
zm#3_)@cE%C&BJ8c*jrOcbxza|Un1=sL-peIz4pb<U3I?M!mw#om-23HpVp@LUcP3@
z$;|bsA=kg_S-8Kuk$uR?nUlx=yifbku?IHBn+_Fql@*lbnpOw)`8adUhzG^lQ{~1L
zt{-N4S-;;Ir2prKs?(QjoM!uFh30Kq@o`Ibt^0us1xvqCI(`^d6hCQ6zdg%e36BRP
z<+@t-8aVN_{z*si`ISA-js6IGothw44pBH4Bpr2H!*7h8JFWLm$`K6;-eL5JxTjD1
zG<W`P`|DO)-Bosbw;Cng)%J^XwH#LAeb4GjudpRsPJXZyrkGAV=jAys`*_F41Bbt?
zt2a;R^V7}n*0ez-_6hG}{}lU7_PqXi;@!D7_oUtp&oCHKGczxx-=YbBIZYk%DsQR(
z$w`r|doH%V7M-4KOX+s$S-kG|ChM*?Z1`>U;<6y;r$J+X8h$aP`st!IwfcA5D-)*~
zC6w1qF*>to{>{S+R<%vuJ7Mj;^uX`SBcitijaj&3N=&~&zi*72(z8XQ*J8ULJB;04
zJ9TfrsrDs#A!`gf)g;}m`R#Dm@*l5My|sAlQSUZ$_4Tfo*803_s~_1fsAcsKnZ0>N
zX++T7O?KA;j1qNp{AX39w9Z&p*zEHArp0H20)x%sbCltSUj8z6eSPHgJ*BsGv+K7v
z-@4a-%7oXekFFbRW_>-yaPPXNz>%Ypd%NC_o}T?=rtyH)cOHm|$1C)6w(IPi-h6gW
zkIpes_4TJiyRV%5tEo8Eb6QK|!S79f+gvlNFtk;;cHX1!<&c!ruEUpp*rlbBsqMci
z_0wdVREu#9DYK{7400F7p^M9sFZ8v`8R2CYJ2tN}b>X7|^KEr^Y!{VR$!i8V#_t+;
zbXJ!u{<CC{JX~A?HXa(U*=^92N$dPOrW7QjVWELlacA$_guM7;)|INc`?C$7RB!G#
zJ7Uzl=05Xo1{P&6AJx*B>K%OVl-p05Rm$NGvj@jmzc}#A$USW&JuNXX@0-kRIpc54
zf3j!F;jK3FtF|m~n$ls&$yGDbj~aUST+o^)FUZ1AZ(DvEVY<~j^tJp<`i|>PYuzj!
z46jX_>8JB1rT@<k69W@6Mt%RtJYrP;_oZ7i$9I_C;l{}<-9__POzLo>XQupE{WOc5
zy~~!3`L;NI-d}U3?EbOumiX0I_f372p7VJ8?uwIc$AS*bX$n}W`6ly-?ScyvW*E#r
zp)7UvJRdl6>Zs$|nZlK`ja9#fj%n=e@78>?&y#~M2VHbO@L}J^EtxoC%kO;nfasHl
z_RL?W)qC*bncwd+nBDJ|eTQA63@#n72<o$DV{%)_m${LC-_M91RpMX%?%A0zle%H^
zKI_asv9rI0l~riCRkZKyxCa_jQ%aIUK0dhItXZ^CYs2AJogV0y9xw}jD^5Dn*~O^G
zlGpPuczrjaF`n2qOrD+-{v`Cm9vjPLx)J?Kchndfj;%HxbK~CdZQm9wue*`Hx@Ot;
zg_d(~-Me{z%%Q2RLrlm3{p!lrnSOnij=M4;|F+4*?NvK-D!NWOxG&(+FCCNSthi&X
zKh9bIFGsy`-JHr!y7~XPdrs+^p?TF0*M0O<T5f**D(UL=#-{L7|BMPeH}y<Z>}6f^
zcq{k8GZxJgPB&T(-R?APyTcg&pN<AiavC2z|J#Wj-`ZI17~lR!qh<J%rk{@={bZbZ
z`Dc^$k)EEx2NPF$<}~QlIWL^rI{I|LjxD24|F~ypW@=4am!~x|mRR5H@qOg@BE^(@
zP076!M~CiRGbcCsjxr~^!Xowdm|Jb9UIwha95p1r=m!4MclNFo7gj##b8F<>#sLmu
zxu52Sp%!&-MmXdgS-7w?PFpzcmAm5Vw$g^+_h(;uYwlb1-glFQ-m0d5qWTOwd471U
zX4gTcEqAmRZoE~!q3@a+i`ME}69?^@7O;HckFlR_kGyWR?O+?SOq*^pBmLva25$MO
zy3#wpu2`d6wx}ib-It+BDHk61$Se_q?e9Jeo^&&&aauR=Nad2hjF<iK&eq1qb5#E0
z;>T%|H^gaA%nw(7AAQtHGx0=Q{JdiQzprlze3gFd*6x?u{TtIJJ=^-(<nz+s?p)<&
ze3!p<Qi#v0jSD_*<y{t>J28A{pq_YX{h^sIUw$#`Kc(yk@5Y=Hi@vQ{AAaF(?UI!X
zW;%7$&>-4dviz@(XqBVMeKo}GM1MOmR|^eIy6@5dG$)qZhdtj`Ek`{l<p{AnLY&nC
zDDoYDE-}+}=HzHFrQ{Js9<fmipme$xHndB#|7}5IoTrB9r0LZA-@i0E5~MMuk4%*5
zc|5U@>0YT)ATw_b@iy1-HfOFD(1%{d!OImlM(@=_=9E&$De5RiC4geLI&y;7!^zLI
z5l<<hSSu7;s>N5z>87oVTy}QXMw*loi?w30iCO?fIrPuAmsLTJwUL}sqOevJ?xPk!
z>Cx1$Yutn`mv|(2X^8Q%4)L;{YN62)%k(rf=r{Fvk<hhf_{*hoG?3~gi`Zom54C`P
zya#s6saX&(Ll0TeSH&YmJaSNruat9w?P_u1YdPvgd)Y@U_L1Id0TiF{y>x2=4QCr7
zd)iB?%pz6hpcX(G=)EgrPv6Gn+DMC1J|nHq$V4rGGU)n%8LKYb-)e~LDCL06;sEVM
zC4lmcM#|@e;#tGxs5c$UDdKpFc&i0a{0tZNe)((VL~W#?y==l-o3OiD07cBI&nsVa
z<&7MDLwngzdhaJQ)dDCZKfg3SdUa)zHsbneh($zRM2yq|C}Z00KJeR=qf{VAN(tuV
z!CWV`07{_IXkm);ptXj`gpTEy%<P!VTrGey{;zi)arwGwax{^SC62hn(K%8HpnTW3
z-%~!{c!(bAOna%9De9?SR01f|>Kb3}f3x7494TopCy3n%+KWm6WqS4I88>6Q+%`nU
zw3nkYv!iq@Dgl()L3@t7j~~`cj$COkyNK~F;;0rtnZy0sHqPtqVL5WAlni2*L0_a2
zK$-V&MukDps~S0)Kq;}rC6-cD0w_QIP~W=geL%S!4WN`{A|#W(Y5|m><%eD#`T36=
zIr62HY@(A*EY$)i3(xtDANt16SdN@1<uI{4Og)220A-omrC+A|hdt6mUFb)*Aku=&
z)dDCh7mZx@`@GN>+DMmDz96kH)G$>7DC>SXFeEa6a-}xvKq;S))+b6)37~{F=3UNz
z`1fo%>PE+sNGuZRB2fvTL>4;y`nQ;M+7LO=UJlAE4$^5<37~A6^Q+~itWl|Y$dXc`
z5s9Y5RSBT{7JO!|wVPp_HqxP#{W7inGCj2bO6-AxvbjTw?`tCiO39FEWyp-w0w}w7
zRqS<0)3et@)|66=NHKkpN&sa~y2<XUtZ8Qr(RfO!lbO}g;i?2sk}b3QFRi=$K#nHS
zkKRUHw$YhZiLaFPHKDIg-_y}XooFw?SSuJCs|8T@cRd+mlhoZqj$CLjIm9uCu4$D3
zO3K{BI$xf>>mWzYl(Lf;@1%yQ5<odJ@KaU!*88psWI<mP!jll*SS^5ZROkEZ{&Oa;
zS0H^#ImRoF@fK<UlrzK1$;#Plm#Wru36YnOPNGTx<*eH{y~RC)%Jq;9?WF=KDyYX)
z380)GGjMisgKnuIGNP0tGP5Iesi_1|F1#t#oYmk$T~JT@MXAI%l{l#dP|7^6Ci?!d
z^t~MUP)aUw%%yr!380iey46!z@$3&hWJ_Q4nNxh`I;#axD!%ngscmUUmZR>pmvmy6
zPUlD^fKqwzM48`=oTohEXfN?tJ09z+1yC*>*>3w?`bz3w`qN%^6XV@<0#pJhSLy~H
z&%g2McRBK)l%vG@DCwaVK&eWbx^zKr%}6;KL@9fS(H>&27C@;s8#OHdr+%(-<VGo(
z#4eK#S0#XQKQhM4X0?~n5E;_p9+sIMrrK2rpfuI~)_3UaQZ2oIH}eQi5y9DtDgl%y
z)z*I+Pap9+k4$MVg+#lM=&1!zUcL^SSN38~FKwhvDQPmTG?}Sd0Hx)J?!>Zb*#aZP
z(c$jInme&VEr9a&`pQth)fJNzs4MLymG6?uyQu|GK7RboIMriUnKtT3DdAWvoJI{Q
z0ThkJvwvI?+Fgp6kRTq#@}t-ohM=X6o4(Fwp}TdzA?Za}5QC9(F!F<KL=bb3a}M%@
ztzQtEk#jTh1H*=Zi@qTT%X4sN=?$vr3*MkvbZuSQym7l^f*6LCVL0$>^z-*+)0COU
zLue1_yiPi=fR!hRS-eh`gmrpG?wKFot5!##6KHi<W(RXC(B)!vL}mvoiGHli>Zr^P
zHZg(jqE^RbcCf(;;xpO6XENp+pUVb5XK^eNVv!y=SV2re1Cv-CB%^`JEG|Pr8PWqk
zDTvq6!0VEZ1hE<otY-0hB)ms@u;1v+G4q&r+`JSMv5XhWcsXc*h~>Oc&f-d5sFbkI
zsUszh+|NZ+by0{!A>rQ`nEoY*MTpRU1`IJTC2l6dW@0JLk}3d(SyF`maHQm)8cSXf
zU&^|@lns*XpjBqrD(f!c1E|{pG)S_G!^rS3>Mr41sM{?x2<$`<Yms3s>JAt}Y^3#X
zXq~ywxny=hY_1?)C0(zQF`3G%d#0O(RYRaH(^7}m5XwS$O&e4}H8kiK9np>osDDfC
zEJJ3RA?pttoXM>$&5wukToDmNIpa{y1VS<*)^N5pTvvEk=MMv%S1o>KM0?2QCuj5D
z!gN^4H+V&9{!vdGE|Sg{i6d+PR!<yqJ1&3VLmTWg?6g4ZT?!V@eIDE|Tt+D0=pr!%
zzKu`L=@!qcTK_o|o4qbU1=Uac(T+D`<!1Z?#H+A)u$DSufX^3b9HE&=%pt@O#5=_D
z4sivA+Nl!?`k*SCkRD7WnyJJH_8LK~BbIffJ3QE}{^0Ug7lz8IhiH^(Hp=Aim4dhn
zY3@P{{sU?LAz_zUANyLMwdXn#@hu+q77qv8BjP(8@D7iY@E1Jn3)>z{0)h$C9zg;k
zSiFrJyNwH+wqW4h`}|4^A|B$#9%AunZpvxyJE{F57kH5y5B?q!C#nin)+DJSiD)K~
z-Vk_F+a;E@WH_h-c+_@Pxs-p`<uS1ZgVzZ(wldr=8vtP!q*Qd-MG!><v4Kz3QYTih
z>|BR6M)RhV`w${-<CNRD9@3Z(amqt1KFujlv-l$CaFKI_Z*u*zuED|gQ)l|dSYj1R
zx<h;|h*iY7ig>~|xclGtC>-gyfLhj7&f+T94SH~Y-setG#ouFSg9C9egjx2sd{y}q
zKjW1F5l`{LDai*BF_#x|S)9k4=JB0j(GW48H_d19Y2NfSiwk(u0u~qY`h~oq^!yoK
z{|t+Zc;h18RBA8gjf+`a!gns=Eu{8RzH=#y&+<Lb@~%?*IlkvP7N6%`&htH__6xkr
z1s4C#3%|49dyy9|vbdb@Ud}6}=PP)jLTU#GNv%$mK!Y_2;(26z9x0_=<O(*pf}O!;
zyw#aCU4R)_o`LP4fo{oI(Tp7cYpL^XFfPcq^|jPlw)Dng<XenJNpCDcz9lR!L;7XN
zLTcZIopxb2coFqA*fSA-BjJ76b03S#uvr=IDz$&Ywx6(zgfoa~2C;=SQJ}G#X(q9S
zuuh=Ky7?($4|%*GrgCPfTvsqvLA=Mg-{X8>FA&6sobN+!l!O~O-$oXP@;yU&Z;1W{
zF_QO*<ozWa#rs4_7^)C-s#GNiNpX=_h{R?PicuFxG<OjrNQt2~K{c+bNJtMxBj0E=
zN{TFEkZ%l&W08I=vXI($VyB&~#`j{+y{v{4uxA2`^RZbztMN8$+s11A1Tj59Y@q@`
zLoQ;PPHZ9Wr$HPs&ms1ZWl=A}nZ>diujAb7SdG_nzV)nz8#vzv7KiXXLs*T6^FHB{
zhN)lVeIg_bGfmys2BO%&I!smh25j25V#m{Q|Je4S#J9xmE$IVWs~~<Lb{`}R8z!AY
zRdq+2iv2`$KXHTdjv(G8mUoFiILM*b6Z-C*GGs|dK@34c2r`l)pHL)(vN#M0VJyx?
zLM~0_)WtGE%tJyRi}R6?&*G;@c#4c5GNREb5}vX6ITD^r7zRxJKo*u~;eOI`dWjV;
z@c_VD>M9lJY~*MU<MJW9#!}B-AhRuy*+XOm<twTqRqYDW9f4-ECMCpLQf?^c9Ll*t
z<OGriVXBH2JQ&z<NYR4XiY;`E$M}xNcwO*@?J>SxI`1suEZ!lDcL59`I_*(a0E6io
zH%iZ;Ikq3yRiMQzw|beEG#w2xw+0qRBe!VeB`u*C<QBu?c;pt3yrk*bf!ubmxEQ$=
zBQMG6l_0kg7FQs*3gjhCS0!?*WN`~}Ye8NR+zVnWa%*MrUhJ@!U6KjdA%Vq-*ddX{
z`>?}47AIkcBo-%Qhh!G-$L9O-*HA+cQ?Pjoi|aYJdd^F7k`0_&1B)MWZVy@9$hkGL
z_z~y!h{a8uTN8^PbB>QWrPO~j=h)2Ra9$bC_m-S)1h0%>aU`#dWN{R)jAC&#uZ(7K
z46lq~@h0AO6YnJTznQn)EMX|2Qqw$)<%h8eEEhqn#EMF623Sj7ScT;?VMhGqwEeZ-
zbopH7^)K^AV6&6-FZP<AZemA!{3H`T$s8rq%tAsIa+FN%5fUCDM=%}W(RHUP(1NiG
z;#Oq56)7bfK8p>`N}<5iJg=bd?7P|$fmU8D;;|Eqk!J6lNif$H7N$V6CfnOw4@hmO
zTjdQl@YbLl*azC(C|10ctVBy)4~C8*m8R=bRRxBjPH%{u!{4{QM~AXq=D1zv0S)ch
zd#5bk8no1ohh^~-C7jI<%a-s=SN}r&8O8Zjy&1fD25$?*nd@9S)@*O=L>ughy$J-Q
zvpg!}`29Pk(|_I})7c?&f)~Rk%n9OgEI*E|U_$A<VZ~+KMZyoT^8?m@KEcDD;K|Ug
zrLJa!4j@dS%KCy8Usz?SYT7`An422JAuK<{nphE56tO0zs&vDHkZ{lkGqF4q_lKd-
zJpn78VIRO+>iRc42=#9|v`s|VL@Xqtsu&JLsDe}CNi09fmTOeCa3KEd`SO>*>r0L4
zPzrFb0zA;YIBedLp$Dha+!yz{$l_|;t6IXp5RKhOdw#e|A57;vrSq0BkL~lSlfj#@
z*foRq0Spe4W>u=vJ4`55H*))eM5w*5IA7nmid?(wB+zQNV>EJ+mPrh9jA3yxax6wJ
ztSfgcVQ~v`Y(XyY=5|-@*vjIa*l4GuBO2ymqg^cCjg5A*crP~J%Nke$Hcw!2A~sKC
z@jh(6kHtyYJc-50*gToVCy34oq6e$CJ$lngCwfx!mO*qfh#nYfdpxa^N%W+6I*aIJ
z5j{vt+oL_59HJ*hd#8xbDWV5am_Q52I=Ms-YRj}Fz}eJs4w7rX$JyLt@qNzbK8x!)
z$9m31;x}-P4J>}hIX+}@Bj?!2;zyk0BNjK&OqazWyhR9a1Kvy!LwSo(7KiZ`VJr^k
zZNqs-i66n+MzA=Nw~b_R6mJ{F;%MGBn#D1^ZH$D$K&d}h9SKPm9D{`zYz7qvfmXvc
zcM~Hhl&=YN^SP1f<xYKB8LwN$cY`$nXGye2)d`Z+<3=oO#AZ+j66gVk=1yV+XCEQx
z_J9xG_J*{ekiiRd;X#k>xK{?7?q=~rsb`kjv-zRaGXoAO2yl+Ro41Y*&7p_GH{gAt
zywtwRu{@nkMjv6tBQ_aTooj(#fpaZAE%j-Z^hGVmw*`%27o%@0i{B&t_sBwO-;JGi
zv#C)M_Do_^p=9it%;G9+R>h`98;R{k!X%?v#59Z8N{b_xnCB9E$@ud)vphBxy3e`a
zXH%g^obMwx^=aaKn^+vi_Y7lGp=jPGT1tHcF^2bvk#Nj#9QI=C7f)A#=8%R5Waf}m
z(kxhJm?|@ev|gZ!G(0FX2hGx~Sl0iZ%omR2+Ogj~*+>?Tz9*X?;d+^OJzME*ka;(-
zI2w6Jqd}~f^^Re2Jo1ibE5SRE_YM{pBky7~i1otWB`mH$-W6=6xDt6+vbY6#x1d4N
z#J3{vRu=EY%Dvbfwmd;hz{&&`Ct_tHi}zvWJ{BioWfF^%u`-#(`?2kQ+)e5~1>2^u
zxD6}Xq@xuY1!Bb~7H=eqjpXamOS{+A&mm^4*VR8o%vi6hpG(Xb!y}q`L>FrKG)^UY
zl~OU8CPbWfJvWH;vfd3We#m)0WN{<s-N@ococALZH*wxgEPl-Oc+B;Y`fuiXG_yFI
zcMs?Lv0l|Zg2j=%dnAjac=sq4NAvE{ERNybV_3Y2@3x8WCH23V@3vXOF;9JO&i!%m
z$N#x@3I<LC5Y;KV<eEMqw@*j}Z@l?o!f2QElV{T)YP-y9yR5H7%;G0!Ne2d7@2vFH
z(a3qx?mUegT)=rkdAI}3(EmM>m!2;orbVO+B;@VIdxJZ~9BLARSVIhISX@gCYFT`j
z7~ExX9WkhrFeLm&TIv&i>BV8#It+K0wvz~K9f7+`1+Hk^DH>Zy8`x%Sy%~3h>WU!7
zVe2^hy!v#3o}Az=*=$mJAM4&{Bc&#6(1fj}HT47=JYn&3Z19}Le`ABcSsY6`#S#mM
zwCVdvr#NB($(=wi7IfN1Ea1+8Af6?Pv&38~WLzT#*NC;G*eYUCMVusjomgCFaW%21
zW^o(QZ6lqf{;qPmSGmqmnT%Ud^kw$7sr~81-;>$jlevPk-F~@KWQ5D(<Fp})Q$%sr
z@I-ulx6*;dsZVJ`8gG)ucZCf&{_XUsA-d)_Xo6KjTuO+S^v)7O|7oIxH*h8!*xYXe
z*JlIw4djA5Dux}t9=@(Swc0q|Ada_^7Sti$;1F*GXV2iM={itd@{p|LGjjWkM5v3x
zM-h4k_wS7li5QM_!;v{8xHN}Ax&`zi+}D!WUr6^aWDbciOcRYJj$ru_HetGj6}Q-g
zNp)caqKRbng%LUe)ol@J1hn?F1)H)4IxG&4mke|d(cD7};gSJ3W2&6@Sn;0C?Npab
z;EfPq(d_aVmLFrY%TlZ;WrOwm*!e!2T{huiO>B0lx{~s><Ce4)-`RTPQ3Hs;$_N|?
zYcF+ht;ILz{@MGV^9J)tb9B{>7ElJ5QhQPD-+L_3khY?b8?zwAkDm61A>%M)B~4fu
zvJXQ_3Fjl@e0uBe>jdT_`+TI7@N;DR99hAH3gUBQ{~Rd+pK$XYQ<gY=C;f&(VqQob
zpioGI8QvgOvb|H!R-H1AZaMa!C!@K2vUjb%_IwyUTTSEj)1+!l?nBd2Ra?H9Mqjm!
zSZyQS!MNxtJaMj;PQdbieDH^_R%JaUMj@RjWDM@P{qS6;NW!O|xvU;(?z8jXGx30z
zvI&w_T4f_!Wn(3L00kUC6C|x1Mk5cSu@b(80&bxRpq2JB@sYJ?Ea1X^T{<;JzS&G)
zl#hjcY$UnIGAxv_xC#qZERN-bSk4HX3Dqhmq_Q}V6Y?YsE`u6`>beq4$C>gQlTY+*
zee|E|haq!X>f20UQnZkSjNc-q6s-kgw_rRRR;nPLz*Z-4AL-L-u|cg=+B*AAv(3@9
zS1uhvId5IgcLC!iVkK{1$vXf(@B7y*q3!HOY5<!^k4>Z}^a!^=sQXggqLPNzN_<+$
zAo%+8N5(AF@zbQ$`~Pr=7vB6hK4WodGF_V6WRBZpgP}(_E2n-d8_ToV{plfgekj#%
z=*io^m$qQBLu^N19Z3|C>^X<(&K3|aKG_qT*xcuuoQQe6kS84@wjZks`7A!o3#VCJ
zzzYQ|F64zm7N6mTGb}FRg(4Og^FlF;OL(D##ihJZ%Hp%UaF)gAc;Ot2&-2207GL0n
z3lfIQY1AZDH`74JP~4!ZdWRM7SkJ7wm<Fn<G`r~de(iWyx_aKoEZ@jnrDgeEX8E4Q
z$B^YQ<SH%86Ug!ei|--Jd&m`RoCes)vR=aQLK?oQuD`)hX%hbLoh~3yM?!3ih_e*V
zMR46BI9F+z)^Hv*G=)^(CKtpPobC(GSc<gcc<VT+z5wL{y7#1Ec^bR-)L=yoyZ6*%
z=X!SUd4Ww{U?*52TI$<)Fwjd?nq#*tKb}AhJQ`a?<M9yZ!8JXq{8Lzd3Nv*@)tx=)
z;m>~dYXU}_@Ayy0adqN+lXuG#t>~-zlD=Orja;4A>307gm*~(a=O1kF51Xck5Q7lH
zoFPXMgD4hnCS5j@UhrwxZZ4bXZ2kLsfu>=o^C#p4LAyZDa65lSP7u7bR~0*dK~7M#
zp`~-&IT$-hjw%dmhG8Q%B-V_;MzAf@T?2bX;h~_a_8U5yo3RlaacIWT=f8%;1=ywl
z_hjp+HWyi3jcuwW43P<)>=Rgi0)M@-)$VKjv!y^n5F4;_1MUkN5X4v5<P~-Vtfjs*
z2~!2tKsu+vL=nvT9@W*!uf$(`t=BeeJ41<a#43(-m!j+I#QD0kliuj@y!Pby>D%ba
zYnJIZ%eqJ_Z!^-}%pPYaApHb-yZUPwaTe*GWiJ?9Mfz7!7is0SA>B5q=5uHJ{)d@&
z_OGK>md?AR^8;WrAoOsdM;7k~7;XmAJwSEWQ(B^}q*p6(gFT?$^UuL=(@bRlX}66g
z2fP2)q^qRLXduc4;tD;`O9w>xP{NOTa<4n+8plxL2EOA4ULV4;_E4u|Dz6V=Hh2N*
z@l=;hVT_NvpFZJcaXf`Ckq5;10dav>J?UWd_Y>iL5N(LV$~Zg}Yzww;8V0NGrb1WL
z=sAbaoD#e%J&@q75AhB#hIZ_9i0=vb`S6XCg>QX!|9|(Pa9fs6km{<eG~#TTX12^2
zQdO{Gs-^$exwa&nF6oikrR#srT&v>A^U$c#z5nO>z?-$duQU2^R`-AI2((SVNp-^)
zKK|{8j5On?3)%y~5M&&JoTY^nf;>WyDB)aWoQs^Lg_Mgta*-(Er^xsza+Vg-Q{?d!
ziGV+RI8@T(?vB^gk_u#+1+p%nZt!lj;H<h*3_}y>r6SFJ#2juRwU;6+YovRHpNnc6
z#`SqIf%cfe56a+&!&0Pok%MOV(%9NJO}oM1<@$*-NJ!xnQv)0h9O-GmOCfF000UpL
z_=hDY+vD`15UguxjQaW$iN3)>V?xgr4v~)y2pltX*eL(uBS(xL;-@j>D@P+x`Z-$q
gIY#>FFZ~=Y{TwO%^pk#$kbVY8KgUWxhe<#G59)87r2qf`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_multiple_d.hpp.CA3D08C20FA799BC.idx b/.cache/clangd/index/device_contraction_multiple_d.hpp.CA3D08C20FA799BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2682298ab8bd6bef444da950c564297113454e28
GIT binary patch
literal 1294
zcmWIYbaP{4Wngel@vO*AElFfyU|<jg;^LB`%vX#I43R*&iaC2H#plm95II`;Zutji
zbzR}!*>Cn=4Q)DbxFw<bMYO5n@jI$v2i&%KhKK%sw|hc}z$_j9$!dy;|IdB2y_3CX
z_UYx4lUFsw?^KQqcHQJv|NF_OOuI#q?)UpLE?q1?V|#5u1*=NOj>qLYUblSTD<H=q
zU!}N-{nYWePW~4E<_SAmxTaK35ZiY(VaMdC(q-qH4@_Qmv-EwG%g4a`Ys-yib@{t7
zOEw=-*FSEM@KErYxqZV&!Llt+6F(aN>a=U;i@n<0{QOO{e$!f!4Hr&H>O~dbn&Nj&
z)g|tfxxkFj%h!IdTUwN0^({2_E87|4tmj4Sr?MWujY|pWjTMkla1(NCI%p-cW^wj>
z7vuRp%Y&EfDzuh)oaVSsVBYIrr3Q-mCpN{l3O_L{3Ut^V|Kl*z<n9Aee;%2c@9+=w
z2=sV9_r+uu$0miFjq)$J-W#__9(tp{?nUa_iyN<r&M&V_Iy~d`vzGY%`xTCLy_1r#
zR-V64*XCF7v;!JTvtt=!qvHy+w{N+|pHQ@};%COvrA+k<xj#3x$h;45^js>G8q6Rn
zZN;~2^{$1x-~82%FXFyj8vB2~$;I+n;lIN^KT?Rl-Z#rMmb38b7Cqe^t8dPcS}U8o
z=H=(Ov(7adSwYe14dMR-lDH2&{Qr;rdid${#g(~9b-<Vxt?=6Pv7z;o023=GkP8Id
zR+?7U3>*-S*Y>K<@vJd_*_fC)7`Rp0RTaSm11A$BI|BnFH)E6Pohz>Xdp`>>vB30e
zv1p0I^e<xASFl<0wTUPb8z%#|x~#f7n1D$arKS}tGa+2T$O*O`W-E+l<O193sZ!QG
zahG#lh@X)9i-|U@96a25%6ghGC-89VE9)ci43rI!c!tV`NIWAwBNK!^V?AReo{6#v
zl1fu$Q#dbnyJb#{nu;<b54Wa_rUuM_)IL_#2BBgnps=d8syWQn6^9=_sbaZ!mYIiJ
zM^;A(rkaP_NZAOkmxtR}*%;1)sR4!qC=_8Pb$c<lyzT#Q!O6p|E~+jA3sj)@we$>O
zdVt>7(nsPMXc-{!47CiAct)m1HVA#jrp8D-6D<=Ym8M#za30KF7A|lk!dwURo`H@P
z!Z;&)BTod+*xne<TWf3;aCC0MbYMVP$ylkw(%w@8)-UY*;@wQZU=`GqM(DK?w33D?
zVkk<@Nd%Quu+V0J+0Rg1lw7P2O0QgyL<=H$5lkKgle;lLJ9l@%+BwWTyj%<bWry=c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_multiple_d_xdl_cshuffle.hpp.18531F18DA84F612.idx b/.cache/clangd/index/device_contraction_multiple_d_xdl_cshuffle.hpp.18531F18DA84F612.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac91121f3a05b53d952a68f4de9d108ff94498a5
GIT binary patch
literal 14086
zcmc&)cT^O~)}OBG#$jlNA%hOdAx8yNlA@xtV#b`-fV#V;b=CEm)<6;iDvDs<2gnK%
z1kpt>i-2hb#33j~P*zYBMO=dHSM1ES`}^yiJ@0(y^BleNvwvOT)~#Fj-s<3yfq`B9
zMWTQOBiAllxOyHf5{Y!l-&Lzuet%IzB$8W-L~D;0PJ6Wck(W%l>dH`^+5dK^xf$;{
zGI}}XTd@dNF3*y#(K65)8fP*yYq>1d^ZT;RPh&;|S=+STx1Da8mFzKKrPopK-OXJu
zKR+qm<F|27ig(HN?H2bot-c%ow4<kwTj|!wv05z)Mp)>{y{-Cl`rCX4yq_X|G9@&-
zx7PvZSyOe7O-|fd=yTzcc2;ly({=iZBc4pwo2lE9AME9O)oS-2Q-66jPIIhTL$Pjo
z!^pH{l?|gBk5s)~ANJfvS};6x_EXKyA0loT$R>Kn<Q&@hYo){G!Uwb6QgU7_**w!R
zI;`rkbLS7UR?Ltrs`PRRAH8Bq-ejgedG0C6lLMh;XNz}VnHga?ew^l?>psL>``Koe
z=G0{OjyH^DTX9~9)4Nd}6~luIckN)FoEST7`H$;!#%9;frcD#PzC<M%&er!lyrzsj
zuw>BVJ8NT6{`Fsibxz)GO8<R%O#b-e2W2Dkg65C;{KK33UJ*;Kc|JY2#C4Y4ir^}5
zk9q$}dz9#JHL2V!*Y6MQ_#fpf!uFVFzx%!+tC!zi$Dc0x#)sT|Dc@dT&bGQ|zFfKD
z{-VmgceYM?^k|j0*Rsl`cVAB5xhcT)j>|7Ovvwbt_2Kf=W69qI1y3$8?lIl>O+{?|
z+uH+^f3e@G9Q)lA@q~4c%8qV!h;KWQzbD`S=lTQBdnB3AFN0hsKE1PH#*^4Tz3dlm
zrn3XMkU=_=)(*`K@lWjBIL<RQLnoqj(Z>97Zi613HXNlN68po_(CumGq~EXJHOlZ)
z;LU&qL&Jj`?AF<S_U*audX<0ukjv%X4|9Lmm@?c>H{hiGbAL1URh{46cJ$u8-@@V>
z!;eMLilK#?hfnJU^-V~97wj=aI^){D8>Q91=EZuB+tzR`tV*%3bJwk9zcX=@?ybuz
zTHMDh+;rEAVWXXXeLBtLRdD6D2}Lb47vESta8}Oi?fsNmMGLkM{bsk%5m(9c9J@tU
zb7owMx|_V<H#$c=Xv6#|?l}dGx9e^#T3)cwN_*atHx4i6Alb+G2;cJ>V>YhsGa+bk
z-r>=?d-^b<i1Amsq7BheYeq+hHC#O#QgP^HcyzBm+kXg38QafhLZQ1&>GBUXlV*Qd
zyyoW2hMbV)R{Mei8w}3q#n)wr-d+`V^g+V+W}%}jz2oM1^_^tqxW%)-jf-ig13Lf9
zq0v)kxJ*jy{WkwV;?lCRz~x64c@#Dt98)(oGX;P5c+(;m(}#gCPXt#f0xtY9lRW&f
z*7Uyp%$Fff)6>%W-S_+y*`+sRfA6MIzt{co<G}@!7EfqfFz5V|Hw!8D$p`J5vlKD=
zXEo<#Dqb2?GOer5<}Gs`(>mvujG){>t5a^x{$jjT*;b^L6L@3L#fM|A`Mg^GLD~1w
zcgFFa`mWE{4^ZkIzA!Sbv}VsY2b`+nrybC$jt|xhv$LDnyHvkC#hvmMKZq&YJ5lpU
zLTcZjsna*f0t5QnER~*G>=hFFZlm#>8@is;2KZM_HyZCx6@~x%``%?k@}AiBv7B5t
zTXRBeSxuHspL2$HMveaX<ouR1d!|IzeA$(j`J>V!v%N@p_56;e7usW^ozC3cXFb-T
z_4b3iN43pga}rI56~nh*dG9oBZpiN|?rAp`ig*6boO|*7{IrA?=NEM~pEWq^vC&Ja
zZ2NWH?j1(`x_WSdvF)y22Tfj;%14%jyk2+z?LqzXeT$2Ww-lK!%+hu<e)VpJRc)Q}
zkDD8{2WTHJ;<h~6Qa`ol6E~kTVM)e{p_W$bYRckG#fvNte$tv(d=%~dSAzrYVltZJ
zy|=6q)z3U@`_r7m(=!*ot8lODiriJSJn6dC!@8@ldfh&4xyGa6>2tZ}u??{~zh1X;
zkd65^^~U0*lN{wQv*XT9yE7x*X2)u?ONN#OW=~8P*D^~z<VF@}Qw*=J-Dxhp^rE84
z;9WIsUb5%>{5|zCPi<DD^xtsv%8%1J+*&>rP$T@!(zki1bksMj+j}~7Ty|9Gk|E2V
z*KLbim{BX6x#Y!KY4f{u>#iLW>xPYeKX~K(yup?sFPsxbSTed&XaAL9cMIY!P@{(G
zuGQMuJ~#VpqILBC+Mo9M{kFk2aO?izP3<x0{loWbog(x~6TE&d8vk81_9`fjo@R9U
zT#);w5iinfPPI<Bxzl1@iEJ}dV5s}@seXw}sk47>M@8AdSsUZCcK+p`Y45iF1KS+X
zV?@coAbH7#xcu5PSJ#Q2ypL`EuzdQ)x=B*Qm#t?{IzKe9k65u~$sg}#1lESnc>Qoj
zUQ;j0m5<_Ie^~#sQ*_AUh1MAxzf@Qh{KB|fv|F${-W$)7HEsP@$-q-j#xK4u^Y0U5
zwJGq`xcLpHFSy?BP4l06bm5806QQftEt`K#ED~YmlBovYoAuH|=KV!nD3XODJGBsz
z$K8LlpMU$TbM<}$DRLSh;yz*XPuNK<fTH8Xt*G3k<!^`#Nh_&T&s0jT7C<pbes^sD
zHna6oWKUY@z}g+yTrGfNcIdaWans7;43QCO<rLNP6lp~zfMT(^$GQ58@_$Q_A8Dlv
z+jQXpY5^4MAzH<<aY4~qNJd)8Ly|mXtQJ7AZJpJ5w8qCyiu_3{;j}EA?yVL;ag)p`
zNgot*RfZH^B5oHg-9_uF1yDTt(wPz8962aOHl&pnY~Dijq7p#qpI}(HuB#zO3$bJ@
zNt7gs(o+kd4A?jBLVf$fMKYvIC@qYvg)vhLp!m0~E4RO~?2{BZkdC&}vaPh0S^#CJ
zbCS1#<Lewf)RVMwl98Qc<ZAJi^22;RxsKZrDH=#xsmDh3*j6ooGCcRf(dQ|b&PmY(
zLV1EspWxnV0Tk}OOP<e~E3XWZ0qH1(>X|~;NF{(WhPq)DI3WBR8R|tScNzLFW2_cH
znb4e6FH3T6m7>0+qh@T}j2+YhC^Hv-d)+R-_9rQ_C6op%Z@}(q@s;wn>sfUD95*eb
zM>@KQBo~p9S^(vnmyzGb+~1&~fqD>1C(`Of`f34`kbagsBLcMorO1tR6iUlNiFQ@u
zE2Uxc1=r42T8f+rr4h>;$yihZC<{)`JCoP8kJCaD($Swt@+UG;3!p62e{pq1<oS16
zs0T5r9kg@@ZKM`JS+t@)B7gYVw}wccv~rT_d6H;XC4jQ>Au2kP=tp*=3mMBxZ1j>W
zqe=i}wco__*zx^bwU7pB<u@exjqCuG0Lpr|<T=>u_tz@(zDG;%5%X3Fpln?<Yj*3o
zsvsHaO*$&a+T~bJEr1eH5u15v%GG~qAx+XzE|TPuj#L6Dk-;S?gST0jNKqfsN+;It
zB->skzEaxmaKDUB(2*iXLTSg^?S!HdK-p$vRv)tTnwb>Y5lS64sw0zD383uE_%e9@
z#e{q*nnWnivHo)`R|}xTEWbM7O8dDCDVj(q&#=og;yhI1D<#su;HQn9C!}aFq5Orl
z{~}{i382KbSJVVMOl#3Y7G%;l8QD#yms$WN{^`c)W4sdv=pi#gIfZ1WkiA*}<<S1K
zpT5jV9ixR9LTNyf24todKsmm%^yj+u0dX>9NhnQ>RTJZ(7C=di?c+ip-{dbv9%Q(!
z*s>M7sRdAy1Fnx=z9gfk7Lt-y&LhcrVwfrcl#D^fn-4!IJgI@S`iZzWsz)5<sum(m
zOj(OW6i1y&Gnv=JZ9y6<8s{nE?8J8NfBz%Wq>*T{J0YBV`yyzD*<Rw{y~$W!VdGcW
zUM--bJWrde<4g;Gks=>Ld5hV%*jFupa;^2?qG7=%q?KSod5TS+5*MTrK)KOXJUgY}
z<N-ZoN@kQr$+9RrwE#-t<hAvklkUYBp&q1_E+p<kGPU?hneX1Udn7wo56MX@)s(E7
zw4xF~DUyej{`_n00YfyNP|~TM>4c&ZKq(n`+;peEd4d!TA{{kh&n7%yEr3$yoiJ4^
zrTL8pGA6A=(OOZoo>~B<a&ga%tsmn;w2%&=d_<CusGnK@<<Y3U1N_E%jF2G{LP=ya
z6N!UY381`i`Lnd`;~RG^q)jOAk>ow<uNFY5)r$T0vvc)lDRL#l{erc>5c5`vuas>s
z{I1O%W@?D6NGr*dSu&ZlN&uyEw(0j@0&*T`Au*x6MUuC~c2xo>T|1wD_qx=5t`zx_
zj>2gBFxp2gzEbK=pItfZ&>g+MXQZTMN?O6G1W-gP7cN@0#a1LjE?RocF_&&oEZ2lw
zn@}(e0m2tXq7GWXmRT+fS!E$F*orI(0j%?oFGL3{7mFQY@nEnNmP^Ivskk3tjSKxO
z{f1{=!Yo%pIh9a>V0A3_C*|}f6<DaYAC^l%P6;RwMxxy(>hR(NeIIGD#Bp1kq-?;a
zv*ehxNTF=t1SLIE7O9jC#7-;;bu3O%HV_4|Ts{?$PmSa|zCZ<B5aL=Ypq3iR52TI?
zs1xE{C}0;F2?JxfXcQ1F#5st~L3*Ge2=frj9mCRNxEH8Gq7L@pEQJ9hi8{pNb$*nd
zdz7C4HFk;&Nx9Pg&|HJ%E>n(|sUiGSu2Y8BDQ6znQ;zl25Pm9+lwl*~%;UYtaW5Le
zPbC%^#v*6H5L=R=U&icZY{m~=6==fHO)9g+ohNhMG0h#Ny^qp9FjSf=qm9dG6F3)X
z?gXPa!C1q#gGHev>VOpHY$4t1y<L3wIkC)Gx>qc14=qBZO2%{x%WmP`{Fqc>D~!qF
zwcT%;Yo7U&78D|dL=|(3L{|ALzk9RAFOs53!$z^RDKy;M|7_j|uTsfBo&%UZEm4QV
zFb9_VgxF8W6x<NYwV@GhXfnh!ED00E_1JKzI)a4onM|z&vnAM^pPDL`hN(F|PhQh}
zgOid?ltYPgC@D;i<!UH#jSxp5aRmDM++{(huNnIJIztoXO}vFCLk~3ffgb*W=DO6*
zgXTKv;hjPp#Y~A}rmcy(60~Fe?;SK3$Bc_(CIBClNY+sm(}Q<*A5phM!*qIoy1R$o
zA7}-<(cMGuP9ctB9HST~VGoKp#v1sL6%ged#nPj~ZKTQ#Kso*1^sioe>-~7L?l!cI
zA*cpY23Dd@6!=xjLdIFh0`@D*#bApVYzL}gxqaATpAg4li&!B}#RjQZ&hvNBraNdW
z2uq1Sq771M`PI8~j6BP9l88T|4NeJh8f}m!#Obs_x)5j31{pkt#D=I^mDRvRe9W6~
zFVUFStV?rBY^1~{5NXm}1vaV>;wo%kg&q0l)!3+7_`Ho_+ZZYIM04#7+b+Z%jBy8J
z%0KU9j5~$+3uF96h`SimE~XdH|BEsGONh5<m~PP!Jm0Eex>blnHS|L@4EguNH1xxS
zI9$UqT*H}v9;v}bYJ45PpI?jC8T~KgXtF(Mof_I4Y^oc3*3bbweoX5;7CwJWdp_o$
zL&8XG@Cs(H2>qcddjuOCvd(av^M&|)LVSiTo?&MQky-8p?R0|n0%L&uk}RewxrD_8
zpFn(kESARdzJTS9;E_l0bl%*uuu~TH<?%Ueat_<`xEkA6V=q{2i8?ohp$@wh*D$GF
z(Vq;pkhUqL`@>cy{*|`5#bZcf2{9T=qp>mcAW^5X&>|efWKOCC7Usm2AN4&lWB4jf
zmMf#!GRlZQr|whieIYKV*m5Cmr`UGNh(EVFD7Hh0J1MqPh?5YTgpA-EBU(r72_a5K
zY%-67a@T*H!_4|ZTt+;l9Z%W8h#}=BQ&HvHUmuJcV0?4<!3pGzVx%ZWwlmc^Gi2bz
zH6Ov!BiItO#B$l#DH{(2EK%p>ptXsb!!IpcDBDe1Y+*E87#*<YNn`cw{)q8oEJ^4^
zd1O)Eyw<WQk8B~XraY=CZ(d6eDUXLj9Ev<bkvFfEFys*?#A(Pp4O#P!Fddnv3vnox
zhw=w5%Y|Wim=K3!dAJZqV0nZPM`F83>;f8Kxoy~Pn-E7~d6ba99m}^1aT;Tv#<;?s
zVYzh1K3#}282bz%&SdN}h4?h%ahmbw-_K$^vV=IBvCkIr&oK69gt(MZlrj!r6vT=d
zMVS!aXB78^xSUax3-JTS^#S9_zprFml|o#>C@O^fN=8x1<0<nFOixdE<U=fEyN3RD
z4I{9SY5K)|=VqGN5SzP8v3DtZ-a<Z6>?g{ew}|72J&x?b0-zQ|w)6okJ%FuYX(Z~3
z5Ev&|x1K~@2ZDXTa*fEc5jjKJ-JM5RHX&!oBUr8(={6%X$f1aFBg+=#EDVV(+mJIK
z62)P?IH9*Yi1iK%@eyowL})V^SSN!I23YPaHaIKXb?33cc_Ds=b)R7~-XHI#WxHuP
z@AMLBgGAaA{ByUzHakH(@VJ21DWH4t{92l=rA^=jnVsvsbe65Pf*duS$gC4NfCgDo
z1v0ruTSFksk|a}+$hbiu%yM@agFB2R7ze!9L!vIL@fug9-9TVTsY|hgcH+YgCGDuB
z-C;yz2F!pHj6dH|E~Asn81uWJhOw++?4jd1VGBO5ae6z41Z*|5RgEw}tDysH=n>E`
z%{`{A9t)p8rUM`I&*yJ)(R>(Nr%TM$nzlB9I1I{y#O7{b_J;5bRTT$eVj=3{AR)$M
zX}r)lRV6|og6&|5x-iJAsu_!$u@U%Wmdl{cGiYa6Oi&qlQB^gBpp_*-tjR9gl3(RA
z+Nq58fW?E<lVobDA|l`8QM%tzdH}q*BzTVN{SR~A5gk0CWlv~hP~Hl!cRx-nUu{et
zB;vk_cyQYKq^zCnyYo<nWV*vn={}+5AH>5C@)0}BWn!mH;kjRd?JI=mysCH!T7u#w
znXIa0$&dLBV&5QBehb%OaUC{<+ThxP=XLApvfs5?E)D6YArs;Gpr0<pTd>g<;a-fy
z@<?n2PNm!5%eM(}7uN2=dXUkP^Muw8qxGQ1K>}x{cPeAYTk2`X{xsvl<1EHLONj3=
zX7`xhV96}^fKfc)Et}<(j6%s{SWR-ct17gx{17UU2deTc=rA-t*l9;i?k2KoHf}aU
z!K9&FOH`{W((<Z(huC+>l=p@&vG^r6ghhrTE-~@zn7uCC3#w8sbR4NJ<q{&PoLmwT
zRRvuj!lDuLiNVqsq4~t&;c-Iq$-qt-Li15#d!^8PRApdjcxMm}f4k=|FGrRXtqhYX
zd9AvhMK%>>kwMu)QLh`@XHafJ+&6>r;&B$`lO?QKWK%xbLR?MxR10es4=JCALL7>G
zLWMPpFys>^#A(PT4LJ*~)FxesL$M+hJMvZ;h81B#9F7&?LL7k=5kee^T_gE|R<~8V
zZWH1ttcViww`0Y2Auh(UVm^9i$&HhEE3E^$1xxNo#G$kf_{r{Fp&v%~<ab3lEe@x3
zpjg0i2Wh>7d>Mo6QpPQf>Cf+ybjB@Rh%*?s3?a^B+%kpuG~;uc8N$Dx#rR|iaW><Y
zE##kJ+|CGbDdSkm^yPP98RJ+c#P=D;`$AmKIF<|X1E${t#*cqr$@EhSaRuX8A>>yw
zj+H!y`wFsE_G0N?UeOZut;JW1ZWvba-J+G9|8-?z*H-VtX+O3u{@ckrKcV}AliyRl
z&P!W#_HDPjFtTw0cLA{n3D@>u=^kvtn_@DSCF7oeCF)BQ7;%jH5`|3d9%k<ey@2W}
z1&I5WUHWTb+3bO&;|xliK^eg2i~U_UYMV=a8hKDj%PMJ07%E)85SORAc>x!8;JM?~
z!Iw`okT+TxlUBwWg3be<=gtn*kt<20vx|1wMSH`U*NsQ*qNnn>gf=M=Mm{BU{}OsA
zxZk*2BTkk_ZgM8Gs$mRj7z=*o-!lg984I|tfgMdwUDeeOjFBbxE*|I6D7cb=lZ^Od
z)ddhwY{23M+zaB~?$V5TD%}?Xf@2HABR5X`bc;;31-Z7M5C~I`Exxtv%G6K`@}L=6
zG^5E7&mJH8)X#jvC~a|f&BaEE?cnU_#@<To%i{`cQz3j_fxRn)&#SOk753rzRd{?A
z4(4$+wy74rUyZ%1`S(FdWC2w-S}^CtrI&;DH5;xZL$9TK*3x#c<Z#DD^nM&mj|;6y
zb>RipbRyu9+0g6lNB-jif&Pi!RTp6}2A2DP*bk^DXpJTJSK?RL2(Gr$*Ir=8FG~z0
z#g1~yxSX=!6PR+!ww!Y1aVKTmNm=j-O($jBNxAYk85t)d3rKLtVTEjykt^T~51;Xc
z$L7Y9sYKE8DB2E=e3G{?22H%J!N`fNsqXDYTCdp|lOM2&CDlX2B+3uUiP^KN+!J@{
z=#g!p#JWms2n)uNqML4o5LaPE6~BNiSB-V6`R8yCNP6tP1k5__bY)?F%x~lxgtktg
z$AXETsSP^sUX<iPma`4HwjmCp-*b)Y&l&IdaOxkIuVE30^7muuery5mhg_^+S&A@T
zJB^)AV?X|R9yZCt4m_^HzEyZMXh5RAr35?rt^b*W8_M2{AeHx4M!%K$n!m#JC2@#{
zu=EgLBVoB*EX&3I(1S#MsR=FSHM(pVBlnO0`ywRxGBt&_lk3!&>(m4u*Hgju)D+%M
z8mTdj)C3;yMZtT~6tELgbVFle(FDNxQgLa~w7N0>@7w%>pOtQ1RBJR*RRU%cuo0++
zICRWbVk1yK%SAHcNJbBK`^7n$<LZhg<`BENh)gdcUr6aM-c(M{@csDmAKMMwD3iUk
z8%uW!4_ei|GOSX;u%-p<r)v@9(%}+zyM+6|m}zpW<aSwzuVc6CLR^5|3WWF`cDu*p
zf|T*gwFAZB|IsnsXk9e<^vbLMCX+uv*&Lt-!}bRIARbF~-wh8~k|bIGhI}Dw>JFzJ
zV)$!Km;!OP|2rKAiy|$mF1`5{bCGo}@`YrfJHK<-$F~SaG`U4mU4`>6Dk&W$C5N%H
zTo<L&MaiKjavzFxP9izjYVqsCTaNwFzf6nfIvBP?ShnwE*iIq-!mwY2xQk)Cg!nIp
z{Y!|qXs}!OvO3Le)nK;@ai|6xD#T$LY?u&-Yp~%$9HGHR@c4GmV*9qAC%co~Tu7N0
zQtteWZ&Kzrg}8|_Z=&3J4K!2c%|aZH%;S+eXoTesBlE*NzOyoQLPzbgRsZk1`JQ4!
zROq0II)5J?LzH+pzk4h2kP0EL!qckoEdF^l9#YLeFY&Y87(7lM_m4L*d}l%Y$zd!#
zEOhQUSeAnafuWPI;_r_zU{jY)U2O8_nq%D&kvHxA^@6eVd)=ymi&klq8C}J)tHQ!t
zA(j;i@f~b%hcD`|Tq%~7^3UO5Cfh;vX$QX@ZZdW^8CQr{S?)b!`JS<Z0H_-~yl46V
zF01%ys*`2W<{sTqVz2Ya0iu-d(xtc&89}tzUAh!EAtOFne~-N1BaRPRT9CK}89_|g
zeE^HwkP#ff-Ibi)iMS6`a+Gdw9_37)yPv2cm(tIrtax*NM(IDJta$U?j`X)9E8bi)
zkxnKugLoG@BrZku2@9`^CM<5k9<Yd{AWoa7(<32PfRZ0^QmQXn;Kk=&B>|t)O{jlt
zE}#oC6xDYy&?3vd!QwaA6OM9{h0^9}bPx;$K9C`<x&X5U!eCSNnGDz!loj+P>W^di
z9@3Co8uA9MkjWyqbRo_``Z>JTSZ)h;*n&O4fXRV~y&~~Y9&f{5+k`j)_e>BPWHDA0
z3oFf)xMwA{=C@!MHtoU+Vf9KLM%(gvdL&~M$$ZVQx~p0JQW<|fuRhHTJI##caTYTy
zONj3=eeN+nd^Yre@qfUK<gt?RSMs>dB-wfID;-y&;WM=B8QPuCzjA2T93H>y!?g6!
zG2TgtcNon(j6UCD6QkM0=)=ybch1Z3FiT4$TRe+4&Z1qQ#Wz6_)7hE+hyKqU^@pQ*
zSKP9!tpBl{ew4j_6SjSv=YO5sxBgtT$*AL+&VSv|=Z@3q#tqA&iO$L><1)&L_XcIu
zfHG<nkJ~BZcFKwO4(-%{c4`!llaO%|a^k&35*m<%Mgjga%PQ9-BCj7=wMOLHh=Ret
zb#(lC#jPf;g*+%m;$mb45g<4`vZ+-coAFz&5sMpfFQ~+Kf4N|Oicgop)QKhC#<JV|
zO`SyjNg8Zw2ttV?QhkU9J*-@~blxyh7a5kW^}^uY=BX4E1>i<j1WdSC5kZ@f%8B5l
zShX@_)yeJ}O8^vFM2J(&vkjc9%eHBc89O>)SkUmmkt0S887d0;N)Zj=e~#jRj^=+3
g=YIzBKS%OENAN$tmV`y3vHbsz;eU2dR{iIH02BNdJ^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp.634AB0569FA8301A.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp.634AB0569FA8301A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c01e8c3066aee8b6e9453358437b7bf230c9ab1e
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-_97~XR;FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx><QT=Qy96v-6as
z1KWd7cPth<R7A$C`E<8kc;dn1C*y0XiYvcXo?4%F`18C6e;5DFxoB+=J*Tg#>GS4!
z4>orFJbPVb$4`sBzrLyK?)z)|v|E3>aox@t@jRXfelGvnq;D5rH}`Jc&S%rFt5ov+
zFqygKz#P$aE@3PST`ow6?bNN(TVwEh0rTW_Yhnbvn`?USElNBcC4F?u)Ma5mntto=
z&-`+Ud%u0w4%@R2{3M+HtzsATZgE;&!sn)Qv1QBGdq2P2TX6rl)9uh(Zx$_oa`)-E
z=}NJe!rVDBilU;~y>DeoH=bL_d+TPH&i%c=KgbDfd=})-B&wCTcpCeONTZ(*HZ9iG
z;)s?_W#lRjoZT$dI@MiM=^)oq!8n&);(QBMYjOlshVQCa&sToL@RiU@w~l?&F8;jW
z_(Zlc_|-o-CmnvZVBO>DD&gDM66fBXzK-w0D_bSE=$4?0Q^x~;$nI_K{Sc+0%Q<z6
zXkF;?5*d!ERxDgLPo#RAM6z{zmUY|CiRsaMr)_gWx6)0~v*khKjZjUV1>x5M9ik&u
zr&z6XZivaa%xS~+Mf%bb(TSH_4wpNu*{f<VI%SI7qG?7}GTSmV<}Ckn_@|@``^r#h
z&6~AC+n-O0y1lhDnmhVo?x%H2?<P7v;W%w|{-U$&5iaWmd@d$aH*Z+gTOoedU0rF~
z)P-G{t<fUcH;!D<)%0NBA9N-|{KI^ywM&GGc3mtgXzaRXUg11X?BeUG;-M2mu1uSg
zdGC15%Cz@^CzK+UL|1aZk8+a?c=){g=aNdlmotM(XUtrp`c+D&cXD=<)wc~IMZQ-(
zpY-UJ?y)#|AT{7<)0zohr{=7-UFh|jJB)pP+ku~wi{&<Io7qUu;mKGM^iMD7-oFzG
z)A#1gE8VPGaWd$<_4(M=T=}A0XAu#HZL6{_B?v!c;|P-Yli9s~!%>mEEiV699z5;1
zEg0Yb!%tGLQe(vyk3DDFpYzLg`p8avzPzNYc8l4B(5Z&20-nv)dHDU{%LgsGTT@%t
zEjgohJ*n3?YsH6TrthmXLi0U~9~iwV{pTiTv{SDCx2I%(boS%I@*JCU_qrVX66K7)
z9XhWyIr|qcm)G;xqF*0hxRUbvd$FWm=9iaUCM&mZ>1Vra@aA64YG%3qJ;#dP|94Hg
z&{=n98GqJC&0q63zvS(GzUhnh&cFB7BA)vhpWxayY0|zs=U93SWG-eDZ~WPIr~S+J
zi=uxe5-06bd=l#RS?{-a^_|cy48@hXNp`^8-BTZLvEf@hD>qYM07JMqn|Lgk0Lp;D
zj+WEMPM!VzpNGL8D6Pz@oB$?Z(nYCh#qSvz7`T2dxYz7$8qC5IF0U-F2Q!o>TtQg@
ziKlI??FKh7T*q1m&U0*k!|?M<$t^~fa3c<*c$gxu?Ny)SS!4b(@`S4>s93>-L;Qr)
zUre-N1qpMi@WX^Nj!l&|t`B|9#1pQltf+x-g|4-(158nKf|T28L*p`bo^T0e33Y@b
zNo7eSp0u(w5>HlH7Kx{Bt!@u9jwf8hS_6rvWvvD0!OUjj0*7huq2CMEJhWEeLGdEY
zAVw~5$ifsI3bA_`FMs_w(2E8v21+nhPYqbVu=9&|gM2QkA`cUOb3kh*$J$saps)ms
zgc8i+FB?8?if+892@DV!Wf`Q1(6rWs2mN1xuuXUKk33`M371lqQiJK`371orL*nUK
z>)FAqV<<|^Nd%>V;-ci@U=RaLfEWzS;WzZx@Tf^fuVRY`3JwViWe5X{gNYynGX%j5
zMKB`}%;emZoD3k>JI$Awg_Vt+gOi(wSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!
yRNG3~OhX=<sF<{jtb(GFimJM%mX5xGp|PpCoQagKk&CF9yH_NWo1dcxqYnVYA^J)H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp.1229C32BCA5E8056.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp.1229C32BCA5E8056.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02f0d483b14da7fd05ead248db17e3980c0ade73
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-_97~XR;FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx><QT=Qy96v-6as
z1KWd7cPth<R7A$C`E<8kc;dn1C*y0XiYvcXo?4%F`18C6e;5DFxoB+=J*Tg#>GS4!
z4>orFJbPVb$4`sBzrLyK?)z)|v|E3>aox@t@jRXfelGvnq;D5rH}`Jc&S%rFt5ov+
zFqygKz#P$aE@3PST`ow6?bNN(TVwEh0rTW_Yhnbvn`?USElNBcC4F?u)Ma5mntto=
z&-`+Ud%u0w4%@R2{3M+HtzsATZgE;&!sn)Qv1QBGdq2P2TX6rl)9uh(Zx$_oa`)-E
z=}NJe!rVDBilU;~y>DeoH=bL_d+TPH&i%c=KgbDfd=})-B&wCTcpCeONTZ(*HZ9iG
z;)s?_W#lRjoZT$dI@MiM=^)oq!8n&);(QBMYjOlshVQCa&sToL@RiU@w~l?&F8;jW
z_(Zlc_|-o-CmnvZVBO>DD&gDM66fBXzK-w0D_bSE=$4?0Q^x~;$nI_K{Sc+0%Q<z6
zXkF;?5*d!ERxDgLPo#RAM6z{zmUY|CiRsaMr)_gWx6)0~v*khKjZjUV1>x5M9ik&u
zr&z6XZivaa%xS~+Mf%bb(TSH_4wpNu*{f<VI%SI7qG?7}GTSmV<}Ckn_@|@``^r#h
z&6~AC+n-O0y1lhDnmhVo?x%H2?<P7v;W%w|{-U$&5iaWmd@d$aH*Z+gTOoedU0rF~
z)P-G{t<fUcH;!D<)%0NBA9N-|{KI^ywM&GGc3mtgXzaRXUg11X?BeUG;-M2mu1uSg
zdGC15%Cz@^CzK+UL|1aZk8+a?c=){g=aNdlmotM(XUtrp`c+D&cXD=<)wc~IMZQ-(
zpY-UJ?y)#|AT{7<)0zohr{=7-UFh|jJB)pP+ku~wi{&<Io7qUu;mKGM^iMD7-oFzG
z)A#1gE8VPGaWd$<_4(M=T=}A0XAu#HZL6{_B?v!c;|P-Yli9s~!%>mEEiV699z5;1
zEg0Yb!%tGLQe(vyk3DDFpYzLg`p8avzP#j`+*;n0(9|BSDV6HpKWYoCEE2tQBNu9G
zZa%p+;&G3Z)}I*%_G)=ujlLwaqo-=_{)w$V>yAIJ3p)1r+O3K^veNyVw<}Euo^iZq
z?}yE&p3JIiZo2fd%Jr{>iRsN>e`DIbrD}I6`M%P9yZE39&u;VkuNsakuitp*cl|_(
zl11{{Up7b8bp3mtQ|0`4r_Numy#L!dukF0dv#DwAk|n?l!l=ySZgMLoXP@GC#lN{F
zuJ&#-mgI9(PQAG2^gq@O0ilN(iYs%I?0~sjV8hbyb~^%OxtRh37{bNb#ACq(PzDVA
zznUh@IMUR_!{85;R%TUB0246jqSUnF_lyh-T)!6FYxXt`X5k5!SC-d<8Ojr`psawz
z)3(-jgPRzxW32<{!E`WjfvxX7^n1aYht>)_;YJ)r@h~lp?Qa-<ekr-d$bus5wY}<d
zJZsEfMxJmL1r;lpY8cJP1$Hvb$Phmv^%oOuSV5{eRrq15GLB7^Hm(nS&BPO~sH~^~
za}Q6ruC=ZMOi^=!l-p`U<1%)ha0z7zb%Y{GWl1ESw6Zi3PgYqLiKlL@ZVxk#CtSl?
z1Bs_)tp(?8@7Zy#gkx$t(3uh}5=t<u4u#mgjF-QD94KtSVxR;w>8Szh7j}N}Zjh%$
zRpep9Up9Q)6y11H6X+=!Wf`Q1(6rWs2mN1xuuXUKk33`M371lqQiIvV6E3GLhs4vf
z*0Y0|%TSb>lL$%!#YM@*!5{{h05KSt!+S0SRzCa@yO1p+C^#f6lpzc(4km&S%n$@K
z6v2!@Fq3mrax#Ei?=)X#7FITP4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`
yQEe+_GYxrcqGHlAvI>eyDyr(5S~~g$hQ_Amawbx`MlPaa?p~2hZhnp)j6MLJTG#Uc

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp.B21239C1EC74587F.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp.B21239C1EC74587F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d809e778fef44de9314ec4b866a41ba3fc949d8
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg`p7~XR;FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx><QT=Qy96v-6as
z1KWd7cPth<R7A$C`E<8kc;dn1C*y0XiYvcXo?4%F`18C6e;5DFxoB+=J*Tg#>GS4!
z4>orFJbPVb$4`sBzrLyK?)z)|v|E3>aox@t@jRXfelGvnq;D5rH}`Jc&S%rFt5ov+
zFqygKz#P$aE@3PST`ow6?bNN(TVwEh0rTW_Yhnbvn`?USElNBcC4F?u)Ma5mntto=
z&-`+Ud%u0w4%@R2{3M+HtzsATZgE;&!sn)Qv1QBGdq2P2TX6rl)9uh(Zx$_oa`)-E
z=}NJe!rVDBilU;~y>DeoH=bL_d+TPH&i%c=KgbDfd=})-B&wCTcpCeONTZ(*HZ9iG
z;)s?_W#lRjoZT$dI@MiM=^)oq!8n&);(QBMYjOlshVQCa&sToL@RiU@w~l?&F8;jW
z_(Zlc_|-o-CmnvZVBO>DD&gDM66fBXzK-w0D_bSE=$4?0Q^x~;$nI_K{Sc+0%Q<z6
zXkF;?5*d!ERxDgLPo#RAM6z{zmUY|CiRsaMr)_gWx6)0~v*khKjZjUV1>x5M9ik&u
zr&z6XZivaa%xS~+Mf%bb(TSH_4wpNu*{f<VI%SI7qG?7}GTSmV<}Ckn_@|@``^r#h
z&6~AC+n-O0y1lhDnmhVo?x%H2?<P7v;W%w|{-U$&5iaWmd@d$aH*Z+gTOoedU0rF~
z)P-G{t<fUcH;!D<)%0NBA9N-|{KI^ywM&GGc3mtgXzaRXUg11X?BeUG;-M2mu1uSg
zdGC15%Cz@^CzK+UL|1aZk8+a?c=){g=aNdlmotM(XUtrp`c+D&cXD=<)wc~IMZQ-(
zpY-UJ?y)#|AT{7<)0zohr{=7-UFh|jJB)pP+ku~wi{&<Io7qUu;mKGM^iMD7-oFzG
z)A#1gE8VPGaWd$<_4(M=T=}A0XAu#HZL6{_B?v!c;|P-Yli9s~!%>mEEiV699z5;1
zEg0Yb!%tGLQe(vyk3DDFpYzLg`p8avzFawOUH>(<O{<LDL{IIRtop~!LUzuErQ6nA
zm^$U=lc+U?ZoHy(J`dtVm%h@?<co3Jd+z=TC$;O33+*)@7iQ<$yzA>deluF|gyxy#
zV{tzsQ;U4}J379s*}JGtX4b5nzjb<tReSAXC#rp&_U&SVRNLLz`K1RQPrM%S?(cpN
z9;=J}x4$^*+PVHOHru=KW6ZQV)$RXpv%J2O$rgDqEa=ko+cO&k+mxkp^~~;1_&wpD
z*{((M%08FOTlSpFxR?5$Y2LlieZ`fzNp`^89awrdW{Uo^f80!g0Sw{dY~rzC0w@Cp
zoA;YcU9+@fA`gQ<P+FN)IRQ+-q>ED1ir+IbFmU}^aIe|hG?;}aTwYmT4`wJ&xPr0*
z5>MM&+YN4FxQ?|Boafm7hT-Rzl3R={;YJ)r@i0YR+p9juv&Q^o<Ox?%P_cpu!^~&m
z0z0<%(C-Cn9$G8#py&<p6H<RM(S{YIms5owrZ?l*RB7Y-(AP{n;fl(N8Ze`H!gZ~6
z9bk%@6QtZ$8yc6f^Mp$%OQ<6hNh(Vs@uZcdk$AGovPe92Yjt~=aXjG~)*47WEo&_}
z52l8Z3mmdAzwDNMEOKk2T_@1>5-bu*Fc%yOv3nUWfBiU6*nq`A38wj}0qYlbe(`RQ
zUqw~qVZvWFeB2b>cu^DRR~cm)q=?Y8)`SQBUxBbqck_=tW9A8$QkGJK*~Akrr!0rW
z)3esIgPF@vl$w(WN(043$;H7S2ABXb7?{IP-xL!_&EKuW77-L25*Er31{MbsK?r6D
zf*FcnMj)8UxhXjrK(2S1FEa}(8#@OlHxI7>KcA<dtFxLQm#34ExUh(Xq=lumjjf%%
zgPy3im9m+JJT_4=X&G4sMI{wgbxkcDeFH;dQ*${JDP1EMQ89O~NG3NwM-N6H0Ay|G
AK>z>%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp.92168E842E069568.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp.92168E842E069568.idx
new file mode 100755
index 0000000000000000000000000000000000000000..176cf4a6bab2b96f3f0783d3a5d2e29dcb8fa764
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-_97~XR;FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx><QT=Qy96v-6as
z1KWd7cPth<R7A$C`E<8kc;dn1C*y0XiYvcXo?4%F`18C6e;5DFxoB+=J*Tg#>GS4!
z4>orFJbPVb$4`sBzrLyK?)z)|v|E3>aox@t@jRXfelGvnq;D5rH}`Jc&S%rFt5ov+
zFqygKz#P$aE@3PST`ow6?bNN(TVwEh0rTW_Yhnbvn`?USElNBcC4F?u)Ma5mntto=
z&-`+Ud%u0w4%@R2{3M+HtzsATZgE;&!sn)Qv1QBGdq2P2TX6rl)9uh(Zx$_oa`)-E
z=}NJe!rVDBilU;~y>DeoH=bL_d+TPH&i%c=KgbDfd=})-B&wCTcpCeONTZ(*HZ9iG
z;)s?_W#lRjoZT$dI@MiM=^)oq!8n&);(QBMYjOlshVQCa&sToL@RiU@w~l?&F8;jW
z_(Zlc_|-o-CmnvZVBO>DD&gDM66fBXzK-w0D_bSE=$4?0Q^x~;$nI_K{Sc+0%Q<z6
zXkF;?5*d!ERxDgLPo#RAM6z{zmUY|CiRsaMr)_gWx6)0~v*khKjZjUV1>x5M9ik&u
zr&z6XZivaa%xS~+Mf%bb(TSH_4wpNu*{f<VI%SI7qG?7}GTSmV<}Ckn_@|@``^r#h
z&6~AC+n-O0y1lhDnmhVo?x%H2?<P7v;W%w|{-U$&5iaWmd@d$aH*Z+gTOoedU0rF~
z)P-G{t<fUcH;!D<)%0NBA9N-|{KI^ywM&GGc3mtgXzaRXUg11X?BeUG;-M2mu1uSg
zdGC15%Cz@^CzK+UL|1aZk8+a?c=){g=aNdlmotM(XUtrp`c+D&cXD=<)wc~IMZQ-(
zpY-UJ?y)#|AT{7<)0zohr{=7-UFh|jJB)pP+ku~wi{&<Io7qUu;mKGM^iMD7-oFzG
z)A#1gE8VPGaWd$<_4(M=T=}A0XAu#HZL6{_B?v!c;|P-Yli9s~!%>mEEiV699z5;1
zEg0Yb!%tGLQe(vyk3DDFpYzLg`p8avzFax(*5a*=YgY9Hwi>-p;oQ&f$F82UQdhe~
zHPvjVSoWMi=T>>)9pSDktGs71PY+!G*?iN4lV2<D_@An{6D8fhJNa?pw=k9>uFo?z
zir0nBjJcd25MUL*euaGdGPUUa^G_FeF7^*!;<<Net;vkG#P83dWj|CrF`ZWZUtg%%
z@5^Je+JJ8Vi~r~N=4aSUUmAb%?SHd_R^_uC!!}Hv^5tonT0+ai4!7CeXU(73E!p3@
z?!|GA$zOOC*Bi|$zq$W%Q0Q5P;>z44J7Dfss<c?@mdh*7!{8sl5U$LsoB$?(GGNfg
z)OXUQH9mx!DG(?v&L$oUCScM<scFUU85tP3el57y>}?v%!V@m9EUyRC#S^ZetboMR
zw$^roTNtimtpn%5L>alj*7qLzy<p8lYX!&lHw-_&l-y!u2{+;}iifFk_efgS!f``}
zktbY&MM4Q?o!9oN&+)7=e}Td(3My7G)ggXD>MthRu!4j+Rrq1T8ONqd8`p=vX5tA~
zR94i08Nw5;Ypv@5Q`DRw<+j?;xQrcWld^<5LXo7hBoa?rSsIBat1OGeQ@2*PhZ)Bc
zu3@c##M83Yg7aWzGjV}~7#0LPC_X<FV)rs${`zsCUkq3blwjsQHDLY1&M)2#@}j7U
zJWTk@hL4+~8!u`Cy(ptBgA@^()|&93|0@u->2Cg!XUsg|Qp!?lFlX?D%PGqt@${_q
z>|o|H6s6`Qg3>^7QF3uGhyf-*3<icU2n8gW!<kHG_gkx+e!vzH6dV#3$`AyXV+cVo
zLlMjf1T#4|B_{*O^-l9;W?^Mx=iub#;T7QL^AvP-Rukm%bP^I57Lky&u(Y<ZwX=87
z6V<j-Hq(&DCMqT^BdegOq@t>>simWDU}$V=E@vX8Yvdv-=I#~A<mTt-!RP}3qdlyr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp.8EFB8D272907AC49.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp.8EFB8D272907AC49.idx
new file mode 100755
index 0000000000000000000000000000000000000000..677001601472e23591b70f15b5b0a898e3280e24
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<#~l9i@Ba%;wwm;23g5c^Wo(;P8M!?@xo5KKA3F=#IUAO4TXSLR
zl$%eY))czk6{*Yl8z=ho<yi~97`NDS_nm6hu0Jlc7k^xson`Z`ulM-P*%9Y8cO@T-
z`=Oay<h$R|F|uawqB?<DRylL)^fs#YTE|XQ`#SC0#RRE_yR-944?LcDJ>uP8|A`VM
zi{!UIYu2ji`u9Ais`=4Qoxfgr|5vZP`tuUc`lhu@miTWqW>n^JH#rxRV<+|f<Dc9T
zS9`Y^OZwS%PPy1~`X6gvy8pN0%G@M7VBT(d*fV4D%H&<#Oo0Im;o@xKQD6cn0|q|}
z-~LsaEVhG(!5=8C%&Hs<CScM<scFUU85tP3el57y>}?v%!V@m9EUyPMlqXz4SpkWs
zZLRGFH!)ntS_jS(T;C~Gy6)9scAjtvWeGKe0g}p+NIYp}X(XPkvMdr$-CEraVYr61
z1`<!pS_{r|Y=6V>^GnGsMwW0R4x<>D7O(A9pW|6${xb4}t0<^g!GuHngw$V5v|$Aa
zbE@#egfos!l{T&qea*xZuBfc2fpD*`wXOq95lobc3mm4shkh?u^Uzv>2gTk)A$Bk0
z<*y$HT42Cppaj$V)PVI1JHL20$X-zud6@8<16ngV*2YQ!g(X-dlwdY|+3;~wbmK)$
zpmSuDWspK#(^?ZA#xOG&xxjG<bIxCZuuXUKk33`M371lqLJCkhWjQ3Cp0yr42pNh}
za}q&mptvZxI2gnL6CefyLl}etkwFM%2%O0rZnf{PR-TCTbhe0~;E=FThETX1Lj;1E
zoSTx90pxn8`7*Pxvaxe;a`W&C@bh^Jx;m=~a(Oxli3^KJNLpB0+t}LKJLrjOTPd4q
t$YT=~la`THP*hS;RoB$g(Kj$OHZ_+sk<vAC5fyXyiez&0bM#>J0RV~O?&tsj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp.4BAF3FDF22005D6C.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp.4BAF3FDF22005D6C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a8853f85e194f3cbc8c2cab75d59482e35267d36
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<#~l9i@Ba%;wwm;23g4~8D;w9W>ItkgdY{6%pWlyNJ!hq^c8O}L
z*-o+SIf3S_^3nUlT`Q}M`<SN(PXD|;VE)Ok6?e`bt++Ews(*L#<HB!h*NVH|&Dbbj
z=QT6ta(+O7Sp51G@-55wqL<J2&hdQg7rw-E@6uY68EgsPpGC`lsCZ&Jt@?kk3ftTl
z$JhQ`z%_r-f6LkT7F6hY@1J!0e^yrMK9j?%6SOs72EX+=z;Rfx*I0aZ|FK$|dg<7U
z{eo&Q7dz^unmpcA&$y^DfVH?XH^~l|xBaD`7YGI@Z{}g}4`2vaW>t;_6F?a-I2Jck
zYwA9qTii^6KxuI{@hC6>lP*e4D}K+&z`*ru!M$d0(_j{!aCv2UJ(w<@a0O)rB%Zdl
zwj12Sa2;zMI8ShWr&Q^>SBKeo!X=a?)DQ+pDoY~qq?M(Sc(Tf}NIZ3GbvuON8rB*}
zJS}T2IM1>D4a3hbCASz^!i_kLVqjXlwpV?QXN~#G$P=!jpkf6RhS7{%;Be_Z^n1aY
zht>)qenRRmCfcxqRCB8E!&GG)n<{NwANrb!CtOikQ3K%(U29zjn4;}HJI<AGOic&s
zm0*!jf+;!_V)rs${`zsCumOvK5==Ernu!Y>@-PSRpg8`i0qYlbe(`RQb3|3-VS2x8
z__!&$@uDWsIWo#JND-lFtqBkMzXD;K?&cqP#>^8gr7VRMpmNG`NIX4jJ$Mi@6s6`Q
zg3>^7QF3uGh!FuIAOOsVFvDPC3=BaCW(a~A3THBhC$Fs*x^3GX%@z?991<4Fker*6
zlL3_XPV;4EVP#|I;N<4v72xOd6m)e~6Xf!A5)u~{k&v{ow6?Levv<%F)wWVL(~!p|
qDkd!>tDvZ)qN=W`rK4|PXl!aOXCkF*<RU8O?iI=8=I7|a=mP)-jl*jI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp.F76E34EF68D8A65B.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp.F76E34EF68D8A65B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9742712adab77adc5f0c1d896116b957af42389d
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<#~l9i@Ba%;wwm;&<jCo-7q2*kg&tdBGqZeC%YF9c&BEKROcUKT
z$!K;=x9PbRvk&#Jy}x>q&0fjlji*<f{u#aE{F5r1ch8e-@_c)b-`)7pqFVU1rEB?_
z8{GCweb&7+*U;!*AAW`3DY$*@<@2iBmJ}}6zO>}tB|9k}<_*=K*YZ`^e3Cl7_rIo4
zv)`A;*J=V<{V)EX<C`B~Gkt0N$+z{UTdm?}Jq+D2b;_5OWoij64?B*{>^|%M*lynb
z-gPgIb4>nnQDM5#tiqf7C+Y?;6j$aZ*#Yx5o4bhBNd~`g9tQsahHzz8<ybHQlmUZ<
z2X^hbV_kKFn<)?|EzTw$1twt9MX71U?->!MFmi!)_a6GaV9i5o1+HHU?lpUx2D9*l
z%PY(4!3^UGS5Q_!;%QrJyTRQNu4AnO=LxRwlqy~K>M%P`xP-EV8o~fcWl1ESw6Zi3
zPgYqLiKlL@Zig^j!&(E0r)8}L=Q*~&Vfgu_<Q5}KxDkg@3`~pH_NveEtTBHXdBRl`
zRIFgaA$~&YFDBZsf`mC$_+i2s$EHdf*N47l;t5w&R@6YaSJztC0j3Bh%ESc@OITp>
zpxC=x_OZyViFTbp3nW-1lwf)eh1k7}m%n}-C~UxDpae7WsR8R3c7E}0kaI*;<YB^J
zHhkO^-FQ(G=o}ek8Ke-`wAO@&@n3<kO?UH;JY(hwmr|BO3OPAtIV7H*wH`dI7>ZJJ
z5<zL8xG1?e7{mY*AO-_N7=!|mK?r6Df*A^DGKZ_2`|JJx=~->Ih@jw*uuz5wxEw=r
zZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$rrEI1l
sk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEglF7}_(Sy+k0CXVCU;qFB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp.1704AA4E0EFAB142.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp.1704AA4E0EFAB142.idx
new file mode 100755
index 0000000000000000000000000000000000000000..adbdc998800c3efc8d10d96e800d5422f30486e3
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<#~l9i@Ba%;wwm;&<cN2*iRpyUsfJqq&*thp{C@D|gBIPbsjcgl
zoKd@;)NA}|*@xuU-&bk)=N~J4VDzf!pV%woopSxZcS`g}2R|+>&#^go??{#3MOou-
zhx|1sXaC~mI`!za=-1{8R}x=;DelzGyz{clWaaiP{cM*R-rTEM%`Dfy=UCDE`mQ4j
zg#Yd^l@ECs{KY=~vb4zg)R*B!zvF$^o|~RlA)2GATDxj4r$nN~qQYlMpD*s{f1bZs
z`<LM%m0FV%OMB8{zUd!z58yAZ%uTWb=Ix++K`SFBp8Lto6d1q|F3u(%1tx$pV6c+o
zKp=no-U1#5f1tE7t8y%ufJqmnrWL<uWMJU>wcuW}w`nj7Pq@6YJTRmohVq0fC@UcG
zw5_$>;3kIaSnI%fg6lh_O4q$Q%+3=op)8?>FhEjS5{W0RERDpIRhC8KsavbtAq>~B
z)<EKES!=<0j_q$4ets#r#mEwF#9<Tz)8g)tw5)~Wh72Rn1uPOuFc)}julgL%8uJ$@
ztfHV|1yc=^X5s<|SMQ<U3)VceR^UM~GQ>|v{l!EZR*+s!6@HlBjAK)!jq5{SGx3Bg
zDl2NhjN%E`wbpfjDT2|AT;KqQS#T)C?q$6E_2WPz4Ok46V5*)Puzq3Z7w-nyB&s40
z6aKQ{<EH4wi<&^2WRzu)B0|$z6CU(`1;RGn%|G&tnI~LISqdpY<&@=+czV`)@E~L;
zO3g_GrGet2<l<lu15AJz49wwXc3b%7Hq`ECiwFu12@7Qi1B-)+AOtf6!3;$(BM{8w
z+?1RQAlEz1mzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?9
s9-F9`w2Z8RqLPZLx~7(nzJZ~!skxkql&+DBsF=G~B$Jz;qX(l809^&wr2qf`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp.29A1243F9B97E124.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp.29A1243F9B97E124.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4ce0e054473e5fedd49cd2f280dfbf4a2d250a0a
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg`p7@l%7FjUMr8+JeVwuQhq`(ppO
zy&eyCm>JtR%!)A%SGb(r?%u1kHn}(Tr20wDFYEVDl@&JlvM?YoPwmE~x=Zi2+xxOj
zVBBG2?x!ITCzd|@&$q)9Pds@1M&E9~)&3tg&&9hxzI6ET?_%9XDVc~hed^2yU!HXM
z5OMU=t8Af|+BxwxzbA{`|2OaR)#<NU?c;UNA5(eo=W@}}@cHNW>#f@#^XaqqBpddc
zS!b;hPI84SYBvQaz0h5Km(Nx_Y{u>d%#+uxiWSIi-s6&gaYO0i-lSVzmo;h*{+)i`
zr0Q+!e(Q`KcIO_rNjUpk#V+jK;<UPi&uz`cmMvfJ{rqxo(f;n%ZL_!R?438Me6n?X
zXNhODtm1-`nYprSGMDGFShbsFq+4&S`Tp*3wa1ZL-prg=R<!g+y;Dp-`K0-D-?kMB
zZ%U@HYMgOR{+1<jO>Suii}n<^3c;6gH=5TiRB(73`}EOXt39fHf#DwJPG5R^&b7-a
zy+1MAi~qYo(pK}CS(bK+J2&!jynYk=bQZ(hd(25&TvxO#uaRlFf6@8|*ZS5=lL9=W
zADa7m9iJDl)Oo=H=XDn~JT_iCl(Ea=z{#U89>w-n?^Hdnq0%W}Be2`5mGR3lE3ORn
z)vZD8r5qEubE_JUH`IC0+M*ORE8~Y;NALD0%bU6`IbR8KJN}rDxApUsdJEge7YbXm
zv{iTeKl&CEBpn^Kee(w4dus2kZrO^t^lixdBx7shn6rU>mV-;MCHLxG9p5k3lqe;I
zO0SyXx5GK`klC_4PsdxorzW}WZuoDKp*nTK<+aoHa))f$cQpFbmX5gmJuAAGEZX?Q
z=eOMx@2%DSPGTp8R(ix%TsAz_zy*xz^tZ)+=clatdqzcf<@JQzIXaxLZXS+0F=^H0
z6OyqnE6S!co?g($8mTySiskhRr>Wo4*KpeLHeMEXvrRJ3sWGyWy6Ca=&)lVZ{!ToY
z`)yO@we!;+sCwJJw<$L*t39!mH!y(V=CVAGX5W@*1&_oJ%XRl15)0mORp_U1L%49B
z`kpx-qEmM~S-9YmQemCV&i~#%vpQ<#_WM^x%H{N?#HRPWnp~-_{iC+T#wsa0H*#UN
z=H{DAH3}Wyiqz%&ixX9S_12>=#;x|;eV2N*(=SVA^u64&*>vu=<I66;$vyL2<bh_g
z{+~HUk(csA0+iy{XGpYe6SH1F|8%aavj6oZo_m+pn#^cR`2WmW_CvuF(`oGg!&TV!
zzBvB&=K{X@i~d{AKDePm&wKx*+yAqRzMk8{ye2_f^JVp0p936^IeSgSZ|fhgb*P^m
zbD>{Q?d54lyHu0so9Y=)2?U-guFOrc1LkhMo;ANk^UEf3GXcXPT%1ij8cYCXz~I(J
z3x^5PhhFnA_yeVtS(W3!1WdXpHLdtPBLf51uLbv-y-kBzc*5nC<@I2O@`NiWD<JW-
zt+n0YCWh-+>%e(19ZXzc>w6FVUa;n&wE|DL5r<JMOp9at8-|}>N^UW-pa^?yulgL%
z8uOQtCtO8A#R{f6#7{{5#Y7uckT9nTKTJ5|*i>ob`q0-*JmHGUiW)GBdBSzAbsb=e
zniHhlRvQ|Z0Tsz9%c&z2=~?U9!xS9~v3nUWfBiVn0s|HUC73XbX5<0~70fG74OqXh
z^NV+boFl3t4^#E#fYwZowXsq_VF?xqC74ZLHhkO^-FQ(G=l~gI8KfZ8wAO?N+FyaN
zO?UH;JmUcBRhCeL>E#KRRF*{INhwPq@uZcdk$AGovPe92Yjr!A!+63qtTm8$TGm={
z9z#)TP9i7`6c;5I2ZI=30>ofo2!l`{G6=y8fis!IGu!9vQaltN%N7w791<4F5DJ%L
zh(Iutb5n9MfL!l1UuG6oHg*n9ZXR9%em+k@S7$XrE>9;RabXb&NefGB8(TYj2R%`3
wD`hhcd2FI$(lW9Nib^V~>Y7?Q`UZx^rsi@cQo2SiqGIk|kxXuWjvkCY04pBjfB*mh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp.020B0D68AE892602.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp.020B0D68AE892602.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8747bfe8e9f72319c5f428fb4e72c738bf1199ea
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg`p7@l%7FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wo;0JcDQ)oHuq@e$WY0}zf$g<6snV-!{>J+C_3zg
z*QuN$-D?($%ed^#>eqJ(eJxXA*IR7&Jb#h><kuf}%<ucSW2>oL`SD|q-{hWoF7m)I
zS^v)*qn*p{FIeERFD}rY%U5?>oc;9MlicR5^O#)a{X4^e_u#v;+pQjaJdv&Q{oi^I
z9;=J}w?8||+PVHOHrt!{F=kqw>h}NFpS+rr$rf=iEa=krTQeI3+mxj;^~~b?emDFx
z+qFnu+2_*q=9*I(=TiSO{qqRCU0j))WCzUM79oxqQHPgk@i6!YFoY|!D#w8dpbQv1
zt=aCle|oGnH&Y-`TAWQh8ce{Xi&E2y-!n2WaQ#|vui4u)n1v@?URhobri&+BL0JKb
zr){n62DdO=$65!@gNZV7fvxX7^n1aYht>*??Qa-<ekr-d$P#YEVH68f<+Z))b3ALz
zUq+sA6$KS5m~e=nkot>>Hmo3FP8EKbaK^Ez(#G|nubFtl6_ph=5RTKe)^&g>YEF=H
zTWx4u22>=cET@i8q-U*X4^y<gXUDk`j;ZND3nW-1lwgVuh1k7}m%n}-C~UxDpafI>
z)PVI1JHL20$T^}a@^E2jkTP+BgA(Rd9uyyb+3;~wbmK)$pmSuDWsrhU(^?ZAXnzI5
zHr>rX@{9xM9AybLm~(i-C6y(Scv8wzNIYp}X(XPkvMdr$-CEra=1iV&4QmY~o|d&1
zoX1d<nv)1h1I0zj#lav3m;f;t7{VYFhzvq7LlDeRIFmX2Lx`+Se6U(6TSQQBNLVOC
z1YC|GIX5LI1IYDG^JQjXWn<^y<mTZO;OFxcbahq}<nnY95*HSckhHM0wz0LdchD2n
xwo*3JkjEw}CM_eYps1vxs;;S}qi<knY-%oNBBg8OA}Z$Y70Kl0=jg%c0|0x}$x8qL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp.A426959DB1641563.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp.A426959DB1641563.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d219e8aff481d8bffd2b76a9748c34c4579a92ce
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg`p7@l%7FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wo;0JcDOihUH>+=O{<LDzMkAOS@n;dh3uRSOSi4L
zFm=k!CsAt(-QJ4S<@}2iRekl=gD=Ld_S}7!dbR713+-he7iMSMyzA>de)IN>^O^^e
zkH!74OfB-=@95ZBvv*OQ$gHTGy>)uqReNn>C#rp&_U&SVRMXwr`K1RQPrM%S?r;1=
ziIPR~+n+aE)pY%Po>O)B(N3MeUU~o5pS)^$nP)@O+9gZgZ#8CA=5aT<6qB=0<~!rx
z+!9xNw;4<H*?&&CIOp^~);F&LuN7D3CfNaVcgpqYuhq62%;I724`2vaW>t;@6F?a-
zsJwZO*;2auH8)crP+FW#JQ_^Eq>ED1ir+IbFmU}^aIe|hG?;}aTwYmT52lMJTtQg@
ziKlI??FP3nT*q1m&U0*k!|?M<$t^~fa3c<*SePQO?Ny)SS!4b(@`S4>s93>-L;Qr)
zUre-N1qpMi@WX^*G$R+-550$eFIe-?S|Q`uRB7Y-(AP{n;fl(N8VJYfTI)K%^uR=!
zxWFES>E%IjPIH2k+iFAOGN4Iv%5v&3Jv`xh)_V3ZlXlBK7P&Rit`q1u2^I+@n4&`=
zb}!@QuO9~r8?YEC!BjsrVEw|*FWwFEil~Y_O!&)&kDH<!FKPm9l2MjH3PMe5O?aUF
z6$smOH~+{p4xmlS5^6A;c)}%>C6Rbi%2G%?X=Q07o~*Jg5>MS)-45mgo^TCo4J4kH
zwHBPmP?VaJ2ucIRMajj%AO@HKF&G%aAQXrULNG(%Oy=-2PFGv~&P-@ziwFu12@7Qi
zh08HSAehOyDLENHu6LR*GYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dt
zo~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X*>CO1Dv4@MsVH<8|O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp.1B1383BC1860443B.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp.1B1383BC1860443B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2948f37983b57768f927cb52015b0be32fbc5d6a
GIT binary patch
literal 1792
zcmWIYbaVT`#=zj5;#rZKT9U}Zz`!5~#Kk2=ndg`p7@l%7FjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wo;0JcDOk1*5b8|YgY9H)*8J};oQ&f$F82UQdhe~
zHPvjVSoWMi>sI;M`@>y3tE?w6PY;~`+5F=ClV2<D_#dyhb4#j!ck<)HZ)VSmxjxL;
zC|>6`Gv;!BK!8&G`W5o6%fzDB&p(~(sq7!V#B=Y`T9X-V3E!VZ%YLYMVmht*f4B<U
z+!x2+{#?K}f6;%-+4nY7=y~s-bo+mHQRz7o<~0f0nlG#0`W)amEZA!zK3o5ItwX(Z
z?8SaTwU?(I?NUviZ>nchGz@%NT$!6>2h81(%*v|ELrOHcnF0eC!o}Idqrn7F1`PHd
zv`f^szdDD9!5=8C%&Hs*CScM<scFUU85tP3el57y>}?v%!V@m9EUyPMlqXz4SpkWs
zZLRGFH!)ntS_jT^Y=6V>^GnGsMwW0R4x?C@B6p9ZWi1>xWEgqEC0HbsVCH&lulgL%
z8uJ$@tfHV|1ydd3C#3#jq75rZm{WxxCJfWX#0B<R@1frd);zRU;6bq|<JeSb<NDCo
zOg!O=%8D8=Jv`yM*18TblbREx+*TVJmjM;YDa)xN6zN&(*~1jUXhtq@P{FJ_6k_)>
zUjF)VpyLc!43uE1o*J-zVdod`2Dw>OMII*nWy8l!(Tx{1fi}q~%OC}zrnM$K(EbX9
zZMvI(<QWGrbd)935YCZQmPF!7DN7;oq?M(Sc(Tf}NIZ3Gbvu{~c)~TTHIR5()>?2L
zLs4o@A}9?M7bO=5gBTGY0s_E%2r~>O#=sDSV1^)=p>QU1_}(s=g2LuE-`FC8f<wYW
z8Ip5Tax#GO-f6zfEUawo9Gu)dyaN1uo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<
uW*YL?M8%|KWEB*ZR8-Y9wRH3i42@0A<xHe>ja)><+`S^1-25Cp7<~Ydg||=u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp.C0ABEB6CD7C15E09.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp.C0ABEB6CD7C15E09.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7e2b545b60b6ad23faf1a8653c82d40fc641c0d0
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JeVwuQhq`(ppO
zy&eyCm>JtR%!)A%SGb(r?%u1kHn}(Tr20wDFYEVDl@&JlvM?YoPwmE~x=Zi2+xxOj
zVBBG2?x!ITCzd|@&$q)9Pds@1M&E9~)&3tg&&9hxzI6ET?_%9XDVc~hed^2yU!HXM
z5OMU=t8Af|+BxwxzbA{`|2OaR)#<NU?c;UNA5(eo=W@}}@cHNW>#f@#^XaqqBpddc
zS!b;hPI84SYBvQaz0h5Km(Nx_Y{u>d%#+uxiWSIi-s6&gaYO0i-lSVzmo;h*{+)i`
zr0Q+!e(Q`KcIO_rNjUpk#V+jK;<UPi&uz`cmMvfJ{rqxo(f;n%ZL_!R?438Me6n?X
zXNhODtm1-`nYprSGMDGFShbsFq+4&S`Tp*3wa1ZL-prg=R<!g+y;Dp-`K0-D-?kMB
zZ%U@HYMgOR{+1<jO>Suii}n<^3c;6gH=5TiRB(73`}EOXt39fHf#DwJPG5R^&b7-a
zy+1MAi~qYo(pK}CS(bK+J2&!jynYk=bQZ(hd(25&TvxO#uaRlFf6@8|*ZS5=lL9=W
zADa7m9iJDl)Oo=H=XDn~JT_iCl(Ea=z{#U89>w-n?^Hdnq0%W}Be2`5mGR3lE3ORn
z)vZD8r5qEubE_JUH`IC0+M*ORE8~Y;NALD0%bU6`IbR8KJN}rDxApUsdJEge7YbXm
zv{iTeKl&CEBpn^Kee(w4dus2kZrO^t^lixdBx7shn6rU>mV-;MCHLxG9p5k3lqe;I
zO0SyXx5GK`klC_4PsdxorzW}WZuoDKp*nTK<+aoHa))f$cQpFbmX5gmJuAAGEZX?Q
z=eOMx@2%DSPGTp8R(ix%TsAz_zy*xz^tZ)+=clatdqzcf<@JQzIXaxLZXS+0F=^H0
z6OyqnE6S!co?g($8mTySiskhRr>Wo4*KpeLHeMEXvrRJ3sWGyWy6Ca=&)lVZ{!ToY
z`)yO@we!;+sCwJJw<$L*t39!mH!y(V=CVAGX5W@*1&_oJ%XRl15)0mORp_U1L%49B
z`kpx-qEmM~S-9YmQemCV&i~#%vphC^)YL!SymI22*FAxiM(<Nt_xJa?yXUOc)h^LZ
zHQOzss<&eHq5ie^S1+>JD|vk4=@qAcMz1)ZVmX(u*l=#yvSpY5-6)y0xAW_qMb@cr
zTIUC;dH?F=I`!zaC~xzHYl+Xl6nC!8xbt$B$;$0p`q?fsyt!YqoLR1K&#_Mr>bs6C
z5dXWwR6gWk@E7}ZV+oP-sV~Eee#iT+y*ItALNrHNwRY89PKiXDg@w<OeqY?t|2%&(
z_b<akDzzpjmiDB@eA8EZ?*FK`GB?Q%n77@2#V)?T+0B!O!9Rc@T$xol7EAzTz+g6$
z-hS&OCj)LKU?7Eyvx!H637B+IYFhDoMg|71UkmOvdz%Kc@Px}N%j>~(@q{ZVD<JW-
zt+n0Y7KZCs>%e(}>pP`N*S$K-#1k&3ET@JrK+jsw4yMJi{SCv<FD17aS;CDtjACHI
zFq)AI?DXD4zZa}|XszJ2z3Ov3Ys_Cpo^TZf6)Tvk5I-UH7ZYt*LBgCW{4n8+V^gJ#
z>qB1y9iXVJsDW^RuC=ZMOc6|!i3=PGFq?Q#oO39|?q$6E_2WPb3|I`5V0xb#uzq3Z
z7w-nyE2<(76Ml0*YbM9qSSg^e1dD_c%!V%;K5mL`yr>Cuj*PMlQV?odYr+HVuRz$Q
zyZJ|+aRBuyOCSZBq_QLuPfA$|i6^Zrjl`2xmPO*JTdTtZn<reuS_6rvWvvD0F%+fd
zB!bdFaZz${Fo*#rKnw<kFbD-CnZvhCE!g#L-S5|I5kbKrVWA8`a5;t$1Tz%Dj6g7x
zb5n9MfL!l1UuG6oHg*n9ZXR9%em+k@S7$XrE>9;RabXb&NefGB8(TYj2R%`3D`hhc
sd2FI$(lW9Nib^V~>Y7?Q`UZx^rsi@cQo2SiqGIk|kxXuWjvkCY03>DB=Kufz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp.76E857B0424F98D1.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp.76E857B0424F98D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b1f67dcb1a7ac6f5623e339ceaff876bf86dca8a
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wwjbvaqjxJj4XxKTFD{v&R9ntdf$BcAZK*eX|7n$
znZE0eN~K@%eR%xZ`_RC7_j(>5NWJp(kM5Q9B3rrd#fEa*mmPcj?ncF&y`5j@EV52}
z(>gy$?fMsKkts*7YrkS&xbpC|m(PWwFYUN2ni-UzDbKr%@y6btVeB??mAy~z{TJzU
zRR2|&ZMWjk@)z~zmYHkJJAG;WlW+THMa|Ve`$6l5m)EZlYhi;!1ul=z9DTO9Q2zXV
zx9~5?EuOzJCU{Gp-SH+qn$iDbab<3j9WZZ;=UOdOnd8;V%@i2G5H8Lp9t9?VGGNg8
z<>8~{GQA-@4E{iAWme@_FaeV;N=++%&&a^Q^=rYsW^dDA7M^f<WqCcAp*-OV$_hw4
zZEI~exQXF9);e&W;QCIf(si#6Gx3DWDa)xL4A8UIvx8}Yi8674-Pn8R_kuMKtrd8}
zjW~>AV2T{u-!T0AQgVxt1x46vd)4Q7)|kJHJmD$|DpoMnA$~&YFDBZsf`mC$_+i2s
z$EHdf*N46aI!94iQ3GZ%Pq?nNt^-UFjArBlhY8F%+k1ALE8&=$4m47NMM4Rt_fUx4
z%Xs<g$AQ8IECxz2%})(jzp(R*cY_=tsv-{){<7iYrs&3tnm`B0D9a!Pp{BJaJkb6M
zgl)Q;f8-el&?aRGq(GBYmPF!7DN7;oq?M(Sc(Tf}NIZ3Gb$DR&glkx9An~-UwctF4
zqSTy3P#P#MN-hotF~9_f!N3p(p+IC1f*As5GKc?4bN}0^&)~`y5fmH}7RnF`mt%-P
zFq3mrax#Ei?=)X#7FITP4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_
uGYxrcqGHlAvI>eyDyr(5S~~g$hQ_Amawbx`MlPaa?p~2hZhnp)j6MKZ)Z6j^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp.2C4EF55D06D154CA.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp.2C4EF55D06D154CA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a9650f2cbe6916fe7c4a2fe8182487f21011394
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wwjbvaqjfji&q@NLXWMmnOVN6<v#oJX5npDrit#F
zWHdXb+w|Ou*@ybq-e0}QX0PP&#?vcK|BPO7{z;Y1yXQ$ZdA_~J?{55PQ7!!1(zX1|
z4Q~6TKI>kZYiM+@55L0i6x_b{@_E&5OA423Us`hSlAV+f^M>lrYxyc{K1rS4`(IP2
z+3(BaYc&C_{ulqx@y!pgnZ7js<lB1Ftyb}~9)@n1I_1mCGPQ)3haJaecAs^BY&UOz
z@46SqIVOL(s4(4VR^iS4o@@Qz6j$aZ*#Yx5-@W_iZJ)Y$@-X-ZFoY|!D#wBepbQx7
zTGg4L+wJg|n<)?|EzTw$1twt9MX71U?->~wxPC3T*X(T?%)%2cuPm<z)5R06psawz
z)3(-jgIgG`W32<{!9*Fkz}EL3`n_PyLu&=W^_^0s>s}pZ;t7{imQzERt7ole2h-!&
z{)XY_my%nIEa65RMlmp9ukBTz<5^?=GV+A0D5zM$ghTv<)L%@rVFd|us_?^vGmcG_
zHm(nS4YWy7Sy2OFldiR{156Q2l!*%*3NTmlpxC=x_OZyViFTbp3nW-1lwf)eh1k7}
zm%n}-C~UxDpae7WsR8R3c7E}0kaI*;<YB^JHhkO^-FQ(G=o}ek8KfZ8wAO?N+FyaN
zO?UH;JmUb`q%46HI+DtgNIWTJDI}h>vNRG;R#_H_r*5qd4^^IU4QmY~o|d&1oX1d<
znv)1h1I0zj#lav3m;f;t7{VYFhzvq7L*Pv2@L2P}Z?);Oy4WIuf<wYW8A9Q53=s%s
za&Aga29WEW=F7~&%Er#Y$<4znz|ZF?=<2K{$mQuIBrYr>A!%W0ZDVU^@1Q5DZKZ6c
tA&*T|Oj<@(K~YIXRb5j{N8iBE*wkFkL`v7lMO4h)E0W30&(VX?2LRO8<3<1g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp.D9C825C0FBC7162C.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp.D9C825C0FBC7162C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7082d48740f58d513f0b555108cbab1ab6e2084e
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JSIwuQje^KXI;
z)83?PUfXD+vTm>GHqV~CZW|#J%S{}?OQtQUd3pVPz@|qop7OIyx>;p9=Qy96vr|jb
zf$hPkI~EHaDk5Xne7f5%Jn`W1lkqiG#g$(xPpwZo{CVDkzl;CoT(mZbp3_&=^m+5V
z2OGP7p1rQJ<EO>mU*FVq_x-hf+O5CcxNhf+cplFKKbQY((zlDRn|rrz=d<b8RVw*@
zn9SUAV2<cImoS!vE*GT3cIsB?tugq$fO+z|H8BF-%{9IE7A2mJl0Ld+>awsOO~3W`
zXMVZFz281-hwa%1eiF|9R<VnEw>YgX;d9fu*s|s8y`Nw1Ex3Q&>2~O?H;a}(x%>3o
zbfwr!VeT9mMN!f0-nTNP8_zA|y>&B8=l<T`ALN8KJ`3_^64gpvJdOQCq|wg@n-=S8
zaYW0eGIA9M&Tf`!o$9WsbdYPQV4TY?alQqsH8}z*!*^Az=PN&A_)6%dTgSd>7k^%G
zd?H&J{OX^alMcUHu<mhnmGEtBiF5BxU&nXhm93InbW2dhspEk^WcN1reu&c0<(#@j
zv@Udci44b7D;6%BCsMslBH6k<%ermn#PsOB)3!OGTj{3g+47+AMyMvwg79mB4$+aS
zQ>@lGH^gLI=Con^B7JFz=)_Adhszz->{YcFoiask(KI6~nQa*wbC&-({8Q3}ePyV$
z=FM86?awDg-QHRn%^m$P_tUzicM~0-aGbU}f6-a?2$%H&J{Oazn>VcLtq?!!uC6p~
z>cXzf)@YIJ8%M6_YI?Bm4?2?}{$ak<+9g6oyDk<LG<IDxuW+6xcJcL8@z9AOSEkL$
zym!22W!n3|6H1XvqAR)IN4ZG`Jbd2$b4jJ&%b7u?GiEMP{VJu?J2|__>e~j9BHyc?
zPkQu9_gI`fkQ#8bY0U($Q*&0^F7*1%9mYPt?Z8jT#c~_9&1|IS@MJ6r`llCk@85}p
z>3ehLm2Ot8I2m-_`h09_u6$9hvxtbpwpCe|5`-VJaRf>H$?RUg;iyR77MFi451w}1
z7L4!z;U}qAsj*^<$DH|n@BRx-wwjbvan8Hi#B@UFR6{NQXLEHPen0r~L5uFz)Yf%N
z&Zu2a>NS3~>_hVF@2fQY^N$riFnU$=PwbWPPPzWyJ0<#~gC7@`=h&RPccjYiqO9?^
zL;jkRvw!h&oqF_I^lS5lD~Yea6nE-o-g((&vU2;DezwaDZ|>EsW|r&UbFAoneb<o%
z!hd&|%7;7*{$ihgSz6?L>dWw=-|@a{&rMIO5Y16ltz9*jQzFq~QQ@<s&lh*}KhIyR
z{mbx>O0CI>r9Eje-}Gyb`~NPk%uTWb=4~dUZySD}+a}7*6d1q|F3u(%1tx$pV35df
zp!#joZ+{*Jf1tE7t8y%ufJqmnrWL<uM3}<F1=ihr==Xv(53Ln=!i_kLVqj)){aSFZ
z+1oUjg(qBISzZrjGf%jJvH}uM+gjTVriCY5$65!@6I|aZRl4rgVJ4n%Ib}IDgaLZi
zdUh}^j_q$4ets#r#mItUxw}WwvKEdTGK@Uo5-bu*Fuh*et3JoG#{2~et0<^g!3+uU
z6H<RM(S{Wy%&Ec;6Nb@@T;NcEc{AhKRB7Y-(APkBDJm;!z|7_e*R|Glfay6DV)rs$
z{`zsCkp?UVN-*K42CQG$`Ng|IHi@dp!-T(V__!&$@uDWsNEu}rq|ni{)`W-0UxBbq
zck_=t;{f_eSpq3MB$XwRcv8wzNIYp}X(XPkvMdr$-C7+Ujy&NS)*47WEo&_}kD(|v
zClQneii?tqgFy^10b(#Pgh40}8H8YlAef<WCUdw<>rO$homC8M5kbKrVWA8Wa5;wL
z+?1RQAlEz1mzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?9
s9-F9`w2Z8RqLPZLx~7(nzJZ~!skxkql&+DBsF=G~B$Jz;qX(l80E~m!6aWAK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp.9ED7E55B00DD6F70.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp.9ED7E55B00DD6F70.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d5a95265df58c67404669153020cc21fd0a10c1
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g@VUwuQhq`(ppO
zy&eyCm>JtR%!)A%SGb(r?%u1kHn}(Tr20wDFYEVDl@&Jlve4muno#sD|66~*rB7l^
zVbD9@E6mk0{pd}#{eKHojDA@B?moZ%oW9-ppGT8wtOWMdZ;4;S#xAz{@kxh*DkA}#
zu#J0u&0-0ce=dIi-;$%-_dl=s`n2lc`S8=54LL3TZ;APM_4()Z-LLhh|M@dX<vip3
zw9R}Gn+}DZ=u()%QFS_NyYu;0(Pw%u4lG%nC4S_Vg8s#5lbpL5i)U=RWWp6+@c*fK
zZ(Q}m_k2?8`i*x89=kBvFFZqe?FFs5&4Sq`in)8s?e=be@jhwcTeaN0#mk@EeR3{b
zDfUvBJ4Z%QR5W|`txW00a|?NI-SpGBzxVeCIiZc;E;%rnY9%h7R(&GU=;wn?i*>a)
ztYuRfxr#5&*elh#)m>BRAlFjCIG0`G-xg$Ra0FC`@2ptQm*1IuMf6gyK&|rP&kNg6
zSbth}>9B1}#9QBGQGNa%c{$97tbhvD9F~7?kjNF(dZ;PS{;>QD-?l^93%96*TnhW)
zJNeQ^^^lhW8XqQPn{WlK*=jILkFTJ}a92h6<G8$&#$29CEc{2-`6V*!-Pp%;>txo#
zD~Ym9Pn>4QJuqBwf0EkTjwxzV``D8n>;7E1V9^(Wty3<3+~Mr}(B`SUpTB~Y$J$$7
zC*MupVLScGv1!-d>bz;$esXtTSwHJU<u$Wy-1&V4q~A2E30w?0cW9N~qgoUBxg8N#
z+q9MmzYCbMp?76;(1dSwOJ`ivdtlES<*E9lWOeF$CegL|AE(*qF1jwQ|4M0PhK}Xs
zef&Rz-rie0!F7|1R*?9;k{+W4hk#KX`FqY~<E2^l#-84<Oe3V#y_>4OZ4leE<dsU1
zTe#Ic+0+L{E0P_<JXBN9na&GX`d2c$$=~_GmzKc(8@$r<d-~i=f;8*XHS_CJ3TE%k
znfG<GYQ@Q*^Va8Mdt>E`a-Bs)9JZ~LzLX&RP>dr;;!kGx`VB`#^0vDCV|j43<u<3j
z`kydKy-JN0TRi5-`@h?tS{YzJsdTQPmglp%YaV>R`0+)@+O4Ur>z1BLo4)7>YiYob
zjYq3Pr_45gYFZ>&YFw{9b*9ep%ZYQAU-tW3a!Y=;-)8g8GlHKi>)HEZ`NWfx>=ri}
zeW`MFK7M=GE16ooOwnHb*vV>Nwcjp2Xma6~`Tds-?aJ#n+WhFBAn|39{PtVTS=C+t
zp69Uj{*Re<Z_@4mcCA(a-Wq7#aP#`Lvs5ji<zvgS$=z=s?C*b=KbiZN!66mix+3Mp
zGpGLLZu)3xRa}{yWCzUCYD}4KB8K}Ob29}7FocV<iTi;GpbQvz*>E;hpV|A3hru5x
zt<0(%2qs|CMX71U&lnjPxPC3T*X(T?%)%2cuPm<zGn6M>L0JKbr){n61~)NW$65!@
zgXv)60$bmE==Xv(53Ln=!VQHDqhVSc+utz!{8Dm@kp)E<Ce6qNb{kBu*Y>K<@vJd_
z8F|806jZEWszUsP)L%@rVFd|us_?^vGmcG_Hm(nS&BPO~sH~^~a}Q6ruC=ZMOwpkb
zyO;6u*N+4B8n75B!Gxb0uzq3Z7w-lcDXJn56Ml0*YbM9qSSg^e1dD_c%*ZbrK5mL`
zyr>DZNk&-)DReZgHR0j$S0HTD-TWiZxPW?<CDdSgdBP=?C6Rbi%2G%?X=Q07o~*Jg
z5>HN94vD93t!@W%CQrDAwFVMT%UTPGr)RAP=M|;qB!XhTxG1?e7)UdRbLXC}6bs4S
z%@z?991<4F01}0QFc^b@Aqc??K`=uR%m@TCIX5LI1IYDG^JQjXWn<^y<mTZO;OFxc
zbahq}<nnY95*HSckhHM0wz0LdchD2nwo*3JkjEw}CM_eYps1vxs;;S}qi<knY-%oN
VBBg8OA}Z$Y70Kl0=jg%c0{~S|zwQ74

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp.87288671AE2510FD.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp.87288671AE2510FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f001fd5bcbedd211d2343a5fa0b27bf82881fb77
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g{?<mW9AK`(ppO
zu8mKMbd7CPMAn@#<-DA&-Y=!Nb|YKlCr_gZ|GsYDROm71%R<j8bLMSZ=5P7^*8c>9
z5ySN7$2eIJole?x?z_C73!D7%`j0oi9)0>;a{nCu^VLn?`@eloeSe^)(&`oGx${*`
z$F;<+$4`Fn>+zdUx94y9@$Gw${{A~>ejY6{l$_M?*}q#gzHZ+StNlNVj_R-T_$2co
z(<---RXe704Tn?bg6(UHtiQxWB))6n^Vzkj$Z-wNC*G>=wzXNgBA3HvMLpzxxBt(w
zi`x2s?gkz{o5{BQkVMs`mqNKqM5C4#8zi-=Em`~h-16^Ye`9sF#4bN;SLJmt_g#(V
zCDq*0mIk%uTT3@An!8Q#h0Lu5b90Y=y1SeIy`o^Y-dB}?g+gXol?`)LC-duB-&)kP
z;c}qZ0t4}_mMa1yDyJw3tqG8N=(?o(A;Yd&O$<}Zmv(=DRp2kPSZ`up%ZqHux%{#&
z_a>SzefM7R$fpFOt3@?VMY-7m*KY2<bd@n`ultdV?f@?TpB7yE7ng4kjpJUbx^U9A
zL+f8p>0@3P)b7yO9<w-bLekpSizOS|CwE^wx|@r?$a`L(hmc~0LRlym^NX#eA{WBK
zIG6Ul7ElnsRl?NI_;;E~rt1=uiyveKvhz>+v#Ks_U$L}@uU%J)>$&Ftf;zT~j#*dT
zyvpVudAoDTwX)>9=MMQ*%-)loSso!^d1%$S;x7U08;|j=U>5X!s<S3b`JU_Ni=7>*
zsVkDA9`YS!+h$nm&2{PEdJ*GWtbdlfX|HUVlea}CX2GJby!R$7cl}cJ)-_Z`bgOFm
ztnY^Lt8P|{Z1R}q;ri;(?P-B*AwPZ=8RpJ2u38o%o2HiOxtC4*@sqX%eYR^_?gW$y
zSuXaLRackX!Lur2f@qNE%;K%;A({4xr5`dm=FB&ICAd57@aBUj8{D*>=1=<c&7JM{
zo7B&1X6qjGo>q0Q@|{-B?+Mw`ixx0!Dh-{?Hk-rTVZxyY{?Rq9y2}bytA0{v(09Mn
z_ap7d_frKYogJ6DKKfHq^8fCQ#lITYtm+B$G<u)Hxu4&UT|H;1u6BuPs@Yy4={dpd
zt@6=(!d-V(2FEc^4>bS0J!t01vlVyxKUdsYCEve0`ElX5G_m5YJ2N(l|1&Y#x$yjg
z1w4D>0`GrZwl2zjzIT-8WAE@Ko`1c5XBhBqd3Uzl>cPho**f3<DXXx}eQ|v4w*|N6
zE&6XcJMOXl^rhRMeB1B8u&>^@hilukNp*R1IVBP;7JYn@Wc=fO@{ijuCfW5Ecs9?!
zb0Q-x^*__r#g?AMmAOfFz&stmds??xJ=KDT!9Rc@T$xol5KI7Nz@Y3T!vT@Lf-Y{R
zK%lfZo46mCfJqmnrWHS9WMJU>wcuW}w`nj7Pq@6YydF#!Pq>1z0uoQ#TH6h7VYrU9
z4xH!M{)XY_my%nIEa8U2hS4xZUfZia$Fs)#W#kE0QBbjh3BzbcF0f;J5B*-S=ApGh
zh@X)9i-|U@Ak~~I{4iA+$EHdf*N47l;t5w&R@6W^PS;x30j3Bh%ESfsAk0V}6z6R3
z*>SFfV`@6k0tprgC77N=A$Bk0<*y$H3LCH(D8Y<;YQXx1onO2g<Q!2Id6@8*4Iei}
zH(t~PI!8uX1}Stjtu^7{@mC;h)7|_d&$xg#DNCrqY~l%*RF*{INhwPq@uZcdk$AGo
zvPe8RWjQ3Cy0y9;%so8e8rB*}JS}T2B%YqN9-LQ{nv)2M`QoDF;$RR1On?{+%;Dnm
zScMNfOxnU05fmH}7RnF?76%hS2xbU^8H!*=AehOyDLENHu6LR*GYcylI|nB>53c|}
zpQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|
Xb2$?!T_YD!F?X*>CO1Dv4@MsVyF0h~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp.AA8697A6E53E704C.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp.AA8697A6E53E704C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b193fa0d52087876331fb2717f71523337ebef89
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g{?<mW9AK`(ppO
zu8mKMbd7CPMAn@#<-DA&-Y=!Nb|YKlCr_gZ|GsYDROm71%R<j8bLMSZ=5P7^*8c>9
z5ySN7$2eIJole?x?z_C73!D7%`j0oi9)0>;a{nCu^VLn?`@eloeSe^)(&`oGx${*`
z$F;<+$4`Fn>+zdUx94y9@$Gw${{A~>ejY6{l$_M?*}q#gzHZ+StNlNVj_R-T_$2co
z(<---RXe704Tn?bg6(UHtiQxWB))6n^Vzkj$Z-wNC*G>=wzXNgBA3HvMLpzxxBt(w
zi`x2s?gkz{o5{BQkVMs`mqNKqM5C4#8zi-=Em`~h-16^Ye`9sF#4bN;SLJmt_g#(V
zCDq*0mIk%uTT3@An!8Q#h0Lu5b90Y=y1SeIy`o^Y-dB}?g+gXol?`)LC-duB-&)kP
z;c}qZ0t4}_mMa1yDyJw3tqG8N=(?o(A;Yd&O$<}Zmv(=DRp2kPSZ`up%ZqHux%{#&
z_a>SzefM7R$fpFOt3@?VMY-7m*KY2<bd@n`ultdV?f@?TpB7yE7ng4kjpJUbx^U9A
zL+f8p>0@3P)b7yO9<w-bLekpSizOS|CwE^wx|@r?$a`L(hmc~0LRlym^NX#eA{WBK
zIG6Ul7ElnsRl?NI_;;E~rt1=uiyveKvhz>+v#Ks_U$L}@uU%J)>$&Ftf;zT~j#*dT
zyvpVudAoDTwX)>9=MMQ*%-)loSso!^d1%$S;x7U08;|j=U>5X!s<S3b`JU_Ni=7>*
zsVkDA9`YS!+h$nm&2{PEdJ*GWtbdlfX|HUVlea}CX2GJby!R$7cl}cJ)-_Z`bgOFm
ztnY^Lt8P|{Z1R}q;ri;(?P-B*AwPZ=8RpJ2u38o%o2HiOxtC4*@sqX%eYR^_?gW$y
zSuXaLRackX!Lur2f@qNE%;K%;A({4xr5`dm=FB&ICAd57@aBUj8{D*>=1=<c&7JM{
zo7B&1X6qjGo>q0Q@|{-B?+Mw`ixx0!Dh-{?Hk-rTVZxyY{?Rq9y2}bytA0{v(09Mn
z_ap7d_frKYogJ6DKKfHq@?Y27{VQ{1sN~{FDfdna)ydD{^Sj{{9rnWOR8EzunZ<HG
zF8gg2>${3RFWbYWw>a-v{!*jKvp?>TfBtbtsJ&eI@netQoYi?IvePhG|4(L0k@I{<
zN6D(ai}pSCjosR3H+{>bW7F4pO#U_X+r<Q_jJvbnmmGLJ@p{C&zwRD9Ru}uD-#Tu!
zb^Tv#w)a@Qo_GGqxApU#WdEm0i00_3{=H)*Y;dT+<?+d*X&>$%|FHd{YOO@#BsTk<
z6EB`Q^^bMQaZ8ut%G@M7V4j|o^1FBsSB5w@Q(yo?xHy}*AD95jfWhTB{fMr^$u&F-
z{y=GER^>o20h2CDO)GxJ$iTq$Yr(x{Z_{8Fo^W|(c|Dk+JmCt;3P?O{Yi&2UiQzid
zI&hw2`x}OzUrKH<vV<E78%Dzvd2O%y9M2l_mysu2MM1?1CLH1?r2b-}4J$~PQ-vQU
zoN;Wbv~hjtYbKs>MP)?|ge!Edbsb=eV4_T1V2}15`n_PyLu&;d6nl5eJ{Gw((XJC{
zfdq?$5=_sb5WAQ0^4E_8g$-B?lwd}}Xhtq@;J}Q0YQXx1onO2g<VsN$d6=p%8$ND|
zZoH@obft{43{vQ5T5H0?<F7#2rn~t^o^b&kpe&&Va{y1cq_QLuPfA$|i6^Zrjl`2x
zmPO*pDa#@8)UDO+VD8}w*Ra+=;%QlHA@TIA_29gs)SN_6%oi6W7YBnFU;@NoU=EL;
zZE^mw&q8swh@jw*uuz6DusE0qLNG%R%uoa~0>MnqP07gsa=p`hnORub*f}`4d3Xi*
z`8)+(oz(=nJe`EZg+(MJEiA2VZ0+nF^hC9-l+85cv5AUF%g8DyDygWdYijA}8yFg!
Yn#-9;=^D9+in)75GP(IVdNBF`00i;OSpWb4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp.FC0AF91C98199AEA.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp.FC0AF91C98199AEA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d3e8469792c2a1643fae5eadcadcac4910b2df6f
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g{?<mW9AK`(ppO
zu8mKMbd7CPMAn@#<-DA&-Y=!Nb|YKlCr_gZ|GsYDROm71%R<j8bLMSZ=5P7^*8c>9
z5ySN7$2eIJole?x?z_C73!D7%`j0oi9)0>;a{nCu^VLn?`@eloeSe^)(&`oGx${*`
z$F;<+$4`Fn>+zdUx94y9@$Gw${{A~>ejY6{l$_M?*}q#gzHZ+StNlNVj_R-T_$2co
z(<---RXe704Tn?bg6(UHtiQxWB))6n^Vzkj$Z-wNC*G>=wzXNgBA3HvMLpzxxBt(w
zi`x2s?gkz{o5{BQkVMs`mqNKqM5C4#8zi-=Em`~h-16^Ye`9sF#4bN;SLJmt_g#(V
zCDq*0mIk%uTT3@An!8Q#h0Lu5b90Y=y1SeIy`o^Y-dB}?g+gXol?`)LC-duB-&)kP
z;c}qZ0t4}_mMa1yDyJw3tqG8N=(?o(A;Yd&O$<}Zmv(=DRp2kPSZ`up%ZqHux%{#&
z_a>SzefM7R$fpFOt3@?VMY-7m*KY2<bd@n`ultdV?f@?TpB7yE7ng4kjpJUbx^U9A
zL+f8p>0@3P)b7yO9<w-bLekpSizOS|CwE^wx|@r?$a`L(hmc~0LRlym^NX#eA{WBK
zIG6Ul7ElnsRl?NI_;;E~rt1=uiyveKvhz>+v#Ks_U$L}@uU%J)>$&Ftf;zT~j#*dT
zyvpVudAoDTwX)>9=MMQ*%-)loSso!^d1%$S;x7U08;|j=U>5X!s<S3b`JU_Ni=7>*
zsVkDA9`YS!+h$nm&2{PEdJ*GWtbdlfX|HUVlea}CX2GJby!R$7cl}cJ)-_Z`bgOFm
ztnY^Lt8P|{Z1R}q;ri;(?P-B*AwPZ=8RpJ2u38o%o2HiOxtC4*@sqX%eYR^_?gW$y
zSuXaLRackX!Lur2f@qNE%;K%;A({4xr5`dm=FB&ICAd57@aBUj8{D*>=1=<c&7JM{
zo7B&1X6qjGo>q0Q@|{-B?+Mw`ixx0!Dh-{?Hk-rTVZxyY{?Rq9y2}bytA0{v(09Mn
z_ap7d_frKYogJ6DKKfHq@?ZCD;9rNZ&|@o9W|nVixzE15S$Nx}X`;I(8O@IC<UO~t
z`B1;^z155E?9p7`czVU`pV2GNJV~>8cf8Uj?`r??yBj}RR2z3&y5^m^!Cfyk^Uj6z
zfPl8R^()?21c$F(d|q|!l8;NZFD<ECT5B?cE$aI-YuOJKPfVv({}&W$_WSZUtU4gq
z|Kk5SzUx2EKYb~>==c5Q0mt`Cy0J!kd)jY5$I@dUbFrdmqvW6OhJSKPJmuYHEIH_(
zmtwMcQ$6GLO3UEl%G@M7V4gNx+girjr6bJ4;2*#cuFR?&2qu6sU~pG;(roFK(oWn=
zfk0_-HgP{N0h2CDO)GxJ$iTq$Yr(x{Z_{8Fo^W|(c|Djeo^S<a1tgxfwYD4F!f+jH
z9XJmr%E$$_zW31Y1#2E!D>$~lVfgu_<Q5}KxS_CNG)$GdN7AwujvF$JJmC^75=t=Z
zytY?;j%SVe3lvsSP_crk4)GIGe=*U96(r25!VeS9I5t(<xIXkX6HmCJvZ4me5T0;d
zYh4GJqC+8eFXQE}9|!6+U@=gF3BwdKae;#a=3E{WS3Wgh{ld;K-VL%>R7D=9_sfQl
zo1z;pY69(*QI<gp9ZhRZczFC32-|cw|Hv~gpbwQL)L{1Ugi9(*BJre@rI2{i%F;+Y
zS!G!yo}98A5>MS)-45m+o^TCo4J4kHwH6Xj&sq=8D@x5t1jT%DQF3uGhyf-*3<l=#
zDMjzEg{_>M#1;`091<4F5C#?p6F~@O2!a`kU`8OA$+;;x89=UgnlCd8D;qlpCpQnT
z06(9npsTZ*AeX0;khrjjgrtR~wT-Qvy@Q^pww1D(hCDV=F=-iD1w|zlRdr1*9eo2s
ZV^ecE6DeII7f~^HuSh00KSvKn9{`J+w+H|L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp.265462C1E86DC0F1.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp.265462C1E86DC0F1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c12d142b726f12317e35795695df877ab1f254c4
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JeVwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MU2@UY^>GOLdptZMXMj
zo4~lk#@tUsAa2@brTeuOLXtIe{vXv}cmAAx{D;p+f0QUx{Lk1Q!N;dF?fB%z2VYVY
zD%L#u^ebB^rhbpb|8MHL`~KD*oF1Pi_9v&(QhIX3XaDKm`~K8?D=qx|v}=Cs#3z;y
zF4x>+oW7<+bdwQ>YSX<9w{kBQC${e7wla9R>K5xpqZ8L(=t^zN+7jp!ohj9-zvKU>
z*>3Cid?;t<H9szShb#Go=W^{W6T>p5o;%nfI%~qVd)D*sNq;Hdcx9W%?AR|;_T0Yn
z(`$+Mtydh3zJA%S6I`uxl`mLsakReO`E+-={5@x-YjIK@3*1z5t`zZGdaKFDlx=Zk
zO<cWL$1zb)YIFGF4c|0eRU#Ie9P0M?en_r_mz6Q=y_Wd<s}K7u++r2)a$GbwvXZar
z*rS}UWqZfD<LbSXRZoAkq$ElUT)VmZlqsXtUiKpy-2t5bKP@=+FD~C87{|R-b>XCK
zhs1roj?W8N>b&59^SX-~9viP6%GhOb;N;O4k79ePcdDM(Q0Ww~5!h|j%2<*p%erOq
zRi~AQtyoU9n(aN1?66<aH%d^`*QAF3$g${8!H%vk6|*!KKQ3%<yHt6~e$G6Hk_k~;
zr%t-7URbTSvNtw*`_>Jj_tf57-Le&P>D!R?NygU1QRD`*ufn30b6P`VkNnQCvl866
ziZ|3p{m!C@!)nX&JRNWSo|@#gyWzh{hU(M_m)B0)%N??1-_ht#TRP&z<F2FzU5qGN
zRxAH$>8-u)il<XLLzd|6c`K3X&^ljV|M}a`P4hHEf6w%c3R-{gl*LZ973uuiiq}Kt
zx}1CAb^K-^n_jckw3OD6pl30UIYV!qj}WerVp^`=Q*}fvr^d)i>Y~TeKWBsP{X3B`
zeQ(aZubWjXP6nN~J|Ei~D_@lBEF$6%c7-={LuUauizd&%El0y|BypKXDb+K5$U5}4
zWxaaMa-QinQ$nIt&iVVl`!6(EZPJ?=eCzm^xoui!?Dq8Jp2?zr?5yNuH%{HQ=E~M7
zH~)C8DRjImQkU~LPW0)^vle|ZZn5X?JJqUPzx>ib{BmY`met#S-{m)FN1WH(6?`o2
zhh}P#@BTo~$eO*25(REqy_r*|w^6j$I(D+!*J<A_CP+2>ot<BL;PAxj5psY0CrW%-
zB)|PxvsO*lzvnq@&5w5K{PoKFzk21>-<SB-PZABfwEWi02Ek*(Qt5iz<a>Wt{4=v$
zF0br!$-Oz|RK~H?|4e*o0m{Xdxk+}wylvn-VX=L~;*&fK{s9c(%B;$<U;-!u28S1}
z)qG}S*~HBh2$U9Q6ORHDFzKSywBq-S3=CYq7Tjz0HVtOs371!wH-hQn30F{8K;mgz
zYrDZM4A-&Nf%9OZj9g&rdk_6yu;!t)f@Av|hM!+bZZWci8*v!Lz*Ko{ulgL%8uOQt
zCtO8A#R?`I;wPm3VxkQzNSITFA10h}Y^t<zeduc@o^VBVMID6WbggwAV2Y+(|2l<H
z@)rY8uLO&P8cfll5WAQ0^4E_8g$-B?lwhi1(o9_7;D9-o2gQ|74OqXh^NV+boFl3t
z57YZ)!^ch0jTbe6&XG}<K?)sBYfX4~{1ph>bT|LVGiIJ}DRn6=n0-9qa_VwOJUweY
zJD9ouo^memuZY{p4)lS#1d<|2bx9<iw7N7BPgY$PiKlL@4mXY`T*F!eiKk_)1?Mpo
zrRF4p(m-)ha&a(-0VY5U2IlY_0k_MI+?LnbB7%ZL!a^Csz~W#c2*C_NFhdc{2m~`Z
zHzg+n$n{S1WoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&
s$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL$>iqe=)vd%0HIXTs{jB1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp.32EBBC25AA6C0CF0.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp.32EBBC25AA6C0CF0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..51cf0aed40d2a3b35cbb7618d433b7a076b3eae4
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nWvc<7#?slFjUMr8+JeVwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MU2@UY^>GOLdptZMXMj
zo4~lk#@tUsAa2@brTeuOLXtIe{vXv}cmAAx{D;p+f0QUx{Lk1Q!N;dF?fB%z2VYVY
zD%L#u^ebB^rhbpb|8MHL`~KD*oF1Pi_9v&(QhIX3XaDKm`~K8?D=qx|v}=Cs#3z;y
zF4x>+oW7<+bdwQ>YSX<9w{kBQC${e7wla9R>K5xpqZ8L(=t^zN+7jp!ohj9-zvKU>
z*>3Cid?;t<H9szShb#Go=W^{W6T>p5o;%nfI%~qVd)D*sNq;Hdcx9W%?AR|;_T0Yn
z(`$+Mtydh3zJA%S6I`uxl`mLsakReO`E+-={5@x-YjIK@3*1z5t`zZGdaKFDlx=Zk
zO<cWL$1zb)YIFGF4c|0eRU#Ie9P0M?en_r_mz6Q=y_Wd<s}K7u++r2)a$GbwvXZar
z*rS}UWqZfD<LbSXRZoAkq$ElUT)VmZlqsXtUiKpy-2t5bKP@=+FD~C87{|R-b>XCK
zhs1roj?W8N>b&59^SX-~9viP6%GhOb;N;O4k79ePcdDM(Q0Ww~5!h|j%2<*p%erOq
zRi~AQtyoU9n(aN1?66<aH%d^`*QAF3$g${8!H%vk6|*!KKQ3%<yHt6~e$G6Hk_k~;
zr%t-7URbTSvNtw*`_>Jj_tf57-Le&P>D!R?NygU1QRD`*ufn30b6P`VkNnQCvl866
ziZ|3p{m!C@!)nX&JRNWSo|@#gyWzh{hU(M_m)B0)%N??1-_ht#TRP&z<F2FzU5qGN
zRxAH$>8-u)il<XLLzd|6c`K3X&^ljV|M}a`P4hHEf6w%c3R-{gl*LZ973uuiiq}Kt
zx}1CAb^K-^n_jckw3OD6pl30UIYV!qj}WerVp^`=Q*}fvr^d)i>Y~TeKWBsP{X3B`
zeQ(aZubWjXP6nN~J|Ei~D_@lBEF$6%c7-={LuUauizd&%El0y|BypKXDb+K5$U5}4
zWxaaMa-QinQ$nIt&iVVl`!6(EZPJ?=e7BlcPF(Z4C$Q4!eG2RT{yum2oTa+jrMjtR
z`-QS&0?b?GqxXlqR#qAJDNhfa{&{=A{F7fxtj-)Qxixc^-|ythncvi|6?eUxvQfOw
zYi7*l{E#5A`1KhQE!+64m(TakaeeG}eTnDZrL`tA*b@Ffi<bRR@WgZ)`~P4Sw!JTo
zul>1zYyP7Dma`8osL=D?Kk4@WtgNs5wj5ra(mLhK$}+WtmX9sRW_G`Ie{45zzwf!1
z$2lf{xu`JRXjb9P{Z0)5MhwN3xk+}wye;9(d6KEB|0fTFe*i<cGOKbdm;lOv!I4W}
zn=k1aDRMIf0;R>-#G}9jOu8sFt@u470|VEu1^1f0O@mo@!sXTFjbOTX!WGmNka*hG
z+HP<Q!*#56;5?WpBNy2E-b23^ta)gy;Mo3#;pdl<TZ}B>MjS>lFjZdLt3JoG#{6aE
z30F~2v4RPQ_z9`Mm}tWa66RFlhY3er_#pcH{O)ByVF?xqHJIZvj!l&|t`B|9#1pQl
zuBZbO<_Xue)^&jCJrrX1GG6}raiCrU76T=iFibHM7dSXz&gDUI<x>OJFYNr{-5}?P
zs>s9ie%bJGQ*`4+O`yFp>M}^7qiL-P50AeBVVmyeA9=>i6E3ALr3JH(CtOZl4vD8{
zt!D=__uo^_1^yLrJK2HWRF^<fB&jZm#FJK+M&ik;%OdgAt<~Yi@q}wwYasEothL}g
zhN9G*L{J(iE=n#A1~I?{h{3=R2BCl?bNIwn=ca1xwLQ!h5fmH}7RnF=mtzP)Fhdc{
z2m~`ZHzg+n$n{S1WoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|R
wQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL$>iqe=)vd%0ALKxU;qFB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp.5C1840A2D370420C.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp.5C1840A2D370420C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3da36c72d9e86d101a5353e785399067a9a6ca5b
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JeVwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MU2@UY^>GOLdptZMXMj
zo4~lk#@tUsAa2@brTeuOLXtIe{vXv}cmAAx{D;p+f0QUx{Lk1Q!N;dF?fB%z2VYVY
zD%L#u^ebB^rhbpb|8MHL`~KD*oF1Pi_9v&(QhIX3XaDKm`~K8?D=qx|v}=Cs#3z;y
zF4x>+oW7<+bdwQ>YSX<9w{kBQC${e7wla9R>K5xpqZ8L(=t^zN+7jp!ohj9-zvKU>
z*>3Cid?;t<H9szShb#Go=W^{W6T>p5o;%nfI%~qVd)D*sNq;Hdcx9W%?AR|;_T0Yn
z(`$+Mtydh3zJA%S6I`uxl`mLsakReO`E+-={5@x-YjIK@3*1z5t`zZGdaKFDlx=Zk
zO<cWL$1zb)YIFGF4c|0eRU#Ie9P0M?en_r_mz6Q=y_Wd<s}K7u++r2)a$GbwvXZar
z*rS}UWqZfD<LbSXRZoAkq$ElUT)VmZlqsXtUiKpy-2t5bKP@=+FD~C87{|R-b>XCK
zhs1roj?W8N>b&59^SX-~9viP6%GhOb;N;O4k79ePcdDM(Q0Ww~5!h|j%2<*p%erOq
zRi~AQtyoU9n(aN1?66<aH%d^`*QAF3$g${8!H%vk6|*!KKQ3%<yHt6~e$G6Hk_k~;
zr%t-7URbTSvNtw*`_>Jj_tf57-Le&P>D!R?NygU1QRD`*ufn30b6P`VkNnQCvl866
ziZ|3p{m!C@!)nX&JRNWSo|@#gyWzh{hU(M_m)B0)%N??1-_ht#TRP&z<F2FzU5qGN
zRxAH$>8-u)il<XLLzd|6c`K3X&^ljV|M}a`P4hHEf6w%c3R-{gl*LZ973uuiiq}Kt
zx}1CAb^K-^n_jckw3OD6pl30UIYV!qj}WerVp^`=Q*}fvr^d)i>Y~TeKWBsP{X3B`
zeQ(aZubWjXP6nN~J|Ei~D_@lBEF$6%c7-={LuUauizd&%El0y|BypKXDb+K5$U5}4
zWxaaMa-QinQ$nIt&iVVl`!6(EZPJ^HL#w|Yyy9`K^w<iUndO_B?z<mfti0{gw5hwM
z8O@IGG}T)%`%wSd`>Pk(?3Fw|@$`z*KciQie^O;7`z*=ot&i{WzZ+kgR13ehbS*!1
zgWG<o&$^f9TAJPK!%cXcqTA<QKCilMQQ`9JOH1xuvXk;*-cbK}EnkJjC#loT|22i0
z<Gwt;Ruj<bfARmEvk3t<)0f7dd|Pk2wJQGB!_ZAmUcW-Dg$)i>I6Xdd^xNV>`SbVH
zo_|Sh@%)uB!CUg|jyLg!Isv-HmAOfFz`U({SD+%t#pMV$Q(yo?xHy}56qo?YfI%PU
zJLjFZ+O&8W{DIQStje)q0w!IQnpXUtk%58h*MfV^-loAUJmK={@<uR2dBPRc6_9w^
z*4l1x6T@|^b>KXh4kj+J^}UCFFIe-?T7f6rh{Grbrp2-S4a3hbCASz^P=vj<SAC9W
zjrq&S6Rx75Vg*wj;wPm3VxkQzNSITFA10h}Y^t<zeduc@o^VBVMID&MJmI?5x(+Z!
zFq)AI>|L014u#mgjF-QD9B8Bgi-8hM)l&o3FYNr{-5{GpRpep9Up9Q)6y11H6KIo+
zx(re<Xj*H+J^o#e`R^>vBQ8L_5-bvGFz5Ug2-|cw|Hw0Do^UC3DJ_^VPq>`A91>5@
zTF(w<-M^=t3;ZkMcCrIaQkOtdB&jZm#FJK+M&ik;%OdgAt<~Yi@q}wwYasEothL}g
zhN9G*L{J(iE=n#A1~I?{h{3=R2BCl?b9lzh4-;BE^W)hff`UWBLK%YKatt8|W+;Lg
zfnX-*rsQM*x!!5M%q*;I>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^
v%4Qn!*hIyoWn>i;l~h#KHMMl~4GfJ<&E-s_bd6j@#oWCjncVywJs5oerj5`)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp.05BFD4E4455B0BDF.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp.05BFD4E4455B0BDF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f5dcda37c76c7522abe7823614f373dfe39a14af
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nP-?87#?slFjUMr8+JeVwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MU2@UY^>GOLdptZMXMj
zo4~lk#@tUsAa2@brTeuOLXtIe{vXv}cmAAx{D;p+f0QUx{Lk1Q!N;dF?fB%z2VYVY
zD%L#u^ebB^rhbpb|8MHL`~KD*oF1Pi_9v&(QhIX3XaDKm`~K8?D=qx|v}=Cs#3z;y
zF4x>+oW7<+bdwQ>YSX<9w{kBQC${e7wla9R>K5xpqZ8L(=t^zN+7jp!ohj9-zvKU>
z*>3Cid?;t<H9szShb#Go=W^{W6T>p5o;%nfI%~qVd)D*sNq;Hdcx9W%?AR|;_T0Yn
z(`$+Mtydh3zJA%S6I`uxl`mLsakReO`E+-={5@x-YjIK@3*1z5t`zZGdaKFDlx=Zk
zO<cWL$1zb)YIFGF4c|0eRU#Ie9P0M?en_r_mz6Q=y_Wd<s}K7u++r2)a$GbwvXZar
z*rS}UWqZfD<LbSXRZoAkq$ElUT)VmZlqsXtUiKpy-2t5bKP@=+FD~C87{|R-b>XCK
zhs1roj?W8N>b&59^SX-~9viP6%GhOb;N;O4k79ePcdDM(Q0Ww~5!h|j%2<*p%erOq
zRi~AQtyoU9n(aN1?66<aH%d^`*QAF3$g${8!H%vk6|*!KKQ3%<yHt6~e$G6Hk_k~;
zr%t-7URbTSvNtw*`_>Jj_tf57-Le&P>D!R?NygU1QRD`*ufn30b6P`VkNnQCvl866
ziZ|3p{m!C@!)nX&JRNWSo|@#gyWzh{hU(M_m)B0)%N??1-_ht#TRP&z<F2FzU5qGN
zRxAH$>8-u)il<XLLzd|6c`K3X&^ljV|M}a`P4hHEf6w%c3R-{gl*LZ973uuiiq}Kt
zx}1CAb^K-^n_jckw3OD6pl30UIYV!qj}WerVp^`=Q*}fvr^d)i>Y~TeKWBsP{X3B`
zeQ(aZubWjXP6nN~J|Ei~D_@lBEF$6%c7-={LuUauizd&%El0y|BypKXDb+K5$U5}4
zWxaaMa-QinQ$nIt&iVVl`!6(EZPJ^HL)z5_rjttN8fy7Jo4e-0_lqB2bgbQyx^&&r
zGta`4dbwXM`;h$l`zj6p{9}brj9wM}6MJR6bDm%O&KZ8z!Iv}ZbF9wZJ5uF$ajwy~
zL;jkRvw!t^oqF_Il(+f9wZ!LNiaXb4+<7_6WaaiP{cM*R-rTQQ%`Dfq=h&wQ^<766
zi2vPTDj)JN_=|nIv4qI^)R*B!zvF$^-kV-lF?G#^Np@GyvGf?syI@hg@o&bR<Da*m
zwEQcPILS`xNsybd-f!_#w*Z^s%G@M7VBY3-TdMj{ah3r$Q(yo?xHy}56qo?YfWbjk
zanYO6?iM@@{y=GER^?bQ0h2CDO)GxS$iTq$Yr(x{Z_{8Fo^W|}c_WyiJmCuJ3P?O{
zYi&2UiQzidI&dCL2NM_A`rbpo7p!?`t-up*#9<Tz)8g3vhT-Rzl3R={D8gRbt3JoG
z#{6aE30F~2v4W`%@e@*iG0}z<B+RM84-?KfHdWfVKJ+ycPq?DGq7KYro^V}jT?d$=
zwsRjQF>3p50_v4ukx+vvf=M%SfkOf2%0nS`FXQE}9|x*7U@=gFsd{R_`h}feyc^^I
zQ5AWZ@RtoAH$^vI)C4*}MqLIebTq9s;o<RDAZ*j!{3FkpdBUaCrL<rU;0c#gmqX&|
zS?k%s%>DP2bAf+F+)j3&o7E+d6iKQ}BJrfvrIC2D>as{Yb!&CFaXjG~)*47WEo&_}
zkD(|vClQneii?tqgFy^10b(#Pgh40}8H8Ylz?sb9y&Imq>V6e<mn|YFI3z5TArvmh
z5P@JO=ceRj0J+|2zRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6
wR?21?^4LVhq-A6k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_BAML$96cC)0JXNdGXMYp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp.BFA2076F1038BB65.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp.BFA2076F1038BB65.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a0b9996ef0ab3246589f737ae537bc992daf413
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g@VUwuQhq`(ppO
zy&fkDbJHpsOm-TDD_qWQckfkNo7|gvQvD?7m-YLn$_g8NS?KUSO(^=7|E<5@(kC&d
zFzB7{73ON0e)Of<{=Wq(Mn5e6ck8b^cg|k_$Iqe~BY{2jTjJNTu}@q5_@qODkt~PY
z>K!|N&0-0c7k_^L-;$%-_lwtjeOh($e3<z1WTAq8FO!aj&p*FUZ~gw5Pn*3b*|68l
zI&b};Npzi47|TMJ3(;Y_Wvleo7?eA*tHp-yR=CDi*?DhK;%hhQqg$pf3;faaTYrD%
zms{NR<rfNS=N#-cXq#8E%Sk%3HSDEKPsC!*%-4HAzufCuFV3B7o>?qytNLE8e4o(E
zN!zTP9424BWfgJR?>0vXU+#r-Wl2BZ-EIFqp=0ZGHlY=PoYGtGIG>wzl6|^tZlL3h
zmm1swGrEtxxvH_o)=Q8xOr!UM(xtr{*khd?8*c4Bb+o#)!dqtXx(RtLFQg^s^4q%H
zn`k~&?w#V1tbC)ZMKw-E$<hMXZtgy1%4oHh{YXZ40H^;?3y%GZ%QuL|aW7R}IBDCV
z{8>}_Y!`a9J2bY(EDoHIw6^tPNnyj}?u$ovOMNf$o)_pLq*$R)R?5ZvqPJA!f?pW-
zQoh#$3gWlEF!eM3oo15hy2Rw-2U`K@{FDByqD$LXEbZw#E-S_LT=Rc{QR@qZty$Wt
zyZs-1iwTmBj@rI?gYZ4I_g1%T#a#L}<b9H{HF3<@z&^{tCD@XC^{$TZ7i&tCl0v0d
z&G6gd9C*lVS)Qlkt>06V+;%toH_1?)I^pu#X?wXtw(L6^{b@@_T>hRF-Afj2eB$%l
z?uqx->V7A&lR_&!Vk<5io@(H-kB>cb`+3tm&CuU7J)?rwAG~D|DRgCX`_>7nq3RQj
zcU}7UMT2>|qb&C(CoRq5bst-_-ug!h*GMs`2fNuOndjKZ%Jya~3Hs+3eDB|h2WHhb
zZC<B)f0(p%UirM;QoHS*+-lR%XvhgRUvg0GkgmWIgCCd0;uE_y?__oU;XJUKbGvYy
z-;cFMF`ohgGACNt&%9TE`qP3smDg6uA(PHnM;?5?`0+)-=<L&6vEDP!>brJ&zgqAi
zx%>Mn4fFg{*-wmKrTq)jGLAf#cG%*4+T33^*VvoQPR&m<Tz=xLMD;^|<;kkGZmg*<
zzIL_sZ7;oQ`77p<mQ?&MHQ!g^x7?dE7ktg%cbSpjJubQOgS?{Q3s?KRt?XC7iTpi(
zlSk_RPMy7~+yB>cef@JgF=RvUlrP1veGYIu=Im7$zkRS?{_uUZ=U)<AJf;3TadR`)
z{w?0L#j>@yGB?Q%n5WrY7A4tlUHpfeDKLN`T%1kZ4@>}Mz@SX(*ptaDJ;FQ;{y=GE
zR^>o20h2CDO)GxJh%kkT3#_~M(C-Cn9$G8#gc}MQM#IeD`nBL*v$tt53s1Pby1Wt0
zW}a{bbp<4zwzak!ObbuAj<pV)=h*&+;pdl<TZ}9y=6Y?f`W(+1^Oun)Ttz{}3Z^&2
zPe}d6L>pF+FsBMXOgQ7%RB7Y-(AP{n;fm^tItW+jTI)K%6v1dlF0glD{+V+9>l8-G
zUkpGaC0HcXV0sUQ*u9LGzkVDjY`|im1k?P~fb|PIzj!yu0ir7MFySv7K5mL`yr>Cu
zfQ-5flE*czHQ`?U_mp#ie?{C*E}%{75?U~uc)}&sC6Rbi>QYENX?1BNo~*ho5>HND
z4vD93t!@W%4^Oy;wFVMT%UTPGr)RAP=M|;qB!XhTxG1?e7{rJG5fA|8LzrPOF$RVp
z1TzG|423h9!}m4L&9qb$Ue6X06dV#3%8;C!l9K_H_fGR=W?^Mx=iub#;T7QL^AvP-
zRukm%bP^I57Lky&u(Y<ZwX=876V<j-Hq(&DCMqT^BdegOq@t>>simWDU}$V=E@vX8
TYvdv-=I#~A<mTt-!RP}3>m<3D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp.16DF4E204049F20E.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp.16DF4E204049F20E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8732f9fe6c321f9bb43c733a3e393aa67621e5a3
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g@VUwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MWRH`!u2GTmHBHeoLRk
zn8KiUzE_y5W%|=KC(7mfT-fBF*MIDOz4`Nb@jo?&cCsAz>~HB$Yi90#`Y}bp!d8+)
zZuO2m|7NjVm)AeH|F36vetrChr_*P%+DFHoJErpB&*i6+^zG-@#@?yf`6xYHrIPQ5
z$^7yIs@k!wYj~VG7nrXpi~bT5k@&8O&u7=FGRHMMpG>Q|+ty0uid+tx<@S*K-Tptz
zE^6!lx$AiNY$4nBLlRY&UUKCw5sg|}Y>?Edwq))1bIZT8{f*Vxs(UF-+*b9yS@}Mp
zmy@<xIXO(ee9J20qTg+f629CE=gN|PzPsE0eL~09>1;wP0y(9(-f=!R=_LDf+1x<K
z8?Q9H17>s|dvjG|jjfj;XP8Fs2c=7UH?YS#J2u?fed=g+=?>+-z;F+9r!T!d=i23z
z-k+H5#s6I(X{-6nEK57Z$Qx`erQg;S`8N2yXEeN}6mlroj_=U>7qU4m*PSvwS1eik
z!FBSbjp`vU1vEZP$Tr~$TC>$)mL6Y0k>Re2@W*j^Cylv0m00+Xtn*7`h>dV(e4Da0
z;ME2`#-fG3`X3}N<a<mGbMX@9z0bVy<Fr2^0fAK=QKw#1+*!D=WZozK<>~^mOTu!O
zrhNCjGtWCK`TDh7-JI6#CwKRi^|MY?UNg(ao!?i0^_zomM}}5<VyO5>yIK8yE^D%!
zMK5!1*SNN!b7geUgl~0AXI#{KV9y)nsrsa3b?SR2(Y5&>r`hN(y51^YIy1y%+MLY$
z?KLaQ-UpsAic}I^$(?WQCK>SX`RUV@zxOOi_qtk_Hfh?G><vbK>o{NCJnVJCG*o?}
z@vcjSWm6hYFX&T^RGd1+@_L2S)NjVuSmw7Kc*(g~ZX>Jo{GL8HlOWCd%~Q798{J6N
zpPl|Uhcj-|mCxsj!;`PK$J`cRWt|{e>MWI`bce~2tJyyHqv*C7hkB=-n9s0hR)O`y
z*PQn?53BO4P8I9?e0lQk_qXJ(^ryt8_jpaJRM-AdTVi9Cq@5eNGFx-=&#4iQo2)ec
z+<3HC>(trVpQLv7?3%fM>M5Uf%P%L+S$^5?Z^<qB*?ybNH_r%uvaDzChvgGbn%26h
zroQ;v)z-JY^s42rm`hqx@w?P~UxnXtZ_ZrsHGkh#Mt=9W<jN27iiR&-?en&>U;QTX
z_xw#BssB55_Ns3GU(5CN&+WvJO-^3Fc9yCow0vwiHo5!lgZ=#v^CuVoGB~8dTUVsK
zc;?i<+*{KurxaJ_CfNb=w8`W0V4?fRPjWK_1~7z+vx)nG37`xZ%wIkG8TY*Xn|K)f
zfzryX%7I`4CS8=8R{V^Sfr0DSf_u&0rok*c;qvP8MleHp!WGmNka*hG+HP<Y!*#56
z;5?WPCN8k`y@!4;So6?YfhXKh*f1KV#j*Vj!_O}zw-{MaguS*`eU4|1`OC-?uA-n~
z1ydd3C#3#jq75rZm{WxxCLDF)gXr_~yO#llC0HcXU>0W_n<{NwANrb!CtOioQ3oc>
z6RvBm>j2XWqZzrtp#XE`p%A;5@$%P?1C2CbF;If3dTPM>g`Ho#8{_~{6?vHOmkl2`
zMK@m51llB{E`t<0n%0`|@c8$XbAf+F+)gf_E7c{m5YCZQmqg-8sY@a8q}8R7c(UrU
zNIW@pIV7ICwYnY5Jv`wW)*47WEo&_#o}RTHoL7{ZlL(6W;-ci@U=RaLfEWzS;rnh(
zQ&`;3y`L>2C^#f6lpzc(4km&S%n$@K6v2!@Fq3mrax#Ei?=)X#7FITP4o+?!UIBhS
zPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlAvI>eyDyr(5S~~g$hQ_Am
Wawbx`MlPaa?p~2hZhnp)j6MJtN#MZ%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp.25AD39BA7B2B9C77.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp.25AD39BA7B2B9C77.idx
new file mode 100755
index 0000000000000000000000000000000000000000..01b0bfc2e855ac318ef54a91fab47fa940876ce2
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g@VUwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MWRH`!u2GTmHBHeoLRk
zn8KiUzE_y5W%|=KC(7mfT-fBF*MIDOz4`Nb@jo?&cCsAz>~HB$Yi90#`Y}bp!d8+)
zZuO2m|7NjVm)AeH|F36vetrChr_*P%+DFHoJErpB&*i6+^zG-@#@?yf`6xYHrIPQ5
z$^7yIs@k!wYj~VG7nrXpi~bT5k@&8O&u7=FGRHMMpG>Q|+ty0uid+tx<@S*K-Tptz
zE^6!lx$AiNY$4nBLlRY&UUKCw5sg|}Y>?Edwq))1bIZT8{f*Vxs(UF-+*b9yS@}Mp
zmy@<xIXO(ee9J20qTg+f629CE=gN|PzPsE0eL~09>1;wP0y(9(-f=!R=_LDf+1x<K
z8?Q9H17>s|dvjG|jjfj;XP8Fs2c=7UH?YS#J2u?fed=g+=?>+-z;F+9r!T!d=i23z
z-k+H5#s6I(X{-6nEK57Z$Qx`erQg;S`8N2yXEeN}6mlroj_=U>7qU4m*PSvwS1eik
z!FBSbjp`vU1vEZP$Tr~$TC>$)mL6Y0k>Re2@W*j^Cylv0m00+Xtn*7`h>dV(e4Da0
z;ME2`#-fG3`X3}N<a<mGbMX@9z0bVy<Fr2^0fAK=QKw#1+*!D=WZozK<>~^mOTu!O
zrhNCjGtWCK`TDh7-JI6#CwKRi^|MY?UNg(ao!?i0^_zomM}}5<VyO5>yIK8yE^D%!
zMK5!1*SNN!b7geUgl~0AXI#{KV9y)nsrsa3b?SR2(Y5&>r`hN(y51^YIy1y%+MLY$
z?KLaQ-UpsAic}I^$(?WQCK>SX`RUV@zxOOi_qtk_Hfh?G><vbK>o{NCJnVJCG*o?}
z@vcjSWm6hYFX&T^RGd1+@_L2S)NjVuSmw7Kc*(g~ZX>Jo{GL8HlOWCd%~Q798{J6N
zpPl|Uhcj-|mCxsj!;`PK$J`cRWt|{e>MWI`bce~2tJyyHqv*C7hkB=-n9s0hR)O`y
z*PQn?53BO4P8I9?e0lQk{H=B?<WKH8Coxs!vtQVU-!Cdkj)dLX6m&f(?X&jAW2~hC
zKQ<n%4xKXF{HbY?WT|nz_SBg%{>u+r_%EOP>*W@Ev$<*cX@<*B_)1nk^jDso{j108
z)P>igyt4VPR~3KRxg^wNUy1LmE9<xPvt4$0eZOWkvs~YvW1k+_J0)G{uDg@Pzv`{#
zulbuLP5u=}R(a?B|1I+B=k~)ZQd*~cDSqv9fa5V|ue$i{gZ1)<@2frklGx%Y_2-G3
zo3Zw9@g-fB{l%5JNp`?I&7aXyV8+zi&CL`Tzz{CZChi9&fHGhpHmfxuc;e2bJPiIo
zX=PUBKrjK5E=o-+e#Xea!1Zgvy=HIIU>2Tmd3AXsn4vu33hD|-JZ)=jH@J!6I@UUH
z9!v)l7ufpVL%$cSd1$S`6K*JM7!A_`qZzrtj({n0Y=6V>^GnGsMivy+UfZia$Fs)#
zW#kE0QBbjhX%6udQhzbgh7}~tslpEv&Nwzz+PFUSH4{&`qPn6E%so8ey4Jc5Fhz$#
z>|VypUq24iYrtZl1QUL0!1{%qU%VS+q^OEKO!&)&kDH<!FKPmflu?&K3I<JUO}NLu
z%Q63*rFp~!s8@nTLJel`zo(oF{43&iash?aCA46|JmHe+l1Mx$btxpCw7N7BPgY$P
zi6^Hnhs0C2R=0yWlP6rmS_6rvWvzw8)3ert^NLb)5<xLvT$EfK3}S!@5QBj^Tw%Yi
zey>w^Fk3`Wa7b7vLl{^bOavjAAqZwDf*FBeCg-N)WB|F|X}-)XtZeKYoZLLT0{ncQ
zg09YLf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=Lbo31jjZMwv
VOr&&;Ttvm(y&{?1{2V<PeE@I#wWt69

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp.F46067E1C6499190.idx b/.cache/clangd/index/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp.F46067E1C6499190.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0473569119788e5b6246dffb8ebe14388950507f
GIT binary patch
literal 1772
zcmWIYbaQ*c#=zj5;#rZKT9U}Zz`!5~#Kk2=nMarz7?yA|FjUMr8g@VUwuQhq`(ppO
zy&eyCm>E|#Oo^GPJ>lin!;9S}UfZU&>B;0LDqq&`pDHVC@MWRH`!u2GTmHBHeoLRk
zn8KiUzE_y5W%|=KC(7mfT-fBF*MIDOz4`Nb@jo?&cCsAz>~HB$Yi90#`Y}bp!d8+)
zZuO2m|7NjVm)AeH|F36vetrChr_*P%+DFHoJErpB&*i6+^zG-@#@?yf`6xYHrIPQ5
z$^7yIs@k!wYj~VG7nrXpi~bT5k@&8O&u7=FGRHMMpG>Q|+ty0uid+tx<@S*K-Tptz
zE^6!lx$AiNY$4nBLlRY&UUKCw5sg|}Y>?Edwq))1bIZT8{f*Vxs(UF-+*b9yS@}Mp
zmy@<xIXO(ee9J20qTg+f629CE=gN|PzPsE0eL~09>1;wP0y(9(-f=!R=_LDf+1x<K
z8?Q9H17>s|dvjG|jjfj;XP8Fs2c=7UH?YS#J2u?fed=g+=?>+-z;F+9r!T!d=i23z
z-k+H5#s6I(X{-6nEK57Z$Qx`erQg;S`8N2yXEeN}6mlroj_=U>7qU4m*PSvwS1eik
z!FBSbjp`vU1vEZP$Tr~$TC>$)mL6Y0k>Re2@W*j^Cylv0m00+Xtn*7`h>dV(e4Da0
z;ME2`#-fG3`X3}N<a<mGbMX@9z0bVy<Fr2^0fAK=QKw#1+*!D=WZozK<>~^mOTu!O
zrhNCjGtWCK`TDh7-JI6#CwKRi^|MY?UNg(ao!?i0^_zomM}}5<VyO5>yIK8yE^D%!
zMK5!1*SNN!b7geUgl~0AXI#{KV9y)nsrsa3b?SR2(Y5&>r`hN(y51^YIy1y%+MLY$
z?KLaQ-UpsAic}I^$(?WQCK>SX`RUV@zxOOi_qtk_Hfh?G><vbK>o{NCJnVJCG*o?}
z@vcjSWm6hYFX&T^RGd1+@_L2S)NjVuSmw7Kc*(g~ZX>Jo{GL8HlOWCd%~Q798{J6N
zpPl|Uhcj-|mCxsj!;`PK$J`cRWt|{e>MWI`bce~2tJyyHqv*C7hkB=-n9s0hR)O`y
z*PQn?53BO4P8I9?e0lQk{8au|?wi&byQ!YsGg<VHot2#I#;M!ZT-iG1<{z&$g^sc!
zbvYm7M4vwMt?Z3)yL;xoSE}0e%P$SiUCvxyXZ5z<clpiPk!LiEf{(@hxP0QtEW5>%
zjJ{O4Iv>Bi>y=EcUZ!ZTe(Yqmui9@HA2hk}%l!VUhIZxk8*P5{PmuVsNPhdR=B(<j
zf6sH+djH2vyEp0ff4kPIe{T)6A{ABt-mwxkI8@>E_~g;D5BHCM*nZNoRw8i{oBhs-
z7tfsf$9mnya$a#|Zjv1^PaAk{vI^0uU&_rC7{CxN&L-{$CV(<v5Zkcz<g`dGSsn&|
zptLfpav+$1Nf)K26+dHSVBq?-;9j%0X)p^<xV*Z&5zJ7Ya0PV*B%Zdlwj12Ua2;zM
zI1i?Si3@Cf@1frd);zRU;0ZSrHjIX8acqCX@bgQ_Ek+g;VXy5~pW|6${xb4}t0<^g
z!BmI%38}xBXu}E;=2YQ_3BzbcF0fBvHf0=}Ds5aJ`kIL+Tv1(72j(80a9wL%2bi9=
za~~!#YWr;h8Y#gdp$1cQD8%k%y!`d!Kw$$G10|U1rv|KF*!jh~K@JdAk%tL?+3;~w
zbmK)$piMIBGDxALX{`wlkAF`&7x-7i?c@U5q%NTavxz5MQe6^>C#5ch#FJK+M&ik;
z%OdgQ)a8(P>elLZF!%6;YglU_@wBY9ka&95dT?G*YEB|3=8KDxi-SQ7FacsPFo&!5
zJYOt6llM4VL{M-@SSUjnSR70QA($ZuW+;LgfnX-*rsQM*x!!5M%q*;I>>QljJiG$@
ze4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#KHMMl~4GfJ<
X&E-s_bd6j@#oWCjncVywJs5oeKUKFl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_contraction_utils.hpp.39AAD30D5F9C497C.idx b/.cache/clangd/index/device_contraction_utils.hpp.39AAD30D5F9C497C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c2cbd1e1a1692f11627d4b6e09a3ea72f71c087e
GIT binary patch
literal 1434
zcmWIYbaR`;%D~{9;#rZKT9U}Zz`!5~#Kk2=nPyB341a+F6?4`G`}#{83fwKOJ1UdG
zdZ}qP@7Y9FkD4t(q8FXlui)w{pS1PE@oUNIjtA=Bze$_zq@ty=>(ec{YMb|aC9S0o
zEH%-Rl$@w=jU#N`R$sB0O~yPAo;zf^{m@?6-N2^Xe?@QmxnfVj5Z${@yIDhKnR`~7
zKk$`wWMg?MyU;68UvLX^_qF!liw`pFTeCEmXBsoZk9lvVZo6K<<$6zA*v#$iJ<9-H
z=@j7yY*FkDg5}38YtAM9+5ElJF?`JfkvzA@{+2x4@7(_MSKj-%=Z75QI&)_886Pg}
zs_8cPe(Z<nL7R%*ytg`MHmLoms${(ss?huL%l=8TPOkf6_RIKdPV5qns(q~yit-h1
zb2G({>wevHwD7^)o$ViAS7^`oVcpF4;_8(7!m7XB{k`j>4PJ2sC!1_8y67BJ8No7t
zmfpX&%hV3v_N&Pk-G2Vu=4EeE-X}(6g+Hv4`mMz`zvxD_Mp;<*^38wuwg=u6>YdZH
zN@x0tMf~b}?@QL0yb#;C?A@P__O`!8>||5atshvvc`)zoenz8q^98agx6_MtuiFQN
z&GfmL{%W0$^7A9@EAtIboH(g@nrnJgVq{wJS@SrK!0PTrYLT9oyf0Sh{SUFqeCIhi
z;GWsMCubgf;cIqQ-V&@9s&Pej`{G@h|C|@ZPD{F|vQbgXWdE5uyQe8f6hdDcJ+)4K
z!MZPh5y#dDjZ%rd<}Du!B03(_igX@|-@|{D)%;+!mr!EI`woW~<|*G~1<gXw)-m(_
zTWeZenVX~tOhp$m1#j4<74z~iu&^<3D6lGsf(ZsrW+p~f21X7>hN9H8;-kQHCfBxc
z*?o(Pl1wZdD$*))FpVr6I?_6Do;+WpWhcj(hfF*i8p0Y1Fhx8Z2Eqn#o|W9kQ`u~Z
zA30b!RC!eSV2W5c)Ogg8c<Mar$UJ#<MTAOS9$h3kJsv$Io<5I0GEZI~uFvDjYN^~G
zm!pAR(BRV$L>OnlX8`Ag_z9`Mm}tWa^n!w-f;vnQkmsi01LGC1y0m9fM?x)7krKa>
zFie<-LyKPv&g=GKZh71P-+~ROS3^M`p-913Aq2rQVKL!G@Z4D3ka+GK?)(V3Ag&;|
zGv_(ow~-6<<^}pi-9+6Crig_@-$WnIo44nCSyVyI6i{@iYG}X|0i#3J0M6UR$8q6U
z#1mh<CLMNpm%{V)U@Hp?hbf;a+)XSTR>D>yFmqWrJcT^r-e=+P*7Vkb$sK<=t6}a7
zvx^{Ggw;i1!a(;5>%)1ci&h5yew<s3*C{Zqj9lQP>ZwxJJ#m+F-ScBh3J$86p91<-
zg;zxirizC{hgS#A1E%JjL}0RKC@xAawg%En93QT9^yf~Q*uuie#m&yZ4w8fbE*OJ>
Kfg8c(L@)uiodZh%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp.F95715A59DC357E4.idx b/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp.F95715A59DC357E4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b1073d7b651f33fc29804402529ec7b1c8d61a16
GIT binary patch
literal 8300
zcmZvhd2~}%;>U9@rKAf>3vGc8Wh(;N^O6^!K*;{`^1_l*!XmT{0oiE=P$?3jq9Pzj
zWf?&>aYO+b6kN#2s6Z`RDC!wCkPZ>bQm~SO%ufy$%P;NkyypAIujg=2KDnRwN%9ij
z&ABP$9gbQ3BogZ^e)*!=%X~=^i6kTbZ|Sln^E$-gSeit#eAmCcPnOjgdi=Ea(nRyW
zUB}*MH?OT)7w)%crhe$LVJ{4LE&p_ngOyv=&Hpp<he6M5&A#}bBdZ6lYZ&<L!wcJZ
zDKGu7{2$7{y*ScS+~eWzXSNQ$zT)|_i(~4e#>*cZwy!W|o?P{Dv4YFZKKJh}l~*1u
z%wK+_{`Z!lsj;JF|M<LRQ{R%lWfy-tcjko|sxoz956$oQp800`Ue(_0t|i9f=e`~C
zF#GObjZLecoEAx)dpmT!_{h9P!>;rkHofWOZ$%dmeX?gx?~AFs$k^kKdHlAFe^v&n
z9$p(if7^)IGtYJI`ff`3&1u)VKRdaUK3`EO8@zMQzTotNIWtxsJ$1)&(4CP}cSTeF
zZRqK&*Vb<JR4o5)#PhiVGks`xX{^JwZS3wTC!c$8b@aNbh;?AfLGzOCTW|GW@ZI%O
zz74|Ws<NVPgDN-QcRZM%|3jCgu2P}f&XIX{cQu*oO1nAlES!C2L&3!XXZOjz>~1{a
zNNRi}dF$C*W8bM>P;dS5%~wDEOLVisZ9VbfQ<E=MZyDE=_kE4z;E6LUe|>3H#qn9S
zyXy;6Y#mi0`OY^s?V0e-$+KT>nO?Cyzk1qVkKDZ?<#N_^{;c<kkoj9per2~2-RcTn
zu6XbCl=QiS4~Mokmo@Phw&f*t8Fhl4Uzo`c{rlDAe)<dh_P$>9M&94EbvqCA`9)~d
zJih+u)g)o&;ZF)bYJOtz&#w7?Z}sBzOMyB2pY3yp+#l8Hb+)|ACjFiqjbtOY|HECO
z${AH_=YAf^*msEax(EM9Qd0T!E0;6Z{Lq}`e_7sjm7z3g`{qeUwnsZEriJU3<&UJ}
zvNnZh7k%Gz*@0lw{XumfcTm;LO_@}sS{ds(C@F1BN%TTpLuaG^&>uV2O`CbN(*5Lz
zmbs<l8I^a+Zs)keu9r5<vn;gyGGTvK(YhVSH(HBi&zf_a>bnge^Y+Do2lm~*G^^lL
z<@$%8y;9z+>|b~NM7iO7c2m;E(Pe#(v&({CZ@pLHm(24F(;qNyQ`9-1D8KSJ68(Pr
zdoO;R`K4p#C);-J*jRe_$d?b+&n{olvVZP?Ukcuuvf_uSp30KJ?`|1dyY;Pi4!z`z
zj9k#N<7jEco%%24)Z23pT%Fsy`<v(Qo&V8Me66MN^GRK7&i<=~Pj3&*-v8pVZ*J(P
zmi~PCPwn{F-fKV4I97G@dD-qo>zQ)*2aT7sD_0Ki9=|mF(ebJ8E->W&Ui;KHV~Sq;
zXvg@eqx<9*Y%1P#`wh>E>o;6qrLNkNx%Bb1_0^vF3y$AkFI#x+%nO=Md3h_=Mq82$
zqYpg1V7U7DqfWi+O?lCUZ;$%*(?_q|`2O+m%2WORX!&6HtrtGMwd&N(!50oy%o};k
zHBDOax1&GiQGR^jl}q0hcTnGZc;kBNsvg>kuq1O_;Mky59qL~$_))U7e9_EhL?TK3
zV_$TQz)nn~-l-B_ZgOr$+YrB6wEd-bKDB@E*tyQm5>I@(Uv@vXZ4jpiNgtgUcmAi1
zX>?kAx*)m0-ZqHS!w%l~`tIA~?{=0<h)?&)?lYuq5T{4I)$;h<wlDl?bYgtEXL8S!
zwn3a8(|qsFdpFCErqRjq>8#|ePHlrYy{5Fc<FHouOd1^*pUz0mNNpR$>9t!Y=s#__
z{%5*mN_=`k#|g%^L7aY{U;VuDH`8Znv?xBElbn;%Hi*;byLR?veN`WS>`C$IUdg>u
z+6Hm@fns5}+ZR`YY4pkXbbfOF$hJY8{<E9+*yXaSn`yK#KAoAInbJ0h(@SPQwX~^M
z;@lUeww?2hr<?b^e|2QPbf3&C^N9};C%<elP(b!vg5+K~m~y`yOnE>KroyWLQ{h*D
zsR$^*RC<+QD*Z|@l>sG~Dz6Gmm0ty>Dxd;W?Nx)R_N&2E2h?C{yc#eyehrwKfCfyh
zR|}@ruLV;Z(1OW$888_?111w-z|?tlVCwujFm(YPn0l`sOub(Zraqttll8J-vVIm!
zHo$^u@EX80_zhqh0tPTmUXxG!&zJ5q`Av&~0;cB@WcHfDH2ck9ngeDqEnW+l7QY2d
zOTYrA)oTUQ>bHVv4Oqdnd2L|Y{5CLc0UMZhuN_Rg-wvidU<cFTb%5#cJHT`V9AI)@
z4ouF^fyo6pFr8i}m`=YFOlQCeChz6J<o!IDe1Hej<#mDS^1Hxv1zcbXUI9$OFMufo
z1TfuRH<)g}8%%e=4dxVP%INlMpl>QO6=)hWt)0?h8@78N%=(I?`+6#Rs>N%vc;WAo
z+IGQDT$@Zm(=GxZub>&x60{&%gH}Xa(1vIa+7TT=2O<~b5S>9MA|K=tT|pP35EKyI
zL3aYh=T1}G)~duSE9qWRD4sEK<%+-j)@3Up$%3-i2wZ=IazsT?fv5~B5miAIqB^KX
z)C4t%+MpJZ2{MSfpbk+V)FZM%7SRwiBv5>V<1fE;#ZA1jiOfV+eEH(iyPKr8t+@$H
zR!d|T-~`B`MD{asjL4dhLsa%pnmA#y8Y(-79HFv0V0kT(V_zO6a_q}vM2>xVh{~}q
zuc316%Og~deMK!%U|$g>3hXOlM1g%ph$^tJsG$n%D<V{ZePu0CVqY00O6)6RM2UT6
zh$^wKtf5NmD<f2ieN`<{VP6#`D(tIbM1_4-h^nx!s-Y_Et0Gi|eRVBSV_zL5YV50H
zM2&rQh^n!#uAyq|t0Pp6eN8RVU|$m@8tiLgM1y@zh-$E}si7L|Ya&#GeQhn#VqY62
zTI_3MM2mfGh-$H~t)W`%Ya>*PeWsQ$*k_`I!9Eis4EC82Ww6iGPzL)<gfiII)e;@{
zby1?jzAi>|*w=-q4*R+qs>8l6LUq{J*AhMU^--e7zCK3u*w=@s9{c(ls>i-QLiN~Z
zYYB^eHcD9RvoXSApAAtK`)mzmvCl>*i+w{aF<{>iB?jypV#I)bLx>u%Z>XUL>>DD~
zfPGVhnCf7CGc^*^ugJ~B^a#0{n$A*jO^JUM3sciY<OXWG1#FHGGxp7m#EgA&GcjY|
zTusf`H;1Vi`{o8}#=a#&EZDa+5)1Y%&BTIzOEtA%-x8)4>{}YB1^d<rv0~raNUYem
zHWMrMt<}_ueQTInv2SglR_xm%#D;xaBe7xM)=X^Jw^dUc_HAKm!@jM7+OThr5Igqm
zjl_<9do!_P-(F4a*tdtN9sBkMYRA4KLLAt4G!h5)9nHjneMdEQVBZm@4(vM`r~~_4
zgmBpB8VQGeu9<My=c*}(eJ)Hn>~js2!@e^@oY;3Z5-0YZ&BTd)XEk+V-x;P(>^mE%
z6Z?FG@Yv@Y36Fignef=>t0|9tK1_M+^9_{8zAHjp*mpG&7xrDv#D#rVHFaU%6{ari
zyBeqq`$B{W*cTd!fPJBv2-p{@sepYUOa<%<4OGCsJ3`#pcQ+C@_TA0IjeU1Dbz|Qh
zrf%%J8>l<cH??f|sd7l=)A29+o_-#!_;t^`_E*Wn|1OKrB;SmIGp2}lC;9Lz@z`On
zmt`wo-5ftQcX+N<+&C&4mpS-W<*1IW$rMjl@mTM?DYr`YRGg0A-AmC+E$*k3EZ<dD
z`mdAM;&*3s&gdcT;Z%fH-<tjozpHnAw?ir$lK8qPm&$<@QU#DwssvI=RX}Q~8b~A6
z0BNOKAV$gn>7+U!y;KjxN?9O-)Bt3Xnua760khN$WRY5ctWqnGO=<(OOYJ}osRM|U
zazIY06DYnP59E@%fCQ-k<d(YINj&-S?Wfk=Gx(0r%0?vK0+~@}0pesrg5<0mOu11G
zrks<5sbCdgDvSy+6`TT0C94EeX;gx#<dk5lSQVHmqY6wFrvg*Ws=-tn)nKYQHJBP!
z1E$8P0aL?iz|^u@FttW4m|9K?Cc`pdGDZeWhGW3gu{tnyMje<sP6wu*)q|-w>cP}=
zdN5g*1(P+hV6q$wrhzqpX)qeVG;jtmO{{4|;^S^Ink+z^Nl1{HHG^q3n!z-4W-u+R
z1x$<40;YwtfN5o|U|Nk<Fs+;wOdD$h(`K}RY2$2Q+F3i8cB36kJ7)*e!8*Wn7#(0b
zI0u*<%Yn%mIWRen1JlVm!E_p(U^+P`m^{ma$s2hvd5#Cu#k#<B8C_tyI2V`#D}X5&
z1uz9p0MpI7!E_tlV7fUsn3I@E25~J&_jwsF&}3$EJBdzwqoH+=Uc3lx*)z!0Yw$1k
zX}T|OKprP<$E5rE59rT{dsBxS*OX7ax@%Kh%D3h_6PKleu?2<VM^N18kMDQ1?(&Nd
zknSsF3fuQP-e3M%#|ew}JL1byR+p@7ar@(Rk@0p(-|*S^158hvK0th~>AvEm;&!Sc
z$5y`?UiTn=Y>pyFEk52?m*1S(kX+@B-`$_@KSO-Z;zAgII<3DdB%V$YQ`G)+ZfH9G
zw^P@g4e?_=S)RV)<BBgZer)Sc4vAxN(SDR69rd*`r#^mc9-lWsJaght61k~uKSd;t
zm3d@EVmRF=_sD@19tDunqXbfUR6uHv8c5^O0BJp1AjZQ0={!0hy+;qkdRQQX#{gvV
zn2N-AINfLVn1L)F3y{@g1+sZ;Kz5HE$l-ASaUKrH>2U(_9v;Z$aRCV)0m$ufx07Va
z?1jG8-+-1b@h_d&dZV~?_`kpD;th$Ur2S?G-`9DQ$7%~Sy5yPDi)WOSN=n=3Br_5>
zXC`j?7tLBYH;%$nV>={uOztF2NlWkCt!o!P!)|>dBbB#g=5+7Tv)AwuBS(!kjxnjS
x<OKr_L;C)I_h#q*rBA>91M&wA9x79)HCi1zEKlEC$=I^8IhUV0#+#kQ{lAc{wp0KB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp.11B1653CC880738D.idx b/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp.11B1653CC880738D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68cf27aaa1202628cbb30a2bfd66eb9d5111847f
GIT binary patch
literal 8298
zcmZvhdw5J|;>YKGMNBR(2}wkVASpXDXYO+lbY{-n&Y3x}W+DfRGMNzC(5cX}?I5(<
zMQOF`vP<c1S6iyI+C`R@5>eR(TU%jN&`U<8;<~HdXze7w=a}ywzdVokna_DY=QCID
zIddi;xZULi3WcLWSU6|K{9uwop~#5;tDZk^Hi^T#X$r-{%^yn_=3W@l>&sI=l-kdD
z_nbewbYSMt=H%I#ZwhN8LuSn@%y^4ugrD|IJLH&J?wMWl-Ua3||84(yEuWVON<sO-
zq_>z2&yVtpcdFWhpZA_;IWVK?Ue1QCyUu?%ddxs&^P%XeoaJ5a9Q*!+ExefRbx?Y$
z`az$r_a-!LnD$HEu?@$c*!$ysePzz7O|w_h`bV2J=e|C3?9BZucLp5T?>N!4^yLpH
zrG0h!$a-7asiJ$2m#7br<NL4tHn#5VBkR}w`?>8k?;XxPJ%5s#*>2>TGS}>1a=+xg
z^=l4XeC+7a+1!@evp?3IDcZPHtq;s_-YJ?jaNo&EIhURY9~d|6wd0NK2(h_(=F}74
z6wMf4kU#%-PlQfnKGTr2EOvO|j1qPBjpoS0Ul0DE<qI39dgZI6qT<i`nHPo6f1O=;
zD=*7wFRVCt_Vs=#>$-f}<w)l9Eo-_RjEy>dwr^5k<mBkhzx3Gqnc<zxo4@P1EBn*p
zOOC4pep|ZZ%9bC;J^e_Cd2x7Qb!qRpJI}A%`=7sldBF3L(dKQ_4}A8D^|EGC_1pW$
zQ-96mw%f&Ty*XsO>W1NFVd)!1D?X5>s1J@f{Lf?4kM3GFcJ66Yx=`rsQMY$xus}6D
zyGyh|rTuR4^)V;*fAR6*nr{2tCEmfmdci!U=kw9ryQQyYo>~5{a`%~!T#?{_maUt&
zuAMGaed%8L^rgD%PYCZdHk?~MD6dVtyKCFU(3VQu`UUH2hWsry@~-Xr+F6>)xfOZ0
z9{tnqu}Q<!tFz9e4nI7hVnF%6x`G9Zk8Qki^5`W?-<qO|vNyL?lT~Le3my46pKkx^
z){ft+9$$Xad?WHn#fQ%vc4aVsUU}|?Q_P#sPrp*RYSW}U2kR@vvf(Db?+44Ce@Q!g
zT|56;QI2}&*uuBVwjY_lx!0ywRDEW3Eh}%Sy54JDL&@yxQ~T3cWuIs3zBP}@TkRPU
zFf2K<ai2AFPsy51l|NVA8hiTOowJ6?_6skHxdTTeztU1x61rA;Vtw9;!k+gJTpuuJ
zg?FNwR29sey#0^om3yP9&+VSLe#@b&xhr;fhe`Phq(6Ln@1Nh#5;ruQ=E}Ij4@Z5n
zY$Ut#!TQ@du8m7WDKAwSkE%E1EqdYj?!V`c>@{e)LVNM8mnYU_e$jQqH-(E!V&~@X
zThjPyd+f(&Hfi4faBFI1>p?!hch#b&lZv+f{>utu`p*?#?D&)T-|E+d+XMeLq`IY4
zTHWpGHG?+<iy!$z{!f#(mA`PvUo-o`Q?8?ZEoJA6U0LkSpnYSPZarUjR4yT#dYpN#
zvi0VYXqWC!^||-tUVr|7Jd5jYUp}<yvo~@g&x-v&Iz0Jq<@-M^DBUqMn*Q#x7Vh$d
zrpJ5M{B`JvPpW3MJn$`CJgT;Cpi8`b-}6bz$&bFBR<dD)M^U|KPWgPIP^A9%jppS(
zOKBRFCMkk>$$1%_Lwv31{7c)pxO&>DZ(i@N_;q}GK=y!Por64`v3o_o1-q^tO{4z!
zbVhPUO6MR?=Z>9u{o~Q^|FOGbTztA9yTH;p$kU_${L6!XtUDA+qowibzR7)4ItO`r
z%tgiY{#)O<mPUUQpDs==c6JW(w6*=#^><np?M|Z;;?r5lS=~AZc{)_LceZcuIa?Z?
z9G@;sE-daG<mq+y?l`ycUA|9OOpH&D>pE^s=O9n-6qYV#Zrcu|(ed%=oaCI8&Ox63
zyl3}d*0HbS_gxmB&P~ou=^W(gpL<DrE<Rb;l12mZ>CEKJl+Hn(o;PD=^^M%bYadAM
ze9c!MX@6tql~D!hL6xKm%8x%zA=O-<u<F?asU<a->W~^tbyy9iM$&+(32DI8gf(C?
z5(6d^V!&j=447I;3#K-t1ydW=f~k{qVCq6TFm+)an0iSMraq(xQy<oYX^;$H8bSsz
z4PgVAM#%`KF=PbO7&d}wl1yNlLMAXxVH22U$qc4BWCqh5HiK!AEMQtf7BDSg3z)3L
zg2{$hFxfB*rcJU1<^Lb)L0iZ+7bt9dHbHjD4yHY12h$$5gULx8m|TbhlM8cTIwS{}
zj*tUPN7w<TQ*wgo3^~DchMi!#Bo~;jkPA##*afCra)ap(xxsXY-C%ko515{i2TV`c
z112x=VDcdzOg_wmDM$jCLP!8p2n%3(B`=uXkQYpE*bAml@`33K`M~srePD``2&Nbk
z!4$(Hn1N9N_rpal7~lh?KxKjHGRZ4^kJQeU9TuU8>SY2eO+<~TiD(d+2!p7NXc2W0
z9il#>M>IqXh{lK!(G)Qunj>aJOT>c6M%V<VwXIq&-K#i8(t~|9eRcBDMm}c6SKW>S
z7Cd2AH8SyhRaO-T#H)M>QnP9>)mAl_YF-VdhSh+nv1-87@ER}~mI0HoGGH=11E!YM
zf~mD?!PN3vFm<dBOr2E+rjFNvsb}?I>aBV(^}HTT18V@&U^Reg;0<6JStFQ6s}W2i
zZv@lCn!q$!O<<aM6PRY!45ry?2Gh)&!L+azFfCRKm=@jwCd;y5vQ`#MmS@4Vv9^(k
zmDOgoaX`GymmoW92h(n~gK6jOU~()6CTHcq<aiED2kQXSVReA%;2mH(StpoIs}oEo
z?*!Auy1;Z<U0}L+7np9=4W`@b2Gh;E!St{mFg;cem>%8(CeQL<@>U*9p69_7SOHAI
zDu5~Q0+?Rb3#QlV1=GuW!St~{Fnv}Zm_FVIrpSt5idGR!kr%-XXaYKUWlawT3;{C`
z8yG2*d}bV9bvsUr6N`{7V%tpMa}=>7auE*E5pf_oBThtD#D(aNxDh=O4<aAo5rv3=
z=#6+0eGwm`7!eaFFS9hY^JG5p%mRZ0Me@^^pI}_-IImAg6MIke%aVM8|2j$SJhM+&
zsz#zZ3CAW?Gf`baZX>E2$hB1UcbZtgRSi`259BCST>w@$5;gYK%|wlTbsJG*UtLSp
z*jG1DHTKm}s>Z&ik!Y~5X(k%%Yubng`<hy+!M>(}YOt?~QVsT*M#5m9X(kNznKr^;
zpQ)t`_L&CCV4sOn2K(AZqQ$<pnP{=EZ6jLjYip?%``QMo#lAL5wb<7+5*_w+%|wTN
zT^rG1Usp?Y*w-~s9rkrms>8m%k?66nZzg)|>)VJP`}$g{$G*OS>anknQa$z!jl_U`
zLo+d8-_S-3*f-Qt1NIFK)PQ|Mlp3&aY$QhP8=Hv{`^Gk6#J;hX8nJI|phoN)qtu9f
zQzJ29-_%S@*f+Hi6ZTED)P#Lg12tja6s0EYn;VH4`{rh1#=g0Yn6YoJrDp7#8>kuk
z<|s8|-_l4d*tawj3-&E-#DaZGEwy0Z(m*ZPw?wH0`)ng&vClRW7W-@)VX@EFQWpDc
z17)$#Mk$MZTa?%?z)`@~N^CzMw-eia<a%m5PUS^4{(Z2C+Rh=zsO>7SJxc7@x3>~I
z_U-M&j(vMQwPWAjMD5tO$EY3qT$FIw=UNGeeXgBw*yrjghkdSza@gl$l*7IwN*vgC
zv=RsQ9qq({eMdcYVBgV19oToor~~`XC~;!n*-D()ceWEJ_MP?AiG61ibz<KcqfYF*
zqQr%LS1WO0-_=fB*mu=a7xrCE)P;RljJmMzjuJQa-L1rpeRn%?W8Ym*-Pm_GQ8)J8
zG3v&?CrUin_p}lZ_C4*ygMCju^<dxAL_OH|#Ha`Re3bCm=UWMneZHOW*yrmhkA1$0
z^4RBNl*hghB?9(^Rw7_uXeR>pg?cJrUudEN_JtS~u<wl$FZR8y#EX4zJMm)QTTi{%
z_cl>4_PsIc#l9~}eAxH35+C+`?Zk(DUp@6<-`7Na*!RV#5Bp-2h}aifiHLo%oru^M
z>#2x+v5AV<7h_aR^liVaI=gn*+6D3N<pE_NS3cddzxadVr@Ql`G$|+zktWESlY)3o
zSTg+8C$pKCUWxCUHzLm=U%ehbHgo9JwWGVnH~UlkJ>`9+*%Pl$+fs8RzByNutCO$V
z$|r*Hv(Jv}!in=ryx*GIrB|nH6E^pYUw~DrhROS+2h~b7kVdHiVw4P!R;dNjDRn@4
zr5?zjGyoZuMj(^Y1Y}m4fh<Z35UXT?Y)ad(#38}1v;%QU4#=T&06CRTAeYhw<W{<Y
zJW3A`ujGN^KM(?Wm0lp9(g!3e#fS9bO;`Vsr(Wt8f5S3*Wn{~jYF^t?WXm0T?G8;3
z<`2&I$d|0rgM$VS^2lGM<TpM3ZK30GzWlDf+q!(wq$``(#P==a3O$MUu~;nj%dd}o
zxj+7i=(yG|?;HPs4hR&;;hhT)bR9S6FYfq*%j%JpEq^^ZGXBx;rwwd69>0UKq_V-X
zogSQ!G~prDkv&T{HLbW8-#15-qm%FN*u_odvE(`_zIl)^XsY~F<PU26;dK1A01szq
zKqcS7clxe(yY<Xj72nsN<sT?-zMr8S{TY+<b$oNakUvg-t>m91@=`m0iby0?e%1KI
zi=y_cfi!*%5aVZnw0<p+&aVT~`}IHuzX8bTHv*abCLpul3}o?JfLK2ZWb@m`%WqzK
z(C)VbaefZS;dcNz{Z1g4-v#9MyMa7@4-oI?fdsz*<n?=je10F0=ocT7V&06ZV8?Gj
z)$>Btr5!gO>lps;Z;E_F(Iq(M^o7yOw(vKS%cndxb=q{rw1;zw>4}?D6F18fH$!tO
zs%FK}H?yQmQrF~e%9OP9?!9{U5HehjM>0|cE;FZhpT4;xMvfXi=25Fno24!u!VVkw
q|833A>p!4i(BQ(Np}$gTbb5oyGCbehk1;y4vOV4qb&s|4MEoDoNeaaP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp.1FF1E6FFC39B05F6.idx b/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp.1FF1E6FFC39B05F6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e07a6300a72ed47e908b2099db0c07114f793f20
GIT binary patch
literal 8220
zcmZvhdsq`^qQ>)m0!E-Jh~SOF3an<5$vsiL%_Nh{Oaj&sh_Y%0qeZt~a>Tk?+o;uE
zJ!Q2X?ZsPLPwU#X?W5hbF1W`lrHDmatAep+aV^pwkXmiEr)W2F&NS~Ihv(sY^5%Pg
zzhoxE%;aOL<nd(Ws#L;UX~m*>%PJF8DpgwSw|d#q1;6xBsrIC(R4casJrG$O89VU&
zna&c&Ip2ZS3u|)HM_ox;kiJ{m^mN|*rTJ;^imde0f!RldS#!J#*1jKMEBx0+ENwr1
zezICp?=OFseRchKx#-rypDRxfT57ACckxcvtGhmE{d&U09QBnWm(FCZ>vOCAn?^nq
zwhufU__yj`G7|4hyZGwt`zPyPZFv0P-5z6A)`o2hUZS5&x}s~jdanLlPus2Rx<f+a
z#WjEYs66HJ+2dRJlrsf)eqXGuBMpbzp1i(g&+)BWet2r{+V_tRJ-e)2%kDKhFQ(TV
zTHRCp{??khuO2;dVu5q#-!9xed9L8KHCp4;d9GUp^K(9KF3*ZS9;%x%dUL~P_OV4*
zsvn!x_|JlQrMV9;d+zaIWBMOXB|LNe=!$v8+RW}NPp^1n<o@<Acqg;*azeqlPls7o
zhFY&?=HDDX*yYHdd-%eRVg0uBIo9WR`ug^ozK5@mKYL+F!qmbUm%e+u|G`g9e@g%E
zxd9(!9vc@GI<lWEd9Q8f-6_j|8)Tm!lV4pjXz{+*EeC)6&+~QOznFR74P(x#Htrv~
z^6EW@N~yed#*OdCy}LWFl<78omtXQ`!HfF?Gqs1u9{s$&;=~8fOkRA}k}Bo9`ky@b
zQe`ePCbQ3_Q;h!WaM#4fLtp&;=-R#?dy0J{HP2dS4p@I_=<SkM($B5?a*pTRU))bu
zX1DL!zH3v3wD7#=rRC9+U5`uee|D<n<=+gyQFQx*cfJbltm3ymv2|_UU#}P5=DRk{
z*HtZ^JN)LPH$I%4Fk1Wa;B(1ij!v7KJ?G<-xle@aUu$bV5w#6jTQIk5_dC^O!v)(4
z;o+=fdoSO7?*-=f>zb|IPyg54kN$AfoyNZPQp>Yv*xl<Z+Nw5eE5CL4ld7V!F&5e1
zX*=*`VBf%RTQ?VEY4=Ue-&3~t__FN-w{2uH<|mfTX<yhiaLcLU1zod7(Cbwht4?0C
zP8|NSckEQt>T|Ds%%vYFuGv;~Z{f|!XIpMvFwJm8o-Z1jGd5{sds%Vt+mgnu!yEGl
z^wf1_FM828T}u|`J~m_T>#gd8my)0QaQgBmzkj<b|2^id=_|i1mtTo`_v|*!VuX~k
zew$9;xSQPGGW!0EQ?oNFX-$FgKx68MGZ$YQ{P2-)pS;GePo;;JZJRVF(I;ljtl3;>
zzuYzZqigpI!%H?NF8`sVZQ+ir#^5tu%{6cSbI|@bik_?7RrI^E&C=?{X6gR!<#RuK
z>ePETk2cHK%tiI<zsOrXb?@}Hs?v&6t7m?0^c1@fPL`kjY}NGC_M^vy>t|bMAAR?P
zHQG9^XV;IsFyh3$ZHxZn`Fopk_D{H5HE8JVx0n3!g>{)LHr!9OA6e}D!L`j%c6O$^
zr~a>(gxqiX)#!)6(fUjJ&gf+QO(AX2*q=8odFGbxr>v&KV`@~@D;LcvBPvz$|Gar+
zo!?fHLd&MBDuyKuOY0qCXNumR0q+bpr@Rm?N}<1tP3I-$jq4qh>EX|vd#ls;`m%>q
zkHn@&WR4i!J1El=-n#!w)0QK_6go9Fot2c;uXj+Uxu0)#{i%KBhbc4=o6bne=-WFe
z(~Inl<+q!hb5d2)V$+W#J~FX)P^R}uYr^dJd|e7HiA`rFW%lbGl<Ct09;zH%e>EmA
zjZF_p8q%+K_;q?|;wL-qo<9GOYD#Q6Co{*^J1EoN->l^i_+P)5Ld#>*g-M03-a(nZ
zH!yJEt0gDfQ>Yx9PD@Jb*E=ZFs-^QDt3Ea~zIUf4_wLn~kN^DUzP9nXsTG=lrc&9r
zu@uxSh7{7QiYF$(K$8hF&}2dkG_?UOG_^r3G_@fuG<5+TG<888G<6{zG}!<PO*Y6v
zlMS)Z)Ccs?)Ccv@)Q9xYGz1LLGz1OMG=vP$GzN^&GzN{(G=_}OGzCo1GzCr2G=)si
zGzZMkGzZPlG>6R4v;-{Bv;-~Cw1h0sv<9ruv<9uvw1%wEv;}O?v;}R@w1sTY<N{o!
zaulXka6xV{q!70%p7;O{O+Ls&lMnIGbOapGbOasHbc7txbOxNzbOxQ!bcUSJ6aoS?
zg`faUAtXT46>ve*6?8$<6>>q-9dJX_9dtv}9dbj{6YxOO6ZAmS6Y@aQ8}LHY8}vfc
z8}dR^42aMagCaD=kO)mFAVE_KO3;)-5;T1QA2fYIA2fX-A2j^|KQ#S8KQ#RzKQ!g>
zvgg-xWQ8osC6LPG3MHl7*swKlXKp=7t;o`48I+T<a*CnJy{Gi}xv4C#Y?{I<C;0x!
zy(ji~NfXv=B#I)nf(bJiwP7trU08>a4YL^aVLe7e*nrU(Hexh|O&HB#Ge%3;g3%hb
zVzh;A@u=*E*ycaDFvYh_9w`?no3C7<c97)WOH;h1X(pO8umdzvqWK5rZld`Pb1l`>
z(fAJ2G*QiI%n_=&44G*r4EmWUVbITX69)ZEEoIQpG*JfqOoTG%*ESO^`n6G_MZdP2
zXwk2&rCRiBo2VB3+6dL6U)M}@=+{Mw4*j}rqC>x~mg>;2Yoa>z>mpQ#ezuvg=x3vZ
zML*k3SoE{Cltn+=L|OE+5z3-p-%RxA*GGvS{rYaAN58(7>d~)nqI&e}BUF!mLo+d;
z-w-7R^c%W~0sV$rYCylCi5k#vh)@Iijm^Y}eq)pv(QoV~M)VtNsS*9gCTc{#F+z>#
zH#HLz`b|+{Lcgh-n9y&kr6%;7ny3l=rU*5m-`q^h=r>1+8U5yNVn)BYmYUITZlY%N
zn<LbWeoHg4px+WD7W7-Xi3R<ZT53VRrHNY5Z;4O~`mN2xihgU9SkZ6oCRX%YYpE6e
z)+TC2zcoUw=(jZ!8~SZgVne^Jo7m8AtED#d+nT5i{k90Tp`VKo?lOtrWN@8?>%!ba
zxO<ozDEBFaehcMVF?Uez8_0Zw@aX3|36Flhhw$j<8z_%{zJ>DW=Q}8men*5j(C_FZ
z4)i;Ehy(qO2I@e+qlG%q@93Zo^gAQOiGF7%aiZVZL!9V$Hc%(}oh{UferE@DqF;y*
z0sTTJ5zsI65CQ!{0~OFOv`_*4LI)Ml?}`u?`dyvGg??8LaiQPUKwao}wNMxOT^-bg
zes_ep(eLggZuGl*h#URx2I@w?yM?;Z@9v;(^m`)2gMLpZ@u1(+Lp<pBG*A!vJuTFO
zeoqJWpx+xIUi5oAi5LCe9^ys6w}E=m?`@%8^m{v~7yV*{i0BtPiHLr&hluDG8>onW
zv4x807dxnkeknpE^h=#YLci2QB=k!SR6@VhLM8M|9aKWUFG76i_jM8<`h7jbhkjoJ
z^`YO_LVf7>bx<Gr{So3vzrU0C(eLjee)Rhrs2~0Q7V1a8zk~YYe*XT33!6r7dLnj9
zEUV?A%I%)x`A*eOx0hX_2^Dw~oKP`s&@@>o6_0skNhbU68)NLSvBLyq0i|?u`lyag
z6B1)jPw6*hfYJ^unBFma=i1}3r-$f<7?f2BWjc0W`Cwr~xvyONw>2Fz-;uTti>(Q$
zHKXIVi;S9qq*ZGn>C`$%teS<SSL-1e)CNdKwGom@ZGvQ0n;}`$7D!gL6_QPDgT$%1
z(eaA_ujU~+)DB2awG)z{79hFQE=X>*8<I!uf#g+tA&F`clBAX(`P4p0ezpJCw0UO%
zKXlZ$w`gib_Q-6n@|7pGB6no2_w@g*6aD$W*R3pX+g=kZ<vH`b@lU3(s8Ck6SXq*a
zeW!S^9;JM$v4bpI&Q(70eJkn`r!0Eg6FVX@`e$S+UonoCPWonc&c%k<3zR05j#S#I
z6@i4nujwQ?ux8uE7w^Q_OkJiydA<6tw#~VobTSZoI#<e_rMze5V2iz-2kT$(c1FpJ
z@&ZaZ_6_gBDp<V4g}EJYf#<_KMn~9z(HVAP6v6^VSJ;Ko9d={%ggqF&VJ}88EMk<x
z5=LLxhtVJQ$K%(=#M^xbWNnD;l_`U#<R}}gd|0u)@?c#rzQOoM-jk-D@F|;hHO6L3
z*;ABbLRt4)nB2SG7iTqgO=0|)&~O?jBvIpsBx_3JiLo=lF&qONBQn4-G6S5}t_4oZ
zX@S#<THv&@7C4<<2b_-60jCpn!0BWiaIBpLj^$Y3Sdj&em095Qc0F)<P7j=3)B~rN
z^}reI2H*^w0XTzb0L~yAfHT^Sz!^Csa7NJxoKZFcXR@1sGjS&1Ori-mlWYRcY&QdE
z=FGsEMKf?_*$kY;ZUN51S%9;M7T_$h1vsnS3Y?X*0%sMiz*%K0a5lRQI2&gJ&L-M`
zv&lB#I6GIU{K3RlAvw+oN#y*HWUe%xcsmaq&+)+VA`cud^T0Xm4&WS|12~810L~#h
zfOFcNz&SZ5a8A()oKtoJC)fqx1Wo`>5Cz}_Spd#ucLC?(T)??R7jQ1w1)SUN2F}g7
zfpd#);M}qsIFH=}oQLxO=Mg=?d1Mc8Ub`1KFXsi$D|&(R%3k0^y9k`fiNJ}X2%IR3
zz)5xqIEj;hlSBzPNtS@~*?qwII3I96(FdGQ_5tU&`+@Uwe&GC~A2`442Ts<>2IYbf
zyF8m@D<r#IsHE6`au!xT__t2=(qMJTg9o!7481>nDz5g=_kRB4pBE)n&X_gx(b=jQ
z%9Kj=X#C;q_`{j;hrQPSjlo587tW78^gmYICm}JZuex7K>O%tu^q11y!f(@(C1-lp
zpo}3y#}<yCFmV#c>j!Jc<=IE){D0484j+-7`<s#Z1*0^K&R{fIY-1j_4r9%(!I@s)
M9O{|u5J}O$0FG+#od5s;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp.27AC774D5173DB46.idx b/.cache/clangd/index/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp.27AC774D5173DB46.idx
new file mode 100755
index 0000000000000000000000000000000000000000..66a34ea03367382671aebf4fc8ee5b0bb88fa8e4
GIT binary patch
literal 8178
zcmZvhd0Z1`!pHMG3Jr%B3X0SNYt1=1W;{!h$u*OV7!}2cf{hxrx?ZWbve}Q;_Vu-{
z_qnU}+Sav-+G_2qNL%Z%g0b|pN<s0~v_-UawYH)g*w2{XA1|MekMqs*oiAY~%nUO_
zBvJgai$dWVC(WBWe%6@g3WcIg?BC2;GbTN0rck_}qEO7+@`=BC+V$S8&wu$~lH=R{
z-}Ib(yw%dSa~DrZ3$7J@7r(mU?KWVwR(kx+f%?H4FXqiz^nJBqbncB8XWTk>exOoP
zemG(mv+>QGyaB&X`Q!C-ZD&}I1^cwxIDFgmi{~xYE=t4YpKmYwe9_};55{%v%V#Ms
z?>Q2lx8jW@KW*N7;Ir%3e)_gW;h*PkITSfv*Ha+DZPVIMx&Qc=`A>hlxpma#f7~cM
z{?DqCglUt;<ZgfK_L}0uyVhS?7_13>^6S~5V<Kl4eEQvS|CQ8<r6ub(<Y(@foaCwD
zpYHPKp1Cybz}5RbvyDZQ|9mZ)HsE-Ahxgf?kJCO{S$Ml-S$(e~qtxpvu20B%u}A&O
zhxXsNzlPPeZuQE^!yk@Ynlqk$K67;0!CmRgT6M9#y2@K}wCAc-$=&-nmbN^Y=<8hc
z@sR}SiI|YrIw|wO?0<^RoA!GPKWM(==yg*6xaCK^5*OuHZs=XMXE8G)bx-rMtBdyv
zA5=bSy<o}Zb$>ddvaNfkq9Ag$Q$@{pZ)Q!N^WHw|q<`OSvoCeknw)cIUj40o-`wwO
zG=}x|{3{!dm92>0QLaB*oNB3PH&Gf}l71oC*fKQBeDQ<5+t+42NOh$pt#0<=zOfzt
z_w4f13qL;icJ9$Y@gR2pEMw@(=lk;3uJQKIZgHX6jyZ`9>)t)}pzGmHU-z7v7bsZr
zSatig?geRuhga{sH23vC{6i)mV&??Q|0k#ICzB^V+H^Z})%bVDU-8z>Z)qJ|+`VJu
zZ&_7h*~(j++mf^WgQq__(c)hEh$R;pv-h2VdBTxj@4ePS9Gx`c{KRO{@tw0<TwZlX
z2x;e*X?MQ;!*=eA;%&yNpF4%}+_}5E&AFNzQD+zYR&{O8_m=7RGKZf3V!(H4#mn`F
z({*nycsQ=&bn@P|6_bYTF?CX%>hW%J!br#ZP5laU-0#g8I<mO^lH-pvPj@!=^L;5^
zZ?hu*Ty4)=b^rI~l(pAyjo=H*+V*@{Hv63atB2V`|MkKO^6va$-){On|KLrLt!Vb_
z!ow9GyI(F@ujt+S(7x}^O=(^@HR03YyI(3Q({^0Cr@wkb=h^;sH!ge~oj73Q*{__#
zoZSx<fBHthId4DP`f#|gY0*D2|2l;|nKL|V{_@)gjv4y3&VEZ_xU_3f!Q!-U5;k7H
zad+j2U+$%SHn)Nn|9&*|apkfbn$IV^b|!RxX@RfD$=!o`?Au@QF5l;G-JcD%4w+yZ
ze}3M?JAD^q?i%ZiuKw)({E~Z|)^wBKKe5AmySf)1*nQ%zRi8NX%GDin@6-(%{q2~)
z*H7v6bWrj3?8>W8zvzAT<kRw%1)ZN9UYfe#>qoDha<nh=JWJ`#zwZ9`fX&jduZomU
zzBv4Yt7|l=(vbD{=(CR70t5HobG2#Pw_$bYjb_Z_tkVZ_xS8{(j_pkp3T2n+y5-&4
zY)e)6hbscU1fQ*Gh#e)GzLJ06etB_juc0Y4KQ`SVu|u1tL7ukkc={}|{<B~T&5KQ^
zC#EMg4e~Vq$Njq9x8@&6p~GU+X^CksG!63f*vaQ6jyn4Fi57~%vFT13oq9D5^7M{Z
zULu>XyzzPp9TA((PR!0}8szB`Y2ja(U+l+H=#bcSyTo=$O@lmrJZaqBOEXOqTPOy_
zraNYIR5uOs^to0o#-yLV7JF+$W78Rl8A(lpJpJ&~zW?gY>071Hg4lGQ#6CjPAWzR2
zKVfEBrb5wdMsZHRlehSesR5Nwl`p^J)PUNj2GaO6Kw6&`Naxc5>3w=2#>W5|d<Gz+
z&j@7lnSjhbGZ5=zfjA!rWc69|<)1Y*VDs64cpndB_t}9QJ_nG~=LB;3TtIH08%Xd8
zK%!3s^7uSJUY{38@=4E0{>YN#rjP3K6Mt+kxth~e{-_Q~Zu*GIFUR(bUo|GaXH-Ge
zG@y{GI8JK68ccOi4W>Gz22<nLfT;;;z|@2^U~2tZFttG~nA(sQOr2i`rY@)hQy0>K
zsrT!_)Ccuo>O*=k89xIi6J)?-LJXJ&zX41`&;X_(WB}9XH-c#l8o@M%j9{AlCNNDw
z6PTuu2~4x!45m3~2GblegUR|?FxemrCL3bG<oq0%T#y5k3vpmt{njz@-vVpUIt?gf
zEsm4TZv)d7w1H_0*}&xeJeYit2a^x+VA}n5FzrD*nD&qzOo!hArX%P8(-Crj>GV6n
zbOxPZIzvt{U49psuAmD{SI7mX+wTU`9dv`~4!OY;`~sLlPyka131EtT5lk^Cf+>bX
zFg<<`n4X{qOi#!Irq}NU(;M`H=?!_ol>8EyQcwa@3Q1rNV~1Jg<4$V8&-#Iev%{ZL
z%H3rf{Ex<+CaHmReY#OTKg$2p<fiju{Ma;@9o$Rq%4f@WVh6Uyb7kC8RT9;gBz~My
z)e_Ye<a(mIhg?ck$0+y_syc^UO;t6(>Pn)<zPgsEv9GQtYV4~^sT%w02vuWWT}{>4
z*HjV>_BFLcgMCds(O_RwN;TNmM5qS)nrf=SzP6HRv9GNqTI_4<i5C0XQmVziHbS-7
z*H%+4_H~s+hkacw(P3X#PjuMVl~Nt{brGt=zOI_;u&=KqdhF|Ki5~m<dZNd^zLe^*
zua8hY_Vv|NkA0?+FxY2m34?v6o-o*FN-2YVCPEqPGu4#AzM+yBuy3d(2J9Q^i2?hD
zQfk1yAwmt<H&jyt_KlUqh<#%%F=F3XPmI_%mQo}3jS*_ZzOkAbv2Us*ChVJPi3$6r
zdSb%9sg#<qZ;DV8_D$8)gne@*F=O9cOU&3e*Ap}L&85_geRG7Gv2U)XX6&<-gvCBv
zOIYl)^@PPfTS{5%vk}T-pRJ}W_PI*JVV|od9QL_-!eO5)r5yIT2<5QPRZ|Z8)*52H
zMB-;zYaOxPMQ$M0r^w~hT0z0DqShafqtyB%u&svJuy3m)HtgFPhz<L;a%#iAt%};P
zZ;MhJ_W2sZW1p`hJofnp!egH=r#$xgD#~M@k5V4{_8MZxzP*mvv2Sl6cI?~BsU7?F
zDr(2RJxcA^chnFE_8oP^fqh2<abVw3P94~HR8a@^9Z~AQzO#lnvG1%SPV74yh!gwH
za_Ypsvx+*g?~GC>_FXl^g?(2Yabe%pKwQ{&l~Wh?T~*YDeOHvau<x!RZtT13h#ULv
z2I9uPyPUeQ@2;Y5?7O4XjeVhp2-p|uh=6^efe6?a%Bg^Tp^6IF7ot?azF0#<?2C0o
z#J<=-MC^;@RK&hmMMdn3Q7U5JQ$sx1_tX&&_B{>6gMCjq^<dvqMLpQ}M5zb+-WuY?
zzPFBevF~jlUhI3zsTcd+D(c0)H%h(OmuiTFeW{K}*q0iJgng-;O4ygGsDynfN~L(;
z{&d;*tJSM#$F2znb{%MzFZUcvepWnwFzW}}JTUU5kp=S0%>zYki-ySL>$7Hh2J_cd
zv2{E2?ckFC#?2p?rn<A*lF*n;8rVv<{gVpr1hy_Z5qmj9pJ9~mame>sS|&H$Ux{B=
zTEo^Ya9Fj4ZHRoBN3@6Sh>oxW(HV9ky236*ci4?6gat$~EFyZs9z<{0iztPqIF9!@
zK3VkPuCaNszw$X%Ke*zxsGLB8DmPAQP7S8oq6SkfsKL~58Zb2$4VW511E!YKf~mD=
z!PE*`Fm;>`Or1psrcTg-sps@y>MeRO^@1KuhGW2FEDV^8z<_Ds3}6~81~3hR0Zb!j
z1k-3Sf@u_tV464+m?nz}Op{;&)6AK{G+WGInguhMEXRV$T39eyfd!M}I50U22PP+Q
zU|KnAKe%SFSe-zEH8)N+&IYE<Vgu7A*udmD9!%cCgUJg#n0C$%rrlx((=OP-bZ`zZ
z9To?e4#5GYlXHUUv^c?Z3QjOxoC{2s#RaBIaDnOO++eybZZO?~8%%){z!WS3n1UdH
zDRLs1qD2H#6htsRoCi#g#RH~C@PO&%ykL4QUNF6a7fgwhz?3W!n35oYIfNb3S3Z`e
z28Oajf%4h>=d|eP!aGIpNn09k`zlpxIJ79$KpLe6NUPKW>6AJky;2XvC>bDw(g0*s
z8i7no6OdVH24a;g5U1pTtV*jIKJH2z5U=Ec>`FV3L+Jo=DxE+sr3=WdbOQ-W0Z3Ge
zKpv$B$gA`MNlNKCt=;;vJyZ4LuQWA~)h){--``ISbne#KBcI~rT@ky{()hq2eyYP$
z?B1Ui&mVDh%ZgYj+m-E=*GUcZ@%0%ZZx;FSLafyIY$0A6#149X%cf-Bv4p`>cZ#tS
zZd%K<4EYX@{N7?G+{OnF@pV;U)k?TeqYkSPHDL{+HmpU|g>{JfupW^KGl+(;0nr#X
zBAUV`M041T$c9-&F3iR8M7}K==yKsqY}1WsKB8Ovm!+_I;d44g4ldk$Vd<mTR%xei
zXOy?(=}Vi(-cCH_kG<Sk>O4x`W%5-@Y(qCbhKX;iJT~w7hQ4V^c<@52_RC`H4on}|
zMgG3>&Wf$u_>3mLE?z@DX`{51GwrX%*3FW#2Fp8V#`q~?8lUIPoDrNkv~gi{<M7|F
zB6&ff7#Uw26<>TczF7Etv032a)-yKuQE__W=#fQ*qh3`6r;eL4F*ct&!Pl&LLgEX`
zq?FVaty{H}+6b;5ZIUHtTKl%`I%M|kmt(Qoc)KB8_sUDX)m{F-S2MC+?9?^8Tlbe$
e8m-=FGIPE9bne-aVcqE&p8jK~IMCrG1O5YtlA&+_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp.5A99B55004C02041.idx b/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp.5A99B55004C02041.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c6b6e83788c81dc0450836fc55c958dcc29bcffb
GIT binary patch
literal 16046
zcmcJWcUToi*T9*XT}4=~SGp{41uTdl7hJhCgRcT=41#DR7_r2zF^M&5G<Kt+F;*~u
zM5HQ+AZQ{YDn$h>#NJRq5os2%7rvp4d(0p2_vDY)=W*x$IA`ikIdf(<aEPyO{AXHP
z11AkxIBW8}37nReRww#@?z}nE!zg@apry4iGJ90;_qSY?mHs6IUrsI?_1PSqJNWR3
zjJ0BquiE=<i0LwIeVX>rnErvY|Bh)|Iw)aCQQ*dq)D`CCOXkj=@@h-f`-pM<24&|-
z)>fI_ON@r!SM1F3xEe9;z1|%ET}aR6qF;<qHHqkjTw6Zud4P(T{XIBkSFFa#$lJ+s
z^606)*L^eR+N|--)SX(qw})ldq)Fblul%m<cy5N}h_Efo-xtqI8M(jn`7t5x^@X#)
zx?eIiIeGYrJ69s#7azG=V|MSkgVC-5iWkBUQ&uh|kqi9hjgQG5vp6vH#Xa6_-ssRT
z|1&k-=nTs4;h%4O-mhrZs{u2=-FYvn%a$dv<>yVdFRbXYWy1zXX|&zA^64hyDxB{W
zWq<Jtjvcz##?$T@s&%<>ZjElPPu%70=d7<a<^%@Ko}GVcOxE24{eB$MpR4uK6wPYq
z;1U!#>&Eu1$I-63=T{AA3hnXB?w3!4JIsm7y_sZM_vN+V2d}?%ytFvvLPYI?E6-1v
zzYVHheB(g&{sXJN3-y~D(mwmxfxUG{&(FDD`^438{Z5^QE^}4FP5bKg?>;OqeSGiy
zubH7PeUl2pe;;Wx)bD6Sk8{e@Y0|-or`<WH1<6s7(>C;Sk@64SDc2+4o&9Nk>WSg=
zqx<eDnD_Nf_a~0mXL;GYf7f?(c#jKT9NJYhw&a%Cb(gxRhshKBs>2R%dha&VYwy|_
znNKfYRn`ZmT>CT0FDqsA{Lh^GW%e1q|GR#*?IrU~UL7h!hCllJtL^41*G8o3JDxVy
z%Xc2q0nNHR=xE?g2kF2;JFTxS1bnW&aAQN$rSc|WZ?89bV}3vIrgU1=Y~I|_bA7ot
z`mn3)r*;eae>HU6iZabRt)b%+@WWTTE2fuKcK3~*W;*PrLtV!NAl)7D&Os}EhHkIi
zwl#R)%I+c3qdh}Ze(k2sS=l{d8a{Ry?_D3+V|MYHn6a4^>)+|zf3Rth>7g!r4}NVu
z{>1oa0Y-JdCnt^yI$55Po16ag!jZ-uTz=cx^^&7aRp{5}t-o}&KlJzUp$pq>ia!0D
zUWbc!hpg5fJ78LET(@!aXRizDSNm*mC*$EON=~ib>u~+bJ!OE!${T^w)`Q~MvD~i}
zle_&G8FVjKdNplAp{2RQ_>#-g*u`BfgSE2r$2v{X7r*JSO|Uqnef4tMtXmJQGM5hV
zw7#*;qM@(tmUW!t)*-jn%zm)d9&fN3;7uZYzWZ(A<pn|Gdm62s={deS|GZ|+#FG_2
z{%o$gxl~cHyz7=4m2Ll$&Z(}`1NP7AWHLxtv()UwZ{{xF^j+}6dDw>TH|i_VT9bSO
z-JjGye`U5g*(Nq{qo2#3(D}RbJDo~bD;lpaP2KQobeZ7$!OXV)!41EhQG*62p1{ve
z89ZNBFtn%p+ij&27j`fSK03j2=)21%tB6l;e+Sz&>WQCuc01fI;&h?$@$7kbj_6oA
zZryhz`?!N^&;slHf@`Yxo!(tv?(_#)xNUXQ$u+kZo=H^ay#67;)1*^TxaN<|I;ZpY
zbJi8(LIMx1j$iuA!>dzE)9TB9e3Sd8=$ql)zv;5_?$eGn{xkh<yqtde&DAu^z#%gZ
zUkce68ew|e_)5_1oS0==Vx(fE*4#z2zR7B*rPV(9iOucw8K)H3(@slzO{`uM=e9sg
zo6ckOU;8Ulvdcd|gxlecwpx;Fdsn}fhgJuS=){L4r2l@v&`#J~OBWmT)RJCeotN0S
zEnsOFr%uy1n!1085_j#TCEel_x42Gi0aLMvzUMan%bS1eV--y$l2=CZ=4}B}vHY34
zr1^8%Z3T9xg>58Ojl{JrU@EGMiq2>Lid?Oa&1ov-oLM>7w=G~Q-ItjxuZbKtLKk<S
zDh*ht0gG(`Q|Vos7Qf2R+ggc5suIR`2%}|b4Va40%aLmLO}}{PVkK4a*74SFc`&iB
z-JC*AifgDowxZ&{c;#Qb*cLDq|8ahq5t}z%Ho~2#%2>0p-CG_^92`Dr{-tZ5X&GS)
zDh@FVv2J-V@#t^9-8|K<VYe=BN5#*u&NEtz*5IScwx8mvZjWwIVr!~$o42~n_h}25
z%9yD4eOG6U>e2dMR&a_H^u4qOOeMsoc#i#duEj>!h^7)?7GU1;VB&Fj&)rU?A9^aW
zJ{6bo$}--fEnq6s4;{&#J0`hEA6rtD3EYG}Ee|H1;rD7i$!b6fEYP?5lqjDPleU1V
z%-nIT^5K%+rbgI?rr>AcXWQ~%;yJ_j8|Cjei`T{NY42(xiY8**7BH0`4+%4u9gh2;
ziwRBPG1hrZ`)F&xRDNFMKX{(mJQrQ8OH~@NP9rvH3z*8X5oJ5Szn$977`LZIjV9XB
zMA;TFl@%TPXAIjCf8Pjqrm2iH8)@G1VB)n)p8j%OAL%J@Kbk@Vv1%Ziwt%UG+P<jJ
zFI}mxz#dfP6|s6noZ14WvU$~SJMGCY7mct9%_Pt)(7ffr#9QKcL;css4fI>OZxyGk
z;%ubWfT?VC;OC`ZZ|bOvIhw*youP&;4<-(`EZ>DHOG_2F2Q5P#v8f~d+5)DsE4bHl
z=U)2Vb+HajVX)3%gO&#q|E}&Au-quQ+6bFc@d&dK<}Htp;@{5iS+gtmoB}Io29JsI
zF@5Q+0aJ;#pSB?6*<-zSxIIlFis(iWgSLRF#B@(N`17z{Z`)xrs`8%bz9;%^0aJ+&
znOO4E4}aS$@Bo@oH8H3r=4}B}N!04qadW=YQw4UQDldrD3(~hOU@9qrP9riM*B(;f
zZd9d)^r#^|Z2?os8dWfhufDlQfxA<cTGFGIs<Z}7<xI`O?iJ34-4*yhROJ!r{D@e$
z1x)42<)D3Eb@ckq2piHG`J4Hhw>+5m#&0eOeKvgQufTdVg(pP$gf?4iz*HW&8_sTO
zT-95Fy{O7_V)dL(lC1$#sW}(y*xfq6lLFgQm6ycoCGCW*0aK~dd@JZi@4T&xd0Gx%
z9bbc%2NS>CrYJePYRnj2tWCv(bOsr;Jec@Z$KpBRRUhUnaDSRX73o|>E7}?`mG?!D
z>P@@9>0*S9X$r&5hMTuMK8i0z-?qDZ&_Rj2Q1LOIJI0%|1x!V2&g3a`eVw(m@D}Wp
zTy}GSwjjkJF%CJh2}h9Pkr)rT3K>=*TUlO>467k;#o|^B;%~>|cF1{H%!6Ej#R8eh
z-WZdHMK1<$f^-x)A4PrTjIxk(7UUDe`2^|9<`Y3WNt{p0+%Btc7f+u9*&PLGEArTi
zM#%|9AY}yVC36gtVvxVg`;qg0<R|kj)aMo&Ci8tX@ID$V^HZdJ3e#JQJ=S87-)1~;
zGY0w*SQ&vq{;Aj_73%+qo&Urj|C8AFB%UDWSBwV~;}Dt4uydKL&&&}lNPiN=pG3u4
zNT<yxVUq$gO0Z6gF*2m<g3v8T2mUNb-bUVEv$a`Iz3%rJJDhH{7bJg`e>XO^1Svoj
z(2cEYf;3z;yc_F|f)uC<?8fF-K^mbN(Tz=If;3V!vKyP=1u3Lwh$CC2m2XqU&7O99
zks!HyxesK;W2RAioi>Ar2?QyEGmPN6umxC<VmZTD$Vr?*5~q^o$(%Zw^OQN4Q|Ch7
z#49)P=4|R`1{cz4Gqc!KX?iV9`(45jXTnMG+=zH?BukN#N;#8K&XmREq+PttF1{<X
zg_wOtQ*1TPSXG!+Mtl7sqBumlv3Dj&MWj~|86a~p=~XPVPMht<-l<Bl!fm~FaT#sH
zXigQ)*|VZCqmE|XYSFQ*t&Yx_R=eDDE#hblvuMtO#jy6u`0JVLMV)_LOPhNmqK^2*
zcA;}gBT_Y@K5|cej#SSfzd))NkY6IzOUO}J6NQJ$@pog*Zpat0cu}4^1nF-q{w*`J
zQaPPA+mzK=po@xOE$PPA8$mkAxgX?)vZ@QxY0mw$%np;@d0613r}Q`}NE777lHsJ$
zn$dt7ICTT(%(@RJHE=-<T(HcIoI@k$2zhuTH$vv;T+ioRFUUilb3<i*!P&jw>>&?*
z!THJjf%E*peI|1g=h*~#4c~7K@5SV-Q_>#MYL>HVIaTgHAdb)fp7w}fi(pIkXXnB9
zH?-><?9z^tMj4MXVSlFd`KX~X1rHAnl<5Q4`2%OhvgD*Du5%OQHGJna{71ReFHh(C
zqS1!5_pauYt2tx#LR<%g4LJ11s^dS3<kn|xU}^6D=O%)*9f{kKi`-M9kR}QR${dTt
zSSU|GVglrYNIVGn4ifJ`zK6tnkRKw$hj3M_MTWJI8<462xy$))z~TlB%C`lJTOjYm
z;!enWux1Yi<xj_AI+P#6;vvXKv3L}6HWsrX=VCDz@)<0ifqWi|=OJIh;w8xUv3Or*
zcgv@RGuusgN(V$GFI4ghwuRxOhrI9*auwgXiWg=1Bfj$^$klx3YREObSq*P4>(}!7
zwY;Isb-Y-|TgY6`i}jEn^XkXEvn+qY_j$timH8>JehTFcywD(**Hh6iZ=7$}MZ6%b
zL&7@vMmHc~1LTcJ*a&$WQf`BDR|K+;kgbaTA40EgvE4GCc5OF3w~vF0EwdeTTDHDy
z;0V%X(&I95l&AeB(z%IPv0a)VrE=<2&XtXi<_S!F5ONx)PJ>*)sS6;V=G3PlpW)PJ
zWM*4$IybcLyIH|}kFBZywsuN1twt!{DU>&1LwnGrNRNooK6Cw=8&19Mldg}{8z65)
z>Wz@MA!!@>OdcxRk#ajSlX(Yn-hn)2-ieeup*##J!yreXfCw~N){jK?k#OinqG6Hf
zD_Oo9+3$w(Xk;G^ITk5nVft}M9S6%FhrHvUc^Hq>@lYO*yyKxf5h)X4dPzu`1UUsM
zQy`}zbt+6R4Jp&0JRK?1A!i_i3}hkK?+#Mlf$}n>EQ4H*l;w~sk+KqU73x_9JN6^g
z^AY4~)Uz6L4eD6~xeh7oV0!gPSr54Z88#qGxxBBC%PZ7h=GVyZHIz3Y!zRe<uyP#+
z{be0iuft&KU60l4A%|jRDAeDGl^Y>%!OAU=w_^2H7=Ig9Z-cxYtG7emfz>-8hhb$H
zOfMWO!y!jv`$&0e6{JW!EE0qM8H4R(pga-VCqhoc$~2f>8dj&l{4=mJ1Ijb8G86Jq
ztUL<&I948qoP(7)kaMvz7jhm}=fV8)u{s}e0ah14K8@9<A)m#{voO6vtSp3l9xKm7
zzKoTZA>YKxn~*E8vf@)#SHS!pVC4fSuf)nq$W_>&3V$3we*PMhGd|(N>5TUd+q}cw
z*?xri9_bv}?094|<cJPF*|R1qf2I0me1|f=BU^e#4%6>ZwZ)U~E=ZHyCV9#Id$QYP
zKvUeN0GjGH70@)dX@I7?O$Ri?Z3dt?VjM@z<>e!u7{^0SAjS!h6Nzym<RoI81bH7Z
z-UoR<G2RckhUnE01Lopswy*VSiGgfi*AcxsV!&Kt0(IBvH4p>l$)g(uqSr_a*fvO@
z&KAAr#DMu>1nNrCdqE7Cmr0<W7QL6mfH`3V>f6(MMGTnRPoO(pz1PHmZG{CXhPRI4
zZQ1@ykoNM{dm+d2*0GS|c<VUG@w|0B<OJS20dgX5oe239Z*hvZdEe%N5u{w+A{TNV
zZ;=N%pSQ?|T)<ltKt9b|oQ8acw>Ts7s0llQllR<mr+w}PZ}5UQW_^xrt?3rO+564<
zs348)Io64FVCDv=+oV?Kx2#q}3~NX?`L(8U?y2&&NRSFR_X3$mw>it{h-mecvxG+9
z5BHs9y2OUMB$7GPWX@5xTXH$mT$#ssMi|8KRp0%?&S&a0;a0~w+ley=grQE0L!IR(
zNWotO50fbldBmZ9@`#H={&8ry%<;%09?Ij9e>{}uA@4loE9(~^?*hn$$fpnulI6w7
zrx@}D<b46^UqC?@P_V3Dg1k$hyaWZ6K=~sS{0Mz5>(`**8pzvl$TmDqmhZqJJ0Le;
zp9VaLZHEP^oD3`{K5TXf`MtJokn(wqjv$>tT~8p5oX|<s^(5pxB<7*svb+F^1&}Ww
z%L~X!mX{#Q63Ew3&uh?>yN-HZhkTAWog?0?V1jg>IGvX{<cQn1W4d>>p<a?GVi84p
zGOvd~9W<u5xUS6aBG4VJO&IUUwzz_n#T#bvUD$YJ!-Mvx*3%Z&RVMs$G5$M`CrfEZ
zJ&zpEBWJlP7m?#dnc3foPRltAi!adg5bMWyI6Eq7w&R^Hlku!Z?3{+CpGFjE#E1!W
z+D>a&jo4`oEn)Lf4NITd|1{gyBN{dk1UfKGS3uuaDc7r%>%&@v6^M@U))O0+qCm&I
z%QHNHO#*ak#x4zbfZSOcu}dQ!z)p5pNE2#3-jNfkBZhUvN}l7>IQKMe5F5LKbcS<3
zBQvWxtzqjS5UV<K4pKq$(GN?XtrfKC(upFSbY%%p=T6HJ5R1<i0jhX}D2@<IS+VsX
z=%eE2WOd!<CfA{Y<Y(k(&YE11Cb&)jOX)YR-vFBEIuXz$?@7QrHQ9SIpef!{08RCt
z3TT@5G(gk6rvsYdJp<58X{HD3Y=ZQy^evzt6h9crve}B+faWUZ0-C3o2WWv}0iYih
zKLT2$SOjRXVlkj4iY0)SDwYCTrdS4OIZ`Y~fL9{LO315_Vin{VViW_-iM_;VFXUKa
z6bm_ySj7=r_G0N71+j{UoItD+ASV*5M94|RDhcvFVzm$QeqyyB@=>CHlysK!%Od(&
zkh6(?Hss?(|2X6VqAVaLa{M#I@C*@UE+S?{#7gF3Vpa_KG0}cZI{w~v;3v@ID(xqv
zBRk0w=;WdOlyqdXOS5@p+CWrn$TXW*rj0}e%`4OAL<P+&(-%Yq%`4NFL<P+&(^o{r
zMpm<VW%`<^Wb-PTw~B`5&mX+iACUL(R(l}F@D4G&1kEvry^v#hhgisQyh9x1c-|o%
zasuy=06CF&NQ9ijo9D>p5?$wb^AnIy^5!QYpW<y#@s4b{Xg1$$b0O#Pwt0~AdE0!*
z1-xwm<kP(EX~<`I+cPqUR(oHaHf7Gc|Dk8~*HC@u-Fv@3tj$TE8Z(^qsWHPzpBgiq
z^r<n!NuL@s%?ynhPWsfCX=Z55uwx-QyEmV?v)P@wk?5aW580VutElz*&;CdJ=xqMq
zO-FVbPfgwCqj(md_EoI?8tb!N+qT`^`giV8-9XdZjEpuzzseKN=n3b}O0xZ2r**$L
zJls!h>S|6|&2?dd9c6YW`X7(2rh*FIw1V%-tce{j$A_%!Xnuv7Co4FI6`Y&A^sV5A
zuHb@XzRj85hUUs`&gC{YfGyTL&kjzwwDxCvTE=s{;W=I<cf?ZOu#{IZLz}I&RI!2>
zR*>Fo2GMD|Xuy^_cF};>X}Mq^H_ltE{T6p(jYF$Sw5v%ccHl2a2RQWsSdn~Aoi8&x
z)uV5%`E~+Jmd#VNSgAyj3ay;h>j|t_ETl(!t#=bx3iR}<<zfg^5U9Ch(?ohP`$wR?
z#I8y{YZ2%{n`Rg9B^!ME_>TK{6M2R?$9Fl$+p>maW)rnyT5mka8UMxi_>1q&9H4?!
z%6BQ{d$1O7=ANaznsMyFTNVROy@~jLUlnUos(64X4#<mvPTNfiR#i3$Q^ic8$RyVC
zJ8ivI!4#8%{{Bqo$V>;?-F6^x2Q;%IktPxzV824DSMcOK5^Eym1LNd{S$x2(7~lDV
zv=fOtk&SFq>_m1uWuupx)kqA3@-So<2IaevW;X)%(Qf3s8@fKDktQ0-qtTFPxY0^P
zVj@gG5!ofe^b?U=BKnV9zC<)Q5e3Mcgv2BmKMC0-!T3qYGYNS^{b5N6oYbcwF%8B~
zLw0E}ej0L1gZZbSVQDaaIug@i{B+ba9ln@!q)CVMPe;D#uzkvqSO(*lA-gh|e;INs
zgZY=C!DX=i<wz`t@yn52IgDS9+{$76ax}Oc#;-$S9gJUx?CN0rI^<Rd<JY0VbufNC
z66<07dSq7*<JTk4df5K;XjnaLzt>3h8ugIh&l{wA1Nkjdy@mV^sop_;k5um=e?Y1a
zkeiUI3G#BRT8=?}E3j$><ds;p67njnS_OGER;`8{ip5Y2#!o0#hhkT`KB3qrR33PO
zv=NINq5ei}w-JL)=SJ+d5rh67hQ%<bABOG1pne#33;R@ma2N*TBOHt2Fn&0;3y1N;
zv1d32<0Bjo3&&u5#9&Pf2IG4#*6f8Gi#4&3<FF<Uay-_=Lr%b&1jvb4lL$EpYmy-E
z!<v1N_hZd|$jMlf3^@gBQXpqwF$2~=1FJLO`^msw85lTnGO?Hm^)s<uCTy=v?3M}J
z`#2VlL;d5}?l{yxj@^!bsz3NRd_Or@%z^Q9uw4#}pM%|UVEi0BI0wc*i^a1r{#k5y
z7REn|-Oj@JXYt^(Fn%Ex3t{|1Y*z^57h=ys*gp#KutL~BE@SaBjDH1-S0G=-;#J7k
zuy_sfbu3<od;^O&Am7B|O~|*fcnk7vEZ&BE2a9(gmtwIL@?9+6h5P`E4?bn}1K2+*
zu~-S^mDsKlwof?G4kt=xPNnp556F)iA4Zq*_r&}?>A}nnf%-=cZ_9QZbFfmyGGbUp
zoaHsY^`0bK^E2k3iCFl>iEUKzEOI)FeB`O62sstWoLRFt(`3W5egFFM&S97I@SjR*
z2U8Dh1#eIxyONnNpJrW749j5|TW@HxG9Ei~B{;>S;laPoI%iHr{g`))1289La26Tz
z#Vt-M;w*|}&UyRbNS4nZ&eQ^bi95f<{aJ(OCI)<`=c~P%mhrzE&iQtiC)VU;Zfqw=
zDM(B~YPlT`An^d?Xsn6GU&!)5u;vfQ7qEB%gZL#_ERi|?PQ*74+vSa-CEAQUH^a5^
z3FrBQ3t+p7(~o{%9#!R1qRUBDyig_INZ_PLyzmHeH7`^{uHl6m$hEvs3%QOL>LAzi
zLOtZiyzm(E6JB@%`6(|vh5U>cp2^HE<kA|Z5Jd_!Y+G;VvVKu0E-de`P{Ft7{^7vm
z{&DAxzJ8*}kng~8|1g9(=`c|omK}iHKm2^$KkhvizB+dKl~--^cB&8=7oyJc%PdC5
z#WG*;?7k$>-+br4?ztsb=LhxbQm~9Pds8i+VSDIJwQRul(3@)6i0zpdwWRXY_eZ@h
zKQ!c|AJjj5bewb2Z1rqFbDidXGzvIrp3^))3)BmM=8x(h0WES~^zrf?CoOhf3}}h8
zq%Ug-PFgB01++}F?4!}bNz0LDIRd;Ad90M3_MEf|d8~qblz1E^o@@%^q%7i*1v#5|
zWt0B0{5bJC4w+teEFuGCIlb;!Ec2B?6FY3&tXxY6zlWE{$6KQ8VyaT7<+8WDpG!eb
zDeyMJ0pxT5av?G+MCPm|1nTZED~4Q(oJ)};%S(|<soWZZbQd|_h4Q<|<t~&rBKJnQ
zb?5;da(@9i3I|5vk#hRGao}#q&#=!kJV@?PWn^HPe0D=!pPYX@H=MZ^uDmL$=u~sS
zg>DK<kywfx<b>}c@h;>nEM{RKKZeC)GBe*I9W)1t;vn3^x89X!!@A7%!oZiwrs!V}
zn){}4_P<qj>`jY$nA05QJXjs5-=5PPfqahBoP&Ix)0~(2ftqh@-_azL7B!3Ski{FY
zU4HX2+u=5Ez&7g-?f**YW3fMmDrRt9GB_7G$0Dvvk<837+N}7=RHW1P6#%Pzb)c@9
zpZD}Vv@?2od4BAS%yCNNe>zpsE6gol3b6R>69MXnX#PGx?zL4!QAJGHHvn{M<<u#%
z-$jt}ICY-P&m#ZIaXe_%^<Te(2IZsLRqOqw|Dosc#nOVGjo+T^_`l3z?T|ma7N$S{
z$Eh*<z=C=No4>A*8}|`WJR&A+;d=QtHQA(k)+T!8FBOTY@Jio0EUuHU6ARLMEUt%~
zhQ&0<6<Dm0nY}&Q3zCT<8Lo`2-*K>B@T$-3+`$J5`v1R%&)c{0IgYn?Hq!3)4r{-|
zR?P5u&+QY^^$$l;!CH=6%jva#cBrMr?&MR)U-S1NavO~D88t|zM(opw2eDPD*_-H9
zLH@(0U_naY{1fC!hK(=k_G|sBge5d*^0yPbY1Ztr<HzqFy_0pz4IeKHLd?a<j&ikF
z8!%g0i`mF?=S-Ztxw#r@Ru?TTC;2D+W`+H8WIwdD`pQ51$Uoga{=`z$wJ!#)-*jW9
fuB)3y?bNrAvrBJBt==CMv|Qw$uJTWf{PX_+wS1c<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.38978BB017721061.idx b/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.38978BB017721061.idx
new file mode 100755
index 0000000000000000000000000000000000000000..51cc782372555eec2c81853e23af121c6bba8bf6
GIT binary patch
literal 14580
zcmZvj33yc1p~rJhLKsLC5CH*^O%yS+uPTh1*;ko>BV_iuccv;D$Juuy5G0Zy#?~4}
zt=4M$+<;b)xU^LZNHjjRt<kZ@Et=@q_*|l~iS-R>$NBetFW;9t^UL|=m-)WAxw+@w
zlNN(k`<bGqM#UPE*9W4$x|*7rdF9VYG<?muWhgb))FkgdX1{*ZhqAff{Nj^eFV)Q2
znEv@<<-W^izd(9-EG=HKWNYi&`(Nz2@MP+f1E2TrH3&|8y`)&`I&o+G@YDbO`~F!g
zXRVaI^u$L!q4RJ3-FnAOOTOIcG~e*%NAG+${MDMm%i3RWc>ctq$+w<Kcx9I=Kb>d2
zBE9qH>mL97oYn7-eGvX~{3>TQca;5#|E)N#{DX4o%9{^8Qo5*clk%p!-?^zTF!H4S
zyM-TKd-3&)2cI~&=GEsTC!cM7z5nq)3fqnjB|rMFgMq+?E2IZ!C~EgVbolk2b*rP?
z_iMg;Va;LVvP9vYU6UW~Tz}(ZUv&T4Bp_FLf7QSH;|s1{{^ToL7yZE$Dr{VP$BBW9
z73QH$LD$I#_KDWbopbREhYy{5>>kn0?_DHtZ~b%6v0LW+LfON4bl#^n{fwT%eSf+$
zaN7RA%|CK<hGEJ$!!~!;;sfLM?w$QNzenCWW9P?vXa4jn$3u?^Ill@YdE))8%O0J_
zY+Q5S&N+`SeD30R550QH*AH#}e!~~{|80NQ;j`U0zihD_JrvwjEZ$-NeD4E4BQ4=u
zzI$`t4|P8|*>lOU2hVtM!O@R+2bhbGv>QL#P=7Bw`-yXfZ```AWBUf>Q)gZG=H`PB
zF8#Vubq2Y8!*3t+pS8Qc`&j#tCvG+MIWkv>9*^om?>)!0-O<A_u526>Hvhfx9~T%p
zKe_yB;R6rYUK;O+-E~tq-MoFxp(~fJE9MqFdBd8+vIgd=U5_3WzxMRP_=&I9*Y>=z
zwSMnk?j3nmc<pVMF&9Nvzr8#0TK2|Omu<b~(Yx+EN`JrNrMXY9slTRlRoCLN^;5PT
z^R_hGrL9MJ-q)Vm^=|tKO=w->nRUOswRoR+An|klTc0fp+0>@rtct(CY0uh~TPEI`
zN*ABuf3a-S;d^?2v*w#6hQkXQ*x0C{{}_E_{&Uw{`EW>P-fht@nEk}|b;8Jz_=`X5
zzvs2$(~3Il6>GoQee#o0R&e%v`eRe~9DC;bo_(?Ko9j!#J&End?|pLbTeFKwwtwP}
z!gk4lYnN=g>D;YP^_ai>;}3_P-NCiZGOcMJfAKo5RD8^M@K})h!L`r5eqHn*k8WG>
z@VfiA<o>eN*7J**Cw8o~{OtaYhj;4tZ`+oBX_4TCy<gn0HTA}>-sGH~n=KEmm2ExQ
zaAx0|vX{@9WvdH4Fg{9NvORzGD<27Zec!$__0P8@eNU&ab<0nl{q)qK(2-a646l6d
zo1e`5<(HjXx=wt*x$DGZPknLp68Smjzk37uLVT(JuLo}^dG}?uk6!RfG3h$^?I-&N
z9{eg{9Qu0A3)>fOO24<fW<f)?eujNya>4kY2PGRzyek#ke*XvdhnbfJ=j`j;vg*vn
z_1h=O)koG3$d8=<!-E@MoqOQW*h}{Q7sg)9G=2Ne#?{;Re*Mkvca|<b(Dy_A{^ggv
zf62jt`}kM$>i&6n%imY-du`s{{L=lCUtOfWbeZLU?tgK9@}^ZE*L}WYm>-Bf(Gy8s
z?>|CnYG$YA7$&X=ywN~iTWi>D_1iX2Z^}oL>5tj}xNm5ONxG?l+RMYs>X*%%-c*Lq
z?-4%LIrArM1HH05ytIDl?CDKqSp4A2$345Y&26Bq<>B@9>-E!{%J8L=U;O<yAEyp9
z(5uSBi|QB6oZeK1SttMKd6zu-UIX1!9$r$vWcKu?GQ7jPMEi(P9B-i3^6>on`Lm}t
zmEnhfd*Ow9F1&d6X*C<m!^;;gUpu|24ELQSUoD={zTQ9`<>B+{&r?osD#OptIr8bT
z&rDkz=;re9;`+t2r#F@1lXrX>ym;GM-)S|r^6;vKtEAJL%J3Wi^VY->gNxTdo#o-p
z^_vaTo67Ked)Hqx+^}O?W6hTG@OCm|JGuPd4;AJ6x%S__bHkrLZlEn?<;?msXH9P^
z!{I<MGD}obGxM*(XFo9hjxK3rTRE-P%Fk_N9UKRclj8((aa=%djvL6s@c?-_ULYUG
z2ju7Yffx=0#Bx}m04D$x<OG4joUpaB3pTP5P6Q~*i2}tqF`zgn4wT>|fRdafP>PcR
zN^{ac8BPW$%gF-eI60s^C;uZ=UN4<J{j1Xb{N#O)zJJMzM)pp^xs%u{?{4K+=RPuf
z`gf-4HQAedmHn@gZDN|*fI>}~Dsk;xFu4pDOm2t^CeO|TlgIGD<b`-(^6h*u`3xUS
zeuxjIz%Bq&zzDz;galv;?Lshxj1Wv=NC>9LE&@};h`<zuL|}^TVlc&w7))_U45q{`
z0aL<Az?6g}U`p*$Fr|zXOle38rpzt_Q^v@^l!atq%I$J6<%}Fmc}NbX!ma>Q!6?8~
zgcM*l+narrgJ&b#%rv(Fg_<)}QrT5tsu&fRs*nmywOtLSno)zP4ynP^*fn5k7!8=3
zkOoYxT??j`(SoTBX~ERlbztfk9hkb14otmW52l{cgQ*Yc!8F(nU>X<$n1+x6OrzZh
zrjap%X$%>`G}%pHnivz9rjQ9tv)v4)nK6TD4w=EU*ezgM7z>z|kOfSu-3q3av4Uw0
zS;1_LwC1aa`_^{Sx`zOFlh(U{9T5lEj&|Ze+tE!NXgecLu$}G1iMF$wIMH@RTwuG}
zi3@F4H*uluj<~^gw-Yzo?r!2n+Y|AC?P(_-v_0L#gSI!~1>4(Byl8v7i5G2O#0R#o
zo%qo9brT=j{)iuJe>?G`?e8Xjw3!G4Hq%ZRw3%+gpv^{Du-SIPqRn;_7VSVJ0Cu3A
z1ket2lK|SmND%B`I|-s4>?T39!;x^l^5?CQ4Y!l<9s=A=!gm2jA`!47?IePBq?<(0
zjz*$jN83ph?PxcNq8*FGz>c+(7}~LJ5<@#4iGv+)Cvmjn-6W27B9Z_*(M}R*C%Q=j
z?PMegcCwu$(N1=gB-*J+3hY!nNuizUCMmSjku=!pc9KRr-A&SHXCfJ}GwmdUcBY$T
z(9TA(U}xJ&7VT^|$)cT$<iO6glN{Q)ZjwVgAIXEAZzp-Q^W7wm_N_#8E7?#vFKR6B
z-J(6nT}0GX<++mwmfp60-2<eNUCdo9telBgrW)n5u2b`j>e<vTBHBe{m3rkbV_7>j
z->7Q0llixkD=S)M>d{yBPE9<jUef?+dWBR^!J39h)4RxHr0H|yKHBsY1;3v*y@Xt(
zO>YBp2M8DY+#$lnK6i|8vCr+JT<mlEDHr?PBIRPAH$ZsU=M51a_IYE3hkafj<zb)K
zPkGqq6)6w<`~kwpK7WYtvCkhPeC+f4C?EU$e#*x_zexGm7Yq;q_60*kfPKLj5nx}?
zM+Mjy^iu)$1w|^rzHopDu`e7VLhK92h!Fe2J}Sh%u%8OCFDz0a_C*6kgniKv5n*36
zMnu>b^-&S_Mg3HSeNmB$urD4UV(g2Dh#33gF(SskxQ~jlFYc#e?2C(3jD5)fkzijk
zL?qakj1dX;C4E$aeMvu+U|&+C66{L{h!p$MAtJ@Tbc{%`FYTjJ>`VKp6#LR5m118u
zKxEjL4G|gkWn)B!eOVusVPDoyW!RS$sSNw_0V2n~e2B=gFCQav?92P89Q*QqD#yOO
zNafg93=jqO6+=XUeZ?42U|-Qk71&qwQw8=FMXJDlbCEQ^O)7`U^5mjKnm<FHB+XOE
z$7%Bk3jQE%eiM0wHje_UibREdRf(vuubLz(?5mDb74}twRE2%j2vuQUT_kGkt4l<U
zef1<!V_$uos<E#gq-yM|N2nV6nj+C)UsEC)>}w{82K$=hRD*rZAk|=BGeR}k*A|Hu
z``QxGVqZH+wAj}kr&{c52dNhO+7YV7zOG1g*w>Ya4*R-EqQk!KIMrcaH%N8Z*Nspe
z_Vq=g$G*Nq^w`%=5<T|y$EhCs`a!D4zJ7%2v2Q361NIFiV!*y(k{GaWI8F`NHw;n(
z_6;M{fPG_;7_o0G5hM1Elf;O9<8f-lzHyKmv2PrqM(mr4#DslQiI}i&nj|Ldn~qZx
z_DzG-gniQpHDTXeBxdZJOT>(Q^CU52-+Y{!v2PxvX6&0ss2Tf~BC%lKQX&@YTPBGG
z`<CO>f_=*%wP4>eLM_<07Ks)6))KK|-#SUG*tZ_1R_t2`sTKRy5o*PLYl*ae22<bG
zDbo5qau;pwrZ80=q^)lvkJ8ppkSA#CcfgJkabVvuMI6|7bWsQP9fQ<?ea9$uVBax8
z9oTo4h!gwHDdNPwvx_>h?;NB~>^n!P6Z_5y>cqaQL|oW+O%WINU0u|Leb*p$Vc#`M
zUD$U`P#5;yCE~`udy2TR@9v^*?7Ii48~g52>c+l%g1WKqDG?9$JyXPkeNPwlVBa%H
zJ=ph*QV;e$6V!u!Z;5!Z@0}uE?0dVY7yI5p>czfylzOr6ouFRq`%1)zecu%EVc*w9
zec1O6QXlqxqtu6e-vsqx-(Mnr?E9yPAN&3;>c_r+kovLjAEkcm`zNR$`%H;2*k`5)
zgMFrpGT3JZDT94xlrq?7CMbh_wnSL$vr~k{KHEiE?6ZTE#XdVqS?seDl*N9aL;~0k
zOpyTg16?$L{lFj%U_UTQ1K1Bt&;a&>B@)DbaEb)6AMBz*><0&F5c|PV8pM8Zf(Ef4
zo*?1x;37=8lZJ1laFM2$hW8^6)9^>g<23vYaz|}=S8etCBNHTo{YWQ`U_a7JBiN4&
z(+Kt><1~W(NJni1`_Ty!#eTGtMzJ65rBUohhiMf1(Qz8Zezc=Div8FGiD5t1Nn_ZL
z_0kyjW5YCt{n$8-VL#SU8^eBlg2b^O@1$|;$9riU`|)8K$9{aA#<3sosEuPkF+md8
zPju1*_7lA{f&IiVO<+GUP7~NqbkruWpPV2`>?b>E68p(sn#6u`m?p8G9H&X_Cp&7B
z*iTK66!ue{G==?CFHK=THB3|3PmR+Q_ER0TDeR{wNE-X;PMXGkx|gQ0pB|=Z?5D?R
z8vE&v+BEhv6C{KEOef7?KhsMy*v|~p4E8hQG=u$2M{NfC*$I-xezuclv7hawS?p(r
zX%_q0ahk<`wxc$S{oDk}VL#VNbJ)-I(j4}4!!(Ee+&Il)Ki5&4!+w5(<guUcq<QS;
zdubl~`C*#Jetw+hv7hg#%~$)E{d3P7x39fDR=)VRgRIy=gp~_+m*4VP&39i%i*<GE
zZDjRr#8DZjWB>EYULAV}nSV$1Hf-Jhxh-pv_uR0M+qJL!6P78KsVaAAH(Sp*fAseC
zGp2v=tQ}-drSD1JeS`PeXFe<s%s*}ZnUz%smBk6=>xZXSCRDE<b`ViV^>XCReeI*p
z2Mu?ZZ_qb!n$}h?6LL9RARdPY#OLsV1RMd7kRt>VaYR63ju=S7kpM|KQXm;e1|;Xm
zffO7CP&21_ZFTam;;4Ys95s-JqXE)#v_LwJ4oJ_@0~t65AS1^JWa5~B%p5b2g<}D-
za;!g6<vlue=l#^Wkm{S$ByGB=`W`eXo0@@iO{OYwrCcz%N-mgO9T!ZVlm{kH$pe$8
z<AKSS^1<XQ`C#&Od@u!40hj`%08D{S0H#nX1XHLKf+^Gq!4yeFV2YF?Fhx2Mm}03I
zOtDf7rdTHiQzDgsDN#zml;|X2N~KaTrAjH7Qk@h`nN$X*Oeq6Xrjvmwm&(DEE9GFy
zb#gEjQU#a_r2<TaP61}KwE3dy{@$!?ZU)jdo2sOes=!n!RbZ-gDlpYjHJECp8celL
z4W>q_0aK&YfT_`Gz|=~$U}}|GFts`@m^!HrOr25(rcS2=Q!mwnsaNX3)a&$M8l(m=
z4N3!;2Au&+qtpndQE3Fzs563TlA6FYDNSISbS5y(QZtxlr5Q}K&J3nSY5~)tw18>R
zS-`YPtzcS}RxqtPE10b<t#-J4;Ar&#v8`8E$<g8f)8TM{>0ljTI$NAzIvq|hovaf~
zSBncwm%|07i*<qNZgGR@cDTWGvu-dwEgmpE4iA_f)&r)u#S5m_;RVyndcpLy_`viz
zd|>)mADI3YKbU@pA51^%2a{=Gz+@Z@m<-E+$+oayvJMtZmSw>Vv;@EmI09e>*Z`Qp
zmLQlxM-a>)8w4}l61G<-RAEQh1H^`}u2Q5W0%pV!0W-ozz>Ky;!HhbhU`E*}n6Z`^
zm@!8T%orO3Gu{#hGwz6k8E4~QCR!3;CL9Sc6Kn#^WJ?mvq$3Gtl1+k{YDs~aa-_ga
zu_-XqEom^*jx?BQHVtN`B?D&0kpVNqX28t0WWmfjvS4P}ESR~L9GE#r4$K^z12f-}
z2Q%-;gPCXZVBSJRw^UCO8`<4Nv>UmTh&ro$$GvMVTYUbfU(rT(#i|v$%F5P8cJ->&
zy2?`2!ia2{wPx&O`8@TsnWrtNtQ4)t<(0Xo*1*E4P_F4VxQ)%taS?er9wI--M-=1)
zh{Bu@QIr!QigRK_Nlt<&%}EhuIT@lnCr4D|6o{<_()t)&+I19&1KC+1PGnbsxRBii
z;zsrqhzHqQAYNo&f%uU91>#3$3WPys3xq`u6i5I$SRlbFKT_I}+WP+8d+Qq6^R(x|
zDarbc>#wYyc5Jq7hQrWxv#vY6ayWT3`Sgs9*Z)#m-UAk$zG&f(OC0{a(z!fSIkn8W
zx(EE{JgTzJqO6@-?p)P&64B1;zWDrR<tLuygD1+baxbadOIB6pNsa72QnwGeht%~T
z|CH4Iw93ccDfma~|70(}+SP{DuF7s)c`Wa*r<PS$Uq*AT`EIx^smiGk)j2hyCZ|Ev
z=Cp{qoDNZ+(<2&k21H}dh-k{05Y0I=q9tcRwC1dc;Q|Ri1()t31rkAy7DyC1Rv<Ct
zc!9)`69tk$P8LWKIaMGj<aB|gkuwF7LCzLP7CBcSIplnS<g5Hw9W$%zwz<2?Z{QN{
z5@BURTsi$I??<N=zE|JC|D2XrcAN4P?OzMvE8<6j8DGzwvuIB_yMrv)L6%p{Pg@$6
zSk%|tUmjS>T`H`s2b!A4S^pw;(Ff%(uQ06GSUKKR7SKy)PcNLW)XURwThnHk=GnMF
zJR1*)Z{q_AYyu#mO$a2iiGaj5F_6S20g~FJKr)*QNN$q@DQpU$W?S>->M=lNQvs=M
zY9NhG1EjTSfpj(<klv;TGT012Mw=1HWHSMoZDt^g%>rb#S%F&HT5qiE59R+!XmbEL
z+nhkIHW!e)%?;#f^8k6<yg<G-ACSMz55%-FKx`We6le<o1>1r^;kNLNm1Ao8e-hdv
zK+(1+P^>Km6mN?ICE5}|$+jd=sx1YSZc77Y+A=`dwk%MtEeDiu%l}9<;lOpiQ-AX#
zVJ6}!Ki0CJU%q&Y@~!IwGXrj)*W+hurpNzvUf}-M!>O&m`t;QI{riKdJgKSiR-Xo{
zPnql4>#i=NDQKy!n^8ZLGpnKTw7GLmH_X$h&YL&e(0s<C`DdQBSaH!M>o+Jby^Ocu
z!i!hS)-M15zPfPf*~?a}Ty^f6^DkJ}#1{y~Qu#TOC0vnu!9txeNVQj7t|#XI1w#t?
AMF0Q*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.015682E6D487BC89.idx b/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.015682E6D487BC89.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c4c25ff48838a1547d6e188c5c8fb9e1ed224a53
GIT binary patch
literal 14606
zcmZvj33yXg!iIZJO9`|pS_Rn)$THaMD+&ghebv$$wFybc%}wGcF_Na))BI|IT8fCu
z(1MIQqNuox3tAl9W~^GR%Z#>K6-A3>TBeRmr{yn>f9Ooi=kxsdaN9TM?R$&KO;1iv
z*Bi9j(p5!8Dz727iEY?eQdCrQe(rC$Ve|EsIXGTkR219yx^2@<pUP(R{{EAPZ+Ylm
z?{zo4vhafC&RZlcA3XTxzgq5l>n`=4#nN5ho0@*wk!YEJV2a2;T(oZSx3TDZN7qkV
zG;NX0^4y7Cq2=Ejb_cIGCtl5aNPg=3kyB@WSbF*I+pNjk?%@CHnfSM_+?YvyxMbh{
zA36N19bc@`zV`X+Cr<Z$ar1Yh$9MI-{MmaSe{v=vs;0~}ePP+psedV6{CV@+?*{Mu
zeofa$4gG>At`C;o*EzT=`P>5sHr&d57CJiGUAwXMuZfr5e`WWV<$)sy@7`TE`~FL(
zO?`7A?~#vQyZQOguYW;leY2u(&KI#GMHQhb+DGpWyj1j$uX|Swe-)h)oV%HQ`h`zQ
zw!W1#K7MKJ!u8rIU#)xi1=;SZ(8K1+*ty%U>6LiDJO1Z~@3@3M)||L-{VfB_R($aE
zp>=m9Uz_=SWsB>v%FT-xowz6Xk$%N|&VrrXb%BxC+ErhOHZ`xSJn&5OIqwa*Z@u`|
zl|@Tap?f&Jr*||TyI<Zs^li&|nj1vDn`2W%;){-Vb!^&r>6LSiUcSRts(*F<H;nE~
z>ASlwir#Koy7SV|+Z$}JH$0o2eqv4O{s*T&dx^03_8XdSS#!mai*NjR_xpc6C$VKw
zOZLqLpX?TvtJyBk+PU>_9r*0=qsxDD&5<RFe|~8@XfTu<n6YK};`=w3{&e?)ul!i~
z?4A!kIC%VjKfCN5<*GG{UfuQaqtc81@zU*|j_s;%Kk&yJRK1_IpWIOV;@9#0^t!t1
zr|+oS?|)yo@{YC4s_>G%+oB(&ZeF~$=C_CTY<Z#drAr>0@#50bM@BF0DvvJs_oJ_H
zsun%=Tkhj!*FO8iH+^5;eEz=fdHf#^J|H@2{^|1*rrzXK$G`^{Ji2Ye_=lB7&dkhZ
zFYoXD)|7gq?m2PAbANFzDckh=1=r0_JpHAo<)fA_pIuqEr@X>bv3|?E$L}+~v+wbF
zn-6TRsE^Ou`pVy@y}m#=d;Y!UTZ-pBvgc*PJD=V(?UPHk_B<0j99teZ6(8Qiz5nm3
zhqj+LuUP-g(+>zob-TE|y9^7)m4~BGEnT_sK*>`J+BZGuTjnY|{8>=#;9hsxWoesq
z2HSpB)fGdBR=oM-tzW;olXv(1ohzI54?UlnCO&o1&r5!K?{>k|Z(k8S_11+4-tGII
zF6H%GKUuP9TT$@Y*=yYMjTiTH&Rh1K_1=xcpPxDXhNSbw#G@|xTl4H^gwey@-w!N$
z^_YJv`Jv_79dG<3+5Se`;qP9!T)trWP4X;Hp!A%0%0G0~w&doM%dDToYMvUqHTL4u
zzgxNf&~D*@sm?QXRo~4Xxb~iJulevj$LI34*1F@{;>n}`J}>*ug35|Ns#jdRsbs9;
z{byJAJb!rl&j;4LKjWq+N8ax3{`1J2$*QpoXUQ!)e*F5aBl0R|_~&yZ%E(jEYllzo
z|C^a}=9)kDt1ADnWZ$^#!{@&V-<DZ**&ACIU-{+ZEjugzRJ!YB;<i8aX*j;gbC?tr
zP5<Bfj@@dK)|OG*bw%F!rSs39Jmj89CjXrOWXqh$6GMM4qcyqViqeW{lZX891v@jz
z)(7Q(Dx({6!*!)~`pH9ncuBnc&MWSId|Mf{=Z2S+E?YBs$PX{yB|OqH)$c8%SLcT3
zmd>3%dB_hdC%*f~!EfSEmeFf+!?Q|fO`SaChdsAE%stpK|MWRUSLTKn&Rn=+@{k|)
zI`}UvyZret%V=$Gcuwh@>63^2@V%*3C!62>;)gO?pBr9Xx>`4R$Pe#-Zu`&Im=2bm
zQ*>2sc=61|(#b=9_=)x@w$o$h9xkJf-0-H-P3tEQ`QfAI9Ugl9l<B%MT9+H1T{?UE
z<RL%&x9&SQbJUj|Eibw@H+&nJavNFu-w%1^rx~`VKfmcO-;~kyIpxCA3#UyU^23{1
zUwFBwsA%dV?~mN|^!sK~?%gnD!@~UV)Q@~ezcM~U=ahTvIrY~3y;SbCbL>D4jswWa
zaRRwGE+99@4dmf?fEW$~<mGsQSPl#1<M@F596wNi698)9G+6WhkITIgP6Vit(+Cvh
zM1f+Q7*L!O2TE`fKuJy#D8)$ur8#M!3?~EB#AyP`5>A%<!uf0IJ~DmsB5HeW;=V)2
zFRv{3R@tgH<}ZdEF;zjJP*t)(TpJfmF2e<r8{&e=v+=;>F+4DNAs(1~8y`$Q!v~We
z;)5x$3BVLE0x$(30hmIY5KJK>1XCCif+@0zz!WheFhwB|m|~k4Ofe${Qydb5DX~ew
zlrR!7B_RozQkxV^DI*0_8j^x3v&q1eF)}b^AsLu*n;c9zBL`C+l7p$RDZo@P3NRHR
z1(?;g>W%s5q;hXHQyl~fRVNFivZ=sSF)A=sAr+Wvn;J|tqXttQQiG|nX~5Jl8Zb2>
z4VYS+7ECRp1ydW+f~m9Vz|=81Fm)jvn0lKYOg*CqQy<cUX|Nf<G%yA*4Iu-VMw<~#
zBVz>97&3xsvYEg%F(xogArqKpn;A?qV+PY4GJ|QcS-`X~7BDR#3z$}$6-+B*1=AX`
zf>|G~&lH}?>$9Z3l>oPq`aQt*upMlBme|p@w-Gzqj<5r4N0vCycC--(+Rm^OY-g4@
z(RQ{GC)%#C3v5@GxX^aB5f|F-up4Z5mblS&w-Gnmp0EdOPnLMl_OuZX+Dw=Mo5>Oe
zZKjPdXnVt6u)SI0Mcdm(ylAsw7Hl?4ShU$T!lLa9`@r^Pi4SdG8}Xs-5BtIPXNeze
ze;e_m9S8@&4rECH?LZp|pxqE|$P|9kHe^XdD*<jJ4SRqi;Rx7~EQz2UX(JJ|8^euY
zH)csA+Kp|b5$$L=3U)M0qG(6kNEGc@I0kksOJZop+DHuTcsLGrJWJwe$J<C8?L;^M
zb|Om>XeZi80_|it33f6|l4vK}ND}Q-I0bepOHyd3+DHoRbT|!mI!n@Mr`t#x?MyfW
zb|y<QXlL3;2JNPB6WC2z(u8(X8)-s2OTxKc;gOLgXaB>m$db0)E%-ItiD)}n^UG{E
zcgcv_karN#jsnk^cyjI?t5^OZH<z8woh{7Iq4Tdm&Q7Y%zW#tom8@!2Vd7M!tf~gm
zSD6aLm2tu3D!E{C^;|G{G9H*bB@axVo(Cph#s`zH<b%oA^T8Cz1Yio30x$)70hmIW
z5KN&`2&Pak1XCmvfhkgoz!d34V2Wj8FvUtSm}0#cOo>barbH<LQ=*rEDV0gVlq#iQ
zO7&7OWilC<GNlYmnO+8_TqXxou9Sl**UQ0F$P{2IlnO8vdIgx(vg%cZiBq+*x*AAd
zZ7PsTrUFx?RDr3|tH4yt)L^QWYB1G$HJBQi2271o1Exl=0aGi}f~i$%!PM%tVCrN#
zFm*~Dm^!@<OubAGre3KBQ?J*9X^<JfG$;*V8uSJ*jWQ#cMx_x<quvOnNoE4mq%?tP
z(wo3E%gkV!m1Z!_dNY_7nFUOX(gLPMZvoRPvw~?=TEVpHtzg#I*4tnb<*0W7vGsuh
z*=y}!+8uT<?W`S4N38=)hr<D;gLQ!EtaXCvbU49uvQ98vwJtDS4i}g%)&-`!)(xiH
z;Re&qy213+dcgEJJYafQ5133X1195Oz+_kkOmD3hOs~TWrkC}C$=0%9vJMtZmSw^8
z)%w8nIecLHSRa`FT0fY6haXHo>jyJX8vrxl2!I)217J4PHrNV3_8J@wE+DocP@qU{
z1k8vd0%nAbfZ15v2xg<B5zI!m5zJ_96wIh23TBjzf*Grgff;kez>KjmFypmxFyoFm
zm~l1^W}-F$X2OvGGr=anOx7mBOgfTaCfOvIsoE5nDMt#-6q^DwU7H3o?MQ=}X47D1
zYBOMF92qb(YzEAx+9oiY98F+0u}xrRNo|&Zn|n=^d+AcRw{tW6+|;_jEkx8pq=m`O
z?L>4taw`$F7C8UBLZ?qIm=-2A_mGNv$kq8wehQP*&aR#ov^&X?J4t<B%P*wnw6m+I
z1#KG<Z6mV0mS0fq%z0;*Qwv^I52<>W<j<B{p*lvYzCa!!Ro^0a(yAjA{BBzHHgZ3$
z`V^SkL%7)I9wS`rb4Lgl``k{-#Xh&2a<R|tr(EpwdI%5uykmrieclM+VV~DYdD!Q5
zQy%tt{gj7&eh=YepMQ+-vCkhNeC+c(DIfd%Zpz0#zn}84FX$lx><f+&0rmwWM1Xxk
zClz2{&`kx{7xYsB_Jut}h<)KPBE-IMgb1-O?4&~M3%jWh`@()I#J;G9h_Ej@Mnu>b
zjSvy`MV(ZHeNi_RVPDiwMc5bj5Ha?}$A}pF;t?XozPOW$u`lkXV(g3isTli`9wNcM
z<QS1)Uot`@*q3xt3HBx3RDyj;Kb2r#+C!w+mmVWh>`O<86#LRnD#gCEn@X`S?Wa=g
z%X)|m`?6z1hJD!xkzrrfNoClVbyFGkW&Ko!eR&U&V_$xZ$gwXUA#&`?JE<J|@@^`}
zzPz8xv9IVM3hXP65e4=YBSe9HMJH8YU(rn!*jMyZ1@^1^N%g0sFx{^nCe^2qCrI@f
z<SttM1_i&5R)35<NUKi*tNMuw`>J80!oF&PsIaf<qAKjG`lt%~szIv4zPg{Nv9BH`
zYV4~gh#LFqE~>`9x{s={uO6go>}&do2K$;}qQSmqf@rX>>7p9!Yx<}L`<g+j!M?Vi
zXtA#yCR*%kCx{mN+AgZazP68Qv9BGZTI}oki4Oa^VWPvnZi48ruj`^Z?Cbid4*R-6
zs>8m%pXjl#A0~S2>nDgF`}!`b$G*Og>anjMq<ZWd`iTMihGAmBzF~qGuy5$12J9R9
zr~&(iL2AIhv7Z>RZyY8@>>DSD5&On2YQ(;=j~cOW9Hd6<oBD|f`=()H!oF#On6Pi^
zq9*K{`lt!}ra@}LzPX>6v2PwGX6&0Mh#C9lE^5ZUxsRH$Zyuy(>|6SY1^bp^V!^&;
zf>^L`>7o|wTl%O4`<6j!!M?ShSg~&%CRXfQCx{jM)-GzrzO|29v2PuuR_xagk@{g+
z{jZ-O^=FV<Y5jJZpLFKl74D_=eaM5f{v`4ktse)r4-q@|?Gwa~eS0glW8dCO?bx>u
zQakqTW7Lj)#}IK~-!VZP*mty22lgGk)Pa4+Aa!8hF-9HOcMcIJ_MH>NiG61)bz<My
zOP$zv4pJxfonzFAeb*3iVc#`DT-bNDQWy4Jz0`$$*C2Ia-!(>E*mn;RH}>5V#EpG-
zD|KVv-AmorcMnoG_T6LDjeXA$@nGLGK|I*^v{DcDJ-yU}ea|5EVBa%FJ=kZ42!nlQ
zf-u-;S}B8lrk66<X9g*QeP)a@*!K<*FZR6?#EX4zEA?XE+e^LJ_YP7o_Pt}&i+y&8
zu-IoO2#bBTm9p4pdnt>3c962zXU8auecurAVc$1FeAxH3QXlqxz0`+&-yrp2-#12m
z*!K?+Klc3-#E*S{EA?aF-%I`2_YYD(_WfhjkNv<931B}kK?2whw9)|f1HCkW{lFj%
zU_UTM1K4jECk?0JeW->O+HgB9y!dWtrwtv*1GM1;@+fWi7jkoP!<OR0*GI-l1pAQ|
z8o_>~okp-98K4pDM@DG``;q422=*JtNh9_fTWBNp8{26k_8SLiBla6dX(RR<n~NK<
z9~~!A>_=N@6#LP38pVEefJU(&9i>t1N1KbI*pH2q81`c=G=}|HJB?vKHb7(8kB!n8
z_G8V(G3>|3NgVs}78=KXyq(6eA0MD`?8irG9Q*O+;yCsb<0OIoL<>z|KhaJT*iQ`5
z1oji7G=cp@b8!Ou$#IgzezJuov7c<GN$e*FXcGI$QJTblvbi{k{nR)~VL#PEQ`k?n
z(-iho12l#G)F@41Kh<2E!hU+3q_LlFp=s==+i4p6=>eL?etMLqv7c@(PGdhaPBPff
zw9pLpGwn2k{mcN(U_UcTGuY2G7iX~FG)|hZ-_$~zu;0{9o3P(BK%21NG)kMW-_%^(
zg#9cT&%L*aXULKknk9ISENRc(f@jH+f!r;4o-7&7-GXPzlIG$pDV%HVzgv6nTybY4
z_bzBNsca^~{NmlZt*44k|Jcx9QsVvf;$MmPE>dw9smqHc-e0dFmUw?hDt<>?`EiSU
z*G)6I+jr*9FkdlWmEW{hXT4zg$van1nLP2dW^!KM&+oI$?an;A$r2U~(^YrCnjkmL
zMdYP<i2O7kQIHlO3e!SFQCfs3PKyyGX$hh<Ek%^2Wr*^$98r;0Al7$~`a>{Jws#Oa
zvZI4IkewaGiR|hiE@XEHaU**=hzFVJAPlm%gLskI4#Fb)I*1S1-$DGyfesQVaDK~c
zZh80Ye%Hc1wUvmr7M6kkr!#xfalo)Iw>z|oQ?;V7ip%A2fp{Dq5TC;b5^w}SLXHqf
z#1R3BIbt9QM*<|}NP%P=8IYVK2U2hpK-HY;6@|4K6-Nc6=BR-*91W0`qXp7&bU=EJ
z9>~Bk02w((AQQ&~WagNGEF24vm1F&dw*FHyZ`FI-=j1N0iWwC%^Sc`E+P`$|?Bzp0
z&~k6(;!0iqS}XT1S-eD---Mk1x90BGvm2EQ_v^1O{JfsKP|t2yE@;*1>V2@drAn(1
z)oC@NCapo#rnQK=v<^|9)*~9y21H}ph-gZi5Y1^bq9tuXw5F|y4SPw$5tsl*_L2y4
z<6hE;9NkNz$g#a7h8*8Z;>d};B!QgVOOnW`y(ER4-b>QRnY|=~+_aZ8A!o_n+z*h#
zEsqV~8o%!Nz8xjy-esC)#=<qY`pVUH`TM)vTX$96)rD8>!D+#B^K;fiu@|RYwdnzE
z?(sD1+*vd8Z|&rVa}SlXyV47ffnWa&$ZxL6X=it*7ql%zw59M&@mk##e|IhHdn5N<
z?j$8U$>RLagmUjLQnCyAZc=hL@@`VHyTGr1zSlEY`WIX7yIo>fvLXLS&F6Akd;YTv
zKL3q;q+_OSzhlOhoIi&<N0@(NJ0Y0z<J9wJwdQVaCKol6h55kH`m#9|_4W7X2Ig|-
z3iI!WojISg`c3YvuW~n68Y-{KUtRgl_0s8+o9go?%{?8~RMi#k3T_P-h*!e{;@9wj
z1T_L6VT}+-R3ics*NA~6H4-3cjTA^$BLkAx$bl3!3ZUwmYIw+~YE(e#8a0rnMgyd+
z(E{mebU^wVJ&>Ws0A#E&0-0(|K;{}Vkfp`~WUa9R)d%ZuE<CjDK|7Em=m2sCoj|Uj
z3&<UG19^fTASTEFd4pacHpl|`f<7RB&<_*{27nrZ4L9envD_~v1S3F=!A78HFbWh4
z#(?6%I8Y*(07?dvK&fB~C>=}#Wr7)?reG6LmISlp7cSb&-njAXFNcITGhtWmW3ktH
zq1|}u(7J<DS?5N#%fl2+j{oNuNSy!qaQ5gae0uil{`-f?e^ONBE_})sK4ms}Zw%zn
z<g*l)OevkpnO0VQ&W!WUHJq<eEjxd@q56VZ6&GGSTe0f$)oZR$uH{{H>19h~D;EBL
rx6Yh9Z+_*X#g{By{+pFme1T9bl`oLY;fmB3&D0rvRBK(QC+7bHEgQC|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.24521662C376153B.idx b/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.24521662C376153B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3e2d48cf35e1680f1a12a15b0a2689e5a1deb194
GIT binary patch
literal 14474
zcmZvj33O9+qKEtc7fPTNQ3eELR|IYLRYVNUzN#e{r|Hdl6Gf4j<R;D9Oa+S;8#Wyt
zweE^KE~8J;aq3eUpI9B7s`G@FDO%K`l~z0BSYFktFVxB8=Q)pajQo1P?`u76ZkyZd
zdb7c><<g1@ZP1+B9E$s=R8&-)Q~ZaE$8K0u93R$HRHW`X<lKC7zjD^m`;R@g<<WbC
z4Vw-wId`RJi@fdqN8UQ!_Q2bB>h>*H?D^XIn;&*$+7|7fD!qQVqHg(UD*5g!>!&Z9
zzD#L<;pCpk%CF6P!&jV@)(9U}jePyp$izQathv8%W%kycqSMc(M-SeV&wX&gJ@<XX
z6J5g$tTnv;;h~e?_6|h<G4}DEzr1$*-M(WJ8L5V{*Y<MCpK_-!efg)@;hymBuh+hL
zB>p$??{5fKKk&!T_hesqXuo?adpz>W*q<H#JN}w^^}U08Pt<Jedi<`v8|FV)JAK+)
zON75W@_O_SAKvhi`pUOv_bwPnbyducOf@`l*Tz>XzWns)>fuk4Q^V)SLeIT)Y|3qK
zXDv_Hrhc{FF!htV$6iwI<wqW~EltgAzxJp+_>YhO`>~xD(kELpzgoX#@S;`kKlfDK
zo!Qst{$XjG_sONP<;za~HhjdiY7uYoE<xSKuTtw)4@ft+)-B!teCt{7p7L#-xAl^W
z6}iZ7c}Kt9(fZMYs@7AZZD;E@Nsq=-Q>C&wAHR9`W`FHf3tm~X!&zne)1uE<<3v@@
zo;k_w))l*IBZuA2L-7MGGfu9py6=%02QHKx-M*=H%i1ft=H1k{_r1TKmD#qet>vx7
z$M#BUbfGr`>&|z*z5n>rudKZI+O7*O{rd^$<7V@e{j;_W&wDUd^}}6{9Q<bKfqn14
z|M<t>9lz`y_3E|D{<Np>3B{Z*U)|pSzk6IA`~P#3_UQ4BA$R4=pQi7l4I6Hlv17x1
z*S{yZWam0|HFrVh?aB9Z(dFyvZ+vRswwJ12z3|CdFR!Tj-B|6LHOa+4K5>x8U-smU
zf~Tv0ec<<>^`3~HbI+d_ivIcdL((DJ4<DYi9?ec;2H!vTiQC=dA1t--=H@SZ?Y^U<
z)?D|77i6<vc*Ao+_2xt8HZ00KcOuYsr0v9kOE&DQnH`wDe%r#QAF#Z0&(jNI`(v|R
z>GN(o_{H=?izV|H-CMJ*a^dgxy=H!=|K{n(F1+n8&xc=3t=u@09^NN-@C)su?Po8n
zG(G>^Ly|G$9>LK)=EdXc7n9GfxWvDI%Cn0*Ha~LRMc(QckB4=Ppdpj{=BXQ({#p9q
zuX7(X)bH|6>)yJ|y-LenGh?^y=5NovuCMk-<9GMZ@z4C|7QgS#i!Jv)b8@PMIdJcD
zM$Ig8e4{LL@UpwacXO+4*Y!22e)FTm{?+gAJJoBKcAw6_bMMp7U7j{QJ9^CjKyLZo
zPowIopGF_veEC;Bh3>ce&sv)M%=py8)3!x3qF2rQ@Qvx`i>9B>&${{A)9dxye=B|T
zn}gQ-x39Qj>%!Wvf7t$A$CKGdqq{FzD7yKXBR`Q{$0~!%_wuGbTsV=Zckg}d+oR5p
zUpp{u!6$z?(tOO7?pu`JdGe!oeY*E{=P$qJf43dl^wn#(eD>JPv6-9J_H%bnzvZWk
zompNme(V2ywC%e$b35J>89yJtdT{!GNzRFFxpv*xs#D)TySRc&Zw~wisi>Io-2)$O
zwJIFd^qOB+unVgeo-_GSJdaHNxsb=0zfN4Z;jD_Qi-(uZU8b0PC>>t8N7B_c?fPIf
z-B3KdplZR4$%oQm_4ohy^6}5o&s0-q@$lTLxzi>eN{54t=%tI+{NY43U0*yrziR%B
z$%oS6oVxoO=R-%nsHTm@!#7slxPJ1Xboee)ed^NIjyKP$xT<(~@!Z9$CLc<N_ji2r
z@a#_()Kybg@$ibO6>BFSN{0_<x$74^)xEfyHWUx9saj*4d?+1$<?I(v9U8GVR8vRs
z@Vu&dGbSHOhd<iT`|Q=v1V^eVQ#|aba+oF`N{3H<F7LYM+|-<!iffC9Zy{4}A?yD2
zuM+aZEa!6{-u%XA)wI5d%&wX}ee$7nxFQz1j(bB=Q8Devb+3G4d7dt)VH<c2SC-y!
z4eR2$U^Vg@VKF=gmYe5><>7f?d3jz~KAsPjpXY}a;00i@JQh}v7lak!g<wT_(JM=z
z$r?7si^1Y}9IQAm4lBV+z-r<(!AkOyuu{AftTZnTE5pmc%JQ<Xa=aX@W?u8ZSLsRd
z0GToQY$&|`{R2;Zyk=<)TkEX#m%g%!i>(dAiqvMyi|^zE$!GaM@*{j81x^8w0#*Q|
zAR+)#=oA7eWQ9NqBSIiWP7#nIRs^IdA_7wE6ay({#XyQ9Vjv|>36K(20;D7&0aEIe
z0x4ysKuRN0AZ1P&kTO;Vq%0x>Qtp%kDQD$C$|G_h6;1_^3RVH6BBB6N=~M!#WR*ZF
zBT67uP8E<URt2Ogq5`tcS?4dETx!@lwk`}SQkN|+jZ*`phSdP6iD-b-I<-J*SuK#-
zh!#klQwOAu)d8uC=z!Ea^+4)bJ&^i{9!P`J0HlF60BMLAfHXRdKpI&ikj97+NR!h9
zq=_{FX^NPDG&{{enprcD=7<?ci_-$6g|z@_iCBQNI;}ujSu2p%h!sei(*~rCwE=01
z*nqS<?LgXDJCOE>9moc*Azwa^H?)w3I|%FoY1jwb#kqiXwGbE5t^#o(-N-cp-Pl4J
zk!~!IMx+^z0h(zc4AM-2Fi5*OH_+}D;zrtCAa0~RoCj!63-KWBDG(3RUd{`&w}p6-
z_7;d2X&>hU+SfvSNc#%JhqRyb1MP1iex&^c;zv5b1%M8;kO0zw0tq0^axBnn3t^FF
z3xq{F$OVB8wvZsw!2$^)9pXYjhgwJo=}>`$kdAWEeEIuqw1q_PAg~K0x({}Yivb;L
zAu*(51rkG=<2azX7Q!LT6$poPoQne;Zy|A{;{_5&I>9A?PPC8&(uo2|Al<|@0o~L>
znviZPkS3&)ToUMH3rQlKERZD9DJ}(cs)eMGP8CQB=`@!HI^9CjNT&-VjdX^~0G(+e
z8Kg4>l0iDlWr5DNkSx;K0?8ts<8nafT1XD*T!G|}ZswYSZf+sXNH-TqGtzBD+D6v?
z`{!A9JCSb3eFu@=QQot@f9Cw1mtXQQsbS{{=1EFFM3-JfoIhjo#Y6e0sk(gKy#&4?
z<~2CAc`Z&|UWZek*W)zg4LFT?BTiG^gwvci<Fw>0IIVdrPFvoF)1J5EjCPV}7e28%
zNenmFNjTi`P7=qR=p+f;O`W6(ce0Ztai=;-3U|7bq;Y3DNd|YelVow{I!O+9b0=vo
z@6ziz`t$3!@@J}*NLz`r^mR~rUH8W#cji@HnXBy~weOPhPr<bVr1k{vQBwPN++DQx
zWeWUZTKf*}ep-7RHou4PvCkhMeC+c_2_O6XF3QI~|1jlapWjdU*cbE=0rmw0M1Xz4
zC=p;^&_xB<7aXPn><jv-0Q<roBE-IMfC#ZK93?{R3%jTg`@+Lih<#x{6=Gl1LqymY
z4G<CbMWaN7eNh(`VPABZim)&0ry}f&dx#kO;sGMYzIc>~u`lkTV(g0#Q!)0%{Zx#7
zNe_`=Uot=>*q4kF3HBvjRDylUVJgACq@PN#FYO^x>`MoT6#LRqBE`P6i%PLCJxrz8
zm-bUB_GLXphJD!pkzrppN@UoVbx|4iWrwK@`?7v2!@j(S$gwXUAad->M~NK!@-8aJ
zzWgwiV_)7+<=9vB5C!%X14My+#VAo=U(rPs*jF5;3hXQTsRH}T9-_p)a)2nYuN);x
z>?^yd68p--REd3MKUHF1)k9R+R}Byq_En=qg?&{QRbgLsn5wX^>ZdB~*Y%US<M1u3
zZiLi*iF<<7{eZih*1bi`-{R`}Xx%Z~L$vM-*qVN#!M<jMXt1xDAR6pzx~T^Hnm($*
zzGjGOu&?bWTI_2_h!*?W38KZmwwr3PukE8+>}!Xp7W=w>qQky!gy^uZn;<&u>$<59
z`?@}=!@h2a>aefxCwlDbM~EK#`U#@PzP_94v9Ir=dhF|ms2=-<eqzABVT2g4Z<rtk
z>>IkN0sDqNYQVl>h#IhO>?cO-8%Kx{`^E`k#J;hc8nJKeqeko-ho}+zrha0=zG;M*
zuy2|mChVKKsR{e0K5D|gX^5JzZ|)~%?3+i38T;l5V#dC?o0_q2?xSYxn}?_w`<8xU
z!M<gLSg>!IAQtRfx~T>GmOg61zGaA7uy5@rR_t3xh!y+R31Y>*wVPV8Z|$R2>|2MZ
z75lb+V#B^|gxIien;<sq+q$U@`?fx6!@g~Z+OTi$CwA=HM~EH!_6cIgzP+2;v2X9A
zcI?}Ss2%$a!=zyZu0AwOkcJ;{x6_6Kg{u|4w4o385N-GZ_c(3%7Pf1cxUlb<ATI2?
z+Nlfsu3qZGzH5lOu<shDF6=iBlSb?}PLM|IH@4G8>^Jt(M(j5Z(MIeyj?+f$GsA?z
zJ~KfW>@)3@!9LSV8SFDdl)*kTP8saGhlv~e?g`??zPp{evG4AsZtS~<s2lt4aq7mt
zXP9`f@0lPT?0ed&2m790>cPHeh<dQ^8K)lXdxwb^``!uS#lE+lda>{ArC#iNho~3(
z-f`;1zHgZLu<x57KJ5G2sSo?UUh2cXZ;1M^?;EE+?E8m_AN&3Z;>W(fo%*rw@1=h1
z`-i9>`~Gq2$9`a#1h5~NAOY+L+Gzm$fnFNGeqe|Oupbzw0qnEGgvCBPL0Igw?Ucnn
z+e=yOvqO}{K08iX><5QQ5c|Oi62yM6od&TV?4?2M2Zv}7`@wM<#C~X)gs>l)AR+9B
z+Gz;;p<WuoerSk>upb(yA?!!TN%UKIxgKq!(d`tjlXcMO-M9y7^d#;v8a<7>wKBS`
zvh2skNeuh3HX6fztb@j|9~-1G?8nAv4EwRx${6;!al&DrYoi?Yxem%<pBtnc_PH_2
zVV`TQ<ggzfCvoh@+h`p7@eUfteteL|u^%6!aqP!iE92NtjFSZR6Kyns{X_>%U_UWP
z6WC9T(FFDrt(6JvH;t1f>^HU1ChRwL&?f9R4bmp;H;vII>^HSmHeo+GPLkM9w$UW^
zlN~gP{p27`Vm~=Xlh{wTRwl8Z8Ye02r`l)=`>76^!hULyrm&wHqbcmCS}Rl7Pmhx{
z_S0=Njs0{7O=CYjNYmI)kI^*t)2)?h>}SSF2K$*dn!$digJ!Uw8KfEPXU1p-`<d3t
z4ED3*B#ZrQ8_i-r+d;F~&koWo_OoL&i~VeCWfuFnagxJ+u8rofpX;DG?B@n)4*R(=
zn!|pswK9kO=5f-D{pL2>jQ!>g+Km0?LE4P{<}uog{pQxn=5l}Ck9Qp1y=r%&c>Q$?
zS-OQtN*C<vZX2oi_M7<Mrc7aXkPCJYSLwhMb{Cnwi)<+GJ!JMCa!q;PMP}bcyrped
z?YVibpnX^IJuJL*p|<pk@(ov>yK-pv<x?l$@$@a^>=IwP#iMwk@XY-k<qL(k66vkw
z3yP%^Xc0SepGO({<$J&7k1a#Ye)B!Wn;~j>wX4dP4f#AiECEjdOUM(#67fW^#5^%9
z2~Pq`%9Fy9@no>%JUJ`{PXSBGQ^Hd5RIuuJb*su312jAhEG<t9OUKi}()0AN3_JrY
zBhLuS#52J%^USa;JPRx<&kD=Nv%#|S?EhYO-nU}i{FSHvNo&|8%a<5Sx82mR%a<=V
zmM-m=-deoGf96h{@+Yu1U%Lyg_ww_6oPxXnr!X(XDawm*it}QelDq_`G%v*|%gb=e
z^KzVuyaJ~(uf(a!t8g}Sl7^?@HIS>5xNtXil1AK2Ct+~AJBb^&r;~Vadpn62x381<
zaQi!nA9tXW1aPyRgvA}~BthJvP7*5b(vzn6R6BE1RQWUY%M-(s!&}o0AK$ZMN)5Y0
zzrs?oYS=Ybt=Uli5_VkexTgGSaMScnGfPk9r&2FZy?XOQhT<t<?##J!OTRKGsl|7H
z=BBIiyZ_~RsPwC&B6jBHt1`BgNVk^H8?SG;;tTJR-tOW@Ss+siWO?blQ^W2gQ+DFs
zO{VO|eJ7c6XL%p`uru&^)f>*@N4wm-++F&zK4N62x9^;_t@!5og87ot*{t+mT>R+H
z-1}Aj=ze+fJSm>~&9t-6yJPaKIA;r4Qu^plt*>5S*WK`7@xTJX0!it&E~Uf8yN-Uo
zXRH*SIEQ!nTY~dGDe_CqORg?G*-H12eNYt6+(A~ti)Tk=?dtNUtX5sCfn}<-!m6*`
zP+oi`A3VOA4<6sd2ajLR2T!0BfG1E3z!R7R;0fvl;0cvN@Puk1ctVp9JYl^MJdsia
zo=7bMPh=8-C#n~LCsvBV6RXAGiA`eg#PwqEBuWW*614<8iAe&Uq+SA^R4D~ds+NK$
zHA%sf)=R;YDP`cv)H3j7CK-6LdKq|fr5rrDS`MDvBnMAkF9%PdRDh>YE5K8j6yPc9
z72qkAO7N6wC3s4c5<F$S5<Hbs1)fT+0#9X9fv2ihfmf%jTU|PX6mPCl*J)sx>a4Kp
z>)<q{QEI@`s5RheOd9Ys^&0TBN-cO=wH7?BNeiB~UJIU1sRK`^)`6!p>A=&~>%h}1
z_2B8%dhqlnJ$U+hJ$MGC0X&1+0G`2Q0MAfw0MDp2f@f44!84kS;2G<U;F**r@JwnG
zcqWqxJX5_1JhRdao>^@M&ulV-XRbGcXHi<fv#2fLSxgr2EcF)ftV%0*R<#v8tH}zU
zwcZMzO=$zqrnZ4+Gugni)!V?cEA8Of)pqdgCOdfcdOLUxjs|Dxn@aIUU#7teE7Y(N
zR=6QrUM`0VJQw2v&lPfk=L);PYjiY%*T^)2*BEL9uQA*R9^+uZV;BZJCd7cpgc<PM
z4mWsi#toi3<Oa_jc7x|}c);^89`HON4|txi2RyIC3!azpg69o+!SjZ_;Q1Ur@O+F9
zJYUEMo-gbJ&+qVq=V$!j`9ps2{9!+M0Y?D5022T&5DI`72nWDp9V~b(!-B_#Sn${|
z3trF>1TV-0!3&0h;041$@IsCdcp)YPUMLg-FBA@e7j;CPr5^}u*eDbA!U{z<!U{*D
z<rQ<pz>6_4@M57Dc(HH{JkG&^$1xmuT!;gY3v=Mb9dYpDOdPy;C=Omc90xDqNPw4M
z65u653Gfo(1b9u3Ch(e=Ch(d<P2e?!o4`vtlHetoBzVbC61-$M30}&P0x!j+z)OWv
z;HAPT@Y0SncxfgLUOJQpFC9*UmvLmk%P<-6GNBB3nQ#WYtRo9vmdS#b4Q0X0hO^-1
z969iEOb)zUC<k6HoCB}f(F|TQ(+pm7s2RNGa5H$@h;$oKlzv{XVQ(YS+i>4bq_>xM
zEOe9q%>UTI#aPZ;{C6e0WA|-uteHc%PYb$zUXMRe;X8Bu=RfY!Un(jlt)Hi-{MJDE
zuV-ri`V%burJ{n}9K2~`aaymlS5BE)HH|mDy5_7|XU{aBqt{+^&J45W-1BDtYTo=y
zSFgEz?G@^E!a22<U7%dG<o|o?-18SMTDolcg)3HGd<kD9mdF&U#qtFLscz0(qa{QQ
KSJs(`?f(G7<?&<y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.050ACEE75FFD374F.idx b/.cache/clangd/index/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.050ACEE75FFD374F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2fe3f761f744605aa14aa612eb86d5125cb3452
GIT binary patch
literal 14450
zcmZvj33yc1p~rJh0t^s9Py!+#f{2-Ytsn!LeeGbPGMRn3S5Xr*vt~9UiJ}C|1hBPL
zGon_-+Twx>OVn1YXf!tZUZ0wnv;pxM0|_ZqEomBkL;8*L_xm2-$Ibla{N^_y_vYrF
zlgssHgQ0G1Ns0C*bMp3^W5KeLl9KuPztLFawk3HuQC?D#%pJ4ee%;Bn7xg@L^g7;W
zg3wa;&Uw{ew>TbmIl}*ZKQd}&J~<dY^p{^xe0xpLY(cP=={~dXlGWd}-SXALwQH`p
z@``u%KYPft;Ee9BN1i^tTf1-1{qN6t?e+ERe~7)`^URYcm!J6Z*j?A2UiMJW+<EEk
zC!bt?;^|$dUwhAV<h3KpAI_~2{z2(^sc&eaa(##QvGzT0_I~nh+n+Z+x?$5;)!<7n
zyPvu<8&*6lZ>xW#pWb@r=$AghZMW4PUwn<H;q`Mj_WtITeIF@aoVK>~ksjXDt2Z0I
zw)|=Nxe05XA~A6Lp(SVLU3~7HSKqy6pYY8xPs?8J>PJ3OPaKK&be{L#-M`)QF#nEE
z-wK}Fcm3&E%kTYVO-u6Q%ICAIqXU%}bM}><d{8mRdFT1RpWUttyjs0mW7#uMw%YY8
z#jO!z?-vzUp5-mnZMyQQ*Z=#uB{QEceXI1P)to;sn|Jk)d;fD{&Qq~NFP?BNJ20QE
zp7!|mdCxC+>)OGCAFuyrfA!>svrpV=d)GDh1=W`QD+e4eS8n}Z^OK40|N5Y$^2pnF
zZr%CA-7nwt{-a-9Kl9631Cd{EmMjxIu1W1~S@SFvnA=|c!u3R8_io|z%YL@{mDoM|
zf~J)#UeX@AxxBRI;e(4mzRw=1JX}9!{@}qgoVm&)tLv^`psw70;#$LRbLy(qmEsK?
z!Iqf3?#RZ&bN_K&`)<>VS01Xf>0f){i|<~2%DyEy^1+^kiL26|y?G|N?-|usn;*G*
z)q_Kq#@aqu^|7O8pzEqHymEhzQ~k>&y!}@^-z@5qzH`S;b=~AH$=i?a*v1u8YuE2w
z*3-1&x!;c~lAf|>r0;brfA`z+Yj4<gc>IBH|3((2UfIrN7Hw3tF1Y*N$9kS$^X<ff
zLCNOP4|RfrpLKcM2UYeR`#wDC2*g*eeQwsmO*>{DoLQ-Dw^y&;(qo@{^!lSWxKfLD
zym@y1vE`CWmo1xfU+L0kANu289y(mL;H6F9b=>mTt!FoV-20DDC2q}McKrO(%hyid
zv*l(zdura59WPw@3+3O(4jn$cnj3t!?8W5|Klk*OtNb$$^&ZfB1l6{+sZSLb-TZL%
zfsLmQT-`Nx&)aY9>3Q&pquhH<R~^2S+_1Ic2SMpayTmich~&lh%Z~Nbo}+r<=f1ux
zuGm#_$1BS>1Xf!X|M|$$tFGDhgTj2~`F*E8vq|6n>aGtTeeNY?%JkytzJ0BCuB<p2
zt|~nle)10GnV#%>@AnH<G=60~u=KmzN9RP07i~B;|FVXe-<1Yl_qwk$?s`?aKlV}P
zv0c+P-m|p++~lr*_5Hqo|Jl9VoVekIPk$hL`p8Wyf5n;hcJ}yP<dI)JGybmq#4Vqf
z&N+MT)E~A!t=zi%_4guU+r%rsHVrO6u>Rtcihmj(AAdFaMyY%EKXY3&$N#wNljZc%
zWyjw<{KOv=k7>W@e{0>wOH#DK_D=4&V9&p2oSyZnESkEViII|$S^s+cZ}(UgH_WCt
z)|N0!rZ1U4waK4DrvA(hjjhT)-??r!b?1j`rq`IKHih9e4@f$9&bZ~K+4P3|@Z#x<
zXH9Ji!`#h#-_0)h^muuRJwLo{+BWsnrZBwzlWRL;d+!a-DY-5`ytHEJ)l-|oaMi@w
z(cgZRdSN!*oFBeq`Xw`_HihBM50!qsko>%BHg)8O*G*q%n%WeGZ}NzaT($o2@!7N{
zKfGxAqFGa$!f;!~qn>A_2S;a9SAO``>9^i6wJ8k0G4If+V<T4QY<hiuxMF(6tf@_5
z`0btJqB~W`zMNB1ogcou;&R2*rZD{8hHZ~*9go(|rq2BEs_CmXPHhUqk(;+f|3^|%
zQX0vuU-$M`Rq}GC(pGt6;Z@}f*Tx0n*?2&F8y`qu695TqLLiY%1SGbJfh0Bwkklpx
zlG$WHa+@4TVN(EA*{W_V?1pkiWm5sEZE7HmO#`I0X@PV$9gyCp2Qt_UKt`Jp$Ye7C
znQdkui_HRLwON5`>uSS=<GP%2)H#5hbxt5xoeRia=LYiBd4RliULaqc56EBV2MW{$
zfP!^FAf}E1V(VC-hWds`;h-&NLiHh_aD5mkQXc_|)<=P2^)aA$eH^H<z7Z%<p8!hM
zCxKG+DWG(H8Yokr`H^OPvgM60EHBd~<xDN7b`yNjISwEv#|h-(xPaUoH;{+p0rGOZ
zKt7HS$j|Wu1vmkqASVdKa2Oz#!vZyM8aBZfof85IbHYFoP6Q~*i2}tqF`zgn4%EnL
z1WIrcKuJy#D8)$ur8#M!3@7s=70&37lUY-z^6Z}`9zSqm{fhia#$FjLo@6T7$~%BU
zm6;-O?OZUqEEh~}hzlmq&I6Oj^1$SUcwqAFd@%VeA54CT52nB_08_vUz!Zc8U<&O*
zFomoTOkqd}rpPV=Q^bnE6oo`!itS=B#jF@iaYziN#4Z6-!b-rDgd|`}?NTtMtQ1UX
zND8LRE(24>%D|L`WMIneaxmqr987sg4yMAc08_y#z*K}3U{=|yg2fYh6<c)&P^c<X
zB$ZtSrixX8sS2sURNK{Hs#!Id>W~^tja>t#hSh+n32DI8+O=S6SuL2_kQPjxT?eL)
z)q$xC>A=+6^<e5*J(&8C9!!JX0H%R8fN2OBz%<&8U>aE?n8uJ1Oq1ONrinFyX$qOZ
zG~3N!nprcL=8zdoi`@dIg|&cb30c6j+O1$(Su2><kQL0@Xzg9a^LTADslA^7XG!ft
zz>cT`Y)3P3pzX*K2inf46KrQQaiZ<a5+~ZOs0(aYGjXBq$`Tja?x-7VcQbLL?amT6
z+McKfY)>=spzX;L58B?S7i@1c@uKa`5--}ms1IykGx4GA%Mu^j{-_^pe>3r;?avZF
z+JR^Q>_9UKpdHAP0NTN55bR(x38Ed$k|5ellmVM*CJfq4mN00uQ5I~rnXqWHS;C^-
z5Np_0{5jjuLK@l#a64&u7&sIQfgNfgA+$s7B!qT276v=qLc(Z=+esMhNGt+&q=iJ#
zj<k~q+R<1P>}U&#q8)7~QM6;R7}&8E5<@%IPGV@sV{x$KEhLV1yq(0+Zj3d8-Pl4J
z(Qa%fjc6xg39u6_B!PCKog~ms#*$zsTSyY^WIIWsor<NvPPLE}+NpMuLOUHxgPm?6
zX|&VrB#m|^mH|7{LNaJ)+DQiO9Ynf=Z1`~!%Cr(`EAmbv-C5+qS&Yt_n&K8OoA!{4
z_Yiv_bJ4^Lm+e+^pCsi>g`h%Gm?#!z!ueNUm<AVLeGie|Lw;6xb>U<F+U&w~xS$ng
z!k5jOngJJ{&R?=MRqiHmUDm`!<TddS`AvL8L6ZPc*d#<0HHi?#O=3hzlLS%PBt?`p
z$q?mDazsUw0<pG()INtl6FP_k+1Wvy$gU3JLUwl$H?pULc#ypv#Eb0fAU<S&2k|2Z
zI!FLH*g=BGOb20**$%=Mxv(g4G{2)SEJ?tw>LHaM!M>^-B$daJ$4KShkvnPSYqYp;
zE4yjshsgc3asZgyL%7)I4iYZ*xnqQjeQqb^VxQYhx!C9SQ!e&-J%opS-XP&&pEpK$
z*ynXp9`<?Nl!tv@KjmSc-$VG==MNG-_W5Ikk9~e8<zt`UP5Id8_ftOh1wBN7eZe3R
zU|%ps1lSjJQUUe_-Bf^mK|d8>U)V#0*cT2GA@+r1M2LN1Clz8}*iD7l7xq&j_C-BJ
zgniK<5n*36Mnu>bby5-bMcq_{eNjIZVPD)s#Ml=P5;6A0V?>O7aVHgHU))W_*cbOx
zG4>@rM1p<EAdz5Sl3!oNzNC{%urKMR66{O*sRaAd9wNoQbdX4~FC8ON>`ObT6#LR{
zD#gCEpGvVW>mf4i%La)I`?4`2!@jJO%CIl%rZViy`l$^2@*X0`zI>3#u`eGZa_q}H
zsT}+AZYsyVyr0UkujnBP>?;O|0{e<FqQJhQlPa*U=%xznEBdJd`&9#^>aV0Y1FRY)
zRcDYVN!9nrU9{?b3Vt80`W$(fR-FP?4G<OfRii|OebppUVPDlnRoGYcQ5E)8!&HTR
z^#D<0Up-3H*jG;yHTKnARE>RgA5~*tJxtZu*9;I1_BEqKgMG~;(O_TGMK##h^id7=
zHN#YceeD3zVqZH-wAj~95-s+%T~v#GZ6DQQUpq{-*w+mZ9rksjM2CIdB++4C*F|;M
z*Y!~y_I1NlhkgA3(PLjfO7z&*PZB-$^<7kteSII*V_!c^_1HHI5CirNqr`xH!z3|a
z-_S)3*f;c11NIHW)PQ~C05M|UI7*DzH%<~G_KjWCh<#%pHDcd5OpVw#4G<IdO{2tw
zebXc{Vc*n6P1rZ}Q4{t}!_<U*^8hhp-#kjp*f&oSGxp70)Qo*|A2nm&JWS2lw+s*q
z_AR5tf_=**v0&fQMJ?F3^id1;EyL7;ed_?RV&6JStk}0s5-awtUDS$wYag{@-#SdK
z*smQSwWDyIT02Q<zemo}+ICu)CFEDMdueSS@-VGEg*;Ac{{`$AAr9<2CW!<4jx2Rx
z-_c7Q*mn$52lgG~)Pa5H2ytTHIZ2$@cV?**`_5kK#J+QwI<fB@r%vp<Mu-dhu1Vs;
zzAH;z*mw0(7xrDl)P;T5ICWv)Jwn{rcTW;G_T5?P#=g6ky0Pybrf%%J$Eh3po)O}~
zzGsqnu<yxI5B5F1)PsG`F!f;HGfq9&_l^)R_PvwDi+yjFda>{ArC#iNhp89)-f`;1
zzHfy1u<x5BKJ5Fl)Q5dvFZE&HH%xul_l;8@_WdKokA43n@nhegrGD)Dd#NA${$c9J
zzJHwhu^$*A0qh4RNdWtSEDc~k&`Sf@4-C@)_5<TIfc@YI31UAuNrKo9W@!-n!Co4~
zesGuuu^$|#LF_Xlguy;DNf_)iS;}Ca>7@+znPJLcpBbkN_Sq4_VxOHPEcV$fWwFop
zQWpE{FlDjNj#C!<4HKl{->{0-&_)}w6js@CwBb?YA=>aY@_E`Yj@(+>u(P!I`p^Uk
zVL#MHL)Z`HXbAhEAsWJd=sXQ!Kh#<p!hU#ygs~rPqhaiab2N<o@DL4SKYX5su^(<N
z4P!qtK_b|Xw9yFmBRLwueq@M7upc>3BiN6$mPW82ogh){N84x=`_UYYVm~@Wqu7t0
zr%~)jTT7$Zk4=ym_G4`{hW%KM#;_k7qA~2p&eIt7W38nz?8hfa9Q*M$8pnP-N8{L!
z579XG<L7A{`|;M&IQAPSNF(+e+h`;98*{V~`;9}i5&MnjX(RR<TT2_UpO_#C>?hi2
z0{e*^O<+GUL=)IgoTmxwCt6Dr*iTN7B=(bSG>QFWjwZ369HL3=C(qL)_LHrpN$jU4
zNDBL@Hk!hIDo0b;PYux&_EYC+3j3+n(iHa76C{oObQ?`$Kb@m#?5Brl8vE(<G>!dq
zYiSz$nF*4?ex{9Pu%F4%4E8faG=u%jd78m~rnNLv?5n=N|I@u!?~Uh|Wt+*0W+Ex9
zw5oTEl#G89`@F1-X(OxJh`TUQ#<Y`*+sWo4|8$|ZjCp`u`~V3Q#%<R;&`=?0+mnBX
zC2N;x3%_Q+anph|U++~;n|kA!&17D|FU)-N3xyYMPAV=G{`74N_bKJI3pXhhwY!P*
z?&8An%{KKLEeH2AkLBN|QdYSR)(w@F8X$e8wMbkU7fh~_3no|31(PS^fyq<yz~t$9
zVDe>rF!@S8n0!4SOo2=Qra&nGQ=k`sDU=Dp6e@*a3iU!TMKTeXBBcmSkzNF*SSAKj
ztQ3PO){DWE$RuD&loBu{dI^|PnG{T^QVOP2F9lO3lYuEy%D|NAWnju>axmpeIhb<2
z9885w0j5H!08^n?fLSH0T30y!%9$!<l?F&(Wi66QrUFx?RDr3|tH4yt)L^QWYB1G$
zHJBQi2271o1Exl=0aGi}f~i$%!PM%tVCrN#Fm*~Dm^!@<OubAGre3KBQ?J*9X^<Jf
zG$;*V8uSJ*jWQ#cMx_x<quvOnNoE4mq%?tP(wo3E%gkV!m1Z!_dNY_7nFUOX(gLPM
zZvoRPvw~?=TEVpHtzgzx*V<t&>#X$yF}1fA$x-b9)8TZ0>0lgSI;)*vI-O21os1Jq
zSG5aFm(vBNi*bSJu6BdzcDliIGj1?F)gCZCP7jzK#sj9e+6$)F=>^lvc)|2l`@r-$
zePH?+ADI4XKbU@}A51^v2QyF|05jkWfEi!{U<Rv$U<RE*FoR4GOs1LvlW{U&G7JMI
zTg`&WI$1DTh6S^srlGcYb>HA>2mrASw-qT=69O~j3V|77Ltut$!eEA7VKBpN7|cjb
z1k8vl0%nAbfElfcf*Ey1!Hlv|Fk>|_Fk`M5m@zg6X1pd2X51AAGtS1rY^-Snv(eQE
zW+U4OW}+qmX2O*KGr=anOx7g9OuCX_CfOvIshSj+DOU>26q^DwU6Teg?Mj20X47D1
zYBFGETp2JkYzEBjM7o`<EnM%EGxrkdy~y_w>3v0RupX*?al+G*zkr!DW6r|D&Co4J
zo4@uvZO-NI5v}A@UJchS94-)#!vo@T_&@@V07%FY0*N>xATdV_B;iPaq#P-bj3Wb*
zbL2n@jsmEPQ*|}`Jmsi>)EqUChNA(}a<o7?jt)rA(E}Mc1|TEH2xQ`zfXo~-kcDFb
zvU03HQsH|yf6;m2*7xE?=T9$rf3a`1>e4l*{z1!`Wh<8%3wJ4$Gb>iEFc$8C|NZHE
zQqB{(_Ofzjm0=ZJgsiJxx4HO*si~=f6VGikZ<|~ABn~8Bn^t}MZw&bZ<C3|TR1|*0
zUbwR&|2bZ`|GxM+wh(Dc@xu5|H*OsEU;f#9`M1BHl-*BO7LUg)Da#`7CS|*k_mZ-`
zMK1h%^7mj|xS_xJepO9XIrzPns!4^YZc-y^nly;oCM}|_Nr$L!(jyw042Z@iBciFv
zglKLuBU+j)h}I@6VndEJyasDOp&SVzhjS#19LbRgax_Px$gv!WA;)thj@+0djmU`{
zNgyY4B#E5LkrZ+|N7Bfd9LW^<*k3!Cq3Q3~^9Rfd^9pZa#}!T;`2*&{f4L~`z@N_h
zJ~6U0+jd`cU;fpL1dAkv`Pcu7r+qVH-lF^SPdAf=&E)dJadhhX*-LD?+n&e|EEX)5
z6mInS_k51BOR(sR{L{<K%c=|C8<Cs01}~i2q7gRgzwlscbKyZriRV8bE^GroJ}hN+
z3r~LH9_jpk2J7+ryn$ee_s7AKsn&ncSA3R*hqvFf_0~LEx7bR{rcIy0nK`?B&PDU)
zn&<1aSIwVg)+|_b@g<8dUAu0*a>K?dm2e^N+EqWh`ttwp*^0}SE?aTM$}3l|spRrS
bVu?&~&2sq?fmFA!!f3gP8aAm-<mdkf;QmIa

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp.095F638AA48F9495.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp.095F638AA48F9495.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d85338eaacd087c7efd5f3c17452bf355b982902
GIT binary patch
literal 19404
zcmbuGc~lg~*1+ki>TVp`v4sJI1{HU<h=7i0tBC??+@eMd2IG=MqtTdX;?o!tMO;zA
z6+}T)6f}s&9T5Z(Sp{)J<4V946%|}j-&Z(O^iBTx&O6CTrugOF+PZI5->SMbtbbtO
z$pID?eZJ~HYwFiOjw2Qp7ES4YGk%;tDTueQm}6&QF>BN5k)Ka<^fNrKdADr#-H+#N
z=`JN0&+jG)_Wpf#wVga{#JXF{o%cpNL`)gdZ`8xzZd5#xE=ltFLE8ey?JfDwcRhQ)
zI~5-lkrB{i?#BJg6Fi!Io$J5GC;HtLUGuqB$*X3Z>XDSTXt348kNSR_)FQWe?a9W^
zx{Y%!tX=IsGsdUCuYK^&XXmZbhCkd8bz<<rJLy~UQUZoG%U4{Y{LfA*9zWRn-N8?-
zmY)bp`0?>p`;I^Nh#NX{!B1ZO{EtOH-81u}p;z#bo?gLjpLspaT-s!&`u)}Cr701C
zK|h~gGV_Nw*Lx3f{3Gbh@Ei9-+s%txQ#H9!$uEV2UZ(Dk{W^5w>aXrE%io{+RhWKF
z?@8@$?4PpZjxOTLOHy_(qif!QjvsaVGI6!{q$?@uh2Ob-`Hkz2wc8%pr*8T0s+>`I
zU&f7kSp4h5Ndv80jeC2s=+7TtTyk<xA2)nY#*Fi6hE_XkZw$YXSo>_n`pT0hii4Xu
zeOi`ZP;+j<z>Po6JyiJJ=c6~*P5rI!=rIq<t_PKUHz(rsqoXrhe_oa~VRU=DN6D%_
zvF5M)ZsP{$W{+%ppy1k~r(Yf3*Jnzfj+a-YcXR#dL0C@v+re=oV{R_ZiS0LQe4I`A
zm9axEZg`b*HEG$zSGx48CvNQ?tE}igY;l(J&OX;y{@8t5oy(&$m$EMIzFD{QsO^+n
z-;6mne&QJQ@&{k|JYA$t{+Zu3_4O3X*hQDl%$#36_Dc5hke-(&mHW)!eEYla()&!?
zHu1%zX+2)-9xJZtR@ik@pC&$k4&6QXSZZmQ7*zE^%GJd5J);KwI)xWIeK~xpd;4wy
zuU1XD`_&AuCK<ioel%!w((&f$>ozs_u`RnJJiC?brG{O+k4NSouyV_uxoD!({ebDd
zQ}*}{eoPv-{_W^=-`ZQ5A+we|d~xJ;+}IzBQf3zwCHJ}EvC}nA9%*|D4>ml?jTsZw
zX7RZ8PdXG@HC|ucwdb99s~@{>{Pz1-jmBgejvPF}_q3lC!goDg`uX}D72ga?h#Gjp
ze`A}(QBS8Vx<7D)pLfEPwWUcb0-HMibS5c0u<tjd@|eYyRWWw6KS-GH<A~#{UR#C?
zT-@sP>|yQd=J)%i@m+jC_j7N%<X)d$2zYvRO33l+jn$+9`YNx^?$4?$A5|{6xgo{r
zVz6I@V_atN#@msN`=99*u*fp*qi+VDj`R4lYfiZ3wgF?JgU4N7?JR9yJhfl{l|!%2
zGz>p;CGxX1wTs$4?)iDl2ZK-hEOf~Ge!`QYET5VpH>aGT*}wg`=biGX`a<_l)=t<E
z`quEQ?pWEjudAQ^{GZHTucbGn(fQaqn}lJ`)@zoQ$Gg+Ivkr7Sq?<FIC!BL=)1u&x
zQg&*;;I1RXtaJ8chj@In^s6?hBSIE-a{6scUfTp+Q}6X*YWSMaeO_P8+kUp$qnVA$
z4sRL#LGkvy|Hj;TH8<_B*WiAua%$%2K6&V{JKAdUL09XiT%U43pNUU?ZJpxiz3EoK
z!T#fpWvr`kpY!TeLE7A2A6k#dzH~ig*e{b-qTm*M?|b;GtaIuQ!|$E(cIY#|Q_uz1
zYpWfLZx5K2e)amv2?u7)8|d!ga(L^RUfyxDm#%FxV``?)5o7f9B_ZWiDG#m>uY4V{
zV3^J}HoNoOZO;Ny`uH`O_`{pevLBA@^tS8$uwj<Yb|0tAefy_F=5LnG#1r%Cetq35
zB<bMd)gBK=<MtO$+dq+f>iXC(o3LZiGMky(e;Ay!i{IB~%E+_dlQ}OPQmU){-McO>
zc{$+mi;moz5u;E2S`oZ-MrOIs8|ybsJske4e#dVfl@;Ln%ExQPtJ!aE{ioT-JGy%I
z`)f|;AC4_wICqZk)BGQPiV7{gHTlyqONzWkHTk`1WyPy^P5xNc_w|g~Q^zl|w6Ne#
z-}XG+BVn{2b@sMUUlHe5q?0+ZpyK*}got>X#IdCv^{72nxk(&vk`K%YQ_&Ah9T^-y
zG($$toh;PrLZj<~lR053P2X%TDE-gl@Ab%omQqRVD@g})!c=Tqr8e>S^txh&8qrcp
zv91(1GbaryJ-6@+Cmh_UM**~yH^li3ZHp;kDs7#nJkCDbDqoKVQ<ZDP<{EJ_Crsso
zp(8`DRX#mui(1f%MmUaW_5O#6eUJ6MAKy6PXDcL8abI0uoA*Br;u{t5A4RtvuSdSL
zg4e|PHSsYgOeG-VsPo>(K`r!X09Cn49Ild9=7g#A8!`B2pU-T^S)s<XlG9js8rz!_
zrZOPoi&14oafhr?bE=XcI3@_4%?VQ()Zb54jy&INg?Oq`h;@ZnF(*uA)W{!g9>4Cr
zUys~rTkaG2J{fIJm`X_FhhHnRI(%V;L|RIqF3{%v4-<c}_HtRj?Vo#FAv-EQk9Fs<
zojGAD<1+gXN-n!;DWf)2<$&mNK=d&uOy%puw)5>;T{vQeFje_j_p#yq4--$XdgNv`
zA##`=4WZ3^Oj<rB9nA?-nbYIJj>^mN33}9@s=Or5FNv=?VJg2kRP4Hzn|)o6WU5j_
z<QigcPMFHvrz4A7`yYOyM+2$Kb>eWHv^FP9Wr^jNi-!9Bvf2viXj6-^u9zC4DPbz%
ztFnJ-QLx^qM@?ucwM4F^6`2yIvNCAPx8siGp3|c~RHd4<s3xA~gsH5$&ifVoRzgep
zfU5jO?EWI{%n4Ijn{lIWOvR{TD<t?>sG+<*l(#b{OeOlIp{~cNXFF_>LYq6>ak$g_
zA12=D@6_07(})dv<Vh=eNaTmq089x}+4QaRxvgGJSJ@&PT0)RxP`md(4dRuvZjT7b
zi_)VGw1jfvSWZWsDPbx}Q^z0bEuIUuLbkMoJgm#ZHs*w>B!B(wxoJTkd+Jd?TFMjB
z@(F#zri7{N^N5l*j=o)EiyUbw!yJb>z5i(tKl|*HM`5=wTBFufyiAmriHbR4DyjWO
z)<p)taF$U!`nX~ReXP*boG_KN`>O+zmEO^M)SI^B1#x~s&B&B6m18^8Z@sRv&lZpq
zt!N%^HIKJ8Crst|tnh1n=Fh!ihXh(mD6tGBvN>TYr|u2+Zhya|xgPmbl{(^FM-9M~
zFqNX@!#3d=H(JZ66|HE6NLGmU=7g!78PfdexX^RCdgMk+sUY?hbmBB6Or>-zf9=WU
z#7@@8g_cqw$Q5+@H6;xyQ@h>5`&W*)Le^BJx30JK`=18!_f4j+T57vPKux_Y)E=bA
zSMPr;8X>MP{Ya?X*{#MhX+?KB@eZP`xIr9l&{1PbSQ(dl4@|6V_eZWRYDTO0#PJiS
z_dgBdU7!C_T+mFlMGjOv)N!cO`yVF0vi{U`r%I{73RzKc3D%X+F=tAc%GJA056Ua{
zf2~J>v<=UR^K&}!ni8gRE!{C@Uh_`n0upE`^N7_vVq;F2%Jt(9a-Y2I8gGm2smc(?
zAx`gqnD}Pfu!nc&M%=POmQ-AXbwzYWH6=`?yy-GIYD+|SYt)je1e0K&_diTr5pf`O
zif!B<wx}%?4{{#V;r$O2SDtE*`+j<*sTCqr9H0xZdH=)2x4!D0<MU=vwH|#;J9&jT
zT%n&1Q^Hj4pZ;uD(2)E4Wn`cwq>6H?=xk1yO4Xj}KQ}6xKu4^emhz0q&*=1IN|?%%
zC4EB|ML8qWND1ThVSH27lrWVya{_yG-(S>2k2=y)?hyG72{9*3<!#Gh-+b`Zgadly
zN>v^a`$xppoHVEe%YQ!kv}XtF_s+pCLEa^_R!s?0nf~>J8CO-!8$pu#JJJ76TCq{4
zGy5d${i0Q<Q*!yAy#+}<#VMz_j_jRCYA&bbLSByy>rn?y9)%21kY6LiYshuTP^U3-
zeu$)QAo>kjizM|Jk&cm;tc8+#lDMBFy%_7vE)r|8kk+}S8jzCiD5)#BZY#JET8pB&
z?$KP3#tEF9z`1Lj&3R>W12it;dK7V=X?%|Bb`F+zgX?|+mRH5eRj|CJ$a^Uoq~%|Y
zd{?6(8pk3z7J>S=A@6Nae;4Yu3xW0@L_P=6U@iYK)cqI&{mDae9xSf}$t92<BJYQ=
z{O8E`Ic(2-EYHWFJ&SRV#TfK|CGNHogZ9K=pBN0<6OX&cW6+*tEGNVA_G5X!#*Mv}
z#7th-=?XQiAuWcqWz&VEe$wKTwrq}*)S)egw!Lhg(j;|Qi(zfq$4gR&w;0}*eUv42
zgzJco%;RK^GN&`U$ED_pm$}E(&SHr^mb74YB&nH1$|TMjUnEW!i6>*7*@<R$X7JCg
z+Vfv&GbgE+1m{bF+gtPSW8O8dGyBx6YDqmoTuzXVtVD^9U&}|tjty-|y}%nT@U7T{
zA*u5P!+fEEZOuM89keM#pQ1HgXZFll)0&n2VfiNUus0#<eDvvj&4h@W&KuMDuB-;4
z#tZiGLNoR;C+bb1^-ZA-dwMPO^Zk}u7M0UJ9^{pSd<T}%<>>Tp9?Vl0afAx&4)P96
z(5lyg)Tdp~=>B!pY;{39yPoLR6MOd5bmlb%>#SS(_L)yk_SIuaUC2obIcw8m%ECfY
z(>N&&@+~CYLJe}yuikWbMsrUzrF{^*4hr7v1=jP(gF>*zhXl_<LPy9$4++CHJ}kH%
z7Th8CKP(K;_=w<kMDTz-;D|6t<1E1|3)Y_{jLd@dpAcM6!17NB{ZGL1vjw+oSbnxJ
zARCrnD5!-(Pp!TpK`nxORtP*R^kZ^1hLG<4YE3iWAlINy%!BQ8J1#{Tn=_5R=*67t
zV$O|COVq|V*L27WkoyAUp~>$c*E^^|-@9MsGQlCZ3nsLU4uygvYX+fhbSQ#+R&Y40
zF<U)~y7#L_t(ljw<t5ydEv6)DFP7IxQ)W;SU4!Va6MMEik*Ig2zd`Jo<3&Goyx|n@
z%q&_`pYn#M8vB}8rPMH!i9VUM(G1hHL}i9)%pGTpoBW3#HOzZNx<^9T0#Q<nd8cB2
zidMsG-s!c*U3*9M-v8RA$-jQIem;waUC;jgSqn+s$SE7SF6<ph>JHAhgBzxC4yWWm
zc^;?aLB7B#7a*5$N*Uz8IOQ+M4>{vQSbjKC!V#!{B~n&GUWW|p5O6M{ks%uLZlvsn
z`iGHn81iwX9EY5bjQI%kuK*bfAeSOzDdcLTRKxsFk@6JsD`a>D+w%q)-auZ4m1P+8
zFA^(}kXK`6HRMfL+4Ohrz6pLhw_#-)l<&sI-5B&I1shW!AH>Fk8h2~{q~Kd&+!Je}
zW{6UTsArxmQICt#ambmXL#C)`@+{FI3-Srk;RNJt(J@<e()3S?)+a@S#yO&Wj_9EA
zDbfBE<Xln76`eJCo~YzO&KJG%MQ=@BAo>)Doi#2Ly$YecNc1U!^3$T{X|bcGUo3hS
zLp~!)XSDvS`W`381h$ziNoqJJhjTvKh*;0b>mhI9<Smf1I5`XQHBP<;`7S5lg}fNa
zixDV)Ka%%D&Ovex<SHarL5{|9GzR&1VtJ>=Y$J@$O6!Pz9sFFHcE#AN)c4$`?i)u1
z%=oRo$G+iQPp!uhoE*V9LiUc}RE<}0@+v4_&B?1FujhQ%!}4M{IR?sOIPVx(-X>1o
z1m&@u91D31=eq@#x1ICY4(FIe&La_W66cWwIhpfJ=8RhVlexah+)$1G;B5ZjoHfqk
ze6wJEIh>pW+jENZI>mL<^iOdgp5i{%IG6Lvh4NhP!(1rO=j41?UI8Z;KrZIw;=i+Z
zF??PnoL31f{|qOef%>JKTnhODCtrYkiIXouzRby&Az$NsufhJ8b8<P9mvi3buss!=
zTmj{koLmX{F6Vm}miK`3cmUteBhKRy<Z8~N8gdQiSp(l+4cE5@zQ328%}e<H79-!q
z2-tHtlEV?SXCy~Jjzn@K<h4j%3wb?qT@Od#dZeyLf!h0yLatFz9);8>DBpnO4Y0fz
zB*#FGLmqMP8xw~H#v!nL+KdKnh8&MP;-UT)<go?vb|i0y<tHLJ5poielOQJ}Ir;DG
zos7VEPeF1Dl<z_E9?1KV?|%5a4<q?7l&2y&74lIeABCKO<cz<wcLuCK6UmuSo`vKr
z$T`S22iBjDJo4f5C`2BGkc*H<5#%$-;|%1p$m1;JbI9Wy<TB)0hQNGMhWeHvFrQpS
zHdhgtPpXh_6>MKMlB;3+Ymi(6xfaQ_kY6JCCFIx0^)<A&*GPR0?X3>E)<Jn4QtP1n
zEt22D^4=l&9pq5#5ej{dP&_ac`XmeRzy*-QutylwUx+;xLSBmHr5KFArPzBZ2KE(>
z<!~rpj^*W$Be5I_c@37=KwgXGwUDE+Z!`wue-oBBL3u2eV<E?3IS%qxEN}fgdvAs9
zO~7&jlyArKcE~%i?@n0%ZtSreK94=vV-MuL*kdo`1K8sL<U`ow5ah$y<1pkj?3spv
zJ*MHlX&BgJCbr4M4fe>+!HCX$`lT(HO7Xx_9L)ADB=st`yoznL1=L@}@-Je`&QREC
z8Qoej9hI?y*cln!_%I!iF@dE1i7o%c4qAyfiRDe=!1ny;ULxNqm+!4Dh-&ywH5v~Z
zVzkc-h#pF}Aui+AmvK8*(C|h9=~KU!S5iT$*eF$O!u-BZ2V1wlyGrF-OKON;NH1;9
z9_=?8&=-DR02<>r2GCf)v4F<;jRQ2^Z#<x7#BLdJV#`44R}s5#$jgb{a>y%)-3rJn
ziQP)b5yUP6awM^fg#4T|eok!ICcQ+DR2sh^Hte)Tq9=8YUlJR3h$qn<-Nvto4Lg9A
z=n-n;*TjacV<mc0(zuS;utOAyZYDK;Lu}Y?lcZJ&t*V3$?Ce8Q9|)};Kz=B+dI<TE
z(CQK7YN1s%<i|p*$B>`UEiTCGMCWy4Yj$Eo_kTp^^^l`P=P1b0qH{Fl4Wjb~$T6aG
z4CIZX^G1z>#;prZ-g2cIwV`vO%{kGI+0bV;xt+$QG;2$D3X?fGnQPd=kklGZu7O;J
z<T7oCgKj)wISn`Hvvoh+2-r;Yo5|;zsT32rn2ctuGanZ-U!5Hn(~X$=vtn&HT_yUf
z#Gd*5A?ik(t^D2Z{`G#r=)-<fLfW-yO?y0#IL{-ltRZx-RA?3}v|;;%bPq*neO>6t
z_6a3*xoB7}wq);&jZYh$`S@CE`bpB{BvDw8CAEMU3gG5`2{DvFK2HqiA>SfRZ_#K3
z^MQ$^-X%@%5@+^^sj-r#Rm7Pc#Y*a4Uf#<)X){X}-z1AyG(O22PVy}^F5q1Xcu$Q_
z^Dd_$pW$83Kz_^{9>e_4d6(x<{)%^b1^Er{@&@uHp~)pdVS8R*RNm=tA^nubOKK=*
z8_IRoX09Wg;}PinzvGPWxRIKCpJ2OB=*$X|)C|EfL+Hks#SPF&+Y~gw-Uy2ppf*+?
zD!~4?q+Z097qK;alT(6o)q8`A+R#zbvvE&%rpAH^=(2Mg(Qngs&voXI0%k#MHlm8V
zh<+Do%Pc@=4lH1bEU<toZYTQf#8y)@MHw(fwtq(zqli9=G}9DKu?9?$1sPzSIkbQ^
zp6(1{%WK$<^--dyy_OZkjvWLqoharOq`MraeLjyJ&f^c6V)(~6>|w-zyAGDrLAHaO
z*g2%626qhx--S_KM*;f0>*s(%{6oM&(`f(EfWGkm0?-)$F@VPUj|DW&e;lCk{^J2n
z&`+?{tYe~nBA`k7Nq{EnCj*+Ip8{yAek!18`e}f^*MARax_&yK8TuK3e$xL0Xr_K9
zpjrA^fM)Aw1Dd0s1Lzn1FMz^0>oCqi>)~R~dNJgqob^%2>6~>s<ax+?9s>F2BkTE)
zLy>hT<ORrj0pv<#T?zRPvc3a(F|l0?zuHTP?GngKiS1Iz%ZSS|(t#Oqy%*sU4tY6o
zSq^yxaajR*C2?5^IfA%EK#n9Xk&tVMWesV<hMPoVYAkC>6BceG(L-#@XQT-`-<Ih4
zzu`Gipx0q|K@{k97+w;Ejq`f1!|;kI(CaX~CJOXA40S|-UWefgQ8ce3mA6jS&Xwpy
zFSy)=Ucx=W<sRhwg3Eo#Rf1=g;0wJL&j*kn3Z4%kKN372L9P}&t06xYJRd`TB6vQ5
zyjpZxt$7uax<+(b19`3Jv=;I@vBNsCBl}|3dpRA}Lyi(VL_v-gJ48d?Aa>XQIY#Ud
z19_v^VWY<Ez0%oo2hs0<)2b<$ip`ec_x&$^GhzCh|M@tEMPbnzwi5kT&9rsqSS(gU
zWVigi#-Ky#bXtAK1;67)voXvT81z7RBhhap&9%XhL8J_=#QK9&uM^}$O@4~FpCa8g
zE+S4v#DlTU9O}hdz(T#~7r_+g#lnyz8kf;541Uhz1^0Nthxt8gPd8n%*z?ZswCTx$
zoGi3tpVxKeZkZouB;2QhG_hHl*oHZg>t2i<9ood{0(Ggvc+W82PxCCo_yJ-3NR8u!
zW^vHPiW9oT3B6g>>+}01o?G;*2YtNpq9I;vp;^u$(Qrs?!M^b9Je+o}K0;0V*p=Yk
zCD@;Z|I{CMyJix1_KcXvM4i-lqK}78>Nz5xgHEa`q>VjHb|5J0-iNYj^)2Sy7sKtl
z2reK3Zr{al0Wpx%Irns^U(5v*LwOlzETcik=9nN!E$0Htp?p3Hnvce7@&(9q0rJy0
z83iQ6?V1b}kOBD)a=!z&cb4LSrEodD8XH$*Q2$mOuvL?9GKcQb7ac?NF))55ok;1>
zh0P)EIWT_36yV36XY8vz3oqG+JL@Dhl2an#5yE!PxSa#>BNsX4B9uSij1QoEJ5sj8
z4WS)K*#Y?oQjS1AgN$cjJjg1ntkRAcB=rC`9>5Lt$KP^|{cFakY1CpK5a|I4VJ0N0
zrMy$AcAhDzZ+WM;8gF&ZlCNI<@eMWQBOLv=zxIwY1nRf;XQOJ{m9FdV-&?zl)^M9h
zx5?+4+h52#74qL|gXabB^g`o=K3AN3pLo6T{~atWno`i2V<<KAEyP_4aWA%~A<@vZ
zE=9N(i&~TDCQ6snxEI@1p=Y|dOEK=noK1;_)O9Jrz1VQuIr3~z-QjONEhROAQzE!d
zTAyP$B?fXarxZgj=ah2D$w)~?Ab$o@G9WL-%2EvETd}fLV-_4rA58+$CqR$g6e7y(
zFnLxQ|H;&?fnjt`Tg54>xYk;eS8;Bu=wU#^P+ZL^tD$@~=eZiLT{dybCMe&;xov{^
zV>u-j%40dtSeQSNGbVDtdrss66FCrgl*Ad6pgf5SOoH-!PRWPm=W}lPu>O3`FQ5BJ
z>rXz{FQ5BF;{r}8fcXnJw*r{Ifb%PW`3tyy1u*{^PB{bfpW)oj!2G40QVQjzoM$QQ
z|0Pbj1m%}Fw@WbpCC={>tp75nT!#9WIk(GD|1#%y8R}PXN(I!f;M^*pekG?=LU|?U
zSqa<sh%-Kd&;Jn@_y|7#YR*^<<<(qZHIzpnB?1Baia>4=2-sHy@{2&={Y0RC5eU4W
zNTfu<{E^5l66TLYevvSLB<dFl^KU@P2AF>Xa@zp&Z$N$<VEzrL-v*dJ1}QNxe++Vq
zf%#*Q8iV{bdx}AWWBz`!V~j_@zT?rrc=V~JzXch$K=~FFxCI6RCL$#f)}M&n5@G#G
zNJ)b7B;=L^<ta!>f$|jOmICE_kg^BL_aM(bsI%6-RHURrc`9;Ch53&n<tUUNMQ%r-
zJQFFIP@ak0GNC*RDOpgSg*>xh{e{R_2%k?O3M_=rrwAE~pu7kL7D4%0WIPMyXHno;
zC_jgc=b-!?3OonpHAty}?W;j<HL(3P$gc+4PYvo<1MR04DYY<vEpn@c`D>A1EzDnw
z`qjexZ;|pA=6{Ra-opHEk>6XG|1IkG7UqA4ly@-yJLL8b=6{FOchG*{p~3H<{f1#<
z7zXwmh6jdWV808oaUqm1#DNQ8;Al8j!Z8^C;n*!4gYmx{E6bsLIrdzRf&Hz)${HwN
zgWcA^{A;k^8Vu%-wOCmT_19vzwNQU8_FD_}W3dto^<%MHEYy#~N*t8OVYfIaPryn7
zlqX=f1SsE*mF-Zz9eZwv{o8|$d*Jiig9G=#=eHLd_d@wz9Jm+C4`JgWC_jV)4?+21
zY&;C*hjHLxD6hhnRaj=*U)0~jmJcA;V9OfFwb-&&<J~^pKFNt48%14?dA!p+zP+~G
zIv^SjXiJwp(-%+P(Bsou>iXvLhFrd#=HKP<hCImmydfWQ5pO7hT*4blAlLAQ8pyT0
zp%(Hp-tY|aOWyDjavg7|)0jmd(=}H8-5JekEXKWyvA^a79w)BHwKdj0^R*hf32VAj
z!!jPM`uU*!vZuTLb=2&kF8iYAls@Q3<8sqPn>6h(gDoRz)uy{T>>aQT9Qty068%oN
zF>AW2!~8*(fu0|l?&+|`OSCiXmx&*+-bv~qvE?DLJzJ30v+E(zi}A6%3&Fd)*WCKo
zon=c`+9%VsAlB6Eq))zU9B3K#|GNKd_lYX*Ci>mlj)Ts8eTbEqdoN~Q?A<P<GEp-{
zDN_p|BWji?WkEh6N+%#^i&8e^lcIDIa*inFKt3f(ry%EwQZD2?QObjyFG~543q+{^
za-k>{LM{@eBFLvj=``eGQ7VR9B1$D17bvs-YBWpFi$qN#`Xnv-mHfke$UpY^$v-Zd
zkbgwal7GAb@{bsIw%tKrw(0g3GrtnG-7mw2IIaKJ=ES!B=?Z=m(Qnei>m~I#k&bIC
zCAz*PPNnc>fa#VPGZwbzNSkW98^)SiYQ7srBbMtghOz%Gsecgp5A6Vhp8fC+p}Y&*
zrmCL}+voBf;cVDG4{{CPz6NqF-@aC3Ru4VmslRa6zs=0>RY&{$Dye-e`+URNRC@8J
zk?t)|{YFoo3UTv7>;YH4&5N)HT=_OXjXmJXw|O!4&?dbS+`I&Pu$6D=?Gw|E_Bww@
zPt*xs6B^c~M4jk05zwSglNvtwM4jAeGN375Q-J1Fuc?5h8K-q+jwn&TH+~OjdiUuK
z;U+|#(R~J>pS*u+2&N?JOz)Y1X8F!)a4v{C+jn*Yu^@3eVMG&swD!()<~w@K2F~8O
zer)fxgCA4HP|g_2frFWSf^nY~yno?kVMWublrGfjmLp|3@?@Q%*O!p80`g;|Jcj%P
zDNi&myZ&svwQu__bh%!Nl~UZ9#dDP1S`xJ5oGVA0@@t!4LHy!;)M36BpTut5P~)t>
zQpk*^{_r~}8Qy}b;eu))m!Y6C1a1kX;h;2lYp4kKF2eqr`J@xqbnTnSuJ+K+O8r$u
zt=iML_i5~}Sy={g&Cs$*^d?J4I{yV56U^G^7uj@|Q7h;=w!97x`6`KJC9z}ofY^Zs
zeL8=i&~)Y-kL-<JYhIE1tg3x6Rm|qwWb@rw)skAxx2e{cWuOnzbT3kC%HKaz_2ayC
zDDSR~^Id}5F6|zOL~r95ZwkG%Q}*XVljnjxn_ud8f?LLm9oW>!Zcth3-e0uTs!isG
zC3C<Xui=K(KrTbW%HVP{4G&9$?t3BbU5Nd)3E>!VJ@)rk;gimHX3vAYX&MG(y3VOJ
z<O;UD0>gmHiDfx)U|Xk>n#Xs_)53n}nH1ltR%13Y=sT{z3#t{Hhg;;~!E7(DemZsh
zgY;+aL3!7bJ};7+asT=Q!OnT<yk39Vl+|C^Y`aI~?Is`45qyO=T;ZFs2SEd=c*9l5
ziv`1C0mzpKh9w%a3#)V})ZbQR<<nw`{tB^Y7gi-Thd1PC7g8kk32%6!@hvaDR%~J)
zP7QUr*l4+E!|a1@;EIh7h&Id-V+B&hwM4%b#*CZp!7@v}<B_`C$NBfIRPhq8T;e-v
zJq#0+Frk6jsRw;RrrWZt59}C`mXJvFiSYQzbXk^F#Y(0nn6Ah&xAW<y)Z-oZy0rP%
zZ?Q(c_h7-&A-@0f0oU`nr+&4o%WCq!XI#|(NSlIv&uG*C-Bqa<b$gQS?@wJpztJU}
zQUWjE#vx-I0w0;p$haADC^m*-AYXuu3pB2)+i|?()%CUX+i@9NUdAqLsmC_l=?;7x
z(Z`WdY*0yRK9TcD2xFc3x-xr@Z^<rcpY@?wTH+#}EaDrpTY2oVBONVUh<*z^7db(s
z6PgQ7y&mG8OL}NEm@Yjx)L_1DNiD^6$C=sM^slFk^QYDMCtLkxsQzsv$JRb9yvc$U
z!Gc;WSXQ=DwP02^WBQjfBI=vom9^i(qMP=opY~@L?N6ijXAkYqj~f0&>d$qJYC1l-
j{j=4FANA_$*Q0y4o?VO<-5V4vdTM`usQuYX`}6+**Tnv4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp.03A5D53DEE4E2511.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp.03A5D53DEE4E2511.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1e6eaa666d160cbcbe949c0b72be3bc51d425974
GIT binary patch
literal 18908
zcmbuGdst1``@makuf0!>?X=Y?X%|XJPITk66~-+ZGsfMxU&b(oFw7XfW(*C=EehQf
zN|eeZ_Yf+%L=rM+2xBTS8o4H>-%4k--}(OYd%n-i^UPU&yzjd0wf0)?de^&$1^D}4
z4=^+9J1t;Q;FsS{A!cS~&FTLOzMVgF9B*b8WMgKwC^m2OthpuK^u^`xwLexY+q<Ac
zud4^U9>a&P1>Yarq%0v|sGhiabepnTKQMpr*e9o|hIQE#{-<qVjIi46?t-U%M@J4=
z_q@93$lytL+ON1Db@YpctE~SvyI42&>CU-#+~x-)^DhryS)5|1nAWse(CMIUF?%Ks
zaBlxeV6t<0ljfffz3h2o&z6N<{)`@(Gx7HGzoR}4xRkfI^RAqU2fS}Ao!hqOxguSk
zvUdr+9?UM-Ibz)1c^TPGvF)69_#{m7ib-wa8)o<@>5j#y_dUOK%<P@-;P3Z&LB+2p
z9O6Rx!CgYmn=3aDeN-wevGJK%Qn7md!ugM{`5#U_@LKuy_W6;$3umXNy5kp?>1)R=
znzQ7_{T654r}0ars0*{4dK&ob`z{~va(d8XO4CD|zq;x_e8lFGz}A0smnKixxS+XZ
zSnYSOY6|+9aRG_m=NE30-)(RnGCsD*XU1#Yff*=t-?G}mJM*`1;(E>)R{TTC@t<yQ
zU(m|>-RpS!755*@?f&u(Uf;vhFC~4cb<M4zzZ9J8Sz3N1HNDG|j9$O}I<eM$$>>Q-
z;<%o*wYMLa{pfI`wz7O)`Wl}Yw`-3dhuB(KOj_yDVY~PFznAWh|FLcA&D#fJrnm9D
zny}~ZORr|xI`%rZ{*&{+ZeJQ&eqr#DEw>8NzRRor;g1u)J!}6}=JAPmRq>Uda=#vU
zYRa9xZx^odwOw-X<R447Upn$z)#x`7$A^~$F4$Fh_*QA)UDw40cW3(ZtLL1ZZM(jA
zO2()6d`G<X=`ei2tN+~8PL0f*w5jS|@$~o7lB9_Tf9s*0{oC1Rx0l3T+cwAHYRPS>
z)vdk9y<bVSLyp88yFGuyu<PvulOpUst%-L1dc=Z*OHa-ha*M|G&h>KQnyejG7n}Vm
z<;=xTABtz3il>&%e$;nU;*4`G4|TexueqF@5IWn}+}bYw-I8HH)Sg$`&a#uTDp$?z
z7Jss?xvSlpRr*2R=K2XUVpg3s1XwT1+F22fE%b{sMo3Ssjs>;*YUc6<-(Dz6G+(;S
zXKsM~XWf_D6|P>k@+yCM^N!i~D_jqaowdemj_>cGUQfIKzRgQ`Gy(l+m@zYaUi`q%
zyG}Fo3c9+gZL1F6`ClJDja&KoeVLT+75M!|ue?rEUrfBxuh-Vzx4!T;$W2~8?)l|A
z&-Ix{?`?kc*X>ae)5k~dOW*W%U()m@*H^YJFWs!zy0z)0eVP0Dp`q(WeH2sTA0Zzo
zdhAr8yy$hd$^4+hz1J-tec)l|XP??;FPXJ7sq2mH9%-jneIDw?&3L=X`lyw6PU);~
zH){Jwt=Rp%I_q{-@U6)&YQI0-Yk|*C-&D^}Tl!7Qycd<O2iuIlX1>aHW!~=a^PYol
zX1H6QU*7f6l5IyGdOVA;4F3G;p~%0QZFJ8q+M9T_>Es3TSC<cIKW(yewM)|kZ|8~o
z@Mn9Z*i@&KIa%*U4Ru;n{n|6&sqYWEo#sh~&+iToda!uq%EXV-E8BN&<!(LJGOO;X
z@r$nO?m7jm4SvX%Sa}?~Y;khn*bb!{mjv6WAG#SGV;%=COAY+X%=(JY>Pw!-ri==S
zIAL~S>i0MD`^?TndroCFc~M-6QVjddORGn`zx2b~m0>?t7?rW7Q^y|(>R`EdNy3uP
zqA$J|6K#)WuGsla;KRiUei<)pdjHigd-ZV3mx<S2lJfjPyKbI+o8`M=l7+jr51R4h
zji%4})}i|@>(&k4v%e-A_0RjJI9?2UcWP|Kv58+d+i?C+pC-C5XG+mmw?_E4u?;Gm
zSGM%rxTo#jeR}d}-Mg}!&pZ+yTI74iV%<+WyFC5k`@R+B0T<t2pP%k|andi_DhoGt
z`KfZ=s40Fw9rwQ4KG~=4bn%H}t`;6|_a6-!Kk9TH@_Icm$!x*mz^Q+0=wjRXN4tPI
z>v<jS*u_jaM_QaCE~doHoQmmx!t?>ndL7@Y)nON^a+!3!Oa_<|rlOrQ$KQVWC$n|9
z169c*c6p?;DPbxWYl`ZMd%y0c!yi$VTg2fO@irw)Md}xSG&^L5w>7rxY^IE|8`b{9
z4->cg&*WskqNQOHZs%^MRP(xOzPTx3Dy`?V&pRp>OcJn$mhgmY`h>GGB}}DF%}M8v
zgVqd?upL!7AT&83*qaij(zf%oz^B7J+F4_3sxriGi2a8jChqXdy>+=$gC0xRiHcWi
z99L_+ObJu*>d}1Uh~t5Gtg$Us8E!Y+{=-kBc%jFCTq+*D(&0|@S)3&{XGwcg!c@F}
z9(qjw{>Ka*?nhN_5{H|_%ak;#lqOUz4*F_bQ>>+xoaD48Ia^c0RJt~~hFTt56(wOS
zsuHS^LNyMiq*3LJec|fre5oncQ<ZE^o6Xsn5~k9-J$FjK9a-pbAKI2X#NiHcGbN2G
z--sig_3<0i6gQ!zoZz%4IN6jil|GlIy;z;n=VVjNQI(^d_9(5WAz><gPj5Pw9CGVo
z!w8oVT^VVvG$c%AP;~qLt2&kc-V`^br3};#wEFPF#KXM;?{{%J{fiENLYsP(w7E*U
zm=dNkF8*rqt!;ik>F^+`l0&R=h`lLcDigU^wn=s+nN4wXT1j7RU&{|aOgu^Z=e_J@
zUdcM#o0d>SS{Kph(U35eFKnG=&1xGU>2Ncu@+*;kB`r(|Q<>hTOh5nI^F)U~rYgCl
zWiDxJN*Yz3{a)hy9DUss6WWeJ+ClmcKTJF`=e4pSYRZ>796(FRCM~mxlPO^;v-Zw6
z@3Jy*mJT~ml|mvF61^#5Ds#ttSLAai=7u$HK`R+%H_ZOS4-?PpXC|gLajn&14_d-+
z#NjtO5gU?5l|MS~{>v#()?oux$tO}i8E;CM%7VwmvjY95Op|aMTFC=}JP=x#5~lL)
z71wDg1+o2fSVv2_L8Ke>ku@ZZD%YD#ou?;e*4UnwGSY6O(}y1>{!SkMmy=U3jShR#
z5=w|e3Gpx`Oy!54uiW`zsMA7gY(q;JYB$vW!w(a$cwzbTB$s=0b$B2x;WTM^nm*Bn
zgsFtw`quDZ?Y&;s*qN3v*kQ2ChaV<hT^Dr9`IgIEODt1y4R2S&cQqwUWpmE&r8Zf!
z-s^CCTFDt=cZM#O4GB|88yR;0v(EQybhs~7xl0`GQk8~;sbuavxBuy~ADcJD8oC-j
zB)W&h#*{FXlP~vm_R6d5YWZQQtK+3QzOB-bFqP~%pC8%XZT5+#SfI^4&S{U+_qHKn
zDmlK1+mXlCDogB0RZ0b^RFF*xQ#sS-fOl2Qtv{M#o~rn3{jEOyF!4>DUDxeV{r_Ww
zHMBQRIrFESWJ;LItz+FaGb;l|>aZ(SDIih-nP5toN{NT!BdtyA*c6Mjk}OV}#kDXc
zOy!=%i`!HBecV^TlAD>*Q`7T{4?kv2FzRn+h7@$aSHa@xi^U=Vx1?n}AWa_-OH;zi
zsPsQK?b`R-URmPSROP539i{t2L(-@c@W;?0msd|~ip{CY0PO&)4?j#?)$C&U%dCB&
zI^34_<qDCm(08sOVJg+1k672#!!k^VEvQN{k&5YDZ%CNRtD!&leB%74ttD<nD;Yz^
zczyU`;+kDu9@KtUFhz$4&=PJFhud^JZAh3(-PU1;2R-SVt;7CQ<r-;ojjn|a2~(N>
z<@5zpeCZx<vI2;xgb-Z_Y0ZXAR8onUN_QXA5Q@qPVt<0TG1i)F1-9|xFzk_b|8^fi
zRC17<gSxP8h)OP!b0LRfeJFNO<r}bm1LUVz{}gf+)>o;_j1t=7wM4g;v|%k0l{6xz
zkya}2C!O|_9*ngn(}lHIINr65U;kgS>_lZf>b@S0R9h5<e524Pl@pPah&ri!3b~&`
zgH%3`dYng}t6Ye>7sB#NkZ%bruN+C`u)O8ib2<K0&A$eFufd<Gyah{JFsOek_S_2f
z6L9wg4BEdJd+o)8)%*vs??DXua~eyhVR`4UbPn=;?0FxSUx~ddVS8R;=_S;!#XV{<
z=zkE`J%|JC3FW*(InbWXobP51v}Xq=?chNFk~k?zWoBd25nn}gtB4aDQBm1L#655Z
z>?57_kzP!$HCdu;#GAQ?#C*NQql6cgA+3kBW7eLi3~fEM9ory8WmxNB?bzNZD#Ke3
zZ^x`UQ5o5BWS4JEn-Mc?v8Zezx-F!gT5p4u%j#w3t)%^6*0RM)RDw~bU}RLsX#?uN
z0S#3-2KA4DoQXPRLj8QyKOf4kA>%dFSIvJ1^}hq<HE2`~nySicky|ZNR8GYG6EUd&
zFz$aCatZEKf<b%V<NoiVd=+P0#ew=Gxc(7bV|i8s0?O8}dUFVg)XZ#|&3)9>WP#fL
znqmGf>4+yGN(~xOqrN#riRF#4ybm)EiSm+f@shV?2LhrT651XToSA9Q{vBye#{;(P
zx7ICJR-0eCPpjU{%bR%@Rw5fBjn;HpPzR@ByJ89=-D2B5B%RpqE7JYcwT|~>Tdhc&
zVmu`DVoed1e8D1LXu)<BQF$S>c_Fwkdx+`LrZyZ!SjCRV=705MncpfzXj5%A^DS9Z
zoq8S2toFI0{nt;4&3kk0htmzKn695_SQ;9kPKPQqtO{}=9#)8fJe?bs&H+Bo^*_!H
zV3vibB$JNGq!W89L^_yWv3yt7A9k9ePoUu_#Tw#zf7iliX9nmvQCW$^mB{kLVU+%e
z$`K?UfqV;#w{WA}ZQ46u2mE7BbE2#f+}8+=gHz9=*9c?SU=d}l;I>xi0(t0KVYtfc
zgpTWkPLKoE34>H#FSxE343Gz{7d}-vLU50O^+yP!BVhfJLdQs0exwi(3CoWXT%%z5
zQNo}oSpE*dZ->xRtuJ2ii-(*f_$LVincSmn+b{Cuyt#Dh1f!0@$d!#TQ8H1-OvtsU
zQ!O&6@)F#!1UKq?4t%&qXgQ`EN0c2x%N>FptAQx-Ld$r_NkYpcmDvi)X+NBY*-(r0
z*xLUjH^AF;qOG?&I;N^~#>;eGrdCBT(FMcJJ%xxV#9p0Tdx_6pGEn6V(kX-Zs(ggl
zA0Y<DTGOeUz25B9P1o|xM7NnZtBMUr@J2;bOKC?M4&G{4@;UQ-&W0^&A~lfA&k-AD
zVX+fAo$d|Cah8!~u+o}N-%KDXdx^9c_A#AUq!XDnjV?Jve+X`W$B6zI<nyHYdEzk7
zbV?JIi=_EQ>VGkr+#)rInqMXk%%T#N4ZO60w`VV?sO;b^cJQ*wJ9+(1zKzOzc*i}w
zo62dtV;bZP-aUi&Q|0-*J|E`4!aH7p`qy~JYmjg9?l)ony@JJFL1uH#r%z;`18*HI
z{`IU@yjBgno16NwwWw@G@>bN1O@C3@jf}g|FqN~BoDJn?kbDMm5t54_UqkXW$aj%^
z7xE)yd<4t?8OuLoQ2#0{uY$ZD>(^soTW`erjgWU?c^A}A$8tL4BUnBHIU5_ZG3Z|o
zHs(OSfQ=U*|BmI~Vg52KmqGp$>;Hu9`3vj+g8Vm@|AzW2IC%vJ`^(9zAaCU4jsIld
zjnHz8<m5;wkLQf>9OzF1XH0;+hcoU`xqGW;`Ll#6&n$_uNh5C3=-5k2_g;;-8S)lQ
z%PksNl}Bh=MnI0#w2XutrLl|B*sJ={8p~*nUga1~ix^EymA7hIY=s=Fkz+Lusyt33
z$3fnvao?u#ROQ<>UfVTYRo<a--vQ<E8n1XLPtdp}Xu7ESJ2h@QA@9<NyVU;sUhCds
zRzU4Hba_nRofCLp^_{-JJ6}-QPuJtjB!B0{B0ZfUX)W?nU!4dfML>>6Qat1=BxOOq
zg```MA0p`?<T@l#m$q7dDwa|);4Ca<K`zBoDdcsWw2lM%ahw#VGIM>>9!C>hw7T<X
zO%71DA@nvmKzXgn?Wr!`=Q;E9TuZjTi}W;No<~};<BX^z@g7NhAN5gQ<~=T}%-p85
z;09-@TJUkM+i|XsYG|F{x}D(qu$G9_InphQ>%;se^_y(BlUyIR$+8ORPSoIHWmfWl
zYq7pt$M#?FQ~lt1uSGr8hr0nu8;~7j&ke{=<uD|LL3ucm!XZZ>?+93443c7?JO+8j
z!17{|6bt2XNQ#3TkG$hydApEd7hG49kRb_jGBPAXPD5^K@FbOn2Bg80(l5yB7v!M!
zFAI5R!TPe1lnvXHgWPjacU3<JeUyVfRyh~B=R$cd`Y0F5^N^GW%R7stvyk(Vl>blm
z%!kkGB67b7%fE!AOHjW6Nd=ILkW>WuI+CtKE=E!@<Xg!57VQ5WB;A4XJIM16Y)>hY
zN}>E7lI}r%h`b-d@+y#_0*+55GE_o-j0}$<S0T45IDS=VKouOnXUOUq9KSl`T?gy?
z8B0H7FrS056bv~8OCgZgVQC%YP~0&TE+(PaFBJQ$<GlfQ+yLbpu-^tK55rOzEH4~O
z;gBP+Ap-7`5%`k`3_K>0_>)M;QP>a#^`o&N8gd+#;$ZpPv9ulX4lM0}oQS2wf3jyH
z2JicBEbWH!BrGLCPQ~7-7_1-ZSW1WT3@l|p-jAjIkPl<&@ITq}Fs%P5mX1PsCYCZG
zXJPLwSbsJ)WW(o?iw(Ju&tStD$mg-)Jmh?A$cKCp8!kdF#BPPqGg62L6hhBP5w<GA
zU_B|t-led8zhmil*#2@Xl|!z;QU&BESb75aDem|b&bO!7?<t&bRk&jnlviQDDky)3
zrDw3bYAjVleuWLMV1HlXPhP?K_Zoll8uA-#cmwrou%QO>dn~<&<-f<C@8S5?VW|$v
zmvPcE4!o}`IB5msU``5#9KuN<kk@hE>o_nUH*(TOC=cVLFv#JY6b?CplOq1fo)H{)
zAEG!Z3d*B7DH?Jd=S|~g)b$~rGsHu?bth-o2|1B7BtlN+49SpFI715LRL+nJIh}J$
z=YXd!og0v@dg4SSgR{!u8s{UsTNSh)?o?GnBabuB<C?1$Ob#*6A<fxEDs}A>T`p<C
zE;L1D7q8#NJ1_%7R4(!QODYd<(myNkODU9k;X*Y{LNyl5{A8E6)Y#fUbQ{!L2d(J}
zml<0l23vNh2=n9HipqFpd@uERPEaNQny5?!G)b8RXtFXH&=h3~psC7KKrcC)mmKgg
zyy9$LL4M8IyoUUSvv~u#hO?=G{FbwM3%QoFsfB!#G`mTxm?uJ{F%r#g5i90p7U?BV
zv)jaq-T#R6g0oo(v0`_hBE6e!_8YNc*R~?HQ=8o(R?LtU=`~`ryTpp!FpA14!ReIX
zqB<x~3r?pYXA4f*kaGm59LTwXQ!eB)g3}qud4f|O<VS+TBcUyGfQm|`;7|$qvEc9+
z@)N<~3FJQohd&`d6&#*It`Z!oR30^D%b2wIlJ0b}C2OpbH8yOreQtHmWAZPy?dWwv
z8j{jfXP2l{A*l*-A(jg1!e_dW6P0vMO6MAxy#{pF{!DZ~lNQV|qcuepuvt6aZ>v=T
zUoq!jJ1Qm|QD%>C@7$I){R!vrgzLy^5UIc0wwiZl4sek=mD?T=x-c)ZsN5Iy_k}j>
z;j=gg>h`V=a!_CRI3mRnTh@_!r+3ST#8Gv6@8&z~=DVmbcoN?s33377p#bt_zQbje
zCq66>FcZH!PA_1BkaZC1sy^C-$nGFCU|&(|4vkjj!Gd+L(3QD<L}im;w@K*Em<49g
zp$H|qP}OIqHHBxe2Q^8SE(_0>U#Gt13q-s?#;MM+6y83CpQU!Kh_^3Nc`6?B(30yh
zl1`_+g2P_Hl~v6GK4{gOh;9?HW)Dwm3j1JHvsIERt{}P<>QmO5f<GD+O~D^n`yuE<
z9h_XwJeRX%t1$}#q2&b;T@ajp4G|%%JQfN<71t5nI%27|upuslDY8p1qWusEqE?*6
z_0Qr4sOQNP(lJGK($+in$H(#$*`s4C5UnqY=%S!6wIO_j)i-Z_Y~L%sHahxDdo}Cj
z!s=r~LdR$|(XA$9S$jq0AdwD|@r<>mSQAzro89za+Yo5NN@Nixv_;`W7Y@g$A<l%A
zz^?Y_qpuG#VfjU6RL4;s>>7`)xU{rLqKj1TBekZO6IL3#QlxKLeZYxYS{fRa2Jefj
z(5NcNg?LmUyiZN%Mx}G$e)R;`{{%NcJ@4-!9rvi?u{5$z$jNS>h12JEiHMiTc-7&%
zm$%=`&t?N9D%W}Y>ngJ#7}~MrM7JEysD@A&)-e_eLmzZ~91QCi3xT2k+z|i51ZzL$
zR{Xi4&S#jY47MI@&$5fk7@sj<!yW507SK4Kae&768xJnrC-j>DXkx#KfF|{u1ZZ-<
z$$+Nxn*wNRzo~$x>!w?)>&*<^3_vq=GXZ_A`x?+}-E2UCx<Ej4b#npD)6D}kUpF7n
z0^I^Y-|4;sv{1JY&?4O;K#O&Y0WHxj0rZ3J2S6*3<qFhN?O`yo42FCdSssR*i7YcA
zzd@F7AlD$v8pv;v<y*+L$g&plO>B7+gZfLbWeMbJ&bk_!L(e(u=a64;)-NEx<Q!je
zE^I9jl~<hOE6A@o$JdbGaE@;v*Km$Ckl%8SZz0!mj<t~Aan|o({qH&J_mB&Tc_Fc2
zi>*ilf6T8C3l<Y3(kLJEVq(DpfJ7SMW`2WMu%$_)yM_6$#DZ-nB8~af-y||L2lcmz
z49!9PZ6ZT+P+vl1wm8+BgZkfy%ogK%b5MVW$j}_r-zBnY4o2~oQR;0ojrS28PeAi6
zOK{ABd{S^c3Hg-Zc1rMu=A_$c$k~EhHsl<^EeCS0;Fb&djNo<#a-QIp2f0kJFH_Au
zQ7ISf%OO_?_7#vH2`-O>E^LpeH$Pn}AwL#e9z%X2xIBUUr{MA@<fnqmQ^-|<OO?uN
zEBoE}YWn<l|MN}h#_kvP8Alz?qMPC?H0BkWz}{FEJxGnel|;8v-FE6P+Qnpe@0U(G
zrNc0oh8RP(dT%ah9un+Uo=R(2E=bFTHq4aSGTrxW>ya<#(v9OXX?>Y^t2cuge9sJi
zxVmxN<a^#!dCPs*(;uBotfVDw)Yxv+I5S&&%j?M#f-LNdX`9Q4TNzQ*63WP+GBR4_
z=e+H6Xuv+_yFKUoFoQPY!oZ|!tCkw**7QQqzYtoh<G5C%U#n@&q6y+^(pTPWvDQIL
zn}Xzx@Pc3$GVVeks_H6|uR?hRGFCu&9G2st&tw~xw?W>A<$aLPW8--kCAE^1SE?5Z
zbd)$_GS^stf<w0S+i%~#qdhu^{0^c3HqQ07t=}dgfQ?FG-x7yDId8ZAzg^!m`f5+@
zky&o$^c0Qc4ah?+F$T#okn@q85BUy~??6t(av}!#4`cZ-<o8&94>^L9Be+I=wx-Y)
z#}HkN8dRt?g|f1ln6~IRKO`{0e+8X5VMq=`ZPg}+A=fbEsd6}y!=XGJxrM_6L@bhH
zp*$A3#=`t@NRET@IOG-w^Cux=5&{NC67o+%AZ9HY8Iz$r8Tltec^;DUVEK8-H4oOG
zhm<_jTkTIC8kmQMsC*X5XJP)c$n`ADe-<fcVg9pd;8~de5|S^${FjjHC78be$puhe
zfZPgT|F0wYI+R~WuGeAy>qxl{>n}!fG1M<cuEkKl7%9b2zZA)(P`?zpmO}k|NWKT<
z_mJB?*uF|+tc1_M68Trc=l>WPA4B<L<o_7TgRvZp!TbuwuE7}0uVAbMV=z9!cwjIF
z;}e4A5STv%yN1C0Ay^54`9tu)5STv<%V98o7<LVV`NOah2J?sEfnhLzIF`d<{&4IX
z4)cd&zi`}7oloI-a5(hHMqy(V2J<@#e-Z_QxuUT#8p@-we>99C-Hzq$u>S4Xbvvwo
z2bOn0`3~&51Il+}c{h~r#;&`eJPFH5P@aU{l5khGKN(ogfbtCNngR3g$MSwC-;Z7Q
zL-|oGABFOx*!3usXJR=M$}_QBCagaf8*}0F$;JM;@cEp<#xqcU2K%3Z@_cN}hw^;v
zpAY31vGF35U&Q_wp}ZW+<*<F_*tHzCzZ@&&aQ>9zf#q=iRA9LR=C8o66)=AVRw`ir
z3Ouj^=6{CeXE6UW?D`Dme}<K3F#j_=@EOcsjpb^XzZ$z%!~E6QuNuzZYCO0a&fhoK
z_y#_oH~5n`aQ@a{V-1woVE-EU{OYh=2kWoHu63~fWt_Z>1M9;w&TSb7=3g);2Sa%<
z=Ninx^@CG_Ik0|&aB>LL58+%xpneFagh2f;P7Z_mVVr9i)DP$6a3~MwT*IL}ij$+D
zJc@ITg7Roij)wAR&Mg}DZzpHm$$|A@C+EKt#%w2Y#zZJj<opw%JcTo+KzR!1p91Bn
zoG}&3Q#t=sDF2l+|CN)NbBi8-IP;s3OF8pW$oDw&dn)hs>OLedZgM0wIUW%E2c!dw
zX`ntNjefOiAEwU_{yMtHh!@m8+|BEE^X*j&G>O+IK~CoN$&gcdeJbP(UY`NEfY%p5
zzRc?{LoVd?g^;iE`m2y{^7@-9AA0<sLoL=+r~K=1-^1PZM=dFOGLQxoZ`4?AR2|{W
zt4l)}8@?h^pII5PEh8Y%?<L>qCGW+I@e}8Z$L#fe{>Q&=8Z&XJ!>HlwB<3(;p)EA_
zrQw?-CZJK3ocTEzRC9)ypCK*T;i!I8J$9=>G@??#dlab5zG<R0Hhj>;3eHU)GN+lp
z`HKJ7%4VNDQ738r_f9OoNWT`-ohB`q4^31Od3~aK_eFy*dHqF|&pnD+7FW@&NFvG>
zjkrY(wIWJ{MvQ<QsSzU~M`^?;$k7@x8gh(AjDfsWBW{Hps}W-%$7#ek$lEmHHptsG
z;&#Y8G~y1(@ftB6a)L%ofV@*9?u4AE5ffF;mlr)~vPj2kXw2O|B5&z~`Nvlk^uZ8a
zg!+{Q`7iG)`7duX`LDn&@?VFMLcjJu4ExMJ;$P3=q8s0h?$qY&PxO#`oNINQGqBl9
zeJEV36P!VH)Mjz5vN(h4NI%K7I>{MWU{+CC&fH_Yu9xeGGTnW8<8g{8Gu&qYn&~mK
zao-`z*B)O3n(aOtXa>3m0-9@_>%;6RqRcbS12o@ve&cO6Q5N_v0Q8;bca33kL|N#$
z5YQs;MU8tTQ5JhIZltT_cTeoQbLeCGLIojX5CWdoV8Ixy2D=x(x%hkY$G>!=k8cH*
zS70~RM(XFl@=C}LvHTG7A6WiF<zN2@8MXUbM?`B}>bw*LW7Qy+8a4Qbc?{_i@Xt>?
zn46rsKh_Z48ZwS)ib^JtGRXwSTGI!L>=?_&lP;tUUni=I!UfL!0%xu6YG;Z0Sz^t?
z+L<1`Zm9oeF~E7@+OT7Nmx@Y1^M2FV9o>DO!oIK5Y|+1-r)+trD_6q@kgR%<u7w>+
zh0bhoq~U*>HfuE<m`|miJFeBZGk$2h%doM`qB9+_BfS0yZ>#R1M|u5G$ai`DUC5=p
zzEtHu-1!$8ixz9?h}{>O+!w4=ug+>slhqn4R%5v#bFY^}Y64X}$jb+L54FY;UM^9Y
zxgY7^Hhj9t`dDfDbdx^n`tLT`KhsZ<h%TG7V3%pkT}vxkPju_yU0K72oNUaQAD1eI
z5nUL(Qfl~|vr#cN^Hi68j?Vvj_?}CTp8UM}Gw=V|qsZPTIz8*ZCRMwBlyf@D4P|kr
zq7p~!;z$6S<*$A`yVRyO+v0z(Z&koi=luOIX^;Qi*CwxP)6-g12Ck>q#TSu$5rMe(
z2yBeNU`dI@#z@GoXfPJ!*Vy=4W#%-egZIxpMQi$!mJQx}k|OT6JRC>;Qx$|%kY?;t
zB=(6dEny?kZG;gk4IkRF63pg*IeSV!nyp`2(+B#hpGG5QMX}3#Gp5R}L(SL-F{9Ru
zy}1kKf3aX|eG|H}c9@wN)jxgIKf9}cdZ~ZbfADMiuNV8n%*?OxPptgzH1Gj`v+EPF
gN6%h<UAuMnG5UIWn|U{CnDtcu?4kbIP5tx#062E0s{jB1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp.FF7059ED53F4F0F2.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp.FF7059ED53F4F0F2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..425f948f5af155f3db87e3685d946cfd63451258
GIT binary patch
literal 16992
zcmZvj33wazo%j9wO-MxIKtf2!K^jQ7*Oo2GI!KxVv31)LgfykL?!yFvmRQGFl4Z}x
zutWqV8-=4(1s1_V1?662w+n0ww6L_L6xecf%eCcf1<P4bit@gLyL$cg`)r;kV}JhN
zuk^8X%=|s`*LJ7PcEPr`w%(XCn~0~w3)<S+maG3$>EuPLRam*Stu1@wGdut6eap1$
z*|T2oo_gPC@%e|Z`?77xfA((OzjDui-S+lFr(XHc{+kXx@Q;T-U;pl7+M=uTw|8y7
zHv7(n&yGF$wB~8e1G#%UzjyhT^KM?*_{*<uxbUsyU)DW*Xky~y5BB+ged-T|C*QsB
zvx4sB?%?qomahKcslt<~<uARy;g~=E_3hvN;va(}AOG?2zRnN-{@r?EWZZPl!f5cm
z<1YF5gRjC9W1~;}=F$V7uKg%u+Fjf|`1{4X-v7y4fqO1WnBQ1!_Wk!GpX_+~uIZb*
zjQ6j!2o>#Z&mTNvQ~KJEuK4uQv+n%e^CL6=>HhK7-~9Wgm)5@i<A-_0o+E5`-MH)D
z7JRVc{TnmyKfG{e$?baU#_7kxpL`?r=QD0T>h#<A<6Ba`b?)mOr?1d_DE#p{W9Lxe
z&}VNOoL3n4E<AqwWpABw+H>c;c*`#@dt#OIz|Sw;vFzmXflq5&bjMu0XmZizCu^Qr
ze{}XYqqiul7i`}$^vXvk9W}ixdhW_Mj@;lo_OZ>c_rAA2fBu~(4isveue<a5Czfnl
zvf@bdi}f?sZN2F250~H5eqi60Uq5woW0A%E;vY`eUU&A66Q1@jKY#P)%g%Uv>%yDg
zT0C{q=`Zfv{SS*4S^jnPB`^Q(w;wOfUvZ*yVOf~I=ls>BGp_!HJpD+;{gdFBH+gUR
z**$MOcF%bS5AM2lo56SV<?pp0I^FpX@1OI+cH61ftor=+%ned9zism5JHNe7`={~c
zGcO%n`^+z2R!&^^VgD!h8Xtb@o6i<IqBmc*>A6oie?q>e=Q@M=K_$E4nG3h@j)$Jy
zchUFPz43Ps9(Cu?_5+)Ly8YJ&KKW>D!KV8!zx(2UPHRKY{p-R@uYSt9=dL}EXs;>l
z`tA4MdHC7yU9xMD?^@@IeGdPZFArY3yyL6F$kyGjedkNhoqrlW#pQg|cKz@7t=#tL
z=dq`MvG_+Tlo#&W{^lh&F8@J!^R=}jtdVyvNtax2|7O=`+NH%8R_xmvS-knoHHzJ`
zT_;_5`lUaO#eVQ;;^xo)<riDdIse(Ar!HBV*r1OW-dKJ1;D40w_>T?F>n}ZN=!_4)
zFRWj)^YycKel>C3j>iX&3@!R{=f$Vo(DlIm=f3bSum1Ylb&u_Q_#VfPR{rYLA9^pg
z-FmNK-%8g_D_1sh>2I!m;dcGhj)#`L@{?;vPdnnQSL#>X7ux^m+rj4!`(D3hPycIA
zRW?8PO8XJdn3E6Q)bZq^pEoJ@N+xQ5&2{Flo?pN9shxp0uFjrzlVbh;s~0-o-n-^K
z!vhu1-ZeLQhK_jQqMu&6`I*;lU-4-C^Xx5$UOlF1U(!7Lyj9<goY4K$_8Ye!dh(~s
z&e-|o)vmh_f3alk9gn>K>WksK_Po}7^c!z%tYx+=2wd^TS@%yp(ewH%s}B#1-*V!s
z7dO9`xOQ*z`Y(2U<Ezt7S(1M-c=YctTl3K|P0jAZ=G4D^aod_NPVT(?Xa8s2kG3w*
zpZWNbpI!0J9pn36^V*O5^4y;<cx~!E&lNM<{&MP_xBdI;sazs5O>J$958d+GWp3m7
zOL*sZ+oG!$u3A33QTHIT|CZl(<=V`>e|lmG?@;4q3zr=^yJ^Mi?mh9-ANVf(<I=Y6
zYP@FEnv-TXt@w;Lf86`4TiP}*;XyUtx^U}vW;d<)tiOKzLG7K~eM|T|YJBX%V;9YC
zTJbrz{vi6yiN15&+b&S!E9i(T==}fr(yF9Z9DM2951skO685Q;Qx=|b?(C)&=kEBY
zFaBQhLUL)_xoUj;s^d4#Zd&p5*`c>SHXYi&wC#K~-neSxiL;wl{OC)kpSL#kU-6}F
z=c)0ARU1y4-L&GzzA{}bPh7BJ2@k39j)gl0XE&|*$;R@I7gz26?GpB=@o@`}TQ<9C
z#jo6U-?#qh<WIay*ssRB7w+CUyJ^MEWjh~y_0pfevxL8`#>*EjKXP`{ij(o-)c!4V
z&t7Kn?6amk{?}XXesjwS?a>Z}I&PrW!=rMAI&KH<RHzg4&I)y6u2o2jxwb-D%)1oo
z!n~_OU6^+()Qx#}g}O1<DWt<(S0NqddWH0u>no(k+@O#Fb3=s;nD;2ugLzMddN4OC
zWW?N9AtUA{g-n>6DrCaktdJRVbA`;9TNJWjZmEz3^InB|Z=lxuqdnSNq2Akp`xNTK
zystuim|GRHVs5RF6?2<HHq31mvSDsl$d0+aLUzm@3OO)$RLFt3Qz0kj&I&m(cPZq;
z+*KhL=5B@Dn7b?F#=Kvle$4wT)Q|aqLIao&RA>Nmk3t^IJr(j`?p4T(xwk@I%m)=3
zyn$M8@%HFog$8d29#Uus^Pvh2VeV7Nhq<ppKFs|J`7!rb$d7qIp#bKA3I#9^Dip*#
zSfL>1A%#Mihbk1pJgiU{^Kga2m`4<fU>>PZ1oNmuQOu(iiees9D292gLNUza3dJ#x
zS16A8utLL_4_9ay^AUwcZlKmqv^_dfp^@8xlL{p<PgW?2c}k%a=BWy$Fi$I##ynl2
zH0BwFGMHy7l)+q7NW@&Mkcj!HLZg_ER%jIStU_7LvlYr>o>M4?d9Fe^%<~H6G0#^h
zkNKEFW0;RsXbkg$LIunV6)IppuFyE<;}sgmd_tiK%qJ=|fw`oRgt=5933Hhgk_j6!
zRn!se6xB{~UZ<#bit{>?YG-m@XHxA<&g<+Y{a)Jo|7IJ}y>!-Ia<#ZM1G<|Q&rW^d
z7@O$W14me~BL^gPOyc1-(FxT~u@kDDQYTim2`yB$q86%JNsHC4L>E-M#4f0INnKd&
zPIN=HTkM8vx73YQT|x&{ov4GVPSRmjpU^{9FY2MHm-JXQBn(hBhz6({Bm-7^5<O7u
z5qqH8BlTd_m@q=sC>o(^l#Ez4B}`B?i6*F;BokK62{Tm9q8X}Y$&6J?!U9!`Xo0Fl
zvS770(Ypr@_+oDkNa~%$qkp0gs(oS~RQseptXdORs9HrURIQQ~tG0v<sy5LERhwkP
zsy$(cs$H~0)h^kw>PR@C>JS}Jbx01ZIulN)Iz=Z`ostu)u7nG!F3|;5m*m2#JK=_^
zTXaL!ExEDUpXi5bzt|7eeyJa;1Bn5s4u}I#9gqgF>PdK@>JdFq^++D9dJ|r#dPOf(
zy^<HJgNeaCF#iw-b3oGIBu+*WLr@(OhoCwn4Pn)n@Ilol`k?BQd|34-{806aeyI8-
zKUM>Y08|5F0IC5gfYo3k2-Tn%glbR<Vl|WqK{X_Xpc;}wSPdt_Pz{S=sD`C5RwIcB
zR3l;psu3xI)o3CL)u<SSYE+71HI|4$H73TO8k1sJjVIzzjf-)p#-%t`hZDn49Ttb7
zIxG!ibtEye2c|>fNDfFEnZ&tKA_>)`n1pIlN@6vYNI^9vrl6XVQdmtV(oju{X{e^9
zG*&Z-3{*2>2C5k;gH<shLRAz+sEU$^)zQQ#R7b^8sE$gbSj{G~P|b>2sAi=sR&$9Q
zRC8husyQi#)qElk)x4O8YF^4?bu2Lk)iH4ls$<d^Rtt#&R10DOss*Wl)$znQRL8|}
zsE$kHSe-~rKy^Z#fa-)afmJCXK~)kZs7jKARhbepL0KkIR?n9G&#8{~s7#WqPQK;_
zij*kg4T>Zd@dibbit}zTNr_3k!6d1#0WeIG`UA}wR<HTO{^&11TTboKwS8;dt<zPl
zuSYGOJzWJ;x+&^-9OrFQ)QPN}A}w;)6m=nYPf<6rZi;ls`YF;Q8>YyB+%rWz$i^u$
zBAcejglwK7GqPohEOXqNU-07D({6LWra~HpzT2ANwobzx5Ejp#f1B%d$lOt6nAMiK
z;~;XCJ8ng;bH@kBP3}01tOy<Z1-QS4&~XU3Q|3<eb(XmkeVtYAL|<o}JJHwK<WBT;
zDncjvIvYYK`m{1@(Wfo57Jb?(Ytg5zvle~YCTr2BRRk^iv<*RvzAl-&(AQPwF7$O(
zxeI+=b?!o6SChNY*QE$u=<8|-UFhqUxf^}mW$s2_ca^)**Inms^mRA68-3l1(2c(C
zhR}^Zoy<D)>B_7_pRUR}^y%uXL!Yk6I`rujL5DtFL(rj5FS8zf`ZDX$r?0Xeefm1<
z(Wh^+9({U6(4$Y^5cKFX$ZSBLq09#K8LDhRpP|kM^ck9LK%YSo4Cpg71OxhdWbQ#<
zPnmnr*Hh&l^!3!a2Yo$F?m=IVBJ`lIry=y9&nUAIea12y(Pyl(5q-uw8_{QMvJrho
zMKGez*bt29Gs$d1pQ+3y^qHz`LZ7M5CiIz_Y(k$&5lrYaH3Sp-%rcwNXD+iDeda2g
z(Pyr+8GYs^o6%=h1T*@~4Z(~)i_8}ES;}ldpQXwc^jYd`L7%0`7W7#Z!Gb<ZL$IK)
zSLWU#w|4vLS>-bK9z?Ek@2$vn?)?C{$-Re>6`^;(0Qc7rdJh5j$=rv&zB2crudm8|
z=<BO<ANu;5+=sqCMd(9cUqk3apH*fn`mANPqR(1oEBdT;wxZA4WGnitieN>bwINv1
zXOr26K3kb>=(APXhCW-JZRoQ#*@ixwBG}MpYX~;<*=4q)&t7Ib`s`J<qt9MvJNoQR
zwxiFk2zK<@8-g8u4w)V3bClVEK1Y=u=yTNBfj&o*9q4l?f&+byhTuS-Q)VanoMm>R
z&sk+B`kZxkqR-i6C;FU<;6$IZAvn?JlG%kmSD9Vtb5+@eK3APx=yNsMg+7-exX|Zn
z2rl%wWp<;_U1m4>+*Nj?&s}FX`rJ)+qtC4fZuGeuf*XDPGWVmezs&vU>#uS@`ugkK
zkG}pU_oJ_05&F^B-w^uIHz4x>`Uc88fWCn$51?<L&I9NhXz~F11{7fceFF_)0DT^r
zJ?Qh4*@He$l|AV5)Y*eRPm?|9^C*G`eV&HkL7!J<FZ#S?_M*>QWiR@?b@rmq+hi~L
zyo%sOpSK}+(KjgbV3FaxPMHS}B3F6vR^&Plet_KM!NbUkFt}fU`)de;hk%D<9zx$x
znTOCfROKP`4b^!FeM3ziLf?=g454qRAq=6<C$kTIzB2pJ=c}?0eZD&T(C2Hi4}CsG
z@S)Gw5PazK%j`#=zs!F0`K#<lpTEw2^!c0YN1tC2{OI#H1V8!$G6&EXD02XPfhq^k
z7pQXpeSszi&=*jI0Qv$AA%MQ1%t7=8%N#^su*yO71?wC{U$DtR^aT|mh`wM$2%;|}
za|nH*GKbI?s&WW@p*n}q7iw|{eIZ2%p)b@BLg)+297bQb%whC}s~kpOxXxkpg_|5k
zUsw^s=nFT5F!~}gN6;53a|C^nDo4;4sdEH<ktRpb7g2-=`XUV>g1)HCQS?R297SKW
z%2D)1>l{U2w8>HQMHL~6zGy>;qAw<M41KXO$Iut6atwX3I>*o#YjO;IF-3@>FV+xZ
z=!?r7M_;_marDKj97kWg&T;g`n;b`9ToK~vi#LQg`i5m5M&EFmhtW4&<ze&<*LfIy
z!%ZGW->@PKqi?t&45M#E=8+=LoxU0=^T<KuDv#WXT<4JwkefVm7+DcU_6u--4PoRE
za8l+Z`jTZ%qAywHB>IwdPNFZ_<Rto%ijYKKvLPhVmy$V!zEqi0=u1^Og}zjsQ|L=I
zIfcHIBBaolY6vOxrDaZ|FJ0y|`qEWSqc2_OH2Ts_PNOfa2x;`C8$ud=8JRQa%al2T
zzD$)f=*!eOgT73YGw91GLI!=AhLAy@D6@z@vCJa+#43yE6YDIZPi(S?K2Z@w^ob2Y
zMBk{)qv#tg^C<d8t2~On(K?T!Z?wsy=o?jpQS^;Agi-WmWzM26TjnhKvQ^HaFI(p<
z`m#;VqA#lmS@dNaLKb~FnRDpNl{tsLT$OX^%hfrDzFd=Y=*uZW4t=?XkV9Wy<~;iH
zWzM57U*$ad@^#LmFW=-m`tpj9M_;}n<k2@K^BDTZ$~=a?u_}+DZ>-K^=o@SD82ZK(
zVGMm^4Pgv@1(^%zE0no_zCx7?=quE@fWAVL3+O8-LIHh+hEPD?xXk0|8!z)X`o^m~
zj=u3akE3tA$>Zo7SA=o&jW>jG^i9Y-fxd|{PoQt2$`k0DsPhE+CYn5fz6nK`K;J||
zm_VN-vxGjW%o6&fDof~->MWs8YO;hrNf9LUNew|lpG-2#ghrXl>Igbzs;VPsm8q_d
zpjW1*I)Y|({f!_Kx@BspBWN$G_ToHyQMDK6*^8>ZIL}^G?ZtWaqG~VBv#Tq6gyKAV
zQMDK6*(X)|<UISNYM-2EpH%IW^X!wVeR7_CQngRcv#V={gvojKN!30H_H*{u+Zvyx
zf4^WsbRVtQNB-8Olm+U#IJ~s7AbNl_2go;<UqhN}=z<ovzN%}B{<glkb?s%qeDKm$
zomU@FFI=s%tm<v8(%#`Y=G6Df-#kJc-b+XBrK4MS5P0idXIp#b;RWi2u8nKAwJvbA
zN7rv$zpb?lx;=Wr#uK)+mNmCWH*MUst+fWabx}sWR`m5U>ekQE{{Q`gt-DgMoPE8V
zdal+xO}&rf5|(M|L$*$n71=gTHe~xW*^wR7<Un>#lM~rBO)g~jG`W%cr>P%#V44Py
zJ=5et_D+*`j`x0KJ^svRuU@M@&dZi9Th&?vc3|_<_nUW4KdKII>f7XKJ$ARQe{9(w
z{*eAk9bVtJ{=C*>^rJstmiz9TH||%TlT$7@#ou~_+oM~9TVr!4Rd+;oTsZe9`|iN*
zXlp`vclO~U&Q08EQ}6fKqmEtGx=wfhS?6ti=+6&4pbj5rIc{5Pp?T}4S+RI_75m&%
zGcrRXkHbaR<P0T|Q!|u8PR~#pIWt2UWO0T><k1-#Mb6Gp7CARVIpq8d<&no`Xbib9
zLj~mV85&2Pn4t+|X@(?ZnP$|Lf;duCM~d@CCe@Ki7-_vJ)YtXb3;pLFyum3Nd>q%>
zOwkasZ;E`#{weY!2c{^19Gs#ca%hS|$l)mpBS)qvf*hTqC~|CyV#x6+iX#tC(eNC%
zUT5mN{Ob=a%>A14cT<n=_}2Tu4KF>VzH+am1y|C>)^4UfdKE3W3VA;**pFPM1?4$z
zy*9Tjp8fE{oUfxs9S`D#_8N5}Yip!M?y6B2a(9iok##lFA?s_TM>f>RfZS7~9%N&U
zjL4=MnUKvjG9z1RWI^t&QSXDe`>RnOvb9E5WLu4F$o3l9ksUR1AUkX1M0VB4h3u}8
z8@azm{m26~8bJ2c$b;;ykr#QeMuQLH%f3cK$i5o+ko`6CBL`{}Kn~U@h#aa>2svD%
zFmj|u5#(r%qR6ot#gOAQiX#u#Xc&2<Mk5d6KCnhf<W!AP$mtrTkux>QAd58;kw<GZ
zikz)c7CBd=9CE%!dE~JgjUg9mR6riD(Kzx%jV6$#8cE19)zp0;juh3A;=GYbb!2kh
z$X@E-OFLU{_UHB-oh)Cu@{Tjrj}K`&cFi5XIyId@T8$Q{OVb6^t?35RX>>q(jULFL
zF#z>wdVq`?BalgB0y1mNKo*S!s8`dwYwkGPr|AQ-YOFvujSa}Iu>(0Y4j`w-3FOka
zfZQ54P`{=hXh1Un<k5J5yc#dipk{DaYd_r{9nuT|`7}Ntzs3&~&;)>jnjla}69Nir
z!axyC1SqPB0>v~jptvRuG^`m08qtjGYRy~Pqe)E?D5XgOr8Q}wj3xslYDAz>%_vY-
zlLg9YazJ@a9%xK622{`#fW|fBKogn?AW0(u$)u49xJa5J@+4^{=lJ<|Uwh>KS08ez
zld==~PPm{opJ+WK>H+ub&-l$90O$W5@Yhw%H)pO}{4@2|duiET(zV`$t<M2&UOfBx
zzq#9Yl&Iq>Jb;#{6Ioj#Epk_hx{$j|)Qzkwkq%j3B0aL9L<Z!Z67?V(OJqbgmB@r_
zE|D48QX&g-Z;5)Z!UJfD`jD+9vLf3`WJ9)>$d2qNkptOTA}6w|L@s1^iQLHjCF(~W
zDA542r$ioPZ;8CfgC!cg3J;(q8bbD!$cOANksmowq5yKRL_y?Gi9*QX5`~c?C5j+N
zOQg=-=JqSG62*|?C5j^tmuMJyq(mcE;Q_QnN#s<CQpo8NrI9lw${>p+5|KwsG>V)p
zQ5HE@q8xI*M0w<~5{)4jN>o4|FVQ&iM2RMlr4mWVGL_T=D2^1>k>b3ONp)lrMw0PM
z!e4(vH<gU0&R4$*(Z00jpWUae{m#NzAQTKo+CpE)fBP3`ebd%9d!xT$aBeg*_ib$M
z+h`(o$u1S$!vn&CBNi^w9J!=@>9V7baxS;_p0a$gbK5b;tvGh|8q1kmzPa^VXPvEG
z`FEQ)nNHjAzm2Y1cl`PjPTY9X$)}!vMn_k-&R{fuqi1cWzHj9!yDQE%?>P?Y|Gx_4
B$`}9u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp.2C9D98DCC57DC34E.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp.2C9D98DCC57DC34E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f8183a1fe746cb103bebc3bd426a61786aafe24e
GIT binary patch
literal 18026
zcmZvj33yxOxvl;ETLdL>AYn*I1B94_v9%>@7DL)3);vN8Lv78HEs|VLWzE}?Y-@9n
zsz4!tP+EntAZW$VB8DCYV^a=OA#iA+1zHHCX(_a~Eh(ZU>A4&C=z8fsH$3e4{nm0`
zTb6!p@8Ke|!4NR?^lT29(>ud)zoe(9XNCCxcsv$eBf{@{dwSCK=eB+9JfL2_Z{r)z
zGapP%{nryWf8Mk7oy`|qcjCW&aMy1iIkWu8b$kEuqwoClZ>Rijwtvd<wb!<N`?!6(
zj(o4JUA=DgI`zIse!KhM&V1?NzW4I0KixI_YCd@Bd;12S9{Ew_?+4<~e13S>tpkhJ
z9ISkJ7kzjr^i2AgZyh<ZPX4slGjjOz6IX5g(}TbK;7`d{UtRUq=<y-hXMfAiu6*y+
zwFe(dX+Al;NcZ64r(ccS|NeKYkFqPW{Xe;M`xWmj)BfkOx6XU;{1YGj!TrCpF57X`
ztCRP<d~fUfFG)|?@Z|o(?>Mh~B>wInKa)Q8{OZhsD?VJ4y41%Wc*b<|@vq!+d#-Qq
z!~MCp-cR1K^&Znt)mqzcg8ysexi{8@x1M%t{JQPI7mqtT%w}`GCx3iTxZ=F+>4`I5
zeCXh7lbf%+{$0ri{SB&n|N8zzL&vTE%bF9nnAaaT{}0C)HvRGIf1V61-ZcI^ck8En
zUo-Tw`}W?yBRyUG+S?~>Jt{qS^s1qUUTyasI_9bSet!KK7jE7A($<&vUwy;XWns<s
z<jvRK+IpNXZ(C$~?v#VUeTTXFhV6ek;TFs7TTgp&WW~jLees;%Ua;ua14~-#&wlgU
z+pj+Q=vZ9-%ZGk=$A@bU^={tX+jsPiZ#HDhFFC0~J2R?VVl$uJdSLjYTlYUx{O0Mq
zcaPn)NijSieSP(zi+3-1qj=BVyBufS|MOek^#Ak4j}|Yket+A}$M0<qz3Z`FsXRBR
z`S%@<D_(r!%}+cBGslEhe|qxC<y%+1w)9Tp=8JXXr@vEbzdrHk>0iHOd$U&k2|eBS
z@%8_);i&sQJ@>iZvyYuI%9!<qA1u0ioGNds{`#NkVM$}j^B)g>V@vT><$JlOLz~Zg
z|4!R?m-i+12e-Qo5B>Sm=l<5rc=w-o_Yc0er?fkJ?N=W;@!S*s_U4TnR-M~)C>=kv
zT>bF33_A>eaNW65`03An^?TPjH=lRc>rW^p4(C~)mCpYCfspdl{if#*@4frUBfD?d
z`PrkHH}|dl-xB);SDo^(s+D`;roT)Dd!laRcP}cezx?SNcHH;)=rQrb&s*MWTxYrL
zT>9kekKX*<#RsilJ?rOxZQZyhcKR=`eC)*NBZ0pkIJ7vw{tx=QF1z{Xm%DGc&G`IX
z@{w~kTQ6C9!(MCVmf`RH`p+l+>_>YJpL=Ed+qb&p7q1<C?}dm$vgfJ;q4O@^b!_Pi
z?yo=I{={`p-W;pm{c`%c3wZA4#1p*qSM2$@%^&Uf>Pw$o_-tnO!k^l1EB_*{w|x+Q
zd)nDJ^7Kba{^1J`-SUeMr`}H8@%QgY4xe^X<?I#QLnkl2%JR{g!r;A^?N^-hqu1U$
zR%!dz8@Jy7@Yem>FZfIL4Y@vkbl+`H{P)_+e)XdAy<eU5PkZk6bt9J@xn<`~*Bq&~
z$4))Wov}3grgz`#MNi@EBhu~vY|DT1&U>~`cdy*^{y^myr?NdK@1J<=j-Q-=+nt}i
z@%m90o)BgO&iLui-f`QJ#g`WzyRm0Hvor7u>gieXkK2D!w5l#%$}YX4C$w_W$`$h)
zaTS^Wv*N+>R}v5W;n}5Zs~9g|w0!aWrW2oi%7!(YezyN7OWCj(Z(Fo&+x(^zpL1x-
zr3cw-rz~Y2G2Xdo=SB0IPQ3Apzx?sugP8}HvdhHygheMDJ-_M1n^vuR_D1_NgUfm@
z7UT6R*RPx3bYk1a7b5rma(!}H&o(h$y>j*1`AsLjD7Z&&xhj5EZ_njoT%w~&<out%
zI+ftMEeC~H-?@G%8xbpKEjr6Kzv;v??H%85zWwl>%X%&n<25VSte@X>;+K=NC+w`O
zetaqWrWkKrwDIEkO(%Y<81H-Jv6pXI*5eZ6Q&*naH^1q``{zFU;Ow4bBTJb>j8`pM
zwQPRViU0A)=%szChwaOH++w_L<+}Crn@)UW`L-Yb`l_EET*@vH;}wfmES}$V;+|MI
zG9KT!@Y78!ng6L?JNL!y&Hd+{)*I?CQ~yoWc>qMNQvV&m17#Y(e4t7Ln9Ir}!(3J+
z8Rqgb$uXB#NsjqonFcW*tkNLniZUrMS5!%Xxw1@3%#~GAVy-Ha3UgJJRG6#Fq{dub
zB{k-nGHEc^R7r!mwoF>gwN=t$t}Bxcb6u5mnCr`=$6Q|}J?2AY8oG%(PqE(6P?d)6
z05+7#fVrVc2F#6RGGcD5k`Z%LnM|0Qs$|04TqZN-<|>&nx0K0(xur@L%&ldzVs5RH
z6?0peY?#}sWW(HECOhW#D%ml2l*xg)qe>3Uon>-j?yQm%b61&Mn7gXv!rWaZH|Fjt
zxiKFu)9_8yS^j!M!&Mr-19+rNBbbj=X#{gmnLL<#s^r1kTP82&-YR)9_m#<qxvxq-
z%>8BZWA3k#AM-$&0+<J?6u>-KrXc3QDg`kQl_`XIs7fKs!(|F%9<EXt^GKN@m`AD<
z!F;q#qnM9YX%zFZGL2zAR;4k_$ICQ+6LsEydPCz?8ovWLUZyza@hZhJPn0Qvd7?@Q
z%#&qGVxFv067y7<QkbWzl)^k+rZnd1Dy1>clqrLGrb-#ivt`O+o~=?AbG}SG=6scS
z%qPk;f%!z0CNR&HDTjHkN;%B)Wy)iouTmcKLYWGf7phdie3HsEN!XaAsyM;Ul-QZ-
z>P(59sjkkn*qQF?OpBfAu1-<x6uUY_u~Y2o2x3R*>Ih;-=<0}XiQ=mLfA=fIm%#br
zD%;`Co>G&R%<m$>du_CT2fWk9`V&BWe|~`mq65$zhz&q<fFFRSEGmPhEGC1djF&-E
z9+g8=9+N{;&dZ@W7#)P>U~CYYgZv;g6;TB=6)^=g6}$qP%BT{W%9s+GN?r*~Ra6B{
zRZIm<6|aJ(I;w`II;MuEnpZ<p6V*Ud6VpIb!)u_ajcTE(jcK8&<+ae%MRm~B#dOfr
z@j7Vgqk3rSV|r-nc|A0TqC-32y*@US0OE)83uK5IplOI1plRR@&@@Jk&@{%3&@}Q!
zXquuXXqsXsXqtEvG|f>nG|e$HG|jvjnwF>qnwFRanik#yO>5K&O>4{wO)GDOrY&lN
zrY&ZJrj55j(;l@$(;l-!)6UzW>4-X@>4-U?>EIpEbVi-fbjF;}bn;GUx}q*<x?(P9
zx_B2f-BC9*-7z;b-Mkx`!_na#u(uH#P5|-4`2`w@jzDuHHUiBNegvAHs0W&!m<O63
z-UCf<)C)~-%nMB~?}er>>Vu{)=7Xk>_d(Mi^+VGi^F!0m`=J?#2A~;;1)v$=1JDdc
zgU}4dg3t`|L1>1eA!vqTA!vs95H!QlFf_xlFf_w_7@Co21e%dp1ey^(0?pCrC^Sc7
zqtG1XN1-_u9fRgrYz&%X{1`OHqvJbZe<wDc0OH5<3lxvWp&5_Gp&94n&`d-V&`iV<
z&`j_NXeOgcXeMJxXeRk2G*i(OG*ht@G*f&En(1g7n(0^?nrS`_%}g`{%}gu<%?zJ`
zW;U9IW;T|EW|q%FlaKPy<YPQEd7g*nM05h06R`<sPVf`Z%tdq1%*Arh%<(yB=A(IN
z=3{wi=J`A{3(*2J3$X$;3w!~ZlN6mK=uT2>lAt|F{3JntilS54pQ6|l_NR!S!u~Wx
zr?Edxv1#m26F-gpB1MbXFH)?C{UY&2><bhXurE+dz`j7du+Tqgb=&yVONUlaZ|LML
zCtEuQb2<wpTQYy7YGE(!29n-Dmv#2$I<FS7_T^!#h1!jD{Eg)8)Xw`}=*(%^Z*LJL
zWm+yCQ|SbqLtEE#OXiPhE&K!h71qCtE$s94*IEBw<R<ICAGyu?KSwTc{bde5zsB|N
zK^7$aS4$T1feITy-$0!Wpl_hb2GBRqW&`LOC~*Vm8>n#u=o=6u1L%`gm<)ZgI+LMK
z)?_mD$=XbYK3R#Ap-)!hWayI#5*hmB6(&cYyw2q4lQ)?heeyPwqfcJq<mi*vI63;{
zf<%tK!3rBh-(Z~$qHnOt2GKXzW`pP(EOCSA8?14I=o=IygXmLKm;!x@I#Zxe(PRqr
zDcVedK1GRBpifcb6zEe35(WB{6{bX=vd)y~Q#P3reabddqEA`kl;~5|I3@a&f<%cv
zRfVb0r>Zj*`czG(LZ7P5ROnNcI2HO-HBN;-l^{`}PhDYZ^r`DijXrggsnMryGd23u
zB~Fb#b&XS_Pc2B)=+jh~27Q`3)1Xh&WE%8o+DwB!O^MT>PgCPG=+g)i4f?bdrbVB&
z&a~*$HklTE+BVaoPg~-&=+o9XE&8;AM2kLMh3U|zt1}(?bWNs1pRUbx=+l)r9r|=N
zPKQ37Akm>uUtxOm>FZ37K7Etv(Wh@SJ^J(|PLDo)jnkt~FG%$08>+COU2Neuu%S8|
z+Kb#|L-!-M+0f_6C2pw9!RObwp*_fgWaw(iLT;!q1NsbgW<Z~z$qeW-w3z{Yh7xB$
zpP|MX&}R@N2J{&#%!oc?of*+*Y%(MIjBRE_pRvRl(Pyl2M)Vm4i4lFK3NxY4RA(mi
znVQUmK2w{S&}S-fCiIzVoC$p<L1IFmxx&ooGuN3JedZ=JqtD!CX7rg$oEd%Q8fQkI
zS&*2~XQ?m?`Yd&3L7%0`Ea<benFW285@$i5rN&v%XAvY8^jRy+iau+dS<z>0GAsJ5
zZDvKEwZvJ`XRUEo^jQUo6@9h}v!TycXEyZNn#_hiTbtR?XDe|w^x0~h4ShC2Vnd(3
z!tCg?*O?uC_9nBV&)#Ns^w~?C9ews1XGfo1kl4}Zs4xfm9ChYEpQFhf=ySB01AUGX
z=RlvM#yQaE5F`%tIV;SGK4+ad(dTS3C;FUi=0u;f#5vLDtZ`2CIR%LmeXa^~q0d!k
zF7&yY%!NKzo4L^EDse9KxoVsXeJ(-bLZ7?B+~{-HnHzoXCUc|D-DYm|xl5cIeeN3P
zMxR@dxY0LUVZ*yvXE$FwGgD{7dy$)L_<rOz8~z-*#0{4@`1~3-ya!p33|}o-$VVz{
z1briQHiEv9CL2NDNSlqIZ=}SHpl_tcji7Htkc^<uQ(+$TdFsrAK2MW*(C2A05BfYM
z&VxQrjq{+-BS<{x^H!J_ecn3rqR-o8Ui5j}%!@v6iSwe*TjRXw^9m9#`g|4UL!YnC
zeCYEvnGb!wHuIs+SK@r=^VK*X`h0@KhdzIW`O)XEGe7$LP3A|Rzs>yU^Orb3`usJ{
zk3PR3@uM$LVFC07>MVf1K$8W~7ihBp`T`{`fWAPD3!pC`NCM~!R#*^y!8!|~FW6*3
z^aa~2h`wNn3!*Pr<AUf53X&lDLKPN5U#QMP=nFMj2z{Y83!yJm;zH;P)wmG)LV_fO
zzHo(w(HE|>F#5tx7DivV&BEvlm$)$c!Zj|8zOWz(qc2in5%fjsEP}pBlSR-MX|o9W
zA|)<@zDSLWpf4gwBIp~fuu=4l*4ZffMw@IDeWPtQioVejH;TT|8aIl*Q9&|_zOf1$
zL*H1PjiGO>$;Qw()@Ebq8!K^R=o_nXW9S<bBxC3sudwl5taH3YJX2d|<9m^tZ2W%Y
zHXHvOxx|f^Ir#h<H@*j1kc?j~S;*rR7Dr#a&f@5cH(4Bg@ivR2FJ9u}=!@65IQrs(
zB#ypBg(c9JsIvt65>1vsU!u(t=u4Ei1o{#+E`h#;AW5JvSz$@^CF?ASzGRan(U)wq
zB>IviE{VQmjZ300DM*s&OI273eW^N2p)b{BDfFe<EQP*QiA$j`RpV0VO9_$``qCAa
zMqj$l(&$S!SsH!mHcO)~UE<Q{OV_wG`qF|VjlN8UWzd(Yvkdw&O_o7lrp+?w%aphb
z`Z6^xgT9O)$)GP=VOjKL>nw}DY?Ectmu<5w`m!Z1i@t1)%c3tUNV4eTD~v}UUuQh}
z_$K4g$F~`eKEA~9=;Lb~k3L?I@aUVUunF`{)Y%03CYo#leG_dqfxd|nH-Wy18aIKy
z2|+S}zFdXn(3h*T9QtximP22z&2s3=mAD-Gay2f8zMLS*p)X%ydGzJ$ERVi?ljYHu
zZ?io5@+B^hzI=_#qc1N=^5`p6SOI;7IxC>B&}0Sl722$TzCwvBps!Hl3g{~ck^=fB
zslp}+jgwRtC+M7{rZ_?CB(=o}dMByGiC2NZN<B$6af0qi5+sv^_9@Xm)n%U&?NeR$
zDbYUFWuFr5Q(g8c(Jo$J0`@7<KGkKP675r6_G!^R-DRH^?bBWMY0*C2WuF%9(_Qvy
z(Jo%O0`_UqKHX)X7VXnr_M&JncG-)fz1U?hiuPicy(rp?UG}1A7q5u{dr`C(yX-~L
zUhJ|9qFv~+3!+`<vJ0YJ=&}o<UFfn4qFub&2JC`p7rN|%Xcxf#jcY#X`TL>x+Y(7=
z7ahNgF7KRXk%X?J<F6y%0(YOrk%YR>*KGQqONx(QGJh%Y!daQ|IU0WjuMeA}IC5f+
z63EFpN+PG`D21G!qcn16jxxyEIm#mQbHpQ0%+UmLZjN%u`8moX7v`vdJV|rn)l@i{
z5+_q#lWB1>-8CtSlVaCI5GMjmIu~8*m&{*mz3>3`x2XRqynwhx1IV%#$&lqOk|Phc
zXb@S^A_cOtMM`8<i&V(!7O9anEz%%sTckzSwMd7oZ;>8(s6|6h;RUxXG9VjUWJET#
z$b@Wekr~<2A`7y$MOI{6i)_gD7TJ*<Epi|`TjWG`waA6+Zjl>#xJAQH;RUKK8bS87
z$b;-{kr&z5A|JB9MSkQ!ivq~O76p+*EeaurTNFl)v?zi++M-e9u@;RXkGE+2DSW}U
zD2|+HQ35&Hq9k&vMJeQTi_*xM7G;pLEy^PEE#i?US~P*2Yf%n4-=aKnp+yDcNot8N
zSe#6Wlc}!Bv^bgWniR!Jv1=lT69Fc6&F-sKer?zF;u-3d+LfC-w|m%~npH0h)!gEE
z^*2^;>Ab8v>!^4Q<G=2h=)AzUI*vQ@_cxq>)clF=#br9S)AStpow=ts@q|P?&E9wN
zrp{^h-q6~<wVOKkK=g+C`ua9??rG=^ZRp#usk0P!&QyzM+P}O5qa*D7e^&Zv%<$}g
zU(<S9`~p^KS8eRvK5@;Thfdb*+w~Q3ynOlcm7R+)uh+l$kY?MXPl@9VTQ-<Ft~*be
zUH<O7@x$VH?UuFf&X@PYkBgbh_t&ozzlXE7o^^TWE9wnx{N_gA!kL+Eo^8Q}<9<7@
z*y-*3R@qEHag=N4od)sWJ>i%WR(4MQcOHymmdxKcvhc0-&rts>xLD560J3a`WXSRv
zk|Ph!&>*s6h7`!k8B!vvW=Msso*^}|W`;D#+8NR!>t;xYte+wM0?%!|@FVZqx1JXt
zz6wby)Yo}#^oFh_$+gJeA<1`;cavoI0(Ty2@u~3T?KBIISNEIzfBo(^Pd@bPM=au_
z-M6Lhvd+`0^V1X`?Jw`PS@`UsSsHo;m-1OMARA}Nh-{iA6S8@h%*d8mvLIV$$%<^7
zB^$DRmh8xmS#lsdXUU1|nk5&qdzRb_{Qj|r-u7>?Y}Sc4?MVAW3s2So=>U*SDg%;B
z<v@edK_G=x0i={FfmBizkXot+(nvKxTB#ODC)EMzrFx(t=}@TilkN=}qy`|P)Cgpf
znt;qwGmu4U0kTT1KsKoj$S$=5IiwCCr_>4LlDdH0Qa8}BbU3u|;vbQY0C}VyAg|O5
z<dgb<{8B$qKpFrFN`pWlX$UAR4Fg4_5uj1&DA1U63}{?B9_lPoy`i`?4wR54fRfTA
zP)eEtN=wr~8EFP6E6oD&QXXhRIsufE=793jJWxSe0GcG}Bmqy6bP9Qzq|?Ynk`|E#
zk_rpld8;^Y$^0F53+qCEgZdx9Rk%R|$g&2>kmU`MBM&xc5LwY61+ub1N@P`oRLJTE
zsgX4e(jaRaq(#;>NQbO%kRExcK|>GVI@cfrvavx%WK)Ao$mRx_ku43fAX^(`MYc7_
zhHP(;9of+!2ePw4PGnbuT*&SQxsit(H2eVWvNmW0+0!5ovbRBAWM6}P$o>ZTkpm41
zAO{;1L=H75gdA>A7&+3Q2=ZuyMv=!FG=@Chpz#NAm$gB0<V1rK$jJsJky8yyA*UOZ
zM$R-SgPd(p7MX7lk37+!3FKUZa>)4x<&g^wDj-i%L)>M>$&@&m>Y7Z8lj*KWQJfUJ
zCW1H-x+dbyhvNNto!yAe8@%|&{^gCZ3oqc|85({CpCB_dg6x?g53+ZLyvV*8@*(?Y
z$d4SDp#XAlhJwhU844kXXDEyunV|^s=nRb_kIm570)O$vJu?IERu750HLK59y|J@P
z-nrFR{Ehp|`+OI^`+kA?ufgR@paEo=Kr&>xKyu_kfd-Kk0x6J{0x6MI0;!PI0;!QT
z0%?%70%?(T0_l+T0_l;51RA;qmoI?~$VP#T$R>eI$Yz1e$QFSt$X0=@$Top&$aaD3
z$PR%V$WDQr$S#3g$Zmn$$io5+UxUw1fku!$0(p?V0(p^r0{M{r0{M{x0tJwR0tJyn
z;(H)+SfDU+M4$-rs6eC0V*-sKj|((@4K7~-#gP-@#n{M6fs)86fl|n6fzrqsfilQh
zfwIWr#oavegg_I>Ie~J>dGTU!<bps2Wbxp#xP0M6Jm4!XUpNsD!ivimPQ(MF;_`(P
z@nEI+-@+so-r@i9M*Q(uaQtHNZ!WZA=LZ|^JN=K_7lplkUm)1y`!fF5pZVs$KJ<2G
zJw1Vi4}%LIh8I2z?hNf16VV#6bCRPL9W7nFw0GI^V~;Vf7}|WsiY4Yv$E`a4gp*cl
zb?2Rb!G#;Y-hbk``qR~Ct^NN^SFSnvl+)Jrt>19w+2?$9KqgnHG^Z-R^0mP&C$2PF
MBFy0YhKX$d2c<h05&!@I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp.66D8D8BBD73054E8.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp.66D8D8BBD73054E8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..91c205427ae80dd2e8daa7b0df533724c75590f4
GIT binary patch
literal 17954
zcmbuHd0bA}|G+!v+<Qx1Pggxf-9m(VdLoG{Wf>|AV_$|D#y&G<EMs4WQ7B|Br5chY
zN>Njmlr0q_yD}5P*PuuYEl9=hRL|*t=KIg@^?i+(x#xNNe9m&uIrp5;{d~@y5d;1F
zZ2gRk`b`_SWY+WrlQ|<Jqh|EK`3vTKFYrc2PqaoxOLm?fJ8R3`?$X<#mx3Pbusr=b
zB`I!flY%p?O*49oxj$y~^re>l^da3-c1=HVyY`)Yu}J~<W&XW33;TIr@6zS%lA3Ja
zITLGtTzGq=-^h?*b0Q`-^PN5U{hX}jm8qxwUOCN=xn`SwEsX3++%r+?ej|QY$%Mg{
zQHR=vR`l2x;B+A`U}4ez`vbieA4m>ebkqKIh5yBq$4?bc8R*mf>UZY`Tv@607;1UE
zjqj?<vy-1K8fSZ}f8wc?S61h2uXQt<+0}mT&;5KqbDMM-b?8{I=-lu;U$2gL-Tw$3
zf4tM=A7k>yAF*0cw9McBhF96Jj*ki}3)gf?e5r50#ASf)+fjvn0U>jx;t|8&tUoZ{
zYM}Sbzd9c|uzcy(ABL`4I3&q;z}oLKiaMT+H;IXj+ircwb>5~=k-~;xm+%&gmiJz6
zyEVe(%d<aq>9wn8udyR$y&sh@XZM1?J+}<{cyQR;^RMskvzYedq!;=M^FuGDSyh#`
z{&U%c>x-*qCA?UB@9p*c%6Yw)e>K7DYR_I7FYbOHnf_LJ(WCR+<yJqrZ7RE?KQQKE
z*3j=go($RKedmkgYi9jt@S>?n7x;CDk1k2v{d8%L<?62@U;Ffa+3%j`gzHyai~L60
zT|e>cs<rE!CHJNkri?e8J<j#b@x*%z?|tz`oEkJ<IrUBRlE4$A_f07OeAk5XX&Flo
z|1Q5B(^Hp!re*4kqqb94&mw1Ijt6<(S(coaH5xCD$~abO^T6-=tEqE(zH2%sC}jNQ
z?qByapMQ7w;T7p8Z>}*O^TP9R;l&4I52yF)c53zZ--g^fw0_L+l?QFdr+R*A{BZBY
zUtWKf{CY!fpYt)dI_H0NcJz{k>xK=zGT-&{@4oo^p5uM5xQKqs=U$Fk;@EGs)xt3I
zfG@UM>%7kuEgto8h+FzzbA8ORe9gm7{Z4;G9nCfbeIM>JX~nenoBi##FKFpJ_tS>6
z=cH+=*`8+}FBv6dT)SZ$<+$$;$Dyy{SNY$vjM8P^^1PLLOPApMU4c)Qwe9hn$*Ti9
ze79snVCyuqxwg6AM^FFS$@WOD+xFd`eUpWTx?4BbE$_$usFb>npDu^;$2Pcrc|h4W
z*2?N#LT)BrFxhlvSZ2$@x#5T0Cv|msaXO&G+?5Axt^{U3+wJYkU3ciJS@@{=7d!vs
zYqoFspAoiGZ_b!y=TzXebE~-T++Xeb%x`+ZVW~LveP7pw-?aY_dCug=?p>G6d=$U(
zK)cNqzm4_?8)iOn#pCq*$puGezq5JYPJa6Pr9<{T=iN;izF^FXf*Jb2m6?<8q+58+
zXmXG&vT55Q@5tZP&4$Lc8fam0sb$T^PX0b4EP4-YUu(44XvdS52Y8ziw$=Kke|hg;
znQSyP-mP%llmpQ}r+Vyc5?yOG{m!5u#clJzXyK+=JEwOY^g}!4&&<otX*TlFx{~sW
z6q~VEi?64iDs=4q<-=}ou6(WgZNcoY{t-TR+zu!^9yN1MaSVQW^yRP)7sjVx{pGwJ
zeuK7NJ9x3rS=Y^O2^PVZEU(}G`^rK7%#vqGZKAWj9o}izw-&t@_`F!_KDO9uK$8Vy
zxj%C^g%?GZ`98V6`aj14?nUh{S(xzB>{{6FLBDOT?4M!&>F0ATOi<y_&`-_!9%-v>
z^WB+!-8;^ixh^&$>QlsLH!Fq>`f>2Lrr$bl5Bp5p_s6eh`cE=Ho4K&O?VKfxUn?J8
z)i(Vz?R@q*?w8enKiK^sBxTONz<0+Ut+;9ZvS{!sd0W!j>vOB_1f8h9{Z-W)Kk4@R
zvR3g^kL1r^I%~=%V<RJU^kZydJG96Q8+SBPGC55qXKhG~Oz5gf{}>CiK6re-dy#-;
zy^(U3Gds&!7!szU`Tf#}UPIEGYH(9mBjqS39p#!E5~gD1c_=6Ma#Lds)=-sHPD-UJ
z4GB|eedqkQpEm4WB;mHSqGG|JSm<I%m`a;x;eI36`legqmQ-b&?YK7e7bbSxma!mj
z^XR^2n5W`_rUNbOFHEd=xA>rhmz=i1_Eh|kw|&TWH6%>MEq`A1mqT6+Gs7HJ8Dcs_
zTYq8Vo;`=(cbagn#tfTK@#m(WTh?EgxKFpgjUq#Oywu<U^d^(I)=69^L(-^{;geQA
zva-Jhn^Tn(PD-Ku(2z8$eAG#gzHXbV!8)pv$VrLZ*M@|t40t*_IL{)#)(SVL?HFY{
z%C7#x#6!<)|2w~879CZAiW7)w0+9^~QyJE4{`K(W&}sooE=G!%vDe7@i%}Dd`qF8k
zpkc>ymK!gbm#4wbv=#B3bv)PJkTjO@XM3^Y)TvBsY)n=DLrnfdBtybfMlM)<J@L1x
zi5l!iRnj>7G_JEDVJZ`+41GVp+j0{P_Ms{#IQtWvhaqWHiC!`CLPd<H2KS^Y$2t4s
zRHY$dDwDt7HKXK!wAu<=&|4mFJHCDWg^7Pmn49Tar76?k{<MSxoaF(|&X6#bIYBPw
zT|Y*JYOpU=+0WVR=h_$&rm}EC1jfG$tIV*LR+2(YQ;4-8VJeFTT;CDjC5Y4D4z!dw
z&NhzgXh@jK;`nuMI(8^qpuxSVN;+qs&bb;Arn2OxbzVzveD$3bwxJb`u^nSqe_`Tf
z-a|k8t51*SW|&ZM95Icf?Py3ERo<-}-Fw8at`c^jD%k>;Ei^YIOlA4j5$K!N6)(-O
zDOE`#rb(oQAz>=3Z2mkR{?nzF7Pu8vnZQlxQh#CM)iLk)ZCQ=OEN~kt&J?6fK{h0f
zDp@YthjZkER@jlM46`5RRDWUOz`?c69_MvyqQU)WUnFy_lj#%FkT8`EH?MpwNE%RP
zhMUn6`kVH*sJ}4rCeNx9eLo#MZ-&jNc#!EJ%lZowZ+WoZYxUE`o>thFipScHwX44{
z@nPKr=kLBvjux;n6+a+m4~V59VJc~R3U=rBHoj(oTT_*2L5dda3<*;?*`sXo*p|{v
zGu(u#Bofm^`n)wHOyx|(wY6;*xR0{JEvU+9+tGIQ7bd<m@%^l%;GG>LY)8eBLX${Z
zNkhU^a;Kkhx)6A>ofWpGDkE)2+SOl}_-Z$YvPb>eWNNS<?a5fqCYDa$hNMwtuJ>A>
zok54JuoW#~gzX5s`U?|Z@4fWcqvyAYge|GKRFFyqdqdKw^7OewT-T5S4emuNIl{F%
zLZ@y+!c=be%iVu()7}Xh>_k-#bIlKP&W41kJQ?wR*3uKDJ0;wXZaM|Lrhso|NE%f}
zKelWAy7pv)N*1Td;+iQ92~#P%-(}yddtTEuxDTx;gR{?|Pg+C5R4VVDn~<@^kv6wI
zRXNDn9;7M_2~&CRux3!!=_yp@Gpcfovp>dlF(i#DVH;ofyJ|<D6Os1IVPbliG&dwn
zrP_7zo%b*M?$zK<w3L&a{YhF%L(-@+1TDJnc;=-={RlthrKfybr6FM|M)Ri6n7_2U
zx+@b?!(PnxV3GDgi!8476GQii%01rY9^ZxaDBgxUCqBB{R}hs<Bxj;d>;V>)^GH4q
zc^%fS!%nJvJ=U&={1|H=Lw<s_PgE8%y0-TAJ#yMsRJNfW+t6sWpq<ETCmN%29FpRY
zi^^w^+Zi-S<%_82Mf9c0*HDjZu)Mp->n<#>2uVe-ylUiLjlNLxufZN`@KBYvU}*~m
z^@m~iFbwo};vPFOX#YXn<scrS=0AeHj$m*<=~zmK<z-?i6Y?L}{SR3FL+tSowx<M3
zB~br0?)e&n`+tXfyo2poO}eZmpgrq}*Lnil6Go&k0`6}ok#?%gW)5d+u-VwgPiKhG
zAyeqc)^JgIF0_6wbYNCk9Xmf4+!!}?TNgDmtn)o%Q5n%{M7yhowM0}#wi?-vtxcjb
zs@14=%sLU3(XB?eV=JGijA=Ed9oyMNWo)al?bzxpD&w5Tbz;_C({F8+?-T6KAx?QG
zi0=dqTbY=hiAANBXljYoFoOkRYOprhaxPUT$y>a2`;e&g^y=B0-7_;nY5p3bse$<$
zOi`A<#h<^K)EqqF&T&dP8da`NM^4$u)4$!>1IW<_u=zc{1$zKF<yWEYuYx06&s%By
zJ=dFDc|;qshL_jyPAntqR%B{0aakK3(&v3wy2?9{-uxQgdJS*Ga<u7v^kiB08>aud
zO{|--X}$5xDi)RT&f_~XJBpdobl%k2(W=sZPP3nD&Wt5d+0JXX^Y*NfqH>(q9#>iS
zD0b1C?EV@eD(jKB9$7TJQy3XhMI;`Fd<~1&aHE{r>a>3v40YB5*J(9g_ISnlW}Nbx
zcYDpdR~r`AIv)F)AE)vg-t`UN3G(na{799{c;_<S1@gc$evr!Lysn(rLmpJlf1z>>
z?^Xlrui?km!1`->=UP~PEkCdpmj990eT3zI<Oh9(<p&DhfkH2}z97Ln2=XSuZ<8>9
z$=Rz59oD+{7S>;)a+H`HB^Jzf?OYUnM4ps0hfbql<Q$B2Y^CFrG~}EHxe~clBE2fV
zft_#QMt%1IPc{fP<GK+}2^4Gs1zXk-P6-lhf*@}aY&NOfz3+~`sUIB7|8)yIy9ADS
zkd;u`N>p|uc{l3Dnkp)X5dAws<#R|r2jy8v&Vqac$u}V1Lh>!h_mO-b@-w77gXPyE
zxfaS-VR;n>?Fqu#APjcn4OqJY@@_2ehWhbXj)(jkmVbksft3sl?(Y;<PC?GXN*3e-
zEEmB1_pp2q@?Ti{7i`Z{tbGc(9LwcUzY5D$kgKs=4S6k**Aj65Yl+udxM^)9@<u2R
zCrUU0_p^g2J0M3BC0gYkE#Kt+AWVKk*MK*I_(t6zIHgPw%OIBvHsylMHV{s!5Ns+S
zR|+<jklzZnZv{J5|DEu~J7K8G?*)tZf>z}!L9P<)RsJBzA0Sr?Zq<UjDz6c`)CgTw
zt`*#Bq5PxJ<s+1T5?ntComBl*#;&W30S6e10mh9Rq_?JL_BVcxOGQ!Hf}|~|i~0ma
zAt?%SERtd&pF+|p$OT9$fczLqk0DnfsS0u|mSQoe?*x`kK+eZfK4gC)`4gZYLZlFt
zz0W+{H8}D@4c#|Wh)oLV!<H}RfuXBho%=;y<zhKaEZ3a91u;(y-RY}{ri$3BmO_JD
zhVAry4E`9ZSWPt5&?IQ^$}}n(yfSnrdrLHL$#_<xsO;jTUEJ4<O%1*pma%`{POo3b
z`_5lnHzpoiP%m{%LXi}TY$3acB5#$$kQ4^x;YbRH9ECihV0qC<iiYxN<Q@&ni$PKh
zl<z^(9>}rCBNmnykM!|yRZBqn1jva<p9uK~vOI$9)%Kr49;aY==a6&`)|ZLgGEomz
zKNI!IM4zjC9=V-|^7E+Ac_`0DQZ_6v2T3`QFCppDKiT~f?2laJmJ7?jjHJs@{|b_>
zK)!*b8<6vmlm|H<N%@cqkVgT$|3V}cLU|!_FNE!RfTRad{t!tIAwNbQk70SmNM8*5
z|2fh>hg^d6C6G&zWhoq=D&$cG%d16FEv(NUOa2(l&j2h1Kn}!GAmnw}c^!Pl*J1B<
z*iRkL_1Jkml&{C$>!CatOTn<bjab?Uc`Md$g}cmFJa{VxuR`1K;BAmYus#Iphhlvw
z<Q-Vr0n3lXQY7RkEJZ<%#!~b@**zMA=QRdPF;Kn-OM4*4Vvkr1p8t3(#Y1@lmJ%Q*
zVkr@FGM1A6$?nOp{uC^wKzS;bQX!wf9w%V^8CahI`{Oj$pN4z}>(4+wkM-vvXJdUf
z<Q%NefqW5LUc_KM$j2V}u>Jxp6~OitVyO`FeJtIF{0K{rAV0>=kKufJjJ+Sj`Sb)k
zKY{Wm*!u~T7h$OgmRF3WV#qJB{sp|B7kKasIKN8q;8MshvHm60e}(n0AeUjO43=Mp
z-OJ$kRbZ(C$}6!{3AqYORgi12R0FvdOSO>wiHAP{^KC7W))MeM2NEd|auAV%AO{mE
z_@C?^Ou&5KM5IknzL`jyA%_r;5LkaW(T78;FM{YJAV(5?B;;L0zYFqiqTdZUhUjA;
z?<JOdN#lHDZV37cTNfcwO@+h6_b}=I-r&}wudtl+cFu+E-y(g#|9T@ok-gG>F;r>(
zyKlsBYKo+hwrQk2GeuaGg*P?CR@53!5|fj}hAE2F=V)?}vtjN?Q3>Zehx2_|4YXjs
zbDGMm9BSj#1!e|1E(+e3-fyL-^tb9inYqJRe1^LEDv73&v{c)_ofEfnb}EN+ZsA-X
zm7_S9D9%ge7|t$+(=#?TM0uF~!v3AUi8q9Jm|$cR-_x_EOPguOlnPBs1#@PlvY-#O
zsr)&OKi7s8VQL8duo6ZMvFPw@n>XK9R3>>&>aA`RlRYN`n&LSH&{WT<fTnp)12o-p
zI-nVzGXND4>mmYrrI=V3Lw-iApFw_3te-<JA=V|3Ul8jTkV}bmDdb~Z(_@?^^Dk3>
z2iNpCXF1Uj$`WaatLX{KhUgca16<QI&XV<=NZrRxPjZ&bvn<jGQPXtJk_C@M>gsNq
z!C5jdxTx$G+UyscShPY^Vud!bkPirL4nRIAv^fYlPG}PcIbLWJ5BZSL<`Cpu!9G`L
zyUY-05|zt>{bk5k1p6zHuL|~8Azu^huR*>}Uz;G`5bSTLJZ5s(xWqm8deF(X*4T2b
zu{E1)Uwi**xtD)7`(Jw#CY<p6@$2@EZ8?z!fb1U-XEsYk8k=ZQz&kP@pGX5UZ6k$F
z%*RKa@`5%?XwABU4Jn=et2oUnuAMsl8={-6U^W-1;wPf{M1W#LaI;bIN%24<abX58
zDuKu<5Or0blvHG!3QgG8NO_IMs`8h-)l0rBbIs7VA>Ou%@4=XP`y|u)XenEGMCuxH
z2}X)qOepdVMZ;B&M!wOI(~wIV)W3v$FG2Y&q})RN)cl3Ww-Cz9(U@{HMU__~*GlB6
zay0gh#-RRW?3)bv26nlDL3_)vZyA(VW2G8{`h$sYFlj80MThA?lo3rCX|9e~L#UVy
z1ao^*#n(jh8Y(sfjF}=^`prxY@nN-VBRKO2PG&WTN;Id9=31+Kh--F;v+rYg-K0~4
zlRj~F>Rz#xH{Z(3Du?pgP##2tB6){M-c^+!<F&`2JcD=0fbvkmJXDa`j9~F=`V>?U
zO$BLOyG13G6GLG`BRQ8yt~YDAsUgbEdVnodw5wKgn$=w6{$y&1bu&d45up1^U7(wF
z6pL}w-Mb;O%>*LdZM!9qzHIvyX{5DVBI(P5up*7kbvsP@vi+SsmbCnaFgPoS#lPwP
z-4Og{0=hqQ4&huUc5Adm&g3v>%|e@W%y?}y-$w1@bG-JP%52)w@LgTVTrD<^w233b
z*`0|>5N8|24P@1@NDOUgLnxhPTy2P<Q$bw-o&7V5ozouvKr|nqkE0=i&U%=c492GQ
zk#m;6u6`+qYr*QL5h2bdlXGBag6hna4&i(!^)`3#9d<xY<2$55KFN1DsdDh=WX;Ph
zpSlkbm0?!H?ATq1%7pF{z`i`O`$RzBbpHm>B%etjQaIUXGN36wQvgl%nF?r{&on^O
zeWnAN;WGo!OwCLyHeRCggXRZ7vox~+&DP8YG*>eh&`+A50L|CT2ed%50MH`MB0!5Z
zivcawECsYovkcJBnx6r!(5wLTi{=+VtB__D0=y1ctV1^HJq9C-V93Xi#WBcf$RZ7L
z8L}vYT#hWtAy*)a3dohnq7w2|Y;hHX`fp&18;~CptH;o8dP1z8Kz>TBo<c4n4n@RC
zowUWop&0Tr;_wXebK>wEatU!Lf&79vyntLv97-Xla3(37Ir9^W^w63~Dre3@i8Q>z
znH=TJS!hzEhX=IBI2qb~+T)ze_WC-zPkVwgu=})WoDA(g?MY6CcAqw#li6}wXZL9{
zI9av(*6|kW)F`2->=7LHKs#-(;IJ3+KEYuh<o$x{e!&CUjjpkf4+yRYARiQ54?>O;
zT;m|e3$F2y4+*Y^AYT;hE~<8&sQfP2{SNt(V0Q^}uHcj_bYiP`ot^1)8S)ju=?dhl
zg40#V*950)kgp3)*CF2!oNlNbT;lV`cQfYI{LfJ>b`D7}tv`RHF0#jo$#K$*83c6v
ziOC7b2RM@hTr*~Zh)Ni*4O0WCRG-(LP?<%{>0VHGn2A-hwa2AoWz6x@)I5HT#=S-p
z*z+E`bLIJ@RVN~;+d6?06S%Kb!y<~ei{gJ!SNBuA-6@r~Uv0K7$hA0umRQV7#e8dK
zA%s2Bo$GUEe+d=55?Z_x9GN8^_HNRI0CT(B)O5(=T(dY&wdgEvP!=~<<vV<fJJ6uM
z!*{zw-+v9~g2FEiIG7i>QcpLw+k*DC&`KTr=YsaR(251mcU2{=EodHWZ%W_ck-Qy-
zBjS+~kH9IAn@GM1<;6%ThVmU)-T_@j5m=6Zd<e^jAfLy|d3dVh1C~EP`7WaDB8~O$
zwa=6a3m4STJ4!{~sc0Y@o;rivyNVykI(L7+d-i?LfB5zPz4iF9w|beL_`%hL+GR)%
zMV-|Wqmdj9`4W;ZK`umcA>?Q*M`Msb8OzC#%dlJqIhe@7q)|U{Nd`Y`)?U9gbUuV3
zISjQ`TM~wJVaQ$Oa3qIAc{p+nhc7cRNREN>7^I7V`S&1s50vjgu6tnq1f(P&U_K-u
zzXSx1&Lkow5y}&hUm}!eBRLzEpN(|cu>Nf1nT<YE?<X4#$VS6d&Ovex%%6jFIWT_?
z^2~wxbI^bsnEx`8FT?zok?u0ge+9``p!^DQy#nt)56O8@o`-aKFn=EM%!BpkBRL=H
z=ObM{)Xztr`B47>k{>|*2T1n->OVyCLnwcUTpz;rJx9uO*#FOw-*ediB}gfO@)G1%
z0_FZ#_Qzm;`D2|w2J_1wd-`KAKK^)sKL+CyfaL&~KLG0jVEzE?836MK-~j<Je=wGV
zVg6vO3x@fFv1c&MAB+bC!~7euyb<Q#h;<ub{*BmsBlc0}(?&dGBlItaU?l{D`5l4>
zhu~4FekfK#p*$4(g~Fq3kyws|^+#e|B&<IQ%TZ7sg>_L-9)sl=D38Iq7%1O^<vmco
z2fOaUUDfs_U^xNG6R<7;=1;_OB9td$T_TjHU^xZKQ?M=t%2TnN3gxNTH5Jx>8Y`z^
z|D48tr(yq`!O9sZKZE_wKzTM+vY|X1`(;CU4pwrYJO}&bKzSjS3t{^Tv91udzYu#C
z!ueB(2Nc5jb05q1VgCDAcOT}zk3H|h{P*#I`!IhImWyEiBCIQd`HQe;5zJqN2Nc2l
z#aJ$e`HQix80Ig=-o<eK7ULnsaQ?o;%1hWkFY(})aQ?o+$}1>;h5cT^{;R-p1+2dU
z>ndRVl~}HX@=EMl3CF(%%QaA5gLO49e+~Amf%jjF<yxp;i*>b7zZQGeLj6D@2NJM8
z1rl8#0qau`k%OQ-i0Fc#d=rs3LHQ=4+XUsCiM$!gHxt**1Uz36M2R3^eTX1_5%B0}
zBvB%vJd*fDLiui@?1u8)#BVp0#}Fk3%43LM43u9YCYOk$n!35fBp32cVsaDmEn;#@
zW#+D+Zi<G3_{_K!|10oN=cs6#$+@VGil@Bpsd^Gmq=zu+-`=W&Hc2o~5}K>;IJbq?
zw*@C=RvzxsV_43vNjs>~b(XU`%XMIfW~u2bXp2>QGihGX%*dXjDrn3loY#i)?Ny6*
z2d~`$IfB<lKu+VeX^>Cy+LMsed2PDNzpdRBx1e^HKeege3FhyF&dkO>_F~zw<{Qcm
z{p&-Z$GfFOEWcSgfF3h_C0M=^8vW_4ljsXZ!?zA>6qtXFy2~0qZD0Z#njj|WFx(Z#
znZ$86%#|oAJNV8!)bJJ!ZSb8>s(kAA+v5&<l|TL0_ry*b)27#bV8Pm#Ir7zn!Y}7l
z)6NVP>_Y_|Gq$thhRts3XR_x1HMTE4k6N|sS+|=SdK_2~%i;MlIu|M+R|;Y!<hO$O
z7V>*Rd=I%w5UU`65X29Vs|B$da*ZI?K&}<UTF4&-@gw9<g7^vYDr0e#dR~lE{EbC_
zmDv+Rd$8f-6V`{h@{+%rEYa}Bobr)qJ`zXff8qWsbiw^s%!AhOj}RcI1aO)F4&?u*
zcbJ}f<}_=#R;+w>vYh%->ppp5?WuG6`X-Y8Y*N;Fr=3I9ATB!xMu+wd(Y%3<`i5^}
z*kZ%Zu~Rp5-3Kvh`3b}~f%I3M%psg}h+2MKkhsf6zAIY*u6wutdBjk=@P9pr*@I3;
zpy3M~)}rf%PjBcQy(gOYsz=P!@Bt3HqZ^zydDXdJg6MSjZt88z^51y<Rpgx(bL$ql
z5Z*C__hL<_9xUGRsLD70SUA?D^@U&RJO-p?0@15|kVsl46202zhe^xBM9&=UH;c~C
zN$;IsOrMgOZZjKS5IN-sw;uq_>O8A)OXQT<oo54@>oyl?{^a%(p!v%D?#z_rlm*HH
zK#RN<H6Bagl*L|)0WEc3+IY;6Q<k|e1N5`U&yCwBr>yW;(MY$7ADl|Qf9!MGU4e-H
z1uo;4yz){##?68P)a&t%Xx<SJ$qDDga5e5hcO}l{SFR_U*`|iCqS(0I?mxB3#w}7H
z9k*Uyy&At+xc&a}<7O|8bfe{eMDj=E%KCsthLHRT@*OPS!9ac&%Xd}&^Fiel3y%)n
zX!%J*P9j}d{NvB3>&6_&b4K)51CV0?j3$;Nr*idx*rV>(`@KtSf&TR*f6^jWzqQD`
zBfY2nymmkRe4_DtrdVDZ3;6=Cy#V<luf3=;^BB>mrQtI&wsW%rsi5JjF(!EG##b1d
zHxH&an<X^K5-gb!P`B4NDHbf5Z>d;+@^BaXguPU84=?ZGJF_BbMqbWRxnzWy?H4}Z
z@1gzL+pBluxRvcZdA3usgX6#MEavH@4f*HNXlnS{jZL6(>3Qku^+P@WXV)b=mqLf5
zoM_5n%rAr!LtxA=f^&(0F~5e7=~xTiuDGz$x;oSRe{R9sz=3Iwxha*jv;N&j3Gb?t
z5}TLI+C(Gexk%21AG~eF%2o^><84^k2Kfb6UO+CzN~y}spHD~OpBtd5;cG$G2W&N@
z8%o2cf@}e<d+Y`$k#X#?qn><Xl0mFgb2*+fiRY}?!5$U?pf&z;%QQ87S;%T+<_w+d
z4c`;88mp@hob6N?UO{JO8Zk*D4y>wL?vR*baeNmQ<ZxUL*OYygL%%@$ak393pG332
zw159H7O^LcosKYKwKAi*@&6CN_8lWu{rq`T=WnSixl3d5MqSmP9*sY-vf;w&(kt39
wH#0Bq9$h@TckAle+1;pHqk@sU`jdVh$^Lk=A4W!<)t{c~PcQXn5B2B&0Ubo}S^xk5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp.3FB08C96D8C7394B.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp.3FB08C96D8C7394B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d5ab954aaf3b6965016f42cfbf71039d4f1cf0cc
GIT binary patch
literal 17088
zcmZvj3wTrY-N*a;r$U6PWoi*7UO-{!T$3iv6%_$<zl&hhCe6vMF~F3ZHaY1fZPP`m
zhKrG@@5`{Z&N-*R=BsmaUN`SP;&i&`rjyqN7YxolV{=iX@1a}FH_!9>Jf8OR|9}2`
znl_h{b8<XpgW=INEiG$9=G5jeA82i9X;~otpWwH~=8DkM*3y#N^Z15$ogZl04_=yZ
z{^ZBky#MICmQ}0op39A07kKt)q4bV}d-JD>N8Z@+`A6T~bIR51)76jt+_O6sS*3Wg
zYCCnb^{IP*{D`e3vFg&7U;oGL=UU(2y4AY><Aa?~_<nnH`7i(T{J*}K-1};J;k^%i
zFmu)i$L@JPwfTP!f4Q>j<IerxdE%>WH~w|UM*g2i1D1nPb#z(Cl^#~iJhcBq=NmVE
zH+k%fleK?6ao>u9;rg{(uQA;^xcrWdR?j=T`P&xVtM&=cY<6FL@bJ~Ie~&j@*R}t)
zj&+OnU3l)Nul>n?#ZP__-15-FpIpD^Da9L$?@B$kbNTz`#dmvN+w`4>cifYB>)B5|
zzqG%w`^BX{J^AFDKizZn#sl}=bIZK<f9Uwzrf<Kp%d@=Yy~1-B#uux;I#zx;dec+a
z{qU^0mX&v$IQjFX@z+Y*e|EWN-X|BHo;rEWy$`(p;G%P_E41^!)p2Y5@KviX{?ok8
z-!qTC{`NDug^8{CGqM+)ciq3<zu>F~{{92$%&+_Az4@c3Z=CL5G^qU*-FfG{JAN9u
zqWIF<VDiG>zGV8=9luU*y7MbjH1OgcZp{t1Z9VVKm+!x7dxz`pUp@buwfEoC(Vlql
zjt%z}3&ES7&e>O9^Tg9H$j)ETUOBLN(HkunC^}Ak=M&5F-~6+3@jbTk_CxpFzGd6@
zZ(Hq}7kJdsenC{`pcB47_wHvWw;S@m`|zzm?a;B-@$YHw?_OAY;>AA<UjE=K&lGo$
zot{{3`2ABKhxY#P8;7s_FU#EB(K{=xyNy#n{lT3J?9V;6>4{%12)VDT+>!gmt()lO
z_8*<U{KxqV54WxC{A%Alhl*=9yt(0>_5-n=i}pYL!MU2<+6{+)yDlud<Fm&fy#2n|
zibV%!Rwx(z=#yVuc3X>TMEBU8zdLy0&~u)L^`}1g8*la(j|`1XpM3Evi~POGrJ4ul
zo&IX>#A549J7*Rux9xst_u3!3KE7rBd(WPKeR|FH2AjdKB73N1;{|X1_1kY9-O>8i
zizA;u@P~`uyWtzr_jK1;-+txXE1#R2JMlcPch^;)ExYWE-Tyl9li@w>=iWYP{blL3
zhhMwxio1e;pW0m5uj$$J)~dZ7$|XN|;*Q0i{LCi(=aL(~zwndK-%rWTON34PEUy+`
zTyo6|E4OdUJpba&C%y8;-Dk{E-8prHKi}}tZ#O-<cGbr_Z$J9U|NJ@IUN~>W^COM$
z^Gm<{uk$|r{lR~Hpt|$W#HFQIH+c81TCVx*Sj)sEhi?Acqy8O#dsMDHEBWbjAN)D%
zldZn`<iS_}dEn3~>cq~UJ=^;Cz1vRvJYsucuRrt5!he19&&9iL{N|C{_wAipzHa*Z
zgF^>~Gj#ukd*6O;!=bN^t$6qr|MrjV8((_)=<c<vKK=5R%UhRx^yJ1ryn4?QyT5Ex
zKFL2caMedA{CDA&#20t<K72bFjE`>JeBYB-buRks=wDc-HGJliS4p)`U%cmw`Qe*}
zV=W0`Gk1(yTIPK=^<LK6eZ_os)io{Q#d8)fnB9oa7_<K^*uQOY|Gp!?p3kljr`zYW
zpE$dzPp`OD{>yE1qoMh1qd0xWoHORlZoZzr|J|=v-0%Z&`dV>%^_<n_*-d>~wP)X7
zi(9+yYHPV#oW6-pxQQ<R&zHJ#$^18uJilehhbOhTMdi$nGtZmd)Tb}~{BK9^J1p#<
z&%EOFX>(4SJG-e*TLKB^X$vlVXFlr@r#H^oXq(;Cr@gQJeM`^I_xGRF;u5Eqb}T(_
zc2l43x&I$;?@<OXKB>hcPM_a#{<*W8`gG{p&PSK)fByb_cBMGIaL&Scvzz+#4VT_@
z!ImvY4$WsiaeC#Pm1|}<_34R^UKp9$blH*l%r8!Fp0n9ByQxn<e##?9o~>BD^VtS*
zx?@hqyxC2C`q}sX_x$OpqT!?#r#O9f$Jt9~H}&aS`-Y#r``zChp3l~c(+lP-IB|AU
zpWYgdB%abY{`UK6-s~Se|IFu84;<2;(-!U&Iwz=p3q;IO=Pu*|b?yb02{P1WIg+6+
zE07FzxgbYfo+CNx@&d_G?-IID@5)gZ>RkouLR}#!P*>zgfx4nV3e=T?5_M&cl&C8U
zq(r@2=tjLeN8PA*7pNO`m7qdhl_M4EssgD{R|{&?)j3k5t}c)ob&a4wU6UgX>Y4&+
zP}d4t)U`R%qOL8F7ImGVLtU369qPIQ=}^}Tderqf(xa{~kREk|V3?p4|9OIF3mbA|
z*o9mm!(L#cU_{-RBO~g@0vS;^2`1D{IWnPcDv$|vvtUNuoFg;p<^q{fw+I&0EjhBF
zZYhujb*o@S-I^mS>ed2TQMU;;)NMJkp>8XX4RyO<N8O$yJL>iV*->{04%8hva-i-g
zkOOt6;6&Y-BPZ(40y$B42`<!KIdY-yDv%3xx8O$Iog+8u?gF_{_Xr-;Jvs8A?kSK5
z^&X*Tg6gjUZQ-6A_3T0}P|sdqpWs8?mm?qQz5@AB_X~d1{W<cZ?k|uZ^<JSD_1+xy
zqTXAeUep6Z0QEqQ0;mTH6hNI5IMlfuaj0_z;!qC?LDYje3Zfn?P!RQy5JEkaqY&z$
z0)<cy3t`m5ISQj5E>IZth!8<NlA{Rfkpe|fj|x%LqdAJA9xYH5^**5w^}ZbSq25=Z
zKGb7E4E0!!VyMRo6hl2BBqpf-THY2;<S4NVxj>1%z`Vet&gY0noi7lNdcV+*dVh}k
zQSUEMKk5U*0O|ud8bE!ZKm({Jg(T|993@dt7AT2&N=Tue%25jSRDn{c3xa^UkRt(g
zp+ExaX(5ezI!9^L(*;VSJ}3;LKA59H)CUVRi29H)g!)j9hEN|W&=Bgw!Z7N?IT}WN
zxIn|Gj|d~EkK||s^^pRNpq>#jsAqDNK|NET4C<pKj1n#wr5ueCE*YhQID(7DNEmBc
zG$t+@Yg#lWE*fiEG)}^J)1q;4(RkCMadFXj)1oX1*``HVaZ$EuQC3`(ZCW%z$L}2Q
zz?cvhO*Ac<5Eo4}Et(|VBwhaXn}+Zt87Ikuynd4Wb#DKB|2KE6x?qOd!W|cN$m=`Q
z^*1!f_j!-Moq>H`Z|AkJ%gc3!f#RK88YJ_|z?5+^FlBKWm~yWiOgSe9Qy!Ot+2!p5
zvy1Bjvn$>OroyWLQ^6^~RKyivD!ocDm7EexWn2knx3?S2Zmt{5?szwtDz6Gm6{iAI
z6<2|&_Nu{Db80ZvaW$A4uLevFrvXzF*MO<@YQfZUS}?V7Etop54on@V15+2*fvNZE
z!PIklF!gagm<F%mTG(Ud3}K+SVM~LIUL%-B&IqP4ZUochHGyg3OkkSgCNRxjGni)1
z45m452GioTfN9|@U|Ql9Fs)uIm{!gTrZsK_)8@5-Y2$2Q+Tu1a?Or>WcFqo_J#Gin
z;dOxN;2dB&;tnvKUMHAN&IzV7?gZ22b%E*PTwuE5E->9*H<)hD4W>Kp2Gis9fa&2p
zV0z*nFnhc`*TN1b*AoVc_iSm9&+7xz$N9kY#eHD<y?!wLoF7bo+z)22w-?M_t{2SS
zcrTa%Zve~y7XUL54}i&eIWRen1Cxt$U<SQGFoRqW%wRkSX2=@?GsK0!48=oWhP`1h
z!(155a6AlV#2W!K!bQN0#3Nuvy-_fuTolY`JPKx?w-3xdt`E$<cpsQCZw$;B7XvdE
zkAa!+Ca#73MJ^EriYK--i1+ef@*EE)ALqgB_x6L?&-H`ZAMXcqz&ilu05<^UKzsnq
zq&Ep>l1qY_j3>cNc~fAfxD=SFcnVCxD}X6*;@gD(90h9&3vmI=v^NcAnoEP3j;FyK
z^bUeK$PI!y7#{?4$U6k)5H|$oP<#l?Vec@Q!`v{K!|`D-N4z6oj&LJjj>Jd6%y=_k
zX1ENPnRo`wQSy!wxTD0464;{@A0_a|$UBDq7;$6hk5POK{c-Y+qd!jEIQruhA4fk+
z-YoiA;<D&xDV{}tg1i&xPmuW56n_2*icg?FNf%Dicj~)sZQ^d7Y?2hn-IJuL^BIdi
zH9r2*u?6C8-I8@nto7r=^#jrD-v^)@dv%j^{v@reudcsBi`wy{(G6`Uow}1Y)V2Co
zaUb#DN2TjxJ$LfF*~8HFVets*6m{;yBcf9zLzYjG9Jy<Xx{wu9q(D|qkrKIkin@_i
zQ=~#xPmvl~GesI??G$N|byK86)=!bX!S%!HJ0<gGkEg@!tWf7sxSgFf>iiNp&pNlW
z#;$T_k#)X}TxOj|kgKfoGvtz_^RT3`UREI)_GC4ZVNaH4GVIBUOolyKnaQvxt1=n(
zWF?6Vd-4j&u_v#Q9DDLSlVeX_WOD4u%S?_vd6mhrCof6l*z2lL7xubp)P=pSJnO<<
zSCMsLudB?uu-8>(UD)d?NxHD7sE`7CiW(`fr^qt}_7p{?z@DPa6xdT#nF4!?l0<<$
zWrdX3Q`Sg{J!PILv8OCDCH9nMro^7I%9PksmLy8-byuhxd)+na#$I=xbz`r)$hxuD
zU1r_b>#nkH>~)tU-PluANQFIBja1lE<(Ue5sv=WiPgQ0r?5V0ug*{bCqQaiKLTc=(
zYox}WI?vSDQx};Ud+IV%V^3XWYV4^?5;gWT71Cf&QzH%bG<l}Mo~Fn&*wd7m278(+
z(_l|il4!7}t&kRb+8SxGr_D1h_OwN&#h$jzwAj;DnHGE6l0=I=U4?Ym)740aJzbvZ
zu%|0B9rko(ro*1D%5>P%l_Was=_{nip1wwU?CJANk3D^n>9MCTGd=e7Ri?+Dz9i9O
z&rl)5Q8<lasFC4I<UBKMXZ4pw@ia`48Qw-NGs6+&Dl>eBT#^_LOB(Bq6*6MaSR*6$
zjCp3np0UV`*fW-y5qrieGh)wJk{GdPs*nkLrW%>BXUa1Z_Dn@)!k($jOxQD3nF)KQ
zlEj2PbA`;<GuOzBJ#(I!v1cwaGxp47X2zbm%FNg^mn3HFSt?|~o~1?>>{;^6f;~%-
zS+Hj*GYj@CRc67Sr6jRn&srfX_N+CsV$YgqR_s}e%!)m0nOU)Ctuib2tR;yRd$tPM
zuxG204STjcvtiFxWH#*C%FKp6Tb0?cXDdl;*t1v2jy-#g?AWvCnH_ufBC})9US@Xe
z*{jTsJ$p%F$DX4?4(vH<<iMUI&m7ou6qy5ijxuv#&rxL#>^Vvj2lkv5a$?U}BPaHp
zdFI5Pv&fv-bC#JCd(J9zV$WHUII-udkPCaR8o99N$}<=CTt()>o~z7U*mG5x3wy4T
z#DzU~h1}S4*T{`Mcb>Vi=Poig_S|LW#-6*%+}LxMByQ|^D&)bQr$!#^dGgGIJx`H&
zu;(c=5B5A&=E0t)B=KOcr$RkP;hbboje5RB&a<BFtZ^8rr^tHVMlQ3SBgj?O^BHnU
z(sNkSSnsQl4|~2E`LO59GavSRMdrhvugrYd^HrG-d%lvyhdqCV{Mhr?$d5gLp82uo
zFET&&{AK3Hp1;cc*z=bpe(d#Ds26*^HR{D)Z=UsHueZp0vDaH>z1ZumvR>@<mL$E{
z3sfk8y+Dlu*bC%Y0DFNV3t%r$W&!L4sw{xLKuHq79#<g_dt8k;>~VR<VUH^^4trdg
zaoFRkjKdyRl5p4yRw#(QV2y&<3+7o6d%+?LVlP-`LF@&qEQq~eNfN|fs6rv^g=!SS
zUMSB(*b5a|2z#M23t=x*Wg+Z^N|F%v!W9Z*FI=NA_QH7<#$LF{!q^L!Sr~iaDhp#T
zT#|&b7pYJLdyyJNuoua*2=*dH7QtSm%p%x}R9OUjk&+~Wy=aA^*o)REioIx_MX?tx
zvMBbVWfsL=w92B`i<Tr&?DbWs4|{zz>cd`Np7mj`ugLna*H>nJ*z2pZKJ4|CBz@S6
zRVaqNSdC)Xi{)7id$A&mVJ}u@G3>>vEQY;UNfN_eqC$zIa2_*Jqr{iUd6wAD8izR(
zMV5FQxy%wrkgF{58FERII4o(b=PSfxkFODrJwDHP?D0j$V~;O09(#P1@z~=_5*~Z~
z73#-ce~tRF*Pmzo*y}H{e(d#^SwHsrtE?Y;{Uu31_6906fW3hl4Pb8|&jzqJP-Fwx
z8z{2@><v`e0QLq-k^$@`E0n}uvPMblCG#way=0Lkv6n2fB=(Y3mc(ANBuQc~RiPC2
zQZ-6pFO_F0?4^n<g}qdnrLdQ(vK01GB}oc<LWKnE2{jV1C*+xcJ)y`1><MKiU{9zr
z0eeD8B496Fp)~f=HA-VIoo8w6rHd?$y>ywSv6rs0H1^UZNg8{D6&l3eV2uW`H<)LG
z*c&XeLF^5d*&y}?t85T^gC)rz_J%4nguS5}4PkF6&xWu!RAfWg8!EFQ><v}f5cY;j
zk|FF3S7;b}!!;Vl-f*4`V{f>~hOswXX2aMUuCih54VNUt*c++P2=+#5G=ja6JR8B@
zNRf?TZ=}pdus2d=BiI`$Nk*`jsZa)cnHpuVm&vmX_A*75!Ct1!GT6&hSq6KVk|cw@
zQL2cal)>UC)x;5Oj#8eD5>`j4D2`xvl*-}=mPe^7j$nI~N|I5+`j}WBYpRcl^|7Y<
zm{=ccs*j2Fv8MW%SRZSukBRlMruvvzA8V?Qi}mrQ`nXshZ>o=r_3@_qxL6-=s*j8H
z@uvE?SRZeykBjy3rg~PaXPfF-v7T+JXT^H9sh$<<*`|6{tY@3*S+SmNs%OP|wy8cL
z)+d_k6JmX$sXigrCz|RLVtt~iJ|Wg8n(7l`eWIy8A=W3Le%Xzcmg+J7oz~Xyc3QT*
z@e!rgaPw)W*6^)#>aFChpMGi$-$tk2MpxFk{<cj#dv*L;hQ?W$=4E#Et>5jq=$4)0
zDYC_y#cS(Vc(^V}9Q^))!slA?Y?ZWgJ)C2a%7EljIZ&6h3rHbV04b$Ppl)e5kV>io
zQcKl98mR_IE7bz&q&gtIR1aj38rH*+MyV0VBsBq<rDh<D)B<FcT7hg*8<1UU2XaUq
zKu)O>$R%|FxutF(kJJOyBkfsV|HHL~eNrEgU+M?ymG%M!qyZpK$^iwXL7<Q{1QeEr
zfg;ifP*fTP>XY^X#iTKygfy`p_By3JP`|VvXh1pul$0icQqmNVAQgbp(lpSZbP#Ar
zIs`N<9R?bajsRt(8K6;;juP+~Nym`KNji?4C21CUf}|4-UhnwEiofhyb%OXf+D0dC
zqf_dS!}@hY;#uJ17Y#KYhdoo&vk#8e_@>B*?4Ke(a_<!NA_t}@fXq!1ha8-uAaZDm
zLdfAM3L{6RD1scFq9}6T6!js;rYP3n`t?tz&YQjNsc|zCGnCi|Z*uq<;*tAjs2_P?
zh6a$6Gn7P5%}@$im>~f<Jws{a!5JDv9-5&c<lz|_Mjn}=5#-DaWspZ{M%)F!kuh;(
ztZ8Ii92svK$%-S{rjZG8WCBKffBw+i($@c2tN2-krAyY-KRnPDK4<AUYwDM9wS|{0
zUACrvmud?yU%Grv{Wo8~#7um`KYo2#<1a8tXHU|y`u(x~R4!`AuQO|C%`Z3g$Fry%
zzuv5&HNVQ-_{rsGY7gGBSiIY}x3_oHKQOyf_u>KdhDRP3hnKHgZmR!?(@UOi|KlI|
zFT~+<)}7<7->RQ{loh;(_G}k#`bx`6XZ<H^3$J#s-q3gw-?)CGxAEk-*|qt~`VaX)
z>ai1Co2L!pzkAxrr*+gnkx>5_uXr1dUliB44U@EJl9tqeBhReAr0hTY_s@&@JhkR&
zX?>zCyn|YIAn&5qUC0G$Ei|}(d&GzO@vH0_J;O8^_Q8oX<1`tOP19sTHcyio*)mNQ
zWa~6pk!{mtL$*(o9oaEW4rJ#vIgwq{<U)2&lN;GHO`Zn7)_U53(Wf#uiofkil1|d;
z_21X)AHMgi1Mfa;5hs?dTXt3bZu+J%c;YQP+H>OY!iyHl>-W==?1W=;Pg!`gIJ}Kc
z-$rNG|CuA6`HSuAVt0!ZXIylKy#8sO`gMxBd9&9iHh#OEd}p$8pONu0AUQ7w>f*bA
z6ubgR$t!`n`EDQ;uL4r@Y9I}-0n+kXARVs*((`&C18+z+?%PJ*2xQ_-KxW<yWZ^A9
zR^AF^<8458-VWs89Y9Xr3FP8kKyKa*<l#L)J$z5H@fh^+J|I8u2kPZ}fdYI0h~qh+
zARhz@@gblv9|nr>5uhj^1?uDbfMR?MD8VO^_2-JVFwgTq{d_;r06zef<dZ-tJ_RK3
z;uSGpe<Cr>r-26fL7*Xi2xyoe1{&c<fHHgrXq5O-0v;oN40)XRapWxVS>y>4uOIq)
z{nqfN!0~H(6I+9cD~^9NOUJ)yY2p6;Mchilulz2(c>Jmxo|+Ts<pRM_OYhfnEy2dO
zp~kn7#<#)Ep-p`vS|fHz>j`t_N>7~Mc2fH(Cz}^szIOS7dFD0WSa|AbXDrfa^{Xzv
z<kHpOIsHQ2GS#=v{{KchzIo=7bIx6Q-uWw5UhwTsnOxDWK1;dyqONtPcNi^EW^i0)
HBHRA~5y79I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp.FAAE94F7CEBB58BA.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp.FAAE94F7CEBB58BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..899c5aff7715c4d209497661b0b6c3c5f8ed171e
GIT binary patch
literal 13892
zcmb802Ur!y_Q3DV&MIZStaMmV!3KCy7Lf~rF;PI#7_n>Yi5h*jC)Su~5JUwFf`CX9
zK{Qdp1`<#aFeum&EH8FrL9l@pvHS<P!=B`Q-~WC3ec8D`&Y81Q&&-^i4E6W(^LAIM
z`cCm*GJEQR$wZ}6wWj~gUl2HB35676m1;@i(UCLeUh@#|4=!8%baUv~y8iR^Zs0GL
zT7Hu_$+p5VJa|%v6ZT(+VgF+PwIAQuR_*g`d!TGut9#<Zqj!YW$NQx&-n;ye_lLv_
zaXuARC%Dv7XW^F~iKBX7>oR4)7JlK#3k!ESS5Dy<F1r<cb7bFTx7||rPjK<wYvL3(
z&Hmizog=R%Wp(conv^wvZ@0Vs3Wt4P=Ws8$WKvYX*4F2iPjNdb9<u)ru}JRy#@oE*
zOZ}NUPlop1y02|r@VA2ukK}E0ZGUglX?)WC>Y;)m<$hl7f1b+dER{WNvDx$GqV|`K
zduAVfiq)0T-sS^64)(fo=f$Z-)AVl+Fy7tLX>MPC{|gU)yEuHp#U(rXhV~oubLi;>
z-PD0k2JAamnJxT!=2~X%>ACjHraSH$9y-h7$B<8<@4hWTYXcV@PVQBs-`V)`*2DL|
zNV49#;^NFN?pIyweDkPy<o3iCyCl*!Ijh4c&vBlHM?<Q5PB{1T(v8*A#-CqQH+y$=
z?v2+!RV^KI{^rOl-yA&g!{YH?8AmdYzRKG$|LV;<%O^w)aB?}A^wZ8^5BF~=+54vQ
zhI#QUk86J%_%5SzO7XXmCE*JX_x#iDwacRE-yN@8<~hq_Y|I7sZaprzN8}`o%FCHH
zclJ4h;6t7dSM5)oJM*#Ab18mN+^|;_PhRdjpOExm4u9?V*qP-$Utcc2^zTKLBc@ud
z?s94JH(t9&?EGftpe2*CUpo%9J)HS%g42}q!}s>izJ7M!vmrTGK6M!5l5%R4i$e^z
zaMZ+(6An37of`A6DXNUhqfVud&8xcYpQ?J^rhOX!d&G-RX?H(arlDn1zUYy*==r?X
z1!YCYJ^Xu2+G>^T`W%JDhgv7*-}SN#Ze#X)!LSwmicPQRzw=3(USYkntXGW7g-+|k
zVl~Njo7P5l>-*L5PS%Gm>|4Le){V4E;FkT`X-1UZZ!`9ddl;|`l^VRWA7FWYW%a85
z)4Nn^v)0Y`eULSDP^t03y*Cd0>-SNyErx1COWsahS#{fQ-)IYy*fAsW2QCq=_402g
z6_p*Eu}&}ftLZjzea<I%6umI_dvi!Ww{Cf{<%(8YlFQsiZrtlI^~Q_fJ&DOVEkC|k
z?=@h}fxLeDBWJ~gP5rZe#k{NXo)mstrDb8<(#+~r_NjAz?f<3ePYX=;tBk|+e;N68
z?iULZ0`4Ub6NVK>AJ)A7(K_JD){vbSE$56o_wr!E4;xGzqqLKwj=X&7xz5x7``$lq
zQ~B0~+)XKHu_s}r_je99L)A&o547HV^!~Z1tg9WaOgtL%yzj>jMW>dU*xH4qPOiN<
zj`UkolDcx#@k5-6W$(f};m7&u<u5kP&419ZB)!1nNAur4_WsK%=IqBf!<Of7pO6jB
z!-^ti|L4bTCHEcqp`@2t(XwNgOYF-sg&SL*9lQLzC}8-pIoabKS1uaXan$?|m!1|z
z7+Qw>Wn{Fv(r8D(ZY$SEr@J|wS*x}EZp-$-cPaDw1gC#Bttfu%q{aD#e?2@n-DB8W
z+hvPC&KsS$#WL=8&Zir7*Tw|Ao?Y?peD_4J@4wRTI;YRuyot-r7Pon4lplEE-25MB
ze`nf4rNUpDjo7hTr4g}lXO+B$7uWE%x<sW%{S^I29{)xx*SvEu#0FHQzd?VK#s?D{
zxMn>oiZu69W3_`y-o}aBI163ER18hp-dnNX?W7?#p%#j<K`}PhB}_%!HqiWFt^COl
zw{%p=8Js$UGuI_d#SDMU{Z4*$*$kUeD*@I4HjNJ^Hjm2r;o9<ff2wgWT8>1{B$2bx
zB}_${`CW_J;J~OBxFxkvjSZ`@i7sI(Hf;u#er+@6nHe^vDub;D+cZ9yxLs_Dez;#%
zsUg;*;(iAGOd20dY&ZO^-#Yd2sfL)RVm|{vlg0-VcVG6$nCttpqea|~ibHt25Wc%E
zX;!&ZZq}xAg^3w9rz%6NhuAben7BvTzsL1!wWr<~TT}5WV!ev=&?QX8`*hCIHnwk;
zint9``HPUhh=nd;D*i)FTACO+Z&Krqw5hgo=3BW=x`e3=C>nPDkI|hg3^7No9K{Al
zX_qx6Ol8p0u7N5a!*+(a1yw1)1_fBsB}`@Tulb{o)GTu`##U5i9TC@&_PV55CHvmu
z%{#l@RO9xvEJ>Vo64zCiFqL6lYc`B_cgs=ZR#YXG6I1C3Yf6~P*cs-AX@~3oFvN!R
zMMc=42wUh9rZQpbW$pEn??<%2Mx9i$M@tX?#)nEDqn;`ilF@eqE7wk%>zSp-9`rRk
zINKeZvo2vaW*Xlq_TKeIYlsPb%|L^Jri~9K4*dGv1~Dspy@>7UApXRuKXI*f2~(N>
z+x>fei$X>iV;gE=G#TB!@xjCkPwDj^b@+0e8Vj@xX`GlwtJ;(>mE|9{wGzI%P^`vU
zs<NH4-Ojn`5~i|VwoCSXHVdn92da|HStrvrZAzF*>XP9J1!wP<8{@XL9Jh&hn@GBZ
zsib}B*?#|)c42DlNv-VUY<E&CO$k%Un6e{1xTW#&rV8)n)O%@#n-Zo{H2!S+oXA;+
z)!2zzN#|_SIa!x5mExy1zx)yS`ij^%O7e+*KCzLT5~gxGWBc;UPPyS`Sfa%pZav(-
z@xjEGZd7bNbR*)d8E#F*BdkYsYJ4#9<q!3Vopx<;H^ll>T!;+{=}2u#n9B9|Wkb#T
z2hh&<qQ#8otl~L)UBXmuB<?tC@u+C1F}9^vjuY`XEoM`~RPKa(4R{=X@`522sfv$*
zk8$IJiJz;^xLULflFe~T`jRJD?+F%l2~&CX^@oVJ39ZMPVGC+usP#~r#z(Vw^!Qm8
zRtNW)VLK`wWIM>A@xjFJHZ7j{<kN%#HTI@2NZ_mzXf2u&rc(do(>I5|AD7+&TT%<V
zIm6wYu`XdMfm5f=Pjgb%24tXH5!e*Mkyzey=g&SwkPjj05bDN8wIJspDF^afWcn64
zDDpaFTBk5uT6lxTC571+<Wj6xid(aFL!f?KFOh4_T)ZG(<TMw#9&Gpt@*__3NZ}TT
zd$jZR%{<y#kk_J~Ytabh{jta^7L8Om6^W^+i^2zy%Rw|y;UlQm5j0xi<EZCxXzx7o
zIuGq#L*g}PuNt{lqd`jkm&olU8m#aqBz}VWp}1El2Jc&kd#=Nv{BgK@9R5bhpNzed
zF{sZjEbhXfKKrq_UtzXZV?hq$)IppX8<7UO6`QS)tmRu8=+<eaYmRZ^G0sM5`V*Yv
z38m=;`7CF2mXj3u1x|B;^HBI9r+ElDj~L|<iLDf^T-L<@D@MM?3-aK$gFCY2SdatS
z26SXQ06`woc1XwWx}{Q(hqfKskxd*y9_~228}k>fez4Zg*lkmQh`fmxHu37_{f=(G
z!5Qdw8Ejq%)PGs-<=QC)4klfKN%v;oUcXn0@ODZ=g51-qXK&UntZ39;!_I{LS)gvl
zw-EPZ9!QX5IL8>S3u|{aqG^Y`$Ljakf(Z<CTNl<LmRI)cz2BMR$`QGg)0Qd?NaSE*
z5lk#u`G{OV+82-x%&9U1v?rT3JWA1;91v3wXhXEm;JUD#m_Qe5r**`Y^^8CbYYRwk
zW>}ERh*24_U~W;6Uy^n&i38hzu&yHpjXNmiwFTI>0QY0N2SJYJ9HW&Mp>v<>UaAbJ
zw(1q0VR~oo(Dp9kq$19Nbqw2zVS~mkm|{PHQzvj1%oVaN87)S`eoRq|=hX3>1@rr~
zHk_tRd95H<a+*qo?eYSrKMwL<MF&z5XI{ivv0~Ww&dROzxM1+v9piK-eNoLH{nWH#
zK}atMshf7&DwXq;_g=PmSy5{uZze9Al^$r|QJcva)+0oYBhGQ88{{EzWSGKRh~pO0
z1+xDZGEm`o;uKFbkO#(-K?<i4mo)hPG%_j;zJEJ$+z#z;C;r=^{dD4#4(+Ftf$7kG
z2Jy}yUnuX(B;J{jvq-Nj(woWU2g#Yzq~f`>4}ww0VAP4ZLD&BG!dqC4ai@KdL98-}
zHOoQdOk$M@Ig40jDa<zHbcJi$jkDpwd<z{l4SR6*XF3M4ZxQatwwr<+%Q?pWeejMe
z<tDJf!M5wve$!r^8S?0}xzE9Oc1HhI70>RghTc5B`<0m>Z$Z))<j(3S$lH;2I~uBR
zHj=WT{C6b%4!Iaf#gI=T=_KTfNV*952GZVu_Maf>36wua(sRgfkm(z^)W1Wf?;x+i
z(i#lPzX3}dAaBIdM#%A4ivK%%#bdCvZNt(wDBq2>yD?~wJy^R3@<FUUsBq6VHK%5_
zm|P<gc|9+zS5^WdZ{URukRy1j2wq|<1(7%MRvRHl@>Y?Mqj>8m-bT^i#1GoU4^}vu
zH;(2_6^`L8Vt6ZsV|j~M$eVd-GjFTN<9I0!@)q7@3-7AP<N5CKd=G^ac$WkyPvV6n
zrGDP(UVlvV>+qw#Acr9_40TsVR5TK!A#X+ER>=F2xF7OaB%XzQ6^U0NKSJUo$gx<A
z#o&EAvA9!Vws6qN(sXKNlcisIqF2JVzVibbO6(SfzEDcM9*OIbHDuTI$XnqEBt}5_
zMkH>89F5$fp}jaH#zA=;a*c!b;*l5+<q1ekfV>sCZH4yIkR}bznRKK{hr9!6c0m3W
znf!`umGbXLZu_CVY$RsG_Z>nmhfq&N{}B4}5E`Ix4sywX@*MPK4wUC2F&Em)Lt-A}
zd?e=oon7-`dmKY9$DsWJBo;vZLL?SKE=FQ8<Ps#7KrTgMDde-r?JTT+84}B&ybQUP
z!Sa+Nu^h@PkXQlvDssCD?OjKj>#+TABF#<6w~*!*<U7dZ4(y*t$n6od_XLSg;QMNj
zSOd8hiM5blBk?ulx5)7=oH1{a_gmzr^k*G%tb_78<Xs2l^+>FT_TD4$J>(#)34%T}
z2!9=ffsb8*zg_`37;A!|ehAitKwgc-)fmjb)!2142J<fri(yb6j>T}u8?d+mawHZb
zAxB{`3UVxVi^X96$73-b$`i1d067thiI7vUnDTdaO@ZZ2#bPRyr(rP-@=olw6TW{p
z*6fDuk%2WCkTbC+6Y>G9IRH5eYqB6`V@)>X9Bh(<oBaWc+i`;Y39CQhHte*}un!Sp
zIUB|Xx@aET+p}mM-Oe<{@R)0*n|kbCfcq%B@j~ogi2Eq#l_Kn3g!`}!&!E9tixS_s
zA#~Gv3b#LnJF$ZzGeF<a6bV$`ki_XFaaMzMaXvvV=iJM=er$^*$W@$smBMW6MpuER
zcp)>y4A9Dd!0Hd!o>jKt1R+Fo?HC*A!iTK#!}NWR&Ymi6pk-de>#yOB*h<ObiuCZc
zf>W=6r{SjPBC|038{^KEk=}eeK_2Tlwzo1J$9aweG~RPOpb4H608R9q2xyY$BtVlr
zCj)wc&0pZ=o}{NLZ2l7RD{TG>@@s7V8uA-#{s!_}Z2lH<9X79nyq#;coikxkMS;fV
zTBUO)ENmyx2xzMvoCyn#3N!-PYA0vHf`jz9M(j(8gK`QyMeI*OK27XTLq0?7&p<v)
z?9W0zN9@l*K2Pk=Lw-bTACdNKvp|Orv8{q!O>C<nKPI-1AwMCuPar=fwof5HBeu^J
z9yvK?%${x6deXs@!keV<=4|kMdoRv!isjGk>EOA*nO@+!C<FQ-XZlcKHg)LCXo^d-
z0WHu;s<)LhW3yV2FLCadxW0;^$DI3Pg~xd(n56QRbN=hhFh0j~*0@d`+S71cHMXtB
zj%?ggH^Etc;ySR{fIy?F?ej@D<p_V5nBFDrSYxu1(Ia(3eB8goqP3BEeU}TekD1RX
z7TcU~@1eg+Sa_JeEfkrBq8`e0$UxQ^a5oT5%%VvTHlt|^66+MwlQA1ov;{*sbtu>D
z%?xx=cUBD+b*IH@2)VOT3i3&;cM=;j7sP5wJx5dMorQE*0cmWeA>gj$KZ@;-;vvjA
z(%mIz9me@HKf+@1w0yywIvBc#rieT%ADcT=F@#fxa2=Ti16^F6DXtyBt$Y&k$)k@T
z4>B8M^GX-Y739$#qk#t)<1q%%w;taD8tXF_L?_4ji~}^@XFQ+@J`(^<^qB}~lFuYS
zlYJ%wnyQ{^##&5}r>Um_ny#J>Xoh+Qpqc8KfM%&@0h*(p18A;#E}(hpd4K}dfq)jM
z7XVtQUI=KhdNH6S>Lq}ds+R)#N&OR`pVdDD3P$Q+1o$~Ndk(#04K}NRT#L<WA-}+O
zFR%ma@CI*f_Y(3eZ1)QCYi#!#@*8aT2J%~M_ZD&;wyT4@jnmu48L>F7KtF@&rE*4W
zB@^hkBE2-uhz(sDG~i6Pa}pb&4c^@}orckLA7vW6yXg*2VpFlfyPNLhB*nX*Aa*C9
z_dQAMPC_mrb|sKYiE}A&gR6n_DafaZ^J&Oui1Qi9XNmJ!$mfXjImqXU^LfY*h|L4V
zTMP0-V)GDkC9$c5{D?R_BHft7Ygj!TsvuVrhib@=iNj;aPl&@4$WMvGQ^?PV!!w1~
zR{30>J}vP5X9uR(7d*Pz`H0mY;btd>6Jp?Ir|G*Mb1_kps_yqMa|Tgg^#3nQ(Pvw)
zS?&CAH(l5^5OD)($J}zv9jAgX^S4(~K{#(2&UaudUd)?G<AaTCE>Xw7mvi3Bc`6Il
zUT)xCZj{17#4?D0WhaQZ2a!I^LC2Q%-&wwDxrUbUB{6+T+A7@<%A1DrZCO|@scvuR
zRg1N@Eod2$v;iJ`(vUU{en&Zhq!Unn9ciyac?6aslu&~pAHdoJxcNIws%@co{rZCU
zv_u)mI|KPE4lspyrx1VEIP9}2U1OTQnX<)%%@}H76{lXswPh~BK=;X%b<p;{*KGR~
zy-WC?mH%zj#V-tUW;)Z#uSe2)Br9c#LsA^%d?e*VE<;io<keVOjY0ktETt&CXUSnc
zV0NnC7(tFeQUq$RyeI-WMIcv&HzH{xly5}N8{q~Z9!c>~9*>;jVg3XpB|v!sa!!Ex
z(~&kEfz>V@`K2T9wQC2`?tt<g$ZrRf=OQT=+RsH!x$ynD$TJsxrPL=E_0L5C3g;my
z59ZH9PI)kY9`ekC`SVc!Jea=#Nd+)}0dgvU`3sR$2<3&yxe(UB1W6@OUV@xTVEz*1
zSpwf*ilkDgUy7Vcp?)dyEQR{zNGgZ=<;bZV>Q^AC0?I3pa|JBlO{BdE+y5r=y9wL>
z7Si5=@>|I77L?Z@sRr7wK~6RB{WZw529~!5^{;{LTZ^Pxn7<Y|)x!L>$g>vauSNZ9
zVg7m~)x-St$f+LYuScHsFn>MjUk~%YN78$k|2=Yg5A(lA-tUo*GM?U}Z{8yiUJ1t9
zU<}50F#b9i4_EX<ur>tBL$F^6{Ms9ar7#TUXBc)0!(e`fV<{ZU!?ANX?xEO^#8M=b
zM`EW)m_HJGMq)5OqOcSN^`o#;6x5Hxo>5Rg0ZR!`KLI->K>b84B|>>3c1ncuR4k=J
zc`9~Fh4M5kr9pWbc20x!%fQ+U*nSz<F9Wt;Ce~&`c_#MDgz_w`&4ThQ?3V@Q*;tzm
z<=NOT8_KU>y(?H`2NFTPiuJBSzJ>K}LB5UkZbN>6^&Tjk(Y<FtanhvCbg9_Q+3e;z
zvjYu1XYi&W$~wswUb-qZUA|$8`$Lm{UGO1k1zpQG@kX0?nXTp7)l0K2!fSVZ_E*b8
z?g!$2Dtp|Y9zw%;lW^rce5B;km<+EM_dna>vHJTthka9AX&H;LRWbgOH6Pog((13|
z)GN8>fRBOh{t#PIkN+Alr<I@H%Ky2fvU4wOr>5&gZ1q0*C_X5u(!Hz|kt28^Lit$I
z@bO332sx4$A|XfdLKNg^UWkSq!wWHxV|gJK@@8Jx3^|S$;vjF~g)NZdc_AKh0xu*$
zPUM9|g-=OK?&>d5^Mw54XiEOkrsN-IaPp6B47*T4&)^NWuUNwg@(9NfGP{0t-n-M!
zLkHW$e)br+Kuq@q75uWAjseeBp4QBI+2w_!y0km-3*7}5;5G$V!=j80*DBf+Vhs#a
zwkg6I7HcVcP&7BMcf~_Bk*B&$ZQl71d78^KK-1;v&CZj^Gvpb7X1dG-nzLMH0h*(o
z<H4LNk>_gX0-EPFuQ~or<Up@LKnq+KG=Kdf@<P{zfEK$gZg$i}UgEZ-nJzxOd1T*>
z>;ZImhazn#yqy$Hw9(4<_DlV~*IySQuA;5z<>l3UHHGz#fkES?IM%o{s)0HO!#(qQ
z<hvdXVHFePIOH1#IUo7vLq3VLCsAKTzYO`7LHTOzyBY&~8?kmH2Kp)3H${<OdVBh>
z*43Hrw3<(m^b|QOhMytnnZm5oX+W*%4xdt^WKJ)cvtoxQL9XE3E0oYO4Uci|)e2v^
zzh>l)az{klYk7y|&0o>jl@MBxcUb)n-pOgY7s&h$JLOV$(sT`w737Y``Mz)VSfbCK
z2Q0>>Z_s#eklAa9Oo?HfB|B3zT-CA4;OvyEI^~?EoO5IAKtshf6_77+noE$YI87Dg
zYEDxP`7x(?tnfX{bj`*GMqTO2?@!M4PtIL=(KXKX8sr+zwFYu6=US`qeHZ>^OCyW5
zv}x}W{kz0OIr)X~`XRgtJNZ4-<YsiY-JMDmuX55=PFCLdjFX-zTs73tdXUeIZL|-)
zy}X;pJZmZ1HGjY5RGXN4tlK`j4ximMFT_WWL&65T(TtyW$An!>qeW@B&dK_e1^?)u
z8?JG(KfhjfV!3&Jq0wi*=JhK7-#eT-@an&!!*A;M?y;zvy`DPJV@Ns%Z)gNzZ4d^b
zzZF=!LSc4;iuOT0R@cLM&~&Ag%>%aEpcAs`J}Ik=Ko7~CZ<DU<)Jz>1-!7Ez%zC+j
z9Yc8+#`X0(j&-{p`;vC%DXe!2+p(>{2eL~zV0<Wv3ifkkKi7)gl~M&xoi*8qlFuH)
zyQa^68HHGofd$1?tdwjwrDDZpi;jvFcYfga^CKE8J2#tGbyI$JReriO|HSg>WnIrq
vycJIjd$_wfcXf5?*1d~H<=U*E(kMUa{|CnYbzwhLs_x3q9?DNQ<>&ta*$|90

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.9102AFA84B67CE1B.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp.9102AFA84B67CE1B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2ab0852fc75691915011e47f09e38435a8753bc4
GIT binary patch
literal 11566
zcmZvi33O9sqQ~=nloHwkEkdD{3KlF1*>`1=?CS+9HhXfDn}Us$ARtoBh#CZ<FeqxQ
z3hGmMg2D(s2XNuJfD1Y=D2{-#1OtL78ZnG)&Nyk`x#a)Oc|C{k<d^TqFTG82@0WYi
zYBQT#!~%gKW?R(Rz93Q|5D4l@|K=~4*Kxj&K=4tuK(J`d&bH2J$91*Y=ii(*{(yMy
z&Lj5Z{rXi8_kHD|q1&Gq9e8d`ThoMD6aMsOc=7RfrVH2XUU<IkmBD*g%sTn>0oAo*
zuHF5g8@B}p|7^VPne~UJ8J>Iek<;q8-<dSwkK~^0te5Uwb@puEN&a(^ztw+ine@?x
zbssP4h~9U(wqg6Qr%r|c*uCq(WyiJ)^R(QoM=qWgHmnH#&#>8NFP@1kJpbL&wMRys
z62Je|f^UQ;RonZ1et7<#oxU&k&3R8gXUEa5?>25RbxaD!+3|BLRlzp9<CPVkKe?pO
zxj_g19zVU~&`;vSO=C|ycE`$T?*#Afzle^$&8`YR_}gO(vTt1or4Bt5cv>|0(6KFx
zR&Bhi?<)JJ8*ZNX{gAyMt-h}H(~1YrzqhE}wX1&Q<m4r5^A_L6kFUMXI_JT`n?Gq>
zaMAg7bN|&BR#r8S9o!l^y?@1$eGeQX9}Zq}^3VPLUfle7TX#k4P5y(w#0G9#F5f_>
z@9O7Hvug{+;Z5dSH%xXVulGH*^{HLsAANeGP`m$3mA`A>!6S}06gOA>yycpEPfj@|
z^3H#G#|*lj+J1h0T;`>|Gp;Yl&yKn6h3nF9wgtw|II~B-o89%7C7Faw;j~p4pWE8m
z@a4@jW{O(g`1Z-br~mx=-qrKgPgvS#(43W-CydjpB0HGcakFHMtx7aV|H#m$m5m$s
zZr*k2@w5H=XLjGR`OWXo-*)uqxOjWpM{{bvo;xeLV89vXWLH(pyZ#1UYu9sA$8MNN
zSN`<3m9rH`+lwEr6qq({<r+^`UU+zuVNIrco#rpy%5@We4?a+LP~Y}d+c&jaI)dWY
z-ame=?#UZNC%-VZ%epV^eC?@C9sL3)zPRd{<+olqILa>6-1k=Zv>#7hH<{b`&DwNZ
ze$!*?w?!X%ao{BHysaPnY@F*GI&RIttD;x6E*-J+1>2@In@8B!&wKfeUnP<U&i4PX
zLEg}`wq|Lcz(&uzUwXDr8T`6y;GtPZ=A30d*A@56{d(mxXTwNc%cC>fP2A~KV;tIF
z-wrnAm)Jey$?Uxm|Hh~HXdm3X&|f!g{5GfM_}J<`w`BT%xbUUy!w>#GS$9p%>DGp4
zUTs~z?q7^6aiMCauK8aB`b}FZ`?24pkq2WBH~!bk5uKJtzE$S`)zI}t_OqI0KVCW%
z{_@0cKWx1(Ft&O`e%ax#$9?wr<%*Y|{nM1%=a+vK+qd^zY@g>D=UewaZu?boeED-v
z?)vXlcfWcm_1AlEx?z8({`3Rk(73h>SD!dIX3Crm&mVs4-e<cve9*Aieff7P+UzW@
z-Tc8Hrw6Ha2aoLh+T3+8e6eot&UdHOP5t;%q5ap{b3=~BmCuB?es<IHf7n0kXsNjG
z(apcre)7HWsWk)s@$$Df$lq;RdwkW4&SNi=$69t=T)mhq{B>%VV18F;^mQT-^xvFb
z95W3UR8zlCz)h^2cysSi+C%jIR9$-EXv(3!wVJw1(~Xslb-e>WJ!r|B+jsBTGrLCM
zEluCv_x6ds13x`nc*n%bHyW=RAedg79$7!Kxp&~F$6W4nfy~`}_W;2yrRnDS<}tkk
zKRtG}{H=7qS+Q!`R+=7OIlQWO;HRftKKs*#lU=V>Q%`AnNac`zy#qfT3rKggOxS*`
zn$9Rq533wj)jROhUD;=)Z)`ew%>Y4bX?jHch%vnbKb>M5&9BOp3#%zpny#&^t?C{4
z>4&GZf2t?X?yjbR(sXNOtD$$`r*{t8_QQu~9NubrTWPw!vc9T!;HS@cerPzpZRA7M
zG+3H$scgBici^Z0S=;vdk;Lwk)%4cVbX{d#|K5S0p4UEe{{1cGXO33&K4U9CxctJV
zZzqhZ;qD%NcaVQ#_~);ys`q&-OQMA69{%^J;Y3N%G6I+;q9=gG2{Fjxq!?v!nut-B
zBqSh9k`k09X(B;cnvjAlO-fOgrim0~SwaS~EGa`-mL@Wk<q0{+@}wMPd78*kRwNW4
zE0PM76=|YCS(#9RtV}9VR;Gy(WmQ52vMQ-US(PR#l+_6}$m*mTWp$dUQPw0hAZwBu
zlr?FhL0Oy7f~-wyQP!r37G+&R2eK}yLs^$5I+XPZ{XP6nq=wTc^~(rgn&_VZHY5xn
z8<GZ;4QXOP*_bebY)l$aHl~RYWmCcgvMFgo*_0+Gl+6h<$mXOOWpkRCQMM#3AX}0a
zlr3puLD`zHf^1D%QMRUu6=hq(2C^+_L)n%lHk9oNJIMB=9c6o(*im*Q93VTA4wM~f
z;y~G%aDwbiI#G6}i4$d4!UeJ`=|b6+CN7kj1OqaYWKd?(ghAPp@Z7`iS8F&=(zA>J
zriteXU~j?;vN!2P*_$R_lzj;w$iAcxWnY^3Q1&PMAp4Vkl>KSqM>&uPfE-8$P!6O?
z0OepJ2y!qPL^+ryL6k#@5Xhlq2<1?kgisDA!XSr}VU)vZ5=NO#upqNZ7G*X~Sd=4)
z2*{CS1m#GYL{N?<q98|;QIw-;5=EIya3FI@4rMM)IFw_F7|5|?4CPpw#893qnR_FD
z{Hx*al-vn4PcrWc@eZl#U9|AW#@frTUA=1JxR*!`S1+!Y^T%?2Ei<sHcMZearCnY(
z(XYdOTsJWw8@q`S+0;!;$mVWhMz(Yl3$nGFSdner#D;9|CU#^;H*p|4yNMIo)lFQ;
zOgCZ5JiOt&W#^~o>PlPRWNdQq%OrlOR@(ZWby|7rJsILzhi}Xb@gn;&#E0z95I=Gt
zLjuUb3<)BKG9-i?&X6!Nn;|T6Bts&|(F}<qa~Z-R$1)^VW`4<8T2u9`H_MwjTQXb0
zzh3x-=Z~cqPS3KlY>DzjbR5=Aq9PHULrziA5(?|AEEVlV&QZ}Z<N_7_0xZrGG4{np
zBF4TrMa9?`XQ>$b;v5xYUtFMK>`U@Qf_+JmNU$$SQ3>`XSt`N4Bu6FKmlUW3`_ep-
zVqaP$QtV4pREm9RmP)ZN%~2`#r3EU*zAR5<*q0TF4EwSam0@3&r84Zxa#V(WS%J#1
zFV7P>_T@z)$G$v8<=B^JsT}+A9F=2VUZ8U9EAm8veMONdu&+o_1@;wLs=&S?M-|vt
z6sQ9G$~;kGUs)td>?>1LiG5|3DzUH3Q6=`31**ipDo<3{R~3m0`>GUGVPBP{D(tIr
zRE2$2fvT{t&J#8E)kUJlzB)zK*jHz%8vE)TRbyXWpla-E@<fAuO_6A@uSro2_BC0m
z!M-L(HQ3h_s0RDmJkerbTO?ZSYg1H<eQlO%v9HZhE%vnqs>Qx8PjuMV6^Rb}x)jx6
zUzepi?CWw=hkadv>aefR6a8^G)ai>ve-1fC^-Cx$|FcxT7dc1u$B+wD{|m4oPYl>M
z6o~=*h7>hm-;kvS>>F~_fPF)O8nAE76C?JGMPkIhF-48oH)g33`^FqKV&7PxM(msN
z#DslQk(jV=N>LN`O<8Kfz9~md*f$lZ3H#<eF=O9cBxdZJQ`C%obC#O1Z_ZIO_RR%q
z#=a#_EZDaci3R(X6t!UAlBE{xTXNKbeM^B_uy4&1EB38LV#U5SMXlJkW~mkX)*Q8B
z-&&wn?A!9hhJ9O+*syO)Q5*JcS!%<+Ek|wGw-u-j`}RDsW8YpRcI?|z)Q)|7mfEpz
z&rv(}?FDMbz9UZ@*mo3(1N)8?bzt9-r4H;na@2u+M}a!9@5~b?_MJuI#J)2{o!EC~
zsT2Fo9Cc#fS)fkryYj?^eOHmVu<uGy7xrCQ>cYM&M_t%=6{rjQOr9{<XNrWuK9iyh
z_L(eYu+QWugMFqz8SH!V#B&@Lvz{XHoI_4g&k_osO0(3n7db~g$B+xu^9!&yPrTUo
z7Ks=8-W2s>-<zdg?0a+6i+yi_da>`z6Cd_{MdHK0FGYRW_hqRM`@S6YVc%DvKJ5GR
z#E*S{k@&IiPf<Vi{aNbAzCTC(*!LHxANzqk31B}^BmwLPQZ#`5K$ZruAIQ-F_5%eP
zz<w}Kg4hohNf7(N6b)iOn59AN2Xi!t{a}Fxu^-Bl5cWew62g8cMMKyRWoZcep&Sii
zKUAP0?1%FtjQwztgs~q^(J=PISsKQEI7h?S4;N?{`)r=D*k_A`#Xg&&EcV$fWwFoZ
zD2sizKw0cZ@+5-&NRdRaA4$;&_9Iyu!G0u1BiN4=XaxJwJc(jIS|m~IM^iM4{b-g(
zu^-LRDE6ZT8pS@BCmi;<BH^&lr6`AeE=xJ=b2-XkpDR!f`>{NUVLw(RG3>`uG=}|H
zmd3Cj%h4G2V+9&3_ops9Be-~O!8d&>xVuK*)y99Et@u9|3@W&;(OoP*=hCiD)QeX<
zS~?rhsB1LvR}*G1gU5dV*u=i2yF2=K4C3uj`+n`=Pq^F0mM$m=MfURNT(M9LBoRu0
zq(Ui>Oeh1A3*|rxp#n%LR064lDj>B`4WtojfV4s_kWQ!r(hK$W^6J$fGyoZeMj(^W
z1Y{PPfh<A`kX2{}vI%WKcA*`}A#?yag-#%s&;`T@86c0)V=o_*yh1OKPv`^k3;jR=
zVE`y73<8COA)v4@48#gqpolO66ct8+I3WiV6UMGk=)@_TpysZ(`;<;IH8<(`(@Zto
zsOC|6{#08HH@bPWo<H>zUbH5`?tJHX=|xvtS38`)LdRd{E1k0GxyV;O8FPo^jw`2Z
z_<rg1RnMiqvLwn7(K<LoBhC;pvLr(!$kGgvBFi#FhAhtzIkF-{6v)a9Q6j4{M1`!*
z5H+$ULo~?R4ACO%GDKHq{%=yc+ShZ*5;l_&wU*zGBDcs7#EN2N5;I~j#cnZ}Vpa^M
zgpq(LaZA9Iuo5t(j1)|%TMDL>m4Yc_WMInNGB9PV3`{vA2UG5rgDGd_U@8~|m<qQ7
zOa-d|Q^_d7RJxU5Dp@6%Dn<pS%B=!Z#j3zmGior^ZZ(){Rt=_x(SWIOYrxd78Zfnt
z7EG;M3#OLUf~jM4VCvjDFm<dBOg*D-E${mDZoMCf)yK+YU<_ay+y*cWtN~0TV+7Oa
zHiBtnjbNG>6PPBq2}~1f0@KWx!8E(gV47Jom=?wYrp0Xm)52Q7v@%vOt!^urR@Mrp
zjj@4gbKAhQu{JR6j2%q7+YY9kwS(zk9AG-!4lo_8157951k>qug6U+PV7eF=m@c;q
zOc(0{lVKPz88-ta!!lrc7*A{YFynE1{6MTHRwggw1=H*Hg6U<wVEPyzm_D};Odsn5
z)6e+9^t=6F`dL4i0VV)uz#RZHzy`n!GC?qd?jV>!HV9^j34s}MhrkT6Auz*C7|gIc
z3}%=OgUK>1n5>%xlVw>jBTNL$h&uvigpGh1Wujn4-BB>3Y!pn6;lSkF9GD!-ff-|B
zV8+}rFk@^C%uY$?l_PEqH%BrDC?QE)A^w$8I%M~Jb*6ln!lmc4r!-BB;O9N1!+ll1
zs-gV1UtVl_zxMFq1;3Z>9%USL3%^u&^~Y3~_uDmpF0Ch8rnZ>N$FqsniLUa#tJTpu
zoqzOAi{9=#z4JwLsWfEZkb3^A4u8F`^vm^pai{#t#U=48Tkye*$*0334}4sDxCIpp
zn)wBA4Y#mjA<&|VMOWzaiXmI>e{b=!(!+I(>Jan)W#6fz?`%2pS9|I1(Z<mpezEhF
zh3mg!RqfK!-NVGg<oqJ%J6Ydz{RRztq$e`8V<aE>p|!fvW$gHm(%r+w!{z){EdH-h
z+LWFz8I?B$m;d~?jjpfi{gzQ#62(QGe4&OD$HhRBxCBTVmjcP+G9Y<e4y1@HfRu41
zkSeYMQpeRmnz#l?8`lEq;yNIGT;ExKz725$kTGrqGQ~|m=C~Qi61M<Z<5nPB+y-Qi
z+kqT$2aq%F1aie)Kunwg^29xz{OhuY^TxeEzPJy_ANK<V;sKywJO~tuhk(NIFc2GO
zfg<q;P&6I|;^G`oEFQZ;^V(-edft`u=W+9I?YR-|8UFW&<!=ZCq4LdMy%+m%Gf!;G
z>Gvy6SH?nYI1&{^%UY~_lk1GlzP*HwnXWz+eJlG3`&ZWts2w!WR%bG_)K%H^gNI!;
zWO#$_#t9QAO`dYIWa#*tMr*Dg`TyNpfAxr_QP(z)xo+$Y<3v)KT&dPxqiPf@j6>@!
N_I7HXHq}a;{{~ww9$EkZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.656EAF4E8A7F404D.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp.656EAF4E8A7F404D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..859313a10d4013b604c653674edde75f22f009b0
GIT binary patch
literal 11564
zcmZvi36v96qK0!X4OA~mvo**fOCt!KeMdkv`yOl&DtjuGW~u2W3W&hT6EO<p4Q_y;
zAA=x>jy^XOcf_M8t4vT)Km~(AL=+587)1r&3;j-+f6jS*4n@AYUw_cGbmdm6)oHVh
zl1U`yq;p|M+kCD@B9Zj2{Fyg@?(CntNhI&oOC$@|@0ix%KWS{tZ+z4L>@%Sgi?=^|
zQF8ya{p+XJx6GLwc;VcrX+tN>obVrS#+ICXdkVAug9X1#`|mynSI+$Qzm8~$L(;xZ
z_^?K~r+i3y{O&<7*T|nba^asppS@JRZqmjFc2C;9XTXK_kGCsBLx%lp;K|k3o_sR<
z>p64sSTww0$>qMiVkO0y;)lmqZjgQZ_vYl3#k$x>yH0Og{`v8>gBLHztbe^``k~p)
zovV*Et@>s|?YEz;TeF~fXxp+gk4@ijduYw!ZBORrU-#G}>!dADcE4rW53ar4F7!I`
zsc@$A%X6}0L&u!o`^xSyU&*)BE?i}Mx}P@WI@f$={hdGU$juorx9w#|d*=P(q^Az|
zmyR3z1YP*dtrG{=ySJ>Vz2}!LKkZ+)>>rcU7ccZ4xi$FAo~5mIGwv-uzjx65Up=3X
z9NYck19cO7j<|K&nZqj|escBt?)~cb>E=lv-;lchwVk7GI@8X*zk2+e)|EH+@1slK
z-oMx3Uw!v|%xUlc{?G~Y$8W9NF(kV5fT#Nbc3VsPsZB3!UpuR9RKEk<q@NEg*|Kfx
z44-O~(KUY6fwtSE_WGx%e8m?>+}2vZ;>~HEmZ@j=sXmCm|H2wsA)f49#V{LQ>bT~R
zWSpj{?-%D?qd)lL?4&K8GY@{fW$l~?a%*myQpdgFYaBOI=_B29JvZ(8&j)P}e0usY
zcbQAm_e?`RTld$yW2eqNzj@k>wzIq6T{gb&6Mu={r9RkN`{6k6lbQY7JPYpK-~7?h
z{DBptPe$~yW6aCf>lS$4KjwJt-aA(u>a+UO-Tmh+TeEh_^iOVYAMvcv(!m^k<%Qgl
z6F)Vs-#?_KJZrOhpy%>eN8F#jvU=s|kNd7$aqGpS!(EcWlOqTI{;zVMjW10-m$$Us
zaX9{rEVk)sb1(Pq=|>-9e9xNhKg3b}QuR0Ab@wJtwV#?YZq0=29=+?1+LeALzhd9s
z?~=+PCtSOJdu;bxmmhgoKfZnAo!fU!@D&e)>>XqOHh#<gi|;Czj%sP?<6bgq)?2*`
z=e}<oW7+#^an+CG=eKBezb-rU*_;ng&bA)<_~O2I?^)Ys+WO^+ZExRG=(#5|A~4Q8
z_qMqkJgK9u{O!g|gP+>i9)G7-dE(Z6{rb-stoq`?ruS!la9Q)u%|q8*+_(PgWMs+j
ze_3}dI_dC3KO~p-z4Y*eP}`!p8&AwSRN7ta`_4MgZ-EV-LFU!YyHaZBS8ap4J$`EI
z>EiKqe@r>GcHG!phiUK9jp3csFAo0Z=(UsQJiYPrw>tjz@QR&Z-#z`8OH}&2XWL`X
z@BE)Le&L$6^M842@TpUO?0;zCkC#hX=7z&dQjg3umxI4)KDu=8b%#ehcUQot&i`a?
ze|0Uf+21(c{`BNp|Ds>deK%xDP1}&kwT$+?N7r7e=fAr#J8w}(;$<R{Fq_9M7$00V
zte!@K5`J`#(Z;St<@ZC^SKV)qf0c3TZmp-&E5n0&4C>vr5Qlrem>qQI3*WzAPg^U)
zjXfG`yB6Z`m{qE`vOQ-e>*=)0@W37e>$(==@b%tZpH14ase6MYP#K=peb$7og*ZI%
zb%k+%JiNS~hAYFZJz7m&3vqbz??0V;`rAb>)>F1J+^<K!o?Q!Z*!0er*Pfp>`q)*H
zTPnjtnud((T8P8Ru;Pu;6J9@2Pp4Lf2lN<F*R>Fb7vDW}s9kyB$9fv73{UJaadOu}
z9Dd@<4=x4oUwZi}$&||Q$fl8_x)$Q__C4FR+7n04TqW^WhKDr`8`-rGhj;XP?fYG4
z-GO>~TV=SZM^jzbLL9!_IPK-*sSm!br?*yyd-v#F+qDpf=eEt5chBhRo}+bLd+dRo
zzd!!m*As>}@E!7wxVUM>{kN#DYrj<`X-c|S-1iN<G%a0D0JB8823VGofh<eQP?lwh
z3}tyr4zfHgM_HaFa+DP*1;~oD0%b*(C{R|Wlprh9N|cpZqC{DhQh}^Wt58;Di3(+P
zN)56)twvd$C2Ev4DGkV)v<78OmS|AcrnDey(^{0ZS)xT*m(qc(OY2b9Wr+@DeM%3q
zKCMStpCx*f4JiZ2hO_}?LzWm&Hl~aq8`DOVjagzu*_1LZ7SEOj-jp^iCxBUES_5oO
znL#$E%_y6*#Eh~fWdYfewxDdu5(~=Kloe!a+KRF@OROl{QZ|rnX&cJ6EU}?%PuW4X
zr|l@)v&4?FBjo_uk#?Z$$Px$2&Xf~mXWEIfGfSK(yHYNYU1=A}t}Jn(>`u8scBkDa
zyR*cNvM1#M*^~C5?8y=j%HEV0WN+GwvNub-DEm@AkbP+%%Dyb|q0FY(#p1902A)l`
z%L!nXuxo$=sQ}1<bO7Z*mIP1^rh*^`(?OJjSrSA!lnQ|yN{3JmWl0F-a4HOPI2}eg
zoF!qDBdG|;k#q#*NR~uUj;5j@N7GT1qgfI~IhKln981Sgj%7&<<#;L%ay%VJIi4kP
zl(`fKGMDC1=CXuCIgv_$oJc27PGm^}Wj@7&%%^#j`7Gg4PNtF|C(}uklUb5PS-3&4
zir2peUa$+TKvRX7NR7Y0IQXH7<DMW5e3Pt6C0>=q_cWDD=9M=zaLM#bTj3hXN<%=L
zG+8B?Uk0X(m4PYaWMInuaxmqr985VU2UFo!fT>^=U@ABTm`cAAOeL!XQ^_g8RQXk4
zs#q16DozEa+OGyv&8opvb80X(ehru!Rs*Jn(}1b<Yr)jAS}?Vo7EGOA2d0kIfvMwk
zVCwyPF!ih+Og*Ou)8IFNX<!Xt8aM-(M!ykEBWncH$Qi*j`Aw~G@ncOPAkLJmlG$$t
z)6AN|G;?M!Eq)7_7S;l$g|mQZ^;^NTvQ{vyoE1!)-v*|QwSj5lY+%~`b};R%9ZWlC
z2h-tqfazczU^+Mlm`=YFOegCE)5$r(bopIix>y&OF3ts}+wTU`&AP#Kb8av_eh-))
z)&r)8^ML8~d%^UwUNF6!7fheu2d0npf$8IXV6uL;6|QP58v^3kWR(K`0GI(b0A_#-
zfEn}$!3?rNFoRqW%#c3>W{3@e8R9}<hW%kM!)zGLFc$_h;*WqCVIyEhxCoe0e-z9p
z8wE4UMZt{uV_?SE7??3G24>tJ2Q$vb!HjcpFgZU5CdYDMavTR{!k+*$!6v{=a0xJZ
zKMy9)@?i2D4`$Mz1T)Dd!Ax>VFohODE?%K3S1+ZY0n!OZkp^D#i+#ub3%x5>uc4Nq
zZgILRPPr?y!YlLb>eWlQP8chGulP5q>zZ(jl6de{zW2&(TU_11-_d->pI1$B+WkZ2
zo3Bi|tCqA#q$gq4C@mA|1>_8s-bbrfNok%+4<Z++^aOH=O3wq!ibRHeS((VNFUwFF
z_GNi0!@jIQW!RUMs0{n^B9UWXUM6zv%QIAteR-bBu`e%BIrilxD#yN}NEFytl!*fS
ziVRg?Uy-K@>?;aXfqg}ZDzL9C5+(MPWunBsGDDTvSLUe_`^o}UVqaOJO6;qOM1_4-
znW(U@%1{;dRe7qyzN$b~*jJUP3j69JQDa|SCTi@fGgOUzb)KrRuP#tE_SGe-#=fRV
zG}zaai3a<c4Ao#?lcyT&YYJ3@eNBmKu&*r=E%vo#qQ$;8L$%n~=BXC@+5**LUt6MD
z?CXj|hkad{=&-NLP#yMld8)&{u0VCz*OjOa`}!i$V_#n;dhF{nRF8dqp6ao$FHk-9
z^(CsuzM)7A*f*4k0sDpwHDKS6rv~gB3e<poLx~!&Z!8ic_Kjs?#J({@jo3HlsS*3e
z0ySdaSfWPkn~KDA5^iRg%EWX5IYUkN(dv6VQ=Xa*A{VIX1agU*&I6l^#EgA&nV7L}
z&QLS<&3S6ZzPUin*f*D`8T*zZv0&d)CKl{lGSq^7OP*S=Zz)g<_AMo9!M?Rftk}1f
zi52_S47Fn4nx|ImTMN{ReQSwYv2QC98}@BwV#B^ILv7f%<*5z(wgR<b-&UeF?AwdP
zj(vNX*s*WVP&@YRd1}YLy+G~Qx0k3L`;H=UVBb+D4(vNJ)Pa3Ro;tAaC{PFX9VP0(
zzOzW2*mstR6Z_5#bz<L{r%voU3)G2yXNfwo?<x`(_FZM-!oDj*UD$W!sSEqA0(D{E
zRiZBJyNkq)eRr9-vG2}MH}>6m>c+miK;77Pm#7>2o+9yJ-%}<Y?0YiQgMCk)da&;)
zP!IM!CF;Sxw@AF$_m+tl``!%oV&9vmUhI1d)Qf#@iF&c`D-s{}eP!aqzAr<4*!ShB
z5Bt6X^<m#vqCV`iMZ%thdu(i(uosXsl)aCNZxSoF=<<|3h+LrT3FH!G&jSaFB!K-u
znFO#O$j|`x19=+2exN`D*bkIw0Q<or31UB3CPC~6Gc<_(V4eoCA1u%y_JbuF#D1tq
zLf8+LNeKI)3=LsFl&2x=hYB=={ZNU9upcgxF!sY`62^WwL&Mk)=V=)G;Q|d~KU|_=
z>_>_ug8fLDM6e&p&<OS;c^bifq(CFskCbQx`_UqaVn13YQS3)EG>ZLbo<^}BEzl_T
zqa_-}eym7h*pHP-4EwPRjbT5Qr!nlu3N(iOSc%54A1{(P_Tyy|$9_CR<JgbqX&n3U
z0*zxoUZQdAb49{opDPm%`&@={*yr+;!#-D_9QL^q<*=V9k_7e>Ws<;tB103{PvmI=
z`-uWgU_Vi!3GDMl!egH=6CV3~hVt0w^OVOvU!Xkp`4Z)^pDdCj_LF6j#C|eElh{w@
zX%hR%0!?B+S)$2m|K{aqCBI&n|3$Z&fBS&6hEFx8!s3Lu=HI5e;+*>Gx~`9oVWyiS
z(k(EvmF0*GS)L<uWJQiBkd-;2L{{a93R#^aYGh50XpprzqD9u_hz?nwBYI>*ju?=Q
zIby7`*RU$pBwP7#<xavN;~=y6k?~YtpE2J(GO>GQ_3YZ&y~IrvZ98l@wwJ$cOyyH!
zM(V0g=VgoxNY2QC6pR8$$tZzTj0#B2sDU($21v_jfpm-xNYChj42%KD$QXf4jLB7f
zlWu0rKo-UVWM!;CHpT{IXY4=@#sTDHoIozd1>|PjKpw^e<Yl}-KE?;cGOVk5=?yRe
zpdb?j3Nay|FcSufFcF|A69tMfF`zgT2jUnGD8VFvc!mc`GRZ$F`ps#lq~Xr(-72@J
zMh-QJr&9wzeB^MG_>4pY-#oI}BtB3fK9*6r1$E__jOr;T3=^8gLr~m%73s>88C3~S
z<>lwCL&hImIjC~@Ha0dkiJ$6h9{OJ6v19WuRaOtT48KMEUx@#H<y61&G)Fbb<_Nn5
z-UJ15B!C>uksxv?M?%Qq90?;wawLKr&5<Z_EJtF<@f?XGb2-8xCvqf#%;yM?oXnAA
zm7o0~vncTO`Ufj}Wb{p=t<{rvqJ5%A+^h|JtE<&79&FDo+}?dk$7Wk)gY~<*Uz2$I
zPuy^o4R+=Emg)vGbrRDSm=&8li5c0_Ni4|LPGUv2brKu0y_49H9i7C1?Cc~?WLGC~
zA-g+?8`;xIJjmWo;zjm#5?__Y4+SfS+Lh;Ysz0W1jWAR^9(PW?@l0&kkv)~I+ELRn
zQe54@&#9RMbbHP1f6~sA59~VG^3Z*ity<T!uD|$f&qp==w%+sZlI4}v0weSnOP`$n
z>g}z^Uv*VhH(Q!n@nHST&eyJdsBu|k^#Iubm3Z!)Qg*-4v)6#XR#wmMKYN%M`M$M&
zkk>MMQ)Tr)*+7+eqgZ^}==!>@r;MsUp;VA|R1XlDAOn&Mav+7E08$D{AeEp3QVVJz
zji3S23R)nYpaaqidLV;f05S?jAd_I~s2<2>!3<;(EI?Ml3S<*(Kz6|n<PaP{PQeM}
z5?nxT!42dQJV0K-3*-}gK&-%ah$mbF9}ogSK_Lhf5<);>Aq*4|B0y0g3KSD!Kye`s
z#0eZwLP!Ad0uPiFl7G_Nws!8yaXW7=KkwEnFJe~~|MwLaUq~d;>dP*@3%l|ENcMZF
z{lq7&J(AIQj7vxoRV`V4$#*2%XI0QW!`rQ<dyk$>ZGFR4jlHgR_O_Zw_pWoA`V8pX
zZ{Rh?@e?Lax^ePN^8PJ1H0!S)_W!MI8a!m^@DU?NT{mXzIH^LZ(&!A=Y6r>Gmi|q4
NR~xnYZgvpQ{{l`JB)<Rv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.793B97DB82C3A02D.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp.793B97DB82C3A02D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..53885d0cde8de127e651aab5d203a3c9940bb908
GIT binary patch
literal 11488
zcmZvid3aN0qQ>)mloHw^n=MoeRcJxUz5_)e`)=5Cvge!>OJX7HQizHM!CVpMLdh-4
zRTS{LA&jFiqK=B7BV2;Y%%~$6Oa;U(Fkvo=I?sgq9P|Eh`#gM4-}Ak{U(%E1EFUxN
z7Rx7Mfxy^qU%6o3GPX(}5DYE<Te@sXywp!1I9MwXtla*_+y$QAi9>Qvyy|)W*}$1U
zo_3_xd^Ebgc6P03;Z)y?-?hvg)z;B=&#RI9dS9O|+<th)mATIiJF%hT+_sO^#3tx`
zG;yp-{$A-1;?wuu@KTj@{Kv~yt-kkCmqT{v&Xc2iduvzi|0>iY*|=%dhLc-H_wLU8
z_?@x+RQL|rpZ=9uzxT?lgFh_M?L7bRkDm$~yTivvEW7mM&+LjTXV*QRzv;Z>?Z38v
zBs|x0;K7zxJ09qIW!&!5HtRNa)%mUCPc`iRXXmr6A$!NJamO5SefJj!hy6qF%R8D^
z{FBeESY3PfdjqE^_gu;xsEyOdcMYA;|9E!hPqxV&q14{ppFJ4dCLZ?LnSCoaJ-fXB
z!;YiR2*O`qf8yOk!)~2N&)gP&{iUbb(%|OX?^!E+VeZ$9cXU6nZO2N{OVy#<XYIE+
z$h|x13H_hD=N#0GZNJdcL5>)f9B-<9=6vAEJ@SFmCkH;c*nY>tp|R-?ng%rokEECX
zX*j!Vr2EOM>K~@RERD(5dv8+BBC5ZATNCJ-|Lym!{>Fj*mol4fD>Yvbzp!nV>z0i(
zH?}PG)gSwiPUybUv(mQpFO+++{$x%4>LqoGPxguC9n+0$e7c1`(9s!qSH0%wh9hgW
zuWugs`>Cr3-#Y7HY~qF8ft{_bL%IzkUmU*nXyb-i)9yd3pa1)(cFq|4&sUc-_f5R*
z{^-pc6UWVO?)}IA&7C3ogK|YI`p)*uhnjtD+d7_IJ8$dLAL-Bf#<l6s4xQiGai?j^
zCFx_bg;l=p8`_>`7q$f6o_si-u_t=}W0LXcu{TBWiTi&y=asG6s!qT6*j*P_3h$V=
zach@)Mq}O6_ZGitP5j;dZt1Im?!&jHI)}8bdE^P{xh+>G9qACyc}?}9W4ne)9P{4t
z^#|{4zEr<r`~xSSoBl;-&l>lRx=E|_TbN0KN!>MPbGEagApM{5RUL=h+m|UvUbkpB
z_e+jD{V!Yev8Mee4xTQI8o&75o}K%<k2lSnmx}gG+q|_tvwP|=_34L-E0#Yre9V{r
z#r7*r2kRewH1*Df#DTMyUViGrr4P5<Ph`i|Mh^__=pt_p-Z;Sc((H9#KKJF6ZKw8a
z-TuX*A9Sv%pT3o9KGzl9GH}BGYPpw3N!Fcs=%<b^=e6CU+Wy=($@|=!KYRYYtBe0z
zyl7SPl8)!ap6`b@-F5Z!#MB{g@8YTN^e&FgpYr|pkG%K(Yn@|P`mb9p{d~`b!E^3?
zb~E?>yK9?%c(pb4;G6o5doJ~^ySi;pVE@3(EluCvy*wv->!YLU&pw^-^ArEtm0hgA
zJ9%(M`1`#-KXm@{`yPL2N9>m?)m>Wk-rZCC9hWiDQHNG-I@{9QH1*Kj)s4|*Yactl
z?&{Ku-T#%}D7|Q%{_drvYo-kk-u?cYU)^|HPO!9V0k?$+1T|N8e4ch{X4TTUJ^|NQ
z-8i&wD1STj{fytNIG7pG(OygEl&42jkErPz_-UnkTI!dN7j@OrV0pT=x^-gTz)w&8
z>C$)4oa@?KOTFdk;nl+j^bP#9yT!M5uHv4tg9J0m(>FHU*wQ!f)9pdoAIG&FI8#e!
zm8To38*2Ine!AoSI`^Gx4!>PX{pIPl>b9wU13#TzbN8UXJU`*VS{f)%&#0bZ=o|Ry
zeaGi7ZZIdF93+@tp1!%^=9a#JpMGQT{;!XG>-5#q+2!f_>iU|#fuBCPW4C+v_}Djf
z0;W9O(Z8duZ{VlTAK5x|Vaa}Skib)(9^Ei{Oy9sy|1xCmOQ(~E&(+eI<>?{SLk9K@
z{Ip=nygQf98drI3x@!7flSdBywBv;@+HR`jL`l&qe%F^%O0=Fp%n(sGWN}gqWpPT3
zWpRdxu`EeSpe#vAuq??C36`ZvDU_utDVC)fBE_;SDTA^sCBw2TLu6Q%C*@F<r{q|c
zXNVljilhR{ij)G&iVRU;S(#KqS(#E|S(zb9EUS_#D63K`EUPj^g=KY84P|vojb(L)
zsIja`YM`u1X|Sxx5Dk{KNiCGMDJ_<@8KT9qE~$gEE~Ue=E<<!!)+hC=_|H)tr%&nE
z6Nnk2?}lth8lY@Q8L(`~5CfKtNh6evDI=DR8Dhk;DQSYTDP_X4DML(HHYd$cHmA&3
zHfM+#%a)`C%9fM`%a#nWVA-0qLfM+KV%eG@RxI0+HYnRtHZ0pR#D-;i(hg;N%8q4w
zhS;&}NIIbGNI9_V$PfpXok=H@ohc`lof+c9vMcF=vMc4nvMWPeSav7fP<E%>SaxTK
z8_V9LcNPDQTgQ1*-t`1xhIqRnGf4)@Op3uWlOYV2eMujbeJLN7eHr4zvOnpEvOnd=
zvOh!oSPmouP!6O5SPo=J0L#H-5X!++5X->~31T^v3_&@R3Sl{vAt5Y>lVK=_Q(-KJ
zGbD`VNHPNDNGgKmNQOkP98E@{98E>B9L<m@mf0i=Wj4iPnavOu%UqI!GMD18%w-6N
z<ybNX<yb0)<yeNqupFNpH~)5Y<>J=(Oh~ii5k3w1Y444jrcBsL>bQDIy@Eepk8HeR
zedFlmq2<j@rY0wUhB2=H{_du`M$Im7t{G5M&mU>|vm5&BSxw~t8*h!@%0Ci+WG$Dj
zo#0eTozl)IU*gYnz9BVzr#O|Cs6a%$a5xr~i0CrrG!?C-{DYK_<~b@lfjLh_XD}D3
z=ptlsfr!yBE)g;M#c3)=zc@$5=ojay82#cR6{BBLAQJRTN<@NwNt#N~FUe5}`XzZP
zLBFI(CFqwHh!p+O5|N@`nx<0pOLJ6;ercXc(Jw7hDf(pvB16BdL}cifrKt@4vK*D6
zUzVpb^vjA=hJJa0$k8t^5jpzhX(~s*JV)i|m*=S*{qiD}qhC=V3iK;VM1g)qnkvw*
z$WaCQ6?v*azoJML=vNkq68*{&QKDa&rb_fHb5x0bWu7Y0uPjm}`c(y@Lcgj+ROnZw
zsS5q7995xTm8UB7tBO>GeszJU(XTEMHTu<Qsz$#$N7d+8=cyX~>LOL6UsE6&^lM5)
zgMLk#YS6FAQ4RVvd8$FbrbspD*A|Et{n`@IqF<Y)TJ&pkREvIXo@&vrEmAG|bp@hB
zzpg}d=+~vG4*j|u)uCUPr#kfOid2VweSzqE;Ve;KBKpgi(^S8f!dYdG>Q7+KQ~ept
zMXJ9D*-#(`^czaVfPO=o8qjaZQ3Lu7d1^qvp-2trHx`Hy{l*e8qTiUNM)VtV)QEm#
zo*L0_EK(!-O$A~?zo|q_=r^UQ3H_!VHKE^>rzZ59iqwRDbAg!AZ!Qrt`ps!-M!z{n
z&FDAhsTuv|A~mDmQXm%eTS~-&eoLBK&~M363;Hd2YC*rHNG<5M7Kj!7))KLz-<qaY
z^jmY(ihgUJTG4MUQY-px1!6<Ltwe0-x235K{k9ynq2HFLHuT$y)P{b0f!NV+FA+QX
z?P+R9zdc9o=(p#o9sTwqwWHrrAP)39O2mPFN18g&@5oUH`W<=dK)<6%9q4x!h!g$J
z5^<v6nWj$kJ9E^DerKLK(eEr$C;D9l;zGZxL|o{1rKt=3t{iot-<78>^t+1Gg?@K|
zxY6$}5jXnXY3fG5J4fB<cju`a{q7=lqu*N~-d?zP@s^19GUhb(uBC9XlcU}fnDf+o
z26K^mFG6Mtgh4-3A`JSOG-c4w<S2uFCQljkGeydv-&Y_$^!rN0hkjq0`q1yoQ6Kt!
zdFn&IuSk99_ZNsC{r(d1qu-yVe)Riu)Q^6Dp8C=6FH%4H0|gR5f1pGH=ntf60R4d+
z4WK`ervdZ_iZp=!V1WeDA1sj|`h#g2M1L?xgXj<DX%PLvA`PNHR3IVrhe{-b{!p5R
z&>zat5c)%T8bW`lNJHok7f2ZW;Svd>Kb)pv^oMgajQ((*hS47`(lGiX1rkAjq(ma<
zkECe?{gE7vpg)qQ5%fojG=lzUfke?CEs-etqiGsNe>6v<=#S=U6#da6jiR3|5ElJx
ziLmHr)09O&o1-lH**s;@&lV|*ey%_`^m8S`p`S}r4*gt?a_HysltVvPq#XKV1rkGl
ztVCkykELl0{jnU4p+A<VG4#iZG*<EJ*MBSc@$$02_p9RYa<+<FICf!>zqqa97L8r>
z+m$zeO?_QW-`#q+(#;alF1WH4XNefIBugZirCB1yEXxuZW_gy#F)Ok}fmxX)O3bP(
zQDIhRi5jyeOEj3ZS)#?P%Mx8ByLFqB4U!GNE8j9`oY-jO*GXpIF>L%-|2d_9dGq{%
z^9S?0C_L|Px>J9S%^6?5fES7!mCJdtPz*^Tlt7XSrI2Jo86>$-4oM+YKvD{okW@ky
zB(+ctNh8!i(h9YZbV3~@y-@F{oTM9s21rJs5t2!0f@Bt&Az6eLNLHa0l1*rXWEa{Y
zIfM>KPN5T$OXz~+7P=vMg<ePH(8~xJNIsztl3(bD6c7d=1%*LKAz=toSQv&B5k??I
zg;7YXkcGqvIY=>K?6(yD^1NM8xBRt!<y%x^nhgBMsg7$N(`?{Zj_SBEW5yWxMJj$J
ztb7aV+A>(>BNuOukLBNj{Hs?kU0V#Rl<=d>@2qm^+LBnMgrD%IM!h}c)Tw1Zm*3;f
zrsf&^b9(-pbeHdo?HkKaymeZux$+)wv$nbTN3Y{%IA(bGhk0S;YyGD$c*;^X4Zm)9
z1Amu{|GUaxA=g&kD$l~3CEi_da>Zl`gV~oQKFt0s@na5TNdR*&OM;j~SrWn=&XO?Z
zNR~t}N3$e~navUwGnXYC=2(`*Dw)42S>D;#RuC(HX1p;zihtP-&A$CYr1|6b%6ql5
zs&fp#xsF><wE$AGD*0Rbpz8X4Yu>tVeR;3Oh4FeG`RM%TukSnkile-FjA;zR?__?u
zd}s37LM8t!{%Va4+^}iLx^lZg(xBkqVrS+3FAo?z;^Fe<`StUgd6<6>lwYxH%bC2S
zyk^ux^t<44-_S!0n2kNeh}qObOqk6*#EjX}LoArBJ;aLH)<bNV?LEYf+0jEBn4LYu
ziP_adT$tTG#9hf>&#b+{ZHhlx{_96bMkx6E!~BBnj`BXdwp?5J^F<!f%*qEr#E1fr
zSW!EqglIt}i9KTA#Ecj?F)Id6oDc&i@koG^FcRP-tOPhoLIRxBBLz;%NP&~GQsATs
zDR4563^*Ah15U=ufRiO;z{x#w;N*-PI5{f^PM(khr|>9%Q!onP6s!U`MM43b(xU`U
z$tZzSvP$5T2_<kUj|w;yqXJIFs(@1^RKTe{YT(q28aOqp22Pz&1E=w5fYUG<;54iT
zI88zWoYtcSPRnS4)3RFNv<WS6I*$%G9iszI$LfI7C3L{)J^GoIuXH`54?tq|?T`}s
z1(js*7=SY{2H*^=0XRd#0G!cd1kT79fitp3;EV|)a3+rlI1^(6&cvF4GbK#GnLTFU
z%#0a0GiwIUoG=4t@mPSfFc#n}tOYnr!UCMtV+GF2Sb?*$R^Y4&D{wZC4LBQP1J1_U
zfU_lRz}Y=^;OvYYI6G?x&YrLX=kPdyb1)9z9IOL4N5TP|)8hor$vA;?vQFTf2`6wa
zj|(^#;{wjbx`1;fT)??KZs6RE8#p)X2F{&u1LyU4XYwZ}<#lSt8-T=m+aV>q3o426
zFu*Yk102ILz%dC1IG@J{oR9GV=VN`q`4T?h{2o7We#Q@+pY;RhPxyfgcmlu$m;i7A
zHUL~85dbde2?7^ng1`mYAaKD%5V(*h1YC#-0T*IJz=aYa;KH6TaA774T$l|57fytM
zi+CcyMVJV15jFx`BoP5F>WKmuWum}E*(h+)L=-sI!ve=LEO0E#0>>s;;5ZKl9LI3L
zaV!TMm*9Yld1AoDm>6&|HU?ZQ5d$tRic9%(wene^Jg$PIiBIHH`Tv%6ve%ZsmoAAd
zoptS(@U`LBp9ueppdYs}cHn~<AGGvT$3l^Clof=2n-fGUzxJtqjj_ts*n;-X`Q_-m
z)7`JCfAs+2z}mV&Lk3@GA8Iy^8(L#G3>z{0`jL$jC$&wv?e?kDr1hf7W3{(5|9@K>
oZfqKL)6HXATF2iyK`fIiR2toA^$ilGslLJLXs4Dt^fu!9H&kzwF#rGn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.A7E63495E5270E3E.idx b/.cache/clangd/index/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp.A7E63495E5270E3E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..43cbd00430e1ef8d37e1679bd467d846aae33aac
GIT binary patch
literal 11446
zcmZvid2|!?qQ~?5QA%jBQc9_S2vS?jzAI(Rz8m&T_DmvM5U>uErNW6K-T(&K1YW3$
z3JNZWa1aoL=K^jhUL$(1AeSrQ1=Ne+3lJCd-k0{AA>Th9=X9D+exH13CX>u>W-`%f
zv+a@!1m>yEMRTUL#VZ5?L2c=0UfbN6KX(%dj#LW-i=H{c%o*2sZ|}4EljC-*tb2Y#
z`dAO)i$iKw)Qo-eIcD|fr}yRd9oT=b;?L`Yu2e7miz~cqwCMEp&lXMJ(*BRT1U(-;
zp`LTF;<6=q=;M2yTRv)O`V-H??F$$DJFg!7-0IbfS3Tvs*7mozmCp{k<JxT>*++jk
ztM2a$*R10fa_#^AY2k<^i{H9$276`452r5u-ey^)J#uKvOM4YF)||ZY#(^JB$+)w1
zVNdgu-*x2o8#XPLAM{C>)4D?&2JSi#ZEv1EA^6ve15P%LSk`eu@>0_z+xM>A9lzh0
z^?+{SnK=vlUB0z$Q7(A$sRtVp^w}>Q8<$>~_tlv7p9dQLDQWw5tz_?tW6i_GXD3aM
zyti$=TKuT}+3YEQrtYQt1c#P?D48{2|F746xZ2Ya+c$Adw<gzvkM()|;{kJjGo9+Q
zvd<sE?(3EvTje{XYuwXz_o5+tEH4~9zx3J?$(!{}#5JRO;Fc5TQp!s2keR=JC$8Oj
z_1P6$w<W5+-tpe@t@HaYZaGX2D(fGba$^7Pjmtm0Ja+3w<K1s<d-zi3k!^c-=f-lc
zR5-h@J0pA5zq;S}7Y^?@X5K#Dz2YDfmQ1|#iRf7D=>B7yTJKi7oD~&{Z}zX6DcGmn
z)aUfiE_wd)rO{KvKde6Y#kSD8rly`7hTgWT?(yUOpO|Da=I>1G*}82)<Mo5{CqFWL
z%ww@TS3hu~>7CX>F*-rASG!<N?8LJxv%1|QH(%`k{^8^5w(;%%{qETD?R3Dmn+Kkn
zdic`|uMH&WiIYCkex@rl?KC~JDf8kzyE2-sqyFuEq4p~i^EGq2_scUQvNzs4dFSwb
z4;}t&+sfnPmDd;h6oHd92S466OkKO<=U0py1iIusZ*TePukMFGoVah~Q#+3ZmL1%G
zzjJHN=+6%NCob$W{Pn8(vCDgYQ*nIlQw?h8#=RRGUzvP)gZJzQO%E+Omau*H*psU(
zkCN9XzxvOaJHKu1IeYz|#lV5k;X3K(hv&Z{+q3!T<o<()H?5T&f8v8L2Y;*$eR^fV
zzBP5t+)EYPM~*$TZiMc);ESLBWIJ|p{f48T?ETBqyFYz=!rQ7DR`O0wO%L-uzEu9z
zzC$;kDKw3pIpU9z`=6UL{{3ItM(_UQaqae5EvHvaKiY8YXmw%hpp8%K|C70Ep>N#v
z^Ua&Hm&RQ9oJo&rcO+gud8Nm>x;?kAxzY2EH3Qw}xrMEx#*Uf#%-S9K-B->ZeKpzo
z)0VL|-9XL$i#`5pWz(OPE!RI>Gw{b@2Od1#+`K*Urv9rDdg7LxTr%smpC9?%{3K^x
zc%|+2QI~i9vEt&{cebo}e#)OWYT8x9TVMK5_m1cGygy>a_2y4cnS{UBgnm-J`~9N}
zp1$RJ!?_mKJGVb`a?QH?zGzsww)xGU*FPe7YrIJ^uYJz_ZA2ib`u+K{%iOw2)zs@3
z#2YFbYP%MtPlm34{cg+DM4P^O;ugX9((vGZgYW8E@WX@XjVtF)h&NZ$P-(cavazLW
z!4D5PcXjRH^X-q_A{bX1zP;b=L%SCI@UZo&Lo0jCm|9IImxk*r>#MpJ{IG1*sKpz)
zy|t{Dz*icc)_vNjt_44=zwy66y?mkljcUr2hHEQpdvq=MVdK!yd$VSX>Z@s>G~80z
zV(41%!&AeGx0**D_^z5xEDiUq>|52f;D@&efBLSuIugA_FrhTuzhD2mx)%KKkz4ow
z_`xN2pqfr94cArHRdp@+;ry;YgHQJRbwf1`mWIbwjx%>H_~E(JrqA;<3k2QfE*{x(
z^n!j^uec~7YUNjEuedlN29hKsK+=R1NS2TR$rExQMM43jOeld=2^Ek!p$5_<G(g&f
z7D$)S0qGO^R(?JAiW?FJAY;M^WJ;KT%n37)C1C-wCagfVgbm1^umd>~4j^a33FJz+
zfZPc;kT>CN<zJy*aVEh4`4T=Lf5HzGNCbd_i6BrY5dsP)!a$Kk1Spz_0<j4eD3*u;
z#S`(H#J^43tGeEvPriTS`JLxR4k>LooM=gT!x1G#D+yqlh&BO>IWfrMq!?v!nut-B
za1xLuNeRl5G?Abz<)k1>lTwtWX(B~g#>qgIC1ohf(nN-`oRfnrPs&l2r->Y81*ZU6
zkyN0pND~FhN=^x~GO0vanI=k<Rh$ZBRZ@krDos=<t2s5u>ZBTFb(*MA)^Hk-HAxN1
znl#a%tmU*IYm-`(wP~V7S;y%>)+Kc)>(WGrvYyi~DZiKXN&QL!m?rv7zy{6$vLR_e
z*^nj%l#QGbWMk5ZvN27JD4RGF$fl$TWmB4%P&RXBkj+Um%H}jNqio?UAX}0alr3pu
zLD|Y#LAEBXC|lFSin5Kffox0KP`0Is4P`rL2icyqqijzTJIW5u0kR|MK-rNd4wRjo
z6J%%7iLx_IoG80E7s#%p3uRZDxKMU;ZjjwcH_Glbaii?zyi3X-uim70B>_wm?<Qb|
zV?bt-49ZNJFev*tAIQF>4`pAP_)zw9evti1Kg#|z@uM8z0w4#H0h9x25<oe~1wjrb
zgD3~nB#3f|3xOO;hENWrNeJaI7X~?;45J)QlQ7B=E&_5S89_OcCJ~gQTomMJGKz9E
zO`<5X91Aj=WKm|*ghe^V#XycFV<^YcB!+UFi-R0b#!-%^NgU-l@;NOx=Ya7C<qrb2
z%3E(z?;CFnTr*0%jr5Aw$?8=6+>D>P)>L&(S<AajZ-#ic<2INf46-jne8~O`@goN^
zB!C>ukRWm>Lqf>m3<)DgG9-c=&5$TEn;|T6EJI?*@eGNV`L_PQ*pD2)R$F@Z11tmF
z{KS=?1kkFk8F~5H&uyH`@N1QypO;oq=j6OB={tyiJAU4G5CgKYgBX!b9mIre?jUAl
zO9!zaTRVsq+15d9$o39mM|N}&2ePw+IFVf)#D(ncAnr2r#~Gz{(Rq+jews7oGu8aM
z;D3{<uA>ZI;*T=^Rl2|P7=ypRH1o<6(FK@XiHbz@3v!BzmQ((9DEV0``Vu)uMdy(V
zRCE<soF`)Ji;F~zeQ}D4u`kY2G4{ndD#pIJK*iXX<cS3Pk|L2{Uy`B{>`Ss#f_+Jj
zO0X{}Pzm;>c_PKWv`D1bm!_x``_e3xVqco0QtV3$REm9Bp2)B-D-s#@WhpAdzAQ^+
z*q7y~4EwSIm0@3=Cvxn|i$sond5X%hFV9jr_T@P$$G*Hk<=9u`i30nIB2i#pk)jIh
zE3#C9eMOEcu&*dk1@@JBqQt(kNR-%Drl=D8$}Cl4Uzwvy>?;dYiG5X`sIadp5*7AU
zDXPN0Doa(^SLLV*`>Fy}VPBmmYV50vM2&rQimI`%&Qdk@)j6uhzPdow*w^HV2K$;K
z(O_Sbq8jXLvQ&e8O^#}?uPIOs_O*GU#lE&kwAk0Cs22O$EY)IPo1<FnYYSA1eO;dD
zu&*l;9rkr8s>8l6OLf@S<){w(x&qZ<U!N!X3$PE+7m5BC<P_B}r*MprrTQ<Cb5wsG
zxj^+-fem?Lz`mhK4A?iMr~&(iEHz-?kfR3d8w%8bePf;&v2QFABle9cYQ(-VOO4n!
z=BN?-#sW2B-;^gN?3;?jgnd(rny_!mQWN$~IcmbbsX$HGH|L2N`{p7sW8a*jX6&1@
z)Qo*|j+(J=E>JV}EqP+WzNJVk*tevp1^bpPwP4?pqZaI23e<vqYo1uKZ!Hol_N^&u
z#lAI5t=PBbs1^Ix0<~h_mM1ps+ls`7eOrp!uy4y!8}@BEYQw&*KyBE!=ZPKr_9C%k
z-=3m&?Ax={j(vNM+OcmhP&@V=dE&snqevXscciET`;IJiVBe9W4(vM$)Pa3xo;b1Z
zED|U7ohj<XzB5am*mvfr6Z_5rbz<L@Cob%}io}I|SBko@@5)ja_FXyZ!oI6OUD$W$
zi5vUwB5`BiouY2+yR+1deRqz!vF|QWH}<`G;=KTe0p23<{(_vM-sKbyU9;5tC322>
z&m$M8_bM=xCk*zPB4M!4q$q=ZCQBLYGdap&pD9oV`@TH!Vc%CIKJ5Ea)Q5dvmin;o
z%TXWpeFf^nzCTa=*!LHSAN&3k^<&?krGD)DbJUN0e}Ve3AIOsc_5(!{z<wY_1K1B_
zX#o3y91UPUP@n<q2lFI|{a}#<u^&v)Aohb<8pM7uM}yc87HAOrp*#s;KU5?k?1xe`
zg#A#KhOi&X(Gd1S1scMBI8VaZ4;M)o`{5J~V?UgwVeE%<G>rXlfrhak$&(27BSjLy
zek4UB*pFms1pAR3jbJ}gpb_jx^CXJ>Xpuy*A5GCH_M=%E#eOtLqu7rYXcYTwp0L<w
zi-g5Ko1!fC*(_zT&*mtLeYQYZ?8ovXhW%KP#IPSr(HQn)SsKHBEJtJ5j}>SP`|&)9
zV?SObaqP!aG>-jvmd3Fk&(S#c;{_To_xX8Q>6o|k+E@9owXLx&!XN7zmtPY6cCGDn
zw~9EP7O04$R1sg$xFE(q#B=ZZ)_$^8YfGma8iqHR`Rfvs$MzZa{kl=zOLtH2IsI1N
zj!ygAy=Ok2;_bY`B^0^J2g72a7)T<N07->XAem4GBp1qo6hZ}%Qm6z{2~|L9p&Cdd
z)BtIPS|FWJ2c#G3UHqynUEvZMfQ&*TkV$9)G7HT>7NG^mDzpOGgf<|%&<^AfI)I!)
zCy-0%0&)x8KwhEORo-ziLI%hu^a1&WexQIb02CAkfkMI%P*@lSiU=b>QDGE_6|z7v
zVGJlPjNhc_|6Ftmdd+{kTj@O3&;cg?JXWvxprM0I{E4bw@gYNpnD{e${59OtHr{yw
zx4exf<cXW-b*6cYN0#o{e*dD<Cg>4ODsO@!Mid5OMbpY8_K3k0Gh#5stQbs*M*^mV
zk$@>-C16TDQZS{A6ig{A1ykmcfhl8TV9Hn-m~xLCOgSS5Q_jl4RCp9%Di{Tr3RVH8
z(xU`Z$tb~8vPv*j9u=4>Mg^vdRe`DYsKHb-YB1HT8cdBx1Ez-2fT>|MU}`;DFtv;p
zOf9PgQ|Hlvsbh3t>R26^dXIim`R%S}^kE=YKdnp#j{!^rV*t~@8o)Gqj9?lWBbY|k
z2&Tzn0@K8pz%;QYFwGt_m}bTdrkOQ^Y4KRVv@jMhEvyAhtH%nam9c_pWvyV^JT@?G
zj15d1YXj5nv4d%6>|okiJD3iS155|w0Mo%bz;t??U^*Emm`>ISrpx03)5W;Jbg?cl
z-5xiXZpIC!n{|Wf^>`<hKU%$vHw?sjr<IBEFkmta117^VVEQ~hFnx><Odsn5)9>+v
z>1X_4`dL4i0Z#zT022T+zy`n!dV*jEnIM=!HV9_O69O~Dguo23Auz+9FqmN`3}%=O
zgBkHez>F{vFe7XP%%~>{W|WD78D*njvK|&pmSMqUSr*KgCkAGWiGdkoV_?QTaWLad
z9LzWy2lD~>16qFm(<?qpJ_~5JeD+Ohb?^7>z7bwpnle@OsH*3`9K`SVO4npMzjjpK
zsfjW~v>mQ5h%-ctEXfcFvNS`a$g&KPA<HvFj;zQK1+p?jl*p<KQ6Z}{M2)P;5Dl_6
zL$t`c4AGT&>!9~~pE=X^N9oxQvJ9HQ@8Dni*RuA&xo1|FW&w@HMq7E0+~R2Qls|pP
zxyOy?-)TD+z1@BMoUOJ}skWxJAAiAyzYJKK9CUs^tGpg($Y<P~9K1hy%n$6~zkgKv
z@3;zXX!&KES1}K0e#QKo^hHJA-H*Qe$nwsa?vUv+{vlspJbQTeKlZswcMq`)@$*xZ
zukBNMu3FPORl2*ctgninvE-HAul2aKA>A3NpFWt6@K34qH+Oz%to%3Q4E)E5)dM`1
znJ<(cTQ93u@fWuQbEnOYcWy-U=1!S6sr0*B{FBEY=%_l^)KYnWI2vJNf(XBj2?Wvd
z&2af9TfP}9-@LzkbIP2lvlAtBPxo}I=w8`F*t5FVExm88an@SR&9zldQ=f*qzPI%s
z-ZFC3=rMYOyk0V*@t(T}|9|)P8+iMmA$JbFt7(`>ELA8~8r{8jXa~sDmim5n*A!|S
JYjlwN{tqjt|BnCw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp.486913CE7740FA6B.idx b/.cache/clangd/index/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp.486913CE7740FA6B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e9ac8a547ff8d3d54d81782083e32ddf11fa9daa
GIT binary patch
literal 3328
zcmZve2~<;89*18_P!b=6CG0z4k+2F0JLDoElIF6+7a)L3SV9q$APJSAsDOgFj*P7x
z_XQoXv?3IosXErGRB=>rff1?7C>0#sij)ec*x)&N?|RPUoV=Vb_uc>f?(e?)-diXZ
ziGIg1EG|!6zCx+Vrehdpru(WjrOOHpFsz+{VdeYo$ud+A!c2SLdv|W#yF7&dxbAZM
zna+^<9s@0VZfTy`@q{nxc1%y2l-;e*^6gmW+;3eUeCgdUd$-tvS&Himv;3<frHOq*
z)Awv|Jsux@v}Wz9%CXN6Jo4?_?X~%9uU~V@O0!Q#tkWL($GI?R)JFNk%7o2l`^NU}
zc;7wvWJ>$5id7e`o=U$y@)cV<o&V3Sbk!+=@5@I&|I&A+@!f&J8_Robzie*hw|1$o
zo$qh$eAzuTP*VB2D>(aqofS(8{vA4u7p4i1c@U!Exo>QzZ%y1iSu)_b-Du*OZB)rO
zBa$;kysD&vMsDobC+*2c)V;^vTm9@R>uUaM3v2tHiImLrv8Rn&nUAHrH6d2ahFuIW
z(Os@g{JT}al5^iz3!BR#dO%IX-a|rFYwi4)(hm0{^b16+%;oe`=|rHpeO>kJAPIX<
zq|?<~6C(R2qwcV*2d^znCN9rAVR!A)Z0q=zn=H?+?X6Asn(&yrl~;M!2kje*EUN5_
zOxm|OMGPE`_o+!yFZd&jD>&NLGL*EkiJoCtKk!X%;ezB7X%C6P!?F_|jhQ~*T81C^
zkM|YDrNbAOsOAQ9#cuQ+cRR#NtUcsQX8(wt|InM!=v|9OVh6jAJ}_@+5PoWA9BMOP
zQ~T?#6k_$J%;&@90~wV{$8l%>02g_}*ZqawFGI_wmJmnp+nfz<d2+LHG@`%e#+dox
zP0D(cBhQK)@E(gE-pzrl-~Km_e{jG^9=gZ#IPvl?Duw;+jz+g}?Xf*OpS~f4oPrI_
zh0;lOLBZl8)wMgWVeOxvYcdz|cEKsZZ<3DIIRBIEdY9W)F`BJGhb{)BU(cA@e<!s4
z<ho#n#|t}l|7PRA&H62+<n-YY_tMOJTzD&}#g=#Y$XvR3e_PMaRBouU=j>afMN0zv
z&D!#xZ0Whb>45C{Bi{InSNL!D#~!u~u6UEw^4(bPP{+k$mRh;yu?%z+yG^n0G^{K3
z6F7`&v*mVYb_WHlY7h@_tF))?OZHoDp4taac-Q7U-4V?QUQ}u5RDb-M%?(#|+TGWE
z!)r91Or`vooonlTzekU&dplXuq(c?#=T}bU@mg*DRMWwM(-%w~p8q~aF)SKimwh=u
zD161MyxX*7&r9d8x%)KrTXatUFmH0(R|1c8v+zC7M&}3(YMs>D6*-G_J5SP`$hGS~
zYK>zl3o(Lc$di0H=vqE}+S7L3dS2~ZhhxdQW`-d{^x>dy{t4s`wxt$eI40FKFEv~$
z`EbxTW2MS`waN&?40!v0ICuKRxB*Tu0aHR&CHR0ZAr%M(P)PJ8nUjYe)=2ooES#_a
z7DyL1U?UU*Vvqs?AV6ptSVoq^i9%3Fl3CwDf$5_6&N?3_=#)wMh@m)>@+pI5nR*|~
zdzX8%Mz9JxP5=N1>ZAmiu#_Sxlp>S~Gf9d(_r#}=?l++uGY|$MV`jlDWXw7mtfP^m
zM!z$ZzglhHg5!iSFh=HW2kp`+Z4Trh<O_Wf3IkzCCh;&now7}WNeHFGbcAw2E}0ij
ztb{Am^-<tN1*|}*h6ZbBB*)f`V(;E+`WbY>hGRo2(+O9OD?(fj7oj*#94SdBAO|9p
z%1K2?#Zi%3bfTP7j!+eiQ$-^=p<R|%=T?717q}Pnl2d6dgbU?V25aG3xqfD1Rr&gc
z$y0;6MYaT%NmLg{=!mSCC-g)p00xkxFH2*7*hbraoI2Dz80kVh_+zm%E5wO`f0POS
zQ6@NnW`YxFCiq90-~^fpPN13KA7z3QA7>&<GWQzZ8D%)L5gAMtH9<au&(h1|go2@9
z=`&F##-6V_eLP3!&H^kE)x{K;B4iHC5sCm2BC1>@h(stFL?fgGN|AmjoX7|H2o-<=
zk~EW!JDWq)8r@Eq1Lnk28aB{IN|7scMTiTz2!()<cuFb`#z`rH5F(TcQxVDr+3}R3
z3aX?OmBVs`s%Wr^M)IcXj<1T1^Y-bKaeyPrw$K*Qa)WLN`9L3pLP02^EQCTtD-k9l
zlm^of$^kiuatSOUrEo$EwFp(yU^S(@GkfXVw%~(Tbjl{c1W~q!_K21{bVtY!`XLk!
z!VzT=6d_v4Fc~2^lp~Y}@(|@xSc*ulhN}_UK!Y17W&PUgw%p9~Nx$WOl*NoJb?*B~
zciCFgt=;rz;e-p_1tD*`H$wC2^WCUxLPD1yB%{j^DyA2^QCGn#`YMFh)9CAIB<s)8
z`OQ2a^?tppb9rYQwM^ZurU(>~cTb$qKuxZ`4^C9VN`z`@u$D&hP@v)4*(~cVGn?6x
z=}b+LZ^?J2rlGJ@IO_|(YfE0+HOb7@t*;p{OQgCuK!*&9JfH_c{?H$xC=iuM6<+`s
zWKfg>QxM9483=s_K1<YV;Dj2gGbmaE*C4c!1~<}3o)UhaYS8<I!OSdOm;y?Y@4^>Q
zLn~Yq0)0W<d&lzZnU@fCX|CE@_iI3Wr`j&+j!hwGGE7Fko?y8v@|G9qP?Dc$KsTH<
z+lVm-H#TLNh|S{TJk6Nm`R3LZY#ZCK@W`m>7y<CN^oihv%ys^M|618QIJ&sHxqEnd
fbA1B>=LPdaop^SBK|)Kb1)rvVB1(!)pus-@MqwtJ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp.8577D499B9483812.idx b/.cache/clangd/index/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp.8577D499B9483812.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2fbe2fad4571cc4a576df95f0d8de595ab3153a5
GIT binary patch
literal 3332
zcmZwJ2~bl>8VB%1f|BqcfCLPO9N|zvL?HnIp)m)Ebhus!_aNaA5L_TZKtW|j4#5k&
zW|SKhRJH^a!BGbVDQBG8Vi*q!7*W?_M+MOlK^a^&P?fy4YE#84e%<Nre}AvvtAvUK
zf?142;wFmnGLjVW6cUM~ul$oMvQyJENTjn=5-G3xfn+3mgroOjE>YSh8_e-&S`uAp
zB2I`tZ1>DxSIddN5A43hKVDF1WX&Afc;;|RVM?5E%z0{T$aL*)?56bSm6%1jjkT@k
z9-Y*#y*2Wxu5aw=)Y~tuT_+D-2r_hE{rvM^&wC2;KhN;83UeCssO@>5zBi=*V{>lZ
zf9rdjyOaC=>V0;-E7sXAA=I$%THnQppSr(z?3&Gt`(yO*OaGGGd#Rfv+J-qlr}f)T
zr}ai%<A-PVejIh|-;)vbCjV@EM3ZYvB=hJo$ZTl3?K;~(egB-|kMV<A2ahwov;O%i
zpwrZ?C@|&JIq<=tBdAgS@bui`M?UxOwpO)iA2=~Bi5z_Q{M0_$laO!)$7o~K5h|Ge
zEiXyjY;@t98&`{XV^0)~MMk#netSm~b&3mKri*ygo#(%8@sss6y^XiIqN_<yJhs9u
z*ZWZ2#C}ewetpiB_NS(1Z_5&bb*f6~&gGkGEZ`CIq<bwmvmj;WQ0lSXT$^yaMeYyB
zOikA-y8J7oiyNLEn8}jt^xGWYTl4Qw>P4Et)#$$a2Wo?xt~d8^PX`AGJESp=+1~AV
z`K^+|iW{TB$&xCbR_)o0lXppJNluK*_c~(kJ{#K6()IgyJq`YyRyMYKk64>*b=mVL
z_F`p2!<1gf{ekygmu7}NU;D<y`L!LAeyl&;_4vWKzj%Lg#@NG~$Gi?-e3L@w^gHfZ
zD`j|=)n3SS;RW&TdB|DU%IOaZvRIq$rJeXb@GAeO60vZks{H<cj;0B!?)2;Luh+l$
zdiU?mfcr*wN+aKs(kwmDne%N~N!7a9gI}ta3|SfHl=Jjbwd*_u%-^1b|JpbCu4cf!
z<J>M+s_nSRk>PUo4~wUQv)j+SvdNBOJ3!~PZOqFJXI4-|HJ1h|!yMe>2d+(O#l)={
z*1sfsc7C9*tXA^+vCYTvX>8k2{>0F`D|z-2!})J$BeKrY;<Y}R4wlg!m4ydVf9W=J
z>NXp@?&7q6WJ{zZ`WKt6M{YI_j-7Je*jMOhd17MaC%Nwi%>D5u|Hceg0seC}LuZ!r
zqgiZQC{ub~T9&@Of6$}(^n32Ur7g>gJ>O5fXAbQ%>h!R<Jy+AsHJ@9f^{8c+F0b{&
z?B%ohqqkmu&RpGpsr+2E>r}UC?VR4k@#ksQ{&;nvW@AXv6?uM!ELQpCnJ4aCb9YJU
zM~oB{O2Vm{RNlfx*|YGGMpnF@oQm<rNWsc(H%&Lu!bR1cFza|boHSmHkwTQ+@tX0X
zg^Q}2l%13;murzo8g4Z=ueXnX)WC2WparNG#o2%zpk4!)0a<`*O4>~Q=yw5Zd@+VE
z1xt}3hQJV!5AZ=I@CCkzwty|_=`cJUq^qSrSm&!3GiRa9vm7puBJzl!D2m9F3v;7X
zc^EyRPWFeIt0@=`00<?%B!t3m2!+B>L@_W%E!84q+Cqi*_xxSo`qazPCH+%l=jp2s
z%6S=pAy$10!_A?2Bq2L!hiElit(MpIXO3x<RnyA7xIh<VFR?Hd*-Hf(R*==Z@hcr8
zeV$YdQz`f|)-rWHDYzBO3XucL0TGwQRacjSi&<hs;jC~(S*$E|ttt2p)(%9)WL7a*
zZDEH*T57%bU&{3M&_06LFb?quqO3(^SVWFc-HSlu-0G&2BM}&e>jQleF~|&>g%Pra
zwuqdeGa_%`Eh47(hyG!NLSP7@Xc&zs6{Lz(F)%y_=7bU21$QCZONM*NYU3kcVlC*L
z<y7pef@5D59K+FqV>nuH?5l!fI9hNFM+=U9Rd5Xdy5N#P1N%QKy*1|_MJkO>REtez
z(}^0$s4}{0XR21^$(?@#s7|h`TiHDS<g1#K#C|u5+Rw<8JEa4<fLK5e=pixy28cX?
zr$Ds~hI;`oMC-wNM4Q28fof$8PXQ^2(m<M8iW&3DR*qbuJR{mbn@_}82A72pvVvBK
z9H0Xt4&d;KP+Z6jAtZ)kMBy+TQ36Qd6A`muRtTXTa0jAdGAt&mUG}T_#8#ceYGq^$
zU`S*pWFoPwp*12$=!nQ2xFeBykcY$ygn@`6U<9H>kcdRihS};+7+wGi5S5T&2@$z6
zK5p`oYh9l*vM$g?BAY@}B$f@dL9`04LbML7Lm~^H0Erb0gAql+C`8F18Hucb3MAxC
zxD!zs8I}=|FKhM=EXuz1gYt!&0Wze-t=R^*$q4Nx!`)=W<z!e+xJTmiD$Zi?j`ENj
zQ;e;N^JGb}M6{B!643_A25aI3K?+19rAQHFQZlWHDBCI95$z#U_K?+neu!uOp>P{l
zc>{ESj*tjq0!<`@SdfLt2|6M203Jdj79a8@go0oYqDUBtC<!D9iI8$Amk=t1g@{VY
zu#~L!3m47FC3NG8`8N=QW<i8xGuRfy(lUn3LN(#vydT3f9v;=0A7V-KMTS_ieUTxS
zGGEmY<!_Jl`1v0q^6Uh;@%%rM`OCsb1CJ_gY2zCR?~o;=gp9=W6lHTjvRH$nxp;{d
zRU6aMqw9+F`CL1FnrMT8@lwNOOpkS5-s^pQfiuI=lf_wK@&8(lOwG(Kt*mWq?N>TD
dt#(=C%64DwX0mE6k74A$F>H|_&`(H){{wDMOX&ar

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_xdl_int8_instance.hpp.78D2B94EA941EB43.idx b/.cache/clangd/index/device_conv2d_xdl_int8_instance.hpp.78D2B94EA941EB43.idx
new file mode 100755
index 0000000000000000000000000000000000000000..750891291d2a03016f3a94b15763bdc025b82b3c
GIT binary patch
literal 4024
zcmYk94~!Jm9mnVQ-U@RRyb{r~WahnjZ)V;MG{v^LrLl(cuceh*S~(Hgs}WIdi@9U#
z9d;yz+LeY#jcsVt4I$?UtrB!u_imM9Q|$`dv!bLROD^%crBVr|B`0mv)M$vkBRThb
z$!+d)pYP=}mv_5+v-;*0D~8%FYxzAlKhU>v%Uv1EvSv?w4s7Y)^dqvYGu*Ns7#!LC
z+DiU~IXmCJwdY)Q?xD5I<{d15cK<_fE*vzuu~W?xPj)<V>75hL9p3-FPk-s1UtJnJ
zzwN=*@2whryYm@m?x}TW7ae`|zz46-`}XMH7yaqve?~<6vX^i8-?lyHN-u1;X7BCm
z{=?a|E9PB&?Y*7%ogW*UwSA~(^O~QZ-uv~79dC|2vV6hP<H--!ztP!waJ=^9Y5V)t
zb?v)uaOQlFd$01!@7|Y}f16xC)Fpaedu74xJ9jSb96xr?=TE*q??HEempF2%W7~>N
zU%lo~ZghYCm#gRhaKVkIznnb0e(##2*X}vKY2%AK*Nr_rf8nCufA;mq*H`{>|5e>H
zMxNMpq&MlD^PAhg(WbgeyKmvP-M?z^j(xfHOUe(;{8{&S?@x!9EnJ@e&g@OkuVihH
z?_R$2i}&s7y>QJ_kLktF58bQwteU@a?LXSydUS4k<F++3P9FH{<u$h)AD&**N8io&
z+&SDmawKCmf4rkt9vmI}+`IkbhYn{yzIb^1%Xi=T#;e=eXWr^RGqAO9gPJ&nuD@Nr
zL$)szpJG>W%f4>fbqlUMSZ$0Z4vK8D`Zx9r{Bf3L@y^FydSQ6{{aG_@C+jSk`oWpD
zn{|OiRshMY1oE;Tkjg3`oz*~o)&~l*0Z^C?flSr_McD`_&c>e*@Bi^lH`XsKo;z{*
ziSv`z<vRs(9;Oz%i3`>(kc(C%0<0*IKr0goRu)L2^%4)PS0E3qN)%XCAca;Z8mumm
zM(Za&Sie9%+8_zQ1_cVxhDiuEEKrEnBnGT0kU<+I5!k3e5!yJ3!NvuOTXx2W`_At8
zX0~hcI8t(+f&*|$<oq7kP02;?mdHgfQWEGzi3EC?l0+{{B++{*dFZ_odFWM23cV_k
zLa$TO=yi!SdOsx}y<Z|9eUMUsJ}6OuK1?Y@AC@RYZ&EVoO^FQpD5VH}RH6udoKlQF
zE>Ya_i$D7B_`wTXE8P<ZRb|ezto2NB>dg5wa)UW%kW=nF&S78S&Jc2qJ1-%ZGtM)a
z*50i$7kh4<x!7|X%*CFYau<7Ug}c~uYuv@2Th6%H6ICX#C+bXKPc)doo=CaCo~Uqv
zJyGKVd!n2X*ppQzu_x<HVox@h#GXvK#Gb5hi9K245___ok=XO9%)_2nXCC&v2J^7z
zrQE}wSK%J^yc+kg=an-a_EeQA?5R3a*i#Lru%}Y4u%{|qVNcb#!k#K;6!vtLY3%7b
z)7aAurm?3}uCb>pTw_nyxW=9?XEgTwD)X`D*O`w!zrlR$`6>6Y=U2FoJ-^0%?D^%4
zkG-JE0_+8K7GN)EumF2O$^+~L6&_$OsPO=MK{*p(FRZc<dtsf0*b5sh#9o;45PM;T
zhu8~iJj7mD&V<-ARc5ef>dauzG?>AjNx8wEsc?fmQ{x7ErkpX@i>fTbUQ}lh_M!%h
zuotB~!d_J25%!`QkFXb&GZFUUDvPlf*IA6cxWQuV#VL=m7gu<Uy|~6>?8W6w-0Dre
zj?#=P|Is~q?4x~$&fVl~{MOPJ3&RICL=&&)Ij3)GH9>jjexM>bgZL(#b7AY|UD&!s
za&aqi0=6PAU@M9wa4T~XwlXhaD~lv?>*YMydU+4FUXeW9s+@wY$}8BaA}QSJoQAE=
zYuM@{Y25lbAGUtphpk^EAGbj+fNhWuU>g)Ez-^cdVH@T{*oH+4acgo0wkB_2Yl>uW
z8|5O{M)?S~QIR6t#<>`_aXyA^T%@?Q-Q0cZU-hqz{(ACWph?bo{6lGyi!7QXkY$r3
zvezUJSv5%^>n3SrzezrF(4+u4Y*L79nq-ipCPm0`lj0U{`Ebc|Pw2|D$zRNNI$O^N
z*LHz~Er6sgfjrv-Qnmupwg&QTA1JT`pwJG1jBS7-I|7RB_!BB!+_Ck$=LUa4GwqJH
xj;p6Gf8rjl_pXVXJ!_!<?g4-Di)oV&S3a1XZ20$yWAv4PO`m2>pZd0C{T~Q-?v4Ne

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp.C31E228888CDC52C.idx b/.cache/clangd/index/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp.C31E228888CDC52C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fe6372ca63cb2721fdcb10f26dee22a4a0f5a479
GIT binary patch
literal 2610
zcmZwH2~bma9tZG;gdpL?5Fi5K$Rz=hYmsmS2oDkhN&W=LBS3?c<%o)kQVyjoI#oIb
z>av20jv{zfyHY5_VY{yAE(Mehi&_;y@c{9ljtByx!nSO1CV78nHZysdPu}nSe&6?h
zhe{@udQcEVoGmNP)fQ%u5Coy?zXgT)Ia*T$QClEL@$vq|;jP0#Rs)N+Syy8NYME2w
zC-+mcYP-jO>^*~3^D2C!8am$JYM#7#;AQv3vZh~kS7MGkTtC`(G=p4aGSqtKuwq1c
z=XS$`%NIu+I_iF}y}{Z#`RK;XrJncmk}qZ=KW|iednO%ej<X$1cKhatRiAFZ?9J5n
zUGIKl2}|d<)i;Ymy{Cr<@`e1U`hTVVFg>s^yno*{+Pe=EVL7|9ivw$m_PTzbTh=(Y
zF26aT@j_~0aNxSh%hO=!+=Y5O=8QmCc8rmb^UEQ+P#)T~SQ^(vx%Tjn@~PsZoi>BV
zt`uI1uo?ea+F<SP6Th%ZxFKzR+sz4a6R7z*El=_Y&q?b#c+k(|tLU^__fp(rRD9&2
zc&tWHr|Bqp`<$%GH2=<q+U|bQLO#pi?EZA^1_f6>mCBG%tradsG+s-QdojCBa_ZvP
z?*Eo<Yp*_rG__SyACKfk2a^ZbU%nqXz5MQnPRT%mpi)@VsXIqIx<0t{NVb-nm(_oz
z<yW!pFYOz(R?OGs`_D{U3N0O9+a8fusjGU^7&gv3BP*xtnY6Q!UYxg@FPeJabSquS
zj&ojfCE}U{g(`9i<PP5<$Jeynxv3n?U%h|B<f*4}8b#;waBtP2Z@;xYJ4zOpUyrY|
zW}eom6o38OwYq`9yY$7poo1bpnt`&zPj(Id#HwcY-ydq)?X`oi9y}X16zW6&Cgt!r
zyX5oHKbO~A$BzA{@|93jIMY-4RqXdUnZjo)KQIG0xr$Ay^yHTgGMDgZ|2@*)22*+Z
zGHO%E5bc-SoFH}Zqq`PdrpNQC@M%|CtJwRcZ3X+*N>P!67As5t;c#2m4&`HsdU3tA
z{Jy}wow3=EdQaThbL`x~1GgIzGD@m<_n)K+qBt>`*QL=(=N}x3n!Lr!+u5@*8uVOE
zKC_P=cE0<u|C_lT<V&wQ-fbyP?XdBj=qS3F+cMm>&Vdyqs8Z)<pUBco`<VCnn1}eF
zim^G(^JhtZ6^^dHlY)@*1=DUKwDZM{y5-&vO1qpYe(bleWEUj)PYsUe7j{y#DNPzq
z>%Fk_u{{A-X(q~(CAOzW16|j)Iy7zB#ox2Y+MpBu;khRzj{lt#;5$Fh_IpqZS`W<o
zWxR?Tnb$gQelkmwzkXp}z1MqzVwRlVixhmBn<>&ibyX8>&D||N9w-u*fZ!Hp77-se
z`aK^%eHViMzR6GLf+F$y-Tr3&(vKU%?tJa~f*vyjG4-$ecSp;_+hr(D!6*u2Q(S-v
z6owv@`IC>z6at)t;x?EKRAh_ULK0yj;~^*>j)faVA3PCmwQiB6pW*^sG{h7ckZBB4
zLORCQY$AER(--*yU#KV*q#BDNbgz&PvxWD`B;1~7ZyZj-IXn&|51t1kF;8raLc#$L
zKvMJ6M#Wr6Oww^H`t=FifLjU?Qw3BhhM4&9?`bSrfD0M@Bmw$K0u+Z6pg5cW{Uia3
z!wC|V)^6j~p=L{$!-7I15(q2=G(&={?TV69Kb7+JmxF;Z<V2A(a84q`2Yg6Az{e<U
zY{cc=E$1@zr$EDKQsNLR%nFh<W(`S*38jVvC?0}^KnlacjLO{+pCx*6C3rcC(=ob?
zXzm1@)P!7tD<p5=4Ji~0l@X(&LA07sJcx(%8Tibo!g=TF#(;uC{a{PXGKv^w59|{N
zaR3LB2k?LtfCWSmqr^a*KnMT;Ne$FSmG?PMZlz`)*AGT96dLRR9H8c0z=gC1tbr7W
z1ww-*Kw@l;;;|qWQW8iq%5YL9l#h>o<<AQz8%IOf)fwrtZEH{T=<}?=RzP_ifg?0?
z6<7ty3wS{a#)6?dQXn<XMDaKf2Pp-l81;|SH>t=J=+xhs7>30WYZn7zQVAu11V|d7
zG3w_;(QoOj!E5>((~e}vHJl%cvq@}7t4XUNMUo=9L~{j60V$D`Xq2HwzgkPLL&Iu4
zt&%v;?A$KauND=f#t<`^fSE{$2Y8S?fhVLOEGWj%9mS(SR3afIP(n%u$woc%HJdf3
z+3j0;))^EQF-pJ?unY|rB>z;K-mNiRDq>T@p(3^*+)#w%Yqw-9z1a)$vkL6=zoz)8
zvfYpNenM@cLR5x#K7wTBW^dW3r(%7qDanj%zTCnRrCQNe$mmgGKAj?qw6?RcwRZ>#
z4hantg=5}~H9}s1JL~`byVB9gna$y@a&=qn;pyZ1r?rAW7k{RgpM<e8TAsK}s*H#+
G0sjM%=O88k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp.108EC38462E7A7B4.idx b/.cache/clangd/index/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp.108EC38462E7A7B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1fdd9321f22b51a53975f51f1a9bec3d082ed20c
GIT binary patch
literal 2596
zcmZvc2~bnl9>yaHNWz7%g-ya1_8m>w0^~wMfN}za3xus@3)li`3xUCrJ|AeMuOdrP
z`m`89#3?9!;;TN2s4NB?C?HrsltL@7)v^e+QBm~y0yDWe&h*aQllwd8-0%C&|K!}*
z2tMDOLLhL(5hZ)FB^e|FfuN_j3nc{~WfBR5NHT#?(l8SDVDAH8-O0CxSpx#^ug%uD
z5AUaBeSPck)!WTz9pk8HXj}LC#J^T!zVa7dL-XWK9{Rg=9{q6I6xB#S8j|8(#ui5T
z$bK17ANzS_dA{_2A0K$}H|f0Fz%)@dqP4%nEH&|>dm7bnT11pZe(`Jn;JK0#_RS4B
zl^*oQXG~;iVf=C>*N<LjGdAxgdb#*!Wl*+IvGMx3l*v(>8_99IByMT1IU$XFGFdfQ
zT<!ko$WpUNN=})a`t!;_>Cw@~oTYER6Ax+9?W7q8Z?q)j4gU1Ut+kSqy#~{#Zb+_$
z7(71AZ&PEYH9wD#jCRTAo7hKUPYU#PdWC;VDND}M%HzUq0gL6M1AV7#v(@edGCxSV
z+a|jp+PBwAeab|e8Poh<d&lwMshZ*41NFS)KfiGPT;J9xV6lpyxe~L6Ug9-<zfiVZ
zy8V85=-=<p3?11gDQPr0)K=qK`mvC-R7w>V9~ZP7*|{E5<bKz+VzqqspPI9Yg%|VX
zcb-ah+mR3%{n^5~*X5&qRcu*{{$Yy#_YZFUSmoaM-u~#%Dngp-YjWE=6U4qfPO|j=
zj)ACsj=P*AD4NpNJ*S5Vxmm8X?<OZxhwjcCaO|A8-q#fNABMf1lVfN`r-kOY=@1qY
zy3%wfHf(Ug=vhnAj|V$%l&4#_=Qi8joS$3?6CBIQTb}EYvg<DA^*`|%(>h$_N|ERW
z&Zkp((Pig9`Hg>-wtC^^Lw--&6w|A{pt@A+TFq5ETI;aOWTkEO8HQOfLmb~f(SmZ9
z%(}@^o!-`C3q>a#&o+rqy=_gqTuUzy+DG$hudcNob0nI*&b-pVn)mw@yU^uRmpJil
zZ*oe9-8a=1Q>pnM`8)ZxJJZe^o8ChfW4<&#dv0{*vF3>(L6Y1(P;FXNS$r%1-g>UY
z>#W0xS3ZJIH`s@`#d8OyZo18p>x<{QW~c5w$@sRC+CaJOT_iq}mHgC0`|0rLXkq_E
z=k@yi9bbWmQRE-kBf*Oe%Ph{9^N;L#Mtg5~Rq@{6U8zyy=i<DcC-Lh_&&y71&9%Nf
z6FC2{{JHFfoY+?N*{QXZ5Yw5;DHdmV?rn7|mws^cgM5d&j1`Bv`cHQLzACU!3Z|X(
zeC{A$&n8DM{$^=7JEmi$o${R4*_UcwxPMP(kmAH?1UuC9^fr8m5MqP`jH%A#Z8eIK
ztxMp0n-?ec`w1aLD~8GHWX@Kj94^StDeP7!5QxmvJ>4BE&(si%f>I*!w=fpUid0%C
zf45P#EROYUG+z9LZAE$2ICm(Lg^>`<05yQ#HAD>|3POW$B?RN39Gvu~&v0~8-&!in
zEr3O`$~*!@Br8qz%M->As9k7PxS%K+rR+wqa1f49LNFl^LX-@WabkBbxa5-DUMO5V
zfhXjW0#fiv!4=DdmvYH{4H9O;Fu~K4Fl&Z2M9vInh`0<cz7`S&7yzPpMm$d8SK<;X
zZ2mE#NZ<e*5>@Mr1+j_B1u1(KMc&*kWs$K5Z{D}0dRu59?*v2M35H;BFa(2xA@2l3
zFgRFTl)l5`+5q*<-6vD1$_NO?BD1K<VA+D;>t!A1GZhJ_D3!08q>Ji8q>t)D<d6FE
zl>!LHM%fSrqJcO`UeRkWdKXGG5ljo!3RQ`l0285#tbsK|&cGQWZ`3<fWyJ+tp^5+i
z5XFOdoIW3>ukO_rHz<q|6oHIQfhi<z18g920WJ{vpgxc>5Abkt1QP%OL<t}Pr`n7(
z`M-CaktvLIP#ws42iO6L+X7pN+<+TIeyAU0%m;j29KoVNG(?FY5vRVmpx@Ihrw0}L
zXiPG;Q5DyUWCf8E$qAwmQizSJoROqRh~h|bIQ{CWF5gZyKDzl&qEReW;w&1=QYj$6
z^vC6tvT-L}1k*;f@q>h5dZ->mX249O5-<nm5HSD)B0FFQkt=Y8$Q`&t<cs>^#}UB-
z&;W=+L8wR>5y8Sh7(`JZ3ZfVg1Ca=bAW8yBIBoE{;)zoyh?~{0qHt7;VOg;_N~_mj
zjIV_4V-_hsY&4V>p_)Vo^mrB71AB-(fCo+m+4&ip-}AzPtU}|>huF<#>mp(WJI~8#
zAZ2!2EFo);IR8UMcVv!$NK)6>u1VHHv~{UE5qhCqdp$}-h`zCbp^2%lA3Gp0h=aP*
yT>Kf{j+TGbYqZ0hZe?v_Yv<tP?CRm^wTtCr!8CK*$)g#Cy%(p(j|vXo2L2DSzU#;U

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp.7487AFBFF492DC5A.idx b/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp.7487AFBFF492DC5A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e8f0dd3b11cf54c84cf631099e1d9fed115af760
GIT binary patch
literal 8312
zcmZvhdwfjSqQ_?~5hIC5kVHHpiFnPP=bqUURL8vc?AhZnNsLDkW2icWsVG6Gir{Ls
z9#!qZp?Xlob?Vh4QhJU<M0%ne=O#5JimNFTqX&IBq4!R5KRfgL<L2|R%y-szzVpeR
z$=ds~c7fn@2D4={+YF(sWadJDf=nh$jr}WKxL|hI7+g=5$;x(|_KEYq9on^S*R?{+
zRrk?rb>`0Ny3b#EeBG5j)zPW)ZGD9DAq8T=8&B#N@8_EaKiGX~9$A$;YpDL_h<oK5
z%e!u$8D1cq?0$aY&Jxr5C06w(e{Q-TeRlg!TIGj(KWTTiIivB~>p{P9q^Vce(PI+J
zKTynzX7#@neRTi!SG!I=eH1+Z{jP_znqGgSbZhR4Ba!XHwl5#Ce8-LDzn!}@GUx8H
z^^tqWXI{U2cB?u0>jA&~&cReBUO0JcY2^J+&Tf7Gr`LC{Q2jOY@<K18`joZSck8@y
z<m7^btMAOKyc%4-W8<L1+~Kmygzmu(&I9XTIU{>;=gP3i?Zxfq^-h}ksO+1BwHKGw
z9AcLAC~&sFZJBj4f2;DxJ)Z2cF6+nF9iQ>9=#90%<k30hi9HHd{P)1YSB@OAytnLZ
z`iAUv#Rs#Ka&zvSi2l>HzmGg?qssD1w9Gv0rmo~2OZNN6_ICRE{<Kv+S1HJV*MleI
zbx(J$ZrGi_x7W^fU1rW=>lT!?XY;vRRohBt&lxl6^b7C$5?vSjejnhUC0*K_w)ib^
zz((fB>M2t%EL`}e`|gX0J7**v&(qbtHD~q8cBWH3=YHe6e5KRUH970vtj@aT)1}z&
zR!uj*H0k`_>km#3${c@c(9j>h^X(UeguPwYL^5|TN_@U$&$%br$3MJcD5*H}S?9-N
zMvrlQmiPJZY>vu5-TwLXp1PN3H-$HHiw7>PKO0@LbCx;%wXG}qpJ*KRWPP~*<q4Or
zR~v7K48_;wV|~4teFF}!&|FlUo43w#|Jhr*zrDA1`5Y}3s}DHEMw<SgbAIhqc0J*f
z1s4wH9@yV$_|?}JKiyn=jbu2Emn)`cm`?mhx0S1R)z05nyL9y&&$xWOS8U>rUiK+B
z{nsA{C9Ahi_=9}(!GhJtUbW7(9vJ&kTG6`SA3I??Tohej*Kj^3xYT-t_l&;!-j%ZQ
z{%MaFr$7Fgw;C_{A0OJKYUtQ&S;O#`nOB-O4!P6rg6tn(9GLBSeq_emCuVH!@Y~g&
zJAZB;aA=Fnn0@S>fBwCCy6T%pkwyKtJ>RT4`sB&P8?K^p)+4e{kA0l?<%W0pz^Soo
z(hqdYojvvNy+LQnGmY>3&!*k8N<LhlzcTaMjJGyt%vrLo;DagOHf`$al6^4mA03qL
zj`wza^(y<ftS_E7KKydu#T$2jDy(c6SUUHY3umtl*KA&!&?tMM&-O9r7O<K7X0o%n
z(F32KzL@uPNdF(?kB1zrvX#zXvCV&U;FhdkWrau1-Q53|nzG7|{pYs*yCn3-q?bZp
zPX7DOT?3c)d~SPqxv}`oeRpP`Iq}-BBOk0-GxT_~<|U>}B{ToXh+q5lNPV)WUzaVX
z&y>|Z*Ia(~Xi43jlB!9T(wrZE|9RM!eJ{21@ZQ@x+!^NnvhS4dHsp@^Zq<s<y6<@7
zV*c05#Ggx-luTbpWU{1xf7Gx_<O-9iZ=%fKJF$0a>kwNxT7Qf;t4CEoeg4;ES`?ek
zNX$rX9i-`z&Hs9^|NA9>N~Ysu(`kum9a;xz+S}#Ar+ZJ&zne^7icRMw=2=<?X}X}W
zPxY(8P;MugH#VJ}p3Sum(sXF$siyG@mq(K6q}cTE#NlG=AWg5D*2{TNr(cvzN5`g9
z6H}8~2Wh&Jxl;UH)5m8!$;QT}`=<9D+&W0pr{5U0rh9$qwq#luo9>y|Gr4t;rf<wx
zzy1E*NwQ=*E;gN+n3>!<NYjrOUE8qG{8MU*Y(i{$Z2PhH)<K$n-qm;X#{6^NC)0x1
zbhpHA9a{%!dcn+>OCM&&k9|Q>>oI@lZ1YElZoQD5;#c?-e(Bi7D4>`J6jUsW6XRpR
zWC9GBOppOn=~IHK3@E`=29;o{d@3+i0Tr04pbAX2PYtFzpaxSNRD-GUX~5J3G+=6i
z8Zfm!EtuMX7EEnW3#QJe15+2!fvF4Xz+`<am~4OrlMS+9>V0}J^#MJY`k)?6gU<k_
zAz%R05Hx_v`8Y7S00$-)<iIrgj9?lAMlg**Bba8N*)RQlr1;GN^E{xSd0Ct+J`0$Z
zfCWrT&;q8_X9d$5u!3m~TEVpWY+%{~HZW~L8<=*V9ZY+`4yHY52h-tmfawT0z;py1
zU^;zHFr5J>n9iUROy0+X$p?5a`5+Ib%jW{q6>x#+3cA1)d;*w4Kmbz+3Shc@ZZO>e
zH<<3A8%)tBf++?>FvXw<rpM<2(-ZK3=?QwkoWM?av2B_1Ph=+oO=2gtk+j0oq}G)^
zz6cet1^H4Se;!j_SbDi5N%8ko_tZ;^r?mY4fh4uA`|%emYKh_!iLX<N2BNr$93_f}
z$Q4v^ih^H56<;BTsp4B;rj{_+XBr5DeI`m6>@yXV!9G(%8SFD*%3xnvOO)7GHV`HD
zl~JO^zOsTUv9GM5O6)7cREd36Em2`#)j(9(S4D{m`>G16!oI48s<5vLQx*2rwM315
zbpugjUmYcC?5iuN8vE)Rs>Z%LOx4)e)DjK$H4Q|AeNB{Tu&=408tiLos0RC*Fx6mR
zTT8Uq*ESF>_O(%>#lE(JYO$}ap<3)~!&HlXT`kdJU)Ml%*w;mg4*R+as>8mnhU&1d
z3sW8T*;>M4pKTy4_Sq<5vCmde7W-@sWwFnODT{r5Ezx6N-$3-(*GGvS`}zv1$G*OX
z>anj6Q$6+#wZwpZLjy5j-w-7R>>Dbm0sDp;YQVlBObyuQY6*vZu7Pma=c0teK373G
z>~l4g!#)?L9QKX1#E5-k12JOX7$rvR8!M<0`^Fk-#J(|1jo3GbiTPU+Kg*gU#N343
zOw7-atEl+`1;3t}zd>%K<{yA9VPe6)B|<FNw=@$A_AOP^f_+OpwP4@UNG;g6hKUvX
z)(Ej;-`Y&9*tb?uEB3AR)QWv;Bei1R7A7|A+akn<eOoiJVc%9oZP>TfQycbejnsyH
zdzjd<Z;uc=_U+BYj(vL-wPW92Pwm*ZH&Q$H9bw|Yz9T{$*mpD&2lgFR)Pa3RJ#}E;
z(MTQGcZP`*`_2e)V&B<JoY;3(Q787D_0)-dXCrlDpAQor`+S7(*yo!GkA1$1^4RC=
zDUW@=k@DDgg^3IMt_X2q-_=Z9*mqS?7xrEC)P;RlBXwb42onMOLWBs|7n+HHeW8j9
z*ca-lfPJBn3fOmti5vUw2ytWI-AvrrcUMt2_TBZ=jeU0`bz@%)6A}AjgoxM|n~8{h
zv5Jb=7wf5r{n&nsh<#6(c(Ct@5D)e}&BTL!PZjlG-&0RL*!MJ2PrPsb?d`fvgEzev
zyWI0;dv(%Ppk-~7?CFz*VVdBd(tpZS>FET2arfd05{=5=JU?Bveq-#ly@&R;N&5@Z
zrCsbYvt=J3ex-;dX+xlMQtL)QTr!8uJ76`nge-{GkQLDuvLV_-c0@<Wf#?i55&000
z=nA<Ig^+;g4!IG<kcj9BdE!_!y4#>1HjQZCazs0NyGRGCxNOJ#X{W!q6MH(fQ)&-s
zOG7$O#12-=wnzLp;DKII@#YVc_X<0D#dZkfiowzWPVqBx21qGa0;%LGAhlc#q>*cY
zv~n$wPObxD<t&h1t_L#64M3cn12W2uKxVmlu=Jaz_$_h^kX3F4vdL{gcDWtMA$I^d
z<xU`8&I7sRE+9cJ0J-IEAW<#?dE}lpl0Ko>mQBl!Q~Z-raEf8^k3eBkSb=zj7$=5f
zz+_Aem<-Q=spOPkDosi-mAn#66{iAIWm18u;#FX(IW?GSlNwAluLe`YX~5K&G+=6Y
z4VYR^3#Qhj1yjpw!PId&Fm)y!m^xkuCd;v4vL+TxmS@4#b9ylKCOw#XUJs^$Gk|F@
z8Nf8~1~55}1CujxU~)VMrjav(X*3zZH1b9;&765y{OfKunXN#)S&Wl~vw&$aS-`aL
z7BH=x6-=wi3Z|8}f@$MyVA@PJFm1dIOgm=>({8eZY3J==IyeWI4wC~+2k!vW$vMGv
znw(%dc_)}W$Aih6crbaM2h+v5z;u~hV7ho0m;xt&DVPK>1zrHt&AGvJo7`Z!c{i9M
zCxR)OL@-5O1k=NL!1S0rV0w5DnB&-SM(NCu;`gyWpz-YZHYz{9Xw~eQqcUS>jZAf>
zUfQdEZ|4AW=AiqJX^KB9H;b1xz*GD=xjDRaYCRZPwPfP09q-1ZJRxs%d|4V^IDCTi
z6+Bc{*?w%v$IjRrq;*b9m$rGNV(j$XvY!r%Ur4b5HglK|gQyHC5mg}-qB^8T)PyvM
z+K?7e7t$fJAr?^|(jyu|21G8zAsR!*I7%B@u@i2~t$_G%R>&5%ZGxRGGTkfgSAQXP
z=q4vj&W(RkQxm4P(fJ$Or#B{^^TiHbj*v4&I_&ky=Z6%o+dL`ubcQ-ZFKxO>A6@Ky
zTkbK$-?xY@YJ1<i+V)R6bjf%-_F8Y6x1aRB(*Lo&x0V|a(rcfk%12bIGH%E0EFo*G
z^qEPwBzh;c-inay*tLUKQ563cFkS{o=~V)$yec5IR}G}`YJjv}Es)Nu17f`_klw2Y
zGI$L@oR<SKdW}G4uenJ2w#Kd<ycQs<*9v6w+JNj{JCMWc0CIYrK)ja+a(P`qf>!`?
zd)+{yR|N8SJ#8dgFmtZI<sMM!f<S3u%Y&&c!~gwEmLAAt#cdDU`L}d9G3jKWdt&1B
z$y28kPm@h+o0Cn6Kb#(a7$})B_mvomFOO=M&_1z)ykl}or><Q(3#ks<z|<tc+AX7d
zkDi%BhrKZ3Mbk*LI*l3L-#ED6|MzTq?>>FAb8_<r3>u<PYP32%m!D<mrDE-A>AX8Y
KontI6;`v`k@d?cU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp.B17F2553F57B25A6.idx b/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp.B17F2553F57B25A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ef4799b3cbd8ab35a55c67b7f4828f12dc89820f
GIT binary patch
literal 8312
zcmZvh33O9c!pCzjC4{yt1-pSj%T8WimY3u~E0R5VF9}Q98Wvd$VI3^wqbOyGfD{<C
z$|9m4AczCV0E5HkSP}h(r9hox8MKuMLCRo(Ktw@DrquJLJ?D}Cch1*yxXmy3*WWwb
zq_^+A6p2n}C{HG{O%uyYr!NSk$YipN#J{oy^JjNSz&B|!S^2gTe$Tw`hIOsqak<!X
z$$j{8y}9$+?(<$fy7uC(bMeWFEl-OTLy9~_uRLs6xYu=e@V%WE=8~0tXAaX}8-BZD
zeMQ%;(<Af6Pr9EOzrA!s<sz&4!@u3V6Mx)zBlGyXyFYAqvN^ly^3qU%H;w4gb<}8i
z#XHKm@!tKe$A7-l_~njI9{n6T^TUoGXWm`<O4;VVD-OlB7HnNMeA%|E%YHw7!IXb<
zNoDNz(dk#}PHr}*eKp|Ue{e9z<!3*+zBu;whbK3`{maswE7-s1)-CWc?4LPnL$}WB
zO%sduueve!_@&UYZR-aeG#)HJp3*(k!MU$;*5|T&H!c>$8W*;o+cS0g&*k5ytU14=
z<^Z$ksUm0lM$61khHh5f-{s9K?@~Fo{^+!y;#b%FyFZ;%A%Ci9#lHp)zIf<><*g+r
zv)1LUow7eKwQv57WAU$ry-zE8uV*c@;^pRoYr4`mEO~Dq+1=@@J1?xvS*augmWGZg
z>K|=i72P>>caQCByG)<S)z2?)&kZ$RuijERd(P+yC!TrJFBi`D`XT6goO)qn=EB!J
z1J*P5&rO<qcEN(>?wilbw@*tw+Fw`y+MHFdwj1$j&Wqpr>n?U$ygGmF@^if}`*rE|
zo7Gdz&rdkB`^vo&gL21yI%wGa@BMp4F=cnx)v?^2FUwmt?K=H1@94W14W(6||F!dj
z(W6ESf9?PAA8d}}A2t4ZVpsi)*>@xBjSB}ZZa5iVw0)*I>!r;r`W<U3cvu<fS2ym$
zm2-S!*f8aaVvN5BvuD7;6`J$P({tBa?mT`?cj~P*%jRgQ=iEN0r-^3$WzKJ%%CDq+
zIREVazWereD!jCG;iHYUmr1tcXoYfW_K0JD?Dp!a9kuiJ)Gl5%$NSt+z0Y&kc(~55
z+VIVv2BoPtk9$usYJbtHBmZT6(YkNU`<W$c|NF=>+rg4}WqtHaerU1vkjp#j(pwkH
zEBa+VSeW(TSC^GPA9!$J2OI6!V@b4d1~aR9{g4~&&dUDz*}mD{mP6BCKQ?V+hu<&#
z+WBMqfCHOkeBO~azW&F#sqD8u$6oHYrDfyv+qaK=ykYOYim5HWnGW%lKE2jeFW8yR
zmJG0SEvMhno_hHw6Zg{r#ijdvPQR^Q-~Fkq#&_#*?44)3H7dUU=g9vo|7LLhdu0Wq
z8ow+&CI76Y>BlejoWFYWm*V5mfn|4pIeYSQp=RTnlr^$va<-2?J)g_nGmV>_Ici|b
ziSzv*4e5C+?ZJ@!)$QiYlWhqc9=NIZZx4$OoxZmBe?BNb{z2fgEj6W+{*=0P=@%3K
zv3={n!knwNrn;L`mhZVS`}1Qj{bstiV)d|t&6?+#F2|X9hlc;w;i-&=yZWr#cw$d^
zZHuPv@g1@LRH=M|r7Zt`_~VEz?{>zU!+LIcr?EizqSvJF*YzF!y=TSV?%S4}A6j=)
zoFyw;R62D5k;zj3^L})t$5@<3{o`eUp7Nd<Z9`(`X#47NpKRJNxS%qPKA)KGFYj+|
z8>H!skJk2h>ENAHY1EgP&X8wxY#XF${@S^b=N`5EJ&l$mrnBYQX>EfvZEF7M-rgS;
zeUwJWCZ;pxnH|~&X}WmB_-m&ZPps}F8=aW$mDOu-+aOIZ?y+V>(Z9ZWH(fR^F+HaJ
z7+c#QO>bH#9*eHdm!;7OiRnUlp}TF6ruV$(x6fGg?)px$F^TECtUP1eAWffmW#sDa
z4P{%>XmMgXN1l_`Hb~P~r&Vsfb8|u>{#;@@SDu^JHb~PgUHylz&O7}>8ZAmpcawMP
z*fvPh^QX@!`!P4U_C={}YyQT`=JyX=e<m+Ip!6#P(#j_&sGJKFQZ7jn<7dERf()2U
zhyhdOSAnSts=!o*RA91x7ECtCg2{$hFx7rFnChS!Om#>NrpB)UQxnvHsR?Po)cUnx
zYJ*xZwIMB-I=>D~T~G(6E~EpK^K)QwK@LnV#DS^z>%r6q^<e5ldN2)s1DJ-O0Zc>4
z0H)Dz1k)Haf@usH!Q}lsn0$~2lMnG=n*HX0^bSuCn1klIKq2#zBw73xFfBm~n3j+Q
zOsn4trZs2<(;BjZY4h8_v;}Qo+Cnxk?S4C$_Mjb1d&mx^!|wpo5p;m*2syxX`ki1p
zgHAA=At#tFzY9!P&;_O|<N{Oh3t$RC0ZbtzfGPS#FvXw<rWg{zbo<?4x`S>o-61!a
z9=``nPtXIVC*%Rs>-U1`4SK=!hP+^o<HkMvWQz=p=f(p~;3hmFX)mFvZ998%6Pna-
z(j+M$?emG9s&$i3N=jZ?AQcijl}Tv@5|o}KF+2k%V`9K$1O`kMuL4tLQh})wRA90^
z3npt~!DIy%Of|0tQ*Ba%sTR~=YIqHp8j}W0ji3Qj%WJ{ZnzUeQ1ud95UI(VmqytkY
z=)mN74ouF(fyoIRn0j6hrrx9nQ!nVjH1Gy64JHGa2EhQPkvD>AG#SA(3Pvz_o(GdR
z@nG@-52l$n7bJHhv&n1)63m_?S$GSW7Lx@`i(moM%3Hy-nyg@21uK{~-Ug=4WCPPC
z*ub>&b};QGJD7IC4yJ>5fax$fz;p->FrB;;OsB~Src-c&>Ec~rx=b!GU4jctffv9O
zOahpKAb=_IBAB8{1XC15Fx|WxOt;Amrdx1>>ES(KdQ2WLJ%R^JFYg7@Yx08W6}(_R
z$3171Hk$N+pYsEa<;FfC>1;BwQ?;H-CO4rXu4t$f7&@1!C@!n(NYVp2>Kwgv9w(h@
zzDH8q&M}ietgI!<3naNuDWgPr4LMGfKO$FA<);+<8mjyfIYO1+0W-CP!9Eiu4EC8g
zVX)6sQ3m@=4P~&;L@0xORV`6rUlk=P?5pBLg?&{ORbgLMLsi&UMW_n<Y%O82&qfK0
zeKt;5?6Xyr#Xeg@S?seB%3@z#OVrp`M~NEy>NrtjUtLAj*jLw3HTKmJs>Z&imT0iA
zi4qO=HF2WBzNU(5u&=418tiK#RD*qOEzx3M8zoxoYvV+VeQg!hVqaTBwb<82s22OW
zTB5_gE=qLR*Tsnr`?@Ns!@jPD>aeeiP#yNUTEbzUixLj|T%2&&=c*`&eXfRb*ykdY
z!@j<j=&`Sl5<T|yaiYh*zKZIxudktc?CT>`k9|WeF<{>iB?jyp;>3V`Llre(-%vvh
z*f&I|0sF>UV#K~NN{rYy#)%R8#wu#WzOjZHv2Tn}Blh`P!egJ05+3_}obcG^t0<3s
zzJ~JH=OdKIzBxk7-;v~BS#yk-?;<x7^JC;{YCcQBZ=mLHk(;Rb7O*8kEZDcihz0wW
zW@5p<rJ7o>Z)u<w>|2_s1^d<rv0~pEBUbENn~4?s)@o|SzO{i`v2SgnR_xm%#D;xa
zjM%VmYbG}A+p4Jz`?dyZ!@jMF+OThr5IgqmF=EHQy_wjtZ?C3y?Asft9sBksYRA4K
zLLAt4#E1j?j%MP(zN4Btu<vM~4(vOcr~~`X2ytTH86!^YJDZ6U`_5|W#J;nEI<fC;
zqE76)BE*G#SB$u@?`kG4?7OO|3;V7H>cYOOiMp^aM2LWWAw~r33(Z8pzEDjC><bN4
zz`oE#1?-CvB4S^R5fS@hGZC>bR#OrCVgnViFE&vT`|b#FW8WPkZtT08i5vUwYU;+m
zyMel~?{1=Q?0X`_gMCkoc(CtjCLZj2s;LM2o(Af{e&Wus2m9U#@nYW_BVO!#n~4|u
z-fHT_zPEvTvF~l7-elkW`|I@^25)#Nak}Tr^Xa6sK+BrDvPTaWL}*F?FJn^zQ@T$X
zD@h}VZk(6JR<2L{Y|mjmZPJ}k>9%9%)VAA=aHbemu7eXnCd?qJ!YV{I%p$78YD7&~
zgQyK_5p`i5A{XWm^<h1tA#6Z2hK-1Pm``HKsBVL9Z5ZCZb+J17x=8CFeNUVkx86mS
z{=+M(mfxDVTin(oab2%a4o;r&G71JrrBDH}3KmGMPy=Zc8X&Dg3#3!%fH(yQq*v&H
z3<?8~QDFq)6+DnxVIG|PJ7!T>fUF8DkWFC&vMcOB4uu2Asc-_h6fPh^ApnUA5y-7@
z19=o4Ag{vvges1{ymI#Rk-3SNC|8}Um+mvKx$k(o;H%0WiKjEVW@JgXW8T_6z??hi
z&I6hr=-s!sOM2&~2lD&oyQDkZ(h?*tn_BOJOD}Kw|M^{z<iwl4^<H>VGKbCE;OxK>
zwjf%=RzzFahG-Al5glO%qBHD7bcJ1rLRdf)!y=+P>_+s2J&4}0H;MaWD;JHwzU|G#
z8ub_Yk4(Oe3yTWJN^5hV{CN8@r5`vG>zmm*GfTQ!l&&5VOV|1lKyp=zx#A~RoF_|0
z+@8{>;cVgn6H_Mkm42Jj1CvuGKcO>Mw@z)6pY|sXkT2#>l9sz6?aYwkwHqfSp3YWh
z>!mAP>19tGxAn7y<Z(;5k|)Q#scrwTLznE=6F=+A^!1UB`#3`}{2ZIzn0UIk*n5oh
zo{&D0=$YE~QA9GS^eIb{i^BLAAeB!A#QInuwNDMC@o9jxJ}r>Wrvu`A9FX3p2Qv5!
zKt`Vti1+b8W}mr4dhyZ&7M}&k>azmbd^RAv&kp49Ie?r#Cy>kM0up=zkmwVE+&(vu
z$L9g^`n*p_Hh=nyf!5D}%H{{lid!E{ZXMdb+6Dgn^&jgphMh3Ur%s$aY03+-iPDrz
zHYNG+h2+CY$%j*u4}+!CUYwOc&y11nQrgQqC_1L4ck0@uvzXzq4a`Utt=+P_Kb4a^
ztl*j9&yFye)tO9TKYnnZ|L@tXo=^A6%kSHNz@Q;Yl}4-68;ABb^k6xAW|qqxq|Q+m
Hfq4HH-f-hF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp.D3A7F488899285AF.idx b/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp.D3A7F488899285AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6cd9679430e41d44e75c1a870c248afb25b527fa
GIT binary patch
literal 8234
zcmZvh3wRS{qK5PTl#*T$%B6*Jl~N^{TqhS=TbyK)OeTr7O$!4Q3dNwP6o@S#1&v%>
zwF@e=uzE^Cg!Q0^x`4RYD!T|JXx1X)VlhAgQ416*3#+GC<b>`s<U7xE`aJzRU;giX
zC!IDknLouIK^QYyp>WLhESdM{!dVFlg(5BXUA3_CiN3LLEJdMM((tL|TM*6Zf8vdk
zCHCVF?>>2AdB2*03s#zIj{W(_e0@T}D`P(!^iH;lyLmb9o2<p_e<|8N?eA-IOb!3F
zU~?re@-G#X%=>ukshVeM-`d#Ev3c$4i+!6u$;-RF<nxv}+m4U?{#MFTYiHT(U%vY4
z(ly8HMt=WR&E=->k$p}3@^5#KANe3ReRtRP--hOGoLMLwDet(nwB^T9M@P>2^^N%_
zPiiao-@mW^;b8|XU(O@uU#C}oAbt4Z+~AsF@Bej1wYtJzdNd=mbwC3t$k_08tvYLA
z_Bq9^gO<IeIafc*OZhdNRDHhm`+*BG%IW66iyH=fnX%>Bv77ej_D??h(Y=%JTYc!h
z9L<@MC(90Ot7X;w`#sRSXV>f%>)91&Cm3c-J^q*Grl&`psZ;KhE<GhC)o$G4WOgT(
zT~$pRYaj48$$r7ff5DutUsv=%|NJdeR(^Hr+VZ9Q2Y;7bXY<ZU8MgV$1Bu4ONs1@`
z=d3zy=k-m>Z3iDu{^$0Zey+>xFO^Grv!=o44%g3{#S~^X<<&@u;(-w-0^#oDmfFEj
z|IwFSq+WYL_wFOPrM5r(YvjPf;@+jTug`7X%QtTx?nt#(_#W>{%sRN<X`B7(yvXZ`
zTiJl|&AuO8D8AZqrfxlR+BkekkooY>V^<a{4=!3+l)LIFZ*h5b|I(f9kAEXRJAcH9
zHySptE%yxnXZ`BX%pa~SbG`SEuUq%tJNN2CYu;aR`t$t46`CF9-fs0#Q+vV&_m*9o
zK3-RFxT#|QzO*glEQ=pHIK9`*+DR`RA5izr;lWietbg&g+CZlr+3}rs7`49l?hlqF
z`3c+Ea3sI}?bO7Qzdil$TFaH?tMixg1D{HMBh#65;Betj-yXkNU2}BGj>U7fZpfKC
z`_1>JY<=5Nm_A*zD=)1<`a^QvEWU#OV(qpJbIpc58yxRUoX<0z=Lh6Wcrz#KzHQs;
z4g{AogD;P-S@auo@5T#%dG^GW^v;BrO$$cs$u?IVS$}OsfIcBf=0E*DWp{Ha`+2Y1
zPdwLW+13G)S8;jQN#V)IX75k0=vvg2_V4P?zy5yPoVmUHe@bfpeOt>PS9a&-j7+`A
z53PT<_@&uPBd-|_K6Q<&yjG!%{A?Z4YvV)yYtLx)MHjbjdTm1bdF5GQ<FQ_gw>^KX
zAu@|Sc&@4IU!C0X3s=uydZ+r#*s6x>;ZrBok-8TXG8B13>I)B5=3jj+kUu~5mGJ3L
zujO>9Mh)q6JxjW0*_ks6b2I<4YQoQZe9Z%%TQz*x`71Bn?Ap=vQPRum;XmJ!jLT-#
z4=GRWx9Z^9=X%vk|J`*$I(Pd<pUm?+uAZrQFTdmcPnxSb+X|gu?FxM&{PM#2y*DOr
z={h}a=c;M?Jq?fXXa9Qg$$l%=yb}EMWYKfko9k|bkLCV8?u&#_cK<~y^YmSb*)NJ8
zD5|RGJu-<X6v;QYoLS|wl%&u}QxxTyiJ56VL+rfK^OgP0nFjO7@c6!p!r1ii^x@-r
z26@`rcBJ4)*R2C7v?Mm2o|vA}Gsx52Z`ZE8{bTh9DO8G04@w-=r)Q9-KMfVUF!0-|
z`V?9mo6bngNa-2m>F|Qr4&5`t-IPM7#-?)<bG<!-JbilhnvIu!C{?7;iLvP+i9=F)
z26_4?<Hy^dRjo-%RZNae7xpf6^bGR!&A!ztVehDgDYP^;ot2nn>lx(fTm7ZorxzUh
zF@^eL(`kumNj-x+efyaSm1WltAM2|qicOD7A7$wo<Y`6aqmNY`7#csi#mPNK_2tig
z+p_aq-sseFm86;_KZRHdsOCcos+Pr*T2ez(9Z*A49aKY;kr-$)0S1~(kb$N~(m+!a
z&_GiY)Id`!X`!hNXrZYMYN4r<bkNiVbkNiVb<osHdT8ncdT8o{dT1IX12hc*12hdm
z12m125t_z;5t_!J5t^*TLX!=!&}4%wG)<BTnx=pWnx>!$nr6uiO>@8uO>@u;O^alK
zrX^s3rX^^BCMR*T<X1pyITzsOLke=s;>jl2plJ)(plJ)*plO%v(6k5a(6k5b(Bvf^
zntXtVCLiRX>5v@IbOapGbOasHbV^QWIs;B<I)hGVx+E7gT>%#~T|pN#1xbLW5D=g#
z1O;fiB{wwP0XH<=K{qr-Nra{t5TPjsMQC~?4>Uai4>UbN4>Y}!7n<IH7n<In7n(lF
z2TfnV2Tfnl2hGXs<Oz4ra^+LlDUhbJQ|~7E9DYx7&$$~;LLt>!0%w|#8Y2^8Fls^?
zjM|VEqb{Vws1NBe8bSt)#*h&s8)7k<LMDvnkQt*TWQj-l@<Wq*E<5qx%+LDA%ZFzC
zeCGL*s;`nrYI%k(!z5pH<ZII=lH7Aqio>cFqB;u4Mb$=BXEApY)z6q4sj7*_kE1F=
zRsX;orK(eq)h$GgesvpBqhH-g)aX|?QZ@S35voSNI!e{(XIcn@ex{8u=w~_!gMOxw
zGU#U_ltDier40HtEkuKUO&igmU(-o6=+`t-4f-_^szJXdN;T-$wh%4)wQWR;er+ew
zqF>ucwdmJIs22U&DAl50*FtpY*R>HH`gNT|hkjin)uCS(p*r;IqEv@|eGAc}U*AUb
z=+}1=J^J;HRF8grgzC|+k5WDQ4K2igenT5Epx@9*4CpsBQUm%85o$ocAxaJCH?|NX
z`i*VGh<;-yF{0ntNR8+>MyL_}#wazSpKT#5`q?(ZqMz*~Ec)3-%A%i*P!|1cl(OhI
zwGb2fO>M-4ep4qgq2JU<P3Sj8s0sb1C^ezq+(OLgH@6Wp`puohjDB+?HKX4gp=R`(
zqtuLkOAE1}-_k}b=(ltd3;HdM)PjCXgj&#ViBb#txhUaIk@))uu7hw_Fn1H~7UnSJ
zj!@{gQtl+?cFJ9ZY>N^b`fVM=hJIT&v7z4<rZ)83TB!~FwsvYmzdcIq=(l$eJNoV2
z#EyP@nA*{AZ>4th+uNxf{d|=0=;u2KkAA+J@aX5mlt(|`N_q72?UYBqBT5|TcXSX3
z`W@ZGfqqArI?(TEr4IBv+NlHm&M0xB-`PQ&=y!G#C;FXX>O{Y@l{(SyY^P52yQ0K}
zepd%^q2JX_T<CX&sSEwCR_a2(tDU;gFGPueexZX1=oh+)fPNuN1@sH8R6xJbP6hP4
zqr{DVcL#B!-`!2z=y!*y8~yH9>PEl2ox0I4Mu~`iv4e={7rTjwelbi%^oy-jM8DWh
zMf7{3#DjiM2l1fa(@i|+_k^hj{hn6pLBFS+deHBU5-<9_9mI=%Z#VIx-y5c0^m|*W
z7yaIL>P5dVN_^<|br2u=eci-|eqWgS(C=%dKJ@$AsW0xg{l5Cdx^e3k#oq1}jV?0E
z?*i>Fb||`TEc}ipluyr^UM4@BP(EYejLC8;7+<>}UAty&jLpo+bjWKX6CWOMfBU-p
z-glHqMg8RVijSAKPk+nPkQuvySE|Ov-vFwWYDkQdfuvDtAZe9aNIInsl3uBYWKbF)
z8I?vztdfOfQko!{m1am5r3DhF<i^EM>o%ngl3i(s#4CA74y6N<Q|W}{Qo0}sN&%8v
z>4qdKMMxf{2a;Fmh2&HE?xyEIT)gUuM+=6={uV=YLrn4_P~GNi+tB+jU8kw#BgTwy
z%eOqK<)g=pcFU_Z?_OS2J@s6}E3s0Rm{kz}gXI?IPL>bE&LvH~i{@<=Vmlag&!BYq
zHbK5^h}~-3S*VKNEfllGcW*U5oA}_xnWMf9$96C+VcHn^A0@TCETQafYCgU3k)IL|
zNwFP__KYr*pPan^*beS2b;Wmp|DfNdG^<K#YD;5mx-Q)$-~Y%*A;#WW2aB^MY{}gx
z_k*GLjXwP{R>#<)!9}CwCnx`m-3HxRB$HYB;KY=77RTg=Q~%Gar#nrVe?*&cKBgSu
z8Bru3b$RiuBDrVLEY7N|s$5wdd(m(z9+If?LGr66#*^Br22RbXfm4fW;M9IKaEz4!
zj^P;K7?A;v@iV|_tQz1noCY|Jr~yvn*8r!rYJt;oTHv&z7C5b63!KiX15U^3fYXUO
z;B<Z+aC)mAI6bEaPA}?#)BE+n8LS5244eTtgJ=NG;5Ptgv>Jgkaz@~cq7gWw-v}IQ
zWr1Tk7C2U9fn)tFa3-q>I1^_A&Lo<EGx<%xnXP8v%$ylGvuFm+>^B2vv08w$a2DV!
zq6IjM-vS(G<#OfMZ)~lO<9J9S=Y!<uCdQM^Y6H&3*?_Z&HsEZ28*p~39XLB@2hJ|q
zfwTMV!0}cdIG*Eyi+#8U9Pj6Wb66d~IXDM!4$%Re!|wpjX>|hU<eb1cMJI4hzY{o@
z)dif3a{=cPUBJ2gF5m>K0Gz-HzzL!NoZuILb6efOxj8p*ZqW^#+wTTWw2HuqoCusK
ziol6}5jc<41DuER0Ot`sz<K;0;Jj8Za9+*}oLBS$=k<Gm^I3hs`8XeNKG6r9&+h|n
zB0JF{UmsG-C$W<tNvw1?$=9&h3gey6LgFWUF2rquJ4su}hS46fW8_0TMn}kj(HU}L
zbcI|Pg^+;J9dct7Ln1~`$b-=v@?!LbeDN6jB&TB5osV^@Dg#v|v7f!lU*(tW*tcqN
zd*ZBVWz%QORFvJB{?FGmc|)O?5#O8{-<%%b?6LkQ2IkGKcs#c0d#s>WLhr;r%A}Ok
zzWw{%<4JQl?n_Jd@B=ai4jM8vCpRyD!h@VmH<)=K%Q|k<|MzJ6u;C*{-#cb(_Wdfg
cMsF~hEaOL-GqtR9aJt+32o)Z-i{zpI1zpngssI20

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp.31CAAB24531BB9D5.idx b/.cache/clangd/index/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp.31CAAB24531BB9D5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a046bf1756e594a1676937928648053bb9b98262
GIT binary patch
literal 8192
zcmZvhdsGuw;>UBZNTYx#3cf0eue@I~KCtFF$xJfU-~${KMZ~CWeWu{EHd;jM(Yn6v
z){4(o-5z_kTCHuVu3BF~l<K;xA}Cr{Ep3r)?XJZ(JsbF)A>TiKoZ~(7;ePT7GZ|-+
zo3WxG{MuWga8DB#&zU~IFixRRbc+5fo?rA@`zTybP$(9cos{lm*JO5l<|rQ3i*FZv
z@k(Oy#*L+Wod>Soe=>Q{+GEEDZyK<>|KksqhjRbubLLob(xROew{GV*JelX%xpL#m
zjyn$AE)u`!+_dD~x%LiAxC2`zE?ltu^``o@`BlO}+l6Y~8$%lN-@CG5!yD_aZ%Kdf
z*_s`{9Snc5dvgA=rq0O)>7U&C{ps0t`IqNadBanFd_JS<(Qf0e8OggFZ=LT|^f%QP
zyGHgnlXPVc8T4#Y(I?WeW3z(mdj9qE$xF1e$K-v}wa2ZLGUD&L?eCkl{pJt2uZWx-
z{8?V+ljFk^k}tMf@+j{?=XqVH&|Rl}+fuG{-4p7+^QiG}x%ZA|<PLo2+`vrz-Qt<~
zXFk}>XgemqbpGg(X(d~kl6yAu<O$agpI@Ea`x{@Jr29|ihjA7D(fd*-lp3FHHEiy)
zrr=PYqP`gqx6QrcQ}<G)kJh_heX`hTt7|`JmDp$h(N#&NM^j3>m1-4(-wbY3-e}5N
zd985Js_(C#TfJ{V_oc&5knP*o1_M`)AKs?=)193jU7HW@b@<^(U|HK&D%K7^zduEt
z>MQ#5xFatb_Hau!tWkYL4dUg`zRlY^b=3Nd_@2iH9DB2LQ{l^f`+R!w`1Fx&#%w?G
z-qp8ri!_T~X%sJQd(y3g^^4TJu_-Ri%KI<5{!(@(bAaZsRcVms51RR*{qqh#jLI72
zJKTT$%q#q<c|ZU6<UVP}Yd_xJIC#-O;pXWl({>$nrj2`lMZay2yU$;7so&LGRkh)P
z5Auf;+)<9++GF^E!oQZ7DzxWbTO-bVuy&aH(%Dm~-AWA8#?9HEM0VWx=g&Q?=_|6g
zKbW5G&xtSnsP3ybD&AjsnoqawU6Z-$o1}d!r#v|LbcYx@@%1bEabL05UM?H_=~YQT
z<{x`i3C4GC?NnwR;#NH^<Ysf1Hh$1`(3)*Wx4A#cc~U0bJwG7$2KTx!Qt+xu?<;b(
zHyW)s8aAw)z4`2aCwxBfMCU<OCl*|O_v+=0><cffZ};JZ3D?Fy4^*Vhs_38m`@u-B
zAJ@D*UPC^<IHTuB+c!1k9e#MTLHNhx0eS6TF+Gakr|-Nr$+t7LZBr;{?e_Oo%%I%;
z9V`B|-c<PZ({Gpm>%sYsH}3Duj32Un&!dNVSz~71&bpkQP!bpX;L3NMejD@kt5pxp
zUq&7`o*GyY`Q`bSOYROVF8ldlZMC*?%f+~hyL+v#R81Y0{_Kz?eQ4Un2T$%UF8h6J
z=fo9Foph^1iyBMbpSt(L*7RTR?4P>klxuy`Ps`kqb^n}Gb52Oj`gPf%_fBT|*O@P@
zzE*hol)|sO{^qPh|6cyQP15VRi{>4d(ii65eemV8XR5y6jd$Gbx3u^t>3PAPriZgW
zK09&V(Xtuby_?lDlS|gE51y(Xy<9c<(_d<?4fsQSDQ=mn<H8jqOil5jjFCCTOXf@+
zLKKSj&-dId9WgjNf#&2Y3cAI2>(nwtuNf^rL+Zl*a8u+=0v#Kj?h@Z6p=FS#9l!n5
z@RvtRK1rbCqSGnyDKE4P^7PC@Zynt+;`}Fx3Mo46Z|nE84D$5JW&XE2-z<JVfsTnz
zr^cryv<&j}>E1uBeCg-6&Lz<C(dnV_Lr1g>@^r<or-Pc^LwY1BMn$K4r1em@4D$4?
zEt%ugOH-T)bV77GBR<2`GRV_+rmfrY{lh#(0?m$2r^Tlwv<&ieQ<n?jGiNQElN6()
z)4kGq4Q?6a>7wZ~icfS`DB2V)9X{;rBS(+K0+nBtBR^+if!ePI()cw%TE7-Z=hp%0
z{dypS-vDIv8-YxI6Oh?&2D11qK#ZROV*M<T&2P(*k8@&y-ERkS_#Hq_zZ1yicL8yJ
z4#@3y19|)&Al}ad34Q^{>-Pfr{5~MjFSe3=%8qT{a;pC6>)-Z#e1CYK=u04}3S%#U
zDxjJR6jUvZky=uNsSc>YR0q{yY9tMqnt%pOO;7`-R?>p04QRpC2DM=7BpsN#fDTMu
zPzR=7(u1iF=)u$n^<Ww#1DJ+@0Zc>C0H#qgf@usG!88VqV45Ton5KXUOjFPVrdcwB
zX%3jdGzZOKS|kgYmVgCJOV9!)BQany0R~Ja$biX8ESPM71(OZ3VA>>GVeE9b1#EMH
zg0`hGvP*U_?EyQO_Mjb1hvWd$5paO%2s*%YN=`7H0VkNwpc71&<O0(baDnLxy1?Wl
z4ooh<fyo6qFx`?HOn1NyraR~c(<6Dn^aMO$dV(G>d5H&;5Aa~}K^{y&62KG!0+>Qj
z0MjdZ!Sn{aV0wdIFny8_OkcnUrZ4CNQ<Ow7#efK=7!<)A$BeVJ-nI%PMgq!Za$8Bh
zG1B%eH^kUQXcRMQs2mtNSGO{|_-Z?nSdeN=W##*weAnDb+PB;;V-HqU5!E*&cAZky
z5!F59Mxy#RayeCfMZvG6s>{eVR8<SCt|DsetLumw`|3ub#=g3os<E%Gq-yM|Yp5Fg
znku5fzNU_7u&-$(8tiMzsRsL+N~*!WriN;;udO0l>}%_Y7W>*pqQ$<poNBSJt)yD)
zYip<$`?@Nk!@jPL=&-MABs%Qt%Bc?fx=O0UzOIJqu&=KodhF}#h#vd;Mxw{QzMSf@
zudk$f?CWc&9{YwWV!*zkju^0SXe0*g8_KBx`-Vztz`mh|8nADyB1Y^R>xdEi#ztbq
zzOkGdv2U!TM(i7Fs1f_7Dq_OEsg9ViZ)zka?3>D|3Hzo>YQny$hMKT%t|Dgao9l=f
z`{qVs#=g0nnz3)Lq-N}!Yp5CfmMUVwzNL;>uy1K37VKNfsRjF%N@~HrrG{Ft&r}fx
z`%E2Su+KCS2K!7oWw6gwQU?1>4P~&;RuLBaY#m{-&o&Yk`)oO7vCmdg7W-@sWwCFo
zCAK>xc9*p^5Ze>v2(kTv9HzDk3Vt=U-A1mbw(o)MwZx8ndjqj!-yR`$?AycCj(vMI
zwPW92Pwm)u)Dj2w9Sy{ReMf{iu<r;{2lgG+)Pa3RJ#}E;SxcPQcQz0w_MH*p#J)33
zo!EC)Qz!PF_0)-dS1oa2-_<}|*mp&U3;V7xbz$FCO<mY`)l(Ptxmv<upKBl-_PGe*
zu+N1lhkdS^a@gnUDTjS`EpcPt-9X&fcSndD`|dDxW8Ym(-Pm{6Q#bZKwZwycPXqB_
z-xDDo?0dr0gMCjm^<dvqPd(V@YYC5izJc)A=Ocv2J|Ct$_W5ecW1p|5JobfJB4A%=
zAOiM<2obO^gsFgip_&TV7wV~ieQz!CV&B_9yx8|fh!^|bF!f^JTTQ*#_tsM{_I<U)
zhkaiI@nPQ=AwKN;!qkU-Up4h%-&aq4*cWSwh<&kvh}ah+M8v)rrXu#mYARx1tfykE
z@BHJP8=KUd7DV6fW%bFj$nOGOtAA89{WAYHjVs9Omp57dbzH%u&XdN=<R7|uUYcRu
z#^__aWp;DRpK)`tQdHk<vbJqbw#!PE?dg8k$&w=nrjChz=w(&I;0>Tv#R2iE5iwG;
zYB1GSHJEB%4W@?GfT^)+z|`;>Ftw}}Os!Q5rk2-&sbh6u>a032b-WHtJ*x*(Z`FgT
z=k;J3SOb^_s{u>{ZvfNC8o@MLjbIviBbX-E1g6Pq0@K8sz%;XFFwIsom}cG#riHbD
zX|Y<swD1-%8I}Q)u`*yXJOd`nvS6}S7EG3B!L+fqVQ>qz+BhKIHX=rL)()oKY6sKK
z+rf0O4lo^72bd1t0j866g6Xt6!F2LYFkP$*OqbOKri*uh$*~-moRtHU<2f+htQ$<X
z)eWYbcZ2C+Jz#pQ9xy$;2TY#j!Q`zxm^{yeDX;>Vf>i)h-~}+ftQSnL)eEMV_k!tT
zePH^mJ}`Z}4@{92!4$0`m?AHNIhGllDPLL=3$mGPpd2Qrm1HOSxwUyQSN_~u^5t9K
zP1-M(HLrpyRcg56DAhn3r3Of=)B@?0Iv~AL4`fgpfQ(8bkV$C*GAqqM7NrG<Q8GZR
zk_ECUZEE;mptJ)ylnx-L(h1~Jx_~$(2jo_|fjmkN5U=Ec1f>AvReFJZN*|D@6kBQK
zv4y3tP4{<?-Wk%2X{@}Sy=B(`XLr^2PibO7&%Qmq@*;U+L3-bGue_>#=!eoJ6YiI-
zk4hP0Mpo<v8=5_IynG-&Ui?+t(R20*(I-ghkdh`ZTF8qS(S?uZ6?^#{NGur5jBZ`{
zkQ-5{dEGuH;Vtj$9D991@0%B+f5n8j34P_0E3sf=+{9Koe`m+k$MNT+=wFd8rsvCh
zAn#X)_AQJ4u_sW4R2yJ{L>*EiYC;-BZAgo#3+WK`Aw8lYWI!~AjEJU?3DF!fBU(Zh
zL?*-_vLQBxHxtflve#_Ri|%k2V;5FlXOdr|=n-xHE+e+X_?ALG38GT-HySZ%3^S(n
zwRvc6`^yW-UEhg5K~`#3Z~3sxM>cvyo4@XeJ=PYomBB|cd&rLH2ssd)At$0M<U-^^
z9HKkqM)ZU{h<u1g6hZ=`H{?b1g?xx&NQ~j{os`zFq3gHN9quXi94((ZMbl>&Hh-T}
zTofqIZr+&GJpA8Jp1h$@<i|EA$2KRnZni1d6j*v}f%V9r<EP|JnwURXF{O1*5tuV=
z_N!4GF~i>`u5J7a%618fNgb0rh@E)%z)tN&E~RVdF5S9k4jXQ@*&R+}s_vzJL)5+h
rzu%_ye6d%bjK2K`s5Dx=$!uW<4^1D`!@zh_)4VSiQX$LbBP0I@9hHZr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv_bwd_data.hpp.81B90E235167286C.idx b/.cache/clangd/index/device_conv_bwd_data.hpp.81B90E235167286C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..908388d1266dc900d7c214b317d6565545662fb1
GIT binary patch
literal 1020
zcmWIYbaVT{%)sEB;#rZKT9U}Zz`!5~#Kk2=nQ4p+4EKO?6?66m#QGmH5V`ArlX+jG
z=$|}}&JKPREwx!cc1(G9YF6o9-`l*LANNNuJD??S&u5y6_47N$?`#A2Z+#OqFIX#t
zYkp_<{L@J)UEycvBz#&eyM3DND={e%Z3Vg2?%NJry87d%SUvki1-)%YmXz>5+Vs*S
zsc-Mk`?Gd(?O}RlRlvPs_OkoUElg8pb8`#x9(&)vD(o)XJ@(YQn-wzF=Lvk}FT2`T
zo|WY=yG`oGvN;JQwTTvc*mJiCO1!ykV`9C?qTsKBl%uM}oyDu#K9@|DwcRN5?#!B)
z&iDWg_GQP8tvV!gMk;FRGlxhM&UJ4!e6y{+`Y%|otb63_v7luF54Xb1RL#wcl%$>~
z2rZit6OqVy+-Y%YFlXm|*G3hm!pLO~6@LGoO`NHB(*Mc($v!h%FZ;ay@A>%OffOtM
z$yeq}99dL$Bqilo<#9zX?sb>n3O2v+$UJNIY>D|7`HwNIcNTW;W0aK6Jy2Ykn-mWW
zb%%f5*CuLAEM;S2W@F$KXBU?M6AYY8z^Gtg<Yb(&+1r`<`1L;mOf2j`5g_1{Vv!PO
z;DB)M-O??M|IDT?z{CoZ*4EUvhDkFNrKT0H2O0&lhmi|xzNboA_rzV!bs>I2>MthR
zu(I=TD#$7*!z|<BRFqXj;wi~0A@P)Dm63QVvMNYCRasReo|>#0oEN*@GABk&MVXO@
zQ&L7!17=lfAFFDEP_YwGSX@b559ZEpFXooF{r@evcsONcWZ{PJaLUQZA@Ss8<dJv^
z8VdR_bAevaP(<P>X(%D_lr@x*cq$qyNIX>yRV1F8h8mm)Gn<hU90xGBt~Is_I660B
zI?xN+GTQ2}D1p&H)u1SWseWp}`h}feyqgIa4uX=>Ff)OA1+}GNdVukklc)rYVwf(5
v#+?l>VwQT&frS@brno4%SQ)62i8DVS?d`71{<=(D9Ps!DN^n6^3Wx*%-RXJ8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv_fwd.hpp.257583CFB6A8DC72.idx b/.cache/clangd/index/device_conv_fwd.hpp.257583CFB6A8DC72.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dbd43ba67a326e5d4aeaa6685271fbc78c79ee02
GIT binary patch
literal 998
zcmWIYbaT7M%)sEB;#rZKT9U}Zz`!5~#Kk2=nF)*x3>SfN6?66m#Cjh#5P9qW@^H=V
zJ$ZHQh5^eB0@O8rL~Qz!$D3R1JgKbjVg2^idD$h&`A!N?=1eYrXY06s>l>qqMrxrw
z@)Nn|ch_iq4nI5h$<1`@9Np?Ox+Yrlk}ZPzYaClwf4r>wuWys%oZQaBhf+r>FK3>R
ztzSLQwD@6V<&~w+p9O~d?*GkG9FnJ}+h;l}t1f8kt6QJ%X}_LV;b>jjykP(K)nz}H
zFJs%1xuP-m)W?16%)i*bcaYAlVB4+s{Vboxw;X$pwg|7r@+GI*D`jjxisYYJ6Tunp
zq1C?ZnA@r)B4?$dM4tsj8VRmlE76^;eT?ryT+-5cr+6e8RHi7aOgnz(l#t2Ui!L!A
za}t-bI_^|EtuE<yRffxWf>(-bk3!Ioxva*q2g>i97xmq7^s&q9|DM8Ei!PSPRd+6t
zoxzgkR3a0YcCmDqt$_<4SN1ke_r14YIUJp;Av5y<_naRY_lqlYlj4ElEtD9oa`)4$
z)dEbc>_9FMaB6F6TQhJ#I2Tw%HZ&ZlZ)am-W@F$KXBQU%6AYY8z&K)H<Yc^j^FgF<
zB;!m0CKi}}DHbVln0|($)U@LDK(#P?7&*b_!^{ow6H<RM(T0_shf_gTK^bNm52vE6
zA`(wYRtbrxEUS#fQ;}6c;;G82BJtE@)!@9??Up$)YAVW%Je-m;k{U3pQu|m{8-$9T
zfWqQ(;#x3=b$c<lyzT#Q!NtQVD<caxgojg3Mh=N5FC&k{Q_xV*hnWlXf`%dzPf0@w
ziKncgjKovXP(k9UYN#Ud)HKxKytT$w0Y~R1Ob5DETSi+Q7N#&aGjf3=$5W-Od*UwV
zI+#m<K@N%%n5w4+tY6sq#k-k+;UFj}4O0a)QcznOW+X7)auPxL5f(fQFzpP*Majj=
cpwz)>D?fGGld@~4nYcLM@eLH`0!cyu0EDJvivR!s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv_fwd_bias_activation.hpp.D4F53CD24AD4278A.idx b/.cache/clangd/index/device_conv_fwd_bias_activation.hpp.D4F53CD24AD4278A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21b3eb1024d0180643cdb18e7450965f1a62b34d
GIT binary patch
literal 1058
zcmWIYbaRtpVPJ4h@vO*AElC7283cj2xTGl4h>?Mzotc55V$R-4{@F~10!P0uuGnDy
zcEMg&!7cR%oG(WAE2bU4yWysfM_Rey%kuZDw%(q$s{N8vLc+b%f9@SnnfxzLrQ-F)
zl|t3r-!nG49PjPi@kK&Jx@diKT)p|b82yZMx87OZ-t<U5_i{=RtHq1h<l3S+2Yn{9
zujxHjUGz10^10qF<Fm``I;L+CZGRFeYQY`3b^b1;#}fT}TUW0<P>}V~iNj4?)5EH;
z)q{U?>kg&(jtl?pN^~<Z6ln>5S`j4b%R5hsC&)5&`$d-vAyr*jxq;hO%0wCX-PUrw
zz*?j3{7l>8(EaY1)g9Y+e*f=$`R%Hr%S(#x>S<p+zO8;HL;v)`RYmnmbDwh0v~fP6
zc)99?i<uI?doqU$PbiP<%QTyO`7N^+o?s09*)3?AwQRy3#`)$!g2k1&Ne#ddw4R-!
z-emA}IyVywI|G+4o309&VBll|1}g(27h^&19OnzxVti~&%s>S~>_SRl0-<0Yt0Pkp
z%i`+-OspJ0EkM8}$|5Sx0OkCAp6+(fEM%7e6B|rgT~pl}CS8=8R{RKP70exsTwwb>
zRm!?2?sBdR@e@*iG0}#Vorg<CRz?|S7Y~=LtSk~wPF4<yCoe0H#8Z$}K;kLNDkAZe
zWR>8&*zJ}%F={Hxj67T-G9nr<t5W+|RU3qgoq)nZIzkq(;DFIgT;Sk<xtE7aPee~2
z=9z9U=9ahp|1G$9xFlsH;WqJbNy$hd@uX#>k$5s1GWsyff&SExMdHb6$RY9MHRO?a
z3K|MXJVgyfB%YFn5}XG!n~@V7PcV<HHMR;kIyYfD(4Xou>gup)duqV?g`Ho#n;94?
zf+Er|%|Jzh>e4VpKwop}3B!C169rlTifNbyMX5Q7pgao;PzIO;Lvc}Zu{Aj1yf=SR
e`uer&QWjnwPHqNHkQfAT!x#(<ya*-_f(ZaARYdgw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv_fwd_bias_activation_add.hpp.D947A059FE31F7BB.idx b/.cache/clangd/index/device_conv_fwd_bias_activation_add.hpp.D947A059FE31F7BB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a7f973a12c294027e4af0b7259908f96cac4f0ac
GIT binary patch
literal 1038
zcmWIYbaP{4VPJ4h@vO*AElC7283cj2xTGjklaYZzikX3-V$R-v|7<2hfur9SS8OnU
zyI?P?;FkIW&KINm71b6?o6Y7t`9@Fa@4LHMYjan=TNuE>bEbOc{py=b_RDkLWBsz~
z$h|}MRyr4plZEw|-HuNC^LoMc`~BzFiEk0lEl<C_=}~;{<(x~5GGBH@#LI{qTz<ot
zz53(4n!i_O><rCZcJms0VVY0-o<}omS2Lv-M(@2JH}PxF@w<nvha}jnEm34y>@T7^
z_l47p!<C9wPhM-@dN2JtfWg5})oI?8z?0VmRvgrr8lE>jhsif}!IEoysnTZ_Gr!)t
znx{$cfx67Lg-y!;lTU2*s+isW!QSNVwyCp^ohZE>zUus;Aj!|VtgLM_4u`!~6yDil
zBh#xWWEi43Z>9O;%HIpF>ZtauIAuR)MWFG~H@xpU)uS1TD|3??fFb9TxA<Q-hu?7l
zCN>Tr7YMl2HPx*dpq!k@<2EmS&#165F|#vp39<{Sf(ZsrCSYhYFmf^Wo|xL&%=A%M
zfQc2RUxY<O9H#%b*x!fSr*a+OW?}*A*JaaH2NMYWMX71UkAP;u+`-5RwjXA1h@X)9
zi-|U@>^xjDvNFmr+jzKSWo40ga<XzrJb77pB%Xq-0uoPARuPG(B&!7H#csFEiBVHg
zX5`@#mJ!x~S(Vzys@fn_>;x1R)EBgdISfWKae)KGQ>CnX;x6Yp9xgo*JtLT}yS<oO
z-uC~u;Nszul#ztn#KR>eBZb71mXSu{$!N&v!z>5-Q$rSsC#NBY#FN*MN8%}HC?N3^
zH58F}N*YRV9?WcDNP}Vu=8?6=Rsl!nCQJwVQ(Z<~9Tp!?4OqXh^NV*g14BhnSQ@4o
zs7O#<8m0*7Yfe33n6F_pBNsTXVL?-rnv)31tFQoNfZ5DYT$Eg_4@!ERAOZroe8uMF
V^SIsE$jrmd$-n~>W?<k(FaeJ_WC{QP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_conv_tensor_rearrange.hpp.90268624B8539288.idx b/.cache/clangd/index/device_conv_tensor_rearrange.hpp.90268624B8539288.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02e20a697ce7129c9a69712aed96e148b25db11f
GIT binary patch
literal 1280
zcmWIYbaVT`!oc91;#rZKT9OE4G6({3aY<38Ark|G9xDSw#hkq%x!Jc31iq#J@A`eg
z(Rxlt+HG-Rrf8!}2QKxPUf3(T_=b4&k!zfu=fu4J|9&#v%<uKad}eRA?>67>%=Me}
z{P6i+u^p|NCvBfD6;(;>y)>as|HRg3Mn@OEKWrt;mbOFOkV*5(9ji9YsY<!LN6txa
z6*Xeja}=?0-SDLOtiUt*>E0Jsm`F{Tc3tqCxbD%TXHPrDS!_RIKQZ&qwk6$NVVtI$
zp7*+^u5s0Cv%6zbnZRUeACT$$;&6W@!*#|pE)3y$cXk=>ZT^^WN#TQ3+LL9gvvw_I
z*tgrhrucHW`Lc4q2Q$C_(fDww`d`q2+_j-GpPAOBFWM2fu4vxj(#V_u=G=o@rn!jZ
z?Z3Nj#>0zC_1*@@^5`rnD&G7eNb7S?^gcU{9M7vlt4iCqYvoSAp;JBcs$@ph|0}<y
zHih1OyDIzI#lvb_O-*w-Zn3RfX=ZJhUsF~u$u;xA`x9l}^&Ai0)QN;CO%lJMyRq$F
z$%4F0(;weB=3NV3{Jt+EpO;@a|IXw<hT4-QYwy}veY-4T<&giS^ZK3Hau=dM$?m(J
z`LeI!ua8sIm94oC^B(hkU95Spb>Z4r<*FQa3AL>`691V0WtTkPDQel)v$2OOW#&Up
z^Q7}4f1P9tPkwF|Y&#mFVN^3Y`AN~F8ODl__xMaWYW>Yd&pY(wtP5}TZxu8j_F4H~
zcx{!-qG`+L&)bk!5j5>(&(cDh4K6z!t#`;6`ldu?elfhW`Xc*x!}d@u_V@g5Z!V}Y
z6j$aZ#RF52(2Jxuwn{%Iu`#i*GjJ)fD=B~p22N%sMg|5(E=Cpa8DH)%n|o1!iIoGW
z9tgNBSuDjFpq!%>A6LDTIxHo?#0Hc0(Dbl|Ni!6srWHR2ngp|lkqd0Tr%GA(#9hvH
zFw+@1!KT7g#csFEiBVHgX5`_rkg?E!sY>l*Rc#O|b^;13X(^e(R975+_@s*E;#p1}
zE<0sAxN06Qdu4kho`bRj63<cD5sBxd?1aQ~R(3|>xhT6J@m!T%k$7&(Zg5_=7jw(o
z{{I#{JY3dF)|xO!^KjWH*&y+3m28oCcA9nu2$lAl_DDPjO$Q{NqoyMg&q>n>iRY~8
zjKp)%bV1^|YPurv+%(<bytT$w0Y~R1Oa}(1hm40hEGl6%Fit>G2@B<?2CQG$`Ng}L
zfN>&dA&qdfhoFZvOb;-=a}t$+i3FyLA-=!m^R)wPZm_t6%M=$S7b^o*GI5DotZm?Q
Vth>#`&B?$CN?{Pd4P!7c006BA5NH4Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp.A70DE3BD61066477.idx b/.cache/clangd/index/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp.A70DE3BD61066477.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a086a3c36bec77231d9368d10b9611ddd59de742
GIT binary patch
literal 30982
zcmcJYcUToy+Q6AJ<<88#UMGl(40=U0Dnh)Va=in_28u~6F<Dz;60_N?$;K38+Ls1`
z6tRI35G;Uz3Wy!BVndCHy(?HykRqrQB^1Ada5yjf&p!Ld?&IFWFYkNGoHnPt=giaz
z<Hj}rXk#<>lL<f0`^SPgoQ;jmEBOE4E|~w>MhuQ!Y;1m9b71<suWouO#gos6ROa28
z7PWTzhCv^B`S#ELW}$ViX2*|yXUW7FwtVE`mED@Eyx!R1IxFhqFFx?^Yg@JM=eq?@
z5*8Nb)R%U8___JcH;2}3`#xp5cWghOZ*7~)hJW>2o$!}+j(b>4(TV9kE4^lX8rwYW
z+q2op+V#{gr_S3vDJ#_}IjSt(q3rJ``*bTGHofB2IkvX8-vnz}M?8<bmTu=TCpvI@
zsL40R;m^=>mt#*J`efLf_KO@In|fSv-}LEs#iz<=EplA6Fyqt4@dqtuGoIGEt?bcz
zdR;}gYo`t-uURnfQP`H+`K6z^-MHGjYU2;n!vB_fX5OF`fpzP=yS<ma&FzQF(W}mU
zDiuz!P1C(v6%`rrkz>jmL+V%O>JPm8<AvO-*?w79N1R+RVD#RPEFK?uWL3vjd>VOV
zfqSycNBd3=YwSL5_X=~9b?@zE>)y7tl^V|G&yQQ#EpAZq4_}-rSZ^#I6T~-v8y`IS
z@skN}&R?50=)}hhrp(@$Tv8qQR{3`a=I*#T{py=PSKKVWQI)yj+2!3;yRQwsy8Cus
zxpreq`OxV%d-f{nw>b%|`(ggb&vtI=|FoR_zuTj>^!#V%v}r+I$87N%^!}`fX>%{l
z-?zNFLK)oK;m4gtj)5LtXFk}mE4Sao8}~|odlKdF!~G8m2CbN0@Xv`e5)SoT@V6&}
z=T@0FlpX$J*S6(JYd;<G=K=qltM7Z(6tdU8aTu^=$I_#JR+u*gf4=*bYiE>RGqyzg
zrFANpwncmU`nnqfQ@g&`xq<04zG!l!Zu(nq^WmLB;_ueY+i-h{o5LyJ&qZe5-3=}$
z7wDJ1IlG5V`o@zA>Q&Te9k}kDAIGG`j@E>*jBCytPd`jKJa?`8i3Jn;PhPP1gE_;$
z=kL5adi@N$iTgiu$X^|_DC3Kht`mhJZa?iGHSUYyr)Tue3IC;Y{fPU|B7R@k_|E<H
zPA?qmM}0o%@%^-!v3<vU-tCW@wU@hpl<|>kVo*qr-wSM|i~Qe1PT6Jrx%hAMS9rTW
z>iOm3knwTy=vUqv;nDc&nD4$B^kcd{Bre^hO0E6yvnArER^Puzd^E$Y_i*D|xzlGm
z-dz>4Ids@v_m%efZjWyXC%iuTJm53lk;#;UwqJCewr%?MJ<dX4@A3;DJzn@%@|UK=
zhf`9=fBm<ihko?f*>Gs=L)N%BV#+n!K`|pXbsx9wQ}*h2(S1IyoEqUjp!wmeA9w0)
zyYlhOZyrpwcS|_w{fT{_W3CH+dh6)Wk6xP{kbJ?u^5%;2e_AGY??1I~Rn&y}Up(96
z5Yl&P=~F*uajY(Pv;V9|i-Ug2o$}+KW9Gj2#yN_8@LJ@qxsh*XefY+YAN-=fYa{3f
zmH+Eu*Qo=KzZ!S0UORiQ>yf7eq_pit@k^E!4KttZvZzEJ$=@CH$vY3nJawJg^!xE4
zU#xC&ziRy8(1E{yKfrd8>E|84C!LIQ*2jy&7gG1;Gw+7`?3-Wu+JIlv?sggL-gKpW
z&s6bI${Sz)+H^%dyj-st-?w@8O^xaOO<O{i=&FDJ>FXK4TypBwe{=AIz~hqiM!@~>
zWVg*T$6P#kc>MYGiAA@LR^8uM9dfkjtz)CPoT92TRcB(Ke6?~}^$#cNKks||i>>>(
z?%iMApP!d?FLmq-n}wF<Yc}8hH1Fg0L>n7j%7j^wo*6HUsFT@7+a@X7r0#8w4Hmcj
zPYMW&bBX^xTaSdnHd>xe&eN%FPP<C<FK;@|h&!ktJFJqZR}%F-+Z>|e@W+5Lcc(l_
z?25W!l^G5*9AEwraWCzU_xK+?ws%DiSp2@j`;ISvh}b3e^8G0i`;#r|g2g*^%1&Lc
zHixLZ_H2tbCHMLYJ>sy+G<llc%O4^(Z(RCd=-7>!Mr6bfx>ZuP;s<SYh>Az=*t$7R
zyASD+ghM#3lTYIZZFPu>S8W0_^9$!WMl=qGk}UN|min|gL}k=%-PHW8*ZUdKAgr=k
z>c3g?ZF7i<Pubk`D6#T0JrZ$5Q{|~$Uj7iV@A2lTca|tOx*`>eXFJSxeECDf6WD#t
zX&(*z&WHx!2-2k9X_8x;)2`w-wd&y$^WPNI8LR9UJM9-8+8m<tZa2AfdSTox1@(8g
z(GE+-!_q5l4pI4=Ao_Nzt9{fOO0?K18i&&A5S1C1?f-FYMdL0#dIcwSvOL-L<qr|h
zp4jt)c_ZB;71S4tn`Ev@?%w7Ql{w<mz{%GZ-_|2NR>{}N`MU0H4pI4s`r3%X{h23i
zQ6H>wlH*QtL)sjoGB;)Hdy`k^Tr;Ba_<>WT9x1q7TAg;4M>Asv?Q1;V71`oYK6Ci&
z)t5g+{LQkzKFK=&?jk*6vG`A&{HLy4n?qFQH!ltQ{>?c(Y*9a~@|UdqC97=?QTg$c
z6VVCFN8UFg51fuH$tg>+v^hlO=UH}L@7H~XRa~*k7Rj_la&2?kRc=4rbN`J0#|nBC
zN3>S&v{pZ)%^@lQd;A~1pFClWfRxv4G^@_q|K*R3gcu(i8%ARSj`tX|R`*?)5e>xA
zY?AtK!qu?VAz>^!;rhw<51xO}74^W;yyx(q<I5i+4q5x))@Xn8K>_u`;scWYfMna|
z5S7pqbJi9-3}F;x!YbEf>6+}=<`9+0$oig-GY>m;MfO-_mcuN^mp?=t<2U2;apB*+
z(^|IkxOqdgmLDQs^L$EL@y<&=T~T-Jd#1xo$Cp1uoS6RkS6~0SI=<C+CgL*D5UtfA
zDoNX&`2EN2F0|J3IN2B{zoNA|L}erU+m`-w|9Q(6^~RAT$_|P0ur`ONZ1Q=-(tKlu
z*4qE<)hT;%|I_Lal`Sbj!8yJSZ*@hEIFt_@K5%^bL&RBQ?-_!Ad342y+;DlMO3ta$
zh&G3)>}j;C={v^%oDq$}D!U}7UDEJ2ho~H#r`80$!^U<+PB@ae4s-jw{2}7pZl`qi
znccID$OVUxF7-~wT|=uwR8A(n@%wsd);T@W;dY-R8FQp=Z4OcS^XD#!MYV5zuSYzN
z<Zb!wE-!zGIPcL<&OiL0?As0La3s%l;&YwS<`9+hJIZXb7i>&y&8~Evl8&>h)gdYu
z-+yh&nX;t&0@C3~>U8=#U6(e8s9frMr*M>O(8snfYu6@O*(5t_tqxJS{_z)%AIzS3
z%7_d&E!!n!JMPt69inn;f~D~9muFPlB4->?fvgnZI@;<GmBOpss7=3ojI(VB4kc4^
z%9Mt+IYj05u;DX4pZU{wT~SXQ%7+dgI==iN;-cyEzkW71qS}c3a0nZvE*mAsHixKG
z{43pCH|n2GdL-iz-jUzw^74m>Yk!aa_N{pFsu7LGA!JKV*^+ykLsXva*%r}f@$yxA
z)D?$tPA8wkT~w<>R2tbXUk_Hl-fl!=aVR?_r=7U}X?2K7Q|8ho7Gtf{+N)QHohopz
z-s%vQ=970u<{sQX&4|X}Q1(bpdnAuGr(LD-Rl_GsM*A6&FIL$tIqk-EwbdaiHuL}S
z>9>zQx3NJH$UXh;jWPBHEuK;1nZcyd8?+mY-3?{{<yuCqh4L7r#vq{Ih}1?jm^6EX
z)`Zk1$T_TC4m*I#53_cMA(yghDGU6UvuZiyI8Ke@fP5{du7w=Mt5N)5@<K9b(YzWB
zIfhqbAXoGD)x0a!ujTD)AwT2opFz$L>@x(QzfG{;26=~IzXS3U!SqBJLc^~XOx2KU
z1ye0$c0*$K>YwVy7z|oA<B`n_qZuE;dPK0pXvRgd9+8kUd5=tf7>y{4_sD{r&3j}+
zjuSoN#9?FvVbG3=9>>IC6WYcl1}#_g$c0=dden)-NZmAO&qR-Bke`bl&nXN0hV}OL
z%{bs@(3Ub@OPOi(L82HXiWx+CEi-s6^9JQ)#v_@TNcjvi_zaX^VJugesZ@T4@x8-*
zM7fmlDuwYqWh_sjegk7|VBV$rLC7lzf%Gj!uPsFrsC*R~z6!lVc^&dxhu)^V5qWGx
zAibr?s}xP61@#muPthRC^=N24noM~i>#>lXNO=c4WCuHm@;-L#KK293Ct2kr3&Ja4
zy$V<mUoktj80tS|m8VcYnDYwefc|pMV>t)nkKx9}aC2yU8@MqWxY?AmIgf0pAIN(J
z^3!O=Tf!?#_(7Cc^4697yOa}ok3@bV<$Jv49zT_G6|Yo5|3QLXkN{frdckYG0OCs(
zlvDx4cTi9cLcS(=U4!~}1<$*}+cdsPL8*lL0Xii>2f|yU^H`&6PcNB{@dhnSG=_;?
z$;;NDrHl6IqKUSso1)`Q(V4cWl~V7OQh(Bd<hDs1dCihZ94EI;-N*pafCq(ko1}hZ
zylBv>B>eA9GCweA!7~2W&QjY{kCEFZcqAZwWq8!IXP@LU25q$8=r_rK?qm*){d|@C
zP066mFwN*krhx`+rfFtBGQ2ZrvrM!4k)fhNd(ZS<KQcNsXz!ce??>i`25q)!c0V#h
zGH4%|KIlhAUIy($(}(@Yo7$kw9XNL|889ggTh-;>E*o&_UiW+bEs{F>tJ}rKjaeQn
zrxi=@6ie@t$L6$ka`$zzJsCc7+6~$HhTNBoc*%3P$S<d{Bs_z*k#XP14597!DC2OH
zain~n@i@;6h3t8r8BV!~@hpPe%y>3K4r1Md*desNEMnakK~7*D5?Dtn-@rO-fV`Xa
z*v$^5^4+ZGZgx23Ox7b4$}?HdOeoJ|wLI3F>gTgsKI9s9Y7P4?mDjRUYau7`?nyj&
zV<q$M$&gcc_Y}x^yh9%ENZV6B?~o7qEbnj@@&n%E0Y8-LKj1wd@WUw=^B%=eUd($I
zLwO1BQ3B;9yk`lNhX}8S2%~6tVZ!TSki&)7!yz9MrXCVN|8!WGdKmH%Vd@dcc{<-b
z-FVVm4O*r+GE*E)oBacE>;uu4Hv6^G__fjm((Dbm8c!~k-X_(UG#lJnTh>%aH8mLn
ztYPBmySO#nmQ1%LSEBE<cmCW8fA3(1!vcrfQcoi2Gji9Sr^C<5|8+xgzVpu7tcgjz
z@!+qBQH$W<?mna5huqAl&5#q3ng|Dfd8{c9R`t`Y=``eg)|3zV3~M?AxrVjWuy4Yu
zX{m+0kyAIq!QN(0-3&QIa1IeHwCF<x=TOLDf^!(;#e(x<$l-!>IOHXQ^AgC11j`}e
zO&b1T!EzY#5y5f<@*<sSkq$Ja5S=Lma+uB(Mwu+3A-Qegj8p~#UgqfYr_M}PK-@lQ
z^y-OzJtA@1$Ci(+M2*w_VfhE3-SW`gvL~6NaoTQq#%_5Q<xJTvQyvU?Ql>nGa*phs
zBfn1hknDX3@^N{@ad{+>lhr+(`j(|Vk|=|AK__0&*^;RdS>?m;&rs1A3g4e;qJ0{C
zf8G!sZ@~9wq|`eSzCT;n2Fd%=y?9-Q`cdvzxM6(K^OIgAezIhUU#qPPg+yS$OMZ5n
zrG8{J(4ak*@W1xgs6h*n@xONV@VmcQ=rPk%;Iuj2=5!~)V0vRt`@0m!{q`qQsVqnC
zks~{hEaJ37a*sohkIOxdQyx0vn0)Q0S$4S5ACr`0Qa94JEnf}_dgt475*Cd08#|GN
zKYUE|n4Nz)b^f<UwhjrNT6i?A*4LmdXVm44Cw+ItFqRlN7GBS&>!Cb>Q4=6<X4K7)
z&oSyb$QKy(0_5wA-F4VO++yr*LB7YR_n`hGMtuaihEZ!EKV#Hqko}SBkHGVUAT<PX
zC{jZqFGuQf$Pq}5fE<m~XvnchjfEVK)Og4nkh%eKDpFG+mm{?tdD0H723cy*RLTuV
zZGiG-q&7niVbu^8_}|5<yC7$=Y8K>!tldF4J3h+V9ff>~HJxHL8vk8Zy$k(6WYvd|
zAF=8q$j@2Ja~7mOj8nrnkiNy7x)^c<r$#{jgH!*29LuS(kk@nSddSI~nhZIWQ&S=D
z<J5hS7xC&M-jjC7%XrH&IG0$(tE-@VHLtFQyopygLB7wc_aQ&z)rXLu^7c>R*r%Si
zuZJ8gI0p-_)AS?>c1dvVvPH1pLTCF1ZI@ub3-WQna$Ep&p__twQvlC*OHgk?E)vut
z$R&bW0{O9EdMs!(J<kNwGsuBDQ=ksW7wb%mA;;_N;&ttFFRj;;Q~wg?JTY?GM!jL9
z-bkhzoVH1C*aUgAzQ<<0O695g9;uMi^gYrbr|TWk^^R13i{5sN-i~sHUd_<=q`XzH
zZiT!}Z{DW&pz`hdA=~xCDDTjlcR=|ry<rzjZ??X3w*D2OPd1T|L0cpm7m3};Bu{SJ
zN+K^EZ{vt#ACBw$Q)m3PU<QUTZ_%!H8KW$N(~Ky_HHw)`<+~VV7nEl*N+#sn$Z#7j
zoZUr+yO4{Jp$KvrQpymRc05I{PvJ5{J!h!r+KbYlHE@Oo$W5G~333px1i`Akgm+!S
zPbO7~yfrz4mL)1#;%jvLTP)fYiyn0RyH2uSCwY?buiUoNMsjdebNH2r0TJKfSDV&J
zYkzH%4LCfSXxV)uZ#)CuGBTz}-N|kenP%eAbIUFqQ5&~^QRNpEpRUJ&o|a!ZEq5nx
z;PDsNc*V`~{k8;4LYRpm%ye2H%b20dnAa&UXO!iP1LYNr#|p-qas;D9KzSsiL_&^Y
zrbIEbsD3o#8V%oN(afZ1W(Jk7W?WZ8c`V}^3wb@GtcT&pGv;_kqyFO=pLk{x<wQnF
zg!&s8Wdr1mjIt5(CdRx8hL_4HsZgHAC~1(lFkQAVJ!yQqn4!C1{$(;lGa;X2lylJk
zMMk*@`4XdCf_#lpu0g)R47>pw$W3P8O~?g|QULV}8Kn^NJw~|)`2pkc0G4MlqZC7V
z38R!iE@zbTj%+T6`BlM~D<D@gN+r~<Vw5V#HH=aNxt1~4LjQG)xeoGkMtKhP>lvjU
zaucI8b!3kw_<R9K2|%DeFGR{h$iYYn?#Sj~1nO59Qo^8oF;W&oUXGOIkRy>233(+_
zRzi+ON;KpcWR8LHtw!e6kmHaN2ldw=Wewzbq{Kr`LP`?kWTYfR-i(yZ9of7Y#+Qba
zG$>C;N;>2W)FlIf`hOdlZzDL?Lgu@Wi;%epav2(0hF+&*vvQ=ABL~VA$fE**b(}{?
zc?9K^NU4PU6is=GW>Npu$h8`dEvwO_YBYn&YmsX$l-D8GI>-%3X@KE3B6A~5e<Si~
zL|~n|37MOqyb1X<LHP@$ynx~Pvx+|po;Qe9f*=R8=3o}&XDF+LLU|aggh38xyM(hn
z>GSPkm0eJt%_`ZD_pr(y$Ol;E0OTBYU=CcIIK&P-1o<#K@G#^f?7$<CkFm-z7+x-`
z<U&5h+MQy1(e$2WJ<qbED4%2P&O!MV*6s@A+pKaMhIgAa--h{hhc(}Ue2-P`LH#0D
zDT4fvRUSevWzD6~e;I2ogIvy<%OO{=<_gG_tWpWXt74TZ$knWCHJocyvy-Y>P(SNg
z*Lo;_!MeVH9Lg!79H>8`oH>*O^(CBB!l8T#r!0XS!6^}tS8>WJ$WfdU1v!>eVj;(I
z<~SJN8qT~1@><Tk7V<jIybf|ar^LhX5;!FRax$kRLr&wAG|1aIWjo}(oU*qgoA<)_
z_H)X9C_lg{2O#HgU2-_ke(E`MJzi33Gwo>^Fq#`6H*w}B$U*$jARf$p7V*j=-huiL
z<~@RWZ^|LO5(4F+yb=m|2|r~CKa1)w<z1J;x!6*E(o%i~l`rRAmqYmq-gO1!RlKqa
zh9AY7qj>OqQM^wSKZ%AH&6}g4Jev23hVmHR90TPsyiW|2$MQ-njBgFEtbv@wD@l+y
z^5%^&yv@9_8Ok$wB?Iy{zRNbgCr$5tUbzqDrMyxKxtv$ZAwS{mp76b>|7zZ?ng@%g
zHN0C5<XYaX7IGc$RtNbR@AeGxbKdPa<iB{kzhHcgyj>&YV8K3E=tI-9RM3_RK9rXU
z_RF9=O0bWDoFFI(0;q3^!oWm0x85KO+yFUQP?Dj3ilC%GP8aOb1&}`(f?Eb$4BjfZ
zZH2r|aN7oXyWqAR@(#gm2jrcC+fK;41-soazAV8m3v!O&ngi$dIl`nI0o4B!g6jz=
zKP$MNg?vL$ZouceA((H#wSWRaDS+}qK`De>Bq&9Y9}3Dt$Yp|32KljI_ZX)Ci7@mD
zTpM{R41Ef@S{PalxkeaT1G!chS_`>O7+MGUxnTDk#`l+C_ZQ?qon4>~lxMikE*$bo
zo!v^v>vYOG9jH(1bpzMa#d?F5pi>f{JW;14LQc|kNz%2qPht~_XR56Rp=1I~CdydQ
zY6eOKZ>-K)ekQP@6Kl{$^cm5g%x&@JZJ*JAeEaye)5OV^?k9a?-o~@r%R1-Fx>w1n
z`#X~FfqDPH3lavrCoCoEJCl_TvcimGZ?!NbHihp_vhDvc!kf1kw6CpSkEGMQZ>-+{
zns1#C=v(WzfEHL60Q%1Q9iZ>6-vj!=`U9Xn$hrrOqywS7$hsHuK4jeoc|Wr5hkO88
z4?sSMtOp_IAZrffL&$mvavo>R<3`es=QL+M4LP5)=0iTiS<gT|%URDtKF3+lK|aq}
z&qKbzSua3t)^%&vIg%wOgZ4t#?FD3iv75gL<N;#00LTl)ZVMp?iroSs2Z`N+Aa9a7
zZIZfB^McJ%r_E9qvchA)HjtfCr7pxg(txeiI_;FY5JNV+h$(g2C3PXTkp^sJ)@irY
zrMb=i3~!K1oie2^#Doto7)hP7q%On|7h5_>owB7a#JtXc?btf)!T#G!-0-fX+^0w$
zK&%7}+I_jteaH{wJ`W%l%YBL=m&koeAU~A*JcL{-_bG)OsP7r5cP5t61}#Y6GYIk`
zea}UZgY`XwA&2ODhCmM0_Y8#`rtcX>nJk>)7T9XJP1^soIjd%-C*0I<o87PPvR~hg
zwApE&&3O3Xx_~2C;OpnxzW3B_=`Fvs@TUDms%V!gI+Az{+I7+Xy4aiaI|eO6a*UAr
zk_9>gwl(dMBRLT>)E09SbCool^b{@oWah_^pGf8>kb`A&unhD=WOE4QP}v+xdDi`y
zai7@#;*8f?D<u00X)ui_P__@0+xdO(xGstO!!Q4Bw~Le$&c9akb@Je|V=EqwS85Y`
z;rwIz?#DY{ZI;aV2^lqqd5!XMrq^-CopK&CFb^)BM6-RPSqtrvV%WYhkmFc2j(v^F
z*Rtwb$SG{E6xN-}H?h4oLC#<YW<V2-t?a<9kPq^G55l2T4&OHi@*%$OA;?F0^(g-u
z?ZI++H5c*;UOfT%0^jQb?@skE@x3lVzQXsq0=a-6SOACeh5W!m$hY}{w;^v3-8P7$
zXbn6sx}O()Xbrp|x?g}CDh&&j{Ai5~lZJ&sUMvk;OnG*e^^5mk?dy!!qU&@$>vRK2
z3o>9ca{D;BFEJ4^V9R&sYw}=XB4p4SWxGbXH>pD8qyye7ZawBes!~fIXqSr53bdJs
z8?+}9{x_ah)lgXrmD_nv$hnj+y&nH6U^7Z)a4zFPt5z~QIGOdJRbeALcq8QV{NVGv
z2Q7+={NRg_FY|*hQzm*kx$U?GO>u^<bB6A9V&{!}WnJgpy4Q(~0d5nz&ii$*6SH~z
zgkrB$aUw0F>tdhl;@h;0BBWO%q`#3eAq()h)dq{kV7NoldZ>ex!Q8e(9k_Y6oai9w
z`+9iQH}70r&>atMyp7&H$$vIzql}{f`51lLiC912(rG>XK|ST^Jo9v8s7cOgo#$!Y
z7_#+mz@_P#uNy;(mz)y8NoqYJLINVMBt70Z7mdrszVsa&E4s#tV`%Q|5bbxsO^?H(
z%VBX8&8;J%%Mr+TMaR2vQ{<`W@>Co}a?_wyi!RlWYebhC$g8E^tKlX}>%kS0$>hWe
zexlanD(!-{b1GQSdPs#lDsl1tO3Trd34Om0kJvl*H-k30>*QhNO+tnQxM5|AN+x{X
z2cq2r_&jSR`?WBAt>;)sGJ<w}ANbinCicK@tnq&1+wB*K0U=IZ%h?u^CNc!Y1>Jh4
zg$N9|y!!tozeWnspsm;UUa#-}bK6t4@WAzYGhx!sarmt#Ur0dVZ?Hj+mcJMtffF=u
z_&hN7{Bro0fW8|36`-$uzJ8N7<Zpbw0W{xdKA>-Xz6G?vX91w^e7*zpz0dc6e(?DL
z(9hz}ooRf(8h`Cd={MtVfPOdr4#=M|`ZIt77-InBg^Y0_<Uqz42sslC&xGUQEHpd|
zayA;C4S5gp*@J-9)?VbZ7xF&jvk&rq<g*|00pxQ4@<HTt5ONOk$$@+b`5c0L42j23
zXWFIaA~6^831mEhx>ET`WIPG^6f&NIoQI5gkn@o-AMzPwJOlY0GM<Bc9vRO=zKDz$
zAzwnqOOUT1;}ytPk?|_z>&SQ=@(pCX0r@01{3M(|{K*af6Y?o;_$kPFoKGGHtk_O-
zKBpn)b3XZy&u~6xAfM%Y&O$!N`J97%p7S{m`2y#20rC~jc!ldqUR?(5DrdY3`5I@u
z2KfeOyaD+pXS@md7H7N#xsWp!LcYxzZ$rMr8Sg;8#~JTIF5-+ukneNG`;dz{V=?3s
z&R7EZA!mFDxr{TGK`!Tv<&Y~lV+G_&&R7Y#iZfP0e#{vkLw?E`pF*zYjMb29IAaau
z23^+%IG=3Pb!~*)r0d!QxmoAbtQ$bzg)ekYFChDiPW~d02Z&AqkQa(h3n2%JPJxhv
zM5iFgt3}(@Vh{SRjTLQUA+Hl{*FlaKm3Xl`)jufO9u#}fcX5trn*;f<XnPp)5z+Pt
z<Wr*UDad)EZ64&)qH<d7PQyPV+MN+q%9WyRrPzbM+p9#|D#(vT+sBZfindQ7SBtjQ
zkZVNS8pzF}Z8PK-qU{UF{*tY~1mX{nlmMwaO;4a?7bvNe7fJSuBopPWlDJjsObj&*
z_?&{cP3lYzJ{a)nJ#o9#nH=nESr4$^DVgAUfc-AX1lI%XcS|O?9$=p-nc#YWeU@Y*
zgRhqL0Q+po1lI%X_hA3+>j8IUr#o=Iephz73;CYxbPsZo>{cWXh3f}y_aQ%!-5x+L
zmfeaWm&k4<kRQr!4<VPzZl#d@^^X2@es9nM^o{|L7wR1sLJrgq2-FWIlaQA6h5<p4
z7wHEqf*h<L5DYm)KOh8hsD3~w<S_k!Fv{U&K3C>`I=}IMXvexN+}2~=hOaX1IXu($
z$zjHkaSNv%Wqgk^<0&6!e2zo;ac25)W+s)#u|9F^IP&i1w6(0yTF42kZvs1>$~UmS
z8z85!J}FQ?g`J+l&ZPRASf5Q$zKNZ_3CfT1K1ca+<Xy#Sxx7y<<P*Hl3CMZ8ZyrCM
z>gV&m`H;`@zGop{;C(JY{}=e_7x<af|0Uk%5|m%!r(c5dE4<GYD8Ir_ze44as!e_N
zsH}<qaczYh?7{QMmIL5qoNUnM_|6$mDPQNCuNzP1QMjdwUfJRZ+Vhr(BTK~5v?p9I
zja@JK(w?P6np{GUKCE2VZ&Z)|W!2aed@<8?F|>6lMqP^$*vs|jyZXbuv2AkKZL&tJ
zy72Tw-!)UOkp;3p@?QxJaVtr~qx)1@NtJsO#Z`A*4vjjPRE7mfdix}OU*cocvyWyk
z?Cf|M&vXt;Zigi+o!K9j-aIT#r@ThCUjvsE*2tb~<S}H4A?i=RjTeJ|bHzEnOmDYL
zZ=%h5z20uU-bCgmq>AGlZ#hm+^Z0)|Zzt!uJmr_i^l7x8y!FW>O7V<WJTro%vE@|o
z<ah?0dF<G%4gc1x4O+)$-NLY08?>9u%$v-6H2gYdb{+E}<woSyh(?fRiq~n-<VN&1
z<tF6S1m#U=aubwy>^2Pl)@|Uun-1NELF?FUU@OB8-9`(;ZUdjl>(FhqFzhxA+97u4
zA@)6*-XrYHBaqA3*=6j9R9?Z(u7DiHdqwdhNH<{6qWH;CJUFu$&3i>dc{D#c8p>mM
zuNWwg;U~vHdB<)IuTpjB)(l$5Zq1-|?AGvtN{4REpmprl@R^Se-C7I7Zq1-|?ABTs
zc54Q$W4G4Auv;@|HT>)v{zIBSb^Po)$j|xN&mnIWW^EPTr~2E4S=%A+6lU#&d|Kyw
zS~s39AZ3Xov)~!8VsUISJma-a8o!R7@gn7l+ji?!7qnshZ(DcL*qc4a9QiBmKlVPP
zyTU6Ytv79uH5YQz1{Snlvq6$h77Fo#QR}T5MDXv2lQvJk@VcD!FE@OIeFdYgfXnL9
zj3pYL?0&$g51_n>QJWxFAhiO4W%FuesYc+W@>$k&mIX)Rt657mJV_nQtHE&DdMR&N
z3Qry{6`Yp}R$41^1WS(4e$qGn$9?>id5Pl+aBfC0Y6Rm<`=<!TB?7MMMlxz7lt(fy
zkx;&xv8-mmQqpQ>+-hj68_QT?p*)tE5DWX;L`F@7;U_XKi7@;`#+t~Crs><js2iaE
z1_mFj1*^~-81D^?57kd))KsXS%DALL{Zz&~73!xkY8uo}V+N+dl1pQTr!k{w_!k-V
zBGkXgxLkzcUu3KoVfrsI>LsXuiE+6E^)E5jOHjXnQ464c0pn5t^$Qql0n{&K)Iz9V
z$hZ_j{X)jO5I%1)qZUK`V#cKy>K8NCVi<lg<5$c~q0e8!s3p*U3FA@%{g*J-66n8#
z@hgG;D;c#C`mbc%D&gf2mCVRWI;_T9c8ppD^{W_{Di~iCW37VmJ!jPCQ2#mO@*L_v
zXRObmem$esL;ZTjr5@_nGv4(u{{oO2fI#~RKrR6Y)W-m14M3p&1|Yux1oCeoQWrx1
z3z6GGG>qo&Lgc*=f%?1<y|EBYrSdSOhC%;f$Sn-|4@2Hz(EnnjE{6Jxk;`J3-o?ne
z7^XK8sgY1W61hY|{YYeug!(Iyx)SQIL@q0#{z_zB3H9TU8VB{`kV_oYk3-ftsJ{lO
zYoPub<gy0puR-2x;PWOSH3{k`A(tempM<PQF#IIsmjufv8L7$8e=>4OhW?Y$@MHw~
zgJd*58G*wvX-G|j{?m|K8Z4i5q^3i8I&w*e;Xgv^Bjilpy9VtMa(RTnE?Xs1E1|p+
zxl}@VEwa?YCGA=?t`^!Z*C9(Cl-HpNb#T1-0;w-x_%D#l3mE<jWPO1~)AYSSelO4z
z%KohC&w}*(vo8K{SJ0pJ_Gf*d{v>~R=SL{3hC=_LtV<~LAIe%oSrC6HJ1G?U4`bCZ
z=s%1d7zQ_}!&plg3rupt*l}U-j*@Iv&4&K7S(j`We>Q8)hVf^!e%UN|{ynU^2m0T`
zy6l1e_psJI(ElFRZx8f;j8%_8|HoLDW6=LG)_M&3KgRkUgZ^__H5dBNWnFTi|6JBP
z7v_I1J1H0D-#ONFj_ps&=R9jV5BUOXx&ZkiYq|*e5^K5y`7&#|4EYLcx&rwsYq|>g
z8f&@+`8sR54*3Rax&iqnYq|;f9;@Di&wr0K--CBb++%(3!I4!Fs}@21BG#n{=2sDG
zErR)7$*Pr5zmj#Sg!+}NwX&nWUnMM`Dpsw6{;OD*D(Jt8^{#^DQ^iiIg5^`sTIyl_
z{fo8y1-XH>G(c`-Esc<ySW6S+X4cXS`2}lv0ok9k_;aAW1aOuB$O}2kLdb!fB@l8D
zX9<EF&Z*%X=pVv4b2ta?1PSLxhI8%9U<Pdor!Il|OE{M$9Ow_1aMmRp=nqzL>ME$e
zigQ^7^;dD$RUP&HR&k*HMsaEs^dH5!L_z;ioHYvikK+8Ip#OMIjfei@IhT0oKc2J3
zL;vxdUp(}mz^Mt)e*)){0R1O$-U%G&KNGk~3Gfc2G)_%}{?j=%9r6}V-2yp-Q!^lM
z<<zZ^w{hw=$lEz}JLDalx&!h~PTdK47pLxmyqi;ZL(b&XOvw8=b$>@T?}zo{0H+>+
z@&la90hm7_yc)tg)0t2R?-D|-Oz>3=yc!DSp}b2dlrQHk%Xwfjxtt%j9NNOK;4Lel
zd<8#Y1zeAa<L%<$9Wimda~!+_DvtMvgX@2B{OCA-0!?omKP3*H1z5wYYhZfU^6Fa1
z>v(k?<n_F|9&$Xd#zWrBtD7OG@@gvNG+s@EoX)H1khk#a7RVXAn$eL5X26+h25-rL
z`IW)@X26+i2LEmb57slvc)K!~e`UOL8O*;j-lGiWUl~8T4CY@MKcx)jUj=Vh0mG}{
zohx8?6}(3U46lM8T>-<Z;HOl;@EUmg2EI2fuLiz<1H6T=f%j~HXGa=%-v;=+4g9+e
z@Oc|~QzPG>hS$WKnjkmxre?@5c+(5W{({M00REQ=&dcB(V9N#P<&akh&MP2C2+k3Z
zBL(M3$SVcsm5~1sod1BlN^o8UIZAMjf*dV4M?;PgoMRxb7Mxc@juo6^A;$^Magf&t
z&TAmA6`a>XP8QT;0kr>QVMwyzMe{dVcr#gehjNObra=7^!6iiiR%<DOHANUr^)m(g
zOaYXCrqDkV-iweacxFOd{!GC)ldi<z^?l*pOt?OmC77~c`DF{HY{+{A(;mos1=C*0
z`vlWI$R`BL2?6w9Ck4w%$bSl!KOvtIET<sn36?y_rv=Mt$oYaLAMzQ&at88Q!EzSz
zIl*!c@_E5>9`Xgjasl#1!EzDuCBbqD@@2tt8S)juas_gMpccULED(ki!16B;#uNx|
z(DE!4)Iz9VD7X~D@-Gyug|Pff1-nwgk@_zcoJ(Q-D-}FS;aQYYVRWeg`u9>{N+~?+
zQZCq)!|=)l=W-Zcx!_R_!z&j?m&5SNg(>AQyn4aD9@fu#p?^KBpY?)gJ*=Pgf^R*n
zpY_7K^{{^aC7Aw#^{YWJH9&3@OpTD61XB~_X2H}9Ib3H7*Ma_OiO#eH@=~2?Ddc53
z(=y1*b*ANzSLjSDAV=s-5s)Kwrbx&ub*7b&|InHKfV@g)S_L^uXNrOxtusYKj?tN7
zAg|V$Rzr@}nPMR)=+p!q=${gFLlShLe@fJ;iBO)Xb4k>J{^*HLe4<mxX_zfT@ICZv
zKZ?c+y9Xr415$so`i(b=^mc3MCLOuA5U+-}To+B2ymx(fGVaH>4z0xIy+;`LBaAOy
zXo_du<5^!i5lqChXvn8{_fxztZ3w4%_tTKi@a|_Q?=Jtp-Q5>IP5rO$bj$SI5%cr;
z3O{TspQP`SM9&To3ue60nJp^Wq6^%-wJQ-{gPXVZ>m?7ec`LVFg-uL4iFpUMm0cnl
zmx%qSeb73w*E)E!>{ii!t2mfObV}@X3f?T6C-%yNTqrsgibH5Zn#Eqt^k!Ln9!Bi-
z0<yo<%U=RF(?&_Xqa+JWc<VjgB;or{UY@zdtNOuzou@hTiz?r5{J$xP(`uxiHIj+!
z33FPw+%sHm=bVNIS^IqBJ@78`IbDx)x>2Owk*hB89!kr_;G{Ykw9f{9=1vTyk8Yas
zWv6js(0|;+ImYbpkGmT73`aA<Z$$es5OTvfuCuN8ijxP=b-DES@?*P~*tT4OscvM<
zwAQ|0)EAHwkeYyRV`x7w5yYxNteMv8MXb6Aavf`_W8b3kXRPHJ<P=U#;eh{5oVp2e
z60aulX4)!}c{Lew3a_R>J}Ov_3U5*UV}j)v<Xpj$3pq$<3etfqCxdmSV9235Qz+%!
zg1C=Mgk#fQ+9h}|hU@Uf$h)}1x*hNB&rhKV*(Q5$qgUv)*d=&p>ZcGZgA=8({%aq4
zp4W5QX1!rEyniTFZ%BolrZ=QPPS+dKA#c$ewm{C%8!{ko)f=`#-ljKfgS=gD*baGz
z-mnAmPQ76#WpWmXmD?_&ryb^D$^Nj^n=}OjzLa&)I(Z1$UnCdRV^i*yTk2_zPnL|y
zQg?FC5?-*C%#WbeMUZR`lH08?PN_fMlYTVvI^5JmV>mU;=l-qp;k2cqaVhPLIBl6|
zTqgD*ofG$OhW(tjLNu<x{ZHGC`J5IZ8Y4s#;Zx7|UwS^W!WH-QN11L%nI5!FB(U8Q
z*dDabZ(zG^fSkv7%j0{{s*%rk%ZGfH?{=0li4eEi*8BQNt3B)e>Mv6#J4XH2-4!|c
zfjgVld;MwVcf3vV{FU#f59)pVcYG8hPuD9?=L*kJ^g69`rQ6{7x?cG@S7QHjzWB&j
zdq-X>8N+EmJN?{;WDTeN>h){;JxQGQoA+<+mL#0^yZ7&a{F!0?^o}D=3t&70=s_(`
zTgX@!QoA5d3uL?k>A@^cJBFN&p+2M#uz418%7uIanNQFiWKKJY%qJnALhh&NIz6Z5
zA@@AU`N*6P_0J&l8OZ05$2ofG7N?y@9_JxnL_;s4_8n?YyM%^bf_w#8uAugHJWjic
zELS04M_$+I9WeNcW8`%M@)fTC6>b2jWY|B~|0?8bT>op3Z*b-tT>C*UPP@sOZ$iGs
znQuWZ<lGCn_Hh)xa+Y(y4fzh|eh2bB&U_F0FXGHakneNm`;dz{k7BO<f>cf`;XFzp
zKjb_fLN4QmmT~Qu+;dtvH?$mb1vj(;awTV}<l65R=Cmr#QU&=jXL$_yDd+W+^QQS(
z&3RQruHn3DAg>ljt)_M%oQBT`#foFePzPU8Ee=~JdQv_pdLI-=kf9AWj}*OgAmcMc
zhp7P_r{Ob1M<C;I(<#xD`p4s@Jjj)zccnOj41MtPir!U_AB*0PA>%VhPsPzxAD=<0
zhK$c3)j-DMux7FSE;kMj#fH6rjK^XA^cWE7!>}Dp>*qsg7u2znYdLf3oFfe)?-^`R
zCA(~rJ&4^N{tSg|xgn1v2YL)zx!k#2?oMrbm+5;i(+?oVcElg<O<TW4LL$CA{_oO~
z<w`IfGmY^Z(|*bL<>z@tuaswaVw;uojC!7NBYhdRVL|E&1b8>A?q=OckAvUvteOcq
zl2ao&AYaL;D<MDN)d##AsbTmwXkIObT*9j*knaoXeF5k{5Yz{div_iq@>PBAz<0tT
zXW{1?={K_dpcS#zZW+;t#>Mc`-7TX17I^9IEz$87yma>usrMi7(%sfiw~&%01%NZU
z^@A-WqltAlo*s030~@r}qAPATq)lLRa?yUf=t}vZH~=>jQpa)C5C`C9Livv9ct<o-
zt`rC0YESAgwtWx>;A&6#u{Z!%d&)6V?-*$)VQixH@@q2WiOA*j_=#G+Aw!<$?(nl?
zpQYO~|MhE_e0K&9=R-te2s|j+`Vkru@qPPb*A>N`Uu!wVDp~3zFFK}rCRv_AUMgFb
z%0RwMwk)Ikz|7a^I(HApF*eFlquhnGik8`}v{v7Rn4p%p?%6V=XIdgwtdV-xNS-vt
zaJhH5+|FfF^$rt#K1;y0YMkG=_Etl_m&401zmr3<3S0B(#7nwY$gBrXHzaYR^a{xd
zgH|b-E8$>ok!)T>2YXM~?8zCt&8hEyoz>M!X+>bz<e~pV%XV@o0XL+M14n!cLbTs0
z4xrWep6GZ_bfwiaR_YxqnQ66a{eTcD{kmU||JJSHaOeMF+UtTRWcNL_qxRp<&y!-p
zopkG$ifB3Bl<aQ8vHg9??mpyFsY@xH2DhBwH&;mBv=#-*=0F+n8Cf|aJCe@%S;N-!
z?q&0q;=5BfG3q9GO>{Y<mP4*!)C$PKNDYSf&TmHQX2`*;8Vs*NzRjw)A>U!uJCH*;
zH5A?>zn4?@Lf*)$8{u|EqF|ROv=6(m`LJNO0rCw&y#bdG*X!)o>)Q375BcV(U&irK
z|8XIqp&=z_@U5sC-0m*x#LGG-QdZ=w4X%bA=R*c9L$uF;2XzZX#{zgzca_w86@5eF
zo917Bfr?baCN9;m!!~m*j`6s}9hW+_eqhqZhSXc!Ew+A%iufSMa&S^QPMz?*Afo*)
zaUg9GMWSPo=tf&aoYXr`awj3nZQrkI54i3A5FBvFIUPQbB-(F-3re>|$J=nhC|c?r
z4Htk~zj{Ri#$UbqVh;WUpUtSYt7mOUjrU2{-?OXweL6#?ujFO{8{(I&9NLfx@VE2-
z{_XOXO!6c-Wn*KZe-5XA4x@j1(Lb&9&yn=cq4ZDh_CJv}C4TEG=WOcx=wBZ>!qehq
S9qv7Jn9b|$3N|C?pZ^cHwT&<U

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise.hpp.00198FF49AAFC017.idx b/.cache/clangd/index/device_elementwise.hpp.00198FF49AAFC017.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7fcb924eafebadba033a0965debada9c2fe9137b
GIT binary patch
literal 1140
zcmWIYbaTsLVPJ4h@vO*AElC7283cj2xTGl4h>?N8j+ud>V$R+P{@F~10!P0uesl2N
zmga5C6`Z<lD>A-5sO`LYk#}~A)4vSykGsvYQn!XxWoWV(RpzJF^`{&@x9z-B^2LJQ
zpGSHsAAJ5Y$9;9d^|RKu9kczDcYKqN`tQ2XP1x+x<133EJ5I9vVq=u)Rip3pPwZ>N
ztGgP~_nOK+&)#GF+B(F1|B{|fhU=Cd*nBTLRV?P)sqC7f)jO`IsrVjPY-X6f>f+ob
zTS}91cz!Fn-VR@srnBJV6E3Tx32ST|Bz_;<cWRQ_v>-+1hAtDQr<Eu7#xyVG-PI($
zPD<aF<I*&ncnzH(n`W*hf?Cr!7WqBjx9J<dSS>G~w)Ddpe?M!9-Ff`~;)cH~xixMw
ztZSMou-9RUO<P<&L!XAjmvyaoc;{Vid{X$l;&XRp@2r@~wG4O98qO}R%uU({3_;QR
zS6J>y9IX~$V&wvIfq++%MN*uB1Hy?q^KJS~mbd>znb^4)cts^e6~P2dy6|HNud0#x
zF99Ytm~I741#6geh2nbgjB_b&Y)s6Y47|eZ!V+MDfs+XsXAF$IjI-V-<n7LrjuU0#
z02(hQDW(J_VEX5DRyUfny}QiK!~)c>&ZaH}CJ_3IQqzi`0nLK>f{_#KdYHX1*E4g0
zT@4fF;gxce@`LH&;gvCyv4inqw_E1KsHrG3^6*N?NNB)>Q~Ov|8-$9TfWpE`!g?^1
zVEzZ%3-UKi7)CR3f&J^LQr10jmvbEtuLg^TD9lq8haWzvV!3#horhOaSrTp&&~eIA
z$UHqMW0)&=c%_x4k>sTHq><!glx2|QWR+#%ylyY%mbd-?Etr8;X-XO(jFU2vvPSTv
zO{C#GSj+%}4iqgg9|2w9Chd>VB5NRP3FE;$#>fSZC77?*8e0V%otrQn7<39U3hJ;R
zduqV?g`Ho#8|WqpK?!M?AwWff3eqq|JiO{k>N+rki&AqEm4FEc78wjMSHWn8XP-|W
uG<~_ZwYVs`*cz0A`9TB(aKji3415SC4}!_dY*9HWT1k+fg`ba`hXDY{V|so7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_2d_impl.hpp.30FD009D4EA9FEE5.idx b/.cache/clangd/index/device_elementwise_2d_impl.hpp.30FD009D4EA9FEE5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8746f853461b34dad76e58cc781c222148857142
GIT binary patch
literal 7354
zcmb7J2UJu?_uo4r!tz*Fy0A;xeMnuZEC{F|MJWmvBw!3h#ppp|EI}leDA<SwrCJjO
z#D>PGks^NyO0n{Zu^jzAV^r+M62&jZ#Q!YpWnNCsrzaeI^M2g>?#%q=-aB{3v5^rG
z-^oOx&?%AgXHK1)sv{DK`qO`N=FXlTPeh`{dLq&M%?<I_GOu}Z7prb>%ZaY?t+x*U
zLNY)+tU^_ivVW<<=YTYE?7YZQv)^BJjFw%i%?K_v^vqbb;-<ZkbJ>SmhfLyAr-uAW
z7VNTl-so7fC2Y@)gkxD<gFe1dZ`t+ih~b~{ArKH0Wt3$uGwQf^>h2`PJl$a*-1|>U
znv=nV@poR;-#PqhMRoREqs;YzHxm1Pp7Ex>vFbDZ$v>E7-Fqqz^UUOoirepwmiODd
zudUO|{&hpS)uC@fJ74A07e7f%8ak>$wZG)?k@(LF|J}O3?eL~n?@;gDO4EOx$||VL
zi@&WqcZ1)c&Y!Gz^%)pfa&_st8!aa@3Jib9cfVr1Gsp6e=-uY7kv|o@SSm}6e(4^*
zyEGlvRxFZL`??2zubO_sBJLA)s>{UetC9J}-=3d2byxYJOE15eH{w+6?wMRCIqjU#
zx;wz`#E|xYij0DFpNuKo<(=kLGVF`~2S%<JT~%g2h>Tpn_d(S9H!!s0XzAB`rsibl
zuKChZWF=L4dKZd~qo!wyzp1@D;MMHN>TI3XnmaQ(E32AJFU=ns-{`qoH@wo^B|W&A
z#J+A!D~MWdHtBfGvz_681{~@7QD^9lQLYajud-CHPT#aEcvRM?nlmBWZdOhGcUQr=
znH3M!nM3zp-1Rot<<rp;&s+aVad_e3Q+Df*y&svSmc9u)`tpTo(B>H~NBZa+^*?m4
ztL;kgq|tYWZ|*m_XvM5WUf-N63og$ZRhF>K<G*nyDiUM6l5?MK(i{6#pWXU@ULE{@
zX*I20AAL7{^V!Di2-ADnZJ(w-_4x5wSm@=K1HznKlSFf}XQp{mSJUsvUHU-(b4Lkr
zmy7u0V0RomwT*~6p5{kv@}SIR&yqYdVk#5y`~cp*`+;Js@H2JMn>JLL5fds_$Q8=&
z2a26j`<1{4RZq-_J{3D!Im)^pz2aA8Ki2JTXpj&$+J7_9UpH-oDqczJ#G7{cE;A<m
zsf9KoZ6ns&2310%c8;}u^x}#EF{diw;&A)!N3Zzl-OoNZd=hRz^r*NL#HAqDHmDNo
z`pLI7g=1Z%L_$?o5&cy}rfpDVjAwH|b!mdFF|nj7Bc&si-47HeJhS}0e|^U}32~t_
zX@P(ih}1Ty@{w<!_!Hl(`NNp>qZS5B2YYlsP&|9auEk3-X5~wWEwykG<R>W?Jx#C5
z`nO%PzHGc{M#R*Dla-UK`+?#mMkT|yc71Tcob;g<%0aIjq}rxeW!;f;b0Wqy8xRX>
zp$No9Ak#LeQf8L-_MG{ABPlVaDkq8kN#dhzP-WZs36EEmZC+?b%&3aP00%|)1I0Vq
z)=pUSOPaGTF{I+NU~m?UwGFCN<$aViL~MOTN(R`7_@zX$l=Rm&sB-AHn>X)yP1vSO
z`cNy|z+fAgXd6^%{I7p$#MWDBQes3^YKcxQvCuZC(zH_58CcsvjEOE)$pvXHWwob4
zl@nLJ^S#%vh%qKos*)^Cw(owR_)N&-=~enS8hVD>4w820&-XN_(y@JGz}dC&>m<aL
z65j;2P2jF=Q040CgWvyXees9^;b<qtAT9=LZPTmr_3ArQOIjC8h!wSR3Z$pNO532y
zjiZZptt4*?j7eW=#a-&|-u*!F&q*`;x)rrq8xR934iksjbw5!2R33foq(}2X2~p7w
zS|P9%qO}dGJPW^IyM61M7+s=cE#l?<<Tl+8kv;*M3xJ2&Q`6_D%|xQUT_#GMwy%nT
zqoHJ13hGd8l<C&{+69I0Yakq-V4o0#VZ-rR_E|wlHRc_vR0mug$k27g+U$dikJ*P4
zcV`_O9>=>FyI3)#Hc?w6HyoukrWnQ};~ZZB!zy3|cERyGz+(saF<b){HDJqd3wXAG
zKSngiUk8ipV2ggh@s-59lGtF*pi>|bb!0R;BgY>l>Z4>RVnffB8)p=IU6XLUqphO@
zW<-v6vUPHpt&M1oS95A(HiZx_#2E7h$EU}oPeAH8ep>9bc!qL8l?y)T<{ZBcxOHI6
za1n$TLG%)B+-t}QFNSDFP7O$Fz?@C~7)Xx^@pX`12Xi+0`yjo~u;JsFl<$?|6zT>q
z!2bmVp`vv8<ekS}U$vntoZ*rY!p56vmg$Db^6~+_H&sWxi4NZZ?j7KPU+eIB#3GMa
zVw%8R#zfq`yRd~%4oV)5);K;TC`EwM6VfL!*|Z7MK4fT~**t58R)TybC>dwv;93q|
zjFSpbS3nTMm7uN^;%1OFgTh(6nQ=7lJGOv|IZ!J&wt@=7kK+qTSRsjG4pc<KiiG$8
zaXdg&OkPX8Yl%NPl*ZwP)R9nz>&cLMhV7(}ZtV>HDTPu$+HkZ9i@;dJ*j}U<cBuBb
z|G&iVr5^p2h`V=gMtz(*&L3^#)M=NyL9(0qk662#V?3*-ya|+(g-!i*co%CIIa<@<
zUF}>I46Py3HN+SR(BVsoqLlCqml5AG;>Yj?;=4hJw~&A>Bnaj1u~SBP@5#!ib1fk5
z1;i5tI{X%**uo5BtpQ?fg@GgTy0dkkbxQLmpmEgT4hTU<^&WY9weEmKbsvps`vRh1
zrpt)3jG5NqHxT6phW#DVM5#u}^XQnyj~PD}iPYia$Hey{|KX*>55BfH`dhbH@#u)V
zD;*pr$8Q9`jW8N>F-Nnh-f^(Sbj$HiK>Y+p;pkD1u5C|Q#0)(I^+Sk6#Ss@a2X0CV
zpQFq1iCm&Nj+Wz-xFi9Z$W0WWY-zR?i?IdL1p<^K&FMu*8D+hv^1@MIH+{t2r565~
z;}@DN9LOv$0h1-ri!th`EA9cw9*|?$inZk!y5a{WtzIenEFEa-nqf6VjrKka+GMi5
z@4d{waX#tD;EWFk@=6_!Rw43j;EaYiTDVv)B|Ii8j!r|lig@6>IKGCM)(|_4SIp~_
zbxje-rWLLVkHMAAXsUAFwB_Z*clrmGjC8KmAgLD4wWn~zA>uYeyVO*TtjjU#F}*=N
z-+sTv$LFR*P(tKR^1-My|M<_h7wDWnYr*l!s$_5EiDp|>iU5^^tQ_o-Y|W-CtN1H+
ztN=GAuLN18P+kd+l|uOrknIr4t3XyI#LZyX491vvG&$F>1&nd^HQ2HR<P5ihVJjH3
z9NY?)tsrN(koXpo0JdgD#J5O@Yl*U!IN=&<a<H=Quh_kg_%L}rQPvCP^~AkiC_hS+
zM}_hRqHJIod7?4VQ^(^<;zmwmqNjStW#afU@GFBzjA@R44C==)5<@sYs!#6YuipCG
zaJ<4yA;TS&<0JV<SBx}{kK&^QD4LHJpcp<zfW}J3nz8T3NyZ6KvQKh=z)^fs1SnOJ
z>da(miZlV5te7l7QxsDKs2mjK;Ecp*986UKgBY#^RizL&gGn=32_3|y1ssG9Vbcl@
z%mGe;-YGD;sx87f{ycE!!362!XyWHOg!lq*7lgQw1QwFvLI(;g65?XwQcS#<yo9)v
z2yrcOt|gw#f$GShIugimJsDKbaQ=g^OCP7te)Eqvo-d!*{_rH!w)4H4;ku4{Prs<o
z+4s&2RQxJs(o!SYMVb`bLDdd}8DH%X-3|#1FC&)Agh_H4Q7<FIF=>{Z99eZ?`64%(
z${LAjBe7)@KTS+e6I(o%VoM{V=WL1ru<@51gB=TQ&86w<JkZ}@?0U;@`}3|?<jbk)
zM2kcZ=Cg@`iQ^FAk(M4bH0N0~%JHt|uC@$?n1skMe{(bfV~QXiLy{g0U|1!LWG-0+
z&Q;*e@G-DB26ha$f@doXW-fRiEbfCHj+-WRVqQ&bnak9Zh<Xysm}wxM4P-Fl?Thli
zIwSE5f9HZ#gkF;mZ+gAyy*%R~o+g@}6FeiKjX1UucPy4T{xq>WO&oDya7F0?^_=U`
z(%$*?edA_skDx_Fl0}j&GEdRKBzJ};T1*tkCJvk^kmXqCC|T~vamnf3fU120;+nUl
zl+t;f1=(3}#9le-CZ;Eu|6n5RC)QpjuzCG8(`r20vR$|wdBio3_~86FKA%YPi8))I
zeBzxi+%g*Jp3hSH;fD(jTdaP*n=a404?DFFRXL+%!!%RL7?+X2=9?LpIUZ4ckbVE#
zE4$WErAne#$%<u-*XUI<e6;OiVs+rJns|SAx9CQ6-Fj|PkeDXNMP7b~qAK6%S9o}$
z?$UF+f?sQ*#Hk3vSd<ilQ!#|$3H12RxJ<(cz1;UYB+k6AxchpB*<TITtHBKi!0~Ir
zdaV#&2J6e<#-i#bSl?v$lvS(KZwuqSY5%iqX7&0L$A4<`X)nT$DK|YgKj>ZUK?6S@
z8!Ib$FT3Y2%}a2#YgtSarMJDe$L)aGE3W?>pyK!;_Cq`jK3@M*<h?WFIQ}p=90s58
zel@lQjXs_9DghjhGK}|3bgTz~I|$a8O2pdRCUowL9*?daj87@l=Xg&ePdU3+2s8?m
zv;37VPj^LH>Aiw{ngA^aqvarHtR4iLgTnf(4s7c{g#+ODi{N|_{E#S)zX2{cz#sAD
z&Xp5(TyO?D^6@U?gITo4yTtdRD|weKb>~@KqGZ1NoQA$bovr64jX6Up_?5y)Rs=l)
z^&=R8!@%904r3ok_JJjiU#z`oLjteu*ysKA*T;v_&Ez52Jp^C0fK<{-T66uzIu28X
zMdSV3&;IS9ul8EPd%XfbpyxQvZ5^|=0sJ<=C^n5I&}#x4Tz`)L1=PPl3=ZnHWu;re
z9V2%-t-s#FaQsW~f5~oTIz9Wo=x1b+PZev3eho3fpx0zc{gcE5llWb?+G-#9p6zu0
z&%ypVTi8cYd5PS_VH>E0{~a#4qfw>3AlVBxjOLzON2D1iOyf#(*BHrvmcOnpbkXz8
zwZ}ZZwddc<->=g9kCv^99rTZ@pss~2i_Cxj!RViVxID6kKlR<qza0->{l0&%#e<oe
zgq&@i?ZLt-J1{#I(d?-isbTaYHEgfJP17ff696|8JY0&989a=MkRsgRM3{Q!%$__a
zOY=W83eBVE#|~dagg<jOQu$pr|484-d7z`xL1AYv8rUlqIj~n}_G-^w9oefBdsVX6
F{|CeLZ@T~h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_3d_impl.hpp.EA5D0D82E2445BC8.idx b/.cache/clangd/index/device_elementwise_3d_impl.hpp.EA5D0D82E2445BC8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f0659dbd90dcaad4f57c9050ae412e38ab978e29
GIT binary patch
literal 7890
zcmb7JcU)A*_TMuiWw|WtQe}5x3oDnh^rm7^>0l61?D`N(gr{iGsL_bMqN2ZOioyEi
zVGz`)paM43M8(7|2}^mXs31|JNi@bhf3vWMx%s?5{K!XV?w51!ote|;keU!1d&yNK
zikg%#Z~Elf6U;;+Q4jim*6f_j*+e9I&q^ekS8yWjr)-12^j^_R^}<<~eoXdLTW+k5
zQ91Wcu~?{c&Yaj|cl^p#L&GkdnyVZ0*}Z>tPc2%UzA30i`M6!&>w~x}%P#nF{P_NU
z!zVueLz(?#>87XGpD6p;%3t+*IX(C0(<XQOov%NgIdkIa9KR#Q8OhpRsV^^n@}!Ub
z#32WFFO8g5UGp&itC{1^?702(mj1`e%Ve<komqBGx|1bq5<Yh>IP%Z!y~gI`^@vpc
zQkUHPVb9=)!%k}33JQKcyuZotXmnoP<dLV2w%&3)r2I52eqyqBmPKghghj#s8gnUh
zQL=Bl&*qOd23(FIrNyzCW1SOx2W+`DVB+;s$4hs+Cq|797LOghX~Xq^V|$Xyr?tC&
z`!cn-F!7F^%r-v4{g>6}iYDbxD+w*j2&wHoQdXJiA3MgjwW8d6d(??hg<l_hcErK_
zMEZ!@PsPPIj@<rj&E!*;MwASC+Cr?oNBr4lm~c@)IBeJ{i<rlS$<Mxi*?pb+$Mzqm
zNJfpjl4f6+Hh1?!zhAfaI{DwRKi549Dj9Un?CQoZmTa<6t9NgnQX$&fdwY(@{hJRx
z4zBa|bxIE3`mlQHcTaTwmv{G1-6iXpx-(&e$F>j7SbM%+>N}%)-L3lzr(E83!!NVf
z$0-TFJCYynT8z7C=;m4C^4b1a$(VAh=D==iZ2FblnIHY|^76D9ZH{H;&5^6cinGc~
z(>5Razxg%t%TMnf$n`q2|K*a|k6JyCMcy2?-$LfvZSNj~c6vlv&uzMk-_(xo^YpuA
z>ax)bTneUNuaf7UU#0Q*HRsS9^@=f*rCDv)-k<4o`&sw1{Wknk(6DNi#la0dtg`M0
zjh*`X@S|wWRhz$Tejzi~%=&nGMm{w}^1R{Jb!AMT195T_@g6-r)E$G!lG-QzF)N<G
zzfa+Bn;nQFCF_(rZO4FQ%lV;?4~IK6N{Is{w-I$42{KK{;yB{Xo3R^2S7gMTvTOzE
zR`4-R$m0CGEO|_~jRPelKqcavpl=f-nI>fES^T<R+otvpCB&DqG(w+7NH9&vq8i$-
zbiD71G#Rm{3$+qyE0LQfWYLz5v%9-wP>6&C(ytgmWdMKEge>momeg#kjQvYSEa+Dv
zrICIe1Cj%6G!MO>K2SIiD@yiq@ltdQNDf(Lr9XY&@|ujuDLGl1tnL`P$ZOMA-vF~K
z65>bKuLHL_@HI`y(uYj$BdaQuONfTDG=s94imP)%mUyVw-}}!OX^x~DWvKwm3Xqy6
zWEqkW^F{O)#}*0krz{QN-ax<7IU&oKyo5ymH^W1%i3{CoF^G#nVVaO7`|2^j)ALu}
zbs%EO;_c$C=opZ^!s9>i^TpkNSQA@H-Us4+AU93Ovi`Vp`1$uTa;%9ZWhn)5DJV@7
zvgBJ-pS&z@<{XGKW#N1BnvMa<+mCD*u)%rQRBIxk<V10zs$)R%fqP#DMn`N~E+gG3
zIanI(+c6-y@*fqRYuD{;k`X&f&X#6tItC=4kRJ1iNsW&*C*3Id7+4<zJJW<Lr-!=u
zrCj~*K`BwFMf?MhJb)gisf%T8)Mjo^pYaZaqhHax=oK9UlFz%(32G{jSt}(nO0FYj
zb;QXuA<KpLG`SC&3|0=rnzHz~_$WFCB-dz4ekgY<u#pfY)#e?L-T@cWge-NpEw+xj
zWC(X4HuMX=F20J60m=8D*PIDyJ+Z=^n7N90m4(W+V-Q&ppiT}vJiPGR@fyXF;pRk2
z-){ly7Lb`H{6_1N$kV^=P3zw&iS0zPo%G;4CuC_q$raqZc%;yp^rT-&5+|uU1|+|j
zwz~P7qS~!e(u0x<iB%y{m?mTq<xHM3>%1e2OPshP9^nBv-e2bL!cc@=gdA;vqsBLA
z9}Gjg<oHU6tb{o97mhyxekUM|;j7?u72FtZ1ph_|$5_MhFTv?0xS<Jf`~~88fw<x^
zINm^F4I~*A%<&B*w1K>b7-OfsxFcq=O=tJZjzY`i_!<3X2+*7nbH*{++~IS_GPD>5
zE{1r9SAoMSkTbj;l-ogz0hgmO!Mg$i*zqbMv=Sm1u7c1iA#MOg186XIa5TPpG=iRq
zqX|5kK#$Ry<I6~N8A)W~C@0b7LVSY+-ynS$y^aLdk#H1`QGERjB#PmB(yyLj^U4ra
zM9lsZgySc3lRYrBa{L<5tbu-bY)nBQHYFaEvyK1ityx<F3_UpB%gsxNk(=Ya-Mn=e
z+Bn|F%}0m+$MJeMz3#NhpE=&w%~yw>#qkkbgpAo*92Y0UWWn)sQsxZD9dZ2J!E@6X
z+6wxu5X96{1Y8lw7%qmGVo1VxY_#H-eUQYg_$o-Rf+JH~El6vH_$5eRf+Mri<wUxi
zbYYvH2c?&5#p%HD`6N7_L?EMd<;1<GUe&pF<M>a!J{iFFo9mG4!_Z=ISPWelcil~F
z%Hb$&j^+;fFIspEdXi{7hI8#VmamJB1`;#A6;fJZ2yVlS-$k5u5oZi9XkJw9XF+ln
zT=C#yQ%Qh98|{)ZR)eG(l#H>nK0rp5wEgbhsHSwvm}Zk^#|&w>&G0US)dMx7&e8!7
zrinaq{8{)_@mn{J#RXmFI&jxP!S>TxY2bd)YKXX_w7`WpK1rW62t%lT((ArTvas6R
zjQ5xOE0M{J4^RbY7&<_t2Z#(^%#1%yG{=cM!xbc?f`l<#NkS?a_8mMaE^y!6JTT)A
z5#K|^9~seE=oXDN4z9=XBfUoYpa28o@2xfOIV!Z9@on`XqQSS#_zI$}5N@NAXe$}U
z4^SgAmPqXDo4{@pxHH?^0(M&%4lz|vRKPXB)qtD{xU-Z(0f*}{L=$ZP@e$ovdVG2U
zE@#G1h@T)p8Sxoi2+J<|?Q<YGCp4zc;tRhW5%H__A-lf@Q?Dr^(jwx9hc@DnA`;E;
zULxJg=yCu2(!8v|r!V!xQc#sbDAq=CHwyZGH9BUNImb`orv&0|IX;um6re0VOMvn~
zng=dSQX4?JL5TA~n$Iu_iW+xkxrnF1SWl0A1|(-#R1ur1MPwXl_gz5xerFxs{d|}C
zp|~)5JJn=o1<7=|AHw!S0v?j%SCi1yBp$tf+|l0C#`bd8nsKz=QXT_OT!^D3h4Xge
zj^Ti#1%kGa_%X5mMC^YeD$E+ws*60o+V<O+chEJ)ernvtnwD*Ol-H;(<DREJ;57Az
zNwM_6DqhtandY5szBk|O+V@TzpY59+gmR!++;@fmRe+)b)F^giA)=`KJNBpqA4abN
zMU{|V1s+vG`Uy~+5YkVA;-nBafK3C)P@BfgZPN%cjF3j`+z3jBo4}?CWEg5W8d;s2
zK*?|!2`Lk1?s5`RF2r?2TSs`*fiaV74S&bJ1`@>R^+a1Qq}LPQdLg}mXd8s|Mxt$G
zcwW=`ov+5FI#6A#B&wAp1g!{@3pKya>j$dI=wo4}5YNnS6$xEM;_$`1#BR%f&VLiG
z;dq^cPJxvg#}C#F_GBKIqDc{;A(|lql&VP;prM+f0yI`K)`5LxoMfB;jhBoUpzPr6
zz5*i(o*_UpwKKgK?L+N{0+gf85ujPxSpqa$J6nJ%KvMyp%+xADUkSksSAo7th#SDJ
z0bGQp<k|>2q3O6bfsUEV9k9Fuws=M1Xm;mX!43t@@om7h3GpM~9tm+faP2}|M*5bK
zfy~{@N#AlIK1#ffk^n|OM!b#*aUJojBmT_v3?$e<`Z8Qkg6kR1Ym2@;bxO|bcg|t!
zo*h5)B+9Mje^Mu!9E~oW_g!X_D@o`|l7OPYi!Z(57;n9dEx<9rjiETZI0aS=#v*G_
zF{EMG<@hRytb%x^&6D7H5&{{n1*ck2G28_HP0*WZZ#i*VPE@!hj=w@2udtlO@%1FO
zo(yKvYb5@Sq&MPC*B?GzQn_INI|Sb?FVe>?wN>1rf&T&MA3$%m_5(<I03#UQL7aC8
z6TuD=x`PbF#PHe0_@WytbA4#)Y9{v0#0@va@$JOEow(t}VC&-r*8iS=cQ%b4t>E4Y
z{c(kDe;ioxqf?%ert&F4Q<9j&eh~D508I^=+J&$iVJ7bALNF~IfUpDbKI)#MO&!ZR
zaK(y+-dRZK8Zro1-kqD5e_0Y1^A2*`6J<~itbX;?TaQ~Zud;i}^xd&mw15c8?8{D-
zIWRL(fQDp_Vkj#pOTd{mFiXIhH6%;GnGf=O(6X?U4_^5YD8#+<q3beH8q_D{{I>MZ
z>8@Kq(E=XWXP|L`*tany;k}A_jj@Tt1e>T&OhTz2S>$(DCC>?>Vk;t=B34iuaa<8e
zz^Gux?<JbOj9&I=ewowSXT@)=;g3JkFL+-O`ML2<H1v}Z>{z)exx*M*49SZj6>(*R
zy2riMzptl<yg)23uyWk!PL@{~#uvJaJGxNpkUK%X6ZCku;`luv-y_6NLH-o<XgwVN
z8sx7TuDWn*RB7MmEpI&?)>sx|Q<IDB;O}Lj(KnqRfFLgHBHp`55c(cXuS8Nx98umz
z99SwW5SodqnI&-CIjuiCJ7{PIc-5u^W8Vz-Q|*x(5OyT7@E!c)?B0}Yn^?=`Z}krh
z6LiVWUK}opO#o_w#tt0*n&U@%jqWN7FS&fj|FJMFkQ&HRa9t|2{AJ*}Oo*3*>vD!M
zLDK!52g!MG#+WQNb@EU-S0Z|Sk~&!N*;}=F?e<3_yj9;WqLDOQ9qxB%%8P@p*FNsg
z@lon1KlR}SH@<$J@sQ)c2i^A&q}Hx$nYZm`6UW~K-A#sX`L!Dkr=@>NqeZA~s1n_f
z;|JIdP_nc$M>)qEt(Ueplye2B0Bj3D$&9xKTx*22>veFu4thK;$3F(o#}I}l!SS!a
z>lK6}u5DR8{NFb`feJWXogT(Q)CBbe0m@KkbfF(t)H<7g<Z}CeQZcq}sH%+}QPgO2
z_?4&^CC>0x3Ey)r^x1gd)|cw+@2wM#-$=qYvQ|oqfA?Q4Y@PBb<4?ruCt`<2W(;#y
zZNv_95pIF*$k>EUaG#U6sUj-(fzR>MAmzT@G)1i^>h+{cI_M{~!0qhG;s<ChBOa~v
zk0GW%dKi<Qs-1Y?Lg=Y<q0Sa9eyKfi#VBq}<d^Skz}T3esn6KhWg9sNVFw|Zr9J~#
z8o&*^AsoM!gsvs;qkf+i96#lEQ0@L!;XjwQwl7;VIGDzOck3?vzVO>z$3M^5zVnM1
z&8W;ed((Kv7k?foacY~sp2{o_<ayx7Twxu^*9q}@kgsQ$cC)8Vh^8I(=o0HY_D>d#
z0MrZKe?{o_c;6J^qL?~GsJU4=6K5?m{!Olzap?SU#|M$fn?3T~-(;BWs8DOfK6)>2
d4^Mi})``3rUC$mp*`t;{>e!<Xd-PzB{|{=ol4Ae>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_impl.hpp.54205731D4AB5FA8.idx b/.cache/clangd/index/device_elementwise_impl.hpp.54205731D4AB5FA8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b152cb6b1e06738df4001a06f593cc411673c7c1
GIT binary patch
literal 7054
zcmai3d0f*+7S9_3hVUaqPzaDf!Vl!Wl|!&nZUg}l6x&*171t_stw=>E%6f29wSEw#
zK-Hqa0#ZOcP^^mA+PdD?^64rqRHe2RwbH}d9YUTH`$ypuW`5ti_nUe1u6Z*_hzbuc
zuoMVF7DZ)eq^(R93Iu}j^gnauig!YYKrl%p5M)<(C-uJD>%m>G@86XhZEt%7!e{La
zDL82~Z@t~Erk&@?Ttl}d&cBj&`bFOU%(j}TvP)??>q5U)t<H}8ugZ7s#ZPV1U6;Il
zQhD9P>Sxkgx@6*W3(3CPvxd*5J<-qSuU|VWew`%lO3!-Tv+E~!JzM=rWX%a9Q@sTX
z{kNoLJYPOJ_V0I1H|o@0KYsT8*E3H{Ty_7EBGltu$)?VW^|49QVh{FzTix7uw|Tee
zHT%VW_Xn5!P`+VK$+5vB{q93=uDJI&*>kIYdR_h;#XIjzJUniGl1_QUjxAG_d3HN>
z7yg=S-!aPv+`CeupG7TiNPlv-?mtlz=U+bHr2KpG?7SHjgXQzq^NpWGEGswItk^j7
z7mFPC->N%XW(?Gw-uKx%8Geh>Ym8p_l_&Pf{-v|DDfXW2vz2LQI(*D6)B4}Mwy3cz
z`QkuaN@(*|+x%yuk9*u~qvl++k9r{({&1*sR$*+n&zd`aT%q2{?C?1WuLnE2#^~>V
z6!^{C?>0<{yDPi-Rz~;J#O+rsE>?<Vi!bkNxE6exYe-&LXXK*Sxaigu#g`K{y)oe*
z2Fu)M-qm?6W1ufbRggEm<M*j45pgrKU4l|ScmE>0T6rO(?uhD#waLp8{`hC&*R@YQ
zI?lgve~|fs(|4^^Ef)gMyPe8SuKB1ix~2a;%RbL_reA@%WrOn)w^Laj;j2YAKlRKy
zBedvVcIBw~+A~IBF;!o^@kf~W^NY%?pDfOHc4vP05A*-MSLmB=KG%79?Red(M!Htq
z3&X6y%a^i#H>^J@i5Fz9&iI=fbu8g8-*4ONGUJGp7+DJVZA7?@m}xozb+eH_U9zll
zQticdV`4~I?tt_T@R|-;xZ}5VBLAz{B_U370bd8!b>ODykVTg8)kaR$-e*8e=_tA4
zT)DbIa!^6+qr&}fl;{y1N-h9{0x;2Z$PyC$+rtMviz_9>g-+540gVu?>5yf{g~4CX
zE%{-ggqTy7W{@_6g{DK6*^V7vZFkN^7!zGO(In|4H+6&Lg;zK2tsPh#AtfS8ZX(hq
zqR@26k|tcz^;9%1*MJyMmK<@8rMf}#`-7Xqt-dyYA|*0~fNujy8;sX<$g(Q;!VSgR
z(;Fnjk&4*>QyU;k(;-W?`)l1d9%W6LNKEJ?*5=kKb%W&9L8<S*9=tO|LR{$xdqBAd
zJTx7$6q;S)%FiroGA3qpgdk~<jk-Z{>6+Uo=LcVK)+0g-0WTjXm#Z6rJ^^ZT;9*l_
zkV`?x+hxW?O5cx@#wpbezQ5^3v&+h#_urmKCQ!1{Txp|jki6R@-`l}2G0T96DLGIa
zs8lydZZT-;TtEEt?^0q+$$N=vFY(rN$g)qcdePiv#LATD&?Odus0gH*4p~}{Ul10#
z4=F}feG4SFsH#UhWNCX4e{9F2_tK5YcskJ@BHcqQH660Fukq$CmF=R6l2DccA}k<M
zO@}N8YlA=N7agB1Ay!nDCQvkii>5=Cqfg$7xwawUkRIXaD6L@73dWiaS<cn&elELx
z-@t(AQ<gk&o<iLq`F!z;Uw^uKEk&x1vwWhTPvrb)hb;eFFL<2)MNPRe(W4_Ql`eHr
zH%NYX?#GJYsD#%g#F>g&54QDG%+U^6hDEBU8MF0EjmbDVf~VBeUELu0kFBRFJXU{K
zW<VxV@-*=@mAXOlinPU<))QFP0I?=lpf_QzpoGyxfdsuHQ<Q;W`??9oM{<$&=$srs
zUpYSjT@BL+WvT$K0xZ!7#F}J+S;EkxsA_42XRn0g-8eU6jA@Sd=lqRvi5wp+54K@5
zCB`Stc{vlG6ql65P)c;lEQXf4F7-xLbNni9l`%s(VL8#50XRN4EH@f68^>3Hv<gg7
zXpY|n(p}oP3#46OiV1?_Pk{6U!-n6-q#WSIDZuexg5Q@Ah>T_@hs|_ce{`mV6w4HM
zloRLG5vwyPYMSHM_^ydUU*q`G;By)xaV*)*HqrC?gI+|)_kvk3$S{V4{5nEEw&;a6
z)At{K>~~2#CZp@3=K%_mcTy<iZ7gjRD29;dmAnl@_rU5NDA7NJ{5^=c2T_RK;ub}C
zHLWV8v-N>XAGjew$Pa?mAbZg(vcE*ne4dY*(YlWjMn;@+4}9=NKfAvPQjHd5fshZM
z86JynFx6`4{4%}NAp;@*8R0%75={6)zKw8guVVU@9iz7s-F7lg8{4*%Neu5N+<xu%
zJ4tjWnaSt}2zNl6evohnwecYmbBM$;{w~6GF*>@wNUYfwQMH^VDIx<$Hyc?SDAD2^
zHMDatc%z{?8ktUazzY?{QLFo|Bh%6F9A8e1%841KMUJl_${ON;l>%<VgyXk>WD8iK
zzl$}SF$xlKwL0MAH^MUYI6lfD$_p*P@iQD|XrX9_Xf3ory1)V*oVvR-MGK`$Q(q$7
z>8TW3L9!KOOp4KMAElVBAJ(0bCjF2GPZQB^B1WiSOd3S2E&!N116>27ar>H+S~0JA
zSMOB%b~iEUCZ_myvdLM`lzlQgN=UXy_Cb;7AM{UuYmzN5<fxSt1>lHcaCBdi-2z)?
zg?vJjyF1f*1({Gml&EE#hbGd|A_V2b+(V;&r1ZeAIX=%fFOo(5DeyT35$FTyC+i(Q
zuKDTBKLrDl=_I3d3T|Rppiu2@2FYenFzt?3E4Zv|Rjb*O|11ipYET#|tWX-PUL<05
z)q-Dhyp6Gq1w(EoZkA{Uj?an7NyOE0d@)Qfh6tRO<4eK46nq)30<$WxX1D=78o&<)
z=lCu#>jG<BA;%Aa=@3{kr!OSa3rPfX@?zp%One!xB;l1L7R|x&JBZH?@*3jop1%GZ
z;pGcf93STr=ZWU$_;{CiEmQ>ZB2Y0Qi$PxeDz+~M7e+4ud5Jc?1nf(+>7^hq)uxw$
zyi6O{gJC@wV;R8Fn$oZVjG0mzK-K^Xh8w}K5sWdSIU2ySMo=)^4^#SKDpT43Oc~I|
z1w>vzR7`1wL|*tRwl5?uj9x_KMcVWtVqc_9FDCM0ZF&ijmuTZ{gw`7`GZDvE68d=s
zyHyfzG!n;eC-kF5-%i}NYtwfSYF90K72&HG#v+?WK{asIAV))qHKjI2!MaZmw;xPM
z7SSl^0c8(NLAS%4M}u{w0B4@R&}5+<bDG5_i(jJB89E!j-}Hx{jN^m6g8Xp=ib8{;
z848nxO=Kv}E6$${7w;9Xg=VQ{IWSs+DnSbsflU!O;<849&b1gOGh70$CEB<iOzOcx
z+rO+Dz)strtQx_N`A;*5n!yNjHOF@V*8wJ|9ggpZfPM(p_OE~eZTu}beG8t9UO;RL
zh$Hi-LgHFTCNo?_T#K~vHsZLAcrbscB$F#i0K?nK<n0Wjd1yKuIS)nIzF)efD`ZXS
zVCan{i&s1!v*ajyeE7tVA=U$97BOfJU9tK|h4XU!a?9n;tkhg-xl#)q1j~cq%yf4Q
zERQja$1$35N6unwNZypUX2{>fU(V{|+?d=XtRgwS1g4ijBr|>)IF^AI!@Iz27bqES
z1dm4WXU0AOW+$}Ej$tqz25V-#A~L;*L^5NQ5XTbY#qf3#zMaG|<5UrkD&miL%a!|&
z{&^&K-=DUyiaw{~LC5L_Y3Q{OSqrhnjk97nW&S!N`E|Mzw}ERL_%qG5L3A6;W%xG8
zZfoz%x54K&1mjM<?cb60R}0p-&@50xCe#pXl!&HRGNFN3<5{-qhg^fbIp3_LqI7_D
z2L#{@HNUiOyk%CZpoA#zsFy)p8+<3wYF4x%b(c<(&`A<ebRmC;%sxbt5MwbS6synB
z%$jR`*Un^HejSMFzzWY$9DfFU&Oi(bv3pJFrY|MFVSmzJT}ZDY_~fr2kGZ<$>?6AI
z8N0)EX#M6rBY@?eXzyq(lxmsEqkE%bXgrOc@KN#>@2T;pYf>udq^%%t1$(SJILbI-
zFH<Qx8y#Wv0su2E8ix`_&-*w6p7$wX^stX}qMuR1=n)?Yo!{kjnw32M7d69QpNJ0l
zv~SP7{4kPgu!Wekuv!pr2k3G>0?9|()zauy0ZQE^5nVcy+&^<niysM8jqfjdR!*IF
z2t-3*%%W%*M8n#+lZZNrF}eyzT}O0?;iKoS&ua;IGLn-1dU2>fx>JqbU7&h6sz>`h
z#0?cY;`Pb~Vvoi>X}`wPcmIq%fAXuIhWK|4!$lj$^sBQEs`6{@`&_0Hwh{9-Vu?!;
z^6kX5oyZWQqtkkQ<m!aU*8@I15QTn@1vRzv=-ml2ezBr))^FcrWR9u)i-D6qNQjfu
zyU;OocKJruTxaE(wbV*ZDkt}r%Zv45K6$#4<EN;mxGx*Z{du)R`D0q7fL$ecw{70)
z(r$Uym*Z=}u9o5J?hkuAl2W$nbG(C*gMyXA-bUUEmZ}pKiOv{T9G|2}(n9G*=?b=M
zRDe~5cFkA~*45yOk;L(x;MfVis4k8_3QkAC5An@`qB(o7Is(mA;SS-iVVK|AcvGgA
zWpRV*<gd@?C<pbUk?Sc`KCSK@aza^kpX-ntiaSrA-^GyOW*LmB?|#`Hm(AZAxlx~b
z^@4=sFM!_#c40E$q5E{4ky$BatRVUo#Dt~ny+nU64PMRl<2{$dE#8Wz-E_6*L3N&~
z1xp+{!X}VxVwX5#&3zh%^rO#B%bZq*2hkDolzBerq>rl)fA0R7%67~(J&`_mxV|`U
zvRbi!b)WGp@5~z0-%c2f$%sBf<7DJ|k6B=v>$DiW-y8m|wb^ViqnwJpz;eOMD<ve;
zoo3|jkG-1@zWES~tD`vqM0>y*3uKNz3qEIICOVN|MOu1lD7{7wZ8rFN!jEg`0_JNx
z9}7@_cz_h(T(~y~Fe7EISeUtL<aY?wNTdEE{E~T|-xUzIczt^ZyGb^-Dr=>{8maUL
XGWKZ49&Op91AA1mM|<{Y!yf++Y`?f0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_normalization.hpp.2F6431A98ECDBEBB.idx b/.cache/clangd/index/device_elementwise_normalization.hpp.2F6431A98ECDBEBB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..98110b189d7f8f53264cefe176db63051d344318
GIT binary patch
literal 1160
zcmWIYbaQK9VPJ4h@vO*AElC7283cj2xTGkvkdc8w2`Er8XYU06Y$ij2qu&=-Y%tHg
z@Q+_fW7?yIVMm;8bq%erZtzr7iCinXzj}N2+qCX&?M{jUpX!W1KT~~mWxMVxb2gLU
z0O?nZpL1r}FP}MaL-v(dx63-^X9T?o2t3d&^zC5y&D4`sbHXo#cBY9I#5O&z=H%KE
z&&}@7rt_JpG_XXtWWLU>rB3P}q<-IOdAVi6yM0Rnm#IxikzsRsyDRYd6?T)X<D1#5
zB&;ViA8E9*UL<=~@|(=_#9g06*<QObM{x2v{^wYm6tqXwxb9$b{$lrOPUrPD^*0%E
zy-i%t`;{wr1G{FqY}VwT9m}UwUEs^@X%E%(<&yWRKWM*x^7j24oKplgxv-|Zj5stW
zLa;e|gW+YLyg3Cei;Z5~h;f>ezrs;y;m$Kc3O1_u&z?M#lJfq4<<a724{9>c{NLd?
z)9hN}ol{&HHXFS+J=pV7a4GZsoUNU%UsQ^>ZL!Jz{@J-m?LsBD&)>u&Tsz#Q_cBj^
z-1?NExH30sJ}?FtSiU878v7IoFtKrhxC{*3TAEtc3>*-SM~~$4Hyb?Pax<}TFmRi(
znW=*b22LhmtT8ZhGkQv#p6NTIPFj?Sos)rEQc_Y8Ou+O9X3Wj5-2Hr!023?Bco`NM
zahUX>U7Rc5?z_*$#>5OXUYuQA6-*$EXDCWdEB*qq3+hfrPO$r68e#5Z<N~|PQ>CnX
z;x6Ypn5&q%z^;L*=HZsLmvx7^iHBR+OxX^`gV8_>KrV!73GowBe=*U9m5qm6Mp;G^
zrV411o~#LiC#NijBqy&dkHk|@RzTt@Dl5WyvD+<kV$@WW8F{#+Wu!G=#-;YLsx}A}
zI{}5o^~J4WPN_Kj@JSWR#j{L6%avu}s)626R)+Jsy_j3x_W!p4xmQcZ5N0G$i>a&)
zf+wdXha@MjC6B~Y&{9C+DQYRgd25ZW0*=m2m=1K9mW-A<EKXn!Wa0wH9xM`gxXne(
zjbJhK)PVI1JHL20GtjSs($Wa0XbEab!%PBd;WQV9X(>w0Nd%P=u<&4jS;SCWlw7P2
gN}=5ETPnkH=Ol14^YC&paDjv%fEUJKVBkS80e1hHHvj+t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_normalization_f16_instance.cpp.B8CDDDF357D7D999.idx b/.cache/clangd/index/device_elementwise_normalization_f16_instance.cpp.B8CDDDF357D7D999.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a761d4c1dd4dbe828d78dc26a536fa1726370a89
GIT binary patch
literal 3896
zcmZwJc~}!?9>DQ%Cjki{hq*I3CqbC7WDo+#AP&kP1ZY4At7s`|5mc^%Ptmrrl?M^5
zVpRmKU8|tB)LIsiT@S=X>H*@S@SxUGDn~t8uWChi%=3)={#c%eNxttp&nMw!@@A4^
zqNDXZ7Axw5n3CMA!lm{s7R#OfEhx<YaE~2}wT#PRmDHb1zOm{?u-mmao|&C-*KY_O
zl?@N3XV#q?Yw6lg)F>+TWcJ!0h9_T&sg!!uZ}6K)o2cuT5MIyCMi-ggQSxiQooDu-
znz~;fO!9kv>A3f>bba{ASB4(Urz5dv;=$IFsZYIUri`xsxVo3@nb`II+DTPNQ^Y0g
z^lK8!cvCxiq-R!TX`3Z}<8NJ$w_Hp<_0lkTto23dm!|K#9{c?=oO9haHf4~THh$u_
z#rIQlpO59QYMotjr}k)aZg|GZOUK7tvRkUD*n^v|B>w9B>~v{E%F6S%hR&@i*xT;c
zzk6?U+nniR)2hr4D^5k<uNki|8vmsHfg6<-^i!*I^`_$V)P)nxF?_Ex`~AmmrKZ*#
zSirKBclxfh94~o1G8sDm?C!X5ULSgbLy3@;V_i;lezj=(!(=zr{Fx<v^8^+Dh-$ZM
z_xOCI+G%if%f71Z_KkbK%V}-s2)#PP{CRVSlMi{cyV#AK*K0n!x{P;iqO0%XP(@c&
z#hEI+{6aL(?;)r6g3Np~F+9He&6Tu)yoOfuhJ;9Uaow&9d7ftNL+wNY+UJqf<ki0^
z=-C=>z`&ZuiLqIp`~ynP2B%f&4Tet#XT)-H*Wo+=`{Y5m=U>OWhpz0u;a;;%No{rM
zw?T|ic~oO#n9?t~FgPbM!!qw_d9Q=L^M__3w&zdf>f5?(Lwe?O$7eTdJs0`@vm*?1
zyZ!@uCGtsWnb<9y6L<eoRZif~S&w@+eVY8;p-juD<c0S~NL(y+HK1vG;IXjR!d`YZ
zazwBrlik<c)GnNO;!!^sym)7YT);i}BH~(XP)n5mg+h<SZ?yl)^~xfuGQQpM<48!u
za^1CwZsiWc%A?m)ayY;Db?mP6E7DZ_=Wx=}{2iZN-sh^?GiSRwL@53`=>(_p=L4}z
z3fosiU)negb6jrMmUL)+jrn*)VC%)jAAMC+Tv@!~K<#n=<9JKKF5h9h2K!@iVt)35
ze-HR>IT7>RDeGX`vwJleLkrY-4fZ3q7k%N=(Hq`R^c8hB@@jwl`-TX|(Xzmnl#O2=
z_LkRCwU?Xz@b)nnU!U)L-fXTPb><JuT&MRjEc6u&Y<3KOgH#MJt2=$Ue9J(ox@9;o
zqHH+tVvwZxZs+j5G~K$C1bOG?laHpKKeuCj*|u%3@6MfkwENVGj-{J6j%zNEM;6B{
zaG{Q@S+dP(Z$tXX!jg}_dV22miBl0?ytX^qmxAo|4->Akmc6k%@KdKNt6**J`)zMO
zvENq+ikg1!aJEe_vn;NTt{m%up0WN2`akJD<oCBfoNbfoX>UhwgZ02nhXzzSzVe!{
za<+}9r`;XhIo1O+-JJcpuG-RqJKHAF(;ki<T<d|E?yOx~JupM!?817Fp7wX}C#(l%
zdeSZB*say)Mx1RE>1m!Lk7GSB)A?D;3LbN&{(>c3>pQPH`{%xIheBm87Cz1oX5K$t
zMt(Rbf*%En<j)6X;wMd&02cs7U=#pFKnQ>$AO%1Xm;^u(;zFPZjY6OZ2_aB~q!1`V
zlMpB(Tm%%6Q3MnbAp(ks6ahtK5&=bwi-96Gih&{~#6S^~VxWjkVxUNH2~Z?P2~Z@2
z1Sk?x0u+f!0u(7O1&Y)t1&WlA0!2znfg&|Yfg;0YK#>_`K#>tLpvXuWP-G?<P!Jpe
z3SvZnf)EH$5E20jVnTo-$K^ng8|6Te6LO%)NjXsDCOJ?PxB@5&qXH-jLID&7sQ`+?
zqyUN%R{}+8R02gwD1o9Rl|WINlt5A8Dxj#0Dxjzc6;M>93MeX*3Mgt^4HUIe4HPw@
z28x<g14V6814V;tfTA&KfTAHZK+%vIplD1QprAMk6x4_U1tn0Rpd<<u)Pw>>i)(?R
zHEMyPCA2`%l3Ji>O<JJna2-%|MjcRegbpY=QU?^BNe2{tfZp(KJGJP8^dX?a^bw#U
z^>abR>f@PW_B>mzb&sFgNDb2sk<3ZVj&H}c?(|b9=@aztF(<iL408-5Q<3HOSTCLw
z4%*mTphVT4-qKxmZ`yQvZ^rlr=G9#+0!#o(hzUW7FcBy*CI%(JB%q|26qF2;fkH3@
zlpK?TQeX;DN=yk#g{eTPF*PU+rU8XwC@3wa1*OAup!B|aJ+o4Av$n1F_S^i0jf=%2
zz$1+L`ZULTXvVgNIncux2jekkmp>v!Aw8P~^suX&t2gsazBBLX?sf64M?L6uW%G4l
zUD*OqLbec;h%EvoW{W{d*b-1uwiJ|%EdzzH5m0ir9F&5s0HtIrL8;g(P-?asl!mPV
zg|bmlTDBIHj;(uF`wlNZ+t*&yORtT4fV+XYYLDD4TbnjiUri7DaC}^uwP6;TzNp{c
zq=1V$S)ax%H+@mhGsH4=);xPWL)w3ymICbq<?n{uIEU_^y3M7B^=!SG8D<`BxYj!y
z<^}HP9iBMxd@AX)F)vKYoW<Pucyg^bJj~hj|C8d!G4GlFoKON#LP`iqM2SF&DKRJs
zB>^R+q@ZM!3=~2kpyZSsl!8)#Qc_A#DoO=PO{qa?C=Do-LP2RMEhrtO1Er7F$1n@T
z%FoJK`u0AsAV0Gplzz0c?0Yr2-gRivdi#`xscEbf<|dKFN}YPTaO!E=)YHt|4{}z}
z8M`db&fdY%iOq3#ndZiGjd3TV)b8AvIUYXKJ!kNPLqfyCXGakJUcN>}5GwtDquxT1
lScb@zDvee*6VnF-;<M;2$?q?7y}aj9>9*0P$oV!4{s+fO8yo-t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_normalization_impl.hpp.D9648470B11E62DC.idx b/.cache/clangd/index/device_elementwise_normalization_impl.hpp.D9648470B11E62DC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c9bc573c0b4b008a69763a0471a60487937ac6db
GIT binary patch
literal 11014
zcmd^FcT`o!*1xmQh;TVvu5b~E^dcTQQUn1}RFGJrf|}%MBC*6CR5X^T5KBO?5CuV{
zh#-neF@h8cf+8&v)YxKdi6|ODEZEUE$QkCm<yv3<`rbc}wY>ZO+?zT3?Af#1Fg_?C
z;H|ApHgtN>viUQ@L$ORI>qY)84qr5TD*-OLGTE}-Ws^cb`NLD${z>f*Dfm)Lrv3Is
z<Fqw4>-ukhb}C)>b#4FXD+M0;iwl3s8{YBoYw!HOtgbsY7mk=UeCG97-)~MkDPBCi
zGutWdpiXee&F0<P9z?3YnY3ARROZ{~AALQ+R%7ptz_Yuu7RHpNeLW(*Py34jmLXrQ
z4vzQo*x%86$=o3~wtRct$Jk2P>!1B;**Q<|lNRowz4nCd8x^*@VwB;Y@Y+FcE(87l
zx}|qK)@1I{$lhV05kHvU<n?;UBm6d;`1aEcd4c8+Y-MkJ&D~CJ85?kPyvMQ~KZPUf
zS{JL0X18bku`+dc#(~Vfm6>M6OIo}(`@adW{w<{GfuhMIXyvrhz$v*S-^`uAZtFC2
zub@o%HwG?-_b+$!eHIe`?quUmQ>VvfNpGE}UKnz9%=Hz2wHM4=HqN2yu;Gy-oAXB}
zt4*8OkvOH%xH;j}(t%}$t$7zrr=%+W@}9Noi|0Gdm*mW@Uf68BeyqkB=fbmdhXgJt
z%IVee>(w3E&OP-~<0{7_wkO{gc5nUWa>I@#ldt{a6>ueWx%F?$3Zn9dPR<$f;Q8WF
zi+=p`;+n=^8=tHRdwccXx~z9ML;ANCuD-k?`KL`C&BynxJlVF_J;Gyn{jSzWW+z{)
zy8KWl_Fsyuul-kCpVoHov#$nj^l0ASGfZd2@h2ATM#Fclca74RHB&iYd%fAb3)S;_
z*N#(d4czsnxU$*fiAB)SaZbHXj(h53c(gdxCZczT&%LYdeRH!`T(@6R)b?O?LCTfC
zzyCT}ZC=HXZxZuv>^S<YcG6E5?}s09*U~%Gu2nEsc{gOXQm-vMzV4@QpYAEp{%v;h
z!9Eu!Ww@s-7&~eG<_e9}2|MS9ZP<~m-16EfR==Ok4gQ_}@T9eiM=kLVd-18)vCoUM
z2G&`-f0urGcSFRQJCPTDF4>fNR=zlUqDSqBk_O-9CyL&D9L!&wH#6<Kx+Bvyf2mlX
z7&4=_=v0#3Tb-S8#pBMU&qk>($+g81H?!{+`D@PGyR$v^a8T<ml~()X-P<=VNDFBH
zJ|P}o((5tiK!w{Y%l<QF|NQNk^kZl5)wqR)3|ON->t*9F!e{HI?$4-8aBA>6_;J}#
z^>foT&JF)D(KPzN)XB!L?ig7I&3twIY}s<7!XGOa6~|8>ts9ba$5&QSHfOR<^`qT-
zo++B!&4(l`tTl1|Au9bl>!_&UiQ7k|dAENt?w0DV%bwX^%<vqeZ##a(rw4W@;tm#F
z*RuDrTeE9NQ{nkb3ljHk-1o%6p)t!XC4NHi4-VC9e8;9<ymS=?RKM1{XgO!R$3-po
z`|$&UU1bw;{=BP`TXphHV#zASqJueY=<f5_ufw7}hdjOOv9HuL()#$p)`t@stm_WW
zKRoY~Pa|&^DiTK5wu~-25c9~v$*iqhuXR9VeE&O7mL=?aXf>@kz2(|<&pWZTMLNev
zKdFBa(ot0xxc%II&5@7%#z*yCTil~`()gQ}gAFfC>lkElbVy3r=83O9t4J8VZT)hS
z)p3IxXN86hihN}f_OMCwwICa^_=4KA6vfl^AE~dJ6fw$0z2e6>-lyd7bEk8^pR>E&
z>rPTc<@O_sB0Ibb3ZKF|PBy`E>Eh+{zv{0hlgUdbde>XIFIPYx3z^Ui#?2szM5a#S
z7x_^vITq}_Y+;oGJc!5*FueheBB3JMZ(w?Jgrk)jXq(G~*P!_tbVNc$lsdQOWiOY1
zuL(R6d4S{(kgZ6%Mbht-ZJXt6rGNqCL6701$KWgyDq=M|z3<1bj_+51I}vFD*Cy~6
z2^Hzr7VPL$veiWs49SD$$mi(3|DbFk@0NpC^{YG`=uOz`k#aq<6bTivy*K%*#*oTt
z1^APPya4_M=!=Ak*oS_8!^?QZBn7w<k*A>dl$foHP!YG{6f^U>#v}y{B_j90_#X5X
z2^H~eITUNW$s$S<<V2T7ByU8$MM6b}X^$z8JB)awfWc&yzro~ha1u$k$n%>)i(K#j
zsRKR9D$AA29p8UYb`V$gL$5UpE-Ju_tnd){hhQNRDl#%L%z0W{=uidd5Ruzpd>gDq
zLPf@AqrI<JsBS30n259j-wK8zp&}DznrH9WaD9dX{D{aMFup@xxQld)Bz&9l>lr(9
zP0%8$u0`^-$V?<uBy?i1?TG{D!xi8|R{0Zj|0KrmB2;AIgCA#GuAOpP6Ew*x=aBpy
z>L(H^vZ&{%dCgp>F8#?Wcft5B*ouUTERox`o~Rw!Uk~)jDzgk{nZN&_?C(%csD8i`
z;`MH1g*MP`0|Svzk=2n?OGh<dS*!#-vO*a)FT<`Pp(5+ICFcER*Bqk&U$V+`(0Wc*
z=^|7lc6?Z5cmZlqf`O$>=m13r^b!dbiO;^7;-c|Vt^g~t$~91310#`iiyV&FmvF!>
zxi|D6B2l0o1xk@nksS-$cdi;!l&1!2W-?()&nd$D519srk7P0=pdEQv*7;SPx~hPI
zWQ|*3bPFs+Lf6QR>PVb6Xa0A3(2KnCQsbre?>{IzFYWd6@V$3U6<|bGXa~J^=qHkH
zk$t_g?Q)KeSA!m8g$<y&0dz$|MG72Z)Siau7wUm75eYXAw}1aZ*~Na#AFXY1nxO;g
zgk6J_H6)UBk#3Q|(JL*UKcB6DVMLDyVDf;Rnl3^`N>3H7+Y%ZWqyTRsaubYi5})rP
zRHVkpz5Vi`%zjExk_Q!V8U>t@NT|rULqqIuB`v+IfFWd+C!qI)tkOlO$fcXDSNHbI
z^6%Q;98QtL^%A-W6=^XDn{CjV)~JBtWR+*2_l&I4MX1Qr#XXJJ*q;nmf(Fs05-TgQ
zu}G-Mvlne)4<80JDM3X<(lAcL`XZqsZ^sl44c65;s0j)p(tzX*sE<gf$h)Y)tBON<
z58`E>rZT}oX<`5VL#DwY*&q|*C5vXvTx?}3lc|-uoB0gSDZ{*wfb0`c;2+_J!<T&0
zx_}pUBmLc|KaI4!a0uxilHxLCQ-=I0pQa#=7xn|c9}MC}U*?5+aIA-c6wAfLlqb57
zmUGn3Iq@!6d0vP>d<4=NCPqSD*o*=;qfzt;ys!lYY?0y=6r6%SWBdX%xBvw(T!;o2
zN^uRct3fW*^N3=|q#ju@+<^QW&<JW?UdRPyF7&2~^Fk#kE2a1>*qsFz%BL9?cp)43
zZ0JK7axu@+H>P<O<YIQEXPOtvu~|8`=?+O^wk3jHsg?;waDEZom)%ik!?gY7uYNZn
z#!o{1l8_(U#SG+>fdUxLME03zD8))KSreVhfG-0JrgK;3raBv3%2$7rb*w+aLO!RT
z&l%7IVS3M|P-E#$Kg0_bG7EWEu_2RL_#G{N?Aa*)K%VA)qG}o_)4-HIkrz&Y#R+Ck
zxmXcU4QNH6CVyWEFjM@7e1Ag$G@bB5EAnkc0kl%#g-hUj2?E$-zUTbD=SEP|@xms~
zZxh4Rd(`D(J;5GShg|B=Fy=0e$f6Os)AUa^5BijYBYRL4D67Dd;TllZNO3(V>!r8>
zlno5i`h^(k5-2Y*cjATLLHRqwa<QtRhN4vs+0$y^tHF$FAQx+!Zo%pLuU1Bt)Dpq{
zNVOjgpk6^77^$)uc9{Og+Za|=0v1}h!7ZFWjj32@<NVsVArw1)GCi>W!Ea-<u~5YY
zR&ha;frVObXe~FK;W}<;ofMzphMnO882<w2d4U_q@D<Mc3g^r4HEzH)&Wqt^+`wm?
z55q4xzn4rgEWG8M-f}LCABp`Uakt`Fh{9e`*qiZV@X#16(RV9$+=`tUe;XdO4G*Tc
z|A^Zg)eJxL=!u0hoa-6RjWV!sg|ofF*)e>L>wAr}VfZCy`%=n}#I})GA|HdDV*c~N
zw%Zur&G(f2^77AB<oLCLdK)Nc#X}2M;yxL`XMi5{0$ykU#|9Wcv0N-?saJSfe<cgm
z{&ES~>=HDx1a-$>ER-U@QZ$&J5YOLt_n6<RZ%Q6r0LlU|p(_qFc~m!7E%Xubf@;iH
zV+Bono!Gw`4`a9n^EFa_4S6%{Ams<v;2?%;F<&c{uf=w?Qu$gOSSyvU!+f1oz7D(9
zVGp)mJ?87B{CezMkNZpc<LhyV6fdgB;Zhun>hWeo|Az(x4Y}9>u;}pz^8Et^u`u-v
z`94EIEKEHC^(UaBr3g>Xs`^t<(ZkG>0!+OfRJ1nZNoAq_3{<q};DwEx-$rgE3v2P5
zUp&KuzwzY~PoInaV6$jzLV`ddC=yvbmWvGv^$VJoNiyo}PS}gBMfPja#}mY+An^Qd
z?_Hk{U#vyeL#jk%!h%W?QYA_8PNdo?#mUGu8TnH8;)P^1AsKzia4K?5mGaY&Ynl}2
zAXN@BVe)xMl_$mdNR=<e<%lmQt(w?M@<IjTE2Ov*@s$kI4y-3H6ainvQe|i28hHXj
zsr&On6^y8YFBq<YQ8n-t!}Tz-UMk-JBN~{zT<q(p;b~u|PAmd^GN{->1>jl$-VC1r
zgA**?cPe641@??z16DOsTn|?DQrrMm4Ghc0?y*~uz45E#les$LDw~n@W;BqQZVR&B
z!Z2+!iN?o(KL&<OgHo_9WwEEzbUr5_i1Dkyt4f-zYrwY#K4$!S@Tr%|H-J|Ilb4Il
zEY&zzV|dy88OqJX>npKFCDx|rG-Qm9O>>-24;4?cywiH*&R%5$a@rupQD77$-Rx+v
zi3SIz(MB-ZDCNh1QH&JFf>Ep#ZvvxDQhXb<Zi6n}WS&$#22Ef>%Osxk-dcA+m$v0R
z>3_BEf-Y^1c~VVlHG?iKzj+ePweEo~4gaLd!j@@Rplt;&?7@zEup7ha*fL$p--|8x
zN^u6Z%#h+tY?;Y0)r+K=OyD!6X{HJstJqe{#UTRS8X7N1IoCBoU;*Vk@;#4&S)^=1
zzAY%2MamnXegjNsxy=iaoL?k2fknPp&M%hXiJ|eI9olnsAn}0)tlNNlQy=(Lx7uZD
zj)65{#37?N<Vk&C$}z8bllxj(lAdS-7;gZ9>dcdl&mf1hqO~tce4J$g=R}h^FQj5s
zDmJ6WrBe`cV0M6F2k20*k&Cksy6|kvN|l+GB8v#-1D`K#o{j*2gyDIWS;DU5_J93j
za;DZJ(YkB2!ps}P`NeQksJW;kkqsn|&yt?i8gQ&((M2v!UZ~D=@}ebwKXGB}M<Cq@
z)Q6@EVs#YMfX1<|DjGbaVK6lXF+MoOF+bsj6fjKz8^+HC{aluoI=e9aN-1sva%NoE
zfjP-po@50TFP!I0&$EK7^8onm!GUy-$ZKPdEIf$fWhIWs_8FwVQ|pYj`pMwT6xjo+
zJ<^k(0R|bYU+Fvm2ANWv1x8um&VGMC80}|x*~weCSKzZ}dhkLt7)66Q6O03+I4N!d
zttQZ6Z@33r?!o=2#+`3qnZ+=z0Esh|0AIq6q+FbQ(H*31CX$QuF1m+w-bH+(5%@;X
zXR4hC<#}oA(KQ95E5>}xt!vrx&hsN)=!g4hc26OMaJ_INIuPQ8rFu(^=nSee0DG*L
zrtuBPV}lfLL^>OhAw8VD5Q7Y3$iT5XZAXD+6xcEqqrownnVA$uV7XDsj{(aVDUJop
zSSj8FmYbybHmGhxA8CL#ZGygRH|~Jy4)mdcurolb?m{1Wt~vv>su}u71GMTM^kD(I
znA0id3|S0G!?tPIh3)Ph?7j#4Fr1EU)1~~q*mkcJXJFe5DbB>UnGDC=A9iup%tdeh
z>zxV>IK<n!#-nr$68pHg;^9!!rvJI7G45P*TV(0VoPP}FbaG5i^e*6cu?kQw&X4K&
zqlbXlxNCUa&A4TIV(`xy_erihg$+(&E1Cr3pHH3gy|&Q>lIuFap#!|=Hg)2W9WaUE
zT+Sd@n)`A&k6g~5=E8*95n1OqthOgQ=3rG0Hf0W5j#cH@ltvpGWQdNP)9WCsr7>|O
zLs!$WHTnGnP$WpBR@V%hzDn}T-1V3BV~mMlInpmjPAn3fK>8;bPG23f<rl?(;s0pf
zjG<SJ{VQJ))Bc0%2PT~xBtJISfmksa4N68K^wEcwmG&4rKRrN$7j`1koydkYJ*j9=
zs&odKh6bfcaXvE5M>cdJlK+uoKJs9=0_j&EQz`CSfgBmGMEaFdekJN#$@qs|2ac^u
zp1Om)+G{X+4L0=9k$E3hm9br+;TX%`Uj@)jBE1f>DU#|_1U^MDk~&&vFBDwFdK=oF
zlIM{&p^l|#xp;v<cahSba`D=LJ@G$1x6=;|I>3ziVLoS<&$&{+BP|A}IL7s6@$(qx
zc8pDeI``Bphh;~ag?h-}Uq`UVmY{(pD2N%h1V)#@WcH$^U{ES8OD=--MR241g1=W3
z=x=<S{lUDm(Bnwr%Ciqf5XLF2dy18xbdE?Ww61Fprb+{{X+T40#Z)X_gAki0fg(w2
z)2_=9YRlrk9afGu^4meub`lz&geEdC%s^jeplJ+eqAxQUE|o=(KcD{E{{Ow^q1QF!
zRpLPr&t~az@zRFA%JG^DpB?so()58j%NPCCI{D3tv4oMx=_PXd)Gtmb2LAMibLMn%
zHjhHDqbx>N+N^dNaP;HMe+&baB^n1>e!d|mD}15#g&p0Fll#Xm&<aprM|A$b16%dI
zoe{~+9%tp9V=nh<Jb=bnEYx7V8muqHt~J=5;aaR$E9KW>*IFsR4x86uOD12BP3p0k
z6#LZUK~g-W9)BstE9&ug6w{+ZB5>Dr7=6`WefoYk{u85w53K&{&9)2EPR7;{8=Ob#
z=aCO>FL<E^ska~>+MG2`nq%-{X>#XG&_&ZlkKP8Hy|`qOo!RL%<o8#ZuW~wa8e3xJ
zsiovj39Ju+Ywz-5CMLC;+vxC%IPfXpPk|9VUA%A>lxNvwLoVLVQFpoE^x*1G6Q*t>
zS!|{DN@Lc0)FJIUWK54*=f#XkHXCqsUV2!Ufh&CsjVZ)1T~~WFT3%|3nz--0fQbGj
zNT-BcRCiaOC16s*!ocPAmkiXFe01@Du9msw`RmZ<hYZjMu5;TU(|+q6ZAW6dot)E7
z&XpcmQp<2oyBKbA?(wvzwtfr|OvM_hSeF$LWmuyO>(T<^u6^-gSL1`}#K#gjgGA1e
zu1II0WQ8-ppJDDJ7q4NdyFA`qT<(--ZuNn(Pn7o_td0J}?Y~|e)1yf|uj^Kr9?55`
zPOt9$x<dQEt`W69(WvVZnrig?b?zbk`|~%ELz;+86Ok*+DmlnB2f4Ddk&8@o8Ghlq
z@Z^Y`(}O<<-89jWQBc=qHhmN?Tt&WD(HItO+L3QN8bhNEPyVh*{V^EQTTIePbAHj>
zSn3eGu$}YU&M@sqIr;nB@Vpee@);#@h2-LyvxwjZv0sbfEMgd5G-F=qFcQj!9n$<$
z^>EE3MB@Y<6v?P#(Q->h0|(71GWtr37foBdt1~Rx)7U7J^<!Uc*jESk)tY^^V_zNF
xSEuf;AgoR5^|tLn$y?sp#lgwZrk|~yy>(xiOSgoqFZ)XV)Rq1?(=RgFzX6~DjYt3h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_scale.hpp.A8B4C8DA3A99ADD2.idx b/.cache/clangd/index/device_elementwise_scale.hpp.A8B4C8DA3A99ADD2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fbd142ba7658c5685117cd4f5ce35f17bf6fe26c
GIT binary patch
literal 1168
zcmWIYbaU%qVPJ4h@vO*AElC7283cj2xTGl4kCB040W$+b#hkqpy|bAN1sc9De{=BO
z7RTQfEP_#Q4mz(6sE$i2d)4M^q{#nP>EHKb;=7}@ZMTRpc|JRRrp9wpV*2If`-~RM
zY${oF{LJRcJLj4wZg{(F?%4&$E<ZT`)A#Ir6^#pZx-vez5zkm({hfLH*O}RA+p;e2
zp0%&2OsCqv_1r_{&o8|Lk{51UoKkUe^X#~goCWXSNbS(Q#!$UH@76U*)#M0HzTQR8
zC(hh{(pdZ9LA}FzX3h(3Ws-#V9uj(^DYDvc3uoQ-j`AC8D?C@SzrB!FacPmsee;E1
z*lV5%8Q*JV@#4@7VAPa3JyTdasesokdDh8OobDD{&3(EmDS`gYDo(|cMGAX*9`7=E
z#?MtNs_rd(;^1G4RPJXF{_k@98#MEFZN+L+DgPYDX>-2YEGl_W{g-iL$2^ww*J0t!
z$1OZk&P0X%Rf=zPidLPkeLHk%ab<4OK49pI7T?qJT<3F5fQgk0$OQskX%=a51`Y^E
zvSWJQ)xVLq1(?`i(khxN)-Y+yt-9xaU2b6#W#Zss;FXY+Py!P$Y4405Z{8hsy3EGJ
z%*ntj$}TDaCKx!Gfbq$|$jdmVv%1lo?cHT=CKjLqT{c}QFo968b5)Y)ylrb<i88SR
zEfAL!R|FF<3yM<Hik|_^g871x6YP4Jy)f4Ug+Z={3Bz2?%msEWOqhpP)=kzQW)u&v
zyn(zWj0aQ0#0B<&r%GA(#9hvHJiNNPx^^&M#BR6DiBVHgX5`_Ol9AGY=}qlpRc#O|
zb^;2EDv9dBjH)>N@JSWR#k1@@ywb|jaMe7#GRiW@JUtm>m`OkvD$63t$?C}>$;m0p
zA<4-r%foryUd%0T`~O=o1Fh1OHb59BV<KaX;K`cE!g(-1F>--p1m@wj##RAG=O#=C
z`b9-XMI9EvuxJ8?C@79#UIhBZO~wylft;C~9gO$Xfb|PIzj!y$O;UnV(lBA5xq>Rv
zFmr+7tE8&~b4yWbPNEVp;lPY%P+E0kwRF%)30MfjWr~ZEi>*N^m>)zy0Pi-BPg|Gt
X1zlm`=i}yK;D(7aFz_LmJP0NL#W$Fl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_elementwise_scale_impl.hpp.792CB7A86911D632.idx b/.cache/clangd/index/device_elementwise_scale_impl.hpp.792CB7A86911D632.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e13d1a762625996fefd2db3eafe6d05370e164ed
GIT binary patch
literal 6994
zcmai3d03Oz($5?sOLz$}kwr*C-~|ag0){0j_W?l=6tD#queVYQ>Zf@9+>1*=(5kqB
zpt7hJ_sSv))D;mdY86W<YPH(-t6CS_g$gcTdk-YzN&83T3Gci=W-@2aoH;Xd4r8Mu
zBS+Z_1mT}VFG^3Hmm(Ah1pVpn+<9}RrxJl+zo|g5sI)%u#;hB@+?9QI$`;RXbv|Jq
zktH6uX6QVx&9e?Jcd9ugnJ{*K^cJf(-DB6wE~YPsNgCB`welx%^u;?}_fy8_6!yC`
z!#1{StI4wLb0yQP-VRsFhpi3iabJ0_r$qe6&i@!Ip6vPD!+gET^m$V{zpQyxb!_5q
zZ3`74RwpgqoZ5IiXNM<u>THJR!^l@D<tI-XFZ>|*|5A;Xv;^-~-7IUglr$c7yE#O&
zBV)t@+iiL6OSc6nAEk#GKfbl2acRfEZ>xtZyUZ5eTR-D&O>SDlg?l{%gBNyg9pyY_
z{3w{vxakM?IS`U)qIffdpMSb*^>pJYJ?5?#Q-Y0sUQfQd!}S}-2YVZ%hq->O-E+xZ
zwfOLuuJ#XQ^#SIyKe@H9E%`PnnE3IQC7`uTZ;9Lz7Cils$D_juGo8|2oK@}{+@@LS
ztUBWSc-PTg9^apA*>or+`c%Z|v<s`6Km46Uezj9grfsOc{wOk`;L?&cw=AE1Gx*fy
z8SP(ed3|JE_`>_DuBmsD+|}wmg<oVPZQpY9qD$RfrTlhrdEu7rq7NqCeiaL6)OG_3
zllPA&R_0w@S<R|K&g1yXqFV{M9rIHkm&tlE|2?rjSg8)b(Rz52$)o!pSgo;AKk82L
zpXC~LPxIYxk{h9$<1Xx*aX)nT=A&MVJ8N=&zvy(gXQgJtpsCwG8?UVGSCV?*)1b@w
zv$|zZoTjxMynO8+lOpbgl?ta6d_LPxo-j4`KW!(xTguy}zD@XQ_XMBUALTqt7k<&+
zygqly6ZuDjhu%)|x$gPXPs=;ks=w`wFPr@A%;lO95kJqXH-7NLst4KEqc&)V4&Nb|
z`)T^enKbl-Kc60*pMQIXF)^|i@aw^RJ;?M#Km)P&r(e$QGZTK#aFG%VD)Jm8&w<wy
z6p=_feY98kyJ9j>F5r)Y<T%*r$vcsZ=AG&#@oyx=Rw3X!K->ZS^#nz%*5*7JQ+GMe
zoLJF;!bRbZeGg<S%7|~${9(_<#ETAc2vmo_S5HvH>A<ye`>|i!8WW+NfLEF*<$Vu<
z838&G;6Z(4b<@(0<z-^xLv<Vlo1>u86VwrKZo0xmTWTaFCe(^RNuW>P1KC6V7;|XY
z@~U6O#GPuW28U|!))N#Nem(g|cR|-BYhpx2HiGF!kmw1Dg#Od!imCgv=@KHPgPbOg
zr-{Fwph*0vd+kO;8upkIYbufjqAZZ>35q1==Uy%O`Q=e_VopUeL6iyhdV(S|bF-us
zPpofCi7^$KEt&1v_dxa{YxBxI1tWJ@5gBE>40LhqdmuYwud+?k|4Xfu^rLJ~iKlnp
z1KDfKtIo&C51tkiH#*T;2&si=JwcJO@t>9)RvF91L`g-CfZ_;v>IsVMik@-#{J$--
zq@+I`q=85ph`pY?6WRYu>YQqm(PH93wbX!f4b`F}D6+?C^KW+NzTR$5L{v+dC`{G&
zK=wXi_1eOIRqJ%&*8t)M8h$#0BKx}vLVx5YM@xt))zC~N%|xLmC~~0Z_3qN-RaR0W
zp(2TrL{;Ad*@s24Zx(cZ6(A)xlpQIFbn1H``{eGLCyV}HJ4H;q=`4<c_!!ve35v94
zu1h@Zv$n{J44@hY4IJdu_dxcyp^>ehCB|@4BBt!EM7Whm^aMq&x?BqnyJqoHLM*9B
z1rb&d8$CghpVwvwPnv(cSxh{s9d)3sgBU$Qk&b7c4xHmzr7_{CmU=L+2dSQ*$kUK1
zE9?p`-IMgCg?wU`PvpFgph(y9^n|dhrN!pNjB3dSQ8p;_1V!egrp=u#V|mC_q|ZLM
zl{x*+!9>N;RkLjX6)q^pKL(G-;4jhV7mlVT_vhe;S&QQfNMHd8MRPg6j98Wt8+0Yd
z*AZ16QKR4Rdmz&1I5fog#1Pfch^l(R@ng6#uDI45KSeb~i-BR`yQ=tq3J2d1b9?|7
zAjSQ|@u6I(6f+CQN6Dj{n1bZ^<RnzU@e|`FCNh*3la|2HY_Hk=xT86KDYsP0P$ooV
zLJXz^j;4x;EQrA@!tr}SvKOo|H*tI=NGc8Sd61k3Yv$HgkhC&v@i=z!m%L~)aQsaO
zx(P#35K}*O%MK9l09zbEq%Q!taJWF!vAXP!F@|YhDC$f1=x>f+34tqN3<eQLbG_*y
zaKJLa@mD}|1xBMGHcj0&MqW=nsD)|vX+G#Q+wTvVzTJ1kj|gd?*gOVXv_nV(gMOSb
zFi<a@$bR7VgPcu7S8H%0PNB8OUIcwBdap*D37v6oDT;Ppkehcrd>PgtI?y%{Zvz`P
zkgg!%Kwh7`)hfu+CZUj5+p86*PRP5d+?*JC1<F^T!nhUkuVCaWh(_!k_sK}V@`bt7
z+D`E71aD*r`M02a%f9Frb!Uz7zzLc@L5EJp=+!vn6=?8_7{)}@R~XR{jt}<=kHy8r
zGD#;{23#4~vq|b|CQdRiHSXyb*W)VbvbBM#4Tj))2DyAJNU{8AJ_z~P(XnH2yM!p8
zUzlN>^4v(sUm)BCBF3~W<S!EIi^P`U7Gl>z<cQG#x*EEYj22)mrVL#n#u<%lFV&VL
zN6a<m_&9Z(A6mfi@#=U3l%P&9K&g^cJB&RViIOw}G*vS79g3OO^XN%)SUa7!a*#3*
z$He_=B+}P<WN>^A1m-|A+YarZX@@9u*9T@1_35dS^;BOoF>5B4IPavC!U+{+H$o_*
zmB?C&H8RG_&JCViW$Qv2|FZj+293rHLM_mB6wHEb2+W2kX2Ex$`3^>+1(S{h&-~{g
zXI{wB6s}kgZs=f+Zg|@caK@d-(K10lKDhHZzMKpwCn`+AnA4~Qy$uT+KNHlMFaizi
zO>ych7=bRE*|g8CWZQ2O{}M3R;)q3ktqs$rma6R-3XlcZvvi#av6=888z~P$^I#-~
z5XWx<pG^?R@LsUl3l0p|f^RJZVSmK&=fUPYIH0FEKATu)6D95_jxQpiMPwwaNhQRm
zgak5NNg^xB7<3TFA0(QCWH{nQXMVYpA#A&(<oHC-#KE|d9G~QwWPmn;d?PqABlAF>
z_h;;y2cC?d5Au9Nem=P78}c`Se3K!+0OSRRxCSh0K)OrcyK%HLw5SCsb4o4P)`EiJ
zI<Tk%DRw8cFoSI!C>VYTLta7%b6PhH={Ce$iF_+@WKJt0@}fUu*COJ{_{BtCY{)Mr
zuEmD@5+W}#<d+h8sUfZ)w6S{EIM9kp=;u$^yOQ|hB5`~bp&tW&74fbz<R2t-T@Cov
zgs)~8gOjG+-bS27ZkShC40gY{ql|JMZ}=m~mg6J+B7<4ViX0Ic!%&Pk#)_dtzr<jJ
znDI+8K;s<8sTps)<9GwK5u7%H8;h4b@XCV#hV#KI-w@Y;tOo21<3U*qE{5x`tOFOe
z-p9f8I9OtX!SRj2HG+(7qL-k331NoupzSuqJ>cF0gBgD-aoS4U*!mX{uObq_a53>J
zHpCUgt%CTn^{XTSl|;*M6$z+fIJaZO)oE#S-oCef;)H2wECz8gD9~df{n-ig=Gqs>
z8lQ$c{QBO~o%h}SXIUo}SN$dP74At6Ut74Zl_vOXP-TM#Gk4)H?x+88YIg@^oFlg9
zh%=Ui!tTkFmRriN&=lSPUJVe;oZSF14KR-3hhY2AFr_~P%|i&o6u<fFsD14#mwD1T
z){y~q#DRIOi415W4p?-HpDZ@tn{j&{HKh?88bOOAY<qQZ#WkB;1!ctf$GwYPJWo=^
zr64Xfbe`^TiO$=(EO$eTI56TbKEcB#E%v&TChj8)ORBgC#6_$#7U>U~sCaky4MkYf
z>yr0Q4fiDNYkQBQI1tCLfWQ?n8W)43w+E)xpu~=e<1d5eGQ^@IkoH>KcWlMJ85fN*
zbca<O0rzV^QQzql|IG0V>=$@oapm|7`wRor4ED|7frD{;3)r_XT(Rgxzxec>kw&!n
z@sHQC+>zj)V1TCDPvvp+O8=1f=53R=P(Rg!ydGT9ZjQED15UFLKxMSt_a2#<xzlCS
zT~Hgx&y>x4hw2_?)Y+_gae(fOzi;Zl)EwTrxb0CCje>JTc8)baCx$xqzq9J)I%-`u
zG0i4YmL^sc)71>)p`8kD2k~~pZe4eR$GA9o{>p?(ZD;Rp{d=35T$c?FtR1!IFOhj>
zfBY<qNYkA6#?`q8#hKggYue~Q7YTRKumf%(TnodkLkBGx8z<lVUW@S7fwoA!R}O3i
z+^pRk7z&kz%CY(HZN6f&AQ77{+S5U3K15+IqsMM=D*!)+E5W7`R1DXFZyf})Y~Knt
zt%k??)x>%=abU|*OhSuE6wB77#I2P0F<eC=t4JKnGu6bmngk<8Il5VN*CsfvwyX2U
zd8p2QMpK@zqpwf5CiKMd*&!|*uXWV={O>TU=TFy&uX6k@aM=a^<sP5!OxyH%GshnQ
zmjeu6@p*KkCUNp6GmiJR^j2W0=J;UCU<GqovLe}o1@uJ4L<2O(a*l%S%`#Ay8Ma#c
z!C^mmG3T8Dw=)pPcIyRjzW_mqanf|Mx(gtj?DxN}O4{4*1~jS0s$+-a&bzkadt2lA
zc2{Yj{QVgbSBu7I35ZKTi57|U_eZ!HehF#~@2;?g_lDjb+dZBe?^=3OjeovZ;P@XQ
z=tp+X@T+gXE)z?eTq;;j%*u%jC)t}$%uW*-R+M|5$1435<vZyde*niHn3pjB&^)8N
z&_b`!Vn<vXh&R>-j&CBWCgO@3@ti{4sk`68FTN-}*5Fg^==|PUc1j*R%3d88(C6U4
zJ6mH>piQ3cx(tWvJsoS>AR3qW<<hgutlu<RzE@-Sn}ZcL9qH?+J5uda-(8E$Nu8N8
zg5HUbs4%}Z;OS@M04q11O9ePG9_!xys{j+b03AAa&Xl<ed;e45*!$4^IN*ywFo?an
nuvgyg@!6JH?b8y>T-7dvoSYpUR01_)tJtd(dv#^6&g}KSeRRDc

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm.hpp.C11C88B2080F572F.idx b/.cache/clangd/index/device_gemm.hpp.C11C88B2080F572F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c28118b71176e92d460ac08d29f42cd5ce71fde7
GIT binary patch
literal 764
zcmWIYbaVT{#K7R3;#rZKT9U}Zz`!5~#Kk2=nWBsg3_8pV3>9<s_Mi4UY#`$DSu?UE
zvhb+%npHubf=)~an`%pD#|dfIoh%D-75(qWF=a<)K=~S-`I6@=7Trv#K2?+Kv2aPn
z$Clo8e~z2X%Q;_FBmUsCoI2w^4yS1^KCazcdMH(3vw94xWA&j@fl_^sNuQ+Oy*N6-
z_59XXKIV>Z<}bUwAnA%pf7;=<mn$r+6;3k#l$G8mop_Q%I#&Fd%td$evx`-)+AL<D
zWK?x*l9<{0eKG52Y5Q+W?q79H^``78uVXfv%X3$iTlTEoEpx2zL~U$#?eiuH29YQs
z5x(m$DtO)UPb^yzVA5xOVvEVGhh71<>YOGrtTPH+%ehIY&2_?Tu6?`f`bt{+{J4_;
zE!WMTUYa}K_oz-VLvdwpQamt}^3F}kK47h!D!|0b4&(v>r;4VEH3J8PGh6D=rwG1N
zs|A=?VAA3&;^HvrYts+B&=RulWn*GyW8f5F7ZC&#44h2BfM;OjWMn8xO)KUBa$>hz
z=ESI}C^PbKiphv+z|2kUV^wVsDs}=2i%5&8!_0%ZijfoS8klgm7jw(o{{I$iJe*Q8
zQpzwxcsQkHq>*?sGBQX!Ss7U*o}7#v5>H-69?pZA2DB68MVOsyjjaNX&P|vObfJoj
ziaN|?Flk0Eu#Y`e%DN}+a;|%7!1{%qU%Z<M=uJT}X_%Qny@D#zFug^oIf+WZ@PT=r
q0cIGCX6Rp(;4fTc_NllixmX;il8ICG=li&y5n&vR9I&tlMGydZ#0-Z3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.236524D86BD9A065.idx b/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.236524D86BD9A065.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c83434fafd6eec7a97a393259688d13bf6471b8c
GIT binary patch
literal 23776
zcmZvk3z%DFp|1P?>!+qSprsV#+9C)j84wy^R_^x{SFS796_6w=_hcm)KoJ@1oam64
z5(Et54wxDhH6lSINZdzcqrx701dod1P>_p?a+QO7%MQ9d>Fgcm>v>L}$JjUT7v8or
zE4N7!*yC_KdFi@!*7Tlhl+0C0wr<_Jt<(Qqc~z}&-V{E*W!<`KZvSccW$z0YAN{!T
z%3k>Hv#<F6)0bR#)~D}%V$S#NyWYx=|NZoDzx3q1QMz{T-*0$xzwd!t+K;~Um5aZf
zd*uti)!ni0v-^(jzPM7`^T08`-*eM%_s&22?)mN~<QF5i{NnW&U%BeSlV06wFvcF=
z@$nP0tuOyj|L`BLx$)1zAN}p63%B0=*kj(kAA0t8FWz|N{SS=4`0eG?C(gG~`SKSx
z-T#(he8*+Kz4YGYGXft<U3Z5k{>bgO-cO}#=f;te^wgG%AKCKj_)vH3{ey44_`xmZ
zBOd$Yod>>t-&cO3I{A$M`TXBr_7+E19aKKL?d}`TuRS_?<f7N^eu&+^c+>mN{HJrx
zZ{K&-W8eGBW%Z|)0$(^y{aXEp&p!S8Uxjoh2OLK{Z+`LOi*7yV!(SD4K5*2&$gl28
zzwxJEeE#v%N$JQ}{jM!HclX;Ky?1@cz?VO~;pVq)*w{aPTkVvq&vaN`HU0UM<qeiA
zALYLI=E=WsTt&Ze$G6JY9R04pHa_t6jeq{sQFEV3-0^xWf3hTg<L{a`zVuS?&P$Jc
z+_dAp+bVbcCVb(wKmW<oAHMH=)B9dK?-cFxmp^+>?8>h_5Ts-J`Gd!P;L~5`f@i<}
zzu!1Lvj2=d4~8#&zwUuYoyXL-Jaq#9J@bK|ow4EQkJ;|Oy2klG{p~;g?AGv=XMgm{
z*Y`Zud2Pdyx(76u1h$#a{?v2VIvS7dzvJfM@uz(1*3<Zt?!WYM>H0U${Ib05oIPJ%
z@%dlB;0NE|{>yy$k;>-ZT#>u^YV|8$eDPyPrr-VMNhdYGc-+q}=sWkHeaHBwZ$0|)
zs^K?V_YX8#;e!X)U4GNmt|xB3{g=P#T(|W8qxL$^qvF2~9GB$id%snBYs)pav#)J@
z^yTxLjrBJa9{lHPmb0&ZvUPsJvh^DqUzcp}{+Ay;e%zadHw}+|bYtk(7x*t3Xzu-G
z_^!`B@#Ql=I2XR{4wdJdmwft``o1Uk{qjpT^;z5All!vu?%->ma=acm=J=DvyS9oy
zd+v>A6Q5Kc`@%<We*WK`j@Q2P>wo*sgWtX8%>3YEFW>g9FMZ2;=f0y(I@#6Ye@+LV
zN#3t{&x?lb`(N1mtBGSF_^`AgaqjuQ`Tj43^Da+)<V%lSck#6shL4TEJN$3n;T1by
z;=g>`*MAxF-}KveIqTf_f=jm!cAVCAH6P#n#HFWqzxMg-j=12Q3;yx?6{mbfKH7Nn
zTMuQ$Z+_>hO)n%02Op8F7yj_)-#0DfF1z~9|M>kA-#udB`Srihr+%#bseH~)UU{VQ
z^F-mk>-VNNJoo7SfB!=ChDYCi?$jIO$_pR<K~?wIGsl_8`rEf1WqKy|x5pm*=&diG
zw0-N)x#PT-Px;33hv)D7OXmlt1@HRvyFYi^+tuUlJN2h`F2ygp@2jtrj&^I6Cx2l2
z=ELE)zw*la&Y_<k|1^E#|CpmcSa->zzqqa>zvWukzw7DS{{DI8bJst8>v!|dyuHD#
zy7uSBzrNvA|8D-T`k(#nsNqe|^#5?-*@a!~KkczscYXG`UC*3*;qJfwNBz{jTHTMY
zKh68>LVGE5&7abHxr?K}zw646DXAyOx^?SMeV}dq(nUKr(MWt<X2<#+8&*%I*DR~Q
zHcs|GQ~#hw+(g4uee(K~x2~Sd^wweXgsbm<>4%%Bf2y~w-?m}(WTuU8zwzqV|M0Ql
zCK{XS6W5>kuGN#7HvRspyMH?V<kvRQ4^8#e>#r`ap3HRkijMB-i+}#tP4v>Ke((DC
zrdLm9I_kaZ`Dpz!*Kb)Doa&uBcJ5j|nd#cCp4vytN1wiBU1+Lz?bvnt>d8!h^@q=G
zA&<THwN2DB)u*gKW&7&MOyBv`Bl8=3*V#7F=u{uS{`g~7PiFdGfAePAOgD0yXke<h
zuit*m>d8#Ml-&6}_3l?M*hGC(ecbxvHmsh^^y?2k_uKFPXlUHLPMYc#Iif`p|NVcn
z$y-~#a!dXR!=E?N%cheLtpC7et0y!4_oKu2K3)CMA2v~Os*hQJ%!bvInO;}RWUu^5
zXzi=3Y+U`S-th3-`|o;AyK{5q_?^eUXZClT(iz8}wMKKAxr=B2_M0<v!kl-F<~8#d
z&;D$iGxNf{ca3-r4~`dja9WKPoL10+(`j_zbb=0?UZV%67xdr^8Ur|kU;t;-7{M6@
zBRG@B1kNOwz?n5>aAv^_&Z4n^vj`S&R*e;$Rj`7yX>8zZf(@KqV+Us!?BE<42RMh|
z0O!;=!8rvdIG4r+&Lz0O2^s;MAPC^x8aFt%;0EW>c))oC4>+&J3(hNe!TB^ka6Z8Y
z&ad%<^9z1(0Zjm0KnQ>nH6l1s5WxjCL2yAK2ri@vfeQ&CaA8dtTv!N$i)bR?B0>aQ
zR1*ak6{6r`ni#m45CbP^Byf@-fs1S6;Nn6YTtbromk<)*lA0vAq>u!c(xkwpgcLYg
zBZHF#8C+VE2A3An;4+#FxQvhim(^s!WrZxboF)e@C*;88HF<D(ArG#gDS#^o1#m@8
z5nNFyf-7lC;7URXTv<~FR~E|P6paE-5fpG0O$A&<sDP_#s^F?Z72JYm;o{lN-sa4L
zu;5*zMWR_GU>Aw7NWd=<%@X=0A}pc5f$YA4T=dS3OXfzh`^Gh%-IA9!uHJ&L-4x6<
z$lNVt?J9n*L+17a&o{{YEoAL#e!fHI_XG0{!lUOqgh#J!5G{IbhiK938bpU)*C9Ie
z`UcUX*LR2>y`e!2=nWlWKyPdiBYI<p7}1*=#Dw0|Atv<Z1~H>IcZeCir9mv{EgfP(
zZ*34OdTWPR(c2othThg8HuUxev7@(lh#kG7K^*8E9pXUmY!D}UXNNe^yBfrW-qj&4
z^g@FO=!Fgu(7PMNjo#fMZuFi8@u2r~hzGs5LA>a_9pXjrYY-oLUx)b6`y0fM-rpg9
z^nnHmpbvCN0KM2CB6_hyMD)Q138D{nNDzIfK|<(59TGwxZjdngaEFA^M;au8KGGo(
z^w9>1qK|e+6n(5gV(4QX5<@RFh=gA15D9&}LE`A+9TG>MXpjW@M295MCmSS*KG`8j
z^r;3(p-**43ccJQGJ3f~Wc29<Nuy79NE&^nK{Dtw9g;zxZICSbY=>mg=Ncr3KGz{R
z^!WzKqtACp9(|!f3g`<RQb1p9kRtkGhZND58l;52)FCDG<pwFEFLy{8z0x2GdZj}Y
z^pyswps#dD1%0(as_3g7QboVeAPcvUwVSDh4q4a_yhs{kk>FyHbf)LvW@)-vI&8Bv
z-7Fop`5f8(IbvM9`Rb6}(`)J(&mL*^$;Q=(m}|FR+jnk%&+PHo=FIjpwx6{|bK+bW
zZrP-{%Yf=+ZV7Ma#Q8AX%1QH=0oBR;65i&CJZyQ1hb><xJZ`n37PeYR3tMfSXmP6(
zb+FY*I@s#!M2B0wsE4gy(!*9?Cwkl(L<4LMk^#1cIx*nZC>mjFl#H-7)`=0fCeZ|2
zlVpOesZLC|HH&80nk6%A&2?hNtwprJ)*@M8YpD|pZmpsfwpPgsTWg(IacdK8u(e4x
z*xKsEhFiO6hpk<*!`5CWcHBBd2W%aZ1GbJjap2Y|I$`USoUnD)i4(Uj(FI$V<btiM
zPF%PZL;<#fB*0dvAG)u&_WkS@-LQ2_ZrHl(#En~z=z*<A^1#+pCm!5-MK5f<k{7n#
zI`QJxC;DLPlYFrC)rk+ce$fwGzvPFlzfSzP4Tu5Q2BZLN19cL>ttg7H6(te2Vx5S%
z4T?e72BjcugLM+bZAc8kHY9~$8>*8KZo^_2wqYp@+i;zPaT^gMu#HF&*hcCkg4?JV
zg>6)d!ZunbQQXGF7;Ixw47Ra4iQ!ffCD=-m1Y4<2B;3ZuIBerm9JcW~iQ_gQCSaS8
z60l9wNdmV?F$vqGl!R@vPLjAyi7D8oq!esZb&|rZEXuHzB^kDIoyfRNi)q-Vr8I2Q
zb&|$yM$EuABV}Nlsgn$DvtkyuSt$$KY@K9rn-g=e%}F`f=ISJe+q{^EZC=X5HeV-s
z+!n+FYztBWwuL$=;I=3hVOx}nur1a}5w|6=1ly8Sf^DfzO1LeHW!RRbGHlCrQpT+!
zDzH@~1-43^D7dYN71&m!3T!KNQo(IitirY`RbgALlPYc(#Dy^2OOh5Y1FDmSrNizs
ziHii*i$q!^uwNwg={+dC#L}TFBe=xUp#u#ru{6EJ(qWhQEZP0pwTJ(kGdGdlHzD6b
zcHgqbC%$cX|FyN3cWm99NvyEMLAG{dkz8R(<kSjFA<HXFMozD=G;(HzWstKgEQ_35
zVL9ad3d<uGR#*YKxWbCar4?2}F0ZgMva-Sy<jM-GAXisd6?uuQuqASk&AxoQ8Xy11
z7cRN=bh0^fiuRN}v!5Kyez@@}-MIQwjkSBEb8R|zBZYgKJvz4^c}VB(L7vdLw~<@i
z+&&I2-{t0RM{csY&#*N=-=_07(zP$xe2>oWM;_Aodypq|{%zzIH@}aA%Xhi?+mV}W
z{xk3m!M7=oJ-$bI?D0d&V~?Lu9(#O?<FUtgIUaj_lkwQowy73-+8)(nPdlVq>}e-d
zi#=_N(_&BC<+Rw-HklTCx;E8ePuHV5?CFM7hdted>aeG4aXRejx||Msx+c?MPv53`
z?CE<{k3Id6>anMvP(AkaEl!U;eV5Z?Pv2yE>>1kBfIUNx8n9;=QUmr36KcSop~V@n
zXXtVU>=~NOfIVZI8nI{WQ6u(@Lu$mHaYBvQGqyM*_KaQ5h&^MI8L?+-Qxo<~J!-<9
zX-G}jGfk)od!`m=!k($inXqSSG86X9ZED7zxkt^|GY_d5d*%r>W6#{;%-Az`IWzXm
zO=iZPrA;l^v-GG1dzK-!V9zq47VKGCoCSN9E@#1>rO7PVv$m-fd)6McV$V9HR_s|P
z)QUZ8i?d?S+U2a+vo@I(d$u;UVb9j1HtgAk)P_CVgxauYYjHO0*}9w!d$uOCVb9*C
zcI?@E)Q&y-klL|lpHMsY>@CiYJ$sk4W6$1XcI-LY)PX%mk2<jD7*YrJ924rmo}<M%
zu;=J<4(vIa%z-^;n>w-Q>`^E7oI~owo^wK-*mJfxC-$6O&WSx|lR2^HYEu{XTs`W-
zo@+>5*mF&&3wy2>=fa+=%ek=UYBCr0gf<nhC-kU*Jz+=%><JSpU{7dq0``P1Cty!#
zG68$;Hg#jq-J@>oxrfw^J@<sVvFC1aZtS_coEv-YCUaxY)21Hmd3w}?J<pJOu;-aj
z5B5AQ&VxNqm-ArH(_|j(dE3;BJ#UYCvF9C9FZR3>>cyV7#d)#k?Q&l1d7I3OJztyp
zu;=ShANG7h>cgIILVeitwKyO4d|l3mJztagu;*`6Klc1R>c^gcNd4IJPpBVz{ubxQ
zp1;favFC3xKlTD`8o*wlM+4Xk3~2y+fe8&@FVNxw*b8*I0QLe+7Qmj^rXu#l9u=`C
z4ylMeaY9Axi7igVp4jC??1@b#VlUXHLF@&4G>E<6kOr|AoX{Zlf-NqHy<nFMVlUWa
zLF|RvG=#lSkA|=p8qyH<LK7OoUZ}-|uovobA?$^kEQGyqn})F$?$I#z!b2LyUU)*o
z*bBF~F!sV-E{wf!lZCMtY10VyB0U<xUSvok*o#bP1bdMd7r|bn%SEskX|f3RqHP+*
zUbIJ}*ozKn6noJLjbbm_;-c7#cDX3_qD>aXUaU=H*o*aO412L5jbSe~p)u^mT3if!
zu`U<GUaZMt*pu2+!k*Nl685Aam9QsGsDwSK#Yxzcx}1bPsmUbl#oIKFy?Bqtu@@iG
zIQHTb8pmF|#l^7~?{abM#hWaSy+oTPu$So31ojd`n!sLSLKE0aw73NJ5?wBVy+o5G
zu$OGpB=(X$n#5jmNR!x0PG}N)$rhKyUb4$2v6pPJB=%Bmn!;YHM^o5K4QUE{sR>PC
zFV*5w*h_V}6!ub0mcpLgrZV>A9+j~t52=hjc|v9E$t_OCp4{bR?8!|gV=vvNY3!wY
zG>yIVkfyPhp3pS*(k(8Hy>ypLV=vugY3ya%G=sfNk7lr!8PW{)G83A?UZ%xmu$Sp_
z8SG`6EQ7sln`W_>?a?guvO}81UUovW*vq!KEcUWpE{nZvlV!1&YttO|ay^>EUT#Qp
z*vn054tu#4m&0DJ%jK|_YqA{n@@<;OUcN{3*vk)T9((x-&0{a$;_}$bcey<F@=cb<
zUZG73*empC0egiZEnu%Op#|&}T3i8pg)UdXUZKef*ekYa5qrfREn=@Yq($r%C$xyY
zVv8$cuh`{^*ef<!5qqUJEn%<Jqb2N>hO~se(u9_<S88!3?3KD)345g`D`BtPre*Av
zd$f$b@{pFXSDw%^_R1};jJ<M~D`T(RWM%9rZK_~T=}`rH%8)A9Qzlfwp3>qJ>?vJN
z!Jg7&3ic{(TESkWM=RK?3~2>>l?kn2uhQZw*sFB83ic{ZR>5AiO{>_e_GlG*)gi57
zuR5Vs>{VM_6?@e#SH)hn$*R~}Xw!uo>Fnv+>C5VRbYVa8kS^SVJfRD3Be%GPeH>iA
z%PriF+++)%dFSci&6!2gri%oNi=;O_2b+szI6Viei)1oA2fK@;#VrynFOu%`9BeO=
zCR-#}Uz*mJ4y!Ls>(kd#z~5+TT3<S>zBH{b9adkO)|U>eFHP%9ht-#+^(Cm!{*9(j
zc^`VY{@PRC`-y5lsmxxiVZQ#Ab^q~~t5!Ic`6r_KCz72Bmbrtd?jV=1@!=P=1}0~8
ze*ES?=;q92?6R}h{r9iU)1Q&CI(dgL`N9uwe60S;mg#K!3ENLyd%ef;Cmet3?1k(?
z_jhW;zkU9jyQi;VSIwOTPnWCa&sux3n^*BbT9p<^r_urGReB(U$^c|k8G%eH6OdVD
z2C}FuKvtC%$fmLZ*;RHRhspuuR5^iMDi@HT5`f$)H;_l=0rIN6Kt7cZ$glDP1ylhb
zQ6&NeRY9PTDg+c(g@Gcf2vAfN1&XO+K$1!VimT#42~`3psY(K+R4E`?B?F~ZX`qZM
z1C&)|fpV%GP+pY>DyRxTMO6`~q$&ZGRb?PWr2th_6`-oB3bdeFIBRyG9lEknEfU}*
zqFP#G_eD2Xcc}M$cKWlaleH&1W-sy$dbXYU$D4FVOwZm&HtZuu%?`NPQ7}D#551vg
z?EpUfIGw$><^vm7-&3<Tn;((+hv0=I{D|<#+7Zzr>qbO}tRE3QvSCCF$i@*dBAZ6U
zglrxWGqPnwEXdXou_D_>#D;7i5j(PDL>$P@5pg2BM#P0IjEI2j9uYUPXGA>6-VyO4
z`$oiv>>m+7a$rOP$l{2I$iWc_B8NsKgd84`Fmhx>BFNDZi6X~FB!(=Fh=d#;kvMW<
zL=wo!5lJGaMkIwSkBE$%9+5P1W<)Z`*%8Si=SC!loF9=qa$!UY$i)#UB9}&_gj^nx
zGO{uv3UXycD#+ClsjhM2`9^O2Z+AX(=q6>E%{_@fgI#9xPvXx=ml=<&U1nNj-7?c5
z>zA1x*|5wE$i`)6L^drm6S8@knUO8a%z|uPW>#d|GP5Dumzf>evCJIE&SmCAb}cg(
zvT%S2EAaPsA7E}|&jIE^_8wqfWZwbiL-rqFe&oOb7C;scFcCR;fCZ652UrL>e1L_K
zBL`RnIeLIakz)r~3|TtBB;@!37DrAjvjqNoB$rtdIkn7E$nr9ik<-g8jhtC#8RYCT
z%OdBNSq?eB%<{;EWmZ5gF0&$XX_=Lf%gd~ctSmDHxw6bE$kk<5MP68D3r`+)ANj!a
ze(MUHvqYAs_a)ahsq^>z^D)T>A9&C7U>dWzhwxw;v-yYMn~5JY9$7nPT4ddr>5%nf
zrbjl6nE}~2W=3Swn3<5xV`fISjF|=5I%Za6+nCvq?PF#~c8r+=**RuTWY?IvkP~B;
zKu(TX5;-+yDP(!fWaRXirI9mZmO;*rSr$1rW;x{inB|cRV^%;ej#&}8G-f5_@|cy8
zl`&J0D`Qqcu8vt1d11^J9y;t>XFUDZSv%*$Y~2^Wao*cMJ9IrfU~}KaZ=C^~|1N&(
z3>c5B9WX7jZoqWN`T^4;8wSjPY#cBnvT49f$mRhvBU=W{f@~cyE3$3CY{>Qjvm-kO
z%z^A2FekEWz+A|}fC<R%0dpgJ2F!!(9WXDlZ@_%W{sHqN2L>#FEDo5692~G9a%jLp
z$l(DCBS!`-f*c*NC~|DTV#v~fNyza5iz6onEP<RHuq1M7z*5NafXT?|0ZSui1}uY|
z9k48NZoqQL`2ou#7Y3|=TpX|>a%sRy$mIbmBP#=@AXf&gf?OT2D)PdBEqoUaokcR3
zUJv6rOVe|f;GEeT<&W67`Ud&6`}%YDlDUWA4H66Yl7)v3dlBQk(-$e?IZNc;=}QpT
z&iTj(ZU5{2(F=Xk$LTxmJ0r746`M1=?7Jef?q1jALpPsyZfNzf%Fcb{g4tu0#f1kx
z^)JspH=c@}v7L$8rLX_9{iONF_nk03d)o=ycFx`ya>EO^d)0emnL{@xD{Sr{UMH-u
zd3=E%zruKA?F!Q(>sFW!S--;c$c7bWKsK&0BeH3FvxjV6VP<5@3bP<vSC|#qw!&=4
z_7!GFcC0W5vU7zwkzFgyg)AIo!ZUE4=sw8Y$ex4DgX}%XyvV+T%!lkh$o$BGgDik7
z9%LeN@E{8!hYqq3a(Md1LXI3{5#;DW7DbL7WHDsvAd`^e2U#3>VTCQMy`1yVL+XRm
zW8#^2j*{6cIxpO~`WX;#1oX+=J#Ylf_sRS{_~^G!cw}v#Xpwb&qC?jAi5}U|CkA9=
zpBRx%ePTj3_lX(V(kB*VYoAz=ZGB=xw)cr0+0iEsWM`i^kzIY_LKgZ&Kz8?u8`;w*
z9%OHyc#(a5;zRcLi61%8Cjn%!PekNkp9GOZeG)<r_emHz(kBt*XrDxpV|@}smik0O
zj`v9%IngHx<Yb>DkyCw=LYDhPMo#xh8adM^8RTrAWRY`yl0(k-Nglb-Ck5nUpA?Zx
zeNsX$_emL9=@SLH(kB(<YM)e*7y4x3p2Kbd`qNtgJZFjYr?&uj&aGtkt;90BPM<wF
zJ3Z76eW(f?YW$pY_T<#`gGKmxCp<;V^FUf&3#8+9Kzd#eWZ(@zM&1Zy;!Qwi-V9{n
zEkIV@3S{GLKz7~^<lr4ZPTmRR;$1)jF95lDH;{+-0C{;YkdOBP`FTH3fDZtPya*KJ
zgFqoZ1Qh1OKoLFy6y>8pF+K()@e)v+j{_z61W=Ms0;TvAkj%?KX+8~<;WI#4J`0rN
zb3l1M4^-d_Kt;X?RN_lOWxfog@Cs0cuK-o~D$oMI;KZkB`9%V}bm$GBz-_X-P0X`H
z;F~YkKNfrL_BOXUbLNIKk6invKYzpdN3MO-e`Lc)j+`a#y5^sc@F{)A^e)edN1nK2
z_RfR5ue{<r)r}e3^zPLjw&$$1pO$=rec~N@aCabmguVar^s+n7-Eq<E^PV0K8vb<o
zslR@7`lq;!u-nLKvp?SEOqZ}O@)rsF;u_C>*fKqW4}I<t9Kp4@y=#A&`P%&6wZ9Bs
z<AJm_Es(CJ1Jc*@K!%zD$XGK1nQA5=bIlB7sab%mH7k&<W&^U<>_Cp11ISr(0=a4~
zAfYAzxod79Pt615t$BfbH6M__<_8MY0zhI-1Pa!IK%rU)C|nBzMQRbCXe|m9tHpq%
zngkTD#eovF1W>Y;1WMIXKypn6O4rgrnOX)YTgw9FYB`{MEe}+v6@ZGhB2cMT0xH+a
zKuS#ks?;h#)mjy3p|-Ghc6c4SM_5}Vz)PgIgnTpEee>E8{Ok*#G~KlIo~NdV)UMbr
zdG<i&_van)mv<d?Tx)vvK630na_a1mntdGD^pH99SzvI;sOQduLq<J+9vm`i9!RU!
z0_oH`AiY`-WKbJ`jA|p0No@i$tIa?bwFStkwgTDIHXys&4&+cffShV4kW1|X64U~a
zTkQt&s69YlwHL^z_5u0TexQIl03@nKprASk6jFzP!s;+kL>&Q&s-r+Lbqq*SOF(gT
z94Mhq043E)pp-fVB&%hhv^ouxQD=a%>MT%Bode3N^FRf40jQ`h0+rMypt8CQq^K33
zin;<+Rab!))C=d~A){U-z)M8EguIXJ-naJl!k1Nd-8g+y*rA_O9(V3>{OpLB9U;>X
zb@;E3!J8c$v*Td;bIL;>em6UN^Z)buf!XKa6*jJZ{@vPaPM*uI9a(eA+!bpyFVAOZ
z|NZG*O=bRyHR5F+9Ix=;w6YeQR?&jf$vSX4MF&nV>%r+2Jvf7G0B2AP;Eb{noKZ1?
zGsz}!CdCBKEStfZ6*D-CYyoFcEa0rN6`WPEg0smsa5lvT&Mw=**%doDhwK37P#oZ#
zvJ;$Bae{NnE^scz1x}C!aDpO$bIWdUZp97GBYVJk6c0GB>;>mlyx@GY51dc&f%D6L
zaDK%PE+7ZM1(X0dQ5M08iU=+!2f+oEAh?hm0vA$3;KFhkTv!Q%i^vgh5hVgHDo4RZ
zl_<EF90M0qV&EiM0w*aFxVRh#7gyro5^@4uLP>y2%1LlZB?&Gir@*C@6gXLy!O4mY
zE-k0QrIj?ejGO_NQ8M7Nau!@x$%4ztIdC~82QDw?!R3`axPn{&S5OMzigFQLQ7M8e
z$t7?lr39`lm%){lGB`z6z$uCXt|C{!Rg?<2s$2zERjS|?<b~|o?ZASvaK#!e5_yq;
zT_nmP0l#$UGZert9eQv6JN`ev|J3ur3ttUS--IrVneZ?^XE|nWWY3s+kiBE(MfQ!E
z57|Fve&oQI1(3xt6On^s7DNt>SqM2iW?|&Wm_?AIV-`h@jadv?8Z!wwK4x*`MKYc~
zNBQtO&r#Mg<>a9UcvsfsD=$Cvo5Z1$|N51h{bt>|Rkey-X8NMb!Bz8r_&@K;UY<%M
z<@7pvZMXX0`Y%kcots_zt*oTW`6+%d>*m-I>))l?uxay_qmMdr&sMwjy<0c#v1~i;
znC-{E`{H@6PH!-p=8k<o|DFraI`#i|?v4{rI(g@=(@sC*%(LHj&bjA(-~!D>r|dpK
TeZKA39nR?oO*_117ZLt1M<d$J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.3F5C595E84E47B20.idx b/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.3F5C595E84E47B20.idx
new file mode 100755
index 0000000000000000000000000000000000000000..223bcbe35d7fd61d2da420538edcd2dc94fbfce9
GIT binary patch
literal 23776
zcmZvk37BJLfv)@i&rg%ySem9m)<$+A1wsd?lYLL)$$GM$h$K1LlapW?6qOQd0%c;F
zAYf1kZJOw~ge8a;5*?9IM=bjajDsT-t+J@7Y$9$q=rHN>#(e!em*+9^p7$%>?y5RT
zDoG`<%i(zXz3bLl)4Q%!GFK<rx^?T0p8Rk3)wRNf6ZpuMb?dIZ>qp_2y)RyN)Dy<t
zz3~0#UiGb?U4Gp;H$M22Ip0_A{cC>o=QF<YyQk-k(uc4A`%Q1|^F4f9`?24B{<5#+
zUb*8Jy1jcox#y_vORKeA4<GZZUAO$=`uXSHG2i{D{8Hq$|9btUSFXO~q*sqN7-LUt
z|Hz5i*5|&f|ITl&z4`aS@BQs}mmGcn<Bxl<f8TSzeCg)h4?R5k>{nJ&AHB#%<tty>
z^w3`oqrF%B;=K>9oEdmu>bkw2`1kL+{UIuScy1IaNzZP%?E71O8XxM8eQ0psOYhlI
zKJ4+2ePREXzV`Vas!l%h-(L9J%iiMh)d!RhZoU8Ji)xQ8fB(|g?thfswRp?B&c5>k
z^IKoL`tfi4@rwGhOMyGyseY~g-RFMxtN#w^P7XK@d%^tDWtZN5{`>z`*zxd@dm{h+
zwe%am{jX0yaRw<J@v7go<<{;#+hY%|4;lFK`#0SB*PAx>&)8Z!<(jh{mRC)`|5$m0
zW%py;XWu;ezZ_T7FYmpteC<((|E2NnFKzt&4M)y>BC+@NSpH;5{PN#5Z`}R6;1}L|
z#1p3NXKk(A`%L(f5C8awKYRbzE;7CAwF^(tzHsGp=f`$`@!=pH(N7;Z_T4vrjtidq
z`hUK0Mr7ZayB-O@_dMOhk2#O2Z+Z3v{u}1~PoBBqs1Mm5x~9hYZ~V${p1eK0``i!i
zetp;Do!2%Tp?g?!d0?yg+#8<%u%q$#zP+~wk3Z#x+uzBb^w4{+ls^8(*`JfQp1<o~
zSAG81FaGwowmp>(f4{Q%nX7WQUZZ~HvoC$<i1a(&Jn5w7XODaG;=Xg=xqC;q-1pc=
zs)lEd-Z#)>h4&m-cjYbDxPEf$T~9sJxo+vaBd>Q{NW~uv9GB<lS@)Ixy5-ut*cUfG
z_VPu|#`>EIk9_mmmF%k@Yh6^Z9R1~uuS>Rf{QLKwIPT5Dn})|exH0t8i~W}mG!H%%
zzW0+q`P|v>nG1hrugY`J<u~3|-}CgIr|!0?&)Ir<?sL}rgRkA-cs+2;@h6M-9xXok
z{2R|DKBhkQ#Sh;4!k?Xv*S`AGKYjI)uU~t1e(<4}KXc#R_gTNN=g5;zcD48))4?Z_
z4{1(+$*^tTi`W16*s&1&j<g|h!9~w}>nY*FD^nl1`}@~j_Tfvy$6kD1_>R|pd+ZUm
zD{|Y8>G0>h>kn@|m%qyN=^s8B_=zQb%(-veard3A^qYUYb(mPn{QB*aAFp4%`Gv#j
zsZW=S*W6x7o^<|A&qhl>$VR_$+?QUz_nyahe*53X51D@QTRxij=ELf}H`(v_+0*V5
z&cEr;zrXqX)Bbk+Q-)*idE;8+-(GzGJ&NwNzuffLGjGvz9{rg30{MZq^^0fz@_21;
z{hcrT=?66Xv(q*6bB}%Mkv~Lk_{6b>`qtfVzxpqK{@gQ9UG?(PL!VmM(l2~*%Uj3g
zLr)ibpSt2k`@Rdmen;8*r2V<$OuxPM&Z@Hc)Gr^o^YbTu_3a05`ojay{$cx%K4QFk
zgV{TN$?!j4x>5JiBZuGl_Vbz_$DcZI)|u)X9{u(?_vqP0y6?U9A4|VBT>twQs~`X3
zUwyya@NZB2=L4HB*!xQLGUZqI?*5RH`UzRLZvCkbx2<<yx?>ZK#MfoEuiw65?P7Ax
zvi8r$@xGte-=h&X(eOl{y#C~)*Dj{|=wb7OYwrKucQ;Z0L~mWcb;H`lR3Ce3Pt;y}
z&Gyafq={~k!&)Tq|Nd_}F~0T2t6%!{hlZPIY%)1<{fURKT}-w0#y`D$#SshN+(h3y
z(PyncE4_9x)t8@h=BAgNPknF`ecwc1v;La$+Qn4gslM*I!;U-ei7o4b6TM^mj?>mI
zrutvM`}`L2_*t)QqMnI9W&J7J)-I;{3(tOkeq--C+a?;F=;PNPf6Us&R6q3N&?`R|
zx81j8U1*|D+kV;^YZp`f@W1SKe|+)NKiWjEnCN$}fA<w@7gPO%XWmSk=|*l74NUa5
z_1lhFyO`?VC3k#7z4O(JH&Ne2AGiLv4Qm%u{r9875B{wBy<cyl;zS>_{+JDG7gN2i
zmdWn^VQBW%y=>#!SM{dvytVJ%=e0XFXO7=-{OQxb--OOQ{+t=j33J}*=Y4Z#UYPez
z|1O&|yubr#1uc+H&;jWMJ&-{#02u`%kV!BBnFTYDMX&%_1uKwEumRZxJCH+g067IG
zkV|j@34#FR7TiD{!2{$Kyg)v|2jmz0Kmj2DBnl!>PzVBrgb+|z2m?ig2vAgr0>y+F
zkR(VzaUl+r5E4L1AqkWcQb4jG1EqyDP)5iAWrZwIPRIe}g*;F}C;%0OB2Y;v0hNU^
zkRm8R6`=xD6{<iB!h&~tHLy9eNQ6ZKyhMa0<eSOPo5`iq>!s<nTWRCkwc70Tm}`)^
z+sN!BnCp<aeZcb#GJhMHofh*QGQSU)Zx9|m-yuADZG&jhYdb`XUe_Qx^tukwq1QKv
z9=*Op^ym!@VnA=`5CeK+gBZ~pJH&|I)F3AGrVcToH#dkGy}3io=q(LmL2v003wmpV
zSkYTM#ERb5AU5>24zZ!PH;5g*y+iEi9S!0@@8}Q*dS`<;(K|cDiQd&9F7&PraiJF)
zL_jZeh=AVRAa3;T4soORG>8Yir$apGy$#|;@9hvTdS8S1(EB>Xhu+^He)Rqh@uLqk
zNC17HLjvf<1`*MV9U`I+Hb@YCutS3ALk$u_AL@`0`f!7U(T6)Ej6TvJ5%iG`iJ*@*
zNECgvL!#(o4H82i>yQ|FsX-+4Qin+B;|&r=AMcPj`b2{y&?h=1fj-$FN%YAMNup0R
zND6(bLsIDF29eRr9U`MoH%J<Nx<k_FGYyhKpXrbc`fP(_(Puj(i$2#NIrO;>$)V3T
zNFIH@L-Obg4N^c~=#T>XVuKXX7dxbgzSJNk^ra3dp)WT`8GX4!%IK8_QP3+LqM)xd
zNCkbRLn`R24N^s4?T{+^g$7x;jm&O#7CK~MAMhe+kVS%vMbeoZ!OhZSvvkO2X|h>5
zWb-Mq^Han)y}8;v`7-XDTvJbZdhfeWHm=?6p51<J+p+ER={?NNnQdooJ7-37;#?SR
z*`&EEfa+v!32*1b`7qqdN%L0#)ye!4-sXusY<Y=?Eng=*ZndHowpvLGTWy_aajO$`
zu+>RA*y`#;hg-d<hpk@H!&YA>dfXaB18fbF0k(!ZG2qrH8ewacjIcG<i4nIZ(F9wQ
zWP+`!PE5Epi)PrGB{OWzbz;V?MYO=yB3WQ-sS^ust)dmSR>=xmYn@neYZGm-wMjPE
z+UmrHTf1n7tzELi)?O!e+&V-DY#ovVwvIY+;MOTRVe6Efuyxjn6SpqW1zVTof~~7g
zT(}iP0k(o9z*eXqyl*)Bes+s)*t#V*Y~6L@#;r&6z}6#qVC$(94{p7p7q(u>3tMlU
zcya3!eX#XOKG^!|#D`nI=!dOe^263&Cw|-p!~kpqQUJDrItk!b6h+vIk_cO|PDI=W
z#UN~hQV_PmItk)7B!*xcl0vWz)kz4qVKEHbuoQ-ExK6^jjffH0Mx+RABXttNZB&fH
zHY!D78?BQlZewB$wlOIN+gP2%a4U%tY$Zv8tyCuxZsTGcws9#A+jyPCahnhmuuVt_
z*e2>Ef!m~*gl$qv!ZukaN!+Hy6l_yc3bv^_N#Rx&W!TD+3|qNQWZb63G;GsS8n)>=
zN#iynW?-9<GO*3mNd~uBF$>$Ql!a}!PO`Yoi8<Kjq#SH>b&|tvUd+QbFXdsIuai7(
z3t|Db1*rhrLY)+FTNI11ElNe$7VD&l+mcv<ZAmJ@wp1r2+?K^MY|By^w&gl0<5m$B
z*ea3&Tcu7E+*ZU2Y%5X)wv{@m;I=AOVOy1|u&vff6}JoGLKyBPNefp1)ycxrA@`ZY
zMFQ(ZA}tcwFOvG?9u$7W(!nbu_=u&0CmMXj(&QtS4*7^rlAWKN{rJB*a|_vd3-WDb
z=WR2d-p`ucz52&vm(x2~C%$Dk@58m1w;#PZlUQYm18jCHl3Zm;<kTulA<L^wMozD?
zG;(H@WstM0EQ_35WjW;hD$64mR#^eLxXOyirBzl!F0ZmOva-q)<jN|mAXise6?uuQ
zvL$kWO~0<EH9qq89hcvJ2HBiBMSIGw=|im3hk0M68`mD+oqi@aXXe^;?q&*iL3?y=
zAM%jSJ%Bu>b8jKHxVb$Xe0-Oiy9>F==03q@e!flTZ>F;^*nE%9??WEa`3I24bp9>m
z7B|0#gOBfW^LHUP+59Kq8-j0B9(#O`^4Q~tl*b-Fraboe7RO_c?{Yl$_$K4Ar)^U$
z_Ov~!#h!LZwb;{+sTO<M7N^CYw##X;r)@GV_H=Ek!=A23b=cDlsSbO(G1XyD*Wz^8
z({(u=_H<3A!=AoP_1M$*s2+RzA=P6~Kc;%@>06v0d-^V?$DY2)^w=}BsR4V29yMUk
zFr)_T8OGFrJwuB#V9(Iy4A?U?nE`voHZ@|;*rP`58Hdz}J>!@fv1e>?M(i28oDqA*
zCNpBs)TSovnR?WOJ=2hyuxA=m6ZT9k&V)Tvmos6{)MO^?ncLKiJ#&wmv1cAqGxp45
zYQ~<q#hI~Z?s8`AnVZavJxiNfuxIH}3-&BSYQdgmOfA^6v^Wd)EM3lmJxh~WuxD*k
zEB35CYQ>&)NUhkjj;R%U))r^Qp0&$av1e^EEB0({YQvtbM{U@%4XF)#wlTF~&(`8>
z*t2yx8}@8XX2YJnP3_pT_oy9v_93-n&pxJh?AcqK9eegJXUCqs$?Vv3w5bDojvjSj
z&oQJ9>^a8Nfjviyb70TW<s8^^G?@c?&Ng*o&)K6+>^X<ji9P3-I<e<$aZc<xyPOky
z&L(qW&()?b?74c>g+14hy0GUOQy2DJEzX5KSC?~P&(&lu><MiuU{B~#0eixb3fL3I
zRKT9l;sopoT~5HB&}0Jk+->T{p1Vig*mDo58+-0Cbz{%n;@sGCcR4rq+)d`jo~KPc
z*z@$L2Ya3&^<d94rXK8hTAT-ao-XIXo~Ow?*z>lj7kl0w^<vLEq+aZK$JC2GZ;SI{
z&)em^*z-1-7kj=o^<mG~qdx5UhSY~W-<bNa=WB63?D@K!4|~2Q^I^~5rhe@Cd(@9T
z|B(8z=O0r)_WUi*k3D~v^JCB7WPa=g+BAT@K#vBn7Z}n2_5x!Xz+Rxm1+W+BasliG
znk;}lu}wwni9IS}PaIMad*Yai*b`fvh&{2(iP#gHOvGNWO@r7A_Gl1$!66M|FF2+_
z>;+p~5PQKc7sOt$$%5DmwP^@@p&kukFEpef?1jcOguPIU3t=zR<wDpCHCYII;WiCp
zFWjSH?1hIkjJ@!fhOrlJabfI*yIdH1;U)`XFVdzF>_vJsg1yL)Mz9wd(+Kt=EiQt+
zNSBLXFVbWY>_yu&ioIx$MzI$i(kS+#V;aR?w8cfS7wvLU>_wX_ioIBy#;_Ob(HQn(
zLmI<gY)oU=i?z5I_F`QwhP_ym#jq!}sf0bLM<wh@Ln>iU8dC{-Qj3$YCv`aqds350
z*o(Jm9DDH|jbkr9q;c%U$25+;c#Df;FW%+i*o!w=9D9j2O<*t4qY3OKhBSe_#F!?q
zmuPVb>?OKf0(*%jOJFbArb+B2do+o?<d7z@mmJe1_L40wiM?c(OJXnCWJ&C$+BAi|
zRF9^xmm1O(_EKY-!d|MyrLdRkaw+Vknk<DqxlLv4$vrA#PaaYkd-9md*ppkFj6J!_
z$=H*dOvYZiP1D#*_h=e>=^;&HFFmGd?4?^=8hhz3m&RVY$<o-%v}p!=nI6qxFEgYW
z>}AF@gS||P%V00l<ucgIG+73F**48$FWaM8>}7{Ei@ofaX0exTaartTyIdA~*(S?k
zFW06y?B#kihrQg8=CGF=(;W74EiQ+>T$jsXFV|!_?B&}ukG*`4=CPL_(meL^W17ca
zzQyISm+x|U?B$y*kG(>h7O+?7(E|1gLt4OIVN46yE3~)*_6l9DfW1PK6|h%q(<1hY
zJzB(GaY&2UD~@Rqd&L%4#9pz>6|q-rvLg0MZCb)!sYgrLD-CH0d!;cgVXxHUO4uuP
zxf1qDO;*BQxlPO1EB9y_d*vZ5W3N1>W$cw(Tp4@iE?350xyj1dQ`%I)p3<WV_LL!2
zu&0cvf<2|hDcDoGoPs^2$rS8W+O&ebN{?2sR~ga@_9|mq!Cs}sRj^m-auw`VnyiAo
zYMWNESMAX%_Nqf##a?wxtJtfyxGMIlU9O6~YLiv5x6q~wH`D3EwVN{wJ-V<Dc}N!?
zKpxYDw~$-h!X6GjzRNA#h1_HdpZMp)!3STF*QSdEi;JW;IfBhaGMpU2>LM9Wj$n6@
zw75lr<weq+9KrS?X|hFv^`%LD>5%%;q&|6(1^kVcCiSI5>PwUQ(joPwNqy;%`qHGn
zbVz+^QeT3)`Qxvw`@es@dX;0DJBjK}lAS)Ez%qM@YA?BR#`}nBAE`{)8W^A1`Hwf>
zrkgWYuq)1;zN&Zg<XJIRC;#N>W7?C4)DONediId|p|$CkYBD?c0_oZ8(CFoNeEa6d
z>#uB?M7N!=?bO-pMUFq=_*17Z4i>t<R2%;E>3eog9=BJ`oiqItOkO3dnm=dum^iQE
zfwU?ukWQro(yR1929*KGs4@bXR3;#^$_!*tS%9o6E09fP1G1~^Kn|4y$f<Gyxl}G7
zK_vjWRc;`U$^+z8d4YT?ACO<=2MVYHK%z<n3aWxYAyo({tO^4~R1u)4Dhd=+#egK0
z1Qb`rffA|&P*RlyN~uynvPuR@tI|LjRR$=l$^zw7IiS2M4^&VUfQqUjP)St+Dyzys
zib?^hs475JRTXGKwQ$b#L^ybzq*^4vOGLFaWA~-ER=2D7d~)&>*2&tF9n%*H2R&QQ
z{>?4A!zPFKkPUmtk<-&P@j@fF{ueube(<DNVRKL8lT9mZ{%L$NXoc~}+7+fn)~zrd
zvVMi>kqs-%fNWf0Mr6|pGa;K-m>Jo!!Ys(v6=p@YtuPz1eTCVP9V^U%>|9|^WY-FF
zAq)GNunK>F_kQL^_UvaKWbc0FMfUAyK4kxX=0^_fX8~k!KNFFI`&ke<w4a5L!~0nn
zIkKNckfZxq6gjq^#gL`_OhS(DXL01j3QOSUBe}wo$f*^ULY7yUjGSI!Y2?fb%OGc0
zSQa_A!g9#@6_!UXtgr%dafKC;ODn8|TwY;iWMzdZ$dwgVL9VW_D)Pb#TX_1AI|D0|
z8)!VTL{=s@nzJLRi*|kUnB*f5pFTO6Mr`g;Jefvp{!#d5;zx`}){dAKSvO)jWc`Tg
zkqskeKsJt;5!p0iCS>!7nUO6cW<j=&m=)PJVm4&^h}n@HBj!MMj+hhKHDWI0#E2!3
zlOvWyPK{U!SspPNIXz-&<jja=kh3F}Mb3>_4mm$!dE~-~6_AS~Rzxn1SP8j2Vr67y
z#1!Pph*glQBUVLT7_o&%;hSlRj3(bYvm@_k>+bmSg>OB1@OpT_=Dv>KIs-QUb^O*D
zFdkVuU|M9|fa#F+1Exnd4447gIABI((}0<f%>!mewhWjB**aiWWZQt*knICzM|KRD
z1KBxXPGr}BxsZhc6Oi2l=0^4mm<QQAU|wY3fccR91Lj8#3|IhJ954|%IAB5K(13-I
z!vhvZjtp1?IXYlb<k*14kfi~WkmCauM@|e_0y#NgN#xXkrI6(TlabQ{mPXDDSOz&e
zU|HncfaQ?$1C~cF3|IlVIABHO(twqa%L7(MRt8K#t_)ZOxjJA~<b?rS_&S_Ai)1jl
z9>ya}lOs!T<OA=q{h#xqm-r?>4(zb+h)jQza&zW1`)QHsuRdPy8o%$>3oi(*-6h(w
zhg>|pOSHJ~@D2a|-1DP}*b&>2n11xfe{Vl&{)s&&Ob%~7Ve5|R>(_63@h-3W`dH@R
zb^a=wJAfzdDx1fb8}X}*N7k+~EwXNv>5%oSOpk0>Wd>y9Dl;ORR+$OeyvoeTmQ`j!
zwyrWOvTc>wknO9?j_g=v4rJ#lb0WJ|nG0Dsz=WT}6_@(}b0d2WFb}f#0P`aI4lp0G
z{{ZtN2M(|RvUq@r$iV|Fh#We=Ldf9*EQ}mEz#_=e11yRhJHTSd(g7wR#}BYL^1>=x
zn7z2{;63-%$v5uo$n-nw-5b|_vv>Bx?ff#Ce-vKb%`X!kS-VWM$hu{sL)I@7J+fh$
z7?6$2#E5KKCMIO_GBG1tmWc(~x=gIdwq;^Nwl5PqvSXPzke$oKiR@Y?E@WYu2*~bb
z;zsr?6A!X?nRt<X%fyH5UnYL!z%mISi_1hr4la`*a%h=^ki*L)j2u}e5#;DHi6Y0A
zNeo$9CK7UdnZ%J3%OruETqa56)G|pS%gaPYPA`)*a%P!ikh9Aqi=0~~Ipq8@$s-q*
zNddXIOp3^*Wl};eFOxE|vP=}@$}*`SSC>h3#?zNwUb1oRH`-^X#ay4vJ%GR9(<k!}
z;Azq)JhHY=w8*+X(IM;mM2~Ff69clbPmIW>J~1Ji`^1cF=@SdGwNI?bwmz{T+xx_h
z?C28*va?T|$gVzdAq#yXAiMj-jqK?Y53;vUyvV*j@ge*B#E%^4lK`^VCn9pNPlCvy
zJ_#X*`y`AU>5~X@v`?bQu|A0*OMN0C$NMCXoamDTa<Wg7$f-U_A<KOtBd7Z$jhyL|
z405(lvdFnU$sy<aB#&I^lLB(FPm0K;J}DuW`=pGl^ofF8>5~d_wNI+Z3w^Tiz#%U}
z?@wN2jz^YAfASJ?JaRkPc{{O8@9In+8l9YK2j3zBPBnhc2@jC+^G<kxl;?r8ycS5u
z>wxsU9>~BOfQ-Bm$i$m~%)A-M!drl>ycNjC+kou69mv5ufSkM&$i=&W1YQ7g^KKvy
z?*a1iULYUu1M>5Jpa35L5_u6Q$OnN!d<ZDahk+t|1SrZ!fnt0NNa7`+I3EW}@Cl$K
zp9D(rDIl4bfzo^$D8pxfvV0aO$LE0Zd>*L47l4X<5vatMfXaLsNZ}Qr3SR-L@>QS(
ze!+<kkn)QJc<JD;ZvwZ;&NeYmPl0=0u74=@{9SEsbLQ*~XCE>9roU*zMMuoO=|8aH
z14m2~_g?$W!+c8LG5Ja0#3N4JKK=E8`*&aU)#}EKZSqscF1G8O*)ty>Wgq<~O@I10
zY~$J+uD}`hAenm<Zrm0gBnyuoa&kX7Il1x35_xcPa?g%T|FDziN)Enb%j`dW=o2=N
z><pw2v-e+~{OPw}u>I2Md+2;8X!z}wr~dS@$>-rSgnfp*bNUI|oaqwQMgA;dpPlja
z$(G4=$ia7lnPul{bJx#4o%6N%>!*Lh$y>41cpz;}3#6;*fb=yzkfCM(GS-YhrkV-J
zTr&e%Y8D`C%?f0z*?{adJCLL10CLuxK(3k#NT>-w?wT9OQ}Y0MYhEB<%?ISK`GEqp
z0FYP{fr7OlP^cCH3fICwky->OT8je3YB3<GCIQ83aiByk0hFvIfl{>;kX)02(zP^D
zrj`N9*0MmkS`H{*%L5f^1)yTB2vn+-fXcNpkWy2CDzyqwwN?dMs4ZMSJ;4wDs8(Af
zz)PgIgnTR6dF$*t<hd6=X1e9*2cDgrQm4gElc(qSuP!|7kB1+5Tx)W84>@)ZIdytU
zP2V7?v~lf?k>He3&s{h>W#%+<m(6HiJ%8cs1e({(Up6CN&4c4LJUFdd3r?%ig43yW
z;B*=tIK5gAPOs5}GpG&V3>pJCquL10s4;>wsZHQa8WT9P+6>ODF@v+HE#NE~3plIV
z3eKvrg0rb@;A|QjIJ?>o&aSb8bEqBQ92y5Wr`ie5sd0jHsa@b)8W%W0Er1g=0ywwY
z4bH7`gY&38;5-@+IIr3Z&a3f)^QnE{d>S7(zuFJZuknKmr~}{vngBRaErJs@BDkPB
z2rj4zf(xlb;6j=ZxUf15F02WIi>M>uBAN)es5%NRs)>S&sbk<`nix1qErF9X61ccJ
z4lb^VgG;Cr;1ZeyxTHD>E~!a^OQ}=fQkoPvSuKN;H8QxgIt?zZNrTI%GvG3s47jX1
z3ofh4g3GCM;BuNAxV$<KF0aXhE2s<L3Yr4AqPhsKs40RgsY~EWni9COx(u$YDT7ng
z3OGfhfUBr0;3}F5xT?AeuBxemTTm}tIJ<&h&@5awqeY@#Bw!bbW|4qjBI+gdOGLAT
zeh=BXXZEYNn=?0&oj1++W!1emPkvAG;B)fFU2q&fz4o8J4_;y8+WX(Zt$;k2on7zG
zDRWoNXkMPr&aU_8mHDe?#LGN5Ug5!MWi2?Zq6Mdub>MW04xC=rgVQT|a0b}`&Y&2;
z8D%3lqhbVSl1<=DiV2)qHiI)OW^fkS0?wjXz*%K0IICg>XOnH<Y>Ew>UABX>D|T=W
z*#XX>IKVk&Cpf3#1m}`n;9QCeoFEI}1VsSnmfhgoiW{6q_JH#!9&ldS3(l)}!TDq#
zIG^GJ=a>E9{E8o3Kn{QlC;@PyEP@ji5nNCXf(t4^a3MJaE~JFOh2=1~uo4Crkt5(D
zN(5X~j)IFSQE)Li1}>(=z)7+MPEsUraXAhyuEfD5<OH~ck^q;Kli-p{5?o47flDbV
zaI!3elNA|UT26yYD`{{UIRh@EWWZ(REV!(a1(%a^;Brb1TwczD%PV<s1-SsOpcKFr
z<s!JEQUq6$OW;aM30zq&gDWd#aEh#eQxpYUMXrFWC>3y3xeBhTRKYFC3)$KA{(`b_
z)r=O2yhy+<5@nHqUpn}HA>fw|z7NGe{r`OH*%yMByc(XouYxdQ!guhAuMu-2dq&KI
z>>V*LvTwwE$o>)YBL_w-fGm!fh#VZTAaZELLdf9}3nNEHEP@;zu_$tE#A3+Oh)Kxt
z5sM=)PF~!wNWSyW=WJ`4a`NEAA-ikx?kf-eOXA?gKmJKg|7G2}HMNFZ=BRgm^GB`X
z%ZDGHy)u<Z%IS6TY`6B``Uj?GN3*kkm6dckKf(88-5fh?{o$$&n>KGb>c}H@9c{Os
zb@awvmaWGfv+ekITsE)O=?zBH+_C5Jr(bf;ssF#D+fO{{<Q=EI^NcgkKKEVcUvS~O
bFV<Xo%FYwi7uk;8?wq{Qw8Lw55#fIU2|u^z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.5836A0C2B3A14787.idx b/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.5836A0C2B3A14787.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0ad5910275629856bc288d9a1e42bd2f979c9520
GIT binary patch
literal 23778
zcmZvk37BJLfv)@i&rh3PSh{H%WbH-}P*NatfI8XtG@h&{>j_AblRY^J22e!Gwr(^g
zrU{!toPcSAG8%y(5+u%58JvNs={w+6aVZ*P8HbTo9-Pq$I&;%CH`LenxjYY{-+XU)
zySnQnsceD04#(;h8#Y+ed#_hA*CyG94I8%2{=MefTH(SOeER4O8?L|WpTe(sU%u?P
zr;XS2!w;Ny^$(uA{Kj)X|IjlFzVF=kr~KsK&-%`*zgjd(H|+b%P467=J$ifR$ydLA
z*>`fUf8{s2yZ7I+|G3^OtF^t49{=0DxBh0|;tTIz?0rUlC35@E-g@QrYcDzVjco>F
z?CG7KJ|)|}=g0cT|MmKte;@q2Kfij(wg;Yi%DeAlFZ}Z>H(&F}qm!?GXC?KSi)>WB
z>Xj{z{K+u6`^w*3@zBcIfsdtbyxS9h;;uU$q0$WtlSoN=_UOx=IQp0Ik?w>?hWEen
z(WA>pJ@wgp4}SCEum7X!w6p)`OMiaNTO40|NcrUU2X4No_T>19OW%CpF?QGTtsgq)
zuP!jZ`|!0-eg6+v)}L(zzH+Ae&H9gDc<#5q2<c7>IF5SB{K{pQ-f{lNzai{;^w|B8
zUp$<C`*%P4^3!LL(lKxNT}R*6J79bAp^YH}U;g-}+x~RZ=D}IpYo}j#j>GbX>Gz*4
zZ?atTB=^;KPWzeTTKcWK?=N3}-2496{Kz*q|Ne8wE_^X@_gk_2X_EM@zi8h6+N;5P
zuQ=vu)6NfWuiW?R@Fh3=^dF!5_`?^OKJ?~=r)yuj>V@-T*ZlRPK{}ydK6JuIK7S7v
zJnyal{q|Xr18493Vfcy<>mGg5d3^ooXHVw8Z$9{wvo{_0Dcd90)j0p>zw@s@xg&hd
zd7r%It-Vim-`sSJ?orL<f$iq=KKJ4cj^<Mb?!GN}(&?YO<4pe4N3OU^`oi1i+#_#4
zfA2R|eg3yD{?QM1{5&6iqO$eZSLbfKPW}2<U-{HA>G!{L>Zz@-p7@iC2hIcM-95SW
z{wF_OHT-(pfuSZVeDu(Ut8Tr{^~`N|{ruP68yg=!cAw)yD*nUJae0n@@cz=Dj=ugb
z_Sc)AeC?uEbK^~gAO6ktE7><b+rFq^+4il?Z%MZI|LxyBed0TXcMMN{a&zdH7yB<C
zY99J|_`X}7x#ygZE`;yATjlxo<)6R3zW-PIfBrR_`rPdw$lYUoAo%9z9B&1VKj}2_
zzHQ=9UVQt7#Anqfy!^@AUix>Z<IV5>^54Gu!|z>xPJZ~Q*Y3RkYxi64-GA(<r@7kv
zPwDWB$wxFFc*U^ez{~r7F?B2jAD1>IF1YB|Klr(D;Z><meC>%FFT3HA;R*5ghyTqx
zx_Z~E{5^Mm^XD=Dt-pDnv(9}#*tlc3^UR*B_4L+ft~jgr*I&NzsEg0P_`lw|`t+~M
z#~F|N(_>lj+uyx*%gc$vp(iBkB|rY@4@^tBE3dovKYshn_l_ERe)+fb*}qqQS3dtA
zUw@+V(?sFnFYHTidhy8v|NfQeO;5i2;u$x`m6t#MqpI$y=T9_|jdyK7*7SVr&rkjE
zlXtv!>W*zA=gtdXJN;WLA78xp58WS~8NBcJ@Bh-B?^aKI_>6zLw-LYe;cvWNI?k<C
ze)S{Mw;vC``}Nm9bUyw3<e$=~{Er3tqYalo`Li2K@}FEw2X;Sq=U=|8eCZ3%-SNHr
z^Y3nQt8VzI@sDpi)&IQsNBvL!eC+7f=Li3C$$5p{>_6?XH+J9h;_l}!xMa^C|D%4!
zKCSNWzi_7ag{4j-bN%nq`?$-pzrF98PbsNq$c7CY&v>+B{o19ww$Mm?LuTj3otxH9
zX4foh|7@Nfc)tEojktw|XY*+rPusS3GM~4NS|?xkz^gyrLjALO`^N2?)=uX0?&UL-
zbF_b)Y@t`o<_~WCV0!IjK3{TL_?DLr9Q)K3dgW~X$i|Ocxpp$2jqkqw#y9`vQ==_3
zHk(h`c*^_MPUiF8lV6Oy9DiuEb%QjU+vKP=N&Mgc%$3+LglF#m%T@g?^kXySx{cSB
z*G}g1;Kh3qr+oLQw~pQroXxv-?%KU}GM~Tk;}?%6Pkr#sEz~ocPv3a@j<u8deDAYQ
zEN<@KXxl=gv-zZrCmp|bGM~TcZa!>yV=z5>LufYd-nsj%wUhb$55ImVZKj*KEi^Ej
zcWm5o{MyNUel@x4`|3S!T)c()X7h;~Pu#S2GN1o)T==2qs(<${Tc|jjkKcIwrnQs#
zyrGuKUUM+C{<k|Io7etYZ+iUQ1NXhC-L*Ay(yo&}F#if>boNQ-uG4~g;llY}-PX*4
zX5lg*VZpmji|WM-=YNY^GmDzV%YcMM?>g~n9!y@tgUJg#m|C?KOsz%>rdH5`sZ;B~
z)M<2J>I5B_dbJ)*y+#kFUeJSSP#eHBXbfN)1Ou2xwGm9C#t5cSFoJ1Po4_<_OkkP>
z6PRYT8BDXr45nEygK1G)z_e&AU|IwVm{zqFOsmEUrd6<lX;a(4v}tT$+5{VzcC{T$
zyT%TtU9f}cP&>eMXdGZV1P7Q-wG&LI#tEiVaDwSlyTEj5TwuBc7np)t08`KiU<!f&
zrd#a>)2(rX=@#5zdej~;JsJ;~9>D{qSM3GUtMP*A6}({j)IKnM8XuTG!3U;a?FZAZ
z@q_6X{9p#u0Wbra0GI(G0H&xG!4x$jn4%zp8B_<s3~GX428AG)A$17MkR}9XNC<%$
zR)@h1Yr<fLg)o>Abp*_aCIV(eh=3VYN5PD0qF_daD3~#I49u7&24+l%fhnmaFeQxy
zrX)yU#?^5!<C-{_aUl+7LY)9Jp-F(55E5V})k!dunk1M>Aqi$modPqZNr9OXQeetz
z8BAFtgDDF#m}zwy%(NyAW?D#tnNerJ%xE%TW`qowS#=i7tR@R)R>*>xQ|G|UX>wrZ
zgdCW8bso&TCJ$y_$b(r>7r-oN3Sbt50+>a05zL~d2xd_zf>~0Rz$|G>V3vdum}PYt
z%(A8oW?3kMsi+k&6^#O>A}C;1)D<u*nhKZ|p#o-AT?MnMse)M*s$edumoA*&iQ1Z3
z(kxvDBrJK?X_=^(3AoEdvrNEVCc-iSzd_Uu^czIeK)*qR2KxKSp8e#~_wMs$ZX$bb
zLcW>oxp|%EcUw!F*Y2*a-|bmwl7-vJ`W={sE?GDLyx1g*x0Cg|EsI^UcmSAh5*~NH
zOL*M1O`^qJ+a+4ubxoqfUDqW#-1SYO$6engdfW|7V!++dB?jD$O=85|*d<2XO-*9L
z-P9!}+|5m5#@*Z{X51}JV!_?gB^KPRO=88}+9g)pZB1gs-PR>G-0e+b$KBo~cHA9J
z;=tX}B@W!3P2$Ae*(FZgT}|S`-PI*7+=V6)a2L8nz}?*>Zrt5n;>O+6Bp%#7UE;yr
z+azAxy<OtP-Pa^O+<jf*!`<H`e%$?C;>SJEBmvw5T@t`uY!VT7u}ehUgG~~|J=i5d
z+(S(g!adX_A>6}F62?8;C1KnnO%lO9(j^hxqfHXUJ=!Hv++$4=!#&m|G2Eplk#Lu~
zM8ZAZByrs1T@uGV(Ig4n6J3(PJ=r8l+>>3B#68s{Dcn<ClEPhX5*c^7OJv;BO_Ihv
z-6d(<Gfk4gJ<}x_+_O!R#XZ|4S=@6?lEXdMB{|&lO_Ikw-z9n63r$kMz0f5E+>1?8
z#J$)hMchkGQo_B|B_-U;O;X0a+$Ck)l_pVeSGq*Oz0xEV+$&vD!M)lfRotsxQpJ6#
zNtSLW>o;plU9xllc$qZGGC^aRbZ6(F)0jDpBb>&}X&m8viR}3jF|Oa7b;+LDHT9h5
z533yhvHkFaD(kmjJ9h2(!2F?ut(hHX?>Kjz7Q}@x+_FgvR|3_^LIZE-#Kkb&%1Mh?
z0@cZ418?(09$a4H!R70OM^`Iq!PQDyaJ6-!MOP>4z|~1QaCLQ}Lsu{A!PQH8aP@Vf
zN7o=4z%@t)a1C{0K-VZ5!8J-oaE*0hMAsymz%@xGa7}e$Lf0&s!8J=}aLsjMM%N-*
zz_myga4mIWLDwo;!L>?OaIJM>Mb{?Uz_m#>aBX#BL)R|a!L>_vaP4(sN7o@bz;#Ft
za2<8xK-Vcc!F5VbaGiDHMAs#{z;#J3a9wrcLRSz4a0N*KSEwJp@3{W`>=xbNx+OQb
z?mBU!>k&QRdL$3Ho;vZM>lMA=dL=Ko-a7H3>l1z8`XnE?zB=)t>lgjt`XxWO{yOoa
z8xRBF2BZMEfjSAGD~ckxq9lSV)`^I2Pz-__l!D*}>m-P7NDP4+l0x8y>Li42SPX+3
zmcrnM>m-bBM2vtNks{zm>Lh}0RE&Zfm7?HA>m-V9OpJjWlVae;>LiA)Bud~)k_4_)
zClb1GF%E8Aih~=klQ_BwF#&EuN`RZFlLWd+F$r!`N`jlLlO(z+F$Hc)N`afIlN7qL
zD1$3YGPrV`$mpiUG`MLg4Q{$l(&%Qy47eF718$~HGU#T-EVx-I3vRYfvgqc-9Jo0t
z2X3xTa_Hv8Jh*u&4{p9r^5_=C0=NaK0B)g93g{NaBDh7V2yU@Xis+Wa61XL)1a7HL
zO6ZowGPq@_3~sqj%IGSh0<Iz{;3{>Zpj#0u;8vsxxRpAopj#EI;8vw7xYatTqPrw6
zh2dV3v~(p<oh&tuxX&am6R?+wv`oNXCiU4pD7-}D@Rbo<qH*{@gG)4KmuMVuiCf5?
zTh@R4-<r9V?70>BcCzR8bw1@?!-sFEy|#1P)=XlRB@VIm8;j&BOCqOMSqfQRWioPl
zm8Fq0t1N?@U1eG1+$zf<=T})Cxv<I#$i-DwL@uqe5^{N!m64TIrXW{VSp~Ve%Bsi>
zvdS9d5S#z?&F?bO&1+B9t>5$fE28=<lAYf}-QE23e|_chJI*3oGpB1$-#dS*Z2sps
zyY%6w^VTnYH&NY9u9{zZ{><JRGw<;8dh6aohc4Vq=f8>BlTUrRZ~%Ek7al~O(uH@C
z+uXu_4ldv07VbiBv4t<P^?k8J7jLHX-}dY&tUg^lfIOm$4<b+L;=9OgZgD>cm+x_l
zcOkdf;uqoDi|<e#dwier*yBf(#~weWJoflD$77H0aXj|;7UQv}?NBZDw0)|@o_0jF
z*wap_7JJ$@r^TMO$7!*rZ80tObRDY0o~}=I*wc-u4tu&O)nQN9=5*N8^*9~&bS<XC
zp1wo%*wgo^9((!`)niXTrF!h?+ngSI`W~mpp1#HO*fVsf0egl%HDJ#$q6X|4rqqBv
zLz^>T&(Pxx*fX@40ei*{HDb@$r$+1<N7RTt<CGe)XKZsu>=}EU5qriKGh)xwp(gB^
z`qYFy(}<d|XPQzI_DpTgggsM_GhxrvVkYdFJJgIlbDx^AXC6^A_RLdi#-6#&nXzZ?
zac1n9Tg;3-ONUypXX#T5_ADc6!JcJGE!eZPISckIJ<ftXON&{sXYEic_N;ws#h!IU
zt=O|psTF(HHfP12wZ~bpXKgVn_G}$$!=9~AZP>Gos119zDYaqG*5+*3v-LO|_G~R?
z!=AlE?bx&TsU3Uv5w&B_KBac-+1s2Qd-fh@$DX~#?AUX3r~`YBK6PNvF`^FaIi}Qs
zJx7~!V9(Lx9N2TTm;-yx4s~MB*{4qIIY-oqJ?E4<vFB`aPV6~*oD+M_7IR|H)uArz
zx%$+FJ=ci3u;-dm7xr9j&V@Z!k8@$q)nYE}2^}h6Pv}ztd%}nc*b}Byz@E_N1ndbt
zPQaefVgmNu9qPuOyHDNNbC0MSd+sT9W6#~@+}LyXI5+m(E#}6ar$as1^Yp0)d!7;X
zV9zt99_)GAoCkZJ9_PWHr^P(j^LD5gd)_|vV$VCGUhH|N)QdfDoAYAN+vB|0^R}25
zd%h0!Vb9m6KJ58M)Q3Ial=`seYjZyA`FflWd%hO)Vb9;8e(d@C)Q>&?i2AYTpHe^e
z{B6#UJ%5k$W6$4We(VK0G=RN8p9ZiO7|{Us0#h2mUZBkduoviY0qg}@EPy?+Lq+U~
zeJWy498nQ_;*^Tm6Wg4KJ+a4$*b`e!#9pvNgV+o9X%KtC5e;H5IHf`C1>0N@d%+$T
z#9pw)g4hdnXb5|uJ`G_nG@>Ewg{Cxwy-=GAVK3C<Lf8wnSO|OJ4h>^3+^1peg-0}u
zz3`NVu@`Q0VeEx_To`-d77Jr9(xDOTMfx;?y~v10uos!q2=*dvE`q&CkBeY0(qa+p
zMLRT#y=b3Cu@@cDDE6XL8pU3;%|)>n?Qv1;MO!S2y;z6Fuovso81`Z#8pB>}N@LiI
zwYeDfVm&T~y;zIIuqSn>ggvQGCG1HfDq&BWQVDxfo0G67^*9N8Qj1C0i+5-od+|Pv
zV=q3UaqPvXG>*M^n~P&F-s9rfi?>)Ddx;KBU@y_93G5|CG=aUulqRs3XmbhdC3;)}
zdx;iHU@zIBN$e&2G>N_Bh$gX@oYExrl5H-Dy=0F|VlUZZN$jOMG=;rXpQf;v8qpN?
zQd643UaHNdu$Ss_DeR?MEQLL}LuKsAeJW#59#I*4@|4QhliQq(J-Nrp*ppjK#$LKZ
z)7VS*X&QU!5lv$+J*8>vrQ2K@d+8pR#$LL`(%8#%Xa;+kKFwe+Gol&nWu`QPy-b_S
zU@z0-GT6(sSO$C94$WdO+oxIVWk)oNz3h}`v6pRgS?pzdTo!xT7RzEU*P%J=<@z*-
zz1)cAu$P<C9QJZ;E{DBbkIP{%*J3&B<vTQwy?meMv6mmwJofTan#W$g&E>I|?{RtT
z<y$O|y+Vf;uvh5Q0`>|cTEJdmN(<O4w7CNI3O%lXy+Vr>uvhHRBKC@XTEt#)M2pxf
zPH7Q)#Wq*OUa`j&u~%%dBKArhTEbqbPfOS<jc5sbr7108uhiyB*emt8681_hR>EGn
zL(AAJ_h}h><q<7ouRNt??3LSG8GGd(SH@ns#md-II#j`)(x(dclo3_1r%b7WJ*CYl
z*i(9(f<2|h6zo+xw1T}#pH{F}8PN*%DpOj)UZu@duvh7E73@`7tb)C2hgPvy?b9mu
zsv}y(UUf>V*sHd=D)y>9u8O^Ci&e3=)S*i^)A@tHvzPew>Cyq@5nXx^c}kbwMQ(FT
z`#HFLk6XG6xy6>g_}+udTQkd~Lzf8_mq~wi4mOv`Xm$=(m&tT?4tAGGn_DJWUM9WS
zIoMt%Ew)Us-k8-JN7Ng$`s|e%@D(*?^~MqP#;o2rqTZO*8%NX|vwGu*dSg~^K;8U>
z*EjsfAFf^HSmwxw(pl!n2glj$1wE<*q%wc7+!~mk-TnJ_{)=wST*<CHZ~h*St+S`n
zSe?AbbA9%bk^g=#%=|(1%fIrYo1dz`e)R0}J5Jtl#`^0;PCEIdGv;sl5PJWtHvHR{
zzrAPnkh^N(TzG(7wRrCO!`r-y2hysvKsuEUNUzcZ8B_)!qsj<mQkj6vDl?EpWdX9P
ztUxxE4alyt136R<Ag9U+<WjkS1eE~fR=I&ZDi4rX<puJod_aDcA1I&-0EsFQD5weo
zg;XJ+uqq4`QAL2FswhxQ6$6q~5>Q+f2TG_CKuJ{+D5XjP$toErtx5xBR2iVGDhrfT
z<$&_4JWxSZ04l1AKqXZPsH`djDJli1qN)H@RaKxR)zZ1^S3k={wM>8;MAcYl_ocU0
zcdGZ_GJDeMH0^1Q`HO^up6%!S>#e$@W@qmwoA#4q=ZDh#|2un@>hOEV<_FZ)|K|Wr
zywuEX{LQZC4<9=#Y~fe<^wJ7j{1ra^v%+{}?F!Q(>sFW!S--;c$c7bWKsK&0BeH3Q
znUKvZ%#3VVVHRZT3bP{HR+tUhzQXLtjuqxWcCIidvTKF8kcERxScR|OeUQ14JqMWw
z*?W+Ak$nf557~c^`H=$$SpZo)$VBAeK^8;~9b_To@Ie+vjvQnW<mf>bMUEY0F=Xi=
zlaS*FSsXdB!V>uBkz8R(<kSjFA<HXFMozD=G;(HzWstKgEQ_35VL9ad3d<uGR#*YK
zxWbCar4?2}F0ZgMva-Sy<jM-GAXisd6?tieE&b|<oBxBeo8(nEr$JU`H<Rm5>Y}}W
zbA0lLkA7fwFiqIPV|Xx4*y3aG&BRX_kF1?AEwXOHbjbP%(<2)u%z$j1Fe9>Q!c55K
z2{R*GCd`6toiHo1ZNhBG_6f5iJ0{G5?3^$svTMRz$cYI{ASWj*iJY3S6tX;FGIDyt
z(#V+!%OGbbEQ_3*upDxJ!t%(42`eBMC#;BEny?abdBV!b%7iJ%l?kgLS0}8Byfk4;
zj~(%?Gnswste^97w&5$^y71kf9KIeNvW4&Ax6Y6)eh<HOhKxtn4w)8NH)J|w{gCOA
z4MS!?HV&B)*)(J(Wb=@jku5`JLADN=71=gqHe~yd*^wPX=0J81nG@MHWG-Z3$OL5f
zkhzgPL*_yD4w)C(H)K9!|B(5S149--7Kcnk4h~rmIW%M;<nWM%kt0JEL5>bt6gf6z
zF=T1TB;@#z#gP+3mOxGpSrR!lWGQ5M$YkX7kfo6`LzY3#4p|mCH)J{F{E+353qw{w
zE)H1{xin-Y<noY}k(D7+kSjx0L9Py26?tjMmc9pv&N3O!u7~lQ#_XI1ob!o~+Wyyv
zqnG$*KQ-*K?~2UdD!(<e+rB$8e{cIf*YsnzU3fug?Vi)F{p8~LJ*VZRM?d$sFT6OJ
z4ZC8y67x%c;rI4a7oXmL^6c#GCvV?1f0_GDFW===?~7#)U+1r~g+qAYuChgZxe~w1
zcx3G=(<19unGRXM%Jj&FRc1got}-LCX_c9f&8y6eY*}R%Wa}!kBHLD(4cWfR?8uH)
z=0J9?GAFWYmAR0GLri!cuDINXm>b!1h<T8`hnN@HcZm6r{fC$zIdF&tki|nxL=GNe
zLFCXO7D5gmVqxUSAr?W79%51C*zA>N$kHJuA;%A~IP%geTUvi{+Tq6!4$Z!CpMURL
zdH$XCk<Du#-LQTaeQ`_{AA=Wx^JBszYsW;3tQ!*@vVKhT$c8a7AREWTh-?}Y6S8?s
z%*d88u^?N=#ENVi6C1L9Ozg;xF>xR}$Ha;38WR_?FeU=BdraKOo-y$td&k6!>>Cpw
zvVTnc$bm5lAd6!nA_vDLh#VS|5OR1-!pM;^i6BSEB#ImxlNhozCK7UdOybCiF-agN
z$0UiI8j}>VJSH-7dQ8&DnK8*AXU8OqoEwuIa(+zm$b~T}AQ#7^h+G<z5^{M=%E-!?
zD9DvDsUTO!q`J=Uy&iM(+DDeG9~KJ(vhX0j4`@IZAH>6CKzL;BfM}6*1ENFL4~QPw
zFdznG<A4~EO#@;=HV=py*)kv&Wb1%fk!=HFL$(iy9oaD;4rJ$mIFVfg;zAY%L_l^A
zh#T26ARc7zfOwI81L8yW4~QQ*FdzYBaX>`m;D7{?Ljw{*4i88eIWiy-<miAzkz)fA
zLzV_aLXHne962!{3FPE}B#~1Cl0udTL`F^zNE$gaAQ|NBfMk(#1Cm3|4@e%lFdzlw
z;(!#9O9N6uE)Pf<Ss4%oxiTOX<m!M_k(UN!>A@pjt3H^$79G!NkiqPg=6KE>WX~PM
zGQX=ce+2cC&1)b31BV*F;DpCf`9&u@j>_{uT3!pJ<8?rKUJqp84M0ZT2xQ_-KxW<y
zWZ^A9R^AF^<8458-VWs89Y9Xr3FP8kKmsoSxp_B`hxY(^c`uNU_W}8NKTv=V0ExT^
z6y$?IAwC2Y=EFb{J^~cwqd+k}1|;zkP@IngCHMqTl1~Dq_!N-L%Rp&94V2+CKv_Nu
zl;d+ic|H$R;0r)Sz6ey}OF(7545aW1P=&7mRrxB=62Ih}Ki+xxZWg~xfE$P3<q6y&
zdpg8CKLqAK4ji?4?IW(>5O|0zJO;NmOAnEy$BsDK9-19(cus>nG&|bX&-wOi^-slK
zysN`)&78C8oMYCHr;9dSbj<qk^odQMIA;F4zwi3LIm)LD9J3!XPC4e3o%8n^JaEm`
z->q)W*k(V_>}7k;U4I(oGwd_((GT|o(nr|`ugxyI^Mak1&OaRO@u1;%SDo?8CuhHl
zI|;j!oH_r>w`O{T^^m_x*jLwi{$$JS%Hi-we8HbgZDHT~7qeJf+_(P4@HHMtThjvR
zYC0f&O%G(K8Gwv6Bao?P0y5XkK$e;X$Xc@k*=jZ*d(95ys5yX~H7Ag(<^mFG0+74r
z2J+NAK;D`c$XD|L`D=clKrH|y)<mFSEeI5<g@D4fFi@lx0gBe5K(Sg3NUBLd@md@x
zQA+?NYe}F~Ed?akWT12{4V0;6fU>nLP_C8(%GdHhg<1iqSStdRY9*j@tqi2p6rf72
z0#vP4ftG4Z`{sw&;VXyQG68OoS_Anuvgfw-D~A_e{;cWNZ4W*>JEV5UcFXhYir-#%
z)F0k=?1}B!+55=}`^g#eLu&rXceCd)4}a|4{E&G~b>Gdix5gZP?&ib`PUPp-{Dbfc
zo7X-7Z*I>XNRSt@>j&6^vT*e}Ey|18^}}sZS-g6kc$tTdS9sX8vKBV2qJ>Q->tNF<
zI@t8G9yYzAhs_`xU^6HN*o?9fHlt#M%_N&(Gbtw6%(59avtowLB3odyC>GePvK2O~
zVuj5n+hDUPHrVX49X7jShs_~7U~?!A*qpKxHmBl*%_X~Fb15#^1X+MhPz2cAvKuzH
z;)cy5dtmb@9@xCH7dEfrh0Q1XVDl+H*!;2|HoxMBEg%PA3n&5DL|KGQR7BW<auBwl
z5`--zhhPgSA=tul7`CtyhAkpTV2dab*rIY2wx|+?EhficizzYKBw2z@QY6^oavZj}
z5{E4zCtyn`3D}Zy61Jq0ge@hfU`r_}*koCTO;%*s(sCNMw33D`BWGaCC>hwYau&9%
zl7%fN=U~ezIoR@Y9=5!ahpiwNU@IsE*otxywxUvmtt6LVD=8(|%5oXDvQmakkrmh!
zMS-m%S756s71*kB6}GBUg>6Y*%C6rWEGbJ@uhTM-mkF%PL|G=VHx9qA9QMZH_bb1*
z|K|^$eJOa!8{yd-bA<^L9>+&MC(MoPnJ^Escf!2Lz6tXo`zOqg9GI{GvN&NPa&W?e
z$e{@fA%`a{j2xM;2y%47qR6ociy=!BCLzZsERMWPCbLI6AAj$W&e><WmXn9?wO&(`
zues{*FA|4O{`;TQ{1+QGtW9gkWj0N0_n-INwa>gSdsQltl+zpJb#Lwe_1~CYKR3Jn
zOIb;m^E3Qt*3Ge_Hoi}_Y0K85k3065z1!^84{qDM*RuV@<9D3&{>v7%I=#VYS~%gu
z{0A;M_l*DFxjRoe^|W2P&phkwbI$wF`4?RHk&88#p1$W~^+mQ5b~<MtHSO@4T}1dl
DyS;6$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.E0256C2451B1EE93.idx b/.cache/clangd/index/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.E0256C2451B1EE93.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3d16fde17d35d909cef6021623122bf04c3eacbc
GIT binary patch
literal 20106
zcmZvk37A^tp}*(*&X-75&0><O(Zn@Y8Dk6@4*M1b4(owL4?71?+fbF@EiT+x(Fl@|
z8bz;SO2jUT3z`-UU98u(#%XS&x7Ak1ac`TtYwhAC_Ik<mkNo;Pf1gLpJMVAi%{XUR
z&cOKX>(^iJ9vicy?1gMPpI~ESV+#iVZqDZ}T|Izn=8cUN?*7$=cRg>dJL*}(=H`Zn
zPrv*puQ;w<b>p5F#=Va|_~|9Rzn%Q(J1?pY($!o3vHg?X-o3Zh_Pz7%b&p>B-kq;$
zcWwLTwxb#cCv*0_i~eB0<+Uy9*T1B0{JVT`<E_8>@Zfv-bB}v}f!=V@vrDcyHdDR#
zXS%2URM_!R;OC#cbMAtN4;=7piN5~(gF7}qwzv1KM<<g1zQ#)B*ugoEeX8&6io6!y
zGjWPPn!I+GJM{G3w?9UutH*mAzbd^vZ{5@Lei!O!kAAHE&_U<CO*0N$_uc*X{ovcb
zR4hN`$8Uc2uIDS={8aYq3m@LGCbzHq^m!jV{3N^k8@GJ<w7bqS9r{83zz_ctDZRYa
zf9HwH4@y6K{gprbHmF_hUq9nb)4_G;-G1g3-w{^qJ#yQ|-~J%=@!P+-<JptRSLeU)
zyI|hV#%}ArJu`!Pe$y4Rc7D2jcI)JYx#O=qZN2$@<44zRnq}U+kNegq%YU;zPw(IL
z(5AvsNBq5f_I<NIy5Y$2o8!Aayy%kUl6e0=R3G2-PT;%Y`Og}coVxI955BbF+^c{6
z%U7=W!5ZV2KUjUd=FQmaXI`}VdwT=4NAH+A`s^F;<pQUF_}?F&ym9v__9r%k&(QAO
zcmAT%yqAyRe`wnO+$pn;y2|?4l{wCL<D-9i?)D9vPyhPn5A6r)AIzGs-K%o=7n)AL
z;f<@;mk;dTwKK5n_#19Nkw5OSa7?=C<J0by7oKVV&ZO7(;W<D3$>QfP+3@t&=Du|K
z#XGN5zW1$zSItj->67D*t9)zebLX_q-+lV7-YpO9yXG?eOAB_lRT;rKH5R+&$_rlD
zdH3@#)vw)p#*tgrucqQN?e&g}>8THW_0xHUyV>_<?|XMmr95-{rBD2*Fp+uxy6T!s
z%?s|I{h?(2(trN^*`=Rc`iXwu*JlTRcaG1|R_%Fy!-L;^;oj4n;~V~Cm%{yk<HlP{
z+g{xE{5@9Xs)Z+Ae6Qu<zy~+1|Ioi^*>ds01>$pWeEfR+I_1%CeSPPfe>;Eu2jBnQ
zU%&sv<Au{MX<zm3e>`;0LzeGuJMy^Y7gYIQ)Ar4Y$5baB)Gyxs)|TJ)*B1j%NweZ-
zt$FDu&kL(#$$z`&>1)?reQx{cbI#as=Lc^eJ>MF+c=4h-f5w;peCYK{E;rus%jf(r
zm{W^R|M<>(?y9Cf`OD5ud~5oTpD#a9%FliC2ztVco4#`8?O#hAcjoq&FZ}8=nG1io
z^u7-teBi*^pZ>W2nDK?T`3vJe+N<2P-FD|IFS?F7bNk;u+HvN}&z3!}U-ZDo1;b}=
zUGYFx`@!G0?|bPGUG?O3;+y0dP4%^x{(d00t90U<e|?5#UO7pn9^ZG{6Mx=#!_7zQ
zOA9xD{{HoUyZ5E%FMoIIW49IOwJ!bMyhBSb3BLH1=53K1ZM#=Le&;63bGFx)8s9G5
zby;@q3HKkl_S?sP|MNZD|NO|ye_rydYYg|yGI{z3_5XX{joO1x9C6p@Z>WA9dVcEE
zQ<OJ6`O{So=-3+V&ky}#>mT)7K6>l2o4)s{_xH1Y{Oo@}GWV=q?_IX;nm;_a`KoO4
z1u`}^^Mt)M%RT3<m_s*)#?nh>E}1obF}P-#{%3Z7_p2qRN}NMC4D|Au%NI;v4E2Ie
z<(MlUe&=U%sBfSb&RjTa`eLY;JrMrYtCj~YpEnj5=oL#=ten0W>Qf?n?zwC1_>bq%
z$UvVx^X$m<#ZVg#ef<7?f4r(Qhh8+$$Id+Vi0O-=UZ4NP3)kHH$=W$IJkY1kJT*0a
zG1NCH7991DplimwvEV?jT(a`y>5HNM&d=VMM-H6&!5r!y=;LP|zj*p$sK5L2)9TsH
zYprwWg#*28=CVc87eoEbOP{1nbT&7K`UiUP%*BhQFNXS^#EKs(*S>$w9O@nDr8Ae#
zn!Xt7_v8oszuf)ct#fE}pf}Cjv}yWcs6YG0^>2LczhKkcF=?Q$Co`@m@&EUq;p87j
zZP@e5Wk3Jp94Zbbi)JpGHGMJEW4Ux@bN>9%S2r|!`m4JAsYAOTd_%KhZhF~@WhV_k
z#{r$PY}JUymE)_2|8wW2$Axjvh}25;>fx7dZdxs<JtN|kJUCw9!D*Bla2i1aPOH>{
z(+XN}I;9SrPSAnVEA`;?f*zbfX#i&s4B(7PBRHdA1ZPs3z?lRSIJ43W&McV0S(FxV
z7Qq6}s<eW$3RZA7r45`-uz|BH?cnT!9h^hy0Ot@K;G9Y)IH%wQCnyDQf*^o%DP7=P
zf(x8m=?3Q(+~7P)4>*tD0q0eE!FdHQIG@r7&L{Z5`IUZfe!&k;REpq4K?E022EYY`
z0Jxws2reiD!AVL9oFqu#Ldp=hkPreFR))cag)q2?G6F6lM8HLrQE*Wq3NEIMfr|+-
zaB*cETwI8QODGfI5<&u8Qkevo6q4X%r3_9MWN;~E3S3G^flDjX;L<`GTt=Azmk~1H
z^2+?`;Z@+=bY94NMpRH1z!ii7xT3NMt|%12mB`u>Iq!>W&h%EYcI$|Tw_+E~p1$=O
zT@8+x$@r~gboDk~C*!+;)n%f-m5i?9)ODiX4a}DbkDjj+9=)baH0U*TqCu}M6D@je
zooLbP%0!1=S0_63`ZCd@*Vl<2y`fAD=nZvZKyNG)BYI<<7}1-`#Dv~dCnogfGBKk!
z*NGXurA#d7Ep=i+Z!HrmdTX6n(c8+zhTc{uHuUx~v7@)wi5<P8OdRMPb>cwpEE6Ys
zXPr3F3uPjp7wSYn?<x})dRLve(7VgTjow`+ZuFiq@u2tAi3h#6OuXp5b>c<uD-$1j
zU!C~S`^&_S-d`tv^kSKa=*2n_(Fe*TfId(s0rbH#38D|yNf5nMCK7t7P9*f9G6|s%
z)kz3_xJ<(6!*voyA1RXv`beEb&_~N8iauH=QS`AgiJ_0xNeq3wOycO{brMIPD3b*G
zM4cqiC(9&>K3OM8^m3WV=;b<*(WlBJg+5g$DfH<wNuy8KNg92oOfu**b&^4!FO&SO
zWOVhLuao?4;6j-c&==~YfWBBJMfAlwDWb2EwN+vmT`g}TYqyPf_~_v#GJE=A!{~Z>
z@ruPK4ew*mO)oxW@v0Gxi{l&M8e1BV0F}u2R=oZe)f?d2TT(}WN<_UCuggUqw!Fl{
zmM;+=w;E9cTaBcFt)@gYxYdeU*lHy$Y_%n##jQ@%!B!{fV5=(;9d7la9=3W(4_keS
z=y7Wh4X`yx2G|-(#DH6)XoRg%GQ!qaB1YVrL=$XHk_on^5;5V{ESh0!mdvm<mxvj+
z7SRG*i)4YVr9>>awTf2QS|uxNttDc`txdGS)+X6tYby~OZtbESwsy%5TYHJvaqAEr
zuysfd*g8tYfm^5OgsoF@!q!<LPTUHj09!#4U@Me}fLoX7f~`w(!PZqGF5J3BH*DRK
z8@BEeapTq_dSL63Jh1hYhzGY`(F<Fz<b|!bM7+55i9Xo+Bp+;jCE~-aU-ZM)FZp5X
zFA+a(MNx#UD2cEYOGLzNKn%b(AO&C>D3JhegJKZ2K`98^V2K2AD~S?pB}sy<R3Z{?
zLt+TFAt?mgP>F<a8y3T`4NGCzhD#)j+lUx}ZA6N|Hc}!H+(yMHY@<>Xw$T!a;x;D6
zU>lQSu#J^S47YJH4%@gChi$w>;<!zS3D_p21Z)!}lE7_JOu{xPC1IN^ktA+qQHHH7
z$*`46M8<7OOu;rKrC^&XkrZyzVj8w-DGl3ni3}csjBaZ(Vg|MuDFfR~iDYn_7xNq7
z7Fo(ifJ!946>p!#0&ELX0k(w_Dd4s!7GYbIim)w~ND;R;k+nCC9zM+-+|5{9M!uP>
zy?MmL(|oIDPoLf!-HfZdMExZEZ{WLxN7i(S23gxBT4Y_9=#ceYqDMA#i2>QzB}QaZ
zmza>vU1CPIbcqGo+9g(GTbJ07?OkF=c65mY+1VvdWT8t0WLKBCklkJ4M)q`x2ie;t
zUSwaF_>lcw;zt&{L_`jBNdP(6B|&7VOC;n_mxPeRT@pr)bV&p`+9grsSeL|*<6RO*
zPIO5EIoTyiWVuUZ<W!fWkkeg~M$U9eX2i!H(w}j4?%gHx=cc7eCQY%?_g`p|g^<IO
zEQ}nPWD(@(B#R=)CRq$QKFQ+9iAk0~PEN8UvOLLT<kTcfA*Ux<8aXq`GRVbAR-9tP
z>w%T!Yd*izar?<+Zu)r5@%G`V?BS<NXHTC-AAK9%MijS^*zjKX@U;5FgWchi>Z9Fw
zjgIf2aBsRv$9E%l==dYZeL8*!xyp@i<KX=p-1yzd6*hh|8+mn&s&~-gUwd#mxk=T#
zkvmlV2y&mQ4<T1M^)?RPzrm^RMy@dR&2SX(HOga;Z&Dt6e24Pb<NK7y9$)2n?C}kb
z#~xo{JoYp-s==P7Nj2Ehbf^Y<nm*NFPgCVI*wZvP4fZq@roo=JMzz?}HmMeS+78uX
zPur(j>}ji<7JJ$Tr^TMO!nD}a)u;}8x+c|OPuHP3?CJVchdo`D(_v57;B?s2RhSNY
z`Wn?^Pv4|^?CCpHk3D^#>anM<a(e9P8=M|{`U=xy&rqWV>=~NWfIUNp8n9>RQv>!4
zRnCAtLxVG5&ro3o>=|p+h&^MI8nI{WP$Tw?eQLy>vC0{-XKZjr>=`S}h&@w{ny_bT
zQWN$}9csd!sZUMVGgUbg_Dl`VggsM*nXqTBQ8V_;O=`xTxkJs^Gxw<(d*&)<#-6#s
znXzZCFf;ZnHEO}0rAaN=vvjBhdzL=6V9!$JEZDO&I1BbH6=uPnwMMPjvo@&}d)5xM
zV$a&AR_s};oE3Z424}^dwZg2}v(=~#d$uOEVb9i~HtgB@)P_A<m9t^b*5GW|vsIW4
zd-fW&W6$2CcI??Z)Q&xSpW3l!uX1+m*&Cc4d-e*mW6x2e4(vIa)PX%mhdQw5=u-#w
z997PNJx7CcV9!xu4(vH=)QLT3lRB~I>`*86oPFxVp0mn1vFB`XPV6}=%!xgrMg{B%
zO)6kd=uiQBLZ1rQ6RMnmJ)yw~*b^#Dz@Do{UD$IqsSA6q4s~JA)u%4(xvHECd#(oO
z!k(+bT-bBhs2h9kCUs-a-Jx#mx%<?OJ$IFJW6#~-+}Lwhm>YYZ8ueh$)1)5kc{<dC
zJx`x{u;;0A9_)D<oCkZJ3iDvkTccj=d7IRWJ#UA4vFGhmFZR4u&Wk;7gY#m~TVY=8
z`D)aMJztahu;=SgANG8G>cgI|%K5P8Yj8g7`6|qZJ%5e*vFC46Klc0`>c^hHPyN{Q
zS2;iS{0+{JJ%5Gyu_xB3h&{1MMeK<kDq>IUQxSV&l@qZiHaHP`Vugv=3)E-;dx0hm
zU@y?20qh0(G=RN8l?z}m(BJ~t3shJDd%+qFVlUXFLF@%PG>E-mp9ZlPta3r@1shxt
zd%+3|Vo$13342nLO4yS+RKlLrrxNz0Dkou2YH$+vqzaR;7pl<^_Cie>!d|FDL)Z)T
zX$X6vDi^|DsKJG>7pkxj_QEw9#$LEd!`KUVXc&9pJ`H0pT;;;p3pcnh_QDkw#$KdG
zBiM^HX#{(b4vk<h(x(yZMXFo`dyxhg!Cs`oBG`-8XcT+VCXHe*+M!YGMf)_0y=aw-
zVlUd@qS%X8SQLA)8jWEu)}%4)#X2;Gy;z^duotUxG3><}Tnu}$3X5ScUZZjB#hWya
zy?BSlu@~>tIQHUIE{?r;gNtJ?USV<UC2BN*y+o5Hu$So21ojern!sM7$|bOuXmAPa
zB`Pd|y=09hv6pPpB=(XWn#5kRPm|b7R=Fhhk_|42y<~+Yu_xE4j6JzYW$eivDq~OX
zQyF`5m6Ne2H#ix4a)rs*OVwx!d#NT(VK3F8DeR^CG=;rXl}lkS)!<UtOI273d+8cY
zV=vvLY3!vtG>yG<pQf>wu5xMYr5jutd+7>GV=q&q8SG`6G=sfNhi0&s>C+7MGF2{v
zy-b74U@uc)8SLe2G{1um&+-hO8q=it-N+r9e+0Qt^M{bDTz(q|@896^cOzF={^l>v
zBF#+~YP5j8LX#G-SLo0J_6mJkz+R!s6|h%ma0ToYDy)FLVvQEDS8UQE_KF=^#9pya
zi`Xkxxgz$84X%j2Vuck)J=0C^js4GG@{=4(-$E3(kj(G|7)$@_2{D$wgDCDGUmJRh
zzkf>o7oU7i=cYHZji(QvR5>?&HM{x?8lJ@(oJ2eP)YH*Pw0~_m?)>SF1Eu%o4L;W5
zV-}w<`YHCZW0svTe9EQJ_`Op9*E=3KWAJQC#rUe>haQ{|Q>a&sPH*uF9!R6m0BIFk
zAe}-7q*v&H3<?8~QDFo!DNI0Sg&D}AumD*VRv??g24q**fgB14kW=9V5)=ZEOW^`?
zE8IXHg$Ky1@B;Z1J|Mrs4<sr?pnxI(6jTI(B!vVNQiOoQiZD<_5dn%SqChc43@EOM
z10@s*prj%RBr9Z~lp+O`R-}P4iVRR*kzY0XhA$`zKt)CI3v!*e^Rgw%ZQmT6GG4A(
zzJB<GxPW`%X@9y!J7e(X5}8#ZM-IOS;%}BOp848}R}UYO6KwoNJZn3_)Gy*$)d|KU
zYbKZmSv$eB$hrxpL)K3)J+fhf8IX+=%!q87U?ybq1T!OBCYS};I>D^Swh3lKw(n>5
zN%(vm`<VmTxt}?ah5bxGcI{^_WcPmNM)vGy9%S!+=0*1HXFg>Ae&$CO_cIYWu%88x
zgZo(!S(;!Hetw||7D5hBurP9Df<=&{6D*1xn_w~I_ymh1Cni_|IXS_S$npe}ky8^a
zg`A#XY2?fV%OK|`SpG%4)!NSrlfcCZRvh`{8vBnHC7#&3d~o3Q*!Yup;P#mMN%%tZ
zJ;ozzdQ5|??J+H~uE%u9`X19G8+y!uZ0s>3vZ=>R$mSk1BU^gRf^6+EE3&P}Y{*iN
zNywod3n7PlEQ}oKu?TXs$D+uw9*ZHzdn}He=&=NHvd5Cha*xT#sUAxqr+X}ooawO)
za=yp%PvRH0$BHAqf{opI|LQ}}9ll0tv+>9A3)^Pu$MFl>W<0W{%{0i`Hq#>O+DwP6
zZ!<lzq0J1)#x^q|o7&8TY;H3%vZc)|$ksNqBHP-`hHP&$JF=tA9LUZ#b0Q0ECLp`o
z%!TZ3GdHrQ%{<87HuEC;+RTUSZ!<r#*k&Sfpv?lv!8QvbOKm10huSQJ9B#8Pa-_{7
z$k8^7BFEY+h8%CRIC7%R63EFmOCrl{CL^cXEQOqIvovz1%`(XOHp@Q_2Wp|s3dqGa
zD~|ZzoYwz3<HB>jgS&t$Y%4YnpDZyqz0$UF<M35VTQ2BFcdkAwIDJFBqD0Oa-VlGI
zxc7$teEp5yK&-fEMSS?)!}pspd-_#Pqr0Eud&u~c@C?ZO9+H0&pE0tB6p)L1NO8nB
zePlaM{p_}52DOF9EL<^sg68(O?)E6RT$Db1r8>#Rr|^4klBw}0)%+ymku{S{gRGro
zT4dcM(;@38nI73N$qdNGNoGVgO)?X*d6JotEtAZGY@K9QWZNXOA={^z{Z+UwaZE7>
zvU7?#k%cKHAiJiR3)wxz+{m6O=0WyOF)y-jiusWJQ_PPnPB9TVFvSAM!6_C*&QG%Z
z=$Qk9$9z+)@ah-G?(nGt%GuMeeuKlPMaCb2!${pC>PPUIJT1Z_Yg$ButZfl3vaUsR
z$odx1BO6-8fNX3LBeJPQOvvUIF(X@A#DZ*X5i7E-MQq6S7O^8cTEv0uY!N52&>{k|
zt3_PM?iO()ds@VU>}?S*vadya$o>}bBa1B}A_rO|fE;X*AhOgV5^|_TLdfA32_r{Z
zB!V1mktlMkMPkVD7KtM#S|ov-Y>_0g+#)h^szp-B=@v;NXIdnKoNtl*BY3sdA_e4P
zixiQ!leOE4d3dEcd|*8|tPj6*6Ao*B{Cs$L&8yFchu1t0q~SF{T3!pJ<8?rKUJqp8
z4M0ZT2xQ_-KxW<yWZ^A9R^AF^<8458-VWs89Y9Xr2_*0Wkc)Q#xp_B`hxY(^c`uNU
z_W}8NKaj|aKmk4g6y$?I5-$OT_z+N-4+BN`2vC%d0>$_kP@IngCHMqTl1~E3ybP4$
zQ$T4x4V2+CKzTlYK0LhU3qVD_2y`P^dm}Lo55NcBEnRid8+X^Zx#`npoi=}T$gP>R
zX8!1q%g)NqAAa};3qP9S&9>GLZhDWMf9#Sko~5<<^6y_ZJ8d1@dsxrbuNs|<x{_V_
z1wFCWpPFH7y*qf{C1));Z}^(-rvm!7V<-G>-{5btgRmXs#Nl6WZh9wSJCW;z)ki!$
zO*Q!bIsDpY`2NX_Zy9|u>YRGZ=#$}dJdh@*0n+BQK)Rd`NT1UK8FB_7W6lU<%9()7
zIWv$YX92S2tU$J$4alCe137XIAZN}AB;*7jSI!0G&bfg+IS-IG=LPcRd_ex3A4trJ
zK!IEUD3}WZNjV89lnVicb77!JE&>$IMS)_u7*ISH2TJ4;K*?MZNY2SXsay&uol65{
zav7j}F27}XfDi6_<O)E=ToI^3)>cN}Kd--ao$;0hkGwoMpjTeBQXbyh{ln@Re>vjF
zrPaZkOXTPhIbnD}4_|OUIA9OI+8z#A)%ZF%U{&gMaKNf~AdN}`q*ZBwbSfQ?UZn>z
zs0=_xl@Z9KG69)YW+02o0%TQLfov)pkX>a5a;O|YPL&f#PzgXTl?%wNaszo(9w4vE
z3*=Mzfcz>ykf;)Y0;&K|P!$A{R1#1~6#@#Y!axyK1SqPC0>xA@ptvdylu#vrlBy(-
ztdfCJsuWOKl?KYFGC+A%ejOaJssd0^RRp?$ti56M6!>=)5AGN|dGzr4)um@G<%h@P
z@IwuLN^tnsJdBRZfBlKW@Gu^{4(^}7?qN6_{+@@8v!{Q*!|=_6hhg$~W^`PSXU8ue
zky=(~Mn}0itG;|hyv&2+vphJBtO2LVYQSk_EjVpf3r;8N!0EC&aC%t}PM_6-Gsp&T
zhO7acQ8t1zW{u!XvI(3iYXWDM&EU*gGdPQE0cXitz*%K0IBV7l&L-Qy*|IipcG(Wj
zp0$H>$PRFhtOJ}=c7k(eo!|sn04HPxa4y*e&Xsk6bIWdU?yMV}NA`g8WIf=#vKO2;
z>jmeNec*gqA2`442j|cF!HKd6PRxqn0&)OcAR7P|l!M@c*&sMcmcU6_30z1HfeU3r
z;KFhkTsRvB7m*|2BH0MIs2l|s%|^k+<QTYEHU=&($HB$3ac~Ja0WOhEfJ@3raLH^E
zoGi=W<g5%XC8xlpvMF$BISnqIO@qtG8E~0w23%gwXNFf^gC9g^^OujPAQ!+DvITHO
zxd^V9Eq>wt_a`sE894X-4TD!_*n7<W6rN-0F$c1<$DGJQj|s@G9&;hPd(4gO=`jzo
zx5vE5z8>=-`+LlfEcTd)9O$tCa<IpO$b}v&JoUvnm0WsL;_%({%{h5<?C>ArhcEv5
zPjdK&v9W12ja>SGMiu*o^IF{znOHKOkW*vwXgB?_{)ws4TQj3S%GuPWO9tr7xHvXr
z<`IfnbLP%F>d5)_1vbm63ufER3zsfhyzEQs)Ecc$Z!nG@eFlHhxvNh2|Gjm|vBxc6
qvGT-|PdV-MFQ0kV>a)*Lop=1&W0Y&GM=v>l@GHUAdrTJ);r{^n{Kt&|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.EED2788EEB0B75F1.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.EED2788EEB0B75F1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8b024a024dd3644e1aa4d92ed950e905fae3a81a
GIT binary patch
literal 3550
zcmZve3tUXw9>-_)?3t!%^z!I|v%5;N*BYjp={*&Z=|p+-Fq0!jAqo|}L}@4_j^acQ
zI?0)YoIEm#uDo(2C#guM4w1(ZQe8RWx-;F+Zm;|K-1&TF&HB!7t@VHIHM7vi+q-oX
zgW(nClMofYaXHRlFpM<s*o`p}cUTNYuO5SuP|y(6zOLQLu;tsB6*u!E9mGFoSY@>r
z%~K4dxfzR`vpK&=1wPInwtm&$6k%MQ_^3H#jL<4Z-#PZhnsA0$D&RdU;k*vqBp+w6
z-XM(gewzF4`uubK{<&Sr?vhU90=M?xO+s{DWM22E#OvmSq`f4|*VS0pHdNR5uYaHT
z;cfl##mO~=ndfe{ygtC5nXH!Yys0{Oroa97hh05+lRYr^B})db1}?2K?<b`K*6IFR
z9mh<XzCWisCZ_wZTkI;4P0%YrUUFOZeP>y?JnM}5+faV`PfLdyj-|I>=f5kl{qp>Q
z`0;IUK947i*=1dFdhniV(EO`)P+0x>#GOZS;)9(+>o@L#v$?BzOG9*gJNIW@ec8x~
zsWdt1xW4m<=cLvj9%p+h<&ULnQVYU!{HM8g7N>9UUl3SS@#AT;=GBpIq?uW3>7D_*
zzAM7i@@@;yXFt5{ir@6<>IBP{+m15_PVc^#VWfEfC_k@q&`=>cH21=Uc<E1-?1;jS
z=JHXEZ???h%eKpt?$sUZB?k0cn{t{<Oam_;tchIY!kH=FadlfpTH|VIg>}Tv>Fb{M
zC^kkvcw%v}$1KCjC~b=3VTq7&ur=Cqv&{0s)+HH?9ai={H_pJ?)>(NzMh|)|4MWzq
z`HvE-GEN-d<EN6ClP{gWZAixN&sfvqcPB2s>*KMK%DC{3D?2v7j_NB{Nfa~2aqBB>
z8okDc?Xqm-Sm)i5$vfsG?|E@<w9%Ve8*UC<+k5~lu4@)dl`2E0uDq=GDLykOTG#on
z^ZaQW?hom|zuiA6iM$e7R#JSSXYSzEySeW^es{bqdvehJOTFc*4lqr&R;Vw%I+FWk
z?g{gi)@yyPaD1ON@BVqe1Rj@1Z?bdt65o8F<S*gOJN1Zn^7-AWR{OHwj=qTMjo?lo
zy8p1OysMg%qtx9OaPQBEJME!nZeeHgS3Qf=UD|!(WuZxVZH)oD&2X}gaC*MO^onwJ
zaK260&dGTnJwm?dhE&}WPaWf$n$YNNV?>tQRGx>w)TQ|i#SfMSE5c`YeDxkbx<w?r
zx9RX^)tXPSF;336bHn;~n;$=$xwh(>^Th=TZ9VaU1w~(T^3^>OP+jx4dfAqUxRl?r
zyRMk?4$B)>w(mOBkYstTEn%HQ{YS;cS(!(gM}OWNr+RXrq3_GVl)F{(w*le7dnENW
z)?Ui8eBFs|C)9iY9eVOcf0t}ai`%Lb%O#igH5ryj7DQN@?f#?q`l~Ny7ebB-e7@Qw
zRee4h@b|Q#`nJorxBH5>Bt(9!eKlyB;Wl(Oh54jUo)w!IrO4Mj*`w>@3fjA`ywk%1
z7Bb|P9Ls>=L8A=6JU)MTTlymLj2<>$BO7oGbcP2?4%k`Q9Pgt0md{wAk%eqw@bEy%
zNp<O1bb0lyQ4D{LY&_1`YIvaJU%6M~;;K!p^sq%5*@R=VWO$%tMojq1*c2{<!6DRp
zy?%DIFyPAx0U@NeX-G^sAtWVyX(9!rkP413C&i=_LQ>L~CLjRFN+1TvN+1QuN+^KH
zN+^cNN+^ZMN?Vbw<4AJkb|O23L?Y1$@rH^zmA>LdJWVjEOd^v>2S8*oi9-mJsZ27H
zNxMyEG08&+gH$F!b&$mbs18$^5Y=H86Qa7a$k~H-*+t|sLX=I6H{2Up*FG1Wdg^My
zm(S$R45GGJ>93ZIt-1W1t%<~i@S@`&T}UtL1il=&0ED0mM95xbFQJa38gC+RxZ!A<
zL^?)fVD#}sE=C+eS%nd$DEDDRKgvoLv7LqN&tVY<P~L_UX*jKu=@^-T(E*UT7<mX~
z6-Jh#+=r3<C@Wdyb{4WfheaMhc^giq;j|9YF@V|wxfnq00Tl*Nd!P>ks6C)$0n{GI
zVFA=0*oFht9!$p|Y7gdO5VZ$Y7)0&CJ`AGvpppeqdoYIuQG0M34rzNcGI|()4{Usl
z;c^eI$5N`=`2V!h-2CSP_kZP{#5JwNC2Wvda7h~^Q(S-$<iZgO@H3wBYQGzs$(P&e
z+8R(+s81$u_?Du2YFXFLRh&O9I$4@)q829`(PGr%0A*N<Lz)$KKGma_*Qm7{D{xYc
zZmhro%CG{5G_P?@*iv+3R*Uvoi6Y{W0HerwB)}+u5R8HdIf@+Jsd%Ds#^Dpf&y89?
zigEG`>PImSP=>`gq&d!a)?KjJVZJ6$#<Pussq32hWb=lvYbs9~XUmv`8oiRrn4}uL
zg36cxWmv|9G^a?1mSzj52WqQgVQWu$uUuyCebUxGPJ0FV5ew;L5J|*Vnn*u#A)OjB
ziQGyP@B@ejk^rJXKZs~B2_l+UBzB_Q;LDvvP9wA@ec$-zUI$WJp<qIWxDias5W~R$
zAs7r1au7K<P={wdEASVj$ilT1YQo88v_efdKp8gSkmi$bl!>yZ1$!~R+*H?8pZ3FY
ziRE(2k5dWN?D<i}-kQW3=o^fqu1;#P=GxZ&w$f!ppS)`U?yp0MhjkhQ5?nHga+ojo
z#63so$+J|?{YJmGYx_iIyn^bFZXtU>$G{|0ThOt#1hsd79NODcqR!Rku8>KTq775X
zBumkTDP#haVIdQu>?(2{`4#$V7<9H{{f!q~&9R|`f{IQ<5kw?SWGJbi5~Im>1Q|&a
z2nC1+A^@VnP>5(S0*+{!{~?jfwSP)tV^+k@(M|%j!+&4F)P%tZqNg%?x_D&Dl8=?#
z`DcB_T5tAZSzu5wV?|V0<Z6xVyK){2XLEG8x_bOkh6egRMzg)Fjd(sYjZH?6F*S2?
zc6F1udw5P83tTK6Y>5B&?YQw1CR+SslGWrXQ%N{oXlpNa{7z&pu$wh@+??-2FmJyZ
IQs&(M0M}69J^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.F71CF786157F3658.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.F71CF786157F3658.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eae72dcfa4a391229a9925f3ec4cf27c9e381427
GIT binary patch
literal 22874
zcmZvk3zS=Bp|<;d`)dibQd$HQjt&;kPLiGbPD052x|1Ne@7%LNC=ex)P>5&{wuU<d
zgN9R1xCAg9AZS2P0g<t&=WyT`$I3xL5kXK<!EjJOipGE1wT5}mzy7X8);xXQnWrsD
z_D&{~bTk}J7uT%|oEAO5vE$rR*t&J=wygZO^W1al&v5J3#W$^6cm9n}O2tnfWjo~g
ze;#`E&n8vdmKP3?V;BGM+kNiBW4E8v-gU^X=icalZNB&5x9`<1UU=tM4`2W4VP~BF
z&Xcw~2IRDL-`@Vs-)?>I(5FL1=jZJ&e(<@<MW;XT_O86#eD>*cFIc+c{O=u;_<cuz
z;_vVH-fN3LUViKS4<B;>yR&HPqMe;LUpnSP7hdr2j=$WqInzFVRJ}U-=reabd*hSe
zt6Vs8U+~DOyFN31=igp@>do`cxc;}-ecgDEbJufs=bnE?aQx@{ety(F7wFFVRblyv
zBcI)RlkX<|Z9g^T!>?@rpZC4}>kBh?U-DA({?PgMCttYsQy))!?&0vEuUs6SP>X!a
zThViW{GGo(<bUA6>r6Yg-u%U{{>F9q&z^Qx)8X~6-?`Co)W(fp9vtYM-TFUI3{T>p
zf8^cQKfLuDZ~gS<4Mz^1W5<2&E5gff{Kt;?w%F^RIxM(;>&6eCxA~ZFw6>pr^!>qG
zAA0Eve}3orGmklr9C^m4j=K59*RQZ|$Q|>m*B=dS`PyH)SKRj1Ge7w)eRuPDGt-;r
z*pDs_RnHafzpL@rpS;-QOCP*=c*5i{>764S8#RA^|E^d+*Z;5T2Q$Zi&i}19Zr-~2
z5<mZu-<Dcmf8{qf-{jw9J?qF<-@J6&R|m)XU-vg|7;H%1{*A#;PWt)Tiw~SJyY+?G
z=7TQ#@UO1y4~|V=dFF$U9&(%c{dZhEd?N7Q>Cj^@{rHcXKixS$`@#*6`j32CieI?-
zjkoT<EquSF-8lKlo1%Np`{n^R{C@rSw;jITarw==>>s%8W4GOP(?i!*|MRl5{`}o7
z_ckr(op?v(82S6^aXlLS<4dvQk9qv1(U#sVhu=2|{ULD6zn}N@M~+Ke_Cod#=Vy+)
z{LJmwzqPZhfA)!kAG-6)AN!)1`P(zbCzK<<cm4A8m%ry-nY;Gh6y09`*~zzlz3}Dw
z$(uLo{`J3a3Z8U#>5{9yvgxc1AHVazHu}=PJ?q@lu6v^T@}7sk$$q?i&TvcScQ4()
z=kd!W_0zL$Prvr1O*h~7z-#xP^1%5Yd&BaVpBaC7vhA`HKe%8&&Y-{H`dx*eU;Ull
zVb5K*^}N47_Th^j&FfmXv->!%^X1`zztjA2`y1J-mp?P#C6tep9=ND`*|ld5wRcQ^
zbkkwN-`$Q6-SqOt9dE=0|7kz}%mr6uj@<Uo?>u$JcJ=Z%?&_WO*0r}Z_g)+O{lo7r
z@ARK^!r#mPTKShfw=e(j9MiM<$4>j{L)V{p>DwRu%B!kc-15rc{I@RL^rObHkFDEz
z@c$8i9n>Z_es{M{yz%9Cj{4r4JHO#P^PO)$b;1+XAH2IE^zhlGzubA7^&g*kL|qD>
z@LD<Z*j<;NecG=3+^0rv+q!tW`X_&XyYo)|<hL(8^tiwI-rW1i&l(+<{p_DNK6&?v
zcYgY*dxlRMKk$!*HvihwuYUR+sq@Ot&o(~s?8@!@z>^=oVByM-?0Db0L}p!O%i&u-
zvj1Ym?EmF?_kUiy^LL-UXCqCm=*=58AF%&oRp0#7j;~+x_|{9`w=TJ&w;sOr*!>r)
z`qm#jF)APM!tXY%ORwn8`c7{D#j3u4?;|(F|9#6N8)<e$AF<(x;{J<O{m2hKPhUPl
zb?io(ThZG#Y)kLISk-@<#Fm~pe@A>Hl~(kQ4LeTRf3d39owMWgo%d$et=q8h;5C6;
zZAX7{Q)NL|u&+i}=){G@8Z8Qo_SNF1%A&X!Un8}k2B#L);532;oJQ1u(+XN}T2Tv*
z7kF^I$b-`fI&eBs2Tm{O!RbXkID=pSXAlkGjDiuIQ8a=x2_|qR(FD#cn8BGvGdPQ2
z0cQ~{-~>ScCx`+#t6&9Z6|LZGf(@Kaw1Kk=c5rsl4$dh!!8t`IIG5l8=Mr7u+=3gN
zTXci-2p(`A(F4vac)@u^FF2pz1LqTc;QWFgoL}^V3kU&l0WknBC<MU;#UQwl5CRtx
zL*T+f7+hEkgNq0ea1k*AP839Nq9}rk3Q=%TF$yjw#K6VG7`V6)2NxIP;F3ZTTvAMe
zO9?4(DKP~uEu_Jv#Wc8#kO7wwGvKm97F<@$g3Adxa5*stP7)+=k|=@83wdyPF%PaF
z6u=e40=S}31XmP`;7URXTuCf}D+^_CWw8vdB2>Ut#0t2oPz6^NtKe!v4O~sEfs+Lp
zoGi-V8bZUqdY8PZ(m1h^SfiHE0@pgR1+Fc$!L?6pgZsSY^S1YnA(e|P7rjTThb`kl
z`wv&W@4tJ?*Iqo~=#@L}30eFWfxGSrQ6p<6M1!oI5G^u4Aw05fLUhRb3DF}PCd7bj
zoDd_jX+liM<_R$)TPDPUEKG=iY@HA*vTZ_a$o2`bBReO=iR_vX7qWXo+{m5@@gRFA
z#Ea~k5FfICLj1^q2?-zvCnSg*nvf83ctXO+kqL<)ixVOuM<*nT9Gj3Ba(qJK$jJ#w
zBBv%Kg`A#{G;(G_GRWBp$s*?_B!?_bh=iP<kUVl>LJG*m2`M6%CZvR1o{%zfWkM>*
z)d{I0*CwQfEKi8M#s}@a<@s)0$m~1UxtuI~g{+<5EL=wxZUJ6Y$l_ISLs{&R#p{99
z3Q?n1_lO$3Mj;yXnjX=h*D6GdUfUyD^t?iN^n8!-=yeLwq1W|@4!vF>di44p(W5sg
z#DLz=BL?(Fg&5Hrd&G#|q!1H&Q;(R?n-yY4Z|)H@dW%9V=q){BK`$soKri%&fZnPQ
zD|%~>Skc=QVnc815gU5DLhR`6Jz_`iREQJ3vqzlhT?%obclC%1y;~t}^zI&UqxUGp
zgWl649`s&?c+q=%#Eagi5FdJ9kND9072-$l?-4)xfI<T313eNzA5=&XeXvJ@=tBw#
zp%3*)2z^)~Vf5i138RlFB!WKDBN6nXLPYdpkBI1_3W=hR_DB?cOd&D!u^x$`k1HgO
zKHej7^ht#z(I<N(i9V%}6#7(;q|m1ol188Iku>^@LNe$xJ(5A6RY(?nwnwt)a|+3!
z&-F+Sy`&Hcz0@NT`n*E&=<_|2M_*7#0ezuI3h0XpDWWg-ND+NWAtm&s9x0(OE2NCR
z+#_Z56@^sLS9+v^zN(Na`f86<(bp7GLtpEW8hTkFGJ3g3Wb}>8N#iSI?HsLf9ckPG
z+`62!(6_E5E%fcnNgI9pI?_geF|k}sJZoofmk`S(YrN&2D~>#G`o|xP$jXr_vmn87
zs<coA>XL;^@u-zpOs}2XElP_;pw8ljYot!7VX2nXuvB-{u+$_pu+&HzSZX>NSZWhm
zSZXCLEVUgiEcpZvOJ3q($#-~I>JmCw>LeX3bsZfn^$9&J^^zWz`i>r!hJ*o@2FU<R
zL&pG1W5Nhaqhy4ov15d#DPe-8NixCG)G@)*oG`=EESX_x?wDa|NmyWMku0#ZbS$tG
z5&|p*Nr0u$5nyReSYc_Etgy6ptgy5tY_PORHdxv^Hdxvdc39daJ1p%TJ1m_ECoG+k
z6PC`76PB)o3zja)1xr`Q1xt6r4NJG=hNZjXhNUOrfu%?Cz|zz4z|x!W!qO{wVd?F7
zVd+cwVCj>5u=I6&u=FSVu=GoQSo%ADSOyXSSO%m3ECZbYEQ5(4EQ3-ImcdRCmZ3xl
zmLVww%TOl-%WxtL%dixNWw;ZDWh4=SWkia=GSZ2_QcQ@j6eSUsVn>8!G!cbmREokf
z+KIw4mWaVJCdFVG>%?FgPsCvvm*TLDcjB;2CX%pBN=aBIJ4slk5-C`wq!cVuofIt7
zi8L(JQW}=&P8yb(L<W`_DFe$)Cj-lDA`8o`l!axslZ9n2k%MJU%E2<%$-z=eNU)S7
z36@evf@MCDhh<*M!!qB=!?KVlz_K6}U|HxCU|CEQVOf-luq<|puq-7?uq;U>Se7~^
zSe6rISeB(SEX$oTEGvl$EGtq4mX%HgmeoWRmQ|?=%W9_z%UYra%bHY!Wvx?#rJRsq
zDN8ae<&F%?My4UbIZ3He1?rN<rFe#tX~DKtYQeVEB`w^xGi}(mOKsS;yQGcVl?QGs
zkFD0uT)sdoUqJ2>OLvWrde^P_+&O<be9NXva+xLfu(eZ=)G|vUr<Yk8IkU_%$k}C<
zMb0g=9I~{`B;@=u%Oe+-Spm7Y%!<gRWmZBiFS9aoWtmlwtIMp4Tw7)}WO<p%$n9US
z_H%6Yf7r1HpXoYm>ANorD`$A_g~-}D-J)YLu=bCncBp|g4h@jjp#|a{Jdn<z1JXP6
zKn8~a$mlQvnH(k{v%?Hzaae!^hX7=CSb=N~8<5>$2XZ=`KrV+1$n9_gc^n=fufq%E
zbNGP#4nI)95daE0f<Pfh2q^3b14SGWAkiTLMIBL~m?H)hcf^5`jwDdZkpfCP(m)wU
z1}N*u0_7YzAju&C<sEsTf};RbbQFO~juKGWQ3k3wDnM076{zN@0m%*-sNrrzR<FdH
zDlK;lsO@e8U1+(`wR%6X`km6(=t29xMY{Hs^ujfC;Y)Pw`oA!u3%4NOLl^Ev-b)wW
zMZSt#xSE5{zkyr05%~(Xa3x#wi(R^SF<rZbEe`17b;x77csufpF20VeaEo8y;Pd<3
z;<d<gZt<VM>Mm7dPd%V&?5W38jXm{@s<Ee5I5qaveNK%%^_){<Pt&Cu>}dv6gFVfd
zYOtr7Q4RJq3a7!Irq5}xr<rpa>}k7Hi#_dtYO$vsQ!VziGpfa&R^hbR)Al(n_Ox?O
zi#@(edF=56%43foQyzQ#jPlsy6^_Rq-{*Mj@pF#Ho~}!E*wYQD4tu&W)nQLJqdM&A
z6i$adU7yooPdDdu*wc5Z9((!$)niXTrh4q@XH<_py~63Sr|)xm?CIy69(#r^HDJ#$
zpa$$2#?*j4!;BiRXHYl;_6&W_fIY*UGholyrAF)-2h@l?<Cq$;XPi+Z_KXT=#GbLw
z8L?-ab4Ki$y3~X{(}0?=XBtxz_DnNs!k$UtOxQE^ITQ9wbIyc4bC;U2XC6>9_RM2y
z#-4db&Db+5oEdxOK4-?BdCr-!XX#Q4_ACQx!JcJIE!eZns0Di#g|lGK(&sGLv&=aQ
z_Jl4KuqO<tfIVSM1?&kkDqv4gI01V?pA)br%sByj)-JVT&pM!1>{-XuiaqO$TCrzU
zI4kz7ea?zK>zuP<&(@_j?AZp?hCSPu+OTJvQ5*Jb3TMNft<Tx8XPa|2?Ag22jy?N;
z+OcOJQ#<zTGit}4UE%E5v-deW_Uv=cjy-3WI<e;*P$%}BW9r17b4H!mb1Iw@d(J-R
z#GZ4`IkD&JQWy4I1M0$_YfN3(bIqs=doG1@Vb9g)T-bBXIT!ZaUFyc3dqCaTbC0PT
zd+r%^W6!N{ZtS`HoEv-YIp@Znr%OH9^9-m5d!8}%V9zt79_)D(&VxNqpYvePGv_?m
z^LD8hd)@){V$VCKUhH{i)Qdf@!g;ah?Q>r2dFPxLd%iC9Vb3?9KJ59%)Q3IajQX(W
zQ#c>?e0|P`J>Q)3Vb9;Ce(d=N)Q>&?nEJ8jpHV;d{0isCp1;revFD$2e(VLhG=ROp
zfCjJ^7}Egu0y7%GUO?dj*bDTz0QLfNE`YsYmj<yH9MB;4f@2!QUT{W(*b6FL5PQKs
z7sOt0&IPd->e3MQLIWDYUT921*bB{Q2zw!g3t=zR=R(*E&AAZv!d)81UU)#m*b9$o
z7<=Iv4P!5?aAE9)`&<}%;W-z^UZhJS*ozEk1bdM&jbJY_qY>;y6fT0jNS}*fFEZyM
z*b}={#GW{yBKE{F6|pDIsE9pL;Y93-eNM!lIOjy{MY}YLz36~Ou@@cFDE6W=8pU2z
z;iA}!_PHqbqH`{ay;zsVuooN981`ag8pB>}Mq}8EDO?PDu|5~WUTn_Auov&rIQHTL
z8pmFIOyk&#&uAQbafORxFW%?k*o)7(IQEiVn#5jmK$F-@j%gBm$r(*zFR5@z>?QkL
z5_`!xm&9JGOH<fO4QL8`sWDAqFEyhn?4=Ygg}qdtOJOfH=Tg{9cWD}X=>bh+FFmGd
z?4@TkjlHzOrLmXpb7}0Q=Uf_lnJ&#>FEgMS>}AF@gT2g*X0Vr0xD56(eJ+E&%$&<$
zFWaSA>}3Zui@ofaX0ex@(Jc0|3YW!Rw$Ej;mz{H2?B%*NhrQf@=CGF=(;W74Gn&I*
zPT_Lc%k{Y&_HuJBhdrrFCG1H9Dq&9=Qwe+0j7r#(6i&jP)aNAZNpnuZUcO87*vk)S
z9((yQ&0{Y=qj~J*6)umxe4oo>FF)t<*ei5t0egi3Enu%OrUmR3X0(94g2ENBSLkyE
z>=ou*0ei(REn=@YphfHz$FzvO;*1utS5&wn_KJP3h`r*RD`KzIr6uf@2DF5|(wLU8
zSDMih_DTv@!d|J*m9STub0zGRyR?kG@_?4HS02+c_R2F_#$H+B%GfLSxia?3bFPfN
zN|#o!R~gU>_9|mq!CqxXE7+?jTm^fTK3BnBWzJQwSMAa&_NoI~#a?wxtJtf~Xcc=^
zg{xw(+UKg+tIoM9_G(>P!(MGbYuKxeX$^a|8LeTjrf@av)%siwd$l=N!=BuwGWO&F
zm9Zy}sf;~&MrG{D3MXSv?sGEs<T)o}uW=1+e2K0-gWDL<#x2PA(8j&Uduiic<g2*G
z)f{~O4P4_!<SSU?%J;s_vGNS@HME7j)`+&S*Sd$cu-Dp4Ti9z|#kH{4x`At9uXP1$
zVXu7+ZDX%JqHXN8@1bq%wfE9C_S#o*ZS1vg;M&-0U%}dIJ>O-ou6yUN=l+6Yl`Dzr
zN>W~Z!kbmDCWl^4<Td^hQGJP=y2^p{-s7(OkH7qrZmN8oeSC5C=W;excCwxC(TO*_
z%)8&*b+={Zn?$OG?eHYFYH|D8Q_yOa8c3tk0BKcPAYR1-=~Ox(y-E*cP#J)XDkG3d
zWdbs*%s>{E1xQc{KvtC%$fmLZ*;RHRr^*TBQn`TKDmRcv<pJ`lyg)vc56G|b0|it8
zpr9%U6jFtN!m2P(L=^!NRU%MS6$Of^VnA_K94M(u0;N<bptLFtlu>1XvZ^dlPL%_a
zR1#2Ll?N)Q3P43w5vZgp0hLu{po*#jR8>`hYN{HLtdfBms>b%UyNH&m1=Lox-=pkF
zTR-;7wa)b`uXDTLzza67-Z)PF&E;oo`1R4x@4MbV&K91*Z{IAj#V6sr$m%7gM%FAb
z4YGELX_5IQ#v|*Nm=0OL#PrC9C1yZ2E-@prX^EMT%}dOTY*}I!WMPR3$kruhMYb(5
z8?t?g*^!;QnR6Nb{;u83h3wwV+{m8Y%!BOR&AiCI-OPvV-_88Uf!!>C9Nf)<$f4aV
zgdE<@!pM=`EP^cVW+HNQH;W?2cC#3Ad^d|DCzn_f|2<MmEQOq2Vrk^e63ZZGmsl1#
zx5RSD(h`%9^Ghs`Tv%cS<l+)5BA1p}3Awz)%E*-^Rza>Vu_|(HiPezhB_<;`9%qeb
z@bLOLYazEEXYDoqgem&)p{Fc8c+AQT=^?hT8*fNcw)ic$BB`fLjjWk64YGF1w8;FF
z@yNO<(;@4pOpk1sG6S-4%8baSDKjCPr_79OnKBEqFl7R=b;_*BwkfkA+o#NqoSd>G
za%##_$muCdBWI>8gPfhREOKtja>&w@Nyzyr%Oe-2tbkmcvLbS6%1X%PDJvsarmTWo
zow6!&ZOUrM@|4NQjfYrcH(Zh053%+dx7fOWyUXzIPxjpdPuRi(c*Pm9#joKNXT;RV
zni10=Ye!6r%#Rq4tQ#>MvVO$$$c7O!AR9-_h-?}$6S8^4%*d7zvmgs2CLmi!%!+Iq
zF&nad#O%n<5pyEDM$Com9x*qvXT&_n-VyU6`$o)%>>n{da$v*)$iWc{B8NsSgd85R
zFmhzXBFN&1iOA6riz3HHEQTB(u{d&a#FEIV5lbPbM=Xt;8L<p<cEqyCxe?1DOCu&B
z=SM7$To|zea&g3p$fXf0A(uz2j9eM93UYPCs>rnwt0BuHCL=c{tnmQca9R`ALT*o3
zdyU)i;Eyx~YjWj0e@kdfe)T6aH&wQVw&qvgAmAc#Z|%zleQw`r(9su|S5Je!@c9QX
ze)NSGA6ohCEyXP-uYPXtr5h9cMa3Qa?$m$57M{b~<}zEv*PN-BnHpKM%rwZ_Wu`^u
zml=<&TV^_B{W8-d8<v>?*|^M%$fjjxLN+fmGqPowS&)TgCLmjvnHAZ#%xuW^WoAcq
z?qSZS;J(SVhq;j5dzc&9vxj+*y?dA!*|&%Jko|j@A33mx1(1V#SP(h1hlP;CdsrAb
zvWG>G#XU?!j_zSm<k%h-LyqrZapcA?SYz!qhAR)gf5BR7yn1z8f6)H-p@7$5-A5L7
z!_Sd4?jw!e_*&BYNDH}rA8D`g>eZDTwEz7pR=@1+7k}u9`~G&*Bl}+A*j~5@9=tXe
zKLZb5n`$6UQv;-JYJvDB52S1Afb>m0kfCV+GB%AsrltwV+%yAOnie3TDF9iURv=r`
z24rvAft*bzkgMqeayQ*To~8%L+w=nYnm!<Z(+?DA27rRiAW*0o0tz?7K#^tyNNkEg
z(Pk7V){Ftgn{l9IGYOPxrhwATG*G6Q0m?SBK)GfPNNP$z`DPxd&@2EIn?;~fvjkLb
zmVqkG3Q)CK1*$b`Kyp(CYP1^{!GqU!3#i?01AUQLzDVACjpN;a>VCHP;*HmEn<^hT
z@B;_0UE#Ntw#jQ(_|p$O{ovIv|F!cUUY~5-7GAlB-+b`q!&hG=Io)~o{hvN)M{wn!
zaXdSI``VL3P1bymR=@trD=PQBIn3(KZqxtsMAT|{Ww!54V%BE=`}wKwXWf607k>WO
z${+r6!Y(J<R&UjtR(=kTU4`5uthdH5yrg)q-E!|QSN;lHi(5}!y^r~=Vf|kZ*nGrQ
z`wnS`U9hnF)&6+O(cj@MubKDVDew!1)xX2aI}`AWhSfjnriz+Z18H~-ke1g1@jMTt
z<8?rKUJqp84M0ZT2xQ_-KxW<yWZ^A90xtkrc`J~Ow*lFCJCKui0=ak>kehb{d3X<y
zm-hnscps3T_X7p^08o$*0)_YxP?!$`MfeDi$csQxJ_;1$V?c2}4wU4RKq)>2l;+bw
z89oD)<+DIJJ_jW65>TGc0~PoJP?0YJmG}}+nJ)uX_zF;!uL9Nh8j#G(Kn=cOSiRM*
z99Q`kP@8W9U0}K3y&sSGt9tJ>uZ|zwweRYFJ6X6Fj<JhFvUoebqri};ku^i2LDmk5
z7MULs9$7aeI%NHj=#dRWVn8+yi4oZ}Bqn6@keHDzLt;S|hD1QN4v7`nHY7G=`;gd?
zokQY8b`6OO**zp~WY3UzkiA3VMfMGe57|E?e&oQA1dxM65=0ISNeDSSBw^&pkVKHh
zArX<ILlQ-f4M_|+J|uDE<d7thQ$vzMP7g^MIWr^~<m`}Sk#j?mLzad_Le39K9=R|i
z1?1w86p>3qQbH~dNg25^Bo*Z9kW`UtLsCPQheSqh+)f(z;yLE+q=nqRowSiJC6-Iq
zenNWnrnquc+xOmPYe%*J{au!S93tKKt55&%cW18L{!U#ueeH0w&{#NojTYs_Ppln$
z7VC>=t&v(*gHzYl;54!ZoTjb;r<Jwfv~?{wUgp8^bsn5f)`8R2b>Q@}9-O|e2WOBC
z;0$#GIHPO?XRI5+nPd|<Q{4p4EStfZ>t=8k*#gc|w}2C50h~}5z*%K0IBVSs&L-Qy
z+3GfMcG(WjUblmD%1&_3x)YpBc7b!%UEtiZ8=SlD2IrAI;5>B?IIrvl=dF9e`D7nB
zU)=}JFZ;py>wa(nIRGwD4}c5GL2$u(5L`$OfeY0`;KFhkT(}+v7m*|2BJ~J3Q5M08
zbrD=tj)IHUqu^q43|y=p0~eR$;NtZ-xTKr}m#injrQ{U2R6PYQEvLby>uGQqIRh?J
z&w$IyS#a5U7F<rwfy>o%;3QcBC)Fizc{vX*U(bUp$OUkPdI4NfE`lr8i{MIf30$dO
z0#}yH;L7zfxQbi>SE*OPRplzUYP|}sCfC5#>NRk(EQ6ElGPuU6jnh}}kyqZ`tkF1o
zjasL+z_l7JaP3pu;M$G$dv5;7FMpG@z9y~w+^BQPoDbl0I#cFCc2Aib*)wGxWbc%D
zk$qF<L-tRZA2~2(0p#G61(8Ej7D5hBSr|DoWf5d?%0%SoltqzaQx-#xPgxwf^$=@)
z|GnoM&)IR-Df=D|?mXwzosoSX<o8|d|HV~4`j7Ti??kpexV}`*7mDjv-ne-6S7G&o
zb?eG&AC}fWJhgG!S!b?b{Pd4=Z2g7<R0nR{^u9wj9~|8h3Vd|SK~ewKBMyE42R>*&
z{sgDX?eXdk(>RW?ZeRHS{qEsM{mX}r{>ZjtkNeo7TB|o0&4TU2mLqvn@UX)pu__Ip
J<P*t>{}%~JpH%<=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.777F64009D887A71.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.777F64009D887A71.idx
new file mode 100755
index 0000000000000000000000000000000000000000..633ba972cddaf3ede8859830955406385fbcaadc
GIT binary patch
literal 22876
zcmZvk37A`Dq4)c~=WPwNQd$HQ)Ik>MBstkn5<>R%BmuIY>^UH^L`g6#5e?8ILH0nv
zV7XxlLJ$m^BFhCthT?J&z7NOhMHZ0@f*XoKMDGPY!=3i|hWX#``T9KMnO}eZ;n!)C
zoHJP_9Sw)G#dYffpN*c^IOUv^*t&J=wygYj_Bm(OpXb)Ci*H)D?!23xmWpQ{YTN&%
z_Yb)0my@b(%gY~<V;BGEU;EsJM{YZ-y<`6!FTU0P_Wa%b{_`H~;)QqJe$e_qA9&iS
z?>ue0b3i`3?q9dv|J$t(9q>%3=)A!G%Ew-;Y(MqE|J;$6o6kRU&iPAsp7+=hiQjkh
z$Nm1!$KF`{>GD55^2q-8y*G=twx8Yk`>RKM;==PEIpuFVH)q<%jH*{fAAj!7=WlxY
zvC4%b_xX>WyyJ`GcmCy-U;O>t({A|f_1`h>bnbZZ?%Yey366jJ@Uw^RJYRRluM5iu
zAM*Uxn|(LyZ~M6^AO6F(zkc{X|Kq~U-CupRd0*%}`_nI9`}yM&Uw$Nfz#lFSPpCz{
z^&io5e)_$)AND`E&-JELw%&5l?Z0sy^vh?Q)pU6Mn|Ez=9J+Dib%TApvv2(Mso@Fy
zOOL+y<|ns)_a8sMWy2wZ7uiu?{+jUGTR%P}et7K7&mS0Ezjfm$&)s~)cU#-eJM6yT
zHy(cVD}Q<CrPGf%iX3v<=MTN*l{YW9Z^#|->o*?{ZTa@!x|iQ}`{|#1K;PYb?#%S}
zv+RczhpHC~_kFAJ_Rn5v@}-YmJUn*tg!InAj*XhXd}K$gpX>jm`mxM0U-m!n)-78%
zf7Q=_>bIrVcmD93TW<DmvYv6spZ|Wz;kOTt^uOtE+&I{fyyLrrpPlgR7Zx8pZT5|q
zW1IKA^pn57qCYq?eZ}byJ-+{K=8xQY@$jj@zo$b_y!zAEHGjHm{)G!SJnlc_87Y3@
z=C}TF-)-UhEbYdLPv0EfeeV4qy7Bkxe{lFg>m8Tfvcvw-+m62NTQ@&^ZS~(SJ>xIm
z-*QjWa_(_=R*sN=pdQzw(br#%9dpE!ua36#ZawJUN$8J(TfcekcOE?|ap}w1Kc1I4
z>ax?f-SCgI%lhY^+VA1Jt~>f7G4o%~8J|)PdF+Pe>92mzyE5N;_vYxf`Y%uX#&-(W
z)lb~KQTO$Kzd3lq-KDQy`L#`FY&ia|pKSD{e|yF`pS}L6>TA0nxu5-X`K;lV%<o>k
zZ}*dzN$O{2hd=YiHJfg^_rW*rJL$pmj(*GXw_h56b)xOk<36@vKgyuL@rE6RXRrER
z@4y!?-Foi-dE%4XAJ6Mrcd&aou5;aRpWkU--}YAas^u@vcL?UVbmQB$^FDg{H=jJ~
z^`~AerhYiP{TI8VOUX~~%pMVJU+}~Z`yHQpYv-{STyoo6AHVU1pPag|&+Fe@_|`c~
zkG$_Z`nCIi_}D2={&-9E-Q%=Jf9;=V9rx**uYB^e=bdSK^6Hl#2zF1n{IuB<yIu^R
zwaM`Nd#-x>_#e>+c5VFgML&w(^4MR?o`e5a_v>dK`{?a23g167dH=>cg=^HG`^ENW
z@4oGn`+u<{oO{#fRFU@LpUmwe)z_vc^y%%7?byD)J+q}QTFCC)`SpWuKl}85E<gYM
zr!Ici`p?PpAMfsd|D<PsUVD1J;hoQS{;=-sFEu{%{K_r;$i=ONPrV<!VI$3D)>RJQ
zaCmy}#fsVcYw?m_eXsQW^R|6>U1CLVIcUqL_Fk-N&wGD;<F4O*Y3D|oTG5*~Z2r*R
zi&Y(X?iI1Yf8^PXG`pe?-f(bn@5QP<qdoeJE9n2`Mk=l7Q#PD((%y?zebsfh9(%xR
zN`2G1^os7R@8tGgtm?Z$H=m^~v0Wctmt4_X588U<-iuYe?yOTzJ^P-_x^){C9=ba4
z4clSDrpm%`3yIY~)27OzxENpk@;6n~q8dmeYJjw&7Kj&lAf2cK(u;Z^gJ=LUibf!l
zXaX{eW+01b0TM(3$SPWaY@!XwF4}>dq7%p^x`5oG8^|MifV`p?$S3-M{GuNyAO?Vf
zVh|`KhJeCi7$_n}fJ9LQii%O7m>2_!i*cZ&m;_3RDWJ5N2Fi#Tpsbh$%85B3NtA%{
zVjid<7J!Ok5vU}VfXZSSs3KN?s$vzWCf0ytQ3h%p*GR11L^f4g$F+dk$F+gBTejO)
zZ!P=2d+STxxRCkW%GLccvhX#sc7<QKo-Eu7yr_`HE6LhTV6jIQZva*+M2%kEBWm;-
zg=o-gdPIX>s}L=EZI5Wt^9te7^F6|&*C|AYUe_Z!^m>Kp(d&CekKUjV1A0S`7|<IP
zVnlE35hHq&LQLpQJz_#{R)`tBxkt?CEef%qxAcewy`T^Qz0e~9daFXL=&e0sMQ>Ax
z4ZW>LZ0PL@v7@*5h#kFCAx`wp9&w^~Da3`|)gvzSZiTqfyL-fq-lGr?dQXpd(0di)
zMepqqFM6LseCT~W;zRFOh#$SbNBrmm3JIVO^hf}GP$5C|!5#^s4=E&sKGY*2^kIdB
z(T95^j6R}}2>M8mM9_;05z&i1BBGBfB#J)TBT@7*g~ZUudL)KEu8=tTc#p)<Cl!)J
zpX`w&`jkRa=u<tCLZ4Ph8hyG)(&#e^$)L~lNCtgYAzAd<9?7E5DI|wJ*CRRfl0qc(
zQjbXJ^9sqM&-X|keL*1w^o1TNpf4(<h`!h(Mf4?wl+c%Yq=de#kTUvmkCf3@6jDK7
z>5&ThszR#ht36UhUsFg8eXU1o=w*e-=;a=f(Kjw5jjxflLrmj((zq44bs1@)Z(UDX
z=-Zc(Hv0DUq>cV!V!4=jRu4>@Dqkg*udeZyotGbS?)3E^j>yW9DzhNLajLXX1?rN8
zOYo?bSWLq)OIj=fbrvsNBXvRzOSPnirMjbrr6!?)rAE@gQq$4EQk&4iQY&d;sqJWC
z$tQSN@)8eAzQe;(m(amdC+T3R>*!#qPv~K(m-MjIcl59{Bn+@LNCsFMItEx86Gm7X
zB_k}29V0AF2@@<$k_nckjtQ3Lgc+7*$qY+##|%qL!U9W+WPzonV}YfR5MU`t0xX4&
z084Aa3QMbGg{8G)g{3WFgQZQf!P3^T!P1_v!_qF<VQKH!Vd+dbVd<2duyl5uuyiF{
zuyjc-Sh_kcSh^E#Sh^)QEZrS9EIkPiEIpD3mY$9WmfnOHmR`vVOK-;uOJBkVOP}O}
zrLW_Ir9a_^rC;*H(%<pJGLQ(sG9U$D8R!IH8B7FW8I*#s40eLB3?)La3`rqahB_fw
zh7(~}hNUnp!<{fJBZ&wsBT@vGkxm4bVnT$aD2cEXJ0dKji6|_iQWTcaP8620L=2WO
zDF(|}CkD%SA`Z*A6o+NJ6NhCok%VPZO2RVPNy0LfNWn5CrC^!rq+pp&q+yws(y&Z-
z(y+`VGO)}@8CYgI8CYf$Sy*PJEG)B~EG%=094vEE4wku24wh0vf~6!$u#`FyEc1yx
zEb~$xmibN|mW4zCmIbK*%R;9B%VMGk%c4|-WwBF)WhqgDWl1W*veYTTvYaTxvMiNh
zS?-i!SxHo2S&=HRtaK`{tR||ktV&f_Ry$Q#))F;X)}$INYn>V_<%A4NS(0HXcVt*L
zG7SmNNlJ|>P?t0=!84Rh3%0FN3%0E;Y2mh=X~VW%YQwhOC2icUeB-wAz17;8%U6iy
zE6819>8|mi@3}Q!KI=~hZP`>wF0<rrwss1VT4pKa^fF5$XO>w8IlIiV$hl>fLzb4A
zgq&YydE~+}D<Bt_SrNIk%u2}RWmZP6EVBx7b(vL>Ys;*LEH5(|x%~obzsOeq!H(ST
zbk~7P-+yh-LvQYd$l5vGqGK_z_Lrn~sDU&N4UpEM1>zk%kj|k4(mV7(28RL2=r97A
z93~*M!wh6`SbzkF0AzJofou*NklkSiaypzqE{6-q?QjEm93CLA!wckd_<;NlKTyCC
z017&SKp{s6DC`IWMH~?z(IEmw9Z{f|BL)<A#DS8IBv8tc0!lm5Kp95{DC@`q<s3O6
z$sqyd9eJRFqX1NN6oE>P5>VMu2C6tJKvhQ-sOG2v$qpH);ci4$uf!`4y}4UJZFd{!
zLd%7&)%%InYc$<=?_=n_AG~+#w_iE-uuYY#$N^Uod3Cn>n7VP_y^p1@%@!wQ@d14N
zZ9>$@nhDV$YbQjD%ufi9teX%WvVKDJ$c70qAR8ydh-{h=6S8?i%*d7ru^<Z*A|P8Q
z#ENX25F4_6LhQ)S32`F3Cd7s8o)9;(XF@#4-U;y{`zFMP?4J-na$rIN$iWE-B8Mg<
zgdCobFmhx<BFN%|h{(|ii6X})B!(QHkT`O3LXya-2}vQRCnSxWnUD-}c0#hqxe3W3
zOA{g?=O-kOT$qpoa&bb6$fXG>A(tnlj9i(J3UYNqs>rnosUgb~BCql4b4hQk+<f*t
zleBibSh$)lTtnB69t$J7a4Ygox^NHjyL90_<SV&_t2p@l8@Yv>kS}KoSFkm|*rkgX
z)74w^$`eHcx_CYEm@eLdJfn+mA}ieDS2+0mKDT%+@|;_IA6VU`YV4^8RE<6Ln5wa-
zo>4XS)C#A@p1RMev8SGMYV2vcRD(UufNHR(8B-1RG&8Efo<`v`*wgem4fZs1PJ=yd
zmuj)69Z)Uyv}3Bpo_0pH*wZSU7JJ%0r^TLj&S|m7cPWoOen5Hb@ngzkkDpN<d%VK&
z*yH;gk3D|Q@z~RKsSbO(0o7qoH>Nu5>1I@iJ)Oeou&3*DI_&A@oDO^XF4bdCKcIT-
z>Bm%$J^hU8v8Pu!J@)i{PLDnPoYP~^(4_|K83xpVJ;Rt9uxFT21NICGXTYAJ&l#|1
zm~#f~8N1YoJ>!5Hv1c4pBle6lYQ&yV;f&Zb_BkW=jC0P2JyVyOuxA=j6ZTAFYQmmr
zMori=DVzy=raouFo@vgRuxIX4Gxp2_YQ~;<OwHIc&!`!DW`#3j&)nzC*fY;LGxjW9
zYQdgmKrPs_jHv~CmKn8R&!TV^>{<Gp1$&k`XThG(r2_VZ0Tr+(jH!S<VMYb)2?{4*
zPv~<3_JlbnV9(m6R_s{^)QUapm|C%Colz_HtO{qvp0&?ev1grgR_xij)P_CVfZDKU
z8&ezhY%^-Zo=xFw*t7LH8}@8-&W1gEm)fysA5c5?>|<)jo_$8`*t08~9eegZXUCp>
z&e^f&>{2K8oCE5_o^wo{*mKUP6MIgDb7Ifg=bYGc&N(OcTwUtIo@+o|*mI4k3wy2^
zbz#q?a4zh*`kV`Ut~uwzp1Vuk*mDo48+-0Cbz{#xqi*cE70!)4cb{`(&pqed*z<I$
z2Ya3Y^<d94rXK8hX4Hc{kHUGd=jn4E?0M##2YcQw^<vLEpkC~G$JC2G?~Hn}=T$f_
z_Pl-0i#_k0^J34}r9SNW2GoZ=-<bNa=bKR<_IwKG!=A6t`LO4kb3W|(yVQ?8|A6|j
z=O0r)_WU#I$DUu|{MhsNIY0LNbIy;wK$ixv7Z}h0_5x!Xz+PZR1K0~FTmXB4J{Q1V
zV9o`w7wpm?_JRW%#9nYrgV+nsXb^irg$rUY*yn=S3(mP9_Cj46!d_@VL)Z(AX$X6v
z84Y1Cq;MhZh5B3wd!acO!d|#b!`KTCXc&9pF%4rcJfmUkg%vK0y>OokV=p}C!q|&+
zX#{(b0gYfUGNuviMP@XDy@<j^uovlb5$r|gTm*Y!mx|aE2UNtKIHn@@#2FQ_Cn}tX
zJ+aS;*c0cRh`nf+MzI$i&?xqzV;aR?bVj4tiz-|cd(l1@#a?vIMX?v_(irw)0~*6#
zY)oU=i_K^ZdohKJVK3I_V%Uq#xfu51T^h$;d_d#ai;rm>d+`~KV=u07aqPwWTpWAx
zITy!XvP+ZLOAcrfd&x0PVlO$PN$e#RE{VNlpG#sdIp>ntOLb`qd#M3UVJ|hNDeR?Y
zG=;sC!lkg6>T@aVrRH1;d+9DsV=q0RY3!xPG>yIVjHa=dR=70w(tR$Cz4V+*V=vRC
z8SG^SG=sg&m}an-nb8dPG76W$UZ&4wu$P&08SG`dG>g6LfM&6m9n&oKvNM{+URL3<
z*vt00EcUW<E{nZfm*%jS8_*o~a$}mqUT#Kn*vlzg4tu#im&0Ce&gHNtb*Y3sX+S0H
zNn<KuPnuB)dy>LQ*pvF4ggt4_N!ZJGX&!s|0nKACKc;!?<!3aHy}ZKZv6t_2dF<uq
zTpoLcE-hfMFrWqO6~?rHy~2zZuvbvH0`>}hu7JJ5oGW0j*ri466$iA4z2cY_u~(eY
zBKC?3SHxbi&lRy(oO4C&mAbTqz0!b|uvZ$>681_nTEbpQ;Y!#m^|=!EN^`D+y>ge9
zu~#0@GWN=2TE<>^M$6bMD_j|S<vv%&UU|-yu~+HR3ic`kTESjrOe@%{%xDFB6@{x{
zuhQo#*sILB3ihg9TE$*<K&#lRj%gKp)fuf~uc~lW>{a_*6?@e=SH)hfOKaGx4QLH}
zwK1(>uQsDK?9~*mhP_&!t6{G;=W5uKyHv)WJfJf6<S~`8C(o#iJz3#o?8$vj#-2Rq
zWb8Grrj2Xp+VdZc5pCRxypuNWL4KDu-b22UYh1;_=ikURZbH7CHLm#JNy3%qSFWZl
z?6pR;g}v5J+QMGzUE0E4>q@SLz1EFf3wy20SqppZt7#j1?GbHbuf3DDvDbc=wz1c~
zl51nHeIwV#Ui)&^UhDZT{qwqa-ah9Aj#aK8sw+r&^?4xn|2!|mD%TLzHRR+~A4tD@
z)b&69+xv7=<#=}d;_4e4HdW4MXMaGek6N!hXujwD6>ATg|IgNOH@?QZ|GwjH%gQsc
zs)cQ<pSAMXwrX+P+9TC!l^RH+(g0~yS|DD<1L;&cAiYWtWKbD^j4C6LNo4{utIR+a
zl?6yp2|!kr709Nt0ohe{Ag9U+<WjkS+$uMaN96(Xs=Poxl@G|T@&g4_0id8N2ozF<
zfWoRUP(&305>+BlR22n^sbWBJRU9a(N&=-+DWJ3}4U|!3fU>GAP)?Ntl2j5<UX=$b
zs0u(uRS~G9Dgl*MWuS_x0#sF1foiH6kgSq{8mh*&wR@75ss+?mwLhTj30sf;!?n)!
zD=#KJf1mR=uil1E{LN*jZTOGFUfOfLf08XchtKjXvBjt1r5frbrbgB*F%7bIiD{Af
zCB`G`mY5D%zr^&&h9zb|HZCzEvT2E#kj+cXjBHtA7Gz<G3CPwZW<|CwF&nadiP@2z
zyO?tszJAv(=0bMwVs2#5F6Ke@?qXhK-!A4u_U~eT<iIW#Ko0I=LFCXb7D5j1VqxUS
zE*3!+cQFw;x{F1TW4l-kIlha<k&{a-iGLrdC6+=?FR?UoW{G8xvr8<CoLgc!WNC>>
z$oVCfM=mU}0&;PQ6_HCztb|-%VrAsY600Crmsl0Kw!~`4@)DDg8&9&vb9fYflC_ZA
zPqOwJf5sGj<bacw9y(&>hV(F7Soyd7zwhl^IfP8v;sbC+Qcsy0Su<rCWbKq`k@+d(
zk#$q1L)K539@#Ku24v%u8Ies>W<oYknHkwKWfo*%$^>NVlv$B&Q)WZ9PnjJ#Ib})Y
z)Rd)=(^Hm4&P-VbIXh)p<lL0ykfkY;kn>ZPM=nfR0l7G3MdZ?ym5|F*Rz|K&Sp~T|
zWmV+bl+}>sDU*>K53|NDxFWS5X6-dzz4Kl<LD};T&$Sbj|9P_0V(b3pTZZ?3w&#v|
z!WJIH>(_`aejBe}Bc?{yjF<*lJ7QX7e#CfW-H7Rs^&_T7HjJ17**IcGWYdV5kj*1z
zMz)NY1z8v|0ogiYR%F|V*^uocW=D38m=oDGVlHI&h`EtHBj!Q&j+htOH)1|y|A_gK
z10xnd4vtt5IW%G+<nV}vks~7(K^8|$M2?PF6gf6xG35A&#gUUEmPAgCSPD5kVrk^e
zh-HwoBbG(ZjaUv@8ZikuKVo_0!iW`+iz8M<E{#|TxjbTJ<jROukgFqBMXrrl4Ot#B
z8M!fGjR)cO)tayta(lwsYut_p|F@=KO|E>`uqCu5zxracO_i;ot@+jWOm7$8tzBo(
z=k}Z}9d^EX^=#=Y7d&+F<1fGR@XBv*DQ-D&^>bG>*>~?7uvc%wy;pBa@Y{>0>^UU7
zz!qM_o9r@M#FvGtmzf$_v&=Ne+GVCi=9d|dtXpO}Wc@PJBO8{P0ol0BjL4>CW<oYE
zGc&SfnOTs9WhNk7mzfpWw#;nE_GM;AcJ5}*U%*|KYd3QtyLU4;vS&B*AbWQ+FS2hp
z^CA0pGe2@*Hwz#Kce5aJXg3QXhj+6ua%4A)Ad9=1h#cL`qR6q`EQTE4&Em+77g%HM
zWxy-n;lIFIYkbGWpLput|GN3nJueGxFKmZ1iRR)L;hWK>8c5UB0BM_AAil{1>6$tq
zeNzu)Xc~ZwO(T%0X#z4g%|Mo>1xRQLK-Q)e$kwz0*_(DCXVVGfYPx{jO*fFI=>hUK
zy+FRE56Iv20|lA^pkOly6l#Wm!p$&Hq!|Gcn<7xO83l?pV?gm{94OgL0;QTMpmZ}0
zlxb#wvdt_|u9*Xpni5dHnFlH~3qZwY5vbHG0hOC&ph~j>RBcv)YRwvu+?0VD?Z)=i
z^O=<^b-M-BZnuFhB9@ED2QLg>z1v;C@7^C#v3du(a`nHLEbM~!Ha6}hja~TuhI>g1
zxqUBbukqb~>VB#C%1u{un<^jO=cD_r-Hr|~9WJlkj!xa@)cscf)Zaeuk@d;OZQ+%B
z_|5xmK4|qdm(#PqaNn8xo)TO+XdJ_i*|zpjQ<F76pdV!2f0P%VePShd8DW=^!&i?B
zn^xY%%C1E25!PGdmtR%9*KWDzS1Zx2#jPi={%!o$u>S22Z9e$QJqMu!&tF*mzrB9a
zVc+8|Z<zP|00Vx(u=Z6g2n+T#TI3fEYhUA{uxMW+HLnJz7S!N0yat>`(16qOT5wuH
z3y$Y`aJ;~S)A2fRIzb0c&+EbI1wA+eZvbZy4B(8s5u8ylf-~_Za3;Y7&di&^nFTXA
z3vU5u5iH;YUH~Tu0yryg1!onk;B34NoK3KSv-5UvcEJwL$veS01t&Nc?*iu%T;SZi
z8=PBkgY)nna2~+}&dYnjc?B;xAMXR_6MW$OydRuj@PiBR0dN5!04~S}!3BjNxDX!#
z7ZO6?!h9H9SO|lQ@DXqkAp%b1MR1}Zf{XG|a8V%&F2={e#e^8RI3EWW7vkWOd=gwz
zNP<i8DR3zv1uo5}!KH;XxD1~Gmk~1HvV0a?R>*?O@i}lgAqP(4C2*1;fy?uGaCspQ
zuD}<-6@&t~B3}eo6pG+Vd<k4hD1j^UWpHJo46edgz*U3_xGG--R~4$@YJ3e`O{jsB
zc^RB6$lw}$!?1c(*i>l<4f`6k_!hX9&;r-y+u+(l8{GMp^WjZ%7g#R%fZkTWd-b2k
z5AE1<%6A7@xCf4Ti$k(_2fpiNNYu!hA<-afheV6a4+)Q~8xkF|en|Alh9NN^8;8V*
zY#I_1vUy0%$d)0oAPYkxAX|sTifkJa8?t>!?8wd`aU#2h#D(l05;wAENIb~iA@L&n
zhQx>L9}+)uU`PVU!66AEhlV7C93GM|a%4y%$l{QQ$k8E*BFBa#h8!P~IC64GlE|qc
zNg=0)B#oRIk_>WoNV3SeA;}?2Ln0yPha`_&7?J{VaY%~Dr6DOHmxrW`Tp5xIa&<_m
z$h9G<A<IJ|BRB3KjeGE%?+(&JZr?%L$d?eyC2K!L<DdIW_x}3KAN}t1l^fQ{3#YD~
zA1yQ%zOY7%^5SRK&W{%Bi)XBnT2_No*VW)OvId-{t^uc&wcxaMEjV80!SQt-oKDt(
z)75q0^s*kDzODynkPYArbptq~Yy@Yl8^M`m6F5`d1kNm*!I|r3a2DAD&QiC46J!CL
zP#3^iWh*#q-3rbo+rZiCHgI;?4$fY;gLBGGaL&3DoJ)3rbJbno+_D>-yY2?(kv-r%
zbq_eN>;>nod%^i+A2?s#2hK11!TIZcZ~-|0E>I7E3(7%o!Fmu}NDhGu)kEOIau{5=
z9tIbYBj6(S2slv|!HIPdTvU#Ni`JvyVsZ>ztR4dwm*e2#^*FesoCKGwC&8uU6u4A9
z1uiY8!KLeIa2YuRE>q8d%gR}B*?JaSPR@bL)pOt^Spp~3C2)B;4=!KNgDc1daD{pS
zTv0B9E7ptPN^%KYsa^tCmdoJE^)k4MTme_9SHM-}D!6LB3a%#Cz}4zCaI!3elj}0L
z#>tIS*Upa`jm8(&sC9A+T&vLn*FL!ouH9&V;O3A1>Ni>I8`8=T4RB7G^Fe&HaLQcB
z?kRI4d#22T?42?%vTw?K$o?twBL}7|fE=8%AaZERLdfAM3nNFSEP^afnTQ;nvM6$F
z%3{dzDT^bw9%ii{eDDDGS*M(F(w+w(&OYnpvm<*x$nUw>`-`hI5Bcks69n$4^`&yY
zP+Yh2ql;I66;?l3x30YQVQKBdlN+Bs<Mb7bpL#sU)^GTbYM+gpKD__t{i0h!flqJQ
zH|pPd@BttB=*R5G9P4zsJzm{`8pjdVZ43Xu-#zHikALE@PaS^bQAaPTwR(fmEZ9D2
UIfOR_4?HLmtJ3fZK9L;v-}8RP82|tP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.9B8C2015B4635A97.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.9B8C2015B4635A97.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e66f128f60d3035ffa026068300ed8862092a761
GIT binary patch
literal 22866
zcmZvk37A`Dq4)c~=WPwNQd$HQt_~Ksc9NXzCkY|@dXga7Pxc&?HA(`Z5YS*b8kVvI
zg9hP-C5Yhzf(8T?5E+YlFIN=C>P0~jK~Pb_APPv)@J;(X!~E~}e0?6m{QCb*e_gVi
zIY}lR4TodJb?XABL@#KZdhW;Ax^?Teto*m*+;i$raqHH_H?3QD!Hthg#j}pI9sKM&
zhg|jZN!7OHx%bGii+}K~KDYCj?dP<wJNUY1Uh98#zW1Ov@6j&qyyGi}t$+E@(@%Tr
zaog<!a>}}IZU4sawmxvklcA#XGxp~{@J!{R)9!!sy1d+c`pI)IT)O>&?;f4_Lq~tY
zA8!BdD~nGozwy3@4!-a0S+sT0j?U{Z9R0za7d~|A-|pU=X&*PLUKM@hsoS5v@$v6g
zc8=T^K78_Z=a1j|kLQ2(`gx~c|GR6yX1v>Z-7|OPo_$Ji{MUPbapc_>>dyRiVfpYQ
zp5A(s?<W1NKQ-mUFKz$pd*A%c&dgnxywJQabb<Zx=dStWM-!iZD169E7l$X*BH!{x
z^xP-D{r3m`_aAVr>C~+^f9@;4bshHeC!N)Fc>Sw)Y;+vCapM;T2Y6>+{_A7I6ZvN!
ze*4uAZT<QiKfQUw5rb#gv7i2;@ZxL#d1`!H?A1>m8eG41<A=`MeDv2_+b=lkzTlT1
zeBrZydF$CTjy{$gar!5Zy!rW8FSl>V9sTQ99|>*w>fgGT-};p^KJiU`ck_8O)9dHh
zk17sT&lK*vv+?(zJm2I?AGmmU{Nz#Tt-~D~HGg^Eb+LY~|KHUQWRCl^|C_Jfymj*>
ze*VM1E49A%(r<6R$-l{Z<`FNye(AQa436=?>TldI*pR&K>w}-1_=~d_?>~L^<>z9X
z54`L{zrLbBI3|6?84o;i@U7<e-G1@#vB3YNLyx}j#Gf>OzGHs&&JB<Fk9bmw@7(;_
z8~5EBzR%Kbob>oj(LLvV<2^V0Vg2{E9k$-F>*nk1@4xjUx88ZvgV$95>#{Td@|`XB
zG%e?yaC_xw`FrYdJsSPf3$f#l{^1LwExlU~yLS@$W8jwmbKch;J~naLbJ;&$kU4hO
z8QZUaV@Fy4^kWA-c*hq$@;Nc{kEe`}DMx(w`sL}b{=mC3ckaC@y1oAMlfL}5!WZf%
zZQiK+(*N8PJn^p5C0BlN)0rDSddGin^re4y=DDX_`&jkGJr8|@J+XYwa7*U*FWk51
zhr1;8le2A4zVi7^H{W~zEBAfu{tG_xn&odlH~#7*+hr$wV8MQ@L4U*b*A;$o)wg?x
zK6BaD^Zxnhhc0>~uWQ}L?&Y}77lsG?Uh}8zuVt@VK7an_W6MWK_g~b#?3y!%+S{i;
zzUffmpKiwoZ+da#sjtNZ|0%yX|H8{NM{IlN+dn&fyL$QSclOSF<C<HVd#{Q8;i0#e
zclb{{{-5P9RsL<yZOcD6$Mm%R(Nli<;Poe5`sTlX@nuylZh2{N!5cd_{kU<=qw97Y
z^ds>%L2Yv5cXsQ<8()0u$nU<s<Lk~d-ul+hj(@EB{kJ!S9y+`9w>xgN{_FgQ)ur(9
zuaq;7-g)WSr(Ach`{c;2TNiIr|Ky)<cHYXL^ybb(j{S%4^}V0?l+kh7&)>Q6@w-mA
z<E)?EJ$&5w{(mjB`B$EN`K-52>AbY#Q;m;5y>dHW-?KgR<d^RL>qaVN)>Tg3aO%hQ
zU#u+qe;o7b_g`Ip;z#9;G`BKu+psOY|6+Cay#3c#?)d$u?%qgKEA!?Jo8PnlVs-XE
zv+bGUQfl_zb;*@^>tS1u*?+M*pLNo@$FJXez;8Cv?8<!jhQo{dFIMNl`G39SB=*F?
z?_HNznYSFa<-_|gR_EE{uk9lA<C;zD(kpXkeJ8j7Vs&13&Z(#ExFfUnC!`1N|C9B;
z_x3Gceg62PHdQ`PRG%j&ul_O{79O}d@MYUkBeJQoz%Lk9e{(A&EZBigSV*kVBEM)@
z{X5!JSriuSK;mM2jnupvOtqi}Q!T2&)bJWGHG&3Aji>=r%WJ{Z3R*C=q83b^=fUI!
z9!y^3!PN0OFm-|sOr59$Q_t(c)C+nr^`ah318)G+AQ-?jhz2l?yb(;JU<A`B8o@O2
zCNNEc2~3k{0@KW!!88kIFwLSFObc%T(;`^Fw1^fk1zrGC5Ckv<Q2^7*TfwvnRxqui
z6-*m%1Jfqhz_f`rFzvh@OuJwQ(=OV<bn;Fxoq`igr|1OJ#k;_C2`(^Qq6<tn?*`K?
zxWRObZZJK(2TYIP0n;OT!1VH7Fuj5oOt0t#)5rV3^a(yNeWDLcKko<AFZjXqi+(Tz
zd;rXV5CAhE2EYvRK`?_t5X_($1T(~kzzhi?FhgPp%rGAYGc1I`42xkfBYXtRh!6oY
zB1XUzc@a!e5Wy5h5zHtb1v4r{!HkMgFk^fS%$N`ZGbYBsjPr3Y<3b$FxEKdB$tS@~
z3P~`NViL?0p8_)_q`*vxDKOJ~8qBni1~V<D!OZX(Ff&31%#4@;Gs|be%nDgBvtky^
z9G?R-C*;7)i8(MOUIJ4RBrqjW0yEF&!ORPJF!N#_%mQBkvmg|}EQkd#i+mBxqEH00
zC>Fsi@g*=zLJ7>0SOT-mm%%IxWiZQP8O#b_0ka}hz^sTBFspnO%&JfYvnp1>tnoE4
zYeEgonpgu<=4CKtK?YM6WiT6j!?1dYw5ie%8g`%)8i_S(@hvc0LJQ2+2`w<&d>hQR
z&<3-8LL1BrEf?BWk54vLK4bX|&_$Mu-lYTg-tufWE@VEja<H<CEPRoy9iS{+OBQYc
zUR229m1OO(WU)sUuLo8uM2)+;N7T4$6r#ah(<2((wF=SVuI&*m?z}>H-1#2ian~tC
zhr6yvbhzskqQ_m|BYNBo3Nher=n(_%Muix0H};4TcauU)xSM*!gu7WGX57s^V#eK~
z5DV^>9<ksqC`7<r=n(;Tt3s@}TYJQcyG<cB+-*H#!`-eBJMQ)#vE%Mkh!b~bk2rC6
zDa3`lt4Cb8yA|Tb-Q6Q@+&v2M;O^-W5AI%tcyagkh!=OCLVURUdc=pjUm<?n{XOEx
zJ)n>P?tvZ&;2u;+5cgn@1aS{3B!qjYM?$!V6%xih+#_M!BMOP&9_f(??xI3O+{GRd
zagQn_ihHz2qPWKt62m>#BQf0L3W?($?~yp}NrfbFPxeR>_mo0XxTks~g?m~dY24F2
zlEyuwkPPmb9?9UIRY(^1Y>#Ad&nYB_d#*=vxJwF=aF=>S!ac8$Jns1($>Uy7NCEdk
zj}&k(Dx`>eu}6xymlRUMz0@Nm+{+3n<6iEOGVT?HRB*5KNCo$*LaMk|d!&kcO(8Yh
zYduoKT~>&UyWAr(?u}if@kO%su+_MhG;RTI?IJDQTi22n?(JQqjeGlA(#HK_V!4=j
z)*iGjA(l(lc+1_FA93FFPv77FMRsOEg2$=SLKUb>7B0m{t;Av)9<!vyB2Z^>=NhRK
zYB1H38ccOZ4W=fc0aGJsz|?d!U}_UuFtw5vOl?OCCZFKJ<Ru<VzQcp5OX$GVNjflf
z9UYkZgdR-2qz6;q(SvD77{D}01~3gB1DM8y5lo|G1k>0tf@w;az%)rFFijm3nC65T
zOtWML)7&wGX-QbXv`7{(EgcJ(LP7vjkOVM=jsT`LVFlAFS;4e+tYF#_HZX0H4NP0d
z2Btk>2h%Ru!L)bmU^){{FrAVUOlQXlrYqqB(<QmUbah-{x)W|N-I5zjcgGE;C*c9p
zBYD8|bUa{s6J9XAk{3*G#|x$};RDkr`M~sbd|>($elY!#A54G84`v_{05c#3zzlQ(
zU<MOGFoRMM%wQ)7W+)K?GbDw;40S?ah7(~h!%`T`a3>6ABoP5KB1OQAbRuAi2@y<D
z62TNZBAC%c6wIg;1vA=-f*DK1z>G;TFk_t<nDIm$%(xT>Gv0}VnM@?XOiD>Glbs}(
zsYD9Ql#~K9)k%SwPNcz1OKC9Eoiv!4L<Y=^lmRo-$$*(nWWmfzSunGmESR}O4$PdC
z12fmjfhi><FeOO>Q|d@y<`a1^^HLtnd?ycPAyELcAQiwYbP8Y=6GbqKQW4BzrwC>#
zQ3A6hmB1`@N??`~WiZQ98O(C03}z)!0ka}iz^rsCU{(`VFso7(%xb3!W-U<zvnJKR
ztaWN&$_W`vS(3q&J2IG!OhbZWl2W4z)Fq8e@dzc;0=HFaf!pem7P{?B8{BrO4Q{(j
z+UTx);kNR%)!LEEXNl#r$X#OTuJMs?yEUIa=g)_2*;Gj`v*aGOb_kMMW+~+KGD{<8
zmRSZlyUeo4xn-6^mX?`>oL^>n<iav5AQzWe5xKO?O33A9Rz|KYvkG!`nN^W%%dCbh
zFEbgr{Y%z<hOPdC9dpnbu0xl;^P;eFF2cPKSv#g%bSwte{*u%VHIT-k0n$3OK)izo
z(m8ZMdWRm!;4lCg9Y!FN!vth@n1L(~3y|OtfUFKHkj-HOvODZRPKOi7<!}MH9d00x
z!vo}Xc!7KlACTYS2MRa>KtV?kDC7tMg&kp_h$8|dIz*tTBMKCA#DL<CI8f4&1WGwl
zKxs!BDC5WgWgS_doFfM$IV7OGBM($?6o873B2dXu0xCPoKov&?sOqQ!)f_b-*&zcp
z+>OZUm3UL7<!%AB-EE+qmYuHE`-#<S^vbh~`<@+Ldv@_Ea>!LgUe#8gD1Bwc+xIl-
znzwK@UHCj*yO}JE=)x_?chiM?koVGsw~?>p7Ovvp^Kal5ZbZJEEnLCY_Qft;yqK>3
zwpX6B9MHvUk;ioLHsl#yd=**Y7C+0u=l8kAYmn#M;yb|VE>&YsJ)mmrsmD}}J@t&L
zv8PrzHTKkfPK`bFoKs^@)1?~hX$Dk-J<XVEu&0?(4fZq&r@@}4&uOrynR6QKX}eU5
zJ?(&Mv8NqVE%vlCs>PmG;k4M(_Bk!~v~y02J-$nM?C}H2V~-zG9((+Z^4Q}Qj>jI~
z=XmV#bB@QJu1j^;(+#K&d%7{zVNW-sI_&8bPKP~RpVMJaH|KQN(|4&Jd-?&@V^2S(
zdhF?ERF6Hq!s)T6?{j+W>F1mtdxkDGV9zk12J9Ke)POz1j2f_KP&fnj41LakJ;R(c
zV9(g4M(h~})QCOfm>RKXoKYk8j0$JOp0Upvv1gofM(mlo)Pz0LfSRyp8dDSYOfzc2
zo=M?M*faGx6ZTAV&V)U4mzuF>9#Av(%wuZCo_R*i*fT4f8GGhFXU3j+&Y7`i=~4^!
zECXu6o@Goe*t5*21$!2SvtZBC=PcN>%sC78gf11ZCk&{7Jz-1*><Kd}U{6pu0eeE9
z6R;=DIRSgtF12FMI-pkUS;y3hJ?o5Gv1e5{EB35?&Wb(joU>xj)}=P=*#^{xJ=>Vt
zuxFc58}@7pXTzSY&)KkNn{zho*}K$^J^O&#v1cDsJNE1|YR8^k;q2J6_c=TE>~qeJ
zJ!h9XvF996C-$6U>cpONMxEGmDx4F0&OYbFo^#GQvFGYi7xr8O>cXCDOkLP>&8Q1|
zE`@Vp&(-H#*mKP}7xvs;>c*aXK;77LkEt7b?iqDs&#iE7?791#8+-0K=f<9=OFh{0
z45$Zto-y@c&oiSQ?0FQ<gFR25^I*?2=RDZ+cBvP8-U0Pu&pW1G?0IL@i#@Nxd9mm1
zb6)Iu=bRUNzAp7)&o`hx?D@vjhdtkn`mpCyI3M<Wea?qH-<<Pd&)=nf?D+@Ok3IjG
z`myJqQ9t(l3g^e3zt8!x=bv+a>;<|sfW5$g2Cx?x(*X7YGaA5NK;Z(|3-q}F_5yP*
zfW2Us2C)|$&>;4LV;aO>a7KgJ3o2X?d%->z#9naD1+f?E(h&AS0~*3!XiP)c3(aT<
zdm)7jVK3C@Lf8w<xe)fkT^hz-ctFG03y*0Sd*K-kV=t_5VeEzbTo`-dITyxWq)Q{%
ziwtN4dyz4XU@tPG5$r`2E`q&CpNn8GGUp=L6T4Kzo;aW)_QWw2u_w-`h&@r^MC^%u
zPQ;!#=S1v9yEKZu=zvDC7ah|m_M$Ty#a>k5qS%Y}xhVFcb1sU#SeM4I7aPzR_F`ii
z!(MDgW7vx+Tnu}$J{QAYY|h277w^(I_TmE?$6kC)<JgPOXdHWSg^Ob^-sj@ji_f_@
z_L5zi#9nehlh{j+X%c(M8BJm@sc=c`CHq_wd&xPM#9pdPQ`k!lXbO9&F->7FHKQr)
zr4%lOy;Pq|VJ|i3QrJs(X&QU!0Zn5sJ*H{wrDrsay|lumv6t?1Y3!xvTpD|sF3n&s
zGoTslWyUmvz08beu$NJ|4E8d8E`z<yoXcP@+of6TWd}5iz3iA~v6r3EEcUVrm&IPT
z&t<WfopV|2<+?P7z1)E2u$LRt9QJZEn!{dB;d0o^^|>7Oa&s<+J*i72>`4PEVNV)U
z3479vO4ySWPQsql=OpY&b56ouzDx7i%MWNCd-*ZVV=q6WdF<sCE|0x@pUYz}Kj-q;
zD|BfAdxZflV6QNy1?&}Ow1B;W!WFPr=yL_^73N$4d&Mp-Vy`%$MeG&Fw1~apj25w1
zRJbDcihZt#z2clJVz1PtCG3?3w1mCVn3k|tn$Z&WN(xuPUa8NOuveOMCG3^Ew2Zy-
zfR?dW9@8@R$}?KVURmMF*emzBGWN=Iu8h4(msYS>8PE#$Dq~u~US&or*sCa91$&h~
zSHWIo&Q-8i?b0gtssmcZUUf{X*sIQH6?;{Mt75O(=c?GN&bcb~YF%2xUTr{Y*sG0c
z4STg2tzoaGa5e1J`dkfrwK-SAp4_D}_T&MTu_uqIj6HcqW$eicCu2|Ub29eiIVWSU
zaW!pxp01tfZH#E+7Ua8W;~wO_wDC6bm0aU04nF?|u5lys<*aeVyU&KKoJhW!wy@V4
z(H8bvcheU3T6<{=d#x+E7WP^<a4qb$E@v(5wXdda?6pU<jlK5Ww2i&?UfRZ9`%12j
zz4i@U8++}`S$nPLyX@t4Z~gt;UvjK+1yNl=%Bv^5S>=D90B8Tt+ih0g8nphv{cj0^
z^WpcBh28MdpT@nUu^XTMyO*?(+xL?88VAyQkG=Lk|Mm{uRQV|T=;G>|fi_ikupRHx
z2{*jRyI;TVF3ZYORH}vTaBf_+xP9%+wpyhI(x@~*T9p=vSMfkPl@3U+(gPV(1|Xx#
z2xL;3fXpg0kVRzy5>x_^Rb>UTscb-Yl^w{bass(jE+Dtc4dhXIfV?U%kWb|U@~iwn
z0aXAfs0spwR3V_SDhw1+MSw(=2ozOCfnur{P+S!UN~)4TDOCz6tx5xBR2iVGDhrfT
z<$xrW1e90hfeNYuP*GI`Dyd39WmOrdqN)H@RaKyxss<#hWT1wsv3>2Hr=@BEwN>qR
zDSP79kGyn^bN$L&*)BZb!p*C<$&-G&>+}u3IqKPc*ZUu`g{SaYp(VEXI6Q}~USeuw
z%@WffYnPZ7nO|Z&vTlj#ko8MUk8D_C24v$BGa{Rom<idu#LUQ+C1ybumY9HSU1C;b
z+Y+-O+n1Of*}0oJm*LOv+Ra?Z?%m9d?Agsc$ll$|i|pIYe8~RY%#R${%>u~5-7JV4
z+RZ}9;oU5Z9NEnx$l`7$B1d<#C~|B!iy_B%vp8~ci6!yxBeld*$mu1PM$Rm;403jf
zWs!4BEQc&DF$p=p#PZ06C00N#F0mqVX^EAP%S)_`Tv=ij<mwWuBG;B!4Ow1dGIHaG
ztnm~+YX69}ehJ+EA#1PgA2&rGI^<(Z4;;O6Lwb-c?8Y0?lr4S}u1M-BQzL7pOoObQ
zGA%McWjwNO%5=#3Dbphxrp$nBoH8S_Y06B<<|#8HTc*r{EKHezY@ISIvTe$2$o46-
zBPXXUiJY3U6mojX(#V-9%OGc`EQ_3*vK+EBWfF3J%JRsCDJvisr>uxvnz9mddCJPj
zl_{$rSEsCsT${2QvOHxna^peP*o{}52U&ZKR}VE-UIno4?apgQJ^%aQsm0d)$DM|^
zf3ok6dcqd&$LrUKEq)cRUn8bQ){K}2Svz7{WPZeWWZj7Ako6;`M>dR@0ogcWMr6~7
znUKvRW=6J*m<3rFF#*{+Vpe3^h}n?sBW6c-j+hhKHDWGg_lUWXJtO8p_Kuhr**9W7
zWdDfykpm+ZKn{*r5IHnrA>{Cgg^?p87C{zAOhk^3SQI%nVlm|Sh{chUBbG!?jaUjf
zJz{C(%!p->vm=&8&W%_OSsF13IX_}~<idy*kc%T$L@te33AsFCW#r0;RgkMARz<Fj
zSPfYoF&VirVU7FY_SKrO7IJ&S+H2g72Y;+7Sd%MX4r~c+$*+D#flZaIp{@DVFF`I6
z_tw5((C79YEgf~CdG%=Nv!8k3;zyo){=t=>-csCh((32-Ubr#AUsODG-+lit*}^k;
z+gxUg_`*2#GE*aKmYD`wyUeu6{4(Q_b<0eLtY2n&WWzEuARCvN5!tlNOvvVCW=6Iw
zGYhh?%mifXGP5GvmYEINzRc{%&OOZeGq`VZ?O`rt_a5d(_UvIEWbYp4MfUAsK4kwM
z=0^_fVFBde9u`Cn?O`G0@E#UMj_hF(WN{A@k)wN96gjqs#gOBBSRA?WOV(I>A>_)J
z{?D-1v+rJ4SFb?E1NVO$i?#dy#R*ybCcXu8Le$8b3DF>HCq#?PPY92!n-CqcenRxf
zh6yns8z;nwY?=@gvUx(x$d(DQAPW;BAX_KIifo$@8?t>u?8wduaU#1W#D(mh5I3@C
zLOjUc3GpKPCd7yApAbKCU_t`O!3hZ>hbAP19G;Lca%4gx$l`>E$k7RjBF82qh8&-e
zIC64AlE|qENg<~vB#oSzkPLEmLbAxY3CST#6Cxq!CnS$tn2-W;aYBm7r3on^mnWo*
zT$zvxa&<zg$h8TnA<GjYukmdcfAFz;|8dj9`(7{HUbtxW2x;YoiOt3H;bcrx4Wwym
zfV5355Z~m1bWI(QzNrT?Gz~z;rV+^0Gy$2LW*|$`0wgpAAZybKWNX@h>`gn6v*`qK
zHC;gNrW?r9^Z<FAULar72jp-1fdb6{P_P*U3N=GO;bs^p(u@F!O%W*Ci~_})F`#%e
z4wP&rfl|#BP`a50$}}@T*=80f*USM)O$jL9%mWpg1)yTH2vlm8fXdA>P^DP`sy3@Y
zwPp=SZpuK7cH<&A8Pje7wcBl=&k@V#$h)r?zU$B3PZgiP@oH{U<^2b||Dd%i{I=3I
zdF={++5x8>wED|`^@4}iCmXkhR}MusAGG<f)fW>_cbt9SSqGjPTsgWqjvcps?PO<@
zHQ%N0W!-<27k=^RN^TcnyU4cH<C{$@??Yu*BKHXEt?_d&DBf$f-1DoI=+@%alUI+H
zerH(!_xEf*{K|btONU;#u==a5{+}GU|GO})KDzx~z4z*u#}8b$@78@AS-1yI>@E(;
z;%)e_HzaCg&5&r2wL_vs=7)qw)(wdcSwAFtWW$gckc~rPL^chH3E4a(W@O8dSdfJw
z5s<AzVnwzMi4EC4Bz9!ykT{WDL*hbq4~ZMuGbA2l?~r(reM90y_7904IWQyv<lvA5
zkwZfgLJki}7&$T|5oB>lMC9m@M3G}d5<`v;NgO#jBuV7dkfe~)Ly|_$3`qt#J0w};
z+>qpur6G}!^FxwHE(}Qlxi}<6<kFCokjq0-My?D=1-UvTRpi=`)R5&Nk&zp>k;Xmv
z#Kvu;h1|Z4w2?0*mP^;(N&T;brF(yU)(?Jv#>x%r<b~7L?nM?F3umv<qP+O=wfmIC
z`r?^uq?XmNsq1RkG_nRZO<e<<R@TC%t!rW9Wga%Z&cmjYb+GB`I@t8G9yWbl51T<Y
zz-Fi$U^B`_*o<`}Y$n+Ro2hPs%`BT?GuO?qS!4@rmbwKtK^9;W>H=(5*$SJrZiUSz
z+hDWRZLry8J8bs49X6-zgw0uZ!se1)u(|3k*xa%kHh0|(n@9G*=BayN^U7Y>ymc>Z
zKG_GGukM4*FZ*Hh*Zr^s<N$1edH}Yd9E2@c55g9bL$HPFA=tul7`AXd3|mBwz!s@T
zU=w8#HnA?k7L}v0Me9-6VsZ?&SUm<?T#myQug76a%1PLg^(1U5IR#s)o`Nkcr(sLi
z)39aa3~ZTt2DYr6g)Ljp!j_YBu;uDG*d$qkO{z<<<>fqV`Fb9<f?R;DP%pq%l#8$x
z>qXc~atXFly#!lXF2h!?mtm{O71%2E3T#!m3R|^ag{>ynV5`+@u*tFvn_QP+Yn<FT
zZS^Q@<@-N18fULj>*N+}twsyB_Q`G7+Ku+R+x+2Q{WfcTMOt~|sB_Ak_v87JDRUva
zr_7D)nKBQucgnoTzA5t|`=`v09GJ2Ia&XFm$e}3<A%~|dj2xM=2(mb3B64)fqR6o+
ziy_CSERNiIkhQ+|?)i~(PCfHu`wn4uoOAMy$i5Hq`!4qX;3}`QPabSJ=+~cKUn=Jd
z#dRy+uz2-HVfBM`>&j~%mexKzxpB&wXRKiSw2yLZ{f76b4%oQqy$5eTD7qyS`1dUb
zM*Uk4KjeMy|A76t<DD+I$E!P3<2c&7ec}K2vxgn|Zy!AB!`qHI_9KgGt=?cX3$_ng
Uj^ItfLl29@sx*9}Pb4S&U-oAe3;+NC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.7475A9472170E32A.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.7475A9472170E32A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..43e827f7888af17d036b5cdd8e22ca639dc810f7
GIT binary patch
literal 19314
zcmZvk37Avmp}*(*&Id*rMX>H|V^J$vPWGeZWZy7kJ6Vr{L!%{VVog*KM5A&v)TJ03
zH56}9;}WVy8?4J!s*Y$EtM}UJ)azQbTGzTX?!DqQ-ht-{`Tg(n_j$}SFYj-78Rq0n
za!gDwS-kkn=-61Guw-3%<$0&Ev9YmvGykqWZ&m4KZfq<%J~p=Q#%JZFXCHaOfv<gg
z(A6(=m!B~2^?mY@4Zr*M26xFZ3s<e(eBkDtpEm9p>^k6I_v$s5+<E&UvpzidjMKk*
z_Jli{q%ijH3m<rU{=)}7A6n|Xc+ne&@0?hF`a}QPd{Vyh^7H3iG<nCm$B&lYtr}zR
z-tqWHnm<i_`JG1&y#MR|l5^Luu72_6(cin|qDNMKwry@^t-Zbc>LpLSbjQm#KKuB@
zC2jXbkFD5ze&?%yd*it;)|_#}+nau6-sarA^X}YhFA3rw_J8rnZ5J8NdiCU~LyvfQ
z{!PA{jJN&8a?;}W7Jjn-zy5Yf=I)=qS-C&7Zqc)^U;C4|bm1e54|;FI;x4u2Z~1b`
zd4KxN=iB`c?Yqgca{kSi-TrsiAul}dT%KM$>*G7;h)2$ub6s;^Z~xX$o@pic*B<-&
z<L}S^^_PFSdG-;_o$S~PuNU6>^!F>H3nCx?<lx|}`E$O%X716yK6l}|qwWvhy8X>d
z|M}Hx#iNfUN1XAKBX54=<Es|U&K>>g$4`Xj-Sb)Ps@rZap824$Hg`?m^2Mq}M=fos
zcAk9yUFFaJ{6>XecKC*tqx+Qn)uG}X-9NvxInu~A{<!?`jQv9YgP-0!f9_BH{14t<
zcJ8m<`}@r|`Nxkx>xd7(xO~Cw&13u@`^#IJvy)qY-TZUn#d9<dozcJb^~l`)uK50|
zKWj9PN&l?)@Dm5#X8q0`8(Pl<{wp1N>dimBulwN6!8w=Ae!_pm^K$f(xu1S{|80x!
zx2-Lo`s__hrq?{M&z5&*J-OhJS>ly9Z(elRZO7eq*G=26UH*qF&id!0^X{$K*2M0Z
zI6D8Nwo_WN<o!1z_M@MEvpuhV%OUr5L;nce^2;^9dhA%~ir2IMSeH5W%HqNszFd8Z
z@#SX@*na1A$6XfA{P#=dXOtrzzhSEP*LQhW=B`~gEm>H4;nZ7yb@FwkQ|Hbx{NlfE
z3MTGe_S22mkDoO=e&-+O_|k8mbzWi9Gt1wae&hl6r>Rw~d6{?Kynp)XD`oBT{RPi|
z^z-qX?|bN@`%iml-Ep7VK6}CZ*HcfpB6hfX(Xl4umK!#o{NmNWsUN)aiur5)=c(_n
zf8r#=xm($N99O-rweLH+_ZNPey?W~W!3W1$qw<!|Pgr-@Rlj_C)%(xvT$=iA|Musm
zmrN#qxGj5haP7rUZC<qX2cK?pTzvU$pMH1CD}Ov)z3=<KRNr;p<RjlYk9+HZ-#)(b
z>EF*=zAL6b?)q<5#eR6x#-|JG&bB;#&Fc>aYl*AQ=s&e%=i*i4rg!hX`t$hj=z}}v
ze0bUKmfZaKC#QH0{g2u&&UQR@`%dA}vy%_ZxkLE5_RQzjzj*g;D<61nQdo22nW_`l
zYCf>8Jw|(Nx=Wv3|M=$hv)1-cNL{ARZrk>YLvLSQ{ATLqN1xfS>-cZFFMqc-{q1Ql
z{^du{4rYIKLG`_{)fbe1{PN5_T_|v`TDM=kX%5Y1#wHG(edy9XH#26>Kc25Y`RLAf
zF4#7Qre^fq*>m^Vb2HRG`rBpioO@8?wEf4V89ncic|X{5Gt^sGq)Qha_2_4Fs63-5
zW>1{9=Vqw4|7QA%Kc13}>_3*A(en?Pf6Shnp?+b*WwnLoFE);krDyc2S*vn;Zif1k
z6|wWb+4#aab7*cxFPOa`z2|1A$5yR8ef1wgV`H<`4__0w^@O8N9-mP2YSZuwo}rjp
z8j*(An1;XQ_=F~`iH=CiYr$#5T5vjE2Tm8(fz$JPaQd(w9MALM_%IL7z#G6B!Uk|g
z-U!YZHi9$pCUB;(37nZXgENQC;4HiaoF!}lXXUNntYIrS8*c+=3){d6yZ}xJ3*hX$
z9h^OE2Pg6(I58}ObMOvuj<5rqlXrr1hMnMCybGKw>;mWJ-Qe6|H#iUP0p|&Oz<GHu
zIB(bs&d2+}`NBSMe%=qxANGR_@BwgvZ~$D84}uGZgWy7Z2wW%}0vF*U;3DA&xF{b5
z7Y#?j#rPPwSU3hQ&d0&U!*Os4J^?NfPJokm37iy`z$N)4xMVm9F2$$7rNSw2X+8}u
z9ZrMG@ELHKa0XnK&w|T_v*2=k4qPsr11Iw`I5{kXEAR!=@SHb3QHT|!5f%9&xMHjb
zuFRLgm1AXat8J@K*n1|NSZiCmm-gFr%WJi$klB4=yOOA{C!>>@dJ|FK0<2MpW+NG$
z&@^?TxdB+K5G{IbooLbP6rw|~s}mi1y+ZWp^>w00&ntvS&({f$-k=ZzdPAKU&>IzE
zL~pDUBYKlUOz2H@VnT0Lh#9@PPR!^n3bCNK)QJVXRUuaN);h7Gw<*Mi-c~0z^nyYJ
z^g^8o=<N!zqqo<I9lfX!5xrO^B6^2H9OxZ&;y~|Ih!efDPMqjn3UQ%#)rkweTOn@r
z?mBU!_b9}J-cu(Y^j?K{(R=H}i{7UYA9`P%_|W?m;z#eV6F>TZLIUUmbrL`yR7eng
zuug*LLkbC@57kKseMBJ<^pQG=ppPmfiauH=QS>o|#L&m;B!)h&kU08yoy5^66p}!n
zsFMVGNg)z?sZJ#HNrfcQC+j4MKBbTp`c$2y(5DrWMxU;eH2RD}GUzjPl0lzUNEUsz
zPO|883dy0*)kzM$tPmN!TqiR6!j+_OJsF*B3!6yc7U1HQq=>$_i4@V7uOwyk<xQlF
zzG|y_MknEmY!~gNdE2f!VomS;-)<l?XW@)mhI8;T^>UyZQD2T{WJ!~T^RcX13RKlx
zG9s;{g{4;3!ctq+!cr&cV5yUJu+&v`u+&R>Sn6dxEcI1AEP08CB`@=^<f}X^4Uz$t
z2H5~hL)8FFqhy4oQ8vQTST(}ZB$;4ol1;EQRZXxoOJ-P_Wiu?zRWmFtk_DC)*#b*T
z)dEYaWQC<ww!+d{wZhUS*<fjtZLqXeZLkz10hWR+z*49Ru(V5dSlVSfEbUc0EJaC#
zr6`NA6ssaE9g+i<4%q=qN7VsKr{sjCQ+C4AS#`qFCAnbfl3lQLRb8-jOKw=YWj8F{
zRW~d>k_VO^*#k>Y)dNef<b|bI_QKLz^}^C8`C#djeX#UZeX#UPepvctKP>%KKP&@M
z0G0tc0Lwr%0L!2hgk?|;!ZKJ5!ZIX<U>TA_unbj0u#89%SVrUsEF;wjETd8smQgth
z%V;$U%a{~{WlWC2GFFYjGA_ko8JFX*j925ZOh^e>CgcPx6V(JPB}sy%BulWAsuC=d
zQWBO)ISI>TH3`d<l!9eSPQfx&O~EoPrD2(t)38ie)3D4)8CYiI3@kI%3@o!!7M58#
z3(IUZ3(K68gJn+6!7^9P!BUoFSjw^tOSvk;vXCjr@KbSFVL4EZ6fVb~k(nZFi_40z
zE!Id8x8+P3w&i7I*p_RgjN9{V=kI;WGCr}+whrh*+l71S$gkbH3s-$`$N}RMktr6L
zW}_d0(J2;1j!m%`a(s%#krPuafh<ii2{}2%lE|qkmO@TXu{3gKie-?qQ!I;|n_@X+
zd5X!%<yTmFCma6NA9Fy_b@1e)Z|P?qF}T$yjvgauL``7yMQcSZkWSPA=|w#dFY-VJ
z(Ewx=jX);R1Y{P?Ko-#gWEHJIHqizohysvZv;&Ev2;>kQKu*yK<Pu##ZqW_o5j{X&
z(F^1geL#NE4-^msKtVAG6cR%~5itT36{A2gF$NSD<3I^90VIhMP*O|+rNk6aT1*3F
z#0*eY%mU@a9FQ!^Km~W<#NlByK2dZRfy(YO&>Gts*YFYR@Cc*(?YXvFfAZ^F?s>y;
z)cC}uM0F`SWjGsN>+Q|#c3<fo?bO#$_0Q?>x0yL?+f;oE@;0iz7kL*|e~rA6Q(w)&
z=WpTEHzHrf)IVb*uc=YZ20FScXqr^B3AsZxTao)z^D(l*X)fj9^BbJzTI2zz`4(7P
zqgw1~n^cQEZHH>Hr|nZM_OuG8#h$joX|bmra9Zr?YE*|kU6bmtr|VE1_H=!!!=6sz
zblB51I34zM15SrMeU0j|r*Bd{_VgX9$DY1V_1M!ZoF04n2B*iKe!%Im$JZ#2J-$hK
z?C~ASV~_7s9(%mP@z~=V9FIMI!135K)TjY_h9)&&&(NU;>>2vhfIWl48L($)a0cud
z2AlzV#u_zZ&)B3!>=`@Mh&^MU8nI_oI3xCq4bF%?<A5__&s3u(?3tR>ggsM-ny_c;
zQxo<~3TMKesll1BXBu!O?3ruSj6HLcnz3i@P&4+-eQL&@S>epsGdDOh_RIs$j6F+@
zTCitnQVaGh9csazrB5x`vnZSedzJ=g!JcKnS+HlVQ7iVWO=`uSwL`7gv-YVKdsc<B
zV$a&(tk|;-I4kySHEP42tx0XzvvsHqd$vBcVb7*;Htg9NoDF-n0cXRWP@@9&geDcR
zCv>QQJ)utp><J1dU{7dp0``OfCt%NBqjv1so79dydxzSwXYW%x_UsC0$DX~x*|BFI
zaCYp8H7a6HY*G<>Vuy;@6Z=%eo~Upl_QVD!Vow}!BK90L>cF0(Ngdd8bf^P+jy`o@
z&!KP*>^T~o1AC4E=fIw`MxEGmHmMVP&JJ~A&)KI=>^T+Ai9Kh7b7Idq;GEcV)u;=5
zt|oP1&()zW?78~Xg*}(Txv=MIa4zh*2Am6f?izJt&)uYM?72JCjXigty0Pb0I5+m(
z4bF`{_keR_&r_ow?0K5hgFR1&da&o|QxEn$3g^L|r@?u!=NWJw?0IX{i#>0Xda>v2
zP%rkped@)YSK++a^ENmy_Phhmi#=bB`mpC~QXlqw9qPlLuTOp0^C_GUd%gze!=7)z
z`LO4&Q9t(lP3p&<zeD}l^Y^JAdwzxUW6$5<{MhpkI6w9RH5$NPph*MR3v_4zdx1U;
zU@xF>0qg}DTmXB40T;ktuttN}3pQyGd%+G3VlUXILF@$;E{MHgg9~CWIN*ZV3)N@{
zd!Z%`VK3C7A?$_vG=#m7!iBIGYH%U!g$7&*dyyKAU@y|75$r`eG=jZIpGL43QMd^9
zA`LEry~u!zU@uyuQS3#VG>W}wheoj%?b9gsq6!zqUbMkQu@@b1QS8NPG={xclg6+Y
z>(ChXVtpFJUQFR)*o!r|81`ZVE{45$jmEJTZ_+sS;vE{tUc67^*o!M%9DDHw7sp<F
zz{RnbsL=%W5>1-GUZO)2*h}<j0(%LCOJFb2;1bwN47ddLq#Bj5CpD>rJ*h(_>`8qp
zVNX&x342n5ldvZZI0<{n8ckv^*`!JAB|9{Uy=0#zv6oc1B=(XGE{VP5fJ<U8Rii2F
zrJ6K_y;O&$u$Su76!uaIm%?7E!KJX58gMD>rE4^ey>yeNv6t@9H1^Vcn#NvQ;nLVk
zH@GzR(gQAyy-bZ}u$O7l4E8b|n!#SCPczudC|m}6nFg1^US_~$u$QgTEcUWZn#EqW
zL$lb+_GuP-S%u4DFWca<*vk&MEcS9Wn!{eMNpslCb!ZNIxjxNdFQ;%h?ByC<4tu!)
zm&2Z1qcZm7CY7-#cc_d#xld*6$qFZ9Pi}BB_T&L4W3O-xE&QC0F8mbQv~UaZHd?qB
zc^55wjl7X7T+PAfZ{Z3zB45P{KihkKa^}L)HMEGmVw)DRSKLO6*emX$MeG$faz*SF
zw{S)56|Z7N?3J&fW$cyPw2Zy-Hd@ABc^55Xue^~fW3Rl0D`T&G6)TT=zAHW)`|9)a
zUg6mPxd6r{YUH3A$qz4xv5Cux>T<GT$bs~(V>kW&vv2A6#EI-g&G40m;}d7HGxt(#
z%Uitri_LfIW-gPd)C=LFmP)g5bV*99(gNvJIv~AD55%i@AcM*PWK<b}OezzQS!D*Y
zs4PHMl@-XQvH=My0m!bh1BogT$f0roIaN*|m&yg?R=I&ZDi4rX<puJod_aDcA1I&-
z00mV+ppYsA6j4QhqN*rROceu)tKvWjRRTy-NkB<e5-6og0i{)Gpo}U5lvQPca;hAV
ztdfBWs=~t2owTSb0+m(ey_8MNKkmJ2owH_Ma<F>e)pLi3=kUe23-{aeBHYpCTTPc}
z9>iCBb%_>P*Cje+eV6Ew`7Ys+4P9bDHg<^-+0-Q_WOJ98ku6<fLAG{@71`D$He{ho
z1Y~=c*pbC95s@8T;y`wGi4)n?B`#!lm$;EVUE)FZc8M3+*Cjq=f0y`?16>k84t7Zp
zIn*T~<Vcr9kfU7^MUHhz3_0E<apXjoB#@;pk&u&Jl0;5*NeVgLC28bLmt>H$U6Mu4
zbx96c?h<*#r~duQGiLwosMmI%0H0>+m+-p&B-1<#|6NZz$+XD2Nv1>APcl6+KgoDx
z!z42x8z-3&*)+*a$mU6AMz%~c3$k^RS&?m%%!VvXG6C7XgW0Fx>k)S_5!tbWIgp(@
zm=oExgSn91JD3~Uvx9k%y*ro}*|&rFko`NDA33mt1(1U~SP(h1gN2YIlPrQ?Uv!d1
zkz<oAh8&+{apc4#OCU>=OhQghvLte9lBJN-lPry#nPeH{>?F$~=O$SWS)ODva^Y!K
zcnKfZ{0A$(0$hHYl}G-^mL-oIblT*@-<vsbw=?w)JaBtV^B{bqwLPXq*7cYUS>I!N
zWWL9EWJ8Y`kc~ZNL^k!93EA9ZW@Jl`S&*$gW<|F3m<?IzF#$Q!V-e(Nk42GVJr+Zb
z_gEY`(PIf@smCPbWRE40Q$3bKPWM<EIn!eq<ZO>+k#jwkLza6?MlNh;g&p{f-OkD*
zK9`OC+g+xw|GfK*)Me_2@EhA^ntSjY+h$s1U7P8U^=+m{=G%-%Hnf=m+1O@AWK)}&
zkj-soMz*w>1=-qWR%Bb7*^q@c6Oip~W=9s=Ohk6HnFHC`W=>>Ro4JtPZRSSyw3!Fl
z+h$&5Uz_=m{cYw)4zyVSIoM`F<WQT1kRxptL5{Xr6gk#rG30og#gP+jmOz%;OhQh!
zSrR$bW+~)!o28L6ZI(gKwpkWA*Je3nxy@wcLYEaDf&;bKWkuw2mz77nHX8g--O0x%
zXP$7(3(Y%e_yBKwVt#1;NyB$ftq<?|(RC(cZug_XqgGpoj|MNj_~8vty#B`anO~l_
zbl$1M&#k|Cqr|UYx^nkP?iHrqiT8ymromUOXs4JKSvSRW$oeU!N9Ly(k8GG?24v$D
zGa{R&m<icD#mvZ-DP}>oPBAO8ZHn2Dg()T=+ozfRIk<0&(@aEmOfv_vbDBAkUDM2k
z?4D+BWY0A7AbY2o7uh$>e8~Q3=0^@pvjB2%ngx+V(=3Esc!d>4uOgaxmcEk}U)y^u
z4WBs7ym({x8|_Aq{{P?Suv<5L@0t7l`=-ZszxZgadObWnt!U1Nr>7MykglQw(pU6A
ze1!)xR184IiV?_EF#(w?W*|$&0%WaNfov5UkWdkT>=iqZSP_976$g;B;skP4TtM!M
z8^}}f0C_83AYa7?<gfUF0+j$zuo46cRYE|KN(3lci2}tcF`#%Q4wR@QfTW59l&mCy
zQk4`?x{?OUR5C!>N){+r$pOg~8K|(fuzq-y%^ZDei$LYIWuOae7wmoU(cK@^E?D};
zjn{DF6Nl}4*a4#hZo#qz`OyKleBb2<41e=`);%&SS-x%Y%&BYc0do)8`?{pn=iGnx
zek+4BXT#&!@e4<nUske}due!9oVDMc_mYn82I~8WdI!7!zi=NZ?7){l-baea<@-o^
z#7}13|H!Lfd}`*aSkKlk7@lOuCoUuGG6KATuni*~?#O<7-nu&c)4cws;=Oj>y?>qg
z0_QKCzhd}je%mzb^L^$Xx^efJ{NUB<;U~X;+EKsZZ68^8-(>~$qS4==M$jx8{S9gb
zEs##o0qF%j5HIjR2EhPi6pTP7!31O$%s>{w0%R4eKsLb!BnSeKU9ba*f(Yag96(ON
z3FH!7KyJYe<PkhTUcn3G6MR5^!4DJ=0zg3_2ow@RKoKDV6cwUCF(C#N7vew(Aps-_
z5>Qe|0;Pl$P+CX>WrPe+R>%V7gdC78$Up_5uxNO%nz_pgMWC`!20G7n-riT@4v(ps
zA0@lrRy{mw$N$ge(a*KJuKBR@@aEme<yNA;7w%k|7SU|QD^4w<Mb@>54q4wKdSt#u
zcw|G17?6!EVnjBzhzZ%;B4%Vui&&7YEn-EswTKN_Xb}O~-XeBnu|-5=M~gU+oh{--
zcD0BL+1(;;WKWBDki9M9MfSCb582-$e&j%l1dxL*5=0KQNC-L7A`#?hi$sxQEfPbH
zw@4g0(IN?CsYN8@WQ!z`Q!SE0PPa%JInyE;<ZO#%k#jAQLzY`aMlNh6g?sU2zLgY_
z%Uekq`C{9}d*8wJ&4KcLub%z8cjTGFc!m1((dkuPR-ZE>O<wck;T>b<{|`{ooHZhC
zUJFiJ(t^|Fb>MU*9XNen4^CgwgX8l&IKIS#Gvp263?&0NW8MhPSTce$<xSvBB@;Mv
z-VDxMGJ~_^E#NFA3pi`u3eH-xg0tmq;A|xuI3X{96G{R&d)^MtUb2G|^CCF0B!Y9~
z9pD@#2RLWm3C>w^f^+3v;9MmaICtI+&RueY^W;6?JS7h}Z{7>eTk?YQ<$d6MB_B9{
z-Ve@S@`DTH1K<Ls0Jva22rgI(f(zwC;6kMkxJW(%E>ensi{_)?qNOOfSUv_WR*Hd(
z=i}hwr8u}mJ^?OKN`RB{5;&<OflKC-;F6^zxKus`E>%i_OXt(z(xo)GOg;lHQ_6tL
z=Ck0kr7XByJ_jyW%7K&fGB~*;gDb2koId(-TqqaL8BuXX5nQoc1Xo^B23IbZ_qxGj
zfBk#*_>bh7H_h04%>EEQ-R&_E+0kPTWM_{#kzGCJLU#9<8`;xi9%OHid69iR=0o=P
zm>)UNV*%t~j|GuKJr+VPZfC_O_ddN{weqafc0b}@y=ukk6L)`b((ao*|8Nu2%d*e>
z^P=36S<6m2>ExwjONVA`?Bvmhr;I*aHu`Wyxo}o-2BW9PIW}wdKB|4^jPHNo+yj=(
z3k7~SZ@(q}`G+3#ox=`aWOq1SZjaY+uueSs_=W2K->)8W<afV!)DIRMbL?>%t=?!d
aTZI$8Z##mw1P?ys#K>~GIN=ME*#80~x=>*N

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5CFCD880070835B7.idx b/.cache/clangd/index/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5CFCD880070835B7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7bf0a8cef4ce6d75b7cb9633826182a7ff77d22d
GIT binary patch
literal 3550
zcmZvedss|cAIE3!+103FDk>toi^yIp-KM*4(oKji(s1gPE}|)vOXfITMi)Aj3U#ES
zH@Zlj@=`C#DX*8@B|4JmijyQIk;|FseRg~O@y_$idiH0&zqNkf-&$+0HJ)zH&QGSZ
zSdKw%dxgQ#n+O(*r6l=AMehu2X0uqG3M|&%6L;4MJ$ft@@BQ#w<;Y9l42H2w(Q1SE
z<A#)sW7|Bk)@oRXS|@f^UESe-am_KFT*sSYv+7sXp#?EQlQ`}lDv3Kh#iQj``l%D2
z<c!6<k^9}<U*3MbooErTGUF4mC}S+OuQUGohxfHMjl&NMUREEvJuu|9{aaV}*BdVW
zU&nWA9Sew5?GBHLu?mh?EL@Tk__^=xM0fwdQH#rg%~f$8BQd;<{YTFlRVB0<KUf&q
zp|MZv9RHcUmTzUm#=}4ynATa;?5jT~KilWb<MD(EgR5a%`cA)i$M2XIvK^##Eh?;h
z=<I*$v3Fc1;A;u%_t@5(zpky;-f-sb{?QXxo;7_IoN*6Yyv*L~eY@5A;q-9AHred(
z`TF(sqEFum<?44W1%!HQi?b)Y7v}xE{pql(=>h9Q<#<|R_}!Dvn|sza-~8P<xHu<s
zC=rX+m&4SDwyMRJUlaAuE^}^BDvI=pw*6I6bfdHNUssbA2O5LQ`^FChg>5+e<f(A5
zomJnd5qML(wI#}Schl7e)_=NscxDHB53l=Dv+B)k&f{6Q+<f_E|J1sEyjZKDF8O@w
z!|Ot^xT7X?Ue<EX{FI-vn^@k08`xlQTI;^r8JN6V$gI8G*X9)m`PtfMx?D6bcGeWC
zZ(Ysva-WmGTEzFh7xlesoKqxjEzdtPXnXKkRaSfZg<X48-xb$3UoIF+k@3^4XfLwo
z2@mV0+t1D2LWPFyB8QERXRKacG54CvBZZ+gm78XI)fQ_`|K)V<@RFB>`hx4ay`_(g
z>;7;Wu)d;mKlOFlXiZtxp(iya&qX-~&j)Q*_~^P<Xz%J0-q&L<`5G8wIaHwbb!C})
z$ii)I^|GL+_4uhwn~GvV)GlL7N7L(%#N3Utt1AAKKihlf-%m`+n+yAe@54Cr$i5Dp
zi+4&~vc=OfyzjjTyY-Bm(-CTr7upj(&A+egRlZtfV*^j-siLl&p+TORL3O2!Z=POI
zith0-a>LJw=Qj2v%{sIwuCC2lPl>A3yXd_#Lnc4EYV=%NpwQ|0erNW~$9%tgPQQst
zwvO-LX<>QOH)!OL`std_M>qB)i1e05j+9XM?#(sHt9z@n?b4<1SAv4Ww#MGhdQy8{
z+4w@i-LW3KypO6O=Z8DOkJWc>IjCw=CimYLEtv_A__s&YzAL}l`mLCI`LBV+V@|Wx
zy`L+-e(84o!pM+7cfMlA*7E+w71IYE?oUbL)GHk|%Q|(_#`4asw$kSlg;}i`MX#QH
zbJ$_$^drjbzHyh;svrD|k&Ulo4S(!?5**8JbMY!;Ma2jM^CWM!-Nf)<X@5+W0=~wR
zCE&^O<R%-5GWl_6`f^lTByv^2y(O}atj^lWhL+pj8VBslm$jYF@{q{Ns>%x{8(Mxj
zs@-4{(HNkBdr4$9Sv9}OhL(GmhJG6RiT#nwS}TzaWemM18(L=V3=WBk<*-<?WL>)B
zDRWI8S3v5MhV-ZsG9@h#awXjvqVy?4+C5i5nNk)Axl-;70eyf(0#krQ0#|@ULVbut
zLQ{xDLRW}H8u5+Hr^*u;^NkVW^Z8RG|K((-*xl4iUQ$dPMy6s+0c09R<|0hRVN@!{
z*rw7jDi>i8hXGUvX&69t7>7YrhiMo@bxXdb9b?&wZ#6|qRR`5~RsMBNYSM%BL@bKP
zUl^i1C{M(mi28*g-~kW~hybEN4~S?`1QE@YZ#p$sT!97OVu}{b*kYs7aHUs9V#<ni
zL~K}5j!0rFfDp8T2$}FrZ0L&V_dr8_a{s3_O`NRgveuLgo`jRhI5Rw%gOj-^m*8Y6
z${%oY1Z6RsJitc!XS2y8D900I0>S815>6%KOaW95PUWIpf>Wg^f552`l*Mf702}F_
z&8Ci^98XXQ1fzo_9H99?4i3<KpachKKJWntXg(lj12iAVW&<=Ih$jG=4<_Ld%?ERE
zh~|SOI7IWo4>&~gK`|Sm`Cv91qWNGv0hxRb$!}SoN2A+uLSV<S+d!{2A#j-K;7=1h
z8Myr9;l`{z8niO6SKszy&RIgTB{-xWvS1uaj~;<5033i2<iIKNUah?7b>>nTj4Lpj
zX2heD|D!u4RQ82tCtGqzH4@YVbZHt1fHG_(Aj5&iZkK0nYN(TTtR|>Bv|}{^P=?h6
zWOxU``mM}x;(@ed5kXa<9g7HnGAtq>!@J#=-vR5*)<`N)c2f4G5C2%#FaBAE20qfl
zG>uH?!hTK4nA`f^jFP^fzN8;AVZM|fGGV>|A?OPcGUJ<>(Zi+nobb?(^$V7|ZXu{v
z)O8C1P=+l8Wcc&}jaxP2`SnsC5#%1kM+CJ8@eu(Kf)NlQbH4f1L2-6K9OKt@A|2-n
zW>1?vgBec8Psg7gu43<HnKi;9XUUZCX7E($!==gF_DulUc0<zFhOp75OXCXc3HvFc
zr&UIN@-Ag|BUz4-b?Ajqj!|{!1yGIwlwmmr8GhLlXP>F``!ng{`6+Lvv(Ww}dtZ7#
zGuhDBg*J-GPhQY;P@kd9s2oo%Rl4vhMkF0IN_|OxTU?hTo$!2&EJZEnV^k?>IUfTk
z!+Z>)Y|Xcxx`n<euB&+-`JaA{WY_>QkTx!%FfyDWDu4>4mqD^s!l-bDKmb5A5C#wp
z20%oEVQ@;5{0|A=B>htowR3Zni}b-q+Whb1OFyvK0?yLt`zO3C4P@5)`K<G0t(%fq
ze$3-~=5ezyD13_q-9uKf2^m>A&NKz?bVc3_Hzg;>g-Y^n%aqk-&YC^P!qR%Fjjf&i
z5*1*jW2Q&`zh717&ePD;(q5puXfXv142?`o&FAyg^^KRSsJg7!fIF{taK%>s4M@G6
A=l}o!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.15161CADD4CF1397.idx b/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp.15161CADD4CF1397.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e68440a9f920628539d11611e2d799498a3278b1
GIT binary patch
literal 13202
zcmZvidzcgDp@;YTCM;oh<<5d2unQ<^GP%u6E`y-SWRjWOW<)lyphj4vMj-Y8DiW)S
z!Jt6E5}(FfDn%f;vUsVscq^^-Qnf|Fdc)!)mWm*>G(`nPIbnOk{{A>TkHfxs`SIc=
zJDW{rx*7<q<{BEL`Rek;*DsrEXlQ5{Qu}Y|vL(ru8su9V8kRrum-w8kJ`o3h^QcEF
z7mPo3FJ7=bbH(hf{T%lXTt8Z#{mDar+OU4o<b$Wb`*rp1D{dIywB^0?6Pth7y#FTa
z>?zEj8m|83n90M9*0=vYVb__bCOg~@-|@wvnUgDt-KP#Oy8YRS%g@Z8_VB;V5Bz1&
zH@6ystESvlxzw^FRs8J8(bFeB368$~`xUznxjy9XaO`rAD`x(qja)kI%#tNjUi{tj
z2Pc;%of>yMDIUu-$6sjbJb&d-?~}8ReemV;Prb6}u7Qykx^MBCTdEhXSd(W*FCPA0
z{Gq^s=5Kc0UfI)ZJ-y(G1K+jfrp#@bn!NdL=U!vaUeBnfuNb!XujxbYrSHCDSa|Uo
z{`vL4+;D4YO5)yU)j4C^H@=x%G}e5>*2RX|tDHCdWXjXGfBVlrrj=LsMW+nicIALE
z&mC?5$(sis`^&nm+uF9@eE)?P|5@%@Av7&qKiz-Vh|t8B?s?XFV&lS#je2_K^CveB
z{MWr*e;u)R-Qb^1eC$7SGk=;=-MaV1rN*a^M7Pdf_m0x~8#1@Nf5wgh?;hSiAuwlU
zX!%=j#7~Z2(cg67IeqrfOOu!0zGw5Y6DuzrOZ$ymQV4igKi3*~@Aa{dT)ozky7A~g
zesS4j_f7I0SB@ST-uUj_HxGX|7H{72o4*fk8@aEAt|vdN=9BVfe$R1c@R*z4)cTFs
zvTECwwd>v(J34gek~2SDX58ZyyL;|GI&IA--)`9V(UzMRHZMG|xVq!9sn>ZA8rNL(
z&K_x)Kg+ypnDEeBowMGTZ@=e@;)o$n?9*Rc_KL9j!tL`!_7dgg#3c*(=94$&;?oCs
z_TPB;nvI*jW<T9f@t%0$_$Etn_3InBW9f&j@9&=4)Ok^O^Mvv%FRr_?amLA0r|0dP
z_Rir4r@#MgWbcOEqs?oJ;kz2Yd1~+y^?mi=+8xU-`}FPKU$4KhVf%fdw;ow|*{?$1
zcYR=Z^r7GHX@88H*X*2j!Lh=&Ez^Hoch#pqzWLd-bm!$`R)mjl*uwqmvRj2)4&AqK
z%%+VE&umJ~%67c=_`c=xgC9Ki@#$w@oB7earkS%R+}U_Mef`dZzuLTdLip6lFF$^H
z`^ml48+SHJ6FWaV|FO^KfBI(AkK@L!{pKUzocNL1ZyohN{`MdC9NseP^R>G>PTJS^
zfA0RLUGt_a`S`C7ntzr${Kaphi)|0@IdxjOW%qXzA9%eq^U*`|onN2!e7*9My)SOP
z_}6288@qMazZd@Efge_EZ7;t4=6u(#qWY7h<NObz7kxZu)H|0%hrNF7$(t7}Q9s<a
zdGGR<H|-zw?3Vqf{?@U4+m*i=yY=ona~;wBht&W1W=L?y3lDtTe&)OxEB<^>Vd3X<
z_DQ+J6CE#|n%)0zmfH&dD4%)e)$0cTu=C}q;<9)4B>wQlpd&kSlNT4+&h`s`(A~pq
zuPlFb-;OoMD#zOIxbxVfdtWmaPb4R<avXW><1_yxt2ZvaZSlM!X=rF1zjK}RD_2`H
zojIqWtF^IpK<}Y;*697!^xei!b7A4CW*V>Qv5jMg^d9PZ$dmVuzIn%yKQ_}VYkFwo
z&;h-Nx}G%fhN&Zd_tL>;I;*BfHI5q8d#LNnzyJ5KZC~8>WHaro=`oFC`u85{T9+@n
z@0Z-7AJ1#(sOk3B_DQ{my1rqs?dA1`SC2N+SWS;>95<x*P}g@Z`04)N_MA>N)2nNG
zV&lZl-a}pAGkjO(%coy@yrp4wP0#2z<Lcf+U2j=A_N`}5mS1kB(`$No<M2Vfhq^v8
zx9tU%KgKmvy{1PtjvUZ?sOw|b?I>Ne^Q(u>YnV~fZLMwXy@$H~VQ~DJ{Wrb(MKiss
zriV2S8_;{G>xL!QFIf7yx$nM_n|kk)!Ykiz+<st6TT9od3r3Bve{O5kK59}QnFN!&
z{{P<6Wl~JxJ~9htaAw5}julvNtipn`2o`V_#RASMSixBpD>$2A17}lg;Ov4OoL#Yl
za|jM_4#feE6F6|3!hz!j9vrXm-~>ScCny3qr{DzVRGi>kf(x8Wae;FSZg6hJ4bCHY
zz<Cr8I8hM6iHZo$D|o?q6)!lS-~;DVeBdNO0w*aFIKSWr=U4pTWI+ZeD>Ar%5C9iY
z0^ou|5L{3Rf>Q(qoT4b;R6zx&Dk`{;5CRucLg2ze7+hEhgNq0ea1kW}E-FO9MU^PH
zm=FUOQ)1vWK?A2L8aQ3h!Rd+)F3!hY=T0tNv-nx(sD3V$ntIQneP7O|3^A=CeP5=g
z0x>-RY|aofdUJu8(X$!CqGt<)MQ_Ov3wld|SkPND#ERZpAXfCY46&iN6^ISJJwxp1
z?FC{-@5m4bdPjjc&~q8Wq2~&OL(gXjkDf0O9=(tu0(zl91oX}faiVt?h!edlLtN-x
z1>!>Q&JZ_xcY(Okdosj>-cukR^kRmH=*0pN(R(w*i{4uxUi7{U@uBw>h!4G#ArgA2
zKqU134DqA)7l<FdoFOuLxj<y}feZ<t4-`lMeK13U=z|3kM6YCsf?g>Q1-+UfDtfg*
zRP><?384=aNC<s6L&E681rkOd$&d*8NP$GqM>8afK3X7A^sx+yp^p_v484{i8hWij
zH1v9g=;-wV(a~SSUvqiiDdSrH+H+KYew8Opy-%$A&Kbim7(TxKB&DTmc>C~4ePq&1
z)8T9qH+2FfO^bliraSt`teIhFj+<d;PMTq7PMcxJYAo#7I14*A$-<6Jv#_&h7T8(h
z7T8&m7T8(R7T8%eE9|UsE9|UEE9|UkE9`8V4R*G;4R*Gq4R*G)4R&_T4m*3?4m*3&
z4m*3=4m*eDfSn`mfSn`hfSn`lfE}lCu;bz!?6@QcJ1)(^j@NkD@o^q@e3FM9pXOmF
zXaeknxBxpLDZoxh3$Sx)PS`o)PS`n<PS`oqPT09L7wlYd7wlX~7wlYV7wp`c8+Puv
z8+Pub8+Pur8+IPe13ORL13ORB13ORJ13OU@VJF5#*ojFIc4AtDomcb1&Kvi_&YSeY
z&YSka&Zqfc=ZpJb=S%uv=S%xwCutJwq__k-DJj8DN=vZwYkt`I<9^urlYZFw(|*{=
znhZNRF2hbv%CM8uGVB6c0Cs_R0Cs_70Cs_N0Cqtw2)kfB2)kf12)kf92s=enV5h_t
z*eOW`c1l`-ovNv@Q{yV^)T9bKHLb!fq=jG?iicnqN`_z;N{3(<*21t0$HTA-C&RD{
zr^B#|Xc5>&;t|+Ik`dTN(h=B2wJ7YO@hI$~$tdii=_u@CS`2ovcno&2WDIt(bPRTy
zrom2&Yp~Oj8tk;R20LBTVW-D+*y%|fc6wTeT_@jp?n!q`*KB?^(AE6a=cxWe^Mt0}
zmnHhXxtc_i1m6}#GY~7XKo-#gWEHJIHqi!T7wtd}(E-GX91t(^K!PX$IYlRsOLPIb
zMK_Q~^Z<#X2;>#LKt9n2B#9D`U-SdXq6`!e13*DB2&9M#kSeM`Au$9L7Q;XhF#;47
zqd+k+2Be7^kS^*#Gx?cQ>rbfaFESmdJ%u{^5)<6Zd1_ik`|f8`iJBfju29o9<SI4&
z2RX->Rx+?JGN${Hvj)>$hCXl3Q#1C=C2Gc=xkAm@Ggqk@d*&Qt#-6#zn6YQh8qC;Z
z^OVINTcRxX*a~H_$5tteJvPU%*kg+fi#;}LV6kV(Qw#PiC2GN*r9v&(vs9@CdzKty
z!JehaSg>cw8Z6ke=BX8X))KX1&sw2Y>{+YSial$Nv0~3!WUSb;W(`*C+49teJzI&|
zuxG1K8}@8fYQvr_$JnrED>63h*|G*3_Uw6T$DX}J?bx$ds2zLuDz#(Jo@4CTvlkgV
z_Uu`M9ea*Ebzsj?q7LjiD%62JN0mCT=g2V*>^X{z1AC6F!GS$4PdV&yCCXust56Pm
zT$OUz<8lm#J+8=b*yFMW4tso_^4Q}`l*b-lp*;5ZD&?`q=NKM)e39X?$7c;Z_JlkY
zuqTwLfIXo?1?&k`Dqv5@F#`64A|qf=$QlIfIrG$sJ!gqJvFEH%C-$6G>cpNi$2hU)
zEHX~)IkN^Q_FQ@D!k(){UD$I~s0(|pDs^Gcm1A7ka}^mE_FP$m3w!Q7bz{$6qHgTD
zE7XlWca^%a=gu*1?754K8+-1o!Hqpno_et7DNzshJQeD}o~KGZ*z@EV5B5Ao#)Cai
z*5JXOn5QE4#1a*;CswG4J+Vqf?1?!>#GY7WMC^%KgNQwEo_ew8Em1G_ycO!jp0`T9
z*z@KXFZR4e#)~~~*5JjSFHe2g^OdL%d%g<wVb52kKJ58&j1POhBICoJFKh5&Ps&pX
zds2x?*pn($!k$#6685AVBVkV}G7|QrtU<z_KTrMG^OvX}d;SXbW6xiue(d>kj30ac
zBIC!NKWp$~PtH>rdvb}&*pn+%#-3cIGWO&gBV$i4GBWn$tU<<JAWs9>3zTR8dw~iK
zU@uUm0qg~GOaOa<A``$~AZrL<FPNu6>;+3Sh`nHi2C)~c(jfMNIVOm`V37%8FPJq1
zv8Uvzf<2`~73?V$s$fs4QU!ZTj#03u6d473O4guYPt8*mduoZQ*i$Q1#hzNFD)!VI
zqhe1jGAj1etU<+IC{IJ!3zcXHd!Y&qVJ}prA?$^6ObC0SA``-1C~F8|FPx`g?1f7-
zjJ<G$hOrl}(lGYIIVOy~aFGdPFPt@mu@}kH2=*c+8o^$qLL=CVRA~fzksK4jUZluG
zuouZ1BG`-OX%u_W5{+UnTA@+wMXNN5y=ab!VlP@`qS%XO4N>gH@-&9MSc%547pu@1
z_F`2U!(J@M#IP4DGBNDMvW6J;v^>?Yr<JINJ*`4D>}gf1VNc638uqjzqhU|W8Z_+b
zd8%ViFHs$PdWGuP)2md+o}OcL?CC{D$DW=w=zTr!s-q33PAuC;3|(Eut~vD|7ef#3
zTE9DY^gLR7QN4Ben7$X)N41U`Q-8rA^6)`++SjEgYH!7xjHZj>M?9k$h&8f67NZ5o
zYP15`j5Z*<(GKJ=I)FGM2gDnBAi*dAIgL&rm(c~}HoAd4Mh}o^6oI@(FObjZ1Copq
zkl*MBl8rJ@z!(4u8iPQJQ2|noDp1H60ty?$KoMgEC~AxX#f&i^&8PwCM*SQ`T^nv{
zW!K(ETDr!%$4d2gXzN#=3~B1U;-v2-18+C+?u2({eci-|EOiqJ+22k4$Z|K4kptZ%
zfE?^5L1d+yD9CCzQISL4B!nF9CSl}AH;EueyGax|)=gr_S~t;<^=_i~@r>Be$zT5b
zvVOJCR>y#jLG>@0`h_^Pm(k8%h12)hGL?yG8+^9RWnxBV%Y;R?l!*n|S|(OxTbbC9
z?PX#|c9e+&nJW_xnJ*I_Stt_$*;yt|WLKHEklkhCM)s762U#o=5!qWNUSwaH_>iSC
zk&yjm;zyRtL`DvjNdP%mCP8GSOcZ3bOjP7hnS_wTWfDe?lt~0RS|(BCSee9-wKCC=
z^)k_s6MW*_JDInH{eL#ia_Y77YMb0PvwnqCOIN$xKC}MycBk^)wHs{Kj<XN2whpfT
z0IQ#0X;bedTzwBPJaK1#wsYklK0EMyZQIt_c3nNN>bQJ?=bg2qYc~&V9eP3i4b}Qz
zU3*dR?6qI@zp{S1uRWB{Uh`GIxptaQnQpEBGg`XLDKik8Vu36v3y?Ks1+t}VK=za!
z$dPgYaVZXnPw_xPN&s@EoItLW3&@>v19?&&ATcEZc~f2>U&;q0r6eGK$`2%`WS~GQ
z02E9Gfs~X2q^4A$P$~ozPKALYsR&Rs6$Of=VnA9-1JYAE&>ViwxmSdr_$qf>=YdDo
z)gBU)222`QzgxF2f3cswcvGNu^O%8STI;VH|BjzNzn{G6tJ=-2rdF<gz2qx1rX5a<
z|L~34&2tTNC)6KLpZz^^#m|R4^-1mK_RjXM`UCqPwtgr24;s0qc5}ywj`8(BUB8U5
z_Ha9U4Iw<-Sd*>(@2WlASThjIvOpHr0%T>aKsMF}WM}O_4%Pw0u^bT3@<0MB06AGF
zkc)Kzxmh=mhxGu7tO(>~y+A(J2PCl)ke~Gf$*c?%U;{uwHVCA!3XsaGKp{2+6lTLf
z5jFx8WurhbHU^}z8j#NFKplL?xp$eqGHzc_#i!0b#~5iE>8#)9^%GR>ImX!wMB%>g
zA*R0XxwU)0hnSJs9>OA9dWZ$t+C!|!wjN?bw)YS_vZIGMkhvbhA@e<iM;3aBfb8rc
zPGnaPaUr{Vh#T3{Lp;c04-xzL+ZiAEgU^+@+GlH=dz@5%Zc)Evw!NwMQd#(HB~0_+
zD<xqDViPQoC1C-wCagfVgbm1^umd>~4j?YU0r3eQNJt1k&V&=lm2d&M6K)_+!UH5G
zL?Ca%3*<}qfTV;3<WKm4<b(_qNCbd_i6D@YP=M5g3KU9&fWnC|P$UrniYB5!u|y0=
zOK3oPLO(}Ku3t3w>}AhOmvk*noc-Y1vxnYaOxK+AZ;b48l?V2pf6aB*Cg#p-xbFW{
z?YC?CKJ4B77kmG%|9&{H?+?uH`><>A{6#m^FuWkj82UB#Hx6iSIdAZwf$9)hx@bs~
z>Ki(8*zi&3yFF7bn>y|CD=Z^S;zZ}g<Nv=~TgO~5wyk}_qzflsa;e#BvvY!L96y?M
R_(!w`L)X)Q=2b}Ke*w`3=GFiJ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.C192E44A2A2B2438.idx b/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp.C192E44A2A2B2438.idx
new file mode 100755
index 0000000000000000000000000000000000000000..078d3d3f180ce04d3953c07f580f6aa056903d42
GIT binary patch
literal 13204
zcmZvid6*OR!N>Rao3Mu6mCI#8R$v!bQIp9vGf8GRG{;OPnIw}U1Qx^y3zk%fz8p#o
z)x>D1O286d<3SWT46dwK(JGeOdeozeqV*^|_)4WJh)SbH6s&L9zG1(AygZM?e)8Eb
zAD`TtooPWuS;I9n_!b0LE?KyIo}r<kVR-GoWy_Z)&#XbdrJ-TvhBxDLF5M>$JNAI6
zmkY*g+)J)pnYm>4LxUW54p}!wp1tqhSJ$uWnt1T!55KM6a>=6cO<UhRJF(@*E&H#w
z&Yr})+Hl$Z?Gsy#*1vo*;rX8)pXhMkcf+Tj&74?C?Ee1H;#E7&UHQ|3srP+tzWdFg
z$F4I3R!_RAa*^fHRPp1(M^1jbFED1+iR*WN=K6rU!STF%Tru;FK5E(2pO!A2^xPkw
zJvgz{_5HY`N$JZ>b9`r0_t`g(@H{r_%lAKj_VE`M-!vq=v-et$xutsE^|$BQF-uzC
zjo+&rXg>D*s>+^b>&Yu0Iq*YAZqmG#DajSLINve$y(5l(;*zu8`D^;Kchk4raF({@
zcK+FQ_piUMG%0cGj^LcJof}_EE*@)M^w1K+?A6XiKcDo(s_(vfF|EF|H!^9&wx11d
zf9go*&tE&Z`OUQtZR^;+;?DCfcth^FUTC^%-8BDABSYu@{?|LK-)_9>0%I&a^V#DY
zhkSi&&tFHrvv%08&fWaYyv(bUst>*M+%n@6ha(ToUi+5X_B5GS-aq}(!EYbhKS7yu
zb7<x3uf&g!zy1u<fu~}#M_iP=Xw{xA%fG$(qOo+)xTOWfbK6sG%DaCayWz4mmeenf
z{Nt|4oA2oI9#xMVZf$(~mKCjUN8`;~pZ;W6$Edw6bRGHeHa;nD;rAS6hP7Y)nm%ac
z*45j#u37ua*fF8cF8t}5<;Fc8skiUWBU5kR_ucxvA8uW7Rr6H`mQ)|zJmm_{LF4V`
zzqQAAmOsn9ZJ2QH>)o^7lUM!v)8feCkL-=Ty!-{>w)3{nm)Hx{U5N{?<eQIQor_N!
zEbjlsp}8A3{fqr*eZ}+b&ZC<w#oPY8p8GO=pY^@nQ<}QZ*S1V3zwq4JpEXWD{{6}M
zd#Ap2=$>iseHVUb{q8a5HAU^F#$%5UTN->XcyP_5%O`*Imp?9yy|RA$9ii7ZTs8T@
z(21V+4G-M=$32~!x%th`spouI*tT`rzt>*+(Z65&cxt-);`ZycqwBYFUr)YHxc0L<
zu4>=3vEj*0sae?>FF(9@rF_r(PyOxWj+bYCIKOG;><KqE9!)QN=HP={ZkwQefBf^m
z?b`m#-0JJcKDc60*B4UH^bK!J>i*OJY|iZ&H)qjbLUXly$``B<-O{dBZNr{tn=bzJ
zpBwLT98(qtEcZ#pt#7gWPd@$dnOAbs1LI%Y-F>6s-Q(56-)wl}q5Chc+<(oS>py5+
zc<m?eANtC5!<!?1n|O;~uvEO~$fwsi4w)n8zqs?Yu5|M!;@0E~A2vU~J8U^L=D(+p
zd2Hs6_T1AuCOz<6(@^2{r(fK0^y<fZA3ibp!rfE<{_2cT-%LIDd^WW1>?i&!vPw%Y
zvi-`RcO|~a>^twVCw~7;;;zvxo5ug{cY)Pki9eV=Jn7lL;^>woPo8JmvSobJ&#5x{
z>r3~oc;d$$Q(pS|u^%tpIeOmusY?#VlwIx<<G&hq+s1}vzg#lENE#X%$3L^y_n@nz
zna-Tk(9_n~Hn{&#d*0~()AYl}k8+xDX)}%2^w`F+!}||)J^Zm-$E<kt@Qcm#XEi;d
zam3*MLtS?bSu|zjAAWzZna--|(T$^r_8;o{;uHV;a@(iBe5{#v*K~Vh`x*U*x{h^=
ze>;3^#jPz3vuk?#py`+OAL@ER{-(QL-Xk@h*)YAPJK8!r`ww-!Xqau+I>SpxnrXDA
z$2E={-hZg;tv8Q-{mJ9yUCne_O}93-4(&hG^`2V`cRqZ=wX>OCR@3J;p4;7jsO$a9
zOXX{KnU|m0Fr%hB+d8}Y4|RQbUdK+B|B`E_v6>#$IBIbJp{{=%7JqX8)h~V8OfRkJ
zvl`DD+<&O+hNTOyT=sYKz-`mj)PI{4UO2IF`+-RvEj^>p89ly!-PWjcbk_ixRFhW!
zzi;U=t7afpWq~ZJ1<0yefo!S`$gbLf9I69|Q#l}B<$(lM0CK8MAeZU_a;t72Q5AtC
zRRZ#;9w4vk1@fsrAiwGdl2sW<Q5B$o8URvN6)30%fkJ8sNK-YSuo?!6s1cy38U@l-
z9Vn*8fM)TtT=jc?{n6)Z>VMP?+yP7(V!E9S+{~r|G2IPp&JZ(tbAg!Avl+soXA6Wy
zZ^;k~dP{*=&|5Rair!ivR`j+Ev7xsWhz-3xL+t471!70<$Pfp5M}auda~Z;+=L&>F
z&u0jao-Ys{y^tXSdZ9oB^v(=%qIVXE6TK@#T<Bc|;zIAv5I1^vfw<9&86u(=3q(XO
zWr&1cDi8_1Cqq2wJq6-H@68Y|dT)Vv(fcyQhu&8pKJ@+!@uT+_h#$S2Au@WoKxFhv
zhA8Nj0#VQhG9-XLP#^*HYKExj)dEq`2QwsyK3E_@^q~w1p$`>E2)&je8hWijH1y#N
z38N1eNEm%2Ln7!S1rk9Y&5$ViXn{o0>lvb>*9$~PAIp#!`dER)(9h-PUR-~YXz96}
zzx*`SUy<{qssE*T;5no9oYwL6SG|^=*3Q<h0W#^PY4B_kH+2IgO^bojrW*#xteatH
zj+<d;PMTq7PMcxJ>MZQoI14*A$-<6Jv#_)17T8(h7T8&m7T8(R7T8&JE9|UsE9|UE
zE9|UkE9`8#4R*G;4R*Gq4R*G)4R&_j4m*3?4m*3&4m*3=4m*eLfSn`mfSn`hfSn`l
zfE}lEu;bz!?6@QcJ1)(^j@NnE@o^q@e3FM9pXOmF=mPA7xBxpLDZoxh3$SzQPS`o)
zPS`n<PS`oqPT0A07wlYd7wlX~7wlYV7wp`+8+Puv8+Pub8+Pur8+M{D!cL5fuoIIa
z?8LMPJ4u&dC&eY$Nl6KIQd)wYNB6+a6ZgQ*lk~vOllH*Qt9xPRjeB9|O?qMHO?zSI
z(|xe>#eJ~zC4I2-rG2pT>wehz<9^urlYZFw(|*{=x(quxF2hbv%CM8uGVByxft?ap
zV5cM%*ePiRb^$#AyFfeuyFfAkyFfYsJ5^U<r^Z#-sYw-fYFdR|P!GZ`7!SfOm<+-$
zm=3}&q=#S^iicnqN`_z;N{3*l=^E^`xCT2dsliT5Yp@IJVc3P^Vc3O}Vc3P!Vc13V
z2<#&92<#%s2<#&12<)PI6n4>g6n4>M6n4>c6n46<!%mOuu+x(|?DVt_yO<t>T`V4h
zT`U=cT`V1gT{qu-`b~FB&uo4+&}IB(r>Xv9)`X`1A7lq!TuqY62QQ0~8HklwAd6%H
zvPxDUn`8sBOLicK<N)F%4v3d{AVCs<oRSmBCAomyk{d{rL?B6$fIN~1$SZk)e3B2y
zFZqFFNd{6R1t=f|fK*8Z3Q9qskQ4&aBn>Dmg@Gbc1Sl#+fpkd+ib*k`nf%Nt^*2=Y
zk0lP&-a?)FPy%k{JT<MR1GlrOL``=iSEy+la+R9CLC!Izn;F;_8PlD}S%c{&!+<yE
zsTq6b5;bGbT%l&{nXA-{J#&sRW6xY<%-Az$4QA}IdCFpsEm0PGY=yGeW2=<K9-CuW
z?6F0L#U7hAu-LQYsRetM618B@QlS>?S*p~6Jxh+UV9!!yEZDPT4HoQK^VEtxYl&L1
zXRS~x_N-NE#hx|CSg~g<GFI$avj!{nY<X(Ko~=Y}*t1or4STjKwPDYeV{F*76&V}$
zY*~X1d-goFW6xfqcI??J)Q&xSmD;gq&oOrF*^7)Fd-klsjy*@7I<V&`Q3v)M73#pA
zqe>mvbL1EY_8djVfjvjo;J_Z2ryTaU66LVRRVar&u1Y!VaXE&=9#>>I>~UEGhdn+|
zdF=5e%43hOP#$}HmGao*a}19?zR2*{<Ff`HdqSQH*b_=rz@AW{0``O|6|g7d7y)}i
zkrA*bWDNrLoO$ZRp0h-q*mG8>6MN1obz;w%W1QG?78xh@oLPetd#*fnVb4{fF6_A~
z)P+4)mAbI!$}ukNxr&Snd#<d(g*|tky0PajQ8)J773#*GyGq^IbLSX0_S{9rjXihP
z;KrVqry};m5*4v0R;Y+Qu}Veki8)5Zo>*i=?1@=}h&?G!CG1HhDq&BmPzifdl}gx?
za*TvMsmMs!ld=X0d!9V?V9!&c9_)E4)Pp@wm3pw}$uS=6d5Vk&d!DSpgFSDada>s%
zQ7`tq73#&Fw@SU(^X3>Y_Pj;Li#>1F;KiOVPkq?)m8cJUz6$kW&sU{B?D=wx4|~2M
z<HMdWYw%&upQnE8`AgJ~J%5GzvFEQ+Klc1N#*aOJk?~{CpEdZgC+De*J-I|>?8y}>
zV^6M98GCY$k+CNi85w(W)*xd~$x{V;N{K4iQz}%!o>HX>_LLl>U{5JB3igz&LBU=i
zPXpKslxP5ZfeH;^FHofc>;-a60DFNV6Tn^|YY1RZ%~KV7YKf}YQ!7-(o?4|U_S77s
zVoxnHD)!W@LB(D$PlMPCmS_-r!3qsxFIc5P>;-d75PQKQ6U1IHYY1X5l&2x=g-SGp
zy-<aQuotS*5cWbjCWO6EkqKchlr@B~r{$@JJ*`AF>}eIMVNa`44SQOS(Xgi#84Y_{
z)}Uc8oTp*zg-bMyy>Nwwu@|n=F!sVZCXBsskqKijoHc~87s=BI_97)3!Cs_7BiM^n
zX#{(b923D_q{u|D7s(nT*o)?A6noJUjbblap;7Ebt2BzeXpV_uFIr@x*o$TjQS9k?
zs$)+tQ5}1Fh3eSTt5nCHo?~?E=|x7zo}M-6*o)<9412K>jbSfVp)u^msx*eZSdNKd
zFIHq?*o$Qiv4Ngv^^u0}zg@nU7<zh)J#*?mE{5)Te%<cekuz!SL-n@S_JI%8N4Je`
zuYX_=zV9GA^<Sk&YQMZS8BG_!k9bBi5Nl+CEJh2E)o2B>8ErszqaDa$bO3Qi4v07M
zK!Q;KavGgLE~5*`ZFB>PMiEFdN<bc?2gqym0{M(SAivQMBpYQQ#i#%Ui~%6kr~(Cz
zL7<Q^1f&@?ps+Cv6fs7CqQ)qYZq$Kd#@K0!xYl3Y#;&=8wDgR1kM-4mLtDQy)IKse
zb)rK3BZKMD5feZE&E!F~@18Mu#?bnA)lXd*-qe4}LjAjI_l-VcdIo=w+eggEY#(8f
zEq%m-Z0#dfWLqDxA=~?i9of-G9LQWB;gI=0!XpcPL_l`-5ht>%kGPQCeZ-9{_7M?T
z>Lbzs*H77~eY$q)d=0o-%EYt{ua+_~BeP|~B3sJDf^01lE3&OjY{>R9u_HUm#DUC}
z35U#=36Ct4iGb`Z6DP8(OkBwBGI1k|Wg;R=Wg;Pa%EW{0EfX)YuS|T%{xb0+%Vi=X
zD`lb}2g)RXtd@z294wO{a;Qu~$Xc0b$l)>xBS*?4f*dWAD6(EAI&!Q`V#o<TarzgW
z*J}PZOtYM^+S6@^+%dEMP-^MvlsjkE&&#<{{o(TUHtUR2kGqZ;T>WvE{nax!|LNlc
z&(<#P=<c|pzFmD(K1Y0O&6wK7BicrsQ~woe{h7RWcRY1|5ZoP8rt9DtJY@!AQ!J1r
zWdX9LtU$Ju4alCd136L-ATGrL@hKiiNC`mBloQC6asjzhZXhuw0!b+e$dmE_c~f2>
zU&;sMPx*o5lnkV#6reyV0Hmf=pkOKp6iS7Fw3G%EPKALYsR&Rs6$R2$I#4VX1DeCn
zIsIYpBVXix*?nNc+S+a3HMna?{bt?1^0`5=C7YDm#qC4d+v*<)|AC*jV354&i`vC)
zrZ%pA^4ts4ryfd-|KOF{#q$jFCe-h4^?!^u^`AU8a0kM-5B0lj?X}?4d1M3M<mn}z
zf#=4S9&a!4BKvxY582;K{K#@Ik&%^Nq96x)NdQ^xB`R{TmjsbRy(EOJ^%4y^+)Kj9
zkzNu(j`or$vffK{<XA6>4e-aGWG?y5@W=PnuF}r#&Yt?U^bgyhZ_gMy>h{{jGe*uB
zU%#g7C*0Lelsa{`-N2P-VokRC%}{%J6-@2{GP7oIX2A@OWm#~nz=E@|7H}590?x`>
z!C3_>I2&sNXA^AT?5rJ}U9f|5unuqz!2ynAIdGi7f#X>o953+T1Xchi2m&}K>jdW%
zoZwun3!F=EfpfENaBjg3PGm)Jq9B5kSP7gYNZ>rI2b@Rnfb+6ma9+U+&d2(|`2-(0
zKkEnQ7yRI4Rt6^vGB|}*z$t<PF2DxB1%v=Nl~uv1f(kCk2Ehe|Ah-}40v8fO;51eP
zrwJOkFdGIJ7Q)~nYy@0Hh=7Z-QE*Wq3QlKraJry_i?K0qF(C$S20!EUFHc)~;(Yux
zePP_bj!KW8dLuo`G|E|j1b;XELx13@GFQ8JoO_(F{+6(Q%4}y-|EaR@08g0a!$Ue@
z24WK|kR@ROvL>uRwuBAHp0EQs5)L3P!2$6J9!N+CK+c2{$dzybxf5<6F(Cp;2?@xP
z@Bn!eULaq>2joxqf#ie?q$CueKq3I7CRCtcA_x>pgn+b!1{6+&fg*_rP&5$*(i1vR
zED<|ROBXJlcj~m~WlMXOB~E>D`Kd$yAExKzqYrO>q_nK-j0JPAxI8g$e!~_2r)q!A
z9r&_;_aE&4xBm0x{DE&+Fz{v1k_C$w)lj=K!WafMo?#r^+;Zlyp+kbhW#9S3n}XgE
zqt0p_eYRViG<nL@i!ZT^G)d<=FBt#-UE0=u&e)F530>z+yznBk)n?}e*EoI*>+p|k
N3xpO@Mfa#A{J%sV1{eSU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.0E4AE68E13A6F5B4.idx b/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.0E4AE68E13A6F5B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd398e7c109e4dc7e3ea5617f56a9d9493716e92
GIT binary patch
literal 13202
zcmZvidzcgDp~m<7CM;oh<<5d2unQ<^GP%u6E`y-SWRjUok{J;K3rd7VYKp`jKt*CT
zF&HWku*B2gC6yu&y1G=Ywq9zj)=SkE1?vrqk60>&LQ7LrP^=Bx6ZZY%@H`It%lp3k
z*4;1(voli-h1PLRP10g@)w1hWE;KbYH4Ur(w_@e;R9hViZB0$99{fXM{#74|L%(>)
zr&o&R?|GLkS(UwF-p&E8dk1eCE6@Ar{@-ohG-c|cGhh9@cGnfxPi)!o&iTph-)ui{
zqkY~q=66k3|9ITgk!Jgwf1dQhw~tMAc^|m_v%_<yR+D>9AGzVSCnvA^cJYh{{$btn
zhaq3wVhXRFc4zfc+wOGflcUGZoct&}_O`EA?>X#wpS#`lf_FkG`!9X;iW%Q7Uq0=*
zUq5qbYI(}(2`5tG@oa12*_N*J*9`anZ0_;*{`$;gFWqqG;OMhGH~Xz^wF_6TFR){m
zjeI9@f9PQA7cbma-P>wEv*eM3Uv=cCEo_^fy6G<WK6CFr-<Zd*7_sk<nZxg7?z(-1
zwroBB%%&f2zNI`ZdH0j*{PCSzUr*gI-g^DcWu|#+-PeC-+T*u<`S;&ul$ZC%rVZb9
z<)Cp-9qatg>xUlx!^WMvIv&00-U~1Oz1+Q8Xj!^xX7J8Yk;%XM!ISorTbEvJHZpUb
zIkk20Kkn}S<EVWbhkk$Z!~a^C{oS<M&VA3VFh71YwsYRbx0Lo@kcE{4vvv=9`^bSw
zq4{eftKN7macbh~ftG_$8S{o;n!5D1z1vrwTyyDoI$*-`V#vSlsrJx2ua1B4>J7H^
zcaQz$N0&W(&y>Ik<=D}Y&2Qgz)5y2uiPjyz`18<?(fixzCi2ZXJ|%DG_nu&ej=S-7
zeZZ(4Yj^F~u<^C=V<U$z`SyD&&3pY~Pw&0QW~~3{%gy^g*m2X+)};rR)pkET{aXJa
z^ZJY4+AEC+=9srllkR_`Ywo-9Z9n*|G-}u*`;Avtz9g)>@X<vgdx`RV@{%Qd>!};_
ziJ60Z2flmcnyuSDXFuLt^`Ct9#5P-L-K(3q<CzES@9vr2(shxxeNyG6=Qdv1JnPiy
zGmG}mc<ac0GvEC(x^MHIvDOVG?at;e9viw`eOEoSVfV_*K7R8z*BP&Ee)OKm8xJnM
z>?e`0yWcZCbpLPmc0SB4YIV=J;COM@j+y`7c-6=Me*KdfnXb#nt=3L#-ogFjvRj0k
z58tzN+_tSvPi#xi&CP!0r~6mQ_r3Sjhi9IAWzGkSTIS4~bVu`v%ymy6`pNcnleE*P
z{`%oH?CE{A?>^luP40UC{D(hX{PF88|DG^@!xtX}<|mHMd*fK}r*FQv_sEX9pKjPQ
z`;>Fjz^Cqg+_Pxf@(=%bpY{9cBcJ^;w#@Os-qUB4oA-P*dCRNiIS(CP?Ed_W@AEYu
z?R#$L#XleS%lIuX{B!A#wtTaCXJ_fn*B5(UD5>8`xz2wtcF~9P$Gmk(Y{aVzPTjO*
zx%&RD?fX_ezwN-7CwCk;{ioThc3t_4@jLIjBR@NK;IR5XUknTHes;^3o!_1}YxVDc
zP+a=y{QXk?$Yj^APR|?oC)=&Xzf`_`;^k|He)IJ6)5Vo<?M=S;*^r~V^HZ0V*sjhC
zzcIWc951bUX#ei@$E(LXZ@=UCL;GGamrkZ8uXP=L<->3P-nwq<id&a0Dv_q9=7~>l
zlz!srXr*)JH+8o+w-4$&)Xy4yzXnaXG!ZvW+~3wTuO81DFzf2RLnChaYU{^&O}MI+
zChGC{=JCV&4vl!&&+Z<3)9$0cZKYS%<KfN22lX8q@sz>WPapN`Uma?tbL;V#<}pM1
z4vpA)H#_aeAN+LTc}=tGaYuVcXWyX_U;gz!kMH{I)}OV~u6jJKdECIhLnB^t$rUde
z>;BYyUeoM)+}YkarSH&)uOI4oev|3tW34n^k0&%w7}j@a#4Gs!y}0?s1)f%Vbv>Tk
zJh`jy(1>@e8UMx;rz+34(wX&mWb?=&eTPPTbYaJ{EPtG9rA9p--8_0w-=PtIGc@tU
zfg4}`td(9>k4H3*7}R%Y#7)buTe9L)YybCVNlV}Nr1;X;TOU0*t)s1b%mrg6Hokau
z>KrqrpDe7!(fI7!x-Ejm3#3@IezLMwXj%m;G_8sink>sglNDHKvI+}L8*786O|U`J
zrr4lqXYJ6m3wCJQ6+1K?tOJ@3!2wN&;((@;bwblAIHBoOoX~W!E@-+07c^ap3z{6u
zL6Z|WXmSb%O`heU$qPI*d4-3jzzWb51Ob|YB0$s4x}oV7+|YC@ZfJT~4>Uc32bvzm
z15GdMg{D{VLer~wq3L6N(DVsDX!;Z%G(}c~rYMNe6crJge%237zu<?aU-3gTzy_ch
z5CYH)C;@0ntOQL-kf13k5;TKs5Sl?D2+g1pgr>~O(3AxknzAB8GsK3V84^O!3@IUK
zhS@MQ!$KIEVI>Srg;k)b2nsY6MS-Tus?byg6`HD|LNmffpcxS&(2OV%XlkqmO-<0C
zsVN#XqihtKQ6UP=s1k){jEzAvCd8l_Q)1AJvvFv~g*Y_hN*tOxt3y*4bZF{|4o!nK
zplJvOG!4ao=4^hp=iIAacY;p<&E@Bwqeh3c^t~cBPR@;2M5(3km9X(}{Y;!Cmi46n
z>}x3!%NAg3mRK>j7Ks&eHcME{*&<;vw`GY9b6b(vFt=xk9dmn;*fDoxi34*-kvK4S
zW{DGXXOTEDcV&qSb61hLFz2#_!<;J;4s$+Bc+B}C;V~DoM8I4q5&?5}mbfu@7l|8l
zPnLKv_Y{c-b8nV-G4~dU7js{h_%Qbsi4SuzOGM1YA`vn7XNezkf06hx4`fLI^FWaV
zFqg7K!dxm63G-l<1Tha5Nf2{6OJvODB9So<Wl0G0P?3Z%4`)dj^Kg-bF;}ug!CWa4
z1#>k^RLs>PQ8ABXNd)sqkwh@pvP8pND-sR!XqH4Vj}}Q3^H`R|Fpm{U4D)!F#4(Q-
zNgQ)MOLWZjBGEB7vc$mLC=vtnYxrv}Z=5sQx)<;Z&QarCa-n`IIr|>D|C}-Mf{_y&
zZ#CPxM|O^!(oYuMG84`w2}>7H%5nox#&UZ<S#>MytO+aZtSKw(tQjlpSe=C(n_ywb
zrdZgq85VXn-3B{b!Uj8A$_6`I#s)jPZik&cVTYYPWrv+TV~3qXcfih(aKO%ya=^}!
zalp>0J7MQcIAP~ZIbr9_IAQ0~U9fW{T(EPcT(EOxT(IMG4t89EgB_RRV8>-R*zr0K
zJ3hg~j!*Hh<1;+$1YLlgkPu)eqy*Rr83A@~-3>c;!VNoj$_+br#tl1<?tz^r;enke
z<$;|i<AI%5_rlJb@WRfU^1{xW@xsoh`(Wow_+aNt`C#YE_+TgMBJ9M32s<$)!cNSH
zu=DGF*!dHF*!fd_*!eSl*ah?e>;j1Z>;kC(>;jnp>?B=+os^JZC#59VNf`-tK|Kh&
zU?K>+U@8c^U?vDVS(jlaCuG>kDH(QhMuuHT55X>!2*ECt3c)Uv3BfL`hhY~^gkcv>
zg<%)Ygkh)X3hb1G0y`z8z)s01uv2vvc4|U}otjc%r)E^xMf3>lB8dp>BB==MBAE#6
zG+l$8me63Qr8L-S84Y$(Jqo*MA_}``Dhj)3CJMWl9)n#h5rbVU6@y(Y6N6n`kHapW
zh{G<Pio-6RiNj9Ub=c_%9d>$3hn=3$VQ1(D?2Lo~J0oSl&d3<B>*Bl4y%%rmp2yDv
zx|+ZG95t?tOls-7xYYmRY7s3Gyex`VAXa37Y@!XwF4}<{q65e&I)PlG3y2drAYSBw
z1W^ETi*6u~=mGMIULc?70}@3M$S?YV0%8D25+$IZ7zC0<87L%%fWl%JND&nvRaAi@
zVgyJNHK3>%1&WC=ptu+Z(nTG}5DlO?{G92HH&l&Fg~M9<E)l{vxtCa;#(!n?5-T#>
zOIT!EFR>xpdx;&{(Muf2&R*g~cJ&e$GS^EuWWJa1$U-j>klnq+jqK?q9%OGX@gn<r
zi4R%qC1O7}E|DIrzq34hffT+j1!`GK`=9=nGPP_$u2Rb`<Qld73pvkN)-bRyF_wFg
zb0*83rvBVopjPx*%hZZKYn58jXRT2y`mA}ziau+Jv7*nKGg;Bc7AT88woF;{u~o{V
zkF8M_eQcg#(Z`k;7JY2a#G=nupf>c`%G8EFTb0_-XRA>g`fPc|hCW+~v7yhFGuhB*
zFHk%B>}6_4pS?=$=(E?T9ewsZV@IF8#Msei&zbD#a}=lpeU379pwCgI4)i%{)PX)n
zo^hbhQDPkEbL318^f?REi9TnUI??B>QYZSHHR?p4GtW5D=PWT!^f_}TC;D6k>O!BZ
zOkL=6RjCVot{Qcr&y{Cf=yR1A7y4W|lM8)ZfpX~M%9KMNSEU^KxEkfq$K@FgeO!s*
z(8uLW9QybI<<ZBNDUUwBN_q70HOixp&oeyw_!7gTkI$KS^a%wjpid}M0ewQ13g{DR
zR6w7QX9V;KB}PD>kTVJBa~G%^eeN=Kqt9KXZuGfp)QvuOo^hkkU1HqmbLUKM^mz)@
zgFa80deG;oQV;q(HR?g1C(n4$=P5BB^m%e75Bj_X>P4ToOugvyR;d?#-Wv6y&zonw
z=<}8sFZ#SWlNWuy0`;NKSEfGn`Kr{1K3|Rc(C5oDKJ@uYj1PUjoXLkiu|P%iiDfFH
zPpncAePWG@=o9mdh(58zi0Bh@CJ}x90`;TMU#5Qa`K#2AK7Wn+(dW-Ie)RcEj30ge
zoXL;AK!FC(7bw#J`T|uNKwqFn1LzCnnE?6%B_@ErK+Y6EpH!d{`lK?I&?i-?gg&W8
zCG<&oMna!dVkGoQIg^CGV1WkF7cA2t`hry&L|?E*gXjz9nIQUtB_@czV9pdopIo3a
z`s6Z|(I;1_j6S(WW%S8;Mn<1pVr2BmIg^aOP=SWf7b?>b`a)G2LSLvxL+A_TnGpIy
zB_@QvP|g%WU${WS=nI!=7=7U?4Wlnyqha)g^Gq0h;Sv)@UpQw9qfaSN1$|1HD(F+H
zR6(CoqYC<zJfombDKQHAl$=RHpIV?Q`qVO2(Wh3aiaxbQRrINOMn#`mVpR00Ig^UM
zNP$Ms7b(*S`XW^tL0_asBj}6dnF#tKB_@KtNX`^NpH`q6`m{3D(5F?YhCZ!EHS}qD
zMnj)gVl?z=Ig^IIXn{u27cJ8$`l3}DMPIZ=qv(s~nJD_AB_@i#XwDQxU#vi5=!=zU
z41KXGjiE1AqcQZw@=OeUu@VzQUo2;ep)X#barDK@G>*P_mB!H*uhBUA;&~>HzIcg=
zqc5H_#nGo1sE$6pOm+0>RjQ*;uTdR+dY;kIr<WKVeR|HMqt7T%1ARuB8t5~s)Igt6
zqXznnJY%5GC@}{5jGW2n_xaZzYdU>$<$hx7?lyPNZ~QAda^DM^_T-P9N9#XpXdgMQ
z|7Q(j+Q*D*Tt|sMaEP7pdHIq0Z73GA<zo0Jp4ke-npq&5*#=}c+kqTr2awb31ag^O
zK%AKa;>|peU>1PfW;c+>>;dwcy+A&*4@fkNKz_3yC}0i%NoENsXbu9&W*I1C4grPD
zVIak<0I6mbC}NHPX=V*5YK{WM%rT(2IS!<obs)oRoTHd$^NsE7hI>d`_jvDksqusB
zS@Gdh|N6tr2Gk#(J!tlj#&e@_J5~J(%h`LW`mZGTdx*dPibGp>poavIr5++72YW~m
zS?(b+a;S%dki$JBjI8t!1zGJODsrTUM3A)}q9I3nNEA8NLt@DB9uh~^dx(x~^bn(;
z8@G|wuP&awhphk0VyO_zF1+wmAy#CzLRe&5h1iho6=FwrREPuFSs_kjSB1EcxeDQs
z`3m8Yg$fan-4)_S_Ed-m*;^rAWM75&ki`lSk^L3oM-Eg-09mRK2{~9HL1ei?WaLnV
zgpk7(5=K@kL_t<7L`9BNNCa7{5Dht6AyMR5g~X8K6%t3*D?~>&D#Soe^2u{Q;BA~E
zX-nTtg#AwvJo(<N1%Gdu>o)2q-wwHBPUDJZTX(12Ij8Zr?+)dw1)ClA*=PT9cFg7)
ze>rpCe|pV}pB#Lq{%}WE$F+^^+7t2xzPC1vtv@`xefR~9pEEbk>-FdC**h2eKOIZj
za!ccLZR@tCtw3y=1+t}WK=!m9$dPscInz!cSK0-{r8yuz%>xN(0mz+p19{ROAaB|W
z<V*X2#Iy+HPy2xa=>U+FmVkokAdsAvfkNpJP&ge1Qql^LnpS}#=?IXP)_|hvC{QdN
z1B$2PKzdpSGSUXne1879E8>s*Eq`m*!3Q_izxPuHO&Q$ywm!P*xdFzqZK3+Z;|7mw
zZ(Jb$H9vFl0D0Tr>JPVD+PTI}!7t64aU?nM{nzRbFElNj)Ofmm^5^UoKOFYhNA-t0
zyE?lYPtLzM2Amu?Wc2#_!?Q=tp4ezNZicO2gFbsFY~$(nH}j*LsQB2~m!Hv=(e6g;
z%UK@;!%tPX`oj~v6QsuZwQ<XAXG`C$vW-8a{vIT0S=9KSt3PLxRv<RX0@;!_AbZjd
z<VZSzoJl8;E9nB_k{l49<bi~w0OU@(fjmhMkT>ZC@+EyhVp0V1C;dQyWB^D?N<hJ6
z5J*nSK%ry^D4YxfDM<xLO{zeVWCTb{YCzFs6eyOA0mYMXAU&xA8A;<DEx+!Dg=cSj
zUa`D;Me^(i3(g+;elguUwm9Q0&)xaUfs3!Xc0qFCqNZ#AFRK4`P5*~|yZ>U}pY`7l
z7xfnw_kY;EZ1D}(*HK#%V@v~@2bu@9ww*V0$Y6DtEL}9LMGXueJ!0gT^S!=lmrb8>
z`4zTN7ICus;)(y?qwV7^7~j!3Y08CDFS*oecQ`r0Gl3t=x`LzH!;$M~NcSrw`X9(t
Bg4h56

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.3B87280201C19511.idx b/.cache/clangd/index/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.3B87280201C19511.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a10692619d00b03f19c3f6a8e4aa9d98a6ae4de4
GIT binary patch
literal 19820
zcmZvk36vY<fo}W%OBmU0He-xq;|)yMy0n)hWH8!QT8IcprClXxASx}TmCa(QMnoil
zAY)~s5K!U(Mnu6Z8DIj0Fd@k#kN^QDguuWH4-bSbFoI!c^5mW;`g`X*pF`1C*B|%L
z-PP4<NqxHC>plC_nHksF{tL_JG|yDc%*-4z`KQsWO9xD#vvOwU!W$nKQfItiJNSbi
zoERQ-)m3Zx+zS<Z^49$f-#+lVb)Mu4-}~*(>$V=V_p?v_Y5aA2e$$HG&wNR|<?~x!
ztZ0+lnBUH%zrX&N)oShUUODpNzu$F?!Ft1`uf7&LW+*=K@$03F?>Xwizn{JRhIi(^
z`}jd0T%ht@x$W}d@tQm2-e2B$>$CS?@U6S}(;ZK|X8DuxQp3a6jXmYP@ZpW^f3MfK
zJ^1tc_a4*V`tinhCEMG|tnk2!#FwsE<+wBc_OpMz|E@<$mme5<V06A?Zsqvs9aneQ
zb>-F12;cL*Jo~}J7Z0DB)qZyFk6->|b9>vFD^Hfr`?~pQ^`56sTzj|u(5L^c_1ZJ7
zuU~p-uza=Y{_DQK^Md|1@tgPfQyaGIdQ>WHn9JW<Rwb`A=fAe??u$SC+b>)Eubw}3
z+p62Xx_te;Z*BS7qkC_D{MuV@+kE?Z-#+@--+ShFa4U+}o$kK;u)tA2`^G)m_jeVK
zRiD<1-T%?91K;`P{J$Oc^tA_Ha@5W5ovHkG+xXU}A8e@ae&f_zlh;1Uuepz$Ie0O0
z$MQeC{^F6|)D?jXAN!5)(WV{ypYZa%rzKY%FCBmJQ@1qVzvB1}bia-DuGewZy=%PB
z{Og7r)7NO^!dw6Gos(|<)>h}c{9A9VUiOEtpSSuCUlC?^-}lPFn-70}CB2S(ew9h`
z++up_UFP8Rl}E$-9k%<*+jd`b?Qb@$3%qvR->XgaQx4l`&$r*&e)S6<?tK2a-RBi&
zi!YbQcien(&aqd0^_QP~%5|uFf%${#$nQOti2u=Z@i$)W9d^i%pFi!_%}2Paj=udY
z8+#o8ka*m=rrD1wZQ=CgC%#yCJ+tek|6*U-Idr`Lz`HkTdRP7HPUG9H8?=9X;^Y;H
zF9&Zqa`4E5*M4<b<fD&2JL~!FPriQr>3{q%^z_ar*3Dhh3tqnLgS!r{`~T?Qd(9oq
zlV1AWznpX0Z+71Pt-xb97Ek(N;M4hMRX_ONzdW_&X5(42=Iuwk-Mwx1>HmH0887|!
zqkq}nN}RHONATU9yN&OhbOCq%Yu_raziHRZPi~Uq3(;Ty==lph*FSsjpFg|j*Rki$
zS`kYgdD*geTj%_A?+<Ue>d4^7AN}>u8TS6n_^}WEl*?~@(>5Qu@%P&jzxZD_x1ZXW
z%Kt8q30^-q_Pi6Vwue>+Z+z<h6{o!VpS!+e_`qB8X>PFfc0b9!_}P6wI^bNR?FXA4
zeIjw0>Y0zmZ@hQo?{EG7DZ}qqQ#<~&`keD$dG__cSuTBi)jx?(n$E7Dc>P<iUSN2A
z?$j^;@_|RUwq{>BakuoybF&XW5z@TA?jIxT?u^~DzJ1?4+kWuiii5bv?)&9E?^f;_
z{piz^j(cMJ|Nd?C@b|XweRv^o-IwmJoqBPwmDgS3?pVY(l^2e_^X{L$Cw^z`%9}R*
z^FRBp{M(72obcRdju+2+_m=Wcjy~a*TQ;ru8uhMy=ZqK5yZiHdPX5(9cYc1x18dLR
zxxKviH19*!PdEK--Br708W)w%>XDflB7d~)!@vIf0V`*cshRnU_PZ#(c$j?MSp2hW
z(@(E;{m`;`mc}M}&9XJi7Y|dt;*(u3wS(Llvs9Sq4a+thvUr&4L+<?My7TUM<CnAa
zs}sFy*{bD>hpFCrVE*L8e*Uw)vot=@YnQD(Xz?)Br+oUKZ{POnMR(59#6+)OwtoM`
z!&IN?yY^xG3$D;CO;7Yu%Z^Gc9;P~fu<oJjRKI#_mVRZTH!j<F$l_tDFTZvE*_Sr_
z;*bMoq7%Jk&6cf;hpFCu#fHaz^3mX-S$g_JuU@wLpvA*f|IR$~XY=`Iv<{ewO!Ve8
zo3|_;ruvODH$T9d-ZswC(<b`xWrr_cJWTcH2Ma%WvGS`|XXzOeedw}7moFZsdZvEP
zxs3<BOCKAuV)0|redN<!x4*n?^Ymj<Ua|PGnBKf{e(e!!H%;HC6WX$N>k`eebGqq!
zY~}nMKNnmgmSw@QJPS_4YQSlD4LB{U1*he;;B>4GoQ~Il)3bVTdR`CCz#70Acmp^i
zYXoQHjo?hI37m;Hf#X;X9LICu%&ZxlnKy&8uoiF@-U80bTESU)D>yrA2WRK);2f+2
zoP&3ObFxlwPTmR5#k#<`co#S~>jvlM-QYZ|2b_oZfb+6ma9-XE&d2(|`FI~Vp5?*u
zJP*##`oZ~mKezxJ02kl`;DT%rT#yff3$Y<^AwC2y!bZSF_z1Wt8wD5Tqu>Nq04MMQ
zxELD)7vp2#;%ppToR5P`unBMpJ^?PtCc!27B)Alt0+-@b;L>awT$)dV%di=689oCp
z%Vxo4`7Ah*6~T$T2rkFwz~%THxICK&m*?}~Bvt|^@e;TKTL4$!3*d@u5nPckf-A8l
za3#J3uFRIfmH9F_nU%rGybP|wR=`#G3b-m;1y|*(;A(6QT#c`RtFv`*b-oU+!8X7(
z_y)Kp+XUC-o8Wesc38f+HqT#Ty5tL*UUQ2p7O%BSU;gJ5GIupu`m#RPC3D{eW);Gs
zXS;+&uTh8wy{1bv=(P&bqStnb7QId(I`q0O(V^EXM2}wIC3^G*g&5Eqy2OCqs1PH1
zW0x4wn-pR~Z|V{gdQKr6dag@2^k#*a(VM%(jNYOU3wleJSkPM)VnuK55-WPULhR`6
zU1CS?P>2J)qe~p<oeFWHcXo*ry-Oi3^sX*(p?539jo#fQZuB07c+h*g#Dm_e5HEUf
zmw3_p6yihg>k=P&ULibszDs!Yeuena`@6)CKA?~Q`aqWi&<7O~L?7&uAo`F(Lg+(X
z5<(wQNCbVPOCso_3W=hRc1aYypb!DQ&?N%;m_lOcV_gzMA6G~meY{KJ=o1P_pigv3
z0)0{;N%YAsNup0FB!xcJB`Nf2g{0A^yCjW1qmT^xOqXQPXBCn~pY4(?dQl-Fda+AH
z^f`s((C4}&hd!^6Jo<c><k3qCk<d$BBB3uRq=3HAB?a_Fg%r^jyQGM|q>vK&QkRs_
zmlaY*U+$7JdRZYddbvwv^c96v&{w*og1)MdD*9@dRMFQIQbS+sk{bHDLh9)2T~bHi
zP)GxPqe~j-n+j>7Z+1x&{pG}TIXPwNqSYp*_7YF;3qCu!H?!}a;L>$#^%1K#P4925
zoL{|V_0}bt3(uVnS2STR0VK_pfLe2xE)g4MVaEzA>{y9~9ou4IrwMCdrx7%;(?}ZF
zX<8cCX~SCBX$39pw2~Hf+Ljh}y08v*Izb0Jouq@EuBC&WKCFkGUeLo%FX>^YZ|PxY
z2peE$5Dc(0NCwy$S_aq|!$#N{1taW?k`Z>smJxQQunBf1!2~;#WP+WkWr7_S=3vJO
z9PBuWgB{o6U}p}SVP_W1uro_$*qK{q*jd6B*jWS%>@1Q6c9xa}cGj>Jc2>a(JF8@c
zowa3!ojq)aon5fQ&Mw(uXK&eI=LkDs=MWsQb4U)@Ia&_bIm1raIRz)|oRSlE&XyB)
zuCNPsF2Mymm*j$-tL1{7JM4y?TX4h9ExBRmZn<IS3437Y5j?Q-NFLaES{~SW!(P~V
z1uyKpk{5Q~mKS!uun%@V!3R5^<b$2B<%1m`=3&PRJnVRhhaKPIVdoF~Vdoe8u=7iP
z*!f$2*agA?*ad_D>;h5%c7av^cENBEc0nNsyPy<=U9c5|T__xaT}TMQE+mCu7ixuI
z7YRpT7ZD<`i%1dJMOqQqMZ;0pMTIErqEZxg(N+|8LRf&EAPBG%Bms6pOMqQ09D`j<
zh`}x<#b6g}#b6f?$6*&2;;@TLaoELMao8oo3D_ls1nd%00(Oa30(Qx85_U-;3A?0}
zgk7?ggk36}f?Y~T!7e4GV3%s8V3!W3VV4%tuuDs6*ri)(*k!^Q*kyza>@rdYc9~WN
zcG++ic3B|{yR4LjUAC2lofsBjCki6$L`j65*b-rv3+G^$6LPT2Njcc%S~=L|!+F@{
zg*@!?QXY2sRvvazSc07-NU)P633gITf?Xk8fL%c-z^)(_U{`1rU{?$mVOJE2uq#SM
z*cDqv*p<R1*p-A5>`GDzcBNJccI9vxc4eUqyRuY<UAa|;og9{7Ckry{WJ!jd+>&8e
z30GiO5h}2&NEO&sS{2w;!&TT-g(~c-QWbX9Ruy)&a1C}fp$5B}RD)fuRfAnUT!&p<
zsKc%<)nQj})nV5NH(=Kg8nA0f4cIkW4cIlqP1rSsChVG06L!s36Ly!HF8$&MODpFu
zGhGI>Kuind3!XldU|zBKAkEUZ?{l^}7kuMwV}Ud_4UpER1=87cKzf@V$Y3)78Er-&
zlg$Lg**GAx%?xC*S%9oIE0Eo02XfdPKu((z$YpZ@xovJBkIe(*wRwSjHXjgg<AMA(
zKTyCH01DcIKp|TQC}N8MMQu?a!6pF3Y%!p?Ee@2hC4iE)Bv8th0!rJ`Kp9&GC~M0C
zi8c`^XUhTQZFwNcCIJ;}1)!p>2vo9_fXcQqkZhBIDz*wx)m8<n*=j&_TOFujYXCKE
zO`wZS7oR-6Gdq1K@@=|e@i9pF?C;RIE9uhb|6HHWeHVF1=WatD)4BJM+sxb*4D5T%
z+_#YzRCAZBmOR^`EcVzwWwFN&DT_UJOj+!)ZHC1j+hbVlu?s2|dzuc_U{BMh8tiF?
zRD(Uum};=6X)_w^X?lzXdzuB627B5L)nZTEr&{c3hg6F_?U-t@r)@J@>}h+97JJ$S
zl@@!t4%J~#*QYw{>4sE?J>8h<u%~M?I_&9sj1GIc1(gna`VQ4&Pv56{?CFP8k3IdE
z>anM9GkWaldyF1?`URC9dxj1*V9(H}2J9Jz)POz1m>RHWXfp=v8G4KXdxiy-0ei*{
zHDb@$r$+1<ht!BY<Cq$;XKXV@>=}EE5qri3l@WWU4mDxV)TbuwnTFJaJ=2()uxDyB
zChVDdj0t<D1(gYVT!(Vl<NB1t9yg>M_P8<Su*bC-4trdW;jqUos5tDIJJgIlbDx^A
zXC6{B_RM2y#-6#&n6YQ>F=p(U7gT2KSvu5$JxiZjuxA-k3-&ByYQdhR%~-H!=`j}U
zSr$|l>{&b1ial$eTCryxQY-eXV`{~owar+uXYDao>{%C7R_xh3)Q&xSpW3l!A5uH^
z>|<)jp1sZ3v1jiwcI?>~RCeq+I@EzZN1r;d=NM84_8epCz@DScII!pFF%IlG7E})G
zIXl#eJ!hXfvF998C-$6U>cpP2%{Z~=>@iO4ITut;?72GBg*{iFy0GUOQWy4IW9q`5
ztIfEu=jt&o?70?HF6_BG)QvrNpSrQ<9#S{<++*s-p1aMsvFGkFZtS@iRBr5fI@E(b
zPoH|Q=NVED_B><i!Jenhc(CW`F&^xB7E~VWc{|jLJ#U|SvF9C9FZR4+>cyV7&3Lir
z?J-{Lc^6b(?D;y>hdp1P`mpC4QXlqwW9q}6ug&<d=j$;(?D-Z{KJ4)w%43i3QyzQ#
zkn-5$$CSq&-)4C1@jZse9>1XCvFGnlKlc26>c^gcNd4IJkEtJf{x;*sp1;TVvFBe<
z`LP%1&;a%VeHy@CU`PYl3yf(1dx16+z+Rxo1h5xaPzA6T?9d?gf_)mqUT{c**b9zn
z5PQKk6U1Jy#{{t#Tu=qE7wXUu_CkFc!d_@dL)Z(AX$X6vHWR{LsK<n`7g|t-uovmj
z2=*d<8o^#<NF&&bjA;aWkv0>-UZlrFuoqEOkp=kvI@+O8>_z)DioNKNMzI$i(<t_$
zZ6=DnXpf0vFRG}b*b_Qbz@E^j0``O<6|g6asenD9%?Q{NdW?WQK~V|Vi*;xWd$B%^
zVJ|kMG3>?0G={xcn~7mB)?;GWiz%uY_Tn8H$6mZo<JgN2X&ig;F^yv{-e%(1i}#o~
z_Tq{vj=e;OCa{<2(**VsLz=)|VoVd*OSG8;_7XiNfxU#HN?<S9p-Jo|`!tEY<d7z@
zmmJe1_L6NTiM?cxNn$UlsFK)Ab!ZBEsXk3%FEyko?4`yug}qdpNntP5V^Y{lDXJ9q
z(jA({Ub;`y*h>#-8hhz6O=B<JX42S8_n0*H(uyjLy-bH@u$Sr64E8cZn!#RXOf%Ta
zw3!U{GCd}Py^NyDU@zODS?p!|G>g6LkY=%$9n&oKvTY`dy=;%kVlS(xve*+lRK%Xx
zry};mAr-MFj;V+}vCW9s6MKw^JyB7K*voZj4tu#i&0#M$q&e*6#x#e$T${;ZFV|yo
z*vl!Z9QN`Zn#W$gPxIKz4{08I`7zC7FW+YJ*vt2rJofU6Dvv#>LnZ7<eJWv38d3>+
z(wIuvliG}gJ*me?*pn2MguOzC7O+?7(*pJiLt4OIVN46yE3}yc_6j|wfW3mEDqyeJ
zp+)Q!`?QF?;*b`xR~*wK_KIz$h`nNuDPpgvsEXJtb!Z8Dr9LfTuQa43?3Ko}guPOm
zDPgbFV@lX7DXJ3o${kw9Ub#=p*eefd8GGe1En~0TX3E$r_n0#F%8IItJ-I_=?8$v9
zV^1DZ8GG`W%Gi_JjEp_G$H>@|6_t#=N{3dkSLxFV_9{bK!CqxdE7+^FnF{tQJ*I-a
zilVAuuiBwi>{a`;ioNQPR<Tzd(<=6=ZKjI7YLBU6ud1l3*sFDD4STgdtzoY=q&4i-
z#<YgLTAQh1uhwH~*sCe38uscPTE||!PwUvL4{05H^)anuuij?r*sJ%LI`-;{s*b%z
zhc>X+=+g%F8bjK^USmuf*lV<z2KE{~rh&bNqH18T*`ZDBHT$%Qz2=ZMvDX~aCia?b
zrir~~k7;7Bsi>MuJ;#-A&3ycR^Le70{{~TigQTWE@D5!6@O4kL-#UO!o{+U>_4=hJ
zWUXDZcK!6BpwJC_+3o+;|FLQEf)e%IvG7A`H4CIsYk;(BEs##F1JbMYKnAq|$fz~~
znbamAPR#+C)n*`z+5%)%TY>CqJCH-|0CK9GKrXcl$gOq*dDI>tui6XbQ~Q8;H4o%h
z`+)-L08mgJ1PZA`KoNBWD5{PE32Ff-rj7x{)p4MNIsueaCxKGx6i`~72Fj>2Kv{Jb
zNK}hJIdu*wug(KWY6+;IE&vtPMWB+p1XNa+fn>D|R8d!es_H6GO<e=3tLs1wbpxoW
zZhk?hT6R{}u-ANxteoFq-Qb!&^FH#GRmc4GA5PkDa`TSmI}VzjGt<xi$zw$OUX8Z&
z7}5WGYE565HhCa*->cG={@UDt%-x31*B=lTSu-FSWbJ@xk#z&2L)H(79@#J;24v%a
z7?Di_VnXHyghMtDh#A>3AQoinfLM|317b&Z42T2SIUr7C*MPW?-2>uA_6&#z**hR!
zWZ!`Jkof`Ok^KYWM-B`~0691yLFCYYgpeZx5<!j*NEBHZ5CJ(hATi|lfW(m#1Cl^a
z4oDI?H6SVE^nj$1GXs)A&JIWxSsV}%IX56V<otl-k);8VkP8D+KrRkQ5xF!VCFJse
zl#%5Dk&!C{QbDc`NENv@AT{LrfYgy21JXcl4oDOE3Szoq>G{Uz2i?CvA#Ofxas}J$
z*&Lg`VsqvE7SER0^wpo2@t<UO>a@{)XY%G9#_5^7aLG@v_{CpdzJGG_=EUaQbUR%g
zyJGRRqDyD4eMIa(g)`SNA`WEdh&Yj5BjQ4KkBA%DGa?>j?}&JjeIw#S=0}7__K%1k
zIWQst<lu+|kwYUALXM0`1UWh)QDk961mxI=#E|175=TyqNCG)IB1z=bh@_CyBa%kW
zj7SDKJ0e+RaYRJq+=%3m^COZ+mPSNEE{sS4xi}(4<kE<gkjo=dMwUlJMy`xV1-UvR
zRpi=;)R5~VQb%r#NCUY!BF!bf@?FmnCq8-2y2)qfsx_;Qm_D(7`aRm@y0Y)}%1fV}
zbMo8;(`(bpc~)kDG_nRrD{FytvJOZu>wyfi0mvvDflRUqh?6-Wvup;k$QB^0Yz4B*
zb|8oB0CLJsAeZa{a?5TYkL&^R%3dI!>;vLu9>_2IfdX;>C@2SkLUITwB1eFtaui6A
z1)!K51B%OWpoE+NO3F!~l$-)e%W0sDoB_(pSs+msfpT&VC@<%MBv}F~$OWLHTm&k~
zC7`lg29jkNs3KQ@s&W;mCf9)Kavi84H-MUQ6KL{sKZShp6rCTx*}f?8@{QL{&Wf$e
zw;nkCgung52lqRze3N%_^ZEnVubDoP=jW!=&)&~-)0>l<*PO7%IDJX|Ba!W|i<|!R
zo5{@`qUw+%r&omOCE&mni!bztlWPx|TRJr-r`I0BB5U>#4YGC*(IV^i5FN69578qV
z_7DTIaSt&foAwYBGPj3t$mTu7jBMFMEXdY9#Ja?Pc}21R!y$LQF!}%AlGrjoz2N<a
zZol{UKj`qQCpYgnY{#bQJ7xN74kll5_Wh0nc&Bi4*69ae@-E?6APuJh(sEiL9j61*
zb9x{HX8<yCMj#Vs0^&Fh$jq66ESv?%%2|QzoE^x)Ie?s;6UfE6fZUuL$isPnyqp)v
z$N7MGjtBB{exLvs019$Jpb!@Vif|F2C>I41H~}ce#em{m94NshfRbDiD8;3K(p(xS
z!)1W7Toy>=M4%j(1IlxGAc>QJ3S0rG$Q6M~TnVVmm4ReV2C8rspek1ds&O@-I#&m3
za1Ed)*92;rT3<XL_D%Ke*HPPD`|cYae!}7A>ACP><T<zR-hpv)^G54N*YwWb^jCgx
zD;9s%2b?P6+*xp{h%Ass)BtHkEs##s0qI3OkU=y68AT(INi+d*A_rs^%|I5>0%R4f
zKz7j%<PaS|PSFYE5?w%U(GBDgJwRU33*-}hK)lEU`9(iaKnwr{#UM~f3;{*N2vAgv
z0tuo36cb}WaWM{*5EDR2F$t6sQ$T4k4U`cxKv^*hB#I(XPRs%2#XOKCN<am%08|u<
zKqavRR2IuXvM2*p#0pSVtOC`<8c<!V12x13P*ZGvLG^P=XYTuTHjVmxL)`a9cHd#~
z4>KROj}5lR?^N%9b|#k<&pc};_kY#oUzw#h7kB@Ii~p_vdGoBLzi{@_oAc$fOZf=~
z&pnk<?YC@y_43)32ONCRf&N20t}h?5!tY#l_@S%Ue#v^`wv$fYeu`al*a^0y%*Sr}
s|J}M~{Sg~BZ#i=7(Z?Kj{G3*&H*%JZrgf~reb^db;2i1=J9rZMKdb8#XaE2J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp.E948F8CEDD1EC4BF.idx b/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp.E948F8CEDD1EC4BF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15dea6fb969c60b54755de5d4798200848cb0589
GIT binary patch
literal 11192
zcmZvi3v?9qy~p#PnZ%F;fdJw1Fp1?MkDd2>h6FHx3y#PjOL#<HySx`cz~vFpDzs1^
zC~gH6j0hDQj1VzPDS9o51fvy2W55+Libb;!i6XZxdaYg3GuiJw=jI&d<g?%3em`Lz
zJ3Bi&Y24Vc)x*-#N<!n-ES<M%4wIIa)-CyO<*F5nm?XTBla{t-*U2f<mYp@ap8jRX
ziHG%gc*(IYzw7>|Q9F<O@A})e-S2E0{q(P5!&}!puwk=x{`jH-?SJyJTM8E#a`Uc9
z8BdT;_MF;&`xg@?Hy&E^zsrxyN1om=e}}%{n^|1^+WPfJZ(H-r+!3={Z+e>F`pYN%
zcU81@o4NnU(Ldk0Y3Cz<{b5<bqVK=GG;qde=iZOU&o8zHQNiaA<b}V#(0c#<<*QG=
zf8cjl`k#0yUKK36@YE~f!QneTd-=Z4Kk00p=zM0z!S!7D+|&<t^oeIZH|jZl??+PU
z*p|WHcfPW0$H$8kN6ubWKFeS5-LcrdTb?`cAv67jZTr(s?)c+|Ppt)auRHkR(vlaQ
zXSBy_-7EQ7mljuS`n^krUwr8FiMl5v?=rhrE_~zAzW$G_ZYzFcGgH;`)4d}%KDl%E
z=noG4)+z3OxMobJ`}aQd$B}bu_g-%Dhx=vaZ(-eqm)8`U&GOPe6ps8Vy!5T8wSK^(
zqniA=#YYA|bl?56x2^u$9k2ZTA77N5+v&vm&3w>z2XjwV{Is^N<?3Z)dhboo+V8%-
z^`Bj=>?N0b>=@s*^wjp(rG2Ze{(A9+qV)W%-%P*t;n?}<+rG$uzAQ%z_iau5cKF^W
z28{gf(eTb5853W8wa2Ngr<V^|y`=4hFUEIaHs_yvRv9vGcaPfTbpO!{8t&|*zPNe!
zi;r%7`=&Fcr`>;jd6jwSHy`h3pK5*c;j@oF+2`cDpFH+p=hcJOo~g}WQ_(B)uTy)p
z6t0_j{@9j^U+-J-_1H+af<4#Hk3M$v!^z@J_ikLDe(8y$1Jf)2F|7X4Nd7Nd7cx&h
z5qM?$=jlVIpSd#QyPC`iPi@r$le%>_8{b@g?a6OHzqwc7(CDF)dKItQBJHl@^GaX)
z``U3I{C@xa^Pk-Q@sjF~r&MiGcW)lpYuNq?)JuU^>aGr^caG{l^w6zChpg1B^W!h<
zTYHu3{9t2r*Sh1szxL12DrL{M@z-|ld*kP)YJ-12yjFWJuWfY2y3eioeOrHgV#rtY
zOKD%v>;6OTrDxrkxY;_KeezF#c=)rQ&K&MCSKV=B3KhT9|M5w3?fNl+vD)*@RIzy8
z%g?@&^YZ2;f|h$n^upx*e|~MxZ%5qr_TIhgE1j8@W4E+A@7f<tI2ewcEHz#l^uvf7
zs^x2ccs_US@?-sXo~(NRqkrtUXZz)!K3VwQ*B{puyz}mu<{gVRZ1`dAsKXy^o;_;g
z`8QW++XImWdSZU!@}|Gc@AT2lum0#%4T}Ce`p!YmJ@!t>(4PryR?E9CeSc2x#M$H2
zz@MMF^3dVv@CUwnU?ewX)nh}4zCNq8rr^`V_gUY(e&dzS4@b7gTAODFKHOFAwy8}o
z?71@cgV*=oQS$t6n$z}wvu#7)w;ov3Mcvos%y;wqUbuJl-%5RLKW7hO=)^Z0uTC5B
zm#_1wlfQ(wHoo`%oAl9Vt4dd{T{`y=NlVKpII^|m1-&qvDxHxQ>XFf-Q~M(MYSI3a
z_2cf(H{Pz^l}$}a#(gvTc57dFakqoN>vP|Mi*IF9ZZhtg(X~_i!ix+2_vZCFxFV}_
z+T>)MpO=4g`@)O$+X$Uk8rYMQHa!_nNuM&Mec{C;e)|5~KYqRTU^X=^8TZcUeM9@g
zi}#&9VP-wQesnf9H5un+<aKXfc=4+TF7A2ysQXAZb$2oznK5#1`@)M)T<hjLowwvv
zHZ>s`_si(lwSD2m4e=jBqo^!8o0^o2b2D<gw=cZ-RN#>#t%-+cWm9Fzcu>ZmyV@6C
zd~r_UQC9g@&ZZ_N<DMBkJGC#o_@6yDR1S-NxTJGhAQ>0r6%1%!c=69&ryM%B?D*H&
z)LqHATSm7|?F%nnF>n6Lp9iG2+_0?nE%sR5PrIwWv<q`Wd4+lXy_J)sfq8>d<SX`#
z^8WaokT2kynj*HCg^Ufbka5KvWL$uQj4$RP;{!ZoLa_ju5D*{}i$%!9fC!mXEI}p(
zB*^4q88SH_L#7lfkSPHLGPPKRObw`zX~i03T0nzLFV-Q`13F|zu>qM8Fd#FFO~}lE
z37J)FL1qOk$n0VpGCN>H<}db-^7ibUkU!v`nj)vzfy@awkh#S!WNyHPY--WeksXI(
zXnN7~4)TuvqOA6#KDCSaV#Ifhq;?%&9r2w+ZYI8qz-)}LIM3D*7U$V!!s0v^BOK0i
zb%eutu9<K+&&LRl^L!oQah`7`JkAR-BH+AGM+BS~nu&n(VvLA5FV+ze=f!3s;=B|i
z63$C?M8bKgnMgP<$B2ycavhOzUT!8b&MPsZ;Ji{t6r5L@iGuTLjHo!T))5uw)n=mN
zycQ!G&TDl<!+EWlXgIIOh>r7m9no=KZzej<8!=+wyirFCoHv?@f%9gJm^g3N5fkUl
zW@6&J6(bhTTXn?3d8?UNIB&;@jq`RLv2osRCN|FdW5j=qqz*lQ9r2$;ZYKVVz)p-f
zIPcUE2j`t;;^4d+BQDOnb;QMax0$#&Kc{GpKXnMtEt(4yDhhRwcfMYi)qb8%9niUj
zx&6I!UQQ@?VD8`)`O18=;7||x<^x5EZ#x|5z6#=d0lAX+s*tOR?=@hyjD>DC$U-+8
zB`kKc6@<lZwvw>e%~lf@ySXwBy15_+-CUG#*v(ZC4!gNZ!eKX8O*ri4%XsMKgFJNe
zQNm+4UqN{6<|_%0-F!9Sv0Eq;pj!wE&@DuXfZaj`5wKgRBm#B|)kMH<u}p++F(^W}
z7$qWhixot~Zn2Vx*ezBQ5xb=_3A&}A1l>}UNZ2h^5DB}bN+MymR81u8mdj-5mV+{M
z%TXd@w_HJF?3OEujNNiIk+EATQ=nT3D$uP&iGtlq1yQhDsU!+^E7e57ZnaE>ZZ)Vv
zw;ClXcB>Uc#cs8dsMxJm6BWC)G7Y-5pa$JqlxWzkRS*rkwMwF4w^mIw?AFV4=+=We
zbn8)~W4B&GbnMnEiH_ZRHPNx#C^Mkj2pZ6BM2UgjMg=jj+o&W4b{o~iz;3h5gl;ov
zLbn+uCU%<@#Kdm1l9<?SRudDutuhO`t)K<nR+L!SZB-BpyRAxMVYgLHEbO++Z0NRw
zHgwxjVq>>mL2T@{D~XNWb~UlF+h69N1*fW@e?Cx@__xD}%3neJFCbSEe-&~y@xKP_
zlsVAt1RdyhqQt>&r-C@x?Nky6yPay{V7FW5Lbn@qq1%lT7rWgG;$pX3NnGr9tBH%<
zGm2((yvxlAl@*l%%`BP;G^=P9(Cnhw9n|-y;$iDnT*{+zLhgZdw>mwwT36q+y1+Xh
zdl$jovf3|xQ|DY?8}a>!=fXC^B6DqoL+0BEk1Vti0a<J#BC^y*BxJdb$jC|?QIOR(
zq9SW;L_^lwh>mQu5d+z5BPO!dMl58zjo2w3G_3h(*+YLV?4G;~&Gd<>w-jt93&dq|
zKzt?-BxDLeVx|ZrWlBJDrVONHDnM$c3Z!LfKzgPQWMmpZW~K>bWm-UXrVZrJ^oyx|
z!^w1j+)THFygi0wwSNxc?J3^pAkQQpzy2wR_jq!}HBi2HsMM<S#VOyX$O+1K5xIu;
z#b}tXr+p`on`qx@<XXm8$H06G<4XXu4HS!hHcqkVXA=~Qezt~Y(a+Y?Ec)3dnngcb
z%dqHYTNoDoTm!|SpNmr*`nd$fp`WXvIrMY&G>3k!iRRGH)iNCVxfX^)Ki@#{=;z}U
zkA6Nu@#yDkXdeB1J<X$^Z=!kh^R*0*e!hj_(JwSm0{VqGC7@qOPy+ge8d^ZVP)`f!
z7n*1R{X#7xpkHWV1oVpyl!$&YPKoFj6O@R4v4$4WFV@o{`o$(%M88<ei0Btv7!mza
z10|tfic=E$r359RU#g)c^h@=$gnp@sme4QNG7|cw7Dht9+(60bm*bR-emOzO=$C6~
z8U1oSEu&v<qGj~UwTz5@xrLF@uQX5!`jt4PpkGN)3i_2AT0y^3Pb=tGnrH?6N-d+H
zUuj_!^s5b&iheatspwY|l!|_}hE~z9*3&Bb)h1d+zgo+v=vP}975!QRrJ-MoQyTiU
z1f`*0tD!aYYxT5-eyxet(67}p8v3;sMnk{eK<Vh$<CKnmJwfT{*K24U{dzsEqhD{L
zb@c1CjE;W2h0)P(G*AZmjW}hX-$+mf`i&aeK)+E>8|XKhXaoI5En}eHXkiTWn+=qS
zelt#)=r<FTiGH(&Hqmd^(<b`OCfY>5S<9H{H(MAJ{Z<2Iq2G#A7W%COWuf1yp)K@V
z^|Xb4tBJPIZ`Cpu`mGkmLciTW+32_9l#PBnLD}fHYiJw&c0FyQ-)^F9^xL(Jjefg@
zvC;2up#1MpskavXIOYEoIYId^BG=IV7!C9FwEqNh6YW2Z{1)Rs&cJ+v@m~aX8Yl<-
zPMmVk?<6P({Z0+-px>#d9rQa*w1a-<Eyh8=lVBY5yA70!em73J=ywy8i+;C;cG2(F
z(=PhmCfY^6`xfJ(-%T)X%0FszYudG|t3G9z(8A1x0q@-#6S^n!9-#X&?*l5&EDw2;
zC6j&}_*DH5za%dpCNL9*csF=*LMxaR9W?%#zdZi_#@5csq1-;X1*wna^7`ZzcsF+L
z(4SnE$3FT5$qDr}`j&W~+fSa5drQl<;pxeTOFNZz@w~IoZ&{GhT=;oL@(#jt>HaFX
zaB(WqomYU}=h9uw7w0VQ>(%6hmgg)7TAj1HgXT)(Uhh8V%_B@w5l#0$i<QxI7kN?N
zMZZc_M*2oN=w2r6nM&cOQ`i5_P3it^P~vP#caYsp=`Ql3oJGA;rOR@b0Y!2m9Td5J
z^n1gmY7>)dH-AiisdufG6Dk-}Q0m>MdvL>%O|@Tq`RC+n%%7G&+Z)>a-I#vn2aoni
zKHRlW*Fx{J)yKZtbtn74w0YP6>j%>Pcv<KiOm|*||Ha~)7I$1GCfDXo%R9*1usyTd
zZ)(C@{u1BB)EjcPgazVCI3T`+2NFsIAhARQl1d~XxkLt1N)#ZqL<Q1HG$6f12Qo?w
zAhW~-vPvu<yTk_am-r{9UT~Ze2gogPfu<Er8}Du6moIHxJMGI|Ta%k{(2PN|Q!8jx
zr%^Y0byaI#PoKDS&)DSl>V0GHJnypVz{=&XEz6oWI{Dmj%(%g+qi_kcq=TwQxJ9u`
zufCdmZa<^nt=`>ZubRqgzqy>+On8{TsZ0u`kJ)o6`P?3S58b=tUN^ZoF{|Lr+sXZ8
zJ~O{#r_Tw6nJ~~wW+l*CW^D&~e?)R0xqkmTCHFNF-#PqXp^>o2TqEI-`9{Jc3ynlT
z78{9(EHx4dS#BgUveHNtWVMl~$XX-Oko88QBO8sxKsFnRiEK3z3)yZYHnP8w_|M^o
zRE@+zb{mO{94rdn=Iu$|>uYkQUcc)PE4AFW1=ndgi^!F8h<rJZD3l9`V!4PYl}m_n
zxs0fkD~M{jil~)qh<dq>Xp|d>X1R%Im0O5*xsB*A_iw><SneRY<u2mgMR#``i>K1=
z{dil|;giYTy|l1&xVK`xW0J{g|Eff4#p1a$=l`5Lb3%CT{Nk%$kKER+`fRedU|K=Y
zdvU%Xq+h+EORq<+>w1=UynuRN!ANe$>)*df>3o{c^!}Ed5Ua63oW=q18V@9B0+6VQ
zK$0c_$(jtLXbO<3sX&^h0qL3!WM~GEshL2QW&zon4dmDSruVXx6LK^M$kkk+(xTFi
z&oVD%Ry|Hx)z{xZ^yGV5-rDn4a&n_x|C&l_E&5xD|3}<CT8V?~wh}kR-Zxrq$!h;5
zORB~f_AN_o3}3{zHbrchg^Z1`ka1xSGA_bF#)o;x_y`Y~5EdX4A_8P$ScFWBh>%HP
z2{I`nK_-V~$mEC&nG#kYQz8mvYFLF#ji``mVGS}ZqCuvIb;$IH4w(@)ATuHcWM<fe
z%#4_jSz!w@D`G)rhi%C0hz*%P>|d7J+58dz+7vlq2Qnw(K<0*B$lQqAk^Sqx&z_&G
zeLE$2Hg=-v&MtVjw5;#40p6=g+KPGMIoD6sD^~<p-gW(#+1D5U`I+VYl17Ji4jVln
ze`4&VjG1$0&7K=fYk$&z@|hiDX>(J*2I1GGq42^a-ah{hnn};NA+u9<PUo&&ZXDNb
zOv#{bS>tZa?bW@<fAuy@dw5ath>={+VOD=_aKZogXkOobHx=Ies{sRV88Xzz3!<#*
W`AQ#F8r?JRw(;|*v3HL06X)L&QAVx+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp.0BA2333B2B00FEDC.idx b/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp.0BA2333B2B00FEDC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa062a145dc334db915273113ebb29c4d27c448a
GIT binary patch
literal 11192
zcmZvi33wFs-N*BvnWRgy$Q@n?B&i&7?A-SZS&0E$a6|?O;fP$jAji@m;Bo|STaYSO
z1mrLvDsAL2frcWr5mrO3EgB;Rfq;rl0ErZpSBqACm-LzJx6kwDd6*}k{r>X(g#XUY
z%<g3D$Pptd2dAVIL`E)KFl)&SCM6}MQ{vy^C5z@V33$VwlCo^a$qAE6&KVsWeja#Y
zjeaIN|5%5oyZkYK``e*$|JbnWuN#Ive|z<imSvAVyvn-x_S_upe={>K=RRP_pX?Z$
zww`>vyKeJ6pN<~?;lXA9Uih|r<oSnZZ_yvPG>toRbLGmTcQ5;S=AF}8dOy#v{rTfQ
zJIY%+O?~yq(bw;LYWr{h_DxC7+^@d8+HXqZ`S;G8xj4`2Plb0L$c%o~)bhw9rAtq~
zci`!-`<(dQnSJ5nrcL|B*M@9q-22dlkK0=(+Mk~C+Da~Ze&YLEdY(ytA^!#brFyAo
z#N`2Bwg0+g%c*%+j-0!uG-f|={aE#$J6<^O0W<l<4X>u0-17T}&sq=Mzv8tI78JbV
zY}M8scfZa~zdEn{spniW<clX7PQ*9H-eq<zp7X}RJ$-(=^q;{uRx$g!oqcKOlN-11
z8utFd-#Ec8hyO6V-6JnO@%y1OkH2*7qfoS0diHA8opWtjuK9_);N{$*pGOzGRc5X1
z`%L~vq0Yf01D<&3k?9+j{^Q>L-+uRL!TIe@b+4(92kvElweL)Ww&L=QYsTarPi?c`
z{ma^ab+9t#U+cPMRL7#a&3}^iEV=R6yr#U=?DStxzH3eO#mO5!&E8q;*P>Z1SAH|(
zrS*M>UVkRKy=&T-SE{<!t!-F1aOwPizWC{=4$P|T^V^hxBX@N@zL4%S?16XgYp1@l
zYS$~zto=*x(?t#LPoLdjp7`~tSJ_Q1HEYhT+t~BuyB|OMc>AUOm!CeKy{x>u?{5>k
zUd~-H_2RMB<&W)I)I1{ADQEZ1i^Gl`{b0P<`@ttyre0ltv|nn)cY|x6iDm!1b`G;?
zec}Gi7g7gJKK=ES>woZ#-n3RP9NVeA`C-k{n;XBp@XPLn2Zs$B+da5swX`eFXBHj$
zcKOKnpL_L@*&8>XnqPTp!oJn&u2ucI4}Ntt^}E9T@f!o^?fG2>9lUGMz{Q$%aa7Zu
z<u|zYkAGOUW5wIg-TYT%iL!gcsGHmOyz%3v<Kb@)FW3H_`OjhHD=t{Gvs%7iKk#$<
zcPY)Yx_s05>NdApTxA{3IQhqy*EIfc`f!Jt>XsuDs54jltQ#vIUpc&Rgtn8JC<bTk
z-L|#c-g_Tye`3V$xd(TQQkPCRb^X%iWlNspH~n?+N4^a&j9h;A*(;k)9=Mr)=~U#W
zODVDbJ8z7;y~AzGB8RqqH0k@F>QB}`oRjtJnH}4{?KA21%Rfl_Un>dn?@ai<V_?ph
z#-jI59DMg$OaJqM^DkUK9mu!(PT2YMN0Ea&FaP!BL*iEdrT%3tm9snSK4}K}JiLA4
zu9xo|H?=Od@lwgZJG@i7=72kV)^}Ied|Pnt@aRh`Pv_U<fAHSH%|B-C8?<dP|HYs8
zSmVb}ZCSqb4|6Xcx}G)ke}^0_3IB`#^2HN3KI?fQ`qGeVYyZ8b>CcZIe`<Q}k1N*)
zuU<OcSg`Bn&XW&rJ9B2~(S7@h7B62g^B_q{Ny|C1w%|oQH-jpgk`n2f*0o*hCULZA
z{Yw9S*M%qVQO9La6B2P&T2`mljTd)%?dhHm9r)s{49ZQ!9n(6tYu$KppZ2kPk8Ql|
z<Mt_qi8v=Sr*G@Vi|_p5t1o}wy!^EcYEmNZk=ElEts5^MIsPKslP-DIpE5ZSPe`3G
zp>^ZMuY7%C<kDjw_sgJ)6LJ5v{^MFVUcBer2{V1?%3&GQ#6+B#mf59s<Hh@%8h3X4
z`0(rPQ^qIa?9A+6wr;%mzz$#g5#OF|%%JX1#6#1D&TQRy@rj$A0u7n->oTa(iMUr<
zua2!7FMj9DH<5fQoz9@fCgRR%ox8Mdy!eY5xkp*$OF4rYlZd;eb!*qU@!}slPB?hJ
z<n87RYFr}jl-8+T>&A;0&6>UV$G*vDu6ugxGxlu!hh6(Vvvd8C%-qaA-v2&9{W1q6
zDNq=g=q;Q-!WObXTp<U<7xF+tp#UTnia=7K1SA*AKuVzkq!y|`TA>D{7wSMpp#fwT
znm|^e1!Nc6K%v6WMDMxtN1Q?j$Srh%Cg)8a>g{3PUY?iUx_2j+GEhwd$4GL)0&x;J
ziQGs6mw?%7!s0v|CoImhjfBN{u9|Q-&&3Ic^IRk0aGtLwJkIlR!s9&ONO+tVs)>N}
zLYxRVFEkPX=f!Fw;=C9qBF>AAM8tWinn*Y=#fgOTQX`RYUalrG&dYHk<GkERWSm#3
ziGuS=oG3W2G!g~p)oP;Ryc#De&Z~_?#d)ooXgIIMiH7r9BhhePuO>Rq>v5vvyxvH3
zoHwe8f%8V37&vb<5(DSWYGUHN87C&rn~lW8d8?XOIB&&?h4WS;v2fn5CN|F7abn}V
z-AHVl4^@-UF_L^ih2kW161kCtE&)5$#KCzdP8^(f8i|ARZZ&an-i;F%=iNr);{1%f
z86odQ>yOOLn+X)ji?oq<URsgfdOk|NpgZSw?&BSu{gKZ7IuA%vpg1rMUh3h%Y@jj{
z*bFc9KsgD#h+IJe`;aS1;1Dod%tALCW}%xcBP@2a<%Gp<wt}$O%~lc?ySZWxy16h1
z-CP;ru$wC<9CmXRgu`yGl5p6~7xU1~hk5Ac%LtF%d^zE<o39`|cJq~l$8MomfNmiy
zK(|mv1nd^diGbZg1re}Ys3Zb*i^U>zi(wJE#WEsdw^&X@>=rACh}~i(5wTk;mY`b-
zOVBNq5ed7caw1{3R6!){mMV#a-Ey%E-EvrlZn=!e*e#b68N1~QB4f8)No4F+iWTTq
z!U}XNWkkVlrJN|(tyB;NyOm0!V7FSVLbn=Lp<68@Dt4>oM8$5kf~eT7RuUDvwPFpr
zwXg==S{c!>TPr6Tc54+x!)~pTXxOb6>(H%-b?Da1h>qQQInl9OuOK>h>y<>uZll<M
zZX;|!w^2q6>^91Yf!#(0F|gaHBnEby#U^x{VH3K|GGb!4Sx!vsHY<pU-DV{*vD+%P
zpxX*t&~23w3%jjyVqv#cK`iXHDv5>NcCiiJcG!k)yNuY_ZI=@pyX^{MW4B#NZ0rsd
zho-@)Djb>(R7OIZ;Y1ZGC!rUSD@bS`awQ2J0(OcW=yt*mbUS6l!EUFVIN0q}5C^-R
zO5$L*TkJx&8+M`FEh8>=yXC~iZnuKC*zHyl7rUqAO=-Kz^+$^Hih-u)O$C~kHw|cd
z-t;!g`XM-Y#iFa3lt1F`O?CIDCYN<-@1;53`Ph52$E3Hu$xEJdLsv-XJ6y{v#6fni
z5I4#F2Y+(3_=&&ecJW67zJQoK1+zXDi1TqkypIPGd;*Z@6M-b31SI=pAjPKusXi4*
z^Jzf3PX{u529W79fh?Z|WczHOkS`=A*V6GhK(5biBkz6kR+8TO206L3fm$kXf=aGY
zpn(dUMsB78myoOIz(E@3<8<IH<a#>rA#x29c$0zoCMM7f%+^vY`q>7GML*k2vFK;3
zXcql!oMzF_*3&Hd*&2pLKikBx=;vxF4*gsM#i5^Tra1I-RWyfwE>3gk=jv$={ag*h
zp`UAFIP~+i6pwzsf#T86H&Z<N`6`-6KOd)g^z-#JkAA*};nB}GF+BQ(T1r5_&_D_3
z7n&&n{X!KjpkIj70{Vq|T0p;0!wBdXniv87Vl5@2Uu>X6^oz}uh<>q(7SS)pX%YQm
zJuRYNtYJj-i%pD(eyNs{&@VMm68fcPN<zO>MN8<H;<SW*sh*b5FV!#-`lTjDLcd%~
z$>^6GC>i~7GbN*6uA*i1%W+ypzg$nt=$C648U1n-Bcorbr4;ll4U~d@rI}LDuT;?r
z`jt4XpkJw{74$1LjDmipiBZt6)>10^)dosMzuHWx=vS*~75!?QR?)B4(<=Ja8b(FG
z+Qg{n*J>#Z{aOR1p<ipJH1um#w1$2yPHX7b>S+!AS`DM2Uu$AC^y{^hj()v?($TLs
zQ#$(fDq2Us9;bEm>-Dsbe!Yg#(XTf#I{J-T%0R!-KpE&ankfVQMip(K--y!&`i*+p
zK)+GL80a^e7z6!gEoGwLY@kf^o6VGoezS@;(Qn3S6a8jAZKB_-VNCR!O^k_ttCq6R
zZ#7UB`mJWlLcdi-Tj;mqw1s}Fp0?0$)i4(NttQ4ozg<h&=(igv8~t`OWuxD&qHXls
zaoR?|T~FKSw`&+1{dN;$qd!zjg-%e(Ylu(-6*`UFOoc8XSJ9z^G|b28&|Ap$bm&9m
zH<{2e2IiZXP&2SoOF8Iw8Yl<-PBZ18->IS<^gD6dLBCT^JLq@bWE}K6O^kznx0Z6z
z?>0~_`rT&AMZa4`yXbf0w2OYXo_5jizR9@gcbgbD>Ca!)l5+FLlCumGnd6&N=v@di
zkze_K1@w^bA)r!UX~df>82f#{O|{?roVc<c&5R!CeZudLEMgY5kvE*U_`7uv#Jl(#
zwe`()?>*VlKJl}idv?xAevFmbGc(8g!2e!-V@W3a%zEOFWEoin-sjxoM|ZyC@`fR)
ziHD2Y6?O0o(=V=mAnlXf3u*pHXj^Ki0xtBOZK)2jyDim4p68#J<<<Bj3;hd$mim{r
z(M)ONpSsMbIl?3q>rz9TLE)}Tb&=<0&AmNY8Ow^b(SuCN)(YW=x?8{JW2vEMp~QJC
z)j@V2OLdXw`semYmX`QSfMWhw8^!J!_V>XPwK0j+%^sdz<lQauM{<Ve6nXD-9)I}A
zQ^!C3?DfPlW>3nV?hUQFKD?Lt{xdxj4|nX@G1vPn^4ZUK+{ZpTY1Xa(`rgzKzKwGB
zraH*;de3WnTa{SN-V58vyA)5n$+~sN%i9_J|Fr}9|7!>IM{o!9_Ka@nt+&IH1%ZOV
z7_U8XXQ+S$;tDt*zJLc33IrgrKm?KsBp|s!22u(XAhkdR(h4*ny+8*t3Jf5#zyz`i
zEFinU1_~8~#v~79PJsjD7Pvr@@+OV)_L03;pIkoavmI*_`$7LH{ii1vD8F6)ZC>5J
zWq(Q?vtaj##1r1*wjP<@+cj^0=#<`iYY~>Zg#^CC4~JU_i_EnU4w-KuJhIS21Z1&=
zh{#e4k&xvUA|op;L_t<th>EPW5Di&xAv&_rLJVZHg_y`z3$c*x7Gfv)z~Y66O44Ty
zODyR~X5@h6L2o`Yzm2@5rP5pPyC&BT51*BHx_Q-CtEv*~-plB9mv>XPMj0DP9lrZ&
z;^D4*SKYhA9UmXOGA-xyUlOms+05*=gNZ*9WuicfnZ-cMndNQd^(S6Bx9<2Rf2Tkl
z37o~}vyQOHTpi(%`8vWQ3w1<57VC(JEY%SSS*{~8vQkGBWVMc{$XXrIko7vEBO7(Z
zKsM`$iEPyo3)!wCHgc$rgwEmzJaxoDcI$|X9L@{h?Y+Rf^HyR}Z`~({MJ)}i#x*Tv
z5xG(hkuT*Dg;D`gEEN%@QVCHml@XOv1yL<k5w%hcQ7_dIjZy>AEHx3WQVY>8wGl(5
zq1CvCr4FK7>LT8scYoU<sxIZhQycajKAG6vi*k#GcnkaK=Z_8D(5doV;^CZ0IbrXd
z)g+|e_(g~Azr8ixt*Gt1;{8d&!1UHXMS!P43z*5HULY9APZF!Kkg-7)GEU<l<ANMy
zyv9Su2YJW@O@K@Y3XqAK2$>iZA(J!-GASrQCTlWea!`g$(G<v(paPkysgS8b6*5iJ
zAk%^xWV)t9rU!M%49$Sd2pW)?nhBX1G$FGz3o<KcL1t?<WOmSoETn}@?_ip^@Cb(T
zljLX)WKPh5%+*}T+@K3tQC?Bohr|BJ#Jq`Zbk(<S9c5MCy58)@ceA{`%KM9!JJMVK
zk_GmzXrLsyKLuie<w;_rEM#nqg^Y`GkZ~~%GCs;f#>aTbgs1?S5ECF1qatKtOoU8|
zN{~r02{Jh<Lng;$$dsr8nG#bVQ==+mYD|Spi)xT*F%2?3szav7bjXaT0htjqATy&T
zWM<5S%!*o&SuqPTJ8DB_$85+#(NIZpe+tDy%ai0p9mt%R1DP9jA#-DHTlUkR&s`j^
zeK{d<b>OT^b)JW7`;x4ZzTVL-WznqYj9Zruix-6#kGu8F^jn+%d`<JdNueWOZyw$y
z9v!qGZR(6^(`SZLT7T$2`P8<tl$ptI!|?5bNOVpD@0oos&7`LN!q+at-@apq+eUU8
zUeLc&`pCOFckj~me|i|EJtQx9=TNTOV5^TdAm{&kG&8GL@7!PB-nZW!0|y0oL6lWJ
WTj|M4!@6bOJ!%#;;=cS4asCg9POQrS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp.F5D89038BFC386C8.idx b/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp.F5D89038BFC386C8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..31e20d101993128d53bf7fbfe8dd0579b5261a71
GIT binary patch
literal 11192
zcmZvi3v?58zQ_BYnby)4fdchHp{<J`kIef$LjxGVfFn9cD38i3;T;44!=s>!xC`ql
zf?Z(|tUOlSV$~Kc%c8eJw7Bbv*5VeCDvCD95>bKGMVIRh-80Sio^v^eA)oxd`F_GA
z^GGJ+$BZ6bJ0v3`5E-*_(d^|jnT(8#ZmGY^mM@*pq~P_ujEt2#k58Pu__Wcr<>x`i
z9@S4p7ar;Io9=%avE!}aJzs9w_17&UfB%d6p~;mGt=(XqeQSQP_RoUcuS(_`@|m6E
zvo@2D_B3t1<Fj!SP8?YI?<H@^hkn0y&NhASx#`@gt83OAzJ2A-vxZGi_WeD->F1C7
z?W{_6o3{VZ;XmK?$d2Fs?egN{`QLqgq5ss@Pu@Fq>g)n*02O}rrGn^pZOI28tXOgU
zy_bG-so$~ZPt}CW+n#z+e0k`$))yZ5^rOz!vCb!_zPyHuelqF(ZM{$BJTu}Mes8l>
zHu|f9-*vvUc-w~y&L28`QE4rj`|XkXeYZaI(g)0x-COo&9N+fGwI5q^?_K@!2a5vF
zIoq`-j=GnMaxN^WdgOO58TyZhTaLw_jJ?C`S~l<X1N-{@cEvaT*EcXVJwM(%eEpL<
zc8z@hz^@&D_k*vF>h$2=hyOTy*3rEe{~nC?$tl{%y7MlsEHTf>i~dkD{EO(KH!H0*
zw>&oD@4<Zkp@9!S@ZgLsE55w*#qWRkEbz$=r@qg$b-p{9`)W?LXsf@va?zO5>yfAJ
zci!IguP#>Z!izn&jqO_2wDmP<-|{QJT+mjUS(NjuDYrdZe|E~2&x)Qc&(orX$@9M+
zx_9#}!@qqjx}!(d-OttaXxh}WWYCI*-|YTuY!_xj(I-zUgU0OYadZjYZ{*y++|@~a
zZo{tU9^3SG-;-r6?*Dv#g?ac_AMR(LN;W)t`iUodAAjeg$JcdUF<{lnqeUyLZp!}K
zq#j?Dte$rE$i}K)?pvA|9qU%S=jz#!M-G25LF{|~`ZbvsHXrVvS^dM1#>ZksKX01H
zJheIW;?_?y2TwV9Y3jGHW{-PnlO7u1t+RQeVa3%azy9>*n?eUh4jzA#fB8mfSDY^>
zd*%C8W8VMW{s-qgx%I<^wI5Eb*{JT?(Ep|(`^QnwhhB_d8A$IK(S7iN+XfF>rdeml
zw(VPWh3mZTMCH!aZ~gA-zaq<(JzK_J-LdcWYfl{we}8b5_HMy9Bdb<_YRxH3{<wM2
z7xeQPiP_yR=U;f*trs^~2Xl}A=?{;#{&ezSms#qzLldb}7y3OhUOu{JRA{vJEHg>;
z&;IA5FXp|lVWFVq-&xr<VgH|B+4JjRx4*r2@0x07TJ`9SN#`BAdECp<*zq#s`2m-Q
z-B2rE{ll~QtCk$;x8r!td(A&=yKn2opFW!RZsNmNi~su0s59H<uU&h2)rf=58)l4H
zf3{(%wlx%+tDm29{^BEV%<0s8bKQ?l&49{lBkvmU%;SHJ82a|e2DPH=qW5O@I)C~t
zs{gg^mmWS?IrO0~9vaR~T>kjr!LLm(d$suEgAZ8eUc2#9=SO2(>yu|@gg)3=;eMn3
zefOSAv)+Gg@122Xe|08f|G6z|3*UTbeiwCLmy_SlDQvrc#g}EiZ?5GIVCeJb)?b-C
z?2SYbb^Pb(rW5bJ*FYcMS5vlZ)uLGkNJa)Va?TdIQ`x^dXH1xy5h*Gty19Lk+FP{$
z$tpgyDX?2F$)(Ct;~rT(I<+snan6ssK3#u@dQUDjF*PpCD(u$2@W$O<{!Q-(Ui!zI
zxs;n4cg^bBseR#%hyC>3*MCf`dO4SxoErDa>UBf=!W*w%y8FZ{)c1+Jj47${#LS5k
z+ZWz=-|1s!&a-Pq=2DYV<ASV$?(GY2eC%pBUrWKlrd(=VYTPHQPuKQ^H~!11%aIXO
z4xLMlPmS}l^1HV$yz!U&FLs_>l5?eVMkqBdE-1dGec_G&F|*__t9&izQg^4uJ+pdt
zYF~KcOV{2$yh3xn$))a1jfZCqpVhwb#{XVCcjpIc>G52uJT)GWHQ=81g*U#|b>e|f
z7QdCqrS3_MyJdCj)V}b>OJ~nncI}q*E7vWj{S|vW{?o3S&+U@Dh%ZTeKM+rn7h#ix
zMdp%(L*|o&M;4MqKo*lkM3$07LY9+6MplwUK~|GQMb?r;L)Mc-M>dkgKsJ-aM7EN|
zLbj8{PP6xp9hTGnPED_tf|7!M-g_o5QqaF(V48fI&rGi{pWinEDCC=zCRSr1$@*DH
zvLP0doW?<t^K+2oLL4M{jfW)f=OM|5ct{GG07=0wKvD<^kQ6l$lA>RPq!<z*DQOZU
zCBFnoDI`Hs)?`S^ei@Q-NQR`MDUekB3M7?~0!dX<A*uRRNU9+flA5MLQuAw&)Iu60
zbxnt)?$;ryhjd69ngL0}Z$Q!r8IUwJ6OyLigrpfVA!%tABrU%MNh@SQ($;K9+I}07
zcF2ZgPz##gP9-lA^an=(g@TjP<Y*2g9lrxfC*(lV)m%ureixE%$c1ECY1#0OopEGR
z=_H^jrBgb{I~tVcv>yf1o1U+p_>PeDM&yeV-*Mzt;yVY-))N-<Y@D!|XIlx2d9I#t
znCIez!#vkYIL!0)gvUG|Cp_l)R>EUms3!vEg*Xu~FSHT?^I|;_F)zl6h<UM<h?tk^
ziG+D6P9)4rtwh4STu)@o%W)!OUT!5a=9PM)U|xw61@lTPQ82I86BYAnoT!*rTZxK!
zt)6I@*WyIOyw*xI%<J_;$GjdVI_C9OqGR5uCkEz?I59A9v=RgJW<4=6Z^ns<d9#(6
zn78VQg?TGZEX-T2#KOE?Pi)NFabjcMZY4J6gY_hMgrqmpV4MVxBe#;^Ibf%rIGA_h
z#KFANN*v6)^~A-z8z(O2-B#jaerD;+V0wd|RXPhOQX1(X@BD3bPWyRVdV|g{$?xZ#
zLFGmA`{xf#lds%29X9o_Zw^o;@oj|--B(3?yOFDjuLil6_+9~K%US4V!z^^Om4wA^
zwu-RW%~lf@yV+X8VmDXLK{prXpqr~C9CmY6gu`yGnsC_7)e;W7`Enk*`7jUNd?n$r
zo3A1~cJtMQ$8Nrs@YpSs3(zfu1?UzkiGbZg6%nvos3rn-3$;YRZn0d1ZZRxEw^&I;
z>=vtth}~i}5wTmWB_ei9<q~vDVF|jWN+MymR7E80ma2(_-BK-)uv;#dp<52i&@ERI
z8N1~wB4f8)O=RqrYl)2AO1T2vN?3tzrIIMvtyB>OyOnC9V7F3B6zo>ZRp?g3Ds-!r
zM8$5kim2GFRudJw)moxrw^pt}w-(l*TdO1*c578c!)~pbXxOdQ5)Hfcavi$$unygN
zCDE~4uOd2j>(xZZZoQW1*lm;>&~1bb=r$^ef!#(GF|gaHCI)sJwZy<~v)qJkGi*Y)
zSxHRnHmit<-DWj0vD>UACU#rp7Ia%-3%adJVqv#cMJ(*Ls)>c&RxPox+b*}E+YZ~%
zZC4T-yX`7sW4B#RZ0xpciH+UC^5AqhRE2|cfGSCFD;%hTRV26@xtauPkZVcs6=0{_
zfo>=4K(|v#9PD<gh=biuHF2=psU;3}yX7u)yI~i)-AdwOw_8PA>~^b(i`{N5aj|=9
z>C}$1+`LG6X*tle(rG}`OQ!?PD4o$kg+KX+tX_JdfXa)wFJ!teW~Nu`ioPp~z2otK
zA!iPkKm50n?x_#***-CS>}0c9ATFB&;<I@mAzJ_vvqc~&TLO}^WgsP60aCM7AT3)1
z(zA6SBijHnvrQl?+XAw)ZJ=OwP)x4`C))vXv)vBzz9-sFa@xNjN}q@L8Y$l~D!q<;
zEtKyha)R=mL$0HJ2WZI0Y2Ta3&9v_Xas%UggMoY-<4XXujTDQ1wuNHR&n74q{cIi0
zqMwb^Ec)4Inngd`z_93N+ZY!8TqDJypKGBw^m7S{LqAtXbLi*dG>3k!ndZ>XH833d
zxi*GFKi^34=;vD~9{qfR;?d97(LDP3IL)J<Z>D+l^9>A-e!h+2(JwSo0{Vp(N<hDm
zpak>_b+mwfAx;bE7n*4S{XzpHpkHWX1oVrIl!$(@g%Z&(CMXg8VjV4_UyRct`o(5i
zM8DX;i0BvF7!mzaBPF3<YM~_bO9@Iszf?y{=$GQOgnp@+me4OXFcSKuHbz3f+(^mj
zms=<q{c?hm(J$B0GWz8>Eu&v<re*ZY4UCL_xs8$0uQXB$`jr++LBEoq6!a@~w1R#m
zPAlkFnrQ|7N&};yUuk0$^s9}Oihi|)Qqiv_C>8x`9j&5YjngXn)n-~nzuLg4=vUhq
z75!QxrJ-MIp)~Ys2}(o1R!3{-*W$E>eyy3-(62Qx8v3<1Mnk{eNa^U;TPPj<dV<o?
zuh-E!`t>-iqhD{Pb@b~EjE;W2jnUC>G*SlojTXv4zmcE}^c!`wfqo-S8|XKhX#@R6
z17o1yXk!fYn~ju-ezS!#(QhUw6a8i#ZKB_d(<b`OX4*u**}$0SH`^E!{Z=Doq2Fqu
zEc9Cm%0j<YM_cH(;<SZ+tC_aYZ#6I$`mHv`LciTe+32@hC>#BDg0j(X*U>im?Ko|t
z-)^RD^xF-Ljefh0vC$uFq=Lt&^f^SZg$kZTPEf&f$aQq^01f##9efkHnGT*nKEec#
zGLTO&!E?Y)BjupqX`vkSI|<4`zf(s$=y&3@gMO!(cF^w}VI1^33C2Oc+eo?ScUveI
z{ceJC(eKvLF8bX#?V{gprd{;AM;I6VZh~>s{t+9J8CS0?|CnJS^RnlKyfauPa$oj+
zKo4X;0927(5%H3N@jv!|s`2vAsZ-){%(y|`6|TI<Qf6rfjotpn6Ys50c20%zd*>IY
zzi2AxT~O>@!nsp_a&ZCs*k+O!DKrWL-Z#+`#^vAo)s~@|sk_TMm38sFGtO?Dn{}q-
z)2!5Y=}%_{tKn4K*^%k&2X>#%baB2QZ$Y6~lNVW%w*+WK-ii*IC5?Hl`^<(zOiHmi
zGq?>aH)p!Y^9$$yB3&6PjCIicOvd(V;isnS&vRX7a1)d`>oOf=cU`87JU?%KuXO3+
zyv0DVyjTat?il&*kV)FzskK`)s;JDn9Gn*^9#vfCT?O@uQs>{-FT;B0-|N;MdgSP5
zpZ__vc8ew#&F~((;oDJt%=aJbow~bg@2(}@mxzylvGXqWp~<tauaOrrgZTN!smXL+
zg8$rtz6&}&52aRe-z6R7z1}@@+AsEc&m{F(Dd4+1eV)z+SRgLI0r3GINC*f(Vn76v
z0uqoMkb#td0;C31AT6K)=>Z+c2pB+Szyz`a7LXmVfr5eH-Rb?16L5gsfD1Iabn;m5
z1$*Jb`c;!Z-?=IE>JFGXU`Bccjp#JuMz5}B<!hODFWNIY^?LQXu~&ijY3rqBOI}%=
zGkav}v16Dq1JfJlLS|tH)edt@>o3&Rr5@YI=yRKQ(Y-+#AITiG=R)f49()hoyCxf-
z;6FdT_~hHEx4;}`PRH9kFA`;<K+BkAK&zNl9pv?=UjFM>^x@@iBEFCDi{d82B6CfI
zL*|<Zk1RA10a<J!BC^y(BxJdX$jC|)QIORpq9SWeL_^k_h>mPD5d+z5A||rcL@Z>x
ziP*@&CKCJ@zi4S94zk-sT;y<R_;&A2;_XpWEA{&CAYi3d_%`A?tzZ$k3J#I4;1Pui
z0a2_F5v2+VQLc~?l?nw>txyrQ3Jp=O&=HLa1JSH75v>Xf(XOx&gB8JzxDG2EM7P34
zytnk;j?Jto<Ngn~)EqpXdb^jElnwP(>}Oy6a`={RwWm{e7f&t@dwWywUnZyhHzn!y
z9XwBhKj0HMPaI_TJaN<9CS+c@q03Fby)Nxp*0G25ewvcnS6}~Kir0Q2yXFbXs=a=I
z?8*1Eyp*>OpmN%Oq2i_UB6yhgb{nZDeEr8N=@1_E^Zw6KKQDqueeV}7x8}6}k|kZ>
zi~1IO?Wuiu%(p5{Y?Or=8)IR{MLC#pF%D*Yl!qA~<6$O51(*pj0cK)Ugqau<VJ1Z-
zm`O1SW^z=9nH-a0rbHE(DKQ0RYE*@p8dG7WMKzddF%4#VREL=!(_v;r4VW1*17>E_
zgqayLVP-`wm{~CkW_HwunH{rX7K{cLdvBrCVLcXHl_n?Zz|4s`Fmt0W%-opUG5eqY
zK7DqA_VvWnQODVw>1>Cyg2jc4Z}B#*jHR=qGp`?#mMsl0yXX3i8P^y8`7_<SkwHf`
z3zvsKf9lArtZ6f+&zKd?Xn)Xu=F>W28MD$i!*Fv^BswpJcg(qyW-_yG$nKPz*STw#
z8^?4T6&TPhXUuK+H+ApvU%d>|9$M-jHk|7@#OkLFEdKxQEhy~Mx8&wu+|vKnL4$p~
aAj+y<r1WN`kv$7;A3K{Geb<N}asC%AD#1Mf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp.935CC5A9EFAF5617.idx b/.cache/clangd/index/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp.935CC5A9EFAF5617.idx
new file mode 100755
index 0000000000000000000000000000000000000000..78a86e8962f37ed47e1ca9fa2d6c5123000e70f1
GIT binary patch
literal 9952
zcmZvh30PG1-pBc$bD$B%R8)Eu1<E9sIr}Qfq>aZqS|=9<H*%YC%kibeP{S6hn`UMz
zskww(Sr{o`bj^yZQ`mMXBEwX)vI}U8W@U?Yt2@qp4)cAV=k<9!4<Eiiz8^UInKRBB
zIr59_galvETC!l)q8Us=LPEFb|Dr_;=P^-u%bk$0WLtgyq{351*QUP)9$lrM2+cp-
z<;m`UAHMb7+sFTF-Hs2|jd<>cnxT;;4?nmf=j^-lGPR%5QqE<~G2}0|jZ0cfzSz~U
zar9Sq-e=WI{(bMe^1<gGoV`h((>9Geab?-EL$@vYYvwJ}B7L9Z*ZlQGzinlaZd3Oj
zJoLt0k8geIqaO=1=Y4;nz5kTv(;uHWarT~^Y$~vQUs~w<mdHa7l`O9Rc;Ayh^gH^}
ziHbl$%Z69Q{X;i3zjFVXFFNNO?R<R7{$*U~^u$j#^*)jO{P5@b-Hp<Ok>>_{-}#5a
zO~>vzfAG{rr8#5HrNcFQZhC&-r_AJ?>-HwpZ+iK`lR0zleqjHn3w$r$u~~cO$k-n;
zlH2bod;Hn4Wazh#HXRMGFaD6(Q8f3h>OK9QTKt{&trbi~&y%}{J+^-9juD?!KXHe*
z`|GcnogUi#=*z=q9@%~I^V>sxk~3DaW9MF6k~QQ@dBF==!@dbEc&9XH*^R4*e|~$a
z_uzm>?|*3ey2byx^VOe!`O0^C>m4<HrY`r~$^37{i6-rVbC)j~lY2eB<>n9HTk~s|
zoRs+&du$rjbwb0&H>Ev`E<bWlOKxID^4})kx~k^v<aJ+VY%g$Yq4dc4Cx-4`d*iT6
zt3z9RB#n8osz<|`rh5l2p8ws>uSRuYR%D#sq71Zl^f+=a-EYL4_wVYYzPMt?i>ueX
z*Z25@rm=s1eVKXmZ^!ns8zQx<PCc`}cm0Q7JiWa0;_Rizk7O(<yUz8|#2)9e9+-Oe
z@XE4B_AG22S==pi*Ojv)4j=k7Pwac&W6KiT*B<JhSpLi4y4A%Qf32CzY*_1mb>o@D
zL6eXFFy+!~E_=fo-9N5d=OLffF21t<!kO!@^H+};H10a@qLtE)FrPNzz|TvqPoCZT
z(CqaakIk<<mS3?_-Lazob%XcX)Jy(X!<PrpTZeZaRDJ89fkj%**-<TfmR{yMFaNA`
z+XL@Dd*#>QB4yXQQCGI^dFzi2M*=^;zEt~X+IJ($9ypUTJ3aE-+JWEDFD10j>i%PD
z`<Aga;)<NtQ|kZz!m8%qkH6k!rn>21K6RqK-!tRnBg;(xNNqbaQS{FGY1QVQuiW`?
z=c6Nc&8yxvN?n|P>{8pgC5xWrH+(SobJx1(t);h}JinoS-<9OHW5GY$5{k38Umkx$
zmur^<4{ZK?(r<q@);B(wnf~;NZCifsH|dRYze}&~FZA;7=l|B#Gj~k$gpZF_e|RyH
zecE&S`Af$=!*g!T-~Qz1!RqbjK6v4PxY^y7T^gyJ-DOw(5Kq4cw@%#g!Y$*cHWaUK
zEBtqt_v=>e8*9$`rG3@UzEiK;ZOe`iuO0sB$JHDENUs>QrHKFb?L9epc~c`x7r!>|
z+<{B!!+siCT^RV4zp(S@<*$363GE(wam~M1wY>fCk;kWJ{jqGVx4rFnvv0?h?e+I<
zIdNj~p@S6@ik2>zSxpiWk}?mj@$J;JQm6@25`sOFdUT2{qGyZPzvSO`oOx`tIzENU
zkH+aq>D^)rC+@cY$=>(x`}Um_YHT#_n$)#ZY~jQ`(h|Sju=@0{6sjN^XD4Nkk1d?o
z^Gd<qO>dlhx^qHaG|ouNxIVUU;#+?I{=&<xOZTTxlcI62q+Zv=7Eb)wCnGDwC%Sy@
zPM92x$0v@@k1d?|nK@<k7Y?-FlS17cjdPQ7XT}yzyz};oiOXIL=XXx<N8`-2%o}41
zC*E`F=#b>?%SNP76QgljQd;-e!ikSw>E>xlo8OQ^+0nR9QlGA|g%iJj;>X}{Dw$59
z#zo`Qq}1-Qg%f`}BkK^WT#!?!G10hZQqNAYg%khLHNX0F;k&IV)c9!JEvZ|l*usey
z&YE46ofY45{gY!`?CJ3DJ1V}uIm;bP%S!9#{QgnupEe*)9?dhv*(&a!$M2aKCst!2
zWBn{-oW?=M`8mjVjfaf)^N<Oe0GZ$yAQLqaGSM$WCTS96l3#*M)?~<JzYLk8DUd0C
z1u|7rAyfS-WSXWyruj9<bWMj$_v?@ungN;NHy|@LbBMDSx`U?QoERrdvmmql7G$<&
zLuUJJ$ntXYhIQ=t!3nt&I><RvmE_nF8s8l}HN<n6#5c1iOg#0-&BW6N%+?SV=h-k}
zah`1^EY5Q^gu{6*OgNn9nhA&Vd=24oo(~fq=lN#B<GfHq1e_PbM8J8WnFu&9)({cr
z#V`?ZUTh{J&Pz2!!g(o7B%GI;iG=fV4UutP4ig#Y<z^z|yi!9HoL9m`!Fi>bC^)ay
z5EbXuFi~+{Z6+$tYc)i}c`ZydoY$I(hVyz2(Q#f66CLOEW}@S~Q9}%zH^Rigd83&a
zIB(Vv^Dv3;>Smaj^~lY{Yy-Axh=uc3m{>S(H4_Wx?HXd^yd5Ss&fCqz#`&qaQ*VzS
zGShOW0nNyr(Lv6&Y;JPw8Wul*QnOO~IcIctFtvZ`fH-*yJk#LN3V3D%l@iZJIIuir
z#IqB*oOmjbD~aa-Fk8SvHydD~n=K_QcC%%K#csBou-MI35*E9;0uH*l00-S%DdDi2
zD<d3sbLE7?ZmyDW*v%L4(9H*U=;liakKKG3;jx=9Cp>oZm4wG`p+JCcAs|4vP)Y>s
z7Rrc#-9kALuv@4k0(Of9B6N!Z5xT`vB4W2#Mnvou%ZZ5HVkHr=TPl#CTM9_fEtL`p
zyQMNBVYgIHB<z+diG<y9fehVpK!$F)l*rgEmk}Ae<#Hlpw_HhN>{bdC=vD#>bStGq
z!EU9DDA=u(69v1KN}^!5TA)I=8c?BIEhQ>;t7SyRZnd1K*sWF)6}z<p4Z5{}2HjdI
z(Xd-9BN}#V<wV17t&(WitrzIftp{}I)=P<w-Fg|(v0E=EI(F-oM8|HUz<_QeU_iG~
zN(}5a%7}s8MmaIC+o&W4cAEv}G&n5>%-KMt#M}rcWwVT!JCVzYS%F+h%mctyfd$=G
zz=Ce8lvvnpl@SZOt#V>vw^d0j?6wPR=(YnkblauG#%{Ze*w}5C6C1njN@8R8#N3G;
zcTn!&q})kBQ*x&O733BG-JN@P2RUbaCOP(W6TfiGa?SF`uRQZz^MMw+7Iskj@7}==
zENoAs+(G-5MEljm_;y*;cTuKOkv;g!Lj{k1l+`tQ<L>f^j@BJyT`Umi;(&M;4<xt*
zAkifPNiGRUcF91BO94_{Dv;*VfOMA*WVj3<(`AbBKh1JkK(@>7Am_R8+T_^tV0?vo
zBE<6xULhldMdl)eL*^rdM;0PPKo%oJM3y2%LY5;$Mphz3K~^I~Mb;ujL)IfiM>ZnF
zh_my2yPPD)9&F+7tfM?fsrUx=G*O=8$gPy84Y`W;RMRjYrakW<H`1QZkZT#wKNy&A
zVLYwCY#qg-pKYR8^s}uLi+;9>X3@`vX%_u#Bh8|ptz}sBvn>pZey)z<(9bnd9QwId
zibFqFMRVxq!Ze3|u94=@&($&<`neW{LqA_f@#yE9C?5TME5)OqucCSM^I@7tKi^36
z=;v!09{qd^!=qoQqXhH|O_YFsp_LNQFI3S2`h_qppkHXD1@sHGjDUWjg%Qv%)=?t*
z#U@Hbzt~EN=ohPK5&dGA7SS&@(jxlBT1G^_*useDm+B}9{ZbPpp<imHB=k#Fw1j>s
zOiSpO8fgjrQY|B)Uut0_^viXWjDER^lF=`>QZoAGDq2Rr9HwRT%Z;>*ez}&B(J!|!
zGWwM|N<qKUL@DT3S}6tnN)@f3UkTF+`jtjnLBCSVDCk#O7zO=m9i^gQZK724tF4rZ
zezl5L(XWPS75!=>t)gG8WmNR5EsTnOt&Y;ruQgE``n6U{L%&u<Yv|X)w1$4Ik=D?!
z)iN6TwH8K0zg|b_=+~Pl9sPPMrK4Z3qILA^VOmGO-bm}{*J~La{dx<dqu;2b4D=gK
zl!1Pul`_z8RM7_djWBJX-)N)_^c%H|fqtWfG0<<;QRY!9eyKK_DDyaSD`mDJSJ7rQ
z4fA2zd<VIaHa|l?%$P?Qm~UmwHejocve0ieQ5O2GR?0%ZRYhCqx5BiAeyfqT&~F`P
zEc9EgjD>!?j<V5jH&HhF?N-W0zg<P!=(oePjefh4w$X1NW^DA^t&AP_4_^^UxN>>X
zNrnmH9V8RPJ4hyocaThQiEByFxl#0u`>p?mx*z|F-cOEZMh|p8ICcjMnZgblwfP^<
zeEe9Xb95-RcWP$*BaO7)X_?Lk#&_!L3)9%uYl%CUZlwF1&lvOU)SJ$&8=4sXc3!8v
zE{=El*_CsWzRWt46#cYwOQKm0_p#R2L~Adwy(Q7ce2#lgx>Ms0E^sdZTI62TK{F-m
z&F(X54>D23+C*~`R<2F7k!PpRz9C*&lwQ<9_b~~Z%Z1+?uKu3O6U{YHVl7X!knQD(
zHu7xu>|XKGkUIoa<Sy!<;?X1iIe4NrCc1Vr%!~=n9j!Z<X=YAv-aRaT@ZjS|zWVx&
z=xWTElrh~Ia{8mUg;(EdIJbo>E}4CXe6qTCbh>Nru365<RZoAj?JoA=Nwco5kyjE;
ze0Z=b60Lpk$Ia<Gr{iHFx{~`Y=pg4%q>^JVNgSp7|Lj)IuN{5Px%$2&K4hLJ<}dg#
zc%E3u_IYB*+1a;JlVh(`;x!(hXH5K_l=ZPdoR0(IeLRrh6M#gY2qgI=AlWAaDLw^A
z^{GIbPXp3@I*{QrfJ~n`##w#tpyjiGY@ZF3pPN6**|x8=Kelwz*W1=a_mAu;+0)~H
zwzre_TBojJ$(xB|7VH`s-QvBj?Um*{FYGJ2_dsFttP#<l9m$Lw5I?-<GIKkq@|Ll=
zHSJYZ(Vy*O^tsh}VO6V)3nrSo+N0m@!S~Rew}0V0@A+w&$KQ+YL;)tyaiF+^^O$)+
z_cHea6*I*h<n%{3>eaVluu&U`=OjL|H4qk=Yakpl-#~a|p@9g<VgnJ8r3NA)%MC<E
zRvL(ctTqr8S!*B~vfe;+WTSx?$YujEPvXAbKrCdtf!N5?bEn_tY;@-w8(m3P-wDD>
zD)FqubyUJ4awQxhU&12_B?6*YA|gs95~5rpBPt~dqFSONY9$(?UZNu!B?h8dVy?v1
zQeq+6B{t&Z+{qnBLPNrR$JSN6ULV~f^Rn`WI_v7IZyp)8u3P1)=(jT`Wd@vcPm7Rv
z`I;`*J#}@uXI{to%Xy<3U9(qTr^1@`dWOdjF4oHeab6CH_wqo3R{#>dB9P>jfMl->
zq<9q|)vE$&UJXe1>Oh9q05ZMiaOWftJ*RjrAlqvL`E&gppF_2~DxRToDz9Fwd-6SV
zoOR;7PQE@l_8J-1O~_LiUpJm&&(b)tAr>;Wn1zfBagcGv9Atcmhm0@gArnFZWJ0k3
znHUlw6N^R2q>u!eR4hRzhh)g)Vi__eq(G(=E0C!n6*9G0g-i=+kZHvlWO_)4OfS|U
zGeQPrMzH~z88QpwThuHzm&VBoS&&)97G!qFhRiOuJF-9DK6N%vyO1Bf3Rr6st<7+q
z52c50bdIBivqCejULlGW28zaC{bKsn#ee@zbG}HRgST}p=$<loO=i;68PlfE3?#&U
z=)d{Yj<JN9@h=1L<$_>nZWKq)zLRDWldf@fN^y7Y+T~iSo9WB$mTcXcdR_M(|I^DD
za`Vt!?=8c)o`ZAxX#+C<zwf4{_vxE;{S7zvziHqg4=;$as%I#@S!qPiwA)6_qDI~|
J{C0B3e*iWh>_Y$m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.81A12F337280E714.idx b/.cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.81A12F337280E714.idx
new file mode 100755
index 0000000000000000000000000000000000000000..86f594f8fad028a91df526f998d774366b8f4628
GIT binary patch
literal 3546
zcmZvec~}!k8pe}Oa~Q&@Tskxa73nS_0TS*@K@L&mkShp5ASeuo9Lwba7zJ5GMNsg7
zQ4qNVgh7WDWyKj(WE4eNM^v039(XVaqO0h-3Hvl%e{7y7mHy?e`o6cms;+K7PdB##
zWg5*T+A||Tl)4e6(P-+@Z%S%%>@^0Bc0-9q%Q(>!nBf1&TJ7WavG_GA`>T&i7(N5$
zoP%S16Ta0RKW<yrHd-?3JpSP6O;cs`i^Q+FqW6!1<_~XwZb_0qNbjiPddD3M52zKr
zpZMq;Gci6=-Eqmr?t3bJ`{li7dfmMx-?q1=M08nyjdI_*y?t>8GwE+r{CQi3!z#14
z^%o`zm-mcJz1wht)9hVqqBipD;7Sw6iX9bCcXnxQdSH_2`Cc+o6|~{`;YPETUH8rJ
z*d{)9$YM9~jyto1nzw8?=p~NaGhEeM#?h@P2{<#@asAu3#D#*aF{{wskA7v1pEmnC
zdXGPJ1GFDwa+3>;PB(qJUHfU_H>1GlzV^($l_lvx*22Ej0+UsWo0;LE3SPs-MOR+l
zkx#y$-Ds6KT<L81udtzFXR%<&BhJUB&qzNk?b0e`No7@oa}+z&;>zZDdqOv)$yaww
zG`el0&&4BRT2cD>!l#mnTpPEB;>6vrLNzN?tN(sEDlu;8iOb&U2=l?@?QU)D9FI3|
z&!i>fVCju!(duzyHu|9wr&oGsYnIh}wg$cr{L;2|Oh<8W0iu90S5~uoj-6(AbyRQC
zt}Z+-uka61zx9#=HG7RtqdoMx_Wg{ZzJ>MEX?n<lL%$nO=-g@1E#Q?Ho{_k(ioYy!
z;@>}kojGD|Tk&zo$EN>t{<Fn<Rs0&${4UIxY%yJW{ycay&Ux&WSd#AW&R07+_gs7-
z;?Tb0h&3^E#ns=&t8r%UfxFw24I`iEG?#~G97vXXQ+{!&T<iqs*_TfLzU3~lob9f6
z^G`&(>8}rZnc=SMYRc9gHW2jPPW>yPEA^L8#sP;-%v@hz?X2{b-|+9Ak=Iq*!Y3_i
z;>wTiJCvn-<XXLEqua-GrExujy`!wb1?yDJx9aj*BLS=TXiC$jCZp_QzjW8Kw{9;@
zIq--1PuN-&BYQua>{>iATQqU-omKy}hZ>h|`uE$JZV>O#XST6dt3-rbCWJ?ZDK%S4
z?zVq_I8>RIT=b)#aiwC9*m7loPm*Tq9+8=YR_ee%JKZm90PuR;DstWWd(k!zBF8Hl
zB>laTQ3K(jTOh9V=JMv}%$JXCC8Cn28?D(bUU4IBpIE6orjojYW4*I`HC{F9&O0i&
z71O!mr(0Q!pC4vyvFzJ96jPw(Q15<0c(-WRJ&V36hwm9F=d%Jc!@8$gGZ&SV8=7C*
zO`je)_Bp#rqhGvidi~T7HF-&?`Za8i7T21hXyfEXD%I~dMWj1=X?!*Ql9cb%@b=}9
z@b-)RHLmP?`8_ZAI^TwiuRr6=XzToEwP{LbLgaDjUF|gU{$1UR%oZiY$B!mZl~+}m
zZKTTVM{w+7y<CvURtfQ!%3AVT!Lto1AAaO}<Nlcvp%SuQDjUih`ph<@JkYviUySb3
zQf1mYsjRN0zGSu`<*zraDt5l9N@CG`rLw7<ssC(4%CuxrOiH#QjV7<(vESvCm4PZt
zfN?M`=}Ag_%o-vO%!?wNgL6rHECJ5Pts(Njy(l6$1k{t@6HrgWgMfMh4uE<BK7e`x
z4*>P#^0-!W(-WBS%pkJhS<Dgh%gJG}7vG;LolFiLE2L8sz>4TtImCDl9WSI)w(%l5
zUJfylLnmOJD54Xv4sz%K)<F>+z`8Zh+KIAk!?T&AB`TX8G`lX1%1KSyVlI?L+!lAC
z3=y^jM8Fn6WX>~pAcK=zPLDae)rzjpoIESBGmr+)L$G{=3Xhc`SUJqK2v!I4B!W%B
zEM{Q48PNC=2387lE{g3!DIL#4@O*@t0A7aR<uKPGcpc1>2tEa~n1SzRK;ugocqz=e
zD837&bRrKS;Cw_GLcsZmT7-b}5t9f3=Oe@n0?tR2FbFswk&6;=K9Gk1I3FlO0Gtoh
zA^^?@CJ_MV17Ze%^MMiufb)S|6j1qA<iDkT{*XF^paMrlM*+FosK9x?a~Mf$j*oI2
zr^^~q>1r!t#?V?Q;>J*=6bXocBABD~-WtnZ72C_vSpx1nt|}Rq41_Re&nIe6!B!$h
z_HL4YYB0=@o(J71eg|HfZj^u-bfbXcNORB5`5U`BWCL4KyaOKCiV`q`RuoViXP2?#
z<ZrHnvO6aMODA1Q!zAG8&~y_Bh(H2>$dYI2NX{||r5&y3&fJmts6z1;*hdvgzznKT
zKyjLv>+c}g(nne~^;POY<hh*f@g@8a*EB#@Oarb3nMNG3(5=zTB29KEd1HQ*Yb*=f
zNfGXi`%!biv+$i15#9u(5m^MJ0dD|lAPYbmkH@nnZLkDZJgYg{mv>;%{RvA#R>~kO
z1hNr?hd_yg2#7!sfQZlITav+xo}F07$qo_8N;!by_u*0wpajfd00k72siZqi_PnEZ
z8obRL--t6qBq#PGEP?hsZ57I_VX$Es8R2Y3r<_khm7BEssw%2l<gs{lz2lfr|G{<X
zSbNmokTlE^IH68+M6RG5bM}c!xxgCeSO@&nXrSXA@Kd6JPQVNr=z!uspXE3gtDk!$
z+b1FFQDhd$)sG`MO*KYV`%`qR4vv3{j@QBQPtggO!6`a`*_LN(M^5QESMGy?s&=7l
z`)P9Z$-l{8I&=1QfSlVmwZOI~iNC#2l$sD?kyK5wSS+3*T!=@K6_&1TEFMo0AtWG;
zh$SEm2mz#lSTLta-$UXz%3exRlA}`GWfuXm=6|0c@`A<?s4WnRuC+{x<bp#20)uFQ
za}q6tx(uc+qY|RyH%rkgW-SAilUGojr^HfLQ&sU)U*%$?&h%WVp*??rj;^(hoxOvj
zlk;*-V7u7TSpWb1tF`C{Jp;DklBLUV!o-woX2G|@cndk^u9{lzt3wet?-d^OHU9;<
Cyu@Ds

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.6BB8AF1DE23C88B0.idx b/.cache/clangd/index/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.6BB8AF1DE23C88B0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..41883f77dbfddd19ef1664387cb98cd4d722bb80
GIT binary patch
literal 3546
zcmZve3p`ZY9>-_)>_Lpj9GNKFBR%$BVkTz1ABjXB;}OH$N|a~xAW?G@adh%1rIJvn
zjz`HgeNISpQ+l6dBt<!tlcO#nxA)@=pU<|}{e13xKC|}z?%#i{|6}d7*7kFEb(N{o
zXwDnmljFj-hoUqZjV=EtY>$s_W6)@qRA{v1;<msz|6VJN_dkYXJ=G4JE*xO^$jnUg
zhVPB}o_5<LSu7pQA6z!t{l{IdD*AcA??Q=tY_w?C`={Lt7`JI_e6<$~+-j>krOC+>
z1M}phao;t4#U*<#&mMX=Jbs{m@J>q06Kh_G?)N$)D~BGo-Omk|Kb=4=-<q~xZ60c?
z=N5#Gf7)UCfFr8#=O~R2#29i;J!pBb#W5uN;-+I)&V1aut;6`j#NDprBBvOgs1nOj
zz2>~Y6JDix#v)78+h?8XmT&18ymu_;!5z>U^X%pK@0O}2&67nYts|idNiyw}iQ&0%
zgIZ!Qrgo!?rtGm~s8f=gqG>D*@3C`fztdz~TpNDb<=A|b&#bLUU-<U6Ykxu%!@4)S
ztH^ST(188Mbi}RH$s}%ppt3kT|DdUDUwP(Mp>IHGU0#jOdcK9;(VX*-D|05zCYpKp
z+Ivo^H=BCe-fsQ)Z(XCV+m_2GYqB2f^KI*^vlNVs8s6B&D~z#rFbeMBoJri_-(<1u
zU{ja2ROz?L;)69~`jcvpW%-ZJ1f5GZ-u?8~?W~L#<urymv#cX8%uA*b@~q=xc1N36
zdipbIr0!uCCB6M?^IK`F#kc4q;kl1?HL1~+-6Livv;OMNaXI16?i@Qk|5Bgy)Z7?v
zb$<_?5^qUdW%1D0>*H(}#i{^XqIWl^zTj|A&(XvWoYAxOcbZPVuT+sP3%ax-O|Qqm
zLsu(qeR);!*{<C;Zbdm*v+FjqM4sF$D;xJFdOknAm;do-*v+QYJ@rvq1$n2{!ndYn
zKTi65z>Zs9A-nhMDvg@($jcurLN518j#RHYraeD#U~N0jF#Q^~EwRI%`>D`RaNpC_
zy;;HI<)f?sBT3Mz-(I}l7@rxFRiUWUz&a6Br<FN)zwz<Kz<kHwd;2>ztQ_t(u8b`d
z9$ciTsaI8ewev{*RkP&V?~5W|#?D&zXVp;2oD0`l)R|8;7Ao?%MHbw;3rtbbSX9rC
z*f^nW<4plETbS{}X<nI2Wkc4J<xA~3+hxtN`gQw{Y9)TqNn2bUEiG1A`oX(vY-n(6
zd>;4CD=opQ@HyS_KPQY9jY*r|@``gs8~NLwN?aRTJ~gZ_<(FOEIBPs{Nrq=l#Kq@9
z?Gg5$V=Yb={VqLa9#p<U`1;U5cYJC8k^2+>Y!B&KQ*-jDSQKjC-12aFdiACWZ~LkS
z@9!b!ZZP`YRlDr0szP<2?~D57p|NkYkxtgLr0Yw*>O3j=VK^&g_P2(!`dX6-BizV$
z&Wp<PhELsiCD8tMQ|H#ZC12)ngPlES2|MG$isY->@yqx~<*S{wDu|CCO{}h<t~lMu
zmFb_m{vqj5d&M&qWVKv2R50|KZb&)nJX*VJ#K(q3^Oeh{OjDofhLneMk`!Ad^PE%=
zf4R(2-~>!Jr2MAtds$jxx34NqD3{qB_JZk#lxgwd5ea*gXfy@=h6B!JmImr9F=m4C
z$Wi4az^ov0!#pU$O>iC=Jxh!Wa4U%1a1V+I69Td%1O#MBxDk*gFaeMy5CF&$xB<wL
z%ja6o6i3YC^B^+go6V5&=hD7(4}rh3JegEFmQAM;z;fx>aftC$I-X6Z!p3vy_;HAd
zR5}6cL@u3xb&yI2unuzR0M;$}mX1`&R(z`&V%P4Un|R^%^{F6*m>(3P1WTa^7vg?U
zcoIB?B0@+&8X+Md4G00G0SSOKbH4dZtyp3UzQqhJP>ZtHYPmAVlm}&lIYVLC;Lea^
z8v-I=10Y(;UusWgOs)e=dHVg&ZT(Vue!$j1K6oaA?L(;Hu>u4;4s!*9Rl+=hVDDf~
zXJ8o&sDC~KI|6eWitR-y9nVDYeF&8RUVz}oVXi>%N|+}Q{2k2c3_ODY_0MPEM_^7v
z@x3Ue6PXAB`y&bv0`^B#AO!4>m_P{FACb-=V1GnDgMj@JX($2v1DObb{ec1m!2Uo5
z0$_h&0s*i;kj?<GKakG=us@K70?MD$zW>lZO>XZ&P_ey|y@;G`RP3nfC?<(42BJK@
zxlwZu8+jHDnmGKo;2bJ{C6q86nlL3CCr4n32_*s|paf=Uwf9ep|2kyHq_f1_SzL9}
z`T7^XL@NAY(8rK(QrA%YAv`tLPy%Lf4Fwd3nYlM<hPE_Jb*w}22DoD#O27>2P(blE
zlvY^9`|@zAV>yb~!X3*|0%lN-0*ZHdxZF1mvha~-z+S-?k=yqkx6kVi^SA+1i5YOs
z$;8MjU3vOUAgh{_d2rsNF$*xYZbeuyR7VjW4AoIYKm<epB0isQL5`mLvRG)cCpdg6
z85xQ{f|HS<1k6B&0*cROV0Wb-OPZ(R*o-AXacstupg1-Y5P{7AA^~48vwxg>lfE-}
zp!fj75_8sa!l>a40}RFFaOabom_BjkuJTH$tEqFyJ*RcG!*Gax_bqu}Thw+wnHo#%
zfI7@j|I5^6huHt>ohp7Xdjnbgf!6mg2%37tR2pS;tP(yj%IJ6{d_a`Z37A0{9l&hO
zx3(pH|H)%cDyq-Po|;oFu0HvgOiR8}rcQHmPFer6xX93$%G1X=WrS7Ju?9H6YC7Hk
z2UtxfU<TE6K=C(?K<V=>H(n{pGY!F@3S-e&EE$;mEeOHG$a>2u8jq!j2q7Sih$bKn
zgaAkb(O^cCe}}||PJNUl#BWGgDZetr+{iDZz8|a&m_fk-fg)PK%qI&?MBN5bx53nH
z5OupDZewhe96chu7${RgQE8S6OI1T%&7Hl%c@bOLefdvwG-qq;SXtZJ**iEcGu1M-
wF|xq*|KC&2JYCGdaQ=dYiwUrV%UfzLu+-zvH8FG1;;dZt55(2m$&K#$H|iR>OaK4?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.6559BA68564A041C.idx b/.cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.6559BA68564A041C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3e88ec4e8e0efa0966e6154be3e3358bb49b50c1
GIT binary patch
literal 3546
zcmZvedpJ~EAIE3!*)EMi6jH}#RI>Ln8I1e2+*5MNZRAcxB$wB%3z{NFDydL$ibRAs
z>NqKH>Zs(DB5$QqBVBZHMC!!p>CAYZZLdFG&ogW9kMCOR_g%kr+g`3tPQCIBhJB=K
zT3l3e7{*{Ql*GTJ<izM3Oa|kY9D|W|qSH6d`-$1yFF!_NS1atVEEO?5dJVY;N4m#6
zD_u5kTh=^WG;BBe@cC_hd2CAL)lk;;sJG$shhG~Lqz@vk6+HKtgTX%4QJ==X*lida
z9ja`-Y;O4@nYjC=@1;gZSJC(FO-Z5cX8%MuZ{6OqIE|I?p9$kRTLy#5Gq?5R$A>KM
z9GdvJ{sgzdy;^VX(DnY6dNyS{%AW6RSKatfFWvQ%XsE(({i`E&293KO7~Zppe`=k<
zsplWJ<M=gfS%1(?7`|t)qN|vzURLCDroZ*(_wVrw1sNlzfxDkvXOEsX_-DA!c;FUj
zImY58=INZS|9ZFj>%#9kzLDK6>3hqIQvJ+Ax|8$t9A!7Lf&*pT1`7+WzPTrzcu}p+
zG=8w$PWyjB1BG@%!GKGQhk3V-W>CsyM|n|sMXg-~C(!8XrdTUdJ)quGeI#mlTbReC
zqe4bO>V^F0qOojqr`p2!-ERX`$`mVqz7i21Gw{rQ?_{W9f8uth<`%BY`wwSQ;<E76
zI)g~1m=SZ$K#}cRjdNAY&bl`Fe)64aUNf>lwtqe<gR@pvaypNl=Crp~Zd9wxKQ1lv
zdsL6vlDxTA%56G(keZhL%z^HOXD3rM(D{e{)E!%JuTedZU!;9T<m?!GCCb+L!3q4#
zQA3NeF9ROtJ>PO)F5au)RhQy*aawPS{?ZE<!240|Q@8k{RO^qPYLVIJWAjn#mK8_M
z$bqX4-sWy~(|Zrx-=3%){%k=*NpRYMM5*^BmzGLJkAcp8smPBl_sQiPXN}u0P__EM
zK5L{0JFKlLUUNiC(0w=guekQ)-@fYl9MLmycz3O>+*5k}-<?D6Dz*iW8&$=Wl<qr}
zA%FD7S(Q4cFXs=(boO@*v-{_-RW#hH&Tk3_?5@(J`i=EEna6(XsOD_lemLpCBg0?t
zH3~XbUgnwAM0{q{*ujscJvSaJU%u_#W2wJhxI>fG%vq%n8f+3593CXsV4`%Z<;UZJ
z@|47a&0f0YvYkScm3bZsDouN$46Id?d;i<!d_@_6ccZ4^YuEKfnm-I5Evpsvbcu$w
zLJr*lF^6w2Z+OLe^VC8VRrEZ}jML~AGt~T*oxEcrp~FAgJ+n*sZJqj@Qo)@KZ7Y7c
zlfnG;aoQG>?wtc0@>H$QI$sRAU$CpssC&ZtM_ST_4Bzyij!E|PB{})phBuazr-zPx
z%dA)K5iXltH}O+dPC~L~700E~p{gKKH*t|d<)@9IsWxuP|L9I7<l5GLcry@k_jT@S
z2TosZ=WF8y-v<kCzT{4;^)pt5C8fuO9~bXx+v!gqYhI@}%Ap=!41uDwqRdPoR%ZS-
zW554=?6HQY9O^BWRi#z^X9`Mgw$gHRzA=7Y4qYdfwWYN^W(rDv9Mw~*xo1k6&F~b<
z`cnGdGX*6NqotLP{=R-yp0QReE2%0inJFllkr=fhDN~lgkk)M7Z-2^EOOY+Wxj2vV
zBqn3r3?diYjV6Li@F;t10bxv-LF7WX(L{1ds3vJlLN!Sj5~>Nf0ICU$0aO#X0H`L9
z&oiAJy}*EP0Fe>jXqH%0CkKUY#@;OPV6qTAAE5_;7a({E#6%WC<Ri3gq5vUEASSaA
z64uEAgoJgFg#cIw1qgt3GrpNEZP}b}K1)j!Hd(8*Umlhco3y~~X^VsfVNV+(El7xf
z1%SwqZ)i<*PAxeN*342XxHf(AoXAp3+<6X)=c07?crl8Xz+8>uH878(_yo*CCcc{q
zwJ&1ghhfgf@Ld?K6FDf6i_!xiicz8j=4zCvfq5JyCSVpaiQP=7eG!v540AR{?80cB
z%t1*w9$Ac%a6GaaCE<AFI7-6tNFkGi<B>&75{^e^V<a38<e&hK2Z~Vu#{<<Ufa8I2
z6u|L-kO|;;poj_Jcpw`Cbi5V09~j?0Cl8>Qz(&?aK+QHLu$yNWMA7Qw!(5xm;yO$`
z+p@SWG#9djE|e)*5+WcAW@(+f^0K#uR#FICz?;KUr1}jV+PhKuxz-?4ywy7};vSs!
z4vd5ubYOtya6{L&d139Xl7>wf(F!+g!bq4w69#CGu}s@>@(+i8$)<?IQ=yT>5vkBf
z;z)=<9DvA#Z(>6YfNGDjW)4_-Q^2`Qz!;@xdDHhKW-BnF5jI<akuZY_4A7k7=I|%*
zH}Mc>P03Nok2-iWU8aH$@$`Kp$<*Q*QC@_h3!Ul=j8Y_rk~{82SrbzRzLO@xo$#Uu
zN@Nf_X(HW8NFy^yNCWNw(m)1)G(MkiM%iEsO!=m>v@hqtq6cFpq$J;dcmQO>j|hMQ
z`;ia<KLC+2-`Ireyx`@DwcN~rC`l@NG2#K7%3h3w8T4X+=DPR7^Z?O`{U}?YHb+f?
z_Mz>s9Ypy!m)0ib5m(_PPOPGWqAGRbQL=cFC3ht~$?z7rw$5fGMDyWIaR4jKN}Dpw
z7T998v-IG`i#~;+k@J#loI>y#_?|e05H;|<a0($|2B#1JvjyM6k_tjS5!TE_O&NUg
zGRv+|>HHJP@(ECipq3Bylw!@?+f;3_rCJ1Ug)P-0L@R8m79n8<wFsd36;JB3jG|h;
zWL2v0G$~Kt=lV83i~r-bthoOWJe*EE9*xJ+M1&CGRG!6i8%@O0M23)%Mn;p620{R&
zfoL$RiN8Z)!z3RiNr@3j&f+Ii5cFcR;UZpCh?IYTkFOuYXZDkY;YYvv(ysyZt3Ul3
z5f>S|NsMkA)-W+CX&Ko$a%}mziVChuj`liAEZ3FFYV+nVP&YHTw6eCbwOg(NEEb#S
tYW~0fsxJCTLyM!mWa%=3)YIn~7#W-5{DoXY2NhN4Re`9J`wACi^}j$#bgKXW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5F5C907654F27409.idx b/.cache/clangd/index/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.5F5C907654F27409.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c79df4efaa8f41d8f5dba06114ca0989e1d4c947
GIT binary patch
literal 3546
zcmZve3piBk8pmhWtYI)NlT4K5k}hj~E;BJ>NN$nHEyOUp5>bdQE>Uw5k#@<Y6bXfD
zyA-zR*&)&CqWiAgQj|^E+1f`Tr_%MzIL|WQd7iVLXTE3s=KsFyeZTkpzHcpGHy4)<
zRSLx^%q=M<JU#@aP$(?vFD^bdvW-TeTvDM>l8V~`V*L87G(P+oiS|&-K2tbITis!9
znmclT%;${j7V%=qP~MQk+nzt}aa7UIgTA#TZqbp#-5;LyETG+?to6}e%y+G;?vx}Y
zi4820632bk_7|7zy*zvP{m6Lsz|h^~)~D9oVBPQaMpkyctY!NfFHfCB1@FurU2Pd|
zYv2@wj!*40d&m}6_^}z|gHeX;(+^u8ZnF>0ytpO*%GpobceESVPTcE0A#{w=i6{}g
z)oaNOIO$oMYb+F)-8tu2zkFN!(Ea?w5ATA`sOPV~e;25lwoDe8ULOrnNbJx~o*0=M
zGo&T*q-!@hYj!*l4|j^Qk~NJb;l0b8AKh&>F0Ko|?3_Oz<<aYE(iXnE<1!EzL9_15
z>@E^)^EP1pZ8qv!>S!9XfL~c0o|j{0(_fyk-P^~%v_7{+X9Let@A%;hPbv>jnoqQF
z@pbndQ*Jf)w!Pc_>7Tkr-FF1bCu<HpJmAyTUoYT~zBRnLn_C!VZD$nJ%RZa1)34dm
zA*Z=pTf+ElvN)$^On*}CNk?AS*}(Hj#(SRq63<MJQck6*)63d(Lp?h*f}giv%xrJ-
zOiOz%*{plinW1-ZUEX!d8qsa)Xn0oF?q)TrvfHNFNzA``4m+Q8V|9+5nSZHYa(Zr*
zm%5+3PKlQ|rm}eW>y2@ivtpIM4biuU-B56}xA%C$NA}xu4R@PQeW+BCI0Rl=k*e2g
z;I69`v!T4I_+0m%o3|tEtXcJ2nL-cFm6c8V5<FfU-N*ZMJoHv`%HD<ut%BS$YT?^c
zGoK`W&R)hTujsh{%PNhU@XeP$S_WV46CbNym9ITNVQ}3eoM!elbVowFEoZ9Gm;b=S
z#jQob{Z-c?e<N|=s^4C|35(5$I#i*k)5ts-Sg(~a^q}d<#eh8f-}?qSHLUFJHLZ*;
z^v+qNsHs;~e6{mf!&UR7J0FTRzlv5`|7X>3$(-72t?Kk=8VeP<oFYq3eJx#B#0cyC
z5gp@g({zhZ%=S)y={T><xw7%l)8$KS+3_7M9S!Rb9@k3vsFS+5I#N=svgD&z_t@~z
z_Sjs`-Pc<DRpE1bV)w_rEgF-wyyF&U3B!0ho{3$WTBjN}l=8~1hAE9Fn528uY`XX&
z@X;pQ&(W5rihh@zwg@a=;r-^wU{7r6z_AAt|9BMKySC=kagi{@wx#v)^0ewL6JEAe
zjb7h_&)=jCxT$t8v#JWweQ_Y-=f}oAkw!X)o+sWg`Kt4@<cFbB^6YPo=k&EE<3>4~
z-#aZT&mB2^^EF@l+bx~j?@hkU8JXp@hZ46dCbUSpt1Z#9CVeJImI|`Qmm)G$F!Y=r
zq{{SV|7Co%^7XDd6=bzkR##A0oE}Jd*SMlmesZA`lj0+l&FE&Urw3Bb47Q7ZX5|>7
zO7WJ;EH-Pw^gznpwfpUVDyW%KLHwjLTY>FAJ&-abHhfduUIvAtpx>D7R3<P`XNoXW
zj7z#niH})9<chh=2{*;L<m8wloR3>U<bu1&i7+J~OM*{8mV^rdSpri4Spq(QEP)Gv
zEIB-mV5T@CE{_Y5InR8Cls}jDr@8a}l%<nNp<<a-`2?_RDpm+Fo<hYlsq(b(EGm8i
zVj_h~z&epdC14$-PywuiEGmF?0Z(8rPuYrRHAAergL4yV-`tQT>5ch9DT=XVIpN;8
zFO;4bPnHwmO+XqUCLj%X14sj60BIIHi<w$6MV36v8CswgVXM`8Wr!|KYANOfrLh!u
zf&wokAOcGPL`!%}Y{?mu9*FYv`p>%lrS!s}jR8|+uW2tLjmX`ADNpYLxe=CuU<VL+
z4p;$#oq)Ll!75>%K(O~Pr_r!<8swix!;Zn6iemdvxsGQb_yL5>S-QLn5c~wp6$oAl
z^8|vwhdGUgr_&(+JQ{us=2R5lhst##10mpeL;*s;@rVk9fa4Jp2m!|<(r5%6kI17D
za6BRvCE$1<0|9V6P=Ek99;iS791lz&0FDRJXaJ4}@@N2#2U1ZWkLP&cUzDlI_+A7R
z*)nW}WVg}(>AgqMkg46j3(lj`^~S((s22<zhbqM&AOZ|9Lu<TdE&l7sGCGwh;wW*{
z$tW9M{=8Y?4}*T1w2RkJ{4rehYbXITxP}5bhnl-JYlgHo${g!ayb*S+M+ulgJqqN!
z1EmyJalbs4IhLb%9qd?+5-@{u6v%m}yYmC%K+DzA8L(EcgyiW?&V;HwefX2RVz2Ax
z^+&lJf7yHuI2NQ8=|xwbeix9|(yCTdu}1j3tES?O@cC9vC13{CR3PWW89DR1#w>yC
zI1pk%P<w=U5Y!$a0TB=ah<H4nC7ET`t731{y+Pr!8g-y}7hIzblz<s@pg_*&(y@Dz
zPbDp~JhozqP##<HL@1A~1VmsffQZlI&m2qV|4!Q#G+3OCFh%Tj>`?jR4E+s7<l`<R
zHPctel)FePt*)ldCQqsBYwSjX^?PnheQi*i`Q+4?B0JP>hPrMIbgI1^`Bj#085OI9
z8&F2YE8zx|Q3;qq85O{6&9k;4gOG_zk5k!uUEXDQgAKe&aX7;IwNJL5f>;~LmGLK+
zKB=faCsVd?v^e_YBfn_`NM3Ba`I;fkG#G<Qfkk4`WISoRg7HwYFj9)dqvb>d6Ocwk
z5|9Rh0i=OQFr!JoL!v`uA0=_I8{<~WZv17#e=i|<L!kuB+|oopLLKsQp5_nG1B3hn
zgp|M;i4r8g-53)V9U(>cO`bHAuAs<JQemoUsH?fLRyZwUDZ4G7HAi!{wvLsx%`#g%
zdj~Tu<E2KHxc>k9mpxAxGccULVBul{m~gmDEcgOF-dt02XD#;1RX-svUXHF*kN*H|
C%M$1S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.8D291954B92003AF.idx b/.cache/clangd/index/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp.8D291954B92003AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..766c156be9ff64e93ea7387b6739a7c1fbcdd04f
GIT binary patch
literal 3542
zcmZvec|25mAIImMImVupk?3S;an7$M+l+mvgh3?xj7pS3RA`Y@N^w1!Tb3s+o{AFD
za~0`w(Y<78v9wUhrG+e!+eLXS_x8+qUdQ?U@yzSx=e&J>zw`Ngf4}cJ#>2_Mp;?*1
zu-onw6A`>G5MwYHs*<1ReNmw`Oa|keB7+f=Rp-;OyTg2L+jq5Xb(g|Td1sSqOFOQ*
z292Gtp2xLFW8DyNoGku4dL-!=Zt@EoZqvPZ-N+peO_HKZjN_GlW1oujNEj=y<fK%*
zk~K~+8yydd+<)NjK;7od7v222M@tTszv;NW3IDq^&)nvXr*&dk*<_ZQ?;iEOfUc1b
zPx4Or9jV;%kCL{!feiNA$_l^awxp7lCxd4)KMbYu-kMZiYB~JCeoMLLC=hU%5u^2;
zISZaUmnf=5ef(4}Q*KI__H!=Ab*4SD5C)6HE#}{+(-U8Ane61(d<|roJ=ZuoGOQWV
z!xp*9Xp}q5PwTH59jLPZHHF=7EPwD>^vG_7zFimQq~^*YuGPB+w>Hgdm#b8X(Ajyo
zH1_MKukuTSI<xu8-27bebyj-G&`5P)gI<{2&uZQ>nu_fg5r=Kv8y`K~A$}|{SaL-%
z;aTY-ETpn<^E4~zV55<TI8JZf+k*Z$4c9oU`>w82#wYz$R(Wz7^q0DHoVeRO9QKOU
zANdC-aphVLzps1Cgm$c~|A14NN4WP&=YV@8&dk2J_T6uGYeKl!{nEL$vg1Q>b^RZH
zs$R2aV#lJib#jaIH>cMyyhIJi=l$Z=$)h-`me!d6MdRt7g~{A>y&LZx?ZT@*9uCX5
z;IHQd@3Z{;@9e0{qbd!KbMfmwpYINhJN<q?v-RcfTNR~yS5&$FJQ#ZBo{r+l{r+nG
z*~-PW&z`cDUwkYSbvnkK?!AVq4%Y9j8++`Zq4m1Hl|u?1RCzS|Dg0+oich4x#V7Z~
z6?>meHg~1=*^ha}wtLCvl*U~cK61sN^U>nmf;9a^!RDA3CXeG4fA!nkSkZUl3hDiz
zaeGGVuG^+=5wE$$Nu^#|Ke;X6nSUcD_U!Q|SCiR?0|m`lyZHXRi5Hd1Hp<7luVo#s
zJvT8H5@q;4_3?z2>!B(Ago>d<qv2Vc>O$+Um7_Jh*W=kbS~)$8-ryI#%U+hg3!4gc
zwEaWxRoSvrK^b-bGF_YO9@%a<{LA^pik1U!JFj*9AZG6$*HYCl2rbS^F&KAi`!e3M
zH|lKBqh2=8Ie65iyn5NVbIga&TBz*zy8QK<zMV5#Rn%3av0U7}L0EHGe){aqrto{(
z9qqRoE;rR5DG`2KG3<P|--*9>%D6|dxn^a{m5HQ_g-z+jp9*g3zgplO;=aG?_Y3;$
z%=+(x|14|KZgXoIP1USln8U>ePUYWCRk9zx6WH^Y!cbSA!wGF~+ud(vbEjJ;JKCqu
zCKKByx4y2zuGt#3#2g3?Kr~nJoT3j!1YMFm(JGF1qe~oHi<B7~JQyNXb=9RagG8D6
zENW^!`C3@PRYW}{vN}uMb7r7q+o?Z4=l32eQbgS)vI<K@c4naD)w?zgBs`QaS46!e
zvMx*4WoDq{(;XYtV`AejD={`o<W(}O+-3$!W<&+=h(0FAV6e2yezeOr(@{|p;T+tM
z$|NBkZVr(EccF>k5Qdb$l8E3D<`4-87n(>83C$#VBs7y0kkCxP0nkjq1JFz$0MJYW
zu7TNXc_KruAw<Sp<5^OeyxNuE!t+#+6cdl&DF|HvE=KV45EJnTk%G{%2{A&PhnS2<
zNLVMu2np*T9s#fp#0Y?OGp?B}9od{~K1)=O$SBP8hPY<x>aF)yI+Dl*2wnz9EI^1d
zI8p&Z!VC%!Kr>Y;sxZ@2^lh!@TW_V3ES(d=XWy^vk&(Dr;&#*;$%$GLcJyYDmLx>L
z5<q0cHR4mZvocCyO|u2dq_@jO@O(IYE<)tP;d2oZW{`^jm@T*#v&Tq12e`sa|I?-5
zi71|g(y8%G6h9Ah9*XC~Jc8n*FefnalT2uRIup-;`8bB3z-XOFM2RGn-WMVhCC<Z~
zhZ6ZPkD$aT%n3~5BokVn&LlEmK8_J5Fj^-QQ4;neGf@)uBlA!a_9I77680k#m?Z2+
zrZY*{k35c%updZ70qh4dQ2_gaJQTowU<3uQA4p&V*bk&L0qh5kV}SOf>XdY-^yNh#
ztId2HKdPnHBn>ep#`e8IF%jIjm<XP<{!$Oe=4)mlCh0ploPZ8Yjvy!-B@rn{LImW%
zEP1)j)9+8UmO+$62J!|fl=;>VKkQ_^)#+kN9*Am;cnTj)HAcb=sxd%wkg-$6{J{HV
z(uH?0q6}Vm2P0tycQ8P6m{rV?tF`Oeq)$);eh{i8f;b3O5<x-)A^=1@F3*N4AQEF-
zxM?``RGK3PBZ}Z0IT#5u$iV>32VB-QuJAK;mu!fty{a$ujyfio+>&N!;4R%J9Rm}}
zCSj|lL!pt$0qJ|z4fmi^<FWWrng}<-gRX&yC63ZWx{;7Z#*&Z*+yJD3SO94zToZFD
zhLXsXYdTA(6VGV1elZ2ot@Oo(Pz+x}2pRj55CLBR5tqv~rEX6BMigfLjZ)nz?#|l4
zIVKF2?(lPrXoYw9IYz<^o@0RK!oh??!uMG}qDmrlfAy_&JY7#+KPuj>m<k#9h#UvW
zeyhw;QKw#nHC{FY0otz`Bx|iOD?KWjl8BG-XX#yc{OVNI+a1y)6sm5cj_{+Q%$Tgo
zf)wcyvJJGUwSUd^Del};+bbv8tpGfTj)RBd;WQBeL=Y8DQgJ8|P7@hGLK+!LLK+AF
zkOo4*tS0#$5*{f1QW71tE!t6f;Vm8h_wl7J7!04;OQs0__D!#NNuQ&PpU~UKmm!>$
z7=HBSwutTFArf@i;mX8hSh8~Rib~3JRpvOU+S@HxRd8B6Z^3*ujfLhGR@QtQ+tsVs
yD=e3o659Xouj*Qha2;K}rOWh5u+qTL$b@IMn5)S#UdL8<boNIb+|~$?_5TGNPlE&i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.9FCCE6CADB35B88C.idx b/.cache/clangd/index/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp.9FCCE6CADB35B88C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88403271dc00401690a38970e2026606530beb80
GIT binary patch
literal 3540
zcmZve2UHW;8po5#Pz<32B(CsAij<j)kN^q2h$w^w1yNcMQ3L@6d=!=9ql*P;B1`o>
z1z7<V7qH+WxQZP`ii+SvMNn7WRhKG)x(VkDbKf~H=Wx&b_}}mTzwf(qXU5aT$?2Xv
zjpi8a5+4~7w;83;Xl(IcY+OwE6$Xu#DMzElSG0QjuX}2xH2M8)gqy+<N7Hsj*NG63
zW^c!*nzbj+x2pKX=vwBFy>A(<ULK}V7GqWTWFX+}Q}-lSlT_Aks(Yh7Gsf$z46|p<
zWlS<IF4DL@`{j?>;SCkspKn<oEYu$jy419}!DcqpIeO2zs(5B}!dG3Q{(0wvD{aF!
z`X0nK%*Hh-g?4oG5(b&^DVG*@<b?)u%>Pq=y<c^E^49p2=x%kRvAd5-d>pnd4l6YO
zpwOD{T^4yP-@wP*=ys7qgVT=gkyAe%>bwJbxBYi`_M5)^nakr(Wird#Sy9!-vm=9o
zf!hY>%9y;Ey?Q#`NpE_SGLqE|8bg#nRJ|D)<aG9#VA`&1q%6Jf@zsO6jqG@hf#sc(
z+a}&UDpwI+|JBi2EzHAJ-9=t6wEtCAd+w@-N+0k#CZE$NU>v%6v+8iceYK*3jZSup
zq^q`D%67&3w0veLd@S8^LT!5G{x+}fX<pgqU3<#Qc7@0`EZOk0DEEQzrwdZyM+YC*
z%Xbai`Lg^nMTwoaGDhvk70h}HAD(zp=X<0rVx5)r3c=pr()Oiyg$Wz<!uKnFcrtyk
z`$l9&#<S7;7CHx%Y&=`^_R?$4<uV4|tJO}lTcNm4>DnRVW6ycY;d-mw+VirWBz8<a
z^oUI?5#9YgZouF1mCYNC+&oPy;hBN({*O(~UzgRM3Gm+?I<zLSJgTpENu&cu_S6Nu
z%TYb}K-F=1H-1smxnJ`>|8QU5TKjR%(80GwyYaqX4%GU4=4T~UmozN=<e%*wy}<IF
zN0z}qyT5ckD|oYNe0^g7`UREe_mqD4n$I1)v!uH2uzr>>ApWW8-Bh_Nz8g2jJ+QQ2
zSC>_K2e0kh6jC8-3J+E^sC~TQs9UGU>c7S{8O07N29wwFD=hmGFKin<nV`N$4Y}oP
zIkn$#>}~hRrfW-!#&(yDPVSJN>Q5evs4%>G%4Xv7*fqhEspFdJl_Rv*A-%74AD@4*
zZ6?zB%17<KX05$JM{cy5tvKKj-Or0Mw&le2wDw%wn0;P#=hWgMy}Izm3Ip!cnn#}!
zV=OFh`vi|2RI6^xHn?<C%KXRp*Y9Ky=WikLZW|_Z>Gk#Y%N8${fBC=<nc4Cwzjh+x
zf_hK?X_<fDzxl5wzY7=7DjOzx=Iv>t{h7YhzGKWb#Ub{5vbS@<^!U9mfGJ;BpI|+9
z;$_U|lvZW)$->0i1WnI&3&;IgvaL!*W`~d8UM;xSY@aqbol@|^`O(>#iM3Mp7r*T^
zyArI;Ed8bx5t!Z`sIkfBbmPaP4Mrif*o4TSa`Aaqc-i<p<c$p`hphLciL|7(gmZ&f
znfuw#{Pkze>x8p%h=*8Klvb3P8%VjT|KrrXWlir{w6$W{NXp1{ZXo5;VfDHVHYqH5
zn!8wLbJ#j_11XPh|LZ$fcq2~^Stph`(j1Sufs|=6AzNcpWN9>Mjpkg(<K~)*ED>gi
z@yIYS2`~$YgqSNuxFOCXi(`p!0d4`25O<}BFeIRvgn)o%5<&u+2@C-=69@n_69@q`
zlgsCt&o@WJ<MSXg;hW48^UJYk8LonLOmQ)(bS#@r6@cZ?u@Z>!R63qbr_#oA=y(al
zL@J$tbs~pOz&c2!16T(+bO7t-d~<s$WedK=JdvIVbMD^Axi)j`^pLeCOJt{F=S7CF
z&mK_QS^utGQi(g}3DuE=B~ygE<DTSt;%!O7lPMzH2}mQ72uK6&0MbAbfHYIS>3l<3
zA~U|(Jn1Ne*{Zf(9+46!WraCHX;|Tokg*j35wHRf8S{;8$r`PVv!!QMg}RdEs_9q_
zygSu&yawKlYB~WksHOv$E%}!7S(1kgW6m8es=rtkg6%`7tg&JQD}nhWg4Mu0ieO_f
zXE3l#1{7b&z>dJ2hGOX`rQ=x$z7L`Lf)^ur3Ct%Eyawh`1RsMrgMnu<p!h-tegx(;
z6i-JfoybB6*pDbi2-uG}i4d?KF^UkdACbWzU_YXeLBM`Q8cM)^APWJoA1Fou><3OF
z0QLi;2!Q=S1_Qu;ppXGzKahq3%FkinAGGQ5xB&ze*{IkAkT((fKgY(ce8kXhqWBal
zzNKU_9NG(69EUn3OF#r<!91;BqpbgXfsGWMCE_mNDv~anUhWT-ey;h9AwFbRQM?Q8
z`c;&G8C*pH#X%-6Eh?Mank9h^DBcVQHlPH|paBIGN29dTTHZ{TB(M_2&%%L~C;>C5
zL;=OSTvy&R@HO)gSHNDy_90&yDZ(!s5A(QQlEO5(retA<vhvg)eli1+Gw6f)K@Ic4
z{h)^V5D)<$01=<hHzSwJ8Ln_QOz{hmWZi?}58<qPPy%Mqg93`l1?bGVR~~t2t242?
zjh~OUO45tMc0=h!;k%*qq6mmU6o5#;7tB9#PQA}a@EfYgMOY$^FDHOnPRm2fi(HQE
zm=tsF^-889&Y+HtHNy|7Iy&AAKcwpD1k9k04k$hy-y-D^S?MI+07V5w4*8V2w%+b-
zphn*<aZ1*xwKmxYmdF;hou?PWsmlx4zdV)fsULd_nc*v*)VPe|qHM{Ys&X~RSh6hf
z3r6x4KotO=o3BdVje|R`zm^r(6^I3qF=7hGA}GQG@gTA>;yn$=BPb#Q2}mQt2}lEh
z0MbA>nAgPLArYG;A0@FdTVkCh7haNK?#B=<qFJXd*wC_C%Gb}!+lS^uDm0omb?HZ4
z`cju$B7-Bs#OS(pH3OBBmXTc`$C6i4RB&Oha@1opT~;VBQd#)JVhc-a8(Ta3<wmLo
xR!hxrjsN#wj`}|^O)YI5U3~&9<MND61?EfmYKA5&RXNT-`Xf$j9E5bYzX5i-)W`q;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.479D1366D0F3CE30.idx b/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.479D1366D0F3CE30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d0bf400cb16767b0e9585be954561b788e3245e0
GIT binary patch
literal 10366
zcmZvidt4Ot-pBdP?BWWjdBIEs0YwyViF3bSkZ>stZ&8|?w^cHWgiPt_s2OOwiK%RL
zQnWf|KG@}GN3n|?GfK-!vC9b$rj;gAC}pS1T0Fx(&$ypI9$qhh&U-%J@8{#r%>I~}
zpOePt=Pwx=7iZ5MzjV>O@>!m^xVWCqpR)4BCAtGMyT!#V-SW}2=@V1LKL6SF<<1Yr
zkno~G?Qt2;p6J_)Bv%i5zbyFP+R>8(y&3~=zh738{^}cd4E@Z1H1x0**>dUM-)C&@
z<zG}Z^uw|bw!S#w!>`YtB-vMHU!Bh#(BzKXqaW`)yh%Oue)dl-v+p}pet&s=wjiGF
ztxO&B#=*izTGt)@?Tf+z|Mk|^x$B<!jghAxo3MK5kv*U69k=h-#Y_0DU(J{||F)sa
zy8f!9?f++J>NL7e`Q*Tp3#XmxRv$`vdEd;6A-i9G>AOPjmkC=)L%8U_w~QY&<=ENe
zNZ{?sM|XU4?;9s?)HAQt4zHXz<E2BM0V|KrGm`_4{8(~v%L6TM#qaM{!_51u;#qz4
zjnTVLDvLJe$6qv)1ui`5{vo$)ukRju^7Gebx1T+{vzqC-ko#f$f^H&juI;`j{uZt2
zp8FHTpHptg8&@`T>6ER19dLJ-N4h*(cHLSzt#sQZBV~;5wI$O|UA9w|^!L-*Gu;Qy
z=yOQ=%D#|V@%*_zxBR|c-n1)xc<ja<k8Ms_xUVUH;`F!Vyz%suvem!+tLxjV_LhIN
z_1gn|)+B6MeA|1Q*RS6YK0W75NrrijX}Du?*8Sh4tsfPivX3oaysCCqL9dM83+`#{
z6Q5H!XUz7OcAsB%Fy*0%iKC{DE<JPTrJ_y8p4gn%|K%=LdHf&8cD;8c{U4i8oPBlN
z!}C(lZ>t?}@~)D*Kg(acWLv`zJN}lxH+RXL3sv%_Hy4)t^3-?U<MT!&lrj}FtsZ&R
zd6UJmP5XOZdj7Y$<KpfdzWb?Z_2l4lts7p?^6vU+ctg_sw1eYrqzaFGkYD&6)%T@^
zm#dWSk8IlU%>y-4@>YG!T)Ok%kHNMT>vC?ll6NIunx6mY?6cRe`?7x7gv(D)`fTgd
zFJ}GR|IVD@(|kXF-?DLC^0e%YlP7&xT4B7H^;-AjJItD=#FW(M4wUa4QSoH=m7C^N
zsk8l@k~SdeazVq@tuMU!hk1cs*1YWVZ6i;24Og6g{n+Qn|1)dI<2?q@M?R4D^m^^%
z@Wx0|;i$t@i28HyiTzK{5qIxc@%^wDRz7oV`MRe*4gR{J?&e<;3X^{5dw5fG$Awpa
zd!>9|?$f7l7_jE`AxW*9#o^2!50^5+U4<pZSqn3c{JQp)yVfo`7Z`GU<EfIvTi(h(
zxiV``-I2-}o63Kh@ON|ciDTowShe?sFAj%CTo&r<wmd#>+`awRjvZUHLArZpWJ0;O
zV&;S2Y%NbaytiiEnX5bY6uvASTJ-6+@7;T3<bg#qGLJU!zyERHpAThDd;Y{<&V}+{
z`trn!0lxne56+)-*SJu>ht541*L(QYZaKT(+&{djN5M&RQCw60kL;TBBNm<s+%@%U
z@t5l6AB7*^UNLd?u!H{?^X5)<LF9$y^HW-8?f*9W_Vs4YJ@<e9?jBK|p7{A&4gWm-
zU@-U7b$|Z3`!Anf`Q*g5L+6(-oq8pW<?|c-8JF#&IXzZ(@6q1<@6R6Cv;5um{fEo0
zj(OKJ^UDkU4y~)a>K&BZR1sIUY|-pdR9syA!20#}vwV6YHRG<hlI!EI?;0JPnI`(<
zyZO5*A8PcBq_}`1-{ifiBs#cq!tdKpuPS6GCsI=#IW<1DXLNAop07OC|ACh;9!{i+
z9JzOV@2=6om2ddl?0;P>o|=|O6+804_<=p5gDdNi_8(q6zwDJn>P|;aiBIVg9bEa=
z4{EFtlb^{>q;7ZQtoW=OqJt~X{^-*MeSY4%Hj%o+k(1+-`$h*>p8u}zIM?r;<B8NP
zM;;kJG8i3P`N1=9Ha_~wv8NKLTOBzgKBH@NaOL%en32?S^QYb7COdLY-<*u-;L4lc
zfBtLV*w21Rq^3FY;P}D)qJt}cHgn@+U!Ho(pGZw{<jnZY?$P0z{P`nl7yhD}NX>NQ
zLGgo<qJt}6oR$9iNalyJiPS_#?ib&$YjkkspTEnz=}3<cHzrck9XTyNtw(fl<*R?Z
zGq3FT!?O~p+Z;JNKD%dhaOFR)o3^L<zIV?jQj;9HPkf)Q(ZQ7$&kL4aNq5NKZ1-PM
z{?T<wC5(?5;U3JS64u88aXt=+_whi2PXH2qB9P>hfMlNxr1%sd)u#e!J`G6s=|G0h
z05W|hkma+0Y@dCN+^<JQLiB6#_&bs9yU)t$&c{|5J@&DgJm#htu?h<so5zC2DI92A
z9tRq)@SyQ|JZOR<fF|S#poxkInwTeoCMgnVQl12wtjM6rc`|5<qJXC4DWIu}3Yway
zf~F}NXj+~Iny%=e>3KS6hGKwb<QbruiV2#TXM$!a7HC$U1)8ncpxJr$HO-yYNJ4Zv
zb5E-~otX*IY0Ta2ob6h|>>{x<%+wKPA96im4kI@Z<~T50OIYk@>j;beY&~JIpKTy4
z_H(s_!+x%gaM;h)6At^i2Et)KUrTuG=j#ZM{d_&)v7c`sJoXE<M8JNbjtJN<)Dr>w
zg$5#EzgSB|>=)~Zi2Y(c5wTxvAR_ilwM4>xsg6k4FVzzX`=tgVVZU5UWbBvgh>ZPm
zJ(00rZXh!DE44(yex;5m*ss(R1^bl-qF}#TOH}Mv>xhc|YCTc0Uu_^N_G`67!+x!f
zXxOjS6Ak;d2BKlVUQ2ZB*XxLm{dzsov0rZ>I`$j2#K3-|ju_Z))Dr{yjRs<1zgbI6
z>^JL(iT!3hF|prlASU)(wZy`HtBzRMZ`BhE`>h6IVZU8VZ0xt|h>iVrJ+ZOhZXkB7
z-@T0~N{HURz_j3JZh=Xmh?xRZ%uELgFtdSzOi7GbKMO8w5eqJCF$*qifCU#e$bt*!
z=fH(4;=qL~=D>vuaNxoPIdI|qJh<>hJh<@1Jh<=y9$fe!4=#dV02iT102iTH02d)3
zfQt|mz(w?n;35`@;35`_;35V@a1nzdxJZ5pT%;liT%=+NT%>>mE>che7uhd^i(Djw
zi(D*&iyV-_MGnf~qWBeXQHm6BQHm9CQ349MC_x2WRKE%?YLN;qYOx9~YCr`SHK>A%
z=GVYQE7HJ4E7rh83uxe?1vPNd{W`elMLM|X#X7j?0UccQpbjpE-vAe*$N(3k*Z>zJ
zV1SDeG{D95o8V#=nc!j;o8V#wOmH!SCb(FB3tX%s3tX&X3tX&#1uj<50vFqFgNt2c
zgNt2kgNq%o!Nm^R*IeAkqh1No$0C?^8VPd>?=Bh%i_A3=4w-KxJhIS81Z1(1h{#eS
zk&xv^A|oq}L_t;?iHfW>5)D~zBs#LuNDO4Nk(kI<Be9U}Mq(ouWEb4%-oXh=*W9~#
zMb2}?c}Ow_6@K=A&ra^M<4Wfauk*<f4qh9@T%uyLFVjIWD`?<qn%RKdL^G$5TWRKd
z<Ot1NMQ-sh7d_w?_A*a;W6ZWuEatNv6pQ(6HO*o^+eEXN&$iMm=Ccu+#eBBK!(u)g
z_Oh7IwNV`Aa~%|i`CK*4VLsPHbC}Py(j4Y<5t_q%uEoP)J{R_Kn9sLSJm&Kq6p#6Q
zHO*r_-$e76&$rS%=JOGn$9%rU!(%=l_VSo7v{3@)3muez`9d`<V7|~q3z#po(gNlS
z5n8}}p~WL$z7X~bm@l?bBIb)7l!*CaH7#Pk*hGt%FSgPm=8F+p#C);EBVxW7_KKJ<
zwNVo0OC6Mi`BF74VZPKvOPDXU(h}xN5n950sl_8<z7+OKm@l_cGUm%2l#KauH7#Sl
z+(gTmFSpV%=F1UU#(cTOBV)cC_R5&Av{4G?D;<=A`ARjdV7}5sE10je(hBA)5n91~
zrNyIQz7qB-n6I``D(0&ll#2OkHLYU4+C;0EueQ=E=Bp7}#eB8Jqhh`q_Nth#wNV=8
zYaNt^`C2utVZPQxYnZRK(i-M#5n981t;M5Zz83asn6I}{I_B#gl#cm&HLYX5-bCw|
zueZ`V=Iaq!$9%oTqhr1v_Uf2#v{44;8y%E^`9?KuV7}2r8<=mj(gx-m5!%3fqs3!j
zz7h5sm~XaGCgz(Rl!^IfHEm+P*+iR|Z?@7V=9>}P#C)^GV`9D;_L`V)wNV!4TOE{z
z`BpV;VZPNwTbOUP(iY}h5!%9htHonsz7_Ubm~XdHHs;$Ml#TgzHEm<Q-9+1%Z@1Dm
z=Gzh4#(cZQV`IJ@_S&(0bM^Oezg;PBBA$|)eK#*~?;`S<_Dz}3_uT5-MPwKmh3?y5
z|I=?*f4u7ZZl0u)<kaL$_my%|$-sUCQ{CIMq>{nugR|UMzDXrn=~-FsyXjl_jra8(
z8QMT4m81%(w)=i5sU*|LEOcK3yTv^dqA#%FR;8UVkHW1A+fG<yuAOkmd^_Qhg?1t!
zi|s^2mfDGgEVmOGS!pK<vf55mWUZZO$a*``k&Sj@Ae-&PM7G+Ag>1JIJH}Hc^vb^Q
zgyMD1NkP|w9`31fPl9u_I$!F;(Fze}3*Hrn2#d^x2#3sv2#+j;h=44Hh=?qOh=eSM
zh>Wa+h=Qz!h>EO*h=#0(h>mQ8h=FW|h>2{4h=pv2h#h1129rvNu4BXw3tr|+3iFN)
zFXnD_4rz{+GsRuSa4&Z{A-c#B>&+UHHO##_oeO(-)+m=&K3xCkp)+R>bUIDTxzlx8
z^`ww`jQnM}bG>KZo;}-LY)C2@di&68RN1s|%gz-q@9I3U*A)zMPweW;;|7SIg!((X
zd#Cmu<i2!SbIzA|YR`yoo!!~}vPIX@O>{1U&gGZbznQ6_nC<wbsi9b8u7=`}`5KBx
z7HTK~S*)Q%WT}RdkmVXmMpkMl1zD}3RAjA&(vbBUN=G(oC<EE7p-g0}hO&_D8p@8b
zI}JI9rE}Q{4ofp(F5$PMnXt%QGvSc=X2K&2%|t*Jn~8`lH4_P0ZYDCa(o7U&wV9~M
zS~Jm*^=6_Y8_mQ(Hk*lwY&8=L*={CwjO$ugExY~fmi5j>G^*>U?(Uam_tH1Kw=Q}v
z-`SngJ*BVv(D>>}I(>c%ebL!nkXyjH7wAD|(p>Mj=UScJ*JoWn(p?&QXNuMm82H8e
z&ThZQpXpY+v%B+!>|EiCor%Fi29I)2ll!>p6m_oiLD7waIk)I5cE;B$_nxDifVA|q
zLGA*dyAJ4_>dr;LSTB>yu<j>xR^)P7ATE~!;&XW*Ay)trb44I2R|1lAWgsP20a9~S
zAT3t|(sOknBi8^jb4?&C*8;M0?Q7KP+g(jbuWoRTM!&3nH@Zi`Ju1!|*16^xI~oJi
z24=^8RSz6CFxRC><b}6K|2ur8bF$L1Z*ylBXR^|UUL$v}bCf$5OJlxF6~#P(^Gp@R
zB6C#~hs;+|JhD(l3CLm<B_c~zl!Po-Q8KbpMJdQ?6{R9;Rg{LTS5Z2$QAHWZW))>3
zTUC^WY*$fsjNNtOu?f-j;n-nG9+W)9{T1!pH4RH15hJFQDT`g0Oa=2WP$gm3;QUm|
zLM>atLM>ZKSghqrIjH3-IH=_+35T_ODG#-L1rN1+CE>AFC>5Yqs1Tr5s3Zc`ilrjd
ziWMT%ij_pfTB%fmTB$;UTB(vqSSy#xP%Br+P%Bpw8EchN1!|QF1!|Q_qF}9BszR+=
zp+c=%NmQ)WN;Rm}Dm19oDv5@*dZ`YzdW8<PdL_}Z)+jZg)~GO`)~F-~)|#az)S4A0
z)S8vV#9FJ=f?BJ>f?BJRSXgV9+E8m(*idU%5<6DAc-}p;I{zO;+2T263!T5|lBI!O
z;pO8-AM)Kjd;YvRbAt=w=5!YP@6UX9BQ9=UY;$gGb9QVq7~5PB+q^rrIcL$_dpecE
z;4QSr8{fs(H8H8%bv?R|?>WvM(lcSa)$97+efr%Xh_a$;x-lwwxHO0xI`IG7+c#xE
nYI;WI;H>OnIU`2jIA$!%r!oCU=H?~$y~)3j%Aa5skemMpq<*)J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.D6BBA9767E8DA592.idx b/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.D6BBA9767E8DA592.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68a19236c4ba8f6d903a5e601a202ab8364112a5
GIT binary patch
literal 10366
zcmZvidt6lIzQ%dinn47-hG+@|qTv0)towZ}5fgSAUI^11ucKrZ37OKVGy_d1F_p=7
zie}r)gIZhLid}S@ky>^U)3&%VtuzrsDcxzKau##Wa-KhqpHHvzo#*#FkFyrDSeM3(
z%*vYICn93ljFAiG&Mcnliin8lWdB!GJg-2vVRE~Ohz0Av%9}X4hZy&tji<MMF_Z-6
zrZz{UymdUjGwD*Edbr5<`SQVIy`5{l2M!k%B)|XBqkT?hAN4=41=nBw_mz~jowMiW
z_W82tiw*CN`trx~CrJABX}4!_`!u;F<LK9053NxT9!~$gVcJs%i=Qs8P8Y<p_bKCt
zezZU5rN$LUZ+)MW@L!*7n6cu`TZ}yM*r=s_j_mqo_wYS8=gsFg{4i<etVjARY<p8l
z+WSxc_&mBw`DWkibMnr#tM>QUxo2`o+V-70e#vp4j$Th{0=fUaeq`#nW9PdBy$8k~
z-SX3uADtMY_q<-&zjXAZ9S2<ri;vDUyLeyvt>E(dXBs|<+}o~#nfX`6we;vmgSVeh
z=C00)yli$6xWJ&N(;nJX-QNHDsSl<#pFgy<oar=&`*q~(b|P;sZ@(+@A+7F-r=!Ii
zJ?_sOUess7xD9_zc)ZO^ZC)<wxLi&u+<4XKF*NFf`FUrq4NFv#4=1zd+CMNU?x6I;
zuuF+0Z(q2({?ASFnr(qYp4D4kS=(dIp1Q2j6F-qNN7CbpmfreT+XG8>7k{<k=Y4U@
zqSw!R<ny&FSFQ@2oqnz$#k|1OJUXxU(?2Dx92D7O4_iEMN#)cBJE#2l;1i8;k^OR}
z58bq5`^AO(dptKfX3+S-h35|L$X#>nwY8buceb&LBkvsB_WAYXf2=)z{{7+4&rH0y
zu`=Prl!C`kW-Xt;vF6t;f6LmPF@O4{GI`C%a|-@g_lx`6nFFE=nUcv?hs^TKv0~Ah
zy`8SUeQU<>h{yVGUzb-+_HS-n^<i)Kw%_~L#Li0EKYR$4bL5MxoL{K;9doXgDeYfc
zv*o8}D#m3l`I@=<*#6&qO^a6an_zX>7ISrC*2~k*cU^J1df}*RZ;Uy);f;5D-{}5W
zzy5hqH?B0SUeP5lef8KeUlx`a@Am$neV0egigjX-#LfGPw+<+Iz5U`f)63Lp*_@J;
z5PR*xn%f)R`S^}`iC$R0@YKeEXWIr!&VG38)VKecn)Ygk1p3Gq(yq=Qd>vRFjLjKz
zi1Jf+cOT#T#&mJ}mPJ?kzO(qvV=t~)_nq(Ns;a--jLwPuHU7|=`j$)Y-+Hgum-oim
z`xBOZm=@c(R_xE*dA^VlrsNdl_nwn-<mU4CrYxU(!JGE&>N5q0)_;<IVsY>3RYyuE
ztttL})Zfj)$BzyFe#!24zCRQga80PLTL0?I;ZJs7?(yWVk{+KN998Ttnf&Ze8;X+-
z?XFmH?)H{lIXk6;bHDrf^CyoC+&6bp&!aW`pLd?R`&`eww~xPY!JoC`^znDSeD~L$
zoi%34aDTVwE<77?U;o?f`fdMsZ~wXu51ugRM$~2f#xA=!V9q)3l<_}`f0Wn%Cj53_
z(deap_y1$)$6M9e!FOJq)uUnR-k;MatTg*Q@${+Bc8T)Dm{XtB{PXOyzKriy+`YT~
zh3~F^bA02$i!Ux1e?5uivud(at_?fdufyW@9h%$!{jGhwUi_?i@1dgGLqBs(K7Fa%
z!4;*q-Kl0>Nkq}Yxzh$w5fPCORIeQN7M~nLO_~x>5FZ)eHniABn$Vx<KR2CSlEaRT
zp~l;CLS#ax(BjCQ-g~9{GdnLIilK6CxpQRawxPw5@BiDhe_hTWpA<vo+j3H5Qissu
z$SLCwK63dc_iYR{)s_cD4)BE*N7jSQzrK5M;d?REW47EqvU{7*;>ZtwQDF@j`(}C!
zHNlp9MfSQsv^es#ufCfdcVqYR80t}5j*E<o4=s-T?75F?Uw-e{x)|zVTTYHlZW~%0
zdClRse~j{+{3C|Sv*pyt)NY~0k+)u!e6L>VzalomYs*>gtb)+u$R{VSe&zI;b=fi0
zI9pDMOlcol?#ZWKQrqyC)fj5BE%%J<85>$0`SR4{4+k>8dSa;2w%j$cYunJ`$jzg^
zywmag=H2Zg#@ce<_`WHj#gT9P(sRU-4qvX0p(fgLVq{{6(BjCq?|eG5=+8q_W2i@L
zIW01+Q)qGIJ00_O)j##w#TaUgE#DV;U)#{)$n$3UimoTylwCip`~2dsI>r`wqC5kf
z9T{7|L@_{Y6br;faX@?&4<tkhKw^{#Bt=O;a+D0DL@7XOlnSIpX+V0E4rD|bKxUK)
zWJOu`$oYD7iVl4(Ui~z<Y5RFO+5Xs>o)Pe|GZ`Q@lLg{3IUqih2NE&`ATd(}k}@SA
zIa3BwG8G^-Qw7p8H6T4x2Qo4ZAT!ehvNElE<b3VT=+M_4K6WMXY$M?h?x`Z4J;>F>
za|pSHc)kT@DhY#rriw7=XQ~N<ex`;n=w~Yli+;9>u;^#235$NVhOp@8DhY>vu8MHz
z=c);Zey)aa=;td5kAA+2@aX5O36FlhhVbYYDv5x8p^6CT7pjSXexZg4=oc%Ah<>q(
zi0BuqiHLr&hKT5wDv5-CsftMGm#T?`eyN5?=$9*rjDER_$mo}=iHv@^hREnwDv5%A
zrHUx%SE`AEex-&e=vOO=ihi|<sOVR#iHd%;hN$S*Dv5@Et%_*q*Q$wzeyxUR=+`TW
zj()w0=;+t0iH?4~hUn-wDv5!9qly^lH>!z&exrsM=r=2giGH(+nCLgFiHUx*hM4HL
zDv5=DtBP3Yx2lPSeyfI9VZZb1J~uk_YaR{@*`9~sppfes2bAxb2;}ul1M+za!o*}V
zkiz6Lkiz6MkivKwNMU>oq_Ej6q_DXxq_Fucq_AEVQdl1gDO@%ODO@fGDO^4WDV&#s
z6wb#%3ZKnG3ZKhE3ZKtI3h(71h4=B0B4i7YBIF8?BIFB@B6tNz5qtuqh}j~fh`A!9
zi1{L<h+Yv=M4t#LQnmysQmzCkQoaN!l2?Kh$tOXIoGn9&oGU|$oG(L)?3E!!_Q{Z<
zWGj%O<SLM&<SUS(coj%dd<vwf*(#){xhkZn`6{HSUKLVQp9(2jwgxF$t_CSuz6L3p
zSA!JIr$LIItwV~Qt3!&OuS1IN)geXq>5yV%8<1k;8jxb-8<1jn4M;J32BetTCZw3T
zCZw48CZw2N6H-i{2`N^#1u0gp1u0g(1u2%-f)vYV-Ai#UBRWTiE+62qQ%gK&@a&?N
zFvx5zVUf97!Xfjughv)?iGVED5)oOdB@(h+OJroFmMF+-Em4uRTB0HAwM0iYYKeht
z))EuhswEcki1ZOdoJ$esFiJ&-E~3I4DJLptwlgFwSoY-FMg2At`-;QpnRD|0E<~Iw
zt;Hldbcq!<dzvWERm#b>PkUP^&mtPQoc63juA@C?kQ-^w733i8xsBZ5@?3U7Ucl{n
z-5qA8iDIyyX`vYGXUb^?`<XhL!G5NZX0V?L(hT-94K4=znSh(Yezu8Xv7c?BSnOxZ
zX%_q0I-13PwvlGBpAFJ1_OlHx7W>(No5g;viQ=%IYoR#o=gMgg`?)%r!+x%j=CGd&
z(j4}44K5D*xqzF)e!hv~v7c|Dc<krPX&(FeI-19RzLDm!pAXVJ_VW!c9{c%#o5y~k
zi4w41XrTn`7s_b?`-M7Mz<!~T7O-Ck(gOAi4K4xug@9YYezA!Xv0rSVMC=#KX%YLy
zI$Fejv5^+BUkuVB_KOWJ5&Ok}Tf}~;iIT8iYM~_Tm&$1g`=vTs!hWfdmatz6(h~Md
z4K4}$rGQ()ez}Q~v0rYXWbBvAX&L+FI$FklxsjH!Uk=hT_R9?}8T;jcTgHB+iBhm%
zX`vMCSITJx`;|Ic!G5KYR<K_Q(hBw~4K4-ym4I8pezl2Gv0rVWRP0yFX%+j`I$Fhk
zwUJh_Uk%bK_Nxsp75mkITg85@iPErNYoRpk*UD)P`?We+!+x!i*05g-(i-+_4K5A)
zwSZg0e!Yp(v0rbYbnMs5X&w9ZI$Fnmy^+?jUk}nc_UjEU9sBiwTgQH*i88R?XrT=3
zH_B-P`;9u<z<#5VHn86a(gyY$4K4%wjey(0ezS=(vEOW=OzbzyX%qX+I@-j3vynEj
z-we_w_L~hZ6Z_48+r)mWiL$WYYN0Ibx5{Y?`>i_K!hWlfwy@s{(iZkx4K54&t$^DK
z_nS+vMBKVwTt{35covbxG_UFTcBc*YSwymtoa5X|&3@xR`PWOXv~$H4#3jb{bZ)@L
z79@2`N_5WQVhd7}Q+qqNS7Qr$CHLy>+{t~2U;R}4K>sQ#wje=B80OrBiY-VnQgWPo
zs?M-|N3QjrFPy406VJ<Vs=_oA2AORpEHc+jIAp$=@W?_l5s<}ZA|gx8L_(IEiHxi?
z69rjqCMvSlOf+P@ndrzyGcl0OW?~{+&BO}xxKW+cFTJL??R_$$?T8M}37@kQ?5)*$
z3mmqVpLo{esk5Ij$gH2R$ef>W$h@EM$bz2;$fBQ!$daE($g-cv$cmpR$f}>H$eN#M
z$hx2C$cCR7$flo|$d;d2VV<9pd8~gv_prUS`dWR*Ik(iEpJX~Z^e`a2wR)xX>g)XA
z**~-Wdku1E@$=O$A3S${Uu#lc-^ZM!C9eyK$H)sW+CR!^6Vj$Rcim$P(kG<fqtd!P
z>$fi2xvh1dbsUlE?6alUh9`*M_`BP0cTVh_>fCf$b|EU}%&q}H+i$0JOB0<)=c{Dz
z;nv3#;eWHIo_Ma}9<C=0GFwkrWUiiY$b3EFk%f99AdB@xM3(A_ge=z+8Cj_(3bI;H
zRAjB5Xvlg!(UFaMVj!FK#6-60i4|sNW7=0Vt&eA%%~I94WZ{JK>sQ(vabVkl?VSsa
z?F&9~KRkDHmi>13_TA&1>!J5g(8;qJ=*#xoBQi#CPUe1Q%nbMN&5ick@x9{*I?o6`
z9j7&TANc;T{q{)L$ezx$bKtTs%vzrWIsX#-K$w;~$k}_&m6$!$`aB4RhNN=N&=2gS
z9~bYwK-mR}$%(1XlOE@JQA%{^IZ@c_$?&kwC$ul!6ptPzCW8TuQ5ewJ3>Gw2VL{_E
zIM6tS1C7t%LE{x3G$BI(O;7~T#0(KMQ4v9tG9=I>MFLIEkU^6b88jtB0Zma9(9{eS
zG*wYS(=s&BG(`hV&(J~B6&*Aq!vM`t4A9IB6EssXL9;R}&@9Eer!_`xFQ=sUSJ`{J
zYwxZ@oW1NEVeIdF>$5-D+ZB{&6aN0IpcrJff?|=m3W`JKD<~dWsGtO7v4Rqjr3y+y
zmMbV3S*f5DWVM1)k+lj+L)I%O9oeX$3}mx{GLfwc$_lgdWv~z3t<OBen<XhJDLs7h
zk<>RS!=YgCodbja9k||ljMmZ~aSo{VF<MK%N6x$U=5Bph8qW2UQJ&XubC*#JGFwKm
z$Xpr4A@gMvk1Ui?0<u^}iO5nJB_YdYl#HyDQ3|qJMybeJ8KoiXWt5I=lu-t<Sw@-2
zRvBf5*?CUvi4Hv<4sVvY)VMU~q{BWB>>D>AOrAndQTPbtDe*iHR7yO{@V8YV1G7vC
z1G7viVQ`i$WMP&qVPTdnB`nTzg&fRsB^=ChrG&#-zL1AmzJ!NazLfAdD-;SaE0hQ@
zE0hueXT?GhX2lW_X2nt>;;d9C!K_pw!K_qDB%GBCWtf#qWSEspiHx&Kp#rl?i2}1q
zDN%4%EmUDvEm2`sEhQ?>YK0oiY9$)XYNbTOS-nt)S-nJuS-q6#IBOIdFl&?;Fl&?&
z182=b6K2g46K2g)V&be-Xu+&iV!^CcO04kgyqQl-ZT<fsisnr(nq&V>7bthP-pIR?
z@O#wb(`L<_KEpRVVtVVq|NfcfyoiXH8Gbn<{Bm0Or7!$)cKGGv;g{3r&Um6#$?-iz
zyWEj&qT0s9w(HoT{m4$khoyCj9%*&%dS6_(`vp-}R82Pqb?Gmqa(y27|K5%7k&u|2
l(z93Z^uGND3?4Gn!}3W?_kkIiUE)V%&!MtLnGce`{0~#k-&z0w

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.F658EA6412E57274.idx b/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.F658EA6412E57274.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b1efd0bd05da4dc07b8dbb4065bc5f045617c4aa
GIT binary patch
literal 10366
zcmZvidt4Ot-pBdP?9vK|X^4eDAPU|u?A-4&A|{+PyoG75qO6fwB;<vjN;A-O6I0pj
zgk%>pAMA3pqnKvLjMB1;n03T3tuzrsDLZ9r@eKPs(|-PVc)k2F@A-bdpO4PW{&r@5
zP8^q;yJ%ok)abe6mMkbPpY4i@it1_qDJx%CqT4XDdsNht+E1oVpU_uK_|Nkvw|z8<
z1Q%p<M5VuQEU_2qU7hh^ng4^;BPRKJHTmBCu&gBQ&3$(bJdt-K@Q4<wz5Mm1^o_mp
z78DHpxa_0Nua5ut+q1_>*43ix^SHg5+?jpk(`|<~s0TmH`nk2}frI4_mN#Sx;^~`|
zDWmos$bY<T&5>(g<tP8wJDcaOdF~n`Pd_?-<-o%`Kif5C_iqaq@tePyQ9SR?flIsn
zrljuqS76Fix?cHg?=$nKp6cEZ===KanH8BkUVrU}eD}$iTGALS`0v_r8IzBm?H%&H
zJL$;Q@9y7se5Bs*YTb~^2{T?h=t^FGq}c53d;F)8i?t86z7xHtdks_kcg3~x$i5Lf
zjw=h+=SE*Ndkb7}_=B0ZZ)u1RJoDvSMIC1kZL4N_&gXs{cTabbH&@5+jJ{oKzVE>p
z@t3~0<cujBxMcF?zbD^&!{awRQFh~MIkoip%SPW(p0^fFJ#}StijwwW8ha-G))@&0
zrEf-GNU3<~+@H0-Z;>}_4<7Qa-}>anzVmlC=T4aZj+`@&o?N!_+P}NKyJA=QC!4?D
zo3JXTcHx~LY+So`UGVgrGbQQfIi~Thg##Y^E_Lnj=)Swz@`WqvX5ZE;{rB7MYfFe8
zoIhvOme+QiUwWYL!xLhMPZ?2q=HP1u8;(A`F(>Ku8?5r^KaOty;A+}GHy%6t=9ou{
zQ_eqMmwbFy$-O6XS1)?L@yD$%=kCg0H0MH<ykY<Rl3zFd;QqXLXiO<nG1KaiQ=KzO
zEZeZB=jE5K%^eeU_mCZ%rZ$iRFSf0Fdw_fU&qEsH=A|AOGm^?b{84WH4^-l7^RHAX
z@sDrV`rSh{lXF&l%3Qwtz)$}6WorgcvwCljy*xeliK4T8)|_luI{wPD6Hjb@_SFHu
zB;7rD$W+fSms;1a={+@T{iKN>msS|B4tOiR_g!YqCb4hIi+jtr4Xt=4e))zuRccWl
zr=%vwUAe9C`sP>m|6yLBm$oeZ^7&z>y9FywzkT$}&;K(!^Qj)m^x==BoxR@rG`Kz#
zmp}Xv6`=myb!^YGbHp86mt7k4%JS!qKDK7l7yjSY)&K3cnEbdO6Ax`@>AdjfwYDj{
zv!6YEOY*9>GvnGeibI$`9w}vnS@|V}1Lmh6{%!RevsN!S=ga(j{i%{eweMsdUp`<?
z{o%?P8_Itk{||G-v7=+YTCwYuuMP!=UJ)AVYo96}bAQrmueV^Gbnnd2_;Po}%*Eer
zE>AtQt7gra>sxo`zb+kI@WuBZ+<$o3-UT!I9cko$|Kou_AMQ8xrDKnt3*^3b^4P0B
zKI!Sj^Cr$36S(=|bBm*H8gjk+;2rz-3~BCh+i`P2RCDf6?5gua=b!P-n)0pqYjw*{
z!cXrmo3L`wfq#zLzfHX-^vYxN`nJyA^L^H|wdUaa9{lqCouWKF_RDt~|8;t?Kl_U{
zfBw1S(J!ukcI^3s=O0@#<!UO+=Qif0Um1O5aF6BjJv!q5@xtDnkG<cq=TO=8QSZBE
zp1g4L!8MiF-5E-AMO4|+1x3TDsHo^$8`h3~flrI2X3UBzNsLbH79Q-GCj1lg`<Bxy
z^4UqT)D*j%9G%=VJUHc^Z#<dw(CZfu#Zm=!xmR?rZsEZx-|}+Nzb_U}NsXln?Q&{#
zYLD>Xl=V=@kFTCz`bI2uw_Q$(PP!pHIORJ&s<DPndM+!Lnr4^#NB6%aJUHc|PrkS(
z;g?;jW2w9Bazb=MVt8=Mi_h$Ddg6_vn_{Uu>~dOkTDS1vlsA0%(zhP(iC<%>sdhOd
zI^*W>;FPydT)%W0wKy@Bnr)YdMi2Fe2dBJu|K0l%`Pyf@M@_QJgAxa&hX<#8_~PVg
zCu*0hiHq{t<y?1eNqBI|CuXjH^5m&ad9l=FyPO`K9v>e5Du4O7dINt^jiqMV<$lrq
z;=+SdzBoJW?P1K1-dJjaUG5Xzr(1Y%%D?>3Z|vb7AFq$4rrYI|=#(De!6{$=<Gtdt
z-w(}>rS7!LnbDa&!-G@)<Ho5wTON4-d@ME5F5eV=Q@8NolouBJ%dV!`l-DvkX;Jwn
zH^!BCJ>H?t(TXc!JPZ))VSzXg2gG}LAi*O5i5?M1@<>3kM+Q<n3Xtkifi#Z>q<eHA
z!(#xM9uvs&Sbvf8@fZ~o{#ZQqUTDjXvvQjK&(8L;k$<*V@#;W1-mwv4vKde^3Ij?e
zhXEy<&4Q9uSWvP#EGW5b4wRh2fs)JNK*?wGpyU-Elza{kN+DYSrJx9)6mkSmirFG4
zMMVUqm?MHx%9cPWDH15990`<iwhT&HkwGcv$e>iR6;LXQ0!k%E0i~L)f>KpfP^vj9
zD79=2l$xS}Qp?dmsb}k;)D<0+dX5fCBijI_p%|bvatu(K*(NAW#RR39V}jDkwm@kq
z7AUP8>o27<ZJROS=^8HEI^x|<A{VE(o_Kd7HxTb3<VNEC9GIyi4E8higu#BMfiT$5
zG!h2;**d~vKU+^&>}MMYi~VdPVX>d9BOLZ~^@PKIu7Pma&ovSb`}sP;V?SR{c<kpJ
z2#@`IBjK@Ms3QXQ3-v_6exZQ~*e^5^0sF-|B4WQ-Pekk&8;FSgVj~f;U#cS#_Dl6d
z!hWfNNZ2nm5()d|IwE7gTu)@|mm7$T{c<Cbv0tep3id1YM8STgfhgFoG!g~-)jFbL
zzgkaJ>{lCziv4OMQL$gEBO3N=^+dydt$}FRuQd`4`}I1aW4~TcbnMp~h>rbwBhj(n
zs3QjU8}-D%exrdH*l#ov1N+T7Vq(8pPfYAL8;FVhW+O4N->M@P_FMJD!hWlPSlDkh
z5-Za0+^!YGgm2SeTFCR>4wFKGcQR0+cRG;ITLk3ymPCljW59(eV8DebWWa^-G2p`Z
z8E|3qSa4wrSa4wrS#V)}EV!_K7F@VI4qUhb4qUiG4qP}N2QHkS0~bDz2N%A82N%AO
z2N&MQgA4EH!9~auz(pt!z(pt&z(w#0;3D`1a1rxFa1jeca1jeda1nhXxQKocT%<e+
zT%-aCT%<w?TqK_aE|OmY7dcM`7r8(N7r9Ud7uhF+i|m)dMafgZMJZ6gMJZIkMe!-%
zqWBeXQS(%AQ43UXQ43XYQGF`7sD2e(v^))5v;qxWv_cJBG@k}8nqLDKJx>Q0y+8*S
zy-)`i-KT?#?$^P^$TPsjC@{dqC^W#u@EPD@_ziF|^Gt9t3ruh^3r%n_eI~e=eiK}*
zJPTZ`0t;NMLJM3hp9L<K-}=kNc?|3o6Mp=IX{U*JPvPA~6Je0qCc+|fO@u?{n+T6A
zG!X$=Y$77E)I=m?xrxZgN)u6#)h41MYfVH$)|-fqY%~!A*=!;vveiT^<gr;}M>@}n
z!je_@Z(KI`MPfhg8U6B4{NF>M^NhTl#DpJ?;o!AX-pf>E_VspB-eoj!HSJx8+)R5<
zA-B=qOUNPGdmXve<-O<vzo6Uuj61?iJH=o=(@8Oy&s5V4<}=MSgZWGw&0szgq8ZF*
zT3rn0GeI|l`D{DIVm{kRv6#<R(=6t*%`}VoY#YsDJ{zK0%x7C&EatO8H;egPJH=r>
z*GX}h&sEbL=5x(7hxuF^&0#(lqB+dxT3sCGb3r$U`FuOYV?N(W@tDt7(>&($%`}hs
zd>hSUJ|Cia%;#HOJm&L3H;?&3J0)Pg&`AlHFI3Y4<_pcVfcZijEnvP7q6N$sT3rI>
z3qiMl`C>aIV!qf(iI^`|(<0`J&9sR5VjC@Dz8IoK%okf-BIb)hw}|;tJ0)Sh)JaL0
zFICeL=1a}Ag!xh%En&VCq9x3iT3r(6OF_4U`EolYW4_!;$(S!!(=z7E&9sd9avLpU
zz8s=u%$HkTGUm%cw~YBpJEdU0(n%?puT;|t<}1y#g851ttzf<qq7}?nT3rg}D?zt{
z`D#0*V!ql*shF=;(<<hx&9sX7Y8$O$z8a!c%vW1oD(0&}w~F~%JEdX1)=6oYuT|3;
z=4;KghWT0>tzo_vqBYFdT3s6EYeBb$`FcC0W4_)=>6ouq(>mts&9sjBdK;}{z8<1=
z%-368I_B#^w~qNnJ7r+L(McJYZ&cF;<{Qnlf%!%oZD76;q7BS9T3rU_8$q{$`DQz1
zV!qi)nV4@@(<bJd&9sU6W*co{z8RuT%r{$ICgz(#w~6^yJ7r<M)k#^HZ&lM4=3C9Q
zh51$+ZDGC@qAkp~T3r_ATS2!K$v0PCin?~SyqUO4@;&+YICl}bOvi?PFZJAP-$kSu
zY5C4e@VsZ=t^RbyrS7h{l7y6me$K1&xRTVHQ&XJVv$&Fsw2T4Hi}ARU{%QROIPc(Z
z=hr`wI4rP^iYrMLl1Dr5m*Ps&jr4ryeYPX+854d{54S2E#QOx?sxTddL1sG$i_CQp
z4w>&DJhIS11Z1&;h{#e0k&xvMA|opuL_t<Nh>EOr5Di)HAUd+qK@4QGgP6!x2eBeN
zd3>*|3r{O<`<#sJHnxXzs+^NxAFZy%3piQ<;;qHI;s9Zg*#KdYxd7pi`2gXOg#Zzd
z#Q+hJr2vtT<p7b9l>kwY)c{eEwE)qO^#IY4jQ}x_%>XfxtpKqiyeL2C=#WD04*O^g
zvIb3dR$81ZnT`oB<3x^D|IGe_oD0vsW{312?$Gi_8lE_K=Iq`sr>TSPcAQo`BcvQ9
zk3ME!%9+zLi<||KxRR`CS$|Pw^X}Sh%U<8!b<S=ao8g?Zl~=|ji=PFO?A^UmdSy88
zQ&yex#GcwY^m}`E=FOR+W9fX9>@(c8z!dpAds~S2GM?cU!XUFPghl3B2#3tK5FS}*
zAp)}4LPTV#g-FP93z3nP7NQ`lEks4uT8M_Mw-6oKXdwo&*+NWYtA$t*b{-+9nDClf
z<gnmGR^PT_>9n)8YwbfitlO}7r)S5KeeOFJyqIh6PKr-Tbe@jhJWi+0Yo#yRyT@ja
z<($*yoO1gFbgj`vlDsvPcMDEPH57x))=(@mS3_~gd=14T3pJF0EY?sWvQ$G!$Z`!O
zBP%tOf~?k1Dza8XX~=pFr6U_Pl!0v4P$sffLs=0%z)YO$9`j<GeY6q>Bo1@dxZazr
zwfb)T>O*_?IM=v-&Ovm3+Yg{!E1AwgbS8cKEOxD6I=k&Jk<5(Y&Xw;x#oD5-bxaVA
z%-|f+H|&gWm+v}9*|{lcDH+ZJrnAmz=XNb}I=yY49o3Zd<~m#5XF#8kj>_q>=ZCH}
zPe*Nk!=<KXMZV!u2c>2^6bila?uf60SGy(xE%Q$2Q)^EITGn6W?6ohUuEo-bueXZw
zK8=@96~!R4RTPWNRZ$!=Uq$iALKP(-i&d0}ELBkwvRp;U$VwHZAgfiBimX*p8nRwR
z>BvSEWgweXl!<IrQC5VVbz*N!czrl>SQ0W4GM#TW`(|TM!q5nLOTA^z2iaaht?)hq
zR7t$6a7rv?K+9Axpk*ovgIczf1ua{_f|ji$ENZz@4zyea2U@O@aH!==dC>9|JZSk!
z!lPCw6+kOg2%r@ziGW(MR0OS9A%a${BqD01QVFzDg#=oul1QkPOJ&f?6*6e$N+P3H
zDOEtLR4AZTDv5$xwNwSITA_kgtt2XHwNeeVT7?E$t&(V{)k}5I>J>U@^-7|n)+jYV
zYg8DZH7bdLTC>yytyy7$)~qBZYOPWWv{r=$TC0*+5$(d_`(}6jKZvr0bIRu1zv+^!
zmp$Chua`XJxwmLu@tnE-d!pua3I6wIp0g1ZRUFxz8`&(1Z2BXc_e3`Djcm?YF!#Q$
zO1}Sg+U1VE!P6}^uKSHW;>YzIGdi<p%s8u8pPLeHzC{paMb&g;c<&)n1~>55|8H+%
p-{h3E^nU#ZWDOcTbi~L}UY1W~l7?mH^iCX`H=oKKZ{9}!_CGX!`8WUo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.D219505C6A3BDF13.idx b/.cache/clangd/index/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.D219505C6A3BDF13.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d33170e11921395a33dc2e7435bbd12ab4944551
GIT binary patch
literal 8992
zcmZvhd0bWHzQ%diVsDv*Bt&D1K|mBI);x=BNa3X6K$s?qvUO_|37OLEq!>v`YM=&%
zSb5aEsCBfX*r3OV)S|&eJK90C(i{*)?35GT#lCmBpFggjkFWc?p5O0zT<;q9yC;sz
z$QT&s<&`#P+~UH5l3CtfUS9pozr`gBU4{uW`g(aSuKPM?+JtZ^<UgBEZToT**-#kQ
z<rVYJvCv@Be@)!sV&@ktM@-5K{yy*E;bK?x2OmE%_(bN>$`|!Jb=Usg5wk8hvoL$`
zSH)j$d~f_$Kb<>H5^v_;oy+gnl|NICezUEqT03+&vHMc~Glxo^Eon>?r85tzQ$~G!
zAgk=ks-w4l$cp%{PdCn4_0}y`nbtCX#o!~mzul9*_m72(gpKXf3+6sPc!}>HYSg}e
zR!+&G8`N+2zp)_a$G(k~;k)+EC{5V0Yv-?7wo`s}q<KU3f7gwRo7{4)|DC*pla6lr
z`RR|3k2GR$)(<V8Fn#AC?}%kb3sU>%mECY%u6yp%r}lk)*Rut0s@^M(emr8waka22
z!+trnzsPSG{%pddZ#M>3zH#!y{H}9N+t#rC7VuZc&Fd=(sVf6_+aJ|WKlQAi^jr8t
zj`ZTeizjb<Gh%+9vOX^tKd@4XD%x~yboeNr4;SV9cs(srjXoUBoeg|=ddMNUJ?&y-
z>E`o2b+_MEs<&@wVyd>hx-NXd-qRTqrhTe7#?h0DSKRuS@4@AJO1|Fs%l?oGzq*Bw
zf3a@$>a`or%s%UiNj=XtKd~_W*`K3U54VT!<w_PVub(wGIOg`)r>=z9hh)tj_4dvk
z7nU3de_?|E@F^pT&K}yCUET8fI!D;9K4VJk_gc1paWncK>yDlKApOOH$P1h5BaY8>
z%|DT`a?z&dt6Tn-u_tBG?2E4`)t@YI{a*8{?Yn|xzaqAD#+V?-8pkB5xO!i|YnyM)
zN%wkk=#HA4MsncYD{DWBw{7np+8i)9>OlHPD(lFX8Ckzlp*t5`e?<)}tKRbSbL%HN
zmVd)ud-A{yXXnyYL#B@Dzuo`Zw2YVY&kb00s&UEq>$MY4Y^;4R{<p9vhYZc}`K{wp
z)vEqEiB*#(epOUD`n~uM1N%Rby1qsVk9>E3$+qOuHv*Sc&wfSA&*asph=A*3o9}Mi
z`pLc2i}aGVB_}rxJL9{d^vp*sC%^m8tc2HsBIqMu%DaO<{ANSdoq(+2O;jb-v**~p
z+S$^MElWF+wk~_C<)u|Mt<FEzHvHuezpQ|(p-t6oe_s6H)|DyFoZ2%FMO1v05O8Ik
zG?cyfVi7CO%yQ+%FNitv$IAC-t}HyCm+)QHkFKV=PZN(Xi=W+aq<ngHN%#1_r;a$*
zlK#W;JzIZh+K_x*Y;34|tswpBu$2sxy;hz-<IeaJTj`AFf8JOU)wE~*s<U^u?9SRH
zA1ZA9<%_3}4BKBgJ@#m`aQogfJuk%OY(Dna^OYGpPaS(NPY8Sc`MDEkrdJMp;r#Pn
z4-UQCcgT)U_6<E9H1>FEq1Wk*8(hVO<OOH*W={D@`h88?4e`dor4v>p9r(wnPqu0E
z?reQ&Zuq5H`+iBBx;k~pQ_r6Ke7B@b^FR4%^FPl#?@Vc3)zh=%udO$~J+|r4g_jmj
zxf#U?8O@n7*VB#;30f8y)D`&mclPgo>GQ6AO~rRdeeONu)Wv~^R+Zni#buo?^(tOc
zm_MBI^0GhNxH|0}A=;msKGVw;Y7g~w59Xc5{m1Y2+h>+%ag+S1DW)7@kLc$fEV<wN
zuZBIh>vEGnm2Jww_F!N4V9EUtU+6gf>c*x1)GSj@wkJE?gC#%oxBP!y&YcqFPvx3&
zlsziQJy^1Fr|ar_7nZ#5Pd#bMVfL^-?!l7N&V8RU?Dp>HzFw0|IVm(L#ywc_V_&Wx
zlRW9IM1N|kDG#y_ddNLka{kw?^Fn^xv(le>!jwboA))TUlAk~O$@ee6-%{gGJ!Z<$
z_Gn-CV9C{oH~-|rocP_J$}#0Qd)z?xV96(DRK0rY$C^xkYO*QE*kc0S!+rT=nbt?R
ztoc(jOgYvb8{i%+`SPsjkA|^V8GmYmDG#s@@O2NCd?m7Zk<lVu3h>G^<uqHG%RN}~
zZ@<PqawO=hDt~I4DM#8PgWQ89-@W%)LGkUTS^m`Hrkr3;=;t0R`Q8ILyW5`m{DMC<
z(Uc#wKj`ZoEO}vpvv_^DNttbFVT($>?i=7@e3)eGI0v{`9~Ox7;edD_9!T&JfJ7e=
zNb-?@WFHww@lk+O9~DUR(SURx9mw!8?vr)Whx@rt_SZhU^Y)H&O0>B%DGcY?8OFgp
z;vqJL1&wvEpm8Z2Xq<xsjZfi0;~hL`LW%&I;1EC)Q$)~2hX|UKB7r73B+%p(88q1;
zgQlb?peYUoG&My9O?9ZCX(<|LnnMFkPtifs9Xe=6iUFG8Fz##CjfL}b-%#KX)Dvbq
z@k~3@K$yMAjf81JZYIokz-&EXv7c=qEcUaFgvEZgnXuT;)e{c;xdy^vKi5b&?B|*Z
zhy8p#;jy1@AUyW-jfBU3zM1gYFVqtO`-KJ~V875v1nd`@iGclLJrS{AY#<``i;YCY
zezBQ|*e}%+3Hzl6B4NMONF?l+nu&z{ay^l;Uv3~W_REb##(ue($k?yc69xN~2BKiU
z(nu8SSDJ}}{c1f?v0rT<D)y_5M8$r!nW)&W)e{Z-wFaVLzt%`J?AMx!hW&ay(Xn4|
zAUgKzjYP+Oy_x9PZ`2b5`;7)-V878w4D2_WiQ(zD9>`<-+z(>7tTLHL;iAfBCIjU%
z(}41rd>|*|@(`QJf(x6?f(x6=f(x6+f(z?p!G+7@z=g}^z=g}@z=g}>z=d;i;KFC}
z;KFC~;KJwf;KJwe;KDn3a1k;Ea1pWva1n9^a1rtZa1opWxQLk|xQN*zxQMwTxQKZo
zxQI>>T%=41T%>FXT%=qHT%<e+TqLIiE^?*}E^@XEE^@96E^?j>F0xYw7bQ~x7bRN(
z7bRB#7bQ;t7saW7i<+r|i<+&1i<+x~i<+l`i|SOtMa$H{Ma$N}Ma$K|Ma$E`MRRK4
zqG#&hqG#*iqUY-1qUY)0qC0hPF)|HsF|rMCF>(!XG4c#>F`UMI7wdYW{M@g@o{My>
z&)9j^>z(y-9OUPIDfSdGt%PaA2VN^-k-1jFA@i+-M;2O%fGoBW5m{;_60+P%WMrk4
zD9CCnQIWM)q9N<8L`ODSiGl1$bd0oiT3lT5^tz=(-X-Q6@#xsB6aV+3Y<&}O`MJLV
zz+7}u%niz#PxJk)jAoY8z*RI;jod~vXOKH-<{EMj&HRbn;mut22ESSx^QO(iY!}61
zK3hh!n9o+xEatOqG>iFcC(U9$+e5RM&vtmTn9tVQSj^|TC=T<vGMdADu8QU`pKGHz
z%;!334)eJln!|jq!<)l=uGYq3KHo+0n9rBdJm&LNG>`dw8_i=r-%0bB&-c(g=JOri
zJm&MYHXieZE=s_Bp^O$VU#Ox5%oo~d0rQ1UTEKjvhZZnj=<pUWU#PVSm@jrwBIb)_
zw21j)6)j@E*hY((FLu%*=8HYFi1}iNw}|;-txd#ysf&^@Un-*|%$KTY3G<~kTEcv(
zla?@F>Y*jfmpZ&9%$I6y66VWYl#Kau87*VJTt&;6FSpS$=F6S5jQMg8En~ji;Vomn
zTx*jtU+JP0%vZ{21@o0ETETp!jaD#U>7*6RS9)j#^OX*71@o0!n}Ydj7o}ppT1Kmw
zuU642=BsVAiur0Mtzy2~L#vptc6h6puh!aB%-6ao4fC}!TEl#;iq<e+Yoj&H*E(qp
z^R*sY!+fp7Tf=;<)}~>;-bLw{ub0s}=Id3oj`?~Utz*94N$Z%e_s}}#>mA-Y=IgaK
z9rKMY%D{Z1j5aXesG<$bH`-_e^Nmj0z<i^JHZb4l@HQ~tsI?iM{L~d4Ubk+RoF?8b
zd~#b0Q-OZ&1u1x1)lkf4cv`VF6pPH&P#iK}L-EK$4J9CpHI#@f)ld?$TtmsoN)4qT
zt2LC0tkqB&vR*^!$VLricsPUYs*c^<Z=?CFjvgJIWv!TH)*f8*&GL@E-T|(V$dFiT
zoh`r>H83jDdZ`I;#YM-(TdQaRu0hd*;;jX{M}?|qLWfnZr2<?LVnmv?P!QmX86A^l
zt*lwa<{bAf?ZF&(6Xq41<8HztbKQhP=DP`xEOZkAS?nevveZo^WVxHj$VxX+kkxLY
zB5U15L)N>Aj%;)j!^4xu2Pa;9UA38W?C^C2S+7afq?&u*yK)J8UrCrcd^fBlEHYO~
zIAp$(@W?_X5s<}7A|gwbL_(G;iHxjN5(QbUBr392Ni<}=lIX}rB{4j_D9h0@G?#zO
z-20?4Nt3PX$~x0%KlhrfXYU6k3`(+24D*~CI%v2>%U*1J`Ow*O`+J>olAg4jmcJoJ
zwvfNRWS(jXQxo#7wb1}q;?%_ZRDODI-L|E>w)Y-`2OM$MF<5auJwp1nGR)i@92p#E
zeete1@8kdD?&M$0%?Se&B+Jq|)y*T;yISn|KQrxw>B1w{PFQ5Fop8u}JK>Rqb|N5)
z?L<VD+KGfLw-XszX(tM@+D=qtt(|DddOOjPjdo&qxZ%q3B~#DUtv2_5nD4Ma>qU9T
z;*V{Q6~3EcZVn3!3$<S7KR8ZD&%H!nHa9y`9K1EN2iS>oZ0YY_F*k?ChYqtAYd@Q;
zU&?#<hr{OP4DXCstJ=CFm|v!P7ur2D8<!9_+?sali-cL!yW9>%Bjb3hsGXbs)3QD1
zDKj84Ix@~$NVk^b%|q9_7Vqh06h^o9&|LXeSRhW}fOv%m5)=VQR74<2k$_}H22vCS
zNL5rIP0@gKMF%nz<33&S*|CO_KUiz-<AC@9BdvY2CX@KNe@n3T(V7DDO6&b0!P;!z
z%cG(aJ@@jcq^J~&?%dgWaKyhi-0Z!i^n}N)^WVIr^u+sQZ8fK~_ZJ9{FH=D=)p+-=
zpjc$Cg5r?*3W`S-DkuS2te`|>se+P_<qAqhRw^h3S*@T{WUYeIko5{mM>Z-b!^75(
zDMS3+KcaYcGbAn~!McZ=kIAHvWDhY#OtI&zVoI48fyxO}fj5gH7HZj27HZjY!eT8~
z#6c}r%0VqxPB^UPi+HHzOL?f}%L$LQLXiNqLa6|?LOBtzRxA>sRxA~vRxBqX)=EVZ
z)Jmli)Jo+<!dkgVhFZB)hFZCt$XKfsDNw7FDp0GG69sG4A{A=YQWa{|a-w3bR-{3#
zR;od*R!%gm)r)kf)k}4#)ys*FwMLNvwMMA{wMIEHJhfg63!a+Q`%7K%!r8?O%ztUu
z(fL77oLs;4s?YrVxdpT5IOloI?k)J=KXa`QUS0*Bk8?a9^F1G(o{#f9ALn~M&Mutu
zRIifde3bUK+57nT`UmuVASiHLzx1?(etzS|1P^#HWZ*-hq^O#1j2_;9s2s-+e)#{}
r8X6uE866WlC_XW1Nb-o0qZm$zV#9`|IQoY^lDUA&7@s<p{N;ZD$O)b{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bias_add_reduce_xdl_cshuffle.hpp.2AAE1676CF63D1B6.idx b/.cache/clangd/index/device_gemm_bias_add_reduce_xdl_cshuffle.hpp.2AAE1676CF63D1B6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2070d7f5428a3a602b251d2e136c51c200675e88
GIT binary patch
literal 16068
zcmd^``(Mq;|G=Ah&-*Q>x8tOAKP3`%P73ut$t@wWZn?E<T^5UV+qEvatR<G*(!~-A
z6-A_tR*F_Bl_dF4xs*s!y17U0-^n>+W`Fwr0o#K!=lOWO-gA4;yyi8pnKRDE+xwpZ
zdU}I@^jW!R_OcmBPfyPf|6jUn$-H7-PcOh+PjBV^^M2(&m%A%Umu7mkuSm5^R@Tlj
z=GQpDE~hL@2A%l@jk+GL@G%}<aP0e6_gC1}d)4oBUnF_{@cq&vvCqW^R=Mw9J@;E@
z_onBL#lxGPJKg(x$fp6x13y_zsM$JaKw{zetNltg6rRj^^K2KGJK2nLzXi=NY^r%M
zYsL?II!X=Z`4cxyDR!|knt9>l!83OYzL6pWB2c1uKR!}Xrr$HZ_Y*tQTj2YLo`$E(
zf-;e7Y}lF?{6~xN!Ex6*=iV6MRySwf@<{)`i<(z2EACO>*ZJ2N{XNUG`_11T{&&*k
z5v8MSUezy~_&49q{Ql9d(u%&1Lv9#N$=%iB{%qZxEmQXz&vtpyWzQ$i8`JD($SUss
zdM;+?O4ZiX`hSF@UFUUhd6D*FUCo7)$YjN?xzVub_(}W4)1bzFu`hfWnbB+Nq$?&z
zRI{?4{{8*(-K+DK#GRdfDl;8)`doFO?Cr7e70Mft&gGT;FNNJ6yY}g?+bZUT>ifKP
z+HvNF(NAlN2OCUie7kIA=*zu-EdM5R$(|*TR@{7)T7@Du$YR10I!kHocMZ>*yEcBW
zX+>ejEKn1Ac8TwYEAmex^DQz*yq$k~jyhlY^VH{$Hdn8V3EX${)Sr76^jSZ-hwrbq
zFMb+Sv}3Zo&D6C{DOU5(e_Z3V`Qp!WzIWO@SNS1)a=1fO!sO9GlY4ZWbYOPR6)ArO
zCf$wC-_!0za^#v)tGc}41(#0edrp|P>U3ddkmc_;7c9(Df1fD0{nGU8==|F$XM)bO
zEzUew8Q8aE*|WvUmk!(CEMDw1*LVBUGV9tImM_;^WLCr+k7>E{Zo)c^dvZe(A2GY^
z<Su9K-1O~>%JPa!ruX*CPBLE*ku>bDW9J38Wf|)y%93hJjx2wdG5x~#&3A`C<GrR{
zv>dy6b!yyNvs2?wg*47R{Vsp)q&QVtNo~OdDaIsr-nn@ZKU`R_zIpSM->wD(4vHLU
z@r~wFSe|vw5tY}`&Byr3HxD>I6OZ~oPv6_*xcbM`x}$5`O^eyT`RwuukNc#QOnwj`
zJ0BZ5wa|Z2p=!eTT*th>SKp3Jah!1_`n{^%!XCfveOLE><$)*jRG(t1mJeBZ;HL6U
zYEZ7KYS6}j7xi_2yJgg5^gX;`f~L&k<)WQMqZW+YFuV8n#na|ZiH>t?iVVps4tla?
zZ<34E__LPje~kcHGY(hjFS&8+_PHayN{{WXUVA#_hG$*n=<|uIu0^~_e0w2cSl`#f
z&$^i=*ae&KDRElS*R;LMkqQTcSp)L}o0T<-e_A}`CLDO9*Tg56>JOT>o%@ra_qw!M
zYbT_Qng3*>&$wli2N%B@|J$h<O{vzde$f$0DL?KEZVdAOfL3J7D{~h5T1PuCEZPuo
zPWmV8_Rc%i-^MQ~>sv7N_;|}GgH<!fboV(K?T?P^d==9vY~#}5y>FLWn@E=(_iQq4
z+wt|E&ix~c6QlgkI_ViScDvT?;CY|4R~Nh&cMeGVrs3ek4;O{^gR+mgpYk-B=pVk%
zF7}PnR&`MPxVbqq(U|W8!Xm3$PHpyTY3#CKQRq#JdF^8L2aF%*Un^`Ge!HKnr;E>+
z!-1YBzT5SUzE{%Pp>O@>ogAKN81uP6KkuGTm%FON{e!)Cdc(3mIzH`oI4rW_!1c#`
zOXzPc-h;M9zVv?ZLd9Lbd(ih<yZz1izgmTCDtV6lPd!^Py6|B4VeZ*k(^|59EF;?f
z6L=<|lVLlnft?3cjoaLB!_m<0>kmGEbk8^4tLNP3TTL(hqMVl1{esVse>9UjIVVP5
zYuEkfikz0dJvQ~TxZsf4v~Bi%_j32k?T@)9IGvVPeOLU?#Y1=IXSeuODkn8xba#0%
z_2Se)x0k=nx!yj_$u6q?>B%T7gXGT}KVNIUacA%Eu}ON)M}z%8YRX3aI%~agvVMd`
zddiVGr<SGFE^ON;{FzK0-|tC#tD?QjI%Fk$`&qPqR;yQYv3tOtGtZp8w)FD~S`xQ$
zdPBY6<cO;EyE5-idpo}LM5jj`OFACQon#ufp=wx}ok^UzNA~H@ZC<{Zds%PkszozL
zw!ubzM8#kW|7BMN*jA+{jcGGR_2omaEd-W&dO!l){T~&R{kSRxwC<rN#UgnuGSnq{
z`uKVL*Ct`)>!JQ8nH`{^T2D&g6$yM7T|!jaZocenov8d(4hK2uNt==RX4GAm5S0$k
zb1v8v988l#PplG%ECP|8E+HyL>f!o>OYd*)0L^egfr27Xu+$|)#k5;=^MtOe6>?~g
za|uGGLAW4oLR7kJkWMxK;xbJRZL!K;WW5)ubO}-EdckV&Ze-M^12n<~`6>LIzI+gI
z_lTA@pC0R9l0z??!wzJ=1F3ZhQE@ole$!T?7aQcTKUVnzng4-0>k^{kT-__V$@g`w
z0Tgi|>p9ta&P10G71tgiJNCKxZ|wj(V3j!DB93>{B}8SI^K0*6)-U7aZ~!i76SCTb
z8>~%;%5cj#uV3nw9vz_-&Lx94&)_@j5~4D?@64V_`m>Vc&;jSN4Vi31_PT_qjMw`R
zoAR(OzXLRB%>^hjfRipED&LODIh@<c;Da1?!ntfgR$EXnT|!i*J&G_^e0MWj4hLhE
z-ALSxdgv0OGX3?qvwhkpqYlutwb4)!2fOJKqB5&(zeL_LvZNzyhgHT~jJNvoLBzj2
zty+1rzf+JLcEyE+Au$Z~)g?q_$+on8`h%nP7{E3-hmTP95t``|qVn7LlOwkc$^KtQ
zD90)jEGAfe`S>cX{n2&l{7o}DKnpCcgo;Y&s7r{-@0(^B%~O>;Q9yI7QY_dO3%zv-
zQTfB>%(CFn_qeJ9aaA`Wqm9T)mk^ar4r8n49sJxRhh4BrIC2X|!*mHz*?zpw<NkhA
z%;c~$R@sWIw&IT0CPXDHs{fj%t}7<Xp&eG)j!d`XaiUGWs@%4|ai{F%J~?#6Dj`S=
zK_0q<sO<ZvUl)55{i_Pt5w~cYfVK&yx`e1ij<Jn>>$P>L0pzjDJ1BbxmAZte92kYh
z%wGEOqa51cg7zTmJ$MFc6QU9|>1Uf0p1xBIpbY2YEA#E}<%5W$Xa5niIz3>S96IA1
zB9Ki4a?&M4C2m^kYo{ZhuXKd%aSjtLCR%;@AmSqi-=uDwGwrwm<goZHl)c5()FwnF
zaqaw{b~a8JE{EN4A%7zAPdtpY2~jyUD?WKz#IDnFI0UOiBKt_xPnQstbN&6(n~vW-
zE<l6}%HRw#I3ry`RL;Mz=r!x@r_~B*fK~1YiaUZ*mk^cQxp$tuRK9F9fC5(e0A(NW
zc-1CEr6})l!?d)BEIAyC>#`qp+m8n55~5Q4zT2?@(+}6mVLz<00hw>WR!N%>mD_xA
zuUZfFAOnc7y+4HHhmeUbAu10(y&PFKoo_2Z1+GgpGKfYUbO}*;w9%_6clL922WX5{
zGI>QNZ>39!O5@tUpo^c$3&odJ<1F9yELPDbMCD0^?e>dXdJHpx2v_$4r+<M{=n|sx
z;;%OTrb6URIqZ&A_95$icqnNTqSErY{JCaE<0b>B#C7=uWuKshE+H!KX0+8m@>`*(
zUE~h(@`Jpgq)mv*`}v*)<4XRQU;y=TF7Ki2Jv7xNL}khBIZH!b@PepYQ=#n3TIyrx
zzdh{S;I*RU+tIfV8Q5f91v1@g`n6Zn-8%If`qz1dC@ohkSCOSil-4NLFth_Gb^wMC
z0EYu$2-S`Sida^D8YoV)d>1J00u?dpqVy0b9<uxjC|<F=5jt#yL#Y0CsMrpf=Odva
zlI09&lmUAXTP{lZ&?ukf8_?(m%T3Uz3HJC+XH<zxfqO^mPmn$(jsz5?+noG1rzB%U
zrgJCAbd&w!ee+60!4^)G!huIP7)-mS7`PRK0kkRCpvN^hm{hf$%a(}wyL*<m6{Ybu
z<GYfjMwBMlOz2A11W}r3GqEc%F`{&f7jN+$$fy^k`@DFc<r-eBVfnBi9u_(fYbr_!
zf|$T^vLGf?Caw`~^es+)i!-B**7`}L(ZoT*-KO=9NVl1&W)xi-o&FB@#zx?=5qQzw
z$blX?(2KNQlxm?{E$mNT1o5ZvgB6^-g0mq-i_$yJ>>cMoSf=x`NL5U;7aTa-{Zt07
zb1Je+MPtZ-7o|F6S4Y|G)+zlDM^k&XM^ZBClZ-rwB9dZw(-_{2Oim=F^Ig(;J7PGH
zRKc58@MfgfkyOo_R<m5ki*<ZEDsSM$29}$6v5DoUy!e#mV}j{1!Hm2mB%KsYPqLgM
zh$%ulDo+)}RF=;Q;#ro{1u>nnjeLEdt@?!z0Dk?*G68ib4YxhF<fq1fet`gCfk^@~
zCxXrcPoI6^{y*7&uB(X4jtA2tPX36qq@zyjzmri%hBYVCI_x@2EaH{j;Vtwgi@qq`
z;hgVqBS|*Ip~s{A9w)!Y*%E;$edI=d<fapr>D+u$9&z(=#Y#AN31><bwLbq>#kZd2
z9lQA$;}SxVO(^muKB_2PLaIw>DtX7asz~*gvg3m||K~Zw<ocqtSX|uU%Z3Yoh|(4y
zZUGGM1L8iGlYp4S@+~0VV)-c$pRybP#Q@0U7Y@a6%4D-*DI;5yuU-3)vwg@pk%o&>
z1J}NRz9CV1&KW=FI#GEuXWY#4Th90`%O5%8k1Pix!(gNg(>bTD{bv}8l(heLBEy|X
zNt`uN+Jy{vAtl*pVDpa*!;zAB(V}z&DUKj3+TSUteG0OqoQ90kP$$ZHNSTKmC|^R#
zODyLjWj@Q7k@7Oj1xQ)Iay>GxXY+4C$`)4s7AfDd{0=GKvHTt>-?RJyDL=5B%(qYG
zEy)T>wvc#9(C#0}lF+Aj;JAAi6PjhBv{tir9GN+y6a_4!fJFHS&>R7b{0z{XVfhiz
zJObnBavBIN1EEBD2h{9<jQjx99ALQs4l00C=mJy%ElZ$8*u(mH!2)5%a|M#_3F19L
zPMl>V-513BELRHVm4YReKM>3xuv{gWSF!w1Fn`E$wO~{&bfnKe5{w?PTqBs)2<B8?
zE11@@{8+GjEGVhGPOz+Fxn6Lo7kW{7gW%G@a--nXC^%F36G41J+t*LtKWDnP-736X
zR6+YH=tGw1!EKXIhs||uD<Y|m_p0LuvkRVA1ItakR};%mdC#Z(AgbTY_ig6;QU1tx
z`pDZ-P7%CPgu(2B=atIxS;6Zp%jtq=x-f|9X9=USgwd2Q3%xE2y$O?01~#TQIr&Y_
zg4hh1E}%h->9FGceIq9iT{={hrnpWSK!@s7*QpH6aGk->OxKwV`K$ezA-+<*lA%@V
zRSf;E{+*@Kzkf^1Tc=*f$O3>mfSowOKpp&V9vuvrr$d1{l$D18br{Q0KoJEjXlc<v
z5zTT8P{gnt2h?%w^KoEw99!N|pgzjV<AFM!<rJV!`8SVFVV_S0>Qq*q2GnURUj(if
zS-(05xaP2&4_xzEz6@M1vs?kx6@W1u_ksF8%auS~$#OMNSO1$wSA%b<;i&=Y8dhEl
z)U_-(19kJid2}=T{2QQt!^&HLx`pLHs0f4>w7ktwv6<y9P_c#OAgB&v4Py`-9R!*B
z219i)D-VI{5SI5t_5OeJ=>6>T2ch~PE02QeD3*^y*W+*iF|63@gsurJCqdUFmQO<0
zlPqUK=S<j#KA#1hvsliC&e<#%Kt%zxp!!#!;tI=!P*KQo2~?D@TnZJXEZ>BRn=Dts
z5mj&kF@vJ?1Uf%~eJDSL&QDo>2A!W#CQ;FLG7<~@>e1xz&N;k0F?*s^#5)(Ue2sU$
z#_}!R=oW89_3!gW_gSvtjcQm95gbFP*%PIGg5y4xBLv3?mJbU?hXpHQCR=xlMhPq@
z3r5M5M~%^#<`3QLix-M8u1gqaPkc8L<mE-FoRgPx*2KQz-9P8^ko%5u6Q^q8`ci(u
znZ4k;6PD>h!^9qwDS*9hZCsf67<iKi9o|825)%}qGS0b-`rYH(4n4nUwjvl;^^VZ?
zj?kVgev@YG{_aFfxd#?J7mS_@CPd)3;?PUa;FgEjkKToBcA+8E&%2CNm(f`2r@cq2
z_msaiy43TBzs$M{c+D`K(wkb<8KyHBx(G}!g5IP-IBWq-^I1L)O^?IgbWA5f(<I6y
z0FG;Shm+so>`0Hbdc~>+ZV+LaE-X%ZaB}=bVfG%IT>M}#vI$0Gsk@tpRC#C$ZB`3X
zwNRd_w*VUU|4AQr)OyZ*J?Bea_SA^RreTWaLv2K9j&hC*ZSP#=T!yX!<yGK9+gk*b
zMJ%6&%G1z=wl@VTQ&_HsovLAPGUukI_E|W^-mVK0afDNu!8wrji8z{Oc96FtdwCH@
z`nsIuoylHal(q`STZK+!jFY#D`=g4JS8*<+<3#BLr})72A}rG#9*}}aP1|61_shuv
z9g6Rv!+YpS;&?bC=<or0lE5n7=|P8&(32btw9XNSa?X<!BuaNV$Gh|mw#GvoD=3pV
z5f1vbo=4DPlaX68(vbHzXOc4co}#=x_Hi13M+5L;x*mEI(LfB@-{2mo<m8oHe_D}O
zoZ=NXkg!a5vO#))GyqRV?Xd=}${tRC4`)I~h=?OL`a#Hqgn0bF8B(bFRvK#}N(*fl
zGIL^)?IMOY0{e}?osOqXz<v|UyMg^~mcxO4ILpPrwwRq!CBU|X<pyBi0NiP}Gy?lZ
zmOlXd4=jHK_8(b32W`*M87WHX&^Ddr9B7{d-N`r+rCeyA%W@I4FJk!`w7<r3Ewrtr
zBVLpqL)*tJ*FoDlmg}KyJ>``bA5^bK*RRS&DVy({&3B`#Mh@ROhvh4L=PN81@|_D=
zzRGvL%5o9kxrpU!eCKN{?+}zbgwAA!iBgE53}HE3P=>R-M^Nrzd9R?{%kn-!xsT-t
zK^Z}rL`v~QY&|-ogK+BbsUzs<nK^LgAWAu4U=A2W-~DAU{4y9pIROq&fFtO;I|&D#
zgo6kZD}+bB_Bf6f`UyIGf`iFPKx>poeUI}dQ;3|?;XShUkd8dK;lkf-AB>v!D_%58
zc*_!ACB1GlZQ)J*b=Vs4mUmfx$Xh;SIa;uc7F3!EK2_uHcB~Vnc)>EB<s`u}i846`
z#AczClb5n_B<-0XncW)?uUkDTKKBw<oNPY1Gwr(R*3-KYBC!r^J05cKhiowA38#8O
zZHY{Gj!1IfK7!lW6!O{KR+RjD__>k%F!l3e0_c7t{Cue_1N6uMZuBK(fPNWZFv}w{
zfG_1Z*gXz9lb0r9JB|PKXL(2*{DyKRR8_)W<mHM|9%q<GL*O`m#u;8>IiEAkXZbQ`
zc$wt_&ai;xE1cmKmJ2z<LYDtR9sZ&*SZpwO+g!doows?sZ63>)c-u=X=kvDtEMMkr
zFSA_0+ZM2Vg;!tUd(-?w1*=e@3!S%Nf>jvHI|ZwqEdME3{mJq!!D<)Fy9KM=EQbr-
z!i63*|46|sQm`RRj+XIQD&yp3tV`3#sT%15DAOG{ld(h&obiKqIr&}o<lr@D_L@C8
z&>lm7eNY$a!w=r)<oCILWSWZ7Gfwf08$ej5JDesD?mEzA%GjV^_u~h{kxe*qCG$X(
z3XrM*d6GpKul-2%f%2}ahFdl})h1vkDv?(t@|{TVZ1)4l3!WLV^;l3Yn3W55#0}m3
z=7*^p+FO-kM<*6J#iBkm>sa(nEb^lq#hXR3j!zUXMe$z55emON{Ah9DS~Z@}=LO^Q
zwBNBI5R6L&8)BQt`@l9zd+bhZ6!EQj898PDx{uk#>F?q!>3Y8v>2F1r<c;EyjI47}
zPg>PC$odWC10A9cT|BZd4L5#_+nAxGJaTxCYh2CAt7)WHraQtXuj9bmq)oR?w=3I;
zlE21(3^f=lG%FYi0h$oNa1<C21twAL3ZSU~W2pJK2Q>FsehoCQSq_8)0^ua8p9wuO
z;W%neE<le9EEhn_0@gn$fdfk5B&z=edOU&Si1{JMx1vmYAK`0X-{HF6VWZ$>T-P#|
z%Q>5Jt{WM;t#hK&UAm)c^(H%2uzZi}e2;Ua`iGGDA!JKV=S3+88RxK3>|A7=%kpbv
z{F>!A$oLIqa(@E9#9B^X%UO|vWV%}v<R!)qE?0VGzl;2T$0s=p#r|*WT?;Z=M0}|C
zLo%941{pbl!J}1o+2U&t{>kb8$&S`7NPi2mBwMQ17F*}g(J4x=k@aiJ$NX*#kfkng
z!sjHvYknU^$4@ZO1OtZSfF_RR6rf3Axf*Dy!6-U@nt`U7<shgDf{c7W)a<8xVr4o%
zcF`g40iv{0vvMG<#!By%qbaS@tYT!VyjL-@b((dIY@PQyMiu}x0f3PP0G|NBti_=~
z6Uxd%fp;iio(}_>FjgK0e8Sl0qk&~K=tAom4IHCEFUm2%GKQ7M0LK_seiUeqviTnc
z-bdN;<AElgmB$00c((jhph;!rslYpxeLf9n(pY&K@JVBzzYjF`!9e=*?gQ`pU^L}Q
zps8f#mB71_mDd1G4J)q!-ZiYe7HDc&c`fj%Wqq4BK=X!`zX9HF*yme-riGQa0G}52
z`OVOBGwed^yBRuehP^0nftFiX`4;H7g_Q?GO)zBI9}K;NA=Ca4s0m@^A<!oTGQQbC
zs5!{W4?^#Q?DJ7j6UEA-pidP0d=~V`qI+0T%7Q*w(3cw1Z0M2A%Cn(QHY>jZEw8{X
zRR0QeyaIbsE`*kath^997P9hEXj#h2OQB;aE58XXZ?f{6(D5cKe+oUGLN8+MvBLp9
zo>5Nh?J+j@zz=({F^oo5(Wo1_vm@etgK;rktjS>{_Cy|V@(0x67Nr(W(ZX)twfD)0
zxh40>a0!n&`D5yB$#l2Mh+;~^>J-y$FOL7`t^%BqlK1|0y>U3c98oS9mD9)|=^4D?
z)LycqOOW>F7*P=Mb+#^tkczIQ3A|$h--|eXIAp+U()ocjHuaou|C~3a9{G8p(|Msg
z@yPzkFa0jD@5@L3xmQ7)VeFD>?<|rY&u|_0BR}*|10Md9-6wm}<s%D7S?q~VJe1<;
z2}9<Ht(}{$F3!eIbru+r#X5C)U|b&XqjoSJj);e2s9i{e;}W4C;mk#435z{%$Ks3@
zx-Dc*{XPPZkL=bo7kcE<n_u{55_GGB1IYBh7(e#scHa6M|6kuH=k=V4NlPw7V^4Xs
z-ROa&8%Sso$0I&`xv)tDcz?p_hqIBm?MQz+G9g>0OLY+e2Odk;<w&X&#7cSwjHCyG
z_<-dqL9Al=p&&kFxmpmbS$-slk65k|#2S`s1+kXp$Ab8n<vKyEW4T@s>sf9P#0Hid
z1+kIkCP8eXOs=fr9<Jl$bzE2KhHCGxk~w>=pZ%I~W30meb90Q`*u_0m%*l({z>xOd
zF6p6C=f?8CC;sp!4pm*}yI<!eYOe0^1Mcu6DOd8ImHcSR2ZZhi1c{h49FP$P#0euQ
zpAbAx&{KP2Hu0#|-VUaVU@S6=WpB2`@TwTTH`(gm{wQ>O^<dv7Y!!AQo1Ms)tdjVY
z1*!7sl_z{^gjDY+mzjDR*2<kkJiU^6WiszbEOq6>0M9VPljV4xFLGPNxWy&Ft%P;6
z(xF>A97xQADAhx^dN`2G^9SzN2ERFB2L5w{i@Z4`BiFaTZpH$IZh=A%a-Xy{7}2Rz
z=thDOt?W=LxDc*3i&KZzwC{mm#Q%`$f5@3^_wZmus_QIo;8h!VM*at{`h)T#7yeb7
z_NLqMaNjDl-6|N-QCKXrEf$POiDdNRwMToqoY>ZytH;jAtr;~2S7j6M*u>uZZ3G^T
zto1(!-Oka~r!|uA_Ly4#TJ_n)-pYhSxGF`+xClAYHU;p;0sL2P&_59ZJx+$=aXr2N
z^x<^1%L9G$fHz&V5@Fv&=uJi;aW8NSbhqAho!|6+O3SFoKn-p|7BJ5OooEhuz&a0f
zr5q2<<6$R~BaU7{>qOX<@Dq(q;gTPsb8riG1CQP8rQZ+0;{!c=B32hy<=?K-i~VOa
zoln}@{pV%|`B?$(x*ASi!v?vuKQ16<mmDczYpDHY0U3v)w3pN0%bC&W+YY3^1DTPl
zro<}ZKG*)xfJ{bmK#6nI{>p&nxSP}8&2^+XZbSOpP)BmC|Ni}vOy|4dudta6<MhKg
zTcWsR_QDxM@SV{iCk(2LpRV%*q-=6rt4CTxLVjO=!a#OBdZczsm&{rk(s~KCSL=uN
z&z*eH)9X#Yy3wya=~vCyuTa|B7;*K0!L0}bkG_3+_13t$xx4h#>-$wf&y{{{{Z)nT
N&x3r?)9XXO{vYKR2POaj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp.4E5648F8B76EED28.idx b/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp.4E5648F8B76EED28.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ba1af861e39a360a3f4b229d7e73fa9e8978139b
GIT binary patch
literal 7866
zcmZvhdw2}@{>L-lU9pn5mW0M7A$8fA=_+%-sz}I&loe#96?T(WS#^T8F0~~aR#kOX
zIPtigr7qjxv~{cM!Qs?T>+$QUUi|0{rK-o>I*OdrbJq45dH?bAJSKU~>tkPbc4l{G
z)>BkixT9-IN`7V0%qdkfCfHL_Qraf|PM<OD`P2k387V0<*BlxBttVS<v-VG?x0IH5
zEL*vEar3@X^<(p1kNh~|vB|TaEj>A8=e|=j4+hJpA3FP9>*7;^ug@;KKjix@@8+jQ
zs8bt0AAKQb&9={u{dn{I{7=qKyT72~`!yA>?Ef)q=gm*ooxOb|xaXAGx4d(gYb{Ea
z{%UuPIzIK<^0OCj{o}{RUSnVR&r3(LUjCqe@Id{a-YEL(y`S&ZxDKp4^Y<67|M}iM
zzG$4ycxp?RzR#`LzHHRE3x~Jzldq(9$tt|GeayO|a`EKrn+gvMd~W#8s#4{-V?(-3
zf8}rg{@~14lYXiF_8RwH*Qy&iyEa!S&%E7Y*TCgl$*!^ucb>nz=EYz37k<!kMZc<V
zwA8_$&gs2vYr(#aOVZAc$l4`juG}@UXUf|A^DFqZWx=~QZnkgt{+%z@*&jW=poRNR
zgs;4Hx~jPR_@Sw#kq#Hi7RBc9ve0Ui!~S#c-GAu*xqc=jc-lm(_v|@W)Wv9bG&*Tx
zwz0Q&=*8f?Pt;qRM>o#7vHsey&0|}c5oVZu!nV^Vh1r*GUu&~HV<6l4Cx5ShrLSGE
z`w#95?sm@S>8Cr5RAvv{eQO<Qy<p(l{KdQ5jXiNhU-NW^5Xipr(?ad-NZ!C3Lj$YZ
zrw-rre*4c$PA<~UY~LRq*jsmWo;_po7aON_-?V69_PjQ0bIs|oKKdZGYKC@gc!hWH
zz3_6sBh#~Wk7q|?{&(YwCvmi2Ocf5Vy70{4Ie-02?=~yn@BR6cw>NxDEn2r|^R&#K
z*Si~o;zwSa_u|{{_>^<0N77E_zF3^*xjtrbr>oj;olXpK=58I!?oTh>_2(nyw{|&7
zR+Mm~JZ)RbUmke*-m!1@=jU#%e64ij<F8!mST$fl1#|JsJHIXtj%>KUs%Wg|-9w)8
z54%qs-?}(=OzGM|C0oZX*cr+Qqz;RHlJ?HV<EJk<{&$7yKk*S~FIn2@pNsmOzB_Y9
z-wn&$_tfH9jXgp~uS|&t>e(NPR?hxGc_I7C{Rn+N6iwN&<F#!c@4HyMqF;~i4(_>A
zs4_=)zx7Fl<KDu>wZ+$HXODjCOJ4qa!$taUm+y4?YH8({l}nCAWP0-5xidIrd~sFT
zrbqkieEratUv`(BTk_4(gD>1E>h{=A-!FP2AnaBe2hD!G@p1R%%d1`Kldf`me*S8H
zeB|uAa|7J7>67A($4-}xKE3w)jR(8lTX~|xZ(WbS(sRu6PlFqJE-(4DvfrB3LwEf8
z#F*iydM(+1^X+377OxvoFzdHj1&!lA7<GEq=?+__4&M`6zu?vRqYCoR7aaP!wCCb}
zpT&(Y=0;zo{@C?P-(>$?*VXm^!~H8Pp6UJOM?Fmctn2@I=i~8h{vDq<`=r<0aQ)E5
zS645wU4Lo7=xudx8I7aoWb`#QE_*(|k(#~u<kWE1?Y@We=FVFbES%0YW(V5@22Gze
z#kZVFNlERtcWM55F(;jjDNpfdre?NCE)wq-$v<iLqd$a-`H|^lbV7GY?b0^6uypq~
zTAXRK@Y=p~GAg0lr?zjETv$4<)kM9+fxlcxCuIrUF|}js<igTs<IStFi?iNMC*Fk4
zPR(wfTv)p5gd_08#xor<QpP6qpk{+gk_$^up+Ed%dB@U{bW)PgT~oWZN-iut-#2{P
z$w_%tEmKA$bWUbYZgOGiB@5ro{_2L{OeduY-7dAABe}5j`|`xl=DNwfTBdjsIyW;n
zFS)SvhwEc=m6d0DrIV2f{YdH~Es_gM|6@YVK9;-MKb;Iu=&aPN7RiOBFTY$m>eJ?G
zJe@q9(0QqO#mR-GQ>IltJH2(srq?DfE%}<fakw$M?VOsE;qTbFV>fHtX84_=Q?VX0
zL2l=;CZa_e9POsTF(LzwaWmjpkp;)PS#X@lf#cj9I9}wz@opZRAPV3Fw*XEQMR1~9
z1Sg3SILR%6lSLVv?3TePq5@8FE8tX71*f`IaGI!r)7%<3UDUzpZXKK<8sH4K0nQXn
zaHiV?SI8BL5BGKd5N^mrvW~tUY00CmX$NqIDCbhD>1B3CDCcTmIz-Xv=?F!mXF?Q%
zo{3NldNxF{=-CLxqUS;shn|a29C|)P@#y&o#iJKOlz?7{Py%`}M2YCd2qmJILX?DF
zick`IIYi0m<p?FCS3;D6UWrf&dNo9;=+y|NqSr!{hF*(M8hSlM>FD(crK2}Ol!4xe
zPzHK4M49N#2xX!l#tk!@c9{~c<RMw_$nDaS@5HdnR6A!i?J#sT4a8J4Kx{P&#8q=Z
zd^HawR0}|2wFo3tOF(k945U;mKx(xLq*ZG`dbJK@R2x8MwFy+r6>HWevECaK8{xr+
z4r~O?SpXYBqk$NW0b(^4h|@SAUgLoTO#l)#5lGS`AX$@v6ioqAH5EwHG$38mfeg(6
zGBp!uFgIAYHbTGmHdSOs{fF8lJ`)_yKJc_1G!WxpfLI3$#5p)1-oXP24gpAXh(MA<
z0+Jmvkm68)REG+rIW!>Mp#vEX1ITok56L=>s7_0sKESITCeB5q>4`h*iE|Zlj5yyx
zjuYo?<XW3^ferqBoz1xdxyJ4c*_%8aCN%cwdO~B5ju9Grbez!Gqibz6_UJkrjXk=?
zPGgS=69#)sJz=oN#0Y~uCQcaaF|{@ZdrX~;!5&j%XRybH35z|pp0L<sV}!*X8z(IG
z*jgKlJ+{uqVvnt{v)JRpgu@<JPdMyxF~VVwixUoeT&<169#?1Mu*cQdIqdOa!efuG
zCp`A}7~!$U#|e);zShQLkFT@w*yC&LJobbz5wIuJ69Icdj0o5h;zYonP-_#gC)C*l
z><Kk?0efPYh}aYBiHJQhMnvq1aUx<*thI^Q6YFdu_QV>yh&?GxB<xA`M8cjFBNFzc
zIFYa?)!HQNNp&^}ds2;E!k!!^GWO(pB4baE5gB`OoXFUdYi%<2<T{&-J-Nm%V^0Ya
z1$#<8QLv}Ph=M&OP894ZwKfHNN}Wx?o>F61u&0KJiaoWSsMu3uM8%#OCo1;TTAPYJ
zwa%tuPpz@5*wey9!=6@8H0)_HqG3;q6AgP>txdz8R%g?&r`6as?CD{mV^6OqI`;G!
z(Xpq;iH<$J)}~`mue0gc(`)QH_KYwwuxHd01A9h{7}zu7#K4|WYcsHC)Y%N|88vnT
zduEuJ*fZ;ii9It$OzfF)Vq(v%wVBv6>ue_W%o@Ad)O#{~CFS1j8Q)NLzvmH8nRQy=
zUUl%z&q7yPk_><Q4(%Vc&KDCGm6;v0trMqV;#*TQ*(DJw!{0^fl5bsyl$NyXb!nN_
zEU~(vML}!pA+4WOTJi$j`T({r&=Y5253bIw)rr%QDCLad)k>72k(nsPAhS`5MdqRu
zhs;MQ9$AP|0<svTL}V#SNyu`Pl981tr68+ON=4S9l!mNFDIM8}QU<aarA*|fxThYT
zs=ib_=r~;_3{SiqIR!amth*K&{@jAxG1gs(SBKmm_f}ums0Z6Ur=Y*Jy{+M9X~|m~
zuz|Nz&b_#Sw^KASvz=m)+3ghD#PjFwtvz)1+=q!;jyGq#Rk7}vaB0aqB~bG?OPfxD
zoL*-+P|!IK|05n6Mrkh%qjZqQQO3i-DC1>dlnF98%6eEBWxXtnvOyL{IS&V;oR@=9
zF3901@8MyT_wq2x2YDP7JOYdgUI9jhpn#*IM}$$)E5fK46meAYNH8jSB^Z^05{}9q
z8AfHV45M;T#!<ziz^LL?U{nbzII4P77*)M0jH*EuM>UTIqncNPQ7x$9sP55WRQKvI
zst0u(H9Q848eRiNji7;}rpJU)(`&-088mUUh$|ATLny=V=G;I-xuFkf>*dfa@3}Qg
z6URZ`$h<OZr`|U6qh`aWY$!~u&Tf^RX<dC-$7;*8<ekE%!-H#}xTb?E!_PNRJhIS0
z3CLmtB_c}=l!PodP%^U8Kq<&-1EnHs4U~qgH&8mV(Lfo<W&>q5v2~|2H!XRGv*~F$
zea<RtlO*2K1J0>UMEhuPbbtoO_!w|ZfC0z)Sa58B1;_a~a9n@`$NPA2e1Hch_yll5
zKmaHDL~vq21Sk0<a8f`5C;Mb@azF;B_!MwTKmn)vRB&oQ1*iEma9ThEr~7npdO!zf
z_zZAHzyN3ZOmJqveCQ5vp2}tg8!jhy;!HZTzx6G^;Zoz5if;dRG_iVweMEQbDEWGK
ztOvh%?YYG2u2R<l)(899^YaEQYx~}}iPgE@T)%Y${M5JE?dGksUQetp=uptjdJyZT
zc;a;6!5wkflU&Yzu<y8NAjZW2u`U*fb8$etiw6>10+8qufh3m%B)eoF#ianLE)_^~
zX+XM52Qpj+km)jk3b=xYcXTg1w!KM|_a2;BWzkuJwN0$2Ptua#1T<|MJabEYFA$jU
z;7fw()BMvb9{e)y!6Ny`<}Y8hhJ=?idYe~1Gk#pf1Yb&W;&(g#_a6_I&oup-e5~Km
z*R-~>=~w@h%D}_~7C$@8W^b0-+|eREqh*`ct%}+f<o9fwRy43()+6mZc9LXO(+zXL
z;~i+FhtRv*|F<?XyK|SE-0pdg_3GnfTz#MD&x>8TN7;UZI%E!ZSCPV}o*Y6A{V!e4
B%e4Ri

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp.51C1492BB7273708.idx b/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp.51C1492BB7273708.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bf1f37f91651cb4e4d6a9ac9116ee2903d94038b
GIT binary patch
literal 7868
zcmZvhdw5J|;>UB|8Dk`ICy7Q#&|=OxJC<|5mPjNM2}YEWlFA4zqZ*gib(Wb*REv;B
zbcLz9Oe4J!*DjW>rG0F-?e1E(7nCYXT~ezJyWQWIK1aTP{PH|`GoSPM&U~MF&pT(%
zIq~M_<~H_7NXVX?KW|3)oJsbCgoKW<-`R6!J=Z!0OmafPyiF&b`ofzocc|NPZf9{x
z+PL)xR<s@<RXo1r)#`6XJwAQGl;Sgadk&qQcQjZs``Cr|+83M+{Qbh3`*~mOd^g)s
zO`YBP=~I_8H#L0n@i%uaEj@K%*8OFrUu`N~boiUpJ$F9ZeBthi;J&l!fRgSRH`^4g
z`qAzlb9&~@nhRH&{`Jj+zT+4D@5K|TFYO-`JW{)5W&R(UziSS;k8D2wm*;Q2)7;GG
zPqY~$c4iEiy6){YV<uiczKfrJ!;z7id-d(HoAXP=Gp}yXJu+nK$UWu7%G8hZGG;IO
z@W1=d|83f}m%q5lec7Y@c4p&_Qf0!MZ5oHv>>`chw%&X0`lh+p4(IM~yKZ3lKQ+g&
zj~4c8*p+i=+iQszMx{0iDeD_YKao(EeQ6zEH!k?o?K@pMy?^iX&GyGmFKgquSItju
zI#*s$a{AcJ;_9xK$1RU8<Yl4VcBlQjetTc&^{IY7EO<LaD)#NWn4e*EIvJU^E!{ZK
zFFZH6<doX9<EaM=Z@+P~aL4#|W;Ii2pVV;fjIiM9-J2cWP9DN`|JL94KS_1V_P*e`
z?CIosj()n^Xl22Wy-k}*`(;DwvRCZwH2$*_`lhFog+Tg^Z!5JotNRSOJv{KcE{>7g
z-|zBi(V6Ai`L_>O4e6&lyDykC{qt?JdTn1`nZBe$U6wgJ+Fu{ame0{{jx6;JYp$yC
zJ5#*7_IY<d$o_I-!8DHciw@!V#>*3iE&Rjp`*m3Ve!ou#-`)C0YWe2nJ7%RkajTax
z^oJ9#ESdY}TV={c$BD!<S#t{#y|>1$==Qz#bGOg(Tv@xuvxk$48{aum($wfIT35u4
z@pf!0|MkdA%^!bpI6G_C<X4Kf{c6$GwDKpHl`>cUdhf>-!O`{iH|CG`zI)7Da<JF4
z&$KVd8e3d9v}o7(WqZQO0Y_o<RN`CPPM^E#{PQ~1|Jetey=Yape=qNU?x%Tk25ep9
zX;us7Kj<Akd1J;8fm-(K{PhdIR-RA4ale|r6pke9-u+6$9}itASU0fumq+*A%T<|^
zd)J;Sbv9S7c)8#v?dm=7jiQ(SQh#O2iRm?)2Y;U@6#gliym8O&%Jaj9y=mK5TsfwG
z&A<=R7QH_4)vljJ7ysw_lr48hdB0wH_SBToyY`IO_TBDVTN{rYY2F<<Q{-P*lp&eh
ztInUy_3w)aN8jDFvV7fF=L+v%Uo*4Ygp0S1ot`?2-*}_%w1L;=jCkV1gCF%-*mHKP
zWs8rWO*iX@zjL+QnBRW4Wo+ZExh3oBHcY6v?<iN(#{ZN$ZVbmw?DBMB`PP1yv|GKu
zYJETTg=@cN_X(B%`@F}r;UE5H-VZCA4ya9o*8lkP7EiL{GjsKhsdfMEeZNQG=lvf^
zx1Vo4`GlwcwLLTUo=muu{n$pKv1erGv;Tc@-r)AVhkxFB`PHf1usuhqV9&tNgxT|F
zl+{oP367ozR%O2-W+subB?<l%M@pOc5IbVTe-iIUz77}gqm#%}F`ePa=olX?-D_o=
z^BpR09!er(V!Df?OS|}B=|1hA)w>?~{pBPwE~e8QY3<{KrQPh5bq5bGy_}pdKBkAZ
z8d?+|ENwox^L_Nn{CAUxFQ(HS>8;~~r9E43u6^tEMt>4{I;Q(L`V_<mOHWzi_3U{k
zWl7tFQ8AsFl9?4BEIorhxS=GixG0Gf#dHrxk9P6F(yvvnPXF6&!IeacW4e>0lQTY8
z`rsSU#mf5geUr%OnC|T8+$KI)`cH@SS!W#Q-*21Xjp?kEtUmF<(*K&2d5Gn{ACyE!
z#&oJ9wM~4mbi%CiDYM(Bwd|YDiSd21^7w;D!$mbS*`L-ut*14&$$po|RcK8{vY+<Q
zK#Ye0Vm&Mn=iz{O4-X`G1R&8P0!bbTNcPA;ibnxbJt~mq(SUT14rF)?Ak$+4<#BnU
zwN$M2*gG-44qK+g6{1`#sg`MQt)^VB1JfajMo+J%X!J~oV$d_IDF!_oqFD6oYKleA
zg(+?o^l{Y`_d8%dO!4UXYKliMged{NP)!Nw#V{qJ7po}|y%eS-^ink?p_jvyj9#v$
zWb{gyQqU{al!9IjQ!0A3no`kgVM;@<RZ|*zJxuB7^=e8-Z-glWy-`gW=*=)?qBpB4
z6MZ39Xj+RT*<ZvJJtFHE)hRK4Y=R|I;hNvFVCV`Oh^b(J*a{YitKfk63LZ$P5P-xA
z5lE_#faD4pNU2bO)Cv_ytI&Y-3LVI(Fo4Vo6R3bI(5xx3jux>=cz76wNzhz5FbNtB
z#Apl<tFb_w#sTpf4<u*;kf@13k|qJknhd093XrO)K$@lj>6#8?Xa<m}nLxw1VY)R5
z)?X?y{yk$&!oc^om!?Gg$2!CgrA}9Wm<=Zl#5frs*2w~KP7a87@<4)901}-dkmQtr
zWTy<II29n(sRC(E4M=zDK!(!*GM(llvc6C&660T_*87jW?MmXRA}w#+wSl;5k#`bT
z6!{KuH6e#=uCNWBzt-leL0)Qih3qY!t|T<}(HjVjee_O3V;_Bo(AY<ZY&7=KYi%_4
z(M#<#_A!-&!9HdKVX%+cNf_*7?hpq1n2?RZK4z_r!9HfGoxwh~lCapvZXhi7u{#Ni
zee4~=Vjml_vDnA1wXxX8F154R$5jz-IqVOvmT()9ql9}4xruN;Act+-%do$=8XNap
z<dB^U!~Wx|2#<YyE#a|`j}jjH_$I<*A0M{y*vHq{c<kdtb{_kLDk5N?P)h{t6QV@G
zKB0*S*e8T-0`>_tHUayDkX^t&v5JV;C)N@X`@|>_u}^FwBKC=4n}~g4jZMTpF=Q99
zPpTpk_DQuw!agZVB<z!#h=hGo*d}40RAZB{PYT&3?31gAjD2z~k+DyX5*hpCCL&{>
z9Ja~WC)e0y?2|)w8T*tfqF|p=OBC!=qC~+yrHLrmr-W?^_9-<s1^bkcUBN!Jim2G9
z))E!_)F@H0Pi-P9_Nif;ihXL0O~pPnWLL3It0EfqX|+VdJ}pW#?9-ZvhJ9MtreU8}
zW7Dut3)wa7)2oP%eR?g?u}_Z@9sBeqqGO*Pw&~cX*VuII(?fP0`;01LV4qP-4D2(a
z#K1nIi5S>tglz`)88tQo`;3s?z&^8znAm655)=E(C^50mY$7K1nPHoWeP)f##6B}*
zH(UA!SKUZxzB}iil-=*`>>X#FC3rR-UHwV;Mq85X@6xr)W7cV8?CL8eE!{ewvcA(u
zV*KK><tr6msZd-yGlhMvno9O(NEzAIHB50)r@mL$Xsu$KbK2yzw|cD$*H($~3)YsG
zbTv?}{cuS^H&8S((?BuEYy-u%uyw&3I|F=pwF@u1g>vo0m)%0q$jlasL1wp5EHW3N
zxF|fIk5D|a5TOKQF+z#RQiPI_<p?DsD-lXTRwI;(tVJjdS&vXUvJs&SWHUmU$RoHB
zk4}?cEEsy49w&^9?eol>%(2$4NcLytWR111>VK1Wf8yE!?lBMdTV~E6YridBeBk9{
z7cL%*ZO-&%K4U%Cx|6^q#_u4&yn9{6*7Av+g85t}Ktb0Mm{QtHLzMQ>5T%1OMj0;y
zQO3tWlnF8zWxXs!Ssx2gHppU>^Kua7d>ll%Acs-j%R`j+@et*MJVpht08zmwKvW0{
z7!|!DL`9znQ86fDRPss?m3$IJrJ#gS*(*a-_Q?>HgEB@HuL4oUr$AH*Di~G0DnwPE
z3Q;wvVpQ{L5Y>DdM75xXQQfOURQKr+)q^@l4X*)F!)HL$2pSkQy(UCWp9xVjXks*<
z%NMM*p6vH<9-!ge@JF=kdU(F?;-*!x&q1Hjea2afx?$c2twzq+nj71k-Yz}Gx)!nS
zq-7<>@1ViQqs&!qJva982)Jgp5M4%tqXRTJri=l{1Q>8^84HdLu;92d4jdQY!0}}~
zI6lCG6Uqc|LO=i~mWklRfCx@1lfX#<37lLegOdX?IHgPhrvwymYMBa74XEI>G7X#-
z(7@?sIygO`gEPtua7MrYXO@}Z%z*jG9pSu_TjgxM9^2a~bjl#>F7xqX<63FYf1HeM
z9%UcZ%UVv>k+p4N{61;RBH-#N?f@JJ`Fe^+7V0ShS*)i-WT~E#kmY(xMpo)61zD}9
zRAjB5(vbCfN=G*8DFfN8r_2`q``&18ens8I*t_i^^?1_y2EK4<$&+h3zV}6JbCxg5
zZ@u|%2ei7|x_#=avCTPMb9!2HW8Ex{-SK>Qr?_QqTyEFEmXCx>bSW)FyJ>K=NP}bC
z3^+z)z_D%?94oTmI5!856FG3an+L~>JUGEEfD=Rkoah$8iJ}Nja!cSOQ35BsWpJ`6
zgHzlJI7L*zscsdVDyrZ#w+2oVHE_CH2d9fVIKyp#GeiTN={CWcq6sdC%XxHLJK3Mh
z<vyb8&W6=QdGF!*Vk(^~SSuj*e}TZHhyNs)J<C74^x=bv4~O`V&HwYA>9*Zd^A@+B
zJmHy%rIX4M;)!4EGr#<KI8JDJ7=Nu_u&ib4<d%p28IuFg#;{;Yq0QdP(c0N2DY<Qj
z_U-aJ=43z7F)@Egr_|0}(z;2qs_BOL<gdEYN^hZG&;M_0N_zK<%&cC09`D=V#kdFj
adJr%6;2vWK4(*yU%u`NsM-0xRhW{U9f#28w

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp.89C496A745057432.idx b/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp.89C496A745057432.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87e5bdee019d15fea17f4f9dd49a0ad678438ed5
GIT binary patch
literal 7854
zcmZvhd3;Rg!pC!-88MPrOA=y9P|KWiIxc5lRTeTK7(s>>F-aGRimkOJBvV6e!IZSO
z2#S#_y=|?nUR-Wnt&g{FS1;Y{%C!%#BKNlMn0}7={qg4W@nnAI_nrBgbDlZp%o9&;
zPR_Q@adDZ&xw9sh%qX(Q#l^Lb{F^>w+RO2gz$C@R&02SK)DNCCxz+l*)1iW~sfBAl
zTGF(SH1+8POIMyB{&eY_7Ya@e+O_}GtV5M!ryoB1LCd^TW#64$^=Qyfq4zW6S5l`o
ze>v)6`nrZMj-S7Cq3W}<(;h7x|I@nh^A4O(*>&gh4QKBit=xM`?K8Gxr<=|4SNv*s
zjXW{sX7KFg@W0MC_8c?s_g9amytc1j<-xb=mgoNc;V%!XT?aRu`R`Y5{q^BPK6jkW
z7#ix-=fySetr|J*;*lME>Gk+dDLGf(8@(ZStax(i)|`U_UL3Zoq(FJ`_@GYH=l$b9
z`_6nj>FVo0+~j`jTyi^o_qOrM^Y1p>Js`M)>@M7V|K)4zW?nszv#<G@z9rvl@dH1f
z+pA$m_Wmu)63z}!*)1fm-94g5-1^K5Yxwnrl@D&;Y1`(*`(JOccQ~=Infv}qzBqik
zBya4A!&3@Yw!2ukc>7#l7Fuj|*njD@r@ZTz`We6AX|-|c-o58?I~i?`ZJe|v&G@L7
ze`e)^&(!d?QH^tNZ@M{T+n5&SN@j??sNwWUVa}C%H(R}zG=S}R$=CDWiR%~cDR*CV
zw{gBqKi7VQGH1Y^@CMRy;ehp-OZK!G^VLy(-E&DoS=#kW0qxzD-3QzrT=q`e_+eW=
zZ2M*Y$;H~4_YN!?&`Wo8oHL{J>n+o|Ze1KmThMBKhB<wEZ#|1GnW5brHr_k%;i90=
zk?h&A*R!)R^T%;{lQ`NZ#tTQ@zWDsWxqts#uU2b6?Dgfd_cnh@E#9zr+qC2!x4Ifx
zcaOfYVCK8^6P0uEM-xtF%*;#h+#0>4{SEE6_FoNhX6zWl9!M<M{nw*o!@C{%Yx22~
zp4QFfZw|ip@c0i0GBb7*zfrK|nR!=IOZqPy&s_fI{;x|aNBrf{+qq*r?;rMz{kZE(
z6I$kFj4oK8mA_-m!d?EPviKp}KTD|Ja^mzA$A7I+eP4aT+4EPl|KG*EPd}J7qtE75
z?uTmL?8a{XW7j9&EqjaoId|=xpOshAu0L8yU+`~?+qv_NhEMlj&Rf&B+mDC#-p^5)
zV|!MAHs0|tu;lf;o3yiA-%a_i4Lxw4{>QcZ?Y~`7{7vz)<11yl^g+c8PMMHbQn<B4
z?_Em|hpz4^Jh$xoV~1Y3pWEf>OFu1MUMB2O8nfm+)A)@0+O>CFiIZ+{dw==1>h6d+
z4=T#I7ZNAkZ9IOuaMbDbKW#bG`Qh5H+Wpq~#JnD(gP&J!?h(xYwYcxPcLwkL^-rUR
zo$9&lz@2xGUtF?bc=qhyW@k5!+c)y`?9=T+Q-<yJZ(2CNYGiiih3v!M74%rr_lvv6
z*A*M*Q?;GHnON$3;JUHtpWMF}iD!D%eA3PI&A#<}{ihRJ{papWb545A&9@F;p8w7=
z+pSmok7{^xwb3|gZc-m(%c_?%8>u-<PEJ{ra<9*k?iC9bSLRIT8q+E}lx0nyJ$Yh~
zii;zx@7XV!H{NO<=NTL4%Sg`X9vvd*i|9{$mycFtZW7ZI$>>O(9G~1QI#_kWqm4iN
z^Y{^oWK^W?6yK?Jbg=5K%bT5P6}Y)Sk&KMgZR6Xvhz?fWy~Rs<yMuqbm`DmEb!vQS
z%jjU$X5*b3+b_?4KaqGNby|E{)97H;lj)CZ$EFtKCzAX~-8sH<i|An0Zyeu#Y3J`h
z9&a8uJW{78r)NY5t6mnUN&EJ;;7lY1k-ANM8%K1o>ecIiufCOVczROYm`I(~Br88U
zSoO!7wpS=?&-6?rBO>)v@lQ324px2jFRKsD{4nd6MDkpu?jGMgFFIKDzlzfLv)qk-
ziDX!$PKi%x79FfQZd%C;(`~76Y!af!WciWCjSc72^duiwPjUMw%aY{d>nR?yP)`Y%
z#d=D_EY(vIX1Sh{F)Q_yf?2JnRLojErD4|VDIK#>PZ^laddiGtYfkYA(YX|RTd5sW
zyI6A~$>(%A`^Mg)Q*<hj+|D7fM7wApX^{q!cGEyIE(S<OWPoJc43Mmg1(Fq6AXzsH
zB<JFQ<U|fg&dmYIyLcdZkq44@^FRtN0Z2g<fE3&UkfKWjQWQlXMYjl~<dT4tL<vaA
zEdeRJWFTcx22ys*Kq@W;NJUhDRNM-Xs!Ih@6;&Wrw+f`@(ty-N4M@$c0jaxmAazj(
zQg`b>8ZHA!Lo|Ri+y;=Q%LLLCO(0FT31l{xEj~F-_;R=$NQ1aRPm*<x?3NHcH^!#B
zvzl@)hflGyhI0NHGF?s4*wZx>jXhIMG1xOT6oWlmO|jUsH57|IS50x)b2Su)Jzq`n
z*z+|MkG)V$3D^rYlz_chO^MixHI#_GR82|POEr{)y<APn*vmDPjJ;A#DcCDDl!Col
zO{v(cHI$0IR!wQxYc-UHy<Sb}*y}Zvj=fP$8Q2>&l!3ijO_|u6HI#|{5N?PWn?U(o
z{*z=~H`^pcua_`oraEWCd2cEWiJ8hkVyCi@xN>e*?0I}S4@oE&Ac^H7B&l42B$vyO
zlyU`<TCPIU$~8!Oxem!FHz1kiCZs$rPqTK3btQ=GgvYlAuoE<AcI?g28V!lj7)Y$f
zLgF+I60h-)1WkY>Y9b^_lOV~O3`x-xNUEkn(liZ{uIZ2r&46TTCZvJfK;7C2eLvVb
zK6#_>aI47m(BbS2Z`(mbVjK)4*1<yJ92_Lx!9x-p0wmEPLXsR3B-tTDQXC2-)uBSt
z92z9up+hnp1|-vAK1tTi-qeKXO&%QW0C6rNu{Z9lCC;}nhlq1K<}h*I#q75^U$?>Y
zgEr@zn5*p0YJ04w1B6B&T}x>6(IG;kj}8+WeYD?3qmK^SX!Oxlb{c(5fH3G|Y6*ir
zCPWzYF=4`>kMY|W^f5sjgFdFp&Y+JC5EgxGEn(5eh6sy3HcVLbv3?tiJ~n7$(Z^QV
zS@dxM!l93=B^>&=5aH0rg$ai~&Tr$;#|3R1`nW1Phdw?)c=YkLghwAAB0T!|FyYb1
z`)xe>_@IqPA75qX(I*6mfIgv?2<Q_+L_nVqCIb2dzfC}&5VQ&C6RPY2`osVc(I?gt
z5q)Bai0Bi;L`0wHw~6QzgEkR;VwGJ)pA;Yx`lMPSp-&1C34KzSNa&ONHVJ)F&?ccz
zs<KPylLJIXpIl31^vNM2qfZVK8GW+fCZkUd+GO;}RdyMDN`NTnQ)-EVJ|#pH^eJJY
zpilAJ6!a-Un}R;2%C4YK4G<N5YAsRGr-q1%J~d2K^r?QEias@HQ_-hZ*;Vvu0ivN#
zt0fxxv=GtIr-g}zKFx2_(5D4$8v3*<yM{hJKy>u!wM0js9wIvW^f1xUr~7R>`t+bp
zN1tA0*U@JLh=D$%mKf+WLc~Cy5he!u48P4lpAobf=rgM92KvkZG0|t%5)*x9h?wXz
z!^A|N>9?8aGlMo0eP)&2jQO5jbUp6jy&2zAcAw`dPoZ^N;C}m1%@_Xb%}J84ZM(J|
ztoy|zUutq{nswteL|i>3nO(M$O7eA*I%QhlLkjZS^t`f4YZBR<-7LGM^^(?dO5}Fy
z@mdO8fa@q{2)~xpQ8Z?zj$$ygbrg%4tD`u~d>zGO7V0Pgvsg!on58;O!YtQOGG?WY
zQZTD^l!{raqcqHV9i?M7>L>%VSx1?ehjK%o+>^eVmvw?J6oy6idwO>IXlsor$(NCx
zG1^*6n?LB$xYd1JBOmYJ^z44t4z8;B==H;A&wU)(obFAZU_ICR-6SEpJ_Y-|fpW$c
zO_F?c14UzI8Yl)c+d#3gY^`r`3DI>ecv+9LAa=Lo^g72vs&p>EBkiF9rM)zubR~@_
z<6!`0ybPdBC4(sIVF6{mETC*9izw&e0Oh<Kpj-tvKlVT8Jv@-Smj{xs;DHo80+51N
z08*$BfD}C<kfK)vQmhbxlspoUl2-y!s*r$`Ju;B8R|Zn9kbzV@3XqCd0aB?@fK)vy
zkg8V&Qms&d)I1uHnpXo-tI&YdJvxxOR|itB(1A2O29So=0Me*1fHXZOkfzrJ(yTCn
z%;j<g>tmDTb8~J;gSo*^lEqL7(bZXNlJB_YpY1)jZbjsi)_p|xLTi>Z%=)Csu*sWq
zBAe4%q$OM5IIOkYjD+YKuJt03FI*FyCDwBzsmwVgmgtEzbaWXF9W#-EjwxfHV<)oE
zv1KfDToG3sGjOHcE3w2E@zC+5Jaj^l0G&`OKqnT7(21oYbW)K7om478Cl|@k$)z%M
zN|6GcQmR0w7OBvwr7Cn<kp`VsszIk0>Cow=I&?;n0i985KxY=2(3z#?lkOnrDQ=Rz
z`C8;?C)3IOtVQP|1;*9!UH<1-Wb<(Q@UGSj`fktmZv2w<=OUXsOP%{$7sIm`7W7}$
z`hy=Ln=`x_KI;sBsZW!8O<Sfcjcm?tm)*sB5$kY~gy?dAY$oBQ+gkjO%-P2){}y!3
z(NIH_4<28vQ|J`Iq9T6_C@XsWhk@zSeACB2elYIw5dE?FmdbCMTUL6uHZ6XB!npB8
z6XT-9AMJ!cemx$ak3EdO)*o?VY-@4sp>J|=*-Md__regHy-9piN3+DF=B-+`$ZegS
z*`swr?tnHaPqj^LFUhK=8)pA!+R;ikp;wpxZ)<W|$4==PUAsTsv$vCR_4!jjUhK?u
UVEbmZOCIPhAvr^z9YhWOKUHauGXMYp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp.DA4EFE2E10A56944.idx b/.cache/clangd/index/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp.DA4EFE2E10A56944.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd5d15c05893e9d71b07b8ec5b0938932ad471f0
GIT binary patch
literal 9184
zcmZvhd3+Of-p4z?gc90v3I!~M7ErEChJ`XSfgp2mW)3lwIY20-Eubwz0Tskn+DJqc
z14gLq0|XHam3nb`Ex73NpuSiik6jkl0}xbL5HKPNtnP;R8uIz$>FdSN_q@O9JDr&{
zlbO(jv15<+$;y%|CoG;@wW!>km6g>o^>5*#1^2X0fj2KJYw?DI#^<2|e}|3nV~NtK
z-K=$QtZY4)uf1*A6RW?z>$d8LW|bZpzw51|i{GiAy70h>SKBK`YyN&>&9(7gCSH-V
zS5rr~d~BR9+R*aRhhP72s^R?;3$86M`*K6s!~4GOy6cAzHl4V9uzv5+5re1pD7?@{
zf9hv9Gx_km3$YX5Cja&I^?p+x{?7vkyFU2((E9z)##c@F^VRRJHZuD+9sj@kF1~#A
z>aYpZU3c7>C>%U{?MrJWPe1+cj$zg3vkSY9J@b;eX~I<Q$P?Sf?w4jy+*MUNeD;Uq
z3l~28-mkA8|J%K1AN~A7$rnAVE*0&5p=|iH=iBU-VmrugYs-~;&TY8=?7p$Dw_Q7=
z>Yst^aUU%0-?C%OTU(#ZIdNCl-E984-IH$5+9;n|J8Yv>|I4KxI(K^Q%BP#$w;o>J
zCU|A_u*&4IDrM^71M^B(cR6jX*uHd_pKZ6z<NmJyp835$9(g>>hB|Dn-MjbXgu**I
z9ol^F)`C0U=pVkne%bpYk}nw7mtK1Ag7(6ccB58%HFtT-u_Nq5XD(mp@KT;c_xPrw
z-@kJ=F5fdhcskf=;5~!x?mlVwL(-n)CenVnv{7EUr_+>A4vyS#cOF|)aQ>S};Q7^k
zrAunf|8&ltxb3yhAL~a}1dhM7?=h+WNKcQ47FB<`bwTfKD<TEUI&3T+wQ&1@k)!FV
zMS%+w%j|JiAB$CZ@<TiJhIU?;znHGvTQaDE%VytQe|p-urGNfY{|@V3>;JKMdCOm@
z6`NMPups~Ti@om{{o}#Mm)-w-^UUEVvk&GRDZXFH30*W-cK<%`TlY`K4=mm>h2EE2
zy8GpWQ<J+r`dYnYa;Rfl|7ZIjy!zqi`{d#sm5-Nh^*wy1Ta~c9%=_(USAJeuKk3D5
z>nBVJy>cKl_08UMX0%s|&C-pd^&L}|?+WMDWNX{s&uQLz_}CfGU)PSP_~h*pxBgW3
zpH>Vw_RHc$gSV^+ULBz<x!xyy==|ItYo4XQny~JnuZG`OaQ@orL8ro-vv%%$yyY)%
zeXFb;(&vkJ_Ffq~!h2}X)9;sgu0~cqs$3W}u+Na^^at<U_w}rU)v-<D_v2aZf4AqY
z-?cMxeB8L_U3*I-lV4mj<n3+`|8e>gT|U}g_wRGF;+O9VeYNW7`?DtP*mdXD?{;3?
zvU~sjt2;Lz(JPkfh5V>(j~zcWwqoyQ_MKNYtg2f3<uUErxi#~;Pdj<>z~R{ohOIx}
z@7^J27u|XLdvAWwcWJMMt(Mold$eHGi|Wf~x=()QySTag;{8+CZfu%Xdo8<aM7Jrw
zbhRdzluYk@w^p^K|Ea*mK3}%J)^+~bKhS&G$bUZVd9V83Ba44rm3(7Fa_G9Be~Sn6
zvOgIWePQ;-pZZ+uS@YZLAMls%Yh8ITIN<EAd3z3Josw@|&+hKEIWgzg2NsL%`>3C`
zUU6o2$+%tbQ1t_9MrSQtGIwT-%F4>_^~O{3b6im_F{fr#<Y(u%NiR}=G16Z-*EW9@
zR)$T=B}PgYW*2r$FC5)_Rh#1-A{XAuC6iOSb9U!;>4l^Fwwp7u%l<!|&Lvh#cgyb9
zKD}`CsOvv`zx~@KujCRtr3<nPTBjF|4lPjntysaWYnydfN*Co96{i=Do;&Exrm5XZ
z^<1K-bkFRb?a~WJFCTog^rt1i9?Z*{lG4*#P1n;4M?V>f7X0lJJ201&rgW$5PM-9_
z(d&O{zS^_L`hVn-yHmPvc3&mEaP*tcZLb@??s&gkGAX5R$-bpcdg185CRSgmeCD(9
zZL>ltU7TOsH@$H5zsif=qD#IXnoB08bl2>zZPE)zXDz6jwXmcxvv0gP>3y>5-RqlM
zPL3$btLWCFTQ6tZ=2dtZ?-1wtrYPvuGDI^pI66p!^D#bfzMu~r!!Y2OAOntNSa58R
z1;;TQI4;P6;~5?tALPOL89z9G&<{>v1aLx704FjcI58-KlNbq{6qLZpj0{c=%HR}6
z0jC5Na4MsMQ-dlvjnTkqK@FVF=-~9A4$fc<a7NGoXEG)@GiZXd7z>;gw7}Vn4bBeQ
z;L1wMxSOwzikT%dZ<6zBAC!}Rm1hn(ZzJVhMP&{vZ<O->5tweIX!LZHqS5;rDIa=Y
zl=7ix8Yu=n6QvmRY$L^@XQLF0o@=By^jwtU(DRKHkDiZGJbHg4<wx(2QhxM8BPF00
zqLhGMY@|f=Vw4imOO2F-UW!r@dbyF3(aTXvMz1td3VJ0<Dd^QkN=2_mDHXleNNMP`
zD5as-8z~*V9;I~jMk8gQH=>k*-fW~y^k$SY(OZp_h2Dx%7J9ppveDa7%0@rCWcH}c
zAv>pJ&P{UOf%0?G?>unG)_Rv@4p_RD2J+SVfS6hah@H<a&ODBr&jIoCc_9CMKafx>
zz)ultMIfnG0+MTGAf;9TQfpNptyTllYjq%Fz5#V}z6oT_w}9;VHc(YbRlwPl&U;8|
z<K1|3hK(2Sj>&8UIzR*Y0zM!nzyPrU7KjUQKzx7)@(27tLO=i#10s+VkbvZX45S1U
zAT^)@X#ovL59mNfzyLA>CXf}dfb4(`R9;d((%E=JUfoufzq#T-=hS<y$2$Nvmxl)O
zd3->OhXG<eED-16fOroN<oEc21djkDdPE?}BLT@C8A$OcK&nRt(mWcF?$Ln^j{#(Q
zOd!i+0ofk=COM~ftEimxc^w>H5#oJ}WVV>MiFltyP7v>Q<RtO_h#Yo#A9cawV=nJA
z$PI39qdVj22%)h@HxU|pbb`>>qmzWj9vybk*rQ`E8hdnuo5r3mLVVcsH4z{7d<o*i
zo-av!*z<*5KJ58oE+6)M4Q?Oym<VC8$21WJdrX2b*kh7}!5$NKG1y~bE(Uu{gPXw~
z8zC(A*e1eak4+F3du)=h*ki*k7JF>W#bS?baI@IsB80;p*F-q%aS6g<k4q8`dtBJX
zVULTsIP7r^ZVr2Vgz(ton+T6RK0$cw@kzpCj}N<e?C~)dk3GJ@&126WA%5)nn}{EK
z{si%3&z~fI?D@kkKlc1Fmmhon2Dcx3LWBs|6Pk#CJt08^><LLCU{46U1ndbhmw-K?
z!7X4<j1UofViOUuCnku9Juyi{?1^EQh&?gp60s*XxJB$q5h7tvY9bQ$qy&+$Cnbr5
zJt^#xuqVY_685A9w}d@8LS*d8O+?0?oFFpx<Rp=?Cx=}!_T-pL#-7~Zma(Tqh=M((
zi741p5=6nCk|YZDl(0*|o)U8@*i#zZ3ii|pQL(2s5fyuCf~eS2lSIXy8g{AJQ)4a_
zduoGQ#hw-+8uqj%qG3-<5Dj};l4#h|!Y&PaTFj+kPit^%*wZ6K$DZCqbnNK~qGL}_
z5*>Sb*rj7nkGXX0=?!ijdq#v9*fW}lfjuKZ4D1<6Vqnh*yA13ZF_(coqrq)p&x{Zg
zdu9_cv1cZTi9It(OzfFqmx(<y<}$HoHn>gfSrKAk&uStT_N)Z4uxBNSg*_|mvan~x
zTo(4M2DgPhJ3?&i*-ga8o}C~z_Ut6Fv1f-}Humh8%f_DF;I=b8@v-w+S1&L6C*`iF
zxuwQ(&Ig0*--&({KHrw)Rdnvs`BvwHN$P4%ezyYWOk3lgo|jKQxthwWDC7%e=L5HM
zsjF2^`VtqMDYa1E*YSgP3q>RQS|}ef(?T&BF4a5rJF_OxDpf0MQ`X*jrp}eK)M@{X
z%VtoEQ{DvD;uMYSi&H*iCQdQPY@A||xj4lk^Kptt_QxqdvJj^PWHC;O$Wok=kmWcf
zBP(%AL003Gimb&c4Ox#<I<gU`3}iD-naEb0vXJdKWh38Ha?j1rRu3qn4-c}~iK)Y(
zXiSmm?BTqM;xWahbAj-Y@z<t5J(!t%V+R$L4Rv-<L){yX9yoFG&D825yJ&{<SVyPM
zr*B-(boQHbDO1QvU&3@&r#_a1yrr3wV6W|+3RLf1hI=|h!zgXjFiO|cIO+@eVAN;(
zVANOd!%-&0z$jxgFv`?3ILd}t7-el1M%g;{k<2saLL5wTHV2bj9S4(qh=)nu=3$br
z<6+Vt^24Ox_QRyV&JUA9NPtPf7GP4S7w}gSLn4fdwg{tQy@;byNP<zxmS9w>mvB@L
z$uKJ0GK|XgGL9-C1x6KHfl;Mi!BI7&!l-JiFsjz8II4v-7}abIMzwkkNA-{nqq?ob
zs9vw*s1Y(?)UXX0HR=r517^sCNz*oA(yTLK(h6BHY1tM`T6Gpo+94YzZQF)PyUvEm
zijoS}c^&3eRF+f%%_^C7lbqv~<fLyLWL}i`ma*epc!_;-!&9l(dEZHWE$1;Ui{EZF
zaqgC}snrGT3i6%L(aueW0XgZL4VfS5E%#PBk4?SmtG)MSh%TqW(bY6KU%3yQui6KW
zDQCbj)eJbcoCU{Lv*5UL4jfm_f#b`0aC|in&R^~a=dbpI6J`okncqmL5$0t`oGF46
zYeaC;ObMJ+BY~4=%HZT08JtqCz;jWm6?iUcxe89LR>5iI8aS<51E-hk;Ph%8oKbFo
zGpY@6X1NK@tTw?}<rX-r+5%^n+u-bK`=;Ap5~^%9X3M$M(Um_af2eZ<@ZHio&X)E1
z$D!2fYIk*S=NNb1sXFGQZ((H)I<}c&GiMrk6<jmLA@j`?kL+)z{K!HxB_NB<l!z=f
zQxdY=Ov%VfGo>J_&6J9)HB%b0-c0GpMl)p~o6VGoY&BCBvfWJC8UFj8?R|!=+;}pz
zwR-YB1?P<F#HnS%nvSo2o?2aO7gsnN@0-D`F1K#q^@-H#vMy!4oNepe2TYxZ-M9&u
z*|uKJJ3MoU(;N-t<9t92#{jV$3&e385YO>Ie$EdhZ~~CXi9ix30m+;Uq;Lw5%Besa
zrvd4l4rFi!kja@q7H0w3oDDRiWX8>VRp&e{Q8N718)wp82X$qgZMmSTru@c@)P)Nw
z7M9)kX8Mgq`pZ?3{b8R|1N%MK(z<fmjOk_NGqchYzuOtV|94|KE%R;qXZ?;dGixg|
z-&V}6teKNSWtQf0x5{qqX_K4RwnO`N6FQEOZ||5hLF&}?md@R}^ZpTmk#~#|d|d_&
y@5A=*_5ZESFX&NNRNTAoZT$ue^fH70FmxE#v*cEK$mlNl<APOW?49CxO8qZxErzK8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.80BE80F117BA699A.idx b/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp.80BE80F117BA699A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..362f06a72f6bfd4683b7e0d2e73b7a605fd504eb
GIT binary patch
literal 13170
zcmZvidz2I9!H4&KN0zX=fWWfcVL`M&GP%u6E+e;GW-^n@Ohh)^iLhu%Em2YV)F2K9
zN})so8*Ht_8=+fPiq?9qipRTFE4_F_KT(b-qU8%(YQ<vw!g_}N{ZY<g&OUkg?Q>4D
znGMOzQDyljTth=>g?huP%dcB@UPD8}kow<iuUnn5)M0u{L&FWbUQB&(>HEI(|2^wk
zy<D(dz4*>i<J*h3XV$mud49LN{q09Ddgy4^(GCBbzQuaW`9mMyw0rSYFHLxFN#@^=
z-Myx<NFJZ=)ux=k^!9G4_RaCxE3S5Z_b10yZ@lr$lCS=G=$jSWzPx7BU4Q<!=k-q`
zn{r#bKA!Q?{w??4`JWdaf8z1!-+y^W+pGiWuTFn(V(Z<1en5Nqg^$j>_SUY0&y+tn
zn62(VeMq=w>9DnTXcs>_+`44rzLwzTd*<z1bLFYU^Ot!7{V!V5@nm~j<LRMKOV76z
zJX<o>#O{*=kA1lAZ=Y=(vF;CVU7agEn|J=~u7{rtj>&9H9Ur@O{yhUWpE`W{)Y5xD
zd|}S}lZI^j>!n|fUwzEFeNg893(n*Hzx!Rj-zD$K_uRZPzvj`-*9K3RdGm(z_RV<s
z<@KX?K6uG%OOM*_Si9!kA)A|?TNqhz>vN0w-^|<n;MY^6$-fU~HjLQ*#>S<ejc%fE
z9`3Nu-1~gwxslgzylztX!&g3AJ2HH9+>MtWzj<_U-%VY6Z~wTtxN+6&`9JpVypmpb
z(_=$-v%w3ef71EdndYbSJ5Mg0!m86Yw;otIN4#nF&Oz(x;3wqW?Jq8D-hKR|yLK(!
zYR~@UAE$4hzPmVi_Q|a$K5uRO%f{<l|8m>B<~@&oaen)V_gY%Ek?*$f8R;JW*^~Kp
z&*2?+H%*HU4#Y-{+jengO1b;Y^RK$X{r7e(vLCv6$Yby9c<<dk*I(Iu<&jli?a$A6
z<<?h*v?cD^z3HN#)Bb^BmX*_fysTrymlJ<|@x_x;kKDWE_^+;Pjm>}YO?ImCbb9I)
z;eKcS5U}puuq*q>DSPs--=+@V+AW>><H`H%U0Yt-!JWw6>3HjJ<2Ty6E}XOdk0<X9
zmB{dQGoShOqX%O*{bugJxUPpQOY<AAS*_kjX4p0?o^$<B-)`$TeX#rLQE&c!vbD0q
zeQdNoXU}(2ue?<ceRT9!mu!3Y*21<({|iR%Ej{!1(qpsSH@tq>bMR+}hCa1o@3Hc+
z{V(kw{KTUZo0{6KpZ{$5{EoHvUV6ob@85a*@XAe77iTuF-Sp~&okw<T+R^iK+tH7^
z%DDw!?)ju6H}KbyhxT*w&R3r58h&P`IC|WJ8z0^G<|Cy;r$2i8)I!@e^0?NekDaM(
z`|7K^-j(flid}mgQ$IL!VDG>yxSu?IuyW?U>i&lh95^j@->!eWRy(mSebYOwm#;bY
z_VNE4yzb!2=39Pr!^E!FUR^Qk$6adoI`^or6BiwNa>46UnPD#-{$}&#tJQaYTPc12
z)O|O<{Ql?@V^2hdH>bSs`s<ck)^{W>Ske9Uv}uD9Hy-P0UNhl~bi<xsj{5yIi8GE9
z_da{-!jXs9{*K8$c4q4*Pv3q2hRZJd?w`AEdV8s3Pj&sL-;eELPOVvZ`@Z6ib<C~-
zBbTjJ{`vP8zs{d(eelCU2ahkG+;u82X|Q{A&!G>?TUQmXy>Zp@F4EA@IN|ZzLic*x
zo9Vno4V`U`Z3B7_^)p8ApQdl`Ih>CPmp0Q>U5{%VH>CGq>gK3L`R3|rJ6jqS*7e+e
zb1&&Vn0m+qn@3;2|MTaY=_PeNv~lQw-h-(}#BXbQHTz_+nJ%vD$&Hh>-h-(p4_Y~6
z_<^U6Hq-faJ+g7+;NF9&XZ`2P6TkiR#s`||g1R2lIHrH^!PKj)PyBMx$hlfG)#`eD
z<M=_n2UG7nP+Iu?7vEY2HgweW#I}i(dk?1Gvtity9{Q&IbTge(*R74MgL@CAez!1c
z=AXFh9~js$x31gU+9&oNO#S(?_CK=x39gwY>v}}vhylF^Q-61U>Y)#=dg0S%dTCt`
zYaBM9_h9OV)t6s!?Mt@4$0pp=`<N7-`_Dc5j$GW{(m8VM$O-1X)6!`btUmLutdnAm
z^^r}mfwL(#aIC<BV-*&hU9f|*D|T=W!2!;pIKVjtCpf3#1m_Z5;9QCeoLg{%b1QCe
zoWOzO6b>9O@Zflb2PX&uI6)D>c?1tQkKzI66};fQiWi(u@PYFwK5%})56-Xn!HI$h
zPE<s20U-b`paj4Lg&??~5(F0#Lf}G52wYeQg9|HRaFQT_lN1S@EXd$wMFtlUBH$uQ
z1e_u$;1opxrwS@KRZ+o3g($eF5(O6%V&Gy*3|w4@gNrM1a0wv+E}<mAX@Uk$Q#5e8
zpo7yD9b8gKf=en%a4A0JJ$EAPoX^icN9Nfx)YN;{?0W)R4PyNT>3f=53&i>>V4Fc~
z=xqgJL(dw7Mb8!pi{5S!J9>M8*wH%-;y~{x5C?jvL7eEF1>!{SGKdSkt3X`n-3D=^
zcNd5oJ!cROJy#$cdfp&BdcHt-^nyVI^g@9M=sgDUp!XDr2ff!IUi97q@uK$`#E0Hj
zAU^bdgZR<=3&f9JG>C{^ED#ZWz#swifdUDj4;mzhK3E_@^dW<U(1!{ngg$JLF#2$T
zgwaa|k<d#8BB7TJBBPfJL`EMmNCbVPKqBZBgDB{g0#VSb22s(g1)`#l8YGH7S|CyM
zF@wa=#|k8dK5mdW`gnoF(I*U&K%Xd(1bWRN8hWijH1xVbbo6?G=;)IMNuo~{ND}=L
ze#xx9Q^rz$={YiQB~87rq|8%>`ATX-{r0o3qs-guzwdSH95|7rtP6lL)}I39tn2#7
zrrTg=OW9y&%h+IN%h_PZ>MZQo6bm~x!@`cuv9Po2cG%fdcG%f7cG%f-cGx*|2kaav
z2kaaf2kaa<2ke}>6L!v&6L!vw6L!v=6Lv1$1v^*D1v^*91v^*H1v|IyhMha*hMha(
zhMha-h8?GKu;Wr3?6?dEJ1)n;j@NnE@hKj5e1?Y|pW|UC=mPA7lmI&+Bfw6`39$3%
z9@u$O9@u#@9@u$u9@u$xFYLT2FYLS-FYLTIFYJ7}4|cwk4|cwc4|cws4|aat4?BO#
z4?BOx4?BO(4?9s8VJD_U*ohetc4AJ1T|f`OE|3bqE|3YpE|3erE~p1#7fc0V7t91<
z7t94=7t%wp3#CG^3uQvE3*|zv3+rLng;QbJg)?E;g>zxpNxB3(DJ8*9%1E%2auV!h
zU51^Ul3^!jWZ20$8Fmpp0=q~m0=q~i0=q~q0y{-lV5g)M*eMwWc1lixovN#_Q&TGJ
z)Qk!{HK)QZsz+fLO+{fB%|u}r%|&4s(_^rUrDCv)Wn!?4<zldl>v7n{Q*qeEGjZ6(
zb8*-u^aSh@sRZm2nFQ<-xdiMqU4xyL(qN}$G}vi54R*S&!%k1>u+uX-?DU)tyQH3k
zT{4w~T{4q|T{4%1-2#5Wx!2n*oeTMeK#Tds=g9ocN1J-zWa#?^x_sF3Mdoi^^FP&-
zntI=A==(=j(Hb(pLrbSkv;naq3uG7VKn~FX<P@DiF3|<#7TrLc$N}*p4<v{JkVo_Y
zc||XfPxJx#ML&=zia-G|02CC1Kp`;%6c)okk|+Vmq6`!fBS4C%0I8x16cwXDF);=d
z7vn$)F#)8B8jvpPKuIwPG>@M*!+em<H;#_fUrU{R+Xx=zJhg71eUG!XM6ElKE7ZCN
zxk{}kkqySWo`HRlv2I7MG1l*aZFy?Lo~=Y}*t1or4STjKwPDX@FgEPjii{0=wi;u@
z9-F5u_Sh0-vBy>@i#@hVS?sX}!(xvuGA#Dk8pC4Go~L%~*-O-pJ$r@Pv1hMRJNE1b
zW5=Go$k?%GuQ7J)Ir7wjJx7T;u;-{y2lgCQ>cF1EU>w+U6d4Eh95u#)J!hUevF9vN
zC-$5b>cpP2N}bqq8jKTr&LZQ)p0ma{vFFND7xr8w>cXC@LS5K%RjCVmE`xDl&sAhx
z*mKnw7xvtF>c*bCMBUhPSEw6%?kaU-&uuVn?754K8++~=<HjDBryTaU66LVRRVar&
zu1Y!VaR$R-k1H}9_P83uVUN#K9(#O=^4Q}ml*b-lr9Ad{gW<8q7a1OVe2w9;C*-Ms
zJ)uMe><JYrU{9!00egbM2-p*fjDS6%#t7K+<f#XHo)Yz7&r_iu?0KrxgFTPIc(CUw
zG9K)CYK#Ya-aPeU&s(Bi?0GBHi#>0Zda>s<7%%p`MaGLgZ;kO{&zGk@?D<O6hdp0~
z`mpD#QXlqw2IIq?ugLhY=c_S3?D_N5k3D~h`myJ)P(SwkRqDr{-(dXM^A{OE_WU)*
zk3BI@MeK<sDq>HpP!W4#m5SIC4MxPCSY$-(i8V&VULa2c*b9_s0DFN74PY-&r2*^(
z3?_iRK#>VxFHmCw*bC-q5PQKA4Pq}?p+W2gt2Btcpuq&O7c4SC>;-E~5PP9K4Ph@-
zq9N>sDl~+>P?d(T7c!U-_CiG_guPIW31Kgsr(x`cOEiqVaD|4k7p~GU_QD1e#$LF{
zgs~T{F=6aUc`9K~Dp3i0QiV#`ld4q0o@6i*_M{>sVNa?t687Xgm9Zz6sEj?iLS^j9
zRVrgoHW(Rua*>g-C)XGmdyzbiU@ual5$r`OG=jZIl}4}^F_;MUB1I;Gy-1CTU{A?Y
z1$#<~D%evhRKcE7r3&^GgHf=j6d473N{vylr{<}OJ+(wt?5P#1Vo$A76?>||sMu4B
zjEX(A#;Dkf=4ljr(GrbfFIu5d>_w|IioK}8M6nkwGEwYBYfKb-u{@1oFIJ*4?8PcH
zhP_yo#;_MNm>BkAMJ9&5SdEEcFP^7y?8Qqoj=gw=#<3T#(m3|w1{23#yvW3{7q2mK
z>?QIvfxSeDCa{;N&;<4pRhqzF!eA2EOB9&|_7XKFfjuoxHSB36s$oy7Pz`%pm1@}2
z3`WDAR%A5nX*EW}o}Q;V_Vg0fv8Pw4jy=6fb?oT|qhn7mGCKD38lz(`nWstYB}+7k
zy<~+Zv6rmUB=(X9lf+)K$Rx3stTD;HUSQKd8cv_O?mco|C%z{Tz2li}Z{+_mkhXNT
zwzZBiuMDVv3g0$zjQNH^{LZ87%zypz{`!aWR*Q8?-+uvZ78?+2VS(%xJCMWT0CHNK
zKrV|5$Zc^0aTX4Ux9~uMMF8?xJV0KH7szMv0r@R{AkiWM1uOxepd|<tvV?%bmN1ZH
zk$_~23>2|MfE0@Yq*_#<s3i&%v&4YnmN-ztk^s^y8jx<$fs&TwIZAkUT-C;IEs~bb
zalUaO^Ap^;+Rz_;wtafP`pq2!ItH6>C72hP)ZZ^UdxZ(SlTaqsJ@|ocnb?roGGUSJ
zWnxEml!*h`Std?oSDCnw-DToN=E{Ua=F5af7Rp3G_LPYS*;^)FWM7&1ko{%iM;6OO
zL=Kcm06ADDLF7=Ggpk8!5=NHFL_(I!L`II3Nd#Fb69ri<6BRjHCQ;;AnZ%IeWfDhD
zlt}_vD-#V_FB2U(Std#3G@qVs{*bJRg<rSM_ay5lkanqko_Xz1OXoys;ym*@q+674
zm+o*nI?n!pwRdplcKywdZ+P~jBTv+CZeP%TnYrC`QX1<&xOH^>TMlg-I@bJjuzrSW
zZ@tJoLz&-tNK@~nO?^L{fo>9b1b#S!-6V({>LwxNa5o7fOWj05mb;0J9O)(zWTl%Z
z$Z9uHk)z!tiX7`EG30nRi6bYvNdj5xCK|HdO?2dBH%a!fdH2+xvuCgC>bqO4S?mA6
zi92fpVzVreJ!=PYWF0`xtP{wUbpg4vZXhnp0r6QLNXQC6o~#GRoAm<uvOXYx)(<3R
zMW8@702IsyfkN33P&gX~lClzzoRxth*$9x5Re;p23KY#ofnwPhP&^w4N@Np2T2=$n
zvpP^Rn*>_KFFN-g@BLroZ(ML>*KPIZ&E(YNMdri0?}jJ)C0E@i*KZy(XiS^=PVfPK
z&We80eP7gXZnL&=<~5Pe&7FBXJ>i{~>o+euZ`maCIr-Kjzr8@%Qa)0@dAx7@Z1Z=3
z`P{5O-OgSR2tQXn#M<}cSO4^&huDzW9>OBqdx#y`(L)@_&K}}KcJ&Y!vb%@4k+~ki
zA@e<iM;3aBfb8iZ9%OG1@gn<rh!5G{L;T2M4-xzLqmPZ*+lM^(e*NB>xL{(Zc~bhs
z+3!^U!6SZAzqw<0#{~1fHLskkKRM4{HVJnkYjv8BV*M^;Z9pu`0@+zRkb`vqIaw!=
zi**6HSvL^JazH%G0|~4E<Y7HPUe*ibV|_q=)(<4IB2a)000r3~P>2lyh1oEW#7aOi
zD+5K?2#~@mKq{*OMcF7&jEw=s**H*wO#o@E2Bfn(P?Ajob?_bMK8pLovTqv|A3XaS
zYlL-#$9&Aq%VH-s^<EAOkA2#@93K0$4Tw#%K=!m9$dPscInz!cSK0;SPP>7)GzY||
zc_1Mz0D00LAaB|W<V*X2{AoXsm==Ko=>SkL9Rv!cLqOql7)VM>Kyq3JiliezN?HL@
z(<)Fj9R-S|V?gnA94L`a0BLCrNKflP$#n7@t-k!H%g$cfeC_JaYtv`HSbFyG!=KB{
zFB%$_ocogLlzx8fSj2JT{ry)gxol~A+46?oZ}7uE==~Nye7XGGq+vzhmz}Ft{B&g<
zV^<`Y^ZGURw+v`*8F>ESLFy1GbkUF|H8^y{u-1{IeEy54&zL!Dwtcu&yudSM!hgH9
oZOqtl?Gq<WzVJs=r`a4%7bkef^P^dJcz9bRdO4N#fI{N`514SoQ~&?~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.E76894F842B8152E.idx b/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp.E76894F842B8152E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d7f1457738c777f154da0dc994389bad50eec7ab
GIT binary patch
literal 13172
zcmZvi3zQS}p@;YXA6dfgDlU%&L10m&Xfk=uOp*bFWHOUUCdou(!<&eUN~**M4<GR{
z5-7!z5+K-8h@u<1Wks~s7h3iD>;r}BRjKrfaOJ9qH0Y&OP~#2TJM8bCb2-OhzkK}m
z^XzPr-DIXk6y*j_LxZ$1a@*o-mM%D_p`oF*{@;?Nw`2y?VNy#&!);q%Pkngh-+bqP
zJ!OerE?91wyXAt>?Zx{u_qJ?*wM$w5{?p^1II{7`%Ku2;W4-76LC>!4ntS6LV-C#A
zeErOW%NvX2*-75Sg!AX$-!0d^IXZRWO|I|%?6`67-k0Wm_1_<SvvA#)H?Mx+t*?bQ
zKMAkSt=;(X<TrNR_sEuSUwd}<vy*=Ka&z017t>#z`tbPL2j6-u@rTzAeYfYmt@~ao
zf4DDO-F4~%{^t3E@7$c2`0`Neyj43}f@`+T*tz`rlXGV-5CZ+j&+B-;y{+-opcmv<
z+X}*c8Ed@j#JQh-boV=-tr~Xs@7}v9S9&?`{K*4PJ|7&JS(Q3^;o6zo2CO-G@YKoq
z4}J97w7*@{y6){OzZ!kZr`Gl7WgfZs9M1pfqxpVU98k91wJ5**>CQa^$6S8b%5!#3
ze)13Zj@b106?^6%v2DI{`CnVtG`%uAJZr@(bGcv7=z9F06XbEf4Q5sjTfcYJ{Le--
z(RUAa*e~DlYWS7m%T_JDDD=@EKf7~y=*XzsuRMC!h~UmUHtx9p<L2V3#ZzZq=iPKY
zU3$kegSyz@C6oTEbI*6pFXT6!m_31wT()M&i;Jd-cTC-M-o13-Zl$aJ_1Vo`M-M%)
zb?#bw_QzkGx_?qvaop4sYma|Er13ASmJRvK&u27mfBM+@?ZXbVw5%iF-^XR-ZQRQz
z^6kRG4G%V579AMSE*Q0LVrD|Q`@2{F<O=oQ(J{yV!CkG-?B8(UuiKYh-+cYy#b52p
zPyXYIKeo2TALv>={-?BmV6bJ;W!Ej}8207ZUrn5NQR=CO?mPO+>xXDFUw@aqRDB_R
z>9wJL-~BFN-LZ0O_NkNh<lCoH2Um2<C!ag<uzllwZ*1@!&uwwM_s-~5wvCreTmRgN
zholl2diUio{p#s`+8w{1{twT_CoA*wt8Ts}@-UffTRC^yvO&Hs>nMGo`=$%t{p~nw
zWrO?E5&E?4-(PzD3SByM<d;{h`|FCrx^VxCN9-uQ^zWr-rnql=^PsTrCm#&@{lXoe
zmOtI~#;$?8pB~%P)NcL!Cqrj;-1*Rz*RK3w|N93Qt-f?_X3d?e|1@UP;SH-d^!(Iz
z<l~Ly+^n0QKe8#lf7|Wjj)|Sq*StA#*2~qduD8jYMSG(2w9Vy#+4y6x54!H%z~N<k
z^wSd@)isNcJh8(w>-2_q-l}dnaFx*Y{<pC|J5K*SdhFPi$Cmu~6u0-&cmMwN_FLmu
zTzhM_{P)_Suh%|(;?@<z+`G5N{<%chy!7)QZCvot!VNEMd3~$fx!{Y!gEPnfVrd~7
zJ+;qz;MU(i`u5(pcTJZHPmCJ;=r_Y=zQ+ZZ|M;<o|MM@UTX#Ko)z`nR^-JEdxpn0J
z*M4|s^oUq`#o@R1i-(r)**Ejxx4Wa3#`mY}+i~6Te@r_)dQokk%RlwT?rZ+kbm>v=
zzil>J-&oB}nRmk!RVUvbKe5%WxFci6ebwIe(67Efx4dO>p<&7Gi?7~D>JJUiEgLWH
zd0=oeoja$Yb6n%NMDL+4djB+zdG_bhL*DjgI-{=J8ruf+9!%YIdfUOg#$VY?Q*}M6
zaa3#X!PKpftr@Xw*XOS`(<|zFP~)Hhy$4gbebTV``_=a>YH65V*VFq=zoPeG>T%~S
znmqKy-ydnFGwXVI<M4sK2UAb^_RHhH`Q-M;n(3^%9@#jufA7K6nZ1`>xLMmj>fDBo
zx*pp$c3khl)QhdVe=%qH^h7gF)b;4b(dYFZOuc>Os6RjPP5FgpI<2mUG!7ZqdocCq
z3)-J!x#ON@nyl+#jl%}?9!&ku7L}RXe__9K8>ZKFdt3Y1-h-*XKR@-vhc~|VNi)5&
zt_L>`9?*L*b;B*!T)X5ATOI%R+|Z`pb5eNa+ig1!Pi$}Lyn67}bIdQEd4?`->OC`k
z|7iGy!^fCk!j?{}YSqlo*V1WIZ9uHb0@+nNkVAC<IaMc+OLYOcRX31F^#E}y2gIv9
zke~`cUeycaQ++^w)ej`9B2YjL00q?`kfcgLAvFXft1^(HDnMa145X?mP(+OYMb#*f
zrfNVjH3k${<3I^D0i>%sP*P0-&E#f!%~NjP>!haMdt2Wrw;IIyGtzgOtp#HJC9usP
zHuSavv7u)T!lGvjghg*Rh#kGXK<wxp263Qw6o>=8(;!at&H{0wcNxTm-c=wj^lpQ=
z(Yp)8joxDr4|-35c+hhO;m~sh!lCC4!lUO4ghwwJL_jYTh=AT}5HEUffq2pT4B|uY
zD-a)gzd`)y{RQGjFB(KdFBXW1K46dl`appM&<70?L?0}WAbQCl5_+jXB=jMJgwTfy
zB!pfzh>Tt?5E;E<5Cy$bAPV}hLBi<61rkQD8bn2}7Kn;IVvq>>NP$GqM-381A1#n5
zdd(mjdaXb-^f80P(8mfShCXhPIQn>j#L*`Vl0cs*kOX?&AUb-zKy>s;gCx->3nYnt
z9yf1_`5@8KIiH(<mdt13m87ZnDHt9zbn7&D#7J3Z0cET=0OhQA_mNGv!OoVl!OoVk
z!OoVm!H(5g*s&=Vc5H@)9h+leXV>kpv#0E^vuEtEv*+xvbLbA(IZ_VTIWi8|IdTrz
zIdv!OoGB;loEazVoH-}#T)GQ(u9ORQu8a$IuAB>YZru$#cghVrcg77ncg_tvkM4n;
zC*^^iC*y&gC+C43r*p95QXK5K3<o<d$H9)*dD!tO9(H_&haI2eVJGMU?1Yp6J0T;$
zPRI$c^XgvMc~f54c{5(vd2?Ra`E(!bd?_F7d>J3?d^sQN{JI}@{*)hf{)``X{+u6n
zqAtQtOo^})Ga~H7oCv#s9)MjS6@XnJ6M$VH7l2(*55g{(3c@a!3BoRz3&Kv)CD=(R
z33gIOf}NC;U>DLuunVO^unT2EunXlvu#<Hec5+IFot%+jC+B3?DY^nXC8fYl$tbW>
zatiFidKh-$R2X*QOc-|ITo`t$uEI`DsjyQsD(uvp3cH9Nfn6jOfn6jMfn6jQfn8LO
z!Y-PM!Y-PL!Y-PN!cNmQ*l8&Zc3MV*otD#J7t>>~i=|?)i)CW4i{)ali|cXN#Zz(E
z#WQi(#dC4kCG-UB5~&325}5?-61fEIbX|v?p3-5bXLQ)<IUROMJqf#HDhaz}CJDP_
zE(yC?+^n-dShsY}=4J!U<>sCx^Le%Y1bXH-Rr3kd{0-}(rruvp&714LaEVq4o-IWi
z5G%4kcF_*x5FJ2H(Fx=dT|jQp4df9$K%B?{@gff-hysvT^aA-rACO=41Bs#t6c7VI
zK`{s<i4ssq3<1fa45Ww(P*@BDsiF!L5hFlRF$$!K8c<A(0ma2QP(n-q>7ovl6q7(R
zxEYhpb8f!WaJYW*&%Dq8XE{%;D{0?(ww9=M6LN)Gw<A}n^#roPSnp+EUu3N7k!y_g
z2Vh&C+OTIUQ5*Jb6>7tttx9d!vl)yHd$uBD!=A0i*s#auDT_U}L|N>y70P0dtx^_y
ztiiC@V~Y%nJ+{WM*t6%U9eef?wPVj-p?2)qtJID?yTRD8XD>2#?AdFK9ea*Ebzsj?
zq7LjiD%62JN0mCT=P(!t_8djVfjvi!abVAxr%voSOVo)yXN5Yk=d4mE_M8Ue#GbRr
zII-ufF;48c^3;VrSBbi?=c-T__FPr!!k)`uT-b9J85j0kHO7TKcb>Yj=Ppq<_S_Zf
z#-6)M-Pm&*j2nCIBICxMyT-V&=gCtK_B<u(!Jel=J=pV9sRw%=gYjU`Q)E2Y^VAp*
z_P9Ldu*a1shdr)BIqY#&%3+T)7!G?}k>Rk%)ff(Ye4g^y<4css9$%q6_V_C0vBw(>
zk3GJ~@Yv&P439k_PX+7=B`RP~s89iWLX`^G6AVVco={{2><KkSz@9fxz1Z`Xs26+Q
z3iV>oTcuv?c@4&kJ#UfmV$WM+yx8;QsSkU;67^xvSD`-a`Kr{1J)gn&u;(i>KJ58w
zj1PPMJoRJGU!s2O`76|qJ%5$@vFA4!Klc1Z#*aOJjqzho%u^A2Vu^~_6Dw52o>-+K
z_C$jbu_qQ85qn~d5wRD@(*X7YB^tn9ph5%K3sh+UdjW$9U@uT)0@w@Gm;m;Ic^brC
zutbB{3sz_ld%-FVVlQYgLF@&KOb~m)8WY5xl&2E*q!N{|Csn9~J*i41>`4YAVNWVD
z685ATBVjL;ry=ZxN;HJMP=$uD7pl?__Cf{|!d|Gzgs>N?F(K^9c`9R1E>RhKa)rv+
zldDw5o@_8O_T(ZXV^6LzGWL`_Rj{X&sDeGELKW;ORjOc5F&G7VN|8~pr_>k)d*M6{
zV=r8yVeEw~G>pA)m4>kwHkdH>!bK*Gy>N{QV^7Uf6?<xls@PL2RK=cJr7HGRgHf@k
z78w<LYK>8`7s=BI_97)3!Cs_7BiM^nX#{%_gNa}-Qe-07i`1A1_M&+j#a^^Tqu7g9
zXcT+VDve?<YA{jkMT<-nd(j#b#h#X@8uqjj)v%{ksD?eQN;T|h2BTq5D>54Pv>Kyf
zFP5h<?8QnnhP_yY#;_Nw(irw)1{1?xtjNT$7ppNb?8Wmmj=gw^#<3T#&^Y$uRT{@$
z++gC^ix-(V_Tn`rj=e;lCa{+%(FFDq6`H_aqDm9kOBhT7dx;{Gz+R%pB(SIFsg6Cp
zM0M=x6{=%TuTmX*y20q!(~FFbJ-x>0*h}VV5_`!KO=2%up-Jo|t2Bwdq`@Symn<?#
z>?LbVvac6d{YAs6lS>bfbN<g;veC^it=pUb;#}I&IizjKNb|kqmd@dA!$+Dg7{s<5
zVK4v3FCMACoolsNC-nUaXtUUWSPKhex7dLk76*{i;skP8TtIG%8^~kv0C5%$h_~=S
zf<*xGTD(9$ix0?e@dJq#5h!2@00k{UAju*Dg)AW;*&+ie76mA52?MDX6)0kf07WfP
zAkCry#Vj$PxFrshuq1$Viw=~uB+pXZyWz$*c5RWgbdK_klFT<}rzZye=(F{c`qgjl
z7|=1$d?~@aETOfj_i}{3R}uo<B=8ixR~zgmL1d|$NXVgX5<-@{iHxjt69qZkO~S}(
zH&Kxz-6Vn>?IuxVt($1bv2GGWj(3wda-y3gko9h&BPY8_vX9NnG3qa)ow*jH@5^E>
z6YF;Pve?SRhRl`;i)=3wJF=rp9LUZxaU#3Q#D(lG6F0J_OgzY3nQ+K_nefO$nFz?<
zGVvn&%EX84FB3nqSSBKJpiBbD!7>RVOJyP<hsq>`ESHIltdxm@94?bEvRWo8a->Wm
z$k8&1B5P%$A;-!jh8!=GIC7#)63BX)=*Y=3Ng}7Y^d$4Xvs??kX`Lw~>yNALa{CPP
zDw~$hvGUj%<{j%E_4ND=PDjU?`(1m7$K2j~*Rv~MK6H3@{pR*r?N^!G)hFZ&{rlFA
zsDH{qZG$c}-zKd;aJLT`Z$5CFf7ib`&s+%v-<(<Nt?<B|wE?kN7Ra8p139t|AZOMI
z<jT5$+*vn}C+h*?vK$bf<$;8(0OZYjfqYpXkU#4O60;&uAR7P*W`jUdRssrTLqKv?
z22!#LP&gX~QnM;hBpU&WW}`q_Rs)J<V?gn294L`Z0O?sBD49(H&Ee*peU0~#WBJ=>
z9p3u$`Zvh9)VMk3Y2A6-^Zk;GA6Du&k34T=oB2xci`=w@{p5#_)o*UIwt38puwI#d
z`O);4{eP(6yx^P#7n$FD?>+UKi~0M?hwC?w_KluuKG+@l*qFM$_3^*eZyq~qY^V7h
z{9n#~C;JZ^_OtrU9YZ_Dn4iSF=B$3z&zTF(;Jc8uI?eyx`VSMl)z?Qh)&|bT+rY6b
z3y$SkaCX)X&d%GxIamic2k!vqWS!ugyc3*@b%As7E^uzv4bIKG!FgB@I1ldu$FUqZ
zj_1JfEDw(7d2j+NfD?EDoR{^2^YUJBKGp}$$NRwfSwA>G?*}KcA~=y3!3Ed=xBwpj
z7i5Frf_xC1#7f{KUIG_lL*PPu2%OBy;ACC~r?3h*g;&6Z*)X^;9|otgDmax_!9~~z
zxCkEs7iFX1qI?vb#%ka+UIQ0nW8h+Z3|yRzgNyTVa0xa6F2N_j>8uV;=XG#NHVH1t
zC&6`a9cSMqZ0SsKsk3y<vU433A3yV>;V|nk!Mq2XPd@dZ0nc0;3isq5V(t5xsD7Dd
z53wP$J%mNJ_Yga>qlY+<ojt^f?CK#dWOom7BYS#?2bt?395UZScx0i62*}<Z;zjoL
z5FfI?hxn1j9wPR!d0Fh(rryh8;meY?UJVavX&VrmW`XQ!JCGyo0CJ|CK(4e4$enfr
zdD0#rF3kb)X&y*O3qanq7s!|P0r}H@ATcch1=0baU^)mSr6r(HIs_!AWgsQ30EN?G
zAT_N5MbZ(VXgUg{r8S^fItCO^$AJ>*1dyKAfs*OuS-Rz#8y1|ow0X%bolDYZKA3;z
z@UK5tnIAMX%scxb)A>1BFk-;$CtvKpaNbq((+jR{==}u$`UkzA;$I(LeRk5Yu<ygp
z#S3p(R7dUFICD<F#{QN8%`NAiKk&Rrt1OLgZHfd34I4aU_ys=y#7UDcpEA`x)GA&q
uOc?Y3-P$(t!cpyGFB*5rk1oB;=5V?^ymvG=f^~<6wuPhDP(=@@B=$d%kZXGY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.741D3DF3032F48B6.idx b/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp.741D3DF3032F48B6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3d85705d8b2b2daf4e3c9e25ad05d13031664130
GIT binary patch
literal 13170
zcmZvidz2IPp~v_48(6~b0`gjr=YnX#WG0W9$z$Y|$4qAOnuu(86Jfy;m8hs(HHd?O
zQYewY23jleMd+55Qnfy-;_+>(l|Fo-SA-*qXt_bkwbEj|VS9%C{&6{nIs3`?``ynW
z$xcFcW@(CYE#J@(T&CT&a``Pw`ZP2&46gsZ`Ic1~OC6@RG&J0{>$TJ;SAFcg=$|ug
z*2@LUjSKI)cwAd?XJ%c?o>z7&+uncbvd52g9$Ww4>AS3VT{PsGjk_1#@cQ_V7G?hV
z^aHCKi{zQ9p2Xyf7VqqmYu}xix$H*g_kXnC@Xk9gE;{{>L*Fgi`pr!n@Bh<3-EV&p
z-k96c`PsDB_iujWzJI^^%(Kr-{o$K?TW1_dpFaD^$t@53>9>hDUOjT|t@m~ve6jq=
z!EAN^*+b$@i-)ecH*v*F!>x-p>}v^ZdT`FZ)z_U_ICqKL-|w<T?a#HfHl7{wg8WKr
z!M!<Sjqg4+;P|I&|N7;I5o>?<-i^7^OL@mn?tkLB!060|)QK@$=04be)0x9(&n$lE
z(^qGGJaO>Wzg%^C+^XZ&Z38opT+&DI{pL6MzE^&vJb34d{OYGV-WoK1`km|h?3?z)
z8|y~xc=XD*79X?SyJq!=gEuw3JU=||?w1z|znZiA(QhZqlO79X){ofs&W6Qbj%uRs
z9&YER?|miw^2l2^+%hrr>6>4!85ue@_V%kz+&L<+?~cyBJ3nhKZdf^U?vFh?uA{fy
z@$``0Y~a$VpLe`<uK9)hj#Kj|v)bjGh8<WjOS)s`j)Cjwpl6lcZLiI5-hJZ8{ks-!
z;j-8M{p`-EyNi=%p4xKqt6`0Q-f-)%Ki@N_dCyZ{U(`0@qn4Jf<onG+Mt)Fu=~TYW
zeR%r=O_xUo`J)$)-Fii4a=Gi=D}Qi?`t5CBz#Y1C@Y5e`|LDU#w_ewL-O-h&_vfd*
zdH0)xTjTfd-gwzhX+Qr^%Zkf?yrg}^Hxqt&#T64%Pd>Ey#4oNJ7M=UryX+M8h4hqb
zLVeHu&TrkjepmL%GhFg7|4JRcyGuUv`%@2dott0Z&Y#TPXMgXn<2KkjFP*jR_op5T
zmdNn6(_j4MQwO7W{A%|9@|{mq7UwtIv`TxJOtY<DIP2CS-rd%*`XJYh7r*=1Bx_~6
z>-Z>r)}HUDTz9t~JaX(8S8o0A?!wk^ze`5#Exq`+($h0sx4nJXeefrThCIJ)@A2~S
z{jcvI^z2g;nwr|IU;SkG-1apOU3Ja+A3k{h@QRI77G^fB+4zU?JC1JOxV`(Qwqu`l
zmY<t{de7(WxdFcnKfa$=cD(s~=kRmWrBP!a-SE`Dcb_aBI(y{(GxKdXDPxB%e)?Qx
z>*>??f2eTxNu7J_Q$9I&VDEry_-kJ{SULA_b^j9w4xE*`cIuz4Nt|4pzT<;o%U2(N
z|HS_cT6=Is^Ibo>Z9?Z;e^@r-$DLZ&TGz$j#xFbc+`PA^FhgHI{N1MItF#Y(T`B$W
z{KI#?@$smWV@^iIcPD@B{LALM*0skkS=ROK<(ChP-+sKidG+|O(+zume(_^B#n0JK
zKJ?O=OGh4F^BX4n^tmmczwp2#>#x50`+w}Z<Nd|<J=Jwz{4l1IIkS5H&V9x0YnfgB
zM=n{T{^M`2eVacs?9opL9z3yhQs)`}#6hl6-G@Fc?_629`Sz7dJ4r)B<M?Op2|na$
zYo>D+G<38!w)XEi)Xx|_f13XF;Ng5!ysDX|>hajdv4eXKW<2<}n?~Kb|EpJ;>6P_(
zNaK+HJqI(MG;qbV;Rl{S)=cNt<B^Rc2lX7xc*eiKIr-}^ZvSmFomY=XH;(Stb1>uU
z@A=!7bB9Lv8PL#Pk0-QFnACGH<CWHDf4*Sk>_jt7)Z=lD;|BH|%=o_RuYGL5!2vHc
z(}neTQsbmV&%uoMtRMTw$G<Dr<5~51SmUrkJqI)X)7J3DP2;lt2Q<vC$8D``6M7D2
z{MC}S-?PF=zL_TL@rcF|{d*2({Pm^miQl!qk!xv~Uyo<^oqc7`!HmDZDE0U!H@x~q
zGrg)F4{aRUzvp1a4Xc)4bMx!A-p6KnQ_o{kc=_KC?mK!#TT92tF(b#D_iamum9;v|
ze{G#as~1SMMtjM|+MsC@ZP2u-HfXXe3r$vJp~<Q&G&z=oCMR;x<WvrtcGeC}yJ&}|
zUA05g!8)Mn5FOBTs19g4Stm4|q7#}<)d@`(>w>0BbV1Xlx}eFkJT!TchbFJ`&=gn!
znt~`mQ&0tHimV7tQ52ymsv<PqtQ(qc(G5+v>V~F=^+3}jdZ6i1J<#;BUTAtnFEqWX
z7n(lS2Th;ogQidQK~rKSXiB04O-YrY>1X}W^oxFI`c*$P18e}A0Wkp0fEs{ikPSjJ
zC<dV!RD;kAu_0)N#1J$?Y6zM#D??KjWoXK(3{8bqps9!oG!<2WW|$2_Gc1Oo8CJv4
zR9O|8s;EL!RaIzetOiX@)S#)U8Z;wp1ey^s0?mjTfo7DALNh8xp&3=9(2TJ$XvV}C
zG-GNEnsGJ`&A1qcW?YR!Gr=aHnGh4uOsENH>Z}e;UDTnet2#83Y!aGDF$vA2nuKP%
z(C)c##_mW7DWJK++zVvB76zMoUITki$5w+_e@1#wv(^H!{sP!$5F6&U0<mGv8id81
zEf5xS&LAA-T!C<y+YMsJ++HAd%pC@CVD2ao2j)(LI5Bq?h!b;{L0p)-3dDsuZx9}H
zzCd`)1%n8f3k4!zE*eC{Tr3a~bGJd<n7a$ajk(7l9?U%j;=$Z&5HIH50`X$*Gl&m!
zUxD~Amkc6dE)|G`x!)ju%>4!8$2?$=0Oo-L31A*HND%X2fdnxR86<>xs6axP%Lb7#
zmkUJ3Trr4(xl$ks=3#?`F%K6=7<1JiD&}f|sF-U8(J<EvM8iB{kO=0H0*PQAHAocm
zXn{mAj~OI}d8|NUn8ytg$2?vjam*72NnoBRkObzsL3GUZ0?{!~8YGE%vOtoUFA^5b
z=skff78YM1^ZoI9($w?DxNzYuoK?5Zf-^?SIu9sgy&foMUE51G-3B{b$_6`I#s)iE
z&IUVHXJN;tSlF={7Itiog&n7Ju;Wr3?6?dEJ1)n;&aT^GXHVH-XV2JSXV2MT=g=Ln
zbEF)wb7UN_bL1SbbLvjmIa5y9IWtb!Ide|fxpWuoTqzgqTp1VaTsar)c%6qGpW<Q1
zXL#80IUaU`F2GJm39u6~0_=pG06S3^VJD_U*ohetc4AJ1om+Rq&Yg0@&Yf|?&Yg3^
z&ZB!^=Sg{B=gD|r=gE0s=heNi^QOG8^JcuT^X9y;^XWd=`BFaE`7%D(`EowkNxB3(
zDJ8*9%1E%2auV$Px*vA_lpl8fj30LXoF8@pJpj8vDge7cCIGuYE&#ir9)w*m6@*<d
z6NFtb7ld6%55X>!3c)Uv3BfLu3&Bp-W!T9n8Fq3;hMk;~VW;Q{?39!OJ0+vQPRS{-
z3+rLng;QbJg)?E;g>zxpsk#a~HKoE%&8V<bb1LjKU4xyL(qN}$G}vi54R#Sd0=q~m
z0=q~i0=q~q0=uXlg<Uiig<Uigg<Uikg<VXK!7i4H!7i4G!7i4I!7i@HVHZ!uVHeND
zVHeNEVVBSouuG&8uuEhTuuJ3;u+w!Nc6v&Oou1KQr{{FoCG{lilBp!@l9?pzlDQ=8
z<_YsIys>WSm@mu+S|}{MK<39B6PtQII)WEa$r^;0Mac%lN-Pj3aX@y-4&;y=Ku*aC
z<dR%Kyu<?uk^m%1B9L2h19>D5kXP~o`6M5ZBuPMi$qy8e0zg422o#b+K(ZtQDUt#d
zmcl@)qylM@1{9GZKv5|Q6q8~=aVZXzkP<+;qyr_TB+wjT&NTB)llej5(fZq@^B)7k
zqnxMK^|beKww9=M2Xcj4_aIlP^%SzfSl2PIFEZ9`$Ti0L1F$VmZRoR=s11F#3bmon
zR;4!d*$l>pK3kEoq0d%hZ0KY2ltmv~qAdE@3T4s9Rw;`<)?irlu|<YOA6sKs^l^F0
zp^qz34t-pOa_Hl#ltUk9FdX{0BEzANt1%q<?0ITOpS?uw=(AU-9ewsHwWH5&Fn09W
zi;Nw8_8MbHpCeBl=yQ~)1AUGPb)e5tr4IBt490;zN0D)$&rxF>=yT?&6MfDSb)wH%
zp-%KUtJH};r@=VU=PWW#^f_yc6Me2cb)nByqAv8gD%6EOSCzWZ=Q0=<`dmfEg+5n}
zaiNdTQyzVMiSp><E0jkcU!^?yc!S~5#}^qMeSD4K(I@1ofIgu_1@s9ODxgoOQUQH}
z!3gLRij06hp~eX46Z2F=pID+I`oszq(I-}^h(6I^MD&S8Mns=jV?^}1^VE$#cZs^u
z=dMsU`rK9OMxWbY+~{)`88`aeHO7rTPo8?v=P6MS`aBitL7%5eJ?Qfoj0b(5BI7}y
zr^a~D=gm_u`n)CTMW44qz3B5+sTX}-gYlxzTV%ZG^VS$I`h0onL!YliedzO5s1JR<
zD)phyXD~kW`HGAWeZCswL!XqV68fYPmCz?usDwVLN+tA31|y+QDl!uKq#7fk&!4A$
z^!ZEFk3N5e`qAgFQa}3q2IEJczsUH}=dUq-^ab)XfWAP92GAF%&;a@ZRT@BFz+eLC
z3ly0E`T{j3fWBa!2GJKR(IEPQ6&gfeuu6mI3mQxieZe9VL|?GR1ko4D(-8VXB^p9s
zs6s>N3sq?deIbJhp)XWqLg)+Cm=OBpJeAQWm#B<Bxk6?1$yF+&Pc|4CeR7eJ(I?j!
z8GTBgD(F*6R6(Cop$htxDpk;@7>t5GrN}7gQ)-NYzHpv~(HAb!F#5t38b)8ZO2g<2
z8%!8|;UW`8U%1AE(WmCAiaxbORrIM9s-jP=QWbrw!Kmm{i;RjswZ^FE)ACe9pH`w8
z`m_qw(5F?YhCa<;H1ugjMnj)gV>I+d@-%|JNQp+!7pc$)`XW^tL0`mRBIt`0nF#tK
zH70_-Xr4yV7cJ2!`l1yYMPIZ^qv(qoOcZ_5A`?Ylw8liy7t7Na`eG#-Ltm^yW9W-j
zX$*ZagNdOpR%Bx6i`AGI`r>&SM_;@|<LHZ5XdHd<DvhHrZZL85#fwZFeeoI-M_(dO
z6X;8nXaaqS3QeFdQKbp=B@8BkzC@8ppf6El66n+OR7am)qB{EY3f0l4SE-IZ-C%U|
z=|x6IpI&2h^d<8&iN0isCefFy&?NejRhmR!(qNM4OBR_V`jRy!+3WLf{CmUMGq-$1
z`gANEx^#j0-^IwiFK&G&|Mvm3rDIs@u+iqF9QErWT1SpHKQM^hcZ{9>zdwJZez}Cz
zVx8RkUqGA12E<xeAkM-8*)4V;hs6Qpv^arW78ek2;eiB;03=#OAh*R0<gs{wycREz
z&*B4;EE164;s**?0zg4a5GZ5`0m&8_NU<nDVM`cDwWvUvMFWahB0y0~6ewni0mUtG
zpoAp>q+4{Lq$PQQ;-2j{w6a@@q@`o5cWlu7E$r;XkRN@yZED~8&F%f$2bs?e^P-`_
zO+6P2^?s7z?;`#u;g_<3E)qZvc99@*sEdS<<t`#4D_ulE4tJ3-vf4#dWUY&6$dN7*
zL5_BjC~~Ze#E|1%B#xZuA_-)@i|EM7E|TnJ^J1y`$7tuTl<K`(tYu=|19yw9Ol-((
znXt%QnQ+MVGO;5&%EW=}EE6ZPt4v(Te3|gbLYWB2Vws4@?lN&Bd&<Ov>@5>7vad{h
z$Woa|$o?|%BL~VPfE+B7AabZoLdbHN$jC~WD9GV52_vgzq9SW$q9I4hB!V0*lPGem
zOk&9KGKnK6$|QlTmx+#?ER!U1T1Zbde|A<!LvLH>x|8+uYMb0P$GirvrDK9TVUGEO
zb(i|D#oHbB_VYh?ZSB0dU3cd*>t8x@^x68&ZS&f$Hn$s3$zyy6w~VTP%OR~p#+bhk
zt)IBthFxZ!xXp7jZR)v1ulLD`Cw;SAe+HkwTCcZc&07D<{0=Q0wyX_^&9XpTmIJb9
z?LdyK1IU?m0=cp-AU?|j30VP1%!)wntQ*La^#FOZULarA2P9=BAb-{m6vzgEg4rNY
zC>sKjvoes9Re-|TFp!#6fwZg!6v;+_qS+`=EE@xgXX8MLYywEn>OjeC5@><2;KB#N
zk9?iKecsVs_tc+7lTwowm=Ei|+n(#2T=}q4zj^e)(XHmk#Rr60%lgU>e_g-1)!NFN
zm$JS*d-{p=_z&Kw-@K&Hl8NRs)O<?TUm4C{&k0YUZer~{Q`VnC-Nc5>b`utv>n0qs
zy_?vP9o@u%?Cd5^WLGzFA@kjYM;5w?fGl<s5!u~M+{m78;z9Oy6ECu_oA{8WZX)&a
zdr$uQ5^;0+X#L(A=N&iGe4#k<nK5(Q;732M-#lU7gbwqk|8qy*GyMjQ_*wnt_Tla0
z&HL889=87Z{rQVw&HMIi%f78tdi4BD_Xz6<x7jjR_4fH|W6idCS?q+Sp37m)o9pl2
z($=NsQ=$Iwq-{WKng!z09FRS22XdqxK+d!i$dz^h@o64NNDDw>S_E>Z-9VnS2gsZD
z0{PNDASo>Y`O|)&Kso>vOb3BN=@5{dmVuPC0u)Y%fz-4Lq@^{WNIC)(O-F%Z=@?Kv
z9S2IJ6F_=e2TG=s7iiV;>zACrwE5;$9XF@Xf3f)d;Xi+_Hos_SSajh_rsJz4pT7JJ
z*R`SFvPD-fPA^&7(DM!cQ|S2?|M_z1g-OG*-Y+{=F1vn39i!L8nLd3R`&s%ow+y&w
z&_Hdl9K39BlNJ~<V(74u7khnIOr17;#!POwRl3AIdHnx(YwPGSW7{T7oOJ1rrd)2b
ZJDj}e87GWlU7_Kv;mC5T=zf*N{ttCOh9Up}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.E3C0C7076B1A5420.idx b/.cache/clangd/index/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp.E3C0C7076B1A5420.idx
new file mode 100755
index 0000000000000000000000000000000000000000..848d751cc49c17637634a137e1a2867c1add9de3
GIT binary patch
literal 19788
zcmZvk37A`Dq4(!~Unrq1l$O#W(9&W>uyeAXB!(?}P8JB1<YdnYwqQ<{lNFGL<v<D*
zE|E}ziVy@EZY~u;xr!hnm(_(${8$t}5k7iFKuqOB5paet9iO!S?|E*Yhr|5xoB#X@
z?MY@PIkPF?^Lb96nQ?Cke5G=3>zo5-W@eU8{%N)v^1%~0dFjl|S8jPy{QcR#vmN%~
zY0b!>tEpvf{><8Sy|2rcFTL~eoxbZ|zW@09cHgjj+y9m>+j!YwD<0UsGgE!)*jKai
zhd=uE1&ey*fs?IgoN!p~>m%>{Ki+n1sTt;esjEKs+#}id|L<@AvE{nI*0<mK?1z?T
z{^;M{zV?PUPkHK|Yrb>yCr>=^-~%Ur`q!IQo%U1d{f~bC-nHL;_TDpo_QdP|eEOwZ
zo_}QU`{$MMJs<s+ujdZG_@*;Xdi02m*(>i}>bm;2_}v$jKFB1_u{aMpJ{x;z-Ks?&
zt$5h`_^Phu8hPXCJNF#&)*mkY^`EX>ed&L`RBQJi?dY$%^#>2Rj+U<!-#+Hr#BB#(
z{lRM=eUQ8T4^M3R-EqsWd*ST&*EZhTc>SUBcRqW7`HSEEZs)+WUiIB}NuhJW{hOa&
zcI>H_Y&+oYQ-1KX%a6M8duKhJ+dX^J#TWc$`PGXbOZk&mJeD#4dwl2j{(gdY!}neC
zw$;}^cV+HRM=ho=z82G-de`Ir$BsPz%GPn7Km7bp7a!@_z2>5`-@fE1*WFuhxa;e0
zF6muaaU{NMy|F}FTYt1-C+GUy$^X6i>3=SHxO3y4)CpYR3s<lFX<?IntK-H)FQ>~M
z^zB^tWNOLIw_m^Ymdv$UrTFKMzJBt~-Ui2>Yu|fk<)U9*dH%{@eJ#G^&imgzY~AWt
zmoB}I%w1!ay|<Yk-P2iTd2PqH7k?qR%o+O3n(I!IPZ*5;`S>pko`dd+rM17kWciO?
z+41Uc?mWM=r1VDR{d+p6{QQcaFJE=~tvk0L|5bXB^Keb!3tv7bw)(H@zje|{$B93@
z{hGJGSy~xNJozGbqVTYE;(49}|M{QJjdyLkMfu?e+B0AHSbXh@k@tfi@3}*J!!=Ls
zFum8lS@+Vf*Iqe$!{;_#|KmNkyZhvbOHX~|TlYU7+WPO&znO0M!7$gkvfc>XK~9<7
zmf3Xv3fs<&Ya+{xwa>ix{S6z3JB)7~71?y>+=-<tBJS6BfAg&Cese|lI{!hRJ?gIh
zBmb-aqtlFEdFD0C^H=?L#ZR`}_1566d!D*y*@O45U%YtT#&@ndA`!dz_Os91_US7x
zzgF0OVn)9D;_bgU_Qp4MY~Qi#tFyb`ykYQAs`1e78&7}bwu?5rYu_Bb`k9lGkB)cl
ze1W73PY1K1n+D62)9-z9#g|`nzH$E3k&jQ%jjyikzV9wm^5Y%9es+BGt9i@Lmp=*r
zQupy+g73b2^S#aDN9O0=dhstG-g)8aXPtMUGWg5<>mOcw|DFr4SZ#dpmhj)3mYZ7d
zeEx=W{;*}o!#6*9i&1~hpS#~q9RH10HyHfr`Hin$_>=Fx@Z1acMBUx{)*Sxbf2>Zt
zWOiLpy!Vd(^Q-=a_xw2j;rHhcJag+!%a4BLiBEsM_NcIQ#T(DQVt@UDr=L%}_Q``m
z&7zl2d;YF3zxTIIAFnOUKX3S=qdGeG7mH7P+xlNOsmq_*ZaytrK23;_Pu|;ei`Hih
z9J}HDbvtkW*4!b3J1gCp=0%k+-9Tn$79IP**W9;T*Day(^vvc}i&h=HaG2a@Ec{yh
z@olelLj2iFs5sGU7Oh#naG2`l_g;O}`S-l@_!4^7M6X!1;^2kDRBt%6aLN%s{mJen
zG%?XfE;@4A!eOepfaZwfPg(ouAv4j5Ubkx9`h~+(pZ3XL-}{e0UUcsgnw;pP7ae`j
z!eOd~SMM7=aeDfzhs?w#di|>P8x{^zz2(zqu6ouvvMrrSP4q<vUUb&NVX7+|AN)r8
z$mkhM=ou5ecG22H7Y<W>`!{#}&C`DQ?@MTAqBktsaK^%6s_)#k=9l;VWAN}2x@n?U
zE?T*4;V{+joU`u7ocTS|5_;xDuU@qJ;Dy6f&mAV-_xtJ-e_TS(p6J6D9e(h_VX9{u
z=bqQR)wlo8ro4FJ&!qd<C%4`G#!2gzZa(suBafZ_zg@a{mYda2f4>tFW<&dl<2Z1f
zz=6|pT5wuH3r@%B!07}XI6bEarx*0#44eU+K`?+baz=1Q!3fU8nZTI@6F4(x24@z`
z;CPM)#|u0-3uggm5iH=WoE4l^u!6I3HgGn<2F}4bz&Qj5I49=>=M<daT$~G>OK^d6
zb8c{M!41yCdBAxD4>&L91?LsK;C!49oKNt9^K*W1e!&k;-~@1jAb<;S0dN5!04~S{
z!3BjNxDXct7ZO6?!dw_!SO|lQa#3(mAqp<W#lXdc7&ws=!HI$hF3!cl#f3Py1eXAp
z5E9^$ToPPTNP<goDR3zv1uo5{!KH;XxD1y8mk~1HvRoEiR>*?OaXD}~AqP(4Byf@-
zfy;AwaCspQuD}(*6@&sfnUlfEf()+66~PsSBDfM)0#_1B;L2PXTv;fCt8f)?6`=x7
z;S_L+pn$7#Rd7|I3a-Z0z}18rxH?w{R~PEw8e9WhLui0&a!qhep$V?VwZOH67PzhE
zt=4^yPMa?=U$T#;kG7@73y-q<@9<|;GW&02|K0j*m&|?>m{SRdp6e10y;db!^x7`b
zqSvWJhhEnuI`n##=+W!DM33H}5(9ceml)6+RboVM>=Gk-lS)kJO<iI_Z&ryJy}3)w
z=y{d!==m<;(OXnvL2v023wo<ctmv&>VnuILi4DE2OKj*JDsiB9bcqAKQzcIH&MtAH
zcd5jM-qj^8^lp{7(Yw3EjozaY4|-3Rc+h)Q;zjT65-)n6N_^;jUE)LUSBW3Jzf1h+
z1(gWsg)R}$2UHS3ALx<*`k+dJ=!0DnL?2Q~2z{taLg>RP38N2pNf>=pB~kRzE{UR#
zsU(Iz)+I6YqDn;cVwZ^M<0^@xk9SENeL^J(^ocG>piinKi9XpSN%Sd|q|m3jB!xb$
zk~I2sm!#2WRFXlT>5>fktV*)zvt5!!pHoQ=eXdJ#=p~g%=%p@^(C1Z>N1yMKJo<u4
z3g`=6Qa~@OL`E-niHyFek|O$ImlV;LR8m4;>XH)rvP#P6%Ux1NUr|W~eWgn(=oOVH
z=#?%}&{tJbMPKcbD*BpAYUpcSQbS)?NgaK?OX}zwDrul^bV&n!QzcFG%`R!8Z>gk(
zzSSiy^j8q`738%24_X~!?(FC3SAj2*#S35l!Q)nBb`v~miL*%{dA1DHp1pKGaS;x7
zoXEkBlR4ONZ4P$Yh!%EQQ42e*tc9Jnt%aQ~qJy1I)WJ?C>tLsA>tLsk=wYW9^{~^+
zdf4gPde|8v2G|)y1MCd40d|JA0d~fS5q3t=2s@)}gq^W%gq<m3f}Ke;!OkR`U}tKZ
zU}uh)VP_W2urteM*qPgA*zpk_cD%^Lj+c4Z@ogS<mWTy*7SRGbi)?|NrEP(oHDZOG
zRkXs+DqCS^ZChbyi`Zai6K$}w$u`*8+BVoZA`aL&L<j5~vIBOGwgYy~h!b{B(Fr@J
z?1Y`O?S!2x;)0z^bivLgyI|*PyI|*zxMAlO-LP}ZZrHioZrFJu9@u$A59~a$2X>ye
z2X@|w7j|CJ3p=mug`KzUg`F?rgPl+G!Okc9VCQT5VCRqcVdoe9u=C4)*!kOj*a;B<
zc7iCtPLKuI32gy(fk*&$0Wko(fE<8bpdEl+FcO4aPz=H@C<kE|YzJW%iiBVn5<{>H
z$syQ<+9BA5BVpKu#W3u`au{~ub{KZiNECKaF$%k=9EDx99fe&i5`$e#jKMA@$6yz0
z$6zN$MA(U<2s=?0VJEgl*u^7p*u}*-?Ba49cJX!`c8N#=b_p>7yM&y8U80?UT{4n{
zT~bWKE-5Ermux3tmx`odml9L3OUWtNrP?Xjr6XzBrNuPt(sCMh>2?}+nMek988HL9
zjGTd8rk#OZHj;&1R?NaKD`#PsZD(PZi{xOJ6LYZ3$vN2N+Bw)s5easZD8WvWCD=)A
z33mBN9(H*#54*gahh4s%hg~63fL%c>z^)({U{`1tU?)dp*vX;{J6V=tC%0wT6(dF1
z6~!X#igFQl#dZ;PrAP^OC9wp%l3apasa=9yIZ}pQSuDe@ESF(dZkJ(KiBw=$5i78(
z$Q9UC+7;L-5e0UNsK8E<71$|l1$Na)6?Rpz3cIRYg<Z8>g<UODgI!Im!LBCPU{`C`
zU{{aSVOJOHu&c{;*wx#0*fk;z*fqok>>6?dc8zuecFjl=c1^JfyQbWPU9;VUT`SUp
zT}y1it|hl%*J`(5cd7Z(ecwD<y7@BmWk8n`^W|h8PoFNZEM9mD20lf%&${6gZ#xI1
zwQGTNb{&x3t_L#M4M0Y_5y)gW0h#S)Al}XcS?m@dtKAA@v)h0ib_bBt?gVn#T|jQT
z8^~k#0D0|RAfMd_<hT2Q1iJtfum^yG_8?Hm9s&y6!$47c6ewno0f}}IC~l7fCF}{H
zq&*3gvZsL3_B2q&o&n0*vp_j}4oI>~KzVx}s9-Mu$#xm2XfFbl>?NSGy$n>bSAZ0|
z0#vnEfok>|P~Bb!YS<e<O?wllWp4ppY`*xE=@+ciCmIi3yzm4g{N>+8W`Bs^N8Lp@
zWbH1ZMb_;iI%NGWqDMCDA_ipRE@DJB?II>*^DbgW=64Yu*|LjRkgdCj71_3n*!J`E
z3DfuJ;)Q2Q;lHy(XSdP)r}=E3&fbVTq_cM-kLm0lWR=Zc&S2kTv)3cfv)NC9xen#9
z$Mq?PJ#I)j>~UktVUJT8hdr*xIP7urjKiL`L$%n`_Nf+o+9B0qPdlbs>}ge|#h$ju
zwAj<mGcERX9je2gu1|H?(+#N(d%7{zVNa(r9rkoRro*0Yp6Rfs?@&GV^nI$wo_<L6
z*wc@x9(#I~>9MEpF+KM5^GuIDLx&o$XXsM{_6$R6z@A}D4cIfN%z!;Zj~TFMm}ds;
z89UU7J!795v1c4oBle7AYQ&yVWk&27d(4PE<2*BB&(xtN?3w!1ggw)cny_aYQxo<~
zDl=iv)MF;>ndX@Zd*%)`W6#{DX6%`V)Qmmzn3}O?R+$-l<{mR+&pgk}*yB5t#~$CO
zJofk@<*~<)DUUr~Wjyxy9^<jc&odr-mJYRG&(fzB>{*7?f<4QaTCitPnFV{69<yN2
zGS4j7vv#Nzd)7X+V$V9HR_s~F)QUZ;%B<M4_LvoW)_G>do~=V|*t7Mi4STjBwPDXT
zrZ()^RA$4Tt;cNGv&}Ob_8cARz@DQ|9oTaWsRMhCF?C?ip)v>d96jd1o@1Ulu;=Vh
zC-$6u>cpONNS)Yoj;RxSPL(;a=j<^j_MG#~i9J_`y0GW!Qy2DJL+ZkwYfN3(bE(XQ
zJy(ypu;-d*F6_BG)QvrNpSrQ<9#S{<++*s-o?B&Z?74f)jXn1~b7Rlbp&smc`qYCx
z&yaes=NVHE_B<-{V9(QI9_)GMnFo8`4)tQs+oxXad56@CJ@1%$vFBBp7kl0w^J33C
z&%D_4b*K+}zCQI~&o`t#?D@vjhdrOleAx5#m=AltdFI2OzeD}l^Y^JAd;THyW6wXP
ze(d>G=Et7D$Nbpy&oe*vgbo$3C-kX+Jz+=%><MEkU{6q)fIXqd1ndd(Ou$~CLj%|g
z^l1QlfgufGFEFM7>;+U7z+Rxo0@w@8vjFyj9U8=5uup^73l3=zd%-acVlSw&AohYi
z7Q|j~o&~WN>d+AOLVX&-UT8=|*b9wm2zw!wg|HXuu@Lq`^DKnDaEFGm7w*$A_QFFN
z#$I?#!`KU}ER4NykA<-po@Zg~MLRT#y=b3Cu@@cEDE6Xb8pU2zWl`)!dn}5*=p2jA
z!{^tr4vk?i)~7M-#fCJ7z1Wz>uoqKV412L2i(xM|$70wMJ5<D;*ry`)#32>2CyuFz
zJyB&M_QW0&u_w+k5qt3tjbks~r*Z7Xhcu48_?X7A7gt#vd+{ENV=q3(;@C@cXaakQ
zK22aRF{BCXCB`&?y@bjV*h}<S0(*%$mcU-JLzCD`_GuD($stW*FFB@3>?Ku}#9p$;
zlGsbmu_X3V9h$;ks!vncOAToXd#N!^VK1e!6!ua*mcm|Yj-{}d?$9*$(tVo7UV2E=
z*h`OT8hdG#rLmXpu{8G5b1aR$OowK$m+8|C_A*17!Cq!eGuX?hEQ7sFk7clznPVC3
zWji#By=<Rmv6mgvEcUWvn#EpLWm)WHdn}8+>>SHtFV~?t?B)72hrQg8=CGF=(;W74
zD$8Ln*JC;C<>pupds2r=*pvEH!k#pw685Aqm9QtNOv0YjV-ohHIVNE*-=TTz<@+>`
zz5I~ov6mmyJofS`%VRI!V|nc5=U5(lg$^xXuh6Fj>=lNzfW5+)7O+=PSpj>69xGt4
zFvkkmlRH$#p4_K0_T(Xzu_uqIj6GRpGWO&ild&hyF&TTs4lQD@*r!G86^FElz2cY_
zu~$@C5qrfRD`KxW$BNi1b!Z8Dr9LfTuQa43?3Ko}guRl=O4uv)SP6TjIab16xkJm?
zEB9#`d*vZ5W3N1>W$cwzR>oer$I93%&#^N0DjizEUZqbf*sBa_1$&h-tzfUBvI_Pp
zJyyY9WsX&_r*x=-J*7_->?uR4U{4uS1$&Ch6znNIreIH*V+!`F9a_a+wNI<qs}5-u
zd(|<mVy~*QD)y>9R>fX*j#aT&>(CnZYJFP6UTsKg*sG0c4SO|})v#CVu^RSjbF7BF
zdWY7rSMSq0_Uc1g$6kF*>)5NStd6~UkJYhPpJR3GH9E9`y+)rlu-6#U2KE|b+Q42z
zWew~#daQxH#vE&4ui2qZ>^1weiM{5KHnG<n(<b(sDr;h|*<(%YHRo6pd#w&_VXxJv
zE$p?1w1vIan6|LjQdtXotsZM(uQkV7`+Lsqf1dg1gVw9$fX&;<;oC`i`hETBB|s-E
zUbysU|M$BcBjVWqeetE6og?Bzc8!P&**zj|WY36rki8?~MfQz|57|E=eq>=p1mwVo
z1dxLx5=0J-NC-JRB4Om{h(wWNBN9UvM?^%9k4PLjF(L`%<cK7ZQzMc>PLD_$IWr;|
z<m`xKk#i%GLzYHFLe7s!9=R|g1!Q?dWaQ$A6p>3KQbI0|NEx{@A{At1L=@!eh*Xhl
zBT_@Ik4PQ4F(M7*=7=<rTO-oi&%v7>x$e2npAVr+H?Lf^^62TsVUz2$Rvmft^oiH-
z&AYi%|MrdVm?xKLX=YE@|6B^E;efOnEs##51JY~sKn9Hg$fz*_nKULKv&Ia>Yj_}w
z#sXy3Sb=OB8<0ce0CH-aKrW37$gOb$c{Cm%uf_}H)A)e=8b6Sr5r6`k08mg91PW<F
zKw(W7D5{A9#WXP>Q6mDyHF2PXCIOVxB!N<z6i`}|2FhqMKv_)|D5uE*Ng4?#ugL=y
zGzB18BLfvRMWB+V1XR|Pfhw8`kfKq5s+uZLO;ZD^YwAD^O#`T@X#%x0t$lR5bw_m-
zcWsX>-Mq%O#y!23C3?n+&;RNAlMkHSy!GI%%cghT(|=E1GQ9WN%;{bC^s>y!%M|ur
zlR3S4a?T9M?49^b?SOE|+5yob>jp%JtRE0PvSC0B$i@LNBAW)pglrxVGcrFQJhEj#
zEXdXYu_D_B#D?q`5C^hzK%B^~0dXO_2gHr+84wS$cR;+zz5($e`v=62EDVT%92k%Q
za&SO`$e{rVA%_Pfj2s=1C~|B-V#wlvh{*8)i6bWlB!Qe9kR)<yKvKx*0ZAig1|)-=
z9gr+?Za{L#(tt?F`2ooz7Y3w&EDwl`TpW-ha%n(H$mIbkBUc8bf~*XPf?OStDspW=
zYRL5gsUtTAq=DQVkS20#Kw8M#h<O`1d3x4e5b`{;F=07#@`$z0yDmO`eZ<nu>%Hsa
z)6c3d6F$!E(CcD*&)jueP1Eh=mpriT(bwO2aB}my<huNHyM2%Mm@htm?NO5txnk9d
zW2V;xOdb-~tvr7EkT`wRo4i0_?={u?Pw)T#^SbG~PwoZwUR%BY-DZ{93#T7r>1Iyh
zfV7GhNT=w4^okzHpcsIRiV?`9n1IZR8HiVSAd6xFvMN>}n_>fUC=MW};skOjE+Dt!
z2J$E#Ag|&D@+m$bzv2fH6agrp1b~7{5GbUCfWk@`D5^w(VoD51R79Y-5(i2s3818s
z1WG9>ptO<($|xD2tda%FDLEiXk%00_9;l!cfMi7mDk?>wl2QUHD`lXHQUOvF1*oc2
zfoe((sIJt38cGAGsWgFFN(*Rmd8119y<F)#?{+RqzH!UfCTG|NaYK6g7yj<AJapih
zl{<Wsn~y&9=vC8K`~1|rY0H7$JKmk#ylUer)AR>AJQh9mZRyxoem1%JDssS8<hbei
z`_d2p<FouVgEuB(t!=Gi`iT4bo2ujb<=^|=<mUCs^_!=60{^W)@PmVvt^T*k&0CMy
zdhGOrOn<_`<P6>W;|<^p<!5cv-|*xN<vAcNuLaWaIv_o-2Qu&mAR})CGVvxLGj9gs
zc^=5ZTY#**70AZhfE>I7$jLi_T)Yd&&AWj-ya&k3dx3ns56I8^fdpOv3h)7-ARhz@
z@gblv9|nr@QJ@$f0}^==D9*=$5_|$E$tQtQd<rPdr-3qj1}MvCfpUBfNa7`+Jf8<D
z@C6{5mw}3W5vatMfXaLssKQr(6kY+U@>QT3UjwT1b)W{{0BZ70pcdZ(YMa~pUQ7F~
z=I-mL{d;>q(_OuBwPkuPOn<Bgzj)!tdB8a#&3*~a35f&JN?IVDqyy4RdLV;j05VEO
zAd_SQGD~J4UgCi)k_E^rS%GYl4agxmfSi&O$R)Xe+>#r}BYA+lk{8G)`GEYAA4re{
zpnwzr3Q9qskQ4$6OJSg>6a|V&F(6SAf#OmeC?O?)l2Q^VC8dDUQW_{DWq`6$7APm>
zfFwx*%1e2mf>Z#KB^jtF6@f}p38*ZUfhtl3NRbqvs#FE4Nj0FlR0nEE4WOpf1ZqjG
zebhL&e9qn<S<`H6Zc2OK$n8D+>sNmI#>`B1-<xdnIY%rizWK;4xr4T3^Ev69FU>4`
zfPcNi!iV_Rn_t>Dnc1@c&CQi9<-!C*=bg?DIB?NHnuC`tJ>;-uhX$5=-N!Fq9B{2z
zefY{FKV$piNhhCj>S+$`5gYBFwVZJ5|8(oBqmNm$ZvAl^KKJ<(zc8!Q8%(@)t@$X<
R=s9AQKX@+nMVtZ&|34)h9^e1~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp.hpp.86A83F148A978F3B.idx b/.cache/clangd/index/device_gemm_dpp.hpp.86A83F148A978F3B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7438ad4a5fc8ab507f991b277118c285f28eed7e
GIT binary patch
literal 4414
zcmai1dt6P~7eDK`QMW_4hdRBVgr20Rn`)F&G$k`=)Tn5P3@&~ZBk47%$w-ALlvl&e
zYbZVRASDx(5Mm4!UAgJ~R==%m=ls|G+|Swjd%x?f*IsL{@^E%^w2&nT`*qI2oBenA
zatMMzkl4OM5b)asf@qc@h~U&)E9!r(pU-=8zAt_7-6cPR9PT10HxAc2qSB%@IOEFW
zQk|s<9OtvHal?Zi8BGWC?iv+9xR%ZPO3E_T^82>-S5?l!lM<#$cQ$0(sk}PwR#bL6
z$o9NRX_S+LN$E@718G0dyC3hje|2j8&&^z$og3%3r7SPmSCyuBa%N+3*Sw7X;)wlo
z-w5UQZ1w1go$mB~ROl1E!bs~iQpKEQ9jE;q&DTw<rMJ92dPSj7f7g!hEdLeWli4kD
zZcNBoaxl60P`O6P`$T<>5OLyp4a@u!tNJQ!T_&H*nR1nP+e7mwr_8@Z37Y;}?+nMC
z$o_kFyYBJpX8gZD{MRctyIZouCNb=hrBTwvkO_fVX31Wf({d_$Zq{zNT%9tvS*7rv
z!kKt}Yusw*z;m)mg3{34s|O|(ANxn5B5PpBs#~{XbOKtp+zG6FZdbWb<XG9-^S*uM
zL(z)8i5EPNg$`eumUrVi|K;xP?n@8V)WuU@eAqNs6xY(c>G8AZH-(NhSMn}iu4%Vh
zx_8ZK`TDxJU)Fpw6c@F<%r3Qzt#~<q@VUFgpjMHMgl4@BU*m?+(~@`UN7o9CH+nyM
zP`B`%(V1F@S8A8r(jHoBOb&h*{8)6@rl8YVv#=%la=vdvY`mYVPfMg&$EbQ+QS^&x
z!^Q1A(t@CC^KQQ`b2Qo2+cwpEZOreF)AHwObl*(juiHKNEcCktuT(lNm7V(dJgaVK
z<KhL6v$CD~Q=(r5xYrxkAAPU0#wC~&Z??p6KjA`KhyBo`!9Qs$VU}xrzgtGJ&A~K@
z|7(x-O^$b7t;>?zPfO0~1&Cy(sd`=2F9c<;zkbx_98r1XeaN8OXGkZl(~I|0Zke2O
zr{JA)#Omw&(yBjSX^ealZk5{^<T|qa%Bg|_6VENZ7$W|0Tj~m@KO0tV_!M1q@n4Oy
z$OUrTZFID9m3F0qOM;GD`hTzOrpIUx2xGWcMa0TxgQ!x6*{8I9k1EcL3|^~~{5{uo
zMom?E))C2yyRTD3L;Gr1Tcv0E%;&rB+g$76zDjOcc9zt({^+xZsLYU}J1btgI^0ye
z+mrAi=cnW8!4Vt%QyfBGmuzuQZVb;;a`o7jS)w*8ziC7Frw1dRTQ(O62Yys}k-lZs
zmNeO>)RYp3fVHN%)#;VD&C5RB=+AIJ_1f)RdZX9xQHM-&hmZYy^knr`mw>e+A#Iwb
zu}^Pq583SZ3<yG^qVTdtjHk0SIe|ye0<J)A^g&3XK1Cm@W>)k2bJ0!{$w`RbDZNu}
z^nvN+XQ(6I|7<)hO-@9#BiB)J^nvM3Q>HC?YvSF*Bc&02A9U`6`PT)D>=3UB^;yTg
z%p<4B5_Hg%AlcCergu`ZenSaDOCC8HX=niL1~B`&V3Gfw^Z$F!0pAxqQUZyHfF}a=
zuL~9l?y}xW5uzXDNG=k|0bUNMeO<7~-cPgjU+?+6XUvb|Fy%PN(qju2`7^`EO<rKp
zF_GlR5VZ9qYlYDVAw{A*5<o+$U*^=J<*V&^q$DzP4WzCiGh++Zk=p98V9|o!O1V*y
zOyFgLE<LtjkrP@q;bDjWD3>Fpkd_Qk$v{aPTd;^wKQGbkb6}kuNg<JP;FW{w*9D6R
z{D0Zr!9{-6R=lMf__8EL$AVrgnBjy|bOPulFql(We*DFzZZ1&t1(3M_+SpZGA(ObH
zm4UNB(Lr>O7NhM17@PnzoJI--H7gV5U?miN9p+sJTbv?_E(MuVP{Joe(RH9z2QzTR
zNVh#u&hS%OguD^MWHIpYDdU3Va7PP~@unQ~%fT3DhN7E5s|oaR{d~bQo4^#~$?H|-
z1zc1s1`ZuXQc>g-te>vCDmipMXAa4sPk_=1pm9VT6oKw_(8Cd^4%!&(60|5H4_m>r
z6;@*nnp5^!?w725gNQIt4TBjtN7{LU4Ijg-qmj)pPz(cQOz7BN&+V9B%^laKxu?03
znUlv5QXQfmqL1IV(rCZGX`*ktJcn)vsvW)s!=XEX>R{teQ0@d(>?()u0_83??gr&<
zHtqrC9yaa;`CgdHyx#}%eQewhO8ubB@B^SUz{Z21Itc0vKLn~nZ2TEaKf@e`9|p5w
zFlX=xn2xadeI&JyIXn&>PEz4)96_o^kSu$VBo)c<t}{LyJmfWRJ<6xRP%sx)x~tnf
zrF)BxE<-;r(OohfUvkT(7TwzH&pUuxNd!_Nkil+!sZo6~7-5WGkmQcuhD?nLGz7D-
z5sD7wg{t8rr07VP842bLjsck%P-So|%#8&*1}DJW1U43eybv^Tzpe01_7tYqTcHqQ
zNck925&PzyV;Sf*T~Bucg>E<W)x-5op_@wg8kpevrqD&La38dANBX>tSwHc$*1sHz
z=9}_unV`fVD+VpB|4S|wi(w^$TVZi4gMAup^~x^Gs3ON4l^v~ds<0K{ew`0kKMoiX
zUxov2f859D=dq!H<*{8vjEw?J;F|^!V*>yaI01+lyYI1$aPpbWpw;xLs4d0dgu=qQ
zIjG~EP*@npxlD)@|2Z<BqUbzO&jUSdi=y*EJ)eyWK)ryCZ-M$PHZBD9LN+b}^&&QY
z02U8m0nQdhKLm@1Y+MBvRcu@h7S(KA0~R%G{0J-_F*v><)y-(Hr$4H2{&#$PoS`pg
z>M3S!IQ%=x*q15jtoOgTggpDu<Tta5?F!F)Ku6mG=&OhGlicg$9WJBx1eH#ZagY^0
zQiPTUtzu9r(5dWtOa+TnScs43aK+;Dwb7x5$hSIBr~^$V2~D8T1e&-cZ~-Il#%3>e
z7xxAt#wIQ%@We&L*qp`08T)$mg|`Qek83m2aM^DkPaW<=p<e)l3$Os&!t)5J9vjnG
zwJ<ne!hLgwBYGoHGLVP+Aj{0k{cf7i5oB?v+D?63iWI7Hg$5?Z8!fwTD8+VM7<Xc1
zTU<N3yXNC!G;>-&z6HMB0|ob<cxLfCo5s~p(kIzHW6tdC(MdTyiqyxadMDHU*U65O
z=tdpqK*~kO!qS>8YGgPxN>C?z9CZQJ#m3!0b+d5~P(5ti3sf%~_W{+%#{EF`v+)2>
z18h79)F2z9Njk*FpMm<!#=}4jv+)Q}BWxT-Qen(E;m{Ez6~W*t>*>FFxTzf;H^;To
zyxfXCrQxU;grF@1V_ZH@j}$l0lH<oC!dk<61}>iGN$Cw&o~MVRGZBM{7##7LyqRBR
z-z%hHT#JqLS;K?xWQ<Wu#)CmTm@sV^4|eh3z~Ce>NMiFzV3)-3O{NlElVp_QP!-$<
zsrw+$v}FTGHGn*B%lC%4S!U`NGmzjPp!x^QW-KIvY9fO>Qge$;u4wCx8^hq>xgwL7
zhfxjDOC*<Q;}#P52l|?!KYSg8(r*<$?Oh4@2Eo1c?VlQ)T>|Ho*e+PRJp#Q;D13R0
zeU$KqAQYLcJhN5)wk7F$gNPNIs{-S>YN`rTl@*m#<m3spZ#<#GY~`4(Dzlx+Z2u2b
CW`pto

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp.AC3765D99685AA83.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp.AC3765D99685AA83.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac690f6bfbc7d0dce10b6867f9b308a4ff1ef47d
GIT binary patch
literal 6654
zcmZvg3v?9a!G?FnC6J4dWFavdXjt93-!BBREbs$vgbhRrfdv8)bRi)GkYtNcwcJ#S
zBt%eAXaxyYB2WmIg+h>e!KFzN?60hKQMvtzbW<p(6(i-GF8R-px94=v$(MP*_sMgz
zU%uIy%*3&Ug+JujY@>r?moAtSnr*k)Y#E8Kno#w;miuhBFH>x`rN4c5N>R=E{Pgy1
zf8!3Wn$uj_kTalan%}HjP<N%azJIT&8y{F{OI@+?++!Dd<ZT+iv**wSbKNP;c|Y8F
zGjoOgy}6qY&mf;_c)jh5D?1%q-#hoOv!TT`i#va@p-evYGI8Zf>MxJ}Y{t~RXN!B*
zwsfwzIq{=-`*3~X-mKRyjQI4@?IXTpCz@xzo!xP+YFw>(=fL!x9s9Cx{rP-z$M=`F
z3>iV5IQ@;(^!W4AhJIU%uAf^!c}cz8_Cm+%z}Do_DberhBB6C>-kkfJ>GN7%$vC#@
z$KIW{iYmj4zZ|gMx$tb}(u>nSeR<()733SAjCplsd+uk$H=dc6z5J=Gy|+)G>rX!M
zWXbV0r;lxX#5JeHR`1zyaDS7_zu6Y*Qx}LHs=9gU!=3f%1Li-M)W2~;wd0km1Lyzg
z|BTs1ZvTy^`#nr#8f6{DbthZ5UtBqB?Wmel_qBBPJ~W6T`0}FMt)E?bsPlo6weCx+
zcl(;o+)DR7@|RaC$E@x%r(FM`dTCp5VdcfhmeKR(_xb(T4>i0xl0CZPm;E;{-*K?%
zO7Y@%?Auq3x&EzZ(~g?PlCLLSc1>!ePR<ba_Dk;lWQz9JCmWlt<i1!tF>}DG@~YC0
z%oVM5zkBZR@fVAhd7ti+6F62r<ol~j-Wof$+P@)pZ_iQJ(mvw0zxC})^Y=GgS-kG2
z4?Md@j636+`s*FBCtI%XK7OaMG5m$Id2rJ=XHH(7e&p%X-=)o-RG#B$9Cmhp&&o&F
zN0zh~9?tp4!OqKt?;ZHx&orr~{+ZO~^qmy_&N`vJmaSiYuEHLEX>#2r*<LyQ^NV%g
zg?bgYtq=H0ex951cIU8YQU0OP@{1K~OHOjnR1EX4aE>V5*Zy(M$g@*-EiB)-X~_Jh
zzf7r0j=#I+{_N6tadq+TiO+vAmdszt_YFQ@GO?}dTCZ1IKc0GR<_z-XnnL68>ipq(
zZ`4-&ZsMJ}`xh>E-+2A_qHX_d{%Gj?mrs|z7E3u_Fz(fxD<_fD%Fc<W4m@*k!;yE&
zGoSL$a-Mx>&n`3f(XA)W>tAg-<bAR3;mK6(v2_K4|Ho%T`JUOibEa<RT^l<GUAVmQ
zh4=qB;kJAH=4YGV-7vm=3UjmW$XoCKIBr#7=0En$`{{aWdB^cF*Z)4ebJHgkZ!bD<
z(Xs!H;x)^*6qiz;&Rf37Gydzw!8^R;a-RNX=(@GivxPm{ZhzD7{txHxy70oM=E%6v
z_v0@I-B6l#RnTjGe<StulGodNmaf^hfAEjX)35!oXZg0H<sWxC?<mJRFBG1*wR`iw
zR<(`#WPWR9JrUV?D7C60t9?@buZF(+R-cE)ee!|(R_65dmJzeBwEfXmvuwfR*@VrO
zl=p}Aqqh$pm`aueY-NrlM{?JYIOn@wDL=jW_3ENQlTyh^i8?1KC!=ey>WqDBoxj){
z|1g!Dkf_s>(v!OetLA_GInnUXzn)36O-a;4dJGxdHCXkKquB*Tvr^Bbl2a3PUQ(Xf
zHCXkC+c&Q@wlCY4N={DH*-6<wy9TRn*!_0nS5H;mzTY-6QRimmy1E9dj?W%=Ofs(-
zspO-Hx^Gh7<gUT0@1##T@YSM|?WyF?6Lm&XMsn9+)wb$6b8E8ucCQbW(zPyiAKZR(
z&!wRQ)5_dl_Xz6^B$CfPraMty3OdS1K}UOO=x84e9ph!7V|)yBte1t3^|8=#UJg3W
z$3e$?dFXf_51rr@pc8xobfQ;;PV|Y;NnQy$$tOW4du8ZkpA4PiRiIOR3UsPhg--RU
z&}m)`I?bm+r+anibe|5L;WeN$d<Jx;*M!dWnb3J2@Z{dxyJemMo`F`f&WFJ%UFQL8
z2furA_XcnW+>b*Fx)*gP%1;4F1t=h?AO$4tr-7scG>~+V29ohJKr#UaNG8Yt$@*C!
z*#HY98)Sjx{2Y*6fCG{XazOHa9!NgG1IY(@AO*hwq!18*6oLYfqF)4342VFAK@mvF
zF99h9Bp{`r1f=Yjfs_L>kaAE4Qt>N5DggyZC8z+Y`c)v+fC{7<RDsm|8jxB*15yiW
zK<a)SNIjqfsRwl+4Zi`T5io!>f(DSL-vrVOm_V9A6G)Hd@mkwGt;{prGZs>jXNr}q
z&;G=zaQA~BP6dZMt9!e<o$djUTyA%Fq8t>El#>FIa#29i4jM?>Ndrl{XdoE}10>^Q
zfMi?@kgS6Rl6A5`vMv@#&cOl6IXNIX7Y8Ko;DO|wJdnJL2U2hdKnhL)NWmolDLO<T
zMW+a)=n{dH91@U{Qvy<QNkGaD8A#bF11Y;?AQguKq~cV7R9p&>szU`*b*ezBE)_`4
zp#iBmH6S&Y2BhxLfz+Kkkh)6;(r_3+8cqX9!({+zI!qu<rwOF#GJ*6w=%K9bo;Vd4
z4-ZN7C|0t*6DeKyH`b|OZFDlF>%PXCPAeNG51V2AjCGe&p3-%9(+#;>2zM)CITE)R
zF~WTwbDVH@U=EY+!zAz#(tQeZlyrX%nQ9>@^iwf{LO&HJDD+cdl0rWfAu04zQIbMG
z-9pglr(*<-emYLj=%>RZjea^p(&(q7B#nNig<#On#0UocOq^iQ&xA<^{Y->p(9c9k
z2K{Uc!J?my5iI)IIKiTy4U;VT*$By^pN*0%`neW@Lq8WIIP`OIf<r$SCOPzT5t2hc
z7bQ9L^Q{E`XV@QnjNsAF#|a+&e3<0X&qqie{d|<<(J!<R0{;F&j1bT-#0dfYLYNfL
zFGNTI{X&!!&@Z+SBKpM`A);T56C(P>Fe##6jF2Mw#V9GFUuq>JTpuY$Na&a1goJ)6
zOiJjNBBX?VDN0J{ms<%L{c?<u(J#jd8U1pYl+iCoNE!Wdl$6o0v=R#Xl^CI*Ux^b6
z`js%LpkIlQ3i_2Osi0qNB~<jQF+xSZ8Yfirt6@?_zZxM`^s7-)MZeZcXz15sgob`C
zPH5=Y4wKpu`25l$q=tSiN^0oWTL~TgdW_J~ug3`;{rX{2N538+b@b~|Qb)hhN*L%j
zVuXQyBTg9TH^QWW>tjSn1N}ynG|+Fh5+?f17-6E{j1wmM%`j=A-;9tZ`pqb5cKb)J
z`Ma(2`%sLq|Hsv6{NRm?vZUIVh_tetAvvS1e|JiXGnuO!ygd@r1<3_{tZUbjqK8il
zQ}ZVzrgKN-mRSFhuYC3o&%J-~($TxqQ*vio(`#<{`WKw7btb0M)6=u8`^Wl*S&urZ
z^S{kVT;lBRjP45<Wv3v~b{Z06XCSe577}OYAn|q{l3*7giFOf^WS1bxb{UdlS0Jf&
z6_RGxAnA4;l3_O>nRfGDvQ{NCrR%W;T=zrn<**1L3KAWnAu%Ba5*uP6aUl*8AL1bi
zApw#Y5+O+;36dO=At@mRk{VJWX(0`g9?~HhAp?>bGVi6`9jlj3yY$=ji7k=m%^PXG
z?LAAIdrVl+P?(s`?wy@w-6EYZCkA`?8afiw4$2`}kFGu_8UDA&@;*P6m>y;y_Mo*n
zqp!VpmCqbnmzehS@?=`mZEBD2d-lm*mzXZ-SCD5-TaSwzDP50=;Ll;W^R3@2t&B1#
zNVGviVhjcnYp{?wgM-8yJS4#oAc=+uNirlzvLQoK3<Z*EsE{;6gQOceB*QQunT83;
z<M3ozOVZJ6Pd({ry8Btsm+GrpD`Y)6OKjY`kIi7Ec((*A#Z!=Io`%Hm3?!CkA#pqh
ziRXDp0xv)kc@dJtOORw<hNSQcB$ZboX}ktW=XFR1Z$L76^Il3kAFP^v_hDg8^%FId
z?jB6PJN*A^hIL@OuWZX6Tg$e}g|m_#n>J&5=}g<Sd!4Pc`*3FW;bYy0Pb>&lRU~53
z-0}C>dnEPjm7JP(e|n$ZV>5iC>5P=IqcXD{eILjfQZUpzZ1{-(k^2drhw|CH|2CVI
n+ke17*Mn|~;doI|HDl1=pP5ccf4E=PBmP-rVe!Z@#JGP0%>tPV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp.BE2169842E629D5F.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp.BE2169842E629D5F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1e2c6fcd3645574abcd9f332d4f4df7cf996464f
GIT binary patch
literal 5340
zcmZvfc~leE9>+6dKnWmA5K1UQaMnAU7)C+kia`mYC}>0!7ve%6NDpX)YJJbGsai!q
zYqe;ViVMXR!KIrk7L=mrX|!>BE~rGTt=+Uf+ZS`r$nTGrbGT=|_xs^<GLxH`+;PcC
zNtZ)ZszhsYasJGC)72`KD#-m;G_NrCVi%Rl&sU`?uKapJX2JFNK*wk8>9Mmat#x~`
z!Nk6s-q(qw%`v5T>RRGd$ey<8=UzTuz+@(@|1oikNtN~B!ntDs30=NZ<&B+lo_v?C
z$;~t_8gw^&(WX}$TKBg~-?aX+`t0+(6VHELn$33&3q4<P54+bf??myQyU+d<WA}Y~
zL0Y`H`gCj5FN;bR)_vEyapS*ZW1s!^dRw2ViwA$3b>l{M+e6El7d?K;@B8>)DHDFs
zWwsr=pPfH3=g6bNxkqo6v~4;tKWE_7+zVG9`MuMy%AB(Eql;tjV$bX;@6;FGc(OI_
zMy2)L*7}|F&3%{D{kNCZs9G$$-;dc+C%&3{@yW32QJX%r|7UmcA1mg|V}>>FR`r}e
zbau{k)A~Nxqn^6`we8pL-aFU&S#*L!;~#%<+kJ{TXs^{*v@RFQn3DzL1AC0*i!U#f
zgCp!qyDlX)Es+7l!#}pIS+AZ`xi`N!!ya%yCrvJ29w4row&kgQ)!M2(SJrt)xxQ|?
zeE!acit3do*2rHvlLEHfqK9+K(uzj>Ry^acW#6CrV8Gex;N_Z_V_Xh8cWe8kk})go
z!?u6Cu5j<s2}Jw}_RO@?`%VtLon~EMem~HdbDgTjGIDwH@I85Q^RBu-(%!9J_(5Xn
z+{Df6iCrPb?z|U0d^R)MN8VtcoYEt*biiecW9+Lt9U)6s#MKBJJHBbYzJHaj#d&S-
z5@7*lx_azp&epJQUrY>n6#ZWIjDX=|*6y@DNV0`Lt$T4d>BR9%mt!M~Rxk4Zblpyp
z%#E@%hmY;~_u4t?`gP;?e*Av7%ZK-!&3f1#ScxJFYIelvG&^6tU$ZeU*FHI;DR|Gw
z^y1qk>oIxp!LKUoTN4JR@A-W|#o-;9%QiI3NKJb>u=eQ0^abTbldhmqrF{;B$8C?{
zc9f&ppZCf@{bpYJ^04~*(cvv0vOheZH!rSggZZ~lcOHMW>13kq+Mve^Rxemuz3GT}
z#NSr*33so{F7=^9D{Idj|0}=Ox4X6`-(Z%M^nZ9L_t}g^zdxUH^@a1kyfVQ)bH~h|
zcTXcWQ#qGsU&tJLI_mR^OF>!eX>_GqS<AlXy0Ey=zg8?6k=y)p7GIjzx2}Im!AURw
z^H+W=pVjBdb+YB&l(szniw?)!x~OpM#gdN~QI3kH-u8;*U32@iKTV~6I2o#GzO?*c
zm)m)xYTFGpyF&{saejxNl!*r##uwh+HfLAeirXh^s+uAjng)DwGiUSDJimg2{ho)F
zb^KoY_tp!u{a@-{4gRXce)i(RYd_joZ5Vd(A1&{cVvXkx$L7rm>&VF-^!C(?_8@A`
z$@0&`_m*``ZSD1<s;DG?Mi{12c}E>xo4CDSv@f1vQDtkrwLYG~{mu8hG=FZpwJb$W
z_r=rQb+~tUkY`Zpp#3X#^Y^(<`{HTtI?y}N$1^B(pBQ=Ih!M2puNvpBBSIshJcCjX
ze);p0-49Ck`{Lu>b(nWpSI?l-m4`T&OWrK_;uGC<ly}r1&!E)pIp>Gm=F`9PQ%!K!
z1~0=<&!E(<>Cwl9xF-f*e2lve@ec9v3`+efaKfRR1!o@k;$z))kav)eXHaTY;mlb@
zPrG+sp9qcTx~w_%a$D`40nvWhgpnAmyhC>~5hFVjX(T~Mnn=)5MhbM4i2@yMq(MiU
zXwWf626T*x0Uc{(LC2a{&><rN9Wo)%aYhbwoQVS+Z{$J8n|RO(Mgeq!NdTQ_6hSAN
zM9@h_33QT40-gMp9P#EO&6XqOXeB9YkkNS79{32%#Q4sOPFRQ;AX$lm&P19?K#~>`
zkffCaBxR-mNm(dBQdSC(w3!AZZJ_~4TWLTtW(JUqg#jdEWdO;VSwOND7Lcr!1tc;f
zKq3nQB(fqva%K*YoP`4<XXOCNn|VO;79Nnil?S9?762(&1V9Q_0g$3u1f*yY0V!HV
zKuTr_kdj3Lq-2!<DaXo2WpVgr%Y)=(kW%CcN>c7l?uGOE)&v%gmI&=!9E6UD1W8X2
zor%<vfFyM!AW1z5NJ>iqlG0Itr1TUZX)O&%T1Nwt*3*Dwv<x5_9Ro;4&j6CuvVdfD
zEFf7u3rM6zfJ8b3NTf%A<g^?hIUNT`PR{|7*Ybembvz(>Jr77hD*#f^34j#z0w6`L
z2uM*U0#ekAfRwZnASImyNJ%dNQjU>HWpTI{4lN^)cv(`?o9C9s^SlBJ$42}UEDjq9
z5@n-6qHQ!tjEw<_wXq-}8v=>5aUk(F9wfmgfF#;PkR+S*CMnO=WR2%}3f_JbMl@rc
zCqOtc;yTPOjCc%lJx(;>fN#W!b1*w_;ws2w6GlQm>BLCrCtVl`{bW5(LO<Dvlh98(
za1#2dCX9l9%860XPq{D(`l))Hf_|zIr=Xv5;1u-JO&AURv=gJDpLSt1^wae?4gGW@
zPD4NKz-j1bnla`)xW0@NW1ydLVGQ&$^*96lOe4-fKjXj|=x3WS7QTPhiLubnx-b^{
z*?OFXezp;3p`Uf&EcBx$3_(9~VhH+?3q#P4>Tv}9s1Zldj~qCHey$1Qpr3PM9Q1Q8
zjDvoz9_OH+Ys5L|=Nvc({d_aVp9j~UcVayB^Dc~se!d>(p`UNWdFbaII1l|oGbTX4
z;KT&z7hISC{XzpS90A`Sp%E9LUvS_8^oz}y2>qfH6QN&pVIuU44Y&yXVk0g>zv#e4
z=$D!?3Hl``CPBaC!X)UI8gL2vrAAzWe#wDLo&KTr4%Lfa=Q%NTc0#X&EahgK+ONDI
zRQPBe=9e8FAD*cE+RGRnOg}0!dby|LeBuHW+Lki>0ynX5ntM88NW?hhPHq3m^man)
zM>_X(U|?XV^5ieAnl?sT*!P*A`xm;J2<p6(k!lhoN=<=8t7(uJH3Jf>W<f$~1QMs_
zK;qRrNP=1bNmPp<Nowg$Qr;o1@$7XvZ{6^Ueerl%N#>o(weFLRN{UKR{-?J1ORu#2
zswDSxSog3{<tgw@+&HV3sp_$NT1#pLWy5zWV^I6lsH^AP)A8!~80GppHlBEdf(NW|
zPs`opU}d_kx7V*-1Hx9jr{j9WMJdw{TK0Thy(o2zdpaT`Vv6!<C^e?>Y{`_fP_|^<
zHJ+`QGVT6s3`Ae$Ec~)b0|^pkpg^JxG)Rnr0f{xRARz++i8F8@@dh3w!61Mn8bpvJ
zg9MVSl|z+td)%$I6z_lF^|B2iLwYOcrtB;IHJ<$>I7dVX;2aSN5``#`Xheg=AO<8B
zu^=IYK;jSw5|4P01SEhYA`v7BNpF&S-<vo6^(MHeFsCT}^}(dqgXh&Hd(F%qxmQ&o
zsoqmFCr`?nqMEE!DpgkJVP@yyl+MGcorgL3*1S3Hm@+H1i`vV(Yd0TXKmWjh?#V%>
zL@G#=JS;d&8}e3oe8K?Zz(IqDh&{Oez4|gy|8F)l;_b+2eGEa;EaIi!u?D$c&p4gX
TXLyg$5$0)l(&(Whu~Gj6O@T=B

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp.C525CCAE82B88202.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp.C525CCAE82B88202.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68c67561f60fa0fcb89cca439de31ef4c0936274
GIT binary patch
literal 6654
zcmZvgd301&g2wAzi$D^DBn647K!?(L`@RsMC}039LIolXfdT?Sv>;)Zq#B{sVN)3-
zA%coxw;*6k1RBB;XbI9TSai~CJxVziwrrC~R~j0W5hKl<E;8qt?~lnjT;6xT-}#;7
z=Dv4RZ~W-O!XI)Sj*;QfOBc+F%yc>&j?Co0sz~LWrUx93uhSfkrN8}fQc=}~f{fO!
zf8!6Xn$<XWdv4$PQ-W5_f|{$<wSBrx-uTc;NBWA5=Vx5(lD}!}&aQ(O%=V@==KpZ-
zovankk7jQ^JdHfF{mquIuI_Yg`RM#V&qfwkEpGqShBD>ktJKx2>A(Kq7t<#1JzLzh
zx~YA|t?{2FT8A17_h!FwaoFdV?;Z&pJ>EF|y_~l5^T$+M_YO?m*|sm|&Yv$dw%xw+
z*1%!p_^EH@`X^tMH}u|8bmRQ`iA!pgmY3RAhqk1aPKteB6OF7p^UmzwOr6v8TISJB
zKlW(9Q#7|>@z;IVyBD6#T6$^f=dUh&qk_DBZq(~5Tl2mcy7A1EoaN74>#=PdQ+wj6
zr%R5lIdyd7W1d+hj#}T2gZt|}!Of0H&zex|(EMAMKiyfI(RbbpDShh3Rk~ig)^FaQ
z{@a{c<PF|@w)dk{mRZ(TTyvs%+ohFHtQ}Ex@`0xI9*6qVlu%xjx8;k=kF-BjvetWf
z_3l9ZnL8PQ$Nuu#+)=B0&MG&4s9f3-UO4ws^sSL|=JovjH;-(8eK_~Qj$ilLynM&O
z`m4o@-*;|XHR{H9zD+x->Po(uaK$sBjy^F>+}k^~$J1&0e?47Se>Lyr>hW29SC!8%
z{mfd?T=TmZ4j+5DXqo@np1Gl;<pXbDTk`Jc(Urjsd3(E#xZeFUe%rg>y)tkA_N$B6
z{q%`%*RU~XJd^)pNBrrg8@rF)tE+4H%H24i{@XJru1!7i?5Xd&&zw-6>#G}bc7NBo
zkFSp|X)QdQ`;UX|R|-Em@Sk7ka#ihf>5UmXY3BWPVrw;5yZn5Gv*DGAHJcRY+^Jt)
zs`)<Bt+-`<C{XgtytMb)hs25s4vkb^u2@@gf`6`JNN|OFSn0mj(^bRIPTsY!eB-8p
z^XmUHX?|+r!!-})lqQNRi+7KI@vG6KV5QJ2{9?)YmigDaz21C!^7ZM{$g5R_=985L
zL-XIRuK3;fd$ad1T<*R3=CMUv|EKY@!5?2aRr*Fe?ZTijuisiZflMhoFP%K_+`$b;
z-Y?I3CisN=?E8CmS$U6dIex+Df8mh-<(fw)($z=T4HAPtJ|8LY&CHuMd7I$b*xvu*
zm5ncb{Ks*3y<;~&-}vE%v8|KXTQx`C{rJZ*t3uQNv2V^#H`2@7j*YtU_o3~Z&Q-j(
z=)fh{{<n+QEPJcCl>U6q@<qO}-_#A*;UAOx?6-s0t(BiI?9y`g+ujd;I&as-mp-?K
zC&XT-zwUoit>0C_toi-T^ixaTZ0%aQX6ya|KQ7O>{==T-TR$j2-R{1p9&5i?c>K=p
z&Hr50GUD94=DD?0bmyV;`4!o%6AJ!)@Q3gAd}PeIPrP@sre-t^oB3VKA01W87Cf0l
zIUFhZe^@_q+kk%Qq$K1hbEUXaJBH*v-|?09(>vd+F6uuaolHp9xhc7s9fMtG?py2r
z)!xLX>1158&Pd5f?HKGjZ{V2&U!6Ral1?Tk>->~_t7EY1VRvs`uWMblFP%(G);TFT
zT{{N5e(d(A)?Z&EO%FQ8C+ocIJWt19*XP<!AKsPM+TPtUDOnHfGH^i0VAqM6{f^4k
zbu*njo~(PN^h)g*?D}5Dqyv9lbfPt#{4!Z*revme40i3PoHe^Dr&s6txY9b-rRI~n
z@9eoexL@}&Z@@ds{`tC>(E%C~6JQ{*0TvP$;2`k<9+D6cAc+AHk`$02$pIOX5>Ozi
z0Tq%K&>-mn9g-0+AejLZk`=HZ`TF|$*$Zy(i2-RHdjXat=$+WPB;JtsNl0PuqRvDI
zX&~tk4I~|=fn<UVkW7dHk_j_FvOyL|HpBwShFKuFAO|ED;(+AB9FTmF2a*r*K=NT8
zNFgWyDTD+dg|Gmm7!-jNLn4r3SOiiEN<c~>2}mg{0VxM%Amxw@q#TxkRDueSN=N}x
z2`fOVK@~_fqynjiRUoyX2Ba3!fYiboka|!DQV;1s>R}y7BWM6=gbW~!umPkQG=Vfj
zCXi;>1kwsxKw2RSNGogs>C=6Fd%Jfp^9}WlhE(L6WGDOim%J73Km6cUaCx&kx4YNv
z?F-4{^>!xOMFUB@X&`A24J6}YfMnbZkc@``l6A2_vThbg*24nHxi}y>HwPr=;eg~_
zJdnJb2a@;jKngAaNWm=tDR=}RMVAPq=oWz#JtB~jO9E1IOF&8<2}s!`11Y;@AZ3pX
zq~cP5RNM-Xibnxbb*VtAZWTz?qXMb9G$1v%2BhZEfYe<&kh)t3QupXU8ZHA!!)*X*
zcnly-mkFflHi0xfCXklP0@89@Kw2IPNZ-Rg+TQNTTY>cnkR+dKCwoVew2m{r^Hvz5
z44Gy__K9Df)^XBzLf$6I+f3Px<mnoxycaMhC~q6)2I4(TfR7UINz5_g{Sq?WMA7J{
z;}ng4IziFsryB^3emY8M^wTjyqn~M_81yr7ia|e<pcwQs4TM2I6D17#nHXWv&o)sk
z`q?<eqMuDrEc)37!lIvz5*GbzjIikEnkWwaT%6+2&m|}h{agd#(9cB)hkh<bIP~*P
z6pwyBPVwmH6BLhrzJc)Q=c9y2KOZAJ`h{jn_%rMeAx;VC7ZQ|!exZQ~=og|yK)(<p
z0{X=!O2qG9j8h`|#RMgyUu+;E`o$;_(J#h`h<>SwlF%>3DGB{jf|AfLH4q8?Qj|#O
zmtsUhzuZj8xIS{6lF=_GC>i~71Ch}$M~RGnIYwmkE6tRGekD#R=vNYyf_|leDCk$B
zL_xn2BMSP}W=ciB8mCnBs|iX)zuG`l^s7;#qF;>>75!Q>rJ-MoQyTiU1f`*0Yakl>
zwJ6chuf>Rle!ZE}(XYoT9sPQO($TLUCi)S0e(6!7qhF5^9sNc#WuV`PQwI8t1ZAM#
zI7|%m8&P7Q--r<d{bn;|qTh^DCi=|;Wuo6~ASSMl86_tA%@{G!Z#7dE`mH!+q2EeS
z7W%CQVxixP5)1uSj98ui;cNcxXulnaQ_eCxjqDGYS!o>~Ea5bYc$ascEp&v2#6%cK
zY=nixML0-&goh+V1V~~;gd{~INODAmq(l@*YD9&kMKnlyM2BQV3`l0ggk(jmpVQa@
z8y98E)vr+9%W?<ij<i4Al@w>O*EaaOB&P?Z4(e$i;7f`gJta;q7?+&R8<AIHui(n(
z|M0@cmo9&Be|l2hbbETu%|M?)XRF=G>5PnwZ2Pa}+U-v~?y4;KE;D(~JH45((oPx@
z<76PQP8Jg9<RI}*9+Kb`Ac;;9lH`;i$xa!P;#451P8E{o)FA0j9g^WRAel}RlI661
zPP^MyFPn1tx9gLuk?+qRZvWVOmNs@7w_tl=ayq9+PPY9g(s65ixJzJrTXNb(yJY)4
z%_k*8|28B4%ah6JA<iKW+rJn4e=4oxC5XM1V{d$PO~@Ktlcas!d|5WtqIJ35wP(({
z<n*B4gYxZZ`;`fu*72&umLz{mrnkWUc9Y5P9oCs>lZKA=)6g*{10CaMpkqxII@ZrZ
z$C(^-oS%b^H+krIKM$Q?3eX9D0XoqXp%eWgbdo7SC;27lWK)Jt_RG*IrUISfSD;f(
z6*|?gLZ_J;bedm-PB(SvbiWRrVH(gGegitwG@&#7CUll*L1+0b=zK0;w!Hz8yW}BX
zo}Jpd?WrZc`uk6YUUV<b-W~Rvv*c6r{%bSXC4yImeId}07=eMr3M?c};2`k=4@nRN
zNTMJ@k^~8oEXa@)L4l+SDkM$NAnAe*$q)=kreH#{1ncLNd_OpU=KYt2Rh3UwO}KwC
z_5SdGU(@Uh#{*?&-`kq?zq8iNNSQHZ+SJnNjwwHPj?&J{>7AD|IxnAE5T0L=j777@
zKH%(<(zRP^diMu2diEHd85qfAri~tvmE-F5Q0~A%gZ)E>4*NHyx9EGMfXn~?X0!A9
l^zG+)*h{m#AgP*e_8;&I%Pkv^_RfAR_yj2|9zKd1^Dh8>cntsm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp.E70AE989E444006E.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp.E70AE989E444006E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9ee5d5b508ecd2334d7e80ca9fd21d81eab4d2e7
GIT binary patch
literal 5340
zcmZvfc~leE9>+6dKxL672qhFDIP0BF3{lXiRScD&f`UdwK_x1*AU&WFDz(q8saiz=
zYqe_CiVMXR!KItHU_mK(9!49t=dKZJYd5XW_Ki7bn%^HU=Wv<td_U)NGBY<fx#N?Q
zk~%|FszhsYY2nPd)72`KDoFk-nOmIS(M_fD^Hr%zH+?-Zv*=2EpyRWL>9MmmS?l*=
z!-##?ysr>R8)GW)F{_D_A$#slKl|d*d?quY`p3l0CRNtG^Jk9*By{^uRWL5^9Qh7i
zo1bYcA9^die8bCi?fcuMZ`yxZdFENciD$nq&gQ#^g`TUrjot2=d!lsDt*3vAvHQL~
zFD+bHbE-Y+m-4a&_20FxU;ppe*r)%!a<_lZ!eQTLUA>xp_rB%y^WHxd4tVshl!-rR
zGw&X|lU+C|_sE0dIY+OT-Q93tUha^b{PUL{_`TJ*!kn`6qmFU6u&4Hvw;GDCKHd^{
zb(8g-_J*DGwS5=VgSM@$Rkc}mzZbK)UVJ&H<MD`^(HlOr|7UmUAIs($Qb)AzR`r=T
ze0J`1Q+0prK_Bh@x`&r<-9Fp?S#*NK$3MPf>m7<YXs^>%wJ#OcGEGGj0(+0*OFI`B
zf+OsUyDuhv+9Csp`+sa*Rjtn3w70M{!ya%aH_cGFG(cQ2ZSxb|iq+fqTwLQF<@&m%
z^W4pKRW-{`tTKG*ObXb1ogT?8Nh^8tx6&DZE&2ZB2ZPVl1TXc8ImYFpv)8vxE=yf*
zAF=J@HN|_6P9)+_u&1Y;+SfGXMw+#{@=l;J_X<^mW#seZ;oF8Ot-I>~NPDMd!3T*I
za}qaJ6T3o=-F!EC<ZNbaPs2L<l$73)6@xo1j&Uz<c7-fn7FR2*@A{_o%KjDFHs|HN
zi-h@<>C&-lxm&_|d@(8HLG-)XGXh4YuHI?8mt+fnQvdu`(uw02I%6YCR+jsJx@IRy
z=0{mt!^idcdv%_=Va<fSAHUb5^YFeiS@$0XZbFeowL4<8K09B&SG&F--##UyC3w%M
z^wJw;)tF)7!LK$ov?mNn-}C$6s>3@nm#k}?F(&QFkh-Ij(&tx}OumRlSM)y+9=9!q
z+fj*Tf8I9(4V-!5%fssL$A-6k$o}wb?%cTT>&(A>y7TzU4NZx*%R?W{Upaqq&4wf5
z5r13BC*19ByVQpcEw4L${I9~k-|pIye3e;LHt7DL{HHU@e}6Xh(y`3@hUE$NnLB3w
zyn7n4k;?6yeLi#Csi@DZE(B$<r_jY7Yuolc(}u-`{<Un;oB6FjXYmyU1L_B*6g7GI
zpS$>5<*fdXuaIrGr`|2#zvyzzsgDZBo-g{ioN`pP^s`qb@0!#9;gd1c4^5#ytrwOa
z>~^DIblpRJ?e5SbOPt@~$7{s{jT4G*Y|YzMzwAa+?e><)#+JdKT+7|~q`<G};J{~L
zYrB50`+Ljz+5RuIFNb|qW<S%h;PQ|573)THyxsOz1=f7_aBM+dSXXZL(ART1>_Jps
zQ|0I3d)Id5wD*6$y`-#gMi{12c}E>xow#jav@f1vQDtkqH9g&f{K<E}eE!^eeMyQT
z-4{=n>u~SzAorluLHn0$=k0Ty^2O8SI?y|?r+ZN9usG~c5+`2v#V5&ily}rn_n_3n
zUi|!c_r0?HzW4;W4)YG{?jDr-`pV|{C46A2pK799>%H{D-Gfr!dp{%e?ds(F{;KhE
z9T6H4<sOvUH9h*c5cgQ`i>JzUh<8X&_n_1-11BE3R&@HFFFsDLgS>-!x(B6J70;Yi
z@}!sN{`B&3-<MS<Uu><rIXK!cn=lchl<V`$CQT$rl!*d~Hqjt4CI%$d#DauO2qezL
zfyA45kOY$el4ue^l1vgvhDbxSawnCi$@sXRBDfecF~M^&goT&^l9ed(B+^U*lC+S3
zB&{SMDKiB~%0dB>vQmJg%`_ls3k^uxN&}KHGk|0)3?LaR14!1)0+O|`fMl&KAdwjX
z5?K%+kre@wGjo9CEF2&?D+fs4%mb3Q@POp4JRk+L07$_i08+3DfE3LlAVrG^NYN?+
zQZh?`lq?b;C94ETL#)B5Ob)+n!%#yqNGXPiN>V=O^2B-dAqNvjLxg%J2cab*LDCU~
zCy^QwkffFbB&j0-NogoRQd$a-l#T)<t)T%)YiU5zIvS9Sh5;m_WdO<O7(lWb7Lcr#
z1thCu0f{sSkVuOFiF62%oQ4A=r{w_2={P|08Xl0mmIoxS;{hpX1V9Q}0g!@D0Hml9
z0V!%lK#DpMkdj6Mq@<MqDd{9Y8e$BjGCAalLmLoCyg^cu^89fh_sZ{?I0=0dvJ^=9
z9!U0aum2v1Xu*h9OmWC-t`j4!!0f_^M=&?wL?aIPW}G++vjZnCflRhwB=nO`jD&vD
zg^|!tHsB=mlg&5@{iFjYp`U8ODCno07zO>53!|W)YQQPzr<!pJ`Y8uaK|kGs(a=vj
zF&g@57e+%r-GI~3PdDQ<^wSQUhJL0MW6pv5%Q!Ix`WY9-KtI!fGtkd8;|%mO4xE90
zwgqG1`)8dP3;nDMW1*jIz**>Ln{gKUSqIKSKWf1c^dl#RpdYy~1pTN1N6?R&aRmLy
zfg|YWS}+d!IVZ+JKj*?Y=;s=64*I!foP&POfpgH$w_^M`aQ}HH#zQ~v!g%QC8*m=_
z`DUDle%^ue&@Z%N0`v<`On`pDg$d9vG~&V$@c9v%aRK@T2QEOr*oukJFFG+1`b8He
zLciFEi_kAN<0ABn4qSwOsTGr;Uvgp+^h+*Gf_|wHm!My2#wF;N9Ju8154U%zp8q=6
ziK(;UYBgqHWl^Z`(HhJzJ3Kx-QTa`kF*cZfu-51$pN{Jp7oe<dwv>_Qxk&@k<kJyv
zM2uH{wAz0%y`Iqikybt(7#J9;e7{y~pO&gA9`MvpUhC9EkY^<$)g(xingWSd(;zWw
z1|(L^f`rrvBu>qN#H)Fb1hoK?s1`wz)Y3oci$_b!GH-5LEni1eQdElahU-ed^hzt-
zo+O_R>lGHNEb!mNjkkK4wm*_jYe<ct>^)9q41Jgrb?K~pI$j+gque^j`V$XO@ZeSQ
zX+sY~uyR`YkNLQ_1Io9>-F{xbb`J<!DbsPi<DwKwx#9jk?%jbx%lFqtysx}n`S-hx
z1c|ayAkj7&B*w;o#M)SpkPU&v**K7R8xN9T6F?GeB1n==`X}9M+w*lz`IuDssv<HX
zrYcvXY;Gca+?yJ3ReEB8=c))JG0c-lJqbF}NP>>iQ=p@a6zFI@4LaILgO1TNpks^-
z=vX}qI@ZX74(So-kP(57({rHXj2!5AJr6qG$b(MM3!oE>0_a4&2s+Uyf=<#)pp%Re
z=nNV|s4|u0Y51BULP?K$)K%bv4!l|hLdcMQ$`eucmGXRhwW$P81`z^y7Kj9iLKH|e
zqCsL10}_i^kPt#3afkznM?6RZ5<n7>2$F=Pf0DfKEtvjl6I@c9Tay0jVDhWM{pyx|
zTpZhS)o(&~@0`pjle4C(rYMz4mE}3i^c+t09Oifqa|^8nd2&pdHKv=|%e#Azp1ywm
zfdRdegG`B3kWcc6;4n?dYvJ(;gN;Lm4tqoF%?;{1fQkBlXG0@ikBrvE5G2haUg{UC
aHw^3(rxp5->>c{1c^aNHcK9f4^#1_O8?#IR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp.F48BB29B1F2C3D07.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp.F48BB29B1F2C3D07.idx
new file mode 100755
index 0000000000000000000000000000000000000000..938419e3cc1f6f99011aa611cd737447743a37a2
GIT binary patch
literal 6574
zcmZvgd2~~CqQ`T8ghE^FmZ3n(7IXJ~fzm*1A7Uy^*@QyMQV<N3rGT^<1eXJd%TrpA
zstC-8ERLW!2tshA$hcsbR-`&g)S$8ngW=%_xD|$Z6Z+1j-#=c@={@)J`+d%*<o5o0
zf84QUrKPX*ayW*C%a+Wa8JX^MI2^gD@486soFknajxRGDjwOG5XF_@1iNc)oFMTiU
zeRO8i+^xO()JzNxXqew{wth|Tu9b~FS2!}4Hy)pMvU9(6Be!=MGJh7G(bVsk8=Lc%
zJKvqPVgD5Jz}EHWzB;?zwdvjC|7wjas$10d`?XWmBhON2&t^XOcHb$LyIMzesc&vu
zetGPN$@9hLf?fG9oGkh5^!2v_2M#q=z0%|2@tP6!18%%IY5T=Ddi?s)iKdG`o!MAa
zLJl4Mhq8Od<I39Zo63JUzIy!PHR`!1F0Kr1N}oI-ezhSMS#@mltUpYe)BIfSfpx!S
zwf$N?H@fKSKC9geTJx5in)KPT3tpH_Ui@VE^DE95d|upmY+{dP4_wIFGKO7q_`drq
z4nBGGK;zw>nH7#TgSYM7v)dEg;D}^5gyQ>ZE}#Bj`<k3S^Bzm<y<<$R>$wa4=l$(F
zYkD~y{PCggcTsuP)Qh7U4!3MMwc_5V?yNh~sktp{-#~^EtIG>EeSX^8*0bU%`t-`3
zf!)V`%?aH7>2q_3ugsoVZT?cb<Xm{c+*7fQ!{*G({`0rqt<MkT-`@6M?+weg?cIHL
z)S_3NTOJ+$!<E77w$<&Z_;%bG&$u1T;VIIt?&(?gXBc1Jzhn2=f~V`p=Jk2Bx@Pi+
z1D3Zm{OPg%2cIrq>VGJ^SLi@>(N7l^zg$*U8(dqktIM62x_u~YdHKpS^Y&~#yJ*$5
z_XodLGUAx0^1rtw?r;8K=fN90c0|8&Hx1hTk7I`~OnU2~qgT64A6MOL@QyoL_jH*%
zdUb5^`O^Kp{<*j9OzFFC{{3@9sax}KW>d~~hJAIFbiSTnv+Ve6XY`rz4eM0r+(}=Y
zYPcHdI_li&P@v+Tf{a(%?ueHc?i;2)J^QJO!@|R}?+7k;mrQ=+{Ks`eTPt5%P~Et$
zXx{EmC)A`T-+A)39+Q)!YDeuH`}kL7q;Q4!yYS-`W6#xG>iT@k$Ca0=rjTdrO05~S
zg~k0|te^d-u{UPzS+I=$as9!CFa6x~;gI*v9G(0^BIAT_#PgR|j3X1L9+!{2`S9Mg
zZ@pTb_dxJock8RKzc!#?^rl1S3cuO7&;NA8UE`Vh1FL*e@V7@Ig@dOT%&goZdK%jX
zo;=g|#Cv}obDbW!;gP0y){Z=LVnNxk%Nrisb>Psn&%XPn$Fca*wO@br$Q5PFo)ITs
z{)U?|<i)JI0Wa<kzVg|>zm04>c+bL8-+J?xW50HHO}<v2(eiVv<bL9dibJ<;pR@e)
z5WjK5yi?Qe9sj96>%^%Z>o4gaXYbD{xo_c*C;KNZP5j^Dg{>Q}4!kyP$<@C$JoL|y
z^>gQs9^V{#?s(r$^`qxxfBA|0(e|2uEq<`9U+MbIGs}Kp_VsRD<|rxpZi}OC>HHbL
zrPg&?zrU;=wq;QNOfqhQqso=$N^d`;X4=0pu5JEyW%<BynPhBgx@TI?-1dV#o%_a9
z?%(f9evnB<r>1k#a?;xm_B6M1;gFW=-&wag?nzC%^WC2IgFUUDoc7Y_;^|s9$N1Fr
zpw5E^wIA&1lIxc*?Kr>mjZ9LJn$Azl@6vv-r>|Ta*0%i#`jbo&N=^4o>pP(RU{Bx3
znegT}3lE>qBx6$3xoNrS?FW0hcIK?Q?_C|Ae^5sI=Wcla`sUY959!~n%4>NG?H`dM
zzqbS^;2qvUv_*qS`)M%g01YN%F<>%&223WvfXP}cn5>@#lMS$7aux?B=jXuW0vwpU
z#e>QFc`*3^52j!VU<!T#Od%kEDOw_!qF)4442WP#mIS8cm%x+)5}2|jgDLxEFy(*@
zreY~zDt-k_C7^(*S}K^TUj<VQs9<WA2BzlMz|;a7n7XBdsrz*>^?(khVHsc=egjM+
zV1Q{_CYYw*1k((dU|N<1rscQ5v;r2GK9?{5)>5nT_4E}0_3`z$$zJQJrEqhh!%_%(
z$9LTB-jH_&P}sY$gXka)CLN-|q{B3rOppPS2{B+YVFpY#$b!j+STNZz3nmxjz~n+4
zm|U0xlMnJ>@*y5fKFotD1O+gKkN~C-7QhsPBA8-G1XB!)U`jy=OerLRDTO64<)93v
z9FoD5!!npkPytg3DPSsL1xz)lf~kg7Fx9XMrWVw|)Iu7VT37>959(m*AstLTtb=I;
z4KR(60j3c)z%+v<m}bZX(+rznT0sj;D`bIbg)K0BhR<)`?%k?<#lAA2a^D1-?B`l)
zDcpRh!BTK}^E+;LuiM)P$m69uh<4Fn(ry||+CziMxEL@QHv=Z)VZdZvESRjD1(Wr#
zU~(=FOwP@L$$2<1c^3~R@8-eeJv^9#O8`@F3t$Q!0Zh>)f+@O1Fh!3DrsR^ql-v@S
zl1BnlcFADMZW&D3BZH~96fhOH0;b|oz*Jo-n5tU^Q}w7|YAy{-&8>l{c{DI}mky@x
z*1^<0I+%vb0Ml?AU>Y6+Ow(n8X}V1?O^*qt<+8xE+!mOY#{$!LyN|JN_taA0d?JwS
z(`>T$aCsT+dpB4L5$`g%-6J#*6JdbZ2n)nTI3Pa40|^lUNQ{U;QbYohBQlT@QGnEl
z3Zz9eAU&c384&}>jF>=H#JWZH#h^0Ucc}J7uy?4{8SOh$dpNaSY^J;|RL9lvCMfR-
z<Rs<2h#V!}{RI3N@g6~r6Ym$mbTdU`pH5IT_UR-=W1o%^8vAsN(AcNrgvLJ8OflGJ
z5)^}dCP^{aXQG6`J`*Dh_L(?gu+KJAEcV$1#bTdLQY`k_C}FYB#t4gjHcnXVbIlZo
zeJ(+9*yoZIhkY(eIP7yV!eO6_6At@)GsR<{Pf$Ge`6R_-pN|qA`+SV>*yrPf$G*@)
z2_M1zAtWdP`$CcuurEZ3fPEoG1ndiOB4A%^rbPVriwR1^zL=y$?2Az%Vqc6A5&L4C
zh}f5!DGB>hf|9T=B`FE}Qj|#8mtsW1z7!`C_T?5z#_J;|C>i^5l9I76M~RGmIYwmc
z%W)!OUumHf>?;XM!M>8D6znTeqF`T%5e557oG92=TPPL#YJyU+uO=xK`)ZV^*jHmj
z#l9LRD)zM&O2fXEpfv1jNlL@M79|??wHVQ`uf>UmeZ7U!v9Bj69s7Ec(y_1aC;D6P
z_e+lv9s7Em=-4+}C<FUOf-<mgBq;;?#(rX8--r<d`$n7?*f(1!6Z>X@GO=$aDHHo<
zl$dyZ%os7TZ^ns<eXE7Cux}+O3;R}*vaoMOiG_VDMl9@Gabk7!hd%i~j<%m73CdaJ
zSN&7$httSGjSKUY`e&$aRXvM(4zvF|S1~G&yRg>ZIW=6AUX*Ryi_7miDpeMaNe#R2
zbXVAq`4x}+<+1lpoqqe~@C0|2J^bX4f!@B>dUt9#CnqQ0-XE{odhcjgZQ+&N)FaR7
z&Fy%A(M}qOaWX)xlLg|O91!p1fdr=jBsxVP$teNJP8mpXDnP1J1=5@vknYrh45tBP
zI!z$UY2Bio7gsKwc>0g4Q&*#}zwc1{W52$nsq>imTT4^J`C0k-_U_=&fU)7ufvp!)
z!!E|9*l*n4uPFY0TE8!jq=x-Y|LykgCH}&@7sR|F4XI&YS6`kzd`|EDQ<v->t5U;7
z-HZC!!}gCR8SO8>I(8s<W7x04Qd_v2@4-4Gujp0mLbobfq=6Wb0b)fKh!Z&=UgUuU
zQ2-J}5l9jxAX$`w6j1?EMHNUBH6UHofeg_AGDQ=}60KY0sGV6e{pQ=hy4w5d#@#%b
zc=PaoUz6<<N2jXtV!k?O(ZAEvrdCd#IBANb^46SVQpf3(j?<|fr}xbd*UU~~`K*zh
zoSoCUbWP9fc3Vz%R#|Rf7@M0>c4uA>*YA4vD)J5S-%(uhUut)0u(y!!_y3LN7xeDa
j-*Y?7aDpgnx?v3*)OUbeG4JZ0e|K;iDIGO*I5pxwmkMRm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.9E7DF32B7E3B06F9.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.9E7DF32B7E3B06F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f513997e262dfbd97fbee257dcd3e3f54d2cf870
GIT binary patch
literal 5260
zcmZvf3se(V8pktZi1HMSq!0=i=Y8jaCE-<q4=^Y}5djG*Xp!Kfh`4|e)T4Fl)>`pV
zs#RO9NNZbLwiLI)*2h|FHwvxNS{pT5#kLBf<)GLKt9Cc$oH5^?vpI)*=6}E6`7ty3
z=1y)-Mtb_XkqSkcHKV4yw7OWSP$)v&e^u3$i_Q*FD2Dkf6g9h#<<HK&oH+82y|;3^
zCMpYB&jbwLTL0!HJlwo~(w5PYm)|;FZ`rZ#PV+S)IqS>nXsvB$#<P3+Z(WJWQf<X9
zHh%YJW@lDjSLxmRgRd^=EB&`+^|F<J8GMY6@7~m*s+$(meQDLw;`nh%QDe?#4XxXO
zKB<Tv@yj3I{CK0Ixq4vm7qs`<wKMlN4*ov4GpzQ{-&VZSaeS`v+uq)admAHue&jb0
zzsL4v+{cSIHQg($Y|d|LDXUIOD6BYoZ_35PE$_q_Ze~`i;$~0WPvxd}vv(DPCn@`^
z#NPOofy<e2W98p(WgIFqtj${Z9va{8^WxsgT}_$C`Rcc`xm@~ZE)iWVDK0wm#@1N$
z(xjsAE)^66@1LSLyxzf=J>NPoaJQeCyxE>QB*?&i;cH&3D?9zGzBrRK_mzcC^cmqF
z^?UxNuO0h3*S@eJwd&N6wn1?rL1EejWf8WXz74AaK1#3eJ7EfG>Q9;Y{d9v}d#*F>
zSlq4^Nk=OW8bUsAnV);Pd*2U7CVdt=EN*$;mX!YJYEAwXPTD>yFSxBJJU6K|=*RYz
zX5&ad`?_sq|FVxhfo17b5%Gab245Nd?ly9LXZYoKF0jE|!tIM}zMj+DaW(D4_TKQI
zl*6Hma_V>I{QAo6wZVJ#B~s_}qAJB1)yCfJxicsIzWyJ%!!+kUY%M(a5LFyUFB&pS
zE~uPWc8x6RvTdoIGfJ2@z2W8V+6^heDM@d+c12t&J=$?|V+3|PJ@-`gw{iEVrqs9I
z-1c61A-7;+?W5$zkn^ee_F1jE>>mzmD)xMkw)CUwqX+6=TgdMGWBJ^eOXq^V_^_VJ
zdHwyqmVS1~i;w25PAT7Oxj$vUt9Qz|`uTwx*K5>-V-G&<oR{%a#@R(PK3=+d+uAb?
zzM(U9+w+nlBKPHdQKdPSnb~<~+xK-l>o-lCG>MpXAn$m8-l1{&n)aZ}3mz}IoAs%o
z-+80;#FC(@(p8<?#^=YszP;eW@$z3EoY!yQsWH>K3gf5Fsq3&^Td^x<>YO)rRu!br
z(j@y3-y6c;9b0ntHSM2ad+Ih{ObyRJkpAl86(@HjzT46l6L9A}?o>sK@!-rAL(VSD
zUb3Xiuw~JOvx_v*XhF%$(g(IZk53OfJ^Ztu^)0)nyfXOql;@Ydk+G&Cwk@bNJ9{-<
z7den$GB0EAbknk$BaOuk<!h?f+RN)7s!Ykd#79B-p;LeR)O7X49nBA`N=j-r9y>Aa
zRqLSzk00cXExUN3`#;yuBnyALdeWAV*EI64a~9>V3>Y)^{qo|pBZbRr2bZqOh)&<x
zaKpTL{%5HN*I;Ge4XmrGEuS~we!5lBpKngvmoUm7&&^j9sa0xU&)}|jKEncHU!oJP
zYz|aRcbByqZM0{Q%YOfExcYKtT&_Q!<1UA(!a_WQTn=g8pk4M!&q;rLn!6mV3ikC3
za(UdtyT5&Sv$okE&vTbGD$P*OAeVDbfAD8xP2)!aiW%;5tWWF&&mfn(9m==;&Og`c
zk6Ya3C{@&0&mfl{2j?IBsp9LK{&=>#9HI*G^$c>kvUGmci6HOwvwoiI-t^VOhNkPu
zqXLQuJrOVe3wJUS<3KVIlf8-5lYk_RBp^u>2}nv$0g^IOfTT<mAZa}fNZLpPk~YzR
zWb_Om86yKo#>4=U)w6(PjVvHp6AMUA&jFG%a)9Ja93YV%0TLMzAdv|HlGpQq<c&Na
zc@qyvK`#JOFbaSaOadT9y$DFrC<0P6iGY;!5+Eg`1W3sw0n(s0Xr6Xnks-_w0aB!4
zl$@SE8GfE80yK_|SOc2FMuJ4yD3E9y4H9ExKw@nyNSuuW3E2=xyp0D*un8cEHW4Jr
zCOu8^lk4vLCyy;SeKRq`d-{Zhm<N)TsPHDzOahX$kboquBp@j>1xU(50g|#(fTYbd
zAZZH?NZLvRk})%YWGoCI87l)w*31HuwXlF>tt=opGY3e{!U2-Aa)3l;1W05-fJ9aV
zNZ!l?lDF`H<gGj)1+xH1!6E=sunK?_%_1O0iwH>3Dgsh6OMsLt5+Eh31W3bk2BX~D
z0Y!$fh76E04f%4C*K~K|JXz2|<EROZw>bzc5ebrxAiarHlYk_(Bp^v02}nv!0g}>E
zfTVO3AZaxXNLot+lGf3HWYi2G87%`yM#lh>RkMI(wJac69ScZK%>k0ra)9J?93YVz
z0TO8uAdwCMl2`M9<h48?c^wZ(K`j7M&<cPQbOInnwFpR2D*{r~iGY;U5+Eh51V~9I
z0n#wqK*`PFZXCvdKoSfwa+2Tn8GfF(JZPLYjA+Na?LoLO;xf!V7;zhBJ5IFXfOp`;
zDVUu&aRFqq4I`nSbYUd)lRX#-{iGcyp`Uc%B=nO`oP>U=4Wpo+a$ywoQ#}|3{gfT2
zpr3N!6!cS0oPvJ34Wpr-c40L1(>)js{j?pYp`Ui(H1yL>oQ8g;9b>)$_m^>D4D>TS
z7z6!`9cQ4Qao`N}Gftd=ezpx`;rX*JjD>!-2V<e1wc{-Gvkshve%6Vz(9gAD9Q1Q8
zjDvoz2jifhv*R4}a}Jz?e$I(=(2v?M1pUZ`A?Qav7=nIe#}V`+2acd0IdKI2d^^T}
z1MWZX!g%QCdoUjQc{|QSKkvYK=;xg{5B)+rCP2U7!UX6SdN2X{g;rcR1fCzkfeX+t
zIB@~`#db`De$j=A&@c92BJ_)`xCs5C0~euRbmAiPOYN8h{gMllpkM02B<Pn~aS8e*
z2QERs<isVff5Q4ZioyHUE=*aJNF>gdSF5QB^%WYiZYvg06qXp4CjVTL%b1^MFCs5u
zd08Wv?$?bzZd&iFA1+h2%!%Lf1@YPLY3+Dzj{K8xeV^&sq>ehRdpbBcSR=ntw=}*m
zO<fs(FTlOsP!b{D1&&mbAW=#RBw9&>#3&h%SS1S*r{q9FN(2(G<UtaY0!X4#1W8g#
zPt(V@U#^{befMVfbws5{Wy&YqRCB~9ySy>oJ*^4SXykSOpN1T(kE!vtds<DYMS0`-
zRo>WJbE7Yua!)5H6GqFo&bi~uZZsl!lY81Q%rHWpzA?t<{?OsiY;;fSL-o<}^vyE|
zj_qGHb((uxo2M<3Z?t?yxSwZhD1Tc4|MN}D8^u6B&j!&85r`0FLqHLUNRTK*fkY!3
zBnB}cv4{nULmWs5A&_{)gCrmUBoT=qNl1E{6qTimi=XUjt13&Xa-ST`d@^`GLyF>z
zQJb5)@|#t23TDroRj8OPmlTRw-opa#VWIbMj`y&n+`4$7J7&(GIz;KC8am9^KOiu8
zcu+=&DUA;C%SavZjCy2Pcw$nrF=gzy@lq)NJQ2@E|G!yH#Iup3bfZa%;gAqB<~e;_
XLX<%(#!d{?Oft{K)3YW_#-{uirswzd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp.8E876A3BCC68461F.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp.8E876A3BCC68461F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..23c19d8bc9fc300947a93b9facb0eb9cccfb3897
GIT binary patch
literal 6654
zcmZvgdw5fIp2l;Iq0kngZJ@O&Oo`^)?-vS$01KE3DO3hZ0~D%gpasgU%^-}r+*B56
zsZ>RA6$Korz(BbKMv9CghHVGIStPE(<<@Ow(t!bW#fm)7hRk#5`^WZq`g_j%dq4T4
z{XXa9awd;2Ec_|gW*ZwEzkK1G&}_TSX3I?c)`Y6(wLD<6eVb;pEr0vt(xRFR1sNUh
z{*61hW=`|`?YRT1ru)shg>`MU^#giMYj|k2EqztP`6n**$lpA1XV2ja=epCH^MAVg
zUe+r6Cv&$Po=HBl{mu4o+IBj&esccTv!SImOS^uzsa!twD$&-K{`-%9GjrPBv&B7Y
zTe?==n*4dZW2C-lZ}uCPMxDEI=ZNp<@#a|{<aC~|now)rJuqWu=f0fVf4R`y`Qz2M
zhK?eSpZ-p2dg?`KQ~#|+H_mUIx~yJqf2nhAU~6hwY4nG>NNB^E_vZfljCn1uWggx9
zbDyr;Mf1Z;za6;Ix#(=x^2;;My}Ia)O7fjA$GyI~Bk!w`4QHn3tbFELpKX)q`jbyD
zDLJ<8^wEaLTysin^`0FE_cyuxTWq1eb%E%is#{k++gYD6aKQ^H0~#k)J6^jsXu*H{
zk1@N*?Z5eK|3`@|qr9`Y?qut>%d4MUKc?o?11(*B4h^OVzM?2^>sMDE>3XPSz5B}A
z-M*$Xw=;Z?{q?o^<JR__Q=$J<y}UiRX#VBMTVv-f==;Zuk8FQ^H2cwx-w)Wba>v1@
zw&JB9+PAG4cjJ4{<{dSSB^Re$bxmocPR<ne_D}7zBu)FTC5=sOc`w&a&KkI;qN?n3
zb5(2IA6__o?B${r-e>#f298z?{qfqe_s5T~_HWAD+jGqI-k)>Z-v9oU1^c(RE#2_T
z3D2%k6VAA%{o9V%l9n60kKJu-41eQn9@6yPnUmLM9C`Ni54~qksmS#-jySu&=lsVv
zMwWFH9?t#8!LF-?pB(tluQaKq{<-w#jGYwy;Rd0jmaSiTzS16kWoq4K**<^9*O%*l
z2=yv%-x%<f{5CJ`gRT+LqJl$X<(Dhhmz?CDs~q89<s4PEuj7lF(PyXaT2#@ndFX<s
zzm`^|#y?*7U`|=QxVm`v<QKmgPZq4^`vqStncQA=z1QolUrf6`YbN<>O`-8rb-~E|
zcWNvDF!}D>{fkz*Z@zhK@w@-q{Q2-tubwV@BbIhy*o4<_t)4<oFF!AyI`G`VO-DYg
z$a=>Ar1R{Ddv=+5k8eHRUNHE=A@9p|k4~j(k8T(y_<w#rRN$GNH)q;5-qp}G_|nye
zmp=XTq&x13Tb^(Jc+<p=Qs!3Ok@r9SdBU2&tbgpA_sfm+iq2!>Zv1^@*XA!PKUjR=
zvSa@{#p_nQRa{1$o40bYXX3@iAv?Sia-aQf_=fe;^MyUy?|j$)!Os@#y7baHb97wj
z_r<q^Zz@f@D(Q89yqSJ_*_$0b%htWSf5^`(Gp_%%XXU#eReaIqysI4Rx>R`l_U<je
zu4y0h<$~7v^+aUnq4cWC?2ah~|2q8R_xnCF;mZ^5+gUR*T1L(Owf)bwniUJ5${}pF
zl>9$!9J_7Epmee%U@Lc|I8u{K;+jwXr2X>V#kEC)r=*ip5_N7$Zf0_^>dbxXoxj@~
z|16!Hl&CXOGE$R^Rgb!J>w06yihb$i)I^<=lG8J}SoMl=TX(*l*3p(uPD|AJDfwn{
zvFZjj`|V#`?$X}2(nLM9$Iv0k#j20Dtu0@2EOq09w#kV)FFVhbT&y}id(ctIyl$kE
zk0<JWDg9EDi&fvvC_V7si%)i>lfO;WnJJm6$;GN|)pO?7<n-&F50#djOWlb(@9nuV
zd{FOlx8W|Z{((gDxyN-U%Ala5d=zxFK|@FTXy_P&fsXMp(6I&!9qVJE;|vZu&c{K=
z8$5KpkB3e$1n2~x0G()v(1|_~I?0frlYA0%vLQn!`()@8LxE25DbT5g3Z3dxq0<Zv
zI?bm+ryDwSx=)ABFbwDnp8=g|n9!L%6FQH>lYRf}EB6fa46>4SJq$@pUI%bG_}x>x
zPXKqo{S>62dvSN7{1lK>fC7>VQb5vv8b~@o14##IAQ?XcBoknOWP%Kkte*vv4X{A6
zK^92P&jHB=I3T$o2PE(3f#d@`kbIB_Qt%5v3IPE~At(ST`b8kcfC!`*6oHid5|C0r
z0#XV}K+1j@NI4(_DF<aB6~6+c5>S9tf(nqTUj<SPs6eVg6-dpm0jUKvAhn<dr0&;&
z)B`$@dQb<_@EbrH0Ru=QXaH&YO(4yH38Wb`f%Iq|uXVcjF87S|jE7X@DYcUI+Ml=-
z?!EBCrQmR9cc1QVr+XkIm)qT)C<g^3<)nb5TojPBg9egz(m>KK8c4>$0LeHRAQ=|}
zB<o;-WSuOKtcwMbb8tX%P7X-U#R17Xcp!Nv4<zs6ffO79kb+YHQg8`CiVhJ-(J2Bc
zx<nu)hXkbLlz^065|FY(22ys)K*}x|NX4N5sW=rN6_)~}>QI4Hohp#3O9fJMXh3RC
z4M@$U0jWE5Aa$n>r0&vzG#mzyhSLDja2Y_F4iiYzX##1wOdvfEdnoI4Ph1L&hleD3
z6f0Sub1Kr3pKst&Xd&FKME42rjuGw)nB#=I6LXk!A0~m1knU5Mqon(5$W#kKp`VHo
z6#A(+L7|@tlN9=?2uY!zijoxi=@x=UKOG}z^wV*IMn4@UY4p<(l14urC290CEd+yp
zCPpymXW|5dekM#Z=w~7%gMKDTGU#Vp2p0Wpj9}5v#t9bvY?x%x&qhcV{cMzE(a*II
z9QwH!!J(gv6CC=vFv+2xi;x`pxhTn@pKm4jzrgvy#|R$%e4OCX&xc7K{d|Pv(a%Ro
z9{oZKA>iLH#0UZXLYxrLFN8?}{X&Ek&@V(u0sUeNA);T55hD7<I3c2643i@I#Rw^)
zUyPC>`lVJv!ud!sLPEb3CnWSsVNycB6d@(_OHoonzuZd5=$B)JjD9&z$mo~Dq>O$!
zLdxiuqoj;}rIk?7ufzxi{Ysor(659^1^r5dRM4+PNd^6CE1{xajS(vP)i|M|Uk#Hg
z`qc=jqF;@YD*ClnLPNh6BQ*4DaY94Cc9_(T!0VS5AvN@CQBp&{-b(1`*JFf^emzd;
z=+_UEI{NhpsiR+yk~;d0R>DBP5hD!r8*#!wzY!)4oR1M94fGpP(m=o2N|@+3V}yx*
zGftT3H^ZcfeltRv=r^OJ+3g>_?(ep)A44(1UOqw|G1GcCnmD9kakf<Z3emefcWCZd
z>nlr1aTarJleb4=dsynQz81Z#=+V={w1P>A?YuF0CDylU>ph4}OMWNseoMn=TK)W=
zUikF#m5=T{wlr^+_1JYceFKJ_t#u~0Gcq!=tq+Rz+n;>gQC;wTX5#Cz-JNMYL+^6R
zPC=sWG$h8(Kw|AIB+kx3;_W;n!7e}&?II+}E<uv*G9<;WKvL~0B+aft((O7V!)`z_
z?dJVt4Vp+x-oCIVxw~`iis@J0-k2CtzBhlgHH|&Xn|n-JxV<p3ozo{L+xld2+?*Wj
z;oIJs*mh73$+~M}O*|_tdGEt=Ck7mHue6kj0f#6^bclw;gcwL{h=s(3I7obmha`jq
zNMcBYB!whMa!7`xgcL|>NQI<@G)Q_#hh&5dNM^{qpH7sF{M!@xU!O|MW`upj!`8f`
zZ+vo%&l+Br*!J}DWLewoYL6d#_RZOl*dEq@SiZGw-7a#ZC2teKB)#rYFi9^3iT2Wv
z7%u~f^|FvSF9(VD@{k0t07>+UkR-1JN%qQ+6t4nF^{S9GuLeo?>W~bt0m<~5kUS50
z@~n|`_S#cVdYbONj`XAYsn&$7J7<Y=?%r)Pm=y1pU{X8<iRNiY49`Gfc@`4KbC7tR
zha~U<B#{>(NxTF}=4D6<uRv0H6_UnlkaS*$Wbg(glQ-|D#Qnjl+4pW1)>J=TGv(gF
zjC;#Je`Z<-wg<}X=XNw1CoUdLd1Cs^8D+C<)9-h-vhKrK-G@(fA3nV>SXG&bMRO-U
zVDFLAvsY?*?*}vb_8Fh)8%t-VjUSVh<LLKL?$BYwy(31B`WLys;CZBg&Hw*)v-1WF
l9OQb~O)(rVDyn7-9`YO0Dd~^)&wkARBw1KIdK@v~{{R7wfk^-W

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F358D66C9DE4EC29.idx b/.cache/clangd/index/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F358D66C9DE4EC29.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f6f4efbf17f9b8a3c1023b7a8c040173039f13bd
GIT binary patch
literal 5340
zcmZvf2~-o;9)~ky1PLHu5K0h3aMnAU7@{H(s~D8vf`UdwaUm}FKzcwURP@!Ysai!)
zYqe^XidzvE1eY%ESWt?dr_lzNy5JJ9wzjrfYx`o(8Tr3+9_MhG-~D}jWG*+8nbAp!
ziC2OZiUey?VeYg!Q<MsYBFOz+Feg9zQU`^?&sU)+-2Bbh^t|T2fsW4~r$$fTY^~mh
z4Irv+dNmV?o1#ka;j4+$!FyV!oPX6ik4cYR_fx_alOp56#q-AkVmo}N$Qd>B0{I@j
zH#^-}9Dg^gc;oB!P5YaqZ<~HydG1Bd$rryZ&g8vALoTenhuv$NbFy&H-DiJ_(N!ld
zN(&cOoN3bjT3j^0`n#qL8~zg={p_#imR^$<4mgo<<3?u7L(AEhU4PE)-TFcD*dH|M
zEywO>=8nrc@+g1y(OX3=8xPFQ>OU#_;<ZP9vuanElXrf6Y1CcpnLT+{P5zB1TVrl)
zw!YU?vy;AAby?YOd+A<9qh<Gks4dmv>)Dr{45}Em@gw_>y9@tVHrFt6P{VFT_qhXS
zWKA)x>xDh)uGwGt`1;*@=bJu{jCH8|`(E01pJEQ$D|Ks|mI|fJsk||PU5E08SLPeK
zgxeQ;FD6xu5dp-*Kenw}r<}QYUv6QVJ>Y&;ilJ<2fVg7vmZ!QEtIPLXUE`&7eN%ts
z!kzVNE0&*JW%$aO7_j9QJ(ydPQZVFs;ncsEe1H1GxN{X<ma3wTaarj6t?lEBMlQDx
z+WyI!{C!8q5`9myXD6SjI@SMnigjJt{Xk<@GgX15W%K0Wdxi-OyQ=?4d9PyrhY2OK
z6E>|Qb_E~1^M2&u8O#VD!+QIK<gO7VaaSykQLpc`1utF}vsc*A_H9G+{uP===k<My
zgn5+d+OeBiTSGg2IWG87<olUZ0|t*=z0>v}(H8c!`sLlklgBS#iH;~(S?vGWnw=z>
zt+h0Sjq3i->Y2)#HDmUDvY^wI!&T=p9zG7-j3V;(?ugQ;cD`P)cSBCLeL`A&mpwyM
z3vU;#!wd@#e!aP-DYk#=p5Nov9^R3@WPR<_;VDo1R~{XgI<Kr?{8copq}PG4nC(&A
zjxsdkiympH&$P>59aernBCPQv_J<d9=ERh*H$VSu=keDYPbJu{$G6T~Id5^r#v|ep
ze_O$)+`SIFl!p#2uRMGF@7x|Ic5O|%!7M83_wZ2mv#G_uzu0%}So%Z5@>u({9n*f<
zJ(<`<WnGzZF@4k-?H6k=2W7Bl(A7?*ja4r+p)nzUFIzMuyWy7%z9gr2b-(1iQyu*;
zTzy_Pz1Ndwvhm);mK^@eHplF0Z5Z})(I>@}V{Ls;``V;kvwJ;$I-L69REVnK^3sDH
zZs!cEe5~KQJ0#B%<9GN;sd%7vO#ba{Gj~-lyM1bJd3{7}ecY!vvo<}=@yk2d=S65~
z+wYbCY`r+c|CQ$TfUk?}=Pu2^{*!&h`azfe)i|pJt2=)<I%j5RTUKU#w@H`mLDbAs
zWnYBtD{Y(9WPDj(P?S406jLa?v`1GbZ0{54i>Fx>nQAY!k7sbd^F3dxKeydll59xz
z#Z%pNm{(YkXOQcl{mV6Tt6XP%@f3F*=oRSW8RUAvt6!e%eo(aE7a!xUL%l-1J%d~i
zDfw%`cJJri{1ju|wZ5Z%pl6Wlv>Ri845I$&=ZlYX*IF-ayl0T>6A|+c#a{1u!(TDl
zU5AH+YdwQpyQV}Q7h<01eesd*I@l}N$1}+F>%g&xZswhR;ERuP*Fj!EKAu6Y75USq
z7d-9UzCPh9&$_HS{c2m~ow!K9OhQlemS3SenTVn7iPV#zBTXdeC_M!_%0z*V*3+P)
zO*H5jJp($%#DI>~v!G*5Ea;FPfex7v=r}zGI?lv_j@R>`<4rv11ib({!6blA)Qg}K
zO(N(dy#zYRB!SMLHiW$05SfMuL!_MKJ;<m$dk_2z%*2@XMJFu8RFJGhUV9?VBp^u(
z2}sgP0+KRQfTS!GASo*aNZL#TlD5!*q^&d{88ZV&#=-!Su`+;U%`6~U3kyir$^sIZ
z5g?HT0TNjeAUQJ!NY26mlCyGv<jp)Fc?%Cn-pT_~FbjYbECL_}s{lyRECN!rh=3HW
zA|NHR1W3su0aCI`fHXuKjPmC2%QVCrl0Zr}jFppob#iZ<H<u=`anwXe`{p1tL<C4W
zf@n{qngk@NApuG1NI+6*3XqhB0wkrQ07<K9K++l-khG2lB%@{k$!HirGCBs3teOQR
zt6>4j>R3P`H3B5kAV4A=0wkyA0Lf`MKyo?`ki41)B(LEC$?JGP3Tgq6f<^$Opc4Qo
zszpGG8WE79P6VW+mH;VfBtS|!36O>;11WC~_r{?O2qfMh$w~g)N>X_~ufWEs$A|{3
zeR~j2jA(}0g%PbV*Wg4g4){8pI1jS}C$51^)?*~}lTM6;e$s`J&`;LjB=nPYI0^lv
z11F)Ms>dkkr<@oC{gex%pr5M2Dd?x_a0>b<2TnmhU60YwPdhOh`e_$NLqA=E)6h@X
z;WYHq4xEO5rU7Fvfc0gZ7z6!`3uB<4slgfOXX<bU`WXk#KtEfLvGDb?PK<?q)`hXq
z&(`29^s{w13;nDEXQ3a}V+i_@6GPCCTo{6WRD&bvM|C)Ye&oOr^mFwX2mPEA<Dj2&
zVI1^xH8=<TTpi9qKj*+X=;s?S{sLHk-ih(h&$}=l`uQ51hkm{e=b@i>;5_sT4VVD^
zf)f*<UvOap^b57Pa0I-6ggRV+e!+nY&@VP%BJ_(+OoV>Xg^AEF*5V@ci*>jN{h|XG
zp<imiB<Po%m<0Wj3zMK<s>LPfm+Ei{`XvW0wfhI!+Y~Q<o8!cknXx@$Gvuq4{1a1o
zZlC3=)$l%Lc_G5bYcRjeu)bjl@?BKgh%WS_Qe#K=bc|0-fc(qyH40aGZn9;G-~aq2
zTk_zG+_>H;?%xf6H+-~w2X6n_)GfB@V~u+{FfcGg{@5)kpFC2X-}{-L`}SB#1j&EF
zFOyV~AW=#RBw9&>#3&h%SS1S*QX-HzB?l6(<UtaY0!X4#1W8g#Z_}5pON!F(Y+mg?
z4{f40S$@Hlg<o|{$t_QGPlt964Us<_zKt1e?Pw}*bx*5FwIDx8$&2Ez@;r{TFC<|j
z{w*urYhWWmqHGjMw2cOdu`wXAHWnmgLm+WB4kX^jgCy7lkVKmZl4O(KrqgNhk0)ub
zop+yAUuEAYd3!iEoP313#I15q8#)=f$kQ!7JO1V!5W3Pm9n&>ND^EXY-1ALE@$ixE
z>F~7hiSm-jk7S)yo<}jTG)7_oSQ;Y<5@n=7qK!02jFAC}HL@TfBLaytav<?W9wfmi
zfFv44kR+o7lHnagxO|SSohnQ4eh1#Xy@JW$p7OcL&z0_f|C@(OaE^!&z&RokBnnX=
z(TE0#K@3PNVnIR(fy5yWBp&e~2}l4*L?TEMlHMlwb8pU+HxI!D`B??2Zw|)489ZMd
zGG`6@thw6}Z^>&?`h@Wr6BQHWN}<STKTK~woY;Oisr@i3*P1ia9h0XI@1X4H<=x50
z*UvvNpmS1?DS--7B@OBlst$f9tZ!_bv48x4cg3z;zaG6A?f;t%3GWsWsf!{=nnk?S
cGg@!x(>+Eb^cvhXWQch(o;YIQP;A(L0iCPp)&Kwi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.B5661BE447C6D950.idx b/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.B5661BE447C6D950.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6fe332601f37bdc3ee3f8af254d901cf23966685
GIT binary patch
literal 22954
zcmZvk37BJLfv)TS&);oqTAFTz9z_Ze#cU`0NqWf1zK1}Tll5c~4BHqoK{(^nfUIFj
z2m}=}4h9V%qcI}sK_dhN5kV{(y|{o-wBU}Y=tUWUAj4Gmz18*2eQrHZ=c%{fuj=ip
z&Uz}TOe`F}y|lCxI6bzzd4_TtTUuJ$zV_#Q<-C9TF}Jkzr$d&OcHi=7zI6U`_QQVr
z_R5`SKXun<&kk>M-x0j=nWumK51x-4dur?>jOq_P`mO)0e)6U>_WmGn#61@ucGb@|
zJ^0q1oi~%ym%gF?`fqmJd-zj#=)ZBy4|Z>M{B-o}tAF~-@AuwR{r=ToIR4Q$&-%+b
zAN%@g-`Mr`%Te=vyE5W6|MtSg$!qT)zVy@$H+=o|SDKF>`QbmG^~dwfqph)TKl7zG
zuHYZ6Tr&FEr4O9(nS1WI;QRmkrNw>LC;s~Q)BjfZ{7HX$>crL`Hy{10tN-x1f_vvT
z{`t-qPrCk`tA23n4?dIrr<--hD>v<a_>CVYKRQ$Py?m(ZwHMzbp8M68V)r*}m*>yE
zL+kzP2R}LrzIf+JZ#R_tQyS@n(~f%I2mgHAp1Xeg(kCt2%Oce$SBu3@-;r2az2>{R
z#)-4tul?ye7eyag4Mg61cIv#tZ~EQIPk(anhpyqf>TkEQhaC~s@BWGFx<|fy+s`k5
zV#ig#?Z4y17cct4_3vh9-}2Sw?km5;e{Aa{eBv+Kz4sSieC>q$j{eyf^N(3xPrYW4
z_-l54;H1MZKdSNZ+4ajO?rVp1?bP{^H+HGt_Mdgy_doZ`%scPivSY5R$0Nyc{(0%=
zx247(yrZbQ>xiSpyMKH{`(5{MzUYD5hUw^z<96Tvjgz-<;~y2boc*Q7y*FJmdo6YD
zzrVKSn&qE;_LDb#Aam{6_dal3@T}i`e#b}S+tMe!=jcP7Uzy(dg};C2cl3&@elq#$
zzyAHf&U1~}^=F>($L~ILwe!Y*{^<{&e@-+W_OJh;e$w~qHy?T6xC`I^*~f1_>e<jI
zj`{vsPWtrqU%d9En~wUmu6WH)cun-J2Qp7Y|ML1-!Q;nQ-ucn=LYMmCeNR9C=`V%O
zohaY@@{XWQx%qEzf71Bu<A=`2pZ?Gl(dz3@njgPx%eRV8OrLx4rq63~Z$0sWn|{*R
z^OO9Y-~0Mm+fFj%v!4r|@)h~2mC|N=`-JSTuR2z9zWv0z<G(rKf>#aiy8MTsQ(PZ;
z{1dy)yN-G?`tN@~IrIE~eecxa=j{LMZ(n=V`Kd2`{+|A`o9}m@{`l(;X)e3q*+;(l
z$bDySdhNF0(Qp69J%_*b>m5IhAA8XkPcMJ@=2IWMUHkrRkDmML>*c4;Iiq;-W$){J
zapgSixA&FO-}!U<uiq{`m_1>dJwEvHi}x0GeWCx#^r8!|yiNS(J3n*z_uW5wWBA}{
zKU)0lPk+7QzFjB(uJs=Nnd#z{n~T>!_E>uP2Isz4c7LlS{@|2<$=&Yz(j#Bkd*kJY
z-jWQTd+6TipAS7!eM55TkR6X*aLF$(I&tS$c3!pm^jVL6O((r@>)DZDV>z2`ynR<8
z^s8T$j($2DnjH7OXCIwxd+Mlk_uZKz#*H^)3*DZ`ht2Q2WV(36yZc9n8o&DXKMqHh
z+pakIh2!qO=hRofm56@!>Steg{?l@J>HG_tA9;4|GR|%K?6n`C-EG~vl#!My7i_v9
ze&A%y9C&jdc}n0tn+onjXm(BS+`RLY11Ia+^X4lrkAHi?^bnd})7v(0d&hy3bzOby
z+{L3WociT=EhX3V_9M3+ci?1QPjCEl?#oI2;fK)tnm&E=>G=aE>-vd1jt*%rIqAZ8
zEv44<jw5$`;K0ece)5)2eDjHymb{11+?qaa^Kp>_C+m9Yyfe-`pUZB1g=EWtuk8AV
z-@NtCUpSB5e9XQ5z-^9WKf1NDEG+L@|3cR&v7Fo>wV>X$UfNnwC)CLe(g+%GnuG?N
zR?vdeCbZymf)1Q6p#!HE^x*UfJvd(A!SM+moIx;vGb9Y)jDiuIF<}H}5=`Js2@^Q8
zU<PMSn88^D3ph){0?sN}!C4bla5ljP&X%x&69fUAkPyJx1v@x<!Vb<MIKVj)4scGv
z3C@{tf)fQ1oR|>7xdazDSHcC(Ex5tC6K-%G!2`~d@PP9QUU1%o7o1P<f%7GN;QWFg
zoIl|Q7Z3vA0*L^)pb!KXOa#G&gb=t;A_Oihgu#UqVQ>*40xpt>fQt%IaM45*Tug|8
zizQ;<;zAr;JP`+%6q4YQi6pp`kOG%Vq`;+xG`Msk4K5>Oz-1B{aFQT_lM)iRtdIql
zO=Q94gdDhBA_p!n<iX_=d2j`x0IrZIfGY|`aK%IsTuCT_D<w+cWI+ZeCuDGCp$x8^
zD1)mA6>ybA1zc6Af~zK~;A%n*TrE)p*AN=J*01NcRvL*$a)Vkz3tTJF0;dQHI3=Ng
zyU2Et<KXR9<zm~#2WiXV*5}(N3DW*6^$uCSk!)O%FAvD_t-$IIQQt^5uFTZ~qP`Va
z(;*u4ngP+E*LH{&y>>vf=ye^UL$4bU9eRC-=+Wy3M30{D5FR~0AUt|QhZxWs2E>5g
z*da#r#sM*+H+6^!y=g#9=*=BsMsFSvGkQyhSkPMr#Dd=1Ay)L(0kNXDb%+hUZ9r`3
zg$@zW3j-pcw|9sgy?sFJ=p7y6K<^k32YP3RIMF)?#ED+)5D~pNAR>BKhq%zY2E>Kl
z-63xD?g4S5_jHH{y=Oo?=)E1{MeiLDFM3~x_|W?X#E0JBA%67!0r8^`bVvYwU_b)s
zgB=n?9~_V%`cQ|2(1!*jgg)FMVf5hv38Rm6NCbUkKqBa)9TG(!9grycSck;W#|9*Z
zKHed5^zi|Sqfd575`A(&lIT+%l0u&vkQDlKhosS`2PBO?(;*r3nE}b5mpVj3FAa!<
zKHDK#^w|N)qR(|m4t;Jwa_I9Nl1HB(kUaWAhZN8k2Bd(#*dayq#Q`azFLg)>eQ7{S
z=;aQP(aQrOqc3+z8GU&`%IGT{QbAuCkP7;0hg8v52c(L=)*&_YwE?N2Z*)lGMzV2Z
z)EJP)t-!4gX`ycoNDIBvAqskBKos;>5Ze{RvvD)kCbsqlZ=ZJG`!9F>{yQVGc1tBK
z=izorUakVQ$@0~Bt0k%PaGNEot3Yj{z8Y`ABn@mevIe%AHqqc#D`{b?m9?<dwuu(E
zI!Omxoveecu1$2f)k}KV>SaA_^=+cZEidt~<z*hWe4FsNHAn{78e{`(4Q*n;tx+<<
z)+ifcYituEZcUO2wkFvGTT`2uaBG&#ur<qO*qYnKj9ZIjfvrWhz}C_x7Tj7TD{QT@
z6}HwkvEtSy*<fpvZLnQCzu9mrNCIpHS%9t3CIW8lk{z~o*$!KKo7i#dkQ}ge$PU;#
z+Qfldr{sjKQ+C4E*(OfhijoLhQ5In<wuy*am*j%2OLoE5)g~_7x+OPk-Lf0D?ly7b
z)+2df>ybUM^{ic<aqE@5u=UDb*m~Q<i(8-MgRM{Y!PeI%KHT~xKWzQ7AGZEB@#8ii
z1z;PH1F#LWNdUJ&DG1x39E5GKO@g=$Ng>#V<PdB_Z4$z5SPH{7EQetmZj&%>BT@vm
z5qa$Z>;FBhv$YawlL&63QWUmPISSion?!LNlVY%q$uZc*+9ZbCxD<zNT#myw-X?L}
zCZ!~7lX4Qa$u>#iHYKHCo03zoO|?l1w`nO2+q9g9ZMsd;xXnlz*k<GmY%^_=!L1}o
zu$5#9wo;o&xXnsg*k<J{Y_n~W#cfW?!8Rx7V4G``9B%Vc9=3To58Hg3<Z)Y&3a~B6
z1=tqaq=4I^RD^9&F2c6hCPmzqq!MgPatXGjHYwp&mSot<vJ6|fO=R4br7~>Gav8Se
zHYww_B2{2pkt?vRv`GcGRjCTws$7L_wN0wHtw}Z5*5n#&Yi&})Z9{70;XaStr~<V~
z<7&JIB(-4Ml3TECwMh%Nilo3+krmh~ZKB}z65AyQAFyw&Txz=%=yGDaoE+q%-gN6O
zJns)jZr@r-uCnAlwsGZ<T4gEZ^eRguXI5DTSz2Wha(0zvk#nmohn!z!dE~+>D<BtF
zSrNIk%1X%cDwC1RtE`M%S!EUE>ME-u*H&2#Sy^SuKDPd!wSEVYZaHu`e*KvC*2*>H
z@M}nU{a)w#;rW->cKZ*`Z|s))bomCl{%O|kgO2F(t;kclya#zfm)}J0am&|n@bN=#
z`4;3ZTfUZUcy*tuZ=f6hed-Za--<k?>OIH{s(ur>$EmO5;Nyp!`WECaQ(p^b4^5wH
zu%{VO4fZrss==OSK{eRZ^f(RnG(%2<Jx!Nsu&3=)E%vk{s>PmmO10S2E~plJ+8(FH
zo_5G-v8U}aE%tPMs>7acM0MEHO{orhx&_r?PuJsg*wYO;9rkoxro*1TPxaW-kEkAd
z`YF|8Prsmg?CE=)9((#Br^lYY%k<de`;^BXKcYPL_$lSF$1f<4J-)~B*yD#Bk3GK2
zc<dSa)POz1h#IhGm{J4w3=3+&o}tGXuxA)@2J9KS%z!;(pBk}e98n|oj8kgFo^e5q
z*faJxBle6#&WJr@ml?5V>QfW;Oe1Q-o@q)=*fTAt345j<XTqLo$eFNb>M|4d%zbLc
zo_R#g*fUS58GGghHDk}*<ILDI4>>dT%w1;2o~2JM*t3kN1$&k$wP4S(pcd>|dYlD&
zmLX@so~6qy*t7Pj6?@hZwPMdYrB>`&7u1S9Ymc*H&pPC+*t2$-6??WmwPDXTqBiW=
zrqqT#+k)D#XX|k`?AeB#4STjOvtdu@QvrLzhzi&frc}V5u%H6=gdQhgPZ)9n_Jl4I
zuxIa6JNE1&YR8^^O6}OQFQ^@R_8w=)o_)yKv1jixJN6uX>cF03L><_3OsNBVjs<mK
z&(Y%?*mDdy2lgCY=D?n_Po3Cvj;IrR&M9?b&$*yZ>^Xa!6MN1f=fs|~%beH~`&7i9
zIHDr<#3>cACoZUnJ+a4$*b|4Gh&{2(MC`fx)P+6Qh`O-nno<|`Tnp;Lo~y^Xu;&_b
zF6_Cw%!NI7pSrQ<9#J>;+*9hto_j&v*mL(dH}>2^&W$~Hm$|X$=~EB(JR|DCo@Yuu
z*z+u?2Ya3#=fR$5$a%2m=`s)YynX7$o_9pO*z-=Q7kl0X^<vN4<Gk4O4mmIOyj|wS
zp07`R*z=934|~2T^<mGqpg!#RdYlh?z9Hwsp0CS%*z@<PAA9~0^<&RJrGD)B7u1hE
ze~<HH&p+h+*z<RpAA5m54PY-Yq5<p$rZj-Pz=8&_7wB;T>;;Bg0DFNh3t%tUr$Oum
zM>L4N;FJcj7hKRF_JTbwh`r#D3t}(WWkKwP`ZR>S(1?bx7n;%#_CgC9!d|Gyg|HVI
zav|)6x-5jfaG!>;7aq|t_QF#d#$I?q!`KV=xG?s@LoSTHaF>O#7wOXo_97!1!Cqua
zBiM^9Xaswa9v8u0WXMIZ7wNJH_M&|n#a?toqu7g1X%u_W1&v}a+T)_wiw?Oc_M%-D
z#a^sWW7vz0XbgL?DUD$-wxBWW#d=%}d$A!G!(Ob*V%UrKX&ig;5shOnKBaN&#TPV=
zy?Bp{V=q4B;@FFKSsZ)GK22gTIigAIC8sorz2t%>v6t*|N$e$uToQZ9E=yuB)u$=!
zrA9P`z0{PZu$NlU6!ua*E``0+kV|1N)nzH{rTa9Ez4VBtv6r6GH1^U9n#Nwb$EC5C
z9&%~yrMoPRy-c5Gu$LLp4E8cpn!#RXK{MFP^tcT6GD9wdy-b&7uqXAYggt3QCG1I4
zDq&ArPzifdkCU(`4LJ#WQkO~C%l2s&d)X1qVlO+TS?pyOG>g4#kIP~&JLIz1%XV27
zd$~T%VJ|nLIqc=8G>5(1g66Q7>v1{k<%V1id$}&lVK3jOdF<s!G>^Ufl;*LQU(h`E
z@;xq(z5I~NV=v!jdF&PXw1B<Bh!(I{n9>6F3JY4mUZKYouvZvz1?&~Ntbn~@pBAxK
z9MK~7ic?y|UU5N-*emw9BKC?yu86&2mld&B>eCYTN+VjrUTI28*efk)345g;SHfOt
z$d#~H>ar5{<UW<LCy%I%J$XuH?8yr%V^8jJGWO&lCu2|UG8ud2J}qOfJfdanm8Z0f
zz4C&Vu~+VKW$cxQTp4@iE-PcN(x(;dRYtUey~>nUuvb~o3ic{Ju7bVFkgH&?(q$Fw
zRr|Dxz3PZou~(haD)y=iTE$+q$5pXc9dcFdRlBT;y;`5vuvZ(=8un^aTEkv#L2KBn
z^|%`LYD2Duy;_&mu-E9*#tn4+@a@_Y;YPG^EAo^!_8>24<4xoq*SL;@j~{Z4Tade~
zaqYpw!&@t@K5b#IHKHx-wWhR%z1D)Zu-EEwE$p?1Tnl@xE^A>==~D%J%7`l1Q>IkG
zp0c0{_LLr{U{4ux3igyPQ#N|OYyP<O#;eLI$NrxK-mKCkhj&SK{Qx+tTt`&bk>WZB
zGK&um9{uavbZaHS5}Nhrif^rajD74NC5FG|-T(QSuh`ZOi>sE^a1dLiR&N}F)~GZ<
zT9p<^r_urGReB&^#RC~s1|Xx#2xL;3fXpg0kVRzyvZ|~=HkA!XPzgYGl^w{TasWA1
zP9RYw0=ZN!Ah*g5<WYHmyecn{PvryhtNcI#RRAcc3Ic^xA)v4-3=~mCfTF4>P)rpA
zimT#4NmUXkrAh&%RcWA%Dgz{`B%rJ+3zSpkfbyz5P(f7yDyoV=B~=MXR>?qRRT-$F
zssL40RiK)x2GmeB)Eie3EmaFhQ7H#0d&-Uv{q9rZrnToDUHHxmx2<0|)?ck*-g4ko
z8XG?>QqPF`A$(%djA)RxGonS-&4><JKO=f%enxm?!;Bb^jWc3IHqD3$**qg=WXp_L
zkgYRfMYhd|4Oy5G0ogtyc4WtlIFOw);zSl_L_~JYhzr?0BW`5RjChc}GvY<|&4>@#
zKO=tRz>EZtgEJCD4$VjiIXoj_<j9OfkfSpaMUKr#3^_g{apdHTB#~1yl0r_;NE$ga
zBN=3AMkM6yjAW5>Gm=Bj&qyA*Fe3%z;*1oLOEXeJmS;ppF3(6AxiTXa<m!x6k!v$j
z+u&1w**R<T3&%dc|K?zYEkA=l)m&lfXYi+vD@=o|U13^e-3rqo>sOc_nO|W%vSEc8
zkc}(Mh-_M6CS>ypGb3A8m<8Fo!mP-)6=p*gR+xZnUtxA+#|m>GJNGi@Dt!IoUM3>D
z_A(c;doOb%d-gI9vUe}@BK!6-AF_Wh^CJiLvH)^$FAE}v_OcLicrObhNA|J^a&#|?
zBFFZ!7;=0siz6pjSQ5V<sTGz&POq>ua%P2PkfjwSA!k=u7CE=Va>)4=mPanEumW;%
zg%y!YE3AYpuP_<8yu!-Jl@(S&uCA~ua&3jxkQ*zk@eJPUTwyI_WrZml{1Hp+{=-ju
z@ZRIsE=Y5>{19G{=1lz%oRKthra{)unHE_$XF6p4oavGIIpdKHb7nv`&Y2O}G-oDc
z^PHKHEpui;w$7Oq**0f3WMR$(Wc!@iksWj9Ku*qC5;-+zDdhB=rI9mpmO+;0OhV4i
zSr$1rXF25joaK=Vb5=ku&RG$;G-oAbdCp|y@|=~CD|1#suFhE%xi)7t<i?ye9)dHH
zGH1#Le~c~N{AJUdPwu}4p0MS6@QgEI>U;2vGhrHJ?SyHObrYsT)=!uonV&Eo*)U-S
zWaESxkxdh3LN-sB8QC&n7G&#$S&?lMW<wSxOhC3zm>t<MVGd;HggKGL2@{cB6Xrs8
zPna9oGhrTN?}T}geG}$G_D`4}IWS=X<luw_kwX&}LJm(@7&$Uw5#;EEMUi6@7DJ9t
zSR6SyVM*lFgr$(v6P8BKOjrh4nlK4DJ7HPm+=S(j^AnawE=*Vfxj11(<kEzdkmU)J
zk;@ZSMy^a)1-Uw5Rpi=))sPz#*0={QIIRh5AuAK6Z1BfV3O=DL+EZ%}f47IW7uN6a
zZ>{VI?I^6j-1^e!qV{Q%F}MHG!Lb)w*B>2Re(}9me*5RYn6Lfy_R{uK*FX09U)_@A
zFD;$1|4Mz8E$_q2<|<R;Q`j`COoOalWm;t2D$^nBSD7A}Uu8V9VU-z>jjPOvY+7X|
zWb-OBBU@IP1=+gFtjM-iW<wTMnSg9xWp-r8Dsvz^_c7=5aNQ*CV<NI^A9EqQ*S>kk
zo_)-N?A^z_$i98dhwR_S{K$cQEPx!`$AZYAeJq3=-p9hok$o(J9NovF$gzDah8*9=
z;>e9v*4TK;;@UB(Ro2?zul%8XLFpH_^tr8-o%x;RjiW5bC6CK)9A!DZc6xpNG3}kZ
z@86Va-X2~%3vN4f+mY+fZoKRKbHCQwaz=3NT)m6!Qg0m8IftEdkk&uP+A|^d|Du77
z2e<$80MWx)_e&o>=BJOX{j07f>}vA<^|SxhwbvuD4)V2xUAw{Sm!3^q4*ZgWjnBV)
zH(7oNUcuA2n=~H6XW8COTFA=XMA_i=Z^YU^d;f1Z*!Z;n^ElVfe|4GnQ``4ExAvXc
zQQA>o|Hu5swCUA%Y}?t}e^EZ-!eiFImfxRt?0vlLWvj4undO&F8{c|0uQsh0w^lT~
z21v_mfpokMNYCqmc%BC`@CG0wZv-;&CLl9!2D0!LAS-VLvhg+`ffs=6ydB8FJAj<L
z6G-GmAQ$ffa`SE=5AOl;@?Ibx?*sDlexLv!01EO!pb#Gd3iDy02p<88@=>4|9|MZ>
zaiAoh1WNHKpfsNb%J3N=iI;$~d=@Cj=YaBj9;m<<fQoz(sKl3mWL^d;^JSn4UjeG}
zRiGMQ18VRM)A|K@?fQgo0V%u!bfN9SgU=#=b<ZnzmoNXJX8()A#pTGx*?n1AzI21s
zqB^p1=2t7~OE*X(YQSj}4LGf+1*cWC;B=x6oKDe!(~Ej=dPNV87kO~J!h<u225<(&
z0L~~H!5I}JIFo1sXHrby%%T~bSuumNh!$`b#RASMTESTrD>$2I17}lg-~>?sCny3q
zyJ!bzSM1;%q63^mae#A*PH;}e2~HG6aH1lDbBQi+F2x1TExN(E6*oAK=mF<ZJm9>d
z7o1n|g7b+!a6ZKc&M*4G`4vC7fEWN5Py*nBVh~(V34#lWA#foj1THLw!G)DDxQG}5
z7f~YMqGA+WREdI%i7{|7B?c}o#=*svIJl&k1ea8j;8J1=TuMoSON(i6X(bIVBWA#5
zlngjYl)y=f1THIP!DW>!xSW^+ms4`!@?subUde+ihy`#3r2wud7Qq#jBDj)R0#{N>
z;ABw-Co3|zvRDRJR?6TiVg+19ser4BRd7|M3a%#Bz}1u*xQ5t>Y+N}vl*XkS)Dm0Z
zT1pF?A}Zh%MFIB-+b7^9)Rz(4Wg9P{e&g%H;jic)zHtA&m@!%21NX7iW1`-JKd=}R
z4YGDjw8*+K(IM-{M32mm36E?T69ckwOpM5;F)<;V$Ha_m850Y#bxf?twlT3G3u7W6
z+sDL?>=+XVvU5zF$l{oY$gVMQA-l)KjqDi{53+YmyvV*W@ge)i#E%>plK^sXOoGUv
zF$p1u$0Up#8IuTdbWEbiu`!7u$Hyd&oE(!Ra%xOc$muajBWK1WgDj1Sgq$6dEOKs4
za>)5H$s-rWq<~x;lOl3yOiIY|n8?WGF)1Tg#-xH=9g`|@ZA@y&jWKEL!N(Y5(n40o
zL_xli*sdhv`on{_4$I&Dv(_WOJ$vnfRbM`H;~vs-bNSp2QrFdIZrnpsH`V8EkfyEy
zr)g@yY3o{W+NKtquC4>8YwE!1>w0kdrXCz$=fUw!9-N_W0B2|#z!~dCaK@$)oT+XC
zXKI?jnd@e7=B62(rEURdX<EQp>sD~qrWKs6ZUbj)+Q12Q0i4hjz}f3|aQ3DhoTKgl
z=V&^>IqObv&ZZNbSQo*GO%a@{?gHm(y1=>XZgB3V8=R-^0q1FYz<KLlaNedDoUiT!
z=WF`F`Rjgg{-z&XpdJ7hXa>Lq>p^hAW)NJc9s(C?hQNjEVQ}GQ7+j<t0T*dTz(wm(
zaM5NIT&x}g7i-4A#p`i!@n#%cvYrH&Y$m~_>M3xkW(r)oo(7k0rom<E8E~0q2Aouv
zz)4LBT(+JCmu+Uj<?1<bxn>SrzMcn{Z|1=j>IHCxW&vEWUIbTc7QvP3C2*x?37lM)
z!O2Y-T)AEbS8kTURq7RRm1YH8wO$2RZC1h6>NRk+W({1U-Z*pP9#W&(ICq0u^%l5R
zvjtA6E8vu-a?t(P1JC_3YkxVv_Ht?GoH-xHV~=wtBD?0yh3uX)H?n8WJjmWT^CJ7^
z%!llsGe2@*&H~86ISV3(<}8F9p0hA=WX>YU(K(AE$L1`C9G|l|a%;|74<9@(zV^F7
z&pB=Xal!M?tDhg)|4w25$^Us<`^CrWZ^iX@mX^vJ@5%??<tn=GvD<e%{*DVa$>l<^
zv{Y}Ne$Lrz`lK^Y=Gdmq?@+z-kge}JY}=u+?V-R4+qcC0J9Zxa?tggCu77ljF1N>P
zI6`;g@pknw|G&RI@~HP6ee7}X|G)=7q|xe)CbLy=EZg48TY^U%8HrbE_!M81B>o#c
Cy7uY-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.E716A8264A34B5D8.idx b/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.E716A8264A34B5D8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3bb46e607be03eb1d208f455637e706a565c55c9
GIT binary patch
literal 22954
zcmZvk36x`Hf$!`4?nfJ&mWD=P6e%{vY&ZK&7;>}kA&})}y;%gqHik?P?syuIH7p5%
zph9#oNI)2f5lIgcAs~ndGDRT|7Z8dTHw019hcW_T7+y8+RM&srd3_E!_3Pii>sL+Z
zR&KV+#KPgPmzI_Sr^I$PPgPE4OG`@!tp9h8a`r#{lv`T*^S(<<J8yg}UpnUn`~EM#
zw{qth&)oIZGs4^5w*{|z?%9|A!Sm@OPKte+QT=|0z5AclFI<1>>QjM(?!93D%YU)Y
zL+|c7<OXue(s$JN{BFm62R?I~{yT?1wR4N(=cDJ}`14=?uzG#<M_1f>^keUw{?{`<
zd(X+=IrhERqUQUL&4^e2>n|@zUiHB6)n~4~_MW$1Z$5GGr~Y#KpUyUqw#B~x+#PRU
z%0E=OaP*6d9z69c_uh8ykN)?oi~Frl{_Tlp|Fv+_@qd2ixYnPx9QNBQ{`j?m`;hPa
z^PR68f6bYfKXvm{UrGPd4Z5S1>vul#_D_?anJK$p+fVi8D<2Wh`qrzl2O73Z@@L$p
z_5SS>pBV*Tx%2q<8p;DHjdaY(hkoo6f4ODXT`#}-1&j9LNcHKpV)0A2C6?B%{GqOK
z+-&EYfBwPw(MQ(;k&m8{I{U!ue}BTWUs(O*m3&wI{Z@AWgQEJKKXYCE=nrrC)g@2v
zxcue*2abE?{9CX2FgxSMZ#8#b_5=R2+a}@Te%<aJRea^mW9~of7q{mhx4e~l(;)Fz
z9`f<y54_~i#^+|&EFZVK9n!T^=S1E<R{fs;^pk(|wclhuc=y&Fb6q_iNsjX`O24`#
zHU7kHMcrKo9VXuW(}UU{dSJ`>58g6NM|T{#^XuO^VJkQONpb5LcQo$1{>s^#sk8p=
z&8=51|Kh7(xc=jrtIoLZ!6SpG|Nf>OpNVfzAODfV_H%x7dgrbG^@HEjOE3S~<Xd0<
zuZKD>G-B7BcIuyg`0y3Z>;C!YKYsBA(YXJY|6TpG?~U(1`rwi09re{GZaDP$(B}{T
z(dkb5>@~Mv_3HJ9zN9N&`7>S<efPo4lhMDvb$amV@nv^@CcV(5etiG4FMjEc&{-4Z
zyWiLmv?(|I{k<<3pMPS%`S?qpyfj*U>uK{77jOMu@yY26uUvnVCim`>AHV)*ja@&>
z-}xW+oWA{dLq7Yp;ECUqFJCEbvA2)OzI6E!nse;O-5vkkG3UNv_|PRk4xQ-w^b?=o
zX+HMQr=$P&UngW<{O|ubY4|n!|N8s4A9H^3j+^f7KfmPx_bE@j^|0pRbDw|oyN}*~
z+CFdI5<Kj^f4}#@cVF7^^Y{_x-+oH@8#kQv(ATv`ZGY^nH{L2g^W>ey3oiax=k}Ge
zwcp=eO8?+5?Z17$^icMgY4+&g=Pp<+9D8g3_38QNU3QE3-4A}{k{`K$^7indlYg@K
z`=9@I>HWu^@cY(B_~)jJ*Ka6Z^Z4WG<!hb0U*Gw?miW|(|C0N<?~X@rUA^v-{ccQ#
z&)RP_`j`C<R$rT3+IPp}=U(`m^N%~^n}=M!_U!48e_JR0^5!!l!Nzhn+j#G;Lg=@@
zEgkl3HZ(c%W6wV}+5XI->F$R!2aOx=#1^_;kx!XFc;R&Mws+@G_A`F#y?-2zEVo>G
z!Y_|};NFwo_+BFV)hnL=<%@U7;iYrVZGQUs^<&)g&g-v@Up{xbFHK8JmF-)$f8hO#
zb@TqO(B<hZ*IswVNgrBDt?L~J@A&xp7aRJ#mpqN|m-V;pOS9|xkS&Lt`2NL)ZeQp0
zE>7Bc!nUQ%x;}TGbK~z{Z0Osc%D#Na5jVN^rMY!|<d!2N?_X@_dyabUcdnaOkJ^{!
z*Yzn|PRYN2v7!I;;$bINZoTk<4=p9v^#KPTaOC?J8+z&NQ%^gG%PuW#IsCqU;1<Ub
z!nVqCVmZ0-C2y;!6YAu~m$0p(Noaty2`!K=p##z<^gw)q2QnlKK*od-$doVvnG<Fp
zOTq$VO;~|!2^)}*5P<9nJCGyc0CFarKw?4!awS|q?t~l2lkfm}6J8)+!UyC}_<;h6
z08lUy1PUcWK;c9fD3XW(MH5k=SRw`#PsD+ei6l@ekpfC5(m<I+21rUsK-okVD3{0q
z<r8_JLZSdvOca4ii4u^Ukb%mHGEgN^0jegdK(#~-sF7$SH;#dAl~$q!q$Cue3v3rS
zHjcHei<@6;A1_FsS--k>$nte$^9sK_Aj>xct2;z}9oakt)B~cv8CcUH8uXe0(V*9M
zh!(weK(y#}9il_88xS3OeTV4L>jy-Sp6?JIJwG5kdP9d8&>IHCfZo_4M)bx3F`_qh
zhzY%EKuqY(9b!gr9uPBnONUs{TL#2}-r6Bn^wt5fqPKO34ZUqZZ0Lm!5zq?*BA~Z-
zh#kFsK<wxp9pXUm7!U_~XNNe^I|syxUhEJNy*MBudRK?I(7Oi2h2GsEZuIT}aijNi
zhzGrAKs@NZ9pXjr9S|>iUx)b6`v$~^-rpg9^!@?yqYrdQ0DWLU0_cMs5=0*ykRbX{
zhlJ3F1|)<&+#zA~;Q<Mwk90@`ePlo)=%XDHMIRlIDEe53#L&kEB!)iTA#wEa0g0ne
zc1RL^azK*kQyr2*pBj)9`gDh+(WeI_jXu*M8T6R}$)J}yL_#kOh=e}dAzAd<0m-7z
zbx00<Za{MA^Bs~$pC6Ds`a*{k&=&@zfWFuvMfAl1DWWfRNC|yuKuYN44w2Ez10tg@
zcSsq1c|gkOD;-inUm1`J`f7(%(N_ngioVt%HT1OssiALlNaH%Pc@xtZkjBlxtqy6S
zZw*Kbz0x5HdSyTq^p_IbrNpyw!?dl^Cbsq_A298H^k44!!w*Je{gz5v&cp4Lyj%rp
zljSS$R!dUn;WkTFSAp6@eFffvNgCK{WDRUJZKA=gR?@;&D{En^Z4)hSb&?LYI#~x>
zU7P4|tC#e!)ysO=>f1z*TVCQ}%ga1$`8MHkYmf}EHOL0o8rsBwTcc!ztx-0@*4QRS
z+?pg4Y)!HWwx%{Q;npmfVQZGnur;@d8MhY60$Yo0fvu%YEV#8wR@ho)D{QT8V#TdZ
zvcc9S+hDtXeY4?KkObHYvH)A5O$6N9B|B{GvK_YeHnHQ@Avs{{kR7mfw21?^PRR*d
zr|g8SvrU}16(te2qAbEzY!eZ;F3AO3m+XSAt4&<EbxUs8x@9+P-EHE=tw-{})+2jh
z>sddZaqE@5u=UDb*m~Q<i(8-MgRM{Y!PeI%KHT~xKWzQ7AGZEB@#8ii1z;PH1F#LW
zNdUJ&DG1x39E5GKO@g=$Ng>#V<PdB_Z4$z5SPH{7EQetmZj&%>BT@vm5jg_eNSj1(
z8<nE4jmlBjM%yHc+n5xCZA^~AHr6IF+{UFiY~ykqw(&NJ<2ETJVVjhbuuZl}61OQS
z1>2OIf^DizQn*b^Y1pRaG;GstlE!UD%D^@wXJDIYlMHSpNrJ5;OR$yNM8a)W%EC4)
zXJMOdlPqp?QVzB`IS1QZo8)ktm-4X9%X!%5+a!<Mf>eNQK`y|y&?W`k7NsI=i*gaR
z#WpG8wj`BcTart#EwxDrx3VO|R+eSh%55U!wk(xlTb9ePEw@P-w-u=Z+lpL)?fU6K
z1-DhH3frn&g>AJ>s<^F5HQ3hV8f<HAQp0USYUJUZM{ZPs+N5y>o&iZM*tX;rY+G&8
z!mT1HuvKIQwo02QxV_MJ;odv!+bS2?E&{rQ*e)S^`OtUVy7SKd<G}}Pt0dQ0ayQ#N
z*-5Rj6moivrI9mhEQ2hqF$p=l#<Iw{HI_rpudzIGVT~1#i)*ZiTv}r#WO<Fr$mKOw
zMy{-}3UYOgRgr6JtcI+tF=aQ~_{Z9K;PN%P_5DXIuX*I1oA3O!^N976$v$1amcj|-
zh%VoZJf+LKkQa3M9poOjd^HDue#kA~h}>n%SFug6?o;)(bo1XuJ)-KHk*8F>3wc4+
z?;!U$_0=5w`5~vi5xL9MSHZu#rcX84(~PJFdzvZLU{AB48tiF$oCbTEA*aEfrpq+g
z)Ap$rd)g7zVoy7zTI^{TREs@rkJDmLJLI(3({`B_d%8Z=VNW-rI_&ADREIs?g6goR
z>v1~l>4uyRd%7;uVNc(udhF>(RF6IVl<KjkUr;^v^gT|GJ^hf=V^7~@dhGFi%43fo
zQ678zl=9f)7nH{y-{W}f@k5Ts9^Yj=_6&V$z@A}54cIeGsR4V21vOyL(Bll)GYmNc
z_6%KSz@D*Bjo34es1bX{DK%oxxS&St8GD=&d&VJW#GbLsjMy{vsR?_g5jA1YG^Hl&
znHJQ9JyVY}Vb3(=OxQDZnF)L5J~d;{JfddonWxl@J@bN^v1jgaX6%`VoEdxOE;D1#
z(x(>eSw_@?J<F6@uxD9N3-&BM&VoJ5kh5UV(q$IxS^LzAJ?n^Cv1gr9EB34lYQ>(l
z$62vw9dcIeS-Z@NJzJmJuxA@l8}@8dYQvswL2cNx^*9^$Y(vh5JzJOAuqX7XfIVSE
z1?&k^Dqv4oPyu^Fj}x#b3^@UNLYE2Hv-hbTd-f5vW6wUNcI?>~)Q&xSkF#UXKIH7!
zvv-*tdyYPJV9zn44(vIm)PX(6f;zD0=y4A0Ifk4AdyX!1V9(j7PV6~H)QLUklsd8J
zTu>+WoITEoJ?D^fV$a!SPV9+&Dq>F@Q4xFMl#19B7gWTa*yBX(i9=4rp4ep~_FR4H
z!k%kHUD$I?sSA6q1$AN1)#F^)a}7Bc_FP@&!k)WN-Pm)Fs2h9kDRpDdy`XOFxqF-&
zd+s6U#-6*&+}QK<sRw(W5%pltGo>Evc^1@zJx`DGV9zt;JlOMenFo8`KJ{YHJEC6f
zd8gEiJ@0~gvFGh^UhH{?oELlEF7sl~*QY-0`9{=-J>Qi2u;*J)ANG7b&WAnUkn>^B
z*JVEJ`TNw5J^zUMvFD#sKlc0!>c^hH$N916A98-|`Mb=Iy+EG^uooE70QLe?8o*v)
zK?B$e^tb@_0z)o<y+D@*uovvpAohYI8pK|3N`u%7E@%*Y!5$aHUU0|-u@~&JAofCi
z8p2*^L_^pMO=$>wp#=?LFVy2g*b5E05cWb{7Q$Y*Ps7*?k7yWs;VBJcFT9{(?1g(=
z7<=I%7sg(=%fi@;^l1cpkr9nxFEXVO>_rwdg1tzOi(oG@<RaLMbXf#@(LRl0FFK-8
z>_w+EioNK9MzI&|aZ&6=hg=kU(JqT(FV?3q?8QbjhP~L7#;_M#&=~e&JuZg5*pQ21
zFV<x-?8W;uj=lJZ#<3Tl(m3|w3mV5>yvN0{7awwQ?8Un*j=f}`Cb5?s(IobgQ<}tH
zazT^WOZK=V_L4&`iM?c(C9#+4(-ihnBbvfqYD!bsOD$*$d#N6m!d_~~rLdRkvK02x
zeVWEzdPLLMOHXMUd+7yDV=vv~(%4H6xit3DU6#gPrcX22%Zz9SdzmTCU@x<v8SG_x
zTn2lYA(z2krpq$elloM`o;0Eo_M|D5uqQ33ggvRpN!XKyoP<58%Ovb&`!tKa?1*Nu
zmz~lq_Oc6_#a_0@WwDnXa#`$UyDW>nT%YE!mmAR>_Ht92!(MJdbJ)xExE%I!LoSEC
zT$kmrm+#X&_VOc|$6kI)^VrKTXdZj{9+$^ne#qsqm+!JX_6mJkz+PcQ3)m}6X#snM
z1ubB&(Blf&D-5{;_6l8Az+SOWi`XlUXc2qGDJ^2JxS&Ps6?<F}d&MDF#9pz>ir6dl
zX$gC!5iMb_G^Hi%l@_#wy;6@WVXrjgO4uuPSqXb`pUT*iM^wh1Jf$-B<OP+nC-*oR
zd-9Nzu_t$#jJ<N7ma$hJ(K7bRQ(DGec|ptAEBClE_R2%9jJ<M~m9bap(+c(~BU-^;
zWlAg9t1M^*dzBtn!Cqy^Rj^m-vI_R9eOkp{bwsP!t4?VZd({Q4Vz1ics@SUzxhnRm
zT~@_jtxs#%tBq(4d$lR8VXwBJHSE=TTn&4*Ay>m*t;=fIYxHU3TDtML?6yi{L>o6F
zPibQp@`5(rLGE#lt2y}dL#}Zna+fu(+WWZg`YYl3w1vIah_<lTn$i~bS_|63UaQBo
zu-6)LE$p?rtc5+LPZjJbBdTCenNkIN%7QA`Q+k|&J!Qx#*i*Vp+3fkQ{L|9gZzyXV
zt6WJAypohRUO&MqU2<TTWH<S0qPm(CH#m@4d}8p}-`=C!DhZa*Y`hM3TjjItvwLad
z^}PFRegAd5@SyqKWcguul||!j(s&r33AvlJkd?cMvdM|zOT7C(zw%An`lIivWi>oD
zu2QQvAKBKZG(cLF7D%Vk0qIqGAYR1-8B_)!qsj<mQkj6vDl?EpWdX9PtUxxE4M<Q4
zKz5ZK$f0roIaN*|Q6&PoR4yR5$_?aEd4RksFOW~=1M;i<Kmk<%D5weog;XJ+uqq4`
zQAL2FswhxQ6$6T^;y_7N5-6og0i{)Gpo}U5B&j5ztSSqXQ{{m2syt9ZRRAigia;e*
z2}oASKxI`KsG_O>RaI4>nyLoWP&L$>-+5Z97LcM+_EPr59iROD7sY+nU#oQ92hZES
zam=0co6hN5etE=;d#?8@Z238SR%nH(pTlQuR+t7^yTY`{x)r8F)~_%<GQYxjWWx$G
zARAYh5!tlDOvvUHW=6KGFblGEg;|kpE6j#0tS|xDzQXLtjuqxWcCIq#8vOmmRVE_4
zR+$Ugy~^Cko>k^S_O3E7vTv37ko~L7j~rNK0p#E+3nGVBSqM42%EHKzRTe>xuCgd{
zY?Z~3<Et!=oLpf^{Qr?!VJYPF3QHqrR#*mET454$c7<h;b1N)|oL^yi<iZLoAQx9y
z5xKO&O33mGlab3Stc+Y)VHM=+3acX5R#*+WvBDb9;jQ*6Ypnq*D@@t+pSHvvIPm0$
z?mK$@K$^4Vhw(s~Gxfu8MbgZf23b32T4ddv>5%nvrbp)Ij7K)inE}~2XGUbxoSBf#
zb7n@i%$Wt*I%ig7+nm{ug*g+D?Q>>FcFdUrIXP!Z<kXy{kkfOPM$XJx23eXj2{}7w
zS>)WD<&g7pmPankSpm5?XGP@FoRyH}Ig^pgb5=&K%vlAwI%ien+MLyp8*|op7_T^U
zrfl+O+0qT)Fun8ip6|dDwtO#MaVAWCFJ5sbOoObQFfFof!gR>`3DYC<6UHMOCd`0r
zoG>G@X~Im%<_R++TPDncY@IMGvTed_$ijpP$o2`dBReL{f$W?xC$czUBC>13T*&ST
zb0d2u%!BNmFfX!i!hFd73G*WdCM<v)oUkBrXu?9s;Ry>PM<y(S9G$Qza%{q4$ngn_
zBPS;;iJY3S6mojP(#V+!%OFb=CLw1hEQ_3*upDxJ!t%(42`eBMC#;BEny?bGJYh0&
zdBV!bl?kgLS0}8BT$`{Oa$~|8_rihGny?nKGGWRlfA09;le(fkwSM>EfY1Sjjd#~=
ztLzBvC~Vwq*f_CYf6dLFH~VegefXbuRW6DyYF{!Lb9>H~jyTV{akg~H1@~R{{a^ih
zzW&<>lnyv)Be9{$*7x5XxbaD^`R$EK{-V;Udu|fe*z#^Xve%dzUm2uXV;W@b8q*@{
z)|d`izsB^){2Jqt4QtGRY+PeTWYZcmA)D8j8QHSNEXdY1W<|EGF&nb5#sp;h8nYui
z)|dm?xtlp(gzs$PZYCnTb~6{UdpC0<dv-GqvUfN0BKvkTAF_Wp^CJg#vjB2%Hwz+%
zcC!$2csC0pM|QIaa&$L~BFA>K7;=0!iz7GISYz{*VC#<v>}IVO_g>RC?r*MNN%y>`
zbmRK8?f<#|{LMeM&n^A>#y+>La!CG=^2SwrTjj{)k=c#Uv8{4S?Uee)J%l@VKCn-!
z`StMn6?pr8+YjD&3EExfob~P2)>DJ)ht9F=SoP*T?=#t%duiiytbbK|-pac9RsGL<
z@sDKPuYT(ApFh6-Rb4^Y738Rmqhs6pyFpk7`6|M$+T>sT_G0fB57_m>`q#Fjw4=Up
z#Qx5-&l?}uen@Z6p?J`Fhi`ngKb(BT{k-ip>z;QB^2?^pzx=YWeC#Hvd9`Wte}!65
zAG=8!UIR`eXuxTCEjX>91*hY6;B<lxoSxT%(+hfVJkNvU1s<G%H-Iw;25?5+2+k-N
z!I^jyIFn!kXXee|%z_!5g|~pS2o`Ww-U`ktSi#wN8#tR_11InTI6)A=*?BuSyI=?B
z;2q!`f&-kBcY<>YPH-YGf)fQ1oQrpXa|teRZr%;fEx5sXcn>&_-~s35z2Llp7o3mx
zf%6GIaDLto&M)}E1^58CfDix|<b&XXLJ(Yt4}l8_A#h<n3@$8$!A1B8xQGw|7v-bi
zqCym0jE{kf2{CYSJ`OG}#K9%`B)FuI1efAd;8H>gT$)dVOABdm89oCpBV@oyyaY}X
zByd?i3oa{U!R7cIxSWs!m*?}~@<JY5fiHk72nBFOz6h=;6v37F61b930w?n_I9ZUv
zmH9HbvQP$B;Va-OLIqrvuY#)zRd6-F2CgR5z%}@WY2zFJ`UB%a<Je7V@hxyIp#@Ii
z6>y56fIH82-riS-Z>yYdJAW^2oKTs!zW<XDHg96oGopSN-xfC`8f5K^Xpwa@qC?it
zh#r}r5gyqvBL-ySj2MwkGh#wE&xje>G9wmb>x@{DZ8Ks+7G^|1w$F$i*)byyWao@H
zk;NGikzF(5LUzxH8`(1>9%S!~c#(ZG;zRb&h#xsHBLU>#j0BNGGZI1$&qx?KG9wY>
z=!`^>V>1#%j?YLOIXNRq<kXC$kkd1gM$XJg23eXB2{}6>S>)V|<dE|-l1DDgNCCMx
zBSqxWjFgb&8Ih68Gg3yb%t!^fIwMu&+KkjT`Hfw#-(9}s$C^FoI^uF<^Xz6>S-xnK
z)S^1Fd3K{#)E8}%M$~}QC>n5DQ43D1Xu;`39XOq$1E&}D;Pi?f953?Vc!dXN5DnlA
ziUFKaG=eiKMsOz41kR+Gz?nreIJ05~XAv#nEQ$r3RkVV$Dpqhd(FV?@*uV**08UT@
zaCXrS&aT+OIYb9IhvER|6rJFliW8hDir_>=1m_Z6;9QCeoLh8*b1QCe9?=8Nqj<o1
zMK3t7;sxguec*hG51e20gYzqXZ~-v@E}#U!1;rq^pb`WZ5<}oZN(fw741)_RVQ>*K
z0xqINz(vI<xTq2Z7ZYRPVoD5LT#SQ@D{*j1F$pfIB*CS`6u6X<0+$xk;L=JOTt>`*
z%P1Lek|=?b6bW2b%!11*S#UWq2QH`Nz~#j}xV(}FR}c%}3Q7T7Q7nQhDn)Q5u>`K9
zl)%ZN3{F;LaAmO!uB?>7Rm2Loic$er6|3N?N)=p9tbwa3HE<2F5!pPuX()}0HmN1H
zz_pYXI7L*zDT)H_^R~~!&kwkm*e>4u`2la=BOLg;{*m+c{9LCoS>6RtUaQAMy$jEG
z#zcdx9TP3GZcKE@`Z3WX^JBsz8^*+dY#b9KvT00A$mTIIBU{GAf@~cVE3$1&Y{<fx
z2*~y^u_HUi#DVM_6DP7bCL*$HOkBwBF>xb%#>9i{9TP9IZ%lm1{xR_*2gW3T92}D%
za%fCK$l);wBS*$0f*c)_C~|B}V#x6^i6bY+B#E3FlN54#Ow!1iG07lHV<I7E$0UoK
z8<QMzeoXSng)u207ssTCTpE)SvOFd-a(PV3$dxgvAXmqvid-9$8ggSy8oTht8)MQ!
zR>nj@zKqx|BjUzmL+|dNzxx-hM_)c;{lKa(pSF31wA@@iYm?M<^=X@DNb08gtWDC?
zHQ+Q&4LEIG3r^eAg45M?;B-wLIDK6YPT$mn<Lf*)zR80#)D7SaO#?V%-3ZRuG=ekL
zP2fyT6F76-49?s%gR|5v;4DoGIBVSs&f2tsv(;_jY)u<Dp)P<EngTd`-44#)w1ach
z9pD^I2RLWl3C`Jcf)nc^II$^$bJbnoTum1^cij!n-E@QV)IH!lO%FJ4-3!j!^n&x%
zec*geA2@&A56<88gA3FH-~!D6xL`d9F4zo$3)Ms5Ld_7ka6Jqz+zf+@)Fa>`%?P+?
zJqj+`jDm~RW8h-V7`S*n4ldq|gG<(v;F8TGxKup_F4atdOV`uj(#<ruOg#fG)69UA
z>Jm7qDS^w@v*5DLEVx`f2QJslfy>wP;PTBpxI(=EuFx!iE7ptPip?UpQoRJO)GUFM
z>oPdGDT6E5%izk*GPp{;0<O}mfUDN4;Hu3kxLUmiuGXx9Yt$R3ZJr@DnvJtIsa0=*
zYc*Tol)3^=X)1f&e?R!bZ?g8+^6Nh<z&U5mNAMY<ITMjxbLK*J&zT$9GiM%T@0@v&
zeRJkR_RpCgIWT7d<lvkIkwbG9LJrSa7&$U$5#;EcMUi837DJBDSsb}FXRSx}J|ndL
zvp~;0dC$`_=bT+XC$i^*!k&x&`)mCNA8-6BZhWw`RNnkh-uod}IsJsEPrCh5$CdlY
z<wCKvRBxVg<{9hy_|s0{*gji6p!(px+dj1a_Wfc9gaXGLur=o2amaxm{)dko`;Si1
z<@R_D2kDMG+O9tQ|M$BGANsMwjyUqDkALEm8m-=FGFt`5vhAb1C3w)mk$9DcPxM7e
G;{O1JNlA$S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.B8E2B7EF254C9C8D.idx b/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.B8E2B7EF254C9C8D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2cb73641125c8aec11a0fe90fd0960d0f6c4aa50
GIT binary patch
literal 22898
zcmZvk37BJLfv)TS&rchhhNc^VQKSG-tSl${N!slDa#)g+^<<UEDhEu2jJE+XkUkI=
z0Rv$Sf<g?VVKL1_luZ!IMutVNST@S)h9cv3F(}aJFx7phsBi9b`+3MyFYj0VHgrzq
zWX&XFu`^2x3*l3e+s{6=-Czp~3tLwIIk$bzCs(+Ig+J}Tu(186M+>EMpLZPe+jkzn
z^NiioFP#zF?7KB`-OrwR`JeosIO?S2Cm1yxc*H;cTKmG_)F*!!KIEPY54z;V{T}+q
zU5DL3PFeW&(%rw=dhfxzZ#8`T$RBRs;QZ;$&%XYrzx>COgW3-+9Ut@P+o%8KtWV$F
z`1ZDU{+O`bzb)hG4S#-N>WT+$e0BFV*WCT)YiB=k=*Rzj`XA1*+`K9I$j`q1R+oRM
z`nj85>^ykt7w@_CydV6ZSNGg+dusl~GsEInj{DPYSMv`Wj`;PZzyEU4ci6W-b>}O`
zU3JzaKfL9KUz9$1gZ`NIVEgi0E2%TT)qdfR2WsAU<^7)4H(pIXpx7@ioN=ozF#pJz
zH%DH%^SF1E_5*2c_SnYZAN<IlM|Vws`|200x{KnqA3Z0_*WP;k!gIav>lN3xw!iVG
z?_H4C^;|gqfiu$Q96b2l3D11t$&dE<%a<N$<_<a}Vc7l?@8@@YfAouspW1rKZ?Arj
z>y-<}SG|v&anm=>-mZR+|MaHYW3FHBxbj2tD{mZo{}C^Kt?-!j&GZ|_EZ;lq!^a(b
z@!`s6zID~2>xCUr{f_jx@wc`uy%RdU@q;h_D)Zji#;y11PZr~;N#UjJFGlIfM{bq%
z(?gE%%vKKBaliu`E_iTsqm<Zs^!BfQ`-F|$<j3;HGrq3eJLrAujdbf@-`Ln&eDO<P
z7<@Q$#ToZLcy#3S-+g84nc`;Yxc47%p!<%UcaH!5z2DL9B|o|S8~^h6hc0_wNnUl@
zsekzX!<V|R`_xaLe(8CS>7al4x1}EiU;oao2ai7gLtlF0hQptYe)h;8obIO2T=lgp
zUL73%vR>}}gx4ni@nGhu#9!V#J#x%Mz4J_Ik3RkM{m;B~?boBN+uPr{ZEM8dzTt20
ze8KeW69?Wmx%Q*oMD5KVS)RCP<9Fq!c0T{g;49kvKc4#V;3vvmKPlY#?{}ZR`8Z=C
z_vOfmca$%AytKiw<JjEGmmH-%*WsFpzd822*Nq2U{B-n0?<b!4>~_nx!+(_c*T0{T
zdFem@{iGYe?D*e*`{twW{@1^9&(+Uvc))kc6K_7Oz39AWcYSBq{ip5s#wc>cJO6gi
z!T)%9>rcg_F8JChmD_GO>7lRcKD7DK*6VNHQ9SEZ`NE4nc-hw;KS%e-3nl4$f8H^F
zr1Vhk*qym!BA>bN$>O&0)vxWm;CyxD`ObU4c<~Q>KYr`RhZ;ZL^S3|EyZ3K9;djmV
z^FQ0U=d~N;s~&qyTD->n!fV^V+w}bK#DC6zHTd;i<0r4X_`sV|vDSf4CjNZjp-b1K
z7WUuz*m<A()djA@?l|m{=bkzJv2W_LKfmRSctlyu<&<}(#pthpT{_~KT=e#%KltpU
zw{PBkxODmZGKWl*x08GHcf~($dGF_T?s+S){l^EIzVXh-ZoJt#>YniPqaV2Eq}RWD
zeBw)&KKt{RZYjqW&OPt!PdvMNjQihy?T?e+p0{&<DrFa{n>TEJ&%TRQv+q~<w%=Gj
zx$v92H!WmV^?Cc9C+@pg)2Yj!x%*@PYi!=1=2!L68;*|eyI9keZ!{h)?7!`_{b^xU
zpR(bU!oG_&eL8)5_qC^|4>({Uy{fk!y7j~RF4puTC;atq-!(sU-TpMUst?<6*opft
z*7VC%x131+<;4RQQmcB)p<9mLcd@1y&N=n8bDzvEENnRP-mAkS=TSG4P1Qwy(Y*HC
zu9C304e0pA)H*HkOXjuzx=qz3VQCwXxOBogX?ZOywSpFwT2Tv29j}9>PSC+pC+c9S
z=k>7E3wl`UMLjGHyaAR5!2nBxXn-Zp^RVOv9+tew!_vqbVQCbMur!KBSekefEKPz5
zmL|~zOEYhVrCBh;(kz-`Y2hufv<MbhT0{#ht-KYMR>2BOt7wI#jkm$lCfHzU6K$}x
z^LAL;1v@P5q8*k3FThd|1Xv29080n&fTcrlz|tW)VCm$YuyhJeSUN=~EZw{tmTtif
zOSkBTrHA*x(j$0a=@C7!^zvR<dIc{my`mSEKHdjQpWuU~PxQgk&--EN7yPjFi+)%J
z_y8;eLI9QlF#yXTAB1I42*NTb24NZEL$C}9Ay|gQ5G=!d7?xon49l<>hGm40z%n94
zU>Ol3u#ECiSVn~?ETdu+mN7mC%a{;@WlW60GS0_g85iQPjEiwtCiny_6G8%(2{8f7
zB%g$3Qb@uwDJEel@**rnL4>6!im*)aDOje26f9F>3YKX;4a>BUhGkkz!&2fUSW1Ef
zOG%Vqnc*|A%m^7+X2c9EvwRkoSs@F{teAymj?cj|C*)w66LYZ4^Lbe2g*+_tVjh+S
zz5vUDP=IAYEWonJ7hzcxim)t-MOezb3`<#%VJV9;EK7U|mL;JC%aT}vWtlI-vMiKg
zSr*H%tnd|BR)h*HD`EwfRlW+#s!)YxRjk6Y#@Aq36Kb%mi8WZ(`8q7?LLHWMu?|ax
zSIldtC!4B@plky=UP-M}lW)SZDKuf(JiZCbHs6M2TWG_weS901=iASBzI$p_y}*6}
z(1rF3-=&RvZh2|PaYFXY)svOW$l`Tm{RCz4da`&6@RCZFt|9BEB};v>bQ7>vC0g{_
zKGCArsYHif*C#skdX?zW>-$8H-k=f#dPAQW(DN$c(er)6qc^I=h~C&IM)W3?n9!U0
z#Dw0g5;J;ppP12GRANDI=@Scjt4gfst$kueZ&Qg4y{%7d=<O=8qqp~o9lf9u0lm;C
z0(yr^9OxZ=;y~|Ii4(oEPn_u8DsiKC_lX<5M<pKgo<8xQ_o~E;-rFZ$^gfmN(EIwt
zhu*IeKYD+k_|XSc5<nm5lK}dlN`mNveG)_;Qb`DXs82%Z!zu}*5BEtJeMBV@^pQS^
zppU8~iay#WQS>pD#L&n3B!)h&k~sQ!pTyB8RFXiS=#vEcq)L+LlYNpzFRDaDFZPLu
zKBbZr`c$8!(5F?BMxXAJG<r!T5_+jmB=i}TWYA~&B!fPyk}UdcpJdVJRFXrV>ysS%
zyh`%u^L>&>Ur<Q_eW6bZ=!+^TqA&JI5xuMu8NJ*mGWwEAO6W^{QbJ!=Nf~{)Ps->k
zDyg8a^hpJMRV7vQ)jp}Buc@SlzSbu-^mUcg(bxN=j$XNpDA$qon=R#fqTB-9yo@x_
zH?Jp6^zF+?8-4qF(njAU_Ac?S-)QY1_8sfIW#{D|_~+?=d~e^K?CfFzZl}tNHJ}}2
z@lw3il9qCCn<Xz*fI3TEyaAK6FxARhm})y(oa!VUOm(sjrn-&}r+P^bQ@yN*slKDf
zsX;Qp)F2yRYUmhn%1b;<d6|bP-{Em;l#DPn%0`$PJ4T$EBoj<cvI(Z9jtQq`$qZAo
zY=)`1W5%gPvcS|LTVQJGSa52UtT468R+w5lR-D=-8%%Ao4W_n^4X1X=4pX~qhpD}@
zdaYZ(N(hnwQ$ZGBDs%*#IwS{79kK(aj*bJTPRR*Vr|g8Ov*X06TXMtHExTdr?znO4
zkvuT<$R3z_Iv$*QB`-|9vKOY_ju)ps$p=%P?1QPV<HM<6^25|G`(f(u_;DJL0x%88
z0hk6l0h|V<AWVaD5T?OS5T_w21k;clf@!D|!f99v!!#_1VH)m)aT<{#FpbC&m_|Ag
zoJOT6OrvrXrqNCmr!gr8)0iBCX{-~&X<UlKG%m+s8t=q$nvfDOO~?tDCOQe6CZ!}y
zlX4QK$xafdq9nppltq||9TBG~DFxG%oPueplfr3QO2afQr(v4zq;V=q5=<pof~nMz
zaGH@aFwMvrm}WW|oMxpgOtW$prrAywr#UGH)0~`xX|9vQX<o|1G%x32n(yRsT967b
zEyx9!7CHr-7NsIgi*gaB#ZD2YvLwS)mSvdA9T}%3sRYxKT!Lw-Q^ILkD#Nramtk7&
zlyO>-Dlo0c6_{2!6`WS3Dom?#6{gir6{j_+2Gg2cgK4c(!)aZr!?Z5fVOsChajImM
z0-VE?l^W0vqFjn+G}$I>o8=~Kn>$Dox9w~jw(W8ow(T9HjoZ)JKlkn(@J-c@y#sVH
zv0qHy<-_0h>CZps_lIuTR86h0)NZzZOO;+>X=G`INywQMmO;+0uq<+Jh2@a*D=d#(
zSYZX^;tDGw%PUMqF0HT<a(RW7kt-{#f?Qo;Rpi<Vt0C7{SRJ|j9BaS8*8Z~Ao>lrI
z-MH`Rq^p+SzU9tex{q3Y7~^WXcnyUU!kg*hEyz3R;$6sl=;GVRS8|J==ivKq<Q8v2
zzML&y!Pfm!k1kzC*Z*FYhIHu$<OyB64S7zNUPo5BrAs*Y{sFgiJ#v>V?SQ{`ZI5cP
zryWu)_OuhK#h!Lfwb;|DoECf90jI^Dw#&5G)Agthd%7XjVNW-qI_&A@REIsC%IUDD
z8*n=8>AFmZJ$;Yrv8NwWJ@)hys>hywPW9N+tDGKt`T?iMp1#ZU*faE~0eglaHDJ#$
zp$6<3=G1^agUT7OXBcn>>>0YufIYrPdF=5+%43h8P#$~yobuS?RgT9VKj3)m@m<Da
z&)B0z>=}pDh&|(k8nI`bQzQ0_Drdx=aljd|XY4W~_Dnr$!k%eJP1rL{s0n+fIW=L=
zq;e+gnFgE*d!{ZkVb9#7X6%`V)QmmzgqpEuo>MdS%qnNbo_WBTv1jfwGxjV!YQdgm
zNG;g2OsEBWmN~Ux&!Tb`>{$k!1$&k*vtZBKqgL!$ht!Ha>x5ddXPr|k_N*#r#h!J*
zS+QsBGAs6MJ!-?AZAfj{vrVWCd$u{XVb7*=Htg93oDF-nF0*0J-lKNx*@x7QJ^O^(
zv1gxCJNE1<XUCp>z}c~9?=m~~gdP>JCk&~8Jz+ux><M!!U{6pv0eixL6R;<AnSeb<
zk2<jD7*YrJ924rmo?}iO*mJ0y1AC4E=fIw$%N*Eq_NWtk&LMST&pDw^>^bMui9M&v
zIkD#)a8B$wyUd9_caOTU=N?iw_S_Td#-4jl-Pm)hoEv-Y0q4e^yUX0z^Yo|(d!8Zn
zV9zt59_)GM)Pp^b%6YKo8E_u#dAiJlJ#UYCvF9C9FZR3>>cyUSPQBRks+<>l-T~*u
zp0~@q*z@(M4|~2L^<mF9p+4;S=G2EhpUU~L=NoW7?D@LPhdqCf`myI9Qa|?m6Y9sF
ze@^|_^Q)X6d;S6E$DY5-{MZZhXaIYGAq`+JFrfkL1?Dt>y@1LEuooC`0qh03EP%aW
zj|Q<99MT~6f)g6VUT{u>*bAy$5PQJ^7sOt$%YxVo^=Jrtp&<=nFEpVc?1knuguRf;
zg|HVIa3So4x-5jfaF2$u7ar0u_QDey#$I?%!`KU}To`-d0T;$zxXZ%Wi}Yv&dyyfH
zU@tPE5$r|gG=jZ|%0;jj8E_HoMY=44y=ad{u@@cEDE6Wg8pU38PNUe1s$3L%(E%65
zUbM@i*o*aO412L5jbSe~p)u^m<}`-An99Yl7aMRf?8Uk)hP`-?#<3S4(m3|w6B@@}
zd`{!oi>q85d+`Am$6mb4;@C^{XaakQAx&T}F`)_UCFV4Ny@bjou$LHc3G5}hEP=gb
zk0!B~9MUBAk`tQ5UUE*8*h{Kh5_`!3m&9JO%aYg=dsM`pIHV%>#0eF#C(fyeJyGRE
z?1=+T#Gcq?BKA@}n!;XcNK@ELO=t>xsX0wyFQsxR?4<@=3VW$8OJOhFqiO7=hcu17
z^n|9dm!8u!_R=bs#$I~BrLmXpvNZOj9+j{s4XK1ZX+kCJNpmV;Pf|Gvd(wcDuqSny
zguP6UX0VqT(hT-86Pm$ZW==EM%cxuidzk^3!Ct1zGT6)ZXcl|fA<behJE2+ZW#=@D
zy{yV*v6mfiS?p!IEQ`HdkLIwK8`2#1aub@vUT#iv*vqM04tu!)m&0DJ%W~Mu_h=q_
z`610?FF&Dq?B(Y)kG;Ig<*}C^aCz+IyDX2rLXQ@(R~XU)_6iePz+Pca3)m~DTmgH9
z0aw6Yq00)`EB0s+d&MCwVy`%%MeG&lw1~Z;$`!F!9B@VK6}zm6J-J6^?8!qaV^5w?
z8GG`a%Gi@tPR5=*;AHH{T_$6%)T1Trm4>v0z0!o1uvePX681_eSHfOtz?HC9>ar5{
z$~{`fUU^8%*eg$H8GGe9En}~&a%Jq52V5C@<t{5@uhOFx>{W)eg1yRwR<Ku@(+c(~
zDp$c?Wx!RiSLw0}_NqNv#a?wttJtefXcc?aIjv%^s&ZBARR>%Zd(|$hVz1VtHSE=f
zw1&Ofgx0WEo6{QhYARR5UTwhDuvhD{8uscvTE|{}NbA_EPiP%`^*OC$udZ@+?9~Tc
z9eedIt7A{Onkv`OwZ~;wU+H!;Rc=AvNtL^h_fX|+<SRMl^BjEtjhu25^5smq;@!u6
zH&vTg(<b(sH`6BenmcI|d(AzxiM{5PToZfE8@VR-nwPUC_S#p|Hul;#(>C_nJ82tx
z?LD-Oz4nz{8++{=xi<FNm$UYIFWCFT!dtJmpW|4yM-J|h%G&E8SoLyp@Z}`8&Yvfm
z&l7o#!<juFx&G1lJ9Ja^cy_#Y?M(`ss-I?`ewWt9WaGX!S*$&1e*BFu^S-}+@ecdy
zqvx8%C3tLGv$VAS$h20Y1=4ABKzfZH$e=L*@fse;s4)VWG$tUk#tdZ9Sb(e=E09fN
z1F~!EK!Qd9a%dbtPK^`Dt#Jc+G#((Y#tY=r_<;NxKTtpu019e?Kp{;CD69zsMKlqh
zs3r;&)5L({nmABGlK@I;l0c$H1WIXAKxs`HNYY3^8BGQ#tH}c8G&!KWCJ$856o87F
zB9N?+fl8VZP+3z3s%R=eRZSJBrl|qdHFY3Gqb#le<Y;P|Ky6L?U0S;iuRd9{_bo(g
z*ZWQX&oPxdaqCBa*Z1tV`aYiX-+TV%wWDRNu=?G5-(j@=-T(6lJn2`Loxb7cN4>Q7
zD0-YN{tTZZT4qa6<8wI6OpB~rW;$g3GSed)mYD&WUuHbAahVyBP0P%LY+hz&WXm$M
zAX}H271_4TY{>RyW=9s5nSktAW)5WMGIJukcQN<s)0hAKUan15&o1Uc_U>X{WZy34
zL-y}te&oO|7C;W}VnO84E*3%#?_y!($SxK^j_zVn<k&72Lyqraapc4<mOxJKVo79i
z7ZZ_F%PfWeKhn!AjVvuQ2|2UOGRWCwmPO7jvmA1MndOlS%dCK0TxLaNd6~(`rDaw^
zE-$k(a%GuSkgLnAid<V}HRSp-t0OCqGv#M^8~!+JBDWuB?REZyHTl57jfd_%X7xb2
zk1alo2hxl!-3M1B?Tl%Wbu*?z*3Xz8*)U@UWPZkYWaEq(kxesZLN?Es8QC&p7G&#;
zS&?lsW<$2mm>pS|F#*{zV-95Jj5(20GnPV5&sZ8+nlTAEGh-R#?2KiRb2FAh&d*pL
zxiDh|<l>AKk>weakxMgHLN3o(8M!iJ73AuSRgr5mRzt4OSRGlpk0}qs6{&q6Yp?U#
z+dcQ&xbJPA>p!R$XJqkVcu}Y_Bg(`0LXR0~BDZIxz0RLz3pd<me)~szf3V-q7Vp99
z<Crboh1bV1(<1A}Ooyx=Gd;3l%nZo<nDNNQF*72Y#>|9l9y2qtWy~zd)-kgp+s4d>
zY#%c_vM^=>vSZ8~$j&iyBD=@TjqDjS53+a6yvV*W^CA1k%#R!xvjB2%%!0_FF$*Dw
z$1IE-8M6p-bj+g2u`!Dw$Hy#=oEWnNa&pX)$l{oZ$f+?)A*aVIjVz6sgq#_(403kN
zvdFnH%OU5-ERS3mvjTE)%!<hJn90bcF)JaL$E=K88M6v<b<C>BwK1z9*T<}mtlZ9&
zd*FC%-p-oH?b}&<oj-G2<SD)ENUz>U*%IASTzh}@rs~${*5ca1+ez%HUu!ny_nw3u
zb-r!wB<$h~?^Pf9#V_w${nuMcTTWX0-m89nQ;P4DPThN3@f=%x0gufUwuCPr(ylNq
zvTlXxko7A}k8D_B24sGP@yNy%W<)lvFcY$Qg_)5pE6jpyU13&a+X}NG+gF$!Sy*8L
zvSWogkew^ciR|9Z+&_h%O`hG%gY4bSyvV-Y%!lmX&HTuL-7J6{+|7c>q1`Nm9Nx{s
z$dTPFf*jq=qR6q`EQTE4&Em+3-7JBe+|826;%+7)E6*`y{RLyI4;(zln(KVW?{}P6
z`sGbmbDOG%6%MPc-{(C#bu_$|@Ra%~C$F9V-?{yP{nBTDHMV*c+<f5XL)Tt{Ha)lX
z&F02aBdgcyZEV}p`aSQn*jew=+V8)5A7bz446NTY{m*;(%Q@exA3yS^kFEZxE+y<z
z@}ae>|EB6?gk6Sw1z}gL^DlmVQJ}x&uIE?(Y+FlPPhLBi)(#WP#(f`Mu%2C-lBN6L
zb?@3K(IV@nM2D=O5<RkEN({*Sl<>&LDKR3Oro@D7o)R;%WlAi_)+w<f+or^ZY@ZT4
zvM?nAvSUgd$j&KoBD<%=jqI5c53+YkyvV*O@ge)C#E%@9k^pjWN`lCtDG4EmrzDIV
znUV-{bV{Pgu_=im$EPHYoS2dXa&k(N$l{cU$f+qwA*ZJ#jVw)xgq)d@403i#vdFn9
z$sy;bB#&H}k^*vZN{Yzxl*q`XDJdbBr=*NrnUV@}bxNwpwJE6~*Qcbu&c8A5_xgJ_
zA9m&5+n7VnKXUCLd;PB0W|fPd*6w{XtY<O4cFe9)d$F@lORlBp`hD4@*3t#*q;+Y*
zX<J%wI+qTduB8K~cj>|DTY7K?mjRriWdO&!cyN4+2WNB{!5Ldda3+@toT+63XLgyv
znOkOX7MBH_rDXwUby>k#TUKy3mkpe)Wdmn-*}>Ucc5s4A04KBra1NIPoTKFc=X5#2
zIa^L}ZkHRJyX6Mwae2UbS{`s-mlvG3<pt++`M~*FK5%}QADqAC2N!S!zy(?Xa6wlP
zT(A`c7jlKbg<2tSVOJPjxD^H$aYevIS`lzjR}@^d6$KY_#lXc{F>rBL99+B=2bXXq
zz$ID<a7kAZT(Xq}C%Qy%VoL;<a;3nfS}Aa8R~lTpl?ErdBydto0+(@Rz-3w)a9LLt
zT(*@3mviO7<ytv#c~>4>zLf`8a23E6S_N=LR}oyXRRkxyWN>mz23K;Gz?E7haAj8+
zT)9;SS8-LqRazBrRaX^UwN(XIbJf7rS~YNWR~=lvRR^bdl=#{?)#~#_ZKbnLO-~bC
zv)u&O_O!vZ+ih^4wSN}gI(iYYU$p+#(YNjv4t~wBe7>;ygM303?}7&cmPTahHara)
z5iPQAM0Cjd5z!+XM#O;3j|h)!91$b3X+%uO<`FR?TSml!Y#k9RvTa0c$o3JjBMT!U
zAUj6Hf$SU+C$f7)+{m60@gRFg#Ea}35g)RDMEuBs5eXm%M<j?G8j%okctpa;kr9a?
zM@J-z92=1sa(qPM$cYh2ASXv8i7bwYh@2Xc6moh*(#X<?NXVHH$slJ(B#WFIksNY<
zMDoam5h)-SN2G`>kBE$18j%umc|^*{l@X~RS4X6ZTpN)Za(zVV$jXE$ci{tJ6VgO(
zPe>bi^&==&KjmTVeCZzt6=p9scm4K^)dTC~#naZ$kQUEgY^~E$W9gLjGo&SD=~L^Z
zZD_%16)iYjLkCW$=)mb4dT@G056;jqfHNotaD0OY$16NIW5Wo}s2IVS8YXZi#RSgW
zFoQEIW^k5<1)N2(fU`EN;H-)joULI4XH#t8><v3OyJ80?Gz4&hB7k!=9N-*^1Dvzr
z1m{$o;M@&2IJe>k=V^Grc@z&gZ^H}Dt9Zfr8a{A7#Rtyc@PqR!esF<C09-%`fD1N)
z;DSmJT&NKO7g9pt!i_Mvuo4CrX+*$9lnA(JBML66M8U-xF>o;@1}@%+gNrM1aEV3&
zTtZ2JOE!|=l1dVs*bu>qiU=;%NP$Z!DRAjV8eCdQgOeH(I7yMfWf~c986^WQ+sJ~;
zDp_#3Mh;v~$$`r^^5F7H9$cYO09Q~7;EIhRxS~=7CpTnpvLb^kHA>)0N(o%KQ3h95
z%HS%E3b=|=0atBQ!Bv$ixLTtIuBOz$)f;tib)^nYIaxVv{R~MtTWPIR^W-MD=Gjef
z?UUQ!+Gn@lb^rO`^S{bD{#aOjTd8}-+z+tzAIP2=^B{X?%!}-sF(0yj#{9^E84Dl>
zXDo;unz0aac*erWkr|61M`tXG9GkHia(u?($cY(CASY)mi7d{Th}^u7HCLbA{O`{Z
zUHweYvl@FJ!8`Yylh2Ls{i3+{;{X0y{Q%>&U-H@)3k#L?FU#+K$yK+E%KLk-{n$<W
zl`BQLv~cp-r<`@hsy^<t6F9cthWBXRyZ@#G4%&QRa!WLP?3Rtm(AL8ae&0X6f7{30
z9<R?IFdm|J9phL!^8fd54?X;YM;vwZhd%s~k7{)Wli6YuoQw7k@Ycv7hsMPkjhz@w
HkmLUgOWmQe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.4104F408858A9B41.idx b/.cache/clangd/index/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.4104F408858A9B41.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fdd230bbd1ae39b95fc2d31ac73fecb2643f2a4a
GIT binary patch
literal 19332
zcmZvj37k{)fyd|ne}8Gj5d^BNwT$4^TrYPtdASdcyqqrw1R1v`Xk>k*MxZW8H4>i2
zqaLYo8;e%cxHTRPH1$TUj$>W%SnW6-c(v7fwAHSSTNUfF1N#a2?S7WehkWMC_cx#K
z43oS!xzdqk%f25S83~;g*}VSjO$9bGGBR)G@5W6V{%MLE8Tt5#k&(?eJ-KrA#@A0h
z>b?C>-@R_S`{Q-X=6Y`nU;paM@BV|YxcJOSkx~7eMPL2r8P_$=e&+em(f40;)TM78
z_UKo;7T!qC8o5t>&mZPLuweQ&{e8zjzj@YauiWz5ryu|IkIyvDc<!?HiBEoc&R@TC
z{yl~J9Q!|7VSdPwcGsG3UzEIZ_l^&zuf6u3&p%ne=a_H*<(xlnFyAs7dHmJezo_z$
zj$eGso0Uh-{?YxnZF%lLKHT?^_1XTOmz%4868(5uIPcH17QJ)XpKe~|U3lMj?|wge
z^>;3P{?_MzB>mHkx)V1wHcxypl|1*}O>cZONA=nJ$GXcu|1h#!wryLv?l!HzfAYDv
zgx|kAx?kS3JEh5-QaJ9LC;z3jtNY%E*IBfeEI;Fgz4`nPZ;OrWt^Haj3-@mR?Bgf4
zu9(~#TK<i7sSOJnAH-k2?wN1Z_$$<ppO-!A=oR|SFL{12`Rmqiw>>-m()X_Un(+SC
z_SIi!>u&n_`pwD{{Q0AIE)#yY{i+l4?|*j6LyO-0*~+IZpQk=EWcb>`lcEc@9VcIK
z@6}_%8{3!awx>2O|H7f(A3Ue<+|9pFAKsfi|3Teaae1<{@~zBoTdB^;x8-%+qZhe*
zQ%7$<a`&vQkF<73E9NiR{L}m5v$@W1^0U|7E<ezy-TPUp{I8$Qu8qC<<LeqHrLSD~
zz#~h-=X~&!`R9srrRcGX=D6<Kd3XEoPkcbDm%eo8&;RA`k6!+|9J#u5_Md<K*k!Kk
zzx&FIZ@um|9`!H3P`?oP^x?@zmRxwkkN4bo+-pm}fBbXjxaiAQ|Ln>S8^^t?%hz7w
zH7mY)B>n7)zkYsB_{5HK_qozOUFyY$UViI`w=XT<x#{6M=7(*YZv5N+>x{4Mne$-h
zhu^BMIOFpd%zG}G{j2=5J70gl@e@t%t7lJYyd>{>Y31GjcF#F;qlT5)o5P8_PQUc&
z)w53BeoFS;OBZW4o+|XjKb*4VQ^S$lUR;{+6!(08v)OUn3oHKh@A33o|MhQY?zs8X
zfB)MrpLEr4|H=K=yf$mM_pCjiKc=~4%WIPlPd-#S?6X#Q(f(iDzu>EP=f5H@-ukn%
z*4%O9nUDTdd&1l&%b$LJ*Q#gl&R=xNH!uI$(;Kvpzp+|+;xF6#kFS0-d&<u2iQx+_
zdS;cQea$C3w_d2U+z%iAqixT5fAhtTM+?8%_qUJx)rTDM56(N5e|6`+Pj1X#{nS&^
z*tM=VKH2=M^W4uTzL)!H;P%P(GuLmMb5n9zdCoH{{xatn^|i^7Bj!J~<>KFO6&BvL
z@Y206pYzl&b(yzsUAH_ek7cv+{_d)!@4T~m(aYJTcP{znYfs)ecltQ#im#`S?#N$8
z_UU#l|F-$?i+Aq(!oT@9bBsUV@7!^VrB#i;y=3?OXMXyt*oq%t_S)NT-Fy17k&Ro{
z7hjvXjr+d*<fG1eTXr5nrOe3q+*xzKcJN}x9Q<?c8t;PU$$ttTL9;V@;jD#;gBJsR
z^Q%X*!fD&)kB+2gbosEdc<^GNpZ||<%YT!9@#`aKZbmPewPg9hi-CUO*?m{OaRt5p
z2)c4cubH)G<-v=Ap5{ZQJrDmuI&vg6qvszp|D=N#1O0(%&%&Ll(y2#|Bxm%zW9BV6
zcrnl;8_q6md~4~*$gJZZxF*y(ZE=2de2gD64Zfopij5_Qq~_J8!Ph%Ft`^ntA!&FG
zIE|<Qr{%Taw4xTAj@N<Hi8^q4UJp($>cR0m4~`dka0cE0&LA4V8F?c(qi6(Y;!WU8
zq6wUtH-j^aW^fkX0?r~@z*%`KIICy{XX9<)Y@!XEgLi;)hz@X0-U-erI>8CN08S7E
za4y~j&Lz6Qxp_A@x9A4v;XU9yq6eIp_k#0^UT{9%2hJz@!1;MUIKSuz7vKZn0%8DM
zkPm_jia~IpnOno~8W|Ep;KF<uTv!Z)i|`R}5itTz<VA3zD1wXfQE*W)3NFUSz{SKE
zxHumN7Z>B;5_|$&LQH^5@=0(>F$pfkr@*De6gY{Oz)7M6F3qRGrNuP344(m)5i{Vj
zd=^|*%!14DIdC~K2d>B$O@o8l=y)+!Ob)5Um%x=`C2%q?gOg)2xbto2pLXalJHE-b
z=@8A{ck5f*qxQ^!1KZ_f?0Pahn2p^)#%={xD@1)Q8D4$Vb)vorSfdaPdQF{Z&}$W<
zMX#+BEqa|obm(<;qC>A&h#tMZPW0$`h4AS4I^oe96k<Scs1pNvqe6`6jdfx~Z&HW}
zy{S%2=*<c-qc_)y8NEdz7W9@nv7omq#ERZpCsy<}h1k&B>cocLp%4dpN1Zs(I~C$Y
z@2nFidO;xqdZA7P^e%<C(7Woyh2E_YH+pxSxY2tQ;z94J6AyZ?LcHj`b>c<uQ-}|}
zuTFgE{R;7;_t%LZeLx`r^np4Fpbsh}h(1^+LG&SogwTiTB!oV!kTCjiorKXx6cRxn
zsgnqLQ6VCFu}(zvQH4a&N9!btKBkZu`dFRB(8m=LM<1_~IQoP_66h0kl0ctSND_Ur
zPLk+T3Q3_))kzAyq!0<cR3{Spv_jJ8({++YpHWB#eWp$_=(7sRqR-Yz7JW`3IrO<Y
z$)PV^PKwu);lZ|e11a7LT)Lc;(3fr?CG_&;L`E;)KxFh=Y+HQ8gYbp63lGt}omYJ0
zd)+@ip^%xQaAs^J9D`3EI|FDt8M_RR$dWn>$K$+u4Nyg0#e=h?fvG01fvKjV!Kqf#
z!c?2r!c<$);#4Q;V5-aOV5+O=aH^N|FxBVvFx6M|IOQcCrhJ}<DPQ4nYLE;tHRKI2
zHB<~ZHA+U98uLb&8Y@Penj{lUO?eYcO%)SP&5{|W=DZoE=874o7RdrrOWp!gOT~gy
zt7L_#HE)HfwPMAoO|rq%mbbywR<YsKAvs{`$U9)_s5o%yl$<bi=AAHgR-8B$Bmt&E
zUVy1k5pe2~TrhRzT`+Z3TsU=0ZkW3BZkW0&Zk&1~4@^CI4@^B34^F+37pC647pC5d
z7pFeS2UB0(2UB0ghf}}ghp9jBhpE5f$7w(cz%-B#z%)<^;4~-&VH(T_VH&IiaT=0B
zFb(BHFb!2gI1Ni-n1=IVn1(B1oJOPwOe6USOe2*DPDM$CshAgGDpo|CMx`iBqxmRI
zqm?L5V^R#Jv3v}su}Tc5aVZYdcs>r(cqNY0gp`13BA<Y1qLRRAQcA)!nNPwrSxMqF
zC8c1R%BNtOs-$o#NfJz@yaZFJBH=VGrD2-Rr(v3|q;Z;&GBC~LGce6mGC0jjS(s+?
zS(s)kS)Asi987ch987bS98Qaw;z~Gqo?biyXgeuhhNsd@3AUxvORz0%Cnem<85y?n
z=`w8P?L@}y_iW!gbkAmV`~uqrK$~rw57BX7dUY3W_|q|SM#m#lEHcf8=S^{niOA6@
z7DbLtu^4iEip7x=Q!Ig;oMK7j)D%l0OH)ijPEWBka%PHUkh4=Pi=3NcIb?Y+liy&2
z|N6lrb6q<7;6rj(Pkedn-M@1!p7}lF8aj3@h2Jf1p<}lq@1$e9koVECFOjd}#(u!T
z=kMUgZbH6-ja|uxUR|T=%jxj{S>2@S8<9IyeFt)%sy{_mIQ69*e13yd-+)|Y>h17z
z)YPa3dzvQIU{BMb8tiHMRD(T@!fCLlX>c0sX{t<vJ#CF@v8QcPE%vk>s>PnRPqo<7
zDx4O3+6Je^p0>)g*wfXh4tu&L)nQN9p*rm8`c#KKox<s`r)zLJ?CGjZhdq6b>anM9
zQa$$c9jeEkzEAbo(<_`Fd-?{a$DY2*^w{HTl*b<5q&)Wc4&|}O_bHD(Ug3D`@ePj0
z9$#fV_6#*@z@DK=4cIeur~!M1J~d#^pl}B485*1cdxk1AV9!{iM(i1z)QCM}hZ?bG
z>{BE5j0$JOp0U9hv1hC@Blb)+YQmnWNln-@b*Kq@ram=c&!liB?3o&z345k0GhxqM
zqh{=xo79XwbBCI-XYNxo_RI=r#-6#snXzZCGBfrpHEO}0rAaN=vvjBhdzL=6V9%m(
z7VKFXoCSN9DzjkETBBC%S)0^~J!^+rv1jd5EB34kXT_ei!CA3qtuib2Y&B}bo~=o3
z*t2!04STjewPDYua5n7O8k`M#wkoq>&rzce>^Yj$fjviuI<V*HQwR1O3g^I{qro|_
z=cqCV_MA29#GbQBo!E1Bs1tk6K6PTxsc=s0IUAf4d(J9zVo#`10eeD|3fL1mRKT9l
zrvml_g%hwRG&li!LX`>FbJeH|d#)yRVb9f}F6_Db)P+5l!nv^LYH%*>xvI>CJ$H?|
zvFC15H}>2e>c*bCPu<vaE1Vm9?gr<^p1aE2*z?q=2Ya3-^<dA_p&smc`qYCxkHUGd
z=V@>r?0KrpgFSDJda>tiQZM$r9qPrNw@<y;^D3Mdd)@};#h$myyx8;As1JL-CiP*@
z*P%Y_`TEp{J)gq)u;*)VKJ59b%!fUHjry_YZ&E+@{2l7Yp1)81*z+r#AA9}==f|GE
z%KX?1)Mx;EfhG-LFVLX@>;?KXfW3gi1+W)rZ~^QEsw{xLV2uW`7i`iX_JSQ6#9pvZ
zgV+lyTo8M~1{cI$u*!ni3)N@{d!Z%`VK3C7A?$_vG=#m7!iBIGYH%U!g{mxsy>N|&
zu@`RAF!sV78pd9@Ps7*?D_j_R;RYARUbxD_*o)L?1bdMtjbJa*p%Lsw`ZR*Qh{8p%
z7in-2>_w_9f<3WDMeK=9Dq>IUP!W4#pNiNM6;8yS*x*F$iB%?IFIuBf>_wY2ioIxu
zMzI&|(<t_$3Kzv*w82HO7p<}=_F^>}!(OaOW7vyzXbgL?K8;~7rf@Or#Tr};d$B5u
zVJ}{zaqPvLG>*M^hsLoN@6$N;;tChXUcAA@u@|qhIQ9}Xn!sM7NfX#hbZ7#5i9StW
zFQIS=>?Imp0(*%nOJFZqqe<)~n>2~NWQQiPm+aFd_L2&h#9p$&C9#*RvLyCWHJZX+
zs!3DWOLb@pd#OH6VK1d{DeR>hTnc-sDobHcs!<7hQj<#9lR8wwp46uj_9TUquqQP*
z342nNN!Ux*Xc~LzCQV~6-JxmhrTa9Ey|lumv6pUeY3!w|ERDTPjb^ZyY0?b#G98-1
zUZzho*vlwf278$Xm%(19$}-r?)@T-c*(S|mFWaG6>}C5ji@mJEWwDoSa9QkSt1OGX
zT#e?imuu1-_HrGX!(OgWbJ)u%Tn>A=2A9KLuF7)QD_%p3*V4h`G&4{1-$IMGBJZTd
zUC8@r@k`{ZxZ)2u`1~DQ@h0ReSn<k3j{}X4m#(2D?3HezCG3@U(h~Mc`)CP!rK`9S
z_DVar681`0uoCv<Yp9Gp`4%c;Pu@vo?8*D6j6L}(PR5?RgOjl*U%}*IFHrmQ$QPe(
z+RL%=ix*tHX7Dr{8^44sxP)Yfyq&1Fll*`~>3t{P@MM2K9UWiBmT3kr`x+fz$JQO9
z*p7F3?|=U2F5S$7W~wnYJfx*ktA`IrX;c~@tx5}|Q|W;8Dm@Ud;(-h*1CUW=1Tv{i
zKxUO0$fB|USyfgbo5}{{P&t5{DkqSj5`bJP7m!=!2J)ypKwgy>$fxoF`Bi?PfGPkK
zR0V-Tst{0E6$Xl^B0!=_1d6JnKrvMeD6WbFB~%HZq$&xNQl)?-l?0SlrGYZ43{X~;
z1<I*%Kt)wiJ-msQR3#u;B_E<}V*a;2sJjoFdH(nO!_S{PxO$%X`^(Rn_4eYo4qWq|
zW@E47!>JQY{UScxIl(l@+6ksb)=e-SvVMZ;k@*S6BO4}|0ogdgjL4=5W<oYkFf+1c
zf?1HQ6U>Tin_xC%$0T!1!S~~wWKLvZk_pJJN#;U!Pck>MXOel4y_3v~?3-jhWd9`d
zBL^l~0693xg2<ss7D5hBvM_RFf<^HA6DOF69GzfM<k$p@A;%|J962$;63EF3mPAfX
zuoSX1!6fAL1WO}lCRheJJHfKZxe1m-E<VkQuj1+CX;wm(pJwuqi<Zdl1%*c+STu9R
zeUObkhF9DkQ$GklXibl4khMLgMb`D04q4x0dSt%Gcw|G58IX-VW<)mim<ie3V`gMa
zk6DndJ!VC=^_UGg(qj>1vByN@XpcpaV?7o_j`vs`IniSY<YbQ}kyAaELY8_=LQeNs
z8adNr8RTq_Ws!3|mP0N+$cm4_4_bbZ$wMBT{AQj`I`BH~;nPWY);*t%+<1rS%NGtD
zw(ewO_v25$&D6W_r{888WNn*ik#%jRL)N#M9+_`59@)@l24rKK8IetGW<oZ%nHkyA
zW)@^?n^}=<ZDvDuw3!3h*=9~;q0IzjSDU$z-EHPZ_OzJ?+1qAbWM7;4ko|4uM-H@E
z06ExZLF7=Ig^<H-7DkS=Sp->ZGZ8u3W>Mr=o5hghZ5BsPv{?c<*=9-PRGX!cr8bk0
z(`}YU&a_zuIooDg<XoHOkc)S+;{9-KF5SsW$nu>`9`Xgz@Uy!7si~P?2j?xFw`%a_
zfYI^!OXsf|ypy1^V&6ah&}7UVI9D$|-#R!~Z@cIL<?-MC?!lRVp0|45nS;;0`kkAS
zd}a072abe$+1ML+TbN>MeCAd&#Wcv;DW*l%O)(v^ev0Xl`6<RD8>W~6**L|F$fhY~
zLN-q^GqPohS&*$$%!+KAVm4&QG;_QHw{7P%b0Q1VOh9%`GZ(UZnz@lZ)69eHon~HS
z-!$_f`=^;7IWWxv$iZnAL=H`}5OR2$g^`PUS#kJG`piA3y{t6k!Oxw{KKO2x!IAK;
zKW*Q#`gb>7!;OwFT)A+~;CMMYz9hK>UP-a$AJ?oM{9<_b=G}*-*8g<b%r$53oVmvw
zddhxd`IqO-K07>fWwbN9dU)StEn9ntCbHfSzkU2GPtD{mVi%n-_@YL~E35*vjj(Ox
z5D)IN%pAlIyq#w7rH}rfJ4=JFZ)SGjZ8gK$|Gp13xLq7J``|lkhO@CA8G8&~tyt`l
z;$!$CkRB-^%RM3w`M2*};;+x!_4>^J%KX*y*A6OwFdg>k*XAy~>cFk<=<|;se8YpA
zn|k)a_wNk92z8gJAH>%|c8LaA+a+3LU6<&P^<APz=DUPPHgt&r+1MpUWK)-zkj-6U
zMz(Z`1=-pqR%Bb3*pMAv;y`wGi4$4q5&_xOB`#!lm$;EVUE)FZc8M3+*Cjq=f0y`?
z16>k84t7ZpIn*T~<ZzdSkt1CaK^D72M2>bz6gk!<G30oc#E}zSl0Z&&NfJ5LB`IX7
zOC;oUm!y$1U6Mi0c1adF*Cn|j|FN+6A>Q_pRX=mIw~sl7w|=!<?HJtbM#nXF4UpEZ
z1=87dKzh3#h_~}V2D<^sXg30x>?R<y-3(;0TY#)~E0E1@19I3MKu)_8NU#e)F1riJ
zZFd8C>>ePm-3#Qi`+)p*KTyCP01DcJKp}ewC~OY{MeGqE(Jlf-?NOkZJq8rF$AJ>|
z1W?kR1WMUcK$2YoO54*w8G8mOYtI7Z>^Y#Kz33Pmac9mo_7ae6mw`6eHXM2(+o!ue
z>8;uJqUONccidykhgZ0<O=FcIsRi}Y;nh!FR&O1WM$mxMlr`YAf)<>%tOchNbl`Ml
z9XP$92d6LV!SMnQjxY1z41xiip=<zW6pY}EWg|F~U;<|<o4}a`GdOeE49+50z*))`
za8|(z&RVvDvk5kEwz3VJLvVm|lpWxlf)kvx>;xwW0yv>8fO82haIUfooLg{%bC=!V
zJc0+Dr|bde6};fQWiL3N-~;C?`@s1HKRAEc4=x}Czy-<ya6usmE?5qN3ke}`p>haZ
zSO|j)m&4#9LIhl-904Z^A~>-uf{O}KaM5xUTug|8i<M*G;zAr;yc`FY5E9@L<pj8-
zkOY@3C&8tJ6u4A51x^wqa8g+Umlo3C(&aR`jF17BDQCcCg)F#iISVc)<iO?1IdDaH
zary87Roqmp45{QUfh%n)fs@@bIC+x{Zli7Ep*O*dj{mdmpAXR&_t+PFqMx|%z?rl|
z#&*H)k?Iyv-+{l<wTK2;+ag+IU5n_D^(~@D=39hEHnfNV+1MgRWK)Znkj*V(Mz*ww
z1=-pnR%Ba?*pMA9;y`w`h!a_85dqoNA}(Zii@1?JE#g7;wul$m*CIY-e~b8$11%Cj
z4z@@TIn*K{<Zz3Gks~b<K^9v?M2@yd6gk!+G30oQ#E}y%l0Z(jND?{KA}M64MI_{O
zi=>e=Es{abwn!E^*CILOVuuuW;r*5lDIv=pA|r3LZFLXsmwa{9%HEsjO}@8s=4!ll
ztTa3$kF6go4@q56pEW!qt7Y|fhomWJz-eR+IBh`-PAhA{=?Xe<I#~x!U(kco%X)Bp
zfd|LSJUBza0L~yAz!?iha7Nh(&QvggGsz}!=7JfVSvG^S6fEE@vIU&AU<GHDt>A10
z8#tS61Lr6>z&T_GIA_5L&M7;=2?YV1APeAJ1s6D%>;mU5xWTz)H#kqh1I{CRz<CQ^
za9-I9&R6h(^T|GN{(>KzU-p9w6awG^asXVg5Cj*LgWy7i5V(*W0v9fX!G+~8xJV%a
zE+R+3i3JgyD2w2tg($eF90eCE#K6Vm7`S*L4lXXo!6ga_a0xj9E?G!|OUg-bsX_`|
zN=|{33KBR;mcXS8X>e&d4K7p2fXm1kaM?l@TvpD4%N26qa&it_ac!|QJR=v^7t2E`
ztu28otuKL-*UI4J_3|P2Uyr>0`|PP7t(<w6jHAaKyV>xd<m@pgve07!va82j$nG9<
zBYS$xgY4}wFS4)4e8~PD^CJg(EPx#Bu^@7&$3n>A9t$Iv9%QA7Lyrq@IQu(=1COt5
z+^}}z@&g~NI&kqne^w7Z7#YbAKU_2X@bp6;a^pw1lE<F6>Zn%_JAKWn{OXal>(Bbm
zx)~iU#W{A^tgopKKVtOAqvp<u%v%~dW!~&aaQ?ytU;l?=9Zr|q<MsIsN9%+WPgNiP
u|NZMR$9;3r;w2}Xbn>?}TD{R^w%Sh{vwefNgpWRExp)R$mI$mMvHu1>xIEwh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_abd.hpp.01CABF5AC7B8D71E.idx b/.cache/clangd/index/device_gemm_multiple_abd.hpp.01CABF5AC7B8D71E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9979312f957b11dd11b84282cb7eb5b50271dd84
GIT binary patch
literal 1366
zcmWIYbaV4#Wngel@vO*AElFfyU|<jg;^LB`%$tl14A+4I6?67ZjL*JpAaeBY9sh;x
zeA6Dfd47v8VYN0&EqVFv!uFt^r29veHmRsss!pl>e$ub&>CvV6%+2fTpTCm3`}oDD
zKbd}a;+j)T7aaMQP<?fWy`fco)E67xpTFl_o>2dyQtUzh!-`WfyFbsL`$^M8LSw#j
zcjG5j14)i}=Gi_Dyq9c?xb00Pp6!?vd$~p|V&5H6l}$@-X50@gshkwm<Hy7;VC#C5
z^QEv9k91=RPnvR$NpHa~?O)uSF|HzIdm6G=ybe{Cxz03cire*@S+3^<9tD~62~Iil
zF5GZeWk;aZ?){w?-#W{z*b;2(T=Fl*cX{yrhsBm#QrO>J6smRB3ICOA<oN!=uC<ps
zpU-^m%K7&DCH`4k5AQOZy!ptjWo_T=7~OoP{R=o+))2#bVq13CyKoonZLLRgW|iHK
z@A76}mVE!IU+zJh#kZIFy>p1VIyE!?QgUPtXJPEi^`$*M6E=1CObCB+G3D>d`!Ng$
zCBj#+#YBb~oi2}9?QUQ>+2@V!_8kFR!=|leee&;+5X-Yo0=}F(Zak{D{pe=?BcZzH
zX?9v-ROa9MJ7#q!_t@pl++1q8A?#LS@I&9g7>lJXtmTIB35%y3?ls+eb++p(`8TF%
zJh$EJ8Rh<MG-oKT%uU(@jA@-YFB3$5ORQmIV&-DtRbW?;0TT?IOpL&|XXIs!Qr^pH
z)@max%EZpiz^f#yqz)z+IDndgph@p9gVD2=e*#Qw+z=67GfgvVnDoA%XAKk^j?Wfl
z;sEMamQ_{<6ENvp8%(nPa%`O}%ESqjR*_Xv2NN*qZ>{n(_sHJ17GPq98LrNvE)J6}
zN=+-aVnX<Vkqhj0PnELniMyQZU@m9o0=pRIQyyM*19dBy(LB6*26}K_>~_nX7&R4T
zMjl=@88r=<qSQWC)drzrC!nx`nt~C`q>952pH#72Jj(_&QdV6FrkaOWLskQcrzxw6
z#M6`2L*nVn>Lc+CWDVdvSimxJf&&z0XSWw~%iI3{7F;~MN}@_KFhBC}DvK)1z(Sjc
zS4C7s1{SzHyxKC_$_Ur#$mk&PbY*mrczWu3dI**J>iS4L19bx=o}r8(T;Ez_tAL|(
z6Q%>h$4tgd9p-eHbAS;5ig}n9fZ=1H0S|H>UVQ_7I1lD#pcYVSKxol4(1a`E;WaQY
zfb*Ujuzq3Z7w={Q#+jg+G{W^}f@ab%r@%ykp$SSsFmsDia}t$+$ri?E$eZ`4fbp1-
n3@pasGQ~y7#mYdHOuQVsRkt13_sE%vkB5N=lH@@oAA$)03+MBl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_abd_xdl_cshuffle.hpp.6B55B3AF2C69EA61.idx b/.cache/clangd/index/device_gemm_multiple_abd_xdl_cshuffle.hpp.6B55B3AF2C69EA61.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a73f60982dcf028b6470e9cd325e65bd9f9ecc8c
GIT binary patch
literal 12112
zcmb_icUV-%*WR7EgX?l#7J&s87M3O;y@>_67HlyQTlB*cOT@0R1!M1u8o`2CK%)e)
zfQm{L5y28rQBXkyQE4IqqKJqse#39~c<+Dy6Q3tJ`#$r|+$rarIp+)`hYTDz*i9tr
zKW)e#^MaR5VMHR4KK*a;l0~y->WD-uEJUI|qBF-nnqSsaUb8rBy<TRYQE6{>m%r#V
zdC)**_mB-eU$_i6WL!K_<n!bwy=N7!$@+Vh!_Z$!z55&)JVUYD>*|4DoBw@xcmAZ_
zB&X`k5xgw_ug`<O%r1P$t+4aI*4B5cnF;6oMt^;H;))GUt9Ad{T5aNXWXQ0cJ7&#}
zxj%A=$M}IumxdKo&-i2M!Pv6CW)JMTGd6?Laz;I!=jg9vU464PbF^-U*dX0awK*m}
z^{Q64P4#!sl+wrN*4TDF*#G3s5l>s4-{cw|8aUrs`E+CDFo(rGMy=e{_$6()vghSx
z+=VsM*6-=(o~}Q|`bTp7*!YPH#z<})Jo3<y*|5-e)5V(eJs#Xwe4aYt`0I5k6U3I0
zrIVaH*nWMpWl3%CC9h6Nif*ePJ)4)?dzpPsS+GOe?-c<fx4s_thi-DjT=!Rr9i3WN
zp5OaM%`CC<-yA6!m2tmIM$pWDtE7&*7MWf%D!-dEdiBhb`OCw5<{$Lh)5~w;p5t9k
zu1@Xybmys<N#-Apj6Ij0QvG7Yrj@(5jW{Rs-LhCx5q;k8X@@Q2P0wz8`|L#CnJ&>C
zPF(Wc_%=4^LbvGSoj&?hIESqqb1v^{kDq!?K2Uhc_h{+nHKzN!E*%?jJpX3z)A2Ju
zowl8|TN1Re_L29e;}h$yHjf$e<!8M~%EJb$CQR6zVze|R&FAMqH~KtjOF7=`T{$M4
zY)PB=Yxamy7yO^Zjs0ux{ZoOd^$%vW4S%z%lU4o?bFwc4|6kWSW1Vo*zlTd3)nUyc
zv#xs0Y*}HkpzG8m-;2pdhb`L~ek{GIR}0=}Fn>i=(#=&|+wcE&IWTy@FSD+sP17yu
z@4sd9VDEoo_FP^UpimrY-QI97_0LWHP8D=BI?*xq_hR2wfu}xh9Car^>c!>O{(Ndu
z)QX;y3Y(^X>=fMRr=bZYEzP})9x(aVR})O0zgIoaxe#^t>fFUXzxS&xSs>eN;g|8@
z+K-)<o>k4BlKRST*Tn^wTJmrGThqsX)BT!|GhSUoCN*@|+csvmcTuCqQPE^q!#ay)
zS^dV2U$W_>b?*Od8t625$WJ~qeAL#llGBSu%=@BTFuFGVOzZ2AZn<WsQU^Sg=PM=-
z^%%Oauj|kgCq&iLjFe~lE$F}V*U!832bCYyIkIkBl6%{DW??U{+rdhgC5jEh%5x_b
zzZkuF;hyQeNy2R7YCYp3S<-L4E^Jnx{(JceWs2lpm*o#*g8~~)m6BslWKQ1WQ{!BA
z+;6J0)bE~DneyQDfGSgeuYEb&%=hP9AG`fY$+`t|16L}J{@DwE^dFR|Sls*cZ__tS
zoD&tlf5e8&nR*j$RV3w~9TwbZo-wTd%4b|+QHNi*7hgzv>+AD5VZfaILEDQLoL5Y|
zdt=;^@fMMD=1-rT{ygMyd`Wg|LVU^1)|F1aAL@T}n)JeLae_5tR#IePx^cUM%;%)8
zyN><bAD1kRb-iK|y*_VKlg_TLGdp`++c2-oPr;pg+?1Jnm?&AL!>x>2am^lTopGy1
zt@IA=X7{)1$3X{cYG!)YWyKUMo>G+**kRI?m0?*{)7pL?*Z0q13Ep3H_9aARH8wSE
z&`G>AIyj<IG_hb!#mlzrsBDW>x?}N^%`4W1i^8{CUUs`O!agK2?a>OKot@7cHc7K1
zRwztdrZfysHXMCp-qj(YuIYwdW$)QD7gE(}`}V}k(qHb5w@B})9{Xg&8n=Fn>}HIR
zw;k;~W`w*d@N+MlJuP`n1FWuC1`QgL_}Fy$-Rt+S>AK&^KK(k@^2XV#%|*Gt{5Cmt
z+sbWz>&NTbynBCXeMZCKUFp+PM>dIyo_v+Qx_;*NzqoLP{^a*xRyPmN+bq6yaZJ{4
zc0Xr~xRn0yU)Kj5{%gj<NIQd-b@{T&fRH(B{tP{};m}8=&C;EDt24@f`FLign_X&J
zl4a?-uM;=^x#ID%-6?aU3YM(QtKGSzB-K7TrSZV9X<Jj4=|zq_C+V6qCa}?)b(VaY
zyvCHd+vsoPwPn!Z^p$bs{Dg6jUay@y#)?~+eXL|;_nn7pC!TY7&}toe-nJ%vT%XgP
zzD{Fq)E4VHA4+gcp7g*i@l>PEu&R@fY#xso6}Pm7bgNpl-!9zwS>!wGM|Y;x`F)D2
zzusuS?3zJyaAkCAhvAD>>CKm)z41VAbgv01#&I8piQnw&ms#IhdbBA%#P;gFhg_!B
z=j6WA`#)W0S@X&=ZRde433E>C5y#c@-A8`T4V`ql*{8m-Iy=>+nOptZ<ox1g^QKPi
zAQG_;pZf<Ir++d+)}A6RMy!ev+iQhLLhmK|uk*{hK^qN%<TB*rCE_x%aVB=q3ZN+8
z)%Wf3rO$gKq@bf57b}m8&9wq3s`<;sAv*T=<;a3k5($$?jI{zNmX(Rsxi@X<6^PIe
z-9VBXNUjw?QOD$NTkth&jvU!h$|>CG6z;4Q-zYb9lOq2K?XHU?^g~yW^a?W33ZQg3
zF?LkU==DFzkRu%>0~=@16=?)e>?;G}|FHZRYlMvHC?~|q6Lduy0hC@-qf88^y7lXf
zRFtxV$afHPtpLi<RV#+ho+>Ytp&|4`sn{hI`)b8E%2mZ-yL(9qM#zSa^0!#^H$4E2
z0LqVHDO0Xq-n>m0b)=MJBuz#ZS^<<{qjy^LF;S(mqJi!r&R$~g^1uIyIx<K!)K$!h
zM_A65FW7MFxGs{muLns_&|T387-K~8+^(Irv<~fzl#~*~$YU5AtpG~Ux(YK6A5`ih
zT}pW?mcA96Y6VaxIZd@%;3?10MRH2XK++84qZQvMH=cTjjLQqrMTV5pDwej26<Pt5
z;NRXWKZWbQmH%%pp`>Fd-CK<S$_#_yNAsIb2gp$uZxL6DWu;hOE51>78(dpZyxC8N
zoapxFV&h!ws1-n&nXdm!&euXyT}0?8C0JI16<Pt5`To6E2EV@?V1!KQbWe(vC#foE
z#5c;}&kMviJk50w>mlMI#2q5Uj#?q=h{c^nB6<W1kLQSIJX1Z_MIGok$B^_GJpzq@
zaTcw!7-iiryss|Or<5~DdIlM51yDkMoE#eGywqEU%;=i#Vfj65q!mEf^6YTX+6=c=
z8FHYMY^=)0-dX{a(7tBKcz9}(F6u--bVMvYBJQjeKnaVRQvTq8@jV$br=t{Lc>(>9
zMgS%Jna!1@Q=776s5_-x$KKcRRILEYF6p6`hPQq(GBk)%lCeuNwE&F(%I+?v<z-nL
zcF52Vl#+%G(y+N!042IC<8#3V*PAlbl~QhFi`&>qD}eG>j_;x4?m2g5s2inR!|XNO
zODlkKZ0z(EA6|RM=pr4u=<`T=p4yQ{0Of@B@yx}SuP9~6g^qFw8(*THK_h^2ve^EB
z-NBm=jgT!JC04A8r3R}JKsi6u`iENq+#4fgL@9A%WgOKQjrc}6I%D_FU$*a+BLhk)
zCh}sU(h8tl8rPpmojPEvM!};PSrntsX#`NN-5dV;x2PF(6g3^CfLIj}FRcJdLBNBp
z>xR6Fmmzz)qFdPF7Tr3H07~JS!jDhoPEU0Zp`%<wy4R3_Rsf~AuCIk`w&Rx0$dXbL
zh&+LqX$4RoUz(dYDfw}Q4l-?D5!NlHu3001Qawf)VS4yLl?=Jjx#VMYK6cXzpft{{
zu}{Du&bmlOM@d7{G}KKifYQ`yWX|}kew8xhOGmkk6_;@rt@uVM-?3LeY|bnha;B6V
ztjeJlpb<c6d1w>u_;tFsF4ChPibv9TYK<BJl(q${Ya(ymJS0O_bd*9YFQgWr5#K0j
zbKhJJ{V`mI22)B3c1fY?tr0+36g++LbhSt%$t+x*Fl+RFG0R;*t{2d#Z|e9u-!x{b
z@c_hfrKD3Sk)PM<2FsO`PUQluy{kJ)wKw#9t=nPgE$MgzdBFwBa{0(EAGz`PB~rdb
zwmeQ`l!=Tjczl*#TRuAo2R9naMG%h&(u>E5#66Mpggcz&Qi*#ik2|@C9-O_`ql{sx
zmRrV&?Z5z78U$D#7u!L6z|t$!@`Tt9ZdaE2QOlEJJ8-d`W+;7U#hYKj4EKhxZ-@-G
zg5g>S+akoRgl*+9#3+RBA(E9MOMa<gSRRJ0cpQnHBC#KjW3WjKb_6WdMmAs*2AL;x
zw~nlMPfx8?Y}zXBJxP0cnLIov`FuRp6EWOj#`rK}0_Mta+lb9JVhdLotQ|?U5fOiC
zVGI|>^o5Je?32IK@<sIu`l%-*=m{AQi-&C&OSJ(Lj9@8S*JG>Xb_MNGiOnjp0}KVL
zp*=KV6fDQG&OW7M$TN=iuo2lv{}<oU|3O$q`I?vt_-bu1g{9gM3O>!!dsLE(6*II~
zEX%cFi#F_>Jm~hu6BY0LtS&B^*|64cvzVb%Q&eJ8n3^@0l-}IEK>FQ%vyO6^d-1_b
zKRVS*$n_HX6(S3W-soAEAX$m<tecU%841oh0y{+r&pHYZj>18(J5p`T2a5%*q`RPr
z^kA{h(RGszKba)ZsYHmqBE*BC!8LH&AkX-v+X=%ZGXcrWK(Hu=%VYvFg*cxH$QR-w
zCZI@&-!VhpF+-uh=iusXI;KJ0{x`5yGGh%5?E_n73UNMTl`q6aj8%~kzhhMI7z^m{
z`E_-reZUtp>THs+Wiob#3!3E`u~Q@N4hQHn<f(7QPY(0wn%5H5TGACp@M-AwXl~x?
z;k3byu`>b{>S^)j<{X_VZwwi(g0L05X;K9sY^4xa5z{JS2HMSV)x@+~h--*xjSxR4
zrq6}AmKf9$BfkF&V(>zU>xgk3G3DD|662RbTu;pEiHdJ;AZ86h+(_ITi6`IwinzZL
z;wIwKMBMoH*Tm(u5H}OHS-9S}gncW-EyTHnxbpqq5%wM5-gDZDSBv&e-Aea24*ABR
zK!_W`8&mbvBpaZG0eO)xM}M^GM$dFNW4D{>2DZj>k&Im=j|aR`ZrFErTR9#2EOI!D
zIEXhPPr)o#jAX^g1WbhGwqyBrtOP98W-nkeL$%opy^=IJ3(Se-{y~oa2(P3&$nK8t
zN@_#OHsO_&%P4b&S5g@92@_sPdx^(h;gxibxS!*%AeKub?rA)RoQXOvO~S-aEC)++
zuz8a<L}0mh*y0`b1ik+$(APM3!2S_*kJcfZb*KwCd`PjVA8*gE5)V6N1z+5i!gBT4
zwjK|G{*Y+V4{OpZ{==z=O+`wW8Fgz|atSLS<OKVo$E(S`pyQ~H12X3Y(@TsV4(V7-
z^gwgk;|04?;>6=}VqZ>N0FRlnHz?s~Sx>r@55(XDQ9uKvadat~91cE3iDJnIu`#T9
zVrq}M6S~;i(2M;VQe8uiaPBM(L`|X?Te$RCnw;5$5;wT?sDcs0vqTM77x-?vE1Hy$
z|KfElxsJUcm0-E|*y26EoiGy%sWwC8JH{b(9C8NhprHg(pG3~E|16h))CtHL;=Fd1
z*FTF?pz`fenEC>8hA52XQjj_YIrGPpj`Y)!3IfvhTS$EgIm1d=E*q({ku%(oELVms
z%a9XT3d>a>s|w`8<G0B0EmHG%9X4Et1vnfVh6`~7){DSqJU<E>MhV*wW5dHjd=wiV
z72;#q@R$%MWBp{T;{AOJ)=$BzHf;n;V>+y#idEnWSne^FKgQ;~FKNJr4Ok6&-hN-G
z8nGjfUt!fNA#TE|CL!L*=<Q_8`0>^fy|u&))P8bh=nS2w>L*my_J|$!h<*8kE+_is
zL=DQ@9u*HRCu4d1gbaSd<EfrQG7P3%PNQ01M$F5I4G^cUcIj9eQLRrKY#Cb<xVS)N
zsoiM`4Dex?DQ(bH7a+rdn1>o<1(H<=&vZX_+K+p~nMt+92Cy0^Hc%_jR2pDJYRTB7
z6szHrER89h4l`b0UJxJCsb~rhFh`Kok~LL_A+{mgcU|ninx)yi(GjsPB>D@wH!U1>
ze~EEN>aUPTJ?gXm(X+p9#{|$dG$4-#A^w0oJ_zwg<nd96w_uMgSkONddxQ${ImYcA
z<KMgSmaXTc2d7vrk8#Tr;%dgNn#T(n(Seihp9<(iN{N0cv4G=hKfX?7L@C6cWrPC^
zhEF$BQ$&HiV7Vm3CLtwn*%?@pffW!pf(Kxv+R_T&F%z+w$W&<JgR{}kyoq1NlFQf#
z5>P0`(0jKQ$!d`Sf9{&P44nHPS1YQQGllt`+Y>d#4J0t?e~{uIA<jjLTp`X!ihLoy
zgA{j!_%2f172+MldIzzGP=V#bh;^6{hZE~?A&wx{5kkC|Snn0$NMao+#EC?eNUXuF
zv)nnNIw!>EX&@!UsYI13#A!s8CdBDPmCoapqrAAH#2!A>MBfm-H>4{Z_Igy}|EXse
zBZi?`!BWM5hIZ^xCXCL?$sp<=`1U7c5OolM*IyjzwC!=~YWf8`H@kk|1UCM?X8BJi
zuK%M$uLVYRf$0KniRCUb))yHk9^YbAw}kEY80&k&_UDYnbH<+Mzhz9{^6@y-=BT=t
zBUw4J1D#>HgIInL+XI$rD|VnmV8|V$+BzMtwI$fJ1Y3b~VYwD;*n-V@$NUZ(z7yhB
zY}hKq@3G-~A#THlZ9Lxo1K!ve{<XU~OREk>F=BIgGGJ-QXY{w&9AZG0R>6#7#pX~T
zW4W71|E4gOyoK~{3Gr>De_M$2k$FC{fgNDEJIMTw5Z^`ScZIkB=@$t57hp*N)`P+&
zOKUZfLaYa69F~?zCHJr%m@UiI;LbICX3BEAh-w$H<&9@IQSBDuJw&xfh$DzY1aaf}
zdx^tdA&w*tkwUzWsP+l_ClQk*V#ym*GBHUO;uK<%BE+f0K9#ug{4`>pCdBE)K3#|}
z5|fL3|LrdWif2w=^!Y!}DgSIv&u;2(S@WM~!8@v=16|ex1ycvRN8DqNxEHVU<s_(_
zuaLlzQ(_I0)d)SYreY0N6#i&g_fDdu#&+{p*tvlDL-a|l@)?pnLxS>9ESAS&f6z&(
zw!97FK{Q1>K0vYuLS<}1@+KrmoHRvm=m<q`I^%~(_7EBKU(i&*ff$jmAG*b*HlEHn
zmXXIYYVa(3E4p6kb2;`UZ73loCBzn7+}`(-C$85sFQSoDDRwEvJ@}}i6b~)MV|g6S
zm_!SM_-LkEG}8~Fj>x>h$M0`m?M&yGMGUieO`?`W4DS(3C_h7ao)W8&tct&nq}n1q
zEC!17bi1A+*;Di_QljAsmLI_!zYCfQK6H#}PWWSjjc+PlYzlHsL8Bl{kL_Qk>X-8|
z>brA<Xn=+bnvy@SOW9bGjm^O`xBE|v7Cs?_{EU{H9wFHyVYz87mai3-7WZPOy~1+S
zVLbS-u-v41NPsngpQO37=J|l1Q7&@K73R)&k=<Qk?z|Z*H)BCLw~A3#33KOg;u9{+
zog;}yq%e0rPu$P*xiZV86Zdo;pZ4lG{Q9BE`{<q)WAkEc&xff6#IS%rcPR7H=u`8$
z!LJ~U@d#tQK+`}qY5b~r>HuPL{qkhvZOx~?TMbO<c7Feg!iK@Lo>)Q*O88thrOI^J
zy|8}2(8;D^%T(;lCooOesfjOWrJ9T=>HRrSMO7#P`6i%1$P6?5tn{lxTX)dA|0QBy
zA_XWRt!W~*O^7#Rb~6^VCo*gzqX3JfHBpAm72+y}t>Q7*72O5RGY;(QrQD*R)4kr-
ze0LWvdykw(2E^3UHN-O>@k|f+batEMor9-G3LT+^s9T5?>^Q@<k}j>p9`Mx@!{>Jz
zDB1X3djZ>~lhwQ{!I52So&NLLMOTi{IqViY>=yTjhJWe~u08akTcM2Mst8-fSG3wI
zd~CH4*ATWwh@TVoxe(V9wpNH=5cY)-*Acc(h+h)+r4ZK>wqA%E2-_gUjf8C!;#Y)y
zCB#jHZ4%<wgniB9TlXX*!b0Ms=?<hL%XIWJ)ClhSby+zw&^+?H=LnB-^b{(QtWxMO
zHBWQkBMQdOHfakvWJy2OgXtlI`wE(AFsOO1gEMGPFxvlY0tobOz&0h=i4P8An65F5
zC+H1JA1i#f5e|}ImWw8O(ZraKX|jkqi#R~cQCR%@SSR(3HTvzvNl6@1Ko-_sevzC+
z3Ss$0l7JLoWbNe^$pxeUH{9;WB`HV&ew^0tP>(b;2)44l+#<Pz6vA9gl8qF6E>@4M
z>yf)KjkRtN;t$CBgAjj2)*pp<3%1^Z1^q*@b*K=ZW7OvuS792f&J*HlMqSP02MYt#
zt19<gq9>YyEmN>FUsrjBonG;72wqaDU1^?C;pj>>Jv7nzU0M8}M^TR=j#ec;I=Y%_
zWhrqfCH~+G7+OhjD<=a1gFB*$vF6<trbay`u}>hbF!%QRUwx0*gO6y(j`xT=;EL{d
z`@c^x5r5YXs!fhL?|iQ3N_XuGcKw1q;lg0Kf3fSoJcgHJN~}h*YNP@mE!DmsgTmLi
zla>y8#@i|JEa`Zb7=Y%o)Qxp4AO<kUTIclBUaI&QYFVGK@h9v8pRVhmcwNVSm`fY7
zk!v;@1S(fANeD+h6L->v1jZwQ@dG0Qou#3L=E0iZx2srk6?;S2L*EdwMJxURQZpzd
z(9og`$;yxgtg5}RG^7H3d#Pf%eb{Lq?!~vqVUsxQ09dMh|At+F;x>)5o*>y1VU~Fa
z%MS^&OwFq}bbM3))vPM<V*4d|68WA)fnd+g(dk)kiB`7XJ<_-Gx`qu~1HJ$AECW0V
z-DS-)J-^Er5ql9S;TTx@ge%F!3V0K4U3Fu%;#0QXe?B8fEUB(NN3!R_tV8ox0<f##
z3+QnaAz6_yOZbB1UxZnL=5Gg}<EKw?S#FObTj+7DLy~pK3d-UCW|mcs?$$VlHk6UR
zW#k92;YGo7rv%Vz0-h{IuuB06x|N1aYsQKp3W3_b2tEd3p9uWG;zhqMj%vShdcnOT
z67}VO_W1S_aeWIbuH`jdAI1g@=-<n)Z*TuT{d)8ieJiifzjyye7Wwl(`|v+|^FMp?
SKL_wX+gGLi=LbJTqW=dLjPIxb

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d.hpp.A369BE7D23825A77.idx b/.cache/clangd/index/device_gemm_multiple_d.hpp.A369BE7D23825A77.idx
new file mode 100755
index 0000000000000000000000000000000000000000..980ec748393331f880145b7d9b9711fee50405c2
GIT binary patch
literal 1094
zcmWIYbaS&~VPJ4h@vO*AElC7283cj2xTGla5+eh{Mxa2&oV^ocvu_(nwAS8@U)aqz
z?V+3JH{D%L!n2L6%6Gqr)pXl<|ESU?l`j*8y#D`AQJ*>W%+m7ahNsW>*tx&=EUElw
zqFT1v;R0L0r{=t+o6j3s)zA7`A@|_Bt)K4!`!6=!AC4ceSvu|B&+UFcL~h7)&Oh9s
z{M1{ZP^E^~$%v1a<M#=kKOqZNrUsQbU-q2v<5S3mV@hY)p08VU`tq@)lBh5TK3i6e
z-|c^L)wf6`)T%GM%OIVjWK!>1pZI&VMk}xOL#bJ6<#&vZD)z41Y@?peVYqOPD~sy)
z%AP#AB$NI7J1@R<mdSXzMAX*#<(C4f{^NCr%r{A&EW5Lq^;cVl;?>(JE%z3tR?q6U
zId7w{opE3MUdE+arDlsaACb83P(7PLSxf7Gps#sD4)2LA+MnL}cNFbvb6Fp=`?dbb
z7S_8P-zO<wU%>z3n(0ef4&~C2b(|aTUW`wO=TQs|=TTgLg8A>y{5oSkmgJbzYz4U)
zuDrJsROKDTLjv}ho_P>t<|Vc3Nx+Y8&IdC(C++4c*m-PuXT?&Pd#&%}k4}=3xbF3L
z{+6ln8#BIHi(L)e5F++kYW=BChrCYm-spbGldY*+=Jwq&Kdt!x;kd<7mvg_cS7j}p
zTU?o&R0oV<rCUj>EmFR43ox;90=YoIZK7#n&A<WS)O-Ajn`EFL%ErXZ!N9G+t{?>_
z7&w_2fk}dqoAFFr<i`hEPZ~v;*f|-vm1LFF!30eI!?(L`3AKf22r#k2j8|n*6^BU|
zrKS}(0Mi%D6^xu<+hMlGZnw;dQBzT7<l$D4QPF@ImfFXv+8|Wy1Qb?KQ80k1t~mVg
zNfpb*vrIhP>ayxeFx5QVda`<O9!xPK7uX}7DrMahcRAN}doj1X?f-AV!NaX2sw4yR
z6%V(jjHWWeBrO>&B%Zd6HWE)qMhA(fE2E3V(^J>eLztp3qYsx`Yit#8bZ)|QpkGX6
zOw?fk0doK|7dSX!t_J$dK-~&qlAeJcoCgyHh9M}3VOpLVuzq3Z7w-m!ii)6$G)xuH
zIzba@m~};|If<bB4hvZZm~MvRqU2&_aI&axd81eDEWeD2hl_y=k~%;n4}u8*oWhvs

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_layernorm.hpp.1996D4382B7902A4.idx b/.cache/clangd/index/device_gemm_multiple_d_layernorm.hpp.1996D4382B7902A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..774fb1235d617418ee58e02f783c34863a3d8df2
GIT binary patch
literal 1192
zcmWIYbaPw4!oc91;#rZKT9OE4G6({3aY<38G7|&COI8MkiaC2HMd#mg5NQ27Cw?P;
zrl{5-roGXbDjm8jqB7(^iMd`&@c-DgH)40TlE~Nn<+*xs!5R0oB9s2keR$@LB)j?a
z^UW9XxTYvf=-GGJZR)r3=`nWv(l*~d+-7q>_V~S&-PaG*=WUp{zrR~6W>@h|_u=CE
z3zpVoP2uEV)cv^mdHc4tyua=h<R0B$b0%=X7XLF3`k724mOa^a>inacsp4m2Mdq*n
zyx1zTRQit0`X%0d+aFH+t7K#Hw_E1X_qBc-)Jpibefq|`XT{%l>NieS)UFo)Vm$Fz
ztpCgHjxE{K9<j8}n0jYZQ}UBTj=EFD4yw#Mdx+a<hK`rFlTp=-(v#cW`+MY{xvRVi
zcl~@^uaoU}@cQe;qIw^5J@-7dn<aNa_NS9V;L)oqobOF=EtM*l>6i64UUTpLobvsj
z*3Z>0eC4uF{<qe<n4TU@*09ahQ(01;)f7c2Z?qH&etPJ&|H3RD8(qr?6(+|?*Hd_B
zHTq0s2ru${rjo4Z@HnY>dhS;4v<-(Q#!U-7U@tmZtn}=@Q(L1J-dkL$xu`mQVabGb
zhb~-mFJFAu?2y6g?SgIkC;msx`u$n|nBkHs&%6}Q@}?x*m?id#vy@5b$1wxvU{$Wv
zgoB2K!d_3NoSZUU@=}RwfAr<lz)QK`@A_|IXx_U=<DdWbrNBrmo9?^m$%MLWme7b*
z#k+hiXRj79&sfjyylU#EEhf7%e`Z;f@s~<v@Xmg~A7UqD&QM&Ln^Xr(Grnz4&t1Lr
ztVEQFos)rELsml_OfYZ&<$xglq2<AL-l$9gCN@rp2)C`Ktu;*g0z=`IiEFl<Wn*II
zVBl6`S5pQP44h1iz(mQ&&3M7jJj^l3JxG9w6{cT{MN1r}zbG}WxPcMj3T7^_?J#3`
zxOEJ4tYBvFaGUCwn!$Ln+bwfq)KruidAK!YG&NwtseP=f4MN3EKw&jqHA|RDFn=&|
zg53`jt~mVgNfpb*vrIsnWObBah5&7nHHGuKy_j3x_W!rw;NjK~)sTUO7!S9Bgn>Ll
zi=l)e63<A&2#IGbVT{BxkuX8xnJSxVAxtroFoVmjHMR;kIyYfD&@Z+!w(2lHz#PxW
z1rBUam9p-MyPWG_W&phi3SyY5rv|KF*!jh~fnL-U)Rcy)0vai3D-AQUC^aWh2^hby
y&|`q<hS3aa+V^+GG@Vu`E=n#|2BicpNOAy@Ox%Hu0UR^qIKr5CxEOfg;tT+k1;Hf%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp.6897EB3658A7024F.idx b/.cache/clangd/index/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp.6897EB3658A7024F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..74c94420ea2d67967b98d7012aeca6d59e4f8855
GIT binary patch
literal 21030
zcmeHvXIK?i+b*nGQ``)j4brO=sUp1yiXe*Ew^%@<i9Pm2jm8#5?1%+LR79Gh6cHr?
zf+BXsjx<ZesMsi?prV3!7TNn@=KH?uob%)Sc=O|UB|iJk+_Pr&^|Y0cK|w*LAu1|?
z(*`Y_H~qKCSVcuell;5rw}rFDc2QB$a#B%Q8ebIl!!M6KwJV!f(Pnwc24{Toc6oZ_
zp(S(617>^eU7L5#Sr&CEtlw{It!x@Lw6~aSwpf<7#W8<SxJB;mtu;>z{y5qG`AhM6
z?^eU+9d!KA>OZqfJzFa+`+oN-ud%b(?!GM6;_K$aPO>lB28k!4CSD9ZzjXagqo}Ij
zi+Spm`(_--e%#QIbw|D3=#1A|H}A@$Uv~ca&T`05)tdFc_u%%HT)zGHxb>SS<T&@b
zX1A;04rMw+<HzfLFZG_hPj7uc$(NZAd#*QL;<rElhWU5LhpD%(G%l4LojQNT&6ajY
z@BTa2UE6;`&!<i8*0tG2HAin`u6xpSb4iuKqfK3<+jrkB-(G(@x8;`-ecPa8g=d}(
zGViuGaBYF>KFhVsX1R{gUbL{6?9iwpdeHLJ;4c>ve{Q6|<oan^C(e3vXPZX+i2Ad=
z>86179{n^u>^zqqIPP5Zq9k|!uP@_UAGtp-Sz$Wg)hl(^jj4r)7T!4+VDn-SDp&u$
z!@IZU@{68Di>CzUy>lOM;bFALgYAFaPMP+S?$>l^;=*}uiQRu)m~2rq;)j}_XN`@n
z8o0{i>}HGN(MuZ-H9L*F5F4>^%96q-dq%x`5qbTF>4C)|$6A|)?f-GD(+GKd<EIxT
zNfUEQ_Sjt1{lffo-7+z1ozsyGPN5M;8bd0BwJYBJG)Vix5s#;1PCIA#hicXjf5$t$
z`l7Mm$N2+Z8SVV#H&hrn)3kS&gI4z^haH(dV0ps|PraNc)#GO@TCr#8@E^N0XEbeY
z%^&$}w9(^Ba|Z-vpN>4^yzA-lBm3i*q_<`l#(b-KoZ;8>#H#PBs3S{KU;h1TsKpHX
zr#*WONtHjlyzHIhmPrRQc9xC#k~lJHf!Tr8iS%z~slS|lknZukd8W~fjjvPYIHq0v
zlFPoiIOV~SEk}Pi{#(hJcOh0kX<vR)eb)BtpC;>PZJ4sDEojz-vL5g2qb|jis@D(j
zeADNGSMcY#M^iMshew9}a;4&P=KZJR^_PD%eac4bxU1M5@_L=|`yH)Av3hxP-Pc*o
zi_L4U3uv-@v+J;}P1M4G!y|hPR5gFHw|C89>(%40=htYSF!`};<eu>(>@D8z^69nv
zu}PUqyR;^0<ov9JnKexM<M8p^YIdPabN0qprgjJW?Ny&InXfXiYw?AqzJqpL+#2@w
z@7~?p{=PhEg4Pz3n0tpd^*0#nTeWM^m}OTVk4k*)U4Q(Xy!z-!vrRJNkOb8=%=F9t
zUGLXAeR|iBmTI~Caq!2Sl%#9B3vE89?cbo6XZ5w@{MVohH`-?3UVdoIpq2ZkGW8iV
ze_5C?cI21!$HQ{N(GZ9Dt$L0NkLiYdIWC{2)^(#ny?e}ooRO9DcJ2;Hf1AIPDwwh2
z)bn>neCFXO|EdkMUFL-uOy2&=Z!|Nu|BCT5H{^KaW!i=s87@D`=QmxtJpXcx+Ombw
zyHoRx>lnkl%~SOk70-=-HB$DA8{O7r`=~or$MZCsFBBzRm(B|N<NfRC+No^Hwcn-v
z$CumIE~fU+PD=Jt@1N!Ovwo(REYmJBEXeTbt)vO(ZDu{3x=Z6kdY`sf>6XxWFB_Nc
zdL4ZrTQe@J+T6Wl=^E#f%o$D75Acp=juCyYdod?3+N+v7Tdm*x<YVmJVd|Do#;SL1
zP8pS-^4P>xXH%8+gSPQ^Er&H?@2}FO=C2P|K1-gMjcm1jmd<=K#DD*^W+VSM$&#;C
zo=cv!7lv6r8#w>|#*ev|m&9zp*>I)hVVG}DW3=(&ta*EUe=u{7{#qOD;PmpWUeos8
zgPLpCoPCm#@#LNMq-jA@4SJe~<=eM?=zdal=cLxfHIt2B-i<a({L22O5!ZggxA0E$
z{VrE_{CIen`q|<?O1qqD3AVXbZK*z~sA^v16Nj<;ANkI?mYh@Nno#xgqQ%ukzfKtZ
zmVFkIV}wI}UrWwb1QlMGpZ=~U{D+<&F3ErB**v=2509Tbzc^>{?<>MR;u2Lt%sjL|
z+Q-FydF(<L8wHvj%h=U4siw!6!ZkBq{3TslFmOrW^u;$1*cJ@9J}>sG`Lqdc3%gl+
zPAJe&Z`3;WY?4WaiOe~&&gWA@v~KyMIYA*zb)OvTYX*BNtscx>6q|aZMsjV&vvF>#
zG^o%u@p-FWUJf6&d#&BjsL|WDr`E9YK_|)_n`^TQ!f#fd{<5%F`{r{Cp88K4YQ6fR
z|L<QkQigv2ag6Ps&nL|MsNr6<rtM?Xiy=Eyw)C6U^Yz}7@<WT7Y<`QfUFFitqBy`I
z{HHR@q&sU8T8uyYv5&46e7msl_5_u(Jp)&{^!++(=i$Fkt!xb5dHk-bZ_lB3tQPee
z)ErWyH}IC;jI%`_$6eW?VqL7oP2RrV(}th_OW@c`@Bi)@Ij%x&>x-isWPdib&XhJr
zRK)t7yV-PnT%i5N{cUxlV*bjSY8^B?bFyjP!|Umb7tEc=AKj9?ShITI@4Zj-y!X*_
z+d-@6+aj}u*xoJ=8T`dX+uc64yvIIX3x%!F8rS^F(OT_J?mOM^^tjhq$L3r%cbGHI
zcwx!gQ`emz@aD;q!Vhe_?bVABo>{iX248;mdc*lK0Rs=+iI5vj96ad2DxD2yJU<vU
z1}||MwP1|#;#~C^RkC9vr%Vj6nqK+Sqr2k|&fK}--O}K!E0b?ou^+aL_u4S_`K1{<
zFZRm}S+XX?#4v68_D07=h4|s)r4!mpA}v+3dL%wSF{=DzQT^=!O)icl&dZ1HpKcsJ
zXwk7-Ma%9V8j`ivdaO+EK-BpBU&lRPRoob)+IZS-QCRKcmgle6*u<>A`8{s=tS1}%
z^i%7Zw>qDU78W<I&nsTM^vPN3mixgUPW|T8`cUu3qy<`+ogRCyn%f*P?2kP4c`ebp
zw%PaWMrri?b8zI&pC?2X{<U$%^1pP(#DDU7pLfP;S*Q^=#xUl@V3Ul8Pu=DwhxvL`
zCnQcD<Fx5p(UBuRbqj46T~q5e*M5?-^~z+IMe#oueA{WW;qZ|OPT51ArR@GiZM@&h
z^1qy1s<QU@KRkW2_>a65_X|c8WXDii3oc!Wxqo-rp+8Do*?y|pKkHd14(`!^LFI_Q
z+OOoTzLRoo#IjGVlkSg5v90^1tpB60)4kWP{^Rdu^1hSTNmiZfKYo1LQ`gx#?Xxc~
zomU_G$?DLu+^&(bSp!pit(Nq(>N{a-?**AR_Bg&e=f8V)2!C0!&3wq7{Yz|S?~D7z
zea?}!Q4gn-<d2zM@~Z6F9>do?K1RfS@4sh$N!6ydDr#urEWbZueZ7Lu1RTAe98i)!
z=kdUVsoykQUXFORXxY3ec05tDwoDs$^X=I0QZ%l=iu^6Me2W8=LWLub$)7H<vS~fW
zC+DakX`qUH4JBDa=_>_L)V^e&IW;T&qzq{jN*pJP<4lwSDEuy$6A!ojuu~0j161TO
zv^0j+Rf<l^=d7aX>Rw-@=m%1lI;>ZRdng4^)T0Ni{aE0;Pl|dI${Vcv23spdC#CT4
zhjxeRvr;sjP#Up)BbF-#P^3TI4#ksmG-Sw=R5zdLlFt|`MJJ`Md8n~NkeU=tBc;@1
zy?X4T6hP4}_)8jXv|d98>622TkYN-uRtli#eciU#w{J=xEo4k6j~Lk_#!@MOVz73X
zT8)d-?^?*7P*RaB71<~SP|R<ZB#o&Z*{FqF2&ELsN|BvX0LAjb`0QQRZaq~)nuHQg
zNx~@;r2vYx!;``Z5jVOqNUa~~I#u?Ks<Tq4bit}J6%|#vs<n$nQjGtr8ZFd=l=B?P
zo}+F`0p-}_?>ek&ZhlP-X>@!FNy3q)QUJv^clAq+Fn=E%WI-tFkjXmatrS49e;)j8
zS#n~QF5*c4-l4cVluRjr;;^qRF5`K{8yWH-GbWan#?qQf0Tj2n@(0(QFO#XMM?REK
zN%AQRr2vZ8wDcjx8}naC(M0m0Z<zmvdn*M{ywg7>{dWF{j}!$HN-O4Dv5!&!#pj+u
z)rPlWl{(0Pd}u2&+)Bn;A%N27how%hCwY5G(R@<M6D)m#&6NTu{`c*ef-V7mQuGU<
z{Eel5V>6`yN}vYT(5e6RsTB1glqPK4gdLRvD8atYt<8z=$4JqSgtC>EZKX+F{t-YK
zbn2yn^$2fU9b`zVyA2s`Bh^(1pbU+kH+y7O%_tetA(SLWEr~Hw3ZRUi-C%Ea?co(2
zWKJj>kjVz*t`tD|(QW^BdgjVHDe@t8*+83apdFN=ld|IGl7z{v52R=&p*+S0kFkYP
z0A=b&P4gvV&V=wPQvy`vtGHD?|M^R$3qz?a=u4*Mv;p6KK6ZTGekmG6%GgG0ZzD}s
z2q<IP+C>_h7nj9Jkr$z?r%l(BnkWQNrXQ*KnR^o(Dn)*T5<%-l&~{1zlwWG+d`#kF
zJ}KthK1RBa(UdC$P!<ikICA$v`=zpf)^riurHD0_D+Ev$$G^k(g7Q~L(LnN{FPQ&=
zeU$<z%W3b;##5GElA>P;<tdgv#imLDl$9gRXC|BeMkdsFLivvQ@3@yzbW(n<jV!P-
z_LCuZLb;38?_w>b0LmZtH(%Yb;y|npawCHnflMOE{-O{-S%n+#Um0J~qK32xr3p!z
zkd;z&QX+KcFBy3|NQ(T)hc?pPHquT?0hI8elG9%k{K|EZB`IYsGFeMzi$VZp%@DQ6
zDE{akHN+Ci8zgyy^ppZ98yr`JJ+);lrD!Vo&^v7V4i8ZZplk{(`svxu2Rc$Thftnl
zljqo8DS)!s(0q2@i~|$3P*3uqQ%H6S*(wE4wq?4Wj{Xw2RSP*1$~h!EM<@ybl<4e;
z_dTzlag`!p@}U+S(1Jsi0w}Q`1B2ZCR4z)<&xG<9Hu#H7ZiN6!0(!@ulf^djDwE0N
zw&U#l{`pI#3yl<$JBi7e78HH#xf<$9YVsCI-Xd$IfHDr#1|vQkeWRs?oH{y=mAzq&
zlmaMe#%m|eE^GZ=icCmNqG(wZt)mp3l(E;tj4!*rQbS!xDYZyaOS)enfRf(h;14JC
zOMB}eQ$ksfOx7bWr2tBXr@of6B-%iVT*-&F(6TM$Lka<ulQr*G?HkhnnG{(QN+d0d
zBou`JO5wwd-64OZt7;)z@}Uc?>;hRu3IUYMNwe1lmu*`sMPa0r_t^M7?x_?&xw>+P
z-e1<C<J1sOO8KrT`L3#^6hJAt+YqYSJy)uOW)R8_WVnOOe}(9z{5^Q-@i_xkrKlJA
z&}*#wnkW?t0hH@{`&Y#2@5|Og4y2R^NcMn~q7Xp2X=?1%xS1}OAv;2;XD#Yk52XOg
zZRbG)H(34Xz#xuLZc%EtsIE!@l=3WBW|i}DM>V8QD0N6uhxC;KDEHT&$X@E?8pj|b
zLfJ{H?WDDo0w|Bh-M(=+NcEKz%_1$T!NxV%Rw;n;Z13t1pX;u*NYQLUsmAWr*iR`s
zDF-kAdaLcbjT8kDN-fr_C8a0?P+sUg@hI)?utf{mkh<(bvR$NK6apx(hnK9Lb0B-2
z6wM>0Jj2pwxSLV{<<0e#HbdX<B8z?!p)_N?W^AJrK&cDotn-pQdP&h-LV1C^zQD#x
z0h9)5&NTlKyT(h=PlU3GHr+%{0EOtJoa|q<Zc9s|3^|kP#&G5_oSRYr<*oj&x1Vll
zKdgqV2;~!!d_oRN0hGpk<1Nvjj0>g6n|$aK4){d&1%&|0$7|z#XUC2FLk;PYQr1$E
zwL~9N2%vnbeR7f6`}avT)QwPDk))MqbqWELuZ``JHfio;DVjh&)PTD;;2@;{%C{YB
zGT$ZY9aKXSQc5l*$tB9ULI9;*P3oUieWj-wG9#2tNV17+?+O8w@6}^M+dmmv>!9w0
zvJshVB+FhQfTFT+`iw=(e2G1vv=}fi|4d0gQ$0YH<K_D>zYp6B_$JoBi9G>Jly(KI
z0&3w3-(DAA_GNgvGw<BBbDxw)sA@&18gwcJ^2fB-W7-eYL0;ZQTes1U0?uZ<vKc>6
zXn1)k)4P-j1SKEL7gnM)T0jxz<vS>D2c-*R#mkE*$0EuXluDl19$X#P<&|a$PxODT
z9_vb#rU6f;DA!C(5asOU<u_>G8}y*?r3Q0%8!y#Wdw0tqWs2Sa#LE*n^#o1^RuV5y
z=G2n~4F6qKqO_ShYpArDD2dWif~C(BJy<`M?k4m>JmV10xPkftRVH2gmXf}uEMW|I
z`DPrj8IKk4Y3zC$4;1hvtbYkR3b+#6R^mP@m5PQ$X^+7esylCro4woR5t$58$SewV
z2cw85+OgSI)EyLOo~WN@+faAVOnLb&#NR?%a15*eZ0I#B)A*b!mS=H%7AJ*H0gI0&
z-A_sDDN~_p8}ZPMc!GcnuxkPCC*U${TZX-bsw>RJ&Z;ZzM2eSxp`>3Zcj3hYm_LAB
z0ZWwDW9N$vLnq}~N+ywUEvFsIX(wS6KGE8rXnnvi1If55Y)lwem_ek3CQ8~w*$SUh
zSeo!D4gK<d&SlkeNUvlw{@F|bOmeN}_Y<d0;ntE8a;aXq)Nt4bc=>tS^E^ESCLxRm
znf(e273#ptt5Hxj3Wk%7mtUmFKOw^GPosU(=-$v6UY<#NWYS&&KE}Bp<2+%Uc=>V8
z^*HAa7))Q%c!k{yAJ(rr!L=R9ca>oIX5_UQdBgsM<&nrMQj9a{UYWEHd=bmz8NGN$
zA0|1LA7&g6Gd-bCI{K(*Hsd7V9L6Ds=^@5GIgGD>iy6marY8&&mKU>@#jF)5_+Xin
z)u%Aap#?lqnyhM(Z<<n{>X?~!^^`x1qEV}#!}5i7x~jZ9g&UW`{Q!MpCSB8OGk5bT
z8MP$ZD2cX(rt$Jaw9z2}o4v8g>$2qUzJ%CarMu)GHw^iUmoHaY-igc$7tX8?@4W$G
zQkrfOZ2+ZVQkw1|0fYa5mMFajLPHvnW+So(4V0I+Ak7wJ4`(3RX0Tx_c7ukIjSL$m
z2-u?U$$}>DyOMwXP=b4e^nk)QBJ{vH%JCfKFU<P)n0t?nK=&jI6mw132+VR`ew+5a
zO%I3VWo<OLQt$VTGs!@;W9xQ25{x!pevbA$M^Ek~D!ngAqV%~8vRbkwx-f7RPxMut
z@2Y;F%7O<5OO$>Wp{n7OT{zVX+Ct7j%D7E@{I*lZ?E(fr4f&$NNh1`L%37wf_OL?1
zTSMl#!c`*>ODM+@Dp=^8G}bJQ^??`dxT4cbG%oiijsJwLKVg4RjCgqg?O8y}q4D<K
z@>!7wYzzpo1zWdZf3U20c_HmtC?Xaf7;)F^Wf&n=W9w?{23jgFKS_I@qysxkRJxDI
z?xOG=!Hx-QfILumjNr(0oYv-Rg8sOGu{?#gNulk4*@1_o(8C0LkhVHVTZ{3agY*!<
zJx5F%=$iTa77oj68UI?QA9M$nH!=QAV*H))|1QQGS^tf!h<}9jIl}f9-cMtF)7U-&
zPG^16#m_TXuMD;qJa-MQiDV4MdHz!olUl|U9&}X1q)CjwGbZ1~cq41FkrnZeuqH>u
zIE^()6XSH&BwdU%Sd$C^!~7sKP2qwQCeUTt_cA>SM$Ns?CCU9|<F(1~CNVgPk-%0=
z_CJO@#7F^q)O$XdTk=j$h&I?pN7(;X;+35UJbQ$61<S7B&iz-S^zXr(0RJ8#Ucj;o
z*bnwIiPFmlL?6@l*XJ;k-%GLl9@f2w^+9*Q@-nPjCdL)mwgTG=&nvNRrTBRj)~yob
zRE|#-lp1n$ar|L1KEfFs;f&y*!16TCB#ko@a5`s_F2)(0Nro6_a$Pey9f5z8>v~j-
zvpA0|&I_J{R;DgdIx&S2+lPYpp}}C+ksTlPO+Wzx&P07PQ2?Bc<kUyO$I)N`pFzXV
zAUQZ-h>C;;m7^hW1a)W~gR9X{LF>3g$uEhuj&#~5U95Gyrw6?k)EQo$&ADfbwT=^<
z>j^>Y=q>GY>8Bv8W#s66fy`f^{;-{bcb3e78cJG2nG173;i`o>06tnmd_hTHP-X&A
z;jjfF90`Q@jgo#76BTY-AoeebtsVW<(vuK(Fx_`Bp0Jbf@))LjjDQE0#ry6bKVVT;
zUcMK(??rw>ukS<d`^5MFaz7x(3CKM`j5CmX2J#a|B@?-4ig7k_&lcn3$o;q&pFwtK
zkee`yXOQn1G(f;b$h8Rd7H~PTD@Sg^D3>GOax_4|709&$^%ih7va3dJ!YrspzSU@e
zfS(}OC#bi8U!Vam&<GezUj7m}zeGNO!6Q!0ooFnL#)dFvynGKf+=HzJydN9x7vp4X
zm@LLe@R%cbGVsar_{XmfLkZRk+36JCbYb0O(vvdjDKNCKk&$Ca;aUgcun}H**9Pnz
zNhW9%Wf?_TgY`M8%YdSJ(`DO;ITFWpiQ_b2t#(+C-4i$m0Vi`d$($qLv6FX?JDU8+
zhm=^zbuHv{;Xzo}t1gp{>33&4rjy=Y<R(m~14!?H7-t~84CE$Emu#e$EyiaMe+Jo#
zhnp`)w&LOBtC1}XS6J&_mzSs4y8jQW1p*4B`xMax=oMamm2$jF^@o#WVqAU0K-tF>
zVpc|D^Jpxx&FW~wI@%FV6kgsy8#ajXJG%cnIuxEy%<naKg0+<;nHjezlUtM>j3jYZ
zF#7u#D=;QHP9%#Y!I<EQ$!3zz*nx3CoCB<75$goT0nzzc?RvHwc>N#{LKgWJEZu^Q
zgejzmiol0|RE9$c@fjt3M(GJeMQj9!lb&t*nR(K!n)JbIRomC9UP5OjbDGIqHyGy*
zcS3M77b@TsE;vQNQ(Om~?mGGGDKbG4Ipajm5{Oe)Id*v%Q=>_$VvVi!g;RC~UeSrb
z4MWDPBQSG3!J&A1VRRlZzf5airhTBJAwomS+KHt*v98c8MVtnj1#ucejKtDNaYrel
zH9(v`zBO#(aNB%B{EV$X<1mP-@bWXX@fkq}hp-M=Ih(L_6Yd6|CPeId2KNUnQ3iY9
z3-im3hN-Vq=g1e@uyq?A1sVk}FQz?<>8a2-m?UI1D`G*QRQ~GSxL{P-Z+hg@SCG>c
z<X7w+GW^=y{bUEv#4efGzqe6Tz`MYKzQmGZonl$PH4)k0?z*K7TYy#er8~FYB`+p$
znhBgKj7i5BsV8!VV(gm8^#u&`k5okwbAnj`y+ayv7R%1!zCx=N0Vrto(u+@PmgCY}
zU5JT-+|Qw2;DRBp1LS^Qj4vSf3u0VK>6TKRGn<z`q;wyO@g>&c5^DpleX`_Oi(6v+
zmi2wh`ooeT#sKTvB47xKkruq9q%Q?ek3<<5gBE~|MTpn2>^klvG*uBL>m-hH>sxWW
zmj{{oS)4`|XAivydK>9QMbHf1fM}VN%$0vaWzfAHofUWv4HI`(;CV5=fC4Xw@h#N#
z7BUg|rIcMMB{HQSQr#a?PEd7To`DTBu$6$zv0*v3f|~O3-K^nm)(UidUVe%7xWx7q
zcecwd)>pu9+2L>5FmY!OZed3QhL|5Q`V;{_p`kk|ZYN~{=MwQCV(u-r0JrZdWh9Vv
zeiW8QVH2paL>Ui+u7r3X`GO(}D16}><#<il>`BvEvkcY~ng&rq@}eR{D7?6p;<i#c
zLe*Yl?lsmaQO0i~(0#vFA8Q?A;w5rCSw}G39oQw2a}{thH;`<a0#4xul1&qEL}7?S
z<X`!#NE>aOZ2Acs<4UKMooL;`)ys#aUB0SD);2PTMHa&Gx*wVDM>YaJ!RVe~%s?j~
zIx(YL&X@^D%WKBqHDdwo=jDxzX(MAJ;BSn<H}Ugztm!&d^n5RCx>t--S%XwoR9-6U
zo+{Xt5Mm{J&sR$Nm2!a*<mCr3e-L-hNr^J>3L^@ESJEx7Dd}rr<r7!MKT%j9f}%j2
zj*8f;FoM^S_I1=lSQTq1t2I<_@X?c|Q^t*AS8Ow7+$_f7*f?CQ4&-3d93jN8WiYj_
zVf%OQKD@jy-&Yh;h*IUFRJ+1iMx4{C{8llJKsph~7*;<|W~|OyWDHZ7Cyo)Fb;uYz
zB|Hg<>a0h`5H012I-|1z8H0Yqlh}#QMq~_O6{7f~fE_3pVv!wq@D4Omj7RN2;|076
z+3rG4!VKDlJa?f!V%&ch8Y19SWSfed1O+Y?d8VR1V%$F!4H57uWOoWV!#MF|nA}ex
zUoq}`3Iz%H9P&Ab0)$~Zk9^LH@de~_L5xe0eJOGg6vR^GQ7YE|A0XQY$O#lio`h1}
z9v~k9KS$=zkuB)=yu5<atf2ITiCRf%R*G>IrCBA$_bJW$V*G&8d?3aTDdUIYsK_Hq
z^O5*@B+W(A8W6`OIvvezp*0}%O}0Loi=s6k=1%4l&26PMzz*T%hiU#W-4(0>;xMLV
zwX`1S6GVBWW%XkGj+VU><4X*8iO~>ti0ce@U5rZ^u2hV#u_o78D`E35VNFWJ_&RHH
zU5szACO5?RCTnt2jBl~Fw^%2kyxXkFZSnIa*0YK21BVHTgt4B@V*G*i{2<03S<jDR
z{E7AaB*ralKnoizl=qqS{46}*Qro|B=8T1H|8Y-OR8ztE#z3<d|8d{n*5~mf99NH<
z-mw*{_Ij=AE9|hz+_+>R<N!;SsPBs87J*oY-0IK}NI2O3V41hN%E^X~Gf&sXQFP|%
zuIxnMw<V+vSh_(_S0u{37&rlT&MW#py>wz9qIvB=_B)U#Y;7^OG&gN@e3eNaJZ5B%
znQm~(?S5*1&j0M8TJj*C(~sw@;4t3ZGHGIjhH(Wk;bLjWSh|<6s$%IOv2>__;~4!o
zu}K%lc*ZgPKzk=%X~wdEaS~Ql0pnJ{^aUJS5`3g=!>S&nao<_(@2r_H9b-A|SkA0L
znLPkWKZN*{l0FriaEjC)=!N(L4$oFzSnTyb4B_~)?}2;v8LcCYJw<V+s2*VWLJ*t`
zS4XChU~QeCwNKE4VFO9%_oTP!$RX+m9gfQG_plA9i5<AtJ?tysGTgmP{JaeJDic4i
zz}^+uN8ndrc?Av;a3$_uDSp2a`&D9p;dvGAUL}5Bg?m+r%Rk2P$Hb1#W1ROf&IbmU
z#OydeTl_qm^UfAOKhE*T#m|p(p2s<F;r$aFe?t8H1m}4|{5*%_bHva8<oG|u_$0@l
z6yscu&;75sM=sY>_+B2z=ZT-^b9}ylA(BRPS4AS1u-u|4E}F6ubk_#VHDD`9@kpM(
zXe0`)h$q&55z;6^7Q)$Efix<Rg>aTWK^jkxg>Y8BL>e!V1)P<K0w0<5yYO}Y|8*80
zdgFE6bQV5F=3p*smCL%oN`z!HVjL+l&7ci%+>xSpVCfFr70zQ4TmC2C4CXxqMoBLy
z(#xO(k_A9<+r%3D8_d1IIuPUsg^IkW$UPHYyiPe@rv}10;pOS9S-KDlfs>h3Z9A53
z7gtS@ng&%ny7V+HpO+FOBbgA0+l*uoeI~Xfa=eZ_V8iV2F$TRv!=Yb^YKj~hksFLK
zIrETRGjah#sv~yl(1P5A*l8QGYeO!O=EIZhlRj&x0b*6xqmk;gT}gSAN3$4*V~=pb
zw(W?;y2s+)(C5TWh+X5bkAM@fdxH4+A?$ic{5%JH<p^$pEH@wd<pY!UkRIDf8}FoT
zz^6gH(wugTu-l!OaNjj0N<)hzVBEu+_plDEeVznGG|R-e0$WxHdjT)6#F~}j=T%s<
zO2D99ks55n(rw}~Q{)Q5nB~?j&((`;%J|oZne#kz_O2+c4<`G4Jl8c|aK)a|J@6%|
z#;_wJU@g^cEfoUWK%tbodnK%9*#F<i<^QgY98_Pj2`aKuVJAA>+g!6|)2FS(dd|Rn
zhT!NVW<BQ1v97Q@cC-9$Ru{I_GuJA{W%)M0_}9<kv+g0&=m9Bp<S2?o-m%DEFf-GT
zXPP)nbQSqtMS+6Tvxaf0VLAiU#e+9nG~D|20_mzWG&l{75bVz^6q1EP1$-5axQa$Y
zLB$K7zn##wTuvN0rL;jQZ3HSFmN(M|&0?I*=w&ko;99}*9L6YzF#&x3GxyQf)g_Ol
z$mDX}ayfJ19M9*h^En&97t`eP)q}WoWH0(Z>wcF$%@~ut@LUpE$*)vxU#SMdynu`m
zQbH>wZKXQB&=O_NGVErM=%FD|W-3Doyu1YIm5?ksWoA0bKcvj+C>vpiucyrF#ki5u
zYNSNLu4YQBS&Z8#tu`@kr?lF|I2>z*i-TpcSQd*#R(2eg#ffnOmL-VsAuKy2#yMC!
z2OA6ZjiC)<Xmhyvb@g89jT^NeI#T@pYcku_H*w(yUVD}@Sbl`#j|kdSM;M$>6XSG_
zPZ#41j?WO|OpebK<D(pZRKQ?_5cNWl0WNfX3ereH#==ZLi8M|kW0;`eVI*f@9VM+3
zT+_r8gomydCn27~uBXIFi1)DVJ#iAEBCi}K188w1F-eh4F4U+QX*MHULA_{0nr+Bd
zP$%NBVH_6u?hawYLju0BFnwg}^LdN@|H1MvJ6OP|BnOM)jsT4KZOav#BK$^a{G(?D
z+`|KegQN@xl!<W#4y(Wugy)qws1gqra1{=y68JFJ)FjHg2Jl7xzvWI5&5yQ>p&eks
zlJrGJn!@P9ZrXv}QiKFbUVfV4PBStvpgXY5X|aL*h_QdfxWWlY!fLEnG20hnSj4Tz
zYE-a#aMpJ0z1_aEw$Md)dRwmy8DSjzuRk&Gsw^mvKALxn9169n>b0tN(8G5_XX}4n
za=?r{2*iO>aH^D6E)I3*c4aj=VXvTWS5Qwl_{jExjOs`(h4L0iM;Or}6FZB;h}~k@
zo)AU^9@)?oS4H5%#0;Xi?c&__TFljA-Dk=x9TiUXkNzApX&0&3C8Tjlh*MR)*fjQV
znVl**voAB&mzmze;ND>DZ!rD>zQ@?yWBLFF^MT|oDz2mmRe8hOykUESSxK}Cu3Ic;
z1KyMlY!}Np1HQk$N}pM5Sot6KvEBa=A2M@q<i`JTUOk9Oet!Hx@+zWLByf%ioF8Zr
zm_*{8lDQzja1Dp(1aBzm8{z1cC@<r{(tPTDC$QzHzUsezrb9j^nI6w6>2tBm{1}!U
z6O;*w@-7jy5GEe+;x$mx1|eIDm!H706Sy-Ui)682*Fqd9Xc6bI{yA(X;0o+pfrr9w
zM_dg5Tug!w)##`8*zr=MCs|wnJv+aGcBuH@XX_IKp4M&=viCt3AZtr;p-D8Mk<}aI
z3vtR0oxrY6NdJD}On+cC^tO;uuV=f|vt407>#(f5#Bg0<roZY@kmY8QnL>#7XzhEn
zJ*c_F7eE^~)2@Kw3Kb!~qNK0Hx~1Yi6*L~MqLAVFPD#Iu^^_`XTP4;*6t}G4#X9cj
zcI0z>2YGQn>a!mOL+eT89QkIUAkc(JRvhE<g6T|Jg?m|~YPFQKR_qH^T+)Is)H{Fs
zUiDz;148_Yt-s<>$jT)73bgTAdKmaN;TjhqZpG5A;sg=JWiBWYx{;iuikn<8Kn<-E
zJ`dZ!!HYB|4H=}NZbId=kZBg`F5s)k^eXBO+b{TT$R4J+_yr}BI#P~jso{eAsS$IH
z*iKz}hlH2kq<wGF6QB){2u<3s4olaG{p5;!V$g=B_<~}m<K|ZX`i0*tdsDx5^9c9<
zcnAhOCnhAV;Gb)C!WiXJR=Jdya1&-d)>@Cd3meBqthG^$H)E~MVtfi~oWd3YzZh#5
zW0CoF5o=!*<EvQvsu<tE+Bd|w8f#W#E7&qSlH@f%VJny$WWT|hE!YZTDjjJ~nxC;1
zB=U5u6V0#KN?0dtShEdV2}w%rShJmEh$yd#6Nexzi=!=J)8OSfw0;il0CTb<h2H!m
z?JnQ~+PpxF3u*I0F+NS3pBCdH+Pp}Ni)r&>F+N9|pA+NrwE1~4-o|KbV=RRF<+2*N
zLYl&-Kd!9OZ9T8?AD>}hVv$`yaUD~bVa1eVF*Q`!1>Rxq9kzjl3X<?d``)6*Ll1v;
zTW~S>*cJc(XEz1}3CNk&LP=Z1Swp)qzgv7=?;6&>hI_#rmMCwF!gkZzdbrr>uh<Vn
z?TezgD9RjG6D%4McT!v)71|s@a}l&IB-w$xTuY+7LMjkHBZtpu0PLaURvmKqBF5j4
z-8U2ny-P04AiH)k4ySy=snNpqmJgK22WkW~g_nP%JU)tX8|BzWiEb{oQ;zLoyas!$
z!J_vgut$U#Ct~+RJX|P03A-nWaVmCA#Y2VXhq3EnF|MOs>cpAp4YW&xfZK~6Jsa)W
zFqU+6A{UUz4F;tUl0t}Hp}1iy^k5|F6^TYc7Zc-va=aqMJje~)xs&@78>0VFwP)Im
zt3y>GvI$wPDlpFA+*bju94r(S_}HR_Qx_c}Wg+?KQRwjkTY7=|*w8UUhlLCuF>=)C
z!9!Grg0Z5aGFtda{y&ZIXE^*&Q5h`!93}i5Bm5j8{2VF#93uP-6@HEtehw3U{x1&H
BZ-oE=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_multiple_r.hpp.D51C7B3444D17F61.idx b/.cache/clangd/index/device_gemm_multiple_d_multiple_r.hpp.D51C7B3444D17F61.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3b1ebace70bafe7d11769637bac2982f4a75d968
GIT binary patch
literal 1416
zcmWIYbaQK9Wngel@vO*AElFfyU|<jg;^LB`%m5|^hG;ehhKf0RgJSb%8;G1Ou5XpM
z(wv#W^VTjfypts?G<Fl)*TkSuMfN;lkqt9*g<P-x``zQ&b<*Rv8E5C~;&Wej+LgWk
z^yOyZ(`^i1MGp7)H>EE9lf3Zq^UrK@k{?`|e>XPP8<c(QRQaB@w6p)1o$jN_8z*NN
z&X-op5R8{n%Gq9H`lgwuXhG3rt{;&M>PD#{-7>dlJl%1MBjpYI@lrP@<HDG-M83P4
zyI;l1#Wl9_T{fRk`+@Dtq@6c!G@5!!CtPT3jXiWO!}qxP59LCsu6GL2#ig9<s_fQW
zmRP5FUOiRqvJ!KBe9QVt<&(beyZ>hvi(XcWqDm^yEz!UR7SA@b38@;5%cp21ajtlf
zv(C+OC!d<|cb0`y#Fo0~eUjS~ecMTxyK?)v#g08$Nx%3X-sbnRt=%%={hQSf3s#7{
znw0Gs@K{Y`m7Xe(?BCU2lfUpz>=3QZZ@9O(GPIYCi(f=eJz(weQ%Bd$De-qbmvgkJ
zHfi_&=*%~D1`-h&&-Og%>%1Pe@ZF^sXXHYP-(2&nfA%J1o<J&Js70Hql1U)5?#W%}
z0z<CPddYrw_5{(2n!qsen&UxQjoEM7*4^a#AvbH<&U*e3gV?s+MsYjWU0FIS>AOLU
z`Niq4cI97KZ!389&y=(WNj=kb*!PI2C@RH1Zpv+NZ<&;%9n_dSP25M|tCo_fPAkuH
zlXIt73icdozg7`AWk=clzl$sd4PwNO{x6JPvDxkCyPm*HQ}&oLYpXo--qZ2e_+*5d
z!#vC9GmO;=H7g$fd{?u?ZjHqu%bVZ+G3{0{yi{D7o3sy@zL<_bvl8q3ATG+p$<4s4
zC99<lCKxz?azGHW{l<}pa%r5ROdLRIO<7HKFaeW}xOqu6>xOQt023QGL<6rbi>^3K
zy5VHyzN8)JGX$8}Vbb=R_SP`z*4gXo_T6ykVq;?FV&GM0S62oT44h1iz?{Iy%Xle(
zXLZ-z&{^C}EI<YRZ2qcX0;B+F8&E+}YFhCZARFcjCN8k+VFvT?8fqDuz#Iwl3o{qk
z4+us22KrVoMLfLb2Ig=cOq7um>@S#>*zJ}%F={Hxj6A$LGCCSCRjGZfstrQLPC#LG
zU3E*C>Wae;pH#72Jj=|(t1qjs1XB%koU9=d&s^3V&VvOU&|Xl0!AyeDKyQLV%TuMS
zd*UwVIv(BtkpKf&V0C*jx4iBDZ^6mKt0}4}1M@Wxua>Bm3@o5{c#UL?lo3ubmN7=+
znaG$R@l0h*k$7e@W=K49b#pz0DHbvoaJjX{Rsl!nCQJtghP{lvI?P8fS2A*eqY~zY
zrv|KF*!jh~fnlU0s3VOq(q7PB8m1TMPtE{gm<dIxIf+WZWB?0w23Y98Xa=d-hi}EX
ltb1Erlw7O~N}fC*0s{D83<f6NWrde-d|#rrmWhvt0RWI47&!m{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.76B6E1492C6F98CC.idx b/.cache/clangd/index/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.76B6E1492C6F98CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..28dd35d2d37073ccd155bcde89ee5a40f5a7ce9a
GIT binary patch
literal 11816
zcmb7qd0dUzAOC&Mc^=j6aohJ@yEbW2Pf-XhjBQ5tWxit_jcpkFB$74NAeBl<B$b^g
z5rx!{N=l)TNJ@)}62H^j`|&)#e}2E$^h%%mex1)f=W{;ivz}v$-`KIn_9D^fU;New
zE&P1}EfR@T$iFqeuU@)XP9(}U6^Yg#t(e}l;+d=Rn>T$LUmj6@b%lqUbM|Eh8*G{p
zQ62PtzRe_dhpppf<y`-wrDe(AdVff^irn~gb=&~$At~*(AHwT%SG?`qyQ0u|Bb&Bj
zqq@?I*aNpLdY<Py+*rG&-Pfdhb^H3{N4<LvpVNuC(rU9n-)5t8q2J$EQxb+<wB5fY
z_JV_EQ%P-nfXmmI)hd~*{$BI9&;PXJk2;;~DoidIJ+MG~&7AZk?O)thrH$)*d*OED
z^@yQPtI@n;>b?8kP2O|KzVNS&M}N-Io2%q8dBq&1spB_ZUsQfERAs2T)>)^Yo}JLM
zo$!Z$Zt(EJE~`_XMt5$^+EKoGeC6lZm1oU%yS-!eU!>0S-LY8DVbPqt#UI1RKG~=e
zW3!+pGjiM5-{#jk9L=&m9_N_)!8+hbNa=%$cl)*$UVWl-SLc80gXACkZ9QWazf5kc
zcAiGPQ}fk*n-h+EJ+)u?)zRoq$)=*~du?4_M=3rZTK49Tg9=w>U)3|5+;zAoeM!mv
z)@!3veV%7m2AO&O;ytJIkK)wv_d4IzYZixi9y@CDvSye0`xoVVkn1|nZJwFq>}|_j
zbaH0zlK<j2aBJ~#G~&<p#`!w*Z-ZN(H2meD;&d|Pq)*e5D|X}NKhTN%yCr^@s(sb@
z7ioW`pIta_cXRXQ_kT1+{&e|$_p(-v%Lga6bcfoFi?v%E65TuH@NDzbb%_V9Zd#hR
zo-sT#t)p%B^uXtrUnbi=+@0{W&yz`7wCC{&o544-jT7H2t3w(MMSDy1Qo8R2G(24u
zwBKgNxyfG}2P|^d*U(!y%CMt*+pBRyt;98c=K?Ltvn|#<PNCvm8@(2nR6pEsp`~o;
zlP%@RT6(L-Xa966XxRaSh1$axS4cb>RPP)evMNqq-+agEL4HAbe;JjnJkxovd9eR*
zE@_?R6-j-xPSNwDDX-^QB*upSX?Na9KL5iCNz~!xBR7Tr5?mL(V@Q!{!*mtNZog?O
zaaF_>lW2F<p7Vy$H!o-B52OE_GHIlS?;pp)tk1@rPL7_NI&}6qzf;D486V7;_0q;v
z%OT*?EA@G5cb=YK>epa)C@p-+m)+53KTO^I?q%T_Ta((KRNhQ?+xyh9BIkp$(pTNp
zM|>0KNw(ztc;~Qj!|i#aCAy=Qe_MVt?RMzQ6)`J^uaEiEUiPQiNQuKgS96`@%KDU=
z7F~X^N_Ww-gh7_a%5Tj4x?o$gsxGx{eMDr~+qJ7Km(~5lrA(@~NQpRa?yGU;n8q}t
z(5W>W%RAOHtEP;!334Bt<?pvJ_4LH2x3cd0Fq7l|sLov{I{k;%gwf4U9_P>MI;8ph
zT3Y@4<>T)A%`hJ9UzE}`RAE_2^7OvQ7NcvU_8%L+SnO~9^-)6U`0??3el`ENVfu!(
z&1N~wRPD#_A}Zp16a$-;mo52lZn(2r`ngA!_m5Nv9iUl#$-LYoy?cB4sL&GQ%5!Ti
zt!@v#LS_8;K3C2v?0q&>AGN9YtDXMD`HH^|w9Fn-`%B()tG<!7atA9<bOc65O_(*N
zFva&spz$}QitL{oCa;)o7<BQ*<qzYNy7CeuYTf6an)oO+%z3H*n%=nWwQ8FbKO|Jl
zy84^{ME8o04T_Z$8rmP#I%WH`BovyTuhO?KonWzYchJyHiyeOD)-G0B&aDX8HgCkz
z2iv!{?Teq)eKNJFINYP*-RK|dbVFUuE0^4x7WXvqRKB%1XuzaCH|JEP@Iwdd8VmOy
zRkm3C+0pXR-d)SbTer{IWl>b>^t*e^4{~`6-m2GG<YUh-^3lr|98q~0`)%;P$oa+o
z`&PX%J+}DU<<_SW^Tv%E7aaWB+uI@}GpfOuxtSgL#k5uHlhUp#V^3|vf1;K5wYqS0
z5B+wP>#YxpZYR_d58k^|94A>h<FTfPT%xGV{fNHd<<|X?r2`MyhsRn@QRqxvqy56P
zFG#;e+j`kO_x0wk;j7JDx9Hhl+&SRDahC$m%<$jN%O?fs7Q_{x(@}Y&mL2d3H28ht
zv2^=u?~-i3nOr@|&S@*KV6o<855?fL4HLr(uF_X5e;;`8t8c)ZQF<Em;<WkJNjE<@
z&YA1;d0Iuu-k;@NceL4cZn*i+!wo&|53#sp*oY|)tVbRi<7>I8KH=C2i=Srf{{4KS
zuSVH^<6m2YKUOPV?7e0DS<O6fz|5KB<*>*7z47l0cXu6(P5CQv%HnrRCS}iCYvG%I
ze*V^rfmP;jC9|EETzS@X^LVq%u%N#@l$wWp8n5HHrEE*RWW}s)S_5ikhYYJbJYwy$
z@D<s5ahEeTMbxWQ1*gZqJ!kj)hf$h4#x)JuysR^1!t6D#W;I!I#%qtoH>h7d>nnch
zWH{zYIn(XgynDv)mjml8)<%7-5Tk?Zm(Ex9__jIR!z^q~^PeG}x6{J4u3b-@c*8q<
z=Ir{nPBleNN#nMdBrSdR@Skr7Eqm&-JRR(P?<ZdE7;)j&!q=h*txplN7nNQeseP&H
zTJHHXfhx9>j*iXR_gfq@vi$kV!%aQ89wFbVZ>~9JcB`SKZ-+QxnXlr^(Q(FhJ`pv~
z<)b`ielhoukJ6j|#r))=lxAnzF=LDIw3ojHjeTpfa`{K2qsQ?;(V7iGKZlZhD|&R{
z@Mnkmp%UchAmT1z^-Fk=jELmPF!_;N6}>j1u0T~0iS0#PH4;}NZ5aWUfpY5iey*A^
zQ-a2lSu(L&Cf1h`U{Q$}yd!(0lDPy~6P7}(Rfw%+1Xxse)n|LmOWmS~<jF!$koXBQ
zkP%=}d${Mx*SXanl#vc$X`}}=(k?OrEE=gL)9cx+QVAML7P^QvFJcQB0T$i3+tqV@
zGMfh=MLQ9fPRXTHLu5pxfGLScL~)ev{R+KGEu(5xq)V2nqLiyBOBsQAj5g>NPcHZ&
zN*QSqmK<6kht`)7U@_HkD@&ibu2m7q4H0o=SW<>HWdvBvCmY;Z|0c#k5e+1>#2|4D
z(vp$yEN^@DYn&#3kRS`Pry^`vgl%L5SS-&Jo2uP*(^EzIWR?b6xq;S`5n!<zJ?j@o
zx%q(-<WFWP!A2$6PDX&mHqdJ472EazRb)tJ`AR8&rOagnSnMu*x%(vOu!jVVBeP^-
zyDU6RMu5dBGSuh#g_EHY<V09*V)iEHWCU0|qn?X57Zx>XAQiGu9;KT{>B|VPc)e}5
z^ncs#rHZr(%Pm^@7SXPh087B5%3b$L+s!qQ5n;)p406aBNeQq_DBL<vY5g`k334Q|
z<YG1#d&mf|{E%mwwb$o$ry?>Tv&15CEYgq>V3~9w=iUVBsE!6A#VeOf8RU|Mqy$(3
zEe@<yi1T_OL1tv3t62Xk9x5ZiGV6Lw+slGe;mT+bF@{uHl1i({2(ZlI`ifT{|7(vT
zV#q>eNL)sIKuUmRL70X?d}+In1{y_J?ohgSh%8b9EK3GG*=+v)f}-lbhM7StXV9jc
zlmN@J>wZ(UdM#rl$dZVgjrFsMxKaWvD_kGt&h>o$lO~cQEH9A!3#2R~z_Lm$G%Y&8
zZHXdci7eMB@ij_SMu26V_mb5k=dINnfHVk8DON1SsxksB8)`o4=e+z-tcvsqOB1Es
zM0O-4z_LR*B53^Fm>0^(fUq23Ob#$kG6F1N7j7(kYT`#pkPQ*{0XBYs9b^PpB1GBF
zlhqFvDk5bvOFR<CBTX3rmMER-sF@WrCu$&d!jezv<`ct|lJ6|#Q?nOUmTi|FV+EE}
zU=>bEfF(w8=L1Le+(8;hmCSOJ(!EJ$krH4@D(_fjJ9c%ZD$*h>cPZt&WYbatESU;6
zkCw-*SRp~9$&S*oPC5zBQUWYz7fczWw8n0o1X&T53t004iFQ%~ESFk*Z#3OlF<Jwu
zk%ewiy0?fdQUWa5NuP}3a@zbPD1gjz8?(3ZU>O0HYuZca1YO<Ut&Eh&EQc87LyUop
z0L%41>gmnBX<dqF0AYzn;%E}cqy$*1GH*l}p87b36^*nNaf9UtyZ-w}q(CFl_`xLV
z)&w~894{Qvu80)K97mA&2$@4lV2)b9*^_3Mj$1SU$q|-Xq*#m8WCU0m?ti$cR30cN
zL1V~%GO%3+iBVDlERDM*R|n_&5#9O{mSSvFOoE}5088ub{l%^U*+mj$N?7h=!}}!a
zO38PYHp}+4N;;z?$egg8$C~Gf`lSR|+8WP(Rxa;5E<rwo<qXz2Lk>|&fJL-=;i5G<
zh9uX@vN6qapOEAe(gzo0xjoos4|eBq3JyrYGXRTaIs3bAuDcj}j>*iWSZ*&;+l%ZW
zutPFOVp$4@T*#7SEMJH<Ay2bg+}+~_jD107xJVmZqz6Mxfn1MRt^-LrkQOh4H1UHB
z1BdL2U3SRv8QDx8Wt2ymL(FH1Cm7wN%)v=mVvR<(D047WmiU`dK4lKBIB>D<u%+qx
zHz}G+X4qs#0=B4BI{a(%gh7*tC|TGj3)@0&V!0M<(}LY>;tNHAr%h~>+P$}ob1jS{
zjQQ9#A5Q~=XSrVN(2FMmBkUftTriddV`JDH(Id9mi-!Xi%W4<cEYvPoEUQ>}y=G(i
zY^(?ImgPFJVJCKk6``V`#IiaDz6f;;!*bn7(v1vZ>3-4tg88#&x!u@iH}>Gi4`A&B
z*b1;%R{emenl=mmzWc-V9&)mw$R!jF2iwpd^?JhH(8DTZpd8tkqbZP#v@6fae>r(&
zD5kl3>|T$#C|LxgxinfUjn;<5Msqi4(;KuI6h^QYVr|kY%g=1x-pFgCx3VNjXk4R4
zK>mPYi>OapX~8#H;{2vZ5C?IA<+_k{7xIS98A-Oe@0KsFBQn=wy;^JqqJaoRuhPN{
zW;ANFysWV9r5E{zg~&quZ;4C(m|R=E<bAN`PGV%WShE)Ez(j+*E@rp8Jre)t%|q!&
z_9U(QAQQ4AS!h>aV_1PDTC_={oj{9b%XT}pZ9g@FNEnZF;?XonTTm&I8Kq?+?1|Vv
zlFvdKz({-+%U{GAU_<67)_6_u%1j}ycpaNw$3H?<L|hXa<>5(Cu@PIt4&8V@SSQPU
z#x|evY{2HSLX<dC2$qCkJvaxktQ-X+gL0IpUs{WTs8EEG8A7on6n_tKVp%B)Ggyxd
z8ee@NC!H`>qoLJk9E2{G>%%sE*c)U5^Ceo57Ob!oONuztSmXxvU4Kxu{DeF~W{4KV
z**_x5N2Cn{ESH4Ulkj(s6U*vb5X~|BW9GzZNybE-*RcFGY{;uKAIs-sLq5E8VeKw#
z52CVM57zGCF{qtL@D)kE@}jZaeyqG7j{u3pvho=uaCz@qwfuhTBw`{~w3P-JF=yCT
z^y`2HZ7MVu%doM0q@=kxhK&>Aqm0f`Mh_w>&BZf1@j{%y=p+bnBBPTi#K#!5V~hs>
z{Uk;$Nr;mfmt<xbKYpCiI?m|uIEB$l5#ke!-U-HlA5UfUQib>=Gvp*QlpjCE3^^sl
zX^dSO<G_!nGi*A)Ur)*KJM+exZ6NvzMi#-y4~zzK5{XdKCWJShdhAk<hk=tqH$n_r
z+JgY2&;7&ie(W)IC3YG|TgTDggCol&($<MQ9#?kMBW~uHHOeekOgR@*BcK9gxf05`
zM2Jf%=TaelOgTRm;^&m}b7}+xO_qB>IlmC%SCsQBA+D#K>xH<La&DzYK-tZ5ZIpAH
z5O-3}okHA2Id=(hF!B#ZKY|WfZXa^khunF*A35w7;&9{;&SU7>#3VHT+p>WwS*{LQ
z*C8jcdX{TM=8edK$C21DlFw``mxI|HtjW9IC#?Mm58~bJHtlko_JXrxxd*i41KNYf
zP4o{<^mNF<q@tnyJLn$)PnbAN>)x0nlZn-&Bhz$b0r3?UB!1h6Bz@?6S-^5JSUm=h
z28qS8RulXdLJ^5o(ms>l`dK7@7OBAsEGg*ZFJU#P2%!O$BQyTneez#?g-l-|J6MtB
z60l){P)GTAdOrS{4@qCJ_7|a!3TTG{+MU-)Gws_<PXQfFRTz^Sv`{&m=-?2eaEMU?
z7n`==(2O%l&p0xW$*5&AnlJ#3FmZ^_Nb(t(foNh`I}FO0u6EaPex|nhfPRP2DyE!x
zhbX1AN`?42rS+V0;tl>4rS(dPTPdwp%857VPD-nj$B<Zw?@F6)$iz_K^rz^5TWwye
zSF!w6tOt7`83G%A<g2zhvG1G5DR(B5&4(c45aa~5Kj(`3^4S(<rX)4gB7<6F4Vz`j
z(%PxC8OTCxmp09z9iUDjwoa=z(#D{7=<;c?theVydqOEaq4arWy`_}iQu@4ddMKqH
zN*|Q*^Hw{B>Vx(DRh=bmsST+D%6Xy_>BRud4}~f@zFs8h6&|1TdH{9=$4X|9-VMOO
zO1BTIrk4G#MP`V?E>YNL{QSjtt~s69OCln6N#rq9e+pvRodL)|iezNei7a7>{;JQY
zONhIXQMV9(LI$6ZInN)BwWF~yzi=7WDZ?gElJy^_`bVtK2Z~Ot-pOOI52AeOg#|BI
z18vzr+e3N7a*>R2Bx4EJ6aCmak{JScU4T!q&2-mE#2iAANhoh8EVmPx>=fcKWD+LC
zyO7B)9<RS$SHF=iEm7!?k(zbrd*Q`$uaM>|A#OyPjY52z9(0?w<o&yV9#kO259mP;
zgt&?JXrjHr4q2|5_GlL34%(xG$6#heRni+9UR5#3GzM`{wDz|#*4=#N&T`RsP&9Uj
zGuk}W$*bb5y9=??WJW2O84U3x<k}RQJvCRhk^w6Rt5IOUpkiWv((52tc*pTSH%>TJ
zaF>{05v5Z^4dOlJA!YcGvIHC&m$~)EQ|~=w;xp91Gn6&V05?fwhV59g9UFjMiDmam
zVCj4P@s8#L-`vcJH&Yrpls=?3mPBcdTuL9pHp>-IMg^1wcy&LvDxe&M*s*|e<MCa}
z>MrE~qO#;>$n`EYT!?+{QUN@!qO7VY2R;Z?QLa_ga3S`oq5^o_L|HXa4sd!bIXu@U
zYPb;lG*JOO{z_SVrR*Tgv0Ob;sTWqvuaU}YA#Ok_4MO||sk{;5w@Br!5H}+IMq~<#
zX1RAr<(+W+C02dOH)Z6uf;K3m&3LQ6OB>u3;(N5gJt4kN8{8M-BHExxh#$~a4`>H|
zzGB*-SUBE7ySC87d9V0DyM7SjR@${yh}&q_HX&}OUE77YgC5yI`||UBq+LJq<NM}B
zpIbg*`axoW8Q3TT+ro_&%e}`o@A;dJ163Q`28u2=_dAECl^r+-lo7gO+4Uxvc7JPc
zT#jrZ^RZkdGOk3naDUO?Y?)P~A$+sdgH(Hjmk(c%>K7pn#wNj7P=17Ar7*rPWVvvx
z9xl8`iN)Hn_<K#jaw%9lMTj%7N(MII_nU=PvakX4YW)qY#&v806<B`*tCEKepp)vq
z;IYcb&V0F3iIpp{K9upqu(0|&Y|Kag_t@Y)w&rmQHfRyzZfw$x?fCJ}*x<8pyce7F
z3dj4fL7#9uhE|H9_4xHK(n=R;JxITB$3>!;^pY74Dk>=V+rrX0!^y!GQ1S(oI)nz8
ziF`wP=M2f}V2R4EFuOPD<QrMEau#h2YaFT@eAD|z$~!U;&1gq6W)N8qb<dv@qNHCz
zvQq`NtHACscRvoOz=1qINo$`J7B44h*OT-pC}s}d_dQv*bE_>e*Irt^mp0<<I+9V3
zWQ-tK!<8Q?W2Lv@u-*P1$hU;@=L6bf%Ih&TlE*J7uNTxvK7iFzzV(zpkJ~8UHp-t5
zSY4D?7c~-a{O7FDr&@arhz|B4@;e1087x3bK<VW>FHtlyjYb^YYWLS9)}4I*N*Vpk
zU{roj-2ZpaXVzb5CIuyrC5-zK`5gsI6W>7O_dCvT;wBZu&yOaIrC7fdTf)h*B*Up6
z;I)|X=5t^2T;(6hW-5_=C7J@>3l9clhE^nL72eHCpA8^!!LtE5`7R{s68iHOH0BHX
zi4VlPu+1+1a#1XMT!4u#HEq14wWs~`f4&`q%U;rWOWz`Rz1QRU_4sG7z00oWk8CP^
z?@KP`q8YVlzRv&`Cr#KlB>9GHd6|!6_2YObV6p5e17rsCCmtky$AG1;dzz@!@9Nr5
z)E9}BBe5>b(BD)kC-4|-f)qW{_YIhVgli<9A-ssIKx_rl;lu5DEPozrw94KMa_&{k
z$nfZR{h!r$!+FXt%$VdRviTy)t%&jlJ7&2W%C$z=?6**^EqwEy=f8VUbIHnla`H8l
ze+@N}_n<mzN*xu*;}$BQg_;B&M4Dr2Y6mqPa9&V#`bzJoDZ~iFkV_aE0cL&s)TG}A
zj+Ng*wD|uH26ws^{gk!(<}otEY0B(0<po;2+ZY=h|JJosLUTtM_9%ai(|=9H#tU%*
z!zKuEBEu#M@iB%yCd5e$n<T`?8TPmkr!Z`a5T9V!6GEKIu&F|Pl3`B@@hOHqCB$hA
zn<m7k8TK@fA9`ABo-$GY@P9u3M}f<<Kdp;5BC7MIy~%%Akp7fDSsQ3$T=*wB+;Sve
zjx>3@t3>jZND~4@e`jV9g`J`J?(fVj5_t^obi_ORAL-5}*yJs|7I~R8so1O$Pl0R*
zxAMdfq;GU^g5*vB$!7}v;3;CCA{~DH3t0XF)_^z!u1WHO^g)lG=>P65S#p<dT83@-
zPA-EUoIwwTT*s0Zh+#KqFTSlVqm{~NExxeprHy-OD@af;Uj=(ds+@gB6uAqz>_T2(
z#dU5EM|YpmrvCHM)N7xxjc8D^C*~hQ+lSCYA&`+~g0|nz<JUu&k8(;{dkJGBt<XrT
zflU43Uf}?v21yfYA>!-pNYXB>iVtG-gTi}i=^H0#v!Ns{_r&H26UpXwBA1=Ki@veV
z&TukF@4vnXrnQ1;JCG4xMTty1uw;kuc3%263R5@9pE-cIvk^q5ht%+glrJBD>L`yo
zYAhc{Iw+40YAl3}=HA(#CdBO=_HX>rsiBN{#yZNdjxy(Q3#HRS8N(O*<BwqnWeyl#
zXNjoWuw)xwa*Abdv>+<PTC!p%l5`4lT@M=5Bg}PS*d|Pv>!dHaFcG}uk`>#Kq)nJz
zrSG}G*nTuS&*7|z*?+d3-rm$ahnmA5$+@H>`E+Cgbv$_?UcSJGIJwUm#edX4{253=
z=>-=~B3K-{S`oMqlocW<C)TX~-<r7ofZ-~Q8ssnH0vvbFaelOCHS6U)!gZLZo4dzw
pXBUy@cMj2T{-ZPh(T)E&g8%5wfAru#y6_*p`Hx=w$6@@({{v41pf~^k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_wmma_cshuffle.hpp.5BE3B8B060B3F7DD.idx b/.cache/clangd/index/device_gemm_multiple_d_wmma_cshuffle.hpp.5BE3B8B060B3F7DD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1960392bb3681072584d40c109a23370ae572dbe
GIT binary patch
literal 15240
zcmdsdcT`kI^glE69>Vh6T{<i*ML<LZrHC{ePem-jj%bLAU1K-VXrf}lL?gv0b_J}#
zZcw8lf>JCXq9C9VMUdF|DyU$v{bt#{&dWc)?>{-mbH1MAoBeR_o0)g#wz)Gy!us@C
z7Nn;qP6%5#W8(aAoSvSZG5tSp{@khA6oi=T=`GxTK4SE&hyG@b;Z@@r*HtK!yZEXM
zWod(q4)-hlA$`DH*@$_^zumvdx^U?j*RPiEP>%gRa*yo!s~y{l0zM8XKHc%cUor0|
zC=->VE(JAMyEZI1_vTT=0w+7`hn0PTEn+h+B!&z1TdIB>;QYgobAo@DzeZ=R8dvp$
z#r!idM=YP!JaW-9y5K$U59R6cdnP;8*1MvW$}0=nKiG*}*XJeP_C8pAXrkHQi-M=#
zex{h|AXo~){cy97IBHejgWFqYlw~jL)9uHx6<b}a$ZE@k%2Aj8wg{F*zxb}_(%R6o
z$FeHO%K<%CY&tdGCgVibXrt%fe|w{J$Z3n`p;jjAzMFpBZ)DM_#hkwLu8hnXlR|2P
zC;r}bj>#{U55lL*b1ukD9AobHOkY1eW>(O`*tNZi6yEB&E>YQCn_{M)Tif;=-&C@4
zxia=d)V-}AeHMDxj#o{qJ9zJoQ9QTuVA@olpFRZ!eVZ3D;@dU1n*Pf6j$85dr|XvD
z{aruy{c4iR@9%fXXRr4!I=^DT^zZ84Y<qsH<mBWrtG8xGl~w*Tv+nSB=gY_avHQrT
zvqyGMi~3uY*}kjAuN%Y9Hf%8R{ptH@YhKl@Ig;xYFk#~8skb|jIbmes&ZUv3&)vW0
zeC*zgli23ni^rqi7*0K%c67C)o1I+r{XOjJ-xuef>^&*iBsRUd+sNY62CjF5XF1LG
z+17R8CDR)->l{9Pe=w$TMzZsdqrYza^Wmg(Z%_4netXBh^r)uDO|_Sz279dhwEV2X
z$NS3O;*0ytgK{UOR_=~CV|Fh6^!mQ%N_$+Z-L4*Rd9`)tVbd)QBlf+>npWpj8FlAe
z+Hlvq-%syYczgIXZu<Ru;|Coyeys=_URLWFS8;Llid|gR=tp0zP;BVg!)c+u=McG9
z<Ne~<CW#aE()#9^{MIY4wwv?d<*{A1H;-4^sTzi_Z!lNQI23CBY;f+PVWZ-!)-6i#
zuIOOsJaNrWbNU(o*>2&HIYl={nvHmxR8*XE=iagbTNXM7tQ*m8mG9EjsjUV^`PQXQ
zRSbMLWl`+Dzo)3q=N`=}9@fDmFWNJ^W4xDn(VFNF&-^F&UjDf;W_IZ}r6Kd5wf{b9
z-`dcIxZ*?AX^Qjjdsg;Hw*Fz&3E%DBrFI_f8QZq4-rVQ+vSEIAqSFJ8Sxjm*FeYq8
zhJVs=&**|%!~Tv;o3|q^@XA}mHYv-ywl(gzcF9@;{SVC#Hjg_w`u85?;Wj7J6ZZZx
zq55v_&OPYB$M)r~vWJAbhEF<P;<xmsKI#7K`PgHvDudeOzZ~?s#wxSnrhCM+%4>mD
z%FsXC6z??}Zgzf6!_kms$<q`|Cwxo|^qy@yS-twfA-?{X<XQ5y_hZJJBt95AsA}P~
zo8uEJ!w;Q^bC^A)hW|wo*Ob;_Rn9L@J8aw)cLYrep0xC+b<nBBt*e(gC!X#$WLZYy
zk|mMD+z+O1-qp$M8!`FpyuW^av?0(vsb*mNMJI=>Saa#eA;*_4&x-6d?!%C-J%_1I
zBv%!@?RtLa&KqT~xHBtzoe#0!Y<$0a*ui-nE`PLX?q0pO<hP&p%z74f_3J?^hRCY>
z-m@FxF|}aNrLyZ+UuE`t6BNGe{nPK&cKJ8MxchIL4+afc8h7!)L)$mk{N5`(d;L%L
zQ{QO+rJYKrMIMpmMO@pSQQX~7|Jo6~c|XjE4mHrzLvAq>&EJRBDY1vYp4dp_jl@{j
z=uxufAGue(V|S}b5hmEhPfvV_lrK>`T>~tJO&-lp^bXl7v68Z!<&|f7wXOjcnJm%v
zN`isH1gj{^3QoC#bJ8`RSrpS3{do7?YjZ54EK^id+*%$$Zk<unG%4f29TVJ^k}Hw2
z61nOcV6jw9A7xwG{K*KnqF-}>$PW+;UGtgcedcAO{so*lHle)?RSmUmc>r0pc)&Ba
zeY@<9us$V!M6!>_RM!BDI`3Mrx?<8l`nVNk$-qV#*i_d5i{1W1V^g<wHZsTVl;s=M
zHx4ZiAUpOP9N;iwmzz26K*@vEgI!x5Kz34yl^<i{WaikGl1HmXw`q9*x$U0|uAh0n
zIno?kP;!JS!nWlB<aUc)zMI=_Zf7NSqT~dgOW-Ya4X}6^s?&#sPWi?h8&H<%s_D)x
z4<LKDezAT3yw^{RaBE5)BpYPX^7u?Xws(w0^7w3hEOgcrd$sD-uH~U;h!NEmBqIL-
ze>@c{dgsb<Ko>pnS7P}qX{T$Tk8V>uWcrGPa1-oGhd9nFkMlOV23SJgtxswDW^<D{
zCX{8GYMOJ)1IU9WI*pk7whqa0M>@zhVz-UB>l$DgIlHX#RhP9P=C~j2VTNjkYs&-3
zqi&x%_($REW_@f;$@_@WK4PM4fMwhno6NH-JM_?6M;ei*QR~n)z%nz{e$v3GyEo*x
zKb^-)V!e{I(KWy_Ywwn-<w5V(m}3VzP?Rdlq2&SOIU7&eW)|-K)&jSpJ!E2oOsv#3
zz_Q?GY*Ktm(eLKioc1t6HNv*#0p!JzZi-D4>_*A49qnNgQEnn;x&~O5^mEVMNBEv{
z97tK>i4ac$bq%n@ypQ_r<<lFray)>ttRmK{=$B|4V2N!zFs5v0(O)Lmk<Md|a*k8W
z1IY2iN5)Ldh+Qnl?P(7wM43X&bq%m24m@ywfctr(Y;gmRIm5@Ct*C8)Ws~gW!M3gs
zo8{P-_OgRG>>wR=4X|u`UsZp4PHdzJZbRo$$0_S*FWLrJc5<WU`-Xa6F~>Y*nXH=Z
z*75-Io?T=7X0LNuCCB!3kYu7vrc>26z_RzU_r{urS-p&~g7y%DWii-N*8t0*QM(de
zeVTDej(un^8;H*a(p%R6%MlZ|6`kfTEtBI;lw}PuSwqyi<}=H;?H?B;7XBc|!IUMD
z2#F*>*8s~Ix9CNebNma;u^FAqaMf_zmIsi}k2UGPGo@mqIrgOFF{&~4Ee{}P8SEW2
zV%6}W=9r`8ajJ1{Ee{~)UivuFeOt|HIsS@{znYk?Cbqf;SS}A9lc$a_iZREAw1@Gk
z@op^-AYYyQ(%IvaONkLSrsS1awi27^8eqBpbZF-R#YbnEV@t|1QZ>@H<pJbkyJ(9C
z*1u7o;Z5hTmYA%iV$wFiQo3r~&jYfaddjgIW!X%Wo2iy-8(=BFppFiETyJKMt>_?;
zsz}?G$7k}V&{lCTI~<W?HSHmlC{yX8Ya3v>+tIhO@$?@(jj$2z;S-X5qW)jo0L#6G
zvNVes_8-i#DP<X^8fM$_0P^Ev&k2Wrm4A|BFFMF}Vz-?-CT#;OPrdy<<VJlgHo|h+
z!!j&ehOKlBu+%*`@;u4-r%OuQj<Tflj_JIQt^t-8v)}kPyT(%$4x$6CBNppOTU`Sz
zZ<>}jcWHCtlnJ(>y?h|552UlM0hWK>8b26p{D;0A52U@s68BijqHTcXU9PO~g0W|m
z33jHv>_^J|$Vt}#i{9LcljcQvuzUi^bXf#=Su%HVP{5_y?nn?<b3<2iqaY!G^a9It
zNrksAz6ew#KFe@_ke2?r&JMx7kDTWO@mFN}EAj+yB#7IQ>2@i;h)gdcPnfqL-bAK1
z8HQ|y&QY7Hz=$v<O3)@Kp9#9mgw9cumcX9{aeCnNFo>o`FAJ?xqb+-5LCoY^XYxuI
zM-a35*4YfV?)3B4X&b%n8Vll3`=KtDx{O2+huIHv0qqjR;r7E_!0rWcg#8E?2m^u`
zVIScFd4M2}v>)jL_9cjs_K_}N2!c4;Wwbja6oR-5wc8~X)GpM07wXLfnS%UNP&X;=
zmx2Z{T#DM2GC>MrDe7K|dNV<nA^$SeO^W-Kp+O8cBcU0Yz>-5o$IEni924wY<oFi(
z1EWCwu5%0y0+%d^b=a;BcZSd@h%d0)3)}-R<bgz{%l<$M1sc7r(ue~Dc0v4yxc@_Z
zAPt7Kpfl2>hHO47k=IJ(!vt{_d7VW*?87oh=M2(?b(=*xXEAJ5cHH3o{?pzDoS1?<
zQrHS`Vj1!%V-n@W3=)t*0-+;L%pw6<|HhG7WE7)klYnd~{U2ia53z#p;l!t$(^JkF
z!Y-t@oJ^P0vM*eR4c1{}@Z<uGQ3h*>F@!HcEFggeMEsEbx^=bH#YZ;py16f!`^<ux
z^F==LB0n0G9kSzAGF?v0M$o1<z#!12aXLWUVXXvebM04f-e80RoqCt2Tz8ndAin1m
z?>S3|{ZyHGhh2O-$af*<rfaTAxY@ALhS3U$fC80w_cErj_VSp3UknPW=x3ZDHYZ3s
z5G<tVtz|8#IqU5uqHby+Q_OlK)=O~%5*wtr8JRXCHRI31Zdup|6jc!Ou}ePg$Z!o-
z)nI3aFLA0%oHG;33$F7Et~(P#BiFf6ir4d<*7JcN6hYj=ciO_R<Jq}Wo-YfIMI8NJ
zixb2ezPC;H^fL|qH)LO2L%~;3{n6}vnBe13&o~suEHDWLCZQ09ccH<%P$a`SsAmoe
zV-{L~0t-+G!=-3&DT)LP&Xjt1&5nu<n@u`rGeZWs(g>m5Spj4Fs=eRW=Y-0s^N&Ra
zu}I#sb*2wN%s>Vi3`4Y_7O2^H!TU63AoEOQ5AzYk<JjOhHevQy!YfL6YcMm<@Cl*b
z2N$g6IdL=BV>1^58p4UmT#sZaKFswvEX60e9w()^lp9pa4T1N)YIh#78JjzcD#9A%
zxCXTWiE?5R@=rnm3~xg2n@|Uaw;;PM$e!UW<d}upz!!010rD?E0Sp%+_d?Wx;UZ*L
z#3TT_P(xYE?vn}XEwX%z?3qP9#Hxq54YQbPtg4pcI&4;lZ5jOqwts<L89qR)4-iM@
zh6+fh0^$#jP7n)8r$UCk`_`<{w~q8{5xjacX9t4Q2wt5m#fLfdVJSYzsZUCADW@)F
z7(S3{pk_}U=KTHI#IxcOK^)Y6P=64Hf5@hgKbjpCUwn07kIxHY36_^&dnV84Sos_~
zGW-U+y}^MDzr$AVup3~Rt}*~&A;2egNadx2Z>S-~Bb#{S2l^z4xyU9L`GLL!+SOm3
zu0O6`;KW>B$mQi=8l0HN3wctU&s*p7HjI9mx4taJS9t3yQhb%SzADAnc#~_q8GFBg
zw<_RO3>WfNg;IQ-x4h0wl@o99!VT$sig}@!VW@>rpQ0^<fF?i%geD`J@(27`5PwGI
zKQlw4ffO4Ykh+SqM17XHf@`PgFR!@C=F{s+($ithPJUFvb;PBP_%X(M;!@9W=&kKR
z+eY=CXD5gYycPuMt?3qOXkR&&qOb)+8Cv4CL_%8<wnRdUL&`X$VoMl@yyH+X!wE>4
zAf+cFWug@CM9Q82#@;)n@24R96y(PGOGV05DgSPy+%3h2k=J1qG__atx`v5IR|w)^
z6m}R5WjGyqrAz7QC@ft{&q2x@qyiNb#2n<EgMt~(L&`iUJs&CarMLtsOa6_$OQi3Y
zA^S4q#`-Hq%5o`x1yWW>aW(R-M%^Hf7Q`ClTO-Afk?&(Eeu8|TFg!54i{<6sn+MbE
zXCrpnh}%N6f#Mofk-J!a7uzsv)mGZTT8A3;K0jlkGL8szD`j|=SAbDy_CPL~ygS3$
zylXb^0eHl?jgf~^?grAK@A0PhcnctmG|l%Kd&J5`pQ@SJf=&`OGg~4-hmqM~)JeK<
zX6aH~jV!B?4-?E|Wciq3s0`}MS}KCfvVO4q!I^2(GGw_7N$^o*c@#Ob)j5SMPf2kd
zQq&<^7!>v#G}3E}i)@(Vyz6n^6AWJvPw@_?cvsl-Yp~lXz7ycDgSVNc@{ebJai<@3
zI^f%{+d4bZ9e4&-XJ9u_1A%V#ts03lY&!&6(sO#ub!54FJZ~1yJ3#t0rfZk2mq~dT
zUHv@boJaaXv>p5K=gIo_9V+PvVkfaXlUO!yoXtCcl4y3+eY5!phA;AcFER`sjk;8A
zwHDSQ+B@u=>A1{4>A)9w+Y7uC^e}l{#A0={&7V}`<A^$rghRnp5Q~UM5g7wBgnBP!
zypHA9u_ZGNZSfam0##qSAJLS2*{4h}oS=Yunn3r1-BZ{Oe766KIYV#Fx1`_Sh&nc+
zZYe|c;uF#nsndOgy&mBn`Z0To)fRP=1hE!-)k<*=*D;6dX6llB$M+bwY7VEj<@ZkS
zZcxL_<c%|VJNQt|hqu1KqLl`FU*Nj~9X^MO?%(@0C_d2qL{orf-;+g#XORdH`NB(8
zHH*07LPLS>lYC#I0G6oLBj0){Za}^bQk;b?vY4yUY>h1PrT7Kc<^|^}-TJg?l;ZWg
zdOhC;bWF4LQEy=wisW?RDzLl)o4}XKbft1wI9Pcqy4zTO8wUd;Ef!$qKe!uUnXZ5i
zFD~ud{LT7K9fwop>gC<5H%JKT>U7{DEH9F3jJCQC?=Q#Y-QESZHRm|GCIVe1SQ8EQ
zqRYf^HXlkYfZ>aLD765<D|atlH1I%fK3#+3$nH3DWiD$qXR}(mom|ZYujYC~4l9UB
zoOKe{hT$EY-44!`;e(v@K`H$NXLmwM&*SX!r1%zReT(sfYo#`!IXnOf3A7mD{1J75
z>VzOZ#Lf?K7ZCK?eq?1`{3rj8f;iZGur2J(1#zbR%(mdmDQY*fqXc=)^k-;}^Bi{=
zT@dGZ&-uKMp{U24P)1vWOx7T4hW8`q{m7k3a6j_i&vr2aErj;ik3tze3z=jgYlbUP
zyGqm%f`ULRU4E6Qs}%RDMEx0li2`4uo>Hp`te4^j6xble7qRh0Yz3kc#7o%tk`(7)
z;~Xi@#m2c(oQI9`q&Oej=3^(;-(_rkSxUcxjju@YRa!KX;!VV06H$P*3A93Qu$d^J
zcrVaH7K3D>fb$pX0Es~gQNYoWK<n`asYFq(E7j9jNK8(#I)xzCbLx7|nR$T*PTe5I
zuQ>H9DSpkVUrX^DPW?uT8@UdRoDb`-iBmU8=^J>f4ZJ<`DjRvLjZ&P%TO~>HCf;h3
z6mRCOHcRmq-gOJ_$@<&MTWw|ZwNHE9o-%3f`~Q4<`D??O>=&Z_qyM~Mu&#Q0xX04S
ziPXJ`Jw=h(bT%KE&FX@%z?AJimfy#wpkXrI2_bmFgpx5_^<H%;W!ym28;Ca?ve9ZY
z2`(jrAreD!K_iv+01<{>KjZwTiN#S}sUMzcFw+d2z{aZfSv@lMJf(lm;H@%vXK<Gr
zo5n^hSJ;-)pqoZK(ntWrG7avZ#$uZwHW8~PX`0hS{F+Dzq&-Pj`tH9KyQBl1`YzsV
z7h6?LMrU@4w}*@ljyNgfJuJV6)l3-L^G*=P)~AKj6*sKHDPsarCy=j>=*|iSv6y%i
zlW#y6pubcY+H+742AqS^u&p@+g+B}8w7_X$u*2LLazEH^aR1h=Il5OAZnC|p2CHuJ
zwje4_yu}N*r1TPADET*5l}PzZd7)HFzs(D`|BY>K^J><A884Jc>E*mo{%_o&ocCn>
z6}(U(rQhL&I}Gm;@2W#CHgEf1R|NbKWvs;VN~wBkPlaJc;0S_-Fzt~rlV}1qNWkW>
zhZ5*noWVL`4##u{It>2td2-S#s(!0D#VXDbg6i)PH@nGBPxqiZo;YL^hrC$q*@<j+
zN^vr>NtWUqWRruuSga{QHYHMg8`<1u7z~N(c{!GsV+YVW>Qu2%kL?(4!fs7C2(V0d
zrVTPUwD251Xhv!uCDlq~lYpGqj1y4%1mw+dBC<)8(i4$KBI?ZOsmLZ(N>4@YQ>E|k
zMmD>p^xep3Hwt3!=OLRs<isS9huY^MZ-(=cO}><#k38~GXGSkaHsw-!Ici@neZK<P
zR7mL+$fp7Y!TX1O0te;p7`usT%|2qgkGR4<L!f81X6ek};TVOA{w9{+#MU5MO{yb2
zWI3TAzQS&=urJVMy1N1(da!?*oR(pE8CxSk{D74oa98%C_96kicy#XSX<NDudrb>=
z1;ng?v<1(hDR7z<O7V4Kc3p~Z5VIRnTuRJJ8BTAPeX8X8>3LKJu|yq9+QW&mAYLVI
zSBW1Cbn@Avla^~5(!Tf+!zsTLn->>9?@JFsGI-MrR(pWcR_ZG?S1g!iXOQ6;EDW5L
z8{EhnSu^neoVPp|>}Xu`bK^R?lr>nW!4{xQG^^x<OPmF`tc=U0k%xm`J^JEH2^U{z
ziBxmzg{@c;a!X>ztb(`+d2EvUl>+2ez>Zz14Iqy~X7`!CLniROQ=d^|iS8WTm7N~W
z_nR*{9Zy4kY0}f~D&$wis%wI{p7UDI1%S{l^!uevUE%DDt!SrV(O<DnCk&V{m=y%(
z_nRNi{M+Ioiz66HL;ccFICvjHtU~>(&>*&Vs6|6-(MZ5B0##?togXHgX|B`!AoORy
zHF(p`+v^nO#B$!DoOfiQ^$y?m4(|&1(t$xUTlX<oNtNUOpWNg$P8xo6ZdNi4bRiu=
zLSb0AWklaE#`0q3@?^TZM&POPpKM*W<FQ|{At&bZLOu)ToOqcRE=%zhUbrH~S9#&8
z6kp?oYf@am3k6bK$P0y1e4Q7rOK}k|6fq2IMCGZu0R<mMZ-KZ@_k(ji2>-@1-32Jd
zR*xO)u{*P=S6J~%8VQ=QS+g_<f5c`Vr8t%t#}XUnUgL;y9I=6Ng(gEZP9QcAAvEc;
z@djc8QC|=b5#=GGV##ejG0P`*;GQ+^T78wcGF(I)i-<eJ#l*2#ic3kSQWD7M&xqog
zbf@-!Q#|0TLH)1I{eAGe>KXI?&v{L_c|%ig&9ydI@1o$gKMo1EP5R<K60X+K+~XFO
z-;!!vH5RIwAqwJa?DkrkduT7<fngTk{w~7ZA#W-5ijByj5qYxW<T7lt47<TiElshq
z%_H1NTC8kSE5$jSLk{N&#mdr-&+q;*eC%)3W4uO+*X+6kR1K)lxP#?)q%}M3#Xp!@
z`KmH2{qIz_|MT4~T%@8luenYLi%Sd8WOz1937}j-C;9-(A4s<k+WUnt(TWWz)yFDQ
zmQZ(_$$Mn--N3zaw4~iJo9_b{E+W#?1MT%gHgp^|h?CybSxXGo5@T3qO+FV`$ZqPu
zg)(}MuDPqox>*plV2tE=W!RDt46Q<5tEAVR4x?d*(Fo~rOH?`<!|*xebxwL6ss{C~
zLH!^Q3E~sf{R!#`xXQmsY&v9xzPM@BSnchQP^<9N=oj(4L%f-Dyh}XqGF+3LRlY6R
z(ur#5N^H3jdx4vO)QNxHN@2N{_7Kk-#`C7Eu#?Ukrt_w-<^Q|GnZrKngQ>KKa$;Fd
zJYXoeFG_p(3(Nmv1#g+|)+neBSQp)SYOjp4rHsc0@z?_9MDIrugI|dSoY6lsI24Zq
zk0nsMozQtg4+vv|xHxEWAC_DkM_rGjzAU+@MM1Tw53HFW{>FLz#sz~rhwq|ZPkTER
zW>fd>>$d~9#dgu8Iabm3kc7~&ttQyBvFBUQcVJ^Lwp;ANP#UsMLk?^vRmi#uIk1`3
zBD-4T0vI;fbpKq6<)zYy`2q_sq*p)cacDh`U=g?xyEWnfhCgAePq-amneH|%EI`Bd
zGZ#7@b9DaVYWqt0<oSwU!hQeqA|c!tri-e*EDVd<_*32zi+9<I|9r&8*swpHuby~A
zW%^~W8$nt{H`X_>{D!om{RAtYU=`B}?d@pL3Q$vOmfEY)ATxnFZs$*^3wUxttj5mO
z>{R33yL}fr{+0BadW4PGU?X+}`~7(S?!#ezbt5R@0^joj-w$E}y=DFFxL)*9S+7Gz
z56xaIjX+R&gp*l4m=&Bw>A`HERG|kYig|O#&)crKrPf{Z(EjTUKlJpvvY(;sXV1?+
zu{gO$op5VM(7o1u!$Q0F=-o4<SD&ukdg+CHX3^`$e)eHMH6Nh+7Y0A{^!l=&JwE^Z
Ee>3<LV*mgE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle.hpp.A234525D7676EE3C.idx b/.cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle.hpp.A234525D7676EE3C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a073ccd3090bcbe28bde3809d443470a27fceb98
GIT binary patch
literal 16830
zcmc(GXINE7)Gjl7&-Mu0&7t=uAShLuAPDx}+p%kmHBAF*EQt!XC`Qy+QPd!UASxhT
z#g3w&qM#yJD8`ClH)`+>?z5cz-5>Wp`Ss;_lEwS%_e@>0+DwM_>De>8i;hmWDShVu
zICai9ypE2J0sS{(PWa4fK}V;~UPovC&U2$i&#DYm{~fVpg>8Q4VYx3h`>if*HR*iM
zfZ5N4xmP^~8}p9sv(@9($<`hp?)@@nu;r7+(!|mBY5TU%T6v~=_TLc~R^1HC=l(IC
z9g|mOJ|V*;Y-8Bi!l!n}yM}bFS?pk|=M?5zXXF&x{qp(#8Q;(EXWX!1(SV<qyZ0Y9
z%*HI`$hmg;$LHoIhMh{E8T(U<6-nPWzIBm)$#b#$;Z&FqGUeLa<HjqFl&<Ys6l66h
z_U9JMBA(|*FWnIk@vLEW;ox}V@W8ZA(P2Yp7CUwrIpA>D#h3Qwjp?0TJbB6PWj)u;
zy6f?*VSA0MrufkOZLY)4{aig#v$JkZa^j=8;pb-G9?;{?<k8MgX6h}rn3z~okUh>b
z)b;8__ZEJiGd}vgw7vby=oNjZZR6+L_zbFg*|^Xod|F)4hV>!G>h2Cb#M=xEp40Qy
zwwP=m+m-g?U&j1(?o8S9+@hZhVh8UYr{X;~jykpY@5I+<((4m*_FrE+Zd>s;4$ngT
z`@Fa#d>VDZF!Wyd2~9>`=GnR-D}UIMG9-OnKvD$%c;~RWcmG{>aIM?pXLbP<<q1b(
zE-vcUF(!6H*5i<$PToxLz4x+DQO|u5yWXCt={{ms*Q{0(w|kRc-0m;%%}ttX;&Jzy
zLy4htNkd57^oh4Ohr0eWwPNCeS1YR9mi9WgV8+;=r$>(+wqU}exPp6SkM(CCEqM9k
zplL0Z+$u}GH|zR=nCzM5=PT!a=Qm{Ki*@J6ZQFM_s@Emg^N(&4{q5<4_gs}~tY`f)
zV?G|RxP%M+ebnt8<0=R5b06I8ap60gTDvyWir*&A{Nv-|##gTg_{8+TXOs~*w$@{x
zgPBj}fSLTJQfGc&;m|cU{*z{@-_BUSyVDH!$l>Rg^;yQP8|>09Ex((e`Ge$bA0lI?
zkDi_Neur+sj%urmXT3cpq(4sHmg4_y!s_^u?o)rUd2)KXX22!)%<n8LI$U@r6o<Tj
zmNGLowtm9so?9Bq4leq0#`s%fKGaQ}*mqBxIr&wWN6cuvV|J-uAGcAC=ia(*wcq2p
zsnPbI!HMY}dy6M+=)OVY6g%8$LHf5Ef1mEdH)IwM$twST`JWxH%=KTjw_?w%{kONI
zuUwkFHmmH&;D-rE)DMzkb`I}xJw<g$2wywVr)zR>89(D{Yh%+U!=I}zixE3}>c!pJ
zGrq&iSpRL;*8kM?#;`L!9bH!B?w_}3S><+rr<pdSdwE)qjRB<vZKiapcH0|~qH(>k
z@mRamEq5Qkd!6&U&6uCZ_Pyy-^SHoUQ?<r=TfskLFXd)%&u`g&F79UVi|~5gw{P#*
zZ(Oh_xQC7Tkuwt?B?URZKapSC@y4ouH7&BTqsO~A-F-gm){`?CQ_EiEU++-qoxdf$
z)z<GV{M`>tO?^1);+tuIZ)uUf;Yi!Mc|Ul>Od7juON$JTZo`BnJ9i|nX!3TuWRw;e
zQ@<g0XX}c!1B<-J^?cD~ee4gdE3Ef?2v4|WxA2qq+TCLZCl{FwPOV#HxFilgzZ>x9
zV!vqXnJqeBa(z7|#d41DUUj*L!GpxQnUgxY{AttQCHBETQv$4x`Rw_l+aQ0Tu+go5
z-Qx~U7Qe-9b*ddG#x7Y_GOzr4Q&Xt@(!#{`ecH7s-TtCOS*G1Fi{}=z{u)(1Gqc6Z
zmKSQrqK4`B_ZtQGt9o|aQWv!_?zQ5q@XFnL%qb_|x;0j<p9c;+y|XB2Ymm#@i8VGU
z4|7|&??LnXMI~{0kJd+f<~`aF9h%qp{NG^Pje6<7N1fDn&S{L8;B8jpdv#=V|FHB`
zb;~x6{uJ2HuENG{%VA_Zq_xJgvEPk6!;1Y=0<VY%%*-W^k;6U5wb3l<a^<_Y1=deP
zR-2f=eIL6BxBEvo+_5O>f@AKRnO>X9=IvXUEX{j7{+~F*s-Uc`qefgWd9ttLtz|3v
zHt4v%Ta~ihFk<$g*G=OVk_&ext^f6c({%ewmLDfga(?&A-i5>L<_j%SQg@viF+_9a
zzWaf$jR`TPPe$;Yo(23iZ@J6!seF6G32T;b<tk=|)UQmeGhH+5>Q6bEf=iWdr`+2-
z`#A4@SmS!1n4m$YO@B0+f5_`rtFpEUS84}VzU}XlF~9N{*Uw=l-{Xf%N%Iz(`l<e%
zw8VtJQXe{E*s318@)jnNjHrx^uFE#Ii~iI|w#K_XKlvi0Va-23&h%R58tgQ8WWQYR
zvXY;~tO=8<r}X&i-=`1A$gDdjXWsJtZF#^qYcdBM>+I^uSvL0Orj7OrzqZ};f$Ng+
z<GF8~?AqroSnYnGP{;n}yZg`6!h8PPIs43rMKfa?oIdFqk1Kfk_V!;5m*zR0K7DiI
zaf44;U8Z+C-!dp-$(a=wLQ=OaE&A!aXkVWreR}3J+HCAr+ZsHx>-fkS&fUsp4eMLh
zaOHiUKkVZdPaGQF!1ej0Z))>lj84RYA14nKsmm+x)O*Z~Wi=9R-BU;NiKssjW3A9p
z(a-dsR(;X}OCDy;k?_<WI+`_{(HhQLD}Z7$cjbrB;_F)^+?lrW53&D;cxnYuEW)jW
zQ|sRUBjImpE5CCVzjLj%0w~rK3vw16NZTXf4z!i`#P&UDrxiePD2ord^tk6-33s8b
z?B?uub6#2jl-7A;yKWrcCs)FbloHRW<2e(p0E+XDC60F!eX9g)-CakspVQmV8EOSk
zT=mOsJKu~-mGEHN%2v*AD>qpyfZ{fwAgrzFw|mssm=5=dV0lDnqZL5$>Y7(l-}~3g
zhPXeaTq3rYNE@yAN-5ra$fW4(L^byArK3sZq*Ts8E51_R&e^~Gx1c6HEYRV;K)NrG
znN|SBZ_M;J>NnhWJ#0ujx<RBH#7ZlG5_r$`oYPc&q=!|sm1L}&jBT_6C?Q>jfB!e<
z+$7<?w3SBU-$=gI3ZR7b{CJc+L@!OS1#M*?vfPKPwE`%eTSr-0Xr%rU4xvl8jx$}y
zwb2TobUkpZWZR+diG(N6j+P3omI|%4;w$CqhkZp8aJalK^_)~s*F_dU>E@T@zdrEi
zUnT5ITX{~bpOf}l0hE3V-VAy&$EKetZb8SAk5&0ttrbA&zoX70^v%`-0&dk!N3#jF
z*o0iPLZ>B%jC6F628~#9C~myF$#n_$q3y(QW-*+rRzN$WjK1;iwD_#M9@eL$xl5$G
zq@z{<Wwd4Lvi+?SMoPFhr92^KPl$_F0A<`doiUSlbjeX;Te`kwe9JQ4S}TAu{`fH6
zUY}12YHUYaNfykK1zW8E%A{j8!58)KQq?r2t;7>`Jh9OVpnS8gN6$MREzgKL6T9kY
zTvRS@U;d-hlE*sJ!swL0Jv;E^i$3qEWf(-ec|rVMkZ-gC+L+Q^mF8I7<CY#)Q%X6J
z%89jB0A;$W+kSPRb7u(;pp>V?^eJ)H3ZTqhQC46zZ&zDG+>Q?RJ+FSxTWJMQ=DIeN
z&!`{qN{!o4$|s)x#G7fwSIPwU2TRWjHze#scPNgtjpN+3;w$Cg1~P5$y)X&;)3H1v
zrjO`YWC4_gc0E6voH>-Jhg(w0bs}A-=SUVnS!B^RxjfS2gdWzVW68w2nb=$_zEZA@
z2=AZUVw8m4X)FH{^}p1p$^s~hCk@RS`jefX3AUoGBq7TrdX8lAm2%Ez(5vygbquj9
zrPLyIEppWgpe)zz{c)N1y5pu;q#Z>GhEamGRsdzi!h%V{z@!2_Oep0Rk!}$OtpLi(
zANL<MRi*4xV>jB-PU5hW_-X}ER!vm3Z}b}XM#A>Al_sKYqBD{OP@?k_n!0WOIYPiK
zXe+r`FBcnW1yDA44Z4|gE9{VjgJ>%a#JYhVSXq3ftlMsR{Mcz~cLV7*?&55B(QT9k
zP&U`+|C&e8OT#Zm<Sef~%R6dh0hH|@CU2QG^s<A1_2^iNh+Yvf(F&mK(tl!hvi;Zu
z6KqZ?3CJ>mZh$PlQquPf?C@spUJ19S!+l3=-%)=gi?5V=^O1Kv0=w#AiMFx_>Fz;>
zS^<>Zv-|zl@qXwTx&f4uh;<XG-IWDU{@U90Zs&;di4vYdC%u~USj~0N3ZTS`gZrLP
z4|^ivP)dpA%woB=S^<=#!B<aIB>Ruo!!4*MN#Udv&QvRavj2MM_}-siEz!efbS&?X
z?mJ|v6+lV(v_4|M(g(p3_Mjc@<<xtr`;-Mx4rmqzZ@SrUhaq;RtyJ*p3TmTd0hH7w
z7c=K=Y;jtRy=W_+aH~()P%D6P=+VHxj<uXpVS>M*tt2DMWNLS10hF`_k@-h|sQpC`
z8_*4?CQ>yC(h8ttx}ELzd)LX<)CAF1ULoCA$XqLcl5Ks(GUAeNwjp++lvJcnr5(uv
zC^^xiew%~B-ct`tJ1QkoDLwYG07{;Ny6R2unV+a-psl<_x-Y2{mIYAqw=CI@bNy~g
z*pE*7PtN*J&POYNay)N;gEY!#8g&xQW5K$qI6y0aa_U9RnnkG%-PPEHjwMU5%o3cn
z0w`xQZXT>2;4?zPlW9k*IEz&@mXZZf&iQV%S`_=RqlANLE02lkV`}SU0hGeYZVncf
zTTCR}kG8Uf6Sr`|S^<=zfCG;&k2-2cZ5?f;j!1RXz{}z*<@WG3Pf!2$lI{_m(P6B6
z82f7lP)eivwn|-E;7215+DavnDya#Q#aGI~M;SjZSxSR82RfDnPMtutP8LA9Ho<OS
zp!>lVhS-L7^qRAL%{glYP|BCBiuH6jF;2n(w3R<N(?4kBAq${XT=U#)Gl8eAxX@M#
zkh*{>gDik@=h^!qg^nc)Ot2-TBqGa1Y8hk!lseD-mlm|%vq{1oX-Atl>rHg0WdW22
zeP_6Od(Js$h}+UuN|Cyhb|ed+JU;N<$KiSjD`?QxJR_w0nyQp6fKu-~HsWaGi$6@T
zC#A$An|RuaEP(PdefJO99~Z5Za3{LDuZh`f;;I!udAE6Gr}C}BcM={*hr6CLT~F6V
z7C`y<{Wl|@RJtxO#FJ=8#|8elV5Aj5`5bcYUZ457zZqh0O1UUlU8He`EPxU|by~zv
z9yDjrW)+;SHkSZ{ENb=&(q6$39G9p`6YSCiM~G>#ho0e!4ad)7QImo_Q}9rTT18D2
z@u?yszv@zBiADTmvmTU~NLnQlHH18(CYiKKX1JC2@3G%)ZeK}6YTfM;kTV#5k%p~y
z`;aq4w6q~)mx!EU`$g)G?UIl)cxq8okHmUpxLE6!AuXaKxP+xk*bHJ#U2Sp%4zbmA
z%MLRSSfAv0&2b_gClU;m*PJBcNhLl_#M2BzR)ux7xfEN)ctMI6tYH<+Z>Fn0CFS_g
z9vV5bM$WveHhAVWH+aVzyc0x~yk@r$uv_Q=KZK-=&Y=iPMantI85+#N>SL(e!tkf2
zi0%s+okGUK93W?-W0Vs%HpagM=`X<&y0ViZgI24Gt7!|FoK+_0>ZMJ5>>j&hx1?z~
z{ottMsQV@5r2lcq51;XDX~;^%Q@&{?X90Xj0qK}4uvCF<*t*O4AdI<nSZ3bKz^l6d
zc~u~Tq&?im(rs+SdXUpe=mGLbURRqvLM|cF_|@e-?+Kx;sM$sAcM%tew3>0-UBnwO
zWSg9*DaTSd4uKv;&0DN~i$ekHYI9H+LA&aG^In|oCDFwz6;!2y1j`4>Dbm%Zq_8h;
zRU1m(+z;)gU6l}{65<Fss;G$;?4kuP*f3GESuou!v;piMlUuSlz4Pq<{6pw~8Ur~4
zhM~~Ok`Co04AKZubDya16LZL<V8`hk<h&Tx+e@1e(-!4yn63MMEbhk^5Oayt`m6rn
zOu#WfnoPA-&XL(ro)XnlVgl)k|FeLKABq|W({**`olRg>!5Yzy<pi2_yopn7;;g_i
zh??`9?Rl<!Q0vC+@kvANI@-Rt_`UGW3v*u6giV{UIRr?&<})_^ti+L=Yb4i>eUIWy
zqc{csAQ2BT8^LQ15%G``A0`%ui4~krUXw;F(v&!zSfnd)2C>Lc;v>Z92r*&5&m=~f
zN}NT^vWNxyo=wcMl{kl3<q#Y8J(pPJDsdk1&LckTdp_~bSK_0@<0$cB-;WWGV@g~=
z!~*4fPZ05h5}zdQCy6Kf{V5`zV&9>}K{Y_GZ!it`Q;_}?LO|WYE$HfM%NuN~Gf?jg
z)EB}^QImtha!?nB&!Vuis0&0I)U=}BMW`>sH&O4Ks4oNv)U%<mD%6GH`&`HSTnK1C
z^;?2ng5UrHqkc=U+s81JTd0F8!BPpf1zRC%uHzop@d$?Nv0FV3X81F9{fq-a)pWH*
z8n7*07MKijKe{!an!(ka-D=Jq(iBm1igP=~1%h%zHHWULT*_f<x=WCD363yya7*++
z%at5B&``;thEy)&uohwjDTcKmQU_+ZTX2CEn{nW7VF2LnMO&T>|KJ!v$6Uub*Kq-C
z6(4cVj~MQCeP{5FF+CzgQIn4R(orbez6|7-p~Tt9FI$OokYA1xpFw_SP$=8pv&iqP
z5??@m7nHaN`4uVg4dizNg|h9xiTrLVaV7GrRN^Y+SH&>wCS6mxyan3_`bu|2E^760
znistxzuRJwCa5NRkpC`i?mA#l2eS)3whf`SE($wF;Wm&Z4r|%t_>WW7>!`(#CoSWN
zJ{+p%n9DhlxG|heT#|_gV5pV}qUJJ|E@L|wq^Nm_)eo@)!_TqXbEUg`jm=-<w(NT(
z7aYm;XDTX}-e8U+zu7!G^{>i~bm-@Z(K%uYBLEjdCofmxm_juYRU`2OJ21-VeEUg<
z%$?~_(~(&^@@0CEjm)x@_zW^TgM8UOUqEITl=ucRyMcV!K2{>LN`|5KM_q?p_+ulA
z7Nlsw7F1Z&92C3`3SkfuL+nO3SS|;`#*WVk`e9rfCr4hSWfq%U+!j`aMyi~7J?8|+
zO{70`EaSbPB~kN{H~z@mLsboWqCLpfL>L;ZCjB7S5TWZyHNQ_6?%Cg^U2jCL8&L=N
z&>VdmBoll1Wi#%bOf(FqklrZ_Llu(lQ*#M2ZI|1LsrehSMa^-}={VO5EX-t|K1Yqd
zDafV!oJ6dXh$FN&d9g>!+neeP=m#g>$sGJN)GleWa^VtYBx+&>DOL!E#fH#>66NwG
zBR(OjC&U!C4FUr}S6jVg#C3wSPOyO%Ma@pZbf@6VaGcOPPS7xXNbou&gu<dz_aSu1
z5qd(}KRdAD=V909n9-$ujJzJB(7|HN;E?6L^J#R1+uz5XqwmMM@4fGENz_#1_SH(9
z#JeT&9T)1vdfY17G6rH#x_NR<753<7UT1r<_s6UBSSJ#LL}Ckz*}S-|l8B`e`y`Pt
zz!8sjy*y#LwdH?~0722{P~_Sz3}vpiXiLrYO)TBSR&eljwWV87>ABj{Ep-rb!Imxm
zYC&49T>kY!kM+WEw%i8<uLEqk=bx#rTgaD|v=nJ6)vp!>K{`hr2J)*@;s?m@ff66X
zrpMR_Ma>y(dPa#W`PP-Z3lzshO%30=Mv0qvn<l<Btc9pqL2Oo#uNW%NR1@WTGn?UK
zqIyhBprQlyXKF~~;xjBJ)oo4{%b7x{P}E%HY%g-|aEKOZtJ3s`axI$mkRV72|1&9f
zP7~UKivBvxH>^A@$e$jv<3#^BX$J}csWIJGxpWOffkma0k?Yp*L8P@g-&d#~l<k)E
z{xD&qx7P^T^d{7P6Y37cD10mQW5+foJWrKUq<RT`Hskh5%78JM^rA+FeNQ31sF492
zl|R&NO;yff+Jm!~b62nspdScp>*cT*%l9r>IAH&|^D0qu7+D-ft(il}M7Ej8h2amJ
z$p_9B>QbU6lDCZH9T{H7o2=s%-#7D?o0Z>_dCO!a&gM<B86OJwbPeUIJ&aJKF_P*z
zv4Wrpl1RD>a_t`05Q_G6!{t&vYhj-t?fah_?v^Ha!G=fo<(AjSd=6|+9X~RON7hh=
zqS-&POF?a*cH4|yQ;?Suw@*QV45uR3ROH3hEENT$qE1TOJr(t1xB$5pATQ=*3s67-
z>ZHWo3s66XOOb0S@?uW66a|!`PD<Rp6!l}c7ImyeT|hsZ4}HfvC4PW9K2YKUY*2vB
z!7PcI6WHK{5}(8dCzbdVHaMlkr?J6lB|d|#&tON^-dSvLR{1@iQ^j-oP=TQ3GETLZ
z(}$uw%>_7B0;gZ9EtgW~!5QYT+K5OCqHSJt9!v*cbN;Wn5GC&Zn(N2#9p2^+@5GGO
zUEb!d65r!(?kVwo-sZj%SMxU2N?gOc*6?1ey;|O;R{4D?F<(mTm=TO5=8;MqMa-j=
zco{KYro_vM`En&*L0nc4PuAW_V!l%O{S#4rBKlB|6shG?{Y&)06pHj<sG5jAoD`83
zOAL<)7EnDAY2n>4ORxav4DP6fuJ&dNj+*#?1>f{waE>XQ8{El=^n$A0S>6Zku0&dx
z>u`e)V<FiiUjGqq#uhG`*hdprnAjh+I|q3z89kLc%8e*!BMN36CzH|1tR5Fru^_OO
zP9`l?YMf1-n@h1@{(+k}YE0z&Hf9Po2+{_YU~SOeztN6vVd)n0Z@SuBI9P&>KOX-)
zwRC(Z+EEHpr66NC^P5TxeqZJB@BsZ#$g2x^dzka)>UJkP7bHBP9}<Xp0&xNp4Fqby
z<ZC>b+~#+a#z*N}m(y4%mGemDf><_}%Joa-Gz>rI%%3YG-sfDvbFM2yz*~xXCtY8;
z*qzQQnixkb6=WAN-bL)dhQhK?lPTXC!cb^#^yPLH2sGgk^v?+vP&bd8x_srY!`>9p
z9@4O98veFOdoM!c0OD9fM!{WH+;>Gkp3p?uQQ|>tcMy+&D<M&Hk2u~Vqak&HIyD`N
zd@ad_vRm-kE%<>8Yp$u=?PUdlU0v^m*beE}O4VFza&E#dkUcfype7v5@MrA&S^51l
z4*IP89?1npavd2zk{cY!4P`irbB^K^?MHDzQQX(|cWLiWsnW@}rfl){2+|%_8-`5O
zLRWit%BK1hnLR~z%w88`yJGCZ>}LtKD^cPr*zSrFKf=b3ur=eSa26??13Q+5oN*y%
z4Ixr<+0OX95?|nqFDP*lXI!Mj7dhjLN_?4kT_zo%<SA;d5vOa!9qt!tCPtjfh&u#(
zH0&i#H;FscC27H#sBaT<@cp9Z4pHAB=5XslgI1!hCFY<^qUHe+9}r`>r=yh`!FDyP
z*@~Jqg6$e5UMsX;D+DpsStr=8Q-0qnIByj^;5*b&sSe6F)@)<`<WzrhR&eD>&pKyY
z$m*R?Dx#^vMJ!!Zdbm2QuEUnFBf8pqZZInlywV@uz|swD4@6PZh}{}-C&0Se+wZSG
z{Gp)l#up~x)VdV&HpP59uqp}NDs8%+{;=c!d-g#o=w#)~bEr-=TOadO&YoGDM$WB~
z^9Aby^%4iF=U=YW*(xPr`y}iR%ScTnwok$CP)2Sx0R{)L4Ko30*gg%rGux4c?X$2u
zg#5I~j}1%zXNAx3x)R^O4mYp|tdXdBhV`E@8&0EdZ2U^Orth%vJ0)Jt8LZ}PVE07L
zAx?dWvt|=7;EW46dv;t-aW<#8whY&B`Zb&t;}`S##k>`41--E$P8GzPsdp7|sv_P@
zz3&l|d&B`uV6*KssUZ$b-Rnq)I?|1)`+C7}y|Q$&Q83)7#G3@eO-j5&Fx;WUae{rE
z;123euVn<MU4l2m2Lz)7f*r)gqUNgLa#iq!n^1b31+h#pXShxf>llWxjjCQHmMWDO
zN%D&dP+oBR)C0=zC_sb2)Y1>~BMSJ%LEnyp&&5sJN}Jx#S?}jun6mC9#ygou1Gj8I
zi#uQ5SU}UF<|xuXimX}eU4ryWkTr{~?<4*D$eP8@mr2lN(o<>MJg<>prET*pBf(1B
z=6RC@D{Y(QZ5pm=A0#!~Hp@H2g>8N<v8*L7Y_lH_(+9+nZT4!xb+zEnHfN3Cx<-lD
z3c+iI-t7B2!F8ST`&PkgD=Y9rKh!(PFH4wrdO=h#h%Xe8(#PHMiC?R4NTZM@Y|w;F
zU{geTcVqBbi6c43Nahb|%)l8$abNivwS9Aa3itZbJ<GzLS-2NF*L9>v9hm@Ah3uGG
zBKbuN+p|MhJcKQo!HD5hF`OwBT(h1o%raZ^GUY#Cd4dwq9_5!Ytj8C~=>-agD>rIg
zaob|t8B9!eK*qL3rO$fPYuW^2l)x&mpcZs_<%cwEc^?tgBVqz2o!n{ni@WPyY`z%2
zDDW5A6(+p6p<|NY*|0Gkz~TXH!OZA7PPL9Rft!$gNp-bwRNa98^#BOGBi-<;Sh|Xx
zVAXWB&w}7I9}RIZsQdlxAJpsJ!{R+`3Pe$}loOY7rcC7$c~v5B2npP=i7f}!6%IQ>
z7w;Xic!z?)7ahA$KKfwrs|WvC6F{|^HQ~$iB)00CIn`#)51tK)nhTum1+Fg`{{p|E
zQ%H~ar}UT{MuCS>XV58nU4#NoDQ`T=P(T^G!JtFuJl=8sFsBoJSGTS&nSGizbPDx7
zg$6N;REUNaA`Qc3s8<>4&n#30C8ObhA#0<?Q+`GTDtPL>>W!;U`)u6&0>l0!aez#O
z*PJG8P7@cvXZ8>NsZ~$aa=L;4@4NwCWKrS;EL~7Oq>$fb!A_rjJ8i__@RNJ!hHXHu
z8&Ee`UMTO;3a9)`3kKU9!%1tD70!)9kB!O-XR6?p$|{_&KU9~>uvEq(CQ<Vet6$<U
z&@El<V=oxm`Ny$K;~oW+O1$PU5f8IF)8;#5F-?iniI}d$8AQxb;v+;nqQsd*%v9nm
zB4#OZHW9OxIERQiN}Nl?TqVvUVxAJ`6ER<jj}q~y5+5VtF^0j+Qqv^AWn)YIny6kA
zQ+QGZaTq-d^20XRERjCLP;KRW;oSzk<K=9NxIS<eE`=u#`B?j7#Q(Yly6m{<_o&ce
z`d@rkmnPhU8IsSq%V#By<VHnu<Jk8ou1gg6m479q&96fTS#SBz_T;MOJM;JR;+pf_
zKtAy6`;_KTN_AB*gaQRb0d!B~mw{|g8i=ZaSV08?N?)|SDlAngV=DQ*AgFWcjky{(
z`@$tg%@-;5kC7+4NO^^<UZJ+&W$CpCRwZFmWt6T;!KTc;q~f4d+zTvYbB3=<!=|7e
z&E8y<g-w|^zmHqr$KJ||n%31yoWxrv@$KNEra4IzuMtb|n9WI|SVk;aAXQH6%83Wu
zRW&D#;!R@70;>ulRuD@TSXB|Ride#uiu4M!(_PY!Mb*_Lq?+`F0EiYriCrD>fMWr6
zm}<TJ%93gQC0_p$En;gQU^hSBGfER|Av`YkdRBR4#H2r|ZF`9HA0it_+u`{QeWFl|
zrDA2mQi^+&DxWAk!*0)%PZXN4Ym@Scg8Z5j_PAoxU$to!e=VklJdt=Ll29<*JiT}H
zN+vx4gUg_Ek{^4rIlV+qFHtZA8_ka>+Lo~R23`YFy_BDTvLEgiM(q~HK)6{Qc%|FB
zRCDy74@jN)KKJ#1>w8j#zQTE4;e1&3bd~eG$}qewqQq-hx~7aK<fo^wDi5re4V-Ye
zdK2A;tAhVkVKC$;54_3i7W!t<w1tm+%a6Pflo#oxC26^n7(p2fJSd^T+n1NCY}QYa
z!&4LpJ4*}B$mul-hNrA_4!G?VwtS-I5w?AV{lVuvbk95JYjYr;4y}YUE8#p~vd{o^
zfAR}h)?+#r)3Jr}wDr!yBiLzM&#BgPCh*4fsVX%FcT8JPEnX(-l!<z?Fr*L#7owh^
z3N#r-!4;?{%$2^`<2>GTfuR0y=ONL{?=SCe*@$XUw^}rStw9~?R;R=dP`3w4T#t<E
z*=Z0p$8g(Y*caBM`6ki%4EAF99=5#4wzoM3v|P$LG1HvLn<n!1Fi@IY^8S^47$lL+
zH;?UW_+W;cc<&~@Birm1GzBMLas9_}?+5Hy+2KpPWpN7GvlS{twuQ)<;WA`VhV0pj
zRUq36<P7-b&b$*|X%0^R`3Qi&ZQlGZ@0XeJj}xRg!43`?4J`$)!$N1UIL(*mMmd5L
zM0%hE^fkv7EL~9=nj5&!4Lq8yaRYX1P#T)|*!(?qW#6MX*C<Z$HcEaj4M*&KnnRBH
ziadXMl9I4z5+2H&M+NbzAS1yTLw%UWpz@<?(4vnY6Hj<mZFx;kLKId-VF!rZn~qo3
z3=gOuNk1f!E=i;>*f2Q7v?=-3H8d4I^#}UDQ0VA%Pt~h2Zdjs0P?&-jB04Z^xNg;f
zpF$Zy2g(N#;onB=Xm-lOz%%RUjQ;v3*2G=^oYE&@{WNjh__4!=j~+2{)R-Ydb%uOp
h=!{{1j%9z2V1JHee~w~*4rPCiXMc`ke-3AV{vY8-nBo8c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.D18AD937E7A47E2A.idx b/.cache/clangd/index/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.D18AD937E7A47E2A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d5c706af3c8c71e6542a4562a3908f5611f03912
GIT binary patch
literal 6142
zcmb`LdpuRy7r^gc=jPSH^*$F@SCltJ<e5q;zer`MeyZt3YI=-E@iQXNJd%>gAbIrE
z!zlEa-YPXeDTTzZMrNwfGzf{S{5IO1^I!jTKiApkYpt{P-fOMB_C9x!+w9rzHF!Li
zwQgaX{6bdoc|4vH{tFHXT3<!+ct!#qFD&<#PxIzxTS5KN*9F^y7yptsC24_b=^8zm
zi1nokPp>E~+ptmT$gG&+x?<O%?Z+LAeD7|{seu{V2ZL>Mb=U1~+aA`vhH0OJjOJg<
zH|^4^?Wv6(dU@wq{84qs&8^vGosZP>eLJ(&A`+~xSe3+tZHW&}4E{k{S7GbJMT>K6
zn>Pz>2M_1v+@w8E_AI+?9(7+F{asX%+%(xeOS}$d8cS>jyc_;pR`c@X#Ydf*>1*|N
zEcoy1zL>5bj%DP=Pf7h9t&hr^U{`*Af9V*H7pb=|8Ta+22Hdl4+VZ+D*>+1}y;*gP
zX;O;K$$86cs&44cHjtmZl%5yxM90@f@71+4XZ>W;^4c^%ob!%KJ+SJ^?-I-Iz22c&
zW&ieHq#tGwc%nkwv0B#l;_B==an_!HLQXzd8yMqr>~8<BTe@c}KU#O_v4Jq!^wwK5
zQGhe`aM}HUMaCAA48E1A$H3t$+IjnShY64O$7Y6k*ZCOC&p)$eL%VO$0J>dtrDjgM
zR%LC_p{;MU^4}Y!)rU4HHTIsGckyjQc=)NtspHg^B-bidy@~9Kw+fwKnCI)OroZ6`
zoGMW~`q!PeXScO&m8-Og?3osl8uELn)8m;NB<BwGqq+k}MC~&EOAp6yTV7}PQO4F;
zz5hUJQ<LM(KBo0r)D4jSwC`nOUFzHN1>NU-_P7KD$_PtJp*iJVUCR!ImDXj4%dSWH
zEx+8{Qc$tswdmD~)aKgdOV=!XpE49<*Sq;m$)V?a!mYQfg}5J^aPVBdOVh%(K!?Ff
z(>wNE`9IfO`##9rMnA!Djkvk~+SS9rc(hv0-CvtgUJ?8<sXAm0J??T+i(mhqaI5o2
z`R@vR{oK7)wNTk(r}u6W2Nzfwd(51-TdzYiL}z8u1dq^X<;4r9z3kuC;s5%7#eETq
zb(a?AxXYMZ&#IYW^6So`6yNz?!BJy;q)R`YUcW~EtixQjtY`U4+xKo~+E*uqZL(a&
zsMW`{TJ@Y!`}m^u#69;GGe3dU3_&+PW=X4zN%%OKPxAUVjQrD*PwY5wr%i6i53-jV
zmF_qa=cmM9yJeeks-=gpK5a&}Ksi0ddA05e`v^C^gtHk{79rEOmM9sHKVIEYJ~psF
zc}@Xy;6J}-IsLLed!<2H``-4-r`e}xxYgSjhR4UQAHSsbuh-W@%och#md!0V7T{au
z{`|_!4Qh9iqf$Qo*1hfa`Q+~}{Pb~L<$|;fr3sGK%Xb#dZwxG;gwhT(pRK>24Iek#
z%IUwdug=)G=Ca76#&f0Z2LJj9JA<oAGrg5QNwsu$b_D-t)9OCN<4If8tKXXGJy8Ji
zN<5|mbUVQ6bHkItYaV~p>w5jWZ<=K&gDh65MuKXj_qib|;>W(0|6A*oAO$j5C30+}
z^2mjVZBDfGMkRf;6pUCn4zkBVn;C70iXAdKK5I?fuL4k1<}uMoAsQ)tZivc)ha0`O
ziI2%DgB<SUGzd;(E29ljS$01BswrQQuLcTO#Y5y_FmfT{{Er5QOs=W9js*&fBjh5K
zM=nHMWOn?>ZwKS8rGPKsG0xJ?)*~05EMQ0R5QB=BTZJk<dmgR~G=9DY1T~=bxgj>r
z$S3wSU)ml&7NoF>r<|wy$c2b6xEtN+^O9%@U@R6t0R98e_}mbc+J+}XKTaIFs|M0o
z#a+`~d*njIO+WO$u3NeI{HUX@BZceOQKJn}`Q+;0-x2ZmeF2QY7J5L?1ES9jQQ-yo
z`3Jk<qE6zE<UU88IALsOjj(XFhAYs_2ax{&0%DbBBp@$gS$<OD&l~nnZbmfYZtOmZ
z3}1eo=G65i+Se#PlSa{LlmdC4&!kgyI+rsjI)i1BZg@naxkg6BH2xwpc(l?I_)Hbj
zs6q^J1)nJay%NwTX{IHaW)dFsMjel65F7@<iFB>~qW|Sav~V64)FMGGG9`VDQ;4=p
zlKF2d79=69BxFd^kt8->ZX}_}P-q-=gBma=2|+VmAnpPK;>}?;>VipxbuR`5{Jq05
z0r9bY^&})D_QyM)yYSBTf!tT;WbV0kj%7(`${0S=f}C2AGZ{LcX+utJT<%0pom}oi
zPJLX?0=F!fN1nHw`$w{raH-9RVNnaxA%bDUqBbsfB2g!o`;e%Q%UK}Gf-ldLXy7=G
zE_LEOGIuN(UFXEf_Bvhnevn$#rOIc9fF1%)%Iu)%9c+s5nP^HFO=*$1@|hS)7{ldQ
zN*K%KI7%4D<#<XZo>F6<PoPv1xV)3n*hvXlc_O8e$mJwTD~S@Z@?=UYnae4ZSqf#r
z%6Cy_ySSW6nWb`hH)Xb)%X=t#4|l$MDS9uL(<#$*%AB>okD~Xna>t*Gs*d?a+{Wpg
zi%fFSwDj30`hJq`+BoZOX8U&odNz1bUqBNsAR7`3nz@W7TxQwBbE-zQ>w$&X!V0|=
z6G>I`3}r5U)Y<n0|2x7oVhX9Jq%z}(j@DxmYA;#W%bWZJyD1viAK4yIBqJU!Dic0{
z3Cj{NkboIsAKjlMw@UpV<BmLqo^GUnpH=&po-Am#$N$cvRI(^_qE8CHl-x+cC;ih*
zJZQ#)KFMjCDFy3Na3%w$nM#;k2@WjR!K6B{CuN#unqX2B%gZb6HZGf}uZQPz9pSG&
zDM2*OalPYUL6RO97*MSRF)_6AMdCWC=i+B$XvWFVX(~HD2`EWGOr|khu;d3po6E+7
zV8`V-gWyAWwWZr7l~q^HV}JC6c0cG5K>(HO*2H-EVnGfv$ze0#N864-uSX#oSdah~
z2{0|xw?h8DhU@k~J}>+DB*_CjwGW{50d$C|Vbk*cpvh&+esCrnI9BFS%a7s**kU@;
zPe+SMvW4Aj{Ue<Ju!cYE_3@=(z<Rk7##eH=4n%b@o;YjR>!K!>BfPAf>r3sdaSv&f
zVj5*kx{k=NSJBGcGJw<B$<E1v_39uj8H9fmwWzrRE7B)fc;n}%%TE^*wHSCZO=3S$
z4OhN?pz$P${xDDO2TPU*fyRO1ve6)X83rFmZT29{A@aBb73pWPT!%e!5SbiAKExw&
zMN#1%$8Oz`qM4^i_!Nz2^RgA`v?2qR|3Sikxbj}4)6117f=(iE^!I~sKYX#5Jl@jh
z<uBj4;t^yblT2hwMv(F^D(-FCkgb4bJk&h2$+DrDT}XKsyDia78Hmb2pN-gA5S``n
zIS`%Wayf{~xm*FF3NBZIVI^3w_RfRoJXii0n8z@V&47Af>bcwiOaqr2fobIO6JVZj
zxd{xKz?`-B6qu*1Jf(eBLx6wK;5Xm&cD=iBy~AbPpWl3&PHoHeG>cm5ha)-&I)iM<
z$f!|D+-dfz1PksV!964<Ikdmk_}cU<#qC(|9`IM6r04#=mCIukwd?V^j7DbB$d26v
zqS3r)w3y`^h`GTnayO984K$M!z0A9Fk3LF>G{wvBIjBBo-8+2zQSE|p<oHAO32C{J
z^EO$4!~Ns0nC-6$nN%SwVxPtbK;vp;NoGXjbKcYvaAu3_Wl+2f8f;cQ2k~<-BvB{*
z;B9I6Fz=?nEoJXTm&jt5X3@+W136zEHaXql=SR=_#NY`*@3P(gN>{OK^N~qDvLnmu
zl$HI0TX`$9u<1zcNCR@Pq4C*TwS)D?>0}f00o{jP`1wK*6@oFDA=&3>xskU9(jVS!
zbs~(37ilII31X2p>$ZJJbsrLwK@6{XQ6@5CIR}YyxSWebxm?aeqC74iM52RS&PSqr
zE>|MON~Fd1UkHkYphYH8{wBO!Bl+ElubycuY|0M&^ziSw_*DEJRNljv-JApvJAU*P
zi8$htqps2)iSJXf<Klsi2X)ftuhsQSPdRk|`PDwJOj)#cjBDW=Y@q>7Za{X#!u1a_
zy@r;PE3k!hN}NvVlI6l@GN_3elmX!z)h<5y{SOvn3wg*SkKONX9$T<kezr{9SM#ds
zy}$RFplgS)1xG_i7c#j!fBn28@3qZCh2i5Z9nBu+hL5*&43}dmI+n|E6dlLqc#4ka
zasovsaCs+1@8ohKMJIAOiK3IZoJ`TlTu!0r6fW<g=v`b+rRY>H;{)JsmTMg+MlABw
z&irby$h#!Sjl2RAN6}0<vM)!DBmo~bY+Y<3z8!`AnE>JhFehi4C-IGp)K<-gZ(fZ_
zg~ACs`sz(qdQ(P0TWM24B%ay_F#7=0NY?Q2TeF1(W)p6;t##=;$)K<L;1A}o>7JL0
z<~Xa0k$Ew)B(tWO!^r$F%YRN9(<7~@k%AR#LADlD$SNAHKC&I4LaOmA(+j7pL?wmT
zU4=-a5SbB+Z)8rUfqm&7ENDcN8<8Vf8}D*2T(v0G)&J^Xy9Iy!9h>Ai<(u!T@3-HM
zRF_;={N{!wk|S3wS~7d^oL9)K&?N<!gb4?D^!F2zd!H^hzxdAPtq<1khZ+6wFQSQm
zQQ5f43ja<t>*UxMsvS{_5lIqq{NxcEWS8KPtl<R*tqG1D{;f@I_%iyzKfsfZI=SHi
tv(9p+%bWr4LOE@bmYTXyLsLbSr}{<C6SB9O?5&8sRbg+n*;_64_WzJ1f!hE8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.46B874AA40214074.idx b/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp.46B874AA40214074.idx
new file mode 100755
index 0000000000000000000000000000000000000000..34046311ab132c871f0dcb71cb0398a815a1a8a9
GIT binary patch
literal 7822
zcmZvhd3Y1`zQ#Mhgpw{St!*ikLRp2(zR#r4LN~%@kR_}&Qb3JLK^6}X)W%?uC7=YU
zTNFVIPvih{a0T>;au8HNj~)~+TzCX!RS1;>vfLYT?$Gy-+vj1(C-3*|=jpFAlbPwr
z!8J9FUDDDjrw*PsecGJK_O!ILoaDdRb7uKlC1F8!TH3r<zxGb_6qU9=RzCZ+_>zp7
z#`K<DN_w#m2cI2z=^^vV`OW!TcI~~=`{a#3<zrdF0bMgU#JB4QA3V5l_N94pL-yZK
zefU^K@Zz?9Tw+JNK=;tffATlZCr&n>>sTHQuYBD8d)M%Tm!lbfyY}g=6~|A{9CUHq
zo@=Lu=T3UEU+K+TOUgces`Tf}=icsE{?hARR}TLBwO=oruRqoQ!<TAneHSBHtNOKG
z_}V?C$6tM6bA>nkjjBVmcZa%w+a$cQ))v_uuWg?dTDaucy2;CGiuP=O;jMeGKK%#p
z^EdC#`sBxlKX83<+*?@bdAatDaPu?XGo6=@TbtoOz3cLsT@{}!+3;R<r~j(<?A!5m
zbW7EQH~+e#{h^c1(~Fsh-kffKd9n7;gQf2*{P%C41>Bz>sV>dkcyC70yXU(-xbI|W
z&x*#4C7b51?$zh3>M0A<%tuy^ta>_sNROS*Zs(fUKGMfAA~5#Dd0X24{_K?R^1of$
z{)ryzFHdgxqBQZ=w>zdd-Z?w`tqChU?t=HI$<d>uH)k9=bF_QS#6?5q?ceLY+I2x|
z`oy~<CgeK(eHMQ6=A7n5&SKK4%dBwCz@_gN)ckXI@vD=TbM+6OKeeXc`saF9TpV!z
zY+m}IXXfS|y2qQf;l2OvHh|mS(cZ=OXTJa2rEdSa6Rk_MW2;`zbPj2E@6e9LD@&Et
z2S(jGak~Aw(DpGC`~LVu?)$&4I`YGYxsPN$a$<U7^ZNex4?O8u*8AW$m3LJIZHMgL
zU*11{-1pTBpZ#Th$DFs1jMy`0x4N{~rn)0+U-vFw-v^Y;t2090ks0REhtJfmUH1oj
zY*lRK>A5GWYkTL1dp&sM-<mI89LU=?Y>m76$TRO3ug~6i@$$86+n;#t8+k+g_m|%6
zbwM3hcz2V%&H2gQPOgeazUWwcX5P-G;FoJx^cvsoh1KTac@-~Q*}EyDObxEDtO<Nx
z@aRJ0NYTARH{2+W?YtTJwD!ujC$;7^$9oM89Gm~!!{@JbJFHa=q^4FM-AdkGI(GH6
z+E?1dwq5rx+p={}-sbI9x9rFNvVZ2j%E9^7YZ8~fiR_q=*l_CmnQa<twr!Sc_Uzjk
z$-mLxRe0f@XWrZN-Mg!QY_4D)s|obI@xzr}&i&sOiJd=u^z5)TE5?0vZTP67uO8Tb
zclq<tWnc2Mw>`Judi9J;ixN_-Z04v1hx4XAe&Jtde(Uh$m$`5F4k}Y;nJ=CHWsY=)
z8Q%M|o%?zQvVJx<_;>%1<(b;MX-;^EbMayNNO<7H1JQ%EV-|Os_S2XNSM_n<fA`pH
z^gSn^`0Dy||C8?Q|G?+}P%p6;-uQIR%<LH-|BHHm)O}Ot|LSAJ!m9Fme>?5ysOGU-
zuZ3J^j}QK7+)%H3aCy!F`*Pt}%zLgb@aoaMzi+uY>Z8KRt5(Rf`mDe8%FT9PPi`EO
zHhcc`x;a!@T6)(HR#d*Bmt+y|gtWSX^n#4kCV9R{eP#Z+_E>1BIx35dNyf$L#W|^s
z73XYO(rNDIv!7*=5y?0=JvSq@vErU>AL^g~@y?T3WNb1nOfPJg+E{V<^-C8v{xW|{
z78#d}i_(i)r#4nR^~TjxbC+CxF^h~&#$D38<fJxMJfrnDqYtksxYj0ZR5C6pDCwEn
zSaGOmvt~{`@V6{7F&TGH?>;`YvEmJjiuZ538r_ve?oGyd>3Qu^8!O&?qW198jiqz4
z)5a&`{;m2?N^Pw8?BtS<SmlD8MLfy4Lwbjd)W(W$w)bv3I^*+SvdG9}d{_Ei8L5pG
z&zknYY<pqL+H0AqH4pE;zIM}z(vs}D!p?<Vt+h*1x5Az+<Yb+K^|NKyIaQ|t<aQ2i
zA)2LOl2&P$q}?=3GAsj=jLN_y<7QxzWm%YHRTd^$Hw%*-%fTe4axlrcIhf>G9wvE}
zhe_Ve!=%6pFe#`4ObTuRCPh|+Nl_JHQgn+jDX|hvN~#2tl3RjFnU!HuR%Mu!-7-un
ztOApYs=%b;R$x+PRhU#&6(&`;3X>YE!K9{YFsZpUnABMvCUsSZN!_i(q`?|6X{ZKF
z8g2t7P1b}-Q#E1Ibek|)rj+S-j-I-5r5tFWGVl&rXO<qBsWVE;&hKoXoXe<|-PjqX
zoG${?4HS(z9j0i^nFfl%oC#A5=4=DSV$Oys7IUtF;xOmJ6o)zAK=GLKVT#9GXrKhl
zg)k*xE;dji=3<x<F_#)B33Dk-NtnwGl#IC?rew^O21>zP2~!H@Y6GQWu7)WUbFG2W
zFxSGAhPmEA>6q(bO2^!2pbX57FlAtFHc%$!W|%TDuTg5sTXvv9%Ah-BecIGyraom_
zcA31+d0nj&dUjo2x4fP$<n%a4!p`J%jtBBP9|a0HpKKx8Lqm`D($J&*H1y~I4L!!g
zK#%b<&|~}z^q2qxJ=ViQkM*+9WBn}j*Z>PX&ci{E^K#JR{2cVS00%wZ!$Xhv^3db`
zJoNYg4?V#nKu_=r&=dRu^n`!_J<%gVPxOk=6a6Ce#DEAr$s<8e@=DN?{1Wt}fCN3+
zBSTO2%FvVjGW6tt3_Zo8Ku_^1&{O;h^pt=CJ=LQ^PxY$MQ~fIR)PM>-&7(n2^J>u3
z{2KJMfCfF?qeD;k>d@2uI`s5_4n4zTK+o_R&@=o7^o)Q3J=0@C&-9wmGyNv?%zz2K
z3Z>%CIX}CuQmF*0R;urib%m3gnYu=7Ik=rJXC)k#E*gk&F+i+~1>#&B5bxrF1eX9L
zx<nw!B>~AU8Ax#{K&ndx(p(yl?$Ut_mjPtDOrQbEfd1C`&blIgEHiZ_42#f6oQp}z
zGB}%v^F`zsac)FT5a(s&kj=Ts2K|W5xf(fWcRp=z$>~NyqmOPPH2UZmq0vVt2#r2E
zWTVkXM{G3u=%AfOAJa$}^f66@K_3$%4EmS^VbI5fYz+FCh>bxX6SOnvV;c#JKDLRl
z=woAqMIW0WEc)1xjYS_Dv9ai5gLW2uTqEJo$2AcSeO!!i=;IQELmwBiap>bBHV%DU
z(9WTcZzMeW_$I=mkB<=^eSCuO=;K2+9({bo#-oo9+IjQ|jYL47&_o3E2{9s|Pe>2}
zeL~14pihX{1oR0(yMR8ik%;IMn}~=$F-Ao6i3uX2PYl^a^obFhh(0lB7ttp*5(#}$
z6Oqs-#fXGHDM2LkNg<nrJ}F|8&?g1$68hvuBBM`kA~O2q7?IH@Cy0zbIb@U3Cr4~D
z`sAQpMxWA16!a-gL_wbtBMSPI1X0kZglr1>l!#41pAxhy=u;bsiaxc8sOVE;L`9#P
zAS(LQkWEFO8nLP9Q-gLDeOe>Y(5E#K4SiaSXz0@tL_?nzvT5kkA~p?uTF|bcPj4hT
z`t&BEqfd_!9esL&=;+f!HXVI>#HOQ9588F~8I8n1pV34N^cgW?pwCDU1ARuwW}wfA
z*bMX;LA!xIvyqtSGn<HsJ~Kv4^qC1_qR$N3O!S!%n~6R%Xg6DY0~eo9yVgAC2xYGu
z<`_27`n0SlIQk^->^OnSuIplSskELK4SDI4l?OuS+mPg~KtWzn%dJ3RL1B?~nKA5@
zlWf@^O>ehPK0|Rh`@kokg9c(83=r#Jfj9>T#5;H(!65*N4iQLlNI<ef22vafkm^u@
zG=~PHJ9Hq!VE~y9^9~KySIsD3mp_+Wq++AE(z@*)eNXP)r=RcFD*1L<Mp-*+Nv$<Z
zUWMI$;06~UQOday@4}-Ljm$(T2APdgEHW3RIAlIb@yJ4y5|G6xB_c~vN<x;Sl#HxI
zDFs=LQYx|*r8HzcO6kZ(lroUbC}kqM75AO{Q0vKKPG;(fBP>#!a(;k|6sKrpCQdQP
zY@A||xj4lk^Kptt7UGnEEXFAjS&CB<vK*&mWF<~1$ZDKYk+nFbA?tBUM>gV=fo#Sp
zvxOfYTJ;q@P8*TjrAn$x##*cNl>5*7Rtfx=+k09`nQZN8))7ZCQ_p2v77>r|#TTnP
zn+KP7N_KM#aywfO0IWkgxtHF4a@*2%);s@VedFvpx}FAN>KP!mo(1CSIUv5C2NLQ9
zAhBKqlIkTOxn2fR>J=chUIo(XH6XoS2QumnAhX^Cs#2=%+$+3&E;N7qiC0%7*QQ5C
zkG9s2y=mT#RwJgbt4Y3H)V8R=x}&g`oytr-$hMa3V`b#jR@LjyC8rDM0@?a&!0ypy
zXM9}`?@hiv#6G0ERX+Z6!+_^=w*HiSyJURH6zey9N@&&Gx?P85$+ye$%eq?A)<g27
z^7a#QOL@+*X(e%Oy?sLJKzGorDeIYe@}Ro?z}%Wno}PWqI_rl^{!V=~5aVNjSRV_-
z`8Xin#{&sI0Z8<TK$1@al6^9e;!}WBp9-Y;G$7rl0~tO8$n=?aXx6kxC*S_N!0cI5
zX8UfxsJ*>Oec9?>%)Xq{qpo>S>nW3bwfEgWIW0BupLEjyO(p-kujS?aEiY5w_CGkK
z<=rXMr#|{n5{Eu8+-7f;-rA9omEER&yS9UKsw;cvWDc&#?QmCKVMoI(?bpAod;phE
xyLxDSy8eIf78G?ZF6q|2XRo{aIvHLNWmWH@bYi8d{DMJ4rjnX_2D+(X{|mKELLUGC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.0EE0667EFBB4A693.idx b/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp.0EE0667EFBB4A693.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2733245b385853455f5ddb8938427ba095e2f356
GIT binary patch
literal 6832
zcmZvg2~<;OqQ|*65duLG3Wy?t3vRjlssTjV45EUzjiO=-Xf-a_ineNOjZI5M4a^u7
zv8J_HiYNN8V{NAnE>owD)oH7B`d(Y-;pi!LT(H_e9j(i}ka?H<-#L$S{N(rj^Y`cd
zZtgm>u%KZ7ps=vq1%+#tFI-iw4+{&64g9HIwbC9QfOXMfVQcn$X`Ne~oYDVscJ<r7
z+Q=8iEgvx?EuCJnansD77mvT!*_v?Z_?denJ0ATiUN&x=m>SjWJEERnbbfvH&uc_y
z^dGMseKBX_-NTuTzb(cw%vJwa_ED$5qx;stY_GfiCH<XL_o6Ou<cAMFeNuPj`U{iq
z&OZI%+O)WNuV!XEeo{N(#D<I?x^5j9nBDkJYJK4!AN-?h{5u=6jy9H++3tFb+cG28
zzdb$U%AOsqIo9xZ^DdaIZSp$iAotc@ou}1T*5Bw_UwgT!d`m&{=_5PdpK<@SUs#*H
z`;_tI_e(x8e}2W9lv}*J>|J;F8`c|xx6a-hX}^BF>&Ee%leNts=MVbt{Nl51UwRMa
z&3W%X>iS>k=w6;eEq-sge)pTo;zb$1TmR3W{^YQHb}2t2ZvTwP<c~UsE;`%cI$c+@
zKkeY^?dhYx$gf-{M=h<NnfF@4l;OuV9bvlnE*)(sam+fp=1{*on<~FaxVX9hE5rA7
zl{-Jr@V|ett<vzjo73K(Q&((B{Fo^Bwm;JvdErL;u!6a-Pg!&BjP-u%x(L(Nk4omm
zjkS+n|Mh#Tx?dlgLiQQ5(p@lV^GAsVe?675XWmw(YDwp{otgW7Ga~2i#Lk=X;TPUm
z9e-iE)!6*;--k|QjttZf(fyiWzqr|AZ@L<hrd;0kPSn^bF*Aw=rqpLh+rN7H$<^!q
zn_Nepojd0HSK@yEk8PK}YhJz7xb*6Bf9t-i=O%R+wv0Ugb?#Gn8+8}-!*-u5o&BHr
z>o@(lZ(!_!OC_gQosu`FAFQ}UkFgxLjd@;*y1&9@of$d4eaVfoy-mN+m$$XlUtfJS
zziecJJAKime=0U#iX;B;)SZ_6OK<!>Wnc9EyIl_+9C_vKuf=BHoyPalZ_BfjMm6aB
zc9swA*yi(mKCtY@nqv(c|FX9(y>#e~?c={)le6RAnS+rN<c<4s3ml&%E?YnDQu2(V
z=0_<l#~yn=ExUL4Ri%69mGmOV<+VR8>AW}eTP1H2u^_+w1M<M;S=$$uz16qn@I(8S
zLm!-uZ#|OtM1SSwb1$6DElkMY>Hqm_PurZ8Z5=yqANury<-b<{dH=Y^mk+;oa&pAJ
ztdB43-0_E<!dvq<rGK@UeIxa6CuZfBt{Yx6ykVO6?8Vq3^`w1?>-(+uIz}C?e(}M=
z>4`H}oV-rIT{8d2Zz8Iu+}YjqMQgjc{qd<4rNg>!HP%dJMtR!zOdsiOKD%I-+PGjv
zd1i|=Ey;1U_KS{ftuqR*-CdX8#xMTM(XOKhZkFs@);7iXukZuA&0kfYu02-2Hr4<4
ziOpxWe)q}96U|xDu6}h_zx>vG<)w2=sXLFhq=&n%pP_Gm`o3fHlPG0F$>fwwd-Bq=
zQDbh5I@Rcm{qw_bmXs~4|M21c#~*E&GGO@Bi!T=RTlmd}5C27c`9jUJ)oYhmtRlj~
z$aUX6-0L{c7|FSF!YYP^4=W80f%8T1Cp`6&y4-iwG$UyZ#EIdFk-<TWqki3c*;OPz
zZ6u!!#3|t^vB5!$`&NqS**`SaMTeCJ;><po^MZpG#~!L3w7T`?pNwQlAdU-<iwq80
zJfh#?tb`NCI*jD3K%5kw6cZe@IQ!wxxA*_J_K=aB9f*^|lOuwI7B6^o|Jv%>u3bj*
znLs=wd`N6?(Bh__YInvo&0o?t?CC(9mY6moIB0S6>nZ0B-}fFjk~0Eve0Y3JaM0qL
z<!OJQrQ4#BEDpp2!UseK2Q7Zw-+H)x#b-Yn$(e!psqm*FgM${YT=;zT$oSCOlTpDn
zcb|H=_u$oxwCIW=L(yDq?X*3R8Wmjm(1(%+C#7mj6kU-tG-*VLOtgv9_H1;8$zm!B
z5k*swQ5Fg^nx-M6Ei_~d%|OOj7|2+fg^abZka08z8E4@j<7pl;-oisB&;n$FMSx7C
zMaV>p2$@7nkVzH^GMSbklPxl23avn<SQN-qT7^utsE}nzS*o_PqAMm!6MIQJI#Qy7
zM@4AYnVf`a3lZ8SCO2W)1xz^!3UkU$P?*zBg2tS76Ex<GlVC7s+ysL;>m*prSvSFA
z&N&GVbIwh0nDb78$DDT)Jm!Lv5HJ_qgn+r|Bt*<bHz8s!ISC1K$xTR@%T7YZTy_&O
z=8BV0Fjw4!g1PD>RLoU3p<<pd<!6U>X@OMGOWG_E6+Dw^yHq=qnWKVdFl{=zB7Sgu
zs&+03(9rl1Au<)4X2N&CYAOY?o0b7NOs|HBDyE=ESt;mIb_#lwgMuDiOhb>h($J&r
zH1uc(4LzoqfgWRJpvTx5=rIlkdTcQZJ=V%XkF~SVV;wB?xMB`^oRxzfXXl{DIXLL?
z#XR(QD-S*1&O?uP@X!;A1?UM@0eXU6fS%wGpeGiK&=aj9^hCP|J<%aTPb!w6Cs`%v
zNp=Z(l0$-?Tr5LRw#v|x?K1RahYUTXSb?5mRiLNX73e7r1$t_+3O&`TLQl1;&{G{M
z^d?D@dQXAT6**E4P@a_6OWNfsCMtMc4;?fnvndz8?q&*zHq$_inE_(WED&erfOs<x
zB$x#t(JTT<W(i0(%Rq`*0aDE>P_~qvrJWwMdy{!l!Fv+ev>MX%CK=i|Qv+$*h1^1#
z_9OdAQx~#JXL?-+eUHwx9oeZjZPbTys)nS{M>UWX`luF?LLcQPDfCe;9fdy1qodGA
zIrS9!=o*qnAKgIG=%ZUm8hx~%q|ryabTs;CkB&wk?bOrgV`@kSeM|$%ppR)G8T2uJ
zl0hHi(lO{`JURw_j8o5`kF6nD^sx;ji$1o6WYNd^Nfv#qOUI&*_2^jiu}(dUKCXu3
z(8o299QwExl0zToCpq+SE**zH&ZFbd$2s*J`uG}>M<3rn^62ARNFIH>pXAZUyL3GI
zc#n=pAMe!j=o4y40ewOPDWFeiAqDgaeo{c6;L-``6FfQreS%XjpiitJMf8acq=-JT
zg%r^z`biOeqDv>DPxR<S^odTrh(4)?l+Y(NkP`Z&7E(f=<R>NcNiLm)KFOn#&?h<d
z68hvCQbwQLK+5QoTSyswvY(XEC%beq`ectzMxX4|%ji>TNCkaL1F4`-X(1K#DSlEx
zpW@Of=u<p81$~NBub@w@AyxFL4Wx=bwS`pCr}{}1eX2{RqEGecRP?D%y&CdOdb2a^
zLHDXlg#Q0LlI0|}zsimsIERR?7&2~1uJ)ukrSW9_SFX;!WZ*_GF+MqTqnDJJl&oFs
zroPodPxz(bK;OV)nZYy~&eR49h&IqbjDZ1S4J;66;DC4o4<r}_AkiQKNd^f>HpoDV
zK><<?YA;Pwx2;H|xBe!u=_%t<a<wO%XQs!Ey8heDK7r}1$gCLcGN`Rg;0oXK01lTy
zFJan`4?JFiLZ-b0jm&rn2ATB|EHdXMIAq>S@W_If5RgSLAtFm&LPD0kgp90s2?bg8
z5-Rd!X>#u)!b?SYUzlbqC4sd_%TJr7y=d5A`E~v_j(xUgiPEw}ZHeBzn?HE``K^Nj
z({YJ$gS7_<t<^6o_*w#1&PSL&!Ikq76f*52Xk^AmFvzTrV39c=!6EZLf=3p7gn%si
z2oYKG5fZZOBV=U7M<~duk5EIb9V3A+NY6VK*vu-^%W!N|Q9yJR4a8J2Kx`EY#8q)X
zd=(EQR0%+0l?WtNNkDRy45U;kKx&l=lq==--U1%D<yu>Mbx&Pj#}1Di-cQ?@gKOIQ
zlq_#52uvsUOHR~o6i-Mq7xc+*x)qpCG$o4KOP^EEOt@i7{q{^?x=>#@O#9MY`N28y
zx3M4mEij!{npUZOS*~$?x+7u+Yza(fC1j;)pVnSh1s1vIO%*J%Y#OKizoIKBnF69^
z8i<h@AXa99IGF?DWgbY71t3utfh1W1l4Ti4krg0SR)Ho+6M7#ZZW#{NljaY4PDcYw
z0~BpbwfAp<qo?QP8*H)7Q~_ITqkw1|4aC?OAlAkLaW)Q!xA8!NO#l*YB9LU0fMlBt
zq}UW7)u#5+%7x3yd)^#YudJ-L_54!SGX#Hh6?cvO=dPv;W<^xavz5(%t~@OGiGQc_
z{`XUXzvhR2eJ=EC@SpuRR)%IPmoHehIDkdZPt)o9ghv=6jnRGk$Mh?V&CeYf8&#MS
zH{hxGq=DncXJlqg$ezd~n9ReK(W(F6Y+~}@l(eD4Mx>7#GnQgGUX;}#(jZ#MOGunN
NWdT_*eUgQk`ag-dq>cap

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp.4B2A61CBEBB69C46.idx b/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp.4B2A61CBEBB69C46.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a45cc88f7db30aa614ee35fb86fe0be7f862780d
GIT binary patch
literal 8358
zcmZvhd3;Rg!pC!-88KN1A&H<Ot(0~Q+AMSS{TxHY5KDwI%1|MqwQE!p!_YEChK^Rp
zZ7mI{7geQ;)-AW&ZB<p>wtBTyDy`^;8`{$Kj(Oi>et*3Ad_3lRzQ38@<jgZ?c~U;e
zG&lE5O49s;7S5b9_kMd)Qc}m*zd3VfSLMXuWO`E4!e@?+t{l;g>9ARP$-DEwfW6h<
z-reKxZ>4NV>Zd+Fa6{jpKHRZu-PD3m=$FrC&1jz<<ogbs{=$WeH9LQ684;%ZN&Sw8
zI;Ad)eD~10uOB(xDR=SaCFdT9^ltq6`1K#E7w<oM{rbwy#};INc%kiY-0H1IZ+mBQ
zR&&Jt*h4pNT)Nhr^Xv5+=T;xzvuEn>{v}KJJCmDz$D0@b`N#RG^XEPCmbc%^tIvGX
zGWT@KN0mR^?Vr+Nu>FlmFOS>Wd?fXy6n4N5O{d$uIN-g=_pYu7rtKd-WL#C(<^4bZ
zt2Ju&KXhf&{{&wMq}MFJ=i0VYS*Ja|X&XLzYh|0K*Lp5Ty_d6oE$jQTEAycoq2O}K
zf;Tg(!UtGY4xc<YZ*R}RUpn@;pZW1e^Q&*1J=(ZtZQaY8=F`KrW|mC<V%0MR;fxaQ
z_TkEymOiD!Hm-W-+Vsx<>al%guhD%%S8o60&5zT@Yf9be*Z=u;?YxBpyKLRPcK)*O
zyX8hR0xJ*QG3)L(O0|`vX5D*g_F}ScZ#zfJrRQgt)eXHodeVMl$Fj!7)ldEP?1H_S
z6VkR93@TW6B&Tffz>FQA<PZBM|FdbGw~d)`f6>f;uM4j_@pf5A_`2bN&JA`)kG<og
zN0MIK`vf`tpXtwi`ZH&LZ+VIB=w$iu?PX&-_9+|id(V}jgXbHUKK_aJak=^KxuIPL
zH{bQwgL54xl<B9QES>+ifBg9Bwvzc(XHFj(yKvK?mWdhrdd^;Egqy#AE@jK#m5=S@
z>3>#^Ib|$fd0}f-#}`fydwcGO#g7;4nso8W&fC7))46a9`SZ-P8%C-_4~~5}XUDWh
zoU`6KIq=$RSGM#EKmJZb;g7YO`k(rw$MU}6g7S5*T{)nwr?Qra(%%joFMRm<kv|oN
z_ivjJSpLxLL0c%H@3Q;K=Vj7cowu#|<=zK#KYg*6b6dl`&0W3a+yA&{#!|2L&G&1E
zulamwC}Uvl9bI>>KYVWe!Nd2y^1+0X<H0X75BYaCN1J!QvpeI}1No_`J)IYWSz|^W
z+S0e>bTz%<%%cacr;ZzPc5vx=ul@I-x&0RP*}3Jta#8&D#=y;amln9!ooeXs+7q0y
zx3FPe^q!`&3(;NQ96#~av^B2%Q?~Ek{^9dElO8LazWvIKlfN|e8Fbn9;Q4(!3oF@)
z4}7{}?$u4tmux(d^4s#h<L_ItDI6;LcmIbs^<KY`X}5IKC$B%>=cO43hb}mLdEk>h
z-(dc=_Wket`!=0?e_LVEQ|T>thA$nf&Y5;%#*L(-+$$SKv>DQ+_H5<x^9{R)58Cka
z!^Z0;W_H^?=hc?Tu&ve2qnv-X8(n;CU-QKti}G5<Zx2(0`xj0+_w@D6KYu)tZ3w!3
zEp=ZO?0NCf)kvg6`YvC_2W`*%KJwS>X@5#DJ^bLd+rB7pZrgUa(*wlpa`|M31uqJ{
zFCYEo-9PUdzuMpM(V*7eTi@UON7~!fJ60#nSu}IfJ}N0GIq#JxwCCJC(@4d*B!5nF
zPD*@;9YNwhslPt^RiLzZL>d_#(>;=Vbc_#{?zrc%uJd+ZJeo#^#dN3SPAT!h(!JVG
z6SCesbUuxYiRs+r+>H2O>3;4<*PQ*li!F_ekLmu&{l~-yOH04~=f@ov7VSwRl`-8t
zxqI9AVCmBIhN~V|N#}M+BV#&0JHJ<au(bd8>)*|L?CQoeGAgF?lJh#o2TRXfFl@5e
z{3eq|%40e!IV&SRSb9y<qD7uI%d^sx#>VvEHiO5-2TSj5NI96?ZO>C_WJFA7CTBY0
zgQZ_A>UvM%SF3lWk^5r$j^sNg#s^DZyuasbuA(0c(}*{wyCio>i4T_kqr>QZXJ-B5
zLK+z!)18w$r^E+K&z>@M&Ih@PPmW2Ae`28{zdgI_EY~yLpW8h*&)V}b%FpeUAg9YY
zz<PJ-erK_h2J$#d6XbTe2Us6ay5C*wrhz=}(ge{i8j^G|4N2NVLy~bZkYtJ(NHQJ<
zlB|n`BwNfvlJ&5V<Xjvixnd5IoQH!X@8TiJ7xR$hJv<}@mjFqjSb(J95g;kLL`aIo
zA|yqR2uaB$K~gG~ASrnyNXjl5l5(*ON!cSqQgJDeREiZyDjo%rs!N5WTC74+^{9~4
zTpA>`Vhxg-M}ws9(jloA>yXqvIwTF30ZF6SfTZCuAZfZxNSehaBu$SA$<m@y_pNWK
ze@M{~prJ)WZ;^EbD@ctWxe{L!XC38SMI|P@GekKz0=w%d_bMte&D|l&y%CtMqiFPW
zh@#Ombrgf12~iAswvJ-avmuH_&(%>JdM-q9==nN|N6&{S9=%XU3Fw6oC7>7UC=tCF
zqD1sk9VMZcLX?DFuA^l1a)^@AD|M8DUI|eOdbN&H(W@a!MX%LS8hR~6Y3TJjN=L7U
zC>^~~M;Yji5M`h@>nIbw8KO+|BZ@{yi3v8cXyh%j&U&#ab@P-5Q_AZshY3~dT#_KS
z*If>isn)$DL9~|!N7vHe7%u~ksb#>iUKSi%%Yx&)95}9)1IK%LaC|KfPVfrggjxZd
z=oP_<wIVplD}j@0C2+D=1}E3b;1sU{PN`MEsa_SFTC0N7yc#&IRs*Mdb#QvE4$klz
z;EY-Woar^enYAW3Z;|)bF(ut!UQ~XItg~yU)cDm6Ob6Df!31F48hqp|4aBev5X-Va
z9LoXmEDt2G0+7gxKoTnf$*c^dunLgMsz4g60qLv`WUvO1$(lfAMP-8ZeX)K|k{Z9F
zOYD;Ij`3rykac~xFg1Q%mk2q7#JPeb_M9_9oEwoFiE{^XD{)>$4%nQ_Z1DWB&AA@A
z-tJs#Pk48bxK|MCJ;yF~BE-EBxskYcAh#0tRpfxpz03yB58K@9k?Za5rS^oUgM`LD
zIznjdqZ<j0eRM0Kv5yYeXzZiIHX8frdOMAMOpq|x$3zH&eM}=^u#aga4E8Yr8-smJ
z*v4QVQ*URmj|~zQ``8F!v5##eEcUUjgvCBKU}Le54cl1kW9#iK_HjYNVILPE9QJXI
zgu_0rm2lX{1#BGlabX*WeO$er!#+Mpc<kdNgvUO<k?`2Zw-O%v_<)VaK0a*Yv5&8}
z^VlZ@iGY1Vgb3ItG!g;(gjOP8pAfJK*e8T-0`>{@b^-guAQ7=oj1Up~#6}`wpV&%7
z>=Oev5&OijO~gL2-Y#OF6eJS%Nf9DppVUYs?2}rFgnd%LCSjixwn^A0)!QZPlY>OY
zJ~={U?2{XbjD2z|k+Dw>*ktUJ!!{ZF<a)b|eM*oh*r!B@f_+LOQLs;GB?|T_0h@w-
zO4z1gpHgpEuulyV75mf(QL#^LBr5i)twhB>HDFV*PYv5t>{ILQD)wnXqG6vFAsY5+
zjYPvft(9olrv+>p_Gw|8hJ9MSUBf;-NObJeBSgnOy^-kHr?(Ov`}BZK$38u5)3Hyl
zx9iwv1c`xtMuZsHXEYK6`;1m%V4o4N8Q5opZ3gxk^>zdM%pftb&x{Zg`^-jSVxQSc
zOzblQHWT~Iu+79iv)*nd`ji!yl5Sj^dz#q&m5$0u*4f$a$8)useP&7jyK~1<>Ha)A
zPqP+R20wLp%}0St?MS*mD?6*3wWtzXb;!=`W?dIq3pJUk@dX+<&o@!dSMZLjiE_V!
zcUVmnjm$Jr3^Ln9vB+E##Ub-e6pt)4Q3A5qM2W~!6D1+bO_YqRG*Jq&+C-_yS`(!q
z>rIr7Y&1~@ve`tL2_CZPyi5Ek@<PYhx{Slw2QE(>?mlqw;GlsR2Lr@9SRl^90r3tV
zNN@;1qC*6d91@W1kbxA30;D=rAkCox=?)#pa2P<Q!@Nah?uJ>}uGMQ}Gph&PL$hwz
zM-I*G{rx(=O>A>%N@<2QE3JtayVkzBqz7{;N;!AnT#8ce9XOYw6phS8DF&I1QY<nT
zr8s0hO7X}-loF7|C?z6GQA$FVqm+!SL@5PXjZ!MI7Ns;~Jxb}wMwBv;%_wCek185<
z>weYxTE=$T&4o_bX*JFTFwbh-3t*nr&_GNL1H{&_KwJ$6#MkgZLX7|<)`&n-jRYju
z$UsVs0;JZcKw6Ckq}S*`MvVbv)|f!Ui-zC2L788wf9M=jGAuT|@{Rlo>k##*=hq1h
z1Nx7;IqQ0s7Fx4z#TBDFb9{Bz*yc{zow{51|CXjw<BP-Aetn@Ou&DCvGf%`0ioH^L
zwNLEe0VxC8TRV8y!q?jjo4Lh|y=3?H-LtKGA!}hd_EFtjRkmI-c1WmjRwiD_UE!{@
zUMbyASI|IA1p~xZus~b|2gFzKKthE8Bvyz(QiTL0SI9t0g#x5js6bkU2BcT$Kt_cD
zWLB6!!-|HLSbNv{tz-9+H&@aVy9Ccx`+KL1T6^rdA7Z;RGdr`ZwHx0oD)+ZBwzR}H
z=Qwi;tv_}g87cnY%lquz*e6tBuejZ6O81YmkGn<RoU4=9b=>=9?78`s`ID_F^4)+o
z*V<;}tcq<e%__~ao@=e?$L7M#1%2yXwm5dJA?)6pN4GA{E+tmVTH=qbAl+Qwx6-k*
zX!oo<>+2e${H$IHa{8Q;tep{?5LM3V1i5|gNr?&Ju5wo=i1yLo=qef<<72=vRSY=R
z$AV+4Sa6(=1IJZy;CLSoj<4du2|fXwP$hs9eIhuqN(3kQBydud1WxwJ;N&V9oZ?f!
zDOCzM)u)0}t5k5BPXnh_Y2b994o<Jq!5KaSoKa<fGkqpFv&y{Xl4eh-zW?UmcjnBV
zJjWON-{ya0-SU1L4;|akcFKhNeG?~5PP)1I|9;}p|9w2c+Ln|wG4XIx;^E}P!zqb}
zlV|#?r^T>zYMIU6Cb_L6B`v*Ohm7`vIvU!Y9a9IXnK_+0cgdw0ju%8p?w93c?<nb$
v_y29p?$-Ubp82=;D(HRJ-F^EHxTmn#-J__h>s~!8d*I-y#2l)4s3HFanP`Ak

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp.E45F1434A64DD3E5.idx b/.cache/clangd/index/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp.E45F1434A64DD3E5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ce3d33aa9b6207cc5906e9a802d5d4bff171510b
GIT binary patch
literal 7368
zcmZvg3w%u1zQ<>;88LYfB9lm@d61OD9zrr<zu$SsL_FeEgA6qqQLjWB(rJgG7{t)w
zFs*uYA{<inXjLhW);)b_sna&Ssx1|SOG8h+dv@lYo%!F-=jQXt%Kq-}?D=J{wf5R;
zt+ImL+>euVI>}zJaMrZ>Qw%ztuDkj>Z~k0IPZiEb>2wRXd@*L+$X-OZH@Ww$yAEYG
zRsPaH`QsDe8*~H36?q%*-#>5IQ$Ib!<N4#;*)MdBa?^O;%-vUSRPB1yI?{{Tbpy|O
zdPFSo{W_=q(&FYG35(vWyZpSbZ{X6|``0TM9XNIW{_;1!sEPahYUfAPs_mzq`J}?s
z;v4$<oCgnX-fM|}c>lrWRcH6@oBr5dS4Te^)?z!`vgqlPE7PkNEIwf!xcu&xpIYZP
zhkrTldVl-0ZiR-ArXHNIz2#)Yd*Ni}^~UB-?`EFyT{9;hnsH!6@dQWW(m~(-+2+q3
zbo9>Vzqxn2qN<h-zPICosX5;^W5bsxmUsH=+Wgyo>uuA+68xY!W==elaXY8xSd7Da
zh!lD6`NIpEQVV}D9_Tvjt1qi7AN+hOux73E;O6RKrQ2h2W`4hNONKW(hw42-7~9&f
zsC47XPwvf({b%yd<>_PkdG7Q+ckHXk=On?|{NX=N*DhF?*K_;cwbe_m^-AzZyOtmA
zGrRvsMbh%ovxi)my9n)X>SAoY`Oe%D=kVKOrXEllmIM}6Zu)a;O;gOI$ekGl8TBXQ
zOA7O%8_uPb{*?CZjMyDxUzn0T>%Z&0E5ANn;$beWdp>rZ!I<3iod2ZmNYiT6{M*d8
zzxfwsII}cIf2x8%-n(Rc_kJark5iX>4qpkpw&I-hRax$*mxm`7w*2MK7v~$l7G_>}
zqp13=zu)>`M^3flV)M!I3pXEaogBSCb?y?y+j8yg@NFLpUm4I}|1oat1!d9ltJ_W8
zcQ==wp8vUJMaG_~H{OWd@#DVOtZnG&#jP7giNg<%e>uKk#^M39Pn^%YcjV5tf!-CL
ztjoGpyLr%sbID6_Z$?@DkvoT^^_Z!S<=#4UHtXegM*W`UJ+NbvYw4W11=}zNUox?5
zLCmo21D;v)$B-8jzIiu&z>al8S`w{gJD&`G;WexD)3voD)_nJxCpxdTPvWli$1krx
ze0<3JpH0d+>;69GsC`$9zh&<yd!s)%lok<@I^c%eG<NjSZFp;Q<**GGUpaI?VnXrH
zg+*7ahR4Gb2ENp9*S0fdEc?rYyf;&B)(ow`u<nP9eU-0tsm|TMH!Jtxe;xIj9us8G
zTkkGCwC9_H>%YGz;&XEU7F###;UrVW^~Ji?6LS_kTh#gHs)uW<>dt=?mPJlEzINxz
z&t}itU+;dT%)YxMCMzJkIP&FhW2aT$`selk`Qyo%m>okuO&s#Umg4Gp<fi%VXT4J{
z+Sh%w=@v4-Y)+ZHu=C-$jn43jL3Hb;iRUc+cb%KGZ^qCgv$oG4So0)i=K5PlXP!QO
zqVb(^|C)HQ&u!_)yvIZN>(g>8C-%5>J~^>C@7$fMFFf<&hgU3b)I4pAdF|ETTc<qO
zy7{fbOX!O~pC1#w`aqu#?Mv6+tNEq)W8>w;iN6dy)N*C<{iok?J6F~hpCbyM+OxlZ
z?-lHHUBjw*FU^{|AJgfu+xx%n;^_ZQB$_xuXYUi%XL7JmN08u8SjzjWrMHKsMxt_6
z$A`s-2MbL{Jly)DtH?4k5{*%Ha#(WrV4>-#iKn(jpPgZg(v4SjL8pQV!9vsB_r0FD
zVDF7nk*HMFJ;Hi~2MbN7cb&nQjvc)ciN>lrAuJ&}SZKO$?r*707nW`5q8p{^w79hN
zV4-R5(QmgJuD-M{5{*-JQdm;wV4-RI<NLoZc>V6iNHkj2DPbwygN3GN)s$ASEysvR
zRHkZEm?=6~Xu4@#_~Ec#`!+?Qk*bafi!lZZP2ZT3dc>T4D=QLNRoyeJXLzvC^qs+7
zS9s5UeI^n;r|QhG%(1~j(@(mM*?)2N->*iZ5vq<2iwzGJnw~pt`n+4FP|Kl+V2gQ9
zKH9qHXDT(yKGryPs#fz_`=vz$+b#5=grtNNt%9TM326!GA;Qggrq(Z_?07z26e7Y*
zWNQC^l%2>Yib6!1N#MwQ5;)3C0Y~Liz|m$JI69vOjxjU9G5HK|teFLl&1Zq*%p7oB
zJ_j6c=7Hn$dEf-I0GyC704JJ7;KY0pILRylC*@1P$z~ZiIbQ}&F)P3+`3i7_*@Z*3
zhK#b8WS4Z1_S~gJ1fRE1v*Jz+Ux|g93-@67M!<v<BUWOeCL=r;u@Nxo#7O8#4@N>y
zIWY=)%7anR(@u<rp7vlg^o$c@pl3W713l}+Sm;>~#zN0IF%EjpgK^OFPK<}1_h3Bq
zf)f*<7d)5%z39Y5=tU1ELN7Tn33|zcNzltqOom?eU^4WI6H}m9JeUG~S#}v0YVwiU
zBSZ9`u@s94j-gtUYh$R|6WRw*?X>#XTJbW_4{Gtc5D`|Q4D^9oqAo<Fl?0BgC4r-?
z6mV261srXqfun0_;20|d98=2x$68t7*jg4i&dLGD)pEe`RvtLMmIqF-3cv}q0&t>L
z1Wv3Kfs?Efa8j)VoNSeWlWS$*6srQ9QmX(rJbQS@V;f~3kv*b=v;ntUMDV=^nvTRJ
z&{!lPf#xAeKom&<qDdMMLo$F^k_E((93Y<L0STl4NF+r-5-9<aNg0qrDu9Z!iy5sW
zYj5i4h~OI<+-M_)zYj;JMvQnLjz*0b37KldD9CgpMnh&AF$OZ*h_R5lMvQ~ZH)1?w
zp%D`xi;b8FS!%>2$Z{hlLsl9wCB)hcq#`0X^8mN)M))!mYGK@m@QsiI2ycMghVZ+P
zU3z?p9$fF$<Le<i4fv~ukSE-TScXD<i0~m|Bjf-g8X&hJ;x1&Do>-y>*L(HEddN-#
z@v0%@NjD;49_d3Q%p(Jcgn48eB4Hls(vvWc^y*2NM>-88%%j|hf_an=Q814RAPVMD
zZHR(-luJ*+Jj$!5U>@Z(P%w{nBO2z>K19PjI)G@HN4Fsw=Fu)a4fANPo`!j}(?G*K
z#*G-5$M_Hf^Oyi)U>?(k7?{Vn^bE{nym|)aF-`*m^H?`xVIJ#4EX-pAh=qA<8)9J|
z>(a9@kM-(Vn8!K|EX?EFh=X~Y4{<P$3m^{Waczi$d7Mkn!932Z=U^V^G;lDFcOxF=
z@jk@EJU)PUn8&vv9_H~bJrDDEubzi_ywkwLJi(0wm?!v<0P}<Z5@4Rth6I==xby<d
z6TEr><_S)N0P{pQ5@DX`Ln6!*14x8<VjB`+p6JqxFi-UAMVKc#4I<2w+(?3Xk`GBR
zPYNIj=1FZxf_ailFTp&?tCwJ&<TOYyPj(|2=E*)J!#p{FWSA$nAsOb$F1-x%WUpR^
zd9u?W!#u@}6qu*@kOK3R08(I{(uNe6r?~VA%u~F21?DMEgA&RUmfh4nxHrEU8UD{o
z&S0NtoH$;4rHXME>jw2t7>`BSQ--BT+G@6N)A2Q5x^8wsQFc?Dsh75jRTp@13B9!W
zM)Bq=X7+dA?yl-2(TMj0?+GK(4~*hQ5)fsi0MSMo5MyKju|^gUXXF6!Mjnu06aa}v
z5s+k*0LexfkYZFisATB6*>UDoYt@QK9+oU=3!_oPWBOjJr#q>q3&RVewQAGqNu6@F
zFZsbM*^l83a3<o%hz2+l@na-p%8yZyX+K6oX8affne}5VWX_Lqka<7GLl*p)09o{7
zB4o*rNswhfCPP;Im;!lJ_Nb1Pgw`w6N@(BM0F_XM*MQzoMbv=aP(=cwswhBo6%B}~
zVgRvKEFi9m1H@PHfP^XmkXR)GlBy&?a+M57sZs!yW|wx%JFAQ2&j*anDOEc`nvzzo
zO+#PFe>iDf=AhB-eIvCnOY0lU?kGvb*;R?^=^k-ClC*`@?pD`J<9^<<T79I_!_&Kl
z>Su8H;I3Nz>{<9>r_x#5a@A{-x+cYGGe2#cNUiDiJt9!k<@mVJHxcE;IPIIF>|{9!
zh$^Q5(d9HCrknx9ma~Akat;t*&I1z41wdlC2uLcI0LkStAf;RZWX-nbXmxj>Z}{l7
zU%Y)?z5kfFm_+Tqk7bwHJ1N^*)zk3<;<L2PsFS1E>$a3{KUJ%AjA2Y~?MtHU;|=3G
z=%>q0zP@|Y59+mP<I*a$#``t1)4k5o@hjESg{Hz3?OJWeO|9wn?KV);7JL|}X$vt7
z)U<^JL|G_6w1o!5SQtR8g$2Y}I6%CG2P9YoK%zwiBv~XtvPA}@SQJ1-*+m_T##ZB=
zHHd6#AMbh&=$WJ4w{}zNW?=h%V(9+iOGDdORQC$ncNIe)G9{T(v{6BQbEKKlLxkJ#
zsi6wQ9e8Dk2pcg~OI9}n9Yke_NE-<p=^%llY!q;mg947W(ZJCT8aT$r0LM5O;8+_A
z9P40#<7^yooPz_7xADO74jwqcCIBZm1mHxQ2%P8;fs<?!aFRm;PPWOw$qpGf#ijtK
zIFt^jn>(#?O8XA*ytx(gZ0cXVJ^sDo&#o%Di#t!7G{rV~YK5--^nZSWAN}9sN!mG`
zZgS{wYUr>cbT}<^STW09IYY&w=_PtYr?Ae(@W`kx-J-h|bXTNjyGIm=G4VZOdnODc
zD4Jn8exPXp*(ax8%KvvZu2<4CscF5_Gy49eKRzgPaF%6ga(1G5h-`|>E1ZsUhYR^w
G@xKAQDBb`7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp.35BB7F78B886DD68.idx b/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp.35BB7F78B886DD68.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6cb217abda50464b900dab0d63363dcc0f212b1d
GIT binary patch
literal 20044
zcmYkEdyH#&Vc+lXcbu)4-8St!{Px&x0wU5JzmFfs34su1?BmDrV~lx}di)$8+ep;V
z78+1ZLRV+K^_G@65ng0AsY=>Fy(O(nNgGxwkhD_&0i{(-XsVzzL~SS#pu`(h>fM#@
z&gV+ky5IZx_<pXe@j14~o@VLl>St>Y9=v&2djGpe@9Q}pJb3U|9{zss`xaJQJb3U0
z<-vpZf8xs@t(KL<Uw-O`zxF4;@;l#n{;&V|FaG|Qo~gg-JHD^}$?pdLQKkQeUq5{R
z#ee(X|8?`7jrouM%wOug|9gM<@z44G;xGO3UyaG%SiH?NzHs>CkACmRKl9WNZa?ta
z|MOcv`VW5d%b$AN-~T~S|KaGXzqbFf@+1H1OMmu}#6SGz8$S2kKmYCj{Kvl+dg=$U
z@AUrP>EHfq){p({RqYK%DE{kTf6GTE@2UCzUby!5@4ocbcVGSK=YLM^uD((E_J`i~
z1?{cVkAE%s+}nTgegE>Oe!KL=uYUIHiLd^IGxN>JC-T4dt8Wgx<3Ig}UwyFrFNKd}
z{@K-Qw{QN=3!Z<{AGx1CnSA@_&FR-Z`{dQt6JPq;FFp1qf9^AZ-J3tW_I~sU*Z)=D
z_1TaA%<ov4XW#kdXEGmYwJ)pxc=bfF5&HZGqhBk3tK*mj|7`Z*!q@-q@#2}k^}%m`
z{)fj0|M4q-<NNRZz`Oe&e0cjgzx`jo^F{S*gD2!;%7e$>_}CkZuUtHMTt2)}A3QJ|
z`QQEd>1T3J$RBx#U-#JSimzNe!f*S+r#z)E7aC8<zwi*h<FR)XU%7aMzxUt1@{@1<
zpM}>xAsr9#laD=FeC6U1UW~@?{mIIM2g>5-3i&Vo(NF)KCm%oQo{S$|>*2Ca-gCcB
zyQg5M)+yLo_YCaJIs-fJo`aoR=U}d`3(RG?z}#IonA>uLdAc4jkL3aLcD-O;%M0e~
z`oMga56s{7gZV8#SfCpK3s?cLU^fUBw1Qw4-3zb_>jEs)4S|KM5Lmby1`AtZut+xo
z7O^5=m)*<pqpz7KAHTFN-*dmzt_r4FDp<4|1&dlyuvj++7PDet@opR}ZpFb8-2_;|
zN`NK1NwB1q1WR>OU@0pFmhPs((pDNw>uO+{rGaI-8L*6%0n2u?U|A~*mh0xga#jv3
z-_3*Ntvs0C)xmU22P<?7U<IoHcGbNaKRP5l`S_J}^`856-Mt38wywd7-6B}gDuR`|
zC9slJ0yDY>m|+=U<!%|QY?Z+(-3nO6s(@9yRj{g61*>&yU^S}-W_C?5(=x&8-8xv^
zs)IGU4X}pQ0Bd%eU`?wD*6OyvT2>3}rh5Z+W8HwYyKS(x)dsum-i{v~l%IV3*1CPq
z{pxjlU_Gk`*6;Sg`c@xo&>esctO3}tI|LhAL$FbI1U9lpVB_u>Y;29e?z(qich())
zq&oqdSQ9X-Yk^so1vc$Y!KT&}Y}TED&8!*NygLV*Tk}`hW1)ZZ?3;fj8-DqurJdyN
zPg5txN%el6YNueQ#wpmDb_RB4oPnKd=V0f?Ihadxfw>G9m|Js$xeYg%NArMr3=f!B
z^MZK|FPKmBf%yy{m|ydQ`3*l<Kns8ci~v|r3xWlWAlQX=0d`?rfQ7UWSjY&0g|#qP
z*a(9~v<O(lh=5%hmsOmm3>7PCM6qH<3@dKLu@XiCD`_OLQbr0ZZKSa@L&M4#8LX_4
z#mX5uth|xO(hVJ}U=*;f)T{XYVf;$F%3&E-)%$g=Uc<cBu3=sq*D#A}5oS><!Ymp^
zm?gCYv!s<^mW&ciLp5L;ngP==447rL47042VU~?D%!*opS<xymD@FxoRjtCTYE_t3
zqYAU8)?n7O8qAtegK4TJOj9#qnuZCpuGV4JwK~kYQHR-38!#JM17^c$z-+2bm`$w-
zvuQM8w$v8PmezvVGFmWi)Ek&L+6~Mb;|6A1ZNqG9ZJ2GN4f9sLjo+W9ZjIaO{pzVb
zu%6Ka>#KdRzR?F8r~|NpF#sE?L$IMS1RJR%u#qtW8>?fmu`vd_Q}4j;j61N2Isuy)
z6EI7)z%0W8o2pZ=sWAncsWY&dF$0^cbFjHFf0aGDr;FFT@+@!t@BjSy&-_6l{F;}Z
z|NUPmYps{hWwto^n7BWGoovL(^Vp}hI3+*bh*R=2Tbz-fZNwS*xh>Ag&o|<n++_<F
zxoaa_<ZfHI$=w^_CimFFL+;rK54qPCUUKh7c*%XX@R9pA!bk46g`eEN5q|Q3Edu0$
zjR=qjZ4o37ZbXp$!WI|g7aMUw9<oJ<JhTxZ@~|z!<l&77lSgb3A&+cCg#6MLmmd@N
zXT!^lxO^U4wS`KqZiGr6wMCRXx)D+Gm@Q)Dv5kn4$88ZOk8ebrJYkCjd150H<Vjm3
z$&(wABv08QMV{J-6nWYfY4Y?&q{%f~Xyn>PXyh4NWXLlcks;68B1@j#h%9-|7CG|V
zM&!uzw#bv`HzH52+d?PTH$o>b*rGsQ*oXr8l`XD5CLSFp9zF%!h^yzZuWfNne!UUb
z<V9N)$%`9NBrn;bL|)p661iaugWT8%gS>2uGI@C;%H$PWRLCnEQ6aC|qDo%fh$?x_
z7B%wPM%2hnTbShLMwsMvThz(x8&M~3*rGw+*oX#s(-uwg=0-HhTefJCw>F|heq)Op
z@|%sgA#dBFP2S#!Hu<eBZa*gO&%d`Dar-=W&lWxM-bVDu`?lzl_cx+XKCs1re6SG%
z@}Vt;<im{^l8<aLA|GwUh<t2|G5L5S#^iUlxFf&Yh&%F$EhglXjhK*Iwy?;pjj+h4
zwwRJnH)2XYv&D>jwh=S(xh>}8^NpC_^Pm3n?|$o_{?P|s__XryOud#T&q|!Bck<+u
z+<STQ0{2m#yu`gyPM%S4{jGBH9JlQ_dD?N$PuKF4z0;jMW$$z^PuV*?%2W1ESIQ}S
zr(5Ney;Iw9%HG*pp0RhflV|Lm?d2JJXGeL)-q}hyWAALMoUwOiJI>fUU(0j$&Uf;h
zz4N_1XYc$d&)GX)Dd+5+Z<TZQ&TYpzd#<%~vFF-J7kjR~bg}0;N*8;smEvO0wN+f~
zxon4vJ@;C=*>mrtn?3hly4iCdrJFtXN^!I2-YRbP+_uBbo@XsR?0I(5!=7g^J?wdo
z(!-u-rFhu$Y!wfC9^2tz&%2gh_PjgkWzV~pUiQ34>1EHmQoQVWw~CiNukG-%=UYo3
zd%m6YvFF=MAA7!|^s(n#DL(dmTgAto&vy9O^RK0!J^xPn+4JwEpFRIk`q}fZ6hC|Z
zt>S0TZ#(?#1=cdaUSKB!>;?8Rz+T`e1MCG>N`Sq<Rtc~dupI&Rf@>LMFSwII_JVsE
zWG{G>LH2?xCCFZIs|48#+KwQ57i)RJ-o;K{uy?VS7wlae<pp~eE9HW{i>-3O-i7VB
zU@x?mA@)K$8DcNAmm&5-M;T%-v{FLsg|<qFy^!q)u@_#;Fni&h46_&B%P@Q4qYSeb
zUMXSr!doTGUf6bo*^8`YguTd4M%at&WrV%RQAXH{tdt0Qk*yM8FJe0)>|L(q<+Bpc
zRCe<6liYiG`2zP*UcSV=QZAoSaQ&@v`5d?HxP01i&(*b5*;98?Wl!Bpl|A(+Rrb`C
zqOzxM6_q{JcBt${*D}gpbSI<iMfWnwUi2uV>_t~fl)dOyiLw{99Z~jTYZ+rNwv#dT
zVtW~5FLsnM_F^j~#$Ifz#Mq13ju?CKwT!bD-^n<8@x6?*7eC54d-0VLXD_~0;_St3
zN1VOHS|->_>|}zy#9k)YOB`i_y~IjMu$R~>3HB1UBf(yBEtBjecQVOdaxatYC66-6
zUUH=**-LJfBzsBQkz_BmmMQj9JDFlHwU;UOQb(C$FSSxq?4`C!ioKNWNU@h*%QSoG
zolLWr-pe$5>7z`umtHAp_R?D=&0gAeq}kKfQe#isNsT>iFE#eGqtw{bR*J@+wpBFt
zG~1!Ems!gUdzqcgu$S4(411ZQ%&?bPDH--MTP4F@#&%@b%dTaXz3fhA*~{)_mc8sz
zX4%WGlq`GMt&(LgYdf;+<<>ICUT!CI?B(_{$6oFzbL{0-N{+qUR>`rKvmH71@@tuA
zFTay{_VRm~XD@%0dG_)vCC^@dtK`|s+m1YY`daGj={u>jr|+fCo_>@%d-_Vz+0(a*
z&Yo^NboL5sSzxcQlLht)ds$$waFhl13M-|+USX>g*elqM0()0$dG)NsD>XZL^-1o%
zyn2EAD6d}PUMW}4D7gMsxq6P<c3eH}xaZevdClJSPF}Njy_eVQT_5E&d)F)Fn!W3-
za?Reg?YL&IxRyosiaS|kueg^*_KHVYWUshVitH7)N|C*y?I^NWTFVl9rJXFXSK7-G
zd!?f+u~%9tCH6{NrNmyzc9hsN*3w|l*hzyuV=oQ%jH5KzGggYhp0QO7_6*x$uvcEo
zGJEBnEVEbM%QAcAqb##mUMXev%3Gz(UfFh(*{iH&g}us7R@kfTWre-UQC8Titdt6S
zm90`?uVOnY>{ZvY%3gIRtL#<xvdUidD68yMS4x$=>Q<?;SG65g_G)WcW3RT8HTG(I
zS!1tulr{EhE2YL>ZL8GStJ#hkd*)i2?3p`hvS;q4$)0(XCVS>eG1)V>ipid7J52WK
zYguQnzLRzK>U&vduYQzu_UbF8&R%`1)Y+@sjyijdwQR80*vSTajlFEJ*Eq@sdySRS
zV6U-N8tgS}M}xiQS~l5h?qrj_=3X}0YaV5jz2-`3ve(=yP4=3$qsd-tEnDohcCy7@
zYcE^uwT`mIUTdYa*lTT-7JDt*(PHmrEpOPn*~uIBZuatqy_=)FVee+8+^~1ERc_e3
zu^l(;wb!!EUVA6o?6vo@&0hN`+w8ShN}Ij*R%x@>wjFKuZrAemS&3Jwck=d=+<STZ
z0{2ngzQny!Zl6(b{jGBQ9JlSbecEx)du!QaueXyu_Ii8SW3P9VJ@$GlrN>@xtMu6G
z*^VB2{k80~*WbxLd;Puav)4b$K70L@(r2%~Rr>7pZAYKI!CDU38|>tOy}@1%*c%+>
zfW5&=8L&6lDg*WgwqwBFa4m=I4R><L-f%C6><y1{$lh?J4A~oQl_7gW+c9Ksw3Z|G
zMmsrTZ?u;q_C`lJVsErkM(mBY%80#@?HI8)Udu6i<DDF{H{Q!Jd*h=Vvo~HTWA?^d
zWz62#c8uA(TgyB4?soEyy}P}<WAE-L@7TLrDR=DMZIwIr?rg^$dy}=Cus7Ms344>h
zoUk`J$_aaul`>&(vQ;MRO>D=6J!>s3_N<+>*t7Q1V$V8Ci#=<lSnOF_#bVF09Tt1j
zwVbjy-N`9?)4iOsH$BQJd()LNWpBDwrtD2^$CSO<TF%&;?c|KT*<Q}rn;qqhz1d2c
zu{YZ)Gxlb-W5(WmE$8gbcXH0&d@twh&5v@<-h8FZ*_&^bIeT;4F~9FUdL8ZIJ5OHz
z6v3l+o>ZRuL-h?m_~lRk)Wh#*IZs}X?>0G4vCf=lSm(}jESJ-T<#xKUJWdam*XhOb
zIel1urynce3}6MFL97es1y;xz!U{XXSP^Fg>(Y7o`bW3IlaH%T6)WnDV#S;>thh6d
zm2f7olFlSn%9+ASJJVR2Q^U$QGgw(?7Axn>Vdb59EZwPN6`TdEE9ceg@odp~ja76O
zu}aPomf<w8%FZ%Y#aY3sI;&VUXAR4Anpkyb9joDNU^Sghtd_Hdb>qCjYCGFlx6a$w
z<I#Y#ht+rXu?Efo*3db`8aYQ;W9Jy_&UuG5aZa!-r-e0jPO)ar8P?o6e|0_j|319e
zz5L;vNB6o>^-aI^;m>}m|2gsSZfBmf@9%aeos)NAEyc-GeDjxi3iq^g3ios=PU)VR
zXK>FtXK>G!;*9ROc@Foya}M`>DbDG-Ocz{N#|7856fU}M(+$_%al>^lg`2L&^uYCW
zJa9cr;i2m_y>Pu9FI?|Zc<K5~A6#F@2iLb0KDvI>57*!E!}TwPpKibmzzuW)a05#b
zpc^!UaD$y7+~86K>0X!@a4$L+a4(kPf^NtR!3}jna6?NGq8m2DaKoK2-0)I_=|;>5
z+(;(^H?kBFx|il<`~I+c*|~fd)>2$P#W$~+DqOXr!c~_-r5iP)aHE|l+~`t7>Bh_$
z+*l_DH?|Zpx^XiOH{OZEjW0!<Zo*8!O>`1)6HAexn>3Sflbs~o<WeN*rpy%FR3`;D
zwG=73X)_Hs-AThuFGZTJW@>P?js{m-3XN{Y%)rfbGH^3Xk)fM4vv9MWEZpo;Wa;M2
z9Nb(d2RFAAIl6f>4>#Y*!_6;6o~~}{aP^K3S6>R9Zow?TEp!TS3rkU;du3j=?@yao
zovU|YEydMSeDj@o4fncp4flE}uIU!dBHUu92)DQtMY<)k1h>>F!7VLCiLPN9aE*=u
z*H{XJZrLouEqBUr%S%zFTQMteE1e45%2HJ5R?RBhYNrafx)fEqHM0h{)~Ug*Ek%v4
zX_|1&jtSRX3X^W#ti!E$>Tv5zQK#E58*m$)2HeI{H0U<XCfsJH3AecvO}Z_!1-I2{
z!EG%?i|&ni1NWwL1NUYrZs@knHr#fn4Y$1%ZMwJSZTr!~jEA2%>)gHzYbkD@;+qG}
z9^7812e-ErJ-U6f54Ydx!|g9cpYFgMz#Vi3a0g2<pgS~&aEF~C+~HCT>5j}1+)-x)
zceE5Ex?^(;cib7n9WTY0?wxrD_pWmX_iicf=uXTD+(~Bwcd`@{x|V6dwK^7DYbh+c
zQ*#P;+L^+gF2$7Y%$&iUb!Kp9OEIH6H|KEYojKh3Qq1q&Pyg><_`uu$;1ka}o_zet
z`ja@}zpnl|96sMze<Qva-%)?ZtN;8r{>K{&Z~B`rDlb2IJcyGQ`AGF3PPxwx;*9(J
zAkMj62jSv&AB3CRa}XYG??HIEeFx#=_8)|wJ8%#I?%+WLxi1dlf;)5&A@1-&gt;RJ
z5#hc(h|3rGNar9_?&v{8xnl<r<BlIhoI7z43GU=UB)L-uk>XAtM4DSW2#q^)5E<_5
zL1eje2a)5>A4Hy8KM0+>a1aIVtAn_Dk!Q+-xaKY%M3K965G8KoAPnyEL6o^G2T|d!
z9z>P9b`Ujg^B_#_`a#sW8wb(gZXQIFyLAvP?wf<S;cg#9oBQ@4ZeQe)_8@xP{e$Rp
z4-R6$Jv@jZ_vj!-+~b27bKf1r9rxrQCfwFRSlrWtm~zh!V#Yl`i242gU%l{es_Ac~
zUcO_iC-M6yq^IgB)|q;Sb*`Rcxl|XHTXkc3R1cO{^<w!{AC_PBV+GUzR!|LMU8omW
zAvJ^*R>N2kHG*}iUdHi+MZ3)1FI82+R80kos!_1076pr`F|e2x1B<J1u(%cnOQ;F3
zgq8qHs!6b<mIO<wDX^550!yoDu(Xy2(^L&i(=@P*ngPpb8L+IH1<Pt#u$-C$%V{~V
zyqX8gYk4qT)xmU42P>!ru!2?qyVY)U_YZh_S`VwQ^|1!p0BfiXu}0bmYpjj2?zB6s
zi8jHqGz)90O|fR$3~R2<UtN#xPY+)ediiIvJbG2=`~Tv*zwyowf8{4${uH2o@|H)|
z!?ig1(EU2qPr*(Xr(kFL8Q9t44D4J#2RmP!gSm7Um}}tzbL(y}_reY4(LG?Eg$K;5
zd%?U5FPKmFf%z6bFu(2x^Dq2h0X+Z~SOma=dJrtQ2!dVc7ho5Q3$TzL0t+odU|~HB
z7G8wGB6<WYvWS3P>X&c1KaN~3E<bd?R9yvA7b;j(kAg)PQLvaE1B)$UU~xST7GK1{
z5_$qGu}FX=^(0twkpxTWDX`Qc1(w#+VCh8~Ow%<mZJ~i>^bA;Lkpau<S+MLP3zpM!
zV7WyOEU)Lm@{2r}uIpg>LI*481+c=R0CuHcy#+4=EUrFuzpnLbu<OM&SWz#66&FRY
zl3oHUElOa9Zh#pJ1FWo<!ODv=SVgaZRTdSns$K=FE~;QPy#`iW)WA&L1Tz;VSY5A!
z)faWJhTZ^cEE-@<y$RM_G{IVW3#_$hf!*jgU^j~!u(sX?YcJYhxBBf{?#~Lhi`x&~
zub$om>n(a<eZ3FXU-ZES`T%UO7=R7+A=q#+1RLoiu+d@!HrB^r<HZ>4PQL@YTik(7
z^a<ExF#)r53(Q(rU{iexHeF1?X8H_lwwQs<^*PvlF@KdUM)TgwKd9`zi@|$;?Bzc=
hU%q(dmwf!+edCjp@4ooGKb5a>K6uTezkcxG{{w)S;6VTY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.6F38C058A13FF1D3.idx b/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.6F38C058A13FF1D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..85b9cc6bb74a225c1257fbfb84e737374efbfd86
GIT binary patch
literal 1862
zcmZvb3p7-D9LFz)Fn4AQGhQ>y7{<(y^}6yVkH+I};fCyJF<vQ&WmNW5(nDymE6cJJ
z<xwrIN*O82<KXPBOeLlDDC{OtGTX*@?5#aB>^XbSx%YoQzx(_C{=dinkN9x8E*k*A
zBiiR+LX3cq0RVuAzA1v_-FyWAa8dz)gJlDO&-PJl)rMD9xQ+3HshF4df~&Z3d$GIa
zUDZkUEpS#bu6KYLK3chd@EW=8QLhJ9*|d+^_iG6SdsO~zarK;ghVr$ufs?b#CG4>d
zvAIi2Tk|qB-ZVe|lSbEH*tAxsW{dE!rTyoHx10iQ3$6-R3&?-dVw=KOt2zeTb{$$+
z;P)jKo$j^g*sTgf-c8p2Xg+`H{(+@G@}QsPXtfXVAq_*J(WxX~&pZ6w(a5oOmFxKq
zR#OS%7eaI<;z2wv^u`dorW~qow|&w!{NQDf8b4k9HYwol1&hp9!L#$Qz`a(W!maD?
z+?YA$u`(L*q?XrWf^8vPOv|q(1V<ar2##gdR9&BMVED9P1cWaijil~bm>rhH_jKa9
zEZsPNmEy+N25;CC)=_>8ldRD4Xq+l6DjmsfH?AOMN1S|+MOS~?r7t!k%axAR`}xf#
zc~Q^Y^wzXI8Ei~@0R-$X%pIt;>Wr3rn^yFYU}QRYENNt>KG9HcBtJ`pzn8Ol+0LXb
z>TLtj4O`8uEbt^=-qdt~6`lG`;6b*Hu440Jp7z#EmR<D^hlXsQOzXvPnY|~*N``Ek
z!3xtqFIALZA^Pgww{+(;>X76NGjVATa=dq+Co0iz+zU9Xy79)Ti<5b0&mD-XJ6y5S
zF~Bq;qy-UpITDZIu=qXGkNXn!_HIeCZ^?UqbV~rEZs=MkkU%2iH&e3X^mjcE+%Z>K
zYQE2?F}E^%sxXuk|D@`+S+n=yu;y5!ElZEjpMl)NXt%|j{{BpbBXqx0BmvoT`!22e
z3(@%1?jx6uQ~i4!yck+(SeheOv45=imoD+!a*=J*<#g*_c*G_|$9xcjd(CG~l-8W-
z(Md=xEhCR%<B+WZJ<Qa;rVRh;juSU5C%nd96^QbLIpyJbeOz%BvTEF%m`-Xi9eIBi
zmk5IeljVPBNR(QfCx*I=E*`Dh!+T%<8@#f7bZI&w5!+$XSj71~F5BtcLQ3JA!si9u
zrJ31`sp*xbn}=qE&Wn1zH5~0+18f-Y!(oA?LzL-ENIivfda68RyzHhizB0@+?VWr5
zhm$9`trY{4t}~{Z$<F%c9t)gisrlU<cjkm_PPb*$>g0%X(8ZdfWx|J6mw0GLmut%X
zARLgAmJmrpZ(L|BSJdw+q=0~*KY(BrvGAu4`sd##IVLtmA`bvcG+~}cr7c|^L~sz!
zRwhFT5J5(&ixnfsVVmu!6A2=@t~rq*vWzGa7t^nkr#0n3gtL<qCUtT<^%__Z3x*Uh
zh_14(uhfh|=*skU#8PI-$OBnqe&02}ibgWR7;&U22X0emz5LEe5jBG<AVC^Qy2`13
zsxY@*Qc?HZksk3PYnRd}eMA97s4$f$^#TzF%veVpm?NV|E1wn(eyb2Ilft9$*CcEn
zTShTk-R@*ut%*eAszGWFYfeOnxQ;X+4H<of0ZJ}eNcvF;1*U|oNep0vb;O3*GD`Pe
z3@flO5BgXu9kiB8a>vX3%+kMzK{KT8q{@1X7PXThqbn&%*ArD^JW;J4MvpCRUwfd-
zR5)$08_hx+qs@?p1QC6V{yJh}m@=BQP!w+<Xr2C;1sSACC9nAOTV~LV^+z)V%1+7r
zQ;}2fq$3mEVEv@Q|4bT0)=yg8Fn299L&QPr!9#f2Q34TlNL@xtIieY_L?{&lBASq<
zvos5I7g#rNjp#6a9kF1Rj8^VEdl9&4CK%1vfM*aTl>o^xd-)%`lad^jqJ^#s$dyj{
zh7Nsef4Kl4K-xC|h)Rgw8;2qsyHx?BsI);@1q<TT2zVdj77t^hs*k(6mIg^%$Ijl-
z$=Stq^JipJ2g6Nl+W%QKDO5eW0mF#N;+SkSx3ILfvDJrk&8$4hnqI!!<+<D3yyZUs
E2WoEVrvLx|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.0B3E6146A79A7808.idx b/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.0B3E6146A79A7808.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5f051e6ab07dfd58ea370ebffc7ee8a0ace5ccc5
GIT binary patch
literal 1862
zcmZvb3p7-D9LFz)Fn4AQGhQ>y7{<(y^}6MeM;?vGUExOA(b9MqVj0D5g&sm-!?LXR
zyOvf(NGNh}c2`u=Qc}?->1DR^-dlTS(m8w2x%ZyW@BY5u|L_0%|1RRf<=Sij0EZx#
z_{d-Z9|HgY5xrvtF?;z^0AM8t0P#g5UgP^IW(p71<hV`Yqj8vL*LPNN!y>SIC2bW+
z7VU6qK5k%y>Ho0&!01JC*PQ_etgQYZb?}b@3N}gdT7LB#`()XRC0^4nR|?pZJwij9
z_O7-S=t;}`{I5#A|6mh3t=f$I{gw|@v|h3DyehcRx|U6TPYY@BU#sdF?Fv1#u)rUT
z$~!e+!7*QJ^?5s8`-|cH$(sk4pJqb8NziI<<9!;(e1qahu8tl2jE4b}-Ex<*EKO!2
zpH%v&O@)JaoNx0OyQUaw=r+68_2AYsZv}p$@O8B3wMwIuPQiEuEU>Q?$Z_k3Zfs1M
zbXa}3=Uy$ZT@Tw%s!PbKChQE-o)a8Rt*N>+-^g%j#|Q}D-kFHoxA5}8Qh5IjJeQ>&
z<{_8g^g`<ud&)G>ouQZNTO5RwhldL%GP-rkNNIac+)AY@{?)4~G$2b9PByr^zl?UG
z=3I7GHa@Xam-ZO&Jdm3)QfqP}Xz9m<yxRmF{n4Y*6LSqw+JYllsl)i|>04LK^|}II
zHxliz)y(p2N8-87EtRaGxYJ(o>`m&@ZFhO9ZYeDD>YopdncbV!2<9>ej!zbhnYDps
z`hTA-D?U$j)wpSF&uLO4Noc3w5^kkC@2w!p(3`J&mdI~xK3O-NSyFm1tp0G>YR?E$
zhmhb+;H57)Jn#$OH+y$5N+V)hv_*U7yQFQNjQX*QH-JbI8NZd17N!~c$ZOY|@<PM?
zI!zhnX*0RLr0{!HR}I>n5Bs%+=xkfQTag3V`_ZlnIYUD!Qb*|SCrJXb<Bm|3hRPuP
zTHleg$EY6tmQD<n1T1X}S9)kN|My<u>*8UvmUD@w1MtKqA2q{K4DJP=IaOGb)2|j8
zS6D=z#D*bmp8d?Y!Ior?>Yn4x##2s{&$EX!ThojEGY7fCDr8N!Eh>@Js6X+p1Q!K^
z+0(`Elb2-Lt*6F%b?TDp_wn8}{0XnFBrVVGiNf~iHRW;choxDSF2v@(%6*jGSD2E<
zn3-K|xqQE&&3aK|poXKGp@sG1eK;&IwhYvt^J$=PPR$hiJSn=Yi!b+cOn7VG@ZrR9
zZfDuZwC$X}a*VZR>0N=<OKMhMPsf{9HmA=xaBX_R+PkhMZ-wxo(<U6+)oU9&?u`Rt
z6Cwj>=pPsT!>}eVL@^!&+&usUD~*LeJ<u=zT`|ESu}hKwAVX`-bSO8YOMnOgBAAI}
z@{45?$KXvnBte7(k!(d_5FtZkF;SLm^uLdpRYtX(-5ftrlHJJ{z=BXPCWS%NWz}6p
zVGKf-rLQBFEK5ud$RhKn-uX4uNe83D5f$l*vAA<|U%j~$_=OYrg%gOZck<e<$a;3h
zN*Z+n<v@aH@jZ`ILuDS)Cr~FUOy!BfAi{td>xct$#1vrS(yqksY(-l@;ZgXXC2SsB
zOu=q;9myAJ0#IKCNWt=R5)mS<BPB>lOrjhK`A;X3(Y;hw|7Z*SI}A{8)@W%6b)vu&
zA5j*FXu(?Rhz+yFl<2(Zmu+O|{c*2!&{QPJcv9qUkoZk7S~zt#ReUVyUUxIZbUrrv
zQj~nKBdXQFXs|^c@Ak5p@u!XUp;f42R2d>m5YfbFt|KOfDW+*7Y2gNf%Bhc4kU^SA
z@|;h<VgSvVqE&zpC>|^j!9h4N2^-&B^i3Y-ph<WLFCHuqQG^u5w46RX=M)9SVL(I~
zQnnT?fX)Q0?)91IFnt}dV3wFxJH{V-ZJyhSR;b0(3KU6znBWNh$BoCv1jeeMZz<$y
zj>_pXZm&xvJONKpO8_7+GAJSpML5Jw3L`DEK~@e6;uHvY7veStU81~;y`qW|Nmb3<
zVvCivjqTR2$oiJro7uGgiz-v78gwm&4wJ>v+h}NHY`V!z6H+%YaU?4{x$cnUZntxm
G`1)Ul)zTaQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.487F7DBAC716F5E6.idx b/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.487F7DBAC716F5E6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f12a5373a740994a81c8ce32efe29379575ba2cf
GIT binary patch
literal 1852
zcmWIYbaOLdXJBwn@vO*AElFfyU|<jg;^LB`%tU4e1{+QWhKe~?LvH6un+x3i&*`=$
z`+6zMi-TcH+%j^wGnwp#L@alEcTDF!c~v$3{=%~Bn?(0tIqA&Jp?yx~+?kou+zm_z
zrZ4^I)X1@EUdYpTe`jhuIQ&K5ZukD3e@z~5|NMEi@X@c_mHTXV8pOx_DO<E-mrdZF
zl8rn5PZIqupSC^z|B}Pk_ovtV{F!(ze_HoROHGUaJ0d<kT|Qg?bNl)7KhOG1)-%d4
zzswi0;Z}4?sfd@$-o)PfNe>QjZtk}-dYSupafIZPY#u(|=&iRjmd!O;d&vC8{=1i6
zgbF0qfB8@lcwn(X+rE<9i@ZymmX~n4t+{YSr|p6BzxxR{mPhT@;@&N?qOyM}e^`q0
z+mfb^WlJ`H-k{{Uo27(5_R77|te@}h{?0yegwIV~DkFe%`j$6}?=_wnC;KaDbwtli
z;pOtaIO|^TA+-e}oe8a%IuwL%9+mr|E6O<Q!CHU+cLnY;lec+@8McP0$v<KUESu7M
zb-B@zl;0C7r(Nb((Z1ePaPRHfeC3k5{*H@$2fG~pT4?Q0{;)yNk$b7?!bvWItIW19
zR!=GI38~O9m15m_ElR?-P2FPYjG9~KH(tx8x~ZOQc`);q`A(h%+Rs-mn7dBUOZL2>
zljasl5k}=U-pWf|9+`^-*%Y=O44lWMv$cK2)4qB0%QM#0F8Fu&s;UZqWhj4C*Kg;$
zaoJg~ubs=6&2cTT+}-!B!8K56&8<yG7;;5K`5dG>Gvv}AhJ=4yI7>fIWlQhE&X-f7
zMY3-!*`vGDqc(15nu_b6Yl|kEIK|2D`J#E@LK}0Pf`7Y6iKt6(y7=0~DSZEr*R5=p
z2{>Wpt|VT`EpP3{df~zJ(?`F2n0wKD`>J~5liv4CBP8{&v+%x2RDCjQ+2j+Fxuvt_
zh(6%X2=Wa)p;~%wdt<=T!!274pSVAGHR)pCj;^`!edpZ1ED8Frwllw8WP|p-Z}WKb
zy?G`*o%_A<H`{L8C%2BOtY}!174>cHS?_;;8~(gb$mc$+wPl*-os3SK`9F)Nyr1%L
z((zesok}lX|4YAK^LB^rFWHmJG$nc!x#OaK)Rq`pC2PNpSSzjB{c>)^<2l>Z_!h>w
zRDN<=`=hnRZQp{cPwn4oixt_*&A(P+^G>G!_~E@DcW7uHH%$Kj^qy_#ie=OO^8eUp
zmu(V$)V}z7C@({CWp0u@Fvm?=%CSwHGvGN7gMR=+5H~BgH<$p*KtNGyTJa|)1_n0y
z9evIV%@vq=f_O}L%;DOCcujbbc>E^(a306|8y^nq-bw~)k>Zf@LTFLoP(b3TbEw04
zT)!6FYxXt`X5tAFF%hvxXc09Lh4Z9;+i6!7y%uBS3DS(wjDsn1Y=6V>^GnGsW)>9d
zSb~tPV+lgG&N=9p>eQp3{{u}@VNub9`8mW-Nd3h`8&;q&4<`>l%z}hR9qDZM748B3
zBjYU-1rz28Qu0<p;%Rzo!g)zXp1Xw;_w4}sM><M65ursRN(0VI^>y6axN%t$&|Co%
z0V{+eK@&kFp0J59oL9)gFmt8B?+ZXJ(%#Y$2rY`<iby;SZw)xF-s4x?Bm@0Wkf#)Q
zjA2@OAJ6gEt&dCvx=<!Y1|FC|PsJ#~c`(I{T;Rmfd+7IqH4m*7))g0Q&gV*X0U9aG
zBC7!lzC$5)FXQE}9|t;4f<;0JrkN*5jztcMr^=!V=RGxG{ld;K-VJh`D33f$%a;uw
zH$^vI)C4+?&x8-20C<9gOoZUPKV28!y62mevhW0nnTXlK^a2x|w^Rs%r{JxC#8dZH
zhx7jJz5OV}>}?p(a>XdcWSEws)SN_+mx_y$i^YL7bI@a*=tigD&9#hy41qAuGB70P
zrsQM*x!!5M%q*;I>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!
r*hIyoWn>i;l~h#KHMMl~4GfJ<&E-s_bd6j@#oWDuncVywJs5oeuo=}i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.0D311D4C9F7E0828.idx b/.cache/clangd/index/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.0D311D4C9F7E0828.idx
new file mode 100755
index 0000000000000000000000000000000000000000..846dadfcc816df2d38ad7673fb0b0e6c16f82f9b
GIT binary patch
literal 1862
zcmZvb3p7-D9LFz)Fn4AQGhQ>yn2ebr>vhW`kG#g?u9ykg(P9{nLd2--*^(Y!ElO=I
zOL<dED-jaP<KXPBs61AM!Y)ZKvz5H|)}9$UXYV=p-t+n0@Avon{eS=e<+`!iwwnOJ
zQQ&qsI#kF<0RVtU?s#F`UcL+fSStX);rv1GNBc?UN>A1l*maS^38<O7zNPGlDD+-w
zJ7t1pGn{r2(>q8H94|UJTuW?w*z1UvH|Qhx{hmugACtawvFwdQvV83&@9CE-xy*?U
zk&$h4Thj_O*|4zii%Qo&=)_j*CgZ?><%7kIx2!q0h1VO`a)|FKVGV(6r5(d<;omPV
z^7~>gobI(`S*$hsy`8T3*=Qm4-l65cvY=n3C>0N|el?^1f&_xQQwu+HJb0pA;YPNV
z>1_05iJ!()B#6cM*N-yG3!v(D^O3eE_h)>R_(`JIv79?4#wo4BN5!zvp+cy@uIlgH
zk}~1AIvzAq!E4q>Hxnuov&(S4g3a^7<7wrkHx_DWZp|nm?%RiB3HugbK3R(F>BO=b
znh{<K7wcZ=ykbt7g?Q5R)BFnr7)5v>Z!EK2uaJ-)bn1Q@Rr${@ZIK~SDsQ6N)AMDl
z3;EnlS5=c!zIv2r0O#QO%)tuNPQlWd#0w8_x(36?W5?#JV>SzqW~U8c?`G^+vCwY|
zd0m6IN0-rya-8s2w>6Y71PMQQA7*aVlx-T~se7a_EXscRe$;$qPAim6?>#w@J8Iqp
z78?A0rLf={-d*dSi36)ngCMmz1(SF`!*y>lUXEISmvc#ROMPnPbk?QIha#$u6s~p*
z(sgl(J~&>+lH-$r$bEA|eKA^5&asxwS?`WHb7)ngwVgmTfr#BfN{`SEf9k#aO;MiF
ze%-pvqV(DG{)EVp(%Xhju15lz!gQUNhl<ZZ4gr+gB36HYip)`}XDUHRblMfJR$U^%
zu5};1a)RvDW933qOGHz)vt|1yF8<afdR;JJ-f%U^tQQ{J>Zf5ejKaL&)2H&v&-G|T
zC*<W5C(sdG4^9s~p|2s?tE}T>y~&iz#Pghitj3Ijz^p#DsFb^=*A$aPs4*COcL@^%
zgE`X$?~|A0nr)^=yL2m$Rqf-wtNsICT{*Tq7Zii;(676|dK{5%eR(nd{Hya%bGq|V
z(rL4EtKq|stD9^-YW0@0)H8L^0lW`KgeF!Y2J?Q^B-ZKK0>8=pn|j!y0H?&a4%Ht{
zon*Hb4o=(68>q(FXkQ)@TE8S`cXzbBX=JjxO+wbD$83Bm%P*|pKD63KLc6={;wODD
zKzw3!Fa`PJILunF_A}@T5b*Q@xM*24{P}@=`R|Gg4U1or1^_uqW0qr)IaM0uf)Hpf
zv2k$_PC|+w3&&2tTX!KM0z|M|7ZD+%gh)%a2H(ZasY-!d)*e=XSY)620$3CYM`ci4
zO?gdsaTvv=%2PKGL!KcaNA5@Zk6jCENG4sBE=%09*ea#?c`2FgONCWWw^O}$Dl=xz
zTFWA8Pyxh=Rer7I>TX@zIWpiEb-^#{g4~UDgPx^@%Z*dV5DOV5^Te4zE)Aw_AQsG$
zP_U_6vkJep5t$5$N8+zrm^`M0LOtwTlCPHsBYBk|C98E29>Q-R6-Y%w2T(xn72~CT
z#6p5ee(M$;SZ4z<VWxzVTt5cn7#sO~ngbOy6I(JT^F0ldz6nJ-B<~?hwgp+#9-4%%
z#mC-=Q4DoLyjmzNrg-A*-nO%Wl;Liq3w4w_O`H<sYNNC_5FJIA(6q6vXcJED^rtR}
zAVq9>&ZpiogyzkVFb2X%4wk6qO|5_O01Igj3t=Tk3FIn6$`V@67?^j7ff7(4R~1sV
z5vN4<71s1#Cn`+cKn$26q1Bd0&%C$I`y$14cse0s3lJ9?#s9R___&aGHRLVBrJkL+
zS$sfKU5W#6#6tjpkZ3_v1cGpwhYU(qZj-zM8pJ5!ux@x~M?Jivn}f2N3PD}N!g9N{
zjji2|uZRX#o3}A3{})vyk+rBgG+jD_rN70<*u-qBxi+L}XzD~%b#dP%&E9G6D)sfh
DioxG|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce.hpp.536083621BE180B4.idx b/.cache/clangd/index/device_gemm_reduce.hpp.536083621BE180B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02cac28c2798749ce09346313dab94531493d932
GIT binary patch
literal 988
zcmWIYbaT7H%)sEB;#rZKT9U}Zz`!5~#Kk2=nevPb3|UMJ3>9<s_8;_OHWYDv9=vJA
z?rZMpl3tE0y~L*~>D}SY{A4Apc9Pq3+14rlb{DScNM2~ID6pgb^v?A&di#R)qR(v&
z{t>78?as}k+Pgd+nxrjGwrtT~Vb!_3(s1`3sm<N3l^YMto%W;U+nhg~*EB`?PjA1<
zobG(&(n00ojxN=tYLhn2J6Y2mo6oS{lGt+8YfFXrt8<IIvoGl-zSz9<c3}5I>tA=K
z|5__r$)l86G-q|($)BP&G1IuCb0;u6IEyXpbn0AD{vfk|p^)DU=PQCLk^X^C=7!FD
zbfKZ_QVEyiHiIjUZ>@hA&D)=)taQ)p-v9o!=e~z7wEr-DN=?%R2i>PCf-^fZjdT?S
zR9__hn0RN(_n(ewvy!+pay~K!t$tX?P+Xat)Bp^x)RM3sQN1bk0!(b2KrRq)3$X}^
zGjKpSt+%$#S>CX2B{vfb2Lrbzo2Dd~VBll|hA{&pH{)EP`HK!O%h|!k#0*rx&(1Fn
zCO`^U!3v(v{bR4;#k)*^i5+Hvs-~(n%z~oSwBlnxvtqYf=ESI}C^PbK3(5#;z?=o6
z89Bkug4veZ$Ew;ORO|#)&9BI>3lpw5{P0N?%f++oJlvwPqDnB$Jlx{4;z&F>Sve%0
zysSJDPeE1ziKi&5h{RKpRf6+i0mH}z4iHb3vhInyoa?&1m|NcV|F_`a;g*n*P)4{w
zQbrPqCnY0=#FLhhM&ilH$RP1#Wn__fa_Vw=2s4#sl;Lu)AYkGGhd<0cJlvW(nl>=U
zt~Is_I660BIxyH&WmMH+QS;P*^$RfKx|xAtFDNLDus~H%RT`!UXp*9)Hq0cLC@@|?
zu?e%iC^aV$lnY_;!vK?DC@xAaRtBXhE)W3$+|$<I-^$&wauX8|7XuGWlz{;N)qG0k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle.hpp.E3AE2B4768599A5C.idx b/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle.hpp.E3AE2B4768599A5C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3e2ccc81a17db29e861ccde08a338a20ec8041c5
GIT binary patch
literal 14556
zcmds;dt8ml|G=A>=Xs=_PN$p7>69*_I`y>OPC_W65L#c>u64J&>}C;mUDurymewWG
z61k+%h0wi3DW%+{O{823C6@~O8#=Qy``7RP^|EK4x6fzhnVDziGoSg)Y)D{0z;$me
zE&myTzlP3SFqPBN(lWsRg)NvrYdx=}Rc)%J^=nMl_^LTo-trq^Iji-n(+_P@HiR7P
zJ*96%?yyE-;kG8lpfA&g3^wGq2R-f9@uIKc!UJ=*RLwx#8IOk2x!x9K^I}H5_)YKP
z64TU;YsHasii7RiYiwIT-a6$r%K6o-nXwb<x0%#!`k_VdFSq-pI$xr)q=BCI<KT%2
zX2HFa@^<a4xjf2i@5&o5x0-z1vAFBZD_U!!>>maXU#Q&Z*;ULtQB(tTCmj0o?6k?#
z?8EDA+*BC@Kd3S$eMlVR9&b2){oI>&r3X*^(5EBwps@6@>B#dA7gw+IdtQ0hGbPL7
zNSD-2=8x>|4{+y9?ETxk>$hE6F=g}4z}<f8U$%$tKKwB#Dyh%L+Q7g)!fM~dcjXpl
zf8;*cQY?Dpp3?cIOYrdfhh>pLQ|7z79!{NoCx5}DZ(erXNcv*(^_@;vcA0c!m}<4?
zFD`XYe^b3OIO)vZ8C_RI*@Y<VbK15W{M^kh$bE51T+$np*n2Ved6)UVN{hOF+on{f
zNR`to-|2O{AGh7_OV@bS{Qq6(_BM8mbDfF4UfA{WiDL}qUHhtC*8g_1ASdK}-pMn|
zd?)D|)XCCg+dDSwoH5n)`1B*?x901dp7`LwsZo;t{5iK;rvEv4sPoWqKfgHn=!dK2
zjR}u_xLxq#=*;9t6~~S}N*h-apO=uTsy}w)#>*WFy|%<5lioy?R~<aRbKb+<-OM+Z
zIBnh4HohrftIO98*Gq059aK=YQ{TPO=bMW`anC1z`*Kg;>kEGUO<5AX@t3)CmEVuq
z8n&oX*|0)Ua=<+2!-0eY@9wl+H-2Gz{d#DL+m921PM`VU^HZwq8~@nQMgGywM|)m&
zPTW4dZ@J5dl&}@T3*(xbl{?(d?6=Hox_f8pg-P4``z@_|t=(jDQ^%#}X}2BAr}MM+
z&be#9yLjA`j-I<pmgMfq>u%qz#OvcWm1X^R`puKJw!b^GsJidjw{yGya(e3>=Ve2W
z4sU66`CBi!Z)wsIzYeZ!_B!O)8gumZ7a5OKCrA4iC73PknK<&9@yRWH-q_rkEiK(W
z<<8>wv12DVeLOJi_275?7yJ07`(1E)>c4hn@zbVM_j5b%WlY;NPJK7JD0F+-$nU>i
zJ=48!YUP>S;SHPm&)>b@_2Y;|W@QGZH?CAj$zfL(B)B>`FHT!nsr_U`aCJ;uxVx42
zs{`&^Eelap*~Fyn-rYjnPmY*z;`vm#)ZO4>zoPn=%58HkM=$?=kmXnBbc?5lt&7N-
z@q@v|73YllIo>;>_jY!=gU^rqpJaczbla)dKXtiM7+?ELweGv=e^hMwbAR<1tqrTj
zF7}#e?XY^~V(k@HJ)GtJ#~qG&ye)7?(AD+B;|J__IM{WJIAn@r;0qnYk%!iX3VrL0
z*8MQ<oGLSDx8JCV9m7(bM=8u4a$##-*wcd}ZBFZZ1sU=eE~P!6P*@YHt#4R-y-OG4
z^2V}bKO3HT>ltpbI@(~8YMS`-H!G8OJ*cbYHoWYUTXsI+>u(Nu{%-SbMBtr;-KQm3
zFSan8Qn|<K^n1UT35v+2+Dqod#=Sr8;oUm5-%VA>T#vZlKQ!MiX<yWDh45F^d+D`H
zdu#l>ma?_3ileo+7k)c=ZRu;1=8X@w&QZngzF=pQ7d-FL;mtP26a3HRIZaztV%y!$
zGVBTZ)p4_tx9*8wPE4@66mr4pz4wZym%>F5X!A6u5-v8#dhhFJzSrKmX0AhHs#e*n
zh2vFyFH~O)TrsS0Lg<0OQ>EwY^txRc_vPuy<v*Wt)PJ~tdd|Bp_jLB?7psTaURvE)
zllga8)`EL8t&YXLd2Tj9HpwCU!3yYLY>?*K+}&9<ttqIcI%G*=tvYU6`>(DG;^lWo
z{*gHCgJ(?spSornBJ^`+=bNq_y?Cuz!xA*=SxflNQHQ-8e=B#I`Qxw+Ne;IBx}iI}
z9jg?(OiKT|%dN3x{$Yzlrx^%ZT3~QiR79_Q0|9dOT2dO)O+!YSL`xf6ZTtt?ov`WE
zIwC>_y?SX$^~k&)*=Q1?(&c_ay?UlctQ?x+QjQDO#|3vyLR54zV~Z=+E!iZ4-nf*<
z$ow(ttx1TA?(-|UMmB%*CQuibGP?U{i_bqqEVl^r`)hJ|iXk+_;%HtT&0A;^qS9^I
ziXH7=41BH&6<DPW>a@Y`nuMrW&jb;(r};(7p*~iL;=4xSQaTf&;uN^J)O75EjdEy&
zRUY!(L*7)A5EYN%0jn2X`|^McTH)5bLYA+PlO`c5gJ%qIJ$PVPxg2)Ir6ddTWI?G(
zh)VFdKC-l@-^`Xni{4t&M<n}*3^d6<Dl=6P;_fFxr^-7{_Kq`<IuoKYA!z8-i3@Cx
z=|UdA=q=QF3zeFLs7wk--(7TNZm1zN#wr;=o&lUS2~n9Ex}r9`+H<iCDshWikYx*Q
zQD;I_ru!rn_=l~ukwH7G@*IiJk&h-JDl?yiZC+qB@TLs*#VW6n{cF@ulMt2fUVjy9
zw$Em)4EDn+Z&0r{$XAmPmC)<AS2b|Ec9=jBS9eVJF&3YHh<I-A_7j0W{<TJcGAwRK
zy6wnNlMt1qZikldysXt*7jn2~-y_+3)J>BRm1Vn!MRZNl^Or$ST#rY{{1G13oe5D{
zt$Y8g`EPec7(z4rqG~9whK`zqsBACU*pT+Y8uy7Bm(qZY8jyu1Au3T>Xi`7jKPn7A
zkGOTbd>wBqbtXh*kDqh*seie@(1kkqMem``dx)$1Ir&E=-Yv*H?Lc>3sEt*Eb%G5)
z{}A!s!CT90dqvJRgeF)V2jy|FmnI=9F_Rjsj{k33hXA|u)RIPY8DaPNN2@CY=GYKO
z;8^d3U#=g&TdfNbF5?Z<d4mU1XF|$ItQ`CEk&x^4GH8RF+lnk(@hIs`h|0M-J&&dC
zo^xf;2dlh9;!EVBNr+13Qz7_ZQ+FE^*d4!UH*db1x6vd-C4ZXL)a_n}`|3ggtGt6c
z@1UtBAu6Q_q3@I?b3PhE1FSMuKGym34-sEYJ+2zG^IK~f?1Ag?6j?k)ZkmLsTr&tu
zjOuZ5iVRv~m1gAGjD~6wqH-so?(K}N)4R%`C02QYES}&}IuoK&rCRYK^PKDVGS~;J
zG$Qjxq|zirrS7-7qx0rTrrjWiRn8#oGf1vUh)RR&qUwz<4Jt#}4Y#I-m)GzLO+r*2
z`xOiKZ*0LYvc;u5L*1Ssdrd-Ao_=&i-~IOYQxm9%O9|;7V)6Nhh~F<?Ry^v~&}Ks@
z$Kt=a?tgQhnuMrm&7b*Qn5&Ae3|(|I>j4=*qNe6#=@9Y6vVD)7$ea`L;C5WcyOD7(
zN;`P<4t@Z!yrNXb>zDDS#GZ>%w9q43=uONnS#O{yT|}~r$dvS+j%L*%FBUR9d-NM}
zB1<kx^X2oM$aoN?#qz}rZ36O5fZ;vBVGkHgwGROK0al(2<jE}G0rES*iI`APx)0>{
zS#AdMW|mh$hgEPe)!zu^8zGZ_H<a&YIRhGHKo>HBMX3-P6|!6ojmlYm3XPsZ7h;>p
z(hGDn3oiLgIzI>8EZK8;BNC-3;1dP>X^XD{&#S<f*02=%ltO<}1F}-%{x3qZBBZ4K
z-??m){wM1-)cL$llNXE9e$I10rzWeLp6kXvvv+(^)m4;6TaE5Xa*5Iyt1&%^sTZXX
ztB{^#_~IePi)VO4Iz+O0F`MN=UMysJl_0JX49O4@rEozEXZd$Q{GBo>7k65xA3^Hg
z&1q`crQnQq+-a+T&nn<YJ1r0T<Uv2uNupE-J?mh9@)?Ndfpj$f1}RaLwjk{-$c&hH
z5t|0>Mr1~;hS{xT?e|C0-SjxzZGHB0KBU_?se?1^;LL(G9sq}@k?juNj!Yv?%H&Nm
zc{9=$PAcF{3s^4V#Ufsh$}jWcWtK~Mv6STvg6RgqjPxNVZ4pejupA|bQGy<o?-0Zt
zEbkJ;U6ie4tNMg%m(~J2zd1!T*PAqqGyr#gr{6`o%KH5L9~xKoTMsy#qgym*PIA}`
zNITc!eO2c_*GKj}9^CTXlN^lu=sh%l4}-`kajcDB)O=y6Ok0#f#ZbdeM_fw_n^qvM
z0}Ss3;!c(mftbkhbs%16xe<tsEPnvv2bQCt7zO_+pL}clsXE<Y@~Mdbi)};&l2srD
zwGo}pFtHJR+}DTvRdD>ZjwmfvFAbs7DF!HFfEVS1Kz$G}@>HNsW%&V6KL8=rX03sW
zHPDOlCaB&78TlTl-ox@GIOr0bNT+WFR8&AO!akPGmwyzdHp@Bbj3Ay7WW;;tq;x?{
zXZfsPepXOW`8mP-9Lwhg^Ybic2<90qX9`A{f(gx^B^YI~oGqAU3+7awBber}oGU1D
z1tpc|35q<H^99#@!Hvob1lIzVF9^;T1c}No3gSiDzJ9X)7bXYTEyg3V2HMxaKw{wh
zyQZXV``)9gz)3~CUlH%m+7-XcESK_rr7TAYeo=xy8B5rX34S|R-X-|$qD*Rr$6V*8
ziHy0S*JAqa{d!2)5K)@oF~OG(ov%H<W@wVfB!(t?OlBxd6~+v?g{p-NEmAFFXo+eG
zOTkONqVHR#TE@thtClkq0aOv}u#N!15rE0N8K^e1@-0BMh2<C^j{)83yY>P3K9*yF
zJeK7+po(Mj$ARED_I-zd>JTfB2da3MQ-CVv-yEF6=1&8vG**5Zs7|w-3p{dJS0WF1
z<gr``JPKL92s|#bd>5$h0%jWD1FCy0R|8cw%e6pN`)>}e1z%A!@DQjTvhq5hs$;nY
zs9OHb!7Xh5R-kHS<*$M2HOp(Dd=2bQ-?tXZ*Rs40%Ga?R4prf-nF@!&;gG3s1XM+^
z@<^zPWH}nDqW{go(QN*`P_>tp$3RsK%SWKc5$H?IoG2ZI9!FVDgdT}3Cqa)SmNTK0
z3H#9e*-*-6IR{EPEMJ21ORziDFM{$SmW!dhnB@v6uVA?n$}3sE2IbdSu7N={a4a!O
zqVx>nzkMh-LaCAECMY#g9yv;FS~O($7;N3bk!?7#C+jWoG_cX>48xEZ1M$AFpfl=1
z1YdR?k`+2rz6BSREObp4^oiXXH+9E^!~<16SWqk&6${;n;47m`Urae+)>FXt%XES}
zo#T^CCoz-@Oml%d9btvQw2<W^(DVp&r#+kqO%o}TU=XjP2>}u7DN2z@I}#~Ki$%N_
z)_#f<WScBXr#Q<~+!v%LMCmqXd7JXLdm5XD%3sHI5vA$M>8`Z)Gn6wJx(t+;fh(<j
z2~d`>oD7x8(3RFc1u9cmu7y2np*#7W-=_EZX`;QIEhpkVs4@*X5Vu0aU&HJjXGcB{
z-sAALQM^QaOHq2j8$aNCkVS`72)pN<(J$I9Z=u6m7(n_>lrAIucLecgrkEuRi{^56
zVMiyPvy0~fX#ZT{?5<EI@iVM=70Ir$9#Ll$jZ~dP(eM`I3X)wx{m83CsS3%fkS}2!
zO>B*1B(XL8hHFT64cSq}&OjSc{BE2w<u7@aK304Td>#Wo(t1%Ufj%X46GK9ASn)cN
zT}M4>iJfsdQX<)<;aI?@XdLaluZMjdM2EuU0h0$&$^!%Pz##hWi(uGA5JdSX9Cj22
z(V>t82PDBkgvmVPbTpwo`qCP{R}Jq%0w$uAB=krUdXbIrCw52@TnR54KDgBBE3eVr
zL}`}wEM}t3ww}$<Dqz0~c+=Xi2KK92-VW@yvm6ENqgcKQtgo`uv;tUHu>2U<KL*~k
z^-qBP6PDY6eH+W|z`mX3bZDJUr@1Jdh1O?T&V%-O(3|w0DCI-@e3na~eF@8@(7u%A
zI%r)-C!Q!hg4T~%u7}q3EH^;w2Fkw{)YdKGDz3;xDTcR+;d>G9P?TbMn^>0Pd7F5a
z5A!yMSw6zs9AWt=Z*!F8W4z5Vmg{+CJ#RxMIv(=8@-fRVc;yS0U-HVAEI0GYW|muc
zWedx%c;zd~B*=vgNoQ1<j8hRW9A+O;4{9GeIHwNIomeOmZpL%15XlPJs8(mtnanj}
z`|uk&BhGZ-w?T(C=ubioqEw1hr6_<*+SR$KU28|qnuphyR9=zFJ30612Nz5(9x6&1
zyds0;d|r{y@<(3rk#}k<4H+}9W!EH8S}!Qp3k+`-6q_lpJ^0&_kw*(J;iil?A8$kF
z%tXtHy$D4Fp|wq0|Mu=7N@HEddXj&})Nd>kiW(a<b_|te0GABlNhe7L=$8TfSq{no
zV<^YL-f>W(gFX&=#lik84~~PwDOW?MYUoCKSd`8qgY(FYbRh00WRSsfCNjumISUzN
zv7C(zvRTeS201LpaE387j3G*Y^45Ryz3JT8&s*<j`2cTyfaN&eI*#Roy!Anr5AoKA
zSdQma@w_`N{|RsLgtw)0<0)_Pl;vl<#WR*0d5cDtn|O;RmY?$$&sl!K_j<v*(DGY(
zi&iQpXCru=Ttc!-?2zv~C?P|hM2fMK)p;&LhWsB_3^s0XZa9v8sSUh*1K)#a?x=Ol
z9hiOSDHfy(W~qW5ak6&2o$~E!eTy5|WB7=iKcYT#K7K^QKcexJ+c>i})`Mu{yxKTF
zGGBKV4?A*g{R$PXe6(O3Em+Z3CJDw#f)!a`c3Caaxt|#4jo0%R$m#_eOb$RqDTQ-N
z;YQOp-sPO`QYN055jG{Ck9tU}MX4NFmm_C7E^i|Jo7AHarF+Qu9_m5m_mS~^mYa}4
z6H*dS<5R^9ULYl{STizcMoM!0AxeL6@;^8WTG4}?{y|Pb`7mdEm}5*^GN(-D9H=~n
zQ>L(-$|+M>F5`^L*!*`m<sDXjms8$lxr$R(u^h?kNAe2t@%Lzs!gO>#AA!+sD@Hww
z*>1axdR}I^1X-1!UbNdvkxeObr2HP4zem=@#Ea5N&iEwTEvGo+Q!L-&jBl}gn=`&m
znV21H$UD#DXhq*ahj(xwIpe}^8FDB=14%`Q9*(PjI;9J=TNJVFy#ESO3N{ILBaPo@
z_-9<f;h)ap88E_SgfGdnFTdzc%ckbLcr-R4t0pv>4uWLPDVdu{KiVD6=?>+6Z70{=
zHjPl~iBgz4Y!tQ03)Bl3iUjINz;F!kjRE7R_FbU93r0~3T?N!tEWZNkS1hlAzH8t(
zs-FpcGGPemI#J4kK3ObZf{IJ5+fxC3E8sY){|x#(gCWGi?_bb=?!v>fif|`I^2$iw
zkqn_j{#8o9{Pw;7@9ZXLqIf^mc@j!GT9hJ?b_BbhdV;i{AO$%v{4{zjPtyI9DBa>L
zZ&5xrzS37G{YPgU;a#F$GLrUS1W-o+hU0)bj^z}fPGPwgsB6JU+Jh}X-NJG>REI-G
z9u3veloNl=;75nX1^9~6LiNG{v>FQo76wyVq+Z0x76mL~WXsgc7}>IbWsGdOdO0Io
z9=M!Y>^B4TX23`{gMiI|$-4!px3KapAaDy{^6vwReZZF1aUXEp2iz#f0!1t<j|Gmg
zto#sAA7aZt1Og7R?~e!Scvc<{0^`~DrvY^uD^CLfX>9(}Kz*8(p9X=a+5Go_`W_fS
zKioYKa1R7it_JFAR$dJPs#*C%pnk~8AA*2~th^4W>sWam2&`kBr&geDW#z3Ppq0)4
z8mM2h^4B2nHJg7eRIG)zw7zSh<67uOc^y=&W993h<2qIz0o4(ZX@3L^h=5G{BcVEy
zl}EzBNXR&Ud!c$SE8hzP_Okh7pgM+?$H2fCHh(ts$)<Z@5y!Oxv*8$ON^+o24lB=r
zfjO+a2r7!8E!8iAjz!RoaxqjCv+`o-Sj@^Rp`wzNS3<{1R(=gCuCelK(D528Z-hRL
z(2tlpZ115@6XnG9p4E{{D^B2rpcz>;BM-XTr*clI+(5bzRB=vKl*ttjY#J+(tdhDN
zI+}|f#A{1(_ZeNdZ_208NISA<N4@A)=(u2foO(#7pDj6Ux~V1MKX-h0#w%s_(u&4m
z_`ot%FiNGNKH_QN&M!l<GUQ5@c6_RW<Tc2Ru#V<t4C&L%q8k&E`o4PfpX(}e<ws9P
zbJ2&cqp>`W5qQ&0%TZoB%JMN@I>z!D-slW(LG`nFqimK7d80y>AM=ind3Wl*xA2ZF
zEWhF%U$MMOFj^&8Q2lVhD4gZr1*6|7XM2Rq;D^LL!=HM*_xOQy`N;-e+3ZP2JoJjE
zk@cLQaGRzpbMx^&B^v}~vo28q2q^&LsU3}nLGf@DwJQlQBms^moD+KY_}qcF58;Al
zdCp?ay4rzHJG*zzhd%lAl7}eOL(h8XOUwuHMX|HmdD)77pUp^nGcqB&6&!L$+Ru;)
zxr|hBbo3m(0PVH^uP1fk&3FGhIX`zl_Qi(V4Ie<tzw~6!%6*T$DrB5=RuIq9lSocF
zCy3`*J}-#pS<Vo|43;wmF_YyiLCj(~TM)BZ&Jn~MmU9I$m*qS`%wsuU5c63s5X1tO
zF9_lVmJ0>3kTMw=xQ9R8RU_>ev0v%%1_lsE6Q3F&{P&*#ba_AfWg(+1cK+uq0y%iV
z6X!CLU8dn|>?tFsTWAoqk)4<Kh>au*8O~UaWaaF>_YQKpgZk5qomcuu#v4*&RqFqy
z;A<?>DZck9-iz9hv%K$Fegx$leqatCOu2*a-NAcNo3KXkT_cR3yh#|iNeCuP_5iqs
z#Yk4nHti;Iy2&=J^C}^!A-NQUGnOJ*DY7O_!MD!Pi2G<VVI9o{Lz40S!<7TK8Jw!Z
zR$#X0Y{sds0G<`B6L%JRo`nO5@erj3=-B`V5GzpYUGD!j(G2|O23wt3yz2LQeHYw6
z=QyWxoJ75!^PJOpmTNhuT9zMhP7f%P4}e2eo!2<Y7kK2#w|3Dtjlg#B0pIljZ$t)z
zi0|cfJuVoLcv1b8V_EUbMvlT&Sq*$vv-k3z0G}tU$xer!>2%5X6qEOSL``;s>Rgh$
z@^~D6V>)M?&N<RH)o{i&l*vg6u5;(TQeu0E4d*{aAY>)TlC<vA%nB?=6DZ$CPPb8C
z%GJoM8uccuqq&AkI)mKf!x4;6cTmauKHUtQvY(qqD*8-2aVzvW^*bK!6Z=mbMl7T#
z6@b14Ab>7`39xSh_U<)zoSQl(z8krBy&CsmHZacyJ*Z*<uq*&QDaS+ec-Vtv6r}`c
znE-ndey+AEoj+y&1>A$%fzNjKwqG0YX`=^cEivbEq*K;*|G9b5(Q!COy0^0xFUsLa
zI~-Y)WuDv=#$%@QnlSl>AGicD#qi8NoMR=&t>pB`-L&~Lf0{ZN!{9{S2gWac8xKe%
zl^o*!^Z%NV(3KV`lk6n4$POZG{`9b|p9*#&rE5W{Xr6hw*D|21>+a#^q;hq3`NF|b
k3tv<we_Y6emX;%Zbfu3T^wF6<enB4{=%YJ*bfb^|2gIYMH~;_u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.7921F720A3C738B2.idx b/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp.7921F720A3C738B2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc1676cf58446017ac2f541e7f0d4459453352d9
GIT binary patch
literal 9590
zcmZvhdt4M{qsMt>Rz#o_L%gI2a?=*gcJB8XNs`4I-VHB#Sru<_BwMXhN4ptWYS~L_
zW-58<WXZ8N#V|F^Rt#M)rD39IRCvMkvuJN>@B0pW&bYrn4xf+DXTS6NvcL72XNK9C
zVR^Z^b2Fl%a;D}jnB#k9k}E1ID#89&^31%Vo;C#HqoNk9Jv6Ft$iq_N?GJ0W9C?hC
z%}Ki%mAc_XQhU;2d0L=k+P7~!I(%aLlN0v^N>ba68M$fpXs_q?AGN=|c<hy--kx9d
z-;f>b<=-gxUb<vR>$S(1-n>xTaOc{+@6weI9~^Rf(emK!Q-{tkzI>^%xUDFir`IiA
zHU2_I{hi74{&}l?$B9p8)n1yWA3c`z`KeTD?Cpg+4<}7Lc71c`{ZFo4i+%gd>%pJ@
zBfn;hkE=KmTKoL&X<yVW`mXZg%&VKGAGmV%xxFJ+e7e8$(Ak|g9a+uxJJj`j@xPjc
zzm1cx?^N1^()#_rd)PmgZsArfa`)Yxz4cQ5P<_q^pZC$u)bHHz&gOydFRAal@Q+vL
zui27+^029GrOteNGUNJ3Z4!fvZ=PI}l+<1S&pQKFPCB|td*$?sl7~KO>^W=bFDt)G
z+O=Z%zPGIr?2?zB_@T6V!?P(1Ou6-j%Ad*ae=hrb!pr8Y>0dms{8Bxc6F-bTf8*=^
z3EM*2tLHZePhS81<F;Q8Up-UW({~5^_Pi-aNBz`T<n{-87A=~2y>EBrh9c&lIhVIC
zfAO~cz5Tv{mmA|xo_s!D?VY}L?dCE0pQTSY_w$KEPaWL(+Ji!4=hxCZKmOtVor5Q>
z-@zu)XD4&PP~F+*pDh0Px&4b<@5;^dJW-y$=Hhcb9-n?ZxOV2iwrk2fa@U>Pwr@Ll
z@6^gEU9ty#dUpNRKl-1WkUx`kPZ3_5Tgg;coxZj%v?uDzH~W)ixxxI#(#8|VM;trX
zFz1`QGjt`vJQM5Q+)CI#`k8LSc72pG|HVzKE*?9$D<=1;He+iCTpqW7<NOCMbXvZ&
ze(L1~<1-RYtY?mvytjQ%tx~_yyJqO)R}S1g-J{?C+*o>O_PKk5ZwQ<I`~B5DGj4Sm
z|A>3kfLo_eWps3J8_W*Q>wWMt*-&;|9<jAm|H;43e&+D6CqLNHIKB&&6^MWSn|Eg~
zP{mI50}kE$VDp~8`Zv-Ge^RebU+RBp)4E-KKIi`a!m7NV=|I$u9m}@4K4rF5^{n4A
z|Es*X?<VzbHFe1PYpJ7G-{?@*HullivZ{VRF)1|u@@${ytCe!Q`GY#0PAM5R<IZpO
zGm7_@oviSl?ziG*hlyV&7Jj>L^Ywz5%@_Zw={&}>^Nn$bxxDSidOiDNM(kIOYi68g
zXZ`q2tJ|URzs$~d&pXv~?dsB(zHc8o>p$M1abex3{zmn+j#(YQ{xjeA-njYu=X~Gr
z?cBq?cmI_%Jik7>`uw5%A>Yho&jy=kd>`BWtve;<2@A7cXtlZcn;##Zcx2$V8_d><
z5igEEzW1j=<9E&7`AgHRc1NEoYP_(1<fY?(7X10oXht4<^V7F3?)$#=oXdX%XQye`
zmhD;jYp0J6^zjYKe&0Ghv*P;b#A|<@$$#Uy#tW@~Te<gz>b*<2f1NpZ{pj3p{fE4%
zH@cUOOs2XeUov-XSSaYxSC=0wx!1SBwfEr@&%HE!-90+(!w;{OESxjB4;2*^-Mw;I
z&ITbRjvDt=R8dlNQk(E#pHad;v3J&=e>tBU9!HI|<F3(N6T*WNCu}e6GXJwH2ji#$
zJ8mD{zD;;=;?5sU{_)C~k;!q?7&}gmPHr0>oY=T~vuW$E3%AEn<L$UhbeGoQ!HI_+
z`CGs2;qPa~QDf~mJvzN}cyQv$N57wuc<b{w;;1L>I59dgDLgo_zkbijSGL!zilc_w
zaY}SboABVoD+A@1VwfNPh@(c?aawd*r|{s!TcZ9kblr<v?!{4)>^M6*dti8Q;vXiw
zQ(F7eDsLP$!j4m;Qy&Zu_v6zqX|07TS{yaOj(bG+hz}1=d^VW5aQg0)Yw=MN?YMud
z{zJlp6Q66yz4&(e;2$4|8g9qEl6s|v2PeKVDdo%FY!eel4YA{n(H+}_2PeMuTaPCW
zxBc#&II7T&yG3_v8y=ka-d_iNC3g-^ilau`ab|R8LU?fEzuJx3QTOyWzs6C+?D(PR
zhuVY(C!Xh<R^m#IteVu=@M?MMz}@xRE-ESUMIO#0I&)`}<}oA0a13OO#z4k$EM%<4
zLdJ0%WSqu9#&bMmyv9Q&Z~|n4CO{@~B4nZ_LMCw%WRfO9CUY`mvL-{Oa0+CKra-20
zDrBmrLZ)#VWSXWyrgJ)Ex~4;Ba0X<CW<X|gCS;~&LS}ImWR_;#&zz(EvDolY4!gCK
zc-|$EUFj(!o(;(5#Pb<2Q%V^0nKHtl&y*7eeYTXa=(A;nMV~DvEc#q2;n3&G2!}pb
zPB`@WQo^Ipmk}O)zMSyr3#CLrUnnC2`a(Go&=*UIh`v}xMD)dSBBC#q5(#~&j7aE9
z<wQbXE+sPhav71)m&=KazEVmQ^p!HAps$n@1%0)YsOYO@L`7dMCo1|{Dbdi^%7})(
zR!%hZ^-`jvua^-WeZ8FM=o_WPK;I}M2Kq)hG0-<liHW{hMoje0a$=%yl@bemtBhFa
zTjj)x_|9KB`><&F69<Pyjwcrm3a@7nP=RMegqR!#OvcNA$rLbPvN<f6td|9oEnvaq
zayT$KF9#-9z=6r<@L=*@9!$P~2UExqz!ba!m_mU7rkEpwDSAaP#R3sbDMtcR@=9Pz
z1rnHYjtr*kmBEw?WH6N+1x&@OfT<KHV5&JPn5tI=Q!P-z)N(X1HLnJyR-l2Y=jdSS
zUL8!mKnK&vF~Bsu2AD>H0j8N_f@yk9FwFuJOe@C%)ACwiS_Rg9)4BRehz(zL!LqF&
zp51u1t{@CDTR~W4u7YsLd<Ef=g$g1dixosfmMVyZELRX2S*aijvRXk@WUYc|$a)3Q
zk&OysAe$A$M7Ao36=89~>vP{-^vFhHUzC|W@_+chYc=O$bTNqyUxmVS)li<JRODFp
z)KQ-E$W4^zI&vlL*-e9AO?!?Y*V3M|$N`t<unYWVm**BRQ$sP>&(u*2_A^ZsgZ)e;
z&0s%MO*7cf)Y1(0GXWQa{Y<lq!G5-eVzHmCqgd=`n<y6h*-Dzlezuxsv7fD_S?p&6
zE*AUQW*3Y7Tn)uxKUYU_*v~al9QJdSG>83MHO*l^S4(r)&jnl@_H)fH4*U5UipPGw
zj^eSOZ=!hY=PPL*`}t~`$9}$+=CPj-xOnX6n_WEi3pJF0{X!ijV875r3D_@G(gOAi
z)wF>9LM<&|zYuT<*e^7@1nd`UC=vU`I!eTTv569~U#z4>>=&zP5&OkjTEu=a;1aQ4
zY<7v*FV#>I_Dgk?g#A(zC1Jl*NlVx-RnrpoOSQCw{ZhasVZYSulCWQ{p=9ir>nIuf
z<t9qTez}sCv0tvHW$c%0X&L+FfJ?@Hx!EOSzfwaf*ss)43ic~al!E<AC9Pn;QcWw^
zuhh~C_A3FGg8fRfOTm7%hElO#t)o=zSDPpm`_)QX#eTJ#R<U2LrB&=#11=T&)n=E9
z{aOvBVZT;KY1pqdQ5yDZm9&QaS~aa<zgA0Y*sld#8un|=E)Dzj8cN50y^hkcUvHvx
z?AI%49sBiaTE~99me#Rf54d#f*PC5B_8T>nf&E4uWnjP2L>bs`RMH0a8`ZRd{YEWq
zV80P?8Q5<$yA13%YbX=@%{t1&ezS=(vEQtuP3$+TX%qX+TH3^ZGvG3@-)wf7*l*QP
z7WP|pl!g6P6J=q)RY_aeZ&lM4_FJ{Ih5c5*WnsV7?6M;L){<*ce>OjJmbi-Y9?Bc*
z+`GwTZ?5c7p74Qv6_H}5<U4npysP&uKlbvq2VC(*iQN)=IQOXHi;_DfcXKYt;)~K!
z(lVSoQ1M0SDd`!`&8|VhJ5ML|escvCU({9Xn&W&Uk1t9!Q}dnMPL6C}*|ppjgG-el
z@m#_yyC7kZ*&tz&xgg<?`5@tug&+}-#UK%pr67@z<sgxfl^{`&)gV!kwII=u^&rua
zjUX|Q%^)$6tst=?JYsPBtj4$1R`x#0ZIj#9*;UR?XdfHCK?t+u_bkE5@-v96pGD;S
z93t=M5e2`1DEdW2$uA+wei>2mD~PIJMb!KnqVCrb4ZneC`b|X3Zz1Mp<qFQEJe&V`
zje86~)Sjze{d$dX?m0T^n2rtKEse}odS-erXW7{+)}7wRp~Wv$zVdbb#oaCMG^*El
z=bc`DPwZAhUi>e6A!m-wob23pj4#R>n{}UFJ-c)5mPMa^+_J~o<)%3wZpn@QU8Ufg
zUF^;6yR}bqzJI^|OHABPJF*+>&6%AtCFd=j1Ia!bTJGaUKC`Eac#h%@t|AOFTSZu8
zu8MHTd==r5g(@N-i&aEKma2$^ELRa3S*aolvRXw{WUY#5$a)pgk&P;1Ae&XhM7FAk
z72&lLe(V0p$jNux+4o^%JQ<Pi7fcKT#Ky2dTnq=q$M8Txi~uCYh(J<|1SH4EKuU}P
zq{gT~T8svy$LK&ti~(fEm_SyHb)TI785JA8M<3b$c)tGR>dOnqUR=A(UOU+%vd22V
zvu#0DtD$o?=GvRPJlG}4`4U)hicXn+nZ9Cg&XsaCXI<`PhfQtOf8$kqb5cf9Z|B*`
zff4%UiQUfx?9BsQ19~{NbGEiWSIZNE$mdGSOzYz;TIcJhEwwx+0I6>p??}!&+Fwmu
zo)$QgbEw%1zvU@`v)SI?A>#QbF8mN-kl7Gnk+~4zkogedk%bTuki`%Yk);rkkmV4O
zk(Cfpkkt@Tk+l%fko6GJk&O^Bkj)S=k*yH1B7A`xcWLqGzfkt>=$6tg&3Pc=JU6jt
ztL32y%$D2J$5{{YMU0yPV%;nd=jMQTHxDGZ1t8Ha0!eNONOsFWidz9v-71jg)_`=k
z4rI6uAk%FES#Ik-U5(kcoRTY6*t5|wqhnuZ7MxkJPxLL%h$6F*oSd8$xll>&mF#io
z?%mD%9z9pq+;T?OGe<k8U;B)%XWb`ft3Ast50@hE>uDgK-*HVg5C)lTAS^Q1KsaQ+
zf$+#e0}+tL1|lL$4Makg8;FdoG!O+@Z6GSL)<865y@BY+MguXB%?4s3TMfjDu=8Av
zi48wbi%d&mT4JVifnr}X_Dal-kf+cyA+j(%KF<uGVoyngm_i0j#>ark6f<D5g)Eq?
zj|G!0X2IkNIWRdN2PRj{fyo#0VDdg5Oum>0Qz#U`6np}hLa_j*SSW%i`b03rVi8QK
zPy$o(NnlFF5}0zK45sXp!IX<-FqJ|DOvR^wsT3<<s)Z_;s!s(|EmpzQ3N<h_p9ZE@
ztbwT)>R{?V9ZbDg2h%7tz%+aYm`1Syrdeo$Y5Ghs&0-TwtIz_|@>yV7#nyduo^SS~
zmgmAH^QM%{YI!lWW%%EpDb9<isOgcHzR1hTk(bl%zod&cUpQQ~pw-o@F*7Gm_f45P
zZAR3L`)Sn7$jd2nrp|6j@}~`=U9F;9$Fzxyf1q942lEp8=VT_t=Jjje@u9>{okdAe
zHQg}#ba2aQe9!Lx-`1puyLL-S?UA05)$5V$NBcg;a6&TMrMKtt4oOdVXHmI>tpQ}<
FzX2N97uEm(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.67A4E982DE011DE5.idx b/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp.67A4E982DE011DE5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1bdf3a987a5214c2d01dd38159d7ab45f2619a99
GIT binary patch
literal 9590
zcmZvhd015SqsIB3Ga>?|7~+y52#7XlHfP_@kt7-1a5r3X85Os<qM>!`YKumeTJ|F~
zE0r|6QF85<Vwjp{BZiJkX_zP)DlV8lgLZT2cMo&#ao#^J&*RtgIq!#gnX}Dz&Ka7W
zl{F_VA|i82_JY~Ic@td`5fO3rzruNQ^LyCfkBx{}u<pRfyupu4@pnF{`RveBq-1u=
zt%&3e#}e9-b}Lf+g;Rfg>&amg+8&><$6uJ-X7q?nvqpJ6cmAxo{K}Enhj@E@({DpY
zpl8uWx!1C#gIlgUy6o1un)<uf@BfsleDdhvJBwBX?wmMqcJY-94F#=5=`8*6veo0x
zrPbY?H23e@ZQD<HKE39`RQ>Rggs)E|Q)BKd+<q`&!jT)BgCBlz{d)Afr``<w{%<+e
zqkUZYq2Rigc251KcF|82=Vx5oG;QD2(=YBGzVgex9f!>7xarUuw(o(?XAAz-DEwor
zd}F)PDwxvukDWvRd)a5)>P7B8J2SRi$Qh!~{^;x8+Nrwj8{XSI;KQYLeHQ-t#{9LP
z<s3g~YFnsNKORrJ@oB60z~WoS*Cr%%)&Ko&|5XzYZ_-{rxw7!FPaAs79CC5hPYF9#
z4%_ptHJn}g>NCF-S8aG9X@M!X+)(j5`QxwU|BQRhoH^~AM^;>@BeP?N(r0ge-!HB#
zsJ(G^gYfK)KR$2$?XWd7q+NZsvG2~Ee0bz<4f*aOe~<h{Gj8<hs@zn>oKt7AKh9ZP
zwzpT2Z@`s?*yG1vidB21Zdtc^bk0|)<Intl?7(yTw=a2AXy~{kwd2#D?A<<S;`(iD
z0)2WC7YNp#e(BkQ&tKfTxaE$lY|k^LscX-_*zM_QM+56-3~0T!#3Og!UbeMt|NRrI
zCU?ph_~q&KTmI~KW_->J);(ERGN*#6syuo9<KV7{Z{O}mmS+WW8j2f^9UXq;O#SR1
z?oHQ~IP+Ardvgn6@2GiQhVJ+@Y5pslR-Zqze@9f-bFIeI^uIE8@5cF$oa?Y+S>2Q?
z3&y3z9b3;FF8pBYt{SCoqj&9)r?2k2cd}dG|GK&Cz^pU(2i+7l{pW{kdZgX%H0}xa
z$o{uao=9u&E*r!S%I>xQ3fWL{R35&iMZZaxX3aZz>G((68pd^^(*3b7{qX**1*+Jg
zuK$7iA8p?CpkD*M@Hh3^v}HxFZu)pf@2|OkzPvj7ciJDZZQJrP*OyFLWskbg=6{zR
z^V7s$Ev5`!e?58Bnw#xPT1P**B)#&FV-thpu2h$KzFQ@?nLn_@$)v)O)9+rcn_jTD
z<aoLFWZ#vy+D-UAKJUjpn{VVsZ9e~?y5nfi_P53!<g&LO>G{I1Y0=*`tet+6o%!o~
zE$#%zU7VHSo_nIlx;4eG{@gZry69-ThK03X7B#3#+NZbw{;wS02V>{&o&9tDk8=+8
z+W8=1SWaC=)!73%gMXO8o(?ok|2ewrJ9i69;})jB++uUV55GP>;n0Axo6MH-;jfH4
zy8E|*<95v1ez9?8o5RoLH=J8P;=<9ta{u}}l9319`tqIgdwy;?`^ukzSt;7}<-1m0
z>hS5l-oAkuA6h48l;0Q?fBnI!oVQ+VIM?#>s@*SF?Ow|L>(rSWhv)oQbikW@vup8)
zM5;^T1#`!Sg@PV=ZN>h=`+e$NyB|OH;;X|xzE7uo^~v?Zg|jF1rXnIDyH+gE+#n>y
zP-CBq$WMq&XcZdlD@y1m`tJI(ujO#VVyF>z+&QvyTxf9OxUI#V=6`i{e+-pt$896q
zwh9eS-0{;%zg`_ZA~A*<ZO4g`iLFC}6L&lKR@RnF@!l9}q8(>MW()`oPMn%CO1XS-
z9{WhdFgxy<&@(wSII(f>R^ygS3%ABl<LtOoWT%#)!HI_)`bXc4VIQW)P-E;kH8Qnh
zXmH|5hku?PfBWmVVyI{BI6g8yAv8E~QQfZNuWzkh9YYPV<D|%>R-wU(SNTgXL@~eo
z8AFY<<CMsh4xzz`ca6S(bjNdD@5e?=u;YF$`V9^ZPW;RG_lj$NTW!DZ;dY!Hnfz#I
zco?62Rck3+)ncgecHAwpTWn}>;;R#rzU{>}GBMO(J8mD@zEx;&;@g+IJ#(=2Pw&N0
zd3M|-vP<jG;KcVI?DG}g-9Isg8fC}bBfH0i1}A>dX5_Zo=YO~qLk+d#$08qV6&jp)
zu5W6gD=~c3#6^dWmUs5uTVHlwNs7()Xdctqt+Dxx#sIMz3&d$05U=q-f+hfong}Fm
z5|FIPK#HaSshSF;X&R8O=|F~N0GXNzWNFqza?XaQqC;l^yzOG*d7p&exTl18HXxT0
z&sV@qF=5bWN(h5KQ%V^0*<!+?&z2AteYTXa=ySz{L!T=l9Qs@-;n3%c36DNsLU{D~
zQo^Gz6cYh`p@azN3#CLrUo0jf`eF$Y(HBdJh`v-zB=n^cBB3vp5(#~|n8@hMB}7JF
zE+sPhN-<H;S4xP2zEVmQ^wnabqOX<^6@9gosOW3OL_=RIAsYHxDbdi^i;0fDUP5&A
z^-`jvZxj;)eWQdJ=o_WPK;JATCi-RxG0`_miHW{dOf2-R5@Mlml@crLJD&pfY0>-<
z0H;N!CksvruV)}ou4j0dm`nyt#>;@o<T7BgnJk#Bmj#o}Wx?b!IWRdd2PT)xfyrm`
zVDerbOg@(fQ^*v+6ubhMLaqR&m??rOdPOkBToFtuQvy@+N?=O45}0zP45sXr!IX1l
zFqKRNOvS5!spKkPs+lU7s#gV5%~iqFGBq$YuLh=;tAVL!>R{?#9ZWq}2h+$jz%;xD
zm`1JvrkQDiX?jgC&0G^qE7Jng@>*b8xz<C|nVr}lV$D+%IBd&_XD42*%L#+bmJ=44
zD<>Q>Uru;rp_~ZFVmT3!rE(%6%jHByR?3Njtd<iMSt}<RvR+PfWTTuI$Ywb)k*#uK
zh1oeNsp!ybEWBB=AIl!&%xs*wnZ50rCuU*E6C|F$aoYt6gUkjAi_8THhs*~Fk1Pa<
zfGh@yh%5z(ge(V%jI0ESf~*FKimU~RhO7sPj%);pfoukeiEIUl6=re4n{(b@^u$JD
zPua|FIluhhOwO6AE+)~TIVx<fYRYq%ayGd=f2yTCXOSBz&kf`X+Ov}ezl!!8Law1b
zr;+_G&p{XXO)k%EV5XX4u%D@=80=>nDF*wQ3Yx)wrix~;pQ)i4>}UKg2K$*N7lZw5
zHN|2-TT8Ln&o)vl_Olf<i~VdB&0;@WL$lb=`duvcvrR4*`?+e0!+x%o;;^4<q&V#7
zDrgS-xhk5&ey)b*u%Gk0IPB+|TpafE)fA8Yd@aReKi^34*w0tcJofWdG>`p!4b5Xe
z?|1Ro&o{Yv>=&vj0sDnoO2B@hkrJ?9sGtSx7piCh`-K`>z<$B+60l!ratYWkR#PJO
zi?x)9{bC~}V!v2Hi`Xw#(IWPXHMEHRqTeNAzu4pwv0tjDB<z=JDGB?fMoPkdse+cU
zU#g-d?3Ze23Hv3#OTvDs$t7XGTusT?FV|8s_REcwjQw&2En~l2Ma$SP*U&Qd%YK)P
z{c@8_#(t%mQm|jCr4;N}8Yu<)l?qzHex-_5uwSX673^31E(QCQCYOT!YBi-|zgkPF
z*snHHD)y@tw2J*|6|G{wT0^VYulijo_Nz@U75lYnO2dAwmeR0aYos*n*D7cY`?V@s
z!+x!X*05jmyEN?Anp_(8>(!Kw{dz5>W53==>DaGV&^q?(RkV)%dJU~(zwUSG*snLa
zbnG{(DFgeBTFStFqmeSO->9Gs>^G`t1N)5{+Q5Fp?=rC8XmT0YZ&p(#_M5eoiT!3H
zWn#ZsL7UibR?#N*n>DnF{ife#V!zqsGO^#PrY!8YYAFl*twzeieyf7Eu-~epE$p{y
zXbbx-zstgYtI1`B`>mzdBmQcdcbd5JadMT#-dfeIH0~pN7LjBo<v91EyleKXIP%){
zM_jS_@m=D(Id`yP^AkHHc5x=NvH2-UDQV9Aso4C~q|`L$R@gw{z2_5py}go(&F?IB
z&UEf(#O5cP$vMtVDMz+vcFi}<V5$-zo(nj$3lIjG4G<QY3lI*O4-g(%2oM2T3=k1n
z3J?id4iFhx2@nNY4G<Ms3lI%i4-g&M2oM9=3=k993J@#I!w0oZZ+KU2VegZyR#~l`
zd9bq++D3<NAHvou@+`%TRm32&MJys$#3AxUJfct}Ac{pIqEsXy%0)7wQlucNMJl3J
zq#^1>I-*fzAeu!cqE%!eW~FBd&PI75=jm$qXnu&jt$OzDIo$cK;2g(vbm(4ccw41*
zPwnX(cJ>kLPVMc`;+HF4|Gw`0&gNGd*>jxpO0Rt&cBv+>{D*xYcOTPzl5<ZpHa~q#
z`a^o-^!9b1E&A&7<~`OXE5-Tamfq~wSqi+}$zI*IOWPFZ-o=|2qhfyBmQio7?%tuh
z<h-PFBH3p{^IhKX-|VR*p2N6<D+z<lRuUGOt0WvUUrBgmp^^y5VkHrgrAi_p%aue%
zRw{{ttX2{gS*s)(vR+AaWTTQ8$Yv!mk*!K%g?Zih%U!=1G3j0#`x{G?CoTN_f{9{)
z*eDi=i{gOzC>}_N5`e@g5lD)XfaE9{NQqK_)F>55i_(DfC>_X%GJwn|6Ud6P9?}=r
zURyZk{JQ1#VUaOBV~q1g%NA6&7&3cfmc6>uqn#3*?|0=V=%i^^=&SbXEGbKK4msz4
z``cghvjFEH+RYA~(xTtSYqp(`me9+260>i(eq}<}Q+|7Oe^>u*&M!OHXZve;^AiK-
zMeOT+_mtkwVd{M0w58_f1|aoG;T`E5H}=BfuP;(|L6@X1Db51}=ea_1bm)0P_;)>S
zPjBaq+xwgIh+$&f3}g()K*qXR$XJerjB|64aU2I3@8%)nIUX{>EkGu40%W3FgiPc_
z$RxJ}nZ!wu$!-}knUf(?+zMn0r$DB<RmfCMg-mm6kZGI-neNsh(>Wb7!)-uja0X<i
z+l0*IOvo&^1)0TJ53_4gWh*GTe5Jj&+o!ef<LqT;uiF>n=4V9Vy`7ktm>!-?B=$`7
zICSsc<~>iIDQRlH>gwG`IhRQLs;j3zBxkLC!EAoG6n<S#J@Nd37tDIXAhY#^Mds=W
zhs@U#9$Ba-0<u_7L}aO+NXT+Mk&%^pq9CjFL`Bx>iH5A#6CK&8CkC=vPfTR1o>*aa
zo~tp@q33Dg%@Usy-`$x+*z>xc@fl(A<ax%2Pc@IvGaabFQy3;Dj{%eMF<>$U447;l
z3nuGh!DI_qFu6PqOwPxF$rW&5@_9U%ypIQyFW|ux@&qsip8%#%Ab=_6iC~I85lpc_
z1XIeBz?6Ivm{NfRrkp2(Df?tF<pLQ@B~JlU@hM;`1qzsIo(iVwQ^8aVR4}zX4NT3a
zfvFW}VCs1~n7U5~Q!miLH1Z5E4W9v~QDA^+=9yrcJ`+r{zy#CEv%s``7MNCn_0XK_
zn>Dfdxp3j!$%QkU7p621|NAr9S%`?37GCs)7bk@mr#@Vy^OqzN_Eg=hB~ddbP4i8j
zGIe^y^oME0jPT;**;8gUCpl9G(ykVfEu&h+#6Hrd^`qHw{W819MQ8VI+y1fm4jn~F
zQ8nE#d$)7TDSVHv|KHk#$2)gPO751LmfrJ;j3@g%#c)C*+o_l5>2?XvcxO^sgRK5!
G!2bXMsa4wm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.18986AD46051D542.idx b/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp.18986AD46051D542.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f28c28dda15dee0bd72394f2b89a008c6e27aeaa
GIT binary patch
literal 9590
zcmZvhdt4OtzlZtGthfTD5aK086cBCEZ0CNTktA8X;ob0(msRl=N3vBrb+n5{mRj~B
zH7k|8oh&`uFD(=`%@k1(Eln{&Gr~)vF9T0%{m!uGjQjrM_<HqqpZz@SW3w~t?99-t
z%*;8d5fK?vvgXhBzcA4i5fKq<|5xzB+`Jw(9EpjDn7{7u$lSq?OYu#goZE8jDN-~$
z`F2Ft4OIzor0t62BL!2xf9uI%6XH%y*ngy;Yn#y{Hq9F4^ETZ*_uDHKuMhF{_^RKA
z^va%rjdHJLO9!`HcXHY7nsZm~-FQ%*qI~k`;HE_@Dw|FpzPR}6<@)^AqI8k|c-iW4
zHL11tCe8i3F|OT&=hMzzo~oayNci$}S87btf?Y=wCRE(qeCNZ@Zrq4^_w1XMzx-SF
z@zH*6=dnBMUixP0S3fT*|N7F5>zk$>ymo%!zTqoBKhR;wtPY!wtzr8f?sPH#U%w0g
z7%Si0rL?+}+_&MIq5r*X3%7car_VR(TQ6r1(Pw}3WpC|l?XC^)Z65I9(%L=??!GZ^
z?Uw9QM@?-jb@uyHsW(4u6<@jd_NlcA30?Gmzt?}&#1ose*Uzjhc<j^q9y5phx~e>3
z_sU`W-?fIbOJ9BF$HKA=FDA`5<(3=1{)IIBx%{88ubDHaef7wS%e7>7%uxE`t#A9q
zZoi|wadCt2?9GO4txJZjnIY}%6J+0=JNd-O>iRrS;7E_WMKf;p>7v|H#O$*dvOdmU
zy!}A0fPcW%`j}IvUW!qBrEFcdd35#{DdR8vQg!&bL%WtdD%5valG5SnPY&!FG;w{9
zO`y+D;wta_eEy|p^S3QLu(;*!%q;IS#VKnqE$sI6w3C(VW(;V(w#X}Y+O>Vh_CpU&
zubSL3ec<Qk*KfVs@51=(8LVfruw>5HOxd0@H$J|zH=^Y2eq?!OWp;gGebvd~6&J3|
z{_g&CU5Pc%MtL^35Dtubq4Ut)pC-+FWz*_Q6^C}aGoNcU=3M`)V-IYc_ef3q70YU;
zT%A8IHMVL!bE4pb9ed9ywHtkFhdh1l;Qce*`u^9gWrt^7crfUeu<1WPT+<`9vE#TW
zJR|!zo;jV`&a-_GJ1DEyp{ry=(Mfsu))xIH)y;b0Xx*ugg7xD%QfWtGUi$9+S@Tt~
zeQp244?fzw_pg5S^nz;j`m|+%S2umUyZ4veKVM#*^$UF@A{bo0-Ss)MeNT_tE%QpV
zqRS`tYB6Q-`WszGt+~~<sCCqnOVai<R873|<7cz{-qKZan|TA<pGhhhIsM*mwbSzt
z6rI}XJJWaN?Y0xXjnDml|K^)H?#-9}I^JQlch_5Ek8)W%Dtf;7Q)*Oc{o3hg*qJ}Q
z*P`jpxL;?bd*+_*v2IP_t3SluIUhLLwtm6Sp9kvICGFDMefwv&|AVpf4$S`H%J*}Q
z_WI_pgkjmW>17uWXAk~v27A8pkLf=|b$RDrL2>MYw3l0K&j0SG$0r;cu>BUZb?5L`
z#+}?(J#gIaIlF%SeP){z&*jzEtRHdt<excz{vFB4gKmHR&ZYf7w48nQZsn|G?Z)!G
ztLoZ+da$>DVETvFnHf88j*7qW*V*j17S`9a{B6~~m&^7o<^FZ{!p#$Nz7HJsb-mT4
za6}^2Iq|Z&d&2@jkG#I(P{D&fS6urZuUh!(u#X?m$=g1;QLte4q~26SL}Zt*muGAc
zlA@`x&qd@VL?*NfE%tAe(4VM#>o2~R%?*pDM%d#{k)2{gi!+YhQP^?b7uOC&Q#tlH
zE;6oFXmQ3JKArT_wb3IIqp8vMI59G@b!c(M9nUTt7j@w8S&u{vv&TIXdUg#h&e*tr
z`}eJN3wA_P<Lq(A$c`;Ti!&Z_>>qv8hkck9O^va~DUm51LW?t=bmE8U@r_@;6-_;B
zkK-fb6GDqK4%F^F_4<zEtD~tQ_Bbgrsa0rk#;cALUv@J;-i@Y4+T-NN<o2P(8Sh{6
z)fTDC&-u~RM0=bbnLZ%2IOD^^j_#k7xNS&G!~}cXuSLJXp~V^hIR3rDbJeST(bRBz
z+%>Z6qoL*D_{^(XOW~RpO^vt5-6FfigcfIfZDLYMFZOpPni_16+eNl(6<VBe<8R%b
zIoi7Xy=W@e9(RuH+&Z*4;|G5o^cUPaG%=bQWskc@c8?7$&iJo3BZEIb|6N@)HPjwI
z7Wr7K(Bh2e`llAS62p7VQ&FM4<(-4~*Kfb1B*o-;J>K5V#<7X>ia?sz3=`vFz+^ZE
zOh#kCWIZgHEXRV$YAl$XhXa%2I50Vl1C#ggVDcOfCa>{e3LXJWffK+KGyzP}BZ4V%
zBAB8kf+=|<FeOd`Q_>_bWseM|%*kNNnhd7mQNUC<1x!U#z*Iddm@21&scI^innwdu
z<1{ceO#@T+=wRxc4yLZ@U>Y6+OoKDPG&BQD(_?~ZaweFjW`b#XEHEw30@Kp0ho<w}
z%-(IAf0e;*TS&a`lkm>zEh63x$i>9_1u#=c81$JU!l2I-69#>@kg({pMTAA4Eha4b
zTp{7m=ZXl2K37aQ^!Y-<qt6!+9(}%;@aPMLL_l9CA_DqCF%i%g3yFxnSVTnh#bP3&
zFBK9AeW{2@=u5>!LSHT<GWv26k<piniHyEdNEGyyBBG$L6cYt~wUDUjt3^abUo9po
z`dT5;(ASEHhQ3xzH1zdCqNA@D5gmQKnCR#mg~UMLC?W>>MlmtaHw%f0zF9;}^vz;o
zqHh%v3w^7ISm;~D#0vY)r?`DsG=GG{VUgj@goDE89SD@;9Udkog8`HAF<>$|447;N
z3nuGh!DMq-Fu4p4OwPxF$>nfh@)<msypIQy&*8xoG6XOMp8%$iBY-Jph+v975lk^h
z1XIe8z?6Ivm{N`erko*zDf?tF<s2DIB|`yI@hM;`ISQC+h6<+YQ^8boR4}y+4NT3a
zfvM$aVCoq<n7U5~Q_s=CG%^e@4W9v~kz;^qW|&}_J`+qc#{|>Lu)wr@7MNC!_0V+A
zE@PuYXO*zq28nkM{<aPh2AK^K7MTkY4w(-U9$5$y0a*+Z5m^cn30V#j8CeMu1z8Oe
z6<G@s4OtHo9oYyH1KA7`6WIz9E6mPWwY^!IPpM(EG!pM33Xa)E!XUGaghl2W35U!#
z5*}G-Bm%P7NJM0*kx0mLBaxAnMxr3AjYLJ(8i|IiHxeD$Xe0)**+@)etC3h?7U#b?
z=lw-bY$W#CwAn5D$NxLYcD`gRCQ+ep7qGc1C~qZfE^iIxtwnC2ytj}`Xzw06ym`Il
zwD&l2HSN8CT;}o~bAjLF^4<kzDkui~Obx|gpJ|{N>@y`agMFr)X0Xpx(+u{RG8cn=
zrpd)%pRJ%+?6Wl#i+#3%VzJMb&@A@Za+<|HTTQdrXUkkH_Sq&Ei+!$w;;_%vP#pHT
z28zQzS3+~x=gMgg`&>26VV^5=aoFdYTpafK3W~=*UqkWO=Nl*<`+Nz_W1laldF=Dm
zG>?70%*A7$Z*uY27b++L`$7#RU|(pU1ndhXw19n~oEESzRMP_Xg)*0beWA%EU|+1D
zMC^+-l!$$?ffBJVme3;h#d2E2zF19**cZ!OBKE~5mxz6-f|9T=)ld@lr3OmEzEna>
z*q6#_3HwqtEn#0Ob4l2jnp_h0<qAs1zFb4e*q0k98T)bxEn{CUr)BKR)wGO#xy&VF
zUv6^A*jFkj1^Y@3rC?uapcL#YCA5NlrJPo<uT;|t_LVZ1f_<gQrC?vJpj7OuHI#~d
zwSiKxua?j%_SJG)#lBiitJqh|Tq^d}CYOqRt%B09uhmc*_O%8|!@gERYuMMyX$|{Y
zHLYP^D|2bs*P2`!_Vo%%$G%=e>DbpBC>{HH39VyaFQ;|v>(#W5eZ9=3V_$D_>DV_a
zC<FUO4P{{8XrK)28zr=XeWRQ<uy0h;2KJ3Imw|nw$z@>Qte{Nnn>Cb)eY1fwv2T{p
zCicy8+QhzDO`F&^%UmY*%_f(LeXD}9uy55+7WS<M%EG=?LR;9k%4rMxRyA#5-zsxi
z*teQoR=961y%F)}A1|CIuDq<rvc@=9^fKAotGX4(eq^6TB$-Ls&Q(O;n*A#(Uc2##
zD<&_#b9^`F+G9*!V*A9-&UtD~UUE`$s&i#ACNCu^CDplzIZ$}-`Gj6?ucTt~I*FY!
zoC_5(d0ow}+0OMsN4C%GnlE?4sY(^`UcnRID#9SMRfI+6stAY7R}mgrs3HQgSVcr+
zsftL*aut!0l`5hjt5rlr)~bkxtXB~o*{C80vROq;WUGo;VIDpxF0KAuwT1njWVXs|
z?Yvdan-CWjy6_8IE8ty<8!NydvH=#63vh^hfJYPp0-_iY5v72HC<kOjC7>Xx0TodT
zXoz}1M>GNkq8Ts|t$>A?nU*Oy8|B69r;mF^^F!=y)w6HU;m&sjXFsN+LRZ1V+bX4d
zN>69Ev-em}N^gf2zx?&<-_~CGrumgd_8jND(rX`xosW}O{=?ppyN~HU$+_Gdlb1Fo
z?IFE!e%HD!i@w;_{EoHBOm_ZpOK<h-BvroM(SACvb6m1>S>nxK-O<&-^egt$-P?DU
zoR@SCB>QM+zCIuRH+xHow-VprQo<m!rG!Q1N(qO|ml7UXC?x{2SV}}>sgy{_aw(CK
zl~ST0tEEIm)=G(ntd|lU*(fCjvRO(@WUG`|VO}@>w=SQJm~_95efiz(O$~p)VB8E4
z>t=yCHwVPKc_6_p0Euo9NODU+vRejH+zOEDR)I9P2Bf=nAj53{nQjxva$67Sv+J)d
z7;|aea(lN(AD%wO`5)are@}}cvo~hiPj`H@V}kSjZs%z_Y1&o#n*DU9l&Lwp+&*^b
zlotIqUbml4NKNSF+zoSHjEV}~*9q@i_*rmpxPEm)m$OIgpY8AJ-_3bR=WqK9W%HdM
z=VzT?;`U26-|=ytwm%fQC--*tW#?O|Ej8cw0jW<i??^S=*vpH*{FSo1os&8zJ2!ls
z`$BfN`KFNbv)A3*S5WfKmA2h3wOt>_a(>D_;+yXgIkx?Ym6(_o{=`b`ndo)s{{7AS
zpS)1?NAsbrcOT^(hxVbZr#&R+S^KDLzPS{BU2h%n-o~S{jxfk<9bu8VI>I6Ib%aM2
z>WF|W))5g|sv{DzTt{SNrH&}bY8_FLwK}39>vcp&HtL9hY}OGI*{UN}n4SA-OjPK8
zT6nX>C&zbpKB?@Jho152Ve;mB$2)tXeKh;M(}D861z}=x888_?116KtfXU{vV6uJ|
zOg5hdlgs76<oq0%Ts{XTpUZ>E`*|?=d>%|8R{&G+3t$TQ0+?d12&U*4!4&gFFr{1x
zOvx{SDdkIG%DFO_vR?*M&X>VdauqNYzXGO`uYjrMs$i;q6-+f>1yjq_z|{O2m|DID
zrk<;Vsrz*>^?V&nBi8`a@Ec$n`39I~t_h~;H^DUXO)#xo3rx#zfobJi56!v$SreP@
z3m43tTrjiw!Ib9ZfB#H&9z;Y;3qSOSA5IECociz~oj0vd+<M=((W~7vCQb8Co-%cM
z#Po;Lh#BFBlV?wv)jY|bI*@j?h-~R@6&>?Po7RtJ#rDhS9vhX_H?G}d@$EZ^lA>z5
zVfJq8k(2ozUH-pk6CUr>IjL*6l+?7IPozKD=P8C064{Quyid1Hc*ZxA${b|%Cj<Tm
D>|R<R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.93B77417ED344C2F.idx b/.cache/clangd/index/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp.93B77417ED344C2F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f30237b4de075b6415ea25e08e7e1cbe004ea4c
GIT binary patch
literal 8356
zcmZvhd0bRwqsMv98F2&@F~k)F*`z`9J^TKuBcy{Hid(oOIw~&3m29+bU2W0GQqx{?
z%Z)U<QF84~B}`2-2_46!G)EiJuyASAk6<^Kw|kg-kMsNE^7;6D&Ub!a&W|(moMj$I
z4$jTpkQ@==oHlsDT=xr8Y!ML=apu3$7v>fBGT~@UM8tv(N5&Km?<U9J+jRcpy1$X~
zxv6&|Qns8*=s-HINIhCQ{kylH8a27Y>B$F<mZr2HH+tKgu`cG`ALoC2^~4(^T)n=`
z*^=$c@@!RlFIzgi^@fwn?p!+GaR27RV`=K9cEj&2TH(8Q=15@iwJVJ!v6391H!fQ<
z@ltyI{i1pQYVOc!^7EPJuS_?*ClbCmlR}NZw{UlD!sHXd?RVFIdh=$~d*|Nr{qk>l
z$H%$(s=B)yUap@0<<E<b?Y%tf#<m%UuKQowU$FACgIz|<>9VbEEjQpu_dv;Czl;Bv
zpagfTZSJNHxK%y!zn6W?uUX{iTb;e*O5O-#?uTFW(a+WI-m+@@ko8OJ`!4+B&H3v-
z&O2S(Pv1eE`|fml@S`^IzQuP=uS-biVf_33L93^Dx9M-3U0M3XM~%H^kGQ(}Si+u_
zqYk{+zkpl%+S5OleZS?!<OThd)?4=eLT>%J{GV~J_nSTA%eE`7)RVa}Bk92HZ*t-)
z@9J*`wusLJZ+#N`)u^?z<b8d2aqrEW>K${TvDo1`+N*fctYF_B>TOlZJ9lyL#=OOq
z2YY+mL#{Q(oId?>jMh7C$A<0W@;*<Sbn%x{N1i>rdr3R7vCERQE(13m+&yf{=3QI@
z?JweecYpT3{7lIwFCARmdQa|P=IM&Gb(df2IdH~F--cO3V%L>3O84EBJ1Y-AJhOUg
z*X*I6`8V(QBj@6zyjh%Ms<`C2z3lfjXK!x2yD#FacXG(`Twh*eS>vgb1t%^x%>DMk
zOhb+9cP`4Yy_I-y><ft__k5H*|J7}4E}uBO$DaFaoAKudU7K)l>-@HtI<Ht(KkeFr
ziRp2tHnZN+4|eW5uhwsMts61$`k@DBdk*-o+sls3x%hC{ZE@Rwu3y_Ly}9eeCmmx3
zHJ?3`-pNrpj2kw%_u*?~OZiEqU`MN*qF?8{Q2XoY4|g?A>`G-Gjd}Uo_vb9oq|Wt&
zjy(Ku`@Tmxjr76`+Km~@Jg;rrxTnt-{6Ak=Gx!(!XvD5v%PVc4v6VHw>OY?U_2B4Z
zQ+l_WHhlBVl(B1XcPx*MdTL2#&8<^Y?*91c@k-|F)k^#MLpz^ME*&%T{%`d&OAeNw
zu5z6nu<}mF$=}2mes^GdFyFrY@}uKj#xc9!o>0pV-gzSH#h=onzHVGM^DH;}r&X=)
z-JN)KPPSv-nO+;#mc8~vhr52y$&QT+fBwwVs4eM~+3A}<^V}ayn168Y4-MZvSKGV#
zQNpOa`t0umNAiY$JB#!Cnr8kG)#Kg!r4?}tGhb=7z2w`Ux=pSdQhA%*QC0Bj#FP6k
z44t^=x!qTPpWWX3Y;ohI&7-fJ{4@X0e?_v&usff<d-=c*t><3*!#5{Yzqx$h>R&s5
zbf}MeX!iR4XJ=If$Hw1$bT04hml`j%{%!UCSH9oBl>h78#h~}O?>t9bDYtu+jZUHx
zldkmJvt^-ZMBZ3&xb$J)2HXB_r(Sw()W(N&>fTK^OBc>9>O)0DME2Ob+_^<ej;1C&
z8&RAPnb0OQm{*k0Pt^U*f!Fi+QPI?BGwvSQJuWm@ajgIO*(-0nd9-cBC^ODV$Vv$f
zRvfpptn2*GuOE)4^3AwIWQR7P!HT<lRP@vJaif!>sc~kU6qyto8m##55AOWF<JW~d
zqp68z+%>Xm>(F4uBkKM!AbZsM%xG%78K*_2bqNhtT;%;>W_<G(Z%0$lm~nh$d_rij
zVo&|P({JoNz9yO)VaCak$!$V|6|X*8amCL5_(wD~#*9-VQ#*$SD?Tu2`ICXPB}LKH
z6f@3_%pMXNtoX-CtIEz_SmTPO3d}eqGNoN;cpRU7O>Zq;*Q2RPX52HfXH00Y;&U5g
zp4#{PEB2U($!0vD)qvrl!HTa>N&c!g_d6R+4L9RXk)7Iv1}kp<t>@FVvBy?LQ-x-n
z7?~Iw8m#!?qeJe}`-i7QQ)A6IBQhf{G+6PY_G5Pa{QS4SMpGlr_=(6T+Jpuxp68xk
z%5(|Onr=~{+4An82b(J|tI09NjDzW8y*Vc7OusO(4i++2XCdPp9Auo%LB>0H$atNH
zOmGO03AzB8=nx?jbrCYjAweeT5@fPNhD_FF$P|YHnW8I@sSXt~RaYU?92#Vru0f_d
zbjWmFhs<yokQuu1IJ3^oUQwYl5<Y@5!n{wyZ#z>?m@UW^g!vqpEh8-YY&l`kXDbMc
zK37IK^tp1vq0dzi4t>6i@aXg9gh!vRAUyg)84=JI%87u!P(cLr#WEtIFP0M#eX)Xw
z=u2fpLSHH;68cgFk<gdRh>X5mPGt1u3L>Mglo17erJN|}D-}dRUo9gl`f52*(N`;o
zioRAxH1xG{qM@%<5Dk63jOgg=<wQqcuOK@5Mj0{CH_C~DzEMGpuy5S~+C_z~{4foh
z2=f<CvnIkKb4`Rp=9>tQEHn`TS!^OAveZN*WVwmR$VwAYkkuxlB5O@VL)M#!j%+j$
zBh1#F%sks#ZewtmP9_)5H5W4!D4!_^6YFHbWL+$nY(5Jn=j6cTTpXBOJ_jc6<iX@!
zJeYhw52oN0z!Y2pm_oh)rsx#G6kQ^iV!jBb<dndaToRa4z67T1l);o;GMI9{45s2#
zz*Jlcm`c6^rs`C|R9z~VYQ74l=G4H{TpE~Kz6Pf5)WOtUI+%LC4yNHWz%*P2m`1+w
z*tDK;%=4q=0S6{d6=ACJ{HP)<GFL@7WWI{<$U+qnki{w@B1=_7LYAwDjI2}<1zD{k
zDza8ZG-SPs=*UJDF~Te@c<Z_M7d^R^n2$33dglH3e@{5pBV2J*=+O<1-b*njsPM_b
z1SsYba*$$fAy?DPS2Xx_G;<8uPcs*gYi-Q;Ht?Hm%spV%OR?C`1}GN$*&xMYKU+<+
z*w5C{EcUa0n#F#$*2ZE#+iYX8pYu{2_HzM>!+tJEaoEpQ(;W75bu@?loS){fpR2WT
z*v~cFIPB-W6p#IUfa0;A4^lk#^VKwu{d^tGV?XbwdF<zFZ9Mk#%{CtU1urFFzYw4V
z>=%NRfc-)>EnvS;M+?|5_-O(Ag<6|{{X(-%z<$w7iP$d&C=vU`ASGhISWS!AFV@i_
z_KSX6#D1~XCSt$XY!k6x@=_A^O94v4ekn*v*e_Mn681}Vw1oYVpO&y+s<lbjFE!gF
z?3cZijQw(elCfV7QZn|-)wGQLavd#WzwD=F?3Zh8GWN^OHW~XBFQs6=5}*|9SAvv+
z{Yo{hV82pFE7-62X$AY0TAPCXO0!MDe$`8<*slgC75mj7rDDHYO{>_i*3l~VtA1L=
zezn%7V!zsKQ?Xz3QX2Ma0ZPMuEl6qDuT|3;_G@*thW(nK*05izwQ1O|HQO}o*S(aE
z{d$1Xv0o2TI`-?;w2u9H9j#-(?x%I^*K2J$_Up|y9s3P0WnjM%pbYFcf|P;%Mm23<
zzfngU*l+l01N)6yn}PjCv&{(i_g{K5;?JfR{KQtA`$X<|>kD@-cV~6aintHWn^<zc
z<UDJMz_s?kiW9HjY-@`tj!%s5X)P?o6eo2~O0@3pF~zCLsp;0TK}>O4a$35zLNQca
z^?X9_cUDp{#oeXuPV1XnOmRxTlss##z>>`eg_gApxK;WHa|It1e1t{je1t>heS}9A
zd_+JNeMCf-d_+Q)eMCl9d_+N3eMCjpd_+UmeMCn#e8dQI!LSaQjqho#%=g&Y#u;n9
z_12pj8x>lmfur{@OYz`6EF$OO5P1)eD0l=!(IX;C9tlzQ$cTzZK~y~|qUO;Mb&rl{
zcnm~mrc<<z^ToV@#~tH@5$54%4ah36zCBq}jgAVfLWK`MEh8<<ni=M#a-{XKXz?q1
z-}t8fa&^m|#$-*j_VoG(QsQy)>VKGXEn|E}k+s|tQ=B<I^D({Y-@W1EMW27t@*%W$
zrdog8(%U)RW#2np&6hhQc1X3p^SpJ{9(`d~c7yqHM&}IK+NCwu%?aDGJ{JC)nHs`)
z@iVI-EHYO^IAp$t@W?_95s<|iA|gvQL_(Hph>Wb%5CvJSAu6&~Lo{T)hUmyf4Kc#J
zVbX6sJ{?{3psl%_WoOdEkCd#P1>)=+5O3#!1iJtv+C?DAE&<7Q8A!1!K&o8@((D?L
zZr6bfyYZMlz47|O@s~F&H$S}Wg6#3upI^D4rqzhKTXW5qySD3^U_GZ-ouQLwT%)g>
zFFR$YZcU8++{kIIa<<+uUrtC*=xwbTTRWqoLaU$Q8GsknLj}gQ$vw^;HGen9mebSP
zrFGdef460QH2k}%8L54&sbzhQFr}7NQjq$l3YKK;kwk@-NiFFTKjF&aFRoH%Z(?#{
zs<kX?t)7~*wq@zmvds^dV?^tKVv1Rg1>!gki0619ffIm4P6U!T2}tH-Aca$aR89rb
zI1NbWbRdH>9@7nb<qArvT4^3*r}R#Jtz)x})x4UttpA3OF)1l2Gkn8J%1UA^dhlTT
zfu}B(H?>?gjEu3?Ic{DyjLgSmy=uO9Eeprt{W1-NxrJwZ17VT52Erlp4TMJ)8i;@_
zHV_e6Y9JD_+(2YxrGY5OY6DS`wFaUg>kULlHX4W#W@`=n$*9oUcla>lQ{ywN8<qJi
zkQJXDCZ>>?6rMheo0$nz!jy)IEo8xD-7J`F2@57y$breZIWV~r4otp~2a|X6VDcqA
zm_nfdrr;L96iNgz#X=EG(Jg{0mWW_Vg%X&OTLM!mk-(G-WiVy845nNngQ*lMU@C3}
zOr=BtQ!P}%RNX3=YKaP_R;Yogxiv7g5)DkfPzO_Y>tO07I+#YG0jA+Lz%)vX$L2iu
zoGC5;i&8poYU%8jjcF~z|Ncz1HX<Tsgg4#c&7$z;^v9cYu{5yZ%9u_4Df_IV8Sbgm
zrq7I+`8bW372cdWciNnmByajq+SV$vwY^PrOxyOc?FPr?I5Xm+1`p`a>52HxT_jo6
zbmQ;+`gC+CsY0(F|KF<#-MS|xr}Ruq&&+x<`>DQvV|g)&>)M+c*fHU0*K8_xSpPv}
G$o~K)5PE6=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_splitk.hpp.CFD1B5B13FAD53CD.idx b/.cache/clangd/index/device_gemm_splitk.hpp.CFD1B5B13FAD53CD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..12132378d9f7fc59de394294a1a5de8b0c3f351f
GIT binary patch
literal 932
zcmWIYbaR`-%)sEB;#rZKT9U}Zz`!5~#Kk2=nIVh}3~zyQ6?69XM`tq`3bcM-{v+Z3
z7RPFSr3%fZTb4Tg-gm^}F7NEKEh%AwFYDj0+Ip<Co4?gb;fb~J^CvEQpRGEanR&27
zMyfe|<B=x2=-y|mb{{{?5dY?*#)0XLCnF|Q=*LSx)Y@Pvu#R~#^Fy@_yRJ`YUfBBS
zL)JmgpbRtT?Tv5OJC{4UybZ8_6}o%7)YV&zlCl%$Z+aNVD}8mtt+IvZIrnZ0UXh%~
z-Bg-oFLgM^v!cPKesk2t%Yri|bjP`{F;5rCzRs!Tydkwetxs}?>?FzOEXNji7fwI@
zGbj1PPtKjQqV;Mtc;^^SOXAb{nxwX8mT@<C!A1io{k@-jrsZDfUfH(ePZ#6aDHC_i
zV4N(u`q;)JVmY&a)E*6f_|RtSxB5$yHpRcW&bqldV9Sx1Q$m`zTUOWx?7p1JH!<Sj
zUArf1tabRcO<U*5bo}8LE1uAz_lrsNS^nhW%G{&|U>LJ#=&oO|;4U9G6AL>7mnoa6
zB$!~}WC8{}10xrs&y{5hTQ_Dr6kuZG0Ez$smxiW>H3O8BAsu{fuSZ=T8xu27zXZF4
zIG8}_-{+k%*=hCEUjj_5F#S?mQnoPtMX71UzCf#B?qK8s+wZAT);)2Ta~;ffMozG;
zFxz6cTjs>5sVFn@a7oEXX~67A?PFDK5Gr;83QH(T=)zRPXeKVO+hAt%aGC0u+Q2OC
z_F`^%+yCE!org=-N!Al)BoCLIlN=IH-bo&br{JW3#8Y%qMB*tqDIxLHoYdgFwZ>Kf
zN9QI?2Rc(jMnfIuN0^mBzk)&t=IW;gtY6sq#k-k-eif9GhN%MT71WT1=>_^l(Nr7e
z7lxwLoJ3F#f&~Zz%rJ)HqU2(IP%LtS2ngVMxqRpAf{mw7F>~{9GVs8J85p<`OaNll
BO`HG#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_streamk.hpp.75A52E82DB5C165B.idx b/.cache/clangd/index/device_gemm_streamk.hpp.75A52E82DB5C165B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b7f740feed43d677b42223d33753e16dd1658ac
GIT binary patch
literal 924
zcmWIYbaR`+%)sEB;#rZKT9U}Zz`!5~#Kk2=ncj>H47Y)D6?69X$9f+&5NZ1zTxR&*
zjDPlhN4DPrf}-qI@tfY<Xfv}^^t3+qVSlvs^G$0vaXV^EvCIFor*`q}Z>x^~xwPOJ
zSM*`^eM=72mqfh{t?zbcw7dPNu;D!O$(HJWVYOEex-PiMsmFZLuRwIgyP}X4f1Rvq
z3^R64U-H`ddgHh2&gPD;j~6<Y?B2ik@~!!@6FxHjl$5@go$#3L+77)hGAE_Yc^7d`
z3wmIy_4So*(bo!(W%u09+Z<l#+ogYa%93w8()f2R;d#r}*<;v$YhrKa<%2i&m>HF+
z*lDcG+jCk*l3~dbfuQ5h@0=9k(Y<gi=5x)<CCrYeCOqCO^Y~S~$uotmC%P@0P9*H-
z7YaVG>(g)1^E(b-zB1!~ruVAGjjP4etk<U4+z=5`+%-?5{Kf9Qb`eVat($L8oBVG@
z)q+zhYi;WTrfAsQd?C!eDB~zYab<2&12AlzO4fQU>S3%EU}EC{a)E$LT~pnf0m_LA
zTV4OKJko=WiJ6^&OPpO?0!%P)G6Ca*fsu<bZmOyLdxd;%ZYCC>0%JB~DKLRhaJ}g6
z-|2GeJ_<0g!Yq)~lC*_cP?VZh><cst<_<<qu>CN5VKfsL*j`VSvhInyoa=bFjCGCe
zV0OoDx6FxAQ&DE*;gXb*)PUItlV;=sI|-&YwU1S`L8#aXs9Ib}Tn}bLw-<BE+y4I+
z>^xjDPBNY_LwLAkon(=Ca!ztcJb5R1B%Xqk0uoQrNfC*s>ZA(itu?j^I660BI?x5`
zGV1Cuufwba1_3C9U=Dj~!1{%qU%Z<c=sQ74X_zXYUO{zfm|kFjC>iU(e9Taknv)31
tIWWI6zzkz3E=n%e2Sqa{h=2gDFPRz~D_8xEW#;DLWZ;1bGca%?m;j)*QEmVL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_wmma.hpp.32E8352ACBB1259F.idx b/.cache/clangd/index/device_gemm_wmma.hpp.32E8352ACBB1259F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24298ab10f4de50d24b0a8124454b61de197ba6e
GIT binary patch
literal 13728
zcmds7cT`o!^Jn+2NVz^99Ui?2il|g+mlhGRQRE9E5;bCtM#bpY#1b_GL1Qln7#m`V
zhy_uk2v+QZ2(clcBA{Z$hEaa|cyHXBe}2Dza*pTxILF=hapun6ot>SXnZ1#r0|snz
zS5f&UA$0Dv@js3wDk>`4^xv!>6DR*p0n$@ZnY+0pCT{wDZ|U2xhOyU7<^?Ufx1)Mc
zz&?%qgL|TCLoV44*I9FBO>5iXeV?9}r>+~*6cll6cVw5h^Iu=N`QYWmWhWMVnC5fe
zeetIL6`6O}#$+t>|8y^{?sUmdf2N1bdp*Q;Rla3f|JC!nE-lt|`W<bavY4+QqS;ue
z`eW|QnuYD}jg~APJ?X|isddWPS!E3&`1h1W=Z>!vDmLC|&E7D+rq%t&T&ra@>IF#s
z*U?_r9fw{V|FFQPT>U$(Zx1#*RZTiuubuS7aqiTzs?~)>vtJBuE%$RSJF79`*|xIb
zt_Rh(7}u{*-;%j)e~#gb1plm*Jrj)F5?h_`Jgg1GJDay}53RD=zoz?#>(7i+KiC|c
zw-vpdF!o}XgK;UxPBp!su*k4rMDvXKdb3aMoUQ4<sq2uJ!wTnq|L9^gch!AjUSYS6
zt~0ufPYc{xd`&X%K)jAl_kFu2Et|D)(X0!>D~7jzK6NWxX7Kr#`PRAjs+W&R*#2zy
zz1ooK*|7u6Z-qpd%r@&*G}mdiS^8hw8>ZLHcJq5$8~s=11vKox@gcka%&q!4K25`K
zpY_J;!+$LNJ<4ePgP%^%y!C8uWt8i<{rOoVA6n-2SLw67&x0RUbzRXmp>j%5U0!L{
z^fybl)$J=q8RKgXMqNDcdgRnRgUs%43MUS2oE?<u@FcaaEH1dD&d0Lp)3-}kX9n0c
z>^km!Y~-dnuZJ$Z^@Hk_qypo-$xgTPAB~jNS)D6i^W*X6giS$XdmoDXx$ecj{C2D7
z(Xq|%ypy7zAI`FS>lqo6FS#&f|HqX%HF1$YpZ^%1pYiI}^O7E$ho&Vb98EbjI(58j
z<4l*mj}r$^K0If^i|f)62Qqrh(6^PUI{T_TNMAVV_47%I`s;?O9S*oHT^AL;z<iBo
z@Vf5}=BAa!yN$i$R@Kw3HgC84+au?;L_dx7T6=6xYVU>~GpwqP46&YZeBO`UtcTys
z9}`#{Y-~6D+-0pRT_^pJ^+9?iWRl;2*Sc>0!LQ$^K3lZiYxM5Wpmdc3BOh({ANZkd
zg4>`aeGM!1$G&QQh_;)=^=#^Ha(B(d-wn99g?k#pUJ+BL^O8UJ=UJ|4j5+F3=oedQ
zd3DExqmwl!Pq_U`BjIi0o7)9^{^jT+mt&4ZITj8cw_g8=sy6?)y3Fll;?WJ2htf6}
zOyehI1s0)|FNdktK6LUtldBfAdbjPTQNK1$$r-qI$gev>UhR*}DtO<so5S~h$T{uf
z&c@}BJilF7zw))}-AcpfO&dbmKC7KQd0u7J;V~=6$NH@(d*$^~O>0X|k{@^4wtIu4
zRzc&%Ia$6%WW?d`)b?A>Iemk>R2$me&~w9YT4(lUyF5s$d^p-a)X%$k$;^S*Z@%$&
z2#oMMFll^jSz+qoupz%KP-_fbIQ870{c@(QHCP(8r8@d(xqZT=>tlZS`}4=7oWS)-
z?E_xs6{~Gs-OoC*b@YrN-{-@x9xglex5d&9*U`H&K7_nYUNATG(!Ab(Sj#T$3wYd^
zm3nZ~^l9<$_N^{m=9zNfZN+!%H5?nKT0Gg>5cN4$M|0<*^L2|WM<>+>lLt`|OO{U@
zw9=z`&)g9ezZ?u5K6}<r)4m(omHMREUKZ~|$7<^1{%$J51la`p&PGLz`YrmWdf0m7
zQ{6sw`j}AiNZH69oehw?-PQ|RaO}fS9c)C&MVz#Vvr%@yVld^U&9Lyr$@*AES)yc7
z)}0NIjeht4skAuKP9LjN@+jG;9-R%4WfKQnDcJq(HBH>rLq%x9k|u1R?106_syVlO
z)A%+C?(VK4+{V(|SWnpji|v$>!e!(7%#>gg%F=+P4Om~<0gL_ifY<%|Z!^)s7W6$P
zODEfOHb8dVaqv^?hm-52SVGCSiS#xxQ+B}O<!=#d@LRP+ANQmzBV;2iIvXGdZp+`2
zRNy*U6Z4e(7->94hRP0Df_z+W^m{e$l>~>-*SwB(uVYhX2P}in-&|d<H+-``)}c~{
z$%dJBHb5R)zwCikOgDE;tVzi&NTUVmDm!3__|Dw0_|bI-eQZEkB4v@Loehv<4hfm>
z?aoe+cCP3eqP~Wh3W^R`M*MVj$mNo4GJUK@B_zlaY&#nu#~#?W$KAW}oenmp<a|z+
z&$%l*U>Vha{RFSBRWTB5MkiW_rFC>Y6x|n=@proC{N<vrgRQBQtw_2R*(f_;nW(bh
zr!@~90(7tqWvNEeYGkkMzOZPm>mBt`FI0l<=|J^ZT2BX3bigw8(3IKcO`13Ku{M=5
zR5sMKvjOr9zowVRPI(QIVmnH{L9A{NS7iq*iHQrxuB)CiQy+7bWt?oBZD#}I*@@xZ
z6O!(PNw6b*hdbE(4t7*_!17aW%X|Nw6*DBbH)Xkjb#CD9$_`i-s3iU;DD-Z$4z{F&
zY$Vc+bX^o3uq^sx!%1KBcUTp7r7T;p<`%50?0_Z7@_WC8BgYo2;%=@gLQvNr^Ug*^
z9V5Egk$_fwZ1cCQ?^!9qed(LjV)I)1vWgC5toH1>>c7pWZ1k}?9VS{9ZQ0oXd5!P4
z_lBGan5T)gD7h7Bv?2p#2P~OQcS_YqT}zeX9&}T;U`Y$sR(8O$YwNb;^e1mZ^|3CM
z5+(~X?QFh~FYGzAF1q@HKGvh;a9Ox%X9MKDj^F*DopPo|A9tbTF|si|IvXJ8Oz0Ka
zv3{(I1P9YeR$$`_Y_05oWq)iJ-}?L~KkH)+Dq*Z_Y>&?73;C?UsZaBt`$}*xDxnJN
zRAF;v2Q0Z~8-!hzKUZmDRVtwwX*AQVsOW$tf3H(*)V$flB-n{cxr@#3Qe`VTU^%w!
zdG=+Es{K0HmP*MZ(mZ0K?0}`<%1mLp=9DZ69zdm3V&h6W8$|~!$NPql>++{*Pbs#d
zQjQYqqr_d=0ZXyPF2fXur0EjuOr=y~gKBDmitY<bRq2tP>$|Ph#|~6Vj4Z~ovjK9+
z#fX;p35#tFa91i}8&=zfrOFOiPDdMO3}00JL?7!@2@$dg)6NFSSAQ$5^?z3JQG)wW
z2{*C(O&p@^faR~NMOFE~WYeWHr7X9w^cHpYiVj$steab+H?7dn$5J}T5ZMsZ&IZV>
z9X0_r9<A#X#+S=Uaye~5(E-cf2Wunhn^N2)*o6-A7v}%MUdj$wI*zUJ9Q@((WL0cP
z2WiBbjaWz70ZZce39~XB*a@qu(Rs4M7UYGQhBIy9*x`k_hI2(|H!|FfY)2`bATQ)1
z!(1_bh_oIeQ}BT3*<yov?KesgFZ{~c{>r&+AAI3O&tvaKaEbGLyiPyzr-U-jrHu1}
z0~^kDtkHR%vzHoYKh6Ujnre}|rEhR<3E_n#q??3#v$tK3bk~b<A<`{Gz1jO-M7kFl
z)_ifrc*l2!!I&5Bk#6^h6pm9~XdvAh81B|LC1c74=Q~nf7-BKR4wjM^hFT1@15eEh
z!z_l`!AZyqVHRO_;4gR~+#=j=m(o4(LWD(x9atSNL|R1J!THAvQFc*|5S{Qs8gfVz
z&nOM~r=egr<8<Vij{L-UP&x`{_zH5k!e-11SCIb|6wEYo4S8Neequc68VUyt>w+|t
z;S;Qg&W8E-9vj4Mp)0tIm@OkcffL@mt43!$1^m2_i~^HUD0{J$$a^IUVmJ*&q@h@b
zi&0=P3WXQpg)_+e3<_fS3W~UbVgW<EMh!<Etg&Ge>=Qg7GBEn<klNq7kGZH3AssoT
zGqn=o8gjbE)I@}K?A?xi;6xz8XYBp?pE&L_9>wS#*t<haKg1ax;*4NZ6JZ^(Sx0Ok
zih=Nv(@+MDuy1*K-WauDOAyQphd9SWoGS=2`RM2PbK(;Ns`mRN_CARNi?xs}_DWT|
z+&~F0u*nPT01~Wpy*(fL8g!wAXGrZCGKPcDsxWcV(?!0?h|nhtUSMN*!rK4H@mHP|
z8ed&Y5JA)Jc??UAVI6Sg8p=2tA`6J4DPswilwe)Ps0gZo5v+^qT^>p^D``eH&Bz1J
zMqVhzzNI(}#Pn>)p7Zidh=j^tsJ2kjnSs!T7mgscBj^i;wV<0t5s8ER-eCzLF8ltw
zybBR-Z~-^CAlNKKxXT6H72_5zphb*dase;JxPu$k!G*)~&X2a_nv6L+oi5c9WVHlY
z!vu-25_zsfUJS27j;l~FhSwnTHOPYD0%TQytYJG5y4^g_ATNf`BFD3+7sKVqyc}5o
zb{Y6^nW}NDN9PR9ZgA%ChVmJj-4){&&a6d@Uvg$I#khkr>tNV3C@m<j-Ae1L&(O;?
zIr84=eQ&4@EJY?ukq4-b+8i<|K^~wA$hW#^bmm!L13X=k9xIv4;e}PmX%%u|V$Pr*
zXP67(g|o=%EOG&{-qKTl<fs4lGadgFCq2a(fQEVDJjb8sbO3vszr6Up>e!b&5gG~K
zNF-pEM0iB_M`GMWjGKrFqc;=dW-)Fd#w}v}m>546<0nMt3DINEKP5&_iHzY^V$>?e
z&xqkOV$A3-2>(L-J}(LXl3~b9>1I`=rA(FYknA0@2T#V+fKaU(>qD#v`6^Y_K`c3l
z4PhT^D6>{@2mLQ>d>qwo<3Sm-IlFAm16&|4?B(qCG8|I1*(Y;k@GNs)nC?8?r}?;D
zbij&XV<-xpKAfRB&T~YxIiYhzv}7brMl!bA$;c%c`7)e>q$y&0Dw3v(@kS)w_)qMz
zQT%*5vPefgn7j-m%@FfvB59@=XCvoq<g<9grC1N?m0eT;C^Q=lXE+Br=ZNV!C^Sb*
zFGkX0Bm?W@g<|AVjC>hBfutwI^im`(730fDdikH&<+AwsYslgn>cQkyAZdk|zY<9+
z#rQUIyN&!H^5lg($nB08*CDq$F|J2$^$bJQM#Ber@q$@eJC13`v0xWN!}=MX58f0(
z57Rc}-iAEE1fjq|G?ZlxCgClTy+t;_$kY5;?FQC{)S4GkxqwtI6e42?1L=#%YamQa
z2af4rFEU&`xMbRR=}L~LM~(Vzq6M2u9?9C>BaRF=5c>w=1UP!^hS;3#cf9EgJ`vqd
z!~h5on$W{bQC5MWN9dM2kI6N%Q>E6fH@&aO+V$p$P&U%bMt#Nhqn9Jbw~^s(<jU5k
z4jI-l3=taDpS*AbWAnmMq<@qhGrX`4%hq9Arp-2@)y7;3)OhG56a^j-G|D$qcSo1z
zAHMoA0wpHua*D8|NF1st3Qgdf!1Ag0Q&gE?Dl|kyHZ8~lOpc~d*td)YEAhv?rjF`v
zYeTc6E!b=e?g0{Mw#FIda<&jl@^qnX){s81@>JPG?<lc^tbEM<l!>Z!miMTjz`lY0
zZ2ShI-9RkCaOAntzy=b-@O?7yKEvO+gcj+JJyAd{u8x@25gT9xTc9&k6v@~KkCDw|
z<P8S^FPy}_C-HD_;}ezDGpfU4EGZW2P*FpJ8AE2KuF+XLWAdlRPaiEZ<fu8(3@vb8
z;L6-+Hj2td(PFcX&q1RZ&O^?5$Q2e#?sf*=L4#mwyikw)>ro(JsM=BcRFv%Cm1Yu^
z^u2vQR?z)?k7(Z`=Ab0G(z?~LLsX7k>WDwkAt$EMhP=WD@_1np(qF{lJDR0o*)s9Q
zd5_KBV|#FHyzq)>y<!^)ii1?e1z1vmO+fb=$`T={82lJDlEYYX7+V9Q+${MrY{_sr
zwkpT&fT;#L>xm!;3W-!Tit-^$1sG^Tsi++Sfmc>>fPtsEluHx#hUALwFs^42=K_Hd
zbsnVOI^xglJ&R~%5knZ67fut))5HPne+j<n_tCq%J|R@+JY5FRxg0ywWnj30gwRdE
z@O=_OHvwRHVLAhOX*1Mz&hj;d-z=1888)TZ8#E%<m`#}&pTssN#kdI@H{rf)ekq)N
z3g-t$hx{n9Ph~hM>zAL0?m1hoPNNxQz8BduD=p<rN;zAmm{QKSlnZ9Kj58?{)30;J
z*EwrOzr~r~;_Ml2;*6Wb^k<y;Gco-WXZ(rLm%a!#^}6=EjCzDpE~b==2SLjRaY?PK
zIy@bCVWP#v?r;Z0k;BA3BIG>LlcCABlO5T9o$NCCOVYy&lLID)FxnENvjiD4ycOAQ
zMUG4>Tan9Fmh$k_?*?o|A&g#tbPA9$!_~;48ufv_O`}lcQH=yK?q7`tGn|FBv#=3c
z?wweBrx<5r?QAjLg|&By@oucWTa5Q$?LA_=7i;eo<3F+XpJIFqtKGs{P^zSn16I3@
zwIIu;o&~GbVJ$eJsPDpR^;iqME%ozQtpRI6vQ0xEtkcHI2E33?%+iT1+fEt8EJKVl
ziCLx?Zzg7&#rO|m_J<g6A!b{|_&D)6P6XJVyiiCy3dOjHcod0oG4UuC;}YUgBE~0(
z#|eg)Kki>MX+q-P|9T4h`>W$8pMPV~_^)UGD;{hPbNVHAJay86fq~zEn~@*Lu??&=
z1a~Kttq@BJ#lBo|g8~i|ZcylfEWbWs^5WIwwV=R8ml^q|vkK&wP)tL4o5CJDgQaJ%
zDYN%;SnC`cofj@*y^G=lq6TZ%U^A$4^R&dSeG{8O^_>@<VCfTV%8slgPA`eGWbSV{
zXSJMjVt5tTcNOQ$a2BVP#hI|@ml3UH!~~qr+G!;n<16C@>Xjy{P1FMiv!S79S-^>%
zkLiEkB}R9NEqL_}@5aP0(K4;3CtnM8YQbJia0?#Xf@2syz!@D7$6W_Fj{{r~#9<pR
z4BTCnys#HtpVLI|G_hb7UPJV1hy_FqAdnuBipwde8g2=y^EHnc+^rj-$2k9rrO<LL
zqhB(>I}_nG;a`jCZG><8CziE|`QH%!jhOzH@NfT#P2LhSCjTAb--+q(3IG0|xYv8q
zoAG}j{0A|;o$&1p@BHSDS<tEW%>Pq0+5Z_`H1@LgQOa0@EoyK$v!dl(pXJ;bumUK^
z)BH+)d&i9RGmiPp@++v+Q#Sc69%Jjk9v#>dR$@=Di1|;qZG26QqlD8c;jF-z;HHm0
zFTdtv&yVvS7YZ@k-k9rt8ppnOqM21PGD$|x>}cAEOg4(~CS<ZnjEj*;F>+={(Pd<E
zS&XkDldBBF#Uj;CE|%nq3nGeJM(_YRbBnr!Punqo-gBfNlN4mb-Y5n2Oc7VBQjtlj
zn4XH9Qjt63&p;*_VtNMZnIV2Y6PaX+>6yqi6Zx>`Pau;M$cD}E1nPN$Rl{iHfJ{oo
z^it$hirg801v04+(<@NV3i0!m$fQzCuSBku$OoR^@9G_X_Kz`XRAY_Uv=Q6GR-^el
z(YwL+BwWbSg_hsOvV~6YP6&mpcE9pkmb!{VSaJx<m`^Ole5v@1F2}m%*aGMp%FA0=
zA1F}MG+A+F%f@?yWN*Y}hAOO9g|%U1x}ID>3M(@lNL)H4U5KK)pAz&^xbCn&<Uyof
zsu(Zh^p=V7Do$?|!}(*h!sZ7Y%A<<9i7jqoH|DNaaek|~;UE|;l_}$KEIH230S)EV
zGR%?QZs!{=d-dm6zkWFCad^|b%BKVA1<75adzY0;;PRXLVa3fitO8tl(>Wf%k^|y>
zrnvZqIf93$5)`-IAfZsAR&j2`!=eB0JUP^6sjth!l00k<CN59z`IF)>;SBaXg9kEt
zIqp}ELm0k<jV@saz_fVK`B4C<w)lK??0%oujbGh-ij|)w&}d5jO#-YiFT|<GX~8`=
zPcI(*ud(}Bm=!e)#g_^&trKpM30;DBJn2H8caIa;^HV*hif(^*Adem5+uWPT<0h+e
z^1@Zl`6|0@FCCO>-FjulsjgILXxuO+G=4~Y1Pl144w@Rq9P_NmSuqUlK!bLmFmRjH
z37|nYQ5eg4YEfh@iUEA`z1kZG7w4mx2=9sIdt$``fp*fpo!A40yK$-q`R5p*ywiKa
zf9N(qEs1Wh|2<SX^LE024kVUsqX%<9uYeGEzPu%4(I3qomFh%jB774IG>Fhl_+~L~
zA$*G%KPLQRF@8e$Cu01R@K43KmGG@%{EYC=#P~VkpEG<Zac{)O2h(Q#zuWh49MFR%
zA4~GZ;r?mNpBDRpavWNY<6sA9D8I0Q4RFo3`~1i-(~Vz!U)Q9SHTTr<SwT%J$RWrD
zObgO#YMMu}<f!<vTJaqaOyuToe;KLHmeu_0>ENyP+aDafwS6JoT=$66J>mzOiO__w
zPXieMxWV(<H}7(c&{tRGL!<0os~>6graD{C^<K}pKwQlW8@S#Z7;fy_rLC)$;c~fh
ziTY8Z3kRq?98<qRbRk88FLP))RQ`n(EDkTscb+e*>K;bUhs9OfcI?s4qI#Zw?cwp6
z;ig{4_q&?y+d)@+EoZovbAs1=qL#A?dmmg)opQW;d;s&9vwUU^U_Np$q6N5t%tzKD
zpIS5kd?GypIOl7eFQg8wA4k0(nwi{B?v#z<EJRM(DBgk%G1YjgB^zRv`7Aqzb|B*&
z$dV0s6B*w`mTbUUWL}Hx07D*1eXZhqGl+FxZ9ZPw=b)ACSGSYb(w3*criQuw>!q2u
z3(hSx_;^z5-xr)5T3~ze@EfYSulr%f$B(;8``q2wMs2qZ$=Z-B)N3F?rZ)~pu;d8#
z0msV=7qIjK765A~KQRRV^||EE{V0#tXezOe1lEy3;8wtFXmX<X@R04kJYt_mdPCey
zb97=^L+oMG%5jex(id>z_^D(2)9-8g=V;#7d;Uud0t<z!EfshdNYhkcY;evhF!rp(
zakDndKiKh-H;O+W_@ScW&wln}KLgoMU-r|F{q+3u6ASOk;-s(E-n-Z>=$k+<Z+{<O
Yzkq(8ekuW9SX91YKZDp$?=L_92Vn(l?*IS*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp.48B71A3DBBF3DF70.idx b/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp.48B71A3DBBF3DF70.idx
new file mode 100755
index 0000000000000000000000000000000000000000..faf863067fad3873b4679ea3904ae6131a328f99
GIT binary patch
literal 6924
zcmZvg3v?4@!iF<nXbo+Fwm=J5EZ2l#cFfE;m&^<^Q5phm34g5+<W{2)?l%Jn2z&DL
z3yYEzgr%S?u22;5msJ))_HfxnL|ItC-PH|2Rt38l2q@|b1tPj3=Zt;-bN-yu@67wX
z@7pK&W+s=fzS{3^cdyM>HK}^htnmED+u3ZklI-8S`H^WK+HJOr1vcBF&4+@ubIy5+
zzkK=ZBmCselMelksw_P-yHS2_?e;%Eyyf+^z8CuJ?H4|F=equJ-x2G-?tOVSHSWuk
z#}<1h<sY`q+%S!)FCJ0$_Mjy<=l1EgbaS-2>`eK_$I6dij120|edgV^@yfd^F5bR6
z|L_Z&J9YbT)Xq1r^=Y}?^2K)<m$vk$cZMbQDS!Fb6Nff6KDV~t%cHMMx^w-D*^B2b
zX&OJiBsJx5>ej)$;K5z}zdI5exoB@o)2G{NmzUf=-<10MhNkOJY~#;1Kd|H0_Mmq0
z#IOmk-ORtWuXg(0Cy(@8-DCFYZvQO&-R0_iwc@f7Q|p_SmEHTXx$g9s^60`#9k<pw
z*Bq^yJMv)W<o<eJ@9;?58g1L#zu(on=0)55PK~1zduHDH`qP!I#XV;%&#&0Qb#)Fn
z`%vQf<E|qch7EuB+NkIAUU}iUu?wfy@2sdPag2C)`UPJv_48%#Z14Br#hKII+omn;
zRr$=owf&aP-CLBnbzk$P0h{VOO#5c*)!`pBbaZb%8R&52(Rl-AeYN!;OMH6-$E$Nk
zuIc;ld$dCR!unxN{YIA3;~S1e=67s5v7~0uq=E$n{F$1mL;lCUd2o?jG4SfmU-j9x
z-na66`*T6Q!<NWIdMo=@>9nT@E$Q@P-To<_?=L;|O7*M%THi+*-)Z?L-B&%ZV%>?V
zH{V~i`WTxoyt3m*PF)%5y?Xk4{SUamyYY5&V)BfupB--Sblx`O)315AusrZm_|lde
zeU9C?!%%yi8}q`RQ;S|cHANvCnzpZ=_TrK*<B}^pEMGlgJU_*^_=f}jzOS#?IepLk
z-gO`C&!4mTwW_%T7d8)UUiC>Gck0^j8>Ut1U3NdaVc3w(M>2`N`5)gr`?NIYw}nrC
zP`^O`^`4_+o@sn1Qd@V;n7RC)^$XX{8{0JS<cMyM=8qci=YrQ?IdS%rl1&4Zmc0v~
zY(M<8dg|HZw{9+-FElsRT>s7c`|rGee%YZ}FJC|KzJ9#r<hYH=f^+Kd_1`{JM;15y
zW6-hPOZTol^LmQ@di~LUzx}nbVDP#|*P72~y*M*af8Ry_?*Hxf+Kn$xpD6vY;-$tB
zH~rlkcGT3!e<~d}{_o<=DNk3vO7ES+d3Mj7Eyb=K?BDo-Yu&;9KknYn$C;-BeHV@P
zj7#Mm*t_ILW6ARePp_I1(O>HLw)E0FW4C|w<7e|vG`!*2HgmWdZFwX17xtI_yzref
zT|2bQ*tDhbwaY7xjvnxvZ)cqUJT&TWQ~x;i+0R4!-rM6zYCo(x^zKi~r|z3_Cv(-f
z{Nct+(F>=3Qf?obcq)IuqWk+FdTZy6ZlR?$`n$iHJkQ+TvgV`O+jEcfn>FIQHGgXk
z{bZYZ{4!&kw|LeQ<&@2qU-^gCRa=!_g=FMtTSMpk&h2wU_Rh~e3T|)t?3r5WkwWrF
z*6yC)y(BkScF9}M^;qy``n^I@m$i%Yi`(Z0%l2lz{c6XTi{C0Fqq25+e)&DQ!Lmnn
zuR1-s-NIA{TYc89D68m~8!S7xV9tiWezEhLB3m$P56&Cx$qkl0dexGejzwS96p}}?
zc4dC0lp8EN{dlkatoYRfh2+1oc9;Av?Q?@=-zg66{&ddKFAK@TS-T{^q<wC%Y+EEe
zd0u7Lw)N>-kXx6=BbhBbFX+9B8eDFdU_FJb^0@RiMZ0P6Xb%k@<7U8PJPdfOn+1>c
zu;6iS4m{4ofycXf@OTdop5PY16FdTVqFV${^oZa|ZV5ceBY`KoW$<K=44&dvz*9U5
zc&b|kPxYwaX>JWX&7*;*yLIq%j}D&UHo!AH26(311kd!C;Q6}w?)~`+Yw-2-^|F+8
z&5H%OYaLE`*fqE9B)cN6UxJFdR#9z|4$}~&BQ!+mD2-7j%s`ZhFc4*;3`W^73sE-0
zLX?fN80EqoM7anDQ7+11ln?U|<s&>q`6!Q3AuK>thzJlBq5?+6un18xB0^M*iWrr`
z5=5nl1W_p}VN?#w5S1e`MCGWAQ6;QEREa1MRiX+;)vyXtHKIaPjj9;c!Wu-ihz3zD
zs$o<Q>k!o=Iz;uTj!`3QK-7pB5H+F(M$NDZQ8Qvf)Qp-K_3^$z)<s>^;4^*IplW?V
zOIh#YzJlDlH(c;u*Wk8m+U0jug9^CnK!scr+Z64kp-1~^=+OZhdUS|}9^+-8$M_lO
zF#!g8Oo)LV>t&(G`dR3)0Ty~}h=m^K<)Fv;Ip}c#4tiXOgC6hYp~w4q=<xv_dVGk7
zp5PUrC-?>E2>}6mLP&s~=oO(S`bFr80TFs)NQ9o`m7pj2CFn^333^gUf}ZS^p(p!g
z=*a;YdU8mHp5j%Yr}!1<DFFp~N=Sj8>Q$kq`c>$u0Tp^`NQIu})u5;OHRx#p4SHHg
zgP!iyp{M(G=;;9+dU{BQp5Zm1XZQ{1836-&M#zAk={2Ee`c3GW0TX&=$b_D+kFURV
zlV)#Vmv11bU-;aXvc6|MS&;jF1vhY-a$SLsC0B-W{eU_~TyfHN<GK>Wbp~~cxGte?
zC9a#OlXllxdt0AQQ#A7E3`HZ4ju9Gpbb`>xqf>-N9^Fc4<k3kxjXWkzG00;w6oWh_
zMi}HV3Bn+cNf8ElOe<lK$0Y3x^4K)RB9F~bEb`bGVUfos2#Y*6MOfsqt%OA$o3yjY
z<I)s|JT60V$m3#!LmrnP9P+pn;gH9*5)OG>(#|1|Pg6Yd_zcA(kB<=^d3=KK$m3Il
zM;_lwc;xX(JC8geO$o>oGL(QkAw~q`2?-(~Pe>5~c|t1@kS8SV0`kN(B_dDEP$Kfg
z7!i>tCWweUF-1h=iLFFLo|v?Y$dl5PgghxjNyw98L_(gFAQJMV6p@f8wGs(=QqnFV
zPfk-Z^5hI9BTtSI8F_Mo$jFmZL`I(6N@V28NxO_ZB~2;FQ!<o-JS9dH<S7ZFAWumV
z1$jy<QIMx3?F#bLG^HX>%}^@x)EH5brzVJsJT*mB<f*MhMV^|ptH{&Rl!iPlLuts<
zVnjopmLMAPv=q^hr?nCdd0Nu0Ax}?JI`Z@kr6W&|5gmDYg6PQ8Q$$Ce-b!@j=}Ehe
zJR?mR$TKpOfjlEd4CEOJVj#~*5d(QfD>0B~B<%+B%rs>p&&*IJ^2``9k!L1|i99n!
zOyrrZ#6+H%w3}_bhgM#--THq1CsezJq4y7+V0~H!ly!5;*!c5QQA2l6_bTh>-pCQ9
zu1jkN<z=U<+E;b5($*$LNkMLN0zO;glxq`wmeO&GM$N=21~nU}Skzpc;!yK(ibpNP
zDFL+@r$p3JoRUz>aY{z5#3==}8mCm$TAb2Q>v2j)ZNw=9wHc?(Hf=2rRgl}1v6jX9
zDnJTyn=#gO(f>ICPu2QPGLOo2*_EysQZd?E;pM$%n3MMGTUMOiPI0&@t$$V2Ks#tq
zjDrEiI#^Jgg9F7ocu;~v03|v^P?AFeB|BtLibDaVI#f`aLj$EdbWn!F0A)JNpVi85
zhE%8@#(QL6e{pefnf3GQjVsSA9&=&y>g;r-Q90Op{W}*O$g7*x?9Wb@cPuZnzU%!_
zte=!Or1?sAx^w@|thJAKWTbidvC1>YveQG`4eeu{ro_g>mj;&VjoE3R!&hoepKtl^
z2iKPDI+vZU7*jFPIujSUyzlSnR36JtS9Ptbw9?j=T4_OUE3NJ1x|}ZBdb-)u<)lF|
zP6ia~WI=IG4ixX?K?zO)l;{*eNlpor?36(%P6d?eR6%J@4V3QGK^aa1l<71<`8xZ`
ztVOuu*tv#iyY9Ywbm`xPv)07gPt3ke+}%uswF$b$z}f_9P)v{k#RgeWT#y6B2YFCJ
zPyi(cMNm>u0wo7!P)bk%r3O_{T2KR}2X#<J&;Vry&7W0vpLFKqcQ;MvMJCRB<nF=P
zyF>0_Z|M1b#p9LKu7Br0HfH?TaT9FgEz4#b({?zn?QlZd;bU!w6K74DIW;S5CkO29
z^78L-v@a~`P~53wb;*z_XGuZz;L`HWUAlG8ZYK^hOz(sDcV)D`@_@?!Z?^2-ik`iC
j_vzQ)MGyRio8upll>5XUtl;ZfHnb*0{39M3Mh*WD!3tf_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp.9F95D8CF713B4DFB.idx b/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp.9F95D8CF713B4DFB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..03f25b73ac9324239f13bfb318e8b90fcc96eb8a
GIT binary patch
literal 6924
zcmZvgdvp_J!p1XiXbmlwwm{3JS}2e(ogFhX&LuO$Oq7N|DPe1cAh#NYaKD+NfUqaa
zCoD=*5Tt-Cu3Qvx#T#<j!^K5JSrPE8mjq;0ki|eiS^0zl5&c5WneFq(*K_)Ip7(j4
z^Ghf1WF~KYjo;s<hs{<!p=REU@SI26*lf1q>~G_o$dpg*Hrx3En{D2P1Hsx^r#(gA
zzjW#$e&U7+2i~A6N|H01<Y!lI{l|kFUt8&0SGlWC_{i;R`WL;2t^Tps#hKKY?~fgs
z@12l;&^CR|6sEptSn0d{7u=X#*>&NDXiaIdZ0#duAD@r(@5X)O-Lm%5`%BK>x;*FL
zx(yw>emY|Nn^!8&-8%Q()r?D9_}lKmi9O0E|9<Si`le@B_IYXKr3tsMeK&J{<AUa~
zV~bOh4yJDI&kOE<yYJOQvElP}oooJjOYP#~TW6Y6pRH-W_ShEw)aLuP-P{_~&L15-
z?$sOlH}}*|-SzmPyO(#Ld7|sT3g5U`v!_;EG;DHx^P<uoUzn>-j4F%Hz0hG(opZ(E
z>e<8hXO8Wy_w@=7x2@2&y!+1Ey@tGCo71srWMb#^n?HWNw590oX^ZpAw{cyZ{Z2iQ
zc<y7@p*4etzJF!Jb9pbXdv^5PsrB2-hZH-8JvjBOuc!L$qTO5j+<$)hln=IO3wu^P
zUA3~$!r8mpC2roc`9i<-_3fwpXVc}OA2oDvZ#Wicf9T=Helvd9^yLEIPQmfY?BOeV
z|L<K|p?-Gt;O0KVOX#r;M<R1NG#_0sr2mA1CkpuFkjVr8$G%}eJGs2-@{K=LZdvVH
zdZz8^Am4stWIVlzeXC^3Q~eioe4%ddB+t(mo_x9Hm4C0URK|8({8hJQ_bpj<wEE2t
zmn}cSrVB4^`;}9dhI%cZ`a$1)?yJ||-JF;>?eaGV8$6x1O#Avr-YqN+ycoW)@p|Qv
zd$t*B_tT@+?L0p3rQ?$nvZi_K@+mJY=sf1+5)aGQj2p{O^3DHcpTGBOOSVtlIj2|M
z=X>*KZFseMcGcX?RhyT6RmUB_@=n8)YQ6K0XVwfJxZzMH(L4W(8>gO<X8on`sgLTP
z(4XIVc+}HPyCb!ASB&Y4|5ZPCRpaR9s$;{tKAb<I-#-dod->?8uZq`KDd%?0eZ1|^
z6YBA2KE8Ql;T&Oe^N?$Q{&4T@56>(*Fyp0b`##h^K6h-)+LHyR)uF3@dZ3QXZ}_tR
zksS+ntxUd_qJLa{xX)jnH=Y=<s>!wD+Zivcn^AwydH;^Tc765wcPEaP{968E)36)<
zZVlUp43Ym{GG^>&;*Cj9RlGv)n#Fl`OrI&muI%sI^pR`T{=L8M*viM5Cj-6bjr5F3
z<?Y+G;CfT>bNf#$n-tMs?C`Gi;_lH~KmYZcIY%2__iULyRE?f{J@yIv2Y+7pcCt(R
zbJNyuY<l(LlEWkW{rRV9XTA-M_~+#JCO`9SXwL^bJtwtamK}Kix5bn9OuC)9Y+U?w
z?S<&s<G(4l4vasZ-*4W%eGj~~{d(8X!Xf(mf1226Zauf+^V(ap5A~Tb?COetZVvrs
zn|$&jV{4p0<FPWzX3MYm+w$s7O3y+ve59?RQ+}tmxgmSy=Y9%qZT#lxTIr!e@=(_9
zmfx*7H&}M@ThDfX;?492g`_TP7v&eV%?*~#Uj6Iu#V_Aj+s+ou+5_?icyfbfdow@%
zu<iT#ZxxaeS-UL1?5^Bk+3R;#b!}IAYDIfnebz27E$@>XEc?C6&!)`w)buDM4`=O)
z{0b>ISa$l+o_ksGhx-c2@3VI2{LXE2gJs_?3hwxN*5U69$%9$DIKQ}UZm?`yBs{UP
zqD$-fDJjUEm!?CRjoZ)aJ=-<7JTBdO3GEta4-Ja(FrZiu3ySk_pm+}tO7IAvM2`qc
z@<^a$j|@ujD4<l23QF^6pmdK8%J3MVOpgi5cek&n_5N9_KrG0;`*0G%uGy_8!4+}+
z0aVnrjB1s1n1(1Fp&?2~X^b*q2BJ)afhZGYFv^Bmh_VqDqHL7KC>Q1+%0)Pca#0SW
ze3*wQAK@X&M|q42VF98-M1ZIe6)-A>MTm+K5u##L#HbXOASy*9h)PijqjFe=s2q_Y
zDo16EDq#hpN<@LE5>+s&hE<5F5f!3pRK=(k)*z}yG>B?Z4WoKkho~OWA*x4pj2dAB
zqDI7ks1Y?VYKBdSnh_JCX4J%}kN5Sr7InJ@pXsXsRqG2{%KBXM734n8V8MG`16tR#
z%kQcI6>!yo3c1F&D%wj!kM`5hqXRVb=nxG(#>+sD@iWk40u1z+5Cc8d%R-O!v(RG$
zEcDnA3q8)uL67ru(BlFe^tccQJ>JVhkN5M?;{!bO_z(|0!7D&d@C(or0s{1ekN`c=
zD?(57i_jATBJ{+N2tCOwK~M5a(31iZ^rVmkJ=rTmPxi~ulLIpJ<d6(K#j8M1@hi|%
z0t)n$kODo`t3prptI$&eD)iKl3O&uMK~M8*(9;4M^t6x$J>9EAPxtH4(*rv6^pFlc
z!)rj#@Eg!G0tWPqkO4i@YeLWTo6s`@CiKjZ2|Zt>udj8HW-nluuL{)fd~Qowf2o4p
z6NhyHTTdL17v!Eatm*9CJWaVS!QIN0p<KV9juBU!v|i7y1aT!%r-<tU>K5X<fjVh-
zowT?1=`=+nkIqmu^5__$kw+&8jXXL<Xynl?ghn2nwA09A(iDR{CPOjEV`79s9+Myp
z@|YB1kjJzT26;@<&LEFXQ!Mh>48<akjS&`kY=W@JV^f4h9@|1#<grORi##q(ameE`
z6o))6MmXeg3Bn<dOA!uvTnpil$0h9?^7u5xBahEeJo5M$;gQED2#-8IMR?@#ErdrN
zpS1JH6VjA`JRw5~$P;2jK%S5w0`i0u5s)Xe5CM5Y(k>uROj9EA#0(`OPmB=}d18Wy
z$P-gUM4s3}MC6G{yNEm~O-aa;GL(coDMlpZNeLn$Pf8I9c~T3JkS8VW67u9UB_mJH
zP%`r57?F`DCy0zZIYngT$t^@io}9GH$Wzjkf;=TdDacb|L_waCAPVx76j6|;v=9Y(
zO46<%Pfb%Q^3)8aB2SGG6?tlcsK`@OL`9z3LR93bNxO<XElp|2(=wEXJS|2v<Y@__
zAx}#Y4S8A%(U7Ml?Hcm*G^Ha?&rmw@^cc~RrzePxJUvBp<moL$N1mRv>&P?Glz}`W
zLm9|3V#Gk6kst>0j1)1DXS5Ilc}CK1AkR!wCi2V-Wg^dv5fgc4f|$rNQ^Z7`*+NX@
znMu3Z%6nkxW!uf4=X^!AX&7|xpmEmCGoY-RRm#SnquMoe^K`4Wz6A~+R^qy_vVUH7
zy1H$3M=Skgt?wA~uw0j&E+1Gv()#wkxaTx;!p=R5in2FZhpPf^fesoJ<6uCs4i*&W
z;6U*X9+cn^K#2|!l;n^=$qpHm;!r@T4i%K<&_L-99hBiPK$#BnziR1!29~Rz#=B?#
z|DvLzQtK;(bsmZfa=RDs$PuSp>){cDj#D&hCQdP^**L|b=He8GnvYXFY9USusKq!X
zqL$*6gj$YMGHNAGDX7&rrJ~m2l!jW5Q#xuRP8q1pIAyl#*DpOif7ICx%d=;z!l)Qv
zEsO2*_T|;h*zC_vmvtyBwVu~3n-t{scv?>)F4y<O`U!ahH($zjJN50vT02{ZhMN~3
zsYo8lP7i7`sM1=liM0nWRF&vW*=e7{S7J?{Irqo=R~Em0Iy+rHs(ifl?w#lIe!i<?
zSu8tU-KDz1N?V&uy$f=iORcZV<#f^3i_R{5w@YYMw37ypcGKW7P6j;2&49-`S@2jl
z3m)g>z~kH;c)XJbk9YIn2~Gh#!7YF%Iz{kAw+No(l)#hR5_qyx22Xa&;3-Z8JjJbm
zr#e;eRJRJA=G4H`+!}bgQwL9X>);tq13bfRfM+^Q@JzP}p0AUy)OwS%3#6;Bhovq#
zwyz-C+jl+`JNNC(Su4QWPt0ECcQzAYB?Mig;JpvhpqL;7iVd=$xF83L5AvXdpa4n?
zilC&R1WFFdpp>8jN)4)@w4eq`59*+dpaIGZn*UYVebVWV-q|#5jErx5=+43DJ45cr
z-f;5zv`^N(^ZbhZM@EeuJ!YJ3tYz73qgoHgv>uLYJ$$6~aQuu3(<f(T?Zkk+O<w+8
zj<$vE+81@~P*Xgx+F4vsGoYlbQ|GSTvfG~h4byx7y<Hfsx7@Gd|C=rCQGR#NUX^|N
gy6CFkxjFtmNx4Vt&I-ORrGtiqh=14vgQ=na2OB4k!~g&Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp.4656D94005983107.idx b/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp.4656D94005983107.idx
new file mode 100755
index 0000000000000000000000000000000000000000..39e7f7f4009ab35944d2097113c02780b5313c02
GIT binary patch
literal 6960
zcmZvg3v?4@qQ~=nu{E?jS{~(9EYE~tcFfE;kIW1+Q65=J30o@!dDSR{_sf9t+>^@%
z79}YNq@Y+=9*VfK>LSP<9(w^%Rs_7RZlbax$YLO%tX`o&L~rQXvHx?=?Kz#!f4<*8
zzjpF{XVR~xG7#w7)!`U8v2wx8$h=4IayT5N`R}>&qEkO~Ivf{@9F7GI2SZh}Py0%~
ze(|%1gh>q(5B`x<v`x>dQ=VSC{jU#hes%4jXM67H9XWRUn(;}m;mL1%T$)A3etqKD
zLjT0VLyj36rm{69!^__4zxc+So}HF7#45|u9XCGG@%;<Y{+;<V{%spCA6{|c)|GjO
zo^5E;>BEsbU%%S({H^m}eV_B_OMZTPNNTV8!9O26xT)^xwY^^)b$R0LYhTS;ICpXV
zxN)VK$%isG4-|wByxHgbBk>Un_MET(Vq4Yn(pzWiGk@Dqf9<ht!e?9W+i`PyNWXA=
z$oN-o6yDrhHEqx1N4l-<GV4^Qe-!`mQsv$%Y3cAOHT6r&x_)A<J2kpvZ2ra8TdLh_
zjt-nN;y~`izM4TjA|o7Y^lfkb>CGNPpL5J>Q#UHLd&bRgzgXE=(rx<k!txz_2Y27k
z9!Nd&zURn>A;S(|9r;YbOV2(%X8yFAo#jJIUBe%oc5YC2?aQTaZ|{Bog&9-d*`_b)
zUa_p-+TKg%>}i?0dC%61eK*y#n)<ITSBAY;+uGZ3BG~H4!*l!2{ASDF7Z2Jkx?Y|$
zVok6Axmzzb&LxM`_a4!f8CQEOI<Iy8@x?>?Pb_+(NJtNzGWb884Fg&#<^8VQ_+8I!
z$w4d6wm2OUT5XO_V772?w4J)B|Kc{!RqvbZ`{Ck~FIB$$&ty+^T$|;ecV2bhigm{a
zzW(m2)yKGO@#P&q@!HC8kJZ!O>9gPa{q?uDrY23la^_I2uidukUwkWg#pS^lA{RGb
z?|JN=9j4ah^yp`IpIq?b$;m3+P``ck)aMqrAN%PFA173fA16#6wD8CMfnKk!*g0+Y
zydKpb?<<_$@XEkB{pN4&w{_L$)%?k;f2y52&}hHwsSQI0Hyp{OdKG?h<FiHb?B5qJ
zdavdQWBu-<qnFja9j&UqYR*{xkDB@G=8mcFcVc*_hYLsc{cF*yFCG8v^U_WI)bo4h
zKi*>4DedG_@87(!WS+RSe(1I5-`#in-Lp#%&V2FO{&$V{&z~5(@zbKy+OXtz4^-2I
zwSVt_Y}b-KYtyf0m~WFud;fmD`NV*Ab)Gd}MxUP%thwhxVAmfyy>k7lQ^(u>RQ^KU
z@Ed{7wL69mRsO5(*l~Z8ZcJWO@iMb#Ht*XtW0oAhdZ16;d!BU%_WiVLyO3a?4E9<u
z$~QJsuz%0u>vg5i95}UVa@2UC^;_}_Z;#pj@lR*w9j|@Ow{6BSEq4C3_y^o?0|k-W
z=?<;VPv5k;?v+a`j*ja4{CCsOei<J5(UiYTdFspX-gkEUKGlC*b@1@d%cty}d^>l=
zy!7G5i?MSje^zfDoN%(R?}B^#9DHNv^-kd>Lyg0~n>5$jetylzRk!9G={<A!_iH}d
z8vfZa<>O1NWA4J4k98#uM`6XER}b8xb}yzQMmcKR6}D^9bjYv#rmvz~o6jt(k{>Fj
z59NnD7j`afI@rUdZ#>=QiPy956w~Vba7kfFi>8A;?9YAo&5o}ZzEMm^=7&2LcD%dk
zU=I)eYSOPybo)SQ=?LYA*O7vC#P{zXHZo@9;-VvGRt1ac!+E5lutIJ+*uxF&9gA{@
zJ-=z?sL2nPmzDQ!I@rV6N4xLiq;KvkroYJ#w=ZnpqUm4{-!2L5`eOFcuZ!t}`Qg&S
z(iTkzd)N_;Oq%QO(0q@iqNaOOcO<uY=Q*Q$%UX}y!`Q!JK6yPNB%jA<PK=uYlJPP?
zGCl@K*3AORdRZV@9}6Vs=78k99FUxk1Cn?1K=NK5NZ!W-DYyk71+M_4;1hrp-6D{p
zR|Hb@i9kwj2}sE+0V(+;AZ51<r0kV}lzlRgidz9v@hU(nJ_ShCtpcffRUlQL3Z&-N
zfYiJikeW{eQg`b>>Rug4-KPU-xD6l;uK}dtGk`SRCXlAr1k&`GKw54KNXu&hY56Q5
zlcZgel>M^IYLlc>l61w~ElIlD*<S1UX?|y+!!(b0<}^=bPt@~UNHNbU(j1ux11J+^
z0A*qfqHKf(l#Q~0vN0A>F2VuIML9sZ7>6hy;Q{5NJfM7xM^uOifC^CoP$4EDDn>*=
z#i$6V7!whdA`+ldR033rNr=i38BjSY11iU4M3smFs1j8GRbmRFYD5K8jjDjEF%?lQ
zq5-N!H9)nPhNvFV0o9{Apn6P4)QA{>8c_pKBW56KMod7>s0pYUGZD2S7NAzt0@R9G
zh$e}UB>nAa-7-Jztt6?$T$LmtJKN8&K}Ai^E|~Uy&w%Ec?Fo1)Aq73vkiwn`&57|d
z&|?A&^q3$6JtoXRkM*<AV*@Po*dPl%Hq1hg^K;PS0vz<XAO}4z%t4R$^U&i1JoNY=
z4?RB2Lr?Gv&=Udz^n{=QJs~VWPxOn>69Xdj#GnX0F)Tt)@=MT@0uuD3paeZBEJ089
z%g~bpGW6u23_Uq4Lr?K5&{F~m^pv0iJteF_PxY(NQv)jW)SwDIHLOBU^J~!40vhzR
zpawlHtU*ur>(J8!I`s6Q4m~}rL(lLV&@%!C^o*bZJtJ&D&-9znGXo~{%%BN9Gi*Z7
z@>|fe0v7bFpanfEY(XzcdL~I9dzs!@#-1eUhxu1Y;<dB=S@ZFtrq7tLjI+dZ86Hud
z9P#{!IZizZ+Pr){De6gM&QQ-q%#GA@19RHx`PA9mXR?GrACn^t`j|Ln(8r`GgFYrh
z8T2uYltCYpb~5N=vxG$-n<Fgx*f?d;$EGNYJ~l&H^s$YUMIW1XvgqToghL;fBOLm;
zIOWjCr6`9!E<-u=agCHiAD4D==;O15M<1UfJo@-J<<ZBdD33lqLwWS^jg&_ppLX)-
z6S71=pO7O0`h++Y&?ls*fIcBZ1@sAxR6w7Qb_(bdvqVInm?I+k#5fhvC#I-~J~2Z@
z^ofmBM4y;;is+NFL_(jGBNF<gIF-;RrKp5HDMKanNsUxOpOkh==##TVMxUG`GWz5=
zmC+}ssEj^2LuK^IjZ{XToOa6SQ?f)spOPaA`jj|T(5IxRf<7ff74#{MR6(DTb}Hyo
zvqVLonj<Ru)Hqeqr>3ZiJ~cyC^r?+hMW32>s_4_QL_?pJBO3a&IMvXnrKpBJEkiZ*
zX^m7vpO$uN=+m=AN1vV}I{NfD)zPP?sE$58Lv{4&jZ{aUo_6Z!GqS`$pOGU5`iwX=
z&}XEmfj%Qc4fGj})Igt+b{gn2v&2N7nIk6p%s4gCXQrr$J~Km2^qGy+M4y><n&`8#
z#6q8yBNqCsIJMAcrKp8ID?=^xS&h^}pOto6&Ata#UUA&~VczHDuG)>{-i>6u{k*h)
zlA?cax!6z4pt^2$8JBp5w5;vy>pal@uW-chww{Y?`xoRduO}_mlQ#AX*iSxM)U-8Y
zzf8;j<9o7d&<XZor8?g)A6!1l-nm)ceY!Pq_ui!?`TY}@r^0@*mbHwFfyBC4NSup<
z#JhM%f=hrTx<p8lOM)c3WJrojfuy=rNSaH7q`P!ThRc9tx=cux%laj){MX=e?ZZTu
z{NG<vQc`CB*J0m>(xRrFtLD4sNf6H__?W^Z2!okT5Ee6+ARJ~sL3qqUf(V$!1Q9Vy
z2_j*Z6GX<WB#44pO%N5cmLM8tJwbHLMuHfa%>*$qTM1${=hrSTTR8e$!|MFqsxT`C
z*wbR?g8c>6Gq(ovmpiuZSZ06b`-@aFv0(7l%lXUg`n2QhJ<TH{tV@qnq>tq<ZzOkZ
zBt7jZo7#BjV!yUVU49@*TuIW_9yojcKki>!`sV5UK>6tM3HF<FfiL*s?lv9c`OE7`
zhxMew=Isr}{35!uyJ)|Amt8y8Q2ot2OQ?OH_PkBm`;Gbc`p#yfP3LER$TQmh6Z11a
z#6V(0EF><(LE=L^Bq1a~5<?;+DI`IXLoy^Kq(D+bDkLqWLDEAyBqL-%GD9XLD`fqW
z^82VW9=)@vIyX9D?n8G@#@snHeK~7C8u9wF|GwDb(ZWYYj~g>~yknd_<Zz5`J{{Y9
zI==bzk>=A0GbhfNl8;rBg3h}N3h#EcC~n!Rq)qF}(!m4WrA3tk+IDQ$zEkJ?HeP?z
z^51`N2UhQ;^sV^+u9kHz@7BFX&)$7JOut`wdEq`;y+`W8iGw<n{d#Da28KT{gbe#%
D95*(z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp.92DF213C04CE83E0.idx b/.cache/clangd/index/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp.92DF213C04CE83E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..117769d3b0c447dc0771d17081281f4e5b6feb01
GIT binary patch
literal 6924
zcmZvgdvp_ZqQ^78&>C7EZGjfBSSXM%ogFhX&LcC!Oq7N|De0{hg1l-J!uw^40>YkL
zF0d#`L0Afib$Kb`mGyxj_wd+@h_bSPXID27Srue45KvaHP#~f=<lM2}KW@(H@67l2
z{eI6Uok#M^uc5~8?{KfpRz109;q37I$2-_;w&Ltx)BMQvkL@<w#R8jcVe8>w?cB4T
zqHngGd4!+RI{EM)sfv>1oM!pCwLAa(@YdJX`ZiVW8xTHz=equB|GM?x^}9TW8vo77
z<BPnL^N-kOZJ5q96xEf!GkEdMd6hkvv_@-6lVuwpEBokTWN<I;OYe@2SKeE3@%GjE
zM>e%~?fLQO-EUm0ym0%%*FR)j+LB-19iG^){N<ld9Nygg+}Z(K#$1_v=la)k7Bwwy
znJ}R^HT6j9)}g%Mp|=M9a5Oe*;l2wkU+kz|UVQspOX_bMTCP8_gFmzVfnB$D2DOVP
zhEII;X8x`HwKMiTd9=^!-g8d({72y*FW2m^6_?gcYiL<odhe&^y3=FJq6;o{*;emd
zbF6yas6&~P2O50+!lP_!v>orf{Z_vbFWBaHZ61@@JL}eWU#x5^>N9hBe)%r0hjY-G
zhZ4_!<T|=x_{jIJjeb7w<xS6xTQH+xcln57N8Q6S&ineRUoCxi=YR(<&YJ%I4sA)_
zie*)62P~PluT$dIecLY$+T752`oFea9r<Bn7kBH)K<A^6HVvBn?Y6%!_U#oMugn{@
zrvLxkqZR7s*AH(QFsg)}(0DvDze~%B#Uln!E_kYdPmY*2^gryaLpsUjRabBRu5!nE
z-^z0x&j$I<TO*U`ZS0#R)1Mi<xa$k`2c~*{y!7<THLv`0eWfy?>+;Wgt$JX^x)arJ
ze6VWuaW-9eW!KM~x-!&n^^Eri9(4b3<DKn^DKoErd8E<PZO6<nzT@4(^1w^sOIvSL
z9=~swq4qvIcGKQd3%8t_s*nvWJ6BJCVR85IpRMq)e9goO{8ZnfpAP!_zqVrcjJ@;w
z)qipze{Som)$^(rY_HnB>hpT;)U~%8r&sIU_dL5{_|Vp)nMD8mPj8-iMw<JF!e>5g
zcuN2M-eY5zHNP9Ft-of>TK<oQ1?!r|wN#y~>-lK@=s|xjc<tpAXFf09T%}ytx8TW+
zBTuWRp8e?7%_Z}N?JXm&zxcs{J0F}|dU*Dh>jyv3Ke}*o{Kn4;&Z;BVfB#TDS=9LV
z!N>P3*|#?NT8jQ|{jmXm_`UJekaf+jHD5(u+%&u4zKi}n|Jn1^8(*J3QSx*7OU-pR
z{k<A@jTj;Scggq(e-m#`eWv0Sdf!~mvuD;EDR%A9z~&EK>kb|GdCyKh&O9CHzi^Ca
zd@AqYzQs41i=RJqdezj3{!*8Bq?g_uxAT*qznp)f@paFRStHfxh1X+$VSnq-3*Slh
z=zL-3=B>@IUS4r*%%B&)pLy=9(CEKT`_r^%zY6Vtf3N2=?Wa|T-}`0xwEa`>WUd;Q
zKi+sLdj8Ze%I(9GPUR0;c>ln|Z|=U)Gqhxc{@(AVG?_autofw&_PnD5X4n0&=C9jB
zzu2aIeVMT}Et>sA8D+EOSNzxN>TOEjLNaQMt+88vw~o0X`{d_d1-G|;xvW-tq>wz4
zwR`3FD$Wg-UHs;Ay`Op`{eB^-&)P-#MICd)ulBJnea^kjjPGn~$lB$l<pXkqWqUK<
zf4l3OMQ;|8(OJ7JzwDmeVA*dJ_DLVva-?e^c{FQR<X1?!!Lrkj_dURh-#$=C{w-^F
z&+pzbH(2)7+qKK8@~LDeTQF-6$s6Ly4VHbUD7fc~xyQaKBoAlp;{4){xxun+k?@qJ
ziXQFfr+-22yfh!pY~6ic@7t--<#f^3JIpGNOK(@SlLn9W(BLsn20X^YfX6yn@K_HE
z9_Qr1<2)RAypsoy_we8eP60f@BY-D5MeszA2%hAWz>_=@c(PLlPxi>*DNY4E#iM|y
zI#uvgj|!gV)WFj`8hE->2T%9t;2BN>Ji}vvXF5&rOpghkubZ#**L$SV*T>h_Qr0~$
z7Ub@Axa48iy!MOiinx9YD(YHAwM#loLzIrt5T&CuMwu`JQ6|Dbl!-DJWy35)*$4|!
zHp*g@3v&?VA{<1yD2Gu#%tMrq@DSyrJVu4E08t?#Kvakd7!|`JM8${*Q86lFR0>NF
zl_C;ErKp5aIV?j|j>r&|qcTR7umVvfqCixMDi~G6Dn!+Y3Q;wxVpI!j5Y-|YM75}f
zQ9Z0fRFCKo)uTE_jj#bxBVs_*h#D9*!zM({hzU_MYGTyK`vzM#b*Dz3>8k-%>kC@S
z`V{vS<UYOOhWEOLwBOS%zpDmRz*P?_<eJp3XfF*t+D}7|4$#n}Lp1alF9SWs&p?j}
zFwkQ{4D?tp3q97)LXQow&|^a^^f)gEJ<iWTj|*_n<3b$tcrOn<-p@mi5Ae|ALp<~Z
zuK+#4FF;QS2+$Kk0`x?$2tCm+LQf2c&=W%<^dzqYJ;^UYPYOuTlR^^oWUmZ8*)Ky+
z4#?1xLo)OfuL3>AuRu=;D9}?v3iMR33O&`YLQf5-&{IPy^fa#qJ<YE{PYY<!(?S~b
zbgvFQ-LFGW59rX-Lpt;fuK_*7Z$Qro7|=692J}p?2|d$qLeC7C&@)3O^n8`Rf!0Hs
zeSlrQDp0@ixh+-vf9tH`g53HF9^g3T+KgXIaf(LG#3=?f8>d**T%6)i^KpttEyO7S
zwHT*F)KZ+1P|I;jMy<pt1+^NdRMc9W(opMhN=I$PDFd|`r_6S3tuLP}$gL~kw4^E5
z6*w)f4CVR>b&R;;r2X^El_0Jp>J)KZLfuAOH&G|;uFvf4eL79i$fGk9jXXL=Xynle
zLL-k(5gK`P8=;X$C+#%ym^8&8kI7IB@|YN5kjEqlgFGfh800Z+gh3vYv@^(K(-ez5
zHbb$<V`GFx9-AO6^4Jt%k;k?X7I|#a&LWRXQylWR48<XjixCcaT!L`O<5Gk}9@j=V
z<Z($mhde$_@yO#d6puVUMtJ1$3Bn_fPZ1t@d>i4B$0zMP@`N-cAWz6p0`i0y5s)V&
zh=4pHMFivtZA3txkhBZP6VsH4JTXIw$P;5kM4p%+BJ#u(5s@dh5fOP}(k>!TN>dW@
zqzok?Pl^!<c~XK%$dgh;LY~w{B;-j+yM#PBP07fUGn9-xIYwmU$q6DOPfigTd2$<(
zktZkZGV+u(r65nqPzv&t7*UX?B#44MB}EkEDQ!eSo|3dH$Wzmliaa$#smN1fL`9yO
zAS&|I6j70<wh<M1YSOMEPfJr8^0W-4Ay11D4S8CEXvou2L_?m|Ml|GUNxOzTJx%Gz
z(=(KgJUvEq<mm~bBTr8e9eH{i(UGSo?K<*|G-V*q$WR9Ij2JPHXC#P$JR?O6<QZ+m
zK%SAb8^|-$l!-huLz&1kW5h(BnII<e%oH(^XSNX&d1lgXw(}lZdDV97$N8UA9U6z-
zKWw74hzKa_=9aSY=c!JOy*$0DtzW^T>PlRf)(+0gPFHuV?rNo<uJxT{9+m5})8#|U
z$5<<n<$Y(GllSglT9n-qakwhrE6_oMVjK)8*1>|}92_X#!GjVU0w~cTf|48(DA^%{
zQXC2>)uDpY92zLyp@T9U1}M{E{;F2~YiPOpalCi-|1T;kDz$z>y?$lcqOs>&S7)ay
zjEW)F|G#_T!Mys}+x^+;vMyz%*22WHNkMLBt9^lj*X2)Q!{of7+plE1-3E4Jtqr@Q
zqs+^XRV0sRr-yYIR%u<A#Kt3+s!H_c?6l9}E3u}}UHJD0*A~BZHalHDwtSLxxh`^f
zKi<=|ES8<F?onM~rLC>Cl7ifpTKoHQxm|+wraLv#ZW<KhW<arS78K{^K=E!Kl;9RX
ziEa^;<d#6mZW)x~RzRt46_n=IK<REBl;Ji&nQjx5ucz-`>m*!p>|R5(x9&a<-3NB(
ztaD=7R6%Y((K<KQeqweBes?p`n$A83LDyI~WkDJg6J$WKK^7Dj<UsL39+VIiK#4&S
zloXUe$w3*E5>!B`K^2r1)IjM$9h4C?K$$`FSC!o-o%Q(LP1B~xq^3vi9*ny?<X-m1
z2|X@W?wmAK&3|m{gmL31+9p_*%{I3EaD4mW#P-9-+7Bnqo;+(>R@P1l*gNFq-{a_5
z*r{_-*Df{1L#v&|1vNuT%DQ#$*(<x9IM^_~58mH{(fZ4SD*nIO(tFGM^zBzUV4#bx
d`i-08ACQ#$#NMpn>rpyvM2Pt79vV)K{9oTWeLMgF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl.hpp.8498EA503EEB08EE.idx b/.cache/clangd/index/device_gemm_xdl.hpp.8498EA503EEB08EE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f6ffd0d2bcfb02fd0bb4484ee937a83a308bc92
GIT binary patch
literal 5390
zcmai2c|4Te7oU5Z2IIllXUsE}tWk(ajHQH;qP<reiZ-b?s<-r$q@+gl^V(^%mYNo&
zQfcp%N=XSxp~#EUrUljS4sOr$`}BwTaPR%j_dMsGd+s^s-e+cby1TQ~84TAz&y68L
z>jIbz216PBhph`;v40qYVb5VOHfEL130aeAuhv@dE}%DM4<o|&;n*Xd`KBwMx{aIN
zby=?_i)F^ruMZ3<`2Hn5bs8U5ynSQpHr?*Q$E+nfI?gwH`ra1sZ<i4>lD0-TxvDfA
zF7VUuZSd*hrE$0ZB&H}dioAB50)pU5=vcHMrAbv1qv2F-y1dC}Ug8Ng6;X@Qh{pLW
z((`&%;=262PQ#KVl8MfD70h)#=kWZunQSsu-mF<@vS{nL2`*K#+7nm!nB2%z-&AP(
zriZn$n7K&B^xbX7iK*{u85U(m;fu3Zt2@4#)$25WR%L;=gk^6%`-OgCW#q%N+S^VY
zS#_i5Q<D6pCKL59SJy07G|8x1AZnBs)mUjXwb)%^ceI2*pHLNcJUjU7Kl+Wy_q!fR
z?B*<ynYcXgbh+TiX#euBaR%zW8`Da!$BsDu*{|HRqI~O8qd>1moj=B&UlRAE&&}X|
zPFqLxtG2>z#nTUvk83)NKGsdB^7i|__P}Z%zla*Lx9M`vbIZ+sei4x_k6z6EdAPkJ
zx1pwkZR@jfUh$;)yYD=k*IT|f`{G_dzhe&;x(ezRT2wE7<T$2cLih5Zo#B_g1E22d
zT<_&rCjWJjn5-MtYFG5-X*O>)@sRNAQgY7pH+#F`wCP_Rt({Dp)4`|KH#&2l1&sL4
z{%m^h?lSY!FYYgDEMB@Hw)9Kko=)wLRm<A_>+PpJek9R5x>BJjsv*35e@v}wL;JkM
z^cA&ntdfQz$)@{%8}8h^*Y5b%;Y;e+Wt<kZ^9{n;jfsISp6qj1-PRTqzqWWX-tYAe
z^*^j@Rft;=b5n0!VEEcqPYz8yE|K@nzHY{ToRJ_Y39fJq-oLtzb>2qvsPHZ~m19%a
z+B_#+G&ac7efbJQubb813Zu`NFD(Ay(EXAfa&UT^l9&4oU)2q>&%AK>dO%pbTVb+j
zhdu9XSdnh*oeJhP`IDz!I|tQ>rx$IWsP`vnyj0Hal#8-TU`Jr5%sWnAtP^_#d2HS_
z=7k)N>1_66y^;Bwn2OI$^7aOp%Nc%nc%*#?Z&_WL-Q7hF5yN+?wTz7SJL1xCMwZ!X
zVD|Kh;+P*XdDppz?XT2GYNu>Jeg8wx-VMu&=eWLSRqwVsWcj>u&9r0#UiCOFr%9g_
z?mXd+{@K@85gy#Kx@B&_+6qNqqwA8g?{f<;U5wbBbaIx0$?kc>?pXc%uGV)+--N<1
zV#c>?-Q@kWb)nY|gq40VYJX`GcQ{zGh+qCwq}NVd0mkh2VPTs>7SDnV2Cqh<sA_nz
zRRzkcG02mEa}wweb_`jR&FDkk`M$u_B&3Q3*+?WpK0>9x!L;%G@invJV*}Np0;0XN
zyhijlm|m5-COX((Y{-Es>I^amDCPj=!48Y89v<u#9Qn_BPQMM6K&}#y<Uof-Hsvm5
z+1@_ut`23;BpzBGWPgL{?JHmTy}G_YM;*!`+EdGu>~AogIYlls=EkBh7G$b2NJkk*
zwf=@72LW0iAVHQ;qNH~?t(*gu(DQ9zL>sUf?63~WM+GHAWw9#<4o4z8A!jEP40c$g
zsN;lk59`7m4pc@WQBW@mS`T(u<ZMH(b)xa|hyiazDn=@k(gxF48@nAR%ka%qAb~U#
z0<A)1$3TZgt_A3<?YJ#pq5?Tcqy=zV0DrK<BF!@_j*bcFNRo%ckVrT?T&2IkbeC~V
zmEyCPx$00B(IPDo+28!4Zwj3~%8uWbhb%;Uv%NL?8%%$fnLp?2hT!umP!ZA9fK!b&
zYoNm-p+QT-xQYzMu+Ys8<8QP&LxL;@R>fdE4mN@;1y-dLmVZ~l%U-PMh6Ke>K@2(A
z*#sF672+wZU>TdSGI>lrlOR2GJPdJE5Tr;)WQf;Ake)i8hBzb$(#z1x3<s&gGOckd
zj*?d)lgxue9#q6LG06f*6i}F3b4s?a;G7i_%m<qJfQ#dpNmc><Dqw(v7Ox4g2h#=z
z8^71dEIhPB0}x~a^i6>C@l-gUnC$*M!tw;#MdKD=jyISfqoHmzG{#XS#UrDkCB|g)
zx7Jl)x);)a0jOO7me>*<#i;SSWbZX0A{Ovtfe8*ey$hksJGTmB022`!v4Dq(5zc2y
zKG;>Vhg>XnBU-F9(Xr_xzxwZfy7ov5n(-bG+ykcA6my?I538epiP4N@K(h?+FkvNn
zwVTEBwe2TV%77{+Y<z7rtvzP+AyrX85Cx2}jrQ8_?*y{~-m5XmL`Wo3JH;gbhQ!}=
zoCJAEkdNKXB$FX8nT}H+FNKa%p;js+Dfu+0k_Oc%oDMb9A&<ftP&0#$Ga)|{3Mf7c
z^0Vl8AGF*Dttfs!wB8SGD0~1~9-#AwAaRJko@_{D({T>u=g{@%LL!&W=fP2V(45kr
z4~cw=7YaYg7pZ>tK^`swN<}~$yMC(AIL+H`Y10ugQ-7vW*I*r(Dd$!i62yrgMxb&h
zw-YL1#nNOo6hku#$3x?II2z+Q0m;6{4%OQueT7i95Ncpz-m5hCK<<C^QMxArgGAtf
z_n08lfk8Tj=jX^(WK8ttBMsg>?@8FAh3CevT3}?Lk1Wao+#FzlbCW=)hJGcmz`04F
z8m`s`ba8q2Z;oBYdZGIql@Fn%&>4rXw8|*QLtSj26pP~FTnZ;ZQ38b*TX|ko4Y+(B
znH2}opFS4EcA)eaJm?0GPHYe&2F^7+^;#w){fK36C0dFYDvO~uRwY%f5C^sB*eVV>
zV~k55(kML*vA%7v)~VOtNS(<<&l5-mmY3o&ar7vQhaPB|6rTV+&@wT`<rC?b9;aCU
z_MFJ@sfTZs$q?ijz&`^FaPUa6@C<OHa53N))A>t)f9e0R@Dg3V6!1&wd?|1$1s+(R
zc=E9AZxVjm5d>Ka1hv2bpF;#$2LyF={1gbD(s4Zy)YI`ZAb3W{4M5O9#~r}7132QC
zA;?Z(+eyc7fbAPPehX~h(s38C?V{s%!1f)5#dV}DBjMa)6g=Uh!pGvXTHHSA$?~P4
zeZLjNxDO@Qx?FWyek-ASUS|1>j`R&eStOhwbj3R)4PM`P>a<DZGrpa7IX894WJ_)`
zz3Or~w;dU_3vzcs0~`v;U;GztQzENT)<l{|+Ed9IIVJKB3grTBE<F`<fo(3Bh*Kh^
zTvYHVX3HpKMmJFF20D~~qoG<f)WOFZUIz+1=|#=cV1xL-7hJbsarE`T^$g49y4DLO
zUhl~qQt`2k7l%&^$#zHU3zH4w;9?@Nwx4!upMNTPK7x!e#)mUHexO<uWfpb@nt$N#
zhv&z3BVypjhlw+h7B6-Aq1I^Qihy1b(8r-%+!0=^x%)%mkf-fg+oH714?9Ij!!D@0
ziz=Yl>1fXWE9796xpceQSK{!oWk|*Nnu(qtxLjhJ@!bOv12@ZG#8=sF6RR<2k?K6C
zn+Nr<Kcx32qXKA*@r}dNRx7y6ZXa?xZhl@m`$Xu~gJ?S(Mmc!kiOT-X*t)OH_Msw^
zOoBubb!<tGaUz+HQy`H-$7l;v={OA%X>^<piF7*7fJ6ozqY%oZ<19#I(eXY=?4#rT
zkl0Vh2Ox2Pjt@fOARXsGB8S2cghv0F;X|ejxmizAueO=0IP5`=a@2M-!H%lkQNv}e
z*Lw0>U8qYs^zvlgp)E+U7>*W0C%iRGR2eMdp*zM+b`M;?9OHr^Ex<KOhTZ=jfh|Dg
zfE%fcW064jWV3u=jT4?g*J0C>zzRn_f$q)Ys(>>-3JJ0qC^Z93s*ZO9-ELrl6a5u8
zZ`9saC394Dl7U$=u%ISN1}@295{0vXSr(no0xnq;-(orZlZ=w47)4|okZS{~cm`>T
z%SAy|oWZY0l}M}wN6~GIOalBQFov2T6Yw)Bj9)ZRmPp^8us?7`k!AP4FyXIJo!1UG
z#Xcm+7{G}EB&9nJjEn;|6y6Jz_5wbI6M--hh$x&6gz0pA4=CLOeC*~AStVC2PU;y9
zxv3vH|8_?2@)_G7KlB^;=ZJr{X!KrF`i<!A&A(#tEaejl#F?@)_3(1>1|n|^yawWh
z|0u0*o@M)H4w8?96XM`x9C)EYs{*Xii}$2s^37^pk#hl#8C;8h{p5gSpMiBU!a|pX
zB}jjj;7Xf;4-XF*3@vKNr-lM*s7Vd=sG-KMAtW_=9`E~em3W%GuAUB;C(z<+YiKgG
Ne+e*jsi6)v{2#15S84zN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp.F73590DBD2CB1EF5.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp.F73590DBD2CB1EF5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f2b198cd46fb8fb21720ed060f296f4f46640b98
GIT binary patch
literal 16360
zcmZvjdAJkhp~lbm<>MF@l|?~81+9~PCJB;EmdPXmvP?2rCBTs-g2Clh9~(dvm(&`)
z##W+@*4Bz>!*SoL@wn8BTNkz0y2Yi{YD*dnsMK;zZE))i=RWbge_Wo2kl(!T%Z~$L
zvYxCG4*%6ZG_<6kTy)-s&Fk3E(9n$j&rO>*p7V}oXvjZxXy~HtyOys?&2t~M!{11~
zb<O1TW7}>#OgkD{wV=LCI<0!ynZNtt3p?K3apflu{pD{@PS?JF_2R2;R4;v{v*{Ui
z;pNkw-ScE|@els-$oAO>f8Eji__>3h%r6x$yZH0{SKRi87xb50n|gKc!S$a$e`#^^
z6EC>jADkhrC~m#%r^me0nRAx^>DP9@@UN*)Ke6lP&zis8vss<?$}?+)UhBPm?|=2-
z=Rdq{?TpTfXW#kL{Tg9ppQ*56dW`KlbL_p_JI_yP9PY90yZ7Bm4=;Lg#a@2yW9L4a
zQCFQa_o|1k|NgGg*W=c&w%oP*LqqwJb<=$Nwp`_WWAX!UxQ}^w>8zRic73+*!fmB_
z@$-JPR9c$)?NhJb`q0vuZ!C{!&)(=+by&Rq=(fPWUK`$j?h7}~moNQ>z4Y+%v#--Q
zOzU>u<ZD}QzVxI|DoZD=+qu&cTqn$J-j`W<$&Dw?ygxm6cj)9NUn@WOi<1NU*X(KS
z^jx}N{>}%lKY4L?+qJJ;P3n6MKaYLdoWB08Thv|0<y|9x9KP~`Evuf?1@5}=>SNBE
zVcLDq3EmH%*?au`SG}|N9Odqlv%R{qXYcX%Zr0v-(Tk3`lZ@}2c1+@mOI%kSv*yol
z|M4?E6gDbHJU8pH`%ax3nDcgKv3D(h<GZ&mxqS86r`V+7?swl9J=%T8ONaZ19zJkW
zYgsz8^~JGeuOE0~$&!82Tl16E6Ft$L?{0kQ??3vlukXEk+nSNbAO7Wx5ANLSdHgU-
z<+yc!YP&B!`wGMNhJGHX-nnJ=@(Yfg|H#2l!oRuK`KJB&P385IKK|LA9Cy!k<~`Tm
zeYt<$qqWms+_yYFpE@2m|G?W@Jxd>&YW-;8J!e}cHNW}$M|Yg|?#}BsJhIgE#!(+!
zc*2Ijw%e@2S^Gb`V#htpzJBFo{+zc@KWF>u8H+ADGJfKUH*alkfBh%bW_jA`;R`G;
zO}S_E_fP(A*1eD3_K2?cC`(q4{X>QSWACEPw)#(E|GeeSUwwV+b4xz|)n(34Zu%=1
zKjl;DLdTX(FOD6aNz;*xc>lBaesSJ+p1Es>arJkvk#0R$Uby;y(%Jj2e|kys!l~_p
zz4n1s4?MBvz<=Cz_Q@Z6&Uo|um%Ss;z3}?$doSn5Hr(;x9WS4`V*OQ{=3e)YmtWht
zXwMg^i)<VJ^zzM?T}O8oN%`noTmAE={`|G4-gxRxLu1wBFFk$MqNl&w=3Js%al=cc
zE&3ah`wI*Ie2e;sxa#-DOJ@CU?ZLmFu`zU8M#)|LRQ>DA-xEHf=iYM7)pIv^pW(kc
zV!7iFpa1vpN2h#s#OIg1TzTNTtyQ1w|9W?H%VBT-`^W<~Bum$PGx_`vidRqD@z{N%
zg{yAt>Ne|6zw*F|{G2)O-TKwfZhPj4SNHx}d&8qQytMmTXY%v;yN0HHI<#r)dF$tp
zp`qdVPgIxuLR>JFE?Y6QcGmE$$>SIOGr;(dDPP?FK{@U?bt*lzA0In>?2Pe?K|FQN
zSIw2a<-eFVw7eg$n6yF}zZk?b9=c}k1rL7o>!~!|kEah$pFDmsi2Yw4*njtjTOXQA
zbNzVE@SMZOF9z`i_pQ%;wtq%$DqY!+PZ&NSJbp2V|MSBy7w_2Pzu@qpY(Jhid)|rT
z7lZhtbqjvQFWP^~RGR9?M-3k}dHiA!e?5KqBk!O8n-8bbOh2A6JY(|s#ULKqxZ%uA
zUH8P_PiD&a-^;e=zr6ka-QES$)^Z$Y82kzQByye!(s4S-bfOM2J*S6EFX|yPa0bW>
zq5(1^XN1fs8X+@rCdf>p2{JQhhRiIQA+vB6$Sk4-GM?ih<3%1affFDTL;*4@XNAlv
zS|PJ>HppzE4Kh1ths-Y8A#-pJ$Q+^rGAHMR%qcn{6FCtwQ4}F_aW2SQq6;!N=Z4HJ
zx*_v$9>_eR2Qn|`h0H5@A(J=>GD(ym^Km}Ne4-CBnUf)tMHw<b=ZDNM`XLK&0muSk
z0J0z#ge)iqAq#V1$iiY6vIrM}EFwlAi*ixOqGA*>g;O9?L<O=K7lSM&#vqGxameCg
z95R(tAyY*avILiaEFmT!OL9rbl426F6qkZ5C8i)tb7{!ZVj8jxmw_xJW+2OQS;(?t
z7P669BeTCfg{~breq_Ob2It+aQ^wD?aM~$w=TDqyxFUBcP>H*SOrWlygQ~8mgQ~8i
z!>YcZhpN7)hpN7$$Eu-VfU2QrfU2Qnz^buegsQP<gsQP*#Hy)af~u)#f~u)x!m7Do
zhN`(}hN`(_#;T=YfvTlwfvTls!75+ip~@F|sPZKqt3pA5s!$Z5DwG7QS_@XFT8mbw
zT1!@}+6p$P+KM))+DbO8+6#85+KYCm+Dmq<ItmV`I*JacI!X?#ItxyyI*U%II!jKh
ziUkp>Vo`*uSQ4@7D!8EPD!QQRD!H)gF1VrUF1n%WF1fMlDR`jjDSDvlDS5E!EqI~o
zEqbBqEqSpj6(p!iMG2}>Ny4hH;Df5K=!2@S<io05kfACUWvI#}8LR$+AFBSMAFBS6
zAFF{v0IGpv0IGpf0IR`55URmq5URma5Ub%r7^>l77^>k?7^{&&1geo@1geoz1gp_P
z6spl;6splu6st->fvQqepsJJ<ti}p4sK$yhsK!b$ti}s*sK$$NsK!fitf~bSs%lY%
zs#;R9nkXconkXiqnkXf(nk*!tnk*)vnk*%;nkuBAnkuHCnkuERnl7ZFnl7fHnl7cW
znki(Unki<Wnki+lnk{6Znk{Cbnk{9qI$|G@2KQLg){e*{3Q&Ax`G5w$k}2a)4+i&U
zgQo|#_m3ZXaxggFe>g33!HGLEF2qHFR4xUS=hjb<PS(Lr7t+B_r|Mv*%j;mLm-VpI
zhxD-1t9sb!^Lp4BWCQFBAp`6TssVO}ya9Gb*$6vh$Ot>5YJ{CJZ-kvmHo?vmGQrNI
znqX(jn_y>_&9F0v%&;@7X4skYX4qL|3+yZ*3+yba1$LIa1$Mm5!;TN}u;W!8c6^?P
zogfRa6G8&)1XX~YkQZQQm94O|hODr&s#e%p^H$i|WE<>kAsg&$sttCwybX4C*$z8<
z$PPQZYKNUYZ-<>jcEHXNa=^}^I$-C>J7DLOov?F;oUn7MPS`o~PS}aE2s<$(!cJ60
z*ok=&b}rckJ6FgBJD2K$oh$Exom+Op&K+{Y&aJv(=gzxf=aD_I^MpLG^Qa!!dGa3E
zd1Wu`ydf{_ys8&=-n<ufk}Sba3Q4e&R0(!cUV@!Z_QB2<^1;rh`e5hF`(P)_GVJ7#
z3_DqsVJGKh*!g8Y?EE1=?EI=9cK*B{b^$p6yFe%ayMP*iT_7KTT~H3fE*J{JE~o}!
z7t9A?7nZ}Y3x~q63#(z+h4W$9MdS$VBB2QEB5DM7k$ePpQ8@~`XebK1s2YV`G#`bX
zA}g>{LJI5@Re_z7S6~;DW3Y>bVz7&;G1$fOG1$fBIPBt~IPBtT9Cq=19CoU#!cGmT
zuv1kPc4}UQT|!R4E)hz=E}<r1m&hkzmz0yRONNrLOR7oOCG$zxrQ{UsQlS*=Qfdlz
zseB4{X*mtMbSMqGw3>!pI-iDJM$W)46Ux9Yqh?^2$!B1fm9wzRhO)5Bs#(}&^I6!9
zoHW7>9=!A);OIw8K$a2ffaZNEn=jt@@$AE=tqnI=c#KUvB#AUw1UcGZQDmjT6y#Wg
z#gO9-7DrYaOhryKSOPiOU`gaugQbwu4VFgEG*|{X+hExV9=t?bqf^FTn}vsH9m4G;
z6JICSBiupc3gxb$kT)s!CUTo{dyu=7`zLZs!~IP&vDbBo4t=^F(V<UQp*r;GnpB5A
zU7PCAr|VK3`gARg4t@F#(W6h_BYO1dD^!m@eUs|Zr*Bg|`t)6@N1wi>(WB4MAqMmr
zdc=S}Lxmd9XJ}Fb`V4JqK%b#Y4d^qpGzRn;JH&`SV~-foXRJ^o`ixC#M4z!ujp#FW
zsS$m~md1!aQ-_$)XX+6X`b-sSLZ7KgP3SYVsR@0iE;XUg)Y6#HXYLR)`pi9IMxVJt
z&FC{XsTqCdHZ`Nq+@)spnOhn&`Yat{L7$~ZEa<aTs0DqNCbgi?(xw*lS-RAMK1)ku
zK_A~CJo@+^;nBxeD33nANqO|~ZOWsM?@}Ipd`rWlPv{T<eL{~2=o2bbK%da00{VnD
z70@SisenGAr4i6)?GP*atUY2ypS41*=(9Gd6@At=wW813rB?J=TN*3+Y#m}lpRGr1
z=(AO*4SlvIwV}_}rZ)81y3~d~TT5d@pS?ru=(G2T9ewr+wWH78q;~Y#+tiLedzaeL
zXK!ij=yP<41AUGjaiGspp$_ypn$&?lN1Hm(=jc)g`W!8d1AWd8aiY)JBTn==E7XZT
zXOlY7=WJ6a`kY<rM4z*zaiUM`5D|T1kBI0KD^x_E*rX!*#5NVtCw8fbKCz_{(dX(A
z7y4X1;zFORLS5){HK_}It~Pa{&()<a^toCZ7y8^C;zpmlN8IRhSEw6(?k07k&)ueO
z^trp#jXrlv<3^vSLp<p7^oR$2o(lD#&(owH^m*FUgFa7}deG--X*}rjc8C{!-X8Iy
z&s(8h^m&`qi#~6gdeP_YQZM?vEsYm_Qin+BlX^r#pH!g|`lKe6&?mL2gg&WDCG<%x
zjf6g5hxpLv>k%LNd==_LpRY-M=<~Iy4}HEa^`Xz#()iFPcZiHWxkqI5$rUQ2Pi|5f
zeR7-1=##rtMxWf$$msKTh#!6a9`U2kU!i{V`J2>_K7X6~(dX|{Kl=PFjURo14hf(y
z&?5o#1u8UvzCe=(&=+Xa0Qv%58bDv5r3s)f*damm1$!ijzF>t0(HCseAo_xB8bn{P
zOM~bOwlqQXg*zmSzHpC((HE}JF#5tx8b)8ZO~dF5cWD@X;g%+hzDS2e&=={E2>K!w
z8bM#ANh9csv}pu=kuHs(FVfOP&=>8HDEgv35=CFMLZj%5Hfa=n(Kd~uFWRM1^hH~m
zDEgERQP8LKh=M+)LKXBWO{$<zX;TG#N|!3=Q(77YeX$OSp)b}WG4#bMG={!dlg7{&
zYttC|VqF?TU#z8xp)cMcarDJ|B#yp#g~rhrZ_+sW;%yp7U%X4>=!>^BarCJjqM}dj
z5fy!Eg{tUNn^Z-g+NLV{)Gk%gr?xaI`Vt+IKwqLq66i}*XaaqSCQYC((WVLXCAu_$
zzC=rtKwq*$lITnJND_U?3QeLf*`!JICEGNKzGRms(U)v#lITlyND6(a9!a4uRiP>L
zrJ6K_zEqp0(3k4c6#7yvO$vSK4oRag-6LuAr7JXzzI2nO(U)%1H2TtAnnqu`rAebN
z(;*r3WqKrozD$K?(3fe_4Ei!{nn7QtOEc)pv@{v?WjiE`zHE<V(U+~zEc&udnnho>
zO|$6Bc4-!U*_I|d;R{~<$<V>>=64BOyJps!QwOgLEy;et{p7cw{hLf%yM!$<4nAF$
z5Nqep=C8fJfB0C}u}cOY_E?sjso(o!X;T02%E>E_9Q-Dyw<>21|7F1k`u_W3b;dmm
zFD~ioOoyzmGd;4Q&J4)LIx`}h>db^}t}`>TrOqtKe4X*gLY)c7);hBy+v?1QY_Bsr
zvSXAv8u0ZvN0}2@9AzT1Ym~W=-J{Hn>=|VqWbY{RB1@x8LiUX^AF@2kWMuy+^CJgF
zSpYdW%7Vz@It%017pb!da<tB($V#0l$gw($A;;@1j;z+1ikzsk1ah*@lE|q#OChK0
zERCG0vkY>!&axA{exdTztaVR6F}Huit1<3o_=nQfm=0NAV|rvmjTw-QHD*LM)tCv{
zTw`WrOO08O`5NPqg&Gr(tu<yvw$+#o*<NFI<Zz9Jks~!0L5|i~6j`Y;1vyq@G30oS
z#gWw-Q;`!jmOxI{SQ0r^V=3fxjir$@HI_ln)>w9e3+C_N{KfkR-#B!J&~o$PMN_Q~
zNUzlc8MFo<qt*yy(wczGS~HMEYXRc5JdmIjfUH_8kWFg?vTN-?4y^;osdWO0S`o;l
zbpg4xZXl1=1LW0ufh4U2<kR|qWUUP3*ZP41+5k{c8w3h#!$1*j1SqPF0x4PrD5i}8
z#kFxDRjUFev<aZ3HVKr{rhwAgG*CvH0m^E#-%^nc-F%Pj%hwLwRLYFI9{*35nGRWB
zW_n~pnHi9cWoAS+m6-|ITxMouOPN`a`7+~?g)$S6tz~9Kww0L;*<NOLWJj4fkey}b
zL>9|TM0S;#3)x*}Ze&lHd62zj=0%puOhWdRnGabmGa1=mW`5*AnFWx8WfnvZmsuD&
zQf3k4XqiQkl`>P1V`UaYj+a>+SuHaaIZ<W_<Ybv8kyB-sLQa=i8aY#D8RTr4WhZ!W
zQtH2vbm*h}6K^Ep<E1U}r7xRTI@A3}K=T&QTQ>NJ{|_#E;>zd#y8FpPhnLS=GdMi>
z)i7oJlm8RHC08Zf&G257u1a*s`YO>Q8>+;BY^)L^vZ+c;$mS|BBU`G(g3MP5k1SM)
zfNZT2E3&OhY{>R1u_HUG#DVOr5+|})B_gt`N?ge9Dsdxws>Flrtr9P?R3#F!uS$H#
za+S!){wnb!2dX519ITQca=1#u$dM|EAV;etimX(Lf*h-o7;?Ny;>c>1sK|*bNgyYy
zB#E4=k`!{fO47)gD#;*ct0X(YSN|h)ocq;l=k{-%)2C0LJ@{1Hw!dvp@!ucWaOlxf
zgK-n@sZLv~YcL(MzQOd!h6Xbr8yn1sY-%tQvbn*`$d(4PAoC5zBMS{CAX^*Eifn5z
z8?wE@?8uHW=6D?LIh<q6i7bvW5!p4yT*&S*=0^66F%PnLjCqlzF(x7V#+VOT9%C}H
ze~kH&17j?J92{f83BK>+^45{v+pGP5zY~%toH{sB-G9+jlhWta!~Mf^j+irh@J{%H
zo4(gLf69iX{lg2{Lc`!G-P4P*g-N0M$Nj^zxLN$*+p3;lChyD7f9Hk%;ni&QNrR_!
zZ~ye!y{4JoZT-U|+L4)q!+UI#x`!P(=STg+E010|e{gv4txU!#<KMyr=Q)A15B|9Q
zTa};#(hGVZgJ1wM3PvE4U;;7=W+01T0pbN7NDu@dt6&AP2{s_RU<Yyt4j`xC1QG=i
z$R)Ud+=3g(BY1$kf)_{<Bp{#Q1Cj+9$S?ST0zv>NC<KARLKr9_M1Z0~6i5*ipqLN?
ziVJZdRZxKvLINl$B!N;w3Meh4figk{C@W-vMvfVoH@F#ntiAs_>iOlNR~?Sxj<OAI
zRfF$@>)(qW`ZhSYb>+B~aO=wHfb=;%kRfLPGUkjxrkn}LoHGMiauy&y#{&sD0mz!O
z0@-plAbZXZ<j6UIoH-|um=l3qITw&S=LYiRJV4%@7f8xUK)#$0NY2SX{+u5ukP83>
zb3veRE({dOMS!BYD3Fp<fMU5AP&^k0QgbR$B9{P4=8{0ETnZ?iO9N$c8K7(~`z?L5
z>%G=8=l<pW4?Q_*%yA3eA=H=?S*$S;*;QjMWOt3Zkv%o$LH5>|7g?$?3E5X;K4iJZ
zWMqGh`H=%P7C;WxSa5<jZa8<{q4WBtjc05+_0W~m4qg1;kJW=KLqj7IS65A3UGwc#
z&06pFz>XEuM*cmVA6b3csx?FTZ?mB@&MTaIb|2$sCNyl)@L}4?Q>Pt1{m3Jf8KEV*
z8B>(SGv~}Y>X>6)Zm;B%{Q==<v*$Zb!~Fm6=<IpN9lzkjlTPOJMw7*Av!CJ+PgpdU
SU%2$>*^$`kG@J}ZiTb}VSDDlR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp.544964C8084D9197.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp.544964C8084D9197.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a56065d3cdb248bcfe33c97d8a5a423949490d89
GIT binary patch
literal 6928
zcmZvg3wTrI;l^{mKuIr3FSG@^0)?FW{hVT&(#XFwluNjcl5&$E5JXm?3oxakQ7A$g
z5WFA?116|@pcKT7je??r?or%?jUXxrLI?yz-IS<7a9ik}W8eSttk1*8dB694`IWEd
z>&f}XkEpG6l-O)F!4Zq+&0aL4i_K=sOaEE8Xu;f*cAJgOw%HbMK2(2Z{+Wv05058z
z_B6-*{@kg@`{utpf9<^d0i&N9wv-u`KjYUYpI%@3{k4a7_%>IZU0FHQR?Q@yH0SJD
zK%TUHEJZi?&X4KwOx2QGj~ppz-}J9%j?Gt;H$OP>mm9xa)Gw8sIe2=XpFOW0AJe_d
zi`VUkeTA!kbCYX7uxZ-AQbShv{cO(FZ@zkPYd|V~rY!SB^4iRe%A3<ZJ>K|_J<-(-
zPv6Q)Rc`syTgwXn<gH(<9GF;A|A~BXeaW7O0!_>VlP4da*gHC~vU$_)U%X7$)n1tr
z%lm%an>#;zKeYc$L%YX!erdz##>#2iGxrS{;Lq6f%D9{DzdCj>^;-4*>?e$-PpdM9
z9bNKXo7nXFOKY;u|9sBAyNc@f)iJg$H5Z>9xO>tgw^BC>yfZ&|*`7IpJCHeY(USv%
z-+VILH+B8_{Xz?I-uL>cExlJgTXWccIRA;uk!}YQgZ_4^N0%L&<#~(8uUz+zz9awY
zitc-g4p)9w6EAJ5Ykk|38uuVGOI|tAxv<XeI`zfKsY~wv_N%<zIYW+?UYe!eT6OHm
zia~#VIe&T9=>^Yyq?eX%8{af`>B>>tHssFQ-Cn=wYa3hKe8+29gASA(=bTUFRu!BP
zR!|e>OI`Op>%ZsyA5SlNul>ZKx3&%2Up9JVK0mblp6<&}_xh%8$cU?DztCTNd&cS?
zj#R!sc+Ii0NiDli4?p_VuJ#{=oqwEM^4h??myUV=G3B)9a?^^vt$)ZVJ-IdUvHg$i
z_E$F_cxv7|H}8L=B5!Y7xqp=SY5(bOGw;k_RsH#`U+XV!dH1n~@?|66YpwMqNa5~=
zf3B#Km(?A(=+E<?T(Nm>`1>1Ae{im>c<4T_FthmXeQWYxIC6Z$>bcjq{(aEzT7R?g
zCHlMGZTkuvmOcF4-q_61AC8Fc6a3F_cHUdCV@P%VaC-K!Q;~5yKVAOexj&a_m2*xG
zVCKAi{qWJpYlTaV`PaXDh^#vry59C%`CMkvl5+=7)_?rnjMg>XH<S0iX}!<A;ZkGC
z4bQiuVzrs0E>=FezOrzqx?{rd&xDf5)<dUeAG~YsfsgcD|FfDf@Nye>;g!RmoL|0s
z{fsACmk;06crv-^mC%?kFHG<G@c8Oi!vmU^j{aReyYTak>Pr_|8&=Z;FF$(Wfi1;H
zx141@EBO3$a&g_=dDp&9X<ILbc7>|GI==SZqf^%ed$j7ahJ~*@KPhT*oZ2{X(Upce
zV|U}|OE0}K<nQGT_BWgTL)YFqIOX8^H=2LEGy2G!YyT`-((T{M_OpTHze{hng{Qyx
z*s>39=B(KoG+kL6va6?8?eb*w+<L<K>b8xC#%{au%ZEpFA28}~U(kCp*FF~?Tw}AJ
zS=cmhW-(>6k%D#MK97I0^bXtjiMBvVQAt_Hkly_rUm1PdBQ@Jx{j$k~v|W%<kl8U<
zcGeFsoNM+Ctji{KX}ecOue^@IvIpP%`pX@cn_kH#{<K}3QQWO#u<Yi0d)MyL%?Go|
zn6#ank(<>qSoVsVv2!!7D*JP6^=aGH)#dFNEIZb6>$&Xx?N_tOq_o{9qfd3mVA-F~
z=yy;q|I*AR<I;AIj2@XCgJu7iTmSpBjqhB}CO=Qxc^P?`9fM`t7R<hHVc){e^~uZX
zSeLa&ZobfVVNk!EfHOilH&WI&krSXJ6pfmRPz-7|Lb0g12*si1BNUHXh)@D*F+z!`
zr3fXVmLrsmT8U5!YBfTssI>^Cq1Gdmj@pP&25K`xnVs6YF9u|F+z+rUs?+FP6sOBs
z*{Nui29I{p;4vx#9^+!bV^tPB*2RLysT_Emivy2WdGL4_51yb3;0Z1PJW&<F6I~*B
zk}82Gxg_vpRR&LX$>1rf0-oYhz*AKfJk_Owr>Povno9#uS9S1omkyqx8sHf&13XhT
z!82VZc<!F=lAktSz+LL@XDMs1muGeCarlu3osFG8WM|0v2&k}g6>gFs4N*EoLzE8F
z7-fPCM41o+Q6|h_lnt^FWkW1P*)WSyF33TY3vm$T!W>5VAP-SK#6y%1^B5I^0z`$7
z08t?<U{nl>5EVlrM8&X(Q7I@vR0>HDmBJE6<)92vIV3|=4$Bx-f(k^HkOEO9tYB0P
zst{E}Dn!+=icu}7K~xKA5Y@sOM)jZ$Q9YzXR1fPIHG&32jgSFRBWz&Q44M!%LncJc
zu!&K3xqGm6Q|AQS)$UQC#<(Y2%6fW*vO1nxaKk4k=M}iCohizB19kdi-zox|AWEDk
zQOAk%66z#zev3M0cV4u2_UQygBaco|H1g;cLL-lk5*m4QoY2UllY~Yd9kbKOV-ggD
zJSIgk$YWXvgFGfm800Z=!XS@H5(arp%+4T>O;9ZI*c8Pgk8L3=^4KV0k;ldfi##?-
zSmd!WJBvInL2=0AQWS?gu7z;O<D!H^9v3GZ^0*}7kjKUB9P;=C#UqbTQ9Sba7Q!Qs
zj}jhve4Oye<CBC(9v`#w$P*HjfIJ~Z3CK&omI%lbqC`NR5GMligd`D=C&cUm^27us
zB2P?FBJ#u*A|g+W5)pY~oQTL1lSD+G7_*DWlM<AKJSjy<$dg)#gghxqB;-kPA|X#o
z5(#-y%q}5MPEa!P<P;?%Pi`SH^5iIyktfHAj669>WaP;)yNo;~K`F>nQj~%`rG+TS
zQ=&vco)RYt@{}Y|kf+4#3i8whr6NyFQ7ZD(7NR0gjS>}kYMiLZQ<FqRo*J{O$kP&(
zhCD4rX~@%Bh=x2ZN;KqYaiSqlOA-xvTFkB?Pft)f^7Is?BTsK3I`Z@=(UGUeiH<xy
zNp$4tF}sdDBS9I+Gg6d+Jfnpe$TOnEK%NmN2J(y~F_34(><0481Z5)6Oi?EC%obuI
z&x{fid1joL$TO3~M4lP5o1MJsm0#FyBp02gx&%A}JOS%zIokDfW0Ab(Nh&AMtD;wp
z^=Wa!n0)riQ&nBl(-oN&-L3TW#hV-NJ@op;^mJ}sZcpn?FepB@@rScFPUNQFuN=<4
z@CbC!pcn@OigmD{I0pxcckrMDhX6`+h@d2g1WI<upcID!N_D89G=~OCcj%xDhXKlT
zm_Mlpecr!{lkOUq{{NQ@DVbosg)e=yefi<@7Y^K>t}mHpO<T+3{qGAhd=l|Et2-CT
z>2(eV<#Ub&6>!e%RJ4bN9_^)}NBd~#(E%EIjE8|9<7J@7_!#Ih0S0=khlL*NWueFV
zSm?0<7J8hAgC6JQpvU<*=y3rKdc22+9`EI$$NPBb@c|xsf=7U!;1!@J_yp()0ReiV
zM}(f}6`?2kMCgeD5qgqGf}Z4+peOkx=t%(yda_4`p6r#OC;MdR$pIO9ibsK-;#Huh
z_!Q_V0R?)hM}?m1RiUT)ROqPz6?&RSgP!Kqpr`pX=xG5Bdb&r4p6=D5r~7p1=>Z*j
zhR1-O;WeOV_zdV70Rwub$Aq5gHKAwvOz4>b6MF9cZpQkRmVTAs+!83oZCJ|so671q
z!m>8-u4~OrlP_$Jq_=yY5q-v3`?GEFYhB09+g_WVF21w4$a)WXqkMd@>(K4j($fWW
zfovVY9GP(U)#-gddN(~?)upPxg*x@zTUUhqL2J{~ZihSHnzk4utK+oH`c88G-|zBE
zTGwQ^?!~Lq-HO7BzE;;dQp?ZkI7;j6I%y|o{ln>#Fq#I%&<rS+W<hZ@2a2b8Py#K0
z5@`{XL`$G#S_Y-i3MiFUL20xGN~d*D25o>cX%m#Yz+Gf5!Zk<RI#Rj!_H(WW-9xk1
z#5zw*pA+6bOoX-ZJ14=~_-Rm#p8>`CSx}sx1I7D!P=a3oCHh5Bl3xNP`(;pyUje21
zRZyB=1Eu?QP=?<CW%|vZRQjCsff=_CO&2bhwXp8?#q`_5f4`<!7dG2HotKk4FQ;~1
zPW$Q79{9uj{ui6e$J#TdPrhf$)M>U^^MViDpZ0zCjkb5`n$gXXnVoY-ZudJ!<PEK1
z^0Gz@$uBPG(X&^Ddr+0<?!o`36pEDt?&A9Xui2uKpOyA2>+hslUXWBxHwU`<l=qem
SZ(-4}k@u3?vDL$=QU3vYpDQ8&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp.6E89B9B9359CE514.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp.6E89B9B9359CE514.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c34436ac5b37917dda93af7a83be1caf3e6de8d5
GIT binary patch
literal 6928
zcmZvg3v^V~!G-6ZfiQV!UL*lh9+A24_r1X(AtMbL0$~6tVZuu$0)wal(gCVObSwlR
z6c`l|g$639T@V6de_8{I3eu(cS7;fG3W6{U42se!qYi>?hq$iU|5|_6;^gkL_g){l
z?|UZJR98ESZMLdF&64?Z7SHTqv)OV}{}wG?_`pfK&BkWhY)e`X)}3B(dT7pv#}hmI
znB)I+w)3|G^WI&sZhqe2aZit0#*E6F`MZ<rH}wDh%ELRntwYbODj#X9Wa3YmbN4JH
zPuV_}A{)Ku#`k`vV(HCC59fDm{`WJ-7AVSFADsB#>%U$+D4Ccwynetho>z~L&+hT!
zHTxlN!J6OS;5znip7HPGh&2O0oBR1US0CEymx`V#Nk5UeGHa9a){IY&H~w=^WKF~R
zn;FUSEq{4?dBI;ibxV}}lZ)#<kq>Mr-t(}(iJ3Qb>ha0_B16iXH}8J%WxBTd^0a8~
z_fNjH^TYRpubpn_aC^@!YZ%v9K4W|Oz7d0cX`5e}aHHcl#||W4uY4`@38U%LinLKj
zm%i67HofuE+Kh9*n!E4L!n%F6jBQKRg=dHCp7Q9;<n?^dtPftcr%&Scr;l0u)R4e8
zpUm-2-*E0Vp@lf_d!utpztzuH9kL(Fd*V{K*Ma!3zjyZTv7=R<zhvU7Cy(em@;+ag
zy{GU{`DazJ{!O)Q@3@l_9%5$8t0p@a)!JR1Uyhl+^!{(J=I+iKakT%%+3L;J#}2O?
z_Scv5R%D!7_}oW&|Nh%1Hr>5!)!1ztb7t@Es9XHCjV)@v?e&ad`%8{<&Zl!K@=psZ
zsYwf@p8KBl-ShsBr<T6gapK_H+eW=sGHy&BKeF_m>=mc_ep5T5=JS#V^%vinx#ow%
z<?jz)d#q$i%kER7k6zu?@uRTw&r^$EAF}u2G0#7zopN7lTDiCFf3o_Y+#3Jb{%3ak
ztF8N=o`2-V{cjG<-P>O38!LW#NBuj@kp-(Ozqt83{lzWsKHgBee9U`o)!sNM*xm52
zl@;>x+Wi-NxxSMtTOSC0e|`N2XG@Ak?(+z<itgICHt&VQ$2YEd;M&%I4Etl-?>D_f
zf7h>lUqQq2N50z|oi*;mn%F+U_k64K-uxXSD(gnmbB=X}C+z%m#fNAATB4QDJvo?}
z`_8pPM}J!_Tx`s{_T9te$)mw*?SGKZrWY<fd*Ec<$KTCtTbtcV-utHQzOs!M8;h^I
zza1N`P9J-r{ILz?1v}LplSY3g6o<DS?40vp=L7pc(sO*zYQm69?cDiS4t;WN#qJF=
zpJ-b#dUNB+#O7Cm<G(sz-{+Bum9K^dH!mCa$2xY=7n{_V&bKwJp@&?0?EJhfMMt-s
zVLr?M;#6Wu?OnN7zD{aeF9df5E3O`2_wLc@YXiO8^x30Am!F>!F*#0coV@sQL#?s9
zaook1-W>6d(gyom&AyTAZXTF+;M|+dKi(dBbncaZ6)x@dA7%R)f8sy=Z?uQ%UwnM|
z2RE|TZVi~OjE$L<^%c9^X??byaK5^2)4{v9UH|nXHQ9s5{{2gOPx`v&V#BMl?WY$t
z&7W07*=%V8JHl1lT!S*nq{%jaep-Hd*O0pUyM8i$c;RfbcSvm}sZH5^)B5Ii4VIm~
z{x?^P=XSoGWvffsuAVMW*I?PXRSV}GyLi~lBvVrMfV2UXU4vy0zwz}~J1#Z7l1Y3i
zyC|)wSJz<K&G+`J-ldxlWs>nJJ0~qCqieA2KiKCtZ<3$8^ETVWlwDj{T+%gI_7^h;
z9gs`EO0g%T?A~d;)4K-C{xPTSPiGpBT*@TBO4+$-x#?YlW!o0cxo^?Hg6{p3m(jIf
z)*Zg_Li_n)gR=Zim$TgZ`?CDBiw4EG7*MQ>1;x2IP`rx=CAb7oqDurNxg=1sO9rL5
z6i}*51*N$(P`XP8Ww;Dbrpp9X*1v3!b$G3_pfscF=)+D3I2*fnf-~rR6jaE$n(CHx
zfQBd?q#;U&XpAxe2BJ)mfhZGVFv<p4h_XQzqHKu8C>P)$$^|)yav=_*e1L~2ALJp*
zhj@$%0Rf^yP=Kfq5-=(TM2Lz(5u##9#HbXIASwkVh)N*|qjEros2r3bDu-l@Dggze
zN>G8Q5>ha#22_ZuK^3BENX4iY&>*S>HHc~<4WoKMho~OZA*zRTj2Zz0qDIhws1Y(S
zY6eV*nn4qyX2`^7S!vmD>!i-|msOUH1vS2Gs->*oR7ThR%sSz%vzcUc-N&ryEWdlO
z+i!KOyIL@#>u%N^a>gm=Ww;`oNy>R0b?R+lHL)0}J4}Q)Poj<y=S9>B;`|nM)b6}s
z@9xuaibfutq-f;PErdoM9U(OG=oq1qM<)o4JUVKpk;lX-26;@9VvxtQ5C(ZngfPfs
zVuV2+lOPQ8n5dmW9vi1v<grPLMIPHiSmd!0!Xl525f*uDg0RSAqjnZ~T%6*N$0aEa
zd0Y$OkjF&`hdeGuIOK5&!Xb}~+BxL$af(MCpQL!?@hyZ$9v>k*^7t6xk;f+pk32qV
z=aDDGDFJyxk`j=YdYcuHCq#&VJRwE|<OvBPAWw+e1>}iwN<^NRq(tP2Eks0~7$G9^
z#268gCnku9JTYn)ktf9|33*bIl8`605D9rwgh<GfVnjlolpqrFq^Mm&o*buS<jF}&
zMxNY4WaP;aA|p?Z5gB=Mg2>2|qjnj2N}N)Vrz9x_c}fdWkf%h5f;=Tg6yzxhq99L+
z+7;xfaY{v=nxs_ZsVziBo*E%4^3)hnk*6kziaa%HSCOa1DGhmAlG2c;wGa(?T7+oG
z(_%zJo|YgQ^0cU3L!KU|bmZwtN=KgFLUiQm5uzhcj}aYtdV=W4)1!7Bc}AQvkY^+*
z19?UZF_33Oh=DvKMhxT`31T46h}sS0nQ_WQo|&Xf<e4qRM4lNTCi2V}F_C8`h>1Ki
zYB#%im8-tAT~91NMfLbUPs?$x^^Jw{+NY>2f8U{ftE@N8N#pa_%THJIOid3>ADV4F
zTkDs!Hr{*ijSH#ioZOr~))zuRe0<XnXRe>fNj)?j&Vld<bkLv}2Lp<Au%I{x2a0#_
zpah2iN_2>zB!>h_cF3R<hXP7<sGu~521<A6pbUor%5<1NtB1Uvzll@soRIqai$@es
zvOa;9J=U?}(7E&bZ%x+~&#<Pg-Q)S`!wKF*+|J7Go#ga5M}zV@?*`>}&gxdQn}#0k
zp`k~6Y3R{@8hVVIfga;wpvQO_=rMi<daRp;9_wMD$9h@lv3?eMoSTCl=i#8oc{%8D
zehzxPn};6n;i1QSdFb(e9(saXfS%wHpeJ|*=m~xSdZJr|p6C&wCwfKbiGC4!l3RkF
z<dL8!c_rvcehGTATZW$Ok)bDhW$4L%8G4Fafu7<~pr?2h=qY{$da7H6p6XGdr+QWB
zseToDnp=aO=Fy<1c{S*1ehqrMTZf+R(V?e%b?E7S9eRe_fS%zopl5gu=ox+kdZyci
zp6M~6XL?QOnSK*`Wp|V@)-5gd8N!uGpp-JhQr3l>o6)tj0T*zXa&DsF!VFV1Y9>rE
zsM#>ZqUORBhnf#lJZd3K38=*|C8CzXl!RIiQ!;8LOev_<Fr}i_!jy(u4^uj7BTN~n
z%`jzl>s?oxo2H&`4X18*18N40w{}DOlGl4qn7_R`HC=RjQK9u6_GanCK+loeucW5)
z>3rE*@H#x{uFvZSe)Mi?x}rzL9o8=IeD3YbLf)`-sp&FDS)Mh0QR|uLm0h$ZH9fRo
z=s;`QT4WlW(Y3e)2Z(lZ*8eN@P*R;nx1wnpJX)o}V`v6EMrFWbX%;+IWx?ZU4m?if
zz~gBiJYMC&6KDZEK^4FgX%Re86~U8e2|P)az>{eiJXw{&Q)mS|MODC4X%#$GRl(C}
z4LnWNz|(0RJYChnGiU=mLp8uNX%jqCHNh*(FDtYTa_UUzQ&w!LD~|RjN%`JeFD$+3
z-kNm=SnG+YOZC=bBAf|6=M*^hJ{lC`V?eP!78K{>K=D2vl;9IUi9QjO<dZ<jJ{gqa
zQ$VRc6_n=FK<Pdml;JZ#nLhJpm0BmAH}lq_>7s?R7uDX{sJ}J*^fS%cu-WeE-kjRK
zIlX&x#?PB}|J1NnbKr#|g=zIu@0m7zhHduzz`XlYzW2Uy_8vXcdO6ZFvu?}DzP%=Q
zWEGQ}Q8Oa1D8F}~zC+7~Rk-gO{!67mEFXL)H}L<PEiC>;|3M{pIBAv_BvsSRA+7<X
V{ba*aP&jJLy`=i?%F)!={{o5lO&I_H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp.2346E4E4EAE3DC7F.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp.2346E4E4EAE3DC7F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa88835956651e2605c811b9cc57c76a08c78491
GIT binary patch
literal 2116
zcmZvcdsLEn9>!lUAP8!B!3l4#w3v=EY36;rAn2r|7%EvxNQ$C)X+X`}G`Wt)Cet*t
zcFppFoG!ZBs!cj-jh3}WGie>m)y+$pvR+EljM2<}Va{*eKlVL`=X}rS`SIi7;d@{Q
z`1^m2fgpNpKxVQajjsto5CM6K(?khN8W6OBgP_c+K29>r#BJM{x48PokJw{TCN>su
zDe;yM<GMoVwmRQr)Z87PV<5Ww`N6s`yy9i}*45_y$nvn`*_mJaq218MnFm-@uS|4k
zO0Qa%fu3n?{`6{XWquUped#P~RnwySZ!<;RI?H{7lUF9DGFgjJeQ!Q+^`okJRU3+Z
z<rd?+6t7m}eq$M5eCB|4-4u096s2B0zI<ua_C&VdxoS0DT&r$-thAxDf@V!vsT@kl
zzk;)U>@5GKz9G^0{!^G&ds{f%GU)G+EKM-H^Xk(ai`2c+n|hp-QnI_<_`$&5PTRJ!
zIAO(aJ*Oy@r5DCLs_z_K8I4GctnZ_*7<R4y>-WHV`p7IsEhtz#7ad@myko$tjpi}>
z<C{XgYA;{%bMujMpW?vbYXUttrd73RMjHdgmq_er{I5vQ8umxzoG0ryhL9%x>czEM
zTzc@yk<Leb-=r&iOeIaF!9U1NwqMrMVBD6BhGl=RW6A06KQCKC_4hi%jjf}PoYdUA
zJ?;yOLa?l*VeF>wyJ%_`F_7s?p%(0_G%UFiBBjF*=}m*yd+$8&6}q<<J~NUOjgoRY
zq?gK=`j(mGjC7HPyw_bM+}TjT?##lJ;*!yM4^N&5s%#^(D}@TRQ^mK_SSX`{!fVn~
zB@fL=fipIrkBRVuv@eG^ff+o-LOd;b{?X6H1;vZbFPkKu(X=Uy%<9P?-GGs&F`!AV
ze0nfqC@Xhyy48%(FcR`0tykx?QDe-F3sd!<5BJBnX)EIMIc#mWmx=#w_6eAa?4q#B
z<>`}tTq~nusjO|B+ZEQiY90GmyRy?^1s}~Mc<ZI*zBi3#yM+5{XLGC)0&bz1ucuDe
z<ayc^MD5SecazwfUDev&j-QP<XT%OD6Bq9cOYjy1Ende~nL6GeMBZFll1DSB6|Y=^
zzFpsuUpSw>ktC<~FfSyIWxY+DpDjIVQU`0b)TKn^g@2Iwp9t~{ZGWzxRk^GsXrHRB
z53~GEmYe%N_8KQwkag>s!@2a!4uMY7mm4QO4i3ssQf~>QP146BE{Pcwj_cL;m?wD8
z8!+!~S<gQNj`1ySLyzO19vi#ZH}v5JYA9j0=fI15NwlHt1KBSO1sQYB|NM9c`kcLP
zS>DheJKkR&aP`~W+gF3E4hDPu#eVgX@tJ<34S$N;r-jq2zyH)Rus<dA&xIsGWC8V)
zF2^ZcF54rQEPg*aGk5QJ%|MKQ64zjpnIp7!QcTKR+$U0Gk#_Gk;sf#MLGO>XJ5dV#
z4VU3AmWk=!pI+XKEjYhY;$dLlt0-NQ6fG3JmsgJKK9N7fl&{ondvT^RZq=ouCVT=C
zXC)tTR)Zk*+KT?eX{$0FDC{7_hl#L-N|65s)q}QZFQM1Ez&LddViVCstV)2by;d`p
z&vXjKsfQ!BDcY2v5@5TJ|I!?i<lKUXA`m+PmB3I5u)Wup7OqWaHRIF|AT|k2(ozYq
zeJ}t0og%a1t`5XTY;&TyrAmOEA5HRaaB$4PsRtvr0op)YCBUBL!#`8^EI8uSLlD~-
zZH!e3u-7zrlX0C*Z*`zh#3mETuu6cvv5nL6GWGs6PJKUO8=?)dDgm}g5GUUJpZ3(W
zRjn2GuGch9dcdGH1S=~87_^08WmN!!mJqD02w>0-f|b<(3|c|3vJ!yl>@`Y7-}0s!
zp3l}~+kl`!F6UB-)H9_BHgz?ng9N<Hs?_N2uGa`QAQ*fBs?1<A9@|DZhiKsW=Dy}(
z;G^(-*av2VO~LcYG;%2DFYmkdA9^MxJCK{qIp$pO7H}hyQu)h(l#Q94sF373f8;bt
zmqY|#u}OTaHC0={K(Lplmo*4~XAe6T<ySX{U?PP`1+7lJKIv0NxH*boayXd>l1%@4
zy<3_a#6mC}26I7Num?yI<r07-bC|}n>h3p{#TZkJ?ZGK<8FE^=6W}yh2Lvm30bocZ
zNa8EgibXNvqsrlkQvP_jV1S^=t?}^IcxY?nZH-Y|WAuMV4ZbaJ2j(8X>OSVkp~I0;
zTwXL3lN_5QL{2c{f;2QyXe|sDhu7JrryD@<qgxTQ1AO(34Gh27?&{|8wWpW2&sRpa
x?mL{VE&d-RnwXlwmgGNCcG`Stx69ta(aFVp4~e>)VMP2UFh<><Mcb>k??0#XJBI)O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp.DCA2A91E3CF88135.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp.DCA2A91E3CF88135.idx
new file mode 100755
index 0000000000000000000000000000000000000000..46dd055c5ef2a0e9f4be5851d2699aa0f967ee4d
GIT binary patch
literal 2116
zcmZuxX;4#V8@((EganW!iH0QszaSK%AhHC?nm{EOf(Q|136TT>60nINi?#Am1}lOf
zE^S;8LPiBdMzBHw{h*>BqLpP(3KannQkE7ZVib|yq<y77dS~vLGv_?_ndd$4GdC<K
zFz|C-2%_<VGKKtuTx|$~jNn&%P?WTw1wkul2+F+B$riHAsOv_2#O2lBq7Fox*%3)4
zxLdyTYf1riX|@h~qb)JVM0EG#{bhfW!rSzBce~d^i-XQ5ryh4fTcLAP_fd$BZqeQe
z!*T;Ua;o9Rhw=A{*<pmw`ID?AZDRTLslqn>#m?TbOQRE+toi89S8rmBqsuuLR+Kx+
zh%dG%$Cna*X5r8EJE7cHMa`pyX;%&{o*%Y9k{xiWT!WL)V9>;q)s&p^vo%sk22%4b
zq3s{J%KukYlZ?ODN8(i677X6(4RjL9l1%T6f0$tr+7umWhX|>~TbuCrd$zaOH<l&{
z&ivedj4UZRJK|Y>XTM@NGC8WMlcq3jUH<hQqlz{(t*gNd5zoX0*$X%JcsKfa4uAWq
zz_8rg-}<TLP?>KLWAG~9kQ!`LZl2LdM{uQ52S4s+LVwN9$eh#G#<ih@v4ARZg-#4D
z<mgb#!_F_#mA>ZEx{{D@<YpT#8fwvRONYa=zt*>6w{@NFUO;qpxRPoch9A0Uc(i%m
z6BLG`SPO%wRsTQYD6Kd~us@lSzeQqNd?{2$BR!zi_1bQ~^Rz?Y(Nxfnl@<;Ya++o5
zOM{KAGOaVxMOyL>50PL~O@3HQmTn1Jh|GO(^axYZXdNaIC^e21-9E@dVDk&E%8nI3
zupls|>^>e4VS4>O8(=dsILf(1KjG}dAB*yf=3SrHNxkCyCUm=Pp7hd8=($=w+SZc3
zJ&^-hhvz37ER1S~Lhm2!&_9l?jjujCQS~vlE3r{enV828)1y95{_Tcu&`eY-nN=oF
z9}9@F!4}E78(+k<hPN!)^1gQ{JszI_)<TN0U0CdVRco<Du%lu+$0jN07Bcwd#PQ3y
zUJm)uJ2Q-_QhSRlIvbiW(~+mJVL_$hqD|pRK78i<HPi)j=W3&<`h^8~9G!Az+>QC=
z^2WS^+4PkZIi)@LZ1PCf@5!^%CHu`PNjf(xQzLUD-gE~ZVS4RqdTN{{S=8Y-O;l8c
zTYc4i`0#7qRrarSnYa3#PNiRTVz^9RtQ~#Z+bcgxxy6^&$sP^4C1;S??pI#xKEZfZ
zlQ?%v+yBpZPP|DSc$C<8VB}oqz?)}?fu!m7-OqkX@f*nAo&C&|pE2Y5%i9ys$LwXR
zvYIa5i>|VuD_?Hiu*9_46XN}y<I+R?i7xC)Aes7M?)cKXADVl1rtbQ00f8TtPx)ZL
zc8QR8Z<9;szaE~N`RUN*p7_9&7?aiD9D$>Ya;)3UV>Gop%HfZdL<R=g>+`l^6GCac
z;x^dIGBe-)!}EGx{%J+Ar-@^SvgEzAaIWyRT=K%;iTr_XnWA#tvlEhpCAa3<h*3zK
zCEVw#0YOM&Q!%Z=jfB=@?}4~DBo3tp_}`#@^xj_{$qRPbh1QIKZF8i#ks5$))&EOE
z%JqJQKC}zANjMTo4WI0}&DnX6>yPL|VX$q9v$RqJuzi*n=B`d=-9T&ZhHV0pprZz0
z=fx2MYn+@j(3&BzZGtq>Qv<O3Y$!#sch~n}ph(z`Ma0t80PJZl=|{@8IcKzHC~V`A
zc$6A|y?8yT+Rb8oNFNG^Z6c0HR0FVA*0FCsPrEmX*4zo(rbtti8h|b0Cx}=7k)DR0
zI<=yX<;%5Wo+OYOEUO{|GDr)SRZ)GC|4s>(RS^Lhqyx*UXn;%$drv;oxwxu<;WD%t
zc3>i=+o?1h<wOZAo4cFSzyzGk3u$p}tuJBO#K>elh%_UZ{>VPkHB<}3we+_PxB7bm
z443Fn3<uF*xFmm4I2bzjtEO+-N5`7sPL^!T81M<`2v4cHAuwfaW(y)zSQ!Xc6ATDA
zaF#mpfrd0aJ{^{++EiOG{GDxKJVZd{3@qcwI11QFaKmtK)%Akjqc4y8mKxO$!__1<
zDGm%c`S@C!>@brB%S1Mj15N?<0B)jM0?=eexAvr_N4+X9JQ?o@s=!xpwQ46oHCP8&
zR_y|iArU`?t2%^O6ffSdBKuVE*AoK>1VydU*fqL$jdIp#^cs!(M76liN$*Bwln9>g
zzP+(g(J`DjC|<}*5x^C}2}~_*1X4#Ag~sTwGc*V?3ZU5-=>_>4<4sK0Z*Zr2e(vS%
z<GUGa@3GO<miTWf&dl6`WM%yy@+P~_9JV+*IlH)7ZX-~((y_QN81b5cEWhm<JN^WX
CYdDes

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.52C4510F756E0106.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp.52C4510F756E0106.idx
new file mode 100755
index 0000000000000000000000000000000000000000..419aa484c41b08ed8c52cae169384d0c80cbca7c
GIT binary patch
literal 19630
zcmZvkdAM73f&TCJn=d8ptt>4q&{|rm;buSC16fYilZB?)PxcU&11+-D6hxT8FpzPG
zBMq2=DkDmX7X=YnLRn-00Y`>GL5U*Epb!IT3ydHU!!qM9?e7Wq{bTw(Jx@O8ecyh1
z?~^CVJvaBJqoL6EV{>x>DSGJzXJ32<o12?EdiwXGi+7&;nrv?F8}oB>m)`V9uAF)0
z<YPX#d4FaUWQF^Gx$?-PBF~(E<;q*1{9foA?F+U_PivlD{!;C+@S{h`U%mRDbmqrj
zIOXPN|L088Kdk)e8xPg}SH63%OZVreo;!SV^XvO-pZxjFw?BDB;kkKN;*#TTKJ8b%
z51;+&OP}of;??z?|Mujiw+{}y$r~@!>z}*g@>`y~_D|ouyzyn{AGf`KN8e~X`EP$`
z9lgI}+h2RHJbYf`lC3x0Olm*qAMxOsx88Bx4L;@3<A46`g}c5xuldDE`+mB3#}|(M
z*%O;;-3PY6*!^|Lc)_jb9`oNPKK9t+ujjc>&bj0D50uTr&sgc+f5}&E&o968`BRR&
z$Gdvf{zu;5|A_ywuibjTu=#}V{b1l8KAYdX{M<V}$1JCO=g#I~sh2~0{(gJ&lKank
z=Gu*(E3O(AH1|Du<+>a0PCawN&fPJ7^(DdXRd>IT+P%kc?5_R4RsP6*>e0({%3ben
z?o{^qewSRe=Yt>bd@s6z?_8V8M^>HqQ7ANeaxQi8F<Wo_^11SR4s8AYY3;ji(w;AS
z!twC!mDPLt4-41aFTdg4?6tc0H;#|UJ#pSG_kBLQN_F|K7k7@25FPAP*KpmgtFAv!
zcg;zUZ@9^SjsC(fy#3+s&ELEJ^H;xi{42jzeP+-6TgqED+_m=93!XdCtH1B~<4%x$
zwC3FB|M{lphyS$kA6~q4OI>~2ZLVSIhx;!5uUl@p_`3&=TzKlJ4U4G@9+Lhl{=Sj>
z&Cm6Bmdl0t9hr;2^NMV3uzmOTKY!M3@f~yW+}2IujC1~upImd~OPAL_cc!@hz_;2D
zh3@%BG1l*Ne)H4~J44x@KDPe8PU%1P=eBI5);o8+_rhh4;9c|F0n_c<wTJZ|ec&%?
z-8-rOFcmkiId{cI$F-(CKYwnN?|*Pv^ttR08{0a&E$7{Bi0hZH`?ssVdfk=-@tx1^
zJN2o4b7f`Om$Xl>xP9?IANu*~AKvFVd+WRBF2Cus=6^l9eD{0r?K?fFx#7T<|MkEJ
z+m>DXQsd+EU(`LmW%oJ1zwE!i@b;V6owvTbhuzz~uyCdLmuEtat?tZ&&#$`7@qYO1
z2Zy&T=1%;|hm+y{9mV%<edD*e$0IL%>F2jS_pLAAdzB{?{+{gqU+uo(C%aAGKJDv@
zXSZvgn3pA<zy6O86i#^P`~$Jwk6vY7x0_oWbZ@@;l&!<B?5}37f9X5rZGSMY%Rl;q
zJFk2C4e8!*|NGO68T*YZ|9aBDoWEoKl(Wx$;>GX&!FJL&$Lnv%?s`D;@@MXQr=!37
z{^F&#rS5<Kr6>M$y4ZjF(F?EVdw)9pv8S$i@xhb+H~#nA4ez}A`Y(Uyc<qD#ENlPt
ziRNEi#})qe$O+&5<uA?~|Lxs3H{9QU?Tm*PPt8T5f5>g4q2o3_|MJdzUj1|8vJby_
zY1|X~)}k)aKIgt??!AJnlj$rMU3S5l$CA0ZWg8!C2X3)%ny1<PTy4#=HOmhjrjG=N
z{;c?D&+nUCteJV5na1mvtv~wEVHO{K*Vop4>Bk42oTtfYylUC1<%bTl*z@sw@7(t0
zWp~X}aT*`L?D)eE9cHm#`5V*CCm+zy)50{~v}_Z9=rD`3mHT{&6AInW%%!LChP4|`
zJam}FKVEm`<A2@1xv_FCH;vCc>`dRG!z?~<#-^X>Ebo}-X<{0$UAA`lp~EcxpQCd3
zzPjUQZ_d-yG+w=I_3}f9Sv<G%>~k*qerW0QbJmJOpO;-ve7xsJuRAxbtSxX0rr8HD
zCHsP7iIkiYGNoM!nTk_Ern0LbQ*&y_)OIyw8cqY5#;$=(%V{Cg+O?4BI2~j<yACov
zr-w{$*F(l}9AuoGgUrAgAT!tvkQq56WJbFYG81Ql%w#t~X6DS0neArCESv>0i`@d5
zm9s)-wOb*xaW=?ob{k|4&H<Uj?tsk6IU#e}oshXW7i2EG3o<w7hRkhuL&kGFWW1e+
z%)@yg^VmI*c{wj+Ub`1EALoP2XZJzo=lqcQ?S9AtTmZ6wJpfsd3qlsO2O$$U0W!fZ
zKo;UckcI3a$iiG0vamf2S%ix~7O_Vli*ixOqV_0caV`#7+#ZK4!6hI|*b|T?xg=ys
zdlIq~mx3&1PeGRE(vYR?X~;5M2C|Gj16h{KLYB2>Arm<fGSMzVmg91e<?K1g@?0LW
zygd(Dfh#~(uooaJaz)6B_9A2@t^`@hUV^O5l_4wJ%aB#L3S<>~1+pqvg{*3?LM9z2
zt^f2Pye4gsHqB`EoO3-{ap=hg9&@FI9ZQe03-yJ|ff@^6BTJ~1lu%XHl~7eSlvq_s
zDyXXJDyXU&Dy*s{HB{AgHB{9NHC8o}2CAC62CABd2CG_03sr4h3sr4Hi&dSZgQ~8s
zgQ~8f!>V4=LsehbLsj3<W0jLQsB(1<s$7G^szEY9)lfG;)zC0t)hHREYOEWfYHS#>
zYLZM)HPuZ}H8o6FHA`lwn(Jn$nj2=US|kfpEp-c2Ee#7+t&$b0*18p{)`k_UHpvE6
zTipg#Tf>G`hva~&qwavJqv61+Q*uJpS$9I!*>GajCApyLs=J`-YPhiKmfTQv*WFNc
zH{4j|B_67Lorfyl;IZnFJW%!2Jy7*DJXrNgUZ{HOUZ{E-Uaa~gA5?vHA5?t}A6EU6
zAFBSkAFBR_AFBZ=0M$S}0M$SvfYqQBgle!Jgle!6#Ht_(P!;L|RE36s)sPf|YN#H9
zYN!#yYFG+GHCzuvHQWeeH6lfz8mUL18fiqZ8kM3@jn<=3jW(iKjZ1N;#_MsY#v5_0
zCZq&Z6ZHgC6O9B`lTs3@$$Ap1$wm^ZDJccjR6PaNR3nAew3LQwx}Jt=x{=0eM#?}n
zQ_ny()5u^oD`laYt!JT{ZDg@3N+MLnx(HRVA!0Qr<)E6Y=b)Nv<gl8T@=(p!^H9w<
z@>nfM1*jJ41*jGp1*{gOB2<g@B2<fwB34UM396-f396+=39DtP4ApYI4ApX@jMa)%
zfoi2*foi2u!D>~iLbY12LbckcVpTFquGy=Gl{JZ%qCi`u+>B;lrv7aDisIlGsql(|
zU+^uxXjl*yB0#Z)1W;|^%q3FtO4um{CG3<jCG3<nCG1qZ3U(?%1v^zt1v^zu1v@pb
zhMiha!%iJj!%kgO!%o9%V5bo@u+zjeu+!8uu+#Ed*l7hV?6fg0?6fs4>~y>ib~-@^
zJ6%i%J6%l&J3X(5onFwxP9M|5PG8f*j^jDlaRLWBF2=!*t8uV1@CMi!1Ox00F$3%j
zH3RI7yb*Rr!3aBJ%m_PU%?LXaZ-SjkFu~3gGr`VOGr`Wxn_*`b%&;@Z%&;@p%&@cY
z7T8$?3+ya03+ya43+$}C6?Rs^3Oj4e3Oj4f3OgHbgPl#V!Oj-5!Om8*!Op=uVCN7V
zuye#5uyfQLuygWG*f|9!?3^(t?3^_x>|DGHb}qpMJ6FsFJ6FvGJ2&r!om+6j&K+~Z
z&RuiEj^}yU@d6J!KE}h2uko<+@E+KC1P|;yF%RrKH4p5(ycc#}!3#TY%nLhj%?mpp
z?}MFB@WIX(^TEzn^TE!~`(fu7{IK)K{IK)a{ICn~0oVnE0PF&>0PF&_0PKQ%5OzT!
z2)ke`2)ke{2s?onU?&Iy?1Y#AJE11PF2sjm7ZO6S3&ldP3)Mof3-e*vg@rKe!m%*y
z!nH8$B76jP5g`J*NGt-oNG$@pC?ADgREWYZ8jHd%T8qLi&c|UF7vivs$KtSy*W$2C
z@Cn!@gaqsou>|ZAwFK;vd=hp^Aql%=ED5`0EeX36pMqUVNWm@@OTjKxOTjM9r(u^C
z(y&X%(y&X{(y+_$8Q5in4D2$o4D2$s4D7Oe7Is-73%hJA3%hJB3p<e)VJ8YA?8KM|
zJFzChF30C!mlJZZ%f)iA%hhtQ%kz2I<%K-#^07SZ^0hqd3VZ=}1)%`DLaYG0LahM1
zB431EQ7FQ$7%RfASS!M=#Ft=K5=yWu#Y(U%)k?4{^JUnTg);2Qu`=w+wKD7~d<AwD
zp#r-~tOC1Atpd9$Uxi&&sKTxqtHQ2YtHMs&EGcHMQm3zA7bG>1R^n!~{$pPAcRSx&
zd(_HWyvO2u+0skVM2{tqlRcJ1PW4y{Io)Gv<V=rckh49OMHYKZM9%eC4msardE`Qm
z6_ATPRzxoKSP8k@V`bz@k5!PXJyu=f*|(H_O;;THUJ|^NACZN9Wa;y=Fd+*cBDd(m
z*C^zDy6_w1AzgR_c}y4nf;^Bd{86^FSB{7heaZ<@qEFePO7toFREa+2kSfuq98)Fw
zlmnR(eX0>rp-(j-D)gyZRE0iOpQ_NO8d4SdRAZ__pK2gep-(*`YV@fmM2$Xmi>lG5
z?o&1T)I+L9pL$Hy=u;15YV>JFM1wxfglN#GX;BUOG<~W;pJqrk=+lg;27Q`=OoKk{
zh-lHLoe(Yhv@NPdpSDl6=+h3V7Jb?=)uK;3kZIAU8xbA)bQ7XOpRPr9=+pJ74t=^I
z)uB%}raJWL1~MJ`^dq82pMFC0=+n2T9)0>g)uT^8q<Zw}$5f9#{XnKiA2%W#`nU<<
z(8skXhd!=PIrMQu%At=NQx1LHK*piZFd_!@879PlK0}Kd&}ZmV1Nsa@YCxZ1ObzHW
z3}goM8Ars3KI4QK(PwN?Bl?VeYDAxLNR8+-j;RrS#(~U;KGTSp&}W(u6Z%XoYC@l>
zPfh4E4XFuzrZF|4&oq#k&}SYIGy2REVn(02Ma}3l_o*3u<{>qs&pf7P^qB`TGx{te
zVnLr}LM-UBw5SDrmOizh&oZPI^jXH#f<DVYW<j5IM6BquPKXtK))uv*&)TO}^jU}0
ziazU@TG3}6$gJqIjff3>wh6JJ&(@+g^x68<hCbVn+R$ekQycnh1DOqdjuCO7&oLnm
z^f_A8fj&o{I?(4BQV04RW9mSkV<2;&&p9Ga^f@QQi9TnGI??CsQz!bIL+V7Ib4;D+
za}H!q^tndFg+AAWxX|ZnQ5X7Led<D=Ye-$_bB(DBeXfDbg+BL)xY6gH5I6eVE$T*}
zyHDNda}TK-eeN-Jqt88%xzWdu2#-E~LU{D?Ey|;h?^7Oq{E+hK<HwXoA3u=s=<|$-
z2YsFi@u1Jsq8{{l`qYCy&yaf1=NVHE`aA=f2YucV@uJT=Azt)(Thxm_Z=ZV6=N(cn
z`n+T6MW1&d^P<l;B0lu_Cd7w6UyJ(C=j&4+`g}v`L!WO<edzNIWIpuyN5qdl|AhF_
z=WkIz`uu(BN1uO4{pj<LsULm*fy|G-z=#CU7nqO$`T{K)KwqFw1LzA3X#jnJF%6(E
zFpve%7aWlw`hpV@L|?E)gXjzPX%KzEAq}E0IHp1L1qZSq`h*b?&?iiYfIgu`1@sAh
zDxgmoQUQI!m<s3<1~LJCp%Dq8FEk+`^o3e9guYOphR_!p(h&MWV;VwVXdnxrFFYb)
z^o1uRjJ|M-hS3-9(=htNLmEb3cud3S3lC&r^hHJ_g1*RvM9>#$(Fpn?eHuYuWJn|E
zi;QUmeUX7Ig1+d8M9~+WkSO}1EgD5%v`?exiw<cNebF(EqAxm-MbQ@@kvRI|6B0*X
zyhY>ai}z_9eeofUqc1+DarDIpvN-w@Ba%R0VnPz=OSEVLeThCzpf53`3G^k#G=aXv
zK$bvXazv8oOHN1<eaRM0qA%H}N%SR$G>N|Cm?qJe9LSRBON~eheW?jap)b{<DfFfK
zG=;v@kfzX=8q*Z|QUh5Eed!TNqc1%nY4oLAG>yJ=pQh259?~@W(qo!NUwR-*qc1Ze
z8T4f)B!j+8i)PT5>C+7QGDDg{UuH}*=*tXb8T4gGB#XZ6gk;f|ZP6_HvVEFGUv@~d
z=*y027Jb=)EQ>yIL`3w76C$EdY*7(?VxNlW6NgkppE#x>`ow`uL|<-1a_GxVNDh6u
z7R{kA*QYu3<%TqezTBAR(3cy?a_GyCNFIIp3CW``-=cZ+<@+>`zWk8p(U%|7Jo@qj
zSss0b5h<XrFd+r>6<V}_zCxcC&{r7J0{RMLT0mc6AS<A+I3h*#6(^*KzG90O(O2x#
zBKnF$T0~!QOpE9%4rE32l}4n5zS4x0&{t~F68cJgT0&oGNK5D|jcEyerGc!3zVe8a
z(N~_3GWyCbT1H>FPs`{l4`~^F<uNUzuRM^I(N`If3i>J&QbAv(MJwp5^l1fsl_9O5
zuQH|;^i>A33i_%eQbk{NLaOMiwrCZ7)jqAFuR5ev^i{{SioWVVR$cP>u6}#&!|}y0
z6IR>)x$WiIb>rFP%IVe8gYU0jx>|~*K>8!a4gdcB-^j|^7Pdt-dk_3pduPX5-8I)u
zPp?0DePH&s%xrws!oF{~4x64nYx!A6&P=m+>l;@bdYAr^X`xLPZiMTw$~I9VtJ*|`
ztZoxEvZhTm$l5m1BJ0{jhpcZCJu=rO9I~NJ49LbdF(RAV#Dr{a6Em`<O)SXPHnAex
z+Qf$JXcGsrvrU}Ht~PNYyW7N#%(n@T>}eAZvbRmV$i6o5A^Y3Jj~r-|0CKQRg2+Of
z2*{x}2_c8uB#azslL&IOO`^#0Hi;uA+9ZLTY?CB%s!dYJ={89tXWAr#oNbdVve+gf
za;{Bs$oV$OBNy7FfLv^oB66usO339lDI-_fq=H;+lj;(mK4@)Q_J>WsH%xy<q{|lW
zfp6R>yG)6!>M|9wy35qanl95IYr9N~tm`rzvcAjo$Xu6k$c8R6ARD{Ph-~UI6SBF>
z%*d85vmjf$%!+L5G8?jEkvV$s@jDlp6WO)MT*&T4=0@fh8ISB)WFBPiBJ(2q7MTy(
zzsUT^fkhTT4lc4FvarYm<j^7uA%_=P7&)@YBFNE27DbMCSsXtfi7rbZC%Y_(oa(X^
za=OdX$eAw7AZNQQi!64Th@9)P9CE(P^2miQD<BuUtcYCdvJ!H+%gV@=E~_9{yR5pz
zXPTl9uRi0^2M_+}REI6x2zLZ!hbfU&9i~E7cbFPk(_tE9ZHH-*bseTd)_0g5nd>kP
z+0bDIWMhXJkxd<DLN<4p8QIcd7G!IOS&?lWW<!p5SR6UgVF~19hb56y9hO2)cUT%Z
z(_tCpY=>o$#SRmZa~+mL&UaWIxzJ$+<YI>vkxLy`LN0e$8M)G7736A%RhL-OTyf(^
zuYUOa!AC^-!p7Mvl$AB5TnVI-tANyUHIPQG0n*B~Ksvb&NH5m|adHmGAU6OR<whWr
z+yrEnn}IBH3y@WA1+vL)Kn}SB$SHRMx#TV&x7-cH%XuJ=+yms5dx3m%ACO<}2MWjo
zKtXvBNRSIaA$bTWEDr-k<Po5#JPH(-$AJ>^1W;0*1WL(MKxugzC?n4RW#w5QQ7!`I
z<T;?cJP%Zm7l4ZLB2Y<Q0xHYPKoxlfs4B01N_94O<L!oze|7MV+GGpY;UiX)DUnr8
zrb1RXnHpKsWEx~`lWCE4O{PQEH<=!pYcdYm&}0T=W0M(?O-*J(HaD3W+0tYdWNVXI
zk!?+8Lv}Qo1KHVRPGnb;xscsW=0@h5j7Rn~nFrb1WL{)nllhSSP3A`qG+6*S*knOu
zp~(c~P?Lp_!%Y@Ojx<>WIof1V<am?CkrPdpKu$JU5;@gmDdcpMrI9mDmO;)oSr%Dr
zG7&k~WI5z~ljV^MO;$iIHdzt5)MO>(a+8&jD@|5Gt~Ob9i7(j_{Dr1qOHRMy+u+}j
zoqeTu<yRj3$H!lL{h@=Wa~rC&(^tPMeD;**u30xdz3Qk{YiIAh-}T3v61ppjXCHk2
z@3DoYcl1mj_<Kx=tm-ipvbx9A$eJG0AZvR}i>&K09kRa1^vGO~ama=qGaws#%!q92
zF%z=6$IQr<9<v}@d(4V#>oFU$V=r?&0FNflz08U1+RI$X?!C;7%<pA9vS%;zAba;R
zFS2hh^CA1E{~6@KUKT(O?qxw_VJ{PqLwi{WIlPyJkt2Iq1Ub5wMVI)Fx0;s~U%#n6
zy?<;9ZQ3$>{{PXX4<DAipc|T=KK_W~*UsMOHM_hp{mG4kKjJ^TtnlNDzS!Hb;_TqG
zZD)4n?2^T!mb7%3(0yxqdW~X@ZuZL!p2+h1Ya4&})b#W=wr%t5-+bYlKi#KZ<=izr
zEy<--v(s-F4jUhS<gs6$o<3{MSsQ1kXTMOuwBpb&5}5r@rgvw>f^zmzOz+GJC6G#?
z0#Yl~KpKSxNUP8S=@dF3y+RMfDL5d5!T@Ac7=cU*6OdV92C^tDKvsnn$fmFXITQ{c
zr@{&3Qn-NJ3O5k1;DI~}50F>k1@b9;Kz@ZED4+-c1r<RcK_LKz6d|CnA`BE!M1Z1-
zC{SDx2TCXsKuJXsD5XdNr4?zQj3NV+Rb+ugg$R^W<bd*uJWxSV04geqKqW;9sH`Xh
zRTLGVs-g-ct(Ml#?!<4&e{?N%+<owx-de?4!|Xmi`)wH0i+2Zq6$ae5#f1Xgw?!q8
zN>l-<MKzE{)BtHkEs##s0qI3O5GQg#2GIay6pcV8(F9}`%|I5>0%R4fKsM0^<PaS|
zPSFYE5?w%U(GA3lJdj890C`0(kWcgh`9(iaKnwr{#UPL%3P2$-1QZs-KoKzl6cwXD
zaWM{*5EDR2F$t6sQ$T4k4U`cxKv^*hB#I(XPRs%2#XL|!EC3b7B2Y;z0hPrvP(`c&
zRmJM3^uLe1GRWHA$xYuC?dUMaxA8f(!<@*j4s#*9JIsyDcNmZC=`atnx5K>1z7F#t
z`#a2!9O$qBa<IdK$U=t+$e|7kA%{CGj2!8(2y(Q;qD#E<?DNk!_#AlA&a*D69lTLG
zc=-Q+Dzh7NbHz_@%4+XkvymS@dC|*<pI)q#%GJ~6%AbaF)uo%KE!{kQ>E>A%Nat^#
z;+Au^%GhDc4wo;VU-_A%jyxiIv=BJy=oL}_s$*9lbKLPKpW<}6d5>4WM&tOLO}X*^
rI=goLXE$s*ar5UD3o5miGZ@WQ`w5nHI#Y1X+VH7o&`{hLA+i4lD!Q%9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.11CF3D24C90FB35A.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp.11CF3D24C90FB35A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..01a0c17c219ecdad25bccf7df7139444060263be
GIT binary patch
literal 19630
zcmZvkdAM6;fyVFm<)fsFEG;e2T3U+XW<S{jSx(lIg{Ik0_7Ii>Ewa=UL{JO^8HYI1
z6f;m|L@Dv2AR<e+EHZ!;2ZljGi6V$7#7No#vPBGw<J@+hc;7#!&qJR4&ilUo(fi~c
z4(Z8FM?<0S#b#y#QuMM5&$;AGHZwDG^uj+EU$W!8S7kFZUtc~mbJ<Og<jR?sPdVm;
zoA+l%K~}i`=PQmpD)P()SFO19iSLHK-o9{)^t9&br7zVU3qN|4{FQ6|MQ48eg;Q^S
z_HSpI{%OTe{`64Yf7Ltpx^(Y7_1xi`nqS*r`{cdX-~8ldh3Dp-iA#^W`Sf4(K795o
zFMYD_bJx^&{KpfQ-99+*I&Zv4uYc~!D{gsW*W2H?qVWaiAGf}LN8e~X<?mm#j^5wC
z^{>5`A3ncv>6V*rCbjSPk9hE`Tkp8;2A}fi@xT81qMhGauKC%?`+hQa$7hfI>E4aC
z?gQIi?EW%jyztiZj`{P6k3BZ`%jMiB=ic$!2g>H*XRdJXzw|4%=a=63{He#?<6X6K
z|0D13f5iXTS8u&v*mS~ozdvvfpUrPtdfpu;G0UmnzO#8)>bIdiU)t8Z^!~G-*|ouQ
z<<-N2=DsJcT6@FYsb@~uu{*}Ex-{6m`tBD}yZ0E5-MRnQ${)H<J9=qOx$~V(oytDn
zZ;~tbeDI?k??%`2on5JXWaWt;g+ikzW>S|Nv*p$=oF~8Mz?Scw-oEQ5?FF*Ej)!lr
ztlHCmSh)6n`3>)6cj?~WFg_x;_xxM#`%HGF>WW{^?HC^+I@oEh;o6;7Uw^*t+LIq&
zf0O@O{Y9UB^TXYnzI**=u6gzNmw&DL)Sl&UC~sMR*P7EVeC|ZA{=VanJ3;o*>hqre
z*PEUn{>z4cdhxQ&b@gqxxrU`5?7QrDx7>2ccMcpm`_xhE=Ta9wB>h$VeIxg)C-rxh
z%Z26JGnah(W!aiw`|fS;ecEmD9dpXemW|<zbNL-VzV^tMuBe}MmbmV~H`@<|?)hgi
z*6(zF_0$bJLfM}@w(h=8>A&{pHgBNTJGa04!sU+OUCX%xrrWn^59>erz+ck3w^ILQ
zDsEbR-m(pjU8X%xJ~ztuKe#;lT=oZzt)1PL^Y1pq^-I_O$2DKMZu5coj%W9s_Ef*Q
zqO#=k+NYP@KKI`bJ-O-!_j%6Q^3HinZ~C<P-;XZe{qDQ_&IoF5IPiskKk&iUCA(f~
ze0;%+y2m%~KKJ*R|M|0TzJA^L>$-c`z1@onSBZalCe+yC&OG@1%F7+^htGL%c*|Vw
z#4mq18SdX+tUo;bb?))V3!i`Tw&%Y2g?q2|gu>sI-T#Z-H~e_F>076NP4VnD?cU|G
z#Piqx@qxk#4_$B|w)@em&1-jabA#^9*POa#_~rf8%=IsQyS()e=C%1pzkla-Pyb1}
z_gnw@^jyY%<BGqY{BIX*Uw-O2=k0y*JAbg9{EhLt8?rke(ERpO_r2B8-+h1avfEPk
zzyH$Sx6cs!Z$5g_^?dK`Gah^D+7}-@`G4YH+Gcp`mDhg$8^^03{C8RV_9vQua~)Us
z`y(fO=jT5=fBg4%USEHI|J5@eo;xiUiT)wCm4=Sn@ceIg-1EwNiOWCy++}f3=$mu8
zMEl(Pp1Jo*vUa7;a`ELCo^>plnOU;o!FJ#l`^M!oo1dwzUb1@Wp@)St!J%KvKHBs9
z=4NYVIn6A@>z1rL`q0BPKH{kNez*6?lXfkqg@t(Il8yYKhiQEDU0+@M`5zs4VmVDN
z#4DGqTzcqX8hbu|_pRGrzx=M{R9uLUUvm86haRT!5B~O*rJ4`5Pkw49y%4Wov;M?G
z57YR-nHztsv%F<qP7@39nk8$N9(tI@e?4;Ril5%Mj;@%=EyQOXc9!qZ!!-WeQMr3x
z+5Xelm($cjylTm+rH3A-@yw2M&b|11p~bI{Y1yH#%g((Y@A=_t&W$T-v-VlX^pC%y
zrnD=8RCX1R+O7uD*fl^}yB0`i*8%D6dLYiu0U7KDAfw#~WU`xp%yu)7#clzz+O0q~
zyA8-;cK|u<P9T@v1?0B7fp|L)<gt5zyml{;&+Y^A+x<WRdjKeC4+05x0Vre-0fp^h
zpol#J6tzcz;`TUD!kz$1+LJ&jdkQFRPXlG_8KA5^3nbb_pqxDil(*-B3ibj}(Ov{9
z*-JoWdl{%=uK-o;RUm1-v~l`9F#RsRo-8}`?F+ABX?FYKt9Z6Pdj(Kq_N!zOm68&w
z%DNJ&%7zlFDoF)ZRb2&DRYQeUwWNlsx~_(*x}nCZM$$l4Q`bOM)6ig5D`}yst!tsG
zZD_HolXOtk)pbzSHFQ|jOM0m4>w2i_8+xpA5(ibT&Ow!Ha9A}+2B;e92B;bu2CNz-
zBUFubBUFtIBUVk4396>L396=s39Dww3{`X83{`W(j8%(dfvTl$fvTlp!KziVLe*Ng
zLe<)^V$~+uplYkzplWN_u<DQ;P<7NDP<1pMSanKHs5<LTs5%=?thyu@R9$r!R9y`h
zR^5^ts_wcQs_upxtGvWRm9O(q<r_R!J(35ip1KFBo`wgjUdanpZ`})3Z^MgKpX7t8
zukM4Yui?Y0U-CoMU-v`R-|%BKAO)Zrs0W}LXauktl!8zV)`L(DHiB3cBmt^IU4W|4
z5U?7OLQoCWLr@JhLRbw;VW@`dVW@^1VXQ`^2vj5W2vj4D2v(y~6spmB6spli6svJ5
z4%K))4%K)gj@5*efNG+ifNG+Vz-m%TLN!@WLN(b)Vl^eDpqi?upqgr=u$q?AP)*m<
zP)#?|Sj|WosAlRJsAd`&tY)PwRI~LgRI`mNRz*pKs#q7HDmFx{=A;}{bM+ilbB!EU
z^HLtF`Fb9z`9>bA1*rhlLcIXhLZg7yqEv)xv0j8~u~EcoNh(3LR4+lb)F@%KER~^J
zu9u-&Zj`ZFkt$HF)GJV}G%8rFN>!*<>s6>$8&#}IM#(k3YFJT|cqt0BS;|dmdRai1
z9lGFOyr9^2(zf!nG`;-)>B42l!HfUJ)+|5koBkFHm(s#)1SmF}0IJQNwMa@{2|J~r
zgq<>`gq^acgq@04!A>QpV5f?yV5h37V5jEQuu}_a*r{V`*r{u3*lBnT>@<P~cAA(5
zcAA<7c3NHwJFTFFoi?V0owlZhosQSRPABMKr;F)er>p5;r|0#s(+hgo>0^4>>1%q}
zaXbe*PT*k2#W>h;H4b(L-T*s;V1S(=W`LccW`LcMH^R;+7-46O8DVFv8DVGQO|UZw
zCfJ!`CfJ#3CfJ#IGwjTQ8FuEF8FuEH8Fm)l0y~Rfft@90ft{shft{7N!p<sKVP}n5
zVP~ybVQ1rQu(Js^*x6z>*x70}*g1Fy>>Ppvc8-_>c8;0@c23?2JE!1;oipZyowMeI
zor`zD&Ly~D=Zd*t=c>72=jPq8a|>?RxnpkFxod9N@jMSZUf^NJ$9UNBH6C^z-UB<2
z;DMbd=7F83=7F7;_rlIAcwy&_d12?Rd12?{eX#QhKG^wUKG^wcKG^wrKkWR1A9ntj
zA9ntlA9evg0K0$?fL$OKfL)*#fL)Le!Y(KTVHb=AVHd0gVJGkc>;yr8oe&dXC)5Pk
zh4>KcLP7|3p;!oZp;`!bVLlAIun>k_I2MLoxE6+8gpa^3B1B*piA7)+sYPHH<)g5R
z3Q^caV^P>eYf;$6`8e$2LL7GSSR8ioS{!x>J^{OgkbqqxmVjNNmVjN7Pr@!KBw?3~
zC1IDWC1IE1Q?N@3DcGfADcGfIDcGg?H0;ts8g}Ve8g}Vg8g>~z1G|infn6q+fnBDS
zfnAo*!Y(UhVV8|%VVA9CVJGq;>_kC?ofs2gC)PyR<@g-zazYMvxmXT%xmpf(c|H%j
zypV@oK9+}FzLtkwfiJ+WAQWI%h!tR0s1;yW<cqK?3PsozV@22%Yem?V_!8_&LJ4-I
zSP6EeS_yV#z6`svP=;MOR)$@<R)$@LufVP%RA5($RbW@CRbW@;tFWsIRoGQyRoGQ)
zRoF?JB*pY9b>T91R#F3LC2mUVKIS!F-0{YmqgK@7Jr<v5i<hE_9!nr6dn}2Z>ai4Z
zy2sMUnI6j^XL~G*EcTd)oa?b1a=yp%$b}v&AQyYAh+OKi5^}l6%E*--s~}f<th&h4
zx9WenaKV4@je5A0ACcL8Wbx}VJ0Y_lBDd)5S1IIuI{Pc+A)Wmb@|e#41$iKw{iAGg
zuN)C2`jiu*M4z%nmFQFUsS<t4AyuMJIi^bVDF-qo`cxyLLZ50vROnN+s0w|mK2@Pl
zHKZ!^sm4@=KGi^`LZ5m>)aX-Bh#GzB7FDB9-KT2wsfSdJKJ}QY(Wf5B)acWUhz5O{
z3DKZW)1n&mY5G)yKFyG7(5D$w4f-?#nFf8@5z(SgJ0V*1X<Jl_K5d_B(Wf0!E&8-$
zszsl6Ak(5xHzGRp=_W*nK3$9I(5LHD9r|=bszaY{Om*nf4P-j>=|@D5KK+E~(Wh@w
zJ^J*0sz;xGNcHH`kEtGg`hiT3K5j%f^l=lyp^s}(4t-poa_HlRltUjkrX2dXfs8|+
zVMGk*Gfap9eTEh_pwG~!2J{(*)PO$2m>SS$7|0ChGmeN6eZ~nfqR-f(M)Vo`)QCRg
zkQ&iv98)9uj02eweWnpHq0clSCiIzF)Pz1$pPJBT8d4MbOk-+7pJ^a7q0c-bX7rgS
z#Ed?3i<;4A?o%`R%tLBMpLtBp=ra#wX7pJ`#DYG{gjmpLX;BOMEPZN0pJhlb=(CKe
z1$~x*%z{4ah*;5Qoe(SftSxFqpS4e|=(7%~6@At*wW7~DkXg}Z8xb4&Y!hNbpRGl0
z=(F{y4SlvDwV}^8rZ)811~MD^93$dDpJPHC=ySBF1AUG@b)e5Nqz?2s#?*m6$3W&l
zpL0Z<=yOhp6MfDWb)wJNr%v=aht!Ea=a@Rt=N!nK=yQ#T3w^E$aiP!EqAv8g`qYI!
z*O0o<=NeNN`dkB<3w`bpaih;YA#U`!Thxs{cb~e^=N?iw`rKpcMxT2ibEA(R5gvW~
zgz)I&Ta-s1-={qK_#x%d$B!wGK7JtM(dQWv5BfY4;z6IMMLp>A^r;7Zo+0(1&oibT
z^mzs{5Bj_#;zgf#LcHknwx}0<-ahrB&pV`E^m)hBi$3o_=0%@xM11J;O^6SDz83YN
z&)26u^!bL=hd$q!`q1Ya$b9JYkBA?A{t5A;&)=ec^!fYLk3Rp9`qAefQ$PCr1DPLv
zfe{IyFEAki^aWZpfWAPV2GAE6(g6AbV;VqTU?2;iFE}DW^aUp*h`wNp2GJMn(;)hS
zLmEV1a7=^f3l3yK^a&#(pih_(0ewP?3g{F1R6w6FqyqYcF%{4!3}gcOLL(AFUuZ%?
z=nJ)I2z{YG4WTbIq#^W$#x#V!&_EVKUwA~q=nGFs7=7Uu4WlpIr(yJkhct}7@R)|t
z7aqvM=!=X<1bvYSiJ&jiq7n2(`ZR*R$dE?R7a7wC`XU2a1bxvFiJ~t$AyM>2TQrKk
zXrD&W7ah_l`l4eRMPGCvi=r<+B60M^CnS!(c#Fo-7w^+J`r<<xM_+tQ<LHYIWO4K*
zMkIm0#DpZ!muS%h`VxJbKwn}=6X;8fX##zTfh>W(<cK8Emz<C!`jRc0L|?K`ljuth
zX%c<OF-@W`IglmMml}~2`ce~;LSL#yQ|L?eX$pO*Ax)t#HKr-_r3SJT`qCqkMqheD
z(&$UKXc~R#K24)9J)~*$rN=alzVtwrMqg$`GU&@pNCthG7R{h9)2A8qWrj3^zRZ|r
z(3cs=GU&^WNEUtB3CW@_+oD<YW&1RXzU+`@(U%?5Ec&toSr&cbh=}MDCqzV_*rFo(
z#6A_#Cl0BIK5<M%^oawRh`!v2<j|L!kR1AQEt*4Lu1|C5%MED`eYr8sp)WU(<<OTO
zkv#hH6Ou<?zD4us%lBy>efc5Hqc1<EdGzH6vOM|<BT_(LVL}S%E3{|<eT6<Ppsz5b
z1@slhw1B?CKvqCsaYTyfD^5reeZ>|nqOaJeMf4Siw1~dqm=@7j9LS33D~(7AeWeL0
zp|8}UCG?g0w1mFWke1L_8q*T`N&{I5edQ4;qpv(6W%QL>w2Z!TpO(>A9?~-U%41qa
zUwI%aqpva|74%gmq=LRmi&oHA>C+1ODnnX9Uu8@y=&KB574%g{q>8@kgjCU2ZP6<F
zs(o5TUv)^U=&O!t6@Ar#th(s)UGwJ5hvQ3rOIQu>d<djJQrz&L@Bf{wsBLDORnxnQ
zx7a(|*XXXjZsG8{Q`QBhw`FGID`)q8-F4W);j@>XedP4b=Q9SatxNu}@%M&>J0!Yn
z_8xfiS=nVuWL1}`kkwtLM%Hwh23gx>T4Y_9>5%nZrbp(wj6*hbnE~0@WkzIEmzj{w
zU1mnMbeRR&+GSQ`TbJ389dpdlgFnA>jyaKCbIgV8o?~ugeva|To;l`0_RcXcvTu(0
zko|Maj~tj|0p#Eu3nB}1Oh69Ju@G{2j)jpUb1Z@!onuktc$dZT>yhZP1ah*=lE|qp
zOChJbERCG$vJ7&z%d*H~mx;)^F3Tb3yDX1f=&}NGvCE3cr7kNWm%FTtT<Nk3a<$8<
zi+q+T`tYhVAARuPUH2U}dn3FMlpUr-R&|&PS>0i3WKD-@khLAAMb>qg4q4w}dStG{
zIAlYI8IX+~W<)l1m<ie3VP<4Yhgp!V9cD$gb(jq~-eGa%M297ilO2{sPIXucIo)At
z<V=TUkh2|@MHV|uM9y_s4msapdE`Qe6_AS^Rzxm!SP8k@VP)h>hgFcP9addrNpt0m
zAHDM7^9Rp}^4SgWmqKzSkV>usQp?pq8o35YE7t<)<T@a|To1&_IUs}F0A!RKflP7}
zkXdd9vdAq!R=E|(Cbt1O<PIRG+zI58yMWwsHxMu9fjn{#kXP;n^2vQbez_kgAP)cq
z<v}1pE&zq(A)v543>1+^fTHp!P+T4dO2`vHNqG_|B~JmR<!PXdJOh-KXMsey2$YlO
zfb#M@P(fY*D$0vMC3y*`EH499<Q1T*y!sESvzZ%jH+=kygD=!3o4pRtSWTuxRyCOl
zS>0r6WKEN4khM*wMb<T$4q4x1dStH2IAlYU8IX-lW<)kMnF-n4WM*VblUb0hO=d;5
zHJJ_B(PR!}XOlURT}|dfb~l+DnQt;4+0$emWN(vsk$p|(L-seBA34xu0pwtl1(Ag&
z6Ocnq7D5g;Sr|FeWD(?OlSPr^O%_K^G+6>U*<?xNRFkEU(@mB}&NNvDIoo7eWU<La
z<Xn^Gkn>HJM=ms30lC;@MdVVGm5|F#Rz|KgSp~V;WYtB!baU`$nu0C4aA~{Vzdk$t
zNZ(aoe(;|kfAzJ84j#^}uTBqN^N#T8Q=hwb?ZV-eN3C2ly?=P;A8$(Nt}LE&@W1~(
zHoJIN!ouh7dQ6F|>M<3vy2sSWnjX_2YkN$Ktm`oyvcAXk$Xt(c$c7#>ARBwkh-~UH
z6SBF-%*d7=vmjf0%!+L5F&naDo;e<Xvx#$_IgwrS%!TZpXKrMEp7F?@dFDa(&NDBv
zZ=U&({qxL^9GGVT<lsCDA`A0OKn~5b5OR2)g^?rkEP@=JXVFETzD|tG4t)|7ZbE63
z*&E@e6J?tykyUM?LRPnl8d=jO8f0yoXpwboqC?iVi5{726AszXCI)0<n;4NzZDK+;
zw}~0q(k2#UYnxb+ZEa#hcC?8D+1VyeWLKNGklk(KM&{duNA|Rd2ie;uUSwaJ_>lc=
z;ztg&NdP(6CP8GOO$6jnn}m?VZ4yR~v`GXx+9pxtc$>tL6K#?}PPR!BIn^d9<aC>)
zkuz<QLC&^G7Flc)5joc;Iplns<dF+)Qa~=YNfEi!CMD!@o0O3&ZBjw5wn=r7?|7qm
zdGWQI+6$-6jiHU3r<cV)yzJq_k{5PE3x|(C;`lYwTeyC7@#lKmmz@(_IBaKj<@CLY
zM=fdTFroX#!r|45)w=2D!uMwR{k09hd1~SCR<?E1^u37}zVVZN>Xpu&3x_4Sv~qg*
zPlm(BhaY+D*A@<+z5478)5Ft`YM;IA&_}f4`&Kcloc`-BoZPut(;_JqO30L)5;B!S
z1(}LdL8exyAyac|$TSKKWExHbnO32NOv`B@(<yY2={Ox^dW9Y`J*S6^Q*e-R90!>}
zVSvoQ86Yz%jF1^QBV;Cp2{IFBg3PQiLuTg8kXaNK$Sj-%GONN0nU%9bW>eT8vvD@a
z90~_y4$c9YQ{jZn$vGi&DO`}bI2UAYg&Q(A=Z1_|@R0Ew51B{dfy~2sAoD7`ka;;T
zWIlxtG9TxI%&+i6=I8v71r!0u0$c#HpdtuakPAX4C<Mp^PJk?=2tgL&LXd?OVaUQ<
z7_x{W0$GHMKo(U*A&YWR$l{7PWN|JISwfM3EWsrpODd9(CAlPIDMbph6qkZ5tw=+b
z=F*U56dA}eTn4hNA`4lT%R(k9M94%=ge<4XL6+lkkmVJ5$nsnsvVx)jS%E7+R#X%r
zD{@81N{SL>C9VWnSy6_p%#|UlC@PRuxC&%dMHR9tSA|SkC9RpBWEalF$4Tp^^oIP0
zyQt&tgKssjQLHgc&%o1o;S+fWe<a5AOg#N*7z-Cx2Y(dC^zgzdS)47vDOpqksYDf!
zT2upRL=BKu)B@>59gtqs192h;WDpHNM$rgl5=}s6(F|k}EkIV$3S<*)Kn~FX<P@Di
zF3|<#7TrL+$OCyq50F>%0{KKAkYDrz1;hYQPz(YIq5u>ULqK6M3=|O~Kv6LY6c^(_
z2{8eb6q7(HF$I(s(?A(91C$lBK%yuD<-{CNUd#g(!~#%JECQ9p5>Qzz169NdP*tq{
zgZ}4{mj_wfTe*eX^Bo=L_!j<;QHME^T^;5^c6XQ?neQ+j+0$VjWN(Lgk$oNJL-u!=
zA34xr0pwtZ1(Agg6Ocn47D5hpSQt6dVG-nLhea28$2k|AdGJE+;vHvSTs!!rbnxN-
zepRMVW@d{2cq*%Xd)K`uf8*`?=EKh@R!ZgS=`-bjgfrE}r>8GIJ!A3d*%wL|Y+Jz1
z=Wda)!<HN_U%Gt7r;a-Ei0IKm;N+v1Mg1#}U3JWH$DeYl)8*zpUj1r~<0PAM!~b`5
p&ALyo-+1Dt&nRY9YAt6lnyvN|ENgY9;OaHu)6S%!xGzFt{|jQR;Mf2F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.3180F054887BBB56.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp.3180F054887BBB56.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2f8932c498f2712eca44b87af074e897bc49ea10
GIT binary patch
literal 2114
zcmZuxYfw|y7Cr<L9w9&sIhY6`f`JeuqPzq#K!9M3fP`mxG=_&IJPZwpz(mCgm+AmE
zID{K0f>3%(iAa$afeKQ5z*wpR6(xc&DB99opePr4=sihi*Zbq{ne)v$-?#Q!XMbzY
z;d410F9QhjPv+)jrsN3qAPB-EPj-$(w5SU~i&zNC`|@^FX1Hpv*-BkkXl<E6R(uVK
z<x6!MMUf7cTp$O=s;UPoLMsT8zIRclwwo1LlmC64bXUG2%e7Y5ROlo_%=={oqoS%8
zBAL%AFEdj7p}s%w)m&Q^S8Yh$Sx@j%>F&DW6Y$S-K_&Y>yEtpr&m%^A{D+&B{b~NY
zSr^<XY2WkZ9u>)_Mo?SMz4^AP#k218zMi_tsf6Wg=@lEst9xp5ZxS67ij~IfbM?{6
z-1OJ+F+z7geY=Fn9x9A^yX8jU09@vNUgQ|um~~-ueZjS0rubQg+fz!)BDM6g$CcIx
z9P;HzFMV4$m0ex1AYE&HAWm$~l-XT)7`{l6ymXYgi&|{;Ts_OKoJ9|Jl=;p-z^F6q
z<<>0en!&l49#Ig=jUjB5?|;5iA^qXWLvv!$_vP(+#a*d~u5yIxu=5An*rFl}*-9}b
z2j{7iw!pBc8=JBHw6vQ&fo~8e=dg^;YRf~yf1C6yig{xf&b%AF1FI@*`{fv8T<_KE
zWB5iVHqHBXlH&&U#@WinSh|y8r7N*FmA|)uU1@U0(zJ`yNPQcNZ4}L#ganl`kEQ+b
zLLS0la*L)FkAEG))nn%eFV`yJgPC`?`!YL7dKBF?qG;s&0a3$@KGl6{#FH3dPHB+a
zcu`KI(rf><Ixw_5K)*R6&f&?)7oBzJ@4w^CKY}0qq{`_oUnxY(Y<E_B2bl(&p-p@V
zpQj%DL>T#UslR4dRDd$gpEPPXPM5cxZ;q|FxRy<;d$@DD4j0RHc%IYJ<b$GJ5iiNT
zaSh|e7RMD4u5y#nQ=LnE_0z^LQrZI;d?G8SbWi=Q`!AzMS59&l%)jmZq(UB%-sT*_
zG$cPBU=Pw#>}4+9BTKKTx>bHwymmIjLDITZ#88KRM3APT?C+MQ-GVcu-d<b7culLd
zUtiyp^(Xxx@Yz<tTdZiRK5?ms+~f6p;ve+EGs<taMIDLUY?leb@0+)F1YEka5k_L>
z(;XrY2w1HAwMbrX$Bz??twY_vY<)pm4vAV^f0deZ6P{@2kMoa?rhj*fR2i5joECY!
zU3HQNF>AaoHpOLeZvUx>h=2Vi>*VtB5Oj6*!kb~|>TtzW)s$PHVzTS_%6{9z1uivN
z;;HzfJ9p1-W5IR<-9tO>4e1i@to+#d&-f7+#?YuZy5nBwaACQm#=v`wu9`@nInnlQ
zGgDyiXZ_iG6-!{%t6NNQNt~@-6-@Y_ZN6eE=zXNq$(u^de_6**bpHHZ*SdF0@A#{}
z{yD4aXE)o_XK0RJebu5;6yoocj1dzucFFAfHgfYqOl=R1(AWVkX_Zf&(QCsn+4-4?
z?m7^JvA$mJf7yeIMMZrC31JLus0HNDpj}35^;2hgZsAx|BqCd4Eb&?ZGXCm0^6}Q+
z2e7Eah)lo`47C7cpY<j6H*@({v8ZT7CS%C@S^#pI^4gdpv}zd#@e!E<Q?_dX$VCM4
z%;vh!s4*0S$P%=~PYXbPy5;nxjEkX`I4B&EZD1R!7Jxi?I^|dOp!paMia=y4Otsbm
zkbf6ayPaIsE?873B3ocA47C8{O|z)>N%5^YEGi6<%`xVNS^%;n<!JW%|8Y;pNb6eJ
zPwVx{nSE5?8bWI<12k|8p*2<k8n}ee8jAq!&tIce-d=vMV=Ro*i*o?2gFPz5uv6(5
z2yMC7(jT+}ULr*_MuH-_ynvq_W84qx8Vf0GN(9I<7E*nwd|;8W(8kXu9Qb+W<n^Nd
z>6tF10VRr(0JeZ>APqF501ZkO11&x8my(fef*An@M^xtZpbuv@a1feoLWaQ+TeGvS
ziH%YM5IRUNhz`!w>&j0?2R6(jG)#k?z?}qNgtXL*7qkSQjnJC00(yG3&{u|UQz1oc
zB5ed<^W1RrjV5VuC_+;ss0m;*m_JfPGkj2F-avT)<<+Kf*Mer@46?v=K(aM=0c1mx
zlnkNf5ZRKX>~u|#pn<<u90(xj;}2o%-=VJ1RsNV<Q*yQ6ATjpixcCIY5h(s|6%u?1
zk9-IdKZHq{$r-80EAMEqt{xhrZ(xWu#+jL#aPfiuc6cK$dy56Zd@E_M=RWWKEFWJc
y(b3D+oo@aASa_SI6?MDK4w}6KW9KgC-7an(6j!p-o&X{obTA3U3Gq9ibLhW>tRY4K

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.CA124F6888DE0DFE.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp.CA124F6888DE0DFE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5e074c84a9cde062c34530787995b7d0f739f37
GIT binary patch
literal 2114
zcmZvceN>WXAIEP|@C8vBavP*TGZZn&@-4|w5iuk+QPE6C_|n9ebZEYfnw37Sj@qyz
zKH}prozFAV%vqbMOLJ!BX|B?iOUm?E*5;g=<>Onm8+NYs{PEo9@I61ipNrppUEk|>
z5DEBvwh06UBnk2}l5?a+5Cq|oPj-%6Hg5<)^H>PV|JQ&xBSO2wYDv=>R$FG06<0&z
z_|aTOP^93KXUKuETJ^1punK~__pSKoCaXd!`Hw54o2n&cE_Jr1!XOD^-6<oO6<0lz
zDSeN!tw?cudSBkIxwM#BwXX2wJj5#$+cZOY@Q>3$CwCq{J7w1=Bu0DvSCEzcNx|DG
zH{4N0@6*L@Ey=e=Qd=&#wo=vXt@&(cx90J9{Nkmwigokl?X|hrh|cjPE#}-)b<r)k
zX|Lj9q@Mo94)IalG#In;!PUTixXkmk%sHeX>&%<AIrqZJk|*gNk0{CWw9<247h3P~
z$>*Zj#`bUuS6w)#SZ%$Vnb4G>bU1S_VxA&@?yU5bHQO7xdzW1}fgZYC<~MT}Gn?+H
zqH+|gCZ}S$WkDzprnEt|>*<zi#di<ySrd!DEpIm}=}g&kkuRMMKfSw+D=W59E|pMn
zaNY)~b4(k}{Ym>zO1rpY`1;JG9FDnFZFy+KuaCWpV_rK%ux>_g#%hb&emcY)HG1*t
z5Wc~MOZOQ_bY90^JyAIy%WyHRbSL(th;|fmD=m)MT6XdqXe+VU2HBKFXizEZQ0ni`
zRH1y9pm?JCm!Ai5b=a9(=W1Kv;EbD_{8*PsMij$UqHOr|Zdv`LG0k&)*qaz>O{rIX
z$fo3_Flzs^+&{RDXWSII-|69zXB`^!x8De79>5QN(B^cNFBPGc_FL3GL6#v_XbV5W
z=PAJ-OQW7I^wn$?^IB2^NF%1B3{~6drr3(JtJ!qTy)6?OT&%$9X-;#aFN%I4b3x^U
zs~;_~Ib0p-uCf?8+OZ&-ebn$payySHB64y{x7S_2^E`TF>4;#?`h3sF6{^UzHrG&=
zDfyRv?k#$<qtdNwc;OYzu*%;~*v@4-$y*nSnX_U02#Qpc<IU34>u|clhrJ<O*tlH#
z?<;SVeTm;md^Z*f=PMf3d0%#uyV+02{=v9)tmW&C;sa4{+Le-sJJzk2d0*aG4=1q-
z7*0{UB^*w{YLxI$)%Rn}4TD`jZFojn3>D9>y-3Nq29LFiMnwll(!RM)stnARPRP7g
zmR(dqtQz*&#{F6RftS^haj#xy9a%gaidL)VUJtpdBdW)%#ytY7A9o&J+GSrfC!i(C
zy{muk%H96!{Sb%#uEEW>2Mq}~mcH-!XWXzGb8sXx`tt3Lp`vnmjfu~FhITA%GOz7?
z6HDUgPd)xl%aPdi7?x1n5~kG4lCgjjO&2UBJr4{zgyRVX&o!b1*U#T}uK6_gjK1jY
zo3>Lwxz;v&jPCr^SIye$BK(b$_lfc8Ta}Kz>$&-%miBvw>0F*$YURTxjHYl*c0oph
zrvU_Es8`AZ&Uw+WDDgf>3S(eXy&(S#`p0awZv2GMBLa(xLa;5y7Oxk8@fS~#54Zl-
zk45c8Fabj_)eFGBYYVeqPZwOoqM{K@#*mHm0&pcgY@Z@UdIJYVBA5zOsd@qU{PExs
z$2_}TI4A<aG?+%y3&2{Gdon~E_|zPVL9h%h^VbW&ze#CbF7C5#SX3B-Z7?>bdI9(k
z$Cq;}j{<+jK_UcGV9F-F0Q|;E-2OQ8`ZN|5j$mtywW(eJmM0&~e)m7=8JOu)EBj%s
zu4Qs34Wx$Hy2yYH(n4%qRKNx)A+|0eU<Zg+>6HVE?+nbPQASZt;ME~s6`8P0=@*D?
zyTdjByee17`_VbZbFZPfl)|M%g5#M>sa$F#2-;jq<I*BQuE&mCDejw?>_j?I#FTjO
zlaup-w%&J2$%sv`BEVqNpa-%`-A#dAf@X-VyCGn=<aeX@X4LbMCS(gT435~EopmYG
zESZPcJR=?hoT<lMl!OkfpFwPx4!eMv1UHQI)?F{?Juy|}r^L5uk<~OY?Eu)p^ib2)
zMny;%VpGLb3D_@~1Eh&=2tbn=la@0mcAG9n8@i1vSOr2rR_kU0tOnzN*t%H&HY86@
zm+B6YEl<o&(-rYL`Rmya3JCh>eHr_AX((+^O-^byX!JKpi2Z1PT)gA}6!$lUB=5@u
z@5_YuWnxBBdJ6JYcre7!2#ql|F~yqWtSl`A_`m=MyqSRefepcW18Il%PM=*IUq2Sn
xnQiaMp#DD>-e_w_+w|dPx}y_w%U0KIZXR9~ce2ZN9uW=-PDJrT{dXJe`4c8T7T5p)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp.5B0B95461D582E5E.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp.5B0B95461D582E5E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d39e9b15d33150190da2ccb56cd9ed251aeeb1f7
GIT binary patch
literal 6846
zcmZvgdw3INqQ)~{N=;L4ZK2j;x0KBNK4WbHt$|WfxdcRM*aDUaq=3rhq#%MqJT4-I
zmEh`%!WOM?M8L8hE#d`r1(p?WT$&zGxt(AnXhj!Kz#xa+6Z$-}eg8Olo_?M8d*8ml
z&Nq{pN$wd_Tf3*&W~&K}SvtRQ@r*2+&6bz>S+uz6{v&pqtsvKCTiWu$<S7eJdio!K
z>AOkB?3PgbE^0U(Jzac~uH7_bJsBUR_Yc^?2&LCAS9a!c6MlJlLf%IkJpU8^@xTOI
zzU@zqUpM8PS~Ig{&dDQO)#F{C4F4{5<>&8DmVdCRkJh*7rRhi85AIv_)wv&1{~iuU
zK3ckCX5?D`Ti!kqd#yBh;qTvED{nY;_KS|s4lb|C;3Au~j)QN1w)(5n*T+AytfnPM
z1-WMnUufUI;gItP^UDKYh<mG_Z#r=Kuyo`JMe4eG&#d=G&Y9TWxNCRlQ%A$Ns*(R#
zI%d~~=-$)Q51gJt*5CW!BOMp>$F~*~?=X)pU-R{y9Yed$J-Tke&CbT3N6%Zi_1qn^
zk3U!C6t<G3owsaWw~D#@)2!A<=f3~uw#MJDFCG5a?=IB47Ju}(lW|wC>HFrBes95^
zQCqV9a(DHB!Nxal{n$CW?N;Y+92@xZhpV1$`Dw(oN%^f!dv?dilq~q<<JPfsPU*eM
z?;iSaH9u_awz}2j>4{6&S=#E~PB_0UdqLY*&+WN#^!wAUskc1wOy$mxU%XnddDZO?
zuN}uY299ca{=^%%ZmxOd=NI$)Z5*kTou3-sHTv(iSNpH*v)5Vq;me13=2+R%zwHV5
z?b}b)kDdG3Yw!H=TKg|ghhKU0$(r>G#&7x~{px^s<BtsPJUp=_f2`Q=-9l;Dt$Vg>
zZCigmaByf=Zn$oc`qhf(-VP3Yt3sOP&aODMv8bVI$@*n4mdRam%}sA5hh7<R@snd;
z9zJ!b;+qu}!BdX<$(x34Yh3X3>-UwM+x=Jiq42?$yl*d^Ene4J_~y=I9jiu+?<#&j
z{9@mIJ5F8N*(^R#v#<Q86(59ZDid3pH;neR?Em=q)~xN9+7B-8sCqT^)GOaFVeenl
zo;_~<!L_Yt9<6VBW6O_pa<>>LdWpWTVEf4G$)lOT;p5SXJ5IguUuO<7{}w(xx7L67
z$>hpssa<tD&w7qu{3&bCC*f;tXO%NKg>%ll*E#vi!!ve1)vqQ0(9b((58d$X!a-N>
zxa>bSHOK$W(9>f#ukF0KQ~VFDvU0_*fBtlFV@K2L_YK}D&gh);`INi&2UoW5%b6w~
za0YVz;9rn-&M7ba$EwG<`#-%jP3||(d}G+gW39`}E_E#^EG%EwZ&mw4)9GK9JiGj^
zxpgtu#q_AQ+N}qAM{9SzcWnD#+{MhD(+=@9uD?b9w_(;y-%Hii&n?}%<-mwP^}XNz
zU@>!f;M&<0gVH5!;--@)#&<McDf!{y7tU8sTmANzmFe5^E?#N-XX(*7o39mLFF5o5
zfyjkB<{clkuc_?9i7y^lPd)kSZqK~ASALi@-P<z#*y=oX?vZD<=FERKF?`#q>y&NL
zviUR9nfG^g={wPym)(PN$)w4)hWzaOobDmxbibUxww_rz?)FK!<erROlwFk9Jy>?x
zYoW)Mjd<|$T=LtDU6Nhm?H(+9$iCd*BlG^exwkElu`7C2RCf=ST{&Rarn1=dSRdQn
z8N0Z!xU74y?BUlgUEF?t+3UHaE@KyF7v9u8SoZfb2EV72U%V}s+?BBlvI}y$2g|<R
zfAX8B7yk8pE}59I`)Bvh=^iY*sd4tAO9OiL-{tJy?{)jHwYJ482lsApX)e>+y^M0Z
zhV>{~qrs!yG<b~0fXBEQ@K}unk9D))aT*66=jOoUH6A?P&4VXs0(gR308i9J@I<!=
zo}@|ONp1-|S(Cw&-7<KJrhuoo74TF|1y6OW;Axr$p61rT(={DD-K~RXXa;zO+W^nh
zOz=#%37)6OGw5IEUxTN_GuTqrIbH7TK9}K;hg=JL4zerkS`I4WT0`|nIz&U14$}~&
zBQ!>t5Cc&r%s`ZhFc@V+EJWEb3sE-0Vw4MU5aq%gM7aouQ9i^&ln?U|<s&>sg^&PI
zAuK>thzJ-JLn1`Qun18xB4ShuNf4F75=5nlgi$#pLsSmS5S1e`MwO5PQ6;QEREa1U
zRYNL7)vyXtHKJlv3uzG5!Wu-ih=x%;q(f8>>k!o=I!29<0Z}7tK-7pB7&SvCM9r`X
zQ8Qv<)MI$6tdqKTgU9O`18SUSvZbukmN>hw&OIl5m8)u+6|#O)&hBfp6|%0;dCu-@
zvo)Q$B1S3KMtny^DH=5sr5My~lwwhHQHn#&M=2h)5TyjvVw4h5OHoQfEk`LCwGyQi
z)M}JcQEO33L#;<C9kmgq4Af?nGJCXjZwfoR?@6#MNy>Ermc^B(Tvt&y6W1DI9g)m3
z#)zvEb%MCQMV%t9%c$da*ExI7-_uEoMjoA}Xynn&ghn16BQ)~p1fh{frwENaI&P<t
z$0R8Rc}$vOkjFF=26;@3Fvw#Pgh3vYA`J4FxSc^Bo1|Fev1y7$9@|V<<gqcrB9BcF
z7I|!nu*hTMb{2VDlH!oZr6~@1Tr=U2$HfSTJT5^v<Z&s&A&-mOIppz4ibo!wrg-G>
z&4foDA0s^S_ypmR$EOI7JU(vcktZZ60eM225|Ae}69IWbj0ngR5=20rkRk%|gt%Qm
zo|vRW<cVoYM4s48MC6GvA|g*r5D|G|iipS)<8~2wQj(I8C#5L~c~UcxkSE26gghxh
zB;-jcA|X$T+a=`5NlHeZoTg;t$<0JYo*W}G^5g`Okte5!j66ARmyxF=DFt~-no^Lb
zG!q4RN{lGTQxZf$o{}O8@|3t;L7tkVROG2?N=2U9OjP8lF`^<*O%N4%YKo}HQ{#3O
zd0LXvkf)_74S8BK(U7Obh=x2ZK{VuPDWV}yi`zBi=}AgQo}Q+3<mt^sN1h%dI`Z@c
z(UGU8h>koxZr71#Bq;-VMw&8^XEYN7c}9#F$TJeeK%S8z2J(!!-9Vn1q)g<QY05;N
z*-T92nK5D_&rA>#d1i{3$TQ=1vxir``Ul(9)Z(wHtOk6x81H^&VWINW(^T(<B2Q6`
z^#EQsA)mYOhpJwgX?KphpY;S1k{;Xm>*=c<{WEt*hpQCsbq*R7<6uCs4i*&W;6U*X
z9+cn^K#2|!l;n^=$qpHm;!r@T4i%K<&_L-99hBiPK$#BnUuwy?QJ+io6%#Z2Up%t7
z&U(~(e8oFY?ECNdyEmpM7dKebuXU|lHYMH?%}kfnmW;EOrETfnUK8hUsm)9m_AM;5
z9!x$gzbDkox1}pHozLVe*2nz)b>8oSr6)elOjl-A4zZT}`12oK5c4b7Wu`q2Prfz%
zt==njQ@@*kpP6<KbeCGw)+g_LXZI&=c=%yly!C(eZlD<&6vHr}ScV0~F&rqK;Xw(E
z07_&;P!c17k{KD4!YH6rMg^rY8YrF7K^cqz%4AGXo&ryywFq5~w)JG#f8IE!3z!1k
zS`+JQhdO8X*A1{XURQO`+PHkK(V+aUyFdk8^`L^TnLUd3($GTtXlSARG_=qG8d~Te
z4K0k9ffmNcKnvq%poIxA(82^6Xkoo9w6H!FT39~|Eo^{=7B<L23+Lsah4XRH!udI9
z;Q}1Aa6t}QcrOnvypM+#-p@k|AK;;d5Ax6=cm-$?d;+uxegRs9fB-E*P=FTED?*Ft
z6QM=)i_jtlL}(F%BD6?e30fqd1TB(Zf)*(tL5mcWphfn|&?5U}Xp#Lgw8#M&TI8S%
zEs9rx7R9GPi{e+HMF}X-q68IaQN1d(s6G{1RKE%>YCwe+HK;<1=GCA@^J&nc`88<K
z0vfbvK@D1TuMRD`Plp!WuS1I-(4j>S>d<0%4QMfZ2DBJ{16quL0WC(*fELqhLW}7$
zp~du@&|(HmXfcB(w0MSi80*@Uxr%ch5tQuFEM;qIoHygfx1fufW-Xd@V`IvVq5EZT
zsPE%AdiG-PBiYmIr%jo9uWg!T*=+San^Su>@9o)~-m^JterVp@j2t(6ygjQ|_DznQ
z+}?fq_v<?*&sW3dIme95zqw#Q(Jh{#l~s3mhmTMPO2b?geChu;TR7;}lEGy|XoeF+
dS<{W%Zg>BtyjU?u4J^EKY>?DWs2)xE{|~Ep;MD*C

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp.144394AC7C436910.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp.144394AC7C436910.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2569757086bb723eea448e9bf7efb8ee62de8439
GIT binary patch
literal 6846
zcmZvg3v^ReqK0#JN=;J)+Csr%r(n+eeU7yWv<6B^<q;62VR%>~kODe9Zi*r(WUhlq
zKnacy6sBl}D*~2rw1^MX;W3P;BS_N$m4}xQ30l#?8!*Ua=7y|wZvVU1&05_(|Nj5}
zUwiiM$vL@ubbbBqGMlX~G<wPWrUz!^*=)9=?AM|Pn(sSex7kVxZMG#XA55CE@Pwz&
zv6sG`Xv}U2weO^c(XmryC+PZ(x2z-MV)XvL+Zmza+NJ8wB5wT8FOM(UyWaC(;U5l+
zw-wv|(DYSv!O7J#TjrcN!qq(5_3^N8)0cnx?nLDW8*kKl4R~q#=j{jgt^6|aefr<S
z;mF=4+h<0u_POcp<MG!jg6IEn_G)G0$*(`_`1Ihink+7|N$WWH_NS}9Jaui{!%OR0
zf>e-uru2pO{p$}ok1#*q|CzX__W9-mmkvuu9#f>QD|gR&Z^WDl?M*v(g+6gKj;$H-
z&n2UGu8-|GHT}S;DP-L}zj?UhLh-oPlCtgQ=gU@qbw|gLuEZm27xZ^F{WNOciY<xT
zXCHg6#wlzeOFD1bw00$P*C%<ckIa4l&8<yOtg9II+uxm^?0R7DqfW+MySmq#kNdqP
zyGL%$`}1A3eFqt5-}<3*RNKwYUpdzEl@HZC)$-%;X%mZEn|JRzGrD}i#~-zhp$jVR
zQGWN}hpYIZYqmD5s?1DS%+AtQ{d)Ylt@#VuzI<-?<<Gx6<(hiaV^3G__~^weC7V{>
z_RyNKjHBPk=I4*Uar36SSAKf2xc7z;%HVTT!#hX)!}e;Q6*um2R)6^NA)Yxpc*$RP
z2mJPJCnt}Y`{`@%JbShM=cmH2Jo0$mx&`AlK1;va_uVrO59&NTp{00?*!$g5Y3R+n
zw`px#emQV(NM2#MVW9fu^5@<T_Is;Jn&r-~I=Nv$W7pz!OJ5u;cgb}(yp<YqdH98o
zkA89Z<e{pw%d3JX9g`<*9J;k>!Bel_Tb|hU7y7~Q!Iq+LE`D9Mwzc%l9Y;G>4j<Q5
z_I~)qUi-G6ytpGO-e0$`^2g;LgzBo3o1^PT`C9gWbZkrBwu|itmvz*<ntt+??-sN7
zt!~dBJOALC*3*wnZhm9)4|Hmm7#Q#peQ(LO5w(*>F@eL!ViUHXeBtk>4>JE2K0LSH
zfB5m#if5>u4LiQ}9J}yi-tLdXSKGc;P8XEUIsIPeq%RK7*zsiVmf}M{?U+4e{Wl8-
zUb+2}KQXnye|E^JF`L$O_U{z`L#wV{{>vXf>1^w0e*NA-JH#2CbN)8vuKmFk?fVL*
zi3gm4g5UcWWSny<OaHm@QSQD^E>4qs&okc`y5VT+vcVU-7L=A&uI;_D{lV$<&x@Z~
zcIVuNxa&e@WLy1~13hE)JKsCH?a%Hq=8kEH_&V2LWB=VaYo_m|+S=!q?Ad%^_#b-R
zXa7wZbE)5&*;NBG<!$1|6UWDOG+i$L{-GDnRZm;>_7~NeTZ=ASZu?io=W{k)ExT57
z`uziu^S94CHgaF{;Pb~nyMG<^_^Z1-^X6Xue&Td*%k-nGirBeFp59V0|C!{lt<PSg
zY>Sr8pP9-2f9F@c6RUgKJ*bdOoMdY(&Mz*=4Ou7ma{khKdd1k=CKi&rv-W`e0Y$mN
zvWH#0cwyVQrLPx~hOAwhUwT7su<Qxy%+!x>IH?<LcV+Fe(z3z1!Lox7^`Ec*-O<(4
z7RcIFJ*sMRgJsVg^@LpT*^1SL<kwldJipwV8!Y>~8H3(aDlgnxNbbzqCHW-<xxuop
z^_leMsfB+zS4bvg?LPT^3UY&GH#f~*bg^&u-(T#^{oQN#Uu|tWQ$46>qs#3YYW)U1
z8)-KUig7cbST_rbb9110HxEj13!p@|2ugBGpk%iUN^vWoRJRIBb8Db<w+_m18=y?L
z3CdIM8DuShwaS&w-0H$f2)Pz^p9ELfwG33mwVLXdbclv19i|~lM`(;PAqJvMn1Lu0
zVKB;uSctM=7NTr~#V8lzAj*X~h;k7QqkM>mC?Dn_%13yN3LycaLRf&P5D_pchD3;p
zVG*KYM8v2Rk{~LDC5TE938QjIhNv8tAu2~?j4B}oqDokSs1i{ys)kgEs$mtPYDC4T
z7SbT9g*Aw35e=hyNQbB%)*-4#bc`Ay1ENOQfT$5MFlvTOh?-#&qGrUzsK@ZsSev?M
zqsQwR4Qi}slBKNcR@j-lUcrV>QLgi_t6dq&bp>^lxK@+yUGIt$S10Nuaeae2O<b2y
zC+x1X_U@mjQxuInIz!ROqoagI9vvq%^5`U?kw>QqjXXMGr;*2`C<b{<hGLM%L<xgD
zCQcaSF-gK8k4X~-c}&92AdgK^Eb`b4#UhW55*B%EoUq7alY~Vcn<gys*o2)$9+#px
z<Z&5_Lmn3;9P+q0;gH8A35Pr`O*rIn2|I^8K1K1!<1-YGJU&Wz<neLBBacrK9(jD4
z@W|s6b{=^`iV~0~WGDf7LX-%|6XHZbo{%I0@`N-IkS8SU0`kNZB_dDEP$KfgC=rn-
z#)*hLF-b(^iD@DtPfXZF<Vh(?LY|bNB;-j^A|X$T6A5`zl1Rvt(nLa@l(0+4lT(z8
zJUK(j$djW)MxGoeGV<gkk&!2-iHtlsVV9Aoq$mY>N`_L9r$mW@JS9#P<S9v_AWumX
z1$j!st{_iMQ7ZD(45cDZjS>}kYMiLZQ<FqRo|+~q^3;S~MV^+TG~{U+N<*F&B^vUy
zIMI-&C5eVSElo7!X$iZAJUvC}$kQ{Fjyyd|bmZxAq9adF5*>MZn&`;W6LuYWMv5|!
zXJjY?c}A2N$TQ-^K%S8#2J(zFF_32@><0486lEgM%upus%qTIDXU2(%JTpm5<e6z=
zBF{|N&2C=ps_$)A(hqz^<u%s0YNlE5mgC${FDzA_e2VJXIKVTY&bnK)?(U16xx0FJ
z6OB=>4e+K($0!;#6QdZ^Y>Z-2b1{lT&BrJnwGg8O)MAVhQA;sOLM_KA8MP9l6x3>r
zQc-I$N<*#3C>^yCqYTt$j551*!}wzE{O@aeWY3nnz}?$AiPm?@nS02w&X#p6K%BXU
z8*94f|5<>L^xF-;oVwD{C;M39a8+18lYK04(4ZIx1B!L9pg0Eyig)m!1cv}hbcmoN
zhXhJ?$e<L50!nqLpfraDN_Xg>42J>AbeR9E7LOhIH)(R!gzUL18&TF^?XX9ezw_9>
z|2*^V_324vjn?#QT`QJOInxr$PM6o0kF`!q+mby!Cd}VlpPer4Ra$D@>wQ>xcc_PN
zb60k{m?>7Q2jBe--fx2y$3MzWSLapVVlC6L=l7l$i>udWr#%i&u{HgT-XncO@BU9@
zr``SB71p%%Eb4ORo<X};(Z#rU>-Wkop5`*U70uA#(HadN!!Y168Ur57u;8&83m(UC
z;BguU9?$UL@fr`FzzE<8ngE{2h~SBu2%f}9;7OVUp3KPL$(jtF!YJS=ngX86sNkuZ
z3ZBMj;Axr$p3dmt>6#9n!5H8fngO24nBbY337)6KQ)(@8b|(z*47602qir1-`lss`
z#S*4Ow|0Q_u|tD1_i+R41h1>Md+l95*C<ea*PWmOuF0T+u9@A6_R`Qo`)FvP{WP@D
z0UBE9APp^ymw^_>$3P3?XP|`%FwnvT8E9d>EVQsb7Ft+83oUGbg%&o*LJQ~RpoQ~s
z(8BpSXyF1Jv~WQVT6ixHExeD17T(W83m@R2g%9%3B6tO85qtu)2z~)tgn$4oLQsGf
z(JMlW=o6tu^o!6U21IBPgCev@UI|(xp9C$EUxF4XAVG^1l%PfS%FrVFWN4B7GPKA6
z8Cv9^3@wURffmK5K#Ss6phXEN(4qttXi>c?w5UE6T2#LZEowl87B#3si{{m!Me}LU
zqWLvw(E=K@Xh98HbgvFAx=)7|-LFH79?+ph59-ijcnxSVd<L`_egj&JfB`K=(0~@x
zYeI|ZGoi)wo6uqgOlUEKCbW2N@i5l4DSM^jJR&IBqgl$<+%#{-^-n<;HP2c!@%q7(
z>qG8kZ!Fv+?X7CR?aBP<lc!CYdXH_IW!Y?#yAP*!AKue_IKBID*8I@Cxmh`O_BeZ9
zkNg`P1%*9t?9;o~=ptVoTjU%)qPTxa-vKvyhE&(w?j1H<?I#U&Rq+-7-)!l?o683c
jzJ+EuL6kM!xb-&ouPVzFb7a5LJH`Y_{rK8Zl>dJK<2Gpi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp.9C26807545ECBC7F.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp.9C26807545ECBC7F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d101717ed41ac6ea03e77a507caae8b1ce6707ba
GIT binary patch
literal 6846
zcmZvgdvsG(qQ-M}N=;h?+Cr_xS}^DRKF8VwS_7q|@(761AP-9fQb32tq#%Mo<~oQ3
zl$6mCg(+I$ihyN#wTKVY;W3P;BS_N=Dz6Dff>v~J0tAtn8@kq+et+Dq)xFNQ_b=bC
z>DnhJ=iEQCwsv=k%~lf}xp;2F!f9DHn=L={yI^7C!$<5kTVbBfwz&Dj36tiZaQ8X(
z@^|Bnna#nrozxIIdaC3EUAys)CNd^U@9(>v5z4M#s_e+;#$I`4Z2sQ$?tg}UJTTT)
zVEa?U*Nr(RS5I%Ab>awD^+e|<L%vI0{`vb8<sWXmS?g8&^3*Td4(?m|)!83X{}l>_
z_b%Q(J$$Xtt?wL<y<QeL|Br93mDis<^LhJc2bWc4bm2`}`@wfUTlLkc>th~YQqvru
z0^D;&FShMpe<=3|bLElG#XZ$8G#<EgSUU2gB6VK9f5!X6W{qoW*tskCsiS^$)v$jp
z9=UUUbkC`&2Tn~QO%ML&@%9S^V_FJJwwqrpTmAJt?SnecE?qnCmfVJ)N6cBV<?P)v
zk3C<ND{LW)J8s>yb|rJ)r&%paXMgb4)`q8>%7*;z_va@%7w&x`mvL3E?)BDFK2PE9
z;hVGmd|!3n0me6P|JX62^|svG9qak>$Ewyf|1@;U_=1+k-Miu=OXq#^amy$=r|d!H
z_m6(GiXXgYYu&2y^teUr3~ki|W6y2Pp4a--^Sdv9@%<_1<XfM7wsObEFI_F%wDPXU
z){JHx{f0NbaQw~NHr2fP^GgN2Hw;q-o|_!nIpQC-*ZQouc~5TTN3R^>nWF<2|8=+D
zXWw>m;;7l5z5ebWueDuS7kYK+Q#DQV#%%l}{aWAm;*SsLI6SVoV3gSVy&`GwZM(N=
zty_LMaBxsoUZ}3W`qlF1-wE`4yF!}b%C0!Mp}4+tQPYx_2Fjgs%}sA72VEX|;gh3Z
z9zJ=f;+y3afs>Aj6E+Us+Awe38xNJ9-Srpx(a^!>{BJLwDOuZ6^wy4}?JI|l=`8sm
z^ir>V+fQEH5fLA$*;oG4@(+VGm5I%f^&`B^`#(OmC2QNowu8&st6obz^Xm7D*oRlQ
zWsjbFa81kUr4t+9-25Y*+$H*pU#1@_+%~Lw!U)EH_*it@_LDFE@9BfgZ$pP?*ZK}W
zm0a;0wX<%=8TYXZKV|LyBy_FyjB+}sXx8cXJ0^U2c-oF<dN&sw`gzC9LF>Pr-~Z~}
zmwaa@=lH%EbZXS5H66Efi2tosRxbZ#{p$-G+8f_^Xuu9}TF0!vO}cM?U`5-$oGIdg
zTz}3VeDl(|v&xJ9wekt>;ZHA4k$cZE-yFQ*Xv?yJ7dz(_6_u~;y|V4msq~da&n>%m
zc3sSQAw9gccFTdD(b}EwAKmt6R|#{^ltX-t^RLlg>t{^&zFb}X{Ng>E4-EZNuZQiw
zDPb=4TQjqwf4a0)+<4;nnD&Otr9V9O;<?HxtKRvtGJR+Mh0Cq~F8gBErfVhF3r~M=
zAbkGrImd?YYaDp~_~(x_QBS?L%ROiI<sZgR^)ydCx+<TYedO6KIdh*&4B6Urow6-h
zGIx49bAD%+y&J80#Wf(0jGthuFUT&)=^8RCT|c?Mw47cs`mXVL<o?WZadvTj*I+FV
zxpwiwwsT9~$Rl-`<)ZAOo4N*Txqh90)$LC{u=i%$eVOHwqLP7KgSGs0&y`ny3chFL
zkq0u%rP-yPuEAQ~>;LS`l?Q$u)YInAELZfXsO}o9<?p8rcwZ^MaAzL5H?v%rU6|7~
zSj*S@OnB?m{J)&bBjYm5eX{%HbPd*WW5dh^7yEX<|MJ|f_q}%iwU*X+<$#{`PR7Yw
ze_$qQP7{*LIk-E~3=Jf$(LmBJ8c2p=fMhfVNXEqg$ucaEti}S#x>z7Nh69q*I3PI}
z2PDt%K=K+7B=6#Z6c_<WK@)%!Tmq0HBLXREB9Nj>1X5xoASF!#QgTT^%8U%8tjR#i
zE*VILQGir51xUrE0I4!6kgBEvsk&4kHAVwc(=;G8mj<NH=s@b44y5kVfixHcNJBG#
zG+YLdCSw9=Y9^4T%LLM0=q~!t8B*^qcK3%=>K<Sv>kQ6p@EfNuZ1A9Se)lGJhMdbF
zg`KOZ?nnn|K<N+-C>^E|Wr7T#Oo#!L2{VYYK^9Op!~)8OSwy)Y2PhZf0Oi6QqI{4C
zln?QM@?joPAt(SUgakl^uz;u-6af`OBA{YeL{th&fJz|=P$?`SDhFji<&X@h9F`GP
zf(oEYNC8v{D~PH=6;L&#0;+~pM75v>s20)y)xsL0dQb;c59xsFVI5H;XaH)23_y*r
zfv6cY0X0J=pk~-a)NQz{tWDjs-tBRZgf!Yc!AjOODU{oFErJc7q@3qrS3A>`^D5>D
zajquayWSZi&JN58;`|nKia0M}j@zAQ?cJ}ZlN60UI!)2&qa%by9~~n!`sf6q(MP8U
zjXpYVr_skGDF%H^nqtt$L<oaECPo<aF$uz;k4X^*eN5cWppQ*bEc)0q#iEam5EgxG
zjIiiq6NE({n<6aw*tnfVAD5&!^l@p5Lmw9*9QwEz;n2q=2!}o{ML6_vaXW`TK1uQD
z<I@z6K0ZQt^zkvmqmNGz9({a@@aW^?b{>5~k`mA-q$vS?LWBtD6JkU_pO7E|`h*k_
z&?m(00{X-xC8AGEQzH7r2ocdI#)ybMF+oK1i76tYPmJ3|^hrrdLZ6hTB=kuUBB4);
z5ea=#f=K9-Qba<Z6t_$0larK;J~>Uv=#wKvMxPubGWz5Mk<lloh>SitZkN%gBq;@b
zN}5v8r$mT?J|#vJ^eG9VpifB=1$|1~uAom%QY!k?G^L_XjSv-mYK*ApQxil*pPC{n
z`qa2xMW2?WH1ug{N<*I(AsYI$7}3zDC5VPTEk!i+X>q%TK0Qh4=+o1bjy^p?boA*l
zqN7hw5FLGbis<Om<8~c=Mv^knXQU|ueMW>B=rdx(K%bEy2KtN?G0<nk?FRbHBxRz{
zOj9QM%m^{jXU2$$J~Kf~^qDDQqR))m&2C@ysvm4uQwzVQvg)gxRa301<rvqq^NW;c
z)=@p{i`~UF*0;sFu?5`uKUDR|OuKSiy)D|hq07(hx{-rxYm{<sfNLoorD)7dlwvTm
zQHsUPMJWz5AEkKALX;9Pi&09%EJY~^vmB*l%u19}Fso5Y#jHgs4YMAlbj(JSGBBG_
z%IwZT>317`Id!$YPv%C^;Vgr%01g@w<6t1M4i*yU;2`l19+Kb?Ac+nUlH`ye$qpHk
z;!q%|4i%E-&>-m!9g^WNAej#HKWWkE;eV4RR*cJRi;`g_b=G!xV)?sI?)!WEy&KaL
zO6sj?>;FJ<yYB0|cMER7*E?4%nG|o1W?oxbTRPf$P3z)4J;u%5T$`CL>Qz)^eX0AX
z{Qh7M@8-_TbOBSKSWhpk{Ym9^JtDA7`|CX41<H<poS`eTD(|qUV=wGIFBVj;%}l!;
z?gDH2TfIl>rrx(aotbv^bCp@s)^mnT_l<`PR(GePwTTS=pBpDmAycSZl=ax5F1PD(
zgGFaHhR0cLy}3;CI!8eAIq!w!cTR*9a8B<|w1<Wk+Dk(V?W3WE_S4Wp2WV(vJPfoj
zUItni9|J9npMe%8z(5P@VWEZfve3f%SZHDWEVQrz7FswD2Q8eJgBH%mK?~>SpoI%?
z(87CoXyLs)wD3M2T6jMXEqs877QrJxi{KTYMeqsGBKQSp5ds3Vh#nDIM6U=fqECbt
z(Jw-a7!aXF@<`Akc_nC(d=j)sehFHnfCMeFM}`*JD?^Lylc7cS%g`bRWN1-53bZI*
z1zHrJ0xgPPffgm8K#S^8p+)to(4zWOXi@ztw5S0US~QOaEt*$@7R{$Yi{{s$MGI)q
zqI+~`(Y-pf=sq1<biWQQdO(L3!(%{;;WePe@EOoz_zh?=0tU309urzjuL&)t&x97!
zZ$gV1FrmeLhnulJvNE4*oLht>yEQA>8XM+JyYUosLF0@C<8K^Hx-oS9*z12?wmCo8
zRK7HO>clCNCO>GKVlCNh6T1&5cOO33eK@uIaK_x=oY|Q;dgd5=R*&qP965PCZ|>8(
z*T{Tt4V#}ka#+DFg?)=}bq}hny4y2ksM=2&?5yC+{=eCx{<oD57<dQGaDpgnx^d@S
YuG`B?6mxjLqI*UKNbT6_5tQ%$0Lwn#nE(I)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp.48B6F9C0E69BB63A.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp.48B6F9C0E69BB63A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..11f9ac7d2dd9f59d49ecc232f5ea6631d4519716
GIT binary patch
literal 5954
zcmZvgdwh%c9>??iQkABtwnUY#<#~S3{W8-?ZwpoF(9~R-u2jV(r8bwPj2g0IZe<fT
zV_4B~=31|vsbTDbuEQ=aO%F3ICxn=?jT3|(J16pbJ?r<6)7PuNKA+#`lg~T9<d-MU
z(~(1Rb2GYWG&znT71K*+Pl(WHG_j%IS+mQhp4DnJ@zEMhMb%e@qi0@D>wIDBucP`*
zu5#4wBl=VB=ETcX?xx<WQNEix(PcO0lAk`vY>Z`wJ^FZ9?6LJ}|91XXH%t?!`Ka`l
z@~Ea26RW0NKFee+YB}Bi*TBR3w=buDwW*`hDPillbG7xymtMMdC-66?({-$3_e57~
z=kA9tdiEyUZ~bw-HMOMa>Q4<n)Gx>i;ar=QhWbN4EWgzJH2=+cIaPMT&a5?Ts6Da%
zl>RJ!^y*Li(d-Y)>mHmD&MuLJmd7I}9T_-fcx~yv1CH-?C3#r`|Ew6YZ@v3y^SHX^
z(P;J9*WPTn8<$@lpSb&la|>4d@=`;3%eDEdW^~h+-XA<|@%C#kPQI`%OV4db6^-3D
zuUd*nd>>IgzwFEXJ4)YKo!tMew{I1ZvyUy(V^j8uPWzWx&G82ZZHxG3M0S^yKG#40
zt#NQoPyKVc^=#_gtTk1?4=5fLS6zPafOkmJjMFEphf-0=W2Lv>_<A{;v2sVj^3>q)
zIrJoD`76V2?ueXGb7|edhv#lLlViFsc|UW{$t{oLH!n5LU73e<T?dtac=5BIn{z(7
zza=ha<3Op`%`wh>ga6R%?7X<+QGMpuAD?3J`Cb*@9kf}syPArImi@5z@CU86kJdOp
znZGP&^^E*YA5c5HeBpgFrSZ)0s<@$i%oheBqvye0O3n5sb@k~H(awS%@}-6A4%xeY
z-cOihitN|4F`=Yo&gywvdWkJ!PW#XO=?@3oJ$?S?Gfk)ZU0>MG-lQuk+?27SbjF%b
z$0uDo@GbR*v%V_!#=WbFtEvtA_ndE7Iv~F#@k{5HPRDmQ-P=>iznXJA_4kEeIdU?6
z+bY)&wp5)sd0~6Ru6woh3mUR^2HyST_8fZZirUD$>GdnCugoth|7_cDl>Y#4OV~<{
zkKZ*gyKpeJow?v1zPo9|Kd;o|*PUm|a;;~U`4_Jx_7&{8ns(vt?-2)2J6mh6N>`!`
zQ?49oEd2S*ggx)ZRK=aTzh`p#`WrKQJbv+k_1c&y>-F^Jp_^AWc5CGSu4HB|e6oIT
zV@*T(r{h!h@Dm!R{5X2V3H##O<59(Yo!%Dpo^?i0KPA=h=h8*Y)bH;Vi!sw)_$*`N
z`RWC|?zPM?7*bcoEUkTG9QA0<+66C{6?n+I!9g{-+v_^GbN3xNzv~-QB7Uj(6q`eS
z=l)O0q=}ZT+1cwVj&7?P@KL9!+Sd~CgRUzl_v;Z%s^K?XzL?)o`Y`Fv+zmG~i<ckz
zIWyQd_U^-)|0bWCvbi<!Y5bKh>s+^9oOWT*@$z1`F8=iDYGT>W18LLB9^M%>&RjL_
z{PI}3?Cks7qo%L*_204ODWRD)Z~DYw=<^+!eAu1yu_+}QjVjcX#6`wMg@=$6{?R|F
zzOp#aI4T;A4A}{h39;cpwfnc;ySwY=yicQ1LC7{l8rp{k)gJ$P*+)4ov1|v8EoAp?
z*Ec&nsP?|+T1R|v+3D`684<D*4T-(NgK8hP_Fg)^Y}L+a^h(H1icB(x2i3kkA?1ja
zdbe*hdO2jrN5)5m2i1Psxp05;%x`Z-qv0XDb7beJ@Sxh|rITmP37z27^(X1W>t1!D
zwYtWenbM(z#3ZXegOD<j8EuNf6!0h$1w4!~@URI3kH$3cXcG-Q1~b58ObqZ?%mR-!
zvB2Xn2RzQi0guN#@OTpsJOK;96HEf|L@WYNG>O2Iumn8GBmqyxGVo-R3_JxZz*9^L
z@Qm?B!(V%)#F%7EQI)#?1buklfw#v&&TM-_NGG`fC>ObcXp@wK0#V9IfhgsoU=%wr
zh+-!OQS8DnN;_x}rJXd0(k>cC83zNRjFSOT#>K!W>tI2Yb+RDJx>y+H92|&pP7Xvl
z7YC!fg9lOG$%82G;$c*92p}pr1rQZn0*s0d5ky6&2%@4(gi*;MfvDt^KvZ%`Fe*D_
z5S5)Wh{`S*MiqwwqKZ=iQN^XesIiYROWji)N{nXX5TNpmg{o3dXk+!^6B*cVeuBIO
zHWe8p$j4AuB60<4+jOJ{k&RIM5P1XY03shi?bVXkv~7>4`~(GgRFI$`kE%oz<WU|(
zK_2Bp6y#9>L_r?q)l!g${RD<Q93(L0;Yx%d4|@=XJnTal@^AoQ$irSOhCJF&(2z$5
z2^#X~N<>2*?Ljo;(LO{&9vwh5<k4O&4S9^8U?7hP5)9-qm56~n#)BBhV|<8#JSKn`
z$YZ=(2J%=x!9pGzBv{B}D-jEMtOv1>$NCTpd29f&kjHwpEaY*1f`dFRNN|wHRU!`Z
zI1l0=kMki8^0)xvAdmBEImqMv1P^(9kl-PYuS7iL@gBrO9`8du<naN-Lmuzd@{lL^
z2?6qiAR$1WP>BS{6Ff+OJi&(q$P)rcfIPvg6(CRa6C&h^K|+K)u@Z@pCwh<wd7=-A
zkS7L^2zjDcD?*;+CnU&|f`kNlQYDfgPx2rM@+2RUAWsS)3GyVbR)Rd)Psorb2MHPS
z<Vqw%p6o#~<jFoHL!KN!GUUl#tqggJpHLuA2@(qADV0cpJjH_)$WwesfjlLE6v$J&
zTBVJbz5I^mabWf@L_|qCnO>}(-SbWF&ooHyt|2;<B%~$es23Ln!{V4*?`5?MO{Yes
z#;9raFQE@#wW{9*4&kkhPnsV$bPQb;>d0hp%GOapVI2mF*3m#QItD0K#{$LaIG}hP
z50s!2fD&~gP?AmpO4i9hDLUmZH79S-k3vzu;i1JQ4ooaiufZ2BJiO%izr0^On=VW&
zQK$E|ES@*oTjdT-C*>yPsf(?tINEOb^liDJX+tN2LH*WLAJs`8zK~a6vCs?UCdiH8
z+mdn<6x7&FV5n&~K|{^B2?lD`O|VdNZi0iFcN09+f}0Sa7Ttshwd5uwsAV@HL#?<8
zrA@1^0@8<XNZOVO_uAK~BOUE5+gd^|Q5=qw)ccGR1?FGv$rn$Crqd$QdaHZv!iUFh
z@o|}}LeoZ_F;1Pn@qD{L`<QO;gr-xwrY5V?>a9<VK77m5_5?_o?4v%r4keUK0fl7@
z6fM(0F){-bE3-gxG6xhd^FRr*0F)?;KuNL$lq}0YDY62TF~QhFU0RE-W;M$A$Fq-U
zJdS@}U8#CEXVHi6;=qc{WOmz%Neej`C@c9gP&Tp%C_6c^O;KhFv``ibv`|(Gv`{t*
zv`}^mv|uv^E!cuV3$|j=f^8VIU^@maw3!Aiw1oyOw3P-ew2cNWw4DYmjF|x~jD-O$
zjFkZ`jEw;;jGX~3teFKZtc3+Ftd#{Vtc?XNtepidoS6eHoP`4|oRtGDoQ(r5oSg$L
zyqO0ryoCoXyp;znyp0DfyqyOvf>{79f<*u=f>i)5f=vJ|f?WVDqFDqjqD2HPqE!Sf
zqD=%XqFn?nl34;Rl0^b7l2rmNl1&0Fl3fBVvRMW#vPA|hvQ-8xvP}jpvRwu(idg|I
zibVk}id6wEicJA6id_LM#@<G(ev(6HGRDXQB^nh~m6uMN@a#V^v&tvU8ujd8^s^!S
zqb*snBxUY~M>n=ajw>o2J!Y(?ShX~oqPD{^ZHHsq4#%|}PMYqRRu+<Zlk>F^?IPRj
zqM|!=>>Sf+NUSA?j@1tt7}qVnOG5Xw^vtXm&HV?+U4;y?ADjIDW(_@hCZ+W1O<{)P
eMdkTEeT}B)QWK>Y26Z*OG}MlAhh+~Ytp5Y3SD|eH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp.FF4791DC4DC38C18.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp.FF4791DC4DC38C18.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4d96fd6b5c46a3c7b4573e20298ac3fd20b85049
GIT binary patch
literal 6970
zcmZvgdvsG}qQ!H*KuKG$Ewn%>P|%+D`+<Q#nOdO+DmBnj!>c?3MxFwV3gXCQbS`Kp
z8XiI!R35_x1c3oW9$v;l@8B?qgRp|I6e$=2MZ79#tnqO}?>cn<akEyxb@sRS?yo)h
z&U4hTiVCC1;dpTBum!WHg(kOiI2?KDpZd`J>F1mdM{c&mv0(jw)sB1QO8@**Z=BzH
zdON>Is@|UCyU;kJ!?x}fgDQ6Yeue+ZuXf(tA6ooN??m~=*CtFqk;VUIaN^dXiRI)K
z$EVVnS1T`7<zE@J=+|eDbvfAhb7R&2?ArCW5od0k|7Oa|^IHDnb@>P7WY^%<Mav?$
zO8+sg;MS4jA09l?SohnH<x7gc*n0hP({r!Z$ycxR)jzp?<JqWkbVBK=uZ!oL*)i_&
z&mDjCH12<IeZhetwF{I(kIt+;-|6tGqFwLQEo5em8T0X)9%l+X&##kbZlS9yu8qB%
z_ucWz8E=O^oBec;uRrd*;PTirf1b1Eaq`-~y{p4lE6)yIacWGr#pkXa*yv+k{-Azt
z&AVGq9$xWK-)S|Dmpz;J{dG^t@J2@{r{U4q{#oB%Ka;3jy1ZZcqMqA)^Si9N)@SC2
zUy73}ONM__R4|0<Y}VbVYWU#9#!KNzFP7IIcmA{Gp8fr3N(fdK)tzlJTe{c0Skm<T
zj-h)_C3A;9bb8f{H!_M(oSOPoeD0;Gb7oxn$GX|mXXd<hvDc>64{(PzKl9bv#hdr-
zxmq=Ed%KN)^nU%F=he;i+iEUWU+!DIjs9S~xT_%Z>!$oEdnS*YeB7O1xTVC?Zuh4L
zUYhP$T~wJldtt$v`A`4Q+t?NQc<!~+Zw*_y;>%U{y2t0daG~&}>hfXnzUQV4UGjd{
z@Bj2<&EkXomrUxre8&6qkq!$}V@A&0XFL^{^!{rvo_>2ya&4mU*xpk{$hYYAK+f6M
zdp6uW*)Yqz7`*jS!yl$?*&Y~sHKQdf*0-wi;ZfZlJy|%i_it~WcFpL!qlY}fb$0)*
zzZWI0=dM4|YxhgrHr=29@U0^|*35c*$@u)!owqNo{_*THs~r8FJFxbVb;(s1|623;
zsS<f!Y|{3({CQ^=<y0!QXJ7kreeIU4D=XX0BBQwvv$p@`#Q6((!s?=Z7mgfkr<^MM
zXz4%CfAj2IW5ep<-xnP}{7dw``uU#?zVUSA=|BE=<2%aR1qbTa3Rj$4+wFb#%`wZT
zwfw8%xe=u+57v@13##7OQZ=+Wdq%-0Yeuh87w<48H17Y5@4ocS;oZM4tgl`5P3qFD
zxyN2Tv8Q}=dCAIgM~@XY{AqFE($IDDfA0M+=bs<mS@oF3-EK;i_n32dp8wxvgS(!J
zEsPHObi{Y18>f%>?ftXUl085D_C<d~MXIo8!`KO%M=pNC8os>xjgik(EbG-WGB9Q6
z#~U^n?@m1Z+^P|a_K)m0VZ~2bL(R3Fex_;*I^CSSBsBdON2=F<w!~shr9b@e>&m51
zEsFi^k;IPIONacH^Tgmu18(=ef8eG`hjMRE-*+MKhbDe@r=oRqYQ>r_s&-$W{6gl~
zy<a|FzN4h2Ltw|RCmr<*XHV%yIUE_q2bVv%(cLGT)I92_>yptWvvo+H0a`y<KW(`9
zd}Y7tY*L-J3o{DyS_j+yUFzGL+paF$olQok?QR*}I<yY9J^j%7FN*8yf9m8Im9~qz
z7WHi%Y<tg&%y%-1b~k2|k!d?GBhS@3*!JFiy$|w7N4}d){As&)M(=^GgKZzph)$_L
zIr(5mM{U~f-@gCg*1@)$C-*rlTQ|*Y@<`gw&&bbg9c=rT+}gdLKk>oUY~oAXxf!{c
zt%Gel=1+U9zM`OQ|J;|=x?dWOr8ewr8rY{}9j(y1{THT{oA$IRMq$8X+zfcE!h*-T
zS@1Z81CMia;PDC%9`EMC6BGeF!7YF%Dk6BITLe#1B=97+1fHzO;K^<oJVjB!Q``!8
zs-l9Yx>fKrMFUTBYvAdM4xaAT!7~&CJi~2(XDTLmrrQM1QY`Q+w*{W3yQk>x5~}m`
z^7OHlz2>c~*0l~_d61sp_9fFHdNHUlMK7b;G!tYX%Y+!nGGU6rEE{AY%Z6CUvSEtF
zEEnV;%Y`_|a$$<YEFa_{%ZGT#@?na{tPm6+D})5d3SmmXtQZs_D~3eKieXB`tQ3?W
zD}^M;N?}UEtQ?dfD~Dvr%3(^ztP)fptArHDDq%{&tQu4ytA<p_s$ojStQOQDtA#Yk
zYGF#ltRB=MtA}*R>S0R9tPwOIYlIBQ8ez)7tQj;RYlck7nqkVstQE8%YlSSxT4Bn<
ztVi(-uorp9I?q5)1t_nl+E(^ieO^}U*&GglW{SQ6i=0kT^v|dxgkDD44gxwx=yRwO
zguaG4N$4L@$DQ<5XIr0XrWoWgDT+ZJ6Cn)pm>6M@$0P`YJSIsP<S}t4gFLpGVv)zD
zC>D8ags{kCV}wN>n;<Om*d$?*$Htv3^0;P-Lmro+IOK5=!Xb}~5e|7=f^f*=l7vGZ
z7k6^V<C`fSd3=iEk;g{}k32p`c;xX3!XuAQ5*~Sc+{q(PXr=_@2`NfIo)94d@`M->
zkS8RFfIJ~d1mp>Er+_@MnG%sFrYI44VuXmu6Jta~o|qsa^28(&ktfETBJ!kWN<yBL
zq9o)=5h5W^iV+EUQi4dxlafS2o)mXV$dj8X8F_Mwl94Azh>ScrMr7p42_hp;P7)b;
za@;8+PidwU<S8jiL7oyJ3i6Z~QIMx3h=M#NNfhKMai@YjwV6_pr=}<sd1{2H$WvoP
zMV^`<D)Q7MQIV&{ohtIQW=cbzmZCJ|X%V6!Pm2)^d0K*K$kUQUL!K6QYRJ=@DIIxw
ziqes%M~IF*Jw|lo=?S7EPfrpZd3xNbBhP524CEOp%0QkGAqMh{7%`A%B#41LBS{S8
z8F8n9JhPcHk!PkT6M1HYn8-6@#6+H%ASUw6Br%a^#+@eetY*qWo|U33<XI77A<v2t
z3wc(8Sje-I#6q4GcUo<{2bO&0Xt^D_K((tYy|2`7UoA(tS3J>GUiJ#rv97Rx;e+<w
zV@*|O_F8mM`}FjH%mF!eI`G{mBcALs(M(Tgb;!!MpEst8iyMFXyycVJ^rM7}E{02>
zivh*DSWujc1I4>|P=ZSUCAvgVl1l<5yJS#`O97?2R8X2r1EsrkP=?C@Wx7mImdm=U
z>_@`9tky?CxVMN>^m@3LV4@U*nvGH{YA#A~sQD<xqZXo+fLe@FB5Em0NvP#0C8JiN
zl!97~QYvaKN@=L|D5aw|qLhK!j8Z0QD@s{y`l-qxCq@7LzVxyvDle+B_fq)x2VeO6
zrKUr7rfZ8P*wZ_1Jil;E)B5G<>Eg`dd+gP+bHS1JzS)~9($n4U>DJZ0jk6!Khh()r
zUc)Xgqs#4Xdfj>H;h=o<C{O`<TAO0Z80axx26~K-fgTfJpvRW6&||$U^jIGYJvP8X
zk1OM#$9XyEaXt=uT!4cfU&cd^_wvx=eLVE|01rK(On{!?6`&{h1n3C?0eWJY2tCm&
zLQnLG&=Uh9^rSKgdXiUyp5&9DCj}(v$z?M1WUmZ8*(XC!4#?0`$`t4+UIluJPl28i
zP@t!lsnAosD)dyJ3OzNTLQgBxpr?5?=xII;dRjn(o?fOyPxtE3(|tPh^neaMqs)Mw
z;WeOV_zdV70RwtwnF&49YeLWTnb0!>CiJW_3woB<f}Z8Gpl1av=y~q-l-SpW^z#_w
z;Xnx<$yWA}N3vR9fVCaKcp2KyW{!U4qm9?o2Y+72yj=UBw|@tf)%teK?jE&9O>IAP
z(~Wex3*AMw-vS=18GL<W@wwya=|SxV-D{U)Yu>vibRO7{p7yvro$cwk*8X;foNkfy
z^nii^#rCxQ=B;y9>l-&4nLfkNG-v-W9qSmH0mafRD30bp@iY%gpaoDOErOD036xCB
zpcGmGrP3-Wjn+Wvv<}Lk4NxX+g0g4}l&6cQtGyF$xOTowJbUik739<Tn!Qi#*N*AS
z-<@}kuy6eIIM_FS1{CXOL2-T#6z}Ij34Q^T=odjrehHN9mq96h1(fPnL1}&sl<wC-
z8GZwl={G@He(SE9KW*0JJGVOZ^8@wOcQ(e}8UF7x-rjIH#<XpYySwSE%RjKS_GW3x
zD;X2Vj2}C0f@8v6&oQxWGcbGVtjE)`^05)lcI`7dxH7XlcFN7UXIS3Q2bsLAVdb5>
zb;<8u*nhylL8XI#SEdw*p8M|SivPdau0=h2_33+W3C;3?q-wg^&)wVVAsa&qx;`{~
M5~-+qz)L;+zr|0~2mk;8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp.6D261DF3AD15EA25.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp.6D261DF3AD15EA25.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c7d5f8dbae2f450cd2acf7650ab2c8f929150c89
GIT binary patch
literal 6970
zcmZvgd3aN0qQ-N+KuKG$Eo~{KKtVbCzF;6MQ!A7}r3PAR*vb+xvI=cf5Jx7XGoYbp
zSY#PgmSF%vU<8rH>nL&uhd~^K8-zz03Wh)tuL~M$P$%?0$G(5uJWs#pyx;r2{XIG7
z%h^T`uc$Ce9FF^^4PP*4dT2^Jhr^MV{%Z);&p79FIC8Tcjs+Y3t9JZ@SNav4diDIa
z(>wURQuU4;--UHEJ8bV-F|cCyZ&&yq`(oG4{h`G_^-Pp+dU@iE6IuM92PeKgG^w1t
z==elB^HSxds)8#67ybOy(fot!Zm+BQ-`%_aHuB7k^IuJUVSdXWUy;98PId`yTeK|l
z?V!JpFZ}k%@ed9jSy%VV59Lcr|Gn+{<))`!s*|r?>8*c!>&8=2<(-LxPJLNA_sq`m
zmv49c!Lx4vyBi7*45?k99C~P0?fFiJSC#C3t8O7Pd+gYc)^s~l=y_(nJZmdmU2$#P
z<-Bi>SI&Gh^y!=@x_$Xk=LMI?o%z$;HII;&|K(jBzFK*9@QPDoix!`|c3_i_dEr>Y
zyqdSSojkl^SnugIju$*z_WgBlpAqXEp`6BtV*6)*ef><La_REE<%_y+_todGy4Gvf
z2cL^mD*KH1s-$oT)!D4OQPp_t#HLH($<LKH9C!Ys<*xmGX-WuImeieXGF!UVJlCh`
znVm!To=WBp9d>%v%vUo?Pn??eMSR|+X>(^@`uqAhGiK$yezC{q)%S6SwmkX8+QnP;
z?Y&wxe@DAbzxRImjpwB;4clujR$uO2y`4TbLEK%K`DIhV)V)*2PdV-`DBjw~({9fv
z2cDndSY1+?IcH(vn))Zc?`iA~eKhad>DPy^T=Ds;yWJCVp1n}~e0BNoc<*ylhc0=)
z%XfcxtY-1SeoH3zUOw}E`bdX`sj;JG?K2(^On(36=T5&lH@P-ZeDsY|M##75l|atf
zSGqUeJlQzgycqoU!^YoD-?}3(?rKI$R;+hb<pZOO9y(b(s^>4Sop#Oay|bG<(RFtJ
z?)OR(*K;?V=&|Sd?VIl{c;MS3JJ-y9WXXhr)17xLt^VQcldBwkpFXhm!S%^i7ynxG
z*{MGA{Mh6julw`PF3PD?YR|s>`G(r9SyxuJn@z@WA7t(L^NI5p@`TkT`z{<g*iJcB
z{Nd7nod4>ndB(=oBfcv+e)y;8yAAc94!-e3<cZ(^_qw-~HwzCmtQD>}x3zoY?bpUG
zpWgD%il;{oT6wURoLNxy>ei~E&Dk>xKVCCtjk<WJ(Rb(mPx-D(UmMZ&yW)o0RbQnp
z&7ODkr4xJ0$CUS3IsToa#f^Vh9Jn-eef{m8|8)N8fn8M(TimUtRC%|#hv)nMRW`WG
zso28kz)wbgGicL{k-xluc6ze=CtpA3Z>&facW)dwam%R1k6I&^SHC*y$%<t?T1Ew?
z4*h83M&s>Cr=MOma?$=#eJ8H?F>9!~w$p8@wy@L9DN8~#esZLG{G}xpYZ~<Z_dl;(
z`uL*Q-yTftd}Yv(UveHDJh}g^p7##eJo!-Wtr`0+1b)}V&*@aMo=&Y;^Y5xX*QY$2
zIqr?mA1UA2r=>&S@Xse54GZT?EutKbjM9V4@89I^l}&0Ma@6H#<Y%@H=`%p<C+o+J
z7oVx@Tb)g+({^!2abD|S+xbsUzx|$k>3_0`KW+ES=sBQuu<hTZzP`Er>cTzQWK7yF
z$|&m4I@tCw>fH&$wzqHU=%`KG{o3~%+&b8HeQwdVC_OW;lVfz+F6mOzyLGVby(==`
z$|%{hE}M)>+j$vzuGYb}o2T?TEL%6tZ1P~*F32dzY#nU-r`*~%K6~`o)okKR+qoIJ
znXQ9uJL;!D+)z>2wtqTjweFY3qp6L%ng;ahSVy~QkNx#K)-i4d6zgU|ac&M2@8&@X
zZUL0&7C}jF36$)XK`Cwpl<HPNX>JXa?$$vWZUdC*HbGf#3zVmar<Z+z>{Vc8wH|%g
z2|>EPZ70wndNHUlMK7b;G!tYX%Y+!nGGU6rEE{AY%Z6CUvSEtFEEnV;%Y`_|a$$<Y
zEFa_{%ZGT#@?na{tPm6+D})5d3SmmXtQZs_D~3eKieXB`tQ3?WD}^M;N?}UEtQ?df
zD~Dvr%3(^ztP)fptArHDDq%{&tQu4ytA<p_s$ojStQOQDtA#YkYGF#ltRB=MtA}*R
z>S0R9tPwOIYlIBQ8ez)7tQj;RYlck7nqkVstQE8%YlSSxT4Bn<tVi+mw-<TGI?n)4
z1t_nl+E(@|CRwfLx3(X^pnC@S?T~%8o1fKsmTL>q%@lnDehGAnqHm*)5PBK08R=6?
zjL_#$CkTBFb&}9OppHA~tIoDQ(@ZhQV^S1@JSIXI<S{YAAdg8926;@9Fvw%#P6l~w
zGsPm0O;IfJ*a%^f$HoYYJT^gC<grP@B9DzbS>$oe6o))6MRCaEB7{R87b6_<xCG&l
z$0Z4eJTC6!kjFPuJo5My#Uqc85FUAajPS_g6NE<|pCml;__&itp3qDQ$P-eOfIJ~W
z1mp=ZA|Ovl5CM5Yk_gBX;!Xj1VlyQoPfSrF^27)cktfE8h&(YtMC6G{A|g+WJ4NJ4
z&6I>ZDMd-hlOjYyo)jYz@}vZjkS8UHgghzkl#nMkQ!?`86eS~1ju07na*W8xlM_Tn
zo}45y^5nQvMxN44Dacb&l!81ZLKNgFF`^((Ne~5jN|GqZQ{qkqd1^DIB2P_GD)Q6_
zQIV&{h>AQlK~&_aNunZ8jXPE3Y0Z>|JS|0O$kQT3L!K5R8uGLR(U7MliH1Ba?$nT{
zH&Z(D^c1BdPmd5Cd3ucK$kP)<N1mP}I`Z_mQ%9cBOc}^CQj~!_BSH-188KoY&qxph
zc}9{L$TQ+j19@gMWg^c^Q6}=t2r-do#)yeLGeJz`nMq<I&x|`w<XO#>g*+=oS;(^@
z#6q4GBNp<k1hJ53C5eSREAF(~c=s*&!qIXobb)GDhgVDcs!3(F-t^m7%aQICk9Lul
zy-0PeEACf(zkTOfQ`MQh79H3=J>5UEe~w)aeDm?h$MPqc>FKNvSq1jp`!sR!x*tDl
z`8YRyEp^eQaAkBcpja0RigR(Gcoz>!a0#G9mk3I7NuXqx3`%h+pj4L%N^@zTbe9gw
za2cRXmkG*pS-&d#5h*XL_0h=QiRrH$rRWWCi@`)G1~nU{SkzpU;!yKZibpL(DFL+@
zr9{+Hl#)=(QA$RwL@5Qe8l_a!T9ndI>rqNaZA2*pwHc*M)K-+T+Vta<Lr#kRdwuC;
zQBqz~WACN#Zx25E-le8PccyDgCfd_GZ#=VbY}1D2>FLtU(!1={vTMPS_P#lrE7H?N
zcNKN9@9yl!&LLT?kDIW|%jj~un?3+udITsRJsMPip5CUIG6s5#mw_JRW1z<b80fKO
zEc93}3q97yLXQow(BsND=y6^SdYq4g9v9%C$CvTY<GnoecpncvKEOjyC=;M3cm?PQ
zJ^^|{K!BcDCPGj2iqI2%BJ{+72tBDxf}Z4+peOkx=t%(ydUBZzJ=rTmPxi^slLIpJ
zlrjZ+idTW2;!~if1Qh70Wh(SkuL?cYr$SE+sL<2OH0Wtw4SJeSgPs=9pr@DV(9^v-
z^mLyNJw2d9&nPpXXLt?h89oDgM!<lcS!P1d^qSB!eJ1qGfC)XT%z~cfwV-GDEa+JQ
z3woZrJ$>x!Li$;b@o=C7k7TR;Z0493KiqUZ{kzWVn3rq+p5L)XPisGP^NsX$KAkVy
z@5PSR48A_8^xX0E^uTrl@3#L*Y|XpZgw6vR)6*W8r?WjB*V^CekW&;%PxmkEUusX=
z@1pL>YJCR<=M0+W+Rg;DLhEgcp&9TPg#nMHS@2ke1&^aS@HmA7kEeO?c!dW~pat**
zMF3BvMesyL1W%$R@FYb7Po`z?WJLx~p%w5HMFCHxRq#|r1y7?j@H9mOPp5V8bVUcx
zpbhX0#Q@KwP4G;`1ka)^@GQjw&y(-zVjtx6O6ck-vDFRNt`~@B@16UY0=hu6SAhN6
zF@3?l^Ue`gf}b7_$KKC?V*M;A&d-73{X8haFMtyLA}Gl(fs*|)D8;XUQvE6@&98yd
z{W>VaZ-6rWCMe5q{i^Dx&z^GU=B%MU&`^D6W89tL|2`A!4TocF+vfORH=T8j=OV|f
z$nuPgNn<CB8$Z!8@mJ3=sckbbXWHyX(z5d5k<NDQGdj33vpaUm&ADrM-q8D*ysY8n
zos04dx)%5AKVaaX!M`a}3PsO7_j0BG-)xtX?mc?-zPk_2@`9vly4ly=)9NN0LkhbL
O8!?$wRNd#L9{3NchdDI>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp.FA9B7FC9807074DA.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp.FA9B7FC9807074DA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4cac5b37b1c9b720faa0571244b2002ac7a8e74d
GIT binary patch
literal 1562
zcmWIYbaNA8V_<Mj@vO*AElFfyU|<jg;^LB`%mQWx21!l^hKf0RLvQEZwh*}bUnKe1
zY!hGJABu%<uH8^Gta6m&oF(Jgq7mdBwC~Gn>n$m}3%eOS#aVMYjoXVY!nK%BFvQH0
zRu|#WKV3L6|K9vgo;>ycpI-j`dUm<y{+Yb`{vzT2Pi{U+Ih<4Y?|kQ@Pmf$aPD!f!
zZ>sq&f4TX6y9<wQPuGmEshjcf_0bv|71rnS&$;*gDEV9Z?ZeN$`1=!|l^-~MatZUZ
zH6^B-j5t)A;x-;zd$BmN^(LF_jFQsFO%YO0M60^nbfa@cE{Dwud#L>;K0ouzCGP#U
zMmBTOKWOwURDZkfqFdC$P%C!jX&FvJd&}+ihQHXp@ya%jLuyr8_mbb$=uVor&C1DP
z^5t7r5tEkWvfOGn%SgB0TJ!nYUs09gWp3(H83CNqTi+<AzdSL&N49f?!kW@4sv2ip
zv#X?9d)+ma4stCOjC0w<ZTlk1i^=5PRXw>E1<rPEvC5GxFQO&q@>h2(TAYzptL5BL
zTA%Xj>5uNot+|d7>D#wvGnedfKa$ZM!0rFlLTmrx@(n_9+)Gs#PI7tJZ8mjr`HE|b
z0gn`~XNW9`$xbwpocBN^sig3B!ustZl1o)AS?pKFPQJvjYeS#Vt(Q^@uN2BMJ#m`7
z_fWFKekHZFCp3NYe%J~`=WF`0noezx@fMZkvo@J>F8J@mPfUxNO;%a!&i(CuuP*$`
zwXkbzb=S0R*W7jPt*w|#--eveGPYYBMRu^?a_C!DCKgsIc|ZO0$-a)#<Q2(QPo-R0
zSI=0cs=BgY`)SXThW}eLRJTsJY(346B`a$G;qr5rm8xPZE>7uOv{mHHyV^C=Z=IBD
z)w%5HGI805jh{7`v;Nc=&VMs6xN2F@*K)OuOCp<>7H(3?cyFuIx@Lv-Nz29Avd7aU
zcks-LxTNXgIkkALdPt^yTJDFJLOb>(m?&>cINW{kX2ZIWRrf8fzKvy@{igK$&3}2E
zahps&pDVteVBQ{cTYyz{f~fClt~Ia4<v$r!Eh|{6`f1BE>C_!h7A^@>+_z_^)t}d0
z;nS{9Ti>?l{T}A)4rac`Ry>_izNu|{^XG$-(OIXx?t0Ahedp3@JZr*-Cu^Ebm;SCV
zN_wz~|46}Yf3{xD`nDgd{@WgxGK%bVTOKLXYjx{l#{PVhQ$<|`M{da1Wz1N^x9WNK
z_fJz~gF=J${=0B@|LvG79yj@A{GB%!SLP<!0rQ_)sf01>k8{s?8T<nog1N=G6~P2h
z1`JI2dsMtVOs?}V1OTOZS$RFd1WdXpHLZ9FBLf4Q{Ej~7h2{#(Ji+|J{9<rj!Tb{Z
z(nvg6egznh>(_#N&EBTLAT4_QmM~!;&zj#2iRZx|4CBGv%ESeBWACBg3)VceR^SO1
z@(~J$DT2|ATws^N6gjrPVfgu_<Q5}Ku#m8jH%wKCpOE^Ci8ic^Ji&aNeEcxs<WHA;
zONHZ&fx@iptRgV)mU+9J(T*?+019(EafiU%(^$Ks<=}&dJ3+!>+_5m>dG(9me|?is
z3=~GTa>FFf?QGKPXM#Mg#IFLg63A2I*FfTF@$0~NheGUL#>-zn4m3%SMNkPQd~N!H
z7g|Ery+C1p4t^n+@KXcUFYNr{-9TYJQ9gN?a8YVbBFH1fMajkDK$<!D=&5V}UvmUA
zGX^mP!Mw=8ker*6lL6#<r};9ou(GjpaB}nT3h?uJ3c5P0337Ql35g4fNJv^(THDy#
z**oZoYFjCrX~<&}6_b{cRZvt?QB~K}($P0CG&VJtGm+9YauF4C_X=Th^K<lI^Z@{0
CTVUY;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp.E10E580D456628E1.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp.E10E580D456628E1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eced73fe4b69d88afc9fdeb520ddd1ab9c0360a8
GIT binary patch
literal 6634
zcmYkB4UCiJ0mu71eJH&<kUI{=)0`6XKJWMQykBlLn6V-gi@>Njw9Lc-<yJ}fm|R7f
zPSTNex}p-ai)_p(uFE#qbc9qkfo1B{#mx`Q>NJr}U8xq->A+5^Zf|cZ{WqcM@7~{^
zA6MRT?aR^y^XI=g5{b-Pz2KI0WBHXV5{V3jzsB<$*Q_Ox$erOAZrOEc>B5CkZ}9mg
zf1K+q9sQ~^cGa~0vp1|dRGxn1#M;E#>sRmlmj2nzk8Qf{m*Rr2^bcna-}C7OPd>Wm
z-dEQD^cVKa@|o|XJ}Des|JQRCzy0hLZ$#(s+<)`(@4vOk`qRE!_is8ky=KkH=O#as
zc=5^=*Y2Bidu7v;?;Tw8z|`0F?ws}T{`=kO&TaqLxcbtA=lv^~GqP;-y{A9<!Je~^
zJ-zlDuV%N-x%Iv;y)bt2)c(~=ZhyXbqH)P<SM6AT>JPQ;_kVuPQx_#J>obq<cJ=Sw
zera`jdAN4x@@v<Ard)Hw){)44M=!ch8DH?t*zaHP`rrA<@*Qe=u~@S4$g|7N9X;>G
zFYlavMS58<b1eR!Stk$ai{?!|zUrS>ZQFFs7Iy8rt%XB>ss8W_-`?`v$Zt;Gbh`B8
zm+sZy`f$(tvD5GU?$Ffjo8@aCwoknD@@>DW|GoU(&psNReey>K<|udYMdjH`j-LNS
z{^+}NPtKXU_=BLf=Kbo(4VPSQY`AFO&F3!rb!2?=x>c`+&wcE|19#7RC^2h*_D_in
z566c4Chj7A;j!@6fBM0rTNi%r>H*pp9v+E}3{Bj14}Y=!!Mi^@w)v3(IuIV78JjtI
z;x5t`4gWeLkq5Rs{QjvEE5--t)bQ|VY;?}VU8IlESa|r+9et0*&VOX*0F8%-hhjtV
ziM#G^-ds8Bx8mY=g8@1zJUk;dqi^D_d-&7AYxcdq;i+Q-baHrjFgDmXao0T@**Ny~
z@o6)A-=Wt(@txiEMEk+LZ(KTS%CN!}z56E&Nh%&tLYdto&N-M|axi&<gQ;>AOjW92
zstFZLjcZ_PQUg;<XkhAG2UC|in0i77)8Gb}hBUx55(bziH^DTe38tAa!L+yqrX?*f
zt%L=p&22DkX@hAeY%m2EU<xU~6bS*Q<PuCNC73cH!F0F-rXw9NorD9X%Uv*C>4NDd
zTrfTEf$2#POfTVq>2n`UU;1GB2_MXW2Ve#=05eDgXUy&^FFb8KFFQDGbIQivsj1|Z
zEkFgLY$HA9ISww*b8vZqaCFt23a*-0!Bq=HMOVvd;A(jdT&+MfboHDLuAbMy)eA&N
z*T@;*8hHa;qd*LF&729YnK!{T3&cd%%30uAc?(>tKrD3aoDHs>x52dw#70--1h^tE
zz!e1|(3Lp}uFOktWr0X^oty)%lXt*%3dBLz&AH&Zc^6!_KwNaaoCmI#_rUcE#6#E5
z`QZ9_A6&mcd~}0c0B(>Ezzqr{=(*jW7&rD${G@=>x=EDdaOx^;qI`l}qRKYfJGGS>
zRbD}EP~~mp7FGU@T#qWpqP;zD5{^CICLDXbL^<|&jdJYq2Ibh}Ey}US>rsw9wMkU$
zscoWSPc2auduolU*i#!+#h%)tD)!WRRK=dwBpUX#Hqo%Bm8gb2twuHMX$`7jPis*P
zds;oJVNY)o9eaA4=-AUsRL7oPqdNBV2Gy~rx2TRiy&l!EXEcd{J)=zw>=`9$V9%&g
z1A9h;8rU;h)WDumj~duBo5aMP*(N6T%n~)RXV$2RJ+nbg?3pcUV$ZBcP3&1sVqwo}
z6AOD*iCWmRYShA>)u0yktQNJfXVs$?_UtCHv1hl5jXk?WZS2`KYGcoCP#b%8i`v+;
z>roqfqDchyM4Jfgi4qmq6E!NZCmK{>Pqe7Oo~TC!_GFVt?8!Ef*pnqHu_tR(Vox@x
z#GY(Xi9K15O6)mJ;$Y8d69;=vi8|PGYSh7=)1VIaoECMk=hUMP_S`0MvFEmli#@kQ
zUF^9v>SE7rP#1e{i@Mlz>roebUXys(^V-D2o>!tC_PiSPu;(?Xhdr-FJ?we)sE0kj
zNqp@2ZQ^6kFHs+RevSIr^BdI1p5LNA_WXL($6nAR0rrA639uKGXn?(-Mg!~x4H{rC
zXwd+BK|LDudW#b~Hk>22Z6{NPM=lwe*M0N1;>|m9vA@kaIv9RFsW4?fgP&g&#*x(u
zQ<1d_(~$KF(~*q|Gmy;+Gm)(dvyklyvynxG31nGe64|LR2idJK7ul;Y581CUA33P7
zpvSBH1&<G}Jba+@M3kAb3!jKG<H%~6smNNHX~=q+>BvTz8OUatnaEa|S;%&o*~p^I
z1hOnMiR_e_gY1@>i|mz|hwPV`j~tX)(BtffM;6~S{IwuFQTiwM&**&0?%Y;eTXvp)
z{o^B@=O?aQ2&YM$1F3NpNQ-MgdRzxG;s%fzH-W6U1!Tu<AQ2ZpGA@CfxC7+IT_7*+
z0r_zsD2NAV=tdUV`E$4ZV&@wzGUZ-85sHi>t3{?FYel9Z>qVv`8%1Uyn?+_KTSaCe
z+eKz0iy{-qvdAQ|Q)CXZTVyV>S7aWtUt~UVP-H=m?|N(360v3ZSm%kTGUWh15mm;K
z)hbhwwJOt)^(xbmjVd#c%_=jIttzvS?JBd8MU@F;S!EL0sWJ!Ituhzct1=JSuQDGw
zsIs8P-E*DvPh4E;exlfOeCy`rZ|u4|e4&iaA6?kB_TKXNr1ZMo^TRLCJZt7T-H#*P
zOHp(CCoVyCf0OXJO)8glw;`pJD}mC=QlP9d)+3(eu;VEXJD%pS<5><nHL1c*O{uU`
z(<<!LtO`3VsliT5X|U7M8tk;J20J~e!%k1>u+!5z?DVV-J0oep&PW-sGtvg^jI04W
zGik!kOqsAV(<bc9tO+|SX~E7)S+KLx7VNC71v@)w!_H3Gu(Q)P?Ch)!JCPKy6Da{Z
zkruEMSphqll(3U22|JmVu#;H{J16PD&Ph42bJ7m%oU8*oH|fI8O}Vgh(=P1XtP49Y
z>A}uRd9d@+9_+lV2RlFM!_H6nu=CSC?EI_`yC50BE=UEi3(^7Xf^2YRw_od9w*Ai!
zy%WA#ho%e-c3-83FJ8KO(&f94hhHA%!=if)yL-Vn{Fk6}wYd8tJ|5qD51svJ=e^Yo
zKEv%=-Rs9g{S()XyLR{@&nVZy3pvAq)Qk$GWi%i?qXQWk1IWynKvu>AvNJZ2$Os^r
zkw8wy0dg}gkeBg*{EQD2WP&pkUKd}xvh$W<d}DU}>du2RZz3X*X&wIGt@9r<|G%Bp
ieGrKZ_8v|@^N<c7+<pDW_TpDQip@A{XxiZP$bSL*AZ!r;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp.030C33F297B655BC.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp.030C33F297B655BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..125ac5b2bd4460db2b900bc8e66b3aa4c31dc4bb
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%yMQ125(LVhKe~yLvQEZwh*}bUnKdM
z@2q8Pe<oOzW#>$m*sIXb;@hX<B(hRfGyd0C*{Dr-cerl!71KyKJ;!p6ak(Y)35J+?
z(&{1{`lkyg=HHv!$&;u4|HI3_U(X)b+&`07-(Mu$|H;iqPY&f2{`=)pSXsEJ!pm^~
z|4Aa>w_l#U{{71jUq4ULx3fL-<7<+g+@z+@{GVClYxevzs{T=OeEoZm&)N?Tr(9wz
z4%;OaDcK_AxIW_XwHG@#9Lj0t^Vzkl&{2o?6W3nX!_(H9b!BS%uKpnT?fiC=s+X<z
z<s{|K8CQrLyD&NLdWQ1c3tDrV1-(raJMUS~uM2;%edCpF9*5MbvhF3ntI?e_ahsKs
z!{p01ts*8Z$z|EqZkCa5y|w1+tGl904wSj6OJxLbN^gCmn11pp|FJ%o5P@q}r-VeF
zEi$$3J#=ib$i#%!OC0)2H(TXObVV7Z-e-yPm+V+5v3T7?9j7mAd(O4ny4(yj+iEve
z@yM?7nQ4~)L{bu?1+Lx9eS3>R_PwCtEu|}mg6;cG<-d?!({kM@({qJN;EzSVZ(lI4
zD&+{NIAJQqy7F3-gm0U=#ZrmbJ7zat>!!M?dNLm5&$_<Es6lVd;-<1)ZW^UF?F>6F
zq^<uTabbOjr#6>YutXjIk!#VXf*pO|F1!%r-hP;mcj@OV^*i=7yii#C%Fny5e%0;u
zvu@owwQFnZ)(>W>*IV}<Z#p43EqC7HcC!wq?+(%$7fj9_TD9)c?<~8Wf*bRAe0`S3
zEMAcydU@LtrLUDsXH69Oa9?cg62YRX<Cm5@IKM9cp;uw$vYuOimQmoAX_IdLmHWN4
zY?Hg<=_j2Um&EMK1XCTF=j%V4e7D%^{FPPz=6dR6UQfuJvxeiAJHNN0_AEcHb1$Nf
ze~e_)YgU`~q*Y^O@w%@qT5pf9;kDb#qNgwMwnO*Ehol0t1>ss>^XL5fX4dTcZO`|c
z|LzLyNy&UZ_xZYmdHg%KDROBkXwRA^8gae;&6D#lW*(Y4?ZoAo`BNTFYM<oR`s35H
zS0C1D$3@0P?~@X8ciAtd{p2r?Tm<{3Oe43$$M;M&-6J`N|J;qB=&)CAr*ewawdP!0
z*P<TBe=X$4f7>~Yr|$`0&)u)PW!E2by>-9N*FWZ;xyEhrWvPC*xy=_d>hn!b6`iU&
za)bMS+l)1QtL|_ArkZLqW&M)8eFh))o3@w~+$?|hI$}FRab<3j9WWoVvz$53Vn17#
zm%%@fA(&f?TM<kEWx!yo^fJGIc)o=^3;{rCURK^<FaeV;N=++X!pOkDCcmT4d7-%i
zGfyzTFuxdFS1`W>zcdn0mR|wJgIUbP1-7vF(C-Cn9$G8#1Pgl$hrtwa{aSFZ+1oT2
zWUd~+B}^E|v*ve0;(71~!+4JEZy0`lDY?bS5-iLi>;)4JveFh$;1Ai%2($o6IK)p#
z{l!EZR-iB+Cm%mdb@Hc6zNNzP#z0|Kc2*IXcgwt8&S*!N1ptM)ow!3_UTCb{(Q@#?
z!<``EFz#5G@P<j8+u5Yo&jdMMiC+a~Hjt;ruYttV;@5%kU^F8aIGAC9awx>^WxV|L
z<3I}pSp=0}s;*5x@Ip(-x)&(S&%rMQ6MkyI`h}feyc;OYC(0)e6D~^4Nd$SMxG1?;
z9K;9$5n#X^{4vT*(cRKJhB1gCIX5LI11RjB=F7~&%Er#Y$<4znz|ZF?=<2K{$mQuI
zBrYr>A!%W0ZDVU^@1Q5DZKZ6cA&*T|Oj<@(K~YIXRb5j{N8iBE*wkFkL`v7lMO4h)
ND}>3-&(VX?2LK1rZ%P0F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp.C8A37285FABA85B1.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp.C8A37285FABA85B1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..56a26dd6e17d52bb1abd3fcbed0efc2e37da679e
GIT binary patch
literal 6640
zcmYkB3y@RQ0fuu<4k5dXu*>eaLm3Kn?z!)K?n{Z<2SJJkks+>vQK!}=LeMEXF41Zo
z#poCxK}DP>6sU-%Vnt;HqZZH(6*?5PmZ|Aj3RE=2hSDl!Ce_s1yBS0NogwVE_xt(W
za4*SThvr;(;ir>Ap;_0>xoz3v>@W+3LVdxn<=K(zM@T5NA^5^=>vs*!of~$0-<!X0
zroCzWi%-6?dF+@y>ryujUGVb_C-UzWE-6pg^Yx84y!PUP&o>UO`OV2YKl}cY`!`J8
zd8+$^f%|Ud@6A5?(^Eg&H+|^^7yY1ns`LA2cNq^&eD12@OaHp*)!*KI@!V&gI&$8d
z`6uX()l)t*UY&dEiFtQTU%GA4Ju43^+MU>W^0EEO9q+8`n{o8x=Z>w}(f8KhZg}vG
zz8&Ws-?onIc=Csj-g4^1vr8u2wKxBd+PQzZboGs2>{+;K%_nCsy5?I?vk#Unm^ARg
z#>odC*xPmd?d{6;{ckK8d1%p#Gq)Uw^n^C=n!S0>6RS@R-1F4%z=1j2rp?<pGIRdO
z`uirY9Dl~r=PtSWtj+QtuAj1dM)b~?#y$T2?q8j?^NMFTJhgS~70d3;*>nGS>&~%X
zKjM#kbnw*RjAwp#>E1mT9XaRJhbxv(p0<A|TYc-5JLYfu(f-#O-;S(n{CnTOSmQf0
zw+wIk)0~rsdw==(HB&b=x~_WY;NACs^}_d7oj3dWcZNTlyyx0CmM*`2*`g1Fzk$fq
z*B_bnbZlA=?H(8E?~n9%wH`uU!Li_@`^$}o@11+jWj(YjI6Nsbsjv0WK0KrO`H9WP
zZhyIl_5_CqBZFgG523Db@Ru424V`ZcUOVCPZ9Vji;P8~lltk+x)Wv8dIJ|Xr*K3i<
zFR$&P(co}jq%YcfX#e5$;c36N&OYJy&@sW`fyh8t>!E%4tKNmL9J=Z4V?A_iaJV<p
z+tqq#9}bNyUb1|`VCOq@yIbGc!@HXsw|w;7Y2*4OE~)MR!hnS20>z~BJH$B$lM4<e
zk8v<%E`uox8B94QgQ;)@OhqVQDlr92m8)Q?LIqQesbFec15*<km|9E&Q|CIEy3oPY
zV>*}yH^4N60j3c%z%;oDrYTG?&6o+M#Vs%`VS#DIEHDKZU<x6?6fps&&22DkVS{PM
zY%m?}fawSaOef}m>2eoLSGZuhF&9jadtiFP1JjFnVEWt#(-%INe#}2@wy(V4v>m<d
z;Iz$1Bb`%I%1SGMazt85I?OX1T%P6N@*Lsl${874IV*!J=ZK81l2O1_vI@9LjwtA=
z85LYLtAeZMh>EV3(ZJQR8n{}HXz1!09b7%DgRAF=j;@h0z%{Z4xJHf`=$aW5Tr+Ee
zYvzcFu9dOCwXzntR*qQcii`kPWCge)M+CZd#s=5U+ThwbVx#M19B`ei1Fn-J4!Ul}
z1=r2G;JP{DqU&WmaJ{Stu9qVox_-t7*U$Rk`Z?lv-1aNR&E2h63OKFnL^=+quGA#b
ze~}ART1h*nwp5|gpOI@+`V_fArT-vT!_u*EXV2?|V~;lp#~v?Gjy+zX9DBS*Ireyi
za_sSHm}5__6B&DQlgQYU3slCQT%j`d<QkQ+CpV~!J-Hf|v8U9Df<2{46znMls$fs4
zPz8HRjVjnv8dSlaQVlEEQ|m;<p4ucT_S6DZv8Pt3iaoVPRqUw^s$x&AhE?oob)sQU
zYZ486T7hcV(<)TMo>rq8_Ou4ou%}hS8us)$(XppDiH<$JKy~ct6{=%TuTdR)dV}iN
z)2m?}dq$lY*fW~Mz@AZ{2KI~!HLz#YsDVABK@IE~)v$p*vrbIxnN4D1&n!?AduD~2
z*fVR?#GcupCicu~*u<VyCl>asCb6(*6{v+ht3oa8Sv6{5&uUN$dsa1UVNcYFz@BIl
zfjv>60(+uD1@=UZ3hapn71$Hiu)v;OCpPx%Cb6++7pRRryFzX3*)?io&u&l~dv-N!
zW6!A*2YXJFIM{Ov)WM!pp$_((8g;PeG^m3;ry6#!=hlgfJ-10*?70Q%V$ZEm7kh4v
zy4Z6Y)Wx1#4ZGO$>cqpI*CZbHyaM&G=T)eOJ+DSR?0F69Vb80EJ?!~);$zQm5+8eh
zf%@3<E7ZrHU!y+u{08;0=U2mir#COQ`ld6jm8;0O{z>Ofn$>>u*M7Mo-K`&-+OL+^
z9(^DaIWp~VZ}4MFiAmd8+Z)%<ON=ASB_<;)C8i*&C8i>4C8i<kC8i@AC1xO-C1xU9
zC1xRu5);UFiP^|bi8;t_iMhyLiFwF=iTNE)pFA+{mi{06!PC|~wtHaoXU@@&>x>m^
zzdZECfzjtDDourFC(423s0^e;6(BXL0%=hVNRR43M$`Z@qb86QwSYua0NGI+$cZ{Y
zZqx<xq8^YR^-t5yEVTA{r}_TqH=1YC<9H(E8Aq1$Oh#7nOhH!jOhwl6OheZ5Oh-2I
z%s@8t%tW^G%t97<CXnqsvyq)VbCBIUbCJC~^N{^K^E>?T$Lr@?E3RBT`b3nO^g2Ef
zWyX=^GLw;&GE<P%GE<SYGSiUtGSiWbGBc3PGBc5_GP96HnF(aO%xq+*%p7F5%v@xz
z%sgbj%=`|&aQxoeul#8JBf%9hB{^k5`_$iZ+m12GWt%PxzC1X7@XYqR@b(>;ncb~B
zFYSL{f+sO9UC`bJl#nh4N=idOX=!nXc$~wICpheQlEaRtIqc-P3_Cd?!%j}hu#?j=
z?3B0yJ0+pOPDv`TQ_>3T)VK;eHKD>zO{%a{(<<z=xCT2dp}|f|YOvGN8tnAA4m&-e
z!%k1?u+!5z?2NbpJ0oGh&PW=tGtvg^%(w|VGhxEcOq#GW(<bb!xCJ{aVZqKyTClUy
z7VJb^z)mCt>_k$)PNW6w?6?g(J7L4lPTH`u(>Cm!xC1*U;lR#GI<Rxn4(!~x3p+RA
z!p=>)uyfNc?7X-KJ1^nE&P#f*^U@yd{J0N0KjFj9Px`R))Bfq*cBN~<s&}9MdvLY(
zjqB@eU!`xJJ#^ie*_)0BU+(AqR{Pd%`#x##CSmj@Y5O8R9^LXNJ%8)yTZ#cb;Iysw
zZPUK)*6q@^9bDuo>1w!;QyfT6$v{d<0a8;cke1Sb^pp-{qzoW4Wdd0#3rM5{ke#xD
zoRkCPrd%K|<pKFA|1|w?*TJKgPd~A6^aY~Gq+j3*M3Hf1xyWQ>rN|Uywa8Rtt;jTF
zy~uQAqsR<ov&c+jtH>;5QDg$yE;1Y0DKZDyEixC`D>4t+FEYQw!F}<c43EA^T0W9q
ze%a_J6Gk6eAGH6ELlX~;eDc^!k%94j6M84M{wK{h2giT&d+=Lt=hKOuPfx%73Wfd;
DUoY|_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp.1333260C2CA69DB3.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp.1333260C2CA69DB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..33cce0cb45afccbe47e01a699b1410f2bef7f5c8
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%t~ek1~*OyhKe~yLv!<Qn+aU|uX(O#
zx{f#R59dd>R$o*z+$HwYBjAarlFLLX&!4IL_qls-lisdzE#&}1^rO3nt1H%ruuWjx
zvB<zrLm)2f@}?X6_6kVunO*zfbNlVx{l#_LCF196=FL0Vef4JB(N~|g8z238b><_>
zqn{oY2ifeIeZThSWU=^vvp!#4ew=lGtnTvUpn`vOJ0Gp?-#%ZwfBw6lZ*4Q{SoX|6
zTa|E9>$<|~rU0cErrCGdY{kQ7#9nChyuL-R<Br1mi_s=J@3#c{L}yyH>Tjsux%A>r
z@w)P)ialo@cuF|?-@3b?cZ<{N5=OTt8BRib%kB1Fe^I{i(l(z%W>s4Ef*;l>Pnx*R
z%E=-4!Y!*4o|nv+-a1Qf*}OMuU*+d-Y?G3!7aQ|#S<%uP#W!Jd$rJNqeJ&vaW>%-T
zM8p;uGxr|4wpe6hLhB_CeWjbNj3u(Fj8X5U#QRHjsLC(So1o+LWp2;8c4n80fhN6n
z(-n`TSr~mS`g88&r80#zn|B{8Wvsd<oV3L?qh)#3oGJeoly4A=<6f$|aFWZz(|S{r
z3sziI40yEgLWananCwIo$$1Y&l1d7%KU%O|L~^N$CCmL4v9m8R?Ap>NbnA80!Yi4w
zOd?LRn-3*B>{nA;dqUGU?+3Ghb-tz_EAQ0y7;jP8{%VtubHRTf{^DBHY_h6ackb`l
zdvzgK&Rx5ftCQ2RU31sHx6EQLjT6dU8|8K>c-b(-D!2z*bFco@Bfsfqj{A{irdJI2
z7Bw$wEA_cNX_Cp2sYPx<2liWUnG|*6rR`|}CeyX|A6Gxm6x^$~H)FNS#YnES|8`fs
z-=_9Eg}qc+<PrSiO7#>*)4GjO=f0h{v<+U_W<5FL(i+FSIcqqrD)@I!FfCQLOfOmd
z_=^(rbVq4zt5%JbMe9DdXuVy&hTATd>3Voi+L38FHKw~HE>89OI(yl!zY`CdRo|>F
z+Y#L~UFYnc&v&-oIIno!%*By2CH3}3q1n%-^ZzuhT2`=D_0yJT(WyJ0EL;?(xNp&3
z%YU<HU7uz?ZT;~hf*13TNLAW%x?h@@bGs+-z2W;5=KcM8?&3Pvr-kmCYBc+9x0JYv
z*29|%-Lp^qu77*>MCx~b`7o9xr)R|dQD?uyv%mhRoKd8|+Vh`teX9PQe6k_m=p+|c
z<)Ite_KrT++^?QL{cDzt+~n{;&G?r;{$FXhl5;csrcUTzhT_WHBs*YElznb-m&tf?
z4i7^>AVV-OD{l~(0Lp+tjfi=B>y4QaybS(8X>KuY1uy}VE=o-+Uc$)0z$U+=&v~J_
z0y9rAzc9ZT+|Xct34UoLo-DrtjK}qB!M$d0(_oMmJ$_4=Fpy`>?}o(l;17oJU^F8a
z*p0o1elJ+_&|1N<{SCv<FD17aS%QT*guP&@f~>T~6Zk_mGxDGahxiGpznEyl3KZt!
z<l~2_PX2Vsw^TUZ7%0rj&ME@)V41he8SMzO0H83pBX=+?IAAmr7dSXzE&<x)EgT9n
za>FFf?QGKPXM%jF#IFJq2J+PSHIR5({5mk+p%A;5@$%P?1FaKe5mbT+Uz>j5g_e+Y
zFHo4DgI@?H{NQne!$iRiw?U2z<&K9r>Zt+i7j}N}ZlExqD4#q`bx~?gBFJ0CMajkD
zAVv^~00ZXWS6MqE6y~hbVhmzP&P~b501A7j`7*Pxvaxe;a`W&C@bh^Jx;m=~a(Oxl
zi3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g(Kj$OHZ_+sk<vAC5fyXy
N3Sn~dbM#>J0RWgmb3y<B

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp.B5F4680BE43A28CE.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp.B5F4680BE43A28CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..278de03dbea57872e1d7a41dece05f3eedb294fe
GIT binary patch
literal 1564
zcmWIYbaNA7V_<Mj@vO*AElFfyU|<jg;^LB`%pzt61`$pMhKf0RLv!<Qn+aU|uX(O#
zx{f#R59dd>R$o*z+$HwYBjAarlFLLX&!4IL>)gG!NpI(vP$H4wy{Ec@|NYcc3`Pv$
zBF8vc51r1661)9fT&bCVvwzLMg6i*;=iZAwer<g8ta0T=tJ@nQV*iyb+p*6o*k*Zz
zU5#&7JOAh1`Sq6`-HtEb_v_P+2isRaE%D(xXRjAt^Y!Pq*N;A)w%xy1{p|UM{>_p6
zI+0JMQbGm1SoWr^y}jW1A+F8bRt7I;U1Oan{lv7YyY1;*v#v~S-{l{8zvXW?vHE`K
zy&q5ibV(b|<QJa1rMFBB%b2QnphL#jW6}3>%fCzg)eE~d&GUwv|Iuo#Ir%O|m%`jR
zGPt6m*)2V9xh<&8XfE3pe!Beqz55PDIr$~a8pWn=2t3VLG_CK?xf>HZxDL6wM>ZaE
zDoI-qc%y2HlF*s}sfVsAyA2xFEmm=O9sBsvM854$lCSVia_!hR`Qp!l1`VsOWfz|t
zwcOxwUl!MAZxMNmucfr?`VKvZ^VLp0+c<hZg!E7TB>#o)+M(=)TU0_OY1G_YTExR=
zYQ^GZBO=w?G$mWNNA0lj9H$<=cTqB5rYS9+qSF4rdTVyDl*9CEflk%2+Ec97IyX3F
z6bDrBytsa3iKyk$lf3l|NzZkUu3WJ2(}e7miyjyDx1Fk7WgipoSTez_bp6ue`V+U$
zcip;qYgcb;?*}v2>$__YG-(KYTBmWrS=^<QeF68AjI8v;(D@&o&E~)DxUr?pYFT(*
zz?BVsnNdNL+I9q;^WgeX-X0djZMipH(rLl6uhlj0ZFCo1Un-tERl~~r;G6r)t5)uP
z6L>;vu8T+}*Zit}qXozG!%pA)XSqCG>uN!|koPRp6TOp79ogS2if>x>PUU$|Udf!e
zlm|u`$*#)<R71~|&I?%jcS`BO<=zi|bzSVaF{yXGo4E3>pp|u(L$=!+-AK#deV(t}
z`@^KAGt1}g7W-}a<W`%8M&p_s>#(HDdwxy0e$lu!TX*7e)B9S7Rry7?a(xtid-mh~
zD|@!>*;r%ytoFbBO1Dk9#%{95_e|#fW4B8tc0){b*vqg}Ic54;F&7$I(%og(YTo|e
z%db5Bu6<3E-~_$6s*mCs`EB>P=O|A8no>0H;K?QFbA9-Jt4&_B;DcJv-$(9kCx5M|
z+F7vv{B^@XQPJ1pcmKvqchA!JEndIO;b(DWZjv1^2fCg&t+;W_L6VojKae4qTZ~%)
zOaNuTV9uNeXWt%EQs-d^07~<+@_K*?m~>HUTJaJ_1_n0y9evIV%@vq=g87B{#o)Sv
z`6c+Jk$AHF3NRkmuLbv-y-kBbTJ-oWVZuP3HNP7Y&x1c0#&c|c!|?M<$t^~fU?E{4
zZ<sL53rt*Kr}rNEy<p8lYXzQQAs?YIn4%CrA@vs%ZCDw3g84Z4_+i4ypDy{93db7*
zg<08IMPU9a^L9C-9bpy#6y|p14u;tP)5XXI4h5J^^XeDB|N17O7^oWAT^lBGZfBET
zKNI9bC4Lo{W*|?EUjvD!#jgY79SX5~883hRIM5_P7C|ML@U`g&UT6tf_X36aIrxQO
z!VexdI7}4Wa2w>xQ0{n`E1w#$eqrYq?*<C<iSo(AR2QY@B!c``T$Efa4y2ibcU5@u
z&6s<noH2+Y2<CGJhUDCooD3k>JI$Awg_Vt+gOi(wSAd_-Q_$5}O_0meNl08+L_*TS
z(%Qz>&fY;!RNG3~OhX=<sF<{jtb(GFimJM%mX5xGp|PpCoQagKk&CF9yH^O4o1dcx
GqYnTO){g%G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp.4093668F9B2A46D0.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp.4093668F9B2A46D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fbf55208db2c11f36383fdf375aa3b551ab94dac
GIT binary patch
literal 1564
zcmWIYbaNA7V_<Mj@vO*AElFfyU|<jg;^LB`%tB@c21!l^hKf0RLvQEZwh*}bUnKdM
zubD6J55>Ya*KVj8RyoRX&XVzL(FpPm+V|zP^_G;~h20FE-K;sC#_h!x;abcm7-Hs0
ztBY{xpDvu3e{X&#PoDb!PcQ#|J-b|U|4d$ef01ziCpRCZ9L_2H_sylSvT#v_m*M{Z
zvqZjczdSo%KI7xt(^K^A?9cr8nq((8sp)h7XV&<dUH`1Af7Bdb|K8)X^@GPLml%t~
zc1cA_wg@?{k9d6T#m)_fa+>*kcI_&3)ZzWawb%9Vw6$hkncBXqKZt%?zulzjW$S%8
zNx5^z6(YwjOfI{gp}h8j*4$=6ZxhANd)D*sX@9vLkrl0yC~PbG-mrY1*h!CQS;Yk>
zGjnBiJTI9sy>*t}vUzUQzRJ)4xF#iAFBaxCS<%uP^=-oDk|+Mh`dmT;u34QD5_z`B
z)VBA~vBe@26Iw5E=quf9l`GLzWt4iKCC*>6Ls5S5x`{eYU)J`VYqxc|7-+K9ZmQyu
zUG_#_i~bxvd1#x$n$5RwZDq8&r<}CKbw%s)Dw(PGFIq=%u5Z0GDZq2l2iMuBE?S4Y
z7SQ-OA=`v2Xx&zWSv>v)T!y=L<Zieg-PIE`Nv3K3m35w(46zaJoNr5d177XmV=P*D
zHtwO}f_TBnVNa&0ng5aNSi4<hdDE;@&g+(T@wLlJc|BkGub`6gqNCK+xzo=5Rn6b8
zot3Tq`kL6aL$^g@pMR6%R&tNnR@o<aOM&YSbFPAWur>GUT|MtN|4eZ|vdi#_;oPF$
zMNOeTmnTiSa(rr$ThM|3QClWOop>qhEze}S_Wncb=b3_g_4Z_(a=94AmG<54n)ln3
zey3?KRTg;!|F}^(h0(Nb-wgk<^P0B7D}Px}j<~eu;3bPl!7bnAra6UOnfqkU#jSmh
zH}k}>`RcqBS+wNT&TGP}w#aWb`%%If6K60>FnYtoqYrW%!nLmE&w2Gtui5w8uHQNJ
z+gbD@XH`DmnZ04Q^Sa!QCZQ8u%TigdeeLG2l(Y@raY?8$>T|DA%%^~$)dKPNVr2i7
z9@RdbeL6h9@_r@zy2hMYJ%OhU->2~Y=Q+oop0iR{yHq*VY^Sj3xeTX6lQ;8SoAPe|
zlOqRG<vSj3o6jQ^{Ez2j=>O_I)09n8y?&c5q)N9ezF2oZW6D#JhaDU3|6DMLlnpsA
z{{E?^^^#Rfs{StA{Xf4rP&w6}LD|T82SagXZjv1^2j;j=4NlVew}F?zKae4qTZ~%~
zOaNuT;IEYR@+Z3)-FO%RfYQ9Iyq;hJCS8=8R=k9ffq_kaN1yXTa|LFeV18kKF}SW^
zehGeQB%Un40*uG?Yr(x{Z_{9q7CnATm@trM&F_Z9^WYDL@f_RVF#P;da*L5ASV&mN
z8zvm$C#3#jq75q}PcR=RA3sbu`O_ueQsH=GpfD>ts|d`=Fb^|vfj!!L==Xv(53Ln=
zf`xp9!eM&Lyj{*{N0<cw^>RCLhrskS*6wII_~7AAkZ>4xEKGRAB+l(@((7k}yrRUf
z0y6~2Q{&e_;%V{gz<7s3>|VypUq22sNsvWQ2_}5*g#4Efk8D?<FtQJ?O+WBLOUSwx
zD9q2nF9cKl)PVI1JHL20P?%4YPaY-=qZzrtaRu{XQEE;iC=iN^l8ePbj35vJ2F$@B
zjE|H`OxE=<1~DY(rsQM*g}u{!nORub*f}`4d3Xi*`8)+(oz(=nJe`EZg+(MJEiA2V
zZ0+nF^hC9-l+85cv5AUF%g8DyDygWdYijA}8yFg!n#-9;=^D9+in)7*FuD0TdNBF`
E0JeT|`v3p{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp.A50E70F0D93B6F8E.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp.A50E70F0D93B6F8E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..17e95b747ee6e19632a73e87eb8681471994773f
GIT binary patch
literal 6634
zcmYkB4UCiJ0mu71eJs5mLXSH+31edRKJWMQykD5QTXa<i_C!d*k5!y=ShOy18o9cO
znR6XUl@yi8x{XccWOlnn78#IABue78#Kbr!(>h%^r>@i$nT{EYS6SA(t@Pi7a=*U6
ze|}uA@3nbZJGx@Udy7J$D>jVYGj?l!Jqv|G1D!wP`OV+nL_(o$od@oD?D@5;R{ky-
zKel`SoaY}LS@zAV((_~I-MMG#Df_RlC^ziC`NK`?o<9A~s<F4%Szo+1b|60WrH``b
zf9LxjkGyo<&c@s4owgSosXV)~x?*tT)SA;jSfif2>51Ws--!MD^T==Ri%uR~HeKKP
zzIFSB`@~z}BWn+DcylWMYGmxkSMOM}?&|Zm{p8ZkpKW+#OK|+k@11)63*(97!`~2Y
z;<jHL*#1!Q!_^BeslE2>b+7#M%=Sz6?SB9ED@ULE!|;ul?`a-=u<>G)Pu}tJ(Fy&5
zGk;mJY;{Ka<*v+`(Y>!n#zT|OFP|J;_~<7?Kis{3=;&ymt(tgl-4lO*YUf6GWZA^y
z{#aq+XLqbV5&zF+A8cGRd3pcKw|u;2d-eJwbG|;dwea{GpZ@&V*85*w^ySvrbZPf%
zJN5TJKk!j(x_RV<b9XGiaQf8Y52yaTsdnmQ`GI$*@B8itH{Nw$D!y^@M<b(Ci(kz9
zFD<=y_eAB;w!43N)7{}EztYdU>1&_PTl}}*((x^0x4hMP?jsi*diaVb6HEJPY;I_9
zFf!OTdl%~K9P8X-(-Tu$S6*~YKke%rUKCk0FniZM{MGiqPEH=*@=QPN?;IYE49}Up
z3-yINKh98S-=lrcMixJ_tDi<YhX*19(b>E1`j0PN@X58Qjm!J#xt+sHB1<lvy$kg*
z8tJT?>z6(!mYodx={cRlLy@7r*}LxgO#J%&$L{?7@qRj|b2uJ}_s!mQ4~I72dfWK?
z;of)X#b&>=hhA(??0xquOXm(MT+zG#gbqo?14<|(J>r~$$t4GqCpefYSHV=J3Z|M+
z!PK}0rY1EowS)$y&UG+#se`E}bTAEWfN4krOe0}{X>t=xQ<`9!2@_0<TVPt!0@F%Z
zVA|XU)0Q@vcESczZ~>-}0!)z*U`j5*lv08z6B0~^J77A}0n<r1V7lA|)0HlmZo&oA
z;~tov^uY8I9+*D&!StmMrl0V^40r%$AOkRiL~z#Zo_U>@?aXNhFWa25x%bjk^2+@{
z1)^*xJ?1$MF3)ptd4X_r)tm~hnpeS93q(a%%W2?hc@12xKs0pqoDQy@*TK~bL`T=i
z8Q>au16-p(40O$$39gwp!8HrSMAyn$;97YLT&qAVbnToCuAR5RwF|^XSL6h^A}_!d
z1tQRuISH=JOK@d@NOYZ?1Fn;Iz;z15LD$W>;JSGiT(>}6biJGhu9x?~^$Ns8*U$Oj
z`gtE*zd(F+gIoY^kPpBO3MA;ci`$o4_ijG1Fy8r%TV={21~1ks<H%~2smNNDX~=q&
z>BvTv8OUapnaEa^S;%&k*~p^G1hT9$iR@IFgX~tBi|kdIhwN9Gj~rB4(Btk~=UuVc
zw@f&{n?yMQ=djWy%IC->s%)pdr%kC*<xS)URX#v&QROspJ**rL_x8L=IQDp(aP09C
z<=Epj%CW~Alw*&#D90YJhdK7tCQ-4cwuy>8wM13ysWqx%Pi;^YduofS*i-9a6?<Bf
zXxP)*M8lp|q8j$J8r86;HK>L?twlBLY4xy%J-tbE?CEWyV^1$p9ea9>>e$m8RL7p)
zqB{2UdRWJv(If`;j5aZ_XOyUcJ)=eq>=_MeV9#h#1A9h2Y+%o95)*r7o0!-$OVq@k
zS)(TQ%my{FXSS$`J+mG*v1c`jg*~fHEbLh&YGKc+Q44!kgId_LTGYaxRS#R(vzx@m
zp4}!k_Usb1v1iw)jXk?TZS2`CYGco?hi&YMCK1>ZZ6dHIN>pG^)TqFoXi$MY(V_x-
zq8=95lT9MAC)-3~PnM{}o~%)cJ=vfVd$L6(_GCRQvF9|2gFUBB9PBwI>R`{QQ3rcY
zgF4uATGYXwQx7}XbDPA)p4%oa_S_P6vFFyPi#@kNUF^9n>SE8Whh6M>P2yqCYZDK9
zUWt0x^J>(?p4Xrr_PiGLu;<mo9`^hu@v-N(iH|+MM1Ab}HR@x}Z%`k5evA6p^Xp+B
zdqI-~*bCYuz+O<I0rr9#4X_t9Xn?(-MFZ>w^>EPZtx7z4=R&c42bnv#=;B3JbUy`u
z>%9kak$09(&3tUAFy$bAY^X4ftX7zctW}tXtXG(hY*d(mY*v_wY*m<rY*(0#EGkSO
z%L<doPK7zhZiTtXUWIwceueqSL4^f9zQrFsJh=XaLo-iAnJJIq6H#UySuHaaSt~OQ
zSuZmk*(fsu*(@^?*(x&&*)B61S(KSTmSrZ9oicNf-7<5Ly)yHV{W9~BgE9+xoIQ1P
z)!l<P2b~v6Y)))w=5oW#ZG(0HuIXc+9i4f8qRIvEF(=A_)Tj!iMKvHjsskBO1IUb;
zKvvWOvZFSThzcMXl|W9^0dk`*kQeoU{HPBUM1!+*7YptBwcGyF%r{zO%1(SC6d6ZW
zi%dn<icCY+i%ds0ip)SZi_Ao}ip)Z`i_AtAMJAACkx68y$Q)$1$XsNv$UJ1f$b96W
z$bud}ePZjD8{U2F;m!+T$%-W_yFc;0_Z&VaJ+@~>=i%Xb!wb6~x4W05E{n}xih_^b
zN#*kHw)2#xl&gT!%37eTa%+!xlEaRtIP7?u!;WV;?9`+RJ2j=kPED(@Q?n}Uw4??*
zEv3OuOKY&xvKs94qz*ehrNd57>#)<aI_!+30XrjQz|Ke;ursm-?98MIJ2Pd%&P<!I
zGqWb_tfU1yD`mmXN?WkAvKH*@qzyYeWy8)++px2<Hta-Fz)qwD>_l3?PGkk_WKzOT
zrX=iSTEb3dCG4D}13M?>z|KiKuye8w?A)XaJ2&OR&P}_pbF(h&yrc&^FXh3`OM9^M
zvL5XGqz^kk<-^WT`>^w~KJ0>I0J|U+z%ED!unV%m+1){{@46kYJo%5#**Y+HAl^Mo
zUs$$w!#T_Moaj6}$OlFDVs-cIBeB^F#@&;+d%?K#@XXcX?!zaddmo`A`({2;4Dlhi
zYjv+555#7#8+Yx_xt~#P>=xz@@(c%3Gb)gl(SY=f4rF8uATwhESs4q+&e%XABY<Q^
z0y!B6$j!JwUd99RGd@s|3C>dIy7;E`Gw+n+o3rEBbpED;@%X`uR#%U|6B(K}Fh9N^
hG=Jvs|86t?J@fzDyzUpFP`vl$g5H;B--AP;{{jWCUM~Ou

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp.74A0C6504550F5D4.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp.74A0C6504550F5D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a3db28d1a6184af53e06a3d68deeb96eaaa83a6e
GIT binary patch
literal 1564
zcmWIYbaNA7V_<Mj@vO*AElFfyU|<jg;^LB`%tB@c1`$pMhKf0RLvQBYwi3AdUnKdM
z@2uHUADkcET6<B=aF;?qi*KKblgP?RQ>Op^DjT)w?he<6C9aIokIp=lIp>wiAlY!b
z_;Cx<!&FPXPiJo%PdvE$q`TeT-TQu;JT>3^@M-6xR}*cb=j6=LJr_S;Lr>pNq+jz{
zdHm!@@9uticDp*N=C{$)?(^S`>-Ns*Z(s7@&*7Sb;`7(-Kljf5{-@W`m;UhmFqm0#
zV9wNaZeeT-T`p)}+hJOzx5nW10_Mr<@?r$MoA)^7U)+$ox;N>T*JZ7m#=oa^d*iAe
zZg+3{{h3Fu#qi6LyQa51w6}OYPdFmv>#^wjx#i!v{+dm@CFXg<#s6saiMi`tjV^_`
zb7T}nMYDTP+S2Rr`vTkBoaNKj$L-zQ^yJ357cR`aT8WFN**?*c{W*D~vY=?o*1k<F
ztvfS({TFZesNt#-vC!mDw@2AR_E>MvhGV;L9c7lyJ8^tPwo11^t@h&2iTr^jBED<=
zQ#m%i>-X6!Q(w4ggD%JGH`~f`8~VO?NaPA;JzOd0{`CI~*)=WKoiaUFs0994<ooo6
z?5a|ZDHSJ7rC2quM@jg$sav#4#NIKx@mkl_P1BR_pnul&B_<7eYZkYZ?Q+v7jcI4t
zc_D562Z;;oJ3O_Eye3Q6@gKPs-4*QU`&99IkZb#4KHjCDL+W?zX?UT~`^wMz`M*iq
z?~88TI<+*LJNsd7*Sg~Be%6V~I=5}y`EvzW-!+;|SiEv>YiR7T?>B3<C@0;Ocs0Xs
zr(57bt!0^>p06HzSt@Hj_%60~iC9t9@k{LvzOT!FoU4p-S<kIM%P4Tmw5FST`@aTR
z-&{N))v8nDl89Z2T&hFweEnyW?-qNV4_Wo^oXRwl><uP<VJvU=wlAF^TB>fDUb6P_
z4;|*|j>6tntr{zf)_rZ!dV72guU#zDbn%|FBVsufh8EHnrh0vqUcT$^^n+^EH*2@;
zk8YZ-b9T??JEk|zDqc5papX*Ky}XI(^{<=lHvRK7V=_B!*8X&p)T`838QP(rzmD(!
zu18a!ZhgA?`kvSSnXfvS`5s&Gct-iAHuL7s2PLDkPJ7+;nCYA6+G;E{;lqm+&8Dfp
z-~TWuxcT<(aaP8ckCuwpH+@_6-}a!SQDm>%a!Hw9t7{iC_UD_ND&opNaznl@WyTu5
zRm-`*|EZA`3Ju!$@50^sTM<`WZt}O)IQ(NMuFOrc1Li<3+fMmUN+o}I8T<nog1N=G
z6~F{g1`J|8_Fu1TUu46>5CD|sW##n%6ENwb)U@Iyj0_BH@;myR7n&<D^91t?^NYcC
z1@lYrOC#}Q`4wP1n8i$7U<-Q>{a&!<p|t`}u#k^X7+le>1^1f0O@l$^>hW8`gn>M3
zem5kZ2Y)b(=h*&+;pdl<TZ}BhLc&7cFyRnCA@vs%ZCDw3g84Z4_+i2@nvo0aFPJNm
zKV9-I6^=Iss%B+p6@huT%-iLRc7#~~P?+11I~eAL4U;&xvq`U?338<pzY5GuAWw~7
z1Bs`_uLI*93bA_`FMs_w&?G??K_!^*xfAkVLOilvfx^g+zc&5A3oRk*UZ5~P2fq+Z
z^@GO^4ig18+y*%(lsg{goTmn?U)cG@yMe-dqI~i&)kUc}i6D;@7bO>qgBU>|0t}dg
zBQq1<IqtkSl`)7RIX5LI11RjB=F7~&%Er#Y$<4znz|ZF?=<2K{$mQuIBrYr>A!%W0
zZDVU^@1Q5DZKZ6cA&*T|Oj<@(K~YIXRb5j{N8iBE*wkFkL`v7lMO4h)D}>3-&(VX?
F2LPr3jO+ja

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp.A977355E082D2CA2.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp.A977355E082D2CA2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..97a5cf453ca46f2ed423bcebac2f093c803476ef
GIT binary patch
literal 6034
zcmZvg2~bnl9>#M|5JOlL*+fJTjpXht1|w($p#dvK5E{2e8Z{L`TZoFCR*hO68{DW%
zwc2W})u+5l>xN5tU941Fb+k@w-HJ;L2A9^_rWoyeF>fyX-^^o%!~M?pefiD3oV%sC
zAV2?|1cf4ZYC&z)G<$i7LZOHb{MOj3XIxh*6iML<MeUX^%Bsd((Z`<s@W_Xsl$5kq
zkKegqQI75a7H7R_y?r@-MQO{~gIhN|xb(}d<0j+0^vFRUuX(bi#q&%vP!YE1j8JtX
z<iSY(%>I_u-Lo9dpEo{!bm!jF_GxbnOZhOO_b#@r<CWLSn=|!&xuR9w>oyXf&U!H5
z)}5P=p8C`EZ~lAkCg#B`&)uJIR_F)iZ(h(m;bZ4MoA384$J!s?+nhOEIqckY`n7M>
z6Vu47jaT<(?7bH{`E=x^iTko*4mC7*My<T7-a6vUhJzRC@d0-Fx41|9HeRf_+jQlD
zbhV%6*Dnu#qPxbN57BR5ODNB5-2K+Tt^1k3KRsWSyZyy}D_UB2*Y?;nN3%NT^lnAJ
zIr%ngdBHjXYwOo%U&D-_Z+^3U^{$@DGna<;-EFFlS=W|2^Sj$(`KW<~zb3_v#CnM@
zw->EEae8~JW6Fx0nil1$j;I3}8jP@wO8Vg9jR76K$E_H6W69pU#<RYjc`uw_S7BVz
zW13C)v-+nRL6`BG^~|9I2bY&Uj5>chb=}pm*>itLX`>pBluRN<>`tvec6^n)tuL|O
zx32cP*orXz(mONr7Av>pL=3JZ@BOtbZO6Ki%Uct#l<|?<s;&5Tdf(0B2WPM6o?QPR
zF)I6L#*E_nEycH&wL45hb6+pi{v+%5god`-s~<HtuJZOsZ|O;$7+cjebsjr=I9X9u
zc5+rkWs}Gsb=~jxdseR5HGWxC#krb$j?3@v_-yNuwN)qmmCa4jjSbSA7r(!nYWYfk
zV#bOB@6C6H3^;x3Um1Ny&{Ioy><o8oc4gGw%1T~ba^zNNbm_%qi)SpD_eXu?{WZ?C
z32AY!t5YVd?zQdYx%ij%-+8Oxp?S|NTm6R2nT;3A8dTmdmTMA7Z&7v8-r~3KjzEJQ
zMC{bJ#}&IP@2l3I9zFj4<Vh&KrEdC(+m&x->3g3q-1%g*_1fPBt^E9J!`kOL=Pmwr
z(_!Xd+@YHHh+E2CA)g=LGk(>yjw|^K#$><SREB=6ElSy0l;;hvh`X?{bk~@LZX#pv
zfvb4$6?+PM|B+Zzw(eJ7+pG33H=J(FDa{$Ue`fR7Lsq`Mu)MWkbM@2s?+3j9LqpLW
zN&EPQe`Fu#=<7|*9iMOOs5>@q#Qyn~%!!`7YMbuV;iZcnCzOpjW;^%sm=hI3;-DH~
z#+=cEo_%p@;QXym>V%B)d+UGE4JvzFr{443mAJ?qmHM44_sn?j;!E1iiB+xh8uz@q
z!1w;nQSC#2ddQtQd?;%6XS%<3rCq*1{@=rQpL=6&%Cng3$G`cYExXc{_f>WOHqv&m
zMR6;4Lh4}2|0m8aO#a}<Z&Wb{W}iC0^ibiWl%bPRZO#0uR|aDWMQBRXs@(0m)NoX4
zR#;*}W5R+%;5rce>3(G9V1Dzk=5RDVVD}I0KQuVVw))Ao>q|yul!T*k0Xs1?F*-QN
zHv3QSnJK5nPLEK$6tFY9Wey7tvOV1Y+x^{l=I;wfWdS=OG@*NNknQ^D`I*DECC!Xf
zlmzVL_~g{!Alr@gVTVJL_H7DBV*_?{XtXLg$o9pBiTCHO&3YKAC=S?3@k#xIgKT@t
zQ;*WpeK8!33D~irv0=eMwx9JZ`~2GM6L-RqDPZ>u?HLvvWV?FW^qMyU7fbo*C98u+
zZROYgZ4Eburbbw_8m(47b^)c+>bn%K!GVYCaNrRd0(gXu03N9!fk)~{;87Y1c$AI;
z9<8B)N9$<dF&YMVjE(^wt6_o1>R8}$8V-1zjsqUA;ep5Nc;E>d0eFH=0G_B3fhX!j
z;7J+@c#=*6o<2q&-?`E(`ec2otmI>^R0mgW*S{w_Cfg{7A}od&L#L8gG^q}*<gSX=
zrmdDgFYq6=YZn6Lz_fK(m&R>4$l`V!WN`<E!z^JVK$fr*AWJwf0%l1Y39_V}1X<F7
zkuXcyD3GP>6v$EzjDlI(MuRMEr$Lr>U^L7!HU?xFI|H(e17l#8wXq<}+F6ig9T*F<
zoQ(rn&dz}>=fF6a<!wC3@^&6%c?ZVBtY8yBR<H{oD>yI#W<{F_vZ7rCS<!)sFe}+4
zkd^Eb$Vv`Of>}MIAJVnl^h5ReKpFKVvXcK|>fm)!UVQR(a*{fDeUzsItI&&S+rhfi
z`Z4WOsGUe#hq~6K)`hf}p!Ohb8`M6e{R?WhQhP_))yKUU4tcmA!yylMA{_E?7s4SA
z_aGeda38`U4|gkZ$RoTM0eOTUBOs4(A_DRV7a|~!@E`*62p=LKk8mpq$RoWN33;R+
zBO#A;A`<dQ7a}2#^dJ)QNFO30k8~?Z$fLX%1$mSoqacrRA`0>-7os4K@*oQGC?BFA
zk8&$1$fLa&4SBR5qalxWA{z2&7os7L_8=PaXdj{>k9I3*$YZ=119^-eV<3-lA_npp
z7h)if@gN5B7$0IFk8vv*$YZ@23wf*`V<C@qA{O#k7h)lg^&l4VSRY~`k98|q$m6^i
z2YH+y;~<Z7A`bF67vdm~^B@lLI3MC5k8>+I$m6{j4|%*F;~|fCA|CR17vdq0_aGkf
zcpu^+k9RA1$P>Jn0C|ER6Ch7;A_4LQ7ZM;(@E`&51RoM0PjD**$P>Mo2zjC(6CqD@
zA`$XL7ZM>)^dJ%PL?03%Pjo9q$dkO71bLDllORuWA_?*&7m^@P@*oNFBp;F>PjV}z
zF5Zac_Y@tE?UzxA1wMa-{Lk~pXkGp6c)D&47GX)uOw5%ZN5&QPN^e`6)h#eRBy30z
z`9@)VbYb+I7>gK~R(Ds&%8#d0$%UJqT<f?H8MsBLv?<^MuEK#LR0L3@iUf*MQ9#iu
z8Yo7^0L7|Upg0u=6tCie5>x_EqDllxQc0a^?x>MJkmkWA6k$oqNg5|V$wyd{M<kcX
zZv~FGnwEamdgDl7*OSVUCd(n&M(W_pLf5_0P+;i1&+ctsGJpJyEvo|Ew6L@&`S>){
z9_wbR`Y1mzoe-4}FW+|M7mZAH@CBpmBee!?j@%6#X`{9fD3i7rD64i_m%<G==;1~j
z^l%dndbkw_J;FeM9$_Rvk1!FSM_38aBMl_zkwy~qND~Qqq?H6c%0Pi0Wu!olGEtyM
zSt-z?4K(P{MjG^J6AgN_l?FY=z<?fOWI&HGF`&m-8PH=5Ea<UD7W7yX3wo@T1wGEd
zfgWe%K#wzVpvPG`(Blm}=<!A#^mr2wdc2hfJ;5M=o?sL}PcR9fCs+m06AdEhiAE9h
zM3V@5qE!Sv$smEAWRyTpGD)B(StZcZr|AdESC|Nkevn=Z6rrbNbs#;gbj_LVcLV23
zbVPJd`8;Wsil=tV`=~uI9ixe%<#*(-#|^t{Nx9S#IIFWlveM)aiLeY08Q!Vn(>QSY
zK7VEJI@hubvd764*Tyf~h+aci1^{}MzLz}RdTWtkZFJ+6z;tqHvPHfUxVdhRyZ1<N
z2BwF^4M~yH@)rO-)WH}3E|Qkf^59%$aG(f=0E%QtpeTj{ie_k_7={6gWmup%h69Rc
zc%TGE07_&;pd>~DO5a<bB=1_gs$n(KH$K0I#A;$Wc~9l<G@{hO?=rw1o3#_c9-DEX
z2r~f`X(oZ9%oI?xnFfk6GeEIs7AVfl0mYknpaio3lxP-#lFU-4s-9L^{`_`VQ*Es&
zdA>36`S9PL$?}FmF`;X7Qs<`9BKAAFxp3j(S)o%VOrALDWyQ;#o?=SZrnPEn<*NZX
zYWiqpNVm}Ls<7~g$eul-3ZnCJ2Suw3a(X4i#P&|i95OU3d)ROT6G!R?4A!RnzuEYt
mzR9Wm(*|nNae`twUd+(-llsuY$hi0y3d>P`(Fh~<;(q`wo?L_g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.2559A4A420F11829.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp.2559A4A420F11829.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6bb143ea472cae935041e38849d25695217489f1
GIT binary patch
literal 6900
zcmZvgd3aM*!p3t?x}+^@D=k)_NOSl7-a;A($P!@M20kbO%F>{0VXF)!4lv^k#Btyo
zQ_uh|jDQM?B4q#(P#6mX4k&}W69j<~H4OyZ2FCzVW<tN`+V_ty&(qWUp7+=1*W5KZ
zwN)cVTrIWPhRv*6+Au40UzW{g>zesl9BQ2VuH9zKDzMp>ZhT`x&B9MByTqP3Gftnq
zapvCbR9WGP`&Ty<mXC^V9lWmmkJ|8gRVUY^PJc7z(aNz0R~<Z*owdO%OZ@a$(Tc1i
zvp4LYN|tPSs`bpZ6GfYkymR%v(C?N!@YAf#0rhx^{doMYU+*cMTG#yU=-j5rHR_9<
zkHpW<n0fB}sa=PEyXDw{UDrbY{^j9c?3(cR%TINmdvn?A)f1=vxmQW5^MT9%kQcS&
zKDhav$@jjUKc$c8dGg#2dB@dG^(Q-?{>?5=(f;P<#QfElb2bfmC${gywS0L<{?2jZ
z3wqz=0bl=++c6c)^DNqVdDMZLity+;&z~H8t?#_Dszh_-@RU$`47YCN=k7}3!noP>
zZ*P6f5LS%b^ud|Bx}q0+wu7sVi}NS%zy9sTFKF+H(&wmLVg0!*y8m0H#*OMv3jbL6
z<LK<ME8c86eQ;d)o<?qI@47ATpZjibV~f3ERMm=;XFgn6^1yZ9pq%l2s;+-FuWR$Q
zf!+_sR=iO9;b+6nY-oC;#|8I>r%PYkGke<Ev%_|zR?L~zrTXiadgZqM^OCgS^>I^$
zAush_d+0A~;+J{|kEhlzed~^SdHU%`<_~|!zHw-Wepkh-*H`x0y576$vu>YG&^tcc
zID>yq-t~EHuXisDJO23BcXk?hP@Y@6c4O_AE3bsd3>dbkn%nOAa&q&frRTP^>|T@X
zT$1P_y*akwwVA2<3oR2rKD(gl{GFdH3QiC0{B&UJvX#64ythNok2fbivF~F?&KR(`
z{L`k?nFkk_-#ck;@~J0YAJ!6X_}`BT*M0oAZi`2(KXG_o#Y2<ddU-@m3F*6Y{<p`M
z_0M{-@x+ckFTd0=`=cLwCF#4)-d!^MRj=^*9fS7uq6e>jX6F-oTL-?lL>hm<tg6}A
zw{E`JFYDhIj{Pap@d4KT;`r@Hzjsnk5AK%#%1~i@<NjYgTpi!m{DY!|Ba26v(;NE@
zD%;dF=Y`szw(Yrpx%1nNJr2Bg^Ze(V?p@Tnv!>w>U-j5_^7;`zEbW^2#EeO|c1O#`
zz306D@go(RKMHOi7rpt=<Uf4B{qJ9&S%3NX^YJ0ScD^j!_wmeAlg1wXX6xP$@}`K#
z^2g^zPu+hbr=+6j*6JTh=2n0E!R(R!HcZ+i&G8P3ULEDi7XNhY#L+YT)wQ{YYFe0z
z=P&Cc9_-U^-P_utbyYj&8<)Qs#k{>JX;z<lcu)D^S-u6A?(F7Y@yx`ErdKXM((mPM
z?%HY3&1u~eULGDleDBQ`*TGkBk#x^(H!j332CsI!HQ0M-;FW)_`q%jPT1Nh-VBxKA
zHobr7bo0uG;zQq0oSOG2b^OIwr3G_}uhi83ZonJcR&{0O>{<I{Uc+NwmZk2w8Ckrn
zVfyvV)tyuJ+L~d{xq26naT9Fyg*k<JZ9}GMd*$DJ_QLOL`i?6gwVAqmPWP^DgH_*?
z{_d-nT9@rAApT5UoKu|JHdys!xo%%#*1y(tu#L~uu58z!w!x~C_w_y~SA1m_kpIcl
zML9)zZG%;R$3ML!vc5W0KqhADo;f`Sv<+5${cri*_bj<Ey`ybxrY<cmy}NC&>Yuty
zc;)=UH(LwHm`vRzr%PVjVAYMYW-k`H*=+X4-wbdcYrUs;hkB>qImL2jQouQ@J<)y|
zIyyi@$M_lOm;eJE>t~^311xl$pM#DIaM1C79y&h2LnrtJ=!Ad(o#+>#69Xc2l3#*O
z3P{k&ei=GBAVa7473h?J0-fqtp;H4Ybedm-P77$z>3$tLJ)lEp_zmccfB~K9H=#2F
z=FgpVBA4g4oxrP)rk`!TIN<hSRh)YJfpof@{o50*(9qE?8ahT{pkrJNbgaTc$GTYP
zIE90bb8*n|3J)Fc;-M210Xo4YKqo39bfQayPEsW3B$ou8tjN&GE*UySQJ_;?3UsQX
zLZ`Y^=rl!xPIGC{>52}W?$V(%6azZLWk6>tCUmCDgwEaFeb>)7l6rSfcW*0MH>ZmH
zwwn?h@}RS^{UAF-&gGE8l(UIy&vcLmmJZRt(qW25mI*SzG9d<7CQLENvOyMDHpBwU
zhA9?VF317Pg*afjFvTIu2YF!m5DzRLrg&t9pa8585`Y!Llz^-l6oC~(BCukZ5|NdH
z60lN80#*uB60&ko238Kqz{+7tMpg+bz$zgHSS3s;$f`jVST&>qtA;5RSuLmmtA#XR
zwJ@b2s|R&p^^gv%9;S3;ji3Rn5i)=^!jyrm88m@4Lng3hm@<)d_jeDsPVx@*?vd_V
zNPhP;D_M`<Mfq)yUAO>}l=BLl<jyqZ{1J16IGafOMc|AP=V{Cd;=F`8MV!|$$L-Em
zd;8zhNs2}vou+8?(GfzUkB$)<eRP7*=%Z7FMjsuw)97Q86oWn{O)=<WB7{L76C(`z
zm;_<a$D{~@J|=Ew(8nez7JY1*V$sJ&2#Y>8Mp*Q*3Bsa}O%WD-Y~0SGk4sV<`nWX3
zp^u9Y4t-pVaOmR_ghL;fA{_d-xSc~EpQL#7@o9=jA0Ht+`uG^((Z?qUk3K#{c=Yjc
zJC8mgNeSo^(v*NcAwmT72{9s|Pe>2}eL{)|=o8|00exbU64587DG_~Qgox-9V?;!s
zm>?qh#1s+HC&uj}`lKW!p-)Ov68fYFk<cf_h=e{VK_v7^DI%dyirXdh$w^8^pPZ&-
z^vMw-qfd?z8GUks$mo+(L`I(+x69~Ll9YlzB~2;lQzAq`pAsVq`jiAw(5Iw`f<7f~
zSJ0;>DHVNcno`lHMu>_&HAYnQsR^Q@PfZaOeQMmUqEAaw8v3*}rJ+xY5Dk4=jA-c7
z5=29vmLeMZw76YEpPr<2^yz6zN1q-cI{Ne&(b1<Th>kuzMRfG(al4K_BS{(PGt!iS
zJ|jX5^cgW?pwCDU1ARt{80a(Nb_0E8k}}a}rYRGBW`vmNGh@U=pP3*g`pgtD(Pzf(
zX1j05s;jme--pgpS@i?U2i93n%Tcbi3ybBZ$EXhV-7C8fvpzlgM;Ee}qMq!`bY)&;
zXN$IeQ~7P*S*>U5OmX>!o9AzQ*d_DD!r?4~N1%g-#5fp8tb>KbIXFnXgNGzI1W2Mo
zgd{m6NU}qQq&O5vszZgOIW$PRLx*HI3`nNK{5j1UHRO%(o^8|3%p3JMhqk{_r`K5x
z$>*$XPqc>ylJ?R-(monU#=`)~co`rW9|I)oVS!}5ERd{^1(Ne{KyqFVNY2Lr$$NMp
zc`pwn@8f|KJOYq{R{&D*2|$V-5lGQ10x9}LASI6kq~w)=lzbA9vPT9|_R2uYJ{d^G
zqX4OR6(AL#0;K9ufmFRJkg87wQuAm)YF-UU&8GpWdvqXmuMVW{(}6TR29So=0MhUo
zK$;#CNYiTqY5Ghc-F@7gb-89fE(>lAlHsnj(!!ek6L+VlhBH@ser|qd@#p_uzpC+`
z66^Yn$xN3HE%jULO5xwW_V8anyZHL;=?SIv*7S}mzh5@#;>I<Z>0WugI$7s_^U_1v
zV;Z)M$V?Y^DlWD@nYUEb&deUZ<w|C{kS>(1MT(>TL6@hMoj#G7_GEebSpQ(`$s?D9
z!U3x@({6{m(3*~G+27}OE{SBOE4x*eS<}|?Sz&(LvYB;<%3QCsle7M>4)rumLt<zK
z5=*m?IGTgR(>x@B79feV2uY$PNHQ%$QfLK|N~@4GT7#t1IwXTOAepoY$z9|wwsyi5
zNAshk|L)s2^&+}Rv-XMg_DO!*nqm9C!3TnMNW12@tsJ&P&M4*F2w#`zC`Dstq7;Lf
zjZ!RTE=qBj`6$I>7NV4ZS&UL5W+_TZnB^!XV^*S+f?17HDrPN8X_)mWrDHatl!4id
zQf7O$R!%cdzT3;D)^2H>wcx(n7tG?u8H>l=-k5ZIXnWb~m!IC(J#exA)|^1ylu47P
z+NM~Q%~sdGIk|mvYWrrOeRD>`%ms5Yv1axtdscQ%t|PCYL&q+iJ5_ZZK8)#_Up2I_
zr09<BcUHOwcm@u-XRy*u>|frGEBpUui|^{uv-jP7oHWY|lB(%uUstb+QrYl!D;`-r
Mjf@yQ<QJ6hzq9L)v;Y7A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.513C881E10886D27.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp.513C881E10886D27.idx
new file mode 100755
index 0000000000000000000000000000000000000000..60e3075fbf0455b4c1c4b9b21b1de5bb34beb877
GIT binary patch
literal 6900
zcmZvgd3aN0qQ-N+bV<wJN{baJwmJKLPH7q%5NH5PQwB=dX;4|hR=Jco!i<i@ao~<A
zXaE;x02LKQ$^fFEFaj66pbYNZAnS;l1_EvaV}J;AL+^8@?;khM!`JhE?=Ss5%{gDr
z;cKWJKKyct%~ms`c5&0p(EV99o2^IY-=a|SoR94`TUIBVZSlJI#@8)4UDZAI!pX7v
ztaUSXZl+2Ljy$-csi3?*`tsn_<$u(M&8<DUGIipI(N9&4*|U7l-t4TkW@+N)(}hd3
z4$NA+Ybtqo<MVANuN*1daNwiMpM-w*@IyCeZV0G{i|vQwcl~BZ$<&FhACJmyiCm$+
z+4^Ms?DQFD&K}#g|Iv+yc5k~9`saV_zjxdCzg>L3*PI(m-W@q%+F$w>r@HOF_&@T(
zeYuZp_-OKdALdUf6FtwJ*&=Vb+_mv&mlMC;<|*9O+M1ZR;$qJFAs@xw{(Kc*9+H1@
zT>plCdvf*nKj(Hy1@k-$w_dE@T~`qvHT#vLgRk_TTUwiFjqINiN{{ANkNDbMC7c^O
ztMS8^pEiW0Bi4U*a^l3oSADiU%MXk5ChxlT<N0rB?~#(1s9a&qnJjw12PMY!kzW@4
zvEY|c*<+Tz-*RHl*zz6C+~R%{H-38Nr@_tp>`nExOOKxXd|B~B*L;I=#+B7x`)Y2F
z)++<OpN*+_wdC`!YEG_gd9U|5_u4;~yt8B0v{R>QwxpKMp4olm_pkTOZTtHLY5u!o
zrwT(}@3(62UsuL2^cJ2;ty%oR9dq;a6Hm??_PBlB(9V^Y#mm>0mA$;iyZo!3r^oAE
zUTmJuza($_x}opK=V}f=^ZlJ&2kw#QG^|?J@a?io;nCGK3rBLBJ>O1ly|DPq#(mpY
zCc70Ux=ZhmX?kZys`1>u3CB;(Z#jGCmkWargtnd@*tTTZ_P^}x+~@d)#FzHB*%8yL
z7nPrGNu7LTQTct7<|Lng_T8F&;imumqG0v$|L(bH_?jd8=T<yE`GYrx*A<igTj%|F
zc*%gQ*P4%PDSPAfF4<rF(l<%pb?WZoVQ+bbukRT2c3*n%iWjy%yR&WJYY$7~cAK?z
z>-tZeCst<t^W337MY=r1x?dZ&`QTM2_2<Do^WPjQY;NB5>nBFWH?{t(DB;MW;pPL)
zm4iyxx6FRE;pV0t4?gDnab54-pWHb6_4@l3wr#Cz`onj<HyyopfDcRC<~}=p(yi^$
z(y^a7AAIIO#fC3}o5x0PJU;mkS2zFd`;%)npL``g<TuVYg!_-rI5ug_!5?1U`B~l+
z@lgJ_yy&q9ujdq36y93#bMc&!KYliAMCIB^>!sP=LD9?gu59s7hmIUPIY3>NySHv1
zQ}N10efT3~m8(D07Ot+{GS9g9Lp}51!lXI!*b_U-_s{grzi?+y|I!yGRJ6Q#@yW_J
zHn|(7y)?URNBFVuxM4eQ>~rmT>lR7(*>wF}?0oQYms^9qdk0?n`|^K``()pUe|1`L
z>xcE9?mf}E?D6=}PZP)HK1CgV?Ja5k?4nC`4Zo{?Z`1M~%<LVjp37@``rFdfs2h<*
zOPU_ImRa37rSGh)dCAqU6B#?+)>x2Jkk>wBn)X-zjTg`TzOMh+PNX4I_sZ$jqkXXI
zd(uCB_j=osZJmfeQy1kF<+cx2J#)f>l6$WXeYmr2T&8wqy9TunR=s{z&CN>c%knO^
zF`2rgsO0YU!KzOO|FO36_@jYNWJ0FylhdcVeX#1}{r&dH72lbi$geVWVNPLQ`(V{K
zyN`eK?1J~(I+4+tx_eIdy!OGWn`h2iB=oe|?9IQeb{}fHr(fqrr{6ina%NJ%IkO|t
zei}MDKtsp)8R(b*10CyUp<@Fqbex}qjtg+m@qQjUKEOjK_yy>MfB>E77oigaB6N~p
zf=&uZ(8+!oIyoRir}!1<lz;-A>Q|vt11fZyUxQ8yXwd0?9XdUrLudF6=!}2?o#{8B
zGXv(oJL^I&$Zx-ZR~$^g*m}PD_GxuF2f&GR(U2Gy1BrF9kT@3yiFfgk1eX9wbcv88
zmjp?6$&eJ60!ekLkTjPDNq6ax43`1PbeWLcecb)5g=Q_+iv0Fv3Liqy+1&9VI780I
zAcZMs3)PY7APp=XqJgEu6pbtsWPoKt46sa?VvuEnEU;{d1(pp{EV5jX1C|SMz;a=V
zLzWNn!15s;SUybg$O=ILSRo_;D}*TlSurRAD~3d1#V{ozD+MKBrH}-y6s9C(<)93#
z9Fl>R!<3Ay5>$XyLJF`-m{O2cgDS9UNCj36Q!27rPy<#AX~1e>N<&r;>cHwD9aud~
z>Bt&E16U(u0BeLP16ea@0&9j$V9hXPBI_RD9%fzSog3XF+zpWY?rB!CZb%FB+iyg$
z0FspR5?tiYH0AsSbA&isNXH^@#)$I-<^*wGz?>q^YnbD9XPdp_{dAI|(MP8#8hvzx
z(CDLMghn5oAT;{u6rs^a$L%!wm?XuZk4aMu`j`k|(8t6GgFYrf81ykI!k~|d+Zpt+
zNs2`uo2FRwu@S<ekBt!)eQbiT=wnlaMIRftv*_cJ6o)=8O>yYsB7{R97b6_{xCG(Q
z$E65|J}z$O(8nh!9({b8;?c)P2#-EKMtJn`3Bsd~PZ1t{eB92XPe@V%`h+wkpihVp
z0ewP@2<Q_ML_nXAA_Dq^xLrV>n50DXiD^nipBNz``otI!(I+N|h(0kzMD&SqyNEt1
zNlECF(v*ZgDMBRlNiia!Pf8F8eNu`@=#%1h34L;slF=upDH(logvjWVV?;)uoFFp#
z<P?$7C&%qF`jjN4pifCt3i^}?QP8Kvh=M*PK@{{UDWafHiQ5(QsYyyjpPHsr^r;b|
zqEC$x6@6-gsOVEuL`9z(x2x#Wl9Yx%Elp|Y(;`GepB5t;`m_Ym(5Iz{hCVHB*U+aY
zDII-!n$ppyM~IF-Jw|l&=?S8vPfrmYeR|xkqt8fE2KtONWuVW95CeTij2P%M62w5C
zks=2AjJVxEpP8gg^qFbOM4uTUCi=`6G0|rxh>1QkMNIUWal6^!8?yYe?fTWwDJrXR
zVEMp_*4?t+wQ50;-10Qlxv^JOuNv#=(Lbtyy%6<eXQr$2s=8V94DqqGH_l%FynE)E
z+TkpPd!2)Z#5fp8tb>KbIXFnXgNGzI1W2Mogd{m6NU}qQq&O5vszZgOIW$PRLx*HI
z3`nNK{CApJKjgjej!h4knLpIy9NO`RI=#-3kbKUDjzoKCAZafRB<-VtWIPOzjF$nD
z@i9QM9u`Q}%L2*zSRgqM2PEg^faH7}ki3TnlK1jJ@;)9&!6N`Ecm*H@p8%xj5rGuF
zB9Nj_1XA)yKuTT-NXaJwDSKofWv>jR?3001JPMGCR{>J-DL|?o6-d>q0;&2`AT^H$
zq~_It)O;F{x<?06_v%3EJ{?HIV*qJ*4ImAl0i@|Mfi%4)kfzTB(p~1}to50BE*9Jx
zB*R^0CF>`Z-@XH5Jr^&i+cjZ(dTKbc((`lkGut8m_xg36_n0`ba&%_8WN3-s+NTLW
z`py&o_0{=zZ%>aeX|$%TpFQ&1w{NW9omo6l%DE1nvFRv9V`ichgPDy|EM_iBahUli
z#bXwtlz>@`QX*z4N=caIC?#W7qLhMJjZ!LRElO#a^(dudHlmb)*^E+VN8WPj_e&<7
zU$-*zY4y$P+tvD}XkENFdvw#r;hE{8u0=)Ge^~ZaG|b2zw((MCx_~Z_t=)`+{y`U~
zm7X|~nf7FP%B<5Ed+xvmp`dz2X4>s=7g*DlgXFjGly#g5{L+tW*;jMB6-P4ft?F4-
zYQ5LmPAkuE-$v_ruakCi)-TSSQ^l!wB$}q7qZJxDhGw8+6b3q$W}#yh7CMgRpyLz{
zI-cgC;}sq{ffk?>6ahMs7NHXr5ju&Mppz5{I+>QClNA{{g;t<b6a_k!R-scB6*`U9
zpwko$I-S;`(-j>$gEpWu6azYwHlZ^W6FPUHyU04pnaib@`z|Y8a<o2025i55`!A#m
zHR}qnc1|<j6}Pudtt+8<=KT9_e?2T}p1x@8?So0ThxV7fas4y=&3bRy!koawDU&8o
zwN0@qn{8sp;pC3PsU3%bj>G9qGv?3E#JXAa_N?rjTt{A~&Rx29>ss4mSPj!7zjkOr
zap4`k?yPcGdj<}=XRy*!98g}#mHvOTMR)b?)9>ywC(ZJLq-wg^-_^IGL^ixVi$;u`
MMuv|XaxdlkU$jk4{{R30

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.43610B5AFBA5C7DA.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp.43610B5AFBA5C7DA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d49faec7331809623634b19965a2354bd780d98e
GIT binary patch
literal 6900
zcmZvg3wTpyf`<G5=_M`qR$8n;5pwSLa|&r7AeR7^1{O-V)1Yz*x5`rD2#zxl$AKMF
z(114_Kt)A?GJq&3EQJLZ6!Go`L108p0|9S~V}Pi$q5B;B?elD&r+?4)e{b`q$$w7H
zIknXzMqDY+$r(1QdTIUa@ZGsNIXPXjKTE<5^WJyl<m9%?$yvJL;KZ6mpI3E`KYMz-
zF=xZ9JzGgd@$q|C*BAF09eZK$x;}r=htID*u_k@$yRnZ}joZI!|AD;R^;Sjlm&Zz0
z<Q|^0e(!X8|K_J#PG3D<vgz=9S3U^;@&5aMnY}5f9V>GjOWgLr?(*rAo8BK&&=|c+
zzS{9f;@r$xXV0D7dFT(DkM7%fHT;i%J#^>JiSJ%|y4$=P%ibI{X~y4rmZdxGyYyem
z;^u+}H@!FYu6K&2^_IL(o!zc%ztXYpM2Azq-{~#c+tid?u=-N|#v$*;U;lWm&?l_?
z;Jp47^ZL{Q-~Lq4Ass68F5Yoz^uC(P$e6h=o)~<!@BE7DWK;Cev~XrDzi#B0o+|PD
z_&Ig&yzrPQt{A!TqtlZom%QZ9*}v+Tv|#GqYu{h^it!yUf1VVGkDtwD`oCRnULW;Y
z@n4F69+Nk2#aoT1_K)wgyMbTYYx3p~&;BsDq1jPCx_ZTl(;u%ayYHHRQ2vD8)z`k5
z-=*p5K;K8>Dqkx9_={nu*Eb&Qe%`bGujOy-o-^ajnPJ=0E9TDbJnGw5dKR?&{i3|^
z&GFO4A+PjWd*I*KBrbLrpGZHx^zB>b7aFG?Sup$|$A+Qp`(2T)TwB@ug~xrXzUcb-
zM5DuV4Ks!3m7QPK_I&^Ruwzerduzvm`;~dMYd6$>z4CHo?0{j5NAX*|Ur%kixb*Di
z=3Q%2oywA(<+sMwzcDLacfNVjr)L&6p1bw4#i4t`J3b%SvTWt9zwK$?<I_#a&m6CF
zBWDg+(&zKW^yvqe^to%wywuZAzB#NpQvYwC6tDaAKf5j&@%ZsW^D7^k`u4UFHD$E#
zjs@QzTh>4K<%Z+idvANCL*6Gp_e?Rjow>bi_-j7#%UcG$-jf-;`q>>%?r9nL^8NCJ
zeO7hNhQ5;*Nd0pEasKEt(GK@<o|h+VJ@TWA{B>~GqF0BCTO0QN?%`31EloeEY9zX3
zgmq6tzd;on8|S`M`^%Qy_bzvRzoGlS4{n_Ma^qc#TXxjc|M{ElTTWa%EJWm;^PikK
z<=0)Yit!(~?tS8L<)%+UTgS(4JT&#sKW=^ZTfu$l*o%oF54g68cYiwT<dktozI$QM
zM}^a*qeT-6V<+#uo?lj3^6ToK%I1yw{-Ze~`>mg{QJ(7?6uUCoohLnW^!Sm}{k63P
z2Wpzx$`>ygBOdJCZ{0ik;&s*A7nqm68_mA6IAx7G`S9*Ohi3a1Uc9wyV8yeODjQ$D
z^hm#LTRgQho}b&YJF+}7VfdaK&F=lL{Yo=Ewp>3SzYx08;n%^w0|PJrebs+Y_@H^@
zKie((^}CH99yrys@}b1g50fY7KT3|h{F=OQZt3Nk+CL6BxMfurcJA)APZicb_H{+N
z`bKogvif_jW%us<iZ|8_d*0ov9UVV0r>;1^xUhA|Hm#4M8_%8pQ%&FT?PzVb?v~%J
zOY30QcVvF}=9QLZJKNDfwl2*tEodF=ddRl>COp%8&}yGEAzQoi+=E&NyS}sbO34#}
zbMrgojLX*LrRBG`4tAZoyVriD@*Ari{g-TAl3!BTI@on);De!0<afR8=%j4jBfrOh
z*1@iS={)h(bBo?;X-CIq>(2R|3tI=fZkRo1iP$wK$I<Zn0iL5Rcl2ss=L)!{*-x5H
zLD%fI!~_`Vm>>fk8(^VhgDiAhfP;<;a?tSs9y&hALnj0T=!BpEofr_I6N4gjQb2-E
z3QEw)0U0_uC_|?N6zG(o0-YLAp;Ln@bXq`zP77+#=>Z)&J*Y!x1PthmpaGp3FrhPp
zCUjQ7g3bzBzjgLPt}JR@z^jjBo@=@=;O585xcIgY(xti#NN!jEw!|<Dkc`Rz$+#IH
zS%w9YRaqceHwz@ka6ob@2PEg_faDn-NM7ZE<lQ`w0wVw^r~;6JTL4mIL?A^~1X6U1
zKuU}Rq@+qfN^S{AnUR5%RT)UxEd!}A3XqDb0I9eYAXP>MQdLzTRksSH#%Mrlss^Oy
z)_~L*9Y|f(fz;hPkOpG_X{ZK}hT8zrWK1AU)dbRXn?PEO1*D}~Kw54KNKc8U^tS_L
zou`}UHb^}@z3gNkin9y+=5ZJnc*xbzwuoI}*K$Y^;%X#qnF%q#GGPW-CPElw*$@jX
z8)kuJBZNhk3vs}5VGdX>LO5jk5DzRL=7HrSghy5g3BU?r0azhI1Z2gK2&@<uffXY}
zL{<t(z)E2WSSdmzWaW?ytQ?ksl_Nw(RtYJ<Dq#g!B|;Qr)sPCT8diZ-BSb}33u(Y=
zVGUR<LNsLckPfUK)`8U{L`T*L8NeE016U(M3}nra39K15fi)w<MAiyfz*=DoSSvy-
zWIg>o!|g@hzRok!Qwu5JnPDgUWMgqr>q!Re04d_S42#^AA+Db>N2#lkw(SJ2ICY)E
zoTRRcnA6mC4RgZbYH_svJ(D5~`j`x1(8ok6gFYrs8T2tp%Ak))QwDuZ!oi@AO%WD-
zY=*GtW22Ns9~-AE`q(68(Z{AKi#|5tVA02=2!}o{LpbztQOco@i&GALT#|C=<I<Ev
zAD3`&=;Kp_M<1UdJo@-3<<ZB-DUUurNqO|~Y09IIPdIq=2`M6=Psk7feL|E9=o8{p
zK%bDL0{VnB70@Rn90K~p6cN!UW{8MBF-k@BiE%2TPfSt~ePWu5=o1qT5q(mMNa&L?
zL_(hwr4st2IF-;RC8>lyDNQBxNePF9J~>5X^vM|_qfd@f8GUk`%IK4mR7RhirZW2E
zghNK3k|GNFlnhbOr$niOJ|#{S^eIWIpifCt1$|1wp`cGq5fy!EhN$RMqf|wo8mB7y
z)Ff5Wr>3cjJ~iP`(Wj+|hCVGrH1ugvs-aJdQw@Dul4|JF(o{pAmT+k3(^Eu8pPnH)
z`t&H((Wl3$jy^p}b@b_Js-sU&ICS(GDPo|{$Pfd4MwA-pGvd@hpOK^n`iwL+&}SqZ
z2KvktG0|sch>1QkN=@{cacZK^Oi~klW}2GlGZPLIeO8KC=(94!LZ20-7W%9>wa{lJ
zsf9i(O)d0U35V6@8?x$3&h;O|XGm_{z&-;f+uxS<Z&K8H&24{M+SlA&idwI=?dj}g
zQjEAZz?UWyBMfFXMp(>TjBuFw7~wGsF(P0VV?@L(#fXGiju9EN5+e#`HAYm-T8wCz
z^%&7H8!=*FHe<xZY{iJxmPfnSE-F<TA0zGSx>a=>W}j;Z#uRfGW8S>%bX8$hC;L5}
zB`sfn<J|R+J7?cEr>g?a44e!k*2zNRoE#+H$wLyH0wmEXLXw;kB-trLQk)7T)u}?#
zoEjwEsY5cH1|-vILb9CJZ)x`EAqOM7x7=f8|52}NXxl&P^0`Jq^1Eu=660lnWPA*e
zjGqCL^|C;+J{CyU&jQJLIUqS72PEg`faJYAki3rvlK1mK3SI$7!6yJI_yr(EuLz{*
z6M+={B9M|-0#foxKuUfINZBg`Df?s~Wxouh;#GiDd<u|?Ujb6}sz9nf6-d>u0;zd5
zAT^%`q~_Ov)V(^8x=#mE_v=6!UIR$OX8>vV4IoXg38d*Wfi(Rlke1g1((+kAT7C;i
zPj3%z@1NPL2+^ZMGCfsxT2!-l(yq+(NOqSlDk#d{0sP;`?`nJ}rOExqW~a-CmIv&^
zWaJNTJp3PDTzK>5^u+QydwTojKP{VbVZ)m2bkD+`9qqN>wDdsU*!s;QveTs<OH1tw
zmgdUZS$V@ZU(QY!GsTL1r|L*x(4`p_r;cZ*y}91r_6v+Zb@-xKJYaQp+T-*T+tUd>
z@5h2pWzp<(RoAKtd)mH9nmx?lyhCaq&o4Wh9;N+v-8`O^FeSQ8*)Nn9wcabW>GuEU
zo5Q|yntjTfw@q!RVfMniZ@yPc8fGpTe{*BX&7t+-s2g<nf#qfMm`VA;$<wAxot`tz
zu5xlFw{1>s+nnCE8Eo5}SwCyx+-$6wGun}xmtWv4Y}dX+=T04~y9^)3b}6bJT3l9g
zOSfCAJOjJ~2i-AP?JD)}(~qzC|7J^X>)xZ+?Y&(LCy2798&+R;&&qPe^mQ#AIcf$S
KF=oh}#Q#6IPNO0K

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.448FFB8EF559CCD0.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp.448FFB8EF559CCD0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..17365c980235aece2c3bb925b54a4bb89f49b419
GIT binary patch
literal 6006
zcmZvgc~nzZ9>;TUSi<H41r!x5<n5~pi9oa}5efuHAu6K8s$g(I2S{7(G8$`j+MqRR
z7hAMdTUV@7>xLsb+R<ult=5Uft!f)0aqVauHPxAzb6)fNV>pL<-uHe#IiI}bzL)o2
zX<>f;tYnocuc~lPZMAJum`bJU9{6vt)ld6Qtx|=>s8n;B4vZ<DePLiX*T?N;{iZfm
z?c9vg;*ZT*QX8L9;#@yuMaEn5@acsomio@!8Trn@QLT$x_lJkA8j$9ByewgU*rBPb
zc2B@Bult~*{ob*JHHW^vdCK;;mtTBby~ZLPO;#Uu_j;uzWkPxLcOxPj9rw`B+un9x
zno@P~(#h=yUtf1*&-Qz^-~M{=`R!xA>HMI_v`2II6^)(vd1|t+>z>Yk3w3RgudVsE
z;)UbU<I_3A`xmzgTW@x$JP~*9)$N9a-ObIO8B01NRuBEwwd>4sCc`HDs(JV`v8y8Y
z_V1B#K5LYrZd+%`p5jdVh?-AN47ryzJ+08w>^L~y<{wF~D7cnAkiA?swetA-W&PRt
z1*=cDmzO7eW>U2-KFZCg*!|${)t?FDv6Qtal3jT*jL7*qrT@dCALHMO|6@e>sQF(t
zo@*`3XsM^?^etca{l#C0)VHZ?OA6<oXg{+k`NapO!4ad=3m;sW-o5$WAmiy#nV+Sc
zxsunus_{VY%h{_wO4-{ob>jK+d0Tz+YpS~y-QJWM+40kLe&)Wi3GC2KeV6b5$5Qw8
z-t4=+m2<v+W_nb=b8pWWzChhHELL}uyZK;I`udf|#a9w9jOiEmN&OUNt+4%CY3g^E
z^NzlIyJwd{t-`d@<xQo(EV^MInVVNvL~l0yQqg>U&c$_YJC=I8CVRT^UyZ8WTji^~
z+&1>?`I(KEdj42voow57VNl21MLRy<8QbUV8qbgFT~xu8+=h$`jlTBR8ZurOH_iLO
zd;9X*?6rUYA%4Z#e<e2LuRL~edgg+PueanEC*!PbGwvRpn-jLN{@B*^Et}%PfA}NS
zOY}PbZ1V6gjqJ5&2JcEGhAjDb+j~1Z25o$qAH8QlVR2Jd`3z1M_S@wnA3EY*q_Q`T
z-hBAJ7JW1%G5U*P?B@F2FD)!`Z)pBq6zz_N`~j2eb%WDZH`aVs`glXjta;kIO}+P=
zdUWa9>KE!dwiVaDd8_w^6AunCc7FTx_oj?{vcs8Hc1k<z-9wpcez0yXb3R&7@#g)_
z-`r;Oo7+Ei4}C?ug`IS^>g2dlhwrT4c{*x5cO-grl=I}QhY`t{2~U>%o;<DS?&+xo
zx>e&=^EJl7&YLCraPGq+#}2pWNXsMl7q^j_pLX`ke=S|N;<#M5qHybs{+)M9$m4b1
z0YxVlwqzWvHqE@=GtoT%<FT2IUv$2$+p;0MbmH2Yju!ho`{?02AGPUQzkGuIeKtJ2
z?7C{b8TVv}asQwjKP~>x=u>S4|Badb<j(5v_n&KCw7@;=d(X+~@1Ua_zvO4uB;6=3
z{afyV4U4;zH7(2EkE&hvOPa6zk)vU5?c@i6qdOvP@6x=r`o1x^Y>cWhJ|aFUI0TyD
zU-Y9-F27NnRThIw19gvx9^He3Qa|Va_1312x!Yr~IZ!7>Bt-@XrQVo#%i>t+TN#7L
z2I@W$eR6|?QXguamA*5*sWwhEDp02+r92xPl-fI~Z>x}bYd{SCOQ24ONQep!O8v0l
zKwWfXO<JsKbfC@*&m0^al=^YEF<)Go{Z&T{9vP^+MRbb_4oY2LJ++}ZaC%eMznYtU
zq~p1i*h;NgJ6?Imfn?EEhZ12XKu1^z(2-^mbfkp@9c89KM_DM)(PkQSw1oy8V`f0d
zSQyZ;W)^g;g#{gF=0L|;IMDHC9(25g2c2LRKqpuP(1~UdbfQHBon)3kCs`!W$z~aJ
zvPJ&0Qx;vH=-^^oa@hY#^VQtGfe)R~(xDGstJmg)5<w84BlHC5NP+|%sV6~45ftbs
zJq0?Npg~9LY0xnQ13E^}fQ}_t(6M?JbR5Bfj?;6X;|U&gyq*W0KnS1{^aAKaLIj<t
z7eOZx66hqo1Ui|JK_}~F(CHF%Nq-(gmAXE<zDiOKdM-M6u!FD1s;v)wAzGVu9!Pei
zZA76=SP8%qHUhAO9TAWvtt4Pc8wps_j!4K-Rtm6`jRGuXM-*gfD-BrMMgx|%BO0=d
zl>sbcV*tz85d&G)$^w?Pv4Camh=nX?<p9gsIKXmt#6gy~@_^-SJYab{;vp+o1;7e6
z0kDD{36K@7B49<E2w2gMM94~339yn)0<2_55@cno3|QGF16H;p8M3+@-Ed`r#a8MH
zbfqAfbrY4O{G?2X4*ryY#pgxZ8(@KH{Yd)<%nqz=#GwVKbz$u}m_1l~9cCZaK7iS+
z)^?~vuP3~SfIh;H2<Rgmn1DXQg$d{*JeYt!!iNdyBiw2N`baM#p^x+<68cC7CZUgX
zVG{aC4<@0H^kEYENVl4VKFW(I=%f6If<DTDDd?kIn1VjagDL2ve3*hh%B`lLkM<%O
z`e;9*p^tW88v1A#rlF7aU>f>pAEu#?cB^UVW4wrgKE{t2=wlq1fj-8C8R%m?n1MdV
zhZ*Q&+-e5;STACskM$!K`d9~Mp^tT87W!BZW}%PuVHWyWx0;1M&Wkwc<NSz&KF)zT
z=;K_NgFeoKIq2hjn1epft>&PQ_aYwpct7Hyk9S}m`gj-Sp^x`q9{PA6=An;wt9j@X
zyhwmP!H)#!6C7B8KEZ_r=o36xfIh*81?UspY61F0FA||o^dk}aL<bh3Pjq1s`a};F
zp-=Q-5&A^8T7*8yizMii{78a6$$=&4lU!JWKFNb6=#zX{f<DQumY`4eA{qK*Ka!zO
zc3>I$WEYm9PxfFL`eYxLp-*<J<&baa;+v|6_ig7<SY=K|PPuY+FVQcbog_3aL$Q@T
z2KLBPt~AUe;;HLSLwI01D=Mq2LRWF~Rz14(@Jw9b%21<C1LtN90TQVpL83GiNVJ9q
ziP111u^JX6PQ!u3Yj}_ZjR2CU5kZnP5=gQ}{xemV3_W0P*)W+6ynTaqSm^C*joKoR
zOxn^=A`Ao|2_peW!bAX)G?0KKjU*sR6A4JlKmn36Qh=mP6d-8>4M^HZ1Cln;fMg5|
zAQ>YANXEnfk~Oe^WQ{B!SrZFL&cFeZGjf3BOdKG20}n{v$ODo$@qiQz0w4vW07$_k
z08%uFfE0})AVrf1NXZ}pQZh<_luQyJWrGYz*(d{2Hpzh0rR!*A-3P9ZSe*<~f89VO
z%`V<Oc87n0J#eT*M@A<q_p>h*8&7iOx{-nDlwm1m<<{E%`rd{AymEEl)9EoOmCCfT
z+;DX8hCFnU1&`IOH{O^#?rPK0z-v>ZQoAUxX`Zt`d}Qsq{J?Zlm!u@+daW(9v?_e~
zx*LJ%cp_d<Zt)ajm+0WVeaNJBB5f16d?TEQfSGh65@yPYD41y{qG4v7h=G}PA{J)O
zi8z>fC*olioJfFKbRrRE$%!PGWhatD`LKC#=ft#g#{zqw9hRN0d|9sd4_#;DbC(3B
zbsAl~GVPYb??-k`b_Axg60_2jY2|@O*XZDb4)9rt+J4H%A6rR?1W2Svf<%cFNVG_U
z#E1+?tjL1Ii5y70$b%$^0!X4Lf+UF&NU|t{r0b#UrR>@bP4hcAXUEg?egcsoD|@Qc
zC_4D;M%iQK*-c>j>4O_(y1shmq^Ik>p?*q3+0%`2Plw>2x{_Q^E<BO-AroOKA3v^Q
zf@-`{sZ`~m&5F?GgwUoXv^k}=YGzF!7EdivhlNK(YNBFd<GOY2QrLZX9@#y*a9DhD
z!ZSU34$RIq3>y605HXR<$<Wbh|8F*_SMNT3pH0^i6vOh8+;2dZJ~cB%=x<C+Dkz$W
L^G6JQ9-00J8B_2y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6CE81023C2E130B0.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6CE81023C2E130B0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ee36d671d02979ab9992e7ef39bfc95b80124862
GIT binary patch
literal 5754
zcmZvgdsGuw9>+72fDll_TMQs-=KXBd5J1orWP>O{)D#3A6?Bb4IK>wj-D=R9`l!|i
zYO7m{w$>KVT5Z)H(ygs^wcSub(b_JSEv1N$3Q|;joQ*j<&F_!p9Pi2Zem;JkJ6!HC
z=4NJOlt(HQX~mfpr3;qLQz;aR5YL~b%gU{tN`+!WkU~+h|D2`3tWDOmO=vm1cs?4F
zt)+%)Oq;XUIdgj!CHCKLh&bAKz8~*+GBBnsXx;azyJ{_y_5G{+ze~6})qgv3YJ5UA
zshLyVGV0Zb74a_`-g)|!&|q$I72f;g(Zgc%z4jme=DL<FN-?o_1GC?FplX^~IOhiT
z<sXmE{qULj#KZnS9D=Q^^U1PT5@UC+z2CGa$n~&%Rl*mi8~=0oW!<B=dzJcIYqKn8
z)cMitLybrC@94WuA3Ix8HDt<}4_6r3!MzcyHowMcOT$|%Z)LPIcfWY{{-rG8$%@16
zon)1LUU16IWgGO(>iXuyVIQZ5HMiP-Yx#ZnQ+@UINb7w~3if;IN>}L39iJw>G;wFk
z^`yM!>>mS`4JrNJy<+~OXHU-$-R?ZOr(c(IFjT+zhcx$&3&ft-(rLzPU1V0y-p$|i
zEe>yud(Ww2c4btA+?>;8yqcq_>2E0SnHHO>IsXB3NIP;MBjdNP6mPEVnEKv>2b%d;
z-kx$YM%vDnD9<K+USAki)AD)RmQMqA-`lnAz`Y55Lt4)+dG~~FMziH%ZRX9cz?n%Z
z_4tnATOHv`xGnE*wi`|loBdPkwQsq~jFFk^e}4Vi6)LzU=oVRXpe}Oqoz{^P`4iEq
zIOV`NYx5?PwYGgwyuNM6J8F9J&}muGvD=cz?s89m*4{C!w(9bS^}BAnoIgI;asArv
z<=P-^dueaO!IZhij(~6Rrb}sI>Fbo=t46(bF+cChq}9&OPr^bzx}I}x+1bIHBI=5|
ztHTeq9St9I-S5d~M&`)+{kD^Re9onT$g&G9qgKtXn6@W#QqhHd{VVqfSMH{$^YN@h
zqn5X?IZrC-C#zpcT=Gs--;^WWf5oktjs97`bZn|wm%?r+oVGkP@gNg>cu@4I;O@K~
z5xJTBmTo0xSSKu)de0iLKW0pKNWrn5p02s+qj&cvOs#yJT($A`^7|!wxi7P-SG85%
zU3`0B`x0$!jpBH%EpOe-KUy;WmYp*FRBLGEtowr&oa%V@Q0~Yz8GhLxPmHcqvrXrT
zZ8sL5yR+%#<PD?quSOp{Q`B}r{cL#b)*;t&3oPBUE|#poEQZ7Sz^35THaxlPLUG6O
zePi-hWx0>+OMUa=eV5cc(7no-vcL1bW8gtUUjudVc5CVJqtVEdt$Wu`X~1_rjhTWU
z7`RYuNZ40gw(y0+!_33%zrMZWLd2R3?AP9_5qFZ?S_^+$RIqCNW%5T`S^JH-Gh1{O
z`3omF`^tW9Dat6B9ro(EwzeC)i&tu%R;b1w>D$v)aO{yXY|-4GhSYlQQ1j!<`__hh
zvvKTOg(bw}qWx{Lf4{Ub1h*a!>)Fydvwy-rzovFCtt_3dM-&SG5vSfrdp99I2%VXy
zkiz`K0=<LhJn(+02Wr}?vc^phLT7kzw10GncaU+&(btDBZ|MFi2+i?ejlU+)JIFZZ
zm$EG<<D;quD{?(p8>t=X9b}yHYQX&MS&t%v6)$`6SiiB;yn~D<Jnp&kL1*RBAk^Z)
zQT|bbyn~Dn6%;KwTDnmTLi0U1&Oa{OJIMH>>XWVm*TlC6E2evJOk~Vx?;zvudGTks
zv3JG<p=J*b_YV*B4l;hKu^j(*+1H&x=rj)w@ec{~4l*uZuyE-tjqmQos=aq@>)FRO
zb?wRV!IDnUy(nLxCmD6=zJv)FV3-jD3>R>~a3c;FArOEOMglNWAORzdBw&<40Y(`q
zz-WO6j5gAMF#-b^V`Knh1r{*Y$O6U*9AKQ01B@4Vz<47Mm>>v%2}S`hQ4j$WjUr(B
zFn#2oTS?N#>f_}k?_Ngj-L>G`W7Adoz9Blh?llBNC!*U4GG@bojM*_LV@?EvGH%0x
zjN5T2<4y#JGGQZtOxOu16HbJHGHD}$Oxj5(lTL($GG(KHOxY<YQ%;0}GHs)QOxtNF
z(@unjGGk+a%-9(yGfsqoGHYXj%-UHfvrdGCGH2s}%-K08b54YVGH>I7%-eY=^G<|^
zvS1T{EZ7Ao3r<9UvS<^5EZRjVi%vv@vi=4AczL%4OZrrOCP-O&i=5<x(z6qvpORoF
znsixU>zQ@AAX#+zzJ!@DU}9zrn3x3vCT{XPTK{)C21~dZ2PSU8fk~JMU=n5mn1qD@
zCTSvpNt#Juk`@w}l!*c+Wu|~hStwxACK{NunFc0pp@GSm7+^AH2AGV60VZo=fytU#
zV6qk#n4F0NCTHe=$yqpH@+KabyqO0kZ{dL{m;_)7W&xOjMF6H~5`ih2MPP~+5t#Z>
zdQ9F%o-Ib`X^>dGASd}@>!<cUWWg5Oi|7u*^K37I!Hn-kaF~g`2;s|eJ)-tLoaI{i
z;rx=?`%sprJ<qycM0X$jYdQz2t3rL-QRhN+U%>1}bxknWD|HP@pO5t*STE?q94H3+
zutpSffxnNrQ4IEB^-2u(;k^hBuMc;iIPAk+C=Ra=ccVD$!|Rne>?3*+0`?IOlz@GN
z3ngG5;YJDAN7O3`*hlsvB<v#{C<*&W7fQlD(v6a^kE~acu#f6ND0u%U2TH*{su88&
z{iEC{1^cLaB?bHFUWA6%M>|j&_R%hshSx{CQ5yEq^-3D{F}(-_`xpnxz&^%>GO&+v
zqYUh0>Xi)aV|x%5u9tP7EbL<&Q5LS3b)zioW9yYH?BjY74qhMUKsngQxlj&XALm9n
z*vHi=xq7gl`Cf#FeY^wZVIS{8dDzFhQ6Bd3bxI!g2|b7a*DE+s0rm-vr~ub1xKRQ2
z33W;V_K7`+2>V0_D#AXo5fx#d=tf1@C)O!NU!QT~ZN;O$W!DjvloXa!ARi`E6LyqE
za$B||!BVs#I!*pPZANw|(e>tdKhN~Iz;Q!mdqr0IcTAo>2Mw0A=~|0?&;?5|=`j}h
z_x#H8kjwLSUJ4oPIf7I=3LGFR3?y8IgG8tZkVq8?5~ZR*qE$3VjEVt?Rk0v(Dh?!G
z#e*cM1dv3P_-9(T=G6Kxe{MhRS-Lh)E6HozcyCfnV$+u4o@tFn6Dj{1A+HWqd*69{
zr!Tz8@|{tuy>C7~iL6FQ-vJ*iQPl_qGhL0)Ff-K%12bEVurPDg2nRD?jqorF)rbJI
zSdECj{L!td%KY~IZ+Mm%XNfD2msVHtxnE9cZH8w$YDiS1{Je3E%`NtuRC~)a9fpN*
z^6Tr_8L2m=5m(Q9rV~|(qvV~_y5}EVbZGKc&$K>3A1Y7(D*E*e8XC3PGd(V1+z5HP
z^Iu0A-&;M^?3s?qi<u|iKlxR0sM`CA=(~?PgDzRV++YbaU?AZJ93;X(fJ7QdkSGHM
z5^bPCVhjvOtbqlIGjJgB1|B5AAb=zqM3D5u^jdkvw*u<6p-IP{{~U&6;i9}!`R!M$
z_P+Ik6-&DLz7^}Nx)NW)Bn&XjiUEd8IAFLH2aJ#izz8b=7%7o}kya8gN}>RxtQ26h
zL<2@!X}}nX0gSOSfUyz_7;9w#<0K9+&dLGCOFUq_l?P0a1i%EV0GKF=fQeS|PgcI5
zWZv`tZCF}fv~>FO4+YN$@0U_qvNk5$#GmN$pF3-ILE#)lp^OxY*}jjnd>`ldKF;-h
zEGjK7S>%aX3#Teoe*S|30)v7FYlaTV44IUMhp01O3XKX29~Ny$Odg+-I$<Io!6c2=
ulOz7$Y@{|OHh$zN9Y#<zD~K<Q8J7?@b~yK9dPL-3CeK4NvW+HW%6|d-ay**=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp.62E7711AB97BA061.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp.62E7711AB97BA061.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7dd5f5275e7c4c60798e98a12663a494347726a8
GIT binary patch
literal 2640
zcmZuxdpuid8ct3mBvB$r5h~89k<_hJ-7g!BY$6r+s9TK|rG(v3Md&EnY6&gf-li_2
z)utWQnyFC?opu<#Sc<M$x^zrcU8jn6v@YEv`#YRJ_WXW%^F7b=eb4v3?|Z)E$LDf+
zBn)O(6hBQIof4^p!C*AtUvi2h?msyApNzqzU2YeJdefaK!)|@o5AFl>g>;rWg;yXv
zn(y~6*5Si+t@iDfjt^gqE`QoFOg{R|^=y^MGj-<U^u?rFKH)U>p0k~hN!fd@5BzE|
zjbUwHvGTKJt#@~8*zETo7o)sqM+T0x{@ZDLgucNaL}AgNO5K5BdtVwqc>l3|puzj@
z;)nM+mWfGY%PD_%FeuHQ?><j%U6f?l-EVC9{d&;Wk6UKbZKtvYq9#(XZXT6=J9x%+
zyz$P%gj}`VO*d27W}D`;GYXEI)5ThSqD$Nn(?9O7Uh5H9E~j1}8Dr*3BUN1{QjXd7
zl5X`n=+tpEd;6tt`lPZWbb+|wZhC%s$>9p8h?6P&Py?gN$8ll<v$*SGCC3xCu?L*$
zGC!}rQs%V{3ZzuIFOgn9ZFJ!AXzA0UALNXLl(WXnPLKT=J4x}yew-X_Pw2<|6w8Im
zc&3VvG^lY=u4b6O|G9RrjFJUinUBnV<TQ7!cS)*JH|cdPq{xy-`vDtSHYBPgKARXa
z;)dpVrakEpE*fSi$B)$egiyD|nP&D}O<Bm>W&q%~NMv00sd^f>uifQxSfNE?{PgQ{
z&R0%vb)V%-PihgK9Xq7;Oei8(r%q8An+H{aV(iK#n>dfFHa)XA3O(tuH(tB?guJ@2
zsKY?d^9}3s-z2I%?1TS;yBN_nBjt;ypESvi$!2ZSf9=_Fhat&}E`2lp4YBGcY%HMR
z_1dgv5_ILv51h;o6wCKHju6^O{Z@wrB)_*or*vn4)fbij2yhoRgk)@%$7U5YjK*oS
zSo|n|dbO%<bobKmYeBuV*!|pN>J46a+Doe^*(A4!u&%*3co+YhpDyx4{BA$!wQmGt
zzf>0(1qGCGy2d>_CfEdvT*q@!#{&Gu!d?*r2fc$eGxY$Ij>3#KT0*9?qfQU6dfCf=
zc(v;{Bk7PcJw`BFQeA$yyjic=b#(AJlbKy3r;EM}T)BIu!oORqww_bozs%OPNN3&%
z;8pk@^*8h3mIF`cul(Hc_U7-VProgDc&28m(09n{Z4~}1SrJz`;fA^CA@gY4W~%9B
z|F8275Db$_%NnArB?Fxo6NIvblh4ckZ27J-vBk6{{*-m|p|7XoxwkrBK1zHvdeC~H
zcrmK?M#b*q$))@cF|_1U<9T&Ks}|OcRo(KdohwaMSsgD(Q~F}E?1WzQix=k0^qPZ{
zOXoe_=kh)iTwP+?`ggtynx0v4;HN706w_Dv)%Q6sbN*>~!4CH-b^2`0>5?NZWkj9G
zOkVDz_S7P`UnYh|bhC6KOUCRqx?>%qKjVx@&|Gr5c;6y?*9nIAigtZt#~=g#B1|}y
zKqV>ytSBDRr^?}6!A^fN;0Mb(1RV`UKxFDf+t%kRUYE!~5G?Bx^o5Fm$c%SkDjl><
zVpU8aESum>d=&wacd*Z6O5?|Hn=rnxtVh!`QUpYHTY5JmA4|VY1_EH2MxZGx0wPaF
zGMdb7W_FMPZ&=nOXc83xkyj|9?}n1v#>jvVEK>*+q9P!&BswPfF#N$o@>eD)@|HbZ
zs;n7tVyK3L_Mj7z9*o!^X9Gd@5K4pC5K3d|u;}Y~hqDY=3<M#yB#{&~L~WG_rf-;o
zQt&ty1@p0BDGEc05Na|ch0v3sd@O{X%(7xRBipNnvs_tx6cVsR2tu|4NQ(Y#9L-$Q
zJQSIv+pr+d(C{zDDuttIM8=X7%|+HC&Bc)vO+{$9loznz6?8%uu+RoxZ#qfQY}B$j
zEI5x^2Ic@@E`S^ywrvH#`)J$+fZb?(3kTNXHs(WfSP1P0<p2=c4{8M<v>((3KxjYc
z77p6*^9;V$r!McMVpYTSMf#!0HePlmNi_4q(=e@Lud@qz0|JCplkqdoc$j7qnd(Rd
zQw5w((>=C6Fs<jV7l2eSU83<kvb0A7fRj`LS%|l4IHUxvqoY~(@*ey)(zv!(q!*5?
ztx0Rf`-rQ!FilgV(U9u4+W1A`J*uW)nrcinM=rs`0N05K!(TgK+Dyp|M9$TJ{-^P6
zR3{lsvs74Agnqppzo4wXxd5h_+Dt=)M((8&NpT-h9kc|kQHQsLP>7|;q9F!TDmBG`
z<DIplX&P&8NBocrtwvH@WRx!t6m0l{-k=`}iNIhKiU9YaP#l=Bfgm1)a)G=dlnW$+
zP%cmegmQu6Ae0LWWEmrQz|}HgZ9^e5mL-BPlIVoUwQo_fWPh^%+8}r>{O<`t1{h4}
z#&FNZFl>E@3;#B?Df!YJi$22MJ)yxNVVJ$^8fL#ZDj^nzf*3EH5}u%}LL{qhqNuC!
zH9U5Gp+VxiQ#WgB>FC-!I61qxy8XpM+td-VWg7lJmPXe%U>Jd0jp2;UEVo<j*lD-L
VM$g=uqfPs(X9U0%vU%9u{{=?8`40d9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp.FC50F49FB1ACD418.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp.FC50F49FB1ACD418.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5d0c0725e97bb84ffb74b510c4f2e006c187bdf
GIT binary patch
literal 2640
zcmZuxdpuid8ct3KNfZfEgyNj4q=Zs+zbqx2NCiREtwxJd!fvP{+EJHQOK9ozGE<k)
zYSWHO&DJP}PCKk#EJfEWT{@<y>r~N>)}?b~f2TQr?D_rj=6l}fJ<s>P-}`=t&tb9H
zL=0wc6emp>of4^l!C=VnFF8dN_g@_RPr_i*u67DSeQ1vAqpkxtj~oQ_cr=EZI=g^(
zGN1o8*8crmomPEY*ZUpgYae%yl1@HzxlkqWN}c~OcO|KggFlD8=VZ&Hs~@;H0DifW
zW@u$svHr6~olkFj*up<PtVH=Nj18S=|BvIY2tEDZ3A~~|6nX>04!kse@a{wBP?OKy
zmG|$mEfSL^*HZp&uV0$A)O(54z9P!7z2DsS+s&ZuAGR%|+stNh1uev2om`4reek@^
zRP&vO2|23!T5hMhnQd9r$|yL+qzN?#1YfhpOn<+>aifoGv6gysY?7WMj#P1;PRX<B
zC)W1cYd3gk^bd+(4~QivXk1~z-SqtOl4BK)5oc03q54LZkK=^;W^p%#3J#}jVh=ku
zWPaLsrNr(S=87p&KLV|B&gk&t@zSS7KS&J|QZ5+NyFCwO>><V%^F6Y)yr3WRQ!JJ%
z<LS!2;-Kagsj5N#q32rt67}QIwWY|cM~;g(`d7uu4Kv<1LW(Td)E}^sWg~(*!n5fS
zBUWgxSK5;<-ikqnQv6t>ZwO^)oM~p?^_1n@o%(>{Hj#wocD9kq8t8Ps8dhkY7(X|0
z(dpW`?d}U6b2FOwXL(08pYa5w>eN|v!>z+AKrwdxYwI}A>(+e>ICWamV;@DW>eJHd
z!lEvHU9Z=SPk$4s^tm1RH{8XDj(IUhIQOJQk|$ZPN&ls9+Z{tuZglDEsc#8YKVf45
zviF2p%?#+`u{3lhKTs$==rD%wBo100<r4XCg3jv90~;?YzY1{YHHBnsmBt<~Xc~_r
zx0(MaeR{pBVSL}}=mfV>T<m`FG36FJJng0BlPscZL|D)8Yei@O+aIrRLiqI$`t6#*
z*e}$$MnM5(9z9cDUDIxO^BjkZQF#IU$*@<1z+s<YjZ9s@q^mHagPM@(<e=Thu3q!@
zAKmEr)kr+zM2q1rlvI}=D{s|pbr~N%MW<)gNNIvE0@v@Juki2HtZVcrA6#?OF;A!8
z3Sd|Go%A>JWt9U@m#+QX_2%|(sh!`IJv?7CTj)1p`6f#7OGyz+Dd85=^e8jhrj=rP
z)&HB^!+3+F(z2!~E74H*l?0w-`ONdOKia;pOl&i4i$81Cdi0yQcvfxq%SVZi#*bJH
z6|Y3~->TSmD!G*NK8Binb}F|aXv5s9xvE!sy?ec->Uh@+;;f#KBsr}c{o)04jaGAH
zX7!TiyBzi>yo+;8$Kal~L38u#_MB9uzGB)2r~1Cf%j|y{T(-r%N}an<bFSosa~VN9
zGLxP2s57<5^`FxtV>-vRBTFXj$i1-+(VuWeI?!Try71r%eAn>?_lov@YimdX`~{eB
z3Z6oc8JLkh#E+GuIov(|B!Cao+IVfU%pkPu>f3qgWO_Xb2!LrSo~k4>2+hw^O^UbA
zV{F0r!L%+_*GOg%I=g??`Er|63<(H=X+69iPi7FhVyHI$gf~#Hf(e9a6Gan0nL+58
zNW&JU_55xU-~-bdcnyNgAoRMr;QNuJj!6>W3)AX&b%M+wv?w|z`564cL+V!`%IcOp
zT&=7bb2L;52i-sqBteyM$PGdvhCag($qqRoB2jiSD1k&UT`mBM!Bbci%*TSIC=4Y+
zs9;D8p@X4(EQAhbSTdZD?N!1VE({I|aTx*x=`<}%GsynMScPz=1~U|Ke{LF4)+|)U
zi&$_8l@yo_fH?pv>2?6TkH$R!*o(%sIIs>UkB1hq5Lyq)1|YN^)DA#sJ*WqO(0WiU
z4wCD61>YD@llD`vD&cwpy-;KuZ`+b2Dzoq$?AEr^-ix#Y89f|PHgc#UFJr-Ls3M`u
zSV)cocodVV^G=Ge8l6B_Lyjz!>(Onx-^Leq>$>X(AmZsF^7F{jJ~9B0ivq|%Qi8`t
z0s4$i9>16S;McL{&36U5;mEs4FhG<Iy*xp5tC0P%MA_&gK{aWuioU`s7A#3srBacG
zwp;U~6g{hEVK>E?!bC><VSvkYguyReu-i<*3`8^sFa0#Nlj10W-3(<01#wU8Qe0M2
z+gbp-=~{FH#EslGC8F$>A(_C1TYy$bEikeJ9pwnJfKVJ{2cbBI27`)-!x?BZXb75B
zt|<l_?re5c!&qY%vNLj_Rf)2TjLLuwa^-u1J|G{31Yj@<MSurUC=N`JBZv*5R3IM+
zr2+{clnN99p;Vwa2&KXR8OBH*a8)LZohW3+us{$-6rB*c`7KHo9ZL4!Tm)~L|2-kd
z0)q*aFZaurVV{?{@Cj9ij2$-0=J*5qLxV%YFb6(!m_x#-gjg7IW4v(+ig+bu0!d|y
zx|%A7?74RbnaFXcY}L@z*0Hm9baHla{foJlsRLv~H~4>EDosz{&<Naa3|C}kvCDGz
Y9@}l!x=bq%E$UyrA^;Z8jg8&+CxvAN3IG5A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp.29C6735FE1834089.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp.29C6735FE1834089.idx
new file mode 100755
index 0000000000000000000000000000000000000000..511d1a6ea3989ec5441ec5ed51d83831254f4c06
GIT binary patch
literal 2562
zcmZux2~<;O7S6(&1c4xg5VrRV<gr8)71<11Oj*j7K-j_#3YMipkhQfS2*YAo#BzY4
zV5U_Tu^<#$98t6a>H$HD2N(h>RINI#7^%wQJTm9On{(!!lbip$-~GRP@4xZfx!e!R
zIGk&&duCFcC`t~8!>J>`G*N28G!chOA>(kF4ZjFxlV@$zejSsxXz%mxnx1mfPwfni
zslVmMU#x5zspQmx9;3LS)wlMO#@-Dp%?bAjKXOK)qocxC6hq?YS;-B)hF2|oO4lAe
z$s1l+er&-0C)@s`)zZ@|uSXgeZeCeW?97P#(Z=py8V$OCC^xp%M$8-Ec)KKBn`MUN
zuG%e?&eqc%L+|+v87-{6$-O~~82$H=gS~-kFH4N(iR=iW=z2p<hl@a=nREX$LBC4&
z?)E*iO~mSs)-W~CQ@Mqb^HF!Xj9<EIzV~}qvU+(SK{`Pmm`muhy|WOm?doF})?D80
z(k9x$wHPnxek*w|4HwVf=FNg6{qq+V=bnxRg0Km}Cra_DA)$(%GYO2m(PovIX>W`9
zfsgO-)XtTOM!94DgIre5aIn?;#V#?+Oo^cvr%Xy4uMICBoDuJ=j4Yf`vIKMFL(g<b
zgh@%jZg%TU#mCy|g`!=Y`Qc}|+J?(#uC_=w^j+Gst6R=JAy*Y9IRzc<t@(@~%FZYr
z*J`FVE_Z5d6iIG`WU|hlEAW^hMeH}#JJjq^<djz=6#G>~Un(3I9am-2#HF)3IX@YX
z!A5^mb5QTME5wVU^yIs%e|2e@vy&Sxv_5q`#%aLBD@5JYYyTSV*6q+`zVqSd&@of-
zljB9FC$jE1Ht4#FDnfc{s^)0NSHpRX{;uShs9jgRB<bCX7R9<z(~Gv*@%ixQxH=~e
z-<I6=bFu1p?g<bbz-X=Y?sM+UPyckTDt~gdCYsy-)ko`i<C=;Gm(EG#%|hlMv$8@s
zm&umDTeZYn*gm+H@t*feBCd$*;+XMhl8iE|rol1Vo6;5A?~Zx|IPP=z{QW?X4l__r
z<z<xVVo7WDFg@&fW_Cs1-KI0f!BEXv+lQGhH7gg|mDwtJ)iK;d(PV43zFt}UKX@D7
ze96Y>0+ci7z&PUhXVrI&8IAoq{p`h=(=hy1`$dKzGGM*C@>~Ac&W;{V<J6ineRmez
z;_vpYUmoAWm)l6V`$l|oD5dAI`Q5g^4_3CnDD#`!`$w$&$?`K?#l#ktc{(fZ#BpPR
zjo4awjKY?f*r$xdE_Trmj5B%Z(|?WR+r90Oc9`CIIhE9PnclPa<gR+Cx@{!Zai-e+
z!k1R`Q7@C}`rvz}bsD$B-`5W>maxTDT`Mn2R`gQZ+S0mNR{F=zCPnoAneszx%hL$0
zNma1Ucqgq@4}EwmqO<p}A){$Mj$`r)vu1<(1=;#!<P@cBwvylF;_+cgWD;5M8d8C1
zv-Az^3mHXErBMa$b&og>;*qoKs`g)kNnBc1QuH)((-jQ*&bT%oG$a#v0XQL5fvPMs
z2qXJQADSemJP+{61V4nPDbUnq2BjTWU%$HeTh?VVfsfFd3Yv;CgVJ&LuT1nQfBg*w
z=a0}P@+N*VgVOz)LWR>(&3rOJfY1yDhPTY1^hBL*Jd4ozK?UcF&~z<2Br_<zt|sW3
zOS$zMnc#!aY6@z~GK12%)VL#Q?#P3Nc1R-0Ix8PsZEAmHZ%7eBmXH-%0)lKHdkk`h
zINJzV0!$>Z0hmbO3^0+f2CTK!wGh^Z85pDo8=?qZSEQ)TMa2}A3}tO6hH{|OcnmDY
zL)91zk^v?d<N$0iD8>V9FuWIbK+{u%FdKHqAWv9;B6_d%VzxT}YrLG$Lc<~yb#HAo
zNw&q9ikI=w8%$A9J^?yKz!ZIr01aUAC;=M7;!8xRgSZ_Jmhk|q2l5F3s|T(T09Fr-
z5&%{YTq1&PJ&)k;AE;iOpyDY)hJX=@rtvyhm7>Ket3cc|8=5P+9sb@_=Ceu%dBn|B
zW~!pSqyKo4?EIp=pYY_|GY_)%MWX30bbs``W~Hi6Mpcii6OcnC2brQNDMBCzwou*|
zeFZlb9{sdAS3nO#=eB3|%ljss=OS*+9hzF`qN5rFlI(_{1yKHHhpz4C#mXN({{opy
zHKwxAb3f?MejZ`)a0qdm%b5eT7c?W0B)e7Hl29!k5@Yrg)Z)Q5p8DqcGc&4vIidzD
z!BkXjX^*^AQI%GTxS5(v1JsS)(H$h&-NYnugFMlukY)0Lco-yrf-xupipHP>C~+GB
zH-J$AJ^-Tv1OTG~A^=7OBmj&G3}9olV5BQ1Y=%J=uqBF+7bP)j^La^2jY;Eg4uUq#
zf4^XKfWw7s4+FP{p<6?ukiGL=UEyG4t3p^{NKkMnE^Ld##U#Zh#v{=4h!;^#UO`bw
znM_epQ{CaNe%RGSo#gI9rD^QcrrX%sI~-y=I_=gpx7%e68~i`77DHFh5CTTVOj9!p
d%RPJd9XP1JUx#JI(bW3HBb>nXc6P&m`X3R4&n^G}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp.2E316B49437298E2.idx b/.cache/clangd/index/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp.2E316B49437298E2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8e13281a5c1355c7bd1195e441292b4b01fe1f42
GIT binary patch
literal 2640
zcmZuxdpuNm8=pDDFf(FiFkuMiWX-rna=*O9yfY?)!N@HsL=5kXgi$M0!ceqbc5_*!
zrd=t!Sz)oYtS(+cuU(2(A=h1@R=Kptyq{yvAMg2m=9%C3`Tm~o@Ao{<c@Ce$VzG%B
z%<f1|nlLIQLJot$P~cy3N>c0#9Q;qhVA8I12tvH+4&)J+{u_r60NOk{Q-#bf;GM|h
zzm2hbKU=GQyQTB}XJe}$caD%wJas-_CGbq0`!IVsv6h2Bi@oP)!()*5UFZitUry7r
zw5?eC*}T@fr!{o``wz>J-t(h_$6J4M*b%O+^9O-f^p{*uK<K`g2KV26=ooD9zPtSX
zUAB2b;>2pozwLBNvle?Ul3JIOGHlu!TVCA=-2P$Pe7f~a7FW<j4ARV{y50_&vz~0c
z^B_J)c~8?fsjenl7SuBePMOk$YW;#QS);~(w5?z7<(jXi-WZ)=<cK1aoTgGvTK5re
z_1S6ExvTaKh~D&z#K-AeVZq(>yz-Kx6%OHNQaB+x`jwAjg*qm&H-vKbr>$cSIn-r-
zT7Rv;ZXe=`s1jcSy?$2z(4(=^Cq+L<^x{*_8!);&4rc5k#ufA3v(-JJAM;Yomn!2J
ziaw&i#$}1JZr;IX>V0DJG3e@IMAk!xh3kDQBE`CCuj|1@=4{#z*od-WK`r6w)UZA)
zB-b<TaVKwCH$x$AwB9F}Y7}dn*?TQzDc48`kl&UhX1V6q(^&l-PFF$;%@X2f$1gZu
zJ-gj)-hFmj4gd7y5w)j00jWB5hOD=BNC_y$u6=10>v7GhcOFNkCqD9)SFb)TsV*$)
z)Y0;M!~FEGB&A;0!~cc57~Vc7;s|FSH;GS*=dII!?cH`qFDW;w^v&eggsPvgF#yGD
z+@xk2bar1HJd+n7lpL@h#di<~ERJxA{67QpHRpi!=apahyYU)=Gqy@%jukYF#Zp?#
zev~}9R#i8)XJurZTQ4ehyYPs5lO2}!(&BLz(Iq^zd+3e4lixQVFLQ$Vx9|7aHi9wx
zRk-?r{$=jnlb)SZu6VN?`wNjL{rMB2uL%J|-a)FFT7Y3^VMaSGKGV@&qnBO1>g6}G
z-u=72XxNb+&7Ci)E<ak{tkvv1Hgt-?$f}Xh1^WZm?w+gg>rtz%cP}4Ub=5RWXWaB>
zSNNXrGx1@S15XyO{@j^Z|0=cP+p-7eYGw+3hb{h$l>b~@#8QaAX=;4LG|IY}YJA1-
ztK36)-Ne$ehDgh#!LG~kJn_<*XJvo2d{>#!V%!p!Z`pk0tJyf#t*)036CREowj3;8
zj_kWxvFB8BDd&AOEjfQOw=Qtq%(AhnM{=!et*Po*=X2tWwvZ$~trhkBx#=ps=J52&
zMUQtm>`!=Sr|9;9U2g+t=ho~vsS3Tt^mR^koBPY`|L9(_!M#qMJzsOS<hWBAK_eoQ
zo%66GwaDd{so_!0V;T`96SkC|82hMCIDH6ONKO|XScYdEuY0d(_t!RhB*0IA38Uhv
z1gU`;=|lWjIg-QO<wpYeFs*^tphyiu>sGdn^Iv^$v<2e}(^@nweW^j{?Y}91n=hGa
zSHc9qw4uDAuhby4%gWn1$wc~X65tQhG(1f~Y7qK`P1N3i7bG(h5D3%Scx|54AoO&E
zUX!WS+)fhU4b!T4Rf5zY^cq?4-Ed<21PSngX)>NnkQ#(eii%D?3V-mB{1u4Oyu}Yz
zDr-g^^pwItbI=mG0gPNhcNv1rAruF>LMV=@%A{@P9mdpQ>LCcJC6OqtA?m0EFkL1N
ziojD?6wJedr6>#~K&Z`-2tsd$@~{wkGt+|Uh#apJ#&l+KP>9PEAV{-eNs>YFE5^!&
znW~zGAa10+IHI(zs2g9xf>+U(x`c&f7zxJ`rR_$-Hrq{<wi$85#a+OHizqcP8vt_v
zR0&!Eunmp70k8*+Z{fgNoNPa|fQ8U{P&NRe^`KS&LhC`@0EE_qZs8!Ao@dbYeica{
z6{{4cEzk}@j`6Z7Nu-$;o`u~Swi>&U=KAxfhLh(U<zY92z)(R-k;--NGTv+L1G}}{
zwEU66rYBLJMU?hZ0JsQpAQRD63WMaJO?2Yez1;i1k2Y?^3bewI*qXFvc^_dF3wG0#
zX*8s|tycU<d5@|Y*iAK{nj-J_1AphKaNS=!VYi8#35eWl;Nnk{MpOqe>}D!5sfZi7
z2tbsMoUB#ow_|*V{E~vo)&e+|q0Z1nVv)I&BTDB{77N;f4u~+Ep)G_$ObsR-5tvb|
zDF*ECY*bd&Ky?SAhYYkbQ98&dT{g&->4M%MAB6;95DJBZ2T&*$jF%yZ4WU#ZZwRFV
z2_TdT6b_+OpjZf{!UUNHNFH!W4VgwLWWqE@5GE-qK4Rlrl$>-h*>7VJv|;}D1S1O!
zCPcQ}D_e$cE^%S|)w+tZ8_Lh)_w5Y{3J%5W+vG3@g^}?wFyuyi;pF7;3W@}h(iXCc
zGKb=^`!foW<3`=8s-~f7Yv<tT<m~bfGj(Ho$eN-1|GYH1wvL`YxZMEG$i#ey#m-$e
W+pM%qE#1{=|MUz8SUguYcF*5}Fy|ov

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_cshuffle.hpp.54159ABC735BA0BB.idx b/.cache/clangd/index/device_gemm_xdl_cshuffle.hpp.54159ABC735BA0BB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..13e853f1b77583d78ca7525219c51a3a6165b8d4
GIT binary patch
literal 5924
zcmb7IcUV-%7T+0Gq+Hmgx4rZh0ciq?w6IjMAP^VS7(fK`1bl!JG(IFK0Yr)@9TcP}
zMT638L<km&J%|NSL`4x4T+nDheY3#4d%yQb{M>K%&iw8<_sp47=dPE#iwj4A!QckD
z$AoRz62N3I7*gmLwI%Y$5Ck?#GZ-<2&8vFDdl#~AN1cyn^;zlhb$BKE{(8>-+R^b_
zGu|!puRP(rTY64rWTyPbzsg^CUybMujZ<$NEX>njCAv1+{by=XWVZ)XLz1y<^O$)0
z?cc!P@1xcxgt)|bB^d7=+qd$mw84(w4te$8)%O-!zSq&BzaAb}7pd6Y-EXe7`Pt=R
zi(SgG;WbaicQG8_I_W%cQLjC0v(&w`<U~zuw#HRw`%NR-!)u)#^n)i{_Y`X$Sg9*1
zVH*=IxBAkz@cT!izPmI(EqT1GaqQ+w_k7dks=NtTNls47fTWXHaEtqQjgp1gEnCTj
z2EPow)fX?~g^k~c@!j<P6<=38{a>#-$4&+A9M5(6eITrnf8xZbZqV|jXIg&#g?m~!
zVpP24;hdCm7$Nl7<-@I_b*#RGH8tCIoa6C*Uc3+vPnfIOwRos&jp*UI5@ie8vyzNn
z3$ZU=)R8b7^s|q0HfeJlweje9|J-)Uqf&V)vTJ|Icu`cg-h&u^SDkdszJN=G*Vf&M
ztY}z!=3>B#OF!#0)fX=Pn{QoTs94v$rmG{uzRsho!<}`W+ZEY$Brm%zKf8;lzIL)!
zVSUwGSJf-+&yDyeb9O8UUCax+n7V#-S@D{6+kVR1*&RM{^Jv5FGoca^RU_AXj#Rv1
zdHy+`+U{)NW!;>2N9yd0B#WB9Agi@~p?5d@5j@TtTP~wkOun;qSe$w`-lOZd<FgJW
z=ldnY(Qc2nmDYa9=#6OhH}WT{Tg`Uu3_3emo~L`WU}e+X^>>;t)$aOJanvrj>4c<b
zPI><Lgi7$4A6BP3o+Y{!6KrL-UtKVn+ZUZ@&~4ILb!mA@=%3bCPds_~Im@WA_jO2;
z(3ZAs^_q4ura~)iC2o+ZaW~nj4(%7ny&+L=ZIS}#B!?_J+g5kI$0u~)LVS(KT`!ni
zermN`lkX<g^Qkokp6=I{h$@HI7<}7&p)r%W$*z7-Ym;5$;Nxuj_}alpIaetKv2!O|
z*V=8gAJ2Sup!|4rV#2zNo?EN?UG<X(+sGpU-|FvCc_bdwA-7QfW9(Sj3K>%)`ONZY
zAI(-JA=Z$Uuoq{aP~Iz6Sd_#aAGfVZ!483jAi>%v(@s`fTD*jJaa~5bj^@y+TYNL&
z<by}~`{E>)4GOuZnod0R9w^;vc;L77%*$RfP08#^rdm(#N#8h|`cJ9c6dUWbc+Z!M
z`Y%qlM4X!t;XIc+^_=<H##q5m!FM>!x-o3hhS8gIw=FaLRg(QL|Bgtymr{0X`O|*s
zogw`%YuLHVxPJt)LitC{%Olr2D}84i-E%LMnJ)QJfBP_BX~c|^cwuL$@aY|@Zkihp
zh}<P2-lQE{tCnlm(e%pq+ON&Y15+7$>h8aI6O$P3uEOhyojk|oDLY+xaCEX}6m%GW
zUOaK#H2YKWNBhaJ`b|eZOAm){zTKDh@aWGGau(n8mmFL*Vo{Oh+A05gv&K`0p`Xq$
zx0)#T#z$?q?B&z&N$L8;p<||1qo+O%Nl!V8=JDRM>isWEF`~AG1^Pn<L%MJeGt<6k
zQVxp9F-R|EFV6IZA%bcW`e<vv<6TV^Qjmrch<p|(p9L1P17<Ob^|W}vk^Y_yr4UO3
z)JTBlvjb)^?mH9rT1+gE4TTX)3S_51a(2KhCc+U)k=_%jY$zkoAnSo-J&>9mFpJsM
zi{6ZR>9%rE3`y}(_TfxlFnMinPheqHz9<_Kh};cCy8&l*z$`%qJM;eT`*2GbGG!U0
z%Ul=L=?g;?0#tSZ38Gfyjx^l<*eed_AQ_%wp3>77{QQ<Y;4!r2(`Iov7m;~lJo)L%
z|Hv)-dc!5(%wt0qB98#g5oD~H0kdq|YIW%?lbjTW;)rE05Z?==X9vu(J??(a@Y~~U
za!?epcq@BzrZ1RW9B1%4^2s(?X()-v_W=7I+UCrFS?X@Rw~7lHxF$VqiB`aF1zP0H
zfLTtp_vr4=5nUw>*+|L|U=IP6*#WZ*`yGGYWR)`Oku)HghCDJeV3w&TC$f#2R|Shh
zAtWUkh$W-l%nX<%azk(wkHuh!ZFcGD%I{191la+29l#$eDL!^p>2RQeGbG4#D3K1?
zI9LcW6G~)KSi&SJ|6f_Ay+Q=(rRJrB1D7DZ)x34E(FoE<%|{1EibROA)sNMj79l2?
z3yEAPiDfg%N=Q`FaWy2WDJ=17sehxEm_Jgg0vJ>PV{8n9JO&JoQ5b6h#HP~*2bkiW
z6CzVbn#|C`MZmTQ*x~1yWD`(m0*W}?nPd;p>H*p~cvU9t47NuO%Of-GhkpCv8mzIJ
z<S&boMXo$XgdCuf1N5;(jWdxON8)UL1x!Sc&jCu9plM%!n!oUd*jMX-a~|z!CgJg}
zaHb=|OqyfD{1rj2X4OBXKqlD^6x)F^cCG*$wuAW?<J}+$GgSaf!1h3dncT;TV&?XA
zkFb)T<+Nz&b|8aA2}J9*13iqf&yfT{X~2Rl{45mAUA?A|-7*0B>S6O(tG?;nBpmQY
zmPuwqBAfC(vLhsN=r|Wz<U%W~JCn?VN_kL)!ue1spN<QlQUM(oLZw1FE`l;eP>y>3
zAe1>s$A_T8A*e*ri=jd>9hX3r5=c_?!%*cg9hX9rQfNle%b-aa9hXC+a%fD^k3ixG
zeZNN`ag>fLp<yMQPw`hlqKcy1mg!d6sJQcyg>L}#orgWLY=xCVyL0YJL<rIf(#6Ho
zThzHZYy&$D<#7rWO@S;dSWp0T(xCx`GofB4G{SgQK$c%kNv}1McO1$bhw_-<Tc0X9
z_LD;xA|wODWUv6&C7e|#CT98pj*9P_7KN_W)z(5f)C0wOppD~?K#W@5z!*PCp!zL4
z0@Sz6mUjPvq!8gp>JOO&$<^mtVR?dDC6x))u|@*ymI-+j-VfdOQ#jDf{j5yD@28N)
z8IY3!wJ;+V2*sv5le>0w%Dyw)Sc3GF^EAZW>BlLx*SaO(8&sW2fp#gd$1zQiRY1Fn
z!VyeHUZu(84YbX4D3uOX@iKyCB{HBg9h+r9dyMf)$Z&!th_%8l6cwB9d-%4%?5fFN
z9yU8cW&w5<;81E60Nny$M&Tl$QbeU9L6!hk2~eTvrNE&SETwQ2aHyi=8-R5KsNg`u
z&OzGFbX%-gw1?9b17Ax|<gIqVZ3o}sCj?E1+d;>jfZIvOU4Yv~VO))vVzbQ}uRx%t
zsW$}X<9<qziBLTe>fur;z~3Z76O3`6LaGWnI~KFU+uZ(k?IH^l54n()3+G`+#KXIG
zlh(R&OeXReft-d30&JQ=&+1I*ih_=!?}x4^=rB%MBDDKe`salb1ep(rd>}&^_W>Xu
z0BwA(Ajm#|zBK$0kPqqj5g;GYaX%pY>G&}qAJg$Dup9-p*hvIA1}w+u_zkdpL&tnz
z$*1FSU^!056TotU!YMDEdNu|}PJOXK?B$<s9p$P$`(l$yeNwo>Xs6!>lwRCL+{M^O
zf}Hltq&o6}t5QjcM$fB}&)NXH4XESb%6dGn#o_m|mxz!G6;q)$_CeNP{@=&5IJZ$a
zrU9cgu!!<O8t_O1JPJ1e#Rhs+Y5<lEz!6ue><+ghfBqV4h&22FWIq5k%D9P8HW8}f
zaKlbT;Wl$b#G5JO_L7c&{9W)x3CGN>Juepdqn3ywpR*U};Zz{d89+9HS_<zTEj4r0
z#7i{{Z*Nf8!>{`4b4a7*$=siBj<}(dO)8X0rRqISHnh~t$rUekuKl*(vBl%hzIv(i
zVfU;5?QpjYxvU-V+NslNi==hk1C#O~L?{4K1wa<NsC5)fopbVaLw5P+8r=F`<gLD$
zX(I}|N~m55HE=XA$!e%u4fQZ?JL(xO;Ucp8i!QnFF8I6OB3p`)PLBGHT)g?qL-}z<
zFD!3L3Jxel9@U_kWIiPF=@_}afQ}0xQAo!{kSL<#gOE5#$A=(sh>nXPQB21rkSL+!
z!;m;k$EA=crQ<S4l+kfHB+BWy0umJzzHXzt!)pa6`>Va&lhg8Ip7i2WWH=Xn7c*?Q
zy9sv`g|{mAe6cXxJ5chZ_FhRWVoZld>Cg_lj)_WxaVB)Z`2H$&ja_PM)X+jaC7u&*
zoBdDX9QH}o_M+RJfY%8EaLznlc%A!KjUxE!9X(YnH_RAh%|{jG|KC$8fI|gv!r}V=
z?=1rCaEziqGnq3nm!*(`f^P(fjsO{|q9j1k1So?m%5%fh$IQuU<WaI5sFVX!YK0o0
zQbS>UA0_F`OCv5EudURsMEVzBKvw7iye{wqwgSKK^m*f24ehVC!Z>?qWLNTv1=9}v
zr+0*R|G5+^KiSIqBID#_#d+fg*%L^|Adw&q>_B|LnJYH?mWxA3P|y|AfI8kZfnM_r
znt(Yj*aWJH^ICx!-YS7!`>c9^J$1|+1gt@zfa|rOuB(3l`q&GfntKO)EMKic`eeXw
zGvE>&X^hAXp#kRTj~b^M@d4S<ojib-#D$xI?T(|Lft6-NMXrxZ7yQekB)H6cl<|bY
zP@$%3)RaR_RjH{WHP!sz6p|tuOX{M2-76-pp{cH<tg6CMQ&41}f4}iZ15X$Xb!z%w
D9i-+U

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_cshuffle_lds_direct_load.hpp.C2FB5C5C6E8BC96E.idx b/.cache/clangd/index/device_gemm_xdl_cshuffle_lds_direct_load.hpp.C2FB5C5C6E8BC96E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..70dc378421b2366b891b68ee03b38132e5dc2563
GIT binary patch
literal 5968
zcmb`LdpuRy7r^gcZk|W3=eZZxBag~6j|fR6q?$3(h^c1Eue^S#iGC_m(!(f<Y4VO{
zL~p$xG9*dGO!`qIKZQJQMF>6Udi=J>?wtSnC!g!=^Sx`GwfA0Y?X~y0OWj>v7t$P#
z^IG?<A?r4;;c_?}Mf@MOIW*{^1cwu?z~OAoxanQLp?(&x=H#m^vFN;Ce*Us)vZa?4
zxkc7KQqPMNO=ez~_L?$pY0#r%BSqDjiRW?(DkP<we}0&zVHc*^edmaF%R(pFvY%y#
z+mxJoSB_{7Y>eESxh*(tYUZ{LX^p<~{C!#+!i;u#86S|y@ZF`^^i0-epTw%ol{JRt
zc`GC*4>yEnS}oLl+<0oY2k1XZ3@`}V?w4f!;PUylU>j{wbk)q9*RGtb)T(ZSX4@07
zmiq(M?npcuw)c<=++@5)E?r;a;Fj-Sw`?&D`g>O0##bNJXNBylF)dTdjk;o%E0>X1
zTr$b*ho(@A^Gmu5u3noydy`URnyGQ$Q}^2RJMlvX{*nV4jmwi_yc>6v9Zk7h?6BYJ
z+N5nM3yXVH|4EX}-@E>#a<myeH846o{mx_Emk0Iz*UmldVJYc*s9=SCQ>5gv{EVur
zen%B7Wc+>7dyOMqiuM(SJ`5IFdiP10_M9onwaA#$_^jGG&$Xh{<J(^|60B`sH?&4|
z*x1U%q`IXaS#m~n;@hwe*$&ZF-+3L+E8IeI%T7wQzMnDhpl;U5zZ;UyE#h4hM?HBT
zW7T!IYo53#V%3Ux?VAca>@){<7+f{!lURIq=k-R#`oVn_O0)m@U#)%pi_{r)`~SEt
zl@rqHq#LX^;%?W`<XvX(C0^_OoAcx5%C*Ujvg@^j4~0tDOa6PVX>!P*hoAK;(S|cM
zcbeiShd#3t4%}aGKeB75uI-);&TR^dALzUc9B}k~o$-fCqVxYer*8@n=PfAxyZ>yj
zMQ=(*oV>fvLqS=Wb??U099OG;YN=}d+RV=MZAPKbH+*vIUfR{`=Xg`_%nR=RrTf@6
z*bJrplHnQD?h~r{OYNU>w+<#-*Ke!+{odUl=^q;^o)(mM4QS<T$+pSotqk5*-)~$P
zlO40-TdS~;<L6{n4?Qp5*_P(^!81qKOSEXKo`I)IN|IZZ)m!tK{YyRDo4>hXx{fz#
z-fwSJb}x9rH;&Mj9+uZH5%?#>o{35gcp<xU=B@+FB~AVmtl37PIt`EGp691$f3i5$
zK6id-Pr#JY#W5FFm;QY8P;0bOb&6oml;2~U?Zt*>KUt_Qib~tEqwYXw4%gvOZ9;Wn
z%|q%R<F(nRSBEcpw_0@gRKy(L-0xOSp62I2|EH{@>i=pRF&_yi%Jj_YU9Ri$eMitk
zYgNMqgX(#eCJDO7XIu3}w=`_ulqHeIb(r4}+^2ec?k!==uGW=O`M<Bqbt*YMAVaxU
zc&BfyeP=uWlzoZQkMxzGux%lJ)riAUkU7u_@k8HBf;5jqZ<X7sG<M-g<F$?d=m)u*
zS15;F;lX4@4qbpI7a+xnhN#SFc(b8#SLClsAd4*|11}jgCmNy>5dVlCD()1^fh1OO
zmvvVgyAW|u)R|QAPn>WbOv2&@;5C47q9H2VLKUmm9IjH71g-*yc9e278M|;~0MCJg
zXtXy#<=cD8CQErBk8RX}ZXH-nG{i=H(SDIs{FRkTAcIv(5U&L3O*BO1(xdqcHe1?0
zR|P4o;->DVId&o9%Yhf^z_b{<ac@LSj!>jWFGPIRlQSdz*{U+$SO<qd<`DM8ctcbQ
zG!jB+!_r4eK;aIq1MfO$Of*ELa{ApR;#0aWc_4>X@`0NV^hEPnWtr)Ehkq_VmIDc_
z5-uCAG<G54&~^S{)jSSI?rnwOm>=x|6x|H+&A=m$rsy`1Z(~?~#=i96B#U}P(Qevq
z2E+^U0qV9vXElo@xby%}12CC9&!q=}8f5b@P{Rz9q~gksB@+oAg=aeTI+zhBP;@=$
z)WcK~yHTEA4>Jgp<U_I(DM#$`YcEL;p3JiZE?tE1hfg%Q^kL9D4EiJz8sa&|5uqJw
zcw!=O5P>6^foAi7Xg%^=j0NS0SB^|ba%fN9Ww&4Y!3!+djWl+nX(Z;le}x9ViJJW@
z;$nf?ZX_UrDNa}OduLV3esM5n9&6n!&u)pE#HF7g$0x{%^vR{4Bgf}#Zby#oZ0<ph
zJ!~F9?jvY1d7i95?5FYNM;vasur*dv;A=D1qwopRB2SF=D16T5b|h?Pa}N^suz3Ut
zM;Nx(>iPTMl52X@xb$nFUNe#B((ORCv$+EV9iTz{&ZXag;0>EQLD0$OE)aCFxf_(b
zL6v#_EhxWba}UrxAYkOZK=-n_4>bBf$jIM;#yd8@2h;ap&d5K2=?6A{1k;af?g!I;
zHh%)@6MMd5pu}t*1d~B9W9$zBHN=(=gY_`jF!B+gMi}|*^z718E4JUm>6nC!lhB;a
z**CWQ*d^Mzpgg&K*(W_4ytc0(gDc2}q$5S=A%i@Izwxl8%Un`Du>~JJA45`#J+|=w
z8qzxa6n`CV5^h~NQBk~PTyBP}<Hbu*xB$yEfC3rVXbCZB24jZXU}_tf686^pMSS<7
ze>Jw}E^v1z-FUA_^3Bbvx5r=akx<?vp+@vq<=L%YX{fJ<J1Ri@0;ErtFoi3o-XSn2
zOPHc>fXWTfA!TuO^S%JdMxBO96zwq0!Iq@@Xh~CS107<2bivDuK$FczBCun#n+UuK
z`&qgdD6c8Hfd|?Gnk}G51Y{<Z?AV^g^cSubZbaNh(c5|3HJK0`Kq?224v7^-Cm>-0
zn$B<{5+<@a2?>+foQ#CYY)(PK6gH<KVJe$1A%#mwgXupO6k<VxtoA^O4EH*|*9tth
zL}Z-EBoUb-E*#^#E}7jXE+_5u3~?onPBSPrgBH;o^^SZCsI%F!1)K<Pl9PV?>}Q=y
z+*KUXk3-AJoVVV2(Hg<6s^C&|HW*}s5h+>}odX6rY`zBu_rQq6o1*W7!F`6w5b>Bt
zj{)xb;c<ula*o{YW#v?J(QYR0cmNazz=(`!`yVyR8p#{+GNm2t9305wqxs<@f**+5
zZg@IJY;UNFOJxgCc#=dg$`&nP$*>40oLg+3E@JxQV#BT?a3gZEnc?w{pGwFG_NHF{
z{X&Y%sH;+taSHM#uG*WuGvb?5ckkm2uSSAuG=;<w?~h2U9!+KV6%xE+%e#?QH(Nf0
z1VfBGc8aC9=+ZJ5++zYVPC&Crk8x7PS^G*tqj0#FP)ig)r#$UNd9f%PS!N?!l4Ws=
zB=&YDjLcG?=xC%A&1{Giod&`*&}Sl?4#IRc9|7SJHfMk^gUy*B%w+RX5FTZ74w&bF
z4P(C$_=TX&WXw(A-(+(U@Qc`d3;4I#TnzkTHs1#RZ8qNr^ZQ^!?8mh)coOIzI{4Lh
z)dRixx4WIS+rIh%;b2RKhv`n=bvW=M&=N6Kl4KH2&hhgwNzTLy#eFfRFVEs9N)fLV
z=@8AN7e>W%i*nkrpb7X*pilBV>Er5EyA?ER@B)u8jj$tGf@yKYa)t^KzmQ$*3z1DB
znoo*<a{0oOkAICc!HcyVRLa>gs0WpL&?c)X$GKiOzvR=AFDjnly<ZvcuXqcOHx(JD
zB0G}v=dA3%yPLT>1zXsvxpgWzT2gqGs?;$ZosTs(`=nddgP)HDVJsLCN037@$d0`t
z5XVvU1!R8#%_cJ@E%xx}s{-kU>~8q$@fQU0{I!mVYjkYynJ=CuuiIQs{iW*7LVO-=
z0_7(7yt9(iH8wr|enCtZmeu&4bLeaPqUj>*rE4d-9P7Yyu0$4<$d1hUcE9xdX_gk(
zv4ug<83bJtd@el<hQlzGa7mf7_a6h5%dv%IWSq=w19wh+w?W=jdhZwGR{FvJ`wOAP
ze_{*v)9jte9PYnPkIH;yQ>DPAJAmq74oag3OX>}qJAvwCa~DuuZ0-iCo6T>5ddub>
zpnBNc3sf(g`+(|W^E;s4vH3kv@7eqTs1I!Z2-HV5_XE|>W?YS<BtE}#{dczE_N5-0
z$zKc}sZu!n<L{2db>yrfB|G-ANWR7klpmo((v-q)Q^rNeiY!P9*9D^zWJxlB!Y^31
zhrx;DFGZgKg%dzCwYVH~%3&HA#gmlcS4K(>2XPo3Z5-#4PM=2C@+JSH{^YAyy1K-y
zmWy>+k=RYmVA>3GNJet;`Ov%#TnWFJ^<U?YxqS3Rf41_EoBSqHFvA&}iOe#QCF822
z$m}S?Z8IjllTx7Lu;LAnxdF<=2csoLrVf-z?RsgFf6hvH_89ib5rjXc#A1hZZUWd}
z{0$4LkVO@mO}yTlk$=nlqOSfI2iwPc{btX859_bKHGSA|FH%i>Tj8r47K=~ZGJlr*
z4$s*~+DDfhJ%(@AH!rj4^u}Mez`Pb%MB)~@Zv7f7{FBduT)AeI?w!jK$y8F$KmVdd
zQjbHDjuRI8udqF%ziO$BUdI1a$qk32_W2g*^LPAf^H%8Yk=4*tS5f7w36zvM{LgZZ
VfVov>ZZ(-(CFWLxxm9Ow{}1O>0=)nL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_cshuffle_v2.hpp.6D300E1E67F0C8FA.idx b/.cache/clangd/index/device_gemm_xdl_cshuffle_v2.hpp.6D300E1E67F0C8FA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8328c74ba6bf5d2d52f277023873192a2924f33c
GIT binary patch
literal 4660
zcmb7HdpOkD8=v=^T*ky04HNT4%p&*ArJ^#K5)zV{w6Sa^bWtd=Dc4;@%Knl<LPDfs
zCuz&2T9=A)so9E3OS&ojRMe*0GdRz9_8(94jC0P%`=0lG&wJj>8Fxo}dm{#g!t-+s
z-y+!VD@CDDl+a)3_K-k-1cDJQJiB_emut`5=`C+ZsJ6{D%=W@a`{QYAS$Wj8tt#bj
zRU)!G8?Gejbj~t6{&C=Cq)njvAM$bW(Osdh6ZUU)*|huP$fn+s43;*V$n5G?%_)8~
z<=w}Ol?g@m;p-AiQ{JSl=w%p2o;fzV`>uWm#U;gVTDMU8v=&V--lE%FXX^}3DKk!b
zPmuHAzD2TWKj$-o*3;Ll&=3@~wX#Z&2^>P(ZiRGhzbT-_|70JvLDqmbPtf1ql;H20
zw}esH`u=D%Q=ehP^nMz*Kh>c9VDMm7fvtRZ716g-Yqpt;%ZrNT-9qAYR-NZm=hI6g
zo?Qz1AwJvp_~YRlS@Y*_6qa6bY5DTneBVgioR7V!!LP0Sf}E|otLxI%#Z)Ou`FxH2
z&ON3oCZLRXzWj8;!!ASru+(9v99h!qYT27by%Bo5BbF_CmmhOb`hrSu@+ZpT;9$qA
z*%6{!>ln>vTlH?O4EYkF_|?5crPg{d{Mnc8Y^L0tnvtPXQaMi#mlt*XbuN3Ov8s2(
zz4iWQdX0@6dulEz7&kTq4TwS_yKeQjheWnt9cZnv9t=GFM40zPyYa<EqgVBJH-0XX
z?dyGOJeZ#mY0(vSy5Y~VO_~0={OZAn`5!vMkMFntLB~R{DbUS7>RgR*UsUw{oUMyZ
zDz(!)vLeXx<kU4c?k^9!ai`^1FWt>nQBwyj{tx$kA59T<RBfuN=(qNGkiEjsW_vZ+
zK5%Kp-AMiZD$lb@W{2+h9J%~eaPiZLBf4!_Z9<uz+tm$c<2u$3EZlIuT){2lMCPY=
zn$6p5g&Mzv%_uIX?$Upk894O)#a9(>#e#!7_8nEZu>7S(V=Dja^2(|QVYWPWuKLZM
z@8s%SS&9DyM!ycArP#lr*cHC^nV;41*34ykL5}5f1=rfO%H=NZ@)ah&Jh_5He+RC*
z6e5x-;s&jMN%hD)=Miha>7|pKt>tNp`pju%_nI>2iHz>QSxOf<oBVHXpIyw4=MG<!
zc@?n7nj$!I=7DeC()NvOLq}wpZIM+y76x*H;GVAx?`=~}u79x&Zy!!Nd`SM9(*lFE
z?v$AQ@BF3R*=bXb9HVccGWi}3I}!`I3?o=#?UA-Ho%5S)UeQmc`gM%7etpk~Bhiap
zJ*GO;WfrQ(7BsMars~VJ8YmgGejlrRGAYX=(K*)D(AiU)tFbmpFKvm27k?4wz3!Vq
z<;y!h_fyvmGpmLb9{uJiTej@fyA@uYl~XigNr(RRj&s7Ak8#82Ug=-r9$56-wxlSt
zH``ZBH5V%RCMjA)+&<zSk(8fOqG~hnsy?<cVG+A_=STJs%SD@UX=R)pSCgY;;qu+)
zgsKm9tGy}CkDh!AJ5_nryU{V&@rhU2``AeCdEq`i^+(YQa}B4h-=J}OmuM?7i&ob;
zSJTWcro2Dv)v^BbqQ0Y_oyZZ_oXSXb%UQi?-fQ~Ls~l>)t`bwN+sl2fxKr+hPw(B>
zP=TKN?Q>s33jb*ovF@C?@I^tnCNJK|U_}1TscYqw(1<NR`4AM!w6kZf>lqFwG6)4G
zg%kl@1nlvNB8%!0`Y?GbI&WrQ+eIgo5KAmEGnOzPpD@cTU@Uze@a6!WP*$Ok<)Baw
zO5+n|vEeojUJUGYWfHPTio2RSdvw6$1v}&0w-k$3(FsLFeh%8t!E$`UEN+9#3^srG
zm`NvSh@}nW+K{PZ6J`niUH(RNo}43tpd*$(p!b3L_=H&wnBH6){Qdb}1~Cn>RDoI*
zQZzPUmb~<Av#&d;iqeEUVo3q{6kv=`n5AIDZm2ryrzB5IMJ#SqH^%6I$tCgePmZ-Y
zWJ?oL(<!9AjJ?+AK#?N=)psDFa@g7XpwmeQCP78}J=8qdqXXW5#U!A*p=gmcosdT4
zM1r13kmD0(5#3$MkLs(fXN-EI2Iw`QLyk?DrPFLFCo#?<Z0uH2K_M00%GiWi`fgae
zFZY;bBTvX6MSLn>Wpu#gX8~(`pFVX`V-j+Ryi#o?dvw6$fgRSi{Lg0P^28KGP6BEY
za@5#_SwaLGLIbE23U#Z!=tk!CM4*ukz;6H_Y(EY_ppJ(f_T98s&OVhoR3Cyy7Jy*^
znBwEm$PzFtkzgF3NWxffVhOCj%4J82&d;7AMdpBQ4sfvqDN+ckLSW&rl_EPprvo@R
z!Za0P<|WBCKSsu8fO-b#<Bc=_2npzqp8qRIA%aQ<%)kU~Tj7PlxvkW1R)tMK#*d{9
z4nAYQ5gU@&eFdq?1)E%0fK{2e`YoMZx-*%OBI|)w4{F$xVr*0oCK%6l=-n$l!_#V%
zkx>uIm>_1<tp`1fEj=w&&6m25pyM6@^wq<vaLyv*#0iW2igOkHdn{2g!D;Hjr`|ya
z`p}i-11BGBu^<{*44h&KUhX<awQ0%W6-c9>j-M`$LhOGecr356;FWSqs<#N}sfa(3
zkV_;Kae}P&P4_&N*J+IirG#=Rp@Ios=jUx+r_0enrj!G#95~p~G$dN54NS2vpT86R
zr5|ZML=t%Vyg69owhx2N2Rxa{lO6?^9AxuYy}=T~mMv~JTw~=XMMLi9)x#?6ZgB?i
z8YbdK;5AOfO~7lC;P4+8`b4e~6(S2;0DTRxpk${h@vruOohw5lGl7-~$`W_q1M(hl
zaBk7a`+&X^a3_$R6LA-iT@&#GARkP`4}p9r!O1Tcivl);j7%C?DT5a-5AZafOd6~E
zda_-OqdW!TTbReQ!9K*r8aa2YL}S9i7Nx`l<EN#FzYgeipoM*szQD}uvG?tDh>!$E
zNnn9%RQh9sI&MwD3q;sUu=Wxh94YC;K3=hk>^tb5Qo%SC=1IJm3d>S~FToYSs+br_
z6<}2Xb~v&QU2`bv_;rsF3Yj68J_MRLPiSO3F+HBp#Dxl{xD<7~tKd`7PzdSufeDT&
z@kXt9LJw~g<5}^98O8;?PSWnmKUtG%#O(_CQQWJrK0FiD(e~uX2}@4<gSUr9x-s2M
z@SZYrYq$ENJ{ibyDZowvJ#7BDmp`9V-8)$F%{8yms^ajjwtfdx+V&F4dnHMWyCO1E
z+!C>QmzwT)mf5_0^395Bx7;t?#=TpBOs@xiy`<^XDOmsBWp=_35wbuj3#Q{c#zBEB
z80)&&hO5PHLDTGI<0p0XpF<l~oe8PSMWXHW?Ra?qt@*l<?yl@Z-+Ts9M%^<5hRr@C
z&)&e^0?Whkjtm`Z`j~L{m%mujJJdvzrW`&pk}&or!1vnS6U|LiI@^gfHUhs9e6g;_
zb8qp6PqE;e&(O;{VRWER(FCQ~e_MG0a0_5Dw&lOAU5vRU5;|@aa@5gh3O=(q*EN#B
z0OvZ5Ob2>8uq9%%KsO6!OK=XLC&4TU&IQF>P?z9*;O2v~1Q!Fhcp`2A#THP<j(KWy
zq0F2t%}2p;95~0pTC(vZa861veorBviyu?i{?~Ifn?rnZ|3n^d0)7+xhy~*V$xz3i
zRd|QER<Tk+OQJ{!hC(pMwTOnI!k`9bV?i|Z!kN<nwz%uj$R1Ga0aZzd7y_*!(8ncf
zpyqb+`%=?m$Rhp>z9shiVD^Pt)3Y-<-y8vN>CgM)l3Xn&?HEJ58}_Jtx~w=U<EKxB
zwWeK%-XRP9Wc_Ah3qu5(ea+F2mc^&!|DHY&#Ru#WT*fFk%yD8-uumzWA?ri;i+}OV
z5D%jtf-H>{3J$ke>d4R3)|#QFq0ZJ+Wl_{HokCHUEHx!dwq&UxS+XQcZOKwqvYaVd
IYDt#=0y|;!;s5{u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp.9F68ED656E988D30.idx b/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp.9F68ED656E988D30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e4ac81032af97d39490bd7f2205d70e60e036841
GIT binary patch
literal 4286
zcmZuzc~nzZ8qa$|%!0B*AYfd0_pKO+AQ1#7Y|5emMHCHDD?(8b3s{4lcIr}T6=+3e
zs#OqLiwgnAg~6pru*I>I=s>L^RSQ+Bpwe1&Y;A3yIp;OsACq&q=l#CllFOHS@4Iso
z;^UiKM536i1XJFMqGb*uk;s+*D=aF={c@B@^sP)JGVS{;J*lvFs>?w2efq?v6=&Df
z`uMNTNDL{<EBn6meLu;<3dKf|{OyXXOKv-=DreV?o|cy#AUmsi`uhR5w;fuu_cR&s
z^|c=k^mNy`SG7KQaJlHMH;SLEt1<?*dyi_rxM1X!(B@+c_wOI8@cv;``zp_l-je?`
ze^#~R%|DV$hHnqtAAGT9hq0luJ2<#&q@!SWLyIwrd-N#3t3%oUo8+11qf@OFo=vgo
zrpc!lEKcw7KeJ2O@IPZQTAh(``I>L{nDEVe&8rR)^WyI<>W{c{TYdQIjo1D?<JIPf
zr0;gDnyo*RSt~7yo-|*)w=wnktni%7-)`3SmPGHF?xwe-{xojNP~y0EcBNv|2TQH(
zZF{K|_?5^IV@CJ!FG`%NqoRxIU47Rch`uyCrFL`ugq`XwfjcP6aGHzbtglSwh)_4S
zWYg$PgsjuwIpBwgE8pyLIJ7S?&y-%Gzq34rsMzA%;AIKF9pm92H}A-VF;TnrU%KAW
zSGM!?;f=~wPY;?SPi}C2H(gwFr+o0Wr7y~?ZIuyU9CX_vn|QioXkodvW^S?my^`db
zN|!=&b9#WbmHvF`g_gGRd`gp;P$V%P%k%6OF=3T%Hy$&S@Af^2%re}s>bvrBZv`SQ
z(Q#{<$4O32@u=HU@nuZI7u(7&1`f)H4!v0W+E`1p-_@;2uiw7&aAusjv;F%1jsELV
z{IxSTmsjOThQd3p=pGiAN2a_{5YjXy?0BcCx^qDoUb*|z^4vWUca+z+Y6F84G7Z5w
zvBkd~jh}Sro%)q0izcLWpAqNptBYC7tUt%pZo82Z)Hz&ZJYM5pkQ$rx$;3H*%ZF0&
zub=g-oVKc}a6|jtbrCyHex9-URMRUlDZ`QVr?021uP9u!n@QHY9h0WjjAuWoU~*Q?
zJe0QP%az0ZZq3@9kVz=#`=9@N>D_eZYQFE#)0w#b((|Vo=NF%I=d9<}9X21^e<96B
zr6oSNJHNAl>-f0LudXWLTV`YK^g8|Ax}Ki<Kd+znsd8uK=zr}vbYx5F`rw>6>G}7q
z@#Jc=6PrfFYR>j44OfDmEm&E7kUnv-SJ{)5z0U2nN$y*EIly%D$%}2D4%e)0vTna~
z>rq9%v@CnWF30M3JF_%jb!T-kdy;>BV`oCaC$~x;2Q7QBf0#p4M|?bY4OcG+acVsL
zmU2$F!{Nr1kiu)N=Ub1|?(u7CQwQ}f+dG`+|5p30@BYx8>^<?~%>iS`4~s|6|2Xgu
z<uA7yvmeVV<+?`tXrx$C-8K1GO~!|pd!9WU*%nnAq?7ObHtAJez#pRI$h)Tti}POb
z#zZ2qs<}L-CTzSMPg@`|x{KYVwt?UCwvX)B1GhFOh0c@X^LW}v?Bi-11lqOnJ>PW=
zgBRp@3QxO;U8J@_pl3Y)>EWS);zl_>pQpXW-lJ`UK*#L$Rpo>aj*;UFd3v09T!?KD
zXhU<rgs=bJxNeMSE>A1HloM=&KySz#+P=Leq0&i|&eK7TL1DH*pa++YKNA@8aEct)
z^R$Q9LuwlY`gfP~<2Um^ACTjzJnbrWmD&b@78R_>F1*H{c0zyLWwySQwLL#j-xo37
z$w=skSm7VwQ6e$hj*t!k4kaSMkvbA^WFiSTN=E^XN~8cs>uA8yi8SCC9RoNfkpUd5
zV*$q|vVh}s9N@S_4saTUM)_BZjT(Q=cmWA(!cS&f3!o)>Vv)TGgn`HbD3>U*Bcw+_
zi3|uRQ7!^XQcnUUX&`};%q4-6(o;Z587QEnaw(vs^)yh@1{x^oTpB1DJp+`CfdNV;
zmjOyv&jKZDV1bg&Wr32@b3n-%IH2TmIiS=8Yo-fLcQR_CGzkDDY0?EG>_SXt+kL_h
zBJ93knQhkzWqz?*Frp2!{~AOWMs!2mj}hNPY{rQO9Ozqc;ylDwoVW}yvS0}2qb>}=
zeAJI2n2*djg88TgM=&2*aRl>83r50xvI`?&KG~0vFrPHzB+MsUa1!Q|R-A<Slm(+;
zKGlU$FrVtjD40)~aSG;BEjR`9DJxFFeA<H1FrV(iXqZp;V>HaC%{UG7=@y)Z`Lq?M
zVLoHQ7?{s=VGPV?`Y{IPGiIED`AiGWz<kDvGccdEU@XjMyD%2!v;7zg^I0>_!hE&`
zXJI~T#aWooSuhUfb6pq*^SOSEgZZ2p=U_h9f^#sRv*MgRUsv)#^vg)m4a~t9IyuxJ
zY$9QK<1*Vz$i7>kXm(gdzE@!B2bh!5N9z+KoP=p}+{k-7raSWGU}><k5GlOCq%zx!
z%N|LXh|OSsArpd-G?5TeCJI8@L_^4!7zkMt3n6FX?C|xZ*ss`yQ&ad^Dx;NY!t8|E
z@~6kkm#8qyjc+z@v3%Ee`lUKunJK7CevI>*cDdA-FT1$7cnL?(N6FHHtrv`i68?0N
z5S~Iaos39=03szMK$L_6h?dX*F%kwKR>A_rNw~kz(V@-78GZZ8`T3}%s<FZs*PG5d
zrsUPe^JVX`-d@5{*b*`~%Q3EYh%dVn?t#L~y)7;3{u0%-i+nlMA#|e9k#GLldXI67
zDC5f-iN;MR-wkpc8SU)7oi7J_2CIa!aN&^2Y!?lok9<S4M1(Lez9Cu!5UC{rqO=r1
zw3Y^l(J}zBS{5Kq%K@Zu*LVq&8j{q%kB1+7xnMmAk04==!tEzOX1n!(xhe>yeXfLx
zm<SLeXn@p&#*UBz0UuH!;6n%jA5tUWBNZg@kt!1SNJIi3sV0GsQc%E0sVLy15Cwdc
zngTvrK?5JHqJfV_H1N@C8u%Cm1AL5%0X_yXz{jW=;A0gm@Ubcu_*ldOAFF18k5h2K
z$Ei5r;}8dYoSFkZ%|s0${Pz5+B{d8{oJK7m{(if9*~`mrVZrjkc`qLpy&P<xQN|+)
z*rcRMqt1wzWEd7L&J-C0N+im#KQ6XEX4)T@*dLeYWvzadhe_G9M>#l(M@yt~r!g+h
zV-sBCVvwsWA==H`-9zD{ofa`YGHS-(CVR5slcrLt|5x==`uUHaFfo82X@*k=1y2bL
X8yDgmsEPITnwhv1kDsHPg(d$9tKzEZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp.FA79EAE16FB6F44E.idx b/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp.FA79EAE16FB6F44E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d96f25739d67f5e06322dce3af6a979305233d6d
GIT binary patch
literal 4286
zcmZuzc~nzZ9?pA0%!0B*AV3s(_pKNbL81tquqlfMl%i;eRS}AcC}0iksiW1F)&i}l
zOtlI^YtfbfV=aT)BEgEINOWMVA})n0t)j(R>ey;^o;l}Det%5P$<6zI-;#Uo`sODk
zCbqhXL~%Ju=7N<a%bY|akvsn_E-A`CH%=tFCKHLw`@YFcDIS>aHe7p${_?GrC)U>c
z1+K|T4zDPv_@VrR0Lh{%#TJo#bJeA#H=R}0a~sCbD98<wolrgc^MJ=@r;gk`twwxf
z{YS(7eGQ&99lzbXP_pTb(%&}Jm_oaJ$90`qIJP{Zt!2^v{Sy_wKaT5K?cF_4_RqF&
zYL>qFJhg1}=J1`n&(`iRHC6X%G`(ZpMZ25YO)=chKNt3PD~Fy+9&11Ts-w!gH9pfk
z_1MBCnf-yscPX3xV=6^!va&8*_V1f8bL(Eq>Vw3B#Ge)qMculoK78rw`hUFiYFkvw
z_d8b4H5|{bmzKm%d0D)-IsM6;nR(eyuh$Qh#qOEqVX&tE5;ScjIq2<O>Da8h<+iTQ
zz0^wlVsyAEtMAC^GS}Le*pfzf|8)mq&(2M&-`Y5Nr+QoH4$3;3;pRN&yx9^J;lY-@
zHU2F^))VL&@?+G+?{+yI+80`2&MY(BT9HOnZF6n%vCh01=M|W+;Pc57Vs`C6d!>7@
zV&}2LTa>FG9kfIreckolOmW?<%Dd}dc~)WTtd2T;&|{k{_*nPIqDouc{8GbvWvO-5
zZpD_i%n)4%{p~BK+B+)?DQ$96iNw@W;N2%;BC9>FK4hle9=sQwW4u!{c=3~gDny*6
z6V|o`Nxquq)v&GVTwK%X?UiRj@5)CGJzKYaqBS<)(mN@y-Mn>wc7mm+>&pHufg4fc
z<>S{^)Z|G<W_DlH-!H9<PJ5#$ymeaSksfnx&%#K&diQ6Q`Fo;nDX+Yv3)LiL8#Q_H
zrBA;|oO0;x##Kj4Ca3is7Z>hph+D^O{F<rXel;zuXSB|Aq%N>1JwD~r;CX{9M$++r
zKki>OV|7jO>s|9VMD0BKZPwPWT9?PAjYc;fyOOrCs(A5kCe`54BF(6q#C}@E<gK24
zC}ZupRii^5ZMwYhDJbuUU;lOX-Av|Eq5sIE*|_2ClSf%6mwfHX*}mRz*wV89RED2Q
zM|^nu<(?w0`;&@*hMJ^n%$EFF4Tkv*{rz`--MHYh+@00q|GwkU=iAaZYVs1KCqJ+y
zl4~q3Yz7gpJu#>>UJQG@a8>O=`pYu|%Kn_(4IVema{uxRA?EA9J=^}-Xx+M2+xxd}
z{9IKit;l_SmvimAJvrL*eL20%p430y*qK!H>5cM-Vax9AALY>WF+cBJqqPgeU78PX
zQqJphI^3KVUVORZWXI?AdjeWJ)nNn6_Kp?=ZtA-6{d4`%fiIuEIcy65amm=p2g83=
z{(hr5_o2L6u5YHlh!!hqd#AS4Wqov^|MC5??J?zHdil<4DX;27o{LgbZyzfzEm-c0
zi9}*mTV-5b<Rm$su~2046njeT4Zr8@AK4!VZfs47SRlt2@U)-U&)wb#w0rY={u`R^
zo|5BfJnbfSliC}Be(A|C_YV!1Hp}ssdD>U(JKo+1^g7R=p-sCtubLp5&(lgD<z#y!
z(AVZ<wp?9X{;`WFlc&R+!y@gCKo3ut_F(<MXs{e##M43Ipm2L5(07+jIvyH+f0`UO
z@U)lMOKNWf`cJpaBi9SR9hT$iJnb%am)aYF78R|`Exyd3cEb3`WcG2X=zMaZaWHC<
zi-|}k<_cYKF`;Ax5SdH@L?u%I(aAJGOfmxyo6G{lC36621GSTc5fNr3KxUr}&?y74
z*wHD%NaO*OPn0+iG9aKtMg){79|0w4Ac2xJl0Zr3lR!xsD4?W_6i`z66j0Iz8YpQa
z4U}{~4U~+50ZPWm040;p03~Z+fs!?{K*{E_K*<?6pyZ4kP;&VkP-->WS;C;Zn6xq4
zB!E)1nF12_0VcDre_=v|^{<iH*S%2Y7mpPqIx)v=5WN`D2k{U_`~a~9Cz^1eZ^wy~
z5ZiF#0>H?MA()SPF$D9`5Qbnrvfv2jqjns@d}PBB%qOiF3G>NbjD-2*5Jtj$(t?vP
zpKQlTm`~bp66RA@jDq=8FGj(9Y6zoXK4rlvm`}Ci6wIe=I0f@*D@Mb7x)-BiK0Sod
zFrT*IG|Z>laT?~+Hk^j}j1^;GKGTaaFrOL17?{ska0ccx?KlJT85_>PeAbGwFrV$k
zSeVZaVJysNEjSDF*>;?T`K%3RVLoTYIGE4%VjRrphA<B1a~7O~`CL2B!F<k!bB=s{
z**($kV<lHHCsV}K2&1rx=0;W(`h=E$h`E^jbbfKdA(t`FgZyd7EN8x~k!oB8wQzYz
zW%f%0>@72~6>J@3Mi7!_5<<#MK}efv2pKa2A!}wK<jkA{zLpYyo?SFOjqjy0R+%C6
zZp)i(+pOOY9(z&FRAvk1vIhwPGcJ_-^JO<THy`2n_&8Ns^v)?$v4lS%B!stPzat3(
zh?I~3Q4$IuT0#TFNEm=v2@4P>;r@rd7};8yHMp;m?~h8Vnkf8nqxpn$T0wmxU-q5o
z>mwXz?cwutoD=Ft__8PA87f?nof$EAmZ~nF;mZ+D5y8SlefPJHpBRs*3cjqBXg!3o
z@XH^WFRpfB{%(gkkBxWreV<oryfrF8CG-fF*)MrP&HocvK*E*pDYIYo4iTXvqJ#u~
zX!S(A10fv(9MU7ekvbA^q@DyErK12x=_$a`IvQ}ao(3GFV*tnK8NjhR7I3Vd1std2
z0LSS$z-c|TKEg=xqpZ*>1vDaQ`~aWX@?uYU5nf@!2nlaLAu{`04;U#0p>&K5p(26-
zLIe$vn$S8BQXt?%Dg=B8A>c!51bn1|1U^zl0w0M;;3L%}@KFj1_$U<xd=#R9k5W^>
zM=NOHqg6EU(TD~<T1^8VqhNrKQ8B>BAO`ptH3NLCf(1TS#R4CTSm0yTEbws(4){0~
z2YejjfR9siz^4t?62jc`D}mH90C8HifcW>@HOpSS>=qZTC|>a5VeyN`{uyU-Z%-&V
zyjhbiUYccGyd+y>6ey7>%kj9x@tEy+T<UmSQINCdRUW3~&K>9EEFLeB%3UV7xlT-S
zPl!YAvZPoKUr#TEpKeCftmv4R{xa2@ojGMXrTTwWA7wz`q{+b{1W7ZTI!rSyA~Go4
UKU5p<?K3<16+CgCeh!xUU-Zkfga7~l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp.F19AC2A6D8DD58E6.idx b/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp.F19AC2A6D8DD58E6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e9328383f7dad384b1177e379093f88e733ccdc5
GIT binary patch
literal 4286
zcmZve3se(V8pksu#Dt){ArK%wm^)7lM35*C8(zvo1BxgbqgI5XA{MX)d+gRqi?x6i
zm917msEQ8)uC){&MS?A^r9?NZRm4Z3N>x-^i>|HK?vQh)e0$Dv4ww0V-|zoF!<~C)
z?wo}9_$FtOC?+$(n7g8InVm=^a^Zgq3iETm9VHUoa1e=%d%sFgD(IQw++Tg4{`|ca
zXV%nu`>oDM3@OVk`>}MJuVg`me3MA_ZpD=)x9ydcv+71q&CLpMIHP>_$9~s$?OL;T
zH|g>9wV(8Nch$L7wLX1#sqme*i=M8lG6c4JjcUIzf9Unl=Hm<Y?Heoi`e{`ADvyqy
zl7BXTRkdW}pUEYIxBKruc(rD`p`o%XIM_PWk-w{<#Sq0kew=6RQ1rc&JXe4EMQepe
zQ*63%(y952(!2dm?^HDW$54b;XJlNu>eDr5+U7l`RR@W=@%I+?h2OcYI&$Uun_o_U
zqd7e3hwZCo=}s@Ml@>-%oG0GXnEGPowCu$%Z`SsdMDLDp)tOU&@t-`9=>PuCR4n2_
zsinPb548fn92sKB=sNLDiBok{bYZ=V&)WUb7iXo^Zmu7{LscBOoiY!mIor?t&S(k`
zb!AK58~q;P(COzC@KgBZ?|0fA-W!-}OfS*hS)M{v6gxF|ny1~4arcXxdvyGmsGa*R
zUhC*B+i~j1CdH~}2ThSDH#luc7uVb=fAHqgS7nyA%J6Rvx)wW3IMp$*pxjb3r%3l<
zNpel4bAhQjJwVe+f4%g4OIvv!rA|yJlo*cZdUT1Huu9kKPnbz}dmlz->hD+eUjD47
z0udMKxHZlGk}oE^*A-WM8`JR3*76I14`c&}U#)#}tU22E$_Ghr-M;f^Mx3d${o1}w
ze(O>E)zdeZS7l2ErgdD_J}N4YOnEy$q-k>4iB4m6=ln3da@V2qoZaDf6xTk`1O_K8
z)(2<D7QH+cKk@MU^(#*nj!)@2Eza9p7qgaGf0n7;dOanmbFjv6qQ);jH8$zMgxS5z
z2U78WKkr^SbyZcthW0t@!grkfI%D$}O|QqK3`W+Ux|XuOqF~`JCRyitT$)xhjy+Jp
zWUrcWIBm_hD+l{rn>E=X6H)e$zy9mumUQMyp3lIu8MyA^i)R_<7M*qDEN9mpF&*D`
zKFwRHAwIr4uQQ+P_^iygt}5XMvneN{PB*8nyZipHU*0&N*ikw9AKMQfElyn@oE;}U
zx6Kkyt~NQcX+*60Os_(JIq3QPmDLC7&oA^Sx-+xZx!yL)d`d3`7;ip(we`?o&Dti*
zM|W;LuE>*?Wo_7LU%jO>Q~h06rj^;9{QKKG67mn+Dt!{P?BTva4ow;I_SiXCJwL>;
z@yI)h*<E%=8dE|FuC|_QJzBflx2a7P)U#~QV6NXg?YDmTQ+u-K^H&@D4Iw`*8anrL
z|KAk9-D=EwBCC{X8|h<_VtKW7((#&%PcC&oe>Ai;sx(L|+i@f54Q;@mqU0HOPZbp9
zzV3yIL}F!gc}z{%I2oQcUu19-yGd;Wzvpc~4!`fewK*wtt_+{cx4p&QF1A5vyEK02
zv##O6c^RI<x1Gh#QrjT3r@#2+(c%80Mj1YjZ+nToM%xCVy|6AnX_=$nQ%6xc-wv`5
z3bPGDyV&3OD9QRVLWVEk+x}wz5ZfTM58<<4+ZfpEH%2svZ!0_%<86b`ez0uZ>A;Xj
zlV!M$Z@Y`#rM5w6|KXf|;%460{W3h2Z@Y+Hq_#n5i}F`w6<p=tcEbGlI@so=tnJ1A
z`rhzyjs`+QgbR1bD=iTVR3b5JL?I0VB+?>4qC^BpQbPig)RKTC6G=c)8VZn<mI5S|
zNCA@8(14`1G$83j8jy^J0VJbk0LdgWfMhi+AXzO7NH&oLB&Xp3$!R%2a)}%u)oyCf
zzZQf+Emtdm@>7o!l&~WC2Qs`I!2{6|3r8lM&=c7}<q(A<3h59aksbjO<sd+kIueki
zo&+SBLjsc0QGlfM6d<V_3Xrsp1|+Sg0ZHf3fMj$GAQ?RaNG68?B&%Zq$?91^vN<du
zIUNT`PR{|7%i#d24pv79lkR9xN2wElN>Zl_N_Yog4z|57JP={;2RqpIyin%1iWwu?
zu#u-hSTUjt>OPG45o!}oG~j@5!HIKFTX5nM(8!D-n2)R&g88TqLogqia0K&F3yxqu
zvfv2jlV*&B`J@#iVLsW1kuaY$;Uvr_TW}KQlNOwW`IH%>U_NEVD40+6VHC`#OgIJe
zsTQ1q`IH5xU_NceXqZo1F&gI6eHab%X%kMve7XgvVLolaX_(KLF$U%{R*Zr9OdrO;
ze8z+`FrR6`8JN#la0ceHW{idTtQBKnKHG<}FrPKyEX-$Ha2DpX7Mz9ooEhU_K4--^
zn9ucL9L(oTI0y5&7Mz3moCW7b^0g%oMZXOdUdQYVp_4-O!X}y(R*~l!So$&MXz<o}
z#|SUEwArrYz3mb9d^uPe>?F{_Uw+FCe>H`@^;S~sckF^GDLk!+R-_3Z$W0rYi_Jgu
zo*FKvD;5i7VMpN(w$G{%;`l%JRq}J3@6=1BKD^i2+1XRr>z^h|^FKImD3I{mSweUS
zQH};AK|qlb5-3VS0Yyt_pcn}Q6f0qY;w0Q(O1OEcgY9EKa?^y7*bH78$cUgMjU<$m
zk%E#o(oix+21?e*LdhAq5qWH2b5Ta`-g16=lv3qbVIJy@XY5mQYvcK{*H|x4;Z@iY
zGAGkMu6BSgyAf`I!qK8FE$aRf<<$#(In*w6g79#^|9k5_#x=Z*FRLYLSD}12$bM+F
zlh;RlIoKmuDU^ld4!<OZ&pN`y3`pv?;nR)}uPb-LJxHL0(@%hd?bIXC{L_>Z3gN%t
zm6DhM6e4J#RD^m&Avpp*q(s1n5CT4=LcmALN#G-uB=C`l1U^zl0v{!(fR9pAz(*kp
z_$U<xe6*YfK3YiwAB||>qg6EUF>(g@7$pOI3}S$fQ8B>B%30uJl`QbFhy^}Y#R4BE
z=YWq>a=^zS4){0~2Yl)YYC>2N{JJ643{ad}B`E%UyL#F1VYeWEdBNP_i-p63?Pru>
z#rVb{`II9W;w2gSg^L!8^g>G{${4v^G;+Ck<Z{W#<?`Ij)o<`JDQng!JA3hHiB#q|
z#@T6Xf=gTsa&btAcJ*>|mwRibhDSt3O@D2Y2Rm)z6iWI3s-6m8zj5Oy1P~<6aH^o-
a$)REXAwGfXSP#z`iA(YL+1i;{@_zwk4~=60

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp.478DBC5276E71817.idx b/.cache/clangd/index/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp.478DBC5276E71817.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e00d3f16d3f890f722752bc08eaf49d1f9221531
GIT binary patch
literal 5610
zcmZvgdt6g>AIEo&fa9XvVK8yt&i$^)CWt=cW(;(wgrIX~A~Y2-z#RFiCoe@AP@?gq
zd4n{y9Qv3UYE6fQ88K}<rie-`H5H96<YQU(%)MTxegAm&dhvUm&+qg3ybovR?03#K
zD?2O89j#EL7iU|`7FR9wQz#VSp1;bf3d@&66pAZB3Wassrv*8cJxO6VciyHxT)X(h
z^5(dNWkrU>y0W_OYTp>4ENoD%Q3Su#aA85Ozov2e-l3DqO2!49(A@uXSHw$xr%JXQ
zHKD7T-@e)1wKsCdsbBA&t9o&D^{*>;n7Q`YA?;`83_d-v^-$sV?ZZ{E-w$bD8r{)T
z^F`~YI~Kg~S8mNf@6FqH9xdN!ZfWe2B-da^#ak_Q^Az@nAIe=FYWE+?2jk!Q=u|`W
z(ToBse{9aYg6@Rlo7FA<HCN-yii*x%jPDwjyl$&)=^ku$)~&hjl<U31{tK6%{dnp#
zttmO*Zd^Luczk|yU{%`KCj+(~ocZvH<kI<nTy5^DN!v0x!swX!)5r<^hLJCCo{3Dp
zQ|oMR+e$7*&!;Aui@FYfUK6@=N?O%{@c0$G(#}rLYhHI?^d@0Fw~=%V<cIk`@s-t<
zGBJXwSvzzs7Sx##I_~?F^WSXt+qaD?vli4CuP@5O8rFxl#5j_B)1wkHXTLvs*p$uN
z&tB^2tJ`#J{~Gnu`+IDuN1hLTwIHDBdi|YepL$g1Y->#Ud{4yspfSff`U~rwO|z<v
zuhrx>HHKB%S_{VMPEnsdb=uxmUrvrUWLGK8hsvV66tuQ6;_}Zlf1~eiYO(3|j=uBn
z_B7zwSt@gR>qzBC6QcI6Z}>93<?{{oXSh4T{reuRcy_oWZN!CFa-QqGes5Z)t+V~o
z_B9Es@T`l+uP)kAs_akhIIq7~U7wn_x+3xD1nuEY>(0(OTD0-4_v$TMQm(5ny`tl!
z?D;0CG^6^D53<JYd-=eUBUPjGx{e2wZ`+%`f?o9r-Mry)p4d6iWIo)KP%$$jXZM&H
zeT({MqW^i&y=2nT9hJ|w&sv$X>BwhA>pnXAbb8)E>VadI@>Vre&V7r{HAWl?%x@aS
z>~5e-mrmQ4zx>N31MY}cU1{Q2y!5+YzCQbE0ezu7zW@F-)Ohyc{i2ieK8a+VpRC+(
zJGA|Dew;>!y?Nuw&I-2U-MSHbcVu6o*H|X+HO|`G-F^F)kFV@5+0;1n-y8S6zkcQ_
zsWdb2<QvW`Vwo+3$;UFrpXgJY&WjJ`EZMn-`tVGTy1Te!WkjzvIKKAWIP2A4A8mMV
zplQWX=j+$6{m@VzSXc7=X8)b9b{3ESs;k&VZ^`|A^``8K-PdY=78l;#KEUEhgK^QD
z2X@X$3^};}MfHp>zx@aE5-TsBI(h2-<}D+Rwh3a-!mR^k2`{!^`}QyWk)979y>Qc<
z`2D=WlRw`4r~0>R2TOhqZVc8Rq&`RuQ0;W_hnkAsKG*%=-r$BQwW2<F)0LcO^yB_g
z<QCjGR#{#4bS$D!1ZZ08)0?!Tg3<gr3Ug#YWT1EOeCEBcpx<|0TbDC&b}%~IW5)%=
zg?k6t4nO!>{K}R)r-RWvj~x~e7U&&hd+NiV?(Mr-eJ~h((qqR4#18civc2oQ<h7b3
z>h2IlfybWcKT+!)Wc%#s)cCN~+Vo(w&|{Ab7@6oDWc$v-QOCK&dlQ0DqsNX4hzj%$
zvi)aR!QrdrpWO^bXL{`Lfbc-?Alr}L{<?6``o*nbidi089itxY9b{WkvACqtFU<G;
zBm{Zim%6rxyAJfFj0!PhIxI#04?W6&P4_8WhXW5c;J_nv1n>w00X$Mi0*^G1z@u~&
z@F)WXJX%Ksk2cW2V{{Dg7y|=5R>uO5HL$?rbR6(F0|z`_#{-Wy@W2yv0`LTb06bAA
z0#7uEz>{<m@Far-JZ+>l=I@KetWD64l9haYGC|(!5}Y0*HrIDTFcVe^lm)BuDcp#I
zByPe%61U(WNf-%`BuoTI5*7j^Nh1l8q=^Jc(n5kHWu!ooGEpE&StyXCjWkHoCK@DZ
z3k{NtkpW4@#DFAYVL*~KvLMNtSde5bEJ$)j4kS4f2a=qH14-V<gCuX_L6W!dASoCH
zkQ7V;ND3AKBt@eLlA=ijNzo#Lq-2ypQu5q%d}ms+NFb?|w3Fq_D8#ItqRj>>M_V8(
zd5uGYylb0$iOFl5G03~7$<v;tz=2?Gh;Of$3&Faeb|ct#P}@+f1qFFKik*boiDKt~
z#vKR_`M3+gAs=@mIOO9t6o-7=j^dDyJ5e0+2?s(zKH)+L$S2$g0r`XtB_N-$qXgs=
zPLzOr(t(hWPr48i@<}&BLOy9jNysPdC<*zb6D1*^av&7sQ!a#pe9DbbkWblA3i2sC
zN<lv5L@CIp9S9Biv<smjpLQcO<kL2khJ4zN(vVL(Q5y0Y2f{!;<3bq7XWR$_`HT%^
zAfK_L4CFITl!1KKfv}Lzx)2ufSvSH$K5Ii+$Y<>+3;C=QWg(w)AROd#E`)=8&W&)8
z&)HB8@;N)oK|be1ImqW72oL$Z3*jN3cOyLH^EQ-+eBO@okk30&9`XeTB0#?2LIlVc
z+=u}Af(;cQU$CPB<O@zzfPB$`h>$P35E1f4HzGp5XhTKF7wxDB`JxjQAzyMJ668xR
zM1p+DjYyC$*-#1cB|9oXzT`wDA75W{SMl3m)n&xboXjVi<VAA2wxK+Rt9=s*F~{lR
z(&gK7{)`CX*2c;Hp6R5(q)<8iTu#PUOkq-<XIh=6&X=FX*1XWV-tleUvB%Q|>iP0?
z&5xNQCY`H|_e_U{g~iD0@jJPJ6|bB&R}S+mxk@bBw}9eG94JCb07WWEpeQ8;6s@Fz
zVw4O}tda$aQ*uD@N**XdDF7uZMW7_5^tY1V3(sxk@sqE7|M;MPU3F34wtCOOY63OG
z<)b=aJ>j2M)|};;jvXEwBj0@NiL;9RGn@N8(~(#tCqK-$<xjc2Ky&eoXFAz0d5pX(
z8HNUVAEf17dP3k#e)F$Wx9EtJI!{-t)JDkN8>0W<(9qb|J=00iNg8=teh3c@@;-d~
zkeD9JkoOm2#`QQ*gq{G3)RRC_dI~67PXop18K77_3lyj4fa3K$P=a0nO4N%$NqPw=
ztxBtw53OH$;0-kS(Bn%w3X2luBbC2b=z_f8C%_S_FtzW9F%32bC>*1J60q?;g{yGT
z!!<bQ;W!R@xPXHmp&~$!&=8<U-~{Lq0s(rYiUd7ULxLWOlb}ZmB<N8p3iK!q1$q=t
zfgUALphv4{(4#dp=+QV0dbB`;9;0GFkI^uo$KVX;F#-d6tcnFaR>OiGi?g7|3M}Yx
zDh~8G4F`H0&Ve2$aG=Mlc+lfDJm~Q_4|=@7gPx!gKu^#JpeNu0=m~-VdZJ1MJy9cq
zo`{Q}Cki6yNh%5SB#i`m5-x$BBuJpA9izqM&xB`9Aha}4tX7bfyj%|t@_v&B9}FwD
z4m|JTRvb#gN<c|kNhm2R1to2zp=7KKl&qD7lCyG9@>U*7!74y0T16-&tK^dvi<d2Y
z{KG(H#iGjDk2mH%9=xw1=B;xJTMiHFr2`fendZ)$uQ17$LQ&+~oaft|@7rA9+gwyu
zyzCi|%qf{Z#Lquqs4_4(WLQ|}@a*u+bUZvLJ1rtMGD;Pvo0Kv+b;{I#@X<{2*d$W(
r|7K&<BN9f99y1OjD4G>SX~IP9$i#SVd`5K4G{aM9)(rg<Nbdgt|6|dw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp.1759B855C9E56265.idx b/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp.1759B855C9E56265.idx
new file mode 100755
index 0000000000000000000000000000000000000000..659176bd2841687b03dbdf8cad434debf694f7db
GIT binary patch
literal 3410
zcmZuy30M<n8qSOml5hvXln4Qwe`Yw6aLN(D1q3As;WAtaYPm#Usem=O&(;H4x)yK+
zWszb9JirQgE60k0RZ6jeLQ#>e3W^tktteQv?v8nO%)igG^E~<Tzu)`4?|;lUGtr?T
zAyPvIBPcagm6KMmPKUu@m{Py|g1pS{7cdw%IShuX_JTYte{8AAg9DH4Pwz<U+|clm
zAUi?oUY1jKuXNX9R(u6-J41h4#mJgFx`L|kMm_JGbd1v}nERuSyG`d(`o41tbaO-7
z_oo|WgzxsL(F2d)&OLuJ<$EEqlvr@JH}1v99&IP$_wP62Rn8*!&hDI?4odx3^1$(~
ztw%1;=gt;v+IoJaIIjOI7nh+I{ds#^yORB#W@d7S`uP)YS+88bK6|Oc{M>T6O4J^g
zBp<cv*u!s~PcFo>6A}h)SPmQd6xFn39R@KWj}j*&_wR_0j@;UKe%bmqN!Z=pS>dt{
zWdpk)aLH<B&GD!=D}6GQZ+~eRD+%1^%awIU{bubpCAI#1PZZ+&q*Q&ir-n>J2mRcW
z6NXP+DKS3aA6U?AYWYcBU|)D-Ls7GBrMOtQo9uoTYofdInyN+O!F4Rzp|=BYezY>i
z9!dtU@6l<h73QeqC9?Y|k)Wd3xYeTD=T6W<tKgV#Z4LeR?C-nTKVDYZesnuOYwmE1
z-<d7OU&@*F_sgGbT>Gv}-BTsGa+q7pv1{+2iZ55!M;FTeR<f$T$|S#~O^$h8vj1-F
z<*uIcT+&q<TEI#^kz+p0aPX|+-g@pJnjC-Xm#TRD&G_KKu?igY*$4l*&6;)AZDC_^
z#ScNPS9X^73ZLjtHNA88E?H|^zTq<H%zZw0u;F@o%2V~Fa*5f&tY2na2Hc7_?N57D
zH|N`H-Q?$OIkt~Kn*Ps$2gmqGzci-z%r&d$hS@>2Rn><?t<`P~zVB8aH2pbH?vdDG
z(v+zz@>CDodhL}Yn@0ubkkt?7BgLbGM(tz+h&FBf_2f$u+3+^|Y^}$Dw_nQ9j{TvD
z6V2U;NhfWa=B=+QvYRA*-YrsdQ3Uq{JGS1&@lb_BM%Id^*qI$8x32(a?+o`Pc*eci
zpZdO(JB&QDyFR^3=bMW8>C>N`R}t~Gg;UW<>m(&dUHw0i+Vmf*nT)Sm!S{~vt8811
z4yz&#9Ev#I_xtSTm?QkkD!r4to4zfM+U$}M%)Yow9YSQc7&yj)<*uFMe8r&itGEpZ
z4%?sZ9pjItrf)h`cVi@j+cER{v*614NUx-a!Fjox`_30%b@h{W%DzkYv3juHCui@f
z#y>7RcK43X4e_paom0=7oAF<la(;L@;Ar{9jo2>3#>{O$$B{MrvhJ=~yZR?zqp`b-
zs;8ZQoIhu@EG75XJ2tnb6aMX=E3TOpzFwo6y;$~_XK`*Xb^AZBi`z3|y`WT<X?%0g
zssCa2+fTOy3qq>v(n4pznZ*@b8GW}W75RlZAHSlWYo?&BJgDB&MjwrlGn8B=m#qmD
zqq#V*>uwi?dBo_W(G<;N@=P^>rrkI0va9dgnP$LPP0>!ePM(@T)5|t)6P_%X>DNc&
zDB6l?<*o@I=r_MjH$5mku8+o2v<1^bPZMbRRQ<&u_R%dz3>i@rZE0a?s|hsyWSvci
z(0$rXAC*zG8Pkld2{ip~l3gcLCAaINkrZvhG+}E3P3NVh=a=hg_ixD2?7OVzO<nW2
z#Ku601z^kR^;1ZSg=-NA01yYH0OIff9O7^(4sk>P0da(sfH*RMgg8=4LY#;v;(s`5
zrN~NTLnHeBm~k}s1zNrgOVlnHQ(zeo%ESt^2*?0b2^0XT1epM;#AP^Ci7RlZ63@h;
zN<v0Jm4t$TDv3-2sw8D3R7omGsFKVip-Pd9$d}%>fl}lz3WZRZNKPaAAK_?vkv<Z7
z!zf47gLK+J=^}Db(3)lc?MMMf(}~(5tQ*045ZZ(4ham(ThVcY~-GgxpinXFpeHV&d
zgs~dM1|ST&5diyxAq2qwU;+WKKWIS#><_w70Q&<q3SfV{8^K|Jd<el|e|!SLVSl^@
z#bJNE3&ml7T#e$eKhcd4us<<`5U@Wnfe^4i(Sj1NKhcE}us@+j3D}?PMo8G7970Ig
zpPWEQ*q>}cN!Xw4LP^-4RHLNUKcM6(<M$T@w-BBG>4tF6id+j}>E96pCC`f&L_ce>
z5nSTYZeLw0Eo2Lg>E8J?Z22|E_@$9l+A`2GmVSF}|7%-u_ucXK_i4GMl1_iUik<hx
z<>Y)_s+U-pxwcaP3qS~G;SeHN1cXQy`2ii9Dk@AEuPvu$v1VHv(RR(MPTk0yh7c-k
zVPs)Jw>P~ACP(u|h1y(&6+vALQ~-v!3Wp(~B49|WNG<Ggk51JMZkVDb<^rye{>t>k
z`afPHxY0|cU36UR=%c&-k4ujnxRNp|En<ndbb8WR_l2JEqOYj5&|D~>)AWm9kE40v
zL#y|~B=pV=l)wu>2=~GvM0gPpBE3imiOfV6^i)%<=3S`IiT52~2F#r4ndxr}kE8jf
zKr7&3eC-M_0cHmw9J7az7<1JkzypvD2mqu5IDm9O3?LoO!yz3mz#$!u<B$#)<B*Qv
z5s;1$5Ri_*2}nnX2}nosNJvKtNJvNGB%~w7B%~ABi2!}3)IBCd4iIt@iD{IVmc8!%
z7dk&LB|ql<AmM$`Tnm(^yw*jG4qg*5ljD;T5*3Ujnqn~GwZla1P@x?rYlkU0soCo(
z7?vKsKu4FU$71Un7@8Oxg_;Hj;ijC>K<*;5g}jfvyd}PV{>%O(GI#V@vXm73zpMp+
mv6YRj9R>({2Pd(!i<^h1wY#Ozb-B643h5d&BqCrXvg$v0w<@jx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp.9CBE0B752A41740B.idx b/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp.9CBE0B752A41740B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b66c6dbde0ce582910746811b7331a1a8ad5c8d4
GIT binary patch
literal 3410
zcmZuy2~<;88qRw{2w@L`kq7~N_rAxHgjJRR4j?E&2%D?{v}__Uw172qX6gbhofdEe
zWsqV;)PfaI>&7yoV3krdP$+IptDv}mU=;<c*3L8Myvdz&Cg<co_xr#9`|kbkx$oYn
zkl<jM0gDxw9-_?4DB7UKVzEr<e_>HU_V<fetQ%YwOId#|HnebTh4KCUkL*tF$mrVC
zxKx;%Aakq8tGHXfa~UVTioczux2<Y4^|rRKCag)vD=(Aab_wTyZ{Tgyx{$f|i~`@%
z*#7;=COPS|efsA9M{nkzO-=ipODrcBUG9r}{+@gL@%VlF4Efb_*xl2gPtF9U|2t*y
zSohYiFD~TI6>Z*nc8w%%;0tHx;pYPdd)m5F{2XUz^M?ln6K^;#UA{bhp~~#cs#vAC
zBQ7cSre)`DLEAz~G0aU!7`$#VV&Gj;*P3$}L<c`ioRHqTEjcoJ>w~i^H?~Vd@9fG6
zlXoUJ>J|koU&pRH7WsOOcUJP7pBu-@0`~gw<UNs3tz4&NR-f#S#C)d8RhN6~s0@6_
z*DWPs<l9SSM*IB&idsx8-fsx#4~u9lX|bu6l!|syJx^kcwbxuxwo2W34rM!Zb^z{G
zOC#cebm-b{t>$`Bo-(#felINoRFxXFnfG|#4qR**6#b2jf#2?Z{Wk{2E2=w=Y!~Fr
zA8z$M^^wtMvFro)DyKeJ|F%NaTO+-6m{-cR?HHJjuT&j~DwhAfZ0&&><HFYVSi<vy
z-FNFRcK251Q!cWQB2LQjJhKs&y+;l2)-!wY<oILXbj72u#)l4$RYA~i7xd?LE6!=x
z#Z9GEKLoa2`n<AFG^ID){MN~<Y`sn8ri-8}|JnS(#%q~rk5v~crKSgSex7w6bS>Sy
zFXLguyicE1v#*!M*j~ZS%zq8uKPtcmWYK-6u2?=Z$PKKosXZiat95PkdAshQ$xi{X
z?unho&DqH%9;y)=&ppx<v&f)4s`maugyiOsVF%R+qD-28J@JCVHoeI`UGF~V<(qb-
zb6-f}L`zR%(g~a91*>a{+-7ONSF6lS9L_tg8*{+g;ZT)*R?h0?nAshpzgz-NURiF-
zVb<NbANxOxwI6+Adu?W?)>l;vGbcYjt0d#=i>ITKHb~2kxcI#<vmQ8FHyK~ETHqD#
zTiw16A5n(yKNNnl|F^j<(O+j)*XW$s)%;Cq<QC_wAl>sjRl#I#tG+`FSmn|+E>H|P
zy^Pzm|FGT3zA?eg^vuoQHe4Ug;&slx`Z%b1A;L51K~O>dmj1J)mtA~iUGncTy)Ex=
z^v>I}w(0kCkKDYX@`Jr<UFKB_W~PEy<=h`$3_4hRdOfDwpecLXPjOV;-kdwB>(~A0
zV>otaN$rf&)rB*LE7S6Sy={GKCgBgid`aD$=vAt6?tH~xp2WGn&>r}tA#V4m)uM8F
zw$Y6t$AJgAZ$A7eNElq(kP$KqPA#k2%Icq)R1_BHz4wy-uGzx&%D@92)_Qn!EGwDE
z=IN>f&8UCeR}H_Egt|xT;ZZcrXY)<efuUEvemc{9zxbFQ9z)aSY;zrTVCcB3TiZl;
z4WbQLku+^#ZegPi44t(z_36?&r&@j1I+}LWcJxpOhVCjp<<wGW(Wi&U(X=Jo(oG#0
zdTN7pr^s!_RS%cbv?<$ER~;Do?WAslT$$3Lheyz~G22*I9T>VGBeSqlN7KI!SKW6-
z@9Tz^ajCU_G9e?vm=)+J0~tUNltBcMG7>?Qj6#suQf$q%$?T;mSG|YGRB|FwGZmpA
zvJjL_6loBU1Edlt08$CE0a6L&5UGR;h*ZLCh*Xkt5~(B=BvMIclSm~cr;tiYK_Qh?
zHicA*oy9&(+xp33KXC|xLdCHRV$P3{t3Ef#mh@mmFQ#dn7{-VZ6i;BpT@<(CL>rFO
zcjLr)6svGz5W%1a1E@b3#sKOMCNO~dgH{}%{-7HNs6SBQ0QHAG7)1TyFa}Y7IDtXb
zAGYEU^@rU!ME#)(hp0c<gOR8|IgF90KRJPss6W|?lc+!0jgzQ9slrLrpX$LV)Snv0
zDAb>tz$nz8YQ-tkpX$ab)Spt}l*Zq`>@n-N=S8<Ltz>7hvw}GdVIEca=A!a{VEV~?
zPktcty~Kp`$cMXpwCS`+S7gLgGnWgStG-l_v!f(RkfQ;V0EJKqQAjFD6jDk`1DTmH
zuKHeLX8164)fI>M6%q7$ECMWIn00LbTYG8Go$-!$=~#<oCjI4F-GWaqrW9(^R|<zP
z(;QL20SJN|h#-<fB8cKp|3ycqONtZ5>nrJ5taPmm8M_u`mv%&6V=$dIH#9eAzAJ{t
zx$3Kv=|KN~w!PaeDqTCMahlfh08hlc2)!|Wk5Yx#`{=Z@ma{Fh8Q1=K;h{ZGT0y79
z95IhcPdaHo*D+f11)Ua|iG)m=xugxa>dRTP0K$`yGOMF^#-CWFLBJCr9PkGS2R$Le
zL4SyFq$i1Rq(6yplqZF7ls|=VVpFj>vtYU_zF5GZX->;d-23>u6KM)eotSQzw}sDD
zzbQypd_th<gAfw72!ezif+U2C1_3@mbU+9Y9e@DQ0SQ2KkPi_Z6hcG?Aw+ag0udd_
zClMVfBoQ46Nkm6VNJK~RDMUvJDMUv>3eiy#3ekye#emsUdWT7|J%Sv?5(X7y<ZgKP
zLKhaK6-K`sB)kji-=bud$zP}9Qg*noQ{s~n5*4f@hGMbeHN!;BP@x&7XohKd>A4$e
z7@8TjNK2co!_n2#H!wCb3^54`geKgO0NxVQ#r&n7UQ!=lzm<Oyn>lzdUqK1~U)Ef(
m%+lJ%mH?!ky`#j*+11^{%FROLvdYYSwJa464)<S!t^E)5)g5dA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp.2A9A70463B0B1D69.idx b/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp.2A9A70463B0B1D69.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9cec5018998756fe410cd3db2cfe729b94f4e271
GIT binary patch
literal 3410
zcmZuy2~<;88qRw{2w@KrQz8WL-S-|#5>_DuZ~#FGLfB+YaA6aHp#`j=bH)u?IxXM`
z${@uGr~_7lS~r#v1*;UXfkJVkt%Bl)U@Hn%t<y)&dCi@3Cg<co_xr!^|K7d-yYJqZ
zu+UJM34;-w9;V96DB7UMU@$DGe_>HU_P2`|jGJr*Lsfr1KD=;jx%q>Gj~!3#%IMnM
z^qwF$N#<FRS8=a=_fl436>lfQXh+rPx;y%UnuultNnR$#?h?%X-oV+RcQJGSSp~Ya
zsr}ni&2qwT=hUr(kKfKce=_BFKDnG&bfquh#YV686Nv{7nDVM;k$Y!8o16|#|1ov&
zc=z^yUz*RKE!wjE+-h;cz~^pm!!HI3_O*4V2Dr@3<PQ(<C*HDNxqoryVwKg|mGLT3
zM?y;cEt}50{I>bjVmvn~Y4C>ih>34WojT_Thz)&|JR!Y*M|^bj_6O%yY-*Q=-`$fF
zA@5W+85RXCTg$9F9{pyuZ<g}yPfcTGLHqqU@}B5lZ9S%Bwjb|}M*N<XYp(Rxks0Wa
zzh`RF$k&(4%nk+w6}4Jezuyql9}(45(rRBVE*0(}d!EIa>#x41QcJx!&Skp{b^-Qx
zHfGpE>CpAPdM)+BJXL&|{C-*#s46vU<M#O830`ay68n|CNx<F%{Wk~3E2=w=?&Rmp
z9Z~zA{>bdpcxL1M$|oPJe^;UDt&v_n!YO4tbPP--R%#k!isk<(Thmx$UZ`%5$9yh2
zezX2kcW-4r=`IT^Vx^wQvl?MIdDn1mKX(#Mjz9HJS3Lf5eCW_v6%P6xL;l)s%R1w+
zxVg0IYH-`-&no+bPmHEo-nmN3*4tNZz684RpU)j?x}KT#RCBRXYI!K<rx~|FkJ2p%
zG9ER|`Ssbh_)DzE_VaIL{%7#PF+MUNi|spo&E~mDZg71~?O{<{tw)pJyS0Zbeh7;9
zO71jo$ySzlYewvS_DNH%qC@h?+6VJd;#)(e9b^-Tv1tC~<VzCS{5JPYz1N_`KkaDe
zfw1I>)}G{)llCq1w$~N8Ez*98T4p7R<eV^!YqWDdT;-INv#KR-X4mM?mw~G!%X1l?
zb#M0j{!inbMxQxcpWdzaW!3!jscq*}L}GpMR7}bSY1vWtfcIr~1IOwn6KhuSC6WHs
z?Q78yRpi0Lk*E59n?3jAVSaUu!O1->UzJ90b;}Adys%pnO601Io#Vht_pWiiV#xJX
z!sdfV98dL)@o%MPZuz?5#%LC&bLRE7km~s;pOl9o1^HY1&y`+r_m_3azsdBqd9cYh
zZ{M2c-_Jkxl*HtRN^0HbH1k%L{MY5|t1kzgtv|UD*KN|Az2k=jvTlFQ-F55Ne(z^G
zc6UkbwCi{CXH8e6<^OWW?)G%je*^Nxb+f|P>r}HBD*pB?!Q-X=z{d><dq-^-mCLiu
zZVtH&Jj{Lj;YT5Y(AtKKuo-ZAY1MW{|D#DoVR7EZSJZdS6tq_cH+tI{p|SA{C5Oo|
z)CP*t{@AY@el7|3iZw!GD4NIQS!e@IuXyw8bjyR{<3?y4MRS>418tz`l#b~257|3b
zn=qm&+L~)^uMISvW>#i;vw!1NW5!yFcF}k7)&`nBy<s@tX5%)s5t=~JHcT5&ZJ_BV
z8|*rTp3@#isGOoLnU;pyK-2Fg4IAXD)D9ywilWV#=7!op(*+rsg_Q=n^;@#F>#pd1
z)6hCDwKG;?KA4oA0fhpwl@O9)5jq5X0Hg$g08#=OfRwlo4k__K98%&k98wZK1f(Pa
z2}nuE2uMl#kdTrLBq1d!BO#^8QpEkU<w_Ax#D|cL$c{$zJEvN*@WP>%$gyPIvM~jg
z1)*%LNQZzNKuVwhkP>7ANQujFNQo<ONQr0TkdlxSkdjakkdnwIASEd$Atk9GAtjkj
zLQ0XF$d6vOu~HNu3WHF%D4s_2J{PdH`x|OW4}$d~x`kuI2sQ%a2?V<bV>ODkp^&~C
z#V){DgJOda20aLX{lPE-V1F=y0N5X>Q2_gcZWO@&K!XC<AMZhM*dHH8aM&N8KycU}
zSED%Wk9VUu?2l_u9QG%A5CZlmh7khxCngXA_9xUR0s9l(C;|Hu8kB(j$sUA+{mEg3
zg#F10goOP`HA=$%WH(B}{-g#ab^d{6PZ__xD7ua4DcwYF3VJt0cvt0fh2{T5jFmhe
zUNHT=#6@z5M|=GAskG2gXhy5)Ly*bV9)8f?QDG&})&MF1LtKT!kWdjYBvqsi(oaIz
z+Vhis!iV82uQ?|!kD}&d9b_Fx&tvD`+e>@yj(04i<E@o+`inJ&1)p3>E!3wDI~Hc8
z+oFI4AcV7U2oWp-LL`g)6CIl>DNY)%ucV%0YiMgq+qJ5?^rP~cLa8*@l*^^RE1E{x
z+9RG`f${&ez1uS;T|cC0imK%Rj*z~&dgB5fuM^zpqtb4AZVq%auK(-eBPWiuf=Y{6
zA`YFNbk%=hV7BCQDlN1U3g|R_nNTfUxJl><OtD&bqrN8=-n1pKbfs1F-NIvQ?-W{1
zb&iMe>9tTufH^=2#~dLf#@uxX@BmZ?1OTc7IDqPa7(jJ64~Ob-0S?vSI1bg}VjQX?
zcmz~O2neW-zzL|15ED=x$s?gUQb0m=Bu+wgq?m;2L=GZAx0-q%gvbd(E+R3F3Nmsx
zEZoqA1!;w`3xlMEp#3dU`rMSnF5c0&otc`Ll9a4qq|g+Dk*FIc>xK&5FjY59%S+GQ
zM8WXPh(&t(Oaqpok+F%nnQ53sNHA`}4h!Nev0Ti1&qpHl^AA|@7m<~-@3Q5j;QwX0
n{G~Q__6`^z9GzUmu5KP)-nO3BLid$c+*Pu5XlP{MYGlnH^*<E^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp.E46BEEC813D33535.idx b/.cache/clangd/index/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp.E46BEEC813D33535.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6a62f99a2dea918b00b1978f8b0b87aae1a700d7
GIT binary patch
literal 4736
zcmZvfdsr0Z9>;e^z-2)$i{J*r0?oX0cY#4%Sil1a$$D8t<mQ5!0<FTKfOM_rloyIl
z2_DfnYM6+6#B$9t6Qwp+QzL3I%zLG*cL^<Tn33nK`#jVB{y6(Qyv+Cge!m~)_RbEu
z*;!d;rA(HwAlqD3RyWT>CX)p^e`@P$Oka<X$<8WdGIQ$}`BQ2yBm`XBe3RL`qO5CS
z`!njwB7J;gRpaGVtH;U<n^ntYKFgX<&AI5QZke`YWMWl0q3BXS{IgBD%;Qk`hR=-n
zlJ?H8_wF!I$;$`MZN54D@b0aF<S&X>Q6r8WFS!5w_|83rn>P8W*51P|?|b{|V8((U
zOM5@EzW$#hL)G`{UVeT5WKBWOJF&5S_j_vI-EA#R<!|4v?&}G249g$Iy|?dBbMR-A
z^3CH86wJ;)7k+SE*zTdyrKqx~sP|0h$x%rwH(C~LCue5eD0b+sT-5A5b^fLO6JO}m
zO}VsY(KN%sl6LR9^k-*zZTx7)ACr?RN``-GztE7rAz5j#&G;=s9MDI+v2F&Id~22c
z*x`+I8Ga%qzO?A%Cr2CnH>al8?FtNiu`RuOT2A}QT~TW_^}-t3b|){ubMm)li*CG<
zYgjRI1*!Nh+@H9vJ8^oQM@OqrWzKIfT$!IkHrM;_4zVR&%orV>IrD?4QK{=Tb)W6&
zZ(Ms|=kl;c54T%VK7Gait$eR7SDJ3UH1|oP{cwx!=yqkjBKknjKw*=8OYTy`KO3fQ
zX$h#cbmkM<L(Ermk5~^kRnu|$>^gbro~q!JGIm0X^88(P+|~YHQWh9*zTbah>xE`S
zb~BlO?~IV|6G!i;Z~i7@_tCeTjtjSZ20EU^BsR>AYFc=N?5e)|aBKVN^7+5m4>jq6
zwl4bVc5JU$|MI4?8*LAhk4JQ*B!*tt5O%Ko^WJM8hG9MWna4l<HvI0W%8b^Q_1nhn
zUN5#MKbf^P@Q3vL@x=!NI!q-iC)iI$Y2VeA2G7W>qSs#=%F&!V;dg*;CvyXL{QB_&
z8e2GAxvzD6Z(_>)od-8%7dv*@if4Zu)iD%t+F044>rS-jgU3x*?(xps63K0AW-Asw
z*O7O7#i^f<k}-)D@z0`)%lE$Tek-3nbtn4t;A)Ten}-JXzP8^?6}B!N$ele;*RV4#
z^+kPT&xaeY7PdSWmN-3SZRaffq<Q-0ZPWL5|9)@(kK4-Ewv7CEO~(iIGnT|wWO{$O
z+MY#KT70=Ya#CDZf0*$^%%g&Zo3}H2k6#Eox1jvxPuk9$s!$%h{rI)awL>}D+1E2`
zs+V-{uRj)-qVF<%Rh|@n?S-VOcc<?7^NXADiMiETiR<GY+J}OJ!XB?ueDk1}3w`rU
zo^{j?)3P55=#3i|U79m@*7wPN7cPxiKN$1f&}V)V=U4xFG4lLi(SK8`H5=~<kLQ^0
zecAZxodWTJXU`jL1?x^lj96tb`JX+(_gt?WUi?a?I%|DfS@v!6)3MF3%esfI8f%wU
zJ^#qLuD#TqO&MDzMEc+}^JOJUFQvC@a5}D!;&I#0E2oT~>4WDwZIzcQ&^1W5`mBRn
z+#&k-%4Rw3SkKrAu0gUV{_)#j$F-#&`QUj@JH#twq-&7uTk|3h3h{%Y4{mVUL0&=L
zu0gT~N#^3tV?N%aWHX$0Xh>+3Ymn^wpSOl%+ThVXc!ARn_X>}94U+xjs&|{gTzbF<
z&vDuTUIE^&L9%Pg%4@rP-N!#h;X3Zd!+*5x>eog3mJnJ(C;bVh(i79%iqw+8BlRTk
zkQM<C=@IZKEd@MEPXUkC(!itjH1HTL13X600FTwOz+?3+@Hj08JWkI6kJs|R<Mll7
z1g!u(K`#JL3=%{By6h!lxELuZ>0XIexb6{fDh))j`{WTuq5>!rQRh~qfdom?NP;A3
zB0&-v5J(~;0!d^-AW0c0kfe+hNKz&WBxwT;lC+TqN!mn%Bx7Jek})zM$(R_BWDP7x
zvPKppSrZGAoPh&L&d7lzXW~GTH}D|I8+nlAO*}{n1_2}mqX3eENdQSPR!o*Iq;H9s
zDrN&UMa-9!^cz#SR=RYNrIntdaIJ7@+PQFS7;zYLUtyvTBTho?z=+FGTX14G4)Ru<
z_!4S6PV@pz+AtFG$v%vPeA0oDkWX5067oqaPC`Cu$4SUXHVi>N>cbG^BL{{cA6ak&
z`N)bR$VYY@K|W=}D9ES!FbeW12S!0YWx*-Pr>r;y`IH@}AfL8jH00BL7!CQf1EV3I
zw%|16(^i~@eA<rFkk8mK2J)FcjDdW{fiaNJSa1gN87s~}K4Zrj$Y*UB3;Apx#zH>p
zz*xv<EjSDLtQBV=pS9yG<a0KRgM6+J;~<}NU>xLg7Mz28&Wdx8&)IPf@_8G^Lq6Y!
z@sQ6uFdp)G3(iA6Z^e1Y=j}KT`GO4-AYbUi1jrX0m;m{L1s5P+u;K#b3wB&^^V1rB
zk^O$Z?mXsEk}xj8C@ubJ6Pl|-gjN5-d`nbXRfhE7&zr8KZmdc6bWX>4$N5X?#ZxAI
z%M~W%IHyC?L-VAs_2vKQthZh2Kk#%qKeR-ees8LG&6`I`YZcCChnxs@KM6=V2^5ke
zpeQ*76fLKLV&n`^tegdklXF1vavmr_F8rlF99X%usK2$zxvvr45q{E6?J{?H=2W$3
zIj2MXLPDg6pfx^sfoEp>fOA?&Dh25k>u_G`%{l5b$DPv&9tqLX={f!HLpNBZuF*Ly
z%0;C#eKp4O{z(5Z?>ML9g5%WEwDiv8t8l$5fd_3GF-h7l-x4y71PY}gpr|woC_0S>
zib-RDV$)clxHJwZK8*)TNE3h(RbrU5!vpeNtMQ~gPw$l=GAKscf9Z1jC|oaf;5f|0
zO7ML^nn@^;89_;zDJW?(4JBh{pk&P~l$@D^k~i~E3TDAArSpm_T(6k!orQ}+dg0_1
zt`|*rh)@w>62W;>tBGi!5WxVYA>!PMRFR-Zs!7lz5ea&vh6FvNLZFA#2=ov_pocUF
z^e7btdX$<1Jql5vM`<X~qg6EM(P|p>Xhee^t)W4WQ8A#$s2R{>5CeLQh5<cR#eyEI
zW<ifdEa<Ts7W6n32YQ^E13eCLpvP%A(BoA+=<#YE^mxRB9<Sj+Pf!V<C#VI`6OaIU
zf<^#6F<K<0)#H37P$CNyFKQ$ut0}9T_w@gn+M4;bGoNl0Jsn)15hcS5AH3Q$oUp{J
zv~YG&u~9Z#vShMC_hzws)9BtTb#KnET2T3dQ%)(LHp0WxYoy%U$9GhKzh8D>W(EpW
zWTz{~1dUcbqfOK$r=(8&+qhsZ>DdHY{r_e|!p4S2Mnw}O#jw04CRQ9jAtF9hh?^7~
P@|=DSo;5vfGB)*pxYsRJ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_layernorm_cshuffle.hpp.7536F92E23B638A6.idx b/.cache/clangd/index/device_gemm_xdl_layernorm_cshuffle.hpp.7536F92E23B638A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7af9a00b282635f3122d5efc564399367326839
GIT binary patch
literal 12868
zcmb7~d0dUz|G@7#=Xsjk9@lNrzG|VmDUa$V$y&-1Vp0ZEWB-n2#+n&J3x!IyjID$S
zV~Mgwi$X+^%AnPjB^uH~w%_62<DB{H_nLW`&V9c=pR=Fy`FzfKPN2W9ua_v144&#A
zK4;p(Fis+oXygAEESx{fP(vcQU?7o%@5~t=GOx-*emipoZ;?G9bH@<xjqbg&`DIh%
zERUO13_Lv#jx5s&^Vjn$_Mh-#McwXzVOusX+|fDo&V%YVCBOKd@;?8rm*HQrTYWv!
zX3kHr2;cLm&yab`1;^$J%S4Cf>z2C=eH()B|KPDNer(~0gx94m<EETf8h$i>a`?fr
zJz*YJX_vi~dgGU$xXWc-wrFvTbgmewIcI|5*5HYpYubkFDI1rsJD9)rb&pNr?q^%X
z^{{`<e2s;fa}Smu4_MlDY{4)2x#_0k){7gKOwm8vQafhmww2o#EwX+Z<M7~^UR$$J
z_vfJUh)plo{&B--hV7JJ2m2IXH++(G*LcbuXQT0pE~ic%aa6--R`6LTzrEWF=e=o+
zUlzDv{Kip^&*sgQ?96-JGj7?;M>SK8a$Nn<`K}%tO&sR_G4Oa_EuDi0&g|cizU<kE
zqaCFM_EWrn-_krpQn`26s^LESm(=b%t;2an4h!j&=pb*Onisg|%+^eoO{NQnAJvp)
zL|pHkG-CN3tNwo;vifnqmvO@1JHI<yaMIDaDJefLen!i)mRXAQh{w0=*Y@-LV@tF7
zmc2_R{Wd4qIOij`Wk$`$Bf%+4uAB}HNn3X7M#j$D855_UKeXjinwV|8eEv^%Ue&sj
z7i%T9#fCaP@ZF|wUv7WweCG8P|5J86n(DmLb9QIJ_@3|U8F%aF%rUZFoxOa~g~sPz
z%Oj#rR;1qgJoAs4b59Shm^fQw+Nq)Te{VbZ_n;jo%Q~DcUVbiKGWp?J=*yRmy$enq
z3G;o=<xT0kE%M|Wi@ScYPs=8aJATb++=gOrm-rW(j<)%`Towu&tm<4AMY~1)`dP5w
z^mdo~)i)K_&;EUI%&u_X+sT<t+S@KCJli<Cu50XpIm5S_wd=mWzx01*n*C1c#s`IK
z3>x(^vgg&d&g-o!wth@X7@GbmH@NZqJR{RQ*JtfD-t#W+`1M2eLZ#QIM-%%G`Fm%P
z<Adl41AG>B%pB?6siE|sO{Hy?UBU2{4=489X1?g}=;pDrM(pw3x+5WZn?ojKHZKL~
zeQJ6QYITiv8hB}A+|9^h4}S95{3mxhh9sHaa<mP86moC1dGdtv>=!@$i5ln6oaEa#
zcV9vF=D%##ogH1_^zi)VAH()Tov`as3s2OB0!@^18_t`xEiLrd#rARI#?B49>{b1X
ze%I&aF~9LK4W-5tle70LuNWP_QV`zL>kt2KjkjZ0Ejb>X+JE%g5N@lL{^h&gR~Kt+
z(zFhFq5ZsDhIW~x=3V?-&F?_orkR`kB6N->HIGSkx|;I9{Cn?lbyNBpTZIo<Ike|B
z{SAwU{nIsk!^|~rHJrALzn2vMt6t=$*622=J6hXx=FV?X-Cy_Wh|?Nf=vh!`-g6K*
zXVPE`Tbc}qjCSwVq1!oThR^!7)+tAauC?SWr-%7n8P@Goq=$D}b589>M^9h-fT2r%
zIWk6K7C-gwP5FWPfs#-4bB0ve4i$eEv!bgP#hBQodHRkWrlYW58!=AWDz)0E3FKUt
zD@8iDZhqXn{m)J-H#c-UP}JhXPYoWOn(IGA>U3cZcOkjya@K^L&eMU!D)IN6K2v`T
zYIU-^xI1K$nR#AvYF(LLajw4UIc|9C;*tz64~fp=8dM)-+9lS?`+(NiRb3RbgBR96
z8mQ%YG&eE4YRRCt=904wcaGmq7`rb^W1nX3z*))b_Us<4CwZ{a(8j+`p0NGN2&<^(
z>7#;b?0Zcu99LyC)gkO)^-pv9nh(w#b|(0Ic%gA(&BE09qEQ*Uj7K<!rW#s1hdNm3
zhR#BF66aL4E%lr8<4*>|Pi0P7K6F*IwW*h%-JF`Ni4WX1+n-BtkMmo9ro!%ro12UM
zQ)5Q-yAz@rpa@^~qUiPT9XIX-L~L69*7=cTQCUfmgXLYDUth01YPw5f^sPTzgWcD>
zyyQPQZBLMS%T>ql^0}rV@BeC+PHs)<t-GjcQjPb!{Dpqrhkvx%xYHnQ#oW)MR{Ykm
z@5hcmo*Yk0uBl)3pvTefnyb8`WS8gejb3_q_WdeztD(}#*KEHZ8ok@eWBlT7KGPac
z-`DeYe`Q$xdwAE*D=w_O?sPFA_f-9?N#0%0@j<(nR{U^maZJ*zlyaYKI`fBrh6Bw`
zY+bN)&QIg9e?XUS9f_Q<cd!haIY`8IWZsS()rmwC`w{#fzohuQ@BE!lb)bR0L`;M7
zG}v985Ead9MG5Xkm#Va&23E;H(hOv%PQIy>!8>Dqv%VvP-Ec*1$e<0`s}rKq`P{At
z&SQ!O$)PbW<tcCVl<%!hh>G6%@@xCwO|{g3a;)+RX?;RE>V&8mNhbKNUX$HN3nHx2
z4yEnTOq~#w?yr5GZiyW4EQ21nEl-irQ`AkJ5EbV$mB4U;tyBl<<5F^fJO?<b6Qa_0
z#Da_ST^81Bz)n3RV!(F+-M{{kbb`QCA^{@s@^-o~Bs@@02K(Y_nvii5?y@Q&Wq939
z<H{|r*6BbUTul^j9L2k-6Qa_u*V?H8tzd)>)W#~SdHHJITAdJ;0flFlt+>B_sSH}<
zikeZ^W@MvIh|1vgBm7`Mn5+D2DO>nXTX<7Zl@Jx5`j!0F=vbu&<ZwlqNGlWRsuQ9z
zaz@&`-ZL()*MYiNB}g9R^!3j-aaQxrLsJ*bkimYq9WRmbOXQ$Vh|1_0xuL)24SLcQ
z%5Vu|jmMgP{qs$HQdTmy&#_BdP{87kQ2G(-s}rIUX!x+#Zsq0sGH8b@d5!d6BRh5S
zP33=i2m0JHtI>gaxRf2dd<X8VDj_Q0$M!t6LVIPI7SzNl@1XP@G*Bl*Ws1~!l!ec>
z7A^Q4R_TD!4%`-1LR4laOmLC@?3W;eR=6#%km)NVs*`Uj1yjanrSFnxK`Aa}v~;x2
z*FQu&PiyFOJLBp+IW%#Ui0?Vsdrn)O5S0bN55q71^l^|3dg4l6AmbOPhdLoD3(wb^
z+iAGW)q)%@C4!ShaQf<msH{rb<5Q_vH(6Ct6eo+q6{!-UlCa5Z_89NFDjjHun_3Iy
zwa`YL5S3Jykw#rQZQUh<mbfi%k@;I}E>sCo*;k)zJgDPYx(wQ4l@4UyfmKuqQ8|*U
zS*mP{+$MuwSfvpeH{y+<N{C8Ur&$fBb~sq-Lk_E)M4Bg&T%8b=)B5oZ7wvbA=?XjJ
z<^~!EntuI5#Kp-rIW}E3H^`wD7B}(oCf-b)5S1Gt&QH9zKXsJB?zoZ`B(xw;bwX6i
zF8y;-RDMv%VOLyA0pG0vZ%|e8O{F3t-M#paQW-SIrTl|T|G|qul@OKcUcpggyL)HI
zpf^@|j&z?RQ*}aAY7Kk+zIn!#i85$`Ro)@XcgR_t5S6;1=p(uvRcSJ4gH>A5z*h95
zIw30cjl27;)qiCtgMF~dGi3A(TQ*ffRG#eVoTxXuvQ!5e;U4Al@_gPzoe-5b-*1h4
z;Trk97UZ$YTPS^tO{gj%Dj!pPU6yXp3+xKJV3l#k<4nK)A!5n=Y10>^*iioh1+fCj
zDv$}8SgG2}5UWGH43esS3o)+(-j??7-~h5?1@Q*5zkvpjf;7&#S$g>#&5;Y@Lis{R
zVxk1`clqxO#RGXfV0b66-wB3L?Y%(0mz8G%c_zzMKwbqLiCYlF2SEOS<u^e7hUI8z
z9}S04{k2fO7Bcy_L-}@=bD(Yx>`6vJ5R0I05zA#zw~XaRsM`p8k}XRdFfUcRT{?Q@
z$f_JUlhG2e&(&;0`kHEYEQnd0UKZDlm<K_u;q+=K6Fnq|<w#bJENMa2$gvvrBP>-r
zb26XOdA_%=Z7*KU31SlPOag=HP~QS>w}3Ytno{Ul3J24!KY?yfpf~ASXP4F6|CiLO
z3b!E8BG8uXdO;j(G1iv26G0qjG0v7ON)ku_sXBZhqd~$4PO1(Xh(N$QpnE>wlWYz_
ze9l`w=ev`VzOems-i5HfV_0cXz_|{*tfPTvG#EsCR|q`|;ULl*!@rMfen>wpB2L`R
z_1VpNl7Z&LPn^Lg&X70&PTa~{Z{=->qb6Yto>5f{Lz}t=X|6#gqt&ibz_v#7A7oOd
zHba7#!I@`py+}oZc$YK3OWD|DML(-P>u=%lY(*BWXaF%Qf|$uUW^yBF#%j*7nzDr~
zvd<dL>-T{mZs$z4b3KTnRrdUuFCx5R0EY$o+c_g5=+-~$Y_rEr>3{CLUBR1pJ=v2g
zTvZ0b|AvzFO(Hiu?w675GBTjcPZh?I<>zpJ-{RL7hsrbsah@<wM-|CRB!U<Vgjm4v
zRv>I;`6v*MvV0o|w^?oiLKDlMf$*8-Bq$`oZ{?Fkj;)$1hNQFj7TUjszQoiDVkvSi
zMI%Um`t*(otS$VjRVs*oDE|ng%PR$#qyTry`+#yEVB{x(@+8ZTfbtOtB<&EyRnTM=
zbf+8-mGO{~?}W;oEMJ8KufmCBIb$mTP0FD=VNdhd#XoC=y_R$08(w(B%ZS(I#1>v?
zVY!tzYUNF+yp1<%W4WC-YG?T^Z}gVscf9U9zAMe&!RvOg{GK;>&l^$s2j1WV%O81@
zkGvU`f8tF(vHY2L`OJ${9--k9p}}yZhEt@5GnGec2+<nf`b*{&O1~Zg{B0J;ZO9(H
zTo6Zh9ql$-9RqsF`d$3d*JdeR*7eY)9u6foj+_+mvc8UF*O3`n22%BTfh=osUci=3
zbyy(vl2ZZ}sE!CkK+XnuyWd2zo3v;_e2DxWqVcpU)xm*OMGO%xQFTrrB@%`2r0UZF
z5eVu4jVK5>1htOf97%L4;HEjO=0(z00h>JKR=z)}Nf3|lU5@Yuf$H#G5G#4hO1=k)
z!G@Oa?7L%v&jKGo{K54HZ@LF2x=v&$)HRf$FxN1K7AY1nduNGa2}9wEaE6vCma!DD
zEST1}T(O*ytx&9BC;=!E*!`IR0ulg|HxVcjS^0XPSkH0_kf#7+TGuWh-^Fq&kf*Yo
z1{7&*{xlGf#@4qVDE71Rbf8FQISVMV{>=edZ2nU~af+3n28z=x7Xa4+)|nRq*Fu(y
zfNK%USAgpkmaBoH8Zh>#7AR_2t^<lXmLCGe!+&$YLl8{u*<+x1%*q>pqJiZWplJCw
z2eh#H+km2tmA3;$JIkw}d=)gN^~FGW49l@l9?NnZRK&6NHVy{FL8iS4P?5mO*FnWP
zmUlqKj(>B&4mN)>R3x+V6sSmH`5<&X2)&8D7Q{o)^$^QPq3cnWk3rXCEayV!T-b-^
z&xg+WEMI`m7g)Xu<yWCG)h~wfVwSH#`8Afyp}d^s3Mj8&xf04NS+0kE^)QeaPC<MD
zonOE{lwU&Umn=6y=O)S{M=1@i_-q}6E%0h&y&Bn(D2i-2?5Mw7YfzKc4(;2aFNwwk
zu>?7nP&YkNedB=*_cbKDhRmroRb77&TSx*`Y&5@IdeHJG`b`{0Di_4C{$T?t6@vbS
zU?5#<SAgFYFpTmc=ywPXqhoyx_CE#(5++@fOVu|hw9*V@lYv}FP=zBz)a5Qcnh9bZ
z(yL=b{f9{JA<J)&_8VkIj?Z5LN$pl-Mgz$tPM*Y>k}yyZ_i<hJaVC@xaC!$gOUfrW
zvlE;><xI{jljSl_uZ-2N;>@a8c{OKN&2l2&C6PBFeumr_@n4RPvKw@gzMN@X%IQ&3
z5UY@573xbS5xZ$*SdV%Tma6ZW$fOcSg<YiTc8QGe*iJq<bEe7H;bqpucWUChkg$Oq
zV(~Il-8qpQf*6N1<4{*JX#%cF^EK*9+8z{^6mocXl_xGSMngA7L!XowtXtG8^r)e&
z26l=DKlG-HC)6O6p#oq~0D9A<Qv?i(SUw014#M7K=?LOcXmFJBr0p+W4wbj=`A!h0
znN4${l}tCA&d_yWb{)9TN=krP3Co$#EEBrWO0u9?7RwKy<pbE8?Bz+P`^=tbXJd^|
zkeSFV6WNnt7w`p{;d{=89GwIlpjdC=ooO6V!0Q$8mLzT<7kPNlRo8cP(BqM2Jkq0s
z-i$PxksgVermf>`BXkxF!h;^rE1)0AHa*DfM6JB43l6EC1JCDR5YuDmSwatgg7_G^
zJ%;^AkH~c)&RB|MrO1XXEJ3`79Pgolgr(|BMABmtKz5RTy*wlh`C^T1OOO-E_N7?c
zQkHKZiyNprl}B(!5u6nb$TB#+40fX)=k$)Ve3#R^%kn)=?;d3`0eJnZZZ66Bnp|08
zvsjE|#mJPjK&rmDBpo9y!BeBUwIn47_-@iLirw>{ILA+1Z&Kyr;X|%F2D^{e7sQ{f
zer6W>Y^&J}MFYEN;6c~xN?^B=<;}ouGs{W9E{Ww^!0HyeP0E2)Im^$1-E-hU=f4ry
zHM0C3*u7`@1F-wR@>ytgmToga%!XFkEEht%Lg+zv%tdH-k>wI-SHf~Bv@2!#F|>M2
zca$JDK&u9ppFpc8EI)--Pbo)Kme)L=*GCJl*Eh)G4H``?{t3?U1UHdd>?+Q&it@_6
zD}Eb!;8GEu{}7{)Zgl)7n@{deXqCsU*OUISypDHQB5#(+I}m$KO2x}cbwN!Q#hOXA
zJF|Pox8NKfq0vV;l5E_#VQA&c#Lph)f*8~@$c@y1$t#GtcnTU8G=|D@K+hcDM&~XE
zc;$e>EDy^8V<@M=9%;~-&P5t@PlNqf9+C!!Q?7-!wK%v_U#bh@d8B<F8IqkKh&f0*
zhvi(Poy&3_(#~T!A8F^ad;w`+U^$u7Nv3x%g1Cn_+r!%sA0>!;d9%GNr}1WKEbrsZ
z_OZO5H`~v0I&YTF@>SmbD&LP7S-jJE_hOc>@$T1HzRtT}XSsxTFJZZqcQ0kRoOds$
zOvVvg3e}ehY6Lr=eFyAEto)aAuK7(Ql7Ze_qP;rS>B(Ps(X8d=Yk5nu%#!Xq<PXSA
zdyWOKc*9q`4RNVS?V*#Rx|mjC#}malMR9$|g8ah6qqy;uKX8U0SV#1MbN|2%B1US<
zRloG|nB@w*4JvuPO5TDF)^lF(Id4I(6O-Q^TXn}E&aAT_E>bQUMc3$J<zj}`0p&Wt
za0>8F0YOx|8Yrv5D7s>6fU<_=R-kNUc@^|t1%s%5F7(WWfpi7uLC-vvuR@cntaU1f
z-sLce>c4=VFKBE|z7>7xJ(5+@_)@C=Vnm!k>OnK@c-iepteDQ(q;vjsd~b3#Hz}tL
zt}+``(6QtH9Z&KN3M>ANWPh_&tG+~$st=C8<t;t^vlG51T&7$$k`848P$mF|(||IK
z<t(7gV)-FZJ_IA_P__VN3(Ij(83!5p4yfEg`Ed9-e)OC@zTSekM7g9tZN?JcB>|Me
zmEnvm+&7$&EmtmQWXpY*GqM%R6^v|!{|d$mCIV$5V5Et_HxV#-*8}BxR=ytiuLn&2
zUBF})u%>O?1srw(k#Z_9NoD1!z#)~D?+42LZ29|v?|!!abf8RU<>|mbovr^AP@ZDt
zr-1J%Hveg$Jk82a1OL-({#u}{1^wy7)dJsI5J0&ODC<~x9q_GV<&S~#F)M!zd>^y&
z2B2(U<qg2Uf%PhFK-tF1+kkHyo4*|>+gW)#@NZ}H$3T-9XieK2107<ZNI4do#Io{O
z=n%`w6QD8yGW}10z6p@&|2n8#$I911|8<b@ipfx!%*vCYZ!()d1u9cmc?$GTVe{uh
z&wMsw&WHZ_a16C67og__R(=8cUts0M(4-hzQ~hG-Pz*)N*PzKYR(=gSTw~=G(4>Nu
zS3rjfR$d8BDp`3Ybf{$IFQMm4IEZ*BeBOheO_Y!H_8fgFIrL9#h(00HPpCTyWd!_X
zO7AK40OXT6R=kR2SLrpaRQ+L`SS;e!aB!shDoz9^8-F`#u(tWofBx8V%KgOF-^*V3
z;m<d(c->e0x1%Nb)Q0c0zkFS%4tF=_wwqHDD}DY-Wyq1fZyx{WigR59r}92~8gY0T
z;t?`{+KhbQp3mM9rbG91dMA}PY)!Y9CBGD4_nZ%g<+Bd$5(vBm#?!^04u_?~QFJjJ
zhJlCSc*1#e?)~-4fIDfp{-51`W^Rc;0M8F>BzO^eUZfu=3lEH*+u2uh<^Su2F1Aho
z{@DEdR2*Q8v>Dl-j9<~S?GedO-OFU0*vbp7>^G7&UT9;vofq0!e#;AQS$@Y0?^y2O
zg$|bA^TK<UKk&i_mOt{sN0vYF!Y7tL^TKDABQ%5v`pttABQ=CbmZLO;D9R+B!b?^4
z{DO?Y4KKT8ful^f{O1)z&L4Qj%8;xKd6C^9h_y&wi@XU-)z3J{<{*I=Zi?#h25AcU
zYKluNL9!C|NPZVN-er&EswW+!MDnBquYxN`c7^>Cse0c*6v@L5tauyAZnHaE^~!@N
zlIt_vlv_x43n@rv@M8lce}H-tmZ~3pkc{;YBL-~H&Zxp0Vz%3C#+R1^w{q5ZWJ9-X
z*q<zaY=WTMQ`n!_F7kE+H?#=Jir5~hUXYOb?t9!A+<w>){O9R=WVlbUzM)G`+|X>!
zF`IKHCPWa=agOI$e!w|CVEG~E_>l5r7ry1YE(US9sRewe0$!KQ?-#S$=_#*E0_%p7
zgE{+_jU0t*TnRi^vab*tfoCJT$<IQ!v((6ZIV!s~&{qgg6lah0Hv4N2uJH_~cZPGI
zO|R$l>M4^K1bC&X-qw(n)~I<n5qf5<$J0Bx-(){(?Jj}7mw+!dT8ClZ!_b$^(#!V~
z-;Uf7qx@>^jPijcRlEeuF9BQ1>Ch-0T9S-{co>==hPH&8lor?Lho)Y{J=hF9H?vO`
z-UH9~^cJgm=h-~x3@e-eJiX*wHh!|8dMre`j!#}lb1nO1;T6(+g-pnOm>?eK%#YJ2
z3ivvVGrvdq{rmlS&edDKocmWJ&DF??oclj=hlCS4r;@Q?CC9DgI+O4G<Ru7pU0>cN
z(T42iLU(gNk;$7sZFbmnoM*^ktp|GltQZf-$!+-W7P*ln1AhhNPLW(jO2{T4cY_jA
z_H@H(+o$`q#{y#Zi0_vWpTA)Klm&@jTG*2`QvyX#Sv8okah;QvyV9+vb8nYkt_mlK
h%Qppy6aDE-fA*$7d(odt`cpxFy3(KS^rsvB`G0C^?#%!I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_skip_b_lds.hpp.31130252724E07CF.idx b/.cache/clangd/index/device_gemm_xdl_skip_b_lds.hpp.31130252724E07CF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..278f35dd60a2b2077b2ca304d2818718f8259f72
GIT binary patch
literal 9254
zcmai3d0fp|8^7oI-L~7$?N+ztcH6Y7^plobNN6a<kTkY2DKic67B6FHGPG#3R7@pF
zxR|D58oN@A7|YBU%(Nhe7(0VRl;u5j?{Uui7oU&&{Jx*(Ip;j*oM$^vkzpYrwoW3^
zM_+_3Nr?L@RxA>UjPd`4UoD8g$%#alj6|X(rS+3q6I%mhP314QcHjQ7|Dg3D{Uk3V
z#UV5Or|4ZhV!ZFf#T8EL?A`q1FD$>dRgk(}8dvfqCr%B#Q>u{OaY(Ye_J((OW3PXF
z(7iR^Pr3Q&%jZ?sk~@8ticd<tmWo}Z^Jmtq{9JGF{w;G~|7`eK$+?}DMX#=`jM*@3
z%DA9ePOf8@m#<uEJaXp8gFM14ihfJq9uqSv;CG37%(^>cM$BC4)3t5nH)|@?D<+hN
z7{?jZgx321v7}VEm*;*j=5ER}rNfwDow9)I4Laj(p3J$fo+VgCH9s4-Z{A-IBwgoo
z_NNQen+`g09W&~7ZH&3@-ulzkp|{XH=P&Q)o^D;^6tLloj6X+4n;vLZX8fo-JNN$0
z%(lJT9SxfOUOAXLcwBvwpXm1a>>=ucQC|<JIsPtu=Gj{Xx6Qgc0+JOs`7x)DuKT29
zuVYEYpPpxh?q{=qNt*Pm+SD(6-MP+bH<EM?>nMjkaI#4Dy?Nu0ekHlqwYN&6CMOKa
zG(Ei7o6mKtHF(?g&$kbSf@Rf*$MkRgBJb>F`vd!@=^Z$l8~5k#hgZ!H<heen_fTIO
zvd6lszJBg8zsKfjbE6;n4--<0hTi=q<KDkr4=s1a&fA@EU59%cxOV7eJ+pr!4xb7P
z{A2u+X+x7&*DqetUC$kFb?=_{OM=nPRkn6b+pZ)^HrHL)vHeWJ4~N_~_^*4=&~eEB
zz}AdZmm$Add)B#K`m?w=Bukk3ebnkU-MTUAvp)#?xBusnB2{62y!}gVp81l!=Qr;5
z9G}zi<(#Sd$9A}!5$=!uHp4C4CD6jxW7a+0F|+mVnm%@)B7Y^Zy`3&fT3EX7>QiU4
zxUl<9;XD5u|L#y*fnxGW)xRq()-3t6VQj<6zpbx+lQ>=Yed}Ad;w<BeX7LR;R_%8v
zIH%^oS1Ah$AmLWTBtb_%KKr6*w`kxJOZ8V5`|kcVFWyY9p7Ph~E!OFKFaFo&)5@hW
zrMFX`Zp`Xur+$4xJU{IEC7+MC8%)d(nAbdA?i_b)+g*bt^JdgVH8nok(N!B_e>7dW
za_O~ePfA}ZGfa2hZMJwlWUanW>Ck?auPra<deo);CN^{2Fn(&OnZeVW@`CVfkXUkG
z@94@^EB?9hc#W+3(UG(XnTF=`qTPmnn-q4nY5unfrt9ZfN)krvE?Rmu`|ux6TGT(R
zYHXa-=j>D63u7&}m!3=Y&kkR>_EL_z+vy|8ZPBiXq96PNMy!du725Jnajt(=`{ty*
z57*rap886i@FYBRwu{N=L9@>93yAmgw79%px!mT|W51=+e(Gy_TL%nkS+@0`ZPiB`
z2EKlkGQo1u5}!NEuLn2TiYsS$m7S?es`RhzwDZ4k>b_`UQo_t=9g#?PopbniFaItu
z2eFGtxQX;{BAJ$m^l?nU|MV9e`?|BJW3UY5&LW|ab7<uJw1kKl%%3sC(kv!O3dVSr
z%^=$haxEbuytM7k2n9@%f(nZifoTy~YY7qQ^L=%Gzu>8jhM<ENszH((q|_23B27M(
z9U3t##SnC{NG+1oA`2}cB1Z8K*L+c){#Xjuc%k)RwH^d5AtKgpV?LQ}cD&jQBzTtL
zR>Q4(f8MjlJ^aLR?lL_o_~ID~Kv4h=T0%r@s(v5*_;yH$37Fv-5@m@Vy+4HQ<?_wo
zGw1(^GXYb~-p9%Ian@QwL;}P$mD~Kx4(o!RlSokMsjPc{L<WHHnIi#(PWvq3^yUNK
znSceJ;{qqUz$vwa%n_Q|-r7FN<%SgOv7%+*R|cV4LPSPC@t*tUs~Pxa`eKo7V7m=G
zw1kL+zx8k38Mi;o3=Hu?BdkVP_x=!egwA>0yO(x|q+p9@*a-Ya@X`_@G9%}!!gD#M
zQbQ2Cii8S~R)D#d5Rn*BM2PF9*B4E|1TPXQ3-##zA?(;4Iii!JZk{v)8D@{P8tKsc
zL)ZzbPIqfmQ6vK^yg(I5tH4-G-itg7b@{faOf3ZkR&xu;wqO&~5F#?KV6F>q<v-N~
z%<&?HNLGkkwS<T)d@}rQWo+aPT@YiDHe}d_OtplFEV^MUMCF=YH3S1JQja9{$XrW^
z$a2#i`M=g2S!x1OEV6}@ZQ&fXgoxyp{d?ff_0fIJfX5=iR>9W2KZL#Ze$oJ6wQj2v
z{P1oHK~acJP(z4FX>`QR**T|OWW9@2a|YGe#x#V8RM<Eer4+VQn1B&pq=d67!Jexj
zL}YhEUDU;rYt@>Q{f?7<#~BM6LPYB3{`}i))hDZEV2NjG=lZsDK3YOV>H`NXd$*uZ
ztqk<B$V*Q4lC#kgB67m_`u)$Brft)OK3Jq249mepONhuR*^Rzem)R_kf+ODa1~A_M
zE?PoF&WzHl4{~}k)DSp4%ONB=ghPae5Rs;*566BMT3aIp2RzFr;5T7=)es_b(R{0l
zoAt#`DcE2UHQ1`LQ)mbgY4g}xv(4t@DJj@tkz%ka#!*B=h{yw@u`0{$qhCuwiAA=8
z?N&UCh7gejabGSxub@#DNwl%`y`3HQvkx5gW4$gfOgEh_5HI0{IMX-=r6bdHB#@xP
z3#*aoY8ICx({dyb`{sq8km*kplK_ra*F<j;dq~*kc;OY2zCtD>8}LFVOvr>!2$pDr
zIZ>MgbKr$%Ncs#}P(e*(CxU&vvP<R^4r~?k!f?Cc&Lr^i!U($&&Lr^gLa<%1Gw}#s
z80kFHg9HSU6LcinOh9z!g%e2s1X2)p<nit37lVRa94|CMP!o(Np@SEmL(p@INhaYW
z+WbK${srm(g3Rc|8$f>pm=WtF`2}yO14%oO1KB$-yhHpuWKVG>m}Y`2!4hp2B4=aS
zvRnUk`9W_G3q_no5obw!Q7r7{9Cve0#KRR$YwkY>`7V4-rC?VI!-<LWLJI`7zzn*_
zP6+Cxm>f4qdh;9I;b|Cl8b**|*4liBZ}TaVK4ouHll+Lr!~pS`^`tyHUk!4tLBZs%
zcp(j3(;%2^+)kQ0C`bSIHpB}JplpDCWZ}NG3uZq|3Cu)d%rI{N1z|W0+gH;ba9;Ak
zqj7KB@%7Uap;D}Rn0}bFw_d?Nyf9yX{(D5SF}8nADkh2(Y9Q&&y<tvYuUcFC(K230
z0a*$-6a9PkBTHj(I@qOyFXdN&q5=ex26>?oTpJ;f?54*QU7J|k0j?b^eh#kBDIVmL
z8Tt3|s^{3kW-Di_Xi&&Q$~>eZx4;W~ka7=FQMccNhU`I;2p(+v^w+t%u}=-f!W)i%
z!%0b$77K4V{w<5&af)}GlJZl;iWD(}Q^kr@7N?1=(!|zOK3!~*E|ycgLTs@@te`kU
zY>~m@OtCUkY(x3kVm@2^UO!U#;QP{)J>;H9y@MU$C6c~mU2PS}R>6B$lxV9UVz{Ay
zl@8fBZPY>sUWo9G2%;w%5fTwWQIu~K!;1=uVtCQAXe%li?HwIRQH(5x;l;{g88k~a
zi$&hEm~rv4c!oDeHitoreHRDOn_e8UID(>N-(-fD9FolNGLS3-Sy9;x<eh;6Db7N&
zES8^*WZ5jvMY7!g$KJW@_<SVGXZdT9Y%PmRkZ%bJA~whiB`Bl>MNnLZe9KsV844+5
z`4vc3fvl)}C6ZOLxC+UtSiA?x_ON&_lI>;jJ|x>mF)4>}ym*78Z;+DsutZxF6N@3a
zNl(&S3=<2)_A@PFI=%U*2~iU%T0Cv>bb8Xs6Ot!VRD!0Jpy_m9WoSYfnn>`+<NYm8
zk1C$f6C#mg19BnRhg9EK;argA(x4;J*5E|pi3X$U6XNjQVc*~kUUEjnyC%gJPTje+
zbuebE5}U3Pn-j(q(^CUuezbJPwMT?Sgb&?DltmPS7F#U#q1#BdNM=w8vM515WD~ql
zhAhe`{_NnO|4esra>Q}04cW9IH=-+#D@@DpI46?ucwDzSR&xVL!ovZSlQ(kqq;MgD
z73Uqz<AM~EJbnaNR)IZ@v_+gp5$8iDCVRpId!8Kh77roULnxS3lRW|7HH9WDQheZ|
zqvt6?heeNw4kL5&!mJUq7?goVWS}s5Qn_eEF2zJc9Md$<7qW!}?YjlXv@Rs=`u`ei
zawdFAEp$uoqlO&fJl*;ur!HHY;ejP!R|3O{sqjKG1U1ttfEQjs&<l#mb>TzNyn*Qc
zYLRO#3L>QuFQkJi9X=)+kd?7dXkJEipr&y^1bBQ+UKJ2P?v2NfANMVsH+dKFxWMt>
z%?+dXeT*|Y##xX=jqQcAZ{zxr`zOvImuMeabm0sz%7FJ?%nLan&jF?cTMhEnEY1UY
z9*axBxCCtI(UpR6DcH~gs|<|Gz=k9qUZ?|E9axjw<b{*qauU2L{uR7_rJ0-;E`a<3
zYoE>F)y(o+z^jGDt>D$l;-?^g%8q{lUN2aFCwO(TxEs8>SzOK;m2*mR@L!J^`Mbvy
zRRp$~DSf9fPQJYFa>lLIsIQv3F&T_aSM!u5r=P0ynCCVByOlULod>7$Fp?g_6L5M$
z@v3lc#iP7;0XY4J_77#24fPLY5^iX4DC19Kq+`seuO;?RWE4#FPh=8cVsIkkcMGHo
z%%~R@qJD*F05N$UTf9#p8qDILLNtov+sODfvLqJI3wMz59Twk3#&=nK4;kNM@qJ``
zpT!T5@dFk=M8*$U{0JF8V)16s-wa0NVTK(C^tXT!2@*Z_W^$6&$-HoovpL8)QO7yN
z*&JeV4QErs;#$t8mc@rTo5L)w<810!+`_3^xS_-mdEp|by2#>7oaz#bFLSEPEN<mg
ztt`I6sjg5=-h9}!dRidFut@g<8&*4#wzCfsO%H?wnu4>&*;$@<s&H;ob22q&Pi(cY
z&Hd-$N4DR`j4sZyi*q7Ludp-bvos^?CY+P1z_SVlQJbxTF;y^`;sVaHfXz__oT`8e
zCaI<9%;@sJGMDwoI<|50HqMToQ3ogQ;Ot1-WP?lbx8b&ibFeDqoJ~3BPIM&E4{u)6
zwjtZ!^r|}JqQxp3tm6#j48~nDkuno8Zcu@g71U30)I~uRwEo}zQL9aGLl^GYeLzFf
zCc=tnnnOg9iphTlNe<5S#EL{KqX$xNoTZ$_pbVtUKm)1I<sxM+#XFZA)`>_^hbVX<
zMj7KsCy8;7@unzN8O!iuJ!AbSFJ2kX@Z#O$nQ?QJa~R$n&pFJvETqgrj#RNM<er7R
zDb7a9Y?hymJhPD><>w=1KFiNX?)mKawMe;^<*!AaYuWLYNLh&-X>h4T?v==!;wq%9
zV)<3by^7`UMasP_e=l<1%kuXj<vy0b54rDS`B#ws6(l3sbC=KHh+j6vtjG3H4%X%1
zN=gvyWt_a7T5R=#tT`KoMs-Mf9<uUAaG^fY1o9>pH-o&H#VsIjVR0+STUp!z@(vc`
zw#IW7zX16Q7I%WYlf~U2@1~geA~sx2%aZJ9|ATM#TdZua{9ym!fND>1@`Z<^aR;J{
zGwtHu7steV@zk1*C-I58(@j%%4S9C!gHw};gJ;k*M9B<Cd?UZm8Kr)JlkhmzI41us
zQY~V#S}{@;(}&X0@EnInXXiKIVtrirxbf6CA|oRwQ?w|2(RgaVNs&pDDJn+c#b`XW
zzETufiY60${D+9dJ|X%mJ{T>?8H**__AS{LE_OKA3!Ffn@3=q4*>`ZR<S2WvTL<Sw
zFlngaExka}7s!@o-E`p7!J6Vs7?ug+Db9g$IS@_pY6x5nV+odMo5mzdoVr(%vgxks
zf`M3g$MNsj#z%^nPoa$mv5+d}Q(2rQ=F?c5F6Pr&yh6;cU~z_+&tP$;n9pQ!mYC0?
zn7lvmJ44gMrbl@k=^tmCo9jV;Jt#=;vBxok8fh~VA1wrRP<$@XWohJi>!J@DDJgWZ
zEp{SlC-Np|$P3vZ%VwXOnie`a)8_R*-*q<`vIb{0m5a)YC{8--*os~w>1+0Vt?93m
z@uW?wgZr{b`kd{erUIYJmgt$_mI*_tVtFtw4`vc9(RStEPaJUWqt2a{=!09ou{FQ6
z{jv@2+2Tiqq8^2jIQfs|j{f<#jRc(au7me=P|@($2HtHfPT{;$IEJ6fd8bl*%S)$2
z&&XmGK8!}rppi4BB|<xA(9W3>Pq=-yqJGEHkH`0v<$hC`h_KvmITI1o$WKip!kzv#
zyKpU_#<Two%D-VCRr4(<-%|WQe`g*H{wW{N9yug3ga(VGAxR-LJd~m#r6`1i1M<fW
zJbO2icC#0$`4a~@??>I!Uw*tX)4wM$DaP5;*&}Tuohe$RSY%JvO0rFIrl=Sxijh58
z88^9+Z7Fgl_;H5e=V^Y+{jmCFU{?md^fA^7L9H;6?z<a;x+!iitvNd2r@l@fw1n>N
z?~e|+Qq+Mhp#kYPpuWTs-iRytJthY>Va6^H?}9!e?Y{+a|A`%mzenygydi(Qd@>@L
zh_pOJ<akKZ6_Kz>97jZ4X5oU*7v}ZM>#m&~6A${?|NSQjX61bqDVJ6qGW7Nt=<YF~
hzo%C}H<9Oi0nq^Z=|(?&=x0Cr=|w-i>E}TD`9CN)wetV~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_c_shuffle.hpp.D2F306B1868E426C.idx b/.cache/clangd/index/device_gemm_xdl_splitk_c_shuffle.hpp.D2F306B1868E426C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..63246b85bcc45fef174f284d265ef684639190fc
GIT binary patch
literal 6604
zcmb_gc|4U_AHU~3cZ%z|*S*$jxmh1XqRc3AwTL28k<fT0nbfqrX<yV-3NKL-U9M)D
zX|uFgi(D#=lq@Y&WNTAm#8jFxde3#8o@f4g|LFCpbI$kq{q8xx^IOk33*6n@uIsQ^
zF3a4vtPa}{!e+5p0`!0VhKLn92w14GSX-nrztwA-95rgaJ1X8EXqk5RL3rBrYTx6~
zWlqE8wTII6WA>>pE~;qq`qcYiE9hp!!0lEGS4SS}l!i`q-g)KYV8`aR!#3oC<Y~)o
z-PFWq8h&2WS+RDxD#^XlLYNSj>OJ-y2SF?AHDvXkS{+trqqWK!9m6~H&Q>jg$|w5G
zwcZ>6v()Yt7`CjW5|ciP&sMDSYrDMHv`;5WC>YG1=GZ*XeO>9ld$vwK(C%<B!{~TN
z(~a_5E(1Y?)cPmwD78}CE5*wH(ru5dn|a=|AlU7^Udzt-)6|B0FAMjssmU=A%zs%P
z?!%p4cRQ!z<68~A9qksHbJ|rWjMg|X+syRM*b`PdaqVT=xA&~PoUq=n>!<9OoU%Tp
zEt1YL)qie1Q^|{}uR2x59WseZ>Yp&z<9*naW>vxT?=#9ht=69ljHsN^7_sSe`TJuQ
zl;$qo*y$RPktzB$9{zgoIs!^+Hy>GB=dt=_*ZYL$qWXrO^w73**{=)oRbO?LSUcR<
zxY5QVD0*X1C$*zEdvgENRmG;ZpLpa!-y>Pm+)G#TJELPu7U`ttwA{&I?afeT|Jjw;
zSJ09k8SHbaX=~?SdXJr-*)(>ZirqZL>vQAM--rn{h3~>wT-_#fTXsK2e?z#eQhD$f
z>&;WF_do8@FIZ|<dasSFQ;i;|TpTx$zOcdd+aK-sq+M-%==eta)V6;m4Rt5}a8$MR
z@r!jT;;Y%*f9ri}|MQPc3Bps(KFb$B-|9Z={>uCtP5fh-jbrDm3cZ-wP;hlk=Lujx
zmizi%+pBC(@7EiPT3*$xxHeJwWQesbvBzf%dwfK2ePc^)@`4x96K3jwvfvekr$VQm
zw+r9DKR5>jhZXAu1m4(PCJYSq|24kqiLmx1unH6S+~M+!E?T@Ri*&nC+LpDlW%q5L
z$eZg*^8T2d5RqS9+<0QS*T=@|xn}3SbDy+<`#4Q$YG{0f@|`tP)celF+uoYA$2}xk
z<$6P7gUH{mV83X7@HmUP7o%3M>-lW$RF|~FW<Z5E=U_<poC5gcaeJ$?y#XtavpnK>
z7Wv2Cx`vOt|IB0dz7p4xckAlnI6+Rm?8nA!uOlu7mY(xCdDJq`ccAA{k?0<~+y56O
z<<eZ8zm=V)jWxD`DpPzP%=~7XG5<`CM$CbG2bT#Rb_9lTbe5@Q^K3Rf5-*YF?OIyd
zTCziVG&^jCuzdbrk84UZt0q*t=0scUy5TpSoZ0fCdDG*XXxY+$jqc5t12YB6Go1}z
zZw?RKc&}VZ=CVj?yHq;hVB8?tYs-J0za=`-S2(t8!_d!v``t>83of?nP4JW`+4IWV
zYxUl>md$bX2+wlUb}hVf(zD5_Xi>-q_bl&Cpx>u0*Wu*dT6Wumk1xrF#Iv?JV@f}c
zKV152)EUdc(R+lMPgv`>u3nl*uvjWqhRwcqyH2P;wmyqm$XTd0+_02T7D0d9=+5!o
zFR~x;ppJ+|<$^J}z;dL+EWE&?9}f>$#%MwnG{aoox%$HmCL25u&Cok&TFHmnV_1|7
z@MJ(R(qR^}bk*QV>vFgApe|Y@4~)q}izqtG;$Rdy{`Ki8C48uc7U_b<UGTe+4zoCE
z_Scu^#)N7@60vyedg~82m@F;{It>51^NbHw5&0@qz6$k6I?S@*@4SBJi*q|QAqTN|
z>3Zo8H<;|p->_f%!uN+&;3z~c1e`)(Fw$X`t)`(N*P6INe5i#sdl_&q1GSM3vqXU)
zzf17lpc+(1vvfj!C)6A1FiUFjMg6!N>Q6>N0<pvbl~|xY(qR^9WXMoHuVB?E$U`ib
zfXXGbkfOsZd3rmww^ZlGs6jqrxd3%9AeAXP%#!~oqQoHN2BiX(5sN>^Uu(F*WLdb@
zi?@p+8#JLNBKwGZCJr~4eC5fu_4Y46%ozno>#?XAqh}ZoH!Nickmmpj{QiVGblz%q
zI1dVuT`zz!7m!^Q9iF4}O3<s~$%{!fC_qb8LVhK(ilW0TuWs^K+s~Y|<3kO^5>FV#
z6LupVX6fnjG&z@-c3uU3gIN4Hep<r~Cco`&aTplNYf&3EeG%jrK@&>RVU~fL>c8JH
z*OX{M0*M+2l;Qw?q{A#9TpreX^7Df=Ar~z&Pj{aFaPx)yTiMFeinY#ssD#K#1V4!o
zjdYmh@BQyK{FdaLq6mdd$jgKRO3`5!Rz%nj>v{4LZe8sF{bb_-Nj(H+55c!MjyUGi
zzdyJx`ZXf7LbFzAk5z~x8FGdr7+Z&=qJUBq(D-8OH#IuHF4cC0B$YyNQwaWt5y44H
zLU1KC=Gw-jtvoWJnN3n&23`~n4w9N{FqgtXPg33n-V_c!lA32QkHU(=xdW|~4S3l=
z7b}@Fl0mSab9+3O6q#_Ape?lnt9D?6_lj2moZ*y$w<LU!t2B7};zTx^Y64D8K#ck2
z*tH3`(KrWc=RhG=Je#V3CKb>W=K{S?&Q{wZUTdL+(};jHVlmc*0dKcMtkUfcM5qFK
zRlowzh<hZFB9-AlG%AS*@BYbYKVYL@XjK6n%xOH?@XWX+9wBHd85GH&CH6++sRhOT
zlj}KOy$hT^k)Cq^FNdjTMH0n&wq7<kMFb<SK{nM5bi08*mXS@xLfu&S1%Ep)J#L8W
z6-$sP?O4SEJ3IqPNrAo;m|{zk)Kf6`De%J+eHZYZw%x1+gJ`qmKvWLQF#+dqq&G$K
z#<s++g9!4JjXz@xq0fp`jR{kA3+m5sVqB3|_5xuqaKbB))F2QJGI2Z<#zO{O0)-M9
z&nfSm8eMa29uoB-(0B+aJdv08B<(8Kqw|qE`+$ERSd0lIl>q${;1`S|0?8<^6<9(a
zWmnniFn$73UJ{|4M5tn?k_R06XGLQPVJ;!8F~;>}6lbKW;2BA(1{l=<ds>iwAnpeX
zX#5$7KQnP66emIkT?)lg8vBJD3CKC#?1;AhGok)7p@k&~y6CX#doxoLBzh(kW<paO
zI3%)y$yI2B%|aq8YIH$EY}O@DVwR6;Gi*gx5nG5S(Sl0|frK!`^UG7rYzeW5#%aXt
zG#ZDDn_Z{jyWJNF)&|Ym=-jq+qWfj_kg`jNkV@#M5+-<ta6O+|+3cNb5Mjo+8BTcN
zb!?U-)Amz6+DHl^NFl~xe);dYse~>QPfR7IV!SCZ$@BWYrO{}<7HHN&OMojo%C3sC
zjsqX52Jo)}O>B%k@S*c3M>9<n)~N)ER>M=v@q|=jB8?@)ETj(_rxCM|J}};!I$3lZ
z?A(R?a0vJhfe^d`Nu@*obod4169w`^#c={_A<lz{>Ph@OS&yWAG=21OIY)jB1PMS$
z2XP`0Br@?qAUMdxhk)P^6DI*d5)&T=g2POl3<Swcd;|!NFmVbHq%d(R5Tr8kQ6M<V
z#I=B|1?u$vZ$i;cXo?FTN!@~?TTEOHMb%7v8;Wi-aSasJFmWvu)iUv2D7wqUJ<z@f
zim|hk)F06P4<_z~_PtE}2HL-2;y!5K$He{6zMqNTL&x`UI-dW)Q>nN0u7EIPc(H?c
z3bv*^cmgDJjY_hsZfMSb+KXJ(U&ntYuJ6e;f|zLQ&TRDg669ZkhPd7w>9D*qt?Wb>
zA|w;SWWp3X+L3`JL3>p78&L4dz<3#Oq!ld#o-*K1<18r5VrGRbXrBdV;4G1HbN1<b
z`*)5*f`5P-AD{u=5%M@fBY`l$M>$+hlsO~ET6&L)qZei%QLyP;21Ypg<Pr*Hz>t<G
z2U_JoJ1hZ-j&+tra3bC_i3<Lt3OJR{Tlb;reW*>l&If4t0a{=|aZs~4BZp~Puuf>+
z3CH7DlkdndfiTBAlH;)nge}I$X0$ntnLmg74V#h@q?F)cWy`U(l(1o97b)RN;|zk#
zVDe`WR0d&==_g#8MKi8_O#h!cK5OSV@Ar%)sfh6s(7y!C@!m*uc-DxgjgebN&62)N
zzk~du9h$YnNmx1f0)g_1;@W_ffn80LGkjgZpXG-IQ7jls8&3*&Qu;VTqOxXH2qxh=
zM^cr5UrC=rNa`-Iz00iNPr<aOz>_w1FA(<vcNz}@@gNh&LvcK0&?QhTVPY9nl|en)
zIa;8wg&F6qP}s`EZBW?8#O+Yn&cwS3%iV-Mb{>-2Ls;%%;=P3BUM7wrETfoMN~lT+
zJ#5;-*V_uU|I?rK)d#{N`@Eyu?{v>bXVGLrJ(;fWB|N42O8;kG|97t}t!fB3JLS*k
zUoASWQK-zy@0+x1g(-!Ju-{$bIp#mMJ-G~dY#k8P0S)XCm>qen{5tAxx+(5ZiObHb
zAC%r%Oq@`J&b?A%k(BVq8fT-(nwmjO!?@ziyfs`mrKtZIT-V?Ku<(4um0yqq&KAxt
zcwKzqMZeDiye!)FQ3MU&esO42M`X_b^mO(5|GOdToXmDC@YYZHYC}BIIdnJUf@&4o
zVKZ=P23}alkCKl+z3}MxPNd8f!a9X;#>!)(`eY*^+%WE#T<0>7BLrV9$}{1~aR*<h
zjz?Pk=k<`JI)GINeO-KR`_0?Ys@idgu?s49L3KJZ;|b+>LLF!4&T++O?L^tXqRq7e
zjaFcZH-}Rnav;Sc1hy>xB_b$ZATR+xJs^jbzdPX1B()DH?W1qq{nFxcn+ry!U#$$B
zypSM@7Yw{2KFuM*KWiwyjw6En(E{^hGojCl#|ljND>HuJ(AyX4NI8YTp%Bc)`TEnR
zQ{^^~Q{JL;LLU(K(YHSN0aZN2#0gNG0KeP{<p)-AI*oC)MoTMRKX7Rpk~Kf^weRsm
zOQ#akQwa|&WJK7ikeTRF1Yb;9IFM%Ms65f=-L(kdIq`*ng|~z2APX<DJ|c8|BASb3
zBySXd7I?s78Ph{kdT34$4e6m7J+!2UCiKvN9*+Gogp~J=|7h;2s?}B*YiVI*Y-VC=
IZeYmzFK~pgV*mgE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp.C1DEA5CF04C380FB.idx b/.cache/clangd/index/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp.C1DEA5CF04C380FB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a3fc917b14598937a0cf1c34c6a332944cd73cfb
GIT binary patch
literal 6736
zcmbtYc|28H`#)=~lMEZ@7;+rPw2hIWlrq&DB}C&*>0VtbrEY_0&}2NBNg2APl0vR@
zMaWHxQX&mE)fKlvm*#GRDf-pB4pw{r{(Aq&hrRaK_u0>S*0Z1IS<l*wT%Da=gdC33
zYS-Y6ep`LH91e$%|AMv#u6vC*oGUyICpi1OXZt3hol4I9Cr6(pZWs9fx%G!7vd7~V
zgyku9^|S=OmS6Qm{l`ipkL5X`GY)I~MJ1oeKai*Y@Rv8mRchQL4?oAP{HU*cY5Z36
z-%IPN`g(q@`q;fRRE~1+DIMdi_3oj}?3l4RHn+GlVmng1gN&kAeR;e(;JVtpWDC)v
z#_Ng&Pl~zw<dwZ#2cL`WUE+4li@J2uxHf9Cw~G!H)K@;sHnAwLXH{9!s`Q1yvkI)5
z@4(s&(`6C5*XkcHnG)vzm>>O9a&N@{oviEEnwAr3BBL7IUHGmpNU`R{r{`m}6HQ*t
zHf`@Ob{(^KL)ZG;kemI3^QvcmZBXp3oK^eG{Pl)AeqN7D>N8eE{NbVEW7}D6ookp}
z;+}WgL(yZQ_omQ*#%)Jq{<gX`Se%z@Qa0n-xUUK^{Fib0sza{@Sw2Z=8}mLCiEcdI
ztsC5uy<o5Fi$imcy*m?;)1X;<qPeZ0Io|2i&>ua&7d7;s6ZIsW?%!$?^P73}Iqw@L
zXD`G~S+4&n@S5+SyuQz>ORZ6H)BO2shhmcg6KXzeH?%gJd1j}MzvtKY_d)_HTXvuH
znrzxv?)BkjU16otcDs%Its9OxSkD>!>6iRfn`+YiTFYLi*-X7+HLzmitL;fAmREf|
zdH7Rm;=xbdEvj$09h)D%yqDy#eon=XyGd=6YzNO=+8+Ms*73-BU+j8QwC)`}W*_Q(
zE(bk$YJNF>$+O%$r>85Jwk&-3aQcInf4;FVHr?npcVS2GS3R>Wf#;Xa?)ZmJs}6S?
z+wBh5s&|C{*$@<T<L?`@q^osgZe>S4I6Qsqq4jBMH>aMBFt|}YzWMq7O&K{<`n|!6
z(khCrZ+jdA*KEox(y_a5TCgEpIecA!e8GuTzADY0brmbmED+aQ3Hl?ps3Z*41@jWh
z$DWo_P3WZ$Eiy)17uPR)<W-S7duhU<%!yZ<`qmi<-QNqXXZ2gEP54FHBjDLrsl)AM
z?(fWQ!i3yK6ZBUH{HqnSqx9SwojcnH^A1!@E9kQC98e!<SY_vTd<|Wg;5Gkt)K3o*
zV=h^`&yw}co7j-)AYSyDvo9d(Q_31!{l1stEB)5Vy-i`!6ZGs0c?T<cTcx_9+V;%%
zT_ShTt#L<eYN5i<bumxv^{&M#wW9po*p+317pGiLEOp<f^tf-+{H!hIW>4I@dkv=S
zs3?<*+<Myl)l!$`YlAM#ta0>AcYE7A<g)W`E86A$e(~Z&|4(kbs+<3)zWi<Cijdty
zS)V*T&o=+OZ`WpLo#UGd-tE}6`C*AcK3!g%_~OQTxpj_dVpGS5!lAZPbv22DN0sWu
z)t>FkzV<}-8J!W7NBCUGJ|4<*ywCqK);P-eNzl%XD@}0Dk6BmuR*yUV!8nM79D1Ad
zHs0`sBaQ0<{*epyy6k5$=p_fos&i-$S&wnU7eXtnorD59t}o<4B~150-9BhFIuH@z
zV#^A(ZC7-ZpaK@jgu+ZXWpp4S(~>Ley>If$6`&fnAsc9A1C!C=o5-;zTY3b0LlodR
zERqGZvara=KtyI-IDa)gV8st|kc%fe0%VT>?a_gVc)IQu`qBULm7pRPDTTaJC>kAz
z$kL;~cWzeRY%T@I2spIESO?4D3r7Y5><&N!?+aqT51DliJSc^Ae1^QwP;+!3I{e#C
z#U;vti%O7>MQ%ghZKyFi5Rrh66&=r4{9MU{DjFO*1uCRK{^&qN_O82Vk?Pu#rUd1&
zmR!iog?gg{5s5ApJ}5cX70QFkSW7;T&j(7Q0};vEuXW&1k?3z#D1)`Q3SC9R7eW^W
z-@I(I+eUlDjzXYN2>A5KKt#$T{X;56MaLDODt3J)(8|QF9~p>9t@F~t2DjDkRiQMV
z#6{>L8om(vQnryucv#0x70O|Hu5hks_(JHmXJdkUwx>8tK?>7(KrRojVq_p9o#Woc
zpW3?djTA&!BuFkudH6!;5B_FrYxkJ0;|;s27|Imm_c}5Vk-_QeH@YWS{iq6Mu?9Dx
zn`rn#=r4~7&N>-)?&U!qrduKYoiaKQkwCvSK?X`V2+W#opyzpWK+*NUupUe%J|w9G
zWQS9T*i2GLMRquKNPtmvkY><$B1F+qKp_eUi4F=!uIUjll>{|KmjYfX5Rthkx&~O*
z0DA_vfd8}rcLuis$2KsJ!QDX64fGk@2ORs@d?XY^LKa^L6$_yT2||j#3I$i8KFJ`8
zZiIqHHogW0*VwoT3YysX6*PGTt%;v0`ZY9p&Bkw_$s0C)3r*fKSngemdfrO4*^r{c
zkbD^8ku;>}NF*Q0V0p`!%=Kv#+qo3&qU~ZpB9o$BwOtKJGElUewwnP735uSpJ=cH)
zsr*{u)OE$8i-=37BPty!keRr22BI?9I15o(43>Z6?o+8J>x1K>379qk3*u{1Byn7f
zR7es6g8QeXKNnU_;&SQB!2U9DAWul}?90HJ!Rb&f9STU*;L;^fuLSCoxF<Ch&oEM8
z$qX7p(@l2;z8Z&Z5+ge?av+P*RtUF=mA?K26J<c73>cHCbpHtSe;sBY1-O_PR|eDx
zLC-i-{bAZI+3z;iZ1rb5O}Muna_M_Oc#nx`F8vV*KQefV*Az9YIg37H8;gKu5ilaA
zko^J&Lpo4MXR=0ibk`tRHEnd)z=CN&A&o6KvUd<cTdns^zexGMSLM>th>B)jDwmEy
zR16!(BK26LLHx+2<B)nB8^<H{cs5Qz!UQB@<P(unB2r;+5>iV->I~kG)b_J+GSWy!
znhbvcX&hkVgUIY4vS9cWWR}9lsc3vEGG+KQG(L@u4<YIhdwstm>Q^?-Koc^MIrDrb
zqB7Zh7Mh%etQkHVQP~Vn>X(%4=vGW3lcKADZWXX5@06lDfMW+(%HS73@B)}I_!Dsa
z#O9-*APRot=U#X=Bf91A0vs&$K&2iSkOw{HPgQH2{p&&;1MR@O9rzG}q6dKY0D~9H
z%s#)-kC%ifyq(K@hKgiOl9FQ(#`sK9a*QI88R1Dy#Mw8}Qjj%z`lflEJlbxHZOTQ;
zxo8}Df(YOzBprsU{`U9-e>jDBSJ`a%$xvSpTaW?;DNvu}JB2qhy<%uV@}0t4lS(Vp
zA=Q-V#rj4X5;B2LXUtluM>-G3P_(VF?Nst_$qM+9NQXQpsSfgDCQ2pPM2u|M{2Vdz
zBzWZ{*K^9gwP*3d!jUK(=@G(TWATmiE+LyR;b80FM&=_85w>rn8zKZ*ASOn7AkqtL
z;c_yHEQfAkSA-#c7}6p$NM@7|M?yB96pm&PydC}%q<iG#WNd+lx`z|leD^H2nsK+(
zW(xjyI#NtWCPaSEk-N$o>6-^}c-Y$5+7ZJf9#|$u%ZV7-2eF|eofr{}fPYONvNcp4
z!^MVC*f#Qn1Wyb{lNc;U6m~NkkH>CiV`nj%LvZAwX`0tTNGLY!Ik0-p-s<;3t9|g>
z)*lsC;l#NkD}$HT1Pz<uR1%M*JA|^MJt1)}4ky4X2C5_z;@yz=&Hk^`gcKdj4;C;z
z$WH!F7TU$%#X`IJyIE)te-8`o<?m&o5Pk>?h4MpLC=Bq!01Jl$emEOP0Dc4;<B&TG
zl!@UKT@E$N*<HpNsCkBsE1+ft8&^WjN;a;7npJFk7HXbl;~J=0!^Zca^?m3-@{Xea
zg4TbraT~O5W8(+V`T-laL+f@n?ts=EZ2TD7K87=i{=}!*^UXrN{P3$4BP}tLu&FKl
zm}s+S#hA#0ygaBw;xp}u$;JQF9)F36FHrCW>XUFv>-X`FP!!$8eNiDWF9f!X4-0{N
zAy~%XR47Pg*Pc{poeCXEs-|CaDZCREG6BbBCsgT#+Qe8~Poc^H)F$l?X|{06NRAyO
zVUASEnKfV8I2BuX6&hZJ)+AhyJGE;%)(>X=PsIhXHCOHjM|Wcl*P!7wMgv&_o^j*|
zLju0U%636Rj?W>iIuD5QfDs`m+>xpDF*cuakFvPmr2hh|ZiI%7jB0WS!WvS6LMpqH
z7&!`&PNL$~t_rod55IkPn^b9C_UrE3uU&8>_5~_`f!{Wdn#Q|cC3YXXzIz7J8)J7#
zj%?%wT~x3w@372YjeSrJ_|-s#xV3ITy4QHp#7f+_Wgwjlq)TFpOJ^a&EM!D5X{E7Z
z$zhKuzI0;2CV6LR@&A0mm;2T%`91JrF1EnV*v^SOfAvLXSk80n+X`Gd7E!TGlObus
zs5mx`M^ro;Cm<?;jS~@-$i{fnPh#W!h}zG_$%sm3;{%8~z{Up=b&!oy5S7BlsfbEt
z<1|F2v2hNfau|GLX+~6WLd8X#t8GBg24<2sK+#>mxC<;~a32u#vH1vS909-aH|-7g
zESfJ$|L#y}jWNHgutm!XZ^uEVK~^N#C8t(V6wqgOj21Aj1$Yws?mC|MH1~Z`7k-h2
zP_q!4kTTV#e&ElXQpHA0Gytat;6a{fOFjCuwC!jJj<GOg7KWyinBwA=#Uc_p6HGFg
zD?9qBfJ6y}TMVPy(46!flCq>T0F6jllHl<J(30Rsp7G=Lc#gNQDMx_K5ipOmtWT!h
zbm~7T0N?FA;w>53;g1%Kp^~p{grx8ut)35b$)+tS(}F^vLmY$q7HFCdtw<ThEhRL`
zg_DR!DSRZHS^{S<En+QHtc7Z%x=O0GPA4=b)%qVn{)FUiMRS~t|Npc~(apfLnK`|7
zS&r!$tEd)_1zVv^D=tN&pGqXPQ>G6p6GuLqP+elBS)7L**$7k`feA64n1sv0$Tu0{
zfe+c$7c5G3^}pN0<Rp)0kbK!8`iF)JE?9J=_uxEg2afIFM`Dqrr4l>`6B#@J90%C^
zJ}B4+zwuwrw|D+v{l*h7Ega1ZM{Xnvi7)WPBj1e3)SN)S4ZgGR$D&y$<vLaV4P6Sz
ziz6*EhlrBRi9<q~BpZjsMo{3Yp#4}EM@w=U`Ozmg9F8_~tIOPKGPiont<kqzNN+pQ
j@yd6fi=LdJk%5-Bo{p}*NRuP_Ccx2QZVj1R1LpR>!ZOdP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp.5C4328570228C900.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp.5C4328570228C900.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68b33656fc29087166b6b8a81007c40cd1bbb61d
GIT binary patch
literal 4672
zcmZvfdsq`!8pbmdh)JL*!JvQ$sLcI{kO*>91c@j@C?Fe=fD|p4a#0Wwwc4%PsDSla
zi&BxQEVWP(sm86gt5z(OqN2qH7Ajg1RJ`F@(RRB-o@dPW$L4u*GQab_!+UbhnJ<Uf
zc^b_Og-8^(bl&QmWh+yyL?V$r|1MaWpV=!Gi4tu@qSY1W5^}V6W{hd8Jfr>c3wn9*
zDTUjitKIh<Yjg?g8k5p@=~g_f7;JpL{?D?Z(5m|OsTDgqZydaQu()q}lt>{uu&i_2
z*iE}HWgZM(H*<_EsrA2o`);P+`*rH@@VX$ALq$;6m3?bR_5FDGV(|IjR%@RmUAp~h
zs{AXjsaH0vtGe{@<9+Lj*Q)NN|9ENOQ>%&%dT#hxm#O=Yz8Bjn^*eu?e(OPY2WFG-
zi+oXUICD1fWM}Tx6{(wg*4!;`T$poCm-(A#z{dS`&PU^_OK(JN%QOD<;y_|u#qzGt
zK00vN>e1$o=SQw9_lL0RFIqxMlv!@4+ZK)6aI^nZRSfa*0mCh-`1kU{y<7I}c^eD5
zU)s@hVcV(}t3w4DXB(;=e$-aZtPGjbk=(gEr0k1nvzsS2e6(QZne5CG`em8C;e)ZN
zYyHVl+rAE?@$>Y2R3Tk?<@ZB5lr$!AvFS;To!geQ7-jogZb_3j7Owx`hrnmmv9H%Y
zsk$H4lxXJ_(N~S7)|wJdJ-s^6?UCrTY5t6s%h4~r0&L#v-4n5QN_=~Q@$A-te%Zp7
zwep~)@<O?H$NY@&albFm_x`26?!-QqXQiPE*?ET(b!p>6?l#4S1TI4DD^I_e=u`FF
zeR==2x_P_jZ{0J=Iad^V$9A|pG<?7#%lq)16^2(A0$OG**zTd*Qr9`Z`B91S+$T4h
z_B7?%W^Ynm`)2!#4~lns)b8iPMwN~8w+p$mHG@k}ZJ&BOOkMezRQL0KrrIGlW>3A{
ziN){iSo7UGtr_>8R!Yi_C>oCa-2P2S_=Ci=p4yUI@sH-{=P1iO(mQ|Ef7`z5am){s
z5+^B7*gOu@$YOf=rLKKHxm{I#tg`%w=L_fB#=V}i#cOr4MX`kn|2jq)Jz7yOyHVxN
z9@*uS?lL#y*uUxb&X$BN$#1QhyP&P<^5Kp4W$`V)X0~>v$bzQi9Z0{J{WQAam!Xao
zi_>lzlAPSEtasckyqI6hJbJG#bhztO&(h6DjR}fkc1=W~SD$J1vzKMns}m}pG~7Pg
z`*U<xL`UI;nirp*Z1`~2+_{Ivr(!=+9{a?nT)$Ou?*{YV^75YV4=dz<z0NGn_Bk1_
z^?0*?p~E@X?{<7u@@@JjH@+^~d@?F|%IQ_R*J*2x4tu%0f2mr#)c><bmx}`iS4RKi
z;+X+;ab<bwZNFs?%kn~r$XBKEmg}#TgKN7S<L-6G#8F+jxxJ(7kJVh?kPx`*^@o~6
zO`~sJOH0hoap+0Qdz=^17`SD9&vk9qu;;4{^~3!;eXiz^k>~U20nf~@FN^Mk?Vs-3
zc;}UO#D)n6uN&-%%p`{gn<O#M+-AFu-zzF8&Uwcd6Nw~l#_eHMe(pB-LY*i@DN)MI
z4POh)FZruOJsTsZ$JyZVyzC-zu{Sq@Y`6FN_qxiI;I~9^yzJ`e>S1mK+5Y%e=fbb<
zU$ntvcv&G)$jpr(JOAOe_hF(pYb*L2FMC^iYs`%x&wBlIu%@s0xDB4b%T5xfQRYUF
z>vzdcOI(hZ+u*UhY%j5wnj1lGS)y@E+q`qY24BR>?h^Mfb0f(2Q{DfG)PpJ;T+7Rj
z5=WW25#&F|Bs6sAecxw;zs<`I5(k;N5#;=3?-oq9walJ_+&pV#7hWH#?G1FdO(C>I
ztS|zebi`r{ky;XHq>cm{rKNyI=_sJlS{i7yjs_Z|Wq`)$7@)CQ7HF)F1sbR2fX3-K
zpdl>+8qy)4sV1sC-Yj2=YO;zDNZ2tG<mR0MrbJI<SSEqUA_@T&5g%d}B=sb)l366M
zl0_u6QhEwlsVoXusUiwmX*~_BbQTS)bP)}$jGh5jCW`@9rig)7R?h+}o5cbvTf{;u
zr{{o`%i@5QE8?IP=@GD^ECj5m2tljLPZcUmzHN$Xj!FyYZB>#$!m}!qo1amzFb0g+
z1!v!Yk&r0^MnR?x7!8>*U<_o|fU%G{1I9r{1`JtPs1f{r87UB8SOh^^wgiEQX@Jy(
zpM^+*1Wpo>z)7k};G_r&I4MK{C#9x<lO|~3q!A6Aw3-G^hG2k`K@4y*Y6du2f(1?%
zvB1fyS>WUd4mdf)0Vk*CfD;i2I1xg?iPQ)<Rg{Vq_5#08kZKwrAJue$gilOvuExTG
z2-VnMZmz;YH(%UZG2$n%eTZ(1=z%<l5sx4napDYaSv{f|Ct4tzaH11(o0zyFw)n|b
zjD&e)H%7udau6e79%;l$m`66_B+MgCI0^H}HZckFs8)=Ec~mz>!8~dZqhKCo#3`6Z
zHRBY_qf9sj^Qbm41@q`ujD~r1H%7xedJv;w9&N;Fm`69`G|Zz-I1TgYHZcwJm{yE|
zc}zFPz&vIUV_+U*#2J{!G~*1+V@x;$^O!a<1M}EcjD>k@H^#y|b`WD>9&5x|n8!Bb
zEX-p~I1BUGHZcqHxK@mVd0aQf!8~pd<6s_V#5tJ9HRBx2<4iaQ^SCxK2lG%XhF~7*
z#t_UygBXH&$cQ7DhnjH&^N<NgFb}nfktJ`o;h|{w`N~erD#dTI-xA^Bjq=--=ZH#o
zV74hPGhD)iALaOH1=+td$eQo=k@<`k+!-!$ZRaA^=knbuscO8?y()4}E35O3;k#W!
zUE_rxkiWlg+;sj{@0pSA1XsP#ZFmws!T(CBGvBQkqi_`J`j-zj7BA|p*v@ymN4m!d
z&vEVQZ>(c-KG*QwPIgX?LUj=ib#l4+FbAs4keK)ntO7$~5;8R;rXbTpVj40tBxWG9
zLt+*(HzejDqaiVZtV&kBncPNFws6N!RRjFk?Z?=U6~_FnI(Dgb_~#G!Zl$LZ2`B4+
z-Ck4R+y63B%9lzhF<F=Z+Z0kt0-~f8AX-WTVx$ZpR>}h6q#Ph5MQ`Xr{LFzAx0{#w
z@%UQ#P7+pC7yvFeABZhe4EN97Y1Yq2jUHFR-|M6FaTD$pPR;!Cj2xLQ_YxXnp74kH
zZ>AvuQ5p&mt)T%i8U_%nVF7U(4iM5HKq@EI1YuAQq_rR80rew$-O<x=nlMJ;2+yB)
zM$YeGtiePG7;7*Ihzh0v(ZMtzCYS-l2D5;;U=9!pMsG-zzbrd-<o||({Ir6&k&C2}
z#{3ee9Q!j?Qe-qmOO_<+7bhi)77LO{q_<oqS}v0<mrE>{X*o-?GkA=AH%e?}Eg2=1
z+1S1{X0+Wr`|vQrUOq2WG0tJ^c&8cufkDADXN62vvH_EQDYyUE>*zAU)!k##6i<?7
dIIn5ms_A|c)y~Lgj?!`N{8U^MJv#zh@IUjD1)=}|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp.8E11BCA1E8B40C54.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp.8E11BCA1E8B40C54.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b78f114d23858bc5848e6955172dc21d90fe8773
GIT binary patch
literal 4672
zcmZvf2~<;88prcqASQvT1cM?Xpz`*sLL$hf2og~QQa~mm0V!IRvM7j%TD8^U0$A5t
zl!{c#Pzx22YHYO~wPK+(DpG7<Kyg9P;*xr-=(L?j&Uwu@=S<Gw@_zUK5C6+|b1xq;
z^EH~83Xv#u+59!x%U7k?h(sa>{wse~UPiB2BucOoiPlt{kI#<2J9D(L@@(`MpVKRX
zPAlAwT<d=5RHKVu-<X)TN4N6ViowR`8~!N!Eu^ZxZED4?&i2Ds4j1*!h!iPAhn9El
z7_(*X<&48Y>t~IYCAR+Aci?8)gIiOFhSmp~94i95t{zx7qVLC}mx3<bTNC{>@$#Kl
zQ{`Vyn|gKQ`l`zxJ~^<yXr1ao+K-n9KDMdYsON^vU8e5;^u5?oso#BX#_dO09hhDG
zFY?8`Va&ONQ=K{2R;Fy}S$n^{aZ&bpUB*36|IG*M#(x%DUD_VGBiHcPi$e)@6)U<v
zDL!=6=J5v|&yU?u9t>vHpSJ{;C^Oy87#EM-c(ea>m6rJM(AL{j(eLF2`?np~_a+wj
zu(YG;;*QlVHb?T)&oxv#{uo_3t1@^>M^fkB;IhxB&uO06P`q%~*{qBb`em8CVbd7Z
z_5P&D9bbjg_yu|as*tX_`umY=N~#T5VtQIL(rsI+R@wH3TjJ!+1sgW~5MZv3dA;sw
z)x*f9gpt$2`>L^&T2uV#XV(V0JrbO^ESTAHCF<ogf4g^j_l57D64w@QIJbSEU$&@a
zojh=vyg=^Nu^>Hc?C&e`ynd;#J9)sxTpFT~U2r^EmpU%^ep5_vz+%+4>dcFY-c`@t
zSM*=6o4<F#_I;Da=ZHe?+7Fe7gbjFPdL6yHa_g&${w=c??)1=YtLt3Q{J6w${-gG$
zeN8#`SzA=szuEa>Q_*gZ+Jju^h_bPMBZKd5Pv_E7+NRzKRabr@)%|>ssdmiK?yDbp
za>-k})_(U^Yx;v{m6Eb!iiYDqw|x^F_9)?;XLQN!xW{w#bCqQtX`Q$9-?puOqWxh~
z!X)KMyC(q}nYNeT>e_dc+EmrYE6b00z8GKIxZiV*c%4qRIHo}1SH~!$Mk(rL?N#pV
zu|3{tF7wil|C@g2TuJECyw;j|3yn=zj&61+i)*=+(b|<P3!IXBDD6_#v#5q&e(PAd
zB=zRjL}xb}+g<kyF6GrSkKd^a8R|OSv+RS<4DpI0c5QgUv_8`s^UJd8HSv{C8}3x!
z`#GvByrbY^&5MsuHM~E2-n^sY(=o-$;~#mK>$fW&v@?H}m-l>sR3ZQCbw+8H_bLDF
zCz|~V9M8Lcx9iK2Z__?%|ElnVQ;|ti&aB?MKDy?!p=mDfUapQ_=J(0tD@Fc;tD^pS
z>Fj{IsIt8Dj_>kc%W^}Ah*zcZmK(2?gX_ASVjpyCW2vs3oZeCO$7^nEj1SoJ`hCrj
zrct-ArzT`&JNBgJKFJMl4A{1z=SFnqkmsw7^+Wx;y{~1H5f}340nd!Du88i29-QIR
zc=wf8_{Ir`Z)|lSG7=phZINirZgaSC`$hRh*>Cw^B9X++ursvE*WC_Zq!T48B}$oP
z;O_#<OaAIe&*q33v37VIFS|%w94vz%JDk`)zTnG;m+WvYFDoPpnPm{<*{`1s*7Oyf
zu*2hd*;(Q|!ZHYQn7N@~qRDv6Ui1%M_OkWTSO!6UPwUbcZq6QOhcD)3cZqwbWf0^;
zr^G9yC8zhjA&TW?S0`5w%OJ@0dt_%ME+@+E@EBfpkT^&!gCIXlasL{r2UT`>G%q_z
zoMe_kkpCDR-_V`=eV-lvCNDcm9A%b4kn@(moj=*$T0ObkQnj*+uaDIB2DsZN6FOpv
zuzvewQbz)!bQB<3M+0JX3?Np=0^)QWAf!WpRFhSNP)=dbCde&&1s0_z(yfakGKm5}
zg~a=q6-hk_tYjt$tYjext(2YuRw|PMR;rMKR$5O3E1gLLD_uxKE2C$CmC0m)l__MP
zmDRJr%4V{_$`-QF%IP^^<uW;7<qA1yMS29RC=&rIDn!t#@>PWh#kWsZ%~eGMdQ+7s
zkZ@Y~jTt^FU}Fe^w$`3N#B@Mv!q-Y9K>{a<NZ=&ZByds$1)LP3fRj>Fz)2G{aMFkd
zPFhU^Cqpp6$sh(e88ri(EWrXNi&)@f)huvw1P7cP;((J=bHIrR1e^#V;6!Q!oGMDi
z3PtBP1gWM2@>b0dNVuQ;<d%C0Y)C6c`~;2z(Tx#3kOwj1F=PWyoW-r%L^R_>3uF^c
zbV4?YiK}9(pKQfQm`8SFB+MfRF%ssH2AqU>WHU~}Jko@dFpo5fNtj2qVie4yx-kmo
zQG*x-^C$yO!91!Nr(hms!YP<X8O0RLqgyc==F#044fE(hjD~r%0jFUe-Hg*Pk2c{n
z%%hEB8s;&r7z6W|Zj6C>%pk_VJjQ@CFpp`*8JNeIa0ccvMll2P*j9{%d2Bbv!aQ~m
zV__a^z*(5bHsdVJV@)^<^H`&pg?U^n#=$(U8{=ReH;8dCk2ByL%;TDI4(4$toP&9s
zQOv<S)QTaPhq^HY^UxrMU>-8y2<D+?9Kk$f!V%0vMlrJH&Dr{^Xz2N>PRu6Rce3wN
z;j{|hF}bBp5Ke2P@19&IRJseZPj;E<5-R-li;Gf_{ksEg`Dt&N_bB1b=`OLx^WhtE
z_-U0?HBOjb9Wl3+)%j@oY1a_fIH9)?PAD$7v>rmF{r{}%y>|^;F5K=tJN({wSH1Aw
ztxv-y_+2d>&rd5xE1ZPOUyx;TOS@wIkwh^@>;`uVS&Wg8sbY+ROc!G`WTqHnAhX37
z3z;j%ILN3NLstIc(dMGXy%jt8)w)NxYlUCDcFi}o+U!p?{Iv5(XD8vp6q+l(?-}l_
zz%OAI6VKtrW)_o>DYKY@Oq<0tWX3FJAhTvM3z;*EImpN?MvzrWs=q%kjihYRuAiy~
z__aHXb{He9`CE0&GTX3EAMw*lPbCr>;D6m&oA1;AGD6BXZ&G5iPyoJplahcaDFuj@
z(tsE#1BjKffH)}!2uac3bTMw$K(gD-EBtzVY<wmOpUC^uZJ&=AHMWGG_Evhk3DZK~
z>L|DLtDy24V!n0JgqDc45~(49Mrui*Q5p(pl$HV-t)YQNYiXb{8U|>LmH`^8VS&bK
zS)g$m4rrW~0~*pGpdl>+n#x%<K_~)$b|<Pl1bQT`{Sf!BAHD#bJe{Ts6%tx_{-6!_
z@1RmaL@=mL5DAD1q5#oBG$1C30mKHefVdzI5DG$nlPGU_R?6`I4f%Pg`LV+fiNk~C
zB~Jc2FtfeS^{HC2G(o>4F-f#UkVGQA^)bQvm}GriYJE)2UY3>4W5nB$VjEk@2&v4@
z{*BS2M$UH#3nd)n^FtJ49mkAwp6M457&L2k@Kh!1KiP+J`+u`eE)!hcJtj@@Bx#15
cHr-1#!*`;3Jo28abegvy1=mE)3C9-x7d8RIRsaA1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp.02602D130B49C5E9.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp.02602D130B49C5E9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f8579630d0979a01b4d3e65e79550f883d734a8
GIT binary patch
literal 7486
zcmZvg3wR9Y!pC=JBsLc%F44FpO2;L(wR6AEl$c85vP5)Ax8hQCkI=>?E<NfyR#8+?
zPJ4=j>cy?9Rb7fxdU|SgEsE-Dqsq~%<p`&x%J;2&&)ENYo|EUv`|NMt-?P6bv-9pu
z-oeGTEwp#JdXFidUiIXZk&Rp~S5ESG@|5bQE>kYohAfwB`s#f{s|H?>n#KOQd*GH$
z-JaHWHP2sHTC=5M-Kb$#k5{iL9`(tN$-7Q`e(T?2ENkAulA4R5lW+W@{SsxjKSkEM
zzU*;g=*EK8J!01`UA?xf@43$|4V*o7@-sJYzc3>9=?eWoe9`S^ch^+z*W|M;`VIKC
z(Zl6ECSHE++;^Xyx_ET^&Kox`yzuaqiQoS|YPXcXV$QWmiyj;?`)}dDeseZ(ZP6c=
z?@fF5)gwyffy|1Ig?szXd+ZNGziOg9wD?%Xc5TJ3&D&ZRF4<c&&i7vVYf}!tv}{lD
z)j7B3e{(W;G4$1C?(4Ok|5dQ_jV^26KGtaX`r`8R;ah#T!c${=ZhUBe&}1Gn<7+Mb
z(a5YrTelSL-MBdO+@Ntgh1{V#%V^h{-WOK#Yy7i+yM8nKfgQhmzP3@Y-@J^HUuyX=
z_3<Z5EBYOnG^93d<I4-}UxS0UwlB$X4;U2uwrJg~{f(=7KQiY`{P6Q#rrzk=ZtmvV
zi8s!*$=iC3DmuGNTzTf3xlPL_F<(?q@56{q+@p`|dUwhE0axxjez;)B!9m8fgMT|d
z%%AmU(dN$Um+v}wv46iu2fj7<Y}vVvWp9-&*fF$M;hymP6^jl$U$tc4*R|tXF7NAJ
zyRfSL$FA<;D)LMuxa^G&-~Y8X<Mqb>(|hT}{(l)#va#jyr>SOb#vMA^vU^F>L5uR*
zFI7*r{$kLbk=2<8dQCVrv}D?dPZHZU3U6@JMpt%Pz3}YRosFZ5vPYg8lc$W|d*~Tg
zn}J26<*XIca`HFaJYV$D<}ZeL*E~IULul%Py5@6F-fA`4R+m1<kMF#)wtU;FvuWLj
zpFZ!IF}%^6)m?siVDtEF`NW;e^(%MHn%^t;c=JsQ0ymC%OXg&n6%UO~Tzg@yxBPar
zvuDTY%GbB`$Y}HB$0diBZ(ecesq#%v&wgo6|8@7J4ePk_CHX6<SJ}nbunU29<ccy|
zA5|?F?b|OIzj5%L*yyu&`kwR36V^6)?-MmL>S)EMRZ9Z-)lWX#tfjT;bd%3^Yz}XK
zKl9>qyQbE*T6KPS@jtiy;61hP<#87xC&a&;ANI$8Y+3&2=O0t6It@FW@!FZuvBR0O
z^VckX@yn9Dp>GX8JZ74wWOdn=8?DFAY%#E;>+e_I^u8O3{?vIt{r$=Q*Qfnnvghqd
zt4@CS%tH@vnsg;=+^c*0UcN-%T-NNP|9w9$ZvCxj)rlWxZ+!E_fbXX$t6v?kqIdM4
z<BxT_8U6a!p|LN`?>4&e{z0$w_>t}2^_we??aY5~S6B7Q^Jj|U1Lo!4acy~l{>S#9
zuX@C>uBUDg+RT27wpAURbt!Rs&BAGok8dCM@p~^$UvejZf-#e8GUcm`@^>CDnHe9R
z@P2gT<*}=pB))tu^Yclc{`vaIP;r}Ee^=H&vg&tAxSu^bc}CT!a>C_G&kw)Yd%akY
zMV9+rm95fSWi$-Q?}dh+%-`0Xom<M6WsyUY({0n+<}?h>bk6pLZKl5c)5lq)FFD;j
zy?I8%;7psht|s34ZpQX3a%gfoFFmhu!{AI8UFkZtewI3_iK{F*-66L_r-s3q?msba
z>9T#b<=L*slG8<LMRvpBOfT5f?}Jwsyt*}u9G;vmOfT%+FgVliEYJ8Lz3ui_vdF>7
z>74W&cf;UJ|1`2-kB3XJS>(XvbZ&ZXM#JDtUmK9Iy=j*gi?hgK$?1ahg3^Y;nZDg@
z=(}~3{{CGSIVd^ZBE3aM!{AK2s-GM?`QN#zvnyscoYf@<Z>`(<ty+*>N!xU>(@2t!
zE(7w@6)E!C-eTuZ&93zNyk$UsZ$*k2n*o#YF<>%&229pw!DM|bn5>@#le0N6IUffm
z=jXuWZ5~YC$AiiHc`yZA08{V@U<!T#Owks>6n!F?qF)44vL!Gjp9H4lm%x;58BEzH
zgDLxEFcn(?Q}HQaDt-k_)mFh&eJYr$Uj<XMH83@w2BzlMz|?IWOx>r0srz*>4ch?I
z@EKqlegjO?Ho-J~CYYw*1k<uDFfE@2rscQ5wDatC_f~49-NEh%)XAnDa=uBDOX2R<
z2rPvFJwA20(?NPFP!0VjB8A?7cYNx4_XfRFfoi;eB2vf%7*LrY11eL)pvneVP}v|0
zDqF*%$^|%3xgZBBSHq#o2Y68VAP*{E!=ow$1W<*b0IE<UpehDLP{p7Ks#qhUDg`7^
zrJw|=R3o7(2V_v?pbV;9BcrMW6i}6*0;*D@psEH`P}QIcs#>F>ss%JqwV(#7R->V+
z2Xs*Npbn~DqoZmB3{Z`r0jg1BplSw8P|ctTs##;AY6UD%t)K;}Rb!!Qi*_&P`p>Sk
zi|m0wgY3s0a&Du^wQ=_z3TuO=yQMCF+C%pQ;%G5NUfSC&b@h8a-kw05S4<H@Ghi|v
z226%yz+`C_OxDAK$#N{19L<5rc{ngRjsugYc`$ho4<^s?U<$MVrr;646gUA)kru%e
zJtCMQCxR)_5}1-l0#o87FlAZ>Q})PU%A5?QLMvb@9tBK=Q@~Vd6-?Emf~j&Um>R8t
zsd+RoHBJLlr*$xOj}E5J>0lbP0jA+Gz%)1mOp`XjG(9GmCTD_a(H5AN#{$#hEHLe^
zcA;~*Czpc9?g_-$yhF|-<MhmiM@3i)QGz}K_X4_(pwA*F2>LQ|n51`;sRcquNcu2x
zoTSen$0+(JmGZq&!g~a6DBe24dlorCcrPP|N$+klwamN`(t8*=PI}KE$0+YnD&;d#
zg25hBM=;o95(I-iCQLHeV<IGjJtj^v*kfW8gFQA%u-Iek2o`&6f?%=7hDjECY=mU7
z$Hqw(du)tivByOT4trc3!C{X}5FGZnFv($$i;x`lxH!pSkBd<p_V_5lV~?*Rc<k{B
zg2x^oCVA}f5t7FqA18V2@iB_Wo)9Gj><M**fIT5W2-p+Cq<}pkLJHUu;-r8*Aw~(<
z6QhKPJ+Y1uu_q=75qn~o6tO2pND+HtoD{Jq#wZbcQk0OeC)E)W_M`+MVNVK^685AB
zDPd2FlM?o%7$sp(juJBV<T^seo}3_L?8#wL#-1D@W$ejuQpTPfqh#zUQ9{9<Qb#D*
zQxb%NJta&k*i#~;f;}ZpD%ewEl!84qN~qXV>j)KlYJyO)r-n%tduoJKv8TpK6?<xo
zQn9B+2@QK%9id@QOAs3Nv@oe*Pm7Qm_Ov*uVNZ)u8us)kp<_?4BXsQP2|~x79wv3{
z=@C-Lo*pN4?CCK|$DR=-4D1<ogn>OHK^WLG!lZ#cBSIS3GvcIyJtIaL*fXPqi9NH9
zFtKMQ2orl|m^86xMo1HTW}Gy!XT~TKdsdXNuxHf~7WS+JVPVe-lNR=@2x(!@ijx-h
ztQcjbdXLP%<hpr%${DgzWl4(?zjJqPRN1dnze<Ok{hQ{Q4f{55uMQFPGPoCeLxgu3
z+$)(7!636Cf<@*+1c%Ir2p(Ao5dyLpB1B{<L`cYTh>($$5TPKeAwosgLWG8_hX@_n
z2oVOd86r$%D@0f+?k_H%l<QesOJrBJmD=`p{t<T0B9Ym!t>v7(Ap@Fs`C*xsmbCk1
z^l9o?<J*?S_Vt}TF?rZ_+xI(%r<E2(3;lvGdALJShoR0VGXGzF+Q~;3w@Dsu-mH19
z^Z$kOv_WPzY?V35?Em;wxBNVJ#;|W!zer|R7L*nEQ%^*NrG-A{9kx#YFwIxB!A=J9
z?#s({p4yyEw-%WVn{BB8T~E>1@#(6b@?OWMpL&WxX6q>ynX9KbWWJu_k%f9nKo;vM
z5m~CIBxJdsl982qN<miZDHU0(r!-`}p3;$xddfgH>nRi2s;4Yu`w9ErGTPZCqrCRW
z`X7_$GN)NiOXnQ!=LU~S>$BnK<l$Biw(>X|wI7|IIa&DOMrlrR+t5vSN_{+Dx3`n?
z@nlyrZU%^Tvp}4i1LEC0kl+@8M7Ibexg{XkEdwcT1xR(PK$=?v(%m|c;WmIww+Upq
zt$TEEi1Fiy{L`N%ugc;^#Sb{we|BYQqtbhHv1i)##!Xu<N}jSlt@`9U(sxHT?R)*%
z{sWVTJNP?{bQX!T1)N-fclUpti}h#s)}^HK&fVpj`(SQQ$8t_|a_!vR4tDJ9|M-4%
z_Jxz*h<7)IQxdJymh%qDU0mI3!AGStK&;LJaXJUY>pYO43qYbS0!g|AB<nJeqANhE
zt^#Sg2Bhmckf9qurfvdRx^<6|yWbN=-rWkHTs?Yn+1(d`yF<edRatWNV&cx=Et}Fu
zRy^@|V1(;&XUgSzBK5K&^>Re&Wy8DvKaQ>%Ghux4W$D=dRHL-?#_o))>?X~c-dCK{
zr#GFGS=_66>lQ8V&y!?T(+$&ls8ylTN$8sY-;U<CZP&h_;{%-^?9#1!kB57DSzheG
Vwec`Tt#ThN8AaLy9_dT;`#*~0AU6O2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp.76573A416742936A.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp.76573A416742936A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..043cf81f0e04a96635475ff1c85b4517ec8dca1a
GIT binary patch
literal 6586
zcmZvgdw3INqQ)~bZAmX&T4=FAw;&{FF*x^ol%^%L<r3P`8mOU#g49TbEtJZ#TR9*b
zSwKnwk4L%73UXP5MK&Op6-7xm1i1(nXhjflk%+de!lLX6InR*ykJIPjYku>-?T^fS
z{bur&6gr)IGEJtuDTOc9PMtj|&SWyBM*hv3UHAM=%4AxfY%;yH=1_U9>#E!%w0^&9
z`@4_NwC_vH-Y{~;cF%^%<9|F=x3+NdvE8%wojHB)Us5P}(b1xYYp<Mr=O_KAAh+vz
zvdQ$t6KBdd_h0iwsPpEJohu41|Kn!gBM0{Te`!4S<n^n8b(e1*Ydm~Oqbt|-yOn6&
zGH=0u=VVAf7tidxcBA9^`n`AVUj5UPw_d#d%jErX_R9I4yOvlg8;|(cf7$NoT=GKm
z!T8^=J)u^8m{c)f$iad|&%986v8y_C>1P$Y^p*Rz?d&~d*}?p2WB*$A_Ux8bD-IO?
zIRBpS%d_5VuU!0|zqD@95B>MPGkEP^K8u^Mv9K&*!j7@`{Pm#&cRq0bt?MFo?j=28
zL}1?W9ozE{ZeE&nxp>-MF{6BMsnxVL@9Jt{t-JBxx9_Gry8EZo>*Bnl79|$_)Fe#l
zXq`H;V$_E-#x=!neq*uoXK%@ltfEx&m}2i&`5WdP?oykVJO4s!%O3~V-zn&~a9h)h
zcP{tI+|fzpx37>^U-)ui_p%x6xw@Bzvr<=c^@)9Zmiflq>UOH7|G1;YPt7@c<kWa~
z@|OH<gEltrJ9=&Os1dHMCGDk`2b69teRX%aW5@x&Z{?B?|5&^1(50qny_yTm>lW8$
zePqg!UMDXEyer=M;QgPQ693%gzw(y9IQq?TMVot_no0HOGwt}vUO7eGi<e|(E!WQW
zK39ByQeD!Aj_IG57tN_W7T&p8e21S?T{Upc;`aKzU4lzeCVf67Q>{68{I{k)uKa2x
zdF7nc>`iyC<o|u!xpB<enF}|)QvYgO+QPH<(yN`?@|T60L93g}cD~*opEKe772Dhi
zaa-yJ-*|LeO^R~n{`VcL_s#P;GESwv`>N;8XH3!jq^B!}J{Ru%(>kW?*We)A?(<cD
z-uXmgpD#WtI^Mi(<?-js-ksUFYX0aA-Qvd&SiMTQC_758g~ng?^dnc6I_;C|#HX^d
zmep)7xgV--zh7{fQKql!`qyJxVDia|PivQXvg@Y)zDKWNub=PwkKNn+yWUT__VT{^
zru5gZOep+%=eNw~hu)ZWHE>3H^UC-?e6qdyt>$O6+JWQGC%%25I@FTXn7wxCYhM&)
zmT#TVGG&gfXie$%JH4O#T~AlhW53+m!t4nIZwxwYy?%D|?KuYz-`%(QlVy2x3rEkZ
zlrMfgyl$ElSpJIX{;t}?lTHubvu4cCAL(o7s9Vl_b!qmS{U%#SEF2S<_D%Z;@$EsQ
zGPNJ-Tc#AQFU}1#9@+Hk8}FyL&XsyuTjmVf`$HFZ89%hKY<}~-1>bD_r2FMl)f1-I
z_H3`7aqGyaz1pjd_D9y$fBeGj|5SCJn)UpKi?3~2aINI@Tj#vXiBCS|-<w`D?S5%f
zqjq?6_kp8M2M3=^9KHMNZ1EDiYS!G^$z_Del#uOzEpMaLKbb6Zo2t?i(i5XYWL1cM
zCH;Ft`@)ezX)-x368BB$n;IRAxEpb?#p!RqkYai!5)X?X=8O(ToVsgqpZdSt_$Zki
z8;R2r(h{SC5kGzJ$MDu~=I%-+%Oi1SLS~ofV8rh=C;ly=@2>wzCQBl5YC@_xIvDYF
z*P%rluh5lUO{I}ID<i9abTHx@llmX9@nJ5RbVcHfgp9=KV8k~kt{Rawyu34+oDhj~
z5_0mQgAw0OpJ<={^2hU%$?=gmJ0W{ybTHyydz9~KoAL2C$z*XP?wQatF*+Ersc!0X
zvu5{<o!#uD=viHM^xlRYUupeQsw_@Rq2Y*-+forD+DR80|Gktd+D%u)h;cHIF>VGj
z*2zM~x>?9LCkGkl<{;ypJY>9^hfHt^kO^)9GSMkQCb~t)B&P(K<dz_loib#yTZT+=
zDv&8|1v1sCLZ-S^$TX)0nda6Y)15kGx?6|Lk?H97a1vKJ20E+;88d%qQgp_{rSn*7
zVi(KewbTPOSQZd5q&;*^?Ap;@x*n*3UO>c<@i0(jybM&C1_rCFhlMKZWueM8uvq0h
z98@_k2UV_t!z%CLp~`!CsPYXwRt1j$RlzGjRcH{fDtbhyie3?_VuOfP$s<8k@=8#Z
z8YHaB9vP~#SB9$GAY)bWC{R_r3RIN_1*@t@g{taRp{h2hSk*ilR5h;#Rjon8s_xOD
zs(W>)>J2(p9g@RgOuv*WN4~=aRP1=xAmclc6kTnNscEdXks147sWlEqmKUq#@z|ua
z*eo;<Z;@g|Tj|GRbCR~vG!RcqF=DI?G#MKMO@?Qn$y!-xvNjf)EYCudvvSbnY#cN>
zo`WWD<)O*jcxduG4^6=;KvS>@&=hz9nxa*Nrf3tPDe@vTC94EY$tFQl;w5OxRvDVI
zO@^k-%g|J;3N#g)0!@Wips89_XsR|9nkuhCQ?qK&)NC3wHC}_JZq=cw+jMB^ybevr
zV~!!l{EN&{n}Y^o9fCo|y}ddqdQXSR7$htwV2!l25ter3Fk$%~*-u*bld&mk36Pc+
z<W|yh0XaljPExTx9VF-zu+Y+N1l^7tCg|^x{Up7gj7@YpK+-M9tt5Q`IYiMXsaT&0
z5)As7HiAJP6DAn+F@BOk9}^%M^f9d@gFYriG3aB11dBeljbPEoh6xsZte<4j#|B6i
zeQYbqqK^$xEc&=0!J&_9BRKSNVS+;+=O;PzaRHJ;AJ<B9=;J~Zhdw??@aW^)2p)ZW
znBdXJ`$-;se1PQ9$G4I^`uGsVqfZDD0{VnDLO`DoCIs{eeo{c65FiEg39Y1nJ|RR2
z=o5p4h(5865YZ=w2@!pwpA^w221pToVk;@4PYh8a`lKKsp-*ZfB=kvPLPDS9CnfYr
z0a8Mr)JjU|lR}h)J~>Fp=#$$B8GUk?kkKdmNf~`|fRxcEw~{jY<PasJPYDtV`jj?8
zL7x&P6!a;6QbC^*AQki}t)zlJB}6IcQ-g$xKDCWd(Wiz96@99oRMDpfNELl*E2*MS
z4N)rkv>>6OPirGI^l4#2L!ag+HS}o#QbV8CN^0oSLX?I+JxJ*2)7uCgeR`PC(Wm=K
z9esL$)X}H6k~;eI5T(a_xxSmGySHawAmgf>J)LgjZW32jFtDJ?pwUwEj11e-CL*P(
zuiQ7!cv2rXCT;MyEA;rt;oQXB?nb+&Z)xaILF0>&!w$3K5##Wjk@>+PZgFhna8_<s
zsqtyj_rvgh%E_gDB8Ss@q-7Y-RNKE>ICuP4YhEK$s<KP7-LczC&d8jx#^S%@r4QoA
z)^2h}em1jPW`=PKG(N}nNQ!=bg<D|<Wx0*Fst$_2jkl5xia};OC>ELPpg3f{gW{2e
z4oW~4J17xZ>YyZKxr36Cl@3ZlRy!yaS?i!QWW9sZksT8q53kwY!HH!}CpP{wa&A(4
zr1mn-$zi@^O8oFm-$f3mTheXDZp7cO{BD-`?VXX`A`h=-%fQ&_rp@#~<8-G~F=hsc
zHM2mRnFHd@Jdj`(fJCzhB$*{3*(?JoW(7z!t3aAr1Jcd<LpnO{see{xpZ_#6cZSCe
zf7F<ODODrlMm(g;SLY8~k^0`n$ce}*%c?Y{<25?|c9-tGmqZTdrsrlGKWoI@lcIYb
zFr|EiWd%Gv(LRD+0Z%uKk6@5lAHgDXK7vE$eFTpz_y_@6^bsPm<Rc_x*+<C8ijPo`
zRUe@uYd%6l)_sH?V`EPz@|W~rLkHfr%TfZ9#6_3DByll7tcwNWTpSSZ;(-L003^CZ
zAju^G$u1d4aVbEmO9j$g8j$YNfgHba3^1<mcjg_-N%g%4D_n*pgEro`@utY6=w?;y
z{p0FroQFv1!M;_jgbS*%ffboR4|cC&C5zoM%=itF%_loO%#cR@-0TbxYiEHtI|szu
zc_6_q0Eu=HNU}>nvRwvJ><W--SAjIU2Bh2dhg3Ip`lJUNakJ{GXO%v9;dwAbzo@FW
zO7=gkF5d9lgh>?>pY>Flo;4zqX=3bUMeJo|>}B+~{vTG?PMKa4c{%dA(NtV~LKkyl
za!S`8-Mba04$rfuCKWo;diU)0NT#f4y8WrAhYd|1q7D=v%l`k4X7ufs)qlXFgDis|
e&w1iWn&AZLH+&x(o1dOBqG&Sd9Fto>jQU^YiRR+~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp.80C5AB0FEC7DFC63.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp.80C5AB0FEC7DFC63.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4b8fc1f16d396374b8f737f6ee3e1b17ff17ece7
GIT binary patch
literal 4724
zcmZvfc~}$I8pbmdmJqNI3W5kkM26YU43&h{h=Qa7LX^z_Cg=q>1ZhPDTy9+(t%?=2
z+FQ64y+vBVrPdls_15jNL5#F6ZJ==jTU^rCTf4dSj=9gs_s7liaGrU;_YJ=@^Ua)`
z$xhR1%Y7x1l+v`CirI^2xkw}uh45Fkc+vbTQi+81lt^kCkLPDp_Qm;JeczVz%ACg1
zBkh>d_uRrQ6~5t_+xAXq2;b<FT9(#T_xN6Q>#L1%j#>6pNzwP+=W3G%Tt;Ds3R=+4
zFPg}f+TR}DSr_@5=5b=vb)#`*)4u7O9-PJc`<Z2j6U>M7O--RuoYgz<U>;kk9Nquh
zvcLcOG&<zO8}IGg|F|=CSw~^kxigb5@#phCd$jD&L*~wZr}32=eu=(uzWgedk^il{
z;AY2wu|i*QY%qJF`C9GmcMs)NXo?q1$yn^BE{aXhXxZGe<xKd%)}5Ej9OgcU`n#rx
zpT9iX5@|os<C4(4jh0^crgi0*rX$?5g*}es<|&~Y_uXi%+4}Ah|E<Ygt&-3s+PNjO
zGPjS#21A1mw%z&m)@hSY8J2KW9`f35ttE5n9#dOb1HTS!BTrW5`M6H*uGza>?YqFV
z#(hnM{6dKLm>-t+#B6tIZE7#C$v5c+N~ZePuk${jIGK<%!7n5==fjks6|Kt}4(~iY
zGP3HDB=+iky02#N;g5=TujI2KS4u8ES#3S_X5!foeAmfGcbtDuUw5rJy;ir;G^P0+
zpQ?6q{+OCoIU7#CY1pE<KDBgv{h-h1Gy4)+u-t`g#3%h%W^_HZE&tN|we)oJtH;7J
z)9Z8I`tCKUPtaU4%fo+%X~ri1kgbWA1FlYcdh=HBn)PuXp`G_nSo$u$)w}EKZ`)U+
zYC`)>hi2wK=gRI8=X)mh*VajUY8I(JjgI~J{2y;$m>!GgA8o0-v#V$zqCC0ron8ML
ze_&Vio^j7E?N(e)&OciC$ftF_Vbd||rLediiN=6wsU_sD->Qpw+t)93kb5Gs75jcZ
z{0oVdJt;rhc=TetrsTs%O=*R<+D;Z0P4iXNcrW16?`AE^+MmAuP8uGwI>E1WeQx&c
z@_yNluFQgd!wfv&hs|@&43w`JpD^O9^ba0ome`-A-=N|bzCWz<Y~smV-9<ZlgF54<
z?0B#`2QM(jvgbZZU)i2V{&l2%Z(G#Hkr7c91-k6t@2f@@k9*@hjWP=zJ2p=;L~bsu
zi)ZLRsuss2%vvy8-&__}ZyUzHKJn#S#q)nsCo#=!%8Hex4ewq_TS0O2mX2?DvLq>#
z3F<MOvJrg!^3Kc?5vt>NUU_rH=q}4Z)Q*O<eGA5p_%V~PynZ*j>#KE#!;2S9+5T{Z
zf9|I1(-#NM?u#_!rgZL}=YDkk^^+g%3450J#)iuQpV(C=)+#kEe_CU1R^JbKa(cs3
z_p1lb{hIpIiPXQmE3eHj@-lv!v3F*N&%@p$y(Oz2Q0<D5y{=Ch-%H9mICxt2@+Ye-
zwslK0{=@&i-~DY#Rc*!J0x*fhO=+%6X^tJ`iRVt282#M*Je-5@{5wDLN4u}B%~0of
z;yHpH<QAlG4x+6%xG`wSfxD+Y@u`BXa8r0V2hmP?{J_!reeFR{JWsF#-2&a6gJ@4&
z%iNA%T}ON3dcjt@DHEN8XeT$Vo?Cq6AH0_&U$CQHqsKc3(QZg;KQ_=maDA90Td;$N
z2ZuQa(Y`xt)MrTLi1x&Fg6;3-@8KLo`>9X<$M(uI-+SWI1l!lm*TXr8wq()lIaO=C
zhOUpl+<9HLoO-;w?PlUAFC(ENvc(e+l%6OWQlyRq9;qjRN9icwQF;n^w2lTIt*3#<
z=osKJdIor`js+g8XMu-w2zW@3fXC@L;Bk5mc)X4W9<S$trw&zzz37fn9j+!sCEhPe
zx$`~&m&-uR8@fP58L<SYYGMsGBuN7aqGTBfqGUA*qm+RHQL2mrQL37PQQAO*C|yQ_
zC|ym%C}Utilqq9Cl&NN5lr^v*%9gPp%2u;5iVO%uQ5gbJRE=PiGjJfvm2n`-RdX=P
z8+Z`q%Xkpwt9ckz$EuUXK6x3{lhit(rm1I$N_?n=9zK7fK@SOn8fp`Ph$x^`MC_0v
z2@)hpM1mx#B0-WOD3GKO1(KAC0!f;nL6Sx^NYW}ABpHGMNd_?>$*35RWC<1|S;T@Q
zt71VC5eOs^LLiA$2qZaz14#~XAjzpXkmLy-BzeSxB(LH@QcbEEu{}cTkU9#ev1+xb
z#07VP+_}(#)^%gV1@Md!c8s_NwF4u5gxZV~9r#c~2@6hiL2bi{UZ|~7;-YjYPj+J@
z<dJraggnxLk&s84aT4-K3r<2FX~Rj#Bdt;r@~Cc%f;`HOQIJPDFbeW0GfqJsWx*-P
zqii?@d6ZR3K_1<W(U3>mF&gq{2S!64ZN_QHqb)cMd9)3uA&<67X~<)`F$VG&JH|jB
z<G>imW6U@Md5i^TAdj)(4CFCZDFb<IH^xF9YsXl~V;vX^d8`>{A&<4-Eab5^oP|8r
zDrF%Lbz=zfkR3yiha4D!JY>cZ<RJ@=AP?Da1bN6RMUcmJV;tmhc8r5O&Vg}|$C+^s
z@;D35K^|wrImqLzQV#O?Zj6UK-j4B*$2%|{@^~}OLmqFzdC22!I1hQeRmu<X#ODx~
zJ2zBA%Rqd1yk1-@CYtU`{(QK&7jrSfvS?#+=Wa{<Zf0zK<#4q5ZOqFU6c>~t{?6xS
z`9=<I8|NyNRURsD@yqi@=2(wU{`*3qtd^+*#q!dON!^S-W~xvQP7cl$cW|bkQ-k9#
zZVnR43LnLA@%Klx<#Ol7PMnFb*)m~7Jy@(r6GlQ!nJ@}!+Jw<iGbW6Inl)i8)X0P(
zs5ukHLCu>men{`Xx3;$6W@DW&du57pvUu&=YC2t~R_xIV<-nH$hl?xo2~~EfYwDhR
zLfJ3E4~g6FFK#WXiW&SP!$UYb84)i26E7nvBY~o16i~E`28xj}K(R6wC?rEbaWW1l
zUdF#rr*g;NDOP@SPMBSsOWbI&Z#Vi^P1vS5)F+gK^MhxKjkz7=`q14wV3SZ*`Ky#-
z`o+#$?%Z*MzG;axaAsN(C`wBKMQdrG7%c-7t7U;gS_BlQ<$&U~JW%Qp>R@q-_hfBb
z@q|Oq?`Ho9|0r?B;!A;W$31@|05jJR6Tr+hBv6!w0*cnqKrtEyC|1J)g)|5#PQwAk
zYxoyx(d_bB&tFTb7L`=xJU=LUKK$=fC>}^8GhQ4@jgyJe<Dd7I-*zi5m{C|{kQBV|
zB!;2G;-SNmiqi6Vg3OqcDRpsmbC-E|dJXgOeko0nnnEb#Y016;FZ%_Kh>K4grx~9#
zf%9h)!edC~|Emrk85}YyZ1k9jNQ!3ps8^!Zv7xFUG<K5z@K@7k;@Yf<ld&oP1xC-B
A<NyEw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp.073EE92A5EAA844A.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp.073EE92A5EAA844A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dec47fbf6d43f255e9d3ef3775e57ad06a22da79
GIT binary patch
literal 4724
zcmZvfc~leE9>+5imJqNI3W97AkzuwwgOacsQIJ$XAhH?21bv`Y1ZhD9Twc{GXjLqz
z)t<tu;49J=TxxAYdG*!p*dQXUOB*O|&=!}p_0_i4^^JKQ^ZVm*4)@IWem{rL%)QKH
zl9{a4mU>DgiN(p)Web+gcaTUV3gNeMS;gW@Qi+6hl}M_aKF>`pAB^_6{H`Ty(!!?V
zBb}Ji^UTt1Wu75v4g03;2-)nBRFd3N|L}HI`%6vH*7-w8l7b(4&(y?^IE=>*<+Y*R
zUo?|#HNW4xxgqQo&BM6nt48CR=KXWF+&zsA4>Kzd$C?i5o125exjuK_gE?%ua>DTM
zEC2r6<A}gxuf4PX!-w5TE4%X3&zzchfj^t`*@KmT9x`?RJDD%v^lQYmv!$1@)ZB07
zdDpu}jAi<=Pe(JCny%D*|MsDrGEL!<S*gpM)CEx~scl>Px19<Z*}nT?iPbb{Re#qU
z`pcI`+rowp^gG11G|<v>-?Xop*nEV0vb5it&@wA%^Zsk?)!W}*?!7&sr(F`XT>Em-
z{Is2uvC$yEgB>@&y>YTGLm3=<SswVxUadJz*IL&Ryo29>8pz}2IUbHPd#m@YQhP3`
zTkpI+RDLeded3R+`XhHbv^RH_R_E4bj1=j-8#lNgP#lkqpXL>ql=Xh1-|F_2I}YzY
z=^s{kK@xTO4n0^sdhaJir(445z)MBvAFb;<Q5$#qJ<kpD30-I3(br#TNvX-$TsN!b
zEsx4hQ|`p-wON~v*BZ8IuIh?+HjaAyYu;dN8<xG44gGld(%hcMmQ`PxzLuVBdFj*O
zw3NoIH@<sC>f!e?neO7ft8VTV@4)SG7kw_zeti8#!1|5RAE4cLj+qC~zcH}q>u)>P
zp(;ZAO_yfgKVFvICeHRxAFipF_E%S^K8c9>=<NMB&&`R#bC0&w-`rC$5?Y$j^wyq#
zPCc+EqIJrX3wsq86LOE{Kk#T@Y}oQ?--Y1lU2#UA*-1s@p5LnqdCS)?bdjy0nTq|t
z9R8KWN*<LSZ8~~BMpN|ugXZM?8y&~<3ub$&s@<1xDYw!q(mzbucrzJ~To>zAyfHiT
z`_f_AuAa2KVZ&V9=f|xJPmPqWo*FyutCaU1q!kT4Nx4SFEPZ!O_vyIfH+l<p5BPP*
z%-VH#T^62ajAGAxkg}#ThkSaZb6-dJX8+Levb>DUKklgf3#YtxmR2vwx9-|H!w|ML
zzdnYc?^iC1jGez^fxe|AxY06(e|7qcHwqX3td3_|I+SH=ig&zyDS0);Em|>k$D`%(
zL5yF2-3bf9H?HbVI~J<?{N|+E)f0NmBjLMtB=27`dE8HFg!$E55j|gRI2=+~F>B|&
zao*WmuFhHJyI?TPke%4QcaihajaQF<&>H+C=e122eLfyi9eZ7=Y5TJ;@_N;sz(*%H
zt#H14@XT*XKOaka>Rx_jae<rhlhl3lx;*X;92qEDdzb1|_zyTfYI-L={ov?H*^3{q
zHCr~UNc|80$7jxON-As0{^o;8Bu+|GePT=0cvn1oj>PEY<mF-;g!|w2BY&{>%Im4>
zELS{BsQsM$6t+RE6$dx_EkAJUq${oyYK4=+#Wsj_{KLD}_8)2vy5c!P?d#<0Y#YRS
z#YZ!P#xe1fE3Ox6rIRwwHaxF?o!7`G&bU8Dk}1>yV*`S1gIKrvH!hqsd2+OyBv+^-
z93!UM2C=?1fBa`iWsPvfGlbgP$=k&?i1lNS+>eIJPyOJE&lYM=Cr=mKAl8zK1q&<J
zyMgtQ+ty{<iHCbTuE&jcGZK2DK)e7qBdI4rqVyC<w4Mfu(K8^idKM(4M<8)}4kTXB
zgQN~o6XF7hd!&@x_6D4yfmmceMWTdQ4pJ4d9<xW%KmwF3ApuHOkr1T}6hNsG3ZPUK
z1yR~S1C%bI0ZLcV5M>MuK$#K-piC75QP#i$lr3Qa%2u%uMFs>=RDuACst`mugK+Eq
z^$+Q0<VrYza#b8ec>@nnzJv!TU&TXI9i>hX_sPwuo}tbFX|{T<n8X`N*u!Uc5!ge5
zpzNnbAR-(j6%l1mBtZg_L?j?d6$wa+pa4lB3XqhF0whh)fTR%(NLob$k|7vCGKc{r
zqhbKb5-cED!~&94v4BJb0wh8RkVu69$q^hNIm7{yQ*nUg2_BF<;sMF4ctEO2H6w11
zuysfs4$@?`T1?_YZkpWoa06S{ixKC*Jw^;+#1)vW81WO#CY<QP?Hfv%aiRxi3r-Bc
z+$SZ@OYJ<_i;>Vr4q+tpkyea^KGKAf&_|ka68cCBPC_5qCncee>cuGNqlPdF`Y0<#
zK_6wpDd?ljI0b!_1*f2o>XTB?NB3eh^wC2Y4Slo~qoI#B;WYHoW}JpT+Je*2NB2o-
z=wo^@2KtyGjDbGJiZRf~m~aOA7&Fd5A7jB8=wte%4D_+R7z=&u5XM3uYsFaTV@)^<
zeXJR0p^vrTEcCH`QWpA9FNUBG4PglSkQGDFhfFwvK4iub^dSq5pbzy)5%h7r7zcga
z5XM0tXT><^<4iaQeViHRppUcQ9Q1L0QV#m~UW|u6ehA~CkGEnx^zkN~hd$nn^U%jz
za31>jJ}GbaO|QEn`Saeg0nEV|6B45rpAKnJjpbv});BRXqhGXNqWD>zo$eVn+AzgY
zm{z%{+(mkke^%e;GylF+m{!ZwzT)(X)ET{uK2j%42P6b!ulVc5NWArNx$V8l{^Wrt
zAi{FF?d{195w#f603JVNEk?pj)nXLPbS*~1%+z8G%xo>j!i;J$1T$BQaWM0>7;n#Y
zza$03oZspvT%y85F;@IM6juO~+g|R(v;Xk+>os}To9YExnW&s8E<{Ilx1+ACRVz&U
zzTi7nd@>(XWfnUowcZw{y+XZ^_{#gmjg^&=qxVx?gqJxP5hDH(;bl%nf<(zEkZ2hV
z5+h?kVr48yNQOY-WE@DmjDMa^WKX?WsQl)PaCXrS(G$cgzBaseT7%-ypfDYf8!%7Y
z#_z)&?>W2sY!Rka-YTU?i*LK$a@$+3xK6^AX^CX}l@U53%brLr2|7|of{xNsprdpY
z=x8ksI$B4Aj?prpV{{DYSS<@WR>y)4X%Xm<4uOu-a-idM9O!s04?14QgHAn89Uv}{
za6<*DgT-`P*0CLrJ@o9Z^bYk77gtaGQXt$b&%P0WRn!pEz`AKjkSGlW60M;@Vl)g$
ztcC>%X%I-9h69P$@Xu4lg3|fVzLr#06jf$DJ1BTI{O>1UJdjA{K0lNi6Q4?-HgFm5
zI~C^5%`Y%W@}74RgZ;43eppmiT)IezsSDGj4vtRFG8b33F&^$OBrB2<357g4!PDnO
zFW+&|F>zBgQ{$&`-b`#rB&qy=vt#`O0>=kWm>3#H(JUW6DMB3;r1C?PXLyf&DP<n6
KO`kp!oAqC*mda%S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp.98C08C6647949567.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp.98C08C6647949567.idx
new file mode 100755
index 0000000000000000000000000000000000000000..abbe2d584a72c50ff74586cc9c7a6abfc501db15
GIT binary patch
literal 8894
zcmZvh33L<XqQ^5cp#)ll7ARXOMP!0td@PgMXEJL90yOMli4a+YAXkM#z>O<H3koh=
zP`RK2E*Ax^q7N;K9zY)&2mzlyF=zl$5z#<Us-j-Kq3;a&zjI#C;XC=y_xsO}X8O&U
zWz>*h@bzn5E_3>jMRRA=#oR8JtGw`Ue%-v;dCKK_ywv4dwDIk6lSY3gw)^QBq4ufO
zo;&6pPvu?r|L6M355l5+SBKei=iCrd*KB<Bk6&N=VYxP|N7r=!zwyWa^<1|0?)epe
zb?=?I{`JXZVhexhwfAdplJ7s1Id<vXncq);HukOoJ!*gI&|}I6KYo3p{_+oBEvg+L
zWJm8|b~f&Q<iQPxwmsD2nb{Y={B7);d)|8LAKx7MugsGjX8$^4=ddSErZ0T+M&j7F
zm;JNvYjMx|!{XF^tta*DCqMD|HvhI<N%UZwW0PK%+rR$Oj&J&{p1m@?Y38=z?Bv?=
zHy>NH_wu{7@m=#ipPQ)sOKruXQxiY>+rsg)$Q?%que<+DzmL@i(-S%^Id=B-&0+eH
z{nPGPyQd-b=7aB>2e-Pa7X;@_i`i=i5`|y!<?Y{oeEQvn+8#YC=CsYt-V{s@4{vGM
z-t&ng*G#Gyx%A}Ln%0j#c`^9&%<De6rbSzKsKei->a!a!=%0n#ZCd#Ht6wCm)Qa~v
z^?k0r+8A_y_q*%E+*x+r@hA2*yp-rNN7!@i`VA$GJ7-SH958Q9G|ZaOuJ8FiJ*2N|
z>ncC_{Q4W$)r@)cgXoF(>Mkwo`?b=ddf@Vj#;RSXCw{n3+4#@Sw|)A_r{A^7t}-fG
z?Hjaf$F$Dc*9S&vQuOfipYM3PVO2k+Pt{#@b6$P!y2nOY@3zrqZQOkK(5-zG%Qv%5
zJ|CVEx#@C4@_Kja&asnxm+$TP#+2T>20XvB8`bUnuszY6?%pdOX*X{8<C`Bi@MBfW
zO&2ZZ@j=|H5B<{aa3_fzx9f!mzg^qytKQqq?;r2{+#fd&n0eje-`~)CZC$oKzT?vP
zO<NZDPOSNdeMjHKo*TEUJaBZz@srVaruUP++Op#553KL^H#|I&9P>+N-hesJjK8$%
z#dEjP6IQ-m(r>Ok<mle&ag*1MdCD5GZRqmXo^{_wF6i??>H6nVnFHm0*Gbb>C+_qN
z8_}g{MKbeCT^+OKv7wixhZg_lz5o5MURlhhDlfPnQf`mN=Y0Qb|FH*Jjv%Hl&c0H$
zt#8w=ahrdiHl^j_SjDlSQ;)w?{rvg=naw^e-@9#~Z}`OZb;^CO4ysu2RR5R6GuyYG
zy6|E0?1~+0W6NJ!K6vJ_J?{zrl}o<Z#QiaS8Tn@=*=s{;`-s!q?i|tm()wYOe;QtK
z>=-xYiH~>R**9^z#`j{wh0<F$>=}Iid;0Rz?H;Wj{GNBq#t}bL>Qf8Pyzo)%WZmC?
zU)fOAM(q9cZsEx(!NpGwTzKlOwJX=C&o2GpUklCOV(~*e-}*GXuFvtsNB3ViHTJ{p
zzeL+S_;Dx3^1nkpQvK0Ab2j8Fm)yN%`j(S=({C^4xBoHa**{xu__^Cq>cfR>#~V5=
zz4?`KcbD~7Uzk0^l+Hb~|Hoq&Tin|@wlt_*96oDV?>~t<Z@)1#fAQQ~>Ij#sq(|e!
z=4P&@lpHs~73)yap;hxxI5RZ=wElI|i3dXd(WT_r!gSY?uI0^xGhP1j%B~CkdG6g(
zGF+H0FDY-;JUG+rWsmoPP8(LVag8cWS5;Q^Y#y9xvD4nRr;g5D(8e{oFkM|)eO>e5
zOb33unA>`0@yn&;_`-DOlFlug2WQ$i8QyTN<rck-YfNFfdu8`t&4V-j!1?^cnLF-F
zl#&w*(={bEX7k`oZ@%~8LD`@GJXcCiDop!Id}{OHOrM;RU;gyE`(~86B8BN8Er#gL
zgEM_DR`aHh$qgzcYYWpIOFFh{9-QfawHx<p_V)MAl#(M0(-kEZt(pgCdftqg^T$>d
zzk989^LM^_|8JYNe<9VB#pobC#QFahQkWhMDMC*wCU4LSq&MsZ(i`yt$po1p&M#dS
zW5Ud6ND*dIG5LdjApK!Kkp74tNH)j<$%a`V*$4|H7vzBC!W@uXgaeWf@<8%o9!NgI
z11SUrAce31q!1B+6oVp=Vps%HjEF!=K?z7HECDG+Bp~IW45S>Efs`XMkV;SiQVA<S
zDiH-pHK+oqhE*WdhzevN7yvR54geX51c20n8jxC815%4<K<YsqNIk3rsYi4mji3Rf
z5jKD{A_kCV&;-&9n?RZo6G$s)0cnLTAgzc6q#d+@w8J)#cEkoU)Fo7PbtT6_)uEn{
zdWC2wIcK@TLcDT{gM}ESXB8J49j6ySO3+J*V)RD6fO_LzK)nesqD+*TRa|aNoLK-V
z!7L?;(I52#>W}*Y^(Xv@vQZXLHqHXdCRjwdC<iDP=K$pr9HM-b2b7QVfbt0*Q6VY-
zD#Qgqg@k~p7!?5(<07D9LPS)GN`Ojn2~a5^Au2~@K;^g$sGN`yRiX-@N?ZX{NhpY_
zQ58@%t^%qiR73;O0HA?*0MI}pfT$ML0M+6epjtvhRFCR_>Tw-VJ)t9NL=8ZVxB;k<
zFc3APCZJ~A1k_BJh+0t#P%CZ$Y9%a0?Whf?9k&6s6E>nDE@U_huq+m`LbZ@ah9*17
zxyBgPy7}4y7AH+#UtE8*kM0kNp}Au6(q16FJ};17#tS4vGuIbaALC>CLt+>XQl4Ud
zq>9;3`=ROg`Jw4&{Lu90DL*z@nuR9oW1-11EHv3X#bT4AIcRb|4w@XpL6gf<95#8H
zhbHgip~*8mH2FNmV^g37XbL_7ngSz0Q^->SHbq*5rsxx)DKa87#XKcqQ=%nkN<Imi
z5+gxV%2N_HWm<-&?31A>Gcq*gJSAgOp%rK<J_VWzqd-&1Qwla!T7{<SQ=zFcDm2wR
zrD8Kc2cQ}71)v#V0?-WPsQ@-LT7#zM)1av_8Z@;$rD0R2b!h579hy3$LsQRFIyMd3
zfTrOyplL7$G>trEVAG^cXqr9~nkHjH)67#QHZ9tMrscDsX)zWwtvqF6)23}`+CCea
zHe*B6&QmrvLw!R1oRd-EMCA+hhcqzccT$%(E^WQ5eDouPJ67APcB-@4xd|d$H{bje
z&tId4jT-Mj&aIHUb@MGy5u%q7^cuX(SVnj;M}|cvIy-T0m{t?5o9~ohFNX;_4SS(a
z5cEmR=Lz}(=G`RyCRsdz(g#WU5a#0~{Tb#X6n&H`^4`OQ7k%Cngcp6@^Mn_D-rb}Z
zecpql7k%F2q!)eOBa|0?Op0I*!(|hbC72VKa|Cl9b0f*@CW|L;CP^{}F=t5TIOa6P
z9Kn-)itwY)pC$b0^XCXZ`uvTgAASBL=|`VGL;BI@Pg8#Mu_=N@ADbmu^szaDMIYNp
zvgl)zB#S;aL$c^&(-ez7E=6$Y<FW*YJ}yUa=;Inm4t-pb<j}`uNDh5mn&QyMrwAT>
ze3szR$L9zheS9OyqmNILJo@+y$)k@?Q#|^F6d|Bb$Pxnjgd8EDPiQ0s^a)8)K%bBy
z1@sANN<g2OB1H6wSwcjgm?K2=iH)R)J~2s(=o2%fh(0k*iRhD3goHjROGxOGa)g9F
zsgab>CnZS<eNu*$&?luS34L;kkkKb+2^oEIj*!tOH<B{?<RmGhPtK4s`s6evqfbc@
z3i^~Rp`cI65eoX0Mp8kak|Y)MDH&2hpOU5&^r<OAMW31_RP?DiLPek2NUG>llcb71
zHAAZCQ`3}+zCelypf8Xm0_Y3mhyeNmjbs3Qfg~9~Um!yU&=*Kk0rY7pLPMXHB{cMD
zIYL99)<|mT(~_iyJ}pCP=+n}ahCV$-=;+h4gpNKvN9gF&8%Z5~dXm)9r)NkVeR`VG
z(PyLx1ARu8Fwke@2m^gaBWa+|NRkHnj0|a@&qz}S`pgtzqR-3{Ci=`AVWQ7$Bu(_0
zNzz20nITQ|nQ6*IpOqpk^jTTLLZ6i*Ec988q=h~!Nm}T$GNgq*D@|GGvr~kPK08a;
z=(BT#jXt}Pw9#iLNgI83hP2UVrzyMW(;L2bUA|OzoN~uTwI3C7E_d9qvF*n~inNbh
zP0qvN!bR7WC&R@np|QipPH=WGbixZCwcYyP6+~IACR7u4E?<UoYi_Uft$MVuxhvn*
zbRG_k8ByUq`<UFKuz6^!p>3U=Og+8g){>JoCn^dLRhuYgnY(yZ;cue+nAs+Z#mqHP
z9A>_W;xP+Nlz>@mqD0J66D47mn<yEx(nKkk)h0^C9B85fn6)NK!>l(^I%cDZGBBG>
zl!@7DqAbjI6J;0ktZt*zZx6ojFNN<f<Oy|lzQ2VbE7dQ;hN-exm0dO7xmqrZb+@~Z
zcMfA&tlF*~?_4C`e{bWmeW$+IUD#9AgsQ1d#W{Y3V}9jvfAN^()p$eRt`d%{>RRZQ
zw=1u79xkl@a$~J;$%GkKj$*x=S%F8f-tEWC*1K8ET)msa%-6el%tF0ez%16gMa)vY
zTf!{YyJgHuy<5So*1J{AfqHiUvsUlcFzfYh9kWsIHZYs@ZWFUr@3t`8^=`YEoxkqF
zjmnjmCdDrqS7zD&S(#-qT$$UxdtmW|FE&10H~_w&FXZfR`=Ym6gy(Jv7B+XjrgNop
zEAtLBYI+NM%Xfv%9eQ{0Id7}>p1x~7d-j)5hr-(t58cZ-@?|lv#|w$^Fufc>S<LV8
zLt;HFB+kP@;ypYh!6QHtJt8E@BSDfqG9<;LKvF#_q<|*?N%LrsbdL_n@EDLxj|s{0
zSdeUweKqYLqkcQ3$FcX_Ww8<N5!X74tt>XuJrdFw_ZUdy-Q%w&r%||-yz=(TInJlP
z{Hyj@`756ljz{%`>S@l3{JMXOOD)@WdZ@5@Xvd*Fobs8&FTVZ!J;Q4Yo4Zfwev7l3
zowsL&eP4NX=Ir}B&-O=1@s%s*-HyFG_ID`fWm?D9%`ed$x^UzbT7x66cp)(gqrs6^
z{E%3Mg~TZwBwpbm35ozoR76OUB0-WB8IqzXkW@v56i@<?G)04?D>@`YF(8?W3CU6{
zNVZ~MO|E$}Zi`)c(>Q<LwE3g2e7O0_@V`G%=Yz{NrTB4j@#EC1AF0?uzROShZdwv8
zxn=T{=*?4Iw_KfbO`ALYwpoQ3nmL?uw<u}pX;oU*rd``>hLqbTUEX?#QPHV=ht6Gi
zL6Q|UpxxN9pV*7-)8qf!TG_3tdri-4uj}o-zV8kH(SN`oKUd9k^$oN-R^Bvp8W|j+
I4<?5F8`?2d;s5{u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.436428D0A84C0AFE.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.436428D0A84C0AFE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..603cd1b65849a7b9512670473a50f765919fdc03
GIT binary patch
literal 8794
zcmZvhd3aN0qQ-L;YG{ktmI6W-)=192Zx9Gupe?k4vJ@duS%m3`3)CANQ3t^)(Yni6
z#Vc|LM}+~egHT0r8IW6^5L^%`7z7*`5TXH$xX+}$_t^K3+vn-`<afUJ%@0oceM!!l
zS`i9O9pv$hU0AVV>7vFPXphHJko{fWxa@an%Hvs`=kcsqzjwy$+GEZ&-#u^FJlE{K
z)mnPvCC{qulkyf0D=iDew(ro_4K1%oY<lO{Eh{SqZ0~tr^}$IKxBoP2*B5gqlaG3i
zaG$KJI#J#0lM%;OTBo|tNsRo<%$@)2GjHq15B~7|h2X_sTSm+)86nMEUs66SQQfoG
z!O3&J7&&fl>h!03cRtei{dXBA^7$7BKDoR%(>N)n{nGgM#6I&9<!7Q-%yU2V_}$&~
z%4ZHZC!Q(ZF=oaJ;f<N>jAQznPxjxsbM76!o2Jh`_@{oK6h1flz`Ev-ANLP#9DK3(
z%Dj7OH*fm#=3P5)jXwVF;Lo1feePWD<PAmrURFL_x%TLUkL>fOzR67*z4rfO$`N|P
z-aF%4Hm`W=&kyCEn6kLh=)G=;zvz$uIrht^^yjXB>&`1D1LNkWHc>qe4SS<|<+qRO
z|G3n)sA^DU`uD}-h(h)F+2arGIr_qhRrPDfEZ;{ZF9(+NuOts8W*8UZC!4IC=>wDL
z_w+$^|GesfjsLZTxqaB`p3}ckzFW3pv|{$Sb-{rhuQlJF_`2+){R5_bJ|*zo{%_85
zTUu-9Fr!}@@Yt@m)}B7wefs`8D@QEMeSOW-zu)`%#s{ym?(2Kgk;6yM_ZT1YuU&Iu
z?a{+>j}6Nf_+NN#>zS!vA3Q$xgQv=RT^)F{c-GX%C;B&@%o(61!s`nerstFgi%TCF
zeyGpkDZiY&(Cgm&tXKZJ?&sbAbL^4W(WgH?^7<-0>I;4LQJHj5U9;$kMQ1l$9I=mB
zGNPb46>q+vj9hrRy@=R+eeJ%+{raWC83U`UCQU7$8Naro>|aYi^fnCKd_8-Y_t5sO
zduLy2^j4pK__OiaM}2c&{3ZEW)sEN33bC_mC++dCywG=Ou=#m@>*Cav;=A@u%?y8H
z%br=QH|^LsWJ!I`DxvbL>Op^*zUlpMhY`;POLH%c@x8ok$8BpWPrtb3q9CmLTWv_4
zzswj}TH3s1)6`#H{O_CpHRRQbxjPRwB~Lw9f7_YARxNE#mb|$C`*;1TxNZH{EttM(
z_hIJAcP^ZtKKJQ8PZd7B<CW$^Z#xG*SiI|<S0+_GF~b*WZmSONJi~rIt7Pq))qOMN
z%V$hWUpc$5anHK_ua5b2%#gqJpL=Tk2YW{?|8!UTwJ#L@T*eH3Hom3$)cIxAwVB5r
zihQ@a=%Bgt`99i#+iIhy2JlO^mgg>h_N|Fur<@<2x#rQ`EsxiXvs+5mEl|$o*ZBUv
z{E3Q7n|B|sS-<ef6WteVTNzQNe7tyCc;FW|pOp^p`0~voO&^T?=;4)l|01@<4!?Ks
znW{hU-Vk{CV$bwh`M}N{-EX-VO}|t5?eLfG9{GLE(I1BXIQ+=Z<##UL@cGu~?>#-{
z!()dU9w9#1xBbzZmR$K!m~v0)ybtd$@GaR^ba~B=X_v~UKRB#)`5jB=PbWN{obuL(
z$G%_=$RlUW^3?ar>D9e+$nFfCKe<0{IKH|{ugxQ;XQ%t+^egBb-06aCYx~{4^{c<<
zk>TugK~6#U&cU4?_0z@FOK0xbmPgLaPWR2}dsXM)PFEM~DBOPe{-5$ab=m2W-9}b*
z4({}v@Jn}}f6pG%!!s>AUD~^JaOdDozaXu!YUbA+&m*tTP7lf%G`e$er}xgkqv#Js
z6?P9#ZFah>ciE86!JYo<h5>K#T1w3$YqHbDImO*O2Y33ib*`6k-AST{XKHr3q<6`{
z&cU7j<(e6<C2!etCXbwwoi5BN?A|%J)1GCEZd~58r0eWDxt(XV`Td_ZY&ki6Kz_YH
z<gal516d0DYk}(gv%821F_2`!3?!L429j)ug(MqhA<5RUkmN!fB)KpLNv@8ABp>1-
z$%lDJ@^w5Ug^&PAAuK>rs1qP5hD1n;VG)vIod`)OBtcRNOOTZ6BuL638Ip2XhNN64
zLsAJTkW|77B$YY^l4?kWq#9Nssn)5G)Iu60wXg<Btxkia9?~JHhjmEmbvh)CkO4^}
zY(UbeGazY(Oh}qx6Ov|~2}vttLDC9akhJP7NZKJAl6KgJq+Mr2(g`_`bixiKojM1S
z!E1vhzb*OtU|Db=(BPoYC3nxsE}D)F2Nq4lzoctn_#6DU14aG!5?#nd7>F_r3`ChI
zgHblZLX>S_A<9NsjB*hUqFe(9Q7+11l#lQb<r{d2@=+e6LPUV5&>%omhzb}LBO*k_
z1`(oSRK%zhksvBHND!5x5=P~S3{km3hNv8sF{(rqh$;;VM3tz5Q8l7MRBccpszz0e
zY7q^hT7w2rEvjKukLVE98+3^3Q5~a3#DJ*LU_jJ}8W=SrCPd8!6QX9+#HbaqAZj&O
z5VfKfM(v0VQM<v0s2#O2>O>rfIt>m)ov4G+pcx$FE|vWH;J9E7(3IdDm)!T}Nx7Zx
z$*_!ket*|W@$-HONb{ROY07_@>SD&nK$GDaXfhfDO(sn-*kpYyG+Ca7CabZ~WYZLj
zP0q(bljAvPavBFsE=_UR<b6Ced7g(Rukq02(-e<Q!6!gd;00(3ngC59O$pc(eIhhP
zUWBHoiO>|&l!#5qCqYx<C1^^T1WhSTN!XNqGBjmghNi5^(3I1Zj7`O-KvUrrXeycl
zO(jh!*i?NgG*w=OrmCsXRMV7-P0go4Q{y#gYMKU3Elp|I)O|WMbzX<2uIbR!)0B=)
z!)HL#;0<URngLBCO&QoUeI_(b-h`&9nb0)Tl!;BtXF=2AEofSr1x+hWS=h9FHZ*PC
zhNi9A(6rN(jZMerK-1wJXgZn$O(#t`*bEK}vhK!_-BkFX1f&FYm)uYE_vChdl7~$t
zPWTVN8?Zk~_>Uu}2>;i}t)%}AvTJMcw~_w+$nB*6C~|`GAH=OFPB7?Wk_3Z3CPgsl
zV_HcDeM}q4ppR)M8T2s;ia{S6Cs_2cNrFWmn<7~Bv8^PFKDLcy(Z{xvEc)04#iEak
z6CC=uB*CGNOA#FUxK@%wAJ;~5=;PW+4t-pL;?T#(2_AiXlHk$DrwAT>d@ISLk8dM+
z^zrQ^k3K#@@#qucgn&LFNeJi@QiOm$p_LTSC$y0Q`h<2;K%bDH1oVk<LPVdKBt-Oy
zDMCb_*h-4%6Wd4;ePTN)qEAdvBKo8_A)!x75)%5P6d|EcY9%G~No}NrKB=9Q&?hA*
z34L;$kkKb62^oEIijdJKw~{jY<Tg@9pWIH$=#vwaj6Nk!DCkp?gn~XLMJVV~T1f?c
zN*k%5PiZF=^eG8SL7y5YRP?DyLPejNB2@ILt)z-RwT)EKr?!(S`qTuaqECwx8v3*(
zp`lMp5gPimR#HQs)<$aR)7nW5eOiLj(5J@<9esL|(9x%-2pxTTE2*PTZzFZ|>FuPB
zK0QI{=riJkfj%Qi80a%ngn>Szl{C<2w2=n-jCRsMpOK&p^qFzOM4y=?O!S#4!bG3h
zN}A|1+ej0AW;<!3&rDDz`m8u%q0dSZ7W%9dVWH1zB`x$>ZKQ=htDUsaXC){LeRiC%
z(Pt+K8+~?)u+e9?k~aG6Hqu6)-A>x*vlEn!J||8%=yQ^UgFYukIOua)Ne6vS8|k3W
zX(t`@ISI<?@{PX#g6D^C8;??S{kWoW)$XSOx_&~@1fa^I%HL?R`S>lp`G+1Q^6UFq
z{l>aq>fI`l+j*_e{Qxnoy70QQkB#V-O^@n6s;8TBF9DLdotOUHbbkGWnhC#s(WyW6
zz~Y=Q1{^QUUP#MO{xx*hrwS%RG01F&Vv)HF#Ub+<ibobQlz=Q|C=pr8P!h77p=4ww
zLn+8=hEkEW45cCK8A?YsGL(UAW+)Tc%1{=vouO=GCqp@1yrh3^;_Zod-<&<kL2t0H
zdy;RXJuNRwKYiG7{$jNMPk8=fG=t2>Xcn1^(Ht@#qj_W@MhnPdj24ll7%d^oF<M4e
zVzh#+#%L8;i_sdg9;0<+BSss@W{ftGtr%?~+cDZkc4D;C#s5(?F3#2&VJg4AWPHg?
z_Zo42ed+kpneLxzetp^avYGDH=T&#Lu6ggo$v3j6yku5My<53IJ-);~_)x#>;eu-l
zdb?lMo4;IN!{0M&QO7IMMEf7WSE7k#kl7}hMdq4l4w-MFd1Rr97LdgzT11wbXbD+v
zqGe>IiB^!+CR#<-nrIDKZ=!W%qlq?<%_iDJwwh=Q*>0k3WT%OCy4YRebZ+N0|E_%j
zcc<(w0lQOvJ?>7M&#%5?*2(n`XJ3Fp6@#kX=i9PkSGVxejiKz}zE}6{?S9+Ytxa9n
zZT!ab*~7ht_Tt@pChwm5!*b*7_f;J)m)AeoeUkio#>)V)UKWV+azMP72NJvjkmwbG
zB(DS{du1TSs{pB96-e`HK)P24GQ0+m={12YuLWd#Z6L?%{6_Cj8+mSC`DgpGi()K2
zc96TN<kyGj5Ksl3J<b1jzsFr1*<HQkzD(DvS6NdTaYHA*{A2L3f>)1a&w1IbvITDE
zwAAg}t9tf%Fnf4Z@u+e)edfT+Z*N*Txh8wKbXMtncQLwmd9v>|9k+N~(|PZfM@Z+@
zj`ybGp~aG$a#v^e-_>!)$4%$gkE<DXy=!uB{S@VP-ty@R`2+qS{0RgYAU41PaRCm9
z5AZ-jKmZa0B9Ii2faHJ-qy!WoHJ}1%0S!nG=s-rm05St6kQK0i?0^mA1f1XK*SC+I
znr5VCbgY9G+P@LEtrnU=W?N_$nQNgrWWI&wk%bmoKo(nQ5m{=XC1kmUmXVbfT0vG@
zXcbv&p*3W^h1QXc7TQ2ITWAy6YN0J;yM?xqofg{Z;_Q9Fn{Md1Nw|F3g5|Xx4<a2y
z=MPo??xBjsL)TnAGUtZbbLT|ndFHrN9?#sahqJpL&g**E`K<qo3zjatX-W2B)s2%W
zx?9dw-tKw%J+A3_bw$DWvA%-biZO+KihA|E*0P6>2#g#R6pC4AuraLs|2x{de@W?p
jfrEw&z0S{Yq9m)jS*G>lm2t(rCsfWSL)D`v5|jQ7EJJe-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp.5798AD2D53C823E4.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp.5798AD2D53C823E4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..009a4c3baa1ac80b73057cc10e0eae2d7d5ccce5
GIT binary patch
literal 8382
zcmZvh4Rq7w!pECrjIxc70YeyL3>XEP#3msjO+wN{Sm*dG+t><NDPMv^P(HLGAbR17
z2n<05Kll=^Ts}k*xLQSg74%}U7SyZ06^9}hRG?U|h;k8qyWNxRf6sY$&YqKhp5OoZ
zZE2H}rj-?b|Enbq$B?NN_s^O(e^Qpi;mFVY&YM5?&a0He@lvkCasS5mY9>^jQafDT
z@_ki#K5_icflst{R(x>KeB|RZW47NqbHbEy|N5wJuNyy4KR5g#v*)m9Hg`Mr+2wwz
z#kDbK%=yZKm)H%<v@xSr{`UJP#EJK>U80|xxM}m^=MEfr@7Jkoe*Ce`;o#+2L1s<6
zU%K`?ws6&7d*A%h>`rGdEPC+j*?~7dxM<6Pb07Zvr%g9|c1)SdT|GBBw*IS8ci;56
zyy4Q7j*5qvc*lWu(XB7+eR%!m;g>&scF=uP{hXsee@2h~^WN?kuMMO2yxr~O{DP;}
zS9g1U`?A`NOKw!`t*JkhTDt7w_Elfb+56+_Lo+#|HvE_4E6&p=jkSq!8{V3E&#8dB
zs`MMfJK}U`cId-TdQGb?DV|?Hb4n~{z%=rm#o;&i@0hakmmhqKOJn=4o-AH2mb~6(
z*p<hno&Wl1(#q0XzmeUKuDkX91^+nQwasek9qLHOW#{W#5B_%HNq-)B=f=IW{&&UF
zg$Mbah3RGOZ+<Misb|TRc87O;YK&a8bKz6F)*l_%C%01@`LnxbcGtiEr%^J#TZt5T
zdQI!|qwiV&;_FqnxgXwXt~=Cw$NVeVpB`VBH#k->qoVS~g3oi>Uq18Q<b9J)nJ47$
zx;;|bZD{c;vz|C&R!?HG51y57JHGsHzm$GEtbJbV;x7yP4p`wIARcXNKAPCNW!*oX
zE*9Bo^9NrJOdP-bTEBkBN`vdh|Lxqi)6?fH9I)%7>c7v<OWremU1U>|I=`UZ)PkRi
zidJmCiF16k{rO!RUtb$~clit>nZI?mSTo|-myVa;SN}4*!n$x{7(e9r(f`R`^>)Yo
zyPx{_&_iVx$G<5R-}KI|BR}pK(R-UU{fi%lyft;W_Vu<A&jyBGIC3blZB_Wpw;!w=
zcYX6&?S<W2p6TEB=fZJ!((Yy17h{jD%-Pdv^(VV)COz<Y{&Q<xPws40Hu?Nr^B)~{
zZio25hDY9x@$VJ$t$Ora)6Va$<d=N%<ZDZWxmym;-rV-Udzw{sV#2Yzs8CN~+M=(!
zn&ZW_T>_uS<?(;M@Y=nHhvh!CuZ9eM@WzHWrXN_eX#e$33-;vgX**%fm{E?`&ddxS
zSNiv^e`3?6wKJYwHMwu)^s<r0s(OFx`7h+LWBUvpbm*T2%)U9lb$a3Okxd;&{Cz@q
zX@4?)feAgId+Feb{#W0;d&j}w-+1@;Q3tgA4Yk_b6?-3?pmtv=9D3!sOA8-yEF5xb
zk!yEss;~BmioZ@4x?LUVO8zioamASWIa>n%K7Rel$8xreS#j;gJnQ=Jul7w|ldD~i
z-sd`)o_KQiJ-7e1`+U|<ecM&GUGh$S@h6LBPWiRd;)P4@Sa<s2pD)JJJEL2+-#zKW
zKeVT_&z`zg{qZf%H}BtjwZr?rcIdahZK_~jZt~F4fs6W2CSP`3>2;J}f1`E%%jzHV
z7R;JFmT)+zs9EqT!PexG<HtF|rP-xrO+)7T(DaktGg?37dATH)tjSDw&hFf%X|Shr
z{=N0|(&187E;%+cU6@^%-!$0M`7f_1y!WLGM{~(QW;#DRzfIF%Pj}9L@~7FK9$eej
zQJI-8>QdCJX|Sh#*MCgyJip-OT(UMZ-8H*w>!!h;p3!#$qjYq%YU`-VO!w&0qj%F_
zPtW~q>6E&xulBZeRA;7(yA+o;4fgbX@4or`{=@J8k>?0zrpsEDjcpq2=?jxe-eASl
zz+7^4X1XA|piR?YPv7WJv%lf)chBdNV=~hnvpcqF8tiGu+-Y~rn_bv)bhC1rj_T?o
z*SEfSMk~n+5AQR4a`T@M4)h5?8rx@VbFz=WnbUOKTNZHp-4*t0mlt+tE@qH|?g=f4
z_R~Pp0UAg;NCV0E86cSe10)k<fMoqFkZgbjk`1yzdi)+BJpm7po}dRv&d&kK1vntN
zAO|Gx=Yiw{Jdk{l2U740Knei?NFgWyDf&eq#efK;7!-k&{1T8-Kmt+<N<hkf8Av%G
z11Se(AiaJskluh7NN>;!q~ceAR00Z+N>Bk(^{YUt0ToC!r~;|^H6XQs2Ba3$fYkju
zka|D|QV;4t8h!&vBVYh&1Pvg4ejkv&fDcGt&<CXHH-R(*CXi;(1k&<bKw1F{NGoUo
zS>CO@sO3Z~FD~x|sZY7fPWF{9b0Rig=HNsOxo5PTZ0<Vuy^tdAKM^gF4$*+pbu^%K
zghrGJF@Q3444_PeL6i-#fU<QgplpOi)D!Xm>Z$Vp>WO#|<w6{wTpb4}7vT`)Lp-2-
z9S<lU;Sm)=0-!>j0H_cV5EVlrpkkc}s2C9ul|mArQk?{-6p;{>Lo%RpoeZcPkrDNV
zynuS^ynuQmUPP6U0;p1_0IEb3MAeWAs9L82szy{qwU7p=R;K}~MKnb9kPfI`rvs`-
zbVQAi0jN=D0BS@GM13J2puRdEpuUI?Q8Q!$YSx*6nh_IGD`Wv`)meaA5ew0Bx%?LU
z1k4MU4=o=JX-xTTcCzm?s&ks|FW}^Kxd*hIe{R;zK@#0Eq%`IJiE7ETi-so6($J(u
z8k%&PqOr-i7-%vq15HL`pvj~u2Aiylg(l0g&}2mxnrxb4vFUMnpy^>f(DaBNXnN9=
z2b-LWgC@sv(Bwo8np~RVu*ti4X!0x%O<v@o$)_nEn}SP#roal&6hr}<LYfk=DY`^x
zimV7tQ52ymrYRAdl1qZ7#7fYVL<yQwnv$?7yJTp}tPD+Al%Xl7DH)qymlv8|)(cIq
z=!K>?O?k1YxD;qAtO89%RG_J(DFvIVONFM&s?bzL6`E?AQn9JIG-zt922D-WpsA%P
z4V$`4ho;Wz(9}g8ntGbjv1zyrXd0{mO+z%GX{0Fwn?9Ehnm*PCO`qt4rY}wTuxYwX
zXqv1EO;a?XX{IR?o0iLhro~#&v_uP<R+_S~S>C^#wl7wh3mIF^K@!R(JK3Lz$egB6
zLM<1v5&cF8He`P;T216MeIRO~+;PJFA>44d8wmGl%qhZs8FQ3$zfQJX^xQGheH?R=
zbbo<4LAg)hMKVs%=%X758hvz%pwUN1Ng920jHJ;=CrKK8bb_MM$HWN+eM|$vppQur
z4EmTT$)Jykkqr8nB*~zUNl*;>*f_zWk8L1W^sy;|MIRd_S@f|nl0_ezBw6&a35rFZ
zCr)_K=V>53=<}oq5BfY&(t|!vjP#(-lO#Rp^CTz_`nWj3p^s}IIP`HTf<qq{B{}qQ
zF_J?cmn1p#aS4h;A0H=p^zjV@k3K#{@aW^CB#%BmM)K(6lO&HmK0)#56XJw`KB0jS
z&?lq_0ewQ06woKcNCAC9k`&M<Bq#xWVw@1sCpHiw`ot6=qEC#HBKpJ_DWXqIk|O%V
z1SO(RiW3t0qy|DlpOhjb^hr@tLZ1{PCG<&2QbM1Upd|FkaY9C)+(5|alT(C@J~>Lt
z=#yilj6OL@%IK35l#D)aobaO0+dz2H=S>k_^m(JC7k%Cs=|!J6NqW)eO;BF+DRDwU
zpVB}m=u=XJf<7fmD(F*Uq=G&rNh;`55|n~IHBPALQyT~seQJtO(Wgd96@6-qRMDp<
zNfmu+f>P0^#R(04S_7e>PfHOR`m`vip-+pE8v3*(si99xP#XI5IH99YZy<E^=_x`-
zpB^Q3^yx8DN1vV~b@b^8N=Kg&Ck*r%4TOO{BSjeKGoqw{J|jjN=rfX}fj%Qa8R+xH
z2_O1=4TKMUz7*j@pD#-K(C3ShKJ@vLqz`?*1m#1Y87EBinGJ-AJ~Ks_=rg0Fi9Rz%
zn&>l=q=`N=L7C{Y;)I1htAVi4XQc=WeO8pT&}YR+3w>6Sw9scID67S1Ex+Wrc4ht-
zRMvkz9?kl%$D>(cd^|c*UUzpFcGcrVUbxU(IK+NPQ$4C<zwg!#Zk5^XZ{u%o-#CT8
zd31XAS0$&ZyzoeRq-yWZ3s=&WkZS4L<}{;wRpRii56;f4RPHSAYFF%M&zbwQ#s|+W
zcVJ`sjTvY|bFv)I?BcGU%;-kZqrCR+yl@p=1t~}eo0I*}I<s!$6YG|BYx>nR-z45Q
zd}y2w4hm$}Eh;OjweK18!ad4*)Y^YpUbwidxYl0Jc4zJw8y|Js>pr|Vy6l~=&%B=b
z?~2A1h3%D=r^||z6RQd{yYoBbcd<WRuRgnRG`nQnw8q0Xc%)&knK^ueD<Rbmu5C_x
zE-qa#?##ye%u1yrODpX=%@^-K)G9D*n?JL=YrC#p>>Jv*#LB6y%C=q1?C#vRGiyKl
ze*2pT<_-Gp*WsNr&$*rMKK3GcVcJPUVw?;l*2zNhI6aU!CkKgl@{k0l07-O;kR+!B
zNp{MRyiPA9#i>A2ohl^FsX@}6IwZqsK=L_#kW8lu$#PoF=}5Kl{luQ1AIrQj!?K3;
zZh2uwXN`tbomJhOzCQbqxi){_sm!4+9#=fYexID<T3u<~{-%|g-Tnf9PkSZ%+mOuL
zr}3*0`|Wel`QjQ<-Pd^b74$9OZ0fSyGd?%G_1(<w9^-mUvfq*RcP0fnP2ZK+i)3C@
z-EG2)s?(4doq@#aEF_QafyC(?Bwpts3AzAD)I~^=E<uuY8Io7`LQ-@ElB%naG+l$F
z>pCPuHz4_RA0$&ZAz8ZBoPIxi>YM68shY;ChdOGA{ZeGE9%>b&pc-sWbEn-ksqrhQ
zd2^@Ct7_Z`H4gvzxy{~iIBsv*oY1m4v3ZjUe{qD$t9<0x580C@+<seVqN8cm|IAOB
zHTA9;na$yMjHI$!Ww&;=$<1rqp?$lG{IVgg{G5tgI^NW&bJuQOMH{ReKJ(@RTJ1Ba
uf6xDKZ<p>xJxY4@F74ZIfSd7f{6I-A77JNnXhE0qVUtP!DC<^Y#Qy;vpoT2~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp.34C01D1CCA40C5F9.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp.34C01D1CCA40C5F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..45ff2bc4b0b7cb6fab039570c406f97bfbe0941b
GIT binary patch
literal 8376
zcmZvh33L;6;>R;HP(oXf0_|d<T%{?=B$*jzn2VXo2z1tRHBgQeiIhu9o<%6Q0sQNN
z2Z$mHth%m%QbmPDRMa)>;*Hd}F?cL!#T1I*as>i-A*=rl-8b!b-`l>|ANkJj^PNvR
z!_1pwQl(|RJka47HMeqE{k)~KIyoGUg6!{-r40)%P!7lHJcnc1<`1UcRMmGxkCwpa
z&ws=bA1x?7NsRogvFQGxg%3YA=ivG2XJ7yEr;U&N`ZM1$cBMHWcgn-PdmjA1vrpW%
zd#rV%b^pgp>h})$dTqn>6F)xHeVLV<KePE?XTCdkWb+o^f$}>Zc)04s2MrfieQ@d2
zqN>FW`<`DGIez?%=-!?D=lp>2O>5JS%1tdZm@iLXnyp%$=gjm!cI%b&cfVVF=aTpY
zX5!_ZnR(*bm3h;@Dy}=d_r%RFE|?=7>m$v$?afP*zx?YfTVE1h&fQFmpEhwp<)a1a
zm}MXQ^2O1AL{7$z7@?T)(T62p6u*7(bn&=%diB}S(0C>HwW`J?hqoUc9;+YvsC(n0
z0sV?D_}|$5?dgrfJUy;cn_hio?tL4Ftva@(cHf@osvrFELg3CJYeRP|s2WxL$k(09
zR(x2z^vbC{6$L{s_AMM;Ipl<Pre)o|2UZ;;W_2C8q={;|tk;#gRtd?&;<?1TdPDxs
zoPDkLJUBS?&o0hq-mi<i&h#2^re)rllbLmMXAN#W@TdCvXUeyvx{Ri&XTR-!_Ttn3
z>YCg*Jo>=cvL1y)8>a1_ynbu&#eGftH_ce}*pQRc7r&S}-hcg|oW6hQ_T7qmHq08+
zbI_$5>!18XPGjyIZrhgKt&@z+Z<Z_5Rs`kDvBPzhl|!~avbyie9$Vg(%XYk?-RPUL
zR68G>IepFd<>gI7LL05u4>ca0zx<9JFMK-rz{33IMe`>uElCrXwJv)eSl)g8=!1EE
zR-W2W7n#3s?&B8+Jz2HY8I+1gJXgJN<+QmapS||%knb07-@NTa>rtmwn^$va`sq>c
zJa@aVWn<H*MLQSmb8lK($Nuq$AGh7S=h2cy6J1qxgU7dqn}(Kt`O}q4vk%{O@ZyL^
zcOHGBuI%IDngz7y-kcxieo%2=YVC>Fc24<Y|GE2bdA4iWiW@eCl)fd6llHnFYS~`4
zsA1e<YupV@P0!p@uzT{Zqr#!|Q|rE-|I(59?=QSK;S=HVVeW`2tr>Uy0NM3O(fDUO
z|6};&ho)}(xqjwLmqwQDdem`v=lSjD4|xYXb?(<Q4@_BoX!VOt<CHEH`^)ycD)$_}
zY2vxSnO!f$*L?cvsg=9`ux)$f-VbkieBa%j*FX6C<2$3T@^_d0#80-wCu_EB|KcS1
z{29J@#Dy;9Ti*6Be(dwD6W%*{Xu^!+p`(u_4>Tqgo^xH@I_RC2#E0`6|1|E0KXXqk
zU0uGZb#Hvg)smTv@)6h6Ys}0m2cB(w@8Apjr@xi@<gR`7pWlA5+nZe)xBOi6Yu?R;
z3-6uz{o}Q7?VqxvmC2mqK7K=7{l-s6u13qQluUkm>o33c{nGE3Ubp`G@F709PzqM&
z4a(dXd3}v}U`=e(t##9``X^rKelRntpjYgx4a-+Q=(&1b`tKgbv1EDu>?wr9LB-6%
zzYxBudF1pOM{Q|NX+_(RJu|fZ<P3<dAGJ*ym`6^{PWR5~-L-A7r*nUPy5*j+!>jYi
zDcR}boZ^DE!JaO7X>IZCyU)CzM+URg1vv#>+Xj33`kyQ6eps<<L%t)FogUqJ^w_q+
zo-RJMasm7QCtmI5n3SFFSJZD%+h9)zu3k*P@a^)K^2lk~={`Aqy0i`UbXE0;H@H^k
zj&6?0+3Eg8{fD#-_Vn8$*X)0D^8>eab5v)iONvScw+;66nOOt>>J!p@9$A&0F3c(H
z+BVqJzx0^8Cw1HYZ}Z5B+3B7+J-fCI_Ozp6-YrWO7I$pjPPuJc^`ZB!KE11T#K8R8
zF{NW>xBvOI!O|e4sijlflfC`T+_vrBu>!Z{uC(`Eeyuy`u7(tH-_(&jmIp{r&;z6=
z<N=blXdvkz4I~|+f%IBlAiY5^klv6NNWbL=(jW8#=@0pVWGn_qCddHEgcu-Miv^Mm
zvOuyS7D&$GfaHQ4kX(oZlDBvu`5+G@AL4-&ECEO%C;%yh1RzCA1X2u&K#Cy|NXe3b
zl!6kFQb+<)wqzjXpbVrOl7Un#1xO{R0I7r&AXQ5RQVps=sv#9f&C-C>f*O!oNCQ&0
zbRhMh4x}E^fix@wNF!(fX@m?Q16BaYKrjGgAQS-7v`iq)pb4ZIGJ&+NxB7J)h*p1V
z5TsJeWheVgmpu@#o#x;`47=+(4mNkheLJM6`);Bmdcqz+JrNI}o~Q>=I!psfM`%Fl
zD2=E$>;=>t@dE0NdJ*-9{eb!-en9<EKcY;S0hEa_fHF}AQ8vs1%0^f~*(i%B7v=!v
zA{?MxltYvc^MLXZ9#B5YBPxUiK!u0^s1Owp6~iK+VnhT~jEaa#VF^$vA^|EzB}C=0
z45%ED0hOaNqDoi+REa2nDp3VdHLL=vMpQu6sEVi-)&SKa8lYNKLsSpzfa(z)P(7+6
zYJ?3yjferL5j7AEgad#EA^|`H(Ey@m*aXy!n1Gs56H!aDM%o8peyugys)97pnqepV
zI%9Hf+w}z;oG!P!<M?y?+$<!)EkVjq?#on1_P9LI^!Pl`^avhkdNPy;o3x9DCheo4
zNeeVI=?q0<)9dm=)9dp>(<^wP>CI4HZ2Db(X!?DAX!->|H2oRMk4?tKK$Gz?&}0M#
znoNdbu*teuXtF*QnykP=lg&^pHaQmuP0q(blM^^-av6%lChy{*$@_R{@&XS{K11=?
z6kGx{1)l&-K@gxRWGDffqDzFP=o6tS3L-Sc3?*Vya!JsXd=fMzL4u}~p(Jd|E*YA#
zPll!}$k3ECl#ETqr9e~hDbQ2|1)55RQn0DIRA{O`6`HD`LQ~C9DmFEj22IVUK~obn
zXlfZs!=~=ip{e_HXzGFvO+7>D*fd-QG!35tO+zrCX=Eq^n*mn<ngL$`ngJmI%|M0<
zVAFJ&&@_D}G)=*TrkSBkY+6Gt+CEujPh>ueg*40>ZYTR+BDd{M$Uc$TcS2-t+Z~X-
zIKOsm`PgB0$G#tWC|kaEH`GD76NLL?xDIir2zLwSG~qsvIYzqok{#zlcbs&8gt?h?
ze}y?oxewyGGeLOJ=SdMB^m)>R2YsFx=|P_-PI}PiX(m1B^CT$``sf5fqmNDzH2UZ?
zL8Fh3ku>_~I7y?AZYF8;(MgI%pEp5x(dSJOUi5j>gcp6@80kfyH%@xd=WQmv=<_Bi
zFZ%om!jC?GitwY)pC<h1^T$X(`uuUyk3N4h=|`VGN%_&oBnSq5Op0L8$D|1ceN2pG
z(8t6{27OF3$)Jx(QVjao1i_+@O%W{m*fhbSkByNm`q((hqK|DRS@f|<ibWrnAUO1K
zDS|^EmnJy$aWRrZ9~UP%^l{B3hdwSzap>a{1dl#GMeyk3(*%z`K1TBB<KrZcKE9dc
z(Z?q#9(_WB5YQ*22myUUnh?+@#7F^sLYx%PCp41+`h+AUpifK?BKpJ>A)-%A6C(P=
z7%8GpjFTez#AZ@NpO~aX^hpUqLZ6f(B=kvXLPDPuBPH}naZ*B`)J#g~laiE#J~=_i
z=#x`~j6OL{$mo+}q>Mf}PRi($n@Jgca*~qKrz8jkeM*W@(5Ivc1$|14RM4lyNd<jM
zGpV3YNm2^>)C8fTPfZaj`qVU`qEC&HD*DtosiIG9CROyQNlHbZmLN3rX(>WOpOz*x
z^l33tL!TBWHS}rCq=r5%NonZQ6NHXFJw@o~)6;~GK0QY2=+on*jy}Da)X}FWDII-A
zf-ul$qzD6jMw&3tXT(SYeMX!#&}TH02KtO7WuPyRAOh$Mq=*3e0%;<EzCer&pf3<7
z1LzAhlL7Pvl2ibFW`Z!$XQl`feP)_4(Pzd;6Mbf!G|^`^lP3DiBxQE^Zdh~9apk+E
zUs0X@>#1p{|9WcLsTQA_j+ZvxR^(gv2$5e~EEkWmpWWDxgtM1X*PaJ=Tp&#=pVoc}
zHhDtN@>3hM&e?87*NW@xtE<}M_s`EcI<SSxuN_B^Q|-n1wUg*ckV15*J<T3g`L|xP
z-il=_<JIvubX2O<YDhI|O?#^ATb+FK#=90~<(AXxV^<b;ay+qvz52<uHPPc`dvShk
zHC+v<hOTK(zZ*L`L5IX(w$iVn-!%IQGQYN@qGX!=@H@Y@e?|Xk_GVdiXYAg0kG1a2
z%KO#ytF<d@G8O%lgX@a3iwk-b6xsLM4}G(_%J=)4dDk|GHqPkyPLs4rkV0ChJ^lIY
zJ<DraH?PmGsdRklB>Mt$*RsEL4%RnX*~NXj_bIZkW8V@c&Fx&#cs9GZcUf<r{UqG}
z?%ADq?WujoPMlaik+;9aJI7Zn5l{Uzws-c?xYJ!~oAPTtP7fs7Nkj5Fy^#D)KP1M<
zKw_ONB+kh};+;Gs!6`rzogyU3DM6B*G9<;RKvJD5B+aQo(w#aa!)ZVYI0KMOr`ev~
zn{1q(IpB+??B1*FR5{q*ZuzxUovI*B=`^K19s6dbxuM|YBia4Yzo!2j`+v<zw)1yg
zuKUe{*~Jxw6$9)w*)LGCHzU{Hp4hwTtaI1Hr26u;BeSrqkhQ7v(tt!m?bG|Si%V)s
zX4(7Le$7&t+xB|JUP*SR>TVNus_ubA>og><?uF#n{g4=)fyC-8Bu?ib@j4Gl&;>}M
zE<%!Y36iYKkQ7~kr0OapP1hjlx(><E4M+h!0Lj$N_S7)%)>+qHK`m*Rv!wdkhw!!G
zKR+|<4-Uu89UpIM|47w-Hgj+9bzRmj$eDH1%`?I?9W&eK9Bpg;Pn=Uf_tv`X$FaAJ
zr#f}c>Ei5~m*1_&b=@lqDn_{qaw|vn{7tXkeXf_45t?oUOn)J*mWo3M{I8`&efyOR
l95i@HS-IQeWmt|MF7+3ReZxi<7X5b2Y|@%=!;Qqae*;0`W%&RA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp.2C6EF29B32E3350B.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp.2C6EF29B32E3350B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2596bf20600584be9fe3bbbd30f53f62404e55dd
GIT binary patch
literal 7560
zcmZvg33OED!G`DFJA@&Nkc5OS3H!j@eZLpTjF4pn62m|QCg4b9OPI2hr358VSsY~v
zEksmATc8CM5NZo_gn~<18VG|ZE&+m){ve9OAHYBm)PItmOWyyS=A3-zzW00Hd2;SH
zU+$O@Uhi|6He2r05sPL`Tky2qX0s*7e&;WkH{&X0v-uNjwneWW9yhW0!r)GK{y1pV
z8&!^9o3p0hvMt+Jm@su<W|n*Pz609J{d<lm`q#atvMUAT)+OcF#MG4N($8=FeSEd8
zpG`lt?@CHI^_Y6*51!1McD`)Vr@JoPUNdL>x5p|g54_pjobw0aPS&eJ(Ym{x%idhP
z`W5G%IX$;E+`08#!-=xN51N;5<Zt}%)o<@LJUiH#|Ju^`p8E5eIp1`=edzw?A0C9%
zN2w#V(=&g4I(}krCi|6(yV7<=y9Z7tebHlYcItth?|f5GHDgs|%k)#;88t5@A6&oa
zgP;C6C%AvX#aRo}Zxp30`kx7(8t%Z>4hwSnj*WXw7=3@>;IhDjz}5>Zb2eqikNR-*
z%`D|Qoh7*bL@qkl?TxDE`;Rzj-}24Ji>5vPW_nLg<&Tc_{>3v!QQFAa%TwG_lisUb
zShwNDlm`>qjehRvTH~P6_uYj97xnUQ`}pF`p?51go)2xX?f=`xq>1yJ^D<WK{Qd0a
zue&GjzD?zw9j$IW|H_9i?5yavc+fxkY}y<bes_9t-SOPAiWkbKb)s*)l{MtbZ?Cbt
z!^Kb1!`|w({?I>GUAfj_?8(I=v!}-IU-{~<j_u#JX0iEv*O_0Q`|^5HzSp&C<=30e
ze=a3$nK#9?^^HAO$J~fq%00d@tMeo7gJ~sWHVk)dyBya`jrd+qq3MpJ*Q90sZpgVV
zpO60O^7oyWKX2}Q=j8`SPF^^(`uwY(e7S#_R^jx1aW+dlqpX}(J?+|-o7pFb^6ccQ
zXl>Q^vS;e;W2wZh3B@NCoYZcmjO$%gP&lUN_}XqGvVNKMu_M@f*92yX<J`VI$0pud
z;3&GX?u&f&Y}fc5Kh=L$aNw`G{OW6~3Xi%fzwg?AXw@INJ!R3mX-iIwX&O?!`)J9E
z?FY8?D=+O>z>oZ{sLz{Yw}14*Kw|UI%=lY5&bQ|sn7eZ1l^wfp^8B*D6?>J*^K?&U
zW>xw2F+c5iclIy+8ehsda3)mW@J8v}tM3%ds;bY}akBXX*D`i*kC&&6-G1bA`la`y
z*Lw+t4I?`gPW-I;+~L8ekC$C|e`jGq^*CoBb53+^USPo5gX;z{bo1f`^5Dq@)BIb5
zy*Jb)33H2b8z!08s^(Pr8otUYom#!=vk{eAWX5t&^{P`#hd122dguCbr!o6VQO2L!
zoo>v%JhQq>&Edd+uTsfRs8!8*FICqp{b)dS<l<C0+t}2v{o;QOtWG`tNM#Rs^e1Ic
z_>NtgUH<6D2euykAao?>Y|(;*_g9S1|Mcfwe9GHGc~Y>leU^L2z2o|2?&mu{r>$Dr
z{<FQ`{8-ZdeeW;!&37j<)Y*lHI<sAtKec|`;zZH2W66}<L%aX}?Y-;G%ROtwVW(be
zcJ5ob;TPiPndy7xKQn9cSi)wD>lt2`yVdBGK#nW1m3EHn+@W=dJr7!c9Oojhl>BKF
zn`A4FO=qQN^=loh>G&VFTv}0}6(^8mW7FN^x+k{|)^zgTRo#EH=evI<kiOV-a$ItU
z*1?+YKYwnpqCEVoM4LZ0o!u^bSnFU-vo+GWP8T=+Hpw<7Hl2~4(Ytl9rXRh3Gy2xm
zXZ9wL<73lZ<GQwQ9jxi_>fU!%d)1-@azbpnPh1~Q>tIcP_jIp=oElXU$WgKBw79em
zt%Eh)w8eAe(^K6(OtL)@o6bzn?Atn6(?4|@_t*M4N3SN3qhr%4aVZ^I2W#3kZ`$<v
zk9BD~x@vsuQLXyu{+8XBhxAG;b$MMQtiLdpe6C_he%HjdM0;r<X&((F?WcicybO?x
zj{%bLGeELl7D(2|0?GPWAUQ7wB<JIR<oq0vyq5=(_whjTejZ4{D*!3@1Rw>!0Ho*@
zffRirkfL7%Qu0bbN<Ill$u9vZdu1SHpA4kzmw{Bg3XqCV0aEcRK&oC9NY$qTsrpqQ
zHLnJw=F@=G{2GwDR|it}=|Jj!9Z17#0BQIPAPv6(r0F$*G<_zJrr!k8)6J9d@b)bA
zWO;f+>g#b@$$CDJ@vYB=whMfC;qVC-WIZKfcR|Y|q78BdT;*+7x-01V4WtU!a-uEL
z0UA&`NCQe&(1<bt22du*0LoM_h_V3|P&UW{%2u$5asdudF317ORd9&%0Ul63$OFn(
z@Q4Zl0Z<_*04h`nh>8IbP%$V1DprVyN&yK_DJTIdRY-`+0U1y^C<7{2$cQQd1yCiZ
z0IF0dh^he<P&KFms#d6oY5@&UEvNyiRcMIn0Uc01r~|53=!hBt15hJq0BTeih?)Tt
zP%~%(YF3ztdJIpFbx|jldh$G@AdU7sX(j7>FE76J`wlL6r_0rLO}jXk2uXDrkQyo1
zZK^HPP8ymtM?;fVX=u`o6pc;B$v~6g7-%vo15KupVz9|NS!l8x3r$vKp~*H<EH*hO
z2ThLSpvkEmG`U8K!zS<Kp~-VRG<lVWCf`W$*c6-sGzCt8rl1PY6dEZ3o1#;MrpSrV
z6jc$LVk0GDQ*uhslsE~Rk}5${YNRA=%1#-YGABb*R%K|)jg*W{#i>A3;S^{pssc@=
zky5a!I#p<@oC-}<RiUXiQYtnzrv^=p)1axT8Z@;=O2ek^)S;<!Iy7}vho;_0>DV-!
z1~d)MfTp1u&@>t;1DmGPgr>=v&@@#Onr0(qV$(Cw!&o<Q>;~sNA|%<PS*gc;5A*E2
z>*<|hYqyZy^`aea)R3LV%!KR=W;SGJF>@h1hnWxAdCWq{E?^czb`i4_vP+odkX^>C
zgzO4tHDp&YYazRaSr6HD%tpv=U^YW`vn^W(gotllx?4{=Yv~?a7PPF}t=);GIfXgn
zt&X*TpBvx0aBqWLwS?<5eD}NR3D+geQNnctbC`6!N4A{;R}JYpiMft+oyQ!ZTxamv
zT1(LAqw5J8eRPzd(MN|#8hvyPNu!UhBWd)}5sF41Q%f-DW9kV8eN2>K(8q*H27OEo
z$)Jy^BN_BD5sE<{TT8I$W9tbPeQcCq(Z_~K7JY0D$)b;~BU$vZ5sF10S4(i{<LU_x
zeO#2_(8q;I4t-n=$)S&{BRTYO5sE_}UrX@l<Le0?eSDPQ(Z`2L9({Zb$)k_2BYE`k
z5sF8jP)i8t6Y2>8eL|EF&?kgR0ewOZDWFfNBL(yc5lTRxSWAfL6YB{PePWal(I<vU
z5q)9}DWXrTBSrLy5lTd#R7*(clj;cxeNvQ=&?kjS34KxxDWOlQBPH}n5lTXzTuaF4
zlj{i?eR7nL(I<yV8GUjMDWgxWBW3i-5lTj%QcEc4Q|bu?eM*#2(5Hk+1${~lsi04(
zBNg;15lTUyT1%+tQ|k#8eQK0Y(Wiz<6@6+AsiIG<BUSXN5lTg$R!eB;)9MKgeOi>z
z(5Hn-4SiY-si9A+BQ^AC5lTa!UQ6ic)9VQxeR`D8(Wi$=9esKYsiRM?BX#uY5lTm&
zQA-%;GwKNgeMXcp&}W261ARsfX`s)jBMtN!5z0WHSxcDcGwTTxeP)y}(PxH96Mbe4
zX`;`pBTe*~5z1`y4SV5x+r1waoTu!iIjK1X*0-g-G(R;T((u&b57XnuhB@in+TRn2
zrQOZ$xz_*NPZXsLxVApKU2M0zgS(^kNu>=hmc@P5>r%_&rHOK_#KlV!MPp`~C<Zgz
zM6sB;CW^z%H&Hxhp@|YOi%pb>S!$vr%yJVYV^*3d1+&^jshG7UO2e!-Q95R$i83&o
zO_bS|%X<_@4i8^CJ9d0L4o_F>_$(~Qt7ZJUk4h}f$j=yWtz{BRGxIaYTQ3F^OSAH`
z##^sHmMsad{P63`@5MeTqa>r$s=UycpE2mn+U~L4$(@qZt$z!5U0?A`$>rDA#di1c
z_9?IyF1r^UYUi7^%^Ta@^^vaW*7C@jw#T==z-(J4;nk7&|E`Y2QoK5jsAHzK%infA
z_PL$=cjl}&z8^F^J72%nT+nhEI9z=#O=2nSpdm311`_LFA#n~467S$42@V00=nx@E
z4hfR%kRd4!1(NDeA!!Z`lJ3wU84d%I=`bIrkDl;+JE`XvCt~Mwn0;6u>og~p=GpTg
zJ!XIGVfy;t&kbFl{Fe){LzY#NHN`s3SHyNdwC~tuO>DP2&E3=bq^qajKD@p1@lmne
znI)N%txLpu#T;83x4d7rPTY0J?%$Dv|I%`^ru9z~Ey`-emdq`$n=LwaVhi(2ER*%#
zIXS-deRCV+a=V7Y8FbT-7&ilnb+eE-HwTG#^N<9$07-O<kR-PRNp{PS6t@CNb*qpx
zw+2af>yQk$0m*cm4^!-Y`pl<WUarrdH)Vct%SNDO_|MOi)`rdYRNLmnw#`WoH>uLw
zlS(#U{VH@c?&*n7JsFr}YyH&!%uks$b!K^Nvtat;l)YVCdq;<a#H3Cg9~qIHpX*GH
zACZ&NCAD+cZswpN*>2CHL-{mja9@33&;PbJy+=l7uikz7^&jA(SwWN(&B#)_b8=o<
Q`eP#}lis3X!->NG17Q*4$^ZZW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp.483FB0E67DC82A12.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp.483FB0E67DC82A12.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1b744d0e9db8f251d32ee19c4a2aad35798d9eb7
GIT binary patch
literal 8376
zcmZvh3v?4@!iF<5rG&OpinL&9DbSKaCgV)T8Ro`JCO}h5frggSfJDMY3t7vhMGJUY
z_`RZlg5tlppj^Zo3d-tkToH9ev2F~yD@cTt%c3h_AY23$-Tl+<nfBdt{+`2k^3M0X
z^Kd4aIbXiiR{4Cd54PFH%&A(^blc*Y88(|OFa57&ar3<Ml+E@~w#~L=<403&sTrvC
zY4`m7#ZTSDr}Ij`Ax8ggYr!MK^B;eD_JMQL&c1c^$8}Hq>*t%=tCyQ4S@n<i>wDmy
zvl|xdtoGgPd*rj0rnk$Eu56xm>gsd7m-u3hGvdFTx%~Z+jhmQ{%kO^l@tRW~HJ@Mc
z(Z%l;)+}n?^Wu{5$&+VBj@TLdXQRjX$BF${m0Q}UvxiS#oTd7D%${-M(|7!oynNT9
zdt2HjvXd_LP2D!~?DFhsUl-0j{r0I_Uz;~uIzB*}zTn-9lMny-jc0cWuV-x}CNxZ%
zSG7J*9k=A8U%xu`W%!%O5hD;WKK-QltHSp#oGu*y{`CW1Zr=J+)|)k3TMlhMCPtcu
zuOG2)VM$TJ`5XVV^V`$wcxRtnb^jmUnDgK|e#P;Y(4O5d)IE0fyyxDsm4UnG)r={8
z;%LUOWuFu-{^`5j6?tVB2Ih~gDmyjmO#7+_K3;L0nAvM|%YLf;l0LW8u|kL)8u@+n
z1HC!tmCQXS?tg4(;GaG1+diBdev7@n<V^c*XTC|Tnlp3giI1OYYT8!5DbaH*O+Ei@
z@3R*+|I#bAPHcU&dRU+Q;mr;ECa-zc|Jt7Y`<|S>;_0$)rY(9cb#l;}A(;dJc-`e?
z_phBfuJ4eGH#hzMsm!fev)x-Z?L1LyY<#y|X;|i$Q^ya@t*R>9{=`EAm-pH9fjsQx
zH%8se)Gr=&&Oc+?!#|Xl?=K6i^SyO&>#@eAcfb777n47ppA%o$Si875Nn9G$bN8c5
zd#@RLAbY^_@7B%@H_o5)zZZu5zUEoGUn<mIsGGmMVNUVi-h967heg{rZaH=0nB5o3
zo_cWF=`ruWa3|BgZvU8tuPod%;>neBx!+y6y5-j0>x&moa@5QnI^jfc|M1eoKmK%a
z)}i|jT+r6Pa%{ugVV@OFoku$#$h<P=qlyO;D^I=oO8xH!&DnSR^Sy>G^FA3+1{QCv
zeS5^}_U*$KHjiKA8}Hq}f7|_eJ14(-OgNZ)Zq?DoT}K)}oPS~B=fb5!?jxr3@c0`_
zWXBT)6Q1w!rFd!elr29u&DeEu^srag+YY^QZu_}|u9D}z|JRvE>mNG!&}+FHntMO`
zNcEh?MP<)8wjKWMms2M_nY(S<7esKxnCgS}wV&O%v`2PP|9|hRKYx2;Ms=$3rB!!~
z{DWG(>#sc~9&8luYR{dh{x*2h^`gG%fxqt96xsL2US`wkiVcaw6RzlEmcP4U?SZl4
zhVL6@^QUua93QqkS#?F*du#Jkv)7bAwfExcs)`4no3!$+d*0dd>yjfcoIN>kS?*Zv
z2j-Xldz?q@@10>kxM$^#+7D+O|6=wt3z~8kT=>(~^W!dRFWskKdwkl@%FpG8e<|6~
zxcTV(&FczAAJ0A*&Rwx8cTsu#^PN|MTic7DJ@N2R$HKQtFXi;Fzfy9okMky5%hIM<
z^@PnvMa=v^63mora@tf|s4TOrqH{<;89M(mOCoE=Y>@_MlT*^u{WAOY>Kv@;te-cx
z-(M}(Ws~*k>B7vyyw1Uz&fB%J@XnoQKFlWl>FK=8yk4DyHO=pCJX%_|Uc1gVIXyk7
zU{G1-U`>0jT}Zz4?b2P@WJ7v-K<0p+or5)f&%dV>>}B7*;W}Ggdb+rvcxdNfO|RJT
zz;9lTUfz{s3#6wjdQ?<*4%YPc!~B*||KYjU*=p0%MFmAeItOd|%*?@mW`yKT*<?+6
zIzKbNSLa|&|JrBD?!<z9-)57O($jr2`}XP_tZ7^GZMV0~FYLa$8CjiIb@gAaZGQEH
zHaI6#U0OY>>wg&Xm-->qm)3VB>-w8no!7nl1Wuo`%6ji|LQcQ44pP8*OLwAuG?28P
z29gfYK)QS`AYFbJkgk9WNXEwi$@m!{nE(SM>tlgr{Vb4dfCZBCaX@l@4oEJ*0qOR+
zfpq)bK)M5NAbB4TB=6^e<O4jAf=>Wa@C!f+0Rc$SCju$@MIgn12&CkbfRy|akWxSb
zQufI}%6=J0IUob6_!J-&zXGHZP=Hi@Dv+vQ1yT*DKx#e>NX@SSsRcA3b)OET?$?3T
z13Hj~&j8Z!8$cQX14xh01Ek0A0n!uj0BQP6AWgpsq!}=Q^!mI&di`D?y#X(fl{Zut
zb>E1UgDQtWDy?)_$$HYIZ^Vwr9NdUO=iKg_%^7yy38~fjKSXz=gEXLYm<E(?r4e-n
zU4Xj6E<jzaE<~9i11J+_0A*SkMA;w<C>v%0Wm{Q9xgZB97v=!vS~*1BK{uf8up3Z!
zs~b^1$OFoUc|iGA9#J7E04jt9K!sKTQ86e2DuzWs#a0nfDJTIdg(X0xRtZr#C<7{o
zWkBUt8Bryu0IGx)K$TVnQ8lOns)kiS)m9ZzEvNyig*8C6Rt-@-r~|5pbwKr29Z@4_
z0BVE{K#f)dQBTkVs3+_J)YIxg)C`({nqd=Av(-e@8}tI|4SNCgwt5k*lqyGCH(*Yv
za%^P{q)C<2tz@k;CTDf7FW}~MI7f8fe@@29K@yx2q!i`6M0ID{K|_;fXlT*`4NW>l
z(b#l3T+nneE@-+07c^Zd%7snF!9bH?7-%vA15GAHG1z1sEHqh$g(fSo&}35-i%rhK
zL6c)RXmSDvO)f=o*mOJG&~!6yXu1VAG~Fr6jZNOcLz8ECXz~INO+H2O*c2QBGzCV0
zrXUE=6jGFcP0=AjQ)EPFih>ADF-3{klpGQ?B}Rg#BuLPdQj~;E*&#zyW@Ko}f(%VL
zMakGy911iQMuDaxD9}_=l!8swp+Zw-RA{P#3QaXdso2yU8Z<RVgQg~E(9}|thE3g}
zLsMsTXzGFvO+7{F*fbmlG!4dprXd*6G*Xm-O^?F^O%LONrbqBV)03h+*fbp`G)=~Y
zrYV@vG*gs`O|QcXO)ukxrdRMn)0?8a*sL60Nm~!A^n;A4<RI~tqLl_-Gw5Z_XA7vD
z(1cMF4C`=CsCHB>q=r!qUCH_qMP_w=dFp=lO(~z!bzbYM)M_HD^E*^G<%|-}&tQ$>
zOc2g?%t^v|4s(QbzD;&NC!KAi^Ha=m()l&!809>G&($bFqmNDyH2UZyL8Fh3kTm+}
zHj+jk9Vcn@(J_ifpDRkZ(C11JF7&yQgbRJH2<bwftBrJ_&lM+K=ySy=7y6he!Jv;x
z5Dfa5B*CDMiI5EXm^P9@9}_1T^f57tK_43>SoE<8f<+&jBv|yZ5t2n8+eWhJW8)-?
zJ~l?N=;NXUhdwSraOmTb1cyE@LUQQi+DHz4T%6?4$HgcPeeNjXMxQ%DxY6fM5^nUl
zBcvOB?l#hmK6jjSqt6|q-00(@1dl#GLGb9~lLU`GK0@;7<J(9ceSDnc(Z|Oq9(_WT
z5YQ(i2myUUk`T}*L`VUBLK`WdPl%HO`h*xIpihhvBKpJxA)-%A5+eG<2q~gZY$HYV
ziE&aypBSS=^hr@dLZ6f%B=kv1LPDPuAtm%lZKQ-gDNai0lVX&FJ~>Lr=#vwKj6OL@
z$mo+Jq>Mhfjg-+R$4MD|a*UGEr$h+_eM*8*(5EB`1$|0{RM4likqY{hIH{mdiBSsr
z)F`2%PfZXi`qU($qEC&GD*DtmQbnH{Csp*RF-k?B79}+FX$e9@pOz#v^l1@NL!Z`0
zYUtDAq=r5%Mrr8NqlAt=JwfQ`)02dbK0QL}=+oOs9esM7)X}HMC>?!9lrYd|BnSh2
zMv^elXGBN?eMTE;pwEbt2KtN`WuVU!B|PZ!BnS`sJW0ZXK2L=7pwH7rdeG;IlOFVW
zVw4AcW|T0|XC?>}eP)s{(Pu_T6MbeIX`;`JlP3Di7-gc*8zsEx^Ck!{`n*ZPi#~6J
z^rFw(MtafbjgwyVd1I8f+vk1wd)rT!7k^D<{MU|Z#((XoW`uA@b%L~RK>@Ss2_h#{
zC>M^g_H`yt>|6fb+EG2yhkd<#xz@@nbn=nL%wvPwI~HX01l2m6UXanXkQ(TQt~86U
zdguC?dn4(}M0KLKyHcmtK?<mWt~7UGUF_YP@0*{NSK2EFSd~Q?whb?H*FNt!6FpJ3
z4(EjG=sHLNI?$EwsvaAq10((EN>N2ogSE?@6DqDKZm@RDb3%hE1~pg<N9$;M0ok#M
z-n}%OR6gk@%ems-$O9i7Kk;_@e^@lNC}f@2noch(J9gNu?!&2yBIUrU!Ze-NC$GTz
zdTzDSrj8AFi_Qs+FCWiar`mD${-sk-Y+RG>mQ5(DwN{$1E_tViziF#4eRx3c0R`5A
zcCS!7r$@!sv+2YAhV^5t?eh0eE^8V2-H+A%(p%$pXQ_3PoDglNA-U`>NQ|9<#M)U%
zoSlQ@w!0zmb{>*o7a)mt5t3w=Ajx(al44gNsdg2TX4fF;b{&#oHz0ZJ9!RF$gygk*
zyV74K8>eTKe6>IQIT@ERZm4zfb3!#4HIOD}Ozuj@|FPU$oA>&W^rap&b<k|<<MR#I
z<8sg3{*R>(`|^Dy)|spyP|}}_j$co#_xh~;)yGNo^^UtQe^@?eQRk$RXme=uzVzYZ
zsl_v`kEHc8OMX`8&nwnR(r>EnG~rFvX-F>J1&PraNUY96;&cv@TX#d^bsmzS3y?%z
zge2(_Bw3dsDY^nl)m2EEu0hgu9g?9NkUY8vlBt`Jyt=n5HQ#o}%#NR+TAF9K)OEZF
zb`1adn{K_Z*>3HAc}v$zD)j!2tMAo+GW6cenYY|JJvhTQqifFAdDj2L*-dlqn45lC
zeftC|qeo^>d#~)A>-yyOuF9(z<H*aZ8r`@5_5B9iAS>D^-SC(<=F@8F$l)db-_e4B
nMa6@M3@sa0KEml@IrmMXG)O39__6r~mE&fSzKPzOiShpjhE{Td

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp.04904E13881A008D.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp.04904E13881A008D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4dd2a49c9613617371a2c75d66213f2a171c5bec
GIT binary patch
literal 7414
zcmZvg33L->yT>y#X$f7VEu|DFZ4sD(Vz>on-?8mL*VLA!rGyd-l?M3~D1?H_2SUp#
z%LgJKprxW;r9}ZR7!*7#ULX>JT)Ea%!locwAP7|u)Gy6FL!Nui%{hHfe)By4`JKMA
z<}Jz9v^yyVL)Mhsc{8Tg+KdK+AwKvwyLMJh05cfgjy4$PZ9O!uqVxwo?$`G?%eK{~
zS7)`m1BONaxjE@%F)z+fteG+WxqNy3*0q2B{^>7E)oE$HJ(>St+yB_%vn-gM@OR_k
z>dpHnqK+o^)Tbvbqr^q0yloF|UVC!ohq3dArCEMWOshP3|LTSMN5A|u&oYejl^!Jb
zI1j8@vE@|D%Cz@t?)~_9?B@r+*!HiVPyLtooy3~or|v1(bkTG7=g%B%Hy%+nZ$!Md
z`7}TINK{3?A>yX<EmVs?()x95Tg4}0_x(F}{XArK&2mp;b&FQx+7SQwx_O5m9kbZ?
z%{o8Bk@Po9!n{ipzWvAC@zc;<-;Cb4=-QBT@(NFR&xLK*_ir!8*Bq^Sb;H31_vb54
zWPROiNS&iiud?N=AC3h7#ZOvqoV#+Y!IIW5VR|=zO{3;2EN*IO?YHU7GZhIhEWX&B
z9<}zJd)jZ+1HOAEqMI>4@t;EK^{sbD{!kp(ICuZv%dQkT;Y4G`j_$HkGv0b)IP1So
zZ*1Rmq+zF{*L3dSvzxa>I`>ppc#mhj<Y<^SH7?`MpfusCr8eoi^Mi+OOdqrMr1ipA
zwGWnLT$LhHhcBJ*{HlFdCY(JYZT&d;<?p}y{#LAS)$<8mkBr#2t14N&db~sxtfxOX
zzw1!Lsv**#l=-#O_wE?5t|a?dtU7J$_62#(gQRS8b**@(xUy{2qXyStWAvV}73N2;
z_xP-G;J#rW?CFE`xl?e^I%>gT@l4#f!ne1-dHjA#m&SY9<l7^dy(=Hao$e{1<Mw^H
z;>LzPKMicny8U+Yjz1?2s~%AI<g=0eo0qiOcRd*2*fhs{Vg0{yUd?bE9NM(}_%~DA
zFIvBxGDP^PY1z9cvu_`5c<Tjp%tP<2Vbk9m|6tWeH($cbmmi88G9xGVo5QK&CT<wB
zExV{CZ|SG+8(&7}3_2OTd57D3JU(NiP_^1I(^OE@t7DnV`>?i_Y+9H1NLX3-m{@e;
zr`>h5JL#@*rSyu`KK=IZnPZQ4DMF^y`F5wYWOVEsxBW>~WtTczLR;SC_MNF8-1(mx
z`d#t$mf_UY37c!BH};N5n6oW&Cx5NA`O@99uItNoZLlrfxpZ`O+rh6mYWc#;jm)1@
zmY{znq5ZeGTZ^u=%q;5rU~|F5Ukel3+L+2s=MK!wa9l|@f7Ea{`lT%gN8h=PKYBNA
zZR+T+h^DPYzhUyWxz|4Y)^@S>pHG%Iq{Q+A-#x&+Q>oRxGkor)FE%V+FTcO|mp|ra
zJ+|3T?fK&S;*Epa>%KX9_tMz2tq-lSE6(*Kv#Bq!HL2geHhoJVY2ku}Q<^T0?0Ec9
zp!Ltn_y6j$<+nb0*x9*skLP+Wp0s=1f|yMC!<wmC!p--N-fz1X@p`gtu_oOsoVH}(
zU&zcwL-S|X&3LgEF&HA#oNr}qXVRn5apeYEVq{|1&=9;cgg#NfH(q!%pDK+;#|F#2
zBYVe(2E82r$@1QFKE8P@8Z8c%<0Ip{h6cTyF>_1w<!f=yXtW|&Hb<J}(4d#4^F<Y=
z%)Di>hLT`8B`Kv}Xwb{b<9q(*YjvMQqvM0+<jCYMp+PTaxAZ?2`C9X)Si_iLxo=Y6
z{-HrHuRSC_p7C+is8~a3u$-EdIv_OY<#T-}y*UWEfyWrig5~Up?2(~CFW<DKe{LrI
zBcf4Du-qfEN7vAxm!Gt6e3W%zXl*n)Ay`h2OwS4pdijsIaeIBQe041veIZy*h)n1j
z8uaq4snxSrCxz#}UsPz$S08=c*m_w=kFntzo~tW@REAfCiO>j8BgzO+BQ-KtpYs?S
zSw>cbiP9)gqsk~yqcs}T=rS7A7>xlnri=kKR%1bpEn`8A(>PG$$~aKtH6GOXG9J_f
zO#n5aOaL`e6G2TZ6G2VVBv6yeBv6wz8Pw!58PpU_0X3ye0X0=qK}{`FK~3wWr93@}
zY+8Q}*Ga!-hDC+06>wOs__XkW#O?STKn{E{5=O#GfL6jzfL6jmz*f>qP75Dq(oW6+
z<RBL#VWg}SXr=5FXr&w!Y^AL<Xr=8mXr&!AY-OwrXl3jSXk{D>Y-O!1Xl3mzXk{HN
zY~`#RXyxo2XyqInY~`&yXyxrZXyqL|Y!$2mXcg=NXcZg+Y!$5{Xcg@uXcZkIY?Z7M
zXqD^|Xq6liY?ZAtXqD|UXq6o@Y*nlZXjSYAXjL2vY*no)XjSbhXjL65Y}FX;dHqa{
zv1!?w1<(uHM4j|oyBZa`rGv8=#|MW`RNRbb0wQq+Pyoa4V__z60wf7D0g?nsfFu#X
z2$&>sa&Y)GCe36fAd+MN1u*hH7G?^kK$0?3AW4xFNKyfef=L>uL6SDpAW4%nNYVj}
zhDiozK$0;tAjyynNHPJ8fk_r;L6SAIAjy&}NU{Npg-H(QK$0_aAjy#&NOA#;gGnCe
zL6SG~Ajy+FNb&)Uhe-h!KvFOZASsXnND2W=fJqS-K~gk}ASsd}NQwbWgh>gPKvFVG
zASsa&NJ;@rf=L;dK~gr$ASshFNXh|BhDiliKvFR)AgPcFNGbtLfk_otK~goVAgPin
zNU8x$g-LCYHblP^gO{mU%LFuBqjc)^*@Nc!@uh1Jqb;w0-ei5VegPPxLT_B*tGBRw
zVZPoJV=L}n45+kw=~L3*&4N4Vd@&2}0Ix^zb?`}5j}VXx3kt{U0{son9u<0X13Pge
z_-U{c+=t*7Ao~&g4rC{aA3(!*65NI2UqkkycspbdhM$3V9ydZjAK^m?=p+0H0eyrM
zC7_RRp#=01UX*}7!h;deN4gR6G<fimK7_mg*^iKSAUjd=0IDCf;H#SpCBKI3Mag!^
z9*jH#?{aR0f<DTJP|!#D5eoV!CrUvd<w7awqr4~ueUt~IppSMVH1yFvgoZxakI>LZ
zJ5d_?XctODAMHhH=%YOt4SkFoVW5xkAq?~}euRNO#)&e}$GA`i`WP?DKp*4580cf&
z2n&6z4`HE?^&>3wu}+kQKGub@(8qdF7W!BZ#zG(GMmXr>d<X}9oFCz!k8`3N^l>hf
zgFeoSa?r<lFb?{7H^M_7??ZU#<NXK^eY_Lpp^tZ=JoNEil!rdvgYnQOxDf&R1Ro+m
zpWsIX=o6f%0DXcB6`)V>q5|{@9!!8f(T#}EC;AW(`b0k>LZ9eFMd%Y<s0e+c7ZssT
z^k5?NNp3`fKFNnj&?os33Hl@_DnXy*LM7;vyr=|yk_VHZPj(|R^vOO%hCbPk$j~P{
zQ5pJV7b-)a>_uhhlRcOWeTo}Vpil833iK&{M1elVi7L>ixKIW96fdekpW?w3=u_Q@
z3Vo^%QK3)uBP#T%PE>_H)rG3ir+QHp`cx05hJ7O&ZW|swsBOoLHu!in{-4LA(FPxn
z`r-o`6<Tx%KiP^4ip%xKUtal#-*#*M?*t^qmY$zptUtaAnf0$EnOCh1Ztl(Y&eDG_
zjwwnYuCEg#f}3-?=5*6{GWp7~mm)8wU+5NG2<X7@B}RQWF*c$DBOsF<7`en4eu_{X
z7zLT`z-Y)!2gX2VJ1`b9*MV`6`3{VSEOcN3WU&JiAxj;Y1X=FDWXMVfra)FZFg486
z`jmPOjeg^A!THckTCzSLbMv#^L(1r4EXJ0SlQLd^Pl>Vh&FMQ{KP)k})ST4u`V!Eh
z*PTm_T)KQ9xTlozl*xLeA&`^8omtg8xH&#9K1u(<v--!a7W2aLshtyEZ^W0u39mO2
zkjZ)@xeQKty^(@U*BfcbOudnT%+?!O$XvaVgUr_(dB{S&QGhJg8%4-cy-|WJ*BfQX
zO1)8mtkxUVFzeG5Jf}OC%fjb$X+h}(efnE&y;)a&dFxxjNi%EaeEno=o%cmV@r))d
zxH<Wm<RtyY?@O{|N<>c6t>EUwfr(~)vGMSg`LpTk5A!>x*o5~7XS9g`M4HI{x**0z
znJ7TCi3Y@&7(lFv1;m*+K)i_uB$xz1qDcfKnIu56Nd}~t6hNv;eM(2i$Tup}+D;f_
zY`MnVXZ2}|u@x8#09lL{KqbbKr=%y4sL+C>eiSbK_;+<({O%uuGnrbRTBRSWtC<lG
zx^(NgGPpUXM^2g^zIOVfLm#|WXbEoaTi*9Y{j$(kHRGc~tC{+~gGa%Fmw=;SApnsU
zvIHCj3k8U_(0~{V1BkV-fH(^Wh_~>71d9Mjw1|Ktiv&ov$bb}!0!X!}fV5PtpFSJ6
zOs#8Bes|{$x5vO9nfff~|8Vep?_9wR&j$P<r7z{iMTM4d!%et^tNI>-w_=F^L`tLz
zrc9y$(Gm@akr+U%!~)_Z4iGQ#fCNbZBuXM6Ns<7`k_<?Z6hNw^J|)AfsW00)SEy&t
zs+wKe`DIe)@V_6c{>5OZ41b*%{yO>TSImaywzOTZ-rpMe;>1epq{)UCpVkajGp4*e
zEr|Kmg_toSvWux}bWChqw`X$WbF%RGsNClhdUj7t?!|I~D9MUCw8s#>KRqbz|J$0>
pC#7$Czh?&wBnD?Z_aB+VMo>&D+1orkyGPQfyed>H8aWy%_%}bVJIMe5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp.BF7C969BEF21A812.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp.BF7C969BEF21A812.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7ccdbc93cc01c32d4ab0a95e1c441b4468156529
GIT binary patch
literal 7058
zcmZvg33QWXqK5M?lytMSw538fTF8)pUs)Q$QfQ<U8lbeaKp?V6r-BQuH+V%I1giup
zh{{-mD>8#4!T{I7QbBMTmMIfbPy`Bwbzq7R4NF;6<fggjC+|Jy_MCnvPu};<d;Txq
z_b2%#6gwQ=UMACsS;Y(H&8~Z#HknMx;s4sY`M(cQCez}0lWD=){S{NnP8W2${r({5
zx(4$z1Lr+)!?b9}nD|-!va;>o9ZmY`zB$GIjbHq>Wnppdj-*#hTgQyr@k`b2v(v_s
zD@`Z4Z&#O`D^2}&@acsEFT_su5BbZao&WAJed{;N?%lmM?D}twgQsT>mZq=G%<1PZ
zO-gMYJN4|4k^6&}zT3ZZMcv)o9gOSyvxmRExxb@sj90r~w`Ww3>HeI{o;w4s+)MoZ
zOZ37uhYQZFNoy*sSRi~hiLE%T?|Qq})}7Oyx6Cb{()!1q-=?e^et31mH><4qb@|uR
z?o5BVZ1cv8Pwn3MjAzxC`6t%wy>caP?E3CKHz`LJE<N?w(Lq-){1g}HpL*zD<4@9$
z?SH|yW%Gj1|GYfz-1s?l{nJ<Xv3CE%e~q~Kj=pa4=P%qjZyz}$xRFXc*6*{}(YIFW
z|GLpKyCiRP=*hH^M2dQ*{mk-xr#?8h$i1|%_8=9wX|L`znmp#O=zq<3zJ6d#dGA2z
zD?P9B-(6qd@TY3#xqge2%70RB&tEWH8IbtQ%)?EeG`!~jG5hGD+=<_hx8FYWQ#-e%
zxoj#k{G;4AcYnV0QhRLqp%+FEo)!1$l6Rln|LKNhT?f9JG56&0lUEaq9M+{v&MiH4
zTuxj+f2Q?=bz3h___6iOhy(9rr#@oel~y%j)hO$R^D()af85#>hDjQ~EG=t=?O2cF
z<L{rpmio%813&)j>R<Q%<Mayesdv9Q`RO9vV{x1~nk}`eOJ=_{yM6uj!3T-z!O0Cl
zU&A$J$gG=f-HFYU%MR8Z(r=_x^e!zKGa+Y^@6qDy2lI}YYkF^<%)V$owqxu5DL3lO
zrI+3~QKTKsi2Lw<;6zE&CnE%J`_eJ{tP8JY^c~jlKEHKN@J`x`2Pbsc-rBOSYVpRV
z4SlNJNhQMQA4>E7Qoix)Tm6XlhGoUwD70*v-}LN~(U(5la$OJ>{jJQQ&Y0hSNLE%u
z^~MSJKm6~f9`xB>JZ)!dec-}6_p_J(S~9O8kon=EyI)!taoc*Wo>{(e?{VhsFRooJ
zpZ4y)cT(PM`nciPo`S;%=Is9B<1u3!D=e;tmeOH6FSFlQWiH)Sn$eL{TQM<or+rr4
zzSW1e7k*dR=Wo5HU08cy|Ipg+cDFtHLCUY$O#XYmEu|N(&Mz(Ncyqby_TuiX19raO
zLp%Izndd?-U%fRaZq9q3kNPoKaBoewm3tdkIY$m^%v?QFxsu?t{Jr+A;v1Xy9(S&t
z_4c{gncEh+l=0upnK-WZ*{9m2<4qTLovc4F;^-R-;~x;)yvM(4T~qSsz3c6pt|x`s
z<-<FhVxPY534Jm8mhGdLhTL_Yy4Uw-+sR*ZUYN7~`>pT4a;fmh>0>o3hyw?AtejhY
z=VxL3%URQpyq0XK-q!u*lBS6_a?02AYp#8M-i&g>WQxgYeq+Q319IcZiYk*kH6}GS
zGKB98k-xZ~*PmHjqL;;!<zcyJOwZ)VV93eamiByZ>kogAC&z{5<e22x$Y97rf4Ls~
z=<@U1;>k&2IU^>cYh*BF$9eUw4KttoF40sLmb25d`$PsqUiR{=ugcD*?N2mK2+Nu2
znY|-}A#Y#$r+?-@kvT7(oE(<(V)BMZ217o7#ku3>R8LicsWL1N>N2P#G8pm?kLT{<
zwV)bLI>T~WOj>MYFy#B)Dn1E3z3*~7IX*0>#H7SV219<3e$BIY)rhHyrio!WD?KYe
zG8nRH{_H1eN2Nt)H!ChOs|{cOvVO~XTW*5e>aZ3Y7Z9dO>y#)l4hA%)k^zl%u%NM(
zENGmA1C6WXK;s=eXnZ9Pn&1#X6DkGJM285PSSf-gIV8}eN(nUCA%iAY%AhF@1vI5n
z0ZnzNpsAHAXqrO<O{>&E(;Yf!dZiAU?NM9i!`0xn<=ZR<8LOL(i>zvRdtBD)=o@0K
zu|5anvA#k?k#R9#Woj6(G9Cu4tcwLJTf>5t^{{B=TpU=r8V;<SheIpx;=#(-@L=UV
zJX!^p09K(!0IT2;&?>q_u!=PzSVfPBR>>uSRjQG|DtRQd$}SnKa*YgD*(0M>aVcO`
zY80?49tEwcO9iW1qk>iSsA$z(8d$X&4Xm0+L#yu6!K&BjVAVZ3T5SVtg~pmnaN9=O
zoIvAkQw=h1mjmM>w@FxU7OOS7q^!JE0@ADlfI^h@CKYAI!a$PY8Avi314$-CF_>g6
zEF@W;g(R!7kYqy?i%HJHL6YM+NOBqnNiIZjnB*-yBzc~PB(L$1<U<sXNx>pOQs4zh
z3Yq{(Aw&t76fGhoMP7uYsELpiLzIX~$s$2g;w4B*ngmHHL`j&GEixo!UWTNs$&i#o
zl#EHmqCisN6-X+Y0!bxADVS6(DkN22g`}#fkW@pIib>6)K~m#2NNSn}Ni9TanA9yg
zBz0bgq^{|Z)I*exNn1Y~Yix_~mf>v@kYdvfGQL7gjf;GBfGy)AtcT$~Yz+|BGsr>0
z`Xh2PY5k0hZX9b1X+4D8Mp{oH`zdQHZZsdkU>_4880=$$1cQA{Gs$2d(?T-X$Fz|Q
z_A!2n!9LbUu-L~22p0R;Ai-iE+f1_9$F`6x_OWdwi+!x0VzH0&5ghh$0fNImE=X|L
z$2F52_HivFhkaZd$zdPor#S56eFTqve1PDwj}H<&_VLXmk9~X#$zvbiM)KIl`zap#
z1Ro(_pAaAf>=S~7fPF$UDPW(_LJHU?w2=b#34ThzKG8>r*e3=E5&Og-A!47{Op4ei
zwvZzBiEX5aeWITdu}|_5681>}Lc%^NNJ!WxHIowdNiC#=eNr1KVV~ruB<z!Ygp7T1
zfRM3I4iYl<$<3sUeR2yaW1rkc%Gf9SDH;0|AE98M5+D@pQ-XwoeM&Q_V4u=LD%hvA
zkqY)HeoDbU)kmn<rv?ZW`_v$zVxQVfs@SKtkSg}6ZKR5Qs-IG^PxBEP_Gtk^!#*uY
zXxOJUlN$DEEu@BhS{tchpXR4D?9+XOj(vK7(6LVs5<2$j&7_WfdJCyzpWa65*r)p`
zJ=!<?wQHt(x9Uz&w0mgxp{2&%opu*?F9a&;Ui6TRUx#E|<X2(z&R^s#av3i$cEGBq
z)A{8qi3E4gfjvhUU+5>6ru?q`&B0y5PZz`%BpGh^h1ciAoXtIx82-uCL0Om3(T@YB
zgJO``4vIzQIw%gA@1S^Op@R~T#SThDmO3a2S?-`@WTk^rkkt-KMb<hf4O#D?^e9*N
zD)a9d_0m(}>$RC}8OHTKOPd-uNxvNFyk0MD{Uctlmu8SzFU=xzUYbMZy)=(3cxeGy
z^wJ`-<fSEK*-OjFikDW9RWGd~YhGGI*1fbI<^L!d>0>MVkE0UYnMIkCjFW-{cUDo>
zB;&_<f;+n?dy=t?7QNWK<g0V%KMPM$W>u!!h`bgm${f_Xyl437<Zj97#y50B4llmW
zgNNwi8|oZ7*_hsji))?y%T=>Gr?;NAzK+vdPcz7DJ<TF>^)!df*V8<*P)`fUVm&P)
zOZBvbEZ5UAvQkeg$Z9>UB5U=uhOF1qdXzU`UHp92`L%C^Cnc{quhiHjTNdo@GH%`m
zNBHTCM>5iluV#i!$3+fcV6I@NC;XqCp5VrvzE_(tt4q;_tKrv8?VHLQM^ay2_|Mw@
z?RQH$r`K%FH);~xjF|yq%`6aS=74xJ4<wicAki!WNoEO1Hp@VYSpibbDv)N@fONC|
zkiMQc<jVA%69>a<bO=2p&sa1G?&0)sppo>*hji}ZABMe|y!~`|BC@NpXBtb_h!Sy;
zBP?TrFG*c)bxrEAEbJ~wE66cimk)2+vvJ{AXZY!?s;n8tZZHnKlHwu<UB-)q=hkVR
z0CVePfLJFB#5p-2-pK<AP60@Cia?T60+O9Fkm6K;RHq80IW-{NsRP-vZM}^vzG~jG
zf-KnH`5}<jH%&6G)HuU?BrbAx2Ul#j4udPUGeE4J1>)=+5O3#!1iJtv+C?DAE&<7Q
z8A!1!K&o8@((D?LZr2~uZ+lK(nAks9(YYBKY3l~O*EG@$GTTVA$Xp}MA@hwik1RCO
z0<zdhi^x(VEg{Q|w2Z7Y(h9QLNUO+NBdsCpjkF$RV||A2w4LXu(RZoPS@`haWd7{A
zk9Quf*3O?<Th{r&)j34|DEFWTpX`e_?e)eyK4sce*L2fVLo%7BMITOyKAawX7<twI
z#hLSF&8-eUEO}xqMR$qmYL1OhNbHvMNO5w}2upHYabZf2?x`7%4jg0~Y#%aon2^R6
w<oEBF^Zz}Y-YYXJw|8EjzQ40FoG8hvJ|J7`$txq%(jOZ=gLIS*A4QD$AGZa6$N&HU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp.11676556038472A5.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp.11676556038472A5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..50edc4f7f25a90def58f072e66e103bcd3d50e15
GIT binary patch
literal 8382
zcmZvh3v|?U!o?>El%Xw;0!1jL6et5sCQcF(@*<f;=+yEY+Cm3#WOxY{L4oRsfauDK
z2oymDAG}>zKu{D`2N6F7T^Jk()K$L`hZPW1U~pU!Wm()Ex+m@Z&iQ%{|C5`0Z+_vQ
zNt*trs?zU&v((`jJhk${S<~t#wQx8b1=)Y|>gV2lm2x;<&2u;&T>s(N3Dqan_E$Ik
zP+d_#9J_nK)2*D9AMG<A|Mc|emu{aqVam9FeA1`qjbEmp8}^9VW9ai6yB_)Ca^Li#
z+L$xue0~0_%(^Ao=#k6+_4{YU@uSx+(JxNiuyN6gd-s0$+tgJ*{nYwk@bavnXI0x@
zyYxM>V8x$%-TLb6j%P0{eE90w0k=N9aMRv%AOG@)4Yzu?O_|DGJvTYF_U|L_z2!@3
z-K8rX6fZIHuDxxen_u4j=-Q3LE`R>Qzz3*?IfsAwf*$q9-Q6x;8%n+MUe^=#h0m_7
z>AK~mCAI4p->BR@w&6hfu_YH@TJiOq-9N27Fq0)}!+$!q>^yzKSe+QR?%kR9oea3E
z%gz|`@Kd4Lp^rc7IjyF&q`qP1lvr;6Y2^Eh!tcDXZOZarfAlRXi|x63qGYL1y1(_%
zD^H2r|MAJB<z=^@k=#$Lx&7$;za8w-dL{Kfb*RIV^9`*AeYfC*KcBpN{q9-+yX>(A
z`?&2znI-LReKNeEN9mQe2X}mKj99pR!LvKo9v;v;uVZWJi+g5v(|`EZC>`ImRE#{g
zs@3^X_pRNvzxodMqub3j2YPL*zmoI$u?6{qVudp*t9BKBncMF2>F+1+nRL=TE`8ti
z@v^Q%N?xD!^dYlmk|$^1S@DizOaJm~*>^+R<+m#Nx~Nb8W&Zxc;Wp+IiOrkV{O!3C
zftgl6=yG7<_@&qS_B~P-Tr>VJ=eC}jK4(Gy9iP<vb#{L8zUgZs8<N!d`E92b{#;zV
zY~wAg<CB-R>{!2lb?AeoGmK=x=Gns7;YYr9ymnOm)9gy?!i}Na;A4mXr(nf<9p2da
z?577FDZe=W9kJw=_jernY1{B#TdnC|{W$pDsl&8ywhn(GFyz9a1BtCG!l%FcX!*G7
z8_#Ml@7(l!zdpYdjk}w6FUh$Wdvba1D;-yUwsY*Hhn^~UakcB2?JdhEpTDR6iJ|AV
z2@kD%{Jj|WVF}l=dyiFZ{c;tz__Jr;TFlSgba3{@HhVwRtm@+vj@(0qdhpX0e$&Mq
zFRbny_%bey|Ko+X?msv*@7X<L$>2wCuX}s?-h~U_xc+(JfWlYWOjtE~q~opAGsDM}
ze!Uu=-f(I4j2Bi+?o&0re1x&0!QZO><$PvL?;!&Zd|T+*Gv{9&Up{zfL;K->ozP8u
zBN@No32n){wC{MotMA;qZQt*2fAIUry;{M#T5ayK-49PtyDjGry#C^)1&=!x3_iKg
zwKF!=SNnA3pC|KOua0mff1I(Xa&*I-O@V(NyZ+3Rxtm5WyLMxqb^Z4@`y{VPHLfQf
zaP7-XJhAh>JO8!we2bs^w5@8h`2B{G&lb*{@>|D63l`tC=G3FVT#RM5M>oB6@1&3a
ztv#1>_T;siPj7R+^Wg5Q?T`M}zVF&L>B2pE$peQ6EbKR#e9dvC=V5N`jaCiss{fuh
zf7awNgu{{3Bib-{i&UCNjveO+cgpG1x^c)p9~yt0^qghS81*yTIBK$XN#~NX#=*98
z|GD|pW5dMiJaSCdF3Kq?XdG<2;I(B%_rH4Ka2^@R+66fUts4j1_Fexez5V?B*Ye2P
ztlcH2ORL7gwjbTRrARsV?tN_>)mgiH=kC242itC_&%6G~uBAuw$njabET^ozaj@;R
z`%jdt9KWY+8%I^vF7901vvIKPZ_jM``T3)7zLoC?X6^Ep<zpHL+rBWV^le5+56B}&
zW$nV8!q$z0ZQp1=_KnoNADqu4M`!I0IUQOz4z}%>JMFG{vx}N%m(FdR)s=^?Z{BrU
zE6oqP{q9QpKaf>H_k?Ce`)Tm#APt_!?*Y#f^nl0s8St1O1D@CK1<xDwg2(z<@Yo;=
z9_Q!4<ANM`yq^b;5Axs%egQloD1ayWMexL+2%hAZz>|U!c(PvxPY%l9DSib!C8&U>
z`c?4MpbDPm*TB<)8hE;22Tu>`;2C}cJR@j;=kxo(^96n2nSK*IGiZWm`7Q9Qpaot<
z*NWn%k8HT2cZJJV_8m&*Hr|=dAF;80$BK5yzE92E#{085<PN!KG=K8kb?*B?McjWN
znk5~gAxhWL5TzqDMm-@9L_KvLh<YL(j4~kxqD&nFQ6|D*)En|b)LZ9;s5j!pC>vrS
z%GR+EWg{#`xey0Yu8xB!7vV6<hj@tcbv#7*2#--ABtTTC6Cf%?1dNIy5u#$92vIR2
zVpIxA5S8jAh)NL&qjE@ws9YyQRF23PRYD3xl{y8YN<_h^8d4#u)~OIxBPvF<kOon$
zPJ^fx(J-oqbcpJ8Iz;t|j!`3IK-8!+AZkPmjQT=8i2CY$5cNfT7&SvCM9n%AqGrUz
zs1>pxYSmc~wIUWqE2N6s>>DsYTrs3#6sXY^ci77QEm@P>_}dX~PM5oX^Zn;$+$<=;
zErH5V?w_e<O}l7l(hLnvTA-mxXDAw*9+wB29>xPrkKlo(CqsF#$+#G3G7JMvMqr@H
zWGDukUY8e|Ud9Veui%BIH$!=`$+}o*vJ4ANR$!sYW+)b$oQs1d$8gZ(1P+>9hT^cv
zyLf2w3=d6S;GxN9C?1=FOMs@p2+$M+0h&UF60j+{L}-eP2u)ECp($o45u1`rf~Les
z(3AuTno@?6uqnG_Xv&NXO<9nkDQ74dn~F<;rot%DR0IW@N`_Lfsk&5Xs*DOvRZyX+
zW+)Y#noEPG#%R#g1Pz*6hSIR9yL4#kj1Em*(4nbkC>@)I%YdfA7|=8X1DZyLGO+1$
z`Jm}ze9-g>K4|(fln<Mx%Y>%Mn9wu@6Pjj*GO=m7ENEJc1x-t^plM|&3!4@FDroy*
zm3@#g6)Y&eqUkNeexF}S<Tk#@!;>sdxIcz>33rNcpF*7`+?P>DN%wxT`FZA!k?v!t
zlcf7A)CtOc9G`!2f<_*lB535%X@W)`9VKbx(J_)n9-Sm<<k1O=MxG~5c#!8w5gz1u
z(u4<jo+#--o+n0nkmpH~9^`ovlm~fCoM4d0qzDFiOqyVj$3#g6c}$FCkjErR26;?^
zVvy&J6JF$bQ-l|J-ZbGwo;ON*k>`z(UgUX`q!)SK1m#5@8z)%gu_=N@9-Agu<grnb
zMIIX?S>&-vl0_bypjhN_ae_l0mm)aiacP1>9v3A!<Z&^QLmrnTIplE(ibEbBCwS!X
zDS}5HpC)+Z@lldT9v>rl<nc+8M;@P`c;pFjLO`C7A_U|KX+l7r5G4iV2{BSYo{%I3
z<OvB%K%N*UMC6GnLPVaJCPd_kQBp*n7$ZgGiAhpKo|vFS<VkTtLY|Z&B;-kHLPDMt
zB_-rZF;YUFlq4nONeN0qo*XA+<jE;QMxLA|WaP<FQbwK}BW2{tNm53hoS<anDRDwU
zo{}OI<SA)FL7oyN733)~QbC@QBo*W-2}(hp8YfibsVPE5o|-09<f&0oMV=ZXRphBj
zQbnGcpj6~(aY93$mLfFdX=y@3o)#rF<Y_TdL!Oo-HRNdtN<*F=Cv@cLDMCk{o+fnU
z=}}Tgo*pB0<mpLLN1mRbbmSRv!a$ypA`IjiX~ICB5hV@e88Om8o{=OC<QWOdK%Os7
z_>kvI5kBPk(u5Crz9{KKo-aoFkmpO1KIHilln;4koG_7RrU(;xW|}aOXGTdAd1j0>
zk!L1J6M1HWGLdJ+2@82vim;Gpr3njpR+O}mXT?Yhc~+9NkY^<*tC?plz2vxdrT#0b
zMR-{6VUz8bpBCXj?*OPVy~i}E5z?A_J2NYuBJ#sUa?xOW{iSAPhrZvh9@H|s+uz#X
z&OTE3%oEdd{$6^D$`6mAN2vDh{BRXr1*(>=ZBjG3RVNPK{_yPVkt&=OUF^!77LMn4
zu-8vyk2I1VDcigA!_{;(s309|QV$Fp5~qU$1KCP(d2y}1ZkiwNUf#XdUW3gKmz0;(
z+G|aZE{ZOB|C`hMv+&|^#bLW*hqFsjH&<($U;jq;9X-H4m8F^TV&(XXqHMRIeL-jY
zkKdJN*N<WrkDGRL3I>fZ>~wYt233Ko9aP(-?AOB<xs40I%{aVB`Tyr4<%jVi+2@@-
z-<#{h%?B9XcQ|jK!7CRZn?LUK`iAWJmW?Q@vZ1>kJkT;QYpXxIyGz?Ho$XiOcZI5{
zEz7rF%<k^grxRnZuD^HYp?L$p|7}>O>{_qW-P=A$ewcRBpgc|wD8|Wv@;besSSJgL
zb8?_~Cl5++3ZO)%2ugBFpk${EN^vTnRHq6`b84V;rw+<+8lZemA1Kplg0h@elR8vm
z{4lY{mq)Ufe`t%Lz3j=)507dw3RF#tnkMzl*+<OP1$$0rr@CZZ$rSsuX;-s<gWO!z
zw=e2Bw&j&p?QU6~P5TS|J?xad&fht=aed!TXD?Is>p{3o0UDGi-~q)17*O7T7Ze*{
zL2&^N6d&M02>}6=7!W~80SS~GkU=Q{1(X_4L1_UElpfGQ836;7FW>`Z224;^zyeiK
zQqj|%yo=6Vt4MXv&4*lJpF-B2%ga)a_}uX353;+vkLy0kek$1CnH1(Wepdoht-DQ_
zYMln<(LJCTodM<5y`Wf~1;yzcC|>753AzAE)J0H|E`gGD8I+<cpj2H2rRf?dUDrVw
zx&g|k`#_nx3Chx~CiVNllV@rMrpMlVlB**J+jE|MlB?CAf@-iy&7F47q?=zs&6_)A
zUiHn5(9Pk$KX=$04#%C%n-iKhCpK+T;p3^xk>Lm4F36cQ;m$ik6CI65{jWb|*3^4u
zWH*Q1HG*o<GN+ZZbzXj(_U+nM7L*Tm735al*5Q_pow{_D6>X4i_{>`iX|?yjem(xb
qy`8%icQ5VPtE^Ao{%((#<pzjSiBQDwLkc@r44q8+M_RWN!~X-m0c2YM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp.E685A8514CD614D5.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp.E685A8514CD614D5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..52d9d83c8271037eea8ebd856e946ef15d765075
GIT binary patch
literal 8376
zcmZvh3v?58qQ)~brG&O11=_{ZQXXvznG7=%W|$W<nFviSuLfFZKmz3@CD$UfXh5$n
zh=Piu!m8^EC{=v0h>Cg*yZA<GJTd6Hpb=9jg3Buqz!$Q5Zs?wAzkAN@Is7}{eBbx~
z<;<&pCex~YzSoA>Y~vPIuWVYhVqTWbX3Nj~Ygy5}<N{^0t;@C9R&M!t#!WSYw7%`0
zFJJtWCq7*={3LP1KUxd#A5rkg6AKQWpLO=FpMTl(=<mOdZlAc?EX|qz$bfza|NHDy
zx9*<kyU}<5XDv;8%f5Y}dDe-apXsyG7hgO#@$WM~oIA2*D|4XY&Icc<Iq`Axg*6{v
zI<>53dGo#(SB8!sKQm_J?)bkJdyMZo+J07UYM;%1efrXT)z@Re+)+>5awYY{?aS|K
ziB4gsUhbD(H0JE;+*#igHJ;vkV$Lf|7D&eiO0#c$=ThC*e|vq~tHNtJTZqZ^Q<qdf
zp07?=`SEXG9sOtMWaNkuh!~%KQu0;Ndlyd^O?tooz?Yj_ujIU0)7o-)$5AoTG~)4*
zo0gRp7hV|k_U`XbZ*n{PUZ=Lb@%qC1Ho4avYiZcG=Y`r0KVR_NRrWyO&LuVDiXQzo
ztNe~nidI}XwWl(_?Bbw;@zrG~#-3?^=-vZsjuG>E-O$oTwO`g7hdb5?@xx=z#XiuR
z^LA$M>sY&CSm0kh?azJG7<!BCUwWo}(V3I!hZfEo)^Xr3O-;{LY)$qYPgBo--{<Vb
zXaC(RzDW!}II+BM!HDMi{dF6+`Cr-Bwtw^NHBXeCoVEOw^zp$Phh`7@pWZ**vG%cf
z6Z#FkbYs)gf68vnS-@}Ky1QeVvE`i#rTz}ToIZBAvAVi!$D`{8t?s+^1G)U=*T>$-
zOkXkfynpVj^*>frw3P)m`QAFzdUWxsJ70e3i@F0#^AgJzPg_xvA}){Zx#z)EeKw9i
zm^*OwsmB^ai<d5Z^5W2^Yqr__QjzvT?b6lt3rqg-=JRDgF5j_b`-zUDc3(s8%tNzI
zk9+@x+nDxEZR3{hT()oI<_8+NKmYXe_BneVFIhI#QPVhVa!0Uj#PF|wxpHa#;kyrB
z)E?h?^r^=3&x&R)p`G_;|FrPq%KMTJoOpBR^gj<?xc}zodzIhe-5gK`m9$RVJM!W7
z9p%fKCoT6)^0u`-w>E!w-L9j;q0}=EeY^P8Ba1&;dU49<!sWyK5mQ<}>H1RH@o3@X
z=X?B9y!`Nt?Y}n7ef81}<+~oY9o~6<$N599(r3>7e&)fa9z3+}m9|Mr&&vJfd)|=y
zP2N29oafB0m!j*x_~O*+-GAD?BXsX4H$S=Wo*o-F{PFnC@EfD=DgR}3oo~$3Gq>*e
z>LmH%nbAesg`O2#-y603i7&TJ`S9eSDYK6Ujy@4T&>CBM&T)0y(D&P8pDb?u%cP(F
z%0IPYUB%{(z0tC(C39KjQ@*YLgt=D^Jm31^!I$>WdN=v`-TRupyzOG|cY3yN{k8D-
z+&KkH@16VOlMV0gpZ;<On?A*V_V$={Z~t=SYPkGLN!@$feyd`BEB>wjEx$i<XtaN+
z<gd;hn!Yde)_U{6`pD*68tbo)ntGwn!SuNN{*iAUTeWV3^Xhe}Z=9^HWmVJs>4eRe
zT^iXqZo4!jmz*)v)-WJ@K(Edr^JM7!$@%r!_O%nm+FWva#xBY(%I_R3JO9-Oif-F|
z=A&HFpRx0^^Lup;mOc2^Uq49ad~eUQ1u}MJkIIRigJnB!euvrqU*A^tw$)|q!G(j%
zItR=4T)mij>HAf$=92Xpdtmm!o}Ghb-`>}*jha)G?rp2h*d>J}!#W4c-nb(7>KD7#
zf09ei%GhPuWtE+SWq&qo@m(9@iPGM-X&Jk?uy|<aVA*Hp4fz`*q(<kGH5t1gyP#L+
zVA;R*ov|l*>;CU^$*CE;Uv|G<or7iDnit*Nvb3oC=+ZfzNA=+kuRgo0LmQIU;Pg4G
ztxuRy0q0HKiuTdq(E%Dfm(K;BE8qf;@iE{r0R}wQ$AZTOSnxO>2ObyTz~g;9czl2d
z&+T)A=MK2R6MO=ALO=jd^oih!0TDdOCxIsgB=BUP44xd2!Bcz+cuGJ4PxYzbsR0!{
z&8LB<1vKz<pAMcL(7`i&26#rm0MFy|faeK#z%zX&cxJ!^&+GGo=M8wltGd3bxa)3e
zs2X17u#|O&nmL_!CR}1cXJhw8<qSD*0~L1OLv%|zNJEqk(GaD>G)7%P7erkl7erlQ
z7e<*N15qZ#K$HnH7-fSjMA;AvQ8vtClnZhY<w6`pxiE)OKFC9q5AhJ?!#qaaK{rI*
zAvZ+bVK+vFpa4-JBtTRM3m6rHB1FZI2vIRCVpIxB5S2m_M5VBVQ8_3>R1V1ymBTVd
zm7oGqC8R)92`d;?gDOPTkP1;XtYTCPY7o^z8br0QhEYAJLsSpx5Y@vvMvb5WQ6pqP
z)Ce0G^#nZ-^@KbS^@KebHG?KZ&5#LEGi+ki8}vfd8}dTb8}?$fN~*fSx&iYVs>WB<
zfSOu0+fvrw$GV)(zl(5lI-DcB?>{Hw<Uk2d2~?VLUZ%P=?VzDaGc+`5frciXrf6)s
z94=_O7#B2Mf(x3iH08o3<6xl4Fbp&qfq^EIrWkCp4i=g$!$Ol4SZK0oip3`9;GoGd
z95gwBgC>`zIBfC`9-2JELz5SHX!2=_$EMrihNhcwL(?s|q3KRjZfpt;0h$6MKvNI|
zXbNdcz^3RBp(!#VG(|y#rkJKgY)TFZni3;HQxYU-N@+^MrtFZRDKj!OWkH6foTg-K
zDh>sj3Zp<%5fo@DX-dJS>QJGnGAcAxL4~H8rc`Wd4h@<bqd`*>G-zsRO2ek^(4na_
zIy7}bho+vUbZi<91DXb7K+_NmXc}qCz^2FHfu@J?K+_|5py^3d9&DNp6PhMtLemsX
zXqsut#HQEbg{GJBLendFq3KOiUTjv4sG_ZhRpvp)RB@o(RidS=-$YL53f_8<Su1!l
zr*q|Q?aphMQ8A;dYpvrS&XljM;=5DM7~%X3)*;R$;cQ2pBAn+@M@Z*hvin)+jFQez
zQ71^}H>l&3^B_JuV+4&nI!VyTqf-QpJUT+s$fKhqjXXL*(#WIZ6pcJrjBp{(l_Xrq
zbEOCu@>~(pg*;c3bRo}`AYI6F#VHr^m>9tzk4X{?@|YCCAdiWV4Dy&L$smtOkPPye
zIK?23jS(#J*d)Osk4+IQ^4JK;B9D!dEb`a{$s&)9Q!Mhh7{MWrOA;LNxD>%5kBg8T
z^0+9;A&*Ot9P+q0#UYQ65j^tvB*7z(PZ2!w_z1}(kB^c(^7sVFBae?$Jo4Nz!i_w4
zl5iu>og&=Gb4N%w^4w9<jXZaPbR*9lr`*UBVuXM^AxQ|x6H<hLJRw2~$P=QZfIJ~V
z3dj@Ulz==jMu^B0lZ1#oF-3^T6C<REJTXd&$P*K!h&(Y)iO7>;goHdPNl3_(QiOy&
zDMCuflcJ=AJSjm+$dlrfggiM$$jFnEgp52nMaamLBczNxIZDdNlM|$jJULFu$Wvm3
zf;=TjD9BS%gn~RJLMq5pqNIX6B|$34Q{t3@JT*qB$WxPqiaa$%sK`?zq>4N>N~*|H
z6QqhfHBPC>(_(~%JS|CR$kS4UhCD4oYRJ=~q=q~#L2Ag;;*^FwJx1ur)02dbJUvC|
z$kQXFjyyd|>d4a*q>em2PU*-qVuXP_BS{#@Gg5?sJR?FH$TOm(fjlEY8pt!^lz}`?
zjPM}OlO#OI^P~t5@;njJgFH`^^dQfZAU(+Q#3>K*%ot%J&rA{~^2`)rBF~JFCi2WE
zX(G=|kS6lXIAtQw8za2P^Ck%|^1La+i#%_H^diq2CB4Y=CP**xym89g&GW85XS?#l
zif^c_hKa)`&bO9GSq=W-eo)hgPw!HbrA@aMG7mjU<TVt@MdPe@H+56`Rh)WkY>&)t
zUoYQv)=Uk@?_ZpKbVxgu*D#r$tXjMC8m7_HK-JUrU249&`tSYcy&K6?rl?cA-IZFk
z7F0kDbg9NcwefduynAT|US+QuXjPVH*`9ityZZUHndm99wL7n&maYXApaWgXdLx@z
z4PJX4+r7@4S}}FBHNou@$H(Zv7=LEo;>zNB>pg5<LrG;xy){E#!{Exn_15Zg&0Udu
zKRDL0H}eULXBIbDmG$Y$V&&jNMVa0Cee(;g|7$<|-If~Wk24otJ8EMm8x}Wn)W%K&
zRX?`AOIa(*tennw#N8hp&uiZQIj?yQcwYZ{cI~Q}9a}bLK1$i-vT4?QyH@_ahrg-S
zm)SkA&%i=!A^ff|ZDEhf*0Y)21Ih<5)+_k;kKfTU=F~3}2V`Cd+nvL$N%9(KI}OTZ
zcY$K;3@Fylg5vBPDBjM4a@*aY1iJuAw2Po5y97$M%b*my0!p>3pftM%O1JBv47&l!
zWA}hE?ItL%-P@%;tTRr}E&Zx3b7m)GO&HdFW^1x)K-Fc{bt&sw$y}1x-tl*z*+~_X
z+|~rgzFTcRmjBw3%=s8RbMOM|vOCH3_@U=@f7p=O?JMw=T50R6fx?{5uLQcMbo!mO
zaMJuVD3{*_it#g`SU(Gj^K+nhKM%_7cY_lA0w~chf|C3aDA_NAQv3=i)vtom{2D0T
zuY)rD1}Kl;1IqN9puB!BsH(wLL#=~%*1qczQhn{(1F)dHfU}O}d8stk-0<xF%<htz
zCG)Jufb}&?K~Cq_D{!cFrwNBzr$M=N7br$&K(RUtiqknzyv~Di>uykjE`Sns5tO7$
zpk!SJrRWMMRaZf2x&}(ubx?+Gfb!@bP^NBz^6K6$)x7AIdDp&zYH423QhRM9cy0L4
z&unYMW}DN!c~jRW)o?lX?ZW-}Szl$(yJ^ns;9T3>F3;9E>wn~eriHgOW;Q3@JekVs
zk=@hYD>tuq-|PBR=U0w%<mXi1(C-iZ2MoMkR<yCY;W0-Q(CY9pBTE0jy@i8{ONI;`
iR#sjy(&=J3ezYhJ7K#}6_=3W!3G+$c6z`41r2hq1BP31$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp.02D5EB3BF76A9EDC.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp.02D5EB3BF76A9EDC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..276951e0f095a24d8ef88e890e28421d48584b51
GIT binary patch
literal 6614
zcmZvg3se(#!o@Q)5yMO5DIy@gqM7$oln_)N`9zQiR0(1<RTSi_vQ<{Jpj#iMrHZ>s
zU-q%9eyg@>b-PtJ)zv<>>SAN8t$m=yR(7?v7#FoztlDnOo{_ufe4NAo%$<Ah{Fr|x
z{~4wv-){E}HW+d$^OrB0U-zidU@$}newWrQS$Ksq7@Xk-!}3>;&ZsE8kTLM~AI3Xg
zYcc(5+@b|H4XY0nhF6YBNwqzD;E?vhsG<2qf4|dHb-94twzB%F7#rhX_4)O`l(ia0
z8peKf;Bt&7_PBcH_gN`7jC|FcPxoB7wZ69OyW@?Ghj#V$X8&Heo%)hcwDHHds$C7w
zzG&W8J9K;J?VI0so~+8a*SqR@{`$YJe0Qhwi41ez%d6g*{m1pS-$va!a`%-V?lq|s
zsAJv*DZe@$RxympeDSM2@q7G9u2Yd;4B4L<duaEY-xjniZ1!zk@R5CC+jG%}H!Xkf
z{@-eA4%U6OXj#JbqL}6XocU?G&9yDGE_=lEke7wxyJIq{T=!htF09Mmk{RZBzxYO~
z(nY5Vmj5G{pBwyI%TuHBPZ_s<`|<MmQ+Fi{&1(F~w8`19&_QWa7O#o1RYtzkxU7Bi
zGcosO_A7q!!wu<&(?`6uY|QdB=k|}kx-s#`#;Eg6n+*rwem=5dX>V@w+TFif-1|*Z
z^xj)k?%87X`SUNn|Mc$q!42d8I(*A3A)dDul(wJ9sj7dvdj3HA`s=CI%fGqG?)8+;
zq9?tcw&}=Un=fAtoqnoeN@iu)!F4bF@9~4%*EfuND)F%|&wbeynP;~&ulr`p`Ol@u
ztxM)vw!OCRO38KK#heq*r^XGi9gZ(A**w{@{ZdGp>YMgz3{6KBua8gpt@YfX&x`M0
zx)!(Qsd2mCeBs`)Qy0!Wd;X<QzC5^Et2f)fIGZY-QP$0Goqu)fjm(oob!K#n-`jFc
z&Z@k1JeJrqv-D)$DeY#=jA2Cug(XAFyo2*oe_8agsb<)onaoPlxdZ!-SKO>K6<yx=
zMV@*#F>L4kj?W4X{W*t!_G)wChnB``iK8aA{DIq7<^M5$<;jvBYwO+*%h&EWw0&f?
zJF0-6@_o_pUDJ1b@WUA5m5C`~H?z%eEIIV}x+#};?!CeDtKTlQD|45O%}PmWsoqg?
zf9G3^e;L{RT=JnaO&y)DxgWpsX2GJCj^v%Edf&6GX7>+yVczr|$3CZ@d)MEUCKPr~
z2`#Mnto7W{jMFEoF1)+DaO##BW>-qBe?zWo^oGM5$1`+qL!F#4w_v_=Tg|Y|?UBOc
zMLC^w#$9cxZJE~jb+)^*b<1b@jhb)anyl95k5*0YymRGt*BW#B;>$(Jf9!X<JLl44
zt%KT*x<-E;OMXH%_vSv=+P3P0(XGC(D&@@dEhGCk{C!Mo?1=%5L*(L5svenk{Nm#3
z2_GNYcKE%fW7%hm>cZb$TbB3f&wKcoH-zfQn#TUAww-rQjJ?GDeEaA4=2iVa+yCuP
z<^A8a|6+XQ#|nlzyX;6DJ7~@9O*0xI#H^hw=j9yP`<L(TboH*u+8|E)=;dDXfpwdI
zA%1R1*tc}WqPf!vgCS(7XJgK`^t5ntM!CTq7ZMj590Kb>@F(o2tryo8Xr<xg^njfd
zk`x^rbUW%dwiPEfvRlK+nE`uv$ndP-pxXn}|7DxLLU=OLP!h0{6OxAo2i=a|-<<U8
zec%5roSYW0qeG%YgM)5QxO>C@`jr*?!^yIMofwkXKRD>N{q{?5hB8^ZA`H%eJ-*-g
zNx?z4Yt7BK#%_Fja-^X&V5cUejtmaE{r#h9hdI@+gp-bd9Ul@O8XR={+Jz{2>BiaT
zBMpxP?39F*5y3&X?+={u=Z@MBuY{Au0XrrnCNwzcwqeQq1xsIu?VH_{u;8q=d~kQ`
z-b>cB2)D&<$=4?@pq!S9K1JJU@MtFu9%E;~W1I|ltepjqb+X`bb`Ctw$$`h)dGL5A
z51wEbz!RJTc%oedPjrglNp=Z5$ti&++hy=%rwpEASHM%83V5nr1y6OV;AwUZJk6<r
zXB})!et5CCts|^vUFlah6BfLx;p1^xs{1|=OO55%pz1Aah(1ZXXo%7^G(_oo8l#Mh
zfhbeMK$NLxFv_}Eh_W>-MA>>4qnwL_C|AQll&j}3%DZ@o@-;j}`Fb9sf=hs?P$NK8
zs24CQx<rVIH6ldCdJ&_NOM<9WBSBQEmoO^3WQfW&GDPKi8Ka6zfv8fWKvbz$Fsiy#
zh^jRzMAdo~qnb;Ds8*vvRIAr8YE8Fh>(^9-+nQ^2fGW1m(v^PI7{h|gWZ(5RA$CH6
z9*S^h$7X}di_Lqe^aYg+3ofK^fWir7`T_K1bzE3*IqeHs%odCOCj=H9&LV<RE$N`T
zDa$RYPt#@^nlwj4lU8YH(%lq|O~%YXli?U>GAaX2rki50$(mVcvK$LdR%M~dc2g`i
zIWq@Mj^m)osT?%9Zi>SuZ|0%Nb38P8m4_zZP4U<i%mOq8PJpJM3eXg~DFK_JS%jv@
ziO>{P5t?E*C1O)DOVE@!37V2BK~w6cBy7rN8JaRDLsM2|Xv*D`j7`O?KvUrqXez1#
zO{JSsu&J6=XsVnFO;uH)sdiH;HZ`*bO^wr_si_(?wQfqorge;!(Z3dfFD7RdLCIE4
zS3~Y*F;6V%N{9@6Et-s$XN+*^HyLTvOp}p8%{CcX)LfI1L(MlCdDKFaQ9vy=8Aa4m
zlTkt~HyLHrN|RARtu`4|)LN5K>(lx()U2@JvlFZZUcz!3Zex}X!g3L{pRim<?IA7i
zkbTo)X(KJCP`8to^Qe84<qWPMUV=s*-9gaEqx}SpJlaFj$fMgx8hLa(Nh6Q;Q8e-x
zFTo&>=^z;7F@AzU9^)Yy<S}g|gFL34WRS=BC<b|~mtc{{b`UJ`SU<rckM)o&^4K<#
zMIPHuvdCk76pK90OK`~JItUJVoS)#3$9YH&d0ZRGA&+Y(IplFZibEdnC3xiV9R!a&
z-cRtz<2@vgJid+Ok;k`_Jo0!S#UoGf5(4ss4njbl;3ov+2_8~Fp3p`L$P?O00eOOt
z5|Af)2@!c>2O%O)^b;cTL=Pz<Pi!Ma<caO1h&<6piO7?@goHe)gOHFX`3VVml82O#
zC$*6h@}zcBLZ0NKB;?6nLPnn4LCDCH{e+A>*+a_6liNrcd2%}`BTx2GGV&BJp&(D`
zAQa>&enLT>;vp5}DQ%>JJf)pfkf-=41$nBMP?4v05GwLiKcON|^^hv^)HYH@p4v{T
z$WwikiagCrXvot#2n~6fpU{w}c}NX;S{tb$PirSN<Y_)i>*Gy&`kLX+4|V4$qZ=O;
zjsNwiXmsPFqJC>5!h-)6>ns1%^v$&i+=kx~5$>dMNjds|K#vr~jJ~=lvtM9$MrcNq
zUUqjrQx)=c+QkR=#U9GC4)2RS6pfncp%~O`55=P9dMFMx-$U`Jg&s;kE%s0%YN>~k
zP|H1(j9TfT6x3=DrJ~k)D6LOd4=ME>oxEys;C!tnYodO>`b&er^Yw!_2KuB$xbq70
zX6m731-V|vId&Qq;ZDv=F4LdMBit!@DP{VLg9vwOUTT?sywxi`>)!w7(mMfMa(S{_
zuRPtImpuN=hNQsm=z-A*`h)wPuC*)5FTJ`kuzR?Dc!7S?-MjoqziErM+XK532P7uw
zxAtRdNoBvh?OlQ0aiiim{e{kZolh(sd$qUV!AEDZjL;7f;igSAD8|HqVofY4&cuP@
zO*|;UB!Ci4A}Glufs#!!D8;0JQcWr-&7?h4A3T!v-JGFcoD58ImN9F%KKBvsNybT_
za*eqU)i?ima^j}wKV1k+L~42JJbmsji~WA+A2n!wV0T7*#!&r8S5CiibVuV<M__kK
zdCFY<bJO1oMTG_52lZV67Kfz-CecBIVjK)8*1>|}92_X#!GjVU0w~cTf|48(DA^%{
zQXC2>)uDpY92zKVs&$xt#$Bepza=yN^x&%>KPq0-&s0D2z%~EiJy+iuXBTFd>s|eY
z*MP9#i>|(q#b%iZ(_*7RF*XJiYhyuiHVzbT<3R~F0hDMHK}j|Vlx&khDK-U^YEwaJ
zHtnHWGXJqhAH1qtx@6wc(gz!^2g84UX6YLS!|cAzioVS`4>u|Ik!<he-M&$~LmsV|
zJ<B!65IpLC{CSHiAFB>*7A%-b8T*CwH-&~rL=KD^kRP3wV~!5X&yE=s8<#kE+<0rI
zEo;I=KAy=KF?P(*|8H-?kmQuKVZ%p`8f~FjL6j6NJylKO<lOj#|CutEv=>d9Oceeb
D{hCc`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp.235E5C5B653C03E0.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp.235E5C5B653C03E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..794a3c1b0a72a82c64268f9541320b55e6ca1abe
GIT binary patch
literal 8376
zcmZvh3v|?UqQ)mlDMMR<BCS|j%A-tSGBKH$kQd1$K&O-f9oj+%WTZf~Wsc?1q6K^`
zTptJm3W~e9pgepcg1UNlT+#J`Vs#vJSCA2>JVaN(fdLd$bnhLyC+&C7xjl#f$@l-h
zFFz)kr2i(>6@LHAgB^~s4HZk8Z&^G$!{Kn`rvJ4rZkczEayTByb~u)7{$ScoRRguY
z9ln1)_pz7wc;2vYi824!p8w#8yoVp3d*JMhGp}F%dBY?B{`JO=%H?Kp*3^gl_d9Uc
znN16JSNg~MAN;hn`HhmVSGCMIdHJc{OZ>6MS@GXa|M<g^&0ClcOYdC$aMj5VTF$Nb
z;KKI{s}{BFeQrtk#EH|RM(&RNtI=nCbA110<))6A?BP=v=BWN2b7$S~_-(%=f4pPS
z-L37D*(n$MrEVE@W_k9EuL_z@y>as9m*&ltjt!7zE_mxg?cu+@^5h=j<*d!bq`E2d
zDmLb-6PA4N`<F+*2!9(nVgw_`#~&4aS@8GsrwS&%bM1f^TDJd^^;*^T)<e%66(h|f
zHjdn|u(&Y)+zo%-{oSbzyt{87b^l*pX}Es_zv5VH{oYrft$FD3Ip5tStAcmVs~TJI
z$k!Remwi;Q_?PcrEz2!AKQM1xMajw0r#sf%_u-0T#Oz*UTK7{O7xkuLt`$P;(5N4x
z@9Hf%J2UqlzxSb`!GHC1?s&f`{5pGW@#&6RPJf$P(=dDJ@eiM9Zr)M4CDC&nO+Ecx
z?=$DO{?;qDL2O%HIlOP)h?cs2wd<Y?ytH@!zDH-Sc)aA>8H-*@ofx!kNan!*zUIee
z_pYBkq2G`T<D38dSmySux!!GCb|0@cHosM>)GZ6hsbhzlDk@5zdE|kC%lmG5S04Vt
zE2GCVQx}gu8<;g?<xi!h`%8iw{I4I}ezbAvoi9B9dF_YubK(mds}~m~iHoCqzPfs8
z?{(u2WDi*W{raYG<NSs{o*(k(swbTRsX%+SX8!WJhN6GG_H@Zli=NrM?d0*JPJey&
z^n)`_jeY0Y+nJ6H`^PTaxp42uM^`m*f4X#e+s&_TELu3lRn;_f((%y#5yKAu{L6(o
zhweFWUfa0y=%%LOpB7A?N4xLKywvbP+5L%CCtuq+^-qHu_TBn)ui?wAM}x}1qV3gh
zj9lCC%<zRR6Bqd>TKo6!xHorq?XIK3!Q@kGzHZ!er1Aav=O=$ATs-7GVoEC~USBM`
z9?75dbdN8@i)*KC`?Y!2o(p4!@7m}%wDasUXAgRcpZek7r&n)Ueei*o`fO_Hz52n*
zhQ>uDPq=m*{=*m3r##wc$Bxg5(8#fs2c7Fby<usO?858*XW!Iww>D-}rW&7LbK9uD
zsFiy@=`s0Wqi{z@pULVULMJ@W>09pm<i#zKeXqRDY*|~jDRFqxC4KDjw>GUmFiza`
zL)~2dR8E!a{nke-E@^My-16Amb)}EJePL}y+5JyVS@rr|Z*Kd2$&qKzoEW&Q&p7QT
z=C|wiy1#b*{WSaTU8`QKet*`n&*wg|pgCv3`M+L1H{pWz{5|@WhiCk%{91bWx8fHY
zw|+f;>xTR>$FdKG`>a^gXHjYVv)z|M+dGP$Jic<MYvJp|F6LY}^-}SbV)u=X)}_sJ
zrV<WEW^rWQ*lp6_Y;xLkM}7ay{=GVf^qryeC+pX(9rsp>HQD6Uv|W%{klQ)fcJ7{4
z1-I`${eCtXNZYxYxxG3E+ZJAGz5iI&-HUP@!L(i0qpY%Xux;O!^U3GGTe>HktV`Pi
zG6(eR9Bg}9ZOxZ1XVE9uIBL>%QGU_T&cU|VEzZ93`L2~8Ws@_~c1dPQS?6HeTl&BG
z&+s32_P)kZowf_}3x{+Lw*BdQ`3>JpyMFOCj@q<6D1T5%=V05XXAk}xBP4IkCacnR
zUS?je&cU{S?>p_)#Daa_Ws_6VcE8Miy*dZmcC_4bYwP@i?xRa*bsp8V?_Jrt>$o;J
zr{3*%SJ)pQt%B~Gx)trG!J~sTcpkq8JWtRA9^+@gV}cBLte*vs4YJ^Iehxe?$bske
zd%^Ptz2NbF9y~tCgD3a}@PwcMp6D0B6N4gnl3xN(3QFL~ei=MDD1)c?74Vdx0-owu
z!Bc}Oc$!}WPYY_`>3$tNJ*b0c_zm!kpaGuG?*q>l^nqvkP4LX137+M*z_WrDc;(lZ
z7j|7u_2t9LUAD5XP&2FZ%7m|2$lcWarE-Vew}WbP{}<6M=@1Q3I!r^9Zlf{k33(vu
z340*wY4c!|2{905!VE;2HU^_?h=nK{W+BS9u^8n-97MS=2T`t#!>Bjpg{U{|g{Zg9
zi%~wrLzEBm5aru=j0zzEqC!}JsL&>0R1Ap_6~iJ##WoS6Qb>ZR6qX<=wMiJ2Lo!6=
zunbYTO~$AaQXr~?6^JTr3P#nC3Q;wzLR4*2F{*_$h-zUCqFS4VQ9YzXR1fPA)!TH8
z8X*IsM%aL;(Pm)O7xF>W7xqEa*XF~h88RVihE0f?Z6-#okOfgIY(dm&voKmNm5;G6
zz?}N>aphH@rj*aLmHqp%HmmdZB3zs<_sH(+&&{|wP=Z?mm7?4iscucXXlT+54NY30
zp-HDG8k-)M2bvzn15J<Mfu<)#d9ca27-%vK15HL?pvj~t2Aiylg(l0e&}0P`nrw<<
zvB|kOXmSh(O-|sT$)zX`n_ia}nqI~WO|RgErZ+`-vB|r5Xz~mXO<v%k$)_kDn}SP#
zroaf$6a)d9LW&ZwDY`^xii`+NQ4pairYI4cl1qZ7#7NMT1PPi_ijuG?yJTp}j0{a#
zkfAB3C>fiIOM#}sD9}^{1)55VQn0DIRA{P<3Qbi|p{b@Q6`Pt%gQmu4(9{GCnp%p|
zu&KLrXzGj(O<mBTsi!C%n}*AProkA{Gz0^hMv5}9>2vv@>0^A*^a(y_`cjk+o2JWz
zrpcJlGzAlyW{NVgX}K(DT8srvOR%76r6>!V<s-^z`(~BCkul{QD85{@)xawTy{zR-
zK9y5HY4jw+p3bSS9$gKpZggFjvNJ?h=c3=f_t}g7wdwh*OMZJky$X&J?oVM=;!Y6m
z4%A7)eHL|ubiYA%-x=NQr2Avkank)2>KNrdfOpa;K_ib&5H#}WBtavOj*v9+=ysAu
z9vvrX<k2yTMxG~1c#!8w5FX@tl7t6&o(Sndo~NDkAkPygJ;?LKC=c?OD8V3)Ne~S3
zm?Xg<kBN{B@|bp#K^_w)8RRiBia{P5C0OLK34%o)n<QA|u@RC*9@|c`$YbLqi##?)
zvB=}11cy8>L2$_9k_3l5E<$q1<Jw6Md0d?2kjKR+4td@v;YFS|L3okpO%h(@c_XA3
zdER!?i#%_f^diq2qrAxDqXdsUK0)xv<C6rBJU&A5$m82d9(jD6<dMh6C?0u2ln{_7
zBnSa{LXr@WCqzgAc|toWAWw*s0`i0yB_L0X5+d@%1R)|%OcEmU#0V)OPi!Ye<cV=o
zM4lL<MC3_PLPDODASC2TNkT%N6d@(#N$sSBJSk2}$dh7}ggiM)$jFlugp52nNyx~P
zBczNxxt)}eC&x({d2)=Bk*7on1$jz>P>`o22?cpdgjA5Hw37<*lsKs%Pl-_q^3*7y
zB2P^aD)Q7Mp(0O>kSg-jc2Y&28YflcsWD1Lo)#rE<Y@^)L!Oo-G~{U!QbV5BPHM>0
z;-rQ=Ek<d`)1!orJUv0^$kUUAjyyd=>d4dENga85oYax0$0!|nMwBp+XCw#%c}9{j
zkY_|l19?U}X&}#tlLqpP7-b;O7bSei^Cbu$@_b3chdf_|^dZmJPWq7Ni<3U&`C^n0
zd1jO_k!L0d6M1HmFp+0QNE3NxJ82@%jFTqv%ot@N&x#Tj@~i}5A<s$@7V@kJX(7*Q
zCoSYzaneGb6{D<fp0)A^$1guF{))<|uN+o6$6ozp)CYzIKusMswM$KsHY~_z);vPw
z)ECGFW9=t7wUhgme!qTnkMy*^m%op_C;M?Inbr9iw0qq&rF6=TcBuZugN>O-2X|au
zr_qyCdpf;NqpLyH(RE#F4qx%+wX@%jq$`uv$yRrzMy&xARD)frX<$w4t?~EFPyY|g
zo#g}U%AyR%rWd#?pIzM(Jz2J=bLwm88c;zx*ro2M92cd7qXOwlVOe3F{m?e2zNoCI
z&i=$X^@GX=)!9#MSKJ-B@7-g^-$=s?rx(`Sm6fToLgm1kg7kE5-`srrFK>H!n32`_
zbh!Jd;YrH*KPM@t9#4|J|MX$M`VhH$2NO#t^7co3@yxwTryt+EE)6f4R8nn2cP)9d
zN1%DTKRrF5_ketR>HD@&-O!_K`<e7~|Ka@^`w{#*CziF2`u^w2{^_N#(>=`INlrcO
zq(OO{9#D*v0mV95P@Izk<#l>N@lGC;;1ocKP7##klt9T&8I<BwK&eg@l;+ex=}sM#
z;WR+`oIX&d(*$KXtuFOmt#N8r@t6D4UzQ0O6NcIcKc~JbqY6}QMs1g}zclHycl8;6
z_t~3PI<4!En`7TBH`nLBd?bB722CF{*Y14F_4u)8pX(k<Py6%y#Wv1<HISdx`AVRB
zm+pYO22NUl2IUEOKrsOZ6dPbcaRCmLH{b=u2Y66IKma8ML{L&d0wo7zP)a}nr3O?`
zT0jG(2Xs(IzyRe7_&}Kf6O<LOK$Q<FA7UT8GtONPlj_S?Z+Usc^Emrho|THDE%jUX
zrKgLg7tOXWeET&^URLMp6*$zo+k``{)1W-M2Na_-pje#+#pxU<ukHoK>pUny7eI-+
z2uji=P_izAQgj8Bs;i(hT?3`-Iw(UoK>2hZC{s5<S-RDwT5h>*_SIKVtu1p~Yp!mD
zt`7hCnQ3o095;7w-qf{8)xUSh>$vl-!r7U#Z@PJAXqIDEm*?o*>wo0j=7!sv(wmjH
zPNFh;WcGCS%Fel_Z=c>3xn*Nrxmgus`dxQz{{h#_iZ)s|eC7>#v^s3mh~ocmHh*AY
l(cmFNONN(@bbDCNd!r}~5(*f8TwZ?pggK;tvNfKV_`jkM2g?8e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp.37CD0D23BDED400B.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp.37CD0D23BDED400B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d065b34c36dc44fbd03ac908fa558a9f271810cc
GIT binary patch
literal 4654
zcmZvfc~lfv7RI}(0h<OD8k+=K0)o{3R4AkcS{Rl_9Fu0JkwsiE&_*z9Js6F<NJN98
z#%NGcbCeidMjgP7B;o?Wc7stgNH7LQSpyoEAvrTe&ndZoOrO(_`rYrV{_c5I_39Bb
zJ3O3FNF<@@vkS8p=P$63NF?^cH!nZe@JK3=ARCFKu>4X&R^*czb~g{4i~RN*`s10^
z3fGF;kGdTz6BAZ$Nm)^rnA2O{clO1qKes*)ImnTzuQtSazSO+r2X+t=>oq3ttaBFq
zVLAIVA`g@n;`%c|%{@PrG%dUTtaGL2fkXMFhqrdG7}M3X<H^h)e$9{kZSTUIVLQtj
z{r%lRb+>=H{;DT``D%6K)Gtb&ds*Lo(X(*Rl<!hCHwOBb-7Qg-{bBps|Eq%U*D0$*
z5(=lB(|?rk)7bMnRVQlZEhm=7C*RULH!Jo9n-aF(Eu%g1JqFhdZ~D1qNoVa({qOec
z6otPgHSf^rGx6iMBX+IrVp{^U_nnzicmHYi!6>5itD_y%^3pv8%F^9CqcKhQmV4JO
zZ#MpDQIVH<@$?aglE?#r2ZDX7E}bRUI(w>vKYQ<VQCwZ^QbUR2k8sQAHCMK%&mY~M
z7%)L&$3<uM7gnsO8<Q0pyyoG9AB!^lSGud#AKRAwy2IJN>bW$meY<+s!=KlWk3Z>s
zFSl?}dbV@NwY^!3RpVU-#@+eREv_v((0KW~J988Bj@HJliwaKeA0mz%R#(Qf$97DL
zt@ON-$evK9w$(RYonK#lZC-W*FQ4rrn^&CWQZHfDU*HcLmrUu*?wgsO@+`RH_TglQ
z9cAe|-T5}jhHFkku_4-LlbRQXUw<<0_l*TRSASU^a>_3GAG^xh)j7j+V^p!u7rahA
zz8N;)e`d^$=e7&}neODDskxOSagHpymfIC>w0Ax9raSEWWB1~`_I<qm;)a6aM~d|=
z!^%bB0i~;_XIkvq7LVo(;L7744c;qEvN)Q1y~jPqr>NZU^M#d-#(`}m?n~S^w{9%m
zr?rlHoY8c(>eG46{E=e^_FPe{7@P2Lb6Ld?pCsl)#}t(1B*io3uAcK9Dno;lip#t+
zRI@7M`pc)cU!63~Z*hwUk@2MY&Yj|LrgK2k)mLh<zpTFN!L;Oy4&2L&1>YS1dry6w
zlj{u2ZJh=6xmCwn?k8;=3hMEXu-Q?YfHx<+m|QSbcV=+1sr$^!)C7a!?vh#Fih_O9
z7{9Kn%8ple8lM&)-??Dpv5gTQsxMygD=1!lwVe;z5!EvOUf}7*8{LyvP54(!zzg(8
z$OQeTo49Xgxym!Xx){;lN4(h|B|Z9|BRwwhf%cvG=Us0+J9~JqZc|>fA=hq2c<^3t
zV~76p&&pOEURO5Q6E>`?-gwt<dW+5G>e%Gh(z8tWFR$cbU)NO(9vPJOKJ_xyR9k;{
zXxYHj@Roq$_JRBF>>76bGpOoKMUGeE1KrkjogugXc56u~cI~>)tfiUcZ+S~s1=M}*
z>hb<<3u2Lj^OVLQk7KetN#63Tg}#_XV&z)<dFVm4n++bHC`nUVDXq<eunw3%@_$#f
zub(qL)&`FgWR;c5-aLr1{mIXq3r=*`+u%`xtgup8n}@e@-R_Dnw}#r;O7w#4W$883
zJczPqVC$8T{iz?1mBb3Ni=)dV^B~GW!-IX5UCU3};0c24WaTu*Jc#n%aaA8R98g%;
z;By7p&C1Q+Jc#nCGV2;E)yX|Jc#I(1TiMIZgD7_|a66BD`gj{WQji_39IefRDF11f
zaQe}*D_u5tv>-cJIar$qQO;fbVV=@<boLzN=2_c%dAOqLae$j`8lfYi#Rdh^6G@{)
z>PVoGdJ<@qjshB`r+`N5XrR%08fc7;0UD!cfX3=rps{)uXq=7%8mH%ghI9yMNRNQV
zzsFCq7q4y_@4*uyiFeFIx%o~3Q<6etj7|b!Ao2kjiM7}$k|`vxk_Hl3Nh1lZR0;*G
zlz{?P%1A*gok9aEZJ>dbHqy|_q%gqB7#Lt>j1080DJ-zE1{PRZBMYru3J0v5fdf{~
z$U!SgLBNU(2w0I3K`Zae2aA($o5qLoIzUl;vPj~ys*;<ZQE*`hf*zfFf+u_d4M~aL
zrK3y|B=C|v3B2Txl!RW2pn#X+Dd44sq!jeh1P#12PXjMKB&DI3AsFCgcm{ZxAt?jB
zEWrXV%d^1C4oO+)<p>UVIi3SvZb-^OFCq}|A|3%R8j>RD<td&OFM@D^ym%jx#J`x_
zye^8DL|hlG<>vKJEDP&M5k_o>&q)zRLZ*r^3Nl@U(U6%UjDgG+VJu{>2;(55A`FeP
zxE}k-&Fe5|s|h1+fVPM>jCcgO7bE&0*Wg4gK6+CT4LH#VxfLfKKsHH<oA4%V!bq4$
zwqYd9BYQCt=8-ix3G>JXoP>E~D^9{Z(j+Be9@T_VFpp}(D40j}Vie4yYH$kXQ4Kf+
z^Qcywf_aolO2It338P^i-G<RHkM6~2m`B&(G|ZzLa2n>(tvC(yXp@wNc}x?=z&xf6
zV_+WBi!m^dslgeT$28y!%wt+{2Ies)DFgG^CX9u7Y#YYHJhm5OVIEt9voMcsz*(5b
zw&E<zV@*;P=5b9J2lKc#jDvYxFUG+<t_J5|9@l_#Fpq1+Ihe<pq#VpcO&EfCs0~9f
z5A|XQ=AjxK!93J}BbbL;aRl>_Ns31Ef;aU_-n__vh*_kmJ=A(}P5D4ww#*T2*@oGs
zsb;7`#n*D2PC-81uCWx#-qzmZMR%qu_U5ICRoOzBm+|k4W#gQ%8*HL)lu&jFafuTb
zrnR5eZup_&@wt(5f=jAc-t?PxqTj78&O%vXr*ITsIO2nj%gr0K(RBm1`)$wq<#QjG
ze=eNs7UDKbJg2Jgf@M_Jp>Uz>G|tIUT!O^Som_6-)Q$cmfrzQ%xwdH}A^}l|0z@Mk
z5Q7*%EMft1hy#QW0>mSJsyJC^N$dD+ri0IfUhVDdCx|`P`NX7KY7g}YWu>PQi5u?k
zfB7WO_v!07GGSSl5gwpv83~AzQGjR}4TzC3fLIv|h?8-EkPN-0%W;9v(p>Lf7y9wF
z@O2lv@0@1&V$Aq=N`$hv(%V%mi~D3@rbjl(V5YT11ZY}I0;04OAX-ZUVzdk(R?7n7
zv>YI$MSyrGexle^kF2T`_dhjqZ##NAP7`|+H|@fEb7aR3desm?pjQnEh|*AiXblaB
z(J+8m4GW0VaDb2oy(LNR;-w2l{!_@yU6dC)a*#4I{O^-29!MmqZx5wusM~U@qB%R)
zYJO5`a>_hO(pyb3Z}f2f=;5NQ^re{s&iU{IsfDH07@4(=?O40<<7V4yLkWBN>=4B}
z4inyWn&B6qnHd=Lmnllt-@})3{eM+Q)kGJ!N$!(9Nt)rNPV?fYtKai+M&4mc$60?}
MfQRdXBe3`X3rb9Z4FCWD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp.2924213EB866E058.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp.2924213EB866E058.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5f129eb71f573834673897c4ea6943a6dc888bc1
GIT binary patch
literal 4654
zcmZvfc~nzZ9>?E%0Zala1ValEL6En<2qY5O%VN~Ag~%GghznLsMaw2wts9C|v?yvj
zTCAu&iWaM*4hE~V6_<%dTwr9Ng<>s?LIVY@%g~;gM|z(5{V_R*oBO@r@8fg1_uTt_
zB+b@paT$WdWzR0jU0S%v0YMNq{->ZYe|e_}K{#gwDXG4kmYeu!x@+^HhQu#VQ13_9
z%Y16?cJ_JH>eJTkn7_74pZ80(_58E-f9!l3dzd8*FSeuvJdb#;8r*{;l8v{6a~m86
z;`2UVmUyVL1WP{`-O~Sk#jRCsPkPoxw7XYd?zr>m+A+Pi_B@LG?&re9Uk@zK8*z2K
zklyxbbmQHhZoKF(T(d!wIQ5f?r$LhY&-xedSAJuNunZ2Yx?iEF`rYMa*k?uGZkE@_
zrj;lg(%(z_e(c4)imxs#Sc9)f&AgND)gn6>b31M4{VK}8(0^#-$hIF_jXkFC2i_b=
zkd^$L(XuBYJqJ5=ch<g5y-aJ=%7f>Wjct$X59{#C&yIJKYby5_$tyqII~R@U+wtJ~
zmF=r<I@A>8TsnKqy&~~Y)S(!4-R1MdCa(ZZ%tw>XmZdbBRxGcO{jPPKyYcD{&Bf!p
z_2CmDT-muf10^+U8^`3v#cb?ozge~{Y@MH?^yIFUFT1_m>Yj??yLM~#b^KU5KJ|3)
zgZz>u*(<%euOG-=su=G*IPTs}-;^KbMy<Z`&Aoa0g5#!?&AOP(fnogEQB7@9S915{
z<l2C%dgd#+;fE&kwS`Uf*B7k(he|qIEnZNb>)nLVnoqC}vr*Zz(i)jP|4B^u-J_ZA
zd#bYc`l)_Ewp{ldPL7@NWOB=5?Tts{e%o5Kcf((+W6!u|{(WCnmnLsyUXmi&>+7I1
z51Zo$!_JMdJat+0&uq`IoQOMlh*x6S_55D#YB!%FulnM@J^3Id=-~UMm$np@cgjjz
zN90Si;guVv<v8rymCDT-#N?;m8+uTZ;cz_vM!#Q@x~zKnk6*7duO8f0;b-*Q{;#c-
z2WLoh50~A#R`=n87S*wnhxT8UtsR@zvAwG1yASkvbCZgy@-kBCYM+3G?zM3-8Rb>M
z%M>$fQwFN1bzPer7`n98A7A#U<=(w=E!{I1(QB=A_`IsAw>>cPk~{nSQqhT1f9r2b
z@${MQxT~kADZlPyYg@+F;pqObS<ZV*X;@3zvnfSW6V44yx!rf}xgqTY)nc3(EGs$~
zNQd^;)pozQXMS9MYVV@0C%4XeS99rVXi@owYh9}7J-XKM52DVREqzngPxyCh_%rVJ
z*a_($Zevf(^pP(6?9!|OEB@+po#^;~kM(<}M!EGAUi7g%Ie+v(!nT5z<@v5_wJ`^R
zS9hm>{882VqnoRS`r}6u>bKqxo!08Sy*@efrRY4}_tOh${1=TiL&t_hzdR1QeW70R
z&5>1uQ?;$(<z0hqZ|ocK_#?XRRZU)yzCB^*=APKQ#dnOA==B@wnJaRLUkg^O4{!X!
z$N%lS4)_vxFQqx!zd>Ap6s*Zz9D*W<laJ}+xWgJ>XDn5ZWXYZ65?jOH1GbO!KQ&#Y
zbEYLbV=26?a8kJ08o_ou{gGGESA9*+n2xt)PBMwD5$x!ZA!}{#n$yl$8gF|#d5*C)
zg8lx%@Y1ynDPvubblwhf42rZhg1upqdR<^N(mobR=522e@5#1Cu*<J{X*+JURXAhw
zc-z;>H_X-u_L(Zl1t-Pn{mxhtZ@W3UiEWKw_bu|h$f>O=XDpGoJ)Arwwnni3a7{bg
zx$0`KGd7pE-JRSewnnh?m%dvdcd@UYRBBtbomWO`>K=ysx@6&cJX4sUOBSIgfFkuI
zP?Vkmiq_LWF?t3lR?h;(={cZO{wiEpPGQd`N^N@uCS|~L?32Qcco9&=_-52D2?GJ5
zgpmMIqL_eD(m;YJX(U0EEGA);GEg8&87UB@iYXYS4K#?-MjAxvVj4yn0|TOrkpWSr
zn1NB&z=9}iWI>cIW?_^wa3IPVIS}QFIT%%isA7c0cga%4sS<$FsWJs6oEC-Dc1pm;
z;5cPpdt8lcfEpIzzlrRczzL8g)C9;9!y*D^Nt^^(QcZ#^IV>V!mcl8JrPLJ2Qo|w&
zW@(%TSz1klEIlluVV1!ekY&^i$TGts24-2D1zA?jf-E~MVquoUIgsVl9LRFRA`WI%
zq>2$Xf#0AYm0D234HYW2-ArIhEGT{(90<G<#k-)kqWA#RCJb-D?AwH!G5jXf?HJwx
zb+ZV+1Gm+J5|Br9q6Fj-R+NA|!h{i!N0>1J@`!effIOmEL_i*CK}pCXJ5dtyNGnQ0
z9%;fz$Ro`d33+5YMnWFhEFvL~vY-^?QJp9Sd6X5UAdfO(6y#B6jDkF>9it$RY8FwD
zM_W)D^5{;KhCJGe(vU}+FdFh`Ge$!m-Hy?aM>mUT$YU%h19?m*%0M1tMH$FrOc(=s
zj2UAfk7>sk$YYvC4CJvEl!ZLD6J;TfwW2KKu_la#Jl2e{kjJ)TEab7xA{O#E3(7$r
z*NJkF$5~Mh@;DR5K^|wuILPDLF%I&$W)WxS#ccZpdG)NY19iyK_-oRI)B294YLy4K
zV;Aa@rI@aW6JB&F2{PjG?g&SII9L)qUP$LClAABjTECJXR*6+_3d6!TDz#M`!jWG+
zC*H#7Lv(!FJJvfzC_aR5iPTnn2x<NeUxwnl!398+p#;=q8A?J;m7x^WbQwxR&6J@G
z)NC2bLd}(-oLz7F(6r^d?uQMdvrY3h2(#Sw>x_w^cXoL3!!lQyhw#D?t|d%rE0OGf
z1YF!N`%Bl%dszK3|EzDU?@ZyDx{|LQb-72h{IKUZPY>b36v`{U@)@nI?7tFM;(@}m
zeC4AgfFhM7P?VAaidND<F-is~R>=azDLJ52oNB7DrRNDr>aN>|pYU^ab9I{_%&}3O
zl<hd<NIyR;50G<0$@}e39~6W<emO_XZ>AXc2a^^PK#^h+C`wELMT=>m7%>ABD`tV>
z#N2CjB_--fmQULaem)@%A%4Q_djcJwjT!$&1wR}t5B3p;g-X?3YO7GeN^9}i_DSP9
zJlU=YEde}2M*xr1lE5Q%B=9IL1w2Ye0gu+wz@v3E@E9!vJVwU=kJYlkV|6U>I4uV}
zPR9XH<*Aw|ECPRoC#fb2s$X1JiG`gRy$3u3JOYIk5{h>Itd3UfV5K7Qx4_y&5<roW
zBv4c&1r!}g1I0u#K(Uc5P+TPUS|RyMS1cO+PoW@xNkQ^xXZ~pO=O<I>Ac*00SClm$
zseLf-TuZ;x!VE*^`~^tHYY$mq?=G}=m*i%z$l>Ljci#~?I693HOPpQCx{e<=+igZ1
z?k1fbD|^Fz!keDcL&GB?qoUtZ%9$|#5Yp%Wje001dizfHn-V}!G&?mgNHt9}N$tf2
S$ICrt{$&xSO^BI=zWqN~J8Rtl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp.3FB0D5B09C446777.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp.3FB0D5B09C446777.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1063932150e1255b5849ee8c9dfc2c408e2dc50a
GIT binary patch
literal 4058
zcmZvec~n!^7RHmCOAHVUiGW265W(Exlp&f%P;n>{WYRE-AR;KB6dXXMu?{FM1QY}m
z`%tWaZ54$lwxXzYE+UGcLTeeVGb&LOv8CwqUGmm7`;V8k*g3zmzwn*A?>T#&ARo2*
zj)=qYT;lU}a@^_|oWtRWnZMN4Da)SoIGk8h4(IEVb739JJKW5FYwitJ#Fs4jxoY!N
z$K{>;WQjvSf%j&)w?vlH`X;vi!#~pJh1*+R#_k%IE-OgARy=W@JMS7NDRAkH@hgcV
z4}Yp1s;-~-x7)|U15w|VZ1lJ>cXZ1^>7^lJs4nH{#aM0i1qZjuo#LsXt2SzP;osUd
z&$+OvyzZ|R34?F`Nh~<<dGpRgj@SB|pWawMIPfTb)tU_j2NvW%O1SuR@Rt!y^W$um
z6g@5crY63stMta{`|Dm5C0tfj4Xpn`aqPBrYm#Nr9)IhygTIdM+%y<a6_$5amyvli
zL%96(?{B|vG21%>DT6Fa2Tf|`Mn0|dd-cVc<ioqEmM6|`zTL?&QCC{V$Axfqo++(4
zUR}O_0=H-Em}8aC>K=5R6}^&)UpVV#mwGh>XkybVWxIAdN4A&a7XoL7*dPBhP?YF+
za-@+6_p{hL#kS`=yBRi=;MU}hKUN+a`6_F%#_W!n-2#hClc#_ANUCi)+~x82EulQv
zX<LG$->A(tU!Ba&wOw~kloiG=yPx;wwf+0eO_%phsrx~))ucUT-qqcCn>7c2Idsvp
zOyf8Ec-R7TrZ!{O_Dh+`+cP@z6D$g632d{I_g&^3QSQOoZ$vmhzS8~;S$pT;#WL?x
z@&I26<)!QtZ0!*aQe|at&1-hwes;opL!9)nUEWAp`0T27pN5w|*M9iC;&$uR<-(-e
zi4AAJdzYnMXkQr{SentJ&2}*N5C=SN%<XlUrtWXZP7gR*5xed_Uu*Uv|IB!!Wqqh_
z$6DKuc}wx4{LuY-E*a4gckfHJ5%?kB`VP7xE3@QTw`j1it6MfY^ru;L(CF@WJ0_G~
zcG>f^|D{&ly!%K@eBqwfud45awc@$GF?C)6zxg~mkr?9EedA*Mt<0RteBbLOxp|t3
zPQ~`_A7m?6UpwsW-|*wMUFVE*L$2Rjb-g3nx--sYU&6zr*K?{{`+t9YvhKVnHE#W%
z7Bj+578L!VJiVgEWZ>h;fpN(HL4<MRl-v7TGu&K$^w!=!|EfD+|6Tct+1aV#>b>@V
zJ(rg!PgK{{mspy;cX$6+@A0Aul|fbJZSS|-tu0;ndTy^!^|DCweiLtJ_RS|@3C}!F
zmu@V7^^gc!yJ|4bN4HPiUl{S%k147>p=*;TwVm3L{{Eiu<DsYn+3Qptx$lyGNBeJH
zP5pWPJKrfrUaPMDVLCm_zxD0Dklj^xQTv)xM=QF|9Y3L29nf`Y#iG!_D<6++ne`Iw
z-8$x8dVTt`^EHZJu68$FC|Hta@!?F3U1)#m+q7zJf5U_~I=+9MudHYBvYL8M%Z4%T
zs@02!rulB_J#c5ISiUrAly1ip%{!6muq~0QO^#J@IUK%SZJuY@BzsdVG@KJ{$+tAt
z2c{?JKPLZ_c4zo~5p0TuFligUjaVPpv}lfpo6mZQuPHX4Nz3@MS^B`H#mBP8toirz
zE2fx+NlW+=V|`%LlE@7eShh4&$eG8aU5s43^?^-GzrW*txWH0m#tCN9QY)#WKCtQO
zL$7s*pJ<PnVqr|$nr}TqAJ}wtq46oc&9P!rEQm>q`C@@Su<7S9_Gg@wx{0RPTqbSB
zw=&iTHvPdo?5D03zdkX=0-5wE{wQO8VACma@u??`4b>ZEqOV%+<)PBb`%~<N(S(}t
zVT;3%h6n}}PDC1rRFgnSYDl0Y!%3i|)D%!s8VV?>a0)1CH4T)sh6YMHoCZpyMnH))
z2q;lF0!oFoV(h0E7Ok*V$N)Jg<Sel}hAGMLb^#@cB9<B+n@A+q07@gaa1BUCkw8i&
zl0ZtPkx)uSQ9w#1Qb0<jQBX=p(LhQk(m+b5(NKz_5Rjro1f(bpL8(Hen8`l7Fk0cM
zm;*?oh-8U<I?P0WL=1IN2~=uU#2y%liT<D%M1-86*#E%nCnYfnP#=%z=NXuklR!%<
zNuVYBcqG(PatdfEB?YupACH1sT22Ekt)zjL?&Hx=i{uDskrDwd>f<4(RgenA7J#Xs
zLZM`d{pOnJ8v|PowlOfJ9PSG2v@lxbtqNs-!?p)U6Mb7Sh=@imaT9DdqKiv(L)LML
zmyl~Q;xuL`FVTPz*CDrIL_6dr9?=Y6M<bVneq<Mygnp!sOF}=g79*h_*?^JIk8Hz8
z=tnm3Na#m3aw+IXb#W=^N9ni}^rLDq3i?qE7zO>PHjILPR1=Saesm+3hJJJxmxg|{
zj!Q#7x)!6MAKie_(2s7zXy`{b@o4CW8o3Dip)M|hen`hf&=1vO2>PK03_(BCh9T&O
zns~_IH*;GrXYlRnb}k+b4{E@q!WC9dJMy{0Xd5>hPqyiW%#jd11yhZfX;))c5xW-`
zJ;+Mn-?Q&FVvg<rkJySEb`v?kBOy}*JPI;Bz@s6f0Uk2&Qk&qWbF()tXZ#cb#b>PF
zDx6bX;W~6{*e?eszJ-1{I0>1`!70de4o*WxIXE)#TEAJ1C|sps{G?vedF(;l*5@_W
zy?MtNW?Et{v0}Se?!%I~PMhY(4KE`TC$iu&GI0_zm5Ec3=}er4j52X#;G-`xwDa$m
z<S~9Se_0?~&C0LO8flWt)y%Z@NNX#$1F$!N5hnT@0O$nHge&`7VKnJX0-~HLK(sRr
z2stA_3MYjNTlf=%acKU{vS-W+#O7iP*7>|LXo-<m`3q*+a=fJzd*{0NU|p)J=WoB^
z`UnUIuoM9ah!RkMXaNlf3D76H95U@$wB6l$#zl!M$FXbr^I`go0&#T*GcBDjUCh3-
z$F4?iM~FsmXQo{(UF}#kdoN;I?(n6EE!GP`<xXsJ_3(bT8gJ#ws@Q81-$Z{sV%5y^
zcn~wd^LUVeC=Ut{?Lh-V9_SPO=Um5w5S1=$xID!;Q4Y6HF-}6Jig5}uU5wL^Q8A7T
zoD!E5GklXvO<9~8JUm!19RBx-WCt8xbjafX%av{Uz5GQHQIQK4a-u%vI1z^7Lc?&8
zVYoPXNzzgV{o(_7xDkJZz}Qr1W-c1(Blhx?i%opKl#Cu_@tL)o`;@63)27dGwnUR1
vRHWVio3*kTE47!6a~w}l^ye<FiiuxLvQ>_8nlQ`K%G);vQ_q<>o9q8yKFlbV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp.7D14F328651204D1.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp.7D14F328651204D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..440b0cd74f175e5360a62d83b8193ed1e84b23f3
GIT binary patch
literal 8982
zcmZvh3v?6Zg2yv6X$@@!T4;H<KuZ!p!Yj$VGw%`FLIZ`8QUg*zAX`KQ4EO-AKoJGG
zfOrqCETEzwDC)8hP<$e6p#<=X7`#O)f;?`dtoB|Iyt|=$hy3@P?KylWzxjUT*Yizh
zGMO1W(lpx_+icpjkqc%|pEt#Bv)OV}|K`qnY}QT6W?PqSvn_b_K-KJu^GwJ1rhOGV
zw)cOeXU&QZ!AqVq83SI<FB)`uLrGOneg;42z~sf}KMJ~!mcEofcKw2j@6RYaUh|5r
z#O6P`JHNlf^};&g$r~TJ@|aDF&UZQe){O55-$>LwwR<bG$a(x?O^=6uS@Qd!FHbhV
z{KLK#N3VXkEf)BMxpi|{fwcI6mYd0AD>ualR~&ld-0iCuUOoQFrnAQ$et6f7%lWew
zWjwWUG=FYw{vNez!J4dHFYnwlaNYx5x164|&^dd2&ESIW3s=N9KlFlm`KA8F^Vj{*
z{->z4{nY2}_MQ#J*b(O@%^X#6CcLHHZgt(;IlDuf|2Ae`eE*F}uhlG9HwalBJ{$92
zKM%h;Ah$#R*U1H6cY9~mvW$^O?3<sS`Oyb&RvxX`6e!Y4z9~HOmAq>E5T&8;gZc*o
zY~9F%oo|<BRLjqAm^p8b^UE_^I{)EZ=dA^!?c7gKp73|JZ`=4z-2(6O4hM&B%euU@
z!@Fxf<0N?=zkk%$J=0sZTnQhV*naSci*FKBc2-Z=bNxi3srSTgOG}v}hs~QsTy|OW
zOC#2oj9cq|;OxP9zb+kcLCh!|wv1cW^<Vn!er-t&6CV2Qd+)w~y2t)*%S!T|bwz3p
z4|(Rus;$pFzJKBymmf-s)6~mK^NH7M*7jSU*Wd8|#kTZNxBWA+^0zME_T|bK5BIu~
zebxQnkIZv?rq-RRDqA%9U~<C9+2ynTdgRkmrqi6$CkuxR9=6n2H~rG)>wOOsGlkq$
z7w#)RV?6cT#eeTb?3h^j<-8;L&*#0|=kc<fv5R+g9lloh$9J38j8m)rqw2ST!IS$7
z<~01dFSDK>)8(W-_0q89lk0ZvUf(OLTCbSa_wnYZ54(v0x9b+Mvp045b>;ReGn~~c
z4rFG}o<8!#r-cRT%i}icqj!`pTX(=-N6vTuXZD6S8WV?dUl|bny!*8uIx^g}uZx{C
z&VO<8>~rRT3)f06CO6p&*EU^D45|5O-F<)iX#QWm`L?ysghTfz#I%J?@AcgA{H=*Q
zf*&VNkDe5IzO3Jxy>Blc_xH%Zty;aHs%Y=Ew}0H)yuSJ9_18D9*?D#8m{UV%OwU-n
zy1n^i-Iknz;|5J1b|9}`eA~#|!|yq8{r=+4H~!wSYhqetex9^y_?6zj^<C6w;7?C)
zI@o%ubua6i)3b3SGh%M!{D+?>oP*cad|0}u#NP7#qYF1aUonl{P`CW4VV8G~a(?k~
zL&?G=OM2eR_1m;y;TO%vwyaz`WJ=xP4*l2l``ygwckB21J;h&tQybmBrDV*WFE%fe
zKHc$}a(yCE`g8Uxo4ZTT+`oF(`+G($SpCU6i-(@Pw)u&tON8yW|8!~7gTFl5X;$wu
z?caR%!q~4)pNgy|4&Cy;T{rWm+fx_(<BarrQE~ZKuev^NcWl>>i`~_OZrbK9oL#$)
zu-P(;VlQf2=-%1ngo(D={LK7zZA0oD(e{(|EH|V11bzEn+xXOUVL@SW+i-XKx6Nml
zl=&*N$*R<JkIWvqZG$zPyK6=F`R`u&BAX1Rrt>oM+O-YVv~=rwa@)m)yRykpYPwry
zw|m+KYkKk8n$FQjUyNmwHL2-7nSIo@!J1z8qxZz`^P5-PYpYC6_bTYsr){vN-&@;m
zZ)T5O8?wo<sp;IzTu0krP4CO?a_Q8hYe%!m2U63XOb^#KSkss5ioQL#Iv&iiRi~y)
zGfL&Q!J57@rT6|p-sJFXvLZEIkXg{KZLp?q{OPSncDiTf-D?|{nl36RDrp<6>EAn6
zeb6-L;KgimOlrDQW~X*-gEeh?Z2CiUm-I~U8#$|OpRD@)*5;kfeD9oEm+2a5txk%9
zu1cVgt0qlu(+#FO=mygra)aqJeIu<!$*J`PeU(5VUrn0)rXNgy&=00R<Oh>BX)x&^
z4JI9;!DLJZOeV;H$%Gg%S(62m4YFXeAr?%|<iO;D9GF~)1Cuv-F!>-4CLiL#6ifk3
zAt-<;gaj}}Qv_2CieQQ%5zK%Y05cE_fEfq{z?4i0OerXVDTO33Wm5)I4$5H4AsI}?
zRKQe%3Ybbr0aG<qFx8+6rW#Vg)JzRbEvSL1g)}gAQwLKI>R{?29ZbVCz%+sem`2C|
z(=<&m&7cXU88X2v?^fRP?vY+wURd4-sJPr|k#!a%v)ayn=>xH9MAi7a(m&6IsiXDI
zi7+j>!mgR=BhnRd%?FCQo+Z-g4!c2hN8F&gqi$4vVc*R3G3kr=<^x52&k||$hy9@X
zBYsf*Q9r75m<E-O(4f*$8dWCDfXYM|P?;!$DjQ}&Wg{%8Y?MWn3v-}y5e`%?%Av}K
zc~JQX4=Nw!Q5C`hs6s>lRfr0xieV8{F(QI0MnzNu;Q*+CNC4D8G=QoUmOzyv5~xyC
zLRAjSpvn;$R5>c6s)Q9#m52hW5>-%D!z!q1L<Lois;FvV4OA_nfvQC{RQ0e9svgln
z)uTG9M%Vz=h!~(6Q3F*oY=UY=Oi<0JiE6pOT(J(ooZ51&d^Aw7yv8EyN<_?Ry8?lO
z)9LD;KK@)@7o8@z(+zWOuN&rkPGA4@5$5yyXrLC#_Y0L~ztaz{-|Gk0-$MD(rJXdm
zw3h~#ZlP#&87BiS<7L2QS||ox*2#j)dRcJU7K%leb8_HvUJhKYh2qfVojka_mj{<`
zp?Gu!rvR?t6~Gl*C;?s3DS|6{MR3IyN<=r{41gQ(2EYxpPyuu$rv$F#mB5u+C<$HJ
zDT6C}WpL#dN=8?4D&Q(!1ze?tQqWbMD!8gw1y^mMRCG0`2CnASz|~qP4PD)-gR6UW
zaP<~SN7rx~;2K^7T%(0D&^4VVxTe<x*KDCobj#i41FeHDbwCa+_gd6#@2}gQ$gNyW
z*lUBu!O7NU>r22YBCGA{9}dl9gewjoj;^zW>wDy%3D-5`eWYtYnLd17M@ZLE<TIq}
zJLKb(>jahd-Ny(w_S|O)H}>2=6K?Fe_mOVwxsQ-;?77d7ZtS^_Q*P|}8VTPq_?Y%J
z5x%p?Ny7Ita*XutBh#l9Ujykof}9|IXOQER?>L@@8VNu4{7r-(d;TQh$DTh%`myJ4
zApO|$CrCf`{Bg>UJ-U&gu}3!%H1_BuL1T}Oku>(`29m}eogiuK(Q%5#9@9uL*khUq
z2764BV6ex;NCtaM1Ib{INstWom^j5?k8LDa?6FM*i#;|;u-Ic`B#S+^fn>4ACP)^0
zY@A}T$2Afh_P8d3!ycC;IP7sTlEWU?KyujQ5+sK`E>3aS;~NPcdwdhYV~<Y~Jofk)
z$zzXiAbITZ36jSiAE$Wi35|q+J)wyZuqPx50eeD>6tE{WkOKCE1Sw!oh*JXg#708I
zp4dc)*b|e4h&?ezir5nyND+Htf)ueQ#wihdfkq;Ly+9KYz+NCp1h5x~kpb)l8pr_l
z0tqsJy+E7_U{7i!B<x8|goHgQNl4g}Vx)vUsezQRCnZP;ds3W|uqQVXGWO&qLdKq)
zBxLN#F;d2!+(63MlM|$jJvmOv*i#w_1$#;pp<qu*5(@T|7^z@SX&@EsDG5@+o)V`N
z?5T}}iaoW7P_d^b2^D*4j8w6wHjpaz)C8$wPmNP5_OwPq!=BbeXxP(|goZsWMrzp8
z8b}R$T7uNDr^P7^dwL_GV^41)bnNL#LdTvSBX#WQ4Wy1eJwfW&)8mwmJ)@B@uxB(8
z2KJ03VPMaQkp}jR2GYQuksuB18F9+Mp4mv4*fX056MJToFtKOGNE3Tz18HK<Opqq_
z%s6GHd&=^^+HU?j?+j(Ht?X2Jzje7|udV7-1ytRs`Yx4Lzxn45+rG*pa%%gO_NlP0
zOh(ab=M)TD`N!1e9!w7{>!0@-#*NN%U0Ne#q&AneE9+nd9{l0C8JXYrKAV@iiPuW`
zmfF)74E|Qik4(2xG&0jlG01Ey#UgX96o<^WQarNIN(snfD<vWaTB!iC)JjRnaw{bx
zE3K4*thQ1rversz$a*WKBO9%hfo!%?W}0XAsEi*NvFPE{`jtD%yISk_c$wZfFys$X
zIkmlvUK6al%sI7%M&ShOMo>;|kx?|ky3K3tsMPJDJ9nwmSFm`U`-i7vOFui;yf2mO
zRo&}B>wTBE7(LnJE4!yQ=XK00u)e3RI{#|LpvBeG@9f2TyYD&Ni}iLtGF@+{k(qis
zgUr_3S!Ax>&LQ*lb{<)%w+qN(y<J2O)Y}8dQoUV5mh0^@vQlqXkkxv-imcV!HDtZs
zt|J@ub_3b0x0`9U)-kpAcOC(x*S>N@Wu>)GcU)bvu)6ux7gLMuQ9butN5swrA7%t+
zzh<U3cWd9R!1^X)Jvk`IYI|akeoL3fHN@K2Ikj$&8_4JJ4YA%gr`GTB1JNEDi19E$
ztcL~SJRA`3;eiB?03>=upnxX;BzYtt*&_of9tBACs6d)W1JXS@kl`_aOpggvKB#=K
zwbZ*^?Z*G{*p|zwHO=jq+u2%6YiaCRZI4;fYl`Q_od0uf%&En5<Hz2y(=v?LuBP5M
zzhC|!>#^0RKRhwlf9Xb9=hR~mhpQO&j>8S)bNGs_Um>-39DX3$K?5-k28eaAK%9dE
z;vGDY;1GaBhX@pK1b`%m1SC6TAjP2osSXuLb7(-iLkBV(29W76@6zYv0+%Njojz>O
zsjaYA^tE<VPVHFxSfKIt@jxMa=q{Z*|BSpQ_r32@3tv=SG}SuiT^N$_>pdO1ypY;l
zmS0w6z3;_i+YfAgYE(sPb76I1t#!V%rc>YX?>tMhcE9x)Ew%a1Gc;@SRmaZNB=_E(
z54?hY1w$;$N~ivY?mSYn?9`rLO1Rd-jR5yj!i`)#qPoiZ4XkHx`B`nx+~7?_mkfVC
zq8rF3`egX?5&b~4NCPn<1H_6f5GQg#yvPFyq5vd{B2YjK07;?*B#SbTA}T<tr~+xC
z2BeERkRcjCrfA-!)U(LCDR&-B&V6j^+{!yICf^zU_h*vz!e)Cg{W6?>+4eu%kgC1-
zR=DH8HoUbxb87g(Nt378rlz&2v!~U~OkvqWqbPeu<~@#f**W)i?9hH>uAw<|vqq|U
zT{?B{+Kpj(K@3Q;Cx0MU>~|Oae_IQB^z7BUPv4S$uKok=8$4v_a35Xh?LKIjo?mc(
N=`_+Dt&AYb{|E3$-s%7V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp.8A9B71EB3B2C8B82.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp.8A9B71EB3B2C8B82.idx
new file mode 100755
index 0000000000000000000000000000000000000000..217d397ea63c90a8fc77978337d8380fbc4db149
GIT binary patch
literal 8880
zcmZvh33O9sqQ`TW*0hwR1&S3)Wogjdecy;sp=qHkv4v8=C}j~00?J~+hb)R!A|Rls
zeF6$PK17k_fff}-6vki)h^SLB$TBh@L_4EpaTzD=IoJN*Ij`q%Pk#A+<k$1veqU~K
z%ZCJmTY7sugC-A|KV!<=YTD!RWT*bknLGR5>y*c{sGY|%fBnJAlQU16oxV7Fxk9>U
z{pA0BZE63$s}kuma{3HkQ?k@olG7`CV(l~eS1;9UDO+zQH_v-ASL_?_G5h@eGiQ^l
zJ%3|Af41!Oh%QqXT)B2-Nx%E8-&e<H#)aGB$Dh1@ZF0#2zudl2cu;TXR`R$LefnJf
z!LrU{jH`D%FvoiL+Hb3JUmS4c_`yW7f8g#rH>@18=G7m!=UzN}?)vY)UHQQdeBe9s
z)sm~9-+RNII}cQTl6kc7YUE(#-QaVxKWaY299_08a_M^cq;Z>-EseYG7$;9zHnZ;!
z4=noQ>2X|s%?sNeJNYu*v;NDUUrBWR?}3b3KINtBIeVVkJm<@US2Gj7E{Feq$Na2S
zYp<>zJLaGBOY^$!e>pyF2Fng9_j!_o&aKlnjC*gvPglD1-g)KNs`SV+-_V5*eR9+6
z>z8*FW*1)DleyWmJa5+AS;2|PP}9I2gE#Fxw={me=kE4tL}V({ePhF!J`ZGVz1=?Z
z==8xm;vG9BR?Vz<Eu6P-%R|rYeD)(|?cRHrbf2!To6&#ngKs}knZD`ls(<Yqy{Kqv
z^}!8i%QJ3ITmH>=!-tiO+){q3;&iWyErlOfF>iLw|K`Js(GP2fuiSU)*pXZMuIbdj
z;-p<mjGD=&?f(3WI}azn{;yB!k3V`~Q_0)C%ZBFMTC9HKS#~Jzo&Q-DJwE8f#Nc1g
zb!dKU-ZP6$t9C-Kw<50>ez5)Ql|8TMuc{Bum~dpn^0zNd>-hXX@(%2Z$#aj*-+0cW
zZhYd<?DN3~vkNv~pIEYg>zAXX4YOCiw`}wiiJV7Ie9IhK(f6r8^l9|E)uXqq`*-H3
z-ESTo|F7NN5#O!)s#yK(#&?GOdg^@HTe}AF(F@CmeatVsaAV=X$6n;#d^I^Ic=Y+L
zi|$#w?r+m~oTz+;ZhyHtRy^d3A%EY!b(HCp?)QZrf8@Q@Lk_Q>(PLBjUggXA-5%c`
zlV0vAS=|es>JU`RZ(Tf9n)OHDrsppkM{66;ZkVw1AKS`itPB^tdidJ@TbHogyFWW=
z^k3fp(*MkW%hLs6*x8{O!^VH{@`-)+kwg9Nt$gvVAGZ!Gd%DsWUKtxP@a^-=!d+(0
zyba>k0q5V?a{9;5KD)4lTUY(q*2hX;{p!7stvydJKC`vvr^(C8o2GABai6@$UjE{z
z*1WC{7k;~D->2`_oc-mO^B+BOlilIch#s$`?YZ*snfspZ7TXZ+_hJrt{Etb^fmasx
zJ{+5D^m=;8)EYVO%5Yz-WBu3Tj^6TC=R;S1E0^y$K5JB{!{SrNekeRS@e;3{=n;ME
z<oIjG{LyO0`=iJ7ZcbZSy7GgiB~RV^{?6go@)kRbD>i>y)STS7xbo2U&+h*4{KysS
zm(F$;WIo%jsG$7yK>CdZ$Jt$*@-L2D`pc!JM;5IRi~4MMzMuNi+CMy>{noG2GiSk!
ziE{~$hn)Pvfx`9~1?@ef#(Am>atr#l9#VTm>rZ+?{i;EmjGNn$V^YIi)4OK09^B!K
z<b{Q4*5#b`o{H3Per|rR)`L5o`O8bE9xc-<+L4v1;qK|(vs(}DaQ61)H_dzV`;Xd@
zq113rdQL{`!5!8<+1Tmw_uKDp?-`vM?wQ-OPwT-Q9`NfwlUvR&*xrt;N)6|w=iShH
zaEBizj^#{0z2;(;XKZSCK-z$ztp|5_$J&gy)4Ol4Ye$x+hO^VNy{!j#cyD&M3t!yz
z<EQP&@u}fH>3!_hgFF0v_08{aYEo%Oj!X^brsrm~9^B#IJ5}yV%=-9zJMxayaOd>S
z8LbC*c=nWg<^+1Q%{`gfI_E17{QA-xXPld}s>@2tD&61TJp*=T>&&-F{-A$|`)g)Z
z`$PT;peq0PHe!Mdm`sQPlc{3BWP>c2Y={MutzyCCf*hD!hy#<W;=tsCJeYil2a~Vj
z!4!f5m_kSZQ>YTa6oVp|Vn_s2tP;VLf)bcgNCH!;lE9RMGMI8m22-w*!Bm0@m`X?i
zQ>jwGRD&v*YDfiBty00%f*P1wNCQ)=(!kV%I+%J$2UD-o!8C#fm`2C|)2K4QG=nCX
zX2=B7tTMs0f)<!o$O6-<vcR;1HkfwE2Gg#x!E}NSm`=z6)2VX6EXphD;Vz7<>Y|=S
zy@2`@`CM}MrPRu5*_7HAO>t>)g)6zcQ)*?kY)x&FKkT2@wtD;#|2&`?|06^jGGPW(
zCc=Qq)G(;BVHQ+2!h*`yu&8oj4pc6}fy&izsPbVRR6fFk%GdCy3Sj|MAtHb()Cj1G
zVG&d@B7!Q`h^R_o2~;T}fhyHVsLEj(R5>DpD%Z%UDq#gwC8B_;)F`N`VHH$0qJpZ{
zsHkdT4OA_nfvVMLsOn)IR6U}Ds@LeK8es!eBVvGR)EKCmVG~p{VuEVcn5bG|3sfs&
zfoj!QsM=v0R6AmWYS-API$;M?C*pwW)HtXX=|x5EQc10~K}92gLPg_Ua^J<R%+_~r
zSVlhot!*pCulfz3Cd&U4)y9mE0hdu3aG55GL6`Nh;Ib+UF55)0=yE;|Tu$Y{<(eoC
zUEaro%d0%Nd=tf^EBFL(1yul7XrcskMV|<+sEXi<O_YeP<deXaR0&+EiIUKjeKNSR
zDuXLGQ8K!UPXSj^6>ya%N<mlkso<)r3a;8jspx7x4O~stz}1>44PD)*gR84LxOx+%
zqigsKa1GS}*Jz>)bWNWLuBn>fnoX36uI019wNwjStBJDEwS6|YwrYcGH&Hgaj?V$t
zQ5|rdCdxs#=$0bJT|B9UB^9YIxmUcYx2=}T-L|DvT3R~VmE7yz#hI<wyKR!cf$$%J
z_a=XW@Sj3X68;~M>q-A!vTY6eW2FBuawF;g8aYn+kK&qbAQ<d134*~MlO!1IG4&*a
zJtjsn*kc+=2764LVz9?H5G?lC1i@mDO%g2j*m{!19vdTB?6Hj`i#;|@vDo7p2o8H(
zg5a>nB?%6DTs_HQkBgBU_P9op!yXr>IPCEa1dlyFLGak)lLU`FzMkZ<$Hzz>dwe6w
zV~>whJobbJLcpGoAO!3QNkYJ$P)`cj6Jn%*J)w~luqVVR0efNtA!1KV5F+-(Bq3r?
ztS3e6i7`^dp4do=*c0QFh&`!+kgz8u2nl;ql8~?`)sqtTq!=k-PiiD3>`8G-!k*kf
z$k>w;gp55oNyylf>q!}Va*UL*CpVHZ_T)GvV^3)y6znMpLcyMrBoyo^^`wG5B}OXP
zQyNJHdrF*Au%|W<D)!U_p<+)>5-Rr8dQ!!n8Y5Ngsg0zHJvC0L*wY#a4SQOG(6FZ^
z2@QK%J*i<&i;)`kv_?|Ho))Jx?CA}Jjy*j==-AVfgpNJEp473Y$4DJ}dLyZ0PmfbN
z_KXI?z@CvH4D1<6!oZ$UPa4=WVx)mRqmeYQXT&K3du9V+V$VzvCicuEVPem$Cr#{`
zG1A1I*+`n$GvkzrJ*$DRuxBL*3wu_Qu&`&<lNR=@7-?b8Y9uY}S#iq3p4~v$*s~La
zjXgU_*x0k{NgI21jI^<5H<C8?>^Nm(&uJhW>^TX-!Jd;O9PByuq=P*tMmpGY8c7Fx
zPMmVudPR?4_FVsY?$;DuUEI02#Jw7zt4lkV0+n?x`;&$nYiH$hD^?R()!ohRgWS)%
zqekTPyRfEzTI%V683Q}IuQcK86I0W_z4=s5>dT~N%D;?myE0&!DF&HsrdVXInc|T7
zW{O7^nkfNUY^FqHshN_H<z`AoR+=dVS#73NWUZOfko9ItM>d)%1KDh*Ok}H>vXJd&
z%0_mYDW{F6b+3r;D_t}*HGf6kq8r`$dw}-Ty(a#8tY!Y9wEsDrzbMThvr(Ew=Atx*
z%tvV+S%}gCvKXaBWGPBZ$a0jHk(DT|AgfVYMb@IUhO9?v9odM|2C^BYO=K%dTgY~l
zwvnAE?X>ZOWrG`-Dm_GHRp%Gyk8!`N$*S&I+;fb337l13P+Ty^{dU@&V<NNlOI-Jw
zcj?3R%RW4PW^d|D<d4mt;J(u1O~pN|qbqJoJ)P4jC)fRey7Jumk=)|3Q(8_?E$x2-
zPf#t*AhWeJi_F#195P=^^T<LiEg*}vw1_O#(h{;<OUuYgEv+D{wX}+?)zTWWUQ6r9
zMlEe1o3*rwY}L{hvRzBt$WATowDIeg9$hf@%=%TS`Rj9gpAvT;cw_#%X`vaLgQ=(U
zI^^ZLpF`Yl(z|4~euM6wo75!pe$oBES=9{B0I@s^#PJ*u&+|Y6F93<W2qf_mkj%?K
z3a<dEyb7f88j#NGKn8CBnY;;P@fMKH+dvNQ02OgXg8N6iyI;oW)kij6OwC(%r|d57
ze7RFYXSUv2Xqz|O-n0JC_MW;k0NeZfYWd`};?0**uiLe-E9c%Q*njq+Ir@cbWi5Z^
zUVk4qlR5)l28i{tK%AEY;=MeO;1z&GuLvZ0B_P==11VkwNcE~fnpXqTy*iNLHGoX7
z31oRKAlqvLIbP>aIxy<Ci+2}%_1DxMGngLS+dUOo)g^QZ&`^5lpLF`%l7VZocbrU3
zXTjKlN$%G3otXCX4IR5ZnR<F)*MSA@E1f^`+P=Rm96mDjbkDIpC%W5~dt)fIQMKF^
za!=|d?;B5%_Kuc!;@rYq(Y4(BM4dBR?+>|l>cj>71L4F47$7#l0&xKjh!5~ULO=i#
z10s+VkbvZX45S1UAT^)@X#ovL59mNfzyLA>CXf}dfb4(`<OG~Q>EHWKo*ktpD_j10
z>uCRG+@$Mh2AQp+S!Aw`=8*Y1nnxDuXaQNQqeWz?j+T(+I$B0n>SzU7t)o?Bt&Y}^
z^*UNdHtJ{t*{q{YWUG$0knK9!Mt16Gr;R<cr`%WFaz}B_>`8MfS{~fha%labs`(eP
zwk?RKZA+hc=iTAECV0Z`kjHa(+rvBC9!_X`II->Fq#2X%o0h_|dxleVTKWy%jCNV=
zJ9X?ZB)fQ!FFSL{?K#~#ceydow4DBd+Xf8eyE1kk{g#6N@7dh$J@Rkv)w^$DzgzvR
bAWDj6^i*%+<iTBYONLG&gCmMciDCZ%T%NF$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp.CF381829133E9AFB.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp.CF381829133E9AFB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..974952dadc8b3965b87a91dbe27e9ec74db1b3f3
GIT binary patch
literal 7646
zcmZvg33O9sqQ`UZ4JBQWF0=(oDQke-eZLn>1C*weHNzUf0A&$~ptAT12!o;q5Lv{5
z2ig$?l|^_0S|&VX8N|^@iOVS3I>7-21dPOiB8xgN%{!O=-#M@6@HM~t{hA;Ba&L0`
zji~f`&lcNkWz#Ab&Y8ZT%5JmSa+3e%FPJwgLD_6$GHtenHAlzH8UCG-9o>3(_*=Vs
zJ>Fr$s@A@SSg>%Hf6S71ADXejKljI)_`5g%_}iKr{a<5xur=%VoRZhMZ!A3bxxEvy
zf83k<2KF<?)m*H9|1$OBiHkRjwg$FNTC(-$)pu@B8$S2$kXQL%w7UPSSn=fZFMa3!
zV1C}}Ye#m4kJSWz{k8w<t)7RMEP7JDeEIYT^_$z&EDJDq{&nw8?AJ@Tp5Hp@#Y<yP
ze0Akrs$lQ_{DUV~k6Kv!aBXbccPAZ38$Xt3z16Vzz~>(gUXZ3vPzUWC(r))FZN7c{
z_jr%m@#gtMeKPfm^nYaTyHa_CR-PU+bHv7{3+OLOpO20!{PvLr-}rjIRB^tuG3eaV
zhXdU&be=xSwyyV%+5`K$RBo}o+j`B|vkQWEuKfM-$d#uW%ARb$OU!rntS>ve{&bHO
z?5fSnzuna-bD)^F=BG2S`}_Q<OLoonFHWxNdwEIghR||b?K>M=t-F+X+DIGS`Ok^J
zs*?Y5%axB-{AKqS$FfRC9X;(oeZEEHA4;be)Nel4I`%%&X8`?-FTLW<*S-GTZ$=ec
zle>Gvvx9P<JN?S$wXgc#cyH&4vNwG@v#0KRcFcmuTkoXPiVmr_7MJNe)(_~P7%=Br
z+ce+d18qJXHg>fV+jn%uke=?I9TqQm<lyTgy1cP6RP@xRM%gbzUfX^%yLR(uKZnbf
zRux|?8Md~wf91Z0VC4Gh@Ncf&YkIATwf)*L<3hZ519ho9v_E(-t5})%y<uC^WBY-c
z=PHk^Z2j4iLtQ72eq_()VTb$u;biYqkFM%>^?LqP?}iVL7bzPCeYEH7pk*yw<G;GV
zE*fv&Ij_gnE-%l_3cR`W`oA_FI#||u^$)9toZ{ZQ-nM(+=eDwif4^n!m@#CuM~)2o
zQT5NF2QN>%wLkaZ>sPzY8$a4rQ`~KG-mbEK6F=%UJ*ut!bkt$_?83qoCnDmGj-pvq
zvhD$|GNRYA8RBEhx+m7ZH|Nz>?|jfFGUbyAAI@15D492XNp_pQo4#sw`oNy>+qD@N
zAG^2srNT`Y##es7?`r2SkG?$PyS!Jw+!;UOS~P3n^xJn6pG60red6S<tDE0BxZG1w
zv4c7^B9y;pleA^xs@%(8bo+hFGv{va$jSKiNyiJbrP|)Bw*Jk0B714h@%0-|9~*w*
z?$x{BY<O9J;L@-TuXnHgIdFdVONEhrv2LDR^7gX1i~jul@M-3jA6G1quIwG^eE(Ra
z+nrV|`&?`H!Q7Ky7af0Y$flb``G0!*m!gK_(<Tgyy%E~ezy8+z%w<oFEkFAE_*Rua
zWwu}5-t$z&x>*Nmhb~-q{HUkPXV<qaKKj<yHrEH={P*M=bIYc8YKU%HesRR-Ek6&g
zBmTBRnKOH4MdQ?kzZ>suE9pAq^Igtk>1(!qKWUt!;Q9HB=1f^Y*lc9^%#GC(OCM-u
z8#&%qRa{WqIW;8T8B$+qCE>MYuj!pK$+5}l{IvY^)L>0#+}(EWnZfGFOma+ex+twE
zCpB2pId89O|HOY?eLs`*C8u-Ka??|THSM|gb9~pOMQ>-4{^WGKw0131gEc*C`GHq<
zOi2r5*~TTO-7VaGQ-d|Ve?$6T(u&^Rl1Yw8PUocMI8uW(eK@D^hc72yKb=WVNKSW6
z>)JatSkoU}_sy+-gwJhd8<m_cDJbcd8m#HBzweUWaHjORR<_Z}>5c^*yQT(f`f63D
z|7MlAoJkH(P8XyVq^AaJ`nT*c2V;+&x|B)&Avv9wmY1Fytm%2vAD;h2Ve{IP8L2g2
zbMoG{z32OM%Bm_GST@r7?X6cp$w;kyv*hx+Dy?5KtIFkbjRf+$CNvZ6rNN|qG?=uX
z29xnJU@|@iOvcZE$$D8ZSsx1~>u15_yd0REj{}qQb71mb9!%cHgUS1OFa@swrr;C6
z6#N31qE`e{^od}Kei2N`D}gEbBrqkv1g7kj!IXV6n6h66Q}HTbDn12F#jk*=dQ~u0
zp9-exSHaZ08km|-15@*BVCr5SOx>r0srz*>4X**F;WNNA{05k&*96n_nP8fJ6HHG#
zPY3HH&Z_cs^mGR5>Tz0Ry@w=E@TNBqIKcz1nawA$E9iOxsM@uhXhu3fgGvW!Q0ZzK
zRVKiI$^;ounQ8`AHo$_)23b(qY8F*4z=6sIIZ(N34plzDgUSbaQ2A;eRUsgNDg*^k
zg=zs+F(85}21QWCY7tc_Ab~0cB~Ybm2~{~DgDMAQP~~bFRVAQ+sst5Km1+f5HK2m3
z231hiY86#2pn<9dHBhx`4OKm$gQ^F0Q1xmZRU=@4Y6J~XjcNl`Ghl*h22D`SY7<qD
z=ILjh)X5XR%rgwg=b2!U^{FUlq&^Mdgm=1nHJ{Ti#iau!DAx_DnQ12tF0Ihu(g})2
zmvJ)SG71AOlb{%MStkoFtFYj*35rFRb8_Hv3I{Hipg44SCl4;K@Zj<Zibq#)3g8Ng
z0Irar1aw8G2(GAz;ED-KL|1Z3;7W=Fu9To8bY-UuuB^!5$_YwFS8*!fDvAQGlAsiH
zRi_HBs;J<q2}(s*b86sfiUzKhpfq%Krw*>J=-}!JN=Mgl8sHj=0j`mt40KJW39hM_
z;F<}_MAy^PLt7_#@&XV&ibX~DO6kS(t`%e_Z&X5d*GfB_$RRt8%!KR=G8?k9$Xv+I
zA@d<Sk1T}j0<svgi^x*QE+NYyyNs-a><Y3Pva85i$gUylA-j%jgzN^g8M2$rY~AwC
z%}Cwq!nI#VxX!>Qnkz=Q&LPJM*H6e{(sh_@UPf1hbbX9mPr4eAqm=6teE+B;XzbB3
zg2o;lCur=^VUork9U*D#(e)&aJvvI!*kkGl27645V6ex;2?l#im}IcWL`VjEOg+h9
zkBL$Y_Siat#U2|YSnRQJg2f&iCRyyU5t79oTTimsW1|#{J+6-6u*by+4trdj;IPMq
zNe+8lgygWt)sr0dxG2S8kFO(m?C~*z#~vRic<k|ElE)q&A$jca^(2oyK1%V}6Y2;7
zdqRv5uqVU`0eeE26tE{mNCA67Jt<&Mh*ARf#5zL6o){xU?1^zg#GV)?MeK<YQpBEE
zPm0(Rqm+m}sg97aC&dT}ds3W`uqTB{342n6l&~k&lM?o%C?#P}t|MgZ$uUC4o*XA+
z?8#wL#-1D@W$elIq>Mc|O3Bz$>IemUN{mpjr^E>bdrFv8u%|>w1$#<8sbEiuQVRCe
zIzq*s8Y5Kfsc}NZo*E`q?5Pn_#hzMEs@PMbl!`sAj?l2D#Rv_1TAa|Zr-exkds>9l
zu&33N8uqj(rD0F6BXsQPF+#_l9w&6{>0wgGo*p4}?CJHSjy*j}>DV*s2m^aYj4-fg
z#0dj?Mwm3PXGBN?dqzEJV9$tB2KLN4!o;2#BTVd>al*u&8758anGw>&o>@<t*fXP)
z+1%^5;z!$^n+qB!dsSIpS%r1mV6XD#d4VePD(}-!eZyk~><jCOtg0fTsLc8+a@4Tg
z9zVSJV2kA8()7~S)-z4MuyRJ)x1G*4-GDSwu4nNEq>-YLnMR61W*aFMnQNptWWJH&
zk%dM|Ko%P*5m{=aBxJdfl9820N<mf|DHU03q%>r`k<yWkM#?}o8!5AyXBLf&9v%3{
zN0aO0ad_HV>tlT<GE#eN)_uu;3?5KN`!yexRaIPGJl5JQ%&O{G-f^t8N0wDpQeHCF
z+E9CXY53W{{`36dWT1Fl@nq{qRwT+hn4i4RK6yAdJGa348|bZT&ny~uzGiLmaM%7_
zE3EsOy$g@D@XguoO&)IdK)V9#uF2Z+%gsn_^;!SR<ZH^g1nXzBs%VY|VmJnf<yat&
z<A8XM2NE~|NaREyiIae2P6kpq1xV#oAdS<2bWR5{I0MM!Odt>I;jNeG-5%+q*PnUq
zN^)&-vUA#4YhwNS_Keg%XY<<N1(o%GE~u<3yr7OLBc`<|-+nFmy!qYpS!>_%#D&H4
zwI6N|Zn|<DuC7)lxylY2h;cALtb+yO92^ku;DH2(03<p@Aju&C$qpGvaVS8lLj}?t
z8j$YLfeeQMWID|IbaGVbl}ROEeV9Dg``i0>v6efls@z@<G|)cqK3V5{as#(%2f2C8
zD+gANv7W%XOp;R5K610<9q1ixN&o!rslG4f?EfaY_$A{?rdm&YQD||qW$VJ#$-|}j
zr6rbq>CEn<Z!8%)JbAd|xQ<h-g0(B1yoj5&q^;65$KG{h@BK~hlm*=j1k18EsdF+?
zo70w^e3{*@zVI@;X&}bU0I_Zsh;ws5yqgCS+yap37J($U1SGp<AjPc!scscWb8A4l
zTL&`S29W7C?~`rb^x0KSJMr`9O`Sio>A{qy;eWpZ)&rYuV)MgE%?}@HemMF5L#k@t
zy&HvBd#rviZOTKF0uv|Orp}o*duDQe@WVqXdyBM|j`Yl|R@to|sLUxZbLM1J_RlTM
zYty!!VfJ~@UFzw}<<q^pYCTK-zoP|39f~`3?$WJ$k6tc@7erarJ1Xs2X+VBK#h|IA
LcUZrH#E}014%vg7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp.117C2517C948C007.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp.117C2517C948C007.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4d2b66e4f3cc1e7a660ede77c035760aaf516a23
GIT binary patch
literal 4898
zcmZvf2~-o;8pktXhyex|K%x*Kq!3YNMv)-Mq9Vu=5LtyFvN$MEWT^yXQBV|LL{Jg6
zT3fAJH|mBJtX82eC{<bPb7@^bL~Bu?Vxc}2wQtBfmwY|vaSk8zyZ^cOyK~n+;R-)L
zOEVtNXOSX5V{y(xoX6vtaKEaY>`#Yuc)U^r9xwl^`l#3tu?N{PQ?=Wn37<C3vTT%a
zzb`|EIao7IyW8vTE_wM%CjIbN%K0(;pxElqN=Nq<K0B2%qb*~LNtFpZupoAlUfw>>
ziGQj$E%&tjePD8L*So>q9$k;vIh*{}KS}wMb-Q1qwC>$B+FxyZMP;`AmpsMNqT;2m
zKbTEwSa~k-#)B_)J>ueb&k4Ku>+d&YvBlHtV;??@KY5F~{hoN^`E}j}8IxgHJGdbG
z#NFDH;?F9?$G?v+a7b@gwnbYibz4^*JaV{p|8=32?0nbzo{+S<?1ve7qD&*F?6$hA
z_J16|h%YGF>|nq1gx8RrdrDm4iHjLN)pItNsyiJE{&O~WYKVGkSV{g(x8oaR)Yh$f
z_<u@XUEevEnX%$c{v6YjuWttUSXMs7WraUk=DdAc<7;B~N^Cw`VfWPYMMp*H)sjwZ
zp`N#@NoV)K^i*3xiF4~+7tft*Mm-%HX_$X$)EM8IR%5cmC@bQ~$vG<yFW+AM`DKHu
zwPmH^C8z5)?9N%1w=PN$&|>v0>Ck(uwQ6eK`XCu;J|R20aqGt2p~ucQUD&dq<a4_#
z^Oqj$?-Vvs0`Y0X!NN7Jmn-eOpO)QsnMlkwzsh=f#9cn{@L0XQTXbRQwEoV!sS1VN
zfy#f0SCP9eE)NV(Dy*D;y*uar-iBJEfsl<xiW<ARv?IPhPVRbrYQJBE#knbqF7Hx&
z-;-**<9D~pdCjWN><Z&A@MdqC9pR?3J-5#I=M|@Iz1mZc?wETPi{G3Z&|bCr_9N;F
zAv6Bl&ygSW)3<j6SL)L9f9YZh7T|~Uo4$1VVx3^}t%kg}RmXZ)l{*!9n^haU^zgH|
zc&Tms_JP+0C2d>3El%FI?e>yI2aX*kEq~o&KE^!CzUtPuQI8i*wso-$p7685&c^0@
z-Q>3GyKZ*JS@QkWqn9_1&MBOB@5|1BEAMyhKmTFNW_54wK>7}*LvLqchkLTP`(5S4
zCpt~MfA1@+UG+n9!_N+pD|>t9#UC!MGrchX&YIixJAb=n`?+gV-?`&Es+AeaJ4X)=
zHElXQe&44~>$myuGN1Z(dA8%dZ>Yto)t1ul-kP6CEZNZYwcB$mPs@v)KW=MmOjaE^
zHCVPF(f;xFS7!6wUk=HR{Yal&dn<qLr8|EGjN31J{W|V>I(@zOZsvDc7X044r^Rh8
zDf@P|Ts^y^#l8I0y_zhS8y_!}WgD$ur+R<j+_-P1e>1o(A=ts;eQ;p&iwA}qvf4%Q
zc{lcc^s3$U=+ev1d5<IC%sZg>ENI+z%Z?KNO`kT@O*wz%QTODaKRpf?xBngD@J{w_
zWlh1OSMth?L!k--X}>D1PVIGYXxjGl)KNo8>ptF}d0FB+r!&#N)xzjV@67Hc<EQde
z1sUuXjK>p5&TRDA>uO`bSH|*KbAh>@CUB>O=12eCo{lv^QzH!c(VQ$6h)pyB$tEYt
zEOU>)xNN`=<z!QVsh%bvdFIfozQfN7P8jfGIN3sAF-jAVte94Nt=>H{-;fu{$r6#|
z6HP#J%0RiO{#M2Yf;XR&r|3?ZtqDjjx-<6RndVGWL*6`2wh`IbX#$d0dcBce>TgIj
z<VA6^waD656Odd}rT4u+eBuiOemEza2uy^UfaDhoZGK>!`{V}v5Ka~eM0%Ql<afG{
zzpkI@*loy*;AAV2l~fau{DF-6u05;qnE^kHlg$KXdYXXbkG+k1)RPW9GT<-ZWSKw~
zqzOpoWiL)vZ6vgFqo=Q#lMR=K_8fieVMDN#AEiKzaHOPSwM6^TK%<p3&=@}kXpE8p
zn%qJ@Ze+w+xvhLWBD4k!^)+)3x)DdEY7<jwR4$-AYBi=sI*taf(rGk!mCmE#t4tgN
zUS-l4@G6tXz*pri@>!^t1S|KE&jl1Jk41!*sj<Fhd4ew(PfgTzl5(b80rl%p13G}!
zIJF+vGCiIKb!lfB)TR4%Xjqq4<20<xjAuYy#+d<inSLDx)@9T<1MA8s$SL$;+(%E5
zJ0lVg&7caidqg<FdPqHNQ8B@WNy7l8NK*h6NsC4Z?N^b$W}kv_YR0G*Fr8F8Ms+~$
z!>B(XpW#zy`P#9fuJNhskni!S`;c38s5UUxbTdZ7d~`cT!+dlfM#Fsc89oj3(bxDi
z%tznj(=Z?1szbwkOf$y7d`vsWz<f*}#=v~c89oE^G1vGE%*Wj0GcX_1s>5jW%~JRB
z-VWy6CvetH=oXCT31_{8UV!|C{(wS+A%LQUQGnuvaex*J7XwNcrUOz5Re+WYmjhaj
z30Gr)%P?V?meFC2>uU~a?b=8tlE<P?AlM8d15gf;GeW_xRaqjYq7oxm@iegyI#nX)
zn%Z}7_SEHyo%EcHQA>$k>rw=dY&r;TEt4J+R7|kmq&FZ1sQ?s4h5=HNN<ax@0-zK!
z1yClL2`HD$1yn>90V>7FQVei8MwV-tbyA!&Ti9?$NU)v~&yfR^U;`zAfWjr=fD$E%
zfKnx?fR;&?0V<Xh1FFCz71#))0mAe(mn52=RB=RWeL!(0H@b46d}J<`;=C`a9fxjm
zqw7g}TA|7W>qGhg3M2yog_GfcV#pXkiDV+6R5BG%7MTTT8MzElF<A^~Ek>@z09RmS
zg_f5E&2FZYE}`7<X60)YjqYHAwf41+M%O*TN_-{JXt5Hkjjv5Kati1#w}gl70vaQN
z4JAS?wPO%XL<5Q^;ztN2;Tjk=6xs$>5GyRS4XnY4H5lOa7_nZ<B}HdSFLpjY%T0+@
zj8!7~40XTnIJdS6ORm_IG!>yMYr~VTLYS2?i@7D`E%hEbsR>pgRRB^-m4FhY34k)C
znSgSoxqwPBX(<M{9Fvx7`AE;2f|$o&ZRGkSi;yYNd44qioNj2wK0mJ5!pK5|ZVhCq
zg!&pQr5zn=66J{Ma(1Nxr#8ZSbP^3zx&o)6%1mN_%2ePCROO6(G8*?<d%by;EqkAH
zy)q$9Mx)*|IEOFN_1)LQ6`RY<8RT~_bQP*xy59upa2*m-w%YE|LK+YwWJc&x<c#MD
zl3$v*P2-BYeuDB6><oMcAYa@UP%s`0C=!nZ6pP0KO2U%>rQvCSmf}kR72pMcR$=%m
z46qu*)mrX+vdU|-NzDW9^KD{mlF%0T%~5x7l<}ByuGq=kY2-kA)>2)6U`22US8N?)
z&7$>=+_0{`#s!0!4W%N$(1p@~m{10g+*&S0Bi184x{2>zGkg<>WFkk@`X8>6=Iq2h
zja;!LMzRp?VYETGO*d?`waMKH_c5qRf}KUo0u(?5j1Y3;Tm!>aT-yLOo$>;`pH2f}
zrZXe-vHn3<q)T5EckuW~d?sjf#z<lSWl6FCt;Hm3u@OdTI1fGS_qBb127vg70U%gt
z00<Tu0D^@EfMB5kAXsPs2o@Ruf`tZvV4(pZ*nb)Tg8iof$X=YjaQI&oRd$jpV)!9x
zIQ;J?9zF1Q3EIa*?PHwwQPYS9by$;@fJ1}cj++HZaf$H>3wcQyi_%j$Ehsry2iFyh
z66zTchNQ8P!o=5SyotWT+jNZC=&=^l+&ny|&zR{o$((Yxb(tXff2AVvI4c|JCw4OX
gi4OE+$0_ouuGY?$jMHp$k$>PKzTezg0oa`X0u*BWQUCw|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp.C13174B8631F1EA5.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp.C13174B8631F1EA5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..be942abff4fcfc21a64587bb25b991ba575ac86a
GIT binary patch
literal 7502
zcmZvg3se*Lp2ssY0mDNP6ciB<fn?MYQ8Vv1N+M7R2*n`AM?}S3sJ#^hTdTcQ<D*iq
zQr%;ZE@-Q()mGbeHNN^#tBYdP-db<CjrCEjj~g4e(yP^WcgWo_zddI;hyTfEzF+1O
z{_~$a{u7I=R%^OWH@3E@xnWl0482aLOH_a7H@-Ue38vGv#_4p;TaH&X*l*H(eQk&B
zdv*_dWpK^vgt9y4uBhRgk~8zJY|g0~%+2BQj!#>9^H>>vuK3;Li5r`5e=;ZILd|=+
z936S?K=Lqy>Frkj^~c9dNp#zin**-=+nn!5KK9oyKCpvcV!Uv>CT;f5%YMuI>eBU1
zHx92pcmL=vulPCr^vQ}8Y3Y>Eli>NaZN8EA(;r-WcK_Cv3!k@LJ^%8{`ySs-p1UM!
z@zzrA+J@wV1y#-KWA|;^dvHYKl!4o>OkHGbc&TP&%AiH7eeJW~w%&brSoVU}8-0Fk
zl6GG{5p(FO!^@1nHnnboz00{R=0HK~hw%p-?Z1>a`i?xFy1iy)!Dc=-;mh*>dNAhx
z@Wg~+@1xD%q<+-0BC6=DzWt55W1sG<IA?DYGsot9o6+^P+_F1g>B#uhGeu<Ti%#`_
zRvc9=ui0GJIM4W1*S7wDy4HV3N~xax@%4*jfBmklAJsQoRwkSpwJY}S@`R7qf5}R6
zBX?xNj)Sv8+wM6}Pwq2v{Oz5{jJ?&94nDl-4-A={y1bY^d&c@ClZ`96{_gmVIh7mm
zDOXQ5K3YEfmJpROdIh^;;2&}iY+qJGJ4gL>|Hq$PNjs9dA}8rBle^|j{+nl8cD%Xp
z$m9?1&JGH-1$UL}7vHbhkh?KySfS-R-SX3^N9M#P?^wC(tF`Z(Nxv6&AOG)H8Vz3-
z)L*VDSu*WZa8gmji*x^e_KRY=-@GfAGV(``US8Nb>rVT_p=XdfKC$K23omvRE?#x}
z-_wyjlPkVzJe#~GX;aq1lK6>B_YEAgf&cw)?dvNGs{W(u*AZoxj-<@%_{&gwBUe7)
z(zuy-Mh9PS-FIMPdTjMLd+pGL*WWmUBg3E7FJT(m20U83``#R5_3Go%aSgMIE`E`b
zQn0CV>$uWA#VcBm>+8`4_<zQ2{-D!;I`O^XO(zEZbfYiL)_#+1oOAQ!rK_u~!*BhR
zb353k&)5*S?a!|{*80LP#}<6|?O(gICY|o3AhnAE`v>n?^K|l_vd{ciN~b#3l;o~I
z^x?|Nf4Kj(WnFVs=AoZH{9)%08?T>x`2N=Qd+#qVzdUNrtf-~y`dDAD-xfcjGH=%C
z<4L)`T}98v^g8~~mfipHKl%>z*SZ%ZNiAdU4f%EGlB^LwzR`B7`*QaohL|_Fb1OZ5
zzWe6U6Mo~!4K+uLm*nU}-~VyZ);0E8W^?_@#iQ@;onSoqSx3&IWy=QlW^>z`7oEI*
ze%sm&`7`RzBn)fK{mmMc`}8-@!R&9obv5nYmQ#N4Wcv!~i#^+whm(=w2XXJU50c)r
zt(*JF!3oXlKL2Rxs7pV!zxGBBzx&z0+-aNs^B?=o9n#fj=UcZXetqS#dmVE6spZ4^
zx*wm-Z2rAVdOtJ!#jm%RK8rcO?}w%MAM&2)<}Ye+wIVuQbf)*6vD>I2ap<JUI#+UZ
za!hzoUn9alvA?!oT~<O=#GzGcoEDvy7#_4Zao_4e3qHPgG7c?M<D}@MnDC&*($j~*
zUAGtQi$fi1oEn|lD?Dg%;lg_@BaUa@kJnYJabZ-U93HgT-H>+a*{-$ey>%68oSu@N
z6&|#B*@N$XsVaL%i9@HTu{qkzh6gQfIg@&^Zf?WwIJ8EMv!b&K!h;sKA1l5!=Fa2N
z-ny66I3pz^J3MId{y5*9$G@*#*;`ks#+fOZIpN`XydmaLblSenap*)fPK-`8ga<9Y
zH)F_=JWFs)9BNnNl<1V0@Sw%N^{x6eFz?juIJ8`i`$hMQ2@hJVdv(_A`A1VC*Cs1A
zd`((TJZ;~5of{JGGFeSU+W)7L!&DO?+=_z?ci<o+tVEGEZSgL`LDWQuw2~kr9VEyo
zD+MyjL4l06(jcQ9G{_h$12V?JfQ+@WAY&aY$T%wpGS0z)jJNV2;~hN61giit!6AT5
zw2B}T9U{mis{}I1A%RS`${>>+GRPFG0y4#+fXtR^8~l8MyKLDuqej|2H7YiIFM-SI
zG}T2eq{(es0I13I77{_+iGwQc#z7Tt!eNzg5_OTwOt^^!fSQQ6kO-1a5>!by394ih
z39FQo0#(XQfhyHR!7A;fL6vsXph`E<u*x_YP-WZ<s4`6qtg=oPR9QC*s%#SrtDKVq
zRnE<UD%ZrpD(~b$m3Q-?$~W<_DmVpD72E=-3QYp6icS$!MYjm5Vv`7~l2Zaz$t{7Z
z)Fi>G?36)OcFUkDH_5Q7I2BM;+zO~FO$w~qNSmUq)OeR|tgRGKnXN`6?WI3IHvEzg
zR<Y4EEV58d783<1gqeQEB8(ex(8Vn{=;9#^hh4%*42vvP!a`7hLKyKg7Gcszf-Y$x
zL6;0+B<xZ~3Unz81-euSqhOad(x6LQXwaoY7!A9OkpW%C!hkLl!Wh_PjV$Q078Z2b
z5XQnTXXHSavv8oxg)k0wc_R<HyoCo{K7{eGD;NdP6)Xbi3L#8@UC}6lu4oZKR}5hy
z>`F!nbR~-fx>5*}U{^NEpetKs(3L}&47-X^0bRwSfUXk46xg-lwh`LGRTuRrn?<A4
zLyva7mRPY4(Ys2rOQvauwNEbY*zhM7u(~@D(|PbhYziQztB`|;=>cRfYC4QYR=lYL
zHJyd*M@?OjeVFM2yhAz>9QyD8f<qr3L~!WCy(kWScn69@AMQtS=)-*&4t+!?LYxP0
z2}A%Pu0jqX!~@7)lsJq=?mVIcCC)<jqeK^EA4Xh&cWNg>LLV7GNa!Pj2nl_p7bT&O
z>_AEABmF1|eWVW~p^xfBDCnaC2nBsq5TT%t@}d;<Q5`4+eUu-ippWul6!g)Z2n~I7
z0HL9e4k9%4(O#5>KDq;?p^x^XH1yFvjD|j@6Jel_2_OvgF+qfZKE{hO(8qM34D>O6
zlz~3RhcVE{b|NhFu>pjIJ~oK3(8qdF7W&u@l!ZRlkFwCm`Y;yyxK4zFJ}!W8(8mQ4
z4*EDR%0VC3fpXBt`B4t~I3LDAAK!`a(8mW59{Tto!b2bLMS1AsJ5V0_ct6TRAMe9>
z=o2~-0s4dhB0!%IL<Hy)yr=+uLI)~9pWsIY=o5UH0DWR7B0`@SKt$*hgNO)yq8AmR
zPwYTN=o9^@2z{at6QNJ)L?q~w0*C~CQV@}#Px7J?^hq741bvbpm7q`ZVG{Joorny5
zasZK`PYxn7^vPaShCaCim7!1eqcZf#K1_x_r4v!0PYECj^eI6^fj-5HD$u8NpbGRU
zepG=z#fK>opR)4rx+jksyD+`0v|p)Bd-u}2%KMcAs_a+!oU{%S8~(Dby|Wcp@BB-`
zuCJ4jcvn_&mR)<pnm}!smy);k&+6ecI&G}>)upmD$#iEuAEh3)##j@yb56gpYEJa`
zL#`&MUrM_%({jCbns^u9jp2}qZj4y2kGyY?-53d(>c%L@bT>vrX1Xy3GTV)@khyM*
zgUokhJY=C86CjJ-m<U<w#w5scHzq??x-lifb!ip8<KvgStj>qcU>m5-hqi;Ej}7l!
zXmhR3jt4QV2cHsf4}wE39ba0Zon&FjxXuv{vJ8uNr5C18(#{$0$|%g3q-|rwyD|$i
zCuzGDe|p2a{L5?C535fny*hom_9xnGt52tAyCm{-;M$4*KWiu61=r5XP~l+a!rDRV
zd6W7krD$IPw5yE9hPPNE=Y^lBT5fK!=Pj+C)w4D{debVnHavP9GU3q^tKi!3=t;<w
zM^8bfJ$f24<IyvaS&yED%z5-2WZt9aAqyV809o|tMaYszFF}?)dKt3f(JK+wE|EGr
zJ=;u?*(o1i{*rc)_uOB$sQUVrchs3R7np6@mD<~UG^(s&yH!1$+9x$d`&6o3XZ4gl
zTT_u!R*tW%(oVVmfA^>oW5at>k&?-5%GWNLdSjY#K!lmd*W~do(o6!P%oHHnOao%f
z3?SCb0^-aZAl}Ra63hZ1(JTUz%n~5kECW)^3LsmaZKU>ZA27wd^yl;2?y6Ik*f+7i
zHVdCwCe}t3ZojV{PR>ov)3!aoxbfP2^3LOuo>gfuWrJxk;D87NkqxHBKmwu+6d>9_
z17ZveAlASF;tU)h-oOJA3<4n0AOeyM5+K<i15ymib2?Ee-kp|t<%~YwRje-_s?ATl
zYl40PpfY_KpbCA(bGml(O?iFd{=cbHmRX%SQ(JVm@}nO0N*M6Adf1w5&D73&`~2?X
zI~Gr{tA{hHGhEuL*S12{RoJr+3Ra=rG!aa<9S20%iHTsk?Ia+|P649rG$6*#0AlSd
zAkNMK;_W;j!7cz2?IIw_E&-D5G9bmS0J3G+vb5Q_Z`ivIW%u{IZ>Hp?<ZH8_?fmwQ
z4e$7Z*$_-Jm<<63L<ocoW<wwWQ33^s7HB|>zyM+e77!<JfOvriBnSc^Q4j%1f&@qw
zWI&3bJSTMzyM9K`hW7kdXU?zac`&VK_~*}5?SW1=J@U{Qc^Ll7Fk-IZBgsSiU#aLH
zJ<~aT>a-cUnUUJehT8f%6-#DM!1Phky$msN@xA*d^eIX#9BWLBEh<PF(69f%RGQ&<
zL6l^3@(4DY#54cDqbX^F(}!dY&B-+l8~(z`{83{FD#J1;Z}hn26kBmEYAsd9BQO3B
D%!Ndq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp.8596E811350866B4.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp.8596E811350866B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3d5815f56c85b2ae288b9a2e24a5267231034a64
GIT binary patch
literal 7144
zcmZvg2~<;OqQ`R=D4Po?C?cRDvF7fp3ZbHuAVrOWA}$$?OVwC(#0BlME>%%e#iiQH
z8<%Q3j<t)+bBS%W)~$_$(Nb%NYOLav(a~6DT5xGQzL0aS`M+}>=kT5U?*E_s&Uf#3
z?@exTq0M$P#bnBxT)3co%3LRHGMVCo|8wS6On*$7Op9YprUl#1m0p{9t#A8pFWoPZ
zUfwqOf8Jka?OE-QEKkT7vaw*9xgept`^u)xsSoc}?k(EZH?Vu&8%bhMU)PGpKhCTm
zH<<p$e)Cq*&7mEqE^K+!vb5J9`~AMbH`6D);H!J%@uSHFv!4vyA<WlnI~A<TcE5Q$
z^;}VhQGFf`pFOAFr;mQCPTJZ3V%<4^z?w6D-1c>a8~6OQKk076?Z?0W*76`L=X3ML
z@0LFNX8LpE4xTRkD&|t=L)SUir?$5%&NQEAF0J^;b?<TUq%pgdz4eENkCCUWnA!8e
z?8Q&t9K)qn{^_HY*WRVmJXe2t&!70;xr|yo<(<a~N7nD2bM@TA7{9sWg})D95LdnF
z;fB$p{<UCGa^k6XebdTWwy@Z23gq41qHQ1Z*}|V&I`%l&a=ALvwfQ=o`RZ3)DjwfI
zkQtx(=t#_N)0*U2b7$Em28K1|9(Zxr@!QLMkJApfi6C54nUo#1w=!nO?HibNYx#^9
z5BS=(_gBv>dEb$|Xz#0UAAIW!v+4Nsr71J?E#=m^^N+1Ajoj5x{qKV#7Y~@~Jh#1}
zIC@~jn(IFf8C+oBTYRJBX7`f4nP<zGj}ud`e{t9SMb(gXCvIH6_<YZe?X4x(vZ{y?
zGuep4-+VjnLg4%V`pQ$c{PeDZV?ByqO6Z-hUN@~cpZv*RR=Df(u1vIjdAn`%%6XfY
z^zBzQq5DzS2brJmZ)iF4p1wz&Uq0dD_BF@uO>4LHAIYZ=dF8p67wougQg^IAU(slr
zAD_Pa@x+2t`>u|VwpXnCY{kgc{)FXMeqhe8?YaJ`K8@b8VdO_!{u49e@W<!I{`+v$
z&>yS6%U8ed{K?>7Z!{JiJ(S10@2nYomS1$Ib7t<!o!rNJ0&{Gaw(eW}@}@0+n{nVu
z>1MjkeWy3S@Y}+_AKo{jZ}y;5=3%Q||7=6yg>B_scNHI3t}f`b>XcV{H%;o7lD@vJ
zO)c)dWU4glY0svu_xoI`s&Cjn;ov_$Dk@*+NZ)he(W%}`+5IVRO&a;<lUFU9``@1-
z2!k75iXJ@n+jp;=$hvsG*YwhzM}OKkct}mD*|E+$H1}8|v*=LYgn8S=ef=9h+<Wt<
zufM*tl-uH5xo_p5J>PwHw%?IAmfYG``Sauz#Z5Ezu6;#5lC@^%U;E8Vd@b{bjVJzk
zva;dHlg2ZzcgbpdZ)n%|B964YcI%ZlJ9)P|dhJXg>z+<(&UtT9j|<+(eY(F{IJHtv
zZW&_sw)1>H=F;;=JDhL%tymsjH*3VOwo7hYevo->;yqrw($#(R+So^Z7K~J*PmUbb
zqd8*Tpmm=wD_B4M<iR12l9yyJDcSwQfabuCC8g*0e?9(-M*G@r%PO)L#=I3fAiem5
zoXF1Ub?l*Csdw$mp4@ACeeqgxK*sj1z0`L$JvG(E_9`{aSy(=CE@3i7rhBUMcJ=8N
zOO6_2awbM5Mu&#rH6rvE^W>cy%Zs#<Sh6%2r$na2hXx~#-@m5IypR8RCYBr)j1wXg
zqC<lbo40;bK6_g(*Tz&5j8l_RyN3oN?*Hq*0(%=5?vEwQf^l+W@^hiVi06Kw$DHlk
zDYcDhWH3%kO3MfhM!fvN^_J3M)dOP5vB5YaG9xQA81aEk(Z?cF_SeLc#lbi}GCnFa
z81YwUzC8X~S>4Py)97HF6_NE)XfWcs{+>@u)hF3)Oe2DEdQy7N&|t*>bawlMQv*sY
zX%EIpkx9{^!H9ovUwX(t>uh5zIXoD5i0lv@8jQGN%FA=+Cx-WZYD{RKuRHzgJ0IT4
z?iS~?*er#{^nz4o85<_X#z4lDF_5t~7BaSsg^aUtka1-kWW0@sj4$IM6Kn!xLYV-W
zXcHk5%S6Z|n*^CuCP60KWXR+)88XGDK&F%_kf}BmGPO*FOtWc_X=NH@x=n{nFVi8j
zCR@8cTMSNXhShA4vG$pm(3*$C<FHH%9}tVnG7qTI@;VVl#=$_9aWPP3DjBS@4i>7c
zi-js%$zqjra8Tu398|eV4y(L_hbr&lp~_eCSQQ)sR0Wp+RiRSAs^}1*D!N3dij^W(
zC5Hr6$t6Kms+6!QJ7lQJE*YwFrHoa@p+Hq}DNt1^6|AZb6{@OBg{oSqVpVf!P}N)-
zRJBSCtGYvns_xREs#oe*wd&RZ#+r(ATJx+!freSf8f08dEis{sD6BWLrFVEqSyW3O
zpeD-lGZkjW%s`h>8R#-i6oXyX%tDt{S?ID&6pLNX%t4n^Ip}gt6o*~j%tMz~dFb*@
z6pvlOEI?OK1?UP*lz?5)EJ9aQMd*r6l!#r)EJ0UNCFn{`l!RT`EJIgTW$4OHl#E@)
ztUy;$73eBWl!9H=tU^~+Rp_csl!{%=tU*^(HRx(hl!jg1tV35<b?E9%l#X5N^H#=K
zV8O*FSyh9Kn==^`x+#a3TkfFT5~F0?qQ}LAZq4D6rIxT<gbSU;Pgrgs2MEgpWDjXM
zPKMX1#Y<W)AlH+Y?~#3!<r1#zT7p3z<0lyOF#&=>ALAhz^f6wNK_633GU#J`6oWpt
zmSEAx`Uw_&Y=B_V$9hN>eXN&c(Z|-4Ec#d<#iEa^B{=kPeu6_E7a%zFaUPOGALk`G
z^l|khhd$0nap>b~2_AjCpWxBQ2M8X0yocn`$9qX0eSAI1qmTDdJo<!MLO`G3Cj|5f
z0YX5Z;2{O{30_h_pHNQ<=o5UDfIhL75YZ?42@!o_fDq9qdPos{qL&oWC)Sf9`a~Zk
zqED(NB=kvsLPDPuASCoj9#TS|<RvBaN%f?JKFLQ(=#y&+8GW*!kkKax2pN5{hm_GL
zdr296ay==dPxetV`jlEiL7(C$6!a+pLP4M6Ar<r~UQ$7yQco)AQ+$+yKDCxm(Wm+e
z6@6-eP|>G)NELmmmsHWG){`pwR3D|HPpc&~^l5%VL!TBPH1ug6QbV8SB{lSE^`wSA
z%|~hI(`yMGeY&5}(WeIp9euio)X}GVNgaK9J*lHl_fdM-H(=F$)8k*}eoxWP+zz<~
z#?_s6=5@#evURXMqal4Z%}V0dZXn{EDScD&j4#6@h9>m7v(Xw6+?*Yq-Ol)=2@R{K
zM*h(4Mw{Tzm1fGaf(~ElnP!SXW}7J%nQNvvWWJf=k%eYTKo*-R5m{=cBxJdnl982W
zN<mheDHU03rZi-|nbN~NEv3YFV$kB5!TDOFteuVd&ZbQ@?~A`)Zk?~2w!Dq=b<+$o
z>!w*`&P{X3yqo5c1vf1qi*8y(mfW<2EW2qLS#i?}vg)Q)WX(-$$hw=>!)*NM4t^_W
z{ly)=;pYv?8)?jbe$k7yOqo86igTvsr;akdNyItR^3z5cpAhFv&rcs^{E%MunrFoq
zH*Xye?#I;8sS}L17|Vo?3H^8v>u{OI{hwtT=fq{Yswux~ze{Vo1V1;SeL|A)9eUmE
zZFX+S=qator;4_$#^X~(GstWe%_4JEG>6Pr(LAzHMGMGc6)hr5RkVaGSJ5)EQbjAs
zY89;_YgM#{tXI)`m_N9;eBtO@+p2@}%@~+bVC=sS7knBqtbDgExH-9Pa+2}2&3FjX
zE++KwBfNVpyhSwL8|P$r28iWZAdcsNc%BCmcmYV{MIecnfMi|<Qg{VO<y9b!*MM|h
z2eNWj!B~Tbdqt1jaB<h&;Edzj$9FVldQvT(9Ff2KUT||_W+G=i-a6Ir>Ky&fqoUUJ
z5M{}LsYfwDY!nN`MR7oU6b~ds2|!|$2qZ;GKys7}q(mt|YLp73MQK2Kl>UrPk9gtk
z`1J3-44!cP>Ha;8eGumyNDl<cr}LkY(IH|&54Ma$Z=8?8^Skw_ma#c_QVp_|8t=Jz
zdui^*_ygC1J28E9`Xu8$KZ+5*JlC$%8^O)liP`B!yYb@tC;q%>h&{MDZFJg1<NP<C
zkp<UY>jSc|$zrz@!vU}}K&+hw;_Ms{Z|8voy8tBGMIgy80m*h5NU<wGs$B)r>>7}6
z*MY2Q*6zlP??rvMp3FMXdT&h1OcITm8jt<j#)KaC!Hjb(xiI4#28hjJfw&wFh|l4H
zgd71#%n^a490^Fyk%5#P1xU?NfwUYANYBxq(SJ@{YZ##iN?XsE8rrfOub?$FgUr^@
zEHYO^bI5!R%_9pnw16zu&?2%_LrchV4J{)pHMD}P*3c@lRzqvZdJV0IxnjyI&ekWf
zb1EjyDQSK2ht?tVM>&64QgG?|l=9w@6UU8r{9%I0VMHd=`0&ed;g=J_FDHgyPAZ@L
z%CsOBy*z}XBO;%RijIwI)4pBX!ub3=b9_wUz=Td6I(AO(+b`Ri^FsezK9R}F(4SBL
r|F$NjbWQEny+_Z?UcD`>AWDkXCr$0b$uA}*6}&Wwv<)3Fh#368<U`E{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp.09DE6214A3D6FEFC.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp.09DE6214A3D6FEFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..90d0939061a1495e00fe4a61bcba7d20d94a7949
GIT binary patch
literal 6700
zcmZvg33L;69>+5?X-F@6(H1DZOMtm=N}JG@(6og@YJrl10m5yhDp0ls%c3BGf(1lj
zL90bjIfUhqVh{_6ctmPcR5Zn)fD{1(Q7CdKZkl~Ve*50`z5Y9&`F<y#&dfiP%v6?{
zOgVj3D&zFB`kEQFHnmEniVOVCubuapA5*ESB2}vT=C-OD^Tpg)cgtb(+dBt6)o)^B
zOht#+p1i|S)o}39S?euxe{J?1y!Fq&nr;qzog6?luiJG}SZlahf9|w81KC&o)}GR$
z-0J2_o)0f$FC4#gE2YKSGP$AUx7By=OgGQHSN<CFhur?Jk`+&{dhw#+qxlJ|uN~Rp
zJlbsi^Uq<cTMA}2EPPtHeEHNzo=v@)msrWWzu&*>{j>A-s+P$wbdEp%-Iaq_((ZkU
z2TrUWTi-gf)w|{532j@~XToD|cP!d}`qPoM8fl_5VtaX?ov-xz@u>&C0j(2;oIjKw
zIk+VJhsZrw%8n4?vsJSzH$0m}d^2>FyE^&D$7|1441BTVd}i*5bBkwM`(MbMF;=xU
zZ(HmBeOYCjRR?35#-FLR-@Wqn<uNNyb{L=TyMs&A4fGh#tUEPe1=YA|>5n@yB1^f1
zrt6=-Vab0fE4F#-HzyhkFE_+=IF_nf-`Nnow$uM?u4Y{3tNsrp-g3Fe$|p+R-TBSY
zsG(!qPFYT!4|Dw>W;`#Q{Y*~FJCZ+~SXL2Ua`*c|e;3WNQO)r?*FQHRe)*|aHm!NB
z;?4KBA2+^Lu|0O$p69A+pNiQ|Xi^SIw-*@)Z(BEfn16W9wceVF!~1)EIePqR(YvQ@
zMfpI(z<!HrA3yL$W!9T39VyRznQQ!`{PnH3Vp}(T^_$bU*p_-Jee{}4%gQ|+cGr#7
z&cE=TO@o@ey}#Gay5P&}z&eW^`|JmzQpHI>=c<|pY~A0yyzI!zn6DZRWltIR_^wT(
z4;MXhBJboAjYU^)Bu+D}|KwPTxPHW^yUvVQ5{6Is?gF)Nf_nSB0avqLo*iX<Yw?ZW
zHyk=(%v}9TWBEz?{Tsde7cOt1lE1z^WZSIraRq{F#IKU&F=FIW&Fy{h2i~}vGjGB;
zyg4;zO2Q6f(WFmvX1L`wUyeO2oT*P<aoojiOXG&5q_6F15-SHSnZ-T1q`!aN`!%mc
zzw=SPYw8yhKdEW5rq7$v5ZkM8<9E@g_V04;ZH>6}<o!i2CU3kjq3q{9S2MqDdwJHy
zgx9{^?mL1X9<*@AoqPVT+@)t0oY--7(>n*27L=51!wywC5_fInH&1Gezx++kgU!p%
z-Pslw@#oXp=jZUPd5tX}4_OeqIPTcG4X2KpFWkF&@9g@Q2lwn8-S3V5t-o2%&v`M~
zwa1%N5RcwjGI!yttIX4fZ2omc1Ak@rDBXugT{(B7d*om1^U>TB-=`c~UcT{GO5#g<
z|48XLHhtn~@0*TY!#ua=M=p70d~w@@3DITOBl|AxTkuTD+Q;^{j;ddKtgRsHs~cMu
zwY}Za>&D1ie^0qN*El1i!@Y6orOMMiPTSWaAFmK==FBeXnpXcH_x&yD+2yBq=#GXr
zZTWd}b#l_G`3q~N)*>pECf&Kl`1;_CNOb%Jl`T<|7#<t~SBBtE#Jw%&mW`CgM50xJ
zI7O2Z7aWv0Zf|4X1^>SKVI*1+h~qW!;lV+P3-15s+tImjZzO67#C<e<dISe0F8V3v
zYz|u#8Lb){h|`nObAp2sFS&mAz2#@J7DTJY1me`B)Xd<Z#LKS#aJ#BvO;IE|F%V~K
zvh#w267O3d{;nov@8(FfG7!gU;<Uj*iBD#2|GMgdy&+0f9f<S7@(P2468}=#+Sz!n
zzAjodE)b_BrDX>PCBACQc#jf&LL_Pq#7UZ@@Zg}tf5lcE@IHC6GZK9y5GQC7!h?en
z&zmuG{>g;Uu}_T%9`mLX_qXgmpPvzB!%cXZvU>rt;1fecm<W&&76N3Xi3AyGAwfo&
zD3DPW3S_j21{rOkLB^OEkTDhpWUPq=8Eaud#+f*faTX3_yom=HZ{a~Em;{gs76D|U
zNd%c_5kV%IB#=oK31qTK2AOP;L8kAc@Aq&r*!0<YokGgoCnAD#9-JO4K09<ma67&L
zP#wM$2_a!6K$Wl)pi0ycuu57<P$lgosFHOgtWs7AR4F?Js#F~XtF)B{RoYI2DqTmz
zDr03pm9aCR%G5Ei%34`aW$i4evUM!1a#jvhIXefcTpb6iyp;!4-p+$6U&q6$U==`B
zunV9n)CsUET18M5?INg(bt0@vRtZ!ky9BCIodm10RR&erE`zFEC&Q{<))y&rD$1re
z>PG{r&`(rIxtQV+!HX!EZ#sNXXiDKCJ{XW6!*60CCUgYo5+VV*gdZbdm(-D<ONu1u
zl75VYT}nrRE+tZ+OZhPhc4-|Ay0l1xF73x?*kyDK=rSS$x{M!VV3*agpv#IZ=(2u{
zg<Vd^fi5R<pv(C&4t9AR54ya_gD&sKc-R$m0_X~&0J?%76JS@=iJ&WrBIt^KOoUxY
zCxNacN}wzGF$s2Moea9ND1)x-$7I;m57ZON1Pe?)PA@8ya(^hXXx_D?=)fn8LyfOg
zgGuC26Oc)VnuJU_)D&dep{5}-4mAUrb*Nd$oI}k)<{fGtvfxk)kVS`Dge*DK5@gw-
zmP4$3>v%LG_-zBckJ=IZb8vCSy$F5|vJb(pLw2J0VKj7Da2JYy2HAt+9gy7^{snw5
zwj%`e5nhCVKEj6(&__5?0{RFSN<bgsK?&$1+!z6UWIIAaAL&I%=p%gy34Np!C83XW
zp(ON?9+ZSW(v6YON3|mq^if`ff<DTJP|!y?Q40Df7fL}N<v}Uvqudw;eRMlQLm%x$
zXy~JT2n~I-6Q!Y#cA+%%(H@kBKH81Z(8shR4D>Nxgn>TBhcM8`I8g@r7#GSwALBt8
z=wsX%1AS~e!a^VGMOf%#eFzJEtP^FSk9DCe^sye4g+A7evCzl0BOLT`UW9`_&WCW&
z$2n0B`ZyQLK_BNqIq2iu7zcfPJHkUB??rg%<9!GZeY_Lpp^tZ=JoND%l!rdvjq%VY
zv?Bua30_2iKEa0w&?h)i0r~_NDnOs$K?Ud&+?W7;Vml&2pXfzI=o5X22z{ax6`@ab
zp(6B&9#n)r(T$1FC$%FI^hsVsf<DQINYE!aQ3?7a7b-!Y<Uu9qliZjDeR4Y@L!azL
zWayK9hzxzQ6P2M)cA+x#$sSaOKG}`QAz#soUsZQ+)plTNTVX<BiSn_fwiy$QfJ_Od
zhosz4q7lIx$<USGRBEbHUSQPV^-m^I&#yzGY$>@ZM&(=n*wOI=etAJ37FeAho*$#M
zr(9S$OY>vKx$aNNE)0JTeoA&>1Z1)cBOy~=7zLT`!f41$7sfzlyD%0q*M)JA`7Vry
zEOcQ4WU&hqAxmAD1X=FF<PgtJ8RKp%{pS;b{ncyry_Nm_$H?LBghj5vqHL+fspFM<
z@F-haaoTw0t~kn;UYtH&Ioe7maMZhR(nCkRu(WWD^19D1c0TvPPv;K@bg9*;Q<QC#
z4_9?W@a7$C3r=~||D5tD8$9JJ{Kfr-eDQqW!0X1x#wRJ?q2Io?Y+?2J<~4!U*~79+
z6wU7XBViRaTTOx0eR}puQoboD&kAB9g3ky-{{>Fttg>yCji3oYBuxUMXbKQb(|{P7
z0mRZQAdcn$@iY%epano8Edr8g36M<7fb^7}QI7k;0pa7;eg68Dz#hlN#`RM6^r%=l
zJ*;@^wZQ7c{)v?G6zKScMf2rf?u_i7?OHq=?7Ef!L~2Pul$HWSYiU4?mI1_SSwNhY
z1H@~2K!R2PBx*%Kl2!sFYvqS@V(ic>lheQZBrtgk)CF0}F^IAaQx5}FtS){?N(YGu
zKL1fp{ZAL4DSRPr-`T+NNv}?yrcCuqY}l<HG0Cd~tMe1{)0K@oKi}E*X2U3RV0Bt`
z+EnFjqWlfKF}j~1DX$%v&ZWk2%C^d*q`=(kevT9>;by!NoB%Tch%}RcC^H3!Hq(F@
zGXsb<vw%1=2Z%TGfCRGuNHmLpB(nraHp_tYY5GiM*RE-IuSN6rbzf<d`X{l<o+?kr
zdPW4Fih(^g;Dun14Fn+4Kmwu+6d>9_17ZveAlASF;tU)h-oOJA3<4n0AOeyM5+K<i
zKcsmx=GeNQ>&>4xZT^_<g{j@ce}AmXf=V?hv^Y7m_-JTx%ELv>c08<g$cCF4TQyT3
zoot;nMK!Hv`kdK;eB{hgm^w_;LmM6$6&)MXvn;OIsEdmz8y26O(5rW!+#&gT!_b04
yI+4iBmItQ)zqLsz{ZcbBvvT?m7=)7y%L&rpG_fzm4^K=g88HntjV>xh%Krm=Xa!#Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp.515130D1DB38B8AB.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp.515130D1DB38B8AB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..829ae55a62e4620ee54d42613d4a0b81bf1e64a7
GIT binary patch
literal 4724
zcmZvfc~leE9>+5iz!0bq9v~u!h>Ww{87hfv#swtG8e~bslAsT&7D0PJ%Tl$~s;PCW
zqGF%$Dq0mQD6ORp(rVp~4MwEysG%yfMKOK$)YmFj`o^3y&F_zwbGVuB{r<lB%-orK
z=aQME(Ukf~BnibyRb{U)o#!HvNEE_f#nSROu1O^l)<Ytxs{clpvgCHO_w|o^vR<8E
zUwo_u3-LL>cvG2ASlZTo(`&=ldncA8b#8t>SlKkIKH5IdmMF>p$$GweTEELUtTp!l
zs{6KqJW&1nlLu=izNvm5*YJbUxT4|UYa4z&hxPR_HJ`<rTlEbMV<WgOuYjXD?2?f2
zeZSYN{O^m%;1locK6v<fM`BHTUi$g7v#;<Ma*hwxJZm*~oJ``Ey#HI|-3z7Hu@v1+
zdG5XTeq))w?5lyy#pd4XA3ti%DN`3NN=;enrpk{>PC2l#d(+vl{w;M^OYG*`cGaDR
z@W<b^9hhi4((Mx4xRsV(ywS8`Lc=j`cyYHqzA<&|`h$0ys<wQz%x_D4XOm>?GR=ab
zd1>1xVFP0Wk2XKJdH;+_8xk6OT^{`A9*rd}W3Q<>w3c6kwvwlo<aoQzwpQ(XTjjIJ
zw0gwqaQVexuL(cD-934`OH)HjX_d~T?Jvsk+qTB*h~iZ2wCTRViCKpd0+%<{)P7cX
zX3WHjE0U<|kLcS~15bWYw0Op!3%*u#>CaVNr`N@u+wZeRKED0JZvE!o#^h@4dQ)oS
zPVb5qvu;AwyIJp_T4&g#{vo4y`?dk^ujk&5J%D8|X2ZYeyH?QoqUY`J%$KES8fSeK
znwGpR>w`OQO1%RYkm>GzJ4^)|{DQZ{T^)6O&Wn5ZgI2GN{uI?cI$^ne>4RIlF5hff
zg(?Znjdu0ie=m>?5*NB>^i^+`c2|`vzl@Cf{KC@@FTNIq>)H-%ey}UQKfE-)e&?=F
zrXJZ9xp&I&l|729@w&FWA@8O)3>&`cx)K__BhEN#PGS+c>-Wk+zUT7Dc5-icrsCk^
z&weAZl0Qq^>f0{GsEZB_H6-QTZ$6cmKgUN|<+X@QewbdKemHsUgCu<Ns#xFRwb_|J
zmiEbZbf)F@84B=GKX06Ww!d`w)Y#GAC+{ChE3yqI-=$&}fBZ_vxwupJt@(Ag0y|<-
zcl^343(qx1vFATcUeS_6{&lQnUvtFzG2szqx!TM>9x3DERt)dV$@|wAx0uoEkEaK;
z1y$&>AOAd1y7ct*&er&A@e_`XG2UNIoIQEDchO2+eD1VcL*EgRhiji^7IZ&aPCOf2
zIYp=3<F)o!nsLG(Z|z>5vase#k>|6S+iNZJ=!v!IH;O0S*`2zr@9Ki>``n9Wo#@|1
zK8U27{7)Voi0J%Md%S7gTQj~NS>kc>y`02^KMxOkAL18wys>=5w3^i48p=-p=Vshf
zlfpi-xp2$ihTat$hC_P4zVc<j*QIruQ{|xpZ427YOZu|%!X|&=@iZptoqFT<CIzt&
zt3Bp!-`QVS_t%JWNkw(pKSyB_iCc(yb3$X(I1fDgHHp#J&DY)02=m|Zkq_<ZeJ@3o
z<$-4jcA#6J!qJGf;^_LoWk(*K@xU_#Tj8c~cQm3s?fI|vrk|>hdf+*N9pDx)!qJF!
zyz$OZ+uFJJUXf%9c94HisG||>byofp&fK}*1J?_7h+9aUqY>>(x`$pF!wK1*5}ja2
zx<*cQG@|`*-nipPX^-^4wSw*E=I8EcMEixe?sMCcvp;#@a|GMR&Bxu*h_<Bs_4yU6
zJ)Ps@CwGj?rqj>&G~bIG=V>IgM5cI#g3=TDPDN@-;E{R~c$Ag`9;K&%M{8-|(Rvzq
zjFtf&qi2A}YFXg1dKP#{i-3pp2zZ>910JX6fX8cj;PHAMc&f3g(7%V@s0vdNq7t8%
z5V_+y0mEe=7CHxrC?S>sRY|PIoRTz<pp+~jK`B{D!cxjWfl{i30;N<X1xslI4NB<}
z8kEwNG%RHd3@Bwv7*NVoGO(02u%MJJVL>Td$-+`(K%f+rAW(`b5iI2l94O^VI8e$}
za<G&)@Sv10;Xx^1$-`1rlqz1lPo75AOqCX>IjRCtiIW<WJ0`Pu55>tmUGA8~Vz=<@
zTQTAyxPOEVBYL5>W5h2|n{lEYciwNpf)kxk_u#}WsJo=ZC8@KXv|=RWkv5EkJkpMl
zkVl$v67on3PC_2pgOiX)c1cOdqpTPOd6W&KAdj+R6y#B6oPs>cf>V%3_23lbQC(6B
z@@OkYLmq9zXvm}O7!7%}8K)tSw%|16(LFc~d32YQhCIfKF_6dDFb47%JH|jBW5yZC
zV=OoWc}x$^KpxX2Wgw5WVl3pbHjITl){e1|$C_~#@>mPbLLS?LvyjJjNm<B4Rt!NN
zvSA4FkR3yihs-#FJY>NU<e?rMK_2RoBFN*c7zcTr4dWn>vtu0Oab}!@JkEl1kjM4l
z9OQ9bQV#NXE5<_}Z^L-V<Lwv^dAu3tA&<A<Jmm2`I1hPzmy~z%W|$sHo;_K53v)5X
zgvIE^MMHd-%jJ%DI9MG_7_kk^e$s@IP*Wz1f|@pAG}Md<W1wbD7z;HrVF+r@gmF;w
zCX9FLw5V-M{L#h_F;8P)bYOz`y2?)XnK-a@imTAAbXR(b7qxIqR@XPPS1uO1RWen8
z*!@<@Oe>?GoFQ}v#Rp}JmoBbQxZJT%aSj0-C)4A^pqNV=1BI-@Tj4LRwTB1atIoYw
zzgg%GNeG!O{#0{ShigXJUX9QlFfzbjT!K$1GmBjl_YMl(zTv(|+{1l)zougHz|$0W
zVNI41Vd9T?8c7)m6eXj8qGdEtjEn(_m9an}83KxvaX|4h{%>_Ud+LM2kQ?WPvx|0#
z9xslX_$|CvUT)UJ;kn!Q?)0sS*4sjLkS=JhIG8_1xIP)-HEM&<t@Kldh~d9?b#ljU
zPW%z!f;B`ExL^$l6s4hnqBS&7jD`V<)v!Pz4FZbOa6s`I9w^mlRgid!gR<r=cx>y-
zN6atWFG4(Hal=W<9UDz><^(}G&zwL+1W-yM3Me%(-Kj`|1UV9sAV(@mkR#P3$Wa6Z
zaulLKj#5$}N2w{0qX`=1Xhee?t)xMYR?{HI5DdsMhygi9$$%WAW<ZW5Sde283v#TI
z1vysDf*c|c$RUJ44k;1HAvFRyj^IF!LmbF)N)F^WH3xD$!Gj!+c#z|jJjn5C9^_P{
ziV>d~VSPcW2%sjZRH7<>y>#Boe@iOLiz>2S9^}7l93QFC(*B3N@npwVx5C_lynKTs
zPqZYG0_S0_^U&ZtEOZ_gl@*sR6lBW$G^vZL+X$JvhvzHaUL%tfi3x;4o)qsh>L0!V
zqoZTurl_Y*o6h+$v0;<RkpI`~KPD)6T<G`-;S(vE<s)8=R7H(d2BJwb{rqPo&&4(A
JGiGC{{{uC)cOd`(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp.252C84373128A03D.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp.252C84373128A03D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ea6218b0cf1ab5aa72ad601c02d33f137f73906
GIT binary patch
literal 4724
zcmZvf2~-o;9>y~hz!0bq3W&%eBI9g#hDsuvaRG_4M>fNfAP-a;L3=>UQnl48)Vff`
zg&yJ6t%wzr*3w34wQk3PF;aK5p(?b+V*2c<uXU-uH|Csa{_mWZbGXcRzwhHWxs$mW
zG7>eKGB1fFz9g}_e8I9JCy7L&5Pm9`RlI&pDv_{m5=nLQ*Ez{cZ%27v|ENFnwS~<k
zM?0`!uk%Ybm3xJzZrL}ZF?79Wf-$jo^Rtnvw%N^5wjyhSB>yM#`I_lNPUEr8+*Y*n
zn-;RQ=8s4B*G_sv{VcZS2ZLc{%Yk_temjQ^4l=c$#+W*FEiFOeoW;ZMNH)7Pc*5Wx
zwX6R9JR<PK+q(`Nde)s#dn_;Q{Mk8I_zT&`pVmI<G<Ba$<d?qpd&Hd!W!JIfoSX98
zyT^tM<+}1OhclL#25Nr%urs?{UAQ<Ud6|nUKQbw~bz|SAv!O!`JFgmTrrS2v_bp)$
zzwK(BWIf#H6tiawExmZ7ZRNz4quh%neYUtgDM9NG+-a+B_;9&TLtJm0Bxt#2QE^e~
zw#nFVkpGeP`#0~MsZS3MiMcKhd}Ft!Csn(*zCEOoUyHVor<P`WI?pjz?|Vz-wYYxG
zs5N2oi-8^!e}1cP$~LFAmX5OOoci>kVy(~CwH}8Rr(&kh@D5DKJQ(l4qOG>^)17C=
zO{%;iiM;-RzFj^1=vPIDd)&FeYsHuTT5UO97klnwueI_C$1d#BZ64T@RFl5GK4s4i
z&&m!{&cy0>GT%E@r{ARhL0htI>#*ln^KZwrVp&Vru+Imt74$yuf9qS*W$Bqcv%d^U
zP1>6I{`YT4J^dGvX|6uo>kBsc1UAH89dmu|^Sk!~)~t*A1nqoqqUZLd_ir^_zS*%F
zRS}vS$JF!xxkxrbT<DuQShHE$S6!j}A|mp$3y(jzI4=^<>1y44zbStxtSqj1N7Mdk
zhnphyPJMA@x8iDCPFLPj&$id~8@{w$35nVsYZx;(p_pv?qpFbazdZUFxi>6Bap2*n
zzmu5pud=S@u1nGC;)73H67%l0pUTUh>!qyrSj;8;l2(y+C~4jOM10EX81IsGSs6c;
z4a&CnrsfXn3-B>NZ(MkGsBFcwn6cj_ef%`F*!m*r4i&xRqgT4m#h$un&fj^<zdJf*
z`){i=@mxbBd;XK8l^xmSe~xzSYY$&PE-bt}H$CIe2g=ykl`nQ=8+$*$#f)8lJk75w
zpfV@x;m^Zm%T8bK?Tot?H}U8=!@V`c*^`$C7O%>Q%bkAf>9<70p~lA<1$_@z5Kl%{
zP0dm6_E>i`)iCkTH+QW_F0H*%?EYldw#J?!dQxNBjgraV?@HM^cy-aXeXhl`PYgAY
z_ao>w-;+m%!+XC-Ki*dN=FIO#8{JO6o1GB<*P$1l2l>+O*H?_1UYqiJOZn-4-Hd%)
zudt17FKiguFtBpNi{OE;u6*J5RoPC>sfv){u0>tvC4-rHp;JD0dmJ74cC+ETdIeF6
zH6C@f?HDR-8a=8)Qdv{}k1?1;;u37y9KR=Wyc?c1Ph#+P@piQj!u@Z5$)D~XcsE&<
z>4s+tw!e$N!aj($;>ddc<%fSc<A!SmTj8Q`wGX1L3X*9bc(z`3!*zlk>=GPnA4Gfl
zv)^oOKh+#@!?Ok3&&6+)eGu*Jto;icP8QkRB{_l};T$o|K8SWX_3GE%0b3WmBFPZ!
z0N;QR`ykrC6pcTQl(q;rJYBGTTzp*ZgJ?hZ%=yf^^z2V=_*}vEa`AGt527upSg^2i
zjXPK$xqV$Woqo2v{ch}dcLSj#^2PV(ZXk6eP?U}Wiq_FUF**h)R>uN`bO<O;#{tFb
zc%W3FDneWz@froo?bimJqMj&qoFZW)mIGBqtic?T)RQ1e8c7f(t4J87^c09vMhZl!
zDhftvJq@C?kp@w^iiS}}&wwanWI&XuVqlcjvmnYESrBEbSQtfm1fs}@KonIW80GXF
zh;l{_M7b&sMtMCCqP&p@QND_YQB|ZWPP|U;2GuN8I#6>}1)>sfBus9<fy8SlULag<
zzj4HAcSCe&v`*}bx6ur_{q}K$2s1`p1Q(gGV#EN{HjMZcY7<Tz!$qd>5%0l?Ua0$V
z;uh2vDRD{a$dhJ_ggnxUk&s8)FcR`e6HY=N*@KgiNA}|+<dGIB33-$mqacs6Vie?2
zHjIKi%7jypNA=(o<Wc=N1$mT3N<kiN#%Rc+tr!h?v<;&nk2c{n<k3Ai4S94wPD37T
zk<yUIm@x+O7%Rp=9%I88$YV@619?mj&Ojd1k28?RSfmW(v1W{gJl2Y_kjL6E7V=mV
z&O#pBgR_vw_Twz%u@)%{dB}_*$U{~PK_0SU2=b5#N05hla0Gd%A4iafEK&q{oEhUF
zkF#PN<Z(8PgFMcJbCAdN;2h*}{Wu4CoJGn(9&g5Y$m6XT4|%)|;~|eX;XLH=Jva|}
zd_T@Z9&eHI4&KcA2a+d`mfgaf{%3bcjoiA_7j66ib2s=$`NxZ27FlUtlZLlUbrz<T
zu1XJaFA;wW`_RiTTyZZc9hYhOdd{jPLRuwL`H3m<D_AbKfAxxKVdtpBh%I1eA?q*_
zYN`&Spr-3E8fvBvW1wd1FcxZ5hasrBI*fywufuqUelvNNnbA$r3TGb>7my{Mb^XJH
zfapsb{e@|Tr@~j<f)9<nTa$aYd9yGb93MPK+}_%&yPdV=do{we-)KKyann7a%qVeA
z*gGOjdxv==@p0;#d$pBQh94)p3Tq)FLd8EKJSfOWpePvy6fL8HVq^?Ztc(Q;$q-PS
zj01|7@qeq+S<~(p2H!X@oL!Vt)CBRm-5Gpm#ui2AZDBefCt$v~AU}pXKN{sRW`i)T
z^ic+jY4J(QM{a+D0#{i>Bs$(1p(QdMiqw$6Bef*(C=CTXN=pHc*3iJCwKVV;4Ffzz
z%K(qnu)t%rEbx#90S{>r@Hh<zJWk62kJs?P<F!2SRAW^E;sOabMUW~)R3ozX20W(o
z<qhW(<`XWip7_8?%Iy!DU=<01a;zJHh;X2kL?lpZVunMJ1POX1B0-N-lAuSbNzkJR
z3iK#MfgYu#K#x*Wphpul=+TG<Jz7bF9<8Q9k0BV)V-N#+jFJI8M$Ld8OR%8FA{O*m
zB@242ngu;XAkaeyfgVyK&_ikjdK|%l9)~#4<CGleacU0qc!CE#9`T^ZD|yi4)ja5_
zNEIV~W`s`=QiTIGS)~$H#e%Y;m;aVjRuorezC6f(IoMxPLtFL!tECC=%yTKsEy&B)
zOY%fZA}Mej<~k1bj>AI7VR3m$S*akC7p6*`oLxrAT;1GX@$?v-s7Qz>6!OG4uQ9KB
z`;Co?j-9HWHhl)?!^DJ6A%p+ltnavh!0{myCWcL-XqFFuEkYF;r1VFVXZiTfPMVKv
K(q_)VQvM5fn1}`d

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp.6BA0BD5269CBABBA.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp.6BA0BD5269CBABBA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..867fcbb5fb16483aa9466be4946d4cee474ac07b
GIT binary patch
literal 4442
zcmZvec~lfv7RIZp0h=X4Gk}O7BGjw36`BSFY!{H0otC9-mTq()q6q4Mh9zpo4PsD?
z;zA;f+e9&fOAN75qrv4Uh)CQ)Au1%nU}yAb5|`*q(dSh9{xSUVPJQ=&U;paVtGc}8
z7_~ZChGCJpF{Op?FP&?JVVEGX;-y6kZb>kVx5u#3x^J~{i|<Wxy1lDAC174%?(rtv
z+xgO>jfKv>37hv#tMOg$6qO&-Uj1gMq+w>=B-7m9C@k}*j!R`z2du{7N7MJqcYIe*
z?=SoP#pAWq64jgV`X3C274-*ZS3SFk_xE$<$HI(9b@laQ{U>xfx*tj77kiKI|Gj+W
zZ*K#=&V01<z@ay%qRLw`5-(kdzA^Di+R4}DuZ|i|osF5e_~Yk+53Us4#^bbi?b08%
z3>XS^g(n7+7a9A?e*ElcTA?a?VSL<Do1n~)*tq>0x;9?$9oV$vX1>XI&lGgOo_u<_
zc|X;AxXUVR&t_I~^-jYIzxv}7-Yx1fMeK<myZ*p~hSE)+Epy!z(cXZKU8c^<nVYax
zfe()LIMVp|?xXXSN!~tTx9z-^>{hoWX!ce%`qWHZE8k4FE>3f@j_xSk_hFFp!phYn
zR+DyDy&V01{;(@}t5rjNQ$eY=GHD=3<GN+7<6&8A*wkq*UQsCrBR!Tkl-C^FaefR{
zd;<%){e-<&I{4yOS(8J=MXy^q*Zy49d9EV-;up?q?Z&rU*{Q4U+Y?)sw7xQa&vvKc
zCZpD`^q(mow^ryks(#SqZrw8Ibb8Lcu>E-IBA)!Z|5jG}+wKo98?Q^w@0oeRCn0uA
z$|v`iNSr+K=tNuBZIxM7u3nqMZ;rY>>+Qowo~zeQ`cl5*$(go$*FNc~y?(c8mAnM0
z@3g4q{3}m7gsyZ=?=P#Cbd?q<zX=Ta>dMPcug(r3w9WggAJ=9MkOdKS+iU+l<#27_
z-pTK7?3Ue(&^Bkhc4}CluR76r!)MaAaKorsQ8{$&?<Lt2yRVOIq4$!>vI9?#J*V;f
zKMR`cny-bbat^+(kI8t{*qV_!%UN0KxNt)3FNsBohho<~jv<0qg}LOeOHKZ<pkKPJ
zJt4hcpGA!NdBeO50|m>cgpK|__KVjEIlb><A26YdcD;A%VtDJLj?5iB9;ZU%w>?{x
zLZll)_)A~Lu4qc5|9iY?U!(u}F{FQCdQ$QqPn6-|E8c0}&-nc79&YsdlZo!lp2gbK
zr#}xCEIoI-{b<Ck2*2ZF43Ac$3umwQEnKOMNT1sC`Z5YURP!=9tLw>f^lE73WUX?y
z<GSMs2ERW(*ttAze))|YhgUPU*0jxKshY$)xr+Nc<G1wR%-g!pHfQFUfm-@;Alu+}
z_Q;@r`!`7^8!A4S{{6^&`?G7(q9Xr1^v>zv#QCQdEFUqoJpOro;kp0Z4S!iFGmUJ_
z-ZWIzx1#Euci-t7-?*PH*r9GM@)>N-Yrcf_r)2mBe{KIVG~}Z?!}pakG#{@yZfn{;
zkX>s#q6jN4EByN?9K&q9jn$ERLdMw>sk1SIi;at|c@e&U^Pk=8-F<7~f>P{>6rt^5
z;~_IIVq11(y~nb{znr%xG(uZuBeOLxVteYFXQqaq%8u9*X+qoG#(ji&5!<_d+Z7$u
z-;nwqmMpYA-8_BFi`X7q!k1sHZGCG`=!CYnjd!?t5!=7a9d}Z$GzHodNkZGz#?{um
zi0$XkrjeC3(SZ(_R%rWM`%f`1V*9O=_N(5-7k;uQW(jR)8)sYdBDS%j_vaPo36oEJ
zKAv{w=d$tKo865M!^b%oNHrNF-l3p0WC~C^GSi}v8UZKLAmBtg1e}zb0!~Up0Vk!S
zfRk3!z)5Rp;G}gla58EJI2jECoQ#eEPFBqVC#zwBlhv`n$*DQu<TM;`aykw;c{LB5
zyoLu(UdIEcVzk2ZuRSv;#wvV(@>L*FiEpH^A;YgD*bqHA-}2DOd~z93CFE+{B9R^e
zE6PW}ib@c)QhEwlseB4psS*lWX*~_BbUqENbO{ZujGh5jCZ7RTri6i3R?h+}o6iC(
zTf#ysr{{o`%jbZVE8(D(*Ym*2=kvhIm+;W42vJ0c58uI{n4w4lYL+5PRN^ea?aVVk
z+!%2NgxZ<EeX%TjG95U16>JLGi<5m&n{e`1sEq{KLRdDDY$M2asJjWW2kK4<c}-&R
zqYfNF9_qyr<RKG|AP*S{1bL{9K#+&J2?TklQ-UCm>cA<;qk3@)@+cEdK^|o!D9EGQ
z2nzD3Zi0e5s#8Ki9^HY{kVp68H004HoQ6EwNYIc+w-Get(cJ_Md32|QhCHSNXCRO1
z#Tm$BOgIC1jFDg<k7*+q$YZ(*2J)Cr2?Kd-2hKts+l#Z1$C_{!@>nCmLLS>ju#m@g
z6D;Jhoe~!ExDK3yJgyh#AdfTQ9OQ9Ef`dG+jo=`U>n1qJ<2ofA<nbLi4|#kq&O;t=
z!g<K!jRX&Qd>g?-9^XyykjHmQcnfcO<rD1Hi={odl_A(SSSNnO2_ah+yU90viaQuQ
zCV51P!!|Y1nHt<Y*;*(E*akR??)hU<I=_uxxkxB0qzZSj{6XA|4o(-W5z3wso~h!$
z`Bf{7Yg->YY#A<VJ@sO_@@bT3=(P<VLRsb{a}&pi*tWGZ9}2+SsKCk1U@o8v96?Q0
z;1tw!1x`cFRNxHMYz59j%~jwW)O-cbTlAr!HD&1!>#Bvj^^Ww87RO6t=_zYX;a;^+
zb|2~PCJwzb%H&+@sJ%l%*@bkGi-(Et9+ejd55A0(3R6%@`iei|U_ep?6eXpAqNOxY
zjFbV2m9ju_QVu9y%KxR#rA~RA?S1Exa6ds-LF2^@`_bR}#R$hyRYEzyHNaaei}yv?
znGar;Er4%~cnlN9{_q)0G!01`HxXe+hhHRD($!ydi9ap4hmT~UTRe^V*_lsaVp(`R
zh-Adu5R{zs2TDnX0Hq?QSrkGD<d7VJ98w~XLn;Jv6rw<ml2af@DJhVnR20b3hz2=Y
zPJ<k+q(P2W(ICel2ILqy19FU#0Xas+fE<fhkYnX6$gxTm<X9C8avb77j+1jB$0<3G
z<5V2T@rVaGUe1FYujE0FSMeaHpcS0B=faSXEBt{{D1t;KoVN?+4j;aYi*kxnhA%RQ
z7xSOQ@Oe+_<b}%3|FFqU&&tTuV;N!#!?G-w>6S~q<ucoHnNyfsFkg^y^AaRh);1%g
zw)PJ1IXRAukwrx!nO#hT^QgbMxR0I`8a`PyW$LsEu3VUJFzx;Ss%~REy~g>B_aiBW
e<tO?F1O|nSReH!3GhE$f#?B$siPNL;`2PV1-&V~4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp.CFD1881950858A56.idx b/.cache/clangd/index/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp.CFD1881950858A56.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94e3ce438b59775367fe76d2328d72c342124f02
GIT binary patch
literal 5856
zcmZvgc~Dc=9>;TT6vLt{4-gSV#Jlerh=OrJi2_0tlm;Q8K3pSc2ebjI?bB+gb*rM)
zj`HgAs<h(LTH2_s7MHQX7_EEMSQXl0F@1LGYh9}E#mwCD`{QMXGxvMG-)}y1Z%*#X
zr6Aj2xE`xgjh&HQQ#o_t^d2geD$@T~y|8NT6}3vmg{o9F+rBQ!U2t<)#MKWyh4QRz
zGY;)TNl|C!uc?em$y>i`bVJI@h%9S%+uCPcwx;phhWVy@vs5KNy3W*&>gX{L9W36>
zZ~mr{+Fkqmqx;Lr*YwXa8o#$#mNf2pW!0~zQF}XE_i4K0psBHOfF!uX;|@&a79<U7
z|Gn<b|2!X>c<imMd-gs%npJmrTK<_+6E2BoryhA)_vE1C=<#fE!8^YVy>qtWD#|Uo
z9#VYwaEGPRRQY9R!F<Q<+8;hRIJHt=Hg96?!l2X=T~6-q)fd*BO6gd)`LfmLxamv1
z)rdd*wq-Zz-G8A+`i}LC`rNgqC4(Ce2`}bf@MZ3pIAG<TJ54p~K3Ej9F0-vkHDHlp
zcKP(YjYCl9fcOK=_pjeOS#L~APQMzG_}X@ZGjGz)`sU;YaXG)9YF#ijqUQux&8|06
zqvq8w>$MCIIhPnd_@_56q;2fc)VQyrrl{W7Q9dbV!}9R`k*(>YMn@-R6@EN6esNP>
z!>5~1_9v?^sdQH#FgI&DAN>-!FD&zP;+69Ae=c>OuxFh9C~A4gpu=akn%3UlkyC42
zSwC^d`w`Xq97Tg`{!#c&tKGat|NW#H8#i=DeD&(h^xbI6d=CG-{Yq)ubI%*!IxebD
z?il}Na$e4c!gp`IrjCf8P2~s2Y^pC^6_dCw<8t4tlb_$cm#}Qbuuu5S4~{u+o`3ho
zmW$W-E#+;5;o4#StN)y>=_1Zv7}H+6R(+wSO8do7-DhVXzjy8x9ahw`d+q%#B^`K0
z=C=2@{A=X?Ekk#XcyVca<mJqwmT6BTn&z5Ued)fGJZw{jrSIgda%#)(wldLkvCm;@
zCteV_=i#TnQONpdMa#C9^TYM!A3tr(o_4Rfby~^fC~ZynJR#@j{HpxDIV<jGV`)p%
zqi3v`Qt(4XyJk~cUU9p*6zlub>RG2cDi)7S@AqBKM^E$0y)Sa^(8K3{*!$?|jMjUu
zlFc{bj}D)>>DQ%&Sg}RNo%tkZ$-b%7e-G{3)hw;-k4u%s#)3Z{XfraFyeOJ!ZTtKN
z+i&HO{J55c>Y^zRf9kAQc;af?!OSa}gAet$+*?MRI)3r?yf=$7i$~pf`YkbZZ^Pq)
z(hCn36HmI{98sj*9=_sGo@MYKuWwzPJE!hadDxS28ylR{8L}b&+KeH$wocs8etGuB
zUBTt!k9BOJ?hj>}Vvir_l-j;99%-__KIXeV*3jc`Pt6+p=iV0)AB%I2&RyJVRNchi
z8Y@ry`+COX`bb}&=CXBNt8On@^&;u^SC_ts`>JBIp|vWxvt@S68C84Xw3M{ZLmv;<
zy|vBqU40}m2Q?fD_PyUxwk57tm8!b7^6!0-N)?pkSUYxyZeS=j<rS4BIw(3gF!;a!
zfmiT)&HbCpx+a8TCcm8&l#~$|6g%YU_S<jgrWS@`g?>9eC_XYUD0bw5mGO)A|9mnO
zo8-46gCc_igJO?*_N%Yy$Jzs-*i^qA7ZlekFevugMM7TPfCes1Rphr*d!~*I42s=)
zOZ&%rW%j|nRRw-KAvPg7FevuV(+3{mwZ5UDn9*;?1jPgg2E~3JQS_O2!KoiZvB`cr
zDkv&AFetXFYUZr!Un07%Phv>mx~w_zY<u(FjDcYm+<<2*SI4hRcuBV+3<U5969GKZ
zKmw06k-(!26!0h$1w7h71CKV*z+(&y@E8*VJl4Phk2SHt;|v_|I1>jv-oOKoH}Sv|
z3<B^3lK?!?AOcS`iNKQ#67VFG1U%Uw15Y-|z|-~9CH%D-EV>jOp(y2=V<CZc+5Pck
zq-2<skg`hIkihEf4&i2ej`9xv&(Ml50?LLjL*0@v6Cg@h2@oZ01dNhq5=2QW38JKp
zgi*>&fhc9AK$NmkFiM+g5T&g&h|)G1Mj0~$qKuURQO3r=C~Iaxl(n)T%Gy{M<;)z2
za#jvRIU5I~yqO14-pYe0Z{uN9Fbg0mSOpLjYyymmW)Vb1s|cc^O@vX&EP<$Gl|WRo
zNiZs#We}CEGKk7H8Af$FU8b_8!YsORIwMe%b)||@ZhI&saJy61n{vA=4GG-llxhFM
zb0Pd3SZlZ!;kTjoA^aDp9T<KX>t36<6T{n}_F(u8sNHJ(yt+G2xDWw(gclKzNB9r{
zd4vNaAdhfj1mqDOjDS4CttKFkbRiP*NG~EGkMtoD@<<0pLLTYFNXR2S7zuf#TTMb9
z<w6wXQC>tr9_2$6<WUZcf;`HJQIJP@FbeW0x0-@H+J$JyqrHfRJlcn7$fF$?4SBQ^
zqaly>U^L{>ZZ!>gj0-W4$9NF~d5jM+kjFSM2J#pu#y}q9!5GM6+-e5$SQla;kM$xJ
z@>m~YA&+%nEab6HjD<YbgRzjuy45V?aW2F`9_K|I<Z(X4K_2J8ILPCi7zcTr2jd`*
zbE`SX<6Ve{Jl>0V$m4y8hdkba@sP(mF&^@G55_|t?^g4WC%BLRd4d-SkSF+%0C|D~
z6Ch7;Vglp|9!!8d!L1e`Pjn#>@<cBZAy4!n5%NR_CPJR*#6-vwJ(vi2qFXIOp5#Ij
z<Vjv6L7wD8668q^OoBYgiAj(rc`ym`B)3|EJlTb0$dkQDhCJDaWXO{pm<)Nc6O$oN
z_FyvP$!@jW%^OqyK=tI&!W*cE1^&p$(`{G~%ddV9g<0Z<#gA2fb4<yPB0JZQ=;@zM
z3r-7Hwum|X3*BE&cyqpgTBp&)Dbug#j&rf5v`PNygv^8~%1_*-OB~Bv@7z87a=IwN
ztW4KG%t{!3es#QmIx-?MR@ruxH}h{lFMl*E`^?_1w`+^<Zd>b5Cyh;-plk=tHAj0+
zs@!StPsjC%i&ZwMW7>ilJ+pRp`KP1tXkK|t_~u?+bz0}++}{3OOM|B<ZxUu9Gz3tj
zh6IYzP(aZd8Yo7?0L5xppg0W&6tCfd5;Ou(qDBNt(nvtb8u>4!T)93WfoBEf()drt
zj_~zhizV!cfSR-;5^BnhD5z;WqM>H&h=H25BNl4TjyR}!JK~`h>_~uGv?CE}$&Mtb
zWjm6)^@%AX@0TTAJLCWG(tD&2QWnGyQqM=d!uziBPp8GCB`MR&lTvs{;0Xy_6eC^$
zK4BvP6lo-ZqKp(!w2=miF)~20MiwZ}$N|M0d7uQN0F-DHfs%|8P_j`5N;g24tejex
zrg<Hfe(>d&G6s*4lrvQx{)UAF9{Ru;6FA*{#yF2lKxuIuP<niHw;~7v^a!2+Jwi)>
z9-${dk0eOYBY6_^NG%C^q@Dylil9J`;wjLhv=r!3dJ6Ptf(AXBr$LX_(x6A{Y0zT`
z2J{%70X;^`fF7e~K#wI@&|`TP^jIwmdaRxWJ&xc&kK;Me<Fp*;ae5B)c!CE#p65Z2
z*Ycpp>v_-<2m$m2UI0BoD}bJ$7eG%WM9>p?5%ff@2zsJk1U-q6Ku_W&(37+h=t+7B
z^khN?J(-t5Pu9wyC+lU<(@{EB`A+zE99|~@HAI)HsH&M2(_j7zq`IoSy71*e$;%<|
zQd^=^M-JaR_vX`}vf|QdC1%w$#ZswCyAO-I56#_&W!;D6l`|^l_+{>_JavzrLA^A=
zp<%rv!uw=LW{o8xL$WiY`u;6CuHUfX86)%~M~xO@*z}Y%D(U~t#`aH09GE<4Fiz48
dCra|rRNVk=JU?VyOzilaS208Wm<ed&{{RkJ+%EtC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp.9CC79FF6A90215E2.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp.9CC79FF6A90215E2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..91d0f1a25411eb76731a1b075f62568d950a6875
GIT binary patch
literal 4654
zcmZvfeO!}OAI5jzlerfV7;cKlo5cNo6xjephXM+M!VnRbIHx9tB$9#xX{Kl{6RbeZ
z9x+Ap1>X(RQK6`XK?o$2flkvzQyeOqFKOA$p67PIKc0O)oSo};U7vl=&biOoAvrE4
zhER$`(R1PovNQ5#^%03g?)+bF-lBO|#3B(=h(rbDHL2P9-$%GLZL8MrucY7B9as8p
zZ?WD{?=_?@E1tfz)Ufbwx$W4a<$sqvjH=*B<CE2GyC08!thrZ4h-53y2WMCJ=^nT6
z^EdU|J}SV|s>ZZ-U0U~J&c)lE%SN|*l-G1Lf4a2q%^%Bt*PXtRr@#N7*$bb$Ih{zm
z`01FFE!WRI>B=j7FHApd)4GSlWfvZG&HilYDdXtId)+w~)_Ik_aDE!TWpQnh>UdOY
z!O-fow^A>;e7(Wz(1{s^#Qdr0&1r*Lm0Kgvr<PnOr33N;9;|q7zS@@6X}Q!ruscy%
zaAR6)Sz=lyez0Z2#+5hOwul8=tA_q_=%?)Hib+0U`qPtpd^BB*&oFu?UeM-wu;hJ}
zZG&v=t(Jm}1X8^%;@yZ<nTDnF^ms`8f+D4Ej!WgTh3yyAflo{1J49<ebMo{7N0(3q
zeuf<_mmiFMvdpz(O_AuUs*Ntqc~8fMN~Zc-@1G7CVz|}!gKq<Jn6FOt_1!lk<3Q*A
zx*slUYfC~dHH7+vU-90tF><eE=d!!2-`iPr<YwZ+%03lE$1is#Op47(+V$<8+UVU$
z6F)ebI-UL6VlG>EdQEmo&E=2Z>|dJc_O08c@HJS<0?z4J+0WA(tQ|{#uv}PGTM>JL
zAOh>D@pCPhv_x}b&*~2^MDPB-c>Ou=JBlvqWzfj?XZjz{VgqY(u0O8&B6PQQ$s04z
zZe4q2@m$x<w>;16s}Ig=Dfs+%k@j=vlFr7GyF((zJ^Xp?im3-HXNO;t6w12Xz6q36
zK4I&Mm-xKQYLPPj`=@QX;>1k5c>UmmW75RV^%uu~k?B0L^`GvZz)LmJLqFWRmYO{v
zVu?C%?EBl$fU3uZ3C)}LEuX8LcsIl_kBu*qbj|vHeB94*b@QfHzP)_IiiGX`oaXt*
zH<V`&u1K!V{iR@lx%S3}$scVlFVV-w1`Mj)aO&aa6JwMG@%^^Ed3OGT#3R2x{AJhi
zx<js0GM4?;f9$4<E<YaKmuz}V)$sB_mmivReVS}i`cGwTH6za7snUAsJ91v%wJUDF
zcm0I6#b@^3*j%-v=xQvvGvIjg8r44akk735mABgb|4}V%elaO(@)*PO9|+C<vZu-A
zQ|`5Vb6+ue`k&rke&O>jRi525r252`8%v{H%1`d;GNn8lx1sg+@}NDf>$AeWn-cca
zN8a^xn%D5_%AAoz{9ncL_`1U$wo&cs_*-pB!>EqL>}#$q`}VH8%DR2>S5e)TI@cpR
z&uOoA5HCJFS#xUO<a>VaoHChGy3-%MJaThY-S;yyMuo^XDI)Lb_Uj}AeOm_~t=L-I
zeDs;^X=y}pP^8D^;)G1gYcECQCTVVNVfO4$OeB){TGmHbg!w7(sRofrB~i)j4Sx>U
zU-G}UUs*LVWQqb$;cYL8m%F_YZ1=-!2Q5By=d1!x;%%iwDYG|%9kd|TQg+_wnF4>E
zxBVo3;r2$b$2@;v+k3O{umVrzZBL13Uwb3iX=m5I_`~*HgR>}&w}(3o*V!AvKCn{$
z*2#^YN*B=--u6~|``a7AK2j<>A@Mr=nF3GdZFh;g)ZPg8omqZgBdtxN!1cVXmZ)X+
zMzH^OOa0QC^W9AaKAE>YBpxz*BiM^F=H{xL9kb^lx6fM1ndjRNTpQ)*Y$EhTvheMB
zWgyZWiqw<9BMl_*C_M!{%0K~+*3-bF4K(l=Jp(+(zyOccv%q5wEburz2RzQe0T1aB
z@Q?ukPvfKUf3-d)O@M|Fl<>q1mfN2cFeOGJ(=iD|7O@zpe4+?*NYY4xD49iqD49>f
zC}pHTl**z&l**@Ils3{JN@vj^O6Sus$`~0CWwIC$W%3yqWsNL|vRN#MviU5Gaz+kB
zxhxJuxqJ>rkr9C?%0eKD@)3+`LN$@X<U5-*(V9e{k~Hap685T>+`dP_!XP5ZG53U)
z2m`7|O#CT!XcCbiOKM4wC40mq%u<K~SxQTREY%~XV3tNS$kJLGWa%C;4YLelK$g)m
zAj|ZK8JJ}e3$m=11zEO7%)%^(IFRME9LRD#Vh&~zLLiH@2xL)@7{RQD(+m?9fnT8E
z8m*v&dBx@SbJ4LRVF6)jLP$6zN6PJ|qa#E#V#Il{YX~bwT!Gq#5#3N*aH1L)nEa_!
zj}r}0x8p<y)J<Zd8LoXJMnWEG#Yo5_Z5Ro8qy;A-kF3W@$Rpcv67t9<F$sB8BSt|U
zWyL7Sqih%jd6WgGAdjlYDafPRaSHONCNTwhbR$MX9&N>F$fIo-4SBQ$ry-B7$7#r;
z+i@E5=q51@c}ye5Kptbo7|3I67z25X1!o|SsmB?}W7=^B@|Y$u19@yC#zG!z#aPH=
zZ5RuAtOaKwkFCd9$Ya}a7V_98F$;NIBgR1<XT><k<7^lQd7K63AdjoZImqMMaSrmh
zCNT$js1ZYuhpZTaJY>TV<RJ@=AP?2!2=Y)njvx;;iIIaBX}&9Z_9(9d>tlkuH6g4t
zM~#X<#+*%FBfO%8U&E9{CHcoEqn-G2uq@bB*xSN8CbwU0guR{VHKnO$!tw=tS|imA
z6jDnjj%#EMp-Ft%JIXsnxB#tu*RuNbZ`Z1O%c<T*p=`cCe(=cV;z4{_>84Z*$N&DW
zRfVr#D__r-{i6J0h2!8r!8cAx**jwRvS&X}wQ%SNA10IAFIU0<`NPAE5vAa8A<Y;G
zHD$&qsA)4sL(P~m25Q!fu~2hnjDs4PG33y~<xVcQU+TcHNP-r=lCy~<NuVf_0*WSS
zpcs+?iX~Z~IFbVjk^E29|L(Q?@=zK^Se|2~Z0g77D{k{s=<ep;Ul`*_ZSowa@jJTs
zvMNx8gbR1=^>=bZ|9Co4%Fm^g2ylEPDJ6lTq!duJlm?2CGC;9X7AQ{20fnUKl{%9W
zaogm3@f<&%&_1C<gwb~fIX&v@I$#}N4ps&G3T5G*%+GZ1B^k_g3=szg9YX>|#ZW-e
zF*Hz23<DG!!ve*{a6q9L1eC^8Gguf@m-N6#c=(atO{NZ12MJ>oF75oo+IwRMW7QF3
z!B}-9P?U}Wiq_FUF**h)R>uOx={TT}4!u&MMH%yF_5P=jyXcMFDZQQPz0H4L=|V>&
zGQR4HP47gE4qE>A!DPwIX~y*FGepx~d7>GP?o3DbjqEw|GkH02Zi2XvlccXyrf_y~
zbL|)BK0caom&ZjZ2YB=!=s9BKsL{HJF=L0S*zkZ*%J=_O)n0?W{rrc#7D&<zH!Ntl
ZCM3*9I|v1jQ>kO)XW=o4krS{<{{n%jKdJx#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp.577D62F548B80AD3.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp.577D62F548B80AD3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f417033e307a71d67ed5b161d9e23bcce7030b50
GIT binary patch
literal 4654
zcmZvfeO!&_AIHyqA06kRR43<<^rUq^U!-#?;zUI$)j2A4CP$boO&UctX2de9Yzni<
zh<PZ_%W~QlvEQ;R)vC>k7!xBW#IUjX-JRd-cD{f7&g<3H{k}e*`}O&DU0qk-;-exW
z2&G6Az91?uGc9M{AdyJq%ztI)EL(I{ED|AwNR(Gzmz1gheVkMCu3G(}O8RYGjnZ>>
zYxiCCL1WUY;@K-pjZ5#9TTVV+{ZGlGunLYeJzd|i|H;HBn)_vhNVevpe`f8Vz9~yT
ze^bBfqdYvhdQyAu<)R<6F5T%_HL=sBysoR|la+&S{aE(9?(EGR{e#cuFMaN0cPjbP
zCzDRM-Z=lXHz)tSApMwaMUTeH+8+1L|8(>j)5NCxeOYZqZly2mp9Sw&USFW92}{Zw
zU7P$?(q)IQx40cWH7B1~JTs*wc|^N%XXwSGlD1OXC&%aEn&%s?b!7CIFZT`ai&N&^
zoYh_ymz<6tX`Qxp?Jc$=WXaCz(f=O(DKoqx#v@37c1FL4rkC*;L+{1&I$V#Gysxrs
zkxjncnwJ(!s&|FF8?r9lxN?ym4~Sk;pwumJs9d$QvrX;$tVF&?w9z#yNAGiD1(oMz
z+|zpH;pC^Q9J>k&L|;{Jb!f?XHYHFp)4Th@*?>{T+k-#&)+dYk>dauzsyS(gdmc3W
za79~R5^(uKphxgk_dQ!f51RL`y0`wlz17EW#VxHIR8dfKr6)EfGRv^<+XMCC`wi1S
zIFU4){o1^ttmtfEW=Y+ZkKY_xn(p+i)8*hoEO7~Ece3o~*%!LIR{mgaTUTEZd5R!>
z8>y)a&6u=AbMwIZ58J}`e_y=my!%~6FZI%I{QGmgYcg2hx~v;ds=o-_uU+xR-0ya7
zyt;g$<M!LG=c*e0b6WF0|6Qd0+`gozspQ_MkSULT-neGwk;?hO*CqL~UZ-z-C6!Ov
zhT;_-FEd)DjQ62g+pgL%vo75@^6;cIu4mJwsb8eqk8l6Cuiy7_UHIq^4{oI94+~kL
z_MQCxE;OwANq%h0_NvtjmDBG97#Fe81(M!*-%pMDIjUjN%*wY{Z&?$&dx+g4@8}EV
znIkLWYqNjJ8@8eT=9U>BZ7(m;M@IULsN8bq(e_i5lzGuZcD(uB;)ij^e|_}JzM6)k
zjtOb2ej7S@+a-q|PgKRHzNNbG@^G&gin%dM7L)Q*Sx4Qti+8KFZu+jQ*Z1v@Ipp3r
ztz-GQgEzNV?<u$zN$&Nji7!-DsYiX<{XluU!~0*=%9a-~VKXKfpZ`E;4wXHNFHgAN
z`ppByjM;y>fBA*SyHt5*|EStiJ8rHFb0|N3pf@$~uPIyF@2vJa(7q`n*u6RSKx61V
zSGz?QeqEb2o{0XtSRUPQ%*8UHQyqP~!!U;Gip#w2*jjb4=o;&^`R{^;9Sx4h_ny~Y
z>mpu!c)IS)@EP~L-Z_(+n%I}}_~r3i>l(hFn>HapzD*H&Uw23+8SdFW@<heX`j!)a
zS)P@K6#Ip`Y%h*YH^25$MDCMjXXj_m55z<wiKlr}ctwzx0-tFVrK%(<nRVdr0qaNp
z_wK9frUxV_@I+pAlejrs2SIi|wsFMrqj$ek;09h+N|Z9|Ajp%RKeQaYm48fuC-JhY
z#C5QB5ahQ{zqmi-F*V0tl+4Rx?Z)b?gCHL-m7S8f9s5*)$MdqY#93+`1i7`Xs%G>U
zy^8{WotM2NUcuHukpIx>PybX(cRGj?c-dX;?rj|e`R+WguaVZGQQ&%BR!h_}>mbPg
zI3<1Ao%Nqv3Va4HyGUGQ)<KY$r7g@>+1pl6F1M~)$+_pd4_}|)WuHnIi4>tg`&80M
z0-}r*AlgU+VvGzR*2n_lj2s|jM1V9t8bVl3Vb4a&t$PKEG7;&vqKFJ)IiOsk0J9-!
zB7v36Ac2+4C83ovQNT)NP{2y%QqW49Xkeu?XkevtX=r6k46rg846rh}479Q)7FgK~
z7FgL_7FszI2drEM2drE!2d&72fE8sRU`4qIS~Y>1P+{@yQ#Ik5I6wwXia^3?ag$q5
z3D_7!_}SK;&=Ns_`o+YbVjGi)1YS~00x#JwCZU%?6!2173V5l0F$KLeqJfvz(!fji
zi)rX(5Cgo7mH}R-U(7%+i&)@gwJh+m{bCk+Im7`kr{#c`>lbs-ix2`{q(#7s`o#!(
zHJoOQunGJIjn!xc5^kvRa_h|mwxkIoE`kF=bYsL-$QF#~gKWl$THLlxL?cdIfZT}_
zU67l_L<`*3CX9r6WH&~_Jko-ZFpo6jB+MflaT4Z{oj3{e$YwDK^Qb0_f_YRoM!`JF
zf>AJ!GUF7?qZ)Av=24wE1@ow8F$MGJCX9x8bT>xBJlcZMFpoClG|ZzLaT?~)oj48i
z=w>ku^Oz=#fq6_f#=tzrf-x|UG2;x(V;XS=<}sZ(1M`?>F$441CX9u7Y&XWjJl2A-
zFpo9kEX-pYaTeyWoj42g*k&;c^SCCAgLzyx#=$(!f^jg9Gvgf0;~H@e=5d`k2lKdQ
zF$eQd6NX?O>c$YvLlz9dJY>cZ%tMVhf_bPDM=%dHi;*oabi+N-UypOTutBLoK0(RC
zX^jmk%~GS{k1_jHw{dRa!XLWCI3@YV=81Ovw7<;XQE;cbB{bJfTfKyz)<`wOg=yhC
zCbxDQ!jWGweM%E+3^ed=_b~TFp?z5Uu6h00->%mVOeeXUgy{_rrj8unQapm6Ryrxw
z!p(lDcU}JL*ULBY(_Ud-k-~LwIPV)fL*||ce%f`2t6I2pga(K29tOG_p%T7(APHJ{
zm+u}(5)egFfM}8i#E=XimSh2OBnJqQ2#|)-Fv7B&BxN%{zF2XGuhrSfd8kn1X>I%h
zyQzD6`DvA}3JGmv{f&3B1OIq7UCI|NC44~9QW6j)r2x@V8W1C80I^aQ5GUmTAt`!A
z=MqEiq<UUD&({+;C~%Zey-)xyw>E>eC5696J$`nN2RjZc;@|aG`FjfQ3VSP)TRTPD
zyF?*IYzB8dS%{I4sX~l`Oc!D_WTp^fAhU%S3z;j#ILN3FLpByVRTsInLj`3;5K+R<
z%P+ryNU#wZK?035kU*m%D4<aW3TSi$4K&(71C5DbfW{aYps^7w&{zWtG%kVz8fV~u
zh9VHqkO2Wr<Ej}c><xc}Jv80|^-2$aga;oVxChj}YCmCxgr=R}z=4h(tdx$J4Aw?R
z0-|&jAX-NQVss23R>uP3bQ~b0L$64*EN$_;f&Ub;m%WjlFz_&YVEErBMR*X2Os^ir
zshdALoa|HdbgN|UEK|ztIigvwG|?Q}<6PV08<`6hr}H>{VXSzNon)|7rm%N#avT!n
zJT;tfmPdsthq(+L?mBM#go(P4Ns~vb*kGSP%Jcuts@+DqdwGv~&6lJZZj9epO+b)`
Vb_DXDqEbgj&%-0)LZ@Ld{{`|r@O=OP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp.FC0BE921C33562AA.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp.FC0BE921C33562AA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2075c68304fbb5b01e516d794ced9ee8b069912d
GIT binary patch
literal 8982
zcmZvh3v?4@qQ^5cp{5TWEwloK7UVGj2_PhynaukMZG?s*5EjCtRA4J8%frBes63(w
z2$aXwD_$O=f}$vb1VP=!tKLc*SYKGY<+-q+Mx#=N)pc*k-68*b&h0t<PJZ+KnqSV$
z`KB|O>I%zh*URD1YbzGczIXmqr^DeWO#jZCKlk^4ryP!r`3}dzt#6N=U3HP|kl1#(
z>g7HC=5(8|s(t9nz|(E|Zz(SKoO-(V*lzp`-t+e4hc6xtxsO+FDz1L!kC$JcQFda&
zzdL$6{KpRz_jA#!HwufdAEk@fZA&lSe(J>;X9rzxp0(t_^X$^TCoWGYo&MwUe|X;i
z%lR!A4zD`?^^x6i>0S1h8!JoXhbN|PB%9W3OAM;|VAr`{zrM8f#Jk(hG~IvyzUx<u
zf4{WNl4nQq=jw~!)W$B{khgEk-Z$@@Ke6+UQ<E0;ojqp4ppveORwcGiUu|98)UVfq
zjThSeEh_K%>{#xhGvPQl{M@9OBdShEcH|z=HojDFAiVvjQS%dTT|Z(RYPWL8qE2PY
zp8IRdjL*m1QC4lWN$no`yVCsp`c1jYFDDn?JL1JH+oXr2*zmpLf#)CXI(q7|da<_c
z;U#mwsDHfQ_04(vAL%soFIUdo{lr1NANl8gm7cAx`H7#m&vv^){DkJz%y#9EPYV@)
zoLfGz*YX8RAGz}2{AVg>A3vvmc4W=LZj;tGj-ZZ(-*|P_#T{=PQ4b7pZJx7f-P#}I
z`6XM9_1n<DVt4OR@uRD*UdbKv(c(OzHg7@RU7uIZG;cdJ=dGI8rbediW9ARdtLW{l
z`t;2wr#d#528rbSgTp^*I=%S6R(!nt*p6X`dI$UWseWQ+oN%ojI{!qc{*~=Vt?tr&
z9slX=pO5<O{FM&NR(|o)D_eej=kYK1)t`Rq<1aQpx_M6DGp{};PMK8eTHW{fN%f-n
zBz1D(9bJ!)f9k;M>m8bA@Z{KoFFbK|)196A4A8&X(s|c!lkS?{d(p3l)Lwg5?5%s{
z+PLjI7kJLpzi7_sAKf`{=c*4*-h2A&h<837`=GnFP~5QUi<Lijf6lo(^~Q}$2OgSq
za_9>$ZF{=5&re+k9N64Pr51-><Es}Gaj*0*9{E$zYq_89=rO3ae#nf{VKZO)@7(*2
zuGY_8SwCvsQ(M=zAGWwx*FBrw-T2g@hl>^tfAE<Z@6Gyt<o+*y{QM6Srk#3hVwZAf
zoAs9#yf=4m|8MVmMY}P5GQVhQzk!>LGcT1@x4gV)Z26C~mVCPO<@$e%jf)NXaNUE}
z{#QF%V?Nxvx%gKx&^h_~${~A02RAibtK7ss*Z2Dq-L8F5zBY93tUjf7<DSndmiQ98
zmtR?XedXS*KR;A-YTT)_mp?uFhr!VcJ(~8c{Vs3DGe?KtpX`5QLx(?2zWBr9{?dj^
zwfnpO(6H+b|I-D(^xgQ^w(aJ6uWVX!;Zobjn8$`x9$WEx_sP5N-g2!=@uqKnE<OEZ
zXu|52=VSFobGzjwzd80jk@Nn2ZQO?k&0JRZtD{VO;r-M8^WtxB{`S7{?((l%H>7s`
zFnQ<0%bGqMvFp*9{8M>bf)x*pds)os{M{$J_wRW2$93m#-1pBl(we~)^Zw&{U+=s1
z|0D(+^xT-YXm-s;!r^eXDm8cBy-;lH82vj(P4|-SJu^f49Fh6SDUUy??_hf5ljG9U
z#W}^fnZcgU`}_7Y%Y(kr`Q+I2bZJg$VP>$W3-_(+y5QBX-peOL>FJ`JqTI}2PxoK{
zw~3dgzb@yK6VlT?b9!o-!Jd|zhF$K`c01M9F(y4-R#Mh0Gu)c~<@@CB%Zv8qli~Dq
zmz*xQWd?hCY%J#ixnjWHd~#xXdSK4LKxVL~pM7ZUkdsyG=CpO(lb$XwDes*b?CIC)
za}VW|?t40)tWHlC<`lXzgFXG#*avzJZ2LM<;258tu4q%CWQJSQQ+vGS@g|4nlU3>I
zlAMy<%wSLdqr=$!Ef4(X<$Q8fdb(py$K1?dPdnz`JAK~rZrNi)=4Fn_+GD?L-+P|#
zQBYGoyn3|#5$q$R<YkUbR-!Gs!rsFLHFSs`4HTv)WXWy0!E}e*V7kL@FnyM<!rrk3
zHNKE<G*H+#AxnPC52iol2h$(+gUMJ7m`sQPlL<3mvK9*_8)Cs^!z`Gb#evC%I54>|
z2WG$ufEfq{zzl=~VDc6ZCLiL#<ik9ef+c_{gaj~!umGlLiC~H$5lk^Gf+<-Nm{Ldr
zQwmF9%9aeK9FoD5!!nqPrGTk~6fl*r0;X!IV5%V%Of{^6saYDBT1W#^3u|ENmJX&K
z(!tcjI+%uKfN6vbFpaPQrfHdAnjsTRGi-usSr(X9$O6*}TVMve1iRUbt)M1Y7VHVs
zE7;d2`&^h_h&NA!un;5k%<N*L>*xhQQF<AXMR&vvs=Lk&sypgN)fe&2%q}-yoo@k9
z)VGYtqCesX)nDfa)gSev%0w7YnK}kkCd#17Mp#hUIu=wm%A(3eI8eDd4pc76p&Ez;
zKn>IdKn+9#sPYjWRKAV}m5=hM3K0QRp-uo*hzh8R5fN0eP6Snqil|Bv2~??00#%Ai
zsLBx;RJl$DRgTK2DiH-#rA`4=i7Ke75fxOmP6btss;Fua4OFd8167M^sOk|NRJ~3I
zRgdbZ8W96jqs{=;h#IJx5ffCi&IHwrny6Y43skGl0@aFIs0RH()n0(<#i<8J0)>JT
zY_hLJ#JtQE2rNzyJvh7mXfMrV$?b8&oZIV$IiJTjIJ?4pULOOLqI^G6S@wJU;QGCO
zaQ!LDk1pe3z-7D)xJ-&-&}BU=xU822mrYSDx}1jtm-BMqaw&>KH{c0?8}J6e4Wy_5
zy1a)6m-q7E@+pc(SMUhn3SI$RAw>!3iXIVM(JO*0rYI3z$s>U)c_nbA6eXc6dt`8B
zuMDo7qGWUxj{>gZRlrqJl!C76QNdNcD!6KjQqk2s8n~KQ16NB?8oIhi2Uqv%;OZ$#
zN7wKe;2K^7Tq8vp=$alAT+?fUYo;g@UCU#EYk4hjtrTUU8ypZEWG}k(g1jr}wW-UY
zYrFqgIC>r7tf}f%HQC;5FSqnX?#(-Z?8-#Rz7<GIH}3<o(wO06#@O#9|LcZeEs>YG
z6UbV06G127-z<HGpwA+IOVHmTA13LyNc)qeFP=Xl>Ep<!N%~9V6BK=t%KGjm!i_!m
z8N!V{_qT)_d+x)e8+-1LNH_M}r%5;V+$ShE_I!<muL=I)`&tO!8RR74`xZG)`VN!X
z6O^xk^nHZfO!`hECn(<uJfSrbe(d>M2tW4xNy3jkf1LDV&)-1$vFC3l{n+y-C_na?
zMuNc}(?T%VW0C}eJtj^v*kc+<27631$zYF3Pz?6iMuNp2+d{C|W0M4nJvL6V*kc<=
z7JF<n$zqR9P%QSiMuNj0*FtdE<B|l2JuXgi*y9>V4trcP$zhL6P#pFGjYI%@ffgcw
zy+D!(U@s6S1K0~RkOAxkn#ln60tqUBJ-(6PvB$R%Jofk`!DEk)lRWnL29n1f-%Rq@
z;}aB*J)w~huqU(-0``O?Az)94lLGdH22#MD&`b*06B3kwJ+YAxu_v|=BKE{2A!1LA
zlOp!S22#YH*i4Gp6BCq(J*kn9uqU+;685AdAz@F7lM?o%22#SF)J#g)lM<AKJ-LyP
zu_w0>GWO&oA!ARDlQQ<?22#eJ+)T>YlM|GTJ*AOQu&1;T3igyFp<qvmlM42f22#PE
z(o8DYQxcSdJ++Zgv8T2WD)!VQp<+*slPdPq22#bI+Dxk0QxlYmJ*|<@u&1>U8uqj#
zp<z#plN$E422#VG)=X;H(-M@1J-w08v8T5XI`;G=p<_>vlREbF22#hK-c0J)(-V}A
zJ)@B@uxGRo2KJ03VPMaQlLq#T2GYQu(M%fHGZK`6J+qN8v1hgrCicuEVPemWlP31e
z2GYcy*-V<)GZU1FJ*$zhuxGUp7WS+pVPVgTlNR=@2GYWw)l6F0vl5h*?WvD_<G68c
z{%OitQ{Az8n0>k9tQpgB3{bdZ_!d=;f8mq%yFV@>3Tk>*_N=n6Ohz#E50rS;tWR$)
zWlMGY8;g5J7SUHW2yN1vD|0K`+X4Ifz?ql%<}`abg`c9}|9pyq8vGPfE<8RX=WLHN
zMd_~!TPfcPr~QrvHU3u0kIb}E3^LnFvB+F2#UTe;sQ@zHO7X}-D<vR{t(1r?wNetY
z+)ByFN-L!xtF4rZthG`avffJR$VMw=Ae*g}iEOn}R+eX$j!wKieChq^0}yltJKG1~
z!JyH2XV@R23TnEW-N)Hq)fd#1nPubbOX7l>a<hD#{dM-E55-sf`P})#>38ZrzWWsW
zPaaE|-MABLx~4Z5bto#a?}pc2+*;*%c>KLL4^YhMdmIl?%;`sFVonB`jX7ClF6QKr
z12Ja+nU6VnWFh7hkj0o&M3!Ps30aOgWn?AhRFKt}Q$^NdP7PU)Idx<s<}{GanA1eI
zVooc|_U=tDg`1BGvP*$pLD2PZ=ipvJxRJxd!(;8k`SRDx7mYu^^~v->7^n>#W-o@l
z3y-u3&E9FHH+N~*rNq9Av7bPc<Yk^bK$52W*$cX$#!b6{e6+8h{l4i%M*D#nngL>I
z7Ko!cpa2~J;%Od8pame27J(#M0+MMNNTC%Vl~#c?S_9H)9mt>!Ad@zMEZPDJx`X}g
zU44Mg9kae^$JO*cE$mR($=;Xmc&lsMm^;5tZ!Ydr?6Du6{OQ6U=lQQ(4|Yo5>AC1$
zuyb8*AfL<E3wEx{55%|_AlAhKaV`!h;0gfoE*?m52|%Jt1d?15knECy6qf>|x>O*|
zr2*+K9msGQK&HzCvRu|JI(CnAb#nQs51j=yBb_7fu=jF7%_!$6pnIJ60F85wyG7?N
zKB8<WeC^BhjxQfyKFwZEmj<`Fc3b<~SEn~u7FU+r?|Zpv&)d&08BvwqTsFR}#y->8
z)9Ks$n@`;A74g--9=4@7-+a<$Z~oe~cO4mc?dJcwl0GGaZOeZ4R-Bi4=4RXJBQMel
z9C^_V<P&`g9C^_X#E1+KE3!bG$N>e!01z+oK!PX$iJ}N3i4u@3%0P;!0I8x1q=_1k
zF6uyrXaJd_31o@ZElNL&oHh04gUNYwr_CFE^Tp(w!~gzFvR^nHQ?f53*_WBW%pg@W
z@W|oW9j)iy&zTmPGHLQu$F!_AZFcRfnQ078A3-_W<lN@U%`a%%p?$lGLR0Tsm{*|{
z-QKZN=PoSA3!)?|1B>qr^zysQ|G%vzrQN#s=y^x)KIwDhfI)-r8tP-pyj?v*jN+1E
Nm9?ZbQXNj*{Xb20vX=k=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.5E4A00F1466A4D4D.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp.5E4A00F1466A4D4D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..29f532be88cf6d8b75c4424b93631e06bde0b128
GIT binary patch
literal 8878
zcmZvh33O9sqQ`R=O1hx5K-pUmA&|T8TLZzCKtUjEB`gsuAW#7n(BPACL8wF7M8>BP
z7!k*XMPPiOQNgG4L{69*bX>4t*c3$|ScJMD=%hXG+W$M}^&D>VyWcPUaqi9cCAoKO
zWhnI0O&(AAjLOCHW-h9xJswYC_TR!q3+DVpc{~s0dpwIb9;li(y1Ui>_@IT)ZCjCh
zPmk5_xAj)8oZM^6_$5=VA200Od)tGDf9&`5L)TaTv^>~<?1rjWI~E--J0G8Q+tYL}
zV)vw%xToLU)IasZg-h3o(I*#N9k40AaoUo<zS#NlrR18^SC*GeyHF-h>(TRK|50mV
z9fgro(_aof@Zp&Yr_Z;wmt2=ebdJ4o`O1=e@6-DYdG?`wlmE8z@V7r*KKAzcE7zV?
zPmlvgz1;Jj19?+=Gi6VIxxLf&Ot<Kvc1ODJD(mq2wpYI$v3kym)aE<?8Jd&)OW|AV
z5(lqezq@AdqA%yYUOazo(c+VnKQRK)=W`d`(q}@>v%;8P``fdk*P_oKTXsvlEN|_5
zV=k4*=lvzZfEUQcpLX4}dU@Z<L-givKUzF<<jck1AKyP{`LG@P_q_0Ex4y>U6(z6L
zP96C6+{%jFhhAOUKX}In-P0S^MVyXh?b$K2FD~Bv;5%*Sl~+D;wBb;#Z{g3wdOWu6
ziTOXD>{htrGF9>M81<>6Prdi}wpiB(t$jDeH{>M#e&@J`56Wl7R?eQ;-hc6xk|}2%
zKFjV%jGO8o{7SENZ|_@i=4|eSLk|uwn~}G7*)w+?*t=z=V=nJ9_lr-zINxqaXuyhP
zC*wyym)dP!Fnz%Dn|6LPcK+ex<sUp%((#7CTb(A2eR9}<EvIsNsj2YBBESE}F)KUu
zT<d(=`SUT?Pkr5G>2h=1tN(fJokPdos5|=1zrNV}s21xNI`VOecvx9BbN$S-n=h3e
zBxaWtuFll2z90{rarr<8V*BKA2NxaEekiKyJ!-_QV@oI2cdab>ecngjn%>(dGk@}a
zx@YHsDL*XojykjE$Po48E_p9qPk%n*_1DVzy0a^8eSg4`vt9ZIS3k$?d@1wOs&}Su
zy}GHg>hzBH#!eYir0#yaZ|A=E`PO~1ZS1KFy@(CLo_Rmq(r@d6*Y906{7fqHRdg!p
zJGko3_rIRIq|EN}@$eU~ZaVPR_4V&gf9=G;?)R^ZEj<u_gWs#&zwjCVSJcb&Yy0L$
z)=YTn>g4U=-*)fXzfWAj^cCk#SUdKobL#fdpUhqIuiqxjoc!z_`Ma81?<e2Vrkz@U
zVBIU{myP*dnl+@&!)w=s+Roitz>McA<o)0DP0g8lEq~(v4Zc3(mrg66Ib;3S&nlNh
zQ**YR<$iheLiES~>Tvjo@m#U`(Y=#mr+aa;_k5o}Ys0=_7cbbqJlXz<cMhC;u=1CR
zzpPKrAKY(!>f#K!%!uDS=w!`Zffv3U)#cdbk;#h6=f(`yu02rx_Tbk){Jf2JrAOWU
zN2mO3ES|6@_xL$x-PNU4fBWFXyDuD6Pkhnu+N%HF^zgVXCk8kDSU>2lua2DEb^B#s
zRVw&qx~$DVmv8R8^|KOc)s1sHH+;17%F7kk?^`^$H1y2or$*Ixm>f?2{@Gbym1p4t
z^X^zgcs$fbYwZi?Hm+{x88_Ke-Ltr7@0Lw=g=qQ7DNU>?f8OYoPfpAZcgpFM+p_s%
zSlc?UYSRas+j%Brhf9h}`m}7^;k>Jxk3Tj-8<$U3Wrw@xbT4e#xWk3JR&={>=ehsM
zC&Ss{qMV}KmW?|+=+{e`SH5{*S3ViZ4tLGz+O}om4p-IY+&!q0-H}gD$qx6)>0`HS
z+~IpB3|u2$9rkpAXHs@}V4H!%TQ=_S?k98K$mzZ-o==X=4j1MWdRsQ`@bT}r$d7cL
z6KLldpB*kOF1@*B;|`yz?)4U@W|Vw#bauEnr#QD|;|^bMU-er0?)SgRC&y%mJLGi8
zZP~cP3ufNAFwmp**pqoJ$9(lae%-v|l+&xAx?)(xIQReWo&h_r<;b^6{*b@YJq88U
z{;+=>P{co_m6#9%CKG1BWFichY={Mu4YOdf5f)4?#DU3$IWV~h2PPll!Q{g{n0$l>
zQwRxQ3Sj|EAtHb&hD0#Mun49Y5y6x~5|~n00#k}eV9FsGOgSurDMw^5m5>6a5>~)e
zA_|ylNCi_3t6-`T6-+IpfvJTxFtvyVrXJG4)WbTMdPE1)2pM1+VFOGfVt{FeOfb!`
z38on_!L&jam{!;V(~4MN+94ZEJ8Xk#M{F>ikOQU@cEEHZ4w%8N!5;3!D5wsW1bYMZ
z3HEa-=YQ9w?96Ijlv*dv&|yO-xRSd%WoK6N($p&XqyE{gv&UcKzYi$pUrMwh6J<bU
zY8X(N7=tPsWkF?YSWwv*iz*l8K;>#UP`Ma~Dj(%R<!g9Q`52F?5EVcbY6MV)n1HGn
z6+sniL{P<;h^iEoK$U7FP^FlJsvMO;m1|^B<(Q1B5>-G|Y7|hFn1ZSrRY6s2R8ZBJ
zimDdXK-FqAP_>wbsvgxr)oXN6^_Y&T5j8+HY79_~n1QMpH9<9NOi<03iK-Q~K(%Tt
zP_3ASsvWgKwQFop?U;?K6Lmm!Y8+6Vn1gCi4-R&xN_MW52S)*ggHv2`-^H!GmUnNM
zMx0-4ohg3RZvZt?{>xM=GaLghqcY$!jTD0}%dz0HDhn>#NU`X090x9^a^P}}6o)R)
z@!;|*4=&$G@#qSi0Ir}4;0leDfUd}i;EJjUuGmP4=t`UfuB1xfN{y6+uFT2c%Bl>m
z+(^mjDx3nYqAK7jjg*3}%BkS0stT^!NU7**oCdC@YT#;(l!mU(>EP<B4zAuv>F64q
z0j{AM;2MpTfv(A!;F_uluGvVL=vtfwuBBSwT8)&2uFcut+Nuq%-ALK!I-CQpqdMR^
zjg*6KkPQm%<jGDfDX6;Se&T&NujP|2Or?6l{~^5X_|t^{IC6&YUqnuj{x`|i>E%z7
z{zJ$Or2i;#it-=E=~z!N*kjTJgFPlgFxX=fB!fLBNix`D8b}6vOp0Q#$JP@p_SiJR
zVvo%bEcVz0$zqR9k}USv29m`do1$3karFd;JuXdf*yAz;hdnMqa@gaNB!@k&f#k5q
zr6>-2d_BQqk53ak_V^6JV~<aeJofk`$zzXiAbITZDT>FQP)`Wh6Vil$Jt0E~*b@?@
zfIT5e3fL1GNCA67iW0CV))ON3#55sdPs|V^_QV7!VoywxBKE`vQpBE^qD1UT^@M~y
zDNRV&lQM*aJt;v-*prf^ggvQ&l&~kIC<%LVJt1RHP7^Zr<P0HWPfn0B_T(ffV^3}%
zW$ei*O2(d2Pbk<^(u9INB||9KQxc?tJtavh*i#xv1$#=0Qn07i6Ds!9G@)Wo%@8W~
z)C8$wPfe04_S6Pa#h#j?RP1T>goZsWO=#HDGK7XbEkSD7(~_iyJ*|P%u&1Rc4SRY$
zp<_=^6FT<v454FBPmntH^dzZcPj4V~?CB{=$DUD77}zt?gn>OHLm1dI5~P7WBS{+A
zGa5((dq#>fuxHj2CicuUVPenB5GMA_1ZiT=Op+${%m&iLo|&Rd>{<1Mg*_`xSlF{N
zgoQmTL0Z_elB9(_tAVtzXQe0$dv-lxW6w?#HumfcVPnrukT&-0Bxz&MZXj*!*(u7#
zo>Naa*mKf^gFPohIM{O%q=P*tNjlhb8b}9wPKt6`dxIap;Q8gsqN5aDUD2UpsQYPv
zt{&Q9DA4c@!~dX>#*=p!bE}>p3aY!C-OJstvg1b;`OdB@Ym?nQD0k3}?q7002+(;g
zm&jT_S75H-|Ktj)F?ajvKhMfJ(d&3o_Ciw=<zGg-f3cvNX`&cpwuxeqxh9H3=9?%U
zS!kjJWU+}7k)<X|LYA8-8Chwf6lArDQjxVLN<-G0C>_~oq6}oSi87I`Cdxv#n<yLE
zX`-A~p51+1YX7i@=4Vep&>QUHo&fhZk=Jq&(met9(w_KM@z;-<4{IIme;N;K9nB!K
zbu^32)zKU>Uq|!ELLDt2i*>Y!EY;BxvRp^Y$VwfpAggt>imcVq8nRwT>&QkOZ6KR<
zw25rh(H62@N889w9qqL8{Ua*snTQ^y3aWb!={eE8XkAcUGNfdpd$FgWx^zhCMEA<=
zqkl>)d++3_H?wD==cJyuyH9z%aYzsA@TzXv-9_z-irp`-tH0VfntOQC%;qyuOZ)$f
zXQGy7kl9+AMdoU04w<i|d1RrM7LdhST11v=X$e`brDbHLmR69}T3SWcYH1BwucdWl
zqn0+1&05+-wrXh$*{-E+WT%#PT6z2V#~zq;YU7&h@#}L-pP}wLuw(JtZNl@mgtEK4
z-q5w!{UYODW$&2Ra&_H(-DeN7&p*I@ZgvgvF+i-31>$@h5bxuG1fKvT`a~efCjrSm
z8A$OdK&np#(tH|_?$dz`p8;h0Od!i=0ogtq$niNq!Tv$NyCCfG<xY6w!{@)t9=F2w
zg&p1F^^Q7rMw=m9&S!Ub>f4EP?*Y7f`ac%xXMY~i{3`YO`@jM8GC-`C1>(FM5bx!I
z1g`)jdPN|~D*?$~8A$OeK&n>-(!3gw?$v<|uK{FwO(4r_0oh&~$niRV&_Bix{BBz5
zk%QT_pqwth$vu(<)gd|rR7qF<LGGH6y&T+pE2s5kIDFXfD)$eZ{OXb5y29PZvd6r1
zQt5Q}IcLN+SK8j#d1ZF@piYBIUHh94x9)#&$;i>!-6fMs?r_&2_r_3mp>Mt`<d)8R
zcdRAt-OX>;#eIuK*K+R@b;xVEKjhlkvm5XS;p_$&AU41PaRCm95AZ-jKmZa0B9Ii2
zfaHJ-qy!WoHJ}1%0S!nG=s-rm05St6kQK0i?0^mA1e`zUxBbUXkJmF*%`fpd?cah6
zahzt5**MK2b8(tO=HoPvEW~L6S&Y*nvJ|H!WI0aD$V!}6kkvS?B5QG4L)PQ8j%>tf
z1KEtzCbAW$Eo3`R+sICwc3Ro9VCLNF<~xcD7ffF`uKB@j&6}1Vs(QS`PrP*M%*LEM
zrcR6AcDpC)4tYG&S|3hreRzB8!#i3ZPM<ep?(8g%xN{^$x5;Vi&CM@p*Z#&EDhr2{
z_bbe+yrrmfhmKvknwC=*7&s`%cVg^5djHb@Z*Ot;9zA>YzUk(^z5#w#5G6%3O4M$g
TT+yj`=<w-eXw=|g#I64Zz(Q3!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp.64CDFD8CF1246F3A.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp.64CDFD8CF1246F3A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c54d140bd4086504175562414a8fa37bf67a95ef
GIT binary patch
literal 7646
zcmZvg33O9sqQ`UZ4K=id($<!pf-HgDeZLn-SQ}c_45bh(70NO|w4hKZn+%FCAOf<;
zxS_JllU*1W49YIxgeegQgcqE^0E!|61KP2`Fix6xF8#lAUeDn>`Q7iA`ztqJlG|^1
zvDbUPt<BbNV)2~HNmUhgo6VM;{5P{|#?%C5vyIHK+2*W2FtT#kHKSSNy?w)WZ0q)H
zo3fX)d>3PZ=3D(E7koTs@*4khx7No${$=6A6+idg$T->cs|MbD)A{b)>uM)k<3*>W
z%EI)c;j7G7?K}lb?cA2jb+_*S>*|B$V{5*2&R=$R^4$gdem%TkZdGWDVO}cfssA~2
z@RemZ{~A616<<|%Z^DV!Kh<Y0JUC|kisB>pE*xLOAA0Zdsp->p*WJsTy0FoLcZT*D
zcO<>61HHfKzeYVh@|z~zx#g$JcK2Mncl)m9^osrcC;L7s-CT9@wbh>$-)Zve<%K0X
z*5yw7{G(@QzIOP)>9-EvxRE|&L+)E6CT%#L{>iJGW}ZKOS^ccC>ucf7iDqr`e)4?u
zLC&2ul55d`cg{rp{?&gPY@63~#i;XDfqOU3eqXZm*u_EfTWuBdx^#{9JHP6$&ez$O
z-+1xz*7g}iV(yCDr~c&c@!Av3)^9p@^yS{aF37qVTx0v>qjgPQy`Ff%NGt6SOB_)p
z|M!nA-L!d%>l1!)R_RaDKWEJ8Bk4^X6He{@aK)0y?WSi=wjXiyckbCI|8RP=KVwV(
z?VUER>3`*Jmp^WHj_CGNw>zu4o$GdN481Y0{O@0MJ2vi%y(gZU-c=e{+{r%d+}@Q_
zZR=VOY8<cnr08rcI{z;({cZ8ljRQXIGRW0=__8S>!m&brd~$l;x+PnWFI)L#+aEJ-
zZvFSOGcu3#ZS(cWK?}!z8y{6zIe6-=6NkM<?&h0QvfgOk^62)kRCRjJ>T5RT-K8gH
z=XZ%qo{2w4bBP^KmV8-tvdOZX^Bsl`8a%vUR87m`cE3$KT@>iB<4O8?$GML`I9_(Q
z$}#l2q2~&fGcD5Je;E6kuAH=_Z~mFwH<y;oxhxJ`)T&n*_kR4><p(EhyZ=t{$g4Y#
z3@jU+qwHC-ZdK=5&ednP55ID&J@Iz$w&{2KcG)&#?{hD`fAXC_pZGjvAG)&Tsiom)
z-JB|6)4GARJzkmhfPP`gx0Mf``5*RDasK2>-KRZ$>-sAC+TsoN5BE>=uPi<O{I1fj
zfB5FY+xz!CeqrDf(@V3q-Tu1gj$v~(&(z<xuX#N<D!4nVAijKJxpVnvD@~fNn^~oF
zA4gB}Zwz!>dt+VK=Z5yX`kwj2ie9ZY9y_$;@L#%=>{weDSbx@jruChS(fe+F-|5}m
z9~@$*&bc=J_OU)s{;kF3yofe;_1`ZJx!bYBqC+RXJG8B?{~y<e7oK#kja-~4_0-?&
zbm+>s$?nZPhqjo1?ekUPHJ;5c{#d)P--AAT<L4TwRe2|8zgt#E?RULZdtqMDjIpP4
z&eah;=M8>kSLM<9|B65CnDemd_6u9?7yo>)|8H4Z!KWkAHpQoP^K+dKwBPm57ULEa
zeRCo>yR_H)-@p1r#(m*vr(J=oZ=X&d`0qwze#u`pb8hAMD#B(<D+sOZw^47OL5_OT
zR*{#M*ElsKKN(U#>GwBWT0BTC$sk82r(37B&Q1;1boTC-Th035$1gHSUvfGpEvIp6
zu%>%GxD(%ceeUiI(x04enbz{L)L=~yJhc3~9@9(PH?fr@r`zVY?T{L*>5;*->Ai}X
zof%|Va=K$$$L^`YnjSr-$KpASww%edjZRMYY}B)NYOtm^c@sHrJU#1&CbkjD>4N-%
zPN~6~-m|9hr)jNsznMV}Pflm2Wjj)XHT~^dorV;%?{#isD@{(f%Wu~)HCWR>R<!?&
zRpN36IV?GypO)V^HCWRRn~nT1HvQQ34Dxr$>D;v3#;L)Yo-yh7GiNn#xb{kV>YA@O
z`e4J(D?Qq0R=B*bV(WjOB%iAU$nPp^AlgfVN&9FpX+I4n<7L2Pd<>Y3p8=EgvS6}4
z7EIR9g2{O~FgYIwCg<nC<h?wYypIQy_w!&1UI9$OCx9vV1u#Xg2&U*0!4&->n37il
zQ}Rh*N`47U*(-x7`(!X>zYM10Rlrny3Ydyt0aNv=V5&Y9Ox3S~sd+UpHJ=8i=GVZ~
zy*ik>PX|-?>tGsQ15CqbfNA&*Fio!srs*@mH2o%+o|c|AkKTS2o_3xNKpj0@EV4d5
zk_){4qXQOrxob+pB6bB_vw*5yFA@z%m(!rq0UA`gnnsl=XFz2F45&;sgDP9jg31P1
zP}yo0Rj!-^l?!m7a@8EFd^ry)AK*ddt9ev~asgB!Ab=`V3#f|aBB)|O1XZjSQI*Oi
zP^Ewbs#GnZDwoTk$^jWvxmre5DOW&M0t%=~wSuZzu7autR8ZAw6;-WV162!Xpla0`
zs(QH&svgim)vI+>jdBB2BVd4PR2!(8<tC_Rzy#H-Hc|Cxo<7#1&aCkC^9%*@dCDxZ
zz7OT})b}7Pc-AE}tZA3x(t#3`>u0KgX_f|;R%meP1Vy9Eunf42!hp*pC<a}YWx-_?
z7F;$#vFLIv2QH^@;BpCyLzicHaCwCXmrqbUx&kYJD<}fELV^;|6<HBnQ4zrv6O@Rq
z#7f{wiUh8dpd@r<Rt8sAWN_sKC8Mjb3b=}*fU6`Z1znX@!BrI%Ts1+d=xVG6uBK?<
zY6(h1S7&u_bwvkPPf$9#25W$8C<eGjf-=xGSrc4SF~Kzxl!>l~@$l9nPrd;}k77~l
z2OfIfjGOt*k~b<ryKAW(7IM%|BQrregUkl)EHW3gbI5$q&Lay!yMQbP?IN-iv`ffx
z&@Lk@LA!#i2JI@c7PM=~deE*T8$r8)YzFOS16wz{&!wksa^bySL%2@CH<~L(xGo{b
z3D<4p5b4@SHe5zmm~@>)j*_m6$Pvo*6@Gry5H$Aa7(ru?juSNY=nzR`j}DVG_UI@{
zV~>tdH1?Ppg25gWBN*&4ae~1f6CxSxF=3Ly9up-Q>@g9F!5&*fu-Ic`1dBa3PO#Wx
zLnMnmHcYbEW1}RCJvKtI*yCyl4trdT;IPNV2@ZQ)h~%)xg-H&3T$JRn$3-X(dwdPS
zV~>v!JoflF!DEjPkv#VJFv(+&kCHt0_z1;gPpBaT><KYKz@88%1ndbRQox=NCI##X
zQBuI35TOL@i8X|XJuyaz*c0P~h&?exir5pwq=-E+N{ZMMBb10esfLiSC&dT}ds3W`
zuqTB`342nQl&~j7NeO#Wgp#l)*AO!H<QO4iPmU8Z_T&&LV^0o~GWO&sDPvELP%`$E
z8bZOI5+fAsDRDx<o)RJz>?vVV!JZN&73?VyO2M95L#WtOV}y!5HBPA5Q$wVRJvB_K
z*i)mViaj+#so2wM2n~B$jL@*B#R(02T8PxJr-exkds>v#u%|^R4SRYGp<_>v5jytt
zIH6-t50N_d^f0MoPmhv1_Vfs)W6!7|4D1;(!oZ#pCk*TvA=1E}5he}n8Bx-}o)Mu8
z?3p!$i9It$nAkJpgo!;fM4H$$!=#BlGfJA+Gb5DQ(Cf40mhIjzRTn9HMc>@M1FYKy
zdqrVxAy83n(Ic|<g~;^Oeo(_VY2Tv0CDsXs=xe6uv&&ZznH8;#*8QwMHAf81asIHn
zXQSj{cVl;!Wn15-_Vm;q5}Xo#r_BHJJ7rej?_`~bNKfrDStl5Gb?M}^%k3}K-~H55
zu9xucr<S6TnOcfLW@{-HnX9EZWWJW-k%d}HKo)B$5m~CGBxJdkl982KN<mg@DHU0(
zr8H!{meP@pTFO8+YbmpVr?f7K94J~iEqUEN4o?f~a?KvpzlQc}J}R@KZDHF{)+?J?
z(XO!FC~K1~v!bA|V3f6C_rmj`myZ17%D&{8+Kz5J&U(s{L}45AtL3edhjW_c<XeAf
zS<gzYmHItC>vc)Kctu4ct>0qD&Bb#^Us=C0`P`0uI}WfG!_GOMH}X|(@+J?rY}zv4
zx`DH{26NI=TY?RzbU9t!tmkG{&`uhNaWX)xlLg|O91!p1fdr=jBsxVP$teNJP8mpX
zDnP1J1=5@vknYrh45tBPI!z!?SC7lOWFI>lm##Xs@ka96WH-xx+`1+QmEjW`6>hqj
zJe=1#kF|D>4_}=(Q~RNAQ2pD$;pzw%$w31#4hD#Iut1!H1L7S#kl+x2M283@IV2$2
zAp<E61xR(MK$=4X(j7XG;V^(qhxv$(j_}+VTX5ma<a+L7@AHIpwK6OE+xr6zun%}d
z|G2iO_v-9D-y|<t!RUes)`GPbe{$j0Z!b5jvcjUmQtQ0m35|YvEUWn|$>+NB+y&O*
z>!-FI`0oWnh9wWT8{KZabt-FDI{6t@za?#*@}^_wt7P{*^`Dgao%03DvNox+(^H$%
zmYsYB-LBs73c6_^#?1h+ZWf4hb3nYC2NK)@kmweHB)0@4yJaB7tpKTR6-aYyK)PE8
zGTa7`={6scZN{W$D(ZLQXU>=~v!wpP`1;|0f6A=~HrrDT563n<9MkY{+@pt7#rmg*
zcL*Lj-79VUn6c$gjk8UtocPR?<ouxD525Uh(jId(&d6-iEURg8c45CR+3Cf7bDHNq
z-lC;p_UP&M^y<y!(cL?0T?_udqxr4dv~AzviB6rJ-CPVWh_b4;Q(Cc7|GfMG11FH)
Kp?!*o!T$lXp4X`W

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp.38A1F46E05CA49C4.idx b/.cache/clangd/index/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp.38A1F46E05CA49C4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5554e14562140ef4f0ab0add3819548242ff4fb
GIT binary patch
literal 6148
zcmZvgdwffGAIEb}G}4=1XsBwnx}N)8v{6l|5|lO+l_82@m8p)F*2B!oVy0y)Gker1
zoBMS&Vq4wztQL#3({7d$#*|TU3B$(bi9D~P-#?zdUj6mFf1giZ@1AphJtybn4l@{@
zbyX-brwuDBoL(}egF>N*@O>AT%$<2jsZg-N3Po9UU0$K_H(j`M(^2E@o#dPP!;w8V
zH@a_Wc9`-ORZLu1Wtwxl+H>gP;=d~&46I>r^W$aB+a9Go>V0<&rU+beRw_K&;m)8r
zpT2I~w7Lu(SKGg-?ZV0*iq79`U6khPTwT}F@bSWq*MC^^oBrgr663vpPo48LJmAQ<
z^B?y=)_C>w<F=9oZ>Jdfbt@l;f#)8!P5p#BZccODy;F2<WlYtxkSD1d=A9^yIy^A1
zj5|8+jl2t?U#*SVcVyB6Y}VKb4dc2uMQ+SEn^$?RicBm?yuak>^2^PZR@;R;-R|T>
zmR%d)v?ga<0lK$w=tu8dr<>DfZ>;71+4o~%X3dDW6ywQJ-niavR2)xkMa!C__g20e
z<yjjz;6`KF^lV(SDgCYVr3I#iGmU7Utl8y}`e~s%7tL{<(<D8q4BDbt5nWVbOguQB
zD2q33X}oxUz~e<>E!J|ym$e^-Hk3RWl&l(?;J$aVPcPGrj_-YwSVVn!ykpN@lcw)!
zy;uMJMeT{oJ{Qg;$E9A1-SSb!4%^m6x0k)Wwf4aEoH;u?)RZ5-*qS|JNYTj8zuA5w
zbKA(_?;XsWNPlHpzGme~Yhh*G#SdQZQdJQCP56aWD>7y_6L4tFPZQ6$TNZwAJGb;i
z&5$D)mSiUe&#)ot%HG$uFMI!7=C<!DR-KN$72HNV7y7<CIpMH{PO2-q`l$A^<Zasd
zuTB1T<BCi3W`wQ35q)ZxT`FlT`}8-3_S2BcR!8OSUg?7#{Ip`p*u6Ware0Ak2y6@g
zI!U$jF<oCVKkm7uQB5W69>4BV05$&n)xGx*sdHLaoge&JK}g@Gf9`mbF4Sdm@9$Va
zOzoOJUz0T8-A!!Q+D8kr8`keyJR@@W?LMZNbXK{lZOV6phy66He&*PnZ!TWDBztqG
zfSCzdXQ~T()Z`v5{<*B{@)OtAj#|CGy3#mgNMiS$YmYxzf24n8SyrbFuYWu1e$Ihk
z9{l|I;re}Hqo*(Wwab8Y=R<!uxGOjRji@uv_q4^cBd(4Q95La?HO+Opv$txsF~*jn
zvD>zd*d1#h+C1;nj%({{x0GKVf^SVcoNJBRrRnvF`(EUY=7fKu7B)N^F>q9W)6?&<
z-n-X4$*mrJxAE(H!J{Vr5&Okwac>dTh2CCAk8HTMaA0WlvF&a7WBwepw&~_#VSCdm
zOKPk$d%Hd3c67kZGrzo3)ECS8t0E|?et&0AKbI!!M)OFXXvrzO64tnD$I8oe_=kU$
z*Keo~JFxY%_HqmM?EPbP$GeTX8~?B4`T1k+OnCVG!1bl|-%Xz0uTRjr;EcQa-Fj8G
zo=rUt)@(e{aPUvhldALzA*1v9itGa0E6)}9esyv2g2JiEh(e+2X<L<9lM)|{jx{Or
zqf}9W{=s)0@P7sUwfWN0;eAF2qhow=j4CF=Kge;!{uSNl?Ys4DFgnr~N2($N{ev8*
zvy-<(X;-&~D8~8X<bY(oe~{z;Pw#tnTwkz17|rv=(W>Z<{y~lpR0SSU#q9qi7|r#?
z5vmBae~{yY&(l(7PagDksA9A)j@85__y;+zZET7PoNf3!7=6_j$E)H~{ev9eniBsN
ztM&8_MvcB$qtXQW2RZ&bJnsv4(SNQ7qoaIrXI1Aw{~*V6r_U(9?YmgYM=v(Wf7B{Z
zJ>9(LO27D!e9V9ilm7xAnXn1%gd1>>;U*kpgn<AVVIn|A8c2|lCK6<nfdUz2qCiF)
zXpqq+8f1)t0U2XrK*kzakg+BfWSoHm8E4`^#v6E$@g^Q*f<XY8U=lzk8bpwZCJ|(k
zK?0d%l0c@5*7bO~((-kQI!q>c_0mE9RSQm!87pW%A(#c52dES)N7@lL<DiOLa8Sid
zaabkH1gH`g0#u1o0#->g396)p1XZ$>gjLE+fhuL8K$R+`V3ju0ph{b4P^C+0SY^x%
zs4^A?RGCr+R#`I(s;q?tRkoCcRnE+TDrez9l`G|7l{fRC%3F9)<x6>370d#t3KjuW
zg;D`lMY9O1qD2H%u~dXr$t;1YWRXBsDwSYWm#oW>*HlQpE>o8SXryj}O!DQB3i4kD
z!Fppcp?yhVS}X;SSBd?hY-gOsK^NEJpo@EzIP4NE0lI{i0A0eXBw&|hNzf&=B<PY}
zB?-F}OMxz>r9hYRDk<2dSsHX{Ee*P~S4qPz!!n@DXc^FDyh;XkS(XJ|R?C7e>s7L_
z%ds5ja#{{_Ij@p~U7qDZm)G*3%X^hP><X*^x`I{!UBRmqU{_>C&=s{J=!#yY2)hz1
zfv%*LKv(iACD_$5I$mC2zQrf%v@*#*ks$wVRbFoLZFORh|8^=*`))%H1Um~ZZI~Ou
zE<yGn*d53=6g!HxuT0F2VrL+`P^<;AQ;9XemFz%p=)>Iz4t=->!J!Ygp*Zy6b`*y`
z+=b%MhdY%x^brn(fIh;F5YR_>5CZxL8%jVQVMht*BU~r}eS}j<Kp*KqNa!Qo2nl_p
z2O*)4w4o&Qk#>}XKGKDf&__CzB=k`Zgn~ZGjZn}>c@PTvC>u&aA7w`==%ZXH1$~rL
zNkJd&KxpWr-3SeRv<IP~kG7#S^wD;dhCbSb($Ggcl{EA*4upX|#*Hx0$9ND1`WPF^
zKp$gA8R%nNC<A?rQ^`Od>p)oOW8DY~eXIvzp^vqpEcCH<l!ZRlg|g7cI+ZN+aSnuo
zKF*DB(8qZY4*ED7%0VAzM>*)@Tqp;9oKwj`AMZeT=;Pf84}H7`;h~SWp*-~Qc9e%c
z-i7ke$2*lg^a&0`fIh*E2+${Z5CQrG8!AAbU`GY$6I`eOeS%XdK%eM9MCcRUhzNb6
z2N9u9w4oyOiFQ<kKGB7W&?h>TBJ@cPM1nrajY!ZZc@PQuBpWJ0pJYcR=#yNi1bvcI
zDYg4Dmfu$V`LLt~>5!k6m^Mzn*k`9y6=~Rt50H@j7+p-J{1JCdP9*;OhiL)6>6E~f
zFu7e2Gul}<bn$H8v`($-CQr{FKFC3vl1KWcV+Y2LksqbrdCRu!<gZtbzL?I7HOte>
z?+xzJx1plDZ#ptOQX_xL+1<8u!K+uQSNW#n2gVPP?-P5<z77~!xW(X`j_wq#k?%(G
z(_C1P|0xdKm2ixde>Wr_$8kUeP5>ft5)g$`fM}cs#NZ4d7H0u*I0uNwc|Za#01|N#
zkc3NsbcBwQkEVQ_ItBTkK*6qA5$r>7L&dEK4w<kb1Z2{RkdP@WLP4gj2o0IBA`E2K
zim;G5E5bqMtq2cUup$Cv(Ta$WB`YGevwV`!Apete`)<PJc?b_2`@z|oo4&Ij5gyS+
zKI_M{xzhp$Z)x*QM<qqE@^kZvtN$ua{{6}DP~V3WHI^tFe9J(M10vJ}AW}^NqSO>1
zT1^9D)C?e2%>v@o93WoJ0}|8%AW<y>lGM^mIyEN!W`57}r+s_XuS36H@_J|$0v>h@
z>$=i6of4JOQ=XPzIr#nqUc7H;KYo}I%LTh`!~qdT0uX5=0Z~Q@5N)IZF-8UuYh(d&
zMh*~f<N*mr0gz}E0ZB#)kS<P_An#h6de3S!^}vgZk|s$b$a^ZkMe+S(ym*BI_E?V%
z0DG**0TFrv5UD2tQF;mxt)~GodIk`yX900~4iK;B0SS5mkf;{{NqXre&7D4L%8U0u
z#dBXP9{plt;)~&bUlZgFg~I%DQ<>jVCCydsrst_9k2g=4I7u=7Wu}<azB##l^R>ch
zvkH7Td`7miLx8HIIxsjSG(4=+u!zB#SVYjUfstK1cj*?b>)S6)pWc4}7e%KgCKEmX
q->fF4M{ImTuUC?Al45v4?A<3NPTQT821RLxWKBU0IT=Hd5&s3?`)kYq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6DFDC664FAC2C818.idx b/.cache/clangd/index/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp.6DFDC664FAC2C818.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f55842224b183ad0c751dfae63544e3152aa19b5
GIT binary patch
literal 7952
zcmZvhdw3INqQ>)0XiaZGftG?TP%2~+GMQvD!%Su-5z}yK!<|UWB~U;j5K3JrY+wOd
zFc07gN(-{+a+Bq{SU_nJP+Jhxa}*66jw>J>Zn`csi$IsFVK-)<vF{(J&%@XJ=KbYm
z=9|nV(~%_u1{k{Qbj2@}ge#|3Pimvn>9Q05=2Xv~{*zv(+nu4)h5vMNbmgcEQu}+G
z4-eb8%J9<ox4XFem5$w!!mVxFU3m3igRrL8=4Xu87sucI+0%5e`)cl^4R1`le_-C`
zMM9T;sh973CvEK)9<wR;;U~-WYmR;Pu=C>IESmS^pG!BhJB&pqAHR5KvU5Sj0aea7
z4P8C()xY(&?z>fU_0hGH5AP4Fe()r=l&z}D{dwm6`Esw6zinpMFMRmV=2^d)^ShH@
z*A2Q`bouVWOJx1t>YVpDL$9Tm_MfjjJ!ewnTuQ+otDlc8E$?zY)RLPuPB>5Rc$Z&w
zyzIM|<3qNUO?Yte((zLxQtuznd$eO`pMz_LPFwfYu@yOQz2UvJaaiNiEl0~|jC<xv
z;R5reMgLgX>rAi8xvwngz4QIK`+JsLFtp^K-?A}w_Rhtl?QRujhuBs8@NFIX&Nx^6
zRm~U1^#v<;^_V$#Z7=%E{!<n>(nGJ0^e)XQE&6apHQTZ!M5hh^UD@FW)7o}lG39K|
znN{r<6>YmaX$dXAy5r20E}p#$*6&__Eu$t<*{}S?->wT$)${JJIGww-!=@Vt%APE{
zbZkJFS^MTIr5z&dTmIFXhHo$4a-z8IG(G6>H^-+Pk&c^A{CKtQ(+O23uC%0Qn_*|`
zhECT7$_J*!s_Tc(c)2Z^z4g?VpYozrpVy4`mUhf~R(Ig`n^*nAN6cR_cF&L1=eKUH
z-gKj#@w4^?e;!d;fBk;{ON;$yKdpQB!q{)qt4389#7IWZf#t&d)!p9ek)FS#cl*<4
zzdF9hdabzcW>@xsJy~-H@BHqa@=tp2m|jwH<65b+$aMMtvc{G$*Q-O@+x`BnZw`fh
zo7$efdHfGY2KVmjJ9e<~@q*o>@6j)8P}i*FBliooW$wCt_xtZZjDNi4mVeLU7i-q>
zq5a&@y0os}ep33mFxR-oapcVHmLB@`&lULYczW-iU3dExb8Og{jtj?9#?M|0p2(fK
zKz05xTDtww#vMoE>y8!QJF=*6^P87@`5d>@-wN~Qd|sb5VmLKsgt)h9&(*Dc<+^{p
zzprfmq;zp?&#LF!{ddcZS^mpwbJwofnSXcTf;$h&-aQ<+JEN+)=S_La*`;49V~%aE
z-S%H(-1IL>1HKKX_MPb3nYUlraOl$iY|EZ~KRC0u^=+kg`}!Z225)ve_xRh_hshHI
zFONAhc;}*vE0-Po^WQ)2^HJzk?&#bL&Ep2>Ud|X@SbOu_%<=2b?S0?%cGlOYg6A%L
z@P+A$<I@QbQ>V@zvbm%r5}#XqZP$g~&F|zd$ewyMW$Xhn{J#ea7k@LeW_itM<GK8)
zKZ`%zuoum%TiJNEs^!+;JIb!bA3wdR<&m$2`S#~wf0<r!an`PyoRMdgziwEPy<_>o
z_s7pX^sgy(Pb76`yQ0Bo?iufNVcu(=_~~M)a?bq9$<Gs=F17mytBbdaJu;}#V{{dr
zQahz3hs2p7`I9z!=(MTnBS&WH$`i3C#WOTHXmR?ZEti)0gi#sP^NF}?YS--KpvBqy
zR(73RcjJ=`YIq{fPR&kB4qDuCbn%6qb#I@_pvETR!qh@va?oO9pMjk>W)FTTOXp9-
z1-S)9$w7;seSAB%>xcRKGN`gdoR^ySRC3Vb*>~?*$Yg6xmTpuc?v~r_ndG3wA6_Yc
zDeZLMIa#`qi8wzuzc4vy@r_A6jxc=8kwFbh#5t)sX~{v0<LyVkcd_ck4;j>mMBE{@
zLt1jsV%_Yi!8s@MT4$F{PtNM9<BzxOy`uEUtT6hGW!mN^C}^DCO0=Jbj1JO}nfxZm
zOhFT5X1^ISbI=T##czSk60|_Z_!-EUAOo4zZ-vYnv_i)ES;*KR3z^MtgUlARL1y>c
zA+rbVka2zvGA_tL=I}cpa|9ib@qQjMKFC8R_yx#>pa7ZZ7a<daB4m<Zf=miZkU9NM
z$eckZWU^m|Ob*JBDSibqC8$8A`c=r(pbDAG?}E$~bV26!yCHK2-H>^^c=LZfhAX^9
z-rgE%ze#%Xh}J%u+7Ue<J$WQ+ixZ#Gka1q?XVe%rE+c>uVtfOb4$&afVH#yRLTHpt
zArr`^unA>TgqTn^hs+?G!)BDt5n@K!60(4730qLMM2H1tCd7cugc+2X2w_mRhO8i4
z!&a275n@G|4Y44zVHRaJLRge-AsfiHunlEfgxFBFhwLER!*-PI5n@M~3vnQGVGd<3
zLO7HiAqU8gumfdBgg8*<Lp;cQm`9n95FTYAB!Dc01(byd5l|LGBFJJ`L|Kdw5oIYP
zfh>h3l%)ufP<Do#AUnfOl${acL|G2WAj@GHWjR7*l$DSIvJzHMRw6_}Sq-Tmt6>#o
zH9}OBT_G39uCNPbSA@7wc8A;`yTfjj-4WtO*(-WI+8H6U!t3=80~+BSuaR~|Pn;fp
zxtzo4fj5e9K2VHqAdj(sE73d+Caut5(jFR26K?|3q?o`oc}!rMc{7-1#SEs|V+Pa0
zTfnp^7BDRy3z!VgfXOHfn2d)3)5=@Hv?^9GtsX0wEYE_;DlC|+hXvEd+rYFbHZW}-
z8<=+94yIkPgK78J!Q^-jOitmz<UAah4&DK#LveuV@HoKac^*t&;lboRJeUG6fGH>f
zn1V+DQ{+W3MMVTt^oU?eyacACNMK4H2}~#N1k<TF!E}0@V9LA<rmV<d${ra^g;&5-
z6a`GhqkyUMDwwLGf~k5`FkQS0Oqb#U)8%o2>E_*Fx)nE=ZjT#GFYWE49l(i$oAKI!
zIIp0Qc8pT#$@}ougL_cNLFJmH-IZsgC-25vB}*-_Y=L_Krj{_s)>>jkW@`zHY^x<U
zWP2^KBXhNcLw3{>2Qptvcx0iL2(7GrTc}D;eoJUwZDWiW{{gFQj1%J%<R;2^jMAni
zaUT?=jHi(g>W%gK)|_r8bPV24#|e#nbR$JK!TaebMPnabucsTJ&lDpjTt8Esn6S^(
zM452?Oi{{&eWrT73H!`3V#YpmoS3oC+(enN&m5)9*k`WSo3YOlBNptl#EAv_EKQUJ
z`z%q)f_;{Hy#@Q27-6uFi4z9<m?nzBJ|;>r*vB;JnL{vt))=v3pEXXb*k^5`tk`Fb
zQdaD<Ht4O`$2Jod_b(eKEcUUD6pQ<pjZ!T3u?>0_`)tj`hJCg;v0<OBk+NZ*ElSz2
z&(@%~VV}L3*s;$ZCwA<!H&S-&vqvdA_Sp~V?H|DWam|FoJ}yo;?Bg0K4*R$$#bF=U
zpyzP?95Ld+{o#lc2lhFdC<pEjN0f44pQAzVz&<`kc<kfjgvUO<iQ=)3k5WAL@eO(&
z`-B(~uuq5+0sDj|O29rLN(tB}H0TBF6JtcgJ~2*2>=T<P5&Og_C1RgguNU$7keZ2v
z`$LKo3Hzi*O2YjiMJWmUq<X!Cea;wh;`%w`#EE^*Cd!HH=ZsQL>~l8gow&VnjL6t0
z$BB%6auX$EpB$xR?2{YxGWIDkqF|pACkpl{O_YLtN|aKtPifFA*rzrV73WWl6BYZ^
zMoPu`Q=^oMeQLd4#rbnJ6BqWm;>3l0u13m*eXc0w!ai4n-i7PujuAI*uRBiM*ynDd
z+_=5&DCNdJcZ1&D+V|Y@X5GV<>hq*c1zyC5h&5HY%xiCu%!;m3*JABUzkg^4<Bx09
zl*HoVwBmN!CBJOzzN=LotE{<+tDnJWhijdI1~M5;KxTs($YQVnF$M<6YOn&a1{TO>
zumRZ(b|B8c0XYl~Al|?O2?ha3G>AZwK>~6boItWc22u<PkZMqYTm~1A+u;6{rae9C
ztB(dPo|TwWufdzA&8hZyGA%v%ShDqgM64lV>upPBg;YZ%WM>U=BFi;IMpkNwf~?jM
z71>ooT*&Sk;%?=6KJS-(%KHx2XIA9*&o9%Sl4e#E^e-sWZv8SVy7li?rftFR7k%)`
zvCCHuCpIvDOn!y-mgPV7&v$<PTGzzl?DpBY+84*a+*~$)%#}Z_PAo1QRamZlmiC5^
zrVOuqXFy_cUfaA}?QTsgOG!^Yoo$^Z<2qt&UF*yWdL5yWP3wpW*}RUJTlr&t<O?bN
z-?^Fi-#Yc`#AwgJkNe$!pV<AcpCuL#Y%{Q@wyQ2*Theb$_Wlcr#ofkqo1$&^_qLRl
zr`mP?Lt=4pPH}h5{^8W_kG3xyGAyyUU`)YeZO*g@%6aL@2g$G{0b_+W3yJLr&_Jeu
z3CJ8U16cwVASS>7Sp!xeHoyYe0yZFfzz)O(I3P#B0mKJ*AR!<Ci2)Hv3P?cCfD=d#
z$UsU!0a61hkSpK<atGW%-frGPZN6_B_P$QZ`+qr=<@Cz2Y4fZ-q3@iYd@>L7?K2LA
z`S#I3CZ7q&>@x#dd=?<a#{gM<Rv^~L0@-{vAiK{F#Q8WNhtC1T`*<M1Cjf~)5lHe$
zKu(_%NcPD<icbMjeJYU4=K^y3+`m%duNgBZ{qm=cIkTtC8I}00uh=cqv*^yv*Hb5r
zn=n2wQ8)gV_&+}rv==(vxYn0})|V4oUnX1sgHtMBm^m%+(ia?}Z<CVxlp!r6Gpl{O
zwk6s9i+g9Mmv}mK?${}>i{w;Pm;2d%)|@_aQD3_I|F<;v>HLBog*~6?Wi**Btj+Ea
X#BO|7hV$m+4lJEa4H){|ATs#hvG-l8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_streamk.hpp.424C1309002226DA.idx b/.cache/clangd/index/device_gemm_xdl_streamk.hpp.424C1309002226DA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e56552be455664a6465e4db8d56c7fe812b7397e
GIT binary patch
literal 5558
zcmbtXcT|(f7S9X_A$)|;B!mP$AXq?}KtRC7N)ZdIYXwnM)Te8CfXj+3Hl$hrjVQ&+
zS}2MIg%wb-ixhD|5Z0A$0UIdF0=}VKzntSaZ~u@JX6E<ZJ9p-mxp!u!yNk;iEf#Cy
zGWTEC_yzjPvsf%O@*fcB|DOwTES3iWzb5C;S+g$5Nweu><Bdb>{$iKSwV5_$@9F3#
z%PMBB`6t{dKsST`bM=YUc5PRk71Wo=xtHyq9i8Z3I;OMB=ggk^yv^-J$`#5H)bE!3
zX|WSjyIyqOdsb|*S=rlbT-K1=L)WiLEv(vBLOti^XL(EO0*;5P6<C_^>pYC^xpGD>
zd#M=rHfsC0=&TI!LY<L8@p(so^tCHeerH0t54{nPs_=MUUB$NU42Sn2%E^CiAEBhh
zpZBa*VTRe;UsfJpA$d9bVf&oyb$hRH{#aR&cGSb`gV_l|PxPTmGw<sA3az>m1yaQj
z-iW<>Xgzf_Lhq5c2`?ztY_{R?8ef~3$~zzOeahp?9vX+JTSZQ3c~fpCU+FkxreN8o
zyJo%G*P^48+)Y2e2>pnSqQhQajh~%(VQga5Pb$;?>e(34YQF!teeb6X`NY|ymKL4e
zv@6I*JEzrn&7RIx31^n(T`zmu+tZ+u*VB6DokFi(@#}XA<0BsM-{dVxsoeR_@SeW@
z$Ok{jDW(P%Za89iC_A!!?XkK!OJ8N|nrmx%{nVwfl?A=$&4bsU%UnGp{o%*FoVcXb
zJ=Zt>{qfb_@il3?s&b5$^DpXT)>b}v9KF*kZo0K3KzPkj^l4eyi~L`5^v;P61w~w*
zUK6>gEhD(HEjjZ@-6@qzO5IyyK587)4mnv^@Z_Ua(57bOv(&ZooTG(*!nHuRfK9=h
zmO4M&eYCCp#%=SD!Y2l*wRc`yUA*)rh2lE+B$dB+XkXGREz-#j{-;K9g?Rr)UxgC+
zL+o%}w^_d>44rcIdPik0nmuoSsJQz4uU+?VuDY%kUMNh~9baEKnIp9pIm{|O*i~Y)
z{zSEg*VgqX{Nn;j!nWm@+ejDjjNcvjJ-MuW)9bjhwpTtWIt#{6v(bO8b59aB&A;J`
zO32yL)5)b9n@8=uu(qszgUQ`gJzVq2hkx6z!S`3z>%>z1=*o!7m#z~G0vFo(j?9Ug
zWEJ0duc=irlI<5+;AXfmWB4s`*K~vOH0qSc=&9{|@4WIUr;mB7P5Cp%OS&s(areT)
z&*vTu4|BCg`GK^&#+PH0hPY}r=}h1K)8d}sC9(5(K6yu$KVO;OF;dHT+=mmw%S-Y%
zy^e2rHMT0_#>}&e#6?YB`4cMAgo;x;Y`@GeJnJb<O<thle5XSz$!F=s?#z!CugtQW
z_HX9=o$2W3t(RL}yyMT*eSX0!eypvUe_Lv*9u)P?>lR;n<Y{n$DSy^l^IMX~n|rf+
za|A{i4h@S^V_#8U<a&24NMQwRT(cw>vsfs1_qdWM2PqHpv{`f(Hq64dgB^=a&N%r~
zdE7Yqq;38Y9#$YMl}Nb~=?-?lqT#q9A!)}EId!Z`Shiu_HY^<MfW^cmu_kSZvx*jG
z6BZYNi?F`|vibCrEoQTn40u?bkZZ9)Ew&!)fMuLYL!jl++w0UZm$0-UK^w9i?0{w3
zp|GvBwgndISdFk`VqPZJ9qfQ*>8r$8gGMC>9@gNq=n$0<_5KFPD|T>7ck6LGcvy=_
zXhcelNI2L5%Yla(sv8=*0(n@INGL_zQlvK60n6c=`YHKdBRv!_MWh5P2dnotKtA72
zYuU>iyD4D#p)A@-(Mh+zVJTrm?2Tx2IY2vWNX&w<%2<xf58?!=_ct*Af#2Wr7Ph*N
zQ^tyf?7{IE+TVO7w>j0N?$B@K;UR=vk9qZ2G}r-4Nz3dxn_YTes$&(x(vEoT$Y`(w
zmU799lo!utnrmSt!s04$74|nkt}t7n%1?jMs)dya*<Ii+>~DZvHCb@0^9tRmg*k-m
zCU6t>H$ZN;6#h52>T@ZtfBhb$)I-cX&;iRQ%SiVFnuogw%pA%MRipbFAhZ1amIv5#
ziM`B<#^5>m=9r?T$WV$bVIoDJNBr|>7+?r>#2E~5a04gfZLw2gZ?ghL@5Z9t*cif@
zqEBJ*DLjh7r?L37EUv_oO8f&zrs&65@>mvEVM!H(RV*VDS4pjE<SE)k-^Cb?1Vy{*
zyBdSNDcVin%^2)Q(eC>0#$aqXibREha|kMM!_#`#&U=%HC^`q}=AiMQv?~A436_tZ
zu14~7CYqax<^h2he!7~f5nm=|i^uGEtj;KT4YRLdbx;bvlQjnpHmu2ie2(3H^73eT
zc{%~vC!jGPx(|;}Kob~Tj<m}W9}bQ@{RJ6)L1G93eeTvVQS1kGWb@%jHyn+Ci3Wf8
zuWa4o6izlDj)sOK9UvGwU(9`PU&i_Fr43iv*RK<rEg@Fei@AHT187Up*D?1xb^t4Y
zuZR`;au?XZY+1LX2>(_nCr`&C_ju$1JCvu>(DXFq$>3@<xf;2G2O-}Pc>~D~B-)I<
z$*J17K$94w63JI04$QDA9eUo*YJEEi9#65S6XXE0aip^sS;0aSosRU<k%Yldkk%7q
z24ZaI+iF|8&g>@gw;<6LG#nW1ew(_M>tzxmN6~u_`8fcAqGPcz7E2h@6kuTimVjv>
zFB4S;k}{~`aka%cqWF+!|H<S@kpUy|9MU|8i~&R1CK3j6Hb|I3uj(-NPMSzI8#F4&
z2GZgzrOEkg{CK;`=X$JEk5wNGX1sa-9I$G>vfbvJXkRcVeREyVd^a+~$=FE(@)uqn
z^RxFbvC$BUByfHy5<`gg8Q-uHnS<3RQUgSdXe5*eiX<ze9%KO_Op%nLnT2Vv2~-)(
z>8~zO4=8#Y;%!4(pg4#liVu_;*jV8C=cen@jph^69mL#&Sil(n4(8s$0$3>g>P~B&
z!x`_1V$;~uj36pvrpSf0?fGJ_LeV?0Xa^QE^TV+yToy-QQG_gx#G*)99EC+uvN#%x
zqGj;~EV+Qk!hR_FB9>f~#o1VrEsHN<$t78Q8A~q9;v6i=k;QpflE>hE#cGieW-aMN
z-w#Ob12TY0Chc6jaGR>If+X=!^H7=K4V@bLHR(}w8RD18y15KV%FqPH?aj!j8Cif|
zD7qP0HzQ}j`|r7(d=wr!inyg1YZhaDM(w9q^C{McV+{g{V+Sq>5CO?4ru89(8|=P`
zP+*Yf^j{Sq5GZmYhG$}PI3g7N6qDa@Sf~$=e2Ogrr%tR9O(^V2`qv$1T03~&N?Dvh
z7=wku!(n-fq$AB*Mv?P10<HhvygiSYRf^eCtjz>cB4#Jb;$+NDmc=(P`v%r#EOZyM
z@5<u)n0=qYpfIsm--XCs{7X#S9}`2fUj=7tN54<|?o1E4gW~jGrC>RV4$%oQfa6MH
zhaXDA>7?j#WLl1_!I>0Yg}7B{D3cVbkaZPuWbkvuevWt$EfoD64Sy~hP3_2}9gT*N
zrAQsL`+}UAByt$59+ov>F(z>~0!)}Ys>%99*ZOASvJ@;x!Dfsvim;#vn*n3q-HQ2{
zj_pm~ZM1xgRiDe=mh_d!vuOHRG>0)wEt+48d_naar##oHxUj?iHSBIQY!92Pdc=n?
z{{Mn|)7PnbSi|ye!sz*f=OjpMedW%2FLT$pbH}5sVe+X8!j^=&NwSIQ9_HS|Vz9}b
z@`pu5FFUS%bMQd2DyRPv%>+h}UXaPxsG-9UFAND8{r4cpJ;;^8@kl40sb~}_|9WZ2
zgrT2B<IbX~46a66)yM$sM~)rxs6}%b{28f!M!HO>#A3TxJdyGEK5V)V+cCHR+ZEu6
zpe;ogV$(uw$KZNwUymn2p@g2Cwt@Ens0|Ae!oaHl2&GQLLT1ho?*HzXf_pAmq3=-v
z^5Ke0KKDH);B%F};%JSMG1G~E4`B8Itifn}8?$d?4bTuS%w&GwYvb$uj^vpu4@5_N
z<MwKuxbl=N)w_Lyk@ErM&bTHU*=8eWz_p)S7M?t~e><r-`|y~3_$S!Gv(sIRt}Lj1
zOb8x29+M%C>g-D=c4qL=cN?ma|NE%eW>s@?A&Ws&4C29#C^{L9PL{pm<)I0A$OY(4
zmP0xeRkdS?YFSt*i?LB(a#yOw8c-}+M&)K&i_(t~4GtpRgUluJ#iyxT&TO&llaP+=
z)0yW4ML$6nPmn#tX{%NZihstZ31b3gC&=cmTbO-|N$>9to0r=-ElMU!eMX|s$OHoI
zeRA$qi!1~2cbf~cC<%apha`jmn1vAf9*|J!{QZ9RwI<KUNg2xZnr}AGMUdv<uE+ub
zP%2q4l@;K>G$4u0WeNM5{y$77vl>n*2lZZ}#KuNKk%6ARp^mPAB>)B%%YYf`Fhe6|
XsLu>_nV}&wG-if+%uvh>Ma=MjcgX?#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp.16580B63D7D6888C.idx b/.cache/clangd/index/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp.16580B63D7D6888C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..46b6915ed3125ef9954cd568d6ae82b1679a712d
GIT binary patch
literal 6992
zcmZvg3wTpyn#XfaN=YvQEfm{w2?9A3F-Xq+evYLLq%BBVh!$uIL{O$JWT319E)46X
z1ELHJpv>SdB4JtW+EOJX6j&F=pkWwfV8jX2L03U8!r&DJVJ38+L;m|bo9F5God5U!
za(?MKeRIBk)jqFx;vk!C{A0eSn&-C6&al~Rd8wbpEsN%#vfFG8IX2r<TMq=zEI1?f
z`_u7n|D%wZ-1*m!e&NhNz2L>>{1KDZPgq4w$e+FD)P{`@-nud5QvRsPH>W(A_wHtC
z@%zl-*aulAIqU7J8Dm!|{;IE=u79<5)V29H9+tZf?7J;jwtsQ>#QOg^oU`<xnjF_q
zztuU~`|^Y}FAsR(OzW=bEw=0S?JWb({@1ZTp1akp9+*Cs8ehHR;>+*Rbt?`md*y8R
z*&UAvJAN#>Hdk(5)oV=C>b7e$7q#<koikcS3p4+A;OnZ7J3Ifio!!~klF7~ZQ$^*@
zwIA;JUH?D)+luNpH+{BX<*aDe#81{YPI@?5w>|5f(h<`#f73qYW{LcG-A~=y&a{?p
zQM0=~DOtL1%M#bR){TBby|nhD_@Ol;=aL`lY<=y?H{QE>-D#LDZBomQRoO1#q1Aml
zpRdxDCSO_BiZ}nI#GDgv>^?EEZE*Hb$L13DN%ceN?5fDlVdlA}`G@-tx%&Q*Z52EE
z>(S{Mm!Ix?;>ra+EN}kr)=s9VQQE(V7PwIG*ui)9v~IrsiznvxtGHU??0v3p%Rv4w
zqsML<HuYEU*VP|sY3Z{mu{>LSEPKzhg^lm++5XG~vCi{xM|;P~LHmlH8!2ZTKNVRs
z+qP*?CDFX3;N`D7z8^}z)zTiiaH9E_KHs8>^@HB(U3w+AyQb84@xIe@ysd`^e)rsx
zjVlZ5N;~?^*mGh-_ai^QSzF!sHvdnJvyQ&HdT(R*RnKwe*=euNy|C?j^~lg&npAvr
z%b-7d_pdwu!VuzUMa_HfB-~$fO{HIMF5Gdo-ZO9H(x3KvN9=mG?Prhf8#Cp^zRZPN
z|GhkP|C9UgA70*kfj;)P`x-je4_V|dt1Qdwy0-hf65{@!nqsf;$BVS!8~e8S9@u?)
zR`@?3?SE={-!Wr9n)>SJTLa<jU%#EvL{2MtKigk@zUhi%Q)1GiSLz0nuUC4tIg9xH
zgSX!E#E$CQlxN|}Q5}_&fBX2YrmrS<eEiKv;mDy#-<r8ke=@+_{0Mh+_X9PTaz8(J
zWABTv&U@bXmy%WUUq9pPzI?JtTy|qal%KzJThTAys3@E^WN~2QwbP%>3B?16j?!tR
zBTo)G7?-%!D<3*rW9**t`NMmSjt@><8hNnji&JLhrQf$6`m_AthUHz84!TZ7F8aRI
z^Dck&gZTVsKd<D6UcVXt?DM@pEM_lkd}`hoW!LHKE3fSC->|N#?IKfB^4!YFErY{5
z-%U0ySrT~diRt@SoqhL>?ce`>+Oz|!e(ZDd>4lfCjr#tj+Z)Qd4qQ5Y<yPDIhS&ez
zDlFvdmKndjucl+BV_D7gE1q|+ZH>Ql&#~XHSo4RjZyIlmJCt#(X_4?+?Z%&a-?&`-
z<Km^wbM7T<woGUArSY!`!*a-)fUSN&=76l8A=Ujouk0J!{`yQMSCd0lr^>~d#d$r0
zRnBW)UA*M2ONVpFDXDT^W?oj$V3m!VUtQUCerbCSIW<)-%q;BPGg##n?Bkn1I4%5h
z4*6)RJSKCD+A~<?OS6aVbF){N9P;5*xgfJ3t7ovv-yOMLSts^8ltVs}DmycsWj%vc
z{$cmA>Fp~I{VdluEmbb-Rc7=IR{2)HKxg;D-<{7P{i$+(W`0)BV3lo)=FVH(P>{Yp
za(2&kdGXNAZEc?_!*c6ghO5;22&v?ARYCH*rlu2RP(V^X3P{RN0qHS3rPf8st@rpm
zRgnCisp&);G?28929oyEKr#jcB;#X%Wc&<}tib}w`dA=YKMN#ha6ocC4oJ?=0m&OY
zki3rvlK1mK3Wflr;1hrp`~r}oAp$A-L?A`K2&804KuSIdNXah&DH}47vQGw5_RByj
zh61GGQ-D<b3XrOy0;&2`AXUE#q-JP9YCa7}&94Ee8#<7>PX|)>>p&WY0i@wGfHeFD
zkfvb*Y5GhcO}`1Gx7a)6ZeP`Vo!;S)MtVnC$@<nvHO-x`3}~8wYj(OZTtU|YNFmqq
zbfN+jkW`QYk_u5kdIFx=>BjH`JqsX(Jj>IG4$wf-K^jOpL<7kL7$BJ-10)k-fMf$K
zkZh0zk`1vyasduVF317Ng*YJj01qS|<bmWvJdi>_08$7FKnfuNNHHJ+DF#I##gGW3
z6p(<Ff)bEYNCHw0$Uw?L8Av%K1E~ZQAeEp3q!LnqR0ArIYET7I4XHqC0S!nkr~#>k
zG$8eW4x}E`fz(4fkVe1&(g+$r8X*HnGhhN~22CK%kO`z$^p3Y0Be&jL?wtau+FNHO
z>sgv=nmZ3tXd0($c)BrMlxs93w`*KFQBDd-ilTs|+!T-=r)PM&F+7xKG$glYTsqNC
z8c3R=fu!9ukc^W7lA#zN88-tY>tumsDHcf9%>v0eIUqTT1Cn!dK=Mu=NS@+><lQ`w
zf>QuepadWVw*aK*6oC{e5lGQ30x3BqASFryQgTZ`%1#+bnUaB&-7=7hQvp(;6d)D1
z0;K9xfmA6KNY$+ZsW~+uHA(|gb8A5AP8~>{(t*_7I*^9b0MeigAPu(xr0Fz)G$|8E
z(`^Fj9qoO<YK&CVjP-gTF<#zEMSH*9wJfh@9g$IA(Z9lPO<N1kmDxQDPUxi=;o3)}
zJI0kDTt_h{3D+6SVbZme1U^E#_F|5ct|OSEc2~@v=BXG#p^r)s6#A$nL7|TdlN9=>
z2uY!jijx%jsHmMnpC?9m_7Ums^&|+-QOrrga|Uym^z0;okC2|dnB%192<E8W6SJpz
zI!4gwqZ0&;J~~Oz=%d3VjXpX;(&(e(B#k~gYNyf1#0Um`OoCw0$0P{`eN32S(8okb
z27OGNWYEV%?F{<Z7{Q{CO%N>l*d)QCj}4P7`q&7`qK}P}Ec)1}okbrPBRKSN34%i(
zmn1m!abc1}9~U7x^l@>LLmwBlbLit^1dl#GLGb9~lLU`GK1}lH<0B-GK0Z$J=;Nby
z9(_WL5YQ(i2myUUk`T}*gh>H?LWC61C&WnseL~bOpihhuBKpJxA)-%A5+eG<Fe#!>
zjF2Mw#5gIUPmJ0{^hq&7LZ6f%B=kv1LPDPuCMEPq5mG{*6elJ0Nm09mJ~>9n=#vwK
zj6OL@$mo;9q>Mf}Ldxiq<D`r}Ick^Dr^E;aeM*8*(5EB`1$|1GRM4kHNCkaLoK(=K
zMC}Ut)EJ?nPfZXi`qU($qE8KzD*Dt2siIGflPdbus9i;$79%wDX$e9@pOz#v^l4#I
zL!TBQHS}q5QbV5>wQK0pV}y=AJwfQ`)02dbK0Qq8=+h&ljy^q3>gdy>b{&02j4;q=
zBnSh2Mv^elXM{-ueMW>d&}YO+1ARu+ZlKSM5hnV~1Yx4jOcEyg%rI%9&y0{J`ph_K
zqR))l&9v{K)-P>8e%q2vtp=(_RMlFmfr{FlC;RUDV?L2v@2qnAt*7%OVZ*|K?zO*4
zO&5#B<E?*Mr%uivbz!~QD>Yr7Ro>S+$@IXRD?fd*c4BJU>+lv?)7GPq%<kE{O0P8V
z1GHW%A1obTTbz1tUcbD7)=K})%g-#W{dDU~sp&CgV|>=B+Mep_HKlpGH#J?@r*NRP
zmg)RvSb><~nwX#3ZF0B<TfaqaJ>{Svc^sa>)|qnaX$K97aWIfr2MdXFaFBQh4@qzc
zkVJ<FNpeV#WQPn%aVU^fhYCq^XpnS=4#{vBkW7bpHyvcFAL~`V{c`Hs3~&u_Tf3Ks
zrs@|PoTrbFx%IvbAM8q0W>nry=f?H=ws+rqUr4>TyrA4^-9c-ck;v}ZUbODuWk=gO
zQtr6(*HPdq7-vz|Wl!DjJ6nzxo%?^5pVp~Ur@XWAXhGJVWB=@)y+&&~bw7ERWSuy5
zKY0q0hxbTuKY1Dw!!wXro`uBm93-CSAql(yN#sRH5-&lLc^Q(zE09!Pg{1KsB%RkG
z8N30><juP&wa?l#`_888;zhy5HFpkX+!_A&HQhR}*=F87wAarcUR}Ul&0CV$P&;$_
zj5=HG-I}c~eb|sb3^qU3^h7FF&YNV<=#|;qk(HBsPrtr>e0k;Lsl04oY5u+a2NV{G
zlA>z5VU8^rBah%qod4hKz(IqD47+dmNY|**_doF9I1j@KL)l_CT~;t~Ld7i7JNcoB
HMAbh4B*n(O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_gemm_xdl_waveletmodel_cshuffle.hpp.6149EF88F8FEFF53.idx b/.cache/clangd/index/device_gemm_xdl_waveletmodel_cshuffle.hpp.6149EF88F8FEFF53.idx
new file mode 100755
index 0000000000000000000000000000000000000000..213dca0ac4b403918dc0b391a1b42a3cbc5469f9
GIT binary patch
literal 9082
zcmcIocU)A*7T%e=%CcOS&Mwk>@0X4u2ncFyv0+JKeFnuSc4JTw0YyNRBF%=PNHYXc
zs-U1^MCB<KG*2a>h>`@*7^-h@_jvF7%fDWJ31`2|cW2I=bLN~gGb<<{Ak2uzo47n^
zYv_uNOZYq<PnrC;VPn|$M@I2@mRdaC)~xDzkJmnSQ))lmaa^HPVN3gzq5$REg!nbj
z9?IXZTl-Y7VCkB=gXRa450!RYN>lOQd#_<`Szbi#-hbUXaHagnE<3%N57C3pmkosH
zoGT_48^84SxF7T9kgwPM`rW=>E%j!t2CDVVIsc?uPCLJhZU{bpvChXnEH}EpYR0#A
zuHAPJC5{=|^D^f6#d}`q{cg1eJO7r{B-;BwF<K(*+~j;AW9EWYMP_$bF4}q6Yl-HP
zYZGfe*tE=_Y;3UmxVYk<QDLi<<@SdBJ)v9m<*k*C%?m32<JstwcqH*pL71Man$d^o
zgKi79KXrVVhcb<7^7dr6pX;M?mmPImnCE7<!$DnKzI3If?S_U`_Ybu<H~B9!>C_a5
zXFmyA|Ma4*Sgf4BbJI`9`xLuQz8W$V#g$CXKRvVanUTN2kC$ghKWk0Yy!K8Mv$|%<
zI_Aw~_nSYBU%kTXgm1;w5BC~ow`RLOvdMD_)4q9lv~A@4y5TsTX&v<n#yusP<r@zq
zp4nQ?Bo#ggt(`8hd){lNxc!Q{LHNpbuS#AkoL*(SX?vVrLi_ysFf^w`Zui`)Df30s
zGTw$PbvL)Hj&xq;?zTGM{?(o;^sVZzgN<Ux)MZAg3VBvxc~?{BwzSPNO}<jH5ZPta
zRd*UQUbd&cKVrG$KRxS`h8#WmPky_;j$UXnde3`x`z=?;ynnLX*wm@8$}QmKbv>r(
z?I9PZfzqby_dTk%ee<R(>!NqvO^bM8X!ZULv5!#eiF8ZB`xC|sy28G*KJM3g^u{*b
zysks3kDtaiC5dk>Z&6kmy4ZjC5AB$!{i_|yhNBKFy_)xGomHfuXwSTm0c~C}^H<}#
z)?Js!cFxub7OzzBv2Hlv7orukBlM4gIiq$SdA|DYs^`LC-@dP#&X`O@zxnMeTP%sa
zeWZ0w`ob3Lg;q^#Om6yir~h#2X61#I|GC_Hyv}mY=(gaVV*7zp{E*_6<H!0h4O*kO
z)Ru3rw4zsjhsB2SggLqEGnOw%dhz0`3;9VY=9#$`zc>$7j13O$>qyhP<N8-g!ugn_
zO$F793TC*TvsayQ<&tp1H-gseIn4!@Dk~Q5(0uc)<JuLvs|2M@Imb6X|JXL8>*l-o
zr>mCyvAwG$eqp#`ZfSnTf%n#V%UZgPs(DXq_qV1y`>oQ``uF7H{!t=_W2Vc;eP}eC
zy>wiI=8t{T)aU-XG34bnyY*R}=er&@rMOJkv~-zTPm<}vD~ldaTKw*A{=ii4fx^{O
zydORE4c;_6@J7qEx$jOcJ8x*Jv)bFFG0{CRK&{7q%*^Xo_{Nv2ZQftcUNyxd?Avg=
z->wgNOkl=GWSkBuanFiSUVJTR$Al&$=SRL1n%>m%|MW5o^)z2OYO3j;#J;SH4`a?|
zJKgz-pW9lfux|IZS;kXl>OLRaVMDL2d{!p6cHo)IlO&tmXQ?*7n;RxCbKSBzp=|4H
zQOP=o6s==#f2F#89nxn#y14w}b=oReHNiBjJ@)Sl9)I}!bGy!az2IHjo?~TMQ`hq@
zC>7B8tr?CDOG4&MWB9k$$!UwVuDL&cZa?(QZF$=`zfBG6I&!~vKeP5w+ZGQ$lg{jf
zUwe$FIeAQ}SFFlw(ch}#pxatjZGg&L>GIfZo}Pxc5BsVaeoWbRb911Qa`aq7owoZ*
z17%{5@V84JKN3#4y2LuHy?frm@l-&v-_Fnz9o30xAssnx8(sd~=##48`i4%5FNwWl
z^vx*OhvL`Uo@({_{_<dEmPt^1h)qbB;DX762`dM!HbyP1*6(>Ab6MOx`*ulQWNDtZ
zX=bP37C+EBN-QyWFm7_BQ})r}ABQ~uoMJkyKdGSn_xj+k_(&LNo-sOQ*<{1DcjK!!
zXWd?OU}ALNt8bp>My>N)S?cyCad`gVL*G|>XBrJp6ox++q-Gx&b9U%BFTjV_5MVGP
z%BjyP^VN5gyu^e2MdRO<ZP*sNlsB4W?5lt8D?jNHu7t%FJf?->w@_l4;0Z`JC;zCF
z-Z4qd94r@M14|xr9qV4lRx$w;@dJlY+Z6fZ3Rpm98A8G#q$3kRk+?`C(Q<C*l&~71
z{7lLJOlivmP;`G<>b>>fdx8`&O(?NQ7>iV70w~tIHS7gGN#BUD4bi0z({<QGCV*o1
zyg97s!cX}McodnX9|`-3E>Zy$=9eo8!zX^#6=5SXO9j@bz@{<*6z4fAe~jPTcS8fq
z5lS=`L}MkH0E(MrszlxPnYKDsCJV)(v2jRSCV=8MHU8lFn3_TrJeE+JDWzsgOC~;3
zwrD%NY%f_U!n$OkYgqRhv944A#eb%E_ddbpyDFbdmqRJ#P)3YY03~p~hl1LI%O6y*
z23e?#QYs_Xl?tE)E!sDY9&TKuf;C4dNU02&$plb>d&S#p*GIG~;4y@90|{>+u}lDE
z34Z#+#zNEeqp>QXG+~7%tRfRYS)qIN#nhekFO;wrp>$IQ-IS9|0Ok8Jdhr)K27VG@
zJ+kQ<tX+d0WdbN6@=K>X9sgsJ2wM|MCDy1UK9UNctj<X?7_6vWuY`51cuW>0%A%BI
z0w|&SCxQ-lUOhA#E12_`aJg`+Pk(sw6e5w07__-*PSf}GyDCN4m@HF|#r4=iCSZ<j
z=XW``4N||WV<obmeQ4}H;%BJ<%J$%bt@E|K2UW0yP;!t`4zaCN044GpgLU;+S1%S}
zL$c5%tbK`WSt@`M9gs1<usnzuPDEz;Kq-BobYuc3`{Z_1?SJ(8oeEYblq95-L>7_?
zpv3ulu6XJ*F;5+<5Xycuc0W;9Du5E-Sa)w(|KwC9tW9*u$7Aw|6Qu$uNjc_`)$*<J
zB5X=1jaa`C+sOn_(!7-K^Cd4W6!2)W&~+rdPNGFBfReF3?$(rm-gcVjX~tuW1;#d?
z{_y1am^Z<MnEzN!(4(qxr9UWOc``>M5;l@KqypwBDB0t0jJNJrz;eXPx3TCp){qIH
z{JduHL2hSEun2n-Wy`Q?879&HNdV<aPV-o%IH6J+E6rHcjFlOw07_%n#marJ4R?z0
zcrwd*ta%<A$^=jz?5f-qc>Ao90u~Z=Bd{<6OJo8l4|D6k>-0MMg9z)BS<10SIf+}T
z07_emrJs8@!&ktHWR@l*Y(g3`0hITCy)U;U1n&^xab%WKtXWEQkqV#;bex}Mv+rQ3
zI`$&7B%rYgBow3qD8pm%7=xVtm7}p7p+sSYD6A?IKnYv1azmvCk2k9NUt|BepInG&
z#!$o1`BUO2f5zCyW`v|TJf>(8@p}7^2^<ERWIDY#WCBT$CLyM`ADKY(L0v&YSv7%^
z2-O7B%uy^linSoq3S~6~l!lr@31yW8W}wNj5nRGraH43YAM5sG2gz3F80&^Dg-~h8
z($X>mA4(qlEEqg}!4dPBqm=k4C4mIOXMUlKexZ!vODIHgLRtBN8E7U2>!n~b(3xgZ
zv0f?{r(wM`F20R5Z({?9VVdc~YJFH6R2@-Wt)Inuq6qiBf}78Y7hYqn*Vr7wR`0c0
znf&Hw-ef?Zr!NHe7*vF<db`Uj8u1BV<2BX<zM*$%d5_y2;TN|G1(i5B3yZR_7rV7n
zSm_iV4_GKGv)~>mv&22p;>y~n1`BGi7UcMmJ?r+d?s1&|_g^=^U;L5;Km>M-z?0xm
z(aa#W8pIR8I8dcWjaXTfigiI1n#sienRpJ1bFfJcc4l!AHZQ_%kmiK4+77A%j|7Oc
zuzwzS>+)z#<6kpI4gud#9jpv7E)dGfJG<316qtr4Lslj>MBZ7*hsAlwI}iCld7zmh
z6j+2Nv$zTcR-wsInrWsEdDkHy7I$I)E<6dm>Y>|xWA*4I-87#m#2STI0z2e0MOdSV
zi;J;FF&7U~^dKdI3d3iHD0+yChbjGG%8(uZNSS@4ELgmYuepn_#o`FQW&{^U@--v5
zcsE~lH(#CQ@8Q#X_@CFG*z|W%MCi5xV(7D!&RNP5S{!JYh-;+X5}4Xg)+UjtIf6w;
zxM@S$EP)UG8dE!G_Zzto;vvf95aj~KqnR|yB#p(eBw0$@OTki;?q+#rF7V$#lek0V
z-|?(F(-4!!#hHl7<l-zuB7nm`gP1dDJnQZ}#N=^tAz})-xCoI5;qWUFQ;Ej2;ZTK`
zDlV=?Of47JAre8K`Op-S9ZFk4*da_v&aSi%gaMj#dqy{q2Q-4jX4oPPPk<Bi)pS?!
zb^pUNiQig~Q42DKQZ{{XIA$$h&5{fZAlm^n3(~{a^8VGKE0hw*zz0hH1EmNdF?UHy
zaKZ7%9%LYvuNuqOfPwj{*Brhpl9-GhF~2w!IkD!?MB+>?K7+((kP~a(LL@Ha;z}g0
zL{6|<n%GKQ%VO9hv4XVKWv%cK2_7PK*c?szPC+GBhn95l1@|?JOpT5BG%2P!QP=`v
zkDPx>l0_LqU7<+<GWwaagStX9_b9b{ls*(kXr}o>Sqsf-&`g;%Q?_s%Xr`Ca@1@M3
zwI0D1y_7xRrQtU6t;sKxi7^fM1`<eIFiZkK+J?jYkekUNIgCYzu`xJUDC@#uN3dn$
zj5I7t<L2OO9F&cN;X|6q!LB)Y5{pk`$!ToP;$m!G%+1NtULF<#lMz*=jXc-}RK~(j
zoqLe+=D4JaiQyz_ka{d~hf)oFJgH;ZSd@+R*@bg(P!3)QSSai6VPR-!$imW|9u|bK
zBH<zJ<Y9m&4XRNi@_^oTq&!)qvWEov5@Os8EXv>-SK9Bx(os1(w|{k};Tj<(Bh6%F
z$VPfP(oIKZfZ=Em9i_JdPz4A?N9iR1258cV8(l{p(4&tuE*2@Qj?ku&8KqYP_%LoV
z71fpa(M^wL5|LgaGKDsC1e+%!J1%xeL~bljLgq=x4pgFvMlMNc92ff~p(!lRLFPHg
z4$Mpf2)X2-aa`<|gQl>!44Ic9JN7h}A(t{Vj*I=u&=eN$#malJ1pH1KSggE{i{r3z
z92f7$%KN!E9xKOlaROFO;NrJf<t^KB(M&m|Q%)H}(xsUSN~eO0D=D2yF0P_<s<^nC
z(y8X+8cL^ziyu;whm;=V4w`vHNgi==D<x^=;>VQaF&96fBu}{bH%jswi=mYu5h1+^
zLX;jbjO#2o`68U;7DHFVS&&=cf=Gh*7#1DFI$&s_?3xJQ9}KPjxT1OSIP!fW5+ouu
zh?wLi<=7aTSNUY%6s2^E(uazi^4zj^Le0reGSEj!`Y2<F|CA5kEsj*wZXt=c8QU~t
zceb`S<0;K}9*eUnNj5hTXHzcOlrLnN)W*Qmcca6tiSlifS{rMHk;JOjOX)%8g!4l(
zwDe{R=Y(eNBf))cjTl7qAku`oN5TOM%CS0J2jM16(%8r~_}mezODigp%4nts8~5NK
z=y{L$KK0U_Ic3b~kq*zK5SxKpMzDJ!_GEDpHYwtc7h(4z?szdaDdvtBWA|cq9FiB=
znDn~MZtMXPJm7BZBceYdO}LJcxW<ACtPb7z$%&73d@uKBe$kY02O_yjdWnY&KjKVD
zGuCIFca*X^N;!eE3Y|P=G-Q95MpkIT+D+Ju{ji%->t>zx^P-22C-y6<kQz{kl?$;t
z7>o4XSh<Lci?MPsi{br%tdNXF$ym(Bvh)=JV!6axesSCEc_XUSV!c{C1Flb`?ZH+9
zcoF#V{G09P#rt~Be)0ZQ=2CL_$L6<z<ObJAsrFHy+n@@O;Lg?gFTVP}M?%&0mf%9q
z-u5r5TEom*KFa^Z8)DvM<d%#ku&F#3x#V*D)^g-h&i0{~Chs-uYFc-h9NSzpITuZ5
zb9z3Sm5=7JxExI?M^o7xUW2CBpxJ;SLP(C1KHWehTsbpi?U(>T)E9&B+Uty6*&Qy;
za(rf(8Z}JGL2BVMA1V4H_eQjfPw!%HEPN({Pe*WZB%hAt;@y0DHy20o=_nS%;}X&E
z5EdQchM4rF2}11Gu)LWAzlU!4zx(~>wyO0vddoFPmTf|sO=udlySI#fjEx-swc@8^
z<Wq?KxyQ(-h>MG{Pce&cdztQ-HC;RPi;;9^UWnwM&DnZnX(zsu2ue!x{Y~?%^snwD
zHX1;J0i?ku(_thSMjDWV?!ELW?f7jBMZO<Kw!>%^4BStv?XXnwj3)!(&f(sW)8K`H
zBwy*99SqQ<85GoG4d@;rTazb;6f8>Nw!7)rKmC8798`0#5hRL{jALHNZEvMde&E&T
zZnr0XD3G8png*PpOs)i<9tK%e2ax^%vgX=Ee;8TA)<<4bj3cl;?3HHHu}V632bowU
zlZ&&lK{mEwcW?qLo?!3Q<g8+~Ty7s-i<N7!4x8~_V5JvW8_p!nyv3$(u_KFnv05+I
zhm1IqA$10@1&ar<&L9^jQHn{F7W@5KO7SeE1x>*V$tmjtFBNS`a70pSk(7ib?xxgs
zv$)-U)H^vvaU3DuqvY>Vs%*#9P04pts!-csS(g_&>Exdv!SWG{KVlo0@r|G$5qq3F
zKn_4U8kdd&*+iF*Jo8Zio3Cn+XAKH~{M0qD=%24Lqg_8GI?Y_9&l2;IZay+$aXHc~
zNBZy~&D0>>8e{^vC#(FD-8lo}FNO=e6O+vz#G-@Tv?6^ih5&@7i`0AR8!;PK$B^PN
zq|NTJ1SysvZC2F=q}YJ8LA8N_+)H*(Qr{7)wID$YGJuGMmB^#1^re|y;eTJ*Xj1=;
zZeuHG2uSBoS>{lV;Axt?#=8DOd9%0XE0p3DO3ZfFZIpf+We$Gjg{@e##E&EqzXF9e
zwLf>xL(n<EMV$w_K`X(7!T||`2g+^;Tecx{Bq7;DR^nkM#@4P%1HXDlyL-6VJGeMH
tIlJ20@f<%hc&_YE7xt$u`_qm6>BRp0{LIed*|GoEgZ=5w{&Zk}{ue^q8z}$)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_contraction_multiple_d.hpp.D8736023CD2927DC.idx b/.cache/clangd/index/device_grouped_contraction_multiple_d.hpp.D8736023CD2927DC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..36d4244faa9a266178eeedf424a3630d0317fe9e
GIT binary patch
literal 1928
zcmYjR3s4hR6x}TOxN%9cG2!LwPy`aSQ$YeGgy;|}7En>^Fe;#^AQhn&f)o&}BM4%w
z0*cI7ohr3sOD$#mrBXgqQYyCMh=^mQ@)2bO3-#Ap5w$Pdxg{@?yZhd`@7?pxy?0kc
zMMmxv5X6GysN9T{?DbrNAQ=3&WoK>nBMBmsPY}6Bn+~qq(i0$Sd-Z$liBi$cfygC)
z_oef6Ase}=+aJU$trdStnRsdC(F=!mJWn)7jT(oa1Sq(tUz&B#qc3GWx<8Srh^lg0
zo+igtUG~(U8tvNl(ed`V+l7~341TpJ)w8hYi>{Jp`$Q3^Eyg>u)zb1~Ez|30&5ix%
zny{?r{dpIn8*~LDbktOR#0&Fqe<|nJrW>_^g=2I<r=c+|@u06|#F;F5IufyOmT%Lc
z@g!BHGd(Q5p!j@Qc2a(^w&O(4MDgp^f$xgS;(D`|Bg)A5)*mLDS0^gl%^Tw6YmbnX
z-h;FEjMmhC_Ugupyn4@<bvN^QIi<JXF6t=CZdjPg?Fy>WF=gh5nW@JTcKa$zk^>ap
z-Bo@4GCTDm+SoO;tstVeJg-T$&O5fkt{KUf=BN{nwp|_A|M24Dj1p^!H6-n)^2H4P
zh4!O{pB+KPLB(fVzOQtyZEESRdv<Pi|LN@X`>Qx<1Jz>}b9#T#Bm|_U#rAcZ6HnK)
zhnfb59a}^9J{ii0?4&M!Of4OcT`T>g{D*OF?VZO#w=$IZ{0GA=DIDg@?5|^v$6T_F
z-8-Tx6dcZMvkP=ud4u8h<2(*M=kr~=qW$*H63mJ__X*h$Aity8@G{9ZmQ-&Tdb2n_
z-DhJ>$G)nbYpa?a;*9X~F%6Qt*S91ZA5}Y<^J})RG4)(63|0*usPPZ>7c@?kdTo(R
zpMPt<-zk4NVav<dP|YET>6_cH*9qgA`6N$FAhAH~#TtV0z#zdv98=S%W6@iZ3nG%1
z5J;+&>RCff!7v_SvhZc$37@_vBBdBcfhdAC0B>qt8W;NRnGg{v!mwN^*RuxT-F~ab
zLv2;pB;*uV3FCpRGp%NTOV{&0TyXb#k%*jzcPj!F5v&2Yd~YDX`*QFe2`PZxCawtq
zT=DvB-PLbw@gh==cdM0ZJ!=5&^70fOzmnl2B0VuIQ_A$L0r<wW!6j{top<;o9}8m;
z81$?mC@v{vh3VPcKgthS9xNv*hCm_O5FKk^f8WB!lP6zw){Drg_;^Oi=vf2s(4N0E
z%YWm4AR=WLRw-3_)&NXorEIbpJtp%*3fa6w`R3Rw8W!xJk;T^%1Oh;gI6^izkW<JF
zUUnr0mRBy7>K;^t^E_$9P>dLe(nv~4#c&jjWRy${8qr8j$;FUN8X*}{!Ra(o$P_9_
z6^&GsN({H7k(yGAArUk(X-zYLG>yWvVQw?C*6cP5*M_^zELzL|vqXm}J;C3128TvQ
zjZqKBrR?YV?UdTAc-b)5Fb~oPoUsP3>H`gOodfssxt$#z_%0L@x!r|Qh$io1o@ssu
z0zj>DD7Ho+0660St^$Ps_Gpjd(F1Rc6c$;+AUBaEQsgGGM26f%mdFth#>CXf$Bmd;
z;RCU-$PyKD6Ir50lOl)6Lo`sFKpvBJwFZy|XH44F7eM;kps;0a$C9eChE|PL2R8<S
z*xS`6;0=f2WmkJZX+mfZA=?k&MI&>#ITHBb`ZZYQ!3@;FC%GC00^ASaJrjtLoy3`D
zjB4=4<QS_O$bkxYud9n-Z&phBdLR5ffwCZg2Rw<bN%2kxvC?A8+Gv}DJxS!PdvoSW
YZRku7h0=@ga(#{d8%lT~2*u><zpVk1asU7T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp.A6618A549365EFCF.idx b/.cache/clangd/index/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp.A6618A549365EFCF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ecbcfc7708aff6165896322881a5279012af775
GIT binary patch
literal 15300
zcmcI~c~lig6YrhwnYowCaJk67@3Jp~qJWAB;)05f3Nh*zHNj}2iP30W5+ewx5Z5Rm
zXjC-d0t$i{6hshE5fo8eP!hlu#Vv@qpuQ$^%b9n+bKd(WUrv&`Kf0!us_N>h>gGp=
zg%t<NWFcRSTrw~E`x&%MCetJTE&P7Lob9YkcF9sETe7|6lbhe%9H73tuym;Iv|TQ@
z&K}#l?vls!DPdRVKc*Me895u#9(@X@Hx1$E)MtD+88Y5@RPxNMS~JDmwwjiW78h!E
zullpJ`-bTsF(FU$yp8g{EqdP_b$-d~{3*f4T8#&H<<pOKD?&nX>x9o|>ji()?9pet
zVZ!)x^AD}deZ0AC-V8IZ{?8xh3=e<#Lwr-K<N0ATwr$~d6#6w)>=~?6@k7exgKK>3
zZh4Q(USHTaD>7=z&VF<5b+Rc}x>EFxPHQa+>Xo!9II-UPR^7_g*+b2Ye#u`}wQu_N
zYrfedgO?t(PMccRo_1-9YkMufagkZZ5B_uR#J^nJ_I`wpanXzwV<yj>s8#QID_LjQ
z(Aw$*yR3x#^|6U@X3^tH`?f?q+7o+m)aAR+mX^JLxMce4p--+<pJ+UJu(!wBgmdUy
z6RlSs(f!TC*80Aj>FNA-#Hyp$=MUU|^tr|TuTKt`{cIz3JTt0h_vnVPC+6JRd}AE5
zW#PK1^mnV@ZfmsrV%$%OH|EXqoY$(y21bP#lwLb^?2moVIy}1GteRx9tZn56)y`Ym
zVH5W}Xd8AiDA|6^UG(MZrof=t&%Sa0%WvJmq~DWoE}yydQcZSJbZOJ6!)v1p@2(hL
zxnN@A?_V8$*Xo}CXI6B-AAjFv*R}X;(yRrgR|-cZ<!rAT*r&C)@wweQSHp#+`_3FF
zeRv@$#(G>v|JS2Dw>>(0)g_%j5K!rQTGl%A{5O6Z2DN<ZTa6y-+z6PIfj?dLVdMH0
zi$4hoiaPUBw`E}~>$_;#43l?no{dVpx+LLDRMOqr`QNQto!mL+$kf)mo%09wc{21`
zpLXx&@+I?()63&>mL8a0omRQ9wtPVOViUPtL%-<5O<BR-le`m-?)@t3)zK+_k%8S~
ztG~ToG9z%J^}^UcqdYl}Jqx1tq5R~$<so&s<8?BdwEGq;bF3R(Id2GdTJ+na#rqG>
z+E8LU*5kHb*|(;({9n_}2hE)D_kWZfhjNYYO|92<_WHK=oM+#g@k8|co_Q1QdM$dj
z<B+dsnT5-L4o}uK&rBL|d(Dv_XP@}SVEym`Tf?qbWE=mPJl?5jS=!*>xw__8)w|Eo
zk>T;!t{`G!dTQr_w6J}4d*e3@cl<sj((SjqM?To!xH4<tMDHmH6H|XL**&&>d}MLV
z#;b|@PTmT8H!#dC`&zZ`;S~|Rr-bbpIxF*CYigga^E0akj#fw1jkAgwRy3`^K_~sT
zY3#l;o7ac=Z@-qA;IT4|J{VZ^o$Yrc<Ekq1Ki{ATP{u3X?f*q)_U*oscie*JKU%4E
zAwh}5JlD>?f5&6l{YHbUxnEcRid^b{|HXaLfntS$amY{2*`M%lCdIk`r~UiSw6>gG
z64N`vB*0}|@!p|N7P+LFXWzY<vT<IDUf%vK{(G;{jZG6At1kPcEv_p_${f^J=SjIo
z^z7=C?+oHck8=tt_P<fzJ$0M;xlgxdr8hhBZx&b{rwZ24=95^@`j}xO7HmkW@Sbxj
z-|(x8HwOM)`1=#>2NNu8I`{n$Tc(W@6-O#fOhO(n-8*Q?<K!8^3kMAv#Pt0oKGx}(
zkKg-*&B{Gdr}c05+rkZ8H&(V&=9z^4Jmb++b}q^GzFO)3^w~`dZlQP0iM4srHL{IA
z{g}D(=a_zP3T5-D=2pGrqmK^yj6Tx)?PTZu8F@AIq)(TfIrXeCA>`Eh(Sb+ieQrHC
zF#7tLDOz)Xo~}Bl2wrJix@7X!?wo=cw`azuzRcp}^?!}(I#wR@a`Tp9#uKRUs=;f0
z*05t=o@Hhr`{najSLY~mA5Y)-;~IzlvheiA>CFy?CD*zO20w3qa-@FQ;xo0=IzH@l
zeOi8TaKp(zyV~_LEl0j-2)&gu*WGvOvhco^*LoF>bl_t{zSU~}qxXHw|G1XzO5A5V
zxPAB_rY%Y};$&oG!Nc`~Yt0*WG?n|vGv|CWbZ46RSc~$eKwo9)goKepiserhTxJ_o
z{*%krU&$QPKey!i?#z<-O;ejq3bo&SlNp@!iOG&{G9DB)W!ZL(u(+Q7WtvT`lhY)x
z0WGcuN9@PO_rJ0#f7zd=9W(n}8^7I(UmW@MB(Eh-mAhx=E{t20-O{>hY1>YAcxpqZ
z{nK1^^YRB~eo@nIE8IG?uG|~6e$8zQM_oRBxK%S+RQ!7V53NT6dt3jVw(i&O-~JrY
z|LdqBf7@>L+cGi2=a;y1e)k#+|5ChqeQy7f%WY}yT)gSR$X9O;2F3>I&-Hu}9Jcm@
z+;ho8+wt0}(m~$!D|XD>viO?QMC^LkHF5s3n{io_-f!7&IWTin{2krdFRVAbXi_a)
zI&WsMQYORNCoZ0P)cw#9O&cWR4^w7`DJKn)Dacty{-N6vI)4hEXrV#_hRFCvY|w~p
zGz3_fXKN>{cAC6Z8!@DpWUNfamKp*q?A~vyJnK(i)J7~}X-3LsWUL{;qDUUI{8i*%
z?^LM&KLgn{Vs|o74*?cEXG>J^r1H25O(!h1*rOKr*AQSajQnZJ&_$<SsnD=s8DD{o
zD{vnT`N;C%$DvP3E<3A{HJM8Zt5w2UXb7;FeB*6X7GFF?g?xi#d<!;e!M!yESS;u-
zvP}84O`L3Mpp3Us*aZCZKbaPdWHbFKp0e_{JMiOrdqX2+N@lVjEB0e&4S_yvUj?YH
zF4jAvLf&MU`&fM+TWAQd*ts7%Gh)KvUsY%nS#ksJ+kj_k2(Z|Xxi@y%SJ!T-ksBGP
z2-_B6Zw&z!x9()~@ez6PYUDCh#wReU1V&FofW>`Q@`_u&>UAm<GFZmHq`hC#<22+W
z%Xc{jKa3wZNE_)9!DUj)OsbcL0E-VBJ!aWhy-*bjAicb&ZQj#f8UieRm1hT?YAKxC
z3ptQpt|0Xl<g6hdS(Y#U@}23?ohsx><kF0dnu+)Y0xW)^CUXmiX&q9beuSkBtJ{d^
z1OhAp0iXQVz2T2LEHWhn?ZMi6u$zVeOQ2_{{*FPrTUE%9NcTE+xsC%h1Xu>H$==$p
zJFZR}>62bsk+Kz;Xb7+jsp<}xnl+d*MCOF$AZ2!tvd|D<3D&xO-a@NxeXoCP?i`~&
z$Jp}%0hUnz`6+)r9OS1)#-x`dxptD=Ttk3mSkB_PL7y*oHAEJKC7UwKCIblsSjIXI
zi&%E=_vT*6hOp#Q>U^@J1p+MNw5=9hmPKx5k&3W1VeKaDs3E{IY5M-fHh~>3DilVf
z+lq}_iR}vHBTGxx%8_#)jZq_xSY0VLD#i910xX}jMl~$C?y0Lnp~M(EY41*AQUU>%
zDO!Qb;K_+qf?<|3s&ZnO0s)rKmfiKX{vq&774jfDx`Zt*VIK_vmS~R=eC>7ahhE5@
ztf-Arw~<8)1X#YT$SgsJ{03?xIoSbmj4F;X(hy+z+Os3*LCe!8Dm0u7bOm!)h%pG{
zBg+e|tluhz&sL$4L@u|n(QRUw0s)q9r%yS3Og3>4i_`;Ud>SQBqxxxxObb)4G8yp=
z-#g!4=-bO?n+lB}{XD|Xk8nQ?fqoY48W&P{|2;AADMWa$XoFX@jfMbAf@_oB<|Lg1
z#5EHYRAPfl;+h2lEQw!iNuQRGzfc<~2}>qcWs+?o5MW8#u}z-2xG778J|zRarq!=$
zGYtWjt-Ibn+^xDl#1L5$`Q}h&ImGS+0xYSaUBfDio)u^#1H!TsD|eC=3FIS7F+bn_
z(C8)=3MY$xjMb0HTm%9vzn!?#k#9L~sS0@!mIqk<fUpPzSTg3MuiDwN-i}39WT03^
zJC^CCA;5C9>g3h)=dORrA~|7sj<ug-eGLJYlfCiQaj)}EtI&UluFqrB^Vm*9KC(=m
zS5|xsb!a0K(#r#+e1MEJ1XxP1KE5~Wahbgu*%6i-EPaDD){u`ZZ~W{0KSeXu$ceB#
zW^Es{z8V57WuLArTK>gfNh;KbDBvA!^^SJa5MZf}-Sge?{=cr#Muw!9+emqv9G?OK
zmYTzzzP}!B`lT1LB`oEvx}2P90s)q*ey88yFAnWgBW+>~dDtxv2WbefT+8?Vx>RRB
z*=#;UF7?==o@_RO0Lu+i&TgaepS@V5O9ncNwa*gwArN3`G?<&Qd&QtXdLc)`vJ<Iy
zA}0+2mez77?bnBz7Z@Tl!g7EzJ3vfIAi#1vtfSxUgI}!HMhe2R4J)^iutFffa_9Ze
z=?CZRj@Cvxgyk+$-bJPw0xb9WKUSQqePf_TMueq<Rd=wK8UieBi3^VJ7!a~ng=UaN
zU%>hoh}8)MSe{nix!Gpmd`gAfiH^Eyt8StrfdI?1tsSzQ@7ROds23S%J63MTei{NS
zo%y@Izt^$dM1?*h1Jz*D8e(+<0hX6#BZ79WHM!6W^(MVEu<8cZTtk4R>qf4&=g^`y
zZKOq5T9C4Z7=u87<@NL#%d4v8Um77fnQjMGbYQiHd}KM#1uU5KxzZ3hlV0{wmivg$
z6bP_%Pu#LPY}cBTD%6+k#)nw_kjO<Kz#>}^J!|1{7nuxsX!m-)sicD9_%!U9hI!B)
zgqU)rCa?qt$?*-8egkC#Cl<%wq4e)aaT3x`LN?&sNykV(O^VBqei^a>W8?Tbq+cgu
z2wW+pCR_!>g>V%qH9@L4qf1zE37f--13@cRYC=_!F%@&E*c5u=NN}QP#)f@0aSMB*
zpUuEwhe|NV|4DiMNsSWMxCMD_L8HVqRU@xzGz!)LG4MZg?umS1&RSn<Yd6<C7*51%
zYi#Q+gyu4tj`y0hueSENY0QzmV|{>fgl)x<_`~`j<p`k(N5TZ_Y|0TL2#&aK>m14v
zJUhpiP$4DMNO7VksgRRWe31&hC|yV$6<R07mnpZ)R3GvCE0o(6Dej;`I;fH2!e3A!
zFQhmgg~m$-lYl}Kq<AfITZ{UL-zOrsL@C~nLiVGPAV?w)6mmd{PoU5fqM$jx7=;!~
zaVc^uMSaBYPa(HcQrv{NCe#apSDmB#N}ZYcJsQr#iae|ajc_FLRuoBbDOQw<Se@|s
zzS1|1RxHPtQLbfFuqd%g%CM4h7jYNm+C>G6D}GBEzNOqnoP%6*P_Q`TJY<-M+yO(b
zL`+FYl)#iALt=<cU@jZ?hD3$q^Ra$D_JCu8$O!8fV-LtL$c(Xm8TOc@iTsqByb9uM
zLrq>qMii1N*e=H94!L(jeh4DQyPEb`P4@-&1ZfueLddZoG&Q-Fr|8Z)=EKnZI%dW?
zW+o&sG@r`+Czbh3#FdP3C1V0{B+Zwy_NA-?%!Pa}AG%H+0>QPZ?!a?HGEY^Kskm9X
zxkD0XdjH<AXnp?;<j<Z8PY?LBdH)Aj3z8D%5M^DY3@=iSqH^je!#XLBM~3mpQB+O>
zGE9)-6Ugucas=h{n4e*>h+#nDl6o@9+~3?vqmOSY<oMl;!EPq(u0i>m*V9tBfToDN
zgnSZ2Z#Cy^{UK%KNn(y&*s=?|gM>N0igl@C2S9Yd@lC8l6YC3+i?u4of32dX1#x^F
zZ5~HEiyv-gtT!|6AZY70r$Vj8%{@A_kXew@f$)#*@ddNnR}M@-H2K~rjy8qwX?`<f
zwwbYn?;+JCQy21FP&FjDq(R7T!A|Xh3X1+7P^bLYZ2;_<j3^|_AVt@f-xoix7_K5x
zETj~LlnPeK@$HnNU5Yc1A_IMFheeVeh1?o~R7k3cc<<8&_eINBYEo+mRoykIH8~}P
zyc+fn$2TJ5Mq~|+hvPTV9-HWXAZL#MmG<~miZf`F4BAF)&!jyvrR{NyX&hq<`H;sK
z-GP?qr*e+w-!XRY7zfByY5qMk?LG6Eh~rticvjM$z)nkGCG8tnyA9IzjqJ3I()JYA
zE=Ag&%1%p_wr8-tGgxPFyuIx3z3fO4=do6Ktc{3|u%8@ZKLzX=@x^eTgFhruG{2Dv
z+Q<ZhQlKP3bnp(V-eF4+jZ#yX03G;@xRc15PWEGamLBk)352d`K7kELU?q4X8?aG|
zQ`vx2Dc-|I?qQ?E@AKHt^Vlx{gD%N#7s?-^vRjdFD+-62L*GFa%Iix%H6WWVNlqup
zmEgu8&>#jUlto}S^=tRLG56fFapVgt+R6|N7^)?NU#O7){{Y9CvKhLcEh17Yz#awI
z6P$m5)${7F*%{AuX#N$;y<$}mN7MXkmU}J5Z&>3utf|=kmNkAW#oerNw-mo)P2aKR
zBL92V^t}{+U`;<r@k+VrO1XqTMy?+tHx$2*mFve!@hZ9ND!IGZ9xvzO<sat{o1Io_
zYC)p#|8JklkwTE-I(CK{5yzL(2BqQ_hQAY;3WXzZT%bKNf>16JNB9kMzhN`T6gW~+
zQM6(ssJ3u?GObRgt-wGbcq8$rP*f5-de8X1X9j>lar{QsZzKB=!{5o&gc=ik7`)Ul
z!u>?bak3X`v2`tWhbV#LQ)s6YadBX1qz9oA1wDlF=6ip9sR$(oR)Ez7*ai$H^jz}8
ziEr%vNY}?0mt#x->_3h_!ML0dF@&H*JVNOTy5{)x%(V65+0Kz|=a(R^isMsRzf=*2
z6~{lB+dgY2O?C%um_(a{+&EGv(RoGdLkWfB6B)xq#$23H9ivyr7=xIGU)~;={^_uV
zx*T6msp}~-@tD6%sV_@$Go@~p;ucEXBE_wgx>btTAoUt#CWZxTk$SBZCnI&T6sI6{
ziWH|Jb*dDfMCy~sObjDRk-AihYmmA|iffU&R*Ekm^#u`!m);LrUHe;)w@7BVWbw2f
z6Vc!J?6B1Fq@e@Zcc9)UHN_B+3^7Tes3<D0896kgK**VT&Z+*J=&_&^a5%*02^B~2
z!<V$dOWG5T$_ZM-O6EnY*Y^y~YIU%>u#qQ@?d9}f(?G_Q#Ho~TKIIQ{A(lw_7D#a$
z<=aO2i&B3``9755oyd15@)xDJ3;FJn;s#{ZfE>V0bNp3gbybQxuu%uLfTeN#3vBd4
ziaW7Urxd@$MlYqf3mbJwaVM?QN$W$olq0c$&P!S!B*Kw0j$JwHEXHuBS-aCxT*2B^
zNO2`=S1H9+tX-89-(W3okT6+Oy&@q7YuPBpH(AS@QhbZGyd}j=tYwpkC#E|79pd7A
z{+}q>?;<rsR7@S^S0}~s$S)oZ5fzhw{1T-21oAt9hKPzOMt;R220uxhl2D5Vqk@ct
zGzf)gaPpBewocC8eKUX<(;B({8o3b|(<l08`b<A$;z;IviSoKcg^P2(NvUs2)3DXZ
zYc&cN=e!B2Hz7$HR)V}rP`EhfGe~_#Ov6B_B&J+VbE|1dWZp;{G>RTqsi}-ZGy)k7
zS(8u?hnRBOG5@(!dpp>Z{Ovime2!h<3t}gXNjc*HUy_`jwJ%}4AQvJB8f$Q!wGs86
zB)3bFdxEcrvIiMasItR>HgkoCR^ygCdPItrw516I6(ADg3xry`xTRZ=LkkLoLRC+U
z+W%M47eMSuJ`@V^ur%WF8IuIYMs(Dbj9Vq+3oC*OJ(;0UmxmdCt*OhC2B8`cmNB<H
z%O!nB=am17Z(y#(9|<-3kq(QKHXRRH!I1#mFp~;_h<yHlr{7Ju{Jk+*d^+7PoerJc
zQ1@lbudAFE9;h$gyk`1O<oNl3JwD(*qCbeCU1DetDGraJBSaiayTpp^a2Y@*FWdyc
z<l+31_l&@Oz$Ri3F|=L`{V|f|_!!zZh7J<jV`;ruY5Qur-)cHoY)_({l4xJRO3lRr
zSkjWyEq9jDmuktt4M@KMnMhrs{#7Y{gmoU_Ua$m?@5FYU*a<r3_-fj)nl=-01?ySC
z`iidXChKyO^#qgr$?QR{t<u_s2>TFYa7YXqVvqA~iMNieAPrVtRztxA;Rb~4t{lwe
zU=zrYd!iGs0vrg@iBfYf0*(RhMTq)^+Yo5rNKr#kfsLR_k~9j(J>KwPfD^~t^|CXC
zM4glmlx}(=@}|ADMbwAxBSAj2j}#gESc#lN^oN*<Xdh*`kFt2JsmyYGE@hod^@a*+
z4|dL_yrj5KE;T^J`BXr@w4zi%1r$he8x_zdttdUD0v<~7P86_HT2b1C0(MDp19EIY
z?owlLyeh>9u*U)1SJXiz<|?s1D2bdim}|uPP?#k-F;?8ix=<S6$W4#p0oH{}AdciJ
ziZ-kZ$qPqn&x(gw7s}Ec*)IBzu$gFlkFnx0)=kyiGjaS;+UTfs7ZlM(MN)j6HaafF
z<+M?`6rZMzPD^nWZB!-2XK14{Qd~`&SJQT)v7e_6&P&76OSI7?DgSF)@tW3!3pk*}
zv9G1MwX~7gQ4#G{B+cQE({9J5xR_QKi@QvzxxRz#ob)jC>g-tyy8q{i?Js{9l|2oy
zzWZM}_=X!x2k6KO9`|ftBP$QszMJYFJc~K8{Lp`S=gr#JxX7995rjXD@kwKb!-=qY
zUde~(OVb9ChFnUKOBsOiZ>iN=v)beSLDF!ZQJ-f%)(N1(O%5C3#u7X~$M0duzY`$~
zAeCJ9vpjYN;H@pLr-q!|`;he5$(nSs4iGzR?V3I<R@b6|IQ|62BZ2XU$@Spy1SU$v
zrx}ye(gXQ4<9C`F3ddRUxe-~H6IQsBl&O_9Xl1Rz4RZWb*5E0@n#)7DTqWL0xKb6@
z(0~;U*cXauVEW`J5iVBYDA^Gmm+<3+&U0ijyRhdj{5f>F<EwM?N~2<}Ny83oy#r5z
z<P>%x=}NeOg|6VRCi$Ci`zo0X>(sz{!WCA}Dr}yzPOyp|?DCZL2AnbX<fzTh4HuB1
z+Zn5N#tkH&u_9?hjjI2!|1yicAvZ0DmcL2={|YR7$q8oMyo@js-nCNZt(3hO-nCM$
zt(1?5Q;~TpvWHlP<e11c75Rwx0y4jV?8Wfz0&=~8d;sS#Bd$FkF~^@AK(&-rt+eA+
z$ZD0eqatKognU1K=o>Jua>w)(vXofHB9?Iiv2x^g*x<3K`y;nU{ozx!$CpSc3;Sf@
za5%bPcPo^dyKPYq!i66&kkT$=AIG>s*u{~efa^KN2gFSZ8*KklHb{&U&$GJcS!3~#
zZe?v+S!dBY*U9bH$z36CJo;$aQRBqF5B%$e+Csl0smm|5k05r?$?A8qAEWm_W;NG@
zDyw^{)_GVj51WHS>A8K;D-v<hBah`{EP_&q9h6bz-$+sKmDHF@YP^WMD87q|6!rd=
z8uOMKFX9};=b%VX9~l^p$wT7-pBQMXcW2GZb!0|qv^q`n#iWGKa=Te05O(orS`l}u
zCgc-!K0u}qP#|R1#b+BP=LB}#{nw=kPaVic7T!5P&<rE@B0Kz_*A1cprC^&B>@Hq!
zreT{j><(ADJ+4A;2et%P(c@EWc42q$DLt0zkcGX$QhP3%T(WULF)*mbdbQY6T6{81
zp&iBZ?f`9YKw87jrVX;C_%Lm7Sc;F(21lg0kTxik;w!ZN71~<te<P#6k+By2^loPO
zZe{{pv-J2Qbvt7LIs?Ne5)|HefS|yW^|;%AUUonOJn10j(2Z5y_+x2Gsd?A|d$ZhW
zMW6l!qkjJ{S9f~fxNmgA6mg`7|M%HY@p9Ir!wXLBA_ldMa@a-%!5q%CrLElY!0(ba
z&A(x}H`0pjTb6q(#oa8|EyeFx?wu6BXSw%M{DI{@NbyQJw^FRI(tM1Zi;?13ITtI%
ztK{4&DUOqKaU!lZJLCNNr%66!)1IZQ&PsjidCKa%6u+me-cu519fPc5q&~L*Srv%R
znp7r{)iDvns~VPMjsHBU0nw37i@Dv{RBF(pv+!imp!Z|1{dk~g(52Y56!(WZ1V^$^
z#cgZ|X(@P8Vkg3j92ltPeok5Yh0CM=IX?Q<QURhrT7~+qLII-xDMI~<P$2l5ivxTA
z$VXTt|LfkOF6t|jSBrP_!~-t0%f}n8OQ!>0v<{hPMgE-3SS2%kVDDVIx+u!crs`*s
zDz+l?R^$TJZ&FdliWF=lbs~y1@?uW&s)r+QWBj+{;Zi4}*oBRxGUu|enJDvgx_3I=
z7t-%b_sYLL9(3hCCRe<4_&Rzt>~5N0PlvCU;tlk$4Rjc^U(vrkrn2}_6In_+rISv1
zf-)Ms+HdAhn*J+^>yA_U$0=J;8P6#FXOyj|j7+4TiEKd`q_Bc?uOmxHj3DVG>J}bZ
zL2?PNt4M?JrV3o~_4sQh>>{(P#0~u4msh~RC^b*4V6r4Ar42f1PpB`E-Omi#CXGEa
znBf`B4AG~pmm95@d&8l9!*o;qj$GZdJyKwXq%)&Hp|B`2qVRA_9I=rs84Z9{5+g?r
zw^1PMB970X?J{V8kTWTy(ET&Ts!z*+iy>XvCe*)fvhe1GMEJr3GI7QQ*rfowfj&6$
zm|)sE@qz(Nk%ZO4voi6+daS6&=Af@TCcnFHyr=6)obf@%;2>iH!XhUdV~{Okc#lSU
z6dt3AJw8H;N5~k?+`Hb)3%Raw66x_et96~#hcl)p{L^|&uqO1s?_QMSYk4q(=xHBg
zypQpKp&=U~wj?}igQ4Ix8)*>Uuz?M=kBXnnO$pleKliKfSdJ|2Empmio_@lUIXL}b
z?-Owe@8w_(9Df-rF5|vX@cVmvQK?t1t;4@=J00q__El>m`u@*TV|e;Uv?4tC1FgLL
zscMDM+cMq%`3xW}BQp@56@sv#OiT8U@QM%`h-<)HIyQxU!jV@xiWY1L&vv@r9?mv?
zIBy*>sAkH%neq}HW-}GoOht$|8JQ;|FVTS}qrhYoA>tZjUW2^A_i%g-3amj9fM4a>
z9y5tA>PI#~3ieFFJRB<EYkD@pKaU*2vY>{htJJ({6xY;%yc&?d=*6xgud7nrg!G$4
zU*1z>vUr4hi$$hR+^-W$O08YkqYL+kne@c}R%z0AG}W|2HSH<3pQoM9(~@HGZq{tK
z_$-3dAKBmvHXK5Ro(DYsRcwfeZ?Xe#vSFgXXkz;|v4a7>eVbY8bu0M=5mg6LbRb*s
zI8Zesq7t5PLO#DBdhU!+l0b!KYquIa{W*$)@&sHo$Uvw$nw(Vzxhfo`GLY`V1)nd>
z>Pg1NLPjQ&jT3)P6n{<<e~uP^Mv6b9#Gm8EpA*ENW5l16KmJ6#SNs#tD<_Q-bv~IK
ZH8x`6xbc%FM2;RK`{W~wEJFPGe*nr5Djxs<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.A4DBB4F8B2C78E56.idx b/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.A4DBB4F8B2C78E56.idx
new file mode 100755
index 0000000000000000000000000000000000000000..af1ae5bec6418f67d98b7b9ed6885ad5d07ba9dd
GIT binary patch
literal 1656
zcmWIYbaN|UV_<Mj@vO*AElFfyU|<jg;^LB`%qC_A1|BX3hKf07Lv!<Qn+aU|uX(O#
z_L*$nAI^_%t=_0+xJ&G(N5B(LC6|d(o;T0c)%H)5Rrl&jj5K)DeP(z2_g{A|N;<GT
z=sIMv(4ivo?5U*hAAM3jWd5^`zY`Nz8~I`N(+Z=2nhGWT>&@J!zyAE)m{epKU{ktr
z*UxDp=JUnE@7FIr{M}xx=I__Clls?J-<%;br#^1&&tIpXZh!jrdUjo<@2u$!=Ra5S
zuZegnl@co8#j-bX?d?U+4{;UCmn2*=UCXUxej@ve81HH8T#?IRv-}=vzlqP!EP2<u
z-@bB(?fD0O63+3v?k?ypaavu%==LPTNoa4m-QT<m_m6wsHsaj<qI1vgLqFYAlygg4
z8kQ~CTsmR$l3bQu?PeM2)mv-6zkAH?(fPKQO)4XhQ+lh7!?~an?#KFELIk#1o#GM^
zTV%{`&K2Ds<kBSSrM#yjMVxQJ>lK_As;=+&@qn*<O7a!nNv<9HCSUwn(6Yj6)v}Au
zl1@2%ikEecr+ZG@#+*2JZ+?vUg74K%J=-{XKO8yk`BS`#UHakHfG8m?Pmz6Zgk;-W
zq-2>)<+ys?9A9k}bDR8-VY$%a?(e#CU#4j-p0=dr!NglzSMo0Ke!XJh-gTm0bI&O{
zC~mRTvSIro{b`A)>QYVKdX}W;x>r{&P^{9}dgMjLoy7}_=KW|t$?C$M8Tvh{>-U;`
z{fW13-rBXcb?pbU(Cf=<k2Glrd|Ib*!CBm;(|rN=lnkl##L)X6lg;Pf>A10_O>J3t
zUf`7tote==lle-Q&YLLm;XL2kC4xm&@1F)R@vgo9`1g~S0(<rLW~_d)XmiM!f8UQx
zzja!!bxn{;(Zp>Z)}7a2mim*lch>v;+`rUT@#U*kdhTZ1S-45<#SGggPS;H4PMLGD
zw{P=io*1^=H5yzyRi4E>=M24dK0>^vibYMlr|ihIoEk$b>5G%S{+t!c-=DVOblmrO
zviGMwQ1!NcZ&R-Ot@^}P-oOCn$eU$rzN}m9`#h87Sl9FDquW=U*82DOM83e}w^I_i
z%%hb4zf|>`UXi(d()6!>%Yx5-PO_=&Fr1V-+i;fuv$;7BreFN|qUY(YO<voB(mrPg
zKAv<<^H1#x_USG+ZmXIXNxn6&-@5+MpTEJ&6(@f!vC8>Xa?5>&U;p=Im)9&fw7lo<
zX?tz8*X&o%@2m@XQd|-u=>0#{?C{BIn{xdCLzPFmpa0db;aKbZjX5)L>7wGw+$4Em
zjy$=0)zU4>Cy($j_y;fqbF*@XfC-=s1QeyF6*Dq1FmU}^aIe|hG?<wuSlC+F4X!O%
z#99Q2Cu%JU=kYgO4ema*I)I5MSc*f+8KFf{SuqaAGbs%}_dhp%KG0k~Yd(9JFi)_I
zi;M??r=qT+4dXeszhU_KrQ{Y93yQfc!N}%@_z9`Mm}tYw$P>)N$-@uR8?kBY1=|^3
z?}66wTJs`VC+#AQWSz3QGTgfSr>cMV6*kQUYT>u$N75qeB8#L&Rb3UXr8z;$ZMC6s
z8PMkf)&fXcgsg?&ymqFZ6^kA(6#=?X&P6T)=E4ONRzG3iDftxSa0MPKnDC(xyO;6u
z*N+2@lwgrif(i2kE3hcQc`#8%E^s9F9{Rmt%|mO2rv|KF*!jh~K}L%5$ir0qZ2Yv6
z+r<6~P*_b}4eph{0%4o(<{x<m6c)4=w1cT;C`!#q1f_}MqU2&_5F-dgKtLdj!N45+
zci-kcXP4yIFa-uNB<H5&WB}#7(|nm(SlQS)IJtRv1^D?q1znxh1i3t&gv5nKBqS{?
zt!-@W>>cz(wXKxRG~}^~ib>1JDkv(csH$sf>F6668k?HSnMmmxxrmCndxbE$`8j$p
G`TzjNCEA_<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.315EC2A3A6D6E638.idx b/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.315EC2A3A6D6E638.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c812c861b6239ba023eb6c71d25311943c3759e0
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hX0%l3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJa3+>tL>jAtM1j6m}&5)`^@h4^6z(~=Q5@+
ztkdWb=4zRKb&XD${5cmk`RDZ?j-Gvb^*QVQs6KV(qrcTZ71{3C5HtV3QGkygTi0^E
zF!T7ykDfh#^y>Eek{{omH?7})NAKs;DpTpnO`rW&o5t7g`?G7$-=een>pedUeaN(X
z-zd5|rg;sQQ|AKnHKoyCVj>dXHS_uGT2|`#gsrmo-lD|H<-JL_ye@0jwElgq+k0;B
zhwtkA+n-C`;Y@zvy<K|CM8AxwdIvgWd_5L@Kezlo@2}qxnbB&A;<l#m*A?s&Jn0cF
ztGGb(MULzfl}zczcMExM-Hg+{zxVeCIiZbvFFG*mY9%h7#(pA_>*t9Z6FayLg}FyI
z9&##q_P{Oibzqmr2B*vx@riE^F_y@RGDf|Z5|_X8u+PHnu0mwX3v0=_{O%o_7UxLW
zPEkDaOu*=DQH|TCEwUW1-_$<d#jy9jaMBjnjF#nfbEecUF5e&+$Guc_;Ut%br}efz
zc35#yG2r3C3mGCyqOucBB>4)w40l!Bf8?02X{_a`#L|Cdoo^~bY?M3e+tRfGuS)nB
zxfY%cd}z4f{si^49aGe#{&5^R7kw(&(fO@n*2=|?3;WwnRlc(K(NnPUSiS3d(4F+s
z>Xj?!u3dXO>IPSSaGC8}Wigk=3FWSha=R3~Y?xvd7G=(94Xr)ef3>DaIq9<0tQm1T
zT>}rREz9+EwR$o&%Pr`@eyc5$+)ljQJKdSfWJ_KD`;(Uyt75AzhMjax*PQwH{mS?(
z{c4d-UK3Ba-a52<ULc#vkE2zl_v^*K_^!3uGxLe+Td60<PR?4OTz#T-O{P`noQu7E
zn>X{su;s4N;M%$5+|F~ttG3MFsQ2RwhmW4b+Ya5EKaxt!7O7ADW$s=3Tems(+ph0#
zer#jWkGxg+d}sE?`_AifJDRvooZ6PWv|KjV%3s+a%;vt<KG)c=kMZtx%`<YA^91j>
zDD;zGzi`(P@7+`PtSL8|n-$li&addUDc9KT@yR`tdH+~h$<N&w6CL(4>{L#jZdS}C
zv)1~)FApaD`kHpfO>gnrXZEkPQZ~)=Yj2(BSH1u8%a8X;W{FODJ7weiKg~&->H_}m
z*xw&i-Pfm;b+-OwnniAZ`rB4Ep40!Q?fEaB?WjGWmi^_ng-Hy>mAOgsz}$H3tl^>L
zDH~IH82kemg1K3_y}$%e1_Fvw(~5sEGB9xcT5zw~+ccP&Cs^28*bS~NSj1Wci6?3;
z3g-!~@02QC_v$cEi-5I&9YTwswIC8t$XW=_Gbs%}_dhp%K2QsvHJ?2~i;T03JA$X8
zuA&X&IkvxH`1z&e786Ub6o-@(OqeBDQCTq-#tZQiQhzbghLw>gn1_>xA0`~JY3l{s
z8D8&!*6~{NB3UQxERAHHvbr+dy8NfAfA<wO%>`=Vx8_IEBI_)Rq(xO-6|SY7sb|Ha
z$4f<kmdiQIg~KdgFk$r*_MMVXL2gpuv4RO73bA_`FMs_w6Hl-Ni-ZzPm?v0)MFGx(
zi86A5W2^Vj?*(ffS}R<<ckP3og-R<>FLGG$poGO!1J*C>{NmjpdqsKVVMhII{IrtW
z#Qq9USWR8c114OQnv)1h3B^Up#mXQ?5QqQ+=HNwJpVm4i$wn~+1~CM}gc%r;b5n9M
zfL!l1UuG6oHg*n9ZXR9%em+k@S7$XrE>9;RabXb&NefGB8(TYj2R%`3D`hhcd2FI$
o(lW9Nib^V~>Y7?Q`UZx^rsi@cQo2SiqGIk|Axv(5jvkCY0I5#ArT_o{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.AA168C235A47B299.idx b/.cache/clangd/index/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.AA168C235A47B299.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e0b46eafc26aa6fa8acf9d997099244aa10a6656
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hX0%l3>9<EhUVtqHWRq^U-Mkg
zY@KZ0AI^_%t=_0+xJ&G(N5B(LC6|d(o<CFf*SVkW6IMOC!9?Pl>$AIucUQmbJ;z|g
z5H50zll9Q)tSGVD{pm{0{+sXr{PXSAuhVnl^^Z+wKH9GSDamxlhKRZQ<u&y5*}9hN
zP1}Bd(xY#e|NQ!$U-IMM@j2`F-P8SXHOo?Va?|JYr$ysycmCS8<9F5J^?4qjg+5f8
zy>3)p9n-vq%c*n0^);o^Ut%H>%N*I&VlVGfC~Ds0mVa@>p382Axl3P)>^t=D>om9L
zaX;>!Jbbs3?fW5#Uzgsp<t`D9TFPyp<Ssnv#oo^^%X{m@wR2Ca<SbtP?5_AZb0x{l
z)yga;taH~k&r!)$zF@n>(fao4r@PDL?H!c1nM-9Xa8=E@@`=IHTTx!pS}4%*#tRMY
zfQXKBGIum`Y`p|I!!&w7C{2pBXjr#c$>DYE<44WbcZ80wkXGpys8wJ5`5{MO*%aT!
z?M5v(WSo{o_MP{bwv9P)?p@P$+!ub^D!N6txK+G5?)g)^iaq<`)_^D>El-hsZ-T7(
zTBKx|MCG`8-5ghK6?2>XkYTyV;_lsI{8ipF13iTlE0&byYB9gKWTlm{KD#w&c`-)-
zM{XZiKjYt7HkqzVOfG(KS17v|G>=Ovi$Bm?R<^&|B;;K1|A(KL7Bz2KRjqrpHf;O!
zS+{Q9+SS|I`@t;r`enN$M-h(7aFG`axfdN#zQC$9>sVyO)$l(NXVuM>%+@aKd^uHG
z#B@#at!W`koa0uesj&Xp?5v&HGUr~k<q`+y*WZ84vyokNeW`fv)R>ZKlWzW>_iO35
zP40@TMLJ`a=<azhnd;CSzh>IzzvX`SudF&e+cWOcI>)>@YdEbowa1*8wQTYU$-6Hs
zY)>CZy>P5+(FE^Tb6(pl^!ly&y2-!)z*E-6vKv*U<9p7zeOVIpf7#0X{b?J{#(keB
zdw<#kRd4I}Hs!kCs!wd?4Gdt6yjiy9%euwB&ofz$bv>6pYQEyM*1y*$@&zX6PEF*x
z9;NjEp{n2XicIrK)4zr-3qITZ`8&U&+ooJ&x5p><Oy>P#WF<d$V@!0|%dk^9Wy)DG
z7tC7g`@TGw^y_Qd9XGwjZ=cz})=Jql&!??*o=?^O$tOSFOPM4(<!#Hx`G1;|Hq{0E
z-LbzvsJgFDE9-3i$ux`H{`9x4Y&@s`PuuffJlj!w!e8E-ixx&P6j$aZ$pdp^Pu+|=
z%Pk-M<zes-U<l@B<@N#-Kp6-qN=+;N!N|bC^=rYsW^dDAW}aYSYhgFIwqOx!5hR|d
zwJ4ltQW}2le{TAGpcX!BK6``~8D|-H1W!d>MH|L*Y=6V>^GnGsCYE3+4k;&?FiWta
zvSKWZ7vd+R{$ipHD<e-Z4<`>lOgLiG)(f^Xyxs$?<F)2RvQFAr8p%3kb!E79`A=2<
z?kjAX3)I4I&5xu-)>#%wi>kURTuVDs&x%Elmx=%_mvfd2hglAz8M(m0*L&#qf;A7V
z6&6fb{e*p|<WrDS6?m**st$$Ny^NQ?ew>LXSb{}D2`0=FtiYlG=RGxG{ld;K-VM?#
z$|Da`^j$c3s_LsPB0#;!LBxX+L_Zrpt>iYbzXH^&rmp4z)B9H-Y}4KRBhP^T6|fet
zg9-Bl3t9^z@r0~};Jl*LoJ3GcC@xAaRt7PGKm-^t2gf{l_-*#uBoC&*AcjDgFatw!
zZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$rrEI1l
sk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEggvrg%(Sy+k0Qz^%`~Uy|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.FA613A1A341E3811.idx b/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp.FA613A1A341E3811.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c86645709e40d3707273de0d0968c1649ee5e16
GIT binary patch
literal 1822
zcmWIYbaNA9XJBwn@vO*AElFfyU|<jg;^LB`%yMQ1hP|8&3>9;ZhThJ*Z6R>=zew^i
z-&eW3KNJh!Ub~@YxJ&nwN5B&gC6|d&lRl;HuWO&C-#J-tL*~ZcO2y9#($3v^!K1)j
z;F>rmKw-}swbYG23lxohSp1!P{@uF#eP@20t+B8Q-czr1etNq2v(u-yr*~bvI`h%W
z?xT0NFWK?iV)xH~oqYHI&-nQC`ftnqae8XbK?VORH$FM7fBpWm@A3IRzv^bjG1>Vq
zx80C1HG4v+qlVy@uB~~!^SHIon7uf#Bs)snC2zv@7rIi@%5P~bTWd1+Q234g_b$D-
z!@b|Oa>xEN4;&?&<9FR%&|BiPx`ff~Nrscq-g3LY>n`LcFU_6V;``;wp7ujO-Bgrw
zOIsS2E!bQ-Ve*n(mR;>;8R^wqYd$|aT<vkB%q?9iBal;i>mBEFK_}dg^|^!yT(dgG
zB_g)SxZ0d6x<AOJNz_YuPe+RQw*{|Pa9pUmzN4a~UHXwkX>*F;qW38!b}|a5`Zi^L
z*~z15_IpC+w8zGimcDg}NZ;Q3n(0fOf5%0!g@W^b6$brly?4;7LU+ov7B8Vz{oPmk
zY!|w<JG8btEe@QJwYK$QN<~|6*Tti`-PL=So>`&NDPR+{+iWUhNv5pmmg`%cR$jJb
z(P%MaZ#~{n=Q?kTQqZi7AIt*Q`I`Q$yj$B>EbZfKx0TX*zVg4puBI0XYhU?!7tX)>
zcImRRZQJs46Lt6a-u$|SKSF@j>7FQS)r%&Lj|?S^9m@*E!oG^yhkQQS+u@s?kyQ0m
z%9VHZjBTo-ne)R}NqRE=e{;c0N^yC4+&9e&7k<g^Q}%CvQToy~T2*}R(v-6O%j;Io
zeG_;>%T-A{lUu&ZP4dD6@zY03zR$gAzBTKA+N9}QvL{HZXE*V^Q54^_?4?SPo4J*r
zo>771mW_)vo=mcOZsot?Qr)3l4=yVk>`L0Aye;MM>Vroc=7qevuk!Wnb{?~vukUR>
zd|PNwO6K#q&(|HT<KMYWkxNTK`*z>dd!<`m+Vma>_?WczU+IbYoX5;g8MJEaPPBKo
zT)1aT_Q}&#%9mF@t2C66-OHD9JFUlUVx_uth4~Jdm<=(}VK2i@<y2{B&AI%nrM{!^
zi>vQX))F?o#ZRBEFBdY3>_7gzRi?kX|MAC%_bTRyPPsK><NQC*6F1qds=Kq_-P6{c
zUG(e2{VBzEzPlr<7Ct-P_1|mR|LY4b2c*|C2d-@6VJNQ5O_B%ZL)MygtAcG7WbiQf
z2QUP4vvRwD37`xF6s4vWZ(>5IW8?zM_a6GaV9i5o1+HHU?lpUx1~c;ni#mvU!%Yho
za}Yz~i93kHdAiG1SH#&r(PrccR`XMff@v`+4L|okH+?<}PcWYYp9@TwCs@W$CJ4b(
z_ESdUY58fvc~%9hc1~(|_5*0HuAgoUOp9at8-|}>N^UW;1WPGP`ND)*f)$k&k$4))
z8gO2SpOE^Ci8ic2lXy6J_+eTiHf_CNJHzWe(1pAXyhtvT_LGLYf+twXPYH>q>8A<j
z#jM<r_d<Q&EMQ<EJB0_uDLlc*PRW#tzw2>{9pou_KY6&P@}H{y-B;K&7w8lL2LU9f
z$oa{^t>Xz+^;1RS>G<iudCdt@ZmSKA%YaT1bPz-`K*&J|i6`P90_PnHv3nUWfBiU6
zizJJr5-e1Bf)!X4ka+4W>Tupu1J*C>{Nml9fE4AChiUn;;p3+0#*3P)Ji+`9{P2Xq
z6D;f?jKq`mlZ}LF;R#mpQ$gZs`)R{@zg>2PGGCj@46;Rm#~!AoC^aV$l!l9ol8cpr
zG;{DFV~$;%YdLwCf&v+WAjum<1|pcrxhXjrK(2S1FEa}(8#@OlHxI7>KcA<dtFxLQ
zm#34ExUh(Xq=lumjjf%%gPy3im9m+JJT_4=X&G4sMI{wgbxkcDeFH;dQ*${JDP1EM
QQ89O~5GFT2M-N6H0N^3rMgRZ+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.F48B669764788B12.idx b/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp.F48B669764788B12.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6cf805a8fb537452564e05401ebc83119f650f8a
GIT binary patch
literal 1820
zcmWIYbaNA7XJBwn@vO*AElFfyU|<jg;^LB`%ra&MhMk-Y3>9;ZhUVtqHWRq^U-Mkg
z^fTG4Kb#-kTD?)taF^Imj|(cQf{Q%fdi+e?zt8=&e&=M~jG&zOlzX;&c0a%Of=7Y*
zgKOg60EK|C$D3}{>=2Nwne*r1^78BE^OyfgGn1LCdCz`gdiQ4T%}1Z=r*~bvI`h%W
z?w@zBFWK?iV(-sy*0THlT0ZU8|888jcSe8vk_Ue-*Blg|zkdI@cl+;rdL4af4@=D~
zpQ?m&TK6-;SQok&XkYua)lF~ChT92DllNuT3V5%sTzfI|^t8DfyCl_fmw(ZI6Q7@X
zR){svU+$g0afQIOOVjhrZ@F4!F85&!)7i@F>GEYpUAWZtjYjKy4yh@bzF+v^Mn_7@
zoKAr)x-Z_?=y)3GG2OkaC-J^&Yt8p}hpRn~l)0r#Wo#AFliJz7XYqvo9^1|+$B5Y)
z+5x)+=HA^MxXV@7RV5-bleNpEERp|LUl)V!uVcp-TklouyJ#{?$LWZlTmM6bz_Ka6
z$>&lzHr{Oa*(@`EqE0OPp~~&KHR21tS39jKl#*(^(m(lAey>~@Q*VHqkd~)NMA=H&
zYNuY#MlSx=i(!tdwhCUG@KDlE(@lQ++Iz2MQ{7ZOTOLfjwS6T|g4gqv3-+!P^_qK5
z(Lu4tRwIDDSNl_n$V4L-W`2h^7Xs&TNoDZ|ddtf4Rhxv!t^9AWtLcT}+E;$wh4Zhz
zUAnAnU0L1R1l<cu-&EaV*XeLNyh)p@_X2C+E9MKVO0$kdT3ikP6L41Dd{JJMxA0}{
z>rq!WbY(^dP5f55bk;<XFY$BBCb^xMJ$HI7m&ukLhs&S56u3A=<CfOW7c#1wf7^cb
zew%W|DNIvnr$_XU3&y7!*!D-9QQ!T3%D$DMzs)D_SrXg4^y?<I7YF2~IbAcE>vC?^
zu_GT{+3cFl!bDnQf}Yp3w`esdUt{&@Kk%aKV&9HQyzAM-nX7_U)<3?o{l3wSl=ao;
z`SQITPF^}!{bb+Ozj8&7okT($!nVn3Uo7A9(x&%7z{jMubrw&en>O}t@_9LR;w*c2
z%Y}PNvQM6_Qog+M+0lZM5~~Thvkhmtw8fcyc>SQL;NY}l8&7QsN}C)V_*zoe@5gij
z_xLN5w{MpG(Gxa(|1I|uDRZAc4?XkwUgA9YcTX$B0+K9~^Z(pV++?@v-_FYZqN+YV
ztyhQYPd=-eQr7&%>103a|Cv7jmuoKXJhzY8X<|!Dab<3jJTMo=l`WoedS7n?4}*UI
zLohchw=<Xk%0NI-YFhCoCI$wsUkmOvdz%I`^8|}Jh<d}d1&cX|A@Rf=#Nj+{ZJ*{t
zFRs@CwMZ#T`5?3?Dk~!KG?X>qJi+yyQl;x&9R_L<bP#k#Xc2M{LgI-yh`@Qe%T`y!
z*+0={<Ox>uRf~jaF)0l{_dhp%J_}DUp97x@OqeHF##bf~!Bh5CM&fDtYQcF{1*>*W
zYIyboXs)iWZZu4bWBVJ1pI=IDF|(j}iX|A?Q!K&Ao(l03Qhzbgh81WM4<`>l%%q4-
zTQAto@Olq)A+G~3k_)AMrQxpN30Cq|LgH!qYQlLigBZEMQPz9t_kuMKtraq*;_rG~
zVh1@~-d7&(@cgH$fA<wO%>`N~;2?lxot&>6+(@2aRbN#ko{p~$oOdY1?q$6E_2a;}
zkz|ong4xa!tiYmx#8YQchx48quzq3Z7w-o7T$D#1rsd0qkDH<!FKV*#1oJ!a!_x~-
zu&{$L5>M7wHUg%FCs@T-1&OEas}1M<cG(rmd~GT-$QA`2dzhA@)SN_68ZItME>;H8
z%)yubGd#X#KKDOUP#{APBzc3#Km;>6Hzg+n$n{S1WoBVzW9Q)H=HV6K=kpYFbygGP
z@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CYrEBCO
QD(3DL!sO=X=)vd%00`6!QUCw|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.B296FF7D04FDFC90.idx b/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp.B296FF7D04FDFC90.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7f0c76b1a2074a515f4b979d1ed994ac455a2e5
GIT binary patch
literal 1820
zcmWIYbaNA7XJBwn@vO*AElFfyU|<jg;^LB`%ra&MhMk-Y3>9;ZhThJTHW#@2pVMti
z(sf^!7YD<ZxMk#UXENCfiCFIT)|r0ilhVn*zxcN9y0Ix*q3puGH5SDd#_4$>i~$Tg
zHktTo2*h1`yy{MkjWBD?tUu?LZx6p7zwA$y*&ds~nhK@!V(Q{)-Jid!cb)DQJ{m0k
z`F4I};qT0<zkkiGYW^pm5|4kM_9x3wp5OD}&*wjz^!LTo&c0i__hGpCl1k1WHZ#8*
zm~(ZVYZ&`NmkZk0cA8e{tuf$tWLJy5y;ETtTV?aTMTw`E^d{Z%nk=&K(7&(K-1PT;
z;CJUQf6jV`JNbq8_UJ7W{W7NN9qdq@HNo}$-1GajzwmFow(Zm*wW_RpwJ&S5H%;7T
z<>U~2>84f0q$Rm5yV}h%(yO=De1G?t{lt-NZt7Am0yw3&-f=t^bi(~upG%0qHLFun
zBGZ&+&XsPR>#nJEkZY-6oXaL|>kFz{OeOcW>dD=C*k<8&S1GdP#pIrI?Y1sA1I^aj
zO;tQnCT?`TsK#y6CeD`9ZwpH`9o~QMkjNE`dbm>F{b~FQ*)=WKoiaUFs09AddRq2|
zW2+U5mdzBY-lmvr-5#~W!gGRp^z2G@T#i1~<28x%pz_w&MUoET*8&}@W7Vfv&2w%D
z%GexU!Sf>giKn*G($2Q~%!ZZTm0B4qZC!NBODx`846ye9+kTq6gYVO-ysaYNBk#s#
zWo_y$&F0R2n0s#B^1a8KGz6!u)Vb))wdj!T1SY|=k9F3((%!fJ*<@+K+vhGE`Ets%
zi!IDB-&^mJ-uf8hNlgD{UhtAqT)sW7O!LBpU%LBL{o7xZK6MRG7e5(fbo+kt{wrbM
zG*ad&Pwd{)`rURhQ^tqWM{73y^?NxxdTX`uN$-265mM^XP2b)mYNj|}_I%QlS1L1K
z^a0n4B;UXjs-@>t{a0M7I}`Nbr{In~30su6B^~ZQc(Y+%$gBGn-``$mGCTVE-sZoz
zh4!RWKA-!1-Qha^o!b<-v=mnFI->PnDA($|vO&n6nAP%pKc^++EanN`aY^WB@f6`x
z+PkOi$ucvVFBLa$zrBU*iFMi%zDl3{wtYDLpr+v1v|F2^%7ervOJ97@Bc}exyQ5kE
z)ydm8d14$_pN@a4EV4<m*Kf6j)K@<D#eer+T%n=4Q7}3G&+WubmMi}4tn4qU>f_6L
zb*TR2vyv&>n!h-m<Y)aq)93$k&E=hX|JaOUTWX3cbCcwOxiDRMPtc<mao2bl`~w(*
zxmmfL!30nS0*X@8iZ?MaFmU}^aIe|hG?<wuSkyt(8?G%_%s~u^C+;8)=jkq6T@h#h
zM4OQ(Sj|^05~jtZH2mEE-1PY@Ji&Ypd@e9yo?sbYnLq?j*;g5fr{${!=UEl3+BvD=
z*$<$(y1u&6FfESlZy0`lDY?bW5-g=G<pUFD3072AMB-^EYruISenRRmCfcw9P2%C?
z;fHC7*tGS6?F_H?Ko{~l@FKZT+E*Ix3Z7skUnL}-rmrTP2Q!F~3mllehkh?u^Uzu$
zQ!4(h$0c@<!{vSD;SSG#s`__dVbffobpj3oNY=^u%E67~30C!0MdIoB>cDwB77Bfg
zNqC?Dj2mR1^Pu>gCm7l1heGUL#>-zn4%8yaBB=y(Fi)@oivki)okbnaduqV?g`Ho#
z8x&ZgJn}FtUp9Q)6y11Hla(i!-+><<`8>hG4#G%0Szp-*m=>O36<-x3p0=+xocG&h
zS19wfsmvf-6nN}mTK)=zZMvI(<QdStf)0YtFkzlxAqOEOo`{18oL7{ZlL$(~#YM@*
z%0QYqc+tUWHOH<MUS<jkWC(&JZx9)XU?%6L<YWN3-f6zfEUawo9Gu)dyaN1uo`SB<
zYJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*ZR8-Y9wRH3i42@0A<xHe>
Sja)><+`U4W-25Cp7<~ZkU;+66

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp.DDED95C8C9D22159.idx b/.cache/clangd/index/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp.DDED95C8C9D22159.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d5f65c9d14407caa0e95bc6287afac69eefc504b
GIT binary patch
literal 1820
zcmWIYbaNA7XJBwn@vO*AElFfyU|<jg;^LB`%ra&MhP|8&3>9;ZhThJTHW#@2pVMti
z*7Z`B7YD<ZxMk#UXENCfiJZKvI?X$;eTmBV*Vb2FC7sS|uyFjn>Cu^oXXfk-WOZPC
zaQaf!LWc!wj)(57tXJduVeqrr|NXjrJKvvYb1Etq?Wj^aFQzV@*8Mr&Ph`5dde?Kk
zXXWvm9=$vK<=N}^B|m;WpQ3Mf|NM`yNp^CRnm+e`W{t1e_0MSc|C-b5^F2RHeJC_~
z*=V`Cc6Fp|i_pO}5zntn?JPKC$0j`EOYZaLh^VvDy?)W#e2!_Q&DrAl()`B$yO(CE
za@E*W?XcH>;27Z@zw2(6_{$}8jZ`BxUNK7%yqNqiHnXN@YhkFEnc&Jw{-wvmQj{P2
zv@FQ=On;u}IjK<a#k{hE71vik-Ch2Ey@K#<sks>oE_?0Bc*6Io&&kg4n@SdIV(y|f
zjxU2$b{8)$X^-mi*l_6*m&n97hxop9b}`8Qx^}$Tdaq*N#VxaRoQ|k0KE9z{!}`=R
z)8$4jH{Llti|o_)2-9mUxM#b~j`hV~Il=5XZD9&uAFucmy|{g<gYuM#EnY&bVVc{M
z-B*|@20UD7a!c#PR<lD=Mm7gEkG^=Eyu7^1eP*Dikz&Q7Z?SsJGfq`%WvqYQ5_G(n
zBY-RUJJ%J*$6Yp(u1k0>9<UV<wmKE;=>1kPYv<y_iT!Q2l)uXB=_^#7SiS3N&>iv8
z>a{CRnx0!PyM}ewoLK+re%8ROh-I6uaA<aTy>Arh$dF5a7!v=n)^z^ej$Lnce0|o(
zWM4d>^(^<K>#e6#v)zI&)T_Q#>78I6t-o6&Bcrf8|Kw%GMJoeJO3Nn7sBZpi`^`Hy
z^@`IqOQD@7qJLa4)@oqeA9F_i_WLROc82aYpS<TuZ1d8un-nw7%T04KGnwmhZq~IU
zA0rv}HRxRvX^jbbUgP96^`N0Q<7ef9%7Yp1F-P0Pn>RDsuDr6p@ZIfv)|{^Gd((e!
zS6z@&`P{bX{>wW4orQ{AS_-Ro9nt#IX*Tz{poCW3x~t{we<mj6EY=CGN?r0yKhRHO
z{hg>!t8SjC$vTm^$Erg1#5!$>UL|M!=pUs8b{5Irk0WxeG`ml(&G^z|re33b?4Y>m
zQ}qjyKbodZ-+$LrBgI&1u2_-DE89M|-_?r)1B0@ZkJWvc-+uDhf?wbN*?PWGwOX-~
z|G(<&%bB;%UD7tS5BZaB^mjko+_qgm*sq$m#}rrQCdmVHVeHWdO3nu=0(ltx0~mt2
zS-D-n1W*P7ic-^xH!(3VaQ#|vui4u)n3*S7)Iro6t}R&1K@5o}?jR25=`LGc5oiBI
zn~^72%}*@~rp2T*{M`TC^!Y42!F&#UE-+!9U>QG|AOug@PZ^1)<);PbSrx3>IjP~<
z51_fae!4L*EspJP7=C^!xy8&9ETt^v3lnAuR#a9*;%O*rz<HjQ^VS^MH0?LgQ^;C)
zP_*y_BWnrq6H<RM(S{XhIS(ffKg^bhO<OP6&hUB<^c1fHFOsLE{iNXz=LuHwQ$pft
z`f0*>FoPJmz~SF}==Xv(53LnaeI55UZd{fGbfKVwpc_mTPq2`K5E4(sK?Kgrl#0LW
zafuxiJo0|>@ZiaRs`__dVbffo3k4hmkX$I|CkJ;0Pq3<=DiTk}PY2FB6k_)>UjF)V
zpi?ASB$W_OQD9L(;;FN!!+B2)Sii9Ii+6(pOO!_*rsd0qkDH<!FKV*#1oJ!a!_x~-
zu&{$L5>M7oHWH?VCs@T#1&OEarw!-*cG(rmd~GT-$QA`2dzhA@)SN_68ZItME>;FH
zf<OcqFbDtpSl=Y!Y4ndND3Bo#Cd|N)oSTx90pxn8`7*Pxvaxe;a`W&C@bh^Jx;m=~
za(Oxli3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g(Kj$OHZ_+sk<vAC
R5fyXy3Sn~dbM#>J0RVsA?7sj2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.15D71B57C7DA4CC3.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.15D71B57C7DA4CC3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..251fc60ead8429cacab75ac55b2f1f10f6b24567
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%wA>&24yY=hKf07LvxF7oAF%x|MK9L
z*=Mp#pYwNSuit0dxjtK8zEZ@~(<X4qG4cIxu8VJZV-a}i_O!g+wG%lT7~~!>zTi<{
z-jKRdSNMR;`t;D9)%IzUH8cKuiEj_TK40uhmE}x6k@xdXoc{XYP*Ls2vIR5Z_(c4j
z*W~S2Ir_H!=-2D<A3l8D&Kg_4FZ$BzuUTe}tV`v?vv>U~`I=h!{byU;-38B74<G-z
z$9#=PZnS6Cd6v&v&Q`mp_8A@ia!=cTN3Qkj$c3u8ON@giD{1L=7{$0N#y|XD<-eBe
z_?>?;58o*@S1&WI`&g#!dtkDO)6@l#id+2`6<=F=zqjOffOfEIqPWbY^7yFho1QG`
z;d-!C?_-XRh=!&x!+X!iQ%(2I+F$wkpR7s(`|?IME6o|4UGl!y5>C0F6LZx(kRSTM
zlp}J}ivGlwH*=ooEQs`2H#3>>@1Khq%4UpK>Cw`1hfc~S@vNHq;$d^QO~iA5wKaUJ
zGp`1)$=^yoWpetV<cU)*3CCvDW-o5Hdv|lu)QK02+H2-at@lVzJUl1Ksc=h?$jsGA
ziym{V2)&p5cwtBox5wTvAMFyE8yk-#9NstCeec0j8&o+3Y=Um<ZDkBe)iK+0z0{}U
z^UZ3FnU~aCC&%sLbqiv0)NFa^*CsIcpcU&D4}K@_9zV6cSFc}F{BN+gB|&lRE6482
z`Cs3v1})C)xhr+UC_Sju_RZw8n}eJLPX}3b%nSXL(jnKltT4^(i>!U==aan(xBVU;
z`EqJ)3U7|nt!Y=5B*(5y*QoqiJ*Dk2M`US7kq%?~HQ6`pA1XNfLi<WfxdeH|ihtOC
zR^{6?x#zm3aGOVT#eL;d4D9=3)<0{1&$PdD)#KTnftS`fF4WDxP~ybT`k2>Gz3K9*
z&_uIy2huLIN^MT?NIB=$9=7mKkIVZ@$_9H?b|^1Xntb|VMT@`ICv)#Tf9D+3tG-)Z
zRu#SHvC7*$rB})}-FIHq!_On_u}W{!shb_M{VsFt5#4fwODiIw-ugX%x=QpML8T{d
zg`WG1+m~5yG5R{~@~WHfe>}W%p~z_I<hw`wcYSnTC-{C>&x0vK>6P5}y?SoRK^o3k
zR}1Hz&5ux?z9Zz{`64?OmrMV6Cn(Q;xb5<OsV$WvQ}#<-N#Aqwvm$5Zz4Q~GFYmQE
z@0H`=cY^<4so*`E75mQrU92=e+TC>1*Z4(}evc~o;uVq{oU?w}bN>0i)p&>eP4n2-
zt2Z$eSLP<k19NB4^Xw{x-s){U4E_NO!Cb6dAz%V10|7;;X~lOK85p>JEx6b0Z5qtN
z6D;B+;tSUnEb1hR#1nH8L*gk|C^*AZ3a;;zDqZ*LFf&iEfRlhLOqeHF&`A)9C+s8)
z=SiqZsugSr)nVibR<uy`f@#ry9VZav?oh|X6U^hp;|>$%36{2ywuSK=+utz!{8Dm@
zktJA)L&_W`>?-nBaG@}B6;Lm~6F-t(ISV<sUYKG=E^t`%9{Rmt%|mO25I-UH7ZYt*
zfzILP<mQK|T6+6^l99wiHK38aPP|A)%2>$2johQ*sk!v#hI1gj3fzt`y@x{VUdGE`
zKMoX@V3AOQ2|s2&|IYH=;<rFyWXC@>VEw|*FWwC@M3h?|ruxf<kDH<!FKPm#hR=x)
zDfoq)gphc$7P9cj;0czukcab%QgaePv0Yr0T&xUY1c3-JU=IHE{<aO<z5Sn=0)rR=
zVZsay$+;;x89=UgnlCd8D;qlpCpQnT06(9npsTZ*AeX0;khrjjgrtR~wT-Qvy@Q^p
xww1D(hCDV=F=-iD1w|zlRdr1*9eo2sV^ecE6DeII7f~^HuMj3TKSvKn9{~D{pSl15

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.1D323BBA44E4F194.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.1D323BBA44E4F194.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5d85d8f7026d737c56224af9ea17322d206abea1
GIT binary patch
literal 1624
zcmWIYbaM+}V_<Mj@vO*AElFfyU|<jg;^LB`%ywo524*e>hKf07LvxF7oAF%x|MFnZ
z>@(T5&-pvE*Y7j!T%WBkUn%10X%o2QnE3uT*Ttv2u?V|#d)l|M-%7#^j5Y_DFGxDD
zC3HJQt2J20ojsNG{a;h+hs1x@v3Fwc*F;^I{j|a;V2}0W>8}sAKK=FSt%KoCDUEZ^
zYu@csIeNGK=-KP{KYaMw&l+35FaFZ&uUTeptV`v?vv>U~`I}n#|L3&0`-`5b9zOna
zkNFypRB2DG^DLjUoTGM6?K3+3<({_xj@+o%k?T~~E;SCCY@?;yVHD%882|8pmH*nJ
z<9GgvJp5<W%)ZRD?qiv<?}5oCPNEAU^|tyiD!#V#es9U|0L@^vL~)r(-vc79Z+Nn#
zhwH&oy$?A$Pc$@z8Qyz7o@%;x*8a-R|72AX=3aJSUMq5j<tV>ew!u^3^zKC>54LN4
z;A+u{xYBOmw9T(b?1GNU^)nkA>i#8c5uDAic5^Rp|HCJIhHO_ozI-_7D!1mdyKq?h
zRg)}@X8yB^sk5Gb;7Q3;GDtqVZ{tO0`R^$gPkCg_aJKV1wSURx4T*k+3oW87yUt8C
zyjaL^MdSU(j{zDhS(n6VFB7%mdlQkgA@TmH=6eaLHzu)g%v+H=`xe706EW#qucOqC
zynJIT;`4IxAy56gY|5IA3q%eTxI1^8OPI?fwS;}4s@w6$_g-ngo=`s{#&N@hu&tMq
z?xojOcV5ZJO5V=4?aa)OSpMG`pVv&7aD34OzD4OvZKM{tFUXuTQ?T}U|6QBif+?4?
z=FEuOIdx5&ugC34vr1NIf8|lt|KD?|Kt^%(E{%4ff{fVo_Xi#;T7+6%3)6HliSYSe
z{m8py)8w9TP2o0==pPr9Pcg9Xk68cgZh7N<m#YuY9?{6Wu5f+Y^$e>;?Q=d#Ef+qR
zd3A-s?h^;nF0^rNPVh)M=hhy!@J^4*`%B6OJC$}QFH?EE`r=WJ`5{m8r&atmZI1o6
z>-(D@+bWKG-l}}PA~$t^dx)gGgxQ2CxhhLH3!2YeCRm}Ju}?H)!-0RL_vFucY_m{y
zInn!QQr&Yt-|~#iRl3Vl_ul{bF6MHQP30NGN4c{F=lDI+$$v2Y;?oyRPfktn+Bzrw
zb9C5gRlUjocgOTIPFypeGcU#XPTnv7=BxfqA2%o1J^or^mGiCSmis)v{_o2!uQ|}P
zyyee>c#(%+7r*)~W3TycXK9F{^?%h_50A`!_Lh~2?X7+CJ9UQt^>5hTIez2KwOhE4
zp|~<PNgkLlyA1BPMRfYi=V9;<U<l@7<?;a&Kp6-qN=+-i!^ptE^=rYsW^dDA7M@@c
zClOz`wqQ{wQ6!$2lNb_D!9u|qrc!Wyr&Q^>SBIH-f(4ueTw%gI!GcbLNIYRDVK`6w
zb(}zuyF(olPcV-Yk2^w(w1u=SjOW<?hT-Rzl3R={!BQMj<}hJbk+*^ig_)~>dikCB
zk@U)0$iel7_z9`Mm}tYw$P>)X$;}Vb3!@piz+usQ==Xv(53LoJ-hQ8CB=JxUXe6%_
zFOrcm7BX-n_h@)(F1@+o9LPuoZbz6=heGUL#>-zn4iuJPkx+sO-*qV1H+TItE1<BV
zg`yWs_%ZYOcb4xKzXb{-`{}6x>lb!@@otdWqTKQ@)n7Jz+!Wn-Q4<(7d`^5w!7t<_
zgv67zkcCGEPq4g&Je*gQnv)2M?c$>3Vr38`2t+_YAdJDl96Wi`YoA))=tWF{K@7>c
zDLENHIqx)IW)@a9b`DN%9$o={K2JecXEi}CPbVR9VG#*Q3rlMoTRVFPJyC5dWit(V
rY@%Y)GO`MaN-C=Anp!&g28PC_=5i)dx<)RdV(wlcOm2RT9*jN!P%W%y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.900C533B3DB605B7.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.900C533B3DB605B7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fba666d07ce5bedd4d00e55829b16eedab949ab4
GIT binary patch
literal 1628
zcmWIYbaM+~V_<Mj@vO*AElFfyU|<jg;^LB`%x-1|201PUhKf07LvI(|HWPUIKP7BJ
z()Ch{#q7}&rF@g4C#yYn=iz8jX;V1xXp#4RJMO?uIk#r2`Q3iW-Fmlq@4Q$I<`WD#
zky)o%4359f49&aW8_D)QegEa3Z?AqmZn-zkM~|)RwECwc(;XWk=I)i%&^yQ8wOnu7
z_W6?^ef#_8*XjI<AD@oTS!Z`w_r=vL%ef1hF7clhjji4DYtN3~CC=;f0zM0UNHlxh
zsJb#&+w<+rgt)by>E@RL(`MYOe9E~!H{Diy%aYy_x2zzSrL2K%uQi@%zlqOZ8TxR>
z?YTC0=B5^$-*Qa<Ufxp9ib-xOgiN%TeDf?hx9iL9X?fyVsjF0Snw>wr^L`yQt!GLg
z<DM(3^LksER8yHA*k<n8v@HCydH%gvC$F4*lM4>&tj3NXnNxV%YEp6nr!f6-V~uN`
zpxXWGf!m35p68ejITf9i;m&x@zo7R9%ZoX=xfKQK+h-&Ptaag*k9l&?s=j4|)tY6Y
z&XP45{EC-#+S5J5)R>H~y@`^x{Zi-O5$5J}TIN$>&_C6=t)|6YoJR$pY8ZtdbvdrM
zFs!n_w>`*HSK(`v;o1uY4auDc+hW!EtD2`KdIvdu@W|bLg)Ly}#v2#nt{W=)rN3X8
z;w3NYSM~DR3r|iTuSBVL94#++u4yiCtuokp<VDJ*#rwAL{cP_p<`BEI>fKg}?~(W8
zluTzGyV2XoTdWp(eR=H>8*NpQjXF!4Uqq_*JA9duA(fuya{ptp`TRQ^o8FrE`K-U=
zaqWQBl3Y(;t1DBh+9pk}cgwWsn-E&9-!4>;aaZoXdVl+akf#yh>f#p@41eDjj!)9J
zK3OqWX<_%4*6+1$k`EqCU;Q-a|J=vsfu$A37rk$pYHWGR>va6h!h}5sUmkg~B`b7B
zpYaCCl7+`KUI@+Hd{I$r%IrxmVyjs6&T+o$Sb5?}jfvdD$z6YZh4c5PZ8#hEeV*)n
zvG*39*7>qq=f3$qLF+N|VYW%Dx|bB0Jz93zgejj@*Q|+4><0hwzb`vdnvI=WVzl*M
zn!meH(jK<#UO|-UZr%IyrB~^_OtJnrIq#1^){)(M$`(tGMBY23zj*V-86u}9M3v4-
z-z<IY1navg|94N|=bN~>Uh(+Evjx{*e*cvt_w$MV&o7espK4Ax%=^yx=}(3J`OQM7
z&vX7f`Cs?YZ1$_$JL^Lp7ng)6hyPDCJ8Wpf`;gm)C*r`o|C1~Kum8qcuKb(p@0?ZF
z7>X-%ljMPU^Z&LiZvOfuSv(B>0Sv)htX#oh0w@CkMX71UcNiHMxPC3T*X(T?%)%2a
z;w0h=*A^`5B#Oina}q=1DVQrb!BlF$juQxScc^3H3FdL)afb=>1WTJs+rW5^?Qa-<
zekr-d$Pz5YA!P;=b`^OmxKNn63aFRgi62R?oVgraZ-}3e`iqG+tc*Ot+??F}Fuke1
zj(Zz7E=ywO2^Me?@PG;P1PeL|BJqTsgyFn8S&5qt@&W8XlN8MrJz-j=a9-A*z9uLU
zD2(j5rMKTF8A&`;13He^i5JOnGUhUH$L-PZ)LeRV!#R-S6u2E>Mji^Wdl@f({Wwrq
zf<;0JCj8WZ^$R<{csEG3D7QRJ_{)Zmo1z;pY69bc&xsEyoQ0f(ka)7@vhYyn36?jP
zhx1?tF>-+;x%bfT1#2E!D-@;XB!XhQxG1?;8N>(#5fBgrV=yoW%lxooPyds&ktr~U
zAvrfCCj%(wo#xBT!pg?Z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@og
urXi0_R7_e%RzXooMO9r>OGn?p(Ad;m&O}Pr$VF7l-7AF2&Ck(;(FXuOnwg>i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.5AFEEB9815EAB2BB.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.5AFEEB9815EAB2BB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cb445fed7936a4f404e901f8e38b187ec2634a36
GIT binary patch
literal 1622
zcmWIYbaV4#V_<Mj@vO*AElFfyU|<jg;^LB`%vNRwhJTz43>9<EhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<i)lV%8nDKxO6CO-ScBp{QZR~$NMHF-QE)Mru)py!!s?57aKS5
z91v^W=EG3DXV$c%-|kADe(?8=zt#TT`+iz1H}C%V)S>XJOWYbZak15(w=*Aodeq_L
zm7|~D6$jnfJO6&oPiFr3f3q%MUH+SOZ=CM(<dA@0i8~jq?!P`?{QLa8mtS?W;+X8t
zq}gsznVPGz(nk2qYSYZOZdRvwE?cj+dAKceZTik>Tc(}!@cp!6i^8b_+18rYzpr(-
zo{}xzy}14NXCAp3*M1zm-D|?=e~BZsA#LXsyUufw)9bEZdh4{*XJ$)o!Hqrbhjz88
zDyN1p=$AI8Co8&p&Sa?Yms-E(<<&oDXP1Xh=qPn#=G_uxAS`m<`C7sW;pFZ`A`i^9
zK5(_@L|myhciQGxBqks>Df?N3gY|y)+#^a3uXB<QFSg$4+UIC0q+3%kk^S`_#zSum
zX1npGepq<3-Dk7Ryoox#ENN?QnwnL;_$w#4wx^}d^3CzUpZz6b*Uo6FiMUT$d?H3N
z*!o+G)KgnAIWE<$4I$Y_w>TYU?9*hEE?ax&HE&v*s%Ogsm)zSSYynF*MqG%yZlu_k
z{(fPCm%NBy)yr!yJULaCPH3}cP_$~U<jUA!>!Dj(V)HIAKC=C9`)O_tu}iDUqI!O>
z$<sGAoptL*ZzFH9TI%)7b_w%Dg+4`yTyp%fW(vDPmCCGRGbJyE{|Pv&Zr+p=)hm2i
z+kDd14Sbo=K@-1KE}gaT$<zCRQhmxPq0dw2Gx4o`{h`lBt|@qFa_-a^m*$ft`;XUk
z`jt)ksO9P;p2;m=<tF*yf%xjDYkr^mI6G=<z41lwTc#(rJ~ee>|2Dxm?s18d<ytAx
zW$BVPcy0wGij;KD-F#6|Ys&0NFJh}$^z=F2b!eXakyK)KNL}=|xo_=n+vePFyS~5q
zu&v^_=dH@uD{@o!w}(i|ONdRFlB=?Gv!MChWr7vj8T$l7HXNw4c<1k~B0Wb?=}A<f
z=l;U>W!76vuC`s?bT7BEeBH`jRXm1|a%T(5_&w6ef3W@H&lf#UPEGLIIwyT|bl7Ru
zca#2a59sIH7`}hP?GtG`Zhxt+y{f+N$4dSAAFGTc!>Wv4%KEFdS5Ni~OV}msRQKtB
z=Z9Z`U*EFLfAUR6HuT2q|C7!Z-03^~&FL@uRr|NM^&kA*e}nr@;~Q<G<^_)!iYs%I
z<bk=eviRRBwGeR~9tQsahF~sME^jaal!1Vv)U@I|j0_B1zZTqU_BIV>;RzOT67hv=
z3l?<}MdFD$i6QY6%oUtqDm6<Y4(-nRZo|kEtZ1(22@}?S9VZav?oh|X6U^hp;|>$%
z36?gOwt?{++utz!{8Dm@ktJA)L&^*$>?-nBaG@}B6;Lm~6F-t(IdeI<-Vi?_^%oOu
zSb_F(b8_>;^rrec?rq$-EQy&XSiniZ118K9Ea)VN#1nQBhVx)*7`ebf)O+anf;A7V
z6{c`r)}Ov6C=qA@vg4NCexGC{@lXxuI9?}SB*)2^%fKDCN5fNd>CFx2K#o)3c7)k*
zD8%k%y!`d!Kw$|M2_=~DQv=p7?EK>0Al0JW@-X2q8$ND|ZoH@oj2b>CKBV9mauPz~
z$(qZ;BZDVc-drBeD@x5t1jTl7QF5^|kY*0vp5#{<owZSyDKLm35Rwo;WDtUxoSTx9
z0pxn8`7*Pxvaxe;a`W&C@bh^Jx;m=~a(Oxli3^KJNLpB0+t}LKJLrjOTPd4q$YT=~
pla`THP*hS;RoB$g(Kj$OHZ_+sk<vAC5fyXy3Sn~dbM#>J0RYnzkL3UW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.BC2C6EDB175EB210.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.BC2C6EDB175EB210.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc08aeea6c2f804fea2357d55d1db93f18bb59a4
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%wA>&24yY=hKf07LvxF7oAF%x|MK9L
z*=Mp#pYwNSuit0dxjtK8zEZ@~(<X4qG4cIxu8VJZV-a}i_O!g+wG%lT7~~!>zTi<{
z-jKRdSNMR;`t;D9)%IzUH8cKuiEj_TK40uhmE}x6k@xdXoc{XYP*Ls2vIR5Z_(c4j
z*W~S2Ir_H!=-2D<A3l8D&Kg_4FZ$BzuUTe}tV`v?vv>U~`I=h!{byU;-38B74<G-z
z$9#=PZnS6Cd6v&v&Q`mp_8A@ia!=cTN3Qkj$c3u8ON@giD{1L=7{$0N#y|XD<-eBe
z_?>?;58o*@S1&WI`&g#!dtkDO)6@l#id+2`6<=F=zqjOffOfEIqPWbY^7yFho1QG`
z;d-!C?_-XRh=!&x!+X!iQ%(2I+F$wkpR7s(`|?IME6o|4UGl!y5>C0F6LZx(kRSTM
zlp}J}ivGlwH*=ooEQs`2H#3>>@1Khq%4UpK>Cw`1hfc~S@vNHq;$d^QO~iA5wKaUJ
zGp`1)$=^yoWpetV<cU)*3CCvDW-o5Hdv|lu)QK02+H2-at@lVzJUl1Ksc=h?$jsGA
ziym{V2)&p5cwtBox5wTvAMFyE8yk-#9NstCeec0j8&o+3Y=Um<ZDkBe)iK+0z0{}U
z^UZ3FnU~aCC&%sLbqiv0)NFa^*CsIcpcU&D4}K@_9zV6cSFc}F{BN+gB|&lRE6482
z`Cs3v1})C)xhr+UC_Sju_RZw8n}eJLPX}3b%nSXL(jnKltT4^(i>!U==aan(xBVU;
z`EqJ)3U7|nt!Y=5B*(5y*QoqiJ*Dk2M`US7kq%?~HQ6`pA1XNfLi<WfxdeH|ihtOC
zR^{6?x#zm3aGOVT#eL;d4D9=3)<0{1&$PdD)#KTnftS`fF4WDxP~ybT`k2>Gz3K9*
z&_uIy2huLIN^MT?NIB=$9=7mKkIVZ@$_9H?b|^1Xntb|VMT@`ICv)#Tf9D+3tG-)Z
zRu#SHvC7*$rB})}-FIHq!_On_u}W{!shb_M{VsFt5#4fwODiIw-ugX%x=QpML8T{d
zg`WG1+m~5ixpKSf@~WHfe>}W%p~z_I<hw`wcYSnTC-{C>&x0vK>6P5}y?SoRK^o3k
zR}1Hz&5ux?z9Zz{`64?OmrMV6Cn(Q;xb5<OsV$WvQ}#<-N#Aqwvm$5Zz4Q~GFYmQE
z@0H`=cY^<4so*`E75mQrU92=e+TC>1*Z4(}evc~o;uVq{oU?w}bN>0i)p&>eP4l?b
zt2Z$eSLP<k19K-=_+<OeGPMRC2LAwtU@lg!5HJChfq<gawBkFA2z88HVENudzZa}|
zXsy8YYr(x{Z_{8Fo?sCt5ns4z!J<y0NIWqoF(jUXg@Q9orQrHbsnT_?4m0xv3pfe5
z!i0H(1)T(uc*0J?aGr#kq*}q2P#s2|U_}c>FPIj`_BRYazm(i!WC@nykTQn}hxiGp
zznEyl3RKO_$;}TFo@KaYWm&V~1}2_hekXo+m@rSUoQ0e%jJNdm`y?ZYhiX7YyiUAG
ziexNg;EMKWcxo=ax#1kh0tIeInBGGnb}!@QuO9~rORz{N!G!Pc^}NaZDxni-B##ph
zl9AFD(r_amGoOEF`EK!Bpk8FJJT+kb!p<+=4YEm;TOMZQmkl2`MK@m51jYfM6CYAI
z3poiP@nkJz;i1kGEN>wX=M|;qB!XhQxG1?;8N>(z5n#X^T=HQ;MMLD`cBa4}hCrAw
z14D9dN=^ol>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3w
vY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)5$<5EvgV6^7q#1;T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.8A9EF4FB27AD09E7.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.8A9EF4FB27AD09E7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5935c162deb67cd51ad81a94a3b8ef45655cfde3
GIT binary patch
literal 1624
zcmWIYbaM+}V_<Mj@vO*AElFfyU|<jg;^LB`%ywo524*e>hKf07LvxF7oAF%x|MFnZ
z>@(T5&-pvE*Y7j!T%WBkUn%10X%o2QnE3uT*Ttv2u?V|#d)l|M-%7#^j5Y_DFGxDD
zC3HJQt2J20ojsNG{a;h+hs1x@v3Fwc*F;^I{j|a;V2}0W>8}sAKK=FSt%KoCDUEZ^
zYu@csIeNGK=-KP{KYaMw&l+35FaFZ&uUTeptV`v?vv>U~`I}n#|L3&0`-`5b9zOna
zkNFypRB2DG^DLjUoTGM6?K3+3<({_xj@+o%k?T~~E;SCCY@?;yVHD%882|8pmH*nJ
z<9GgvJp5<W%)ZRD?qiv<?}5oCPNEAU^|tyiD!#V#es9U|0L@^vL~)r(-vc79Z+Nn#
zhwH&oy$?A$Pc$@z8Qyz7o@%;x*8a-R|72AX=3aJSUMq5j<tV>ew!u^3^zKC>54LN4
z;A+u{xYBOmw9T(b?1GNU^)nkA>i#8c5uDAic5^Rp|HCJIhHO_ozI-_7D!1mdyKq?h
zRg)}@X8yB^sk5Gb;7Q3;GDtqVZ{tO0`R^$gPkCg_aJKV1wSURx4T*k+3oW87yUt8C
zyjaL^MdSU(j{zDhS(n6VFB7%mdlQkgA@TmH=6eaLHzu)g%v+H=`xe706EW#qucOqC
zynJIT;`4IxAy56gY|5IA3q%eTxI1^8OPI?fwS;}4s@w6$_g-ngo=`s{#&N@hu&tMq
z?xojOcV5ZJO5V=4?aa)OSpMG`pVv&7aD34OzD4OvZKM{tFUXuTQ?T}U|6QBif+?4?
z=FEuOIdx5&ugC34vr1NIf8|lt|KD?|Kt^%(E{%4ff{fVo_Xi#;T7+6%3)6HliSYSe
z{m8py)8w9TP2o0==pPr9Pcg9Xk68cgZh7N<m#YuY9?{6Wu5f+Y^$e>;?Q=d#Ef+qR
zd3A-s?h^;nF0^rNPVh)M=hhy!@J^4*`%B6OJC$}QFH?EE`r=WJ`5{m8r&atmZI1o6
z>-(D@+bWKG-l}}PA~$t^dx)gGgxQ2CxhhLH3!2YeCRm}Ju}?H)!-0RL_vFucY_m{y
zInn!QQr&Yt-|`hJx2Y~q-FyGzyO_&KHkD@#ALY&#oa6UMC;!3pi%(xPJvlYOYwMi!
z&(UG0RrMzS-yPG>IC0H<&b$=kJ9)qSo3HveecYU2_xNjxRnE7PTkiAx`oAx`yyigD
z@|Hgj;zb^QUHs~|jJ@W&ouwg)*8f#!Jv=h^*;`g7wzu}l@6;Lo*S}$V=lG2`_us;O
z48@hXN%FvanLn+ZYg*)SOCARQ0ES>LRxTef0hED&qSUnFJB$nrT)!6FYxXt`X5k4I
zaT4)`YYP^25=G*PIf)_h6f6{+VJZdJcS@D6dv%zZCs@Eqz!fIU6D;T?h{O|i5{B~}
z+utz!{8Dm@ktJA)L&_YcCB#oi{l!EZRz{v+Zcc7~m@tfH<O2Jk_t5VJYaUuF%re}v
zvaH!~0~1d$zZ1VZOchVCoQ0e%jJNdm`y?ZYhiX7YyiUAGiexNg;EMKWcxo=ax#1kh
zUIlJPnBGGnb}!@QuO9~rORz{N!G!NR6zrS3{+bm~SkXe!3nqMjujftPR|%a!n|Pdf
zkZh8+kcQjznECuW%Xf?40`($${HX!!7j}N}Zjb{+x#eL-e%bJGQ*`4+O<>gUIq@L{
zzmSs<5>M7b79JTq!SWXJa9&YrP9i9_i;I$rl|hUk5CH*!Fa`s2@B@u+M;)0SKVu3E
zVo1(S$;kl9d8hd@v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI+SxnkiE3LZ
vn`y{n6BUz|kyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN32ia`SWaVDte1DmkUN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.DE2A54B24DD9CFDF.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.DE2A54B24DD9CFDF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a34e1b08e6fe9ec651d316dee95228c314b7f806
GIT binary patch
literal 1628
zcmWIYbaM+~V_<Mj@vO*AElFfyU|<jg;^LB`%x-1|201PUhKf07LvI(|HWPUIKP7BJ
z()Ch{#q7}&rF@g4C#yYn=iz8jX;V1xXp#4RJMO?uIk#r2`Q3iW-Fmlq@4Q$I<`WD#
zky)o%4359f49&aW8_D)QegEa3Z?AqmZn-zkM~|)RwECwc(;XWk=I)i%&^yQ8wOnu7
z_W6?^ef#_8*XjI<AD@oTS!Z`w_r=vL%ef1hF7clhjji4DYtN3~CC=;f0zM0UNHlxh
zsJb#&+w<+rgt)by>E@RL(`MYOe9E~!H{Diy%aYy_x2zzSrL2K%uQi@%zlqOZ8TxR>
z?YTC0=B5^$-*Qa<Ufxp9ib-xOgiN%TeDf?hx9iL9X?fyVsjF0Snw>wr^L`yQt!GLg
z<DM(3^LksER8yHA*k<n8v@HCydH%gvC$F4*lM4>&tj3NXnNxV%YEp6nr!f6-V~uN`
zpxXWGf!m35p68ejITf9i;m&x@zo7R9%ZoX=xfKQK+h-&Ptaag*k9l&?s=j4|)tY6Y
z&XP45{EC-#+S5J5)R>H~y@`^x{Zi-O5$5J}TIN$>&_C6=t)|6YoJR$pY8ZtdbvdrM
zFs!n_w>`*HSK(`v;o1uY4auDc+hW!EtD2`KdIvdu@W|bLg)Ly}#v2#nt{W=)rN3X8
z;w3NYSM~DR3r|iTuSBVL94#++u4yiCtuokp<VDJ*#rwAL{cP_p<`BEI>fKg}?~(W8
zluTzGyV2XoTdWp(eR=H>8*NpQjXF!4Uqq_*JA9duA(fuya{ptp`TRQ^o8FrE`K-U=
zaqWQBl3Y(;t1DBh+9pk}cgwWsn-E&9-!4>;aaZoXdVl+akf#yh>f#p@41eDjj!)9J
zK3OqWX<_%4*6+1$k`EqCU;Q-a|J=vsfu$A37rk$pYHWGR>va6h!h}5sUmkg~B`b7B
zpYaCCl7+`KUI@+Hd{I$r%IrxmVyjs6&T+o$Sb5?}jfvdD$z6YZh4c5PZ8#hEeV*)n
zvG*39*7>qq=f3$qLF+N|VYW%Dx|bB0Jz93zgejj@*Q|+4><0hwzb`vdnvI=WVzl*M
zn!meH(mtyyxG>6ex9<J<(yR1drdWTRocBi{>&R|BWs4<8BJUm2U%dI^43Se4qDtqa
zZ<fAxg7w{$|GOvf^G)1buXudo*@EjYzyHdS`}suw=NHNRPc<hT=6z@U^ru4q{AQuk
z=Q)3#{I7dxHv84>o%JD)i%UY3!~dt69X7P#eaLOY6LDbP|H+mA*MDOzSN_fQFJ{#>
zhT_WHBza)oR9Mke5^-o!F%N@(07EbrD_1a>0Lnl>QEFQ89YzKQu3ro8HG7)|v+x9q
zIEnbewFQehi6ZgDoWziL3g!w<FqJTxkqc}=@1frd);zRUaBP3W@bgQ_Ek>4LDGn(!
zn5qyzA@vs%ZCDw3g1I@l`C-DTzK(kvH!e$J<_Q*X67YZt^8^bz2_o@?orK}MI$4RE
z4)Ou)K$8^B6+K~Erf^=?pS~t25h#po?<~VDE6bV<H!$%8^E>go!&LJG%bCmBz<5h<
zzfUrfc&G+c#OuV1q)5hG2Cit2hNtGzn;Xu7e4xPX2-ABg#O`Ih{Pp8NVF?xqC7AI2
zy`DFDUnO(`jpT9SK{8U>TpDiVQv=p7?EK>0AR|S&<zafiZ1}h-y78hWFlzXm_>h8M
z$Vmu^Cu=SXj|`q*d2@L<uP8Mq5ft0SMajj=AVwgFfPf$vgMm5ts(GQ9biDUtrobSE
z<lK~;44|BMnlCd8D;qlpCpQnT06(9npsTZ*AeX0;khrjjgrtR~wT-Qvy@Q^pww1D(
thCDV=F=-iD1w|zlRdr1*9eo2sV^ecE6DeII7f~^HuMj3TKSvKn9{^K%o$vqv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.EBA14072309C61B4.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.EBA14072309C61B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f2115b0162ea605a77f1da4c4643c1135144745b
GIT binary patch
literal 1622
zcmWIYbaV4#V_<Mj@vO*AElFfyU|<jg;^LB`%vNRwhJTz43>9<EhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<i)lV%8nDKxO6CO-ScBp{QZR~$NMHF-QE)Mru)py!!s?57aKS5
z91v^W=EG3DXV$c%-|kADe(?8=zt#TT`+iz1H}C%V)S>XJOWYbZak15(w=*Aodeq_L
zm7|~D6$jnfJO6&oPiFr3f3q%MUH+SOZ=CM(<dA@0i8~jq?!P`?{QLa8mtS?W;+X8t
zq}gsznVPGz(nk2qYSYZOZdRvwE?cj+dAKceZTik>Tc(}!@cp!6i^8b_+18rYzpr(-
zo{}xzy}14NXCAp3*M1zm-D|?=e~BZsA#LXsyUufw)9bEZdh4{*XJ$)o!Hqrbhjz88
zDyN1p=$AI8Co8&p&Sa?Yms-E(<<&oDXP1Xh=qPn#=G_uxAS`m<`C7sW;pFZ`A`i^9
zK5(_@L|myhciQGxBqks>Df?N3gY|y)+#^a3uXB<QFSg$4+UIC0q+3%kk^S`_#zSum
zX1npGepq<3-Dk7Ryoox#ENN?QnwnL;_$w#4wx^}d^3CzUpZz6b*Uo6FiMUT$d?H3N
z*!o+G)KgnAIWE<$4I$Y_w>TYU?9*hEE?ax&HE&v*s%Ogsm)zSSYynF*MqG%yZlu_k
z{(fPCm%NBy)yr!yJULaCPH3}cP_$~U<jUA!>!Dj(V)HIAKC=C9`)O_tu}iDUqI!O>
z$<sGAoptL*ZzFH9TI%)7b_w%Dg+4`yTyp%fW(vDPmCCGRGbJyE{|Pv&Zr+p=)hm2i
z+kDd14Sbo=K@-1KE}gaT$<zCRQhmxPq0dw2Gx4o`{h`lBt|@qFa_-a^m*$ft`;XUk
z`jt)ksO9P;p2;m=<tF*yf%xjDYkr^mI6G=<z41lwTc#(rJ~ee>|2Dxm?s18d<ytAx
zW$BVPcy0wGij;KD-F#6|Ys&0NFJh}$^z=F2b!eXakyK)KNL}=|xo_=n+vePFyS~5q
zu&v^_=dH@uD{@o!w}(i|ONdRFlB=?Gv!MChWr7vj8T$l7HXNw4c<1k~B0Wb?=}A<f
z=l;U>W!6_hW>31j>0WMS`MQ<6s(1_^<<1tA@q46`|6u#YpD%izoSNXZbx!)`=&;kQ
z?<W1<9?;LXF?|1o+b7a?-2PHsdsThkkCpoKKUNt@hE*B8l=W9@ub%7~mat3MsqWMN
z&JVu=zrJOg|KyvDZ0L>I|0kU-xYKv|o6}$RtM+ei>p%Fr{|5J+#y8r=&I=wf6j$aZ
z$pdqx#_7WU&K}auJPiH;48dHiT;5;;C<6gSscFS`7#SG2el57y>}?v%!V@gwB;pI#
z7A)!{io_Fh5<}uCm@7EJRKjRRF0ci?hkh?u^UzvBvn1ls?yT=Nj6A`L=8B#$RgUd%
z7=C^!xy8s5EX5&Z1``hP6H<RM(S{YMnwyiGA10jY>$taZ<FX`Xo?rnd0S}lkPq3hq
zAQDg5Nf^$X!g*PL`kJ6bph?IU%re}vvaH!~0~1d$zZ1VZ!X!C!IU5*n>FxJPMiLLz
zfQopXc##yzn9INw?a}bmTzYfEIgnEoxE*174~5vhjF-QD94IWoBB2BmzQ5P=Chx0+
zPN0!IPCQ6PN}Ef=jeKgr`h}feyc=YsD7QRJ@0Sf9H$^vI)C5KipA#Qa@C!K!A@O9*
zW#N&*6D)5o59bx7<|Kk*ySOO1SQ$t&2jAEfn!uqG_?0Ozh#?S?5I|%Qf|;C~l9K`C
zdZ+m^v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI+SxnkiE3LZn`y{n6BUz|
nkyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN32ia`SWaVDte1LCl#c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.5B618BE41EFBBD23.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.5B618BE41EFBBD23.idx
new file mode 100755
index 0000000000000000000000000000000000000000..95bbb25faadff3982d5aa4c25f0ec4c52b8fdf9a
GIT binary patch
literal 1714
zcmZvb3p7+|0LQO+cwWq4MtRM68(Kmb4CQ)D#P!%}kx-^Mjz>1nyux^fh{8_Bv|7^^
z>v2@rFrFojP%Ku;W1~53C}vwb+B1_qXTNjqckb_hzyJS#?)Tm6>FnU}lQ00-20LF0
z4~Ypx0RVvIUC}X7)C?p5aK!-NQsq<c@Q7(s%)}09IPd0@3zDB~Urp6FXc<1hZ#f;;
z(8Sg_tvPLlbaprn6$ELsTblN5+Aq#@KTLR+KMWWF0df-(Xhet!DfcN?qUdqf3MVo4
z`=k%l-iXES(TZ0yG)h3VU%7$7Ts6tT_goGvpZn0jIr3v*?1v3Ai^R`Cz1M%>Ed3SV
zRt`Ep-s)o)8yxn#TrqdK<IQAiZJgLAa@FRqY2F>9EOCf-+K07qA^$saVC!R)R?^ru
zF{(c9?kIKW&*5*`c_l6IHg5A?Ym=5^>NF#ZI|Qm^G_NzK^J9zLrbuuKuB}CtmaQ_z
zaal0+=`D4BVOP6oC8wfTl$U3WO>03{Rb^FbEU1lGt);#H*>}h+Vzx8gF6>;)?NlGV
z`I@4HeoOJKo);x`xQu36BZ4V0`m$Y-dFr{RU0P|G4?R0rev{6r6CJ6ENKB0R(X7Yt
zhC_*c_KSgKLnY#1K_m4^8>ni^y-F=xFfL}kJwoQR<Z}?~$%WY_Zno-Iw<Ep&W{!5d
z(6NI~kR}Z=q~9FmUT3I!eE!zKDP5;#(=G)gN;iqUywEb9(9U>Lr$NI|Jh-N%Lar+m
zT+?JTbZ5aFrIXPhg8qXZX<wbE=9wA#yHK{W#UN~ByQX^n4x>A;Fmq6SgW%wcv8wf+
zp_1PH@}T|s^WpK~d&(^IgKK{7+!VA&X#c~S5m7%3(;Dz2)=(?2z)5WBaY`HZ;@HU>
z<^F9GHJRS7)5SSW&-BLgYUw+xO(ln*+0{p@F#Q%=xkKGOEBm(HsLz1?l1{}8ry=0H
zHUr=2Qq*rh)s<E4A$;f)JIZh->D>CUyHlMNuQk^cUQZ0&pd%Cdudt{Vv`UkFT1hUL
zC>+kx5pSy<-jWjY7S?7-l&hqU;uy-2nu`9C`oinAo;zB1B0A6BAMpQJtvp+nQp~LS
z*q>C-wR+uqMW-%>6=vyE!wDUWQ%G&iS-QY3&3+Zo_jx`$jBx;t52;<8YwKDOH}GpN
zd&Fv0Y_puC7-TZ~FvyoJ`x*1W%OS{4P+vs;fLVd}?NXzgvpFZ9QLlH+vEyT$+P7)H
z30pH}7I#KYeD0~BT*MolDC}EesuU(XTX;V|d#Mt4jZ@x0_$&QPW3pFLXSaKCLV=X-
z`Qgh0i;1%ZW1A4tZD_~Xe8-?H4Nv3>JJwVWM#qJpA@Kf5$Zz6w_>hb}2)G{yU<tH@
z>z9f5<lhk$ayHr=0RZCL^x5>ICNL6&aS-m9zy{+X{2r2r<OQUbyww;n64`<PVI@%|
zv#%CSQB48qz4ga<Ivz>p<w`=5j$aokla;N$k``G@KsH&cz}9-q6aw|1Tok-Fmv0pk
z|MEWhU0gbA-jZl9w}!nXi;U;hVIa(&IwF}UC;(v+nIs@DHyQ2aA_`gvgh4S-njf5g
zTEU^-&^VTtgoUsUd>MpgAelWR2gwPDIn6m^v~OOEXOV)W_LQrTRRnb)tWDMy&?6mZ
z67%K#;oWi~Xv1&m^0HnDAWZi0wpIb9pvIR$SRJJ<pz&7x2Ki!x1kbBVR(0VodUFuH
zE+Q?TyBmoE34EDD^PZQKB^2^xU9zq(UtYiezVTG*C<z6^(vY;^Y(ZESlHEfDGQpN_
z0bw1oj(~uukg&krZz(z|C|Z5@#&dW2GLZ0`g-}Bwv*#MZ!xQk}%K#7*9vl|R!=q=9
zAyH^y5m7M-P!c08<&3qr)xwH9Lo)KRa{Km^%`GWb)`ySi<Mb@lO?3$Wr{Wcrl!>Zp
j8k*W9-2(;(4UJ9AR1PV8V`PWJJGyuw9NcW2gk1jtgpI;3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.6161CF2711103FB3.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.6161CF2711103FB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc2186bfb77873b2f297f92a21426f5b38b524a3
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=na#`$4Bt5!7%JwR4b9EJZ6<K-zvj7~
z*=MqOe>gw7wR)qP;V!YC9sy50m0TuDdHzh@zt8=2U$XkqZCfP1xjwslcz5-?-g68_
z4B;ZjI9U&!&I%Hn{aswCng8<rpKsp%dUm<z{+ag0#z%JxSMIafZ4e)~&o;Pc-8qrv
zdSd1AlODZ0{N~x~{E{EPj`!%>-#`EJYmuSkq^8gP-Kz0*yZ#yN|6g-@eZJ>sp$~;t
zZyQ-x$270ua_U?VzNR$$OH4%KHb-`~*vq>VikkPh<zL*e=dzn&?$Vbc`wsp4I?e5Q
z+>f^x5C2VN<3A+v>(X1c+$F+MOSuh{+=VB-*!%fqd25}tcJ67FoJGr@*@~PqSCY(J
zt;}M=I(KdJ9F<(<3$|Mvt#7Y>I@|pHeMhBj##Wady1j0!cq&&k&B@NNOl48ihHHUg
z3pNPMy}3K^M%5H0p)~<g4_#GuZ(xsgcWt<~``A(U@*RqOf#x1!iog8ao*!mh^hRU0
z+x*i4Nt^j+rdir4N8aLVDJ{FbL(k#;_YR3%L8*r;<=vmgzmQ$ia@{G@bA?Lak4@fh
zZwP0VaD-H7m`Jg%ycQ+l+m>#zR3i4ywjI}_PxW|B;y5V0H9J_+VfwW|$Ld(^DOU5G
z8=NwV11fl4gg^1rR$AKGX3wzk^R!PP7Xqp~qF%iyxUqO)(Y#;$i?uuWF0FdEMdW+r
z-MFl*^z7Hy#I7B>JtfxvcLQsn!jss*1?_qrLi`Tm9T`&TiJ|cyolX7oTymnEg)eJg
z*T{-Ex<z-T$L~EsXFa%nq_>A%;;`KNd6R!*+qL%}WGk&))_d#U(%Sh#M|JaW+i%`)
zQ~I64G=+A0ME|&8e5!$Me}vw$-S2z$tqlEbK6%fQ*ycMHk-{Yj{5vO@`l?Sf-gU9C
zXkz2(1-+uKiqo%HUaxeT`c3#6>wLM!i=vBVH>yg<@#?XCSrYVLZDs!cv<)ZYzR#1r
zH|>F{xAl9Qa@}v$C${ni1~5k6EL-zs-D2P8nJo9T=B*bESL&*l{oby;sn^Kop_liQ
z`#TT4<kHXGI&III?b17~ukVrQlT+qNxt-RNHn~z=`bTPk<&K1%+tyrKJLTq=uqzhJ
zL%Hh3UtLt(`s#V0^o}00+5es6zY52HIud>&#a3?q^-7y}vi--~s~;DJ1#C+`R`=n4
z=gD7zU;loyJNeeqYQ+lE|EjY;CtCO2;{IyC_TTv@f9h{=%=-66`a1tyCx+t6+$4Em
zp8U(<^)#&PbS)2qe*i-;H!HU{m;lN^Kv8O1u_Y4&1J|zw_nN&;gPD1PMV&-_;o5@5
zoWziL;!fglp5XdUsnT_?4g<9aItjWWv<Nu~A@M|<MBqH_*Kq<t?hbW8Exb;=?g%Zi
z=CU>jo|?HDoM*5-ncFHe#*B$4ScXH!453JsLlw?*Y=6V>^GnGsCKeP$EWyZ%Tt(gr
zE)-_20$MKMB!FbOg1G|RI-X!nb4@re#7{{5#Y7ucMxJ0EP9A=k<x6kBPco8ts0K8N
z&xsGoBsp_AxL%%Mb#rw%Z;ytj=F*!R&Vfu);BkbRbST8`WxV|L<3Qg@u}CSwgn5FM
zS(M?ttJ}rCT9j_&2I^HbSM-D_dTPM>g`Ho#8)T#?k33BHeP`%HUSauups<#?Rv1k9
z%Z87eq8l%2vhW1+JMqIKktbN#Nf?PIZ!Yfy)4~(1VXgt^!9*Fkz$vKr(C-Cn9$G6f
z6s6`Qg3?lPQF5^|h!F%Lz<@be<_ULw{DLjXOo2fRfiPhPhUDCooD3k>JI$Awg_Vt+
zgOi(wSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!RNG3~OhX=<sF<{jtb(GFimJM%
emX5xGp|PpCoQagKk&CF9yH^O4o1dcxqYnVuXUf$8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.77ED46D094B0B411.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.77ED46D094B0B411.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1362c78f922ae925e199db0db5e5b12f4e40408a
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=na#`$4Bt5!7%JwR4b9EJZ6<K-zvj7~
z$vWA*Kb#-kTD?)taF^Imj{p@_!9^Z#J$|O{-{*e1FIoNQtt}GYT%X-Nyu12c?>Pn|
zhIo-{+^kHgrh469cN<SU`1|DinyT{J&xNP*(+;l|K6+cZa+}3YgXow!^EKA#`HLLa
zdsaSw@}qZuzdSoFU-IMA^(p#xchCR$nq((8sp&KSXV&<dJ^yTW|1UYcKHu}R(1%2$
zmyMRIYZga}wg?^6iF_R{wX@(*UK6L!uVr5ypM-smT6}rh)43bFB-L}5f6;yupP%X5
z$7IgV|6N?th9mo>_jc*GN^`daFJsUy%5@eJv|V<8=`r6mJ=&8u^m&}E*03>Cl9WtW
zV#(#UvgWt+Jm$Kf`WEZs*!9!$@9nAP?!0mBg$wiAutToO<rB(9K6P$X57g$+7EWdL
z5?{30+T3gVilrSa+PZ5U1W)R{akw8SxWM)Pj~_4m)^{eG^s03WY@B>C`G)fp+0@H3
zf+d}D$`mi_98XvAzTQ}HFIVb5>x;i~0@-ug+!Ve(Uhv0#ar;sS<tY<eyo6Y%#bzDv
z3rKcaaL74vOX!5G+!oW6il*SMi^qLemsh#Z4D=LItXT9dSBrVZrAn=g_1UdK?Zq4c
z9Le9gt~frPWFzUigy-S`TLJ5mDa)JMZaH5Ga&JE@$2-;im3>UC)0YWhrK^|jIRENf
zP_T8l_5Pv_LN9{KY~R|7xhzzC-{quS;y9&>zofBaS)rKUSJC_7pI^9L@y*Ums(R|x
zEj7(B*IVn7-ue*ZNlpKDU-rAiQMvi+Chx{JvF$&m*~lu*5^{}J6`#8}W!wJtx|M5t
zqE2YJDv4)u%U8KcUU(qh{WInF+)L(Lv%aTIn!Y7_LbQH%)3-N@;+vMeR4KY<qSklb
z=!4{zjf*s%OtN}ztr8+Bvso{pie;Ui<l7GIn;(h_%oeCm{blY~``fm;^xLlQIsdn@
zOwoC}$NgRO-S>*-dM=KvDXF(N3e~#LK6ja8j`qCuBH>D1@_oNA3*L;1T=wOX$Jzab
zt(QgPc4etnM&@5Ds$N%FSolvOQQ@{(vdL_oy*UrIA3XY?=jpMHr?v#8O^yzHEx9)6
z$M%4JzLlZ-`)(ab-*WShb?sIEnv=@$ij%))ltewhxT4%f>h1epw@AmEeG<P<)~l<%
zcE7s)_HUK#&tEK9bnVad9KZIr23h)7{_ps;f70jwiOLzjHm^S}KG%w&xH30M9+)Q!
z?zf(Q!g4c@hrvI9A()$$+Z#*(Wgws^HLci^iGhLZ*MfV^-loCKJi($)qP}o#!D3Eg
zNIY>TaX3%=b(}zuyF(pN3$GKeJ3@=BxvUL>r)I7O=Q*~&Vfgu_<Q5Z4undQc8B7sN
zuquZtoaZX?R&b#(a~05B0Ve?@a}~@L;3n||Ynp4qc_Ds6>MthRurl%l^KkO;!_0-z
zj9g$(_a6GaV9i5og{8OOCmBgRR0CST=fsC(ft<M<+(@2ab#rw%Z;ytj=F*!R&VejY
z;BkZ*d?>{3WxV|L<4iokQY=zRFkzlxWfo;P@4`*%u97WNd4Y<MJ<WsS>8sntzFL%S
z<OW)xXs+l9GwG=T>lb!@@otd4qCE02;rE@P4|#><`+>q*=2~Ge;V&CLZi;TasL8?;
z%<sexk0_pCVJBfEp1ird6HE(Fu!gw?ocC8CY}4KRBhP>#Bj_aPiclovB!t8haT0;^
z7>ZJJ5<zLHxG1?;8N>(z5n#X^to!|a<&vW-8<+xv7y@C!3=GM+DLENHu6LR*GYcyl
zI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5
fO)VXL14Cm|b2$?!T_YD!F?X*JCO1Dv4@MsV;!f8M

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.07FF698C2C4EA5B3.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.07FF698C2C4EA5B3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b37d375e7c66b5614b42f3008aa35604bfbc934
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=nJvr=48J)U7%JwR4b9EJZ6<K-zvj7~
z*=MqOe>gw7wR)qP;V!YC9sw$<f{Q%fdi+e?zt8=2U$XkqZCfP1xjwslcz5-?-g68_
z4Dlk@xLKJ}O*OmE{^p)|u>IxxnzHiR&zZ;apB=t9^U>W-oA`OLGxYWE&(m}ZPwxsA
z@7}HNS@=D%?DJQ1tD3*b$EMGJ|NKwY&KW$O2Y((vI!WI?wtnXQ`h8EYZ(mZ$`NLxF
zw*yU~`!25Ga_US7Tk|!$H^w4y`#~19+RMKbidI*yy_k9W)47PFJ(ITu{}TPSe)|?R
z_r}|c+kc<tk!#udW$EqSvWb4TG=m#NcWzVc?2x<s-s`bh*s-aeH(ZuHvlWq(p4ii4
zEZDS-b?)5uIVz8pF4(?p`nWFqw0Zu${mm{p=_Sh=!?YhRnpl0pn(OD08<SUvwruU(
z#I%(A;?2F%OQWv@xin3k7Pg=xMf}@>_bWIqEXuE`DOqlQ#Bi3|WY>-xCtn!KD4gos
zlzC<+kD}S{37OL#8&6vL+QH)ew%+$lU+VfhO#PgdC;Tc5`M31qVXp;(r#zfgJDE=F
zZB6!HVW=4JaG}91Esw2chop>b4r?BL@p#+S>b;B4tWfC`unGKaHkHxmrH$y8>sy^x
zKDK1hXxYfl`f5R;>%1PNptg($j2*I8r-B`w-zsLUTzojOzwMOrS9!N^0ox^^cQZBL
zZQix_%9V3h&#jkT!@6rqtpD!@*1(JtyBA&Hh!t3B#}KQqD05D0aP3k4t2R3YOD{*w
zni042+M>2lpWBnVww%|#+OuTA|2JDEn>gKjv*%6lf`vzA_bL0gcPw+c7<SV2xyH=9
z?>EM0sV^4W<Tdex>#alW^8(pSejF__yjz?8#dqzk{WG7azLk1%?ewe#%+)7a*JN6C
z&N1!noZQV*!<M^7gKOuKb3gewL)p~VaL<ryyePU@wop|%j#rQE%aWk~YAf^ir)@YH
z_kEu1y=e`Tmd^ctqIUIf+b73bHC8mN$*~Sw^giQqrBvhguIJfDr*=GwKmPY&N6uoM
z;HuOm&+hL$^pdOm>aHn!)@+yFX?@*Bzishx0p^?ABo{wSd6CMzpGS{fT*tgGblFs+
z+4s7l&Sy?PbpL2oiO1ewop;#m7T<mL|3c*7BQtGE#f&2Rk3U~3*I#}9@yCbvD&~kz
zc|T*L{GaBeO?4r)`w#n1s_x^{$~sqn@>$KjzRTa7nEg)wH@5t*Kdnhj;WwK{(!5i}
zmAOgsz&u&;%KEUl;i|bj4E_NO!Q8CeK41bU0|7;;X~mXI3=CYq7Tjz0HVtOx2^Mt{
z^@VE-7IP9q;)y$n!+A<|KR4xHDBQ-x6D-3aV~)_G%ApG9IkvxH`1z&e7847KB9>re
zMInAd>MthRurl%l^KkO;!%TwFj9g$h_a6GaV9i5oh2{h)x7CKmWk9D2ItjYLRPh80
zISC>0M4Uw6yjg}@R+cpzZUAZ#a1wAwXi>0Guto4REi~c0rMKTF8A&`;18U)O;zQCR
zXCVhy#1pJ;p$_Nm(eTt<dUL}$kT(=~9AV}j3bA_`FMs_wFr1`Vq?BO7Ji*E=%5dJ*
z?P6aoN;h%?^(tB@dchRk-|Kmk_f<kC&=tH+yhyH)wUC9|%M+|-p$6wYHDLY1&M)2#
za)l_5Jj|r`ouLnTh2{H!!dezu;V|JZ8$ND|ZoH_;!V}Ez#1D^Ko?u}oVI-csg}gIN
z3s10yg$A6*P?VaJ2ue%EMajj=AVv^~00ZV=qxz}e{%n0A&lDKM5C{`yU`WnQ$;kk6
zz0-V|Sy<WFIXJm_cm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f
n$SNo*si>-JYU$`37#f?J%b7^&8o7vyxqF2$x%oMIF!}%hyUN$>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.1FAF079E735BA4BC.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.1FAF079E735BA4BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1e80350eeaa73223ebee729ec01c36e975cd933b
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=na#`$4Bt5!7%JwR4b9EJZ6<K-zvj7~
z*=MqOe>gw7wR)qP;V!YC9sy50m0TuDdHzh@zt8=2U$XkqZCfP1xjwslcz5-?-g68_
z4B;ZjI9U&!&I%Hn{aswCng8<rpKsp%dUm<z{+ag0#z%JxSMIafZ4e)~&o;Pc-8qrv
zdSd1AlODZ0{N~x~{E{EPj`!%>-#`EJYmuSkq^8gP-Kz0*yZ#yN|6g-@eZJ>sp$~;t
zZyQ-x$270ua_U?VzNR$$OH4%KHb-`~*vq>VikkPh<zL*e=dzn&?$Vbc`wsp4I?e5Q
z+>f^x5C2VN<3A+v>(X1c+$F+MOSuh{+=VB-*!%fqd25}tcJ67FoJGr@*@~PqSCY(J
zt;}M=I(KdJ9F<(<3$|Mvt#7Y>I@|pHeMhBj##Wady1j0!cq&&k&B@NNOl48ihHHUg
z3pNPMy}3K^M%5H0p)~<g4_#GuZ(xsgcWt<~``A(U@*RqOf#x1!iog8ao*!mh^hRU0
z+x*i4Nt^j+rdir4N8aLVDJ{FbL(k#;_YR3%L8*r;<=vmgzmQ$ia@{G@bA?Lak4@fh
zZwP0VaD-H7m`Jg%ycQ+l+m>#zR3i4ywjI}_PxW|B;y5V0H9J_+VfwW|$Ld(^DOU5G
z8=NwV11fl4gg^1rR$AKGX3wzk^R!PP7Xqp~qF%iyxUqO)(Y#;$i?uuWF0FdEMdW+r
z-MFl*^z7Hy#I7B>JtfxvcLQsn!jss*1?_qrLi`Tm9T`&TiJ|cyolX7oTymnEg)eJg
z*T{-Ex<z-T$L~EsXFa%nq_>A%;;`KNd6R!*+qL%}WGk&))_d#U(%Sh#M|JaW+i%`)
zQ~I64G=+A0ME|&8e5!$Me}vw$-S2z$tqlEbK6%fQ*ycMHk-{Yj{5vO@`l?Sf-gU9C
zXkz2(1-+uKiqo%HUaxeT`c3#6>wLM!i=vBVH>yg<@#?XCSrYVLZDs!cv<)ZYzR#1r
zH|>F{xAl9Qa@}v$C${ni1~5k6EL-zs-D2P8nJo9T=B*bESL&*l{oby;sn^Kop_liQ
z`#TT4<l269*OWbLwoC7{zP?AIPfnR9<#t+6+T==g=^v>DmOB!5Zd-F{?Ub8e!md~>
z59O*Ce|1rD>#OI1(mQ(0X8(7N|0*2+=}7pA6kECZ*DGz_$@U*_uYOz@7O*Y(Slx&F
zohN?<e*OE+?&Mobs}(Cu|EtdaoM_#5i~Fnn+JEPt{HedeG3(zO>Fej`Ix!Si<|fGl
z^JL&^sdr&v*TQ)i`~w(*xmmfr!30nS0*X@8iY=KK7`T2dxYz7$8qCZSEb1ic3)dDb
z<|Ky16L%7a^90v-N|mmAbr`5c&`Ho0p+(3^2#F`+Bm(CdY)|I4%8W5%;t7`FkTFAO
zQRPsD^BmjXF#P;da*K%tMG;FdvLcvPMlP^_dJp|xu;!t)LWrM``iqG+tc*OtJe)lI
zFjcb*x2!B{HrxPos(_P#J4~1-SixMu2Eo%b*M#$y-hQ8CB=JxUsD;mo4@rxhxg1;(
zPq4bVI-Ivh!&7tV%?;;3mMic$!puDsV)rs${`zsCr=?h=lwiU<!OASkaNgDJVqYyv
zH*y2@Dw->L!W7-#>v@y+RYE7w6}(QoNUo4ImxbHQ6Rc*g2IoCBVEw|*FWwDug(#0a
z%%u07p$~b5<@<rcTIO0|FySv7K5mL`yr{{-6U^_#506`(U|}a=B%Zvvyc0|dPq2o$
z2As!Gl$w(WN=wB>$;HYbMi7Vq1Lk0d{8{!LmApnwfk6y`FkuFU<lK~;3?SD#&6k;l
zm5rT)lbeTEfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`n6!+nf})a&
hs=B6@j=q7Rv8lP7iIlF9i>R2pR|u1vpQ8t(4*+eT!Hoa_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.FD07473382303B16.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.FD07473382303B16.idx
new file mode 100755
index 0000000000000000000000000000000000000000..942505d1008d5c9f2a8c780d716bc37be3a1dfea
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=na#`$4Bt5!7%JwR4b9EJZ6<K-zvj7~
z$vWA*Kb#-kTD?)taF^Imj{p@_!9^Z#J$|O{-{*e1FIoNQtt}GYT%X-Nyu12c?>Pn|
zhIo-{+^kHgrh469cN<SU`1|DinyT{J&xNP*(+;l|K6+cZa+}3YgXow!^EKA#`HLLa
zdsaSw@}qZuzdSoFU-IMA^(p#xchCR$nq((8sp&KSXV&<dJ^yTW|1UYcKHu}R(1%2$
zmyMRIYZga}wg?^6iF_R{wX@(*UK6L!uVr5ypM-smT6}rh)43bFB-L}5f6;yupP%X5
z$7IgV|6N?th9mo>_jc*GN^`daFJsUy%5@eJv|V<8=`r6mJ=&8u^m&}E*03>Cl9WtW
zV#(#UvgWt+Jm$Kf`WEZs*!9!$@9nAP?!0mBg$wiAutToO<rB(9K6P$X57g$+7EWdL
z5?{30+T3gVilrSa+PZ5U1W)R{akw8SxWM)Pj~_4m)^{eG^s03WY@B>C`G)fp+0@H3
zf+d}D$`mi_98XvAzTQ}HFIVb5>x;i~0@-ug+!Ve(Uhv0#ar;sS<tY<eyo6Y%#bzDv
z3rKcaaL74vOX!5G+!oW6il*SMi^qLemsh#Z4D=LItXT9dSBrVZrAn=g_1UdK?Zq4c
z9Le9gt~frPWFzUigy-S`TLJ5mDa)JMZaH5Ga&JE@$2-;im3>UC)0YWhrK^|jIRENf
zP_T8l_5Pv_LN9{KY~R|7xhzzC-{quS;y9&>zofBaS)rKUSJC_7pI^9L@y*Ums(R|x
zEj7(B*IVn7-ue*ZNlpKDU-rAiQMvi+Chx{JvF$&m*~lu*5^{}J6`#8}W!wJtx|M5t
zqE2YJDv4)u%U8KcUU(qh{WInF+)L(Lv%aTIn!Y7_LbQH%)3-N@;+vMeR4KY<qSklb
z=!4{zjf*s%OtN}ztr8+Bvso{pie;Ui<l7GIn;(h_%oeCm{blY~``fm;^xLlQIsdn@
zOwoC}$NgRO-S>*-dM=KvDXF(N3e~#LK6ja8j`qCuBH>D1@_oNA3*L;1T=wOX$Jzab
zt(QgLtqRqwjLg4QRK2dUu<)NmqQY&nWRuxEdvhLaKX~*(&(mWYPi+ZGn;aeZT5@gB
zkL>~dd@Dou_uV>>zUAg0>)NaSH7Awh6(@hqD2aN0aYea})Z6#HZjp{R`y_s!tXEfi
z?S6Ip?cXZfpTAhJ=-Qv@IezVL4YKsF{NM3w|D@0V6O}W5ZC-!;`dll9;>z44d0?JQ
zv#EA_)S2~(hrvI9A()$$+Z#*(Wgws^HLci^iGhLZ*MfV^-loCKJi($)qP}o#!D3Eg
zNIY>TaX8Ph{SCv<FD18_Sb}9ZWXxb%Sb|kKRN=f3KOyxO6Kz--d4hR3dH7+9U^F8a
z*g3t2elJ+_&{|=Z;g*$U&4wF*t`Kk%aEGbl305#yutD%N%{Ae?rMKTF8A&`;18U)O
z;zQCRXD$a<#1pJ;t`6tz(eTt<dUL}$ked{E9AV}j3bA_`FMs_w6Hl-di<A;fm?v17
zMH$Y!aMQZ0WXn`updw^X^PqV8>UOcO7Nr}xffgv5D|*6Ay1&=+Chx0+PM}kHop_O)
zDr+tacN|Z!nz<UB_tb#(3p>AfH^`}?Jn}G;-gkyR<Q10h2MTMMYlXptzijxpDZ25Z
zCJRq6zY{+^qIiOZorIBi^5*hRFfBa68s-{s-d};RO?UH;JOjE=&`Ho0p-9L{2#F`+
zBm(C#6s6`Qg3?lPQF5^|h!F%Lz<@beJh%EmpWPK(robSEK$tKCLvn6PP6m+co#xBT
z!pg?Z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%RzXoo
iMO9r>OGn?p(Ad;m&O}Pr$VF7l-7AF2&Ck(;(FXvm6~z|-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.B6F367ED6FE27214.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.B6F367ED6FE27214.idx
new file mode 100755
index 0000000000000000000000000000000000000000..643d64a7aa62c695b58fece978e6829945bc0e08
GIT binary patch
literal 1658
zcmWIYbaN|WV_<Mj@vO*AElFfyU|<jg;^LB`%ob(_1_>?(hKf07Lv!<Qn+aU|uX(O#
z@|kSjAI^_%t=_0+xJ&G(N5B(LC7+2>o<CFf*SVkW6IMOC!9?Si>$AIut83rA;89>M
za7~;Ops;6+TI$BX1!_h=EdEYDpSOOU{rMkXYb;)A-m{<h{PexnpPwG>SuPR#O~gM|
zCvU&n(YMEMe%)^W;=|kZlJRvrXMbM(G|5PX_4#};?|px2zLwT~f7%v*cfvE(2bW83
zF>;6Pl!_E>5prA~`Pl5m&JBm&H1YZD+EwWIgzYESUf08)!laLGnYt|W2iNcG(Nc4(
zAC{~0my1i@;ZA<xy*+!&M8AxwdIvgWd_5L@Kezn8>95!|TXZLH=$&_BxBi^<u11%_
z+&MD1tfJX1J#V=!sLp6E+ZKGf{Jq_t_9-{+tytd3Hg!YbY33(7tu-lcJX}~47cXAp
zkf<0{T(Ee<M-5k%h=nGHx;@GsGVGc*i=nrCYd8P90&h#VSmnr;7txY)`Q1BKEzZii
zJ6-X}F9D;oMK#k>GNT2q-Hd&Ei(&44;iN6DD_Yg<{7%(hu#Vtf-+F0MfajtQsk2{L
zsD!*0Sn)9+!-Okn{Z@lnJk=IcC1UTGR$Pc))e|*IsL4I+`VzASy|s(m%67SFlv=ej
zv|f0&;)BG6^(QBXO_{RH<d0y-+U+6Bn`XUOctvRO;l%c~Q<ab8-KGo3E(yJxt@&>A
zzP&D4+2Pl6b#hv^uiT~gR$0uYal&@jM!8!GUN%g*3tTQ+aIfCgC%^mW6SpI~3{4E@
z7BwfazMip7)%0aR_$x_I#{YR2y`&VEe~;r+y>Q`|?LPl`d>3C|>R#rh7v+8QO}&53
z%Dry_PiVO+iDz=Z-`!)h;Gq7r?w9|p7w0c}^&xH2^ex#Fq}9V&WH-0RoDlU@S4}Tj
z`}m0t^K?gH?Jg(1PRr|6PE)@LUt^mu*LYENvFt`w={UY~Y+sfL)lWWh`@Lk&=IeXY
zk8fA4IC<%u_4(N5yK+Uj&LUSF!pwBF_hx5au9RvlKKi-$k@S^R(f{R7%sWnQJ7v(C
ztvj*)yU=p)JzK82eJa{|`_$jcqo0#3syK`%_)a#I3Vb$q&BN^nk3Q&mdTW!Hc~IJA
z>5Ydhb3=Z-&*0bV+-!a}|AEo1qJPToy1G9vuX9!le_Xi#_v1qG6dSvH(*2v8PRSd8
zYyKbQvraDLzWU$5x#urBy2kuHlk-dYM+HNC!s7Esqo4gf{f6tU>o@)EudB`#SLP<k
z1M_5<UHR9#q@;N~4E_NO!91)y;a~zN0|7;;X~m383=CYq7Tjz0HVtOx2^O&yaf53M
z7PS^d;)z*{!Fl`*SA)Axtqx$~36|!Nc1CDXQdWwC@w8vZ2?V)2)B(*EuoiHD3G)QY
zxyX4UcxviuIxwDN`x}OzUrKH<v7ngC5{zuFtH@iyg~H5LKy&%5`H;+&agjkXS4CX~
zZf=O5kot>>Hmr<1!MvQj{4ncaG$R){WO@(%Ua;n&wL)`(l-p`U<1(P{1g!<_VXAn7
zg{_6*ymqFZ6^kA(6#*J4?;;-oQ?y{h>L=_wC7*(9QsA|M2`|0<KFLVpp&HQ5{MP(P
zZkBbCMRK#Mx+>hwheGUL#>-zn4s?Yii=-0FB%WYJ7DYJksR8R3c7E}0kaeQG@-Rg|
z8$Ye&HnG0~6joPP_kaoi6$smOH~+{pps<j&kR42zp(r&c5tJs1i;|0#L5v^}0S3&$
zpC?W-;qXgVWC{#o2!shUFeK-u<YWN3-f6zfEUawo9Gu)dyaN1uo`SB<YJyyzPD0|s
zA`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*ZR8-Y9wRH3i42@0A<xHe>ja)><+`U4W
K-25Cp7<~XtZ?8-M

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.BE1B4F87BC348C49.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.BE1B4F87BC348C49.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0732144486a019e863162d3bf262ead834d74476
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%sOTU25v3}hKf07Lv!<Qn+aU|uX(O#
z_L*$nAI^_%Wp7Y3+$HwYBjAarlFLLX&!4IL_qm_$6IMOCK}X`7>$AIut83rA;89>M
za7~;Ops;66TI$BX6>3I5O#aW+&(mMOPv^(gA{$8!JDZ8qS08MB`s&YL2g98b8uNDT
z*!Opu$oBbS;rHtoAO3DHR`d01-bwxN>zfTVE&kW+eY9GC`u%C;@%cZV@@B>{*`0rG
zTX0tEy29$F0HqhE*>~A&#lvRAUTE~ZzD2L&j>7tj(Iz_Yw*>k`XIizc-%!7I>BXJm
zb>*K5%I6*2YtXi@WVfUA%a*X0Og$$qwy0FSulW_b`2JC^+eVzXUv%#Iedwo~igIpg
zOT)4Sn@cB5UXshQtKBRky?SfS_jixkJv!g^vPoqGa!PNtaX1%r!u?pEONfA()hR9!
zu|>x0=3LSJK`u?AUdnqqQpEWdyk5b1q3ZgM9}oDpPf5PQJIS?U-{gxw3tCoKty*@`
zS<)$oPw}$O@pR8=+n5vQ-Zfpvec`vQqFZ!}Tg9WtOa5^0ZQlAJN<)`(stQ+pS!bWK
zQ!ihm6o0GRVui0!+>1{ZGz52EJQ~~0zIW-V6{?zw6-&x;vzT98vdYR>pWPa?yqF_^
zBe##MpYiW3n@m^FOs~W8jW;U2OSLjO<h*nZt>nIEteB(zFR_U6qGQz6z0)4qy}s?e
zq-@)^yyQgXJ-%1JX3B3=;A+{U&GqF1Yv3cn7fh<NjzvaXE&p@lth%|9+1iDjFQ-b2
zn662_HSNk0=h&6$Dy)AtJ8NgQ%(+)>ImN;G_4glkl~yk6z4dQt?R=ply7|BD$EDvk
zxht*~>5N&TyXU>}sRp+F5wYjL-)F5A4sAD|yk|*l^Bs#w;gU0QMT*(8{JhS+a65i8
zl1;DKEKH;|MyPn*=N7HE>1(*{Y?&r=yICiBo9&aB+xx;(^<P@&_ITqPY3r-c_vL$k
zn6z|m`Mlj?zipq~YSUQJxF*Lstm^v3WuHx$-iw|$@5&FD9{T6_#QO?rx3vz7tlR4H
z@3QwCy^oi|q)(o%+CKGn@!9T=PZUpiY|A~C;WV>+Q)9f_^To!~vQKN>^_qD$-=)_(
zYQ=}|2jyafQuAltJCVBO@gLqjtv=T-zufTna_0LgE8Bas{Wdpc%`^I@|9|C-HS!_$
z>c0X@&tD95W%)b(CeQJAA1~}ZaBxE7_A|f#Z<61_eUpD}>9W4!%G@M*U~V)ITVwEY
z#VcnX2LAwtU>;T;UoZibfq<gawBjF(3=CYq7Tjz0HVtOx2^O&yaf53M7PS^d;)z*{
z!Fht~JEcn3y*dokB4{mWhtMKqEri4qwibr-v|q;w1i3rZ0ksHN3pgOO$T`b-Ab4u(
zYC13;jArBlJG}SM?*(ffS}QoVzhU_KrQ{Y9ORzMDv=dAfOR$o%QY?(;D)Lrvp)hk5
z&^kVAJ|yd8oMn)#Q&CreTNmOdr2b-}4J#v0FfS)BKg^_drk)jx9xoLE3d=johr@&y
zOj!MdeW&D8kZJ{9E12-o+wYT%Bp#{(?d7-TN3vJeSr*A&RdrRky@x{VUdGE`KhDGx
zEXg9N1T%>zSdm2$&bxT;+6O%gl~$l4WKZ*;c>1XU>lb!@@otctM0w?5CjD&uw36Gz
z{t8f7U0ogSilWq<L{Lg7E=n#|1~Gy_1Q;*}cRqKmtNAF%&lDKM5C{`yU`WnQ$;kk6
zz0-V|Sy<WFIXJm_cm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f
n$SNo*si>-JYU$`37#f?J%b7^&8o7vyxqF2$x%oMIF!}%h4@Ih-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.F02158140F1CDE9D.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.F02158140F1CDE9D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9699169ee86b357a11b13c9a5de1efe42a1383ec
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%sOTU25v3}hKf07LvQEZwh*}bUnKe2
zwwu=)<b|%^le(?qw%t{ZvrERaMI*>NXy2Rb<ry~%7cGe};T8L>RQ#-9U3%L&1|x=W
zkz<^!hfZe&iOuF$S8DFRZ2$Akw^zSD_r#y~*Jtng?*HUxQ9<H98+&ODJ$<&W<$7Y<
z<0n7*_V~@O+wvtpzFqIp-+$-q&#Og-ma44J=Zkso`&09^wC?-Uw)ndfo~b^#Tyl$<
zJ8UOgq-cwf<NAo#*Iw-0a44sl&u7=NQpYE3mA&^CB~~u)O}gcES+l0~?`z%Ob9+DB
zUfjO>GtawL!!Jv3^WO4M-{K{n;L@k2a<TgP<@a2F&BAU?^St5W|KzUtIddh+%+<;)
zCaiPUHqTMXRlZ=m#nJlq>ZiNQ<?lIr<)oWjaOn2BvEr#*(KIJJqcW96O&hKShAr42
zF!$!}z#COll!Vp<NIi5_*}Z{1*4?+^+U{dV*~@n<lv$iNLC5LK+@5po%q~|0O?%}|
z2_$XipP6Q9=e=o*EXV6N+X`bF?!6Z@yrpDvNLb$esr?JtH7(bjGCfzQ1pe6MUGhda
ztAyi9g@%a~>&oj<625KBbG&-=@}lp&)(v&j_2fLLyfr&m+F|;&K<DaM?I~9CoEw}n
ziUTTmUW7mK)b{cUmiotV<XrTsU`OXK0a=|FFJG49o%;EReN3D~$%L@d<x6**|Mg91
z<=nMvqoX6Z^MgxkZ?WriFe$$8VzMo9oKnHC%b{~wp;*{gS^JXDCwn`5von&ao=Um$
zuAZ??)ig6c{FS69<NtRTyrdMDm&bimy>Q`|>^|>#d>3C|>fYw1XXSnLO<jM@%DHa>
zPiVO+iDz=Z|J`G>;Gq7r?w9YZ7w0d^`kyvw`j+eo((2hwd^er-QhYCao|L@%!ou?8
zfz%7fx)x3FPC8dQFJS54mePaE-5-2vy4ZJP(%iTnKesPSgz6_Bx&2-;XVdk)>FwK9
zD^6ZIXMH}l`>$M4uCvG$hp=t3+IwGTT&|RAG(P%y>!a)|siO5(pQMkZNT<#?G<DjE
z`ijoWs&TinbSoqC&5O#{RTkEL6G&CLZI*1Z@Qn4%1Nn<TU*tT!HFav<(lg7<7ag<I
z4f*jsfnTq4({`WyCq}ml|FPY3QhWWe;>O31JIZb4<lprkk90Kcm;8OSUQ6vY`_=x=
zx{$lYB_V>2|5MElZ>+W{(JwGad7%6G-~Ko3Z=Jr$&b3|ER$Q5zBoEAur);X1Z2q=}
zhljyGfFYQNmB$xM0A(PcC^fD42O|Rm*RKWln!QbfnR$XmtVP`5+JZ%`MUi-7)?#p;
z_Ukx-Aa{p4pcVmZ0SAN@IcGT!1W!#}O$WwvY=6V>^GnGsCYE4n4rwQtFiWtKvQjLJ
z=PL46aG@}B70_HhYd$1%Wt?S@%vDiWftwrRC#3#jq75q}PcSbhFF(wrcBY;ciyki(
z0Se1I%ZI~+7fe|Fgng&vQ;=!}UMraJ(%bKoj3gea0qy0t=0~zu)>#(GUR8BfxV?u$
z>|VypUq8;o6D-LhsRT2LCs>h15zd2&GID`~zW31Y1#2E!D?Bw|{ld;K-VJhvD6c$B
z)pz0Gsj9EGhyaa54k8|uAo|((X(hLb{S}~Ib#-;P<NgYSZMvI(<QdQhg4Tj|Fe7<_
zg{*~;c*54ga9&YrP9i8J6c;5ID}xw8AOZ}SgWtWLCH`qjB{NfC5JMnLn1LZVHzg+n
z$n{S1WoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!
oEhDR-sHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL!sO=X=)vd%0ByyxKmY&$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.9325EE22E541E919.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.9325EE22E541E919.idx
new file mode 100755
index 0000000000000000000000000000000000000000..076ed46fd7f6df789c5da07fe0790c82ea5968bf
GIT binary patch
literal 1658
zcmWIYbaN|WV_<Mj@vO*AElFfyU|<jg;^LB`%ob(_1_>?(hKf07Lv!<Qn+aU|uX(O#
z@|kSjAI^_%t=_0+xJ&G(N5B(LC7+2>o<CFf*SVkW6IMOC!9?Si>$AIut83rA;89>M
za7~;Ops;6+TI$BX1!_h=EdEYDpSOOU{rMkXYb;)A-m{<h{PexnpPwG>SuPR#O~gM|
zCvU&n(YMEMe%)^W;=|kZlJRvrXMbM(G|5PX_4#};?|px2zLwT~f7%v*cfvE(2bW83
zF>;6Pl!_E>5prA~`Pl5m&JBm&H1YZD+EwWIgzYESUf08)!laLGnYt|W2iNcG(Nc4(
zAC{~0my1i@;ZA<xy*+!&M8AxwdIvgWd_5L@Kezn8>95!|TXZLH=$&_BxBi^<u11%_
z+&MD1tfJX1J#V=!sLp6E+ZKGf{Jq_t_9-{+tytd3Hg!YbY33(7tu-lcJX}~47cXAp
zkf<0{T(Ee<M-5k%h=nGHx;@GsGVGc*i=nrCYd8P90&h#VSmnr;7txY)`Q1BKEzZii
zJ6-X}F9D;oMK#k>GNT2q-Hd&Ei(&44;iN6DD_Yg<{7%(hu#Vtf-+F0MfajtQsk2{L
zsD!*0Sn)9+!-Okn{Z@lnJk=IcC1UTGR$Pc))e|*IsL4I+`VzASy|s(m%67SFlv=ej
zv|f0&;)BG6^(QBXO_{RH<d0y-+U+6Bn`XUOctvRO;l%c~Q<ab8-KGo3E(yJxt@&>A
zzP&D4+2Pl6b#hv^uiT~gR$0uYal&@jM!8!GUN%g*3tTQ+aIfCgC%^mW6SpI~3{4E@
z7BwfazMip7)%0aR_$x_I#{YR2y`&VEe~;r+y>Q`|?LPl`d>3C|>R#rh7v+8QO}&53
z%Dry_PiVO+iDz=Z-`!)h;Gq7r?w9|p7w0c}^&xH2^ex#Fq}9V&WH-0RoDlU@S4}Tj
z`}m0t^K?gH?Jg(1PRr|6PE)@LUt^mu*LYENvFt`w={UY~Y+sfL)lWWh`@Lk&=IeXY
zk8fA4IC<%u_4(N5yK+Uj&LUSF!pwBF_hx5au9RvlKKi-$k@S^R(f{R7%sWnQJ7v(C
ztvj*)yU=p)J7KefJ{4`fed=%J(a%X1RUF0>d?y=91wNa*=Hd2(M<4V&y|u~9JSgq5
z^u|M$xgkH^XYlKFZZ<!g|G?;0(Ld#PUEQCT*Ey?&KQ7$=`*ERoijCbp>Hf`4r{s;l
zHUE$DStl2AU;S_3-18S5U1R>9$@!)Hqk<tmVe$E+(a-*#e#7<F^_zaq)>Y?<D|3_N
zfq7DS(+r;HWfJRo82kemf_YeZ!odVk1_Fvw(~23H7#O&IEx6b0Z5qtX6D(pa;s)0i
zENU%^#1pd?gY)<st_F9XS{=Z|6D-Xk?TpZ(q^uMN<H2Y~F0j45hkh?u^UzwsvHcCh
z&o3pnm{?HsvIHaR4e=9Fe=*U9m60cymy?$tW>j;6l-p`U<1(PVg4TleFkzlxVQXPH
zubru9#iGYcMSyzcUF0KRie?#ZSy|R>xB=)*K5IS)m@rSUjEjsXf~TUcq66bCn6UZ@
z`%cNHAXg~xTET>u-hQ8CB=JxUXcE6QKaxqZF0x1_sj92OO*#}}_cC7o`f;GcC0Qhu
zU?%YdE3zoUdH45v-sF9i&<QkGz*+#wTsapxBy-i&)!^nnHDLY1&M)2#a-k@%Jj~pm
zjh|L>o7i6g3ahKDd%%SM3WROCn}6gPP*})X$POmVP?VaJ2uc&hMajj=AVv^~00ZXW
z;F8Nbn*)3Om;!?s0%5`o49U4EIT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQu
zh=intrL~Q%oxOvesJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdrm8
KH$O)YMjrs&sjFxJ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.BBF8A6B4ABA52D27.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.BBF8A6B4ABA52D27.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d5c2ab70aa1e717b86b75bb9fb4e1975ab2aa30
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%sOTU25v3}hKf07Lv!<Qn+aU|uX(O#
z_L*$nAI^_%Wp7Y3+$HwYBjAarlFLLX&!4IL_qm_$6IMOCK}X`7>$AIut83rA;89>M
za7~;Ops;66TI$BX6>3I5O#aW+&(mMOPv^(gA{$8!JDZ8qS08MB`s&YL2g98b8uNDT
z*!Opu$oBbS;rHtoAO3DHR`d01-bwxN>zfTVE&kW+eY9GC`u%C;@%cZV@@B>{*`0rG
zTX0tEy29$F0HqhE*>~A&#lvRAUTE~ZzD2L&j>7tj(Iz_Yw*>k`XIizc-%!7I>BXJm
zb>*K5%I6*2YtXi@WVfUA%a*X0Og$$qwy0FSulW_b`2JC^+eVzXUv%#Iedwo~igIpg
zOT)4Sn@cB5UXshQtKBRky?SfS_jixkJv!g^vPoqGa!PNtaX1%r!u?pEONfA()hR9!
zu|>x0=3LSJK`u?AUdnqqQpEWdyk5b1q3ZgM9}oDpPf5PQJIS?U-{gxw3tCoKty*@`
zS<)$oPw}$O@pR8=+n5vQ-Zfpvec`vQqFZ!}Tg9WtOa5^0ZQlAJN<)`(stQ+pS!bWK
zQ!ihm6o0GRVui0!+>1{ZGz52EJQ~~0zIW-V6{?zw6-&x;vzT98vdYR>pWPa?yqF_^
zBe##MpYiW3n@m^FOs~W8jW;U2OSLjO<h*nZt>nIEteB(zFR_U6qGQz6z0)4qy}s?e
zq-@)^yyQgXJ-%1JX3B3=;A+{U&GqF1Yv3cn7fh<NjzvaXE&p@lth%|9+1iDjFQ-b2
zn662_HSNk0=h&6$Dy)AtJ8NgQ%(+)>ImN;G_4glkl~yk6z4dQt?R=ply7|BD$EDvk
zxht*~>5N&TyXU>}sRp+F5wYjL-)F5A4sAD|yk|*l^Bs#w;gU0QMT*(8{JhS+a65i8
zl1;DKEKH;|MyPn*=N7HE>1(*{Y?&r=yICiBo9&aB+xx;(^<P@&_ITqPY3r-c_vL$k
zn6z|m`Mlj?zipq~YSUQJxF*Lstm^v3WuHx$-iw|$@5&FD9{T6_#QO?rx3vz7tlR4H
z@3QwCy~5~Rual>%wom<Ce75`J6U9>=+j5U(IL$2I)EMvfe6jJg?9*Czy=I=xcj@(x
zTJhogLAe;A)cl$EPNZ&m{D*f>tIxH|FE>2CocX@W%J$xDzs*fq^Nhaf|6e&{jeLl`
z`mey!^A`hMS^iGH$#eYO#|wK89GuX&{mk$Go8-4}-{fCsx~#9bGB-&cm>UmrKC6s>
z<^GO`!9Rc@n1_|e7fb+UAfPBUt@sBc0|VEu1^1f0O@o<vf<>%F+~C@RMXg1Vcw*LK
zaGv1$PN~v$uMPvX2wDr;A+!it3nB4@t%cz{$M!c2KfjdRVqyuF=8$%RX<-RgQdWwE
z@k0EB)L%@rVP)hA=H=w&hY7bc^{iO*c&P|bSl(Ga940)=aLdZFX2T6YSMXW$IlzQ@
zf@Pd#JP<q;brl^LZ^4AsPuO=#J_T8zz-t8)UV8g|l99wiHK0lS*8E5&$vVp-nWU<&
z3O5O+m5~b^s=bGPFIe-?TH#QL-OG6S>&KaRf+blbm0+rPf)!a5;k=9Yu6@w6P-z7!
zLJkWal(4wJ*YhUttAtLVn*^){klZBaEQjPKHFY((o1Pl5eqrYq?*{o!lvf^R`On5r
zE4fYVuK<PB)z#s8i&AqEK`EiQD7jb}#0UZrV89&Qx+%7J)@k``Oo2fRfiPhPhUDCo
zoD3k>JI$Awg_Vt+gOi(wSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!RNG3~OhX=<
qsF<{jtb(GFimJM%mX5xGp|PpCoQagKk&CF9yH^O4o1dcxqYnTE2)l~_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.CE7231939A842BC9.idx b/.cache/clangd/index/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.CE7231939A842BC9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f444d4b643a6a4da69705bfc4752880721253729
GIT binary patch
literal 1642
zcmWIYbaP8$V_<Mj@vO*AElFfyU|<jg;^LB`%sOTU25v3}hKf07LvQEZwh*}bUnKe2
zwwu=)<b|%^le(?qw%t{ZvrERaMI*>NXy2Rb<ry~%7cGe};T8L>RQ#-9U3%L&1|x=W
zkz<^!hfZe&iOuF$S8DFRZ2$Akw^zSD_r#y~*Jtng?*HUxQ9<H98+&ODJ$<&W<$7Y<
z<0n7*_V~@O+wvtpzFqIp-+$-q&#Og-ma44J=Zkso`&09^wC?-Uw)ndfo~b^#Tyl$<
zJ8UOgq-cwf<NAo#*Iw-0a44sl&u7=NQpYE3mA&^CB~~u)O}gcES+l0~?`z%Ob9+DB
zUfjO>GtawL!!Jv3^WO4M-{K{n;L@k2a<TgP<@a2F&BAU?^St5W|KzUtIddh+%+<;)
zCaiPUHqTMXRlZ=m#nJlq>ZiNQ<?lIr<)oWjaOn2BvEr#*(KIJJqcW96O&hKShAr42
zF!$!}z#COll!Vp<NIi5_*}Z{1*4?+^+U{dV*~@n<lv$iNLC5LK+@5po%q~|0O?%}|
z2_$XipP6Q9=e=o*EXV6N+X`bF?!6Z@yrpDvNLb$esr?JtH7(bjGCfzQ1pe6MUGhda
ztAyi9g@%a~>&oj<625KBbG&-=@}lp&)(v&j_2fLLyfr&m+F|;&K<DaM?I~9CoEw}n
ziUTTmUW7mK)b{cUmiotV<XrTsU`OXK0a=|FFJG49o%;EReN3D~$%L@d<x6**|Mg91
z<=nMvqoX6Z^MgxkZ?WriFe$$8VzMo9oKnHC%b{~wp;*{gS^JXDCwn`5von&ao=Um$
zuAZ??)ig6c{FS69<NtRTyrdMDm&bimy>Q`|>^|>#d>3C|>fYw1XXSnLO<jM@%DHa>
zPiVO+iDz=Z|J`G>;Gq7r?w9YZ7w0d^`kyvw`j+eo((2hwd^er-QhYCao|L@%!ou?8
zfz%7fx)x3FPC8dQFJS54mePaE-5-2vy4ZJP(%iTnKesPSgz6_Bx&2-;XVdk)>FwK9
zD^6ZIXMH}l`>$M4uCvG$hp=t3+IwGTT&|RAG(P%y>!a)|siO5(pQMkZNT<#?G<DjE
z`ijoWs(GtRH7g_Y&5O#{RTkEL6G&CLZI*1Z@Qn4%1Nn<TU*tT!HFav<(lg7<7ag<I
z4f*jsfnTq4({`WyCq}ml|FPY3QhWWe;>O31JIZb4<lprkk90Kcm;8OSUQ6vY`_=x=
zx{$lYB_V>2|5MElZ>+W{(JwGad7%6G-~Ko3Z=Jr$&i%Wrt++BbNgkLRcdc$e{`MSG
z2M>dP07EbjD~~Uj0Lnl>QEFQ84@L$Cu3ro8HG7)|GxG$CSc|y9wFQe>iz4yFti|9w
z$M!c2KfjdRVqyuF=8$%RX<-RgQdWwE@k0EB)L%@rVP)hA=H=w&hY7bc^{iO*c&P|b
zSl(Ga940)=aLdZFX2T6YSMXW$IlzQ@f@Pd#JP<q;brl^LZ^4AsPuO=#J_T8zz-t8)
zUV8g|l99wiHK0lS*8E5&$vVp-nWU<&3O5O+m5~b^M7@W8FIe-?TH#QL-OG6S>&KaR
zf+blbm0+rPf)!a5;k^5MJ#X^9O6UYyCtxjrWSyL|9Flcv>S}Q7o*J-zVdod`202`m
zR~}~Wcj4fvs;{<)0Ifp~RUVX3{n_|wCAW$F6`)>qb#=G}e+9xe-OWGp4Co3$Ye74h
zkvzdd)<Q@;VQXPHuP8Mq5tI^&i;|0#L5v^}0S3&$W*mnPOw-NhWC{#o2!shUFeK-u
z<YWN3-f6zfEUawo9Gu)dyaN1uo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?
qM8%|KWEB*ZR8-Y9wRH3i42@0A<xHe>ja)><+`U4W-25Cp7<~Y%h`|j2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp.5A5CAFE59DAED728.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp.5A5CAFE59DAED728.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d55f2b519c6bebb6eedb942d42494dbc0974e844
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%sOTUhNqki3>9;ZhUVtqHWRq^U-R6S
z*=MqOe>gw7wR)qP;V!YC9v7Z?D!ELQ^8A^)f1mrwK4Ep~4JJ4Co!C=dQGR~!MM(#?
z2cPa(EOe-d^y{wqUZ`R8!{qne^ZD!7?LYJLYm$jwa7~4heptG=*y`@@>0KY6&V018
z`{~>5OLqJ<+5hjKwCukBhR;{8FXyh?8s+D{<iVfMMMuNq*X`4Nx9{GE)zX*tu++>u
zTa|FqD_mK-IY8+JZ}uHATk)_NwHF#a!_D<NwmQZw+@5h`*Ht&e+@+bK`wsp4I?ZkU
z-VeVgx9{%edDm+AW$A6xTOR6Lyq+f<k@EFe^!?oO`@Fwub++nG+R#03R{4DE`0kV?
zYx@K)JSo}6cTFWzy7ApY-di{0bnox|{ew|x<GB|;%)DBOi>I-lh~)Zt;>N@du0vt&
zkxhqUwmf^_miRcZ%VUF6W{dd5H-{KYbVV7Z-e-x+-+9<);TEeL+43S<axTAn$EL+O
zS+#o39bXkwW<CAUlad%MaP4O8+gl8K?+Yhwab3~6yl&2v`i12i1mn1usxF-5^03=%
z>tlx%*AxRDDqhbJSrV0<Xd*f9;gqD3!ut;#^EHjNJe64budMS-Wr&S%XMJ1R8}O=x
zkFjXs*}#W}3+_))U)wQ7P3j-Vk!#VXf*pO|DrT)*yg0DG?Na3{`yPEoE05K;GBw|A
z-nIA2m1|+wa&>cBw`=a6_f3#n$vt9wWuM$F1uh$=T!lrMb6P`dkM>`!DN;_lD=}+E
z+)mfPgKEoiJzcGyOwDo&I$$5YWs=*8m*-9^vPwnWe<)jN<+7ezf0ohC3|-aBf9L&L
zYISq*gw!q<y-Q;EzV{d{ICwt%bmaHBm(91n`j9qh`j+esW`1ETvNsP$o#<LN`GjQb
zYYW-52Syo5o`EM+OV4ec7qIm2l&uGsyC>9~%y7>+=r+CiVZ*$TSNBc6zP-*<cJuYU
z&5v&j?McaeKKJ>$gLV8nw<&UIDQM4{CVFq{mX|iY2da;LUi<61V!fKg?Mo>Jt<t&^
z?b|DxDicqyed4<H{>gVSA8SmWExx-#((Z@+iDf4xcvV09UHfqQL6SjAc5dWKY0b?y
zdsh@VT8sQsywR+mb$+&x^o|~}S^FEePdJ%nWV!xR%B9`&eEPntP1aeSqu%rP$$g>8
zSDRnGpZ!~7_wyGEmR$QYJ?EVJTIns~UG-Xjql5nbzuw?x_nXn%XG%(OWp0u@FgI@e
z9CFwBpJ+A@gMR=+FgGi=FPH$zKtNGyTJar51_rKQ3+^?0n+CJ+1dBR|dc(B^i#doP
z@x&d(k$4Jz3Xw3Cg6lh_O4q$Q%*+!k=pg6}6XppPau7n|i8zSBdD^ez1cKZh>X>+f
z1snui5nAN@<bq*5$M!c2KfjdRVq^)HQkL?C3A>8C6<jFHTm{t2>%fboSK3b+t~bO_
zNd3h`8&*c1U>;5$ewg0$bEaF;BsyLIg^`V1di#Bnk;FqappkqId`L#h_{qSHgqgv}
z1rEpFL%$cSd1$S$pKVKu^33dwKqD3X6r*6O4u#mgjF-QD94IWwBB=xuermw_g`Ho#
z8{{rg9(kDXmkl2`MK@m51jYfs13ywY3p)rS@nrpE;i1kGEbk`|=lyos70P^VDl^Dj
z1s;2txkaftiJ;gnE=n#|1~Gy_1Ox=a7!1t8M&~OhcWCkMU<wRkNX|{k$pFfEr};9o
zu(GjpaB}nT3h?uJ3c5P0337Ql35g4fNJv^(THDy#**oZoYFjCrX~<&}6_b{cRZvt?
iQB~K}($P0CG&VJtGm+9YauF4C_X=Th^K<lI^Z@{)>%2n%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.B9B0155CE4F94655.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp.B9B0155CE4F94655.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6c57d54ea50d7e45cb65ddf3a1ac158c00b1ff71
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%m!u#hPRvy3>9<EhUVtqHWRq^U-R6S
z*=MqOe>gw7wR)qP;V!YC9v7Z?D!ELQ^8A^)f1mrwK4Ep~4JH!bT%X-Nyu12c?>Pn|
zhUw3bak3sdT@>|Xcfa$*gUfHeuc@mp{hNAjz4+pfGanrmuH0y~dqYI*9{b=P{qrLJ
zu{!ztRgb=1{`2ei`!7EH?YE4t-5dR3_0=LPCD!Njr)%%~RrEEt@cY-c_`4IHsXq9e
zbB)<MY?oA|WQ&kve#GN#FLrJ?^ro55XV<Pm$2Gj4xc0gpzP8q^D^uHd^#{>!>$jU!
zy=}ce-{wwiN`ZdQLjAkzF1lGQ47FlczLw!6w71;uZ{CIX$t!aww#0t9vgh}mpHV8x
zxuq=)YRfm5Zcy{Q%~8Uid*R%-te@}h9_OFXQR<d1l@Y)xz17CyT+j*kV|^|m0^6)k
z35h&gWXf*NCEXw7(j@Aoyr&~YoNqz)N{$N^;k$o4;4433_)6-elFR*5FKR5DrpTrS
zzb=+^TEnk+S?74V3is`%4d>ovUuS#qcb>pvT_;_euaB4f;ojTa`yon0mvgELSNxWh
zvg}S<`5L+Sce*WBxEjU1_~Zwk<syr_f3xvddCv^=6jH2MQs%40{Nj+6R>u15mZ0Uu
z922;6`?&fU|F+p=x-K!f_`zLa+r6N9Tv=KCf!?yRy;UYt&ISKZtl_-q7<JXpyKw&1
zw@a68+qNk$J5hU&@6E4U`XdA^5A8Zz{N(~$-~+)7W<lS_I%{5O*R6RrSz7S*u?t7O
zoYL*$GBeEe*1Dv)KE!wu)4!XGrkXhUmB;aEUbyg!@BN9(idC@{7pI(b&DNax?)w$}
zt0(1Jb%IrjCT{z%FMSFlZ{5DAbGzQp*|##Z-F)(%C9%z03pXjfI3ah<Da>T9%efa(
z$6p#SPj?jd?sC#wS-kFRi`LuaYk2K!nI3z)StoIu?ULbP|FR_L|FV_&`_nd@j{81O
z_Wra7s@~S`ZOVDSRiD_(8yLVCd2!2{59^ltKF?&ir#&zK>VCeT#s_aqQ)|||lhOId
z@{I5q?&v&K{Z)Vc%AR*UUu-D<)-CZ0kGv93%H@=vW0NY?vwx%(7)or~x^2ys-YGZV
zxP|Okpv(1N;N3xS(-*xfd1KtX&)z>0e<H<n=A7$QMla|3s~xYN>^Uv?nDDW>Pv<*N
zzH<EfUAJz^{o;~<pz43#X3v?g_ul4yYA^a<cjZ6%X^m|5zZpNKPWw_^nVTdJ%#V?6
z_6yf<Kgh<z;2*#c%+1Q}4<>*z5KxqwR(yw%fr0DSf_u&0rok*c!J-bL-f(TfVh&<R
zJaGqcB%XqwLL^M3;QCIf(si#6GxG!sItV($gn5F69E6Z~A`T*Op7!fFfgpE>IwqcA
z0S5tBgcdnJxnLO2vHcCh&o3pn7+Hd)l%;%O!mc831s4i4R{{0%I`AUtmG+Z{>kaV}
zQhzbghLw>gn1_>xAEp;ZGjf5$qW9461#2E!E2N(@-I6BJ@dBtC*{1x7Yi8U^p0)}o
ztmvm01Jk_p_WL9wiHB-H2k<%YAvr+CPX_LQLm_rA<K?d(2kMn%kyL`|eQLn^g`Ho#
z8{{rg9(kDXmkl2`MK@m51jYfs13ywY3p)rS@nrpE;i1kGEbk`|=lyos70P^VDl^Dj
z1s;2txkaftiJ;gnE=n#|2GY#Ix-Qxgv;MNGF$D%O1VIu4hzvw9lXFvYGJss~G+$;G
zRyKAHPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fTdj~yHZ7XFn4S8&$V$w3Q3W`c9
hs_L3rI{F5N#-`?SCQ`aaE}~-YULj0wevTfDJ^-GHoWcMA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.F1D87E46323F05EC.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp.F1D87E46323F05EC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cff8e92891494f96237b51e0db83c11757a9ceee
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%t~ekhO?Xu3>9;ZhThJ*Z6R>=zew^i
z-&eW3KNJh!Ub~@YSmh|kIZMW~MI*>NXy2Rb<r+5&A1&d|33{{d#2m{xGmXuym`^ao
z%#&6Z;m|+rIWgaUPA89@-~T6@XP;iJZ(L_8-{&XtUjF3grzcx~e){vaGpVF-QN^32
z!q2vvd3NX4)zvII+@Ei}?@yKW^XaFLrg(6jv)7+i^YiJa-+#WnHr>BR?QHRZ=P$Rg
zi>)a!-DJd}+7!3>Snh@5#MYavvNK9bA2&rvJ`t_zZd)6bD{?t(R@g)CH}QFyUrurF
zx31b@d;Wo+gme6^*u}jiPOD4!+}2!ZQK@=g^DB1o{i8v*l{jy|SXo)V^Yc~@w_C3`
z7=8WHUnjUu+9voy?$(02Z?AqjJDdN#qTuTuKa&d!gv_#bw$E8Qfxkz#bA>`q$rM$M
zGp^a+@<hz!mUggcPjRadJgN7_@p_Qpf{*LUKHihf>rB34I?1(T-_(mg3mP=6Pc6H)
z+^8kuz4NljK7S9fZLEoN?_OQUd*QdOl3R33RK=^~fj`Zw*s~vQ4Tuub@)X&Z5p4af
zB}<k`RIaGk&2iOMF}KMNIsG)<#LIc#eVt~sc#?|K2Zy(_Lf9^}mWEu2yUww+@40}2
z_$?U`Ip$xcm6y6aG8aAMcgWeh)JF8m6}d&ztgK|dWoXP@{_pS$rbW$LR?VIE$ZqBB
zV&B}{)Vs$HxmC>G^g2_1gF;bD=<Me&7V<7Sz<hyCY1XmGh^yg$0?xXdD_!ed*!gm*
zw20`M<Xh81mej_rOjBX~v)WZVvt`b^J#T^+Ec_~aZ{l+2FS|Z0(wy9VEy(EZ`;GCd
zCfjmFF7tSzl>0FDyauz>pPHRB%I-_n2Cx2>uU6@~n@#(1itB|&+chm~GORl1T-@3>
zc{5K8o3D;0*Uly9c3u-+wPpSWvmakLdh{gTcIe*tkW^r{Kz-^jbI;n}%FVgoHhq8d
zVH=Bn<gLo*JF_?3cV3s<(ZqG4YguY*xo)nNzw(E$k2Q<`m7ciHk}N%oXIa5o)lbU}
z{0;PTc23;0qAdF6T<v-M?%(IJZ^|@sdwgQgWYay;bNJ8Q@QMz5<#sBkN>^*n<!>$Z
z9L4Xur9RKr;Ei$fJ`?{`J!O+rzkjxs+;_j@j}PztXmMKeuwnB0Kf+0yc7^}lHI@HU
zRTrOE*75q2&q_Az+E}&h&GAM5{Z{>de&KRMx;<ms$6jC#tjtZ42j)cEBd4}KOOOiY
zVek)N2<B$xb_Wwc83-s!O)I{`$iTq$Yr(x{Z_{8Fo?uZ2QE#}mU@-?VB%Zi~I1*36
zPazVfQgD5zROz}`hnabT1sw#PVZuDYLJmSmJP`*GI8Xa^oIsGfLmd-Ouz-VrD?*E$
zpIk7E2csFezz**{^n1aYht>*??Qa-<ekr-d$Pz52EaeMR<tp-4aG@}B70^gt2VNv2
zrTwJgMtZNCd3+tK+ge7RU`0R07?|D=KOyxO6Kz<5!aSTj{4nA4bEaF;BsyLIg^_Jq
zdi#Bnk;FqapiO)Zd`LFQ_{qR+Iuv5}GG6}raiCsF7D*+Tkxva+zp(R*cY|ya<&lR8
zf7$SHQ*`4+O<)}GJMbfgv#^6O5>M7o79Q$6!Sa6caNcj1U7^g^rZR)fRp7CQnOl^a
zlL(6K;-chYWgyKQJb#ZR)7;HT;!J@-3_*~D03rhs%;emZoD3k>JI$Awg_Vt+gOi(w
zSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!RNG3~OhX=<sF<{jtb(GFimJM%mX5xG
ap|PpCoQagKk&CF9yH^O4o1dcxqYnUc9HSBd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp.B674F131151301E0.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp.B674F131151301E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..43f502c7223d96dffaafd3d9ebcd052e52fe3fb7
GIT binary patch
literal 1612
zcmWIYbaQiIV_<Mj@vO*AElFfyU|<jg;^LB`%o=6}hKHOC3>9;ZhThJ*Z6R>=zew^i
z-&eW3KNJh!Ub~@YSmh|kIctfElgLWhpnY$ymuuWCe6)l+C+N+-6LT!*%rrK)Vm`r8
zGcQ_Sgk#FnB`3DapL1c8Uta&|=GmuLr%Udi#qX}(^}YYe&reSd)l~lb=3=<ZBG9I6
z<Bp&6M9k&W!|m-ZK0Ll%GQMu_+|Q?<CK;))KA$h{z3)%a&)T}*UuVU~O@3zi;B(0>
zX6~?GQJal8RGT6<AImo>df0N8Rd~ji(#Opak!Q7g{i3&J^@g6Y$c$bZpYXrPJ3CCH
zp!`cg^}K^~H%!}Cviq9eB~R<rq76w)b3K$69sjcX+K%p}g|0%|1Xos;?>ru*qWsvW
zWkIfA`twBBNri$h=9L|+xW4-7Y;*a0&PvxL=VmOp=(Qu`iQK0?Cp&|0Dp{<F$%}Lx
zU&d@Hd++v=ZL7${gr!SGHI#0)+P;X^VlsJ=E&u-ROWnerW!F*!m9`mG@vwJnTAY(r
ztmWMC)h=b$(;qD<52FOm?T&qWn_=&L<)q%r0o+UKWTw_<f7u|G(PA_yz;n@oMZRxu
z2(R+x2&p*HyNxSoZSRIQ9{&O_!(ThHzr?-QFxK)^V&T8C&oh-F*utOnZE0`7s|r5G
zqQp0jjaC^klhqe@Oi`4oU`Xoj`xFuoT-_1%>P5kg#S6c9{B6I;>cXEHYOQ;;Hfa0v
zS+{nb`q|su>o9%O>rD9#3PrrD&CZ)TNxLlMcaZMLkV=0T693WJRQ_(qt~?!IpY?Y#
z0uNd(%k^}<^?YiUThN92t+za-6!We0e~V;fY&>lJsiY%=OT>3(%$B}MFYEgM2F>k>
zI$`9hB%aCrzRFGV!h`3hk8b&W?#1k=t@Xwyy>FRDMCxld@x4h@O>w>K`DD#X-#K$q
z9~fsOc?O<PEj_p0V}+4LV(x)296fpxZ#%Sad`K!VTcAGmm$_%{Z{_CPZ=1fq`LK;;
ziq6|T?(d}UzE?EYb8%z^8nub3)_eB3%N%<`ZSJqAm;GdV@J5?jv*up2pwG(<{0;PD
zc23;0qAdF6T<v-M?sv3R-{0j=xt!9IHfiB>(>>C2_|M%qWf3;Z?Nm;cu9od`QLg_2
zukVSMJ}>_y^`qz7%>DNwPo$X2&7W3b^Uk*aczgBZ!fC<ll8?pxoPQ*xbp5Z=ba|EU
zBC=Di%=ka)?8`Z&!QXt|_D}skH{}0!*#(v7>zJ52raURG%uSL9=EX!)hsA~LRWUpa
z{s9cZ+^pO_U;-!u0Y#~4#djDP7`T2dxYz7$8qC5IEb1WY4c8Vd<{*Z|6L%0t;wktk
zM8Z@GuJ4p8UH9rRGf%LfgP=1^m?v1sK?sQ_;vfR&X}^vW2y%C*W8w)Ga1d}sXp!@i
z3x@F=+utz!{8Dm@ktJA4S;`kC>?-nBaG@}B6;Lm)122+ZX+LSWUeBLKp0{?}doc0@
zEBYzIof_gNr2b-}4J%NXhm(gNW+aSe<N^m#@1frd);zRUNIz$~B~7B^1yD7z1D4)?
zpJXKQPz~q+J_kM|2gvxzz#VWX#O`Ih{Pp8Ny^<`FN-!Iq8nAw0=NIn=IY5+09wz){
z!^ch0jTbe6alr4uj}*?r4#G%0SwC60S9pTu{p8`i-!8jCnXgS{2AQkCV-GX8C^aV$
z6x+o`$;HYbMi7XAfIt|7fjM~Qo-NB>%scdyDKLm3IX5LI11RU6=F7~&%Er#Y$<4zn
zz|ZF?=<2K{$mQuIBrYr>A!%W0ZDVU^@1Q5DZKZ6cA&*T|Oj<@(K~YIXRb5j{N8iBE
Z*wkFkL`v7lMO4h)D}>3-&(VX?2LROwe1`x4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp.E06437315FCA8871.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp.E06437315FCA8871.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8ec805915378ed5d6b04d579c69adbeaa96cd6bb
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hKHOC3>9;ZhUVtqHWRq^U-R6S
z**f05Kb#-kTD?)taF^Imj|(cQf{Q%fdi+e?zt8<-pRl_2g(Wxko!C=dQGR~!MM(#?
z2cPd$Ep%9*qwc!zuYrhU&76NH)3=9TpP&Ax%5IlMV2s_w=cgxge|~!NxA4)UKRX|-
z>^}PT`jQ=gEq4F=#xA?>uj%8{`sKxSTVs6OmOS`#IqBr|`Rn(cd%y3_ht<-T_OR5<
z@~cib=oOzG#<tMKAp6>nUN^lt8+Io!P2QJRBcQ#ya_Pm)(_w2jc1fz|uKuF^CO$v&
zEEns$KDl@2jCP1$yEHxT`Yl(h%;i3OVQaSVdb)g>QFmRce51)ak3(ikrtj^(M|7v8
z%;^-^a`xpLn>8LrdQ5jO>q)$?+FJAd-C<Rgq`4P6*`(ICbi2J%e6R6DIoW?vn1Gqu
zDWNInT(iIDP00)J7UT?zifUAv6nmq2-%=%q=6&Bj9-Mnm=y*n!ub5(z_Tu9O4I0*`
zmYFR#YKeI7yezWsyocC2)`xp;U){%h;kT{QH4EOY4zG>}{_J<-&u(<nQ0APf!liTT
zN}uh*TP+T)?RPR(D_p(Bk$vI=%W{#${kz@Xf9*3`JW0jrgU8$1A#4V%pF=LhUFTTZ
z_gp|h{Ft1`isoacl}242l8YMI6}G+5@MksM+P-3GU!R+;lvn?i`W+VSFBI2iEnm9h
z{Ht$4!O`KC`wKS+z6dI_eakH7vM%xJvz=}k%^F`AGnfT^AM4C{rCqn;*<|U!>%B`m
zUrv?Q)rvU2MR%pg?>#|hJ-B|o53}}Ee&YK)bv_fXZdrx<JU&6+#zmTwyRQWr-F?3?
ze%0ha?#N{xPn2>W)}Gg3mikk(b4J;H$=cx6-}2QeJ$JKdzfN_%&}h4+Wle@v=N!|m
zos$#U?3&f4iL}N9J+J9*(P|dH#^%$1;Hc$d*+N$7x%_R+RY5E3Ctulq-}pw_`s(w1
z`Q8pEFP*D?vhU(wxuVBTA|Vc8Cc4@ewr_c9(|e%$=;z*FFDK-49y2qM^m(~;;w=02
z%BISPr+S~bZoPf#T})x|N%z}<#x(`@6MQEd&hl)FyY}JsgB*ht?Z*+BQkt7z_OAFc
zCrsmy!nJ1oSLfD0G5sL9*0_H6<_RaWZ0vMDR@|wU>*xP|?6J=BnBzTrKiYSmeC7D{
z{oL9qcZy2_f|CDvn=LQC=6jQSt9{tN@Swl<%Nw5A{^nb|dh!;A;>z44d0=jw;C~}G
z=|W*Y4}*UILohchw-1;A%0NI-YFhCfMg|71UkmOvdz%Kc@C1uGh<d}d1&cX|A@Rf=
z#F2Okz6udAl|lvYE^s%!UC+o9ETt^v0~6MM9VZav?oh|X6D;5$;0hDw36}Gf3xe?+
z+utz!{8Dm@kp)GstH@iyg~H5LK)t*UyhwVbeWl@gL;Qr)Ure-N1=`EQ$-@sbGS$~{
zZ{x;gNz6RKf)0XiFkzlxAqOEOo`{18oCj0G$OR6f-b23^ta)gyu=Mu(BqNE3YCu=;
zIq)I5LdI7H?uz|vTT+x~W^V+#LeW<-5@yt)5WAQ0^4E_8g(X=em0-e84OqXh^NV+b
z+$_o?4-@{f;p3+0#*3Q3IN*2SM+#?Q2Vo?htgkFQ)OmvCedXc2-!8jCnXgS{2AQkC
zV-GX8C^aV$6x+o`$;HY*nmPDZ%-ITwGn&p!fk6yGkc0pt0};&R+?1RQAlEz1mzjl?
zjh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8RqLPZL
gx~7(nzJZ~!skxkql&+DBsF=G~2$P$iqX(l80A}E;V*mgE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.159F0712016AF057.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp.159F0712016AF057.idx
new file mode 100755
index 0000000000000000000000000000000000000000..34f2304c152954022aa6c0bfbae1289fd40741a3
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%z9=9hL@ZS3>9<EhUVtqHWRq^U-R6S
z**f05Kb#-kTD?)taF^Imj|)#cm0TuDdHzh@zt8<-pRl_21rv#HuFvis-d+8!_Z))}
z!}RCJI9U&!E(&^bo8Nij!Q(II*OXP4{;fQ={`ul=;iFH5EBD##G>D(K*FG?2&U}q|
zyCUw?`F7pz-~2ki{=&oC@y7f96q!E{KYcXCh3lNX?zEbpKVQB6^X@e7{#}Y^y$>vZ
zxrJS9O^N9yBM#N3xXs6MFBB)X-ei@XQBwN2DMIpzXjOOH+NfNS%VD#^9%{dd&(HjF
zihI9x)eh^^58@=8<9Ed_?k#azUBc(K=0b}~)%%)Xv5W5?4Z5wwdHco6%I`ZrZ}o7y
z^@@Yh*Dw8bg6pJhf-mH5EtvcE>ZiNA`KM3lD0Nep$_U_;-g?LRT+j*kV|^|m0%lgH
zghZY#GOae}lI{<3X%h8P-qVpH{%t|_N{$N^;k$o4;M;!0@Rig_C71iBUes7POp#3u
zejP07wC20wWu4>cD%`i5Hk^Bxb)D_SUpc|Wx=y+_Umvgd!@jq9>xU=}UCya0T=830
z%2qpV<!j{P-|4nk;c683;*%d(mWwRz-p%&C%6n#@r;uXBk}_W{<`;*ov@+Ibw*(z8
z=9s{p%g5Ev_;;2~rt1=uiyzn(w%rSw$CZ`EALuPB+goKa<y`RphmT|yHE&rpciJPn
z)wjKUb8}Pg9y{b#F?-YNO!*B8MJ=JTpTAhhyXXM(1vaHw$08%HhW`mTt8T7zt#@JP
z%c;^LqHB_GO$%948@Dn|h4s&FSMAJ}Iq&wo30|=9tL(jr%bmaM`mjiIa`&|$qr2}n
z#;=-e%N4oI<B3x4!`ky2%u;`9cFrifFIgMB`dhwQrRQ!o?Z+vu7aDEXw5-Xn>YQ_N
zYv1HVHoa!GX(FvLLC<4ebB5mXkC3jZV$nM%`L;v%#)}#QnG2J>{?1;q@9(^WX4QAA
z%POLqrt7@j^ZAbHjrWSz&0HK=Q(P}^I<$M5+1%xV73(YZ1=`Dg3UAuT>&q6r<C@UV
z;u-2Q#JAn?hzr@h_RTrb>(gtt$hHO=*Kin2@SSWp%j?<PH4nERyqRz!?AE3r^PsfL
z(t!`>Ow;(IaI0DW)%n{-(mQ&>X6<j?J>g`Qjh*h7iaWpM`uV>fd#tm(=6KKFNA|*#
zuQtEB|NFPb-sdkCEP3{4dXC@nwX?U1uc}}3H#+FAd=8Uu?KfGU?rAOz#g(~9^1%Ek
zCo{ETTDr+W9tQsahG1@1Za**ql!1Vv)U@I|j0_B1zZTqU_BIV>;RzOX5cP&@3l?(_
zL*j`$h$Hb7d=(;KDuoK(UEpqdyPlCJSV~#S2PO=o8M(mr_8$7ZV9i5o1?|^y0zvK$
zbxb_L0uBPMFjYLka=vmwFrH)k8-|}>N^UW-pcv^Y@>X!6Fmn}9FRueHl3r<FX}I1H
zKOyxO6Kz<5&f($Y;fEQS>g%|-apSTiW}aX{2SGQOFi)_MgAfu=#6bkk%b&Pr#;xRO
ztAHjc`YJ}lv@E^-KFLVpp&HN?d=7j_u8{GSfxF^Rh~3M0`Rm7ldL>yTm0)_G8nAw0
z=NIn=IZl*E9wz){!^ch0jTbe6alr4uj}*?r4#G%0SzlRrsPhEN`^v+4zg>2PGGCj@
z3^G@N#~x;GQEE;iD7K4>l8co=j35vJ2F$_1Q=~UFaVZus1qLw$!h{(Zl5<mXGJss~
zG+$;GRyKAHPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fTdj~yHZ7XFn4S8&$V$w3Q
l3W`c9s_L3rI{F5N#-`?SCQ`aaE}~-YULj0wevTfDJ^-z^kM;ln

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.4C82308AB1780753.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp.4C82308AB1780753.idx
new file mode 100755
index 0000000000000000000000000000000000000000..162b541d72508544a2227a0ab3a690668c0f483b
GIT binary patch
literal 1608
zcmWIYbaQiHV_<Mj@vO*AElFfyU|<jg;^LB`%yMQ1hLfBO3>9;ZhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<OOS`O2(u~9Dz&9JbrG9zrQf$c;{rXjG!(0M(KR%&(8FoV=!Wv
z|NI&^D^sea-lwy>g(n{TeR6+IRdMCl!c*(h4u4iYdQ`b`o5e1J=sEjrgXYYa)s)*k
z<9>~2*X-|?SFitn;o<A&n(;Ne=6*bVw8ln-^||~r?tMQhex{cFemN^XZt^qB2ZvKG
zF&BsZirQ?%q1qI=@z{FPqK7Sa*@S0&DSg}=5p`C(*DrcoPH*TLi_GYy@d^KnytAii
z6jXmGsGfCj?uKdmN_Jn<yX0w|TC^c)Wv++PqT^q7U)s^Tw9r+^OmJmo`Of1}D$0+2
zS{CH`r9V&foKz_IVqV$7itDSN&NhFqry%&cr_ba<hH8vS5qqV(qP#@4(AK66W`WZd
zl;}n6E?!*H9M$Eq;nF28k%?~(@qN)1Wt94|wg3COFKrfX!P%QQ1fwNw+iYEK2AXZP
zoThlBO5EsdQH|540$q;hcej1ZZMgTIapSSf6|6z_e5d}KRz+}{I7xc0Pzg-9xUB3A
z$5k~JEt@B9(X1=4xkWg)xmzrisJ&xWwY+?a<Wv>GCiblRON<&c=PYmfw#!YU)S{hX
zr$O0z2U(MKCng8FoD$%@!@SYey;5t%N?RA*(h`d|7Xzx6{5$+p(uF@W)LQpwZP51T
zvu^D=^>b_UR)^`EUT41Fs8Gav+U&fkleEji?+(%(8B*yFL*hTyn#$kp*p;W_>$5&4
zJMf^^vRqHsTTiEExdmPLzx9@<lw!V>{%(<sjE#q_Kb3T3aEbWNjM>sR>17@N-=MiY
zQ74RCmBcf--&eUwUU=}l`{$D1elKT7ZLKyw>3z#IB1&Jo>D!w`%@o(mo=?`S^p%;P
z^1wJF$usbTYU#P{9xIG2(sB=c;g~bu@RnfshKDyFq&WC%ea)Zq>svY7>^G_JZ~n{W
z)QG%Q*?-4;*L~-_n2si{6J5(vS$|D4o4Z`FVs*v7FnifgrU!4#Qft=SYZUZ(xq-id
ze$38kdsb|dzByMrPw$;9f6C>Qo@EmkPB;ByY9Tvk!_vYvCcRT`K8Xscayu<tC-C&1
z_}8b`4Wxebgw5E0%=bi!sg$gEiOH+lKDXc1ivy<x*C`*1`^kSKrF8wTI(hjeUqxi6
zT<N)g@>#{IT{Ekey=tHJ-)mKU!K2JwoAd8&>j_{euFOr62j)Y!YworG3|7nXF!%>B
z1aq@;yMYOy3<MOVrWM~|WMJU>wcuW}w`nj7Pq3(is5e|&u$Y4w5>MPg9Eqpks}KQG
zDOB+80(aBf^^82hQp!?3Fk$W2aRNc^4s}dC!2%8ft}tPqU^!p8AQ;cF{SCv<FD17a
zSy1%4io6wED9l_1)XVF@i=<cDR~oJtW(Fe{*kiqielJ+_&|1NJ-OS_bSl!kFoulZh
z7!6Yu;wPm3VxkQzP?(34haV=K>g%|-apSTiW}aX{2SGQOFi)_MgAfu=#6bkkTYCF_
zl99wiHK6189QcqNC*vyvcif>6yO;6u*N+3OlVp)pf|>i&fb|PIzj!yu6{0-yFySv7
zK5mL`yr>C`1AYg7q;M8?5Juw3`pUvXohMk{S02v$?XoMB`Px)wkhuyx_AqmcQgaeP
zv0Yr0T&xUY1c3-JU=DuwApGQ-Z4x0&fk6y`FkuFU<lK~;3?SD#&6k;lm5rT)lbeTE
zfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`n6!+nf})a&s=B6@j=q7R
Zv8lP7iIlF9i>R2pR|u1vpQ8t(4*+#Dl868R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp.872D44B53EF52F10.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp.872D44B53EF52F10.idx
new file mode 100755
index 0000000000000000000000000000000000000000..73c15dfafcfd75ed2c2d7b5ac868a5520573f68f
GIT binary patch
literal 1612
zcmWIYbaQiIV_<Mj@vO*AElFfyU|<jg;^LB`%xY!^hP#{$3>9;ZhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<OOS`N=A{2;35xQ&!3y(?=MU_-Z@z;BWR1hQ957xvopQt7>pRE
zKR?FFdgydf%#+#W>PpSuC)fY{^6b;A$5Z0Zwmo<6>URH>Q&^I?&7!t?iG{s&kWHE4
z{-2XX%-=5$kN>~$@b&Y-`+k&IKc3$Gv%-Vxoc$T`nvWkpoqqG}vg!UkYG;cNw4YqU
z{%lQ&=_Vr%)uy<O$JSmfPHer&COe~~^l?*!)DzLF?l#@%T#?IRv%(&7zlqPw{BntV
zziq`1+p`b+B%I@S#V+nGaavu%=caS9MWyO}&9B&n_m2hLR^pufVr6Cd&d*yt+-|+%
zVD$A%f1ThtX`A2+xmye7zP<YC?Cf&!2_2<w?ot^6oYGtGIGzhS!G5gIB}Cww)hQv7
zXNyd&&AFuegIt<Ky_ENKq==U-$X>~Dp(1>D#d|*UBZjY}PAa+FKlP&KhVvBJ)Zo|2
zl1^*N6))=?S6AV_-L&CcUiNji7k}jh7VA3c+I)Sy;E(*?=BXc|G;}$qs&K_`St)Dn
zw3V-si+`uvVuh<w+>1|s;8-rQxcfF+d6oCfKu;mXiUno9TFfsFS!rdg&u$57FXouQ
zo%@ZepYd;(O{VJ-lZzj!6}H_An#Yxu#UJP`E6Z1HGUZ(G|A&8g7Bz2KHFw%0yOp<#
zeRFeD@7_8TRWW<h>&*8X6^dF?XFq?jkay7m+Y4+;vyMeZTn(>@KkII;bgg$`=gX<t
zBBE=OZcPhW5*xQNO@;N(dROhtmO1bCya`&c@T=^-iOZe8?E0`sb8`2!AfvnQ7sjue
zY|9n7%;Skt?!(yg8q89EY7FO>+}ErPUR{>2R_VE$&HHhR>xFr?I<0FmtUBjh+}ihe
zGEWSfuZ||y&L!t|UK3umMSio{k1rf^<{RD;jNb6@=7SUmf32_ibAEj*XPf;d_5IC%
zxtwvEwtPNUd_7@*d(3SCR?!KfzNfj~N8Ku!=k{Rjqn}ZKOHN#8NzR_dv#emL>Zj!f
z{s#IvJE!egu}%8sT<vxGJA%FLRm-1Pc2c5O<+I<l57!@*6dam%Yg3SUP}=0|z=u6%
z8h;e7?A6aYr@zFsKvLJZe(U-PC$mhf*1gZTVlCJE?YrAzo#j66J$pa#cb<F|_x0~K
zyAy9Utu$8Z?w@@2<AhS>Z$2;iTmMfD`TvFK!iMzw45|00m@pJq<|fGl^WxOWOP3hj
zKM}#h;2*#c%+1Q}4JLpx5KxqwR(yw%fr0DSf_u&0rok*c!J-bL-f(TfVh&<RJaGqc
zB%Xq=LIg~uP{F$k+)Z!SGx7vWDNFglgtcGC2?V)2)G_e{3pfb4!i0H(<$UFWU_8h6
zHw-_&l-y!uLDB0f@>X!6Fmn}9FRueHl3r<FX}Dg`pGKaycH4UZ?N#(uggYn1Pe}d6
zL>pG1Fb^jWKg_07U&p<T8<!<9^8^bz2)e<9d4h!;gphb54kB<KObsI!IOKW{{a&!<
zp|!%&+wYT%Bp#{(9mnUuhvYaJUm3XL4u#mgjF-QD9Ow#37D*+TQBMt6zp(R*cY_=!
z$|DaG{<7iYrs&3tn!q^Vci=|~XJH3nB%Z9VEZmzs!ScTHaNcj1U7^g^rZR)fRp7CQ
znOl^alL(6K;-chYWgyKQ+@hp<D`ug<H>SWKh9F2n0Fi+RW^!&yP6m+co#xBT!pg?Z
z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%RzXooMO9r>
eOGn?p(Ad;m&O}Pr$VF7l-7AF2&Ck(;(FXu$1g?1i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp.83CDF2CC13C4A415.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp.83CDF2CC13C4A415.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d216ce4118f07c6cb21d6330b9737452139907f6
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%sOTUhNqki3>9;ZhUVtqHWRq^U-R6S
z*=MqOe>gw7wR)qP;V!YC9v7Z?D!ELQ^8A^)f1mrwK4Ep~4JJ4Co!C=dQGR~!MM(#?
z2cPa(EOe-d^y{wqUZ`R8!{qne^ZD!7?LYJLYm$jwa7~4heptG=*y`@@>0KY6&V018
z`{~>5OLqJ<+5hjKwCukBhR;{8FXyh?8s+D{<iVfMMMuNq*X`4Nx9{GE)zX*tu++>u
zTa|FqD_mK-IY8+JZ}uHATk)_NwHF#a!_D<NwmQZw+@5h`*Ht&e+@+bK`wsp4I?ZkU
z-VeVgx9{%edDm+AW$A6xTOR6Lyq+f<k@EFe^!?oO`@Fwub++nG+R#03R{4DE`0kV?
zYx@K)JSo}6cTFWzy7ApY-di{0bnox|{ew|x<GB|;%)DBOi>I-lh~)Zt;>N@du0vt&
zkxhqUwmf^_miRcZ%VUF6W{dd5H-{KYbVV7Z-e-x+-+9<);TEeL+43S<axTAn$EL+O
zS+#o39bXkwW<CAUlad%MaP4O8+gl8K?+Yhwab3~6yl&2v`i12i1mn1usxF-5^03=%
z>tlx%*AxRDDqhbJSrV0<Xd*f9;gqD3!ut;#^EHjNJe64budMS-Wr&S%XMJ1R8}O=x
zkFjXs*}#W}3+_))U)wQ7P3j-Vk!#VXf*pO|DrT)*yg0DG?Na3{`yPEoE05K;GBw|A
z-nIA2m1|+wa&>cBw`=a6_f3#n$vt9wWuM$F1uh$=T!lrMb6P`dkM>`!DN;_lD=}+E
z+)mfPgKEoiJzcGyOwDo&I$$5YWs=*8m*-9^vPwnWe<)jN<+7ezf0ohC3|-aBf9L&L
zYISq*gw!q<y-Q;EzV{d{ICwt%bmaHBm(91n`j9qh`j+esW`1ETvNsP$o#<LN`GjQb
zYYW-52Syo5o`EM+OV4ec7qIm2l&uGsyC>9~%y7>+=r+CiVZ*$TSNBc6zP-*<cJuYU
z&5v&j?McaeKKJ>$gLV8nw<&UIDQM4{CVFq{mX|iY2da;LUi<61V!fKg?Mo>Jt<t&^
z?b|DxcIs|#ed4<H{>gVSA8SmWExx-#((Z@+iDf4xcvV09UHfqQL6SjAc5dWKY0b?y
zdsh@VT8sQsywR+mb$+&x^o|~}S^FEePdJ%nWV!xR%B9`&eEPntP1aeSqu%rP$$g>8
zSDRnGpZ!~7_wyGEmR$QYJ?EVJTIns~UG-Xjql5nbzuw?x_nXm&XG%(OWp0u@FgF&p
zUYzTYGW`n=gMR=+FgGi=FPH$zKtNGyTJar51_rKQ3+^?0n+CJ+1dBR|dc(B^i#doP
z@x&d(k$4Jz3Xw3Cg6lh_O4q$Q%*+!k=pg6}6XppPau7n|i8zSBd5-OG7=C^!xy8s5
zETt^v3)2$fC#3#jq75q}PcRQB4?j#e{haBRG>MKEKw)InFlk0Eu*Z52{a&!<p|!#+
z!!0YznhiHF@dWca@VdfO@dQizNe9DtOK-nVGLm?x22{l7z=xzr#!m*WXg}MQ6y=%O
z8-eyJ`YA@i^d1Vadl@f({Wwrql0{MpCVYRd=S|*M37tSA1snvBjFj_}gB$tOfb|PI
zzj!yuhoU_4Fuh+keB2b>cu^A=2mB8FNZ~B(AdJM5^^=8%I!~~?pFEuR+htcM^R=nW
zAafOX>|y2>rRF4pV!OB~xmX#*2m%oh5C~&1Fb7xv+2s5oit#&BU=TxcZc0uDP|iEe
zmzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8R
kqLPZLx~7(nzJZ~!skxkql&+DBsF=G~2$P$iqX(l80NQTA2LJ#7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.A4A9E7E93B446A25.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp.A4A9E7E93B446A25.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6abf02a098514305c7b586d23a7692d77272cd46
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%m!u#hPRvy3>9<EhUVtqHWRq^U-R6S
z*=MqOe>gw7wR)qP;V!YC9v7Z?D!ELQ^8A^)f1mrwK4Ep~4JH!bT%X-Nyu12c?>Pn|
zhUw3bak3sdT@>|Xcfa$*gUfHeuc@mp{hNAjz4+pfGanrmuH0y~dqYI*9{b=P{qrLJ
zu{!ztRgb=1{`2ei`!7EH?YE4t-5dR3_0=LPCD!Njr)%%~RrEEt@cY-c_`4IHsXq9e
zbB)<MY?oA|WQ&kve#GN#FLrJ?^ro55XV<Pm$2Gj4xc0gpzP8q^D^uHd^#{>!>$jU!
zy=}ce-{wwiN`ZdQLjAkzF1lGQ47FlczLw!6w71;uZ{CIX$t!aww#0t9vgh}mpHV8x
zxuq=)YRfm5Zcy{Q%~8Uid*R%-te@}h9_OFXQR<d1l@Y)xz17CyT+j*kV|^|m0^6)k
z35h&gWXf*NCEXw7(j@Aoyr&~YoNqz)N{$N^;k$o4;4433_)6-elFR*5FKR5DrpTrS
zzb=+^TEnk+S?74V3is`%4d>ovUuS#qcb>pvT_;_euaB4f;ojTa`yon0mvgELSNxWh
zvg}S<`5L+Sce*WBxEjU1_~Zwk<syr_f3xvddCv^=6jH2MQs%40{Nj+6R>u15mZ0Uu
z922;6`?&fU|F+p=x-K!f_`zLa+r6N9Tv=KCf!?yRy;UYt&ISKZtl_-q7<JXpyKw&1
zw@a68+qNk$J5hU&@6E4U`XdA^5A8Zz{N(~$-~+)7W<lS_I%{5O*R6RrSz7S*u?t7O
zoYL*$GBeEe*1Dv)KE!wu)4!XGrkXhUmB;aEUbyg!@BN9(idC@{7pI(b&DNax?)w$}
zt0(1Jb%IrjCT{z%FMSFlZ{5DAbGzQp*|##Z-F)(%C9%z03pXjfI3ah<Da>T9%efa(
z$6p#SPj?jd?sC#wS-kFRi`LuaYk2K!nI3z)StoIu?ULbP|FR_L|FV_&`_nd@j{81O
z_Wra7s@~S`ZOVDSRiD_(8yLVCd2!2{59^ltKF?&ir#&zK>VCeT#s_aqQ)|||lhOId
z@{I7A+?~2A`m6r>l|AozzSvOyty|(19(g66l*=hS$0k*(Xa7hoFqGJ|b=#UNy;E+!
zaSPe8K$q*kz`KLurZ0L|^2WG%pS^!1{zQuD%sJPqj9$+5S36!k*>hU(G2vr%pU!ul
zeC7D{yKdc-`^6;zLDm1f&7Lz~@4e0a)L!(z?#h4i(;C_Ae=~kko%W@;GB-&cm>(Z0
z@OIsGI=zsG!9Rc@n46W`A4~vcAfPBUt@sWj0|VEu1^1f0O@mo@f<+xfz2Vw|#T>+t
zc;XJ?NIV5Ug-Do6!S$U|rR!cDX66YNbP#lg3G)OCIS3)~L>xrmJjeDo3_rh=++t)2
zmQt4Tg=q=#6H<RM(T0_gCzywmhaV=Ke$I4DnncG7pfIxP{E2I3+)AFd3Mj1Trx*iM
zJ<D**%Ccs|4NN@2ybipsFkzlxX+P;;7;owA_en+)57mH*_#F6<6v_C>z!e<|v3nUW
zfBiVn0!bE0C79m(dp&RRzDno>>J@MhK+-GcCkNLHQ_RQ(4)xwczZa}|Xsz(nfb|PI
zzj!yuhoU_4FjZeReB2b>cu^A=2mB8FNZ~B(AdJM5^^=8%I!~~?pFEuR+htcM^R=nW
zAafOX>|y2>rRF4pV!OB~xmX!UGY7wJcr$s<+r{-vfk6yGkc0pt0};&R+?1RQAlEz1
zmzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8R
kqLPZLx~7(nzJZ~!skxkql&+DBsF=G~2$P$iqX(l80O=K|%>V!Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.729E5DD280A9E536.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp.729E5DD280A9E536.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f1551a55494c824336362fb8a0397f1483ddcd02
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%t~ekhO?Xu3>9;ZhThJ*Z6R>=zew^i
z-&eW3KNJh!Ub~@YSmh|kIZMW~MI*>NXy2Rb<r+5&A1&d|33{{d#2m{xGmXuym`^ao
z%#&6Z;m|+rIWgaUPA89@-~T6@XP;iJZ(L_8-{&XtUjF3grzcx~e){vaGpVF-QN^32
z!q2vvd3NX4)zvII+@Ei}?@yKW^XaFLrg(6jv)7+i^YiJa-+#WnHr>BR?QHRZ=P$Rg
zi>)a!-DJd}+7!3>Snh@5#MYavvNK9bA2&rvJ`t_zZd)6bD{?t(R@g)CH}QFyUrurF
zx31b@d;Wo+gme6^*u}jiPOD4!+}2!ZQK@=g^DB1o{i8v*l{jy|SXo)V^Yc~@w_C3`
z7=8WHUnjUu+9voy?$(02Z?AqjJDdN#qTuTuKa&d!gv_#bw$E8Qfxkz#bA>`q$rM$M
zGp^a+@<hz!mUggcPjRadJgN7_@p_Qpf{*LUKHihf>rB34I?1(T-_(mg3mP=6Pc6H)
z+^8kuz4NljK7S9fZLEoN?_OQUd*QdOl3R33RK=^~fj`Zw*s~vQ4Tuub@)X&Z5p4af
zB}<k`RIaGk&2iOMF}KMNIsG)<#LIc#eVt~sc#?|K2Zy(_Lf9^}mWEu2yUww+@40}2
z_$?U`Ip$xcm6y6aG8aAMcgWeh)JF8m6}d&ztgK|dWoXP@{_pS$rbW$LR?VIE$ZqBB
zV&B}{)Vs$HxmC>G^g2_1gF;bD=<Me&7V<7Sz<hyCY1XmGh^yg$0?xXdD_!ed*!gm*
zw20`M<Xh81mej_rOjBX~v)WZVvt`b^J#T^+Ec_~aZ{l+2FS|Z0(wy9VEy(EZ`;GCd
zCfjmFF7tSzl>0FDyauz>pPHRB%I-_n2Cx2>uU6@~n@#(1itB|&+chm~GORl1T-@3>
zc{5K8o3D;0*Uly9c3u-+wPpSWvmakLdh{gTcIe*tkW^r{Kz-^jbI;n}%FVgoHhq8d
zVH=Bn<gLo*JF_?3cV3s<(ZqG4YguY*xo)nNzw(E$k2Q<`m7ciHk}N%oXIa5o)lbU}
z{0+80T{dCQin8dNbG7I3yMLd@zA4kl?eU2{lTG(X&*49J!z()MmD{PDDqXEPm%p{t
za}>YtmijzfgEz*_`%L^(^^{Fg{r=fja^L-qKR&$oqs3{>!-mQ0{|F~-+7<qH*Hr#b
zRb6~uS;y;7J}cR<Yh%^2H^&$K_gnS<`Gw03>Gq828+(B{urfDE9+(p+dpRV}DV7Z3
zVek)N2<B$xb_Wwc83-s!O)I{`$iTq$Yr(x{Z_{8Fo?uZ2QE#}mU@-?VB%Zi~I1*36
zPazVfQgD5zROz}`hnabT1sw#PVZuDYLJmSmJP`*GIM1>D4a3hbCASz^f~AzDd|_I=
z*Udb>j@4}~BTulRpJEJ5IK)p#{l!EZR-iBsCl5bNIQ^XImNbcu7eHZTLtxU3Twwq9
z9{Rmt%|mO2S%zCymNgr0VB!hpb>MY{sp1Ki_LB~V@s{3xpJXKQPz|Vv&w&p~k&K@V
zT+yKryO;6u*N+43m1L1rg6X}#*YhUttAtLVUI7OIB)xKea&Wy*4OqXh^NV+bydugY
z57YZ)!^ch0jTbe6alr4uj}*?r4#G%0SwC5LsPhEN`^m$3zg>2PGGCj@3^G@N#~x;G
zQEE;iD7K4>l8cprG;^?uu&&(MO+PL$1qLw$K@tLp3`8)Kb5n9MfL!l1UuG6oHg*n9
zZXR9%em+k@S7$XrE>9;RabXb&NefGB8(TYj2R%`3D`hhcd2FI$(lW9Nib^V~>Y7?Q
c`UZx^rsi@cQo2SiqGIk|Axv(5jvkCY0B&%Hw*UYD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp.649F61751C48D48D.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp.649F61751C48D48D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..55c61152843612c1a428cb8e90573e0adca4c1a7
GIT binary patch
literal 1612
zcmWIYbaQiIV_<Mj@vO*AElFfyU|<jg;^LB`%o=6}hKHOC3>9;ZhThJ*Z6R>=zew^i
z-&eW3KNJh!Ub~@YSmh|kIctfElgLWhpnY$ymuuWCe6)l+C+N+-6LT!*%rrK)Vm`r8
zGcQ_Sgk#FnB`3DapL1c8Uta&|=GmuLr%Udi#qX}(^}YYe&reSd)l~lb=3=<ZBG9I6
z<Bp&6M9k&W!|m-ZK0Ll%GQMu_+|Q?<CK;))KA$h{z3)%a&)T}*UuVU~O@3zi;B(0>
zX6~?GQJal8RGT6<AImo>df0N8Rd~ji(#Opak!Q7g{i3&J^@g6Y$c$bZpYXrPJ3CCH
zp!`cg^}K^~H%!}Cviq9eB~R<rq76w)b3K$69sjcX+K%p}g|0%|1Xos;?>ru*qWsvW
zWkIfA`twBBNri$h=9L|+xW4-7Y;*a0&PvxL=VmOp=(Qu`iQK0?Cp&|0Dp{<F$%}Lx
zU&d@Hd++v=ZL7${gr!SGHI#0)+P;X^VlsJ=E&u-ROWnerW!F*!m9`mG@vwJnTAY(r
ztmWMC)h=b$(;qD<52FOm?T&qWn_=&L<)q%r0o+UKWTw_<f7u|G(PA_yz;n@oMZRxu
z2(R+x2&p*HyNxSoZSRIQ9{&O_!(ThHzr?-QFxK)^V&T8C&oh-F*utOnZE0`7s|r5G
zqQp0jjaC^klhqe@Oi`4oU`Xoj`xFuoT-_1%>P5kg#S6c9{B6I;>cXEHYOQ;;Hfa0v
zS+{nb`q|su>o9%O>rD9#3PrrD&CZ)TNxLlMcaZMLkV=0T693WJRQ_(qt~?!IpY?Y#
z0uNd(%k^}<^?YiUThN92t+za-6!We0e~V;fY&>lJsiY%=OT>3(%$B}MFYEgM2F>k>
zI$`9hB%aCrzRFGV!h`3hk8b&W?#1k=t@Xwyy>FRDMCxld@x4h@O>w>K`DD#X-#K$q
z9~fsOc?O<PEj_p0V}+4LV(x)296fpxZ#%Sad`K!VTcAGmm$_%{Z{_CPZ=1fq`LK;;
ziq6|T?(d}UzE?EYb8%z^8nub3)_eB3%N%<`ZSJqAm;GdV@J5?jv*up2pwG(<{0+*V
zE}O7tMOpOCx!Uvi-S23tzQ4<#ayg|ZZPLQ&rhBC4@SnSJ$|7u*+o_x?T`k+?qFnz4
zUf&ZheO~@a>POGDnfvcWo=7p3n?J3>=ACW-@%HM+h0}u9B_E6XIsZsX>H1%#>GCSy
zMP#R3nel(p*_U%lgTMK_?VtL8Zpi=dvI{ED*D*1BOnFjVnVTdJ%!`@M>T=8qzbEoA
z_y;fqbF*^$fC-=s1QeyF72jcGVBq?-;9j%0X)p^<u&9HmH(Xn=n1dJ+PuxKqiKpPF
z5D8N$xV}@Wblt1N%sjz@4uZ}wVV+<i2O%V$h=T~6=h*&+;pdl<TZ}BhQp!@kFfE=x
zjXZDdw)bG<30Cw|gj*2eC#3#jq75rhn1_>xAEr9}oavS{iH;XQVPr#K(u`bS|MnjG
zy<p8lYlT^cTUM4e8*X6Y3FdX+b%m+o36}Pg4u<iT-hQ8CB=JxUsEE&j4@r@XpA1~l
zp%A;5@$%P?1MQV$kyL`|y}#G<Chx0+PM}@^2LU9#a(;4fy-y8Tzp(R*cZ0kl$|Dcc
z`(?w&P0@`PHGy%!@4$}~&cY7DNIY3TS$LrE1k3x$!+F15c7-xuo5~C_SAoYKW^Pex
zP9i9_i;I$rl|hUk5CH*!Fa`s2aFb8$#%sCznV15D7?N{Sax#E&-f6zfEUawo9Gu)d
zyaN1uo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*ZR8-Y9wRH3i
a42@0A<xHe>ja)><+`U4W-25Cp7<~XanR-wF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp.8A0C33E6B1E03B1D.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp.8A0C33E6B1E03B1D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cc5dc9dda0859e16e3f64555172ff01a2a7d6a75
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hKHOC3>9;ZhUVtqHWRq^U-R6S
z**f05Kb#-kTD?)taF^Imj|(cQf{Q%fdi+e?zt8<-pRl_2g(Wxko!C=dQGR~!MM(#?
z2cPd$Ep%9*qwc!zuYrhU&76NH)3=9TpP&Ax%5IlMV2s_w=cgxge|~!NxA4)UKRX|-
z>^}PT`jQ=gEq4F=#xA?>uj%8{`sKxSTVs6OmOS`#IqBr|`Rn(cd%y3_ht<-T_OR5<
z@~cib=oOzG#<tMKAp6>nUN^lt8+Io!P2QJRBcQ#ya_Pm)(_w2jc1fz|uKuF^CO$v&
zEEns$KDl@2jCP1$yEHxT`Yl(h%;i3OVQaSVdb)g>QFmRce51)ak3(ikrtj^(M|7v8
z%;^-^a`xpLn>8LrdQ5jO>q)$?+FJAd-C<Rgq`4P6*`(ICbi2J%e6R6DIoW?vn1Gqu
zDWNInT(iIDP00)J7UT?zifUAv6nmq2-%=%q=6&Bj9-Mnm=y*n!ub5(z_Tu9O4I0*`
zmYFR#YKeI7yezWsyocC2)`xp;U){%h;kT{QH4EOY4zG>}{_J<-&u(<nQ0APf!liTT
zN}uh*TP+T)?RPR(D_p(Bk$vI=%W{#${kz@Xf9*3`JW0jrgU8$1A#4V%pF=LhUFTTZ
z_gp|h{Ft1`isoacl}242l8YMI6}G+5@MksM+P-3GU!R+;lvn?i`W+VSFBI2iEnm9h
z{Ht$4!O`KC`wKS+z6dI_eakH7vM%xJvz=}k%^F`AGnfT^AM4C{rCqn;*<|U!>%B`m
zUrv?Q)rvU2MR%pg?>#|hJ-B|o53}}Ee&YK)bv_fXZdrx<JU&6+#zmTwyRQWr-F?3?
ze%0ha?#N{xPn2>W)}Gg3mikk(b4J;H$=cx6-}2QeJ$JKdzfN_%&}h4+Wle@v=N!|m
zos$#U?3&f4iL}N9J+J9*(P|dH#^%$1;Hc$d*+N$7x%_R+RY5E3Ctulq-}pw_`s(w1
z`Q8pEFP*D?vhU(wxuVBTA|Vc8Cc4@ewr_c9(|e%$=;z*FFDK-49y2qM^m(~;;w=02
z%BG##=B-a$x86SWE~c>fr2Fka<C+5d3BHpJXL+{8UHfqRL5@L+_Tz|5Db39<dslp!
z6Q=P;;aao)t8?q0n0}C4Yh1s3^MsRGHg>unEACXw_49u}_E=|m%<-PRAMHC&zH<Ef
zes1lQJH;gdLCOET&6XEm^S#Nv)jsTBc+lVb<qgklfAg)&p1g&jxH30M9+(@OlV`4%
zIGZ5D!{8sl5X{ZW?E@x&G7wOdnpS*=k%58h*MfV^-loAUJi($4qTX<A!D0?#NIY=|
zaU`CCuR;V&rBK1U3*1d_*E8}2ODRkFz=R#!-!T0AQgVxt1w}Z-Pe}d6L>pG1Y93A=
zewgZ1U&p<T8<!<9^8^bz2)e<9d4h!;gphb54kB<KObsI!*kiqielJ+_&{|=Z;g*$U
z&4wG8c!GHycwJ$tc!H&UrGsF+rMKTF8A&`;11jQk;6qX*<0}JKw4ZHDit^0tjX<X=
z`YJ}k^d1Vadl@f({Wwrql0{MpCVYRd=S|*M37tSA1snvBjFj`0gB$tOfb|PIzj!yu
z_o6)VFuh+keB2b>cu^A=2mB8FNZ~B(AdJM5^_7K(I!~~?uRNUh+htcM^R=nWAafOX
z>|y2>rRF4pV!OB~xmX!UGY5A|lrEKKsj6lQ3}Og^Bm@u{h+rn?rsQM*x!!5M%q*;I
z>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#K
fHMMl~4GfJ<&E-s_bd6j@#oWC@nB4pvJs5oeaxa)5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.8C4CEC479CC69615.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp.8C4CEC479CC69615.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1c122fdd680951b7416ef5c3c424c54fee3f546b
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%z9=9hL@ZS3>9<EhUVtqHWRq^U-R6S
z**f05Kb#-kTD?)taF^Imj|)#cm0TuDdHzh@zt8<-pRl_21rv#HuFvis-d+8!_Z))}
z!}RCJI9U&!E(&^bo8Nij!Q(II*OXP4{;fQ={`ul=;iFH5EBD##G>D(K*FG?2&U}q|
zyCUw?`F7pz-~2ki{=&oC@y7f96q!E{KYcXCh3lNX?zEbpKVQB6^X@e7{#}Y^y$>vZ
zxrJS9O^N9yBM#N3xXs6MFBB)X-ei@XQBwN2DMIpzXjOOH+NfNS%VD#^9%{dd&(HjF
zihI9x)eh^^58@=8<9Ed_?k#azUBc(K=0b}~)%%)Xv5W5?4Z5wwdHco6%I`ZrZ}o7y
z^@@Yh*Dw8bg6pJhf-mH5EtvcE>ZiNA`KM3lD0Nep$_U_;-g?LRT+j*kV|^|m0%lgH
zghZY#GOae}lI{<3X%h8P-qVpH{%t|_N{$N^;k$o4;M;!0@Rig_C71iBUes7POp#3u
zejP07wC20wWu4>cD%`i5Hk^Bxb)D_SUpc|Wx=y+_Umvgd!@jq9>xU=}UCya0T=830
z%2qpV<!j{P-|4nk;c683;*%d(mWwRz-p%&C%6n#@r;uXBk}_W{<`;*ov@+Ibw*(z8
z=9s{p%g5Ev_;;2~rt1=uiyzn(w%rSw$CZ`EALuPB+goKa<y`RphmT|yHE&rpciJPn
z)wjKUb8}Pg9y{b#F?-YNO!*B8MJ=JTpTAhhyXXM(1vaHw$08%HhW`mTt8T7zt#@JP
z%c;^LqHB_GO$%948@Dn|h4s&FSMAJ}Iq&wo30|=9tL(jr%bmaM`mjiIa`&|$qr2}n
z#;=-e%N4oI<B3x4!`ky2%u;`9cFrifFIgMB`dhwQrRQ!o?Z+vu7aDEXw5-Xn>YQ_N
zYv1HVHoa!GX(FvLLC<4ebB5mXkC3jZV$nM%`L;v%#)}#QnG2J>{?1;q@9(^WX4QAA
z%POLqrt7@j^ZAbHjrWSz&0HK=Q(P}^I<$M5+1%xV73(YZ1=`Dg3UAuT>&q6r<C@UV
z;u-2QZWjf6#D(l$`{tbJ_31TRWLpD`Yd8!h_)a#Q<@IdtnupsD-b^?Vc573Rc~IJA
z>A;6`rfK|9xYexx>ilgZ=^Z^`v-Y>{o^UeD#!mN3#hu@B{rumLJ=R%XbG+y8BYWY=
zSDRnm|NUEI@ADT6mOT42J;!hP+SyyhSJkii8y)mlK8MM-_M5D)_cRxV;>z44d0>8Q
zo><N+F=G}x4}*UILohchw;z}Q%0NI-YFhCfMg|71UkmOvdz%Kc@C1uGh<d}d1&cX|
zA@Rf=#F2Okz6udAl|lvYE^s%!UC+o9ETt^v0~2;^f5Y(eOUW%p78KzSKOyxO6Kz<5
zs(Cnh_+hG3eI55UZd{hc%o8l=Am|1Y<_Q*Z5JKXKIEcV``4iX7xRpF@70@I_U&Uyc
z78uRQ1@>?6q2CMEJhWDrWw>Q!S+n5=CZ1qk2VPg0DxP3zU+EwiZ|Uv#Nk$S6)qsll
z9QcqF$@t2^6&(t(dl@f({W#EGNft>ZnBMz)J#X^9O6UaY6>tzh(ktgH2iN=5fb|PI
zzj!yu_o6)VFuh+keB2b>cu^A=2mB8FNZ~B(AdJM5^_7K(I!~~?uRNUh+htcM^R=nW
zAafOX>|y2>rRF4pV!OB~xmX#*2m%pcz#P0}whY&!^>YK60)rR=VZsay$+;;x89=Ug
znlCd8D;qlpCpQnT06(9npsTZ*AeX0;khrjjgrtR~wT-Qvy@Q^pww1D(hCDV=F=-iD
l1w|zlRdr1*9eo2sV^ecE6DeII7f~^HuMj3TKSvKn9{@qdi{1bL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.D16C834304C1756E.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp.D16C834304C1756E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1953972b69bab58ac424b0d1243ffa67dcf9abb1
GIT binary patch
literal 1608
zcmWIYbaQiHV_<Mj@vO*AElFfyU|<jg;^LB`%yMQ1hLfBO3>9;ZhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<OOS`O2(u~9Dz&9JbrG9zrQf$c;{rXjG!(0M(KR%&(8FoV=!Wv
z|NI&^D^sea-lwy>g(n{TeR6+IRdMCl!c*(h4u4iYdQ`b`o5e1J=sEjrgXYYa)s)*k
z<9>~2*X-|?SFitn;o<A&n(;Ne=6*bVw8ln-^||~r?tMQhex{cFemN^XZt^qB2ZvKG
zF&BsZirQ?%q1qI=@z{FPqK7Sa*@S0&DSg}=5p`C(*DrcoPH*TLi_GYy@d^KnytAii
z6jXmGsGfCj?uKdmN_Jn<yX0w|TC^c)Wv++PqT^q7U)s^Tw9r+^OmJmo`Of1}D$0+2
zS{CH`r9V&foKz_IVqV$7itDSN&NhFqry%&cr_ba<hH8vS5qqV(qP#@4(AK66W`WZd
zl;}n6E?!*H9M$Eq;nF28k%?~(@qN)1Wt94|wg3COFKrfX!P%QQ1fwNw+iYEK2AXZP
zoThlBO5EsdQH|540$q;hcej1ZZMgTIapSSf6|6z_e5d}KRz+}{I7xc0Pzg-9xUB3A
z$5k~JEt@B9(X1=4xkWg)xmzrisJ&xWwY+?a<Wv>GCiblRON<&c=PYmfw#!YU)S{hX
zr$O0z2U(MKCng8FoD$%@!@SYey;5t%N?RA*(h`d|7Xzx6{5$+p(uF@W)LQpwZP51T
zvu^D=^>b_UR)^`EUT41Fs8Gav+U&fkleEji?+(%(8B*yFL*hTyn#$kp*p;W_>$5&4
zJMf^^vRqHsTTiEExdmPLzx9@<lw!V>{%(<sjE#q_Kb3T3aEbWNjM>sR>17@N-=MiY
zQ74RCmBcf--&eUwUU=}l`{$D1elKT7ZLKyw>3z#IB1&Jo>D!w`%@o(mo=?`S^p%;P
z^1wJF$usbTYU#P{9xIG2(sB=c;g~bu@RnfshKDyFq&WC%ea)Zq>svY7>^G_JZ~n{W
z)QG%Q*?-4;*L~-_n2si{6J5(vS$|D4o4Z`FVs*v7FnifgrU!4#Qft=SYZUZ(xq-hy
z`O{q!_N>?@eRHmMp58lI{*=opJ<BF8oNoHZ)IxU7hNXpTOnRr>d=eE><#t-QPT=W1
z@vl#>8%X`=37fJ1nD2=cQz=>T5|dZ8eQv+27Y9xYu2Vi1_mlrfO6mGvb@K8{zKY0B
zxzcn0<g<!ZyJl7`d(}Sczt^hzf=8LVHs{|j>j_{euFOr62j;`2%xk78zkGR$hrvI9
zA()$$+YL+rWgws^HLds#BLf51uLbv-y-kBzc!EV8M7`nKg2f!fka*$_;z&FNUxf&m
zN}+;x7r2|=u4m*4mQt4TfeAadzhU_KrQ{YP3yQG!x|zq<vAV4Vs#f$>jE1QW@e@*i
zG0}z<D9ppj!w(Zq^>y6axN%t$Gf%LfgP<Esm?v1sK?sQ_;vfR&%`)7wvaH!~0~1d$
zuLG|uLW{JobP$XOqZzrtfzo^E_kuMKtreEuexGC{@lXw@m(PI@Nw18r3|#M_5WAQ0
z^4E_8ohr#9sRT3X{$9_Uysr{EfqDfT1d#N~`O3ldJ~d$d!p<+=4f3Wak33B8mkl2`
zMK@m51jYfs13ywY3p)rS@nn5v;i1kGEbl81=lyos70P^VDl^Dj1s;2txkaftiJ;gn
zE=n#|1~Gy_1Q;*}dv#r#S01EU#}pXE5C{`yU`WnQ$;kk6z0-V|Sy<WFIXJm_cm??R
zJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f$SNo*si>-JYU$`37#f?J
X%b7^&8o7vyxqF2$x%oMIF!}%h*>tMB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp.1399C252DECC9B8A.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp.1399C252DECC9B8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f2c3bd813c11afbc2d7045e3f6a7c795120ab922
GIT binary patch
literal 1612
zcmWIYbaQiIV_<Mj@vO*AElFfyU|<jg;^LB`%xY!^hP#{$3>9;ZhThJ*Z6R>=zew^i
zU$fa#ADkcEicX$=<OOS`N=A{2;35xQ&!3y(?=MU_-Z@z;BWR1hQ957xvopQt7>pRE
zKR?FFdgydf%#+#W>PpSuC)fY{^6b;A$5Z0Zwmo<6>URH>Q&^I?&7!t?iG{s&kWHE4
z{-2XX%-=5$kN>~$@b&Y-`+k&IKc3$Gv%-Vxoc$T`nvWkpoqqG}vg!UkYG;cNw4YqU
z{%lQ&=_Vr%)uy<O$JSmfPHer&COe~~^l?*!)DzLF?l#@%T#?IRv%(&7zlqPw{BntV
zziq`1+p`b+B%I@S#V+nGaavu%=caS9MWyO}&9B&n_m2hLR^pufVr6Cd&d*yt+-|+%
zVD$A%f1ThtX`A2+xmye7zP<YC?Cf&!2_2<w?ot^6oYGtGIGzhS!G5gIB}Cww)hQv7
zXNyd&&AFuegIt<Ky_ENKq==U-$X>~Dp(1>D#d|*UBZjY}PAa+FKlP&KhVvBJ)Zo|2
zl1^*N6))=?S6AV_-L&CcUiNji7k}jh7VA3c+I)Sy;E(*?=BXc|G;}$qs&K_`St)Dn
zw3V-si+`uvVuh<w+>1|s;8-rQxcfF+d6oCfKu;mXiUno9TFfsFS!rdg&u$57FXouQ
zo%@ZepYd;(O{VJ-lZzj!6}H_An#Yxu#UJP`E6Z1HGUZ(G|A&8g7Bz2KHFw%0yOp<#
zeRFeD@7_8TRWW<h>&*8X6^dF?XFq?jkay7m+Y4+;vyMeZTn(>@KkII;bgg$`=gX<t
zBBE=OZcPhW5*xQNO@;N(dROhtmO1bCya`&c@T=^-iOZe8?E0`sb8`2!AfvnQ7sjue
zY|9n7%;Skt?!(yg8q89EY7FO>+}ErPUR{>2R_VE$&HHhR>xFr?I<0FmtUBjh+}ihe
zGEWSfuZ||y&L!t|UK3umMSio{k1rf^<{RD;jNb6@=7SUmf32_ibAEj*XPf;d_5IC%
zxtwvEwtPNUd_7@*d(3SCR?!KfzNfj~N8Ku!=k{Rjqn}ZKOHN#8NzR_dv#emL>Zj!f
z{s!Bh?wYV?#Wv}ibG6s$?+Es~S1o^H*-43BmCt_HK3soLQgCS6txZAZL1~k-10VL7
zY5Y;RvR6Oroc<Ei0!dxt`mO6HoXj$@TK7KVinUztx9@I?b(Z_M_w4<|-+A&?+}FR`
z>`uJZw9;6qyMOZ8j}uCjzxll6Z~Z?t<o_3@3mek!Go;O*V!}{dnVTdJ%!{Ww9?m*7
z+hj8jgMR=+FgGi=H<$p*KtNGyTJar51_rKQ3+^?0n+CJ+1dBR|dc(B^i#doP@x&d(
zk$4Ke3K1}sLIv+Ga5ueO&&U%jr7YzG6LxHW!|?M<$t^|}6k*SwMxM8J+j{_2EBY$J
z4GHlRQhzbgh7~Bx!^y)BGbGj5ac|?sWl79D!GaEgZZKh<U?B$~B%X+a2%I;|aLdZF
zX2T6kJi)vUysii>(!SC`Fy7MJ?~{xq9;yKq@j37zDU$J(fh#%`V)rs${`zsCn<QBz
zm0)^dG$R){Kzk4UUa;n&wZi?qo;P`4C3FIf6mSqgGE&Z04sPUA1J*C>{NmjpZ;JBB
z!;Jc};p3+0#*3Q3IN*2SM+#?Q2Vo?htgkFQz<7e?edXc2-!8jCnXgS{2AQkCV-GX8
zC^aV$6x+o`$;HY*nmM@l1G^Th+V@RNfk6yGkc0pt0};&R+?1RQAlEz1mzjl?jh%y&
zn}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8RqLPZLx~7(n
czJZ~!skxkql&+DBsF=G~2$P$iqX(l80RO(emjD0&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.3060AD718DAF4F4F.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp.3060AD718DAF4F4F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..63bf47e7b9e05a74f3a95a03a620fca4b15292d4
GIT binary patch
literal 1948
zcmWIYbaR`+&cNWD;#rZKT9U}Zz`!5~#Kk2=nKjG|3@14m7%Ju*4ZU4-+f3l;|CF!|
zN#VXGi`k<mO8F*7PgZ;E&co55(x!0W(IW5t`?!O|XO~%bPR@+icq8}b-A(B!j42G;
zG`6HN&3JrujZWEn{RlSsXZtT7z5Dg-a@M_(ed^3dcdLIYvfZ&ECVro+hFUoD(O~h_
zyY)Q_zb6)c{%USh^EcV?`uz9Pzf|p=)f?#e^7z$R>+ECeBk$GkTY7!_ib~EO7W2O!
z=nC03amgnigY{-hwno3yF!p)B_fyO5Wn1Sxjk*+NrEGd~@uen>!=)mhM8B=yenqSN
z%-i#F@AQo;5~3c9=bK+*vGG(6VVw0e=(mbhdhD;e9^3h+8l?)A9ZdWz$MsrjTF;a~
z#ywY9=k>PAsHQSKu+7}FX<7JZ^Za}Jn_bozOJyvW+B9>)55`Sw&UQxIG`tw=Rx_=4
zIML)<TcDisS|yz^QE_KlAM2LS>=)d0n6{iVGn;egq-+t-s)-p7o4aiyp7#r+wcj+!
z(rD&CtC%wD>4%<_OeKTlv->t)bbkN6CnHtG<V>Rdxl{X>Y~GOQXSmQ}k&NrJQ!{QA
zDqPWczwtwW#!A*DQQFHyrT7Yb3~y}6f9jZTV65Y*#L|Cdn{O&Zs8l!W+frTOM>XH%
zxR!khe7IzKUh}*u4nbU*JCp?$8P8bWz;?U&N|5{U$8uk%mfonJ5#zXFLfF>JNq5pq
zt5>ePc;#3gZ_Z3(&E50934UHPVZ!l66ZjUTFSTL1sC?1n+?kGjkB(>A?G#M8oHb`g
z+|H?M+I&53PwFZ;p?#G{RsVm_r2-kn)mHlLA_W<-?d4BC3e3@x&0hUPNy=y0Z`<$Q
zC7UMqglh`z@`(O%!T1yd`~Haa&+e8t-oLWy@a)dGOY0oh>8=;F+SEShqttTYlbKgn
z80<cAAnigM*Wv}<Z|3y!1uneP<MRHJvcXQJ9m>m89<RQ5lw*F#oBKLn-(Ke_yZQRw
z=Et`m$*N~QU%PnQ)W32;sWk}^4nf-{PC4gxr(~AfgS88ehSijxkmqzOa}riG@ACd7
zv$W&VrI2X9(^dN)-`%;a(s14!nf{8~X)MbpTAY^tky>K8OJS$ck|5pZH~&n#TGV;^
z@qg7bC&j-#eQqFilP7$}z6a@l&eqi&d;PK`;^NCK<yLd$-uo_(Y~=Lg{CRM{!lBob
zU(J{Psj>d~ixr3R{!GtlOTNl`l{?j5^l!S+UwL)8<<`HYkM8a^FRsi@k_YC+X$1?O
zZulm;hljyGfFYQhmD?Rm0A(PcC^fBE8<<zQel57y>}?v%%o8l?AnFa*7A)o<hQt$h
z5Qp=$U&jdqxjWRc@&pSw2)H7&$oa_yBY0|lYDhdiKRqO#nV%V)r@L%*MV$Q;ZAPA8
zbwBkem;qJ=t9DLmc=iJ*tna5E0~2;^f5Y(eOUW%3mS8DmDPNc{OR$o%5)w~KSqq70
zq-+G|xr)3MTqw+31#~^H122;6rTwJgZsG}6_ESdUY5Qp-@r?b9;k*z(A@vs%ZCHWM
z<l*Guhq)(a<%YZ$>icG~@Sr%82gR8@!N|_!2}X8ird0f0k4x+z@5}qi!@Zv+skGVB
zY*Qal*xb)N5@ua<f|T28L*p`FEDAaZy1;~af`uG}ka!{vB5)o|4I>vgVe}sQy<p8l
zYlWq^-zOPKJX8Y)51#`cQt-(5$-q6$6RhH=g2dDD(?Q~y_?f_YheGUL#>-zn4h#oL
z7D*+Tr98okEQ&}xO%_cgo*|1NocGj#^$R<{csD3~MS0|527KA@aZ_~TMNKxIV15UF
zc+%tv7IqLu;>r5S!o!FsSk+GziKpwQi^MbaGllbhyX*>OzBZK^WTyg;J<LvqqSTy3
zQ0^%%N-kCg(#*kuM!)u{sGj9z3JhWhf@CZZ8Hive=ceRj0J+|2zRWDFZ0sDI+&sJj
z{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6R?21?^4LVhq-A6k6qQs|)it$r^bHJ+
YP0i&@q;!p3M8({_LYUnA96cC)08a=_(*OVf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.8A08D10AF6807EE2.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp.8A08D10AF6807EE2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..76ccc1117a34ccb36f5631e19a16583cebd217e2
GIT binary patch
literal 1944
zcmZvb3pA5`9LJ|k!ZS9^Y}nXbVn!RAL|I`oJYtWM$~EL#4(mOuoD!9C>2(QjcNJYR
z73m_5l{a3sBvMmSDT<;}Nc2vf&GVk~KIc6D|MU4jzsvXcJO7`6&*kQ+ArNknd{JVQ
zFdT(IAh7Z`MVK5bLLv~&>Ij7BX!EMX%{^?)<DXX2oFn(*(Gza3WEU@4FmC8idbaoT
zDen11{$x~R1FzL?e<6B}AR#oecWM32f!;yD1evsT(@{T@dK*iRhG~&0bHZWz*n^K%
zn;Y&;+)rJ+1WFZA9+Ygvzin$9-<b8V>(Hz7+0D=E%C1d0+?<>-G;a9Gc-<x*PuX#6
zz0cmVz8_yAT1!H!8t;!Z_I@5(*HMp3J?^_btjQ~oO3U!046@>%&0mUN{a0}NE`Fe+
z?Y^SrAt&o%o?PhSG7`&UaSf80uf<$({piiY%&9hjw3{(j-oui4TGn_Cgx%2qJx`5G
z&%R)N54NeL7gXmSKRll29GV?jR(Sx^-Y>kWwc0G!aM#C^XxY;c`iomr6Dm5Fq=P<4
zv~~FD;_WR#8j~$g^7m;-xV)fr$@-e;5pMWv|18hy6|uYV`NI;Dkd9L!j>O^r{Fvlz
zbh;|vX2n=#jx-bL{?wmcyR5|Y&NedhesU`Jay5Oy&~VisiD|tf>I@mBQv$jA{?pj*
zyhc;Jr=!4+e`?g@i0qPOz!2u#XkNU7hM=-`(P7|0k{9E~wfH1)-9bW;88)34*ImX;
zv#mIt{_O^L;g%(ti0kcdOGgdb#jS;_-Mw_tI}s5>^f%{A&wRtwc0PPr&2xL_@|DaD
z#2A;Y`cR`c9`rmUE-s!n#MWm^dJYD!oJv#8r&h>Zc-n~=RZ<!o^Uy&mxzO+`Gj7>+
z?4?Zr=p%ksdM?cP(bY1_hc8|%s#V_!`dU7<UpNqTVjjA3Wb}ZG@u}`Tc<|8rn&N_Y
zJ628$jt%>_Fg+;vSWWUaE}^||psl%PN;-ICZS79`R&N<=4Z)I@t~DHP9_jIj-=9;p
z<>>mzK4we=-@nCti?sH*U-|Uz+NQI-Fky2HFQI8~U`Dpe((|2-{P#U#Lc*2qTN^X$
z%tfsoQs)Q38PlZXL^r(v&T0obtv){X(N^NF@vH&EuVS|bbAul1+j>>C?lVSjc7K)>
z8@6W0)fGLz{%SHwIOd!O7BYz3x5p}8mNdQTORubnD&JIg?@Zk(iPHz)iOhw((9Yr$
zD`NVXYr|&r9Zd_nM5!UtPuIqsNaT;XcY}Q2<r22P{;6Hq<^Od&IQ8J;6rrEMZpNEE
zVBZCHYOZp9{9_|?`sd%rxXx25<;4BfDO(dG$nrU~Z*!92axMn~h?RZ_NK;jl`)iT^
z`TZnEMW@i@bE@&Zv^T?>1t9?l$H93k-5`CAK8)}jypl9&J{gk1?j0xqB2tMyb1oDr
z1xD6XYZx)9OeJB(`-xks-rttz#B#81a{+WXIxy1X=qZUd6oL`(I6G7U2*bg6%(;+2
zl3<Q3Knobzg0?VXgKQ<y&s-FRxc1Xj0LT)w3{nJ8>?#_+nw2Y=Dy?|wY_cTj!u5i*
zw-a(}JJ4>O!fJWuFk*5{C_|zC(+3m;?*j^g_W=dL`{1GX0eeyxxlLXanuCTFWe6H7
zr2-HYq{4^>(qLo{+AAr*SI2TN-$hj}6Ho`V6*6Oqq0c=#)MnL_L9$YP9D#DhiGBHi
z{MZq6T%)iq?-_|;8oacSYtrIq!893xM#@Y8WCohS$Og25kpt+UBuT)sccJ9&yYjn2
zKoKa4EC6JRGKG;9$_hpd6hldG7O9S@Y2&kJi^2h9MS!vDPnXy2dT2EtfV4T<%Bu-L
zx*S~?8H2{k$^gh5G=~u#q{D~>vXnICeq$B#aS?JhrwQPy$cadfN(i5wo|NPbDOCB6
zghJA?pr+{83G@5~2*K~cnqB?pJ~t#rCdA0mI~qiyROhQLP}c;|7;P;+*2B#LtHEEP
zgV)tFAS`xrc3Ha2m1B#eEupd(lK<~i-_Xd|#FS!YZfRw0W4DN5&vGO=5a~>JoIcMh
MU>?_Zxu=THza-@wiU0rr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.FE154FD36F572C54.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp.FE154FD36F572C54.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b066f8d4df1255d21c2811a2e580c64ed541180a
GIT binary patch
literal 1944
zcmWIYbaR`)&cNWD;#rZKT9U}Zz`!5~#Kk2=nH9_o3`aQ`7%Ju*4ZWRr+d|;#f05*4
zo@TeDJ~%(R6`ef$$ct?!lpQB4cM7;>d7hg7_ZOdPypVeI0?&=Vm5QGgq@BC-f=7Y5
zpfYbxfWjV~<F5OD7N{Bhu=qLI|NXjrJO7_&b1F&~Rs46Er|YM8Moj#<de`dH!bgL}
zKkr`eS@^xM>hoKEtD4{0r^MsmrT?iil;`(6_;dNsCVjj3`nh-OcRszoeMu$f50jZ+
z4$PUl&Mk~>q00sBYdcJ<^wt>MUcfwgU0#fUck>>n{EHh>SNA5}^17^5)B5+dZf{)G
z!|m?;-^C?uIFes@Z_nN`(Jy1_xq}_5vnIH_pX<M0`^)W!tZ0=)VO!Jpisk#nPI^Sk
zDlRzrGDlX&^O70UU1#YnoA*ZTtN#AbRwZff#ZES<kV9^|<r6lGe44mXSx~fPYu_f8
z)}0w=&-ETUwpe6hLhB_CeWjbNvM-{vm`v_n)swsPu+73PRynfeMYQBx{_2iRi*vGS
zwVXS?>ZQzj`lBZ$F<Ri-&D^)Q81~**PTJzSqIG$l%+&u2$~OqbaW7R}ILYN<x7pTY
z_Z8O^10E_~&k$J>m7QoJIq$)gq>{qh5B%~qjkP?LSo*K5^Gs!kjc{juTiP4&s)CQP
zXyMtohlUH{C#tXQn4%_C$A9En^r>J+-?xfcD;FP5>~FhN`N}>f)~RGd*w)KIchXC%
zSFT*Uc1?721b2RLneAI!F_*p#IiF>0w>XOIV87+ya@m4=^{=k?tA9RmJF?5r#Bgs>
z@1mwqpWBnVww%_!+OwpgetU*$?}W>9r^j-cY}t3%`cp|qoOql`YSG0guVue&zj>E#
za#uVp(iyWvx8{XpszdX9{b!TkSuf4cy87;{ihkzxgv>c>IBsojUphh5SAC-Ku4{!g
z(;81N=u~x8oPNdfdX>}EZ^_r#=F2s%_I9&Q;x^kOGl%zu=hA<@!T0OVCrsY^&i3|x
z-GkoK%I;OZlX~-gLbmjx1x%Y3Wkx=X^`3q1GRGY4dF!wGx7T<ZYzu7DjLGc$V|j}6
zX{+w_N#a+_ly?^&{oMVz`F>;0te%SoM(<O2_w(wprRS{F)h<;|HQOmFdM=~<5dR{X
zn!xBf&V`MqGa7$J-)Zp)TYNFo(tYv0=NVh}n@goyE;Ubj)BE4vCyf2-{_>B3QRgo@
zdhY$9e{;^c5c5F&kiP++<4=D1FI_J6_Ko$_L!B#%D|3_Nf%$N;V8k?$wP6)J4E_NO
z!Q8CeZeRi^0|7;;X~o*We9HA}!M$d0(_m(vU{MEAZ@9K#F$Xaup16ZJoTvRdP9VtL
zp^lX&SinKR6`@7WS1t&_Q}b0r;_3P7A@R(7&EP!UWveUV?4M{e@&v2<sz<^Muqs%!
zb5g^zA3$M!U;Sv9uw(lhhM!+bZn3ZgODRkFz=T<Xm6Vl`cv{L@NIWBDBRJ1h<gMUB
zVdg5J>v<h`kz6nBD-CxOPq4DDG7?YQR~w0E>}w3?h4=}nznEyl3Unq9Cl5c&J(*JR
zcRen#gIq7~D-U;lmZZ{VOS4UVKw)!V^9Yz~7|qB9j-=j0zZa}|XsxjH_WL9wiHB-H
zr}8=QAvsmXR|f6?o?sPU6(pXHuMQH=#McDQ+p$pSV@$#W1r{EZpyNRaI-X$UpyLTf
z4!T1jb}!@QuO9~*Aju-B1amb{up)~h5>Jyw6NzWYVhHCwHDLY1&M)2#3M^3`d6)rT
zHhkO^-FQ)xjVGAjfghgEc!Gr;gpqi%zOwKT;|W&vRYl_I`syO_Onpt^yx%UnLYc2k
zWd_-)z+(@y^RGbIrn~t^o&nRnpo5?@OqeHF$Uz8+C*mLi=P?we<|Kl0PjOLlu`-Ae
z1R@|H5XN9&4qo=};r`p^%M_UcgBX%?Q*ttZa^7ja%q*;I>>QljJiG$@e4c`?&T4{O
zo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#KHMMl~4GfJ<&E-s_bd6j@
P#oWC@nB4pvJs5oe&42Fy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.56FD421A51C8C5E1.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp.56FD421A51C8C5E1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bf92734dd74b0c6a9e404c22c2952bc9a6ead8d7
GIT binary patch
literal 1946
zcmZvb2{crD9LFzaNOO_Bv5hg7G<IVd8is~3mdngVN!iyz8uL6$D!fytJPIv}PT3n;
z=A{&#I9iBMyyWFIlI2;Za8i%-@<yGRZs)vv&i&u}`QHEU|NZ^`|NCPx+}(?00APC*
zBP}tS6OIA^Kv8(7aM<zbQULH+768)99tMoCM{GAuu0;j%B6j$YDrJHTHGCq2iw(!=
z;D#gmM|IoZRCu*HVOZ*=sgAJY;61-#%J8cOyDd^(Kms$a2W8P&9~2~T3gLlHy@zY_
zTINd4F&Pb7zS1WYd?zL|lEa(s9d!2dzAk;D)-3Sn(V{o+2H#Ar?oV6((wTF;jc-2l
zA;U(qGH0N_-e%9?rZ0`ZuVr}M*w1{gw2<yJpRLMGATTD+hlM11?7K6-j49z~PRic#
z*+=QQU3Ht|zPE^8jb<L5+cH5Egzo1Cyek>GMENT;W}*GoBXfRIm26s(`xQx2PxVIz
zr>}Xbr7#J%HM(10kaK^EqB3M-f;Fy`Oygp<X=2Ep-Kh4=fwX9!8&mWAtk1ylow1~+
zR3h5iWl<_h?&!g&1m<bU&w=QW>{IBxPp{p(h8TU49j@$NElk|s3YYVDqT((lCfGL>
zJ5Fr&Cp_zrQBUXf;qXu78{+@)lg3@x-bLye!#*CMg+02E6<K+VKaYLZ?GaaU|4`YP
z!;IMQ`(+WX2WaFIPtI809;aSm@*~H}{a6@}+kxIhlGf{z2wwPTcP>9HW@iy?gAO;b
zv>T`+sbOBu?J|6@BA}>v&s26W%wI`UjI*m$29oKzv{(yY3AsiYZgoLHCuNp6$6K>#
z$_c|?+gtnaZvX5@7F@~Af2A>O70XJcyd_Kp;qfQk@Y=L~&vaA;y&}HsykgeFlg|&b
z30*bM?JZl`#2~rSQ;i0yGpBacVLd~>V5`UMT?2m#kvD$*v{`<>YaDMM;4l>U*o{=;
z{ZHn*bJI~q*c&W;PGRFhdF*Kk`9r^j^tQ}Pws8ZF*(tYH?aAF8975nkDUio)Afuwe
zI%arkVRr9jIaPO;bHij!?F!c>>~VfIJ&f&^O!sCy+EJOO=)@gv3z{9dffd|*(Hh%%
z=b<As&=*WkB(74bx|iE|N2(GC#uxMN&XDrkGhOnoEx*8zD*Y<>#n3fL<=hh4Q?a_{
z<=L{c9}ac4r1-p#*hVJnEH?ITfCvZ9+AZ`<vpFv8Xpib$eGMz7^^;2HI@h+B6q^fr
z`ol>#vDYoKigA|~GaR%7ISy|xLbaWlgYki3Ij}2#S)G=a#rijkfSdfFr5!MwzgE<n
ze6_heqg+s1vF2T~HvV9o6YC6gJ<78OQqmJ6aKcVxEc<;~X=?QZ0bf4=#-K6uuY>T*
zw~HMelR^}B>c&r(`MF*;uoMWZQdPagVXzui4MFNubupRN)Tf0wPY@+Q*a$M(B?@s!
zu4^hfGPNv}%^~yMBH87b09aY#yhNd34T6S`NJhbW1U&?8C2U2I6~RhOq6mqNUn|J(
zSiO98%u&}LQZkw@+=dp~fwG{2Xx45nFKOE7-USr+gB|#T9S9@sKp1HkuMzsnV?buT
z3><_Lm$4VC*Fd(tSs;|{AiE%uU2d3$d%dM~s&M-hsfsj_48lrOB?PHZRm60`@)kEW
z*OG?@VL7Ut+xHMnNE7}}1c-njB1A-xHDoQO>XG>ftIcDE!sX<t@`&ZMAT9AE5Y~tE
z5o7|HAczE!#MI&MG!u$@(J9<NEDEdlJs&&@k04`|F@h{nmSP%hW&9~O8+Kfn&nD_7
zL_Ru@j@Si+4Ix7WnL%a<vW0BLG--iel2K64TR&!15GTrKiMMz&IQyBgGzcqD6~w<L
z2rE;S5u^=ii;oI~4Il#qnL?%rvVm;Gv`T9ZkQzE8weG14I*U93c637cdiSKTBU1>%
zrxaYD^LD7fJ5y7V$pTp48f$%Iet(Tjj7o?TqF0QM6bdaZvq2UE%574RXDGUDH&oon
za8*)QR@scTCp&JVI5|_zREZ7*+by{NQ`I!Jv~}@%`UXbETTRR@EUj(qbVwSeRy0*L
Ox~IRSJLKXa;r$;G4^zMZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.C29BF8AC2D59E669.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp.C29BF8AC2D59E669.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f56ac5818db7a2777a39403530bfa968089abf67
GIT binary patch
literal 1944
zcmZvb2{crD9LKL_B-~-F(~M=9vBVgfVMc{amKlR9T#@V|yBwyD%IOIuNzc;qT9hp9
z3`I(o)09`9npBn=5h9ABmqMa9dS|Bl&UycH?(+Hle*fS1d;a(Syt!<4z8nBJMsQOT
zB6(qG0030PR|+pVCKUw$t?~endZu+#!uDP!w)=;P@bKwJahO@hA<@+uBU*#zoJ+?r
zm*iAH;Lb(1G;`W4a*8mUz2ie7`t2I;jP{R#dZ?sbiDx|Z8qJKIn-@~`>9bY~#ZNv~
zZEt=w`#5dQIyfzr^yEYo?rnSfOw^&LT_=Yw=eG9Mm*1MVx-++^rPKU__NJXbld`ud
z(EUL9!1vGLZ6`viS{_ff^#45`@T?J?cGhE0=v`MoG9}ZKG{%T~wcHNp`$tg6KCYj&
z`N3lwg3s4Szqr!HrX^IwXfz8JzZ_@t8>j9RWzV;Rx>>a8ie84u#kkgWH1xg#<Wf2#
zy!e{&Ey%2nT3DTT_S8(iO-OD;dF2so$1v}vlCMFG*1nJDF`}2j)Ynb(vofkR!ZCLg
z#x$(#_@0OU3Ud!%6dY6#usQx2g238slkBiJUWZ(&9b&R@1rq{Y9#unzFd2*c{bQ1w
zc3D+{nZtDDVPQ7P>7^I5Zv6@U`@4zs$H{5z>($g1;}cao6Vm%9<!K^PrvSG1_^YME
zX0sxH|Fc3*Zt0ZsX;F=_&p7tdRDPV5f_G)z>Qms8Bv)F)t+*t9{Zafe1C<O;Y)?5o
z-TYcv#@7b+svYaHf!iH#OQu$K@Y{-fom|y0zX9Rn)RD_27rtWaI-d?!a~$8<ej&2`
zusY?NKGdqu`1b|J#>P>`nVL*N@6n)*^Xak$<ZB{Zj&cH4RyUoAeQG5XTxlN4j$MCS
zr6v)C-1)CcE`{nmzgbTD@Y$6?HW@fa-NGdg^F|}jEyGk!P93q;Dec*hgH8t49xr^i
zcjJP0@r2hyx-$tEqe$G%#&-;iwzt;K3&&1xsr${c%}vDEj5nraC{2VJMmRs`4j-=C
zaV9WgfF2#r^?GQyLs)m#vtl8u?(RiSD6ch|6My%BUuLe1-Q~^}?)zRoKK@2eQ&e`n
zVQSkmq0N(^%mv-#1V?ormai3+(ij)>d>3Kg%%M@OFMP*l!<D_J_tdNEoEEi5vXTQ5
z720xR>yP!_9-2$yP21!{MKl8YZSl3i6L&`jGAe5$D-z2eU8pY=to`6Io4tw?(s}%x
zL-YWBOX#AewSG}oYFe=H)2*4cy2Oup5Bxpe<>B|d`Jr6Y<@IGIDD7x>3eVHqV$qE`
zYS{&KDsHmx{vJhN`0-~k+h$&-LX=aTvMV8+D4s*0qTxI}X+a4DHhKcEqO2nO=WQwT
zb0tS^OQDG8R8l9S@H3w$m@m0DXrrg_t)xHGjRB)TSc9eECaD5pO_nBta4ejZ6e;eR
zg0Y^xXb>il3GTnPkjNwinUYNrL?_dwq{7c3?5KKwPpqlJQgQs%K$WG6Aa$0yl&BZ3
zrUu&&Q)EEc7&7*kY+09d<#u8E+gWkg0<s8@gg0bQ0CUqk5gPp;hiDk-5Dg<8qG6;%
z9}iXIu>xCJabHj#RF?FOA%yg~Jd<0JBSJ)}Tr8e+V=eRg5pmcWvfeDIK1Hj&yfc$l
zEepaJ7RFf;24O8o%lj9RAu@s}5CuV&kfoF=dMCr_D+i6l?UYzbh<4hLwzLR@4Il#q
znL%a<vVyFnB=A}PE`-?gK>QW)XgukcAAPhwf=tjR2%@2BQhMFQeTjJ=nj`juWnmG1
zbRk`-3J4oPMhG&8%n`(dm{J;9EjulzjLTg*at)9u@iSfh>3YDvrzXomSed0P{Rcr<
zjirVl9Y{xd?jURk86t=ZQ4z#|7*d*dYS@J8K89M-)C27$nm}@7eAv?Pq$F=mA&Z|V
zxUA_)v{9{H-!d<6z<Vk6@A8to0KnFSi1=tRx^06{XxZg*E94bH3|3i*tK#fvq@uue
zP{paKuf(reYh!D--kxQyL0Ly;t|I<lR8vb^M^B$*U}$V&YG$#TX34PDwIWdIP8ymV
OSD$5Uj}0y|?*9PBHXHl^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.9341E131FE1393D1.idx b/.cache/clangd/index/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp.9341E131FE1393D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6eb4a8d103f2a25c10a1ddf6a5be0ec17b354ad5
GIT binary patch
literal 1944
zcmWIYbaR`)&cNWD;#rZKT9U}Zz`!5~#Kk2=nH9_o3`aQ`7%Ju*4ZWRr+d|;#f05*4
zo@TeDJ~%(R6`ef$$ct?!lpQB4cM7;>d7hg7_ZOdPypVeI0?&=Vm5QGgq@BC-f=7Y5
zpfYbxfWjV~<F5OD7N{Bhu=qLI|NXjrJO7_&b1F&~Rs46Er|YM8Moj#<de`dH!bgL}
zKkr`eS@^xM>hoKEtD4{0r^MsmrT?iil;`(6_;dNsCVjj3`nh-OcRszoeMu$f50jZ+
z4$PUl&Mk~>q00sBYdcJ<^wt>MUcfwgU0#fUck>>n{EHh>SNA5}^17^5)B5+dZf{)G
z!|m?;-^C?uIFes@Z_nN`(Jy1_xq}_5vnIH_pX<M0`^)W!tZ0=)VO!Jpisk#nPI^Sk
zDlRzrGDlX&^O70UU1#YnoA*ZTtN#AbRwZff#ZES<kV9^|<r6lGe44mXSx~fPYu_f8
z)}0w=&-ETUwpe6hLhB_CeWjbNvM-{vm`v_n)swsPu+73PRynfeMYQBx{_2iRi*vGS
zwVXS?>ZQzj`lBZ$F<Ri-&D^)Q81~**PTJzSqIG$l%+&u2$~OqbaW7R}ILYN<x7pTY
z_Z8O^10E_~&k$J>m7QoJIq$)gq>{qh5B%~qjkP?LSo*K5^Gs!kjc{juTiP4&s)CQP
zXyMtohlUH{C#tXQn4%_C$A9En^r>J+-?xfcD;FP5>~FhN`N}>f)~RGd*w)KIchXC%
zSFT*Uc1?721b2RLneAI!F_*p#IiF>0w>XOIV87+ya@m4=^{=k?tA9RmJF?5r#Bgs>
z@1mwqpWBnVww%_!+OwpgetU*$?}W>9r^j-cY}t3%`cp|qoOql`YSG0guVue&zj>E#
za#uVp(iyWvx8{XpszdX9{b!TkSuf4cy87;{ihkzxgv>c>IBsojUphh5SAC-Ku4{!g
z(;81N=u~x8oPNdfdX>}EZ^_r#=F2s%_I9&Q;x^kOGl%zu=hA<@!T0OVCrsY^&i3|x
z-GkoK%I;OZlX~-gLbmjx1x%Y3Wkx=X^`3q1GRGY4dF!wGx7T<ZYzu7DjLGc$V|j}6
z>9!d0N#a+_ly?^&{oMVz`F>;0te%SoM(<O2_w(wprRS{F)h<;|HQOmFdM=~<5dR{X
zn!xBf&V`MqGa7$J-)Zp)TYNFo(tYv0=NVh}n@goyE;Ubj)BE4vCyf2-{_>B3QRgo@
zdhY$9e{;^c5c5F&kiP++<4=D1FI_J6_Ko$lrp^_`mAOgsz<lVn@LyC(u;hFm2LAwt
zU~X1!H!uN|fq<gav|?>wKIQtg;9j%0X)rTSu&9HmH(Xn=n1dJ+PuxKq&eL7Cx+2d0
zi8doou)42$ButA{!K$5;8lL?C3hVpoN5h02+utz!{8Dm@g(X-@S;_|{%o41mtc1kV
zQr1G^87UjVc`zM}Twovc9{Rmt%|mO25I-UH7ZYt*f!6VG^6<k{WlF{0^|-_ia)rFF
zJlqvol1iH`%{KJ`h0T4<BVej$8E#ow)@-<el_!|jf!7r#%o8l_D;<R3Df=oT@w9!l
zk$A?w#&F)!+wYT%Bp#{(jpK9RLo!arR|c+#Cs@T-1&OEQtAoTd@il?-b}SV77?bco
zfrSSp=y*_qjwcv7=y-yWgYHm>-OG6S>&JlxNU}&OAv~nWqKL%PWYI+88L}9{dH45v
z-sF9i&<XUAfP(;%hva<a;LhX;R`XRu;_3P7A@R(7&EUMJ2CQG$`Ng|IaVE+m4|C6#
z4Iei}H(u0a;|b<>;D@I(o?u}IVI-ccuPi*`c!E`ZRgrkQzPd;}Q(sd!@3+gYQ08k>
znL&0c@Yuub{3{T)>2Cg!XTWqX=pg6}6XppPau7n|i8zSBc??CVIf<a$Q(Tl>tPElV
zfd~i)gfSSHgU`o@v<lt(8paeD#E_hul9K_H^G@?+W?^Mx=iub#;T7QL^AvP-Rukm%
zbP^I57Lky&u(Y<ZwX=876V<j-Hq(&DCMqT^BdegOq@t>>simWDU}$V=E@vX8Yvdv-
P=I#~3<mTt-!RP}3O}7jI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.11490B4C8E752971.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.11490B4C8E752971.idx
new file mode 100755
index 0000000000000000000000000000000000000000..268450253592331152859a554aa7c374a3d8f392
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%syrY27N9DhKf07LvI(|HWPUIKP7BJ
zR=BUpV)p2XQohO2lhq!(^Kdk%v?&~Tw8(p(9rwgdIk#qR^Sk|$yY+7KUEAtrqXwP>
zVy)YJ7>f7Intt@(-5IGL>OM-x-iwQ^kGb^vYKm3Bo?4~#X3gBI&mQ$>KKi9-LB*@2
z+TWI%dH40#@Bh=)uD@UZ<+ICoyW(=Medg5C_`BiWg|o+3pI^P(-~7ufHdB2@d9iN3
zh{>nSCWXx7e0Fu(mNIqO)V7z+;qwY^Z@IQP#yiS8eTiCSNR&Y8jy|V-hyHz?7WK4m
z=i!@&@4B*m_sRbA=&jqV2KCGqEr-o9S@T_<ukqd={gQ8C&@!Jxb{0AJY7=7*Pnu|?
z)i8b6fz2ByEMDTnut&W&{8~xYzh`H6Z=cX1+sh`Eu{1$V<DT=igcI(^#9TEGybpOG
z$`QF?MYXxpHoqb<0kKKh&mtVG?>Czz2`(tzw(;S?xp$P0FOcdKv)gf^x%3}PV%ZGe
z#cf7EE__kEtka&Z;w{`Lna$VRD`&Mo-9=k@VItq}!mzrj=MGJ?JlgW8qex_?c5;@x
z|BAqS$qyF>1aW(84fD}9k$Es9X~V+{r{eMrjCDMfSo*JQ^G#(4mFi}FTcRucsOFm-
z*Rn6350^~O>z+5oA&4t;hp@mR;~C2v*lssp336AT%=dL_>5cjsaZVd1gneCcbjSHu
z--3b{uk5`mdBZS0s7&^)@Uuuo#eP>sS(o!(l}wA=7iG?wDHway|7uN<@`+_vEez)t
zsYmkOp0G_-bme*Nt30au|2r-f$SAJvi~A)~ka1Vge$sL0f>kAJ)=uiqFqrxH{m%GD
z8rCN(rYbG$-qQMAb}>)Ehwiggg@5!u&W^gef94C-w^C1TeUe%zTzx`veWBIFIbpqA
zp9|T{n$1EaSYwl(=ah7CJ$<H9HC?XprRqUhL)DoveaDn*f;QHFzOp^u=tlbb>hpg2
z-Q{zZ%-!C1E3B;AW9rAoL`Kh>Zb4Skg_kp>8o3=m&t1^O`fdB;e+4d^4ofOJ<%v%J
zVqKoG%h_9bmF)7=zh5^zpBr*H$!O{1zeoIUeRSR-n18$H!2+T5N@n|hJ-6gZ+Rj>6
zi{|O(Md*g_2>*8;m~M{j`|EAvZgzO?^7@<ke~;+bUtE8_!ge<AZo}E7j%k}O@4cy`
zy@AKrsqTY)&xc=*U;iprOnF{h5|ViSpSRiJi8lOCxc5qwD2e`^ukvqyv|P-TH_~m%
z8x9p$<|fGlbLYKS9lrL6?b$pG{s9cZT&!H-U;-!u0Y#~4#djDP7`T2dxYz7$8qC5I
zEaD{M3)dDb>LiN96LS(n;we}tIKxy5hF;1jSi-Y`i6@xHiN^yb%o8kaA#Dfa39j#y
zDqZ*LFf&iEfRlhLOqeHF&`A)9C+s8)=SiqZsugSr)nVibR<uy`f@yJVf5Y(eOUW%p
zmS8CkDRY>x%j7RF0*i$20bRlG#E;|(ISV<sD?<E))L%@rVFen=&B@IV)4NB*Q*-If
z4d+0@3fzt`;X@&IFXQE}9|sCcut+GugfDw@7R|6pdIGeG*NGR&CK(GExJ@v{j9lPQ
z?>+Q;!J3EG3Xhr3zq5R|_$|;#WXC@>VEw|*FWwDufGD>-O!JowA2&reUep9e4WAPq
zQt%5o2_f-hEo9-5!4oWRArI#jrRF4pV!OB~xmX#*2m%oh5C~&1Fb8i?`Ec&dyj9<s
z0)rTmb5n9MfO6hxzRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6
wR?21?^4LVhq-A6k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_LYUnA96cC)0NBBlZ2$lO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.ED0E8CE7C6A61531.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.ED0E8CE7C6A61531.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a49d19b6aaa3ef9fab1ea9930c14192a836b8521
GIT binary patch
literal 1624
zcmWIYbaM+}V_<Mj@vO*AElFfyU|<jg;^LB`%ywo527WFEhKf07Lv!<Qn+aU|uX(O#
z`k8FjAI^_%t=_0+xJ&G(M}Uf|;3AJ+{koUy-(TGHIKp$Iu5MfH#LDj<;^tdhF`r<_
ziOf39VsQL*W@z61*bO}Q&i`BRdH3tt{l!(ACH&dCwx@qeGTpHuBKDtAfX_O%u43`i
zyU%+Tey^+h{IuMr=4Y{A`22U`U#hmw=?ipx+5WU^on2gg+@1PeORsNVQOWtkVD{Go
zldf!=y5yUW!TM`UHcP+MF!p)B=Tpn=ZJXzbMqP5VQa06Gbg4<>aH+^A(QoUwUs+Xf
z=B=OHJALDdgs8{j`P(nC*mx?3aLy79`mJJ>9{VfSdpq}5qfnu32Nf&%mzL^!yPaCW
z5cf*x{4uA-Nk)tx=9$DrUS9okw)y*ePF^|rCKnvkS&bb(GN<si)g<KvPGS1v#v0c=
zLACqW1Gf|BJkK#5aw<A2!=3S*e?jjJmKSq!b1O>Jx6eopSnI+qAM@m(Rej3_t2N6)
zoh55B_!Tefw5NN9sWBN}dlMyX`=!pmBh1a|w9Kc%pt`B&4$ZPW+VZHQNMz>fq(zS>
z1gze3{Fw8~C1Mk**3JmausE>sNW$TDliBwkJheepQ?X*vw%9D@6{jL{GS<JIA+UV&
z_W+}%{aosMOR~Q#VNqLZz+2AZRKoUJ<bsmzjHsj%n|Fcn(d~Z@KjCU>-Lk4$_gL+<
z?bAJaml>8NCmf#RdG%|iyy1B-PFEf8mkL#3CzS<k)n*-=DS5H{&yBO2r5(4;J=Xbh
zs`N>zi1sbID?QtH1)UH0^l7~Y?{UFRS?`|)F!8Ou{rLBjj{<X6$-I8`L`g~~?SJ(n
z?>DLHlIw$2N+xdmu<*PFv&^5Ot#jV)FaD#piZ5TS%5yhc>$W|PFJ{<^^q5=BoiQhD
zE!X94o*cH=6^2~9T%P5WbZ|XgrcyOsuJNSkL0Lml$@pG9w;xNC>MtL;oiCZQ`TE}U
z<J*r~`&~L0?VhXt+jhyR9|txxE_tQ2a!&7#mnOXjmM%Ez7I2X1_w~Yh3#FS6dIS`&
zcX|JsYi)9O;Z(t|ZI@U5+;#g?TvmqVJexMjkLylz@M#oI+y3G8i=q-ou2he;R_8X~
z)Xjdv^!?=j*d6ZeC$9US%TF=BQ~0a@uW9<fA2aRee^fi)9((@ty~=s=@1ItNCEc`C
zwEq$BTJd?sFT0!mud4g{vYx!ye<sbMYWczH#%4z5^e2B?j{G<GZsML;%gj7y**S*d
z%G@M*V7|1D4!!(MX4?%O2LAwtU@lfJe=q@*fq<gawBkFA3=CYq7Tjz0HVtOs2^Mh@
z@r7#(7IhLu;)ywlA@LL}6r5oy1w$`o6fEJ{z{C^G<HX|u6Xpq)wve`i@dVd*N|mmA
zb(onaSiniZ6(-CREa)VN#1nQBhVx)*7`ea>?>+Q;!J3EG3Xbh>7=C^!xy8s5EX5&Z
z4pZea`OAyIBH?>LSMWRWBe_D(LJsbV5I-UH7ZYt*8F_-aIl1{^diQ8}YA(IG;T%X<
zf!h%#d?>{3WxV|L<3M2v76~Po@MUk#q8T<xPk=V@I`JaeBx4~1x9P4!!M?fcuUP^0
zDq1Ld!Sp_6KL5`0-Qu@EVPwZYHDLY1&M)2#a)2neJWTbM4Iei}H(t~PMh%}6A5!oO
zISC>0WG!Ujk--xzZy^uo6{Y4Rf?~V4D7jb}NHYg3toe53R^j7&Oo2fRL6C$1A_Ec3
z<lK~;3?SD#&6k;lm5rT)lbeTEfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?
tLmr!`n6!+nf})a&s=B6@j=q7Rv8lP7iIlF9i>R2pR|u1vpQ8t(4*-=>wf6u3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5ED435B77281FBCD.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5ED435B77281FBCD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b008b50fbf2cdefe41c715a52a623fdba733f7ab
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%wA>&1}!cIhKf07LvI(|HWPUIKP7BJ
z*7Z`0#q7}&rF@g4C#yYn=iz8jX;V1xXp#3mJMO?uIk#q-`Q3WS-FmnAu5ER*Q3KC`
zX-l{HFcj~Z7xwhq+Zm}J^1iXpy&o4_KjZT2qfe3+?5I)F4{PS`zIs!i`RLJ~4j-=^
z{q!w3=+3^`@%8^?`RD(ibNTA=-K=|~be9W<1pKPoxo9<i_<XT)`TQ4Gd9&h}?wwDw
z-Jr5HS9PVW@R`-7FW<OXo#MG{z24^Gwzq52Tc>TAc+SJOYsD6YQw6fEHLZVN>u%MO
zDc(J~efMUbcQdd3IC^`n38Vidj?kvG)+=_M=OU-qUBC3!X|2!HmfV6HdwyTqHBD7H
zHH1OGv@<<<p@Zj4h6;bF^=n>U{d0Eq_URKk-uALdy;zc<rcu|h=HP_>9^I~>hWjfH
zWD9Ic36#IXantf-q(g+qx|zw0Z~riE5$tA|yLl}y|GOt*mTXtOwtP6~D!1mdJICh3
zIZ>ul4z_<2FgjaQ;kL<0aK^^8>$hGk{9f&}#YiaX*#>#_r{XVl*31m{(dg4~pSV+N
zS(yyS)>G9wHdDN^7*}5F%4}_8mI>xDdt+63LHZ1j_au&k%3I$CNjiicOLDA^J>Ft5
zcW?V7&Ds!)mrJ7;tE*0!;w1HsqvZw9GtC7@zbS0Ja`EI6_u5?9Z}NKQ6eB!V@4B9J
zC%v?K<;sgD$MSe`W*BSkp7%}g^O`9EM>b92Ta-T4hUud6MU!)9I`%y}o@KXFFy(gU
zoEdRDC$4EM^|(E$tK@|CRUTFS|6P{~WE59_i{sOLaABA2KK1_g2O&=*!qvquCK&y`
zAH4raT-l_LR<2IsncVNI-6S78n7;aH&i}cO&9}b#khW<0mh6Cx)6y+{Zx$x(Ir!40
zC~B*g&vD5cJi7vtMM^s7Zoa6fHD&gs7qL|=dgnObb*wz`q{c+<;pDErvzP7rJMExZ
z_1)^Sn&>@`Ro?C?y;8R6zVoUcejaX*ReF<7-Rzj{cbQ`jXH*W0sLqD>eZLP2o@7gJ
z;n*1#b1DDM!k2t&dsdjw2|GLe?zHbg&R13E=BU?Q>Xl#ATX8*=<@pqg)6qXlOU$Gu
zJ4ptu?)#imc35iXfvunF#lId@-1heVLd#8(c{YEK=gn5n`{|umdu+Ci{5GZYza=W`
z_Dp{s+39+M^XEZ(gF~~~uWna=5IF1n#lT+K-}*Oso_{py)!&kIfn&|j^H2WPN6W-a
zdZR5Qxc(VKab<3jJTP~@nXvt%Hup4X9tQsahF~sMt}rkGl!1Vv)U@I|j0_B1zZTqU
z_BIV>;RzOT67hv=3l?<}MdFD$i6QY6%oUtqDg{F?WfUyo*}%jT%;Uu40Tbp4mNu8R
zh4CEQ-!T0AQgVxtC0L3>$_ys#GWpAkz#`#$K)w7<{78D`%;n&EL;Qr)Ure-NW#kFw
z=H%vw=}q-@+}pTuSrRi(uz-^Ql3qb4K_s5AlQ5iDCo6H&K|X*TXp*A2q9@FxDV&$}
zr>_Z01PUWNZjXki=F*!R&VlSz;C6(mJ``g2GG6}raiFjSi-ZzP__8-=(F~iUCqS?8
zI`P7-<O!BBmw|ibsR8R3c7E}0kddO?@-V$Hnvn|}N4<xBFIe-?TH(uvkDH<!FKPm#
zhR=x)Dfoq)gphc$=Cbg};0cyDmxuF;QgaePv0Yr0T&xVFnS+leR~E6PN!l|71~CLe
z5(0<}LNJqaQ*ttZT<<hrW)@a9b`DN%9$o={K2JecXEi}CPbVR9VG#*Q3rlMoTRVFP
zJyC5dWit(VY@%Y)GO`MaN-C=Anp!&g28PC_=5i)dx<)RdV(wlcOm2RT9*jN!Iozyz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.069AF0356549719D.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.069AF0356549719D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ad5fa11289998a4db533db1fb07065837b6d5e39
GIT binary patch
literal 1622
zcmWIYbaV4#V_<Mj@vO*AElFfyU|<jg;^LB`%r<5Q1}-iJhKf07LvQEZwh*}bUnKe1
z>>M-E2kl3;Za-$$c`<jQdy|r8r+{l#<;SJ*_ZOxf@0*mgJu~o4_nDc8cg~ZvVm`r;
z6Pb0I#o+jB&(OSk^EUF_^ZmbQ^X}KP`HicL<>UE8!u_B8EGkIcXY)UI!Hzu=8uNBV
z+^O~I+FibR_WAlp4?o{G-u0)#e);v&M^k#Zyzc8yEBX2G*X2L|9_#Mi=X%!r!0(q^
z*u|D?TdT4)y{U4m(%e{Y_nC=RucxZt**5p<nhTT6E=jKR6r3vNBDrqyg!LQh_bv@<
zP0p+SQcyqdAoFFZ{S~{Xsy*<OTIh9Q&4gR(7cF03dhfL>cg56|lQwYjovhYan;mHs
z<i`3TXnKX27}u03oDJWXd^|NP?(F?NmDQ}88|PkhV4f>-hUMt@38oUCj@($J!TMt>
z*B&-SZ?UiUdMC`CJcqTV^~o$tu?y$kJ4$clykL3z)`u5<c?QQ9M0JYU?Ksg~`i~*8
zG{Sdvo6(O8@0^!K@|~Zw#)l<s&CP9QRWJU^39jvFnP&Oq@rpl6b6Z7=yEu;uKGiS^
zKk9PaZedtue{Xw`r>?@+D8scE3L26-54Pp1f3Iqun&=(mRIq4UY!>s1QxQ2C>tD|h
zIKKINfYDNZF7>@7+FzEis4X?%EoX2lVS6odLCJPTR8on}yTJJ9_CJT8va)bzhL%V5
z{9co<KT*mz`OU2Zvn(gCdVSe0Vct}su83)u9KWoY#I8`~W72DEyy*Rpg|q$hSZ<a{
zsV@s(r<$E0dTHAdwYRUmY@JV@u3uy#<9=e*-08KP7F+V#-%mR3T(GKW&FabB85?HY
zeLqou<K($1d;BI|aJ_Y?-EJ{Y!H3gNlXm?+_hI&mTl;6eP<<;Ua_bY@Lg#M*2KOG6
zBo#$%)$%!Kv_Z0D;W3RDLNhmCRMeU>d%}y@Di%F`j&~g^Py9$KF*~F#`g69*y?-YX
zChvV`YrRMJUg4y&dsbQ2H{UCU9%FCgRS6aMe42B_cX=kuIab|kO<ZC(%8&nj*pbp}
z?9>vYsrS<S-G!3&Ft;n&eyg9IE<0Ska^b62hNn8pW8DhuIydiWz3={fv2l=wbJo?u
zdA|9P%FlQB{ky-&*2-m9U0#*)?1yca|G!yR`KWLHuFsPBpK5M=vUy+o{O7~CdFiTR
zbA>-O*N3#2_piz~t_`|ZToR(}yx(Z{!#lFe-?4JBvCIkjF8t}=ejN@y=Wn804lMmt
zT$!6B56qQ!KmKAp?YiIz4}*UILogRBmmin_%0NI-YFhCfMqsY}wcuW}w`nj7Pq2uS
zh%a1Qu&9$L5>L!Y42h>;uHXbyDHwVwqhJZo1}2_h9w#0Tm@rSUw7Ik`jHg)=acFne
zcN<2YU`2C9PnfV{`x}OzUrKH<vII+UNSVQeT_%5d5m+RA4`?L66F-uXa^`YyBSZXz
z)L%@rVFen=&B@IV)0^t+xVLfRvLt4nU;!rqB)x)8f=E1JCt)~m3g>10>1%=#fhHll
zVvmNW=F*!R&VejY;C6&ra45v?WxV|L<3M2v76~Po@MUk#q8T<xPk?Ubb>fAa%@ZtR
zE(3QnOfe%DI5K(<{a&!<p|!$O1J*C>{Nmjpn?$+gVXD4t__!&$@uDU$YWSS^kb+;x
zNeGE2Yc30q44z<lb9p$gC^aV$6x+o`$;HY*nmJfRJ;h<;F~caPz#xV|NJ0RSK?r7Y
zZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$rrEI1l
sk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEggvrg%(Sy+k04S!=LjV8(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.511C2FDB97E265FC.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.511C2FDB97E265FC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..02fedb695bf19c2491459a3bdaffafa59ba1acd3
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%syrY27N9DhKf07LvI(|HWPUIKP7BJ
zR=BUpV)p2XQohO2lhq!(^Kdk%v?&~Tw8(p(9rwgdIk#qR^Sk|$yY+7KUEAtrqXwP>
zVy)YJ7>f7Intt@(-5IGL>OM-x-iwQ^kGb^vYKm3Bo?4~#X3gBI&mQ$>KKi9-LB*@2
z+TWI%dH40#@Bh=)uD@UZ<+ICoyW(=Medg5C_`BiWg|o+3pI^P(-~7ufHdB2@d9iN3
zh{>nSCWXx7e0Fu(mNIqO)V7z+;qwY^Z@IQP#yiS8eTiCSNR&Y8jy|V-hyHz?7WK4m
z=i!@&@4B*m_sRbA=&jqV2KCGqEr-o9S@T_<ukqd={gQ8C&@!Jxb{0AJY7=7*Pnu|?
z)i8b6fz2ByEMDTnut&W&{8~xYzh`H6Z=cX1+sh`Eu{1$V<DT=igcI(^#9TEGybpOG
z$`QF?MYXxpHoqb<0kKKh&mtVG?>Czz2`(tzw(;S?xp$P0FOcdKv)gf^x%3}PV%ZGe
z#cf7EE__kEtka&Z;w{`Lna$VRD`&Mo-9=k@VItq}!mzrj=MGJ?JlgW8qex_?c5;@x
z|BAqS$qyF>1aW(84fD}9k$Es9X~V+{r{eMrjCDMfSo*JQ^G#(4mFi}FTcRucsOFm-
z*Rn6350^~O>z+5oA&4t;hp@mR;~C2v*lssp336AT%=dL_>5cjsaZVd1gneCcbjSHu
z--3b{uk5`mdBZS0s7&^)@Uuuo#eP>sS(o!(l}wA=7iG?wDHway|7uN<@`+_vEez)t
zsYmkOp0G_-bme*Nt30au|2r-f$SAJvi~A)~ka1Vge$sL0f>kAJ)=uiqFqrxH{m%GD
z8rCN(rYbG$-qQMAb}>)Ehwiggg@5!u&W^gef94C-w^C1TeUe%zTzx`veWBIFIbpqA
zp9|T{n$1EaSYwl(=ah7CJ$<H9HC?XprRqUhL)DoveaDn*f;QHFzOp^u=tlbb>hpg2
z-Q{zZ%-!C1E3B;AW9rAoL`Kh>Zb4Skg_kp>8o3=m&t1^O`fdB;e+4d^4ofOJ<%v%J
zVqKoG%ei+&nd<V?zh5^zpBr*H$!O{1zeoIUeRSR-n18$H!2+T5N@n|hJ-6gZ+Rj>6
zi{|O(Md*g_2>*8;m~M{j`|EAvZgzO?^7@<ke~;+bUtE8_!ge<AZo}E7j%k}O@4cy`
zy@AKrsqTY)&xc=*U;iprOnF{h5|ViSpSRiJi8lOCxc5qwD2e`^ukvqyv|P-TH`49g
z8x9p$<|fGlbLS=J`v(MmA5r6B@DE@J=3?av2NOUU2q;QTE55_Xz`*ru!M$d0(_j{!
zU=b$~U%0klQ72I(o|uyu5>LTG!5OAfaDAs#>AF{knR$W*oCI89!aTu(PJ&20VJBfY
zPeM&ntzb*24kJ&nqJ^RtOp8LBv!~M9r+G{~!Te7A9x!2^U^xppI~dQg{SCv<FD17a
zS%RfFq|9N$A$~&YFDBZs0*&P6<mQJ7@6qtoTzYfEIgqddw<ApWP>9{jc=_wcfx;3j
z5=t;(7|qB94y)cnzZa}|XsvMBo3m(!P0|yfyLg>=k=!L?Ap>{U!ykETe>;Rq0TuB$
z@gOOZwvdJ^ddz(Oo#ng5Z-MqAJN~Hw>lb!@@otbSMY-i+&iJz7<EH4wi<-cw;dA0c
z3VtCcAtauxg)BTWc!K3E<l(%c)SN_6Y!??L7b}ApK_CJG0$~gW=HNY-yJtOEm9my8
zFo+>JHzg+nDCeE#%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$r
wrEI1lk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEggvrg%(Sy+k0JN5~i~s-t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.32DD26A1AFC28FC7.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.32DD26A1AFC28FC7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a677cb2d627c948b34f0ca9d19c08dd0b2676cd3
GIT binary patch
literal 1624
zcmWIYbaM+}V_<Mj@vO*AElFfyU|<jg;^LB`%ywo527WFEhKf07Lv!<Qn+aU|uX(O#
z`k8FjAI^_%t=_0+xJ&G(M}Uf|;3AJ+{koUy-(TGHIKp$Iu5MfH#LDj<;^tdhF`r<_
ziOf39VsQL*W@z61*bO}Q&i`BRdH3tt{l!(ACH&dCwx@qeGTpHuBKDtAfX_O%u43`i
zyU%+Tey^+h{IuMr=4Y{A`22U`U#hmw=?ipx+5WU^on2gg+@1PeORsNVQOWtkVD{Go
zldf!=y5yUW!TM`UHcP+MF!p)B=Tpn=ZJXzbMqP5VQa06Gbg4<>aH+^A(QoUwUs+Xf
z=B=OHJALDdgs8{j`P(nC*mx?3aLy79`mJJ>9{VfSdpq}5qfnu32Nf&%mzL^!yPaCW
z5cf*x{4uA-Nk)tx=9$DrUS9okw)y*ePF^|rCKnvkS&bb(GN<si)g<KvPGS1v#v0c=
zLACqW1Gf|BJkK#5aw<A2!=3S*e?jjJmKSq!b1O>Jx6eopSnI+qAM@m(Rej3_t2N6)
zoh55B_!Tefw5NN9sWBN}dlMyX`=!pmBh1a|w9Kc%pt`B&4$ZPW+VZHQNMz>fq(zS>
z1gze3{Fw8~C1Mk**3JmausE>sNW$TDliBwkJheepQ?X*vw%9D@6{jL{GS<JIA+UV&
z_W+}%{aosMOR~Q#VNqLZz+2AZRKoUJ<bsmzjHsj%n|Fcn(d~Z@KjCU>-Lk4$_gL+<
z?bAJaml>8NCmf#RdG%|iyy1B-PFEf8mkL#3CzS<k)n*-=DS5H{&yBO2r5(4;J=Xbh
zs`N>zi1sbID?QtH1)UH0^l7~Y?{UFRS?`|)F!8Ou{rLBjj{<X6$-I8`L`g~~?SJ(n
z?>DLHlIw$2N+xdmu<*PFv&^5Ot#jV)FaD#piZ5TS%5yhc>$W|PFJ{<^^q5=BoiQhD
zE!X94o*cH=6^2~9T%P5WbZ|XgrcyOsuJNSkL0Lml$@pG9w;xNC>MtL;oiCZQ`TE}U
z<J*r~`&~L0?VhXt+jhyR9|txxE_tQ2a!&7#mnOXjmM%Ez7I2X1_w~Yh3#FS6dIS`&
zcX|JsYi)9Oq1%<&U6)t=+;#g?TvmqVJexMjkLylz@M#oI+y3G8i=q-ou2he;R_8X~
z)Xjdv^!?=j*d6ZeC$9US%TF=BQ~0a@uW9<fA2aRee^fi)9((@ty~=s=@1ItNCEc`C
zwEq$BTJd?sFT0!mud4g{vYx!ye<sbMYWczH#%4z5^e2B?j{G<GZsML;%go}k>>NXJ
zWp0u@Fkc2&@Wpl8mhI$W@DE@J=3?dY2NOUU2q;QTE55_Xz`*ru!M$d0(_j{!U=b$~
zU%0klQ72I(o|uyu5>LTG!5OAfaDAs#>AF{knR$W*oCI89!aTu(PJ&20VJBfYPod4(
zQ)%tfJSLuCekXnpgcdmqIXf87vHcCh&o3pn7+Hd)IHb&B!Z4bV3+$iXL%$cSd1$Q=
z;wPm3VxkQzBTq0lCpSM#)gBE`&80UtoC66fa67_;4~5vhjF-QD94IWoBB2BmzU<9e
zG{Yw83D8}<PP|C&lChA1yX&q)!M?fcuUP^0Dq1Ld!Sp`-k+=4@L#Px`FOL%sl3r;G
zX}I3U%;(=(zFYhjs2AB!PYqbVu=9&|gIp=fEe|vD%Z87eq8l%20;7h{i4Q6Gg`9+t
zc(N9<@W|i^mbZ|H^NLb)5<#(DT$Eg_45XQZBh(-9`076S!xR|A5ClmGATkiaOwLWo
z$pCV_(|nm(SlQS)IJtRv1^D?q1znxh1i3t&gv5nKBqS{?t!-@W>>cz(wXKxRG~}^~
qib>1JDkv(csH$sf>F6668k?HSnMmmxxrmCndxbE$`8j$p`TziuNR=J{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.72C23D8E55C505F5.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.72C23D8E55C505F5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..39f18410251c267159574e496842ef7c456ccf4c
GIT binary patch
literal 1630
zcmWIYbaM-1V_<Mj@vO*AElFfyU|<jg;^LB`%wA>&1}!cIhKf07LvI(|HWPUIKP7BJ
z*7Z`0#q7}&rF@g4C#yYn=iz8jX;V1xXp#3mJMO?uIk#q-`Q3WS-FmnAu5ER*Q3KC`
zX-l{HFcj~Z7xwhq+Zm}J^1iXpy&o4_KjZT2qfe3+?5I)F4{PS`zIs!i`RLJ~4j-=^
z{q!w3=+3^`@%8^?`RD(ibNTA=-K=|~be9W<1pKPoxo9<i_<XT)`TQ4Gd9&h}?wwDw
z-Jr5HS9PVW@R`-7FW<OXo#MG{z24^Gwzq52Tc>TAc+SJOYsD6YQw6fEHLZVN>u%MO
zDc(J~efMUbcQdd3IC^`n38Vidj?kvG)+=_M=OU-qUBC3!X|2!HmfV6HdwyTqHBD7H
zHH1OGv@<<<p@Zj4h6;bF^=n>U{d0Eq_URKk-uALdy;zc<rcu|h=HP_>9^I~>hWjfH
zWD9Ic36#IXantf-q(g+qx|zw0Z~riE5$tA|yLl}y|GOt*mTXtOwtP6~D!1mdJICh3
zIZ>ul4z_<2FgjaQ;kL<0aK^^8>$hGk{9f&}#YiaX*#>#_r{XVl*31m{(dg4~pSV+N
zS(yyS)>G9wHdDN^7*}5F%4}_8mI>xDdt+63LHZ1j_au&k%3I$CNjiicOLDA^J>Ft5
zcW?V7&Ds!)mrJ7;tE*0!;w1HsqvZw9GtC7@zbS0Ja`EI6_u5?9Z}NKQ6eB!V@4B9J
zC%v?K<;sgD$MSe`W*BSkp7%}g^O`9EM>b92Ta-T4hUud6MU!)9I`%y}o@KXFFy(gU
zoEdRDC$4EM^|(E$tK@|CRUTFS|6P{~WE59_i{sOLaABA2KK1_g2O&=*!qvquCK&y`
zAH4raT-l_LR<2IsncVNI-6S78n7;aH&i}cO&9}b#khW<0mh6Cx)6y+{Zx$x(Ir!40
zC~B*g&vD5cJi7vtMM^s7Zoa6fHD&gs7qL|=dgnObb*wz`q{c+<;pDErvzP7rJMExZ
z_1)^Sn&>@`Ro?C?y;8R6zVoUcejaX*ReF<7-Rzj{cbQ`jXH*W0sLqD>eZLP2o@7gJ
z;n*1#b1DDM!k2uei|&QY2|GLe?zHbg&R13E=BU?Q>Xl#ATX8*=<@pqg)6qXlOU$Gu
zJ4ptu?)#imc35iXfvunF#lId@-1heVLd#8(c{YEK=gn5n`{|umdu+Ci{5GZYza=W`
z_Dp{s+39+M^XEZ(gF~~~uWna=5IF1n#lT+K-}*Oso_{py)!&kIfn&|j^H2WPN6W-a
zdZR70bNw@h;>z44d0_4=|0bpybJS-I4}*UILogRBR~VQ8%0NI-YFhCfMg|71UkmOv
zdz%Kc@C1uEiTJ{`1&ca~BJsqW#E^Ij<_b<Ql?rXno=R(<<}vXE^E>f-z=U~%<;>-5
zVLZq7Hw-_&l-y!u36|oJGJ^?+_z9`Mm}tYw$P>)X$;}TFPW5%%+qiLA5;IS*fRg}{
zk%CTwNIYRDVK@(_m5~eV&E7-57p!?`txzW`annIQfE{RoqPe0cOw|<5%lgyT1SJB6
zk)5+g!&7tV%?;;3&QaiYgsDChV)rs${`zsCump>Q5={8AH)qido1`Z|kMlb5!d=1>
zEMqPM_xQsfd24?=gh~Mw@i_4yDUvpqhAVn%!1{%qU%VS+uPC=X%z`f)K5mL`yr>C`
z8a^jJq~I5F5<=q1n#;l?gC|(tTprFVO3g_G#ddK~a<MXyW)6OLf_Zl3r?Nt(z#xV|
zNJ0RSK?r7YZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#c
zpeL$rrEI1lk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEggvrg%(Sy+k0Dc{(
AdjJ3c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.80E125D05CA137B9.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.80E125D05CA137B9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b08c378f055cc8e30eb019d39674f6c22b9e85c
GIT binary patch
literal 1622
zcmWIYbaV4#V_<Mj@vO*AElFfyU|<jg;^LB`%r<5Q1}-iJhKf07LvQEZwh*}bUnKe1
z>>M-E2kl3;Za-$$c`<jQdy|r8r+{l#<;SJ*_ZOxf@0*mgJu~o4_nDc8cg~ZvVm`r;
z6Pb0I#o+jB&(OSk^EUF_^ZmbQ^X}KP`HicL<>UE8!u_B8EGkIcXY)UI!Hzu=8uNBV
z+^O~I+FibR_WAlp4?o{G-u0)#e);v&M^k#Zyzc8yEBX2G*X2L|9_#Mi=X%!r!0(q^
z*u|D?TdT4)y{U4m(%e{Y_nC=RucxZt**5p<nhTT6E=jKR6r3vNBDrqyg!LQh_bv@<
zP0p+SQcyqdAoFFZ{S~{Xsy*<OTIh9Q&4gR(7cF03dhfL>cg56|lQwYjovhYan;mHs
z<i`3TXnKX27}u03oDJWXd^|NP?(F?NmDQ}88|PkhV4f>-hUMt@38oUCj@($J!TMt>
z*B&-SZ?UiUdMC`CJcqTV^~o$tu?y$kJ4$clykL3z)`u5<c?QQ9M0JYU?Ksg~`i~*8
zG{Sdvo6(O8@0^!K@|~Zw#)l<s&CP9QRWJU^39jvFnP&Oq@rpl6b6Z7=yEu;uKGiS^
zKk9PaZedtue{Xw`r>?@+D8scE3L26-54Pp1f3Iqun&=(mRIq4UY!>s1QxQ2C>tD|h
zIKKINfYDNZF7>@7+FzEis4X?%EoX2lVS6odLCJPTR8on}yTJJ9_CJT8va)bzhL%V5
z{9co<KT*mz`OU2Zvn(gCdVSe0Vct}su83)u9KWoY#I8`~W72DEyy*Rpg|q$hSZ<a{
zsV@s(r<$E0dTHAdwYRUmY@JV@u3uy#<9=e*-08KP7F+V#-%mR3T(GKW&FabB85?HY
zeLqou<K($1d;BI|aJ_Y?-EJ{Y!H3gNlXm?+_hI&mTl;6eP<<;Ua_bY@Lg#M*2KOG6
zBo#$%)$%!Kv_Z0D;W3RDLNhmCRMeU>d%}y@Di%F`j&~g^Py9$KF*~F#`g69*y?-YX
zChvV`YrRMJUg4y&dsbQ2H{UCU9%FCgRS6aMe42B_cX=kuIab|kO<ZC(%8&nj*pbp}
z?9>vYsrS<S-G!2NslfHxeyg9IE<0Ska^b62hNn8pW8DhuIydiWz3={fv2l=wbJo?u
zdA|9P%FlQB{ky-&*2-m9U0#*)?1yca|G!yR`KWLHuFsPBpK5M=vUy+o{O7~CdFiTR
zbA>-O*N3#2_piz~t_`|ZToR(}yx(Z{!#lFe-?4JBvCIkjF8t}=ejN@y=Wn80GnRfT
zuFOr62j)s1Rl^<&%b*K94E_NO!Cb6deqaJ90|7;;X~lOKfw}hAf_u&0rok*c!6Hr~
zzHn{9qE4bnJTWIRB%Xq~f)iXN)FMVMum!z`elJ+_&{{#E&Dm3F?bAFao?w0_eh;|n
zU^#O+TNqEXB;wHStnW6AJi&_Qik>iG$M!c2KfjdRVq^)H;*c_f35WOzslS+L!wOW*
z&B@IV6HfJY+}pTuSrRi(uz-^Ql1+k6f=E1JCt)~m3g>10>1%=#fhHkaut&pFbLq_u
z=Rg)Ha67^*I22;{GG6}raiFjSi-ZzP__8-=(F~iUCqN(aI`P8I<_VTDmx256;g7tv
zza2uQfQopWc#srHn@htLJvCtc!p<+=4YF61TOMY?mkl2`MK@m51V#;?6CYCW3poiP
z@np?q5s?urZ!Qn#6{Y4Rf?~V4D7jb}NHYg3Sc)#&(y;z8Q(zE7AS5Ax$RGqWIX5LI
z1IYDG^JQjXWn<^y<mTZO;OFxcbahq}<nnY95*HSckhHM0wz0LdchD2nwo*3JkjEw}
pCM_eYps1vxs;;S}qi<knY-%oNBBg8OA}Z$Y6~g4^=jg%c0{|9eg^B<G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.663BA57C2C8487A1.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.663BA57C2C8487A1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..232e13b8fe0778cf7f7bcf32dd82d46d461306ce
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nf=TR3=v!m3>9<EhUVtqw&3~pUsA&K
zdR#vD1peG<ce<i)Cv0CR&l))?_u!@{i=XUybG^Uo^Q9%ae6Mp4*A}`2smOm)oXVrX
z{J}MGZ-7ET*yBwRd#!~f_ssuuba{FBb@}Cgs?7G-1n#k&l>T`#_vW8J^VPdfcMBga
z?3lg#yl3I}x~jiV{jF+#R-bx4|6Tf@tU6=4Nll;m&5P~+SA4B3`~PxU{M|{<EDyJz
zTw;D_!zuGwp^h4YXPUOwt(?fM{>Ey?fhGA-@-BI%>rJh4BkxJBo>6$I%xnLl|6k`_
zd!kWL+fh(I?_l$WY5PicU(~zgX`RZoLFu}gir~fMU$Gf6$4~7r;#}R~a?knUM!}OO
zErb;_qA%R6JmG1i$8`6yp2YjAtu>#Y-HrA*VCJVTm9bSwPim)WrMsrQWwp@OrVS>6
zVGBwWt#<P+F6oZy@<>~{R8&LhXY1OF<yTlTE}GlDJN!~Kv1i%ECxS|4MpZo4!lCY0
zO`e6c3CzwtWper>=ZQ<D3UkU|A1i14wS&Fe)X!OY!Ve3L{aIf&2xPPvO<FU_r7?B-
ziyKT?J{(snG<vpi1+DMh&?e*mCM4-g;oVnj?**n_nZ&{&x3YHjErw+!^Q3OQj#_x-
zp$wDAp&N6VN-ykK;Nd2y;cQTl-XUwXD%jEai@=s6FA8o%J3pQHqy02@huo!Ad0VD@
z_q;z(d28|3&))3bje4Qi58u1lq+#+THZY?_uOsL^Ppra%%sH*Wwa3fv+Uyc6y&N@b
zM%>OTi`ra$Zcm!EWv%wto+S(ZzuYp}#OYqnq8#r93y=2JGSA}^^zB@<HmKuVpyBNM
z%<);ui^Vp1&D3zYb*Nn~kj>=B(LZUMe)_$fz2ersnIfuhrJh_nJ<G@X%LMMo%Oye<
z*;2yC(<OiK+=>Vk<MKT7lkemzp1w$N16!u);yq<Yrsh-_T1a1*?DeNtIKMt+!>PFM
z^WNT@)-Y-5-0vr9SO2zsa;#NjMdO+r>##-dGcH$3H5MQJy!4Tm%faHwCxm|7WvGkE
z+7>iTbFV?r=jrb*P5oQ@E@Z2Ys{ZTsucy9QI_vq-4@L9K6>gg)o9yD*oAcoN!JiL$
zo*vtHYD-Ys<mkZHlDSbow(t2p=l2Phm;OvPA0%zr|LkVI?%w#c@WF(WRVG%d878kb
z_qgq@UK|(};Oy4(_xOKvwby%JJ>6Lm@*?>40;O+ts<SR@-Yx%aIW2y{Blo8N|KEwc
z_5a3u&gsZ<hT_WHBza)|+^V_4tuX#)4G)8V07EbjD^CQN0Lnl>QEFOo6%zvk*RKWl
zn!QbfnRtT5oWy+L+JeQM#Nj-JHfK+zwNLZdc!C8y1pE++1f2vu5IhBU1-KTTU`0no
zcZ8g#uBI`9r=_l?3*$MqzhU_KrQ{Y9ORy}5tOHD#C0JEiH5$hAnz_IKT=N@V7N859
z1dz-XaS}n|DL5*)!L;xMYpQGNz<2@QE7x&ny?hAtC$cR(D7J+738}xBXv50L6U@uW
z%Ma6=tG#o2v_yy>P*}oA!V4zcoFL`4+R(TR=q4d2AtZZ+orIBiA|4_>FfBa6qE4c4
zxwVh){N;ZvYymV^TU{G23{%X=1&*@bL%$cSd1$S$N5fNd>CFx2K#o)3b%d!p6k_)>
zUjF)VpvR<Hq?KU8Ji#g~DsbLqZ_c6_Hc3yId4l<!_>rPa-ccSY_%zft;L-B%N8Z}s
z4xv&&EqqRVNLu6^<&d<ftE<DcJT+kb!p<+=4RWC<uRP2tKf+Cnq>r&X1BI0wm4acy
zMX5Q7ptM+Alw7O~Vg!K*FklY0tW(X2Xj!&`DKLm35GKsPker*6lL6#<r};9ou(Gjp
zaB}nT3h?uJ3c5P0337Ql35g4fNJv^(THDy#**oZoYFjCrX~<&}6_b{cRZvt?QB~K}
e($P0CG&VJtGm+9YauF4C_X=Th^K<lI^Z@`tzwT}T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.FDD27F96EE937A50.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.FDD27F96EE937A50.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d03305666e48649e0e25ea1440e2a926b276693f
GIT binary patch
literal 1788
zcmWIYbaVT{#=zj5;#rZKT9U}Zz`!5~#Kk2=nSIO*3?W<$3>9<EhUVtqw&3~pUsA&K
zdR#vD1peG<ce<i)Cv0CR&l))?_u!@{i=XUybG^Uo^Q9%ae6Mp4*A}`2smOm)oXVrX
z{J}MGZ-7ET*yBwRd#!~f_ssuuba{FBb@}Cgs?7G-1n#k&l>T`#_vW8J^VPdfcMBga
z?3lg#yl3I}x~jiV{jF+#R-bx4|6Tf@tU6=4Nll;m&5P~+SA4B3`~PxU{M|{<EDyJz
zTw;D_!zuGwp^h4YXPUOwt(?fM{>Ey?fhGA-@-BI%>rJh4BkxJBo>6$I%xnLl|6k`_
zd!kWL+fh(I?_l$WY5PicU(~zgX`RZoLFu}gir~fMU$Gf6$4~7r;#}R~a?knUM!}OO
zErb;_qA%R6JmG1i$8`6yp2YjAtu>#Y-HrA*VCJVTm9bSwPim)WrMsrQWwp@OrVS>6
zVGBwWt#<P+F6oZy@<>~{R8&LhXY1OF<yTlTE}GlDJN!~Kv1i%ECxS|4MpZo4!lCY0
zO`e6c3CzwtWper>=ZQ<D3UkU|A1i14wS&Fe)X!OY!Ve3L{aIf&2xPPvO<FU_r7?B-
ziyKT?J{(snG<vpi1+DMh&?e*mCM4-g;oVnj?**n_nZ&{&x3YHjErw+!^Q3OQj#_x-
zp$wDAp&N6VN-ykK;Nd2y;cQTl-XUwXD%jEai@=s6FA8o%J3pQHqy02@huo!Ad0VD@
z_q;z(d28|3&))3bje4Qi58u1lq+#+THZY?_uOsL^Ppra%%sH*Wwa3fv+Uyc6y&N@b
zM%>OTi`ra$Zcm!EWv%wto+S(ZzuYp}#OYqnq8#r93y=2JGSA}^^zB@<HmKuVpyBNM
z%<);ui^Vp1&D3zYb*Nn~kj>=B(LZUMe)_$fz2ersnIfuhrJh_nJ<G@X%LMMo%Oye<
z*;2yC(<OiK+=>Vk<MKT7lkemzp1w$N16!u);yq<Yrsh-_T1a1*?DeNtIKMt+!>PFM
z^WNT@)-Y-5-0vr9SO2zsa;#NjMdO+r>##-dGcH$3H5MQJy!4Tm%faHwCxm|7WvGkE
z+7>iTbFV?r=jrb*P5oQ@E@Z2Ys{ZTsucy9QI!pZU>DpV3Ik$T*J~8^TsqsGh^Je2|
z(#N$fd(HICb6uOhCF9}!OYef;DGM*QXMcYn{foqh_kGdwN2j_MI;*|D_+rZ3i!1z%
zq~5;kb&GU-D`EIe{{NjBYre0V-u@_XuK(KRt{uOpX85_6+<)`jqOP%1p7US*?tm>e
zH_cDB9F{Mx%uSL9=1zsB$E&ybD&OE?@DE@J=3(Ut2NOUU2q;QTE3RT<VBq?-;9j%0
zX)qH{u$YsWFI-!&xRW@XC%C>-s&w6}!z?_(LQX=i2t~q9!bm(34-s#KoT!s1Tu!0Q
z*;8rl(>ykwU;z&SKZGJdCqWMcPr*&W2f<TxP;^7^G~G0j<g^^L;BqiEj9g$Z_8$7Z
zV9i5o1;_R`3_rh=++t!0mgSJOhpA!-R^w2E^Soy6??2c4h8O5F0Ve@>m?EBF5hoEO
zo`Qn{+%G)Anhu(9UWlKN`iqG+tc*OtyqvuJFavV6cTSI%2=N07OE^h*!Gza7zVny=
zv9JYDSldB69wxj;!&7tV%?;;3hA8kl!h{cn*u9LGzkVDTH_|N9N-$xbU=<b>IPbDI
zXVDCsq$kWg!Te7ANP#KuAP)~Bo?s0J4LI-NkG!?N9YUpmTKJs!khI7-$iWry1gkr!
z!+B2)Sii9Ii+6)uD9S4jbIOl!6C>$k?9M=8B?qM-nD8%;Z&{merrriR969KDP=dZF
zH76027K@9Li<N;ib8wyvzfs-2omNbNK@35VWC|h!5zOS=l$;D8*E`LZnT3^&or9B`
zhgX1~&r{IVSxu13(@98NSVTh7!qVEt*3RBRPgL7V*-S$oo2ZzyjI4s9l8UOjrk0Mr
bfuXUfxtxiVu91tVn7da9lbfHT2cr)F#v|-!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.33C6E967995C7F50.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.33C6E967995C7F50.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5885bec4d7d8df3004251276efd37aa00a06fbc7
GIT binary patch
literal 1790
zcmWIYbaVU0#=zj5;#rZKT9U}Zz`!5~#Kk2=nf=TR3?W<$3>9<EhUVtqw&3~pUsA%<
zd;NCS7wu-=c?Z|#Zn%A+ze(rhwuHz}fkkoOUiWu>zO+o2?{)6s+CrBg75Pt!Q+X7a
zKe#6D4NwROo1VI{u3pXPhr!Ru{`u?I+57(dnp07@sN%cFJY7G%Gh&Okt9NyO7C!n|
z;PY*L&%)n{Re!(oTh;u{J|#Z?ecGQa!~HUonm+e`W{t1e^~-4Y@0!Ew^E^J!dH6Wx
zlJv7RpL)|m1-w``r>u?7d~%3$H>aAx%h}i1Cq|#m?)8hFR(?z?ZO)dtm)3u%-@Dw@
zlIxC)@||9z9pcw6P0usG<!Y6=+=nsj$*ndOr7EBO-o3LUB*WDbWhb84t#4zdWGR`h
z#FEQxWzBEtdCYY|^)1%NvFoSh-`i8q&Uxe7OBd#~VTW9m%hld%JY`PzUlb+~rgn;J
zik?e)dE%73KySg5Yqo4*a#h*C!T49-EC$_Q$BsK&>{aWzXfn&j>4=|Oe`5ZkS1V>6
zn{!GaX|wmtG)p_i$X$FbpYtyN&~uRgp0V**W(JFpy!%u0Ub$&Zy#a2MLMLf#cq1fh
z?a<5FD8=7;G0gGRR>5l%9(T_5>aoj<+I!i0RZr9;hJ(yo^Mj=iOgR_qTpg=D#cG~&
zgOkMW@CAHQ(?w2dFr{*^z2n{J>|Uys(IM9~O+uEhI(vnU(Eox$zKhOLSJzJaWcTuR
z@shH2Wpz&zR4*)jRdu5}TA;OUk2Y8D1=he<dlz!EnDlNgSoQwLzS;eGE_b&~Q(G3k
zZp(^9$;;c8sJ(sVWvi_DV!!oUmEH;Fw>5SPWn>g)-)~y(>=L{vW3|Yl;*c}nzMq(W
z>(q+YH9=}z6SsZX=RSpzw{GJM|84g*YlTC1n@?V}B)0iZRpjK%ez_vW>{)(H>AuO2
zo<_3SHJgP^X>}4R-sikYt9kkw?iq58-NkO!kG#xm<md3d@KpV$C!D`OeZ#r9@AKZ?
zpVlyG>D=!pYFGcZeR8Z-V@2bd9P6+}?=vn}N;OV=^wX<Qb<u;Jo++Gv@)+#(Ory0=
zPl=OS`O~}n<*A2u<*TBmPMZF8``M^(L1()!-gds#m~*@5;uE7Un;P%4Ej}oDx;J@h
z=F&52=8LWwM{W6#pUJPuKe6-T{O0lp##KB&%8##=KRUJlQJc@T#~(GV9~aI~vAK6&
zx_@)iEyJ`o_WyPoM#_ii^B--tnjgh3TKHZ2;+(!03vZq;{KF({FZAbs*}{y<P4x^r
z>JE!D6j$aZ$piD}aqSqt_uIbe@-X-ZFa-0k@`Qs4pbP{QrKS~EF)=W3{aSFZ+1oUj
zi6>ahNz50nEm+)19L`f{bM{nP`!tV@Cs@Ekzz?BF&`Hn(!BcQk@Imkt9TeRVJWV%E
zBsnbyEx4Rx`x}OzUrKH<u>{L<$lAlSumr1dsKI$&GxztOYktGa0(60sfICbPPq2uS
z2og`hLBSOvr|F;x=Y{wQslS+L!^+4L%*)Bk4^xz@y>oiBM2H_ySi(ud3nsjR*D~Vm
zuH;$3P(gMR4~m=CKECsp|FN(IP_MRwc05e)9t}^;r8hU61KFg&>j)D*6k_)>UjF)V
zppnun(n>I4o?sOg6*vzj%E$$dt=>bw7p!?`t#H|!vuK7*(i3K$V16flq!5>PkcWpO
zPq2oA2AucsN8Z}s4xv&&EqqRVNLu6^<lu^Ug4G?=;k>5?tY6sq#k)Zc7v+_Qx#dT=
ziIMa%c4wfll7mtZO!%)r*rvPrN1g#=M94`99?m?$!cM|SJP{8OZ<rRIU{NPgxLi?c
zP9i8R78fNKD+6if;9je*vrWtmmM{edF$6)9DToY2Fq3mrax#Ei?=)X#7FITP4o+?!
zUIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlAvI>eyDyr(5S~~g$
ahQ_Amawbx`MlPaa?p`5GZhnp)j6MJ+f8?wH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.59DE1C78AB7D5F80.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.59DE1C78AB7D5F80.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b6a79c1e826f48a441594ef00bb3c47e7e98a95
GIT binary patch
literal 1716
zcmWIYbaPw7#=zj5;#rZKT9U}Zz`!5~#Kk2=nXSwW4D4JC3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2U$XkqZCfP1xjwslxVrY;3mygL
z0@uVj0SbH8sHJYKELSwDu>ae9{@uEKyK_Iz)>v2t@3Ecu{Pbq-&rhG~r*~bvI`h$$
zqmSMd2i@5_|8C9CxqR{e=6rm*{I}))I6XDzpn`vi8=su+zdrx@_xX80zv^bjG1>Vn
zx80C1HG4v+qlVy@uB~~!^SHIon7uf#Bs)snC2zv@7rIi@%5P~bTWd1+Q234g_b$D-
z!@b|OYRCRF4;&?&<9FR%&|BiPx`ff~Nrscq-g3LY>n`LcFU_6V;`_zrp7X;R<w+B_
zSvffbU$|*?!t;_D(_LrjEt~g7?W_F!PIt+XGPiW8j6hE5t#_Qy1)Weo*5?u;aLwuz
zmx$OR<7#uR=>8y=CQ&ctJsm0H-xj=H!Fi$T`i>tj_{^syU*VnP+Oco)#h(QYE38&6
zyZGFw<;EN5Ws!a7C+Xy}AF|q=8za8pd$mi?Hjds8M~(;n6y4iA^+S|~F6UGguK2Rf
zzFMbVzD6njR=33pU!%AepZvhHTxfCkZL#lF-ctj;gPcCNyuEdW?ZVR1D;MIfb1m(A
zF0er0mW+rT^DoiLOI?#LO=+`d*!X$ctB?x<wo{^#N-W-746yY7bNB_<qUJ5DtaXpp
zUfVw1<84mv-CKvEDrT>Gec3L-QG}y1T;#<<?nOtKFR*%<^lsj;>iv&}v(?R&%+@aK
zd^t5+#B@#4t!Y=5)W)t%S7H6L+F3iZWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rTZp{nhQw?nUBlMQ-e&4gNbJgG3o^hAfIp)n-!)cYmzjK1Auexe_
z$>PUPl$fVGN^5sH>2+FOuXLLFP52t?e7VMpqKjoWs!GT4>al%UB2+*5$nAW|oK4sF
zrnhfbtvGq<ob~zG?!R(Hxy~Y29KyEAYVUoWak)~e(fjD<t&hB1KCXZKufSznV4G%4
zX6GO4aEsW7OGB?lecJSI*WD9wt19Do-1`IBHsu<-eLlHoGVdR&9WpT+Vxq%dhMmf(
z($1Q5*{XFv_xBeQUi{kmk~hXp`rQAO`+psoZ}U~$D6;?fb6ffT>gSI?KD<{kM>OQ?
zjE(YtmM3kh3#py|O8u*?JG*G$m-w_|yNcr*s~ekx+3M2H{h$7Jf$xcZjN8^NzQs^n
znVTdJ%$48tC(K?^;?>8);2*#c%)`p#4<>*z5KxqwR&2?{z`*ru!M$d0(_m(vU@<2#
zU%0klaVK#ko`jPGoF^E1DWhNs&jz3tJ|{j8gcdmqIXeVT-9jDCQ>y#9DgQ#@HYT26
zSq@oqgd#N#H8{_){SCv<FD18_SWpzP1S2bQnf&ENV3F`Wpyh&2f=HGtS}4M;;|bQX
z(1P<q{DjnBOtfKT<O$~G<mHE1-kc!iw%X9R3}}*&laL!sm?v1+Nf?PI>Ld#1?a}bm
zTzYfEIgm*TypAv}heGUL#>-zn4)mQgi?k9<m?v0;MFq~g?9Ev;!zSqo&^mr6ek6y>
zTgbzm$`h<%p#kSz-7fakqI4rS&^jdxB`=t{PYqbVu=9&|gIpoXD-RQX-x>OlS6IFu
zD6DOv9S#%zvf<;V=*EkhEIh#iP6F^q<OvpW5<%iASSUEdwD1ILT4=&~Fi}P>a0==@
z^n1aYht>)VMX5Q7ptMw6lw7O~q?v<*1K<DF7tGRP3JhWhf+RB#8Hive=ceRj0J+|2
zzRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6R?21?^4LVhq-A6k
k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_LYUnA96cC)0C&UJ1^@s6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.F3F4F9DA4C0F8BDA.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.F3F4F9DA4C0F8BDA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0952cde8b8bd1055ba630636a76a371bf6d68ba5
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=nJvr=49r{%3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2U$XkqZCfP1xjwslxVrY;3mygL
z0@uVj0SbH8sHJYKELSwDu>ae9{@uEKyK_Iz)>v2t@3Ecu{Pbq-&rhG~r*~bvI`h$$
zqmSMd2i@5_|8C9CxqR{e=6rm*{I}))I6XDzpn`vi8=su+zdrx@_xX80zv^bjG1>Vn
zx80C1HG4v+qlVy@uB~~!^SHIon7uf#Bs)snC2zv@7rIi@%5P~bTWd1+Q234g_b$D-
z!@b|OYRCRF4;&?&<9FR%&|BiPx`ff~Nrscq-g3LY>n`LcFU_6V;`_zrp7X;R<w+B_
zSvffbU$|*?!t;_D(_LrjEt~g7?W_F!PIt+XGPiW8j6hE5t#_Qy1)Weo*5?u;aLwuz
zmx$OR<7#uR=>8y=CQ&ctJsm0H-xj=H!Fi$T`i>tj_{^syU*VnP+Oco)#h(QYE38&6
zyZGFw<;EN5Ws!a7C+Xy}AF|q=8za8pd$mi?Hjds8M~(;n6y4iA^+S|~F6UGguK2Rf
zzFMbVzD6njR=33pU!%AepZvhHTxfCkZL#lF-ctj;gPcCNyuEdW?ZVR1D;MIfb1m(A
zF0er0mW+rT^DoiLOI?#LO=+`d*!X$ctB?x<wo{^#N-W-746yY7bNB_<qUJ5DtaXpp
zUfVw1<84mv-CKvEDrT>Gec3L-QG}y1T;#<<?nOtKFR*%<^lsj;>iv&}v(?R&%+@aK
zd^t5+#B@#4t!Y=5)W)t%S7H6L+F3iZWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rTZp{nhQw?nUBlMQ-e&4gNbJgG3o^hAfIp)n-!)cYmzjK1Auexe_
z$>PUPl$fVGN^5sH>2+FOuXLLFP52t?e7VMpqKjoWs!GT4>al%UB2+*5$nAW|oK4sF
zrnhfbtvGq<ob~zG?!R(Hxy~Y29KyEAYVUoWak)~e(fjD<t&hB1KCXZKufSznV4G%4
zX6GO4aEsW7OGB?lecJSI*WD9wt11m;Y}q7FY&$E_tNz)~?8EB^MFj_oawC^UYi@qI
zc15AzZIOR^w-!4my<VPXULg6_{QvTIuY}7#9f>`WVk<Y_yvF8TZU6E1>c@p)0cVqs
z)qTkCI{7Q`Ywf)GQ@+c{hPYg)pLzD@71{K6tXzJaKW19~_YZ5nrtq81XwIUb;>z44
zd0?L8TbQd7k^c7)4}*UILog33j~|!-%0NI-YFe=+69WU+uLbv-y-kCed4k29#C+k}
zg2kQ0k$4hL5^x?&4I>xW$lgQ07p!?`tsoeBDWhNs&jz4Ld`^5GFjYLka^`Zj2%fsR
zI-DoCzEi4n-K)buEkaI0t_VfKPQpk$Q72J2&tQ8pw^e40852*iEQhQaLW>%Q8l30Y
z{)XY_my%mdEGUXtf{_)uO#bpBut@kG&~ia1K_vGonk&Lx$P=t(t_A0X_z9`Mm}tYw
z$P>)V$;%J3e2<2w=F*!R&Vhs#cpYKFheGUL#>-zn4s@zCi?k9<m?v0;MFq~g?9Ev;
z!zSqo&=ven{79~lH<yRo%M+|&t^wy=-7fakqI4rS&^je^B~O^SPYqbVu=9&|gIpoX
zD-RQX-x>OlS6IFuD6DO+9R?Hrvf<;V=*EkhEIh#iP6F__<p~yX5<%iAm@7EJwD1IL
znrp&&3`MCqiJ-JpT$Eg_3}OU<2nYy-F&LPGzwUf2`sqUQ8>YY@hUDCooD86xcbYFV
z3o9Et2PZcVuK+)vr=Y8|njn{_laRQuh=intrL~Q%oxOvesJ4}|nT9+zQ88&5Sp`KU
i6;*XjEggLWLt|5OITI;eBNtIIcdrm8H$O)YMjrs->&cG*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.A371B626E37E8E07.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.A371B626E37E8E07.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7246b0fcbd7ccd6e9e03554bacdd7f75e4add066
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=nJvr=49r{%3>9<EhUVtqHWRq^U-Mkg
zYn^P~AI^_%t=_0+xJ&G(M}Uf|;3AK=9zU1vuX8`$m#ltt+ZKs$uFvisuC9Ief=7Y*
zgKOg60EK|C$(wHM*)JejGvnX6^z!TG^V9yM+3m3k-eW!S`RUEvpP#;XJM+=2Pd<gO
zk_x}8YQD2Oleh1mTl@O@Ge16k{yQgLUwpEnrp5mq5uctef35%d_xbIAeoZscXOv%l
zSuSG3Df3yOjv9hzy0+G>oye{J#%#ucCHYbEE_tTwO|5bx?@6wnQFy7$YyY8tU#DGr
zqES%&rNI8w1IGyO_+581#9uC%Ys4BcF<W|4#|y*1>rC>JQ)6bf__`>)U-;ohM@q_^
zPJu1DFW%TZQIQmHe3vQSQ@%H9U*+e2vP+JXxur{GY!%Xz+S$Hm@dW=K+s-J*h}jz2
z0lNg|-rXIz%T?D^B_cDEwacR{k^fiUEC$_Q$Bw&~S2<g{UC!9VA!sdW+s5p2HPE!z
za=PM?G<l=5MK#k>URnv9+nxJ(H^bih%1OPKFL0>X^_|L>vR%U>wXjDeWRk{)H$t+%
zTcm`UOy#&1uWh=Lt-MC@u<;zP9=p8gz1OW*^+Zh)YEsX-zs#&bbMErCZ@a=YN^RO1
zS`EG}aF8`wcVcqjlqm~MJ}?~NlfAlffnt@$)*~+pZY*B-$>Y!AtEw*inW5!TUBB1l
z>rcG3^VZL;&08I&uX=siF2PYm>eD(66DM((mHZCk8W&B@9a_2XaeLmpI~}_+b$xx-
z$7BZ{)>@Y9>3i$?)GD{23-x8UJf#%#%i_MNUbygx@4l*kd&e@Di(w~SpRbs4_x(iu
zt&;<}BbRzSQObSjJU@`l<j2t>!@ITVUwlK}-uL-5sf@Sib*k?L#p)BSYcj1gEwiPB
z)x{-$@Z5?B6e;pN^OKJ=ludmN_YAql)!uH_kG#zG$i(ox@KpV$*10|2_(sb5>hpa0
z-VP@(ovVJb@9JN<qQ_1mR~*8&$!cFL-}2I?_khR8nzcKG0{^_;VJ|Z=cS@?)`kpI)
ztivs0A1)2P8ue+@zioF<#I33{tokjHqHx<R*<>Ei-kb;14`vh`EP5P~8l}1U<=Pcr
z=WNsb6T8K|{m3={vw07UZ{`2<zk5~v-ILDU6HZpy*yaDIxbs`Cpa1)@$7>eMI^Og5
z!F`d*Ul+gnZDX(b-qLDC$AbSxvp=u2K6i&(WRB1W<Klnw*RZX1{3fo{u*j>pGB-&c
zm?xV|MKg2e2k7xI_y;fq^RV*xfeD}t1QeyF6<abfFmU}^aIe|hG?<wuSj<Vx7p^T>
z+({gXC*dRk=Lv>h$|zXEvjM1u&xy|ip+(MI&KAK_H&=)A9NXV8{QOdKi-{#zmP6JI
zridk2jYAF2bD8|*MPQNeJ)pUQPJ&40Dw-?8P2vgGGS`ChLi~i(Ure-NW#kFw<>cju
znG2&Cxxk+8J@k9Qnupd3do(;Xm)_iP4y0Ov*Ab@bP>9{jc=_wcnRtSwS)`R<!aTt$
zEGlr`g`3u0C0nNQ0u>?qfCt3~m%TZQX4oV>0s4U7i66-a^5*hzH}eE*m}|g!SGS9O
zwJ6=l4Rn)|xsoT$@}~x@U)cG@yFpGB<&}pCzwZov$SW-04;0om*A9aTf7$SHQ*`4+
zO%|SD0Ve@?MDYZRIEf(f6wDQzU|M*BHO)2QyuSiro9^Zxc?NX2kdu%rLXohOFcMGH
zNfgdwC`!#q1f`|oqU2&_5F-dgfB|zb$JOkDS6^m|Fa-uN1j2+F7?N{Sax#Ei?=)X#
z7FITP4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlAvI>ey
iDyr(5S~~g$hQ_Amawbx`MlPaa?p`5GZhnp)j6MKUji7q~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.AB40F2DD408B7367.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.AB40F2DD408B7367.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d69c3185b2869b992cc3eab89fbff78732cd89df
GIT binary patch
literal 1716
zcmWIYbaPw7#=zj5;#rZKT9U}Zz`!5~#Kk2=nXSwW4D4JC3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2U$XkqZCfP1xjwslxVrY;3mygL
z0@uVj0SbH8sHJYKELSwDu>ae9{@uEKyK_Iz)>v2t@3Ecu{Pbq-&rhG~r*~bvI`h$$
zqmSMd2i@5_|8C9CxqR{e=6rm*{I}))I6XDzpn`vi8=su+zdrx@_xX80zv^bjG1>Vn
zx80C1HG4v+qlVy@uB~~!^SHIon7uf#Bs)snC2zv@7rIi@%5P~bTWd1+Q234g_b$D-
z!@b|OYRCRF4;&?&<9FR%&|BiPx`ff~Nrscq-g3LY>n`LcFU_6V;`_zrp7X;R<w+B_
zSvffbU$|*?!t;_D(_LrjEt~g7?W_F!PIt+XGPiW8j6hE5t#_Qy1)Weo*5?u;aLwuz
zmx$OR<7#uR=>8y=CQ&ctJsm0H-xj=H!Fi$T`i>tj_{^syU*VnP+Oco)#h(QYE38&6
zyZGFw<;EN5Ws!a7C+Xy}AF|q=8za8pd$mi?Hjds8M~(;n6y4iA^+S|~F6UGguK2Rf
zzFMbVzD6njR=33pU!%AepZvhHTxfCkZL#lF-ctj;gPcCNyuEdW?ZVR1D;MIfb1m(A
zF0er0mW+rT^DoiLOI?#LO=+`d*!X$ctB?x<wo{^#N-W-746yY7bNB_<qUJ5DtaXpp
zUfVw1<84mv-CKvEDrT>Gec3L-QG}y1T;#<<?nOtKFR*%<^lsj;>iv&}v(?R&%+@aK
zd^t5+#B@#4t!Y=5)W)t%S7H6L+F3iZWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rTZp{nhQw?nUBlMQ-e&4gNbJgG3o^hAfIp)n-!)cYmzjK1Auexe_
z$>PUPl$fVGN^5sH>2+FOuXLLFP52t?e7VMpqKjoWs!GT4>al%UB2+*5$nAW|oK4sF
zrnhfbtvGq<ob~zG?!R(Hxy~Y29KyEAYVUoWak)~e(fjD<t&hB1KCXZKufSznV4G%4
zX6GO4aEsW7VWF>sK5hE9>+XrTRh4l(?)`ymn{tiaKA+q(nfH&?4w;w@G0|Z!!%pQ?
zX=ly3Y}LA-`}>OtFMjQO$s6M)eeVCt{lAXPxA`h=6xo0LxvhME_4CIcAKt5&BO3B`
z#zy%+%ab<Mh1AY}rT*2{on18WOMF_fUB&T@)s4-;Y;|ep{!f3q!1u&H#_eH?Z!r{C
z<|fGlbLB&ZN!Ipfwp#Hp_y;fq^RV*xg9)Gv1QeyF6<abfFmU}^aIe|hG?<wuSj<Vx
z7p^T>+({gXC*dRk=P9%~dn&Dcng`S(=p^WY(4uIeXouiwS!ls|N_9Us<zFb=#>5jW
z%OPuyP^8A82Io1pzhU_KrQ{Y93yLC^U}QxhenRRmCfcwv@&xm8^76w>YEF=HTWx4u
z26U5<laL!sm?v1+Nf?PI>Ld#1?a}bmTzYfEIgm*TypAv}heGUL#>-zn4s@zCi?k9<
zm?v0;MFq~g?9Ev;!zSqo&^mr6exx9fw~&W>nkQJpLIcjbx?Sw6Md?Ovpmj<XN?tH?
zVKgHbIDUE${a&!<p|!%pA9-tkJA_IBE#PzFL$W~5LJn>uPq4a$I-K{^fb|PIzj!yu
zKcc+yFoWNBhCbvKmhT4&Yg=fC!-T(V__!&$@uDURPq2WK06cDaf<>G}ka!9f3eGSs
zJi(e4ns6RNQEE;iC@mEiB^N6LY3AU-?WZ->yymQ93JhWhf+RB#8Hive=ceRj0J+|2
zzRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*54tk>6R?21?^4LVhq-A6k
k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_LYUnA96cC)0FVvCT>t<8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.18F54267E4ABEC46.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.18F54267E4ABEC46.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8556d42bbc7f339431c0bc8088b275064e56ad03
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=nJvr=49r{%3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2U$XkqZCfP1xjwslxVrY;3mygL
z0@uVj0SbH8sHJYKELSwDu>ae9{@uEKyK_Iz)>v2t@3Ecu{Pbq-&rhG~r*~bvI`h$$
zqmSMd2i@5_|8C9CxqR{e=6rm*{I}))I6XDzpn`vi8=su+zdrx@_xX80zv^bjG1>Vn
zx80C1HG4v+qlVy@uB~~!^SHIon7uf#Bs)snC2zv@7rIi@%5P~bTWd1+Q234g_b$D-
z!@b|OYRCRF4;&?&<9FR%&|BiPx`ff~Nrscq-g3LY>n`LcFU_6V;`_zrp7X;R<w+B_
zSvffbU$|*?!t;_D(_LrjEt~g7?W_F!PIt+XGPiW8j6hE5t#_Qy1)Weo*5?u;aLwuz
zmx$OR<7#uR=>8y=CQ&ctJsm0H-xj=H!Fi$T`i>tj_{^syU*VnP+Oco)#h(QYE38&6
zyZGFw<;EN5Ws!a7C+Xy}AF|q=8za8pd$mi?Hjds8M~(;n6y4iA^+S|~F6UGguK2Rf
zzFMbVzD6njR=33pU!%AepZvhHTxfCkZL#lF-ctj;gPcCNyuEdW?ZVR1D;MIfb1m(A
zF0er0mW+rT^DoiLOI?#LO=+`d*!X$ctB?x<wo{^#N-W-746yY7bNB_<qUJ5DtaXpp
zUfVw1<84mv-CKvEDrT>Gec3L-QG}y1T;#<<?nOtKFR*%<^lsj;>iv&}v(?R&%+@aK
zd^t5+#B@#4t!Y=5)W)t%S7H6L+F3iZWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rTZp{nhQw?nUBlMQ-e&4gNbJgG3o^hAfIp)n-!)cYmzjK1Auexe_
z$>PUPl$fVGN^5sH>2+FOuXLLFP52t?e7VMpqKjoWs!GT4>al%UB2+*5$nAW|oK4sF
zrnhfbtvGq<ob~zG?!R(Hxy~Y29KyEAYVUoWak)~e(fjD<t&hB1KCXZKufSznV4G%4
zX6GO4aEsW7VWF>sK5hE9>+XrTRh5P^wrr9oww;yeRsZZ~_Tlw|qJo1(xsgkwH8;Op
zyQ0wVw#YxdTZ^5OUN28GFOYm|{(t$qSHk6=j>Miwv6Y)|USspFw*Pp0_2a^@fV0WR
z>OSOmo%|K}wRYb8Dc@yeLtHM@&pi9{ifsBjRxUrzA2Tig`-e4OQ~1qh?6N4RxH30M
z9+)Q&i1QT2&b_*UhrvI9A()4i#}7;ZWgws^HLci^iGhLZ*MfV^-loCKJi%g4V!m)~
z!QxKhNIVHA2{=!1eWz6Ex>tvRT7;a0ToGD?orIBiqE4c4o<f_mr_$P|c|a|KPJ$i?
zEsExfwg{e<xfYyfusxaEDl^86i6>Z=L)HwTNR2}c&U0*k!|?M<$t@-p6h$n-$cjSz
zgw$V5v|(lB3FhVG<%gNHN5fNd>CFx2K*9>Vjxb>u&Bz4~u--$z7p!?`t#ByB?q$6E
z_2WS2NV7;Q!Bp`CtFWlRd6&I8i)Pp)Jpp=?--#c|oAT!JaF6i>YnW@mc~`fKeYGgv
z$PILblDU#6%-n}R^49)#2$cew#OK6^WRje@99%C?u)4WAocGj#^$R<{csIyDqP+4j
zliqiRKI9da?*|HNn`?)`guiU~xGB2vq9zMZuz-^QJZ^b{MVv&CcnanUPB1Mz!J6ip
za2`WZYEB|3Efp6f7b}ApK_CJG0$~gW=HN~9J1bHXbeot0gBX%?Q*ttZa^7ja%q*;I
z>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#K
fHMMl~4GfJ<&E-s_bd6j@#oWC@nB4pvJs5oefJnu`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.192FAAC8A46AFF2C.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.192FAAC8A46AFF2C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68b38b8a820917c75e4668b374c74d22322de6fb
GIT binary patch
literal 1714
zcmWIYbaPw9#=zj5;#rZKT9U}Zz`!5~#Kk2=nJvr=49r{%3>9<EhUVtqHWRq^U-Mkg
zYn^P~AI^_%t=_0+xJ&G(M}Uf|;3AK=9zU1vuX8`$m#ltt+ZKs$uFvisuC9Ief=7Y*
zgKOg60EK|C$(wHM*)JejGvnX6^z!TG^V9yM+3m3k-eW!S`RUEvpP#;XJM+=2Pd<gO
zk_x}8YQD2Oleh1mTl@O@Ge16k{yQgLUwpEnrp5mq5uctef35%d_xbIAeoZscXOv%l
zSuSG3Df3yOjv9hzy0+G>oye{J#%#ucCHYbEE_tTwO|5bx?@6wnQFy7$YyY8tU#DGr
zqES%&rNI8w1IGyO_+581#9uC%Ys4BcF<W|4#|y*1>rC>JQ)6bf__`>)U-;ohM@q_^
zPJu1DFW%TZQIQmHe3vQSQ@%H9U*+e2vP+JXxur{GY!%Xz+S$Hm@dW=K+s-J*h}jz2
z0lNg|-rXIz%T?D^B_cDEwacR{k^fiUEC$_Q$Bw&~S2<g{UC!9VA!sdW+s5p2HPE!z
za=PM?G<l=5MK#k>URnv9+nxJ(H^bih%1OPKFL0>X^_|L>vR%U>wXjDeWRk{)H$t+%
zTcm`UOy#&1uWh=Lt-MC@u<;zP9=p8gz1OW*^+Zh)YEsX-zs#&bbMErCZ@a=YN^RO1
zS`EG}aF8`wcVcqjlqm~MJ}?~NlfAlffnt@$)*~+pZY*B-$>Y!AtEw*inW5!TUBB1l
z>rcG3^VZL;&08I&uX=siF2PYm>eD(66DM((mHZCk8W&B@9a_2XaeLmpI~}_+b$xx-
z$7BZ{)>@Y9>3i$?)GD{23-x8UJf#%#%i_MNUbygx@4l*kd&e@Di(w~SpRbs4_x(iu
zt&;<}BbRzSQObSjJU@`l<j2t>!@ITVUwlK}-uL-5sf@Sib*k?L#p)BSYcj1gEwiPB
z)x{-$@Z5?B6e;pN^OKJ=ludmN_YAql)!uH_kG#zG$i(ox@KpV$*10|2_(sb5>hpa0
z-VP@(ovVJb@9JN<qQ_1mR~*8&$!cFL-}2I?_khR8nzcKG0{^_;VJ|Z=cS@?)`kpI)
ztivs0ANqy94*Im|-?qCa;#O4}R{fSpQMhfEY%-5$Z_b112Qvx|7Cnwgjndrwa_x$*
zbGB*ziQVGfe&m|}*}MnFxAOn_-@U5-?n&qF2`8&;?DBt9-1#ln&;R|{<24Ir9q;-3
z;J(P@uZv&(wz1cIZ)vrnW5Iu;*`HThpS!~?GDqlxaq++TYuMI0eiK*LSmae)nVTdJ
z%#(c!SXUo*$+*SC;2*#c%)`p#2PS|r5KxqwR&2?{z`*ru!M$d0(_m(vU@<2#U%0kl
zaVK#ko`jPGoTt#{?5VW&X&z9Epp&2nLW`ogqAh}_Wv&J1IkvxH`1z&e786UbEQhQa
zOc6`48iyL37vd+R{$ipHD<e-ZFDEZQOc9J`<O2Jk_t5VJYaUuF?9uSlTzYfEIgn}v
zUPqX!Lm_rA<K?d(XW|K#W|3Bc3G)Q2u&BU!7j9a2m28>H3si*c93B+sT=wQHnqial
z1n5nECw?Sv%A3o>J;oEPVXgt^UEMDB)uMDGH_%N==1QJ0%OC#8Tl?D~R0?PkpA#RF
zNpj|LaJ@Xi>gMWj-ctkCFYNr{-5}qI^2)<ZdfyrPkXKl~A1JJCt{nyw{<7iYrs&3t
znk+oQ0!{+(h~fzraS}n|DVQrb!L;xMYnp4qd4C1MHr>rX@(k!gAtxbMgd$-lVI-cY
zlPH|WP?VaJ2ue%EMajj=AVv^~00ZXWr4bX)Xa+o#VhRjm2!shUFeK-u<YWN3-f6zf
zEUawo9Gu)dyaN1uo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*Z
iR8-Y9wRH3i42@0A<xHe>ja)><+`U4W-25Cp7<~Ytg3O};

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp.F56F1203139F8DB9.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp.F56F1203139F8DB9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1bd12f4cceced40be843622eb05940181635900a
GIT binary patch
literal 1776
zcmWIYbaQ*b#=zj5;#rZKT9U}Zz`!5~#Kk2=nNye<7`(U`7%JwR4b9EJZNc;Hzodj|
zw*Gb26YXZ++a6pqyTSb{aDwVhF&*!Dai2ED+iUgBR(o5pJ+ItiL4iw<iu@<Vt2_$K
z8(bIa3LlVJua>&8viO)$h4tU=b9w96*`ImwwZ_6KFh=g==cgxIe}4M&wKJ)taFGSy
z#vOm>h-}YKzaD?@;ltC*C1dOM%)Y$(X_Ao(>(ck)-n;%3e66he{;@6Y?t*8Qhkw7^
zVyxqlD($Itp5=3vYt-&(eMX1BT+{a7ksI|oa-HhhrN%*%ZM1YdjAGms;~)Gl@?TqY
z{LVLlhwp5ftCyM9eJoS<JuunCNpwM^-d6ub#n+bJpIUM|Kr>h^QCMbDd3?n64NsQz
za6MS6_aR5;iH4>y!+X!iQ%&d2+F$kcnUK(ixtBZHqBLi4cD?t#mT=1boS3WTf%Tye
zL^&citoWYTk~8;-&Vopfbu*J0|Ngm`p=`z&l`btUcjzQ<63?oMFCI2`+eAFC77lB_
zYVs<CO@3DGDU;I=IZs@2NjNquH+gZp-n)Z~rcS(I&|V`m<^Gb*8<PAC7g|JFcAc4O
zcyS~96^-{BKL%*5WL*-cy-d_<-i?h%5)N;hTD|w+sST=}0yaUn&9*XzWa^l0xxCe<
z<M7RDjTx8r9`e+`%cZQ@xIpAk!Cpm|^bLNDyqB6U2rWLGnErKY=?VKDJ%tF5)w?by
z-BYiv?z@tam3*D;+L@UlvFy9ISB80<NDlO1zqmQ*TiXl445{=qm+K$*&2~3$%8BX~
zzN~qD($x)InbAR$zm+bXxA4i+>l(br1vh1hr^Yk!t-XD-{J}?oIa;#Wt3{Ngbke?8
zKk$B&qAs~UNR4OWwh#N9r!ex>ZH)0Lzb{!Q9QxZ_c=3|h=9uWX#Vg(UEgzfusW)9-
z6_{8wt?}#vUeUCJ6Hd(O<y-5tle6>QQul-s#{zdB&o=Sq&5U*{Puy4e_${4n_M6oA
zk^i?<9QV9c`Fcfe>h|^!NqGsg2~%=amTnd_pSw)ZLfj;tTPq^Le(v|_#*=z4xSW{!
zXwtt2Tt~I9bY#UAMP0kUP5Pdvsai(zf};g4!bY>ir;GgmwT1cKHpzpHs^>nj{gar(
zf6ika&&rU~m1cVrx6XN7_jCX0{{jL(e^39&m(!Nd@OL@qtHmsTX0r4@`rUK=@kdSX
z$A!~#Y~H=^Js#PZD=GY``~Pi@_3|P1a((V!?Y8o1J^8TTrPyj!yJKzJ<Yc!x^=JRz
z+C6-ia-UJtYVY;p%G@M*U>;>{{2}kjAUBnV!9Rc@n1_`o9!vmbAfPBUtvH<tp^lLY
zEZ=+R_kuMKtrfU_Ex6b0Z5qtX6D;N=<_kA1Slme*i6`MC0p|&>@02QC_v$cEi;$C$
zD?*E~lQ0rb)JYW1Q)qMcR9gEq52!`ZNzenKMcGr?AHma6*U^RX9NXV8{QOdKi-{#z
zo<kn4g(X;BSv?u1$i1rQj`hX=l}tP+ig-{Ih4=}nznEyl%E%MU%gM_R)4M*ZLP^tT
zPY+O7(n%8T)D4q3x3fvFp9vI}a*}d^sotaEsk!v#hI1f86nGtB!iPfaUdGE`KMr(&
zEQ_oXOqeHFjYSR4yX?(bG{Yw83DEcaPW*7M@B}M*DkAw_OI-`@`>Wf<zFL%S<OW)&
z;;9l1Gxy<-ytTg_LZyHv@j3A!nWW&UfMk-Ux+dJDrv|KF*!jh~K@Jz?m4})1zBBY8
zudsYSP*_)8*9|88Wy8l!(Tx{1S$Kj4oCM%e&J!%+B!a|K@>B|eY2gXhR@XLy@feCy
za}q%*ued0=SQ*3!0uf-q9GsvO_JQrH>k+2FAcjDgFatw!Zc0uDkn5f1%gn;c#?HaX
z&BH6e&*v%V>Z~Tn<>@3OE-WG;X<=z?V{2#cpeL$rrEI1lk4;oeT1HkuQAtHrT~kX(
c-@wq=)LhO)O4rCmRLtEggvrg%(Sy+k09ER#bpQYW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.4C04B7436EBF8811.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.4C04B7436EBF8811.idx
new file mode 100755
index 0000000000000000000000000000000000000000..abec3628c9e4d15b18064042e7af8593ae41c09b
GIT binary patch
literal 1730
zcmWIYbaUIq#=zj5;#rZKT9U}Zz`!5~#Kk2=nLW%547OYh3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-k%HE)6xJ&G(N5B(LC6|d(o<EoF-{*c>f8*q&jBC8Vl|H|FxYw?{glz)j
zjztE38Uk@)lQ%`|-z^}yXa2u~>EExL%cuWIvy+*td2il{&qr^z{(SUly7AE$uY3w$
zJ^J>qIOxv)`FCr6&gF~$H|OKg<?WXH<L>yiEqU<g@Y^PRySVy!_BFd6UoT%$$@#-z
z?zaQFt7Dqia5;4@n7*b|`%6qjqM0MRTI}Ur3PsI(-10AO*mK#<Fn8%ok$s2$eVyj^
zJnqNYi-!-Vvh^R5_;u+mTkaC!sHNNnO76mwUhMt+vb?iSSUdN$%9_RMpX9ioOHb_C
za#fINR@1p_j((H23BHiKwP5bstDnx!F3(p~-DYZa*`eF(#)_wWMbn(@49ip&HEp;Q
z7`9-8z}%ZV18-DKQ4(4cAob8yW%mZ=Sa;usYrBsfZMKdH?h7>c5L5i+=l1-t+@e=2
zX1UEjEs(UCduE!Yo%g0KvK+79EGvv{xc6Sr@RpLvAz^v<r{Q1trX9*&xJ4yol19x<
zRjV>iQ!AD%8xg7ArYqUHJ!*%I=Xmw#<wf0jsT=C1>&bagcx!gBw8QjkfzH*j+Ec9N
zIX5_EY!0vBc`^Nor?!_@uv8uYk#o_jf*qZ|1Y~tye7M2g`KirE{>9oIe3w?`Z58<*
zc|T5R>*lSc+T7X?b62hNukL3JRCp2_xS(CHL+HDMxW+}3bB9*#d;B|Z-kpvcTiVo?
zg|7>^vY{t4I%slP>C$-<MLw+OTf0QCsOoiUI1}&M`;Trvc`2}0Z*RuxCyO?RocZ?s
z#PnOI<XYDRsT57z_F?0B4Q8o7Nrl#T>(amYhP=J!^J!8UZ_(ov-wSrtCtBBJT4`D?
zmX=i)m)yZ~D<V*&$n(t3bHb~(%-^8*;|oX6I*GR(x;H;$6&Nm1pZa6A%DuYN33K<p
zv%S4v_n`N*vU`>9qTYO;kS)Dv0o$fcZzEpDy3am$nIlGP-g=R6B`&!o20sp8tPhIW
z7S^U2lhygBTKuI~{mkglwO%Jr|9YLKTfOpGrJ;<h5?{*gv>vr-mFm(zN((G^B<$R_
z=F-|JH^2B@u~^Q^T`#_czoNj~<e&6Pz8JUP=j+d8*FFlJSH(MX&GE;@ZpR+m9jVx}
z-&`)8^=smp8}@(9lQ-F|*k{k-zpAE>PfO|T{xi>NbokfSb>5#|@nh!9|MKP?+f;wE
z@pJF2EUwH=k_YC^pu8pJUndLo@-X-ZFa-0k@`Qp3pbP{QrKS}-F)=W3{aSFZ+1oUj
zi6>ahNz50nEm+)19L`f{bM{nP`!tW4Cs@!)&;y}J(N56?!PBzSg7b7N|7<>T(AAHT
zCs@f&DFCL(vHcCh&o3pnm{@{kIb>~N!YsjR9BObLOq7ue?B?D>zZa}|Xsr<9C#3#j
zq75t1NM24}eweDX-FAOEjIy2peSmBv4~mf)$EHdf*N46a>XmSkaEIxw{ySyT);!Pc
zKw)h=?O2#_bApuHYD42Ppnrv&gxp}lJi)?F!bm((Cs8=BoBhDTjzdKaEIh#iP6BWX
zc!EWoM38t2b_&i2<23Cw;k-Q>o|;Q<Za4?>odT~TOwpkbyO;6u*N+2zC(R<Q1QX^7
zR$)<r^DcXH7R|6pdIEGWzY{-F{K(tM!()UeSi?>O&U^SHZ|!e~P${4mJ|{jTEpm2p
za78@9>UQdI-ctkCFYNr{-5?i=^2)=UQk0sL2uf4MMajj=AVv^~fPg?4gMm5t`FV%V
z?B}7om;!?sl5<mXGJta4X}-)XtZeKYoZLLT0{ncQg09YLf?S?XLgK<A5|S2{);6|w
z_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=Lbo31jjZMwvOr&&;Ttvm(y+WAW{2V<PeE@(u
B(E9)Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.D8BA94FD49F7A399.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.D8BA94FD49F7A399.idx
new file mode 100755
index 0000000000000000000000000000000000000000..54db1ed90414c6586060ac87d6a9e4ed4b98bbe8
GIT binary patch
literal 1730
zcmWIYbaUIq#=zj5;#rZKT9U}Zz`!5~#Kk2=ncd6`43=CB3>9<EhUVtqHWRq^U-Mkg
z^fTGKKb#-k%HE)6xJ&G(N5B(LC6|d(o<EoF-{*c>f8*q&jBC8Vl|H|FxYw?{glz)j
zjztE38Uk@)lQ%`|-z^}yXa2u~>EExL%cuWIvy+*td2il{&qr^z{(SUly7AE$uY3w$
zJ^J>qIOxv)`FCr6&gF~$H|OKg<?WXH<L>yiEqU<g@Y^PRySVy!_BFd6UoT%$$@#-z
z?zaQFt7Dqia5;4@n7*b|`%6qjqM0MRTI}Ur3PsI(-10AO*mK#<Fn8%ok$s2$eVyj^
zJnqNYi-!-Vvh^R5_;u+mTkaC!sHNNnO76mwUhMt+vb?iSSUdN$%9_RMpX9ioOHb_C
za#fINR@1p_j((H23BHiKwP5bstDnx!F3(p~-DYZa*`eF(#)_wWMbn(@49ip&HEp;Q
z7`9-8z}%ZV18-DKQ4(4cAob8yW%mZ=Sa;usYrBsfZMKdH?h7>c5L5i+=l1-t+@e=2
zX1UEjEs(UCduE!Yo%g0KvK+79EGvv{xc6Sr@RpLvAz^v<r{Q1trX9*&xJ4yol19x<
zRjV>iQ!AD%8xg7ArYqUHJ!*%I=Xmw#<wf0jsT=C1>&bagcx!gBw8QjkfzH*j+Ec9N
zIX5_EY!0vBc`^Nor?!_@uv8uYk#o_jf*qZ|1Y~tye7M2g`KirE{>9oIe3w?`Z58<*
zc|T5R>*lSc+T7X?b62hNukL3JRCp2_xS(CHL+HDMxW+}3bB9*#d;B|Z-kpvcTiVo?
zg|7>^vY{t4I%slP>C$-<MLw+OTf0QCsOoiUI1}&M`;Trvc`2}0Z*RuxCyO?RocZ?s
z#PnOI<XYDRsT57z_F?0B4Q8o7Nrl#T>(amYhP=J!^J!8UZ_(ov-wSrtCtBBJT4`D?
zmX=i)m)yZ~D<V*&$n(t3bHb~(%-^8*;|oX6I*GR(x;H;$6&Nm1pZa6A%DuYN33K<p
zv%S4v_n`N*vU`>9qTYO;kS)Dv0o$fcZzEpDy3am$nIlGP-g=R6B`&!o20sp8tPhIW
z7S^U2lhygBTKuI~{mkglwO%Jr|9YLKTfOqx(Si~ViwU{24W%NU&CPjm{ovCFO;2xa
z^4cDhcG)`c@tkkke`3$r-8-ne@rQnjRe@x#{eQ#dUps><ADN_VlIuVJQlkGi$K#Lx
z?p4eYz0zu!{QtxHo|C^8{Q4(iqxs3wYDLGU|3<Sv&$KD|rCJuh=wW)=Kl>Q@sHtzH
zP0w!s$WUCFn<Nj+o9h&ee&wADi{WAL4`2x9VdV({6F?aVC`wH$c4A^+VEwEoUmAJD
zn~5h_mP6JCt}R%NLk-U3`nBL*v$tt5P?4CEm@h(+xRW@XC%C>-s&w6}!^}LvLQX=i
z2t~q9!bm((Cs8<0q0QM-Y3<WIpcX+VK@Wr$MO#H@1W(IW3(nKA{ImJUL03OUo?s<g
zC4ZPA$M!c2KfjdRVq!t@1`CQeLi~i(Ure-N1?uJH<mHFy%{Vqy+PFUSHBeZ>Nx~f_
zT>W>-q^)_L+kwK`w%Rc;;coT=3p)-KHL&mm3pfeDRr3UkIEf(f6l@ioU|M*BHElKF
zJeVjW7dS?G5B*-S=ApI19t}^;r8hU6136BC*Ab@bP>9{jc=_wcfk7|LBCP}y<_T6|
zQGxR=dvg}eut|CX3`c$^exz`ex0Q#798a)@tp=R;@JHU--wvTtKrMVud`Md4Y~|pJ
zc!Je!)#1FS2CQG$`Ng|IE)?aJhdHGvH7602rizP_i<Lo)AP@lt%)w@R5Am~{G}mPc
z3}Og`2{SMx=ceRj0J+|2zRWDFZ0sDI+&sJj{Cu8*uFh(LT%Jxs;=&>lk`|WMHnw*5
z4tk>6R?21?^4LVhq-A6k6qQs|)it$r^bHJ+P0i&@q;!p3M8({_LYUnA96cC)05mq#
AfdBvi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C14D5D3E331E58A4.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C14D5D3E331E58A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21ad6329a1804c7cef5c0927aa1c9b2a3e7e3e8b
GIT binary patch
literal 1730
zcmWIYbaUIq#=zj5;#rZKT9U}Zz`!5~#Kk2=ncd6`43=CB3>9<EhUVtqHWRq^U-O*L
zE1hiKAI^_%t=_0+xJ&G(N5B(LC6|d(o<EoFuX8`Gzj1O>#x>sGN}u07+-p}}!Zv|%
z$07qi4S~3@$D1Pd?-!7)nepe`^6%!`=P&zHR#sWNtfE}$yjZ$;TKB`-#z((io%!g>
z(LevZgYMMNxL5bLmoNVRj88|Ow@dDiJac^dk_Uet*Blg|pSSOfoZbD8v#l@fVX2wp
zQ<ZQ|>$<|~rU0cEs@ZqBY{kQ7_+DuAyuL-R<Br1mi_s=J@3#c{L}yyHuHR6<cj?8Q
z;&sJeJ{X*N5Gdgszw7RT-V&$PC5&!QGMt3=mfQW+yO^J}G<Rl8=@*xK$q#FkCr#XD
z<>U~2;ilCI&r4=Zcb%oTY~CBSuk!Oh*(FD|xur{G1aeAmz2kT;=!E*QK9>-IFsoBs
zB4Ue-t<AZj`-5DXM7@;vbfk!vEqJ|x^Fr129Tf%b*{2L&v7J<MxqtFS%?;-e+0e@`
ziY1+L$`mi_98dS0wv9P)?p>30+!ub^D!E0sxK%uQyy8#Wy@Np&x>KgLcnP)Y?{+z^
zx6rNKp|{;>ao~iiwXGLZD%yfOFCO*nw%)t+)CyHi#fn8`xmnCFE?H$|tj}%@I$q2X
zz>)intDo`jB%4fE&rGkw?;CGadY5WtbjW$>8dg=_T)ePo-jDXz+8unCR^@FK`5t*c
zPHF4rt)<%B+7EMAt@E$$XAM+%5*xUnU9UsvyMwsKMU!)fR_%NIJ8#~djvHIr)Ru*>
z3%IhOCo?)|a#`upc@srGoabA+M6jspb!s>h@7nv1Za;Y`uvc$y#_A`FHiw+~_Wi{4
zTc_k&*955)P2Bcj<9Q8csXs}E^K<qG|5963makUnxtncg;U=|=`L<7-u9?i8GUsA%
z-{Wqc7`EIs8eBV7p2a-p487GKDPB{>vTnV^+Ya5EPp>{ma`4ysXzsV?@2rDX)px76
z)kil?*Ll0=^PR0X-YZ@=b8+NNNxi-C((XR9xyuD5wC1fB30LBhTVn9z@WuL|m~CNg
znlV|Of2ze_dezU24qfYY^7OCUX}Z-bpB?@1R`G$uw%lVG2{X$#HO8|YKFE4{YwFaz
zrDxR27afy)8~$VWjM_cT(MccUPn0|`x>fft+3&UR@=r%HPo&sN&An_W^;OVq@!!1{
zHMCc7D<Av!FkW)<SI4h^H7kOi6qf`Drv6Vgdv09$;*0m2eJ+R3pZQb2!)}Y#&3IMO
z?JpUMD|3_NfqC=L-Cx1#=LJN082kemf_YeZLcj!21_Fvw(~6y#7#O&IEx6b0Z5qtP
z6D;N=<_p&rEbb%@=P9%~dn&Dcn#ar&Ea)WYfl#DqtLTj2Y1wMQc{-MVHXk|Y>c_|v
ztYoX?4^!mW{)XY_my%mdEWxrIvNkYbmS8mwH8?NCPe}d6L>pG1US3XKewd<+V^gJ#
z>qB1yg(aLM++o7if2T~^n&-J4D6DO(9Rm~YW<Rj7<4{op3s10slK@;bPq2uS2og`h
zR>29Tg(q0kRuj%!^2L<>FaP$Bz>q<9Di4ZN_h@)(F1@+o9LT8(ypAxF4u#mgjF-QD
z9H>{CMOq0a%oD7_q5|i^L>aljQPX?q_kuMKtraeNa~938NqPbdY<?$xq`;Q9m4^o}
zPq2oq2AucsN8Z}s4xv&&EqqRVNLu7<<=~2Vg4J!+;k>5?tY6sq#k)Zc7v+_Qx#h1w
z*rvPrN1g#q5^@r9g$eTn3p)uT@kE_O;k=^MoJ3HXDlSScRtD0{!3Hia-`7XT^D_ko
zF$6)98;A@<Fq3mrax#Ei?=)X#7FITP4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfj
zJ9`H`QEe+_GYxrcqGHlAvI>eyDyr(5S~~g$hQ_Amawbx`MlPaa?p`5GZhnp)j6MKT
Cyx9i;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.066E2D36BE87D012.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.066E2D36BE87D012.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15bde06a442a46d6da38e606baad7d30715f2e65
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%noJ-274|BhKf07Lv!<Qn+aU|uX%3E
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0TU*~?hPgwov1|8{duFvis-d+8!_Z))}
zL%7H>PS!)Gv!cXq^E*#Gc>HF5O<i^A-^z3A#TS2^`RK23<wmRB8zN%=lq{2opQkx*
zSH!(K->%#JpI_(izx43-`{I4Sinc!pU;VVih3lNXZg|btpRZm&dUu+4|1QO|-UpW7
zT*E$XO%d-VEe_SDxXsscFBB)X-ei@XQR4f$=>+d5slBd;Klw=?-7<Aq@DHxv*Q2H8
zRzECP=PwtRyu+RR!h5^)mWh5DQ}qsX$oP6J`hITtebZmDYqsc4-q1U5R(ZU|{f?9+
zYx@K)h?H#OE1I;W*WvpGwzoO!r@g;dTfyIX<K7D&W?iks#naeNL~{K+absc!*C99e
z$i_oXCC?tXCB6>q^4Q>%*&;si%^}7TSy9HQ_fq2WcOFi&aJ#Dz+490#axTAn$EL+O
zQnh-{9cL9&W<CAUlagsIaP4O7<6R7U?+Yhwam{F1UN>jT{fpKS-0NE}O$zW_^dWWn
ztwMz>F9cS62*@zu3R=6>U>48$0x!c|JF-76*gi#as)}Hfd)D=3Mh$vv7dL&|6{b;Y
z)y~j*;n|805*OB=oE+xjCCppTlJs2n>B<EQzX@!;a`EB{cju=zU*(TY7m!^NdN*71
z-R50;uUt8I?OLu*PRn-9-SfT)axZc?v3rpNf2@G2T%(@AMU!)fR_%LqJj-sU;Koc|
zU!V1N7OY5=y}WISlI$lhQ)SHu^0HB$%1^%BJFUnn6}A8P_a`p}_Ui4;Sp8(t<|$|X
zeZQi=b-FEA<Wi3(O1Tef&ucJC{Ylz7=k5ODU(3Q|@A-V1RK{EMIK}tEMB67$*G%TR
zoO|JR{G|c&bVq6JE+@T}#p^z|XuVy&hTG1T>9Mz)b&|K)E*T#7FH3^{FI$<vKV`$&
zxbO31?@xQ6>TUhrrd;=1^@**#fdPz>H_O(1S-05tc_xdV$a(Xw{DA49e{N5_ub_7O
z!W@IvY~6|fKZyBnugDDDI&sgM{qMJJnw}NcbF5~IY-^x(4Tt4~+}Vaw;m_vgJa~Ta
zXu^r2+{mTTnwxL-t|)Z8E&9*uL$mj*`}>cTJ}|mf_wU#|XSLTSpX||}>?xmSWO;A4
z&!(m>!^Ai8|K=M;+OMdy;p;c5?c>vWb*KL1vl<Ke{%72OE3Wl!SNXR;MmcKc8~??V
zc04Ms%uSL9=F2tPy?br6KHKsz_y;fqbFp$of(f7u1QeyF756eSFmU}^aIe|hG?<wu
zSjbw)4X!O%*jgBgCt@uE=Lv>h$|zXEvw?{xn9rKe5urueO4<R&6I|aZRl4rgVW7GE
z*8FxbVV+<CYXKylptT^J=h*&+;pdl<TZ}Bh5*!kiFfA^Vzq|-6621qtj>nn@$vR0Z
zNw{?(enRRmCfcwv@&t2pa`VIV)-V3w8~!my9w>}#<fJF;D|o+e4FL+vSjj}gR4<sY
z`U(3^$)_Mg6u7Nm!Z4bV3mk;Khkh?u^Uzx1P>9{jc=_wcfvUw=#FSvFE_-tp&9F&&
z0(35~H7}BLrL3gj&V6dY`h}feyc=YbD7QRJFGEpkP9i9Nii?tql|hUk5CI0v!CmhX
zg>MHvQDX`WVhDr@GcY9QrsQM*x!!5M%q*;I>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq
z7M9jFws!UodZOA^%4Qn!*hIyoWn>i;l~h#KHMMl~4GfJ<&E-s_bd6j@#oWC@nB4pv
HJs5oeZZ?h=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.341F4486D5AB6E5A.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.341F4486D5AB6E5A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e00d96c3b5eebcf19aed06ecd2d50ebd9be8ff7e
GIT binary patch
literal 1540
zcmWIYbaVT|%D~{9;#rZKT9U}Zz`!5~#Kk2=nT^a04DwtI3>9<EhThJ*Z6R>=zew`2
z*;j5$eQ<tsYwbqgq>|_-6BnGA=+dDSI_cBW`1=b}k9ST^+ITbJqWDbX_F`M(61EA9
zI~E!EX$ZuHP2O~4kG-&D&5VCX)6J)c&rko8WhXON<DTuL=cgxge}4K?UcKw$(auL3
zyN{l|zGTNwgWbQrP2{`(Z{EkJ({EeukI_?e4l4MUx$(*A^V8#>Z;xO1^Qmd(JZ3xJ
z<+d9Vre;qFb<`02(zP|OcOJL)8M7A$mSjhXyW~x{{z6x3TKO%FWou359umK?|IVcs
zcewZ4R_(Ao|G-beIeypO1-&Ist4kQ&o@6))?Jc+ayY51M^3vRyExun|?!`W=QJyq$
zo0XG8@P(UJCp<5iG2L~R-m-ab)V|8k|74dODRWbo$_V6?-g?LRT+j*jV|^|m0@tig
zafyg6GOjk~itZ0`X%h8P-qVpH{%yhQ6`U8UuJ5RL&u2a*`3mnO*N%OYFa9iOSYfql
z*~R5XEjQjcFN^H+pQMw^e#mNfZjAVX@6|3n+c<hZ9628NlX-9R)DKY_x|~y0xZ=w?
z`)Zwf`5LA8Tiq5be2wB>eDVX!a-qfDx5d6!c~1@W4s!b7@b=afwhK#3uUv?`&b74f
zxxfN}TQVYY%)dk{FLh12G^Nd+VdLj%uR<;e*iMN`DzSKTF~HLQ&*2wLi<-BrverFX
zdujV~kGDCwcW)hvs+hg%^<}#RM-h(7aFG`axfdN_zQF2b(z|)Xs`oz@&hF21$%%3n
zzN~#+BP-(Q7TuMezjp<l_u%@m*I7HWWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rT?w;4iryAJyN9Zlv{k~^k=c>Q6J>xE|bIhBwhSMsAf9C{KUv<^=
zlEsgo=rB)rl-BNY((AOmUg<RToA5Q(`ErdHSr^N0RF#h7Imh;8iBSFIBe&;E=4_h2
zH@$tkYQ@P*=d91icK?+t%5@gG;t;k?R(tR3jLVf$jm1YlZ+(<~B~|o)>67q|lh;ht
zeI9ywKiR+M=w;owsL-{l`mg10_r6&=OZ@P~+7*pCw|g$mFnXWDoX`He*?3y^X|20n
zGkw>)_7>k-_2E3@`&gm;_46JdNZs=G&s3W>pKFglX8u3^_}<@&njPkH>8zsmX>aWR
z<r_xcUtQ<>yP0eLqUNp{zw9?#${uQ0{=viVA5n4kNd4;4g;!71F}dDY_l}{sGB-&c
zm?PI@dSyg9)-&=j_y;fqbFp#-f(f7u1QeyF756eSFmU}^aIe|hG?<wuSjbw)4X!O%
z*jgBgCt@uE=Lv>h$|zXEvw?{xn9rKe5urueO4<R&6I|aZRl4rgVW7GE*8FxbVV+<C
zYXKylptT^J=h*&+;pdl<TZ}Bh5*!kiFfA^Vzq|-6621qtj>nn@$vR0ZNw{?(enRRm
zCfcwv@&t2pa`VIV)-V3w8~!my9w>}#WNYOAcd_Zw(LiAtD;YnS>ID;4KVjb~`4nV`
z0=E@R7)CR4frGI3(C-Cn9$G6L3bA_`FMs_wP_-C~m=a9YWpB=+88%5zfX?N$=0$R@
zl$8|RxlavPzp(R*cY|ya<(7x(WhhF`Nd(1DaZz%yGKdibA|N0T#$aF$Zn}NM<;FFe
zc&5N0hUDCooD86xcbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQuh=intrL~Q%oxOve
ysJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdrm8H$O)YMjrsjXLN)B

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5A4CE3D80E2AF506.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.5A4CE3D80E2AF506.idx
new file mode 100755
index 0000000000000000000000000000000000000000..082cd833d7115efafe88b7f9bd8b4fbd3ed61a7b
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%ywo525T+`hKf07Lv!<Qn+aU|uX%3E
zbRBQrAI^_%t=_0+xJ&G(N5B(LC6|d(o<EoFuX8`$C#-(-f{FAu*JpPR@2-B=dyc_~
zAzb7bC+nfpSwUj6zdKJn*#1(!rmVX3Z|1S}&lh(KAAKrZxzA>&LHxYGMT>S=*azB_
zZrt&AlE`-XwCnNp%MV|-Pp|p;Gw<B`Y27C~MCR1T==}V2`Reqacb`kwl{(Lw-q3$J
zlV2y|sZ>g+fEUZ&q_xqDpC97d%xh)va@IB0iPBF@tGe5s&Nb`G)b?HeLG;`D?Iu>=
z59RmEy*p>LL;To<>3QZE%5yJh#Wo09&YIxzey;z1?Ju_@GNaWJg>6mW?<?3Rc+w+U
zR&jymiyYY}Dw)!a?-ugjx*4Z?e{Xd?XXlM`FI<>)wGtOkt3DCQ_4CAyi5*;rqTC}J
z4>^^jy?0A|9oXfu!6~yveBzr!awW2&j8X5U#O3ciY_o8?s}R}p!dh}JKYPcf#W_;7
zTFxD3?NVkv{n3+>X)SQ=X71x%414b@Cv9=fXjxt-Gxh!j>j=*Et(PVRcrN;oI{ns0
z^(!v~R(uG^FyRVXyVYP8kAH!e;jSIIh3mIZk({a`*yNsdeVI{%-rB`Y-*$y*l-jg2
zv|f0&;Df}4^(Q8Wxp)cl)-xnM*L}Kj!NP9>Td!QaxWe7}sm)h;y>p6I9;<g<4!V<G
zTD@}R+_h_?qa(QUgUf8+GK;x1PAGS6l-s4?Wy27wuqbm*YiR9J{;M@b%1M`{X3dD(
z=@xicYgw+R>#gTgv)qCX?6=x7$?e3;xzlU8Ot#duzdvzVu`0IeV%SO7bj_J}-*1f1
zQm+=-<Tdex>#ak-=LNEv{5V=<c(*qFi|^W7`)58;eJl0k*vVN7n5$2;uF16OoO7|a
zZ*n4=Ub9)4NNY^c^O)zHp||`a#A~Wp^v+4X?a;mXqQ*ex!ep<%vzP4qJMExa_1)^S
zis+{4I&b%UzO(hld&TQ!E{?1zskb*?+TCY1ce$WNuuZ)7KDXGgkL}9)9euVfvrOa)
zk5c;oPIq43$4jfSbSoq4Z|9}xUyX_XbTLiD?d=D}M-JO^k7Y#7EZ@`^&z|0_czSE<
z)V!r<mYFX)W~m$g<M|GLy_L29#jZA<&iMZGwiVyZHOC)cUe<HGcFM;e1^G7TxU`PW
z*f{@>aq_0RfZBR7In`e>vLRQt|DSaBWs2Rr)8e;2u93|5{PTZ@^A`P^_ov10Fk&dK
z%uSL9=F6WeN`Ezb8eZXH@DE@J=3?cF024qN2q;QTEAC}vVBq?-;9j%0X)rTSu#mNo
z8(dqku(dD}PsCaT&Jzs1lu@vRX9E*YFrPJ_BSMR`rL;Ya=h*&+;pdl<TZ}Bh5*!j1
zFkzR;UtR<j3Eu<i<+0{L(kp2x3D+CqC#3#jq75q}PcSzpH$O~os;}eT#*NF8fKKJN
z=10;iU@d^e6SNkD^CmrEU%~r*YY5OJ8B3Wcm`MvJtbW42Q}QXu0tId>nDC(xyO;6u
z*N+2*#aP6YV8WNZIg4i4Bs~GTnb(>ZZZ=P_l%*8h%})(jzp(R*cY}-+<(7x({jfB1
zZtpUsQlKz$Fu<f4xxlf~d+7IqH4m*77>ZJJ5<&4(T$Eg_3}OU<2nYy-F&LPGdnfGS
z^PZRx%oG^Jker*6lL3_TPV;4EVP#|I;N<4v72xOd6m)e~6Xf!A5)u~{k&v{ow6?Le
zvv<%F)wWVL(~!p|Dkd!>tDvZ)qN=W`rK4|PXl!aOXCkF*<RU8O?iIr1=I7|a=mP*)
ChnQIa

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.403AE32B0F021E71.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.403AE32B0F021E71.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9cc1dfb06b718971750480fa8ec3a774d9afa040
GIT binary patch
literal 1538
zcmWIYbaVT~%D~{9;#rZKT9U}Zz`!5~#Kk2=nf1&J4ANW-3>9<EhThJ*Z6R>=zew^i
z-&wxAKNJh!uHERH^rgv0NW^jyN8plaOKM(TFVDDH_-IK)j+994M9X4}>&DLM43Z7n
zT#s9r9;TY=b$|UWJn>-rN&A|r;>xd?r`D$({;YiTsB+~ti(LlMbN1N=&6zK&DYtvZ
z{Tkn{+5DGR=ht6;cze2Le9gYOA5R~xu~A`tF8`c+-;a`?rDeZ=&Wewl{LJ#f<CIIx
z#bG<yB1Kz-9Q7k#hrQUj;ZRN!pU<vkrH)V7DtqrON~~Pon{><TvSv-=-_yFi=k|V>
z-Q0frGS9mf!!Jv3^WO4M-{K{n;L@k2a<TgP<@Z8=)xvH~^Neu!e{xs!oVk)@=4xdY
z6V|zFo9C$HDqpbO;%I$)_0!qr@AVZ_x0za9a_IKDvEnIT(KIJJgEEyxO&cx+hAr42
zF!$!pz#COll!Vp<NIi5_*}Z`|*4?+^+U{dVtIKyRlUbZMLC5LK+@5powk}r#O?&O8
zDIQ4^H#%EXGbtt0THxAE-^aTc_TFbs+TxngqHf1`>ir8@o!0A4nVu_D0)K4sE_oxA
zRl;$lLc>Igb>;Oa3Ewt%i>VT^ced=f9KEV1YLZYBd)D=3W(|647q@-e6{b;Y(az9%
zA#ME!i3|ECCx=a$vdrX<T*ut)A<LWEUM##KwD@phf7_|bNAl;+D_D7~-gPzTj(Tl%
z-<5OMuC3Ks)3SZ#F1xq3VlIsn%3T}fb}4w>VUAtka@m4=^{>A7yMMlLJF?8w#Bgs>
z>!P+)pWBmWZCS1TwP#7g|FR6#-U*kZ^>+(pWc(Goujb$WqV%O}w5s^rrAgcNv+uhS
z_Dv&Ys?x;nEv@CXZju)sh<E=?`91fNdEnM+<CEUEOi%PqHg){{Mp1mzvX@Jq^yHPu
z%ujh>l#%QjctX|ixt0HlOLb><J-Dn~P<JfDJ?CUwdh=vP+m%P|TYUT$&i49E>ie7j
zayjERZTWnz_<GX*_L$oOtfmvD<{tNYZ*{9=p4)?vk2OnwaFs?L{V)4dG-;FQS<c`c
zSA>50=YK5eo*oveee(3L-)XwlE1&&(WOvD7Tkf$8ubJhW8q1q6ALKl}HFav<(lct;
z7afz-4gK-H;=5kww(EX{4~%Y={<Gq9R(pN%#g3ng1K-!0+`2E_yQ#@b-tb#@eU#6&
z@2ln)|66ci`HKZg^Xk&gjtl=N_&0CiLr%{1Gk?tAqP<1@CVyz~Iy;8q%G@M*V19gM
zdW2zBqRK)Z2LAwtU@lg!05AcRfq<gawBlYy1_rKQ3+^?0n+7xU1PfUUxxuvs3tI~#
z@kFde;5@<5OBn@Acs4Nc1oK(*IU=-3TT0u*c#iFF7=C^!xy8s5EWsgR0TXta{N+Vp
zk?=jBULI>6B)yWBl5o8tenRRmCfcwv@&t2pa`VIVrusVWZQQsl3FuUQYknlX0@eaZ
zJV9$gIIlJG|GU`q=xCrxGL|yFFq0NcSp9^3r{q(R1q$3&FyTWXb}!@QuO9~ri?N6)
z!Gte+a~938NqPcwGp{u-+-#m;DN8B1o1Yr6eqrYq?*<tu$}JDm3!@piz`@^p==Xv(
z53LnGEX|zTyG*GRs2VwR7>ZJJ5<&4(T$Eg_45XQZ^Y^70ES&l?jwvvRAqWz6ATkia
zOwLWo$pCV_(|nm(SlQS)IJtRv1^D?q1znxh1i3t&gv5nKBqS{?t!-@W>>cz(wXKxR
uG~}^~ib>1JDkv(csH$sf>F6668k?HSnMmmxxrmCndxbE$`8j$p`Tzie^o98V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.9C8F172FA791F403.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.9C8F172FA791F403.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0fe87938410225c1992401f716cad528d81a7d82
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%noJ-274|BhKf07Lv!<Qn+aU|uX%3E
z^fTGKKb#-kTD?)taF^ImkANqhN-h(nJby0TU*~?hPgwov1|8{duFvis-d+8!_Z))}
zL%7H>PS!)Gv!cXq^E*#Gc>HF5O<i^A-^z3A#TS2^`RK23<wmRB8zN%=lq{2opQkx*
zSH!(K->%#JpI_(izx43-`{I4Sinc!pU;VVih3lNXZg|btpRZm&dUu+4|1QO|-UpW7
zT*E$XO%d-VEe_SDxXsscFBB)X-ei@XQR4f$=>+d5slBd;Klw=?-7<Aq@DHxv*Q2H8
zRzECP=PwtRyu+RR!h5^)mWh5DQ}qsX$oP6J`hITtebZmDYqsc4-q1U5R(ZU|{f?9+
zYx@K)h?H#OE1I;W*WvpGwzoO!r@g;dTfyIX<K7D&W?iks#naeNL~{K+absc!*C99e
z$i_oXCC?tXCB6>q^4Q>%*&;si%^}7TSy9HQ_fq2WcOFi&aJ#Dz+490#axTAn$EL+O
zQnh-{9cL9&W<CAUlagsIaP4O7<6R7U?+Yhwam{F1UN>jT{fpKS-0NE}O$zW_^dWWn
ztwMz>F9cS62*@zu3R=6>U>48$0x!c|JF-76*gi#as)}Hfd)D=3Mh$vv7dL&|6{b;Y
z)y~j*;n|805*OB=oE+xjCCppTlJs2n>B<EQzX@!;a`EB{cju=zU*(TY7m!^NdN*71
z-R50;uUt8I?OLu*PRn-9-SfT)axZc?v3rpNf2@G2T%(@AMU!)fR_%LqJj-sU;Koc|
zU!V1N7OY5=y}WISlI$lhQ)SHu^0HB$%1^%BJFUnn6}A8P_a`p}_Ui4;Sp8(t<|$|X
zeZQi=b-FEA<Wi3(O1Tef&ucJC{Ylz7=k5ODU(3Q|@A-V1RK{EMIK}tEMB67$*G%TR
zoO|JR{G|c&bVq6JE+@T}#p^z|XuVy&hTG1T>9Mz)b&|K)E*T#7FH3^{FI$<vKV`$&
zxbO31?@xQ6>TUhrrd;=1^@**#fdPz>H_O(1S-05tc_xdV$a(Xw{DA49e{N5_ub_7O
z!W@IvY~6|fKZyBn-w}Fs*_1tN_P^h@X?j*%&#{^<vaNyEH5`@`a%US#g+H5{^Wgcx
zqX{R9awC^UYi_>TyQ0wXw&*{r56#}M?(aWV`oQQ`-M?e=oYh{Re6mM>vZs8Sk>$PF
zKAW1l3=`kT|C?_ZX}_Y*hOghKwvSKi)t&m2&uT2>`=4?Dt+>{^UFF~Y80Dy$Z~T{7
z?RZpNnVTdJ%$Lg!_AXoVaf${HgMR=+Fc&LVB$xooKtNGyT5&HU0|VEu1^1f0O@o<v
zf`zPw+~C@Rg{_5=cp}y!aGv1$PN~v$uMPvX@LTiSA+!is3n1|Xtp(vcg*InTrL|A<
zn0SJDta%&}S|qI`9bi1i_BRYazm(i!WC@nwkg$XahxiGpznEyl%E%MU&B@IV6RuzU
zzc>72j66^nS@onR>??S`Zw&zo%UH=o!&EPru=)x6PRXYrLln5JV8Sq(kqaC;y@!4;
zSo6?Y;ZTU(%Xs<g$APNFSj3cIsxEtT7R|6pdIIzXuQe}{FQlxb;J$eHBX8|*hfpb?
zB0g(ABt_Cz(r`sj4OqXh^NV+boFmFD53_)wC^aV$6hFm9$;HYbMi7Vq1Lok}GZYR#
z%`Yfp3JhWhgb6b+B<H5&WB|F|X}-)XtZeKYoZLLT0{ncQg09YLf?S?XLgK<A5|S2{
z);6|w_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=Lbo31jjZMwvOr&&;Ttvm(y+WAW{2V<P
FeE^Pkrjh^v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.2C419A4DCD2F2168.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.2C419A4DCD2F2168.idx
new file mode 100755
index 0000000000000000000000000000000000000000..148099c580cdc7b61032efccef0028489877aa17
GIT binary patch
literal 1540
zcmWIYbaVT|%D~{9;#rZKT9U}Zz`!5~#Kk2=nT^a04DwtI3>9<EhThJ*Z6R>=zew`2
z*;j5$eQ<tsYwbqgq>|_-6BnGA=+dDSI_cBW`1=b}k9ST^+ITbJqWDbX_F`M(61EA9
zI~E!EX$ZuHP2O~4kG-&D&5VCX)6J)c&rko8WhXON<DTuL=cgxge}4K?UcKw$(auL3
zyN{l|zGTNwgWbQrP2{`(Z{EkJ({EeukI_?e4l4MUx$(*A^V8#>Z;xO1^Qmd(JZ3xJ
z<+d9Vre;qFb<`02(zP|OcOJL)8M7A$mSjhXyW~x{{z6x3TKO%FWou359umK?|IVcs
zcewZ4R_(Ao|G-beIeypO1-&Ist4kQ&o@6))?Jc+ayY51M^3vRyExun|?!`W=QJyq$
zo0XG8@P(UJCp<5iG2L~R-m-ab)V|8k|74dODRWbo$_V6?-g?LRT+j*jV|^|m0@tig
zafyg6GOjk~itZ0`X%h8P-qVpH{%yhQ6`U8UuJ5RL&u2a*`3mnO*N%OYFa9iOSYfql
z*~R5XEjQjcFN^H+pQMw^e#mNfZjAVX@6|3n+c<hZ9628NlX-9R)DKY_x|~y0xZ=w?
z`)Zwf`5LA8Tiq5be2wB>eDVX!a-qfDx5d6!c~1@W4s!b7@b=afwhK#3uUv?`&b74f
zxxfN}TQVYY%)dk{FLh12G^Nd+VdLj%uR<;e*iMN`DzSKTF~HLQ&*2wLi<-BrverFX
zdujV~kGDCwcW)hvs+hg%^<}#RM-h(7aFG`axfdN_zQF2b(z|)Xs`oz@&hF21$%%3n
zzN~#+BP-(Q7TuMezjp<l_u%@m*I7HWWzN0Io6{YfUzh*TtF&@i@2!7JYv&6c(apbY
zKP@fW<gU0{q%&rT?w;4iryAJyN9Zlv{k~^k=c>Q6J>xE|bIhBwhSMsAf9C{KUv<^=
zlEsgo=rB)rl-BNY((AOmUg<RToA5Q(`ErdHSr^N0RF#h7Imh;8iBSFIBe&;E=4_h2
zH@$tkYQ@P*=d91icK?+t%5@gG;t;k?R(tR3jLVf$jm1YlZ+(<~B~|o)>67q|lh;ht
zeI9ywKiR+M=w;o!(5-r^`mg10_r6&=OZ@P~+7*pCw|g$mFnXWDoX`He*?3y^X|20n
zGkw>)_7>k-_2E3@`&gm;_46JdNZs=G&s3W>pKFglX8u3^_}<@&njPkH>8zsmX>aWR
z<r_xcUtQ<>yP0eLqUNp{zw9?#${uQ0{=viVA5n4kNd4;4g;!71F}XFYd&f{*nVTdJ
z%#qK{|Gzq4UH*oL!9Rc@n2VJw5KI7NAfPBUt+<zwfr0DSf_u&0roqfS!9vzTZg6eE
z!q&n_JP~UVI8ShWr&Q^>SBHUG_^tWv5LyJR1(0}x)`D=JLYuRv(%PqaOgzCn);x{~
zEs|D}4ltf$`x}OzUrKH<vII+TNLa#zL;Qr)Ure-NW#kFw=H%vw3D+<F-y8lhMjj}P
zthzPw|GU`q=xCs@jFpTZO!a~ZtDms%lza*@M1k81CVVKw?q$6E_2WQcF%~f;nDAw9
z&Y~GMNl$>z<+bKTa;}t>6x_KFf8?$G?GP#jRK#b^honf_N*b={sR8R3c7E}0kiDYZ
z@-PcvG$R){R(cQpUa;n&wE{y?YEB|3eu|5di<Lo)AP@lofiMOGbFjeiNoT%kXKrN*
z3}Q&mP07gs%6X^xGPAI<v2$>8^Y9Ar^LYxoI;#nCc{&M+3yVldT3A}!*xK1U=!t4u
wDVu4?V-ppVmXTFZR8mn@*VNL{H!w6dHJ3Ay(lv4s6?69rVRG|x^kDP>0AuT~NdN!<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.217BA51B10EED1BA.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.217BA51B10EED1BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a71092c3a33e0f75bd29e51e0df009f0a02aaf4a
GIT binary patch
literal 1546
zcmWIYbaP{3V_<Mj@vO*AElFfyU|<jg;^LB`%ywo525T+`hKf07Lv!<Qn+aU|uX%3E
zbRBQrAI^_%t=_0+xJ&G(N5B(LC6|d(o<EoFuX8`$C#-(-f{FAu*JpPR@2-B=dyc_~
zAzb7bC+nfpSwUj6zdKJn*#1(!rmVX3Z|1S}&lh(KAAKrZxzA>&LHxYGMT>S=*azB_
zZrt&AlE`-XwCnNp%MV|-Pp|p;Gw<B`Y27C~MCR1T==}V2`Reqacb`kwl{(Lw-q3$J
zlV2y|sZ>g+fEUZ&q_xqDpC97d%xh)va@IB0iPBF@tGe5s&Nb`G)b?HeLG;`D?Iu>=
z59RmEy*p>LL;To<>3QZE%5yJh#Wo09&YIxzey;z1?Ju_@GNaWJg>6mW?<?3Rc+w+U
zR&jymiyYY}Dw)!a?-ugjx*4Z?e{Xd?XXlM`FI<>)wGtOkt3DCQ_4CAyi5*;rqTC}J
z4>^^jy?0A|9oXfu!6~yveBzr!awW2&j8X5U#O3ciY_o8?s}R}p!dh}JKYPcf#W_;7
zTFxD3?NVkv{n3+>X)SQ=X71x%414b@Cv9=fXjxt-Gxh!j>j=*Et(PVRcrN;oI{ns0
z^(!v~R(uG^FyRVXyVYP8kAH!e;jSIIh3mIZk({a`*yNsdeVI{%-rB`Y-*$y*l-jg2
zv|f0&;Df}4^(Q8Wxp)cl)-xnM*L}Kj!NP9>Td!QaxWe7}sm)h;y>p6I9;<g<4!V<G
zTD@}R+_h_?qa(QUgUf8+GK;x1PAGS6l-s4?Wy27wuqbm*YiR9J{;M@b%1M`{X3dD(
z=@xicYgw+R>#gTgv)qCX?6=x7$?e3;xzlU8Ot#duzdvzVu`0IeV%SO7bj_J}-*1f1
zQm+=-<Tdex>#ak-=LNEv{5V=<c(*qFi|^W7`)58;eJl0k*vVN7n5$2;uF16OoO7|a
zZ*n4=Ub9)4NNY^c^O)zHp||`a#A~Wp^v+4X?a;mXqQ*ex!ep<%vzP4qJMExa_1)^S
zis+{4I&b%UzO(hld&TQ!E{?1zskb*?+TCY1ce$WNuuZ)7KDXGgkL}9)9euVfvrOa)
zk5c;oPIq2j;p$bTnw63DxARi;ug1iGx|k;7_V$C~BZqCd$1<X3mTzi|XHRcdJiRq_
zYTnW_%gh%Yv(yd$@q7or-pbnlVpkhaXMF#8+lp`In&XczFY7s8JLTh#f_$5ETv|tG
zY@GkcIC)cDKyAI4oa!$b*^n#S|4%ymGR1D*Y4KYh*GT4j{`tSdd5iwd`_tKX7%>!A
z<|fGl^QCcg`bX983m)<?_y;fqbFp$ofC-=s1QeyF756eSFmU}^aIe|hG?<wuSjbw)
z4X!O%*jgBgCt@uE=P9%~dn&Dcn#aTw%wx^th|nTwDQOSm!DvP<u)V#9elJ+_&|1N<
z{SCv<FD17aS%M`vBrIU6Li~i(Ure-NW#kFw=H%vw38(rx?rq$-ED7j1ertXtn*^){
zka&XDf^goXC+sVDzi$lznj~W>69qGA!GzUM*mp`k1zDiLZ3PoP6k_)>UjF)Vps*N=
zm=a9*vNvbZ44b4UKp*m2^TN&M36`>yg8T5{kG!?N9YUpmiukPgkQ7N<O2ZXBHDLY1
z&M)2#vR9N_9%jLZrI~YkmnoG3g^`1Sp(r&c5fnegMajj=AVv^~fPg?4gMm4C#)_ue
zAdM~~robSE<lK~;44|BMnlCd8D;qlpCpQnT06(9npsTZ*AeX0;khrjjgrtR~wT-Qv
zy@Q^pww1D(hCDV=F=-iD1w|zlRdr1*9eo2sV^ecE6DeII7f~^HuMj3TKSvKn9{^+l
BkH!E1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.CE8AF0B7DC8618CC.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.CE8AF0B7DC8618CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4fa095847156547d88f8247c0eeb50c392081bd1
GIT binary patch
literal 1538
zcmWIYbaVT~%D~{9;#rZKT9U}Zz`!5~#Kk2=nf1&J4ANW-3>9<EhThJ*Z6R>=zew^i
z-&wxAKNJh!uHERH^rgv0NW^jyN8plaOKM(TFVDDH_-IK)j+994M9X4}>&DLM43Z7n
zT#s9r9;TY=b$|UWJn>-rN&A|r;>xd?r`D$({;YiTsB+~ti(LlMbN1N=&6zK&DYtvZ
z{Tkn{+5DGR=ht6;cze2Le9gYOA5R~xu~A`tF8`c+-;a`?rDeZ=&Wewl{LJ#f<CIIx
z#bG<yB1Kz-9Q7k#hrQUj;ZRN!pU<vkrH)V7DtqrON~~Pon{><TvSv-=-_yFi=k|V>
z-Q0frGS9mf!!Jv3^WO4M-{K{n;L@k2a<TgP<@Z8=)xvH~^Neu!e{xs!oVk)@=4xdY
z6V|zFo9C$HDqpbO;%I$)_0!qr@AVZ_x0za9a_IKDvEnIT(KIJJgEEyxO&cx+hAr42
zF!$!pz#COll!Vp<NIi5_*}Z`|*4?+^+U{dVtIKyRlUbZMLC5LK+@5powk}r#O?&O8
zDIQ4^H#%EXGbtt0THxAE-^aTc_TFbs+TxngqHf1`>ir8@o!0A4nVu_D0)K4sE_oxA
zRl;$lLc>Igb>;Oa3Ewt%i>VT^ced=f9KEV1YLZYBd)D=3W(|647q@-e6{b;Y(az9%
zA#ME!i3|ECCx=a$vdrX<T*ut)A<LWEUM##KwD@phf7_|bNAl;+D_D7~-gPzTj(Tl%
z-<5OMuC3Ks)3SZ#F1xq3VlIsn%3T}fb}4w>VUAtka@m4=^{>A7yMMlLJF?8w#Bgs>
z>!P+)pWBmWZCS1TwP#7g|FR6#-U*kZ^>+(pWc(Goujb$WqV%O}w5s^rrAgcNv+uhS
z_Dv&Ys?x;nEv@CXZju)sh<E=?`91fNdEnM+<CEUEOi%PqHg){{Mp1mzvX@Jq^yHPu
z%ujh>l#%QjctX|ixt0HlOLb><J-Dn~P<JfDJ?CUwdh=vP+m%P|TYUT$&i49E>ie7j
zayjERZTWnz_<GX*_L$oOtfmvD<{tNYZ*{9=p4)?vk2OnwaFs?L{V)4dG-;FQS<c`c
zSA>50=YK5ejt*Vxb@KGD-)XwlE1&&(WOvD7Tkf$8ubJhW8q1q6ALKl}HFav<(lct;
z7afz-4gK-H;=5kww(EX{4~%Y={<Gq9R(pN%#g3ng1K-!0+`2E_yQ#@b-tb#@eU#6&
z@2ln)|66ci`HKZg^Xk&gjtl=N_&0CiLr%{1Gk?tAqP<1@CV%Mfb#@HJmAOgs!2I~|
zyPeW0?wWHv4E_NO!Cb6d0bl|s0|7;;X~n&a3=CYq7Tjz0HVtOx2^O*za)WCN7Pb~f
z;)z&`z<CO7&YntZpXM>~1oK$)I3l!2T1wi(c#iFF7=C^!xy8s5EWsgR0TT}K6H<RM
z(T0_gCzzX)n;#~e>g%|-apSTipey*T`H_qiuogh#30e!nd99KE-^Hd!M*~fgv6S(J
znY3WS>L=_wC7*&UP~f(L2_Fivdl@f({Wwrqj73ZdCVbhOvuK7*(i5Opc&&NiX7dC~
zSxUjZ0#nS$1rEaAL%$cSd1$Th@JHU--wvTtK)rm{d`NnwEv4alpBk`!Vdod`202HR
zTOMZAhozZwdzUGd0)>%7hoLAnClM4s#YM@*%0QYq*w}4r%j)+PUQB^O3_*~n1CfCU
zW^!&yP6m+co#xBT!pg?Z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@og
urXi0_R7_e%RzXooMO9r>OGn?p(Ad;m&O}Pr$VF7l-7AF2&Ck(;(FXuL(3B_u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.8B802336DBA274D4.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.8B802336DBA274D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2539d7d22c3f255e1c8274ac0181d9058053eba8
GIT binary patch
literal 1662
zcmWIYbaN|XV_<Mj@vO*AElFfyU|<jg;^LB`%ywo51{E#_hKf07Lv!<QTkw4QFDYT_
zt-qc1M7x=H-qE$W8*X3dZ_>HAEg|xgXHnd@*X~C>FE5kjd)@VLpM{d<B>tZhyx1l%
z?pS2try&p*c6rl{J@&$qH8cL6eQqv3eg3mQMTRo-HSXC?5<fkeyZh;jx1C8fb&D!w
z9=&?z9dxH=Mtt4hZoc^cGe19lew%fFM4WH?k_UetU!A0HA6pZ7r*7Xf@#{+}&wVH~
zd(9iZrigcw7Kdt6+{SBbFBT`Z-ei-VQR4f$=>+d5slBd;Klw=?-7<Aq@DHZ{)2pTC
zRzKW+oMZc9o_8&VUzXnHz2%|4#Y;TFrB6-eV)gUO@45b(h25IwdBes3$z9QN=1P{C
ztCd+ySm&;7o}-ehe8F~$qxJ38PiLFIzwgMjt=Q_KL$}wB6;JOy(OFva;Ejh1YvS6)
zYa9|4tv2&7-tbk!Rb{hNW{dd5KZowUu*zcEa?e!nUi)F*#GYM_pExd-8(GQosjcc;
zmHA~QkD}S{DVftA3r|@3)*&K&d+%$eFNOVLi*=opZ9W!;)H%OTP~9Wu<-N&Es8xTr
z%W=7Y#m)<kIWN4Z;j#DHp^PG%jlrE4kLGq)?_GLog(|0jO~7umt&Am^vZh<EZ*^Mv
z*pfw~#f-i6ctf4*#4Snzvl4zZE0*00n#(1%sC|K|Tl?XAQd-YD>m}wpmrMvNUA}aW
z|F3T{E9b6V8yy|NogZ9k`=(j9gGupy7n5y?<CF@6lE#K*g<@e}W$jBoU+nGh&CW=w
zdMf40yL!epRnyG)@K=(ajQ`(W@RCwoULN>O^}>Z;b8Fe>@m+j<se7B3qLug2H+B6r
zE9brmJfW4kNFbB@eU*#kg$LrNSKoYRy*z(W*8jAiDO<89NULWX$zE{QOYyzz`JgAS
zWX`<Q2Sypmu8|rlhR?11!<PP?QhIQ?dqT~*4ELOqZO@xGGuo~^a^K|Rw{*7IZ&Kgi
ze3;7_w`t4gbH&$__P58}7GO1<Fg5qM*L$m5CG*@8mQ?HutLfdf=8=DTUF)2j#XP|~
zE(!hQ*Du_4#C!Fstxr<_=I)-jH{|mkiN5#7yeYTSdX`P8RG0pdY#=pfV@!0|%dk^9
zWx8207lK;<`&GzK%BarmV4i+)-RJn8h@N|Pe6mK7{%XrjeAH^^oqSS|Z*-C?>rl(a
z_&>jsHth@lJL50=QoF5uT1jup7g^3d+CKRQkK*!@J!<d%U%qsqDdj$6!ISm(7>X-%
zljMQ<a^HJT+sz;BLU<Vb0~mt2S-GRY1W*P7ic-^x8JQRuxPC3T*X(T?%*+!kY%S~t
z*A^^dErP@owHAf*_#3VUcb{4vz{C?Q#UbU4(4wfU7zg7ChF;1jSi-XbXfD4szavbT
zCs@`+)(gQ?Rae!8@f_RVF#P;da*K%t#axzPWOH36e|ZsDBzzBOF0VB&lDX0@(n#hi
zt1H9J4e=9Fe=*U9m60cyhm(gNW?gfFl-p`U<1(N(1gr(@VZuDYLe@fXUOQ9Iibao?
ziU9S>xyVJp6fKyr`U(3^$)_MA6?m**!Z4bV3mn+Jhkh?u^Uzx1P>9{jc=_wcfi_97
zNGQQn@dPWdD8PA_y*Z0!*d#pxdW_GS56NRPE;2|SQ&Cred+ezJ>lb!@@otbSM0w<4
z=KgH_w36Gz{t8f7O<m0cCj3_*Y}4KRBhP@sg4Tj|FkyzG)SN_6nkX(xE>;FHf<Ocq
zFbDHZ&ERWGdbE-$Fo+=#Cd|N)oSTx90pxn8`7*Pxvaxe;a`W&C@bh^Jx;m=~a(Oxl
zi3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g(Kj$OHZ_+sk<vAC5fyXy
N3Sn~dbM#>J0RS5YxmN%H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.17BBA75937C47055.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.17BBA75937C47055.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aeb9b1607dd1938801a8fa9e9beedfb3a7773a5a
GIT binary patch
literal 1644
zcmWIYbaP8#V_<Mj@vO*AElFfyU|<jg;^LB`%m!u#25~M1hKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%Wp7Y3+$HwYBjAarlFLLX&!0>8?{h!hC#-sOgO0>E*JpPRSJ%FK!K1)j
z;F>rmKw;0CwA77%3)GB$Sp09+&(mLTr~l(>k&UFrJ^P8%S08MB`s&a0#-t*{0Gran
z+TW&{dG>ni_Wkv1&#%|}@#^&5N%8B}r6&g!{Cin-G<?4Pe!chm?tC~ceQ6I%&8+j)
z4`zj~a|~l$=yD<Z+AiKIy)_2A7cfs=ml-Rt(=l%G_Kb*onTuy^yHvszpZLGZdvUS;
zpLb6lzO!krPUNY5S=MZJiF55GR*8v=IX%DJ`}t+J+y3sU+fs$LT@<eT-uk)M!|m29
z4n|+c^w$dNp0_zl_;W9u`<C_d-QD)@6NGL{^O{^(BxIIV*)V76g#I4c&J_wdB~wH-
zbX?N;@<g`DE$v{@p5j&^cv4Ts;d-F(g0JfeKOU6L(>%UHTBTc{R(<j3ha7>WA-;>-
zjaqKVI4z6pJ3mP$m;I2{?(7)x1>dWkdbV-&emHX6^QY|I=B*#1G;}$qs&K`Zb@n+s
z_3|}J@wd7yR`?plz4&B7LvZKCqp{uWdzYSCp{l7^v7{_Fi}}SRtE`Om*{wm#i#Y-~
za{IXY8UN0*$#nJ1^g1lxc%#z0R4b!H&P&(OO745ciaF~45{nowI!0aHJMEF(>)YN-
z%C>FGOHNeY<9qdMru;?)u9iL8TwgA*20jve!K6CtSY*W2@;^7us+%jBtzFppa;mh5
z>6+wQ)2=LWj$N6q!un^mvvy|7oO{)lQyiRMfB#`uY2~usTmP2U&KEkOoB!K>T>5R3
zyW(n*&X^^-d)^zLYGB(R5qtjoeb!px(0235dzQpD-?4}kE;%Dtq?kR+&+FU^x8pY>
z+4P#t!bDnQgo@XFZqa(1zJ}Y*mT5A#n{|@6**=N6y)Qgf|D|<qzi)ISZGH9mzI^Ww
zla|gcpSN4=x9yW#Z5k^Y*W_4-Rb9Wh?6V2e^RDO8N6lB9*82DOM83e}+^LCN*Q1pF
zzf|>`UXdBPb>f~i``>TdG(9V>N1eaXZBwqXoAaqXlX?GGS;^1c7!w`#GVD}NopM&p
zCBN4F{ofx<il~0f<F2>Z_j&$fgOpA4{My&f_p5$?`Q^uZC9_0Beoom~|8x4$l)npp
zSw8OPs_E;~TJ&=NnKX;~{^jpkrP@^Xq~80#*mdF36LkzFO{?b>SLP<k1M{P;VQ$zV
zo#ad&2LAwtU~X3K05AcRfq<gawBjF(3=CYq7Tjz0HVtOx2^O{%c7tmR7O@sV;)z;|
z!g+$Bmof^L@N58T;kV{@L}-z9mi0vNRMk~=VLZY0ol>RiUL6K%5wI4pg9-Bl3t9^z
z@r0~};5?WbMlP_!dk_6yu;!t)f@Av|hM!+bZZWY0OL0g!!BnvXD=I6-!gwx|zq|-6
z621qtj@OzO$%WF+(nu~;R#%3*FvL$t{l!EZRz{v+9!?&9m`UwSJu4PHUMd0<mUEU1
zhY2s3u=)x6PRXYr)e1aTFyTWXb}!@QuODaP36@}yP=X2b1S_y8z<C$%UHhPCq0$Oe
zgzN(z6dzpn<}8|Flk^1W13qg$Bp=8)%OLqcMO_8%gQo_pU)cG@yFqRe<&lS3{<HDZ
zN^TSTD?nj2bu|x|a8YVbA}A#k7bO=fgBU>|0t}dgr#SEMw%WFJ3sYbaLm*6;fgw3J
zB_{*O^-l9;W?^Mx=iub#;T7QL^AvP-Rukm%bP^I57Lky&u(Y<ZwX=876V<j-Hq(&D
rCMqT^BdegOq@t>>simWDU}$V=E@vX8Yvdv-=I#~3<mTt-!RP}3^!A2>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.9C7E5EE7984A7775.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.9C7E5EE7984A7775.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c3369396945c2f68b02cc40408beefd2ab1208bf
GIT binary patch
literal 1644
zcmWIYbaP8#V_<Mj@vO*AElFfyU|<jg;^LB`%m!u#25~M1hKf07Lv!<Qn+aU|uX)br
zl}<MA59dd>R&P`@+$HwYBjAarlFLLX&!0>8*SVkW6IMOCK}X`7>$AIut83rA;89>M
za7~;Ops;7nveb=#E7XjBnEal5K7akX{bzoDee|bj!H*9vackJb#a4g(-T3HJ(t?Ut
zNtM5KHS_GxuHXOPt^IxdnV+v--))LtJMD6DP{F^-osU-Qhu;?~kI(;cl{Yhv$?klb
z?S_=8*%Lw?H3YxR+M3rqk6ZhU-irfEvZKUZ@+Mq=p({15{FcVDwI*{9h2PkJ@6wAq
z-21I7cSxrf==Uttzngc#&1zxjE(YZy6UEMZ*7NJGzbM~$X`9a>vnr+c`wI36p7e;8
zRa~I?B1iU#N~Uz<yM?^BZpP`}-~0OqV^C7{;zqWpkV9^|{1Y~de44mXSx~g4w{H_u
zYwL?M2YU})TP!j$q4g4nzS7NB<`-63Ok3`m>e=l$tou}=v^hm^(fgDVI~m1QeXBCR
z1hfgvc0OTp`eWpzrEeV~(#v~aGkvM^@3<(oP;lN43(fik<r~D}xR<IfoaFNGwBE%+
zj};dc10FBDkRh@pE<4dglCMC^aMzCPj|;YkNN!b8W$C}N&Nr7KHp-p%ZRy&8S0#Lm
zTno<zJ~UhqKUsaPhNf@k4`l)C{FQ#JyceBUs=A#&Uu_a{PWXR8CF4cMsH=OYJ+ga!
z+j~jbwrzRIiOPF?uYS#x->AUVvPYZi%LUfJM}jYyRA(KFjJR6<=f+ueb0xF23p-y<
zl@>8wlYDF1l_k!xE7Mh2|7>>F&TN@;uiA2ogY)a}KkO>4T-JN*-_qLoLPvD-f7_2s
zzio0?TrJWWvqX2#d*f3LZ2Kd0(|^C8v#)d2;n|*Xm)1Gv%~``~mBznwf~l{%YI@1y
z$6s`qr#nh(cRA^GT3)Yon))sH8tZ(y#?#zx)=A!Fb%s{b7bkoD@fFVBpSIy_-1m91
z_oqEj^|pR*Q?C21`ovb=zyRjRn`LXhtXu5+Jd@>E*K_Hk<||HX{d;^OUtn_X)I_f9
zQA+<`s`^c@$PC>&anG9iH`_K%&x-50Z_gria@$!6Ud_*bW*<&x*jOd++_vV@+9@~R
zxLw(?U@lj^{QHBd8^0Zs?2B=meeV9T#1kpDGkK!zXMVY-KKb$9lPcOPK5-`h|D@l2
z@|WY+;$!lnKV)P>T`t$pJX`Qj&i_8Q36E#R={^75yPdoz{$)*3TRp9~GB-&cm>+%9
zQ?q$by0q~y_y;fqbF*>>fC-=s1QeyF75`vlVBq?-;9j%0X)rTSu&}kT8(dqkh_wh3
zPt;lz&Jzs1lu@vRX9G|Rzcs%jLW``ktS5q}s;;UF<2km!Vfgu_<Q5Z4uoQ=s6HJ&T
zSW#Iq7RGa#{N+Vpk?=jBxxCiANajjAOCy=9tgZ|<7p9ew3+%hzL%$cSd1$Q=;wPm3
zVxkQzBTq07Cl5bNRXbD9ibao?iU5V>oaMq{!V4y>e!{*}@+nBQ0*@6;_)v)5%Xs<g
z$C-G7C0HbsV8T4X3M>k6-eqsjq8T<xPk?Ubv*tr`vy8J0lABf3Rp4%ZYQXx1onO2g
zWSuCFJj~qh!ogEjUu_WqT8A7&JSai*v+>hPZWH?}K)q_}Y927Xe+9xe-OWGp4Co30
zYXLi$Fi)_cwIC8t$XW=_D@x5t1f_)HqU2&_5F-dgfB|#xk{QP|KJIvu&J-BL5C{`y
zU`WnQ$;kk6z0-V|Sy<WFIXJm_cm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy
v%{1h(iHb?f$SNo*si>-JYU$`37#f?J%b7^&8o7vyxqF2$x%oMIF!}%hFjtW!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.15DDB04E41B23392.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.15DDB04E41B23392.idx
new file mode 100755
index 0000000000000000000000000000000000000000..23a7bbc48f125287f920a66862a0a7649543420f
GIT binary patch
literal 1662
zcmWIYbaN|XV_<Mj@vO*AElFfyU|<jg;^LB`%ywo51{E#_hKf07Lv!<QTkw4QFDYT_
zt-qc1M7x=H-qE$W8*X3dZ_>HAEg|xgXHnd@*X~C>FE5kjd)@VLpM{d<B>tZhyx1l%
z?pS2try&p*c6rl{J@&$qH8cL6eQqv3eg3mQMTRo-HSXC?5<fkeyZh;jx1C8fb&D!w
z9=&?z9dxH=Mtt4hZoc^cGe19lew%fFM4WH?k_UetU!A0HA6pZ7r*7Xf@#{+}&wVH~
zd(9iZrigcw7Kdt6+{SBbFBT`Z-ei-VQR4f$=>+d5slBd;Klw=?-7<Aq@DHZ{)2pTC
zRzKW+oMZc9o_8&VUzXnHz2%|4#Y;TFrB6-eV)gUO@45b(h25IwdBes3$z9QN=1P{C
ztCd+ySm&;7o}-ehe8F~$qxJ38PiLFIzwgMjt=Q_KL$}wB6;JOy(OFva;Ejh1YvS6)
zYa9|4tv2&7-tbk!Rb{hNW{dd5KZowUu*zcEa?e!nUi)F*#GYM_pExd-8(GQosjcc;
zmHA~QkD}S{DVftA3r|@3)*&K&d+%$eFNOVLi*=opZ9W!;)H%OTP~9Wu<-N&Es8xTr
z%W=7Y#m)<kIWN4Z;j#DHp^PG%jlrE4kLGq)?_GLog(|0jO~7umt&Am^vZh<EZ*^Mv
z*pfw~#f-i6ctf4*#4Snzvl4zZE0*00n#(1%sC|K|Tl?XAQd-YD>m}wpmrMvNUA}aW
z|F3T{E9b6V8yy|NogZ9k`=(j9gGupy7n5y?<CF@6lE#K*g<@e}W$jBoU+nGh&CW=w
zdMf40yL!epRnyG)@K=(ajQ`(W@RCwoULN>O^}>Z;b8Fe>@m+j<se7B3qLug2H+B6r
zE9brmJfW4kNFbB@eU*#kg$LrNSKoYRy*z(W*8jAiDO<89NULWX$zE{QOYyzz`JgAS
zWX`<Q2Sypmu8|rlhR?11!<PP?QhIQ?dqT~*4ELOqZO@xGGuo~^a^K|Rw{*7IZ&Kgi
ze3;7_w`t4gbH&$__P58}7GO1<Fg5qM*L$m5CG*@8mQ?HutLfdf=8=DTUF)2j#XP|~
zE(!hQ*Du_4WbLYJOP{3v&D}k5Z^-985`FKDc~fqu^(>oGsV@B^*+6Q}#+c}^mtm)J
z%5<}0E(Epy_p6Yflu@1A!94xqy3g@F5k2?p_+*VD{neJ6_^8#+JNcv_-{>S))}fY-
z@qd0NZQ2+9cgA1#rFL8Sw36PGFS49_w0-gq9>wJ)d(__jzkKOJQ_6kD!k+c_7>X-%
zljMQ<@{{g??wlD-e|Z@E0~msNSb3tr1W*P7ic-^x8JQRuxPC3T*X(T?%*+!kVlCnZ
z*A^^lEsDewvlfH%_#3VUcb{4vz{C?Q%^~fK(4wTQ6bIuev^je!t$msYG?&ks&k-if
z6D;E*<Ava<sH^D0c#iFF7=C^!xy8hSVlGQCvbiCCLh3Ij+ORV61oLw8^278tCrG)i
zHZ(2+x=GMl&>kkt6D({k4Cl2o^{iO*c&P|bue^(V1WeI_39Fy5@05HBGE#xp3MPCg
z#O`Ih{Pp8NBPCfRm0-d=!HO)3aNcEa&Y~GMNl$>@<hSNW@}{hdERr`>)m7o%eE1`8
z?Qe%rDWDbsYXKxJaxQX6TGZ6l;98y<uzq3Z7w-nSP?T35<`fvs$OVqu-b23^ta)gy
z@U!vLN^TSTD?rui>gpaaReuG-Hr>rX@(d^}WG!R|6J{t%%}E5MiQ=N<Vr38`2t<GZ
zbMOgWg~qg=AKXlVK@5Q~VFrfe+?1RQAlEz1mzjl?jh%y&n}=6`pU+d!)mcrD%hO3n
zTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8RqLPZLx~7(nzJZ~!skxkql&+DBsF=G~
M2$P$iqX(l806`SUT>t<8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp.C031B76FFFB0A20D.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp.C031B76FFFB0A20D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..acf2197547d4612fddeea1cafab60c70fdd67ba5
GIT binary patch
literal 1702
zcmWIYbaR`>#=zj5;#rZKT9U}Zz`!5~#Kk2=nLW%53|?Fe3>9<EhUVtqHWRq^U-Mkg
z^fTVPKb#-kTD?)taF^ImkANqhN<I^%Jby0TU*~?hPgwQn1rzCSuFvis-d+8!_Z))}
zL%7H>PS!)Gvx3BKe@|Cx=D%$J<IT5Mzb^O0pKo70@6o4DoA_^fXQn^j|3zeR_q<0d
zyN|w|zGTN=gWdmrS<CMGXZiT_`SRkrtua1rOCJ1roOE*f{B`@zytljmVYT$7JuEeI
z{HzlWYK1GWZVFI(VVZr1%~m{YM(l+~&v0|SjynqR3%6%P+{;=#W80-IT=9wjtGpK%
z>;HN8<l#G;=ITVA+LvX$W|uhEUSgG)xR}%P%e|jpcDwEGp1LhnXxqi$&+oXOXHV?e
za#fINR@1p_j((H23BHiKwP5bsE1&M}{yu#|$J^d2sf++l>8&;n=YmeSAM0}o5iqkl
z#U&!P$duijOS(VErAgFFc~3`*INyTol^ho;!gv38z_<N~;VY?=N-p<Ly{NHpnj)JT
z{CXdc;<fe(nbRH{Pg<Jm7?EDSHJiDl&cEZL*g~;+zY2ruoZlzt?h*6yc2ey;G(9)#
zcw@j?=LLtH7hcrx*m~_yhS8gcQ<6#wuRn0i*EH7hRATABvd%M=AvVgL^=(OSz^f8I
zMy`ct10NbLh@Y&!wquH#)IW|R*P>4aJNmv=%v!nlaAJGgsmfROJ$edO9;<g<54w|H
zTD@}R+SthL8#f5vTlU84mb*>|lVW~1lkFDADHSZY99%A2aIfCgCBORT6SpI~3{4E@
z7BwfazMip7RrKYA@KutYjQ_JPdPylR{~pJudEvq@+kNW(?Jr7Sx(2I?uU(q*ZU6GR
zm3!X=p3rht63^s*Z|x>|;eq(+qb2|6UNqnO>O<P3>07cVNUMjl$li3;d(yRR@(Ib<
z*A|wi52Ri=(zR%U_p3SCwhO&}b7nXD_aAuCbg}QoBwl^Cbmpp{mGzUaY`<@GBW->4
z`M!Mb50jS8EuXiWZMW@{TWuOE8rI}Qg;ia@wCuA9)AO$9(nrl#oYwkR@Wj01<hD}=
zt=YO0>%R*v_r9|#E7&V;-M+WmHcbxOBhgp!=;-H1lf=(Hb-vP=Gppy~5u^7hy!-ji
zv8U@?pBB1nqS5So9irzl4<52-uP^Y_{bT-;FUBpOp<esF`g6s}t3U3zzu@DJZWB5A
z_r1p>9aSZbzfG>s@mVJqa({le`$;=@cG0iz_orN*f9bgKFDLI`yp`Vf{$DR$boJyv
zwsQ{qrWIG_CdmWyra!CtPU)lUTs#c^0Sv+1tlTkR0w@CkMX71UYD^3aT)!6FYxXt`
zX66YNwib4SYYP^!7D3{PT8qMYg6lh_O4q$Q4AdfEEntVxB4{m$#1pa>g7XyGoIRD+
zKFtGa;kD*<L}-!YkrP7j)Opn5JjeDo3_rh=++t!0mf?``gDGMOR^?EI^I)QkTwvez
z9{Rmt%|mO25I-UH7ZYt*8F_+vIC=PCs@j=)RxEnFR0Jrj$fKwL6J9W3^%M4;l21XZ
z6?m**!t0|dlr)X@^Z<p$ti|A7*)WN7JDc?SnLuH2YjInc>O&!RFXQE}A7|nTmST}o
zf(i2kE3+uWdB@_;@UH8*kqlIX>?a-+KVA0bESh1H^aL1IeAawOVI|Ka4-XxlU=1D(
zIPc+)ytTg_LZyIO_^tVov?%Z>z!mWXYw~Eqc~1>kzp(R*cY{1F$|DbR@6X0hE4fYV
zuK<O$c(in3!bPb$iJ){-T$Eg_3}OU<2rys{_B(dMB|~PzYo?$;hCrAw14D9dN=^ol
z>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3wY^EWPO;k);
nMpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)5$<5EvgV6^7io%ge

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.C7145A60489EBDB5.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.C7145A60489EBDB5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..933a2277d3053c7948f0b47ab0a8e2c51dab8349
GIT binary patch
literal 1644
zcmWIYbaP8#V_<Mj@vO*AElFfyU|<jg;^LB`%m!u#25~M1hKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%Wp7Y3+$HwYBjAarlFLLX&!0>8?{h!hC#-sOgO0>E*JpPRSJ%FK!K1)j
z;F>rmKw;0CwA77%3)GB$Sp09+&(mLTr~l(>k&UFrJ^P8%S08MB`s&a0#-t*{0Gran
z+TW&{dG>ni_Wkv1&#%|}@#^&5N%8B}r6&g!{Cin-G<?4Pe!chm?tC~ceQ6I%&8+j)
z4`zj~a|~l$=yD<Z+AiKIy)_2A7cfs=ml-Rt(=l%G_Kb*onTuy^yHvszpZLGZdvUS;
zpLb6lzO!krPUNY5S=MZJiF55GR*8v=IX%DJ`}t+J+y3sU+fs$LT@<eT-uk)M!|m29
z4n|+c^w$dNp0_zl_;W9u`<C_d-QD)@6NGL{^O{^(BxIIV*)V76g#I4c&J_wdB~wH-
zbX?N;@<g`DE$v{@p5j&^cv4Ts;d-F(g0JfeKOU6L(>%UHTBTc{R(<j3ha7>WA-;>-
zjaqKVI4z6pJ3mP$m;I2{?(7)x1>dWkdbV-&emHX6^QY|I=B*#1G;}$qs&K`Zb@n+s
z_3|}J@wd7yR`?plz4&B7LvZKCqp{uWdzYSCp{l7^v7{_Fi}}SRtE`Om*{wm#i#Y-~
za{IXY8UN0*$#nJ1^g1lxc%#z0R4b!H&P&(OO745ciaF~45{nowI!0aHJMEF(>)YN-
z%C>FGOHNeY<9qdMru;?)u9iL8TwgA*20jve!K6CtSY*W2@;^7us+%jBtzFppa;mh5
z>6+wQ)2=LWj$N6q!un^mvvy|7oO{)lQyiRMfB#`uY2~usTmP2U&KEkOoB!K>T>5R3
zyW(n*&X^^-d)^zLYGB(R5qtjoeb!px(0235dzQpD-?4}kE;%Dtq?kR+&+FU^x8pY>
z+4P#t!bDnQgo@XFZqa(1zJ}Y*mT5A#n{|@6**=N6y)Qgf|D|<qzi)ISZGH9mzI^Ww
zla|gcpSN4=x9yW#Z5k^Y*W_4-Rb9Wh?6V2e^RDO8N6lB9*82DOM83e}+^LCN*Q1pF
zzf|>`z9aPNvMGDk?0>&))AX#k9(Ddkw@tamZqBFnOy>P#WhFm%V@!0|%dk^9b;?;W
zm;74y_kVvdDWdu@kGtMt-{<*{4N^AE^J`x_->>@p<(D7tmCO<i`8j1{{m<z~Q~oaa
zW%;<DtER6{YthU7XVNU{`<K6Gm1<MjlX~y}V%LRBPt-A#YOS7ET$!6B56q83@AI;i
zSsFL<F!%>B1oN=+1b_*k3<MOVrWOBSWMJU>wcuW}w`njlPq2u!h#Op6u&A{t5>L!p
z49*i=-zinP?$u$S7C~!4JA@V?Yat|_u(dFp2UEky1$IU6q2CMEJhWC&Xmj>dTKhB)
zXcC__pCe2aPq2)$j3<JpqOPI~<2km!Vfgu_<Q5Z4ur!CX6HJ&TSV>ta7RC$l6H<RM
z(T0_gCzzL$mmenF&eXGF(c`5eKw)`j`EZ!<f(fgiu<w+73R11WYXuWN6k_)>UjF)V
zCZ1qP7D*+TFi)@|iz1wN@!qu$dKN0JKt;&T;X!fEWpB=+88%5zfZpV{=120Ttg|eV
zH&xYD;of}sBX8|*hfpb?76EGkBrS5za!6X#)Yafxo*J-zVdod`26;o2S03ixpN*eZ
za+}y+0Sc?Dt9!tNi&AqEK`EiQD7jb}#0UZrV89&wcBa#d7Zw&1m;!?s0%5`o49U4E
zIT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQuh=intrL~Q%oxOvesJ4}|nT9+z
qQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdrm8H$O)YMjrr{sgIrj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C5AE3E9229A8EFB4.idx b/.cache/clangd/index/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C5AE3E9229A8EFB4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd55c7f849cf562da8acb17b6f14b23fadc2e921
GIT binary patch
literal 1644
zcmWIYbaP8#V_<Mj@vO*AElFfyU|<jg;^LB`%m!u#25~M1hKf07Lv!<Qn+aU|uX)br
zl}<MA59dd>R&P`@+$HwYBjAarlFLLX&!0>8*SVkW6IMOCK}X`7>$AIut83rA;89>M
za7~;Ops;7nveb=#E7XjBnEal5K7akX{bzoDee|bj!H*9vackJb#a4g(-T3HJ(t?Ut
zNtM5KHS_GxuHXOPt^IxdnV+v--))LtJMD6DP{F^-osU-Qhu;?~kI(;cl{Yhv$?klb
z?S_=8*%Lw?H3YxR+M3rqk6ZhU-irfEvZKUZ@+Mq=p({15{FcVDwI*{9h2PkJ@6wAq
z-21I7cSxrf==Uttzngc#&1zxjE(YZy6UEMZ*7NJGzbM~$X`9a>vnr+c`wI36p7e;8
zRa~I?B1iU#N~Uz<yM?^BZpP`}-~0OqV^C7{;zqWpkV9^|{1Y~de44mXSx~g4w{H_u
zYwL?M2YU})TP!j$q4g4nzS7NB<`-63Ok3`m>e=l$tou}=v^hm^(fgDVI~m1QeXBCR
z1hfgvc0OTp`eWpzrEeV~(#v~aGkvM^@3<(oP;lN43(fik<r~D}xR<IfoaFNGwBE%+
zj};dc10FBDkRh@pE<4dglCMC^aMzCPj|;YkNN!b8W$C}N&Nr7KHp-p%ZRy&8S0#Lm
zTno<zJ~UhqKUsaPhNf@k4`l)C{FQ#JyceBUs=A#&Uu_a{PWXR8CF4cMsH=OYJ+ga!
z+j~jbwrzRIiOPF?uYS#x->AUVvPYZi%LUfJM}jYyRA(KFjJR6<=f+ueb0xF23p-y<
zl@>8wlYDF1l_k!xE7Mh2|7>>F&TN@;uiA2ogY)a}KkO>4T-JN*-_qLoLPvD-f7_2s
zzio0?TrJWWvqX2#d*f3LZ2Kd0(|^C8v#)d2;n|*Xm)1Gv%~``~mBznwf~l{%YI@1y
z$6s`qr#nh(cRA^GT3)Yon))sH8tZ(y#?#zx)=A!Fb%s{b7bkoD@fFVBpSIy_-1m91
z_oqEj^|pR*Q?C21`ovb=zyRjRn`LXhtXu5+Jd@>E*K_Hk<||HX{d;^OUtn_X)I_f9
zQA+<`s`^de5qfpmls#+e-)!47Ju9x~zCDZ7$!%vPcr`!!nSD5&VPloNbK9CrYp2|N
z<921og1KDv^6w9-Zv1vkvM<JM_PP7V5>KSq&g6-<pZVpU`sBxZPpW9I_{5p~|C4_A
z$zP6Ni;u~R{*aLkb-7$W^K8LCIsg0ICOn=Mr}zAK?{@N@_?I>1-0Eq?mAOgs!2B4Q
z-g5YGbNnM72LAwtU>;VU05AcRfq<gawBjF(3=CYq7Tjz0HVtOx2^O&yaf53M7PS^d
z;)z*{!FdX8&YntZpXLFz@LBUYBDBaj%XlJqD(Wh_FrH)k8-|}>N^UW+1WR*BJHdom
zf|Zn&Vqv@xKOyxO6Kz--d4hR4dHG?&Fq)AI?1SDzzZa}|XsyuB)U#sI<E0`%)$-2r
z;V@MTCaiwKzEkok$PfiyE12-15WAQ0^4E_u@dQh<NGidEd4d&L6ydzf-ke1<Y?7V;
zJ;ra%kK{2~XIUhVsj92OJ@)WN-rC;|p;ACC0@eaZTI8JNkhG|&tHHHAHDLY1&M)2#
za-k@%Jj^NIg@dQ6zS<%JbRlvO@t_3J&&E$HxlQb^0QIV?t9!uo{uKz@bT|LVGoUL3
ztp)91!aTu3)<Q@;VQXPHuP8Mq5tI^&i;|0#L5v^}0S3&$>unx*AN#!|mnkraArL0a
zz>u7ql9K`CdZ+m^v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI+SxnkiE3LZ
vn`y{n6BUz|kyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN32ia`SWaVDte1K#jU2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp.8F52AF019F02B944.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp.8F52AF019F02B944.idx
new file mode 100755
index 0000000000000000000000000000000000000000..465fc541e4e1582f2ce211b116cd0fc753afea50
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%m!u#hL4;K3>9<EhUVtqHWRq^U-R6S
z>1VQ8e>gw7wR)qP;V!YC9v7Z?D!ELQ^8C4U|33HAeZuO}8%!j=xjwslcz5-?-g68_
z4B;ZjI9U&!&I%H{&F?(%;PIF8HD%SMe=Cozf4=zR%twEPEBD##Hi(b=Q?yJXexByM
zT@iQce7kP<Z+@M>|MJ7z^NsiYDKdW^e)?#N3)eY&-Dx#Hf4+MC=iO=E{ks&;dLLN+
zatpiInj+p!S{$lPahtE@UMNm%y~!#&qr~@hQ-tIb)2i;awN|+zm&0cHJ=A^^pP%{V
z6!(7X${o8;JcyHUj^A~6L2rrE>Jmn`H5Xb`s@~W9id}sFsMl>H&f70mR(@~&-0R_X
z>lFv1uVeb_1lLL11YgMAS}^zR)lYYK^S@UVye&O9bAhXB&XrFLmfni;lGZ|jjyGOt
zXa_`eoRhhukz?y6$Qh>5`$1_^tVP4R#YzsZV;?_iw!R~De1)`1w?M7>;?EB`0?VfO
zE<T^ivC*={XS2-wi8{ILhpcvI$A~ZZUhUMgjidL&mHx?}tgG0wA8rkZ64LS%+4m;M
zny*DlmPu5OtJlqO)mAaL$qyy{G~L9@b>+THGg>@J#p#2`+gl-Q7nYWWT!_2Qwbbvq
zzyg6=GNN+Kzf3DHb$Mhidf4yqX3tU^QIjiji>6swNqx;&F=zR|!#{);HE&r}t$VaK
zZ2R<Cw{G6r)!W+p!7TOqWxFIt5suaCG%h%cyL7rQ;C3<T-MnGd`X3u-`RBRhL^%sz
z);8D3ia5SSccn-Bo}jZHTtC9w!!B`H?)_fm(b#tF{YTqME0^`&`nR-pzR*$K{NMJg
z_uJHdr!Y;SogUE@_l-|Au<eh~TlV{X&%TwR?dFsBEQxKNTDVE=#R<7Er)ws2UCzC5
zJATrDdAg&tc9)ah%HnmOTeRM)ui>_{WqR!GW}W11wo8VG{mYV||7t7q_or?+8~1&l
z?EPsERK2a=+m!2mt3I)nH!y%P@@CnZFY6ZjKF?&)6VW$&72p16W<rkh<bxvXwz~Y2
zdB*u{Y3NnAPn-VjyL~EdRi&ZS=Zj@m_~e~<Qf{a9JeyRhF8!mlz))h-)NN}nt(|i7
zi`x~8<)&Q!1>PSNFMY9gCvS{f_SyT#5>KSeeg6FExzG2i=gGf+S{W8_&NA8l=l>%q
ze;54vt!MA`!P08Q3djGdvp>(g?SF^+t9|Hy-If3Jzcr@W|7P@@JF}#?GB-&cm>>Uh
z8gE~gs&}4;!9Rc@n1_`o089X7AfPBUt@sWj0|VEu1^1f0O@mo@g2f!fyy4n{#T~?v
zcoGg0NIXS9#YmV+!O%+?1xt7~F!2NnItaSKgn5GH{p3SnJi+yyQl;x&9cJbU7IF}B
zh6(co3p)rS@kAX&;XKFoHw-_&l-y!u36@rt_JwJ2nf&ENV3F`Wpmlr>d`Q;G_{qSn
z3-J?De=*U9m60cymy?$trZ@eZ>6SE!ju${-WFz;pZAnp{nY|GxtmLN@1yg+}#O`Ih
z{Pp8NVJQ|VC7AGKZ_c6_Hc3x_4&Zm-M{<CypDf$~PYqbVu=9&|gB&2rD-YBAWy8l!
z(Tx{1fq^aHAb=F0A`T)*JUKr(c<}NBEBGnEdB0tDg)(28$_z4Bf!7{pE{ta60!MQ1
zq2CMEJhWCQO3g_G#ddK~a<MXyW)A*nDCAase~&3sU=TwPBq4yvKm;>6Hzg+n$n{S1
zWoBVzW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-
ksHCE*uBoM?Z(wL_YA$CYrEBCOD(3DL!sO=X=)vd%0KN60oB#j-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.3C3BEC3EC7808D5A.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp.3C3BEC3EC7808D5A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..113a03e7260bf0549f336acda17ce278236587c5
GIT binary patch
literal 1618
zcmWIYbaV4!V_<Mj@vO*AElFfyU|<jg;^LB`%tmGghM$}a3>9<EhUVtqHWRq^U-R6S
z>1VQ8e>gw7wR)qP;V!YC9v7Z?D!ELQ^8C4U|33HAeZuO}8%!j=xjwslcz5-?-g68_
z4B;ZjI9U&!&I%H{&F?(%;PIF8HD%SMe=Cozf4=zR%twEPEBD##Hi(b=Q?yJXexByM
zT@iQce7kP<Z+@M>|MJ7z^NsiYDKdW^e)?#N3)eY&-Dx#Hf4+MC=iO=E{ks&;dLLN+
zatpiInj+p!S{$lPahtE@UMNm%y~!#&qr~@hQ-tIb)2i;awN|+zm&0cHJ=A^^pP%{V
z6!(7X${o8;JcyHUj^A~6L2rrE>Jmn`H5Xb`s@~W9id}sFsMl>H&f70mR(@~&-0R_X
z>lFv1uVeb_1lLL11YgMAS}^zR)lYYK^S@UVye&O9bAhXB&XrFLmfni;lGZ|jjyGOt
zXa_`eoRhhukz?y6$Qh>5`$1_^tVP4R#YzsZV;?_iw!R~De1)`1w?M7>;?EB`0?VfO
zE<T^ivC*={XS2-wi8{ILhpcvI$A~ZZUhUMgjidL&mHx?}tgG0wA8rkZ64LS%+4m;M
zny*DlmPu5OtJlqO)mAaL$qyy{G~L9@b>+THGg>@J#p#2`+gl-Q7nYWWT!_2Qwbbvq
zzyg6=GNN+Kzf3DHb$Mhidf4yqX3tU^QIjiji>6swNqx;&F=zR|!#{);HE&r}t$VaK
zZ2R<Cw{G6r)!W+p!7TOqWxFIt5suaCG%h%cyL7rQ;C3<T-MnGd`X3u-`RBRhL^%sz
z);8D3ia5SSccn-Bo}jZHTtC9w!!B`H?)_fm(b#tF{YTqME0^`&`nR-pzR*$K{NMJg
z_uJHdr!Y;SogUE@_l-|Au<eh~TlV{X&%TwR?dFsBEQxKNTDVE=#R<7Er)ws2UCzC5
zJATrDdAg&tc9)ah%HnmOTeRM)ui>_{WqR!GW}W11wo8VG{mYV||7t7q_or?+8~1&l
z?EPsERK2a=+m!2mt3I)nH!y%P@@CnZFY6ZjKF?&)6VW$&72p16W<rkh<bxvXwz~Y2
zdB*u{Y3NnAPn-VjyL~EdRi&ZS=Zj?`Zf`#*o^aTfdo06fX8ESZcy{q-#puk_T6eu>
z`sTaz7Dp}lz%M9Y6WD)m-rWPKTW<YX8`I`<t+Foei_QDL&wqZ5o0rb|bguEY$^V0V
zuJNzh|NiS@k@<_8yJmff-()#|&)++HFEj@Ky+7&I{}k;E=1uY<3+J#f6j$aZ$pdrb
zCZ@tBS*?VZJPiH;48c6CJV9UrC<6gSscFS`7#SG2el57y>}?v%!V@g!Am$C%7A)={
zj>MC2kU-)o`YA@jR0@V($|zXEvw?{xSkOVx4JOPJEbk{D0^<p;@02QC_v$b+Pq2`K
zkTXn}Cs^1)7>Ot9APVO>w!dNc`K9C*BTKNfva~Nui_7FMF9M5%?*XmjbKpa=PR36L
zZe56<kot>>Hmr<1!MvQj{4l-g=S;VxNp!pb3L_htKXJ{BTglT_0fm+Plwx42VKgHb
zIOKW{{a&!<p|!%H5WAQ0^4E_8RZFo*DZx}-_U0^_VUzR(=q`Q-ek6Cv`pLrG_0)j%
z3p>AfH^^P0yz(%;Up9Q)6y11H6Bq{q4gyHwEaD)7#FO)rgNHg#u!5fgocG&hS19wf
zsmvgA6?pAo<`$*qB!XhQxG1?;8N>(z5n#X^JdH1}`t@p!My9|ZhCrAw14D9dN=^ol
z>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3wY^EWPO;k);
nMpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)5$<5EvgV6^73}cQt

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.2554570633260D87.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.2554570633260D87.idx
new file mode 100755
index 0000000000000000000000000000000000000000..098bb82fe29eb34ab331b244644ed28b3460bdac
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%qnIEhMSxW3>9;ZhUOManG5XuKRGfb
zYr5K*4+qsQ81&|<`5u!ub-K0t@*KZiuULKRYQtnCq<7C<IX82^%&pojyU*|IWLm&b
zkeWFsKw-}s_0Y=y4}C;F6#inLe>WzscHYO;KMP70Rs46Er|++KX8QB}b2Qw<)w`DK
ziJ8Z5e)R0|n^(8rm;CtlyhnfkopV2*78y!TYWnQotr}mq?~m2~zeQ*D*L!{z`jBb$
zwvlypZ1);Ir_Ke}*Az>CiHS(u=E$xVdwZus7+YoYy+w(um-Hsx@|rBN@6f-m)7;kY
z`EYx2`|Zz6?^+DMEWNFJ%R_yO*Yt!VT)rn3eLwg7zUi;rHCxwiDv-A|eZQ_`pXg1G
zXj#PtnlE!?*LYkqW4h}sy=C*>sD0JnAKIoQSuYmmEm_gh8}&{xUGs@@vb&NNhqP=e
zFPHbkGxvHAom(t2F`@MmhrZIyR@)asqKsDWqr~O!JhUzB+2y*4<D$CcTz>YBO^b7)
zYPFm@-rA+id-|g%B~_T?^_y*_nhx*3cSz(4Mm=09@BXy>3*WUv*$cO*giO+?nObE1
zttCp9NmS0O*UfR=Rx!894>|ob-Nd)czWX}OXz`>Jrw<-)cg3(>m|7ZgA?`Zo(!S>c
z3q)?oh{!SjGOfJS<&nAQA-{vqy`XtqQCa+fr)6dNs!gt(3;v&2!g$d!?&{oWkL*_8
z_V)GNwlOa`QF+hVn_su^M+m4M(weGT<>44o!4RadsB%tgXzfw{t2IT+Nw+0u&4}CS
zc=3SMv)q%ex1LYUb_+VNUnt5``N@}ar|np!qV^vyfAUgb@47uNUcFI@^h*0){bu^D
zlX9(VidBjxZvU_@J%oX`ZfA^7`F+XS;?>{s)hbW!X48J0;(Vdec1=r8hE?aBi);HP
zZ)S>Nj9sJ2wR6e2oliJJZ}~@X+1WEK=617A(l*;8F^Bhs=hA<^MfdBpAMC5RTU}NW
z-85b2{hrTvN^iVZyl&>=$eI#5J>qHXBH!ngEayb^&0od0*LWLjTX-^o>w1(@J>O@Z
z&!<*ZDOX;txqt6;TvmmpjID+IiFMi%zAB&nwtaa0pr+th*sV=b<w0VXvoAiFBc}1k
zc?o;`mD%%;XFcGGO8plWzHq|Xikhf9A9sAWmXm+id)(EtvQM)5@qWR{XPaNWul}X6
z_4$hh%a;9_p5v#UV;-*`@-O21e9y1-^Xg`OdBpzx>m-5V%G@M*U`}MP-Tp(i`{hX<
z2LAwtU>;T;FE9a;0fVB{wBkFAKqgcjBNte{_t5VJYaUuFaQ#|vui4u)n1v@;%t6c>
zriv$6+(8_PC*dH0#8dQBjD)Ea484?5u!Ls=6Hl<9gP<Esm?v1?Pd)_36I|aZRl4rg
zVP>9SAqOF6m@rSUu!ArXPt-va&U0*k!|?M<$t^~fU}<G(Uziq`$zNUs775=2TF2+W
zhh&|MpA6hO?{zbeuVZyv%g7U~<fjw^(;MO^r2b-}4J%NXmy?$tCY*lGbW55<#|xk^
zvQ39V>|VypUq22MmST}of~mgj%~>?VCg};#0sId9NDh$olZ89rsR8R3c7E}0kWHez
z@-V$$HhkO^-FQ(G7}x?10!RTW;vj;=lk<~<2QN>sf}aAM_uFMxDD$<c%ph|Wc<o{4
z7NzDSf+D-PD7jb}#0UZrV89%lmw!cAgja7HQ(zE7AWWEnAvrfCCj-d!PV;4EVP#|I
z;N<4v72xOd6m)e~6Xf!A5)u~{k&v{ow6?Levv<%F)wWVL(~!p|Dkd!>tDvZ)qN=W`
erK4|PXl!aOXCkF*<RU8O?iIr1=I7|a=mP-5G@@Ss

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp.D8F2815AC5D0D7D6.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp.D8F2815AC5D0D7D6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..62ef0954b94146ba0c0bc194fd78334888823d35
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%sOTUhPRvy3>9<EhThJTHW#@2pVMti
zR=6+Ai-TcH+%j^wGnwp#L@alE>rB7%Nh$K@FP^QFQcveKTDe`EE}G7l{_ISz76TVU
z__Sl3tcOll1x@`ff6j$betG?en`fV1eJ;6w7QefC*LVLXKR-P=R8#rybZ1gY;i8IH
zNrj(Hwe#%Gt=sq4uRXv1+>cME-_D8G6Q8W8Y4Lwc#HXjvPp|)cTmSl>PqR$)8Rsv%
zEEloi*47iR6sB-goweHTc)nFMO}ZrE($=}$NvRXeU+7A0^SY(6Y^{msL+v;5`I%o%
zaqqXS+F`5zz)!+Ce%IYay(LboOE}%uTxij$dSCM|cJcjVQMa|Yzh4ag{Lbrn^u*pR
zR|T17O*(h&V3$g+@&(&1j@GwVKiyp}f6rNMo3YhJhjy<UE1t>~O>?p{DpOh1wBcgl
zwFMhQPQJOT@y5nWkTXo9_k+@;*c;7yfr1M@uPgm{Q1+hi@fFc3-2%0fF8(ZN(6BzW
z?CNr(mK*P!pGEea_Xsm<+;A>0dOFjWy8e!fVhcs*{VEKpbAO+pyGP8++ex+aP`Kvy
zWd9Xc6$2hFypkdEBq}@6L~`DPDZl`{{V*<H(^$(>jivv}y3?r)v61eqZ%gL}ysF@1
z^jdh<@uA^@_{r*PJEo{f{bM-tMEB{+1xvpPY`t^w;Rbi-w>Dqp-NXfCpM>7c)O@#j
z*WN2vQlwv-iJ2X`Jtx-xcLQsn(wd;iMeQOkt?UY{O0$kfMqCa56L41DTxr|fg`F>_
zN{jHWNxnBNW=U<_&NLO)KiUhYnmGBr-}5GT!NRZBHSY8HF1|k1z3!aeq@^ik`;XVH
zockv5gq6FJcqRAyDmT^(52l|!TJnAFMf2@h|I;Q--;=$;$j^*L_Qqka6K%^TpOC!!
z%ED6nfaryzvldP8el^E?Ucl17ElVGKRyNp`utj-W(&6rdHyh@Kyt=RQ{q1(9ciq?b
zrnhfbtvLDVob~zG_P=sPxy~XX4q>I8q4!j8mCSR05c9EU@xRg&^Er>1nMnFPjPibR
zK5<@RTx4jz%KBOFZ&%;+=GAXkem7+jzaZPD*IaJNC-+R&{bOYzKWF3AZELR0opSU`
zl!?W1U9SHE?+%KazUW=a8{_DG_WqH0jTB?K`PZs!-p%bl-d_E<Ffiy?^0B&4=Q~e6
zbNu>Uw{FV);*x;i>VMv5%bB-JZx=sRuk$ZD=&!vFGjH8T=5GlzzZ6&ICdmVH<C2pa
z>bc?fm3bKa0~msNSb6-x1W*P7ic-^x?=UhjaQ#|vui4u)n1v@;%t6c>t}R&HK^%!E
z;UIy;Q}k1egsBt^y_8X~gl7X2Pq3hapc_n>Cs^K3J_N=ST;C~Gy6)9sW}aXn2O(#e
zFi)_sgD?_L)Ik)^b8LUZ@bgQ_Ek>4LX=Q0&m=>4GUtR<j3Eu-+$LGL@WSxwk4BR@;
zpGKaycH4U}@&qgSDZ!l@;wPm3VxkQzP?(pKmmg+i`Z?1rX%ZbTfWpW&!K4|vz#-Rr
z==Xv(53Lmrh1k7}m%n}-s9K6eN(rXwvNvbZ44b4UKzH#w@FTfP)=w7huBQg9U)cG@
zyFm^R<&}r&{j%ZXrs&3tn!q>^a1cNWXAuVxB%Yj~9Na5B!3ur~aNcj1U7^g^rZR)f
zRp7OUnOl^alL(6K;-chYWgyKQ{D7fvlg*Fkg-n4#3_*~D03rhs%;emZoD3k>JI$Aw
zg_Vt+gOi(wSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!RNG3~OhX=<sF<{jtb(GF
iimJM%mX5xGp|PpCoQagKk&CF9yH^O4o1dcxqYnU>n44_?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp.0FC1DC5FE52A0C60.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp.0FC1DC5FE52A0C60.idx
new file mode 100755
index 0000000000000000000000000000000000000000..de60f391d277ba3cc605023fc82260c7996e5565
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%z9=9hPRvy3>9<EhUVtqHWRq^U-R6S
z={nxLKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2pRl_01rv>LuFvis-d+8!_Z))}
zL%7H>PS!)Gvx3BCe|MgEu>ECxO<8s6-^^p{pD*qfKKfL+a-YpkgZO!W%NFghun)8;
z-MHiLB$4g%Y1iZbFFAaDe|pW&pLyrjPwPI}Au^{vM(5|J%U7rWy!%|buGD$f^oIV+
znfy8tPo+{q1-w}HCasNL{QMBtW?n0Um$R<1PLzIPTGie5bgo%frnc|$52D}JZ#S{}
zeki|R?%g?~9pc9>OwTjVP@a21E4D$<a@GWw_jCRCYk#>Nkr}O)C~Rx`eqX^p!IK`*
zvWg2dU*yO>QOT5Ue7BJI*3CHG`+KYFIXiEhd*Q;YtChHTTJ?!YuAe7vOzhw~6y+Y-
zc*v<F?Y&##>%cCL4NjRY;uGH-k}HuFWsG_+B`$yGVVi~9U4_V&7uJ$<`Pn-*EzXgu
z)pG7QYnL+X>5rb2OlyH_H*+8FV%U3MIcbY)M$7U#nW^_LSVwTKZ@n}rz;n@u)akc2
zs$Y2_u;N2Nh6z{D+N}n&c>D{z40r9wEnL5SisV!k!6x^t>&uK9^wutJ`nD@fqtvFI
zq4mPE1s^0XtUob1%*9KXx1J&Cx$e`I3l@G8*m~vS#TD+(Pi?-+>zz}y@>sp=a?qXh
z((08f=dN8F9UZ})A6#bpmRZcDaYDIkquee9FB^tfg+-ZjT0?7(@?WhfQck)oHETxP
zPPf3rTFY`hU2i>~n&lRBpx$cBB)1bU=T5KXGTBns{{F;e#j4n<i(w~S(=}(_eZMh2
zOTAiTlh?!(uD1^To)^ew^5bZc;oaKwFTQJU?VtHX^{v#CV<%@VV6HyVx+c@AbI!%y
zzR8Jfdd+5GBCRn&&tsl*hTigz5U;6X(K{#kwnO*kiy8x&3zNP6&R(+b@5F;@)px7Q
zDx#aF>%86b`OelG?-j3`xj3?>q~6|mX?LI5+~tB6!FwWB%k%v-K6qo6TC-+MX6GNv
zGs0)QSFhUoB=v9Z?umOtifd20&knS%=`fg(JKIpo``O%_2j36gOgK@L8@V)EbMwpI
z6@`A$BL9@{H0x(wkPepK(KBt<{?7dqPFC62t^HbY$6CIh|NF7WYZmkz@A>;=zVPI)
z&9Cl9|Iyh0{KbMr-~M=+Eoa^?y<PlNz0SYrpuhGy%)E8qWan7TIKog|nVTdJ%#U(&
zu02RuYtYEU;2*#c%)`p#4<>*z5KxqwR(yw%fr0DSf_u&0rok*c!D0?#-f(Tf;tt|S
zJP8L0B%Y$LVgyX3VCbccf+ai~n0SH(9R%HA!aTw9zVg8^o>0NN3*1d_*E8}2ODjwJ
zz=UBmBNy0ly@!4;So6?Y!Lj`f!_O}zw-{MaRJ%<6@*=QE_#V(EJ_kM|n`C@t;5LQ$
z38}xBXu}G$iI<a?A7(?UujAgvjmwgld4h!;gpl+KI|w83L>)xoy!~ujQj}+AZv>j8
z<f{}3GwD!>-OG6S>&Jn@QY=zRFyYJIoJBKilAZt^$M3)oH-slx)>jtpxTglJU)cG@
zyFrc<<&}r&{j%ZXrs&3tn!q>^a1cNWXAuVxB%YkF96Z!{f)#uf;Jn{1yF!_-O=Sj|
ztH5gyGq)%;ClM6e#YM@*${<D%hyVlT;IDkI`WJFcv1AGiVhDr@GcY9QrsQM*x!!5M
z%q*;I>>QljJiG$@e4c`?&T4{Oo=!sI!Xgrq7M9jFws!UodZOA^%4Qn!*hIyoWn>i;
jl~h#KHMMl~4GfJ<&E-s_bd6j@#oWC@nB4pvJs5oe^%<#m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.0151E7B579E72F93.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp.0151E7B579E72F93.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88485d710a33272dd3d80fbf6b300b5080a93007
GIT binary patch
literal 1618
zcmWIYbaV4!V_<Mj@vO*AElFfyU|<jg;^LB`%tmGghOe9q3>9<EhUVtqHWRq^U-R6S
z={nxLKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2pRl_01rv>LuFvis-d+8!_Z))}
zL%7H>PS!)Gvx3BCe|MgEu>ECxO<8s6-^^p{pD*qfKKfL+a-YpkgZO!W%NFghun)8;
z-MHiLB$4g%Y1iZbFFAaDe|pW&pLyrjPwPI}Au^{vM(5|J%U7rWy!%|buGD$f^oIV+
znfy8tPo+{q1-w}HCasNL{QMBtW?n0Um$R<1PLzIPTGie5bgo%frnc|$52D}JZ#S{}
zeki|R?%g?~9pc9>OwTjVP@a21E4D$<a@GWw_jCRCYk#>Nkr}O)C~Rx`eqX^p!IK`*
zvWg2dU*yO>QOT5Ue7BJI*3CHG`+KYFIXiEhd*Q;YtChHTTJ?!YuAe7vOzhw~6y+Y-
zc*v<F?Y&##>%cCL4NjRY;uGH-k}HuFWsG_+B`$yGVVi~9U4_V&7uJ$<`Pn-*EzXgu
z)pG7QYnL+X>5rb2OlyH_H*+8FV%U3MIcbY)M$7U#nW^_LSVwTKZ@n}rz;n@u)akc2
zs$Y2_u;N2Nh6z{D+N}n&c>D{z40r9wEnL5SisV!k!6x^t>&uK9^wutJ`nD@fqtvFI
zq4mPE1s^0XtUob1%*9KXx1J&Cx$e`I3l@G8*m~vS#TD+(Pi?-+>zz}y@>sp=a?qXh
z((08f=dN8F9UZ})A6#bpmRZcDaYDIkquee9FB^tfg+-ZjT0?7(@?WhfQck)oHETxP
zPPf3rTFY`hU2i>~n&lRBpx$cBB)1bU=T5KXGTBns{{F;e#j4n<i(w~S(=}(_eZMh2
zOTAiTlh?!(uD1^To)^ew^5bZc;oaKwFTQJU?VtHX^{v#CV<%@VV6HyVx+c@AbI!%y
zzR8Jfdd+5GBCRn&&tsl*hTigz5U;6X(K{#kwnO*kiy8x&3zNP6&R(+b@5F;@)px7Q
zDx#aF>%86b`OelG?-j3`xj3?>q~6|mX?LI5+~tB6!FwWB%k%v-K6qo6TC-+MX6GNv
zGs0)QSFhUoB=v9Z?umOtifd20&kmH>dbnMJ`Q|ps#RpUFon)(%pTpaxbA4LquBk?|
z?{!E$-{Stzp2_aVBC)-eI~z}DZ2T!*$uo1!@yD0D`j7AZ{qaXZzRfwVQ&nef`2TY?
zjI>|zZ|}u^qnbWGt*nddQ;Mz1b}y`6$awtIe$}i0k9%GCk$#^s{@d&jhT_WHBza(t
zoY~%$CX<pS&BNdyz!1#C$`c4CfHDwJl$ut2hmnDS>(_#N&EBTLEIh$t4r1PLZNcIW
z;z&FR2MHveqOW2EOr>DxrHq0lJR6vJf(0D}-C)8z!ScTH!7!ds!Mh9GO>fsT@&rpO
zOZ&it9oye9{QOdKi;)FI*k$sU7lB2>_kc$7Iq)GFDdQ^xH!{ReNd3h`8&;r^yqvuJ
zFe6ia9rreFT$aSl6D;H)grry4K^TcA>L3c|<xgBQ<5u#tRX~%Je3hbMCc$V%E^x^8
z9{Rmt%|mO2Lm_rA<K?d(2db80ky3)Gy6nwaG{Yw83DC{_4*YPld4gqqW#MjqYQXx1
zonO2g<YrM`d6?cW8$ND|ZoH@oi~|7&0i<viaS%b`$@$8`L!BpB!B+v!`|YwTl=<3J
zW{|lGy!J42i&AqEL9tz2lw7O~Vg!K*FklWA`?D(f3tvt=Q(zE7AWWEnAvrfCCj-d!
zPV;4EVP#|I;N<4v72xOd6m)e~6Xf!A5)u~{k&v{ow6?Levv<%F)wWVL(~!p|Dkd!>
mtDvZ)qN=W`rK4|PXl!aOXCkF*<RU8O?iIr1=I7|a=mP*<l#dMn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.C058D2A1DEE10686.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp.C058D2A1DEE10686.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8bad777c3c2a8328e2ffc0e8ae52d348e634a25f
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%t~ekhO3+m3>9;ZhThJTHW#@2pVMti
z_VrSh7YD<ZxMk#UXENCfiJZKvI?X$;eTm-BUp!kUrJl}fyy5t3)1xyF&&=5w$m+oO
z;Pj=cg$@hWxQFhnt5@RsVeqfpKX3gyd*2^lzkGS*QTWqk-no7~pXcfOWi|EA^^5q&
zp2^#va`bKc$*<S#Uw-)do^yWvzS*Bv|I8`z;yP!4?pe*hm#<ELdH0!j|1QO|-Up6<
z+`*o<=G)q|QUNcP%_(!^GoKvd+|8+G@bdOK_Kh26nM=*xmNT&@^o&L3<V))})bCwt
z$}O_PvSNp|_Jg<x@AzGJv&3I6nQNpPvGIagir~fMf3aD2j@;Uz#r?YD%AVhslC?KY
zEL3y4ux-ia&l?m@?%*i#-<`1M_0>;zcYmKgp)<F)&*Va;YK%z{d!@Ugykxb|)}{?c
zf!7w4XimDjGjNx)uB%E!W+rQwM_D5KFBjJazS`H1o2~b%^<CUDOULPm%Hrc2+BK|C
zEi*lC)N<pU)3eCF^B!S(jRp5|qvkVxspIc3^>fyq@T)MO?(&O7%?li<DhnsMG=`pf
zdqXhGmm{P?qjwuu(c0b(Z9M%2UWUJRy!_(#Uc*?+Q;ntn%D&U748fNEtZz%_2D~cZ
zWAsXV6W3^!5i?1BamN%zsU7S|PVSXj87pmFbj!bd+2Zc}cH%F7@#g}zPeSizX};UM
zYwwjSPo$oQ^Mx_Ro?Ewk@9`!LQ=Op5tQHZMrQaR6Tuk~m7pz+UW8W<QJeRv=)6|xQ
zuiKK5Ao+RQ6Q#GGPMIoezNnv6Hp%V8?6uQlxlFd~IJ*7GOM#0a8oR7^d-P4Z`A_cG
z(rrdrimzolW1i^lc`qsI&>Vj*{qfs!r~7wS-JR_j_i5e1Lsd5gwxsiOD@xDuQ%d*E
ze)M7qL%f4{c9)ah%3?dkg{K~LEN%EHxWgiSi}Jdp<<FZZGup1avOn<MZF$x;-PiY~
z|K6^;Af@uTZPERgb^JRE6}hw&R_{8Z^`+Bn?sGwl;5{)b<@tV2OUQAZd~iyf)XJa5
zGty^-`(4#Ox%%1dG~Mcz51)NH_}!r__F#sBWYx_>@0&Lt<cyX$o^{zXvM1kV@$xP9
z5AB(3e=NH8$0Ue3{({nH^D3^^oVLS{Pw}<yz5TGDKHoye>r~N+8`A%~X5`3){J(d>
zeWtBDyXn>^;hQYw*1s*-8}P&VX1&_g|K(pY&vd^}V4s-4P+XatBoEAqVmXZe^Sb|k
z;9>9&U<l@6<?#d)Kp6-qN=+-i!^ptE^=rYsW^dDA7M@@+2QhEBwqS7waU`CEg9H*!
z(N{47rcyBUQbxfNo()Vq!GaEgZZKh<V0mBpU>HxR;N1o8rnl=Ed4i>trF~$+Fq)AI
z?6}@TzZa}|XszJb{)XY_my%nIEGVj7CVzPmSR{N8XcM0UACgTnzA|u|yw}Y<zK+#x
zEzl+<U!`c64IzF)>MthRumXj7IeGbE!l}NFdmA?{OJe2;7IF|mGE&$<7>Ot9APVOl
z3bA_`FMs_w&?G4qDJ7Umm%TZQX4oV>0XmM~fgi4#Cs@{37VfyG2CQG$`Ng|It`Oyw
zhw1&Y;p3+0#*3Q3I1q3UKniCO2N5KmoUa@_)Omsxd==ol-!8jCnXgS{2AQkCYY#KG
zC^aV$6x+o`$;HYbMi7Vq1Lk1+$I(}AD+kYH3JhWhgb6b+B<H5&WB|F|X}-)XtZeKY
zoZLLT0{ncQg09YLf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=L
dbo31jjZMwvOr&&;Ttvm(y+WAW{2V<PeE@!4u{r<%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp.278F3C311504F9EC.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp.278F3C311504F9EC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c59c78b58e8ccd853f025a90c21c55c2330960cd
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hL@ZS3>9<EhThJTHW#@2pVMti
z_VrSh7YD<ZxMk#UXENCfiCFIT)|r0i(<1MEcFt>4Zf=TJ=o9>sdUWRDnU=+ixf+-b
zOkG;&)c9eOT<FtZe>*iE9R8tiw{!2Ve<lyFU%uR}eDtYsrIr0IgSqqnmMz&~VIO2u
zYPkFFB+>2iYS-icFFAaDzk1ETpLwU&Kihm!Q`6$V#q7^Nm#@~py!*WQpI2=r`i%0&
zAIr_ycx&s4R|-=&s!m#McRJrHnkHP5aB1sY?xe&C<}Y-mwt3vrShm*0^P%>e`25T-
zhq(7!R_w6Weh??&9KY-CqTUjx)g_#6Iu}}Ws@~W9i(Pd8NYrgD?$<9?R(@aldF=_e
zTdz16eLd4(Cpb>pCip_`)`GcjuYS6_`}_0>ow>bzCKnb8nPu&4p0jj9e~)bE3WYT#
zQ%p5<T#c;FxvcwxT$)6^l=pO`h?gz!4q<t*Cp)&LpgsGML}_=5;39t`EBS8%sqQzo
z{EBE3xE*`K=JZF&lc&@KuHE!~9n^5|J)_|*r7eeo?fFjSzmU~wz3!Chxk4rIht|`w
zHym57ShQ@WNcA?wWb5{*9af$b)T5VYb?2pSs+;P`mIo7WmG0zOp#6Hqg1zg6y=2cR
zI%saWscpmd#rx9|(TSH_4!>`lv1h4`=$0#Ti>6suRoz;=ux#Ehe(~o5vQI+qW@*0L
zyld~3D=E^i!^FZ4-JTQcU)|3dsH77Vxu{*lrL|gtRcY4o$cU@qHSuTF&6T#zUD)|@
zs<sHPPV&8JF-v0OcBZMY{!w2z)x^m!fA5>11q;7M*Lct4yZHK4_qua>la{7z+uvTd
za_yVI6ISj@;+5R*tKC>HJedCZX2tKh7tFV3eNUS-eNXlV13xpCw>J+vooHJ&`Gn-%
zR~A;92ShI%owaC!_p3SH^8%Ls?O6KYvvNV<xeWK5qkiYv)tRe;R@N`Rv)!I6LVSJo
z`EU8&A0|DWTRv|$-*4L|x7svTG^}|gw0e*CmX|jD2WCEc>h-ta#C*<UVJ4D34}-j)
zoKKvW7#A73US<8P{MlRIENy%3{cg%6enz%UuescUPwttl`^U~gdd`N{ZELR0opSVt
zn~B9zRj&VnZx4!peV#j$H^x!>?EOQr8Y#vya^h7MZ)*A5e^(z)3=G=me5~%%{LYil
z9KXK*TRY`maY;aMah>Yy$C<a!-QvD#f90R<%76T3jm!4^;PYac>B3N4nVTdJ%#FSm
zzGX%9HaGDw_y;fq^RV*xfeD}t1QeyF72jcGVBq?-;9j%0X)p^<u$Y6GH(Xn=xPv$n
zPr^Y0iKpnR7y(l$7<wtAU<uC#CZ1qH2SGQOFi)_&uY53!Csgq60(aBf^^82h(#q04
zFk#2`Hw-_&l-y!uK@oPD{N+Vpk?=jBk$et(NJh%|%D|2E{AuKQYqz}z&`2d;CAf27
zMl*7O{o8x!_kuMKtrbH2gw$V5v|$CR=H=w&hp9^Sb==#yaaj^GPq2`K5Ry&84#G%0
zQ3p{t?@)-{%Xs<g$AKnEu}CSwOuFpNSv12Y=?T!y{0{tZ)jYwnzOryPKQ&<e!p<+=
z4RVDjuRKiemkl2`MK@m51jd1Yg8))Ei#UiN@#K8v;NIj3R`6AT^M1SR3T3`Fl^JBN
z0<S&H+@jQ+L{MxO7bO=fgBU>|0t}dgz1PgC-l{X{Hd9~_Lm*6;fgw3JB_{*O^-l9;
zW?^Mx=iub#;T7QL^AvP-Rukm%bP^I57Lky&u(Y<ZwX=876V<j-Hq(&DCMqT^BdegO
jq@t>>simWDU}$V=E@vX8Yvdv-=I#~3<mTt-!RP}3v1_o~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp.8DE51E7D513808A9.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp.8DE51E7D513808A9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..54811546169da0b4913fb06d00c144ce5db5166e
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%m!u#hL4;K3>9<EhUVtqHWRq^U-R6S
z>1VQ8e>gw7wR)qP;V!YC9v7Z?D!ELQ^8C4U|33HAeZuO}8%!j=xjwslcz5-?-g68_
z4B;ZjI9U&!&I%H{&F?(%;PIF8HD%SMe=Cozf4=zR%twEPEBD##Hi(b=Q?yJXexByM
zT@iQce7kP<Z+@M>|MJ7z^NsiYDKdW^e)?#N3)eY&-Dx#Hf4+MC=iO=E{ks&;dLLN+
zatpiInj+p!S{$lPahtE@UMNm%y~!#&qr~@hQ-tIb)2i;awN|+zm&0cHJ=A^^pP%{V
z6!(7X${o8;JcyHUj^A~6L2rrE>Jmn`H5Xb`s@~W9id}sFsMl>H&f70mR(@~&-0R_X
z>lFv1uVeb_1lLL11YgMAS}^zR)lYYK^S@UVye&O9bAhXB&XrFLmfni;lGZ|jjyGOt
zXa_`eoRhhukz?y6$Qh>5`$1_^tVP4R#YzsZV;?_iw!R~De1)`1w?M7>;?EB`0?VfO
zE<T^ivC*={XS2-wi8{ILhpcvI$A~ZZUhUMgjidL&mHx?}tgG0wA8rkZ64LS%+4m;M
zny*DlmPu5OtJlqO)mAaL$qyy{G~L9@b>+THGg>@J#p#2`+gl-Q7nYWWT!_2Qwbbvq
zzyg6=GNN+Kzf3DHb$Mhidf4yqX3tU^QIjiji>6swNqx;&F=zR|!#{);HE&r}t$VaK
zZ2R<Cw{G6r)!W+p!7TOqWxFIt5suaCG%h%cyL7rQ;C3<T-MnGd`X3u-`RBRhL^%sz
z);8D3ia5SSccn-Bo}jZHTtC9w!!B`H?)_fm(b#tF{YTqME0^`&`nR-pzR*$K{NMJg
z_uJHdr!Y;SogUE@_l-|Au<eh~TlV{X&%TwR?dFsBEQxKNTDVE=#R<7Er)ws2UCzC5
zJATrDdAg&tc9)ah%HnmOTeRM)ui>_{WqR!GW}W11wo8VG{mYV||7t7q_or?+8~1&l
z?EPsERK2a=+m!2mt3I)nH!y%P@@CnZFY6ZjKF?&)6VW$&72p16W<rkh<bxvXwz~Y2
zdB*uHEOd6zr%nI%-98n!s?t#E^To0&eDY2_DYw&lo=vJ$m;O;&U?{O^>b5nP)=s(k
z#qEm4a#ODV0`Cusm%doLlQ+gK`|SN=i6>I#K7ao7+~<4M^W@(@tqcn|XPIpO^Z${Q
zzYBi-*0cBeU}?2th2wwK*`H_L_P@jZ)jssU?#h4q-x|~Ge=~Zy&MYae%uSL9=0~sH
zdtW6r^+oY8_y;fq^RV&+fC-=s1QeyF72jcGVBq?-;9j%0X)p^<u$Y6GH(Xn=xPv$n
zPr^Y0iKpnN7ztA;xV}@Wblt1N%sjzD4noc_VV+=N2Vo?hsDmh+r_ko?skHWK9urS6
zp97y8LW_)_ObCqU*#3s$=a-UOj4Z*@%F@0tVHnNG1@=$xq2CMEJhWB_@e@*iG0}#V
zktdj!lb0W+D*c@4mNbcu7eHZTL-w<6Nl~7ey%8v^<fjw`Q++7J?q$6E_2WQcDHbUu
znDAw9&Y~GMNl$>z<#*sma;~hOEZn&df8?$G?GP#jR3zvih@?o~Padx5sR8R3c7E}0
zkSj%b<zW_l+3;~wbmK)$U>pcI2q1;Ch=T|cPtH#c9_l>73VsT3-fx#(q0HB&GK0)j
z;I)UDTa=oU2#W3EqU2&_Ak7>cX3Ms=WT*WarobSEAV@+0k%0(ia&Aga29WEW=F7~&
z%Er#Y$<4znz|ZF?=<2K{$mQuIBrYr>A!%W0ZDVU^@1Q5DZKZ6cA&*T|Oj<@(K~YIX
hRb5j{N8iBE*wkFkL`v7lMO4h)D}>3-&(VX?2LMk~tw;a>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.B3BD1E3E3F6E4877.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp.B3BD1E3E3F6E4877.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8d46b8ee94b618ea5936163cbeb9c2679fa5977b
GIT binary patch
literal 1618
zcmWIYbaV4!V_<Mj@vO*AElFfyU|<jg;^LB`%tmGghM$}a3>9<EhUVtqHWRq^U-R6S
z>1VQ8e>gw7wR)qP;V!YC9v7Z?D!ELQ^8C4U|33HAeZuO}8%!j=xjwslcz5-?-g68_
z4B;ZjI9U&!&I%H{&F?(%;PIF8HD%SMe=Cozf4=zR%twEPEBD##Hi(b=Q?yJXexByM
zT@iQce7kP<Z+@M>|MJ7z^NsiYDKdW^e)?#N3)eY&-Dx#Hf4+MC=iO=E{ks&;dLLN+
zatpiInj+p!S{$lPahtE@UMNm%y~!#&qr~@hQ-tIb)2i;awN|+zm&0cHJ=A^^pP%{V
z6!(7X${o8;JcyHUj^A~6L2rrE>Jmn`H5Xb`s@~W9id}sFsMl>H&f70mR(@~&-0R_X
z>lFv1uVeb_1lLL11YgMAS}^zR)lYYK^S@UVye&O9bAhXB&XrFLmfni;lGZ|jjyGOt
zXa_`eoRhhukz?y6$Qh>5`$1_^tVP4R#YzsZV;?_iw!R~De1)`1w?M7>;?EB`0?VfO
zE<T^ivC*={XS2-wi8{ILhpcvI$A~ZZUhUMgjidL&mHx?}tgG0wA8rkZ64LS%+4m;M
zny*DlmPu5OtJlqO)mAaL$qyy{G~L9@b>+THGg>@J#p#2`+gl-Q7nYWWT!_2Qwbbvq
zzyg6=GNN+Kzf3DHb$Mhidf4yqX3tU^QIjiji>6swNqx;&F=zR|!#{);HE&r}t$VaK
zZ2R<Cw{G6r)!W+p!7TOqWxFIt5suaCG%h%cyL7rQ;C3<T-MnGd`X3u-`RBRhL^%sz
z);8D3ia5SSccn-Bo}jZHTtC9w!!B`H?)_fm(b#tF{YTqME0^`&`nR-pzR*$K{NMJg
z_uJHdr!Y;SogUE@_l-|Au<eh~TlV{X&%TwR?dFsBEQxKNTDVE=#R<7Er)ws2UCzC5
zJATrDdAg&tc9)ah%HnmOTeRM)ui>_{WqR!GW}W11wo8VG{mYV||7t7q_or?+8~1&l
z?EPsERK2a=+m!2mt3I)nH!y%P@@CnZFY6ZjKF?&)6VW$&72p16W<rkh<bxvXwz~Y2
zdB*uHEOd6zr%nI%-98n!s?t#E^Tjd|x3?b@PdIGLJ(giKvwTxyJiB<aVsz$dt-D?`
zee+#<i=&o&;1`sy3GBZ&@9u%rEw}!xjcN0_R#_MK#peCr=RZHj%}Zx}I@kEy<o`iF
z*Z5cMfB$u{$o$34U9-N#Z?c@f=kJ}p7aD{A-k<d9e~NYn^Co#w&p9j%#g(~9^1vLa
z7kKJ-ly`zA4}*UILog33PY{>@%0NI-YFhCfMg|71UkmOvdz%Kc@C1uFh<U@c1&cd~
zBk?30B#?NDeu|MWm4fR#rApVmI?T)yEaV{M3=`%F7IqLu;)yzl!g&g9&YntZpXM>~
z1oJuYxgoU3_{oI8crcog3+(XTL%$cSd1$TR*#3s$=a-UOj4Z*@%F@0tRUv*t>MthR
zurl%l^K$a?!-Uh%nQlpw=y(AXMpm6aam|cd$<tN=g_Zo2VqmHdh1k7}m%n}-C@jSy
zr34ec?9Ev;!zSqo(7F5${7BA~^^=7=_u-GcwZ9!grGSbA9R!gS$@|H}6+Jaz{ld;K
z-VJi4D6c%sf-f6BZi;Tas0oY%0S5u3a29b8LE_2z$-zUNCs@Hx0nYpFvMZGN+Eiwc
zxeC1YFmsDia}q(ZU0jr0tPElVfe0{Q4tDLFlV9d2yOSv}h#?Rr%)pSGo05|O<a($1
zGPAI<v2$>8^Y9Ar^LYxoI;#nCc{&M+3yVldT3A}!*xK1U=!t4uDVu4?V-ppVmXTFZ
kR8mn@*VNL{H!w6dHJ3Ay(lv4s6?69rVRG|x^kDP>0An?i7ytkO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3955AFFBF9D5A2EF.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3955AFFBF9D5A2EF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc6f51d6da540c775391534da6c16fcb7bd281fa
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%qnIEhMSxW3>9;ZhUOManG5XuKRGfb
zYr5K*4+qsQ81&|<`5u!ub-K0t@*KZiuULKRYQtnCq<7C<IX82^%&pojyU*|IWLm&b
zkeWFsKw-}s_0Y=y4}C;F6#inLe>WzscHYO;KMP70Rs46Er|++KX8QB}b2Qw<)w`DK
ziJ8Z5e)R0|n^(8rm;CtlyhnfkopV2*78y!TYWnQotr}mq?~m2~zeQ*D*L!{z`jBb$
zwvlypZ1);Ir_Ke}*Az>CiHS(u=E$xVdwZus7+YoYy+w(um-Hsx@|rBN@6f-m)7;kY
z`EYx2`|Zz6?^+DMEWNFJ%R_yO*Yt!VT)rn3eLwg7zUi;rHCxwiDv-A|eZQ_`pXg1G
zXj#PtnlE!?*LYkqW4h}sy=C*>sD0JnAKIoQSuYmmEm_gh8}&{xUGs@@vb&NNhqP=e
zFPHbkGxvHAom(t2F`@MmhrZIyR@)asqKsDWqr~O!JhUzB+2y*4<D$CcTz>YBO^b7)
zYPFm@-rA+id-|g%B~_T?^_y*_nhx*3cSz(4Mm=09@BXy>3*WUv*$cO*giO+?nObE1
zttCp9NmS0O*UfR=Rx!894>|ob-Nd)czWX}OXz`>Jrw<-)cg3(>m|7ZgA?`Zo(!S>c
z3q)?oh{!SjGOfJS<&nAQA-{vqy`XtqQCa+fr)6dNs!gt(3;v&2!g$d!?&{oWkL*_8
z_V)GNwlOa`QF+hVn_su^M+m4M(weGT<>44o!4RadsB%tgXzfw{t2IT+Nw+0u&4}CS
zc=3SMv)q%ex1LYUb_+VNUnt5``N@}ar|np!qV^vyfAUgb@47uNUcFI@^h*0){bu^D
zlX9(VidBjxZvU_@J%oX`ZfA^7`F+XS;?>{s)hbW!X48J0;(Vdec1=r8hE?aBi);HP
zZ)S>Nj9sJ2wR6e2oliJJZ}~@X+1WEK=617A(l*;8F^Bhs=hA<^MfdBpAMC5RTU}NW
z-85b2{hrTvN^iVZyl&>=$eI#5J>qHXBH!ngEayb^&0od0*LWLjTX-^o>w1(@J>O@Z
z;?=KwRV%O7+`o4^E~~;)#@0gq#5!#WUzN{(+djO0P*ZR$?AE5J@*uIx*%u$o5!3kN
zyo5de%Ix{avmS6orTz;GUpV1xMNQP5k2}6w%gMj%J?`pR*(X{3c)#G}v(2yGSO3!3
z`uxR$Wy}6d&+${wF^|^|`4{nhzUSBad3CeCJYxSbcalJHWp0u@FefgM-}q&M*0k?D
z4E_NO!91)yUSI+!0|7;;X~lOK85p>JEx6b0Z5qtN6D;N+<_*^tEbbtV#FKE4K;kL-
zDMrFn3a;;zDqZ*LFf&iEkb{sjOqeHF*g+VHC+Z*y=P9%~dn&Dcn#aTw%;&)8hR`D8
zCldnWIkvxH`1z&e79&frw6e4>OxSzf%;W1=-PSVl1S|O|#lVC^{DjnBOtfJI3iERE
z^23DF&zWvXljwK>6h<}#Ce6qN4wT+QzZa}|XsvK4#O`Ih{Pp8N)lw`{N-$NIy*Z0!
z*d#pxI+x#pAIZ70ezI`qKKzll_P0Z*6i|_%gCLS3c|Uo$qNfI|U)cG@yFsoL<&}q7
z@MXitP0@`PHGy#;;2?k$&LR#XNIW?|Ie4h^1S|L{z<Ix2c7-xuo5~C_SAo|aW^Pex
zP9i9_i;I$rl|hUk5CI0v!DU>3ChcVY!oU<5#1IG*W?)FpP07gsa=p`hnORub*f}`4
zd3Xi*`8)+(oz(=nJe`EZg+(MJEiA2VZ0+nF^hC9-l+85cv5AUF%g8DyDygWdYijA}
c8yFg!n#-9;=^D9+in)7*FuD0TdNBF`062lH&;S4c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp.C97D071C6295D5F3.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp.C97D071C6295D5F3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ea24c856958788adf953739f34b8fbc33692bd47
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%sOTUhPRvy3>9<EhThJTHW#@2pVMti
zR=6+Ai-TcH+%j^wGnwp#L@alE>rB7%Nh$K@FP^QFQcveKTDe`EE}G7l{_ISz76TVU
z__Sl3tcOll1x@`ff6j$betG?en`fV1eJ;6w7QefC*LVLXKR-P=R8#rybZ1gY;i8IH
zNrj(Hwe#%Gt=sq4uRXv1+>cME-_D8G6Q8W8Y4Lwc#HXjvPp|)cTmSl>PqR$)8Rsv%
zEEloi*47iR6sB-goweHTc)nFMO}ZrE($=}$NvRXeU+7A0^SY(6Y^{msL+v;5`I%o%
zaqqXS+F`5zz)!+Ce%IYay(LboOE}%uTxij$dSCM|cJcjVQMa|Yzh4ag{Lbrn^u*pR
zR|T17O*(h&V3$g+@&(&1j@GwVKiyp}f6rNMo3YhJhjy<UE1t>~O>?p{DpOh1wBcgl
zwFMhQPQJOT@y5nWkTXo9_k+@;*c;7yfr1M@uPgm{Q1+hi@fFc3-2%0fF8(ZN(6BzW
z?CNr(mK*P!pGEea_Xsm<+;A>0dOFjWy8e!fVhcs*{VEKpbAO+pyGP8++ex+aP`Kvy
zWd9Xc6$2hFypkdEBq}@6L~`DPDZl`{{V*<H(^$(>jivv}y3?r)v61eqZ%gL}ysF@1
z^jdh<@uA^@_{r*PJEo{f{bM-tMEB{+1xvpPY`t^w;Rbi-w>Dqp-NXfCpM>7c)O@#j
z*WN2vQlwv-iJ2X`Jtx-xcLQsn(wd;iMeQOkt?UY{O0$kfMqCa56L41DTxr|fg`F>_
zN{jHWNxnBNW=U<_&NLO)KiUhYnmGBr-}5GT!NRZBHSY8HF1|k1z3!aeq@^ik`;XVH
zockv5gq6FJcqRAyDmT^(52l|!TJnAFMf2@h|I;Q--;=$;$j^*L_Qqka6K%^TpOC!!
z%ED6nfaryzvldP8el^E?Ucl17ElVGKRyNp`utj-W(&6rdHyh@Kyt=RQ{q1(9ciq?b
zrnhfbtvLDVob~zG_P=sPxy~XX4q>I8q4!j8mCSR05c9EU@xRg&^Er>1nMnFPjPibR
zK5<@RUg+MHD(h#xzg>OPn^(VG`Q4OB{DN$oUURu6pWHK9_m7o@{G5$bx2?H0cgoQ(
zQ6?74b-DfvygMjv`l5FwZ;Ye&+51Q0HByY_=3lF_c{jKJczgBZ!oZ+o$;awGo$oyP
z%<=1Y-MT6Fi%SB6tN(eMEoa^?y<PlNz0SYrpuhGy%)E6UnZI+){8C(*n<Nj+jrnqF
zSM>S%KJhU42QUQlu=4nW37`xF6s4vW-(h56;QF=TUbDApFbhwxn1h%%TwAcXgE$gT
z!a)Lwr|72`2~#P!zEi4n-K)dQJi$T^Le4N@o?u}IVI-cYgD9M*(B|x^wDxHp6HhRo
z1D_j0i;SO42#n|0{)XY_my%nIEWy&s(!MZZ&!0w~w|3imF!BT|`6<C&5#lGL{$ipH
zD^Qr1lb0W+I{lpKmNbcu7eHZTLtxU3T;M?IJ@k9Qnupd3heGUL#>-zn4pc40BBcaV
zb=jM<XogMF6QFbX9r%%)E9)l<ckaU<d24?=gh~Mw2|5TODU$b-hbww&!1{%qU%VUS
zN>N^Um<3-peB2b>cu^A=2LcWPNZ~BvAcDk`^OJ-7mnT@kPXW&R?XoMB`Px)wkhu!H
z_AqmcQgaePv0Yr0T&xVFnS+<_ki6S-!~X|UU=TwPBq4yvKm;>6Hzg+n$n{S1WoBVz
zW9Q)H=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*
guBoM?Z(wL_YA$CYrEBCOD(3DL!sO=X=)vd%0A_iW=Kufz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp.E621A4FA60DE216B.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp.E621A4FA60DE216B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6f7643f8d8c84f39a7aa4c7fc6f516ae99fb2066
GIT binary patch
literal 1616
zcmWIYbaV4yV_<Mj@vO*AElFfyU|<jg;^LB`%z9=9hPRvy3>9<EhUVtqHWRq^U-R6S
z={nxLKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2pRl_01rv>LuFvis-d+8!_Z))}
zL%7H>PS!)Gvx3BCe|MgEu>ECxO<8s6-^^p{pD*qfKKfL+a-YpkgZO!W%NFghun)8;
z-MHiLB$4g%Y1iZbFFAaDe|pW&pLyrjPwPI}Au^{vM(5|J%U7rWy!%|buGD$f^oIV+
znfy8tPo+{q1-w}HCasNL{QMBtW?n0Um$R<1PLzIPTGie5bgo%frnc|$52D}JZ#S{}
zeki|R?%g?~9pc9>OwTjVP@a21E4D$<a@GWw_jCRCYk#>Nkr}O)C~Rx`eqX^p!IK`*
zvWg2dU*yO>QOT5Ue7BJI*3CHG`+KYFIXiEhd*Q;YtChHTTJ?!YuAe7vOzhw~6y+Y-
zc*v<F?Y&##>%cCL4NjRY;uGH-k}HuFWsG_+B`$yGVVi~9U4_V&7uJ$<`Pn-*EzXgu
z)pG7QYnL+X>5rb2OlyH_H*+8FV%U3MIcbY)M$7U#nW^_LSVwTKZ@n}rz;n@u)akc2
zs$Y2_u;N2Nh6z{D+N}n&c>D{z40r9wEnL5SisV!k!6x^t>&uK9^wutJ`nD@fqtvFI
zq4mPE1s^0XtUob1%*9KXx1J&Cx$e`I3l@G8*m~vS#TD+(Pi?-+>zz}y@>sp=a?qXh
z((08f=dN8F9UZ})A6#bpmRZcDaYDIkquee9FB^tfg+-ZjT0?7(@?WhfQck)oHETxP
zPPf3rTFY`hU2i>~n&lRBpx$cBB)1bU=T5KXGTBns{{F;e#j4n<i(w~S(=}(_eZMh2
zOTAiTlh?!(uD1^To)^ew^5bZc;oaKwFTQJU?VtHX^{v#CV<%@VV6HyVx+c@AbI!%y
zzR8Jfdd+5GBCRn&&tsl*hTigz5U;6X(K{#kwnO*kiy8x&3zNP6&R(+b@5F;@)px7Q
zDx#aF>%86b`OelG?-j3`xj3?>q~6|mX?LI5+~tB6!FwWB%k%v-K6qo6TC-+MX6GNv
zGs0)quDZ7LN$TI+-4pkQ6xW_~pB-pj(_t_ncebIF_p`Y<556C~nQ)>gH*#sT=H{2Z
zD+>LhMgA$@Y1YrWARR2dqi5Qz{hj+KoUF33Tl=-*j<tM0|Mz2$*DUBc-t+g#eBsGo
zn_u0J{-d$~`HKaMzWwnsTh6>)db{|kdYymKL4WOan0f2I$<BE+;|N1>Wp0u@Fh6e6
z)oMtQ^-Seq@DE@J=3(XW2NOUU2q;QTE55_Xz`*ru!M$d0(_j{!U@-?VZ@9K#aR+fE
zo`i!05>L@rF#@JisNmfN?xwfv8F_-Gm8E@P!U}E9o=R(<<}vXE^EvRj!Gw8&Wqf6V
zVLZq7Hw-_&l-y!uLD3uHC#3#jq75t1NM24}ewgZ1U&p<T8<!<9^8^by2q76M>>!N9
z6Lk=U^I%#Txxl{fJ@k9Qnupd3``NanD9_B^2(&=SS1A&v>QIQ?%Xs<g$AQ99EK*7^
z;mh8fMKf%Yo&deV@4yc?n<rS-R~GJ-hd=Vx{&onJ0xA-85JXZW?<)^i^wfa$3p>Af
zH^^h6yz(#$zHIooDZ25ZCNK^J90ZWUS;RpEi6`eP2M=|gU<F?VIPbU1u2AM{Q<*{L
zD)8FF%q>dINd(1qaZz%yGKdibBEWz-IOjAo^XnewSf;=rhCrAw14D9dN=^ol>z(Gy
z%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3wY^EWPO;k);Mpi*l
jNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)5$<5EvgV6^7<t~GD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.06F3665DC4884ADD.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp.06F3665DC4884ADD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa7a7231bdaf23f43df8b2def1cdf7393bbfecdc
GIT binary patch
literal 1618
zcmWIYbaV4!V_<Mj@vO*AElFfyU|<jg;^LB`%tmGghOe9q3>9<EhUVtqHWRq^U-R6S
z={nxLKb#-kTD?)taF^ImkANqhN-h(nJby0Tzt8=2pRl_01rv>LuFvis-d+8!_Z))}
zL%7H>PS!)Gvx3BCe|MgEu>ECxO<8s6-^^p{pD*qfKKfL+a-YpkgZO!W%NFghun)8;
z-MHiLB$4g%Y1iZbFFAaDe|pW&pLyrjPwPI}Au^{vM(5|J%U7rWy!%|buGD$f^oIV+
znfy8tPo+{q1-w}HCasNL{QMBtW?n0Um$R<1PLzIPTGie5bgo%frnc|$52D}JZ#S{}
zeki|R?%g?~9pc9>OwTjVP@a21E4D$<a@GWw_jCRCYk#>Nkr}O)C~Rx`eqX^p!IK`*
zvWg2dU*yO>QOT5Ue7BJI*3CHG`+KYFIXiEhd*Q;YtChHTTJ?!YuAe7vOzhw~6y+Y-
zc*v<F?Y&##>%cCL4NjRY;uGH-k}HuFWsG_+B`$yGVVi~9U4_V&7uJ$<`Pn-*EzXgu
z)pG7QYnL+X>5rb2OlyH_H*+8FV%U3MIcbY)M$7U#nW^_LSVwTKZ@n}rz;n@u)akc2
zs$Y2_u;N2Nh6z{D+N}n&c>D{z40r9wEnL5SisV!k!6x^t>&uK9^wutJ`nD@fqtvFI
zq4mPE1s^0XtUob1%*9KXx1J&Cx$e`I3l@G8*m~vS#TD+(Pi?-+>zz}y@>sp=a?qXh
z((08f=dN8F9UZ})A6#bpmRZcDaYDIkquee9FB^tfg+-ZjT0?7(@?WhfQck)oHETxP
zPPf3rTFY`hU2i>~n&lRBpx$cBB)1bU=T5KXGTBns{{F;e#j4n<i(w~S(=}(_eZMh2
zOTAiTlh?!(uD1^To)^ew^5bZc;oaKwFTQJU?VtHX^{v#CV<%@VV6HyVx+c@AbI!%y
zzR8Jfdd+5GBCRn&&tsl*hTigz5U;6X(K{#kwnO*kiy8x&3zNP6&R(+b@5F;@)px7Q
zDx#aF>%86b`OelG?-j3`xj3?>q~6|mX?LI5+~tB6!FwWB%k%v-K6qo6TC-+MX6GNv
zGs0)quDZ7LN$TI+-4pkQ6xW_~pB*T%^>Dic^UZCNiw~yUJIPijKZmzX=lZnJT~m!_
z-|LWizQz5aJ(Jy!MPhp`cQ&5R*!WYrl4s_c<Bu<Q^&j8+`{R#-e4BGzr>f4}@c-v(
z7-_%a-`<P;Mm2qWT3Hv@rxaV2?Os^Dkn#AZ{i;|0ANRWOBmF*O!o1lb48@hXN%FuP
zxxsi!&H9%%_B;&!0Sv)BtUQ5W0w@CkMX71UcNiHMxPC3T*X(T?%)%2a<{;(`*A^`9
zAdbY7aF9UaDf%i#z*Gtqyt}~N^maWXPq4JIv=2;Jq0QM-Y3<WICZ1qE2R=8LFi)_I
zuS_tE=h*&+;pdl<TZ}9ydPDq#)L%@rVFen=%gM_RQ=RJTxVLfRvLt4nU?B$~BqN0#
zgpqin4x(^g{=_viZY57!1vE*?S1B515{zc#0tZU(q2CMEJhWCg6k_)>UjF)VplT@=
zDJ7Vy%if$tGi;Kc0KLNRzz;W@Cs@{37VedYKl0Z8b_kUMDiU-ML{cR0D-T!n)PVI1
zJHL20$YY|s@-Pd&Z1}h-y78hWFb)J91dzg6#6bj!C+8~%4|Sek1z!a?@3+gYQ08k>
znL*|%@Y=)7ElSNv1jTl7QF5^|h!F%Lz<@dUnTqBg?vJUxOo2fRfiPhPhUDCooD3k>
zJI$Awg_Vt+gOi(wSAd_-Q_$5}O_0meNl08+L_*TS(%Qz>&fY;!RNG3~OhX=<sF<{j
mtb(GFimJM%mX5xGp|PpCoQagKk&CF9yH^O4o1dcxqYnV)7^nFF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.8A48A77FE8BA877A.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp.8A48A77FE8BA877A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2ae0b599b1e17b7fa2cb5ecd1e669f2a10b5bcc1
GIT binary patch
literal 1610
zcmWIYbaQiJV_<Mj@vO*AElFfyU|<jg;^LB`%t~ekhO3+m3>9;ZhThJTHW#@2pVMti
z_VrSh7YD<ZxMk#UXENCfiJZKvI?X$;eTm-BUp!kUrJl}fyy5t3)1xyF&&=5w$m+oO
z;Pj=cg$@hWxQFhnt5@RsVeqfpKX3gyd*2^lzkGS*QTWqk-no7~pXcfOWi|EA^^5q&
zp2^#va`bKc$*<S#Uw-)do^yWvzS*Bv|I8`z;yP!4?pe*hm#<ELdH0!j|1QO|-Up6<
z+`*o<=G)q|QUNcP%_(!^GoKvd+|8+G@bdOK_Kh26nM=*xmNT&@^o&L3<V))})bCwt
z$}O_PvSNp|_Jg<x@AzGJv&3I6nQNpPvGIagir~fMf3aD2j@;Uz#r?YD%AVhslC?KY
zEL3y4ux-ia&l?m@?%*i#-<`1M_0>;zcYmKgp)<F)&*Va;YK%z{d!@Ugykxb|)}{?c
zf!7w4XimDjGjNx)uB%E!W+rQwM_D5KFBjJazS`H1o2~b%^<CUDOULPm%Hrc2+BK|C
zEi*lC)N<pU)3eCF^B!S(jRp5|qvkVxspIc3^>fyq@T)MO?(&O7%?li<DhnsMG=`pf
zdqXhGmm{P?qjwuu(c0b(Z9M%2UWUJRy!_(#Uc*?+Q;ntn%D&U748fNEtZz%_2D~cZ
zWAsXV6W3^!5i?1BamN%zsU7S|PVSXj87pmFbj!bd+2Zc}cH%F7@#g}zPeSizX};UM
zYwwjSPo$oQ^Mx_Ro?Ewk@9`!LQ=Op5tQHZMrQaR6Tuk~m7pz+UW8W<QJeRv=)6|xQ
zuiKK5Ao+RQ6Q#GGPMIoezNnv6Hp%V8?6uQlxlFd~IJ*7GOM#0a8oR7^d-P4Z`A_cG
z(rrdrimzolW1i^lc`qsI&>Vj*{qfs!r~7wS-JR_j_i5e1Lsd5gwxsiOD@xDuQ%d*E
ze)M7qL%f4{c9)ah%3?dkg{K~LEN%EHxWgiSi}Jdp<<FZZGup1avOn<MZF$x;-PiY~
z|K6^;Af@uTZPERgb^JRE6}hw&R_{8Z^`+Bn?sGwl;5{)b<@tV2OUQAZd~iyf)XJa5
zGty^RhpyK;x%%1dG~Mcz51)NH_}!r__F#sBWYx_>@0&Lt<cyX$o^{zXvM1kV@$xP9
z5AB(3e=NH8$0Ue3{({nH^D3^^oVLS{Pw}<yz5TGDKHoye>r~N+8`A%~X5`3){J(d>
zeWtBDyXn>^;hQYw*1s*-8}P&VX1&_g|K(pY&vd^}SU)j=p|~<PNgkLJ9d<r{6!Lki
z2M>dP07EbjD~~6b0Lnl>QEFQ89YzKQu3ro8HG7)|v+x9qIf!|~wFQeih$HbN93+r<
zioS{wFqJ|D?=EmRy<N}96D+MP?E@25Xmj>dTKhDQi6@xPfzJ&l%o8l*D-#Uk!DvP<
zu;Y3U{a&!<p|yfz`x}OzUrKH<vY;60y>905b*ye{fi@}mDn-LIhxiGpznEyl3KZt$
z<mHD6r}{eXZQQsliJ2!@$Uz9n0m2T#NIX#oQ8@2Vh~3M0`Rm7lCP}eKDZxy-?9Ev;
z!zSqo&@227{BYGg!Lq)xaIZZ4k+=4@L#Px`k)VShk|KFudAOpd2CQG$`Ng|IP8H>q
zhgtAt!^ch0jTbe6aUkFzfE3Om4kAcAIbS(=sPhCX_$t78zg>2PGGCj@3^G@N*B)kW
zQEE;iD7K4>l8co=j35vJ2F$@3%3r+&H)(b=1qLw$!h{(Zl5<mXGJss~G+$;GRyKAH
zPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fTdj~yHZ7XFn4S8&$V$w3Q3W`c9s_L3r
dI{F5N#-`?SCQ`aaE}~-YULj0wevTfDJ^)0}z@PvC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp.85809BA7EA818954.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp.85809BA7EA818954.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c72734627f57cea8b7c71694305fa6487375dc8f
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%vxp!hL@ZS3>9<EhThJTHW#@2pVMti
z_VrSh7YD<ZxMk#UXENCfiCFIT)|r0i(<1MEcFt>4Zf=TJ=o9>sdUWRDnU=+ixf+-b
zOkG;&)c9eOT<FtZe>*iE9R8tiw{!2Ve<lyFU%uR}eDtYsrIr0IgSqqnmMz&~VIO2u
zYPkFFB+>2iYS-icFFAaDzk1ETpLwU&Kihm!Q`6$V#q7^Nm#@~py!*WQpI2=r`i%0&
zAIr_ycx&s4R|-=&s!m#McRJrHnkHP5aB1sY?xe&C<}Y-mwt3vrShm*0^P%>e`25T-
zhq(7!R_w6Weh??&9KY-CqTUjx)g_#6Iu}}Ws@~W9i(Pd8NYrgD?$<9?R(@aldF=_e
zTdz16eLd4(Cpb>pCip_`)`GcjuYS6_`}_0>ow>bzCKnb8nPu&4p0jj9e~)bE3WYT#
zQ%p5<T#c;FxvcwxT$)6^l=pO`h?gz!4q<t*Cp)&LpgsGML}_=5;39t`EBS8%sqQzo
z{EBE3xE*`K=JZF&lc&@KuHE!~9n^5|J)_|*r7eeo?fFjSzmU~wz3!Chxk4rIht|`w
zHym57ShQ@WNcA?wWb5{*9af$b)T5VYb?2pSs+;P`mIo7WmG0zOp#6Hqg1zg6y=2cR
zI%saWscpmd#rx9|(TSH_4!>`lv1h4`=$0#Ti>6suRoz;=ux#Ehe(~o5vQI+qW@*0L
zyld~3D=E^i!^FZ4-JTQcU)|3dsH77Vxu{*lrL|gtRcY4o$cU@qHSuTF&6T#zUD)|@
zs<sHPPV&8JF-v0OcBZMY{!w2z)x^m!fA5>11q;7M*Lct4yZHK4_qua>la{7z+uvTd
za_yVI6ISj@;+5R*tKC>HJedCZX2tKh7tFV3eNUS-eNXlV13xpCw>J+vooHJ&`Gn-%
zR~A;92ShI%owaC!_p3SH^8%Ls?O6KYvvNV<xeWK5qkiYv)tRe;R@N`Rv)!I6LVSJo
z`EU8&A0|DWTRv|$-*4L|x7svTG^}|gw0e*CmX|jD2WCEc>h-ta#C*<UVJ4D34}-j)
zoKKvWm>0TtmCE{A`LnmaS=#p6``wgD{ETdyUURtxpWHK9_m7>0^qdW?+tys0JLTvP
zHxrAcs$BmC-yRhI`aE|gZ;Ye%+53lLHByXa<ix8i-qiB9|E@lq7#OtA`B>ej`JE@9
zIevZrw|2_C;*x;i;yTsYk27zdyTyIg{>neymH+t78kg<+!RNJRrVB%HWp0u@FgLdL
z@}2ZiFFnD-;2*#c%)`p#2PS|r5KxqwR(yw%fr0DSf_u&0rok*c!D0?#-f(Tf;tt|S
zJP8L0B%Y$LVgyX3P{F$k+)Z!SGx7vWD@*&pgcaJHJ(boz&12#T=5ye4g9-Bl%lOI!
z!+4JEZy0`lDY?bSf}+>+r;+Ea-S!?pBb9uW;5LQ$38}xBXu}E==H=w&hZ&OU>$taZ
z<FX`Xo?sydAtalG9fXm1q7I^P9!x7E7dTLQ5B*-S=ApI1p%A;5@$%P?11*qZky3)G
zy6nwaG{Yw83D7J24*YOKc!FhpW#L|V_#<!aZ--DRpdvvBK_o@;zVdKIPYqbVu=9&|
zgM1*$D-W~a%Z87eq8l%20^>lyK>#V7MI1zscyhjS@UY+sR`6AT^M1SR3T3`Fl^JBN
z0<S&H+@jQ+L{MxO7bO=fgBU>|0t}dgLyrDA_{DR#3R7SZLm*6;fgw3JB_{*O^-l9;
zW?^Mx=iub#;T7QL^AvP-Rukm%bP^I57Lky&u(Y<ZwX=876V<j-Hq(&DCMqT^BdegO
jq@t>>simWDU}$V=E@vX8Yvdv-=I#~3<mTt-!RP}3^|YP0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.FB2E0A9CD009454C.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.FB2E0A9CD009454C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..06ac96331eb69e12791e963b0be56407b51b6b75
GIT binary patch
literal 1914
zcmWIYbaN|WXJBwn@vO*AElFfyU|<jg;^LB`%w}c=1~o1QhKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%t=_0+xJ&G(M}Uf|;3AK=9zU1vuX8`G-#J+{LyYIQ(&u*%_u7@0uuWk6
zv1mh_hJeOu&&V9Rm<}F0zyA+5?|wb|ym6heeBU{d_w!GDetNU@=ciB8jgP)~<@52%
z(MRvHgYMMMxm)vlF5mrsb3Q(u{@ZeYoSvF<P{F^<jZaRWzaIbmdwkx{uezCWOm@D@
zZ8s!L%{N)y6reOiH2ZIqvbf)kT!Ti>`&;Zf?qtPfm)zR4$2iRJ@shVw>RSK4)?Irt
z@WHny5C0i8vnT4-zARfDyJT|NDUk%%b-5l&i;jQUeQ8JU)WTFDGr`V1?T3<;CrvC=
zbGoo?!RF5sCa1)(>{=dcaPMo@&v$p*zfTalEjc%H!Bwvv8BgRs^*PxYeN)L|P0U@i
z#^I$?N!fe1mttE*CMGOhDypG$v(@&6br#c>7t-?m@4mDZ^(?!XBB)emRK>&Iv1)Oa
zRI#3O$633SSx<lHoV--3FsJ<Xv2w<${pv@gmM;|asF^e6zG+nir-_rK=L(g;gqx~X
z-#AUxSh8$H+@e`mUU!RdZcDeADp7l9Th;pV5Xr47sx19i_W9;A1Y7y@zAarF@T!E5
zkt^|yW203@%tZCY8k)|T4;VXStyTp)I)4es>b&@HgS+$7i9gyein{P;hL%Tl{a%x=
zKk?SiTR*opZ*`cy>h)#21V<66PwO;HoWxyL@;iuYTr@d%Xyv}g?Rj=P1xqhS&6*Ln
z^V*`eP@mhAW^Gxo{k3Pwg8#3!Og3@4_h!!<?*$8w%I@=?$0z99xoGXA?sI{LyYCCf
zXQ?k1+vGLzgzK$C?RJZKGCp)aP1^PQ+>6;OZtb7>MD?xIlWV7EEpYxefjjbYNl?)>
z6E#0QqYsi>HZIb5GHJ$VS)o-tei7mZwoHq)-K-yZneCB^;d$Yy`cJKM`+cJuDeJ4x
z^W}RxoV;|d`pLelf8~lEJBeIz2-_yBeX)GYOPk&UOCKGL`gliW9?QRyH|CwE%uX3C
z4R_1>cgWPKKYVBK)m>Betl6J$zG-^aJn_R94R1MY%RQD6CRug!!20IX2U$<|CQr>=
zdS==7Mc0a>)_!Q7aJ;~Mxy$|A%Pl`h=GxX9&sUuMwW8+S;fg!#68-$&k3C+qfUigL
z_w@e>KG)o@+F$;^;@I;S3zT;KJDqc`-O%)xYSet;57RXN-xq6}ru<u4;MYb`hT_WH
zBza(-obdR6xc9GX?|B&f0~mt2S-FG31W*P7ic-^x8G(70>(_#N&EBTLOgzD&4x-+0
zZNXv=VsM^7JLg%y=B>TVJi*c&($)w?${fl_JZ%nbI8UL?*;8rl(>xZQU|t7aH-r`$
z8yP1APsK(BiKk<u1LrxmzhU_KrQ{Yf3yPgA!N_*91S8w&HFJOex#l;#tUST|F8rPd
zTZCPNk$7@8a*hZ&H5)Y~o}P^!oEPFJr2b-}4J#v0Fb^jWKTJ!+N%h`i0V$0@VFMcj
zFPLzy_Ri_i5+Qy-VQ~j>514Rsf|T28L*p`FI0!ljy1;~af`uG}ka)sg!hQ%j5eE^t
z988pv3mk#Hhkh?u^Uzx1P>9{jc=_wcfl(sKBB=yZ#S^T^qKL%PWYL84j(ycTdtRir
z9Oyy~8x6P%FMD$q&9F&&0(2pt10PaM$=b-mV}mDH)kYPGr)#4N=RN$9xAwO~s1ye<
z*uD4z5Ka+r5I{0c&Pfh#Cr_}vjXc}{o?ta6H6%H88+9Z(JtsXRIei;_xZG0%)-UY*
z;@zMK66KMHS^UhlW7nbG8U{dN1ser;1pRi|70P^VDl<s60*^gRHA7KqP9i8z6c;5I
zD+6if;Nv^uGE3V8ZZHJ}G6X`h3y2ItFq3mrax#Ei?=)X#7FITP4o+?!UIBhSPeE5_
zH9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlAvI>eyDyr(5S~~g$hQ_Amawbx`
SMlPaa?p`5GZhnp)j6MKh3uplV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.21AF094B4B66A48D.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.21AF094B4B66A48D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..337051c91a69355331ca726759da8833d268d7eb
GIT binary patch
literal 1912
zcmZvb3p7-D9LFy+)}w32Fpm(U!5E7n&vKdW%rFcQE<Ges6r)KoXq7tE$)m`7snm+}
z5b1%9M8|ILjTI@6l2_|3&)5k&+B0MJoc*73|Mz}=-{0e$@BRP1xE#(wc>rJsaS!u@
z1%YS)0C1uwQV<b(5(NMhB>*^__u7Xa-eIlMx@6dpobyMR@*KOT{nlOTx@zLQvxlBO
zuE2TJdA_LbnA0oMq)b8x5foCU#=0LdVgx^=$;=hp@u#WKZgIEoXZR(84ELo^_a6>l
zY_vFLK|a@jAOGd~RDav)`N_!~izfp^n<iSG^o)ghb;jj=I4Zn^=~~R2=+6;usPm^&
z*<J@1L+|#zYJ9irudllf%+HvF`pYuhuG`-U$PTx>b<%^9O9)%=)g`&V=bN42c9k;{
zJ9++Lg7$kg7F*W!ULDOkc6wphr1XX7r`f*lMe|egfhKW-(S<v&W=6ln2RPp5J5vgt
zO+S3|ASL@yu&OmKsV1R*k8Vc#Au8+7%KZCNX_-NO=<4e`%|6cX=EpuHd`(mBP`?zK
z|JLO}NT2rPITpiwR*xqa$S){?2U&vXan2yI0-qjGR79Xr+D_XKKs?#<0Znr1*mQhE
z^0lTQ93|R37L-pcF63!M(%Ffx89^8N=Zs&xiPFi7>3A9=6I*BSyT<h^IOn5zSzT7K
z+Y~KX>f9T8o&jz8wzG;>R5Z^H&)Fe&^Brd!JAU8#?A39dm9b?p-s%;4`smb_S~os`
z*lt!HjVkpR5@2MM0!Nr9Y(xCg4GL+7?9fj#iCVVB2$MbS!AolDzT?F1PoGO@A~;|}
zePVpi2us3E)7z`s=POQCz&XXk&5QshUSsk$&5}iOxQT^Hv?j}j`kS@A2?rQexTt2|
z3%UMPtp)Kun>sQsl$ROJBwx*68Y<7UWPKc|8F9_O@<glF^)d$MmQ+xhakjy8-NP=m
zT|9$A;ysu2=Z@|535-H8<jdLq-POW|fr~fV>(2M-k81R_zPgalo=sb|y;<ND(P~L`
zr<;FKFTYzr>&wWU(o!u3-M5j<>}3m$I8cv^b&ZA}IKb6kq$d5^W&S>>>_{Uf+!cQD
zrr}6m?b`+Ih+Y=H1>RDrWz-Z?K4_QnxTxoe??k8GW%Qz8#3t#b^-J@jtsLHBy%j5U
zYZdagI;*CnF5Ko`4$mKSdKQ3ct5h%a3?2wZ)nVxJBx9R%KG(9%FMd4Z+ZcMk9!t?K
z)Vid4B>Zy!<9+2*-%9=BH?{UoY*6eyG=AKU(W#kw|Kt$aV%Bu2Hn$|T`1t2V&!omC
zren}TPFvpQ-9uBZR~h3=JFb1(*g<Sh9tq#xQkD|=6Mr{ZbR#Q^_MME3$5w#AW)A>a
zCAZ4!`zBKUT@k^1BV|N4R_X8jkEdL%5Lpl*zyw!G8A61K62g@wX++%~`V9>tI2gxX
zNm7MXrAW<1&22@)!+43N6VK*SO8XCp8YsicoRy?C&>H3n=|Or@WDFTgh?y_~%ufs2
zQD{U{LDN=Tf<|-|bfw5h!AL?b=j&1%PK_wbfe6+C%N3V^h?;|%6lp`+@QOx-s8U3S
z=o0dFS2Y+&x090r5l|7tic7Y=Fi3go5p+TnHiOJ0efdUJIXhN+yNkjknB*+Z7Qbbb
zzW<GsdSQM^H>6YV%^2K8e)kLoBI;;$iZ}~I)}q%+ks;bpLQiMtuN(1kF`@x9h~^|N
zX?9hN$hF(uCo&oXW2B7Mg0v)Nf`~q(FGVJhiG;f64wMRE9}NYOYc5z1abFNofmNj1
zX*0DYGeLw5ktGd4gvz8!X$FviltyRLr8HB>RHF5p$xX{+NoSY+C4gk{;LrAn)%6c)
zW}>hTq$Bb7>*U^VRK%Dpa!mR>fd4rUAR>Jpgbvj{sMc)M@&Fxhqj&%i5xhTe`A$Se
z?1|J}{_tGhzAqGVH0$R_!<cwwSuan(OB@D(J^Y~kAtH3$>xM$h{UEQPxC&HO!D6^L
zM>Z9w#AT_H)bOj-*KdIA>=_$jI>E?R*P2HDzbH|2jh4<@ik`lK;g80qX6r1hY_u&k
TOw1h!L?;(-S&lo?S!UBeR_+v7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.D78BE8BB81161A14.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.D78BE8BB81161A14.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c69162528d72ddb5f6f7b7101437df2373583daa
GIT binary patch
literal 1914
zcmWIYbaN|WXJBwn@vO*AElFfyU|<jg;^LB`%w}c=1|=>AhKf07Lv!<Qn+aU|uX(QL
zwN5tg59dd>R&P`@+$HwYBS1w}aFNGbkDp8T?{hz`-#M8#LX79P(&u*%_u7@0uuWk6
zv1mh_hJeOu&&V9RxDFmW-}>j9cfX#kZ(L_AA2(Oi&Sv8C)0??JKYj6c=A&1id_G<|
z`siJB(4E>@cWeIl^4<SG<Kxrozb*I2>8Uvf75uB*_~f+y_507h$LIh2s+$?dWaq!!
zc0<C{e3R8p0ZKDOv;RgZi~HTkHE8s_zs0WOPF7rY$*oO$jKd5cFL^tquJ!M0-L)qJ
zAAEcAux`r1<_**KmF&K#cgfQ_m1~32bu$&gi^;!YGw&QbwZn*Wc8AM7=Z6~wPoA(4
zR?LXLaI^A+r;#4h-OG9s@2j@fe1G?t-J|nu&pDF|S*kH6Mf{cSit>`xLR*_Qm<5I{
zC{eW9UB0-aJF3fL!=+1HA`{;nlKUcS${6*fx8J?I%GuKGa>gbOL2F6dHfEQrfu_Bd
z(-n`T$s3(5s+pGZ(n{dm?%c<_8TQ^6PU^jUfkVa4@07fh?HU%Tg*_@ElQcHG5t9Ag
zA|=dZD#x{WZPS%(<u!_jjpumv*yTm<y>7j#Cu)*VlY7?vWo8YUbC<V$+ZCo!YSqrr
zYVd7=gRIHA6O#j{Oj&61f#C?B?A4VE6st709(hr4WAVaI9)AvBRdwOd3@wl9`n@J!
zf8wp3w|;JI-s&)Y)$7Z4363ICpVn!ZIElNg<aZF)xM*_j(8_&}+w<&p3YK1unl&SC
z=e0#`p+2`K&DyeF`)kjV1^?e{nQY>8@6Db!-U}8UmEGq(k5ACIbJ5yK-RA-gci$I|
z&r)A3w#jSa3D;YP+U*weWPIp;nzZZpxfion+}c0$iRxRaC)ZBTTHyR`0(a!)lAxk%
zCTf0qMjs@%Y+R)AWYUb!vO=qP{365+Y?;ER_mmx(np0u9L*l|@uRp#*`TNs0oQ(TE
zPxjulhDl53em_yW`nT<qW33u18rS4lhb?-aak)~eapI$+Q6KN9%wzfY_(s0a)Z8hl
zUh8|V{AtZ@ky~H1d{wDtW#s=Gc`5o=_ekuuXOlj$?W{zvmb1R}k5mJxm<@L<!e)k@
z%Bj-MvR!_bt6un{#M>8BZd8BPVYgd6_j&$Pqm)f@{pUR;`m6sx{`l}-#T?O)|1&n$
z|NMS5r8eZR>GyuGnm#@)rPuq<JgX53o*%FDwetUcqo@CGZo80{V$bNAx4xC3xH30M
z9+)S+d#4<~dwj}d9tQsahG1@1?jSG$l!1Vv)U;woV4mgrwcuW}w`njFPq3(is5e|&
zu$Y4woTt#{?5VW&X&wttFs}oz8$ywcwTvTzr(&&w#M80Xf%6>O-!T0AQgVx#C0Lq6
z+6tzHC0Ln58HuOOp$+GG&D`IAuK5iwD^D=L3%@5!3s11Hi!c&T&RWg^A*W`ohQ!me
z)`Rn4Y8bh|;m~{N_kuMKtrbH2gw$V5v|(lB3FhJC;fJY;IH}%yEFh&3C~RPD0Jknz
zd*}3Mi4Z@au(*S`2Tb)e%}EKtY6VZ3c~E@EgW@}$U}WDN3bA_`FMs_wP>Upsq!P>|
zo?t~5MI@dkizb|R?5p0{^CGq7K+83(HQZrZE_-tp&9F&&0*o6z2R@{@k+qhEhd)oS
zs<kQ-PuE%(&U^SHZ|!e~P$>>zKzs2Az|7?d7H|+iGEUA>4sIt;u)MWA+yI_nHAgih
zIdyAwBso1tJtR4OYkj!fQv=p7?EK>0ps*L^k%w9Q%(i3Kq1_q=Kw$-I1$fy1cG(rm
zd~GT-NVNiwJxujqfv`<?^N&0OdSB2%&>1Go6D;H)gv1l}681&Ni8zSB<rs=ma}q&$
zqPQrzSQ$t&2d5t|w`g^+jbI82WC(<07Z4eQU?%6L<YWN3-f6zfEUawo9Gu)dyaN1u
zo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?M8%|KWEB*ZR8-Y9wRH3i42@0A
W<xHe>ja)><+`U4W-25Cp7<~X1=s#8f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.143C90F3DEDC2341.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.143C90F3DEDC2341.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3194548487954adf1c7c62f6eefab3393a751ad9
GIT binary patch
literal 1912
zcmWIYbaN|UXJBwn@vO*AElFfyU|<jg;^LB`%qC_A1~o1QhKf07LvQEZwh*}bUnKdM
zui0#=56+Knt=;IGR1*DU;sOsZC6|e>CVW~Nzu)=v^dp|)H*z-I5T9w>UTkYz!Zv~N
z$D$2!8Uh-t3nSOq&*|i`^ZWPU^6uBO^@Hn-rQ_yltebaY^U;&7Hy{1^+xh6zBbSd?
zj{bQU9dxI5&b^xdF75B@{eC`r{kJDx|Jh?jO^g3F@gI&(f1SVi_xkOBeoZscXPmD-
zS#HMWQ|7Zm9W?~cOxjwvbRxI<8?_k+mgGmtyX2X!H?_)*yeGMOM&YG0ul<MqeVumg
ziAF*3mxB6v2b(uc+gGytqTVG>>r}1{O4rR)1TQB4ip{!n<kSu$&et6-_naSY6g+vt
zLRc{)`ohi16P`wTOm{EqNxZMxTJ!zgVbLW=wz;WGWo#AFliJz5XYqvo9^1|+$B3C4
z+5x)+=HA^IxXV@7RV5-bleNpEERp?J-z)~*U&oF&Tkl;avpCZvvZZ5gkKAFoMXy%O
zIyUE&K+<OJnQ4}G-kVBfIiBBL_A$2M-h07~$1*cmgyh|y`uECBW9kiX6Vmb&i74yr
zi*@SdY?R_}y%^^BYOCP22@g5_RNdsw*WP<68|tR($$3zDYksiwfhp&LovUNDr&!H%
zZg7&=9ln51YWkCt8eXRYq;{|;wfDUWxgcOWB`T@J;?2c?N|k>PKQS$8-?A!N_h{|4
z?bAKp#=QM=>Y&#G@2g)k-)~gl;;jr9ky^;T=t}hh?kO2k=?_Ehf2=j1f2U(trmnBg
z`k2hX!$!+;J$-Loom%A<biv;0txE3%^Jx9uLKzu_-T9N3JG%rg%2@qmQE|wbZ{JT$
zzjbOw>zW{yqKVr+EIhBlEcGYJaDLAI;9qL1%JS7JJ$JM1{JKdkW4`SZr)ws2r_3?!
z?R>nMr-m(ejRx0Fm1i~WEn3aV*I0e}4?JaEEL*559mjKy?aLCO`o~9Zzn5II`TE}U
z-`iCeq+~v~ExP}zj(=yNB3G7z_U*o@FRX8s%yWAX^D${@MV^p9)BnOZ=AEamoibWF
z-7V|iAyX&*@SVX|S54Woratf5rs-K+BJ24svTe#WcH4aF!fCcT={fx8BJyNb`)H?{
z?G)WQ?_yT#e$Gmc@&d0fduwMg&%b!>v-xL9qsV@@<u*NTwG9_D>hn!balJawv9bOq
z|FM+XfWNc;xtrR#vx{E*8lP5d)wA-V{>5K&8n*lV{J+g$i_FdVV|zBG7FXsb$pdp_
zlFIsI!{F9J9tQsahF~sMu3#_$l!1Vv)U;woV2<VbwcuW}w`njFPq2uCh&NnYu&9G5
zoTt#{?5VW&X&wttFpmR|8$yw^jkFVjr);B)#M8FXhVvZT-!T0AQgVx#C0L3>${MDH
zC0L0=35lo0p#|r8&D`IAuK5iwD^D<=3!f)U3s10+ix3h|)<)J5A*X7iip0~k(S`Fu
z{DjnBOtfKT<O$~H<mQKIi8!g=dn_QO5h$#0qwfV1hS7{%;85#5^n1aYht>+IzK(kv
zH!e#8hJ%2E08%&zItU{1guH|TU?%Vc3p)tI<#M%mPLGxd@dH{X<{;()Q*<cA?q$6E
z_2bMu!4fPIN-$xbU<DQhB%TI~2Ap^7tKQl3BDLi}lhke0;U-=7<}8|Flk^1WLS6@6
zxV=2VGBz^s*x(6Pu~9+d>DcJNc@KZ&t^Mr~D#ZZ|b}v38Z}2<tBN->_Bn!8bCs@u#
z4sHNXu&R?PlAM~28j_r@lP;2+o{b({?x_Lm7j}N}Zcqe?a?8WQ;F)d5u0y*u41mJ&
zHuCTYdNnWj<<VTDUSNnJCnz421jQ4KoS=TY><VSRHkBD<fdaQZ%mRj@)SN_6o+vI#
zE>;H8%)!^E&2`xG*YFNgP#{AfB)fpfAOtfxHzg+n$n{S1WoBVzW9Q)H=HV6K=kpYF
zbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CY
UrEBCOD(3DL!sO=X=)vd%0Gf2-F#rGn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.426F56794F68AE44.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp.426F56794F68AE44.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7fd04b4823a2b73300e63d3687cddd3d0e90790d
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhO3+m3>9;ZhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^Imj{p@_!9^Z#J$^3Tzt8=&e&=M~h@hPKlzX;&c0a%Of=7Y*
zgKOg60EK|C$(wG}>=2Nwne*>pdiizp`RRYs>}2L@-m{<h{Pbq-&rhG~r*~bvI`h%W
z?xT0FFWK?iV)xH)*0THlT0TCl|Gl_wZ;Vgdk_Ue-C!L%=fBpV*@Au#N^g8;|9+sL}
ze$@#FweDwxu`YBm(7yI-tDD}O4Yw1RChyCv71()o&$Sm@QY+KfBpvg-t@)Sh_x0$g
z$%_xXJ$d+VDjWZys9%@fy6yH*e>+9<Kv&G|2`(=EFUyw}N^3tprIMo*^jVIpzt_X<
z*h~(`+fC=rIr~j26nrtS>|n+9)lYYq%inWW+GcEZ*<tn6f{Q0?pUAe>q`dLG!gc7D
zOJw7v))#N?^<L7=(o{O=<*gmyvWdI;LU{;F#zk{Gxw|iU3wxGbOc7KnGpgcY@7T0B
zN2*xQx#O%|%B-h9dQx6m37p%V`*=6Q-uuEyy_YX=EUB9_<-TcE1gD9Uq~{8iz=WGi
z%ieIBs<CL<h`2?wuDs?J;oO#Pu~ee=&bF%c<x?c5st7i@XWd_B)Sx+cdDFLDVH%}Y
z?F_93-xfH?nyfo9Ind>l0Iv<hMrZd*t&A16F1n>97H=*FR4(~<_yyOZ_ARTdb&uAD
zZJ$2t*3Mf$w>EEen7-+Cru+s4F5b#;5vhgTi>|OQ;C3<T-CVG0{f~vS{PSGyZkeXG
zEPUP86^Wvkw=Ge6``OD>S@XqytG6n>6U?pke~V;f6t=%VaanPZX5g04qKlSZY2T|~
zO}}+=MeCX%m7<B;J}gY1!pK{<F~+C-zGQ9i>TmgKm7crVc7ENY_TqqCkz)2NKc#fv
z<VR1GnCCl6Yj-*6tt{T>>@@Wt=WC{A?g=GFGu&fNwuLu$Gup1avfuO7?R?3cP1pCP
zw{KTnkdpb_w&?z=I{ux7id<R>+PC|rzOcSkGSBTn(8r{;|4L76XGxY$oq0)H*|ff|
z$f@#W$ku6l*6hzW-!wgIp7`O57r#4f%RQE%AX#<u!20IX2U$<|CQr>=dS==7Mc0aN
z&HunZv2WiBQ~Ta42hz8k__KDMliKT#KlW^ST=>4+#_o=E|K_GkGREJI{Z~?Zz4+C8
z>t7mMpTAhJXx^XcIp>~do5$*h{EPTL|Kykdvb9p5->|!MP4FnL%uSL9=EP}h-`(eB
zVh`bA@DE@J=4R#g0uw+P2q;QTE8fJ!z`*ru!M$d0(_m(vU{MEAZ@9K#F$Xaup16ZJ
zoF^E1DWhNs&juErU;zgKH-r{BKe-SDPt{KqiKpYI1Lx^3TU`-n|3sUSCs@r-EefW^
zs$kX5Ne$0_0EKn^bYoz`j_q$4ets#r#mo{cr7YzO6J`lkR8~adX((&Jc`lQ`ya+53
zz6W$6uLCcV3#I*};jZ8bR`OFq;%WM6!g(QnLh3Ij+OPth!o$hK53@67<%YZ$>icE^
z0}I(HJSa}#2}X8Grd0f0k4x+zPs#hq!#&lUAmz5&(6|g}lAwd23(NwZU?B$~B%X+a
z2%HB~!^j1W$=*Z17p!?`t#ByB?q$6E_2WR3Bv~YtV5)e66<8FIc<L<baNcEa&Y~GM
zNl$>m!{@+<6g)D1GVmba30C$~M&fDtX~B6<4OqXh^NV+bf=84`9%kp44Iei}H(u0a
z<q76@;D@Ido?u}IVI-ccpKK&d3s10$p9&IB+fN(L`|YwTl=<3JW{@okJoYdxMX5Q7
zpfp@ulw7O~q?v<fehG4Q-k1A`DJYO32$H-(WFUf>oSTx90pxn8`7*Pxvaxe;a`W&C
z@bh^Jx;m=~a(Oxli3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g(Kj$O
ZHZ_+sk<vAC5fyXy3Sn~dbM#>J0RS&n<+uO<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.1C2FE2C88C10C9D6.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp.1C2FE2C88C10C9D6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b92e640c5bc5e985caf4bab1c491e5a8b04244ae
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%t~ekhKrmG3>9;ZhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^Imj|(cQf{Q%fdi-3vf1mqF{m#j}5kWceDfevm?0$al1&;#r
zhswOY0SW=rCM(^q+aVxXGw0vI^z!TG^V9#N*~!e+yk|f0`RUEvpPxR}Pw%>Tb>^d$
z-AC_UU$W!3#qOWqtY!E8wS0V9|9f%W-WZ>@B@g~wPC7Y#{`&pr-tWKj>2>s_JuEe|
z{HhZUdc|jlu`P5l$iDWY*G+HEhTREFllSG-2xzaaTzWC{blBRBU6Sg#tG{T!iO<hG
z%f<T6Pwt(*afQIOOVjhN-*UCeT<*gcwq^^jr^}Zab=RfJH=3;TIAo?|`d;ySM0ZNc
zoKAr)XJ5XtS>s`($8`6yp2YjAtu^1@9j^8`Vzt<q*CdKly!Vdty}%Rh$L6`LRmky~
zqN*{s``Wv^8g~~=QxaOUWebz5%I*#9wcf4`Y_)G6Uz~eS=y*n!ub5(z_Tu9O4I0*`
zmYF>_YKeI7yezWsyocC2)`xp;U){%h;kT{gH4EOY4zG@T{)}DRzSTi_%ET5gAy%>6
ztmE?na-9|&a?aZlIw32!#Wba&Ex7CA@z}-RtK4S>dI~94EcxcE#XRFsrB=rJ?3SSA
z#T*m3lli%>I6j_bBk8(?=i&iof#{Mc%bRB1a=sGe-hNn)cd7d;`<`_QRUWHvWoo|L
zyld~3E7wBL<;&!-?3xnm|GR;8Vs^yV%5{Rb6u4>`ZaKJIw%}g=t4n^>&nIqIN|Q5_
zs-C)ab6qpc_13y%w?4#pQq#ZHSN$$=R9>#Ux!l3ITlU_><<2g_3o=eUSyVd3=X>?5
z>910*I8D<O+UXJf<AU+21~&V6y=8aH-R@sm^>?;s+@*B~PgO+<mL%|ND{9a3Q%d*E
zee^_$dA_5tcbAji%Hn;lPE!wZW-~8yPbfK>;U06aExfs#(RSsP{hqIG=S${nzP>lT
zeY@&{l+5S0MfYFS@$W2D<kC{mo;6MMMf9zbd2SDaKGrP$S9)SQOLBDT%uCtIruBVA
zPL&Tsc23;0V&D61o2F~c<9Bzz&V18Ea&cwKi&WnIe0uESIjJ(*zRIa)dqt<tyZo!A
zo}=W8>)fA_bJ*+_n?B3`B4QNTd;EE+Y=8Ck#~&ZwtC;h2%99xz&Ho7+M#_fP-#_O*
z)7G6`we-pUlw!M>yBT`}ek9+lSG)Yb^j+q$&GLK#eFqqdD|3_NfjLojhvF#_?iX4-
z4E_NO!Q8Ceo?rqf0|7;;X~mnE7#O&IEx6b0Z5qtX6D;Z=>J8TxEao7F#1nTAhx53#
zeVPxwxLya;BBd<lgV3U=tcb+ZP}YF+1Vb-n6fEJ{z`_$O;2_|J&?4t67mVPk`l=%F
zbbNK-Ji+yyQl;x&9R}Jd=pg8f&?4j@gv1kZ5P|bxY8bh|KJPvBd%>EA)(X1IR#(K?
zKhb9730Ct}i-f7NDp<91Qp2+!Kw({9-DsGwWBVJ1pI=IDF|(lfl_eP2uPnjHes!7r
z<wanT@I9bkc^!C>{3`7$4fhXEu#&G55>L}t6V40q6H<RM(S{Z16dq0<ewb4-rQ+{;
zTw({gP~KM_?!rSMb}!@QuOA17j3kSs5==EuumXz$5>K5)9nQP#%~>?VCg};#g?tWt
zNG_D|m4Um0Cs^568HuOms|DvhHDLY1&M)2#a-k@XJj~878$ND|ZoH_;$`j1*zz<I^
zJi)>a!bm(>U)czl7M@@gUlk;twy!pv_uFMxDD$<c%phA7c<f<Xic)hDL20<SD7jb}
zNHYg7jx`p3eSB*ZQ&1p75F~kn$Up=$IX5LI1IYDG^JQjXWn<^y<mTZO;OFxcbahq}
z<nnY95*HSckhHM0wz0LdchD2nwo*3JkjEw}CM_eYps1vxs;;S}qi<knY-%oNBBg8O
RA}Z$Y6~g4^=jg%c0{~0>!BzkO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.BB1A75845E7DE245.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp.BB1A75845E7DE245.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1241ecb56209cbadbf8cd52001d6b46c0ea155ce
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhKrmG3>9;ZhUVtqHWRq^U-Mkg
zYn^P~AI^_%t=_0+xJ&G(#|0Ht!9^Z#J$^3TU*~>WzjG3AL{LtA%01gXyPw~C!K1+Z
zp)zl8fI`5u$x8R@_6SJU%=&jOz5Kd)eA=HhyFITo?QAAKKYg$DXVI5^%XjR0qd9L^
z#GP92uHF8dXTRTn@!@ZI<9&as?4MsheKf^`>zuv*w3?q!f8GA`@3rp!eX3`@4?KUl
zg;{LPm#t~30$wbe(|Ye`JUPU<n_JD`<?CbY8#m1gmzujR=c?4|8HJa&dF?;+@9VT{
zYXTm;d-Jev%E9Ih)Ap6@zNUA{(>k?iL(<Y*52Z!NzwExYqkCzgtI#&Vm6iNEk4LE}
zKlW)^kn5NJJkfPhq2P;oWd|#+uY9_@T;AS6X`A`0%mo*{c4RzZ{M6@UXY@@ai#0KM
zk&ffbm@VHPxV>cCDl#!)=~7V*rJJqHFQT=WOkQNm`@j1#&B84>dlQFXw4`mDvdhgt
zv#l>r3nbNY%uKViQ;ysr*77+o{I{9I^Y1+ok4>&Ht&~rH%HJz@jj1=lO-RdABx1`-
zS$3zboQ+)kJ1>SguG%ViZ9*dda*@UTyWQl!^cgLlq~i3!<L&GaHiOpBAs76vb1Xgg
zTtGqmn4IW}=3}OnMqM6~iyGY(w!P5sXEoj0zG7(~U%RZ7SO1mz9lKgyD6V}K=Uq7e
z>f5Euwyi6wdz_$rVd)#ITkbj?F^5;3E$+R*7WhK>0-MsTW04kDL;pmaRX1O>PTE`e
zvZi@d$cDbm=%9&wl}l$$6!|iLZP_HZ6SMbDcjhwLvg7#oCocsqifGI-+L<A%y7|BD
zckj2USDdD43hnfW{&B(ZR0ErRyxy|k?|b&G3~e``yk|*l^VF}KlwO>W3v&uHnd@?H
z)~zEaUD@oK)uxHG#sodD>2A?#7QV*j(|_Qq>SEbKR_QoCJ+?1Pg8r+m%-^53;cVRZ
zd9wGXHB4GM_xp+3#k*~v9Bb8B(Xb{XDr`Z1=H*JM#)*%9y8V4UA)oWuwo^t+ue)Xa
zlX=?m>Cnm|<;uXnH*!<eLv7^u+V7M*vFxP8S!HMa>>sHHQZXB*7OpYropSR_SV)!M
zYvF&27fy<oK1(l@`q6W3#{P@JCsIu1zQ4ATE6+Rr`0?J47N<2&8YbueQBK-qwesJ-
zx&21feSBW8PSl@VW*_tWV%5SA%QyY^S@!?+oy$Gx^^E^o`*svp<|fGlb7BcAv-Pcu
z;q!PH`~w(*xmmeA!30nS0*X@8iZ?MaFmU}^aIe|hG?<wuSkyt(8?G%_%s~u^C+;8)
z=Lv>h$|zXEvw?*tSinKR4WUKOS1uUAQ}tCv;_3M6z<IjMR#(K?Khb9730Ct}i-c*h
zDp<91Qp2+!Kw({9-DsGwWBVJ1pI=IDF|!0qDNFglgjs?Wl@*bA8p;}Qp3CGfF9M5%
z?*U!N>%fcTLTO)VxGQ*qm3)<uc$&VNa9)U?kot>>HmpFW@Nn|*!|cqIiofe|i5=uZ
zd0%<B3t`fXT;MS7J@k9Qnupd3I~EFkj7fN)01R4Whx4E~oF^FB;fF%(UdGE`KMvF)
z$s(x)Gngk>fkgp{r_Q1d=Uw*ZESh1H^aSWvJ_kM|zsmT^!2QD$tn90d#MAQCg7cml
zuzq3Z7w-o7Rg^~_X6Kg;A2&reUesje3Fdd;hetk7u&{$L5>M7wHUg%FCs@T-1&OEa
zs}1M<cG(rmd~GT-$QA`2dzhBL0%4o(<{x<m^su0VpfgOECs@cq2#F`+AOhzVrRF4p
z(r|H6a<MXyW)2oOxTj-fc2^8jP#{APBzc3#Km;>6Hzg+n$n{S1WoBVzW9Q)H=HV6K
z=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_
YYA$CYrEBCOD(3DL!sO=X=)vd%0GUSM0RR91

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp.A1E84AD887D755DB.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp.A1E84AD887D755DB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..31fa84a2ecb9b0774098ff1269ba088a9e5dc487
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhO3+m3>9;ZhThJ*Z6R>=zew`2
z**RvW5897x-G0oh^J4Bq_a-IHP65}f_9^;*fAQ(Y3#spRn7rt>Qt`8bv~zb}@F*~U
za82ABpb!xDcvHlFdtu3%8Gnv0FTZX+f7zcjvpqI}F}4%ayC-vRKKe6Xz3X(h@X=uL
z&%5<K3%@5;eg5iiRr9y{l=%Gj>3^yW<@r4i{yhG(N#8EEe&*f!eNV4%UsB2W!(!%_
z19L+6U0lQE)R{1C%~$Q-7>h*ngDh&bmwzb~t*%^qG4u4Na}h^-CT|P=CHigs_AP4e
zjk6cG|NhJ)*Ru7?(%Zdd6a8*!1~-WA+@{#sA$R$G&||f*V^cjN+?PDF6_Jyk*wbSy
z*tCsx?%eh{Dvy;e*uHK0xGwy3`TKkEj!N51tu8vOo?3A6gzOX9)|!+zo>#aIopOn6
zywv*Q&AHx7x>=e^2fe(t16($7TVE&-Vad2?ZYOu=rD<W$vWqE#N@Ye>Jk=eW7UxJ6
zYdLqE)k~T6^hZz1ODlnMyL}(;X4rdQIjQ&Z1&$?kGE?uHR7G%^I7xc0Pzg-9xwPyJ
zhp8HimW_y8H0#Q1ZV}FH>K02SYVT~RT3$Xya;l18lY7?vWkwB}bC)-L+ZCo!YSYfp
zYEZV`LDod?<m5n?Qv$qqm^V7RS88RfuyxTbEwOlWF`#nEzr#;iUHCIYw?}pTUX!P9
zy0v)gXKi-vM!nSQm+vJzib#E0r(xnG?y~Z`gLp@VRQkh^_>Z-w^Y3)*%GCAsSs#-b
zc-UxJuBYp*r&F`sf-d|ox#cOPn7=KqO!LBpM`HIUEq8VaUX-!=$)e&ZXTE*Ep}%!%
zAa~?ak0(mG51r3zFiZVOGMt~YKls<Ou($VoK20j)Eqa~mdtsjK6Q^q?b6w8OI(Fn^
zB%58cS(r#`OwjY1_7<(?<ZG-x{Rf`1E|x7+m5$>%$M$7O(EnpA^XpSKoQwND@9q6*
z4U?A6{eGf$^>5oJ$67U3G_1+74qNm-<8q}`WAV}MslOgh$mcwE?Ud2d>26v7WS+Ks
zy0kJ&wKB5)_L@}vt53}0WNY|SZm0FIO<Fje_m7>0{G5$<EW&1noysZF&az$nm-D~i
z%Lm<2pD%mx*0^<_x&KNlWs_X5pSQKt*E;vbfA?P0(AGFBn4JGdKWUTQihp|-_Xkz?
z@oBv}xjv=XD*pD%s)e80cm4NS_W%8h%QMpNGhTi%aYk`vZjwAOCmt5BOMD%4{1^{|
ze*i-;H!HUnm;lN^Kv8O1@g^n)2CiQV?lpUx1~c;ni#mvU!?gvAIfx<g#2v)pJi*XQ
z83jvtHn8vn3pfb4A+*T($%P<zs(z|SJRLtBI8S%k>WVn~C)$iW!D@bLQ7|o51*>*W
zYIyboD6H$J8v_$|Y=6V>^GnGsW|m+nWhq~nFiWtavLX^sLs<jPbD8|*MPQNeJ)jGD
z9e9ykDD5W=cLh(dlAjV1Pt#8m&hxaKx8}&EX}^J?f@~)bik&>c$aaSK38}xBXu}G0
zJr5@jKg^k_zK(kvH!e#8nk48Th~yL@2O%V$h=T~6mnjv0*W(g9C>-Sd<l*6PD8%k%
zy!`d!K$9d{B$Z$m@B}NcC?N6FS=8Y?m>NbdaOC$M`n_PyLu-Z0-ke1<Y?7V;gNM(7
z4{k3{u#BG!JP3G#mHm{Fcv^m1aNbh`)-UY*;@u#>it@<AtogFx<EH4wi<+!F!Tb*V
z@btnHEbJhR#FO=tjf83830Cn_LE>rqX~TKHU3P^sUz^GdvPFT%9;T%zH7602hKq}m
zi<N;ib8u~OYDbNesS{IBAVUx&d4tG61T#4|B_{*O^-l9;W?^Mx=iub#;T7QL^AvP-
zRukm%bP^I57Lky&u(Y<ZwX=876V<j-Hq(&DCMqT^BdegOq@t>>simWDU}$V=E@vX8
TYvdv-=I#~3<mTt-!RP}3QJw?9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.43CB1A6BDABED53D.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.43CB1A6BDABED53D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a81cda35c32aa8e09d086b2fc9e5480ec5149a3d
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhO3+m3>9;ZhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^Imj{p@_!9^Z#J$^3Tzt8=&e&=M~h@hPKlzX;&c0a%Of=7Y*
zgKOg60EK|C$(wG}>=2Nwne*>pdiizp`RRYs>}2L@-m{<h{Pbq-&rhG~r*~bvI`h%W
z?xT0FFWK?iV)xH)*0THlT0TCl|Gl_wZ;Vgdk_Ue-C!L%=fBpV*@Au#N^g8;|9+sL}
ze$@#FweDwxu`YBm(7yI-tDD}O4Yw1RChyCv71()o&$Sm@QY+KfBpvg-t@)Sh_x0$g
z$%_xXJ$d+VDjWZys9%@fy6yH*e>+9<Kv&G|2`(=EFUyw}N^3tprIMo*^jVIpzt_X<
z*h~(`+fC=rIr~j26nrtS>|n+9)lYYq%inWW+GcEZ*<tn6f{Q0?pUAe>q`dLG!gc7D
zOJw7v))#N?^<L7=(o{O=<*gmyvWdI;LU{;F#zk{Gxw|iU3wxGbOc7KnGpgcY@7T0B
zN2*xQx#O%|%B-h9dQx6m37p%V`*=6Q-uuEyy_YX=EUB9_<-TcE1gD9Uq~{8iz=WGi
z%ieIBs<CL<h`2?wuDs?J;oO#Pu~ee=&bF%c<x?c5st7i@XWd_B)Sx+cdDFLDVH%}Y
z?F_93-xfH?nyfo9Ind>l0Iv<hMrZd*t&A16F1n>97H=*FR4(~<_yyOZ_ARTdb&uAD
zZJ$2t*3Mf$w>EEen7-+Cru+s4F5b#;5vhgTi>|OQ;C3<T-CVG0{f~vS{PSGyZkeXG
zEPUP86^Wvkw=Ge6``OD>S@XqytG6n>6U?pke~V;f6t=%VaanPZX5g04qKlSZY2T|~
zO}}+=MeCX%m7<B;J}gY1!pK{<F~+C-zGQ9i>TmgKm7crVc7ENY_TqqCkz)2NKc#fv
z<VR1GnCCl6Yj-*6tt{T>>@@Wt=WC{A?g=GFGu&fNwuLu$Gup1avfuO7?R?3cP1pCP
zw{KTnkdpb_w&?z=I{ux7id<R>+PC|rzOcSkGSBTn(8r{;|4L76XGxY$oq0)H*|ff|
z$Z2Qr)m>Betl6J$zG-^aJn_R9FMfB}mU}EiL9*)Rf%VO&53-)_O`e*$^vtsDi>?*l
zn*V`+V&A?MruMy84y126@n`KiC$-lff9%=vxbS_sjolsT{>@F7WQ@NZ`>&+-dhx6G
z*1t5iK7X-b(Y!y?bIv`_HjmX0`4{nh{>d-@WoxBAzhQSZn&44fnVTdJ%!v=(r)|E<
z{A(8vgMR=+FgGi=7nlIbKtNGyTJa_(1_rKQ3+^?0n+7xU1dBR|dc(B^i#doP@x&d(
z;XH*lXHTWIPxDxKf_WWy-4I%&{iH(>JS9IRB%Y?9CY-0cY;{GP{S$3Qo?taUwJ4Yt
ztAbTKCpA3#0TkBt(~W@%JGQ@J`1z&e7Bfq*l(LjBOqeBDQCSg*r=hF?=Y{wQslS+L
z!wNKshm(gNrX^<OhP)T*`(^?Cifj@Oib*`d$R=e<#ozU~#13+kyq`SWP0a~XZmSKA
z%YY^cItaSJEZ_+iau7n|i8zSBd51#mUdGE`KMvF)$s(zQ(4xSifW%X0QHS#`dvg}e
zut|CX3}`+FKBRz_@soiECQq=kpE43p%TEi=gIU7J1&;jQL%$cSd1$Th@JHU--wvTt
zKyw8g1dz;?^OJ+?<q1~xQ$^zG`02oTPYqbVu=9&|gThyoM;>O)mkl2`MK@m5WaSCw
zci@Mo7oK2Y2Vo?hte<QoObbu2ik}J+Puou$&in1ME0p=#RA!JZ3Ox2OEk&t0iJ&xG
zT$Eg_45XQZKh_;65!Mcz!xR+A5CloyATkiaOwLWo$pCV_(|nm(SlQS)IJtRv1^D?q
z1znxh1i3t&gv5nKBqS{?t!-@W>>cz(wXKxRG~}^~ib>1JDkv(csH$sf>F6668k?HS
WnMmmxxrmCndxbE$`8j$p`TzimyCI7J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp.336E2C49F8565116.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp.336E2C49F8565116.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6ecb7e5ea71e9e58ab0a7f7c11ab6e8cc5677889
GIT binary patch
literal 1856
zcmWIYbaS&{XJBwn@vO*AElFfyU|<jg;^LB`%tmGghA*593>9<EhThJ*Z6R>=zew^i
z-&el8KNJh!uHERH^rgv0$i#9IN8plaOKM(TFVFZn$>JvKED_iElyfrY&dija%b3ET
zcfMDct7ZDr6(_F0KkvdOf9(H<lW(tnUCz0G-ZysjuG8+HO6rOdt?cWo7fHm-(U99c
z<8GaA*KPjEuk-6KKfFDibAJ84*`HSb%(<e@`dmJ}*zSMH*V3x*KilH(PIzYd;Bmzr
zM&qzuQjwA^LXP^8kHcQ<+;AwTiO*-(u0ls0o=;qRT@Q<`HS5aM_Fesf^_%{7ld6ZU
z_hoGF%uO%Q?^&pScilxdtA(Lf?8;&pPC|Rj?f&Xr%uiaGJF%to%auL9@BEBXQO+%G
zX;53fxpaeq=WUJ>{@e@a%CdewJNw()<H$BQb*YR1PU)?89M1)vU_aL95+V?0bxKI&
z*&<VGb1v!rAeSalFXcTQDdJ@dvR86ks0iO(QP7_KNTReiMQ~AmYKh$&#Z!HoR(@T{
zqj+uigv@D=g(tP<I!2_MZ_Q>dsq62!D7H{+-mk)tI`{VpntQ~&yq#1#4~gey9q$WR
z>%8EQ^SX-~9$T*+$}qAysCo3oqh$x@-U~_#ndHJDw<31(DTZAe`nYbrlv;S@qYTp%
zr`dZCB|Gd_@sASJ^fmb-*Rgi{l;uscZaH5Ga{v98k9X<kEA<jF4kZ)9N>?x4;s5Ge
zOmK8`<n~P)gzhbSV|A-qr$Z+3;xo&tjAo4j{w<9i%L>K9N=5I7f4<n;QJS2QWc8HG
zmG$(DZK|S~)5BLudNTgcz33&SxO{tDndXHHzjXJh`nSI*edrpjCcbuQ%C`N->sGFP
z6L`YNRY^RP`@OZB<b?;%Ki{nQJ@<n7)~xSolcsOU-eBMt#`5;&VW$&a%O;<YjD2lk
zb^1W+g(H&|P4Ip-C);+R*Kfh>X8-;J51KCa-Dvu%UzKiK6|}N`@s;iSjUv+ZtIvPS
z_x><x>D=;pyV-WzKDpJVv7%v3MpRhE^-Ie>n=n1^danKTvSPe&&)TVpT+^eJ>br`Z
zc7{G(HDS++efifmP1c(Co!!RpVP$b;$HmFMlMSVOp3PnJF!ueP@?&DRHU*gnr9GAo
zd^qKs#vlJ9&H7mv_0_KzOX`}{Pd&Z(_=J;L7B*{t6x`S;&-eX%+hLvMYubDEe$wwc
z`D($h_ipwlK4{+3Ug^4j^4XU+a@*g{WRp4dUwh_1NomK+^ZznL$Ir20D6Y&+k_YBU
zi^{F{Q!Z3K<zes-U<l@B<qiZBKp6-qN=++1$;80G^=rYsW^dDAW}aYC2T^ahwqP*_
zF(jV2gE*WgxV}@Wblt1NKrMm}g3bsnLJmSmJP`*GI8UL?*;8rl(>xZQU|t7aH-r{h
zD_I8wPt{5liKlC&3+L%BTU`-n|3sUSCs@-;6K()ZD<c=!cfE&xFIe-?TEVJd)y_!`
z&wc<^8(JB9z*IT5zhU_KrQ{YfORx-wj3rE%C0K<+1&OD_p#$fI_z9`Mm}tWaG>M0k
zhaaXTQ!4(h$0c@<1&UURa0@m};@r+Ay?!Q8Si(WV38wl`h~3M0`Rm7l0Vc&Fr34e^
z307iJLgHz$Xu)}xy*Z0!*d#px1{j|MAKbq@!E#n|@Ic`SR<lw=;^|rG!FdmV<gNYf
z5Gn;UK)^u&$p8f_1-K%fU=1q`B%Xnl0i5^Lfb|PIzj!yu=b}9FFh{?ha;jx*VG$G1
zFUVoYgA#^3!N_6wWy8l!(Tx{1S$Tr_9r)qNh$mRsK^TcAZzb;vvxO&E-AWyar*EYX
z=lyos70P^VDl^Do3Ox2OEeu7eIf<aOU0jr0tPElVfd~i)gfSSHgJ1d^CTvWORbvVY
zWJu0U$;kl9d8hd@v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI+SxnkiE3LZ
vn`y{n6BUz|kyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN32ia`SWaVDte160#sH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.117E24CA3DC21528.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.117E24CA3DC21528.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fa4464ae0369b187b769133a6775d0434d371244
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%t~ekhKrmG3>9;ZhUVtqHWRq^U-Mkg
z^fTGKKb#-kTD?)taF^Imj|(cQf{Q%fdi-3vf1mqF{m#j}5kWceDfevm?0$al1&;#r
zhswOY0SW=rCM(^q+aVxXGw0vI^z!TG^V9#N*~!e+yk|f0`RUEvpPxR}Pw%>Tb>^d$
z-AC_UU$W!3#qOWqtY!E8wS0V9|9f%W-WZ>@B@g~wPC7Y#{`&pr-tWKj>2>s_JuEe|
z{HhZUdc|jlu`P5l$iDWY*G+HEhTREFllSG-2xzaaTzWC{blBRBU6Sg#tG{T!iO<hG
z%f<T6Pwt(*afQIOOVjhN-*UCeT<*gcwq^^jr^}Zab=RfJH=3;TIAo?|`d;ySM0ZNc
zoKAr)XJ5XtS>s`($8`6yp2YjAtu^1@9j^8`Vzt<q*CdKly!Vdty}%Rh$L6`LRmky~
zqN*{s``Wv^8g~~=QxaOUWebz5%I*#9wcf4`Y_)G6Uz~eS=y*n!ub5(z_Tu9O4I0*`
zmYF>_YKeI7yezWsyocC2)`xp;U){%h;kT{gH4EOY4zG@T{)}DRzSTi_%ET5gAy%>6
ztmE?na-9|&a?aZlIw32!#Wba&Ex7CA@z}-RtK4S>dI~94EcxcE#XRFsrB=rJ?3SSA
z#T*m3lli%>I6j_bBk8(?=i&iof#{Mc%bRB1a=sGe-hNn)cd7d;`<`_QRUWHvWoo|L
zyld~3E7wBL<;&!-?3xnm|GR;8Vs^yV%5{Rb6u4>`ZaKJIw%}g=t4n^>&nIqIN|Q5_
zs-C)ab6qpc_13y%w?4#pQq#ZHSN$$=R9>#Ux!l3ITlU_><<2g_3o=eUSyVd3=X>?5
z>910*I8D<O+UXJf<AU+21~&V6y=8aH-R@sm^>?;s+@*B~PgO+<mL%|ND{9a3Q%d*E
zee^_$dA_5tcbAji%Hn;lPE!wZW-~8yPbfK>;U06aExfs#(RSsP{hqIG=S${nzP>lT
zeY@&{l+5S0MfYFS@$W2D<kC{mo;6MMMf9zbd2SDaKGrP$S9)SQOLBDT%uCtIruBVA
zPCJ92E}O7t#lH93Hci)>$M5cbo%yDT<l@Sd7pc7a`SjSub5doreU($q_KHrOcllRK
zJx9qG*SS9<=djr=Hhq@=MZ_qw_xSTt+5YP7k3T-VS25@5lqWMbn*S3tjFb(nzkkks
zrmZ`>YUz{xDaCd%cQf_|{7Al8uXg!=>ATEho8|d-^c`R*uFOr62j;{DORhv|?kaf3
z!{8sl5X{ZW?FlA;G7wOdnpV7tiGhLZ*MfV^-loCKJi($4qTX<A!D0?#NIY=|aX61#
z+o$=^i|chjEmF!-J_s#}%8E!l4P^~DPjG#wROz}`hk;rI9R!^bT7(>gka!{vB5<BU
zo3p3V+NXIeJi)vUylx0B(!SEc2%eI!5)x0-R};?DUADR+&i;uuBTulJuUaHbi&ep@
zos$}#{QwH<`szl*gdN-8F#P;da*LS-#ZxT7$ev;eM)p*QpOE^Ci8ic2lXy6J_+chx
zO2yyxxWo>!K;Bm#ZUIc1kqaCVy@!4;So6?Y;ZTU(%Xs<g$AKXu$s(x)Q^gakz@mV}
zQ)f|!^DcXH7R|6pdIAg?J_kOekdg6~f%~2(SlL$@iKpeO1?N5dk+=4@L#Pzc009R9
zBm?Ap<=~2Vf>nK0k$5`3I&j`o1J*C>{NmjppNsOy!yNr(!^ch0jTbdpd4l;J_~Ge=
zCs^1)7>OtAD;ojR!V|3GtAfPS_SJ^-e!J`nWxh6*8DxtBk3CFFQEE;iC=C}EB^N6L
zY3AT}bGK=Pb$p-56cor11WDc?G7!N`&P~b50CK(4e3@BT+1NQaxp{a6`1w2qU7ghg
zxjdbO#Dzs9BrPnhZEWrA9rQ%Ct(46)<gtm0Nz2G8C@QI_s%vWL=o=Uso0`j+Na-56
Rh>E#;g)q7KIeIYq007@%2nPTF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.59CFAC251237A8A2.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.59CFAC251237A8A2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a5ea514f8b07d233240d9f87d953c0e0b393e786
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhKrmG3>9;ZhUVtqHWRq^U-Mkg
zYn^P~AI^_%t=_0+xJ&G(#|0Ht!9^Z#J$^3TU*~>WzjG3AL{LtA%01gXyPw~C!K1+Z
zp)zl8fI`5u$x8R@_6SJU%=&jOz5Kd)eA=HhyFITo?QAAKKYg$DXVI5^%XjR0qd9L^
z#GP92uHF8dXTRTn@!@ZI<9&as?4MsheKf^`>zuv*w3?q!f8GA`@3rp!eX3`@4?KUl
zg;{LPm#t~30$wbe(|Ye`JUPU<n_JD`<?CbY8#m1gmzujR=c?4|8HJa&dF?;+@9VT{
zYXTm;d-Jev%E9Ih)Ap6@zNUA{(>k?iL(<Y*52Z!NzwExYqkCzgtI#&Vm6iNEk4LE}
zKlW)^kn5NJJkfPhq2P;oWd|#+uY9_@T;AS6X`A`0%mo*{c4RzZ{M6@UXY@@ai#0KM
zk&ffbm@VHPxV>cCDl#!)=~7V*rJJqHFQT=WOkQNm`@j1#&B84>dlQFXw4`mDvdhgt
zv#l>r3nbNY%uKViQ;ysr*77+o{I{9I^Y1+ok4>&Ht&~rH%HJz@jj1=lO-RdABx1`-
zS$3zboQ+)kJ1>SguG%ViZ9*dda*@UTyWQl!^cgLlq~i3!<L&GaHiOpBAs76vb1Xgg
zTtGqmn4IW}=3}OnMqM6~iyGY(w!P5sXEoj0zG7(~U%RZ7SO1mz9lKgyD6V}K=Uq7e
z>f5Euwyi6wdz_$rVd)#ITkbj?F^5;3E$+R*7WhK>0-MsTW04kDL;pmaRX1O>PTE`e
zvZi@d$cDbm=%9&wl}l$$6!|iLZP_HZ6SMbDcjhwLvg7#oCocsqifGI-+L<A%y7|BD
zckj2USDdD43hnfW{&B(ZR0ErRyxy|k?|b&G3~e``yk|*l^VF}KlwO>W3v&uHnd@?H
z)~zEaUD@oK)uxHG#sodD>2A?#7QV*j(|_Qq>SEbKR_QoCJ+?1Pg8r+m%-^53;cVRZ
zd9wGXHB4GM_xp+3#k*~v9Bb8B(Xb{XDr`Z1=H*JM#)*%9y8V4UA)oWuwo^t+ue)Xa
zlX==wwEUE>VrAgp8@Z|Kp*He+?RUzZSawq4tg^Fy_K#EpshAB@3)h(RPPzFdETqcs
zweUa13n#@(pQRT{{ph(iWB<kA6Dg*0-(OqFmFFFQ{CMw2i_@AX4U_Z#C?{>QTKR9^
z+<v3#K0dEkC+bfwvyb_Gv1;Lm<(vNdEc^fZ&gGu;ddB~{eLIRPbCcwOInkqd?Vi``
zSu%MT`~w(*xmmeA!30nS0*X@8iZ?MaFmU}^aIe|hG?<wuSkyt(8?G%_%s~u^C+;8)
z=P9%~dn&Dcn#aNu%<I7GhR`DID;<pBDfucP@icuk;XK`Ct1IH{pJ+4k1grU~MZ&aL
z6|CAhso~iVps=p5ZZu5TvHcCh&o3pnm|23Ql%;%N!Ysjx%8E!l4P^~D52l8Z3+(CM
zL%$cSd1$Q=;wPm3VxkQz&;lM#9)6grOsV+09+%ib_R9Op!|mO%Q0QY!!UF|h$ROLx
zgJLgFFtWXeLhN3~%U?eZ)FR0usRXltCs=_+0g0#1q7LU>_U0^_VUzR(7?^wxd`N*Q
z<0}IXFrHv#Uu7homai6^_wYyF+TRYLQa}R)90ZUIkn@#;E8+=O^;JdU>G<lvc~1>k
zzp(R*cY{Jjlt&)s=r0>SZi;TasL9F`%<sStk9?kBVFzI(o~*BI1WXH0u!^q=5>MM#
z8_xUfvMZGN+EiwcEebsLFfD%t!ZzK_Kk^J1VuB8W&M;w~U?B$~B%X+a2%J}xnv)1h
z!^K6(#mYdMIrv7$N9&ulB63VYfeb;A<P9PN5zOS=l$;D8*E`LZnT3^&or9B`hgX1~
z&r{IVSxu13(@98NSVTh7!qVEt*3RBRPgL7V*-S$oo2ZzyjI4s9l8UOjrk0MrfuXUf
XxtxiVu91tVn7da9lbfHT2cr)F!s!55

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B76F39C8DBBE2FD2.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B76F39C8DBBE2FD2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f3e1cdc2720aea2a4a567515e56a074c353de0b0
GIT binary patch
literal 1824
zcmWIYbaRtnXJBwn@vO*AElFfyU|<jg;^LB`%qnIEhO3+m3>9;ZhThJ*Z6R>=zew`2
z**RvW5897x-G0oh^J4Bq_a-IHP65}f_9^;*fAQ(Y3#spRn7rt>Qt`8bv~zb}@F*~U
za82ABpb!xDcvHlFdtu3%8Gnv0FTZX+f7zcjvpqI}F}4%ayC-vRKKe6Xz3X(h@X=uL
z&%5<K3%@5;eg5iiRr9y{l=%Gj>3^yW<@r4i{yhG(N#8EEe&*f!eNV4%UsB2W!(!%_
z19L+6U0lQE)R{1C%~$Q-7>h*ngDh&bmwzb~t*%^qG4u4Na}h^-CT|P=CHigs_AP4e
zjk6cG|NhJ)*Ru7?(%Zdd6a8*!1~-WA+@{#sA$R$G&||f*V^cjN+?PDF6_Jyk*wbSy
z*tCsx?%eh{Dvy;e*uHK0xGwy3`TKkEj!N51tu8vOo?3A6gzOX9)|!+zo>#aIopOn6
zywv*Q&AHx7x>=e^2fe(t16($7TVE&-Vad2?ZYOu=rD<W$vWqE#N@Ye>Jk=eW7UxJ6
zYdLqE)k~T6^hZz1ODlnMyL}(;X4rdQIjQ&Z1&$?kGE?uHR7G%^I7xc0Pzg-9xwPyJ
zhp8HimW_y8H0#Q1ZV}FH>K02SYVT~RT3$Xya;l18lY7?vWkwB}bC)-L+ZCo!YSYfp
zYEZV`LDod?<m5n?Qv$qqm^V7RS88RfuyxTbEwOlWF`#nEzr#;iUHCIYw?}pTUX!P9
zy0v)gXKi-vM!nSQm+vJzib#E0r(xnG?y~Z`gLp@VRQkh^_>Z-w^Y3)*%GCAsSs#-b
zc-UxJuBYp*r&F`sf-d|ox#cOPn7=KqO!LBpM`HIUEq8VaUX-!=$)e&ZXTE*Ep}%!%
zAa~?ak0(mG51r3zFiZVOGMt~YKls<Ou($VoK20j)Eqa~mdtsjK6Q^q?b6w8OI(Fn^
zB%58cS(r#`OwjY1_7<(?<ZG-x{Rf`1E|x7+m5$>%$M$7O(EnpA^XpSKoQwND@9q6*
z4U?A6{eGf$^>5oJ$67U3G_1+74qNm-<8q}`WAV}MslOgh$mcwE?Ud2d>26v7WS+JZ
zEngL?Ss7V>drhkT)hA|gvNilEx6^vqCM}%K`^U~ge$K`_7GX2PPURG7XW1_P%lTjM
z<%901&zC)TYuvie+<&E&vPrJj&)Zt+Yn}Vzzk4rgXltAmOwRwKpR~zt#lO9a`-7_c
z__SV~T%S^G6@U9>)xuBhyZ-wu`~UvM<r(Ss8Lv#3IHR~SH%T6t6Q{_3{1(IgZZ;2t
ze*i-;7b}++m;lN^Kv8O1@g^n)2CiQV?lpUx1~c;ni#UjQ!?gvAI*20i#2m!nJcTxA
zPo=d_^H_L-c^r7$5L%@Cq(TrpML$I(o`#<WoTs~Nbw!;06KzJGU{ycWD3}(jf>k>w
zH9Y$P6xQ+6iGc|_w!dNc`K9C*GfS|fvZOCem?c<2SpkWsuB;B{d0Ng}b7a%B-#||x
zYvDoB!V`?FCB#oi{l!EZR-onFoZS2{TT*=;_cm@^mIO3Oz(D}XuYwMONIYQ&VK^^S
zD*mp=C3cWE<ox8|-hi3N$OVpw-b23^ta)gya45v?WxV|L<3I}}SR|BSs(6CsS>%y;
zYAkAS-eqsjq8T<xPk`ai>%a@QmnT@-PZ}P^Ji$tSN=Q6SKTSCA;g7tvza2uQfClh8
z@FN)@>n96W#1pLIr-H=O_S1&*o*J-zVdod`1_hQVw>-?zUp9Q)6y11Hla(i!&w&q~
zUU-6q9E6Z~GJZ0VFfBa6%6`g7JS{&hIPbU1u2AM{Q<*`wC~(`uv=pW0B!bd#aZz%y
zGLU8tmd;(EHPzU3FH=w;Ll7i+gUCPxGdVXUCj-d!PV;4EVP#|I;N<4v72xOd6m)e~
z6Xf!A5)u~{k&v{ow6?Levv<%F)wWVL(~!p|Dkd!>tDvZ)qN=W`rK4|PXl!aOXCkF*
S<RU8O?iIr1=I7|a=mP-Nq63Kl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.9F991AFB3119C4B9.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.9F991AFB3119C4B9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cb5d1e0def03047a452ffa40395d46f25a679d26
GIT binary patch
literal 1842
zcmWIYbaT^UXJBwn@vO*AElFfyU|<jg;^LB`%qC_A25v3}hKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%Wp7Y3+$HwYBjAarlFLLX&!0>8?{h!hC#)*H!9?Pl>$AIucUQmbJ;z|g
z5H50zll9Q)tT3_L{OL-~{h#fBJo@(O)#s*o-Q$xRAN>`s+$j|kvF=>Gq=uR}^U+}O
z>)YceKl=9g(W~3<OMZNN-n4$-9o?T#Usai@u|A)_F1+UJ&rh#Ey*tgje~;o>?*q$!
z?qOdS`BW+;RKSa6Z`#`13!We1+RSZb@N(8Q)``+jOsl%vp3XJv%GCB<{z3HH`t2rG
z-w(a_lY4j0c!&703)Az=GnD6E(28vkw461;<^5d${n=l3M`T8;B}&_xzTa1{Pw=Ei
zw5;L+%@;YcPgF9c8{aMDy>&BA_x|4BALN8K?!D;1tgDr{cpCeONUonJZcOapI^^aa
z*?7pQ<k<ta#Mgmc9vhr8Tf`^6ImB2ZE6N!4UP@g4&ckUIZg&+TTV7a8&gFOS*t9rD
z%65w4k!J!%XNzjwHf@pRc>QKwVQj;__nd~eluQl<+xwk*|3Y?6>vgA0&lM_xKQ?*4
zwGhZE;RvbFFp*+ic`Zu9w~cvDP><fbD4Q3$scx#CEe|H%+Pacwf%oeb3-+!P^_qK5
z(Lr&GrIrobm+VhVL?>QyIV|6Jqtd%lD`SPNi*9L&&AW>Mmi~X+PqMnOXNG={>iWGV
zPv3Ow=B=gD+|mzoPpw;Cd!$K2;L|#d3(n#$o$d>`T}*m6Z&<bd$HrOyc`i9o&cc_q
z%{8(jj&IRj>CwI?=&T3VkMQ=eOB|MazZZElwq1Mw(YDgcWxcolEv=m|bW}J0xBcq<
zHnra=OjBs5M|8z~<5LZ6`y*n{f4|RK8@!q?U#-$}H`~s_O=>Sr$Q3DO&+=1BFIoKf
zjSlm4M``UYC%u)$>pr(=y-i=kZD-3gncK}e$=ht7#N6H&o=g9o4Z2@<I$`$Sced91
zbPsw@E4x?uF6zzq3E9$%7BFqv^fuyUto!V9mpSHW$K{Cd2TA<7-dJy^a$D=L$hxgA
z{|?DI{jR>VDl6D4?%MtG=$qcA=es{XQ9S3dE%#W4%gpjkjq&c!7aLE@KCN}vYv$Q}
zmtOB%D?adl{1w43w=y-~r{IZERPjGEK4-PpFH0hRyxj8LZm!&W-{p~xo&J))PyW|d
zd(D3J{`cR3P3JELy2k#}zsa*)KlJC)HojH=r`7zof9<$>!e54hx~0#HD|3_NfjLs|
z<m<DG-;~SnF!%>B1aq@;2Y?Bn3<MOVrWNmIVqjo@ul~Z-xFwa5C)mK!z#XQI>(_#N
z&EBTLOgzD&4x-*LVV+<y2QfHL;GWT|#w)_>fO-`y72tXm+MGR=);`T+;R)t-;B`aj
zm9dnuNAOfERgic(mO5}AObsI!*v-9%elJ+_&|1OZu+p*nI&Azv>ohDiTw$sl+utz!
z{8Dm@nI%}7L)ro+%o423p^U`S=Fo=oJog2cEIj`4FEbB{79JEWJi*9XLi~i(Ure-N
z1zOI-$-@t`CF9ssY2*6P*Fa%$2XVNYniHhlRvQ|Z0WA=85Ojg5<_Q*Z5JKXKIEcV`
z-RuVzb{r~dVC4zscj5O$Xc2Z1M&ij?%Gn|0)GXDIczTw4aNeO1yO;6u*N+2TAju-B
zgwUeMqKL%PWYL84E_-tp&9F&&0*p642R@{DleLtE#|=-gs--FtPuEfx&U^SHZ|!e~
zP${4R0uBO52FP2=!xixat6Qog@$@bA;k>5?tY6sq#k)a)EXpGfbM$YQU7^g^rZR(s
z6?ovP8H!SK5<#iCxG1?;8N>(z5n#X^yvF;t?D-GV8JL0s83JL#3=GM+DLENHu6LR*
zGYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY8CeBI
jB^6b5O)VXL14Cm|b2$?!T_YD!F?X*JCO1Dv4@MsV9e56#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.A3477CD2DF481BB7.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.A3477CD2DF481BB7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dde2eb592d778f0113532e33a6f3a6c8f4423d51
GIT binary patch
literal 1842
zcmWIYbaT^UXJBwn@vO*AElFfyU|<jg;^LB`%tmGg26iq6hKf07LvQEZwh*}bUnKdM
z@3q@fADkcETD#FV=}VK1kcs6aj=&{eLHoYEw%+o_B5)ITj+D+f*Jo!Qo|&`rqND@c
zgHLxX7CKZ!`gPU({^yhOA@P^^{JSx6wevo%{#kG(cu%#;`RRMPKR^9>dtOq>#bBGV
zjXQo$6ET-h3%_rF>EZ9|lJRvr=YBr@G|5PX_4#};?|px2e&*KwemW~YZt^qB2bW83
zF>;6Pl8TgU5prA~`Pl5m&JBm&H1YZD+EwVdhUXL4Uf086Yt6bcwS8BAVEv}Q-K6Si
z>;1Vlcjl!RsP`<?zq{_Do7KWlD|Y3u3@4$z<#vD9U3{OkGIwH2?w2cjcHj9KrJ|f$
z+R~u5d~@jrCC}R&CH%P;&Xr~Te0KJC`Gk(Qy;V{f0i4oXZ5+-8onSxK=Mo}dW_3zP
z<k=!qc5^Q2{vekoQ7`2^9Vy~`3$j;oT&M`&UGbi8`w_!eQYV#M?w@*5W8pMKHZ}P5
zJ|4wu?GrMmJr<tSn(G*mUcNP(xunj&<D%F?v3b7=gZ{a^Pte^X=H>0A+IdJkH|uy~
zz*^@8hn&}4)bQAP?NEl%n}<`9N(!$(aLm^<*78(h;lHxZGnFAW!kzVPNpHZb5<bSF
zg=YgF8ZL;RtiHBmikj3vjw9EiPX#;rzE#Xxx%hBmd)uYTSN1)63RWJgcU=#<qh4CQ
zZsppwYi>u};K&aylYOf!=F+zz>$8mQ7Dtf^mRk-kmo2zg@9L6Y{qu#}kzIx+hI5OW
z7d4go+@91`azguR&yoiFs4bJ+PQ2VZ-I>c|OI`o_lb02%Vk<6AIq90MIrH!PmGN2n
z)gqgGCZ2G;b!c~AAe+gLpGAgu>(jsZuC>}T^NH$PsWm;5MIGJ0Z4lbD=%q@L+jT3y
zb4CS{CL5P%Jeg$mJj#E?rMg469$Z#7*p;wFd0WEa?t?cQ=7qevuk!Wnd7f=IPv6`8
z__olVl+5RIpRYTZ$G>x%BA1qe_N-~5_qJ|%Y14aP=cAu)9~D{;*316*?!0NOk<Y^@
z?<eQ)C~nz(e65yw>l4?n#oMRO4f$AOv5Vh$g70L*Spm=Hu6fvh@aKb`wOdk`<}Eq1
z?E0cZmbxK7+6!V=RQF%mc>Sq)fuycwy=cAS<f|2T*8i%wV=3R?|NYowo#k7O_xyb}
zUv%=-y|3QS{;RNm`HKZh<o-_2`L%d@*Pj^c>ZA3(`~J_qyTJBD9pjT5OX3-dD|3_N
zfjQD{f)!KV+qX<S4E_NO!Q8Ce{$K(q0|7;;X~nylfI0TP`U_X%mQ+TbV0{aHH<&iA
zUkmOvdz%I`@dS%Hh<d|>d4k0p#Na%Edq%GsuL!RL>Q%5%fa?`p-zinP?$u#to?t--
zL1%<sAqOEOo`{18oTt#{?5VW&X&wttFs}nI+$5f084DRZgaOLI%8>}3wuLra&fu`p
zvHLn~{6MFu2dl@!6v1dlF0k)=5B*-S=ApHMWBVJ1pI=IDF|!0qb4Z)RRIvmrDJumb
zcv>7<aJdjaA@vs%ZCHU8@Nn|*!xUv4n<{NwANm?7Ebbug3KQ;TKd`XlP*DRbPcXj=
zzb8zXCs^1;7>OrmA!mz_Qw>&)K=5=ebm4M`LhN3~%U?eZ3{goINhO3LMHWRQo(78s
zoOjuqvuK7*(i32a@;UG!g{Z8BEIibBf>nZ5kU~_)LI*DQ@JHU--wvTtKm!CE1dt4n
zw~&V`;t5s@RzosC&q5C__tb#(3p>AfHz>qJdE{Y^{>-oaYk|$KD4<`E!=48v?0JHb
z!~VC+u2AM{Q<*^~De&0COkyZX%}E5M=HjB{Vr38`2t<GZbMRw{ocKMD7WOd(1u_J}
zgc%r;b5n9MfL!l1UuG6oHg*n9ZXR9%em+k@S7$XrE>9;RabXb&NefGB8(TYj2R%`3
wD`hhcd2FI$(lW9Nib^V~>Y7?Q`UZx^rsi@cQo2SiqGIk|Axv(5jvkCY0PG~`6951J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C60AD9A6203D4F40.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.C60AD9A6203D4F40.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2e9354759fdf73fbd5cffcfe0075ad8b55231de2
GIT binary patch
literal 1842
zcmWIYbaT^UXJBwn@vO*AElFfyU|<jg;^LB`%tmGg26iq6hKf07LvQEZwh*}bUnKch
z>8xDdABu%<uia2H+@<@;<H8dUC6|d>o<A?GzrQf`c;}?pgv$vR#b+9~8y7!wOJ|U5
zI9>d>h3R3crRJx%?Zy)i9)CH%rmVQ~YwEG}&klFbd-PYia^IWS8T#(|)m`16=RMlk
zee~@1B|Cmv?EdwylkfijSs$NXziqicMo-N-sNi4a#wVxsr{90JJ%0Vqr>2?nnC<+R
z%Wg<mx_aVNM-9O*vrO}v=W%PF(R*=V$!clwBc~MBUx+rzxtg_j#<ojaxZ)H37kMuZ
zkNa`=<l(m(&DM!LwJ*zh%`S1Sy<{n|(MfpHi@l#;miN|)d*^nm<SbtP?C#TZ*OeqQ
zS1YraSk7JBEb4j7Z9#QLbJ@1w)8_ehd)i%c@~>QWX!g3X;wfX%G$%WQGL=P58!iS;
zTVNo*)^bPSjfyEsLTdt~9=fXRE?|iDc5OJe`_|Fw@*S#uf!9666n}-eJwGhC=#9qg
zsPCy98|_+rHp|SPpp(OW$ZGfMSm_0ytDSnbacuo?rGN4#^(yx4hg$=pgtR<G_GJWH
z^R;BjGKtC+^}0E(+A8KYIgx+4$l~tZZ2VQ;GXp(^6e|{#`D!u0IAo=ju{^saXn8Tm
z1n%5Eu71Y9vurY5mzZ4q;I6RkUeG+QtStUOZ&_JBYm+JGg8x5!B(tb_%c^SKqqRZX
zpU=8=>(s8^*4__hsn;*sB{@FfSZ#LxqO-J1qx%A07n9!28&<9Uv2j*^p353(XW`45
z<{DZN$G7ON^l0A`bk>9G$6{CQ%$7O#sx6l|IKTe>W1fxdqU){VzB6OCOq+D`|GZyI
zzio0?JT1~0vqX2#Yl&2c=K1>Km;aW#-M_Nx@NCbxOY07vvWOHcN#WO4)Sl(1lwPv-
z@e>{9>5jtQT~2x{i`RW^(R!=ChS$!PN&I<F*%7gv3Oftw3zNP6&R(+b@5F;<)i<ll
zYNDH_>%86b`Htz0w~E)zTpU?bTrY1rw0oM_+~tB2n)C8S_=6<=oNuhRRk^KoSY+K+
zmw#tuopx8>SydG55f@@#9(~hO^t^li`@=jbms5Jwrc|nDSC~uip1a`{9rnuYR8Ezy
z)||^ut@WJOTX%i;kgblbW1fC-(&zAx;zp6Z$N#SCKR);K#~&5>Hs^}Gs?Xfe{}*W(
zd3W`{oj3VERd?}uy}J1S<TCrJl_$<`-5mbs`jg+c->|)P{APaTz>;u=;>z44d0>vb
z>G<`{CfNxSc^Lcy7=pQ3x&6TePzC~uQqzieGXZn#d-WHt#x1FgJi+=F`ff07T)!6F
zYxXt`X5tAJbrAK23G)PtIf%h|0{4twHC_>32h^)zp#ayb(B|x^wDxHp3r{ew120^d
zCs@Wp#tvbUa<FnFf~Rev4VN=GtaR+Y4jVtvT=ih}c$gx`_BRYazm(i!W(k(&kT!=2
zvji(CD+M8VS{zz%Ilt&drBU@?Qki*BwD6#4;R#0865=PM{$ipHE6{QtP9A=kxf#c%
zN*mXQz6J`5JBYi&guB@fEbKT`)WFIU%<sbQ2@~cC7IqOv;>lUa*&^grgH<CCJY5T2
zxZI%-yO;6u*N+3;E6F0Mgixf&qKL%PV9|i{E_-tp&9F&&0t{_F2R@|GmbH+DhbT|5
zO0Wu2XzN($z~x|;Fmi!YNbjNF3)VceR(SX$Z|!e~P${6f0uBO5=E_^h!}an6s|Bke
znX6}^2bX(l!1{%qU%VR>kfJ>DFl&Ch><VSRHkBD9tiWRr6aFg@w&`yEk!L`k3pxlo
z!-RQ)g&c&Ccp?rWa2`WZYEB|3H5V5p7b^p4=HPZ8*?Uj8{;M(t1u_Id5;}+sL@<+c
zQ*ttZT<<hrW)@a9b`DN%9$o={K2JecXEi}CPbVR9VG#*Q3rlMoTRVFPJyC5dWit(V
rY@%Y)GO`MaN-C=Anp!&g28PC_=5i)dx<)RdV(wlcOm2RT9*jN!an%<;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B74C847DC291EB3D.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.B74C847DC291EB3D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0ea6f449406e2af231bdf5ff24cd482719269c29
GIT binary patch
literal 1842
zcmWIYbaT^UXJBwn@vO*AElFfyU|<jg;^LB`%tmGg25v3}hKf07LvQEZwh*}bUnKdM
z?<`;5ABu%<*KYJp`qE@0WMVmqBXG&IB{eUvmuK88T$Cc7Ath2f(X!aWI6bd~Z35#C
z8*@Jmfw*apmG1Ac7nH1-@#pYz^XcL9m;K2yvw5$nCwFqQ>-$4Fg}=5fx3GV$DYtvZ
z{Tkn{+5DGR=ht6;cze2Le9gYOA5R~xu~A`tF8`c+-;a`?rDeZ=&Wewl{LJ#f<CIH`
z#bLXoA|+dd9M?xYzV>40hC?~cd_KE&6*}the&X8edU)Dev#v~S-_;*jzv*u`se0LZ
zU(WW<ytD%Ko`w2%*Ijh8S{Q1@t~@QnNoa4m-QRT=@{?EQPHgf0a%Io%J3pgTlygg4
z8q}6=F5RHwd7GnzKlj49vaFxa&i>~1I8x?j%WJZtr8nxGV!GxN=45vzEskv2R8Fqq
zi?jAhxAwYgDjnonDj4UoiM#s3))1DAJ*#8)l($Pik|^y>5nPm?T4E=oaH?<9%F_5Y
zf!DPsOiq8~Jb7rF!kW#uZ*66?+V6fOqdS1x|F4DC{>9}RgyOiDsxF-5^03=%YqI}}
zYl;C66|ZNAEQ!iaG?DyP;AOaLM{dFT?NcPDst7i*XI)=n)S$O!aZ}kYH;q!8c7~l7
zo-Ozwabf+5$zd*D!o2kiNw0OEu3WI}o50p97awkLc7AE|RsP(01uKu$yDkUaQ7^4t
zw{q>;HPO)#-1)&}vTvEiT>3WTe3r4@;wVzVaLd8vvIY0*U0w34e!g%!vdhrKaBflW
zqNY%v+mpJsoY%hEv!vnwwhYzY374()e~V;f{N=kpak=xCT^|-{PVT-IWVHK!^8PD+
z-!xKYDoyO((poONm?z`I=c6^7e*3+g9ksRI_@wtO(+DYb?IylAiJB>{mpz~KtSgn7
zpYp&cBgr%Hglg%zt@8qw{++S);Bxl|e|j$V-Du)H&!)~?6|}Pc@s;iSjc%mqSD)w0
z_x><x>D=;pyV-WzKDpJVv7%v3MpRhE^-Ie>n=mc!dY*l>S%j_L>cjf>6zS9%ho(+D
zF@L9XMtOg9=sNC`-LHO^DO;~xB=GT(-8ly{-(wjrGs`!%l{a5L$Qhk=+Uu^zOy6~`
zt;V-je6atxE8@G{%GC8f1rLm(ivL;hIjgO{xFY_?#TB3Jq-5Xqx<&eR_DlRe`d?dZ
zb@kQy@4p?J{4Y9s#{Qa{^Xu~TRexfffug?q>d#eO&`PmqkZfLV&QM&Ln<Nj+k=*<z
zH%>eg8o|ThAHWdI#mW@`CV(;!P?VZhyqk%Ef&IPu3s>WoR7Rd)eM^0Jm^Q9o3+^?0
zn+7xS1dBL`c*BHwf<+xf;XIfqBNy20-b23^ta)gyAaKv<RpS-mbwDHKE#=`xDzrI!
zDy@B*$HEiL<G|wv)65erZ7FS!;3-=wBk{B?wc$L2!%D~Q>#*?y%~iKlcZF$jY=6V>
z^GnGsW|m+n4k-(mFiWr!hY}J`i$e>}3-J?De=*U96=)JSCpSM#ORBHq-o}l~l9+jd
z1snvBTqx)uh{O|i5Qg(Ij!l&|t`B_;G)c@s4DP0G_5%w$4iz=9@&xm_@Oi>4;0YFT
z5klh0TFTlX<WwzHk$Ae6x^Uj15WAQ0^4E_89VWpdp@h()z@mV}(_qnn^DcXH7R|6p
zdIAh>UI$*ddwGIoEM?#!$`h<&se;7QvDAU{9{$K%``aN@3TObo13!`ha+Y#%MLfZ3
zmTE{mJxe_}@2LUn7j}N}ZcreLa?8UU{o!E2=@O%bszARWM;8xDbnyfuN7rwcU7^g^
zrZR&}Qs9PLz)+N$lL$)9#YM@*${<D%hyVlT;JOg?!oq_)9GHRv83JL#3=GM+DLENH
zu6LR*GYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY
n8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X*JCO1Dv4@MsVFz($&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.91739D7EC5D8C4E5.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.91739D7EC5D8C4E5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..702365b4cebb59253388ed2cd181750bdeb94156
GIT binary patch
literal 1868
zcmWIYbaQiIXJBwn@vO*AElFfyU|<jg;^LB`%r<5Q21hOihKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%Wp7Y3+$HwYBjAarlFLLX&!0>8?{h!hCwp9agO2n!*JpPR@2-B=dyc_~
zAzb7bC+nfpSwUj2`Nfr*`!C=B@#f#JXP<lSpJ`t#eDrtvr<%f|#C<#dymT?#Wff>s
zUReF#RP&wvxx9UUz1s8lpa1db^x--2`g!WkK?VO_mK_wIzkdHY`F(djzP7$JhuO||
zxoyE|t?LS_n*x+xsAk{gvK0@T;d`Oc^ZFLOjynqLFGic_yx$V&6P;<*x_(3b-lZ3J
ziq{o?_)swO;9`TeeI>gcrC+v$y=3Y+aj`|E>V3_x*hTk`c-=PQeEni&Wq<4EUJtih
zuQ(We9n)VcsC(Y#DB;h&aPC{y&v$p*zfb7M?e#Ocut>-(YiIMEr4!tb^|^!ygjt>9
z5)oTuY;DdJ-5=!AB<iKSrz1tYY{Bal92cst@A&bA@A?tLS8OMhT<)KIQFFt2ifroT
z7r~NFIc18Mb&iXxc;9Z^a4t{kI_rzSasrEWos@09K3?#r@7}?n3f(EwTD*i>^><(C
zlU?Z6?$Fxqv^a1=*4oyKDHTn@T^En~c3baVdS-=6r-01L*x9ETc5Uh7y7fA0;gyFn
zOd?LR_Z~`i*stOrC8+6Z@<*;??)E9mo7!$UUkP$=KP<;P_4Ac_iCD*y31OwHm+sJi
zRkm*B+_h_?BO^HTgUf8++KRa}PAGS6l-s4?b%!}tVNvFs*3jCc->=q`DJNZ)nl&SC
zr&Hizqh-0CuD6~}&2kGmU~jc$lG}-wYp2I@nQW=+&Odosu`0IeV%SO7bj_J>-!F{M
z(ykWS<Tdex>#ak_=LNEv{5blkY}em&FJ?z=tu{XCearMj?_^WQ>TeUcBQKW-SuU2A
zRTr1s!E-AjP^8FnZt-08kWBl;*bgrSC1MS42}W;vc=bV&gTL0-{7JvQd9%HKQ~Lhq
zzg*6^O<O*nE54qzzdh!*0ITSPskz6!-do)&ndkN(=ws5-iV24VVvoB1+kSX|Xu`Iz
zHqDr<&Oc{-o%-{OSFZ}ytc<L`za~}xs*Qfz;^i}#Z*G%ZyfF3NNwzw<7(Ts-b-Jrd
zv{TLY3Y(tGJbozN#kTLqlj%`E*H4h$p>wz1d%NP~uN5`7rhVMeZz%Wveedx|$E6a+
z-}wJup0Q?rNWC4ed*(hDcG0hI_oozF&0cx!`K;<~?D;2u{om=dMeip66P{gN#g(~9
z^1xiF@I|!$W47iq9tQsahG1@1?l3R`l!1Vv)U@JDObiT6xu1_Xicf#e$P=t<uj>ob
z#`SB#y=HIIV4$$LgSa<LSg@?%2!F>0FQBlRy&7CtF6zj?a}@>t%sj!;9MX0$)jYvU
z97;$$O%6>sPod4(Q)%tfJQki{UI$(`gcccl8CL{P*<Klmr)94N=fTu4a)CY8d+7Iq
zH4m*79NXV8{QOdKi<t$*DJ&>XVF^Zdir38j{pXtB@G|iP3%Uq;!VKmK7IG1S^L#Yp
znWp^M@et^9d3$+yIEDBLslS+L!wMAU;pE|m>1|Gsa$9X^T*eAC(o4_}CJf{S2!<ed
zLS8~hazX(@NOB?$A}%nM-RuVzb{r~d0J@Ohg&*!Fo?u}YVL0zlh~3M0`Rm7lks!$;
zsRYx@6Rg0ZfW%X0QHS#`dvg}eut|CXj08RhKBP#HwU>nlG*7UKy$TXf+g=;ayS4RT
z=!-uRc|eDWI*588>=bhlgYzE#$Xol{Ayf)zi-3awk}Y!fa&VJ)f>rHRk$5`xI&j`o
z1J*C>{Nml9Ko;eZhuQMmWmhQkwW-V?VFez0m~c^QP9i9c7Z)WLD+6if;O`v4w`N~+
zk7o)BWC(&Je-IgnU?%6L<YWN3-f6zfEUawo9Gu)dyaN1uo`SB<YJyyzPD0|sA`+4o
zmew}5cJ>Z>qS{u<W*YL?M8%|KWEB*ZR8-Y9wRH3i42@0A<xHe>ja)><+`U4W-25Cp
G7<~ZkWeK$a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.612D072BD9DA9C78.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.612D072BD9DA9C78.idx
new file mode 100755
index 0000000000000000000000000000000000000000..62eecf9d13bb8cb83cae21822c7f8c88fbb2b037
GIT binary patch
literal 1868
zcmWIYbaQiIXJBwn@vO*AElFfyU|<jg;^LB`%r<5Q23sx$hKf07Lv!<Qn+aU|uX(O#
z`k8FrAI^_%Wp7Y3+$HwYBS1w}aFNGbkDp8T?{h!hCwp9agO2n!*JpPR@2-B=dyc_~
zAztJfH!D-Bsb=@rcJ7G>kH5^XDXXvjTzM@2+2M<wkNz50?z7lw5FhieWRXPdT#b3V
zckKN?OXRzJy7~ROrH8lQPp|p-H1XVe@$HipH7)*^?0(RF{(Ag#{`hx4zRG5vW1PS2
zvfPf;sre?Wn*x+(sAm85QWp2S;cL+7d4G#t$DORW?2=oX_85m5K3?*6N?q&U*Sc#@
z20nQ9;Ne4~X7@zh+LvXEW0y=WJ0+6fx-Qp4Y0>d7yRYo%oLZPFbWLz&Wq<2&Hx=c_
zJ}nD!9n+sHs-N7!QR2TlVbANUpYHDFf3GNbTWW5`f~#ITGM?~#>T{BltQOkZw81Da
zY(a^l)$ZcOCEZb79vd!Q;u4wo<`CZ(Sy9HQFTMT#@4n2kaJ!tbi9^s@(zeak<z}E+
zujN$5BWdDBXNzhaH<ie8Jioi_V{F5{_lz5lWoEDh+4G%xFJ-I4Dz&gjC1jGuhBrah
zWi3*|OrmmJi`O=VWGk;xJS;pXsK+iZYVT$1Q$1djI1Wy{wS6T|g4gqv3-+!P^^!fO
z=%CnhS37{c*ZWh7$V4L-=I;(~E(FfwlFH%_^p=(7t2POdTlwE$SJMl{wXedw3*)c6
zjS036x87g0LFh$LneAI!F_(pk@4K9oOB|<E@s~7qEGrcA`zm@r{PWG;6|=8hI`ZYz
zR54SX<Xh81mc+)bOjBX~>%M&M62VVtuT!Tp@#^M%)UC8~F;VT>=A~!leln+?|8G#V
z<kktHT`qc;#O%KH7%e!cFSh;i+j6)2S61Df?HPAzonzkKHJrCLx5u2AwQTZ<$jdV<
zcAhwpe&JZxq6yxw=DeQ}kXmqPSHnv|iCDv1g5jGUUVV_{;IH*Hf6}jS-fXYml)k_D
zFPBpz@>XU4o$b5sJLko8G;y6cwJlloSD)G3<$^mx_C#pPE!p4_UO4gp+YkAoGuBQ`
z_0sRX^5=}NQ-6N(@>QXlm67%L*QDxS-6PRA&xtSPc3Mx`w1v}o|Jdx1iP><+B5Y>Z
zshleHEZgOx-1UNAEFMgbfAwVl!O{ntOaAiiZS%SI_+zI3vB!2tD)#I*mrG|AO*(VK
z{*QC=rn;cNKQ&}he+tQlT-o}6(%F{@%Yx=-zr888NA>mpr~fXjOR;BQy|a_4xH30M
z9+)e42kWWLn|r{RhrvI9A()$$I}}U+Wgws^HLds(69WTN?&l+p;?tiq@&xPJ>H5I5
zas66wui4u)7$_|6Anpwl7Az|`!r!sM3n;8+r{)D0Uf(HIy6)9sR-Rx%FF{|JFp%dj
z7=+*nc?luO3Hb{l$%#0KIKxybv^je!t$muu!V}Ev!0QGR<_VUulW{@tl<ky}cv^N^
zaGqoP8-|}>N^UW;1WR*B+rqT41S@eUA@MXhG~qn2nfv?CHNW9y0y^47&=aPGCs@cu
z2+s4-jAxqiW5+|F_vP*6;ei$6C#3#jq75rhn1_>xAEvjP{lLPGLq!chBl%tU;fC-8
z3%dxzc`(I{T;TZWJ@k9Qnupd3heGUL#>-zn&dd`m$s(x)Q^gakz@mV}Q)f|!^DcXH
z7R|6pdIAhJJ_kOeP?NQjg@+7Ju!@}u5>MMs8_v76^<e0WKN5LB_li1*dLZl+a}a~`
z9{$K%``aN@3TTUfg8-5(a&~fXlX!wv?NpI?I(9m6-ctkCFYNr{-Jrk{<&lTk^5%2&
zRtFa?bD)Qjql^b7%6Nj2qwKfKu2AM{Q<*^~De&0COe#vvNd%?w;-chYWgyKQyo~Q`
z!|z4g1et;Y8G<0mA4CQsn8~>*IT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQu
zh=intrL~Q%oxOvesJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdrm8
KH$O)YMjrrT^6fDI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.91C96AAC7F5772DF.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.91C96AAC7F5772DF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..67a6b90b1c63fa2ae11eabd841f566fcae6801ae
GIT binary patch
literal 1868
zcmWIYbaQiIXJBwn@vO*AElFfyU|<jg;^LB`%r<5Q23sx$hKf07Lv!<Qn+aU|uX)br
zl}<MA59dd>R&P`@+$HwYBjAarlFLLX&!0>8*SVkWlRYlIK}Y(V>$AIucUQmbJ;z|g
z5H50zll9Q)tQfJ^{o+c^{*&W>e);$8+vzFy&*`oA7m=TLa`VxXtv8c?{BtR+EnQUc
zE~)OnuI4*?-~9dmo!Z~;_x<_k_2HiTI?o<^2NnFQ-1nec-~4_WfBgF&XJs?bG0s<?
zEH`8GDbo|H6sB-gos8P<biP$IO}HfClIdD*CG!*6U&MG%Tjz>g4x8ooQ2R}MerCzL
z*84IxHax}^BF8RF&oj?Zo_j$nwn5Nx)&!ULbN%;wf0-ST8LgHmYis&`W5GVblOEBs
ziVHMf<j6iz$&_w<w~+VN%{blrdw>666xz7<VkcWv$RRh~@(G(oKApHRv4iW7mwRO6
zA*YhG=WdCw1G_vnIAykoPkeKTuS8aqG3vdP`20H$XIZ%2RfuePVJ$hAzq(`7;vA`3
zE$5E2dMUG>{%A_cv=+E_)A#W%hQ0ThleV~Kv@EZanPPvzI)ZC`>!nEno{K)DPQSHL
z`N|7{6(0gJOt^y9ZZ(+2<6q!qxNApd;qvWMB&VtfHYsn-4wiJ7el5_kI#zp%)g0#r
zr;N?v6+ADdKk?L7TH4umpLyfwX`eza1XOoKy?XI9wZH9D<tzJ`c*l|nVWq2=?$Cc#
zwr=IzwQHjzBRKPe%WU7;in%mSD0gj?+oj-jhdEYZQRbZ1(AuNluhx_)Cta4BH6w1P
zQ{Z8vWx1ZNx1LVTatk_e-+Id=w-YbdPLJg>*;3b?fAX?oRczJ8u#>Lonls<NUl^aI
zT`jW7YvKvlTZfL%3uH6-aWu*9X5I2HzH4voo%uxdt<;laCuc3Ntv=DZCex~O&c)uo
z$K5<JY`JSRxOOf%xAUCvsx9)H^?rQeh>176B^bTw^y-608|H<)x^MCI?Q@=MH(%e|
z{P(udo|MezbDys}TF1Y0n<AH%g7)pcsrO2^ytL^(5b!Z+X~l#?0<lM3|7}0KKQv)m
zSes@{R_C8HzE1u5#j96^YF0-6zp*A&|LPuzef3<H6LM!8&We0CH|N3fg9QZ#i*h5E
zMr&?9Id?^&o3_|Lr5A+<)c0RGS>OEnK<dk1(p9`O*SIgfJg>*Cw&P+(eZI*lF0CUS
z8|#0xA4~bW<k#mwi&LLetyZka{;xXw@<E@acGquh)~)cq_W$(Xg|R39F?^NW@vFEp
zH%T6tD|znBK0ZZ``zsHFe*i-;H!F82m;lN^Kv8O1@g*h(2BzH4M;yhcKWF3#*0t01
zfobFVwcuW}w`nj?SlmI}8zu~+8M(lw^&a}YV9i5o1;Mg{Bm5m3ynw3J?9{wqsubFs
zJ(boz&12yS=5^q8g9-Bl%h<`dAb84l%1AsdJ1scRvHcCh&o3pnm|23QIizi2T3CXW
zIFyihnjD&Np4ZI%{pXtB@G|iP3%Uq;!nE)N3%Llvc|Mx)OjCaBcnEZ_yq!GU(;<FB
z>MthRumXj7IC=PCdb`;VEbKT`)BrS+--RD;2v4xEi!hvbD8%k%y!`d!%sj!8ERsqv
zy*$AREDA_Gbry9v?~KaP6*q$ooq$@9eaD01JDy-<-(B|RESh1H^aL29d=7j_Au4Mp
z3lBA(U==$RB%ZdNHk@~B>%q_$e<bpNei3yL^*}gN%s~v!d-x-7?Qe%rDWEL^4gyHF
z$l1xkP2vewwNpjn>DcMOc~1>kzp(R*cY^{_lt&(B%Ws!mq0HB&GJ}K_c<f=qe+9xe
z-OWGpj1?GVUV^?bVIa?6FbKgD@)AOl6Y>{Ak`r+dafYcZO3g_GrSamT<YHwY%^d7~
zzeuyS>D5=Jpg@KoNb(1ffe2=DZc0uDkn5f1%gn;c#?HaX&BH6e&*v%V>Z~Tn<>@3O
zE-WG;X<=z?V{2#cpeL$rrEI1lk4;oeT1HkuQAtHrT~kX(-@wq=)LhO)O4rCmRLtEg
Mgvrg%(Sy+k0FmknwEzGB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.A06BD8EFCD977C2D.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.A06BD8EFCD977C2D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b256627d3274602e0e296c41d9f99e8308cd9746
GIT binary patch
literal 1868
zcmWIYbaQiIXJBwn@vO*AElFfyU|<jg;^LB`%r<5Q21hOihKf07Lv!<Qn+aU|uX)aA
znvOT`59dd>R&P`@+$HwYBS1w}aFNGbkDp8T?{h!hCwpA_f{FAu*JpPR@2-B=dyc_~
zAztJfH!D-BsaE&Z-^LRUwttkbDXK62op~sK`SWo1uGi_GO6p1yt@i(T>QPu)xuoJ<
zQrUl1&3E=bdHetQwXd)D`|;=V;VJjeq%STGD)?7u_WZNIdHgc|`R{+6mCf{Hlt2Dh
zZpOw_=CeW_H3ZKzZLM25kz4(Z)r<p6@}uNk@=Vv8TIELGlUzNc@KTxA{zL!1PP_I*
zqhRm%4+RqsF5WP0U&-!^dY3${Q@J)MT{lw^yqNqeHY4WvsU1d~t2?gjIesWvdD6r}
zHKz;P7Hs}JVRA|g%dX|I2KT;Z{d{-#ceck7GdFjsjIBa?QafAsES|uAY@W+n1s&fh
zq8hOs=ko4q++8$HNodWMEljQ|yEm}by1O=r)joba*?O;1-$j#II!;IY-1-yQ1IwoP
zCikUsY`od#vsos7f=(>^p~~%fHR21t?{;5PC?(Z+rGM|I*>3#YjcyvsoKsb}bar3q
z(_MJ1#i6zRP{wM7tG76^Pki85F0#1)by)t_KBL8xRGbPHe9P5ho^h#CD`S0jYtZjv
zjsTA2a;__mkDF{HU6=4&JRmDzT{2~PQ`;@)D?#q(&sUp-$gTWeP|A7HIqGWfv`70^
z++Mk?Y+YI1%LL5}OW#!8vexNva=s_p>h_{Z<160_CPClFI&)rW?_2$Bvvi>O+NGT@
zr+VvVMI70pyVB$Ko}jZHT))=)S$ir!IeYB18LL#(od?pNN;)#UL~^Ibl=Mxy`RCp5
zrDjH1imOFBW0vUFypc?GXpTRV{`hRU)BP)}-p=-nyR^<RZ|@q;TN~SBPRv?1`9$R9
z85UNj52Rl>*0pGY_p3SYCj_Jx9NE?IQgFwvlr75ZQl3w5-ppvb^2+|eSGV&e*KE4J
zH+}bZ)deY;&uxqDzpCTkS*XaRrJ#MgZ|V!{TP5?{9yonWT3IpSkWlPV*MEl}-q$jS
z);v8WPGaTH&e<(;_n!HMo(}r7>EE`vlX0swWA@Z_Sx(5EZ73Dh7H9V1^@F5>V?~c6
zQlm6Cf1JDG>y$E`Kh9SUclg(>ocf>ZX5;xwf47%N8%6TBAJ3KK`z>+!;k$bUGekqW
z6qD_L&Oes&cge5cnHC<ORjpR6$p5c8`|(Ymt#()MOt!7?zV`pT=%Uh-`<R!l+<mLK
zGB-&cm@B&{$n7m+xV(yo!9Rc@n2VJw3`_uJAfPBUt@siX0|Qg;=Od2d)1Nc)1nb!A
z_`<Yt{aSFZ+1oT2C@kh6<_#0($>?F*=r66p%o8ldA!P>_4p!t)MB-_1Xux@bWd%q0
zJ2rR$O;WX2g`1?%=Ip7o_GumqPcV-Ij~l`yX?tl`1W(Cc35lm^uL<Wlw!dNc`K9C*
zGYg7SSWuk85{&E=ubKP%&o#f{W#S1Ia1roC*dpj62<Q1|#xqU%vEw1oy>j+)@Bj(%
z6H<RM(S{W$%+1No57V3K>$taZ<FX`Hppjkz0We`8FGwH)!4vcnM3NH>5=4>{b`XYp
zznlHQ!j3~l4L}$2x$wc=#1ky!A_V8bjAi5kM``b&-wW0}v{pD2V)rs${`zrX+(@uU
zD8W?m1k1C?Bk|N&)Zo0!-ke1<Y?7V;BZ1d}7w!t4U>SQEctG<6E88n0@wDu<;JjN~
z4~D+@BasJmuZV+)2f|KK2T?fh;g7tvza2uQfVS{E@FUqGYcC5oi6>aaUImG#ZLbaI
zJvCtc!p<+=4GMcvZh4q3zg>2PGGCj@3=&r0wucE9rRF4p(s*%Ea<MXq5d<P2AP~l2
zU=G%txOa+Xt^*5GP#{BcZc0uDP|iEemzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx
z(!$c(#@5c>K~GfMO4&?99-F9`w2Z8RqLPZLx~7(nzJZ~!skxkql&+DBsF=G~2$P$i
IqX(l805xU=3IG5A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.0F35369ACDDEB6D2.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp.0F35369ACDDEB6D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..92f36103e52cbc007e4d5bdaab112791a780994e
GIT binary patch
literal 1998
zcmWIYbaOk#&cNWD;#rZKT9U}Zz`!5~#Kk2=nVrlG3`@8e7%Jvm4b9EJZ6<K-zvj7~
z>1VQee>gw7wR)qP;V!YC8Uar{m0TuD#s9e!e}8f6@y^LP5!-lfh|etNfB$Z$xfSyX
zhMdT()2apPuRKHZ?B;yrvGc1xoo+roe17`Rt2Gu@fqUvz&QDhte|Gxwd_R%J-NHwc
z#k*&(_bmKeSoZa8xK+*X+G+ax@0|O5`e~ApitD`kD4owwpPyd;`L_P`pHEFQ=P}#)
zF1Ou~ur+t$)p_3Uu6C7{_3Axud%ZS%9_Q_{YoDEx?=3K15h*xznvdkd#Ye(#?7tVZ
z`cUC^+o~P5=O6TC`^N9OyNI*INja2rmiUTVp}Cv&{;s=_>%7)yYKw2dl|8#J?V6^p
zoEpN=ziY|njR6yu_%Q5hKc<~swYB2=yTjEUN6OsPrCtOS2~WA<d@bMv`>|;*QyyH`
z>X1^IX7g&Vbmv^>gUknoTA%f?Z~9z)!A*;)<lHv1JG&3_7V@le-Spw0tKgc?>>ZmH
z=R}oiHS@o<OPTfbLrY4kQbF?BZCfuozyIEmktz`NZiT%2)AACrusfQ1Qx!rcoQM?-
zw*J-<_0(2W&Z{-bY1P(8Q4=3>`f0LBZ<BrZwasYpBo(I*4!OHS*aD_*47m_@y->41
z{r$oxUh*P*w!1_tgC==sE_%r4&~Ywco={X4f1tOlEMN7N>(>JRCzdcCbc(t<ciJPn
zmA8w1dzTrO<tCh+>v{9*7XBH}Rh<M+3(2~i*S2A}D16c6+?^Bq9<^uL?GoJZ*2LFm
zeT;H`g6O4fOO)Pz_A*sI`B+}CcY)xhtmmoonRwUUerWxvgu^eiue`KKkXLo{Z`*I)
z?^4=hu4oGF^oXvwZhWeNZGVj3vfb}{_N@&4Z9aL=lGtXa=!*g+3H+?$yuRuaFRuzs
zESlDMb^*7ktK#%4meR5dgLZOu-&^XQP;xNC-REeVc=KjP+m%=Ld%n7@FPXFH`rh>R
z?MJN-UOE>&{Wjlk+cQ%?HU=|BUUdz#l+L{T*@Wr2k^Xg~?@if@em{5+aD4ySC%hIL
zdaw9A^5Q%*-%#lb>)|^ms$x6$WbMm0Uo?Hz9*y?J)oRQaOC=YtOgq;lRwp-y|6IgE
zp4Fl1E@pe5@t(`%{5xNuLh(l}N7nD?<pvufH&ra=uoFI8QFH6;#~u9Ua`Nwbk4JiL
zl`{Uu|6g=w#=B5`{)^2$^A|PGlKLIL$x`-Z$hYTF|H9w?Fc<z|udApV^Nm>{`ms4f
zab<3jJTPZYIK=RH@6MHLc^Lcy7=pQ3xf8(zPzC~uQqzh9fH{|Gub+k3QuaDVo?vkY
z@er6cu3ro8HG7)|Gw}qAI*5A1gn5F+9K_%}g*InTrL|A<Sa^bY9eCXkie$WG0uVeE
zFBK%7j+YLc=h*&+;pdl<Tg)uM(j3yBFfA;>${fl_JZ%nbIL~Y5{{C~#Z+KaGg85ze
z;lez@!Y;x{JUKtPV3=_{!D@bLNIX41JvcAKPe}d6L>pG1_jx#Z_+eT?tpW@8=D&La
z^cjB;e<)0tCs-~<E*`;Ci%~=3>BZ>5dEpreR@#z}w*$>p@KT6?X-UfP%-2ir><0=P
zcp1Qjn-iqmRvQ|Z0X-$?Am{>9%@Zu-AcVvd_7e6($O#7s!=sBQSj0gDt`eq(kqaE>
zy@!4;So6?Y;ZTU(%Xs<g$C-J8C0QhuV5)e66<HLKc$zGlaNcEa&Y~GMNuY4xbKpaY
zVOcL(cr5Y+t9q#-@pQd(;k;Y*f@z&xv7dn+*6`8@gW37;N8Z}s4xv&U!0`3rhpXlZ
z7H|+ivPCXLE()fFCs^J~9!aHIh#Hcdx|ceVoL-0?lAOMmK3wjp0qYlbe(`Qle2enP
z!wmoBv&uw#UO@~nXpj>k4@yGh2}Vwczg>2PGGCj@3^GZ9#~x-<QEE;iD2EjnB^N7$
z7(pNc448wNws&imXqX*i3JPQhgb6b+B<H5&WB|F|X}-)XtZeKYoZLLT0{ncQg09YL
zf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=Lbo31jjZMwvOr&&;
RTtvm(y+WAW{2V<PeE@b?R_Fi#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3DF5C2CC3145E799.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp.3DF5C2CC3145E799.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9fe62fba00dd0bf56203d3b8eab3b3be6b2c6387
GIT binary patch
literal 2006
zcmWIYbaOk$&cNWD;#rZKT9U}Zz`!5~#Kk2=nH|gw3=6mz7%Jvm4b9EJZ6<K-zvj7~
z>1VQ8e>gw7wR)qP;V!YC8W)~;D!ELQivM#d{{G_B<DHXpBDV3|5T9Aj|Nh-hb1UW(
z3^|cmr&SBoUloSt+0FULW9L_=z07=i`26KZujW`-1@5U=IWMLzp4R<&zMshAZsDWJ
z;-6=)_bmKeSoQU7xK+*X+9;iUcg}U4K3Zd=;u=^pOTX*r^V91$-`2l=^r>m)JZ3vz
zHCu!9t+^Ag&hvhEwX3wOcin@w*K5P)ao#Sw_Ic^CMULqKn^Qxgj8a?rgyIwbS9!0U
z*0=lZ#lvqa*~(+Ge_eVj=DR>xYNgkOn$BD57cF03de3#&?8>PS&l@c6XLoC_&8{>G
za%26l!rMxk+k4U!PKWOoDpGydi@m>BTfyIX<J=1uW?8SiCZXyRQd~a|+?doM^e2in
zu=&U&_r2xH8;dn66e1Q~xbxAWtd9Sdhik*Ro7dW^%XcXDId1h3v)iH3RQiu0uxyI&
zYQNME3-7l3Y?hffVT}(<-kO`+%&K1el@nac<8)f~>*E!F`pzBdvOc<0sD(>t=jx=a
z<MRSm?^%Axd1ZjN$JVeIZ6lk5nnx23Z#%sAUO-yNBo_|36}i(-F|0D0#&zqpmEO~j
zZ)~T;ylg(S(0#j-zn74vtH~Wkj*F6~f}OnIDrT)*d^oZH*Db48_C0zE5gw~|T@JdV
zURu3w<wcWY>v-4fG}hcb?;GRiJuVT)Ho350{LIDg@WsWX*Vug1`X38t_2;qVSo;cJ
z*1mp1I-+%p?n;l}dxFk-eEQ+;qMXsP=2qp+<qpoT%YO9P$TkH}eH=SC#>M&Mo4WqL
zldNwlikEh|=v@-K_qE4p!9jiT&nMqmFU`-o`tPiYe&+Rrm1{FNtTwc(iu3xaPrSS;
zG_ht{<Jkq=rml+9uUJaUE)3df*nMxQdqT;<40oTSZQ{+F8Escy*<bkTw!UP}rt5pt
z+qWOJK6vR|bodRv-?lnaJ~jq3MqYIdvy{%f+-<`2e3QoYO&^-F7yW+lAmI4^vrl*}
zHuPTcdE~{ZGtW@z3+v%SC#qsQ7iI0sH(xY;)*_j{cbz;g*He4iW?5=S{irOkvq<(%
zjaX}?xap*6eu10yi~lPBxHrCc4!C~bqPg+xq=UCvKFrMc^k>V|%6smoKOf#}bKYxJ
z$hjN#f4ox`?F`>*<H#>m*Tt8W^}7D#vyyqiSLbK{h%bG7-{(R7mc}i)H~AOGJX~K~
znVTdJ%$FDbt2}Tk*foKN!9Rc@n46V50Zag8AfPBUtvC{xZ<+S`S%@uVuVds17IzR2
zfobFVwcuW}w`njFPq3(is5eZQCs@ou49*jrCQ@@qpk9@kCs>+8+5@3TnL`<gr_G@a
z=LxRwlqy~K>M&4?po5?@LW_`t5E4(sK?KfIXmj>dTKhDQg(sNTf!7V8MaEOcAHh@c
zR6*kDc<R7;j_q$4ets#r#ms`@OqO6|XR-t%JJW0C{{C~#Z+KaGg85zeJrTADy9guk
z<b36V5OQk1YDhdiUp+W4#7{{5#Y7ucU?B5w^6<m7gjxj_?#+Mq1n4vVApTI8Fi)^t
zv|Jp5rxvY-#M6t`gY&{O60Ec(A8!YmtKg{+4%3p9;hC?O-q{ZnHt;ln3%7r8m9aeE
zBL$38VK3nTm}(#|NH_w<I}~E~GG6}rabN^VvPdeygn5D$Srn0Ynk<@d9!w1*7dVOa
z9{Rmt%|mO2%if$tGi;JTzT<P?Ly9+9Pg!`}@C2)Rsv_}pJ$2!{TlIozom{b>f!@&Y
z)PM)i!ykETe>;RqaR5Wjiyz4w0uBO5w#Ws`MZ#Rh6D;p34>y-5SS?r$Nlx8U9Z60v
zSPw}~-%}qh_tb#(3p>AfHz*cGdE{Y+|906G%6x4qGe}s0#~vokP?VaJ2+C{4Majj=
zAVv^~00ZXW&2!^EUfc23izz6OArL0az>u7ql9K`CdZ+m^v#_$Ub8vF=@Cxwrc?!BZ
zs|j*>Ithsji%3XXSX$fI+SxnkiE3LZn`y{n6BUz|kyTJsQc+dc)Y8#6Ff=wbmot&l
THF6OZbN32ia`SWaVDte1BwAm^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.B3BE2F1E169224A4.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp.B3BE2F1E169224A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7e603dfdd03284a5aa453ac97e0767e242e878d2
GIT binary patch
literal 1998
zcmWIYbaOk#&cNWD;#rZKT9U}Zz`!5~#Kk2=nVrlG3=6mz7%Jvm4b9EJZ6<K-zvj7~
z*E-p(Kb#-kTD?)taF^ImjSDKOf{Q%fdi+|tzs~)1pRoGP4LOn<^o`!j-Md?yU&1zl
zamOZ;KJNz_uL?u+?B;#svGc9dU1mN#Jbu~Ht2q^gi++6Yn5XNfcSdaSclECB&%#HO
z#Xrx^_bmLJS@rd=xK+*n+$f!W_s(^lK3Zd=;u=^pPrvKv^was9Z?C_8^r>m)JZ3w8
zHCu!9t=XchZKvOPE&BRSRN_Hh^XT=q!t377dG5Wq%WYw3o9<$>NtxO{Q(lRFTfg0;
zs#9jWJOB2@JnwcGepz~Z?WRV*Q(m!654}@sgal=m-Jg0Wcg5DVlQwXAoZYQ0yUj*2
zb2W2Cs+UzXyZ59mybj-|e7$uuPWRs4-#-|IHqO1+$!4XxlQ-&(V!GxN=4AIplO9~R
z>X2HZRcTUf9#p!5g`dS+mB(G~#JTs5;ww2WEX<DGQNG;lh~bN{iLLw+t_3gn1ybE_
zZuu3$CV#8;gvsfLoF`AYJ-Bu=_w}Ood+!AeZ)s&%2HE$WlIN0LW09q|RUu@;iH%(=
zWviXmew)X|zckFvY1P(eZj&E!`f0L>Z<~ATwQj1Ls%OiC##^N;c@}8DUa?^By2n9s
z=kGb6+-j*2P}!>eX^HB@OD@d(4jmbq{+zm7+gB{@<7>BlmEC@&euqW-1BJD({Jab2
zUwON8ncMA%?b|jy%Lp#Bearl;a%F<#=^#ay^V+2hi_{lo&e=J&_9*|=niAy|+l);N
z_ZF?KU|v08o2qE${P0zho^gNWriz+4`MuxsCV0WZqq6&y{o5OMIo-Q@(sj#}Gw;6N
z(BC;ZurP4B#}lR8hq321n5F&{8P0#RFZkE8u($VpK20j)op!9VZ2@z2MA5|yB|?^q
zqh*h!OYY#=6%Z^^<T<xkkaMckah)ok{sTuX7t3xGm7d3cj_u2mp#R5K=G&)lI2reS
zp6orb{-&jK%OiI9eXWkjc+I?=Y169iHBRE0m%B}vlq08PM^+r<y;xh|koxF;+9x)d
z8*Wl+SG~AHZIWGHaJRKi{kluY_3ED+c`o{<u9j8x{wWH(^pZ`k@XXD4xcp#7!Lez#
zHbs>?rA*Ggw?gvmg8!3$=pWg~?-*PEn1@Zzn=N1XK>DJSpD%|!tCZJ0&;R|CjrPhQ
z%jEwb^e1*ci~IVC#n$78rqzm-YyYdxzFg@STEFgL{ng%j<Hr9N*)DBMt!J3n^>8^u
zab<3jJTPY-``X@_YA}5{4}*UILohchcLJCI%0NI-YFcpsFy}Ju^|KIL%3jCF6D;l^
z9s<+G^=rYsW^dDACZ1qX2T^aBFi)_UgBYAAJYB(dp(*=%W}aYa4rvdBB4rL`B%U^h
zHk_x>=Ip7o_GumqPcW|ouNy*(jHir0f~VrCg2dDD)PeIH+utz!{8Dm@nFYm8mSAK%
zS%Q)6^qRT9|6KDMURIu9eiwdEge}4@!bm(hU%4QJoSLs15>L-p56%nm6H<RM(S{Wm
zVmzEY{4gz{R)K|k^WQxI`iwt_KNKd+6D$`k7l+`fMXMq4^rH3Pyzq<!D{aZg+kxgP
zcq)X$v?OJC=If<*_5+0tJPqK&heGUL#>-zn4h(xq7D*+TYMx+47DXhUCW|JV2UEky
z1x^aRhkh?u^Uzx1vNvbZ44Wj7!}%QekfKD^Qx+cVJi)4-sz^LtPhB|gR=r?aCs*ue
zpbIrTHQ+vf_#<!aZ--DR4q$+I@dv<c=Lr^Y5J0j;E?6!SAt&!C4>y-5SS?r$Nlx8U
z9Z60vSPw}~-%}qh_tb#(3p>AfHz<NcdE{Y+|906G%6x4qGe}s0#~vp9S0HTD-TWiZ
zfT=^!LC_f{%o8l+AcVvd_7e6*$O-!k!_xy#u!w^QTxC&eP9i9W6&EEJD+6ifV1YF@
zPXu^#RhfbU8G<004@3qcn8~>*IT=8%cbYFV3o9Et2PZcVuK+)vr=Y8|njn{_laRQu
zh=intrL~Q%oxOvesJ4}|nT9+zQ88&5Sp`KU6;*XjEggLWLt|5OITI;eBNtIIcdrm8
KH$O)YMjrqH0YFIr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.ED09CF7C1628C3B6.idx b/.cache/clangd/index/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp.ED09CF7C1628C3B6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0140ef4eb2a892d04d592b570bc269990bc42606
GIT binary patch
literal 2006
zcmWIYbaOk$&cNWD;#rZKT9U}Zz`!5~#Kk2=nH|gw3`@8e7%Jvm4b9EJZ6<K-zvj7~
z={nx5Kb#-kTD?)taF^ImjSEjam0TuD#s9e!e}8f6@y^LP8M}CHh|etNfB$Z$xfSyX
zhMdT()2apPuL48Y-Jkz~$FAqUhMM{G@cC*-ua;EQE!k0_;{V)TeA(yA@qHrC)7`s{
z>!q2;PkQw1@RL`s-<SOO^?b>U`h9c7rhmSf;nB6|hsA91&;HZnmu;WFzWa3cB^&k{
znf~g8V_CPAR@+X$vs(4_8?M-cyw|1I+X!!ayXHA}uWR$U1!>(YwxpcmkX7Ax=-=0A
zQMxkavzyy*U*>tY<Jy;{w{>SZsG9^$UGS4>tN)_nYfJB|?($tRb>*ZDe0{UZ!=tW$
zN?Fpw^&oJ1iJ6#~il%VGyM-T5&5F~#x3{{UvGc~Y7aq*AUU^MI)+eO6ejd0nsYB?G
z7i(bikxA}*i<LJPYg8yiEV^*#qeEF8`z;UGhI2QswN;nzQ0sHt>LF&gL!+s*jz6$$
zitp;a)DH{qw)$+AnKxmL4@=&fn_JAPUi_64Sj*#dTK4PX1%KMk9hzi)bg57am(b4D
zNm<AH0#@%?e#m)cfVaoiuo!J4n?ssM6AmvsJojEeTF4|94!IS%(@!z1GMdJ9>$R2M
z(}!<tr^LM6duXBib`^gwAx&44J8~QsB~JxAdB0W6TDkaeV*jsOR<G>m%y--{A*^)u
z(jES<zQqJD&g_Ymj3`VGDzklS`>e8aLq}>S-=gQD?;2lpW=N&yrNn=%HI={1aieUO
z+OqI<6U;a8WJU)~EUR2PYvPlS>pR>oa71qXy2-n-?ON_f-AXH#<yyygTkY&<n{@M^
z+%MJbo1FDtiFC#+(cSY}GS#6u{#^Rwx8+XvudKQ|+cWOcx`R&97X`MY^K<J-&GJ)P
z9=iI_3l-+sj^bHePI@bg&8(c37AcGE_3l6LQ*yDap{VpcUO%=kOM?Ept<2w_y5U^h
z_jzyci@j&@w$6`Dw0`q_L%=b1Kh~5`@yMq+OMLY&am+cjDCXoL?sc8-nwy<v>t~#l
zE=b5;IrGR;p$MB~mlxb^trNel5^}v-b9;@ezNw2+%pGINj&-Lcd`*kPwtRT~prqiC
z*r^RsrA{f2v+q5aGVQ{D-bV4J`|S(1?JwnGo~_z#-u1xv;*+0+=AZZQPk;Wqx^mvs
zkkxZG&i@mhvS?@cUK<yFqq;7>tgJWnC!bX;3BEZ$`=|ZZ$M=06{NK{DrQl|{;FBlq
z#g(~9^1ys4b1^wMgFosQ4}*UILogRBS0b1I%0NI-YFcq5FyAum^|KIL%3jCF6D;N+
z76Q}8^=rYsW^dDACZ1ps2N7?WFi)_kgD9K_6J_KAo85cp_kuMKtrZm7oIRD+KFwp{
z3FdL&af7Mi36}Pf4nXjfy_Au7+Fsglo@4tPhM!+bZZWe2OL0hf!nCjiD{&|x@w7O!
z;5@IH`}@x|zu{%&3FdR*gA4Nn3%Lj(@nrpEgJH(;1grX~BJp(nbm6=ZKOyxO6Kz<5
zp~B6{%@5NOY86<xH~-xepwIY%_(EaAJi)RtvhfI>YK$rpPd7#v&I`{-u+o-%yd7w+
zyqA0gOiNORXTDx~XFpI_-%B4Zoa*biw{hdLBxasq0S5u3fE08PMB)iM2*Y`YLhN3~
z%U?eZ)FQzmp#-ylCs=_+0g0!<q5<b!_U0^_VUq;%6t4p>+$5f087~=ll<)+rc&Q-q
zbi8!nyj%5xX`NiLpMkDd_fijo+4=BC-rC;|p;8>c!1CgQtL6#jci=~|MK(k>3Z{i8
zSk6lhNu_FtDw3R<ml~3sZip_DoSv5+T<)m>>lb!@@orG$iE_)s4FB!2E0p=#RA!K{
z0=GR}c+1lnmgU?>fgy&R*mzJ98&5EDV*4u)w&`yEk!QfnA><|G3o{AG^A`$&@feCy
za}q&$t+*(;SQ*3!0uf-q9DLrs>tW*j!>5>n0vQ5f!VC<_xhXjrK(2S1FEa}(8#@Ol
zHxI7>KcA<dtFxLQm#34ExUh(Xq=lumjjf%%gPy3im9m+JJT_4=X&G4sMI{wgbxkcD
ceFH;dQ*${JDP1EMQ89O~5GFT2M-N6H07=jt_5c6?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d.hpp.CB0E4FDBA2354B83.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d.hpp.CB0E4FDBA2354B83.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1a8bdc1f85839ed72c45e5a0853db2f46b0c5ab4
GIT binary patch
literal 1476
zcmWIYbaUIo%D~{9;#rZKT9U}Zz`!5~#Kk2=nF>q{4BbGviaC2HMR(se5NQ2-C*F`h
zI-|?n>(cHUdPg?i)P1|GRd35u9k-3k`#N+vBm1PM{ri5Z@3MyZr!a$vf6pq;z3{kq
z_2}aXYTRBsJwBG_BuIUFv^uG6^Phfx@qXp3FfLYYfn)P;c!e!Hs=v*m;KONMKJWJZ
zw?3UXxPI<4MdLU3>bAC;`~9AKDEfNB@=xbB-q7S>@AICL+<559o+IbEw%%PXl702U
zMw_)ok7l%747|Iwyi_Sb{@dmfj{*lXPAd&7h5tpx(XxrRx7;anzj3kS&a?Da>PdyN
zYMTP49{jsfw&tzxl88e+s@=0*iUe7!i-hLwIQ2H}g)oEO8VQXX7PDCOBhsS(@A;Vc
z>ae?f-ctViPns96nV4S8?6`PiQzyTD&5wT%tS1!|TsY8PP{H!;ckS=^bidtfuN4>h
z<?-h2ocj4%64U#}qd~UP-yc|>nH73rv&|+O&y7E~*|#(wTc3OXvs26`mi#SiM3(s~
zIL<XwFT2hvSXG|Wm^?Ss_fEWuI9usyyN;zX3+^q*3~l+Y;k9z_v#eQVnX<b#FEq)W
zt6>{#XFcKVhwkSdoQlPf3Nxqcq^iaXsb*g>6xNBKx<{qcHY;*dWZy>F-~_)N9-p*a
zC+VEnXMZxp^_p_Y`if&-Z_ZEsWuJGm>TqFFxy<=jCVA_U_w4K|+-2Mu()Pmg<fJE;
zJZ4VM&}&y-wA5sY*4Jx4UutaN-)OqbG`g~$QP}W9aB*dBQXMe8OgbKQCrl?mPLzqA
zlYv`BRz)35FmM3nfMAb?r{>a|8_uyYv2ZYOE3qqSfC&apCProk21ahiBOL5UH*j2i
zFTljg3DLr>p`~F9(|=x5Kz-3K&3*wUHkh=%roA;xx+pcRcs($+>Zbg$mSw6v#K^*}
zCaWe7vo&_RWloHmiZUY)w}y;{2F$M1K33HRp<*YXu#%CI9ZYq_;fGJEST3IB=i$~C
z)|P;&=Hb>A)<xp!3F{&8^o8}2cm~1-NIXMfLu8(up)$fuV_{<?Ib%6vBso)IQzSVv
zVKXG2xv)7B&qCM&iDxNniNv!Kwu1Azy_j3x_W!rw<>6KlRgr;38qjN!x(WzK=}GD#
z@$@D2k$47@21q<Z6+>-=K4TSQB%Z0HDUwPvNi!s#xuiJ~&qC4yiDxNkiNv##w1V?s
z?q%fyCoWio@o=jNsmUOC+G^Sc2uB&(8+sym#`eZ=-dba;fTMF0rUT>OUdCP>md0QP
zF>-;^sHaL<_rzV!bucp+Il(CtrV6GDC=5!GFyW^LtY6sq#k-k+Nls8h8m1X&lc2pc
xOb<g*YEB}kw1uTacs^k$E=n#|24w^;NNxa;JP0QDj$id_Hu2s6!^FeI0068;E-wH8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp.62AEEAEB4FE47662.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp.62AEEAEB4FE47662.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fb08a94c3a047111e1255e4a6e7d914cf15d98ba
GIT binary patch
literal 14888
zcmc&*XIK=;(`R~mMqGvkS;8(Z89@XD6AGdXK><B;PA7P04yRz&J3$o1j0hrzQ{ljb
zs3-~w1_V^Z1eiq-jDJN>Ma9J13tKnsr}x8sIG*RW_IFiVUEN*XT~*x<4DQ*}+Dk{r
z|JUI8)5p&a#X34VEy#c2v%@C8;B<8EDs^<`Zz~u!{kMA_hBu#nll}fIV$Mvv(<uvD
z<e<6JQhdib?unY4-{0_Pz`bRy23YSsG5K&qUUES16xYXpg?F%?x#wwiSIhN9D>lBX
zE&sT#U)+YpJ#K6lCVTfeoYLiU+T*;>6?Kn)?`5P`9AC2h^$3UX6T{B3Z1)3|yWHO#
zH|gHK($4OaVeiFy{XAz)TXuY<BHvEe<;Bt6S5k|obzi@-VBXycVIJk}-e#)8za_6c
zZy!6g&4oE5uT8L+GAY?|V*Y@&XGb-xX*JK~lKuFLI?eWXda~<MNYX}~4-tRlubLG<
zq{Eo=UT>q~$Ihs+9(Q(1uhY-xs7KX&ihgr{r*%J<uyJu|rVH+5-JWr8LRaqYoh`54
z&A;z<qV~$KCbu%bb>FzDa`%v&6<J$@RS(Lq4m$a`ON;Pxzir-D`NrkaGW)Nq?DevZ
z_bj>7^OsNMdo8<N%)ZKu*)Y}8=;u!so$LE7&`A$o=Dw!irStQq`NWFt<!J{Gc~8&U
z#&<cVelg^P@zX;OLVS$Fit|Q?WmI2Pu0EPSq;=1+t%|CbcKH1JoB6KI`WH9<`r=Hk
zqvz|gBlB+83>#Ejac^IN)xq(v_uM&>l~APrq0=#!&5H+?g@u%d4V+p2^9&q3ujBda
z;;y7u7Wu)sp+z?oeMV%@n0~E`<A4Pf*8>_d`~Dqzyw}d|FF$;CZuofRWYWpx$1}#Z
z<lEo&_1L_eSvz;b+g0xV<spRyAv^EQTiR)newTf#r#YS2(o=Cb{j=*ltA!84rp8x9
z7k4{e72Iv~^UtOgch>HEzp(SOw);;0UOb!qzF~q-Ny*^Eo(7{vjjXMfeTyi$SwHE>
zfnkpmMx_-^Sk%ow=FcZL`|ny&c6{WZ%-ZfDrNeinO-es_E#cD8cII_+Zv7fK;#Fym
zL7z?zzo#Y*-4W-~-Er391^E}ZT?p}vpU>?5xm`lvs8-E0SH5|%x68QB9b2{NbHVZ4
za-$a2^|_^Pe+|?xc($!{W>7%o`_9QPHhKP@cVW{c&!~+qEiVtaTXW@Tmu#DWIn^#3
zJFXRbeH%M^xPMact$DtA<q?G&a_)V6aQf<~w$tv8FaCCBWbMQc1&78ca)$eLFdiBb
zb~CJDO7@aJCaZep&n>p_>!4PvPt;Yt-;_7kDb+GQ%kVwRjoLTu>V2QlYleRw)o|>~
z$*zNv&rUrP7}3mTz}1l7F4x~xZR~2kA^ggbP0O--W`r-W{9m2H^`~75CzVdU``v3!
z^lq1$<X69_l4Ck7jQ-z+x89Cni3NY>9;@1xdM$j_pc41^Z4P>Yu`i+rc6&Pd==Rm?
z!mrP1yYLEkf2i}h?RxD>Zfu$reC~*6+v3SZ^LzBVz4%0E$H?TkU$*uyobT^{+~Buh
zTkDLL7XrPkGS8n~X*|sFC0?4w^|~fY-y+6#k1=f>(#pn3RPO3Ogvq$uI{KkpHr%$I
znqRkVi^+Vx-O$$J*tIA4q0!q_p%V=~BFA=}mS;A`ILN$(adnyf5dSTi``)#AXOeGM
zRe!AZOK62#UW4wp%O3ZaoGSTH>vLp=V3Jrcb6N9Vtpavj<s#3OTo<C0e)ZPINu|0;
z&6an(G||Lk^o-H1-?nXT*m3@`0@EpX5>Bd}^X~d>b)MTbx|LaSK!Y=rA6M2&eDTDw
zq%_WIyTuFN#aG`YxHwb{tTfbH`1G%7`u2Ti&UV}LX7I5sXF5C^-Fk_i?Rp2_P0P>g
z4Xt#tUpnmO%1etfZjam)mSE{pI{RUj(-+kMJ@X!2dM|yv!{OQ(t0{e!%kM`kmz>>u
zZDYWpZ(p___P$YJTk`(qBHy~^Q68_EpxkEGp}(h|65Hgc%$A?o`St77n_gQ>Mm3j>
zwN{v?T_05HJGIrq$kqX`8-D7$-QaK3d{Wy>r=Kr*vbdMsJF7a6DAPAy?;g!@>#1|`
z>zqITzU>@1%GG8>?4vOg`(2#&UcP2rm+iejJ-W9MC-hqHmfOSdRP38R+^rW|``)Tv
z6YQoqwW@CFJe?($GmV!Sceki+F!Xz-TR+NkbYytUK;h$F{=~wMw}Lhn^jK}{*yHmM
zdC$zSquJ4+dolth?mv~Y@5<@(uhmb&7fc`Ls!O*1SI$=s#_WibqjoMjVm$5<k4I{W
zj)qPCnm-uO?bl!T50s<6MC1dj_`s^Q1Vr@Xic`WyR5mk0X3jd|du;d~+i3}i7!7VW
zyX<?tn;dl}TE4J$U)XM10wN|}ay2_Se3l&95s~k#(|0ylOF%?5tWPulxcV7-h$RjE
zD*M&A@dMb7(|mhPoa9k%f|P_k#bSy>;|H*vE~!%EyatyTp_YW5!W*XWEwuzhoMk5`
z+-z%AEJwYG4ewahJJwc9K*arUVDf_DAI`{8dm@s6dnAw{Ya$@hZe7WzN8_fX$&ob?
zS%b~iV0$eAk&ed#oByQ0A>0HR5nCphPf$000NW#CM7GQqC-JCx8y&Gvvpyb;A3Dtu
zL#Blhna)MaHnmLqEmMv<kS0E}_McfFEdd=~i_(+!cVGNKj#?9uXzUk_f7TKZ3G7vC
zIO=fXYdyrf>WI6sd^a}L5)kQiEx~L;SdS1nYD=`l;I1)vn3jM@-?utTVs_3kkt2H|
zvKE`ICEaZzATluHMB<cp(+BFI7Nom{jI59`)e;aH<gnU$Z$QgxJ)}=WB9JTsb<z?L
z87+)!pR{*dPdRcT4Xwkfb-1;bfJkV_FQtouA9xxfbD|}cw@u|cX~{n#8?HV)H1gnj
zJ;V`_GmPvEW1%G=GX7BEIQQ!9yXB}AX($f6#$i!QKxC5X$St)4l!Nt<K(w4=Wak*A
zmVn4)-;7NO4u_A*kqyxji(O-}ua^8HlHtP4y5*ygqb@`w3JXy<KubVms*mriUe97a
z>mf`U%41}CjG2~z$lPA_$7c65-5^H}M9U`Zx`~+EL_lPT<sX~+)aY6nA}cb}DOjF@
zTWAT0L=1A8wZo)&p()~tmK8{M1v1nU5LuQr+hw%=nN%L}#FphqZ#hzE35djeaWm@s
z_EUo#c@Qm$*fo)Kx`}|u>K#E#zncfTG}-bB%U=;&nh1!j$?}g4Eq-`Ojye)8kywbt
zJ+uTw5^u*3H~)0m!3e2{E%h9(=M1$3M3P=qG2O<r-7ZItL`wtf)IfHWCITXV?(&`A
zPJW?C4>coN&N8yI#2K0hh@^Sga3vOZe>FkIL}ZfrBz5Blu#abr$*bHvt&bj(5%yTw
zSfj=dVCVFGVlcX{eW@G;6C1W+r)}6vOF-n5iQn!4-W4N^kO9#!&v2e|;|H+w_v!{j
z+ih@{qadPTD|Xt7J+%Zx3O*(qjQ;4bP>%W#kvi73j_s@^AX51A>%Zsg&77r&<fM^&
zMwU<9sEL5c71zJ>hF=|PYJyCN$Yk@$>c$UX-!RqRpc`|_*$}A-`vz}wgYTdvAX0Y0
zW%lLw1D2bhene!N`84;&4`AQDkkcYlcyZkj84~s_-tZQ0p(P;luqfDUdaJ}&JTfIB
zJF(tQY@{V1Qd5|D<dIi1L5@0+&Thu8n~8@t5fJ%b>}}Kj?Ml7m$eoC6!A@InH!T5?
zmy0TM1{of%*F$=wp>K@r8{?rRAo6$e(n^c7%VG>0kF|GP^LLzDY$71?w(9!(-qUwj
z$&oE-Xd`yrNNi~$Ao9Na>xUL8A>#~@B@wyL>fdLLv;;)z+$X=i@nsb@K^DZ8P>WE<
z#t&eBeEiffyHMwm9Jvt<N!Td~`)CP>e5qX4^G(7;dpYVwMCw_?de&S^K;-M9PC20)
zyH=PW1!-iW`9yW&2e7}n+hxDpG2yx#{X{gpXI1aX$=5_cL?>+ggz(#KNKD|IcJS^n
z`6WgWr|V8{1+xyp1QNtONWKTz!Hy!+h7_=Uuwje3dgedzA-}K3gX-}}2q4(x-Dd9=
zeDGrh@jhyQA9Vr^g7^Tne?a33)V@N(5TuZ%|3vaXks16@rVUk~X?^#VNmJt6-7^%#
z2~HE-;20LfiB1#UV2A{9lG7wNSoeZB*=e#HEDk~3#|ZlvBUrS8coun_Mc&Y$Am$^d
zeB>tK0%TW!oM`M@fC434h`bAtpM;B$cM***BBzVAzG7roOzSH~zQwe@%gFmO{reK+
zT_R!qiRPUr?^mB;u=t1<9`SNmlrRU3OdGmE4ANrAjY%_8ghOPuHgf@+xn3X%F&!E7
z0#;eTwue5*w2>VQIz)C@5HGOG3v3q<6vPVFv4RbRH6n;rtVI>u4X{ib|A8X2a}f!d
z6^Y+;F%dMcV$~q1!M;_jzl3X8^BS68!}`|H{AaA?GuB#?f6iJyr|~P+{1yHCTGqT)
z!pd8Rb-$(^Z;!Ef2oE@f2g3P^#Y{XR6OWYe8*KIlE8)z;VjX8!$JxW)0ue14bxk-6
zV{TqrFiYo!_xS(3qgL`r&lugyRb=ubS;t7WBTSujw12Q|`r{d7@=tL=r?|f06@qw^
z3%DtHiA_$}q?Zxi(Zt~*S=&gq6=<;aIhy^p%Qe}*9zzKGv8*xNhcOhyaEEX=*qj9M
z9;17YX#tTaaW&qcidTaBLvlbSqA4=~HINdJj&4KpZS>aClpla|I{)uy!-0X~A7rQ!
zxPSyM7)HeP*RNeI(C_C+N7r&4*K!`f(BR=~xqr4lEUx2R*KuwVe;wC%9oHW)q#neS
zy-2>7-e8*&5HQ8P+SH>!@BS~ulx*BO8xMY5|H9_bc-<C|nGg-TkbD<v36mky<|m*5
zHbx|C%u8S}1mfO9GVw5Q?@l8+nRN&p_?0x2iY-#HBQW3x<oBBFCD?xM>ew9rKLl~C
z$=Co0nIMTF)|_XR=h+UDHBA`}SmSxK(=2^^zxzZm2J6OP2gphUu?Q=Rum|+d%ku9_
zQ~A)p<yfp_%`2q^i^Xczyqd;OS<9#Nf`7@Hzohy1dEvgaJ7VzxFFc@e1+S{$TY~#w
z@gc8zNaIRgRY~J2UR6coNBpoy{0Qmy)x1$PZ!F;&-o1uzFX1P=@I-1Kk~=aYS6Ss%
zX<f>+DIN@+4^wn;`<V+KngxukBGF*TvjXDYKN}FHs3010xS$+*0he(BWzqr;xR~%_
z@Mk*@BKQcmeT02rjs@`vZu>;SJ&a$q*ri|>62WXPD4Xj8ek+KjTtKO`AbOr!UOVmO
zgg?nKU%{GGuq|O57Q}RHl8#$SE!1!=YB)1c1o<i}(`Km<E(#=<QKn#P>4$Hy^Bdd&
zw%x$eZC$sH>=CY65R5@Kqk`ib9El?gBxIx3;DTgSkdQ;;lSnv8<2+=Thg?8jrcH`r
z-23$FY<8hX;sDa(EynH^;{Y2EWXoifb|d+2dI>h=%%A~&U?9_G%CM#c@jKJvJ7WPM
ztY&6gL?R2ASwXzaw!6%B1JeZY2CKfoQe7)&)#Ws<V_Vd*7LY{<#Dy%Pu@m?h849e-
z!S=Ao1Th!e<YHF|pT{=mX<UG93TRx2I~HP5lCQ*yN^B)f))UU~38#ix9@M-??)347
z>oh~hH^0Rjz#@PSl1-(lJ|K<AQf#~w`%KoR(}VjOIX;W^=9&qlCZhkDRlrUokjO;;
z3#)+8g$yOD|D9Dpgd&jiP``myK;|fj5m-M0D<B^g#HCn&DONy3nxWE<#tIm!p`r03
z4(z(;MNEIrE1vVFU^<j8$gC$Lc`~wq$&qOb7%=N&l3u<FG_2cA*2XK;=@kkBK|!oV
zooZ?P8g+Ut;W5WOr;T>7w<AH;a%8a_Ie|u^2rFM<dsqts$zkl?ac&TTkgc0HPUWqk
zbQ1Dt<wX9Wbp`3ePwqeYz!(FY9FR?Am46u8dI~vyn~EvGAX^iodzEPkj;tAM<L|7a
zG}z~`?K#{D`~s>jWQ<agJQdkM*Jaw;3-oB>+F^?<#<e^_EIg0h&f`$f07V&2rY*%t
zrd(%quQR5SDGiL!z?e%*XeX=O$vQxb8ox3A%-VwAKq-fqmV)Fd$QA@;+JX+42980d
zuZ&fevHd}i%mQoiiX8%5olINifgeJdhnU_}<B?21!}y(H`h{wTnIs&H-&v*~R4xS4
zRljpgKj=G2Kp4M#rXM8yGrPQ=HK=s9nXVviW!$zh-B-&Kvs*37tr5s{xt(FU(~8~B
zGTmv#Zs(Zpv|_h>raLGmu>fk9iNtMD2Y<i+(?KN}8n$Z++e^a7Ip^b?*Y5mRr-$-$
z%ER~FIJI$Q$nRv<D%hY3_9qyMx$e`gCQf&=B8<~q&}ptWYy?o(BBP#$<Y}}+G*!1?
z@aFaZIONLytO-QJBi`T<?*YqhetuQWLVUf%M-Ue%7ARrY7sQ2%g%nz(SVW;1q>ho=
z62#R=y_&{rka`V`*CO>=8m~j@bu?a&)a!r5j_aj^Ll9$;I+o_gBF9*oABWU&G=CFP
zZ=!KLQpeLc0jU#cycwxC(>M{S6KVVbS$vQ@n%IN&W3dsqyg=$1`WvwklsW{m81>_@
z5v*2$R2ua+VI!C<fh7C-@z@Be0UE4~#}*P!!1@W;2!dG>En{T@wvg~<tiKr>L77J&
z)kFP6Yy^cxffRD|w_qcv@d%_8qMw9~ph`mA9P4kzMo=~*VLR5}hK-;QM~(na{e`oE
z`k^3x<J8}1{GC&Or*Q+PZlG}lua4lU-$(N5$RDv|B<~{WkK)x)G(U=WjH3C=dG&Id
zAI+<yX}p40ub}ZtUcHjWF}ym4#;bVsDjM(Q6?=Il?B9a8pI7Xs@c~|OfX3;(BAvzu
zdBs5*XYh)QAF*WyZzt(L%qtGl{KLHEVVa-GD>7;RQC@MB#>aTYF&by_iYyu*=M~3k
ze1cb;kT66QnnbR#%ndOj?8Jo8RNRK$V_EW&g?;uE6e%?2sFwBE9`+v%_O8c0B%HvR
zByg4z-prY7rg0)?l1Sq%oXHj%r*bB#G*078(rBE`sj@j+Nq-Kf%AxT|PIZ#Tr#RIq
z8s~DVTpH(bsyrH>=2WL?T*{f0((RXVCS^1(=S<3Je3LV|DPb7EW-{$1fOKa4IKm{<
z#df=~GsN!#DKWKqkK0Qft>Ze^aXwN=vYt0s&znhWD3`a+<y(V~!CpWPhW$vsU)sE6
z+N%bzY;`Yo*{k&pU3v;)s7a_A3ax@T#bk;a>IUR&P%W^g&IZ*&3N2DCqR?t&yP7V(
ztwFYHXuKBLuBGuhWV?>W>yhhv^iL5*v-i8k(l`#;#?kVdknJWK$0OT#8Ydvz1R8Hf
zwwr03h-?#Sd<_|0Ln>Gbf_NPnU8nI4WOReZWyq+E#^uPUoW?hi(M=lPMn<=3d<Plb
zq48a0beG08$goB#sR-gzWc*Z0iAWTUl+TfsWbX%L_W`w$a6PMA&l*7aTXS|PVzD`W
zb}2SubNcL3#9?##>{4vP=JeTR8IN1hXO(3FZbhF>ip|)ZKARMY*c|*(b2cfqU~~Fx
zQY2w>`fO5c#pd+cq}YbdrL!pmn`U4O*uw-dijzliCek`t#>tn__#0>YjdPOL$#>57
zJB=GS+Xfm(@U{^=B_GMVM)Dn{-$(JTQ8Zr8+b*Z&qj}qC8n57OSI~GRZ@ZGlF}!UI
zjaTuut7yERSM8Tl4?#S@s}9gOomZvP_#m%3NaGCNK7)6LeL)Zp^Y({noXM*)Y5AkP
z>L`tm@v37q&f-;BG(OI&j??%AuR0-NNU6w?(sX|bXU*!8SmQX==7%I8&E<k}xqdK%
zkVumJ;5w_k&JK`V`59~Tj2#I<u}pj22|L-^>TUg8my8%sH2&oKlRxaV>sA}J-;w-V
zmN_Q(4(|EL7xZEA$9o42Zc~7|(tP?JLc<R#KqDmnkM|N-EJXf=Xt2cp@ty*UMaaL1
z=3hka7ioKbytlw&F=|sx+w<c+hK3(hOxyG0y#^L9Bmc{Edq3WDV6g=Gm(cuj<X=wX
zo5=qrjqjkqJ7~Dnzq=^#E{$WbV+?kJf)OV59_JYBD&ZpBxd?kgDoVx!cP+x*0B<N8
zi>v>#C?`i&G<J^0B9sgSu@L(e;$a{dx1*J)a(wZQO!P|5WF_YXt`jfrD%(=hI)O0q
zxS%|#Ko?)t=8Tw^@`5n#@X9;9J!C)eA4A40HBgt6l0XV}O~IbhzLJ7_r$}X7Qd_{v
zTDn9~i#=+wANY5|#h|p(=s#SD@wvQlE^j3*xH8_jjJJZjVc1B>f@r#t2R}_}cXaRE
zsCI{lhLc>-NxC*z&IOc9wZWtpB~uNqDf^Mm6kxXk90KD4B^k1EF0#st^d(XaYf(dA
z8Z}+<!^+v=_t4vNKyUs3{4NC&Yf>m_dSL*ABZ$`+-D`BAq=7YVU>%{$B&El;`BI@|
z_aCd*UzT_7fv{MHe9NTEJuKcvJ#M3(;KV8Z_bmL*ect;2&f8!2UXN?9ta2cup2`KJ
zO49+CprqucdHEn&d=w8kDml;|mjR1k?o7})?MH0z5!-;%rg!k_U$8x7GqGSNR`0})
zutSofByW5}@|=UWHqLHXj{=B>?~K)VW(bsKGQw6&+0t$Bd%`Hh%0k=*9OCfk+wFI+
zF)$+DUCEkMk{fI7O}jv9{U+5ke#)wzvKA8mC2R6h;va2&HmmH)@o-{e9cxv`x=4l4
zT<n&My`jZp|66#>Y}MO+|N6!%%i~DmqU$e%$eq|7UU5e%6T;01aS_c+5^1=1W7TeK
zBgN0>u<{(X2E#NCqRf|Yj$HRz;qaP1|GCRUU_*vB1IaU_NKU4G0|Sfnq_<6rnw1~c
zk~;hmBpgAe;KF3bK|(H#lUZdl`%eLmTtcvhPgrwla<e$yEY1iV5W;%mz?zpe(un3U
zc6m&9D7&3%9qjbiokLfMm-W@}+ZM!5T__)s-u2z%e?8-;!>VgXe_hz0EZIlc;*nJR
z#^e&k;)#Uw_w=8k-&1$l|GQ)%h9rUqk^CUCgb|Zz9|ytEp8qgmNJiM1ouq5MU3v#X
z*9snN9ru(!Ka~9Mm7MoV>6Q_^pKP={kbDPHf?`2Dh76A(bHFm~6C-E{3WLOXn;sTH
zQ)I0(x;GdTsrtGU2}_X*>iL?gyzv*-2CDK!&yymyKl8ZW42u=KP$AuVX>QqshcvF_
zg-RM%@j?}it9hZC#x=Z9L*vK1@L0mdVS5LBsh%GGpRb$Ch5@&}7`*iQ&kea^w`k?k
z?t=`-CR@b@SFwGhbymX$*U<Pm>+qayBk^ll+gjEEFub23k%8uA8Mww(Z->PL`>7ND
z^#Z$obIvel>%t|CWp3-OOk29lZGDDmOP9H=&oXV{G}hG5t<N!S=`y!<KGPPwSySe=
z%|z}}Sv!SwNnyPte4JAs=h{G-8$uz{(WW<Y(ACo3&BGGC$JLMpm&OH<H}|kb!4rk%
zvc}hPATNl22!GhZ{w0WO7-0=V;q8pscE(lWr!zu2&0o%{mrHMQNZFQE$I&>AwN7I_
zU@;3~7OT#Z_~mXd?_~}Ni6ho-VH~zF9ibrzE=m8IUhsjv4|9X>=)VYwA-_oE0urSd
z0*-vLFEl;qgI|!aly^+w+rrxma%aL@m+_7;gc|Hr#=8Tq>T=cp;{he}ubrjp(~R^l
z;t+F9=eU3^(p1za|8m*z#GozNR7<e31goWp`3hEEk#Nn~Gq<<yG_xZb3K>-)(*gE2
zC=Cz|`;dGeU7BioqX@g@V|VU-GXt~LL@<?ap2{nvnED3a{06UpUO#ooPVZo`e>Z9B
z5^i}3w*#A?%0M(6K=K20Q(GFJJVH|tN|7_L>7k=E0TI|Z0*jE9Lj8@*ZPPnP5PZ8W
z`?T938~cAf@=%E(Hvfg>f6+EKJ(~oZVc#I8G`*Pw2GsD3WsQ#{!9vZuJjFi70^F3m
z6hUq=$W!v~^(bgP8Ytmd6ckJ2Bgp&+vX}U|$UK+E$*e;%>n`!1u$E8Q)?laRWr|x2
z_Jn*;bM4_7kNxRVl4k<;mr6-R*rx~wOCekqXPU)X!=|9Qrf~bhc|t}-F8TSONPeJ%
zqxhgG8t>&D_VVtM7iI818GNvW5A!~UB@87gG7;%Wo-TD%rhQrp<Nc|_to)#Zh2sAk
zxj+$)Xxxe9JEgdR<bsV4QbFUFFR7>9?k8v>`df_dEyf1c4!pl4cc7a0snXQ`n0jl5
znohknkAUdZTZ8G;8=kfj!?q*&cB$M*UOJL2fVBm~!p2Vv4Is~30}kjtGJd^e7z6n)
z+^y??B&0Puu(v_{r~^iahy4<ssrmOa-kOi5zwYowN5@b4>M4Eo`R6MV-A5QEFRFfc
zM&2{9hj%wWpY8$vzMfsZbb9<FqeK2FC;a)t7ag6h($_%gtFQD`(?#uHPxzvv(@pyN
Ee+V8*9RL6T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp.DA7CD25895EF8940.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp.DA7CD25895EF8940.idx
new file mode 100755
index 0000000000000000000000000000000000000000..762868bb73294b09ec3c1f74bfe42ed1b68c2e54
GIT binary patch
literal 16464
zcmc(GcUV+M`!;*#%sK0Fc7<J5c3~+l2oglFVF5*95zrXyMvZyZ*rTx}#ulZiD6xP6
zktQOF9TY6s#S#lDih>mdv49v9)TpR@!@K)<#(%!wmG_@`ZG865+&lG|dCH6fLwffP
z?rCNgFfL^7wDCWWVa?3UEXjYff1WkjmNPTU*P5BlO*!-J$mw^ytxD<_VTW#`?e3mC
z&~aChI>*CFt~-2Ex6^KEM0%^+E!r$+Le7N5eEi$)UdZr1cfHStoya&g`%!s*|8`|Z
z-<ISaDcXGC+l(d8@x|!i)o&Z;6va&ct#4UiM`uakroOZO9J{_CQfBY|A}3<n#s^vI
z9z!0Ku9%e^JWD>^t^Df&*_LIkm8&+zmJQ83Kg%||w8xF7M|W4Jd{gcAsmkh?#z7mn
zj(;o&Em}RG_t1w$J%?7b+mj_1vPz<sbS_-I`OkuOGv_QQxOX~oyuq#us&x-(-MV~=
z>%>Ffk90q~a9pl?MOv|A;=~uj<=>Bckz9T5q2tMk!%}u{`+0xg5bt+?-B0-XhYej1
zZ~nWJ)2qX+#w-e)*QZe0$E7N2QAlD{RId=trTN;9I}2MBZJY2S_qX*YvpdahwQ%Un
z#dCg3U#Rb#|K!54XU?r6dvA*@O^F)Y@7?Q!K^5P8_%`^0=j4O2Bg2Do+onA%c+%_d
z*5#J`&jEh3*42HzyVT`Cr6RM7+JDUMh`A>ktbcKS*#D^ZNqXw+DvJ#>Zq9Rwd)a*C
z+i5L-{BR>7^vCw=b|;qg`DX8F+txR_FY1!=gY*ew{iJ&rpFVHBQl{UoOe)xx`S-_i
z-JCmJlMA}m?0eBUD)(kx&VSar6|7g+<aE25anvnezFJq9xQs81ne%wiuD+8$l_(~s
zDQ6C^pBw9!7iT|o$ner(_ZPN$TQ;}UE3(_x9*eBTEex{?2nyfA2X~z^c!)f){b;Ma
z9wR$GlRwA{$@OkFxbto4>#D20R=4ZiZL4>?>Ju4}f2Atb`6<oSgC{J{sI*A3^Di20
zb*8#+S;H?&Y|^hs6>N_1yZ&&sYS5hlJEF>>{}^>(fBJ&CQ(j*`d+Y9*XLrXp4((9q
z)9vq>V?BBu${y){YldNseRa{aluJkRFOQn@XvK`_em%zhSaW**uExR5hqkf5^PA7p
zxIL4qrf#!s_+!}gDJ??3wKx-3;*-T3@7#O5Q@@0<^DlQV-BQr5_nGb6-Ei{5_3bK>
zyG)*O_F&Y-!htzBXG@R2vr6qxrx<!)t1SAecGfw}_qS-c7PIi_`4UfQg~VQOeo`0Q
z;(N1Q<}2I0x-ySH+4XX_r|4OFXiZGP;jFR==F8G%N-kzbe)=o@%741a*4}&WXgjd7
zvXbqz<6^%N`N0$BkEn~d*4+NZ{ARW6$8hO{tJ62Fd-`i$pP>)eg@sCf{VwSUIzG03
z^O$1WD;uAroU#0+c*60}h7Z2?kFRWbYeww6w_DcrUHH2&a&Y<L)|)#{IO4VCe#>q*
zoUd%ld+2zucyHta%gi7A`?jCxmT&#j?@@`u;c?X(>roFgCVuy7WlZ=?N#yjP{FyBi
zW?HptyKUs-%PBRJ9$r+}`?t=pa<^*zEJ7G``Pi0?X47|{FrWNEyNo;8G5P2WMpNe9
z<+RU{iA=-1epAxT8b>9C=M?<3_t(A8D$EuwxH)E{^Y_8|mo38b{6|do9+qx@?(NWa
z+eU9b6Q3~yw+MLqylRU3(wqtHx*zDVMR`HSI}cnQ@_w75tIx%_$te%w>t1<}8U9n!
z$b{Es?+&kh<JPgo@|RYJ>NEQco_NVJNcnSiX@vjVlDUBo7~G2;Klk~&HFIyi{&0PF
zQF-0CQLz&eBR4v4*A?|&Iqdw7{@sI96H@1yC0`yo=2L4)V4Qc?p2;=EbLw9O?F#VS
zIzH)f0=|A%R-r7vmuPth#YUYzu1fP-JE=pjp)&-FPix+9vF^60W?=EH@XR%uzb+ng
ztu3sXG4u5BpZBiJ_;YCOo~)vz*gc^YMYnpyh5OdtJ7ONdm9#v&;i^^HK!;V0Gd&Ul
zr~Q=C#czJ`wqqM+%>4Dl`Fj%c=eAYa@6NBCxpto>rKEAnh6RJ}j>~Bf(&qN6ba<vX
zwq*a4=)KQBZuqN}TAy|@Jmqm#<G{In7KE((dU2esMcE1UR*$;SM}hm9;4Kc#f9EIU
zZ+{duZvC>=2`KOs+G2KLwz<RVeQ783lk0Zn&rBH>*UsT_^^WiFSN<N-p~aDFuYX)D
zEOsgKeVX|~ZSx^Y7W-pxcAXb0RR*!#6vd0m$d*9`UCy<)vbDUNJuhcO{Eo4TQ|&h_
z@9lA9-e7f?qy7cIDO$hi_z5lWoi^EQ_Pl9hrOnLD(4V!1NAm)%m@^ihW`<}ai$<MH
z!c0nR3Hc)%`7D0^oNNbkMnNb`kZcL^HVL4}T~|-+=A6FJf@#^#%n-*Z;y6o__)Mu<
z;?m~mr722gFsbD^WAU76Z4y9f-p_xX$EFo+6pY@>%&;9>Z^zCi0ThcL&)Jq-Yp_u;
z9SP+Cb~}K(n*>m-9YO{qHq?#a7@jl~i_Bw@$|Qhd>o}lh;E?fmEf^i4oZ|JT_)aDP
z6o-1VYfGNL%TX}Sq@f(_k%N1g1W;Pl`%k;2JhofGd_ySl*diX=m;_K-zr6TI?BLtK
zsu?wD=&{7+u|#JQK=C;HZKub5<JNJEmQapi^J7?P5<u}%x9^)i?5Aucqa>6!5{owy
zcas20`$r?Hk2@VwDHspZmz`L<6Sp=Apmb9HF+b&2+9d_!O(=OdFb|I~37~i{dO!Wi
z+y0K`j1_6<C?h+{v@i*vbZKBpe)f9&K*97QwH(53hj2HO07|#m4h^+mheRnDCql`_
z9@%8-i~=aV8W%|3^gDN5!F)|9`PeNVcQpy1_%GjehbezoreJ&tWg`|gVuMKl#lXxO
zp~{xbR4{IYvKM>o#XU^|D8aeK?j?_FD$E%f8PQ54TZuF#0hEvto^~C2)V)$LuB4V7
z*m?)iI->weAMZ)a&&AY^Q#012mZuV%r(_w80x12`nw@Vo_sVK>MoK837}+Pr&Ln^m
zs(-p=Xx!Fi7K{y{tl^uj;dLefltIIGEX-*gp*Ck&LTO-R4UDx(0A=v^!N2d=C#*GR
zED0qH$-<DeNdRSNn6urzeotp889PEb$y=P{ZA}6w!-facj_LK?R0Y$840;0=HsC;$
z0LqNJC(hikZLwX+*pgZjcs7BzFbSaiJRz^y<yY-GnlsG_rIC>}G8&Tr%CFTc)&zJ6
z2h@y)P^u(0Rb=Xn0w|07{Azp1!K%oDu_u&TzGW@n!6ZIYDhIozFJBR?U|NyxreL=e
z9AXkc2|MOA;6St41?CJQwNy%El@cqH0803fk^wQ&p+4peM<@|U7D1-YD1Z|4WA~8@
z{eR6?Fdaxk2eI2h>}L`{S+hWx<nX-ZTm|DpDC@D)dU6CA1yI)hUgkB>!8Mg*gtlge
zixU2#q>D+IHDe^^W@ZwDB!2#awe`2HzcQYg1z1slEe%Ei^`stp?I)jd=U3%_HD*0$
zv7U1>7zI!^xIUYfnwzdSXXIo|_a(CX5~WE1CFAy-tf%gdpUfGa9L^`O;skDC5<tlc
zU*s1w{-T?L=|skqivx4X*<chvIk0!w4gV>7<O-$@p=4s`OtRUG;xpy&{xfCITC6Z<
znvuRlB3UGvU!wrZ@oO<t^aqw+Q81yTmep9h8tY90C?_w34<Ed==Czs`L~3~^v3Vx(
zG6|p*_+44?Y0y@Ug6T$T*^fQ;lM}`$fO7uppq~Oejqa{s1`<jfR>fhRNdV=NO?=kk
z{fB2On9hWfh@BJ3S{nsWE{{|_UU#qk!xjurC~J{)EwV5Ppj=5wUR)9W({cslPAKcJ
z(>kIKMgf$o7Ox8v@)KEe29pteWMm)7*<chvDRH%Y9%r>fr(nJ!wWMOVRHEQU0hGIk
zo%rV`Cr(l`4y2ZHiCwwG)g*v&@7&hPhJ(>VEf_~q%Uf3dmQ|VrP|BtieplPJ&mRh=
zJ)vY_pDaAgB!E)6rQm_zKr3=Kv?i22*kcb_X`=wjlb2IZwi@JGr(gyUN-Xw>#hp#!
zGv$+OXz2FVxeCUGP!h0H0+|=107~7PsK~mI<~|ChKcTF_+BHNOi~=a{y~dZHY-iY|
zW?GOzKatoxA;*VN0OiBxb#14dIc~3Dx{_M<VUK-88H@rbjT?40bh)&Ap9P~LwY=gj
zUh#G&0hC$eC(N$*B;JmD*6!QG6c;3dp++LDk+cGJM4K|K>*u}kM}pxgYX1~<0^JY{
zRj7RxjjK`nY7v8ng=8kb3Qh;XkdCd=u^(7E!BCC$)wtj6Fqf&dmT@uM;ZBaefxC~A
zuGBJ4wM-AW$%O)kjMQlK#>5)auvHrF38y>h8rE0gZ=gnS(<CyJa|X_b+$%b1N_@LJ
z76S2393D$rf#V{O$#!@mX$9**1}|}VDrp54k+dLjsFJjTV_q<9V}xyt5?mI+a0z)|
zLcZc;T|sVFkf(^RBB!g!jmEvMq974pL%!FLzle*GZ!wK;Ah#QIeWl2$l&-H7^(v+7
zyNP^n(!aljd~b<ZKGCN0<Q=*b64nqd6~d(oP#u_iEHgQ+;+#hyrwHT&#zi0}o>L_9
zIbb>r1;bTVca`l(jeyk^i@3Swy1!R(*^ex1nBhI6^MM0FDW_)d7{VH6`ONYa(QKdD
z6#B*Im(K{iW=1gVLW*6;7Fv>-d}nZo!FMLagGg}@=|!T^kp`mWkZY4>WC}S%@$xwG
z=>v8M2)xYXP>YjbJR5-^5c}eE0=t9h5s20LYCYE;DiREFeCIgc4_qR_u#<1TleY$+
zLonRt9dGlk!BMv^2usK<OZ+Hh4If#XkE{+(x+dJ~BO4&%Fl-Zssq!$~D-2WRQP@5T
zJBs`z*nSC(mt&jd^zT<-n-wC~l<tv!$U4v-v4&DSpcD@TE6y6?IqP^%16zzGBXo-6
zoneI9=Dz29ZaHy=%;~7Eqk>^!ZHmv#G^_R<|E~`WW||xg#&`&p)<F^8b-A>#oM_ET
z)^(+*xPqY=I~3zKFar)%uKSzKdEiGz=x^pP`)}w-{(?Y?{C*Eo?4h;J7*&DJI4%tc
z(eHZn6CvK>f^Tu(h{{sW8R|tXY<Xzbq?$0_r3_1Yq+iLl0zGC)k8Fyu9rQ@=mzQ7H
zrAYRV7Yiag(v_V^u@h;a1(_+ngFeGC$;eDm9n6|wsAsy@Gk$PvxIg>eVqj3uMPzbv
zIr0gC!Fi1P;A6-9RzSA@cCO=g&KnAv@VDE!5hC8fdF<dkY5dI&u0LQn%2}BydK9~I
zl6O7Hw}q%cFudd)U-GWtb~j<Sm%JBXIJ(K0jG?AjQ3K=D!1RD11!7JzOW8<~jqJo(
zG6tY9OAvVy;y$F<hnz&BF%11o9Nx*gL(sqvWR^~2yVKYeR<cu#_pPbtU-Tyh2G*bk
z-GX>kAu~m(Vpo?U%caN;ETv#rjx3iWJFvxqVHI**h1x)PC>S2Fx(94)=$K%rWNj;1
zR}nvB9iOpnMErtvd_m(_Y#EE~z%L?Z4_jTrj?jNnKh`|J&R~cILj`VGfjvaLgi|fy
z97KL3ZxPAsU~wVpCNp>#DGnnC=!nb|d&3NN?XYxU*{L0`$wa+mtY0$iK>@p#bebu5
z_P<MN$imVr>;jG;@pG`I275zSy4k<H@FO?orJ6N-U~N9oM{Of((@5h;Y#&J<ozd7P
zT0AOQ!#iGh$A7k<tl>Q`yr*#kuWjIM!NQXEc<l!of8@0vY5a-Texh+BuWIDg;_t(x
zsxT>q!==N*rNhPY2&r|1R4d|0sdXfcqolS`QaiDHiPUxpjh9NjmP*@;<;$dA%V-=e
z^@x^wisdV$!V1xYfV2o%fgGgBp$|f1as<|}*F1G-o5Sl45n?QMh{bJ0t*XGD71$S+
z3bH7KcnB#DA$w6Pj42hU5j;%dyVkP0T6QoH$-cmL%kdB}hcZ)!1%3z_7Scj4QsknR
z@I#p?&+?hL_SMq?vs7h-Sjq*L(uem;&hSz^ybb3Qst3P!@+QP(T-#-wpQ!RNT-z8C
z8!q;YKHaOMFVW=hT)*oG1~bs=x#dO`nooXT!UdOb{a}j-h8G<9d_GuvA6fF()S3w!
z$e7==>i4WIj8QOL#OjOKR%|$&w+!d4p`xJT6rbb~fwM7jlbGfQ87&-$O<PKM$+Uni
z)pW+&)-$a{{E87?(dDlh;kAh2upw@{F%c!6R+-o;6YC+w77Wj@<{5T@b%qQT8S`$W
z*iCzQ#yl0Q0(j-*=r(4j#1`tA;q}Z&=xV>v&erDw69<sVJ;ykmV_e`Qg)|niBE}q+
zSVJW8jYOfMTbqilQn4!pt|TzP`lq-P^cnmf;!hYeT~N`WW`Reijkj1untsJMd&M`0
z{Q_o%c-Y3o7yLpXv9kY1CKSSGfppFP6B7!36-dg-zmW-rY?MIKVg6w#6f$LlzfrnA
zjrGON1QLJB6R_%uDKr&GoGxF7RS?1pB>I;pVihC?h}Vkc>#+(lJ_1QO$y2auu_+rR
zkO*C#idEtqq+xj)R*B(PI+mwn6>L9&<V)n4SOx109VC-&%;Sk8UV(%aNK2cm;8kb{
zXi!tftk&a}(6OOo;)n0tc&8hg)8D14-=!^JPQm7rRWPOjp#|t7*&Wrau9_VJ8(T0$
z;g(T&1nd@>DJuvn_H9FAyT^wkJO0n%%BZB8=Rp>)Gf2ztP^Wh&SZw({>hzw*4X9It
zh({ghI`umjXD5~%i^wh(xj~mmFu`f!IA?IZ1)_0IYj{skL4kO*R;PGJaB)XhL{H@I
zJC>2zs$!&7j5R3#*hL=AZvRnfNeTkk01bE<lbs5z9Oa5EPB8pO_MaAT5($RUR-+w7
zG)6XtDjRDxmMZ%}_5)Qm&T8D}vKd`!XAUa<*;*<Xk{Hh<rbnAMC9@wMK2+PJ^`0M@
z9<<hbeqwsiTJPD&^q{rgGYs_rr4kIsQM==)=d0f?9ln#3S3#@}>yyLw5%G1-{W{le
zfy-8l=FO&;K<GtO#hB{_Rhhm0+L5^E(Z7*ddCvyFXTJv3fR>5+8dJaU3v#q0VHvU!
zy|10DW+&?+u5=={O2m5DXLG!!IZT}9=|Ebz%LU)%`odBCYyZE8T;8#N0x1ZWHV>D2
zgI+_rfppfGONOTBo+_`L&tARcCm7yJRBt63@$h~pQN5$_dx`2jjpLCnUNmszY(Tno
zG)_diL>jM0y7e?pLb@axCnH_*m)JEKQ66L}(xuYnsmL{zF5iH38|d<Mq)VsqMx@(F
z;|!$Bpz$W8+eG6`q|2mn1Zx+;e$D|9A0ErIu@bx#qApmTgOw1zk?0M}_h2O?2L<8)
z$oG<~D$^YdaS*UP7b_u?+=Ml`*iOWGSe}QK5DSph0@mbVI}z{0@_kqd**t;dbLIQ7
z5>mYax#g1|z)DDB5aW#H2eA?o_5!&+lILS3_(a4d!tz5{37#`K+j(6g-xAcEY&Tw)
zMB`*$mrUanUYA1SRN{`)IE~k(eTiMu_%>qw>AWtTE>GuO)9LaIUY9|aZ{l^EXq?IG
zGHJY-*KMZp7GAf7##y{Bi^gYp)mdHxhaNHXyy_f{&-1GDG`_&AF3|WQuewO%LS9w)
zCAKf*oy7Vt^Qy~q`DNbzGF@K8tBUCItGwzejj!>lYcwwARmC*E&a1A|_y(`KA!3ML
z$%$aRL4p$jj5RSEcUa9Gwu^WQy=PtDvq5m2kpq{t`^a_&EHhm&fti5|CgQ#sZ<nA2
z!ElU`9%GbXM##y_NRNvcGK0h`GhQ}{KTN<@3AiQrBN2xNy05r<peadU@FYnT)1RnK
zi|6;-dgf2drrd!#m$QfCu_-;L&ZF@@PQ8!D`#JS~8lU3Sr)YedQ=g`B38yXL^kV&`
zoVJw4H#zN18sFlyw`hEu)83}>9Zq|P#&<dGT^hgO)Gz4vUvlb~G_K>+bu_N$)b%2c
z>4T!4uWsxTA{f3=exnsl@krT7OA(EdjiS(0^;Df$RxVMOOLTCx(quHWZzYbj(a^r5
z@q3B(J&o5P{W?0^oQU*^G+vMN>uH>X^hq>MMjpwC$_=L?k5n3OK>7_dKOO1QX}l5X
zH_|u*=`(1&3F$Y{I1}kJX?zzc?;@?ZCGH{RJsOuGWf_ehAmsxZmm_64jVq9{g2t6d
zSxMtZNco7ykCE~*jb9*(7sy)dPd&1#7vmkl@EU1ei`fss5WzY{uwRM%cr1;_%^`o+
zWYkpI*oHQ0svK-X8#UD)Y(pD0)n4+7gEDINxwsW=#O(8ME81wO_F)^^XsPyN8`@~8
z4qzMFXsHfj8`@~8^05tVv{Z+%jcBw=aElT#p-FCxI7K?A7B}cdPO*{3NxVLZcN2|J
zGOtgjaSE?bp>Zm&Po;4h?~%rL6n~%2d!*AigV$%!{7t-m6OA)@eI|`J^ZLy+-oop*
z&^U|NXVLf^uRSMv>Vn}suRTxW3%vFMjW6=ri!?6eoeOz)+9)|+rg0IkEu#5XdF@pi
zU*omcXk5%|i)nnF*IuXb4PJXg#IUX;P|rt-eB=y@DKotx09z7sDb{MOcHMn)(%<HS
zZ*!rrqAMS!gnBF<KAsd*G1@A|7L3ZO7-jpkP1E<=u!bwh>k8@xRkMaKZ_t|XS67h_
zT~6PqHI)ZnMZ?ANFK^gb!!;Ce4Sh}#k)-+;H*QThpqMVdfxK?e{rU0+j$D+WuS)6u
ze0d|+R32PP_vgzSI@WL#1>B_D`|`$)HQYi0x9IY66i`m%3KUR5<3}jy5&Bjf|6>&N
zn8sPyH4D2zev~EoQui$EA>ta`xdwNIL<dXqr#>~fJK)u~=5>*q<vnjQ;w=mwV8kI*
zC#rFi)!by;L;Et*+YwL=a1MyKV0<_d;yfoN_E5kgfjsL_-<R};7awcWTJ_YrR@ak;
zHgoFDTsP36KQ38DFY|b^gA^ol7Rj6=6vX%3czyk))(NEG4i|h!Ov*tjj5trm2Pklv
z5MTb)iJr$YtI03k@R~QgGX!Aqe~%s&(Oh?fcvS`1qX2h>Dw=TL0z6E_YdOtYIyhg;
zd9UUCAy`j1ADmUZbWs~JFSmKC+q{FgcVF^WFL?)W?2_8$ZJQI>E~jY%x!{|0;;)V~
z)QO3|lvMYMg$2KK{+}}xspF%TE*$jwJQ?(I<ggqKhBQp-l=IV$4~uAuQ<ox#rDzBQ
zxG=$F{*144U=h>&@B7*h=qvx%ulGP{IGO2j5cII=rdE3bw-j$`pJ2@s><H@Il-JWd
zrEwM3RMEH^YpO+@G4)uVxR>f#WR4<v`$*m$<|boN%<4i#=Rk66_W<>JAYQq%hDsDz
ziF(6^fSwRP+xS)ql(9**GzIjfN<8Nt&vgJ*$_%(;=U?zX`Tta<&39_Yw%3%qkUpN`
z$mes;80<W$ArC3?=!+`jV<_09JLew8{ikL0ChGr+v3|ukKw`BCyS!rBiug5S{hBU+
z&A7Z4%Xf9?*8fcE=yjyUV_0_#yNZLY<*jN(_22W|&Gs3;H&>DtKd|Z#tQNu)lKo`W
zjWmwL+DL3CmPccCH2%!r`>@Z!_GdSBAcJ3p+O9%DkW}3FZ2ms$6?NPH@onjT@4bl&
zuGR#THxqAo)f@hE_7U=&WN;O%rb4_rmYE)o!E{1WnVfmX=VGFEuESRAa7)P09#Tk)
z&VF0j?_b|fhkaXGRz|;BNpgbI+35-~pg4wvW2gn_GI3y#aGu85tR|bzx<p`$2y7!R
zRxu|n=9I8$j<gPO+k7kMGFhxbiM&wK7AEG%f0~6>o*k4+3XU?`ql^#u2}jP|7{1G=
z{?R{P#pAOc{`K95`TfcMT*lcg6YnBfLkwpZBVu^MM^<$!Qfw8yb(!fQAFL|8oN6XB
zJ>P@QlSo@4e3a<LFe`=;V#uR!(>rI9XJ@Qa$V=wWcqb$5q|0MjU94zA1;Ykbw}HmF
ztYa?gEv`ZVt1A%8!96D<IDixf=$km>BSIL#sjdBglJ}NI5#s;fIGuhoVMz9@6X|3*
z&PkllN&H}_XYM7BeZrj{N(we}zMIAC`aiE-y?C+u)rtSvO8r@t6qfqb`>KL9H1I+L
zeXsg~7e3JVBQJcU@h4vRMB_$YXrytNR0tFAEm=c^REVH)q*REcag<buqVW=`u!P1-
zrNUAgFOv$(XdEpSqG`NbDl8W<yyPT{umdS}(E4V4)(NY1an`m04G*Wy{?`wOLJLD_
zL-U$$|9VFt!A_>Gl+~279biYuOz&J_>MlDih>qwns5!ANAK8$PY(Mcp3d12`n8Hi2
z%M$#RSiS=5S6~;w&;ap-zI+jO#n)xwz))SnKR%b>=qF<+VKpV<E|!^I*}@o#hE37D
zos-(+olRmVDZufD`xA0YVSI86TazRYkZ%MEh4ft0?VeQ%c7=o?dHRU;PsAI|t0lh-
zb9X$um|P<!F^)-0Tl%KP@gvg~T57tVbo|7$rEh8+8=1CnU($3x=@^FEiZ?Z@kaQJl
z0fxBg=0<-Ud5JeQIc%F8wws8rbGqx?S8#Jv+_%}RMBlNMWZ<Vc^7;H+?}lg1o!lX#
z*OG!n#wC&blNr;23Ky2-4(=|ixl89n!?0Z#rgEUhH^?x@H(#X;oRqj?+5dcML8i@0
zW_r3TuEsV-wT-cZE((Tt<Qb2;f<lp41_dXhfg(;t!KpMphHQ?B5qFd2wK-4YY}O^4
z^%8#{f$bx3Yf#*#XEB~xxGRKtP4^dFb8!H;giZGqUGs2&nEI)~el<8m^plIZ7R6$$
z)pXzBnaFn)Qz>bDa2h{Q#OZu+I*rfrE@yc!G5jgy{R;UI5nty0E{hoYO*WeGoi{8I
z41hdeID>>Uh<ej>771r*{3jCrr11r0aY4L+l9`@`L&f)xhTM`@k6uo+X&*;EpS|Yt
zE>{Bn-l>uN<E>r(_NQG9J%`wk8uxSL^I35#HJjV4dEC4$+4FT+Q-^h8G+B=|^&*DE
z9U<;TioN3Xmdx~E9XkKW3zMe?*2K>sm%+_W@hWj%Yx!oiyb8k1CvEa~b+Fr!LHbmS
zZEJBmXc`{jkUg^<DYlDm0tLfKWN{MNfM$`mvHyJ!4^2aMpEP}g)!d*DFynK0Al9Yi
zANR~@>HLr95#G_07PcY9HrhfOU(-Vie>Y_Bw631Ek_@hnv94p>AxwGOVdkme-Di9L
zpUxk=11B$ljj!;<{ri&X^OE@*YHZq1bQ+}z(`gjqQQ^?rcmPaq(=#+Wt<r?)v`Rz6
zmgAmx6JC+7pJSxwn3mx1Lhh9eem_#|r=2<DKLUWkL+XO;7~{VHfUcX(8b5W6f!GVf
zPV)+@+Qq{p5K+RD12d=t!ZkAp)4)MDgUV*l`eAl%lkMr*^q29^=bIcev)=Ik%oxMc
zg3|2hy@L)b`t%Ly*{gSee{fJ>cfTHH-9K~80>rOD;@1%IYY*|OpZK+}_%%@c8Z3V8
KBYy2Ae*J$S=n8oN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_wmma_f16_instance.hpp.FE4A66EFBF344CEC.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_wmma_f16_instance.hpp.FE4A66EFBF344CEC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..feb8f66eca14c4923200144d07c5e2c37415b378
GIT binary patch
literal 9704
zcmYkC3zS=RoyYV0|8J+6N43+bL#In=X)W?_PSK`Gu6dD1liaFZ4mf2wgq*yZ+{?|&
z7Cp@fE1n56lTcj@3CRp4BbxSHCIe_)APvgy3hq`}w=SY6JrxisE-1@G7ZjCE<4N=F
zIsKn}@Av=7XD0XN=Fa`+8p-Q@VY^P}j!B17`)>{lI-PE`^Z($@2b8m<(;XSm=?)$F
zZu!Q{=`HJL|9IKsC-)o9e*USg>hf{JgP&1ez5M7q^Dn%#Vb#dT%@1DphkyCtZ>Nt*
zQ~RGA|H0;W#np$N=NoRl*ZTdz)7uAk>^$^gV|&-B>$Gq16E{4t>)|8CR}@38?N2uE
zYpnZv=6R)g)9DwV8r{MFtUP(e*c+#|{XMr1|Kz%7u6h3#yB|ON)z_vLhwi-kxy|i6
zesq^Q^?F5{yzb@F@f+RE|9mH?hW47?mu~&=Ti0Kdn|Wx`lKu7PH*VUv=Hy%RBmc+B
zkAKxr8w&j<vGFepl`DRI!&AqeDr|n~bGN+zG0znzmC29pIr_+CPJ38)^Vq2$|LEma
z?LWM8;hAfr7ZvyX^fY^5^pls|@z|=<C+2Q{S}R_<X3BKGd*Zj^%<i^X+WyvV`-R)3
zKRIvt<Cl*f&0Kse@Sfo{?SJ<^viDoFhg%PR;kfeg(uvPJ?U$sBPMy56C0@LAYUcMx
zhW%Im`|YLMhTS`FdiNsrr&sNkrQK@l-##ThFmgt}#iH$5TdRKRnTgON2cMmDz4pZ`
ze;odwH}*Z-dUWO&|8T|e+-3VRKlsk&&rURw&L6&g>W4?i-#zniCH>ODM_-hFwqxhc
z-S@u2{%gzY+b?Oq{l4RsnZN$W$G#&yvF7k)U)=fa7w#$UOKG92=Jx#KQho2!_uTR3
zU-xXf_qo0A?iiZ7W%a>aDx!Aw_Rycc_RE=1?z(oEjeJlyv1wq_Xm`*Jc8&<0jwWdI
zp~A+`J^1Pq!%W|qFgh?g*d6*4He51sN%o(oKDtJ?sxx8h_|{9iLwCa34|JZv@!p%i
z)6V7|8fI%c6D9^GR(A*8An*9kDGty6)89SwhVk-YHr$yoHZZ2|4t+&=@FzQ}E7ydF
z*}Bezsevg&chC(A>~GI-;y&%xPu%_dM~2x@XTth{^@H7^{|slY%07PYLqGe(8r?u=
z!shYKmv)E#5BU;5^}6`_H-0zFMmrNG2PSuQhrZ(X-+c4<4PW~2%ZAzN&V=!S@zL(k
zSNvt|zT;;yPo5oSgPjTM2G$LBhyDcJf&KA=v{k3$2fkntpE&EFk%{#q>&JROd~a8{
zhq}Aq%inwNYb!5Wwv0>|>SVZ|dOu+&H_7k_WMiF-*c+Q<#NJdV6ZWPinXot4$&9_Z
zNoMTDI*HheO%k!U)X9RqrAZd-t#z_uZ*7tldt04s*xQ<9!`@yeJNEV_*|B%j$$`D2
zNe=9tb#h|wY?2duSDjqgyPD*}-d!g*_U<ORvG>%;gT1Fo9_+n!@?!68k{5eloqX8)
zn&iXYUnf8I{wDdcm+B;8FEvTRKCncAW$L|&MkWF)6nGSJaEXH02UjSFeQ1e7*oRgq
zgnf94!q|sbD2#n%i6YoXRw#mfbcv$aM^`9{eQb$h*vD2VhJAdA;@HPmD2{z%i4xc+
zRw#kJyhJkg@(Rh=D@&wcudI-Qy}Cpy_Ua0$*e91LiG6Z~lGvw~D207$g;Ln3mne;W
zdWF*czITF_hPo$aIBh6~B%Cy|hJ%puWSE0fh@nP?dmtMXBe2G-5m;lMjK~^mWJK1a
zn1D59O~9J+WJ1<dBNMV_#SE-DYX;VwCo{6<8kvz56%kl5D*`L#Nkmqxk%+8Cu>fny
zT7b3W$%3q<MiykPiWOLE)(WgOPgZ2DHL@aWQ*6N6vNm9Cd9opEtC0;^yJ82{p0xvO
z&yyWldyVYKIur-6j;sS%N1hzWI%?!V)~Pswb!MHwI`ia2)>$JbvM$91tSjpR)|Dq0
zvaTArkaa6=VBJ|au<ks$k#*O|jjTuU0PD$mfc50bgRG}U9%Q|W7g%rB3#>O!USz#B
z@*?X~e8Bp$K45)$@*(T1kq=qF;s@5B^#kkAlOI`sjr_<;iUh2bm4KD<Bq1x+Nb0k_
z%Q+hAUdHuKC&t^ZKQZJK7BB8xe&`LR{+Yz6H$pP$O_0obGbB+jLbB*BkgR$uB%9s_
z$*#9Ua_Ak9oO&lDm)-@*t#?E6=sl3UdM_lO-UrFA_d}BO5~P4WVCsL62K7NmA$<r^
zSRaNI(MKRf^-)MMeGF1uABU9CCm_js8Iqz`AgOv4Qc|CUl+vdlrS<9asdsJuo6hb&
zcV*t&-JQK$WriB-?`MO?4EJMhF~bSWGu)8pFu%+V70eB8IEuL>8154K^Nm$zM4wS(
zM)Vn5%!oeY3^$_BSms9b85`V)K4VERqR&)iCiIy!W<sB-#Z2fk&2SU?Ol59DpQ*u3
z=rffB6Z*_mW=5Y`V`lW3Tg;3;^9(nm&s^qa^qCvnj6QQoFr!bbG7){E#zgdqEheH*
zoZ%w+#4;DrCpNf<KCvW-=(ALr1$`EcS<q)`F$?-EGu(ncOPO2HXK8Q?`Ya{Ef<9}N
zS<z?Jm=%527PF$yI>W8#vzECPebxrIqR(0qtmv~<nGJn5joHv=YcU)8Y%|=3K3kdF
z&}VCK8~SV|!G=D2mD$l}*O(oB_7=0F&pyNL=(Cr(9ewr&x1-Nq671-6RG9;P4vjg`
z=V&np`W!Rdfj&o>JJ9E7a0mJvCBcC{XO%h8=hT=Jea;qhqR*M<&fDSq;Vg3}`kW2!
zM4z)HIML^-G8g(>8grq~)nYF6x$@kFK3AE$(C2D!7y4W!!G%6|mATR9)|eZ8?iO>S
z&z<LP^tsF2jXrmSyV2(^32yXxs?38vkH$Rc^R$=;eV#n`pwCn09`tz{+=D((N${Z0
zTV-DKc{S!mpSQ)l==0{e7k%C`_oC0+;9m53OM(}DzAE#f&!;gT`g|?sL!U3tedzO*
zxetB52KS-QR}y^a^H-T4eSVGk(dTb5Kl=Q6?nj@$%>C%|H@F{t{*vHFpHyWM`Xr4>
z=#yGZLZ6i968fYvm(V9QxP(5bBuMBBEU>^L>tD+RmRX>Qxy=HPVV>iGc@FbyJg|US
z<AM7zSA@XbLVtd6fd$bQTxLP^1=}o$zTg}WqAytELG%SR9z<WTA_UPFT3{jcg_c<e
zeW5lBp)WMYL+A_DcnEzVjfc<|st6(Ug%?;Dec@#mMqjwi!srXn@i6+rH6BJ^SmR;z
zg)2fBeUSwgL0@E<MbH;%vk3Ykb3B5+NR3C(7tweGeUXX~L0@!%MbQ^sW>NG-+boK{
z=p2utFIwYK^hGruMPIZcM9~*pU@`Q?mRSsau{MjLFE+<x=!?~O41F<;$Iut62r=}<
z7g!v9@nsfAU%buY=!?(ucma<0c#X%=7uR?keesGAM_*!rCD4~xW(o8q+AM*-#5_-+
zFHz$O^d&T&KwqLFB+w@>Fd2RFGLz9Kx0#GSd7jJYlWSZ?pR92ieR4&R(Wfjh1%1jg
zQ_!chnSwrLo-626YFt5|qHzU%N<~o6r!FuRed;n((WkbViavFotLRf}Tt%O%aTR@P
zMNrX~TwqD`C6`$eeaSXUqAxkmljuv<coKa{jVIBUtO!Z;r50EUeW_)ZLSL%QQs_&~
z^A!40HJ(CWO5-W?r7A)Sedz_3MqhfFrO}sevo!kB^E{2dbd9Icm)3Y1ed&sj?)yG%
zDrLqe58v5Y64_+l<nAqn?Ekd$yI=a&u}Ek0*vQzr-a^sb8@n$QPaWRa**v;-bi8*5
z^p4g4df)u_4&2t+JicLkQ*XtlN7;tD%S!z}aRYMTrv3+1AQSj3q*)5g!#(oAA_a~>
z4$47bgP9<(!C49-8(gFyvLQJHY$y`~HZ)5iWJ8M-LN+XifemNEz=mfjjBI$3!pKJC
z2(Xb%1lY(dMUagwQUuwk90fL-i2@s)r6{t|MT#ODlViZfGBIFdvlK%%wn#B#<8mC>
zcqR^Pe3s(K#uq7$Y(h=|o5&=9P0UgP*~B6xkd<W_SUDpDE6<XQth`7vvWlz#t7H^l
zm041dRTfD>R+Uv?)r<<PI!h|D>LRJgCgmis$xIU1<SZqTO)gRr*_50DHkC;Mo0_E*
zvZ+N%A)A)dz@{^4VAHddMmD`j={|c`efw1tA9?j{HZrkg>lR;cm3Cxe+tzKq-kr$a
zo3e9@>fF+A@BQ0+Z6<g9i$}iFVJB^quKt~*DeIK8$E+NBY}M7NuX{VEgozJKjQ4JL
z^_E=M4t3Xwdr#R}8k-K>+G9s1g401rq3IB$@N^hbWI6&VIvs@+n~p(>Psbr8rW26l
zX&I6-tw2(zRY=L{B&5`I3Q~GHeLj8v%sr7;2A<s4`Te&^+ivW=q`eoq^OBxh3-A5@
zz1Og_`P>S4Z*%WNx^<|#<lWocxkp$aLm5|%3S`7=Dv$}Yxj<&jVu3`=mI7HYTMJ~x
zY%7orv%Nre%#H#%Fgpw6#Ox}N3$wdGZp@wnc`$nm<i+eOkPoxJKz__pfh5d<A_Xe=
zp;4qD=1`GBn8QU1V~!Llf;n2GDCSs^VwmGaiepX`DS=rol8jj?l7d++l8QN5q$K84
zky4n`MN0ScpGWo9lam+xs&lMuk+xjjJH~q7ix?Z~ejfsko}A&9-Vrx4VaypJnQ|sb
z=A0Rlm=hscau!I|oE4HSXM<$V*&#V{4oJ?N6Ot?Eg5=J*A$f8hNZy<mk}v0j<j?sb
zNjd3!>U<L<89w*@nu7<T2d_Q%;6vwzzx`X^d!W-@(0@4Af4JfNLpJg9G5_6np4|2G
Rfw2qLe`v$FZv1?w`+pinRmcDU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_wmma_i8_instance.hpp.22BF4E121F8B1D63.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_wmma_i8_instance.hpp.22BF4E121F8B1D63.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b40a34ad4aca94fa026aad16948b6f3af4cbb4aa
GIT binary patch
literal 9704
zcmYkC4U}7DoyPNj-^(=f(bA#q&=I?>tsJa-Dl3M3n46@@4M}bVPPbST3^89#?j^aI
z*|xgPAU${{W+nrQ4GH<6CnGShmq{#I-8P*rt`yj<yXVMpE$Z4Th}$CVf^gL38#aw6
z<moxR=RSGf-~DCgCU5TCH#=mX?}aN31}QG@({}BR2nK^;wDW(@-redsG8i5hFc|h7
ze4=#Q#OV#I^KWhb(aBw=b6@?>jj6@Mrw)ByJ^#r^-k5#izO|Q(Tw^`-g}?mkrT;ti
zsJe64tylbH{hPLB`<~}(@0=H(9z1>J;MIHDSDBa0_g=0)%#X=msNZ_yGru?0#e-KK
zd2aiT!Jpkb|HGaC{K{)@%~t6BYhFAhynEr~4-C(q_;cyj^RDA3kA8Z0Em1QJKJwlp
z$A2Xr`~2BZ;n%x&erogSYk&A3cmA<FaBlCPo<HT^IDhtl<;^pP`PFA`d{}yzZu!xM
z=_d@k9{>3AN7>r9YrYQ>XW3i7IQ;VN7vI^p=K1|A{(k6+TZiwy<}vMauin1JaN9GV
z+546K`9tqH{6|k!k3Y2IFGsfi?bF4dyqI|MtFJ~LG%ar2COvn<`Ryz2_-XjSmT&DS
zeC@!cez~}R<(K!z)R(_ByZM*SpMLiI@%z4eQ|n__9lFqHJ#g@UK2<jD*m1a!jITZO
z(F4!E<bFFp9QeC``$qAuVQJ6pZ~jf{xvkeJ@^z{9lh@1N9l33IgI(XYvRb+Rlr4N@
z&oBSUbKxJ?{^5Ij_cc>zPCuD$t+~aurg=2pC|?kE9NBR8<>c8N!|$H^^|qt0A3Ocp
z$NyNV)gJn7S~$4%%uUzted&WuFHOJy@1On``KjgiZ2sUgKX~E7+|8OE-a51G{YK-)
z|7za-`u}ZP*En(G8|x3Oxa0JmtQJjm&i2sT-}&wI=dQhJn2lU&7+*KAZnV2G40c`-
z1_O=L=;MWteC5#jr-qraGhlRJbg;YW57@l-izklX{)V{RFwhyWer)|Gx|{BRl`A?&
zaJ}>TEBd+Y<HKxuXTbQt__FTCFvvUpi;Dg0*8ca__1l*%H(b&guyJhTC%T*d5$?&~
zd&Q~O%%2=)!<_-E238rnoBnes4gUP<N^3`Cm|fNxuybIisk<=@3hb|nV-M<ge)j(7
zuNr1UodK%{Ru6VJ{UiLIKl`ff)rVghW}}?}TL!jV+uih!@aONne)vmY|JmkYwyZN?
zY+!7(yXha{y_Gi~{>{YG=Z4u}XTW6xmko9|{Q-vEyApe7qrt#;-(j~sb<Ra2<Euwj
zuj)Ow-l>p=x~JfqC*Ju^>v#4IBjct9nI5CwGwkFhnT|j<H^_{=xk+a1Ee*0@Z)uVR
zduxNN*jt-q#opE+8}_y)*|4`a$d0|eNp|eT28q~<O%kzpG{}Lyqe%|zoegqg?`)D2
zdsl;8*t?qK!rt8=H}>u(xv}>&$b-G7NgnK_21(dUO_H$pHpq*;w@F^?eGT$q?`x6|
zdw+xc*!!F0$3D=Y0QP|<1+bSJBx5f(Nya`nPr*g%y+=mIgDnakg&dlv5cZ)Kg|H9L
zQyBYji^AAP<|%@Gq(u?zqw^HSKH8!v_OW@2VIONz4Ey*z#j%gKD2{z%o)Xw6T9m*(
zIZsLKlPyYOugsHzz0x8Ddv%^v?9~>j*r(<xg?*|;DeTkpl*T^YqBQo}JZac#Ez+>h
z%u@#YOp7x8zBfV3L*0oPrVZ7UhDl@6v<Fg-OfxWrn5txY0J2#%18bf%18dHa8Ci3c
z%*a|)3$T_+3$T_PS&+3<$%3p^wE}CMv;u3*kri2Mm8{6xR2#6iNgJ@X9NCbyRmp~|
zU9|&ipR@yO&ygKjdzI|SimC{#I4J@v=14?VtdfYVLv;Y_m~;T^$dLnCN0l7NI#nmI
z&PgY*&Kx<BbymrVtV?wP>zZ@{>&lS}Syz=@$huWGu<l7Wu<jhWk#$$ejjTuY0PC6b
z0PD$-2U$;*JjhC_1gtbE0W0N5LRPAhgsfNf0_&ah0_)9@7g=wWyvX`gAF#ejAF#e0
z`H=Nh$%m|8^#kjl^aJb9ksn!qmHfyC)Bv!7$pEl{90iaKR4IV0tjfU3lQOV!j$~xz
zD#?AecR5Ex-OIS%bYi~imSaP1q5iSX<%iK^>CYr)qZyLLXn|xkS|QntHb{1(9g=7i
zAvuf=NKT^@lFR6V<Tkn?d5j)Nl2L->HF_cWj6O(yqaRYh7=R=jWk^9|(9-`P4H-j_
z!p1P9h%o{wYK%gP8Do&*#yF&eF##!QOhQtO3MAF2LP{A^kkZC9B+aNn$`~^rrrx#r
zA3CS|;+1*tba&2jg_)|Xf1XV`Gd+g6%}mEIPjgd_!}t<6l`$`H(<7LRg6Te?Ki*tn
zX7rhLW=5a6&CKXCPjfT+%q4C{pLv0s(Pu6SX7pJq%z{3P&MfG&w3!8cmT7K5pQXet
z=(8+v3;HZY!Gb<(g;~*O)tME2);6=E&pOSm=(Co%6@AtPZbhH9C|J>Ft1uh-Y&x@{
z&(>x(^x3Al4Slu}x1rCrz-{QW6$Kmm>=kB5pIv8m^x50Yjz0S|x1-Nq;&$}e7q}gL
z_M%`%pIBic`b3?H=o8yaM4vd#Mf8a!E}~Cd;3E3Oq9CHrQDF}BIdtYgpQFtj=yObS
z2l^Z(?m(YofjiLWC<+erIV;SGKBvx{=ySH26MfEU?nIxn#GUAKE^sIMoJGNjK39dg
z(C5;b3w^FObD_^Q&0XkomADIit_AKwpQ|Xi(C4l&H~QQ<bED7QW^VMkbKLz6m_OVl
z?na+`fxFS?E(&h+c`D3<K99~k=<~Fh2YsF#_n^;H;vV#Q7Ptp}o}%DEpHyKI`Xrr6
z=#$z^LZ6i568fYPm(V9Ia0z`<QIOE*tuQb8ygKuu&)a5R^m%jKi#~6Od(r1z;9m53
zi-H$@z6$f9&!;mV`h0EXL!U3ledzO*xDS241@1$iuPFG?=dUn7`usZcqtD-Fe)Rcs
z+>btgiTly#U*LZ9`HO-deSr!Kpf8}a0Qv%L7C>Jh#{=jKlz0Grfdw8wU!W)i&?i@z
zj6PXsGWz5;lhG&VxQsrz#AWo!3tUE@Toh#V1#2u=XZ>rL;35k)F)y*;am+J3ILl#t
zl?Q8>bsl^Sb6E)9FZ9QUYAl4l&>{<=FSNu$=nKv85c)z@9ztJ8=OOfk%0dWz;Tj90
zFTBXY=nF5gF#5tXJdD0@m50$6)_E9x;j$1$U!=w&=!-0}2>K#REP}qs43D5MQsoi!
zMRXoPU!*KV&=;+-DEgv{EQ-G95{sfQI>V#ri&l9QeNml9(HAWXQS`-XEQY?=B8#Cf
zw!~uSi_P#D`eIccLtjkiG4#dCLJWQJ8jGVZzR2R}i!ZS_`r<P@j=p%6$I%zpc^rN5
zvJgjKqQ(;FODwVk`Vvblfxg5nPZZ$wo~ZH!`Vu-%pf6Dt66j0TSQ357MV3Tga)~9;
zmz?EE^d+l2iN2)Hljuv9g(Uiv8dK1xEHVXs$`Vu1r_6E%eM*%p=u>pApie0a3i{L<
zQ_-g`G8KL55>wHq&T<ufYL%<#Q+2MQPb~{7`cgHPLSJf;rO=mJVkz{cW_b#IsVYyQ
zFQxMo`ch>fg}!u+rO}sOWNGxJmslEo=~<peU%JZE=u7K7jlOhQNTW}yF%5m%BGb^P
zEiny!+AP=5r&YOzK27Hu`n0m3p)XTo8T4fqSq6QXC6+;7W|n8rm#Oj$`Z7Atpf6Jv
zGJW3-mg2<NmV3V4SrS=iT_^RHLU!HYe&Xxje>B>8c-6?N%X;fTy#=GrAFUTxi+X>w
z&b)TrM{UpE^O4Tz(UqfPy&I!<FZ=O>vnO`n)p>Yq?by2BLQd~sj-l>iQtu$0+v7^`
z_WtKqa3c5*kn$9qg<I#rIt33x4k;mELlYrjLwO1z8>&+X*{~7@HarmqHk_w0vf(;~
zk&P%3U?UR|U?X{oARDPu1lgz(1vWYn1vZ+eD6-KyMUjmuF<@g8F<@hPiXj`TQw-U-
z5(hRu5eGJ&r#Q0lI>nJqC<$N_6A54wc}gIgs8a&jq>=<SIgtc5nWrSO$vP#GRTKqS
zWkLZ~$&-StQYQsjRZ)ReCsbh7JgLa4byAT{DJfu66DeR*c}gLhs#6Ntw2}rkJ&^`B
zou@Ri={lv6)f5d_Z9)T9%aewzRwoVFjFJI1Gm!x{lcx-_nL1_q?0t<Zw~k+R{!eUV
ze8a{K{@yLgk?~C%H~D*S{@$CiGYMQ=@a?^Sn-`|DxBTwlw>s<=#}-fj2GdS)r@P0r
z_8q@uyY`T;GbN0#7$5811?#QEt{m#F4)+e(nG&XgclOwk@z7KVQg|v1DKZs-6rGAf
zicQ5J#i!zs5>p9C$*CkHWlDjhPN|SmQz=O4sWc>QN`sV{%6yof`OO2-mj<4`x%2aH
zk~iJf`<3>7S)E_$#TD@0&)=IdHx6}|yn7FK?hzKqRKiuG0+}&e3S_}-Eszzntw1)+
z_5#^4iv<!fI|}5$>@1KIv#UTZ%<clYF?$N+!7LR>!t5=O7qhQGKFt0C`7s9y6u>MO
zNX8tTqhJ|79p)&6IXp*U%#k^YV2;jF6mx8jVwmG|6vv#HqXg#U93?R;bEIHa=Sanz
znxhov^c<xzYjdPw&dgD!pWhudiYK>R{$l4vwL#vnz4xx|eJ^6wQ1|-~@aoB$zSeui
zjf|VKW=NK-1(G#ug=EXxAlb8aNMcrm<j6W8IkQekuB;1^JL`tz$$B73SqYLi>xJaY
z`XKqUen^3A0Fs=QKTMr(f}|rCzhARwcWlp17kAcN-2C<5>fVmQaCv`sReyJFe|PM|
aT{eF9(ZK!RK6&l023B3Zdd=D~!~X-*)m!5L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp.8935210F01B214E2.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp.8935210F01B214E2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..146798544a26c57283d67222919074d5dbe77bc3
GIT binary patch
literal 16388
zcmY-033RJvnFsJb?-#G(UO+G46>nu3m!snyN2Vr|eGd%6Ag-Cj?0ZH<+k(^f)a3Nb
zzM_HAFrc)erLL%yGIJMnYQfUlQun3wmLe|WR_Z<~9Ok*56Z?NThn)W3=i&eBSDLnI
zlefI#nrohM3=DWuZy3KReM?+3FfeeG|GoK^(SO~cfdSv4fr0UlJaqdv-t+jCmsBS%
ziT>#R<2T>`<FDNL>9y-$cEg>YzU~LMmkq!4<1y#dM^kTqBER$S!tK7dpI-RTyGssZ
zZF*Yy^FN-t{`C*84}Hw|%f^j`Q-68ax4(Jo*4otzL*G8~nx{X1+c(Y{fBSr;>HFaO
zpS$`6r~dHTyMJfAV&`?XyHC?|o4<Wz;KunY-l2cTcO$tEz4gJfZ~od(zx}Nb(XmVF
z4&AM<81Z@E{L7Qi4BY*u!iPf5%Rl(c(Vu<d!mr=+^3Pv9-uX@1eD%o>x@qCnzj^=n
zu37uo%);+H4_xxi(hYz7*=lg{ZyuU`=Jm~Ae{JQ`vv2unq38SFT^D`s>!+@r`^2-C
z+&=QS>+c5F-;;Uwp#AcH|JH|d%7fpz_=UA=uDav3nb&O&H1CUKR_;&U_CNpm+qXS(
zdBXg==uZxttGD0q{W87Lo%4Lx^2YHyuKPpiw%@PJJ@weBw{M*g&zb)IzrMe`@TI%Y
zbiVuGP0!uqzV^|pzB+r~K=0PkpL<K4ZS|FJd-KHbSK{}c`1obv>#q!c;qWVtHCOIb
zGLP@Q|23V7lDAhY8}9w;az6Rymo%n-KXdp+um0z2*MgQiPW<8??U!%(GyUd`$@?D)
zJbm)=Yd`7y(|xbBJUKJ{8^?+3zxd#{FRvap-}0#+eeBFtKVE$HsZUuie|A{^jhR=q
zp8C_O>9j5T!{@|xe~E70JaJRvUHt#~@SkV4OV{h)dPE#~!NB10L&uNq-wd3`p9=#6
zG)PB3-#Grh`+mN3M4ZP97ah9jy#1Sl1;<~0{i{#@Wa9Hj#QD5%?9j32@81lZr*Qqb
zhBy6j;=UvA{MnO7#8F<j@z9O-{hNXFG~$B&g_};C>;*4<+Z&IFL%c9}Xz;@Qn**X0
zJATb){^9lic0@dn7mgk}dU*flVBwP&nEbywmAvzaxR@7iK6JBf|K=R=qJ83XPi;N>
zf~PJvUNE5Lg)5F-q2IsRZ+PK_{7(MwzvzYrzxsU71q09Lh0Bgz_UiqcgM~j|`-gYl
z|AX}7N5n<E@R~!fvF+cSBQDt|o_(S3{<H6W_$NoiVP3fS(8cHN-yFP?1EcB8%~x3u
zKIYlO`ycDhSN<~d*(WVmo<BG=OG96vzQlQ+hQ0(@H%mI4>*h&^bNwvoaju^yJ<biY
zWWc#$o(woQ&XN)5#(6U0+%!ujoSWv!gmd#OnQ?BOCo|41vt+@!Wu7cJx6YCk=hk_$
z;@mb%Hk{k$$%b?LEZK2xpC>!c9kb-XxnrIjICsvH6X(u(a^l=IOD>$d=E;R~_bj<_
z?w%(%&ONi_!MSIiJUI8xk{9RRdGg}iH%Gqv|9XEqf6%u;zW;{opCdoc{R`yBd0>tL
zI1emP0O!Fu3gSGtKtY^`<|u^o&;o^U9-gBx&ch27#(89pA~=sMPz2}EIf~*ux<FB!
z$L1)8^VkB#a2}tdIL_k>6vugDjuJReEKmaH$vH~mJh?zgoTug}h4a(`rEs2}qcqOb
z3zWurW{xsA&n!>|=h->R;yk-RS)AwQD2Mah0_AW%c`r?Vk@{Z@&mWw8fF{2TIe#zZ
zah`vG@;EQtO9h-49-sowr|4dqBHWmw2lxbcRNhgg9hG-fX{X3LMQNwVJ4I=y#5*Nv
zr^GuYX{XFPWof6(J7sC7!aEgdr@}iGX{X9NRcWWnJ5_0?#yd4>r^Y)qX{XLRb!n&0
zJ9TNN!8;9Ur@=c7X{X6MO=+jeJ56b)#XBu&r^P!hX{XIQZE2^?J8fyF!#f>mr^7oP
z*y;ZS;5)qkogcnkjSSra->k-mCLy(G=vMgFHPoY_J26kw&^?fKBRc5Pjq9LG*CHKu
z>3XEYF5NWguuDIphc5lN9=h}`(qostM|$kiPm><I3?l~UGK?Fb%g`bNb{TqPz%Iiy
z8L-PZVuUW^xDmRHEiz)4u}4PiGES2byG$b{=rWC)pv%-E6Ly(;WWp}fG?}o=JYt3}
z^SBwh%q=owm$^q~>@rW28M`ba7U;5!TcFF*A`5m|dSt;a%QRWA%Q|9(F6+1zx~wg-
zVwbf?R_wA)lNGycBR1%=joYBh)*>5r*?MHdF55KOu**JThc5fL9lGo-vSXLMM|SM8
zPm>+H93u|sa*R8m%h4hSb~$?Fz%Iu$Ik3w);)E{exD&dZEplR)vqw(sa!!*IyIdnK
z=yHv_pv%=F7k0UN<ial3G`X<LJ>rHg_qZFn+%0lrm%B%9>~c?&8@oIs9_aFnd!WnH
zA`f<XdgQ?_&op_k%RAzQF7LP(y1Xs&Vwbl^UhMKtllP#j{}|{e;_&{1ANU%%E`}Zx
z@O5!p3_XE)R}4Lk`HV93kOKQ_%FsIIEoJB|<|WP0!<vJA-MY}Br`r}f^mMyIho0`V
zqB{e9x-~_Io^DIgp{HBY=+M)z3q5-JZJ|d`zbo|U>CY&7ygvP!qDN1^rRdSqFKP7X
z8P<gXJ;Syzpl8?>2J{T46$4(MVNEfhXV_8<=oywY2K0>U!ib)6TNu$Z?g}G%#?y)s
zJ>!~UM9;XT7|}B>X^iNZ)`bZ@)3z|7XWA7e^h~D}6MCjK#e|+|OEIBmTGE)%Gp`FX
zdgg6mM$fz}%;=d<D`xb}Yl;~?^Oj;p&%C5Dqi0zc7W6FJ!h)V<S6I-qoK`I8S=JN_
zdX_E4f}UkbV?oclF0ANTw}llw>#ne(XFa1>AA<XXbxpCNXWde)=vkLER`hJ^!iJu0
zTiDRE?Ft)uw$q9Yug|uo*wC|WDK_+MOBx$`_H|)L&%Q0}=-GFL9X<OQ#g5l!UsLSp
z*|!usdiEua9X-doaG>Yd77p|ryTXB<<BZ}!&#|UB&~t1l4)h#L8V7pLb>T$Mxh<UN
zId_E<J?Ck~i62krn&L#yxurPKb1rF|=(*N~3q9AiaG~eg6)yB#XA~D+pKDEVq37CC
zT<E!$G%obq>%xtmdt12CbMFc_dhXMT8?Vp3rnu2_Zz*o{+)EladY*ORLC>=-Jm`6L
zg$F&)8O4Lw=UG!c=y|pj4|<*@jR!sNy6~ds-4<T-yt~4Sp7)I6MbEpYc+vB2DPHuv
zOByeFz765~y6FGPk$*+9BYaO{-V?rGV?L$$PAjm#s`$Q!c~kK{j(Jhz`-<jZ-@hUJ
z==pbqA3gt`@T2EHsrXMppMO>HqvzjL{OI`?HGcF08zO*SU`GVd3+#yidVy0)0Ix5w
zsszvrY$^ft0*jgedch45L@&4_g6IYJL=e5;NhOHa7hF|>=mj^GAbP<?O%T1%h6te-
z+7Ti2LVF^FUg)F}LNBzcgwP9ZDk1bji<%I6;SCW+FT5kd=!N%07`^aGC5&EpRSBaP
z-c-Wqg%>qp^dcJ~f?i}tM9_=ui3oa;lS%}=$f^=SFS4mb(2Fc;BIrdoL=?T~j)<Zc
z-4jvtq9>IodeK!Sie7Y6iJ})>)I`yXZHO3pu^kaZFSaLQ=*3PcvD0vWh^;Cy^kSPz
z487Q*CWc;oL&VXG?}#{h@jVeoFMd*q<MqW?l{k9wO(l+Ad{GleFR>vK=p}YU0=>kZ
zNT8QEr6lnB601rAy~L)HKrgYVNuZb95J~isJ0gi*a!(}DOP*4a=p|Q`BznnBC5c{g
zQIkY3wINdIrFKLLz0{sap_e+Tr10aJT2)f$r8bomdZ|TC3cd7(NTZkD5oz?&dm@cq
z`jnE!>r1aHY4p;YN*cZNq9%=AW<zAq%j}2@dYL_uK`(Ps$>8;6R+S8TnN9wi9en;~
z7Bw04vKu0cUUo-h(aY|MEPB~fN*1p#yQ*Z-%Wf)J^s<YZEPA;OkwY)HBXa2F_CyZ7
z+$kl8UT#&%p_kiKa_Hq2H97Pq9}|-s0)FNEgqYmH{Ir<d!+b`Wd`N-)b!GA~%x9I!
zZOjjACLhrp?B^d7dGzv6h&+1vr$rvU{23*WUVdH4qnAIc<k8DNtjVKScuW+~D?A|z
z=oOw81@sCJDTOb=`(IdB3g{KiDh2ck4{HkOP0?dwicp-QC-?--DSDbuP@SSP$`s*s
zPSHA_pgcuq`2_7LdRQ|>sH?22lCH9@O1jFrD(Ncgs-&x|a~FIZRMu5VS6Npjy~uh|
z(u=GYCB4XcQPPX77bU&OI(LCyWW6ZqMb?XwUSho@=_S@nl3rrHB<UsAOOjq<ox4CU
zv0jq&66+;NFSA~j^fK#ZNiVZrmh>{~Wl1ly&Rw9FSuaa^nf0=yS6Ht|dWH3hq*qw4
zNP30!ilkRq=PuAItXCwx!g@v0tE^Wgy~=u3(yOdjCB4deRnn`ha~J4U)~k|UWxXot
zHP&mAUSqu`={44Cl3ruICh0ZSxeN3f>orNQv0jt(I_q^wud`m4^g8QxNw2eBm-IU8
z+y#1_^}3|jS+7fagY|}_H&|~-dV}?bq&HY^NP2^H?gG8RdPCA1tT!aR$$C@Lo2)k_
zy~%o0(wi(d3BSfRS?4a$o2)k_y~%o0(p#*zB)!FYOVV4cw<NvA*WQw@y~R3rf!<=h
zCFw2JTaw;py)Ef&*4vWaX1y)xZI;{8wYOR4F3{Vow<W#JdRx*vtal{6!+J;3JFIsk
zy~BD((mSkk7w8?<JCfdEy#xAn|MBGyJXZSU1+VAlw{&Uf!}v$QF6l7qyQIf#=#l}m
zu}enGrY@N<o4aJjZ0V8(v$acB%(gDsFx$Ii$L#2m1GBSBPRy<@xiGuC<i_mjk_WT5
zOJ2;rF8Mx;&*$ipA9J8f0nEWJ1u=)Z6v7<tQW$fjOA*Y`E=4iNx)j44?@}CdqDu+P
z$u1=^r@EBFobFN@bEZof%-JqwG3UCJ!#sH#P4@8l3%5}obKy2BV4kAe_}K_}qVkC<
zofP?`D4mq}q$Hh``J^nJRQRMKomBaxDxK8$q$Zuz`J^tLH29<;oizERDV?<Vq$QoS
z`J^qKboitLCpQ>wd+)K=wm(Sc4<0uj_w=9D*Z(~g9o~P+UH{Kg(eaPGH}=q(r+EA5
z(W964pI+C09-i*-{`2qp+k8J=rlAeIpDvRQvwoTMm<`Khz-(M5BWBYwnJ}A|$&A^u
zOcu=6WwK(nEt3tieVOc-9n0jv>|7=%X4f*gFuRw@joGtI9?afl@?!R_knbCKKV2a|
z=D-RCFb7vCh&i-EA<W?w3S*9}Py};yg`${aD-^>VU!geW#0n)aCs!zmIkiG5%;^<M
zW6rEl26J|WvY2x#l*2svC{2DH@28Ja9&_PQDqx<XNBMq=Cn}$)(n*m|iqc7mPfF5B
znNP~nNrg`;(n*z1s?te~PioRholol0NrO)s(n*s~n$k&&Pg>GRn@`%(Nrz85aMHi4
z@pEa;J&pC?F6JBd-O|6e@^7uje3Oux<hvC<-F!Xr-GzCEe4l{qANE6+f6NbE{wDdc
z%ikkEcKK(>k6nS`0CWY$0?-v`QUJRGJqln~V1@$N6&wyiS8yx{UBM;=u`Af4Aa(_3
zD2QF5;Sh9%#zN2)YElThLOlv$S7?Sp*cBcQLsxh#3|-+Sg|RE#qcC=bXDEzak>Ln*
zMaCk~6=_lgyCOY`U{_>@BG?rjjzU*-EDBxGCPlF;+M_6TMQ13AU9sU9bj8ME&=qS^
z47*}IieXo5hGN(iAC5y;d@K%K@g~KwE8e3xcEx8Xj$Mi21au|F63~@sQUbdYJxX9#
zVuljfl^jk&S8^-~UCAaTu`Ai5Bz7fdD2ZLE;S_YG#!}FgYElZjQawsxS89e**p(hm
zLsxn%4PEIbrLimBqcnD<XDE$bnc)m{WyUhlm1$B2yD~k>U{_{_GT4<J&O%pqEDK%P
zCS|cJ+oLRYWoIahUAf^Lbmhi!(3NXa4!d$a%3)V-hH?j8wY{rfYT9TY=O0ckI(pHI
z`_DeSy}f<aTdsK3N6$TM&<>g4882ELB)wJ-$)GhrGHQ*GOj;8pv(^mBqP0M>YORoL
zS{o$0)(**`bwF}zose8w7bLgV4auYRK=Nw6kbGL7>EL^TU+aey&;}p{wLwTBZ3t3W
z8-^6oMj%DCQAjau3{qShhm_DJASJa)NGWX!Qd*mal+k7&WwlvIIc*Noq;}E-zwOoL
zAr-U*NK>SpBFHLfRm?@w7BQDdTf$r>Z5eZgv=z)%(pE9oNL#~PCv6>bgR~9IP0}_o
zw@Ay+%RIQYHfh_KJEZL#<lFClrSm1P`uT6g`Gc38xa>{+m#3URc%|mbzwN)8rC-KB
z-k*ETOaG^d&-|h^5qR<=w<_lko;dHsg$H+)8?U+X`h!oUTh6=X!hXYN$4@`+O*h?h
z4d3aGU3lzR|M_fvh=Fsjndt+MoV_dY(?bvYc>Br|S33Gz-&y$N%k|BVKFP;}#zANQ
z;Y9y^=aO@;tLc~Y-z`2q_iCE{@iRxY)<<9a;vaI!WpBAG(EpG7udZqE`nlKE^w;^T
zfl5PVd~l+Y4zpe*J!XSS2Fym4jF?UQS%lfFk{Po_B@1S&N><D^m28;pD%mkRRB~W;
zs^rA%Qpttct&$tFM<owtuS#CbK9ziB{Ao}nKjwf+0n9;_f|x@pg)oOz3S*9_6u}%-
zDT+C!QVerkr8wq<N(szKm6DiKDy1-|RZ3&dsFcB+RVj-(r&12{<mlx4`k&_fl8A}P
z_d_buWCb3$OtxwA1DNli$-6Mm(B#J<=STC<ou9}<cfLq@?9R6-kKOq@D39Iw8OmdK
zVYC3<g^2=m7m8HC?n0Xi*j>1T3fNtkp#pYK(dZOG`xH$~5%f>-L$4`*Aa(GOn&JmG
zQ-sD8-N7fQOwkOVprg{LijGPXDmp4Vs^qBbsFI_yqe_m-jw(4t8ZDwzq=_OrMRtmk
zQ)H(oIYoAgl2c@-C^;n>EumAQi4r;`-Y-c`iJg+<l-MarPKlk8<dkW&j82&*%IK8Y
zDN9b7owDSV*(pm-nVquaRA{t<PK72a=v3ILNKS>FisV$-sYp(Ror>gCX|#$?l_sj_
zRN1LYPL-Xi<W$+IN=}uXs^rvYw1!TNCTi%^*r`cQjh&k0)Yz#>PK}+K<kV@jj!vB>
z>gd$jsY_0sox0@I*{MrTot?VmG-$MePJ<>I=rq`ANKS*DhU7HZX-H0kordHzX|#z>
zlO~$zG}&oNPLrLc<TTl7N=}oVrsT9}w1rNKCR*sU*l9^li=CF_wAg7$PK%wE<g{tD
zjZT{;+UT6S@1OfTKKT09=G>N?Hal&}X|vOooDPk4(CN@b2b~T(9m(l%Lq~Et>~tij
z!%hdB(e%6H=U$C{^Jwzse>(TZrRQ$`|DR+1HwMJu%OCmhFSJj4|Ma|Lm%jKV#|K{W
P*U`X>58ga}@aF#knM>jx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_instance.hpp.7A273F666FE77357.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_instance.hpp.7A273F666FE77357.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ef82df0d6f989f50d6642d5b1034674afb6a4a51
GIT binary patch
literal 22174
zcmY+M33#LhndkF<U(r<44P$fDAk8D=3XVLhPDrR+m7`KFk1PTT#$<A*l1e;C7e}Xr
zko&62QRvEyz$TbUQTJWb&30pS5_Wo=PNSeRJxq6j?h3m)Pf!=#Sv+w6iBsnP@;ns%
z{`vpjU%u+3j?`Cm-<?k9QxBAsH1ym#^hMXhotBc4l0*3Sp@%*F?@CF@llw|ahMxWA
zS6}(k`J1ks_Fmck<Lu#wvW-vPe*fp5ay(sn|JPo9%<;oJoPRrd@X@c_Km2c!{QQ4R
zoOGW5@OAU{%Vt(qx_{|<c%}CXE3W#ZcU=CLkDvIHbGN*@c;O?B|LtQx*#F@p$*;Wf
zyE`8Gkm-9W@c8fQK6UcZL!bKci`V|}Q`K$%BEP@z&OerXD)(UX?alvmxBG9N`@+a`
zSGC{wiSm7a_1T-Ac=e-SoeYnT|MMGfKKAqR{}cQ1U;J|PBg;?T{nL^}h3D(v{(QbE
z^yxRw-c##uIM`Ny_ve2y?Y?bo;=|R?U;o+tm;L0nH;=p$Eq&YZ_Q1_DRr;^rx%S<=
zPW<3sPOZNBt7Pr<`Lo}@b*{DMy3^A2n)``A?|bYG+dU6_)^UA!_2Wx-l#c)H<F}7|
z|K59k{%=*EeCluR9DnDZ?!C%6+i!p38GE?(?nmEy-@ZG3@yYA6)q@+ux7a+d-5CDl
zmwtc4NdF5bemXw*jk~u0)0ckxAMX0gzl%NdX0C4hf4|r5mv`fRPs9dq{Y9btzZ}?q
z-QT_Ny$}9y<G#BS?({1chpqR2W;vd|<4?c;)R#UOsQ%0)e|oj=J5QUvkNy7fTjt-r
zc<$xYH9xy+q2<%*TNYkQ9sbO(@4I^4v45%UV=X`MzHNWzm+${^^~>M+f5+bY?(b^r
z_y5Bq4|?8r^{)fpdhK(*FE+-$e^1@a%U`Sd+6~WNIQ7Qs|M1Rh$3A<-!7pd-XiGi&
zf6pCuzjx7J9eC)ml83xs?0Oa70hj#xxz{H?dHbgi7?&O@IdbE^8xQSXl<dcMNlA$u
zkwdAu!;d6?x_Q931cl4@UA}+!Vz2OlC*I9I_`RdY4;WXVaPPi*Z{NKr*>5raWOu_y
zPjnxs=>Lnd1ICpod}QB8N_Q`I3fJxy?7#WNcK9uC>VR<>3WxU{zIgYdWWR|oaOZ3I
z#Fsr6Dqs1%_keLV3Qx<vr^Wg2|Jm7c&2DL;<FGUKu|I$GfN>E@hxQ%1WcOn43r#f-
zec;xQzc_!u*oVTE`>x!-d$CtYe`WuPeb=YHe!w`0!d3gOvhH5&6~6acrSIYV*DgI^
zT!q3T`yMIXz1aC0uG{?@{_(`szq;d*%z;ZwE=J*|t8cnx_hNU$<?q9h{A}+2A6<II
zk)=yZtSEfo>JQw$d$CveW%?J{ma$iEJz(rd;hKHd9NN9u`x<`tz6Y{}2VZ*YfN=nY
zEB0NnfA?aq;Boanbmx70f3s)hlHI@3g^k}l7ynV^O$U#Z2BkDD#lIHJIVoM|c~eSX
zrnxLAW!x{zNg4OcHl>XF<v}UuetAyHxnI61<=nRg#m0SGPHfz_ZHkTi6+x-senn0y
zxL>g;72K~3N+tIzb5hCu%1x={epOJaxL=i%D(+WpN)`93gHp}?>YP+_zj{-uxnC2M
z8t&KRq=x%7n^MDldr<7$x97ypefy@^xnCQUTJG28q?Y@&n^MdDx}el?zb+?r+^^e|
zI_}p8rJnosIjQG<{if7&zac0M+;7N91NR#?rGfj7pg6ei$ccmdj!kiJzcDC{+;7ZD
zBljCOrIGv2pg6hj%!!lx&P{Q0zbPn9X(`^$9X!&Mlcsf^H>K%inwx{t%>Cw^G;_as
zQ<}No5|kG1x8$US`z@Q&!u{5uv~s^SC#~FX-IP}Dw*{q*`)xUC<9^$wv~j;ZDDB*D
z&q+J?+c%}1`yD~);C@F=I=J7lDIMJJ3`!^WJ9E;>{mxD4<bGFBy13t!lP>OeZAusS
zyMxlr{qCG}bH962y1CyIlpgN)<fMoDJ)6?QeOFLi+;`=~#eLVNxVYaNlwR)l=A@VV
zy_?d@{l1{|albDoecbQcls@jegW~4CJ11`LyEnzn{r;fzbH6_){oL>0lz#3H1Z9Bx
z134Ms{=lXTaDVuD8O})Y&hp@q;k*ogi|3bQ_`5WZJTD{MAIZxI_eWlm5$=!5^D-(-
zj7lB}Go$hn5~jvMjVo##)VQL?L5(YF0@Q?}CO}OnY68@Rq9#F2Dryqcq@pH4O)6>%
z)Rdy8KusxX3e=RMra?_BY8uqEqNYJjD{2PRjG|^h%_wRH)QqBLLCq>^7SybwW<kv=
zY7W$#qUJ!&DQXVXoT7Z7e2Vgc@+ry(%BLtlD8Hiop!|yRgYqkC9@M;|=0VLXY97?Q
zq830cC~5)Jf}$2cEhuUc)S{vmK`kn35!9lhmOw2jY6;YmqLx4{DQXqeDyZTE=z~XA
zL9K$~<Rv8ErDP4%nxfV~ttn~^t~EskK?M~R1Qk?N5L8f6Ay6Sjg+PTA6#^AfR2WoP
zQDIPFMTJ3y6%_#$QB(v}L{Sk?5k*BoMHLkV6;)IeR8&zhP%%ZtK*bal0~J$L98_FU
zaZqtZ#X-dtl>n7cR032&Q3+59MI}Kc6_o^)R8$gFQc)>TDMh6~r4*F{l~Pn1R9aDK
zP-#V_L8TRy0hLiy22@5-8BiHTWkF>Xl?9bmR2EcLQ8`dKMdd){6qN&&Q&b+*&R4Ye
z)0D@YS5zKUo~Ysz>Y+<^pGecwvZJLB)048H(#L4>OX(s#B`aN#(p8$vj+PNyHdIDz
znP1A7EnAT?X3LM36I(u1PHeef%9$-+k#c5jM{UH~hHS*z{9<F)wjwrWD~?tWTQO8Y
zY=vJcn5|fm3T7*hRuWq|R7q^5Un-fcT#-s<tBzI?TQyWgY?WWCn5|loDrT#XRufx2
zR84HPU#gj{UXf~MYmU|sTQgKcY>i)Pn5|in8fNWB?Zn!L?8MsrVrSOAB6enLkJb`f
zJ5)<-tzT-HtzD5?X6ugD5nDG@M{J#6>X@xtkveAUkJb}gKU7a_y<h5?tzVIPW*d$+
z5Zf@+Kx~6w8klWZkp^ZRM;*jEh8)B?{NiBNu_6v;8;>>;+c?xnY@=TqnQdH=MrNHy
zoy0naoWwf);$+sjB2H#k#j`4p?A?s4ig#5W=lMC@N&hw7(5#7Pjfpk!t}(GD&*2W5
ziJ*9bOa#RnWFiP6sECkwLQI6j8)6~^BBY40c*0DC#T#ZK3?i(Eh<GAQM8q3mA_5|!
zh^Tm?Ohm;SWg-e9s)(3)Vob!u8)G5{BBqGAc;ZaN#T#cL4kE6Igm@B6B*dFwA^{?y
zh@^OuOeDpdWFiS7sfd(#QcR@8n_?mbBBhA5c+yOy#hYd#4I-_GjCe9kWW<|cA_F3$
zh^%<BOk~BIWg-hAtB9O<a!lmJn`0seBBzMFc=Al-#hYg$4<b)Q@s~CD-RI74q>Fd5
z#XD5QJHLl6#>HPi7tBj`fAd_-@vG4-qx7s%#4s0((${(3HcH>-`IK4uqDlQTX6Y8s
z=gra!Ja1S^H!XYpvMr;GePsosjD2O>Mj89cPMT$>XkXbGvy6RZ=gl(qm2Ftc*jK(~
zl(VnAV3f13eA_5zU->DsoR6>kj9JdU^7Cdn`^q;g<?OR<88-IW3Wkk+wr#`4KHEvt
z#>ZznW7^nfJ8#<9XWOvY*jKS-RIsn2U{tWLV%w--U&TqYf_)Wd%nJ5ZoHr}jSFvHK
zU|;2yQOUl_f>Fu7%59^PeU&H8O7>NrF)P_udETsKU*(3Sl6_TMMiu+23Pu(Cs<w?P
z_EnuUtJqg{#;jsr)p@gueN`KlD)v=x8P)8oE*RD9tKK%M*;jqitY%;J8MB&w)#uG>
z_Em3Ks@YewWz?{*reM^tuV&k*VPDNDv*tzmd8j#K*08VUyjjD(nhi@0`|MkWoqhI#
zVP~Ix+px3Ge$uq_@!8LqcJ|rNn|AiuH!ODc)ovNJ?5iypwd||iHfq^dd&;ck<EuSm
z*0Qhmyjjb>+6_xB`|7rgI`-8Sj5_w!Z5ws$t2<@Zv9IooS;xM*^JX3U>NYHO?5p20
z>e*LcFzVS?zirgBul}T2&(~-D8MB^!_2<oc_SJ7#>e<(@Wi+s_p<pzyuVLG0U|+*2
zvw@GV;f&e9zJ~K=1N#~_EDh{)Y#9#rISPh@eU5Fz!9K@H)4|8*IAc24=QwXV*yq@=
zIM~;?Wi+y{v0yZ^uW{RGWMAVcvyqRl@r>EXzQ*%rBl{XRERF1QZW&JYISYo9ea>ye
z$v)>P)5$*P8Pmx==Xuk~KIewT$-bs7qv@<sy!XaSEd`_Lb)L74rnh-MVK%*BQvbBs
z^!GfUGn@XI=XFcdw=8@8<}IU{ea!`<nSIUMMl<`GkDJXWXkYVbvzdL(=gemIHLqKm
z+1IjVw6L$GV6?EWW!q?BU&{%zg^#c0wAsSGmUCtc`&!m5E$nOEGFsW!S}<DK*Sc-A
zvaj{H*~-V)dfIGdU+X!um3^)2mR9z)Z5eIsYbzLS>}%UL+Su22+-zfC+iA0neQoE=
zHuklxTiV#yzGbwtuf1Tjv#)*IXlGyhakHI$?WfIl_O+ig+u7H?ZfR#<$ClB-zK(*?
z!M=`dql0}N$ITA*b(}Uk*w=B+>|kHVx}}4Com)mH`#K9oC;K|LjZXG;9ydGL*Lm9P
zWMAhwvy*+D>y}RTb!{14?CUBRUF_@HHoDl?b;9g=fqovkPMclq>pEw4v9D|0(#5{+
zEu))#-36nYecjteH~YGeo85eT-KWiN_H~~#yV=*hZs}%U&z8}{zMg{7!@izvqlbMx
zC(Ir`zMj)&5Bqw~nLX_5S-14C&$VT^*yk!3F7~;$4Hx@dCrlUnT&GPJ`&{Qt7yDf6
z78m<^w~Su)^%jg?_VsQXz3l5fZuau^*?ZdTWnb?(vzL9n>y}>j^=%n_?CUESeeCPo
zHu~7tcf#!B<Lf(Z_OY+;oY}{|zI96<``lZGn|<zr;bxzE+i<hbecW{O@wrc%ZuYs)
znQr#E*DY@L^=}#d?CUQW{p{=CHu~Auf5PnN<Lf_d_Oq}5oY~L5{&h<~`v$g*0rm|P
zi~;ryY#RgY8#rMOuy5eBIl#VwbLIg12G%VD>>EC748Ls9OSP{X!*B5XwlVxOo==&>
zFPhZfGKbIde8C(p@Vsdme#x@eA31A`uy5pbV}yMpZyO`*8#!f;uy16`9AV$c1#^Ub
zBb$~H_KnI}V^r8UDz779=cv4mgsr3UqB$zxq~jQsEhKCnl?zDNJt~_PyzERzI}ZD}
zvX8?)uI%Hmk1PAQoHEDp73^JW<FI3wN-iK_`#9`)`I_t#uumxa1nd*aJ^}lLvQNN1
zq3jc|V;9*caO@MxJ^?#k1}FO@?32nq3Hzk7Pr^Q_?31uhD*Ght*hTh9*e8{J5_Y_t
zPWCC-r<8pP_9<nbf_+Nar(mB__9@u0i|kXdPbvEp?08w8?9;GMEBiF;)5<;#`?Ru8
z!#=I-)39R~*{5NjR`zMw@$x>|XJDUE_8Hh`lzj&F8D*b=eMZ@5V8<@9&%i#T>@%?A
zGy$^D!al3)v#`%9`z-9U%03JGtg_F-j$LG*g?(1pXJN<b24tUueNNfuV4qX=IoRiv
zeGc|HWuJo`yU0EV`<$}R!H&}w$nJyPr|dr1eah~G-KXq6*nP_GgB`ob?t|T@>^|6W
z`UBbhu=|zW54&I4{jmF$-4DB8+5NC%7uo%=`<2}fJ5Hk@`#kLP%03VKyt2>3KCkTa
zFwd)7+IiTqi|q5T&nx>p>^L2R><h3jDEk8J3(CF#`+~AB;Mf<`u`j@mU1VQ?eL>k5
zV8>}4WM71RQP~$^UsU!**cX+35#~j8?2E8t7ugqKUsU!**l~IY*_U8nQuZa-my~@8
z_9bOsf_+KZmte;(vM<5Dr0h$u<1`bpufo0xJHJ--I+8b3@-~v6spOQoim!mFEhJ}E
zasf#}C7YI2*!QS4*w>VO4adHw>}xpoHDzDJ-}{=fui@CSi(G3s_BCZ+!?EME7qSOo
z4=Q^Q_Mox{VGk;M5auAR`MskE!j4@kxqyU^JqSBapCNk)_K>oNU=JyK2=<V&hj8p6
zb?hP7v5V{>9D7LFL$KpC9I}UD4=Z~Z_OP;tVGk>N81}HThhfJqvWH<0D|;AroX$h`
z2<#DMkH8*L_6Y0|WskrfQT7Pz*hTgT>=9*;z>d>`$R33~s_aqNqsktIJ*w<c*rUoG
zg&n)d9)&%s>`~ZpdJ)-Uu*Z}=2764|W3b1RJqCMB*<-L{7ujR5$CNz=J5Ey~dmQ$-
zvd3YMD|;OFxU$D#k1Kl|cI+a19QL@f$6?3mPGnEOo>2A#><ML0z@AX{1ndcAPr!~{
zWKY1JQ1%4uIBkmTN!XLho`gNA>`B;@%ASNhsq9JEv5V|U*ptehgdL|}kv#=_O4(De
zr<6SfdrH|;u&0zg1v_?;Jq3G8*;BCNG%m8IVNWZ28uqlZr(sVkdm8q%vZrCkF0!X#
zPb+&GcAO4I_6+P9WzWE#QT7b%8D-DFoKd&58Q8Il>>1cI%ASE8r<IXC3wu`Cv#@8C
zJqvqQ*|RwItUC5A?AS&2EbLij&%%z=)5xBKJ*VtB*mKIBgFUD0Ihb?m*mJOB7uj>L
z=afAMJ5F;Wdmi?@vgcvXD|;UHyt3zE&ntT#cI+a19`?Ml=V8a`a(}cJ?;-JW{U6@+
zReWLoKbL>y!K-hZ56HnIH{Nk$L-9>)?L*H#*zwI5|0m)@hYnp^d`DQ<$L~CKMd$hp
zhjvbX{S_WT{PvHAgGbhotf?f3B&d=Ql8{QmNWv<KAc?3XiX^I%7?PMu;z;5uNgzq6
zB#9)ck`$7ZO43NuD#;+ps3eOdtCAd&oJ#UY@{}C4lzct%(cgS;_o{mfSKT}6s(TAp
z9VbC~3s)T{AtWJ{gpq_*5<wDCNfb#`B{3v1mBf+6RgyrGP)QO=QY9%QDV3y=q*anZ
zl2J((NmeB}BsrDjk>n{Ue)l#0;eCk3Z(Y2mHQ4m<-YZZ|$C@6a$tO*V^j4y#WocTa
zxp}ae*ydx+#5Vh+nc3!LX=b)%u!Y!`V=ctC_@sr|mSt&Swso+T*w$mM#J2jRmD$#1
zX=S!;u#MQZV{OE?`J|26wq<E!wtcXj*!E-X#J2mSo!RzfX=k=$u!Go+V;#hH_@sl`
zj%Dd!wsWwP*v?~}#CH0mliAK?>14KRu#4EPV_n2{`J{{4u4U<BwtKLf*zRN9#CH3n
zo7wJV>1MWPu!q>5V?D(7_@sx~o@MD_)-~uN)^*H9tji}ZW?jqTVzzg%m)PE8y~OtV
zq?g&=W$9(MZ?KQpzGHpF_W7ib*}i4zW7a+BCf0q-O|07|Zf4!f;%2sgu%FoeWBtVT
z`=p=Q{$=TBc3^OT*nwjM#18mmfZ2g%8Q5cIe%^S4?Yr}basPDrq06r+o(b;1_}kx@
z-SCVezUJC%-d}tR@K@(w|LA>3KJ;H5IPJ_@YNM0HtYtKnTgz#(S#2~`SSx6%v{uqo
zWv!yA+FDIhjkSg*yVXupt+ke>I%^$G_11ct8mtX8Ijjzv8m)~qIjv5bnygK>y_>{l
zYcowP))tytt*tb*S=(r8x3<&NVeO!))7nW>m$i$gZfiG9J=PwYTviuNz1Ciu`mB94
zxvg%R`mOym4Oj<g8nzDG_I}krVjZDrRIH;y^SD^Yd7cpK1kaOVo#c5+tW!Ksi*=gk
z8L`grJS)~&p6A3m$FonYKA!zz#Yz5qU;VsT=XqWb>jKY<VqN5UNvumeuZne5Y{iFt
zI}gsRYdi<Viua1|{r`|yLp+DY8s<47)(Fp0u||20i8aP^T&!`P6Jky9oD^%4=ag7e
zJg3E)<~bwQ49{7yW_iwu74LQ5JD$8)^Lz7Gmp@p0-G_eqj&bnFO@DUNJ;e_fIC$h{
z%gwhIKS?0=uM^(S{pi_6%fTbJ?Z54^y<3TUtM9Gf`zd;K|D%`f{XIP+`=61^ir)Cp
zsf+IU;;Iw3KZh?ne0A|`*y86UTz<*!=OpYsfLWFI@JMj%q9VWZ_w*hf333ud5>!bD
zNk}DOBw>|AkVI4xMG{p>3`tBSaU^k-B#<Ojl0=eJNeW3yC21sSm1K})RFXxKRY?v>
zP9=FHc}j|(D}gI==R+m<ik!G;Ri@MxIdRdNN`f+RQBWlzBq5cAk%UzeK@w3(6iHMi
zF(fgS#F4~Rl0cGBNfJp?B`G8+m86lRRgyuHQArj_RwX$kIhEjzwin?Iuz$RM@*moN
zf8gvleu8h*s}5atRq=ZL?}cZ&{@1>j9zcBapWSRP{$v!t-SO;k=c77`S9|ft2=UGb
zbQI$^-*mY%56t}@ORhY0<<-S+{^ByUcRsD7*e>2j<H7yTXLJ-}+*1dn^cj9*RzS*l
zE)PgK&$fWrc&-RY1<#cMspPpTAXPk92c(+knt;^sY!8T?=h}eO@>~~?I-ctTQqOZk
zKpJ><1jNB}V?Y{tb_T@Bb5lT?p5Z4j0cqyBB_J(4w+5t@=eB^f@!TGecAh%|(!q0Q
zKstHu3P=~v-2v(5xhEh!Ji7wo;<+~<y*&2?q>pEJK-@g{2c)0pfq)F~Jp7~#Kg;iZ
ze^N$x9+fBYzI9H<k&LTk0?C9*CXr04WD3cYN~V!at7Hbrj7nya%&KG#$(%}jNPH^s
zBk`+b9?85)7LY8cWD&`tN|ul;sbm$&DiXe@c@l4H;v|SpP$eNGA(e!YgjEtj5>ZJM
zNmL~<Br%o5k;GM!K$1{N5=l}eDI_VCq>-dml0lMDNft>~B{?KHmE@7+Df#UoYt`T1
zcFq65jo3~1-SoNQO<nObgPz8%*v=;g(aqJEluq)^)tHp=Ts|h{Jln>^#&gA(RPbCm
zCY3x_jY$>H)nii4bIq94@N6FwJI}RaQp<DQnAGuHKPL4&H;hRG&yF#1@Z2~ijXXQY
z#L08hm^4lD&DEGR^V~8fEj+i5Nh{B7W75WR`<S%z+%YB{Ja>*sC(m7D(#3Q4m~`{p
zGbTMeyT-)DbMKh+^4vEjeLTCz#LaX6nDp~JFeU>$4||3mDc+CbgFL*$kJB_R!&7uq
zHasuGOFTa(!+%Zlh-ZX$jd(|B*T}ex@UD@08R1<c&&dey8WlWUrfs9*9TnO)ikr+)
z+yL%<wWGK}8pX{EC(j|_t>fYu=dI)79p|m%*gCGZj$`Y%+B%MRqK&Jq6XKcRtrOy%
z;H?wrPpGXE*gB!MPGIYV+BzwoN!~gs-bvm%iLH}r>m;^Ls;!gQI;pl!iD!zpPKkGl
zw@zW}l-fFlty6026t+&Ot<&O}=B?A>o#w66*gCDYPGjq|+B%J`(`xICcxHI(jCf~w
z>kPKesI4>DI-|DEVC#(9IxC)8-a0GZS>8H}t+Q(DEVj<7t+UuVtG3RGXO6ediFb~-
z&SC4E+B%1=b871xw$7=oKJoZ?t53W>-s;0vpW5ofR-fAH!&aZ#>KBinxBA8F=dFHh
z^{cIZZ1t<Her)xtt@GlU=dJVNo#(APKSw*?Nqaw^^O)z=)_H85S6dgvv%p&y#Jj*-
z7qE3fZC${I1+{enTNl*UMe!{1)<y9y^43LcT~u2av2{^xUBuQ!wRK56OT2YSyi2@w
z30s%c)}`I!=i96$Y+X`YS3%)9Bt1l01+|J6KWG}qgC_2*p|hqsLF@^tP7s}->V(k2
zGf&zaLMNm;VRZ0ZlsaK_!m1NN2hUEa6G11UI#G1+e3d#;bfT&gLkG`bsS`sdraEzS
z@SK)9adhIUlRyX0a;cL*C!sn?bnv{FI!Sbrs*^$o&xEOyLMNp<X>{=1m^x{6(yEg|
z2hWzNlR+n=I$3n^{Fypgbh4_GLkG{Osgpw|r#d?yfwl7?Q+wYnJ0E<s^U*@Zq{QR;
zLg&t>fj;Eve&`eUuW_XF+ZSIS_~z7I7ahLg`uE>ha>LI5{lh<pi+?C7xqk1D@8A35
IjeCFm|82BxtpET3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_scale_instance.hpp.A306BD940C6CC8E3.idx b/.cache/clangd/index/device_grouped_conv_bwd_data_xdl_scale_instance.hpp.A306BD940C6CC8E3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d25ebcfe9b3a2719f7aebffe40fcf38609c44a93
GIT binary patch
literal 15542
zcmY+L3vgt0mBw@aw;6gKfHTM-GY(8`t?R<9tHf^m{cd0>YC&|3-P`YXBMM1oG>IWi
zZoi+%LQ*Tjl89uZD}owQ%p{00B1_vbBCf14gy^b>j}TeGDin0jWUI~hr;1zo&N<)r
zyZ=t#zISeJf8g4yuYTCt)8k2AJN&WKshFmxr{^I5fAZAeO&iqHbIqQfp5gnxc6aF3
z_2X|Xzkc}>Pp9-RKG-SgKji<6_cOW=O@4Lsl=#mpd*Ao;k8gYESGQkw=O;h${r~v(
z6IZO=X6--!-S2kVKXvJ|&42a(&CTk$FW&XY?_Vi<fAlYlrH7u|bD!q5;mPKuPu*iV
z^3Z3Of7US=fA!o|)?dDI+xR1+kAGSD(-(fxq5IR1Exd5k56oZqwduEm7qooig?ZP@
z7d?I1TduzAJtwA*|NhKv7ld!jP2ShI^v>7!KYQ)5>C`p38-9G=J&!zl`QDA!?bipR
z=YMI?^XOgQj@^Cfv-3AUxb*6!PhN9qx$o{P&zyVn*Dr3}^4K~4yY=Uml6TL3@Nna{
z-aF1)p8MW!Zh7jui1DE>-SX-8{m^^mrF);h_Kvxq|K_h=`)l7P?|(<<`fKxR&$*8M
zbMfZQCsqy(?#l$eeo-O+`ZLGYe{;@-xhpK%Z+-NK>o)~94*mYw&pY=2?~AXwzPI|$
z>k6Oy%FDOkbzSL_nMa?w=-K6b&xxV`ddc>_10Q++H7DNn^M&3YEkBic_gB7f<Bk54
zKiK!+<F9U3mOp*_oiDxd(U+ck^wmdiy5hv=Ub=VBKmF+s*Z%av#!KfscGbPVcw+Io
zP1g(G|8)J4U*7(3W#5U1LLZ*{+M_?b;q`kbC!g0n^Y+R<^Qnh__TR5PaQyK<&VF#j
z^oN%}l6O54zT>f5umAI@`!E0V-kaX^|9j8Lk&nf1=HK!;zx(^kKJI_=y8YtN1wBWv
z*mH$>XVbG+Bi`E6Lq}=fAEsX@T&MrwezA`Sj_f&daA(uAm)|No0|y_e9r^gfzxw8W
zq2+-?dk*d0+3XJdo#~CaC%<vk_2>2M;elg^kNxe=W@q3n=QG0b$5*yK|Ki9a`^A19
zIK1cZxjUPly$bK(*?mYa+TW=5e&lES#X%mpVb2Ygoz2;uJa1<p`_$yo_ul@YU+)(O
zc;M)sqx*L@yZbQu=EPm|w_N+!c|GUyz-5Oo)9-9{_n{Cya`hLleD5dri*tD3;GTo~
zb~d{S55B5D_st6)du6}4kOxlgIceV6oZW|ucJ|@TxAc8=<JKo%+%NX>z=eA*+`F^c
zP3Rd+rBD9Uxci#ZzMa>)wfx8Fhkjx?-hHJr`*vR8Zu!8`-WlrsGIf7btmmlr3DCM3
z(qUgWM>_25XGo8I{T%7BZ<rwi_6>7nz`k*YjMz8MkrDf*88TttG)E@vn`g+3ee)cd
zv2U3n3-&E@WWm05hOF4P&XE=Swi&Wv-!?}!?AvF^j(z(a*|G1KAqVyybL7CjbB3JQ
zcg~R$`>q*sVc#`JF6_H!$c=sX9J#UYnIRAMJ#*y2zITSa*!Rwn7yEs))c4iD93%&h
z_RUk@5@_En`LOSsCm;6xv*gFVf1do<56n^k`+<21U_UrZLF@<TDTw{hEQPQinx_!<
z!?P5|et4e3*pJLo1pASBieNuFOHu4c=P8Q)*eu1cADgEb_T#e@$9{aC;@D5jQUd#l
zc}ieEIZH|GC+8`N{nRX_u%DWz6!z1zl*WE~p3>OQ%u)vXnR&`!KRZiV>}TgGi~Z3V
z8vQbLKfn$g9i5}mCqR$Q&=~f|=4cH2xf#k~KQ~7??2pq7jT26c(;QcDM&%h*no)U1
zm1gohlb2@lJd>AZ3OrMgW(qu0kY<WJQ<P?kJX4frN<34NW=cF$l4i<0Q<i4RJX4lt
zDm+t>W-2^Wk!GqqQ<Y|_JX4isYCKbuW@<cBlV<8XQ<rAyJX4ou8a&gGW*R)xkY<`Z
z)0AeKJkykBT0GN|W?DScf|>4l|EhgE=X*F^4fLLZlhttVD5wVY-T|ks-YM#R0JLsE
z2VUK<4!pVs>Cmg2A{~151A6f4hxOpqH%O0O{S@iZYZx$q*D!1Vuc1K(^ctqffL`N(
z5xmA>BY2GsGNRWwMMm_R229{J4V%DgYLE%NrYSO^*F0bbuX)%EUUP%Y=rvD~8NHSP
z3wSNV7VugcWI?ZGiY(~04p_l!9kznk+8`@>ty5%0uWi5vUfZw@ytW3}&}*9_8+z>n
zcJSJV?clXH$c|q76xq@17;u2sG3)@Zqd^YzI;O~hUgv-lyv|`Kc%2P$qSrY^PV~A4
zT;O#LyTI#ekPE%8DRQCLJ>Ukfd)N(LcZ1yMbx)BSy`BLNcs;`&@Om2LL9b_uJm~cf
zc){x(_JY^jATN5oQ{>(CcCYmL!uITCe)l86@Hx9SedFt&q60^d8jsq#7wg^2{ujl*
zo$LMXIA7qdh~DoA_|(`Cy)U5N61^{@UQ~J?S75%Q^sb;@S9+gEy{zf|re=3uw<2`N
z(`^VH@^o85hdkYaqFaPKT}RO&Pq(h<kf&SL=#Zyh5qjk5H-sK}`YoYHo_<l$<NEX+
zMUOoFx}ry(ep#bOo?%58kZ0Hs2ILvGgaLVm1;v2tGjtRK@(k;W0eOaHjRASa6=6i4
zaYGo9XWSA-<QW$fBl3(L#fUuPx?)70aam(To@qsxkZ0NuCghp6gb8`31;vCsQ%5l&
z&$O<XkY`%fn2={)5oY9>H-s5^<}G1Ho_RqrBhTDX%*ZpZD`w=Gmo;YOSyqGvd6o@f
zL7rtxSdeE~P%Ox^bQBBnEbEE|d6s331$ovLVMU&GLs*e#-4a&hSr--S<8VA!JBk%~
z)^){-JnOQ?iagtjup!U3A#BLAZ3!FlYzvAF*JtY}HssmX6&vzw%NiT<>?^{KJo|>Q
zBhS7i?8vh(Dt26-y`$KXXJ1$B$g?kN?8tMh2nX^U8^VD+$Chv)&#|aDkmu+q4&*u3
z6$kPh%NhsroGZeKJm-dRBG0)coXB%7C{BDmogKxAJm<RNM4oe5<3yfoMYxdX+7K?}
zxweE0d9Fpph3j*56c_Sb>xv6`u4Ro2dF~bAMxJ{^xRK}H5^m(V7Zf+H&)ren$aAkN
zZsfU_HE!g2R)hz6o(<tao@YyVkmp%cJh(njNAV!fv#xlM=ULWxkmp?yUgUW<gco_<
zE#XC;cTw>o&)ZSF$n&l%UgUX~HD2WPt%|<yitb%DzJahQ`hJXhTlD=F^%<paL4o<U
z()VrDYf9gbP%mlvzM<Kj_pJ&a@_d`Zhdkf5@FCB4TJfEMJYQS!A<ws__>kvY()f_)
zUlo4j`8S0hdH!wTN1p$T;>Y#*+ln7~{x!vqJpYo$kG#OD2p})8DFVm~Y>NQ$0;iP#
zt}oD50>}%jDFNgKmNWt61y@B7dBIH)L|$-P1d$g!tpt%5Y%4+J1=o}y@`6j6Ao4=1
zB80rqrU)S~v@Jr&3!PR%$P2ZV5b{E6N(gzOB~1u<;Z+evUU*Z4kr&<;VdRBRD`Dh?
z+e#RD;WZ_Uyzr7HjJ(LIh#)VrDI&;=Y>No;BBzxI@*-^|g1pF@5<y;MNfSX{bX7!=
z7u^(5<VCkd6nW7zN^}8^hiF@gA}_k8M3EO=(nOILTNN?n#WqC@d9iH~LtgB(62tYy
z+DZ(0u{9-zyx5W^hP?Qyh$An)DdNbBZ;Lqc;%AgNt}otJ;>e4yDRJb*mo#zYC00cO
zd5KMtKwe^7B#@Ulqa=`*Xe$ZiCDxP#@)Ap$1oD!rB8j}@rbr?$xh;~&OP*Ge_<APW
zN)ma=H6@9><dP<dyws{lAuqKlQpihfixl!wXOtAKFV$93$V;s$DdeS=G%4hzS4A3m
z=}nPFUV2-kk(WNLq;Y-ewvt9(dQC|qFTJEmBQLWmGRVtpiVX5H+aiO!%o!zv>&vv2
z4DvE-N(OnEB~1o-*;SE6UUpMtk(b>TS>$EUC|TrX+e#LB*)=7LyzG)Di@eblG5Q?=
zH)J=&=nJT~#OTYY7nIRO1?D@-=nCp}W%PN}%bL+|YIf(xR>T<c#x}$l^2WBr81lvz
zlriLub(AsWjjbzV$QxVMj3F<#B67&fZHOH5a$6#YyxgLadmQ#Z*HLoF%dIOp<mHw%
zIpmGgiWny(j?)HLkU36UTtVtMEhyuJ>l~*JSCBkT>s&$hI4x_&3F#`+Rf(=LU6tr6
z(^ZMCGF_GED$_X&UI&%wszg_ru1fSg)AJHN&-A=R&oezQ(eq5tOY}U`ISc4{rspMk
zp6PjsUSN7bq8FH6kmv=b7bJRt=>>^iU^-_3y}<N>L@zMCAkm9VFG}<x(~A<l$n>H_
zFEYI-(ThyyET9*eUX<uXrWYl8iRmSYUSfJlqL-LnlISI-mn3?L>6``h64Of(y~OmA
zL@zVFEYZtMFH7_?)5{XQ%=EHEFEgF9fL>;LS)!MjUY6(;rdK3-h3OTEUSWDgqF0z+
zk?0ksa~9AmOs`1v3ezhRy~^~eM6WWvD$%P<uS)bP)2kA_%5=^GdX?!_iC$%TRif9J
zUX$oGrq?8Tjp;RsUSoPqqSu(tSwOEby(ZCXOs`4wI@9YCz0UNyM6WZwF45~ut`mO7
z)|t*(K(8~sF45~uuS@g>(;E`K!SseiZ!o<f(Hp$>hP3tu(>V+14W>6FdV}c=iQZ&-
zQ=&JS-jwJ~rZ**elgUkK?M<e07SNkaZ%Xtg)0+~##q^d$Z!x_k(OXP!N%R)eTN1s+
zbj|{Li|H+i-eP(S=>KTA>(;~XYTn7W(2p38c)E{@M27FXHTt#3f6C>92M@lj`w&a_
zU)p>h{_Mjt-5cxO8}GV(J5R@S%Y47RO}(r5SGYFmQ0v>IM{Q`60kyGBM%1P@nNXYC
zWJYailLfW4O;*&lHrY_y+hj-WXp;lAvrSIat~R+)yW8YO?P-$-wYN=P)O{W5dj=n%
z=#US!ze9f1fer;w2Rjr*9qLdBb+|)e)R7KFP)9oyMIGx<40XIiany+pB~T|jlti8C
zPzrUrLuu5R4rNehJCsE|+NRM}d|;qWW2kd&%Ap>oHvjn_D=Jr1sgmbPUaAzhQjjV|
zt`wz8i7O?kQszoos#LgAkt$WLRHaIdD>bQ7=Sp3wG`P}`Dow65rAmt{EvR&#Gvg!g
z>_cYVL#%sX^54qOK3di-^PRT-zEj<!mDNz+D5yI1-2v~yzG>=v5VWt~2VUQh54^rQ
z`Oxc|CLenJ{eJNJhy396*U67w|1|m08|V*!H!u_cZ=g;A^aiFWfZkw#5WK;mAb5jy
z3ZgeSO+oaA`a|Fi4TZoPs#6HPp=k=CH{2fvZ+IvS-f*45=nYR(7`>7H2zVny5%5Op
z6hUufnj+|p_D8`R9g2cCTBj&_qtg^cZ>&EC-q=tKys<jP&>Nek7<%LVaqz~6;^2+f
zDUROwG{w=I=udz*F_ZvrqD~3)CZ;KY-ei9gyvd;?c$0NXqBl8BN%W@rQ{YVvrNEo2
zQwqJQX-c6t-Jb?;dMFLvbe+=ZO;1xAy_x<Dcr!y8@Mh|iL2qW7GU(0rXTh5t%7Qmr
zr!0E2)0ExyR(^Zc+fCnX9^q5>p@WAm?mkV^ICtr@)#btQ_wreJf_m@4v+@M#Q0pg1
zkJ>On2Gqt0GNLw3kO{SUg3PEb6J$Ycoggb}+XUHA+b775+A%>6)XoWVqIOM?3$=TK
z+^9Vh<U#G7ATR2^N$UFye*R6854C@i{HOzy6hIxEq#)|hB!y6iCn=0NGD#8C(MgJ;
zj!jYwb$pWIs1uWvK%JbVB<j>8rBJ6QDUCWaNg34HNy?%gouJWs@T@#RW2kczltVpE
z6MR<2ipmvLs^qznmnsFW6r@U#D@CbN;z~)Xl(|xtDiy9&q)L@5RjE?rN=>TNxl)%Z
z4X!k#N|P&1snX&~3o3UvUwF?4E_>&{ojs_uy(W0@L#qR&*XltTv<6T{tr3(-YXW7~
znn78#7Eo5L6_ibD17+9RK{>PzP)@BAluPRZ<<`1Ed9)r-Uac2YpSI7mdqLsT`at=$
zeoz5z08~&L1QpVTK!vqoP!VkeR8$)U71PE*#kFxz32g#YQkw*o(xyPAwP{cpZ3a|U
zn*}wh9W}vCPwg0}oHhq)oV4QvS|zQDI#1d>>H=vCsEed6qAroPgt|=HGU^IxE2yia
zt)i}xwuZV++B)h6Y58Hm-L*AI+eF<WZEIKGeg8Y0m%Q^=zY_<J9y@XD{oSXs4jetM
zIevNf3bh-?SE^^f%+S3g{o=0+BmSS<cZYJ|=!v~2&fh&$Zt&dj54&%rQ+rRH-%WUE
zc=4R~f9!#)`Al~B{KJR4k9B=(<KFl$_B`3g<>M!gJG$IA4r;AWzw6@v;qtK$9P@W~
zuKVqY;e9*bn&|EhAFwL*7V#{ok`A?AB|U0`N(R(Mm5iuO{EJ3yR>_RoqLKx*RV6EG
zn@Tp+c9ra?9V$6cJ5_R`cB$k-?N-T++M|*OwO1uC>OPhFig*@O$%opnk{@+Ir2y)n
zN<q{il|rb)Duq!;REnUEsuV>XQz?czu2LLzLZt-iq)JKDDV0*F(<-G=XH?3d&Z?9}
zJvuo0@$Ng3A8j8Qy$w{JMoVyp9BtC*r%+GP=mVg~2FD<4Y-9|w#_}|VSz}Ea!>q9>
z8pEvIU=Ff!BRR;*<tc|*xhCZ>D>p?s%o?Y`ae}mQ8W|_Z8|QQOIG=%cS3b^X#BsvR
zahl=^rm8fkVya3bDyFKOs!FLUr>at_%BiZ9ny0}$rsioRkEwZ1%}c3yPR&cHc}~qs
zsRbG=U}}Lz3Yc2p`GS;M;M9VYTHw@zlv<?0BBmB;q=>0SPAy8QMNTbBsYOmLN~t9p
zEMaPiMoO4k;?$CqTH@4_lv?7{l9XDe!7`?nX{3y)Wlk+isbx+rOQ~f}Ela5t8mwSy
zg+?lvTH(}+lv?4`ij-R6)QXf^rNJtuR%xV)sZ~y`N~u*&txBm?POVC*H5#m8YK=x}
zm|El1nv`1O)S8r9<J6jzTBpG}rq*esj;VD{txKtOPOVF+bxy5IsSO%zU}}R#8kpMP
z)P|JW;M9hc+ThfNl-i`hCZ;xNq=~6#kE65iq}|VfChMk@+T_%xl-i=f7N)jnq=l(1
zPHjo4Ely}jsVz=zK`MXO@uRV`Ul2Vxm^gXO*&p6^cJu%L9q#_n({u6ekC*)AM{)F;
Zm!Eksku-kzoWpOs_>v<%NB+|4`9I}AO*8-i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_weight.hpp.FB767EE60994A05E.idx b/.cache/clangd/index/device_grouped_conv_bwd_weight.hpp.FB767EE60994A05E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..831fc9e040f0edf39515f88276f2ea6a31912609
GIT binary patch
literal 1098
zcmWIYbaQiJVPJ4h@vO*AElC7283cj2xTGkvnvsD)87NROXK%o5ufq-^XO@3)ydQP_
z+V3E)y&WP~xVY@va*KSXsZ~$gYtr^;|8C#h*;m{4aW<+NKR+{P=7Rc7uP$XcYKj}y
zv@QE;^YG{Dl(x-pR^PRX{JKI@m1CN7ww7FE=e3X@N9F3-<COKjA2~BisxUNsmcikB
zck++MR7c5`K3-XQB{%lHU9(J!=8m?REKlFL2m0Qvyw#h0xA^>v2X_=+n44V9`>VN<
z?d`9XiMzA*T+Ows|H%;b`jD_i$#(s+ltqfwxk)`cx3FEicKGmz3;~~nlgbqV8LCH;
zMNS6_T7OnCJt!P~AV9V3?Z>7Vi2zYEQSX?BTTZW+@pANP9eFD{<=9lcrS31%s<;0-
zXuP0hf=IE#$~42vHJ&PB3!Ky<!m`$RK6$ht?K6*4`ILs90>M->HX+>;`N@+!yv^>}
z`&*<-oeR?Ze_Z#b3a_%b=^XLGh>*m+%mGr1f+o2H9#=?{@>zacnd_p!>Z|uYY%wpA
zFO&)HntIUukz7X+LvdwpQams&TJFr>P$;_Pq5u;sJCF+moO+si)(jjF&VmW6pRn(g
ze9Fee%*Mbe$u6k~CKx!Gfbqw`$jP|M)bw1L_MI95CKi}}85S9Fn0|($)U@K~K((>k
zEpuYjRFoNcIHhHzHDKCO`&d;Qgo>Sj!jjsO<}m9j4nKTS#d7g17Z0brvOHWh52u2%
z0uoPASrLh+q^yL*Q&v_+;;AUBAn{a{RgrjV%4$eFb!BxVo`$jpoCgakMow@L!Q9&I
z#oY3?|Gxz{(5;&C1~8ui-Kwd8#8cE%MB*uFDk1TdHI<QgDw--tJXKv)6NH&+x@t%~
zbzOBNo`$Xl5>H1)M;WFM=2S*5aO`=ilyy(s<y;4o26_k-XE0T3jjaNX&P|vOj1xT>
zJ$0D3pBk`!Vdod`W&*~bptLkhGf<JBo-|AmFurpVm4HbD7O@O4Q(-hi34`8TrR(<>
j6&EEJD}z!QG^v4@Tp$7rm^isRr^!#PJl@5`#lZjooUM#H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp.2CEBA1D92F4EDC1A.idx b/.cache/clangd/index/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp.2CEBA1D92F4EDC1A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b627f54ffe42a121497784b7c5b71b74e4e4c791
GIT binary patch
literal 19286
zcmeI3XH*nP*MRBj>UJF3c3@^mGYB#!P!K^eAxbM^SaWuDbq$zScVAprS6xMsY(PcC
zgo+BRAQ^K;KnVswf}}waa|Cnx3TBF~J@1eA_-p@soMYEK-rH5xRn=9u?yVj>bnxJj
z!wn1u&KbI5@!a2kh6V-(M(E$t-@O;CVhjv&O$-cH?9H6?<FDl|in>v+GTw%dJiep*
zv0ElDkG$lvM~-fhf82C^`20Od4shu05g}jRSq7gDJ-%Z_+#brkU1a^fj6+L53^}y?
z!s-_dlZ^`!C#TnGOuF_OvvJVc8EI=iR6YCKbEMq3u(;i+wp$iNd!L$-Yt<=Lu`~Uy
z%ee5nId8k$1XpA@m7M)!b?>8#viBTXIWF#_yEeA!&X0d>od1J7efG90Z#!*aK|9Z7
zcdnn$Ke<^x$T^{06@6`I_nfr(1J*y=94K>5?78sL0t4H*)!r+H--`rhYu8NJ;U318
zGoQTN5|>V~^;$l@cG#hlN1r^j`)N|}g7VwbN_KDFdiFnO`{#N6rCIx=S7L4DADg}n
z>fJuNVtQQ6FUhw@e{tQo%H8->)ZI0jPAvn*bWkdLcApZr=}MtXWQ|Qy!rE;MX88O*
z;#ZF$e#vKkN%=9*U~0gy1?!$D+RW@<IN%6iuUOCMsWGL30)4ue%Y82_xwz%<os3B{
zK1ZMUp)mGh^@bEHf2#|2?p}izzxkj&Jnj1M8*4tbEA*ZAabJfsno|vDmeKw94X^%n
zaGTsCYs-9oy?$ivT<_kcH||%&K3T|{?)fD>c>b+??cXvp?HlU1HR}0B<AZ*mRoVGQ
zmq~XQM!h<_qF4Rn3;jx~MwC{SloVUtJfms1_EsP5^5NshoM|67xw4!6hmsy{HtF*#
zwoT=ppX@jI_;ZsR))taaQOg{To}X_S`n&zn;3*+>0a3fR*-y!S6ZNQg+1+8mUg_3H
zwEb^BTQJzAclW6$(?$oB{E#qe=G)z8TK*XZ9j3M#ztC?@pP)#uqR@jefrSfW<W3Fh
zDK2gI=iF@TUi_2A->+oe1A-@l>q0=lkCB5n6bH_E6_NF-VdPEagiBrruUcF_J9ON6
ziA(o98@tE1Jvx`ZUAcT?VPN|PXXV&J-}K{dqct^)&Bx?Te9`6f;&W@OUyq3D-|)J=
zhJM`IfA*!zMmyGR@VMq1JL7`oA;+Sw$8wTA{z|^yoxeSj^&HXS_}|axg~deoJv(po
zw9B`Wyd7O19|(T`p-<UlSI@c0k6Zc7d?nlS;N*&tArI7tjJ{k3eM}Z)Er>{cm0&&V
zvn(6#<PEj<o|+F`qjLKmppPXT1|~U2{eOBaGaZ&RXH(kbdG0?R7^)i4XMTZ)+-iiw
zty<{_t!vr(EkV=f?2Qp_B^}Kkw8o#Nqqg)8dCgTBQJcV*&CwG&Ij0;ce!KoZl_$TX
zoLw<`qG7?2al7v2+CR;lo@4xMYHm(a{JXJ#wb*~;W=t1%IBwkBz2010j}1Np_g#Fp
z?cyBmVpaBv@=Nm{S?~1R8suIV^!)As_fa(w2Clztq;q?*nn^SBW<TszAFkLk%%>sL
zcK490W2U-zC7RYvzxkx!A-D9n>knh&ubWRdVQ#lMnC|4!ZIaeIeVfAj?B(co2R4kE
znk)(4+`hZ>95c=2!pH}I?&%&~HF#U-khhul9`OU(wY_y{PC?<4al7Q^Pq*K^I(w*W
zZreBO?I#cOt6`Ph-aOo;?PhBGZE?njQ<bV7n}>wMJ<f|qkGJmP)oNi<*I}84OZ<9v
zU+>lH#pVgC{#+EP^mx7g<8wp%S<jEJnAhirJ}1geCDHTurY~5&_uXUfNm=!PN{aBj
zF=*ZLmCwO~y64|E4!z;KdYyrEdf2yuPrE)X$hq<D(II7G%`ch%F|p0M<neyo>z~F-
zeP)^Yw5lk17wvFAN3QzS5+*7tihO>UvFvo@ka73tjx9RB<{7WOL0i%Z>qFLlrX7rX
zsiEPTi@x!*ukNav*84x>D`Jg?YXg1m5BquK<zd#Z`4jB}!i~2)I8<#Fy~6d&lN<g0
zy?PXRe{o8v{jf0b`2B10jAL}_VYkvvaOJ`o(AO)-;J5Cw%^|rh&)UbGxO>;<D|NAY
zL5;!EKNruM*Ah*}uJJ7oPYO#>f>AvUgy|O3yER=5BuIq*z*+N*U-m^*^Pqb#10fxn
zq{9yS1gp?)zjk_7KY6eWbVfPkvu*QPXMKWITDBhj<bCAli9ArCT&_TqE6`G(U=_oa
z{vJbXB64`(h*Zu)tMjmnKEWz{<j<W-*RFTsffA`?K$8rVi!NzadFcM?)yzKG#-Kf_
z%L&Hf1k*#GU=`(|AJ-R^OnIaP9gxZr%OxF}F3sWw$;KpbahVcWBC(f+m%8bK#b)E}
zA4v*ZXqBKX5-+q^sBXGovBj-U6UJM2duRwuk$9AJlyTDqi`BQ%_f@hlHO62Z67Qqn
zKFUO&V3oFGV-nX-^SomW?2yV4R(6C{=@YEd@!gG4^Xm5VhQJJ|T%n~`Xcv8gRXRIQ
z>h@q(!G1#^Ln=YEG>A6UC(SC$nZGw|e)G~8j6o{vD7cPN>JzNuP}NFU*MDTOAuvQL
zfwVM`R_PP0(yd~^>+MOee0VSs^+_@`Plh)71gmsk^u$2<={Hmth;o@KovLWMV6n^c
zokL~TUf)cB1m$p(mfWOe`UI=E=9LGiZpcn4fep%Gk;Ni)(*=v&vf|%Ma@J(<K#p>_
z3{5VhaiL4Fiu<U0P9K8?Jmo<Tq;ec~KMrT<6RaZixD__!eaLhfutkN$vM`o4)hAeG
zi01m_FTFxq@L)K~B?mg@zybOMtBmxbLCJ4VepP~zD3_lte(u(E!Q!zydd!Bd-D;Jf
z4HD0@n5S;KVDW_R!%xj}thD6852%1#=$H%L^$Aw_vD3*N0Ty2yWWc_+fsn-TNsN&`
z!79^-#wtUWougPFL4{nW4X@L3eS%eHcXc{vJokw^4<@04QlLBqs`Uw0nP=-8)A}#D
z#t^7cE}LlSCfY=wG^?D~6noBgO;Q4Lq_WszafhZ07SFFKNLU?MgbEpf3dw@DS+Ku8
z!72+1Tz`(vzhuM%2c&WiTAzdV`UI;iKG=aROiyNwK}S?b93_jR`s)*{@@vFrui`hp
z*2cgBsk{WTm!OS4!7AR1r={%Fz6>=498$@prMa}3KEWzW9bbk?G#QLB7>iWAp|?rX
z1&e<-aM)*4QXRzt6Bh$vV2goWn=S?|0fpvCN}yIaWFA}^w7<>}v_$uhm5!A+UGV*@
z2IOD5SU*l}3_2mP4-0)*nLfcPYwnv{o^~7;ZwPp#GD$kgs_BBo>pI@J-7~cCvmuZo
z@kHrFtENk{*y-!!r`1^_3;~VAezepNb(k(`R(btrPRsJ~3uHirR0>#G0c)vGu*&+B
z-@p20ui7pH?U2exrsYRQtxvFuU-*tti%mB5N??Ul=3C5HH(jteU_s*Wu|xKI83G9!
zp#_jHfF}9`t889eZzx~$hbs@9QMX@!))!E>>k_Q8B_nM-^+)Aq9`r#fXQ1sF=%G)r
z%FbIKCQ2nvfCuA|%2jB46}^tS1gq@-T3ugN8QVq%x}drg!B$1kSf5~(!)xyQWp6+G
znFmfNmy^)+B(&8hSS8l}Q0uf_l>t03Mk<-mBoj?Ax};fU$n{4H>P^@4U;<J}g62tR
z9@HgRB}MC+_Wt$zAw1}cR8pZuD(s<8u*!+HO7(l^vRN$9pq9M}4R50A>XK%a#TNT^
zS}uo%pfys-rlr|*4}F4FPEGE<%VqiBXqp;^3d)AI*{EG~30AqZO7U%;*S#`B&<f=;
zPC8EBbiv}R)nNah7iv(O^hG&bgO1nWAbo;Wa$eNmvAOfaO$n?~4ht+6sGBZWTy}Xs
z*qb4(;z2i*LmIS5Ll3J<u*%(tFa2h^kDP7XG}6v7vU7|^&?Q*qzQ?y8g!Wgwc`zE~
zk_S!lU>kjcRi3`dUi>L*RwqMXgz6DOOG9W2eS%e<X`W^8&-+7Z2pFU?Mmk2`biv~1
zc~vWSzYN=A2wEU<04)tbtDG*uDzEE5oB#T7<|RYGB9#f!306%PEPh{Ff4w2_`ZwL2
zTnqVHXe8(otYY9j*K6tEzG$Y<+gm`|WM|=7g%b)G(*ni;HwxGay2aPr5ok_G0Ez_A
z2~UHZkO&lsBqsqylE~O#V>zLY=IiJ-czxxBaL9#2waB}mY!|c^IR<u)fqg}ehi37x
z6K1L2?8B=IHv1^4-q^#BLVK_3+dHGKKg9``nQ@nysrZ@LGd-+bJrd6GobZHlc|wg5
zpXoI<_%-#j$RDX5AF0tIuLB<Iz)X?10GBO*xIY{W4hO{jk-#OAlph7XkAm^y{ii|C
z(_pN~mq3q8fGF=O@VH8rcMHgFfzG1-9U!|y@;xBCCo(p7At$_``4_Yjw;?A4Kt%vl
ziM$`m_Cr(5QoT`(+ZY?g0JRv+KcSm%l<G}mtk}viaQDI;-O5{Xf|r$7d!^op<%D@w
z^V(yhgA?Xk&2NuqZ%$ZXwV*wIjX7bV)x!4Jy5)pLR*TwWvx^fJ+b-@pPrul6!pgQQ
zJL1KX6An|BhbaxVY+KE<=(!+HeUXAfCCgQ^{5t)TgN=7Ipr6zHbFxkRp~7Em6RF;$
z$B)MAO?ouWboM+RXRXbO1joXvMKlyX(%nANPPi4J@P&5#LU$9{2X^y;&2pm&1@nJ9
z%q5`#I;Z!X{@7$Qib>3w`f%kNRKv}b+h%HPXT62Y2@ff^hZHefN*T9OW-K1pct?R=
z>Bj8@t{*2v(UK_I2pdwkQ&?%!jzUxmpgjU;Pdp8ALKt)m6Kl%}C2(K~^uS9QCse?J
z6(nonI4zuj<tEM<uWrijN1&(M3vKs87u*cEb5TP!?ptvAW+lmz&oL*QBv4q=eMt}e
zFi_A^{k7BpJmwpDgq9j9at+m~hO!lTKeX8oHMle=yoICQ!qIqcghDD~n#!1A`xOeG
z8O>*=9kw^{&WV;b?U-=wIKdCf{Gb`m7R{5;ya2YvMuvLefT1ocR37MQBcW}i__TOH
zqE6B6v&0<lQ=0o!4_t98eqHY%NnQ=g;U!ePgq?7Wt^EfMwK;Zw2@?F&>nArnapMgd
zYDV4e3}@W-wD<g{KCb>~=)Z&}FQGZk$fkeVsn@-3O8?~+#@j~J!n!>pt~e(IGm2nl
zfa$>8r_udig>k|G#^C@n71tkcCK;)I3n|{Xi<a!7m3V}6!Zn(^Mw_3|+bEonM@#Z(
zB{se}w1h~~p%SlA_yt3Kq}zXr8P`&pTB<wlBl|g@`&xp(F1J9l3e&xkalx-%BTudr
zml!BiG4@qVS5aQYjI3fti~N9beZaVh{E%^dNb+-L*mLFwEO!_+$I~fpc?jy$e8w!F
zv1k_Px5iMa-w=yc-$_e$(&l2Dq(Mm<G{*)JCj>EyAf`QT?5@r|y8Sh_REwtgGf;8{
zcEE-&Cv0LAn;0jo=y=b>CDX8CHX4rKV7qT{G#213ILhHJ&EKVo)kL=+$1|zZklJ9@
ze2NR2Fds1H55$Lo!b8UVA<54f^XDSt9XuNHjoWwps6D3iFm?7E`x(`!4qDd1j(EV}
zmlf&j_Vc&~F2UCg9E$(^51L{>(ncT2Mc@l<^o3*}XygN%7tY?!{$r~$;@-&6>OrZV
zQ!aR9aKal(^@ijxl<Et~8-Qv9AoRBZ)i#oM0o5*&4*=Bxk`DvbVUmvl#WB!TEI%12
zl1aV-R98s-T%gJ&xe!<tf<EH?w}I7dk}H5!1<6&ws!HTOZC>RpVt;;R425E7QVh-T
zv=4=HXi`qH7Mf^DE@8P6aXp4YDa)0TT*jK0u`0Y?L*Wi<euw08*1VkL3f7{6Rg3zS
z?2nb~bdjrA<0@7may9E%%{qy!Wu3Kb50PtF#~M<8mvz2N%I~qZ_t?&&el2TTOL85{
z)rrr8jTbZq8y!Y?%)0&5CD!(K+3Kc6Ub&fa#dA>OBCmW%xr!ZA$|y@2SNy_bClh+h
zp3(d>ai~i5t|mOj`%~H1_FcT{g5HX&wAEF5Jf01)4+<6ifaV|2uJ|?KgriV)6uM!S
z>OE69N9>tGkD_x+;fEQtz2V^)Z+R6eYF4jVZrG3+bS?bx_>USFRMl9feJtaG-%C!2
zXWGY$>{+<C@1Ch1O9ybmsxGTsM4N7Pm(>J?QjVciAF-jfQjS|m4yR<{l!YkYNy&DS
z98GnJrd;rX%n316ml%?tQL<;0h1l>fDA@~=UsAG{BsWm9hJUhC1KDwJDcM_6{*IEp
zBl$BW`%LmzO7@lHZ<OpC$-Y453yAvp0w-TU3=}^g^CRW{K;}<!Adm+FOY!-(0@+ql
zz8%Q6lN<(QVI)TYS;RltDS|A2H<0Zn<$Hi^56K6B>;TC}fb0m#F+dhWay*d5lbir-
z6M(Z={{-Ng00xVk2y7Ecc_MI4B;`rKHi?uc0oNo_o&sbkWO->omPYbfAUpd{c04Q2
zh@5Z^IG!W<Jdm9y^)CR~1(Gv>EQ91sAj>2<3&^ra&H=KVf3i~!+5UMzmPg9-fh?co
zLZB!FZN&B}28v>mOMs$;<O-muAh{YSs!7%Yg_h)6(504KVd_AaI+7oNE)Ph41musz
z@ra$;XtsDm^KV3lfmH9>#?#P<Q9VtsdPIytvrQzeiKIJV69>;b$c$@rc;nfr(WT*3
zzzoGRDBgw|N}GJ%qFNyB5lDODU_+zPXr4!_M58f}?wCj0io63l?jZL^yP(Z3*bdL0
zoUjWz?jq%-a9}C9U#f%yD@m?_y=q_|ypVFjdpP(#9D*m!aV<U0EuJgej2^pyZCSv!
z#w!qZ)}w}dOY?7OD_jGq-er#)ZsN~7ro<mC?}Ku9#L6GBCOC)5^6YLi6U^EpKf$Wj
zt2&GGZzv@XrMikU>JCc2gXCzcbu`r(&#03>MRcn><J<kaH^O)?h87lGAOZUyam`yu
z^>G8Sv;Jw&HEzc=1+X*Tj-i*4u{*~&;XdckF0bcjra#{0al$sX^)}WNJHt8QG;4jD
z?Sz-O>ARmk^OU`bLQSQmjI~sI{MJlA(fgNa9olJ7Q(dPmuhX`;iX5uCSrXF@KMS%f
z8O=wgt5{+TtB7H(aS1c;2F_!vtsB=&cMtdR_`z^M1a;;;n!iUH(YojeE)Ng-R#HJ1
z6TvxfLMCmLNn7G=R^#$-bd|Qm%Rh(04efGiM{yiSz_t<49*-JMI09QA5&a;X5C;`;
zq<`)Sls$oJJTB3LLdTELRb(GV;lo(tZ2>3vGM2uKt;ic0$Bm4u$PG;E21bRu3j24Y
z(xxDe_*hF^mI&A@(AX_bTFMP~3n$c2b~U8A-9ULXke)qX;NeSp;v#@Y1j#wTBZuVM
zpzm$qDdt}V`c{dILrQ2EyrTJ6H1XDdrq!QC9}wDRK}8U>!>udT2cB>@;=mK?@5eO%
znDngaqEJ{7*97gFb+IS0Vcvq)Z-E;QJ~etOEd1b5Tq6E|G%M;NQsRw%w2dF_xK1Ch
z!2<#*>f%zO;z=ku2~BW>!aKoq^5jb{^=RlGfXV}~dBVXH2g;!_Y$e`*kJ8)|JstS}
zky>N;YG}JZLbjeBxP}u(DM$6g{+H#eWC>aR+Y3=4`=ISU(V>FFTWH7I7~8^`aY869
z2^GCu9E$$($<P$<N;zRYqgc;a;$gQUqo#Hxyp`Xb6PDX7C;W*1R9k<lgE;3!Q8rOj
zM?A+Nw>KqEq|9)4px2&~Cy|^?$&*P=q2wtfr&97%lG7-88p-LDJe}m@l>9i!Cn)&|
zl220dlO)#zc|9=0z0L^@K;A&|8z6r}av`lQq}$@r!U;vRx`^c4wE8y5#k9JZ<Put4
zLUJjsE+x5)R+o`{hgRPqxtvy)lUzZoD@e|Qt+Jpoe&IM2Kx&l@jd5%V&2F&ORcMT3
zaLDxo<=3DYjvR5wc#!8nGi+U;*$>Kdp&4E@(d@*kec86+La>2VZy?!^Rr`_b&#L`N
z4q(**BnPtUK$3%4br8wHtU8$F5LO*R@<vv@k>og56DKZAXzRpk;z>TnYL1bdz-kgm
zPGmKSBqy<&B$AU^O)|+TtR{uzR92HpavG~i6B)M~TDo<iWc)s^nN!lY=)neW^!^@#
z?GM4Oc!c5LGJ5+PBg<GHz2>xJA8m=(H5_C{YK;+QQEe|R*-NPDCoeP`=t9h*VgT(D
zKntQfJCJq>qy^Eu3!+_uXaPGFu`PriL>G_78Q}#S33O3tEZA^p?aC3cnb{KLmzDfv
zg#EIdFheqfpqY}H1kIAnGQxH$C(M@2CTNai4ndb0)n%p~UKJZ-Sgw~D56n1vf_$Nk
zfp1(VwA|1hL9{1+s>WcOV>oGqmcfB#q|sRg2UdxU?I+Znx_~*h#L&C|&5K}L(K@Y#
zX0^~-T=i3#&Z&$uHh+T3qm~@)e{jTid)Pu+Ts#<f+1m%LtlMeHcDfZF5S$P}OCm_l
zq$g(5KZ<rvD5D8wI^hXEWEj2S*_Lm;yg6Zsa*0~B9K0>P+u@ap6P8*oCFpm{-w9e~
zxs0GcEdL;Ad8g%F#a#a6|5S>!j#96qG~yDno>H$T*_TrLlDvUZZy?!^Qu~qYPj&RC
z9K`$rDL#-=iX24oK_mxLd@#u&6dywJKFV?*)lOW}A}PyAlJ`@V`$;}PSsozyAZ2-w
z<S43B6xCJCFP`G#DW$kX9;5hUBqvaO0?CP#c_L*k>L*d=NhBvz=E)?dQ06Hlr&8vr
zB&Sj4X(Xpp=IJCKr_7I&e1bARLGnq;{3OX|DgG>3-*Xgyj^y(cf1c!v6n~NAOB8>J
z<jWL)ndB=Je}&{siq9lDi{i6LzDn^|Nxnw$*GSHx_#Be+C_azme2UK}`8vg4C;2AD
z-z51K#or>ifZ_{CE~5A%l5bP|ZIX*AzL?}viZ3O(jN;2kzC-bMNUo;%YLaUxzJ}!2
zfPW2?qRCSa%<F-*$PK`}f#f&9{0+&4v`ryxix;Ry6Ue5B<lD5(ZIX*=n_`knXqys}
zOKF=@lFMkDGLrAmHg`xar)|nfuApryNY>IuTABzwexOZ0(3WC-qo8#Z>?rbO$X|v^
z>?!6@Ae+yCN*wv-P;j5W0+sl50NOA@J`*bO=^_qoQRG?Bf;740+0cSCx#U-&1!;1b
zUxU`9$z`4ctx1#1JQrGvCRZNh^PmzRszLr9X#7SDQ#P7pHomMaHZ3?|18cK^WIxu%
zk7R$=#-HQ>)+T`DK-MOZ<RI22h~!|_CYa<9)+U7HjjYW^l0#Y3P_~U&zc|)5j&;O4
z`$iMZHlE~TtnD$96Ij~>k`r0mM3R$O+a!{cS=(fiQ&`&+l2cjRRFczJ+cc6-vTaVX
z?Q#7=9uF#>@8$jZyTflAU!S`4VxZOi{|Z8Gy0>?f{n{yW8+{oQ-6vyX1fQ-$d)ucp
z|CHR@>dx5Vy)6zkGRQGS^N;A}-E!m60dB8&_NQ?#$n7IJ5^|9uOZ6x7aK@qjxA;<*
zu(SWQQw=Se0q-5Lyp04K8MlqhIBYd+4cQbmdQ}+x-FC~48oP@F&L4h^ax7xair9A8
zpxp6p=JfTg)whvBQw!~DVQ<l7s)fUA;Utlh8M9>4gOtp;Br^gwslu-fi7E77(*-r>
zX;yKXwZa|B2{%~94b}>ugu*W-DzTpC>*@C5%&;D+*F$HKL!crAcE&8#pIXGtw?`Hk
zw_$qad-P0|Z0kz47k;L_`#RrUeR)~W@AmZV`;h2YY8qmRZ02W_>KR3N^FLFn&m?aJ
zs;#8AJ_D#SNG<|aMWCl>G*kksN|I|q&ss1_l-~zE?~5EMRJ*nrGrW}?qGK&4CI5&|
zLeCO76dSrIJ_$`qNqH&sEG6Y-(4>r%mqE|6f67Oe!O>!V70{%D)USY^6{LP8G^r%z
zmC&=2lvhEMDpFnrJ*)mHA6Z3~r-j|L&{-^B3nyve6p?G7Ne!7_4Ro!6Zlru*jkrU=
z|Bv=%x?`;34&fn`J%ma;oDYP?)wVcgejg2<L9GUL#FN6|f#v8>{Kq~2a};B56B@-2
zY5pNi91x6vTm&Qz6vo2Nv9O;wigkx)@hFb%GG_JD1L2MCs*6nPi%eU*hr$aUD&Zx~
zzZ65MoDc{Vfusv`KeX5no1G+5{h{0D67+{|(KFT1d=2Te)g8UXihoZkaFd=`WRFU0
zpi~W1H*t9S0+lZyI0C35NX`MO9FgN!oMA>Ujv71!o$RAjFDQ+e(F@A%1?432OG@>U
zl)t3xUXt>+l<F-he@ofDCFSoZ)jLxDj<SCz`tms8E2a8M%D+-}U&;HwQL1mG{2OKW
zjg<QVl^-DL?+5Js08xK`pz<f>{=nWJG@qJ6yF{ScPRh3fyY1xtVL%l|%EN$N7%ATk
zRJ%#}ZeX{Yl<xtmJ*0dOu-`+L9}N`IKrOa+G|)tgM=&`d8aPFhQNd{79u0<)`lF-4
zWRYWmB9_cA7HDG0{9=JqESX;{aE~SPiv^=&$^24*B9+W96=+h){8E8aDw$s@a8D)k
zO9i7-$^6oQDvfN9bf8Kn`8ZG=C;0?Wogn!nP@N?CJW!qgCp(`9eZ}^;08|%9`2}Ei
zf$X15pvolWnZPcSlxG1|7AemHc3Gr652*4;c^<IKBjx!(l~2m^fqg#N|HZ(n7<3ft
zTMQhE$tY(r=vNFpM0p9YDk1etfI|tXUjq7-kowiYs+!cV1`gGvel_S<P3mibm6p`k
z0tYRruLb?Iq<$UfSqJ)y^{WGe>p=7IN_0j9^n5_dAArFRMENo2KBF`D&D@EmzMD{e
z6Lu7rgM3zzFB&BYCEI@gwgn7Avv4MDl}S&)ew#$^fQ7sIjc-7Ljf`R=(*c`e$<s=m
zqBgcRMy6OXG%AKly!3GBh>=k_$y#Wn6&ZivfM(fRny)2yQelt_BX>$iVCN%nu()#Q
zzJ|clby}xO$BV9?Sc<CV%XIc-df}?2KUtY>8vOd`zkHJNxJydJs#{NoAcMS!l^2Nt
zUEB%Cji>u+1Fya)xW{%p!aCr?qm8Dc^$oToHXR$;_6F;S83zqf)JgY=hnUfEC^-(D
z@nIrP2w@Z<On=-$c(aI(|G%dB*Q5>ljaGl72^%yRDuP8Dv@xdC;efcBbHYK`?x6TN
z0IEpSha|X<PV<tRS(9r={J$Y5e4inA_Ui2^$NJXa|Ch%w{sIUM)yB_%#PJnIOTx%-
zUJ8_?KqDL<M@|7ou|W*yWyFqN(rU0|!~ZlZue|k|km!B!0D7kfxDS|y%frqKbbkZQ
zH_$fXoS^&c2#>eyf?K({j~gz2_tq!-(Qco84_$6?P$*-$GVz39;|WLZ4$0*#S59&T
z%T<tE#d1|7SF>C-$y%1vl3c@bH6-6<xw|CaW4U`I*RotK$@f|AzR37f4HOV*{M-t+
zV6OMyV?Nzmy!8KH#P56&b6?-<x+mk&AT*2r)BAB<vnp`C+ql+f0W5<P%ivUT0jz=(
zt4OYa!)xFuEWhE}VfENi>hOR0)xKF)exoD5a20yha%rbrdN@8|aI1LPq|Vk^Yte%R
z&}{<fE@Y6dO(5Nc4AQj;qPyU5R^jlL-Vv^|YSCLdqxX#dV#sZF@7c}Y&)j~Wd(Y@e
zf(y$1AW$6y_M&YQ1yoTa9|EdFBA5D=n6b+&ivP>MRobw3?EHw}fbV|i@nL&piQK38
z`=lkZ19Ce^OC%b0jwUS;-8YJO)~V@rW8k}ZGwNUN(VF*5j(r+1!2<Q@%-%Db+aGVJ
z(Tt$`ZW7N4wPr`Ugx0loK>f9yaoEl{Vdnt)xQKBG6S?-%#nL?oOf@LuO|<zY+8Iav
z?mIFKEn1s~pd4b@mNBecbb95pE%RA9wp6e;87aP_`FFI5IDzZF*ThZvpv$Ra&X#eF
zN4vH%rdt_%TqpbmC-Q9QKHtPofewq&l5MmZt`q)_6D?VdA90FmTWHA^a>+`9d=fN`
z)q4jxVI8AbCoWld(4%a0UwMidh0~I7^3#n}C`pCp*sewH2}Thpe!B5`@2Lx2liIfX
zmj?&_I2E;D<Cm%8)9j%od&G!+z3kCb-=I;Arv$(6>yNj8Wvwy!bgA`s-}vrRV!Y|&
z#hcL3I15x~$&Vz?0o6H?@kl`9SohT|wxd6NioVdbBD`^r5=l!U>9%+%c+dUi=RxRW
z+ClM#cNH(zPNJ~07M~L|z{3u^!whh{;GL5JE@Y|q?4>a%F9Wy6i|(&6Z~dPu{$XI?
z-ux2?)!|d7Ejn$vwUy_Pp@ZG}4{{&iG0?T2!H{MZgTdm@0pd>&@n=8r=TPzIK=G%i
K__M$G^Zx*`hAF-P

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_weight_wmma_instance.hpp.D8A887BE715EE7ED.idx b/.cache/clangd/index/device_grouped_conv_bwd_weight_wmma_instance.hpp.D8A887BE715EE7ED.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e926f551d0487bf57aa124e7cfe6979a506b1a21
GIT binary patch
literal 10682
zcmYk?4Uil4ea3P9e~Zs+-+|xoU21UcIzxDm&`b4(-l}(FOn7l4DUx?3tro4`PX?wY
zSaOri3z47!*7byd6J7@DCYdtv&`bk`^3u|QHYLG8htkuTp(QDjB_jcyNs3hYGV(pl
zc;}wuYkfSrTj@^Pn|FsoCob3N{Mp?H%ku5%^K?4hrpce%Za<Je!F0OE*XVQyAAU~T
zUwrqP3%~WrEr;&^7vnEZf4TX&OZGnFO&{^iJbU4#gW6SB?eRVE{;f}p>qocfZ@BJ7
zx%lFaZ-;;V{##c*bm$@H6W1TPXZNx9KEL7BV~3yr!uHdzes=l|-+hn#;PnSyJ!zj!
zw6<MweWmfj+iyPl^1@o@MW^&@ojXoFRjPda%5g{J@!ZbWW|E)I{`WKdp4MZ>E;;N^
z?#T#&o8J8Czs~>tjR%iD+`Q<%cV><KR}*g~wa@+H`}Mc4{@}(RFU~F8yyo`*k)D~m
z;nwfDy`MgK)f;<G{GIL6Kl|wUT^mk)_WDZ?es%wSmuxB-pZj6+ME$%s+qTM8&r}Yy
zjQ<(=+R6UO`!l<r?mWFgde-u*$PVv6-_4(VMRM(YcIMF1kAC*OkA8OW<>=m59{S>g
z)<0e2{F44o&kNuE(S1?%>&Yv!N8Wg8Uu1T}9~}Joqi>skV^eNS=v`Z%Jn_oQ*Zq<H
zSNVPW3Xgs`_t%%b+qoru`ml0)VdwsLzu&V2{^>=-^Z$`}yZYLW2R~(*th{ym`%j+<
z9{crrkN0F_?;VHVx$EEj-`x7)FaP1r&%6Hc=8=5jfh*TmuHE;`$8Wjhu;<hbZ(NzX
z=EJ*x8#;Ez^U29w`rGa-XRe!k<<|f9k+!<$k8hc!*$Z?F+h(?HS~=*}O}@H19a~_V
zp8C$VJO1XwmuG41B-l8!aox&cHSqlIqtnMfxbvx5nwtdkGxKX#4!U*R<mS)b!r{lq
z_eMWT?wF;Gli<>sORXyh-MaJW_rW*g<KF&UB5}TM%_O*Z{^BcE4lBfl^C!1(&$qsN
z(VJ&1SIpA-Nw9flGrw|J{Wm=DUu%z_|G$5`?|j|NB-k;(!@6=<y@j7Ydc5v9x_s#@
zt(yc}X11(bIjr8onJ4nc?tOCC9kaA)65KR%lV#<w`j7l<<Cjk?RsZpWS>h+brkPFa
zRt~GTa6rEGHp>@uI_|(-*V<qHz{BPiw#;qWGL@%y^r7`DcksP0{qB*Uym#%+xdnYq
ze`tCaO;$tyopbltWVV?8C~QN`0I{KAz}V1Y28@k0BgDpr5o2SE88J51Oc0wICX7uj
zX2RH9Gec}{m@ziDm>FYB%>uEdVZqqaVit_8H7mr{h81ILi&-(Y)oc*k8a9k=EoQ^m
zUb91NZ`d)mx0oGcN6i7Tqv62V(P9paoi!)K&W00XXNx&8cGX-EyBaQxT`lIq*j;l&
z>~6R*cDI-tV^7Tkv8Um|*wbPjjJ-84#NLJ%V{eOjG4|Db5c?WFjD0QU!`NT*L+o$(
zG4{8ZALBqR0CAuZz&Oxi0gQvSAjH8&5aVEr1u+iQLJ)@<A&f&U7Q#4O3qu@kgfR}c
zSQz6-Edp_*5y3dpViAnvi}Aark1cZx@p}Aj2D8cHe+fIWn1DD@PhgyAvING-#U#YZ
zdJ^MglO-`076pifx`46JWCF(Gq6o2A7cmx_OvE^~n1VP}Php&DvJ}SY#Wcj}dK%+&
zlch1vEM_3i)H4`onk<8{v?xI=)g_FjCX+DEE@mOl*0UIAn=Fg5yeLB~*JX_5CX+GF
zE#@H3)pHo<nk<L0vZz3;)D?`CCQ~rZFXkc6*Yg<Xn=FrUVX**lp<cka&}0RSON%9l
zOZ5`Qr6wz3TwW|gT&|ZfE;m^j<H}+M;!3@Oaiz&B7*`jo5LfF}jH^vnU5z*W)l;v1
z<u_mbWbzcdbI(q3`s_dbQN_Q0<wq2FijC<L@DwZRGq7a6vbqd01C$2Q0Hr}TKxvE_
zp)`s{D2=iaN>j`PrAahFX_8G)nqy`t&7v7fvuuXa60<;Q5iL+!WDAtmm=#K^Xob=$
zTcNbYY*5-n8<aNL2Bke_hte+Ep|s0(C>=2eln&7Wr9*Z=>5Ms{bc#+Wow5^3SIh;a
zOLRf$l3h@`V{Rzjq8m!L?1s`4^FZklJy3dN50u`R7fP?_h0-f~q4dRkQ2In4ls?%9
zr9b9}(l7d<^vixI1F-;<0Wkn&Kn_3|j0K?#ia{uYauCW;ECgjp3_%%^Lr{ieVJO35
z7|O65hB6Y1Kp7DuP)6j)xza6Mb^P+T-Z?~b3)^>X-#z_JU~*pX+PQmrUQd6+c!buk
z{AvLw_JHY!@Z+&wCH*&O^>o&2q(6dtiS%vUJ<|7ak4Qhp-R1PBxz*b<4445ugGvVU
z3>q2GGc1t-JwuNS=ov<2K+n+S4ComL%!r;*B_n!9jg06Sm&k~ou}4Pqj3Y9lXY6uD
z^h^V0LeHd<2|bfWCiF~8WJ1r>BNKY25t-04bvYAy<^eOKXI9CKo>?O^dgdiEqi61s
z89noe%;=fBoEbgKfLYM9sANIUqLBqX%Mw}8v-HS<o@GQ9^ekP@f}VB2tms))vZ80z
z$cmnIiLB^Zdt^n=IwC82)-Gp7&o*E-^lU2G(6eb|L(jHEHuP*gvY}@ikqtdtm$RW~
zA22(5c9rbt*)_7GXI~;adiEaK(X)@pj-I{C+0k<hm;*hBN)GfK8adE&ERh2}M~@un
zIY#6_&(Y-^=sAbX`7u1cI8}0@=hVoFp0iEPWqAB>_Q;8zb3{(`oL$a|o@>Zl=($vK
zq36=bg`TTTF7#YIa-ruMkqbRnmvf=#9x^w2Zk62Vxixa5=Wdf5J$H}X=($JaM$g^l
z+~|3R%!8gsB@cQYjXdai+T=mc(<2Xho)LM_^K>~6dfp-PqUTk~i=I~_FM8fKdC~Ls
z$cvtLL|*i~UCxW1Z^(S;`Bd_u=hMiCp07<l^n5+?q30Ws4?SO(^P%S-GCz8LmHg=W
zHS(k9Z<8NAe~<j=`A6hO&)?<z=mmx>fL=hQ0D1w90_X+W6hJS~qX2q=5e3i-bh!X}
z!66Hx7gQ;TUQnYTdcigY(F^t{h+c3+LG*%ME{I-e$U^9aR0^RN(kO&ps7)dCLOlwh
z7aCCry-=46p%)&qFnVE?!svxH3ZoZpQy9H)kHYAMM-)ac+~vaPMTRVbUPPq`dJ&Bx
z=tbHTK`+vy2zrqbMbL|Mxd?jk0gDge1&+8%@ox~k0HRU+2<|q;mvQ$geg^lL;(yM;
z-|ukoF6_jBCD2Q#lt3?`Q3Ab0n-b_H`jkK~F{T81i4K=QFF9aI^pYwi(MxKSL@(K<
zBznm{CDBWcDT!XP!zIxZ224OtP)R^f&`3Z}Xp?}R&?f;sVN3#gLWdL369-I0PgF@n
zPt-_6Pi&Kjp4cZ5J#kDTdSZtY(Mt_j3cZv{DfChrrO->YDTQ9DPbu_LV@jcy>ToIa
z(gT)8FRfA<y|hMY^wMogqnGYe8ol(G(&(i-TpGR1fMw9jsFXo3qfrLEOq(+3W%`ss
zFEgeLdYKNFK~EYm2|Y<A2|Y<82|cMz5_(dfB=n>)N$5!(PC_p`U|ICCDrM2jYLrDU
z+omjf**;~_%Z@3FUbe$!(US*EMo(5rMo-pAMo(^&jGo*l89jMSGJ0}{lhMl!S?*(a
zJj<z+LocUM4!ztm<xaulSFTSv^m1d$p_l7$IrNkvQ_xdXQqWU0QqWVDNkLEPlY*Wy
zCIvmE!zt+Hhb)g?UZp&Gd5!Yu<(DasUcOIx^zviMqnGb+dGrcHRzR<yQUSe!Mg{Z=
z%Tz$G(5C`=g)tS-D|EO5dZi&Np;uC=gkDLb5_+X&Dxp{EQwhD&m`doCI$R08@{pC$
zE2~sSudGoSz49`Z(JS|<j9z(6W%SA&u8dw~$SUYnRH~p?(Wru6Wtl4IRr*vxuQH|z
zdX)}WL9aSwRrIPVRne<zR7J14OjY!%eX62Y9a9y(YKN<?dV8%8ROhe0=f26yzT0fu
z{L^<-<-LFM+}D2kWM=Z|&2yVKPv0!QUH8RfyZ(LovCAj#qVoD{;6)tX0L#c5VVQUn
zEHiJ0W#KKbth^PLjkm$F^LAJc-T}+WJ7Kwa7c4jLhUMWsu)Mq%mXG(r^7DRJ0X_gL
z$OmDC_z<ixABGj-Be3Fp{F>Dtv=V#*R+3M`5_kcY$cwO2d<s^YPs7Ua8CVi8!OHSk
zSTZle%JDf^3a`M*^LbbWz5uJlmtd9oGOP+;fmP+J=hpQ5hOJ+@YGL}?<Mfrv$*0d=
zurU4fx6cGGw!ePQw#oP6#*G{2r|(Eizb75*S1w$b){{2`WAU%7-pg?@o`#i;-@3XI
z(L@}UkVwIj67uRwMw3t`g(Q?oDG8+z6`&LZ0ZKs<pcJDbl%gO)DM})gsb~tyl#qfl
zC8eNDN7GQIg*23DDGg;Nnt?JSWT4DQ87QTw1f?WMP)d>nWj30HGAm@E%t~1(<){p$
zEXYvGk_=@onu9VY<e<z+IVhE=0;M7-P%4rFWj>mRGB4zz%u9JF3(*3U1)%_CK`KC5
zik6@(2_+~?QVGg(v<zigC_`D6%1~CK6(}o01<Hz4fwCH{LRl56P*$btxpMmQ|Kt=q
zdqvOmr1^12y)*va!+$Y(%iywgmt8PDKPT66t-JYx=~?iTgU_zHvHVzQavsiKFh4(i
z2Vr_K&V}n&F8`U{<m5ce$4k?{JGYQf62;Y(RFbd+MSvwLBCM2>f|XX%urf*pmZV6q
zvPu?~tjMr(N)DEyD6sNM9#%mqz$z&vSY@RQtD;n3Rh8<w_4XZmm#+W6^Iw?UZ`=25
zzj^v7H@!0^Uy8GrMosUx=`~Rc>sPLYntpongj>{CR=+TYq5+n%XoO`dnqZlWW>}V@
z1(vmFg=H(+VA+dySdO9tmb2)D<tn;hxr=UCo}veqx9ElCEBavhi+)&vVgOdK7=#rn
zhG2z@VOWu3<lH(Sm(pjixV!CubX)lB2V2e_e*fRr=?6O9=G719&wWS>S0_*1U)X>2
Shif)(oxgC4?!t5bOZWeVk0ojV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp.7B0144E3F93B6769.idx b/.cache/clangd/index/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp.7B0144E3F93B6769.idx
new file mode 100755
index 0000000000000000000000000000000000000000..295f53117dd6104939fdea443c6783a3e886001b
GIT binary patch
literal 32782
zcmeI5cT`o^*67)N?{kjZE$4v1=15hfDT0CpODw2~(W6P!7){g^OJa)2O>ah21T5Ia
zE+B$n#ST~ykYb@XdyA-miiINh7CC3Rrtcf$d*hBbzW2|4cZ}T0&pB7wWwp8H+GpIz
zkP!Y)E2|-&jQnxVr{7Iut*op%p#Q%8Ztg5!!OAMb$;#@-jhPeYd{q^w+!<4_sNGEM
z50Bsa&>_ZqYQHJ7k9_l2pP7|5+#7GD_)fIt7LSg$ZK><G!g|=f%<s!mm}sw`YmeT1
z^kZs2F6~9l=kayP%`MK4t7qhIs{U#3`pi(l{YB~A-<vCI>KCmG`DDnZZ=S3P-s`*e
z$2}*%nl)8isyS7%M*ZvKqQXAcZ}A_--8!@X)5W8GKVNfqn2W=Yn>+UnSo^cFu-%e|
z-<B8dTGwT7FWZsh<L4w-<=>w;x@Y~dmJSiM6OY(jT>j(zUsg0s{psk91M7ZP3a>@P
z_WSE_#MTE-lUCX<%>UrRU)^U%Y;t?+$DiKTk2IbBxOvo^>x(;YOqo-i;9Jx6$*@lk
zK2Lvh`T6%hO`cplr)tul^PjA_l;Hc_Tk9v-stG~M<ty&)Z+7f?^!?7O8e+cQFgH3i
zu3*KUalx~0zP0gaRgE#W?+B|Yg}ctgFD~piyKqD2PVV+^zcc>};ppi<hW9^jePnd_
zZ~Hg&uOBhOy8OhrclyTn`|kWZ9~KrI|D#|_fbC~54*%ZwVN2PTp9XhZU4F!vG5<#2
zbH2|jg@o}pD)0BKE;)0s>|oLFKTQ55?ec?TN1iSH{%+6ggFhtQ=<AdHuf#JRl{aP{
zs_bU#(#yY7C;Ode<82<Ntp8%wv8dZ$R3~+dQ|?dyGrl6@*p0LU)93V?9QWIp1>GAL
z{J!9mJr|cwcAwXM_BHPE7rC2m>{;GnK-cHeVma^a!P{fzF0$)6{{5Kub?V5FD!X1Q
z@_E<qza9Nf`}WC$r+;X(VS8O@#jud*Z_7un+ca$Z<$|}ZW(~QKy3MeoqRW-JKc8E<
zE7LXZ#q1vLmp-_?WN?=|>WU9nI7jb%@rh6CvwMs08GPRS>3gs90g?LX$L<f0=s*5z
z?X?574%{~rUC+0>G4i8LlNR5wUsaSpzUGa->^t1Yw)c6r#)C8N<$J$45c$zBsY@r^
zi5TzR_|u&W!=CKTKGdP?+=xDt-{!vlVQJZ%)KmADUUGf*#^#`e{JPmc|5-N9H*xI!
z+qF)cdn}z`cWT2?!<SRD?;Y;?%dJ6IF8J3a-I{AS9MNC4y7EbN>A4FhT*@5BT<?8H
z`fMEC-p8%6#`EV2zmP*4bJ<yge*UqiU;m&<iH)H%yvw#`ZJoI_D{|)g<UFTe7Wa=_
z7`69yaH(JU+}YWK)Y6|m&M5l%h+Wg0KgQ3t3IC(yGp^^PDU0W5S~CXyux>@4xjko>
zXDqDwtiIvUq6xF3kCz{e8J0K8t~7Fv<A5E;9mPuWl$iHDrS_4f>EC?O@s|lHiEb0~
z`pj%dES#O)&8_Rez_0D^8-r)BkJuIN9bUS%Q>W9D$M;)l|6NH#-pReX$oz-{@pnrO
z%__?1|2Vqa=acpXzEkw<djF!7SqD$f_RG86EiWZ^)YE%8S7vGU|9sW4WbXH)Oh;Q=
zD?De^Zi$|Ir}Zb#8KtA%tlT$bhI`G`ZADSv*UvaNtTLv*KI;2h%`>*e&-5#9jeoe}
zf^kI0#<wp2(JtxRw|^ZNw#GJg)`F;G`k|Rop9p{PIdi>#9MLgv*1p;Es^5EiMA4JP
zgZ2v--buXiZo%@QpTB2}GEVPVcxd*w%5T|G8zu)2oBro%x1`M-E52RPKi+Or(eaZp
z9i^(Q;<ZJkzXi^m*7b1jI@Z3<E9BaiiJC)y4fWjg-Hxhr8@oI_y(GqF&-8%-N79rN
zr{e<swLKgT7aW<NHKk|mwnrTtoML+D7Jj}rX!MM#-k(YF=>w9#X_C7P{OHycCHL&)
zmFzvkcg9%b(Jy}RJLdlR>?14LeRtjG2UFp@_<bFwt$6XPef{ldF3PXFUxxLAA<jn+
zWYrx#{q7N}@ng^HOP94?37L6l`tzGJPc46cub*qyoGH61SFCN<^YFU*NuR~8yWJcf
zRd!;?=1)(0>9&TY?;2-i{7gSO#&9U}anG9Pzpd}w(%R|Kyphj;{Qhy#ufY)&v$XTi
zmB^>vcUQQIE^d48exiIBd2Q3PDP>dkwfolu&qpI2E*<T0T(Dj|qG36o;GMsJRsMBx
z+NQ9DUn)OcFOA=Qm_N7b<A$F#OzSnTtKV$g+pz3C$AO1_>YU|L5pv~Tl<ish<b_(^
z@eTXwIo&5~_RJKIK0M-m@4?))j{^I-mwBiE;xerG2jBDw(wa#h9XfGxz0WsmysGx)
zmo)xXd!hT+&gEyWck2Fd#*yzoOx%6>W%Y_9ay9zbl?xs(3m*7w|FN~V|JZJSv-5W!
zclg{ke%Ti0Cr_%7xBi3ss4A{1YkF-L?;omvy3ps>NrO84cBT55^zo(3J4Q~s|GZXd
zJ?gC88ggnz^BrN-(2M$OH*W|&u?yPoKfUY3z3YC>Y{(mIGWB6c$Lup@gIoiuh9*XP
zrmx$4;^?Mp5jg{$t}H9A8gX4Sekb3#i^~GBb;0<xUuiB(UHs<Pn||1pRGb%dwz+BJ
zAl2^suN9m6zXm%`$y2)8DawG+ADmZ9=YDy8#^anu{eW0^gB3e4X3Y5jVby@qlg=+%
zQ~#m!;N-x;jUD3}=Is6>FU#(`Y}%<sd*PdPqSd!Q&6$2!XJtjN7}4&?wcp=U=zw8X
zrp;P*v$mr}SczzKMgL<zfA7pM0|QRztfmjOGI@%gzW@B6mBi9kpAVu;6t6x0N3XQ?
zH*IJM<xxmE6jHtxf%DLKclywFhtEhGT1F=2ltVc(F$-+c?!3#7=`#{Mw6q49Y^5|?
zDZNEt6XnL+HlLi?n`1*eBa;i1!v)ISBCv_x*>k2>s{Vyc8$+y2$0XZhlF=ftiDATh
z{Gmfhb_(5RxRvR$Bwv;~Si~!n$=keJKTh|yp+7<e9itqMQ9UdIo7fx-KKSwXKV-@D
zY*f$=NxwsKv<PfsJL~HYE;)}3lIb91vR_j6O9qR;CLJAICw|gvb(&1iLMCyNeVpWK
z5!l48uuryIeUZ_IHlVtkryS0si#Cf_CXFq#o95|1l<8oUOT6R|FZHqrY-0c9y=be+
z#V*#gE6OEIP{M>x7J*HSRS(Zx8G5RtOpiploRoT=ltx$tHt7_1Y3tp{4IkOiE-05g
z${~;HViDNH<Ne3w=VKxgZRoDZB$IN;q#P{*n|PKV@hb0JZ?d69WKv8y6jMGHfla(L
z(vTB>PwQz-_e3UF1?8&XViDNH*S&V|s>jFrXlZw3(kN&e1sjXNCf!P=-VR8q7->!S
zLM9uu$_A}!5!j^rsh$1joX-!{(mXOrpfm}T!6LAU|GDwEr#H{MF4G^N`;sR4rb$CA
z0-FR3*uxI`>ihLJ^c0lKLCWDE)y*QXNzk-m)fWf;HD9KOpj;BAo{7?Qi+E)c!MNoP
zdEjqN`=VTSX_Q?WH;cd~eg3sTtoBGa$J3oqTU<^F%TYmQVI?utXI56IJr3A9cHH0d
zuPWAb50p<RtAw%xEduBBM$I?nhexk_X+wJ;lQWdV8T6FQ0-FprX3y%XS$9{a9Z-c+
zCEHZglFZ_jN$lG9ww82CwV^*mxg=2zNtCxmV3VQI7pMI=FX<C=UD7!vof~K}3v4pH
z^T_tC9`8=j(w$K*^F+-&(bgic$%wx*I(v2+mt{?Vf=n(`>}AT<BCyHG0>jFz$JJf5
zv>h@zB5IC^y(|KojOt+Yn-so&qLyxl8h(x>=Sa2|flW*gKC$mO_VIR^o{26xR<e&p
z?a?f-$=LA?@0H!ze#?e#j|!@w94gS;YZln#oxaYgU!`fj($cobWILtVj=D0lz$Wi5
zt9_^C{JnM7^vB5LKE>XrY%BtsOdR;@r)lk7ezT!Hkx3TikcFO+Szwb7)`YIu_vPwb
z<)1gERwLDF98G3{O(th|zxJ~W^R_|<qFgeyE}7Z@i@+vRJ9QO2D{qCGJH*Wzd9$X2
z$t<wRv~Qj4zp?rHuuKm?_hqN#uoLw#W`RvUHo32g^6&P$me!+!S_Mt3pjyN$liV5G
zkH<(Iw6qg4*-U9RQ`Q!NO+MX~XywbTbg`iYWKu*q6rq8_EU?LpkQbNOp|d{F($>f%
zj?%<YP8NYpX6wF5UVe2@I8V1jCh4L!UDR6yHu)ymv0HTBUN0NE3o^-|95T?|H4AJq
zH|gMSBR&y5c-jY<WJ}s?$=V{Y$+y3a@MyZ85@1b#icCUjHk7uth*u_SV-NiJ_qsA0
zx;rw-r5ti8SBt<V-@V_zVpsIBom#pBGTBCHwxRdMEMA$^@jXuEy>rT%?uJa_G!Ai^
zUKW8({uSM0RlvjPJ~s4JWU`NP*oXQCv%n_bZwLr)SN(Z{`K{YZvs-B!lUZPsAB{zi
zdo!PxYH1sk%W+Y2T(q+YZ1NMeCn2QIFW0qnM`W^t((FJFz$~!IZv|Fs?ywJkk?9%e
zeK{fdo{)xG1U6YXcaZD2Pkwwuru~q~K1tb!Mn|*2CW|&Zf8S1?@z{p;LIq`04%z4#
znFTfpxAMz;_wqF~3fLo)mxAV{(9t4ZnZys+vFOmZMKawF6_g@5r%1jQflVT#gEoeJ
zkPt7^laR@2spn~Fm_=ZdB_B>%_5LNBcVyZTnH-mFkE8C;EU-z`-46}iz4~Yy+8fp7
zEah+(Ro5)A$%^~SyC*$v_l+&x4&`!E6i<qZMPQTFBYvtEe-GVd9;vqq%2qT|n*}!6
zuzQEi>9wBk+R&3xLFtr3I+~uC1vc3l^7H1uzu14?n)XK~Pc-ZkjbahlB;M`Y!N*EZ
zd@R!=P+j&&o_o+mo5d@WQJH5Od%pQVrUxREB&lPP<Y^JuBuPJc;`T}X$D6z6D4LBz
zEx;_W$!Y7bJ-H2e)6G|OL{N^PD>4ggaz@V0$Z;GTVs1yNG@FWYF$-*xSG1(~*s^G6
znSLMD<&@O(6ndr10-Ib6|GDnU_Gy31^k8I?AlW8JT`U5d6s%33`h$~0FPWZ(Ofn?f
z45^bvV3V@neyCWqyx&Hdo`_7+CEs*ZkXc}ptF5Q!Rt}8Jw5EHbyIUYA1%jJJV3TVX
zMs9ji+s@5A@a57<E<Mm>7TDz0^Nah3cKY2)OLHifW<k>|*jfZOx%;dAn3gLu8f98V
zbvY{89z}1RSzwdu@7ljFr~T@#rFAHm7D3a3`dzcYCilu~9xOT-*j}OCkV&pq$<-Py
z0-HS8Yn<Kby{)J<_D2OBl$;JqJ{EyZYPU~#qu=amKP|0Axx`YMSk&yz0-HPxocP(B
z{G)1_egoxlNOC@enwnW)lSk}dAMZ%<xNAczD3>zIp$v8KW`Rv!Y&|~y^ubf@6uK`O
zr}LzCd6JDqU=ypkpMLgj^K7(!iK2TXU9TL>>r5GpA%p3P@4L?QCujJR>xw5sI@3JP
zFc0z>-f)KRiYF*KQzmc7gq+12vIyga5~Z;$kML9sFOO)=Kg*(#?%zj59UJny&h%Jv
ze=Ln$?5o-2@m1n(yhvg+mZcF+uS3IIa1k{Ik9<1QW>MTMs(4z&AMfQHJZyKGkIvLU
z2R6`S$rZk&LtfI;2!}Jh!Wk3cC}vm`GnMd8CU7SM((hqH_Ans*VW!(*W+bsc!E`^t
zyh%8RQF0h>!WWoc7hrj%%&<~e-c3fi3GHhbr3UhIMtKhTBpY~=9ZQ~AHXD-7P9uDQ
z?R9}Q5iVtim9kR_-(mxA!Swgpkoz!wJ?mf3zD?|3usvU}V+b$eltr93;U!$JB^)Sk
z6*p`Z2g=*TDVv~uJg3A%-p48XAkX6i=ka678yCihgz?h|FXwwL=S_sy^265hQwhiM
zfpIW>5+9NT)2Hx(DNuis?{kuWm!!YQ2VLaH5ia9<mBI3^@WZaa^8V(Pzj<$B|A<!}
zL2l-iX2@lNu1xU2oukfFBe>KE!FVl)eE_JTm{<O!1tf}MB6;K6`GrN#UfrWgMs1pM
z%Ah*qxtb0wQJl_Do$=}gwOPt3lj@AuoH{f;cFLkU<C(J#O%|N8sm^#Fp+lpw(^;xB
z9;kI_9_E=#b^q2fMbnvXQ7*SA54_yanZjvfIPHdKc{<a2+GRcMfyX>lKiZf<yW!Cb
zr58ru65hpEtTWvf#@rU(Bm7W!>!C0fbGsQ1L9=!_o}^e)iB?ylm2pRneIcC2;sGJI
z>H^j80yPwSLUgA2qBvjdfIABvTCB*iQb)Xw(V3n|h9{B{_d@OMiU<1@UHP16O~d;Q
z55|2LYZ}pS1fY@qMgpo8d(?{lc<REM9*KUB#2$p7h&`SVJ+2z+=*%7(Qujj2uaN4C
zdqN%R_WjDJzIa6F@Lu7}uVZz4DV^yQ)#ns75KnVD#P7+Sbob|{sE>a}4R}V4!;2@K
zDVZLSOpn9UTGZ##10K`kNDr094oG9ikrv&`4QS;+3&__D$k&VuvMju@^N7)zZiw;?
z(FR|m&eS9tnnV?|#^O@K>kRBtLW6_ZpM;kV28$yJz4K--61f?tH0mjhKQ0=(lW3i(
zM3hU!j-)8FZwVJ=8<6qnqH=p3x=T@#XOt9(4Y9Wg<#1V)FN@Y#p|SX!a1Pk#gba&C
zxfm8@c0j!{JpH^Q<6GyAih3xh4~6ddy~QpmbX8_Y6uGJkl*a|C2Yv_a%Jzvb;t%^P
ztf__`SVIrOPo|AW)zEK4o>W7BK)8<XRYwOBPGJY8u!HcEWKAjTs1(wdxACMD_5-3n
z#P&MG1`>YC4SdQC!hIQQddiJ@%DoMF(o^mO!q2&0&$&Rtt0dP|k{iCKtf^WWQ!SbB
zgJ(_KG#$5T?C?^WH9ge0J=D141-hNZL&xe&rJ`Iadg6?9rp1zOvE)HGTJnyTLNIGA
zemwk4?Ar(D4BBxr6FrmQKEWe!1IF$>^hC_QJ@Q1F1=nW556?BRlMg-oyP|wobR`9u
zJ$|?#?CP`ESR8yL#~Rwdh8}<)jm}g@2i4I-2+w22&SRz!UdfDI33(qgb|2(&W^6g+
zyG*~k%y5$aG1Koc<P_FFg&lw&t<H3a4LZaQA)LdG&0(hyE@#J<L#|=R)<9mvja|b5
z`|Vu6?c8vZK8fp>1o<iF|CAemUk#n<IT!St8$vjo9~;h3A-tX+yB_jxe(Y|@*Z8s5
zAV1*yJ>Z9v^w0Qy&j>ro^ZG0n^RJ^f;FhG`lDgt|(5c$vko5hvfk@HWs<X!IytlG)
zUN~f=Yx3<c(LW!jdLO3-;6BOa*xZ>9{tRA-MxR@f?Jdb3XXH9C<>=$y7d8L!trpKK
z&@gOvWRo`VQW*A97>>OUI#Yz?7a_eGBXy>W(vXYNFuamMtzH^ZN*H^<QFArBze&bh
zMRBX>fX8#(<03<|6Py?}3NDR;KlXxq-~Uh<H+<0V=qVqe`W&GKV3#yDKtq$+M~-vw
z{p3YIBmLWHfi)FqdKPE`@le#p6ALtx2p4L63pL##k1o`h2v=!>t26@$U(y6$f_y_0
zazir$>#<iJ)yV9U$2IC6(5uI9<I1n0zG;sn?vc7+4~5RuDj8ZOe>{u9PJ5Jt*<+7$
z=y7NBZgpzrSLl@<_SUdbxOem)aVwm+e?O36O;sBEDvblK3TwKgvA+cQhQ|H|VeI-x
z-S%x!zAd`r-V%+7qT@p{>Cu^5L_>?{h4mWC1_AC(@zX%h(Y!l=YpgRJ(*zvT497L<
z`L7A*U(T7a1wF^zRG-~c|G^d?d#?w96<_2$G$Dlt>tTaicVPH=t8S$~eS@YfwPJ@_
zIAwVxc6bE&iP+%@VZ0xK5>|+Eg&2&ht}|_slr7Q#%o@us1-=wqQ*`sqTNL<GdUu{~
zHLd+e|7x4f3ZqMeF?b=+d+^%92bx^k|8GyFPr$-)S2A`z4b+*MX|<UQ_BzvFjQSVz
zYVg;YmN4oP$g3E2735uvx(o7tM%@qjD5D;Qe4J5_Lq5$IPBY!fh;xQ9oPk`(sD;qJ
zf>A3V-)5X{Gkr<=JB-sE$afj1yO8S`r#i?F8K;MkA2UvmA!o2^2ARyDWecn3LM~+0
zLdX@YS^@cQR{a}tEvwc-e#)v(AvdsU1LT*i;U(LhjEH}6hQBz#%Q<y9_iD1GGp*&+
zwUD=PPFuLXB>h&-X)EOIoYQv537k^`<V4OX5pojelmt1HS3~(%lQq;E^J)a-mAtwV
zax||-L*B`&J0b7k)jg08@#-PS$9VM^<P6@B!FMMkb~bOwhFr<3mC*h=uU>~-%{x`|
zeM$Ozywg3%4|t~skn4G;ddN?Brzen~@lMYO;|7X4AM;ir@lkE2Y&KKDA6xci`gVGh
z`#C@Dk%2XxlWfmPc6j_^O+}Jz5#$QVwgU16t?mL@zp|!_THQs+1zP(8t%}#Ttf^3I
zUkJHKYhMJpSnE)%btLvB+7C*!QwW!8txL5A!ev^uOlu@uu2st+S7`kzwEje2sST*q
z_99%R^{ayVOWJ@-P=8tLds*9^*k955UV(f~tGh<7H(2g}=A#hzpW0TewDOD&AYDTk
zql7VlqZuU{atx!yK#pUSILKES<tpI;ANuU{EUEbQpQW+QGb$J^jct<YV7N54c}xeB
zilnhNX>2fQldYUhD;G@O!F-KPz9ty=_Sg<hh0Ghu_@&aBmI>{b2?Ov7L}!W@+Q$n6
z@HU3dbV6u<LKuKI0d%HD(V<ZsiH8^L%Mmn|{b!PKt1zrp7>@T9kuyf}3x{*83(}Aa
zWKM+^&eD()$dyv>N~te?hjpg=QpkO21fJ#M{cSY1-xB3pWMtHtUWkSlWU;R^t&$v8
zNiMj?8p|#?9uV;ejhdHv3!GfR0gb}}jW@oBLsvezJ9e)A3L3&sP`yu31F(Yw-*j|y
z&0FInhyS~qJ2Yp_y?2}30?~|Wji&P&O(4E_ohe$=Ihyd$yg^}S-sl;Os@Bh=Ur#)7
z8fK{JuuaeWjQ;r$)%y_DAJ1d({8Xc{?9`JQ9~C<s6^(d1US~=bJEV$6ysfA+EtQm|
zk|VyEs3KDPXz7(_MQ7S78MaE!M1NIMu1b!i;x&?AjTB6{UNY26&Uk8wmbIGpcQq=$
z=fjIO_S-ON*thBE5^hsIx2Ztl+`B{h+=1Lc`87~|NegPE{2C#*P<}0tTPeR*$dR;9
zBppbcg^_efByA!bP5VUCfy8;Zf%e$|c@OQkhwe*S>|WY$FXV%?a*%c)`a`sG2y!Ov
zlSu~>XJi)blLh%1-SZh-w>QyB6YW5Hgcr2(0&)wjv_PK6DDxO#KacUB$ABJXKBLTs
z`cOs*g&fB83}b-bAd*ocp?)c&EQP#`QI<ho$tWvdXaAM3{MC%I8tT_D${NVgOwVXo
z{zgW>kuj3%k70VoKz$t3GY)bhqa?!ok{Bfkax$YNL*B<I`(9`NeXza<807%er!Yzi
z<fDvo6mlw~q(V+(lr+ev808e?48}8q3>9q~WWgCs2)W)&M#+TsS&WheIhXOtg|pmT
zW>hXSo}|xXM&&`yXMFOZ{W-?x9OME<DS-JGF-j5SVn!*3T+S%vud{zSynmI9QVI1{
zj8X;pD%0~Se15kX!)>M$sed(NsD^xxG2DY(#~A7$*E5EC$WIu<6UYsWPXoN(CdQ`;
zax>%847r8Tx4`%BG^?Cu9q>qtHal454CG8!$%LH4DmkySe-0ec^H?Ph>hoD8A94Y!
z6hJOwl_JQ+tWpfQoK?zSXa90oUL~tkLVXphR6)MX>MyfKa(#DL<qp)}WtF>-t68NQ
zat*80yw3hLu>3k!se}55tnv`@Q&xEjxsg>GAvdu~6XaG_X@&eJ=lLfWK<fV|7yKs|
zLU<nMIS=aRal!MTem<wn=fLv~<&;p!VVqAGTs4Jpqr$lHB>%<SsKt=OIiGN7AHn%V
zKwicv%V7RdoDv0j1*fclyoOWOyw3h>I50z8$0_Teem$qGhrER|Y~ebQ`o?gE7|5}l
zAr^81XGnmY%o&m)@8%4<A@ApW_QUH-;e1jcALM)vLO#st4|Cvso6jrrc?Ue%(wP?U
z$^ys>d1WEwa9#<2o&CdkFhh>yl}M;x!YfN4ui%vxkXP}_D#)vOWi{mWyt4jv_FoUn
z+rTRupnfB-Y=pd-*Kg*H<oe=yB_8S%cqIXHBCjMu-pwn!UuXZ_u>8HevKQ+2@yb5P
zhj`@><m0?@9C8}3q(MH-E2kl!;XTjr0i^zC_~0{q2;ofLGZX4F`QS{b&*D9^pgxNa
z&Vu^0ymA(nm&+@;kPCUG@OAbpB&%o?r@{LbK`!Q%VrXB&D<zOCc%=ez6|Yo5zQik+
zAYbE^Yp=8aHF*DT@X8ISzsV~%Ay@N;YQ7V>Up2g;268QLsD)h58|ooH=MB#xH}HlA
z$S-)G7tm|b%KNlJe#!g1ggj5s&lAA=ST1xe7yPkT5cwU2t`(3gg|3y5>jb4vaKJ9`
zwoORoA>>Db@`x~=*Q0%BbAT{ek{n^(j<CH*$JEHVHFCZ1jlxS6l+YYDOcEwaVxkm)
zU8QJYDD`+DO~z{t98b(@EK$UG*H?$4l^pXVce3=nEZJR_oXMhPo5p*aCIIgi;@x_*
zOe+%QBDf=Kjy=XpO&oiSmYU{>V^Y;!qPUCf8@wY8OPlklvJ|a)a<x*fwmtS-;Lv1L
zXLCR@$zh==E`*T;J0x+3)B$fyBTt58ct#=#@G=SAHFNYb$>=ESc9iW&Zg&&s*2MM1
zw|o4w)$b>5tLlrYc2ldrskOybo2bw3F?F|HXQ^$a*8Vo-jh`qAe4_LXRCin?a<)_Y
z7RZrw`$*awuC&@m)81sIwTIU4p}UinQwpt5ft*RV&!oM{%IU7)a90S#ZypXOL6>R{
zzb2PjAPp&iE{8H{NEzXY`ww31zwXUCbg9Wy?_??%y9y>Yuj_H=`23##gJBr&Px>J#
zEMm?*RQ6qs@vg=TuZZ!MH1gk@BgOI6Ob*_l-_G6r@_%#F;{bAGZ;l?v_EVxC)C^ai
zZbv%^HMDgN-I>g+Q&{U1wljGHo^sYtxz6|vm~yz!7n8fVyCKg@x@b%nJ@M;}Lf<rY
zn>FsZU7$S;ja#jzJFc?Mv`K5&q;<xzc-RhA%^a?d8{pI%VKexv&TXES1EP39?1=pp
zIF20^T_(z9a34Ng(uKo)_zjZx2DlGzj&{dYo&LM8RJrO_2UJ2Awu>F!j=|0<)V$0Q
z@c3q-eST3~EZXCW=uq^$xKpyn6+(L;Qjcb7Fz#EhOBT7s%yIH0qef9_6zzzEBuBE(
zkz8<RghEv$k0z-*&ZsS8k7md(B##%6TO^Mb!Z-{Obus3!ds6TgQQShJv~U<cGBn5F
z6T?VRj1-OdQfIlHHaNGF<B^YYfM$R#_ATj50}TV62@TQ=0=hR0Zvfq3&0wG#Vi@vD
zho1+!Pv$UxQVVl*H8#+pKH4ou@+S9bm&R)s3BW<aizeuyW*~Xj7HZos)ON&y)H>5%
zt@B=OSN!gMaXQ{}!=|Sn{FmLXuWom4N&n`rP3X<oL-pB1cJOc%8>*`LoB^q7k<hb9
z2qYt8i!iK37>*-*+hRog7DLbHd1=Ub(#@ba9%)E1<O=DX3h6yO{NPrBUP|*J22$__
zt;Ys!H@qW@;^nl?*;)_W<lERYTkD5;?(U8gC!PwdN0+u*vRN%T;g=Biyr_pXpMfAH
z9An*%!LhfQb89AJFOIxJFOm6V1WA}jdFD|)a2N_cB7v^Kd@zC-hKgb+9Jk^mIZo<`
zqnvc6r;_0*8MnT(9Hc;fhWQi)&Ili$Kvgjxn!t*-En4*$+5g7F2)bDF(FtPs-;aDh
z{G#m6TDY<Z-M$r)=L!<Kg`J})NArOTl4Fu2CP}vOmN<Mg@<}~fq=C3`|Cn+8*3WFg
zxp#D?>y+zt7`a$Sxztfz$dDOEyN1yo_(36mFzvbqMl#0HE^%}hd~GNMk9IvpdypaV
zJgq-Z+mSB*0<FIQ`68{q2)TgP7eFqg^@Wg&XnhgnVp?AexrEl2KrW^ArI5>LeHr91
zRv*UNks)U>t6vN`oYjXzj$ri>kRw@rB;+NmehK8ItbQrvWvqS~<mIe>Ipio-9|id|
zr$5cv;UN$?<2Zc=<TIT94CG8sp9wjO(`P}>=JeT+&vN>+kaIYF4&+=;p9?vU)8|2s
z6!ejT9ofZMBIuVuUMlF9LXH*mv4S1h<%tvYagcWi`W=u{1%0YuM|OeI1brIhbU~jE
z`KsW1Rq)4CFrDd|;Cl^nhUl0fzS@&P&N9*Q4CGAFF%xo@=$HjLTXf8Zd{%Tk3pq!0
z%z>OMI_5&o6CLv)=ZlW{kkh1gX_7S#AVD)7sa?8cjn9^#9U!UQ3CS9tlhUDsw)&Hj
z9S+Y!?N8F5lI-xI7acl{q(3d$;e$#j=uG4IT;oDUrv{B<1LQ`HV<Y4yjbjt!W{qPr
z<QE#p7m!;tjxCT|HIA*2Uuqm*LjF_h_@@@sXPee-8yPKirtMm{?T}-%ZZVK!wQjMH
z<FsyZkauX^c0i8Ty2V4@sdd{4IYH}|069_XmPi;MVMCq0`QRGqahHhV5;A$g4+u>d
z{}$!H#m;!hL+c^Y@tGJv_@!ugDSBhpSdP2l`N!{{l=UnAJ9IAU{MSjH*Gb*6;a|tb
zd52#*+!ndz(cNP(o=MJM+uN8={r(5~=S8C1BGCs=r$S@?`1$Rfnc3~z78PPL)d4Rm
z+9HO;-H`WC;vUHRDRDp50S|O&ZcK>>Ag54b3gknSc!=tN_xyCG!<2X!@)1fr0{J*4
z9;Z6s_(HVxM~SJB(<m_w@(D^jL3O~ppr{Q{;z`J-DDf0wd|(i*k}rw!B^cmrJ~oJN
zH4gVjL%jL0Abz-utG0icIB?tB|8}qxj`KqswB{p*I7hT8EsEiy4SqN}(@}BEQE@W%
zDYR{W_D&Ut;SpYkj-!bQk_|qo7H&D4h_301C|@C)&Kk=JMSM*WZ!)1Z%U<@rgx&`~
zYd?E3Z|<w<+ksH9CK%9V%H=ZEm8_$$P%c*>U!`2GLcT`1T!Va_^0`h0kn}ex`6gvU
z_&())pYkA+j0cqS1IRU$a}DHL%DEPD9pzm|`IGdIDESd(L-;WzKZaaS$@P#!X_rvC
zE1A?Rpj{R~UP!wvguIA$Sp+$Z_6ef{Nct$+C5rA!{01v%mlcp#(k?3@ucBR6L0(O}
ztcJXX_E|#*ko?xs@;cgv@K)M+EA2rhUE65qZIHLq&f6i!(9SWCV`=AD$Z@oH9PLl?
zi>Kvy+J^8>THXmcftC{>C(?2v<W$-vmF`L=foZf$8sv1^B^~k!+T{e~leEi8$fszZ
zQ*;2yFORm)qwUG0GM~22hkTB<J_q?cZ9GrA6Z;Fa@dD(FwDBV30@_#rxsWy%LN20>
zMUabWV=?3s+E@a)ls1+^E~AZQkjrW7a#-I>w0w!SA@#jX%a<Wvq2()(uhQ~W$k%B3
z8szJ=d>!%)TD}4K7A@a`e4CbUL%u`HcOc)V<@=By(DDPwb+lXuc>!y^fVC%+?1ilL
zLdc6)>qU^mSYsILPV5)6#>J4sSz|cl2-X+@Ig&L-LSDiemq1?18ka&|#u}GFUd|eq
zLylsNQIJ=#)+=Cr*Rb*$)`rw~Ei12uyq=ZULyl(UXviB`c_ZXath@>G7FON@c`GY#
zg}j}Ww?mF$<rv6utQ-e<2P^M@ypxr8LQY`i1jxHsc^Bj)R!)MP%*x4-_p|bT$Ol;Y
z0OUigd<gOh&iVvrPZl{RIqQ><PjS|#AfM)pr#W|GpTQY3AfMrkXCP;C#!SdroG}Y>
zHfPL+e3mnwg`C3~b0Ft(#$3pGoG}k_K4+Z|>wA%tFLE}dz6G3I0J)Hp3n3SCaxvr*
zPA-95%E_gW%Q?9mas?+>K(6HEO30Tu`4Z&IoO~Jb6;8ea`5Gr*gM6KnuS34U$u}V1
z;^bS9Z*%f($agsT4&=L>d>3*xCs#wh$I15~Kj7pCkZU-(268PY*Fvu6<a)?YIQa?W
z=bZc;a=2g}F4&VrZG>PQ0Xb4IMhfmkzeF%DfxJ{OE`_{Zu--1%lSOZgU>yTFRxrj2
z?nECa7~>%C5R5w@9}}#P3HD@BjJ}L^9CE5)OcmUTK20#DK~5Kp>5vPAUIjuh@go)p
z!wZDbgo_2aSg;{{RS38$^o5J;fNPMi3$E7%PolpexZZ%AA-ZNjANm>5^$g@p(KQot
zmgt%VIa_qihJ03ZJqtNU?2;q;;*}EG%M!cfLe3Mr<U!6CyW~SYC#vVj5jUObyr`as
zTqxQUiXaxTMzpOFjikJ_lJi=)C$LWv_et&XLPUoSR*L(j_W0x_n*2-hQOO2>v;jE|
zCHa_SgO8P>16Y!LT(ZHR5<!atNluk)a3rb@9UAGFCUt^NgO2G^C+IZjctYxg*Q9Mu
z1LH}_9Xbt+rzCgiG%%i)+=<g5Ly|Kj8~hCnv@;}G-zDe1+MEim&o!RVso>fGxl!ZV
z2)Rk)+61{-<Jt`Qg~s&-<Q7er7L6~-zg5$v74l0>mzR+L)OPt(3+zL+YN*zgq~E6X
z+@|%zF8MYmgXeb0F<Q?U$gx__Sjcf&&p60Cw4OU4$7_A#wLMAxJGH($Atz{k6CfvQ
zeG?(?*1GQ2dgAmE_lB0v{A})vfBP5`et>9|cu$n?iLd<b8p}s&@G23%S;$*)Q<QJQ
znC52D&`g$>Z86QtDhZr2Hh;?IRYo^Aj`98D{ZG+q^MGi7fb9A#tN5vRJFAqMHgAQk
zhcEV4EMHh4mj|v&MgRQz&OlVelI^BnRx$tfk~^yK?!$S<f7$(C_E+%XeYCnae_IEy
z#c{Jmir2S(R-Nk*w#fJIU8r)&8YNldjJHB@M~%Ed<`3-Ptk9;}Yo}&cU-vpWC_T1q
zGc{jpm#=lld)BL;Po45td&d%V!!Jv|m!&>ri~q9p)@5k|;Vl}wEpTgdizaZ3W-#8u
zT$3{*zG&g^KIjqe)f)C{o$yQ3wg+R#);i%mnCKoU3E$7}(RNCHKh=9b)gQZaq6?$P
z9zPtm0V(E*_VdIM*e9_;*}N-s%H;;6DABer(e}ob$Bq|t`R0!Yk;^}&2|T5F8{emm
zo4jxQdiq~M|AX7rn;Q2lDDN0yv`71wwAutuOhhtjBn)FsWYk2+1&mq%`ESPQZzhN=
zx*jo3k03WNK@H3pqHkn^8X=!%)zd5p^DAK00?2n*^$z4^oVpB#{cYo%wsApZ5wVMN
z+68$(7qp)PUmZEX1s#AqpI7I@Fs2o}x&m@Muf{{J;MEGq_j#xLd=QzWKINUBLVm#q
zz2L_XeG4Dd0=ZlWC>P$slZ4HKuLgG-`&PS-tm(XDdmf(oIWG-AFO9^zjI8N`WP1VX
zFG#~LK>bC@_9E0@l!jk~`U1(e0O|{*;RUbjM;AyYQeLrSTMX@srQyZUzC^Muf%+0@
zcnQ>(O17m?Un&hRh59ndwhZdaq~T?+>qnQt`cz0gDx?5XzY1wWh4en*O3Aho=2s~N
zS4sn*en=(R#l@o>>dwrc@4~Hf>+)T<groMiQG1S}+C@<wSh4;0h-LY5(6Il~WyTJv
zG7di7wBi5Qxx`h~P4~CE-US&Rp!yu32H?)6&G*~qFgaC>Pv&wuQ<W%Jk%O;oU!~AR
zNnLSg+_u%JTLF2cq+3Z?WBKA39(eFhC+biuMY&Q02SJudx+PL4lJO?Vdy~{3Gdg_p
z&*#Z-MhO!N2WSp|?yE!HA*0S?dXNUcl2KPe-p8o>AeS?0IpiEx&0%|x23^jo<&bMw
zwFdGUPF=$R`*2<j=X;O_x1LwmL*C7+yCGlW)oX;Yw-9X(n?I|DA2&Mb{ols5CH;7u
zGtJo^QiWcW7qt3<b|bgo1?}~M_9xszt1VF9LVLA9{d`89&w%vv8L#;aNFU0mp->;n
z_=PgPN&ZV2bt%*@WxSTc^vf7^8PqRhyp}=zYDQfR^{W}L)lk2NQP)8I8pdx8EI)}+
zlb}9{@k)Z}lNmJ`>XR9-WT-#Ds0X0_0ONH4>Qfjs1?p26zZ6)0Dx;=CeJbOX3e%@D
zeNve>$o)!V)HG<H#(1Sc`!uFc8nn-3)J$lf$#`W#`%I=!CbZ9D)GTP9#du{w`z$6n
ziy2Dllg|X@GXo)y%!gkRJI4f_gZgt!$T?;-u`goOB3ND#<5dL9D`wPUs4r%`ilM%e
zQ7fUolJTm9`YJ}Pg8C}PuL|D(YR0LW@gVi7X1Z0wqZrjp|7vC!)Q_rW-X(mGak>Z7
z-($MngX!-v{qMo__n1-lVETH-sUD`UXS&tH^z}^tdYHbR8C4I{KVh7n!1Pa;ZckwP
zCrtk*F#Qu|)DxJ#i3w_g&!>qAX@bwEnF(r!`er7i8S2ll>KWFJbb)7BuQRMa;Y?P|
zg!)X@D--JTSTzso^H{GusLyBBe5lW7{qot~B>y5-ErR+Y)~g7nFJ{$Zs4r%{ilM%e
zRV$&slJ%;D`YKkfg8C}fuL_obmsRgV{ax1UE=*s|s?|_m&3aWseI2XTL46(TRR{GC
zS@j{*KV<zL!txtgwGrwYS+7QzzLD+I$i6}DS0g*35q>GJiB+3m`X<(^38rsi`!vDy
zP3(v!n0`K|&ga1MpU-*C=fLxy&-Izlf%4~bBj&>|=Y?`=C`=#9d4<CCp<Hk%H<Z*r
zlp7uT`caplaBd*chjXLCxpxUia6u7JAHju0aHEMnic_Ot`B9u#6fA!Qr>=nd6`a=!
zs9(pa>!5xe=d}*%*K_K6s9(?dt>?h|8^bxpa2}*SF<iG8cvK~Z>mS1ngZfc1@cLpo
zr&yRimg^P^)5miCV`2JOZd5ExpUgQW!}Q5qw`7<;nd_en(<gJIl41JYoYQWYemB=`
zH%z~q>%SYO-_4EM4b!J^K`HS0q;MfA@cA6%f(}CcK`!JV)Gy%G1-u)XLM`CE7V!Rr
z7xL;ts9(r?Erj|=UX6tMNZu<F>X-295~yFo`z_&nll)il>ME#T#e1!S=~wgWYN%h$
zd##504ZOMm>NoIS8=!t8uWp3;jlADRSbhSpCO~}x@09@4C-Q0{)F<*@iBP|nSNB5w
zUfydj)bHcfeNexT_uB`{Pvs4%yd$|Esk~b%{PJTe@1F{f#-#FZr1B%7y(yKSNI0E0
zq{IBudAD?!Upnuf4)aUr-$;k~rSqnAm|qTW$btFg@NPLUzZ~8_2j-W<zmWs;%i&Eq
zFuz=0&4u?Pk5}^`=ksbl<a4}w4)S?kJrB8<SBqcgfMUKMxgRCGS_1VYyjKZ)K2^M0
z1@%?DR~6J>;?+w~e~I_H1oby~^#;`6;Jt1@{Y_rI3H3L5znk#+*YHj?ya%ao4d1N>
z9`&i=``7Tppng;h{PJim?^FxZ*Ye$JVftFWe=SU3%a5vs>7VmX&tdxKe7EN?{d2zm
zbC~`)Kk7M5-@rRH!1N7#w+5KLf$!e{(>L&=8esZXKB$!+NUpDy4{7B`6Mo4Dy@dLg
ze8@|vuMh$%gnrnkg4VV|Kqcfyg8E2s!y7Hh0ewxGo2ITs%kXT;F<bJ$5!~p5r&>cM
zaR%<0yJ+_M{_j3PtK}xi&?I%ifx&H2|AuDBFC@bY$Ssngh48+~<^J0v+gl?y?K!E#
zImrflCD4f^sY4Ou3aLW{;e+>oKG-q*apHgcx8e>3?qByy!GjUV`I4{I=aXYi$7Jzh
z*2G(H|9=OX>Hn<*E%j{4`^o+4Z~w=?nt1;h`T5OX-6tNT*S9%Nd;H!bDCNz#|KdQ*
z*lGH*U5FU^e{>+?&@SYxGJn0F6!qU91j#Ne$jiI;<mCV07DIM*^q-sV1{TQZ1BY5&
z0r^@}+t;Xcg^-K1x+2KMT3s>ZQmw8Oa+y|F2Dx0TD~DX6)m1>Q)aoiBS7~)skS}R<
zmmpu(>MldRqSal2d{wKvN;q$B!r11UbH4rmJ<k8*9tUr{qh?Yq%GKg4XSK%iLlxLL
zo%i@=_YE1<Ls3I}-F<%E?U%5>2E5ZAjl%`f_yTDX8HdZH@nw)JrMD`jF<5^gxXZ8O
z#yGC|kMHa*PMB@i@ckwy<n$iK4wLZ)H4eB&hY-s}xtu(jwokX}BH#z9mP$cOrQvvA
zL1)@1d2f{Z5#B0|+bT`Rj6Om7&+m-jM~j0i(3I@HDBl-tu_21%mkf(uFX)}6(g#?N
zP6ht+Yb38SD*e|4Z|AH((5uiQbZQZN$Z^0{p;N2igU_1kP)vHKmx2!tQZIT-n!7G|
z`YrSprc!-UsevRQ|0LDtB>Ap-N%sd;2i}{y5*gkSjJE`DJpXL_ZoTm~Vf34vH#^+l
z9f)q;F-ASc_!7tHGgf`Z`Vz;*b5?y0`7o~@=6%UT`3SEbfqax#j}pd_E~ub8qI?IQ
zwx|~!>)~mOR?*N(PFu8v9dubuc7xFQHpzVrd1`2f`=8%&!8f<;cG&nGMV?gKp19Dp
zL>Pdl=H*W|j+?nIe9`~lICOQ-C-(Oy?^ICPD<#jB<QU-9w2>9<?oC~c6e(1n6teq<
zD~Co^^KZ?NQMF#s*9%?9gL@|EpAp8-37uOo|MUz=_<#3dx$C_z40)bpNBzflH`g1!
zOl}_Zo&$PjX;hyy(%5c}5@!~j^#27Zj*3o4#nJdk$Su1#pXGPjcSEgizNXuJjXwzy
z3DtB9C5%7Wj*h&Ve;bHAycQv_MHmb_guqr|FzgTlUkZb9hj81Edm^;&7=cRMq?I;l
z^(3|=Q!8a^^*FZVZ=a*d0mfb1QJv>$I?mJh;;P|@Hk8A4QNB*%iZqrVBf{+pR|(bG
z{1Zf^&S9b$2G768OJclak3Wpxwt(%?M9#nC7!|bpZT`I?oDu$^B2<;DqI{KvG@&mm
z|MROwxGJ?9(!coJc%UupB~^4v6@%~@9Q->*XgX~E^&*@#8VW=)f}H)rzh;Drx+cok
zU{U7ZG{QyUU$8*!)BH<DB*$_pxSSe;fBFO+{}IIm5>54Z<Ixk{_qw?M$2X&TrTW4A
z#bbK@+o5sKe#`pZw)s^1e?Q^Ekw@Hf_di8bF!-tP=goVPI^LVJL_oo&Oi(WK7VbjP
zH(i*ZJjh{OP#E_X8TuA;L5m?5@@gUc2yhXv7D28P)H?E;8qLi+PISMt<_UUT7m4B`
z(FKQZ;Taxk3Fe=|!p<P{8(3dV8;T}4LzA>O40V4@pl~o5{=kM6enD}_h!uV@v4_bD
zXZG#fkH3v?n=FmOGbStAG%>Zu=gz;)`l8)?<0rg3>g{o3M!z%m&9|)Hd1YcXj{Ny1
i`Ev~Ub3FO;UGnE>^5<LR&k5wuvE<M9$e(YMKmQluJ*RB|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_bwd_weight_xdl_instance.hpp.DF771DCE13775534.idx b/.cache/clangd/index/device_grouped_conv_bwd_weight_xdl_instance.hpp.DF771DCE13775534.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8c93763ad6ca2407f64de90c0da166595dac6f45
GIT binary patch
literal 17894
zcmY-033OxSc?MvA|0U+mmM|;>j59riG6zVTPC}?79a+|9Nl;2UEeQ^hBk5X_m)meC
z79qsFKC&!tl5B-0qz%X(Th(rDfu=j0X6u}0YxcDVQreIjr*%Bj<WR?&q@SItet*u1
z;QRmgz0c%&Y+c*;{=vN-&r9yEE?4;8CmxPGrd`$5)pe5p7kzAE{0r38_3rMjt|wMr
z_~`ro@L%3~{rt}!F#hC=cYXQu52kKD``^!eD)^N1=-*$TdhCtIzW9+Re`xZ*8e4nn
z!xtaDsqo_muX*l?$Nu?`{bxSE`NhpUzx<w$MVya*_nI~1gJ1dU&83%qket*iH@>Ss
za5i6l^hXvY_3bB(D-O$x|MQj$Z+!Hn&#q;j`1WVl{^F7MHvaPc&wusJ<+@&d#pbC0
z&Of~MQ_i<sN?&~8wu|oT8>`QpnZEFqT6pYP@8!4OWqI;DlYi&l?R(+%-#C%F<B`{1
z{^0qqdVc<Z>Bg@o?tjOtpYeVAO;hXbUw`IX7eDyJXWw@1`KO-wy^D8LtmpsZ*Y2xj
z-<>?Q7|FkS<*9dTm)BSRYM|#6zTcon-*c?{8{hk>^R+vAAFg~fVtM+z@BEKy>Nh_3
z*n0xszy6cg-BSD4?|0q)rVri!r*FCazj{CZ{=d0x>2IF9^XQTALxu1B@cb8E_^o%8
z?s{MLmI>Eqo=AM@2c>sjeBqnUd;aOI7mH2$v&_WS=T7&$ICJ)0zyCe+%YSrt{mlF3
z|MO^FB<F^odwKey8!j$<{o}D)?tcEk=kGf4(wP(K@84s5{%x<7jjQk5v;OUsE58`E
zK6yTIdEwmnTq5*OzBk|Z>?gnZ_|(-`-~HI-%i*`avhtsg{G0LGi!Xkz`^w9Xq09bv
z8o%t$ylLc~=!YI2{ww~79{u_HONsk_`+;NP#5G+#w|3up^5CWG2!GW1pEpc@>w)wa
zYUhuMYk51`9d#VMbRD@$>~EiJB~O2#x%Js&;wW!l+kNekgO^U*ws+~jDnI&A>zH^w
zZ{N~=%hd-jT}Kp_ySfs=)1KPf-v0P8@jBi<*?scp!AqyTAeJhB{_n->Pjp?)+qa#%
z?ac=-2iqs#(A9eH6Sv-Z*YiuqL^p4r?mm6w;HA_4#7B<&b$3te>0{zJZ=dQur5?O=
zwlDg>bfUTR-Ow>{9dAEH-A~c){@=ek|L4B;(zkbh%XHh^F>yUd&Uc@;AG~z7fBK0_
z&-vnCyYrYh!rN!M&zwAX=>V@?_rS-_Zhi9lW8wzhKG%KDe(=(1@5V1zkKJwi=7}zq
zw{Jgn`#TO^I{)XcT=U>_FFx|Q9~~3Nc>B8U>y8|}bPhfd8Hv8=IQ-~+nvNcP*p@%D
zw_bnQa?A0a-Z=FxQRffAbB21K#yLy9C74ZdGGX47Art0JSu$bX949m8%^5Od-kc>f
z<}GouVBV4;3+63ZvS7Y1PJNi~%TOQY`?A!Bd25`kn73xgig|06teEeQQ$ObWGt`gy
z{w(!l-WDet=4~0WVcwP{8|LkCvSZ$!Av@;nS+Zl^5hn-c9T{?9-jO8-=AChJV&0h{
zC+3}5a$?>UCl}^j8FFFXl_eME2jVn<`GE`#V16J=1DJQm$&Gn;hTNEUXUUCuPn<lM
z_hiU}c~6!+nD@rXi+OK`yqNc9$&2~HI1MgQ=TrE2&tQfIpT;>$gC&@Kaq?l_mmweK
zeOdBheke{um><f}5ax%nG=zD7ocx&gXULCvf0q2155y^e`9OvOm=9zrfcapYf|w6x
zD2VxBmV%hq;-q0-%aDe7ElV2ahvPJi`QZ!=V}3YG!<Y}nDTMh@hC-MRWhsRDaGb)J
z4`(Qh`EZuPn2*FMg84{>BAAb4DT4WtIE`R_Bts*ZAIZ`P=11c+iuutDjbeT@OQV<{
zi_;k9$1*gA`LQgGVSYSL<Cq`M&^YGDvowx*Jx)62^$h8l*R!N!elkvzOVoM$KHf8#
zp~<Il&eCKFW+P4p=8X&)m^ZRyV16o2Q<$I1&=lsUvNVPHSe#;*k7X!^`B;`>n4hLN
zO%pbzDZ?w+nWikSU@Ok8xYUYsD=xL-+=@%B8J=gP)(p31q}B|#W~5evTM4O^;8sFv
zCAgK4TC?1mm0Gjhnw46!+?thIbKIJfT65f*lUj4!nv+`d+?tnK^W2)3TJzkRms$(l
zT98@`+**)Y3*1_eT8rFTlv<12T9jIg+**`cNp2;jR+3vusg>kbQfj5Rm6BR1Zl$DF
zid!kEmF8AjYNffAmRf0UrKQ#iw^q2-xn4Zpv%;+vZsAInS4yz5&aHK+wa%?|skP45
zy41>XD<`#b+{#I<9Jg{(E6=UG)XH-!FSYXA%1f;Rw+d3Lz^#JRDsZbHwTj#-O06Qd
zic+h{t)kQ_ajPV?O57?*trE9NQmf3ZveYVbt1Pw3+$u}03b!g!tHP~{)T(f+BDJdA
zs!FXYx2jUB%B`x@+ThlP)Y{<IhSb{N)`rxoajPb^YTT+xts1v#Qmf9by40$3t1h+b
z+^S2h2DchgtHG^?)M{|6A+?&^YD%ppx0+I`$*rc;YH_P2wOZV2Nv#&QTF~meKk>IO
z``=r3zI{1aZhhgso0C7I<2`rWcSoS}y;A49P3`ExcbT0uU)*VfVYt-O2OowhI{5hE
z<kNhR^yxlG`l3Ea4rxP>9MXp%ITRg&q+jzx(y#j=>5uv$8PEcd4Cnz!2BHB-2DKn0
zgL)8>!DtYYnx;Wg(=|wHQ4NyA+At)C^<hX3M~5L9(n637=^;plq9I6zwJ;>ZdKi-7
zXc&?aEdt4i9)V;e8iC}9HUh~JeFTyt(Gf_FYNL=G)kh&Y8Xbk?m^KE<F?|e@W6?24
zj%(wP9M{JoIUXH{q^{|Z)O8(_dQ^wx3TZ18#y1Rog&xK^x<Zd1&g-PDV}70Vb<D3*
zbRF|K(sG#3k)FeRj-olt=Sj<BK2Lfc^LdKqF<&69fcXOH1<V&HTEKjfv?At<q!%$?
zq-YWIrTy<U;QW;KuVvx<lqg!le3`T|=F6m)F<+)=8S@p=Dwwa3Ucr2Yq7}?nNvmSM
zN_rLZRf<+Izd_mt<~K;+!2AY9H!xo#t%ms;={3yPC|bjOowPdU>!jB)U#Dmt^9|A(
zm~W8Yz<h(E4a_%5Yhu1hdK2?aiZ(IdBCUn_7U?a_w<y{=%$q+jr@nf7^~-ktDU{lK
z?(k<)Ce;L!SvA9CQ7tg_seLe6RVz&WYClXi)drJYwZr649WXgnCrmEY1=D~!0FzsF
z!{kvtFnLukOoQs+Ie4>CeJ~BFLooSOKTH8N08>y6!lbDhOvCChOd&M{Q&<hd6j38E
zji@6qjjE$Cjj3ZWjjQ7@>8cLXq&j)7^PYXY$50KJrqn5zVrmSgX;P;N<~XTwoM%X#
z!8t){0_Ry$XK|h*bq?owQs;4AAaw!fMN$`WPLi6$IYnv;=QOEloL5L)p>y~JT3yFE
zM`{k|JgIq{3#1lsE|OZrxkPFS=Q62foGYYOaITVC#d(9&4V-JF)^M(qTF1FTY6Isc
zsZE?)q_z&{e*bS)?)=vG3~{{Y?5$_-?fhJxFV1eg<KE82S?6c{pH_|@{E+|f>a4d-
zy;tBz&Al1X`(bhTWuSLW^seK)C3=61^PcE!<D6D{mlWvlD7}|pHnquwo+%?t=$Y1p
z2|d%6FrjDK6DIUbX~l$|X-6@kXKs@jJ#$8w(KD|JGkWGNVMfoqC(P)X(~22A^NwOh
z&(bCfdX|i^pl4YV7W6Dz!h)V<Pgu~iq!kN#mL0`{USFH~(Cf>HKJ@z5L?3#6TcQuW
zzCF>0USC@2L$7Z~=|j)jCM$Z@jIg3-T@zOHtXsm0o^?-H(X*x%D|*%)#fn~koBGk~
z&xn5X`qxB1di`6XAHDuP(T`q#TIolxe@E#@&(<azdbW(Pp=VnYHuP*;!iJu0PuS42
zr4<``wjITWp1n<W^z0d7N6)?{?C9CIgdIKmp0J~5Pb+rx>^q7bJx7}y=s7aNfu3Vc
zIM8!!2?u(PJ>fvlkyae&Id&8Wdd@aE(Q{^m6FujeaH8kj5>E7-d%}sHGp#t$bM7ck
z^jvLnq36m77kaKW;X==~C0yva_Jj*PS6XqQ=h{(R=nb@K0KI{X7(j1eO$?wnuq6i2
z8`u*A=nbTm0rUoTlmYbIZE~aM&ImVp?ls{?&%Gtw=(+cV8$EYgaiiznQQYWx+T=mc
zlMx>DJZr*(o@YyV(DUpG4|<-o;z7@|qj=Esw#kd0HzU00dDny&J@1zAqUYTcUi7?a
z#fzSINAaRJ*rvfN@T;c5Wigl$hnI<it77m1&YNQJRh)Ol;LmYRDuXEn`rFFjPhj@7
z$%meAS@_WNtqLD{zD?mn&$lal==qY04?W+u;zMt!O+)AnEsG)ahE~N8dPAFH2)&_Q
zF@)YwQW-*TXj>UV&)+6Ldj4hMN6)`1{OI{Ng&#ftuJEJhPbz-&{M(8jy+E4+=mnNV
z0KLGf2%s0(6an-CyCQ&IAgKh<3v4R^^nz^)q8D5iLG*&FB8XmaQv}fq?usCK!K4yI
zFSxA)(bL+bp{Fei4LxmDXy|F1LPJm66&iY4Qqj=UwiOM%;WiDUH@qx{(HmYB!{`lf
ziedDIcf~Mz!%1Zrz2R+T7`;%NLg<B-MF_pnstBPM+7u!5Lc1b_UMQ)A&<kxVA@ss+
z3ZoZZ7Gd<lt0IhEcvFPY3-5|Bdf}uJMlZargwcz%DS}>PSwzr_tcnPFkxdanFS08l
z=tYuB1ii?%5<zdIO(W=yEQ=BJMpnfLdLx@+1ig`6F@oMmQW-&SWLp_QZ?sLL=#4Im
zQS?Sv#VC5Cn_?8b(Oof$-e^)8MQ?Ol8AWfbO=IYdEsHVq##Y4`dSjbn485^kF^1k)
zQW-;UY+D&aZ@f+8=#4LnarDMl#W;H7n_?Wj@m(>F-gr_OM{j&v8AnfVla8LgEOhkr
zRiUG&ZwehfeOKt{=}ARLPv2H_^d_&+<S*b?&6CSwG9wOuj4*jYOs?U)B_@B2b6ZSa
z!FfrUTvnieNtwJ1vvGwC^o(U;pl4hV271PpFwisF!a&bhQVjHrONxQs)D@aSZ)#ag
zp*M9wOrbZmC8p4uYKtlKrk0c`^rkK;Q|QI6Pz=4;vWTG<yC7od#kNEYy;xhs(2Ff8
zG4x`Wlo)!`bcKJ-j^Z>e3;uOHd`zb40<WMtO<TNz?liS|1!evPzrx3YkKHs~;uX~6
ztj8ri&N}~EAM`luaY>J}9+&hu>wGM!wWRQ|K##K?m-HFdXC!@wb-q0Sjx)phjHJ)7
zJ|pQftn;y?){?@<0)2+{8A(sDo{;ne>wKF594EngLedkgCnP<=Iv-1FEh&5~&=agD
zBz>0kSxKK|oo`ow<IJ)?E9tYW&r13%>wGM!wWRQ|K%ZrOR?_EKpOf@C*7>#vIL;jF
zbCN#C`kbWCvChYmT1yHa3-meG=Olfe^?6C3XPs}KfaA=wJ}>F>tj|mOJnMWcskNl=
zu|S_^eO}TRSYMFz1=jgC3^>jL>kE>;!1{uuFR;$Xl3Gg&9}Dya))ypwk@ZDMUu2zc
z=YZoZvc4$ki>xn7`XcLmEUC4m@UcK&WPMT6ldLBtJ;^%X76QjfvYwRmB<o2@PqNO(
zl3Gg&9}Dy(>q$vZv7VCj6zhC@2^=TIdP>q$tfwSB#X28LYAq>zEYMS|rzAbidRo%c
ztn+OuaGW&jX-QABo|g19>wGM!wWRQ|Ku@!tmh=^B^KC3w;5L>O*7<f9arlGp6<Xz$
z3)0Faue>U)?DERbrIn<zLMa80v&}0%ft7V?^X)LGud~j#&A@TiSznj*b=u?=)Yoa3
zS5RN4q_QsQ>$J@)sOMPENqUZTzWoOD9P2qr&#|79^c?FsNzbvKlk^<xIZ4m6o|p7I
z>wFsz=y}%jlAdQhFX?&K^OBxtJum5b*7K5HV7(ye1=jg?AkYh}7bLyFdO^|)tQRD`
zz<NQ_3#=C;y~uh|(u=I~ZAG9LSuaX@k@cda7g;Y#dXe>_q!(E)N_vU)lBAbd=i8G&
zFR@;d^b+eONiVTplJpYmB}p%_UXt`O>t#tVv(C3UfnH|4Ea_#|%aUGZy)5Zv*2|J!
zX1y%w71k?~USXYYmjb=QdPUMJtXCwx!g@v0E38)}y~27$(yOdjCB4cz-*yFhmG!Em
zS6Q!0dX@F6q*qz5N_v&`s-$nQz9H!wtn=+#pl`6gA?X{eZ%FzE>l>23!TN@zZ?L{0
z={44Cl3rt-Zvz9p#(GWCYpmBKy~cV?(rc{OB)!IZP15VE*CoBqI^RwPdY$#Uq}N%m
zOM0F4x}?`xuS<HJ^}3`tSZ_#rgLS?w4fF=<4M}gX-jMVL>kUb7u-=gL2I~z;Z?fK$
z^d{?kdmHFY)|-;vWW6crP1c)|-ekQg=}p#~lHOvyCFw2J`8GMwTdcPvy~TP<(p#*z
zB)!FYOVV4cw?M!D9Upn*)LWLG<a_FFHQnm!JogoNV&#$G3(x;MFP}Vl@}|xsS)t2c
zzyAZ*1>VlT^1b2e8?N7fT5JEG<2}#N@n`7fUrmQkaJ}hQPi`Il&UZ368HVrBCng_1
zoQ$9Wd&YzTdsD$F*qfS|g1uNU279rI80<|`aGGFunkJ?R_W9`ue!$@H&k(1HIPT9-
za0d5hXkrHU6BJC~eu5?vxIat5S=^tciCNsAqu?Cw&(Xvj?$1+j9{1;IVjlMwD7b+8
z3pBBS`->D@#QjB@Sj7D#1(Ud+q=_W%rzn`h{S-~4a6e7KH14NqB7L~uxg&sY<J^Bt
zuk+VnVaNSO^H-Km^WUv&PhNXt=l0F?)t|~Oeuh6j$9vA)bmmOwIl9`96CZl$<(14;
z$9vA*bMF4b-|wHl=luPLzwL<L8-3^D$MGq8*HiRR=en}~#NWT}{KJ3l;UnMj`ddzQ
zZb|4oDag+k?ms5j`Qz>1%KfpcR_Nnbb^6D9)_G-JTFLQBPFl(HN?uwi@Jc~iDe_8D
zS}E~LNm?oMN?BT|@JdBmsq#uyTG`;04QZvuD>Z4Q&MS3krNJu=X{E_4O=+dYD=k>*
zJVD8)ZU6B}JZ%^Fw5?00?E;@RT;U64K5e+d7q5KUaD^`@`Ly8*U*z#=!xg^p;?ssJ
zd@;nQ4OjRAg-;u<@I?ckHeBI1az1Ui!f&p8+Hi&6@c6Xh3cm^QX~PwMW8l+<EBwWp
zPut<jCvQLU?kB&TdzJqZo_fQn>pExgAAfJF@k;Hv_w!fOo3FlE`igq?8M>?U6}9tM
zkw2Aw^-yZ(w0AxOcw_%j)Xws^KXCtxzuP|XDb~*3clHlDpNX%$5*d5;!Q?M_`SdlX
zZ|HmmbUxSkAo~xjb`H|{T;t{a2UR=Eom+c&dH(^`&T{A3)Htv2KbhKD=iB?V-f-tw
zJn6j;!xZg(944bTb~u?d6ZA~F33{fe33?{O1U<86hMrkBL(d#FL(goOp=Z%7(6i_k
z=vkr`=vfR4^!l_u==JG+(Cdr#L9fr~gPv8hLeHvOp=XU+p=UL$(CgRwq1UhXL$5#D
z550b)A9^;;20fc@gPtvFgPzT>LC>z)p=a0a(6dMF(6bwM=s7e8^c=badXA_AdJe+@
zJ*Vb`o>O;1&lz<>&uKWJ=h9rzbLlSVxuP!UxeOQd2DAa_4d?^V8;A}-Z@?IUo?CN6
z&#k+m=Z?Ce=QiBX^JpIEd2|o-JW&tyJcb8)Ud;<VukMAOH|m9+*YHAb&=`z$zD^wP
z@fkjthKwPY{DvQ<fDwQxXar%>3=O7XV;H885rQdfgkg#p5tv4d5tv4eQJBVzF_^}U
zahP;NhiOuueE9Iaf6|zY9ZrUBz@A|kus5Yo!QPZH1$!|)2756h27A+_PZR7;lQB)O
z&(CDWai5=|<i{Kj|4cKa&*1(H88f(_AU%Qm2{ID6KTG;7?$44ji~Dn=&*A<Y8FRQl
zPx?IW&yz8a`wOHm;Qj&`3%I{X`XcTxlCg;UNz#+JpCluR`zg{>xSt{;h5KpJ)3~1|
zBaQnjWUNrE^Om@Ob!M#NoFgNLbDoSm&IK|GI2Xw%;#?wrrWKA~CZmjVg^UW$RWhnL
zZ;-KpbB&A|&UG^CI5)^>;M^pmiF1pL*5TYW5qZD1e_arr2t~d8Z_)F!&wl*wf86q(
kc-^U+ZoK()*NvU4hOVxg4*&Di;eY=BH)>s7rw>>E7YiSZ+5i9m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_abd.hpp.ED9472BD269F359D.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_abd.hpp.ED9472BD269F359D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d1baa52097344ba3a47da9d6fca2b8120fb99913
GIT binary patch
literal 2508
zcmYjT2~<;88h&{JVnPCWJOd;IY6T*z0xpmQ2#`Pw0kJMPSg{4cF&F|QAq0?UYhg-t
z1U;lu1S~p=wN=WZAd8L#L~ILq>`=6dWpf!pWMmZ4j&n)RH81Dn&-?Co|L@)VZzmgM
zQBhq~1QDmm^y#VEBnt#VF!0rA)e0{%f>Z)dU)d3#e&ALZ^XbFatmok)WkVg+)8U2V
z8$Mbeca_XCejj^!N-bM6sXA{zb47Std2DL66q9(Q{AV(FWq<H(Re<2r>6Px8Hg#pn
zw4X*|;<q&HI#|kTFH7s*Y>>$KoaM9vixX#^$Te+W99h9Ukx_>R+5Yac@7JaK`bVY6
z%@kX1Nl(b$!sPl@%;toQ^ub?|%S`|3ZToe*dW*f~)H-*gQ6l4!zdqG_O&jc>&8y9S
zylRvkOsT#xAC%UOzqYwsm|)#??u>oVr=ISq{EHsF@^yEHFaE1)&~2lpTOJfAdG*bM
zsAp*3F?A7boZBQ;e~^5tcYDZF&zr)_yy9oq4kc&<d)f!=9&8}h^4>J1k7XtIyt>Q3
zB(1;{x~j5WJ<&}&<^oQuX2cKUkN*B+Ov{MF6He7YrRwUyAC95wb5~bof3oDk$6ucD
z7+Ali5c73>b)z}8^2L=i_%yxIw`_K`&gu@Y<?#Hv;Z3*D$6u8j2j*&`x~CqLAo-5t
zRxf)T$Ga)Hf9=(l%I1>J`zYv5u9sSlQx9n>HEOTP<C_poO;i2Oty^{nI6l|K9jv}L
zznU6p9P9sOuI)t*(rl<IO7H#RXv^&?`^)z`>nWl~BR|_W&ULyR&wJ)tHy8CyqJGkN
zKS4hlIFyk4?+Zf-<(2xxBJOkpZ~KqO8iR!@z2VA7+9K8S<y%w!IgFeaK6!t}rD5jQ
z{?qxNzifZWxa+wFSx5Okyz|W{WiZ39xBpRdz^h-{UL-FrC_52wX4Ydkin_g+e(%Gx
z1+tE96Vvx<nBBdPLaubJ^{u#g(lud{(b2d3gY+eWl9Segy{8ysp)uRa6N_r<lUo0H
zTbQJ3``gGT6Xd1EA_vQ&gY~C;4*Ra|zhuAVyDz;rdDiLc2O}B}RvxbCr^U$f3N7)?
z@vHa^lSdM#`|Ptj1*E4(`dY5dG@O4M*BlwBs<_$yc#pC6SfgN~@r~#@Gc}Q(R#Q?Q
z@z#};?wd3i^QbJe&~Ignu;M$$M)}3-4f#<wB2Hck<fleQcXH0nM>ogqnD^g07CrOd
zwY%CxCt{f!OYiK@o28Iti6TvIdU65@K|Jf8Fl$@8eqf>Ywg?u^314b5$Rf~H;3K{s
zALQd{$YG&$Kueq@%S;B+u^IV@sEc!fiBfI<bifo|iV#T0?O2=KG-Of1L}`#Vcp8L|
z-ttp*`?tgXlh&ve5Raxs<0b>?oujjhZ=`t(tx*b~V`wqB$v}F~)Ls$)Y4BMV$^vv4
zCv2(7K)QTGiWj4?MqrIP*dQ3wifLmq2-OnI350}^deCY6(t^Yo7Rmtrk<O9JOa}Vj
zxz7q8pWpL@g<1nz>?~epGLU{PY28d7G?J~*ML<5363R0fgld6;r~-LJow`%g3I-Iz
z@AHZ@qyKLtIws%=e4wNAg}I5p|91Q;O2_03IS0mqj>R%!Iq>Q*J7k(EE6|sYNd%Gz
zm>b~Y6zd9Id{f>Coza3HNEVpXS^6Gy3G5mh_KAD{AK-OjFEJn94k&p^_)rP{qBK*M
zp)YtrNURjT0Op)THRT#^g1H9E3ps|%@LoW2A;EB&j-@S1!=O}RGhexl=jIARq~Iz%
z;Bx9AOZR<tnh3BQ>K7Ueec?fX<U&!vWjeNd(e8JW%8t(t{jYhJ4}AF!e0R7>-#Id#
zXL0jq3`E^az5<>O+|Vm_1r(-ZN?hp(PeI32xQc)>@C*XV#4`yf3(q1THLfNi7qzQd
zQ-f;=oW?~%;Iz1wz-8mv1f;`t1eAm45Re|%6HqRm`wqcg0d=&{C@>nJ76jr5b1lf7
z<gPFz2cpTJO_jV2vIU!1Vy=0UgeBp^LZM@ktVk~GBoHaKYKi%d8SD%K%4BB}P!>Cj
zfYfekZ?m7qO+!FhwwACno1INSI<}60a@aWpq-X01D3_i44#CktqM1g5%uRRTaI4tx
zRk6!o>K}i53zQ7=JthzC`g&_=t&MDx8yQp*kMD1eU6^~A9}FxV3wICqgY9~5rC@z`
z)yGGHl7VYj23rHh$&zN8G-f5Kos?@vN}kfs91OKcEi)sHNb?TClz={17*A#ad5ByJ
z(>UR0IZLs1H~{)yfhnBKN(Ri}1P=yXK|mHJ84QRzRhhK#?+W%I0{0;r^$v}2Vc;_t
g%%303d+5W5eVjFp7K@p7II<Y7BM5GOv@<{cA0o(-hX4Qo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp.52A5FB3B86490259.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp.52A5FB3B86490259.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cb3f51f4ffa97cb56f8fb997055d92d225db8048
GIT binary patch
literal 20290
zcmc(GcT|)|_dm=$O?X&=Wnq`z5o85KL6F!hmQ}3SHEJ|!Y-mhk6oY8&9g!kM5fH3^
zib@m_K|xWmAlQ|nsEASN`WxIGX7bPPJ>Qda-ZzJw`+U52XYMoQ&b@c;6EJwd0AH1s
zR^OR}S1gznG@a4X(lS8*FAZ8UcOj>xb<SK%YeiIU;KWg7-HbkaUg*>Q(cb8E$94%0
z#c#XV1YTo6+H~XP-7U7v=(xf0RQZmX;O(2$PrffOxw;|K$X~H8|A|lKuArb!OBOA^
znYeFS(AJYNHd(VR`>*#qThTjSzD<8yvh}N~>o+73j*AOJcV=0v?ZJw&o~xIfaJ)2d
z|MWAnf84NB?!5Kvpeq-9gO_RB2PBWbC*OLY+sHw;1}#}Le(hGrz3cOxHpcpx*o^G@
zX~;F>AwyQLT(<X1V7ym-Uf37OkL{;V^cy+VQ5ke~;+$q#i}v9McJtzLj#d4fe&|iX
znb4!R^DORoxc2y4FQsUH@rRh(XXeS<H;nzmrfm3!g}E<QuK1Mo==&Y2g~$8sve8W}
zz56xeW=c);{<B-=_Up5;(dnh}?1HC1vF*hb$NpG<`{I+a_4&bddLL|7d;t49XLk>I
zY<S+=)vw6VVsw1YtIqq1;Ez|!&Q(00KeB9mjMAg7r_LPH-{v{K3N$F5P!KdQE$mrz
z&a<SzngvCJV*ES~d3<_0wB+<qw}N9w46h%5(mj6q_%vthBOi|Xey|_f|HabYm7{+t
zNVeUyvaV)GLhp-NcX$6@!14)CKSpgFnv^r^bbJOo=Fcq)dM6rv%v+$hEM>0loT7;L
zKdy*>cKTkB!}+Omr#5I0>U1?@$}f2(KTn>?jXQNwx3{fN;`15K9ctT$^mB0;J|}&|
zkPS^QCa>IB(ACN>|BRc~<eQ$)M#V|zp1%5%i=VLAB_VE-&*Hvy3tu~Tvi>b&c4FYt
z*E@oRl1#5#jgPpNxY1o(mJ~e7inzYPZJ@XR_EAc&ovRjI=*~aA(XP{tbLYNVzZmxO
zU-M3uRSdT~U$eNfcUpzl&+p!PRV`fn!-?{&<<}1_s_eHZMXJ0OX;AilN63WCb8&ui
z$DXR*nen*PCVF<CNfmzkUs{dzuNe~<Sedf$ug=X?D@$&NOh|gXN}PFojnS~Y-~ncX
z9{f1;urmDCG%@}0#MB3;gAXhyEc@|7X2fqV^&TAvU3Gq8`b`Uq1EKHU$o8ChV<;Pa
zWn=G??Y4N^bnxU(G;IG_EAl4TT<D%Y^!<95mYl-E_&al^@3s5>{+7@xN%VxVUFZIE
ztbEzqlN%$o8~$u;aZlr%9*s%cGG|5bu6HTL!y?K$22DPfdVSB+8;8#)pUSW-^Z9bn
ze|};8P1E@|ZuDJI=CSc%yUN--69am-OG?_`!?(pQyV+*Xls%5z@{M_oucFd~u$qs5
zuL)|LvE!gap<;LGEKyOy$T`QH*qQmpj}(*L7x`|mXnFmM{R+cTJv#=?``V?Q(I!}G
zC3<u-H+fvNr|6>p*iBhQiD8}hcRMgJe5-3g`k!IHclgazu5xLKa><Sl-4&5ArpMjI
zTYrvzacb$q#aqwa@YvlbtA8=RqUO#Wx5>Zyk4^VuvY&qpFFugJXo&Ob4pEUaLL#2o
zd#8F$?L545c9OQ+*g2aw{Sv$Fqua@GuY$TIt;<=tLJ~Rd#?-+t^@lDymg9HjefmX%
z&+ca!{+CYUW0}(pqeI*;t$MxWWNEHV^MbNPe~v5J7uf6>J3;&Imfj1eIW!f<?$iE_
zSzYF3KI=`OJji{FenK~U_t45w^ZXayt$uY(wy0su;A53PXqQ$$vRgIuW>3#RgYHE`
zBM16lJG@DMtL)~D%AJ2)s8})J>0|9F(9-V*?S9hFf!l{UFF50uDvmaHvtEDv+BItv
zo3(LnY59TAx_aLz>oa7hQCC~1!*-1a`<DD=*}q|6yKX^cGlR+|rT38i{&471t<t5F
zf4<uHbti|~4KwfmRCxcX%y7rkv7KKXdRM-)(7IFh69c_T#=rD_KI;C&H3_5Topx%E
zI(@XqAah;I3xQW^mlk_|@f@Ea=j}I~FZUd_J?~7+eP1Ed_j0c8@X)<og8Qz1tSDfX
zn=IaGl{Igra#Lkvi}sM}S<hJG9qXU<F)>UX_50fw?#CR$3!6>G4;b5j%F4f8H;3@Q
zXg%rW^XXux#0yKN&$Yc~GQQWK*!hp2hSz`D^s-MU-IF!4RTV4eUM>1^u&(E$^Q*4*
zev+|uZ(yjU_Gqt#r_IL%ZvVOdhY9^|4eaH-_VDSoKi{9QqIP`dj8~ofYBHYt1vg!-
zUY;k5>gc^A{(0jO&)}&hGnZENf3<GVd&ZEPEO~q|CvilFw?$4P`Yo!H7XAi83QuR>
zUi~mTbk5P+=IhF~as3*<W`Z3Lem^l)<hJhlC~+Nd-}24n@R#>H%WoWxo2vJB!S26%
zW_JDkSFNSLESRxDM@x%mrsw?n=VqxPu<fR$dIj`ffesp?C7|;O{nxJU?67Q`ow*Xw
z%STHU$@Y(ACuj&3F-RL^Jt2a5D*;`R$VS#=BkQaoSVa2C|CIBWv6T|g8Hwy;mHXIU
z8iGYkj-I^`anxh2F%XbQCKNKEzJ_2C^U~?ZFYLH>fdc|k_cYW!4W$}_MJxiQl-Fbw
z9hZRVsFh|$znL-BkT#Lm%HQX8?w%(B9g)Zm)^Z2iQA4nZ<(#isgU2eT7y=vAQ2}5I
zfSHD15hs_A{X_ds@|A%8sFh&WIGA<RkT#Ku<OvH7HqSH!R;ZQty#9OMTtl!(haJAD
z%L?uV$bcyli36r_z+6MHi1Wmtyr{&_vn9YCb+nVU+=<jx6D;E5Ymz;J9Wz!JaHy3S
zs1pN?H3W;exu||f%5v0|fZnK;uZ+o8rh|rHkuI{uJvQDmNR<FDB(jUO-o?6UNSjE@
zX04r*xmgl03yCx^(gwy{L$FBKBVgq(mGa5Dz!2$j6zUv>T{Q%Yc%98Zrkba-K?3@s
zR-#yC6x&-vut>KZH`XNZ8`kIp7hm*7oFs-b&=4%rE3^B$(;X7l>jD;aR1b9Ofr*A-
zkv{u>4SGEG@OBAMqK@{m%Dt!~HNheS=8nG}cWY>@x}#)PlFS;Y)C7wRntjA<q1n9I
zGGK-}iU+3gXpYnbiww@VljS_Ywu1yrMjgem&T*`-hF}r@37HNBlhU3@Kp!NsjuqFj
z-8H05WZkn{@7*VT&;{*KM@dj830i0f7E!r;?lAlEYF}MofJB;rP7{!82o?!gd#bo)
z`tMt0U>Xuh1g43=Nkg#6MArc81uecS^no+drJB*JW(+k1i%eSOu({jj=0g(jJrb#7
zWOa;{hG3D&p@$?>c1^I+1$wBXG^mpX<r;!Trh5E3`oy{rUwvSKM5_6A)o2H(2^N{N
zWWQmR^3!n{a6}@Jz#<Z<t0q|F$NS%(`xJ3#odkFwUBX#qI5Hk;f<@*Zj40cG{fV6f
z%tRvfjIy5b){r)ll3j6SzlEQcfexsnXkZbII#LrX64cl${h;kJjsp^JE!BF!uLqtQ
zqSX!pZ7nTrmG<(c7(cI5O_egx1GRGym>vY3H3YZw^Td5k*9?36>jEB$?1MV{keyN!
zEVA-?C%2Ag7JKW0_DCcQ=!5}74Z$M8p~ubFn&*c~KzH;`TUl`{+d)IHNXUw%{PV!g
zt91c`T1kL9324Qr2^LwmG5y}KvcgJTAR>`Ss1u2FQ4=h(VS(Miz76+&lYyzN>H^b!
zz*a-B$fjvAtCf1`BMgB(5{VG>BLp)I!6KW_Y#4DG>_;14ibR@$PBYq%YJx?!+AKcL
zzHINV4;+xl6~X3;;He>4WZRF4M~+Tgwnzr1Adzrj8jf^P6D+cQW?=hyI}g7y1a?TI
z6fmX8IIBsU$kTSOqcb<uN<dFEbP;U-2y|$v2^QI1)=45A`|?j+V1!x;hC0E}L_@Gh
zWaJIu)Q=7eWI%yLVu5Ka+W2aMMPfr`-tu48plQ@aB5_bB4q09`!6FCm_^3LiRiM@D
zhjfWyonzRZ8iGZVgEwVZ9bLI!7wDr_QlL%>^wJP4l5&VmxBM(wrwe3AWDV3=gVa?M
zEOMgi=NAsY77x$`0<xKjtR#^&)(|Xm(jtAp#qEPfNWcK3%X+rUdUl3}w22I5o;w*Y
zSStZtkjNLt_zUBxAy_0`d->VZ_cP`i0!ySzDqvF41gJ@y$bxt)*vPi%0v#mM2y`0J
zVo(z-a%$4?v0M4`Jtd$M>L`RY4`DlN2o|}ZmDRsvX4O;)n1Mt-GqTUf2&oAcxfoI2
zW7WLm&Jyq)64}o>?`OZ$kT#K$o2MJ|o2ofrj&w<2brV=a4Z$LL&b!a(`@7jofG28Y
zEo;6O>7piBq|DhXC($q^RR)|;D^b8A3YiZz!6J{o$lkgB-8yvrP%F`_G8!E>YJx>x
zIDfeK-tCp1E-*o@?1DPG(0Wl5Eb?-o$%w}p3$Dn39ErpL(-^es)C7xE$GN+i#jTO(
z1M5~@1luT~tA=2akGF&N6)r5vH2h|H4{`d3I6IY^V3Ec?%EV2Bj6L;%F>0ki&@T`y
zGz5!$u~@%+{-=?K5-<t9Q7r2m%lc?Yn@Hy8v`A|hF9EJdB#c#tp_f(@Eb?{b_$7<&
zU6)IM8xm<@j9bu>QWGqqwPe=prTKnZT0nXK{IL8v3EHBnnNv1%9{5GzlHa=f<W%$o
zqUtEu`6%au8KUYG*ZCB|*rx?L8c!BmbrBtuo=so~K0~o*i(32uB_E(2_CR$s{w;Pl
z_1yoA_;HU%nJB9M0FHkESJGk(u!;c=c*;a%9js!312!BYa^<YzfCKjOMC3MD#RCWI
zQHZLKK>P>{vC5(<ml=}FjKqqFsyt>$9*whD|15UI^p4l&22Bx)MAZe>{{oFm_#q|y
zNHR2~{E$+D^?oq*o||HRQJYcS6U2Lh1TTU1@;iOpiyzP98C8&1koVLSC%e)OH=7vM
zYTwn}2@2Kr4b`UbCT-a!ZDYdUuWh_v+l1g4ZQ~dk$7##rh`hm=+jABj6py3v-@}jE
z!v_+zb9hw_KL%qQUofJo3QDS=H5s4ljN5glH^Gk?x5otQXrd52k<zRr-3<ZH>d@la
z1I_lpo_IUr@C2#y9!lOr8{DFfCP=|8;D6WF(S#;#j7z~`X^+EO1V;6iad^u(VIQj%
z`@Chk6a0>Gct`W!F+T5T{(Gj|d&ZZ@zh_3gX95VWWgKeh-`6rewdD8m`zM6v#M7M_
zMs<Qke<QGm%c#;>pLEuj;CRj?o|9u&8!5`$z2)t(&!%YaceU4{i@9j9X8X(@glE;X
zICrtuYrk3l^B2V<&+BMHK(Z^<!6OzB(IIG;$aci%qloOF%OS2aUS1-4aqqW$PyFJd
zDpY76DwsHF!Z%TMLa;s|bih{E{CPvgeC?4#(7>c|MroWoZV1PoXcgUtlG{*@SBs7&
z0>v6xWYm{$DT~QwMdWc<1w&WtjELw+w^{>T@o6urc0;}0&=jvzQ56C8BA_XDFhx}q
z)Qf_q*vYfdgtkb7*HH2rnh_1uaW2*X2e?p26V{T!YXqK+z&~AcfQza}(D@PUg&VSZ
z<#eju@~1t~%MZ~Sf>H?T|E$;b9c+7~{3rC!uQ=r^&J&M`ZO`<~58ZC+{Oje2<847y
zJ%y5|&=PCZY7_=n!9aow8Mi{FC&9NE^IME7!H*cXM>PLAWB#1x*D@VynQnN)>1d*Q
zJal+jAQPYt>G9BED~}Eybs&!!BHDS=(%Bh!=i!JR%~Ca#RMWGhj_Yw(IIc&tqz>vy
zi=P?cGh>7uBvF;a%5zwId@yvZo}wQxu-7Wo!yeva5ATJi0}1lV9KLPscT&w9*g56r
zO^{I?;Jpv<ZKo=uI>>t;r14R{`%(To+)$_W1BPyOPPRrp?B+do6G2gRiC13YyJA6{
zctG>^1WKMjD>4t~8U6E2o2eC51&mt()3(w?)m_H?E~6y!4;i<IG{1r|ub}zwneOkI
zp=93FSq?myIK6>1egh?MpapJNN0a1e6I?d##@7Wu#Gtj@$c<{`#^G1#qRD=s7H&bw
zEoeeoP^Ut03ta=dns^QjXhtohuofw-6P|gT9AR`cDG@TE>setv>xmt7QFWcQz0MB9
z6Dq3eSlc>+yT0*yF#qcN5vYe@%rF_=hB&i=H2fP%{)V<hLv@-3Yv_IX#bZ~M%|Rp>
z30xw94-UhTMFB2Rzz4g#_;f(4U7dxwsPU$K?sZ+PXW!=`6jfVTVGHYs!&OmL#LA0U
zFQR$_D{mm!*RuB7JZ^feQR~TQ{Fag931d|67~^*|e$RAx&$tkNEn{3u^KS~`P2$8b
zs#}71i^jJF#cjb1J2{N1SWpzx_>Q2sL*u)G;x3Kv3B&FQD)RgLg2{bBL2!v+QbOYg
zg4qMXg78ZPvr-zD3GQV=XTpCdxId)vBf;m9(4Fuf3(CiW8^KQn@hN#fKfjN<C#5Z_
z)>X#orf{Zso#M<GG8^i&7+GbBP)vjhoHh`V7cR6gGVGw^d>L95@1W!zy?oSJGrWBI
z8rY3|H9}^Mx=sK>0<gqMk$&B0{>5KilZP0MoU)N?Gd%rveXJU}MEV#pDmnDm=G^tW
z8uog0qn#J(b1&zxm-E7V1+7fZA(G&J1-<^f+`Ee(YGIbstS;Ce>2LJbAWjN@{noIu
zk{s*(zg~W5VG(c}F{(Ia6^SHp2WUyF6Mc9|W49bJ)agAuZaBS%bgYMxdSb(nrT>=e
z!{dUJe8^L+fRYOM9o~PUs+ch<X8IERk}-Nou#P4b*w&CH6^N!#od?9Kh$t)-*0N?e
zVii@NSldr*SG?Q&_g&a_Bjk-9QZ*R%3x<Pf*P`DVI2gMY!&KfT`TfI30tWfQz4pK<
z*vDX0VX#LSoP;qBmyjyzd?Qii2o#S%Q=GjJQ9fDt${1sdAK-B)!OLaiOEduu+{gy5
z&A-PPN@NAq=}FS!WguP#3Y=g-qXxucpupaWsCo{>=RkooK%?6A&ss1`e=F*tKxkJW
zw8vXzjKrt(!;5#rzjdu_;JV<gF~)b~;;e=5Ube0_e)L|x&9<Y^jqe)Ecf*2Xr|$_&
zjVtp-ExZt<F9c)U^!Sq*H@^$_uR)9oE}()Ni+?e`DcqytTu7JyY7B*KBU-ZRG%j91
zIDd;=%pxc$BEG1oie}`|OxwwTVp2vPPvawu{t?C$uT&jP))zO7v%YBZ)G1#wc^g?_
zBWsMKHBohgmET|;@ybMdhn3e8Jo&WG{7H`Xc4)7hh8CxxGwwk|CyhLYv&Tmd@{)MF
zYTk|b=TU-Dlwghh@hLAW2WyGT&Y?NI#K|skZLSAeKAdSW=Zj}(s-QU8^xKa9XzJpj
zK|Hj^**n~F>nMK9QR8=-7Vu$|Rto<at^E)vgg{4p(6$DUt(%5m^`^mi#wumI(%ml&
zP*+2lp>q7s-<NN9+U&aNZ}iVCoU(<u$@utfhjI_!Qs7vFR*(O?2sn^<{4LP31$MyG
z)0#2}$YmxGyoGhz!g`R?IEx*b#ZDx%P{eu{vAqdyV7(h?T*3#G@RRVE&9%E^WUVJj
z<V4jRojGl>zNngOG}jv2U{MvWEeY2)C3v5<e4n;G!4cY$2$~<MEsrF8oZv?@SOX<B
zboi}K^W&v~)BMQkxCtdU=~aD-(LY76TJ;@3+YhtWZvArlqYoPRHNa#IaKPg`Uvn#9
zq@%eDzznqNK+kG0nB-<Y0_jI!LGU@~b`JWGfzO76vS9$hIdD)8jZZM9Cm4IeuVGAU
z7<;^fTivLhTiC&5;v!g|2)6HbO)485hpbN|+ZSi8THP%47f4T}s)5Gac;junHJ(W1
z5%P|I^X@o&K}MPH*}xAbxRLMKNaGm6F-CC5y9}N5LeB(YFu{j}o`(otSnkmB*Wy2-
z(H!pP$L{7Qkd80$0hjnTJGrP^&EiqTK_+NL)d05|(4&2)(Ziir8Q&LGAxxJLrf*XC
z$;`Z;^mdD?Fs4fwjZZQjCz)P^pT>Bk(Re%GaXarDTT=#RJ`bqFGl}9c^;OPb`{kQ<
zob9^`H4J27pf|pQSuP(mNIziXY^3@w-eVW<O;pe3mDzk-FpIq$Wc$?jKDY}}bq0uM
zfIT_5ihx)I?8(9P7>JL7JwCYbZbl8MuZ2iMnb05;wpm|MRm%&tyaAb|XhDb;4Dc+h
z$a(hiSLW9B;UY@FJEa0wqQMd1bcDu7fzweM9|KOuXnY(v9jEaL;B<n<CxO#R8m9rL
zG#aM^r*s-;0H+KZ7XYUM;7UfM5I7amxD+^*0#`gjqN)rymC^V$kiG_Tvi+)owECae
zvKrVCegt%hfNprCTQemtQP2%%N<`H;Xnu}Z8c~%6&9i9y7c~Ei#uuRZ1sY$3<`-$4
z4b8LviJh_`m37F0<~cM!2Rh}@{9I_BOY^Tl^D8vYgXVcO&WGmtG`<SWuhRG$G`~jU
zFjg;&HT<GU*CYRj)!WM&VwWG8a8@szHN>V^M2SzmeXJqQjG}Xh)r(*ai8+d3<q@m}
z!I7+9Bx^{_QzR>oWGx7eV)dd}LmZ(ZH-y!TW)1P3iHI(3^!Brc_&x_Y0jypOYlv^p
zkblPN#UlMRHxekd&g#XnhS)W4HE+_tc{$lApLppf8h_@cpJ`mjOY3M{&r9oR+`vm4
z{)sIccstU5BQI^F`Hj3~Bh7E(rA;)ynU^-xxP_Ot(D(~4{X*lfy!0!Lg9T}@K)qj#
zAdL~^*bt!c5TtQ5ju)iyG(I3m5779aAU#Oq1VNhcPi&bW*pc=R3DQF}|BzsLh~_5>
z(nOk{BuJBJoGeI_X?$3a9;R`MAWflhsvu1z7~5(+9nCE{v6!{Mq!w6_jb99{ilGxZ
zXCFbUM>Kv4t)9~OJmYkpaU=4HtRj)MA=%<WR$j<jlXLbuE5A<TkF5M7jX$ySPc;6_
z%0JV%j+NIDjL&<NAA1EQub>N2<SR7#3f&3b%^2-wD4fLTC((J!WJW)k#+i(MCXFvJ
z;svJ7bJWq)5?~kUkFx<CH$FVQ3b{zb+=lhRE-TIzAQxC&SU?6ogcU+qGklMWF0)wM
zk0cUZdtmji!;*6Il%nbgXLf|MC#UKW&hH4<pWvgM*-@IG!I@?J6Z>V*@~1enQ#Ah+
z*W(m7khE9M$;vrP8h0z_dJ|m1$tq}m1?OAA^(FjDPF6|tD>?5<t{34~ak46!U&VP<
zaXktD6(@T|^BXu>!#}Z41J{ShH*&H@n%~HEZ{+$BeiJ8aqWMjncN48oGbd}N`OTbH
zGp$bxCu^bkyLiPe-j=k#i}%~bQ}eN#SL~+wyZIiwd1`$3@QOV&e-Gb%4^QbE#w)^T
zei-i?##7_Fmsjki`FnZ4y>$PvyhSYUNc79*W!XG6pV@r3Y@V9WOT6q7&A-I^UZVTU
z;bl2AKZp0u;i>t&%*!s*{L8%8WuBVPTwaz-_}GO&&YQZ9hU|#5(BLfXynWz>547{P
zUl8^a=MC50v`1dxx1t-;P!#Zu0)222wl$aI8w~r<xtvaGpcl^Nv|bZA?S^hRWz*{H
zm_<Q5tSLG?p<68U#`g)W@vmDv^v3Zox*3NKO^`~pM=&-Kj59Gkml%gjbgDg*b<Six
za3&4icd_!@#HOLUdsf~|<2$^~9o`ww=q7)-_U*Q>-6n~uafag*But#CGt+>e**dc+
zv`DtdobY~E{7!|RA=-)%Z3;&MStKyWL3L|RTNVY($#6ykSu`-mv7e|)0#-@D5$B~^
zafc+}Mq`g8(2d|!piBi_XkS}-gvLjK@+ggu0p&3o9|y|gG(G{8Cun>UC{NNj4JgxS
zoDP)fG|m9Z3>p^zWdZ0y``XGv8kYiPDd<A`+R8E-zXpoez?vMX)xf?QbR;+n+Gas&
zdHe-!|Dy2)XnTRi7oqJ%8fQagHta%{R}NI>&^Q;`=F;+4pzReJ=Rw;%8s|gXd>UVc
zwpVF<4ccC#aWf+{GwtzxEs6(OX&7rt`^M6}tSRjqOT$@H+BcT&V@+w_*fN53pnYM>
zNY;V&eWg*XDee18qghkh_m%EvO=;g(8pE2>zOOVE=|}m#(m2+X_`b(k!{e-iY{y(q
zKbJGX+YObGaK?F@1;P28aXyW&a>iF_e2p`{M&o;&{yn<A63)1U=0D(!AJDjzGcKiZ
zHD_4ODM<gpykRg;;eEXEK4kMXfjf$2d3h>tMQ|Q(m`BTh;uW8GYvPN4<`ti5T*oWw
zXk5=L>S^4-+c)qX$?qF^`$if!@rou|zL{4v)3}9Kw9xnqulPdaue{<bje`Y6ut4b>
zCn(}b0uz~IK@m^m1A^iJjSmWngEUSM>=T5Jv~O;Ih{lP6B9WF)5)?@^P8Jl&G(IdS
z4%0YAP^8c}RZyf7jEg~0x>Q{WiqEZ0*LN80QtW+#T*z{6L^-)R#kHa6$W&K^;$MiW
z<4`;f75HKlU6sQ@>2M6rGNY4^6@pnAzWv|)Av3$CuT}Z~`Mw&Lo+6#qb*ESZTyTmG
zVRe})X0+an`bL62<P-^por0kkeF*!mfqm#h*lahnBZqJVG>d?CBm)r(?P6g^`Vh8@
zhaKrdSkXjB?CTku_5VDC6_@Dy@=VqulXW1w{uV30MIX9Nth|ZFcX-(yzRe13TR(We
z&#XsR(PYI!gIH*d72EdUmu`Am>2F&D0pnpx90=gkg4NL!+meB*1va&y8#x}DfoC)D
z$CKV#u;Kial#qz3^Nh!NW)L?2*leQoTP>8-(sg+1l3g63>|Bubb=IxP-yy}4fRF@?
z@PUk{1T8OhAum?#KizG$rW}WbFcWY}anCcCi@p9j_!{+aU65ZF?D4J7p1Nt1|7efO
z5kz!ns7z*k$fzf?Lz3BX1jlmnSo-oJmh+0``rylqu&aX-i`T7kMK60oFghVvkxiU0
z808CAc%g;|%yiL9i}OVbHd~v^)|TMx<-QG;<4?RwDMmVGYg=V&J7UQQkM!i_As)vO
zqlz0|#f`%D7MCNUrqtETczB|sJ3e1|X_@zb^~!3$(ZOXWS0rpgBM`=Wgpo=SQFWPD
zUgp~>KKJ)|;%7N>h~EFa(kynu_C(3x(e~)pVGVFx13d8fp`;mb2nX)ihM=SwaM%aj
zv15<7Hj16qHR0rkb&OEQ7~xww^y;kqGO6E*^Or?Wx%^-dn#yucSx)YW;w`fEA3h9f
ztt=_#{EE53c#aSDEwkuzsWI|Dt}E<RAi>X2@)@>WNT33bZ?)-ow<T<TeR_sux&#UC
z;z#e|C*rQqy5mP=ldIE&vYHv4<<A_EPEDM$iQK^Bi(#}a)iv$p^+Q=9l<k7A%SBZo
zYg@<;CpnhStnFulv57?usY~8*L;rbyhcUy#$o@X!={q7fdfHy}P6@1e0_%jY{zY^u
z7+oQ|@5H2%&heYu8>0RGmN9tC$jCL#JI3H0jo&kN?@2i(vSExtE#aRCjZO+`iC%*S
z=bq61p5TEk3eGQ~wV^K6C#oc~@?_SUSd@6KV?5`HEeI|VLaJ3lNhS2db`;$RF-Ap9
zAA+AUMo$Sw*C5{t{_zB#)KvPTi<n1H@`$deD~J8d;TY0T9^;kA3?R6SaVw)M=-x3M
z-q96v>dyvnM>tMGud4n^08byf5(b|2z#kW^phXRxi^)lWixANd)!z<~A0CF{VW{xe
zlu3x_3kBf|Bg3zhzS`}Dwb>F+ROp<-S)_1Qc<p9X|C(X4@x$SNU5<Fl>ty(Ex84jw
zJzN*0*GYW<PKTi-_8dx{)B8*PjRPK)%&1;R!*3h7pq0_g4R0pq`m98_d1d^|VgGZx
zch1ko;N_p6x1iMIQf8@y1mCNKRfYtu)fU!jQ@91{x6nzDos9lYax;t0Jw`mi7?Z_R
z!-zGEG1*|-cwrlFh-d8l)OM<uS4O3wVG4t0VemWbU!Bjt8+fGq-{t>$Dsd4dDwVha
zB{%3x<&BJdBV8J?jgfDoaVn#qN?uJz^KA;A;S0l)xl<=R&m+|<IAsMXOu&I2S|gvJ
z<P%+IrT&5iw}5Q|GJNXqR>%g35S$`}PS_O}Rr!K-zTkwthgR&IFSuizlQiNdy#c~n
z)Z_o(?YR7T_LyT!F2*9A*8%%=peODT7aXA$UP8%BdR?nOl)*aZ-M&>&@cQG$|Eg1w
z_bMVdy2|U89;3P~h_^|RTx-#vSWM$Pf_R6<cLnh-jqeNMeHxbtVhN2O2;u`8mkMGj
zjmreFjK&WI@ga>L3F0FfKNiHtG=3t8PY5npl035M#e${(b7N4b-|v1?`!~M-b4700
z{r2Y{Jw~<vb~N{X%k(Ff;~mrc9gW{J<KHur2)~x;T}${yevT^xROVs-+JM{~_oKXi
zd9)SUGxM4G5*#qyy0>gxN9!xA4O$BwtZP6gQs__*%<6$NIrbWXStD@9$6f1(N@h*K
z8CTw+0|J;e17{LY#Xv<2bix4}y8DBQIOv4kP4s~S6b?aSe076j7APb`V{%%iK%W#i
z5c_AM>KGJ`L1Q`y7t*0I3Bp4dhY-dSM@h&MGY(-iKFQdgWZVfqjj>Im@pj&PJMV(?
zIkzhxpFjR6!yk>;CC>X2Hw4@I;vwypg!|2Ctr%?L&|h1T=v}urWocuk?c$L;2HL-+
zt8j7|qg>h~=P^ckG(OLEJI@Xvp5<R`x4&py!t0li+`gzP<@HMmE(yIa=a!k?`#<L!
zN<Kyf%n9GP{=aGkae^CFk*U8-#KybzLg2e7IFMWj>|`u=lDojx56uGd{#ixlxrO}>
zUrx4uXg2b~KdZ?47LjsvR7}M7ZJ==}A5coF$Z)0;9UdQ{<Rg9HslSiJhsTqw!4LFa
zPuqY5!}x$O5@fUnhr=)PZ87b$Zbg0TQsvtJdh<T3|LI6muQ8^m$GyDAUb^@vmsjSJ
z;vejHDA4x;-@e-<gBuHdVqt%>9pj--JnT<zN9#k-mEMll$<UQ-$5d#L3N7gEXnhR2
z(%aEG9lDb3_yQWf_@}#WSWe?AXj4Vs=3Hf@R~hQVi<^x6CS8ekhmqf*aVaA$rRAS7
z@@F)^oROE)xSlbnXDo1W9@^Ec{1ORfQHdgJlglc}9(=>f-_ZOT)~1H$@8R0-;mk+?
zRK`oo_%<INRpj#07X-hQC%A5}Xx|AP2=`g}eb$`(u!NPD(0C0eU&B%SKREdx1mn**
zITQ+i`|Os?a4KV#%Cvnkf^M`Kvm-Q4XS=4e-lXHToXc9ymEcg$C6wTDclIOS-lR1J
z5+$^Y5~L&vbVX=)MUY~*t<p8~h^IwLJQBRan%$vq+`h18UkJtxpcIe#b6~9chp5a8
zZpUrx|Mfn>HX8My{)`y+fCE~@P=7;=8BP7pj)C{~KxEDT1bTmx&t{vN4qb42820f$
zmY(Aa5A^zVP*O*hI7(!F`xY6S4SaKh1m8l*TiU*;KSjoZ_)7;A^A<x%F=-JkD%k%8
z9rNZgUb%G4TgteV(lPH_ro&r0=2d^tj5}H~YyR|p$P?(7s{7oiX7xC29KPeSXDuv=
za}rwE*~jCig$Hu!lJA!`wR#o)*tO8o8cH6=kjFss7(gBelE=~HaRPZ9M;^zvJpxr#
zb=G8^%(!Z!apMCgOdRV!Bw*;6VZ%p`9yMa%;6YlW+9b7xlgE+daWHutL>?!S$6@4g
N1bG}!9*2;}{|7_~apnL3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_d.hpp.79D64EC0672BD2D2.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_d.hpp.79D64EC0672BD2D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d26b8e93e54eef8f282ac92cbbcecce91920372
GIT binary patch
literal 874
zcmWIYbaP8$W?*nm@vO*AElFfyU|<jg;^LB`%mO9`h8gS(3>9;ZhWhs3G7xxrX8nxA
zauJ-WslR2{hBfg<hh~*)moAOda`m%e<W}GIXQRm4Z~L$RY>?n^-Mht7;LoxAXJ<br
zusQ#jsnup#6LiRHNrCn$^@D2@7qwU~yykLo;q4Hv<_|kkq_zZ_8HXwEV0B$-pnOcw
zU8b0MXTXF>IiVj-QWi}(vSCVLn%Q*5n#Ud{O;;^io^^igOuZqtKI{J4m*J;*C(A6~
zY_TFs+EYaHfnD0s?^?1C3-Y@~L_1_kx&y)zR?f|mEt&FCe3Rg{1J8qY`QOuVYsn7e
zR-J4(VYLE>p6$6cPyaY<i}*Gn$lZ(Ib7NlQ#yrV=J(bcm8<t&uc)F)kv0u7#ZNRH(
z33qN;x@}=go-2Kz$iO?NMn9@M;g$CNrI}m*yUx2AXS~~`aC-9kg?HPwEnPS1`qep$
zt%WzINy)2d1)aYlvZ2{Ar*MC7+`TtC@6G0YH<GQJ_nPgh)SBq5{pGd?JCgk71P7J}
zC7+miX_?vPvQVq$iDAZuJN8w~b6K<GhqmFrEBksLKKiM(Drnl018+;NeR$0`ukyoe
zzI%NUz5CyO*vY>C^Q8Q#*BZaa{n~qgegA>DCyT?A^bIPXhQ2J`bJO&r*SvdAZx%lK
z6PL$!-8QZ7;RSd8nX-z}-)%DLS}K-T=Uk85Bv){+`RlF6>h{9+iJkiwOW&D&aK7H<
ziV15Mo!DU=v?ArZaEDS<@uG|g#(jYjZQg+iJFc_{ot@{Td2NMIkkG!1KJ8aq^-Mf1
z)wDxYKTVh_xn1ONj;Nd(zxbs#-+$kx>PY2!K0BiNbC%j|ZVg3E-f2Pq@+JQ}v1qHj
z4}J8nbmuzOKWs^?2fd3cbCVQ->CgU_L+v^jk#ue*7FGs!BQ_&tFu}md%*4pRz{t+X
zP?VZhTntpnYtS((YpUUGMjm!OH9Z}eMi|Y=1=i@PQr10jmvddt-j=?9+rL->RqKlD
zs=`z)h?=mR&oe{?B&?=u2owJ8vMZGN+Eiwsu#uXPD@+(@cTOVEYKG#X<YHwY&BSi;
YPW(B~q!@E1HVy_hkQ4-Pz!(e+08_JW6#xJL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r.hpp.4DCB681EEBE1DA60.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r.hpp.4DCB681EEBE1DA60.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cc53f0da458030bf660b2de7499762770647bc3d
GIT binary patch
literal 1706
zcmWIYbaPw8#=zj5;#rZKT9U}Zz`!5~#Kk2=nN3U#3}1k96?3)*#roei5PAE#u2p`<
z6~BdkH_DGpxTSVr%fV|rnX}bIW<I$TQJR!&{Z`;s{rt~?uN<_#8EFQsJ8${@-s5Up
zi#xkN*F-J(V14a{&4QoijwY&a`kTMK<EyFFUZ7yP;ADEyY46y>`rAKTX^PK%sIcY3
z^=O+-56r_pHe?hf`k8*brgdNOUL<cwd*%MrLwpZ+hJDWKdmhQ=#AYc{<tthvsBo5}
z;^K-ml0N0D*GxRqRIvTrtcx*kr`=S3JzbWUuX@|F1?}9MC%jZn*{sIhC8zlP^hWQm
z3Ad+4B;06rtvz&)_lE5J2jxC@gT9>K{dl#1feA-~tFXnyJt7m%Rcy41Tb{0ZImCd&
z%+Vx+bwyA?SlG@NGftaMT=}0zF{&d-dH>(bb9Bxut)FZ0;*O&+``jS*+U3*xmc}s0
zrbzi}UtAvb<+7u_)uaQt+7kb!%;x%Y$+7V|bKmjg$?0adRkk@ME7vU9zVza<<F#u~
zRCkzMO@G0xWGH_4B*$Kd@K6(3NgI=jNlv9^(>d;(N}3xyvH1E1p-V=L<rT>$?+j#m
zWlX-mD?EQ}c39xt(8a~&A(2_<?-jh;x~elMmGA1CwbMF8RvJ#3D5@}*G3e^E$iL6n
z(^EKwcWSIQ|1kBkQ}ddQ{b>umCrH28bEcVFLuCEo6}*#|zJG0HqIYZi<KLNo%WD2}
z{<=SJhGwy^<PV9*J<l7AU4s_A;NsrLCONrjL$s@?UxnTdrEs34<lPKynG-h{vCh%?
z=qy~J=ym5uM*T<8K!bBFKhEFH(7Unh>7t!gb7t&5e*Lw{x!c~y!!)dy?YIz;RxwfK
z_pIV&4)3ekBZ}4)u6%Os$miyjKQ}xKn~<@2<^HT^$ybUibCc!*QzWl^;j4ZJr$kXE
z4lV{BZCPz~Fu}k9lmh|*ccUYZ^Tn34F)?#8@My4W=z<9bP9{cRnr7r-l>8GT|MZcH
zgeVg`P`{R}mO7Y#=~sQY_WB+Fvr+;~Y+Mi%dE7PKtzpuAtHYd=XWZ5iU}A+y>uTxR
z!lW6BQqzi=nGvpJ;sU!4<|-Z@BPk;lxFJxtGID|4=&4fHJ#m+F9n6I+Tws?W^ctuc
z7{K)M@R-}1dm?xi_7-p+%nyv5V7J2zh}~|P6Qic0%*ex|E2FCc)0^7Is@fn_>;x3n
zFxPN~sjfKu@JSWR#j`>@JO;uB5-`<3AMhFpA$Z2Z#z=A|!X`*OQ(;pio|&*263<-N
z9GNF)u8c6#Lf8UH&O*)tNzPK(5=qWV*b0efEo_a%vk|sI;@Jw@BJu2m?T~o(!uCi!
z2Vn;|uiJ~c<!%3e3jrP;Em18QSj_P7Xp3sgz)~U)kFlh&0>b$wk|szzQ%O@Ko|&W>
z63<-4TpOX!Ld61!XDMlkq|!>#3W;YeX^q6Qk+ebL*-F|X@$4k+ka+fz_DDPjNe4I&
z<_chv0i|bHB3Nr|6>xNJ!gOHTaF=mchb6412CQG$`Ng}LfGJ5(R~n`ns7TOV8m0)C
z7jhDnkg_C$)ZL6l>o=-1!BQMNvlbU67b^o*GVusXU5hB~SZ>b5%gw+I$$ub{7r_Jo
DDm8jx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp.CF5C02CB8B4910C1.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp.CF5C02CB8B4910C1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..33afdfd0402fd66a23e339ba59572a5857f9cfc4
GIT binary patch
literal 19906
zcmc(HcU)9Q_da{?%)M8*EM-|>=~AS*AgCY$V!>6ch&^f&HAYP|F}5ToF%}R3rHR<U
zf(0-lN<@UENf%KR6~%(mG$s-g8`k(6+#P1}&-eG&`@ViY=IoR6+?lyk&pC6hVIeA2
z!FU~=5ub&8w{*eExj;up#|Zshv2yuheNIP*v(nM|E~RqD%&*&g3>zo@UH(Ej>s9T7
zzNRPU`wGVXnbxP?s4f{V)k_ZEerWllwD(00ld40LQ~aT6ujdy=uC=;2uJQGYj_C{U
zMR@#RcXXiT4`=PuY7hQX!e@Dw6tK_DPsVe{1{S8p1x$0j?cQRv;f+(kUc0v^Djo<!
zhF`3na(ULvF|fL(sP~{@rm^!kPZ)c-X`U>5+K~S3sXjMT28>J%8{p;i+uIfQtE9hd
z>a=saFVEW?+fuPmqIa#xxoK4e=vOx)E22g*@XOxEFIH{)JTUmqzMPqz;BmoB{e(4Y
zqb&uy!-tRk-Er~!uOAHgB<$x}#=}N7WZaHB7w>I}{PuR7>^B3CKQot^8LhBuE8Lg)
zvHPTq?TZZVob}V&5ItVHc>Bk@E`1Zf+2d<4%d2CFY4q{u&sV;QXsCH`adpePIuIH0
z>6-Fjx2qctHe|UD@&G%I+&g3_diq-oR+fJrn!Oaf^(t;yRxL~~y%Y8(@{f(}U$063
zwt2*Z<I&5!k60GU2PC=oc~-gi)yXO2W}jS^<&!9S-ui6JlfW+Nz6Aj$>y9_&F57i!
z^{DZy^;$~D&I_~eNUX5FQ?m2mnRPc04V<`ppqcb}=DKZP%CdgGxj!iRXxb-M`FD@s
zn4Q1;$$pUfg_oP+VERKVNoxPZyWf_)>Dy*K_)SqCwZ-5!E=65enu3-z9y|E!&sX<<
zGGgI3?QPNHCp;JzbKdeojBoAIm)qBW^lIg(?;A6Iyipag;&`u%zO(ZWx8};8es7y>
zVDv?H$fS-!dHIvfY4+Dcr#9|>+?cm}?T?p^o}L%JWp#X^#9>kKMw6vx>Z4~vua8)i
zTi^Oy4wrKEQ~x>g^S`HDT^&~V&beLxj%@MOuJQ9fndox0YfSlw`8xs@6%_~BoO-dc
zYGY&b-Wky`U(Fak@+*61^`#CE&|%_NWjiyyzqN(VQJv9x8?%33RNue6GVgL(U-fTS
zx597ZCLFdIu(+}7*wNLOv;I2&-LQRY_ReHO-mH<;Zdw@sY)a4<4aXh_Pwh<Ow{IxA
zvv7LW>Br`k_baV=`2APeuH!$xv_SgN^psJ03)!i)3)9}d{z|_-;z8`Lm9?Wbo^Z`S
zwP~KUGrVLWUzRXf*t7l%i(~%z&(}?{m^pv-1HVVJ{~9#!tB-2_ytDX`yXS@>iJtet
zcTRric3oKY+4xI+K5z7!?0a?9F8}(uEhCpz1;-EA8e#nLnbXE|!!oudE}U~JD<*c;
z<y?<7jWDT6DQk?FT`{R5V%}HZ9Cy8wbKs#*;CIoMW16b<yM~{+VPZK|T)3;{v+9<<
zl~1?sAGq2#d-u%H((1gvNz10qxcF;_zSZ=ZPZJ%!w23rz8t3a?=D>XNrLl3OPVb8P
z#7jC(@%O#;%c9RjITtM5rn7eSxIuQa<|*~xk6o4vryJA+mgzno+x~g4<`FNOyjuD_
z-#aAMW3KL#NMJI2kVpUFgKwQ3dRQ!7*ci;Sz87|#z1Qn+_lCdc|9JoVXI{;}9vhSL
z)Gk`?kib0tW&3ACTci8BZjlUiI^*D8A6e&c{J|e5E5l<f#|^7IXR%j4Y4O<>`*lgH
z7VMrdYkB=7WAjBVj>9Gmj+h$!$<I1{f1A4VotsX~H8_9O$r%GpewgWKw@!b_t>(Yh
zHCcwtaVzroZ|y48)z5tI61dp;+47Jb@yANPt@-hmo^{abU7btLhPg2-QXglInswJX
z3dF6w8+`c(b=S*L-(8#6dS%c`pT{P4V^{gQFN<G3b#u%_aimXO<k+9@Zm#q{eSK2O
zoTAR86+>n|{ATay`K#ZDk3Ur3b>-liVeJn*Kg}AexIFcKv(vC?6~aert9~l{adNxl
z=rM;?+uwy-b0*PEY*b?AgZ}2{K3n_xE@!Lobgg(8_gUx_WygDm+1Wa71-s`Qs|~OD
z=&t`#d(RX5O@{j2C^dY4+VVxw;r*>f)innKrU`81UsevfaY;&_7dO8tcDwkKxk2&j
zg46eAH=Q&&oOidt;pM`OOJJ7WfvvB<Ue-En+xno{8}iPCzHV&3XYbIiZ}Zt|(KEca
z-u;B_V{F_-TMNVHHQOAjW}UWA)NefTk?Ai#ns0D&pJva$PpGo$yX94tvHW3ke(h@m
zb=7AkIxD_kI`3vL9UaN0C24;pz0s304*hggc~F@LhiQe5fKCte$MBCTpZN|ydvlD;
zQ%6+=^{Sw;R$z+JxyXKvP9^6InciMHsy&=^4`-wmm?A&AYlM^8Z(r*%JSwG()h%Pq
zv;tF14X1m>tsj}C$3Rp{4V2bEQ?2Ns*v-5*|LuliDdU2gn-7)waHv*bidkLw{BsJ+
zRy{@+Rg}f(W--QEfhiVu7NupMbY?iFmxqojh2c|}!CIlyi;<vrmZ&6_Ig9U$$Lo|*
z#u-(0655=E?plG%v3VY6aeDaJYAIunC?}xJ2}IEdOtCNj-RLp*N3N6^j3`A=S%msm
zBQV9`t3HbkCybh}$4F6wt6AM@)>JEcC=Wf2!fiSxN*NziiW-ho!;iEAQ(WgdCfZfT
zJeM;zsFV|o^$EsVD|#sNbp5Q7vR)c9rifC{8`kqyT7fC<%WkWl#q=92XQm)Z0b@~s
zO3?^R@!(5do;k?;sK?kLN(HN1!OFD)Q#@bT$v$!4JWI+<M$IjQu4QnrR$z+P{(X18
zJ~!m8F=K^F$z~0+StqT)ltBF{S60kg_ezh^LzHq>w;atFjlh&4$_EE$O_>xWWqeUZ
zg>Ymc>KBc`l+n7Y+7{ZJ_LDMw5G5I!Cc{2jfhl7v^m%#pyZy$Dh$?!+7`|a#v;tFR
z**%)IX59I!#*6|{IvLQ(SZW2P%(?sV61&0cQl-osR8a;LGoYJRV9MM{PlsP=tIXBB
z^ez_cVihWlz?22eAH93^kyn2y6NXAT3ti7bKdr!&FZK_w-Pe?|QI7$rEv2k(De4!E
zz?7w7>NNEyBdrV>OGL>RdglxMwE|O?FC6VsKP>Hmlu@FJa-dxf^wbJWS#u!2xpe#p
zLn$*GQBFbIQ&6cDnDSHo_m2kuRJ%;dxFSjcR2HE1s1ca5DR<Urhsz@uNEs(YNrhIa
z&|51oWpnns)9a%<tK^I^YHluLk;_<V1*Sy#ZG7q)ayM4W^hc$XLf29lq!pO5<$6TF
zWn-QiNtr>2vJ;x_gw9%lDO*hg1M2c$`)~|I6-5HQNFdV+Oo>ea=N@mFYir2ZqEf=4
zemL6gGy+rNZ)e)y{%1KFqmK~fS7`ey>KBc`l*A*p8P6xA8cP{JL`j2U8uZtS9?BEh
z(Jw3XQ>2VPq7*}AG1?h40#mmASwCyZU%pGE%nU?HgtA0vr4^Wx{B7f{ZP#BdFl8X3
zL;)cR7-~fi<(OpH?&~jnq|5+BDS@sfFi<NnW#_Ef$<>oqerL$&qkbtC42uP8t-zGs
zojbp73F~UmV^~Ca$LhXgt+fJEQd$k3rq?|xlQMl#MZZ9^U!bd2V9Md4b7qVh)}S(E
zOi(Gce6L!x1!)AP9DP35YW3@YF;ZqKDkTmo;-I}&V9GJ)Q)bm;N-U*}H=<<0pe#5`
zD=_7hVtJQB?mf$xF-H|OGoTqQSdGAx-1E~@4_JAlvt$-3Wj8e0jVjUzOv&rj_mgR9
zFMg3S0f@2-n(jiIheq^J{#g-yA#5#L0d9!07h3H_(_14jrC@zqR8sk;QO1lMwWWjs
zC8#YLfhk2lW*rnJe|JUBSR+a{W1WrmU5&t$in6m8xL3DyrHnhOXa_Xif!d-Gm{R$t
zTJ%{oX}_E?K&9j`7CC4gX#}QJ2iFBmwzYgHX9gll9%GTmcxwfw)c71bXWstPNjYPL
zD5n^UQ>YY;z?9lHOM}mSHWo>lX{asnP!W&1P9rd-KKkPE1ik}JZ%0Hq1C?jc^wtPW
zx!UMfHeR0VX8d6?q;ZC6oP$auFr_h5s{Uiaqdj^IhuX3Z=&l3ST7fCI`-npR$GJ=O
z7y(hjfo?c3*9uJO__fw;;fs{sQf4}8O9E6TKwquEln3qK=-m1)`LUD<M3lo&JdBPk
zjlh&YUWYHt?)vJ2F=K%$+9()q6fCs@Q#y}-VmZ2AAIKR)MERAm_!V`XMqo<U<jY5@
zdolAhrcy1G)<PqdM)XjwE!`z;J)I<FW}>!igUW5Fxf+2fuhVV^Uq7|-r6FU6DvA*7
zBZPrkfho%ue7@rI0Ak@WqN*B5tARb<GIX_e9^Mog<@LhxHH!{QMAb`9`I7U(<HIih
z!QZd^eh@0*DCd2Y^TP^JmBV@G5X`9u+69a}UW7{6#L70Y9{6MxzwMg1xc==Cq{!zz
z^ZEWgCH&;)w}(^RMT&CXx}0~z3jKbYcYm?n>lQ$DFt$026Q0{5>QCEW87JJ;BAS}E
zrx+(ZYejT$*yb`$_)O7XXyLo~7pqbUP@NORbAlA_b^5s(B`$DT5foM4z+3!a>Z1RM
zDqrABp#Ts-p<ob9p%4(#gAB~C4)1rR`AZ(CawR^w5?|vfBU|Rq`RQk%DwOyX(zsaS
zUn~hA`VxtM360N6{Lj*OKkKod^~A;lP@Q2t&aj@?f-<-q_)q6}`%pAG+o0_>*tbhN
zeMD6Ubnk$J`oEu1ImG{?M1v<oe;Vaem4Jxp&?g;E!xoFEdICM4z=@b>ILf=C?Q7F?
zG$^Tzek#)++soJ!0J>WHA1g$(4xBd%9(aj}Dz#vz7M!si(v4l!LO+a+rZy~Ec1S#d
zni9uPjpJwGVHZ^y{G<$iI>y+skmzdN3ETwiPN1e}d<oK&8?11Hm1C<~RBZs_2B5$@
ziHP=0VF#4s?G!s5tghD8z%__yRhrjAJG@{;)n3khFV`P8$yDo#pm#V2q~}0yvaU3K
z30_yG?sG3zPf))`-MNk(xQ-3PtHP{oUGlHZJKqUFwFx?Hg3j3X=*E7Vpg+Np&?%DE
zM?$|yS|0@mMZv*@9|c3BU>L#C&?%aJJ{tN(ljpI^)Js?ED3Q+00Ja&x8NV7@jKDUB
z#<{>Ym&Rv+@fl!2_>Y16W6+=A8mOp&y>X|Dsz0E~AJ7WVJW=%snmi)de0azB>WE|`
z)OG)`9{;dm*a|Vfeo}awp6AU0)jn3bk2N5xBafBlu?Bc`0M&Ik@j48}qX<;Dpx-Sx
znBcvf*<Q{Zk19~r@eXyoBQ{hN1_9@W9x1IvEjS?YJ0Ka2m#xL@tc)bL_6?k<a&~Zb
z>3kJCdfV>3<3yFKgDZuG2t)KcxmnzVCzl*W)mUMy{>dfwhMg{)hp3tiCij~6bavmm
z)~)HHYATr8E8>jHst0;5Z5CH5ztQ<UV8Q=<X5fjy=xY7r5XaRq2?Zn8ifE?DvzXC%
z>R@jft%h<SEeDo($aS?oGhPi=t*`DckxUwkIy9Rz$>#durH-9xR6-e$meD0>Txwi`
zt#rL#lu*}#N@#$V4bTOTrR_uK)4f*T3qp#G5}l0_UA(8;Z5k0`pVjddQgm?24zB0;
z>0PjV(W7+%o6zCV08Ja90vBZ;bSnSJpi8>{@%e@01L&<U0O<w#bk;-(@LS{kOrfid
z4d9{aUK)1)uzm5;5LN4e-FncMY<S_oF&y;6c9N(n0LBHt1`m*^Dg?%bG%f|bOM#N;
zZvumxWH>}s3ovN`c0~Uam_G%M1iuI7?`fO>jS`?WHp04(DWh%BnjBO6p?p8=ja?>D
zbqE?Ag4Wm~>prbq(%}GdTGc_rI%tK53bhBCv_LzuhqOWKHt0%l2dbRLf5P5>LM74v
z1+D+0^-o~$C$v6+Gf3bpNqyzKK{;=U=dFV_HiJg+JdmCT4!9$9wGkRTf{tHA_5XWg
z#t<|Mt64=gI}H!1)8Q3C<AaVLLblF+=&&D7p}WHN5DX<<@h9~B6MjlJ-Tf~(8)MgD
zOUK=eEBpo35DvV<!6-cDI5>nhcui!8OrEX4VJkh@FM-WV;DX<wyYpwH!+E6hvta)$
z7>GSEQS}s>KZW+V1)}OX40;Yj2)@c&T;&~IwV@|b6)jjq3y#=+#s43j@0xfNZWs<k
zq4Ql6e!>b-b)6BfGYaffiK_dIc%M;VhYz)Y5uY;(9I<i#<730HQA5|F;fdqD<9J^@
zZr!GaM+PxPaa;=-0-B%}sqrQ&++<~VxJ2Y63cH~U`;Q)xs&TQdnRcj#Sl&C9A53aU
z=RMMSAB=HW3=tcEv;o-SUPd;|hZq^2DfsY0P0>WgNK-0U?+SJlZmXy|1FX&fFI<pU
z{<Gr~rya0EGie_b_CXsmlhUD(4sFOdc0%(`=t0KmZ)pBE!2?!JzxaOX!hL8az2GLl
z;Cc+QfscHemYn-56e%JkAVMO?d+cDFziYqb=Kd|~HeO6OL33h!L_*U@8b?8wDCkD?
z(a<!S)|U%nIkD4$szMMeXj~~MDg{ey#{*TBps1p8wV<e`agCs;q47Ck{5fGFdA?RK
zs}&RkpBK!|)3{EstP`w>{(@k6fyNhweisFAqOTYF)zkQrpu8k_5`BXpHjwuTkOo$K
zqH_Em^=KN~Cyfokrc6M<Grb(yyNO7V&grFd7I+@vz$jW9ns6wY9!;#!MBBbQp|BIm
zab!$Hu?t}rl;f}l4xge4SqG$b^wy;bqT&fTWUpt3Wl~B$YDxj`S-=m(O&MY2FzI2a
z+zu(uFv1ze2FDkI{XZ+_&#kXUiWi*n1=q6(g10^UZPIdCGg7p16Wh3{xTxTB$-hr|
z?cjqNvyF4w#`%y9cL(RRgW%xvLpN6p8xVj>*e`M3FByPMvXLg8Mh9f<nh!-Q+c<Ar
z^vL%Wx2>(i8jzx$Q??V&cGQ&=|2-d#T!D<Z&8*>O)&iU6qAH3tjG}QgYZy)AG<HlH
zI|Yvwa&TBzHQOJr+-@A8W>pyDC^T|5Gk`RMK5YxY$O14GuV7Ji6S&_5{seD=E}P&W
zyjpa%v1>dsxWQ<v&;+i@TsXuU9AceG3yWBTBGwtVXngSK-1wHwf1^Gr;ysJ_0l2fq
zPxLjb8M$*3Qe0yluCaaaz8B_o<gkxhbURW!=O#VpX5uF%_ZnHWbb(<kQWOiliUk8a
z)u&4RZa!)IbHaxwl+U@II;PFtKJ(atTZ6mn5M<W`Q~bmXS+&=v$IYFPLd}Y5)&)0r
zMo#Xfz@1|rBSkA0*2+!8mCtyy)9b!^<ADFV-;EC#v<fx%0i@^4S?_W-1TRw*kpQl_
zU?>^V$6(-NFb?B6yB~FqGJKwfUhN%Y_m1h0$8=7f-;&vV939YZ9to@?fh!(#5jn%=
zdpJkzv>;cNcc|q(vD1R~X2B#vu*0V;-t}2gRRW|Xpf9c(*^a>EAK;C#uJ*zLS2U+B
zd`9l)Unin9xsNs3$J&wEl*gLnv37WPK5l;yq9d+SBVsWpFXnnS2vJqfS=4hqXCf{D
z%~#FE2kEu{`)HfDR@v)jQhT>67%l<f5_5d;%<qVI-qJVbDN?-Rl&^?!kFQctn=S+C
zWwPdUwHGS*Ef=-9#&%7&@7{q6dCx+=XWv-tP-bGMFHJW@AtHNkkM~UH((IJanncwA
zhXI}h`P%z>^l{s0S~AS!m8kM_@be^Sh;)br@&ASf4h<nF*gn|9zHZu~lf!>FFRD}y
zDo=vOO2=9dPjP(U_z;2)N^}oOO!2OV>}83`A&DKqhb6j)Y5ft2$q`zgF40Y=^%)YA
z45G)^OK4teE|tg#U1x>s^y%`R72mT8;&>!OAsNc?UWbD{$hOnmK;eSE?r?fn{LO|`
zG%&IJj97j)Ud!m9<ENzaGcaB@=<zp`udFmf#QR|2eGt6lQ}3bHLn_yxp#}qA5sWXh
z5U~PCD}Xhb(3+bqJk+brM~^lPi~bxDxANXw`9ZiUQFVm(I6`*8)nkL}+-LZNqP4e<
z?Y)leS(M1VVtcQr@dmc{1{!Z<dv7H8yYiOyHQ@4vX=v>+&LvD=Vz`zv&ZRUy%Q&B<
zaT()WM&okExtzuojB^EzD;eiX8dovSRWz<<oU3VE!#LN_xS4ToX8ICC`Yz*qm&Top
zb0^am?|Y)^PsaIA8t-DAcCl`Fg(EYWb=pniY}PBA^}$<)s5-%VouF|Hu!<pJG&DlM
zbSv$5#saHYS|10j;%FQXtm0{$0IU*doCvHE|Bam!>Af%sSS8W=B;cGx>yv?1GOgbM
ztai|NC$QQ{<6XdN7marVtKBr-1FZJYI0aax&^Q%XrP4SL>gPdYoUwr{E~tM78sm!=
z6k3P+`Op~0C6TEH^$VaezM@4ZD%4j)W9+YXV{<jMCb$sl7eZrf6r#-#nioQAf{UPj
z5j4gT3$)Wi{bFdGsJ+KT^Bd}yK;vZXb)$%`XZ1^=F}}GK(XOk178>Ji8HIVEei<~z
zUJ-JNcv%B)j$J}gb(Pn@${Q1Wjh9`c^^Lr&k;d0~*>xJ<;AJ;x+{DY8{*7&#cn4Dc
zO<s1B*5Bl9ZqoW&yzCaOzs<{T)A$ZAyF=q<Ue-+GyS(f!jqmZYdo*t0Wi2#r<z=lj
z-Ym#A3+DJx?{=m2qXlDvV+2_Yt=}TZw$OO1AlpjgSV0y`<2XSU_it<yCpeJu;{{nf
zt&bOM;%R+?AWNY2iGnPV#@hthHX0`hvLqUB7i8OMoGi$aX}m*_?V#~aLAH}%oUMSi
z_9`H)LTgofHLk18S-@uiUJ!^_3#7GVg`rTxhXe*pT>IH||Emwemm}hScJzKWjBIUZ
z*wJU$Fuav*C<t?pzMZ!g84wW?&j`tIY@381{CUmzjLPZ(qAHWK%;X%&ag)gfWO5@3
zKFV1hrS&H{%ai}c0VipG4riG|>vOmvIov2xUJEC0;cRF;sD&Fwa4RQorS+}c;8tz~
z(YJB(Hd^1t`L=OGiN2kax6}G|&ZnIVBKi(a-a+f1bMoi^#(vMa;e`K!lfR(#FF5}f
zTrkmhaq=!&-^KZM(d~K3$zRg?mz>W_x;?Kr`72t#l~-)#?MeAt`M|Bbir`pY5lic1
z`Jh;SG||WLia1&y#|Omm)Og18ig;Qd&-=&o)cYjxiUeAJgqI)Tsqs3(`yJt_{z>QM
z>9jtb_fMzG%i!f1v_69$l)+Quoyp5HX?-T|o5@q-d6btQrS<u|Sw3$^-Z!5gluwse
zz?&7&`T~A%0Z+{zHE*V-^=f{Qntr~JH!Gy|g}iSe-JT-etcd7w5)oR8`9PXaqRYD4
zR3f})H{DHg9${&j^}&{(SjbPtZ-DQEku|DGFv8X-KA%w0mw@yVU9=|62p1hQh7IqG
zdq3z~QRU>|<VJ3M`Z)BVkh9Wx0MWTPxKKJ*r7NZL4)hMDyof;GVB$9f_yhz}o<d-7
zF!2pSd_n?=H!ulI(tG%GhC)4N_*qdE2Ez0%Reu~hKjnrJ-7rZf%gGJX6fi~4dR+9i
zMTeJvkM5hKQ|0)+NmM0Dj1wga;zR6{=<bsk5qv<Rdw|9#CGwLJE27Vr$n$BO&B(JE
zE1cGXY&}MPg0aGJXk>dZwrZvij`DV67d7KaV=px`h~N@NS;7pUZD(aEjn6X5votPa
zlw~w7XO!hMu3(fEG_GWnl{BtmlvOmYW|Y-5u3?llG;U^;&CCGWc2?e{aVMkfWCqZ-
zv+_?GzhQ>HVMbxk9|h`I_g$<vIdFEf?z?H6!b($EInnQBrF&_-kL|sWbs_p});F6S
zf)A{2+uQd9jdNIO4l5`6Q>^q9jZd@E(=^Uwd*_kp407aveJpUrM=pxB1N%4{#{>I#
z8Yckz1R5s-Wg-|ro=*bGBpN59n=zWd1K97N@lIgBlg7J%{Vp2s2KKvYya(9tp>YbZ
zPoZ%tuur9N9+c%l3)=RUoq-m#?JdiP7PRdxD}WZ*)abUgZPd_-wzX{vp%ZOe%Zi`{
zZClHVp#^PQ%SxaHZClGqp#|Q%x@~LOS!hAq*0M5aL2T<rXxvDyHxZvRJj<Ek%@JMN
za;D{+HNh2}X$6fdInzoSS8=9QG;ZJw8|d<`a;8^l{WZ??8jTw{(?%LU;EW$|3R3@D
z&iE~jxACUic<WGwN%OK#S2l~PW4!S(p5kBQ71ww>V#_!3ibfh==M~p!e1liqpm7uL
z*u=Y&=Wp_kH)(u}SKOlcw|T{F8sFg+cWB(qE1GG1msi}S@jYH~kH#&$qJ_q-yrPxH
zF@hq71oTAJ7D2Iv##;r&RvO0&idY)Q3661sJ8jE5#?v@KP$ba&L_v{A<86Xs8;z3$
zMG}p-3ySSDP8Jl&G~OX7cF=eyx?mwV=0bwWHpPzaG_V#fw1r$ZZJD#<<dX3-lF;Ft
z0BsXsUvi7w2Ho38audGdKmqnsKzfSYt%<6~z~nJ-B3sKdVD*f?@r;80QE(gyuxm1h
z@u`m`ud7WI#&!d`BKVMPiwPpiA9ARH?j*M*jdM#Qmk{WVh4-!FgGi9}A#d=IHzRXy
zpJ2C7aKU4Q8-xrNP5LpeD6Z-ILHatUIzM<DrV$b)_BL={GU}{KAg!c5AWe2ME+KJg
z(fb9L=LDh>)Qq5JOz=62uM1J2xCls#Kp^S0C&1(h7=f{_HkrAnL~RfNRecdiFVblY
znv`Zt#3{{)cmzm~05ejeCchaIx3{>J4lO?PJ8EO8U|uRX;wysfU7ya`U|@9(MdoUu
zvKIQ0d!kx6wiZq&IE6D$p`+v}oKFfj97oM@&8UO_J(75rMQyAH(t7%3G}+U*jmhUm
zr(fB$wl5;?6HN9Aws`DNzX~R5!4`)=aobT<n&fIciFSLP%3QtC>jZksc5XsD*JG~Y
zpel-!YjUme6DTJS3aQW%Tg2TdR5pJ=FMN}mlIq_2W7#U-|N4Ch&i+PorxZv_=>(Bx
z;NDE;lCCx<9QR`CtL%syW-->N4>q%wn@Q+&ugRfRIo}<QLERhAd&iTDbmUF*9+@Og
zzjyfE0Gml;_5bJTQU}J!Q$Ma~8jW71g;TbWyNERFvxc{CuY8LX^;|$b7lKc^Lw-XR
z|M2%iFQicOo@$bgaA?FW>)~au_xyk77T!<L+|ne{<C$`J%kO#fq;sXHgsuFvt^6EZ
zXLqbS><Hg;e|PxS<9XiZEqzcMx;SMQ*|KmdJ}TiLkRGHLt0pfW4<SwmLHDR<f%GhJ
z!)puO?*Z`|olnvT#6}ul2jX=aKLRF?KtKFEN=W>W-QQEucWM~AG);wC6vbLbu@1yy
zj%F>R3C`+2IJ7eL(;bLd1+A)}GfC^r7fkZW(#(4B>RsB$hEveXuLmCMK^V@PIyw7>
zciI*MW3<m~f<~L59B;_zHVzs^(l`n_M3L-D6qtiX(L{eTA~j>>yVUilhI4|!Il&9>
z2H0CbYh3eX0v<UO8fCr9*btoL+MRaknnP|7PyOuKW@ov)KT2Cm=d9B?TRgHjB?YxP
z4@mRK2_UNO0QWn<5BIC6dJo3F2Qx9&)qePZ`#JZIHMwTdPmcb_^WRSU<m~+6@}tq{
zRZ0a}DY*&8ISOb><pOCgxxMS&UP4Mq(?!*7;C`FFF>eLtt@M<NhAz?c4Z7w74*aTl
z+7CF;Oi=@=n)d%SpK)N~na}Ul1?$$fbSI4Kdz;}?7bT4A`%)-CKY&8P`oR<m(GTfC
z`6)w>?`$w~L(}3VH{m7O`U|AOrK%b2<NxQo8TPi({M3A{Lgr^F>s`u*;H$#!qz2av
zGC$QGE9V+L{Qe|*wR}D_pYO@>PzUrjYTx`~t01a^gdk%Q;20;2GbSO8Qxb4WA|=5E
zwFIbX+yM+b=p^Hp!0;uFH$!nVG$rdbh8JRZW4uTTYp%`A@_*X#AD`Pe`XNJUTpzyZ
zA^jHvJY(q9m<UEEf-t-qMO6|QokZgfVADZQ>CMn$Gjt?xcmRqApeZ)rP;xGG&VpWe
zTShY*`l;av+Wqt^gd>Ri`52l%{&yB%|K~6e+Z3WIhBu8N(YWrbnm!GDKVrHz@j*>|
z2*Ee`pqn(_EcDqd^dsg`oDdWzgb*Ar1jQ3vG9i;YeX@Tcnm?_avXx{Llq|jZ%d(+2
z4k5(`cHjnf2;Tn7GeW=8R|(;05&z$o7*6{~i#P{JbLefQ5R5FO6LD?<_gnO~5(!-*
z>1{;w?G@e~Dqeg(HG6sKe$=!*jN=|A2oE;?a0R7MYd*!o3K6BE54;6^vAv2@r_jZ_
z=DRFHJkN;d8ArSxQLiv!E8~dI4^j0uBmT`e;^<O!!{zhme}7i`UpE1(AMRe4`kT*X
z5vVE!v6AGtb?4}bRWz;^#A+JX2x1M5YXz~E#^(j`JdNuFv5v+U1n~lmFACyC8rKVA
zJ&i93;w2hi7R1X0<By5Zm}x$IBSUcuetHYe!()a&ghTW05Re|CEhNq7aCmy0U!FPX
zP218H|MN|~&M^4;8-quK|K|<4=<vhl4MQdye6X*EZGt1oZWalLMbbD5&WM7ui9Q+*
zizfQ|fIdHlO|(k>kMHL=+J^?a21sjYzfkjGBOdHa)4wo(xjNMr?FCs9kR_2~Tj6r!
zs_E`_RX?NiuavbYWnIV#TE<$Gu`a~oD`zdrSr?K+R>4|Sur40jkMFya{)K;7Q|u;o
z8?VB8V2X{`?u3~>_rU-<VW!V3f*T$*7oND2I|j9|nDZ^>#^S^A%Gh4ZcLvPsPV(#G
zCU%kZOiVy-hUODZyd-dy$V$+BpNSRdgDEKNg_hXB>i)RL<`MB!Z#+v0TeLGe^1r@K
z;jDPnV$J8Gq{Y8L;TLF%U0G4}44OP6=i=R>kS6`#KaD~YAfBHRPdun@hcq;k*g5yw
z)(uTCI&lm2;zoAhMm7kaQ!Rt)M|2%Cm;A@)YRk*7vfd1xYJr-Xz<VdqiB3m(kE0~f
zsonf=-}w6m9*FoFdc20-cxw|?Z=lB;f;;-bXS{(~w>KUk^okH<<n+!LdgTi;9PGd$
zIur;v45WvFIWcP<0+Wa2$ko+;gNv8V{R^2zhkqR3o$$P#9k`yH?hpFrXZ5%Kr8`6N
zB{X{pmAH-g!w6JWF_0F608-UGV0aG<!B|)OEibMLf2fFNv*tryJOv)PE%sb;V4)mU
zrRGD`WJvM%rsxUH2fg@-Cn@<Qo+o-c{>O*(X+w+dt++4&HLZXTD<Dn7z0n;3`0zb3
z_I2?mHmG~jfi#^SiMPPyEj<#N&x>*Q;Gsn&YQ8PTFZ<@rkrL0_$<I(5uCc;3wl_`=
zUcO+-++eiL;crKE@Dstu^jb|G|2$oS!)`dMRR>pwUG<)SXlBLo`70iGZ<)jJzNw=#
zg8UpqeoiDmN0Xl+JwF+h{C2xt_D^*|vhfophE5tgZ20JrBZ5bTj2SXCXq?WF9->Yt
w`HB7wHT-8R{-LAO{c_qrBk>O%ol)fHQ1UZ~{G3F7jw3%Oke@xdB04(%4-3yjy8r+H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp.EC51DA2544106248.idx b/.cache/clangd/index/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp.EC51DA2544106248.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a878430c418d3501689d43394aad5211a0324616
GIT binary patch
literal 18826
zcmd6PXINCn7dJC=@0D^{mR%Our8gC1ML>|I*iccg1zRjJYW{{K#@-T5G%DCY!2;40
zH5O1~#EzgSihx~XhoB%$uqzt-JIn4cllQ~>eoT0t+_OK=x#!M3Gjrz5nR62r*{j#$
zNF$>j|BL)_^2Ax=IU^$@OZ?x=Su?(i7L1IN?Tn0m%qSW;c}Q81{B`t#*e<V5ly3`&
z{I&az#~odQ(;`j!pI%=vJ4917Tw#8=X3hVi?&f}8FzI}bM{(`;L953r-rli^-2GSd
z#A$P`yw4psYtjBpm*d|!Mihq~tL?VcHbI_{?fkI*#wA&r$Mg%aYmYm`RJ0N&FZfh7
zL5fKTJ3Z~<byE-DDErxK;!m|QNxa-$vwgzI=N(p8^>m4FE9ECxhNO;lbjxhJ*m_J%
z&tcbnJV)>K{#d%9?A2w(p_+vg9zNatT7I}l*);oZ_YL{6t?ykKJ?Nt2NHd3pJr-^a
znD|v2?JJL8%DXw7u0r?rzdLzY`?-C5-qPoZw?b_9&ujBJWx>0s)-KY(ZQ<Yjc4X?+
zao6Tl<o|u=NlwYmdpX_zbTNy5XY^_E7Mt+U&a;(&E|0Nl3SVpYMGs+azaPh)obM5~
zUO4epl^g2PA;ZGugw3<{CS&cMXise!U-?sVeU_EO<>93hEC=<gs0gr$j=OvI-482r
zKM(HuXZ#rd$jgDLdk<Zh)^U>hYOv$#w5HeddkwCs@r%oD6<xUCfAdD(xY76EvLzKJ
z=BHO)6s~#wao%lU;n0r3HnX}V9ZB=te_`&z+b{Rdi>-3LUvVS;V@Yi9<Y%ukR`&0<
z^Wt8QjJt)0Ju+?=uGaoA>0n;Kor$&zZCP2`?)8t$=45``f5m_kmljR_W@Y?`#^`=|
za`&}nUhRi3R(uZXvu$gY+p6U6J*F?*A2+bl|Eq?|<MHp|6_3u!E~>jvvP!444(l_=
zYhrf!Us~%E+ux_oTWOZ{_3wA4+|2mN-5FMGmkuu2*==FlVG#wlXB;U@pXpdFIQ<g*
zZZ8*n!_Td;)!gD4MmZUU(NBK-JofD9oJWow@7yTb?;WfCtFO03dcdTwf4aW;E9Lo%
z$#3S&_HH%Yd*`U(n`)L;y@(gWgEy-}N3Z;IUeVs7$lQ`qlWKk);P2A@oWsDEKD$@G
zd3}z*p7;Bcf`bd|^0OcI(hi;6Q}DapJ@9eyuf0b1xjL)=h=7!<--`$DPTX^6(#gRa
z>yEV_x3^aIQxB)_wtV%}Sv#xW({Hzo&g*4%<F%Px`>5ijJDZ+u15>7kN+TDR)di;A
z>K`#1eNw!<clknfb}zH^E$4pi^zwr?B*|;iwLNvOt#^Gr+dtIlWZfJ8Nz1QHwjcW>
zJN5X~tKY?6`NcLgVNmSueom{rkN>{E=<=MTaNgf;U6;glqwC%3o?jR}INT(wZRRZV
zjR#T$@2#2HiJqfvFF(!r)YkRouia<GyR=?WRXelK)-w+?jcYBkUuB&5qP6Gl;3e1Q
z4a#vGx~Em_p#=e*EyGr1XIlPsH(Y{66UI%-P{4|q&bK<pZQN3RyGP`}jcb}_bsu)F
z+~4DigR`R6)Q*nMZ;UUybo4=d-lZQetDlBGK+$6P{D@fp<hT7gG{u=qcTZ-_ly4fl
zc9Pe|Au}T4elfG|Q66yEw}aEA;5e%rrb{%x&9Km1xVg!+m-*k{Y(M*>fAP>iEQW{X
zmyC_wIP-S?j4mE))-9b@CGS7`>8Y@@&kmlkeBFT)TK&{_)aID`Z-*}FP`E%dCVEP6
z>XJ{Y`h}f-w*8`V!j#;SYxzsECxF;7M&X+VY74$gZm^j*=EIt;w_YAN@U-cPJluKX
zrkl5FZ)e_|wBS<q5XZACwgwHqKRYMUYVo0yGvQ<V?!uI3i}^&e+tJ+ue>+%N734Q)
zdVY)@7_LsAX4zfQZSoG2X|cD;g9e*vGq!y@J<Fw}-SRHA@;QI>9{g#Tk?i;M_?s6a
zb`QC_>>>v?cfK`yMA)_`eHEE~<~%<8v-w~bOOq+?>!xVNwjZ*(Q2une&^|8T`FVEZ
zmy5O!wS5u|hIHE7rMJDAW5KAR=R@+MKg|8{#svFB_ZPzlcI>vPYQd$Ui8F>RkBiPQ
z{o`AsWuFfETQ@bff8ZWI@7CXw+K#QC6r6n1;e*<G&CbfGm&P}@pLjhzWr#`ZnV;3k
zYjR2sOx+YaecD{^=llV8uCGgzK3|@_eRkW-@xTS#{IL4#s`r<pOmn?UDsC>U``PnB
zw9~^c2fW!^*Co0$%=$KMim%rJpQn2U9$ap|a7mQ$4+j<>`Qu#fi!X8~)U@yXbo=5r
zc1LaUf0#SwRPGF=XWy?D&JvzAF8%>@cK%Cl;#e8eZSJDEr=Iz?U%$G}ZA{N;HHr)K
zV_Kz7U9$c2&c$&h-M1H^BQo>lukH;Fo^T@>4QW&q4~Ty`+AqGg=-Zhti^ne*Km8M$
zztKFxXy#9oCmc7y3(sTuk&c(o3=*Iw&`5fM%$^{HffyNMHvWT;NA@`$_t07fy*n96
z_mK7;>SiEB1eKW_Z<p|7l??X8B3F@k6@?fG5iyzRn)f_@<wZHP!w<T{o7~~;4TOl8
z<sS=nn`u)agFUd8Qe<6<S{n!vu?U`*+2r-p9ThaeT9yLMQs8bNM8tB;^wgQJnx@O(
zD6FLcc{bp-=m`<A@=a@SbaW3_!O>VtJkZ30HU>gO9E(B^7f&5jtAeeu$UMzF=jIDx
zJLQk9{XFNl5;G{r?2XWLBeXG)7LnsV{GPCy^)lEUx1$zmYq8n%gowEIe-G4|{k;^>
z32Uj~<rUZ_deR~iZIZQjMb={(48dCJk!L*$HxMGy#x(orWZ#ku861N}ijlk+X$*vj
zcp3Tq?B(OtU<QqYjHG8s_6%7Y2odRW0lgV~xhy~q+hQ$;k@;af5cGtIbbVBx_HNeW
z2MXwhoBI)PAAy~L5Rr&c;IE@Ue_?C}EwD%iG|hmv20}#o)OYEYf2;HNispe($jJ*i
zH%U*3NMz#NILN1uw1&pGEq9^uT_`sYA~G!7`r!Ni4aN%SiXXHR%2&d+20}!Jw^b?T
zdc1W|Ku0W+2;_;t%|Kd2Liex8ox5q740gZ|szlmK)YU+U$SC>Hewp*btYq*@EOH8|
zPa#(WAtGNEikn+M9xlmXH!M<vv^BV=^@NCw?^4#JSY;=d!S-0>0#aN+js`+RqQ|H6
z4#_@CW$-I3avHTejRFjWh<qowaNk@g?k|I5u}C3O7vi?)2@(0;DY9NQzn6^wjqyw0
zhGw^+(m;sFPxXC2%06t*mBC=F<rWIPg+>?%5&5-b*s^nn)N@o&jkV0y%(id75cVQ-
zS)=m3W~u_ZVD=>-zl1HTCoLk5o6k>+K3}SWR#;@3W}1ETg|HVd=w%+7c~h=}JZAr*
z`NgyOLfElWl6-u-{(MOeJuv&I;Br*(HxMGS@~_+s^?{<zDyYOFQ#DiVn=gbN|JS}B
zr0ShsGU$h2@*+}P#0J+BB9hQE+*O@9Hc$qAut*tlEJN)KgovzvcgE%IwdGkd7>Y&e
zkY^q0Y#>A=`TQ@LUFOfjzQh!_V}WLYXY+-yQ{I~{7=7`{Q56is>=i(>0{9sS5lJmq
z%M>FA3^IcXEV3DzZiZF{LPRo_ROE+^I5b8Mov}!~;2JLk7zh#Bk*OGD{;1%N3~F&x
z?<4JfJP7rKh-9TLli8S$ij_l8>;>{TSsrI;AVg&Mo17In6NI)h*ad5;MiJF$lz|YD
zKibYy?dbI}M+Vzrk-Nz8E;gH<5RrW$t-?HFZogAOTio1PnpyVEmlpQBB5>`Y)<gxp
zFnhjczEkssuyfX~$h*|A|AY*V#1D9gJRhQ910f;@7Om!%UAP}9gT7d#3TdnGywwvT
za;zyR%5TcMTm^K;4_PV5R|<9pLPSp7etX{EtLG0Y*anMCc9`6z`9jzQe3E%^W?iHl
zT4MHbL4I6tFc2bgii?hoKhSQu435L?ID?dDkiCJlh-hwGtlKw2kim{vq#QYyBOe24
z5n1=7dXo0_WEHf=51OHwVc&cq>>~5CUHT*r>1zf>%ua`<>3B-(NsEZH=i^hop0_cB
zGAuI5bdsX^Lf9n(V~e_6kxEoh!0bgpvj}(^2obq9)a?8os~1QGZLr8p%}o2|3t`{z
zS{5Cbvpq%y%`kg0&@9Fp^n{4q+*hcv2<+k^gPrk<K0px<@X*l{B2rdX8~AJ1(Y`9k
zVJ$I06N6W8Js~2M>*Vh3ZTcKnL3=DRS2NeO`9j#$6P8^x`Yo}Q0L=r8q%T{2>DPQQ
zGC>d+8375@RPQ{$Z^XS(G8m3qQG#qrP+J2bIvS>J-0%5$+$I_H$0FB}|8*2$AVlQR
zcOy=H+A^n)3aW4`rfa6#H(v<*b%LEw5Ba!i{X}|!WG|4Vq$fn=?Tr)tg3AtfH-oLP
zhE!;piaSkDh{&hD1^0^luRoE&uDBhSk$4$}8VC{jJUQ%-hSRHttDpnc@{{H#=jIDx
z8_k&b&CF+=aExY%;y6=7>_+S);vwH=rO=+t22naDI3E+*5C^Nnp2vg^1Vinue5vHy
z?Ld?gcx?jjOIB+V1VT}YN3wY2O4evoLnufV9yDWqXz=OUaIB$}8&b-RA{t1Hh(u{A
zk}XAHM1v^(fy6&h2*IX?$dPCi@&cU0BK98VL}@=z?g#D3!Y)dMKv~G*`#^agv?t4x
zC_MwpXB3kV6qp(UPhuq!c=Dq32Fl(*N758gT84TrLnA2OhobhO2^8m}w)rTC1W%@h
zaFsOOyu<Re?~{DXEJbOm(^L<#W{J`?r)eIG3`?XaO?R5^L57?t&2XCGamwHlL}{ke
zOb;?SMQN7PEDz#cMQOIvY!9+ji_%<=x$VgG5v3x~wum)N5eP2=5!7_WAgCC0VR7GL
zFo@!ppzTX)I#GHF!e4?2YT70c)C9V)xNj2}L~#-nlc0ixY39+IPT%deKM6RgL=Z~^
z8F8*8pvNOuAI+0K5v6jVDhEzf<3(tE5h}=PDN5G_i)(_KD6$-O?b~VF#C>>xX7NL^
z_>pAXiP9;4;3<9>!DN4c1vens2INFK)6}p_Ae~2c3AkzcEdp&?77(+5ni}gS5N`rC
zZE7qOW1*VF%Ql6-CLgFs`Dn~ZN0G--)P@Wf9S%5(I#Ha9JaSonE(*wH`FSWP4~0<q
zJTxc|MNyoOJo4H5^HFd<3MG8oTRFxbb{_FF;iO`qEvCbRlbV3GiFP+9y+Z@vp+Usa
zanffL@)?Cvyo3u`!u<n}Uc!x`{1`4IhUK5&Y)){tBm&{2GTyC>cPBGa^Jmd?qlU1F
z|GI_9whx&$Z}(`A|Kk0N9|>ATY_0GnIvL1j5Z|QV3DVp4K;u17MaOI@GA>0b;%}S`
zTSPn**CE+D)~V_@iexA{FYOWO`g_e3>{LH++7FyR8O*NX2XkKpT`~R7^(fieVqdG@
z&=Q-t8aA}}os*C(3Axkhr{CR@cP86fEZ78PP0*GK>i4)rkZf+T`Sm+nvKoqbO}2f3
zoXMyZr4wAc6P!O8g<h50=R7~tTZW&o2N>@GvgULLzKGIYV0@S2c8z1@QN6n^z=B!4
zZx$a&%&+q;+EctA!R?pz>c84E+ZhY)<b8MY!Gs}7Cwc8jK7im353B==%*wyU9U8~E
z$8m#xmybwJc>3cc-`Y)k?Zy>N#WP_aaM=f3iId@^d=Qimf+;=@+8qbp6c+%;0^me(
z1#qbVuEZ&EQUeHT0KpVDf_9C-o8rg7@iFzc9cPcY^m%f0CT`&>e#k0*1ZjG}$N(Gv
z-ccVg;}vpvh1wAF26f-i{lGgHi+{c8!Tw95Do*ci!o6?^*c}3a#3rKj0N6bMf#jWo
zZw5>^ckf<~1#cnaTc|bJ8jDgYXO+tNkzNULe12|{F#frX&SP62MYhDFa#Aj`&Sh~P
zYMY0=C_f)r=d=7mK`f-pA19p_#M3N3BdE^^c4SrKq#{9G#NxAp`Yel!1$8ltO9Z^a
z4y5;&3W`#}isEyE?Kwe1@p-}aJc};~HW%os%1M_5@e<p<%Yt~B#a9H)6~UhByDErR
zS^hOabB*QyDTsfv{Odx8>w+J>|ArvmpnPIS>_+stC^G5E+KLxveIknZ7g1UZUDm=N
z5_yQ&C;M-NeTkFSg?OHCU{~TzdRV#*emF?wjED0|V7wAIk{x@`(EpuiRl2CSc>-ux
z(MjKP-K+Y+Gn5sW@s1n#j%$f5dX^?P41MPoh=<bw&f@?VNM?p89pOBVP#j_P*m8>!
z&1=@6eaE%<50XB^u1udhqi?kv8t;bIbR{W4#wEy_Y<D6)pDA~6h`NR~zUQ>>xt6%S
z_n#S|>7ycMDn#iZkRJpZIu{NCzk?u*;=@3GnB^Y<@*^xh1LSA^iT%#7_ZI`FV&F;j
zl>m7OD_;uar7W%j@+!O)HRO^+sS5a2fiQ}zfxMdK*8q79i(dfwi+^Ij7wr8_z^Ms%
zQhl$1{53282FTw~OxznDY5HU&F+CY5n4!-^k|8J!>||3Ckun(DJO{evz}BR5qI`F3
z5A<4DjT!O$;CLE}4l#*1K6#>iHD(+aOpXf{WY`ar1>b$~;K_jIJGAk9N79;MA%mwM
zpZY|nY8O9r7xltqc;k_xPkxf-ig@JsXTW}BDB-mYY4#vL@*1Rn@QwPED7`Ta+NHtv
zWalbMZ=uIq7}Rp3Ay<k=_i7|tja&%B)Q~YHb{RfCY2=<QWg)oDcLe1f!J53~7yA!g
z>73H{DQ3LoqTX`D$y<^fD^~p+%AP|Ts#>3LC921Sr7O4c^;7=!CFaW`!9R>??e2#6
zq_?5NZRkl_DB^^gZ64=NVhn5p-t7+WLn2)~Bm}E?!I@<ANbVTF`dTDg%Z7mZ3^GwX
zwz@9TNSvLA1#fVw8(d2$AxgEJrj`pK#vE6*JX*NttaDFeI*bV?-Agjic>Dbp%HBc^
zv9XAkw6;f4N8)Nt4S8wu;IC^E+r)dvy~O$wd2J%!5_^(_HP-kN%3jg{O_a7F`8L$@
z1XDvAn`n$)HDbQQ1iPd7RW2ciOK2d;gosiy=b6lnCd)3#cVod9Q1*fy7y7I>(MV`3
zQ$wzsyfu#Gnokdi08R-Yw`Zswi5YPUf%7V)4p_t+lTKxPIPI>Nyv0l2hJ>@Cv_){<
zBD5vlH?_v&<B92uH{urV<cII%Te2OxfG_GK-!jCf1wEZU<j+|)*jS0sClPjunpu{9
zZ&FMl)`)z!pspPqSN>Y9Fnc6QzwzyU<3m0#D(<p&V}r{a+dh5dQ7gZ}PuRix?%+F-
zeio$yURyxB<HysL)$_Qk7bW}z;FAbCl9?__Yk<!h7Ow?9YgxPw_^e~`df>C3#Yw;?
ziN(plCz-_?fX@aNrvRT67N-KAR2E+XK9@j8>X|PCpUW)%2-<uE-eidvrB9&ECl<#+
zWgN66E)a*%&@m3WQ5+AI@hm?cI>xj7a%f);Tl|+ORY3a+7T<&R_gH)%+TUk!CA6<(
zaTT<$`X_Ev1-<D6tA_T~EWaAIsb=}L(7u-C*FpO_7S}`jdKN!~_77Ry0PP!C+z9O(
zS$q$f-$M$L!V~eafcbr-ASVVQ_TT1}NHN^t;PFa~%&U-sB-wG)jLaV(1xfJgu<Zln
zKyfuPuSN=Tb|B&d1KVokKyeK+uR#jp>afE?=Cw#ccE2J%Q8BMW3Ub&Y;`2@OdZZvb
zMG+_e%^xBKITsLdp2oZZDM%s*x1U#5^0su2KH!xPSX|93t65ybD{EL>%PVVHT*oWx
z{)rvycsHuQo>$hh{CeK8p5;H}l@D2d1FvjgaU-v6Wbq?j`H01jdF5jkKjD>6Sezgz
z69il0gYf(olxtYLR#2{G@j5}dj>YQ*<$4w;3Cg5@V#g%Gjp|Ppl*ueVS#V5d`5Oe~
z29}>9C{tLRDkxJ~oF*vKSez~>(^;G$C^IM~n=(O^mLu76<V7b|8WPh`TZ+?>n9kx1
zBxbO9Cz9_(Efdq!aOz1G2a-yKrp>3FbPb8&M#OL<Ni<LV7LK6wXP#6r9kx%0-N=EB
zD8(TA8016RM-skxmC_xE(%W`}fZd=I_1AYm;2jV|o}=5Gwz-8oi5+yf-7Tc0IF<8H
z<pRj6soT7EO67tHCTT1DTDpT)`q)@t91Cp7p-jwg3+G8$Ma6A&iF><J&YdnarCeAk
z7eVpgoa%4Rk;OrObKNL@%c<V7{I^`lTdoJ?zvEQzSpGY%(>ty!<-g}t?^*tPF7Q1U
zPWc}=)d!Xz&#U5jrakd|a6I3g%CF*8t62UjK6Dk|lk!*ds?{uiHQ#A9&$K6jS0%9g
z1U@i<XWEm<s}foM4qm;3ccuDw@L@Z6#(q0_^-h+*lkc*V?@i^icy$)b&*DR~*!E`g
z>TH&u&4*<3jD2_U>Rl{<7az8ZeSQJ2Dq!1Nzy}rZO#hwaRVP{gNj~HxtN#?QI>qu&
z@tsbw_A2C6g)G024=iNub(&Y5ru<mLVL#qr=}-E}2qniQylL~%KKW<76#@af*!glB
z1m0%n%V!k$83mE~DoRVZz$IJ@j_ZqJ9w&Sb8X-!PRg>+>DKO4+0jE`<HA(yHaNAYD
zhsD0DKoG@=ARv+50IvZ7YgoJ%1gvH8IuNjq#p^-9dKM>vfFu?tgMeffZvX)sSeya^
zQdpb{0#aGL4=DDrrxce!z$JErd>I5>W^n~jRIsNQAA$BG=*Wg)+D|NwgYI##9UFGJ
z$FsNsx>i6>;{8PF9(29O;``9`K8q`%YbA@TptcHjr1w`tZ8eK)p=&KGUk6?5SX>WX
z>skB|x;|ua19WX*aU*nXWbq?t`H0=sJcgE!SzL^ai;)G%Q{(qT%6mw|hFQw{NW+F%
z%1WeR!z^VL(y(Ec;{)WuhE<N$$b$`|lr>1hhEd8|q+!D-WgXJ6VU)5SY1lAI`4DN?
zFiP2gG&GEQixh8>nogVBocuOtL*3F{&iXFrKyevoUB=>a&bpk%e{t4-vG^e;f5_@<
z<g6Q6{v*!%5sM#l){j~IiBo*y)b#o5c*Q!N!Fzb?J-h?iIN>8uUQxvIAMol2yfY2s
zs(E!ai)(mw4U21ebuEkQc=tNK9lgJvcduvhLtg!mm2cqH4J>Zt)r~BE#H$~%_%W}3
z%;G1!`U#8I2<kO7bAS^Df_g2B*9q!%EM70D*RwcDa8DB2v0<HiGK)6|>J6-Xil9zm
zajKwBWpSFIPGfPppiXCThM>-%IPT&at99xP_i!HJZ*I`vT+4w1Db>PT<0VM8gys!I
zoLq04%Vs9@1zaRvTlFAgpmyQNiMmWjw=q4)`mj1q;kW6xsfRi4{|cORm<^mb94N7Y
zlMb_i6DbhGTkn-fwi4M9S7d4^6C({ty#5vUV7tQ&EO;F`Tt|MS#6*-*xUdv%FbT!h
zPCov5;?*xZ;})+1#;br8Sxm^J#TwQl*?Kk&q^~+7k4mcaI@$Hawg#-BK(H+k+(~jR
z>CLz=7hBlhz!|9*Nc#c>(|P;?^?QLvP@K=%=Cc{9d@eAb>rOIO$)$aE{<(C4H*U%n
z!D@@(L<h_<!RnacMAmCk`G(_4UD+EkKdzYo#>;>WIoT)qbeyTzm%mZL3?OCzcN+KN
z^z}91PUGH1P+SDv$pS>us}`n)>Nu)s8FE>M+L7)NrF}@gj~z$*k$gXkPb16I$c`ki
zbZIEdOURC<p{^jyE69#itmt~x@;b7ky&BJ1#dD6NS8=Tj=a9g)p*WdyNM`W{&S3+K
zQ@9Q(TnLrV;Vg1E#`9j|Ev`||mF9aed(INyeYnrxa|7OULx>AZYq&Vc;)?A++*MbQ
z!xc1uws-^Qxq%x+o<O<~hgkaJL1Ma1R@-Lp|8aX9){xEnX43;ooEYP^g|z->Q}=se
zj)VJ||LZ68Tl%Utjy-dy4>rc%oc3>e6u#Brl)SurmhN1#mJ6%pB8io@^(b@bUidEk
z|2v?!$2T0AARDhc58lBK-NBD08g!|)K?OAVzP;?}1V7u#*7#AYdF^U?{!7++tYH<B
ztzxIKz7&&8<1FvNKR?=>tV^mEa~8#%3wacog}A5m1)AhAJA1^nE~ZB&Y4I-jAh0_K
zeCf0~1MJSQ_$;tH%i=0vR|S0Ow0Z&TUa<HTuzN)@i8t^&tU<Ch$c;=ZQ$ryrd51ke
z=JA6jZ|x-)rNhAPFmR(SIt;uH13!w70J|eB{|L|?0Rfa>0_;jyehKg@Vec;mcBL%8
z6!@2dP<nqgu&V}c^bM<lS2ggXxCYqOu>2aJtpNd){~Fl6X8Esy*K79vH^A-<%YOs>
z-+)kZ|L^`GgNine+kh?h1ld19ZAhQt)Vg4GlzL23I*SEYAlV9bLD3h`64U-MW7+qq
zork`{!y|>WO5s|QiA-w0@bJ);)sktv&rnf|cS^c)TJp~#t{L#mr)56)*a>KlvxaQ|
zo(;6@N0b_YwvoPg&Y0VNTjMPhcuhHqERP};IU~@m%$B(<&O>f_wD1T&7+L01eopMh
ztXUs7F2e6qB3P6NzGT%R!6>#!GLj{umdZO*L&Yz#2#J!hU;>gQu%}P@x?duAu<fbC
z*DoEJi3Lw1httT5diey-Gl2^vraaU*?~u*P7rXv*r89>E_ovLi`lt`Cl`9aG1@!zZ
zH>2y`lxvn=*oq1KfCRo}9XKX4zH;`9YUy8h&ZNu~4^Mr?H5rL{VJ?={%U`d;F<vGT
zGm$lMD!2d@iN}yN9a&p><E^}c_{02?8zX-YeOdFL>m`ylurh5f8YeHRJB$dAgOTj6
zI6R)kNl=*t9VovX`jo?9YL6<|rwT?<Tn+nFvp5rJGHKP6&VHI>$er5n7V^1;f@yN_
z0SbPAda%37;A+%^czIp2Ff^5mAhSZ0w({0nd1qpOU2?s3CEtPWdg}P_IzE!(dOp0K
z#R)>|1fc_oIB=Dh5S}DNQk*P=CsRxu0Uj~>3VUM6f&tmW;T->bJlNlI+PCz)lhoYf
z3H}PoUa^OQ`T~672~HU*@3CMUlEonhVpCH?{e6p|q1YA&vby4X(gI!HXv7kl_dT68
zXo`6+;~1>({~Z{gdHv0>KW3cVjGI2%dvtH2kEDfhAgr(6Ck&j>g2ua`B?)jv{Gp2R
zMPx}fM%ad&V-k%Yi<}MB{-(`U{#4CcXtNeN(V>+FEz_Wy9_Xh-_jKq@+M%n2ae4#W
z5VPnGhV9<Mmi1njzE;GrVe=j&--BA}@N{RX>SxG>o~5qmtk!e(bSNcpR!I~WKTP{|
zV}0ONGfp}qh-YY#kFLl`EMoClK|IUiVnHlsaj77dviO`Jo@4QOK|Igm3xaro#TNzf
zB8#sJ;&m3^6vUeplg}sc_|ShlK`kB&>|%izNui&gkv;hRgUK`hbw7MT-t*Rbi$@{<
zdP6R`&0oH_%Mgp^2%_^*)Q$S7T+}(2#d+w9JT!*#^HH~a6hZLiu+~3E4YW`G&#ePj
zZ_Xak&iU*@{Ov#@bWVir>A`*mw99~=v=C<_wA%<hX(8Gc<hliQU=RN7f8)JK{)!a+
z<58#k9EXm%7%nh|X0xx=Rpjsc>rh|pY;JI!ZgBmG=l-*w$&8e+3A)mu)%?KKv;dxb
z%Yzl`zu=*&mjSzFz>DnMZ}>baJ1}hAI-Le+v6?PgWHG^>FcHZT*)4+pgCF8q$;7~M
zss1w``l!7?wU<__iqaWqe1@&=|A-tu(%MqI!4)j7(7MEX$0P5UH;h|>Tae91Wz(&h
zZr?hfkcLK;L6>{H*<%a-bK{15seumzbzeDA!yE+N4uVL!x5)<~`Jg++MPP6d7)fyz
z=vD<H$zB8>j)RZ}(4FF!VDL*YlHkg}r~LlD>o5&A+%DdC7h6tpn%ADD<s{X%+r1NO
zE!tzjO<em;oFB~}WpeE^DJCt)!%6o&R;2r!xHZRmED)ths!3hRQdNEG<n8p$Hg0$*
zZH6|Rp)XnMs&Bk;Of;EX0Xf{m;uY$2IsAQxc#Yzk4*aWD7B;#}MZ91Vk57mV-$d#*
zStdsXB}q$>l1MxuUqjhz*z$dcDD6h#ZW;#QDotc{0C~{$FAs(0q5j0>m>RwtBZIll
zJNI{ghuvH8J7jP+8Jw0_gye;A)0QFGGUQA&ni@VUBO1szKG;;cZ_0?}MErHdh|j3y
zfcZs6Zh_A}7x(|%n0>u8@~Hc{{Y`jT`FD%ciRE$BX!z{T@Bti&b#RQO|JaV&I2#&g
zvoY2MWPAZxlCn6QcjX+{(-`Z0&tt>T%_M|7d#w3bZ=&Y?`yB;7<;l8@c@8wrfi7em
zliVTR;KU+XEb2_g02V}I4hkjM)bNQP8586aKU__q`@WAnNR$?MFYqHNR-7dQ;l=dx
ztr-)47~d0rsL*qdS(Vk(g(Cn-8<Dz1Bch(<CR#obBrj$}o-}jD*E65$zI^DTyXgP=
zk{?D!QS|3<`g0Kd89{%Jpg)JwpF`-+{`BWS`g1V-*}LT@l)jr6^RZ`}uoudq!-kLe
rVo0CJL4Bk8^&dERKt%6eM!i}DjlQ5i`_Z2R=+9yFC;mS#k-z@~Tv`3|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_wmma_instance.hpp.377DCBF55D8BBCBB.idx b/.cache/clangd/index/device_grouped_conv_fwd_wmma_instance.hpp.377DCBF55D8BBCBB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..47c0b2e1333ab35cb903b7c8ec39f05196dc82c3
GIT binary patch
literal 12338
zcmYkC3zS=RoyYV0|1Z<DGgSskffhQ@7P=H%sz4FgIeD0O%)>l*x<bJ?A-S16?<EhD
zo4`>Jseof(XEqEk&6FL6+yZswKn$sGR#$J_J!ek^6+u7|MHX~jR^2<}xy-lc^qze0
z_xHJ<bb9CJasLso%e8l($@Gy=ddrza>40D|nI`qW#ZuwMzmUoF!kEca?)1Nu%^ur#
z{I~!7vA1r%H@V}DZ-+08Pnb_U=DPc}+P@t<|C>kux@FIoAN}=B_s`w+t-w!?jP~!|
zQn1`#|6y^@hWd`@?>t(b`s~L)__bXX_3`&#aj5%&tIq%GXJ6^ue(QmKUmKYJ`(bjP
zC0zUUA3yx}Z*99)F2C~P`%nMGm3Jz?DZl*U;rqY-z$Gh3tG|r?>xZs&+-=&pcd-AR
zW4k+P=|>NJ_o>VM_q8v);TnE;?)%DU(^%a8nZ4ONU%2$mZ@#wgwc6X?{msT3@2mV>
z_3TI9dCB&(=k}GK`24vSUh{=-AKG;9S=(PYxc_(k>t6rNxF>n$Tl=5+l=y|6H@>y&
zsdMfy#ZG_XnP;rm+Mb(wJ-hw6Q?9t|$wKlX+xFBOfrVdPfA2#dxcwW$()7VU{`T%$
z9=~+gG3({$7H=E2zO(OOxA#=}^-J1c{!8%Ur~Y!!K2eTcaO9Hse{R41NKH9%|I(8)
zn=W#H@xCW!o0Iq6bcXkPo^#GR>wC$M9=rMvKYZlVzet?@aqDd#3?30WS3h^?mm9>=
z##0}>{q~1{x%E3&Z_E5@#q{{CA8kGR@oznJUGt^>yM6z`KW*6X&OhIFQ}@|}C!TiG
zKkoAWc>ncNM_!%SDbJq$YVWzPO#aVTTE3g^{MyqOynNwL{;TW!qWQzOJoe*-LxXev
z`oxY4ryl;p{a<?Ce(zfsw~rhAASu6n(vydJFaNRq$XV~a`sOb_y)b^$v8(-mKJE<f
z-CD7dx&C+hS+MDUANc7l{)?jH?D#FFsU2fGCf7Emjry5lGSL)G_CB{wdhFNV7-!@9
zzzJg~Y+T!{51f$u?T^N9exo<ePSFPrj2$?CZDZOfu)iI_^6u~7IP^+IADGk!wvBDu
zu(mO6<hp-VVLku*H@p$Qthm{<SsyrW>v_A^HfxHlC+Y*<KYhr1>}iW<vuUF~aL(3q
z&RpB94V-j>KHz_B{yUEy=UUlpGV23pY&~Q5+Gc%V=HD-R@>8eW_S!f*Ss&OxwtwH+
zX8lQK?|$^3j@@w2zic*b&<D=mdiL(M&H52;xo(?l`GXh8<Lo$nU~+77)7oZjVC`RV
z$7k={{vU7d-#gCUuMcb)+hShZtPec(;Gz1Z_xx<nID4NyaMIXG8`n1L1B0<AKe=l@
zeR#8JOdr^_b=S_d&H8hATmAjx*R-^A#@Qx);N-EBC)YOXNBHycmp}eO_M5Mavk86R
z#IX}Mu5H!_Ooha!ioJ79CSIsrcF8whvCzcSNt;jFVmvqFbck<SI}HvVefNP!U%c#$
ziK*#%n*K5wKSn)wY5Fi^%RE`Iw{*#Zy>*_f*ju|~#ojhgHtcO(vSDwZCp-4`F4?hn
z%##CqN0%JfJLk!Xy|YVB>|OKZ!rs*-7xwOXa%1oAk{f%^JbAG9bjgFgcb>f1d%NVt
z-ZxJ^?0sGGVeg+OKlc7E`LPeoQvmxwmjc)a=P8JNuuDPgL-Q2EKGdZU_ThO7V;}BP
z82iXPMX--_DS~};o}$=CyA;JfHcv6^V_k}2AD^c<_VF&ou}>~g@+)LqizcR$OO$*Z
za%zE6*r%2#g?)N~(%7e$D2=_iKqB_y5{cMn7AS*#W{EP`XBQ}oeRhem*yk1~hkb5|
za@gk=D35)9iSpPN7N~%IVTlUZ7Z<3AeQ}A3*q0Wlgnem=O4v&aBw;Ttk%YaxKr;67
z63N(?7pRPVd5Ox{R~D#(ePxL%*jE>*ihXs7s@T^SsD^!QiE7x_7pRVXeTnMWHx{UY
zePf9l*eeU9V6QBZf_;+~s7W}{q$PdN`i${Hs$W`HUq#{4nwU<)g;ku+LQ<x0hO6xK
z98KQ?*^;mTY7s4fS`-VSmN~K@YE4)HwTf0it%?;<>l|4TwIytT+C&?mHpPahZH{b+
z+7os_?V=q}yJAPwK1X&$9SH}Z4$%RqLvbMLm?H<G&V&<Cr|1OKsW=gJ&XE&QSHcCT
zOLPJ1Qe23-=E#MpJK+Y@ExG}9D{e&HbL2+UlkfoQ5j}u<6c3`FIr1RtO?UzIie5my
ziWgDu9C;D-C47MTL?56&#fPYGj(mvv6MjJbq90Jd;z!g!M}9;Di2$GhF#u>l2_PDn
zqX448L=e!R7z8w^1Q89+Q4rBkA_Qni3;`NaLWqXuD1>M@5e76Yh5-#LVMN1o6h<_X
zhyWT9BY;Mf2%?cWiXa+IL;;P8Q9z?g6w&A$MG=i9Vt~fP7@#pFhG=Y#Vu;2QaX{l@
z9MHHDM>IZ1@paUAC8kYlp9YMp-6^O3(cORU?MeM@=b4T({l@!t<0HZn+qCxCz!=xx
zXR2)aFoO%?kWC-Otg-14<_@2}kHh>5pMDH;pHF`s^Rh7gsIWe7QJDpOmLapC&!RC4
z`Yau8L7!!XThM3ea|`+`%Yp@cR+U-NXB{#t`m7qWqR-mlR`gj{xD|cYKDVOJx-3}H
zXH%ICeYPR9q0go<8~SV=ZbP4Kh1<|)>vJ3WY|DZTeRh@E(PtkrJNoPzv!l=6;db=d
zSGXO0_CB|x&%P|!(dST^1AUGmbD+<mF$ek_9qvG%V}(1==jd|>`W(xG1AR`FInn1F
zGAH_+8gru0+2Ky~Iajz7ea=32qR+W3IML@)nG1ccA#<V6r7;)!TpjL0pKFD?(C6xN
z7y4Ywf(w0amATR99x^xj+!}ME&)wl}^to5K8-4CRccag}EV$9<QJDvQo+0y~&!aI9
z`aB)(L7!)Zd(h|Ua}WAF%Yp}eUX^*#=N&RH`n($RqR-pmUi5iaxEFoiKKG)}yDWIo
z=Tn&veZC>{q0gr=ANqV9?n9q%h5OLw>vJFae9M9leSVes(dQpBKl=O{^P|t-;ePb_
zSGXU2{yz7k&%Z4A(HBrz0DXZW3!pEcu>kr49UeemV1)<J7wGc<`U1;B0DVD~1<@BA
zvLO0`8VjN?*x^C+1y^_weZf8tqA$2C1ko2#SqOchAq$}|q_GhCLLDALUucDg&=>0S
z5c)#PLI{0fm4(q49<ng{!Ws*sFWli_^o3V=7=7VB52G)<EQHY)QCS3iks*tqFQTys
z`XU`3L0@ErN6;7P^9cGP%R&TwQI$o}7ag)F`l1?(qA%LvQS?PucoconK98a=x-3M|
z7gJdbeX${np)aPf82Vxz9z$Pjg~!kr>+=};V#`7deQ}k=(H9@GIQrrmi=!{z;c@iE
zS9lzK@jj2EFTO0q(U(+N@-Q>red*tW3|aChW{o9Bm=}5Seh%|Jo>Vapc=B1yJt3(I
z>+>m<rO=levK0DK8cU%swa8QGOZ9jPeW?LYp)b`FQs_&oERDYOkfqU=)>s;S=|!GK
zU%JQB=t~cH8hz=WkVc=VG7){^kcsFMH724@T;wA9#2y#XCl0uXKCvf==*y@qgTBm=
zWzd(=SO$HWMV>)lrpGhr%M5r1eVLw+L0?v7S@dOxEQ`LZ#<J+kF7hn;vOS(fUv|K=
z=*#wmEc$XP%b_nfWI6QZG?qhOZjtBEm+SEy`f>xFLtm~Z<j|K_Sss1)A<LsLudzJ(
z@{2r=zI>18(U%|aJo@rIA&<U-$_nT!3|Rqv1&tNZS6Ji)^c8x%fWE?j7tmMe2?g{O
zRaQh_amb43D{8EWzTzS;qOaKFMf4R1yokPHPbi|Vq_PtFN<&scUrA#n^pzHQ34Ns=
zFQKnA;3f2xdO`_(lFB6XNkb-~Ptur#K53Cl=#zR}LZ39?68fZ`AfZoInT$Sp$Yk`%
z8k5l{FLD`ua*xaClLuTzpWG8<^p#asMqhcz%IGU=tc<?$A}^z_+~Z~Rl?S|xzH(0}
zqpzZ}3i>KTRzY7yV-@sO7I_7Il^(C4uQK2j^i_I71$|YORnb=+vMTzj8mpqOy2z{O
ztM+&meboW4qOaN$s_3h!tcJeYkk!yv(^w6CwMAY-U#-V$=&KEQ4SltqP(xo`Wp(t`
zhpdjiy2k40t1t38`szJiM_+xw>*%ZZggW{fDr=yxF=P$&H8j>hUt^Is(AVhk2KpKU
z-aucYCp6Hfs7yhhGGq$+6pbn9Qx>^`KBdPM^eF?bpik)u3i_I)vL>OiNke@PI-8{F
zd(hgXMcyRzHmRrYL35J^`W|%aGlD+z-n$v&KJ>Yp*6u&Umrxo_{|vu&*T{m|s*x45
zO(Pp-yGC}*4vid`of<hYyEJlPc5CFu?9s@B*{hKkvri)*X1_*$%mIx8n1dPxF^4n?
zVGe5)#vIWof;p;D6mv|Y80NS}am>jPC4Y`L=SGymoE}jcvp6CVb7n*t%-IoTG3Q2<
z!<-*c9&=$t1<b_}6)~4aRKhHcNWv_SNXA?qQ5kb(L>0`{5mhnQMpVOGA5k51V?+(i
z%7_%q`pK`~T3X*@tdG#YY+YR(VVt<ex2yW@)g=+e?!d)g%x?X_9e3&9!S49*j*l3(
z{6F#Xt6tND^t7OVHM{dny8%y3?cTZDZY)ihnA)>*kKI@-F)_7w=U%&EzqZ`TUe|x-
zCH=UQlar?y3nmf)TfcR<d%eDU%jPX78p|Y%uUz#fx4Im|cyh*}_1&wBFpS-@-|_u(
zuiPKj4|1OKyeo{~RKuZ9udc{2rf+@IdD<mEykna_eZu4kCmVO(KEL7F2X8xCXzII9
zKIP=ojnx#!GtzI&t*+WI4x)c{nn~Wc{z)e}o6JFKQSufzg_CVc-U~T3lL9t1n*uh~
zq7<^JHl>hF&!mA(&!&M*w<wKlx=m?h#TgM;aaIIYY>|kp*d`I#%uEK@%xnhOOp7we
zX4;fNHan9AHanXIHrt{sve`Cek<HEIfX&V3fX%fihitA*Ib`!Qd0_Lid0_J`$|IX^
zQy$sEOaa)!YysFpiwejV+EhTcI8y|+I9mj^*rFn`#Wod@EzOjGEzOpIEw!kGY^hBp
zWThDiSZP)QR%(%itkfn6S$ReVR-ToCm0KhuE4N8Twmee?wme$~w%npJvgJ0Fk*&;B
zfUV3{fUUHsf^4Nt6=bV3RbZ>LRbZ<vsv=u$Qx)0TObyuDYz^32i)zT$+EhcfK2ry_
zK3fO2-l96P^)}U!ZOk-)ZOk@+ZM3L?Y@<yLWR)2OSY=iLR%wxftkNb0*(S|23Fs!x
zHVOD9we*(_JVKM&`kwV;+;!>q)Qd0wYG1!X?se?-7+3a(P9Gls{PU&%(|7N7?EaYX
zVzB)BEtRWY?0kVuOzr>h{%PZiJ~4IR;sgI+Tq++eKYd&{^N>qtx4m!MR^v{evEXd$
zrnN<IaM6{M6@#6aO35imX*mr^ltoAxIRhyxXCdX}9HhLQhg6UYkcx5<Qb{gBl4J>z
zEX$C}av4%Zu0X2FRY*0t2B|LBAvNR%Bt=#rHA!yLd)fFe&|mmg*9jUwzR{~ceMWuZ
zh)@4v?{Vz8(s(ib`Abdf-IMB1^xfxOb>0oetLMwt-}(d7!*^}itKVidPhYV9|7S5<
zAX&{;NH((#lHF{F<S;uRIn7Q;F0%`g+w6wqF?%3+&0a`8vk#Kr?1vOE2OtH_K}aEU
z2vXP_h7>VJAVtklNHKE^QrsMelr$$VfOFHFf|NF=A&F)YQpTKtlr?7|<;*!qd2=38
z!CZh;G#4S2%q2*YS%M^+Wk_Xn8B)bufmAhDA=S(^NOf}^Qp4PUq?i>*O)@v>y==U%
z($CG+C91}`X{=4vf8$pdrW(8Tb5EKs!?`C}AXz0VB%5S|WS8ua9FhZ)Q*uIbNiIlk
z$qmUPc_4WuFC?GjgXEX|kOEQwQcwy)3P~YIVJQqLB1IrYr6{DB6oV9(;_oGW-L80G
zbwzfua6>V=`e57Y=5K$t8xKruYU?{cntt~x@7&m_r*GT7<22Lh?{%ir)*qg_{&2_o
G!~X}Q3P*_m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_xdl_bilinear_instance.hpp.4827161718A2E399.idx b/.cache/clangd/index/device_grouped_conv_fwd_xdl_bilinear_instance.hpp.4827161718A2E399.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0ef24882f59298226648174268588a99eeb3b831
GIT binary patch
literal 21842
zcmY+M3wT?1ndbeSBT_XNnm{S(Ep<~0%+PkZxK<@u*1Zh$=>^Jotd(`Y#Z0quFH9_}
zlC7(cnO2k*JFz1n`KD`3OItUI>o`DRW|&Frc$sB)p8}mOopz@S3^3e=PMQ7e8P)H<
z&x7#$exH9oC$VMeoFjeqQ+M9^=|Ao0um(Rhd^GfkPu0=Uae)8#@FNcm{u6a{+;CY(
z$8ho7S6tDPAH6dF_Wfu6CUpCkf`9qJR`c1e?{*#iyThU1PJQ#TXZ#QBvHnTlug`Co
z4t{0%)4zG&<@Cm@kFNgiJC&C=GxvV$`Wr0|-f`&%ZymYepYOZd`K()c<j?zWJMrl1
zFYde|@Z=5qK0KBF?H@j&fAjFuzgT$bUk)X|_iFB^zc_up=Qlt9$`Ag0`F}q9H@Uu#
zlqPTau~VMC_O`X>c8~n2CHKdrdnffjf9(HW8?B%I<m>03=nGwOtMj3+zo75({OkQ)
z@Bi$Fy;nDX*L>{{-+X2DXYYF1eP&E~-^ZgbJzacdzWLXm|M;CZbl!dbl=0}nH!R;d
z^3tidr-CcWd!M)8c=g5aKR5KoJ=zbqKGED4Ts|~@$4~6)=l=00#oyfZ*603fesVqM
zyZpwx-*sc;(4Nr4)fwAQf3|VkFaPMLS6_ekUDv!g`LS;_{_f@P)nEJ7H*Vf{Pv51^
zU)G+#{`3E(^{-Do`Q6_gyDR&*ul)M+hi^=IAA4^}y62W3U-+V-dnNnW$G`L1=I+0=
zmaaMK=uG{r^Zw4S-#2^99se-&M&ZvUKKAdG%}a*sPD+}8&c3vM+cmc*Kl9eNkAFl_
ze)!izZ#NF#{+{!}qi=RU_tCF?XwNle>&YuW^MO})f9sx&;M*_%zWgsH%kDppKK=Za
zSFL>I{PZpMC%*oZ|Mh2==4PKP$E#CaUH`|P*%g0!@H2P)H~ClcC$4zr?v96JNBzIz
zfAQVFuly);*X{T3k>35@j?RO-4j$OP>A0M~EBL?t<CkB&C$s$c9_cC`-@5Bo?e<N_
z<tk}weBkMcYrgQcH(%N#?cwqMUHdQJzG;89@vUF{wElYR={?dFJic<*m6vVbbX+d8
zxufH=9dF+JzyCd(+}E*>$2VVf^9QzXwm<uQ2Riz{`=Q7G>b_uVUx%8<omX{!VEbl!
zeC7U*qaQhY+<EYG_w12&@%ZXpSHFAvrtQvnuer1Ju}}QP9_d{?zGBxEySH!J<Hx`C
zR^`;I`c6r@@jY8#OGiiNzTNxY(;iR#(Uac$KREP{d!&Q>*$?h|@Q&@9_MiLw-=F;1
zu2(Mm+`f+8JU)EY;TyMa+F#mV{ra|x8&@Wd?d!OV$A_*uba?xwJ$~JwtbFUm4}WxD
z$K^b}_Nr??uzk}Wzxmbs|LHqNKhA&WSM&HO+VvD^|MUOa{@?yG{BIw;|CSr}Nbl#B
zdw1P?$M#Ko{PzC7XMY&|_Un73y*xg!>%isPH|;;%L!t1)uiV$sAwLw?YhQZZxPNcw
z3>}-HWZS&Avq;B^m`il5gt<(|%9ty3tb%!#j?H48qhoWJ=jqry<^?*ofO(OQEn=?H
zu`1>o9jjrk)3G|{1|4f)Zql(P<`x}m?d0|kL5g;7fA_b)IresT6{)L6?Vke94eELx
za(9usao*jaZk%h1q`|qSK^mNEi=@T5wn18)>x!hqxvoJvoa>9E$GN^idYtzZsR!ph
z4eG(Up-2Xt8yaN5d2f+=ao*dYUYr|?WW>3#K}MXLie$pMsX->3n~P+|xw%1RoLh=y
z!MUYD7Mxp)WW~9)K~|jGie$sNtwA=N+lyq!xxGPlocGO8-y*et-1c_%)u``T$o(_a
zkMsT-_2b+zLk^rfYUIGVbB3Hach<;>^MM%}!1+Lp25|10As5bFHFDwHJwtAsyKCge
zxo3tvIQP`ZgLCf;d2#Nokr(H_8S>%WS0f+J{WIjpxxYq!oCjtofb&3&0yq!OP!Q+A
z8U=A4nxPQRLp2KFJUl~ToQG=^#(89hA~=uKD1!6hlQdkV_HTo|ox^n+eh%`;NgBcV
zNS#J-9zRKOoX6`F$9du;C2*doQv&D7la$1HlJDhkK6;WyaXwn7QJjy_Ng5;E7^6Bj
zxRc_Yl(3WHos_VX=AE>#ljfbYu#@4PjIfj8os6)P<(;gsljWVPu#@ARoUoJQot&_f
z=bgN;ljohhurtm(<HF83?~DsO3hyYwj>0>Nurt9s6T;2}?@S0gGc-#xbc%L<gU-;`
zxH$_=k#~y1PLX$t!cK{IO2STwcS^!enRm*<PMLSg!cK*ED#A{McPhfpEbq(;JF~nq
zE9}g1o)dQFcxO)7ndhB(VP~Fq=7pUF-dPZK7I<et*jePAMPX-=cNT@6D(_T<oht8C
zg`FDj)P$WH@6?2yI`7nlojUK-g`EcPG=!Z7?=*y+Chs(bohI)zg`F1fw1k}&@3dg2
z{rG|(er!F&Za@5JKf3<R?(Ij_?V5iJ4fZ_%zmtag9)^^mzC8Sf>MKy+moZOL-wDY5
zgZ;qjAL<8Ie}?*z)nA}~Wc5!{Ke8Ny4q!Qk9Kdp9$bl?JfgH$kOp*gx&Os-zoI_4v
zIWy!$ma{-kWH~3viL8Oa0bmUb4FGE(Lj%YfD9`}11}13$S*}4Buv|kfV7W5nLYAvQ
zE@ZhT$%QQUpc`23Avdtx8FC}bT_88I+>_)+mS@ldEYFY!Se^`dkmV_m2U(s;@*vAQ
z=mnN{$O|lQhP=q~7RZY%?<9GV<s0+?%Qxf$mM=p-Wcdo@LzZune8}<-`hn#i@&n7C
zAwRPG1@a@yKS_RM1qK7a3Je8+706HkS%CrtkQJDu0J4IEL0|=kg1`!9D2S|Jfr7{i
zPErtAp}`QaLPH^7g)$UER;WNBWQ8Uvgskvj7+B$<FtEZI3L`6ApfIw+lN3f)WH17(
z$WR1WkqkwU6)8{zS&>PK?63}A{o7BT``cd~puL^f*{-{*{cN~>u;82W?(L%lJ9l?o
zOH$X@r1qU8KM7lty3S)>m%27EPs?31GVC|yt`_DMx$8yDbyZhGwX@&7Bz2>&drj&_
zU-!DyjlS+_xf^}mO}QI=-79i8`nv0?ZuDuEBn|pBYmx?insrHoKFzeOL7%26YtW}z
zku~Vk)KwbvX_q7|`m}437Jb@vNsB)1w5&y+wkd1Tr(Kb?=+o9!TJ-6bBpv#6YmyFq
zx^+p1KHapeL!YiG>(HlLk#*?P)m1w5>6auu`t)m(9)0?CNsm7Lw5&&;zA5X`r(cov
z=+oC#di3=yNj>Q6S(AFu*RwA5ps#0I?m=HqQ|>`u&x+iGzMi_O2YrSm$$&n?nq)wq
zVO=ty&oC_;&}V4M2J{(LWCQvPb(H~qy-QLr`g+%-Ui9^@OTFmpotAsi*V~kP(bu~o
z_oA=2uIfdfaY-_w&$uQT(PvzjjOa5?%SQAWo3aso#ueF!K4V>FM4xF%GNI43CYjJ@
zT9-`dGfm4T^qHEn34Nv&*@Qk*U1dU_c}X&(&%7p?(Pv(l%;+;u%VzYMo3a^w<`vnD
zK671VMxSL#vY^khCRxyDS(hy6vrNkt^jVs+1$~wk*@8YxU1dR^bxE?K&$=dA(Pv$k
ztmv~&%U1MRo3a&s))m=`K5JcNMW1a+vZ2qmCfU$uTbFF;vrWr3^x2xS4Slv1*@iw_
zU1dX`eMz#T&%P$v(Pv+m?C7&k%XakHo3b5!_7&NVK6_ndM_*q{>N_RDL;O{#?;Pd}
zQr|_)$K}2gGVIsoz6Rzqa^G3ZRaIY2wX@&flKRotzbf^kum6J7kG}rnazFa|>vBK(
z`p?My=<Bbl`qAfTNe=WmRwW1e92X=9`W(k)2l^a!*?~UC8QFn8M^)uOpR*-7(dS&1
zoal32keujq9+#cybJk@i`kZHEC;FUKl@omfEolIK1FO;i`UWmY1Lzw#E)Sq@pe_%f
zZ{UnPfWCpMY5;w%mgGX8YgKZg&vij^q0e<(cA?KzmtE*{osnJWb5&I?^toG-8-4Co
z$&Ehu1<8#*_i@>cK6hPqqtAUtcB9W-Rk_jUX-OXRc~&J4`aBmT5BfaEWe@s1b=iYH
z&l%Z+K2KHUL7%rJdC})xmAvTlUXZ-#^B$MI==0WPFZ#S^WH0)>Rh1WgzLw-epKn$2
zq0e_g@}bXnT=t>QSC@V0^PQ1>=<`)oKJ@upk{^BkRmqP&{{_j9KL2sqk3N50_M^{#
zM)sr6Usd_h7idWV^aWO>0Qv$KqyYK?$K?R}0(CinzQ7qdfWAOg6+mCGB?ZwJT$O_8
z3to_d=nEc~gXjy^<skZkXXGIIf>l)zeW8{VLSJZA3ZXA_K?<QSbX*RhFI1O9=nI{Z
zL+A@tRU!0+TT&Q(;Z-S&zVHPpjK1)3IgGw=T@Ir!d`1qVFI-iH(HCh+5%fh?r3m^W
z7o-UKBFE(j`XY5Xg1*QZIfA}ORTV+s@F{6{NrHp$=cM5^%onBMb<8K^;b|H68}e`y
z^I3U#1#?w3TvP4rkDQW5&^K~U8bRO4MQH?mBPZk$^o=y+5%i6ml}FGwQdNzhFMdjj
zqc46=ilZ-nQHrB4enO6;FW!*j=!>6~<LHZ5RdMtsPDu&$CC*6+^d&Az3G^jS$O-f%
z8gc@CiL-J7eTk|nfxhG^DT%)1IVp+0<V7iozT^oxiN0h*PNFY)R!*WXSyd&`H+o7M
zMc?Q-X%u~<7o}13jh>K4(Kp(VN6|NWRvty)XjL_ezA-u_jS(8h=o~lb9HWcepmmH+
z$YX@wF=}vw<}o_U4Z8Vcj%tk1o??4Su&3Cb66`6qrv!V7?J2>YVtY!kr`Vnn?0lL?
zl@jb}wx<Pqn(b-9o@RSmu&3Fc7VK%Zrv-bO?P<ZzC!kbm!Jc7zMzCkto)PRBwr2!;
zhV2=_o?&}NuxHqw5$t?wOO+AqS+-{ddzS54!JcJ%R<LK;o)zp_wr2%<mhD-=&L_=O
zS;3xTdrq+D*q#&YIkx8ndyefn!JcD#PO#_Lo)heRI!=`n?0L551$&<DdBL7%dtR{T
z*`62ddA8>Td!Fri!Okc8RC&QZ&h~M^KF;=W!9LFRalt;$_Hn^J&h~M^KF;=W!Oo`;
zRpWwPVY?#O6}BsaU17T-*cG-bf?Z*|BG?tSD}tR*PO22aKEd`0!9KzE3Bf+W_6fm0
z!S)HkKEd`0!9KzE3Bk^%JyjEeeTJTuX6QKyPjtS-&9{Z-d))lJ(0o;%p|Xsh^)xrn
z2+etJz9BSADn5;>+WD*^+lzv|NH1}N_9A_c8?+bct8!7W7wKtk&|akT+@QTkODaA=
z3-%J*OM<<`_L5*PvArbNOKdL*_7dAmg1yA{l3*{folpIOz0CHqU@x=1EZECzFAMfE
z+slHz%=WTiFSETY*vo9^lgePPu)QMKD{QX__6pl8g1y4_ieRs>y&~8vY_ACR3fuW~
zHP~m_J}cN~**+`SXW2e0*k{?!r^?|wYL@M@f_;|lvx0q=?R+8}>~n0N6YO(rpA+nJ
zY@ZYCb8Mdz>~n0N6YO(rpA+nJZ0A$#V4r9EykMVa`@CSEXZyThpJ)5LV4r9EykMVa
z`@CSEXFH#~2m1os7X<qP+ZP1;0^1h^`vTh+1p5Np7X<qP+ZP1;0^9kr0I)BzeNnJ4
zvVBpoFS316urIQGQLrzveNnJ4vVBpoFS4C4H~@Q<?Nz~EWqVbySJ_?_>{Yf`1$&k4
zRl#0mdsVPk+0K_TfW5}{nqaT7y(ZXeY_AFS8ry4vy~g&MV6U;gCfI9i=Zhr3UT1q<
zu-DmM7wmPm*9Ci>?RCLkXM0_+*V$ed>~*&D<riRYu)QJJ8*Fa~_6FM<g1y1^hG1{7
zy&>2eY;Oqm2HW|<4zM@b-W2Rjwl@WPlkH8x-eh}Ius7M>6zom5HwAl>?R?1y*jsFG
z3HBD-TY|mC_Lg98vAreOTWoI$_7>Y)g1yCdz8K{l`=_VRe`!7QlYJ&W$*NG-<9Ma3
zLfx1(3TZHF71CnXDWt=!S4fY!N1-0f289fmdll-%Y*fgI*`$yOvsocCW{W}=%vOc0
zm~9H#FxwTfWA0O^?{U0NMxlPp4uu?;oeDWI4=6N%*`<&Rvs)oIW{*N1%wC1On0*TQ
zF#8qqV-6@3z#LR4h&iNC2y<AWFy@Ft5zNC14L^=oIw~}RIj&F~b3&m6=A=SN%%cj8
zVjknG?8XS26gMfMNpq7HnhZA?p~-TS6`CA3IibmOlNXwCZpMX1;YJaf3C<HjGeckG
zYx=$<T+{bOzJd>%67Q6Rrp!%QXe!)Pgl3kTS)rNZW=?44xtSN51#T9EW|5mkp{a6H
z6`C41HKD0<Qx}>BHw~d_a?=!=7B?+u?$$gRz3SG-p5W^`4&HIl+P<j1{VEUh?(J81
zz-1kq)b$G9!)#JFX3ZvPFl#qSi&?ixI?Vb_(qrz~q#n$MO)_Ba-K1X3#!WI}Hf@p#
zvw4%um@S)R!ED_mD`wj!*)ZET$&R^igZh4m_b?mOkJ+(74$RICa$+9XpaIOT4RT?2
zZ;%_aXM;SLy&L4k?Ast8X8#8HF$Xp%fH}B9LCm2I3Ska!P#AM$gCdxRFVXOi@E+z8
zjbM&nqB!QnB}!mUUZN!C(MvRnd5kXcJq$J}Zc;*%<|ZvP8E!H{ljSBWG&yc^LX+nv
zFEr!aj0=szjUqG?+)N0~41J&PVO|sNVZP7zFxZrMrzA9GZpuPa;ie)qv)s%I%^Wv#
zLNm|JywEIgvmi8!+$;)Bm7A*2)VQe$O`V&%&@{Md2u+ilrqHywX+hI|MHpWvyY+Ig
z_FaklwKqQ1vG)rvs`!Gk!`E5b7nJSoJaqFRd;7ImdpkQn)cMIaVQ=S6hi|g9U&iGb
zE=IlQR)5U*Tvr^p;$ZuwW1+jv=Tcv<`tRF%eH6b^;ONeQh#88`&=)YL`QJ@}f43QW
zf}1I!ndYVpO_3r+q!uY!L~4=0H;V$b$kd`hEz&eMNG(yMgwzs6OGqs-wIonWOf3o2
z5>xr01$?b#ij<LBrf3<dWu}$|YMH5Jfm&uNAB+HMg(4NCRw!CQYK5s4fm&f|MW9xg
z$_E*MI!lpRq|Q=w7OAsLofW9FOq~^|vrOH(&VJ|Lc#a}-NS&kT98%|)Iww%)m^vp=
z=a|X|1%NtFk$I%fQ*<7w^Gux=sPjyn7pU`0<tO_<U7*MUQWq$?fYb%1E(p{GrY;E7
z1*Y<oc%UwBy;uf*3Kl84h}1=<E(+8|rY;K9MW*tTbD&l!QblT&qE)0;nOYU7Ri^T5
zS_H;4H%R3d^3;%8qi7AOHKx`CYK^Hifm-9Qm7h$5r%sVNQtK40BX#RFF<Y-6+4-)i
zGqo;I>rCY*$*|L)NCT-2iZ+niU}{64HkjHFs12s_lV6}VDbhr0lcG(eHksNKs7<Ce
z1!|M2{G=ABEsC^|+M;L+sV%0q1Zs<^ErHr%DnD6$huVHAFF!=wdL3{3ex09$o~Ewz
z_z>|lbz|0?CJkooY0_fWohBV-{b|x;?m108m<^}NfVuZH^<p-jCL?CkX)<9ppC&VA
z%W1M;ww@*{X4`49VYZ(pJLbM+>N|%I5tpeSvtyYYn4Qby#5}M}1DIXQ<ihM;CO2l!
zGI=n2m&uFSw@g0F{$=uG4lGjub8wl0m_y4H!W>?vFy_cIMKBM)K*MYJ5b*^X!5n{q
z;+PXJPy%!E1xjKbeSt<XkI@VK5D}XcH+=3CoM~?Od?qv*ZulG{G+A!=ydN|<Zundr
zG<k0L{1`Oj-0(RoXcTVvJQ6e$-0-;_Xl7`YA4Yupogbd97i+KbLjY__-0*n`*eP?v
z=L(>yaKp#%p_%1|kEug5#|<CnhGw1{KK2aF0ylg-7@9?H_!uoTRc`pWDKs^1_*f=1
zb#D0hA~X$d_?RCwO>X!&95gL%_}JJx%_F}zJ-=}H>0Mh#u+&{$a8bXy8<Iw?fuvPy
zA?egQNP4v%QjfX^l0j{N)T{1=WK<g=nbamoX0;iTMQwp(Ra+t1)HX<VwH;EQx~~f^
z+En*La;P1UoN6be0rdbRm)Zr%t#(85s6CLpYA+<8+6T$6_CpG&1CWC1Af%8w1SzZz
zLyD*)kcQR6U2sv3dIVBj9fy=qCm<!&Nl2sWQAlH?9wW#pQl~JdNu9==A$10Gmeg6y
zIa22^=SiK%JWlFy%nGUbxa!WIYJ$`gm}f{mLtQ(M9*U&qcOclwB~q6#mq}g5Tp@J@
z^DL=nG0%~D4)Z*z=P@skdI9qysTVO<NnOQUBXtdPoz!*A4N~(vChUCuO;R^8w@BUE
z$$xQ|?ZQX;yYJ)YVK;y1W^Mb(QTzEAzYcQieLUdd!I7>9c7DRUqFoO|;(y0HJap(P
zP}h@~r>JWNa`%yLV0A~kfz_R+Ze(>As2f?`Q`C(t%@GZ-G*Jz(G-=WxOH&{XvNThq
zL6-K27FgP-7FgOeX_2KZkQQ0mDbgZKcSHv)T~r4wU7B>r(iKRDEZr38kflGO2bMmn
z2bMlfdSvMfq(_#1iuB0pIno2Ho@fuSdeYQ`teyh(AggDJdXQx}VgQyQY5<lYO$KBc
z3S>Z*VTugM>OIm6tlnrZuzJ(fi>%%P^&+cxih7Y{JYocvF=_;sF-=Bf84F}YmT`)V
z$TA%<0m~FM0n3yo6S7PNG9k+}MJ8mKkC=gFj+%jGPLml~<^q|KWu784vMfg|z_LUw
zz_O&tf-FmcEXcA<kp)@SBUWHpqgG&9(_}@KwLn&6S*OT~EZY$quxwEquxx3vA<I@E
z8?tOuWJ8wyh#gq=s2y1LG})15FOVHs_9?ROu*QFGzfQaH*fo4_xBtNY_qSi|{e<qj
z&uIEyILGx(wwwN>eT=I8JA;pXZM_{)`%ZA~`fu+4!4Dq!H{L$<$wQxOAMI-Wzf8>W
zYO%nF#BMd;dgspF;=QJO?K=+|9yC2@Z-e)CK1JrI=+5??^{L?(-*w;7xjUJB<-4!E
zs{QiJ_SeH-$JYBEwZD$GnCq?gJ8IYe{f!^3KJ&xnPcx+RQ=P8%;kW;>T)*k|FP2{A
z`szEbHnpGs{lkyPOkY0m%*$NgtiRdbzW;9@&|~)2n<cf`?Jtz;|9F$6_MrXe<H6Qj
zBem=HyCQA9%g@$(BDH&dq!{aZbmy<AJJt<J6VpJ_#<Y-hF&!j*Ob@9i)&t29GeGK%
z^+GbnjF3z*6C`uY49OC+K(fZHkZds<Bzw#bsV~;|X!|c^Z)bn3ACe>HfaHuhAq~U^
zAh}{LNbZ;$k|*YY<c)bD`C>ju{+J(9AQpfWj0GWuVj)Q3SQt_y7J)Q8IQ+oQxAE}M
z@WYT|!;eDB&~P4}Obr)k_{*55Xm}d($lwT&M}|g#JQ5oL@<@h8kUUbL5hRaH(Fl^`
zgK;3ohvGnv$KpVaXDE*3c!A<bj!#h>$%(-PkP|}*ASYr8ASW`EKyso$2_z?`D1qeU
zU=qm5p(K!#u_Ta_8A>8KS)e46lT(yL^620wkVl6`fjk-;1@dTyMv**Ppiv}`PSGfm
z$7pbjfIUV-V+8auij5KQ$M~uJ7(Y?p`JNf00ypRw<D*%8={D@7XfTDA6b+@&lA>4&
zEh)C71WSr7DZ!Fr3tyrQmNX5f(UPX2G+NRWOQR*tmb73=vn4H9(rn>NvB8p|!3<h5
zG?YP0hGH4CWZ04sEE%?B1WSf3e91LfvNV`QOO}SRXvtD6i<T^Zepaw#*^(73S+?+{
z)nLidU=A%g8p@$1N3k4Qa%{;7mK<Ahf+fcmzJwYqc^b^4B~L?nwB#w4M@yb9dBKur
zOJ1<>cOW-t8K=Q<w2afxI9kRjHjb8Ywu}pwakh*LmT|W5CDHI_QD{&>i$X&RS`>;Y
zXxX|RW9#aQou3zlEs9`K*us}S!_EW^PM~Fih9=N5L9q$6Ot57_uuQOJLa<D*g)eb_
z$I|go=mFo>J3Bu7P~hSFw?4Rj>*hcHySDv7N5^$LA0FEI@S2?uZ+PcJsdM<;;n*MC
RFMf8{HHWUh_PQH7{y&wq954U?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_xdl_instance.hpp.DE209B7BB2124A4A.idx b/.cache/clangd/index/device_grouped_conv_fwd_xdl_instance.hpp.DE209B7BB2124A4A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..44b29922f7d187ad8964b7dff3f70e7dc0121a5a
GIT binary patch
literal 25076
zcmY+N3wWaineY34UmQMW2IM;8C<7D`L_B&Fgl=o^O`9U;;DVmBr?&K#CSO}lvkc6T
zUAnDF@3*5qJ{_E99a?)!nl^D@yTAqwwYB#)ZS?HAuC5$BsH;9M>T~qy!9D+0v;Dur
zL+JbF{r!K_eodMtZ$AFi<#hhEE-R~~(e?D>?kDRUSy@@v;@`7RK4Je-va;^JA}j0Z
z&?{f4Zu`d%-O%&aBQJjc^2K-j!^7@R-uUI@y^l9k{OWrje)Gz&{Og7K-v9B*w|;*1
z^G&Yj&wc3elH#QQ7iTX0z5n&+#(rN?`oBJKTk+Wk9{%Uw{oqX7CtNL;pFRFJm$y#5
z_m<dSywxjjmY=xzz8|&!W@+TTzrN$lN1xtb`uXX+r#|reufF#DsWYF;{qlL|8#On5
z?EY_m$^DG`wyVFsa{g1vn%nMr__bfY^vkT854C^mC%=9CbL$_RJooM5=P!nS_OnwD
zy|HKA_`jdo{c_0*&wS_ZPt8AZ{L0O#*k)VkpU1y=`S%q+Kl^^UadWHJy!UI)TZ+Q5
z!QBVH@lxttFEs9acjk-h_rLGIjX(RV)xkecT=m$Q7e4aN`%Ycn|7792&t><#+4@ZL
z6+c{f<mn%u>$&IU4|F|xb64oadtUhE$Nx+5^}oNA_||{E{ElZHeQ^Jw7e8|9fvaA=
z`Jw+4I`iV>Pu_ES?=MDwH~+%J-*Eof#ZUiR(drjBUi<#Xr&3@2SpK!wS>JfzzT>`s
zxykdNej2~;FW2r$fBei}`(E@t{_d&=pQ~ECx8j~B?>T$*1K;b}9sA}#cElgP|G~sp
zauyZ?ufI3%;-f$Q#INs}I`Qm3bd|sUi$}UIU-z}q+2eOscRljZw_hH(|D64$_gwcM
zYJd9Ui}yU{{<pJdDjt5W{jCRoy;OAj{~cTZ)96P=Z~o5J&EI_L7r)&4!syFIiPH~1
z^@GUHzFqy?EpI*cd_&e*|Ks%^HL|jf|Mt1>1s}ZckrT$-Z^+6%bL`BuhZk8_;`;-C
zPQ3r)!KEMl{a>9h-ifOpJ@)7a4==K=bb!ToQLO*;Qw#s{^%KSkTz$i_8?HRO$j~2{
zo_o5c;0N!&D(e`ozWvng*@qX0SKoMT)}QzyFMYoAjo)0Abp@`z<J2ARIlRbRUGTd%
ze>VGuKX$^n23Oy7?4~OYFS4#Q@jaWB<;wX|bouh=?@k!+!qv~pv1jE&fBaA8ulMP2
z{J;O`&aYlPVY~w)A3pZs4;@})uC`A<_T7s8E8w?q^{uCFeb3=V#^3zT?B{bndU@ok
ztSfQ#ou}@+?eHRVb+GodbM8NW<lG74ZMgdSW7i))yg0lXf4#ove|+qdoxi!|s;sMV
z_1&lLe$U~>;nk<E%Nkz#-E+~`a&I+^+pj<PY_hVluR4C!bs0MT`78hS*t=q1JYie~
z`kG_cSceyxt6v<y?#)lve*e}J#%WxA?%25x9$sWVv7dZlqig=ppD<1sH{t4+<=D$|
z{~!O88TnQG&C6Av*?Zp!<2Xi6A3J^R;YCL9)-|8H{N1+YH%=HQarKSIZoKmFB6IZ<
z?xwT%RcB?HPyA(W&ez{4xaMT`B{_dd{wi}&PG*nG`Ejl%<opEJlX8BN>nS-u#r3qD
zpXPc-&d+c?E9Ymq4$Jv4*AY1%;d)Na&v6}<^HHwn<@`L?3vzye>zJI6aUGZQ@uNC(
zGkQ~wAKrd4hxcT5PDpZQB=en!IxIOaaJ?uwuTq^Gl3bq84NET1=PpVv&*z0CkLUBk
zlE?FTi;~Cl`60>Y`TVft^L+lI<nw$%ND6qqAS?wuU$7_zJYN`+LY^-SOCiq}E=nQK
z7lovV=ZnHp#PdaqQpEGcAt~nh;;<C+eDR_b^L$B2N_f5`EG0Z&vM41yUmB89o-Ykc
zDbJTKN-58mg`|w<%feE|^JR-t#`DgQIC<U~7AMa;7sbi*<sm8O`SP%o^L+WDl=FN=
zNGf=~A}kd=U$H0^JnssLi|1Woaq+xsQCvJ<8Inq#uMA5i&sQ!=CC^udq-sVoKSxex
zSB0hO1+EvR>Q$<%LsHH2)nTdT`RYZf=J}eC)bM;wSZa8_W>IQ*zBVMaJYO4@TAr_6
zlv<vz3rQW%*M+5y=j#@wj_2z`QqS}CVX5c&`bDYd`G$}*@O(p98hE~8Q5tx@F(i#V
z-x!ugo^M=~MxJ+v#Le^Wu()~Ny(n&;Zwg5h&o_moiRYUZrHSV~A@T6MCoCSG_biHs
z=e;5E^1L@JUY_?ZikIi@kk~wLhsEZ3dr@qjZw^T_&o_sqndh4qrJ3hjLej$XEn#Wl
z`Ibd#;rZ5(wDNpwSXz0$bx~S*zC9%EGm`m%eKNZ}EbT9Fy(sOkQr!`f4xaA_O9#()
zEJ_E@cS=Y)g^5lHqr*(6ETY3y08~Ix0Z;)&1waK96$BMjR1j29Q9)2aMRkGdQdAeH
zE=6^L>QYoUsBT4dgX&gPH>hq!^?>S8R1c^gMfHH{QB*IeUPbkS>Qz)Ps9r_&f$CFK
zAE-V>^?~YBR6nSGMfHQ~S5!Z!enky{8c@^#r~yR{fErNLAgDn_4T2g})F7xqMGb)(
zQq&NrAw>;=8dB6Ss9{A7gBn)UFsNZgjer_a)Ci~%MU8+OQPd?l_$MMqk2Wt!6rI0S
zotMyArp`F(aYc=T8dub~yo3%@6QCv(H34cuQ4^ph6g3HIQc;tjCKWXaYEn^Cpr#Zx
z1!_uBQ=q04H4SQ7QPZHN6*UcNT2V8gW)w98YDQ5rpk@>`3u;zTv!G@bH4AE1QDIPF
zMTJ3y6%_^*R#XI3L{Sk?5k*BnMHDp$YEDsepym`c2Wn1H2j4vh*+<_|QPfdIML|Us
zH4kcDQS+FXSJXVHc||RNT2Ryis0BqWfLc&g3{*@}F;Fo@#X!Xr6$cepR2)=XQE^al
zMJ?f0xr8I1`91Pv_7aZG5^j~;Sw!bm>MVm=R@Cz05oKx_x8-F;t$<ol)C#B-<yrx?
zqNr6+tBP6$wW_F9P^*es1GT28HBf7cS_8GFsC7{5idqM?uBdfT>x$X{wV|jDP#cQc
z0JWj0O;DSP+61+!s7+9tirNCTrKl}XTZ-BOwWX+SP}_>y2DPoIZBW~a+5xqrs2xx{
zirN9Sqo`d_yNcQcwX3LIP`ipsfl4VV1uCVe6sVM<_CW0^Y7f+&qV_=TDQX|ozM}R)
z?JH^@)V`w9pwfy;gGwtZ4Ju7k=64}HRdDcxTIQ*O%yVg<LI2>Hv`inrGTBwF83>iH
z>MT`3sp_F$pQ<iM)kUhS?P_AHebvNP2c??X>I+iMY>iz*Y>lsm*qWf!Fk5p$YM8CH
zYl*G()e>79lv-wMFGwx3b#@)Gb-p@c>w;3pY~2N^W47L|C$`>KPi%cq>Y1&-Aoa{P
z*bT%s_!@|92ucI94Hu+=*+#pO*hXI?v5i4#WVZ2wG&1Y9-Nd?mZera*aWm__AZ})x
z>?UHHd`-kQ1*M7ErVG-<tjG2c>+yMr^#sMktmlGwnDyFTV!b{uvEHC~ne|=}FSE97
z6KnfyV(p;V%-R>kX13XGCbrqvOl)&dnwf3BAkECS*e%4i_*#f<2}%pIEf=JP*;c!i
z*j8UFv8_RAWw!N#v@*LS_L8*nk51o`oaH)*AEiC?qjX6wpmUKr%VIBc#<KX9Ib#{W
zbS<llWn?U?j1{q0IAcY8E1a={j1`r!f{YcFu`2c|XRL~El`~e6v8pmwk+G^W*2G@p
zj5YDCamE@l)>OtCGS*bay4dTSu`a%K&R9pry2@Bb#=6Sb5PO3&HpI8V85_vhP#GJ@
z*iadpVsCQBrua5FV-p#hDq|BFn<`^V>@CjN65keQY$0PyWo#j1OJ!_}z0Dch;@jqo
zZDee#jBR9WtBf77cQ|85d^?=6gNz-Ov4e~qm9Z=KE@$kDZ<jN6k+G{Xc9F5GGE!ov
zI3p##6lbK6ky05cWTaHap4fYwu_wMg&e%i7p32xm#-7U97ki&G_Qki)8T-iCR~h@r
z*jE{8vD2K97GIh((#S}wj5IRRlyTELe_j5{#;>lGli9bG-u7VTMXStfYk0J9@WNW=
z(L&}mHM9?2P|LJ4ucdv(as2Q_v`inbN5qVrxIwQ^B#oR`xlS25dt8s1IU$qgqh`)L
z*9kM{6|N(WoH@tQd~VFhWnXU6$Yoz{%E)D3?wFa&zTBvp%f8%%najT1h$EMMc`+l8
zeR)YEk9~P5BaeM~V`d)v@}g!Q`|=WI9{chljy(3|$BcaT<tL4N_T{IHeD>v!nfdI?
zkDB@H%TJj3?8}ch^4V7qGYZ&OkTeR|SCBFa*jF%S7O<}%Y8J4sAYm4;uOQ+mU|(U(
zC}dw@(kNtKVah0EU*VWp$iBj;S;)S^gjvYG!ib}geMK>&h<!y#qlkS)DWiyeMPp_W
z`--Aw5&Mb~W)b^}B90>V6~~NX_7x|MV)hlMjAHf`kD0~nD~_7Q>?=-~#q2ANIEvX<
z5;IEJSCTYJ*jJJ=O4wI2W|pw8Bx;thuOwlXu&*TIC}Ce|%qV4FY0@ZVUunuHWnbx-
zS<1fBs9DOs(u7&czS4-Jlzn9}ql|rJNu!K?WhtYKePv^28T-njW*Pg+5@s3u$|8<3
z_BmsQlYP#l;bfmPWjNXA95bElb4E=k`<w~W$v$Vq;bdQV%qVAHdD19nUwO(XXJ7f4
zS<b%ls9Daw@`PE=zVe8poP8BBqk?@ENuz>&6)B^FeHCM71^X(ZW(E5y5@rSaDk6>w
z_PJt)i+!%7;bNaFWw_Yq8Z%w&b45)T`&<dr#XeWW;bLE9%&25vWzwi*UuDXuWMAc&
zS;@Z2s9DLr%7j_TzRHNBl6_S%qbhFDFVsn+>Q$~&M%5nIV`f#zr1_{>HP3azta^p(
zh@)!GaWr2YGpgBFoiwW1SDiAd*;hSgR<o}<YF4waI$>6`uR7wWW?xOrs9|4C(x_oy
zP0FZYU(J|V!@ioRS;M}XgjvJBnuw!@eYG*8mVLEJqn3TODWjHswPR*2`)Z?RE&FN{
zW-a?_BaT}3)y0fD_SGegI`-A2j5_w!jhS`qtBaa-?5j(db?mE)IO^C}A2aINSD!TM
z*;k)3>e*L6X4bQ>K5EvpuRdYcv#&nlsApe8%xGX=L(*tqUqi}hU|++S*}%SrsM)~2
zhJ@L`zJ`dSfqjiJqmg}$Nu!Z{jVYs%eT`#gBl{YoW+VF=6J{g(8Y7NI_PJw*n|<!2
z;bxyZWw_br9y8tSb4N`#``iiB%|3U;;bvb`%xGd?Q_^T+UsK9xVqepk*~Gr4sM*B6
zri9tVzNUzyiG7}!;bEUAX?WP@Nf{pYdB#i+`#e$8!#+>K^svtpad_D0jTv6{d6R~h
zecqJeWuJG<^s>(zHNEWfCQL8;yb*_&eRj;S*=HvWn|*f5u-RvinKt|EsA;p$PM9|P
z?1;l=Uvtc8W?ysCXl7q?%4lX^^O)JpzUHXe%)aJ?+04G?h@+W(Eit2oeJx3&g?%k4
zqlJAfV`dBcTB2qP`&tra3;S9kju!T{#*9|>wI+>L_O+&rR`#`ynXT+=jhe0OYfYG~
z>}!oUTG`hgGuq<@?I$FS_E))18SQ&qkD2Wuljfsl`#jePv;7sW!;bcd<7mDkW^}Nx
zBWZN7uOnr2u&-mx>|kF<)a+niN5br2Uq{%{!M;w38J)t$PD!G}&Q3|8!`4n2GqG7g
zpGT)e(P49^B+y|uHZL5V!u9~{0c8)sj;#v%iUwd0D0={QY-o@@0DC~$1F#2_9UB~u
zfU*Z+4=Q^Qc5G*m9Z#j8!=HN)c5Hf(JqUYH*@LhLl^vTPj-ax4!QQ3pU9fj4dl&3o
z%H9P#Hb%(a1$&pWcfsDJ?AR!AbSZl`?A^-V4STn;cf;PT?A@?q^Mvf(uy-qaH|*WY
zj?ENDx3c%Z-lOb2uw(0lzVbb=_b7W0?ATx-dk^eA%H9KekFsOK#nGeey|DKxdoS$0
z%H9imud?^Tj!hV{_rl(*?7gt}Dmyl19KFik2Ya8g_rZ=W82a4%VDD4*KG?BQL-sz{
z`;@&8_C96D#*L#-+52JdSN49`u}wqve%Sk!y&raL=8(M~_I_pWhrM6fvAN^uSM~we
z2b6sPc5LO4eE{|WWgmbY8$M(ofPFyO2Vfsic5DDS29$je_CaMIgdN*GWFLfmP}v7z
z$EFb32Voyn_CeSOl^vTzjzMJ~f_+HYhhWE+5ZQ-dA5!)q*s*a$_957Zlzj;HA!Wx#
zl4D5OhhZO9_F>quZAA8A*oT#U7<Ozfk$o8UVPzkNeOTGC+2j~j_7T`elzjwtY%P&}
z1ojbSAAubkP-Gv0eMH$uU>{L-Y)Cmql>L%iHZIBY#?ki5CHXo!%c_$`=bNhYB@=s8
zCQ~tV;;QpCbdstw@4)_w<A@rEeO%ecVaFB~xyB`p4%^4&OC~n5D0y6B=&*fUzJ?Cl
z$7S9zuIv-APbm8Y?AXR4`vmM0%02-*HoM3^0sDlqPryE*>=Uq0DElPrlgd5`JGQ#W
zJ_-AzvQNT}4KcD$!ak|&ldw-J`y}j>%031Al(J93j_okAPr*K=>{GB~(~Rs>uum!b
z6zo&VJ_Y-fvQNW4t?bjVW6O-})38q~`!wv>SR?y1?9<9V4g0jRPs2W~>@%>>DEkcT
z*j6L^4D2(?J_9>8-^e}#`;4;Bz&@kwGqBGn`z-9U%03G_w%*7-3;V3H&%%xkI<n8g
zKCA4ru+J*{EbOz&9)>-v>|xlkJx6wIp`gRJxG?P4#3Oqc_OP;tVGk>N81}HTM_`XA
zdjxiD;gLN8dqmkIuw$c->=D=_${v9|qU;gaBg#Gp`<$}R!H#V{vd_Ujr|fgEV>6KK
zbFj}T`yA|Z%037CoU%t@k1Bf<c5DTbJqmkN*`u&y!;tLQ+(U=&>!Pqnl^vUe4(tWe
zJ?%W~^U6LCJGKkSJ`el6vd_bgO+~WL!#=O<^RUk=`#kLP%Dw>mg0e5bjx9y9FTlQ_
z><h4C<B{wOurDb40_+RQz5x4zvd3VLDSHfdY&()Y2764|W3Xd$lI$_qW6B<bJ*Mn2
z*kj5bhdr+BaoDjnN%lDGab=Igjtxw*$6=2vdmQ$-vd3YMEBg{2T`b|z1+R=G(Ro#M
zQt0fd&X|c!PLru9I`gWNK<5?Ji8z*I&T&L7!@jKS%Xn<FtnABpY_qKF%Xn;qjZjKn
z#$%ghWnac)n`LER#^ai0WnY1PMcG$i$2KT^rYo?oDEkWR*eoUc3hXP&z5@G-vai6t
zqU@`%uPXa0?5oPY3j3<EufmQERkE+bzN+l2u&*lnD(tJuz6Sf6vai98?Ns{Q*I-{$
z_BGhC=}PuB*w>VO4fZu<UxR&3+1FuTSN3(-vE@qkb=cRHeI0ge%#wW__H|`nhkaey
z*I{2*_6^uKlzjtsY|D~;1NIGN-+&#Pw`AXdeM8wdVBb*o4cIr7eG~RgW#5DyTeoE2
zgnd)lH(|#HFWEO?-&FQZ*f*7Z6ZTDI--3Nh*|%WF_Ac4CVBb>qE!eRMO!h6<x0HPg
z_AO=Kf_+Qbw_)E__HEd)1x)sB*teB^8+L3IlYJZZZDrqveOuYLVc%Bv9oTo2eFt`I
z6O(-h_8n#4fgPL4WZ!{(N7;8^-%<7**msnD7xrCc--R7p$z<P!eOK9cVaJ9u*>_>z
zRrX!jca?n?_FZL9!Jbm~6zte;CVL9@l(MH_$EGydQ?RF$Jq3G8*;BBmlzk8OJ!Rj6
z9b3|5--CTm+4o?_#x>dZVBb^rJ=ph@eGm3MW#5N=U)lFz$F?=u_hH{x_I=o~xlQ(c
z*!PuvANGA^--mr)+0(G6l|2nRwzkQhhCQw9Y1pv=PWCkHX=P8to>ulW>}j%Rev`(^
z)*sGdkXg=p=>E^Qo%+DgXR-Np=7BRMnfI7ycBwFba7u&Bbj=Tb=*l|z)ZaO<CzgF%
zab{2KWcKZM++LR17CV`J$DMbSXO_fHX5W>4S8?V%2ai1+YO8%E@j7(ZUwi#cne!Vo
z{9^y1r>=kOBgV<>w_kyiMEvp3$?T+Dk(7V0-47ML@|mwp*B2g~3czzo9?yJQ2X)&e
zd5Y@*-j>owZ??IFowrN!*VGvo&p2m{OWQbSjN`3|<0@ku8RIHrLOc_kF(GXeoH2on
z36(K{j0u%7DV|Brn3T3j&X`2Tq{^5?#-z%a63-N8Oi9}mXG|euN@YwTV@hRAi)WfM
zrloC~Gp3O-tum&OF|9IY4&JFt$7M#^W;kO888a$l1{pIdV^%z~oG~kHvz#%Dj9Ha2
zi;P*7fp-XpIU_7>Va_;sBkI8$DUZHm!pI1#jEHz5oDq?>2xml)5m6ZtWJFZPoOtFq
zV@}%UIAabOb1GvF8FMNlDxN53M5Qgt8Bt_JRYnvUQI#<-o_Wrgm$rG%m`BFE%9uyS
zyvkS*&jM#GNZSHuEFfb+Wh@|LL1o0m6XT4Sw8c0hhK!iXh#@1UGUDQib4Fa+;+zpj
zMqFjYkrAhi%pDi*-VRR5lDT`!+)JT-a4MEe`^n!HzA%{`JN7U(S*@IVGxv=rvvaLn
zs`9Kns`9OTstT+EstT<_s*0>4s*0^*s!FU9s!FX=s>-Y~s+^XSs&cEGstT)uDwpM=
zs?w^Ys>-Umm-dLPYN~3i8meloTB_=-I;!feda4?%2C5pZMylMFo2n+Oi7Jofp~`D{
zsj@Abs%ER1surt-s#dF&s&=dWUV5F_>Y%DqtWKdiAXb3upjbh!yTt0^x?8Mnu6xAl
z;ks9>UatGZ>f^d!tbVQs#2VmwP^>|&hr}A<dRVMsu1CZg;rfzTm*n1~JG^nR#<`vl
zYl7=Zu_n2m5^IX<X|blco)K$?>shg8xekjJ<~kx)gzGu6=D3cE73F$fta+{%#9H7w
zCRU8=xL9$nm&96<d-*vBYnkg6u~xWV6>F92HL=#XUKeYf>kY9sxZV_Nlj|+9wz%FF
zYn$sGv39uL6>FF4lvpXQ_r%)cdS9%4uG3<rkLq&|mTtZOzkc*#Jo38x&bxCmr;f@z
z0xSB%nWZv!dAO(Z=CmH&)8(|~oTVxtIX!efp`204xkz=cH<#Glwp?O!1Cq;Z?x^H4
zo9E3VHm@y@*t~$`F`G9kdCcZ}^NG!G%O^HJAo<MZk4iqX1>OQ;3)%{ZEeJ>fvjwA4
zz-*znkl4buLShR8QpjxKs1!0=<Sin$sI7?DqJR`JTQn*~%ocl#i7jp`Cbl>r#mp9u
zN-?t~-V$O<+DeEm2}lXEC8JWpY^k@D*wVIAVoL*3%53SVlrmf9EhDz9t&G^RfRr&?
zHY#PzI=xO}oo!BHodIz&>l_s)v*q4$V$0jgi7gLEIkV-XQqF9Jw}RM;whCe^0#d<j
z#i&#;>+-sYb+x&Obp^!5tZP(U%vO3UiLGp_B(^djmCRO-O63vT`>$p1&e<J0ja&N-
z*WU1s%=>;bxB5>UKRi(!-P+TV^BUiR(~`?|URv_F&QD7|*9B=Q;JPp^g<Kb<rHJd|
zv=nn)l9m##OVd)yby-@<xOS$+$#r>J%DJvcO9j`iw79sgOiLx#RcWbujo;FomTInR
z(o(~9ZCYx%u1iZD*Y#<s=ei*+4O};-rIBlQTHIVWrKO2$Pg*=&d(-0O8gBrzxo%EN
zGuJI?Y2msRZy4jcJuU68@fm{B(!q5n-Y(WD+zFr)P@RJ}nFUp+3!N_2=|-npb$Zb0
zQJr3NdR3<noj%p+N2gzP2GAK$ok4U4Rc8pDA=Md1XIOPc&>2ykOY%SQ#K8~fSH*F3
z##Ltmoe9;ML}yZUrqG#EooRHYRc8jB8P%CZXI6E>=!8`#f=)zr=Fpi_ohUj{)tN_U
zUUe4GSx}uAIx*FWqZ3!1C49w}@D<}nR%uyQofXWisLm=ntE#hx&YJ41qqD9$8|ZAP
z&L%pWs<Va8mg;Pyv#mNi=<KM@E;_ralR_t@I(z8ssm?w+`>K;hCrzEfE0W`VOHXv-
zp7``lr*F=j6Z5lwvL3(xQ;V;leOKvS6`9{WCvJZAnt%GIC;uPXx0m1kXyz%kxt|36
zfA(5v<ZUOjKm5RlKX&v;?9uxl%{%&i?cDw6@{WGHc~%~HRz7z0c>iT7cv(s_M=ARB
z-@WbO$0wXH-E{R$r!sHe%$yYT+T({O13mf)furnLl5eTk?9zB0jXOBXcx{V2ILdgv
zi90yTcnyd<ILdfkhC4XQc&&sxILdhaf;%|Mcuj#jILdgQpF23pc<!A$ILdfFoI5zm
zc#fJoILdh5m^(PicrKSaIL3H>^{Dg0omYPHv)ki8##i&2w_kI8=Bt=F7<fYA-~_6f
zujc*_zn6%A_irjOeTVCg>daH6-+#Tc@N?J3zk~MOxp$XkepqLY7()JVn$=89=6eZO
z9GqM=)6Se%_248Q2WM5y^zoS9pYzPo_fD=qm#RE}9##4Ne5wlk1ymLK3#lsd7g1I0
zFQ%%*UqV%>zm%#ne;HLyzmuwRe>qhZ{tBvGeiv1h{z|H<{8i5!-5RR>)l}8^YpAOA
z*HTsIucNBoUr$wozk#Yoe<M|HzniKie-l+6zlSQX-%FM4x2bCOH&fN(Z=tHy-%3@x
z-QJq{mcm*7eC=nc^0z-jRZ!Y{=;wL+1!=!Xb%)(SNgcipO6u@;P*O)wIykB0f^=|F
zr`VlBS)JnR6iVw9f2UAhCmtho;wS#m-?bA@&UNC==trG^*a6NAh%dmI0r3YoGl0y1
z$_yYgpfYjJ51c~d$PpA@kTZkg4{~M@nL(8qL}pNBc8T4^nO)-R;><4bcX4JH_%4;%
zh0HFM*)4WAXLgIPn=`w`-_4ob$m~{`-N@`#nLP*RJE70LM|?e;*(3fQ&g?n-%=sAg
zAhSnh_KMxhnZ4rc<;-63_i|=0GJDnEvlp4YDzi`QKF;hDUms`oiNB9C`;ghEGW(F(
zr!xD+?&r*Y@%3|Nzxew(vmcrLDzhJ%{VH=n>;cXk5Z?f24v2q%GY614pfU%LIiNBJ
z#UA9$LGcZ8=AifoIdc%1gDP_nnS&~GNbDib91`CUXAX&fh%<+fIixa&kU69>hs7S|
z%wh2jbLO!4hdFZ?nZqh`7@5N=b42VB&Kwcn2xpFne}pqfkU63<N02$9GMB`^B+q1i
zcAm^$7XLEWE8<__dR6?ZT(60Ljq7#suXDX2{td1-#lOk*miV{0-WLBh*E{0h;d)p6
zyIiNlpW=E?{CiyQi+`W%wD{9Ub=DK^R{Y->@c&<UqT%c#2Or#WaPfyfH)lS`$~trO
m;jKp>zWeCI+y4DSBl~^FpWJ<R@Se_N@4oexo6p>q_5T6d)328R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_xdl_scale_instance.hpp.F388F6485863F7A4.idx b/.cache/clangd/index/device_grouped_conv_fwd_xdl_scale_instance.hpp.F388F6485863F7A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..215542b142ddfb6fb940d23d1b4b2a59759be18c
GIT binary patch
literal 20762
zcmY+M3v{bxneY32U)Y3wk-Zt*!Ui`L+<Ml+&7jnT-0$KZWpG4#n#ujXGY3w}&D}Jl
zX_K@~riu(qrZsJCnl{OorlR0wQfV0(&e0KDT4s%N&ddnn(&de%qcbYrm^n{-Ciefj
zmZAUu^F04wp}D-@_rC3iFTC%4SN%y}pCfu<<i6NLAyr>r-zon8gAYBB`~mg#ow&TO
zZ)9QXFMX-`Td$sa<L(Fk;_H8Q_G_o_9K0u%d_4Kjf8~1NC+ELvR^9xSFW&my$UXo0
z+v9)v?WYGXeB+j@q!*w6^~ld=Z2#wtXG%Xk|F#bwyM1r-^>4iVt=gY!qVKN!@P((n
z?|C@)jn}?);{9WvuepzI?SAW@Ub;FxIeq?9s$V%~TDQD<gU5XAr=@S6*}84*?;0;X
z_{5Lh?-)GYz2TPJ=dSwY3&pD{`WtS3{naa8`tMs$zB2y7|9IQSU+CL8^S~Fsc-_D!
z@3`qBQ~$nIdeck)mOZ(iA3pl-n?LeO(K6X}#eVq6?bbm5ow;{Cd;XIrKb+coT=)9T
zU;E%w7xs>R`u4lWZ~x&Z-V%BC7p?0*f7`Y1J^l0_ee4eDqmMs7qA{QQ-8apR`u}YF
zb<Jaczt8q}<*$DDnS~$fU;5Q=pSbyVkNvXv=$FpD{HtF))vxUS%S$7F_ygO2`_T(u
zsK0jmnU~^q!^*4w``6DuH}h|e7cG}v_S8qLw|=F3f!_J(72o^rn>?SJ{@8nV{^6wy
z(;F8)egC~b{`y}(mi*ArpFMe$=2CC`?3Sl~dE(5&Z+-qc&#S+gyft;ce*ce8J^YEy
z?s)0r|Jiqo<(}r?x4-|$oo{IV>K|W^zV^@2pFDo!qi^<F|L^m+-SbB4KfZ7L&|~=*
z=Wl-b1DD_T!vA>3%^NQcef0R72b9it|MgStw?F*xk$<cWf9D&&sJ!wHYp(CEUn~Fg
zVEVrB&-v$l`M<vN2d>+Su@ill|3P1W|LOj>A3XG3!Ow}SPxc-CP3@J!2i|-4G3f-C
zuQ_tfse^}uGIyMO<vl-|_=i7u@<gAS%h#R0?(GK;z3Ee*DW7@x$sav`OghTtlSfWo
zaq!TazV-PVzw)hj9l7#E-w`gKJ$?4ZgNNR<<+nfkvGTq2XO2m4;_}r;uD<->q3;Tr
zzwf@j3(wyD(DGM*_Qo;kG?&jGIj=c*=(|EC?U(=9Q-AnRb3czAldj?N{YUP9@4-W_
z9J%Z9+qN%DI!^Q*=kmGJ=iYwsaKL`^slJO}sa$tk?w;3<Nmp|D+9TH<J9y}Q=lA~n
z>4%>C$qVl~(RT%xZ#;eD+YTOj-+3W;=6#)a-}cdC(q&w}^2n7(4<35cAOG9m*4Dpm
z=$E9oUA6zM_Vx9jIC|o$Uiqib4S)Ti_bi_~CSA+ZPtcJkNb_I+=zaJ9e&x--sTsfY
zr^lqXaLZjs?t1URL$CbC$-$>zPW`X%9g~i6`P7k9R~$U_01w3C53YT%uTOp;W6(YS
zJ<G}C{Z+bDrLi9Qc>e-jT0p%>mljbk(WNESHM&$oy-b&uQP=5G9rX%bT0y-^msU|X
z=u!i9lP)z;x9Cy}wL+H^)NQ)dM!iOt)=+ooQs+?jUIcXX;K|?n`ShOr6YTFl^?UyD
z{(*TKXj1PP$XcO+&w<v=lLqS=g)~^#&XX4FT7|S&*Ugg->pF#WSl7>!9_xCA^jJ5{
zlL6}ng$!6X&XW=AMum)6H_ej?>n4RvSU1m;8S7?+%viU~lLhM*g)CUN&XX1ER)wrs
zx6P9c>o$dKShvrU9qV?5>{xfqlLPAxg&bIS&XW`CPKBIUcg>Rv>n?>{SRbs?V1s(E
z!sGpeEgF0p^iY+Cus+nHA*{Qr<i@(YMQ*Hns^r1Cr$rvDd#mKdy0=AMtoy3u!@93U
zKCJtz<j1<dMSiRYsuaL_phW?!2dfmsday-7tcR);!g{DhA*_e16vleEMPaN*suaO`
zq(u>|N2?UYdbCARtjDSp!+NYmF|5a{6vukJMRBYrs+7QbqD2X;k5p--LA{Td<NYHo
z8hIM@Xq85>KH8#DtY@l}!Fr}e8LVfkl*M|sMOmzm@tZZ)$N0?<>$xiBu%2sC4(sDo
zrE$WJacZ%HJrmqBA@ode&xFvE=bpUKljoki&{N=^g3wdoo`TR*<es9?Q{<ka&@;(B
zlS0oV_e=^sQ`|Ep^h|Njl+ZKHJ<~$ZH1|vkJtgib2|Xq5DG5Dg?kNjBW$q~pJylxd
zE`<(1SgW+g&KJO0;GPAcXMuYbgq}t2SrmE}xo1)6Sz^5;^el1DlF(D*o|@28<DQz(
zv&=oqLeDbyEDJq#?x_nsb?&JPJuBR^BJ`|q&x+8q$~~(>&now<3Ox<(X$U<H?r8`;
zP3~z5Jx%Uu3Oz0EX$d_o?r8}<3il{NkHS5Q(9`Cgw$Rh&p0?1l|FhYjJN&t?v0f8;
zI^5F{dOF<Efu7##nwKB@OY7e9qqnNI9X(i0_iX+YN)Fx&pGm`m4}vPt;1qm94OVDy
z4)joR2xf<dhhTQ7Ktnh?RG}f9btm00>mGK)th+#NoOM^ojkBJl2WCCP9+>qM$b++<
z3VCqWoAko0ci0QF-U4}X)>|Pj&iax*nDq_&VAfY4AI|zJ<ilBi(hsx#VL#0J3*^UH
ze}()w8%PFVHZUB3*+78;I2))?0B3{AAj}4bgD@K`P!MN>6$;{PC>es;&~ON5Lj?-q
zY^XvZoDC<#FdH5Y!)&-fVVn(DD2%g_WCUg-!x5N`6exnTkqSj{HkypWY;-sZv(W-Y
zaW-0^D9*-`F_?`F$6z*Apcu}^Dip)ncrp&N@!>ek#tRh3*?5KGIGac&U^X$FfZ0TW
z5;&WvP~vd*nrmObW9x;VpQ7Xa*E_HOK<~xhyGZ6Yf&DAw-m6`^Vphlh<kz3z?rU$q
z*3`RR?OiiJCm%hySnjp+CCa)q@M#GyS+=BsXHoA+123YUlLtO2Lw{QySVO%b4}1xA
zOEsXV4*NChk_LI2ElGns&5oo&o@P$gAWzelHOSL!$QtBnS}G0lwCj=<dD<;Wi#+X)
zq(z=~PSzq%+m^M+({9LG<Y`+fE%J2hk`8&gElGzw-HxO~o^DRoAy3zqb;#3g$U5Zd
zS}Gm#^y`uydHO9$k39X3q(`29PSzt&-<I{r({IRn<mp=~J@O3ek^y;!Ey;j9!;WM?
zo?%WlAkWa24ahTW$OhyYS}FtbjO&sSdB!cth&<ztWJI2EPBtRX*p`jRGj7O6<QZEk
zBl1k^k_maHEy;vD(~e|9o@q`tA<xv7O~^BC$R^~OS}GIr%<GaFdFCz2j6Cy>WJaEO
zPBtUY+?LJAGjGUd<e6J4Gx99!k_CB|Ey;pB%Z_A0o@Gw9AkWg4Ey%NM$QI;TS}F_j
ztm~2$dDbn-iahI%WJR8JPPQV?+Lo=zvu?;%<XKxPEAnjXk_~ybEy;#F+m2*Io^4LH
zA<x#9ZOF52$TsBJS}Ggz?CX*pdG;;Ijy(I0WJjKTPPQY@-j?mivv0_D<k?#)JMtXs
zk^^~;Ey;mA$ByJco?}jSAkWd39msQR$PVN=S}F(foa>SkdCo1#i9F|y<V2oxPIe;C
z*_NHib8g5^<T+a^C-PkDk_&mREy;yE*N)^uo@-8aA<xy8UC487$S&l$S}GUv20PN=
zx&#a2O=)lo^|my)gL+mToRgtnkq6tTyYk=$>ZWS2r8?{%>PSP#8`_kHkT<j~4Iyu6
zRvtp$kRlHuZ>TE|A#bRu8bY4CBe{|1-jv+Pb8ky-<hf^MH}c$y>_(ouE4z{BZmQhK
z^K>K+@;sZ82YH@t$%8!4tn5LaN0B|q^K@ko@;ptI2YKF(<VBu$Q}QCuyDfQ<=be?k
z$nz?)7kS>U>_wipsq!Mv*O7e4^KD8#<oUKGAM$*&vJZJaMfM@j*Oh(9^EFjI<oP?2
zA9?;w$&Wn$w&X{ie^&M*&#%aS<oUa@A9?<!%8$H2M+zV>uqg$Q7uc2p$P3KM0ptY~
zIe@%CR}LUA&{PGG7wkwu<OMgSAo7CSQV@B;SviQjpdts67wpPG<OQ3mAo4;TDTKVx
zrW8V6Xj=*)FElHMkQY+q5b{D@IfT4WQx!s9xFdy;7v7Y@$O~^vVdRBp<uLNXiX28>
zxGRT|7jCM;$cuEO2=XGEQUrOCZ7G7h$gCVeUPO^2$cuF42=XFLRRnp_jub^+bW@5V
zFS;#7kr$nnqsWUYauj*dt{g>Pw5f_BFV>M_$ct@CG33Rzr5N&JvvLf1F-49cFV>Y~
z$cr^qG33QNQXF~lO(~AN__h>BUVK)LBQLJVapc9javXW_rYerSL`O;>FR>{lkeAq&
z639!;$_eBp6gh#sL|0BAFVR#bkT=qiM%E>`WxXklY@yzkMs`ro$|G|!^egg68+BJ6
z*+AV;jWkt<{i7Xe6nUeY(kSvqx1~|!jn2xW$QxDUQRIzw<x%8~HdLd?%XFj+@-mxJ
z26>rnDTBPstein!Mv*hf%XH-o@-huo26@?zlto^4Q_3PQyDeprmz|Zf$jd5n7J1pO
zoJC%?p~@m}tRsyfZ){T<L*Cf7G={vfS$Pb3V~RY6ys@r4hP<(cY7BX~j+8@QZd1x3
zFSjk_ke8d4bI8jnat?X9uAD<&uA#~yZ=5>PI3aPIHrYYuIBm0o)Nz`X#|gRPq_BhJ
zaq6;zY~K1%jT6!*m_8xUCzw7V&?lHaA<!q7J|WO2m_8xUCzw7V(0NluH6hUROwSAS
zJk#?6J<s&KK+iKhFVOQ$&kOWC)AItIw`){+fnH#GL7*3yUJ&R7rWXWyf$0T-USN7b
zpcj~45a_%Sq$&vXBGZcky~y;UKrb@ADA0>cFADS`(~APV$n>H>=Pf2xQJ_yUeNv!L
zGJR5@PcnT{pieS=QlL*VeNv!LGJR5@^X8RmQlL*UeM+EDF?~v)PceN;pieP<N}x|M
zeM+EDF?~v)^R}33N}x|OeOjPTGksd1Pcwa5pieV>TA)ueeOjPTGksd1^M;#hTA-Jh
zUJ~dfrk4bIiRmSQUSfJlpqH3l66htSmjpU*<*7;nz0CBoKrb`BEYQnLFAMZC)5`+A
z%=EHAFEhO?(0P+kRTk(~dQz&=XQadJp(_0iJ6{%@U3T^arz%%zUdE{oJL`h;3_F{G
z(^m1Oqv~*Kf$0kZeSzr<0)2sY*+KdORpkYNzCayzkiJ0Au!HmkYO8oV6X=UfUlizz
zOkWh}i%ee>=)8?8FA8+tJYxsxi%ee>=)6&<;*B_<FEM>dpf53fNuVz=eMz7%F?~s(
zFEM>dpf53fNuVz=ows0tUSoPqpx2mQ6X-Rj*93Zv={12~V|q=X*O*=t=ryME<}T2e
znZ7L0mzlmS(3hFMEYO#kzAVs}nZ7L0mzlmS(3hFc+r~hzGrca*>rAf;^g7e)0=>@k
zx<Icpy)Mw}Os@;{I@5VW8t5xbUlHglOkWY`D@<Pz=qpTL5$G#SUlHglOkWY`D@^CD
zZlJF+eN~{ZGJRE`uQGjApszB0RiLjjeN~{ZGJRE`uQHuC(ShDzdPAT$nBEZR4W>5)
zdV}c=f!<(xL!dX9-Vo>wrt|hZ(3?ze3iKw^n*zPb^rk>>GQBC#n@n#C^d{4r0=>y}
z-uMT4i|H+a-eP)7ptqRb66h_aw*-2N=`DfYVtPxUx0udH3V^OKT@mOC(-ncPFkKPo
z3ey#Vt}tB@=nB&nfvzx}4=Vt@&GfcFZ!^6u(A!LJ3-mVA+XB7K^tM26GrcX)+f3(U
z5I|pJ`kFvrWBQsvUt{{3K<C3T@|r;BBN6Q2`}`Wy*91Brzfke<3ZQqG-Vx{>rgsE-
zhv^-G-eG!2pm&(w5$GMJcLaKe>3nqK_w?Q#DDT+rpGxj+!QQETJazi~#~$My+iPyW
z#?d=m+Ph-r?)^i`y>9<^zki|c_`}brcsKaQ>+QYW;N$&g&z*Jk_Joi3_rJ6Mj^63j
z<NY_?c$2+%G&wM`kP2>Xeur^ap1Sgy-jU?qxo18+vVZ8ghvWU>JsNlgkFf5M2DNsN
zw5WA^q(iOWBRy)v9vM&@_sEFav_~e?<~=f_w(OAwwRMlIsBL>>Lv7z9J8H)sIZ!+I
z$cfsuM=sQZyEOPR9*EhcA=K_&a-;U_k_WYSm%ONbyW~Ud-z7ilz%B()2X`rmI<!k6
z)ZtwUqmJxS1a)+mqNroL6hj@~r8w%uE+tTp?9#}~c!Y76Mp0*WDT6w@OIg%oyEKM6
zw@W$H<Fv~M4AGfjXF_oD?BoTfz)nGMitH2xXOf*s!I@%bN^qvxnHHQ9J0-y>vr`tF
zD*Zhl)_WC>aV@a3AUKQcEDFvNJ4=F7W2YuK%j_%*PMw{);H<E-A~>t;tO`zpord5v
z*=Y(+i=CF>DC{VL(`KhFIBV>z2~LNd4mhzpZCevhHN)2ZUmxHTv-cfdVpJ(rrH4`P
zpA;>@f5<9T*qH}sff5Utu|TN>%vj*3#)6Qsz!`i=0vU^xSj3D)N-bi>B4;cL8H=33
z7Z#AQM2RKLSfbPtW-M{Wl8~{)8GLyF88u4OFr!AP8fMfuqb6k3ID;4Ykg-gOWz1Nn
z)G}r)bH=ievCJ8~RELZ@CF+<_r&Jv?>YPy*GU}Yc3vS3*p~MPitWatNGgdfbMaWp;
z3|=-v#wsONF=LfdtC+FM8LL9ZDrfNG7cv_AN3!8L)}T}aGa8)H5HcE^!An@kXi}ny
z8BIzxF{8;DO(CPn8N3jM3_c3p!i*NBT9~na{(1kb@Zqz+#TmSu1V^ESf*A^>6wFXK
zLlH6*&frBMWV9*K#*8+l+L+PijJA-`<_uoiLB<*-)-Yp@Qfrv8#u;lu#u{hv0u3@c
zl;~hahf*EP=x|0y$mnneFSCB1(L21xYm)unw1PFs1`RxmYmyDppw@1X7PW4Jbg1<k
zq(^PoAOmXS1{qPCHpqn9yg_EvmJPC?wr-FWwQYlJsO=kMNA1`k2WsaAIZ?Yd$c1{a
zOM_duCh5`;YIm31s6Acsp!Rmji`v&EA8LP>{HOz63ZM>lDTq4Mr4Z_Hm%^wcU5cQN
zb}5QF)}<Kgc$ea+6J1K69_iA^7OqLUG>SUYr3~t9m$InGx-^D5*QFe4-WlLE2|9eY
zp4TMk@Evr10;9uswRuf~4&Q0!H3>R=_m|fs=<pp`UX!51cS(6of)3xg<QED$d^eHT
zB<NJ>8UEn>^6y_f`@gIA41Wxw!*{~?Lk%6iJH;PS=<ppQ{=h+p@8a;s2ReLbg+C0?
z;kzCD){YL}f#5f4boj0SzwM&KSL*yGiVj~r^IIZ1d_~J|aOm(=Ccj;w!&iR%=7SDj
zjqzK_q0_ti;j5|r-^%R0l|1wt>*rT*{LB$vstl+H26|U$$NM#E4JfTz3reTffzqq>
zpbTmQD5Kg4%A_`dGONv?ENTlVtJ(_6rnZ5ytL>m1Y6mE%+6l^~c7YmH4-UX7B=ry|
zx7rQLqxOLEs=c6mY9A=S+7BwA4uA@(gP=m{5U8*^3@W0IfQqW4pknG6sJJ=~Dxpq*
z8c~l7^sbqY_m8SaL1ok#P+4^r)R=k<R8E}(HBRbr0zE<M3DkK~=TR3(T|iwVbrJO>
zsV7lSk$MXCG^wXimq=YgT_$xIb(Pdr8aP}~ERcEu^&+VkQ7@5t33ZLsHPp+bUPfIf
zbshBzsaH_1l6n<&gVYVwO;R^ew@BSWt&p03LH6+bZ<D%>dX3a;s5_+Y9O{pLz`1?%
zkmf^tC*#~Z&*^%1FnV`1c-64~8`iy-I={j%4%~bA3Li)fJP3+^j#IFr7^u*|CqZj2
zYG78A(!i`HPa2%nR7iuf+KXD4)uyyCtId-ZXSEg5;;inX4rX;J9n9+Tq{CTVg>*Qp
zzo>^<eM%3r`aJ1zR$n1K&KfQnVAhZ_z^oxp2Anli$bhrPi$<6=ri?Ia%##skjTJKD
ztm&c&W=$y*%$o9K!dX*=OgL-4XogvH$_%sSJehIUTp=^gS}s~(){?TotR+ttoV8TQ
zg0t3(R+zP>tT1cMlND#J6|&;2?V=54Z7Cbf+VW(>SzCo{IBUOXhgo~d4zu<=*>Toh
zAv?}GE;?Y=k#fMSBTo*TbyUcKv(Af7n02O{Fzd{d6K9<ja^kG(q6=nSDHqJT^5nu<
zSA|@Mv(rCyU9a2y*cpDcpFDN)ExliMo4NeU%cZ}4V0`}*e1--d!`JK#X;5otNQ+uG
zLps#@8PcOR%#Z=KafXbjO*3RdZJr@BYRe2+P+MonirO|qHq`bRvZHp)kOQ@IhMcHf
zGvq=&Sf;^8@v5LqL#W+la-;T?$%EQkCNFAVnS7}IW%8pAlqrBZSf(KAP?<ug!(|Gi
zj+7~aI$EYE>R6d#sN-ddqfV45fqJA&Bah-$L77HTXUdd8oh?%q^;nt4Q0K~&Lp{zb
z?r}nAf}IJ$$+MFeoB}%q!6~v+6r4$RCIx4TohiYYW@lP(O6-&br_4@SaH=%JR|Ss=
zR|PYCRe;VS_bdv|5<5$RQ)8zlILquT3r?M#y5Ov^vm!XF?5qk-gPn%pG}&niPK%wE
z;3(`Ug41TFEjVlJtO-tsoensU>A&__?cf);_&wz&=S}bLT_n`s`rOHHe)FN<v3>TA
zvmfd0s&rl}q}|_Mc$BwZ&YRCW4i^u1neMV5F0Jo3-ETi!h(1B)C&<xT;;fH6bJ>UP
ztG|zvuYS|jr+dGZ_xIm_W%IKyf8meV?!VCQ>uub;`kis>U!3~vzp#DIaL(0R;Pk@y
zTIT=%5t!Z*r}sT@VgFCR^z7bWeYyYW!C!mn+5Co?9(d&N-$IktfYPS5pmb>+D1BNF
z%8)jIGNz57OlcD+bJ`5blD2@drmdiCX&Wee+78N*c7SrGouFK47pTGX;3K{NCdd1S
z(nFx!X*VcO+5^g)_JZ=IeW3hlKd3-D04kUcf(oTWpu*`es7N{jDw>Xhilt+q;^{c3
zL^=U#Bsp^L;WKSyc;rD)>5)f36=-A%-Wo?LG%^QzG&u@Mqr;<+G@2fTq|pM6V$x`Z
zMlmUq%s^6RI0H$UbOw?#1<GJjra~D^$|kdrlpW4OQZ}81q-=q*n3Szh7L&%3V~{j9
zJO)W)=`lzeE6^Awja6t2lXA%%B;|&4kd#a3ASqX%946%|l*6QPN{$m`jnnWrLE1Q_
z#|iSr`9ox!KN1dqrsI6Kl#jTAGeOA-%$%U%3Cx_J^aN&3aOQ-NIl-BHa1}E1l+0sh
zo`&<7nWuCfGxMC87c%pl$;VM4vp~rLW)^6;fSCnK7cjHHnFS%Uz?po=6f%pHEMjJn
zhKrb4q;wH8i#%TxGK-waM@J!Zl9H2{IZ4Blm^n%5Nz9z&%t;}0k~8^0C}d7iatbr2
zXm|=UrzkyznNtV<=J<1;;!Hl~37ONBoW{&)8lJ|?X-ZFH<}_zc3*YB7XY%1q$ShH^
zgqbB8E@5Vg(k09+ab`)#EO90u$%M=@CCivurr|PXmMLAv%>9Ma{xazBIbY^XK4|%S
z=YiP0q5Z#;_TU4N2k+kh;;s7+2mh1$wfFu``-|5;b@IrWvv0lb`WyPL@0I%ct~>nl
P?BSPZ4!^wN@XP-PmhOpg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp.043570DA9A94B804.idx b/.cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp.043570DA9A94B804.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a36df96b394124952083077b0948b10d52cb243a
GIT binary patch
literal 7636
zcmYkB4RjRcy~g+d&blo5fW%@<APGyPMe3nlXPEkdhwovA5rvYv)?ucDAT3-7!%W7M
zp!L!Q3ISpSuh`s1AX)_hHEI$t5I-QB7|JD3AgPJia!P>)Lutj^izwd7o{sZy4sYg}
z=iTR*eZOZicjk;4M+-usX^+iZS+dA|G#m<rvc3PMuCwGkg+kn*P-tas%a07}xA9@0
z&ptb<`t*HUUi<I8yVaB5IoU97W9uj57MkyGcynCFwI84UdLp-|uHm5<mv3CN`iUny
zTh=|X?D*-66GuP!!1`0HtRdsw{p*&V+EMV<!Dq`Kdt_|Qs{PR!|5`Zr^0>uYr>*#`
zs{et9esJYnsB_hob=-Y_|6pXtYbQp$w0rLGmUwdOEk$p=|BG*3+jx5ZoH3{7ymsN$
zsY4IOE*ItY^^JP_XCEG%_F#@v_0x@4Za#f>!5xQ;{j2_z{^Czo4jR1YM<1=7v}V5~
zK7ME3OP$YL_|x;|!?U8V7B?(7_SMm};+OL_%xSIp)re79V_y23bENx5;#Xy_>}Z>M
z;^IG^SzUE5{=)3)C6gK-9QPj|UcGJP+~|_LHTjQh?Rjq2zRI@uMi%^YVNd$KJzK_q
ze%o8IU7fwhAG_n?gSYqf+1IAb%{h7Z;O{T#?tWp`6SpnppW^QC`)6_WR^_Jf!rGer
z73Y65|H0+;;=J2$KRK)a1@||<{djo$f}7mTwUM<ipWl7+>T5mb-}$9qPpPYW_U^`8
zmQPCmtm5(UOIlvKTq*x7Yx5UX_Kny1eQEEVUw(P^{)TtIvvJ;@cYeQn^bc1=9<5%w
z``X&CI-lAR8)THtZhmgakg@-H@{x}H^ONtsv7_SG>W$i(8Gksw{^873Pkq_*-KxIt
zO@5D_e&^+6X?e+lD_(E2lFwSF-1_{5%+R23g<^#{g`-nNXs|aThGm7)zS!|e#r<M&
zI?MFz+(>SAsz}*hjf};cB5%$fF*=>)dG_MS;`>rXXmFT$ZYaBKMeES@^=f7);@Jf`
z1x2aC56^z1a`@w~_Z?1WX`Y=G$r_w0{P04p^~SN4D`uy&9M2vb8GB!<@WbzY^|2|x
zet7Nr%+O%Z9+@+8RI2d9Ctd$X|E}}pZ=|y!o;@rwY*4BQ4d%RW9SY4FYqJ0Pzx`)3
zLs8Ejkuzd+s_;KuF@N}sJ$DLa>Fg%Y9vT^%mMZ*Tt?JPA9edACh%q+m=7F?OD3+O)
zd9!aH9i80uQSqHq(^;<f>EcN7gjC`G<~@xcJaqBP+~w(PsAuOz@`j`e|I=qP|LmN-
zV_$1#D8sXha*9T$ij;jzcIea3<~2^9{@uIM*{z;EF*0#Ns_?_V%1XR-#%evE&eA<Q
zJCZ#(Rruk~qMw#7$qV|O(o+5He<99!V!}O#&rb?x#H{UPy+VEuXT<F7WTSSrlY`pb
zPHv$6fs>t<8Zdr-Ml9Y*@gDN~HY3J&5|3KwBmuS9Ng`_L7)c+I-v}8o`54Klm1Cr!
zR*#X2T02f!FZn%`5z~*8j@md*25R#-nW(KZWL?B&IYTyT=L|Wh-81BZ{`Z9Ontfol
z_|4)^6z}%HEcxd4zb8muT54kWw=-f=sZ<s;fmAL%396E$P0&D6HA!1gZzHJ=bv;RM
zgO*EW$d=1x$d)Tf#%#HoWXzVgk&M}LJ;|7@lq!&|lq-;}RFZ<(N;N5%t!yI&vz2;M
zFk3BEAzLk1AzQ5^6|>c9QZZZIMk;2j^`r*bxq1JRTaH}Kri@tr^!)q%rNbY?UZ)K#
zPJZwB>ymeSU{UgK`|FXNmRg7W+unpY!s5SYewe3wS^Q(vmstE$)J<G`9|!f@x%h{u
zySexuQ76Okws26NKf-wA@x6>k9)F4P$m5$h9(jB_$0LvL=6K}s$uN&R;Rq9uC-gD_
zdBP<oAWvxG1mp?roPa!`n-h>HB*Oyo#3M{Zp4iJo<cXJ<h&-{06Okvjb0YG@Zcap=
zm<)@^lRB7mgu(plVNx&ZJ|<m4-N;Ez9Mo^)q;}L@oYalFH7q5=L4CP{$;gv?n2bER
zkIBfB8#x(yavLWjPwwJm<jJjJ8F@+vQ;?_hFa>!^A5)O0G;#{^lr~O5p3=oB$WvOw
z3i8wrrXo-6VJh;}KBgj1ZRAwsscoE!Jhh8ck*BtXRpe<MOgqAY4MFQ+S}*E8rd>ka
z$Z1U+)NkXocGO**){XjbSZfUj_4N*>BTw&PI`Z^BrXx>p<aFfeZJdrgy^GV4rymaM
z$TK>afjpy!8OSsGn1MW_ku#8Iv~dRVj4sYVo^d#AAkXYzCi2W4W+Kn*V<z&<M$SZ@
z*~Xd3GrKqwdFJ7;i9G8hvre(#oMN40)?ZOyW!C4Yn>cG92ld-I>qFGtob^Z4$*|QH
z4(i({nT<UA9J7&UUu8D(>?Y1ep54yb$g{gS8+mpzY$MM($sFW4=a_>$=PGlM=QME+
z@|<?gL7vmiImmO8VF!8cDdwJr_HzHq-1DeEXYT)^ZsJ_zx$T^bJhz*3k>@7EZopG(
z-m5-V@o8qv+xV8n?O>aVuc3G)o)eex;PPvTM^{)TfGeya0bOyK2(Gw>M0BMR$qANM
zse+{ScpNH`!IdjWMpr3Oz*Q<pL02tN!Bs0rMOQ1)oM2hdDo9(8=h6}#T)l#Hbd3@N
zT%&>vbj=bIT(g2qbgiY<ieNZf>&V)K=h>w;xb`}-(RG$O;5zHbLDyaCg6pm$H*oz^
zlXol{c-r#MKM&ls#mX7C=^5{Am3w#YH2<;TC+9BA3}rliz<UnKFUp_le>fvnFs4BE
zPs<sx!ZC%i{{)*6ixtJD`ll~*Wvw;8<=|hu+@aY+bN%PJMGs7A`Q@9hEXehqZ*1%5
z&<v}r)zAu76<N=s-bz+2>Rn{Lj=F`c!=P>3hH!h8jp23`*%)qbB^$%-U1Vdp-9k2o
zJGKMi&MF7PohouL+}TPFhC92+!EmRA91M4D7sB0DE{3~R<YKtHm0S#Wcae+XZVS0V
zxckNAhTj#v5y|rAdNe*hnCpC$2PH%WP-0XBB}Jw2!CaN2GAJdgfKsC>C@rdu59XR4
z)j=6i1C$vxL0M62d@xAts13@AI-uOB`;98QfBMloe*A+u-ry@5Q*@92K<^Ji?{RM6
zje<V~yw-5yWkGAi%j3TQwSnSK!#Wt>O!4PX*HQds)D0Bh51My)$mYv=$mTZ?kJ<cY
z;xU`ABObH)2I4VWa0JK}$_2<4HjseX!e$aMTc{%evxNo{Fk5s)$QH{*$QCz{h}q(1
z5;0q>BN4O31`>npbyuYPdpb7__j)HQJL^{eE!4(A@7JwA<g6api4yU7&;xrYUXLxB
z;KAkh5Ra~q5Wp4okbth35WyAqkch67kmd!WTiQv|UfhWiGPv?ilF?NX3b@KnQqWZs
zD!A%SQqk2C+Pq+PXgf*Ui#t(52Up)oI=V)}0N2<_2D)a#1lQb2Cc0L_nitG2tAVU{
za1Bk^;Mxsjqw6FbaGeHn&~+0oxNZZvf&0w89rbr5np(X6k4=v${*Lit!3S9%f9!tg
z-7c6}knlH|J(nuVA30n5w6_zDoj!JEFf1lco2UexRXnX&37T;6^u;s%FSd8(ft%)(
zyg0+l9X4cGj{hFOuk7uI11|{u#o!MYJ~>!_rcw237tSm+{LQVaKmPpnin^qCJMZqi
z$^IKe|L3T*)Eh)U#2di}D1I2mY&=Qv{{hV(ARc`_Nj&<(0TR#`k|dxn9v~5YF-ao&
zQZq@1VC+h*Bpn4UH<OIM+)6U~N;4_wE3KrUuQrp4zS>GE`dTw-hhPkAZKNFotv8d7
zzTQSU`bIMu=o@Whpl>#liN4uJCi+%8S)DMJtrKK@1ln#V8-4o(+2}j%<e=}IAP0T7
zom}+Y6XXW|pZ~IY^5*RNUT?G%O(>H5{m1|5y^|eys|=&%GR42d(Q%o0)WT&FP>YvI
zL@o7`bRB!QpJdcZKPjlyeo|3u{iI#T-svYDwb4%oYO|kA)Yc8}3=5AW_6_gU3fj5h
zol`-(Uz1z)jrJ$8Hxmcmru)s}zexXfTIvP5f7_c|wG`iutyN1rYN3_{)M71(sHGZ`
zYO%FyNJg#Hkb+vRAr-Y&s+IYxM@CF5*PaAbN!li8Nv)c+EvRcqt3|zyv^v!Fq`eJV
zFV%sim+QdND@jL|UQIf(^cvEUrEeo0S$aL`$TCU|U>W5Gu#8GFkY!Ypfh?nj3}hMG
z$Uv4+PX@BgQWIEaxd|+@l1yZo)np>etRWLw<~A~sW!94!utLtFWseU0g{ah7SXw-A
t;r;E^|2M+F2!-;4%fjGtcyKxLn@bk^_NS2vB}U2A$ne5~5&0uS{|}T?&UF9)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp.8003C44E683CA082.idx b/.cache/clangd/index/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp.8003C44E683CA082.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bfb90c0a149d38c30e3f7d390ff06ca6ff6cffeb
GIT binary patch
literal 6892
zcmYkB4S3bm8OQH?aw)f8&}uC$lydo~WokD!Lt<nK3@8#4GeuE0(U>gKZa^!AkSwmC
zI15{gtbkNXrB&EUaV@Q_Z3l%FtXpyTg~j41;so@;&EcFMli`-$=ZW`!pNI3@U;gL3
zze~=^XYZU}Sa^19C^Y*g^POeZ$_K-tP$=F1SGICR@s|_|Jv=fLa(47=d1LwU>ocCX
zJhl77bB`A%U)A>dTdS6(GkI=T=Y5xUwO$xpxPvb^etLZVqFXbCo9<}L$dh(Hx#oYj
z<bU_)GrGoIad7gSpVG+fBPI-8p4L<}dtuqrk3VHq@42*BFBK=9eB-f)62s!olW&||
z^vgdLJhSRzZRTr>D-JFD&$H|D&z|_O=MMMks{6OM>w6*vrH8uy+bd^nNzA+E7q{3S
z_g?wimd6ghJL>7ng=4pkZ!h}uuX`3hcXGvzv#hS`Ub=YweTO~{KbXH`_Dd_aPJ8d3
ziOcVhR!O-Hqc=@?VByXaN@P!G(Xkzua({RAyDv|jGjU1M^2dvwd3WZy^_zQFPCGY!
z%_}pfZ@$x<`1=Fy6#YArGUoV#OU|C#i}%eqzIxY;UvBz&!o>V1cH7y<S3Ud5Kl78n
z-t_t7PyVQS`7gS!i(L5hbN-s}&5^l_7k%%6w?5DNtnaznq8l5kMtmVAeUMx^?}Nl!
z>v}$#I_FO6&HB^Tlr#P6n$6x)mA++d-}Bt&QDvL&eSff!hGsR*dT>wo(GmK=Pv71<
z_|BT>TmQW2^gB6kPk-savKy){+_K<{#+_rXZ~bh-8|UuXy5a6Gy9@Kb{>i2jKWkHm
zR(<@p@Bb>}zCR8gscG7H@PFO8Q*WKQ<k;rSH)qwZc=f2f{LZIx_m3L;*6M$Kv~$mO
zpN;!u>AJ7K%s=tiK*ILW<?w{>zwVW}%Z!Wu{Ij0hZ_M01vOJlMzA_ZePRJe`YeI?s
z8u33#U)7&pciXfD$t=~6vl6n>V@(jJKb7&w;j}e*$!v@t&rX<K7;8d_VKzKEy=BAL
zRm~TUB(o$xPD@Bjj5WdNE4M{&ZkRT*C7EUU@uGx9g|Q|W{p}sw+HRgP?{I1;#gFqc
z^QOd_*yyX$Lo+UY@{h*zu0MK|A7><FjEps*M9%+jg+jMo-*{$p@2`%fh7$dFQs$(I
zu_iV;BQ11Kk$%&)+3$Czh9Z7EE_2+JSQErX(fpEcFWK;9YG|Y%M>C^SVoeaQ)5jL>
z{q7BRG8^H?qZ39a#hO^`yQ>ags(=2y=~2ceUNQVzLZN7CQtA~!{LY&6hc+ZF{xmg|
z;E&GD%$*c#g5qB(P5hW=O7BZ%nSNZ5P%tmn1iR%@`?(rvb#h-a%l6}vgpzr&CWtSm
zEo>VqdFgyIOZMaRg!IH%6S!Ac50#Z>#^*gfDK_81FHVW_`-y*zg5Qu5750;WEbb=}
zS=vt$vfN4XD-<k@l&I243bNWsDzer|8nWI=`YRMHjg+X-Nd~gnNhY$@Nfxqwi0tDO
z?4*>abBG*d_Yk?r-XZeh%&qXIOGnPn2&Y8t)pj|QS?vJ1t6d;(wf7AL2h7Bz*ntB3
zwSoBOVZZXt#CHP=4J2SLG?Re2*gzuYVl#=DOARDpE;W;cxm-tbBg~83L~;kPQb!8r
zN)suVt97Jet~Qa1xmHIS=2{bJnCo?<H^RK?O{8}K8+Bx0ZZwgBxmiah=4KO_m|JyZ
zVQw{%g}J?(?0v8<>^8D{fSujsVD7Y$gSoq#T+H1zaxwRIlZUz2MqWG*jt5GLogjl{
z5u6~ilVT^vU~@`TE|!-DV|*%=%Ye$|M`7aRDv~!LSChOIxrXFgV5L}r(n_fUrIm68
zODk2RU}>eA6fCXOkb<SvViiiOr7DzG%T+9`R*{OO)oN0)v|2+dmez_jD6N%hP+BY3
zu(Vc18kW|oNyE}w4QcVx!I{RNnc*u?Fa=rJU(D+{`tNi~iRR47xh=R{1lJ`0v^ji1
z3VQpe%~Ql*z{B|z3CQ9p5|O1-Bq7U#B%jAw8YBf-9V8W58zc=`A0+)e9!`T~Ae)0^
zB3pxGA=_uj{umF3Gvpw<XUIkN&X5=9;9C4}QtTQGyQPcq-7J1J;rkhX6!|3M-$CBR
z@w++b-@)-OBKL9p%g8T;`Sx(UztF`5^a=e;K%a1u3Fs4caRT~;4o*Oy(8mer6J7`l
z=o7n`h(58OiRcqgG7){^E>1+B*ujbD6Z<$3ec}sY5q(k@lh7yiGYNgtNhYCB+Qmue
zlR7vFeNrDMp-*}tETK<sXL1LFb59?WUq&8a@(JX6PHy0!e=8@qA@_3fVdR#uye}N@
zue37-eM%ow(5DPA1$|0Ar=U-1<rMTOy_|wRr6sJOPi<!^`qVzAqE8)QD*DuVPDP*E
z%Bkp6dpQ+-YD-u}pVrPa^l5!eL!UOlH1uipoQ6KFmDA9t^>P~ew3e`jKE0jk9V|G<
zr9|~UroW6l!1NQy^_<?oLH|}xZ$s|o^ux%_VZ9|B?{Bm-1ARsxGtg%YFav!?J!hcL
zXypv_8NHl=KBGBopwDb)Ci=`iW}?p=U?%#^dd@_j*~*#dGkZA`eP(mmM4#2pEc97@
z%tD_vz%2Ax^_+!1tCh3RXZ3Ox`mE-#g+BWLvwK+lHq<`G>{pS8n0*?#k+b)3(7&6r
z|A5@j*+-E(!**9V-rqUE9P~NIn1eoNh&kwU8aW4jPB-VE&*|qJ^f{ej2Yv1V=AzF%
z#$5EdL(D~=+sL`-bGtbgeQrPJqR;INyXf-{Fb{p+G3KGq8)6>%yhhGLpV!TK==1tH
z4}D%|*o*t*rw={(>bk$Drup{=>xr+z`-Al)AdBlsM3&Z*ge+H(T#5Gw6{H}m6{I3-
z6{I2S6{J_<{Xqp8$YuqZ$W{ee$aW>!zrp*1N^+3hN^+6CO7h|yJdFD{S;J4_!To_M
zR+nTJRBrNbva)7nDZ#T-aI@w29)2(ndf$KcgZZJ9Rr~#Cvz$peKL~oIM04|U=LXNz
zDbZ+tbYAfIo)VorY4Q&O`vcC7620fZ`@U{;`sl3SAwKBr>xLin1D&PL>Up&Fw}wA?
z_7Zzp{1~xI>{UQ($$kQMj9o|ebHL6L2gW)j4vclyl7nNNI&yHVyTpaDZix$H-L>T6
zShtQ`9P2IdV60c-!B}rCc{tXqBQHL7<ySM?cTH+Z7+$pzo`;*OhyWx;L?9_50m%`W
zhl_ed0a7C>kQUK^^oY*G#Ux??nGqAnidaB)#OC8u;zS%EH{t?$5$_waZ=3U%Yws0)
z=&y&o{5&bRw-45Z|4=skW+zw={@y6!m&W%7U&5CGmGh6m;mvO(z8ZN8@ioY`#P0+a
ziUcSvln78-C>OA_u#p5TEo>nHOAECmU}>>PgwkS(2&Khx5lf33NyO6P780?vSW6<7
zmWm`OEtN=6S}K>Yw6u{VEG=yz2}?`0B*jZtT%4ODy;eEapP97uv@yYB@}`^m_lXN%
z?D02Go-_H@;JAKvd~e#RQ!D@LZ_b^UdslEO*!$5schP%0Hu~p_0<GY#_*~7tefEO*
z>RfdDq6P6aSE7~N6|AJ^o&6(zxa`?Nzi!5;jLhI6=SAnMUHsZo&EK3oEBo%?4eFa8
zoIbE`s6+NQN9RWu2WOSw*Z60b;WvH3YC7}&+L^ye-}hI)ci!~8xxv~Cex%=f_{Cq)
zJNQNZ=HVB9!RFvyUqMpr9bd57U#TwtNKnduh35)DqALPPt^_2z@+0v*qqqu?>Z(AR
zs{!f7`qKD5)Jyd;AXk3`s9b*(E<<`1>6?(NN#BZGLwYT+QEY(7C^bN2xCV%fasx$1
z6&WZps>wi+Q9}lb%wiKnW~m7x(=|b4mYXOttH?x=SxqL2%o;LLWEEQ=vPvxwS*`^l
ztK347RYewxtZK4QWYv&`BD<RGt#Dzow~@Ua*r_H5b7vbln7h^FV(xAu7jv(gJj}gq
z<i+z9*3t)uUxb&fFv}JUe;GgA#6B!~<j32W51uVunlLtZ{J5M6p`0KJg~r9d=ElGJ
Of6xj(6M`?H(EkB9fn`Gg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_conv_utils.hpp.82861A708A167591.idx b/.cache/clangd/index/device_grouped_conv_utils.hpp.82861A708A167591.idx
new file mode 100755
index 0000000000000000000000000000000000000000..10ea0f430d31189c53c1c26876e954de1e63161f
GIT binary patch
literal 4012
zcmY*b2~<;88h$T?l00q}Ed&T8=zVM_EFlvTv@uAEEP(+AY*fU-xI`2IAwZ2NTeU8A
z8`L_~QCjOhs4X5_SH_OonX&CrM>{TcN1awrtE1J99ql;xz3}ti<edEP{O`N>zW@IJ
z_kZ`6I&*TiswgUJy3<$VX`CXZC~64#H#B-@ouVn~Nf1T(HvZ#{!s6Sh>Ziqz=;6OJ
zB)vIj+_JCt@9f_<tY&0+T(`PsRn6t1%gnQfF%N=@Tr*X9A0L=!p3^;}JKk|(?Y9O0
zaj!Uj?u@qQW>b4{ciZ1T-v7Zoh421W-x~e)-(2kd>3N^^R>zr%|Jtx=)0Tz#=~EB9
ze0Duz=A4a_JI!e?&VRkdRoi}LL~;MAxRigc`}CTM=}IaK@9s;#)swhjS#j~&>s_y%
z`BSEC_O{6trT41$Y^)ve+5F_9bCGjzpN`&teW^4Ge^j_4WLMyeY4naS;|?t-u08T1
zCpBxND*I{r{vS^)-InlhbG@o?fvUosX8oi_ay4*n`Fj2FnnQ+Vd0As+TTXAWu}xIi
zqxF%U5C6Pi$;N;8v^+kzB<`2~CRJ0#o|TWjt0=GdDt*o(->c~UUVQR3vnEP)Zf1U9
z(sX*AKB2Alc**!fE%$zyrEhu?a5S}adRlAk<+il%R9$zj^nLU4z|Zx?%~!i_WCR(u
zN%Cd8ey!O0cv77)?Bs^{x)rncuCHh=-9JqmefQg&X$iP#vzFf4wf48aKeeUiZ(q#B
zvwJR><~MHd9X4Uglea7mLt9DN^T2?{iN8v+mDZ&7v<KIf*bc9_PN{dFji4Il)l9t|
zKv5d=iAQJEce|95U_FIf)Ga#RP%=_=@{7vKbXt9p*-D9u!0j|@ry~X)FhrMWed;O*
zzoe8XhEX_Qk+0<qz;X9~^ff;yJ*Jj~krjNIzCzvrY;e^rzH%e#kFQA7x>xX!pdqom
z0oZ7^m|99N-Q?gvTohCk%Nu~xPu{+sR6KkuqO>{+zfEi2reg;llw2M_MNu@SM_#bh
zD`H;1%B?|R6og`VL&+JA-ef?=1Jyfc_FC4cB|}MwZlik?Z{Q<ivQ*Yfo8CL3k;DwA
zutuuEya70;8ZW=7>is||QIgqhY@3NU08f@h?0Y`HGaFGM<m>O#n)gYF{s$!wqWvMd
zTQ48{^kV%=wIqV9;m!0G@&?wZoK}AA#Oa08{1Jpp!!hOzz}0<=R{i(DlaYvuB%jz#
zYj)Ey0}o0rllVVT*Bki6p0N0?S`tpmWR5pmcmr$9jo#ZCGSL)63Qb@qaz^lG5T38z
z`<KOQ&nqQ}1lX!>)$s;k-^r4nE9vZgYKex-PK-~q@CM+P8=v(bJ};A#6cIR&%`@=^
z;CGUqD=yBT(n2^$;2bt5l{Wyd(yl*yD%hbR3?i^2(^1G9fLAYZc8~O2JB6qrWcO=m
z%^H$m|3gS{XGhuTyM2G1NfJ!fuw~i`c>`;_yV+=wKl{)`5*$rom(+!M18}$M;I;lK
zyU(d5p=9?Sqh}Ov0RHmvZQ~zK+#BPcJytpvW6l74BV|$c(O<GnB=cl;ReY6&H-qp;
z(<AOr+Nmb}L*Nct(Lrkm9x$Yz`8@Z*uGgQE?33to+1ym#0Q`N^xR(2OyHnJXR|(9<
zvliX}OnE&s8n($PN|CyBL)E*cD-wj4(W+&1=wSQcm9%Q5;441ac`PBYu8%>Om9mIe
z0%8zOlqQNiv(zl|Y@s%tFwzm~5P5E?TQt%m^@zMGX_d(HhI&Qwe4)NU50okZ;WDNy
z3`&OZRMXT9Xkms|FHjBbq&+*=8ScD=5DsU<6QGR{j$k81o|e^$yjWSR0y2tlvMgBv
z>_*rkvnZgE5YCb2$bq{E=Sp+sP$-1+q<M0nJ;LMJ@d-k})v{^@kOJX0X`3801mrdl
z;YrLS(H<L3jTs;k2sfLWGeB3F0Xc>}Bf=eB2W_}9Kr4`En&D*`^cJR**PJH9Y=8q8
zUWT19Z!YgVoYjhW#RHIJlc*1PFqKT<^jVm~^#Fi~?<eqpcjV)K&4x)J)nqMBr{Oj7
zp9ur(LjEIQz=O(@wK(O5Pk?CClK%@G5=}@{&Xk7>HK{GAErXgcynw?>`R^TDeR|>d
z3eq(jXw?QfWKeMM4qCN?4grOSCyB(b?+`u|jsTpr!@@vE5~4Vtv)JPy!xYLLAqdAR
zW3>Vm<L&WI!An*qYXxE&?MA2IS(Fy7K+|-4x>N9MX4_~`Lxk<D-6(htvtu-<Ey7uB
zmQnEB*qs8oCOyYaH42^wds09n5w2t_je=LLtkyze5w5k@It9;b_Kt?+Al%3{8U@d1
z_KgOT62`O5M)2UcBkA;?b&%(}5fAIfo^K4J5$>f`y|f-i!tTio?%)IF!kI_9Et-iA
zgL;sxnCuxsR<cZ48IT8tKNVpCg!6Uz2I#Q417{@R)<kBasCg?*l^Ku_hCerf0#Gs%
z7<XvGx*!~6B|W1L7XnK!NH2rUF#H(`Q$UBvlqe=jw3qmT_zA!ohCf$f%JTexML(>5
zIU<Vmb8>i^@Rf+<2$82v){4AHHqs!BWUZ_n7AL2JHPH?`LfC9I+aXqjZACVhkWxpH
z!v!fNI;zYGh4_h%Dsw|cI;zYIg>)e7lsbc<Y9#s6{9s5v5g68G5c+u>J5J<{PadBp
zU=!F0BFt@fj}cMX?imBMLb%FWWf#%4$m<djw8-ZIN+R4MYf-@og>b8^RRwe)3a@Mn
z1>r$NuBGjCFep7I*U}C;7~~qVLKwl_bf8wCIpnqfY6F~yoDzV$)xdbzEYTM3HUoTv
zosrSpEe2k1Z@94Y^UX(Z5s3>igjj@Ayu?s4=pCGK^Q$q;7q_@efr1Sa1|=#IMV?t<
z7I`+kEkPLR&^tt)Tj3Us^e8+cuS!uR^1OPlXr52+8}yEh8b0LC(w|mJ5ss6^4HdYU
zB1;)6aM3EW4i&gqBdZw-RfW47NtF1!0+=L<yRbo`pi1Pm|Dpx}5SFx`j?szQ&sbob
x0C6z<tqn>9v68j8D;vBfx3%giT*?}}(;B!ONU(SK8Jn*Dn_QZuSV$&C{STQv?d<>n

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm.hpp.96241388E9D6D4CD.idx b/.cache/clangd/index/device_grouped_gemm.hpp.96241388E9D6D4CD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ab02bf01c84537690f8604e15065b0c65007c566
GIT binary patch
literal 1524
zcmYjR3rrhz9KIfX(0bhej`D^Mp+HONfNpW3g&{+`F$Y6dx9|!S0bv443$zY-MM}m{
zv`(Wj=mHxCTqp(~pgcB4=1`o&8mCJLI0CpuHr#MiH#Yp==})<&U;FRxyYKh<?lMvg
zhGh{!Bp0UG%8RN@LV_Sz{IOJ-OPd7*;lO{`Iz}D2Pfo;owXE)|(XDh;t`5I4UVM&8
zj2S_?gqqIsv|X;Rnf2k!PoLWsHOw?mzrX#Lg}S4Ig)<GohKlZ_8|0zy^<!sTbwRi9
ztQ_~bK)&Rf|4F>|$>J(;Of;{vu3c*Q;nmg0>Q7d(0cF+EZf$*EKApeiLMO2}d&0ic
zqeb8P;iiC1lJ--3MWs2`wA4x4!~GMp6-Gg6hHQFGa;Z?hyZWPTj&$eC)l1Bgm8;g-
zx*Y{QCmQYJ8OZ~UT@U`cliL4!w&D7{dH+4BS<XMUcN*zE=PQhM|98X*u{YCwZfsSz
zwH($S$;sVPHqxc;9qlUFd-t5<_VtqXWl3HU&30F{il)04rGo=q0duEgowl(7S4NZF
zFqLP{zb})SHE&H$34125I6YFUQ%-e^<t^K@<FXFkORTthz*Bo`aK`+^wVq6u@o-b}
zH=BpXUwGW3>v;D=eN*ElOTM=fS}*$ld^R{t!+g7L>Q~D1_gyr8oLW8FTQD`LT6(s}
zW_hu^z=n6hn0NhHtMRweq^AdgScdiIfpEwB!dHCUXz$!twSUp16o(lv#)k(8XAN}>
z#2;M{Go(B9&kaZysnbI^yEyu(tl@r!Cgq%XARER5giV>@^7;n-O>a_05lE>}s(65~
zS!g<a%+xeQlVUD6!+7!l;o8Oza`etjJ574yuv{Tm@c`jBJCokXziYlMCWQ|Wh?dZZ
z2SUo%DL{fAr-x{LXud;Cx?w_yB*c#gLQ3MX5STFW`N*B3lb_Adq=?J!#{}>I2_^;a
z&IKR5Z)Hdc^sW}FeIfj3nstA6=<~nPq!-tF1RKKxgv~|8mSXb%`3q$H+@b2tg==~R
zS$9StMB&QtaHtj5f!W+HfNf0kR9$hzi^LR?OXUbALm{<7t%N!$6sd^ZaMCC=8%|M*
zs0}BrLi^tdb}W$cyM<C<w_M?Pu~16x;3H)@5n)OJg;YT*H6({slQMqAAVMKcyv6{e
zqIOYuqo-vDzBeQ^BpOKI1FmUo0?1^WxV&&k_}zKT)<kNeK?It|Y|azFZrknFeC@UM
zI3a}?cLsqbg;;kM!9#$@=OqsW1qvy>l|fKB2;mMmGo^&pv^l;=uBVVTOsj?DP%=)=
z845^!TOUn7`(FOrxKv$`P6ZPMgK>>vcR>aa!vp6G1)7VoqTeORw~fS;WmSQ|!jh~m
z5G28MSdz0GNScevOrhA-Fku2j;hR{@j`pgQ9q_Z5pRvSXCW&11a^g%<#Go6)(ge*v
PH9T-{@_#P|eh|dpB)I<D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp.B5F90340138DCBDD.idx b/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp.B5F90340138DCBDD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6f6f303c6c042687f9969b309192bc1e59762567
GIT binary patch
literal 11156
zcmZvidz2IPp~v_48(0Fn6jqkk3NA*_WG3^-WM(2DM#;=%CJ*Ih<RK6i!o{N;1%eW_
z4-q3l19Gjg_83^%wwD&F)LI`!!O{xt)uV#9A|i;O5VT%Vkr)kn!}iSX_m9gt%-K&q
z``yoGv&n8UJIyxDhq#7@@FIKF(sX7)pN58pmiqq{ndLWs&onfgX=-R#^~9^mrPp6P
ze(<@?FJAxXLOSZ8d7N4ED}x_AmA!iQ`bj%)dG(ghtK%-Me0tkUVjp_vfouQyh_-W6
z_Fvh}%Ur)+GwsIKz5R{*AN%Lcv#z>l!r(RY;=8{*^!1_Y^CLr>H(Y;iMEmCbU%h%X
zS3O?7eZ#WlFYQ|;9RKoLA@G#A_*CfX6UX;eM$Ui#=<C%($G^-x)cnTon}lzUA3k)X
zcJkc1&2uk3r5<}N^B#F@`*L*$>l(Lu-XHfaePzXh`wuma{zYcxiPbaPj&Wy(HP5+x
zY3tj`U;gUYz*8Um<Usk?!+)~#k=s}3<(H+#4_?^Wc=58n$<wRb{?KQLeEQ|DsyiEZ
z7an~sefcF@_I-3@_!|e^?FW*PwpUtTcy7J0@qjk%{eIe#|Csn(?aYU7M}9f*mO-B_
zTf#)&-I^XW{y}#0s5`9HTNZS6ZQr_j^4PZ@och+Phu!Y&*R*|T{B6aP{pRi&^VZ&-
z@4XsXcIIcAIhWbw+JD8!(jaEd(I<m{I{IHXuU+}Hd~o?a3l`lxWavAGmVP#^v-JMx
zj%WMMZrV_N*Z%|kygYjq|MTM7=-B(k>DNVWoAUecE3dpg`M#;jzS&J(UmUw+y<iT1
ze&E@A9^ACRdgj`*{ST|-n`b=s&Vp+u&b#dA@BKd9dH-9BQ>$FZ4ti`(iFS?rNkhk%
z!_LomcJJBFo5y#ZSoT2fzq7B0CXN5i{{8R&jvxN@uC<46u@iTskL*2AIHTQt@bO8h
zYg?~*?6;m*-fq?VGY`|YU%obL=0A;l)}H+1=$23Rre4jwK5*^Ox%=Md$6GJYAHUe@
z`%U29M`upH?)hnV4d1ck#}7Wg_k{bdxp(cKF-w1@9RKdf$FV;>+wanC+xr_I2`_!_
zbGJ6_o_W)Chbzq=jKAUh>$>;FgXIHPzBqZwj!nY-H!WDXWtn+!!lp|Gf8tu&asTWo
zV;8?MZo}3C-~asA$G#eT<@m+#UiRFg6DRin#T8*sey*JQY|!NPzaKt%NSd?yF<<4S
zuijVR`ufs?ntm@?lzwF9>A8EZJ2mcqR8e0$?iXh_|7q9s_USv`x#*j1V}3d%H~hpS
z`#$*HN$#~zOUE|e+O}uMn8)vY<~O@HtT+4J`dfZm=CW~*&&v;e{SQCgv-{I*>jQ_X
z{f8AsJw744c;%4iYl7Nn?V5J#-Rz3iv*Y(w1M6Nc@AN(W^i#Wq9bbPnbAIaZ2RCfp
zwf*nEKm73LeaQD$|B-Io_K!uOYag+%SSgP@m$>|+m)9L#b@Xn}kfS56>e#huE%)#f
z(z(5dAOFYDtGfog_R;y(=l*)Ixbl>6%aP`Rf4^hTvuDJ&ez)w~muBx6acX|&u6+-^
z`{&jVx^DZted;?CcYL4Sde!N5kCortaq_26+<f21pDo!o(>re7kvICR&D^v0uW!71
z?xO5N2XA}e)Hl-K#?Et>8~)PY`o8h@qcfxx*`*zikcNiF_7~QNpA3#^qBqWN=xA+h
z?caT<KTdT28}M!C(feZR^d_3D>rsuPTDlL8Zh1Q2w(_Y@{@O&R)^$r`OaJbJqeuNP
zXZw-8lSU3~m{Hg5!`d(HJ~%q`-4~y2J(+#FiO#I+5sf4Ibsrq9(^*&iu;uQ(0~>Cr
z>xsiAj_p1;`u2Xux4b!ATG&MA)U~_OZFC<Ty`tZD#@%}c-P=TOs_RLOlV)@u9R10H
zF@NTj&jL;K`nn$4IJAHF!O{QQclBkLT=L#M%?-2adP?6ZR`<ctXLo$@%$$4odz<K#
zx^8Q1Yw12Xy85g-<O|=PyPD|qx*pv)`bXUdN1q>@eCEh)e>>Skr`7dEjTiOrJ~+Bz
zd3y1R^J9B&o8W-%+oZ7PyUy(&P8`$R;bA?zbM@;K@=Wd}FY5*84SB(_EDMeevEVqC
z1IL9pa6Z-t&KL55<5?aYAL79YtN>033E)Il1Sf_>a1twllR^?WKkEnQ5Bb5#tPD;L
z$>0KP09+sx0H?4DI3=WjQ&|<98dAY&tOiaCY2bov5L_@61Q%w*yz~9p+z}3iC-;)h
z>frQ{4$fc=a7M@gXR;<ZGh~8`un};PPz0RCTHvgZ1<qz|aCXQB7iFX1qM<0b7#jl@
z3&p_2**Lg(C=PC#JWcJn=XK1GXY`QsI8?tYTzC+IyF%8phV*{-dve6H5!jpcqW0#9
z7d4w@QL{P1qUN$3YA#1O)V{0_wJ%3}sQE09n$Hm)wU8B13ppa77PBI1F-JtyQdUAO
z<%opZpY@~m=ZGJ*oRv|_IU=JDWCN%JITApvWEIp(jwq<rtcqIA5f!zT)lh3WqM;6E
zgQ$Z!5=0%&hS!kZdrLS+!W)70td3gG5goOWHBcKlVxTs&CTcTBOw^HV1a%}wBB-sb
zh1$vy3$>lKQQJ9UqmE{ysG~U&MIFn=P{(p4hB}^&qmJiD9Q7=D)}-D$)@*rp4>>=<
zZynJ6lN;_>rpJc+Rl+j`DCJoQl=0lzOJ36pJ8!}ZJ8#MhJ8#AdJJw`j$0k_Vu_+dI
zY=(s$XL7LP5*+Nf6bCyl!@<sH`e5fv_+aNt`C#YE_+ZDIJnZ-c4?8}^!;a7JuoFxH
zc0xjcosbe>Cu9WJiKYlUF(JZEOo^})Ga~FHQ-Yn8kYFdJB-lwA33h(d4?BOt4?BO#
z4?BOx4?EeEVJ9bK*vTmwc5+6BUBC>$E|3VoE|3bqE|3YpPB9hODG3F3N=kv9l2Kr%
znkwwngbF(~rNU0lsIb#a4R%^WgPoSrV5em?*agiX?1G6P?1HHv?1Gsf?80W)hR3-?
zcnVM|ybvf8zO$EfQ-__N&|#;iblB+`9d?Fkz|Ke*urpEy?2L>7JJU2_XC_S8nJE)?
zX2yhF#Eif$l8C@Al8V4Cl8M01GA-Cy2@7^s%7UGhv0!JLHtg(#4Ldt!!_Lmwu#1{e
z*hLdj*hN!O*hMo@*u~5k>|%)+>|&`H>|&W1?BZq|cJV|UcJWjkcJWLcb~ng3^!&Ee
z+>w-%Kr`i;J>)zmYZ=h}TnrxYyNKs0{MD(8c#+vI!Xk5BghTdq5g#(&MR;VPiwMYK
z7ZH)AE+QfOyNDlI?jkaBpo;{Ml`f(nt6fAz*1Cv>9PA>&UUr^~K2HX8KM94eOO1HW
z!`H=Apq}-#_XnP*L_M33%hYokxk^3XBIg;;ItKO?#`6uZw?@3!^A@NVd)^ZDV$WNq
zUhH|R)QdfDp7CPOTVcG|V{3%P9$TO+_Sh0-vB#Dvi#@hVS?sZShQ%IRVOZ>OHNs(!
zD^L!5T#0hn<I0r79#^Ft_P9L5VUMdY9QJ%Q;=`V=Kz-Qrm8cJUzB2V;&sU{B?D_JH
z4|~1}<HH_bBRux_0_CyCmne@tzD#-S@m0!WkIyqa_V^0JV^63N0eeD$3fL1$RKT84
zrULeaDiyFN<QV~bLWL2qC)S9FJ+VMV?1?2RVoxkn5qn~lir5qLjEFt4!id<DYDB`G
zRG<>}q!N{|CzYv$J*i41>`8e>!k$!NB<%TX#E(6Hf%>uMFHt}C{AKFLp1(@{*z@NZ
zKlc0;#*aO@Mr7>C1uA1tE>RhKa+%85ldDw5o}6c7?8y~I#$KRC0@w=_XaIYG5)EK4
zP^JOw1*$ZFy+ED`U@uT%0@zb(M8Td?pbGYs5>>FLl&OL}rAig-DS1Z0o>E~H?5Q=P
zVoxnl6?<xls@PM@RK=cJr7HH+JfmVytuQL~v>MT{rxmD%J*`AF>}h4HVNa`44SQOi
z(Xgjg7!7;D8VO=ASfD}d1xqxDy<nLJu@|h;AohZJCWyUYg$ZIWTqEK0@Dw{-pyBnj
z_ZQA^iH0{JmudJka+QX^Mb0zfbqwq)O!ymMy+(BG=>@7|PcKm&dwQAb*wd?2$DW>N
zbnNLBM#r8}BL?=20yVH_l&FC{qf8C#8C7av&&V?d_KXT+V9%@(6MJTXn%FZ-)Wn`y
zrY82xDmAfZ<{1-vW`!}a7paj5_96uu!Cs_9BiM_SX#{(bDve+-l4l~=i&U5h_N*GQ
zuxAyhg*~f8E$mrkYGKc+QVV-lp0TiJRTv9<c8%EBvkTP5o?W6g_UtmXv1eDQjXgWh
z*x0ixjE%i$jYP2*Ezl_Tq9q!|UbIZ3*o#(a6noJ;6UAP%!bGtbtC1M?Vg(w*UaUl8
z*o&2E412LEjbSgAXJXikRhSs|;x!V-Uc5l#*o&8F9DDIHjbkrfrE%=V^GqCj@d^{~
z?Opr8zZ$A%Gw;zp9rK3Fo94Vo)u&^DYeCW>=S)!3fbKIuy}z4>J4yH{cx=);iH>Y^
z5(C-nBqnmClSGiMPGTY3oy10tc9JM^tdqo$<DDel%ZcEB-ZqSX=wZ^_G3v@uVdpjM
z8P{EOMS0^@ed`xb?LT#pb0wW;1av_6d92<miC-7zVYB{D7fxyQz7u}A?uhBHb2GIm
zw7H{wa{CPD8DDe9L}TKO&e!tZAHT5XFULN7vHn@KC)?*b@ASYKW7POJA8M;#+%mXj
znDg@BgZ*B4a_yewKdWEdHll5`bN*}VQv>_{@WSyYE?ku{Q~l0W+4kvu*;yYxvA+HZ
zxD)Pa&NXZ9m~_*m1<qC2zG`RRDNDDQ_3elsjTq)U0o$caTh!Os@@f6z)^V-8bAEEq
zjH^%2Z$G%N{vqe~nd^2+%^mam%<rLN12+uZ!@YTBeLMQf(bqVq;f|g7*)<zmb{wu>
zJaN{<h0a&tW2x`ieuIWTSigAc(5dathx~Ys(7J5EjU(FXXVzSv(Y>F%*X0FbT`Umi
z;(&ZE9}w^2fdrQTB)UW($t3~#U49_hB?ARq0U*Vt0I4n&NONgGL01qc><W)|ZjR;-
z-K7H=E(6GPnLrU&1juq(K(@;Uin^jeF;@&I?uz%2b2@g^fbP?<z2Da!w`UyO1a2=7
z>t=yCHwWZ%`+#^i4<xt+Aki%XNp1<q@Ad=9ZW$=x4ge``1xR(PK$=?v3c7<pVRv{O
z+yrhN$Z#7#rrQLHxFbN8+XAxPHc-?Z1&X<2Kyh~*XsSH5=VhW#UE4R%2|F%4b`2dj
zRB~<~=T1_8mE^)nap!h&9uMlL=PsNKcP_5qgGA4S-kZ;p_T1h}UeOEAoA!cZMHU>J
zX2EeH2aZc~;C!MFoG<MI$BR5TKFxy@L;;+T7Ql(32u@6k;3QE3C#5BDe$fxkpZ0^3
zMH!r&mca$Y0JuOp08SAVa7tPMr-~{#HLZfvL=Bvl*1!eDAh=*U2rev!C-i=MhSTBO
zdr22{aC%w?XNU$kBW-{)MH8HvHo-;22)IZ(0?raGa8}v^XNxvCJ8grDicxUUbQD}n
zjDd@#W8mUq99%pd2RB`w-g64Ax#LFp#vXEh=+qy{F8tvK+!cz%vk6{D^cIO1nJp3)
znJW?w*;gb!WWGpvWT8j|WU)v@WT{9bWPg$Pk>w(hkpo2%Kvs%GK~{@IMb?T$Lk<>6
z5II~V;Z69vT9N3;Mv)lEW|5f4ks^s8TSa0a+eKm{M~fti94nF-a=b|5$aCa56P$-P
z=Pz_F9?<=lIdCH;J+t5uEa?SelPnOI<bZrh9}u79frO+0Bql{5DJcQ@lYSsMDFX$P
z0U#x*0I5k8NK0x!!DJ99oD9!`dw5a@GLi<6nKXeS$q0~@w1DiS4HQjAfnv!RP&^s$
zq55Az-MZkypG2)#zHr6#3vbN7aQOGXdCnUR4XNHY=k~nGbd27$WY?Z`+a7IPI5#zK
z{(^??5A|=qu;<+x7A{?M>ykRg7AKfKeH;6^`ZqNX96abpc8d|dyk&qr`J&-NT1Sk$
za{R=rCS4u6Mi|OXxJ>i3|9_VbyLi;-F_*f>j=RFk`b5beP=l8!ZM>`x9Ttf$qULqi
HS|t8|xTxBe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp.BC9F6A61DFACFE50.idx b/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp.BC9F6A61DFACFE50.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a084a82038d83c059e4bc2cd10e9ab8799a13028
GIT binary patch
literal 11156
zcmZvidz2IPp~v_48(0Fn6jqjpE4UazlbOsTlbMNt5G9$(OdbdeMjis$g>dyK2Z5l(
z+Ex)GQ37(cvGy2P-L{ujM5@+DQ9u;1w;n2ZD<Xmj3PH;S6^XII-mpExe*d_f!<_x(
z!<Wx)vU%-nn`IcQxw^XGd~5l_RJy%aU0q#M?eEg`lG}e|>gvum)YUD2dRJoMO;?N=
zcz(m{H~n%R9e%(##;p8}&JUczPMEo7{MI{m-EpQe`l|9*i{21>(R=pa`1i-Pt?M)Y
z$ZS~b`0dKJTbg(G)gOBD@3+sm{=TsTSI&uT|K{L#2P>~$9@wz<rt`yEH|+a%*O6@H
zSm~~{i<i8yXSs0fn;!(<W^uvEz=Y$+_LMK5`^k~_DhH2!lV07pcl)iv_s0$$JX}3-
z{*eu{uR5t7{cHMT^5mu^>K4{9dc~ZVb}!tqwEdxj^&@_jUUq!N^p>OC*&&Uyu36ap
zLE_iHIoki^r$66c`t8u4Z+-l(rJ>TBQvIi|ZLPmzaqq;b6)k`4HAp`7=C_rt_1p8m
zd?j_wmCx?^{P57d2b`_@6XBK}&9A+(Mp(CBYx|^+cH4i9d!>5zvk$_*?tjOC(~ECo
zA|Gu`4H)w%yJ7g<=89+AySg@QTrqLf2ais9fB9oh=cXH4KGWY>`b?kMJ4e31d+W!$
z!i&$|uNkwM^^SelUS1r)tUU6J|4&E$>-JU4UXTwgxvzcx?Slq?cyQtAw$9=wBRXF0
zJ+onL<s<LE=$X>Y<@^(cRgqB-iBl(q7fpUXxMRl$6Ca$S?3vln_4Uy!*9gYYSNos4
z@6q+`=8HF;>w8EY(>U$P58H1TH|OdnK7Kyf`Oy0dlFJ=O54bE&iF955^SX|2hFqBT
z^6qn;w~y&MzWCwne`nqcj34v6efvK713&b;ZL1F5Va4xG9p1e^e^&eHfv3hNZ*0Ed
z$=|zne9#=~%REL~e*M=O)BmpDx9Y@8BbvV4o!pgvum7sSv-f<$k1^kzJ7$5|`+MI5
zznngC(yMLv4&8Fw&mMhs_i^XFv+vzEZAR$DQtY3Hzli?n<vv$Ex2doGx$wprkF&X9
z`}A8U9V$0|I_BmJ?}gm2A1Lj=?)8bcZCNiobZh&vXBQg>#;(6|;Fpet9S_Z%JZiz-
z(Q7yE|M7|69{qOUbz>HMboDFqk00OnmLtrbIHR0AJz!$%yN6C3lxD4X(o=rp+fUT@
zzq{&y7J7ipPdz^U)a;#;PLBQ`RSc~f{i|~u{<Ljs>(nhDUiSTSBY!bDJM{SDdp`Zc
ziMw}wRXn=x&X%29Mm}}Vi@)2xc8$^J&foLTrLP|S)STSl_x|{co!h_4G(UW>(sxLH
z_)}wp3ziLfwJNCf=C-zzA7z#{pBuBM;(O%H(pJw4FKpf}Z29in>2s5ZKD~M4woUIo
zf9SC@y~vLfUZU%s`^)^mjgMQ`E|V`mAHU}FHy=5&{K!vTgN_WlzGK_^Ror7wOXqhV
zdg?EOCv^4u>*p6%oPYa3VcALHj>C=p-@SY1%V)*+|FHOnH)d`bc5-g#wmqvq`g8NA
zU5n1NPWf=$mLD@4uRrz3lcm3HIq{39Z-4NM)3-f0-938F;k~_9rSDty_TF9RFUzbx
zu;}5F-%Ia|n&T|hz17$JvHq^J)1;-Dg&mKRx;p087bdTo^ym4Fbu(txbxiI(+3Y^l
z9w)m0)VIF2CiskhWCOjWrkm@V`*t5}-S3CaBM(N^sSPwy)5GhBH+3Iu-Sk4PW!dI0
z-)^8&YPzYusc-ke)`MMVI_Ay#vaNsJw3==m(t1_*!PbF)etml5iOdTPbb3t>s~^^<
z`(W$)H?Q8i`ktGA+CXR3w6or+cOPuM_rxN1fA{~M?q7FvO^+KgZdCWd*2le1T)TSX
zgb5Aw)|wt)KYm*G!PZ~4kNh*Qoc1-)n`(M+{ouac2V0-p^7V_e9@yt@pp$F5rM{)9
z`(W$J%j%%7Jv;Ahpi^slME!`LbRTSeVPN9L!;9WI(Lmd3`m*}V`gR{|UAH8)VCjWX
zJ+}?%*L|DhcmA_;(`VyG+PBH@e%-f)eQ{%lE8v=FUn`9r?tmMJ4X{95fCKUbJV1Pa
z2ND7TkQfkwq<{qE4S0d%fDGgd_<)pv0;C31AT6K)`2&8SU?4crzTFx-LV*yF9?*e|
zfB_T^gn`U}31kH<phzGB6b(dyVu2XYG<lk8-*@bLfxln(cR$=98P`hEbBnmL#I+9C
zopGafXNem%n_*G2S;C^`G8}3yOE}b?j0d$ROFXFg43C=65+1dX5l{<RBA^yCB5E;9
zMATA7LM>&9gxZ_&qV{Hq7qy&`QOj8(qxNNdsC`-DL#<>K)Jm2psMU;$TFnv_wU*IP
zYgwY9_GkR4{aNBi9n1t*lAiC-V3q{e0f#an)S)a1q1H1xYCTJI)JDcYZDfgoI-Cik
z4rfUiwV5$dn^|I_wlWrKD@!cYkxT@2BugTwqnRk`XqH4#$1*Y0u`G$9o*~Z|Z{L*~
zJ7&r=FOmKHac95o=MA`H87>R%S8>;5prmUaP}+4*54jCD?A&oT?A%E=?A&QL>{x?^
z9UEt1$0k|Wv1t}|oWa44i*vB!k{s-~GzUA6;enke?tz^r>4BXm?SUO{@UY|KJnZ-+
z4?8~1!%i>+*a>j~c0y8sosbq_CmJH`#JC7MF)6}MOpCCS3<-8pT!NjHlwc>NCD?fl
zFYLT=FYLTYFYLT&FYIJPhMgRjVJ9bL*vV-bc0R)gJ73%fJ73ZVJ73xdJH=37r^FT5
zDM<x(N?L)PYN)VN<0|abqzXGVt-?+-G}viz4R%^mgPoSvVCOgdu=B_Lu=6MVu=A(=
zunQVN3m)g<!O1|$;5?vo@SYwD86nt(;vv|Dk|EfI(jnOCh7LPDuES1G>af$(I_wO?
zfSnOHU}q!^*coX9c3~q7yKp=VyKpiLyKp)TJJT>>XU0v~nMo6NX4-_EWmvGY;uh?z
zqy;-GZNV;LL|_+*M_?C8Mqn37M_?B<qOgm`qp*u6qp*vnqp*t^G1$f8G1$eDG1$e@
zG1%QK-+bxSQe#I#P5@1pr(Yub9YIsS?so$4fZs)2oAISn7jYxAU4%vEx(J8t=^`Fv
zzKih4LKhK`#V#TuOI<`l_I433vfM>vWM3EYAuC-(K~}qnimY`J4cXsC{5@>H6MD7w
zPT=DEAh=$t#B~9#7gwIT*3h0OJXevr)+3jw>lAW@x_&^;F|J1#*q0gC_rUHdabwS&
zr*7=Ii`0!hcZs^O=dMsU_S`wfjXignabu6I5*B-Gp0e0ui<HG4TcRxX*a~H_$L1Io
zdu*9uvBy;jhdnM&IqY#o%3+TyQ4V`tg>u;Aatwz(uFP=Q^Hhlkd!9V?V9!&e9_)Ea
z)Pp@wg?g~($uS=6dCH6jdwiAf*yHn*#~xp#Jofk!<*~<CD33ip$MD$W%M6b_p-Keo
z33)1DPbg9WdqRl{*b^#Lz@Cs}1ndcAM!=p}B_j63JQcAg7O99mu|!4ei4`hhPs}kQ
z_QWzHVo$0P342nWO4yT%RKlKAq7wF`3YD-Y<roQjQkjvk=dBVi_PlxO#h$lFz1Z`X
zs26+Q3iV>on`6A#^OhMe_T(y&u_x!Lj6JzXW$ei%Dq~NsP#Jr2j*+n^ml+v*zAEuy
z&zGk@?D>k+hdp13`mpD#P#^YuImU-QUzzb?PpJ|GdrF=v*i(vB!Jbm03igx=Rj{Yz
z7zKMunNhH(R*8x|HBVLSsYR+{Pc2auduoNM*i&<iiaoW=sMyo0M8lqzryBOOBGs^`
zm8gb2twJ^IX*ov2o>pcw?D?z2k3D~$`myIPQa|?mCF;kXze4@k^XC{p_WWhWkG)`(
z1TVl_>|mY-*U+99oWUXuu17A>;3?z^4gP?fV}g$`urD*g?}0;A62e|6Pea%X6=?{2
zp%M*YFI1r+?1gen2z#M26T+TeB|7%>Jk_zM7paaty+n2F=@qJDPtP$r_VhBNW6!7(
z1A9iE8rU<6)WDumq6YSi3N^52<QM~cMwv0N7p{^p_QH7@#$LEc!`KU#Xc&9p3Jqf~
zoMXb+3zwNN_RK0Vv1jI~i9NGOP3)N^YGTi<P!oG*jxn)kmKhU!R+U)Tv+~r!o>in4
z_N)@MuxC}Mg*_|BSlF}5jD@{Ol|--?$<qk-B1Ia(UZg}L*o#zX1bdMj6Tx1j%tWvk
zt&%AAqInv{UbIM~*o&5E6noJMjbbmFW1`rLmYFE_VpS5uUMx>z*ozfu412K>jbSfV
zp)u^ma!d?+u`(0u>D~D7Kk6#y(jU`a9dicFX|sPw)vKf3(Vnm=?*Gq4L-^H?k;abU
z*9{NaKMbBW>9T7}>#pxzyLd|9DFf_(eRnkU9yeW^OdC5|C$>(r-|#ecjMK;6VxP0H
zUw_kn-KR8qUhd;(@xaesTlv<}&t9(uS`)3a?f>cFv-<Ece_P#BySQm!(-8YdXpi>U
z@yx27OYW~-+%l|Xgna^I<L3Uo|Ml9jr!Rh%Bd2)n&+@si9?Z=6?CCYN&(axpw%MOg
zW5@Vg$G6*$j+>Tm?LB$nvqo(@>?gy9*l$v{DQ)w6>(72wySRCDGjE@**g0*&skyBO
z_SF8%*}Y~v?NVdM++K4pk$nX;^y@xZ)ANM~J4tXeJkEzYNeEf*Bs#LuNetw0CkZ2)
zoy0`8I*ElG=_C>4XeWsx$2v)@hmZDO+kYqbx9e(O@`&q3++d&nIePs58`d>#IaIrN
z+>CMa>@Uu~{%W`G#q&-*UnG84eIfNe*Jr@cM{C<DgQv9GSJxM_gyzNlZW-26JKN)M
zjp(^y+zvMo>tKO62M6SFcz}2Z4<tAQAkiTLNe&6f>+k}}4jIVj@Bt|f1xR(MK$=4X
z@;m%MK}T?eeO1-Y_BcX7x<dyt90pL>5e70HCXnT@fFh0vP}C6xiaBDJ$UYfZJHK=B
zOkmFq=5o45_uK?7*2VXb+vx`9X5HXeCku{cS#X?_1IMu(IFHi<&ck}Z@lGBb&+^~|
zrvOf11#qHM1Shg0ILRr2lUNCy*Xaf4Wxe2JrwmSJWpF;H51fznfm56cIE7WfsZJG~
z%BtWrrv^@AHE@2XADo}{g9|!?qkF!^f^3lQp^!5KF2shw=}sM-&g$R{rvc7j4RB#+
z7+jbQgEO5bIFmKOSxyU_#aiGZ&Iq^&8vz$}M!`kdD7ct21}?_Nz)g{-T>6n*V@I3Z
zc8R`nY+6glZn^kmH+b}5$-ZaUPbal|$i<W4a1Rk(V|(rsu9WMp9&(FraPE{F94oTm
z*c1zn6FG2PiUa2nJ>Wbk4>(@r!SN{`oFEF|gp>eI6h&}iN(3i~5;!R(f%A%9aNd*`
zoGi-V<dh7~C;GtoQa*5ssDM*a3OH3%!Ko<~oF;1Ew3G(UFZ#jxQ+{wkF*vs8J`qd>
z@9LqD7y=hcg}~{e4o*+$;0)0KXQT{pVKEFYoC<?8MH8HvGQnA*1<p!Y;38rKTqG3%
z7Zs!6qNymjm>2^WOU1xVm8V`hOWD|Qi+sx^vY-5GznyXM7eC;xP#~`L@S|FHfw+;`
z0%4K40^yK71>!;G3xr1&3PeB_3q(Yg3PeKo7Kj&FE)W^nS0FxQr9c#9wLny4tw1zn
ze}VXsg9Q>?k1yd0B!sLNh>mO&h=Cj~kT9}YASSX^AQp0@KqAP|0*NBW3M7U+OP)2>
zewk(ehRzlJx_>hVZp4IZ20Vf#+(2xC1>zDMkSE~*;uAcOkPv{xga{-hBp`3X3nV9G
zAYZ}<q$Cs|HK77&2@S}f@B;-C!5MH5PlSN<gbrjR44`l#3}hxuAS+=3MG_I9Xd(&}
zOT;cw?YE%rY`^#`QA?N1TRQdP2XikT{{3f;{Xtz_vggCump)`V-Yu~^dX4z#o%(sR
zlXK>_*LD9_|Mv4P{j<7x3+LZ?TMeTN;!LmJ^?e+D8yfo$9PksXNe^Dr)X$oD+0a4F
z!!EyW%(&~vPYB!~4CcmOt+`tNze|T)F?__xtDK`oU+ZQ)qU80d{wtLhUJeZ&5{}HL
K#-tlf68k^GG#DxX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.F7CCE132C2B5DC0C.idx b/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.F7CCE132C2B5DC0C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f532a55879e20bc8ca92e361bfc8eed77c2c6bb2
GIT binary patch
literal 10826
zcmZvidz2I9!N&J}2bQqA2rJ8N1s9`eGLyMvGBXhnBV=YWlNebpMnEoMA$)k0qd-t%
ztrx^d)PQ`pvGy2P-L@aC6t~_j0*V#d>al{YhzKGm1gS?<Bu1ltVS9#s{wU`#XMcH~
zXMYPxW|N(rYMbUHwyrL`&|bMDm7d?HuCA`R_IG)D*$qE3b#<p2>*`iM{zhWS<O{|Q
zI<xWR$-i7cNA5F^FspxK@PnqXSI$~HVe5@=-1tRh%thssH@_nGp||aw_^(H_ts64`
z&TL%j`t9ne*R|~GZ`}9jziya$#U0}Yt)6@Jj<5EAv%m7<h|tD$lg|uq+qmcJHx6Yh
zM@qM>Te|F(-7AG7UwtP8wup<4hps$&WOsSQyblk(RoQ>!tMo%nZ||5be0$`;{)5$H
zXCBx%=c424;WyJCkViK!Q@dH$m{oJ1-L>S^<@4{`Uq9+s=@my;&1gN$of_J7{iRD<
z-c9`aH-`rt|LD%WrQZ&_bL%6wEZ0lhr23Cu+FF0X(!Plkt6KlmXRv%?+t-z?^*i#v
zd?9t|g-`DO<lwNk_qp5lCL*n`w!HMhT4DWOZR&^pw8j53{)OtPkKc{_dcchXKVQ0-
ziN3!nHE`^M?8cF|TC1L%-_^Bw)2d0Ay!+skcUC^^c5l9_^<(2N%b(~sXXog5c5VIO
zjmXkdcWdSxW`k?bWh06Mnbn7$2tI%4|87{b;%Rx`vODH4ykYQ=_x3OOd}?R$!%-d2
z^_|tYuJXSBC;CNc)=K^tg*DMj?h~h76S=wlkKtEeeRtBmQ<UAa8oRzceBoNb9QNXX
z(|0_$VZQau#MAu`sAHR^Kl<MMtH#g0_!l4iG2D6IJByMlU5EF1Y)*-GjkvR}<Ex=R
zOn+|I>CPL*b{$=MfA;U0w?Y%f{%+5n5C6aq`)2!^12@|7TT=&j?aiOkezxzi3CW2q
zS3UZB&#UjY=>3_8Y3r}woH^rP#vN;pJv*xTvt7wI(r*n|Gi1*05Bagyws~V0S$)3^
z-1EyBldgGj>g~h27ytah7k3?X-#+K|J=16E&y=qI{@|yv=b!6$(NmlI8=nZTeBpDq
zH13!&`<esgrjN#6`@>th_vL-1y_df{X>s=k;lA1PS3J4Y+&6B+g@ZnGE$O&#R{JH3
z-X61V)7~F{@!P{+54wEpqW3R;Vd2rEdtP%z*kfNP$3Gu9sqL=^j_sGOU-hW3{L0rK
zs_%Sr(LPPThb&AzGULRYo!1;6^G{XO*Npkq>5b2CpVl_5`@QqOeQNYw?b%^RAK5)`
z+Y#=~lf}d9Z))AyJ^HcRp84I5b!*LjH~pS}Dt+;o$L8jSy!EHMcJ4TtX}N!YrT@_U
z$j8Qo7p)llVpUM<t?g5fzn@v&a(e9UO5lNQrLDfFpWd=V=>F#G8S|0{KDu_(_RW9&
z<AI02=tF+I@>#n6slP1@O?<?@Y=u1HO#ISMwmooY<)NQ>1|J%JMaT9HYq*CWm(J`u
z@YvsmT-nv|<|jX_I`jIz!iwX<jR%_s{PotI&z%zA`NPuhUYXTB{P?`i?Ykd(|IaNS
zb=~|$+m!dlcmJ5#bj67W9xeT&``BHN-*E4zpD%uDhIh=|gKzg)lfGlk>u<kt=KRb<
z`)<Df__xwuE}82t)xFl=`mz3&&!$VuGfO%iA$4{2Z7;12KM@?=NUxh!*U?hn(!b|W
zJ5TidY51=5(7iErS|d%=;*s?un|lsU-28N|b;Xv?UT>sRYH@RYbN`-$6NkS4^7Bo{
zGEX<s8MSzL{qTN02Pd92rFHqqs|ync)LmPP#}6HUNzcKFlkxNP$4|}s`+&OXwYY6)
z+eJMGCtm;k(4Ve9^vvo;I=dE6sGl&s=itO!nlgv}xy`J(udl`KdbiPYaN>9Bhd+E%
z@OS%~>SosB$$ckVJqIWLZ2ss!^UCLeMmo6`52+v0zvtk@r@Oy==K6c~cpGVZEpDxE
zZSFZZapgI6@Rz=ww>Q#hwRlwhsPlRbPW;25#4`tP{>!mOI<*#`Uw?l8o`VzDElVw0
z{@lgA_s!&np8F)f^ZU-tACDj1)ZuY^#yJ1J8nGT8DCC*cOJ27ZnqJlmO>f8xP1emq
zlVw?GvLP0loSTCt$8yl*LL4-GZXYy#tPh&LkPn)?n};UP^3dc%JTwKj08N1vpecj|
zXo_wTnj$MgQw)jFl-v?DC02r_6q2Cncl)8~XZ_Ihhy2i#-7+*~R)(e=lA#%J2cQ{X
z1JDeF0?<_43N#f~fu<5tpsBi5XsWCVO*N!KQ*&$3)L0FgT1bOt&>e(kkPSjJ7z#o&
z><*7{o}Nt|VK&SIg~F41Nq6he)L9*xdPs++;WnUYum&`ZkO588Z9>yzO=y}S6Pgis
z1ey^x0?kM$0!_<pLDOO_Xj&l)nzq}9rp?;Wv_m#DqwXj)qihtK(NGkcF?S4_F*XLx
zSSSX~cDY?WcWUdHB2NLDCQmy@&UsyJ=sBOmsU_oCO?uBCo-FaK2li&X7<;qCi!qyF
zF=n%b#hA-*7;{;|VeHHJF!p7M4`V*VW6Wm>kFk&uFcz{zz*x+P7>ijVVk~7OjHN7*
zF!pEs82hutkFlJQF_yDL#yF4(U>wMj0LDs2!C1)>1!Fa%VytF~im{f_FxIj}!#J1;
zVjRqpAjaWLcs1#LK8CX-ydGH3=osr+qGN1i42+E|F)%hWCdOu#m>5Sg5sV{Q62aKY
zSQuMbVqt7&Y>e$Ju`!Njq8LZBB#Ln?6T>)`B{7U=$TKGNK96R~GtZIpl4o&4&r2P6
z=9nHEo;Yz&J5bWI04VLbt(Ux}7k1vb7k1vH7k1vX7j~@4!j6rzuw#=f?ASC5JI>@_
z$Hh6=aY+t#T$+QO&-B607x%%=m-NBTm-fMqH+k6caUOPjl7}6i=3yt80_=pi06QTm
zz)naDuoF!Yc4AzFotPA1C#FT%Nu~rlDK5cIN=mSk(h}_arXP0xxF2@@q#t(vv>$e|
zDZ@^V%dnG^GVJ8E47-3CfL$OSfL$ONfL$ORfSqD0uv6j+?3APeJ0-2aPBm57sc{u{
zYEp%rnpR<_nHuc0xCT2dsliT5Yp@HNLD&W3LD&V8LD&V;LD+@Ouni}&c(@%X8D0RC
z4&T;Gx~aoXkL$40lRE75v<^GNG+<}M4cHk;19nE*fSqZYuruQ(?98MIJ2P#<E@DPt
z7l}t;7fD877fDB8XPFl4thfa`D`~;bN?WkAO&fN0+=iW<v|(qbZP-Q4DD0x~DD0xi
zDD0x?DC}Zp40f@240f?(40f?}40hAy>E~Y7G<94nUkh}deBC*6-i&sWhMsq!aOSQO
z&kyh-$CIa?wY2vP?I}{v2ILa;oItKn&v(c<#`6FJ`!eJC7T8-QUi5kM)QdiEk$TbR
zEm1G}ycOz2pEt*N(dR8QUi7h5!lIAOQx<(}k+SGxOO!<)TcIra*c`*6k1aDS`nW3L
z(8uK|hd!=IIrMQQ%At>|P!4@uj^WV9l^G6wzAEvd&zGk@^!bX^hdy74`q1aAP#^ky
zImU-RUzzctkFOFQeSDts=;Mo&M;~9JJo@+w<<ZCI7#@9mnc>kVREdB-Ax{PL2}LTP
zPbg6VeL{r_=o4~`fIgwj2<Q{5L`0vMry}~qA{EgmmZ*q6u|h@ki8)3@pIBx@^hs4B
zp-;+F34KzLO6ZeHR6?Irp%VI}93!DmDl-!L{8i#dpFdCi=<^q;AASB3^`p;Up?>uF
zbBrH-{xai7pIjv}`s6&7(I*$Fj6S(UW%S7vDx*)%F*5q(G9#leP$dEM1@bh2zCe)%
z&=)At0Qv$I8bDtl#{|$9C^G@{DOIANPsvjSeM*rk=u=8mL7!5g3i^~Bqo7YIGYb0D
zDpAp==BbK4wMbR;sU@nSPpwcDeQJ(T(WjOf6@6NjXz0`OR70Ouq#F9P64lVBRj7tO
zEyrl+)5?s7zF?IE(HG3qAo_wu8bn{PM1$xHR%j4?!5kArU$D#s(HE|g@DK2&FPx|0
zwY2vYQn*ON8<0yhd;+;b!`~t2nD7G(?8{8}TVTCPboA+Ys-sUYQXPGIiR$RnD^y3H
zo?~?M>19SopHU?S`iwj^&}S5>fj*-|4fGimYM{@^F$VgKGGn06tP&G_W}cepGmF$j
zpIM?N`pgP7(P!ou6MbfxG0_*Pk_h@Dc^W}qq(~#^i<D>teUS={pf8eRBIt{hnF#u<
zDzVUK<*9`}t4J;MStV+r&#F)heO8XK&}Wqy3w?H#*yywK)JC6Oq&E8O61CB1SE!9X
zJIC1Qv&)Q)zG#(1(HG6rDEgvB8bx2UM5E}7R%jG`(Hs*+U$o3b(HE<d82Vy)8be>K
zNMq=Wm1qoou?mf$FP39s=!=z^Sg&v5{r|43oKAm0`*h41JZGx&Q;t3zNmnxAP(1j*
zHxK0>dYCkIjJ$kg*!gkc^lQ$)thD}$zO|d%`?n8t+RlZXhK8OCIK8i|!<{6&1<ovb
zC()6OPGTUNoy0_rbdm_N)k!R5yOY?+(M}RYj&+h)FFO~Mnj3m9B=vroo-X3q0&f7k
zUBrvbb`cht>mnSouZ#GQ`7XjE3tdD&7Q2XuEOikH+22L{$Z{8vkpo>MfUI;81zGJP
zDzer^G~{3x3HI`>G5sxWhSpA-I@%_+O?Up6nmWcC<F9k>+IxQf((2a^fBbUoo^MOE
z&2dWipE5>{{l`PCwVRs<H4k-uIQC$_SD#q3bJ^Xsn_Gvsj&d%SIo}d!L(iqU-a7yv
zdYiTk==;-4M;<@>7#rR0cOGL;oxC?Q^W(?Y*1k%2+&#7T-PMHI6XrXgc=O7wecP8j
zY1X#G&l^6}dCTXNYu>XL0(-sq1lX=jUD(%n@?>q>GNy%hE{yJ+e&vaIZToiDytDhv
zb~|5AQ^(vsbI*};YwfM&*-MFz;P8NT19o!%xV-i$M_oSZD(AA`;iGq7wZ6IgK<(!7
zGsiD*{)<1A`kwALaM**jo7;!9w>h8l)9Zznr482&A5pu~=kknl{w+-%UY8e$b+JI4
ziv#kxd_cU52NGNYkmwSDB$ou_clm*2mkbne1%MQn0;IZBAkC!#1zka)uq!;uxhtDG
zbe9ffxC|iEWdcQ95g^ND0og7aDC&v=#ayv-<XrKs-Gyf__d0jsN!R9ebX@n@6XB3C
zLnNo=JfdsYPtIQbb=tMNNc4>By^B04&n>;=6}^yoQ(nkek%f#+v5;{h2N{>*AoGbn
z$b2auWW2~j#;1751W|xYNC}XMq6nFo5+RdB2{I`qLFO0zkoi-7$YfE5Oisy=1;hYk
zfm8r8MN}YDQVL|Ms6wWuRLC?@gG@_lkOjpcWWiJrvalE)*ZU9*r^2`Nk}m3y=_wsD
zLo^^WQU+wEXhLSDOvoZ)1hPmf0+}URkXb1UGF!AEvr{%?Q85ZxG!=y`CdMF(rDBjx
zm8YJ&%+b`5kQ3*~Idfgm(DRE7@Nh2>&j$R?Ss-3ywm?{9u0S|sUxE0L`2yjQg#r<f
z#R3tLr2>(V{RQGjmJ38i4ircLSt$?&SuGG1St}3?IanY;<Zyw6H{c7L0@0C;0x^)y
z0x^*z1rkBF3dBOT3&ci_7DyC1Rv<CtS@NuL&NmL{*GyU)dVakG?zDtwCY+TLULZEX
z0&xir$d~W|@d+MCNC-e;LIjc$5|BUP2a*#qP#_TiQW6S~noxnXga#B$1cAbd@Jx8t
zCUhVpVE~y46DX3109gqO$WGWm(L@v|mWZ9BWvQFypZ!Ii<;xZ<pLX_xd1nv*`7_t~
zpsp_2`{A5(A2J;W&s#ii?xpj~^$X@C=gym7*Ym0VlP@?|t6Q*S;Z2Kc7+Vx)`t+^u
z=jz|sG+@xc^Xz6Ld}(uoJ?Z>mgIk7=xP0vRD<)hSx=I+rjk{R$wEeeRhh8vp)aZ-c
hmyEg0%lbsgA5eoADy_V%4;dPXE~Mr)6D@M}{{x5tEqed}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.9DF3794E443669AC.idx b/.cache/clangd/index/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.9DF3794E443669AC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b70f0cc02ad62e2154c2f3d4b872a402f8b8696a
GIT binary patch
literal 11156
zcmZvidz2I9!H4&K2bRDt!pd@S1s5Y|GLyMvG9ch3$=ouNi*hq^5y&otPmgjG2uiH2
z6)_SeAfIilJqA{{T3V@6YrPZ&OAFZ6qk>;WL=ZtCcsZgXF*evQY|rfe{wU`#XP-R#
z<2h_@J3BkwHqEtMLxZ-+Ub!Tl>FC$c(9lx<yF9b(rXQGwhEq)q4J#kpm0WV|g_DP#
z+4S<YKVL{k?>CP!tAAziL#MG<&slfXwwrg|{AG3g#g)%)eMRg?@7Ndq=fe}XZOne2
z-L%y8<m&0yx9%Qj-1o>oZ<>AOT~mgxo*&=w^?`2>RA0OxxM}^hXGXPe+WXC}!@26w
z@@?yvE_-FqO5y0&-wT0f#Kk9qS06jNr*gr94-UUkJ#h5v%-ZHRcibR+cl6+aL$%{)
z9@sSR;u8~&{3G)|d1TA730qm$_*L`&uzSgC%RBBn&^YE7nH9%Y&1yTsof^?R_tGV;
z?<9Zut0RL?eE5@n<tIn}WZT2HE!WGhN{t`Bw5{>NrTvp9SGE1V-!S>)tKU?&HSQ?<
z{Dt(T7d^4(<3l6g-0yDRmyEQ%*80*5>x2#aCQkogz{DH>=ZY6<r#^Zo^2@<D5BXy0
zjZE~t&FLYNA7nR;zTH~&L`Qe`md&fwaqm1h?d_Ehx!qerZ66tbUH<rhc{|6xy?fjH
zyCO?Z{dA%^kJ;$jd)Wo0A<XK-k6-@B!@s_1&5CE`{mbs^Saj2{;qM+;^2PM7(g$NY
zpYK1XX?^uQ|G(*%<vA<)pB2|c$K5B+xF&Mz)Zb~Zz4nfJ?=)r4oTl!tj$E`(Fh{;P
z`1D;5ZtSq03!ff%aKhx~nUB2N5xQdjB|m%rcUsqdZ!b=*bRF66u{kB$eZfx}I=>!q
zcINZDPj}rkx%=4C`*VNKz7f1?@>6^Fe(+m<<hR?`9K6|1+@3zPdtc$y#2@W{^r}?2
zHT1}DJg>dest;ryqHVwY$Lv}EH11k+{10PVKHZ(#m3d?En&I>Ie85k(UR^MGvDN>(
zz&$^orC#&m^gBmxz46BnzPS6C`_6fH?wvVXf36(=*P%~he|&zx#ZPY;XnZWZ@}<w+
z+O%WV4c8p3G=Dhxy0dTS-k0~6_f2|Py>aVC;l3L>Ry?uP+&^XGMMFPzE$O^(&eU;>
z-yFYw^S&Q`_T-UohEAHi_`OSBSaj^z-aoq{?C~#^6JHEb+h0F;{D3rf)g!*jE8l!D
z;q7lP-ak>lhb&4zJnQ7Vo!6Wg|9=xiea-k^oZj@u?K9eEY<>6q@17oe_tf0TV-N3H
z@aoaqcYRhmvf-Atom<B~ddG85?O4Cg9B|8T_@^_MjDK`~e)t=|zkBD7&$6xeAE*u-
zQ5gN`6m9W}VK3H%360kF=_lUHE^j?Oc~3R)z^mnLzGt6(W{0r#+izwqNFDs}y3O0S
zy#Bj`4}IB>{BZRj=!U2NwkQ~W*uHFqe8HK-r60fgz~PmLf8-f<c+{1h+c&P^9(qhV
zv-{wqe;a;v_n?1#e0J5Dzw9rrI3e78sCn?~x9@!Zl=${<mwx}soUNlyEa=+4XYG4`
zYW=YL)-T(qy?e#hAF`XTJo&&Q<+rvTzx%P9?)~J88=s!#9Y6ojoBh^g?ppJgH+P*m
zKf8ASt@oe!PWtP(`R;PVp9fk$G~QO5DJ{<~>3o<pG&Huqv`%~c^07_y`Z*1qt&Ob%
zdk*#2iJm`$zVAAGZ*0PhCYr44(T$^9dJc|mc{bm+;+aqX(nP1#bxUK*z@CGn-M_Ag
zZ!LWH?k0LeT~BG8GPCF4XnyStAMw*JG6pwXSJ#t9Od8j7aCGorUwyIpc=p*QI;*Zn
zHI5q4b8vLmk0<Uq@YIh!AKWmruG>enU)*zW^xt>A_vC;Tv!80BbL-mO=r(!|j{dY`
z?4Nk$i$D{-wyuXa4j<TaaP;R-Wm+cgKe)WPVRl_l?LXD(IXL?C)~}wMd(U2P6P;Su
zZH;X$JqJfupPw-7E8osLo9K+X9@9AHyq<%j&kjvKcj(r?9&e)4>-zl0^9S}E9Nn-i
zy?FWAaeepA<%4?elfuq_b#3|Rim}a|9=B(_^Z%<;(4+Q|*X;%84SK<`ZWbIHWWjN6
z4jdQc!1>%haK4}q9Pj4A@j)J(;1<9MK>?iT7Qu-@5uD_fz)3*~oZsyS=MVb9$!-~(
z9F)NY+yQWbU;v!rR=_Di1zgA-8t*)gn>$0nklIIKcNkna7zU@hRd8xh1*f?+a9U6U
zr@M7<dQb;vxD9Yd&;Vz;O>kz=1Q&5fz(s-)aF*KwX9X>Aw%Z102W@ascNAPS7zGz|
z$H2vcF>rBr99%pY2RBWgHlg>K*Ev(3*-OsrQ2nWJ?nMZm3R%x;()Z~1<cMbjus7>P
z?adJ{YBtNFW^;r^&1E^%T#j(4eOVuBUyk@t^I0A>pCdeKAuFI3azsEaW<}Iuj)<tG
ztb|(15ec<F>qqU+5kG1<E2EZkL`EIR22clbB!F7UDyWqlQBa4nq1B}CNfOGD&<5ae
zHjFx)BVp8PRz<Dmh>BXvYN)jw(NODI9kre#I%*?ppf+;EKy7AC)Mk#Ds3X}3>PU`6
zP+M6GwUr|lYCCJAwsXWr9nD5jM{^{KI+l&0j^#)UbvzqK9nX<C>e=$_tNNZ;bL2U_
z<h+C5GN|XB8=hFE$A;%s!ZQ^p<yi=n@!ZizUegOZZ^8>ZZ^{chZ^jEd)?{JFCRo_9
zDHe8YhJ_txa<JnP9PGFh2Rkmq!Omy;VCPHtVCPHuVCT#DV8@$0?Dzx^J3ht3j?eJ0
z6HEbiLPCI@kP=`gWCYlWrU*MRA;L~fiLethBJ3nnf}NC*U?-&{*hv`)c7D?jJAc9t
zJAcX#JAcLxJK2<BCnsds$tf9jaz=(-zzo1HkO;sokP5&qkO{y}F%{S;2?cgaN`ak{
zQD7G`LpHoPB|=kyQlW)Fna~}56gI=K3n#*`3#Y=c3unTxQ%x0iYC?sbno?n>W>nZ|
zrUpALp}|f|X|U5W8timahn=3#VW+2b*y$M^c7|!d&PW)rGg1cZjEn(0(==gcCQR6w
zDHC>P#)MtOjKD6Eh`=tAioh<CiNMY>E!bHJ3wBn@f}NGIU}u{)?CgXMJ3D2=&d%7d
zi<(i`MH5ljMN?7OMKe*@#mpG&Vu={+VyPJHVwo81;$|Fn@kAVU@l+gk@k|_c*U8uQ
zez(-znUs@2v*cO5<UFG3pq_KFz8CxsS4Yx$k8sY+&KuNoM%MR9p)L}72H(uPNElh|
zA}X@hMKol+i|EKk7cr2{E@C1_x<~}s>LM1h-9>EVXcvhh$GS)iIo?I$ee9fJzDNf3
zoLct%KRq?#ISb!MdJ5FDj`qFFc}mo?5xGn~Cy}ev^F4B&@jSr5zQTCE1NPR47kl0U
z^<vLkqF(HI%hZcKZ<Ttn=gl)-?0GAU7kg}tu-Ib@l*JxfqAd2<GG(#HRw;`;HqWrw
zV=D}cJ+4ML>~RIkVUH_O4trdga@gall*1mEXE^L}6^6r}uSR^>^A)HMd%hC&Vb52l
zKJ59b)Q3G^p7CMNS7ChE<7<S+9$%n5_V^OzvB#Gwk3GIhdF=6dhQ}UXVR-BbH6mb7
zC{O`=LWv646UtP;o=~L%_JlkmU{9zp0`|lj5wRy0sE9qWL`Ce0Wh!D%tWptsVxAGP
zCsr5{ds2-^*pmuW!k$#3685Arm9QsOsf0Z#&q&ylDvX3Ze~tLD=Pyt{_WUL4$DY4T
z{n+zYsULg(Jmbfnzry&jC)bFKJ-I+-?8zl6V^1zq8GCY-%Gi_hjEp_G!pPVQ)JOn(
zfdUO+FHoWZ>;=j+fW1JK2Cx^%GXd-cDog--N{uMkQwmhUo>HO;_LMSJu%}e1f;}bA
zDA-dfjDo#TjfBp^`%<VtL+hyXURVFQO^Jp!BA03CByyF8zDLe8p$8b)SD4Uuz~LGR
zV=r8wVeExVG>pA)nTD|!uF^2}!g(f)y>NvIV^6IS6?<xds@PLYRK=cJrYiQ-Dpj$k
z<{1@xYK2j;r`3ptJ*_}B>}e&cVNWYl4SQOZYS`2AjD|g}!f4pjYedJMUZ6Vm^b*yv
zr<bXYJ-teG?CE(%$DUqcbnF>5VqniGPy>5Li5l25%GAJ~QKbg<j67pt&!{j4_RJbF
zv1b;ji9NGKP3)OvYGTi<QWJY-o-wg!Ru~g|ks66$FH)cp>_tj6g1tzYMz9yD(g^k<
zc_xCrNQH@D&#Dm%dscy3*t1I1!k$&87WS+vwXkR984G(>g|V<_*NBZhyFhL1*(GXY
z&n{CNdv=xD*t7GDjXk@<*w~BKNECa~0*zuXTB1?xMawjby=aw2u@}uVQS3!4OcZ;u
z8i`>qR-iHL#Y!}Wy;zyXuotV+81`a$CWgIOg^6J=UL$er#S1i!y?BYnu@^7XIQHUI
z8pmEd&&078uQ2hxUikjc8>*)>@6&#r^M}o!?wob>>%?=A#O42c>j-}BL!`NL^rX?6
z^Aq5i*PMS@dBc_c>o-puIBkfty*;MC!OfaDl{R;_tL-zLi$l$wlZ;8%JEt@E{P?BS
ze?Icj%k^!0vVETON%x;JMo)fgZCm~3mZ2>poF9ojIN-I%*X&&O)B4SAquRze*L<8+
z{oLr>MJG6C=q8?L;GDtRO}xl#H(`;vZo(n^x`_{&?<PF5&`ktnv73m<Qa6#1{oTZm
zEO!$bInYf4$VxX+`grp*gZuycrK69X`#O)E=6Al%Pk(lAcJ@b)t*d{X-3j;fzB8IB
zH%#ep-Y>VT+}3~Uk|)gicGP*JMmVQg+m-2y`WsJtR=>G*d@Ju<dfYkl>XQrF_wT9y
zlk@t`b33Kx&ISDz^pX>*pSGU6me==H#|O}n!RrU_<ldT84~&^KCgfa(J96x&p$#ot
z57uv<G<(uQ=h5(q)c^E=AtN8G-#l&jw05Wa$y}jz>7eUJwbiezxjbW>aC4{E<ppA0
zED-17fP5|=5bxrF1eX9Lx<nw!B?0+eejwQ;0|i_GAjPEsg<PRA&OOoG8Fqz%RF?{*
zxilc%r2`o*1ITolKoM62$Z}agw#x>Jx}rcaR}3ibiuaOpId=4*p3AU(cdv)_@NgHf
zULclbfjE`}^07W3p5=iARsa%N5lCVsAV2E|l35ujzy^R6RsjmJAs+4mHVmY)Dv-u%
zKsu`f8LR<hvL;Z3jR0A!1!S`}P?U`V#n>27oQ(rbm#6psnCLUtmi2VX)^o4E!^aPo
zocqUllGJ}-a_*wIb3Zw+jrGfO=PrgjH`kv*qGw9q-RDVrZtEkj=mqCZd%>|H3yw{*
z;5d;3$E7)NKG6rxm-d0<MIIcV=D`V~08U5?;6zabC#FSkk|=?b(h@ko=m+Od`@zYg
z3{FnV-~wU*Tp%3)r-%wTC9Qx9iJ>Wd54un~bXy;V#W1*VIt)$~Rd8xr1*eG`I4!M#
z(?uPep4Pz`q5;lG8{kaQ1ZSpAa1k*AE|QLbvqTG=mA1gyq7BYY+u)*N6kIeN1s4-z
z;9}_*xVRVx7f;8*&5&pGUP5c`yk5S(mz+19`YYMFU;KclLXmhj;<pDy;zeeQghl3x
zghTcfi4U1C5*}G75&>B(5)oM{5((K~Bz|PMNMz(dkpz&HB2kb-MH1Qwr`+Kp2_vgT
zq9SWWq9N-=q9YqcVj!DEVj@S1B!X-eiG^$ziH#gBk|=VlNMgwGB8el<mFG@zUf!JF
z(7A9>&u`|yotX5@hF7qp7l=)=KwOdo@+Eyhe3Azek^+#J6oI6q1msWpf#jqN6i5bu
zl%xU_N`_{`Ye6y$q$X7$EvW(NNgc>Y8bD^!1d1dhKvvQMvXeGYG#LeoC1XJGWW1N^
zzXf$m$GKmLTE1-I@)_qoSa9y}pFi`R4;mU$eIL&2{gCN=Y9e|4@yb1~H!hr)n!li<
zq356ak6+mPX$=dPEV|{!I>r_!n11~m2e<|{H4h#-<UG5@&@OEmWUJ?o9M(GOf=QFF
zxbmv2gCSu!H|3Ivp7#HC>xc_Sj~RQhd))ZTysS@@{DBFVU!=70vOau7B)W*2*Muz+
F|9>V=MZ*99

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_fixed_nk.hpp.6F8F8F18341308F9.idx b/.cache/clangd/index/device_grouped_gemm_fixed_nk.hpp.6F8F8F18341308F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3feaab5c0cf7fc315a6cedddbf8b4f2188029830
GIT binary patch
literal 1692
zcmX|A4Nw$i7~b34`{VZZxSzc{@4yfcB;?0LZV^x?0tS&E4PxZSi3*GW3673~kTM8U
z1cL|-CBdYC0Fg`?7(j%GiVOy5V5A0-$}9$Qbixqm+htyF=APMkp7-7Nd%x$|10g{{
zl|l>)crC<|o|K(PU>K%Ae^Yi=N<9z5gaQn+obTSB5OLE}Js&Z}?`fo;9Qr`Q%<;Yv
zSu*bDO;x9-kC~rUu5~5lH_wRWPEKKYIk&q1nEtfxNa`pxAAhV?Ky)A8d8U44mn^m<
zT=uX3QhTw(-8(Tmb}i1NM)+N<3jZZ0aN6>~cc3Mst$#XK<2%rO__?7vEOB?ou^2&h
ztb5?Zc+aAZJSR8B&)e)@+v*$dx#Z}rI}_+7NW5g8DJc7*ub^mSd&p0P!rl+llY650
zFFF7A!;Y6rA1}=YFAeV77)dPn71veO<K9IVEwdMH3_5rbU9&;8f;s=1eM?%q%cjuE
z&;+mNJzhWM)H^t=G(2i7ebMVm_4a!$3#<Q?4R`8l8EHV=3T<x;E#CB{x_?o8^qyO=
zNWSa)uv->t(;)M)U}iirNcPURdkfxmiW<p3($RNiYo*siH&I12ezt`1*f7KNFZVF}
zHy=89*`uOaV|U{Aj9>KqwS@1|KA{y;W!+Jlztql|`N{RevdOQ^qEr_*_tMt+@SR6<
z+U_>|oV|KH*S&ONkMxrGEos*8C$ER~4S)V#^gsLpYK*~5x#_PwL~cn69R6%-HcU{s
zScTC#QYYg86C+2_PqD}M^k-2G!4$p$$x3UbjsuXl#OPC6N~0GjoI<jJG^jWL*=0a~
zb?idZNeY)E*~ZakD+eHlnZuVy`pM%Iu0yiITA||r<bxe?<!zpV0Bc;RTt5xXs2CTJ
zGtX5FnI4SvQn)qhwk2&<9Dr<o)9&w#a!rOcE`@HR!srNc;iJdnHDCXFO^6E=80{)@
zeTD-}LZC~*cr@?S?G&bP<?Rt6E<y!zk=%|0Ov1<43##7eoH$WK`cb$Qny)3bG7ezC
zIgCz{U&X5_T#sbc7S%QmKyF&;c|yFO6G!0&B&)5|N)AA7aVQ?8%SVT-aS~4Nqwvv#
zJlXisa!dPQJcUb8w}#ZnH~@K7ycT@mnjn(GWk}YOdKCvCXC)<@GSC*_J<p2s95x2e
z5zQJ7G<c0@)@Y!?V??tC1MTHMU^ST=tLBk3Lo!P65-dm6tlv-#GEc%;4CW+U_eqO)
za&+L6v~ja>8(4~7>B;y(4@vtNe4OFJAQfl;8wnTyt!Mz70vKTFzcZIdjQ)s|v`V6)
zz%!Dzk=RhM85E;evN;KHG?$cj1{9R+<rmGLUSq^0tsyj0u!N+wgjNdSAZZ<;lfp%l
zw4TsQ!5xw|5C$n+E1D^{(ZEcSwk2$(unbz!)5i!a!cOQsY%JkCFcQ&h7@-PwLNpsf
zpc`6cg0+s>^T;1pd)Mu-Bg7I_vzdfy7>%xvjUiOS&ZwHrAXLMGh-R|~G?tZ=mS~SY
z9%#p2h*_FDYwEfQhk2@)vR*ZLu;09zh4r5X9xzP%^bx14J#NHCW(w>1T8$d}A1m25
Ay8r+H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_softmax_gemm_permute.hpp.3C2429E11C534BCE.idx b/.cache/clangd/index/device_grouped_gemm_softmax_gemm_permute.hpp.3C2429E11C534BCE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..400472e974812fcf303814e4d6d1a83b4ab41dae
GIT binary patch
literal 2222
zcmYjR3s4hh5WXC5yo+E`E+L#mNSX%)E+pj<9%H1WhVcbfkqX7Ap;Vw+KtV-BMIB9*
z0>hvc1gwG&6iZbsNOc&oYDWvzjylj$DYh_}js@xqsr~Od3&~8rx!vz}|J(g`|3s6~
z*x-s`v8zo5*_J%BBZgs~=%1UH^BIj}Si1{`6;yXrq-FLp!mf#gH{Bh_7FS;wPLHoQ
z8AbA~b(^WZg~P7P%gYANr4HrSt_-uveq4Pvfw}*0jsI__Ha{Kw<LUdRgV!I44;3Vg
zs($~j)lm3c?>=~`p5mqI#zN{|kR9%WE%j-Mqlw$3KK<<~%|PRp%#z_9%*;(UlRh*>
zJ?S+ZESM9LcXN^6M|)&TY`?htjCnxxL)8_-0Y!s3d7IxZQ%_q^-<k33s-nn=YXL8s
zCPF{mTTygxy<X9_!f)&G;*i9aw0)w;F<WSxTfgsB!^o^<C68hpe=_vVe_-BSoO-}#
zTZR97#NDd$n4$aCmkuuwz1}{$e_iJ~U6Y|Oq@g#%eYw%xusYQc>Tuh4;G4$-8(&<+
zyD}Yms)g$sM>Zvy<p-HM@8_{u7T%iDrSWI7C5yXte&P2HWTs@KY`=cX*gjX%6>|3a
zOuE^YVz_MW{M%PhQ@JNNeyt^`uQ_FU{=9<I-#;x4exMmTKhV-qx?%Xv=Sx05dV*Jf
ztTU%};n(T6p4}NMc|(aOYNM+iZM#aFN`$k{mHBiVE56LQ5La1<@7yDOVomo>xDjw-
zPwFmPqqOdWdEG6yu^xP<t?tsvgx;QUvN=1xHf+g1nqNnrzX@+`yw+ATdpz;H_GR(X
zg7I;)ZKCoMbF%nosIqaaZ*KA$=Fc0hDiwM1yk7WP=iNIkd7o+g81M1hvt3@t%haz)
z(b$6&EO%q}$`Kxhkue_S*@~A9e4PIdMk#3}!vTf{i~R}QPgyTMG#}^V1fq3nosk2e
zHAP#8G@>2@AD@BfAUY_R1E51{zj)Ppc*i+s+(Cd*QWvS31I!IaCu2N{r|onWXYmMv
zk55DKTD8{50mMfiiCB0m`>vjkyCPan%Y!)p+O)uw)|8iV)dP1Da$k#4ajLjTpqJ^}
z^`)lXmCm@c5TpF2`FU{w^r|aP>qQ|ykMnU)lsu3Q4CVmnjL~n`CPn^b<>PLMX4H(4
z1EBLy$t!KkoAUTLiD(6_U^oD}=+&m^Hm}^veB1-k0dxSv0nlF+&M%W^%69Pa>4;X-
zYK8-#4=uU#*wz@R^uQgVuyDt43h3rU`k9}`a+{C4plBH_3+4dmm#TeLmeZ*gKJJca
zjap;m0O%Y`dM<`83s9ct&M6Mv0>zP=l^fikG;*^tgBuh^ZdP8CY=gC|Ew%`BMQ&D9
zcpLOYZdOcigN{5mPCv-`(SjL@!LlePF+s^lnE<LLC^;z?Kq`V#kO~2)K~PFkDS*ld
zN=>Q-&})JUAOi#-3_%5wfdZHTf(jyo1W+SEF(e~^0YH|R@tv|Xq(%TOCMb<q<2z+(
zNv!~CCn&90>pNxXNSy$hOHewo&UeZRr^Ee0bb^YYBkU|at+%rx=}0>(ijK0kR8hfG
zS)%D^J1$xkZO6sXF_RpqfV*?+0o^C47+K6@NNQ_DR?53Fe;z@F`GhIq6zCrklwBZ5
zW!di}FP@P*2fZ!TNoPY!=plNW-6wb(#6xa&kH8H|Ave1};I_s)l-Nq&h7)KigffZU
zR0!n~yQvT=B#;<Qg-|K6n+l;?VmB4S012Fhra~AffvV6{2!kYcQz2y(Py<q4s!`Y}
zFV!mSw3g}=pe~$*>R=Zg>VOuap=Wm+dURT=U({8zq6)Px&L>U*s{v)AB<#LJk|&JQ
zn9Fp^iX&M(!z7Rs#d^j`AQnCn#j=+H#O7GC%zo(Fgp;wEJ-_suuL$!&y-6lFXH{+_
kd%=qSzhznUntJq}dK6DS;?%<`ee(MIj`tkJ-d-Z?Kgz`FK>z>%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.75AE4D40F988002C.idx b/.cache/clangd/index/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp.75AE4D40F988002C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..de53f1b7204ea938cbdafaf6245b1fb6f625e0d2
GIT binary patch
literal 15912
zcmcgzd0bEF_y2sJd++1Zr`xA}-#2MTr45mkL{S&XG8oJ-ma#KmyGYhjls4N~N(hBQ
z6j4O7Oe(aIWG8zweoyY_xOe8C-|O}LzFsr0*Ub5x&ig*=Ip;a&JkNv&2L@X7mPq<d
z3|=^W(%kX1L?Y27|IV2^drD#_iR6g2M6xjLPUNVWwWqY2-+t2l?Jkw8`m*MYZSclk
z%(VTbDx>X+eD2=gZ)*^HlzED`47}y?xvpY}j%~R9sQc4V+Rg!`+h?wM`0M14nJr1?
zf6ac}_t~wVoYGRA3fSETk5G?e&Rz0&(m5$~>dn;AK8J#<^5+guc+h<Fc>c$<o5=yU
zM~>F&`(pey$Lo$IH^WC8$g6IiIq2>f(y(d6u%0t>S=mYHVdmG=<+GSmyNA*iqx`!4
z@!-Yr%a`M7mS-L<j)@C&*V}<YXGUCEE87yeYx1Rrf}ubCG2`^C$?W|Z$(ohJJX`ON
zHd~w;?rJgqm2H(*)bFz=x}V$MEi6(k_khL>-w7ivFPX|tt{hZjc~Cxn($0HLpEp$o
zIh^@(cwBlx{H(2USG&Ge)YgaW@^aKo^}RBCZ+O~>8PiMpXGYepTd>#vMRDG>&eo%T
zsSVBYADunMHY@Ab%>KhvT*f%3Ms@nw^k$D%MfUTg2~DS#eYE^}sa=?ZmUh$md3X9h
zpE&RA&<*R`hkV#@u|59SmtK1={;=9mKXJ+LuS1`CkIfERxaytv{yk$asp6ZaXD9WZ
zbmjW<W3}DY9ahhtvHak;9y#ZRhS{(wiT<Gl@uN0JM;DcCjV%uIeU&u&_NzPbKfR5y
zQBhsBx%bysP6k8A%%6Ju%bU4QNg7v;v$ZGw_TcI%!>j9EPM?SzzUaaEnxknO3*x6A
zP0Kg^xGl2Py5f?<wh+f{t<zHmmuGxDRFpRO%Sg}N-l_A9PDIa{KUwEp?7|H56(?}p
z?yL2Y0d=~X^G5mqHhfIx($5<L(K7iibXx7zV9$m377sh*T;=jhg5#{}yw`(YU!I!%
zQ^F;^XXo#%9CAG4x@WZc%7r8Tvy^>O^`VPhxbCEJD_nPQ)efpYlUlYojfv2^^=r|5
z=1^;q?!F}linGllyq@;gx&Gx<wCYy-ueOW(tz3AB=;zHNFY5-ZIT6)#=lQkwOJ~j5
zte=#0Zt2|_uiwu&`nkTR?TVKRgUeUHZ3?|ao4?sPu6X|FTV8vss`if?yQRecmKXE$
z#*}_thfQE>?IY8RPvRSkey+ExS@s_%)5b+=shMN<c;sAPRhyHnt9PjH{A}F^yMLHE
zZAtyzm4nCp8lj*4bY<;|TQ-l@OsnoZXH@h-y^-n5Y8_7Y8`gBM*~sI>^v}Pn-;w`V
z`}3wzCqkE?8^5il9?KHUyzZM6^gG#e-<-`|Mhsk#=s2Qvi|vTr$8H&grtF><G4Xgq
z${9x8z3lE<k6G777A<)GJQ6ic={{t7vCk0?w<cqQ>3(hr-R4H!<%-sZ+68&7%<MI1
zOzHgVKEKx7%^GP~*XPjZ?ehX!3_LxK{T>t@`nu9(qVG%xE3Rqh@RuJ`EnCZcPY?E4
zb^Dsf<`o4OZpZx4>XrYl&%4IN!I6%s(-yh@rMY>UOYP>qinN4#v+4(&$bI+zc7^|y
zEW=-~4DUUqaEsp7_HzqWYRekGcrHt@Q>2BiIWTO0{;t4`$*cEPj9r}5R=Y}N>a!_|
zk9y7%eMX1B7`Azn<!MxA)VKG<n-b^y;SuxoBNj5dJ)%ymXf;?gs_D)WNnUVxuF>?h
z<BE=#Ps%;E?OeC!PfoEn2i>r3Xba7$z3hDV)9oL$EKH-y(zm`f|EZz9#VP1b(ihju
z?VsGf&U!Gw=hKP7!=Btf*|McKV|MN8s?XlH2AcH`$athN)otawu5D`B2lKn08eFNX
zyMDrs+3fTy$p%eZyKQsl%<LbORlij6ofgqJG&ntOWPX>rxG4FHPSuar9}a4%>0kBZ
zTvI>WxTcnrnaAA+C+}RkD7~uP^JnWqHcF$pY*vuLZR5Faxt5)xnx(4Y@8&;c%kpYA
zdgq6%j7%L7JM`{|CvgWiPkmRo@#uRsE@GKXrz!2z9^Y*}`%Df=YN>Jx%-Yl0$$#>S
z3CYjLO?q~9gH?>?ruvh=tY2W!bHLCx#=LM`#$$s{D~2T)jGUAd=4h^7@3W?E#^6Hg
ziDgpBiMO3<N5mY|3sVdFw94&L#_~7Cn!$N%FBO{2?wfqRbXC*it(}&6eSFg#R@!^_
z2-T2j>r>Dpzm-e(1kQTBUq@n6T#&$e_n80Ow0TbavMJ{a`X5iJ__ODQ$8WV;XFA{1
z|MfsfJu}+;=_1#hwEIzGr>;0Q^6Yuk|8iW$u$EWUMa6#9?pJKkxTeyYyS&)T<VC_-
z#fP+y(VfCBxPJ;T?A%So@9lE+y-SK!+_$ZCU{~E$b=#phOGVCJ;dbCypT%Zb_Zt&!
zpHzHt|MS5M)!kZKoZi2$o3g^{;nK=z%a(aVi^tBgKeONR-d`Wa94+79;BswU$jt?L
z5%ya@Sd|Y8&kWAG*{XYD_a7&^{OOo&6fotkul%Fox~olDr^b$-Q?fo$t8UrWX<g=8
zeoC9zXj_$VWa3{T$Gq;WiE5g0$=EO7HP6-q#orh-C2~{!X}dMCb}J%nJ{1<;>OD7z
z8bV+Hka+uucW_DE^BCWXFTv+M&Gt9iE$BHh*lLa8Pg{>H+nn@jrmR2`WxL1bb0?RD
z4VL@6n|$mMcu9X}hK`5zIsLrSkI@$oB+sqlp3fLc)B0CmX4DS4*R8danWj$3$mlsI
zE>6;H)psmmf>*~kC(e%CB)3{)@G1I2qo2+4U+kA2Fx21jLs84JVTZKd&RH~lg0`we
zB4yqkTh$sgK^4h@BwR9*CnGZ@k;upyPX3`DrT6ht1h#7+Ljq~Ts%_X<NdQ8(*T)5I
z?y^#Wei$I(ifEf6x`&bg1pl_>@eqf}1`0GOM8a*S47XEON&*m>LDi)4FReRO5lcEs
zpyUaZfsz13H#>CB{<+<jDUg4#gsVgbmB>L!zC&(3I~-EI^OG(zBdxrW8oiR5C<#FH
zHU)lZ*$}%y7nu{tYpKy|0^tcj45d}C)P}X(Qy_OTqbs!e71~)z0Al>YJT|iFuT%w^
zNM@8m8KqD*N&*m5=M{$@uUXYo3t5tmG8v6b##%`LVpe&xar}9m1sG`sO1Kl!PA8<E
zN+MCAq$G}#a#D-49xHnlXD(zV;|Rz`X5;nkKN1xhNq+9r0r-?4tDcb|SD?XUEc+<i
zeUzV)Ku7k|`fNHBaKKO%$;nufkUR;QDhWVbmLDJ3%`oGQ2GSv`n~&V`QEw#yi0kpu
zCv%D{J}OWT0=a<<Zy+Zn0Z7*&{f7JAUeXyOW75%4O7$pZs3ZV!n<|Z|Rh#siD$*v9
zds6v5sg{xe#ET=*NbfOO3gk^@w1YC)L3LM>?~vRXzB13;FB-^skc7KMD{j%6N&*mH
zpX+Z`T#F+WsQ*9-S3sE-kXCpCkiO-=pWmKW)(0aCGL{NjwSv}C5`g&Yrk`FC5^AP}
zYzSl%*4Ttilms9F+l}1aB2rrv$b$@bCuO>ma#s?7^w(<1ACz-kUlpm5R<e*h3+XAz
zcSz;F=M(enIRy$Jt?Z`Q-IR}#0A%3fJC`1=G?}Q1R7oqTNS=xel>{Jz&K|nFZ_|V2
z7-<klHd4(-c1i+}(Dhbfp0%a*sz{wcwj=p=q^~63At@Qd-X7X9Negu*kd27mh^&+Z
zAVaek#pw-Pr>lz`2&7zUS}yIPBmfC7%JDr`(>_BFbs~0jk&<1cG?WA&5z8ccbJiJ0
zX(2n(N;J}lM$SqCkP+q7^^89TA_eM6)+L+L&!+5^1R#;wj$JDs1-{ZkR-~0ZjL{y(
zLrDNK%67!Wo^Q`wQbjHV@>VK;E7evKfc#KXo^^5N_L&+;pQv{!aw{bkz!QM{7-d*+
z^NWVFDsm#N+?UGlOLdh5AQLxEOsud89i>2HIwno)T%;|O1R&Fc<34m+K5?4{(j^^T
zKyDY1pOOG%j?b$Mjd}gcRS_l~r673<*=e2tWZvM~ca7Y&6{^UHK<*&<9imb^0my=q
zwLXP2RGur)Ps9RpDbrk{Qal03qP<&tEJ+@2qCmaK>K>-d4pUu}1R%?%m?qx1UD8Pv
zbt0x#Nh>O8JtYCi@_S+HhP!+YRv<4j=~CLPly+1SfJC>2zc`%O?Wh8cC6Hp;rkETP
zJON0|s;ME<qdrR%$dy1Y(`J`RD?9;6?5<a*|N8M*y8`);rK>;&732)y2|zYlGS`j_
zw~5q3)?{7Ur5f#0XC?U#`LO7)*X<GW6sRkibOvRTK|BLb01{_$Q(aB%-FA#RkyeT+
z)nZCpNdS`Qo8Y>8fVG+e_5QYW8iy$pvUD8;AW6HTx0J7s{Y`-;5XfFiV=rZ^BmhZv
ze_?G+uN$I3-N_!Eq)bmz&PoE16iaWs;92@2jc+GOELMpnE{G=pNmHNkmrHrMg#x*e
zr7NQ?%4k<50m!~W!^#^o_up1U44Kh3B;Q8P0G<FO6EAj*zEgET1KATu8MZ9LZb|}>
zBelg-PjqWF#K?p|wo<BFiG$|}Kyu2)bp2_;A8lG_Jb~;$MmvzJk^tm*lymIWF()=C
zP+v0KtF+lwveP^PNbaMty+)mEjMPIKWTy`!*<oa(Bmg<p_~YaHkiXIts2hP?M+Vo4
zGVlZ-XYGGHSg5__fCBwUHXxZYOeT>BPXJO#wQVSLA6u-76lA#RNS;nijVIqBGjB<X
zq(gRVAX@@Cgy}<AUr7K`^!#<)tZPpvXdpc@mO|uKNSp^x08;*BxXP0Wbb$i(BRhST
z(mP9aRuX{x(f9n1P9c)f8pwcjRDm@ru&I&&r1~}XO3B)!!5@5Sj3SNE<ah#*nl-(O
zMtpj`LV+feNvBfUsYEe(0+0s1fGOdJ_O8@Lj-;b&Qqyb1d-DV!O?9*O<k#ETsv;Ev
zNkj59WTYejdA8x{;G71dJG#h}Kwe6XUXoafCjfb|w)oU_yU|1$x)4Yv*2p9QBu@bH
z(bGNFxaVP+0{N14IYpVCA{M|CfP6Ynb-$<YV5mTz1d>mg=2I?80+7!+W9Lnrl{Qd;
zMiZ4Pp*2cqBP9XIpD}w!&3w5jS%Lh>x@1zinIvfA2|y&XCrzFc;Xoqamp#nli>@9;
zEVmc+*^4;PYL<(qT;eHzaK8{lQ*veK3_%^sH6mFfG65h<LIT-N$^@bX2(6`ZWk3xd
zLO@Nr-bHKdqOIT@l`F&Q@4iVHR+AR^m>P@|f@v})K9Yu}S#og7vMD_{Q(0~e?Yf5U
z1wE+zqN-Y7-V!8bIa__(uHRAs@*lOXYg49f?N*Cfvb*LlrM9r~ED5g7UrB9sm2o#q
zB5Cv2Qd<bJS#AYlS0F9$$t)L(*jS_m$3Dv?AT|MMg(~9>wV#YUryMjnk7=$Rv-Ma3
z20?QTSib=q3Vb6rXv9VW--HdC#Qa07|4@v7gxN<zdunSo#zcCTw33O&QQC2o1?-Yq
z-hp#nm|xG*ED3&P`A7rSfF;%}D-`n=knDn(FGaFafmdsf{c2!jo=<wX%-COM`awRR
zZgAVb>#YaB&@?x~W`r$-_cRx26Di_G+K%ij$f9hbzRNVeB;>u|`Y-MnJ2HwxUa)$a
zy7T>-9!>iX5hPF{3Dh{y1IU~xmditmJY)gRmF2D=?JLL@LLa#@pMoe-x6-1^ift8t
zeQVJ;fpYuaqH@C-xiY;HG%TJT7Eg}=jbgbo%%C$&D9oPa&NAU=8BX8}nTSGW81Q-v
z<(HSdUm8q?x(nItLf(Si?M3!`QE!2dr}X0~C)hYpIimA?HYbdzj*hCMM|&wV29~?T
z1YKf+p+|$Nqq6q>C%bo|xmY?lmL3R0r@068*a!3v@FC5$(EcrSKY`!D=<i?*z$4RK
z2@_Jn41$JfZaubMk8R-irn&dn|2^&ph7V~bvGZG$;ud8B9m|!;CfK>5de5>x`_7k;
zUFc@o%?Y%^@NskhN!s3<$)DY2?#}RMNKJ`X;PX=O3NS7r;ImO!A)rXEOg@8c=C6-0
z=(m(5d6~gI=>SOC%oQuVV`b$H3=vSq_oyx~VwP*5j2py!8f~6NI|=eC+PI4D0!zno
ziHva~^PM+;Vs}Die!U-=iiN~N4mJl<A_MEldxc@WqIF);0YiVxI=FB2lNpczlZo-!
zFU;7oVD`_?miJu4&}6)N_b5Xc4{;}yUW1rVqm9yNQ$b!u>s8T)FkYIBS1*zI&RhBA
z=D+c}A^+EP1`>Iq<$M+odq;Kv$r_L`m;+1lW!VnO7*c8HiS53oXh|_kb4hey5<Nhe
zYBC*|Eatb<f!oD=3LTgt=I=5AcbP!YCYq~e0;<J)4HH<y3=rhCOn9we@HF>=v3$W;
z3-T9?_Y1~X%!j`a<6knCF9kf*DadB?MGEj$P@N#2BNr)hk)`0|N|1I5>I`X@Tv@Y#
z520p3w2UuU2y0qQ>lD*gkh*kh&Nys5u4JZ^=0e;<JU~0U4|uo^8;<Nj9Qs|x@-AZ~
zOr@H!tQPY%j8%;=6`HGMENcaM_Z_o7oy=I3OB}*UsoF_tS79LcrIz=l0g!%zJrXV8
z>n8%Rk~XZQEg@xNxjVGR9ohi&*uB?9`Q1fPI>gS}Fx`gb;EOz-^{Ad!_G$<bSkYFx
zV1Qn>6C~r+f0)-XW33Qvn6a0ClK=6~mg@gl2+%dMHGItm>=3#q0=`lMbr&DA*H?bQ
z<6oQ6+-=O>7S3myyMx&~V*U?um}6t%ysyHBRbu`wHoPn5tFd9Vn6JUwHCR{pz7}iO
ziurrk@SYfdADi9B76SeOHhUoE>#$B8))VCQ*rOhM3VZ`*8^q&pz<v$5uOM&4Y@=A-
zi2F2Re?i`a*(R~P3HNQn{RH_#%sv#$A7a0U;`Se5-$%HYfN#e7&Ek0IG1h-9=3B6F
z3pN$-Pq6V5G2e>2w&L!B{3-7GRLnoa&d;!`Ab)|`7sB|U7DqIIufT!PfhN#$W!+6s
z<xFgsiCrLpV7aT<;%djA_#Cx^x4(+J0uOaM(gR<lgNFQj7O*!4nI0#qy^Av6MGfh%
zEVKoL8~?uey)1x)dZ$s3(<l&J3(K8HJ<g**$mv+F1bLRAAaEcZQB#jAC=jA1i0Dbz
zeDM$TKA>z{EBD#D2QlakjQs}22MmEkZj61Lz(aYE7&KoSgu$~UKapipI*_`M?1OfS
zr@g>?!(2$$e8mvD4m`8^#k6OWcaX_7Q@YKRF?_+2VwBoNS{v$A<S3_g@6pD>(DE3~
zJVqa!SdjOYA#YwtI--oXQoXm*!B9zqVk8+eUx^gPd;~cjK|X^2tw6F0qz|>5kaLgy
zVy<O$_?kyF`-nD$uE8M_UEm9sLW_L<4*-^&^Rj)E4IKO|w~_YTNJqdN2Zwp;m-bH{
zN_wb6)^(^eoQhETBxC05o<a{zNY;dO0Z77BN_Ljgg`6EiZ1N#r4HZ62Mh?kn0HlFr
zl_>YOR5%PJ%=u6(msd8H3}rdeUyfWL6&kuh?MKCp-`;nu0#<*7oxyaVuu4K}zN#v8
zbwjFkLuxAKJKvCY6Zj-LIEm(<YnDr<gOddw{!Vm-uf;-q8KKhu?DR<*o5)Zquu271
zhoL~4MrO^IW<^~gha?mP;Rwn5D7V+t5a@yn)v{}e@5yvwN#)JEUFrw3X1PD5-hYbu
zXyhG@{Gcb6TaCO|i}{UY$&fp&Kg(@GE}O)BE~S=BX+x!xoSc-}NlF{4YAh)vsGXv;
zp~gV6I!Y~{(uRTpOWcuK0i_N8iC71vcTot%iRn{%C1U;(rFTipU#9dfi}_MYuk`P{
zRjD{IE~E6y#PTxAs!S}uLg`%*%Zq5WB3c`20xT(+suk1P5R#Dfz*d`Z7w~>8w;5Y)
z7V~k~Do)JDW2<;EzXe-u`8)5j1-lFFC19%ru{;60B#7mS*eX#hKZH#WVe3L=#f;?+
zW7ET8{s=ZbBIa|jX^xmbicOFHowq-VU4{0JVbfz``7vyNOe{Z+O^*xm;p4ZC%-&V&
zLsa`U)_yG<*iZo>#=u`Gfa8qxIqkDKxOM{h;<+^9xpXW9U2w@jzTj^dz!)OGBzJ2*
zy~_Q+PUyj-YRQ#163{ZseU++zl^O`vvmCL@k*>hUAoUnzAXv{@#I8lU0w0Ie<B)-1
zJ@JT*N4f&P9XV`AZr~PKa>d|~g52QtfhC^ZAr-lSt7l35&mj%DL1~aBbyA0PQtMQf
zP&<k#vN%dp@S^W1=XX>Oa0@K=k#hY=eLuZe?la~3S<JUnUE3*d0sj}}`j;Rd6?ir=
zET~Tr*_$-vk%q!xAfsuohTfYaClNtB(vL^3FvHP3Q?++9Ei?WRsg616J#BPn8*7rt
z7E`9hls&jY(lBF?#@IkkPYPDpIu^S^4o?bbSSK4>2>RHDt=g~)#L*BE5GCPnj^M}|
zSLW~IVEjuLGQzJ?qpwm=*ux3kgHLOZFFZvAc~ZALsSj-agr&|ZHHi;3iNKP!G=PvC
z)R7#L{QVQ)NmN3aUJ<*JL~NfZHr3<U=D1)g5CC*yxhjgSq6~ypuckDssqe`!%T-fu
z)s&wguc0(+#PS--y@u)|$eSs<X37nCa@>5o1%ox7VqL6bp{Cf+vfN<#U_D4GST0m2
z)I#9G<Y6M&5S<|+*--gVk!&b8be!;Aggind<8(ODcloGuKI$&Sd4;HRA?hx~c^6RU
z3#hvg=ar()rKme>>a^nh_NnRbM*Slj0yx=o{w@#pd}f&c4-2k1rIQ^@M;_@Y6zt^J
z>GMCFJ3ObE2vU(pDhdMwn?*~uXLkK^<C`+r<OydxSR^@#`CC9?*AFAd!{T7;I+9&S
z`fxWgC+p^!Eo;Xu`L@eWRa95lm5yDtucn;Ed`LAFCh#?sLk;C3<_FaXM?49wDC;`P
zPLS78esxqo;K2vT<jPA)VZ;ekuLLR>N{Jm|x#lj)6hc{+TS@Dzq%EK!maC`D>uE=b
z5=qUO?*ED&3e}&E`l|6|#ts5jlG-vFZHx)<^Pg6!hK<kcOV)G`(%yr*Km`$E2C{Gb
zohalaa3@Lx{Dml-4T~e({O;s<dyuB;v3fmrfIw-9QA>`c+}xfBzS25h>F<I5(%kXa
zf8LWdiU{r?!#gMdmIX3&vZ4HqDkMUr7)Tr3qAi5!?qD2uFmB+HmM``^v*AcpM_1dC
z$96ObocHphEf<{H+B+Oc3i3!np`bHxIZLvTN=i{Fqyof8eY>dz!(2IPb*QC#Q#NTK
z3mIl12S`g+7KA&lyLWOa5m>oe`GYHhAcQoWLu=%SFDm$pU1%6Cb_wtXrMMv;0e`&<
zKu8ft4d*yg92Xa=&m-;g$QDeMWQ*VKe!<k%52T`>Z~p3K%W|EyIvWb!pqsp#roi=}
zd#DPWC+#WXyl5{mr|)GZAp2|gHxjsDd9bL>f%1VOu3V~HE;WG|o+Wo`y4R#8aNX7s
z4)<x727&YM2#5RpDdwY5pJ)^$9N?=_pVeZ1Bl6gYdcyJ65e|E767%`UFJGL=79zhw
zaVC2K^|>I<WJ{4>sgTK%%PvZ`mr{p&YZAUuvcr@*Ty>Mu8l{s<8H4F`gxfkNDPwWC
zt#gVp7Kht9`INCZ+}0_ejD>J}HLbT=sJpUU46PR<=GW4CYsLIJT5p}0FQH{6v^re=
zkxZV^I?Wh@?RVThXq^%BXBn-tV!n{kDirwjtpR`hJb8Be|6Iz04JPKzzd#VY-!|m1
z4F$n%Jh@S$++GT9b))iW0`Wckg9PwBo2wUjsY!C5ci5nwl{45NSPpTs{38XSqrWG4
z#Q16JO4<<|52-9L4w+1MVX4+*&-K_(h;&l1dMefzGJyhYQGmM$k<WW<^&UGw<P(3h
z<<*MQ|2h1R!*Pph6+3*wvoJF3C$!BI+8g{DsMI$n+wo=sCYpFnb8U?C<AdajgN(*O
z#sWsTwZY+hpTgfe?(8bDK_#{U@4NNGxG^i$P419d$1U3Vme5BBA99Nx0ezBlm@!Bf
zM|bH=k95W#eCxKd0sC*QS?WamQ~}m0z~;hauVS66*c^&YNs5ANH>%w_GUga$6ob6L
zV8L+6toYX+Fe|VSBH-V6z+nUHO$7Xt4%k3=(m@7ulTzFi;w`!IVFzq|TDoK7;^KLp
z|8@Scd+GYk<qFUK|Jb_U`_-EI7k^Iue^+Dg7n3Jl-u-4rR-=tEYhxT?$3V%674Q#D
zU~c66M6!DEZAAtp%b<)w`(d5QqVTU!;KOWJS*?Zc-YaCQ8?aG>P;e!;?bx^xdx1`{
zToX2K67vtS(L({Bz3?<MWcuzvGPyRXtWBx`Nf^t$lgi!+{2@1=AtmYKk_fPZHmRWP
zgoC;a>y!yPaYTEjG+^ZExnvdgBkldDE5sR~0;HhBKa_#Vu;e~pr--%^lKU1qq=g<S
zsOxLG|7$uzh;-iLkoR~bm>hIKE-v_2HbTQ$$UO@M!}$%*ZO9yMQ;OT-^Ot{g1F8fi
zJF@P7uejws{Vh*_-J7g`{le=+k7LRIseCqo=w>C>t`rjZQ<Lk<1LRj4h$iGC&3t4G
z9<k%1Ub9flUqG4{#C$2zEERZ&lu68S3@MI@V~*2E`!uqJbV06sTLdx3nbJEW5Ba`*
z^slc8JZd7l#6M*M4TJoG*c<<VNl=KR$nmH+OQ=M$N~8}D1HjLb7Wvmr@FB!oM8H30
zg2@-#E$!x+J8<tme8z<XLuRN2%A)^YpK<wS<<+aLZRh`Mr*-*B%JTG<9+g@&_XlSG
z5Uv_%t_rhNV*V~>?~3_q%vOu}8qC&+dGgMsR?OeS>^(7mAG7zx`~%EB5c74ItrPRj
zm~9sGk1_jL%(q~+Ma)0J>=QBHirH2%{}i)N#r!kOJ`?lLG5cKLp@dDgkbiRqCr?Gs
z&WpptOt$^&`g>!<RD+N6)6I#ebfsMtV7QeReA^oPOgAHcZl*PwX>0I<mCeO73;b?B
zCc4r>4QinVLuyEKPpCmp#C$6i+A8F*H20JWd`blY{&tP`zRzop1dtxqBK@_<6^?B<
zHOLzAkM2PK@3y6d|D3!o=D$|cYQy-O={f49q{~*idn?@&CIvAE>5_lF2j>dB-XjA3
z%^p}cBoV~A`ImYyNRkQC!P|vGKS==T;1q#x@T%zdA=^OuuWvMz=g5fp_k+TSbC6>W
z@`dw;<b_Cf0~x>_{zHSkPI1lZF2s$r(>m?K1vW{TXq`U={!v$^t&_U`dIHSGD%n^Y
zJQ+&@2bD6c4UxoSr~E^1rU!SEefUJ{f1;gXAK)n)Ii2~Ji^3RV=#UutyRxz5IZ)IC
z@jIf^%%IcamrW&15c%J4AozicOR|srdq-&aO<Mjr*Bncme|<aNYCL(iDmJX6Q2oC>
zRCvGS(o(%IMe6^1Lu-Oh*<3mHuHzIDy{Lfnkt`km4isD#<dH<cKLP~-B!Hx#`;cNE
zQhH$W<U)ydT}_Kh%Ij$Rb@cb*BzY=FYb4M{fR`)Zj6y{8<;%WvuJ^XJ5xuEHvN~i5
z1)M*<9NP*P*-}KX3;FIsK@g3>Rc*(<e|usI_DkNDP8%OUydAvrkbp)6WUDsnyj(Uy
z3Qhs)(-Ig4M4=M!5_4w%IA?!{=Ntm=Qz8i$ehwFYh6+DN3O^%-pAo{(5yH=5-+v-*
qy-VT6!6iEeszr?&Ib`VY@Q7g}B16L@Vc#K=QNqtC;pb4{=l=mRNEfjH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_splitk.hpp.6F08C0BD625D58A2.idx b/.cache/clangd/index/device_grouped_gemm_splitk.hpp.6F08C0BD625D58A2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..77f39df6833ab46e3a46bfc76a5d5fcfb5801824
GIT binary patch
literal 670
zcmWIYbaR`=#K7R3;#rZKT9U}Zz`!5~#Kk2=nc<8K3@Si@iaC2H#AhEe5IOogu%;k?
zm-XHM3Ww%AYHqgO@qN<L^4yGD(vw08H@)BI_k32Fs61P<!wJjp^W-fLuGlSWu>InU
z&pg)r{x#06zh9K@x}B)%Z}W50q~+Hxuh*)5cJg}Mk15vwg7%di>lKeT&Z~7eyYsVv
z>EnCNXGGFw?B2TN$C`@;>YDd?dG%}_ebijvd;HqULW8VtS6`c~{xrqed!EVfrGH8u
zyKl`1Rj!NazILlwzV1-|-`^H99(i%ik=8PCD?YPnY1k#*Ce>X>TdXJQJhf(1N<EX9
zce17AMyk%Ow?<`$?3y&b-&^9u#=x1%(fL$xdlBobYZ*@SinngO)U?27^3(NPhedq*
zxKi{&1Qk0(W`%T=GFE^1E_$rGKw|on&n~b3FK_$y`fw(%->pVZ1+hi%JEpi^m*+Vs
zV5zZYWkUL4&&mZ~R3g(Sq}O*uEG$TA@-O(_EX`0{nVaMQ3}>OnyAQssV0a|J#KH#T
z0s)7jprRZD2ZWQ%!kc?9DanD2iJ6swLzG=q4oom`G6Ca;fsuo;C^fCP7|4W~!^jEN
z57QUB-7+UeO+}fJheJ_bQ5&WTCe6qNHqKL}tb5`v=elk$=9ahp|1ChOO%$zRs#mPz
zW&OY7(Hx+#sD`K|Ofy4KYEGgO&~CUudXHMQ`kHssU>e{u#YM@*`XC>%g9r%Vf-xAF
XI3f}btvt>caFdyni=BZJF2Mi*c-j9I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl.hpp.46AD692E57980C90.idx b/.cache/clangd/index/device_grouped_gemm_xdl.hpp.46AD692E57980C90.idx
new file mode 100755
index 0000000000000000000000000000000000000000..721a30ab4eca4c3f0d5cff855c8aba1824ded8a2
GIT binary patch
literal 14358
zcmb_ic~}(3)86ixA<Hr>%eAoFr*eaUfPfdE;-I1sZ%xEIn)p*RYP@1p@B{-YLIA-7
zBxn#7<PZcwR19iR0S}@A#skraqTmto^}1Wmd{6$!KgnaPeeHL8x~r?JtEwkr<nZBA
zKanVS_Q;jr%vlyii$o$_^8eCh(O)$%BGEgANVGDoeBzx&cl>039DclCt3YdoVZg{;
zCc{Mce$!6#8#<?8t?m4+$C<DLLn2BF9#^=#Cyu(7_$>RDDAqN-WU-%DTJXm3|E*j8
z{)6KM1$W%HI4Jzk{R;oPsTt+kf!_Dqv;*gAtUGbGug01MewMcJ4_ilT4l;i3dNw6W
zH#X_<8e0?QNb7UAvq>KgDpKvASSVU<=`DS~i$3hvrqi$5cR==*3wGD1%=NilxZ}Z&
z^ShRp@9LeGmOV7S+StBR$G+1k^hny9>TfG7=jpE7)SPPg#I<Hc@@8k(-WgjrYic$P
z{cy1PT8e$ynVQcX=eLwRyD48+uqiU^X26L}8^4-ZJw0Hz<oP#D=h`({!!As<zHvbR
zPW$TJ;wQe#j;&~1dZKk(T((J0<s*YR9n$M>x6L$&OWd$z)Um~<9yn_p9`V9tLd!n$
z+5VY$daJ2PRPM7YcRH}iw%(c@ic8B^6?Ij#{;2OCAG!8g{PKmLS#vVO_f2zeXP>T~
zTUR_Q@>y?j)0eWu-R~yqrUpiD%PXv#-`+BGWO~D*aH;3-TzltNsndfacBUW9@B8z-
z(}kwHR-LJyJ0fCm`yuHshpu!L1m^xeeplV&Z#yG5Oer~?6}WKX^T}Qpw>3C66zWy2
z8dN#??JV>6yV@?d#}8i7W?HQ^L;u%#CL^MymlTgqY(IBCImWHLAiQl?aBSz3sFOV#
z-L6$^nSRhOwS9_i;@_n!olE9>#~e(Z>|t5=aIuHQ_bD`D7kzec>x%H!i&`Vk7k76@
ze2s?B)M1?t&31PE<wVh;Ll<MpF0QPdFy(G>xX#d{Q4TX59Y;O&U36>e*H2c2{gw3l
z>iu7JO4mgH51C&r7a51v{H5W2D`>%a-|78lXk0bT3F~ZZ*S>$KcIovim*>B^I5oPd
zbNP!E7Zye>nmTsrDeb1i^-Hd^;i*w)FJGDIHuh!b8=I+_{q~xD6ZwL@JFxDe-R22e
zsq(aG1$tE$1Lu`KySB*ghxn#*KhJ$-UYEUDQFv&9(MJF2(KB-29DnHcZ2Hg9$tT*(
z&n4Vh5Z5t3qkVAeuvs<1isNk~BQ&cF<G0b*FW-sjw{monchxWAm;E13@lt$Iw$?&t
zdD#TX(UAAIj$LcMx_Zw;>mA)vc6RBUFIt9-w(IAuJ!+JLN9M0Vxf`RjbsDxBwB4vY
zr^z^tFRu-hkKPvjmw&BIP6m#5*gRuXVt%OXLU7f9k?wOQRL+m7em!xh?YO#($}NY!
zJG<}-HEWIK=AM&vrt!0Fht%J(U$tvvzdLl<v4yU4{HR^`Jklp6dB#WWNN#`Un5`HS
zl1taksY8dCWZylbTPlq<m^pH3Wyhi!9T!Lb{)cXH$A(`mitC=&nnW(H-C%Q7^AEpV
zt?0PM;ylB+rQbEqt}$5TxW~Q7?#emci8p`im$oN1^}`K|d7Q^rXOD>E9xX5*`o3+z
z732N)!xGT=dkM`4+)8iV8(ZdDgWK;9DHycj=bq=c7fYJ-Lf_hsa$L23;amIlX^~=i
zQc+CQDcN1m&fg1LMy;Ca@V{H<?;PmLUT|+?^7h41+cy20|GiGoLfrxPOKxYZbiF#U
z!t&=c9o^kqOjF;LOc)=NvG87V-PJdZjZIg-o%Cjp$Ij#i9mO))^rFvKHil1r8%afc
zIo|km{EI)|p1plJ{pRF*-G3R5JYF%=A;{%S_tKZ^9YYk{zDac-#<Jd3zXcUsn_}Oz
z^Q4EPO~>YiQMG<9AsJiiA3Y7JOiTQbF@9e~!D`o7O{Wu!Uk7RJZrE|D-m3jc(55r3
z5sPE=$7SuEetgmR?^BHCB;-zU{kABm_WiN^+|n;Ro_^WneAMLdch?G{dW`Ll<*(UB
zXAG(`8&-d0_o?0X?N>kiU9!(?KzrC1l4ZlM+q5ikn-=0(;xYSOdvDOV&vy-R$g<Pl
z%#<kX2XsHV^kr<zYB#qk_Eq=q`@%I_)}4KPzUj;I{XIQF^6k2|y&JN=zrSIRMf<Ny
z75*X1G=DIu8`YfgVAk5zv6(lPnCeY>hHv^`*>lr-kMHsqmJ45Rb`~xD?wgs3nj(=#
z_y?VeWeGd9u;u^}*NP;q$Ve?jVsZ+Q|1^H_-+6XfPPHL+Ae7z6d^hT=7N00DN)DA<
z&pob%bqM7lk~~BTwE&9Nc|XO^R`Kyt>>eQEu42ViY^fGN(bjz7onWQ)ofbAEy*xmY
z2V^WN0TjiSamRuay#AJAUw;wzg0gx+xvB+FOpctd>K^^BL5@uVMO+>h=V5EL0E+4S
z2%D7817tF6OkPySXw@;MYVnEESbX5G31%bousxwvAz2lfv`PTQs<2P(^P72nW!REX
zVj0a?#!M}MVm<bQ%;fdoOZ0FbLP=(1$>c>U0Tk!?6CMtpI<D0an-WSAGD$*aY5^2i
zx=+xXS2u#Ru!PJg3rn-GnOXp)e{Wn)ne2<#QrwSdw;7u?V^6gJijP_D&Uv1*&PlNw
zQQa%b^%XTxEr8<RlX>;5(-;>i?oWDY#Wt<jTP=Vx(Bqfy+};_FH^k=386lHoGF+AT
zL>aq*j?PPsG{oZx<tJqF6H%Q?0A&b17rbGHQ=}9}kg>d>WN#=FwE#-cB%|}!e%Sk~
z6bF&9+{51Y@JzJ;N?1p{-uwE288U1`UX;erX^fFt03~9^XiFVO6sU#8gpz_KDcD>s
zfD*aYBwVNT#u6!JiFRwTK`k*%l>o|wCDXh{MmhJDVh=LhKPihpDQC3+%B0N$v~;ag
z<Qmw5yyzXJ^^Ve03!qFs!ft7>^jbi)OI~yl8(qYXY5|ll<4aQ<g5+3+`;uOA8S7le
zOD%vhV{R#RXQk$48Fm~f;%?E>TePlP0A>D~Sjjn5J3xwEi0bZOvpd8XR01f=e5fv4
zaljlYb|sVsEN{Sl)B-50^P2BXd0Xfv$Dff&Kf&TBSf&<0*%)XvJf%KAMGNbZ;ih3p
z8un2OplsfMam)a@!!0e$5K0o3B#{MB37~APKR)r=LEZi`tWES%h@A`ZV6_0s_W0vv
zUp=#)r-$uGFE^0v269jfpzK(smlpOozMB<&?kD1G#5R^6{}O4^STxZGaVXB8ibi?I
zrs-iT(nmUyr4zGM3Fss4cASgf?uKuq*qdnOH>~)Ln1V_GCE?(3)8%76Ymj0uqLnVn
zwTtW`l>kaoTuW$N^3OWNphz!m*r1K{q7t7dmE77DOI|t2un`$cJ|oL#%+&%Yd*lCP
z8FuFoEo?w28Ca5m`>O>|l8;dVvu6ioNbw-@q9$zAgdNlZC~4<+d-^n(5>psLDD~L2
zp6qCq080AA)nEOTVZT9!orwC&v2{82Pz#{!3x4Rd=I5tdweV-e<~~r;50qRjfRb_W
zp?HhQj!9BHl)UIBmfs`>r4m3naBkY5l_>|Y$}#6)aSn0JDgl(k159R39OHLSik*pe
zpHmjk$wpHNpcM35b2y=3={=SG_fpbcN|#d!pcECKUHe1Kf_G9pj`Z@9a(PJ&RSTe$
z8oFd^$7<+F@dQHYpky7Comv3pbkgFsx5pI2^**}y`?TynZOf^|CrZnh&412bq%XyT
z$)xXM_AU-k3!t3w)tL9=O!R{WHX+M+5NjR8c4`5X%BKnaQ@-%1l;U90%QdWj4J*_F
zD3_D_jC9&QHbe{SlU~}8qzxIX1yHUGLXtlRmCVt?3PO2;Bu|iBEr4>>>(Z;E1|>fl
zVlzV7i%j+s->nirxi<0pJeLVQVH#LYD0Nt?j;xVN0HtwV+8@vQj3+DYNru~q<&9*e
zRRSnYLpCRzxiIa#6h{*6KBZ(&DKoVIN^{xm-<PDGs*zzSvASw3ug11&@riP0<(Q*|
z%ziy=OU7~x$&Qf?s}ex@W2&3M63Ikk4J;#@J{ElzOJ<}JB25aB;Dv1bN2i;5_pV6E
z(ZU*}pB5x(Az_6|KtJup>laVWI&eh`>k>*Lk|ZKMwE)WV+Gsn6lSA`0@Mnbb2x~pU
z`f34`p7R@*K1lz*Ns6_IX4<fI8+K9)p!9mLnN_sLCq@rjlU}wnvh5^hPzj(!&zZY4
zuCGXhUA6Q&6Uxsamir02{e&anEQ64ZlBfeU&;rY)Vm1{kfXI?%5jSE(STP9Xv?S{A
zjeqb43ci6xgPUc!1{B;N#EmGpQHbv%{kzDV=Rd{$pW?xAWUyQ^Wsyud@c0mAaY%^k
zD4ROUg>R2zhQ={rpdFUm#{}+UhQk4`p$;#xL>*M}1MfiM4rBzzuenf5>wewi5kzM;
z&Nd+*!&mYb#GXi^jyS>0A>#Z@qK-59j`w4BKURPjX1Q!UDjQFRP>khr@yJ{}0el)u
zB20%o+#iBZme{gWHTDHFXNfNq-^7OCCbd1b{Ip<CzdIT%iS;a#kUfNfEZLBjdyzft
zRia8{nT+fq?qa#Sh`o#Sc-6n7hrOf2z+180dwSS=9&67t@%`$EStX*m8iuW5q)F<d
zN4qem!hu;_PO;>y6&GR|D2pY|UtBK4)mU83V;$ogK_1udE+joZqDMWVN5iXS(V<PN
zofIpG@)IbZ1nQGH7**Nj{}H@mFlp$;Hoe#p*4?=2xcI~2B2R|qvgoKRdL}q`n#-pn
z^Xbod+)Nua)5Z||(A*Zrb_-(%Avn!lVS=tO!4TLhbOLLKWLMUZH@cX+xIk<G*#l;|
zP8`~a$HPD*>Kp<#x#H2oVRQ5c>?F^-JGqC$^RT+aj#POCcz1X&X;5VoU`M9vG=h<+
zlLu&Exn^YDjQsd@e1Pm8pa8H)mOD&&AEv@!ge-TIvN%fh2W%<*VZc^#T?<*h_t^A3
z7VNeCwB>%<1yn~YmA1^~vE`G#r!<$}S1yKy$U*|E4w@u0P-QYOP)K1&gDP?P)KD5d
zzhix1EJc$M$lqfH42|aY(<b|Aa~J_lMqrZ5W1GPRB`^Jcll<qa4tt0=MpYgKLuI*Q
z>{yJQA<|@txlPZazl0IOt|tCPm0iKQvLrqiU&L|<-CbvY2r$QMt1!*|MNj^Vo(i^4
zbG`KBULL#2>y`{JJvsCvvCm(0Um>wiFOS{AXOHkavSK?S=Fo$4=%FwsntM$5c}zRN
zgPx%ewrMCP`+V%t=P%k98kDd0>E*Haz;lwD-%T+f3tY$0b&Ldzvww%rt*>iej3o_L
zw3Q*OJfx0ffmL}UI7dhr$!^&~$+l2du)q>^+6cSF&wi#TO6SXOiAN~LuEp35^z3iZ
zd1;|WRHrV@6*6og?|f+P6vLhp;vz;-#F&C}A{{Y`Vj(VJ6eU7j$|y>O_%x$`nla>`
zFJtt}gt(kBDrXdY`x(aQj1X5arWK4i-(Ja>RtoW1#^WsG$+uTA9#ukI&A3)G?tJ?>
z#`T;KpJ&+f{CpwNCOTJT*`V`4zw)2&)vkU^)+mPZh@m_|=ivW|iK?<~-nO&p>Djz3
z4^(I0v_zeA^TeY_=O{9RJc#AaAe}SFtXUmqv0OdUsYhnuvLHuiSnd^OUtu{6h2=^p
z=Mvr(v0N$TTq?w6l&p+00lq|?*@L+Z{mwAl<@DZCVpsLpvL3rYl0wcxs&5kI4G%&I
zfM`jT|AUr7B1{jhc_R)XS}Mk}Vr&Hl7gC$l{`oswAJX*|+VKkQ54(xwuF;Oycs#r;
z=K0qhb2A8W17)~@GKF^%OdD>XP2s^2*V6)0rwm)lvfKgWeE<#O)prni9~9!l$osGm
z=OFJKAud4P1!xej?n2~UD8wbmyF`dfk$0&OUqId$&>&bI;+l~6MIo+3-gQEJ8F^pk
zF_%)Y^UB5!Z8GCml))=X0o$7;>8n8(rGU&ftm4t&&6oEn-9#SE=J9?TqA{`@s#*(J
z*646wql?3m$B`$xkZl+01K}71qQq!ZK9*kK33B>jaWggr-xH}hto)lfvUt)%KBJk>
z=)gvT;tb7lM=*N?8}R|xam*gaMsOr6Ey2AQ`-3G&)FmA7<e!hx>-#t%kkMmF4MY9}
zoAC?Yj^*t_{1nTd3ULRPcL?z_EPp1%FDdPpls+V#O7&{LqV$F8)$XG7LA@*q_ieT_
z4t%V;gR$8m#Bq#GoDjz|Ht|Bdld;(;#A%Fq8e<Eq!*c12dAbnqW6bvnaRy_aA;g)C
zc_xo1MD3Y;bpIWHvJ|z9ek~)1rGO$BF`!QD+==}_gA#Qq4EFaV{R{nO<QdzO*!BQ2
zI)J?REpr$d9TwsOWK@8>_$^U_j7o&~0y4UQym*tZLq>HxhMF63b*h3JZ(hkv-(<!O
zViyuTF;+*JzEECMV#lM52jEHXllwh9z0vz0AuXKyWGcCs&BYeH`Q>9aA6xL|SA_M8
zusxU`%T;3iN^B3s6_%^P`Zd@d@&Spu@CW8Ob*|gEDW{Lgh}IJ+pF}=(o?bNI>uG)L
zY>5|sgv}peN6<1$ihjnoXgdgrSrWI|#xm{@$dP1_F-T>sc)O@$Z0Z;{hz4NpWH+no
zjG#Q|kTj^Oi=gJ2>s&Q&?|z_5#%M`f8bkarJ7H_=D#!J;{~GE*S(0?DDn^1k|9aQV
zz9B8#3Npvn*zz@Y2Fqf(ZfxI;`|-F3+xH0ZTWtSUh#%3iN3<D)MqgJQc1%rwJ@p@f
z$zs1hmqcD)W<;Xr+sOSk8nkLiYC~w!^<b7uq57v#Avx|R$9vt}v5G{BlvfrN1UQ<G
zs?3P`p_u4k6XmgqKcAQ8UMox7GGpz>IfT5%q5PP;aY(li_u!BoA%2TP-U{&s%5wwd
z1J6UHmS{~?rUk7n4-2Vtp5&wS)19<+C+!8ZURmApU=>||nHcR2<b4AT;B$-y<lP{|
zjmW!Eh@WDcr?@Y?n&sjcw>ZX&cUAis$9;?&nAEE0zoZ*z);}N~VGCuwg>vN|+)7z*
z<?-tE+0z2@CdZJW=1?v<{N&&PGC5U&8LWFvq}Skv`~m&R6E%!Z4dVn}XsyZPd~1oN
zqq2cEsNqlaAJ!IcI(xsR<fA(JchZ5Njt!*|&fA)c){zD)cdH;UTc}-=WI<K92K#5Z
z-?8|2?0s5Y+#`OFGN`47z*sjO{C?HwoSF+{)|p6=iEMcHmW3>`P+!1M`X;BJs^HDP
z;SLtx!KSbwz-bUIsH)%K6}EHKre{0<_VZvl+dj6gkcN_^R^p<|qkeQhEgpH&o&w}W
zdkK+|mj%xpqCdohN4t^XZe#{EBH})fWh&|e;iM8fq#}1A?w5-Ec$|(L(vdq%o*YBS
zCmjtG;*fL{#^Yni;TUq~eg84!a|{g>;*et~jK@{Tp$fV4XF(P6sX_yVIHU@N@%RSv
zzk!AdZQs8^h#QfAqY&reQF%C$cj&d4t;PBffUu;H$F^a8DDEp=mB(xB3s#|YRX*K#
zAdh>nPmd74#XfI^coQYxM49va7)l-^#Ick-R)`ZQd4dopQu0J0-bKlG32`AMFBIYu
zN?szwrIfr>h@Vp8r<4wqZdp<d7k5xPP$*Z1H72hpYd)-bMcKZhoOs+tnRHRse3;Wk
z*>+J*Jbpuo-%vV`RgmM1wtmE?w=CB~%X(-d*tM{3Wb>%5d-%=MhP>KP7?i}fw}&*&
zpBw$*U-M$Hbkd{pc4(yC^6d$w!RtuiNZS?D0fEe2<1+&s2km%Sa`w+)%Rl}#AqPt)
z4n%c}#E-BA*|nems5`S<2IZSUje(_xTPCvIRd-9^(c>PUSQ0t(sj}+y|Ie|Vx4!e!
z78N{Gx-hwwE4VPwE%8KKDcM$Gg08xe0&kjd?eU8rPOr}U$2pj&)x=Ji*%3}UK0$k)
zphrO3llV>f-*f7x4<rq#NSulcV4v^aKQwX1`)Fm|=@~YEhV5ZiEVq+!+{yIg4~tC3
zE|YPA!(#Vk-7TA49~~h*w$idz+6q>BPm5FKkTZwcm1AIxPcU}i>i2Zdn7%>BtbxS-
zy_9P&H2?;u#9_VEL>}L!jc*G>q1&|2Z8``d8Q8N#;i_vmP&mAaG^p<2U<s3ILl0k%
zUFSkHo5dJpF_th3W#njZl(B^93F^bds8m;U&;v_SR`EUT1vxC-)sa`JF6v+mY3a_b
zYpTBW{YUHxSuuG+bzuijKuk#*RJV0-T%^z06uTz!buH0kCU(ojvtV@T3u?cqn7F}`
zG^At8bUYpoKe*T<HmthUgIG_gV&h)QidS(n?c7Xzfl8r{O~#<Q<b%N-Shsy=oz!>O
zKTHiGXYz#V&JRX?DEN+fP}Mtfb@4B{hRZ=h?8MSe{3&ObsILgY+;j4@nZJu$)5vsV
zQJ+}k3yboLm;czB^cj0e5AP_mcT`_~y19%&F2AfN+PGrR^9Q|1582o=8;3!K5cEN2
zp}KQ~9txkXDm2>mGWS0(YM$~rk-WOTBa|rg1fzd~&&MF)Bs;Mev%T1w--(TsK_m64
z$}UmgeS&(5=iaXikz8saS}4T2h1eAQiL&yqTh3$f{p1bnDRw>YUr1m`8N~7TITH2N
zDZJryptbIUEpOt;gUOUEnJ)#g+#yPKNQmnwgF4Crh7LY~s8@B<3M&V}7->*lvx2>q
zeW@A#;C$psvQkfw;t2|X+C$mJhRMGKy!zumUyw825wn?L2cDA2Wg@>!Gz1O;;=hp3
z31P*i2Km(R6&Nsh4T<^!7$!%e;*Xc%Ku}&LDZg6caVr+LVr!@xvfL@k;1nNTlVp=J
zsN``4x7Fs^<;7LRah^crPdFb$6`_bCG?B+OXhaPf#RpB7Q1~VEIbgV%CR$frOM}*{
z`mXEeTfk-gW0zOwj9sKXT)gRDH*<AK*}FMMPBkm{DCH5$CsOC$%pHF$x-yeA#2~vE
zG#IpW;adH#zqWTNZ?gXXz0-v!Kdn!H>{BnLxgv%w;;XF6Dl%Iv#3c+{BE+Q(TPnn-
z8TPagmoaRa5SKG-xe%XW*fT<0!LSuVT*<JNLVT8C&kAuB!&V7#HN#d5@i~S)$72Y2
zNWDY#2>_^u+?Qf@A2tEE&XO}m+=S(D=#WATB|E@xGuR_~B<=e6p#a~}UTn1&yFluq
zyn43!3A;ccPkDamrekydMVZ(t6T84cqjZ;c+1LZzrLt^dor_(;`IGV<w$8_{e3V^}
zb?dP?oOQ|yup)^P1Qq#|LB7!OoTLm+3UL*sUqxAh{~%>w%Ipg7m{{&QWqh5o<J(hd
zgH&2@tmV-9IkY9Xb!9LZSWJgPFbKD><bGK7odmC_gDCVM;`qIvg9hiIp*$``gA376
zey5e9&{D+l_#z6uh&b43EO!|VzKn(fz8ctPO+>g^(tqx~ew(<!xMz8~G8Ay4ouuIU
z>#GC0T8DgNLH;?9w#cLVfkx^Zzn$o8b^d!zWu?UGHge-DC2h#G4LO2Ql8b9B-izfz
z(-HrK<-F;n;Q?uQgwS-vnOH6~9X1=A@TQYO^+}<;q4J`v#Mos~o`CBgR{gJZ@Qp_j
znp9$3Hu2t%CO0Q8F+v<mxx@-_0_B-Nd4WyPq~hk8D8#!c&s{>CLpkK|$0^M<QRAAZ
z2>$t-)VP~Me47fpO^t$f@R`Kzt3IURZK#PdZlc_P$dYr~;3ee*2Rq9hrw1OVBVj4X
zy$c;!LPvmc5@Tn=t}szh(Q9z;xN~y+j9p|6Zz7$W{O1|){ReU5s_%4QNRaH2O3`b~
zzQ)G<P`a^hHx^{4Jy^F#h~Hw}w?f>DoqF-7qLoDbwGRXYH)EQNHNG`z{MV%FW>;Fo
z{N!z$|Ka3sDH5)w=j&V~lRHTdIZ1P%6{x6^$*H~~f>RT=8i{RH9}V$(Z$);kC;+~r
zA?G;do5|;AEq+&nyN?>9|9tN~FwS!M@u7|zIUK8KgDTn>=0Ls{p$)2e{NTRd*VE2X
z8UI>R{lkOl`b@`YJLxf%(M)CZ`8*|;(adG^Ay0w1lRQQBc@&@iCL+T`<OE+pJ#s1e
z#moH20ix#$+Ngqdg)u@lOVp|Qb_%*AF&<`<v4Y?D4OrZO4WV!eJrf_Q`ml<Buo)RO
zBTGp2m3gvjhVULSlXA@z;=`2FVLnn&eiC7DloH(Je4rE`C~HV+o{Ep|#Qym|l0YpB
z4a`EJymKx>0YzvySP08qLIIc1aB#lQ-cRcpof_-=F@92<Kvq1l2w4;%dmh&yMGdlo
z2bCRJTtfDM!TiM%^_OD&)c=3ElO^ew?FY&khR%}k*Xcg(3Gpx~urj_|n8AE;e>bDE
zn=#^(?JUMBi*ewW{|aMsg>eGc@G`BW!ad*G?msu_SF$G^8@Gmg|7-Q-&D!(p<nPbw
z{423%%!&z9ZT|Oy7~n3-XBU4H(vxppWE@kXED7($ZhLVA7#w_dM?8S)n>%oC@84%v
zxHl(tkte#4xC>cB-18Tm%bw6Xl};Kq(ex%-8;(KHC+SM{Q64OqD0<G<QOfVhLXK*+
z8vObFL<E^IT+oPMu_5~u!LyKciXc~98a-=iD(MSz@o$(C`)zMS+3>MrhL0E(Ix=i@
w$S~1x7>-CZivRiPYc`Q+H2<$-_@5*BpJDvZVf@cf{^wZ!XE^_J1po8@0N021>i_@%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp.37D37DC699ECCA33.idx b/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp.37D37DC699ECCA33.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7afaf0be7d49ab01c48821f8672f0ee122ae8c1c
GIT binary patch
literal 11062
zcmZvidzcgDp~m<7MwaYF1q8(9B7{qlnaoTwH*^7`napG+)5y9k;SO0bs1*(hsSuAh
z3JpdLr~&KKLvUfKmm?y0E2yo4dTPCZdW6<`Ibtb|qPA9Qh{n@~Jx|#8kHhmg<hSqp
z?$0O54#`eBZPOIEh6epAd-bwZy8En#hK9kl|5m0~T>B-{(2#FwXjr}F*NJ({+b$lu
zHT2Z<%{N$2zCS#-e%+;a{c(afzB&5FAFf{FpZe`ianN@s%BzJn_nh389@f<O;5U^I
zMcuT_JKkEkY~O?rb1UW_oBrVCtIb`WMem>d_V1s6uxY_t+4Be9y-_T^|LG?OmfmpV
zO)u@eD3pAmEBN9q*4-bR`0Vh;L#t1Heaq{^%d6xY_dT$2{cFiD-ne=H@c$7mT(YV>
z`HR0dUc8(vnp7D6z<<A+xYRi`vHausb<dOE?mc?biOmlmNIdh8n;%{C_)8bOoxWi6
z{aqJq>s>eg*v6l~_v6p@e(;s@VKH^=-s|$yUs$lFX*C@`^Kxa;PruEj4?cJD`Nm(J
z{Y%MkZ~f`ZpH4e|WaiW+UoZW-XXVt9&i$6}JlDs}pEj%fyqiz`?X&Ur1~oizROi}b
zlf3U-cBJQb)4K1PV0}87z2iRXwak0Pwx(y8{mg+O>yB?2xaphG*NmUE=DmXnZR%rd
zJtu>YG<O^rc)qdwV0n<a{++JR@0oN-SMhi6y>WDN;b*yZBd>dQbAId7H?6uO(P+Io
z`pDwnR+|pqr>?rwnK)>ewPMyQT^kN=i)A+EJo_kbzy8dpbGLm$HjHZ=`K+40dF{R3
zvHSF^HZD7{sd0Pr;@~#^iQ}DL9sIbkh5JYzwfg$&-rF&u{g-c_{d;Cy``Q~~o!yzY
zpXqw%=x*bdYeye>aQ>D;(r!DPz3awBfBy3E?dDU`jf-E)PAP1^?xU}QhgOb#<AJ|?
zVx4^9_Az&i86E%f&HPgN<Ey`UHQzAv@jZ*@9&bE#=T7|>_ix%E?Aj#nm|7jV-?eny
zV{QBb*B_qVwSVMaz8<uN9reT;JKSxp4^=)o>}XFj!s}Nx&D`O;dFiy*UD~a`d0_f4
z{xbBzN4q{c7~!uy@>=J-J>f^nTmAcHcdy#Ma_&s_?&;E%a_9OXkN3oOZpoi>#iJiS
zVE*;pA$u=+_@eO-^}PJe{@*psVvj7}^7*&Z&8<)V>DxIk&U@iJ<*gTf^5&7Z500Dh
zUjsKZ?J!1Pdvxh5_SFNr4lf@$<zVaP-3j`KOEylk-}?K}SHD>`bI<1P!<WBx>3csh
z_ndg-sqOCD&OO-ZNk7`k&*UT8zN6d&NAEc``EcmE{cm2l{I=v%xx?>2{hu%2Kkb#d
z`u!i@Cy#pQ&?Cp5`;VWef3xq0=Uy6jdCN^}m}<xNJ=@>>>dQ&rZyop2h7|{{zEW))
z{m@Oyk2mMdAAbFgcS`3!|C+mapz+FpLv+cl)13{+?s#|nqKCS6{!RAo`Q)};rNEr{
zlxMGacT)SJwbu*}GcT^4`Lj1Zoi**$Imf5`=$ps$N2h-=y1K@`@!yX;@$>x)PrPXU
z^W5bhFFXGqw`$y@JAO9k)WT=}sE>Vl#{5oA8Tr2#Pr0qQ@aTH?r0ZYqeg1-f+x10g
z$3E_zhLxFRSEflrL*s>eH|X0Yk8PoeISt)I8;3Ua9ct%=zCX?1^?bb6QD?N!nYDOS
z<EX)Xhk88tiCo)Hc76VG3%#Ni4{jXX)OV=I-eEs18GGg7OIqmMT0E|C+|<59Jq~^U
z^_Pzv&pgpWXVu~njUxv39qRFnHJhG$sj~Z63!PVsr#4P?`VRH@`uO+zUTxgHq_tso
zEuKDLdZ_PEk00LDdu!K@Jp;~e=&Hr-!`mnH9qRG^eCC>GzkU3&vm0jA;xWU=jO#np
z<IlUtKF2Fx$}MzyEgsf5tf}u%k5BIY`pJ1~-to22j#}K-*fzNDP>+8Ylz8%^<*yuX
zp`EpONaK*EzC%4;ky^6y(1iZ`esXi)eV#w?eb3ILQ^vM-dsq)&zwc@k@`U@z%X%U6
zhP;sZSRZ7*kPkAJWg%ljEMy$ZLB@qR$o#AyGJnVq8PD>N@gW{EffXPVLIPwWD?%oQ
zM92bc0J1<R09lX?LKX}KA(L1MGASfMCbKeRa!7_uVHL=fkOG;?s*tH66*7(0Ak#t`
zWMMYU*Pp$u-QiF;+)p~IL#Bsx$Rcb6vPdWbnZX*686g8QlQkhTLndTVHVRoZ6ot%U
zEy%2p1)0s-kl7&{vKSkKEEbAE=CBTAPRN1m3h4^<%#UsN3~9z0s-I^%oBPf$@S~9N
ztRwx;eovNo?nN#T&sJb>#*1}tmUyx5Ef6o(eHkCteOcney01WdSZ6aV*4Zp!vCb9<
zi*+u;VV%no4(nWja9H<e{8;yAi686!0`X&=&+u61vxLVwUm!fzg^Yl8Axi|T3k4!z
zUCfAB7qdjfx>z70)&rRU)&p4*z<QuS0$2}bf>;k`Nf7J70tsSW%1Bt3vP8nVR3H-8
z<&2DVIZI@$%LO81UCAg|SF%LGx>6tt*42!Pbu~*=tg8j0VqME<Sl6;d!@5==8rH*^
z@H*0ewg_iQ_+I1!32z0~GdkAwEYY#97l@AaNG5{yNR~vf9x0Fr){TsTbt6j*tQ!Sl
zVBO4^SU0o8#JX7^Cf1{wDAuD{62*G7K%!W;G8WdYEU~a|6^Mm(J7Z(r&Jr8zc7fPf
zk7Z(5k7Y>=>#+ifVcp3%Sa-6-!Mal*4%TN&v#0f+_~uA+faXeb&rtpK)Y9g@*HLg9
zG(9$)1>>F$prmIpP};MqpS-3QcHX!bcHX2HcHXoXc0SVwJ73%fJ73ZVJ73xdJJw`j
z$HrOMu}KzoY?_4~XL7LP;vDR_BnLY#&B4xZ`eEmf`(fu#`eEl!`(ekMJnZ;54?8}|
z!;Vk$uoFxHc0yc$osbk@C!__~iKYlUF)qSROp34*(<1BwW&n19cmQ^RWB_)7bO3fi
zGYGq2JP5mBG6=h1ItV+-lwc>tCD=(x33gIif}L#2u#@95?Bt{jJ2@@GPB9hODRBjM
zN>YKHl2%}+nkwwnxC%QpslrZ8tFY5d4R%^wgPoStV5g-u*oDoo4QJ(exC1B|UJR5D
zuj(h=)M2N`b=c`i9d>$Jhh4;sz%CMxz%G)Ez%G)Gz|JrY*covHc1F^Gosl+RXPPGL
z%(w|VGik!kOq;Nano-zA<5Ac}lTp}3(^1%2rUg4IZo$q<TClUy7VK=(hMgU^VP_|8
z*x6|tb}=&syI4F1yI3*?yI48~JI8cj=foY@IY|e0PTGN8m(+FUwQXznOlc<2ENRvm
zs-LN8bKg7T{uBO!@eAhE|HY`kPyQ8Y?t6pW-||$6=LdNC>&a8k23r57wYQE%>e-51
zqMpx?E7bEHa*pxb$-utMcuoO(tHg^wZ=QP5=Pgn%`n)CTMW44qz3B7i7%%#~WyXs>
zUzPaK=gU(c`g}#|L!YliedzO5s1JR<9OFZuugv(+$5si8J~mHT^sz<CqK_?67JY1m
zvgl)T42wRt%&_R=s)R!ym!};1xFY4y$CW6DKCVJJ^l>?cLmyXWIQ03e#E(9Ip8C<}
zFH%4H{3Yr~pT9!===0|oKl=P-#*aR}N_h0~dCH@YFH#<Te2Mbt<13U$AD?4*^zmhe
zN1sq70{VnJ70@RXsenGAL<RH-6)K=l$T0%?gfb(bPplFVePW)9=o5=nM4woqBKpJ%
z711Z=7!iG9nGw+!sFDEs0(lxhU!X_>=nIr+0DXZ94WKWOV*=<4l$ikff>jbkUocOD
z=nEEU5PiWC4Wchtp+WQob4(C@!7>v>pHw9h`lLLS&?gnCgg&W6CG<%ZDxpuxF%tTu
zG9#f+t`Zr2a-PcQlZ#YFpIo9c`s50g(I@8^8GUk@k<q7AiGn^QPZjhjMXI1rDNzM|
zN`)%uQ*w-gKBdek=u@jiMW33dD*DtSRne!GsER(dLRIvsIYvdFT4q%AX;q@3Ps>vc
zeOi%f=+jD6L!VZm8v3*xqoGeLGaCBBRTBOI-VTTJG`xZKzx)apX?QDgiH1KzuF&vz
z$T=o_Cj<L36Fvp3SBZ{3Jx_J?=|!rePcKm&eR_rJ=+kqIjy}E2=;(`7Nd$e7JdL0)
zQlt^|MM^Y+zDR{e&=<)u5%fjMOay&Kl^EzV^3*_|QKSa?j1o1_XH=+xJ|o8%=rhWU
zfj+ZJO!S#~YNF39QWJe<iJIs$E7U}vnPW`!nPtXAU$jc1=!@oQ6n)VmjiN7FqEYlk
zD>RC}XpV`ZFIr}z=(DQCLZ6kV7W%9rwa{mksD(bOLM`-JImSYtRc0*o*;Qhr&(2dD
zeRh%B=(9`IMxR}wHu~%wW24V5GdB8SRT4vAEKg(Tixp`MeX$aap)XdUG4#c9ObmUo
zG803eQzZ`ioIG{V=M<@fKBq(-^f?vkpwG!M4*Hxj<MjJ3TmMBv<z)Jg^sH`tIeu36
zqVpEbuD>{sPrh&YaQ^OlNNe|~i$>}7Uu4(MA+;Cwr>|V~zhuX4*X2VzrF*9isBJr%
zI?k<s%lhB7m;0wLYt^?mI*~WI*;)r}?H&^z(^bEM)!N-|w9l$v?pkxp-gWywIl8a*
zEypCrEU1^(|IHY6@u9ohYF{2aXz=j*@0soz_|mpJ4y?Gn_T{z_ZKLa#wd&ujrMd5#
zS^a;Yb}6HWgm=ODR_`G?a-@eukc}Q<Ae%kJM2_~5D6-W<EM&Wf*vPRS5<_-+h||xH
z>^ghEKlgsR<@DV-wj)@-8y`EdHZ$kwmJPN0ZCqp=y#1d#W9sbs<Fs@2^8-4TZ8vM%
z5$B8;UVqE=8>REA0mk+dwJ#5yIFzqnEv?_JgPZ%V8us5oo?hbF1#if_y~K;`>m@#9
zwwJKTTrc5}{k_DG%=Z!=S?DDKve-*R<UlV8AP0L%5LxOa60+P&WMrk6D9CCrQIWM?
zqV@BEu1h{!bm8xRTf3*{pEciIe*jv$7oN5741Mya+b_L$@a_+4cWL{a_Qmz*+Mfdh
zP7XZx{JTy+c*8m_tZ)B3PZ)Yb^Q;kVwKuaa&*=WQre2p9$mj9_u`U*fb8$d^mmi3C
z@j!x001{myP{0)c3c7+ol1l=TT{4j3Qh-#K3Z%I-ps*`Ey8d)*?bcm7P{b7hGF%3b
z=`w+$t|*Y@vVd%t4HR?5fE<@|hU!<pM>Y3d^@gXn+cOa!0=E~)=k@`yZWf4hb3lH#
zABcDJK!RHU65S$Dz#RYzx`RNHTLO~ZGLYg{fK;~%q`5Vqusb{v9s;)x6mds@47UMf
zx=o;{I|^jEEg;)%1I64iAjj<hbx0j&eob`3wR0nF-+lTYnPC%$1?rDa{d`^f0Kw_c
zAi(1!dfNLRCr`?AeLs0cFJ#`77c!sdgUpxmLB@(KWNeCsj1xJ?xD*GOU-U!fPx&F^
zMIJIf#X}~D0%Sr;fJ_uc$i$QgSwIXx7Dxpk3yML=f~g>6k|;qYr6kB?QHD%T$&e|c
z0-2IhAX7yZGBu?_rimJ4T1tZ~EQZ_rpNHX8`1*d*MIACdr9&1GBalT>5y%YDfXql4
zkeQ+hnVB*ni;7XmqNylkmS{m{r7Xy7(T2=U*^tG=7-X?j3^GS_Aaha<WSvsynQLXO
z-3cjihUy==8QI+TQ5*PCNO<PJlRx1F@+EvgY=Q;i5*&~};RoUqJdltOfW(9d6i5Vs
zf{7rIl#qbrgbbu46d*OB0%-{iD4Yn-fhR{o2Z|&jKt{p<G7~0HG!X@|5*Co1uz_NU
z7?6{2&QQaO)Y9(LpK)5bV)4our*AAgefYmW3+gu-nC@Gb4!Zj8hEXpzUO9i!f`!S3
zMQ5^x`TaMO{Wq`dzqxqXRZFj_p|d2;oHd|vpsT5+_3T0Co?{O-^htx8?eLKE&l@`8
zf{QMmGIiP|p-Y8fT>E&<d*T20)!`#YjUGGBJz=89$NI%UP*x|8QQCMZGHiG>wwRii
JUuKaX{R`C&69NDL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp.1D6C6BE3C31CA169.idx b/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp.1D6C6BE3C31CA169.idx
new file mode 100755
index 0000000000000000000000000000000000000000..31e701f6796b2242691686246a1f25804137d2bc
GIT binary patch
literal 11062
zcmZvi3zQS}p~m<BA6c>wl|?{Y-r<pCCNq<Hq6-+CWF|A2M%HBsZ^(-AQK2ZLK)hN@
zp~0vDHDW!z1Q(WCy+wStptcI?trZ3J3av${SWBabt(6*L<LwQ5PuTCCb2*10zkJ`v
zZ_nnvvt71n3S2`&_!@id@>II#yoQE`w%XrS>6JH}Vj3FqEe#E8xBe=zU`6|7!?p#V
zp1I{_>#6rf<Th@&;;uhT^2WDB5B&bRYyH!|+9eM8`eb>nu<oAI+tb6F8Xx+y^1c{0
zE%S~yS1sQ=>HXZwg&)s+=*qR`ZqJhUPJi|HPmkWe=*{fKgYVufmfrj5!~M%{UVY0;
zdoB$oU+9)zyv@4%=*dqGZ$7m4%on%4KBBxvS-tnc%^P1!es<v2eIx!TjJtMCdFp3>
zZ@g>;Su&+C;=%uVCvk;yXma^G@f)8fzu9yAmXlk4wm<RgKW=?=-D59Z@>cqiEe~{G
zvb}G^#E&=s_}%Y(vghdM>ifmi$M@ctpZUU~bxmvO#MxJ>OMduOE`9L1)6X~l^8BC5
zhI`u&U;b#ui6gV8H~ISL7rm>dk8&Qcd>6TnF@M^k@e6J}^S4hX+MBe{;L%;{Kc3=!
z`&&nPe><b+u1VHMQ`tN2vtG-*TWoK7mf6SbAG+bhros1rIp+F_Q`WtEFriQX`FhW3
z>5=Bn{ev$yo;X+@Vs3o9`_p@-T;5&$?Yjq#Zz=pJw_((czu1!B_RKA7?npFRuZ}sg
z^f%R}gZF7`?sO&(8E&nd^Gf%ogWF@7%{k9r%G<9$`_cUEACgTI8b|#?OW(Tw-k#Wf
z;cGT8KY4%Sj^?G(cK-1bU7sI3R@lmYpp0I7(~a-$oYe8tx6c0^GofSs&9SbY%v;ZP
zzkU1(<F*^d9C>Kr)<V*5Kb*a5^^!lIdTfXJw7h!hYuRaq9XEdPxpZjN_yZ6A<wNWA
z3qKfp$JjCPQ*Y##DaWq+^3{C9?8ly5I{!rDnLBrdfAYZnJB8i%D?6uGN9}VhoAC2?
zev#|<&+Ohe>Mvgm*~*T7{J>6kd+WoM4-Px}GmP;1HBGa3`fgn|<8_z*{a-&g^Cy29
z_Ryo<9~_MEHynAbYr&JDN6Op$d*}A7*|BQ=Z1(P%^3_V$#-We(#&&JZUvSl<?>}h%
z^_`)6F8$f16Cdt<`OAI3ZJ5IzS+VuguV$Ltp8C^Q^Ilx=!bR$vFMRKfBX1p?FzY`D
zZ))0UjJe_XvRCZu26Z1^F>2bu)-6vY=<hGzJi~tT@5f*La?Nb%>8lT4`Q{bxe$RaJ
z<Ree-aNmC6!A4K|(N=yoAJO+7=N>$M&zY%*gE#JbW88|{lTYUkzxT|4y!^n7SLTNw
zICh^h`r$*5eEi(M|2X~Yz27|d(u6BpZdu1vJ9j*}<BiWxP5I}x2`_D0x&OMWwe~R&
z-=cnJOWyqES8sc#b^YtFxyuF{uM9dwul@c^XVb@byfbmh!`-|7rg)$H@b=xMz`Xdh
zUtIOhl#V6quOAU&UR*!>M+ZKdGvmw!C#HS-%g6G^XMQ%Oy3StxZ$}>g@xH|;Uo`)9
z;fiC+FaF2(b?(ufKN@mo@w0yjkAHdA!Y*AM_1~9GyS=#h_(u1Xn_ljF{*v$R{w%oj
z0QYvos?74M)1;xHaonCw;q6n$x6s7AhMr-K!<zaJwdaNYKh0nF9$W8dvs&ovT0FXO
zbX)(S9=AQ7YyaNvPhW1KSJmRS#<r&ZLp>g+{c+*-#}<BbenWRH?ikTAssB)qm9N^~
z`NH+>qb+nnEuP*u-RVEn<KREPIQ7Vh%;PO|PAwkUIC60Rp&nnG81~dnM|*~}Hq5QX
zGY8EK_8;o;^+)#&IsM&FkGIhIwRl3~gz5c<di=9-kz3cUes;(C4YO+T*b!qV^dIW+
zr#<7J<JD723!Pbuhc^yy>Oa)u(@%Wy)Pi+y`&wvcEpBgYZ|gtQ<8Ou}p88<LD<@iL
zS1lgeIJBw%P>)xpu3dF#(!lqAbaVgrJiq^+y}OQ28{gXF33@{HufMg&8}tJCf<7QN
z$O3Ue4#*$$1Mxu~NC*l*Vo(GM1Oq@)Py&*JGLRBffYhK0qy;q~J*WePf}v3T#&7Kj
z2g5*-U<Akr8bD^y1d0ZuKvvKKvV%5IEEofFf)3Cud6rhcE7kAqUCsUXa=1Y<o(*K+
z7V%_>=U(Ij@oWS3X1rMUW{DT;-U9Jr-IwuU-IpaktosVYhjli?Vx7$r7VB(*uvq6Z
z9M-uk;jqpX2#0ll#*cM>miV#mFAzW0`3#SBK1+D4^990VUC0Pn7qUdax=<hj*2Rp7
zbumjstcwLAVm*)vU_Fo}0jviKB!G1(BVk?25((>4fk;@FGcwlYERnG;7l@2?C8J<n
z$r1(YN`WX?S2HTs)hto5t`>-jbuFV|UCR;;>so<mSl2T;*7YpWv91@0j`dI`w1Esf
zLPJ>+x)-@XLfe4DnK0JFSrW#2xIn^Kk7OcPk7P*%>yZM9VBN?VSU0l7z`9W&2G-4t
ziFGqeOsty)Vq!g-iDEsPB~h$L3nYqlD`R2Z$`T9fR)JVpw=*`@?JTjeZWoA+^;jl`
z^;nj~upTRr7}lMPgLNlM9IQJ9;$VHQJa<O@sjjtWo;(j|zC8aN)nC<@HTS=o!_%PY
zvEf-T?&$<ddX@sEJ!=NYYkFbljeB9|O?qMHO?zSIGkvi0#eJ~zC4I2-rG2ntO%`@+
zoP`~mWMRjqS=ezV2Rkm#!H!FEu;bDk?EI!5cK)~@cK)OvcK);<cD%{Mj*s)O<C8q>
z_%shY!4zO8#0A(1Ndb03T7aEsim(&oBJ9MZ2s<$?!Y*J2U>Ar7U>8URU>8USU?-Up
z?4-B^J1Hr_PD)F#lT8_Ra$JU;oRncFr)AhFrUE-9uE0)7DzH=13hY!<g`FB#VW%cl
z*r{n1cABZdPK#@>(~=tOw6q30-PB>H$9350NgZ~2T8CZ84B7Ck91nE@B|}Sr(xEj2
z6gI=K3&+E-3n#;{3#Y@di<lAEMdA_IMUoNNMbZ)28Kwa{BW}RXNE)y+(gy5I(}bNF
zH(_TcP1u=f6LwKE3cF}L3cF}B3cF}J3OmcRU}wcG*jY&nc2?Sgoo(8%v*R}G?4%7l
zJ8i=*X2xI_i^pIWOU7UqOUGd6m=5fmxC1*U>A=oOJFx4PyU)G1ZS9#Y&jy+!&pAi+
z&uJ~q{h!YUp72AxB(xh}=zB>RInqla$VM+Qkj-9VB1d~k6xr$}7P8$-Y~)xki6J|^
z#2MiF9h^4ze{LN3vX@-6WN!V1uKp?Vm!!G><KsZfQzf2n;9a{XPd%IHz^j(0NIlz-
zOVslTa)o-nM$R#wI~mxQ8P6GDZ<Tn_=gm_u`n*NzMW44sz3B5+s26?S9OFfwx6F9a
z=c^JQ`h0onL!YlmedzO*s1JR<3iYASmt%bB^OYGN`q(OA(Z}W~i$1nUS@f|b%A$|0
zP!@e`j$zTqmKhd(T$OO><MNb4A6KLt`nVG1(8pCMhdwUHaOmU842M2{mH5%;&r?78
z{6*?VpT9)?=<`>oAASBD<42#r%=ppAR|$_kK2Le{@kPp`k1tUkeSC%T=;Lz?k3PQ4
z@aPk&L_nXArvmzfA{Edll&F9{p+W`p2{}eUpHOB5^odm>qEF0I5q)Bjis%zdR79Ux
zp(6Uk93!GnEHfhd0#y<~Um#Bd=nE8S0DXZH4WKVjp#k&-a!deyfie?7pHw9h`lLLS
z&?gnCgg&W6CG<%ZDxpuxF%tTuG9#f+t`Zr2a-PcQlZ#YFpIo9c`s50g(I@8^8GUk@
zk<q7AiGn^QPZjhjMXI1rDNzM|N`)%uQ*w-gKBdek=u@jiMW33dD*DtSRne!GsER(d
zLRIvsIYvdFT4q%AX;q@3Ps>vceOi%f=+jD6L!VZm8v3*xqoGeLGaCBzD$&uW=c$f9
zy-0QR=_RV8Pp?oNeR_`3(WjRg9etrH34H@E?x8#lZKCxT)z+R+k%qP*muTn{<O&Ua
zjhtgbcQUXqGodrU;VKEEFPx`g^o5HwjJ|M*hS3+U&@lSKIVOz0aG43CFH$8D^hNSC
zg1$(RM$i{2(Fpn?6&gWbB*#S17b!Cl^chuRpwGxt1ARu38t5}h)Igt6p$7Vl9Alu*
zC^H87%qlU_XXdGiKC?(o^qD1UqR*^Q6MbfmG0|t1854ccDv6>mnx|3pMT<0wzG#U?
z(HE`IDEgu~CW^jjnTev$suBx*R-Rhuvx?M0pH-q3`m74I&}Zcs3w>6ZvCwB%iH$xx
zPi^$sMQWqZE>RnOc7@vLvvZ7%KD*4==!;cJ41KXYjiE1Aq%rixN;HPPScS&W7t1j*
z^u@|d41G?OIOuco)Ipz9qz?L=5_Qn$RH%bKC&xJGbIOb};QQ9b&l)PH(|@4n^(>mW
zXkPv8_<23}<~%<2z7-?*yYC^bJ)<uj9j^cUz5B|cp3=S32i30bZ0fwQzODaTdnbSP
zbWi=AeEl7F?PI{%(>(R7TmO%DwG(-Ro2z%y)}FDUvEB94Kdn6-M#r4`=bv@A?b)#J
z!{d8v-m!_Xi|VC~e=|m3cIfW*+SP4C+D6oWJa*UMm$u)rf8`HqSGSLBA5%XkRA03<
z_n%Q3_;Ngb#IqYdZ+QEN7unZGe8_AcVUf8$!Xf+nh##5nBRsOuM+9WCkBG>DJ`zBd
z`iO)q_YoOc=_3lV+DBAmt&eEPdLPjT_>tY`5Bk@hkG7t@O~-di_1p62C)a1@9pAdC
zcDqc7On{$?OrJG<ZvC}n*V^X?buQmw*0v)r7&)T;k?z-O*ENHT9Vcs751TxUub;}?
z-+lQfOUC`~H?{wA;du+)byI84;`0`tqYwY|gDdWBd*Wzq+cB?WY5gAX=fI%TgD<@J
zuCrg|@Xm4d?VlG2!)|V#GqSyQZqDTyQ@_X6&ds^JKt7ibh;^|*oQnhUyZk`Biw6>1
z0+8qufdZ}okmQnpWS0!2xD+7Or2=U#4M=zCKp|IXO#S;_J2&SF14UdBAj4$<nJyD3
z>WTtcE(^$Z*+4N@49Iag=cs-Tc4~9~xmS1~xjmByZUPVM;Rnd;_Cn@my^#6bKFEBm
z4>H!xLdLQzWSpCWjAJ>-{BA#Fe%22e@8%)nSspUMEkGu)0%W3FgiK^b$O7&FWC1n+
zndFurlUNBd*)2mRvod6gTY*er706V#3Yp5PkZEoWGL6+B)7?5`I;%q#a)%}l+{Hp{
zh##P^I}BNv4MP@jM<9!^5y%X;0hz%XkeO~1GLtnSi@KwbMcF81mfM2NVlBvQw+)%i
z+K|QEG00+U3^K><K<2OxWSw&7xgW{3_FN@jb&gKDc5S8|Pn><796ot?p#I3HzvR>&
zA!mOF0UjZur(@tT;YoRJ8X&Leh0L4sLgo{Fkoi(R$XJnuj7_nSaUus9m*ODvi+;%b
zDL-Vq$V0}bc*q1%fJ{gUkcpxQnV1qG3y1;80;vFGk|;qYr6kB?QHD%T$&e|c0-2Ih
zAX7yZGBu?_rimJ4T1taV7j?+=lnz-)40Q}VCPJyuO#>7b!;pnjVaOt41hPmf0+}Hi
zkQpfhGE+1mGgBsHQ85ZxG!=!+5-rHAlm(eB+K|~P8?u-fgDjScLFR}KWKPO~tV`}X
zcb2lXCm|=!QT<nLMm6{Ustr685}tW*=TCTnd<h>An_z*s1PA0#_<{HY4<sZ6ATc2V
z1rh-uDIo#L2^mO9C_rjL1=124ke<+iLW$5kxN{`JK#@cQ$VeDKX2JxDCZa%A!UD1r
zHc%`P19B41Iciv$TGn&+cbrzOT)Jx3*$a!$9{%^wqWXmfrl(6f-LY`csaG4XUbtk@
z;$*{;b6LZ}fs4t3i&qa^T)O<4W!Klxxi-$6H>h#2tEr{+{2>=!V7D3JDQ(SmXz0Zk
z4I6pMrI$^cKI8J>6~b_?W1{XI_y1ixV$|p{<0rT$P4@U$zZj4dZR%LHotGoSM?_;w
Msd?qMEb{ID0TYw8<p2Nx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.24FD07A2963099BE.idx b/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp.24FD07A2963099BE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..51d7446a6bf7c5e1ef21eb3cda15e643db72e350
GIT binary patch
literal 11062
zcmZvi3zQS}p~m<BA6c>w6%Y`YhY%h~W->FGC%S;KNoF#WX=GiN@P@1ytrac`sSvN$
zQfM%0K#f>WFTsVSR$Iik1+`UBuhs{s6<X`#g0&b#Z7tLg8*gvedqTc@&gC5E<jeQ_
z_P1^}vw3yfrYUeuP2nr;HOo`!-m{vTnugZ?Ta{jU%@<5lQ@*XKY0Z}3Bo?gbxOn*1
z;1e@9UvEAB-pJgB^_SlMr%B%U=IHBxy!tBt^sje{L%um)UL&l%^W?Vlh?eI2PgFh-
z!=`24`sS+TdnSF5Te<L)nfG6|#@yvu^4`g>|Nhzg_bhrdd;Z`%Hi@P8K0dr}+4VQv
z_`>dsg30H4q~~w8?s)(Bv4fiqtU2}7&999tuU2l@bKj;7uO>f#{ieMm|4X>=s@3JG
zpZ~r2;uU1cl)}jSe*aG5Qs=<r^7rG{K1*KNedNaDn}4w{@zg(VdT8y#FI@0e`hv~(
z_FS;7fBnQyHvRP7?;qR!{+H?p#ndNvU7Mfz+@iHDYv{z;m#IsB{B<t9|Cy7|HvjtU
zpUZ}O>yKajc*fB~v!}QC`sr7FtEP{3?zMd9xjth4yjkNH+;r-1$0phvwb0-(-RnM?
z;(hx&hx&duqxben*2h!XTkp1B&AeOeXnBg+%j_Gr{^-WR_na8}y@^xSzPmr6Pyf|A
z&q?XQ)~<bn&o>_3Umjv^c)RDbJEvUIQ~ce#uOHc5_(^X4=xd+eoZtH7jjL}>G+Qr^
zJ+$<dYRmq+wbi#dlZTA3R?c~;XXE~DvCO8NXAkA=*Pi-#{<gzp<AmnXPiyI$*4@<`
zyE}a4rsc=)Y2MztRNBTra<u!){T~&!a33mT)?9b(yE`Uz{`{@8|G-S>Tz7q}yEpUJ
zQ$24VdDOW1nz4uOU$~`^v^x%FZ@*#5U%q&FyZMBC!_rr?(+b<K{qRfaz^d`D-}mWZ
z>*RAk8h7isvGFh7$S+eqy86V+`KH+qKelxK(dJXP?F|3y-g|ZkyY5kTOs|gK>smJ9
zR~`H!*B_tUwRiNVUk%y9j(Ozu9qx|y2Pz*Pbo3_~;k7GUX7BLbv~0#}F8znUy>I5v
zJ{^AlLp>kvkMP$VdbNAOW1$DjTm5_H_O9N(YW{5Yj+ycmO816g5BJ4(Zpoi>`9mMv
zXa4n_VY@H-#YGbz=zH<R-rqINVGpg?^4Zrj&8?6B`RjSlFL>@e_08vg@W!FH_D`7g
zpMy8H>@db&b7a{|_SJ)W4z3tIZGZdbM-%jqmu#A0zxnqgFQ54J{Kq!;9=z<$OW*y0
z`PlIXpV;oc<=p+vp7cZQ{A@m=?>WNVcjV4fQx68O-TTIcD{e_XkvsU_lfQiN-We~=
z58wOI-O88;4m|kDGyn0^^l$fk`^*ayE^E7SEmQ5<{@C_6zWid!KetYJVdKhuS6`uZ
zjD6rn_4}Li=C{9j+dHlMUw_SAJlJ??&;feY4`(_XKe_dtiAx^n+4(oc``F=Ic9jD2
z;?thK{GBPCOV)jFWQciw-Rz&d{_&g{r_MP#?Yk!)&L5fi`Pk}O`-Xo%^vF;5E<XOe
z`LA<Ve6;-hfBaDA9@_DfA*U8U^{4Ro7iTT(*45Gfeetwgii?kIa8J4J#r|h6__tl3
z2Y0;8z1_4bv;2xQX=-Y|aQDXWwyERWXkuPd@9^f~Edz(zd12t+;8{ndD=I&}^z5dd
zTHQIabJD<}UblYJ_t82>o7F~V*Xl9NV}=eK>h;h^aveX|_1TMU^zvFgw0UUDz@c9I
z`krY&cF}X5_NKYDdgh>+!GS})_W%Czr&s>c>}jL(YxRWY3DXA-^*Z>^ufBNjXy%bN
zI;U2TY92Ls;83sUzVb5f8oQdb(FL`7dh>K=;83scEPc9^NgSSbcGIj{J#OT<2?K|E
z{aNq$XL$7srH#(4)gzimv<w{T^~p!SdVIm!w|#B2t5$b3cMKgk)a!4DBp&~8#Y;!q
zXm_n1);z3b;83qurmk9bU{d2}-`zU!Gtck)XWz~v)5f>=dfcAL_5b@?Vm&;hpeNKw
zUbh!WFY5);8}tI{bNhhwu|6PuK_8H;n+1|(Ss>XU3nb^}faF*XNG`|$>392q^s{~-
z{XsvFyqgD-XL%s`AP=PA7Jw930Z1Vz04cgfAVpRLQVfbf2HXK418e}uKrjHL<d%Sx
zSP4ieC;=(EWgul%22u{nKq_tpNQG5^RDueSs#^t8WmO>6pbDhs)_~Ml4M;7h0jaxn
zAazy;QV;4thTNga_4`+QZ-@=?kb<F5BZb{zAj51J$Z#+WWW*f-GQviHj07V<8g2tf
zgEfFOf(DSL+XT{NO(4yn31rkA1v1J;fs6*DKw54KNQ<?Aw1O6pw%Z2MW^Ewtpbccq
z9Ro7P#(<0kV?a7?2S|r?fOLWmkX>?@cIE`zd%1i$q*?N;GpT-tZ)+Vmvo}scp*|AY
zMc}6FBVo*uJ`%xf^brHI*+)#w(LNHzZ1oWfv)xB*%&|Ta!|e1Cr;+P_d1~j*(|>jv
z-?k^?Sx*`#NKck{?!sIko~@9*8871AEb$`lEf6o_zKjoXUzYd~_Z5f_aW=yu&SnXV
zI9nhr;#`J9oXZjpajrl(#QhmR;{Gi0BknH{KjM6bN1V?R9&x@vc*KQ_fVhw)0^&k}
z2#AXr5pgj~M8w4c5fKk$0*D8)B!GCJKmv$M83}PIOC-dl0+A4xGcw|GmdJ?91tKG^
zWE8}eEKv|w3PeF%&8Ud0S)wAY7Kn<tmeCN`vP45%D-aEFJ)<M8XNiuuULZQ+p-gB!
z+)G1Q61od>frPd~4rjuMhqEM%c(_2qh(|IJ#3NY}K|E3*5yXv*fw+++2I5A67>Ju0
z6LB+3OvKFsF%geuqKHScB#L;nK%$6S84GbMODx2#0<jRcGdALOme`2f1!5x}%ft|m
zWl0S2Sb@Y4cQOv*PL?=`I|bq(K3ASQqjC10C(na4U!H#^)gNImYaMuG4W~iVW5Zc6
z?&*S*^elyx_N;Csujz%IH|~X<H|d3)H|>R;&-B607x%%=m-NBTm-fMqHCfoPaTa!L
zl7$_cW?{#f9PGF_2Rkmw!H!FFu=AUK*!kmr*!h!w*!k0b*zqP0J3h|Cj!*Ki<I_Cs
z1XF;W5Eo!4Bn8+BX#sYkDZ);Si?9=uBJ9Mp2)lq8fL$OSfL$ONfL$ORfSqJYu#@5v
z?4+awJ1H%}PBvxO$#EHWa#Dt!oR(pym<sHaxB@#RslZN2E3i{d6?SS|g`JvIVW*~5
z*lDH)J1wrkPD^UA)6yF3bW?|&9@k-~Cw18AX&rVUGh{a&JB8w*E=bAHQb_60>P8Bi
zVc3P^Vc3O}Vc3P!Vc12?2<#&92<#%s2<#&12<!~gfSnOHU}q!^*coX9cBW~<&WxL|
zGm|Fl%(Mx+s2PP_G#-UrG#Q0mG#!PVWm>Sa;uh?zqy;-GZNbhqZP?jy8+LZmhMk?Z
zVHY!Fu#3fGu!|*Qu#2T*uyafYc23-Zos)E6=cFCj^~gPE9+J2B&X#9Gnj_CSlj?sI
zNbA6BCG``2{RB>12VNtoU)|ojXyT%I^-cX1lV8`yr(Z8=P@XFBd<&0?J$dTcNb6s;
z_7X&qdbVONQO_~V73%p0bB^)c#=ySJcuqm~R*4sV-aPf9&s(Hk^m$9vi#~6KdeP_2
zF<$g}%ZwL&zAEvd&zGk@^!bX^hdy74`q1aAP#^kyImU-RUzzctkF63GeQchx=wpkN
zMIT$DEc)0AWzomx7#4kOnPJh#RSAbaE>Ai1aYf3Zk1J6QeO!fd=;Lw>hd!>%aOm?_
zi64FbJoTf`U!;EY`AgJ~K7WP!(dW-Ge)Rdvj30e`mGJ1}^OQ#)U!*+x_!8yO$5$wi
zK0e3r=;O-_k3OMF1oR1cDxgm&QUQHJi3;cwDpWw9kYfb&31vn=pI9X#`ougH(I*zE
zh(57IMf8akDxy!!F(UfJG9#idP$dEM1@bh2zCe)%&=)At0Qv$I8bDtl#{|$9C^G@{
zNmU}DPs&pXeNvH1=#xrRLZ4Kj68fYZBcV?!GZOmbDv{AA=c$Z7xkzR7$t5bIPp(iI
zeR7VG(I=M~8GTBXDCkr2R6(Coqzd|!5>?QrRH%YJCC4b}Q_75jKDA0z^r?BOqE9VS
z6@6-ns_0WIR7IbfV^s91WkyAxRwWwxv^>?&rxmG&KCMJG^l25Up-;;(8v3*{qoGf)
z5*>Ydp6ck+i&RISUZOhs^a|C{r{@?QeR`SE(HE+c(6{g`Ih3cNjkNKoEL5bSt(Z$R
zbPRKahQ7g^V?wtvurD*AQ;@?|5=LJ*Ps8X77ik!M;Svp_FI=Hv^o4Uw7=7U~6GmU8
zN+Rfs<Y@$bks^(tFH)ir^hGK(g1$(OiJ&i1W+Lb_s>DE_k*5awj3PDAXOyUcKBGbn
z^cgwEK%Y@&4D^{*VxrH?Qxkn=k(%f;OVmW4S)nHS%p7B)&nz=0`l3}5MPD>eqv(qk
zX%v0Y5{;rSTA@+&MRQCPebF)#MW0nA7W%9_wa{l3sf9kPL@o4L6>6c+$}txDtTJPv
z&#n?1eRiJO=(CH|MxR}xHu~%ewb5ti7#n?dnX%CqtCAS{VtE=vU#v)D=!=zT41KW*
zjiE1=V`Auwm6;g&oGNk9=j5q_KBq_>^f@K!pwFpL2YpVCanR?K8K>d<&W6vMDksx_
zqG$b|hvR4UE;(<>-1>v__|&^sjO6dQleG7axoAwd{`z*$Wy3tByQUASUES5vb#9%m
zpIT~<_fNmI*?1(6U-AYg@&-3o@1pI!<3i(l>Th|r_jVecbLy}Auf2Kq`n`vb?5PRH
zCB`kPLmU2PjJf#09UZl+hYlG!vi>^L?So&~cI&>CKdN2bF{)#1{e95-YG~`gyQqzC
z$J0+dyWok1x1V@1`}&CwGuuyC%v?X=F#G$7A2Z)ic+5gS5ipDWM8q8ECjrb-KanuY
z{Y1vB^b-ZM+D}x>T0hY+>-|J;<Og@1J?LM%Ki+cs#v9)y)o;9C9bcE3cVx@P+U+tS
zG67y*m_BRz-1;rMbIr4Zx|VM@Yui!hj2c;g?)F=?`^rJa_T#myhff~P*WbS1*K^6S
zB^UnTmD-nFc-BI9-PGQ@_^ice(&0b<=+e7}KKg!b+c~duY5h0gFM&ZP2cLWX?We!X
z5nUJ7w|`k648Oj0&Zv&s(?6GIY~#tC*X4!ebNL{#E*28!;vo55en`BFha|WJNTN%G
z6mSI~NiGSJ?2;iVE(Mb6QXy$B4U+ECA%$F_vGt#OdvDkkh7@r{AQ>(LlIb!bMO{%y
zmdk=<yKG1?R}7Nla?Ygs=MHMm&`*EppngX>?%KJDc0PLg<S=6Lh(MjH-~Fbx4t)He
zPS<WQ(bL(u!8|F?b&cc|z0i46Ug&(H4?17U2OTT2(6K2NI!@%E<5C=Se$fw|Kjnvx
z7kTLT6c3#s3eX8D0Xk6>p%YUgbOA8{T_6>JP7)>Pq?81mEXvTyDH%FNRG?E*3UsQd
zLZ_xw=rmD-PD^Rf>7ovup3<QUiJ{KMFJvecx~`GJVi>w`Dhypjj6fGjMW8c813DvR
zKxc|3bY{wgE-FT$i>9K`S)v7<m9n6-MH@OhWkVMeW6;G?G3Xr8fzC-e&~?k*XWsf~
z?@h>wGpYWOo6)TUAGLwILc%i-e)$t#NWO#*5}RNlaS0BRKjDYOCwNFgLVzSDL`Z=|
z0FsoDAjt_Cl9Es$sR<R5me3&S2^~@>5t;|T9EmWbNFoBsNEndJgb68{h(fXw79=}i
zLy9G0keq~bCN-^0E$cn~8K+e%m#&(1`oiMVhyVSzsD7b|>Am~%6R-7ii$l#<EL^f^
zak6R2nXYMJ<6^RL@ruU9rOU5e_Pttku8K2f4Qd|jYH4dfd&s%x*h7u*l%cJ5XxRDZ
z4Ig#EMHf$-KI4+$rNRiVbE59O@c+AZ<mfSD$4_uin(XnhelZ{^+SGAs2QNoPjEu&X
MQuDI!Sme9^0|2_)ZU6uP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.32B3D11C83B5AEA6.idx b/.cache/clangd/index/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp.32B3D11C83B5AEA6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..951c4901861bf6ac501ac6cca0499e20533befd7
GIT binary patch
literal 11062
zcmZvid6*Mrg2wy%N}6=jDj*;>cQ}$%rK(a@K+9z%RjEo<Y9ejZgd5UgJYi9YLR=kX
zV1iKtYQ*uGCD_o6$BK9>sH1|q-XN|p<9Mt%4iiNk55^FU>xAAX<o#oL9*h3vec$JY
zm7`Og?yyZ$;2IkAi|jSalc}!L8X6h~RR3F*T6x)5OhZGiUqi#1Eq{zJSkXFV&{p;F
znVYY&9(#XCcEkGMEq@v3jcty+`R7Y6_D}y|r`Z3;W2H61+S`tAOAT&nyyyG!hoWv;
z=1p&}TE1u8huM`2Kbv{a`D@Hwo+a-e|KYnYKDcwy+nKZb-nvOFzW?bb`<7jG&9yJ?
zK37dV*D1Yly>;sc$38!_>A;#3-(LU5kkV@9nmu=K+VEQ9%QvsvJLG?aQ5UZ+wSD<r
z<CGO-$%OonyZ`&$c+fd8zI1u)if75IyN_IZZ1a8l;!pnbx(C)i^x`@1q|VuVSLZp~
zde)EqY}4=FyZrOrAACLW!$R`2+poyYd~VU&rZsfz?DHos`OOd6)c$9VKil~9>3@(7
z_txLM^y!SFhi6Z3^7YVfyH`yg=G<lZ&T@Us{B`pre!+DozWIEty>XJ(cX-FT&n9>e
zh7NbXKBMcFan`48>`ixAuchBBv^G7->}B>1Tz_<9-#fn_ap~9zYv0=+pFI7+b)MtW
z{mt$B`krk(vcJ^d+;Fh-i`yog*I9V|y*H0+&i^*Me%KXHZO(0d;@Z_W#T%_xMjT%H
zYNcua9g|ky?2PX}*jhQ~<<5=!w?)&NvYtJZx8Hd3)A`#zAsfdu4tr`+>biBecSY~e
zFWR*H*qx2ro0m%4_=k^ne7*nU{1)ybW%!yauXu0AxT$}5=kz}@W2UaVD%#PNe&@-~
zgGU}QuD@)=;d>Tt$tUdALz!EyS@O599@=g`E?=|swam2q_A5U6S~{?5^qY78{S)i>
zb2pB>Y2=95S8wH(DIZ_*{VTbK*$+LsbpFxC6F2YFe}C7VJA_?#Dm$iEhV6AN8}nc*
zzsU9HCwA=}_V;i5Z()Z&{N@gKYs<amj}AGLpJ0SHE^3;+!*|`X8E?2I|N8m6Xa4^0
zgYJ2t^P~M?{<6cbbu4&PyT7#6zh`dO>g}uM&t`9(DPO2`Y#8`ZcXa2L+!+@<@ZsI&
zKi(a<``r7^9eZ#0OW*H(y<rY}c*T}4ewb-)eeAD4%zI(Mb7xI_`?;&%I{eQ5F|+=w
z@5ZJb#)!*~EPL6$q)+Fe6~m_OZ`u4voc{T|O*8DbzdQ2E_fOB49>4I=`ELi`yV`v8
z*!_=hci(X4{zgygffjx?AD+DD2zU39+fK9{Qm@$i)~FRXBp%NmdjE<4eCe(kFVENS
z`uGlI_`L`2|LmFn_+9GxJwHG5;+XUMUAva4v~PcO`&(asHQ}eNV_w|2a^EEvPHG)-
z@3j*z-<&gl{^LRKw2oi?k)6`lc)8C3dhxGkIvYQ`>D{qQ?(N+9jpBXulN)vw1M_0j
zp1R=O2~(G>yL5=gys&QeZ{PfM&Wsah9G&*7?;pw?nfc|2%3AxHe?R>2@AfV}_JaA#
znJYeCe)d0qJ(+u8$8Y<eSp4K)^wBTPTG%mp;;{dnGVO-K;v*Z}6Rv!z=h<`qZP%CT
zjz4h+8&;*4Uzj2d4UMCAZ`8N7jqXR|^BTGaH4bX(Jyc&8djB;4*!}T3XVR>Gbapi!
z-Z*?f@1YhCcsSd7^{y{o>PIiA#seA$H1!^8@mX`OWc+*QOzKDHSK~2_W2W~WYVpKZ
zy1rcHJT<+gVQw{^*=MHOd#J_gPv3rZ|Izfr{pg%(JhX9W-`+zlzCLnr%j@yncc(XW
zR^zrIZR2_mwfHgDW4rG=gP(YM!>npNa>&Rry@y)-#{T3DJEm@}{>v9s<0*|(oZdq%
z{-SI2GyKG_lzwz(H6GkJxT*J0i;q9@?PCkp9`yC2?bW!ov2{T2p%(w#KmOQ9D_%a@
zk9Jh!fsF&3dJnaDW%A-x2gcRUeOq(yInV9;se9*<X`@@ZJZ{hU+PSNe>e1@tb$cQ6
zs$R%^ZXaYm)dv~tW+7u$7BbGwLB^>ZWPZ0FGQaAFjCb>p@hT6Q;1(bgQ~@&4EkY)$
zB4hz~0J4A@fJ|~rkV&cpne3J!lT{fq#jQZ5s0w7NTOD6}Ftl{3s;bo~=ng^_RD+O(
z+#$$9Y6vpTtwE-#8f3a#hfG&>$inV0WMMT7nc+4dGgJdI(``a#swQL+cLcJC8iCAm
zTaa0*1)1%(A+uE*vZy->SyYWe=C~cm9MyrWU2dOr>Pg=<OP+O#YOgaL&AqQL@Ki{9
z)|2|9-;*Jp+mZ9cvlZBz_F~<eAzrL|^TdmFU)qOtUxxUw?#mM&*4Z?Rbv8p-th0H-
zVx3EKSm!c?!#bBI9M=75Ki2&j;>Wr_PyATt(>&Js4B@fP=LwH>AuV8C$PfYRLY@d%
z7t<ov#S9U#F6N1d^*}m+^+1LMupY>h0M@0ngmo!HB&<t$B4J%l%UG8)M8>+DCo<NR
zw1RadLlmqld7@xlO{?ok{YkH8h<ZD6o~T=agXtjFgBcRUdN5CdSP!K`SPx}L2<xFd
z31MAJYgpGZM8mq4CmPoEw2pN>Lv*a`d7@)IoDO3>oFQSXhw~(ibt7$H-N+CF>qed!
zSU1xq*3Aqtv2NyxiS<Z2g7rv-M6e#olL*$Ww1ssmLoBRYd17JRPTN?wGsMQaohLTd
zqv<HtqZty#dNfa>Sa;G6)}0J-u<qoEgY~)c+!^&3zIpOIp!xFrQ&f9DwXC`KeH6S5
znjRZo1!JCepoC{BP|CBqPF~XsJ8#SjJ8!}ZJ8#MhJD=%;oiFBtoiE{ooiF8s9c!|%
zV`D7r*aQnZHpRk@Gdb9CF%EWIf`c8G;$Y`D{jl@L{IK&U{IK(<{IKIq9(H_;haI2b
zVaKO<*a@ZpJ0T{(PDlu_6H)@~L{o&F7!zS9CPdhYDG_!7GXT3lEC9PeA^^KUDgZml
zlwc>tB-lv_33gISf}L#2u#;ml?Bs+DJ2@r8PB9hODKQ0hN<x91l2Tx&nyL*iy)m^N
zD4{L|N~x>s6f}dd3&w)53nqfF3#Njw3z;F<g<>Jtg%Tmyg;F8dX{H7{EvCUvOK7mu
zQX1@ZQ-_@%(_yD4blB-B9d=<e47+eF47+e547+eD3_HU#U}wY(*ck}}c1Fs8ooSk|
zGh-&~%!CO$GiAaqVn$#WiA7)+Nkm{5Nkw31nHKD<m<2m4VZqKyS+KKB8+LZghMk?T
zVP~go*hS4K?4q$K?4pS%?4qeC>>Sg9ofC6l=Oi51IVlHropR@?_qHuvv*p=9bL2Uv
zsP;-tn|rT}>o53=#x9yy`;Jljn>6=cAJ<B?>*J@Zx1YQ^uH9a}VyqC)&+x9;lcS!E
zwC1Yj1?t&~T%?}Qk;~NcBXXAU+|0nf#CT2sdn?3?K5veC(dR8tFZ#Si>P4ToOugvy
zW*IN~yd}nqK3|3S(C5ohANqU+>O-HeNPXz@m8lPXzAWQIpRdID(8pE?i#|3-S@f|5
z%A$`gQWkw|nX>3(vkZ$qw#2aL<0^zhAD5#X`nUq+(8m=ihd!=MIrMQ^hC?4$VmS2q
zE5wgJe~$Xm=Pyt{`us)eN1wk;{pj;&89(~`CB~0FzCw8P@j1$)k1tRjeSDGf=;O<j
zM<1VMc=YilhDV=JAp-h@92L+f6sUkcp-2Vv31upvPslO?`h*fApiit25q)Bgis%yy
zR79Uxq$2vnG8NG$W*HHEVu=ya7pRZ``T{u`KwqFh1LzABX#jnJG7X?FkYxhs3zV1u
z`lJex&?n`ngg&W2CG<%}Dxps*Qwe=imXXjWl^6+qa)rp~lXFx?pIo3a`s5;&(I=Ow
zj6ONb$mo+xjEp{|LKO5VIjW#fDNqG{N|7q)Q_569pOR%1^eH7qL7!S7>d$bAsphD<
zk=Ebss0FHSMJ`hHbL28re?-nQ>dg%7ON@E~I9MS;^aXP?h`wNf2GJKR(jfYRWg0|Z
zFv|qd7c4PB^o1%UguYOYhR_!(&=C4UMH)h1s7yoX3uT!Q`a&fpgg&i8H1ugXs-aIS
zPz`-rk!tAE%2Y$2mSr^bX(dKOpI#w4`t%&t(We)vjy}Ceb@b_Fs-sWOGCKP75~HIp
zTp?lfg>y8FzHott(HAb#F#5t}8b)6@%Y@MvE-_*B85Lrn&&W{&eMW&A=rfAcK%Y^j
z2KtOFW1!C{F$Vg~3Ng`V=BSB2vp`MsnMG=%&n#0DeP)(1(Px$z6Mc~iiJ&i%qY?B)
z3N(VgNRdX+7b(*S`XX5-g1$(JiJ;G_5DR@)j#}um3e-ZMRiqaBtTMIGXJr`+eO8IF
z&}UbOjXpa^ZS>g%YNO9CQX74CncC>Hvy6>CyTsV&i&jV!ebF3^qAyyYQS?QNG>X1x
znMTnU%`#E+MN3Q+eNKfq=yP(^L7!8g4*HxTb<pROse?Wz%Q)zBN{mzYg*JTIP(Gge
z3q7q1-=&|{wdAZNb8GL_V{Lb=7{cFr8)@kpe(rF+_Vew|^9Op0w@>d=y}7-q{mfeX
z8Ylb~H+OP7ZRr}RjqI#_cGc3=X0*+zeb80=`;6w^54!4a_tkEq?t*V0!EO>n4t0|d
zver#BWWAf{$l-1hMmD;MfoygY6FJgNBFI)Zv5@U<Vk1YpNfg=XCQhB#UcY<&-cOF~
zsh*;d@sW#a=VHS*#_%ZzZf&jJJfQ!8A+?{B-O~5PZ8z;(d1LkF)}gH<Y9Gwqzw7iq
zzwG{W%gOURx?QTB=Le6iOV2y9Wn=X`j|q=~pIA(pHDzw?`{m9x&-Q6wzTK>Dhn_KX
zNbNGK_7AC^{*#|3)=$5uhj@0u)t0x1c#(ZQ#D~oG5Ehy1Asn*5hxn2C9>OCFJw!kj
zdx(e}=pg}QsfS3&au1P_l^&wh`T2<*7xgi=AFKY~4;nv+uYEqeuk*ammyG(;tJO2N
z@U(^Q+6isxT725#Q&c;H)vMW)pQzSP5k50M`Rk3r+Xp=IK~>N;uWf1VQT?|-pW}Vc
zJo}cD_t@a}Q8mHe76^l`YMwK+wR+X+@{FiogL++FAfL+z#JX4@&cy-wU49_m#RCa0
z0Z4R-Kmk_(NODO)vP%Y1TndotQb*LzOG{VK6$A>oLO_~J1JYeOP}mg)GF%3b=`w*L
zt_YCjvVd%t4HR`nfgG1}ifW&J4{z@M)VqE^dRPw+cLD1K^07W3mSurBmILy$ejuLZ
zfdp0n5?K)_zy^RMRsxb)8AxFjAeB{lxC_`IP>2lyX{-jMvpP_i4Fegh0c5f!P=t*D
zS*!(Qvo=tajRHBW19X9W!Koh;9dqs6M5jJ-^1I34@q+`k`=|D5SN#QolfQ!i_mk+E
zTECw>Nzava@`_%_yh$%)KG6r6FX@Af6<NsGBnufQa*%OJ4l=*!hs>Y!L&l3dWPFl`
zOb`XggroqOD2kAYNfEMu7=SF03_vD{5@b?Rf=m`=$mFC9nIbBXDM<w~RaB?eA6;rv
zy|PY0F$h^O8H6k(h9C<iLy&2r2AP)BAk#%1GCiq778b*hg_B{(4AFqhNE(otq6wLq
zG$D(K5y&FR2xOLML1rZ_$ZXMu%ud>nMa3v&(PR`dM|2=_k`81Ya>uF9%38YOa{Lt4
ze&uFZbMLR(z*8aanFkO4xEIJ5_W`kS7Kn>;K>oNNh>!C?LR<h6<04QX9srW!5|A91
zft0ucq{h{G@Gy!8fkN>RkQUc~^tcWbj)#GaxB+CwO`u3T0%XN4AUkdYMdMK*C+?i0
zhLy=>T_=CXY1PW5t7e`2V)4nt|NU81`=WvAT5<W(U*5_7^lsyY3zsZfoM>2bDr;C+
z|1wek^1}L;OP61??9wVa7sr^>`ZV@+HT7#bz5khK*aHlG!hmL58+i6vgNB}S?v!cM
zXPl=7g~8m^v6H={{=Zv?3>!XT^ceTJ@g5)R7Xy+qscq!MR$dMd9ukQzrRMn|i~Q<e
DWcS<M

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk.hpp.4C6A715EB09F77E1.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk.hpp.4C6A715EB09F77E1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9eb6fecae2defa8225bc597087307c4ea488e866
GIT binary patch
literal 17172
zcmd6Pd00;A|9{>0Imgq}>FH_TON-K;qM}HVI%NrC$YkHgGWOjJgOZ&PvV|;(2p_ad
zp){#TQdBdRjF3iz@Vg)9ygO%p^T+4<T-W!Haa}X-=jDE%bME_nzxVr|8N&7J$M#WE
z>obvCJaf|gag>^xnhyDA-u$`Keqq$q94yq-7Vj$<K6Bm!U*)|)Z(<(xuTGrO@hG<^
zYVzWYsOa9FYqOW%k^Qi;Lv?fL+`Sj>eEoU)E${8;R(W5k=^uS)#lqp`KaR01EjVR$
z`poi;Q$x{@EwgUgZg_G#y5-N1WzMgr{~mXGap$f}rZ`(o`+ewdAGU^i#tf5ZR@Rv1
z?wMTg%goPS%bZ`38TIjQZa~8nudUw?^*a#TP-*p{>!HN9ry5Jm1`j-XBPb~;<jCl7
z&62Uc$(1kWr%qKC+x)pW|9$rHbvrlgD4JGTH9uo?LwN_6)6>jr=A<7jk92O$Qh&eV
z>5R#b-ZzF#VSKL~@gEjZcjftgD!{nzzFPGF+_>q7uo}-T+c#AE#kBUw99&vhq9Z>s
zVdRPhD|+W_UAEx-LPw`bBf^tMFKz1XkT_yUuQBt&YF1fkav!v`hR=0vx*Y4^8}O=X
z@UfnXa)YcVZcmC{WtPmkZgF#r)0$&5q8EjH^;%%}M<4fyo%d=|d(K-o$hT(99>;`x
zmI;YlJ0-Neb9njc*De8fa^@5?1kUvwynEN(&qsdk;aanBv*$yNKMSYr9~B%iC@g7q
z#}iXNU36Z(H2cQnK@kBb4!zrSXyC9*GYnrf4?OmyN87D6VOK65@jpF$mRZl6zSEzd
zw0B&$;mgY7RRKpD{T!+vhJPtCsx=*LGw*EFoZACFjxGPA=uW3?S|?|xy_tK$Yhm(@
zifJvEmyYo|Q$l_&c}bN`4qsaEZtAP6N48C!k{vb2Fy!t4#i*nYUyBo7%nriWt>XL)
z>N;=4)8{`qPrq@lTdqusY}3}!`+e<%oy8l21Kh_PpEuRx`s(WEGae<B%=yW;&3Er_
zO^s)Kj3eI{&vT7g8#MB#;YzK>JZH^)Va<UC%M$&k^_i75N?FpgNPTmk$c@Q&d)&^6
z3eG)x&Dqkx>sIN=PH!q}?B<<1T6g8xmWk0OPJi?&44LV9b4W__MVDVs>P|2;xZ_%p
zy4+{_xg)iOZ>NtgojZSnPs8DTU$jo``>@w;(9!{2A8gYv?!Rc5^_|w;_O3&_On5J^
zKT$NV;<uJN=DE|ySDt-0KQ4OMwX^*X)qQ{La>L$EJ{7;uxoGVB-L&&*9sd}Rk<ci+
z_+fDQy;Ze!Ava8(_sP7MXI`5$;;!NqUYg{1VNk~05k4tSuZ}p>mZd!NJ+e3A<%Awd
zF5Nx9i@E<>ua_$lox1FeUZ5XeHSE;HE|Z4^#O6%(v%D4c_F<&5V0-ewo&7WSoIf_J
zacK6aQ9r&M@yvH`!oEl54Z*i-E1VNj3q#*dxOZ+>d_#H8wSj3-HT9k1Q-9gAt!s*@
zTlv{{=eK&Ed9!6y=mg725py<2E!{oy@!psv{mv|5uPt7^bh&|p=ESF~MwV1{{9(d^
z(f-qdUa#iLX0$}<a{ioUVB6kz1^<aQtJYcXUbeDzV51)QsehPZltbbJdEEMNdoR<J
zljlFMlb$SX*3o*N^(?|QvCCVNwoC7OkMg&e(lm2TdDQ!Lp$U1ro40<Py+|3JG-O?Q
z&))m)miM3Q-hXGdNA{lJ{6zV^WUC58z4Td!Ul^v(I`X3Fz|77emE#RVbpF_$9$a^R
zL}HhI%FeAT1JV-dGw!d9C%U$B6R!+<xlA#r3ySvIcPGMQ^Q`#rz)mwFmnJp7e0R<0
zV4cC`ud;ES>*E_Hm>;eh*w)u(Xna_<!Kz!@@%eX?dhT2C{Mndjo3Gw^4MQL2#Z4Jq
zere#PVUCHFbo8)}R<m;peKNUTy_MYNvb@rfM@QCwKYiP@oqw9yP4kv>my3<Ij|<6r
z=ex3|-@x~l-J}1Sc($nP{j<e?rc^Jub9sJ7hq})iZ4>|UUvQwYA^yj<!i-_@FXA3f
zUfVn|;L`@CCL`#H(Xg+()E3RG9({i0gXw`$Hkr%(#@BXi2@P!vsn@7hSL9KVy${bU
zH>)1q(0VR><`ldu=y&5Z<=z9QH^<tR91f@){I=$m!NwK3F(n`Oc;6hqzj*AcVD7>#
zo9km{J7*s2<^SogS0n7UXRZIeaa;4Hktd__9Q7iP#p=%*J@(W5fN{z)qx%(_HaX2t
zwIlkuFaE^-H8j0#%&YjVS#NC<=|7VL8^hg^(X6G@bpmJ3JZpQ`{os@y{YR%A(m$(A
z+j}`FW%9NQA0CYEUgD==Zu;O}#`o1vH*$W4%U`_Mve4?wj9mW)x=GV}@voZOmULdc
zXyN?+t@0zTUi}zu;4XT9iQ$spXS}3+GegE+-xa>mdvRjx&|?t^GedhC|M82pPKU7@
zvA)q`#v$j2h~_x;&zgH5C2Ok9TQYO}K}|I^b!N$oegm?5>LLd}HLi+LRxu_5QInB)
zMgC}nlm{%^9_(O%3<*o6x@o1lxj=wLGc{+7m-nX*3N+MTjk`+eU!^<*0xViHo{hZM
z5dK_&oO`Kphj8aZxQjr5MY(rkg7xG3f3Rxb_fX>;WsaS{{Yy=QLTV#@)j4(D4s%;w
zJj+ZCkU>v1?lQ8xj5-Pg-l32Dw_aO3K&e0xWG2@r<u%GwAi!cYz9g(gtJg9m>P*&C
z!f2E*<^pMFIhKBPD83!0K;6k(GN_Ijl&e61#q9pgA<buJ%M{3*%w-c+Zo;|(0T#;{
zvcqxj<u?_m8+l6sr7WPV1=7y4&&VRG$@r)y(js#yQ<s;i>k0%|Y=%6U-EDg8IwjKR
zt;X#@&O1<mK!C;JYRYbH+p94KNT1B*60*ER<iZnRaSG9?)E+i%lOd83mULvEjyQn;
zi}TO}EyH_veWr<Y$y^?&%O9v43IteO2cLX<*ec6ffvm|}He=;xvS^;Pvuy5Np3?fU
zS&4=dma9m16)6P*EN;Ckzox5I*eQ@BQH6t8c@P^51X$dCbceS%E=f}$Q^Ilp8y~=?
z0s)q;%J)mcHkfQzqK;&`ag22w<0TMa@urG{t_{oeRG{yO=*}VKIb<XdVDZsTpVKzY
zoy=tj*`rI8@)Bh(5Mc4$9b!`X+kG7dbs!(Qh%_%EUx5IN-|&UOk!wbUYa&K97c7q>
zE8+>T^xQkCt?_P`JxU}ea;ZhKT4W*+VCgmVT5(F&>LCi`O{RMmDbFHvfdEV2;jNE%
z&&$YT5KBy?K)pkOdMAOXX<&6?t;9@1M|*6-hX<u8kq((nE~A;tm<a^l6Xud6d))lR
zkynRoWS&jbfhWK+D1F7M-m4GZQlMc(GB+vZP0Cmxz%ta_{w6#2gSH{EBNN_$%r~GO
z0s)p`*xNib;_Q4SQj)psVU&9q3xNR3h%50uhD4gU6UmUbBqQ5o)KwtBGQs%uytUac
zn-$1~s6!%FCKCDbq@Bg(xK6HRRG<QNAuM~a{vM(+JOP%;gSFcxJi0PM`OVkWFd8+C
zF2|F0mhla1qo!UrV32}LcNf;&g{=hwEK|curv)vo$|LqjHu42>enIS!C%`h}$-Dm9
z`*tlMeu}*1p1SEh;xTywEHgW7YI-)W=?#Nu!jgnFldzsZfMvE~d;XZuk<XM!o3Pwq
zG;c800s)o<x4hQmJMJNS<VE%<hjPuKdI|(s7VUl!p|M4`PKn%zboV3c{m4Tgz_RjA
z>6LYhf3eX;R^&sg809LWFFXO3HGQLOCzTfD5D!CGD%4FY$clIZEL)1E_8V&S#z7P5
z5tbcTz609|1X#AoQU-E?Z#FAXAlah??2>@J1k%pZ>(D3N(11ls)G0uXE2R{rl#W1v
z<>1p+&!5hypVLM5<U<!3<weF;Ai#3S^0;oURccQqGA1mqX^q#kxj@=kewaTqBGhS<
z0u3gj%cqq2l%+s`B~fo~)ZN@S3~wX_$h?3U2~U9KaDH2B=ydfry5Bx@lUClOZ8@H_
zvy}W*YF*(mO!;lP8H`2-2{$|emSc^BcNh3>@X$mwS#$}dD4`4m0xTz;ZdJd|xIR>Y
ze8@JYBAZmyO(5+ol}Q0R*gq~PktJF5T83K77zqSePEDNmX6yd!!G?$?i_XKcJggK*
zJ4>oZr{ybh&hq~I8j@Wj{+lQ5EQPPyV&dN(RUmirmMfIy70Ov4z;b5Lq{BZh{{1wG
zj^F&Z@&E}G2e68?vn-e&l~ZeRLxH*ymc3YiFELo20L$4h&y91OA}{b8P=d@$$OiBP
zSkBF~_50(z!AS-3By&lpEYrzccmgcvBcGl34Sp1=M6QJ81lBr%O#}ig=~JJWmuE(n
z@Xomcxm6IG<_WOmGFc-GuiM!ZJ0fNujjf}xhd_Yk>H!oP@;-GJKiy1Zo=K+56JRN5
z+3;DblXr>%vLSnPN8RF%x`RM~<p#Ey+sCH+3kDexmSRe?n9>ypu>AIOGP|eh#s~#6
zBXik`m0PjCK!D};y37eJi)@n>D2PaRA6D!mEIa|0(yrs``ZUiRs6bAH<sddah;0M{
zEM>*3M(=$W@=$?1i0HB@*KFcncmgbS7M~i663{jUG9h!>f|XmakwDs6l84!cIP6MM
zB2B{b1Q|X-wgLf`$5TfRYq0g3q(CFchptn~>m=Cl1X%ux`~FnSz`CDwkp+><TUz;+
z*fdYtS)Q)X?05Y~H{N+%r4?6c9gZi!^5N^PMOOlTE7wGk#ExQ-A{JQ)1X$X7#cyY8
zOU5(EmV77%X~rN&fdI?r%+Gck(%p!3t-q<a{sC-A)LTV><;$Bs#`ZIA>nhM7vZ5@?
zC5s9a2(WxzHtV!y<+edYE@Uo8k>ODy7oGr%+T2N#=heBBxbAwZJNiD$O`YZLG3-6Y
z6tY-I;ML_q+72lKr0ryY&)M4trfX`}6*q*CS2=cd9H2_^Re$6uO~3i?{qhvEoRgzd
zm-b|w`$~2CO1W+m95c)1GlBUG2T2B`{)pvDF<Xi|gO8C51pq`kEsu43f85!nT9YM#
z!>U5P6T}p<B6X{K>YX5JvShwimFk^f0W24f9OIEYB*HAW2ifmIE>i5j2L(y-kUeOG
zh%X`AOUMzjWR{c+wwIA3R1+*$LkHK;;o!?z?jargP{dkOOns)EGS5|~xHAlUhEYIH
z10@ZSFkjD17P1GeizjkA)3dX5SdtCNHXtQ0VKp_DB<`{uQoJ9@_Dk_mBs(g_$w-zg
z#py_vF2$KhmMO*ANR}<dxk#2P#d%1UC&dLwRv^VCNLC`n6-ZVg#gCBekrY2cvL{mf
z0?A%TaWs}i<M#Dy8{ZA`xbtW_5oJ6Ijz|4qkCbymE0;N$EFwDj40}Jr0iZ>ohDgpA
zgmw!zY)kL+Gs`tl`VCaq@CUb#{Z&zR0QM4-X2Q3!5R!J5D?{OBXaIx~ma9jj>(N+9
z$Dv-NSuO{&IoJ@Q1<Ms<wip}2L|86{@`#~&h<GEV+(?-MmJ3xTsC&D8%7xkz%+E-u
zEs4Pt;m{&Xx>I~mmJ1bT`y2A?&OYh0D1gZ1F1ERg9f8rf@|>*g_$5ybip!^s@@eD0
z1iwddZ)n>$v>l{B6qm)=W-)e$g%pJ1@)*55MjvdKJkNB@XRM{zH=pS);zCBRkkN-s
zl;UnN9d9w#B7Ve}JYq~Cu{P26FYJ9Rw}33o$HK=2_KxKy7)`JTWb*83@FczNTgb9{
zIQ19=e+~6CaX*DfQHYd~eX-nY<oz1?Lf$PGs%+R{sItjN&r|Gq$^yhB7iw*gr=?J9
z6IJ31ZP*c(<XR>rlr^MbR*F^KV`atlGL}olW~sQds1x<rsvf(*7g+8+wtbI1L>x<5
z#Zr>-ZIsnEDc((4?Uv$d%Bq@j5q)btWmZph25j}*;XfJ+>r@)hNv)F{yoh!B+^LN#
z{i9_RnV(@QHi7w3Ts=0e$L$iJ$owqdV{4cn#l=#lv7&@1ZX0E~O^SC@rn{xMnli1X
zEMPw5{gh!nWdhhH;KG$=-%|O%P7%Sa5YNmvOdzI144~vfe+0Bh<tJRqkek#`1k|HI
z@DnWe5;?p?j-Z9mPf<rg7X@AfT@;yI=$(jKb{Q)#V{_OnmivnJzv7OtIV3hv`a38I
zK0_(bQ0C%zC8b|Ubp+wFTot8PMf9sZFgH-$8mK^UiO{S0W>5x&$VfD(Cgj~DH7LG=
z11oTy*yd+}R^(9>S4(-+Ql9Vyg6YUw`UeP<6jw(_)=6;#?b$$gg>hG*6+~pj_krLG
zZidBk`lX)_#Hu&i*U}E~gzAm<byD0w+ceO&Fb<s~vafvG2qLcApot-#rxtBM6p>0$
z71dXqa}gb0L=S*BKx2u#f$u888@fC_aK>zG!B1pPliVh`gIl7wx5(=)>I(Z#ajnR)
z6*&R+41K(wF^Tl1DXtmWHzNlaptu&~)gpcHJ#u_6KB!u{cP;G$6H_hSyH3Q=-Xas_
zdt0DKU3=!s?=KmtPyAgjL*+7Z@YXD84p283Ie6!8jowu=3Y#K`_i&&b48VIpQ;g`v
zR>W>adeFsD<yj%|XdoDmTxgYn?Le!HVYxKiFAWb7H8}@6<zR0Smt(dZ8;hEI7dzd>
z-XgBT{i^T~z{D26_1(Zx`*s?yHcorYf^S)ZcRcce8Tkgq2c7?Dt^Ggl5vZL)F0}1H
zl<fa!yR1;H!2>Lpg><t}02Fqt&>|DZa}b+@x<faO><BW?MI0zTOZ49ODhd{H9x~5E
z+$o{PW=UbKdkqDMnC~u#@8^qM5tkxGDe@DYR2(*o!=ccA>|yz)XgV|QjSj^fVAum<
zrb=-K8TOzQA7V@nF=mj-Qd}Zqk|@Q88I!|Oe1tJMBE?4;y`zkQ_<Rzhmn6l<7^7p1
zi8y|oF*+{ACm6F6jD<LUk}*3e#mS6EGUF+ZpJF^tNpT9}n!>n?<EI(d(^8zuu&LsD
zK|l;Krf<zzQ3H11;2n4<G_P1L1G{A49-=C*raV?ty+Bo2?i}TEPMQTaQQexTKHzly
zeX~y17B2fhjPn8Zcz`{{2WxS^S{w=(;tDZuz84I83{7BSo?8&R1sO}t(>EUV1!Is4
ztzj_FzC+B8E`KNMOUykUDdSN`V1%`jyhmslw<pnM>IP-%UXU2|E!_3zcOPxNS(e*Q
zyKbj@ffHr9INCK%isNb5cq!gZyY80a!?gKfT2eC}q0Ns-@d?`egcP5o%}+}4IojzQ
z?I(Uejdn_t;!CvCB`Lm4J6#rWzs$9-W;9MdNOs{NHh755K<XrurVQ><X7J*WJNx_+
zMh4F#jC+vY9%KPBW4V=BZzZ+>Zw@||$dK=+gHToFxbr>657K>b1tjM2?R41Jusu0j
zN>(>&kxyk)`q`8T^j1jWLFwmECeX_bADCJF`eAJkvX2kcO&+KRK}IoPkdM*L;N64C
zz<N4pJsk|IfKEMe%*mKd#`>c6reN(9Y$W0g+&=?<C*l&!mSBBRJgcdQ)zk>Ua-rW3
zj(Mm?aMsL8%FSd3dl-#9j5gRG=qOn=e+mFf`Q1QWhlXwbv@T29jD~qwlFeMhhS#Jx
z9~<UNaRD|gkmBpu@VXRdQCeA)9<*6hQq{_)^rTYN%Axc?sw`<0*}SDYLl=qVT4<XV
zDSk)Wyp!Vhw9R`dZl!HnrFbo4v6ittDKuSKE}F53mf{%3B1VeWF&68jIF_-96){MX
z_yhis1QhK1aXTYZ6023}kjChxF^0ei$0p<n{>%hqKH}5vE>AO7cl}3^{ljFp!6P$L
zmE=_o*t<b2vZMSH^bXSvv;N1pI%;p@%MfMDVWM}H$fgo?gElJ5-A6X}rT8_nd5yY(
zHL+Y1vT2gyG~782`-ys4g*#VaKhVQbm%7i0vbVF<V961QMG|%dg(Lx%Hoi*RK><u^
zMcVc??G6Pn%dKJb*DxK$UCLlAG8jiFh*c?+O)ldG=B-MhI^STrLJCDHL&l<naRloo
zts+LRhOvTzcg*5_KaI`G{DT<vX1d#E+8<1k<(|@Ap3=UM-Hrd*RpY^q$2vsx_LRLb
z<YI8HMfQ+C&4O%nf^eFpQipFxSuikhTWs`F$CZWubLtFJCU4}Ayg<R=&7^h0pLf9k
z%k85)_EG&{7f3cl>1R<+U~;gV<c<9C7r46VL&mt3H;;M!O)8d~=}uCqm_DVgK`Jxi
z#yj+_4_ibS%du5Cb{0QXL~9q(X6=l5#~l;)xBT#rj1YQ;WDflK8cZHEgyeku;TjC6
z9FgkS0~l1rzikQS48=c7j&Gc9(w^{Xmh@D7-p~P}N`0oaKhs8{BiYP!+{|<qRqP^T
zcad>{Qz{}grb`ynT~xU|#vzY!7ges1aVTWmMU^XOT*{fQqRLe<rd5m$^m$qC5##WP
zafkhc=ZFOOBSKi{Y~R;&hTfTPL{x5+<0#)QgLc*T-%_iXOO4AtIBwNtBJ=y$<Nn_>
z6xUkp4tp>!xjb|Gy0I%rB2td~mW!#2DyVusl&147*rNq^g$Kdqk}o{L$|qulD;JI?
zLG%m5`xZNo@KzNlcWCQ7v=_)6&Me8A_ybFDZHuqfKVC|e{nmjc$A(%7SQkzdS<->k
zI)rthB}=;0*d`x4z?@jF5gRn(_HG^VK-i#3ir>+0?`SV^S7R8*7{(2Db?K|q`}H--
z9+Nda#@3IqtN7w4*!qcxm#sV&<##4>En#fIomxb3vfL-^@d>y050?-13aB{Sy&GBl
z8AkgI;{<+vg=xcCYq^yp8Tdl!f1!GS^}sPLkqUn}3p>4P#pN|u>*}w3lVi6#v_Hu4
z*YqLI>uWCkOa>g>9fH8?td94!C|3XZ7xLG!bWkiE4qsT^^5O3Oi;w<81`5>O3)Dm4
z3(%@3pXLvJVPHcTUh{IxSKodt=j`t64OKJCxwyMXkgL0^1i87pNsy1+M@M|lU+ynK
zf$~5J3X%s&P#<|835t+MNYHp{yr%fh3Dg7$nk1hj;k=`6a7W!7j^$WxBkHjc1wuZf
z!h<%V??s%9%#x7}#9o%<;Igww8<KWX9w6B{qzyGTks}H#K!e5B%XJiXU5am@up3f*
z6NTNB;@imcHtH@Ge78~1ZN!PV0@+s}S26HbpxzZIM2f>J&=3(nLiUf4D^wpOD?++&
zkqNkR($GZgd!#4gR>ZbSaT{XWr1&#pKTGi!#D0<DO<1-GYeRFFgh?#hjJ2URMQR)@
z+k&;>5RtSbuxu;Vh67HPG-+f9ur?egv!v}KJBYQR52P}Cy+hbkH2Xv>OT^j`S6On-
zYkC4(i-|=V)=d)|rYv^>>t2xJi&*!f6qjR<a_l2o`#sFw!+IjF!n#%1M8prU?gJ^V
z#=6x~{0tjD6C1RuM9i>3il1Y{=Th8+&6=={`2I_j>=LC7tr%6-Ynn}2i&<|rWt&Yo
zi8zNc&7rKpC97%;+Z@VC#Mda<HA)*!L5SO=Wcif#A>nca(O*hdKxspdoh4_ovg?#K
z9M_UyNy%<f+E9UzBMrJ^k=VZ@B@C^6KpTNWB0WpSv5N5!gINvJxke0T8~*HjZ|dZ^
zZU5)pw2iB}?>#l=-H*hxJ*EAg(!F6u8>^@JuKv8Nl=$oCRJZ5UK&a$da@OH^iwT1G
zy0Q7n)z^Kjp8VGi<EDnRIYR?qsLnV3p2R?pfz*)v?HF*PB*mb*y%9gXrD}<<mfD4v
z<P8pn4zA#~A=VKm%wL^>%njT(8Q?F>z!idR<qb`%GM1hBN7}R_F>w3XPjl-@Y`=~z
zuH#Okg0<3)t#lVL4#qNev5X7E!JYL^xxKTG{z;~mz!)bmb`Z68z8^d4S8ekOQVWz)
zuBB9Wn3M{Kmr}z;oJSkyNmKSb+B=U95>xn6#<rAk7xSM|Ca{zl0C-nn=<z$zKf90(
z+sx>1W~^XpEVq-<-^o}(<8<GC=iis-E%5nA83dUb5f^`X2=)O|JTkyv8EPL8%z-Q}
z3$t0+RP-k~*dzx#h<GifTuWJs;*6mzW26T@8!5|;B9;p`jvym~I6>;@zjp!;|GZ&K
zvBD?#A3F(pOy<mAJAyewzDWl7%SJG8v`@7~P~OLc|0+Pph)H3<Ur7SBQf;Ymsq`cw
zpLWhC$&+vl33ru<xHcnpGqMuH=PqQn3%QA1mP1JI5VD7SkL8Xby<^B8E=I6i3ernK
z_7HC6!i^_L2Vv{U8u{x^uttzS8Tfm)51m#bVE#@M)GVijJ559?MOaZJwv6P$%_i8Q
z(_TFy3igkUCyd3Ec`@Z6N+pibj}!F|ViuVLfAI;-^^9=wi45@9onQc>I+@(x`@V3=
z>D%O-i?A=OPcB?_g2|;-n;Oi?zdnMPX9XjxV2r?OKtahj`3q3+#Kjj&FB+|HPX5;&
z;r+)uZTGUW#!#a135;HX*!Q?Rxvsdcyr^EGx?G{N0hxh+R9&vn*&)UIk<NZ8PDeWF
zQk;o&GNm{h>10cBF4D=B;yk32C&dLwr$CBJkWPsdS0J4VDSm`>9!c>Fr1L_Gqp?mj
zZeI^1ZDh~*t7xFB8UEHfkJr81M7(zt6{Qdp-B(!oN?Ox2V`a05GjCQzp7#6e$-i!^
z;c6XOD1WaG77FJs#31-<bTE+T@N*ZR3t>tBaEI4UM9k9aCtLGho!hl5nI9*e`cDmE
zti~SIe=mc_)Z%eqV)^eT4@;e!dx)4+zPeq$dI0zss2Q~6!UaAUP}Qa#6R;D+YgLkJ
zehAx(NoqdsmM?bSRJCd6M(GS*Var$IMJiRTYS)C_U`8z0j4hj`<1M&H3+^k9Kc@yg
zry|8n<{cgQjt&=NdJN+q!}J61smgTv++qemrUMR*oRRal55<M8VjNa6ZcvD-B5uc>
zi~~em6?WXocmTdpUUuu&pKnzcr~ZE1=f?BBEB7~emnkUj5W^l4&uLWWP;8<Uljh=K
zDL%rmN2K^D!yc95B!*3r;$sYZOp1>)>~SeR!LTQ!_$0%gl;UKDO_t(Q40}q7Qy4Zy
zicd4_X%RymLvE53BG*FH1p++FeMCMVk)K#-eMA8tQ6DJ||A+>OxD9!<iKodd_ZfM7
zmf}@7WEBnvez|ZR6b$CJzy0zdgUomR>uewHsS?}gZ>K^qgxjfPfWMmBJ}~@8<JN`y
zt%yN(qq-@;`$Ne?4EixvK9+W!_zSBL@yhNk81CHhx~lsz%5+p`=!=y-&HE`m;O;X`
ziuA=??u#8Aik#%RJdomQ>{2bowb-#1JAwI;gMA!Si~C6N54AW-#LuwjGwdbe2JG1&
z#m}+lb181Y)-7TUNRdX8%O~;s_t#b$GYd@b{nx_e{@Z;+rtV%J^AB^ZGTBzX|E%^+
zB9VK{kb6uNWDQW3Q!4c}FGRMY`aVDw50D*%DB_ZkS2OA%svmz57}T#$xB*NylfV88
z1NFYQ`@ByvR{z)O<zwR{m#t5=-H1AUqVzve#$x@_M(MYS7)oK%l1#(?X}B+>TB`mk
zdtZ9ER)zhmq;9$v>(z>hrb=(!9!uM*sg!=Ic>hN(+=m7|eq84}BPt&s{GU$(o_Nr2
zJ7^niAWK`rXsluM#6S?oXv8sk;J=@_TsiG!aq2M9i&o00RlEb-yzfem`&nzdf8DPA
zrF`DFI(AUk|MUI<#7?5X{FQTviBR*90sf*nM9+6CuK#TK>8kes`BdITbC>Ok(pvt<
z!@Aaii@yKC=BFmIna{EJb20sZ8}VcX{B?M!u0DM_mg8Qt>n#ahv9v=h-CHz)^|Zr!
zDUPE9;^-0L_-;C2w-g_t-Hy;9;`mA0?W7c^(f(=lFme1c?SEOsFd<Ti^0)Cp_^Jv(
z_1~NU1E~P8{SMq0IxTQbpDc~PoDVSsj<AT)Uch!2aCgyYFJilkQhXP)cf}i&sxvX=
zK1$LrNu?}Psdi^hZj4fXsZ<0wce!xoAB+|%TQVd5qCZT1?xY#x`jS}IH%0Sz{nx(?
zR|my{vrjd61vKf^KnhTis)6dwn>%6N7FApfhTy2C7B2qmC;ki-e};%Z!`go$Znodz
tp~}I>uV@Y!*gv>mSV(AipT26Ui3)!P!VmJlx)6W1|3?LCYW>B({eOqW*?Ry0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp.47280D4E92EC5D1F.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp.47280D4E92EC5D1F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1fda3304ad1a5ceb010ef04490eb18c95373a366
GIT binary patch
literal 5776
zcmZvgdsGuw9>+725JQ6UP_Vr5LFRGifrM9xf)7+!)L0P+pd|qm7v(@fDcVX6VbRtq
zh605y;FDGwtXiwq7Ar+v@vMtg=~huG*4CDy6)mmet2?A;;P=Po9DbSa{eF|r%)K)=
z_p&5z-n=ay63LvbxZ>OuYfUnVMB*y^6|OD#`*W#8a#SUe6xW|kO3Aw&=5l9e-<6CD
z5ph>q%+sev6>a;EDeFkb-kI-K#cY|J=oj^*Usu%<@#>NZ3q8`>c61DN)a5OzlekF!
z`BQf=cHzOpJssgYD_%06f1DO}@?`7hjVqIn+`4A&eRj6wSVTlnTD_0IqH~e6j8D9f
z=U>@d6nyf@1BZ#xU+RyXeSVYqV%Y`0ob34I;BRw}efH%WwKs2mm-SvaxnA`+((kpl
z`kG@kpAIevI+*n7;njV?D{Ff;{74;NmpLOXr!MibXLDaL&F&9gHh<E0um1Tw{Ag<A
zonx<$eT|U$)&2MJnJY&_u;j~%YRk&_=8fHv*~>Q^+;_cvN!{e!u&+!H{R;bYYB*{y
zUff)`%v%+IgBmZn7<aomw4wIRCxb&iZioN4Y?gUZwT^G!4_N*64@ds(aeT?9bX!}z
zOK|qf_q!C=BF?AGG{k;fmx<jU@?AAWj(KIT_HOJtKeoW>qnhH=V<xUY)cMF|!{(l@
zU-F%!x^(Y0*PKR`|9GqIozUyu^PC%cX6&i9F4k-wx~lrKva|hd?b)*Lw-nq6PCiq$
z-FJ8D`U>;rWpQl4+0<zV2ENC(b8_!v+isj6yP<qiTv>13)iBCwmojtOu7(!V?7QFg
z%;~K9)nU<wzNGzej#Ha<l=qum8fHa1bk+Lps|k+uDca8;IQ{JOHA<#xPE4b^H@Y7`
zUg@=P#og|<_qFe~_OE(Wn|-2e#f8GYfDK6<ep~$B`k-LtytS_P_YQp@vt!-f!x@pc
zQa@29O<EVPyxe(kg678hq3o*iOvjXOa;}}&G`BysHLzl9@6ua_1hxC6U7CBZ@8mE2
z+K|<{pl0#hU)s-Ycg2`zZ%p$)evyhw_^3!T{^-6}EKSE+XYbE;j{g^GSv2*-!~^YT
zU9tFGj)AGOs%y30zI$e$P<fq<pVND{Dx;#dbew+7e5p@!?WwI*$D<W%I-cH1iB4<1
z9rAnYy1Lw~Z#UNkM+G?ru<ysJvZoXY7x8&#!O^dpnqM6G#<_d%?2jkBcTJT)W$=o5
zO`Y?_;B`GG`)S*Pr2C=H#aCS3c8_cLG4xqfgs0D?rr=MPj{Rxh^QvumhQRErb6+f6
zt*>cZo3Xt_e$X74X*+*LZ!kJHmk}@C&&_qpDqr~6A+hI1Nt*HWho-1#LUs4iW;fO&
z@Wk=e4;un%Htu+fX!Y7QE<xk9E@%GU7dKYL^sGLwO^NaOBc&}MT2p_!D5g&r-BOkO
z*t@E>w*o~Do;uLcncn@R&-?uKTVJ=IYWt}Ds|N=Y|4<d*4Z7)ZXZ~H^XSTSa*Dpmo
zmuEY?@OtjK=EcPmT?u*T65>>y51y_*`W{w(F5=LdRi|Wn?CHJ7$~PxW+n$aOZt9&d
z<xT0&FVCoy6t2(BSc^#{a{te^&)FRwpu&?=B_?;dyQ94b^Mw7Se6p*@5+9nN!jlBs
zSMKX-FQPqFvuL-+;sYrvJWa3z<N=BHBHGbUAKcr2XZ=AHZV+rwx#t*r5p9dnYo2YI
zrbvY^5^OiQo6=rH`<?O%@2-I`T7@SHwyWG#VK1WHoL-tbX0XrbBuNo$je{o6UPSxr
zo%K}@OU_JG;Y$SDNA5GuUPSvBQ^4nJ$UTh;Uo6-jat}v)5$)e-hwc-7%dV>Mmj&Bf
z?(JePqCMDA%M7?~d^A?FRIvR#{HEHAXg_yJ`s8NbnL8>xL9oZm$2;1KXcw$lS?J|5
z+V_){_CDX%{&ZK%EnR?<NsZK;*tLSvsKbDYRmYDi0udldXb6xbVhNBW5ebr{h6G76
zmIO%(Q6NcaD3GLLDUhTQ4U)8m21z=W21$evNFogaNfe7fl0gheG8zUXnOFuSS;T@Q
zt6@Qsjb%ZSLmWtQ8V)46SPmq4#DgTS;X#s*<v}t;8!~<57Bq!wLxBp{hL0%m2@(c*
z`0)XQT!yJDu+gDZTQPM#)Qy<B8EB#mBVeAeVg$?+jTiy*WEn=nJZZ&9m?s-C66UEg
zjDmT}icv66HDVOZ(`6V9^RyMCVV-WpXqZQ37=n3Z#SqM+Mhw9`Q-(1x&sZ@A=9xx}
zfqAwJV_}}PVl2$FjTj5_Tp7l}JZHr?nCBWX4(9nXjE8yNit#YdH)1@@PuEV58NESf
zXlDQwrHvX<;vIl1?Q`eo4KiCXJ6Ze!aVDLww9lZUA+-fpm*L`f2$MxMuC9T)30Hpr
z^#HDZ0JT}FE|r4MTczqss7qyPi)_><EI0vqL^V!89?^sokVg#Q1mqEBDFJzeRZ2h}
zQ7R)KkF?+<<dM}l33+4_PC_0zfRm6%nx!P<kya@Qd1R@KggnZEQ;<hh;}qmkO*jR4
z)BsLF9%YtNkVjdi6y#B*G79o&3r<5GU5(R_M>pX#<k15-4SBR#N<$uPmC}$$m&$0!
zLlzuC9;(I><e?@UK^_{w5#%AW6hR)cN)hCtQW=6g#)31D$5i7C<S|V+19{8<&Ojbx
zmNJmXSfvc)F{LsF@>mPbLLOU<vyjI&;Vk5_12_wLtXaxJ9&44dkjIwFSjginI0t!L
zHO@gE*MxJB#|_{d<Z)&x2YH-T%0V7iD&ruJx8OYF@zppFd3+PjLmoeX^N`1zr99;E
zRw)m8e5s5d<;~iBS28rXwi}a~;A|9={P>63!Y|sD!hEAptHJD`Ab=tjBv6!s0*Y49
zKp_PJicv5?u?iL_PQd}iEBFz$FuXF)gMDWQ=4A4X_e~Oicb6`5BYw6{cMzIG97CMN
z2{paAKF@fv^|sLL=IZv6xD;KnaMnrl;SVx|X4kQ<<HaRtQGCn=WC&X<H2cl<OBR=Z
z;-@D5wQvQqvR<7q+AHc3bt(KI*Au`bN(g9@dJ>pq2?<R~PXUuEp`c0YX<*VNG&GSO
z0TY!VXfk>Rm`n)+O;*nWlPzJP$>}*@awQxzc|8wIzJ!Nngf?P^I2KMOomMxZEWZzR
zH5~0ZHauVfArWF|>zKCPTiOdY;Z7!R<Q*g~(49;^$R|i#?2F^&q_l4pz<?OksiOlz
z7zm(90|^vmpn#$cG*HNZfMN^`P^^IkiZgIP@dh3!tyZfOuS$H!l=jVwc*3@R%ledC
z_1lF$2v{DFA)a|l@fQw@bN9>>nmxyPdWegHFGH4OIn3SDFEqOcy0hYzu6=1#U%LOd
z=Y-~&vYC@bl-O579}Vw{M5OR2>(p~Zl9P$h5kQeT5-3VX0Y&R*ppXs$#poEISRD%#
zr{jR)bv#gE+OS#T3HRRKG}G#O=(=znekp#L;=}i2ki*~@=ZRZ|<`DM~e=#j??1Ua2
z-qneh*RN=)z#|V0&ki1e9zkMi<Tg^;_jh7i=!|^z2GAM#1W;r?2^5u20Y&H2K%smD
z6qC;Y#pbg>arqojd_E6Uq&9MnczNR0tCaQ~Cpcjnral2@85>4GP1-OLYRZODP}4Sy
zh8o#01T|yB7^qnr#zM{6Fb-<ohVi3X+!4Dg?K@#`+D3KeXg{lS)VZUIFcQEcatPp&
zMiO{r4hcNUNCA(^p@2slY2eX0H1LoS0T1OM;4wx9cuWohJl4npkIi9$#~C@`aXB3D
zcq0!yK8GLi3RbK!4ga^GupqNAVfbMAaQW+%E*?lEuZ$ip8#$Dk(#eUH<@Zc-dB(Ez
z<*yhe#t~1FF?yJpo3&<@Amdjql*%0BV-$`mr?D>1<KkTB&QZH6<7T@}96#YD&j_t<
z`i!XPnVdU{oDxR*|G!obZy(=)Nt34r5){qwLDMv$;eH`rY;cUb$Nax#;`8*gVzCAP
E1JQ?gvj6}9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp.1DAF74A63A4FD2D6.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp.1DAF74A63A4FD2D6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0a9211148c0d943a8d527c89cb75e958826ee91b
GIT binary patch
literal 6768
zcmZvgd010d-p6xK0z^O&6;wn--<rFxQ7}eD#09mc7L`cFA}H2O@ufv;VVbEiF4a~I
zf?I2>3RaBbR=46(TdcKpW-L}_tZk(_b!^?}*r__!S|;RuE`0xZc^=MlzxVS?J~`)}
z+}v}r#*Q96C&6IIm^HR^ZsGiVlfht!4gSxYU*h`BXfPa(G8jtzXR@a~^ZlT154K#c
zxe#FY3zJqvF!OgLNO7l^pZ)9=N0;5JH{XA6!;|~3`3H21-2J!uIoZ3qJa#{@eOI?5
z+CzO+)sTHX*<XA9T)~q^!<7#b7yfzg|9&Cw-S_@mpZ{ytGcWyGkoNZ7-W&h2y3<1w
zefBl)25!gEn6WRtR`$xh@1w-m-KWmYe)wwT9e2vmUKPoI{?PHl<vYLrmS6Jf!phCx
zm0uUH-Y)%?n)LkTgtC8*l1@DL^QU_o3hT`$@12|Hy8Y=N)k_=2hMtY?q;==(w5+p<
z-@V-OB=zt+_G>Bjj<1f1PFni+U$$Qy_|9;CKpi*yE!!Wq966t}p=td5O%vHkJN`O&
z(72|V`P0QO`xj0&_>OHod_bw$tWiHEChV)d`N_{0Yrd!{Uw(MQgQps1NKa2JiOR{a
zY?*%K(3!;F#)ajiuU@+RkGm~rKi$@C(bdHF;vJ@AKOEh<s+(z!f6v^~tT$f#M~}&?
z=AVw}Ubz4K!_H-6ni~JGFlJcew93Hx(|pByE0R|Z-1PP6n9WVStJbCOTw3kEvt;46
zgZsNVR^*Ny(SPnze&)Ed>nH9xQS(v8wuwKK7wlP*J^$qxy8Ggeo5GKlWW4)qW|zLD
zmLn4)jXS@b8aao4apb!{{k<up@7m_DPD79NcV*euWQ{z1fB5qA<y&lxYm=(hYl9OP
z?A%vd{oCo!RZ~=8$~5-!pW==ltw{K5{a5Q2t$WE*cIiU1Dn~9U6C&q#KFJ)2sG_R^
z{-vKEU;gld`{jN!uFToB(ec5E1D?XnD;3^D_xUp`7bLZ;ob%MW@+o`P?}?gxXz!V6
zKOZtX);R`D$%^f0yZGU&kIsE_GNbSASt}-Qm)v)Hs9Wph*%x1ae4^@waerGdyJCIg
zi^`>J$$abb>D<r4rqF9H?#jM)$Fr!gVac1d<>SuzSFLWyt+OVb9G`LZ`o>AVjTJGz
zhzMgsVC|`cYmcV&sJ;C7Mo!wa6O}{mpLpf%x%+RQe!6ELQ&{<i3D@5{xMuKB{=nXY
zBlfsl@hz8cMcloYxM5tu1Csm0uQ!LM$jj)r2j6YF-*3gO%{LYmC43s_S9kFJW#2`N
zS!^8UT;|-hrsD9H_wTKmdUeUM&rZ4Se*R^p<J^TWt`6K^R5bATFB|$Te`9j}hToqn
zDvB-|G2uT;CpUfj`le@3*W?cyT4=rX@QpWx^>N3K@((u0uK!c{iWf@Mgu6Af2i$60
zJ^D%Ah2<wNRm__@I^jXi*5ov6Zo`649@*1sbEO|scQPOMaUS|<;j}fIPo21bF_rn|
zT0{M*k9Jjkd~e^B2T`R(>XrB#WA5C2vU}`;*Dj>REPEmB_k<@si+=y2u5nWF*-2-{
zvo{~Vc=)X0bVkC|ltX#m8qbpjrRiyojoW3a?4vnv9ymq)IAdtVyt26k^C^QNyw3-#
zGq$BBM-k_9hWxnjxDKHvc$WzMMgF?AY4P}hlcGp=5cdx69UE$Noc8$M-JLhe_C=AY
zLEJOEXGEycaml9Ink#*WjP7Wd9K=cSNl%3u9WS1dFnYIYT@XbaLEI(0OJu0g@ycbs
zUXA}s;iAZtAdU@>HHR7<?;jKUj5y*}r)Wb?5T}Htj14t9ZrI{q`}6CcbdMrgL7W(#
z*eTTLxGDZf!;p8k&xj(^f;c%mc}l3!@vZ#iW8#3j)+jPDh~vZKJA@h?|C<Z@y6$G~
zS5f5IAnq03t8=K)@srNk?_VkY<VF;k6vSP_yLJdQIxZ=kJ@0YX_Ob6B85;9d^^dpK
zUb81h=UaG-q<_F*%CwAcPc%;hqcdq>49@^#G8tej&jMpJSzsK`0pl_`U_8$Q<1=|+
z0xtj)G6i5FF9H)YMPL#y0h2N%U@|WQlQU&t3a<cDG8JGduL4suRbU#g0n;)yV5zp$
zA#GDBf3R(EThiw~9~qkSaCy9xWfj$aNi06f@(yM{WvPKodnp?0X&*&nJ?*DxtY^Fw
zgY}G$Vz8d^Qw-L#UW&zf)<>~e&-y7A>p3sQVLj)gIIQRV6o>V^m*TOW_fb67^L~oQ
zdcjKxSTFb}0qX@nC1Abir9`Y3eUymxqMs76Uh+~B)=NH0!g|S1Nmws?DH-c!A0=bG
z?5AX`SG<&h^@@*DuwL;~3f8M$O2vBBN2yq^`Y9FbH7})Mz2>7dtk?XMhV^N-w2|#o
zbhvFeq;y+)Thi~wxsjo}Fiu!vSxjKURuaov%+<uQ9dk3W{DirMSUg5}z0YVV$Luj#
z7Ma>P?Ikq&=t@GPkFF*(`sijtqmOPOH2P?tkwzcwG12H_yo5m?Q%M-~G1Y`YAJa@2
z^f4`jK_BBYGU#JGCI)@1m$2w#D+!A}wwkc$W19(!KDLFh=wp3G7JaP8#G;S$5)OS_
zCE?J=RTB<<Tr=U&$F&d+eVot8p^x*JIP~#e!lRF`Bs}`~YQm$BZzeqY_!h#WkM|jQ
z^zj}Ok3PXm1oR1&L_nWVO$77_%|t+-&_V?C2|lBMKEY!W&?kC{h(58Bi0BimiHJV2
znTY5UTZo80(PtFVCwfdG`Xn!r&?i+A34KyEk<ceK6A67%3z5(#`HT|!B#%i#pX?<v
z`s7L?qff3TGWz6ZBBM`kAu{@8pHW7i>@mscQ@lh$pHfK_^eNRuL7&o06!a-AL_we8
zGb-p)JSGKws+XwfQ!9yzKDC;t=u?}Ciaxc4sOVFDMiqUk$E2c9^AZhxS|!oYr&SXT
zeOfco(5JN!4Skx=sG(2un6!3Z`Wtr)kA9i|HD$^lWgg|!@BAa3|GB<#-tqd5!6!nq
z#R7L*GYyF`Gmuy_3yCvxka#l>NiYkLM6(D<GE0zTvkXZwE09#P3Q03-ZD~SkMRB~i
zvYd*}?>)YEw*GK6+0ljmam|ph;OW#3sWEzI^Gf~2GftlPK6tuIY?mJTN}+H<`pHEH
zb`%6p$99bEs;?0Abw;Pi(E0*?_yUyWC_ae=C>k>ppcu?-fMPLo0gA)S2Phu15TFFi
zVt^7cO94v4EC(nVvl5^b%xZvAF>3)zYtIYDkNgWiHDzM3_erCYoO-YI!3i!STGvp}
zdxs^feVSO@77q@Eg9b>uX@rb}0m!%+gsg)F$hujCoPz_%xjBTqg9pgFd4z&P04TTx
zgrY+PD7r<2l0yP0xg~_MLk1|jWrT`D0jRhYgsMXYsJc~znnMGqxiy4#n|-K0lcV#8
z*@m^H#rGdIZa&;}q;<L{rw-DK-ir8W+w%Gn4~fq2#rIP5<z;k!BA=+}D_MQE$3%wK
z!El|OmYnwMOgm{vjFW-HI$21ZlY_)Nc}Rj&fFwFaNRm^6Bs*nDic^84I#o!TQ-d_b
zHpH&?fA{UhWjWXUtAoRmJUzKU?@Vp!@vw<=t40S;_w3X&UVpyTUqwZRR{Z)vd^jL$
zR@kVj+rjF%{&Awdvr#{J*v-5?m(B)HXP7ej=`YqtH8`5Bn*;ir1aGc(ONRadqVs7x
z4T-TckXSnliL-N%csmbCunUkxy9i0LOORx{3`wyokW{-0NwaH^Y&KiE-ZlNxlgQAX
zMtj%r=j7`5p5bd^_kR-{mZY4dnflG-pGw#-5i#B03Z71lOYNh-rM@2$9Ie*PkoLE<
zSS=~KFF0CO8WLk=AhA{!5@+Qg@m3y^U=<*VRuPh9l_1Gh8IodEAgNXrl4jK)4YCbt
zTjt+3*ZRod{jEPt@%`f!y{GzjFxcbP{T=9Wv84=pTueh^iWx|3F$;++<{<IKJS3r5
zfFu@+kfdS>l3Xl9Qi>HwYOxAQE7l+lwGGYCFQ&dd)hjZzEd^c6v$TEpf?ouer3gBj
zM+2i>G%zNQ0mir(U~C=>jCHZVxI7LR=i-3zc|0)Q#RC)a1Ym+o04C;%z(kh_Ov;mh
zNiGSPoF@a5T{18wPXVU56kuwe3QTpWz_dIKnC8;jn4zSwD8F?-a9+vGd6QZXrnff#
z_m`(17!1$1ALh0l8uPE#t;w18`nOzoL2lmk=VusZv@t_L`{B&Fvx?>f<M`PVjHa;g
z2y=(1=#HIZI*pATm0^jE96O>*_paS~^c-Zf4;eZvZMYQ25B{GNrqBO(HojM4@8rJy
ip6XAtoFFTzbzo}JfCNz+85cih+)Oh1ne<HR>Hh-g^K#1o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp.CD8E0FEB6A298EEB.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp.CD8E0FEB6A298EEB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6364304ca89adf5c7cb14ad9d1e118b92c05129f
GIT binary patch
literal 5944
zcmZvgc~}$I8pbn+5Fr61D!U+yNHw!BK!^&+7FR4*6z~elqNyS%>P1_860mBm7?ieN
zxuQsLL8;OzS`{~3xOzdfbt|oVwNfiuTw1l2JER@*{c-a=oXmTEFT8WknVB;mNu%TA
zX%Cq!CTnzIZuY_%a+yq~7JmvBE-)Zbw5nvX!cAwAQ}XVGx<1&_@on+;%9@TnSY-If
z#U-sXviAKaD7yU9*j0g({YJiiq`RyRnXtNgbiIAn$J`+8Z%(Cjyyy4$x1T$<45WrE
zIq-(fa%h`%tK)9Y()kTl9{;`cN`3DcxAag%d;PS^%>$VWZEHQ-;-=mFA>?f9hx31Z
zroorJoqF-gjmPUJOxaMbrr-H<@uBM7yF32udf8bqVXx_#_tV_aIi|5m*T<}H;!ZZU
zm$lE^W@(|1{&v3qyxr|f4_saU`8Qs*=OTs?tJrA?0k`tsJ`QO})jfFG$MN6}he-O1
zJI+Tm22@w09(JnYi)}AIUU=c~^BekueLrHC1TH=q-Q~wV8uq!%Pz!pQG^;+;eXl8L
zNXGS$(kWN>&wbnR?S{5+ud4$i?%qyS<`m!c&2rxP;ezXEsQt?{w?flLEmJR7dIro1
z(W+#sb3y@rLVL>;I?6Tw@!afFmheM%J|(*bZ+dz%&adv-2>DG@*k=*jzwA4C2YY2f
z;f$<#fjMU@in1-uU;JQK{n@&&XS#IF&v+YLdB*3#fc~rP&YQAG^HAT)=+C<RkeLf^
zI5G6<YJ+_rbACQ^y?$?YbJ%UK<%7MJZ>i6%Y%)%shs*chzcOyssrC;xG21>^Q9k{!
zvpxx}Ny=F9$mCjyMcVK2ACZxI{^9J))yL{xy3B2D`Ssxq@1L?iO5J^#nq(PTdc(Uh
z`@-NSsrx=?**EU)q87Fz|2e*4gzE^;it`)2F2y&RAMEpuPHtGS+ORS0+?Ynk+O(vy
zh~r5u#;&fmcbr%D`)b#^Zw|@g9(gx1e(oo0fB5pSKCl{_Rq-Ssy?Se9R-?WmD)#55
zl~)6oZ+*J&%Kf?})a}ZoE3M<o)-QG5+i*WP>gbEY)<VZCj%N*%YiX~*x#~SJSZdVZ
ztPAmdc21l(s`{S>hfQ1557p)6m%hsgH?C8@`gcv^pQfiSM{_s7*mp)ZA*n4bA(!}m
zWX<{CPxu=L?+rRvS$eU9-!M7u+Ui%`b-(%^URCCECb&rP%ku3T+cCrW9FNxcpUlU1
z)oj~(CA!OaO$eM=dOBtLom+i&lmw(`Vw95$)@~nF^_`%(-_QHuqK`_~Cd@CZFtc+q
zF15VR54w<&@muZEpH68G9Av5vMLzwod8|igzt>%wokxznI#%pe_TY{3b@wsZq%L&h
zJL0uTRy8Q2I-zl(**q!K?_hQFy=Auw*W9$UYBwf)EuU?48vgEPOgrSqQV;S@@bD6}
zzI$18Gc$Wv!;6jlvDc|xIkWkv)WuyHRe#L!dW&=s<HFot`xPa0FJ6Ch^NY5ad!x4=
zkBw}q=q|{4u|3SP+h6-rK9%-3`-G*xXHF;|b~^LKGTkBLO4->vD`b6L$Av!+E6I=D
zU$7*1=2w_ZrWo|~+L*0j{wkC*O_n)8F~HH<h|db^NBO$yX31D#k_t@~vA@D!ZEYm1
z-utolqCG#KQlZHrRx8ww)<(kg{&%ynx`M|l^sb1#72dAaM#4Jg6U85H`d}5BAmRW;
zfTOjMaO9g`AMbpyWUmUPir7=(*~i*QxH&Ff<Hr>VPO>x+Ywfk;txYd}lG6H(r@B;y
zl0@vQ@O7~^627=*RQ2^~g1rj;L&U*~;1p}qi??k0v}@UE4;6Y(#C{4t7i%NoAIFU!
zyCe0>G0w6nA`Ws78f<MO{PPU|!vz1BQ=y3>c2l@HS{n&J8)n~rsAJl76`CYsABB&r
zwUO|juE}5B$UFT&g_1?wPtniO+DLdoc20rO)wXg$O6!W1Hod8;yQA}W%JfzH4ru;+
z^iG+$1}CM#ii*QSfzsn+ZHmwkz)5ffa1wd~I7tl&oFqpAC#ffalhRPYNpTc#QhEwF
zX$=jWG)Ds`t*3#L(J;Wta13xVdImUI4GWws#{ws-XMvN`aKOoN9B^`a4mf!Y51c&5
z11GQNfm6^3z$tJ7a0+?>I3u(X!+S61Or2H-)Clc}UM1a)9!l$7Xj^i;7{g03+fw3Y
z46gv1D8>lL2{T4OP8MS%<fIuRA*YHl3UbPfQIONc7!5gX#%RcyVvK>DF=GtmY%#_{
z&YCe6a;_NTAm_{&2RUDi@sRUojE7t(#stU(GbTVjN;@jnwg55Om|i74;R2P`CmWdZ
zL_7&hcru;>R1t23v!6%+E0Ih9D^Wy1E15_FE166JD_KNBE0stAE0s(ED^)~6E1gIK
zE1gUOD_ulGE0f3oE0fFsD^tWkE1SpyE1S#$D_g`uE0@RtE0@dxD_6upE1$>%E1%2*
zD__JztB@!FtB@=Jt576B>qza$-n+;tGg2D~RJ1m_S4le-Qd)N^+j9H(`h-Z&ZE4?f
zQd;*bn+dPM@EvgBYcK+8vIZlerfM(>YPtrap=N3@25PnjW1;41Fb-<I2IHX?YA^xn
zSZ!?YNTuC1L22Dp!5~ZsFGk>EE=TyMP+JgQ19d0DpFrJ>aHAdgeX|{22DMR+o8&f5
zm=FQ;h;l^0Ji>wqm`8LX0_G9jh=6&7*^YpDgi%hwJko?nm`9c)66TQ>M8Z6>6Ok~F
z>_#NaBh7Xs%p;9*66R4RM8Q0&98oZjvLFiPQJsi_c~m!|U>;?*qhKColv6N|HX$13
z(dCGSd9(%5Fpus;G|Z#B5e@TbvmFidXrr8ld5j4$Fpnum49sIJh=F-bCt_e8(~TII
z$C&LHn8z6949sIqh=qA<IbvZRYe6i`V>=NG^Vn|0!aUY&$HF|;C}&|FXF?pz<H`{S
z^EeCQU>?_rIGD$EBM#<qW;+h%aYi`@^LP{DVIE(Oc$mjq5D)YCPQ=4Jz8mo{k2l-#
zFpoFNd6*}dkO1?9awNb!!GZ*sCv+kK<_X<MfO&%1PJnrWQ7+i>qF4SRd-Z%_J0{Q6
zJLuD-O*eMR??3*l;7F5Fd_{G@hk&Pp0|6B2KmtWMP(aZRG*FBK0~G7P0>wFSK=BSd
zP=bTdtHx|RZoW|Ri!*Y{^zrZ+C_Vn2GX32Ayf6M;A&d`OpXW}jDZ`vH{looZq*t`}
zCiDw=xNf+;*d5^*(N`Mh;<2$UnyH}^#qOZEpea%vQS#!|<;sSe2YTH85#f@1<<q!<
z+O{>`Vz=5=?Jm8GRPCD8bou0>OJa9Gcz{mYj&?sPS(0*R(^|26h~p3!X%6cOkJwMl
zt&SJFJzYHArRV-JJ}JvSuKJPKJs@}hA=QCR?~Uw8A9Sr*?2ebm2TErmePc>%ktOXG
zQkfxcmpv7h)Gf}Ck6(~9L(++erk*;>W`Y-hWES_-o@&cx!gY9z^cS2m2^|3xsUv}+
zbQDmujs}X+F+j087AQ`~0mbWhpah)&)Nt+aXlW*H-!n$7Q`dhlPE=4zP^Pq!(zi-!
ztq8$5YBBsEd^f1Y2&l<gjD(u1#VDxhT8xI8sl^zm*;<T+nybY)sQFrqhgzt`1e?B3
z+duEq*W(j$G!bqQgQP)9g{-*jJyonUnnw<GW)xB1bIrR2yU~)X_Xbf~>siSy-qLya
z60nAO1W;ri2^5t_0Y&H0Krwj?P;4Fx6qm;V#pm%r33&ogQQD{&X?Rkl?XI*|*tQe)
z?(gj`y$Xt_pN`M8ojz{B=h_sJP5?$22*Aj65-`$00!F1%fKdhtFgl$Ej5g4KG3g9o
zjDZ1+O=kgP4J=?>ItLhM-~i*(dBAuB515cH045lOUbY~6?u?%QNfa!|ESS`Dklxe$
z^|6A#4m1AllF95cqpqFL@l8&{6j{?}q-V^OP48v0nYP0$+hJyI*4){m9Gf%VPHwO0
z<KU=ra(3<OGFlxMGeoT%J*uBa{{fy};bO%+Vq|0#>86PY3VmnL|Lb-48R+XD5I8tE
eWau!Qq8UyI^W(h<HrCBOVN5oPPmI=M<NgPxDt>eT

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp.C65F333685492D7B.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp.C65F333685492D7B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fbedd0aa52cdc203057cdae04fb81c35e6c81b7d
GIT binary patch
literal 6938
zcmZvgd3Y1m9)>d~X-IckXsO*xS%l2Km$Xp2rW6epNG$~e1T0Z7Kmpk*Aw^NPAOSBR
zN+CfBw+bQ%1q7+G=w(wBMFlCyDij2~7ewuy#yidLkIVCLn)ke4pRY4#GFwGiNlAQ)
zMpHPhZ04l#)5hpD8cm$?zjoSG2d2RBNR4LZx<l3~=OwZ8mG#$8yWe@sfBpUWx$=yC
zDHCI|-z-iKTbo+8pijl+7T<yWg=teR`X67btImHzzV~|V$x{m|3IfM#CpGzheC=-I
z(OtGz)}1<gf6k04T+V?#M^9Y!oz@<{T62Hylx<`7%l-R}UiWq?{r2MB`<g6E-aLEg
z;mNJLA3VI&z52&<CqDfA;@yc0mf(+OG#yQ>^geoc>)dp)spQ1-pX|E6BaEza7)Niq
zx?;}GXLp`+{cCs2?5nG4nyov3n;Xcgp3gO$W&09q*n-k<r`+~$?+7C8#l0Ud_`zO~
zvHQ$}!{-uKyNq<j<ze@FPFQcg+F1URuzS?wtNO%$5FO+F&+_WZU#FQLOW54$txMlO
zy{Bl<n*1plNn0BoXD^<9ms;;&8*`?8vOzGv;Eqol7rkZf)bF~7ulRjw#72Mo)b#1S
zOA2n~-*IizZD_iLoqQOxBd$LG`5I$N`k{k~lcKhu+?M*q-JiZX$MA8Xv+s2C(2Oll
zj^A?mPV>>`qA!*Td%q~LR5#5jGHQ2RjJfwg&0k?Bz7k$sRNk|5f_ZA?XDgSTXc#oh
zGymNtL^u}hyf~uepf+_>lz8*GYY+5!pDyoFf33?$?+s1b+;_<#*GK!-mN(hvSZd#&
zc5?ECgYP}HV#S1&1>DHk1=pf}@8fv+`oPl_C8fT^-!@HOJ#Y5xg8EH?;VI;ey0~56
zu1G%q#dk|Lq_hsc`@XRENbbq`r2XxB-<S8$?9ACbzlxK#FP)ijyvBrv?L555u_on$
zaZgO3x?*|10~N=X-@Q|#%igx2@j3rAa`cyb^v)$miYITGypQz0I^oosD|=7=W8#EP
zBUhGnS!5pd;M9ns#}i-Lzk8i{J96&3!=`gxHu<0Nl}|bpvElseJwK0q`J2}cJB9})
zi#@)MAF(igXM_3Jx}`Bs1%7;N@%c?zlV>+?D;<2Wq<4LbrnK*s+xX>Mu?4$+o^-Eq
zJ#jw#joIh-96NLB-(%ic5}T?$x~Sjw=CiN9^>)IL;b$wRw5=ER7k2Y*v!d-apIC3;
zmDN3tZ#lc_vhVZ}+s;8ZpQC1F-dIaFmRxDveC>AqrGv|fRbw`7c{#I7pU>~F-8y<%
z^!+`}B}aVn;Nzuhgl*%V6V^o6<lcES?Bfqx5`H?HboAKmS^F<-ihOZd-NS@$M*TF`
zb|K1quWtUI^K2XY6qmK$eC3Z!!?=vm8+(rx(lg$fnEOJp=g6P?e_g$=^v&8KMf-Hy
zU60M&m|Hsa*mo(^$X?b5QG*8m^(ZXua6{nCnVK^XZam-l_^AmOYz@O(@*jT7{t-4~
z+Ughj-p~6z=C4~lKF<4P{ldfEf_bmKv_NRPdhEs%&XUT{BeL#1^U?Q(4ZiHuXGLE6
z%>KzsrhZ2jX7#%BN`9>2N%^i+Kd)$qrgqk(vD-0?MxU{JS>XmLGZIye)Y!Y~yM_k`
z<y;Z`7x8<;`TBBUXe4@C(V6<pxZt4bxUH`xPk;Z$r;%v5qT}>&;lV-GnGfvIQTM;-
z8l|aLbZ%Jg6Tv~%^p2;-V~w@9BGFTdPSz)P4i2g|F^lzoB$#_fq9+xdrOyfv4yqpT
z=+|4DuguySiAE?oNuSgyIH<bdz*r)-*XMg9(I`dt)c33k4yt}}VrA7l@tyE!O{Jo<
z6SI2+2UQ=rU3L7cq`0M#s6x?c`m~tfpz2fW*R8z!(r4Ww(J)1)>(gU`gQ{<g$=pNm
zx41|&M9~TQgz(^?>U({|zHhod^1DbhRM9E=l+M9H)qi%jZa?Sz>`ElEDmq>tA08Z3
zJ$3xEwPU)5RxUjvxMEA2A2l?#m@=d6X>n;?)xST=jvH}O-4_a(@$wK6Mglm(OaMn3
zN#IB`2^?jlfTPS5aI}#IjyBW4F-8VB#>@c68d>02GYcGN<bdPM9B{mm2aY%MzzIeH
zIKeCcCmKcIM6(E-WR!rD%o1?2Q3g&n%fRL3<mGl8^Y;9l{0>sTM`;nk?@ws4a5si8
z#X?JhdoX-0V8V?N&=VevfSz<?B=n>QBcZ3<7zI7$!6@iyH%3EGdoUV$#*H!1GaihA
zo^@j^^sEPCq37Hf2R-M(IOusd#zW70FdllrjS0{T9!!8<bYmj)q6ZV9m)w{Hz2w0p
z=w&x1Loa(U8T#U!;-b)!mgJOlkb3SjMg-4!U`Z`_1z1okUInNQcfr-Q5Fkoe2@oaf
z2pA<TB#4q$5=6;55=JQt1)`Ld0#T}tf>GK+gD7pKL6okeVU)2jAj()75M}BZ7-cOi
zh_Y4|MA<qPMmY-yqMVfjQLc`IQQpFXC~xIKl&|AqRImsjDp&;&73u^S6)hr&idGRs
z#X1p2C5r^2l2rmxsZN4X*&>6eY?VP&u9IPOV9vmf(?pcLAg2IOQBF|@sYf9e5xj`1
z$D4W)MG?V^r#c;FPf1JZtzK!=OKEIG@Ddsd;eHH%7aov)jDSq~F%mN6$0*3OAEO~N
zevE<4`Y{$V=f^n6ydUEs3w}(1Ec!7KvgF4k$g&@kA(!TqcI+_q24Q4G@Wudav>V~|
z2z)8M2ww@=hj2gSR)pV%+=g(M7W}?Pi!X=l(&6)UAy2pw0s9CqB48ikLj>$2S`h*J
zh&DvPKEk6VU?1Vq5wMSRBNFzJUPQt^(uYXcN46po_K|Ifgngt(OTs?Vr6XY<<wg|j
zqr8ZMeUuMTu#aj*6zrqg5C!`vkCuXcluJj!KH80F*hhO24f|*xqG2E1ifGtJw;>w#
z(H<=g`)HSrhJB11F|d#EA_n#`KE%L2rWG-;k7+{;>|;Dy2KF&79RvGVH)3HQ>qRW=
zV||E)eQYaYVISLuSlGvUv@GmnT{;%_ac;!HKF*6c*vI)02m82I#KAtU4RNrK^JqEP
z$GLPI?Bm^thkd*k@vx8gAs+Vet%!$xd>i6nAMerfu#b1?c-SYnkpTMyFA`v%;6noJ
z6Izh~`-C<mz&^pF6=0v>(h0CnbR!Y=iC!eaKGBCn*eA9k5%!5~NQ8Z&M=QcU(WMh%
zpX5do?327mf_;(?Nw809MH1|j+K>eMB#%~teUeKj!9LlIWY{NrkqrA}ACh68+=^t_
zC$}LP_Q@Wt4Eto4P7d`ATzFIS@cy*#F`d2KP+qNG2G!H6a@}ixxC@=#VW_C`!OyE}
z_cli>Pd)~`m-;(VcEUgaA`K)U%0L044KyIezyM+mEFjLn0pbljAi*F25)C3C$shre
z4RQw!T64g2qW)$yQm*FRQc~5cd6YdpAw5}r27N+W;Y=i!EXShknQ~^K`lK=J$@t#a
zUe66vrt`w{V%6O?qrB*t@o900GM!zVU8z0>t3P8Rf{($Wy@>}fd>6by2w((cGJuhg
zsQ^YnrUMuanF(MFWHx}YkhuWHLFNM(4_OFc0%S3OiIAlLCP9`1m>lBz-yVMb*3R=E
zweOV7JXzf-3-1)C=A2xTtW3vsj!RVU>>IY!G=F_?`j^UdmYii$Z=(MBOZ}{>mUYXN
z>0aTzV$_4GapvBzA(J+iDAP$XNr~#+$9}$ITv+kOUzF*tJ-ZU>`<CWm{jb|HzCEH$
zm+8v7tAC05+^!te?Jpfd>yVz3o~^F9`ct`^Yk&0+3gITaQ2le1W86dlB26S9%0vO8
zO*A0J!~kMVEFjLr0pd+OAi*R65=|l?$s_@iO){VXIRgf&+kfGbEC2PlZ9gcxCA%uy
zt{#?`>9G5qV!JI?rt=c=GSsyCo<-Ra?Jrrtj^J<+><EqkL~<k`ilYG091V!!7(gt?
z0^&Fh5YO>|1Wo`Xav~s!lK{z_45)uj|Bjp2Ukr^N)Nfn+w>qI`0<CVTdJ0gsxc$8m
z*kUI>3v98I07N=TK$MdLL_29fjFSPxI$1!RlLN#%c|d|w03<p^K$24eBs*n5g*k<V
z>SC&|y;38Bue(B<mfR&dQGLo$&f_+GY-mGq2R<=GgpB}>a1g+eHWE0}K>|nFDBvgu
z1srXofukKXaEy%sj&U%+u{IVs*1-bD**M@h2L~K)<ALKHJaB?d08VfSz=<{yIME>j
zC)p(6B!>i?Y?FbL9dd`$OdUUQO#55M+Nt*1q3s`R?Ze;yg5vLwqyIOh(P-_7RZ$lX
zW^TTvANR}{+vu^HXF8l_Z0O^-&`0~EaT6yfSpMu2T3wjFlOa4ZD!OxQOj%rUVXwG|
zvV!<-UAiVcCg+%P^ZF0SClieQvc<<U{@-k3N@`kWR`(t~d-v&!Q#8W~QaYbZutf=p
OrGv(!63aj{Hu%5as*c$J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp.D0AB01FA41C26CA6.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp.D0AB01FA41C26CA6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e6a50f614a0749aee2e2d0ab05f1c5b135d7ab3
GIT binary patch
literal 7272
zcmZvgdw5J|0><ZjGl|JXBnXknO@uOY?sp@RNTlPMZmQ#2Ba*P`BG#o#+tf_7DAhq!
zTMw&D+U_W&b!kg%i>;P*t!uSZb?dr9!?JbDp3Lr%_m9o<_{@9WAHU2w-?@Dy{fmmm
zcD7gw%KOipR588OX0ccjj6c(+PpynF!oe7eW#-DSJR_=pR^l(NzH+SQ{Z(~WKBD?`
z&3%5+!P4^0-(=<2E-Reh<(ahHd)HjY6S@ysu(toEi1OFObp3AZBCaUud*`v`Q3pD+
z-DYjOFO=I4zyITv2Ny3iJ~xEz7`&CW_tJ#h<7R9r{kT_-K5pgeRQ}HkhoAeQ``&{S
zU;gcu!};=~w~w7Td%b4N`sK9=+!GJCjXrwf#8(e5+-X=bX0-K^eE$3CXUM>k(*u_7
z7I)R3S$t;ld%*+T&b#~DP5$uA3)@aEUH)yS(7s+h>G{IwCo_MX^5}Z^ts`9*@3e~g
z^aFbzZo^NlcQ`7^`nrUkdsb|C@6Kz}4{g7F*8ORlSA|(!p5K+<m?m87xje2%klZPm
zuu18#(O=SS?CI`{hMn9p@zIrIZynY;o$TD}!ugTW<7+OYmbd=POH)si9uap&CC>D}
zdSLkRRY{qD=&r|DV)n_?oL}lc`jx)%Q?5^}*b~$~x2DWl*LCI1UH#J5-^#I_^Q&+6
zdjI1#&wL=9m^!nxd~%oZdskFf1i#+!gLUnjOFkPH*Z9Y=kFr;Nm2$Cty9L(${&FVJ
zBXw2&n~iBCAZBm<vd75<l@YB1Q>O5zH*Ku=T0P(Cm9Cv$S-ok}s@=Y4COd6gemOB{
z{+=^0uH@f)ac=FH?XBG<<n@xVbFca1S5bW;KFY`$J97W!-ydK5*G+Mafy0NVY|^$q
zw<lsJd*RN=l3<TTXOrtI4t2dTa`TG^HV?Wm<A88w%5CRcIq^A3EB3$B>1a`X;Ns@g
ze9zXo3o73kwQoRuRA^Mm;$B~r9Pl+Z9)6<rytbdLTk`GamVVcg>-n?}yFx#FyxrYp
zEj3}qjZE*_HLJ?&-7E46ukN0Avdb%LZf-s?^-TSxn?piZ$9Hr0-#PcxnZ~y7o_nEw
zeElm|Z9i{WH}vh;gl+ayYb<m|TVX=bI`j!O^XLxAcIdGE_WXVOE>Gk-f9Zbp!Qhf}
zkMd^DJEt7J_uYlm?x9N?g4>o)N-wE+X>5Gn(o;+Raj59>^c8<!xAqTTJ(zv+-*L|t
zezWMp(p>BFeUA;g?vpAuFB~7db>~kn^!;$@^0}T?`>bP5&CiqDPMG!T?DUm!bIUix
z_859l-yTl}ZyeF1<B`0Lj)~KIFPL-urDY?m3*Ywyznij`>?l6rN*VL<&PCM02mVX0
zG25qfJo1(||LngG-c3B<lYSg_r0H(N+e<rMJ~a7g8=K{!rCrX2UuM3v@{?abOUy~T
zdw*})?X*dU_VwPKoo6jS|5{`0H+*JQ{gf#7r%vq`Ke!vy>y7-UH*7du_v?vMOVxs)
zf41KfVg1GX_nRG0A1&P;I&$~2>)xlex9_Z<ot9i2{V@Ap!pR4##Shh;ot}-@r_X!w
zyXRPI+Nv+P&1*i-OtaqkeqVNP8yBbAX8OkNUh@T2no+pUGHuqRai35YOJw?IZxpOi
zGh)by(U!9Ik?o_xgK@nG|3u$=_uQO;^3yS7xKU?BW+a3MvrgFfTJnsKuI`B;&lq(=
zWI|MUFl%nh?<%PE)2_#m-x+msWO96XFl!gTF!Dj7J3EF9G3w07%&72S)_v~(di^gK
zXKjojBaJ#KGO1N~Fzc^xj5zphQo^DbQexDpk*RUv!L0XxROq-@G?b3Dj56w+h@8RU
z!K{z0Ub(FC<-c`|Aw!KiEix@GJec)~N3E`2ypcY!wPl!5XLZQx8XnC0YH7xHTDmUA
zkf)3~F)}eKJec*Zo)Kq0zcTuC40+n9QzBF1!-HAh?pH8mc4Y6;81k%9XGdm_2oGlc
zFy8aY*{Z)?j3FMQZX4M)Dm<9=)Qa)b)+e?cIW{_c#1`$o|L*#qTp6)tsR^m=&7Bil
z=5#n2Q)nd7sX%f&2eu^IK?6yPG?28L29j|wKr$i&B;#g)WF0J!tjGe%x>+DO2L~i4
zazJu!4oKd?1Idd#ki44*Qg8@B3Zej{;1+-s9U_pTC;}<EMIa@I1f(QNKuT^2NZBC+
zDT^|YvRejHaVS74q5`DiR)AC;Dv+wE0;#%HAT@^uq$X-WYHkfk-Jt`ii#m|HTL&^%
z&+YyAyejLX_konJ=RZ#7RU<t*e6?tqn$8-^xrl0+a?SwdTmhM`p=j*s07YZZ)KCoe
zOn_prXKN@Hdp1C^*mE@$hdmddIPCcvipQQ0P(1cR4JBYN1SkP}v4#?{7Xy@ty;MU<
z*h>LQ!d|YSWbEYtC1bDDPzv@+fKsqmYbX_aH9)D@Yc-UHy%wM}?DZN-$6gOmI`&0+
zQDMsjcv64zaWd}&#vyOM4ZtBUc9y^a_c%wiB)XV}j`q;dF~tmYjE8}aEoPx(JuGxw
zF$W#z;h^J-dFXf#51mjfKqq(v=)__XI?*FSClyQ3NgfG0xmboy_Q=pF#R_zaM}bZ)
zR-scpDs)=02A$^7pwo+W=yZ<`U0=QL<MT7NtUxb#oXop4i4NbLTMk`HYD#zWR%YI-
z<D<j(=oZLXM>#*hQ@D<zF*9`(gPE<PSj=1<#bM^_C?2y=M+un4I!eSW)lm{=xsH-C
zD|M8DS*@c~%vv3#Vb<#?9kW|^Ki*U36G>Tg_(=qIf}c2RNXx`_))MD3%t7L;!`wie
zH!wF5r_T!iK45h&#_Y2>{kE2#_7fU?bS<ILM+XUwKDvR>=%bqmjXpYHrO`+GY&7~9
zKVi_v)Di}LOpq|>V;Ts9KBkE<=wkv_27Qdr#-NY&6Bd1JEn(5e1__Hkwt=wdW19$z
zJ~m)w(Z~91Ec!S<;n2s`5)OS_kZ|bZ8VH9zu8DBy;{sL=eVot6p^x_y9({Z*;nBwj
z36DO$f$-?#n+T6SK49h1$NOwN`UF1_&?nRq0ewP{2<Q_Uh=4w!i3sQu0#*Tig3l(P
zPxKQJePS&U(I*Cph(57_i0BiWh=@KhU=`6P`fMWlBtMbRC)E-OeNvD}=#v_Vgg&W>
zNa&LSRtbHQ&nBTy_7fR>axIb3CkKg)KDmL&=#!g>j6OMFmC+~rY%=;3KT*)9)Di`K
zN{}e%QyPeZKBb8$=u-k#1$~Oorl3#t6BT`GEm6^@28oJ3wSlPUQ=5p2J~d!f(Wm-s
zD*7}((a@*W5)FM?kZ9=B8i<BIt%+#p(*jlveVWgvp-=Y{9esK&(b1;|iH<(Kf#~Sd
zn~07+Jz&+*r~7Pri!XoPKP|uAo_>b1l?}2F8f88}6%PCNKi;0UV|SeKT4r~4gR8rp
zhQ!zzNUWWO#MwDWyq$+6*ab+UU4$grB}lSehNRdPNUB|hq}eq{x?O*q2E6k{;Lx0Z
zwkEM<DIHTfn~x2#Woe0N$>z)UV0CF#2m1BJRBTy>mQi57CJY_Yw)^EJy(5gxxly@o
z%>DEHz`_HL;mT9S=B$2M!^~ImIsf|Y_^Pevwl&ikxtdAOyV<X^e)#odV{<}$LI?AC
z^xbt6b|2q0<EXJYQ_FOjk1-!!n=@;~Pb=RrHg}8a7H97J^)q)wJT+-;k+C@`E~$fg
zzyGULQXbK7?KNX_`|S3#xn{C^Xzmqn`l+vt%>!%$x|qL-`R;8zPc^Sfv^*uGrKDw<
zyTXL2=<u?H37NI=nAN-{VQw~Nh|5`E{+6+2w2OwsxEM&Ri-p9wI7qyUha|WJNTN%G
zB)KF=vP*`fxD-gLONFGlG)THjhm@!1<(uEzJeqBy!wW0$MMISHQ+%EbQ8Z>IL@}7z
z5XEBVLKKIY4^cd3Aw&t7#SkT8mO_+-Sq@P$W+g-^nAH%aV%9>GhFK3$dP_e4voCK+
z!lv(yDU&rKtIXUBzi<(^Tea!9(Ab=tn450yBy*X`m?X_>Oy*9yW?vs5y*4$UMv2*p
zoJpDIoH2Krmz+%6m^)R@Sul61Xh=*I1BtC-A#qh4B)*D=Bvc8I#3~VzR3$-@t7J$@
zl>$kvQXy$o8YI0+hg7H+7MQ!m{4T}`*1UMu@<s9LYObQu(c#sz7AU!0atHI3$JjSs
z=Q#8KGg75<BBW}k4`-s629&O(0i~;HL>VswC{xJ*%2YFmvR)QYwvq*ut!5GByd0oh
zB?l;1%^}Krc|iF}9#FoTM^x|%fC`lYphC5PsOS{|6)QzR#cB~z$twXWRZ4(L)e@qz
zR|ZtBlmV5iWkeOP0;p1{0IF0gh^k%{P_<G8RIOGK)w~*@TBQc4R;?kbdv!qdN*z$W
zT1T{>-tX~?<J5|YrOoTN)25b9d%F39w|NNvSj(2(*Bd@d8a^Ykd`zi#>^Mu9{BJqt
ze~|Il*p`puT0WMyd@P$(KJoWP95{Zk)fN%i${rOH+d95YT>pfA1>F*&`}b|zv0eM5
zPMYrOos-)qk4bd&%2J+4|9@LMq;yWr$n4TJyL*qGPL|_ESxu9YX`wK&!;=FlNKtXV
Hn;P^Vq*s(=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp.43793441438DD042.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp.43793441438DD042.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5b2e2077d082157c13cd8637a7bda2a007a9bea2
GIT binary patch
literal 7270
zcmZvgdwfi1!pC!-8Dk_72}x)|L?diFbMAK&ammGmR0uo7rGp5v!dQ3J*fDlxEjoy<
zyQ&!~3EC>PtzE6?y3}oHx7+f%t=8QXwOd}=>}1~enBO0p&&M<KJ>OsQbIx<_rA38>
zuXc=y$s1oZcS^;~vF4bVn56K(88fF>n!<1@E+%H~#=pACJlB-O>u=rqVqCGK=!?;&
zdCu4GS~eU$HeqGIh236S!WGo#y#DAPb9Q76+?ZTYJ0|W}+ljf;Lock4NsQTW<Getf
zeDHARvHkAF%RZg?u;+%08{MN$?OFQd<|FTeColALcdS~=zTr>k@J{!Z%Wk~LctY*2
zP3Upp@spJHy?%3Y&tEUz%W=3zZC<`&WyNKt{Wt%-{cOR4`ExdHa%bI;uUr^^j;z_c
z_s!t8iN1OJw(q;>T*DuJblO(((XL<J`KI(lRraQp>zTkg-^+(*B|mI^<MM*ihHd3j
z_pg~Yzxc|+%H^K_8C}$FTK4im`lJt6pYMK7zjJO5eUVxHdGeC4Ej7pIAKTS1yhm_Z
z?3Hb&e?7%+Us>8MCRBUGHhF~7^yt=!iO$7xO@cOPz@g$ne_ph+&zmI)BfD;k`DF0I
zHJRg2Zw-BBEpr?%>pk4LdtEto=1Icdq(w3NK6taW|DWHcPKmEyHMg;4`+4u3{xR{j
znoC1(udsBkA6QZU-TlU6jrn7atRGzwsBwid&VKIPvG(xmch~)<XyVvY8^`tib-upp
zWD376Anqu=RDPw)u+li^`>tYJ&RlO`>7m7=X~)4nXSlA%Gv_vbI4jn)qj7lrq!SJ6
z5})q-afP^T{^HtKKbHHKlI5k77T;NzxT#OC7DLx&NOZ+a=~si@BWj-1e9_LDTH2tD
z`ts?EQ{T;&*T4CGx+(p7@x$HnfulJmvhC%Y2edr<LEnG(el%&)%M&N$r$5MDe`dD1
z?C)u|?G;nIjy%*OIJKzH?PvY@yw6Yc3P>FX9IG$ve}km%ntcDr+)?ItVvpVymi*qb
zsq)~{*Ox4wz3PEb**B>!uJeec)-hLaEIm>5!{)1wZS@=0{UZBcY4&~J*?#-low3Is
z$2PQE{zZN1)hg$wZFl4?88N8qgwusJI|hZ86@Gs%QS15k*IC=Q1ZodW^Jb;|Fz1gC
z|J`ySJ?}!!orz25Y}~$iKl8!9s;AtLBP|zG!GcbeSF1i7cqX>&m*14~O>fBu^V+9O
z-}R@Ho2ij+e=uswN~vMOus`|-Uby^VuxnS_1Lp^1-2b>+a>L3VCv1<ar{#1ww@uhP
zRKD@;D68<%1e<=;^4jRJfop?1ZoeKlnKoQ_e|>SykkqfM>o489|A%{3@#Am*m~;8T
zx!*aC9(?wP+q*pn*=d&GlY^iB{0Vz@_13O)X6G$PH1BD{oo#n8&EIB-lCgj5(Q6AX
z&n><bIHi~6|Jl69`@-C<IsK*=-%4jkeLn1QeE;H~pSI|9WJ}#wUrjv!`1a)E8{R(q
z=_kD#8W%o%x8iZ!-J9ISp*QD0>h$bF(QEU+>eXg>_uu{8_9yG)pU<rO`-V%)cCSj^
zc<nD$`_@o}{j`%+lcxUusQ0JYGZ(fgSaz-T!i>a?)4uhLW%KJ}X3U#1ZVwd`W9j^-
zH}c+AGvi3vsF?C(OLD8o5I$Q(eq#T3%cVsF<RNincvxpzGLs^Mp_ATSk~VAi?Za{8
zmtmb`Noo}t49)EwTtV%eaW9Sx4(l{aT4H1{w1Z!2d6weLjw1uZx{IYttH@yJUQZv~
z+i`u~yK!VhSa+~=Xc-v{{oA^~tk|?@Yp`|9u&~Zb&FUH%4860udc^w!7Eg{Nqr*De
zl3f-V41M(bvQy_eBrT64rD5I4(kUS_82WE-ZCw3u!DsE`$k4FPuw*1e21DN*ez0NY
z{B39AV@8H`&lWvji42CmJvQ@WTDm93kwIadVo7Ng84Ufy%Pr1-bZgYtab!qXr(4n!
zBZHxTN_6eH==tn=9C3wpTT9zkk-^Z@D<;g?nG&73%-G0`EpL3fW#?5#W_)?4q)y4k
z>cyAaZFa`k7GZMQ2Skas(cow&4UVxf;20+Zj<vDiSSJgPvvJ@!CkKwV@!)tT4^FTN
z-~^`tPPB>OM5hQ&vPs}1rvy&6$>3zC3{J5r;1s6<PPM7vRHq6~vuWTorv^^9>ELvy
z4lY;E&3V4tl=sp5JSXEA>l_<7wxSbh_fhucRCK!Re#*WcnD$XLdfHFX=oufyplAFP
zgP!$KEPB>YvFJG;#i8f?6o;PoQ9OFyPx0slA0?m{{FH!R^id*u(NBr!B_AcBm;983
zUiMKkdf89O=oKHOpjZ5qf?oAeDtgsVspvHyrJ>jSl!jjSQ963vPwD6j^}_t<l=sv7
zJtyP58J_aylO{}giM<pi++{C|5?w-rqg^yOri1~<xEOG32@8&OvEaB84jkv=!0{zK
zINrsB6G{Ydf=d7=mWbd)mk3TOk-$kV37lLagOgn{IHg1Zr??bwYKaO?b*bRA5)GW@
z(!l8@Iyl{>gX^pJeSUn#m*?qu&&jx^k=V$!EIM`RozlA-XLjStmKYnkqD4dYAZ33C
z58)t1BQrsYL1u##i_8Tn4w(;9JhBj^1Y|KtiO5orl91&fB_k_AN<mhGl!~kcDGgZ<
zQaZ9zcRt@z#`U{AHgesDjo>5pMI^do+iQt^HFAL1gUBIb{~o!C*u5tBeZR@R0@-V}
zFEmFz?ISey(Y1ueJ~}{X?4v`3#y+}<(AY=&O*Hn=UNen-jE^wb$J7!A`<MV>u#X85
z2K$&M!eAfcH!;}9c+Cv<u|C3LA6rXU>|+Cj#XdGfSnOk)2#bBJ-^5}c>ov33$N30{
zeOxW!u#XE64*R$e;joWuA{_Q{eiMg%oY%}@AMYbP_VKlZ$38wlc<kdtgvUO<iSXFR
z`%OId@m@2JeS(h&*eBEy0sDji5wK4P5dr&zCL&;;;5P}_CwR>Q_K7|sVxL$`MC=m-
zM8rNZL`3Wpn}~>gqTeK9pXfD<*eCgjgnd#ik+4q+5DELF5RtG?Y9bQ$Nq&=reUjHK
zVV~?HGWN-}M8-ZjKxFKbLqx_txrxZwC;Lq@_Q_tejD3oaDA=de5(WE|08y|{2@wVR
zlqRBJpW-(u*r#~S3ihc!qGF#~OH}Mr14P9>HAGbGQ=5p2eX8H2VxQ_YtJtUch=zSy
zEzz(~3lI(av=Gs-PirC?_Gx~ThJBjXtYM$-BRcl!wM550JwSBq(?dkZKD~+P*r)qV
zI`-*avmWj1Q*$@w$)lO)DRX(TwRog)1C>AQ@waPd9B51kKg3w=-QetQrGXeL1H@Wc
zAkNAG@m3y4unIt;RRof(5|C__ffTC(q*_%V&8h+ER{c5kU-yas^rE}1Nql*F`}B^+
zZ9{x{MoLDS@$C7Ey2_JEFIz#ymuG63dB$VH(1C5c-&~o~BD^@aRc;$&`^+AYf6_Kw
z85CZeRgg8zcsgG6uP0w_+IQ*0=5%JRW~2?>CN}cW4A<s5%Dx|OQ0gcenW>`~WVViC
zk-0jGL+0x!9$BcP1Z1&}5|O1kN<x<FC>dF)qZDMdj#81eI!Z&<>nJ_SHTMfT>Svav
zg||jhVp6Jc8@{D}V&j*GW}OW$?xJ;Z7<X?CcNWbnySnj>@ZxT*x+NI9+|IcNS`3=9
zwJ^N6LqdmC<2J|8;Tyx|w++!vihJ=vsdRjcf~|MLznh$$OdD@s8i(fIa(Dj6(eUE_
z=Ke1lvt&HQhyP^r8;a<k%t+73GIkT=GkmYv{GK8jvODZ~#(yfloOaMajDrDU9V`&%
z;DC4s4<tAQAkiTLNe&4}cE~`ALjh79Dv;*TfOLlr)LZY}$JiQIu6cW}OnUD^cq3(%
zWtAJd@ONB`M=jg5Um9MVo08kvNE@G2Y~<C9v002)GvTe({9eXb96qUtb_KSUNCPn<
z1H_6f5GQg#yvPFyq5vd{B9J6XK(Z(UDWU?TiYkyMYCyWE1NG8-J-_?DW8LW|J>F|R
zd!%HiaK?HYr<(BkH@|p-_4nB4!TNh>AjZQ0u^twP^Kd}ChX)co0+8qtfh3OvBzt5a
z#iIbJ9u-LQXh6D02g=v;^Ne{n-lEa5k+)~j6-{fGmTEl2g{R$Z9~WJ1d!@Y!sM_wu
zQ=yxNC|yZIl&+>R%D5SbGL;NOnQ8{3teb@>TggI{t!6RGxjBe(l^jI5Y7V2kn};Z0
z$wQQ{<}oU`1&9ij0z`#s0i&W@gs504LR734F)Fzwh)R_bM5Ss8qq19us9Y&SRIZjW
zs<;)1DwPUEm1+f}s#}GqTB$-*tyVFrxiyGtl^R5~Y7L{hTZgD#sY6t+)-hV37d$^0
zr&m;sZGL|{V|w|FA<bXh%|qnJRNiXs8|kxNy1v&k{?)PWG2>z)$^VnD{(mU^F(&$T
zT=eVs=-2Wo<EtixalnLEOy(Armey8r@vRfvBorkT<aJAmE$Z90eY@liFKD_Wr)O@j
z-b{+EN0#!7&i`#`YI?^`nO$D&n%(`Sm+dUai?W&_rO`ruN@~CU6{N7FkCQ6?5Blrl
A+5i9m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp.2F31492B9345223B.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp.2F31492B9345223B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..46c89fe115d502fe02c189d9494b056321ccf4cd
GIT binary patch
literal 5692
zcmZvgc~}$I9)>d~0U@9)$|6NX+-6@4TL>r$Xb=?zO~oyMxJ2Qm0v5Q})Fjqw-GZW4
zXjG6^i{id-)kmzTwJvo*>$PfeLlNBfUha_l4E+AMJP(KY=6m6t$(+pOkdiofaHy|L
z7MGP+m@_3mLm`vNyu|Ok{9MBuxlC5=CX*G~f6=8GZ$^4PSbt+v=DEl-Q^(7Tv~%_h
z%=0%aidKduCVq{l{904{pAX%Ju3j6NMwfMAf|Km;Pam8%9mUGB*UkJ@!Pox!GDJT&
zDpS#?e*2AAZ~l7m&$5Z<znRuyiS5Fp8=ul^vcJlmb39TLlOF9iYxc)t|NSKm_upPP
z@a{i3Wxq8yPrdQWr^j;^E}=i>hFy42=jXBe@W<~P|8iDu$MmP_I~~~Y<JR6xOy_f3
zdgn9_J+!3j#)wHpXLD|iz2G};>Z<&~GY%~r>bBXvZq(@r``nL-TYZ-u{B6>STdtNN
z8^eB{Rlp2CzAALk{qh@wW~|pftWLW9WB&Ie$<$t%-2(eG3;VZCZ7A+Xn!k&x-Me!~
z?9fu#f#iVg)eYA_)htNv-lOZapr|{sJ)KkopEFq<H_yqv9zr<1Tctdr=#l69TDPQl
zX^GwG>-qP})V|N`x2@rsp5?Rl?aSypV%w!t<uktRJLik#9iGhbs3>?ed}}Aevx+}A
z58E3!$v<t4%ek?;t8!jlJaen(xUQ~Yd24JjFS|F;bZP1^d>KAu#let``Pm&euX^iW
zczV?plv3?;S?LjU=TYX4jLWB)+vU^zS7docXWZ;F9~+%PJMXCv8@BN8KmK*P^!J#W
zMDCV%Z^AN|2(5V?yYPJZdNgC0N7?FrQGs9n#BHtlRC9w+xK)iA&ptMJ?XO!AFnmhm
z+2f1UB}bl|YZ5Nxx-MSx@b`hQs*7VhV*<*~mUg|AaK!dtcejDpmY!PFu!cHAj{W8e
zLukhzOF38k;zcvwvDSOn=2g39V~eltEu0z6T2tpw9$p$BKkxjW+oqr6lQd@vOG7Ib
zUYT!Oo1ob7?B)Bb2TaM&*QV@UTZ;F79YF87?=$G=>mBJwPCSmU-?Mp4<baTL*U*Z1
z2KQD?tBf0s7mTPZFu2&)rZjq`UcB3IXS%tk&&z^!FW&fB!isK2Kbf|;z`k+aes=fL
z#t(ezsSe*@KO}TE+|52Y;({`*E^XVj1EE(vW|}tjC-0@qDvgoX4QdY0djBl@?UpUo
z{nC#5Mc)}7hx_)Ly}It%!#i1D#QUASm=b0$dERl$J@=2b$2L#fG`Vl5-7h=S0b~2^
zD;xIdYo|BGGYbCNe&f`s*x#>4)pYu_B*plTW9N6?yxeR3<V^!s{<IT6b9CZzqr%c?
zT3vWz%cVzr-^x*aXNNw1yL9lqrHkw(2`783cXEH;`KWqD&!M*qD^e!iRpG+!X21ML
zUyc8E?cqJMFGsA{A6GGdu0=OrUO#tcqu=$NzvPequK7~IGh?5Ir)e!Qp6gb=O?owL
z?90y{jxrW$Ru3Kb#j?6>S0_392i2GFsyHh9^7)7Qci!C>`PaJ1x&=obJ$>w~Rb7`?
zl#@9IlgXUBA6OPw5fSW$bZN5mF3w$C9fNp(aC|90l>a#|DJ<0ujS=f0=O8b~Ak|*m
zzYUnZ?QxA8N)c->XD?UBAk_mtK7abtgQD$j=yS35clPh#7^K>oZOO8|$c^bJ8!gts
zzQMg5gH&&L8ntb~wUxDQXq;GwIERdO3{rjS^}6Rok7ZljWn;xU%qdLk7^M2*di%;3
zU!C-GL!-nx&^fS^W02~{8Nr8`(5EUllq}Xh&OWY=L8@PpPJbM1m~h<<jTGwu=KxQ~
zAk}|+>UP~So_yejQpMWa+1u4INVP0?%GA6@kM>tLMCo`{m!9}oUVT>+?4FKOI4hl>
zm{fQ;NLoCpJrNWENJ2#blF$->Bq<V*q>2P2sU-nPQ4}C46$MC2O97ImXh6~`8j!S>
z1|&l<fMiq*AQ>$KNS0y&$*NdDvRW399K`{WQ*nUgv>YILiU%aG;sMEPc|Z!307yY4
z08-EjfDBiMN4HI~^aynXNNTmZElKYY@gldrLBK_}V)#<5{ZisK3@?M(j^Ud?Caf3%
zdBTPfkSFXI0eRAjk&q{C7zugOj**b3tQZA(%7#&pr|cL7dD@E6kf&`J4SCv*(U51X
z7z266hB1(5>=*-i){3!^XKfe@dDf1xkmsxz2YJqhaggWi7zcUYit&)=Z5R)E-j4B*
z7p#~7dBKJWkQeNj0Qvsv{_*V-WPo}ANU`eJwj|vp#R<}Siv$xS8BYPzLx-n<WWvpG
zZX^=`C3FNp2@?TPGMNM@sUrbOnn;LJ$rL~-9R*OzL_w5JrU6RpXn@it8lp@x15ifC
z0F*H?5M`5DfU-Ikpsb07D3{Cul+$ql<xCtz`D7lTyp9JbZ{i^;Bnto)bOJyHlK{~u
zbyVAX-aTET)_@eFj%iEMQ*VgU@w5Y%+k$W_0++Z1;VWUTMEDOdHzB+k<`#sT<>2>i
za=aL3vjVp$+Izx+2<Rh95CMHeB_g1YXhH<^5iN*-KEfs^ppP&s2<Rg%h=e|}1d-53
zRw5Gm$R<QWAK8LP=p$`%68cE9f`mTGf+*;tN)QEoR3)OIk7_~`^ieH{f<DS7r=X8A
zD=6rrEr^Cbx&+bCM^_>m`sgM^Lm%COXy~JDavJ(*vx0^`#)25=V@ePMeM}`{ppR)n
z4D>NAh=D%FCTF0JF)JA8V=ah<KDGq0(8pFH7W&vG#6ln2f>`KdZE_a+ShIqKKF)$T
z=;KNd2Yp;6;-HUfLLBsQEr^3Y&L-!ek25Pc=;JMjhd#ap@zBRtA|CqqCd5M@--3AP
z<85*t`gpT~hd#lA1n3h=kN|x`B@&=dXhH(?2`xx~KEWmzpieL>gm&M+1&y-zuk!!E
z6zTCU@jB_j6#x0Vdu#FzpHPY)GhOgL;O^i;fJC~GAW<$9NVE$L663;v#JaE`aV{K4
zybBMK;3Bl8;Snp0zRZ$h%so9QGAK^^1T}huci%_Lqn*Uv;jZBx(kXnI6n~By8$MFp
z9h?ySxwLGMe!3|g%LZ_tt-|pA@Yz{~5ipZg7zr~~g;6lmRTvF3Q-v`wvsD-iGgpOi
zF!NOy53^8(3GI2_>-SgI?)`IL>q!QOM@lET;8{YT`ofX`akrPJm#_4hVR47!6$?+~
znvr{YfL}nMv@UT^5A+EPkk&ZTMMX--3aR~-fD>WMlX*pHckRnWLCA!V4C!~O3lBSu
z%-Jwl-0k1V-&cBg{T!N-<&?1DiMYE<?=FnAOgu5Ve}lezU9Gr#kYZ4XbnwzoOzBvI
zNe3VupS#kr_L6psGf{)bNeAGbPG|^_NDT=RrJ+EgH8e<!h5?Dyupn_74kTW~gCuAK
zkRsKQ1Eu4?Q*Vx4?zQ8lczVHU!IPyKcAs~8)xpDWvA8?jC%n5vOKV8+tXmh5(&;^M
zskWi09j%kwr?(F;QEit_=~zliwD@8e@gneI7zvO_BMA~^q(GvLG)Rn*0f{xTAaO<x
zB;Lq_Bp3ydqSeuH(&<TyTpy)laSIMuk7u?YFmAxJ+Y_NDKt~t|(2;r)bfkd<9i^v0
zM;R#4(Rvzmw1Ea4qh~<J7#Ps8dKPr7fdw6>=Rn69IMDHW9(25c2c4i7KqnZ4wk~%{
zc1G(z0P=Dt=cTr8OlTb(U-I-F|BmiCJ?Bu9bH?~d6DI0qlcY)}o7lcNzI{{QzUla@
z|BI7zva+X(n@LlL%N0(}9b8=9+&g-DbV~F}i0k8}OdRCx*SU*-*GTcRX-xkCv8)dj
w)iWIL{{Qy+eiqm*xJO8@-hBy@=6FFB77-L0!1Rmv@f|WW6Ad0QP>T)wFN(*a^Z)<=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp.A7598EAA8CFFD492.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp.A7598EAA8CFFD492.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac50a242f6e17ffdb56e42263b373fe6cff22fc7
GIT binary patch
literal 6686
zcmZvgdstLu0><ZjGvII$g%wm(B+{I7zp4y_pn^0=ItDu4$N_H~22HD7)FOwmbk%Hh
zM_Y5p8)GXNm5*h)nYp1^k||!YG)+yd@G@eGyXGc4&OS%pKbGg=d(Q8?FY|l8b3V?Q
zbBYTK3Q|%HhP>&8bE{@NTW&HK42j|2InU0nXfqlN4RHp;+`#9~iSF++yPnu^a?8|%
znFnT0HqNsLc6MI7*Ish*@r?&o=C8@vk^R!u_s(vbesYQb^V-5AWc$81eSsGj9Wp3}
z$=y!Pjyb*L@&}xAL&DtX?ib#>bLra2J8d&w&gr@Ss+^M(9B^!RXMf9)^Jy0@&h_0W
z-S^G4ys4+Bl&)XU{@AkXXN8+Bzn|%s`?o3U5|*s|+4b(<@7((BYWeIJ7Wg-u>H71)
zb6?FpO^t8ZK5%=}oCCWjZQCy#xpexO-M5?6@+<r2=Nwye_(;#=l_{&=c&z89;J(3u
z1$PVInfdyzw$Y7e9*_2wJ~jVXD0S4v<kb4u=9d<o8gSHl<>(7FPxoPFH03vVgBORS
zocJcYPjNw1YhB!=UU6#es+gg7o9E6L^+ihGoR@wxAJ}%gIjvLbs?(<agH{aO{Of`8
zqJAS=i<0xBlH$_Z3SQf{f77{|vL(akd|_O7vsZILEYS{~7*zTB`K6WdTMM$zFI<)}
z;pZ3s{K>=aSvwpHyR;R|*m|k8@ynz6=k~9Ct}3{AbSU9SQ~L4KkLNspa6$9L3D0cV
zIo3O7Sicu$Jl67EkFo4^x^;}@cv5jg^0!M9Lgz0|-By0)+#>gvYHXomUcar%-)PvX
zy+KahNEI$UeQij$(w6n-n}d^A^y`1XGGgY=H1m<3zm^Oi)$-7>;)2C{6WUfkw`xJA
zby3+vJ7#^=f4BJK?McPKH+${Uwp~2b?|RnzFCM%}%2Sq+KKX}>uGai=pn2?}(_acd
zR%;usnI{Zh%lv(su<~$uV(~|t)2se7QhvDU!r8>-8<VzgUA_HJHA_$InEO;V?=5-j
z!l>2x`FV#vIQzt>(~E2uYe%lN=U;C8;!tgTUF))zZ+2F@j@sUurK*J^FArQ(!Tw}w
zJF)k*W!7;6dTtr_PGifk=s^n~8N2z_{NkS~pM7QIrO^{_37u>F(+Y-69XTjFt#|yB
zn~U#_F8SeJQ`zd>RqNZ{|J?Rj@!`pT{XTkE&cW3;YDx?fPv$$bdetAhnRF#fJihde
znfLZiI$Q56EVb0!QJ?AG_I%!yc*lo(=C99e8{Ilo8@0bK{mu7l7v*+(bfIxj$)a&v
z6Ek|9ZMBZBTkxN^=YN>>;g1$u?q{_%U+;JRvHLx5r+X`A{j1S+bI0<BipF1gC972}
z-2IQ=v)+2z;eTo5x_e_*?%&=Nx~ANHBXPx^!$}*K+<EeY4gMo_{;brty9L3k%a;cB
zDo0lwHop4rB|mt-Oq}NaWXzjuj;VE)_pf*V;QGsxm#o{>@Z7P%mQ$OW8}|%-ySC=W
zo`JXHu6|ec@jGAr>zD8E4)))=-O!$9b~N4X|C`!yXZg*#k3Q*Bl+3=m@u_cT`KU#U
zwYthDMNicy_cF7$EKYpn7n0sdSv%$Ao~D)7Zw+(iRZSg784S^VcfOvtX>fWRaZWUp
zb&Kv66B)wShsaOtZySGDSTwLCj*JV(Y0+tkkwK3W-&>OU+`E_e#*yN1oEV)L6B+cl
z*BbZyr?2OYh$EB2aYl5;*vO#Axp!||tNUr*dvRoZIPMYMqf=zi<Bc)T=T;R=Z;m%i
z2*;UGnfAz_$ITl8tA2fXU-vljcsTAI-8&&N=<(T=FRZB~8xC|fj19->De3(pgC1Wl
zPyd+LuBmav5ss6ilVc)-9^Yi5PW_{$^t(7RCLE_mr*@4Ddi;A==k_!1eLuyKl5m_9
zofH!p^w=<a#>_ci|DogP_KA%g)#Z(MH#VHNrN@_9Sc{+^oiN!gMIA)5G-R}$hKyku
z$QU~V8OySev33?Rj^!ZZ>>Ok~%R|Q7dB_A-fK0Fpkcq4anP?XwlUNBd$u2=Avod6|
zU4~3y704930-4IHkg0YRGL6+B)9e~#R;xAp!CETIwq`#deRZ?3kyQ<+$4gn3Qyr(o
z;-f5Ukpq-v3oz}aXzbHIipD-2plIwfUW&m!<D(etGXaXhKI^4e?6W?K#XcLLSnP9N
zio-tVqd4qy0gA&u@1=O`^FE5lJ|CcX><eB>z`o$41ndg|O2EG8r9|wDK1#&C7@$P#
zOI}LCzT~4M>`MVk!oKXKWbDg6O2)n%pk(YTUP{5f;-eJoD*;NuzUrk^?5jRX#l9M#
zRP1YBO2fY9qcrSm0ZPOEFzc}VjwL$WIvgm^n)iV8>uq?!-oMzwf^}GmVYxaj6M?EN
z9$c#q8mhFDhALf6W0i3*P-UD9RGDf9tE_{CD(hsS%2u;j<s2MTIVT5IuA0Lt@8F@z
zJ9((`)jU=OhX7T<DL_@I7O*NhM5u~R5vpRfh*ilUK~-`}P?f4BtjZ1<s<Kmts$4B&
zRdFa#Rh$Y`m1+g6szZgU>QtesR;yUm92!(Lrv_E6TEprP>yQT*!T7QqYYxy*>(B?J
zFCr2fxp8-_w@1y7I`svs-^NR0Be(C4lEq6b3yJ<2;oF>_SXLniiDff#h*+*8w-bxU
z2!HQ0T2>%?OqLo`M^1YQjXv5>X!OxRLZgok5gL7TJE74>`;0XDXpf0TALAtq`WQc9
z(8mM`gFYri81ymigh3zUGcxF7JSGNxte3FpWBr6h9~&er`q&U*(Z{wE7JaPG$fA$+
zm{|02Uc#Y|^AiqzT##_+<3fZ(AJ<Me^l?5Thd$0@;?T!?36DPBPk8k4LBgYt4-p=H
zd^_RM$NP*t`go6tN1xy&0{R3$5zr?DiGV&KL<IB+?L<JI;4=#76FepXeWI6$=o9@!
zM4uQWBKpJ-5z!~M6A^u)&nTi#^q55SNnRqMPx2E9eNvD}=#xT3LZ8%5B=ku>ql7-m
zW0KG(dx?xb*-vEj$w4BcPYw|oeR4aI(I@+iGWukXNk*UIB?|f!KT*)91c`z^B}5eT
zDeXi-pW-tr=u<o<1%0ZQsOVGuL`9z(Br5vU5K+;mwi6Y7s?VsRPxY8o^l4tAp-=M@
z4SiaWXz0^IL_?p}PBipsKBI;{&12F!e8U!BHQc%N>?z7rhFACe@&CTKX3nn0&f)8~
z+42ZnZOt?gV`hL@GYiC-IUwH50|{mUNHmK;l34<h%`%W;R)AEq3Z$8}2Q+H%U)(AD
zYb&Vuvb4;!JpCL0*wIM?ep#6v72a))v3AjKWzQGoA7m#C8WY}~J|ca*{tUUW?auLa
zJAU}^emmWoskaxmj_7SY{90;wcVgGX6#dKDvQGcp^ykLe9ui-c+C8<m{wNt=);qa(
zs{Yife+CsBd8pNInr~lOIB(+l!0X|U%qYz$*N=9?++9&)s%i_uyL%+`NYU@OpK8U^
zqej$T3GeRKzZ<VV%QueAX>s-a_KWcD5vCCtx=L4fiH$rc>&ozXwOR6XQ+ye1qk$M3
z1H{@`AkM}C@irbvun9n-O$3r`5|C_@ffSnpq}o&<&87j_thQnL)btaT5F2^OhpE+5
zmXGkuZ#_jLGxZdM%+^yZGFMM=$b3D;BMbGEfGpNiBC=FZNyu_NB_k{Kl!C0*Q!27n
zPiY-|?t*9N%EWEohfi$!#Pn(UtS?DXw>ov{zAU`knr!W>9~Au+NBDf-f4>0-MYRlq
z^R3cAjLHD9DhtG^91ySaK!Pd&iK+-BsS=Q^%0P;$0I8}9q^TNEmNo0a_tGoo1|P}V
zcK^RRxqq^x&s6_c3(xreD;t=x+cFPk?52SjHv`1FSs>2M0r74gNN@{4qFV%#+!B!N
zmVp$v0;IZCAkD1-<yv#|^pmN-aq1NtdD8?_b6KW#OwCeZsq7%yMMFkc(2y}M1~R6C
zfsA#rkg*jkWSonGjH}=v<6S&td<74g;1VDcDg?+xmk60yAwnj(B*>%+2{PFwLnc?q
zkSQ(&GNnR+Om(S{sTC?@noEOBtI!@~;kSB~<@aCl&6z!IPRad^()&Z?$5@uq@^1Sd
zyXE9YmrtHj`h?3cMUM=ICptDKcWk;kHY1<<Kb%%Iz4FQMX3@-1MpIODCv!|(eCMuR
z5(*PX<UNuYTR1$a`yaaX=$UD?We>^89V#TVSq~4g^!<N(Q+oC8lm1Xfzy1Sgh7%=O
Z9XL2mOXZdP<djE8P9+7ShuNve{tLyWe0cx>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp.FC712CF5F768A94B.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp.FC712CF5F768A94B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f7a4605f39c46d90c0c669f1aff421c927baafe
GIT binary patch
literal 5778
zcmZvgc~}$I8pbmxA%p}KP(VReMVr}YqCg0u2q=n(s2CSC1aV2LjRIO+pb0LxfU(9a
z*aj7G!>TAMT(vHxda1p2sn^>oR)pGCMWwcGsBO7Jo@e0u<K}tzGVgicnctj~Gs|J(
zm{FrP_(&viS!41R&dD{(Boc|I@RyUDZITG$uu3Ayb6lEo=fgXZ9=}|BGAULwzUKDL
z5rJNY#$BiSnl?mx4bg5<r0&_aXXWkcdlgxi*4uaA-SZHw9e6cl-qm!|NOi@{IolVH
zOl_I7c-y?kHrK-Im)li4n;eGp<TsZV<kbw0%&*;X<#qVCzm6_iQE0zE?8_YQ(umh%
z7HPK6TKR5u(7kix-~8FO^k3H=l+KImv*_<P*;gOvAG|((=gbEM1tnX4uw1*p*fqVT
zKJCJW;H`?~m*1~L*VcH??VX!(BkiOr(`YGerux>-Kao0ahRM&x9NF>8K8pF`*=OZz
z?+h3jw3=(bc-Q;ZIzz@kt{(7Iu_4sik7CZ2Hb)%09-gvs$&0ctioSg_c28VmYofHw
z{rurdRl3Pv-6r)valCwLLHo1cw+_mDS~1M+aIBmE$sb2v-F<m*Q2-tL-O(GVK9gey
z`5jNI-1RCrA%3cLNd3lj59)n#d&sTdH>J@r2UiWvE<TN-J$IY4a${b2EH7%By!_+j
zxI?vrCQ%K(19R?7FqYq)e8la<f6l}e*Nz}sxFB82@aBM-%IY;Yyi(UEwLBlRcXrZ&
z8zC9ZrhQ+&_i5(gHQ(he>NVQ8a^q)ywb>hY7!$oh-o`&^?IS;!^O)ULe#mF+({pjV
zwqM`po2O_lJurqCxc8H_zgchPZ7jg28<%gM{>{Thw<;<Y1wKzn*<{Kca`Js5;SBTB
ztLYQ{XH2iF$qoA=xoOFX734bihr2KH&rZ0776r~nUdg|Zhgw6MMnAv3Ns+ejXl7K4
zGJZ@cw<Jk<<NNA9nsk@;GZ=Afwd2E+3ua!}{gP_h(qMbIq%C3QK<|=-d5y1W!@-#=
zJZciRO-N|1FWTf;F}3~npV>dg>~9}e)p&K>%`MAI=wf@=$YCo^d(>QPjn5vXkJ=I7
z(W*E++d5@J*xai_15@|@)$fDCl(^=f%Z#>Ljvh8wS84yMvh&Bc)}2?(wAWoZUQjDz
zrnN46aO(4l$8kp%u0ztV|2S<t^-+(dQP(OC=I#5T-SqE{1ux5Al@;#Yc44RWgE@f_
z&lcp>AAfryUmwO$)k`<H)*NyBsxYhR@`k^)k$(M;uBknA{p8lHmDJcWW5;U!)9A*$
z$wjx{xD?Ab6_<zDS<loT_cxZWCXOvm9{A|}a--`{C2!rl4fj@VpZ(Q-V@$)`^UU@6
zvo`DMLJe-{&j(+xEJ_Nl>0k2dIPbS)?C8pZd-LO)<~FSO+TZa<bY}6+vMY<UnNLb0
zaC_L_a+iIt{@U>BtU2k}^KB<*t@X>&c~5VTnEKyHu3&1`d254@W})$1LsQee?OAoc
z`5%WytJ_K+R%bjP-gD26Wx7)3wxsp1k7%&afWV*jWFI?=1>S$z(BkrU-%KyFB-hg1
z9@>Fo7Dt`VS++1U7n4ZjLI2zoR}m4cLi*_vqqp4K%~^yyh4Z6)yQ^tc($GmNG)1t3
z<UyX!BHEtS>;0A-Y5hutCJMHf+{?{bM0@zV-=7}%by>9vO%?2Z@_w$)BHG5^VqcoS
znLkK{W(f8G`G85zBHBy8H@!rc!o7M((gZuiCuE?rh_>UiwBaAgw?A-~OcU&Im+;Zf
zBHBl{xYf!1tGBDrWWn~7dn%kow6E`UZ2hC)qOS_23U+`zpr^BlcB?V?G&SU@Muk!Y
z+ehx><}9N99Cv9v^(g(83QZ7ff4RShvxxRz9#d-WnlJvULX!l$x4gHTvxs)~oVhuE
z9^GRfs&tNdN&UNBhgx*O?nX7Crp3V#ltvv6RH8bmTj2x_lDGy3Njwn;NrE6ilF$$!
zNhA^=NfIPTk{S{u$wU$)DS`q?N<)Dpl}LdkP0%1oYiN+96KRlS2nHk>4Fi%)A_I~v
z!Ga{KVL_5jWI>W6IFRHt97uAB97ys650bow2T4AW2gwL+M0D2_G)8J8fr`>bbt&-*
z5`N^)%Ln|(g_ybo>;5U#Hcb5~)DBF&A85P~!(kq`VK~g=4h)BRq7Wlsp0HsA%o7fb
zfO)bIBVnGjVI<6x4vd6(st}`Kp0Z&S%u^1Gf_b_SqhX%5VKmIs4vdC*rVwLbp0QyJ
z%rg#*fqAwNV_}}PVJytE4vd9)t`Os3p0i;b%ySNmgL%FX<6)k+VLZ(94vdHS;o9Nx
z-4kSlb_7td+So27-a~zr&U+`AASvpJV0z@MEpRrZ;K0Q5acB}L1Tcww0-9tB2~0Ad
zgeH|j0h7w7ph>6Dz@+nOXfi1bFqwP?nrsRSOg5i|CYQnilgsCz$*1tZ<nwuG>a@D9
z>&)F4qmAiO;*)2z()q*zgK9<ULIj4@j?`PBu0-lRP`4rVZ%}t2wM7c{w@KBTp|;4>
zR#~@?TM-U<xE<k;hgTvT^6)l<Lmu9NaLB`LQXKMdiwuW6!iosUBkYKPJfad2kVmv3
z0`iCsL_i*4lM;|eSY!m`kyb=P9%)A;<dKz#ggmkhk&s7rAQJLOo0Nn+(jp@vkFp{P
z@+do^Adjj<6y#BDh=M$-15uDi*`yTYQ5G2md9)SLkVo4Q4S94Wq9KoNLp0>k9f*cJ
z+9stTkG9BY$YZRCfjq{J7|3HP5d(Ql8)6`j=|BwRF*Yd!d5lHIKptyFEab6v#6lih
ziCD;E+Yk$RYzJZ?kF`lz$YU)s7V<bN;vkQ+BM$PoO2k1P*M>OA<2n!rd7MqkK^|w3
zagfJb5f6F19r2LIS0Wzr_%_5t9^Zj@$m4BN9`bmLjPK^f7Cn)?d6C<Q$&B&6<1@tD
zf4si^=dzqL^-AGoPN5zI?hXnZC_+I1MJh<3C<O%+t)PKo6bw+Tf(43Ga6s`2zDp%X
zY%%*##hWpAV^CyJocNwHdA!%)ho3~d2+a|05xvANi06!y&Uf_g8!GJQ(xmuriK*ce
zgndI2LiFMby!fqBI$z+s_f=P6>N@yPtHN-oi7JeMnykV|sHrN9f|{<vXsDSgjDec1
z!dR%eDvX1gufq6lz3Szgo8>2)YCA_FI3iLUiK1r-{k7MM{e)&O4=*3_@x9UYLdClJ
zYzuNX`uqC#7gq@G#sKdCKXJ7o9s^T4*DB&jeExXVvS}@jO~N4u%ova%j`g9uGcFSr
z?j9vH_v_iuM|`G<D;syEb44Q_M3`H8^)zu`cO$OHfg<z-P^6v&iqca+(RvyvM$Z7n
z>RF&TJqHx8=YfjWM(f1ho*3+ww)slMBcUIj9-e)~^FKFaVwOw7?#Dv2cbGRNu0-o6
z$2>9w{m>vZ$I0SC#e;}LDD=8>wFY{vQ^$c`>u{h59RU=nBY~oH6i~E|28z)!K(RU&
zC{D)##p`&Wv|4Sfc*4#1Eh9ehJbFjyN6560+2Td;3+wX2wU_Tkp*g}kB1lY&3r=BX
zbgnkV?mbo<vZ1J>ofoN3m=7zax@MQsx$YFx!W1y8mw~}B<3JH+0w~f<0!5iAplCA<
z6k}$9V$Cd2oS6fPH}gOZ(+-OhyC<%Qy_L?DFgRg@I<xzP)h6|VZiO3g;Nd15c!Yrf
z9$_MYM;b`rktPy&lz{>sWukyb8))FsCK`B*fdL+4Vt~gQSm3cH7I>V210HAMfX5qn
z;PEED%gdg#z}WfUf}HHxIg>g!W_A|mM`~Q#I<6`C-L_zPX1ZbKtPF`kv?P*Q-J9v%
zn;G4kncbVS7iKM(FUX|1$x@k%+*RSGa_`~MtLGTcgt$ST$}uCoe0%rl+b>dh#~v0l
zd<5-HL=6akFX;cZ`uO(`2o4Dy7&cgqlMKsghDHPq@uT>7Z=ca)XQNT$V-vA){{!A6
B<SzgK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp.6FC4AF50AF38A41A.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp.6FC4AF50AF38A41A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b12e081ff00d4b68219d09b34b1dda43ba236476
GIT binary patch
literal 6770
zcmZvgdsI_b-iC8_0)~sqMeqWMh?eHuuT}`6maC$Mq7p?aBHmDJEHYL)6HMc@YQ2O?
z$5Fpn#n{h}DMdxAwziJT&bO?#own+X)=t%q7ep)dR;pslgsgSs{bRBgd!1)LzvN}_
zvvYRNDIAxRvv82XkU3}E^2Kwj%S;A?A=3Z1w7RO|H>1Juewe|q-24AUi)~j%Ms#fd
z(Dq)VWl8*uHGK<i@WnIfO)-hd=iV7slwh52b$(d7;>xauf@!x0H+Afy*R<}N|4z;C
z1{#b@YF_dd4EdO~z4?cqU$%)O6IYz%Uz+q)GvBiR<ede%OP*AY+R|RRc|zZgS4_F9
ztnOBtceK}EzW(IjKYo|;_+iJ-L;vUE_mf)Av|T${XPVyesA}uPJ(GS)IrPf8Z`-tS
zGmROZ6A4RB=G&8ml?6j@1nv1g%K1!_cx#98&&_{%F+y4SZTu(MpUz)D;>|_OMBkm>
zt(R&~-`LW>_LF}s`TeoWV+Ul~F1*;V{=w@-ZSm_iH-uEQeR99;lQDnY@b1Cv;-;Ou
zua3I<cl*$U(?z-F1yzp^_L+C!^ayf0Gw#6tZPRKW-nly<YG23K8%ztQ9M9pNtJ^>4
z-pR7u{3#Ke(oRM7yS`J3(?S<{3bWiX<C8yLUCmdHIad&|qocX%?h6UctIN;EoL&>*
zO!nPxSoN9p-v6FHvdFye!sZDd#Fy5wmB!;~N198bcbqwrz3!m-Pj}w-yniS0L7&*;
ziy96Ln{vAF;m&bw=Og}7ZSJ@HGi^|i<)6g^CiZ$|NtN``;pUf<4_zJj`mjiM%53{?
z(Zx6UT+3RDpZ(z1K2%l1MB&bWuZ6t&qO9@Jy)PN}-QDzLO2G@St}Z@xW6Al3hU$0j
z_03%sk+7?1@w;<=$^O^O{DtMM|KvFK^+P9*Y*^+AdMo*ht?9XaCq@KyjEP>ofSzAA
zwg26WtzRw6YFd~Xb0*}^&tBiVx^YHf-H6(EH-AH4zSO&T-0-W_@6EkJV%~S$-IVUk
z+7l++lp7bYsb`KJdQ^D(g0F6FOBP+&KcjlxfPx=d&wT!J=f^wdkN?NKc^fyKT_h@#
z{vLAk+fOHamO3@%E#;%rcjtIsD=#>)v8itSDet26hVh-z7fsvWIQC7_?!o(l5(oF`
z2v!R%D^e_u71FeQhrZljd(b4*+^lI;?B(ZHZ8?|}8ahO{`9}YX+S#`0uTA~awO6yZ
zFZ*r#`$x|VwjcgFYXAJp<F}mXJv}4+)V;@to%Z{~8V|nNtEGA1ZuixTCqsXWj}Py7
zIAO-jo5kZt9F1M-E$I5A{dW3=<x_IMeH64Vc*FW_WvxGyWWBt7;V<bWYd%;sG~a&f
zk7;!;owe<H`1tzgZx71oecz&WI>+U#-T3{7!#huzKdasF#VdzR&$Rl!d8sJJH>qvG
z*`MOppa1U1<(}-!$r)?=T}hd@w0OhF6<5}uy5_B2xv{XbHv7oML+_ri`Fw_uxxQsv
zDEB0;Y2nrZquG{A$I4f3IVfH#ceMUAq2;@^M@L7#)9;u32ll;>5_=VX_3^g+;li$6
zzdyId75PS|+E`NQaGhBG>%J8^qbF_N73Yedr5-*1*9=$MM@wrK&#tBnhTz1*8!{W7
z8xltB#fGw&;Fyp=<G)S>UZIb7TwIy2O%5YPew`Sc7#V1E9l2-Cpx-s${4|Ud`gK%r
zR7jxF^)O|RFzZRv@ZN^0ew`GXG%V2QnlHXsIdAH#6T=PD{W>)$H8;@cdUWUA+y8gH
zW=|NI=GOy*2lfgyx?bLVtmMZ%zitg9GyFO^IC*lQ(e>VqA%}wF_G}3wQ~WwIIMN(w
zbp7>q@7wok|JFZ@yx`Xf!3li=jjnH&4f%*uZ!2NM=GU>ou_1v**Z1k53m>+XoDU-f
zejOJa7ZGT5{cA)~%lD4IT@NFZ{W>~0Iwa8Ox@zvcrJd0|bDt0znDh0=J9jkwU>y=(
zW??M6K0H6=S@L^`W@vD99u1CR7;sD;1CC`_aBLn6j$=4*TpkCGXLxXY9uH1n1aLy0
z08V5?aAKYaPGTf*Ql11(W@K=3o(xW56mUwO0#0RAaB7|kPGdB1TAl_jEj2Cu=|U_U
zl{)Gv>1&?}4Xk-MJ#Nafp6WRz77u0FitME<yMbvpMPohfp=hk9y%df0jGJPxp7Br&
z)-zs;!FtwBu~^S~C>HBkFU4X#=cYKU=R6dL^_-XDu%35QJl69bipP51OYvAQxG4ea
z1rH@)z2Kz;tQXysi1ng}60u(NQX<w%Zc4&>$wNt4FL@~m>t#13W4-L5WUQCHl#KO?
zn^Le|@lXoZD_%;$deu#-Sg(3073)<mrDDD2rZlYAJd}p@nwQeB{`u7BvwD_jW@;u-
zR%+H$(yzw;>q+;e7#6I}QV7e{X>sA|w9z1GCykP^F(4TygOateAXz7ilCyCjIVXpb
zxA7o(Cy!FF2_OZhfKs%HAVsH$QnE=PC8vZ^w#gu6r;JjuDIgW6f>O1qAXTS|QnP6w
zHK&F$BQ@jc<s`gpZ0gvjq;HLVLIYbOoU2C4atKdPBSj-KjTD2-Hc~7y*GO^5d?Upp
z3yqY3EH+XiveZaP$Z{hkBP)%Rf~+=DDzer{X+5kjTPifLwd=#wxAv0Iz?QCe`}aCG
zv8*JpR_lr7ZDb#@>_ToQmY<Njh{a`u_j`<%I%Jp0@|vl~({4gz9$imp%%gpT#yq;6
z(3nSe5gPMokCDba+GV0Kk8u+Q^O$<VU>@Tm4CXQIguy(fi!hkSc#I6@F)kB>d90hT
zn8(%=7V}siVKI+wCoJZ%U4+Fv)?;Kbk9C<?%;VgI!#u8@aG1yW2#0xGJK->o>mnTH
zaULUwd7R6{VIJ=$Jm&HBgvUJIM|jNR+X;_(d>7#{kM|gP%;Q}q9`gh@5in1vCj#aP
zJ|bYA&`t!*6S|0id4k6%V4mPI3799kiHLb(JrOZa^bryB#C9TLp4df1%o9CE5%WZs
zNyI$KO(e{d>WPGTl8;E3C$$p^^Q0~!VV>kMN|-0POcLhFZX#ozTu)@olYK<SJh`37
zm?w7;8S`Y1QN}#kWs)&ZaT5jelzO6Ip5h}4<|*w&!91mlD43^sj0)x{E|Y?Js+*{o
zr`8h{^Hd*EF;8tLD(0zOM8!PSV^lFub(vJm)7(VEJguH+n5X%OhIv{$(J)WzA{ypt
z9;1eNn#-j1<c(eRqv6qm>I;;qEH^rLmVU9%vOm1KdFhekz5Um7v*j7M+L~z~#>@b*
zW)_Gub3nYA2NKKzkZ2ZxB(nr0n`I!ytN^KI6-YB{Pif+F8y&ISx;iSnEOBIFrhXeR
zWl~hijknT+{M~6GX?^v<F3ZpQf|)j=z~7yeon+Vdb$!qzG;mkiv%}#;t^E1X*=-+O
zJlOq_L()d-AGzvIc6{nr>jwF|qavbW^`CWXdmU?7eZ0y=!pq|N$HnXSpy6c+F$sh8
z``Pzyt*n{;gLi}fy~#6@XX*b&)AA!h1&ep)_`3)885pbYzxqvb-_XF#anFBgv0J9=
z?+q`b?KBW$XMk8c3&hzuAl}Xc33dTUw2MHJT>_HrGLT|dfK<B*q}er~(W#@Y`XnAm
z37J~=WkZ{P#E}t^{q&I@RSV|?W$(P@?~WNB!|9Jdj!(&GD@{Ci!rz@^${DJEh(4?S
z`RRT{0P|zDWWxMdX&}bR0I^mUh_iA)yp;zMtOAf|6@etH1SDH!AjPTxsa6$8vuZ%;
zsp(_&p<TY>8uM1<-tYV)PMV%nuCLo`V$g$LefzKVcc;aqCF<4s^AP_;bUzG%iBK#f
zU?LP6h*20IR$+lSg#+Rh9!O9GAW;#4Bt-&}6&XlT6d+Ymfiy(}vZh*}{t~}sZt{?{
zz1?5;*x|9FK2rVb<sWhP^Bowm!%_nycF;hKg8^b4ED-15fOrQFBsc^h(IEm!4hcwh
z$Uuri0a6_*kmk^U#-xtP)K8}Vcq%S5@K_3lR%)5uGc-$urLu?UQW_jxL4#vT8E{Mm
z1CA|a!Lb!AIIfff$5n9P_);DmU%`VDN(FF2g#b=06~T!WA~>m30w+~S;N(&noLnJ;
zQ%V(ZN`(SWEmgs(6)HHbR0F3~XiuG?YHnp&_w&G|Rpm=3cOSgi-2`67vTNjkbMBWX
z?gY;+DSdI~EJLZT42GFKhb29Svw99^_Z*flo>RHNkNNW^8cjjLz04tD;k_gJ_8AwM
zo%u{;=(y*j`bYO0FmR;*%P=Zq^cX&dNlPB_Y~p|Gjg5;>7?Lz}*zgn!%?hHVXwMB+
U2XS&%Ol<D>a*{J?Y#ue?KZ>%q!~g&Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp.B22CC0B7C42EE2B4.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp.B22CC0B7C42EE2B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ba17cc6860b2fe54d7a1a1a5f94b3ae8f91bec5e
GIT binary patch
literal 5774
zcmZvg2~<<Z9>#MgAwq((Nl*~k1bg=-0>&W7qN0d#L83(spjOgK(}E&c@=UPS1#4<?
ztA<)ZpQ5<c-Rje#LVMbZ;?}zGv{tLNqT)g=t-MR#x$uAIyqv?F`_1<yzqvDa=FTu-
z)QAxuX=JkGnWKvHXU)w~$Ye4f@vm@hfw^5SlkHH;WW~<2nb+UF6XSii@wqusJGN=i
zJE307f*%KNEMN6QxI>Y?azIk=non1=?Jt5yXF3Pt%j?f-xB&&dbcOc_ztJ9BneSHk
zEmyJS<#AaH<9$CITmSDrUfjHU`G@F}8+tXqRQ$T<<;7Rc_6L9cQfzpy=-9yGUAaYV
z>lO1W9_|b=z4!D^d-SG+8-`bRON;mv>{rw5!uXR%OP$RLH(I~TT7A4nX|O||U9U1?
zWZAgz`;j|us^S7N*QcMik1}3nmM+GommSXjvFgGIK7XsZcj>Ly4VJ(YPwFg1fh`7i
zt8e`+KZ7wlV)rSca(sj)JL%ej$E6F{uPwQo^FNPyc%LsZeD-Wj>be7$S1Gx^yLVNq
zZ<vGd&*eb}_E%2+$Dhw%IiqqPZ5`^dE72q9;EkknRo@1z3w}58!rrFwnvBG#!2Oe|
zzjzrsJY}+dP+j@bwmQw+o=W?b=1ENa&PCA$D~_S#K2?^PbK{?TFIdu?v0$Mgx%N=h
zM7kk%K;iXqIh8jvc6%JS_*L?XL&M11e2D(G;mJy7@zo^rvh8PfJt?VZ+83Gez?*wA
z=wf#7^nF+A@{S)_QeA%GY<kbEqr)4y0n5CtVMX&)^2<l(^*KK_Q@yq5p?6-?fd_8=
z(I1g-H>pOMx7_|9G^V*~)AG#q{kN<YhV(1Sd&AlA*Twx41B26dT^oI&W#f?(>n_!=
zy|Rm%=3%LrAN;7lI$VA%?@Q0P{IYL0_t;=dnW$6mJhrs$Lg~Elq)9=w>&G5J<C|uU
zx%p|?ffwXi&-vKvIj2TbVcW9ktbo0F*7C!@yzg`3+==bwTdMZmAJg7^H|^jDUP+p7
zeRrfz`{NQH-dmlpv9bQqv%pRGr0bt8dvn5`3ya_XGG+U0LqO8n$(5<+?vHLLk4OsM
zJdZlB+HvOlqxBIx4xG$Re`0obRtmrSY}|kQ!L2-7g#WYot6E?7*!)iGzLG1AYi~uS
z4SgIpKYFh9n}a>aXUnEckEPmne?9hBU50<-^T2h7{<@m!b)ca)bx`qIHI;l;YMXsp
zySwqaaZ+nj>)YGrhd$iE)sDPZS$qHSqKdD5vhu!DO-mb9f4J&H{E(se>i7GX&7Qa3
zdD^3AzrOn9sGRVs;xTLV9<jN1Hl#K!nYaA!Hn%m+&DM=G_H5I|>jxIjdC`FK@41z{
zN48dm)_eWzyObJP{kL<)h{~o5+Y@2~S2=!q+P_sfJI>MP=YY4gH&*s(zty(-Ti=C?
zkoJQkRo2JXlYT!wYuCEX%Ad}CsCn&{8GWvNVeLNCsTt*;9pb$99d4T)_0yx_3lqGi
z_ioF2elq)6$H&#57wgxk)}}2BnKm1x42i4P-Obg%-gNrK^{4)?`{epIB(x43eEb{J
zP;XtStgs|McP=KADMOB|Ox_$Dsz$~)WjX#ze-Bq7t`e?~>d!Bl7o|l{RHIDM4pD~q
zxC+Vk*|RKg-tON{tI-6}_Eq|NxC+TmXn*=>`@NDqYBX82`zrhPa21lh;v+}Ur6)0j
z8ch;yAEl4GtB~wfh3ZUh@s{YGvJBA<(}WFh6_ULz^0%p*jy10JluZ%s7`K>_u0pcU
zZ*s0~{l}RAH5xD4!OGxXu0pbZ%LzS34|=3kBZFvblo}6LA=yud4w=>7+`3YYvP3&V
z88OjSNcJ<r?b^`?Q-4&WaiSfh4DxmrlKr=L=9f1uXYQ%dMA7zB`gynt$u5}nR$-ub
zci)GrTz&p=UHcccxAmc(IXH<kQVT?-#bba<#nZYKK@uQIXbF%cQVEbGNfIPUEeVoj
zDhZMlNr5D#r9hHOr9hG<X^^C~G)U5^G)OWe1CorE0ZAs60ZEo*L6X(7Ajzh(Ajy#&
zNOD>ZB)L=$BzcktNnXo?B%jKIq(BNFDQE?d6jB9{jMc@)bv=SPgLQ*}8loH0rKA}o
z4sz%80fStI;U8n&Ly0>ud;`=@4BrMcQHBvPPdG3F<_RZ8z&u%okuXm>FcRiTCq}|N
zRfbV8PdP9O<|!ve!8~1t(J)UtFdF7*Cq~0OQ-(1x&p0p!<{2l(z&u-qu`tg%Fc#)n
zC&t1&SB7yg&p9v-<~b+E!8~7v@i5OjFdpW4C&t6PP=*OGFE}s(<^?Avz<h!(A*K5P
z8KxTsRH81iOG#^JfXcOYf(OKaPXLdH6}Q2s!9W0$uoBQD4J0s0D+x`?Kmn7oQqZIg
zG%#r^4Nb<t0F$vY&}0oPFj*@LP0qjple2Qr<PAJ9c`FZ1!5{!runN%B>-1go%rhrm
z7vH6%H_u3w>x~0C)sFBo1iG~X;j5vpM)($}TM+&j>JEh4<ly@bIsOUMHU(~1bo+!I
z5s*hzAOiA;YD7RD(SiubBRUWPd4xkwKptUJ5RgaO5ea!@1tKAjtVSf{ku8XXJhB6k
zkViV?B;=7c1qpeS9Z`@+RUiuTsA@z(9@T;<$fG(C1$mT1PC*`JQ&5mc+Yt?UbOoXz
zkFG{E<k2mNhCI3h(U3<w<TT{bHU$lNj2$tM$5bE&@|bGGKpxYA7|3Hf5CeIPL(V`R
zV^c7Y$J!AKd29t@A&;#_Eab5*h=n}11F?|DI^-<mu{H$@d7K?_kjGUZ4)VBa#6cd{
zf;h<IIuHkWoI}n*9%oZ<kjL8*4|#kA;vtW(Mm*&4Er^Faz60@)$2;Ua<ncBI4|#$e
z36LjLAOZ4(Y9v6O(1HZW6FQIpd4fYOK%QVz2;IEICBMsFJ)e6GQ{<%hrDRLXe~R&s
zU)C3XRi_d+bMAN)SRLF6ph$NTD9W7zigu@gV%!;^Sa%jE&Yc5_cjti;+=VWc9{Z_9
zL$CM*^UMhu9Fi>UQ!>W-4!plA&P}Y2^@#P7Zedn>;tBh{f98tSKAt|l($>8wE#-T1
za?CifI&65DQQEREdhzN~<-z7doj(#9J6QUWB~OO;(>1OL6svu`eKpdXue`^p%}eVF
zY{)YwC?Kexw7K@o3HA>Tl=krw(NpC*B!Cs&h))4MY9xRnjU-T%kphY~(m*jr1}N6Z
z0>v3Qpm-w>lwcHqiqpmErJI%dLalNgTR?+rF#IrlmDgYd)MO1tLQU0R6x4JLMnlci
zU<}l34aP#v)nFXdd=17!E!1E_x0afZsa(e?so6UpE-IOF+qqI4x`?cZY^lR*i@$Ok
zmtQqPtnS;ZuSR<9{%g>LnQp_Y9*Wiek^Z!F$gInVe_#r^+8|aZE0V({lr+TR&~~2F
zB&0Z5^mwvF^2{Oh1W=@&1d7sAK+$>{C`Qi!#p+q0I6Vgxujhdh^a4;ioi0&o@U1(x
zVXJ)hUKj5pY)aS+X{7IRZqIvo1(b``vHr0kQd&A{iapwS&XjKNp?j?Z4cXf{|1^;r
zPD)Ali&U-?r<C^mp9Q6Bil|)Y&h8Ly!AqoH5PQr*07Y6zpePFk6m6k_Vk`_$tc3-N
zvv5H179J?UA^<g1H#Ax51L;KUuX3G)L4!?rZg+!mGd`zV5hen7gqZ*yX(E9~nn~bM
zCJK0znF1bdqJc-7Y2Yy?26&8_0Um2&fybIz;Bh7nc$}F79&h4-$D4WJ2_^w}f?4SD
z3TDm8>HKd&VZn^TiJceII*b2)Owxr+p5r}r!Lfpkes3#tr<$fs&z4Q?@?_JyFSENZ
zbGt8R<j<UwC(5+9(&Y*_We;}`wP#OnuU?~kh9^h)s759E2Ke>v(|54A#~vD=FpTjh
wheX7@5%T|PH9`G?L&L%cL=MCWisg81bZq}YfwYj~uNgUd1{yIoF%=v0AFnE?IRF3v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp.E11E98EDBDF90D0B.idx b/.cache/clangd/index/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp.E11E98EDBDF90D0B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a16ffc18d741f4c178909fa551bdab262d369ecf
GIT binary patch
literal 6766
zcmZvgc~}#79)~kOLgZ2e1S&`Y@63IwkfKr$VNv5MYdlZ{uf~c|JPNDS7!RseqsFbe
zjaQ>pto5q5Tf3#LuC3M9wXU_>R@b9mfZe*K+G;jrpMlRG%kwx--g&>|bq*$(8CH;&
zH#x~*$eCDBSw3lYsnK9C#037$nq6Uk5@Imyjx-o5y`PLIx8CT}{^u>nt?$&DXLK5~
zEPCjj>4*PP#4>&J_GT6A$((zB?vsNbJUd;Gvu#FZiudR3^no7^OxavDxT7ItM%8rh
z(6qm>7naX|`nbP%tkZ(e_D;Vb?A>>8#o31wtuvmK_1jpVvf<pSc1^|}XXd%q({CS*
zE|@W|YR->0BDq)RfAHz#rdK=<=J(1<UYqvWsUcIoe(>bC(u!9Wcs6~%`X+bne&s)@
z;p0w3RCQb&n=Mq;kGP|hF8SJ+a%}b=>RlU>&n4X9lKMIh{l0jNJ!whfG5@nS^Deyq
z@Xbw2Pj)f6>z@1L#k&#Z+v7*?m?f62>c0Kmrt>?SVxM+@<7$Hay;YmnHf^w!ESp_0
zCMA9AS$lcPIs49XV`fdSm!?<_{PyVXPv)*=^*ge{FYQ{qa`g1{DQ9oK_t~l=yVCZ|
zUe~SXxd9W7Ozm*_`5mtrnRR)UF?YvZaqZof*1vx2t$%qki{sAhVzwuzKgrAcAt!px
zD>n!4XliO-c39guHRDw0QCCch=gb{`ynFwhuhcwO9z$O@r&_LCAGLd<?&WQVx)xm=
z@ylBUcdx|mo*mYq@^4yFsQL2f=LRP>&8U!09IKtl9KG3bN!J+Hlrq~+(aG=eIp-c6
z-ShC68pH7~-cGEHy0d<J0n>TMvZaj+uT&PkO7H5KmCH?<-RHp<Lo=<9YdjNPYd7qj
zBf}2<N>6umJA5GW{)mJTb>G|^JMiQg*WY4p58mvn>HZp-w>3Oh{nLB<cdw{S{XDTp
z=MjguEx)^%=@2>c^d-)mKjwpB=Uz&Dc)|4U_SlLOUmcvYaQWq<m4%xVzPw+z;=~|2
zsqp-v#<8c0PB!(NxRG76{L&J4N*7bnFOR2XPTyjCG^qCK`9T-g{;}=t(`Q4ov*vAx
zJihf_*NQw_=H?V}M%2BLYvYF%Y@D^SN0EKO={K+ay?vjI+n2f=x%SbG4?lF~(YG4e
zTlLYVmw!H0b@{>@*ScA=8t31h`rwK9_g=Z#)Xf3kKCC;rG&76Yzx!bEKD#~9fAwCQ
z`k6}#IcXlX_fJ)st8abe8MNidzNu@BUkz#Oe_6Xx<o>5+)Q63Wd>>L<UjD3bY~k6{
zy~`HHP2KYF?TtO+toM3mr1nr}OnVS@GHQM2<uyge*Iyq!tvGJfll?<>`A@CB_~thb
z<=%hPGdp)q%X|0P(6yhxHL+>>C#RIUgV8g?2l08nr{34|HeLL3SHIp#2b#YA@%a@?
zC(lnXe|+KG?L#RupPg?suaA7v_IIB?I1^t#@B4(ZA2`SRQQD_-yB~Te?B8u4_uGV(
zPs#IR&K+9vIA-3w1(Vl)JMK}_PuE7x{2<RC*~|ZZpL@%W4zInrWZk1L=X4*x_<H*1
z=EQG)^mo|%eLLaKnQ)K#Y*|0w^-rFWtkjJ8v#QF=W>W@(DfQT@oNc|+B8hFZp)}qU
zA0Au;t`WhXh$maWTU4kGk0c`kI@OdK6I|#zX5X@;IeYJY6iJ2!bgU^hJh;$xzh8f>
z-*vlcUnChB&>c-3+XNT79(1pmDXIL&^(e!rfX)cb7#LjW`t_ym<Vz2G(UIg20UcwC
z2@5WCy?w>V+gHE($<)?R6wv7j>0N^hU7uk)bz1P<(oK<MOh9)rbr~L9==zH--gW<e
z_2a}y^80{JF{QK%E_8jbH0=bZ)+>?38qf)*gz(@(*FVvrSC8K<z7k1>26VD1xqWb<
z>!<BU9Q@Yt@$E=5JfP!DapA#*t}7-@p7m>7>%LEk2=4Qhb-!-ibKQ~_RcdC;y#D_W
zP`<gam1u?rN9WVv7={7I<TKz{h6Tswv*0*}1IOia;CO}y$LI6l1V#WS<O|?LMg%A3
zi{K<i0w?85;ABPyC+ExH6h;B3<SXD*Mg^zltKc+71E=L{;4(8ZUuZcIOS3YvT1Y?l
znTX&y4~NG^nO9P+hs5lr%zsAqQs$k&w2Pu~o_13-&eL9s#(Bm?F*wh-DF){mFU8<I
z>!MhkXWbNw^Q@O*ah`Kg9L{rYio<!%OK~{QyC@##c{j!5JnyA=oEKb_fb)Wz5^!Gd
zQUcD4E=t6C(M^dsFM25v=Oq^<;k@LgB%GJLl!WuLi;{6(c2hFW%U(*xdBsI3IIp-V
z1?LqnrQp2kqEwt$-IR*+s+Ur6UUN|z&TDQ;!+Fh1X*mC4#*4YFr)W+_4p44JZVTyG
z<G}T#`BDrgtkpaWPS<L)6VFa74U(>=Q8HErBvZ|xWUVYnwwguBSvingHHVV7@*w$Y
z9;ILvKnm3YO3^BU6stv)l2rmJRZA#ks|-@EmQgBJ1*B4~pj53YNVQr;saZ9UTD69<
zZ${si%Slvec1Cs!=^u^75y6i{I94uVUPRz%d5C!(vX7XzBR3FpBeI{Eogwi1?hx~8
zWT(-*(Aer}7oo97dkBp^+DB;Y(G7&g9_=SI_GotqjXm0Fq_M}i2!lPwLm2EaKEhy+
zX&?;t7(ZdK$GAfn>@iLwgFV(oSnRPL!eWp05f*!F17Wep`U#6Y)*ZrPk98VZ>~Su_
zVUP0=4tt!BaM<G-2!}n+PdMyx?hp=poYTl*k9QFsd%TD6*yDYK#~$B6c<k|h!eftj
zhw#|rokkvef{O^)6Ffw~p5P+__Jjr^U{CN90egZwM8KZlGz!=gT|~s5=piEZL?02c
zCpHifd!nC+*c06$BKAb5QN*6)A`<o_50S7Z`G|x)sewq?ll(-&p5zXZuqQc<682;l
zk+COxh>ShiM`Y~D4MfJC>?bnzWOs;+J=tlLv8T9*f<47U6znNJqF_&HAPV*rKT)u!
zxI+}|DNdt;J=H~2?5Q52Vo&uE6?<v}QL(4`iHbed9in1SbsAOdX)dB+PxBBBdzz1E
z*wY$_hCR(sH0){a5Dj~p)2OxfvTGg~9zU9Wl`@tNj2k#!|CG+P{c>}|tdn(Z1K-}m
z%-!K?8%6^$VGIx(#sYC+91tJI0|{XQkQgQcNnsL@93}%PVG58MrUGeUS_|o$K}7^_
z$n+1V!M)cy61dk_Q&FX<eNuDu@5e<$Vtd?K^FnA~b!K>GwEoMJ24}y&@W9(;fz>fl
zF|qpH+}y(4bIizILj$YR2c+BdySPRFd3<Tp;qQ(#ze!qVAN@^g9uDY~apCo(!0Ooc
zu?hM&oE2^Uy6v^P3MYvwO-@Yiq~DH3m8Qg}B<Z(ydNDd8c;BdRMc|Ow%%k9t*k~Ze
z#sIN47KpQPK)j6y5^MsHXcK`Xn*=1=WFW<+0I4<=NV91`{WAJl^xdNGwRRD~dtTVI
zTFQJBzueVQG%{05G01E!#UgXH6o<^$QarLyO9{wgEhQpLwUmS`*HSXFQcEewYAvNA
zYqgZt%J1A?R5j|lcU9o%cNx=Vynf8~RGthSTK-mEV0FiK9TW7gQm52m6GI2Qbw98=
zzFRz}KiQ}&>U*~&^&e*ftMiO`o%LO)zZr=LJ{M@+h4^=@pGtvKsQD2=>yX)E&e7i?
zs+6|SK#YX}Vl6BXXW@W&3lAh%1R&8O0!bDLNVdp8ibVlZEh>;^(STmacp+Qg{%bd!
z{ny0o|0b|o(nqCF(9ft_V(6nb(TOVpt25&>Q}t>6c}QSKG(QZ19if<e!H!U9AVy(;
zScL`R6b^`2cpyO$fJ8+Ek`xI@R%9SWQGirM1=17^$dX}c`6hlpY>%5{?r*-TC3H&=
z^)1z}DS<6+e!c@+>@Zis7CUGl#=!uw4i<=Wa6r6+2NE0tkmwMBB!>hfJ7gfmp#Z54
z6-aYvK>aiN=jaDhe>{~O5qvBKn^s~jYuz-n-8`+8=n@(nZKuI8B@8&m&VXY}Sa7VJ
z1;>?e;5a)6jxXWC@pc}ZP$Ga6>;gEkL<A?=MQ~Dy1WvL`;N%h+oNSlDDJ2Rx#jb!;
zOH^>GT?MC=Xy7!v*5V8mlctq6KM$N$F=5v5<`-j|7ytVy(O(!sO6RTaEWPnXO_r&w
zxMb|O@rL3SXBgM|a(wH{veuUq$|p{HDS(BO2ZtC#O>M%$Bcs~3k8W2GGa#pXOhmzp
zv59dVp6l2r@W+tVw_krgp2_Uet7q!}Tbq#FDJ3nvbJuP?%rq;AlA`r~UQOcU-1vln
OznehvhGgecgZ>LGOVS<y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_cshuffle.hpp.CE155CFA093515AD.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_cshuffle.hpp.CE155CFA093515AD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c85d894aca9093fc94544fe4d0d6c20e5b0472fa
GIT binary patch
literal 12684
zcmdU0cT^Qux4-Ai5h=sv(k?|Vy>}R@qS(8lVmBsWK1=LkuTg{2#D=0#31aMs5=E@2
zfB{jI-c>*p5Rn$6MpWLxd$%*+%OCmUt(CPNOWpJ1>^XJ!Gfo&WbZDrXNHpj_Bi1aM
zy&{;1L?TW4clnBtAEJARL}7X&(V8P^lV|Lz^wD`6kdwK%<b0!2MhZ71&~ou1bEV$%
z%Tt%c+PK{{kFcmph@1SWbwR^+S;i2jq|nVrS|4xRaBFkb?OUtYJ+n)9i%j;sIxN6L
z*MEBHchzxmYfeYp=+l01K}TZThGP-m{qQJz_|X#^?r$zTw4S+I>gboZdGV_Jd+LQH
z_IH&&56yenzxN`izS;#3;zNBhetSRAXH$I8c9O8{{E5|d$+4wJ3f6>7Z9Tg>?ALo^
zl822OertC4q{xu{Q~E9`dbOuK`@4i>=?}+iDws?M$D4O_=goZHRXNURdvVp(=QGf^
zy=p$9OQX$tF3FWxzt0lQRccI8U+Vhm`LmrrkF;C$cJF(?TX_jB->T}am_2>t+d&bV
z23Fm4KUEsv`&@LH>8K4k<8r4QbYiye@mCFWcD<3^KAjjns$435u;N0VyUf*YX5G8o
zY{x4%eE-wn9x&Fu{+mFrsvD=eG}mvv=oZn@r>x)k*XwKtk1mZr{K6~F!s~nI`ChaC
zoO$Zq@E@i&R(EdbnB?k_=eFo(*j&$*<FBRlrrn--G9|3p@K9ya%x&6Z()yi$QeJ%6
z!fV)}57tYX^5->~S08?Tf9AEKlxYY2x4BMfX|{6_JM_u&at}$8mH08imrZJpy>~nR
zY0_%FwE7F*zJHXgF*@?v_jjlC8lQBjcb;v9{_+c^htADymBd?U<#m*fKcusei*E2a
z);z-LkNG#oKJ&OZbhX*W!NJVOU%S3>E1ah^tW~ciBzbB^`kIfy&#vch)ZbxU8mm>^
zF)w%9pLx@*nT<cIhy8p!D9OmFGkK237|UO>o@J{oonccLJE>vZm<ffatPeGhurDm0
z>3PrNZ1M4xu^z{~m2JlSa(Le8yn3=}v+|5mP1S|dY;y7@o|%`YGliMCyxj2@&nT<3
zlC32@dYpLOvqF3Rlp80L82<q?8w_4{{&{rUwP_2|e69utG?qy3y1mZoTY2%&yb&gu
zmsDQ*{mLZu9y;){o$sYdr@|W!zWr@-$-R@a<A!jZ^P|c4Pg}n`CUtst?Reou<KbHe
zY1QpIxa_h|N<-|4sTqIb=2K>KZk&p>(%!_nSvuG*evy>=^wy-+73<#Snj3{oV;?_l
zNxe9~(|ydq4OOjWD_5<bP<Fphb?bt`M~mW0Q}(`$yi{p%E%s0DW9g5dmlu~_6kn`V
za}9bmanTaj71m3iwrqABG54ZnNa#5C<uSXJT_<ci8~NO8`ewIp6FhV-neBZwz2BmU
zu-1Kn&58G444ystQb^Y3y>qhrWjbah?64a$aAmpfj=L2_Hut?2Z}mI()-S)8nn(4V
zRY8?XMdE|KgDQIztr_KgaEsmQBFlu)LAUJ}-}fJB*fGpw<nV)@e;~KD6GHZ;pPMN*
zKA2t`<h!otS&!wfd@F0R!g`LJ^<rSo{kF<3tytFk+h+#@mHHibTe_^^`nkd`EfbY*
z-gLO#@DslYb(9RX-Lg$Nd(&SB)b0}U?Gm-(6_L4a$6xGwvu@KW)3m25@84XTd3Q)f
zVw6vn$?>SGPLD=?n4WX%-NGvIywKO1ElieOysZ792Nym|r{A)SMFBfCUsPPI_|MvM
z%eNWD*{N6d#Q*q{<+aVfZVo|#zRH9BD%N_s27CM^Yhldw25v4n>Nd2ZRe4XQ>);=n
zd+4c-J952p<qpfH$FGh>m4uA7)vp;Ru9<By*<#U!n3RJ*g&39oH8FEx-Ez<31BSQ0
zPmh}SA@P$*_{d|H3(s{Xnn*7X+STCwLF0H}Q1Dx?_C9q<R@!sEixekD{*X~y8ayTU
z+PLP9LnBXwXNBzD^W$$LoHwfdX+Hb?@rdMUK^OYAp4h!)q*2rXS@xWLs(oH8Y^Yl@
z*VaGvZun!nj=udq^w7K6w)@YEQ%8j-`I55_caNGp`NzZqmdajQvDMd_w$w5k2JIbv
ze#Y5%?rr{&e;nE)TE2GCEDI%(Nc~FCh9=#j1Xk3~TEv+tn^}MPN2EfqsJ|8Bka}U#
z0@oF&rxrG~5pm5#(oAFuB2uR2o&JcIp7B+C?7Loy)$ByveWZLJ87c^{Xr%pAH1L{r
zkQiG~mV3zf9<o&sV9{RFQsK8}#z6@-vlVeyvHDePpdi3v;J4Pl@>#5@<jY)IiApPx
za&q#O<p;}&MSI=Nm9QF}?p>sI7s(U^SWJFzk9}Pn(IUYn_9Ctci<_{ff&h!vb#3cd
zt|zC2d(d7gky<6vRuEvZKE7b@(Gxx;x>%dCL?Oc{WTYU#Vp~^hR$LTvT8#baifXZ>
z7F#L^u-JvP_6_m)wn~h9Q<gexTZerW1X%1{wAFTHZs=0O8g!s^q?(Q_6yz&QnXT7v
zfoYv;SWH=VA=O<-svyAPoNm=`$9mmO65NZjq!5xq^b`bG`c8bUIw)4$tcz7BOBs@t
zA$tV@mI1OqqmqoBX6j-k%5n$E?jTDA0hUn@lJTyT8~(ajLRrF*VL06vIRTb{tD{<;
zFN=66!7|E{PR!B?ry#&G=jJJmseM!(l&}tEX~b%cSW7{GW$q5^^)}~PVwJEemE3hb
z*V%pfN2J0aDw9-l=P!z0rc&`ZT7p^HM>dgUQ^}PR=wpq#-<tAgA6&Jt5oJjxl4R0L
zL4aj#;NGm@S7~dtu#~dAVGQ3eo(cji>ki!-y*27B74E)NxU-O67G1KO0L%L6eLTi~
z`}2GW)}_6~6P0*kq#(euIds;Dsdc4}T3=W`Fp>|91t%xKvPEl^d3s>n2Py7BrR51y
zeu5+l0xY2)ZRW<5Z);Y=26Uhdq?&>B6$DteKUnl$>siNGU93e}B9UPvovxe!OT^_#
zUCZ<iAC(VOkCf}_Kym^sak62jJ`LS(sf&$iFWZr9JF-!buPi<*dgs;T)QPb^oo+Rj
zRAW5_0hW{Tr>{rv-e0eVm1!?Gk?KvPry#(RI6AXq$>LKd#MqDaQiCNm*hoQuC8_3h
zjYs)}0x`DyVqGNn=!)*4ZUQW~=Om9QamkBO!#(LhX-G8<87l~|WCUmq876D`PJ-1b
zOFdDmr@|m7z>?oX1qG)xwu^BeI+y2I@*Eo~2(VN|I|f#~pW&*DHEAyq$S{J=MNWX_
zX@im9u|W<4w6Gy%IZh<UsW8Y1u+;8(FnE5T%wBF%A0Xui)TGJ@u)LTuYvjVq3$Mz}
z*lR}on$hIs1X$kB8PF7_nG&Fj_2@v+NES`cNKSy|!#9@~mmV=bBgW=*>#~t~Hu6vq
zU=fAPp0oUzB~>b?2flc6nloa#GmPyS#tST|$^tdDn%u@=w86~XY*e>T=>B1SYKrP(
z-&A12eZsb%us!e-E{3s+VQc`ayM`Z{zu&Eju+&nRgd;03A}rlGlL%x5o61r-HHk!4
zAi^xQxF%7^3XHM(Ttn|45{>R4!c`Ntnux&$fSc=~`o+DCuv`_j=lY;cEO(pHzs)a%
z<#HJP93d`c^h<@fhS9I#F*rx0s&I*5l)xp@2Dv{pQ1~Sc{o;YHMNb#d8EzrNw~!H_
z_TVA0s=_bA8!T6d#Dz#7L_!iWyllO_!D_nMVc0ti_XlMK2TI43dr$CUx&@}TTVnG*
z(>KzQO*--im!g+pnfzkVYJb}B4jI2gR?wAx*>B1pPhD~ENw_#>KpZm=PKt1eOut0N
zUx<e#GQ)Y?#^|*%`XHQyOCS~r#1iZu;Tp-<MiKzJM!0q&YbVCw*$p)Oa{8aV^E;i9
zoy^V=jt!dW(&X+Q_Sqotif{YDvALM#E?|QTSOzQT#`YJmGho<FRaHea06nn(>ws8=
zn!yV!b;mk&*cv=FOLt^s9P<sVnC0#;-glT0JigC(-xuOC#=A_2YZ>oaA%01uFNrPe
zh(SaBklEV4yXX!$*gFKk4j5fB3i@ks=mCaskBHqPVh@r{xIAK)C&c;0E}zH7+kZP*
z?NeYtU;Kz|J_?mEg0YEU1WLJwvDw386Y&<`2<6;bs+4a~uQ$jRwvy$#ut^to<X60d
zG1<Wga4cgI%VXf9hbIrVz`2{ev`JN2RnskgGLcMGZZt$9v+E)Anm7Axrz-LWN#7s?
z7}(s$^t5Bp$Y9F9o4D;J{Xloj{Zp=W_~fhp>*5L;rmm`p4tc4PduL!^xmqNyMN;tK
ztiqe}?M+y}37f)Rv0NMD(8h3}M=Xs5>_0H>uws_#m3IR12fbpsB%+Z-^uV>STpBS=
zBeoFsu~hHO8;Lzg8B6VzaWk<71IAK4bZ93WD4NyB;PTDC4(&}%cPQf;%5X3-!fj*5
zY-7d&2FptiR~`ey4nXkBu-spW{R^4GDYD#8Sm!7FRh(26Au((PD4KG&QDWaAfl8mp
zlJnS%?<fu1q+uVhdsGv#><)J3aXFThV>2FCU`d4#S7J#ek5v_cH%tVqB=s5cs2i5X
zQU|G=jm*GzvRp3em5V&ULQ@lgb>grSKhYwrQ-qyhBKC@KU0YQVr1Kr$L>4!Z3q(aM
zjoe%_kRJqf)ZJnGDD1^|yaU_s5aMWTA1%Z?vHeaV-i@7hV_*LL3~Zl)y<lQ2mx=8&
zg*XfAWMNajy#>3p;68BpR4TE~2cgu4G9IB!f4==T<9(YM#N!;sJ4c918Shddu3@}u
zgm^cx*iBqO>YW4rGt514briJ>TZ#2nV*9m0k=;-!l&3bZqaJ<ls+O;vu0u^mJt6f(
z6%@c@+~KIt>SH2kLk8-Xfri2!d*ofXRoV9YurG#|eZ+?l-G-LU6XJZr=JVKdnx`H&
zY{Ey{;~&`U59|P|_H6d5SeVl`jy9MO6J0QKy^Y)NFHj0@H|>s&^gm%k;P1vxpRkJ%
zPyU3b^7u2>|19MHj2%8>C%!$B(T`*V{Y5fUBAID?dlaJ|C2Wsk9HJP(`+FJvy~6ew
zMnC4iv3(40-wBsX*j#=i2=|b%4~6&<ae72tgdWQ&Pl)r0Q@#)v5Q75Ji{~#S28BXg
zL=1|AxR@9e3vmh2E+M-7`=vy?REW!nUKugq+slbwxe!;7UKK>fw^x#0l|o!aT&svX
z-~O1mJ{IC9#OaA}Jx__#Qz5P)Yz@CZem-eus&m#;E6IIEuAh-Rs3#=kG$fQ~;~>@D
z-o&#O4FqojX*r{+NXB`_7~DGs4}czNe2aVM-~ph}1Mgfn4>1T!{m1ZW=*?{{3!CS}
z(_o+z>2{)Ce2-~ZHx2iK9;xHPx^1`@2ph{q6U}I%2bwy(IMHy?`=S|Cy}Gbr7nT7d
z;dU^FI~W0uWej6^3@HIsad|ZX);K1>TkqbGgX8HMa+wLa%p~ZMMsCc6IvzvWf%YJ;
zIY1AADnrs1&6ez>i_0b|*+c_QmZh<l&Ld*T<2<61C&c+gC!fa?$7<O$?(yrP&C;w~
z`UDxl0kPauBz-Ex)ks<`#5G7-BgFMcT91tQU2Z_q1|e=l(ncYEiKH)uxDKn=VQuiG
zEKM)Xj}u!yJUBtjPYCfzVt!JHPZ9G|LVTK-pBCa2B1<6_e0Xu0$Sw<UDv_lM@f9Mw
zBE(mT>?)5Z2k)PDE}^P7Re?I9T}Py_<5N2iy4BzK$>VS7(u_II<EP)aAfa!BWAAW2
z3z~Vwci}hImKI7Z4asDc$PU~&OOs>$*Ni0}QpFRCc;do`P?w0#C1T32sh*hE6KC)k
zEKNQ1T8RaONI}oS=PFg3KBY5QVY|W`glX3IzZ`q?FgN!fBd1xLoK!0JJ=dh~8xv!F
za0}lP^+iu-&FW8oy^dQq61s($nW!qtV9>zQdX-HwafVVB_2I-clh}Z?cVoLu;tCkv
zqgv2iGlNWcUi|OLNA#uSA1wi#41Hf-OM?y<#>}!ATsv|#joVs~Negn|g|`h^wIMei
zw<D`|A?`p{9YP$B9pbS!&wms<92MdO?2sVDTZm)}F#;14LW1wa2XDDbckmN({lweT
z<!1|T?hOyxNJXOs4Q@fh`4QUC;5H#{M}ymixC0ID5aLec-iZWpSQ_?C!+pUB!dcKg
zlGhYrkJe<?K3_))9@1StNt{m-j#tQ3VwXysLHMBRNPCc%9H9rOHBz4-uQ2j?Y912j
zA#Dgk*R>?Y>8KPur@L?-o1f=1VUTy)wY&}qU2h2F{7aI3J?R_OM5CHdAvYQ}CY!06
z*wKbm#vzp%4;B`7oOUHIR6<u^_h`d^=RsSqPjHBQd}R~u%EZNFAPD_W2Yh6?Xw#2$
zUow~p8O#hAf#oWg2^BnsD2=LtytWA=vUEF?vyc?_r#p@B^#IwycCcJLHj2mgaO6-y
zqBE42KB1$q_jg}T9RDtq&d^4~Mhf{iyhr)vbx<%XQ6rhq*SorWTv*QB$=n0@S<YH)
zt;-`jRXa@{IT0r{9yt?d0dgTOLZs(n%yar_`x)|RJJQ{bjGz!pJuT9WMn+Kdqv;UR
zy@QM(IACdAN%<bqfMO}jJw&}8B6EI+9-&^3gt!dpmLVgELRhW|X*MB!*okf>a&18Z
zHEly)ZKyBb-j2N5g}4KGbqMh%B>f~z?>-~xXCdxF(k>wm!_qKeg(n<K!-Y5kOCy9h
z3QMDecn6m55aMVojTYjaSh`b)_hIQiA&$k;SRp=ur3ZxgDmJ{zr;}8au`~nAz$kPl
zSJF%&zK^B%g*X>WbA|XJmOd2XYOGw1HR2Rib(R*vm20pD)W>O1h?Q%x1|;4rElevv
z#~M(xrYS#Wo3S>pT3d+O7CwjSE(tg_6A!*5&`KOz`I5knj%(?ygG^rhE%}PBI}+gZ
z<FwgSV^;XC=mUxr-FR_X;DA@VXjJuov*)`i*7~T6QeJo445^6|*o?huW$ffx%>ne+
zzc7}+Fut&`z1)M6s!MgPv|$%A*o6i|zP&GDz}}$FknW0B4U*L$D=<?m_Y%oo^5svK
zJ5B6P6F1&EUnQ1Ti6dC&eGfIm!<-rt=?EVg$w$T%4B`G-`#b$_9dDuyIYd8)Sc13Q
z|1N0eRt=*PI&>3uYQkS9+l@yx;VCeAs#T1Bi!e%QVZ2(Hfe^LC<_teow0)B!wUlW@
zCykiEoVxwIP9`w{KYt|7;n{|amEM1goWNetWI=vw0|(aalLnk5qxe;)62DY36h;8S
zQ)9VTNc;-v@|qrvM@Qp89`D3scH)UVj=_Us@JJq?!w%>0*F1;iu4DE(*5)-O1CP(Z
zQ+b?;CuHJjJkG&`a_~qVm*L)Jd{(2XxMTto{mZ~=KjZPE)YS<08TbE;1=Z{*rhk;M
zx*fyxiDCM}VGwQ~qrXp>`R-%ruRQto{fz#8Vf%i@d_Q9;Y<J$zxC*i7e#Tpf2khtT
z9faG@Oy19Y&Gdh{`EJ$pr5pCq8d?{2>%t!JB1@Acw;hZC$1-lQJU%|CN;WX#!{L9F
z5~wuMa6*3h#tXwOG~^bV0QY8KWxm8tY>Uthv?0wlp_y(+n(ac|fiydW_#Lu-#|LGq
zin}~GtR$}YxZmS~4$!GLVxvZE1CcMiydXO1ynurr`*LB@ebvWDLIp@%fDFO0U0Am$
z=kl_RbyO*DBDb5UA1wM}m7(skyIE7|P=^@y5F<!h&oKID`1`5Luhw1Gi|jc2uh*AT
zy)GVHU(h_9Dsc|c&f)W->tf~ntf|k({J#_FoA*klCH48^#lOClAjmXwl;8gHlAezG
zr=y8b?q#_ctP_L3)~CB`COSFT7OV~AgWZwfmpfwUDD(8V#p*+qL;q3fGOykIFgr2z
zF%`%z<l4pS!rixXCS3@*bCfo$v|Tv}I?gUEc<`X<P4}(lf8QBqzdW=#uF<PN(rwsG
zK4UYm4}^<iOrjV8E+A|HUveT`Az=%JxQMVtLR?JPVj-rUt3-%P30o?}WrQsg;&Q^4
z3vmTuD}=a`u$4kg?|!R<_%UH03-J@eJ`v)lgni26`~EGTRtM7>`#)V*JP0{6zO!!8
z^8a^wljmo>VM2gW?7yxz5I(BV`UzrRBPs92W3gkb@N(e*c03@&KV!+y*pTPHf!P~a
z3VKA#b=Waic-im}J3bWRJS@q>hQP0?_|O6zMgEle`tMi8nNVf1AvR(N3kx#)ym{7t
zkum-CVMcP8u>hSYsLxn(-M_d&h1S=R<0tg>dWI?wa{MgBUC6OZh{G_wYWIMJ5L$7f
ze*`!Ja}h!ukH^O20G|IS9(z=X6Y%f^JObLwx0UHDtu!qCTVP%O=E#J32P4D&ruP*F
zdkW){HSSSM6VJHEGk&l}mOIA`ImZM5hVLe*t&)F&0wx7wJlY`t@B|uaeewpqJExES
zbvXlXO=l$k&IR0gUChL#?LT%l(*wPQly4ztu;kD6PdM(b*Knr0aEWoa#C*LlVYw8>
zA%({;TzkCfsi7Cu%}7)(5p9TQyF&_<bfOJ0GX&Xm4)PCaAogf*yq3ht5|7b2{Kn|~
z#$S>*WF23cZb|R^X(oUTK4NRwdHUrFV-Udz@E*os505*JT)XX(Y-ahd@ALbyt0&Je
zU(!aG5<zQdeBl)81~i_Me<B1TL4^z{$0IFBqu|<;GRnUff|Fvod?e0Crf`&<ohNU*
zJdS-slb#I5JA=P8?yl5(=L&HJ<6R-db&Pi%k3W&K>_x5AIC_AggoF}xP<HC835nrx
z$n1r|gK6X7b863YT7R5^;6$M+ECTBRWmXZa65>t~ocZ#Q@0VZd9>@hoqaQ$heZ<_|
zM*kN*M{K$y_VIOabawLcc5(G^bN3W^eB}{&^N$Yvqa*)F|GNbIc)^26<jFs}^N-H_
Oqc8vH!#}$6kN*R3&oWQ|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp.5121BAECA46B0C23.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp.5121BAECA46B0C23.idx
new file mode 100755
index 0000000000000000000000000000000000000000..62b2281895d1caa4885c3d1461cd807135542fa0
GIT binary patch
literal 4952
zcmZvf3s4i+9>%kq5CcJZ2uLIHkjlOnNPzGNB2qvF1q&jEB5Dw-HpK>7rKmL}jc-LF
zrdF^K+F}tkwjxxkwbWbc6j6NCS_DyiR4G0Xuii>)m)u$S-<it{XZAba_swrPC+wcQ
zK0P5Jb*w-Tmpgr>c5%@{kw74D;eHE?bW3gt1%hQ#fna6ji41LuztZ{NpWdrAY0w>A
z@i&ErTkL(rrDemmotkqG)V&RjN0olhdZW6=q53-V%d+w<(c7Y*wDuZPXW;M7%d1RH
zUpcSx?$sW%f9&s<2T~><|E6hha*z4R<EwkRpKSVT)i)~^oK!`0jjgPDU_ZYu^V&7+
z`?Ht-=)XJnjlsd-o}Z6*M0}UYwzc5hgB^>`n6C#uxZAkm%HU1eqypsn7Z=B5TASyb
z+UUDO^8VSSjmWu=-1Ej4Ex0hJLAq%1T5~t!d35QK*~wW;WOfgmUjDs1;n25_I;xBv
zW2gEUf}fwh=>F?^$~q(e(t7ug%Z_$MeY)mKcjMvnVJW*?eh)5o+V!Mjn)%4(?c$(U
z4j!nXt}Pi)>=BMX@^w|_nm-3_D`uYlEvdv+UFsUP^lDsN^_hJ0gtuect?jct(qjW;
z7x&LjEh(Egv3`o|M*QiLy`!>AV=WCUDmJCu?{<ku$VR#(?Miv1k63adu7Eu8U9Dfn
zi8JXxKXyNJEi64Vzp-K&+BfjPxrRAEV`Xd0tmo2iYnI(;NL5cH-=S7@eO5GJASa;W
zJ$bUt)x<Aq&n@GgEiVc_)^uR6IcvZ%u3XZdzi4i4P-<1yI;1bRA$4iv7=7N<-Lc(%
zYjbJ>4wUUkkgX9c>S6x;^39L##V#dkdfr)QY%Xq%iH`cbz5P<%0k^)B8#}V5Yu{SZ
z`QwExkHU(tw0m29<(B;cKcBi7u-;2-v7Z)F@wK&X({tyu<FJY84c~9PXBh81DZAwo
zLrgo7L*z9{AI%kJEV7U2TzW72RP|G)tGp$CAo5u?TfXtfYDM2eHm9z$#QCdfJ5%HD
zH?Q03vN!YjRb8<5Q^l$;(w{2EHhea}hU&@PaQ@Vug9iF=<o3$;imi7O4UtoOGcE4v
z8+H|LNlahzLE9(Y7oEeS7XRuWmfv%+y<L(>-g&6H-RCsb=gf;cSA()XG}BMNeRA#c
z`KNwY!d}xOxy7B<YsbY=Rd-Y8yuAABi`F2!WsE+?{5Wy(%6(h5PCEBve#-T_Tc6*)
zJgZ~cFQE%FSG7jXYrmd|)%YL1u+#n6&hq3#@%Dv(R6P2k0%Ja2Te@Vk*RII8yg3sZ
z2Ty)Ex@@4WQnPGzeD>Vc&*;M+Jd}o%kGs{T{mJRWvuB-xz1P#2>94gZ0aa~ha>JFo
z%6@3g_58_R^Kca=pLa7qeL?e>xknyqv*uK-sDDac^6fbJ_6;X%-Ql}04;XXt<pHwi
zg2?E+)4zm1b}lJ-nN;2xJ8&-g@A}VjYF>`oz0EsV*!kWCH`?ra{gXD)`l3~vzi#>0
zduw9s?ltx#4F;WlwXGm`=5pWd-yucee*T5U+C^cAKp^&M+8Vc4=_^I&WD0WJ#BTPs
zfqNR*ejNU^bd@HBrb*Fs&h`>}x!4BYb~&^`_U^&^KS<Hpob4iZv9}Gp9rNOG-@ZG=
zhooo*XM2i0N7)A6exH6tX)Y<SI11)+w%kshU>kUQUw><8z09;*iq7L~Ke1n&ZQ$+u
z3w@6<A$<xdn!?%cVt0Gnz}t@~UD~mz*^#3KX`Jot;q7M|czeM5_^T{^J0V3=Ia?-{
zx!4BY{>wSztImSccckd+ob4)hwYLqtty`Q|7%_Te^}HNxtG1>2g{A&Rl&@nB7Kq9D
z`{$Griv%heOBqpkAP$na5(h~<83#!skN`<SNq{7gOn@XANP;A(Bteo)CP9)4q(G8V
zQXoksQy@tP(jZAIX^^CoX^><B8IWX@3`jD`3`nwpEJ(6S79`nZ79<sNMd<LyB_~84
z0#ukhY*_K1S8hXwKCECvj0m<J8CiDBgkXE1wjkJkpm8ID!#r+6aG1v}2oCdv5g}ln
zFd+oY6BdMkdD4iGFi)Bg66Q$@Lc%;{L@1c2Ob7+@lm($+o;D&h%+n@>hI!h8&@j&!
z5eDWN6T-keV?h|0XN?F8^Q;MBVV<=hEX+sAqoPMPP$gFZ6)le*R{Rs?RR`OX1Z-d$
zmI1a;jb#I+#g+r5M=%52$}}7_aWxK_xE2RZT#w+eNu&{=NvH|XB(wx*5_*JyO)`xH
zO;SyQCaEPslhh+5Y*J|yXi{nlG$}0wnv@=)V3SUxL6cU~ph;_K(4_SU4Vz3F1DcGQ
z0Zm5BfF`3y7}#XfSkPqEENHS?7BpEs!osFft{ncab<7EuhXWNUj~rI~J0!xv_U-^5
z2qTKEL&3J2QS4)=Yf)?;)cq*-1nN>DW)OnDNr>%$+91L{5{>w{5yc@7H={V@;k76Z
zd3ZmHLmpl##32ti331574I&)!2qQ{B9$`ia$Rlb|0`iD{lz=><R7gM`VG<IMM;Jr|
z<dH^{ggnxWl8{H%q9o*z{U`}}WT}vZJklg2A&)eONXVm%C<S?x8KoePszoWtqxw+_
z@~BcF1$mT7NI@QD5K)jv8&MkaXfsMf9$kylkVp5UH005xLK^aDlaPiy+90AKk1?VQ
z<S}NHfjp)bWgw5~M;XXtN`(yMF(x4cd5l5CKptyES;%9}C<}RPEy_Y3+mEu4$Ce6N
z$YV`H7V=nwh#ld@uKPpqpMj#^5K+z)|0xUjNBc}=d4UJBxeRg4@e22f<6p<=vs|&?
zx2x>9X@$MQiBIQm4k*FuAMBKHM<fYGg8wxM4iq6FfFdO%P?UrMik8qoF%kwSR>BUe
z_mW;~r82`(x%=^+>YdJ?75V!@9Bc<s{#m16-DLQ_`^Gmz=?w2|KE3WyyjR4z%`$G<
z#o5J!e{*dbb!zYWX5CtDTIMO6z#laC9UN`<vgu|eH?8nh1oJ5VZ$=$#hg%-$_&=+<
zzqhn_-i^wwT-PtjFO~1quRLy-s;y4orai}adhk!mlOeC?+QnD*a?@^;+!+3Z-<&@A
zUXIU|7H&F56f=>(f!jfL1EZYAZQ`aC?g}4%nm=80OEGk`=I^UlQojw2v<~gB`y}@u
zKE*GIy@TzL&8N9-wIbMY_%600IMjp{A)qF$2njW1MJT9gD?&rfSP=$l){3wrnqT{|
z4z|NRxGfcy$gd)IlR9kmh{9Dk@NgXtJVHePkI)goBUL2uNF510N<{&W(ow*pRW$Hu
z9SuB2#Q=}dF~DP0Ebv$z3%pQy=<tckF(*PEF|7D6Hr!4PeX9XWqQUaOB4}`+2n_)g
zsUd-)G!#&@h6ak!FhH>yc325?i{Dx}^lhq8rzuPuy2u(D{`>O=e<2WLk6fyUFNHbf
z4xe-uXdF@T0(ExQ8}kM8hdsf9kxPv>_pLWMnUptED6$ielGsZfM>{)>neGxF7vSPB
zeX8rYv2LEPghxcFq9?~pVcn@n|1jL=|IK=g_nP4AH}Ta;7(vp3LBWbprFV#ok-z5d
PkvL-^nlLMN8j}1E=rY2y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.1597C09D10EC33F9.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp.1597C09D10EC33F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d78992b5c1774c07de6f952e8a27260b983afbab
GIT binary patch
literal 24842
zcmZvldzfo=eXsZLw|;B`8-%^(Ca?<>XOc`NS-Gz?S-Ibn6>?dbTvjHx39z|EG}0>K
zsf5HOR<wm>ja5n`q>Ua4DjKyL?NQX~R9cIzmkP?MRJ>v_BA|FV!#-!&pFhri9`npQ
z-}l4&xPA6aoIMjt=|qBd4;+YIk$yw}${V}vz<~pY_x|&`8wc0?k~?tVZx0<f@P@m-
zSb6l-k9yAf{#So0haHzc@E_l}<>DjXeD%ovwmbgr%a?_<%ieZY{Mws7`owR4eKb?C
zTz>A`bGKSLU-_?#22Vb`wf?RnnP%~P{n_WVD*yJtr(U@665Vgb7yj`Fllig7fAE{@
zZGr!I#hc#rKzsAKmv6p1`PQ3n)t>NwPCn-c6QTZ^U)}Vl&;H_&<xhY2o?m?bmtVj2
zZ!BM(_s)LJ?|%8F-O1zm8_unM^Ss|YL)UoHKK8Lc@ErTg>u%W!8xBA7-sP*gyAPQj
zcK`6Rv+I=ezVJ^6KK!|c_4>d3cu{IxbLXwq$KL$g+`ZM-^G}V`4_y|y;Vri3zwmEo
zeCsyTM>6OA`)%jE=Py3_UEkll`J4aylOO%sCmw#^gFm$0^vSpU;vHX}eEjQo9JHQt
z|KeWLW7k}0yqml5;~)D>?N6V7`bU@j^B0b1Z$9$vTaUP}{rk~faqZ77zQ=C8r+e?c
z_kG~j$g4gk-#h)~CGl8o=zQqyw>|gJ%Wph&6l(nU1?T<u*FSf3aQkuc>BINr?cR$I
z>OK>_;eyZnulc2JpZ53fe!_9*-{#7-W1oCScKGZYFFxyD^S>tE@bH&LPxDW|^ZvJd
zGjh$_KJbOb(^vk}kG^&I+8cG}zd^a$`e#qpuIoBmkA41;FaGd*pLok(KX%Wt^Y4^y
z3mkdS{np`+zveytYW}*nl+OIgSyx~B@mI1xpwGLm{n}~iYi9rXvP;%)cqAXcP&%{p
zl~*49^P8RJf9d`1N4vInyz|M*Q?sj|t3TGtT>HTCk1xy$4?X(RX8z{S-fFnz{!;Ol
zx$(9m2QNA0Rqy>uz4rCwn{Injd&S}RexvinXC6A^w)fuO``NoMn*DYAkFNUEp{I{g
z(?`d5{p{mE-Tv%7zdZWgJH8Zu=J+%BEFSo?D<1i)pFMYJ`+q)q^H~r4)4L1tpFQ;G
z-G|>_d;W{7H+{qZgXbUq;J5!-`u96;e#iZ{Ypw`)^YJ4O{Ns;fgM{F{{7l^+k3aFz
zi0K1I96z`-S-vKCrSW-@UH;BjK6{(_Kc0WXjn;?m__JTV@5xX9HvGe*H+@IE@T%<7
zcYgow*Wdl%xxWtl*SA%J^_Kgte}(BY*>C;m)8oJU)mQ(qbxbwZt~>QNAHDr?<L9H-
z-FV+kw!q*2-tGFZ`G&&#o-3^1wf5W_lMnPy{k;8(@9wI<3wSR$`jNl<#C=~|yz7cT
zx%Apv=bPj6p8C#ppDTa*viIL{nd8`(KYh&;?|yvtjRU*?{m3bwc=Wt~f8%R^)qmUK
zufO!{kw<_1cagVW8NdHK>TfR5J^ftw$ZhxOwcp+@t$*g(-77r8U;V`=UwQlak0&dK
z9=rVfo4@aS*X{+udr$kfU;jz`g@1Yd*0+3L@#>?89(W&j^f!m@yY{Zbzpq?lFMFbA
z{Dyjc58t-P{@ncRv%mY&6%USn_&3KMx>tGgJOANj@B94?uR7xk5AK#fTFSiXaqlZG
z`P$Df{g;jN9^8HN#rHq{@~=Mq`Li_7ukHwbjwF6}Q2e6jzYe_arv8qb95`_B;!nLZ
z`hMY(Lv*Wlpndkivrj$ou=nxd#4p{h?wm8<{=^4QKTzE}Uv<h=g%c0^=hJ>Y{q891
zk`K}H-ua?~7aczFuzx=M;kRAz$M^sE^M`10?|k^+;ZsjM?4N@#{QRj8JbBZH57Elr
z`TT?DzwE@r{<-+!M-JY7$N17iw7z$~_~6B*6A%06FaP1|SMT}Jc^^FEfV_7;dhXFz
zop{(k|G0a}akKL&$03^EJD+p#oKsIc?4O_c$j={W3?DHbqJ_Ql1qUxUeBxpM{M)lC
z4}9l_FFbjOmiEp^4jwu6#KZo1aOG96`?hlOckff=v=hH=?|A5i>3#p{yX5rt8oy@W
z|1Ypd!kQOITFalLb$%VpbzvRMb!i<;J+B8-FX+M4OL{O3ya7yuU;xt~8Nf91Mlg+n
z5lo|G1apJm0CPjw0CPjy0Mo>qz%&UaFinyP%uRk1%uQhv%uQ(%Ofzo=(=3?5G)rbM
zExZLxi(moMB3Z!Xc^*t&;KAf29!x851=A{6!L&+NFm1dIOq*Z>(<a%#wDWc_?SdUl
zyJQE`!8^co2o5kEk^@X9?*!8+IKgyEPB2}(3rv^b0@EeAz;yF&Fx`S1Ot<6)Q{V+K
z1wjB)kOVM2ya!B=-~rPkdBF7YUNF6a7fi3@1=Gj-!1M_|Fny8_OpzDC6a^7XQ4+!Q
z^L{Y>f*(x3<Oegr2fz#n0Wbqn0L&mC1T!cE!3;`4FhhI@%#aWQGbDw;4D(?y!$KI$
zuoMO}!biZ22oW$NQUuH>9|bciM8S+oQ7~hC49u7i12ZPYz>M>8FylfT%(xT>Gr=do
zOb7`u6H)@qB%cH`DI~#6N=YzNd<x8zkODI$rNB({X)x158qBnm1~bEFz|064Ff&pH
z%q*V;Gb?1l%t~1>C0+tk5+pDsNdhy+=fKPfIWTil4$M5C2Qx3^!OTl}FbjME%z{t=
zvmh0~Eb>J#i$W31qErO4#FxM<2_-N~QVC3%m%)?;8BAG{!7TG-Fv~(2%(7GlbBo^s
zb4%C)b4%I+v%*)vtOyk_D^dl_DqjV&DpbL&N>wmxd=1Q+Py@3j)xfOtbujBf9n88^
z2UFn{Fcm=oQ;`%f8+-%IhR^`BAvM6<=C{Gz7Pi6MmbSrc@=Y+CLKDoU)C9A|x4>)(
zEihYB3(Pj(2D2@+!E8%yFgtt)%#P3jvm<rD?DAbOyFwStuG9r{hu;BnN7w;#N7@1N
znEjaRr6<MhtL;|<U2DJgCE9oPo*2IPta<;5;r<im7oRY`_;h*y{d-RlRkC&mIr+43
zO(Sb}BTvcNhk(~rvX1_`M%K|^pOSU-^(xV$uh)nkef^Z^(Ko2XfWAQ^2J{V6VnE-h
z5+nLXjTq54PKgoy4V7%5zoC&0^f#tt1AUW9Oz4|5VnW|EB_{MYRkDfxrbaf=-<*<7
z^vx<Uqi@!T8GZAVn9;YW#Dcy>BNp^6Q({4%R|$_kuMr-7eoA=sttzpiZ`Ftuee0B1
z(YL9@hQ3WBHuP;%Vng4q5<B{Kjo8t*Pl+9Uhe{miJ2c`z-!UZ)^qnekqVLp*6Mg5D
zIMH{h#D%^~BQEq^Q{qD3tr9o-ZjHFncTb5MeL*Dx`hrFT^o1!A(D$gsgT6;29`rp^
z;z8f55-<8*jd;=bPKg(NpGth_`!wQ1-!~;b^hK43=!+T=(HEygMBlFxKl*-+_|f-I
zi68xdN&@HyG!j5RFeL%>gDMH4AJj+?{os@Y(GRI4gnmdPA@oC25<)+$k}&#VjfBw;
zPe~a4h)N>pM>G;aKQbi|^rI??q94^r6#eLwMA46_B!+%WBQf-2QxZczu97(VagD^$
zk55S){e(&q=qEIiKtC}h3G|aHNur<BND}?zlqAtlsU(GdN+T)sQ&W;cKdq89`e}`%
z(N9lF8vTq)GU#VCl0iQ+B^mUyD#@as)kqfo?385DmsBF5FKI+VUz!pL{hUg2=;t(&
zLq9hqIrQ@?$)lgwNFM$Cl;qJbsHA{?K_dn93sX`+zo?QT`bCWt(JxL(5&e=%O6Zp~
zQbNBpB_;G_mB{GJ8j;bLr$k1-tdcVNWsQ{4FHcDs{VkPjp}(b(E%djhWDEU@N-F4A
zG*UsoG9?xCt179YU)4wz{pyrd(XXkbhJH;WHS}v!QbWJ4k~;czjnvVvPe~npMI{RQ
zibfRll_^osZ>Xe!enTS-^cz#sKz~~$+vsm=WE=hMDcMH9sgfr8O^r0sZ%#=Q{gz5v
z=(jY|LccX7E%e(eX`|oPNE`k3l(f<BsHB5_M<X5dJ5$m@zpIii`dy85(eF-47yTWT
z?4ZA+ksb7Rrep{G8|*ju_n&m1-hPe!H9)VmzxE~C|C;nOd(XgM{E9R@1DDq-@cdg@
zYXbGxUIR2FYh!#aF0aFKU0H|adVd|3>qD}ROTDazrC!m)Qs39ZQa>bmTpDBpEDeeQ
zmWI9omWClQ;L<1?VQEy1ur&6Kurv;d5tkeC1}ry}4OnjUH(<FjBpbLi$tGBu6ca2>
zeG@EALt?_^ro0KuO=T07oBd5#ZVt&NF3qwTmS)8aOLN~0OY@MJacPk)u(T)^SX%lP
zSXzd}f=gcJVaY2zEcreUOMXarTv}x-EUk(ame#%%mewJ$;?gGDU};lqu(b7Uu(S<{
z4VQM=4okaYho!x5hoyZ;?6`Eu4p=%A2P_?Z2P_>!;=rX-cEZxBIAQ7RJ7MV@5+^QQ
zvI~|j#RW@O-vvw8khpN^mff&)D{ff2`)*jehs2FbK^9;sC;}{nz5q*MNCaGZWDhJo
ziU*dSz6X|`A@ShSD|=z-RlKnD_Pwz54v80+KG_FLpW=h1ukVAUZ%BN&6lD>Xq9Vdl
z?2E7zheX7sU-rY&ulQl<@B3lt9}+(<19AYC0VM#-KtBM>z>oxR8I*&t3@Slb2Kzx+
z28Sew%a9y`Wk?CZGSm;jGBhM1T!!T^EW=6|mf?OFmf;}@<1!*gU>Q*&u#EI0u#5~z
z1eZ}c3d^Vxg=MrKg=KU|qPUF7F<8cw7%XG`7%XE$62oO&j>9sp#9<ll$6*;Ck~l6C
zasrkKB>~GsKLN|ckR))Kl#{SbDoI!-`$<?Pha`#1l$?TPN=d;o)lb1PH6$rqrsXs&
z(@Gka>3$lP=^;tuG9zbTnNjw>`Q`upP4(&ROg{t5%#dVonU%A!%qm$}X8T!KW``t;
zOG%bsDJc>xrM?79X-Fho=HwhKb4m`Dxqc3oxgp8nGB4+0nOE|#%=h!K%nwN(mj$^1
z%YssXWuaeyWnoAPxGc&=SQeEcEQ|djEQ>=@#AQh?!Lp>3U|H&yU|AZH5-w#~hNY~?
zu$22UEaf4Qaaoqjuq-QOSeE-`SeA#RjLR)~3zl2T7A&{=Td>?3k}X_T<O(b+N(Gjc
zeg&45A*tZ9Dpz4yRjRP8_N%b04oMZ4HMs`Mno@&htzUy>ZAfaktjl#+)|EOe>-{<`
z>qAn<r6Mb^R1^i4N?(DcG9(Hv8*&4d4W$9gM!x~e#*j2{xh-$Qa$DJk<#vA?mfJ(J
zjmxIogk@7{!m`<K!m>FeO<cC*7A#vz3zn^Z3zn@RY2mUhw_({<+OTZ*+pug8NgJ0P
zxdY3N(t%~C-+^UkNIJOe%3WA?l`bs1{VpuKL(;|Nj=TfQ9c2fWJN+G4?hMHeF0Zp+
z_tIahoZi0Pem&5u?XP}`E_@+idgI_H=bn9fJHKT4PqLFAmJ3T(KrSv>5xKNvC1iQY
zWaRRam65lWYzw)vWEJG<l2wswOIAa!FIgQ~SuzE=v1AS8?IqhrZZ269xwT|1<o1%a
zkvmJ)LGCVD7kOvNc24sC_htPpciM?>$b$citkSi&(v!chYZJQm9^^S)`vCHauKfae
z#I21v*lXO{yO76h?Je-X+qz2Ev9~^<>)2bL({=2vujo4V)<@hr_SQ9S9ee9zwvIi$
zO7+;&PpBSy`Z?8OPrssi?CD3G9(#I?(_>FRW_s)yRBFJUVL}bqGt8+0dxjM?V9zk(
z4A?VhoB?}=F*9J#s8S>Lj1y|ao^ei%*fXxE5qriFXT+XS<BZrdj+qgA8!Fwv-o}J(
zU~gkiH?X&{q8r%T7;zie+t9cT>}`zM2KG!UHDS**p(gB^=G25e(~6p~XBu%P?3pyq
zggw)knXtF1(oO7bPUt4~Hs^E`dz&k|iM`Dcw~4(?joZZD=9q0_&#Y22_RJG%#-4dj
z&Db-qs2O|a5ogApS>w#uGmn`Wdlr>iuxFW23-&B?YQdgmMJ?E~j5rJSEE;FQo@LA|
z*yB~oV~?Lu9((+p^4Q~7l*b-F;&|-w8pmUgA2S|%R+U<@XPr<h_N;Si#h!IVt=O}U
zI4kz78fV3xb<C{Tv#HdEJ==uZuxFc78}@7~YQvsw#M!WC(>NRUY-47_o?WGO?Aa&O
zjy?OF+OcO}Q9JhRBhHRJyT;kEXCE^=_8cm8V9zn34(vJR)PX(6iaN087;z5lIW*3J
zJ;#_iu;)~%6MN1Hbz;vsr%voSSJa6;=ZJG+&#7@v>^aBGi9MG}UD$I?s0(|pIdx&r
zwW2QUxkj7|doGQ0Vb3*YF6_Bg>c*aXLfzPN&#4=G?iF=o&pqPY*mG-~8+-0Cb7N0X
zsenCULIvyzb1Gm@SWy9c!iW>FCup32Jz>lQ?0Hn`!JcPAJ=pWisRw(W74=}xGvYkh
z^Jtt0d!8}#V9%>kFZR3>>cyUSPQBRkuBaD#-Vx`;o>$|%*z=B=7kfUH`mpDlP#^Yu
zbLzvMZ$*9B^Nlzk_Iw)W!=7)<eAp9JDq>HZP!W6LoQl{JS5(BFIO0U?i5e$jPaHE5
zdw!MrvFD#qKlc1{>c^gcMg7?Gk2pW}{2J%So`1~z*bAsMfW5$k2Cx^H(*X7YD;mIF
zV8jKm7tpu>_5x!Tz+O<LLF@%5G>E<6oCdKMT+tx*f+H@7y`aVgu@@Y(AofBk4Ph@d
zp&{&r<}`%8(29nz7aDOP?1eNgguT$1g|HV^X&8Iq2@PW}Jf~sog;z9;z3_+&V=t_6
zVeEy+ER4N~N+Z~dOlSmqkvWZEFS4Q$>_tXg1bY#Ui(oG@W)bW~RT{-!bV8%pi_U2j
zd(jn*VlO)4qS%XSToilJF^gg^rqUSpViOv}UTjWd*o&=b412K=7sFmm<6_v0jadwP
zah1lg7oX5L_TqCI$6kCz<JgOjxH$IW8W+c2e9Yq5OQ<w~y~Knju$P$A1ojdun!sLS
z#3ito(6|Kl5@VLYUQ(q=>?J2OiM`~UCb5@X(IobgBQA-(q{bz&mmISs_EIWMVJ|hI
zDeR@@G=;s?il(ra8gVJ?r8F*uz0{bcu$NY88hhypO=B-Tr)lh^S2T^i^oUDiFRgKD
z?4`#njlGOYGuX>aXa;+kIn7`%v!WU7Wky^Edl`+(U@tRf8SG_Mn#EprLbKS*&S@5V
z*%i%VFFWG0*vo2M7JJz-%VJMbsf0agLM7};b1Gp^T2Tpm(uk9=Cuy97J!#A&?B!IN
z!(MJebJ)wxX%2h270qEUH{x>G%V}H=d$}>oVK1-JJofSvn#W##PV?ByuV@~7`4N}L
zUS8w!*vpSu9(x6q7O+>C&;s@fb6UV&VMPnrD~z}T_6izTz+Pd@3fL>Ew1~apgch+^
zoYNxqiYr>gUU9?~u~*c%BKC@7R>WROr6uf@CbWdT(wvsCS6a~$_DUnJguRl+m9SSD
zvl8}XmCD$YCsf9sJf|}D<Q0{%CyzK8d$PvK*ptUh#$H*aW$cwFw2Zy-oR+ayUePl4
z$|J6fy|TuYu~#0mGWNDqx`n;13Ejfp)|_r(Z)-)ju(vhhwy?LQaa-8i8nZ3zRa9ER
zUS&cn*sIKG1$&hhtzfS*;wspyXj}z*l`*Seud322_No(F#a?wztJtfqXcc?a5m&`t
zRpYAItBzR}do`8TuveSV8un^)TEkv#MQhlrjkp^2Y8qF=UTw^3*sH6wj=lPX*0EQg
z(>nI*D_X~1eZ<wVSJ${Y_UdC+$DX261$)YbD%exzRKcFIq6+qu5vO2J(KrQr%9ttG
zYpAq=y~cz#u-BN=2KE{&+Q438#5J(j(6|Qn8e`VL-nL4&v9~>;+t}Nl({1c+ujn@R
zwnyAH_O>-{8++ShwvD}}N}JegPG}Q*%{gsiueqX4>@`PR6MIdKYhtfCW=-t1RNBH`
zYeHMtYt3m3d#x31VXrmfTG(r8Tnl@xF>7J3t<pC3+7sHwUVBd4*lVw78++{$*T!C3
z<J#D3k69ag9hG*l*O|}`_BwOg!Cq%YJJ{=txDNI@8rQ*IXUsa->#DSiz3zl|vDcl`
zF7~=B+QnXX#C5UP)wnM9x?|SG-i}Ilu(vazJJ{Qq(;e*XtmqE*c1GL|_I5OG2YWkX
zwsW!<e(O&U{OXw-zfD;C##h|f-v4gRoN(t2=bCT-3vzn<qU=SL{U4Fq|63w@+KC_g
zKlztXxGi6PME}D(#8dXR$4)(V=Kepj^S#@S9en(f?}~drS2bg6AH%;=nz40c{fz06
z4KrpyHqMw4d1J;lkWDjYLf)LQO=R<onUO6sW<lm>j7PT4m=)PJV>V>_jM<SLGv+{c
z&X^O~HDfMh_l&ubg&7l&Ju~J(_Rg3W**9Z8WO2qsWdDr$kpnXpKn~7W5IHnsA>{Ck
zg^?pO7D0~ASQI%nV=?6TjKz@?GnPP3&R7yTHDf8{^o*sEGc%S!&dyjCS(-5kIX7cD
z<ot}~kqa|cKrYT$5xF#DC1iQVWaRRUm65k*Yzw(EV-@7;j8&0qGgd>c&sZH<nK1>q
zF=Gwn?HSufZq8T}xiw=g<o1lUkvlWiLGI327kOvKc24rK+t+W_-}ke3zxW4Px3;nW
zG2!&~x^5jvuhRn=bOs=!&Iq)j+W<1@OhB8uO(3(*3}n$+fOs7bWYt-LY&si|U1tY!
z=o~;!ofF8Va{;+^ZXiJ?0C{vCAg|5~<kR_pM4bra*ZF}0x&Tm67X%9FLO@|%7$~BP
z07Z3CpqMTO6xYRp61oIXQkMiu=~6&xT^cB(%K&9{Ss+O#0p)Z#pu8>*RL~WGin=0D
zNml}rbuv&{R|eYBZ2?tu6`-oF3RKh8fa<zBkfKw78oCD1wr(4!scQnYbS<E^t_{@D
zb%46MF3^r{=OubAJMdQ@w7&4Kd*3AEJ<hyeg3pY;<IIOF9%mx5|2Xp_2adA<a_~3{
zB8QH%5OVl93nNF4vj}qZIEy03j<XnY{5XpvCyuiOa`HG!BBzeC6mt4FOCx8FvkY?f
zILjhS$C-qjJI-<^`A@R3hfOtiVegs8(fHBo{;Lta_38LUfArPcFW7tc;j<2(yZ_TB
zZ-3cm-+$XfgIo6AeZlz`9NmBQp?5rSSKj>Q`jv*g4{uAhcJf<r_CCNZ**da*$@IvE
zB{Lu!m&}N~v1A*_rX@2WZ!Xy;vU$nO$d)CuAoEMcBU_iuifmgl8?t@L?8uHKb09mH
z%!%w;G8eLY$=t}ok_pJ3CG#MAm&}XoTQVQAxMU)-f64sFfh7wd2bU~}99psva(KzY
z$dM(BAV-%hiX2<A7;=2c;>d|5OCTqgEQy?2vJ`T9$<oM~CCeaZmn@4cEt!OzTe93q
z{=ofbobvlmeebRppGRME%(nkL`a?e%-Bkb2yWY9??u(U++xwq>?|Z|8rxg14B=+8Y
z{>#rlcmMmv{=r$g;*|J3KiPZt*=uK;_g`A<p>p6y?Tf$prM-8*j=k;=_y5`Ze`5Tn
zop{}@lh5zh7G&)~{OPI%Sx43{h#uLnAO>XPf*6rE7Gwk2v>+zr%>~&+HZO=7*|H!O
zWPU+-Wb1-hk!=fNL$)u79oexU4rJ$oIFVfo;zD*Wh#Og05CPe<ARc7zf_RaA3*ti-
z7eqw%FNhyGupj~C;DQ8^Lkki@4lhU;IkF%T<miG#kz)%ILyj*<967Nd3FPF0B#~1K
zl0r@|NE$h_AQ|NBf@G1U1(A?*3z9?5FGwD_upkBG;(`>BOAAs$mKQ`uE-y$Kd22zo
zkShyPL9Q-H6}h$`HRSq&)RC11QIH!8(m>u`kZt7Vf;5p^3(`VvFGw4?vmhPh?t*lY
zcNS#lB>(RxZVBCS_#=<*{b@g1Kib>>Q~rJHDbKv@%=6y9_wHlo9J_e`gTVf)^zHpH
z=8G@UxBo%lC%XIooO(a<;+F)@Sv$wNZ|#5Dy!o^fFV?qj?|t$V*OL1moc3t2_CHS2
zy0{Lv^}#ya^r9Ya`hgy92GIaF!@vMHqiBSiabSeohPVN@jll-oOri;Hrhy4=o8l(i
zHV2z<GmB=pnFnULSwsumECUPNc#(%2Kj7hJ6|Hcy4y<sqi8i>|1~$0aMLXQ=13TOt
zq62P@fdg(%(Fr%_zzH{(=z^PT;DVc5bi>U(aKlXy1-J<V0d5}A12@mW12?beg`0Qa
zg_}?G!Ob`D!A%rJxQPQ1Zhp}ZH~+v7w}2RcTVN1?TTl$bEjS3mEhL8E78-=$78b*B
z3lG9@i--}pMFtVLMa3xGqJt>hVqy$#u|W)OaWM|J_#h6qgqVO^VvvAaQcS`vIY`1S
zC8pq(8l>Qs7SnJ`57KbUh#9zL1{t_z#Vp*igDl)6QG%N^kl>aRb8yQIa&XIwdAQ{V
zdAJqC0^ACN0^EvX5pKmn5pE^11h>+l1UFff;U*7cxRu2++{%M8+_uCmxNQx#;8qbU
zaH|X|aI1<{xK#&LxYfiO+-idw-0ET-ZuLPOZi=YDO&KU~YlscFH3kj1ZHwD*+a7Gg
zttmF))*Lk9))HHAYYkd(Ym05TwFhmub;J(bI)e_}x?&e@-9Z;_JK_%9b_P3eyT*Ra
zOD{NZdb@A$zeM{lEW7ut_r+J0g->C<wX5KhR&O0h-_rvbdIlh4&j_^9+W<24OhB8x
zO(1j63}oq9fcPE{WbIjjY&{#0y=Mn<^c+CWo)gH`a{;+~ZXls20C{>IAaBnL<m>r>
z#GVM`@A-iOy#P?K7X%9RLO|hO7%0+<07ZLIpja;k6z|1>61@aavX=x(^-@6TUK%LV
z%K&A2Ss<w=0p)r*pnNY6ROl6eioGIGsaFD$doobDR|eYZZ2?t!6`*Rb3RLUWfa<+E
zkkV6t8odV4c5fS~*=qu|dM%)KuMO1cb%46PF3?VI=Oy~xXCD1t+5J@I#V>E|GVj;$
z%Uiq5hb-<g5!t`X{K$b_7C;W}vLJG3mxYkSyDW?x*<}&r=q`&Q$97o^Iljx{$cbH+
zKu+$mBywt(rI6FRERCGmWf|n`F3TcIyG%mP?XuiS9$b0--oG<o@81$I=)F#U@r}ld
z5C8Yq_Wm0O4m4kSlWV`~k=H-<A>C=?gB_*W*lr!zdwuNvU#<N&4jky5e6!cT;`*!i
zFnd*wW2YQ^neNm>r=M}wnJ-Ttjz`xIpO%gsIqwx`pZ`kFANfRoAQ-ZoWAa|+HeCGw
sd+)gyUUc-5OaJgyYkK3xCU3PlT*4nXFEHC<=bW3&bZO%9aEeI(2kt=`Z2$lO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp.FBCAD14B8CE81D3F.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp.FBCAD14B8CE81D3F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..264093ef7c96f231fd4ca8ebd054037cc372ff04
GIT binary patch
literal 6612
zcmZvgdvsG(!o_pWO-tGWZGl29R0<5a?<deen*ybjhg7MiKuN6yQpx~En7}9uh-4B2
z%0nPRbOa3GXB9!?gNGuB44({39#%wTcngYvyh<4v#bH9eb=&>N*R^_@efHk@<@TQ3
zb?<p>_>dt79<W$SYKPDAPMkU3YOz={<G(Xz`X*msES9Nh7R#*FJ4bs*_HbnU>-9_Z
zVGn)zbm{j@@PcxgWbS+6rB$B8*W6dOY}@MS_RE#R)19b)qww~!6-$a<D!O&xN~ruH
z?$2ZER+m@I8oT=9@$-=$#XlWw8rgrx$J_7rKOebu^Z44cw-)_A_v2^A?Q-@z-G23&
zYbj4`tUP&={ruq3pBpcZdHn8O{rs0ZPW1b%QaQAbJA3!Ug#D2->a~m8o;h~+`<$NB
z$eA~fw5^Qo8MXVxyq9f%IXGn-IsAN9UFVtOz8<wDZNkL)k+X94)+rx7T2?hVC+Yh3
zd;fWD$a|mOII$*lqWz$*LH*9&BU#@q6k`uf`etF)E6cW?E_{97v9sIW|0;jvYx{oE
zXQ#h<Yvr)WM@N?@YY)7;d4qUza%X-m)A^$h*Hq5?wdsdm5AXfwh&k=tfp+;*j+Y!-
zzkgCB_o?EBSi_?oD~jzoN8Wj~e9p4&-8T))Iaj)O&bn6Bf#T?vXI3s6dHHN+zaiD+
zw5`Em6m(6t?JSul?EI|0+vuJ9E55v$wf|&(MdhSzE2q+{P1iCu$X`7)>%hK8?xcOX
zVd}Xp<?imnpTxPR-<a7H6mn_u+Pa*@>-lfp*-JY8GJU50$@b0bB2`VTOIFw#CQTSq
zua&Q<T0pMWZYiI#t*yUq&}+qKyUnlJVBfs#<smuqEECSlzrOv)=Pwmc;WwQB(}K{R
z*#`y`75+~{!#5i@ceuLi#S>M-y-!X)^~KjqZsgBC*6?5~&xpQb|8n;c`@$~C(Uf7m
zSAH1VxadyC!A@-ViY=eNcq!O9qi6NLZ)ASh&KkaMd)kdL%;*Uz{Z37}RK0usZTa+y
zeWgtWzpPhQy!gfXURSRxH5*UO$@pN{s`ApydloFsTvvJLxKH=K-fQlg6}Nk}-}1&2
z8^rUqFMPFo?Ylwg{etDI8&)p8I5b!==t^ZYtKx-MXDk_7G5NVee>;05qi^BF?|S4<
zI)9|0!8TO*@w)SetLcNf?*Hw_<66~UBhu|pZ=F2))$MM_@(24#?d(&rlRJ_N*IX<g
zb?^9hza7w$rpo>Sk()y&&U$Of(w>LEm^AXt#tU079DU@(uy6Z}ubg|JaBRbwq3nhp
zTfbhF^~tIgW$%@y%=mfbjW<`a@+<QLlNWb+wV<SKRPMIByWVcItm)8d&(vp2tH(V1
zi}e0;*VB5h=yc(b_aEthJ$Nu(e{iA1hJT+w(!S=<{@T8dSC{?$KyCIvQasn^vc}l&
zCsmBw({{{9*S%Gv);zQ6w(w2fiCs^fOOI`Q|Kh#Pp;~UbJ?D<4pr~%|w|#DA%$aj<
z#EMhJO^1vAv+0eR4fk5T_R@nobL#1@J4lgsXa07`x^U*)#UJkb->2sdNV>G`{D`~S
z-aj0gR{QYuyyc&fnM1qvm@(TsA)i<*$z8WEEm`NtOQWMIEj1mIJESCr_|+isllp7)
zbYMguR~oH|+g*~oWF`jF&V28MoIk&N`R{4;(YT$NoSBjsOnbm@H?O|+<Lvj+=;*ke
zot)h&F_`wif9*K(^RM;O)9Bc^otK<fk{C?e(<Z6<>49$?ZfzM8xAT+oha?8mp7+#j
zd29zYyN$&aw{tt@c1sMVeR+J|Cvxwrz0&B&xSf@pm68}tyD4__gDQUmpGM2$c206m
zW@0ey-!n#kaBAA#AJgb#al2h|yOhLW+P;Z(Gq$vE`S0bXCjM(n_WTyzbgnS3b&XxM
z8|LA~RhfNci(*w4JhqGlk5f7DxH1krUgg2#%XshvRRB*Y6TlNy5j?R>1W!^W@T4*c
zJXw{&lgniA6jcFFDO12xRTVt7Oa)I<HSn}D4Ln`d!PCoh@C?-e&nPp%a~O_3&1<Em
zz$j=|=IRww6RQ?ZPl(u;la^Cr4-@-Z)KOx82Q(WZEaurTVKL7}35$6yL^#ZIVZvda
zixLj=e2DOv=fi}@JRc=I=7kUuFfW9OfO#QG1k8&eB4S<)6A|-bl!%y@LPWy66ebep
zr6`dwFNcVXc{xmE%*#<CV_pdn1@lUnD418GM8UioA}Z$9Fi|nDMv01fEkrcTYhj{c
zUW*b9^LmKrnAgKZ$GjdTI_8ZKF)(k0iGg_|N({^oFa{L0EYX3+Kv2a-akDb7r+rfs
z*HKuoF8gR$u5NoZD6f4wC_k|Wap}5PXtHh=nyi<FChI3GHaQmuP0r0hlk;-W<otxg
zChy{*$-8-I@?IXAyr1yc6kGx{1-AfA!7D&h@Dl->qDzFP=oX<VdPQi8ej;L1a!JsX
z+!8b;uLMoWPb6&0E*YA#TZX3Wm7yv7iHuFfr9e}0E6`ND3N#fzQLw4HRA{Ph6`HD7
zg{JB!DmFEj22IVaK~wW;(A4}y!=~=ip{cudXzE@an!2Cp*fd-QG!3@_O~Y$I)9@1m
zn}tSU^L4y+jni;~>TmRKR^~m%<Ee>z3;4=}sC@y2FGhsgUqM|@?QfxOr1o2=1B^Y$
zKtIgbUq&6Y+Ml<!_-u%>$YUdvMIKvES>&;eltmsJU|8g_VTMH>8?>^><3f}}9v7h;
z^0<1+A&+aM9P+pT!y%6gGaT}`pp`=&AEG?+_z2~Z$JbLHd3+<~k;exZ9(jD2;gQD&
ztvvFC5EYOoM5ur~p`HrJ6B?<2JR!gc$P>bhfIJ~+6_6)}sE9l<LPg|>^;ATj*hod>
zi2+7Lo)~6C<cUG6h&(AoCFDsFDj`p*rxNm{Mk*ms3NRA#q%b2PPYPNk<jEl_BTtS{
z8F_L&m60bmQW<%2fRT|WhZz}pa?mOxPYF>4c}j#T$W!X6f;^>>D#%j;jDkER%qYlH
zf>s52YKW@HQzKMGo?1^;<f)BRMV=a9ROG2)Mn#?)w5rI{LR3SZ7NHvQw0f!`Piv$a
z^0WY>Ax{f48uGNDRYRU0qB`>Q2-T6N*HayNdLz}5rw14vd3u=9k*5c(I`WJVHIQdS
zsDV7Ao*KwA8mWOiBfuEQGs294JR@i|T6pGll%^)00b4HPgL@1fXWlv$FZkK=@1~jm
zB-R?7e%P^MT1R>DGSa%H%djpk^RB<*k#_7)%biK_X-A49-Fy@CFG)>2PMV}ii|upR
zP1jbni9fE{?1K4uTGy~P78GaWK=C#nlwcD;i8c|GWRpP2HW`#+Q$VRU6_jSvK<PFe
zlwmWP)zc#e9}p|^%i{-|J1Dopd^k4$woOeu9GeFlKgJlb@4(x{7-3O!F~XtdV}wU7
z#E5`ej1dvF6eAL9IYwmEN{lF|)fiDxYcZms)?-9RZN!MtqW#Zq4}O03+{gEiespfN
zdDIJTly>QNcyUgAIx{1)qj`6|sMYRu3-|cu$ES0$b8^kCfVbXllXP$U_p9U6jy#8M
zu9J7J1ZIytw|Z%OI&VZ?xw$Yl&DxPv?p;45KAqh*yQ6s<va9!FwMnJxuf(T2^z0y;
z+ber2`d_N)dTd{Oy2M)2-F)vKw4|n1>75qEryW_2uI9A4caff&*t3AM=CBu-A0WOm
z9V{r$!GYo(JSf2-fD#=dD9Is#k{vQA#i4*w9V#fzp@GsJIw&K)z1i|P`x<?lZ`ZEa
zHoZg(V)wtRSv|8f^GMBe5Z_3-zZU~X?6ePUc_+Ki{%nh4oh*2)j|Go&a^P`34m{q;
zgU9=L@C2s-p5PO}6P+S>qE7@*a!TMyJ_$V8DT62bWbhQH0-oYiz*C(nc&bkYPjhPE
zX+8}+-Km48`*iROrvaYfGr;R-^lN?+ZCz7j6g4YzJ1l+~xxW_%r^jQjgA?LmL2(`q
z6z}0d2_6BI=n+9l9to7}kwGaQ1(fPhL1`Whl<v_%86KlqS$q?p9Djc&Z-&n^!*%~f
z)&1fBejYbpSS-~oFWt>AnVP<~6tz{AcYg9XcXid{PgtI4_AKLCUV6N>PyR73N7Oyc
zSd)@l*;3M4x5-FvJ3O<r#GaWtd{Dbi?K@;Y(6?Wqv#9@ofl8KG&?BGg`v1*#?A#?c
luUq#&^kjKKQZ>C-AIF2eb7W(1R>z?aji*B%DIP}3{tLUG%I^RG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F03BAFFDB1120A94.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp.F03BAFFDB1120A94.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2f7adb5724a0400c2781d3335f29347efc6edf34
GIT binary patch
literal 8954
zcmZvhd3aOhg2wxOCzQ~F(6S2jwum`PPR_mxDSP1p5s)B4P%0A^0V7UCF9@TAgop|#
z93@^wte}P&2?}~eMIA+4P_B+EDk#^hBJNm`MFDT<OxpL4>GN=&yyy46{MxkHCypLD
zviX9tvf(pFFG|c>IKx*~R#sK~ciqC|++UfpvWF|n$`);SW%7Y*4!B*udi(q6P-EPV
zk3U>iTfP69w0-N^M_%tg)a!q1mxg(Zw|{Zwx0*4N<#9c3zj(QP<(psjO&;ENDSKD-
zm@6*n#dYgGee!!dp8kDdo&QYu_gBBp<>wsw`t;S>ppUyQUAm(&bmF|w`cbzmS?+yN
z|5CL3*EwhEwWk)}bo`e}`KGh?{PNYWA1r@BemfuUcJ0|;m+s9S8o!{&W&3-cK1Kry
zX`J)KpWHcj&tG=wP`>Key??ug*;py=vyZ&^(EM1>J#Umf_R=)<>bsx3Br<L8y5*N0
zy!p(<n=hMw@`rTx(f-30tk6#Gd9Tw)E5*mh^n8D1_j~T#`Psk+Zr=Zoo!j4iYTtdk
zj%bT_uJ~p3>$xXCSX-{P+g8{t9-P}p*vRyG@`>lC+;sBjH~s&<=h?AKsy|s?ZO?lq
zy`jOoLmqf=`KB40H*ejteAtCgL^n76TGudQO3K`Q`^po$&s*3bJ!INXy?g%j!AmvC
zRnL;=sy2<+2h^7PpC7)U_w)aq?`IG6o>~8evF_oECrz2NbM@HN@f&NqY?l6K<f47A
zrH?8{*KAv{f7skxx9lk#o%O~yA63m;=<l^Cc9nYD;VJWHn9~ow{LU*!{`u63`wwoK
z(`#L1<)G?a_ARMz-yP8LDrNqPiJiaea#g=4&-eX_p0wt@TQPd>L!Y;=V;8(LzM)T~
z^TapLueoE1Iq9wV*>Co0tJfT!{KG?6otS!X`j~k;{&szzvCi%T-(N9)$&1VRW!olR
zvMevGtS+x>cj3KnPMz|>sHH0pdtIyU{qV|Hj_vNW^4@LnAJ_MN=>EowW<OndbPg4_
zXEyx!<oB1qxarrL&(^-yaO~M*n+iK_oB7UPe>~Bz@xKl%>9XUE^%EL?+<jnU)q_(`
zzS6w(!}_mJ?)%SA{vLUM-ICSYR(aMiX34YFJ6`>EL~^7vU|eVa^_i!(4->ak8(*&*
zHEHhPS;EP%Z`_*mkFS(IKDlV2x_j+yr|vua+?k<AY8HPQ?lXJr(RE*K{L98&J$@VX
z<tJHQUAFb=uHy4!Kl<jm%p<4Xes%hstT1JMhtu0v9THw1K7ZlX#oC}}{;;i`)`AK5
zotVIG2%NZa)Q&`lm-U&S?aiJYG@y6Q<99!`_1(f<GjHfOZ_1VXGd+L!bpA_|p6mbM
z+Wy9z*PomF#ri`JeOR{lr+3;tb)e__H~#yn#I1$<UproX;I~JH-9D>f+o##nb^fC#
zW>l}-%6Vs2McC($@15y7f4y_(`Kx*z8a27{;J98(z8ZMf-rj>ZSG@Pz4Gq7)^_NRO
z8u;b_Zg}O6`<R;3m0Ra+sQP2_Tz!%|ywhnqVBo%$g|R<N$B&<VZRW1@kq74N-W<Dm
z%?s`C`(wd{o%Zb7J8ApGF-s2(=vw#g&;8yC_T07i=GtwC&U^dl%UwM8)Y`!>k&$Q1
z!|%IamR+|v@ozRMD=V*kdd=_$ow`bT>6Eg@ZspxNv>b}}4=rC6zcqc99&1G_>7*j=
zTi&;-<xs*^?CD8!>z|(5vFx%U4z~+WXgRdv$8PO?{kCskuB4X~aaDO$hn7PLSFb&B
z!SsFM?VZY^MO@RP=E9ak2@n4L=O4BlUi?@kom|Ad%6qkMIh62Ce}4GIT`#QdSV^ZA
zacz0+#Fj$|e><b@S;_ptsHEeIxO;i`4lRcgKKA&}JEo=H5i03~BJN$@yQ<|-!e_cn
z-tp;zJ%=ml#3HUPukO%tDB<L++1I_%x9!>!D_X9#)w_Rh+WPsxx{i$jC7_qS!D5O8
z#<wL_VS%v`78s{+z_<tpj8}MIe1r!kC;~7cA^;02L14j15SXZlz{H3METn{hg(4wf
zk|F_<A`&oJk%7q(8JMCdz?6srOjT51YD5L5DH<>>q5;zt9he@`ff<Sc%!nAkOvMCd
zMoeIqVga)v7BE|}f!Pron4>troQMO=Ra{_h#06Hb*ITVOrp6)qkk(YX1rDibx#hv1
zSe69VlD0pqfJXuwF*lLGqmbDwVX@D8gvCDFL|E)|S;AqT^9YB1u8DBi=d*;zKJO78
z`+O7Ou`gtafPKLu0``R_B49t5B|+>5Jrcxzu!#h*FJ_5|ebFN#_QfV5Vn386A?$}d
z62g9{iG;8(Wr>7+$s-c>r6wX_U(OO4`?5!5?8{9=#=epz3icI`DA-q;h=P4JOH}Nu
z9#OHcHW3y3T9#<o*F2(OUuz;7_Vp~$v9EhX$G+Y~bnF{hVqo9!h=F~hi5S>7v&6){
z=@AqAW)m^7Z)J&veaj;j_N^vjVc*UY8~e6LZ0y@j#KyjpB@Xr-k2u(OnuvpaH%naX
zyB=||?=}$^`we<SecMfbq&^bTD1B6GDs@Oj%VTTl4xsA)9$P0>v^=JkP8T0+qJhb9
zuMi7d0Vxr<7E+1?GI$RWWnsz2SXi<N7M5&^u(;%+94xsQ2TLx&!IDc64wrnChb155
zVaX?WSn?^t<5GwUuoPkfEQN#sOCd!BTn3{-SO#N3SOyb8SO!xhh)Xdl!cvThuoM#_
zEX5QNaT$t+U>S;qU>Qn;U>Qo05H6*t1WPF<!BR>{u#{3n!lfLQVJXLCSjq_*mU4>7
zxKyGFER~o7OC_PeQb`d7muggnr5aOVsU}odswtx4Qj2P^)M6SewS)#sEk!h3>QNn*
zdQ69<p3q^br-+VABWl3Xh#9an5(X@d6ftmVMon0nF%y<%!i1%nA|@`as0B+aX2H@*
zSg^EG#KNT=wP9(;Y*^X}8<uv8*tm3}4lJFR14}32z|u((2bXTtg{2#FVd*AZSh^|V
z;&P}ywDn0$$Hrm$Fi0cx5v{58DxiBs%d>yDBP)==F1&jx5Ee66ARK1CKzPhTfe4s`
z1ro$87Kn&BR3IVDQh`XA<pPm0D+QupRtrSMtQClcSuYSBvr!-hX0t#{%vOO|nC$|w
zF*^m~V0H_{ZOf%sPpgZM8|Pj-!S$M@fm<j%4CH9w9?W?f*n+v427bYuW&#-o_8t?s
z3v<R7SmA5qY?iX<V{?>6ADgEv`q*a5qK{28Ec#fFVbRBCd@TC7EalM0<tT?fE>Ai1
zam|!NAD3o0^l=`;p^wY>IP~#Z%A=3ZQ67DKp7QA9n<<YzKF#px<2{B)AD{8@=o7M3
zK%bDK0{VnJ70@R%QvrQKni0?^c#MEPA>$L!7tGQi`hqzcL|-sZgXjx3(;)hSX(oui
zpvMH!7tHvA=o7P4M4y<WBKpKU711X)QxSb)ni0__dW?uZG2;`_7s}EQ`a(GxLSHCP
zL+A@N(-8VXX(oiekjI437s~iT=##QkLZ6hQ68fY(mCz?OQwe=invu{ad5nZUDdUsS
zCugaQJ~>Ba^vQWDqfc(8GWz5+Bco6D7#V$X#wVjs$x;P<N{%Y%Q}R?npVCYf^eJgZ
zL7(C=3i_0cPeGrWr7HT=997Y$=BbK4wVA5uQ`3x!KGkDX^r;!2iasq%HS}pYs-aKI
zQw@DuGu6<ir5O!<n#XA9(=t8{eR`Ja=+kplN1vXjI{Nfxs-sU&GdlWokI~VmXM8&P
zj4U<KXXL1XJ|j;J^cl_6K%bFj4D=ZuW1!E-_zd)!S!$xs%uy44W}cepGn=W2J~Pdj
z=rcXWM4y@Qndq~!)Iy(?qZazCJhjkgHB$?HR+_QUXL*c;J}cw1&}V0<jXpa^ZS>iB
zYNO9?rZ)QQG-IRB_81#|cE)F;&&g5;eNK)#=yUSaL7&r19rQVA#zCLsF%J5ijL$)z
zo24%L+#Ge$=jN%4KDU{==yTJIi$2$5T=cmapWEgeddv4^r;aWBg!mfAbRE-JdUbrU
zvu;5TY1N&iV`Ja3eJ7XRHJdc9nme*K+^%@KzC(TIQhUxSb_uuj$MxNcFPHs+AUs#`
zvyeDH2Z{IdkOaQ~Dd-PE68$2ikUs=T@=K6pzYIz7E09#b3Q6;8kaWKe$?zMHOuq@q
z@>`H>zYWRpJCIzz+nR0|J7S+W#hy^S-f9|ZE-O7pFI`czqUF8zw&#@itGMN8L*I+u
zUe&v}t?E+Mqx43_?d@NDaOLjgvf}C9y?WP_-hkh-ty8-{p8n^CbHAawdad*udi1;W
z;;El+SW`S*8>?+BiMKA=)owy!)5zlKUg!1dQF@a5f;n+!yM|5Q6;F2ybdyRSYV4jg
z=$ppc{jU{IkM@nep!9$L<BMfO)>b`!pm@4wYE8WK!+oW;JJ!B)&)bWq>$}(2mfEEc
z5_(j$e1OpQGY6c&fYSHXv5|FHNSwn#;vF86;0Ta{P7splh>$`~2$JMTkYq=Oq&NyB
z)lng7js{70bV!C{Kr$T@lI2*CY{!P=I1VJ&aUl)T2erPq`_8}h4m#lRb2rQGf$nPQ
zx-9+ji=UgE`{)I(({Nx^+jSa92L7Whv0)Y%n`D7;VGbCV<bd&E9vGkGfeB#&n2;2J
z1;asL!DJAa7#4wvNfB5m90C?fhJZ<737C|WfXQJQn4FY>DPaYel2m}HVHKE~RDo$>
z4VaeHfazf!n4Z*u8DRsMku-prVH22{G=W)R3z(I(fZ1Uin4PqNIbjEwlXQT&VHcR2
zbb$@l2e-aa+OcuCKD;%RKEW%#t#Iz+JNOfe2WG<`Rh)&y#W_fPoQEXD1xUeo5Rw=d
zA%)^0NK#yaB*$e)N?d`Y##KmKT!W;?bx20sfMmu^NLJi}WXEktPTYay#@*JGoOSh#
zbDwTrmyBN*J$GT+xx@c{U0%9SR(3_}MW%7cD-{zO7tFq~{L0uB(=MN07Hef?)7vgy
z*>*9Wn0fV8#W;5M#f-0Ad3%3{%8s47bUtr%Rl{($s$%r8>YiP@^*Z0ZXkfU0(BL6*
ycX2>}o3H)<&i3fjx2CS&1s4W5AsA9r&9Iz5nY|@_ME4$}#>}82#|`~6iTn#jfXWa6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp.EADFE46F5C6C50DA.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp.EADFE46F5C6C50DA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cd64544a59691a06d83087c8010b6f6a068d5aed
GIT binary patch
literal 24924
zcmZvleVAiqVYlab&a+E)7llPwgq1I;GnssyB$MPM<mCH1ISDg4NoJCh$=5+y3Q-J5
zQD22L@n|iAdXyAPYtU%$l_*$(xTv*QilbGE)K+P01*?J<k*HAA+IQGq!~Xtw_qsUO
z-1*(R_r*XaCz;IIVkVg^wU?G+mu0T+zW&-aTUuH=y!+2J*Y>V>ky~2&_@Skx>u-Ok
zqFj2^cjogS{oeHJ4sz=!K5+ewt}8YU#Xt1k)Hm*V<nOnBc+chU$b9YhzX|;1&eaF*
zs_gs6?4vimBlz3fpRKP9i_uS+5B_lYu?N5L(~0Hh!Ji+9hVp;C_}-76{@&-$3H|vk
zw_L{_zV!KDKJ=~ri%+(0NL_vH{TE;Vf1mu_V)>9Q^v!R*^Z)$aL$^G;_T~G({E=Ti
z_Rxzvuikh24Zh|t9{T)afBxcke!VaBrSMMfE2F3HJ>^@~AD!DOpPpr3>dchEPaZt@
zR|hV!KmCz^+INrZD_8#F$f?S5_sqvCADR5%tQVR$JoeNl&VJ`F$4~t9#(z-Srv{(9
zmh(J$@UCy~b9W9b|HtfwZ~f$jgTGs@Jbvq2uQ@xg{Oo;KfBEo>AH469Z!dlFEAzKp
z`zQYH#SdN68eDhZ+21(rg5dd&Nq0SP@5OgmjqQ)$s(h{Ldhw~xebs6__n$8L`g={r
zxKnW-tesYj56?<p{KfZN&tDZ8*$wyp`ZI5Su5sSKUi9#tN569PqTl?%$A8(r@gMxZ
z`sAYzy#3tHU%79)(ignA`gr_ZZ@KQV`_DY~*%yf8aoQa|{C`Bl|Bz>&%3g2%(5>&u
z{_DZ!CvU&mdPnlzH}<FR+W+xmH_wkGfB4XOui3uI_Qu+Gzixe-cG1r8sf+Kq@|suQ
zX`6lZ#^*m*G5-4QSN4xY|2*=8i*9V*u{?YKhmz;MbE&g9^53O<|K?X0&ToJ8x%*SG
zABjKy==tY;`>A7Z{7&J$*M0P%XFf4mR+P`*cdhk{-~9fs?>)ux9{QB?tbNyi<d2Q*
zrDFJFU%35U<#+t;yN?BZhwr%J6%~oE5AJ;0^y4?~L>~XmhgI!?r5jf6J|p+FrwiZs
zr+t>%zdv!_TkG-9=J>}=pH3h7i>JST+Ka`1`~Kkl<@<)F-r3)O{wBji5AsKT^4QOP
zU%kBWo$KE4v0MN3%4@H<{jS63e)q9Yyx(0}`QX~|N1wawmi5%Jqvm`6*yOmz_SX1^
zpI@=;`{=vB^5{ihJ^!B99lhY@=ab%|<AxRAn{GJ#wp2rSMEl+uU%c!KKlzKFeCAzW
zKJ5$lU3Bo&3(f!Xx7laUxb71Ejw{~z_;-Ky_x62eYd>`BbN<&~f9!oXyZOIMebM{#
zM*f>ue&cBA$AJ$WNbP*-@ef?3Jb3rP+y3ZtcieFEE0b@(>QmNFUu%BdZH@Pw`{VDr
zo=9BtLe2iFXCJ=&Kfd;l=Qyu?{Qf6|Pr1X3&1-M|rgp(2kKFmflZ&6xxBDM?X5aI_
z&cDm^jrZ0caKBhM{RhuJ_`1`pmo7hW&TZ1$ex&bQ{YK|SZ~ENXjxXPG(~<lC!v6f%
zFE#4!-zU%d<gcH3?{)8Z_Puw%@Zq-vwIi1a`!D&gv$wwEeeLf*`|Zo_yn1-<f9<>W
zzCSyBbo8Y1pU?d6eLL5k^5_k>3{xNe{cGNFn&+>(uYJWAANZTnJ;m(r9>3&O7k>M>
z;LWD9ANaxz7tBJ>Kk|!*PuIOKd?5HYFD8D!?`oU(kCv`^YqwJ-OH2DNxc}zZ9m0i&
zXr;Q;I&1$~2TmMzzeb$+b?`Ui?+<cr`4BDd&gbnv@9>Gk-hBA8e{$|yKJ(MB9-<q&
z^BMckIB?>yH-~=v+%tFn_^qEkL^pTmbM~Ke%8A3?T>IK>r~UH{{P-11rQP}bBj^9Y
ziNoIf=JFTs@%;DY`XO4|oiE&f;l_!>-hB554}5<A(a+v?h|0V3;r)jVCk}h_kItNL
z8Sc`*dg{{F?tJ;a%P&1~*qeXazVNu!^~~BKdg<<bWdD%^Ck}h_ukL>Cb4vfw)kAcB
zcRsrR=;0HGz4^aSuYB&`-t@OWK17SV^Q-p1>cEM^-n`U%{Tr?^C?`L=_mhJsKCSP2
z@VDc;p739IYHOKacJBQP?vk*436QjW=}B7QSHN5mR=`}5R=_m!Mlg+n5lo|G1ap;N
z1#?wc1#?we1=GZvz%&UaFinyPOfzo=(=3?5G)rbMExZLxi(moMB3Zz+@>Vddf)z}w
zWChd4+rYF5HZX0H4NRWr!Q=%VOkU!_wDWc_?SdUlyJQE`!8^co2o5kEk^@X9?*!8+
zIKgyEPB7Q_H89tNH89ttH85Sg3rv^b0@EeAz;yF&Fx`S1Ot<6))5Ckf^avg>J(34Z
zffv9O1OZGz62SEGUNF6a7fi3@1=Gj-!1M_|Fny8_Oh4}j(=YhJ^h<s)MP39`6htsZ
zNdz;%2fz#n0Wbqn0L&mC1T!cE!3;`4FhhI@%#aWQGbDw;4D(?y!$KI$uoMO}!biZ2
z2oW$NQUuH>9|bciM8S+oQ7~hC49u7i12ZPYz>M>8FylfT%(xT>Gr=doOb7`u6H)@q
zB%cH`DI~#6N=YzNd<x8zkODI$rNB({X)x158qBnm1~bEFz|064Ff&pH%q*V;Gb?1l
z%t~1>b9@fWoR9-EC*{DDcnM5Nkie8A3Cuj72Qx3^!OTl}FbjME%z{t=vmh0~T<6!p
zTo=~CT$k3t+~7CB+z>Xv+>kcFEb>J#i$W31qErM^=4CKtK?YNnWH3v73CxmE0<$EQ
zz%27+Fv~(2%(7Glv%*)vtOyk_D^dl_O@0&1O<@zvO=%O%Eq)8kEny4HEolqPDqjV&
zDpbL&N>wmxd=1Q+Py@3j)xcDE1x!Uyz*Hm!%sO8Ovo6%ZtV?w;xA|=_w}ovmx20_`
z8+-%IhR^`BAvM5k@=Y+CLKDoU)C9A|x4>)(EihYB3(Pj(2D2@+!E8%yFgtt)%#P3j
zvm<rDRCyIlRZziHB^AtL&SUPEPd!^#IIjS@(s|{}w0C0t(wX5)r`2#~sFCFlkdx<s
z%Q{*9OXM+G{xtB4Mpn>Y(a8$>D`T>PzELAa^o=?(qHi1%Bl@cvSw(+UC#&eMj>#(e
zCXJZTH|fNLzG+NM=$kcSM&GOxGy3K+F{5wMhy{I%PAuqK#>9fYRU=mPtva!yZyggW
z`ZkT&(6{NthQ4h~Z0PeE;nC-H!lTcR36H*ABX;!dI<cc~9}_$J4vjd_cj&}{zGF-r
z=sPvyMBk|sC;HAYaiYJbku~(!bh3v2+L)}N@6w11eV0yL=)1<mg}z%OZuH$caii}Z
z6F2%Ejd;-a=){A*XG}ck3mOs77jz<^FN}$RzE>k&^u0RqqVF9OFZw=>_|W(1#D~6b
zOnm73HR4C#uM<D|{xR{RFKR?YU(|_+zBnc#`T>mu&=2S&fPP?10_X=d5=1|!lOX!R
zF$tm{(ntvXkWNDAhsGp?epn-6^usy{qaPlVF!~XVM9`1uB!YfqOd{w<H4;TXs*@=C
z(J_glAJa$-{g_T-=*Pw+hJIWlarEOliK8DMlQ{YbjU>=d=p=!DVoVa~CpD5pKdF->
z`pGd#qMy=83jLH$Qs}3~B!zxjBWd*0I!U9S9+Ncs8I5Gn&*&tBer8NE=w~&OML(;P
zEc)3o$)carNDlp+PIBny#w3Tnq!9^yNhcEe(wIo-=QWZ?Kd+NK`uQ=*qhHWS0sVqb
z3g{Qcq=5dqM%K|^*U38i>tnKx{)R?2(BII>2KpOgvVne4BSrL!Iw_)G9FrpYvPNX|
zWu3_A%VQ#=U(!ek{gO^f=$FQ%gnn5gW%SEBDWhK=lQQ}hja1OD=%j*vWlSpQZ)#){
z{Y{;0qQ5yNo9J(8WDET*oou1MH6~l=S2a>azp9fe`qeS1qF>WU4gH!<YUtO-q=vqt
z5e0okCkpz?m?-GiHBv{vu9G_Y^)acVzpasN^tW}gjsEtSY@^@MNCW+bP8#Sp#-xFM
zQzK3En>uNt-yD-B`YnyL&~NFag??*HTIjbm(ni0nlQ#P8F=?aU(MSjVj!ru0cgCcH
zzN!%weN`ta`s$dd=wI!;I&ks~{LRic1NEG}muc^{>(B3=fxq<1HJpLV%N02P?k-;q
z)F;bBJo}bcz+LICfV<KsE9e?!Be=${5nN-R7|~soSHWHFu7bPTC#&e1WD~fit_fUI
zpP0}!%Vu!RT{F1mJ~5+fkuBg_x)yLPePTh^DqF#|cCFxA`^1W_O}2q+>)ODz^@$B#
zUgp8&yF9pjpYZ6~Wjna`t{q%^pV-lL$PREFT?e?1K5?Myl%3!@yH0SOed0uSO<n_c
zt-A*9TA!?;>ylmIy1FiKU47z0*Dbrjb$8w1y8Fb9u1EHO>*;#H_4J7cT|pMW6}ker
zLZ1ledSx%T-mVv1Z=ZP4^~pYPeO({8zCQ7x>zDoD`n!H`{e9v`SCmC?#jXgh*e4>o
z0XYC}pc?=;&?f<OgK`kuU^fVEuup>MhU5^qp>7D=P@ja*4a;G0!`(2r;XVnY8<8X6
zM!FGjBYhG<H!4TLjdr8pM*Ad+ZcL7W8|%ivjrB<k-MAbFH{Okd8}E}ix(PV}Zlaq2
zH_;~vbdz!t++;TiZn963=%(ZpxT$Um+*F^W&`rx}aMRs1xamGgqnnX4;AXlRa5H_9
zK{qRB!OeEF;AZ<Ii*8QNft%~*z|HkZ4qZu>z?HfZxKf`;=;q}-xcP1#+<c$p(JjaY
za0}f6xP?9`pt~-wgS*~c2Y0<s*3sRNH^ANKZh*VdCmZM%<s!JnZV}vKpA^xRWf@$#
zD}yWdiHvSZE`eL>mcT9bNeSJuTn4w?ErVO`lQOy$xdLvbTLHJyClz!z<xOxmyPM!{
z_Q@u?Tk;mTTiq>ixB6rY-Kty#x7w|OTkVr7x;425ZmnAbx7H^$bQM_vSLrI?Dt)4$
zTbJwL*1L6Z>wQv3cU#^Dce}d{?slJSquY=h;5NDqa2tKnK({G3!EJV%;5PfDiEc}7
zf!pe~z-{$O3*ENd2Dja9gWK+tHo6_T18%3=0k_j99duP$1y}8=;HrJ1qI;F|s+V8)
zJ+<{F=bK)pqrVNB-`4xtkuy(it<Tx|=h?{<&5b$RKrYT%5m}xy8M!oPCFJs)m60oR
zRzcpJvrXi!Iom?6&RG??HfJ?tWzH1j`kd8~x94mdxiM!A<mQ|;ky~@tLT=Al8@V%Q
z9b|RR)RVmTPLoe?2T#1k1Wx8Py8Lc>^7FbpqRW4VJf+KbA}{FjFOdh_@{ohyuXD>E
zLLRc^cfo(R6^*W7Z)HSRu(vX$E7)6E&=u^h47e5St?1ke_Ev^$1$#z~8nI^_Q6u(@
zQ)<MXaY2pPGY&W-_KZ4b#GY};jM!V%=qmPBM|2f?t5dp)z10O>#op?GTgBe0&aGl^
zb;wq+XVRz%d!`XJVb3(BChVCO)Pz0LfHPswq;n?hnTE`SJ+ns5*fWo)8GGg_HDk}b
zpl0lu2b>vuW}P!*&pc#i>{&Ew!JcJAE!eY6sRetM1+`$$GT<!Ov*?@!dzK-yV9%;i
zEB34-YQ>&)O0C$lE~piI)&Xb5o>k|p*s~6q6?-;~+OTIEQ5*JbQ)<JWZ9#3=vkf>K
z_G~(5!=7!(Y}n&9%43foQ678zl=9f)7nH{yKj3)m@jAz2j~_A~dv=Z5v1cDqJNE2T
zYR8^^LG9SH4>&va>^f)1o_)yd*mG#qfj!5FI<V)MQU~@N3+lk0W57AE=g>I^_8ddz
zz@Ag1PV6~H)QLUklsd8JTu>+WoCD5@J*Un&vF98zC-&Afx`w^A5naRH+LW$gZ*4)>
zu(vkg*08sxb8FaJ8?rU*xisp+o@+#1*mF&(3wy2wbz#pn;9S^q>6{CDt|4<_&#h56
z_S_@t#-4jh-Pm(4s2h9k0q4e^Tj$)^a}SvtdmfE?u;&?35B5A$>cO68K|R>>3^)(=
zJUZvWo@dBB*b_7=U{4rP0eixf3fL1CRKT7v-~{XmIwxRH7%~BSUX6OO=N(Zm_PkT-
z#h!OTz1Z^(I4}0RI_JfncgVch^J&zFJ>Q7>u;-gnANG6;>cgII!1=J}(>WjZd_(5L
zo?oMW?D<F3k3IjC`myI<P(Swk1I~{<zs~uw=N~dZ_C$?}*b_%o#GW{%BKE`u6|pA{
zI1ziI&WYF)hfKs?K%)Wd1x7T0y}*<Puoqa+0QLd{E`Ys&&IPa+7_tENf*K8CFF2w>
z>;<PZh`r!~2C)|$a6#+^buNg#;E)Be7t&}5d!Z2xVJ|eLA?$?~G=#m-fD2(Sq;nzc
zg@!DIy|6~Z*b9$n7<=I<4P!67pkeHV2V59?VVw(OFFa&n>_s#h!CquUBiM^fX#{(b
z1&v@YGT<WEi|AYgdyyfFU@xlCDE6Ww8pU38N~73|E@%{c(E%65UR39z*ozKX6nim^
z#;_L~(HQn(QyRlwY(Zn#iw(FK_F_60!(MF2V%UpoG>*Oah{mxOpVBz?;tLwbUVOmC
zu@~35IQHU07RO#fqY3OKMl^xF#FQqmmsrpQ_7VdwfxU#zC9szmvIO>$8ckv^IigAI
zC8sorz2t%>v6mchN$e$cE{VP5kR`E~(r5~MsS!<KFEynp?4=eog}u~(OJOgib1Cel
zhAf4>v_{j|OOI$8d+8}nV=uj+Y3!v3TpD|6ol9dcJ!EO@Wi*<>US>oy*vm|5278$W
z&0sGx;4;|D=v)SSnIX$yFRRfk_Oc_I#a?zwv)IcnXcl|f0hh&IR_C(V%MMu<dpV8f
zu$LRr9QJZkn!{dhL37y44Y(Zkaypm8UT(;8*poCWVNV)S3479%O4ySYRKlJ#;3Vux
zIwxUI8Zrrcd5z|=mmkqQ_VQDj$6kIx^VrJ|xIFgqI+w>@e#r9JD`>QUy~2nVuveJU
z0`>|ETEJdmz!k7p(76Kk3PV=F-nvHDv9~^=>)2bL(sk^uFX%e<)(6}=_SSW79ee9T
zwvN3Gjc#CXV?;Nww=ty~*xOjp4eV_UxDD)W=-dYOHim2idqs^Fu~!_?BKC??TEt#)
zL5tWc4!9!riaJ-sUUA5Z*poFXV^1DY8GG`S%Gi?^RK}h>;AHH{IwxaK9x@qwC5@J_
zR~pe0_DWM)!d_`XOV}$7xDxhCI#<G8X~;_0D{Hijz4C~bu~(kbGWN;~TE<>^z?HFA
z*10nF%0pJhUPYr7>{Ujzg1yR=R<KuD&<ge{1FnL-iq2KAR~fPj_BJ)TiM`Dc-NfGJ
zlx|{gb3r$;w>jW8vA3yno7meNvQ6x5X><#FTO+!Ky{#$T!rs<`Zeee0z-?h~OXs$*
zw>4y2*sE%^ioNQHR<T!|(kk|<3tGiqb--1zSJk;H_Nqfx#a>OLHSE<!w1&Ofl-96U
zThJQzY6Gr@y_(L|uvZ(h8uk>8D%ev-RKcDyr3&_x1y!)83^)aQiq0w6Q-(~zUR|Sg
z?A1rKj=lPn*0EP#&^q?&1Fnv}y3W<HS0A!E_O>;;jlJy=-NxSblx|~hdqKCcw>{vt
zvA3;r+t}M4vTf`&G}^#kV?-O+YfNbadyNHcV6QRY8rW;-TmyTJA!}f-snI6(nj_l8
zUUN#D*lR9m6MM}8*Th~^=bG4S4p|d>EseIY*Ba3l_F7Zg!d`1ZTi9z2xEA(WI@iKp
zYsgyIYiqQPz4nN<vDco`Hul;J+Qwdcz_qd0*10zJ+C$dHUPq%H>~%)8gT2m_cCgo3
z&<^%G1FnO;j?Q(k*BP=7_Ee3k*i%PT#hyB)D)!U`Rk5cII2C)U&Z*c_hfF=$i@f_k
zmtOqUwT~0lddsWc(%5@nSzh?yn~qp-dOta}bzbhg%H9uilnQ68JoN!_-|o%F4jeme
z&))kv=fJ@eKi>@Bo0^g32k^~gGqQqgoDn1P>Wr)+n`Xp>Y@QJ_vSmgr$krLLBHL!f
zhRn|hk8Gb2JF;U&9LUZYaU!qH$QrV1MqJ458F3?fX2gRm%!q*Ooe?jxZ$^B`{u%Kj
zi!&l32WBLI9GsCLa%e_E$l)0YBS&T=f*hTZC~|B@V#x6si6bXwB!QfqktA|zMpDS>
z8A&5&W+a20osleZZbov*(u_#R`5DP07iOe@ygnoA$Qv`Vfn1!CBC<RqGID7~O339I
zDI-^Aq=LLTBb&%uGqQzToslYXZANOy%8V$;^%<!nZ_mg!a$`ms$jupPBDZFwh1{Nz
zHgac1I>_pbs3*Dg<a;jP|L+%mU)=rq{RvzCEBq$#gsmVOC(MYvI$^8GrU^44n<vbS
zY?&|%vUS3&$hHZyA@dW)Bikp;j_jB)2eNa*oXBevwubDQFc-3W!raK73G*Ne6DA;g
zC(MiNn=l`;f5QC8;)IFFfe8yB2PZ6u9Gb8Ya(KeR$dL()AV()GiX5A;7;=2V;>d{!
zOCTpFEQy?&uoQB7!qUi@3Ckd7CoGGco3I?RG+`2Qe!}v|g$XMluTR)I^2UU1AQvaB
zh%8T-j9i+q5^{OM%E*-os~~Sq*e3GUgl!>LC#;HGo3I+PGGPjGeZuO<+Y`2p+?cQi
za&y9($gK%$A-5;2jog{A4zfC7>Pf!*rscP<-1YNYU%De3mQ5$$>#<^30WunlK&ysT
zAd|raWHy+AECvgZ)nEm(8Einjfd{f1>_85K1ITG`0<9U=fLsO_klWw}@)$fof<XZC
z8oWS0gAd4W@B@hk5h!2?00j*}ppYR16gGr`B8CW1)DQ)V8Dc<jLmVh!NB|`bNuZP=
z1(Y_Vfii{+P}YzI${BJ%l0gE>8}dK}Ljh>runx3g*Z?XTia@eK1}Yg!KxIQ2sA8x9
zZ5lR#whUW9RYMi1W~c!v1_h{Yr~_>qwt*Uk22j(`1Zo*tKy5=CsAK2=sRs3BdNW(P
z?Nj#O{=@Fu+Wg0v|6zD!6pu3zIdGf>kb}or5IJ<5g^<I?Sr|EToJEkM$5|9PcAUkK
z<HuPXIdPmNkdwz*5;=98rI6FdSsFQWoMn)+$5|FRcbw&rrQ=LO&L3xa<ic@QILUio
zis-=;mp(fA%mJT_T$hXgy?M*CzWc1>{KWa2d(SrB{foqFe2?68?(WTJoPNfUy&re^
z(^I~3$9o>^-MD-6(N`ZmZ|_1!@B88Hms)SAzTUX|HE_<BPrkG6)Yi(Jtsooc%!s@?
zXRFAjIWr-f=gf?3nKKKrb<V8FwmGvQ^K-@{+vm)V?3gnLvUARy$ZK=9hU}U%7qWZK
z+{m6e^B@azCLnv~%!}-sGas^l&iu&YoQcSRISU{M=PZaEnzImcc+SGekvWSXN9Qbx
z9GkNka(vF>$cZ^iASdT6iJY3V6mojb(#V-P%OGdxEQ_3*vmCNCXA*LL&hp5GIV+sx
zJ3sS^egAX+lefS09Qwjzj=ks5pZ?k4t<@)Pzj^oO3#%74_U;&WUH`zo_3p=#yEmWn
z%5#qFy$y8l+F-$hC$0&0@_w_dEN}1q-rfJRl@%bPVgy=MR)I{43COINfh>vz$f{U@
zY>EwtS9l=1Vh3_44j`xE1X@$pfLw|T$gQ}6Jc<WMPy`^a;sx?4J|Ms12ND$#D4+y@
zf=Uo5q=bONN*E}jM1Z176ey;|fZ|FVD4`^Pl1dUNrKEt;N*XAmWPq|t7AU9WfFwl%
z$}4%Gf>HolSJr_xlntPwQUsC}8K|U`fXYf4sG?MWHkD1FEoBR+s#Jk$N)1R+6rj3N
z2ijJ)ff`B!sHrr8T1pG3t+atUN(V?)RG_Z2>)N|(-n(+3{Qhs=@$~L}@bKw}&)mE3
zec4sKY+vGIKij?etk<1o-Mi4>gQeip%?tkd;oY0BW!JuT?+^LG4{r>A;PBo5vU~IS
z)$`T82eG|*_h&R;x`f}}9r;K0eZM;8w6p(o*FJXEu?zOBy^H$o+Am$vZ_oaj;jUZg
zC3nB{YTc379kK6Odk=kYKX~HWetY)r1E9E^I{8Inxwrg|le8kPz_!v`fz2oyVKerO
zu&s)#u&wr1VKa#)*i1bWY-Z66o4IF(%_3T0v-B*mSw$;s)}9qMn`nd0*0aIJi#%+6
zkB7}J+F`Tz?65gR2W*a>12(7Vgw5G=!nP)^!M4_0gUuznU~~0cu(?GyZ0?>LHjn6m
z&C~P1CWr!TLQjCrD|%t`_Pnt9L?3Lvo)0#^=!ebU^TQ^JB5Yz$ge@QjU<>pDum!~+
zY{6a-wvZTtEz}Fa78b*>g?nMxB4Px#NG}3gRE)wF?L}dWi80t>y%=n9F%Db27l$n&
zCSXhS60jx3By7oE61J3>f-Tic!Il=&u%&xx*fL@UwoESrTUN}%mhEL>%ZWMIa=jdE
zk|@C@^(5HxVji}9FArNmEWlRi6<}Kz*I`@lt;4n<ZoszD+kmYo7GW#)im=I|44d4O
zVJnFx*h;+;Y-O<wTe(+;ts+)ntMn?cZHk+)ZT2={+Y+~6+v;t>Ru!wTReM#~YGMtx
zTCWD1A}X*cJq5P9Sck3NtHZV}Zo{_S+lH+nHehS?8n89RCTz`K6SkJvg00nS!PXYr
zu(f+_*g9edwob1Dn<}cXsXZ08KXm@#%U3Npwe?2l8(*fq>(K6gXX&Mj(85EQy8H%s
z&{9``jH(f6Rb2%#sU{$^Y6h~X79gu?1+u9&AYSEx?5Z8ep*nz^suO5UT?2BdE+Dt+
z2J)yLAVC#?ys8(-r}}{Wsvk&HMWBEh01B!>ppY5@3aeqDh#CQks!^br8Uu=}aiD~n
z07|M!pp=>dN~>w0jG6(;s#&0%ngfzl2`I1TfeLB?XkA?g+E6!uifR!^R%M`)S^_Go
zWuS^$0oqhIfwt5wpsHF0s;M;~MOA?6Y8_}>-3Dr?4WOpl1Zt@*ptjlu>ZlzcRaIZ6
z-+$>}o-BEusl4<$)(-Q38(+uTVIp#1hXs&>J1mGC+F>E&@D2+jM|M~QIl9B5$gv$3
zLyqsTIC5f#C6JRlEQy@jVJYPF4of3vc31{EyTh``xgC~6mUftgoZn%2<iZXsoaEl?
zuiE`v2zLM8fu4Ge{L%;Ymk$5?YkTj5rKQHpA95|z6K}s{eZL;v-&Pv+?dH<%)wK71
zHTOPPS~~Hk{x7QC%dR@M`(f@4d5-Pde~RJ2p;KRR`f0Ds98ScR4<F1#&p7*4XTJIz
zug@<Af}ybOEb}E7dyE(S|86~U^fl*Qc+qSB;B_mjCJS$Otht5rUFTYz@w1MkvK^W%
IMAAh1e^sSpKL7v#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp.5656416FBF150D8D.idx b/.cache/clangd/index/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp.5656416FBF150D8D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..75b128e083a6cdbc5bc0c0d463ca67929d4baa56
GIT binary patch
literal 24924
zcmZvleUxKmVgKiO?z3yy1$Dt)<z-<hGLy+9xk)C;otT^V_v9u_GB=sz<~<i>DFh+B
z2vwAjaJ5Ql^;%LZm7vkkN(7XEU2U~mkB%?Z+Wxc!OGT=M0#zg0f>b+f&#>P=esj*=
zGoO4vJD+p*Op?3V%-upN9+%roOOZ=cH+J8AU7IZ}EgjnZb?tS%%U|G@mOgfHY3ath
zzEY}Qa<1ofzq<RMr(d^|J3s#Z8*g!5o<A7<;CmC_c<|fbEI;|+WjCk3_L~=dAKAJ3
zwEIf?{xQAtxBk0t{=2t*>a6hUJ^y(4wrj%qC*FL0;mF%A_%C+P-JfW_@a*FGXCJ%$
z?r(iH;l1_Am0!Qm`S&lr<C+J)aM8w_esD-EUh~eI^nIhd-*Q>N|KyLp`d?=^Yt3)H
z@br&fy7YGXcVD=R|L-3?_Vl-2`i~cHy}L2~w~K%Hjq4ue{{6n*84s8Y7vI`=@O!hr
zxwd`VKeMBMa$WYBTg&JC&~@WM&7!;Cd-+<n@U!>+`u;Cm{`_|je&MF?Tz!4;shMs6
z-8Y^f=U%g~^z^OgJjl*keDCf5{)?X;cxk!x=$$uR`+DE<PanSeuMfTOfrsDp-K9@_
zaen=Ezwhl{^p$H{gBu=x{Wo57f&ct}kna23Ll@m+F}6Q;XZ33p=L=7L`pXvMIp2HJ
z*WYV0Mjchx0p&G?=<smzqR+qQM*hms$Yyxxm!G=k=Z&L3xbUm@p8LgfFMRQhANxi7
zmhX6<|HR*a?(OGvKI8h(6(0XZl}DquU4O$reD-z6@Bc4if0TBIUwwyY_$~SAQ|TKm
zAH4HD=^q?ue&Vh>Ece9UeM^6O+y0LozhizR{^VDVo>#x!dX@4|U$^|8cHz$Osf!-G
z;@UUdYn^@hmS26kWc=j=ukIfSe>n8R3vX%OvpoBw55~`V=Tc{J<bULc{^t2(^Sl23
z=buePek}gv?(>g+_o?Go{lDCMZ@Bv_&wYHdT&+Ix@O75Yf9vsIKJ+T{d+1Y+!~1Uh
z(_4-0r9$wde|gt!#hd^3-N*f&L-$;MMoHppgL|Jg{p6~h(4#N@v8sJ;>0K)ioSFIB
z)46YaZ=YHD*C&qNREyr9;U6`9GI`|Bo_^;Ai;ti9t5;q4$tzDQ{+spl&z*hfy|*>Z
ze>eS?Yqf9R^%u{5Giv$E??3h3jmkyN2d0?|uedMw#>=kz*poZw-*d-z+@<6Dik+X-
z7Vj}VcS-qx|26%OMKoV0<MyBT+#g&$>HH?wd;B-Q^ZXs3zv?4reCFaCcY5FUe!G6~
z<_mh=$m`w^EIs`DfBg8x&tBIE|HDfUJ^1|(wGWgO`Y!X6Z~OYszH`peD;nqhqVtY_
zy!pd_`u#hFd%yX1<5_on;O+B^{`uiy`>^Aumqvbj-qG8C^-qOr>FHA1@mH^Pee=P~
zzx2MJ{O7#2{P-V!`)lw2^Yu%=_s}(8erw^1TmJ2lr_THE8O@K}o&V`qe*FD?zq;Wg
zj~#sWnGYZ6-g)>hZ@6OdXUD$yueW86eflHP4eO2%J^%F;{fQr38h-pzzVnU?&bD0b
zZ9n_R_irv=@zU)k>6yn4KlkF}huc40d*&UdTfTPH>DT>?f8Z~!iN4hQ_|E1N2G`FH
zK6R!2BDeJ`kNnePFPy7C_v(-S_WaZS(DXl^7tg=)osUJ&x$$j-Kkt0}ga7^T^~=G@
zlQ*ANwBB&-Z~e-<p1u2tcVD=C&3|5W^R2&l-%AI3jg7Z0U3*ivQzT1E`!D$H9g%y4
zV+U!eveY`f|L|!i54(>MCqE9nIDUMPacv!>#ohVn{-cLZ9`@!#_y7Jm*MI6izI>48
zcjq(rpLyEJ!`^H@<4ccRQdKHvEXlj``A5!w<H^I`9QgImpS$-bH{E}bZtu=#??3xh
zCl7n`+bsKj?0-!9%;`(z-TAV8mtAu5us1*Ofzuw@fA0MsI!L#6=R^As8BQMd=Kp9P
zJ7IA?w|<abvO6Eyf8?~2hrRjv2Y&wPYX9$757N!u`P}{I9y)p0n}7NJ{a4xFa`#0C
zsj@pC+kY&7@~}7m&udGc{+GA@?N1KU!tQ+5{<BUydDxqmdT)NqwT9}cKi$ug11JAj
z|KLl%9^dzb_t@#JWq#SQ7uh9Y`AtC5@+GHeg<k=4MOXoIMOp#V$Q!{l3Pv!Ek`c^R
zeih7BVHM0(X%$QpZvxXKn7}kiCNS6dH89tNH89ttH89P*8BDWa2GcB=!L;xeFfD=w
zOp9az)5=@Hv<g-*t&$Z?p69{j1s+UZ;=#1>HZX014NRM41Jlmi!L$o@Fzu2ZOb71(
z(;+y(bVv>`*ZFlY*M)U3*QIqZoxBrFr{DzBDLKJ(@h&i3f(uNS<O0*pyTNn|ZZO@F
z8%%*0z!U@lOhFRB+~7CB+z>Xv+>kcF^za@qJ%R^JkK_T<%X`7}3SKb1k{3*o7r_(-
z5lm4M!SwMyFnxj#OrPWf)6e_C^b3A4{gNNd03QG|AOye+NC7Z|d=Si_5Ck(Q1;GsQ
zAuvNi2+WWa0yE5q!3+ywFvC(9%m^O=Ga^L5j7SkMqkI(1s1OA+Dn-GJ@i8!CLJZ89
z6azEP$H9yXaWLaj9Lxlt05c&Zz)VO9Fq3=|%%qS6Gbts(Oz|l&Q$h;Nl#~K9&8NXk
z3u!RZQX0$*p8+!?WWdZw889VY0#gzsFeOO>Gs|be%nDgBvr-n!9G?R-C*;7)NjWe#
z`Asl4g-tLwrA;vNd>+iakOwm_<-siQ1uzRj0nCC_0CS7q0&`2)0&`2+0#oK?Fl9jo
zQ<h{fi+mBxqEH00C>6mh@g*=zLJ7>0R04CG-v)DA*amZ3+6J@Cm%%IxWiZQ98O#b_
z0ka}hz^q6WFcn?_QxOy}6-fcJ%2&ax3RN(xQWeY^Ujwry)WEDsH8AUZ9n88=2eU5K
z!EEpiFdISx%!bqev&lEXYzj><n^F_Z7T*H1CA7e7Ni8tjd>hQR&<3+DwZZK09WXmW
z2h5Jt0aN8wFjYYXQ<YROk2{XLUO91XUGBIX=nBUbuh8C!^~-05FP~P!nW09O-%n1R
z|1Ile`6I|<viwQl6^*Q*zoL^B^jF4Y1%0DNjOZJ6Vnp9KCPwsEHL{BSs!mqXUmcTG
z^i3Kup>NWO34PO;n9yI-$Qt@<I$1-1ZA{kCH*3U<zF8+`^vz>pM&F_l3;GtFSkSkO
zi3NSDMy%*tbz()|Iwn^1d5!Ss^E%<t=f{Lc-=+~8`Zk@|(6^0=4Sl;t?C9HdVn^RT
zCU*248gZcS(1`<m$Cx<KU)RVw`s+GbM}K`x*3oxr#EHIBCr<R8W8y^Lr4bkUE}gj0
zca4b)eYZy3=(}~|M&CUqZuA9>2<Qts5zrULL_mK-BOB;%=wt)^jWO9k-=h%^`W~Hl
z(D#gq2Ys(byy$y%;zi#(CSLSKjfm)rIuX$q$3#Tmrx73eKArf`_l=1UeZNNh==*iz
zN8djte)Izx37{X)NdW!8m;}%dY9xq$P$xn3gJTjzKctZm`XQZ!&<~AC2>q}|!sv%}
z5=K8fCSmj=8i}AE(MbgT$e2XXk7^`}epDw>^rK@EML(vI82T}t#L$n8NeunCM&jtl
zbrMHEJ|=PW6B<dNpU_DH{lu6g&`)Y4iGETiN%WIrl0-kHkretVoutrDjY$gqv_{hC
zr*)D>KRqUC^fMaCpr6r62K~&KWYCv1BB3woL_%L06AAsSMzZK<b&^FtJ0@B5a~jE^
zpVLVW{oI)3(BIU^Ci<H?*+hSHOg7QaYb1|;UMG3<^J9`nzo3x<`URa7&@YTh0sSqF
zY@xrUlP&bO#$*e9StBy~vQA|5<uQ@bFKVQSeo-ey^owIsM8Bkw68a^bl+Z7YNeTUJ
zjclX8t&?r^x5s1~{jx^N=$Ca;M!!5JW%Mf=si0rcNd^7Nm{ibLG@_ud=tMzZ850Hl
zsz$2lS9MZFzd9yW^lKWap<mNU4gK1f)X=YMq>g@FCw27eV^T-Ip^*mq4V^U5Z;VL;
z{ia5m=r?uJM87#EP4rtDX`$cJNelhfn6%JuYov{STPJPw+hfv3zoU^3`W>Bg(C>^%
z2Ypo|D*CEURP@y`QPIEJakcN%8Td7hYk+!=-Yc~C+Vvy5XW%ctat&wT%5ojfzq`v<
z1NF)B5YN7q6>wL&E8wp5$qKqg#R#skYXsNWCq{Hvl~r(8yQ|=?_Q@)`CdCA<scQn)
z)F&o%*OWDI*Sc%quJy?px@N@;uDNRl*W4#&bS;VnTuaviuBA^b=voykxYn)}Tx*|L
z(d88$T)xYL%l8S7u1&FlYwOy;we^V&UAtlj*WR^*Ywr^~x(>wwuA}P!*U={qbk~)2
zaM!!*;I8+{I=W8939hs21lQRoPIO(03tU&%1+J@4T<E$LH@NPu8(epvxX~390bHRg
zfGhNgfbNE}0q#b31Kf>1*+AE$c);~^J>Yu!#DlI^@q+8^dcpPfi5Fc_5y2I^BDi9o
zi0JwhAGp4*4_sfL_|Ww$esKL=Ke+xr@uM410^kO^0dNC-5<oYo1i=k<gWv}HB#3TE
z34t5xhQJN=NeJDr5(YQi4TBr*lQ6mwB?4}w8v!@cClPd`N)+5^HwtdFPon6?lo+_N
zZVcR5pTy9OD{*k+-8i`MK8d57P!ixKx(RR-eUd;osU*Qoc9Y;H`y`2ON=bp6>ZZU=
z^+^idw2}rl-A#j=?vpgS86^X5rkepb(<d2pB}D>P>Pp~BeIlWoRkGk_yIF9veUe2t
zr{uuRb#vh6`Xq<$rm_j{W_J_Z%|6*gH?QQu&3E(Q=KCa%Zb2!4Tj&<RE%ZqN-7RGc
z+^y~wxLbX)g|4i~;L2SYT)9tVbc;$6++w!~Zm~~_=$4ccxTS6h+)|&E(A`$H!QJj|
zgS*`)+vt{+GPvb#8QgN8l+mpy6>uxv3b>U%si3PU3b;yF0axi01>LGr1-II*f?MsA
zD!MhL25zle1Gm;EHFWDr9o%}i4sN|q>gYC<2DpuG1KdWRG|+7-O>mptCb-Q$X`<Uw
zTHv<2EpS_X(n7bbw83q6+u*kQq>XMz>44kmcEIiQNe5k3QNdNaD!6K&sOWYb-B(`s
zJ-zi-$6H^abARn$`@P;XM_zY&Yje&vA7Q6XH1l(oM=s1+0eNfAwvgpHlaY&aRzxn%
zSqXW2&bE=ub5=&K%vl9lnKK2sI%ien+MLyp>vL8|Zp>K&xjAP|<kp<EklS<CM()g6
z2U(pn^%U>D)5Le+<U35@Tz*EDKZob*GqQqgoDn1P>Wr)+n`Xp>yf!0i$mSU_BU@&~
zf^3};D>6SLJhE*@Y{>Q*u_HTX#DTm%BkRb{8F3=JX2gZ;o)I^)Fe3u;#*AzrduGIg
z?41!WvN$6mvTsIw$o?7eBL`+AfE=8WAaZC%LdfA62_r{lB!V2BktlL(Mq<eE8Hpn&
zW+Z`}oRK7QYDQAX=^05QXJ#aWEX|07oSl&@a&AU)$eS~=iJYI2JaS=13dmbCvV|<q
zh>TpEks@+wMoP%rGqR0bo{=(gWkxE<%8V$;)fuTG*Jh-KT%VCTa$`ms$jupPBDZFw
zh1{NzHgac1I>_pbsHb@E9bzBn4xD^z7~HQly8Lc>>fv&EM3?^rc}kb>MPAV5Umy>-
z<sk>ZU+0!Th&*J=x52~libhwkw=$wD*jt&>73{4n=nD2$2HXnvR&;I!dn-e>f<2=~
zjo34es1bX{DK%oxxS&St83&vZdq$lzV$V2aM(nL>bQOE6Bf5&c)hS)Y-s*y`VsCZ8
ztzvIg=T@<|I%KQZGilU>J=2JquxFZ56ZT9CYQmmrz?raT(m50MOhabE-kL_&u(vj%
zYuH<x(lzX@E$ABd)&|@f_SST64SQ=twuU{kM$OnWkEj`Y<|#E}&%B^!?3o9g8GB}(
zGh@#@WM=GHG-|<~WkfC5vrMT4dzJ;YV9zq(EZDQ?oCSN9A+uo5s!=QUtRrg0o^?vC
z*t0IE6?@hJXT_dX=d9SX4w)5uyheHK@gvG(kDpQ=d;Eg(*y9Hrk3C-Jc<k{*#$(T>
zQ5*JbBWlB*ZAxv}vn{9%d$s{*!=6p&Y}m66nGJh(joPtiA5lB@>{DvTo_#^>*s~8f
zJNE24XUCp>$n4m2Xw-o{$A~(x=a^Cl_8bf9z@B5kIk4x@IS2L}L*~HVx<=Qrw?3ll
z*jt~{b?mJ#=sNb+2i!XL)^%<jd+S5Cjy<PFo!E1Zs1tk6DRpAcxu8z$IR~5*drqBm
zV$V5bPVBif>cXCDL|xc(O{oiet_5{r&o$s&*mLQe3wy31b79Y|Q8)J7BkIPUdrIBd
zb1$eHd+q_}#-3Z}+}LvunHzh8Mg{B%BPw7|m{I|I!h#Ce69$}sJwfLL><L3AU~faC
z8`#?z(GBcvOz8&pHWqXPdm9681A7}fw}HKlA=|*7N24C>c}CQOJ<pVSu;*D&5B59*
z&VxOV&Uvus88Q#{yc+dl&pV=C?0Kiui#_jxda>sna9-?rb<T@D?~r-1Cu&s0o;ac+
z_QWX_u_rF5h&^$@iP#f$PQ;!#WFq!_8uek%H=;i5`KHu|J>P=*u;&|aKJ58)&WAnU
zkomCZ*Qg(R{t@+K&p)Mp?D-eek3Ii@^JCAibAIgkhs=+?fJOt@3yf$0dx0qpU@x$s
z0qg|^TmX9koeN+uFk}Jj1vMJPUT{Q%*b7c+5PQJ|4Pq}i;DXo->Rb?e!66G`FQm~B
z_Cg~X!d_@fL)Z%~Xb5|u0T;qvNasS>3k_Kadtr@+u@@fEF!sVz8pd9DLBrS!54bS)
z!a5hmUU<mD*o$a1g1yLyMz9x|(g^k<3mU;*WWYtR7ty&0_98<T!Cq9OQS3!WG>X0G
zlt!@^UC=1@q6039y{OJbu@@b(DE49+jbSe~qA~2nrZk4V*n-Be7aMRf?8S60hP~L3
z#jqFGXdHX-5shOnKBaN&#TPV=z4(BOV=u0AaqPv1ERMZ|MibaejA#OTi78EBFR`Er
z>?H<V0(%LaOJFZCWC`pgHJZd;azvBZOHOGLd&vb&VlO%1lGsb?ToQZ9AxmN}rO_1j
zQX`teUTR8H*h?*F3VW#mm%?63=Tg{94Ot3%X^p0_mmbkH_R>?D#$I|s)7VQ7xHR_C
zI+w;?ddSk)%V;!%z08Pau$P(A4E8b$n!#RXz-6$P(YXxvGDDWZo}^I;d(wzX*psGI
z!k)CC685A4Ct*+0ISG5xkV)9fYBY<z?1*Numz~lq_Oc6_#a?#6WwDplxh(dwLzcx}
zPNO;O<wi7zz1)=Mu$NoV9QJYpE{DCG&gHO|8?qesHZ{75z0DEb#NOtVZenk9K{v6t
zIp8+2x2bcR*xMYkP3+}0n#W##MDy6oPiY=|`322mFF)Y&*vso&9((yA%VV#g(E|1g
zBU->-VM+_wD=cUMdxZg4z+OS;3fL<QSpj=n8r{O))`)IlZ)-}ou(!3KTiDwga9h~h
z(zz|{Z4KEL_GFF9*po+8#-2Q-GWO&Jm9ZxeI2n7g&dJ!5hfKy^QKLod6-Ts)z2cM>
zu~%HsBKC>{u86&&&K0p&9I_(zN*XO;uQZ}1?3JdpguT*&matbEa3$=ObgqQG(vX#~
zx2@4_>}`+eHukosbQ^oy3%ZTH?E$xqy=|S_#@_ajZDX&j(K7bRBU;8@c}mOJD=%mn
zd*uOF#$H+H%GfIpSs8m3jaIN%8PN*%DpOj)US&Zm*sBb<3ic{GSHWIo$ST-VG^$`v
z8Bqm$%9JYDQx;Ugo-*JR>?t~@U{4t`1$$MER<Tzd(JJ<;Q(DDdbwR7xs}8s-_NqEp
z#a?yDs@SV(w1&Ofh}N)Io6;KgY71J!UTwhDuvgQ$8un^KR>NLhqjl`nN3@Q;`jpnO
zS6|RN_UZ$!j=j3h)v;F}vO4w}8f{>&F`^CZHKw$Iy~ctzu-6!H4eT{^u7SPAkTtN^
z)MyiX%@J*4uQ{bn>@^p(iM{54Yhtgdb4~0uhpdUcmPT9HYmI0Nd#x#LVXw8IE$p=h
zTnl?GooivQHDoR9wKdwtUVB8_*lSN|8++{qZDX%J;M&-0>s%Xq?ICMpucOfp_Btcl
z!Cq%dJJ{<iXa{?p0oTD^N9Q`&>kL^3d#XlN?5QKFVo#k?6?^J}s@PKpoQgeF=Tz*e
zL#CeUh2H(_(hJXD_b6el>(9ErvG;b-tnh)i9<kj1N96R@(ah1(-p^>sxieRudcU}D
z_vYiL9e>T9-TKjcF5CaF#~v4Ve|~?$mOq34S~OuR$i@jXBCk%^Dza(9Ovq~!wuWq;
zFf+1c!Ys(v39};e6UHOkCd`IxpD;VJW5OKB>l3z)?3^$svTMRz$nFVqBMTEIAa6|A
z2C`?uJjmV&^CF8ACL;SL%!llsFh6o&!UD*_2@4{JCM<*;p0F@-WWplI(Fuzp$0jU>
z9G|c_a$>>~$jJ#yBBv%Sg`A$SG;(IbGRV?|Nyym=%Od9{EQh=~VVlVL3CklFCai$G
zHDO!G@`TCA#R)4SmnN))yggyt$mIzuBUdJ@f~-uKf?S=jDspYYYRL5ot0OlitbyE|
zuqJYA!dl4f32P&FCai<3PMCU%FS~vD?JM{F=bbO#kqyhHQ~xNuVpst(8jL`zhE*Vw
z!34BsSOYQ}%s>`{1;}c!0`UeO$Y!ts*$s9ehrt1~ZdeC$8k|5bgA2%Qa03Yj0cgXp
z0pu}wfV>7TkZ2Hrd<Gwo-{1!d7y>{+Ll7ur2myr+VW5a10u(hwfntUjP}~p)N*EGA
zNkbASWk>;~4QZf^Ap;~CB%rJz3zReDfHn=AKzTzRs9-1nZ5g(JWP=P;G!%hKh7!=W
zVH>DyC<9dt6(Gf+096fDpqiluR5#Rt8ioc?)6fKJ8CpPXLmQ}L=m4n(^%c5?Eq&<E
zZNL7`?%UeDCz$uE@XRQlU?Q^b1oI*LPcT1n-~<aG2T!mda_9sLA%{<}FmmJsiy%i&
zuqblu1dAcZPp~+0;si?|Cr_{>a_R(2A*WBUG;-zy%OFc9n1q}?!LrD?6D)U%Z^%Ty
zxmI>>?w)y^A3J}0?_F+tXCS*jx$^R5k@nsd^X{L;&hvcx_H%aaGhch=k-Z;qdEcwP
zc+Y#j)VpQ(=5yb0?$Ny~AN|3TcU@w6SLMyd-ABDSTR!z(y3<=LbGCwPoHHZx>YS}2
zo94`fyf$ZR$mTgSBU|Rof^3~LD>6T4JhE-hY{>RGvm-m^%z?Z<XY0t$IddYr=FElc
zo-;SHFlPet#++>+d*;l8?42_&vN&fVvTx3O$o@I=BM0UzfE=8&AaZEVLdfAc3nNG7
zEP@=JvnX<G&SJ>%Ig2AF<}87noU<fyYR*!~={ZXyXXY$}EX|pOoSm~Qa&FFYr}*Aa
zow4t?pZ(EYFP~l?J8s|m=Kka}gPST(+;zw9&BrRo8hg*m_ucrpeVg5n#&>T%`_*S3
z*?WiB-qpbT2TooEZ0|0y`@JnMm-c?|?svAl0%VkpK&$d9kV!THt;uUZX4wp6ku5-0
z*$TwVJdjPc0oi3ckVAF=t;_2`PT2|Ml3hS<*$pJf0?>xM0pyWAKwjAkB+4R?Pxb-%
zWj|0r4gdw^AW%pS0fpr-P(+RZMdc_^OpXD?<v376P5>q4Bv49D0j1?MP)5!GNwNf#
zm9s!OIR~^UZvy4zJWxR{0By-zK(Z_Y73CsONiG3x%iBO@xeQd1D?o~@09EBGP))7@
z)#W-+Lv8>y<t9)|ZUMFBHc&_I0I9MHbfx1;=iXiOp_S8$fAp<;p5DC=9(wJe*X`Z+
z{@Pi%bYJYF&+OiO_~OHsy-N#zN%lY8yx{v^-M#rbcHQsneUTqNc}wv9haPxr_vZ5}
z=c{|qZ9lf{d;V3gdHwr#Z$5tb_yv3R-sSvupYmS5mfzm}`x(Q1e@Nf-z{`JQKXUOA
z+n%-e9QXDEC$H+aXYW25h|7snj|R)V<(p5@ins#XN^b==qiBT9*fYYmDz3t|+FOOq
zB${9|^-Qp>iEFT}_10iBi)Pr&Ju_?;(E^*LXMxQsT4A&Htg!JS4;$a(VY7)g*lax;
zY<AHOo4sd;%^^BqbMzdrt&8iht@qYpbBa#boINLOF3|;>tLK8vExKWI_uQ}vq5zxF
z6JXmAH(=Z7ZNTOcJ+OIt9@xC17dCIt3!5m4u!%hpHlOH&&DZn6<`@02`FnoY0%8ER
zKraAWPz=Hr>;+*9i6PiRy%20+F$`O{7lth&MqrEdBCtipC~VPQ6t<WcgDuvJ!4?<e
zu*G|E*b-s_wnQ%hTT)EImh2^AONlAiQoR&xX)z63x|fD6BW7UB^fItXq6C}NlVHn=
zS=h3@ENnS32V1U}gKbmXgl)6830q#w!<O&mVJnCQ*b2P@Y+K?MY+Jo8*kn<LP43CC
z6~!WK#a<D%l30SR)GNWZEpEfM-P?w(ES6y__sXzU#0qScUIjKqRA5tj3T#!e3R|^T
zg{>ymV5{|Nu+_ynZ1r9pwuab%t<h`1))bqtHG56iT4D>fR<8wHTWrJD?zLg-h#lBE
zy$)=usKTc9RM>vk@w=~Fkl^&zRgSA(p}prNdf?=hXHPwa;S1J#*P-40v8R_WLVHSD
zR+rzh_Z4=}des#mqiO_NRab#bstIUKT>~<!W+01d0kW!AAYSExY^n{&uG)bdssm_U
zT?cZiP9T@+0&=TvAVC#?Hq;FukLm&Ps$L*b6@h%J56G|jfdXm(D5wU3LTU&otcHOi
zY6K{%MuB2#3@EO~ff8y0D5)laQfdk)t)_u8Y6eJBC7`UD1<I*8piOlXD6i&$3Tgpp
zOWgvJRT-$L7J*7?320m01}dv%po&@nQd9+~s#bw&Y7MBa)`1#o1E{Gsfm&({sI9hv
zI%)?<Rn=GMH-Gimk7W0ArI%mF+F{=B;_FyDOhoqWFdwpihxw5MJ1l@4++jiF&<+bB
zhj&;QIkLkd$k81ZMUL&T7;=1v#gP*`EP<TdVM*lF4oe}YcUT%Zv%@mT(hif5vpXz{
zoZDf!Q`~#=mAn57!S4S%&{MD7dij&u%ZLB_sPBEUwA6UzQ?7O7=!y6h>G}Kix2ug>
zy}7h|HSN8R=H4euODDhT|DoEw^vdJApJv{Y<=DRcuQHr=@boiY`<hp$4#gtNhYqB|
zXTJWd*S+EF4Ubp!`2#`g;k7qi<ThUL|GV|bx#t}{cH!^5@#2+LlbN^K*ImN-&T}k|
Q=;0%Ybce=sp(K(17t`X``v3p{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_bwd_data_f32_instance.cpp.6CBCB642104CF767.idx b/.cache/clangd/index/device_groupnorm_bwd_data_f32_instance.cpp.6CBCB642104CF767.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fea89b1c7c5f9ceff47a0f485624855d7da4be5e
GIT binary patch
literal 1390
zcmWIYbaP8%Wngel@vO*AElFfyU|<jg;^LB`%t&SihC3V#3>9<EhTJZ?ZNYQ(e@VhI
z-&wh~kK2!IjXrPIxqSM`#cfK8oylESnx@FUzb?LI&5lCj%emhYzq=|Ga6EW$j7Nd_
zLu%&UD+-=hFR!YpC|5JCu>Wm5|9)I-{jBHLyFWhdeDqeha--GG4H0wq+XcqN&C!^*
zE8<?AZ`bYbpI_(8|M>9sv`u{NzUU|6tDlyf>9YC1Bj)L9|L}P6^7;9nR!d*AVXv8U
zwkkn0>;8fj%~y<O^k)1O)12;iE7rhK<Ng-A6Ni%axfNx&UX)zjT6n5VDn8+Vk#{!N
zq%GD}JFL$?h>JQMzwhpX?i#1%zMNrV8O}nTHcS7rWr}A%UZrB=^61&!>GPy!_VgGF
zGToYV{+#o{Nri$x<aQ<Od42WO)z#(sih{SL=VmN$)r~cI#ZY-%vEJaD&Z4#rrh!)%
zY@9OJcE`mVLRylN*H)}ik-VuV<8Z%FaDnRmA3t93$<IhO;oaogab@KTO&)>t<2Oy7
zg|rFGc0OTp`lI@!sjnTZ-k0^hXZljt-(kwPaN?1_7Fz$^K5dwovc#+_@aUo?tGwS@
z2xR$i#8hbTM)9t^<`#9hP2FO-MD3le6&I|}^>|O>ILM!Mf0<Fk$~n)QzU>Ojs<mln
zSZY!hAi%bSx7uiyr{pq6W&!J>Im_GHZXdo9<bM3I9Pd{3clI%H4qqmOm9Brf!~W{q
zoy*Gd-u}6EFl>_d&97V7bxsH^d~r&MbC-g)OoN`mMU!)vUhaG3Y&!o=$Bix1)R%?l
z1+GgJzPxRTlI$lh(`3yD@v>G&m7kowcUqBEF6!3f>Q6NtDO!_q=f;%uO}_cR|L>*h
znE|V;T$98zx!+g2Nq%@Ref86v|8qZ^Z+-P4ZPWBE*#QyjukrA{SvVo<?B$bB)@1t3
znV0&&I3xLN;0@LCbLB@?n3;&iF8IO`6UT8+Fue2Q(HA*K{I$O4&-(SvosIwI>3f?W
z=5j7^xmD@EGk@d#!*_NFsEP-K<f_M18Sjv?4aq58mpWy??6>=DB2wmViE$I(-zs@3
zc=Gh;>-}<Wt8$GGE>=IWn(bcCIrsEM27XtTg=)>dd%{$&BKXjQ_dhe->+f{j=dk@G
zsh{@mJde||Ym-lA8mdjcS7jvm&Q~q+<P>$MKl8i)#RvXzW&ZqSONIHWnM&s!zR614
z==(7gSLP<k19RF_jlc7q*7pD6Vek)N2;^erG5`}m83-s!O)J)6WMGio{$)j#U~oSp
zOCY}<e<)0wWBVJ1pI=ID0fqVW_+ntfp6?HCTzcNUpOFVeIK)p#{l!EZR-iC9CpSM#
zbt==L<Ndb_gh9dz+*UB*J}0{+r}rM01qvfuc__s0WxV|L<4i!aS@@J-s)34F_~AU5
zC?glxJH3Z~FIe-?TH&by>lb!@@otb!qTKQ@ReuG-Hr>rX@{AQIY|U*469)1exSf%B
zuG}6-JTGovB%VKa5R6xpnv)0$z2c(eVsQ{72t<GZV<7Kdt{*QOY&;o*7?N{Sax#Fz
z-f6zfEUawo9Gu)dyaN1uo`SB<YJyyzPD0|sA`+4omew}5cJ>Z>qS{u<W*YL?M8%|K
mWEB*ZR8-Y9wRH3i42@0A<xHe>ja)><+`WRC-25Cp7<~XGOkuA8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_bwd_gamma_beta_f32_instance.cpp.59A23F12AA5E3659.idx b/.cache/clangd/index/device_groupnorm_bwd_gamma_beta_f32_instance.cpp.59A23F12AA5E3659.idx
new file mode 100755
index 0000000000000000000000000000000000000000..031556422218f670ef9d870ab172c98bfb855ca4
GIT binary patch
literal 1394
zcmWIYbaTsMWngel@vO*AElFfyU|<jg;^LB`%s6HS1_@3EhKe~?Lvr(PTkw4QFDYT3
zz5X`S6X)5g+aFvryTSeI;sn)OVoKh2US6~Q{t7JIv*xD3nxhSeYd5;uDgFH5E9t=Y
zVAUy$g$@&PmX&_}@XROWL)kZf!~J`A)!FN>6MuYK_~>i*Pf4Xki5sozD;Mq9V-aXm
zwsFtjDI(kRpI@K<|JmW^{hyP*KFkY@Sv&2rp?24!yG5aE?#0(e-KpK7xh^l_v+;ww
zIoFs^?<v%M)+_E+vG=6)c2}(y(X{7V&o9lke(svNAwoIQq*8Kg;LdHATw2d>sNc8r
zqAKStS=&2v(+l)_9_ruEyU1p@FmxA_vX_Znr%T_Z`sOX%z1vEK<d_OSzteVK>*03m
z6@#4b(e&2~Qao=n)bOvpe(qb=&qqgRU!TzNw%5<(!a^aljLL>NPbb`$v=$3&y#Go;
zIv}DmgD+2{tR;x0P1H--gJs)i_6x3hOeIgE#P{ty%>GQGq&w#!v#X@cPsT@YEM|wz
zKlQ-5nqy{~<sXMl+XP!m%dYQwvG9AfOV2irwF>W!2mWL|dC+Zwn3uPkYIoDC8Mg`*
zu4KL6`5_?Fgez$6R)bl*{sn%9J2t%juweTf$>}PBP4BO)^G;=mjXuu$wsda5tBP-O
zS_{uEQ262@{cVZp<V!xyYzi)$<}7cTcKh&^Aot^s<-YbRzq5~tb10Y)R=WP_4*M%_
z(|vo(a_?R{6gI*4=GQIkXS%r-_Jp>wT<n%&SEy2(bu2RCYIx0xXN#p7Z=d^c<jX1D
zF4G*RThl_8ILEC_)2aMpzG$kAli&NjGC>a(F16jKe!ksdm)b?sldfClocUM(X8Ntu
zEly^dLc2YpEAAVgYGB_VrJMfy{ha+PL)*<K?_Cnxe92;GT1Eo@(h1q~{M6D*Rv-SN
z!#v$lT&nMo-p*n@p~YU8P5S3uQZA^R__^<fm3(~9Ik!JcivBNKnZMqs;>5h~^LX>S
z)n_iBTOP85@3*bamamQljGuRP@AS=Q3t!xRYOPk*!LwRkJdaKu{O@xk=AoOnPFB%I
zZ*l(*i>&H@p1ti+7CY^;!LuhR-1WWZ+|w&p=H7eZcPgh$`Pz;LcRw~?k9W1Jddg-y
z^Vo^hJ==~3pQ%y1KKW$kO0~)Nj;9#ineDTwYx0aIKmRXP{jYoXepb=Md8+B|yqBM5
zp4`X8Hrd^*xH30M9+=y*KR60^L~qgHVek)N2;^erG6NGp83-s!O)J)6WMGio{$)j#
zU~oSpOCX;fUo=dclFz*`ZdH~kj68t?$^zjqVaN723_rh=+ybgb77p<fQhzbgh83uq
zo0FR#rW!^wa)IscJ@k9Qnupd3ec$SqPAi*U15}M{<v!!*A6WDvy+MX3aNEN)9}2O1
z883hRI1^Ac3!f59m?w~*g&)p)YQXx1onO2gq*s(%9;WE8K-i|c`A43y0)-v8ongX2
zo-4Np63>g<7m4T39fZUS;fDLZC^aV$6ne!)$;IM8nlbQqNv6s}&%0ubK@35Va0QXc
zxhXjrK-xRamzjl?jh%y&n}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?9
s9-F9`w2Z8RqLPZLx~7(nzJZ~!skxkql&+DBsF=G~Fq4~~qX(l80OV9kg#Z8m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_fwd_f16_instance.cpp.E5229B139F1A9B82.idx b/.cache/clangd/index/device_groupnorm_fwd_f16_instance.cpp.E5229B139F1A9B82.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3b94e1bd944d1f73a549d860fb20c8b680ccf759
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%tU4ehCLh%3>9<EhTPA;Z6<K-fAAbh
z{gv0HJ~%(RwR)qP;V!XCk&I6&g3CPKdffc@@0XnJ%Dg#ZH!iimbA5F8`0n5D#Elzx
z4v4pY4Qf^iJ$vfW$G5F#9^CzNzNWCe_H*LI`#Q%b8y}sm{;8<8EOF<)KW}{ucUlG8
zly2Phdx}W;dolC*{~sKFzF+L;ug8Cl_Q$PHOAdST^Y>M&`1&3Hw(R*|F+2YLrf1a$
ze*fIV9Jc6_X_uCVZu8B~WA7&xy<B3;BHZ$+^l|f@h%?(g_}1Roax5!x&Zf1;#Bc1s
zb7>Z@r_~&rJMn2B6xKe~zrXIHtKGuO%NVrHGn|DwZI=G$dnus(_>_u`+oNZ;BJ;9l
z_VgGFGTmxAf9`Ob>SLu3wj~EEuCKhhIz0crgHoA^)n$j~s9hN${FTQQ_ZxoGS=6?{
zJTP>@#tCcR-p$D2nA*dW9TGOFCr7;M!uA-Jj1$-G>Pp-D&lt{<x~b%n75YVltz+}!
zoUCF!=Z;c&qq9XdZktMUJ)YlP_${~L+<VTA#}uzj+99w0)ZRm`Y*Nn^H^WsCQ$n__
zl(k)WtK~pz`=JF_CuHR=G0mxHEAF~@+_`!0y~4DZNiH1w0)MMbW%PMzBfjObY0&G-
zmMj`eHa0gID7?Kf`J&G$174|y8<QT}m|nRfw``h~m0Wqo3Y)_JiAAgzoujTUp7z*o
z_3iJoZtYt8xthJ&Q!n-UMY#<M+%3M*A|4mmA|EkqY3y89IL)tIRDRXZCvHb}ZL~3*
zThw2~y?e$sRne8_wXgOpZK&V=LZp9!xx0QXm(7-vBmGZ)3Y?sxGFxkBhOBz=pZUL*
z_8Tn>3GF(fcS+2yq{ryNLH%p1e}3HkAU`Ygf7+($Te3H7jLVgjx^dWTm9D?~<ju>>
z3d<%oo?p-@>Yh0Lj^%x!#kVf_Y-zltY_L;kkMcT^$Ez<M?U*0(?!L~~x8Hf>Hc#K1
ze)zVKN9UH$^NQ{#Y;WILsOY7)Kx1~;jxT8irdBI%y#6oqQ9X&X*Ltym{>kmWPLqxF
zk4;`*rhL_8=7UJ#C!xIS-KQ_+-e~9-vMf+*_We`3dX=V!9=wX|nXIk%Yd^Qer<=2;
z>pfOIQ~zSxlH1SLe?F~#O)>cP%P)JlE@wVZwUT}3yFBvXn&nP^{;&GW-SKxvW@VMt
z9_dh@p6GeCjK`Ifv>1vjbCc|VIj*VoQq6=n+;eys0s<L=cvyKXzyweR47Ozbo81?+
z&6JzTA1KYm#>E9DVA4gYX~laOfq8d*r&Q^>SBF`7f_Tk&t>A_R@!Ig(Bk>%0U66Qg
zyq-urZ(ctb&$0at!_O}zx0qRi1eFDSV8SdxLV7|02%d<t2wcu-Rf632A0N*%@&q9Z
z`<hS67t`&T02Jmj1SSTkcS8Jx)L%@rVFe2Fa`N)Sj4ZtBQk?N(hb>ST*{0%2(u$pP
zjx7MGR^T;*sfN*vT;L$<J@k9Qnupd3FgsWwK?M^A>J8uxf$8B167&@eMDT<Hg+dWL
z5nmCw-2Sy6V)k)L=>RQ94vRw}b}!@QuODaT2@+rtP(tVxVi7{(39|^pc~1>kzp(R*
zcZ0kk$}11kQk0sL2#U|*qU2(6Ak7?PaK2t`)ho_8#$bkESWGi8B<H5&WB|F|X}-)X
ztZeKYoZLLT0{ncQg09YLf?S?XLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_J
hRMj=Lbo31jjZMwvOr&&;Ttvm(y+WAW{2V<PeE{_fq{sjO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_fwd_f32_instance.cpp.D9D1D3F585AA6CD7.idx b/.cache/clangd/index/device_groupnorm_fwd_f32_instance.cpp.D9D1D3F585AA6CD7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8c15a7517138238bbab564c45f0eac6f413a6ef1
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%tU4ehCLh%3>9<EhTPA;Z6<K-fAAd1
z>zUW2J~%(RwR)qP;V!XCk&I6&g3CPKdffc@@0YCZ%Dg#ZH!iimbA5F8`0n5D#Elzx
z4v4pY4Qf^iJ$vfW$Gfd(9{l~&UsG6K`?+x9eVyas>0R6XpFI8b=1`GsW$m&ZyKI7O
zN;mHMJx8Scy|{UN-SflK?Ztooy8PE@f86@C<gh0{+pk*1*YEkaXV3qV+41)`J*z&@
z{&Np=*rHFSU0NQx%{M!by`Nb0a)~XAaLcFC$IW*l&TRMKTYF>6v8=>7o7Nr^zp?+$
zrCGe5R&#9b_#0JBxb|s!{`D71^DkruH%yh@qS)y&?^1nxNypT}R3SO#!q0MC{#$+A
zj?Ls?%srU?Jn`_PLct$$yAt-izVhno>hgR=L2IeGnF}0sV@+1I?|D4ozC^WHp!1E{
zE3_^|oVfNa&%~_7%dO2+OWV_J8@KI?XgwyAC)x7*cO7<rCgIzgBdD}htBQ;H$jw5t
zt(MajkL==~nPzFH9Jxzu$>(<gzs($;e{YF+ERf}Sr+@ONd=-9cPq(ba5}`U?TDPzC
z%?rqNdT_`&@j{lzR<lb|W;TZ{kG^=k@L=3~3u8S`C6@auYA2s!SXMHR`_@aTm0w@V
zFo|5c;piwaq3q4c47Jo5Y`hLRCky9EW##b)d&|o9TbqQ;+ws4kl=Gr<)YZk)9^0+H
z{e9N0U28vAvsZiSrCz@%w?Tos#Wz~S;{sdcBZe)Foy!WR`IU>xulo7K?Z~c;HimPH
z`ir=C&)B9ay7IjC)t;pd^;=(v^iME%*RSQW*-~<(|H)5*lT%b?YwgUCRWJTC|JTxf
zqlF=%T}SjTiP@F(7(F<se{J>8kGmh_XNCSx+cbSk_J)mdxsp;h4!f<=^;e&~d6`*Z
z*~G^43pz#J6Q|#?yf3u))&-v}jhB=QN;NCoV>FLVU)(Kdzw*xh$XB=XC2i95tIxM>
zSDhe~`P}yDzJq!E#gCn)#5e?+tu1(Y_CZ!jz{czUEg$(G5t1r*ON^V8Kf6UOHO?ol
zeyjA#WhYtA&r(>QRbf1<NAJini@=#~D|70!x9@mzm7{rex}o2?9WVatJvfn`9k;7z
zy21a2(UbCK??0*Tzs7mlw#zRoMVDPZce=#t&TPNUO_BbsKmLdQ78m$kc=^+pk_z)x
zGbFa#{AS->rliDBT$!6>2h4GO(J}1Ttn{>b7y<$rf_PYYEWiX%1`M`j{hQqvwat{9
z$sZ`q#m2=2CScM<scFS~7#SEC|E|9i*q`!;ktaw{S<nY&sA7fAqWv0n%0OWeWf8cr
zWBVJ1pI=IDF|(kkW(g9~6AFOoWkJ#6Yd$GoOt)tO&?GKHV5CBw72+qP{$ipHD^Qr1
zlb0W+w|J7YV&|M=3qZmOyk;<A7|qB9_G9m%-wW0}v{rx_#tQKpOc<y)fHwrD2k0?h
z!9WC0C{QRA!4vTnfy*5Vv3nUWfBiT!PmlnMfD%HH5Q`8JPnbm*&b!Hw8+;~C`yJ4|
z$U*Sbfb|PIzj!yu8=}1OFq8fYgl)Q;f8-e}P}rQ;3MLHX+3?yU@f>+wka%vqo=7}z
zUOyPGC^aV$6raUK$;IM8nmK4^l9tYn9luH$gBgNhG0nh`oSTx90pxn8`7*Pxvaxe;
za`W&C@bh^Jx;m=~a(Oxli3^KJNLpB0+t}LKJLrjOTPd4q$YT=~la`THP*hS;RoB$g
d(Kj$OHZ_+sk<vAC5fyXy3Sn~dbM#>J0RV<Aatr_f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp.4F6E8B2271D69F1C.idx b/.cache/clangd/index/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp.4F6E8B2271D69F1C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c72d19680bb8e560af4a2e3780c800e20f6c93cb
GIT binary patch
literal 1592
zcmWIYbaOLcV_<Mj@vO*AElFfyU|<jg;^LB`%p7J023Af6hKe~?L!$GgtOf4<59Z}l
zU8*a@&)FJ$Ma^-qM*p;pdv}Y)Ra<XU{BnK1;b}RkORrum;lEm$@bS#^nLnRBV_LxQ
zAvJSvfI`5v%d2WCsuj5^%zqzvzCHZ9{PRCmhBEUt@7YWeU%i>T`|69QGm~m6msPw;
zs{F00op=B2`u+c0+t<&3Ui0;3-of=@r(aGGnX^BB-H%s~uTDRDce!+3sq?J)jPKJo
z^F?gBWj^bwqlVy_X<O@-PvlmAV>jc#lKd$7Be$~lU3{@6^>xacq+_15r_{Cm{j9q-
za`Az`ClB9U(NLYJTl=zXaqN=GWv8Yj9F4j?!PTYzXZ`9z`RvE5bZlIfJi9wxPI_W*
zkFj9WHr2Uv+f6->xh|-_#rimQ`855wJ+;g;ZybE#!@M@^kgIb4gmRHj?8*L%!UW3H
zPDxGCbKT@_?zMfz(he4F-L(#aC-r0;=da+nuqeOg&+)l+3Vjz%X6ZN`QCWQag7XyF
z)XOu9C7p8m6hG@6S69)V?qKo$TdzN()qeFOQp*<#di=G}+@JkrgJ?#J(WC&+MF+H=
zmdS9Ms<CL<OmT~5-FeL|!nrNoVyQ&!o$bF?luwbIsv_JZe`Vj<REA(nf7Z7pYXe@D
z@G*KN=B#&+HOV_UInd>l0PmZ|85aWQaY<$I2cDLd<+C=qBDeCt!M>IkitArFdKb>W
z@;2SqH}CPEa|itvoW5DQrCO&W=&;q)rEVUMA-mXv6c$y^X$_7&`a8>Rmtg7Ts97`o
zc3w_mzCB~Rs`$$p;p-$l9qYYsc}gkf^TqXPUbygx(SE{mXP4lG8L!_cS$aMDQ2%PW
z)rl3YYl>BhCT{<*FFAyPw{B<DIllXrwZ*IZ^3^I&?q<t={j_btJli#`cP^9&S!PQK
zFBfP1!E`GkSft2v?q^QUP`2edtY`WU{FrpHZ$}reKD#(`RnW@%%XhY4XSESIUw!`9
zcGU$bmCtR9?!U_8-&v^0rKPa?){&Sx;}ue}A@8o#_x+IGIO9^Zvg!H@%gq)?o?O|u
z@Y&OXGp8?}e&_!D-erevxd$_TNmSiDu=-#|!NF_CHs0D2B(_<4<3;YMX&0>Pg}y%N
zy7eK3i#fjL+;py6A^%Q$XxwSr|7dZ2p#;~?nLg*b&1aU>sZaiQFC}Qolv^zu|2t3q
z$8a_N`>f~YCw4pf9?1ym+Rw;+)9Pw*Wp0ukFz==O{JqPesqQ!rLqH%y5DzPl7nlIb
zfWcwQ!uBl>cDQge`2(f7*todC1WdXpHLdt4BLf46!+QbO<&R%7@&t(ri+aEe6<psb
zRl4rgVJ4m+UUOb6m@ts%#_I{=i8^{pWv!ap0@N$4EF1_ER(F$o7gzoGJW!a+h|3)2
z5Xbg63_rh=++t=45)>AM+r)yRmj%TYA$~&YFDBZs0!`xO<mHE10HYbXz&`Fh^n1aY
zht>)OZ_A3`b5A%4RE=yw@g!-*&N;^xfDBRKHG^qB6k_)>UjF)VW}YAc76B!gFi(&W
zix3h|ghd3-gZYG&3mlX%dx7Q#@P;A`5cCxEL-2%ygu@X$QBP60+*1SAFYNr{-5`gF
z^2);${S^q?bT|LVGoZtGZFucr!a$xQuL}~-n-^{pLs4o@A}D5yi;|1Qfi!cF?}km=
ztoO1#V+>{phQ&DpLvn6PP6m+co#xBT!pg?Z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvE
zX>DU`XYZgVs%@ogrXi0_R7_e%RzXooMO9r>OGn?p(Ad;m&O}Pr$VF7l-7AF2&Ck(;
G(FXv`bE%vF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_fwd_swish_f16_instance.cpp.3151733ECD55CFAB.idx b/.cache/clangd/index/device_groupnorm_fwd_swish_f16_instance.cpp.3151733ECD55CFAB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..47f64cc623aec1072675293f099cff4a9de28dd4
GIT binary patch
literal 1574
zcmWIYbaRtqV_<Mj@vO*AElFfyU|<jg;^LB`%w%Q;hN~P53>9<EhIkj>HWPUE|Kys4
zt9`RF+}NWhO8F*7Pd@&mqmg045e}9ctO}8Tzs7c?h@3vB_-^6e2{-<3xW6xuL4;wS
z&RTIUmQz(>Q~TfRyD-W>tN(QL?$@)+S@%cwsi$|9`#*X5>&>B}+Ml;)CKZ(~tMJ;m
z>-Q|ta(OZH`~Mak-mcHR?_bUSC)ZzJ&GF$nXTR=R&EJ>5UVr-cS$F?F)wA9Qj{n@l
z9Jc0LZ(6B<7t7}4wfD219OA6zP&0Tr`x^VkjkEGit#WUwNv@tzcu8$(eB%Eq?`$)T
z4+Tj(_Uk-wjPQ=%bvHx&<&wEZsu3G6nxzO{O#b)YL@p^eW@_v4jw^fGFC{B)npmjj
zbYa_)&7U_Yp4`Dv;=em#&+99n&d&b+UV**L#Okuc>Zt`6Psn}hi}{!E#`6l-p?fZo
z&6k#XY+D|<%T?D^B_cDEwacSyA$#p2C5P^PA3s{pty?0oIMXDurNh&$Ke0i>`qVPR
z=SD3zzUY0{IiBtqrq@_-PgZI^Q^_9nBT~y33VQst(ENYxMWXrzj#QO}lUy1@PrZG?
zoaMt2QlZhijjL#F?}j#>^95dpzbZ0H*OyO`oT|dkV!xtx_GyM?Mf13By^LCT<)sYM
zltVWh8?7?#2u=<ZnX=HJfFa4ry;3V<g{?@`yB8IA7B2ke@wfe?s0)8(Xn)kF-)r*p
zO-*;M{anpn-Kdv-z123saf+1LN*z-tu0@xaCor+jIv#0pHMC~Mv&qtd=4+RBzMN96
zyKO^vWpvTRUwewqdT{+(zhv$b!B5XlTw3nn-0l8H!bVnU7ME*ude;+;nUDWp(J!46
z$Q_yP?=o@whlSlM7^MCb?VTmRpZnLdu&?=Q8=u5B_kJz%z93+m({krhNl?Kx6SZ@C
zTpw7sY*f>C@tpgalQWbpU5E8d|A8l67yEW}^{#glSFQ?LS-<(t_WRrwI&t6U@#cFw
zocwgI`pLelcjby6JBfrigl+BI`6n&F)M~|@)$!kc=+_u7EqBX`Tj-zbws}%8v-8>0
zAD`_0>ai{NV8$bfs+$LHUwrzaDYqwiYvxj&XWNrjPv4@Ds8I9lrHkv*|DxNK<7-Yj
zI7*+aKh<|A_h)q7PbN=|CuLT)`e#aR75Xkey!Yi64b4`~<p0;%{;)s%yQkk~zvjK(
zEUunq3x2Tcis)QrD6Y&+vIFM1!<L2ZTORCi;b!s=WC-G7<KhMrKp8MNn|Z}&T4YcT
z4?_S@nunFg4otwLi&E2y_b>u;@A^)u(si#6v+@M-n)6!0bp`R-@Y*Bs9C=-kcy7F&
zNIY*|KNwHlP3~P>_2cu5JV9JWT;?!0I<~)I`1z&e7Bfqbpt7J3Of^f8ke*Ngf+wOZ
z0+(}Il_2;1$48J!$ig9hLh3Ij+OPtJc{zFcVR{R%x)f);*kKD4MmDl|lC)yyoMQ_>
zsug(6V5;}8{SdQ{Q%VOYjO?64A$Bk0<*y%S<_QvD5m17u1{%pCgv1kO5r*?%9%SVL
z2Y&CN-wW0}v{nFG7r+~eFhI~(FbKgD3KR-M@I-t?;BqigMlNt<!Hj!q!1{%qU%VUS
z4N+csn5v@GoJ3H378fNKivwxqAoF`->kh46P{kO`5DbfH28QI^l$;D8*E`LZnT3^&
zor9B`hgX1~&r{IVSxu13(@98NSVTh7!qVEt*3RBRPgL7V*-S$oo2ZzyjI4s9l8UOj
frk0MrfuXUfxtxiVu91tVn7da9lbfHT2cr)FPJ)jf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_groupnorm_fwd_swish_f32_instance.cpp.7C71C8C8383AE2D3.idx b/.cache/clangd/index/device_groupnorm_fwd_swish_f32_instance.cpp.7C71C8C8383AE2D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2c235eb5142fbc046f2bd5a4ea379821e66c2c75
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%p_(8hN~P53>9<EhWHj=wh{RDKhZ49
zLvzARnOhSao-NI7>wg)z*kjY$PMx2zo}P!-r;8qMS@zVy^1#t+?-tj7i)M9Td~o_w
z)k22_Yt%z4|9|rl`H=TZe*WE<xY{`%SO5I*$mQd+Mw|F~F*Eet_5DS}pR0GZi*(P{
zk1YJ0SN8R*t?a(Kc^{uHziqicMo-N-sNi4X#wVxyr^i3rK7ZZMr>2?nnCyI(%Wg>6
zns2hYIY4O!Z}#6XWpTe7u?CHv_qW(}OubsO_G0Ggu#*u-dnTJ{{^j}oJUVK#Q{!&s
z!*5qKSSRY%zARfDyJT|NsVNCZy>3r%b?N`P|Kg7Bt%adt@05c+zw_$vJ>hn2CWqtg
zN$1WT?oxTIbiwv*)5mqer^WT}?Q4{|aqoo-^V+aOuFCuq%Ab5vPxfCFCa_KIl++YG
z*G=qFOQWv@xin3k7Pg=xMVxPee+Wy)MRU7<3od^bNjB+K>lQFP`NHsvK&tyqlV>4q
z0<)V>+noLwIAN-{!kqHm$Cfk5z8BnhEHi^8$iDAXzLaeai`2p%m5@mq8@zVPwl_!#
zGl|N1EneFcldZf)@v!impdPz-QMDJWPxW|BVs4s$W#8FUhG0v7*0&{V1720|F?uED
ztap$#$?G_&!F4J?;s&$f!o@bCCRgORbjwRD-dqT%Uh?noQA-#8%FypopMI~%^EWl!
zx%G22dvv2-`t{bk$&OQ`!dB{-I&m$!WIKV0b=L7ni>sk~)<0{O4m4l8wDaYZXx(cY
zIxC}#Ch}E2oi$P9%YRR6Pvs|P=RKXk#H;)6W9^5Mjtni4+_{>Q7N>mspS=G{oSSaS
zTw6icdxv)GEn>>}aQbP|uK(v=%#JRtFz!@+FSTaP^z5czFA`Os%vv_NY4b9}!lJMS
z{RO?Im5N@k=J<Q8FtSM0OsHegJ16<JWA%*>H3eo1)TjP3KUrHHd!Y8?p6_ewqnW&>
zmETkTzV_X>3DV+=7BFpE<$LC-dGo4^j!(n>ZaZ%O;pmim5-TfP?Y2su_44E8F$E^Z
z`Uw7%+hRRxeG8}a?vd(YOV3G_S?#NxYPeS{+jcoCr|m;!e|IgdzwdX*{&-Z(#&yf%
zpLdx^-r4w%7ll<8PI&$C#}<Rfg~v_hzJEXVc+nD(V?BRY^Zwv|_;**o&3?^$yjfg5
z%NG3LQ<Tv;TU?o&WCzS~rkYtrJId#M<6#I0WC-G6<*@@3Kp8MNY+2a8<-raYZYF=A
zG#48eH<*A)7p0~Z?_p$MVEnuOPGEn^A4Z-aL1jT7n4yXlI*aye*eL^rMU+M0!s>2v
z@8YT-p9cza8F87z?11TF<N`aP_t5VJYaUuFIJUoG`1z&e7BdTq*(^aqdO`s(JuD~|
zg!l=mznEyl3bcTilb0W+sCbgJV&|M=3qZmOyk;=rLm_rA<K?d(XXXhKU=dJ)3G)OA
zu?QjYgjs~)JeZSNxxm2zb1Kl>0Nzl90fN4QK?t5upimfsC*mstm%GW38+;~C`yJ3>
z$ewy?!1{%qU%VUSUQu3onBKnvVVmyeA9=<K6gKCzf(Zk8HoW#oJV#y^B%T|uClb$_
z*AK=kO3g_G#b<F*a<Mp&W)3<D%=w}dJm)h8GX%q8nt>rXHzg+n$n{S1WoBVzW9Q)H
z=HV6K=kpYFbygGP@^lgs7Z#C_w6L_cv9+^z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?
cZ(wL_YA$CYrEBCOD(3DL!sO=X=)vd%08e~?&;S4c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_gndhwc_3d_instance.cpp.B02146942667831D.idx b/.cache/clangd/index/device_image_to_column_gndhwc_3d_instance.cpp.B02146942667831D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e841ec95cfbe787ec174e24215c71731d5f473ac
GIT binary patch
literal 2168
zcmZux3s4hR6kQ+@f}6wyF$4%8gujFcQ3()ffB;GOMp#7BVn75*Q-MH15alBxgQEP@
zsT6CWND5LEhIFV{6(}M_M6@a`RtuKk2h^fA(n2?Mo}n{+Gka&xyYHTJ@4I(b#19Ba
zHGm+$9sF!rVul0*K@bUj<QeI^_Ul2=0uF+*YrCUl@-dH9Zx*N58QS)L_2Y}RXHpYe
z`6oZ}QmZO#@Bd^ayi@}}+Lsr7@8E%s(Us4~DfM|4&*xi%ATo4f>W4I(Igjn~sy~b3
zUGTQuXH-4@dZ>R-PP<8YSblTTtN6@x$$8Umj?ND!>@UlW2EQ97eVgm^z510t?b2Tx
z!lF;lP^Mp}uxaFDjh4roGhV-Z_s4Gwqq3Yl;WgzUVtb5mUCgvopJ3SbZB9>$GU{ad
z;Lr`}@PX!(sXaR*yO87>Nlg8b9b<;0L74{^)Sm?;T}#XANc9msEHLwn>-x~qOx4hG
zn{86(I8lE{#a#Gg^^C7VMk?;lzkRXf7fD}zWv6F;wB6bE%wrV|GN|ip%iSBDZP|bP
zK5pE5`HjHd;I<&$tV89HH1OCcG4RXEhf2>w<c8~e5<J~ScbYInf1xcLXc07=)gWy$
z2(NDzdz{>Ibaleb8f&~|8|liKS;y<;7Zh^}cyr|EgC2vAc*lxm`LRh^HB#@6PmdlR
zLAZSt@6s@f5vuJy5po6ge4&3i&*<k7Y~-$4#fxLoM0)Qw=_f45x);^)sgmesg0h-g
zLtlXj+aBHTmd|LZFYMep^p$sAcS)^uRIlmcRea2?%EDW#SG@60zA36BJEahd*LZoF
zcec3FM2nKWxGP-o9gbjM-0kPFA+l#<UFV8j0?pc`Ju~cKIX0$h%J9s~i{YG+Y(rML
zyYj-h?6{$%O@&{l_!*5B#`fK9!==T>S9D#B9}0T)WHum%Yco0bKyX4A?=qv#dy!$#
zv~n_&8p1kWd4JJj-uA7V6XUw7G`QDgZ?cUbe}zw+OVa#o?WOZk<EQw@Z>`*&wOftu
z1QMwf>1UyaH&lX>X`U{lz_WU?I9N0#F6tE=+458N;iH%TtQ;x7SBQ7jr|9K~6#GZ-
zDStJ}4!)5){7%_dzddb=*4T<Xw`d%DcUh6&(qz`?oPwB&ooaVE&8O^7%8eJvI>M)M
zwQG>$u@&W$++3?Co)Xq>`17WR$}7pKSKMSV3hCO)`z~?K*mEU|bpu6fo#vnZLJ<3&
zI=K*TKUgwR^yOlf=Ty_-2bU0D9A{+J(}BqjeidJu;5I!QV#g1CT+4h@wSQM|3-RXL
zqzS&M>wPZ2d(o%9w952!Psew@Wo1D>EB1Mazqh)zm{Xc8sr-i9dfFF~?~%ptMZZUE
z-i5*HvW8L^ir4}nHdvcYO9u5UE#tyifAuK$CJcq6ZWFA@rX>U1N(?^W!9L4g7!sjw
z29~jL$pH8DW(j@ZAfkq$Fw~95;ysoOaHl8kl%GZa1v<UHVfk&(I(1-#%ph+8Q6M!P
z!=;ptT%G@-<&l>a>6=p1%+@D8x-dfJQG?!Rh)@O85P-<4_-7nzbtMKy2v!6(phQEg
zsBCZuM(9>_Hjsf4rU`RBs0f@v#Whs}aXV~hwYiiQiKc3=Z*L4L;g&pu)N>hQ7|eev
zz#udRO$4Yghy%kx1mrLXi@}Nja{g6ws)%%>2L@=q9D<)VUvf10LoK3msD6MBMra64
ztVMJV-4Bq%hy&sv)*|KxrWcTc5mpo{79gQ6!`Y5;;F{*5A_yNXlM16QCl8}9rwd_j
z0|XHR#EBy34+Q9qEH5uIlx(l6@WyI*aRlNrUL2LUENL8_s7=!q=Sl>;i6PwXg7b~>
zXq%ZFrXScw5sxVZXt%I4x@}%ufQq~8yT5No(RR~-=JoEMqOk-Xfv*)O2nk^TotoN6
z3x#F|&{&c_>3uAqMjNl0XVCbok(;OEEbsQC^^h@SYe0fF0YkIa*5iP2u-4YY!mz9X
z-_?mDeG|?{51?}9`sVND01+yu`96T$V<vV^pGYawT%R$)5}ZQk*O+Rlm7^QeEwyq?
zJd*<Cp0aa%dbhn~qH@dU?C({`h4J0JzoR!p=h1_-v2+1F1fajG=W@T@I#YzkE}vg0
zJ@I3SX5!`P3340FOO(d^yCT60q=%GI9?Ej89q;KzMM4R(9Umv7KH<&~Jq&h*z5#9}
zj9*10@JT#BJCY&amrOBQZ9?^6dwFv<`+UGKcJ{Q{=wSJOyKBtMX;y2k>9+QcPU~Em
b>)khOvS69IxcM8e2^2)@2842hbb|i@i2_+c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_gnhwc_2d_instance.cpp.BAF41643A25FDB03.idx b/.cache/clangd/index/device_image_to_column_gnhwc_2d_instance.cpp.BAF41643A25FDB03.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88a61b4431e18a0f1dc7969c5d6954c718a2f7c0
GIT binary patch
literal 2168
zcmZvc2~ZPf6vsCqkt-x55HUbFl4v+W5+s5g0SOpL!Yx^pLj?g7k=jB;@d8UIR!mhu
ztW&H|1V!W!1Upo;A{4aNN}0+Otw<3?j#v;uv_Ln^e1y()XZD-?zqjxA{`<bSi4X~e
zpXx!7U!rJ#YLX%z13?fGxziPC@+=()n$d@#{Uxo^)bzgf#?uS0Sq7r7B<BYF1OxYD
z3%RWsli_qpN=y!GmO&}__%1G9M_+!)Og)s}?R|3ka8x2h$#ygtRH6rx*4gKG&g`V+
z5PJG&$J8HQ_MGT{kXKHBK9cLT%_{C|4(H*??Cv4r-@mp0rIw)8-VZ*xH80Go%X;vB
z^6gy7HjY`LEvh6dmNh%s^X}~I+^d))hb3Jz$4>Uh16Z4fb|p%U#hBJ))sxa!Z3b^m
z72ALNU1dmf>xvq7GB3W-XIELnt7$Wt3hm{2%(VGXSx>hm{%v!jjPqx0x6q?Z+5*wj
zc(z|Y^Sp4Rclbg=Md~?a=WtafA@^-lptx^D4;AdYbJT_@dZMQwq&|9>QGaMgJx9Ly
zW3xCwk9S$Aik>cNd@*^hfleeVRbMrrb8}_EdxY_W;l~573*<_((^G=bmg3`(YF|_S
z!Swfm&cZmlTsa&=IZyCB-cY`K*I}05vAu8B{*rcxoG6bme*DL!z9y8%54u0e^1}>6
z=hkOumF=rvxz(^@^|b>3fuCKQ$_bYr+lCF4J(8JTj82K?*;`E%B`7AU?pWO@qMTu1
zF~4r_y|^};T~q#f#%6kvM~PpyK-Q?^wPGrz;UXDZVUN2_PZ)I?sw%CV4wF@t-;MA5
z;(q3c&k44zM__f!#(Qt7TMpM%)x>SNajl;Cta7ZSeqXFHR`6^k_JAzM?z4dFN72E9
ztg9mqx2j*}TM$Qodep9XzM{k6v2&`#A>u;R6^aP8u)AAqZhPql{J~Tvn=9^L5t=u=
z_w~k2t4(Su5}R-59jU4r3-5Bsq?yPKty_+*Exp4j&VN1aS-xkNbRoIm%=MI07R)IB
z@$d3tyU4*z>`q0^;0u<{Szi;!Nx_<`^xAjr4&%(LM&5jMxkTuhAaT8ESz<ISs+bRs
z=s)tyb7#+S{*zXxJCfpzXPIjAvFTGO7T=Yk8ye*2YOA(*d9csRQ$8kVF%Kpf^hVQs
zKF^d_t6sd?Rp)goxX{_$?dAEa1847aJU=ni*C2L_NbZTb*5;lX@cwj1tL>E!-`vE<
z3J-Q=RTh}t?cU1y$LWj-kNvR!a#`KC_Fcd4KlJq5iSJ#Ut=_0rom+g*Sh`FT79M;Z
zJ6>iT+R@0`+L&{^Z$3*{n9SoVZ{8L-j`GJ>&ns%@a|0I!-<3)4-9mMnlw?8anW+hf
zknb@*oB3;@o^3Y{3f%<3Mpz@a#RZ8hUKv+1GR-Vh<v1t|39rJga$Q_NIPZ`9(x-um
zQ#dFB3A3^6b&CrK9}6PgZc0kqfP=!3Fda*GU0gsoEh#zu67pYwY8QlN*N-)$a4?BY
z+5{{S>%KJKTRU<Wje{xvlz=5GFhxWO0*W-8Fvh3N7GQ8N-Ujaua*-4piaU4+2QzFK
z?n`c9?kcV`u)YS{Tx6L-jYNtwb(v(~%{Olhnw?VgVKD#kgn_ASY6Re7U^}*51mI#|
z9-9{qxEoE!iDf1_2lNnQe7rzwj1)zR)hY^~A^<oXOogd2TE*Zq1OSeM?O?kYt>UiZ
zdVo?mm?z~$10_MT9oMjxU7C)B;p3K-3ZX2AhcK4og1B1%Ap9O-FvoF&80pCJ*NY4x
zF^5lkVKrQRJYgAEpF&t>TAxAC+T`eS2!IPRPzH-5rfU0GdUe3lCMnpr_2{+A1mu{x
zUR(iq8d3gSF;IIX%~ACqV~--9Ty$NQP6?Dnkeb;9$eiuw!fI!i3<Qby#|MJKh!W#N
zfC6YFS2MUE7kEduH3JK_zj@_K4T0465CcSA;*x>3NblF%tb03=rjRfs8ek(QiJ{Vf
zJEV^oI~o8Zr-k9sK=rmkRmY&?NhQLy)U{mV4so$lx&*U)O|!{(Yb}>TrdVsa3^K!7
z%jFtz%>Z}Q{eVw<^aK~-E}zOdW4PhZJ?$@%&M^EL0oq)Kh!F(TT+!6Q?>DP+klf|d
z4y7gSj@Nu2>1jLC>6#x>&9Zpuz)zd#7;59h4zx59iUivbv?Dcfw+xYz<X{~Pc7?8<
zz9G)Yn1B}%H~Ot18i)iWGZWKQl=beOUVLw#4Q#Tb2i?`q`u}!UT3AwTY-x<uOnV0w
chwJRJ&drKv?zGmQyfQ!>g%*bR2BL!g1xQLq=Kufz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_gnwc_1d_instance.cpp.3EF12405C93313FD.idx b/.cache/clangd/index/device_image_to_column_gnwc_1d_instance.cpp.3EF12405C93313FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7c4189bc59a5da66cd83b477faf1f5db6767ff46
GIT binary patch
literal 2166
zcmZux4OCL+8^0n%*cAlB#6%EC5D7|@<p&mGC^G!^`jKTiE!3Dx$YnOC*;a09UpC#U
zWs!NDSyrZ7X`V=D=~il6ZkDFAoR%M(V|7f;T|MUt|8xHLocqG_KJWAU{hsH2o*O2Y
z$v#{T0MaD+UUjlI0Sy2E5q>q=jFcR60Qd_B0DDh2MXEKu8?0W>kMb;(#fq;7rM`ph
zaVOR_Wln`M6=_@YdDAS4e(vq%+Ps5xj}*PP1hN_L-Zxv1nI!?2Ifd312TOmwz;oz7
zQvmh{&p*HZr25qnVUCzq9F*I^p4y+4k#O7l(`-~BKPb1-$)@3~^+z)~nR71)p8KcM
zek?sZF+7|4Lm7x0n7x~Lzm#^x=i#JJ-(S_MeQ$5xUl!z*c!wLstUeIJ7;#T~x=g$Z
zDpPmn$)n!mAAU-YKi8m&?oZW~)pyrf#Gjkrr_^Z#v#(Wy6W;mR3E(HAnMI;*TunE|
z_LvSQH}p>%ZpoU9PSz)r9L`q`_pN%+!*EWiW}@9NXKM0Kv((H^ah=5L$(^a6Pkmf^
z6!paU4860QBuf1~{K4@K|Mp_P2rH3iyTwYI^X^TQ&x-t=8C@0?#BF}vE5xI*T$Ee?
zuY8A}2F+V-_Y_?qo=;EK`iwcZ7hkLTw+eK8+|=RL@psJ83biD9XTyy&b8Y64!Y-S0
zHs8M8jh>fGZeQDI6R1dBQ=l(Zr~1=EYNX|Q>A5aWz>UXxRhajnbL+?-^>-W^N+>5;
z81#|tJr`Va`8B!&nbFJ=&u^u<zN+h3?`4LxYZu6v3TLpHo%m|)i>mUAuR~N-y1NN&
zpSNed+*rt`dj?eBl-{4be)E&+s+z4a4OeT4{TIh>*1n0e!ua;vVD_o<IlF`F@=(el
z-ldn^+gFB9tR}wtv9m=xu&mYM@p`p_8+JD8JBr+F{==?7JIBh8p|>kls=4B4nBYSr
zdqy{hud%MFNXl$(%c(ju7J84H#js9UF?s8-Yxy0)S6_}!c<ItpBj<M<KUtS{bTvE5
zZ(MgM?jBjm#_ZJA3_a(WmrAT%rhJ{MG?)Kv;f}K}S^9`kx)n07M1`o)_O#`QyyBfQ
z>{-sk0mtmV?D|i)94O_DH~#a~{kL()9+95}la&0(b!+(bz1ld>M~`2(zS}ll;2f7$
zFpD0p=nL?P!1m?s`lhg}ZvAAEWwYf2>`q?KSCKc5xt{7jOn#8!p??E9fT3OSe`p6>
zdwt)T7d6@z<5LHpsT$)3XU98M8eZsTGTa@`igtKU@Cw^zvWIp<A;h-Ax#!PpPgI*5
z=Cbydv}iIzLfhsas=Boo*+<_*sbZQci&|*KzX%Q+nz!1=9v^dYYyG-hlci4F4^Ja*
z!<7&_#}iLMAS@h!_!z$1;sSdXpR4iun9sJkzYhYzu-g)2DOy|*_ubQ9Xwx?h1cE>)
z>}Fz^?u!fJZbx^4+?4N1K_CQn+hT0p7Z=2xk-S4w34aO9T6}|ZYsc!%K#0UAg(IAZ
zG&74uNk#gZ@=JqWbmEl)yFHg*-bH~BMM{w^aX^#+iUP4nBZ*^T#`JMC2;u2?4+Ihp
zp;J7NLm<SWvpkk^fCO}b2SVzEsV}ilqej4**jP3hkt{hh222~Yy=e6RNJT?bJ~a%%
zMME4uCk(+wLqfhV6v0)X_%2TTgZCDM1>@sPgefVp2mmZh$%;h)Ku8cS*p2{%-X{#@
zyKD-Cg_*KM1d-Sy$GkBMxHvq)1Q$mknBd}A1QT4GfPmmeU%SyME(@Go4g-ks5+eX9
zinPUOQN$F95snH`O+|}Y5+gVV;+Uo&7Dx~(5E4cTV-PD@DQU{PdNC2cPk~e*N48;0
zAy_I~x}rX+b|x+l&e|30x^x}++wEl4>xG$cEMAKDH_F7z@k+!x^{0p$1WeqAV~JSe
zQY_+z+iRh(k@nI@_>9KWdfMPZlF%duatef~Xez^)IS0*Q7;_b(g$zWveyFH*$mPgE
zn9d$+zeI-|hUpgO1ffF^;q8U_K(=43!x*rHRbMV#CmByOs-Tc5G$aOY6Ecg20K)|%
z3uuUnS04K|w#5Fu4pw2>esfmPwZ8P0VYpx{DNAOIWd*Plh&5MY$o@WhA|H-5ZEGMS
zc~`>1AZRi^&@ka&b4bifpW?P|A)q27fJkIU03gEn_XBm(E){Glb|}r!m}S`IxD}wK
z6#*|NZk9R`E#$tWRn{wQDH}Yzyv05nH}T0Xo=g#k_WyQmR@+nQjttfswlkL}Shrs2
Y{+@%I-C9>avW<UW6iOB(2{3#AKb_PzZ2$lO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_impl.hpp.C46D2CA10F5E5642.idx b/.cache/clangd/index/device_image_to_column_impl.hpp.C46D2CA10F5E5642.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f4eb285b7cf157325f333da512b56b5591a9118
GIT binary patch
literal 10726
zcmc&)d013O5})cB5yr=G9|OZRpu@=HJ{1*$TnZ?7YrGd-6N9dz1aFLZa4C}YLNP1x
z07dbR#v4(H1`M(yB8ng)8jq+%v(f1GFjKyc|Lh<6_LJ}P)%#7?>+b66s;;Vb$biVm
z+iH=hf8v1Dq$zVINkk%%3H_HmXZB~K#Ujx~P9#d*a%#k<Gi$xfU;3UYGn!^pr2OKp
zP14u@irCSxJY&Sz<-BLFxRTySrv%%IpEdPLY4CHK5<RW;T*ixn*U$bQ+<NEjnp?w)
z6sNrhUx_Ttw|UZKLrB+-w`)%QdxCBJNz%JG6;yA=T<@3pGFUm~^N^OUSBx{tDyBEz
zY?wLlo5Rn>4=8Zo*|K0+NYb;~<oyYj&;IVdc%gX7>D4}N?!PQ_oqqq^*0i1r9~4$+
z)urUm`oX<-s?nzFFYlZ5svq()ZN{EsSt(hM?*#s=_^D~&*#1d#r?;mX-WpTeF!|dp
zs|#Mt%WvrjBW5p_4Vu-_p8Mo#LBWi39d~k{7_~32d+wTJ*jMS9lrZvNeKORKPi-|Y
z*wvWp<m`ORTwY%HV@Leu?}vPSDcWahBG>Pq{eB%$^i}rW%F^^@a@SptgYVSM9)8w)
zRF}M1`?WuKY<je|J!5{{p?0@ZQ3t}laEaS6)5PL*cGr$CLZc>!Hr@Fx?Qg$X%$@tF
zcTcPM;dAySE#KV%eb0Me{VmnZ-k~FH@x@mvg~P1y+P#wkCTGtyI3y<ycdNWy4t760
zS>kcNH0G2jrpJsiw%_huc{1r&^@Ou^Ym(i~O&up2{^zH47J60%1qbhTza8}C#}iX$
zF7f@mzM)6T>X^14PK;T*COffP>nCIVcN&gL%oy=me#P>^wP#knJG*Adcb^_<ZtBb1
znkw3*d-SIEnZNOhUFyOE-JTpWy<GP&uWEVu&I4mB+i%@Uw;K|?(xTnJ+nkB^8IOr-
z|IH%(&+=S$>~mDxT{&*HA){vGv?_}Y*9RSQ*}iRD<2PBy>`b@$jJ}_n*zNF~t=Gm!
zH<y}ib69_|@yPYaA6-g^kGP$XcJKSvnEXjk1Dui;a2E#VMz1}QQ=jLdjxF#z8DrDa
zPnt6K(zeeBZ(DmGrdFPDn~~hcC8qAqdYW)zrAxlm&4v~BWfgYy3u=hb`r-GMUpun&
z<b%-q$@}h0TCQ0P`nF9v<&(PpQj3b(Uu=E*Tv>2qTFj>Gki>Vs<Cm^nINRjy!pTEt
zMJyfu=FE(C>w}}-djEPfCtn`+*l_*fLo2FZ>{&Rs>Vk{SZj0Ch*^_qv+%3V?^6DCu
z*PeQ%W9l^@gJcKSmJw+d5jT{pbFQpEv+evZM}D_9`oS>8WbCuJcPl2vF4QymPZ#Bo
zQ;S8(^O7bH=|WBIM$b_r-Yu5MNLM=%UjsHZpwJPKK5eFd+>yB!pDF+0W=xC~B0iTG
z<PwFByl43X?DP(`4>BcYlqFmeZrl05W~-K**HU}8$=!&IHm?HdDzMTKWU+l3UgEWL
z^fU|7jj{yU1SvWn*z6=->G)+~FJBqyO{aMc?5=@QN8Yo1m$=!x=6jcJ#6UAmkmiC)
zN024xSW(;Nw!$o9VoF(V5y>rLr6b4^dhioFMaLr-8L^=Q)q+heT@Nimmj3&4>`yjd
zv(Y0)w3k-s)(WONf-Dim+udwd>s^o$2ii+BsGA`|N04Rk;kWZ&{o|L(#zbx};xmXW
zgP7<DvV0mMmK`>?$ulPUbfBeT<E3Id9eL04u6z+UX+(%tH%E!VQK}m)L6-3|3V-(h
z?q6ptND%F%pG`ki=L4H3tVw@ScUCV_kLcNocvV-Gedj}DK)_HW0v=|yt$Z0=as^C@
zCH+1{k|OVX;P=TV$@cL+f5ga07b@yiu(?VXOiPeu?%w`)E~RET=n*mPp#ZuSfSHaU
z%gTl?bD~Fmf4&>BqAZyp%>-*5L6)pui`7;^sfIG*PX}rQuSWPpN024AHEeBMOvX`D
zVorN$AjS>ER!5NKpE>$Fd&Jzi*NsRhOEyTe=|Eb7EF0X~evNN7++j*gD9Z_Ae1fj2
zmb_=#7x3v_=lIDMB!IH`*!ZYAAK1M4`Gs{~%+K?X5hps;8{l;V!gT~$PPWYWx;8Fr
zzKr-$mYd*p6GC+aS<b<n&Erxwn8`>l%F+Z@O`z5hWGVj}88X{)+OIO=NLd=ds{tZ)
z1X(JlE@^RVmG_sCV9HVl)^%X7Bgj%+waN5pNlBIkky2>_Z2}dY4{W|>JY~l4WY;Vi
z@u5A`gI_(w>Ikyb7i~%J5;LYoyP4ZS)&?fLmLN-mOyoND)-V@KBBs531Nz^9v5p|i
zEvGM2%d$*w8xspUk2l2W4e`|xWVySf<=5$r(}InOHSOgx7+$8frzP)M*4-@L#a|v{
zL6o$YV4Gl{&IdNPcFjKhN7*;$El413_OS6#bw059-jDhc`O=_h8S$k9+y=|r)Qe~d
zvOL||Pw&Fwp*JlEM|<dH(@W9$z~<RgrY3_U^_x1+N;rNm$o7IUu7FhMU(qbkzk*cf
zRna6lomUl0JH6@qpZN@M%z!}jsvLhD9FIdF_HXF5D*y9!-Z#Y@ALtO+9kt8xK@LIP
z(a&;xFNa>;Q4}NVe}{S1-kuH;ewg1de;nAfee$Ee%Kzj6$GaN3_P_~P`ksIGlkvTc
z^ydmtS3m%cY*m)F{?P4>e@G;J6<Ah*HL6R4-KxM{z|~+`&9+wqUkx7Ej-pAVx?=)|
z=BP&x%Z0&CI%m&O-RR#23sf7&KZl^_5P=)gTK+2P)TIs+I@Lad`V7CT+Z)!`59q(y
z^_`EtgkL0EWP~nG!l!{O4H*0kVxB=P_9o%8iB&eSM&B*rcM#JZ#0)hk;fsi65wQ|*
zF|jHp)&f38oR1NgPjz0~ruM=t(ZfDd{_=XERt=<`r>3x4;}zrWQJWk;Nim5*=@5|)
zQ7AjdmqI`(1fecDegg^LKq7JO9KVr7Y$Q<v-b4m&BC!I_C(-$2AY#WA{Rb%b-kU*J
z-aFJg63q?|a%vU(LAD>va6GB*I7h2M4b#o9J<A25euJRjU^ps;;|~&}gM`O{l_7_Z
zKM(pz`jK_(JYqn8jq%AjRQ}~)w;Viiq514&za^4^DYPLMdgj6>D5-nmn_xS#sEkPX
z%_M#^8G;f?_$?%C3kgT;F*q^GXU~^eRKh~yS4aY|L9-v-cMwkj7ZL9w;w#`{;!{lg
z1YAO_N{F1rZY9K>#p6oI1Ob<lfKn2K*wfeFd(n{kJJi!`hOo^r5k=u>u%q7$JunX9
z_*W403PzyLe4<-3_2i=i^fVGE3yCxKsFA>S2T=&Ph&U7xM*$ZT$713n;1Uv3BJ4s5
zUrJm`Nq21bAMJ1D8#Uw&UF>$SYX=q1ARx0^<WV<o7Ht@2GRy*P1S3T1i?!h)?z-Og
z&n`{ZoAiv%c!ii-A(mlqCgE3#xs@zlC6=!e+Y0R&VtEFOGsW^u7H5mOY~gzhK&gS0
zf~=H1Hnd?V8c2T==Y$9G7LHVcKKgy^(Y0Z8N?oTmG(`zG8VRWyhzABq9RHR$z9lL=
zbv0Q1mUtl!|GJ|+VYb<A+P{COf8+*T$P;1y*kqfTcs~8vFVrvM{m;6Hmu`g$;3~is
z?V95&fvXg-R2K)M7!mKv8}00dT&4YahI)phpNOj661;g#L^555bl}p#63s`0Rq5c3
zIEHw*_3%iigYf(({2-JzZg8MgWyHn=`tv3jw@KKV=)0+gwLv#7*ieIrQ%O^dSJN++
ziVc>EjZlLlCaoJ?vh8MX+OS4!wnl7;4WrB|yeICn?oJz)fy*)o#W{^Sc;u(P8wWn6
z4Zfki;kZVlUT^enJF(2~gHv{NQ|apy?RL|}5-M6a^~%_d29HgLaRRP{m`aGnrNwzr
zQMAbc3XkUtZP4Zd*f8$9USEvu;o>aeX#8zg0BY0%M;F67gSg=8aP&jxqr?+?;`l~l
z-bftLKE~ZmA9ZBv%si@|MdphfafTc{K+V(OJ;vOFI&n?nfz#mlS==l++MLEg8Z8F}
zb~3*8{s0j-_XJ(NK75~8+}QXr6-tw(3~))wTKj$f#Z%0Wq}tE@eWxDXKZc%`xPLg_
z%g~EK-iF=`@-_6eM6J+$Z0N@ze?$NGC~5!tAO3!4pqz@C1134(h3NywKLO(>;EC=8
z=S~IHWJowyj^;4FnGlWX4ab**Q@Ic@Ve&-#-v_dNtg&lTDAW&9eW@;S5=PL5(#XKk
zu-m!~9C2bC4O!f>iLap8T;iNd+|WUC{4o-EjD!e^zC?^J5i3lBIQ|iFd_-JP@hLIk
zKe~+yh^NX4u?SJ2f>X~ox6YT;RC#cGq$-jZR2!v=Vo<aynn43p0~j<=HIP9usu%{v
zs$v-wr;20HV8g+t!g%q9@eE2ZOkhxI_tc)KD2|`keIA46cc0Io?V#8WPH1NwzXKFI
zSiBPyJ6XI76uVel1d1XS7lWdh#k)bVo5g!Tv4_PapeSK+J($&lH6~&l%_YrlfHkI>
z)U(q#3al|3<Y*#kb`z{I$E3y!W{qHt88pYAC9Y?QyP)H9#PuAD%ZO_ki_a6+^DHhW
zuH`JgKwK}d_#$z=$l?m(TEXH<;#$e#7UJ9@>@kkNOPuerxRp4!viKfxzQ^J=;@rmK
z`^5P^i`$8FJBuF>=Lal)NSq%EcyZ$O;D6swpG~cD@qoocg)Orb1}ug57>&BCR9C4G
zY%&L;a$vaNEa;)K8f*ld2Vr>-FGNdCpx*?RnDA?yOXL$6CfINX1a&|p;*6VxGj@k;
zi=#Hk#xAH{I_;w%I|>fCf2F!I3Hpevf#USXYu*J#QycaR2}T>HsBcI(gCaxX1QgyQ
z++8rUh#nCPigb_k6WXHOqZkzJ9?hTu?gJP!(0w3-V%%dG6zd+#pg8w91`ReFY$J>p
zZx+v>1hWJN4fPxvAaIT}9m$+JDK05)f^NTb`I14YZmB*(+dQ{<44UsYpFs;{3oQiB
zMNW%c1++xAglS6yr!-(%w}I_8P@<AJz7T8+S^O>7e#_$Rpxh2>fqw@mcd&RTD0i}W
z7bthJxCoR*EG`CRF^hMDayN_jfN~FuOF&t|;u<im0XddAI9l{Gy$W(X(l}Z*HLV3X
z#<LtP8OiIxk#%D78{o(~F?j<xvQA8X6C7D5CT|2s+~pkq3z+^wyVRAQIQ|q-ogyBB
zQ#(ymr&)Z4sLrtXEb%x?{007V#N!-`%ZNuAi_a5}^DHhW9_1{)Ks+w6_#*MR$l?m(
zQNiL$;!(-sCZcR2YCLE;{ti*zVR182HnX^es9T7az<-yh@3Od+s9RZlkErjlxQ(dW
zSbU$T@3Xj_sM}fmfT$m^_#sh06mZtV@VaSJXTSNsr*zrRkDq(m-{Iaz&wf{Dnfh$~
zcIHt7O`Oh_QZ#s6I!r)EK$Frxl~yD)C5o+tL1@a74@xbP4@xZ({?DZr3IFF(i>6)>
zyO|8am<}^M8YydQ;CL{t&B`f^U+_=y2M0%9(y4b-YR2`|#B};0=IHbBxS}zwwm^>E
zaC|<5<+JD4O9*-?oKNeTd!7wFz5OBWZ<*M7nb-vrm36Nsj{C|;agC-QZQ$Mpy#(WJ
zgP1lLDd0?Eoyn#nnIs^SgkcoB{)gyYwJR3$^r(A8%pVa4!3y3I^S8tSlUTHFx)j<9
zKFWa6KaqA;?(t`i&m*pR#2=S3-}`WJ$};c$R1~jJuL#U8&}q@}4}t8E;H9OyGCv-*
zTMFD-7M9Ho{2=Aq8rIauJ|V{NFBkr{_rGZuR`&bm|G5x*eS61WD44i`GPZ)E6;zn^
za<sy3{zh1_17qub3bTz&DPt9wRDlK7WI0-(HK`Wxk)P)uvC4eD`y<!cjs_guxS-}y
zG+p6kVzXsJcK*Fg|EG(iTI2rm8lSAZHhORH3ys71n8ji|2&rO@fb0m^qx+ER?u~HI
zok?og^N-N_?HY4dhxnj5V{}O?H08jR3*T!LE3XDORJ2rg$%Or#FN&XG7^$EBKld0O
z-IVb#$PR;xFhlLt6!utgwdV5WhaG1>u!@QY`Dq310X0-(tHj(Yp(ZBbGsIj5i!;Ss
zCX2JgT$X_G&Wldu0LTu2xiAs!1sG1`^6cV-*LRbW|MDLB(K!72YokZK|8hsw&I?w3
z)o-W~oq84cRzV<YMWZFZY6wD%*E-ZYYi{fW2UY<I6)+Sf)|B7av*t>l9t(#IR;>TX
zo4KZL?npPsi%aNc%mDoiu*4f5O~u4E6V#Yo(fS40W`i0JBMcwu`Wyt=LFj>kN_AI~
zsM`AF*RA#D+SGmYFaggrdIV{2MA7rscWfClZDZ!j53B$MqWpg@?ckY48?<+$sPm@M
zMES$9MaI<ND?wc;l$V-=uJnIZVhtZz+0E5+zIrku*p|-kIXFItKwJ{s%Tyxm9WE*l
zYku@})?DDCpd6pbCAwffhlep`(p=*TOnJmAkGS9w-fF#r&uugEpgZOb1igViI6BAw
z0YQHV_?}mnj;=;lS(LGn7&H<yA*gsm4BikkQ~+K!(#?AuxZ_}lD%04kyb5}vTvFZ5
zF^YAcKfc%3Zciayh~L5KcVQv$9#&7<d6g`PGC?tcxv`sQJJAlM*VMt;Ad_Cf{3!_J
z=%rD=EhH3|;rW*1XFT^ixqRfk@WS}Pqh%R`{r>N1!plV$7hAqQXY_v`==JNJXFMC%
zcTlBng0M~SDVB9av!{G9DV+XC0K7&Qp<csFy0w@;T{03dp2fPn2zLsGYa(<O$+Q2K
ze2DfX>aKZo{t@2DW^MbkQ%>cw?mg6AE}q`5?mazJZq6caWD$wzf3CzoS9}qPJcL(g
d;nhocRSB;y!fQ|A)lGPH7hXMu*B-*_e*p&Vw1)ry

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_instance.hpp.94D5335C0371893F.idx b/.cache/clangd/index/device_image_to_column_instance.hpp.94D5335C0371893F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f0ad3d5929e02dadcc19ddaa97084890de5b383b
GIT binary patch
literal 2682
zcmYk8e{2&~9Kdg-bZ>3P)|=TFGxT>GU`C+!=h<5~<J#Vp!63ofU)K)rZome*4%{Fl
zj4fG)qGk{iOid7?s7!vv5HeUIus|jT$i)3e5QhW-MdA;L8^LSC>wPcjH}|>E-Rpbr
z`+fIrb-1~?OHB~3y%p~1<hwIUf*^F#FWcR<-daKsKg=PBp4`~#%Ns5%t-OBt$w%w%
zlBG>L>)>j2-y};1I&QD1AE_D6O%2R$pMG`s#;RC53K;}Z;M!Imx-oEhcb3=u_rjw1
z_|shZwZ}F5#|!s1jp@Zh$t!E7Ck}qrk-2&Noh=*qhCSN?m)7OVZ)g9V_3b|~{pp>K
ziP(BmU9$4Z8qeL_uV)W@GkpD%ivy>2oT#$wJ6Aq*d!T#NwRU7`@BY)*esNCS4E;F!
z?(nt;Z%;8*r8_5UcP(%0YwbLDrb4ZF>7MOtG+F!5bo7;>-Mg&i-!sV?<w;*|aB`+@
z{y!qB<1Tzvy6WP_(^FTCTL;N2%XZ8vpG_S8X>q!JwD;vz4bL7QB;ISv&FGsSrdCkZ
zzdcx+ThzBy`EA|k=%G!Yb7*MO1|r+r+4e-5wUY6sl{=3=IHggj%LuYoRlB4Rh%%)@
zvV`u~(1I;T?_JO+RFWoDWLY5)WqC;u4Kt5^|ME`nF^vLAx>{9TS_nkh94Qe5(K=Bv
z-M=KjXcSsW2ULN@g@7CHPcLXbvdpnnqnImcol2)J1l;(2@3w~V*rTNyMY*I~RjrE(
z0YCk;a_z|*o5$~H6s3}`QdN}|0)G1GzQ4Dhj{ecFQD`Jxuc}{M2>9tPUdaAjDv!TJ
zT^RSC^D_rVZczq2ZmCQ8LF3C>(m^{Kanlm9jux4mv=cZd>jch4yMS}CF5uj>8#p)X
z2F^o!fb*~(;JmaKI4|o3&PV%z^RYhQD4GI}VyU9T`P8cmd6b835Do3PRJ5c~Gy*b-
zCO~GvEXsFU1PdT5Zxv)`<86TKyd983bhLws6`g=wq6?5)aEtPt9>D|1%X<ac`FI~7
zil+bt#b7(GRxKG8!+;op5oO2ntRPc_j{u5_(RN(5S~4ca0L8^PpoEYR<vWu?5>Sdy
z39{pO4p5p;1IqB3BH=zsg`8h#ICWf4ZgYC<>~x8i)ai8d@si^{9QbSMJ5}>SEondo
zy_^&yG6FIo6Cg7(1F|3sAS<#0vLPEFJF)|EActN~iW4~jxsVHx8@T~_kOz<#c>(#5
z4-kbYKtUAL<H?ayi^6~y!~kLu3n+pjfTAd>my;4hF+gz?2b4ewKuMGYltL*$9O3|_
zQ5sMNWs1bFXdLT1@kJY{C8JcdQSL@G6m5|yM#X@Og<`<PsW@=)P#m}fl>jagN&uIn
zlE5WHN#Igc3b<4#1sq3lz;PiCxHOdpE*(k(m!UGiWkQ*v!<GAdlfd;Y)vw>sDA%un
zF|^2J^c#UQGDhG`eiLve#sr+%ZwAiHn1QqSEx=hA3vgDy6*wzn1<vNT0cT@uz}fwF
z;OvaO=m!6M{YCH9_ZLX(m(-CpcwahlzrXj*Te%OETC&zr>ybxXzr5ZpudcxIfLpHp
z)@7|OTrML$XXdWzJlHI?TRpG38o&6lmfj-y4HwUHaj#Vu_E>ybO9ts+qx=U!HrOIl
zm<|IMX2ZZSGy@#NGQhDk3mnU`z(wc?a1l0Aba>EG&-0r-?qV0;k;#9GWV_a7sr*Sr
iJ`@fG+3ii9zx0M<t8!jtg^sAif*|I}r}<CI=g0ps!06uq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_ndhwgc_3d_instance.cpp.709987A54287EE7C.idx b/.cache/clangd/index/device_image_to_column_ndhwgc_3d_instance.cpp.709987A54287EE7C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dcce595072a83b02812e34edf3abcbcdc7dec91e
GIT binary patch
literal 2168
zcmZvc4>;3#9LImyV$9YiF^pmUXBMjr%gq|LVJv=<-DxDUTVb4k{qZNsQxtLalys$J
zM^5o@kJ3M!<Rq!9bnfctObK<W!yWgVd%otL=YG%g`~H5P@8|n@eLkP}@3V<`zP^b%
z5X9NeOP9r_NKg<25s+J+lDu=j76i>>ASk_}ElegKbYJ;#VQh`AP1nUAr!3DU#@6#r
ze&IQzEU|g^6Gd>N9Q$(L-teaf4>S$vy&W>D+H3xHuHGLaLMKLlNWz$Lna=OK(u}-v
zJ~es|%nW_#?aIh(G%glq*T%h%PED3vF==CI{BXkVmfWD{S|8#2Ebp5$@3qM{{#qvt
zKRaPG_91~uCLXJ{__{Xb!@Jo(ew!bVW$qQ+Z8?N*j1a7e7<268_t|{PY)@zjJDJ?m
zdtcgjpf+JNV@GH!9A7Sps4Ch%s5{`FdT@T`lyBVKq_n0)Z~luMQ%+Rt=MJXI3w4i}
z#+44kRfm-H`CY3fd=xT5epmLR>jghcI;%>XJ+i}XFEyqfE4d(pS})Z-x!>H7{>Se_
zh8?#)^6hjU@smxPl=gAmuMA@S&X>Mu@i;`ha4#dq!%g(K28G`wu)+G7`}d_?kk;q~
zSJjH$Pi{WCD&|4CmA*v-;r6*nhkM0W71Ijr#?Ui8?maKL$MR*_k#T9|Qm>}3jvncU
z*_|b`NvMSo<+k<^xdMG9cT+Lf;CVkfbmydE>X<Z^+Obu-i{Vf?RTiBn39rSql#$A*
zD^S91;a#rTw7RO?<}JM!y(-%ZDx?EiHP`RxM?5UeeYk4H$4&7MM3qFx1bqH#PY<)^
zIv28NL9z#Pn=O9K;_r)k^foe3Ha^&TIp5jOv{BkV!R(WxBaV;io_lvanAM-I%P4kh
zxq3N0syA*!?zagXgMr-0&L<6+q{!%!*6Y!|{_kH;`bMy=N2Y(_pU~8Io|xG?m7-Il
zH<C&UWPDxvY{7ia=98-<?VhqQpu>4jyfr_2g?E&5+}vcvjVob8r+M&i_3RxLTMQoi
z;Ymi)Q$f1-mHdJ+t|l$VqiiHUKr|}O>);n{{we*)(OZ9(_7^|R)pyZ0(#jMm_76O5
z`PCpj;C@!$Y)fa=wxm&Vbv^vnyn1l<mLj{Z#<bcg0X7jkR_t_~OV}Tm6)loA1&?7W
zwnpsID=8jfXHi~zNEo}ZXKG%w+>TGY?JA2<NG(gBIY-r^FBdFScIR0-&b|2=C-ymg
zaz5Cur(ihm{6d<?Xw8vdZou3qR{zR3-6KuB<2-4M>)2$VEidR*1^wgk{W}Bd@DD!4
z4f9M~p0Rmt3*J?Q$4$<*H(m2NeAxfFVxPPCCgtHmW?{Uf^gAE+SszHAA&cIF{2nbc
zUW^nfnQd4oWHSU?qpdeAE{JFGg3T(aA$F|iSV)ApX=vK|#Ra$}=HpY6X`w9^5+ZJW
zw7&b|0^B*>gubhc$Fo={7;zh;jW;YVz?~esLw*VQFVN`l5oR}xH)~*FB8|8iKndR!
znxxjhZr8-ZBreH+2?mq+q(ER1mC@rYtC><17RFI<Oc0BNP)JO02n$mwROV8vFx{BG
z7Ia2A$l+OCezi*n0kCi!HGmjS{9J93SR{@bM~2B_wMAu7Icm5)Y%f+@^mTMkkO~Vk
z!Wfaj0%tUI7ZoqMGxts1OHYa|y1C3uN)AQDu+z3P1X{5R#-W+%ltC2gKN_K6GL0+(
zrJ!JYn!N~sQ80tX2m$E?nlvX{+JU<&p+az9EbB~2THcG0y1e>8`c?o4>BbM|Ieo@Q
z3R|8fG>B-YEb&6CxELIM85cvsgTjyxfT7|+VOZD&<AMiV@I-V_%@@!CfRJs~T!HP_
zKz3WsmFj3@DCjIY2LvM)mo5O-Zb5T+!<;w=Db-EeZAn8=4T4n7B0%PCGx4om>-H6b
z#Bp&vkQlK9I3ci3kFF;NK@;5wlAujkLf#48I2L8`qzmbuh$32n(^!~{B3prch+rst
zD|Pn_6vGOXdS|$(bJ*$V0fcL&ZMMXHMMqC-<BblhR5ip|sJSFVl7*T}HKbaoxpaNH
z5#YXIW_ovQeMd*Q%ctP)mGIS}-5tLpy;8YUe|0RCPYndt-(}NT-*1`7Lt>Xt7nB_P
zr9?H4^5ht~wdy5IwfuX9f)_{&W<z@GP_m*~&9G1?Mz;OSc*G~z5vYYiuh7=P=wbC&
z;&D6zmt#xN<@pee3|1MF+?k$Utc~8E(F~nDtk>IH{NL_sQ!_Hf(u!(h=is=;g}&Bp
Y-3D`piL>h_!_|KLa82JJw!cQeKV!pMKmY&$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_nhwgc_2d_instance.cpp.E16A024C295F734A.idx b/.cache/clangd/index/device_image_to_column_nhwgc_2d_instance.cpp.E16A024C295F734A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..57d39235b904b71862dc9ea6766f580ec116c95d
GIT binary patch
literal 2168
zcmWIYbaN}<U|?`g@vO*AElFfyU|<jg;^LB`%z9=9hVPsV3>9<shUOOEvJm+8|D_I>
zXP@`Xht5t9PsL0&*jx4~P}w^>x>@I^isYo<Ut<%QIWw1wzkAw!(XRhnZFmZUWP>)>
z;})iesivCU$G@j5HS=G-|MSheU(YW0+&|O4Soy)LR-1Ua*ctlz=lUBiJ{A0M>BysZ
z-c0Z8&gAX;=hnWy{>+b0pa0H@*B76x$T+7yM(5|J%U`Ge{QEq*uGag+_m2L{neI9f
zPqR`&1-w}HruE)l@ca<hXOWVGOILfjmCR3sztEMMw*8jIvb82_4+($qziwjn{gAz^
z&Yk#_4+=dC_3y5`=w`JrbQgnik%?mGJ?r^(*I&#}TADkv<@k%0mHe%rw|aQpdd0!$
z>zMvpLEZB<M+txKg>z*aYQDewE4t*sHZOM$lNBw!QQxL)e)4JNCS^g<mfpTiOs%aM
zzWHv6j|00rwm4<Bh)?|D<iBfL7ejCP*6!-!A0CEovC5e(FQO&q@>_RpTAY(rtK-~J
z`rhblQH^LyX0*Vyo6B!)WwhGOek`LqfZPACMMM3<@(p5f+)GtAPI7s8+HC7${uS30
z10F8CkRh@pDm&3cQ|{xGq>_&-56W$yA~{usmF0ZG^(96PdTSRql|?8|v6|!D;FPhM
z*@o>)_NFDGtV=a{_A4j7)_uBi!Ln}w87mhbPCVXrs`8b6kDj8H$Ld|KC*QB$Rl9!W
z+SthL8#f5vTlQwxhVmEzR;PPntW_D!8inp#8ku|_>#TXDwQu#aZfU{S$1WTxdBAp*
z>zZM%w^q>n@KutYjQ?{kdPylR-yZf$BqQVRwEJxS?Jr7Sx(2HXuU(q5ZGU^+%C&C-
zPiR#t31@Q4@9HpGa8Q3*_sehAi}SNW|EEowx+Qypw0gM9+ePg$Cq#YKS<_3_K7OLZ
zJl#=P`;(K}%Hnljk7&JBU&Cu>%e2_r%{s~3Y>&(w-WQ%r|M>>puRHs|(E6?I?Y+7O
zy{DDkt9&Q*=G%m9=|u~eHf_qBxo%~ojg{Wq%<SK3_3G70T-T$N?3;G3HCy`RdDeQH
z3BHpJrOcnrUGs4I!J`j)p5EHzWge7vIeOzE%Ut&#6OZmyT=i@A9=;g2>J#<LCW#-u
z7^nY7cdD?7-nsiL7OtPCdE#nC&7FB4cNiPWy?@txJkoKhjPW=A|7T{zyj>{oe{{1>
z+~(%4TmS0aCa0O$w(QyOr2X3Vt?Iv$b^lsD{y+Y@F!ba;#*4e>)H4)U<|fqu^JC`M
zi*;=ama_0L1OzjLNw7+ofeD}t7)(jtV<foR<vkBW5Kvl=RnHVmz@)FYi*x(r@9O1Y
z@CQosvhtdP37GUt>Fm9F;+?WQ41qvtWmaV~FaeV;N=++X4lD~8k2v`^Y<RPckta+@
zPbe5>sNniesnT_?4m0tDNxDe7!(AFC=_%<0;|aY;dSk2fa}o<r7@q>4B}^|+i-M#j
zOc76*ih_zI+%;ha5(b7apQ*8Kn<A4f8v?Xkon2iN<~GOnHw-_&l-y!rL2(mHn5>>`
z5X@YbFbzG8Aegx<VJ3PefiQEune?>P^jc1U92USAi{vmr$!M4XK!^FMM8gc=2{Z69
zh=dssc#{8dv$lg5&|%27g!l=mznEyl3Nn|Ij~~fgPDy?wb2(M`5$5W1>hr_QP4#u$
z+qiLA5-?Z|Yz&-W!a$yjfftMiGl-E3oE&-&{a&!<p|wK3_qOm2zhfGKZZc*!h6h%2
zf|T28L*p`_k$f(E9tbz_dGYzdcrc@ZMuO52%*a`WTUM4e8*bnL>b2!_f~f{3g>a!5
z1W(dd(h24$o-o-kS>!OZRdGW2RU=FTIYdnjtY9|tgqehzM8SBopPR}0F}yno^ot<7
zAlxs9LhN3~%U?eZ3|}D@Ati(hWLadD5H8SQ(NIFzX~JTn1T**0<KCl>wfkFu=E|_k
zz|Dmj1oS>ApTNBT)PVI1JHL20C^|*?<YB6S=~q-z9$}q|sERzoIs;w<ahP>5QJ}e?
z3<fjzwPlt4k*JRbKvy8A%D)0(o9^Zxc?OIr6&DqEm}Vf)Q^g0y`#a@J^^S;l%|N}#
zshy!HH7602KZ=Wzi<Lo)P!ItI%wbWRzlA=KDY?KD8p03)6J}sY&P~b50CK(4e3@BT
z+1NQaxp{a6`1w2qU7ghgxjdbO#Dzs9BrPnhZEWrA9rQ%Ct(46)<gtm0Nz2G8C@QI_
hs%vWL=o=Uso0`j+Na-56h>E#;g)_PNIeIYq002M8MsxrG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_image_to_column_nwgc_1d_instance.cpp.D56D4C0735B48980.idx b/.cache/clangd/index/device_image_to_column_nwgc_1d_instance.cpp.D56D4C0735B48980.idx
new file mode 100755
index 0000000000000000000000000000000000000000..adf441a213a6db4ab07db44f122d3ceb2a77a645
GIT binary patch
literal 2166
zcmZux4^&cT8^0nqg03hSCMAMEf=b*HWvO683`IWt@3qWUIyKT*UqURiInB26OwG!s
zTeU1QkDpmqPWP4OiFB5JtF$dQOUqe)TK?G@)iE`9^_(kw=X{*QdwJgH{r#Tb^SjSW
zv{Iq?WHA89Q<QtOsrn>5000#9tJ7z0&#?u7KZpRZ=VWu7R@b}M;mz!bfTS)~eKQ~r
z8t6zowz}EyDvG1Z*px4r<kC$u?=IHm9jJex>b(vtrX;;@Hy^P{0WR_j9Z3huemN&_
z?LSoj^+(J;yZWf+<zZ2dlvx~}+sS*iFDo<YhUD{XT%j;Lx60kA@wDSVr*aH4&+Yy9
zO=kR9c6xkhI{n|}5b^2st>imp%u9jyCIb8Zs96zoWBtDJ@POo-p7ETT{gLcp-;BqL
zq|1<U?fpDu{0G8Ak6ByJG-@{VryI)~y6Z_>&dlyr8+GvX8_mGDBtJU|`fS8dB<?2G
zc7x7Gj6|imf6{zi(Nc80A(iTSwrZ$v+1(zt$MzZy-UmEYn}3q4<=mIn%K{$VeD%|b
zPs>Vhk33GX?pIL7>A%I?J=z)4Q5+oWAolMdEp<BU+f4tWD8z%^MXIEH6x_W;I+Eat
z^XdOZ==z@l+cxLjMdv7Icc$tCM?E@<uT;OThTI-Dclvbxxv`{DE8DQG@%s#0z2R_S
zm(v-i@80dg&&npYu4-}$Qzfq~FqLW3Lzt1Z@(PpuOcy`&`-dh?v}C}eZTR<wo34$e
z^y6F*e|T%pIqzIyt#Q9$1E<vgJ9%!9<|;w5$eeNI91X1WfLeITFIPRUt~mcDQd4cb
zmDK)aN7jpVg+i8pXw5bGor$a0KC7v&-Mq2!avi1r{OGm1w}}p5P`?wnSCh}*6;Ypu
zQx6I*yzsnnY3SGr%F7?`x9XoRY9l>dqg8oEpN{_*U1>A>Nmtl%x2jK(cS|*znc^p4
z#KGa+BkN;UI@VUE7+TtMs!xqZ-S*64J8oYxas7~Y#ZCC@uSUiLj62igX48%yug@r1
z!HW+bGagL5O;hv0ZTi~5X9C+YnWNXMAoptB#n-K#W4sG?fl{1ti6S6bC2n#)X*aB_
ze6Nmvl5_8=TXtV|!{?j!mkGw2{(9{D>zG@QI9LXzq(1Q89CLM#KGFZd!#8d3KN>6W
zNX#gh#t#`{!UAIneR(^+E$pgaGm&D~Vt1FYP0;go+_fX#C;AW3?r!%py@gz%sID!)
z>z{h}2EDf}YH}_nrw=^QG$js9k996JKQ~Tg`nsMLr%A>Ih3!+>gS(JON_*kVvnS5S
zYHZCjS$j%bb%w~O_St)yZvA;)$=i6%#^$P`R%Y?f@F8={W|xGcqh7vg-&E+bw8{I>
z+bB{LUN>}bY9Syp20)xaXW#h+&CGu!ZQnw0$uklJM4@R8$nl+Du=FO=j!EUg=5h##
zM$<x2=rg}y>C3Oz9*@v>a3CN8P1}KX;`s$jXQrm<s?dJ|o7SL++`7>Q8wjBasWBLe
za;aeX?u##O;UI)Arz;keBJ@zY3X7=2$)i&C<WW3?kXd9u42wcobU*A6gm770zlB~9
zm<9V`(%r#^QkM)yELxmL;L$K`7}SO|??;hos^Y>x01M+lx97~jwq*xLfl0Hz7mxp+
z4)6#=$cV<M@CaYXkH#4Ah)5`k!c?rEZGFsOM0YPn+mMoFmNuyI)J>L%E~U#bItXDP
zj08)>m2zbmIfU>Leu5=}rLYVu1tFq1(MBv%;e=l9!UMPG{EHwbT9s1Mt;y9~Yq}VC
zD@G7?$9^#1>%%bAKkN2l!>PRCBNA|qmq@m^;w94Utwa;K_Es{9usz0$JyECU<_dE<
zRV{1IyL>(w9ST?uE3s2(q=FaR+^K4aubWEDLp6C5ycgbS{?_G+nl&N=iY3d*Ar?(!
zC0UI{C;lsDgahMuP%MQ&S-@gxRL<NSV{%Nffg@X(J?&_lR6Lc9okHIYp24=X$;b2A
zmNrFr5gQY38Z2rX^g4V1E$2dTSt$2>HE1T-FZ*iFJ{p;6;ic2)ObaiU#${S~VH(WD
zcwhSMUDulM=W3MKdd_AXkSl#VTZd2!xN@$-g5`#CRai7rYR>*)!+1W5wVq!fGj&JO
z-2KpHeyro5Pg^8%syye1%RMSwTr3cWT^#_3wfy{8o3cZLhN?8REgoD%SWH|3**Vyg
zm6Y{zcM3@vL|x{%)S154FF+y<T=$`n=H<^3^O^r|*J*_dgXPBNuH<=m3gFdiM7|%m
W`Yd1N9ZYiy35&-m!eya0|M(la)lG%~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_bwd_data_f16_instance.cpp.7967A43F40615422.idx b/.cache/clangd/index/device_layernorm2d_bwd_data_f16_instance.cpp.7967A43F40615422.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd36674a6656001d8db67cadfef10945f85d77de
GIT binary patch
literal 1384
zcmWIYbaP8!Wngel@vO*AElFfyU|<jg;^LB`%xGo?h9?{h3>9<EhTPA;Z6<K-fAAbh
z^O@JVJ~%(RwR)qP;V!XC5x*xVm3$^jdEWf^@0XmecHSL^=*{hKT_4>&zWet(apMM_
z1EEt3S1{~dv;6g|kAGRyJ|uptHr&5^-%k_$Jh8_oXFj@XT)EL|_lAhreYU|q>&}Us
zkJWj<Pvz*{_Md03-~aL9Yk%d&Up4zR)`g#bIYZR-W8qG%u=(ru>%6nSFLr(Vno7<e
z7PG$|2)eRPF_1k(;!D_;JokML-D&bA0+&qJaz{vMnLW{!3@p2)v2d-)+(qFx_TRho
z;?m;#^KI_Lr+sKR_F;Oyd5ZG<6I!v22PgS>Ec$+K`TuEu?5^EfH~E8vt@-<fMf(n>
zELqzpa3Z8;8=u#-ExiZ6pJ02Nvwm9sy*>5poi`4?@L|@~+!A<N`H4>ZpK~`hdT<@`
zbB}CZ#ARJvvUtT;6+@#n9#UT<ZtBT6UN00}@O53`$CI+QM#oo3Cv^+dsxSVW*kECO
zZrR1>MlCmFoR>xR*-wbM&3D0ScYcicg74K%J=<7#tKJ>={3%|^9{ojX>a-p&v5V_>
zU+Gg`=+=IqwcTrS;DoHT7c<Ok4qG05@o3-D#rKNRVkWt8><iv)HkI+qEi1i@_1S?d
zFI%##2$=2oP;<ilP3miBOi7cfV@P_g`+Vhshu;jgUb*;jhr9DzoA2^!;R3QtLTxiF
z-|?4LuUt9zZp8M@8=Uil%WU5=az_YGi4C3DuJ>RT+X8MMlV0PftKl^(o^6(%IP2Pn
zBVSI*K4jZwnCq=|$#8v$@ieBtHz!WDaq^SjBNO~!;aB!Q4)gdgzTSF#o1dQ5>60@5
zme;Lpe#4nMS7~GSmR5e-#XJ=sR$nbD`*H5)?5NTj<D1^MOf`BZOSiFo3s`jLLy3{)
zVrkjs;*vXfZbcM}l%1SgJU2Zg(_S?8!%M*#F`V}VcZ)q<eer0={E&C|O}@Up&Lg+^
z`rh=zw}tkIWIms(ALIR3PV81&M~uV9JnO{EvkzW<v9M@4SF2#X)tmji8X~z{mmGR9
zPxjN)?XNhsSABeP^0>#gT(ycHE#;dU<K3n&<_^r#&bsS0^K4$BGxxoz1{F1y_fE=h
zf3&=*Kjzr!$Uni8C!F-1H-Eay`Rd;`^6#Ejh8ez%>G)s%^1t?^`pd_ZYz@!#Wi2}~
zc|t9-#1z9khT_WHBza(7J7BwY%X6*|Q9KO(0Stj$tXxK50w@CkMX71U%8bB#xxQ1X
zblt1NtUQ6-*4%b*ZGqel+|EcmS8fj^o)@<-63?GI2*wj~`l)eBGx-oBPawZCe;`cQ
zvHcCh&o3pnm{?GRSpo(01VUkooL_QHU!(kQGf*$GaEPCf`iqG+tUzIIPHujf>Qtsf
z$NO&=2!n(bxUFEqFq)AI?BCu)zZa}|XsvK4#O`Ih{Pp8ZJb`>Hd`d7?K$}<u;Jl{>
ztY6sq#k)a9igL@t6fqQ~<|Kkbt+*(;SRBL%0uf-q7?{Vm;@50L&YO%u49U4EIT=7<
z?=)X#7FITP4o+?!UIBhSPeE5_H9;;<Cn0fR5eZ2POKTfjJ9`H`QEe+_GYxrcqGHlA
mvI>eyDyr(5S~~g$hQ_Amawbx`MlPaa?q0!6Zhnp)j6MK+Y+N4z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_bwd_data_f32_instance.cpp.BCCFAE80AA603EB7.idx b/.cache/clangd/index/device_layernorm2d_bwd_data_f32_instance.cpp.BCCFAE80AA603EB7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b34215b662bf216bc4c9a3c3b3081913c518f0dd
GIT binary patch
literal 1392
zcmWIYbaTsKWngel@vO*AElFfyU|<jg;^LB`%xGo?h9?{h3>9<EhTPA;Z6<K-fAAbh
z^~h^dADkcETD?)taF<x6h~JZwN<I^%Ja2yd_e)M!Gw%*V_T~1su8-~>-~IcYxN!r|
zfvHOiS1{~dv+VV!k9Q^0KGc0nx7)va-%pG4`C^Y(3m?6m{wb-nC~@O13;V#B*m)ZB
zc1PT+_3PUG{qt=7`VS9J$A7N*`m%0f{Q7B^4Yfc0Ib1Y#y<J>=+@1PeXV>RPeir(W
zX!g2c>53TNMO-TkUaZZy(^u0ccGm8N!jjd};u{R6<e!K(T9mOh&?!1I%60vQ`n^jp
zwz}`PuiCNy!~@3Ohx+%|opiIC7`lt4nRVs_m-lo1|Lgv&U9%-t{h?u1-aY4+HN7W2
zqGc5)X#UKR717C*ZhSwH_ts56-TQkgzl#ZNY`)mZCKZ&ic$)1Ko%TP+Zfx}6I^^db
z*|<n3`tt6;D;1L@B*RpAt9o+8t1fJfVaeFDI<}_Bz5k5it64XcT<)KGQ6r&nu5a_o
zuPb>JuhpI~IsMUkQub}8sPyk!vzbfk`a3R8Ti|y7uZ7ls#rFx_af>fa3OuRwt916Q
zi^8i)Ibtd_OeI-YUX$ugJ1jgWxJU1u^sBGaj2}-@ar)4E%XB5rg6r1`9j#-Vy=2cT
zI&|LR;qq_%GtD;h=n|95AD9KA^DX^ZO}DqNT-w)nTvkf!`Og0ZrJNTXqpmhjdt4X3
zefq3hw|YyhxvjsLrCz@%x1mAR@!~Yq$_(apnGClMxLmea92$GHJ<D#l;L1E6f1l-d
z1oIN4FK=6-^tE#7tc@Zs-cMWeM6l@BgQfnB?bqZ#Fjv~StiLWdTWe>Au6pqg`QJ<V
zHmOeu?K+}&NzAUK$LPU9{qWN_A6h@p&kFsYwrTp7>;MUMa~|F|3mtYoe3??@cHPSF
zyitLq$;L+-Zzfqkk2=5NQk~YV2bYx#N;NCo-$eE3yQeGH1?{ZAd}aH6<BBtJ-{<k>
zdw&R8GPm5WoVT=kMz*w45#!NanXSJ0Y~hQWyXKw}b^G&j!+xm%jaylsI}h4gKhb{v
ziu3g8?$7G-$~R4BK6og6@1#_nbZpN#mxXIrlqjc~?LNw@TM>Tf!N;c+?)7)N=5yH<
zN$RKnTdbxydG^UCTT~`{rrXct`K~s3P0B4{#eesg{$KC$&s+NQmm?O|t7a(qJAD)7
zNij@gD6Y&+k_YCs%C)PtI~3-;;$iR)U<l-5<uU>jKp6-qN=+-)VFc#OxJ^H@cb;rw
z<O$?g<`0BvlidDgMU`N1KO;+^fSy1oOxUsg4a3hbCAWaW$f`s9gw$V5v|$Ar!p+Ie
z4^y4Wbm(~h?E+ztumZOgOt^HNZA|8SRx_Y5ve__cMlP^Rdk_6yu;!t)!l4kmm+|t~
zk2CQE^0DwK!BheDvIxL=PYqbVu=9&|gKQGzmWL_&D-gEnZvK&HtUzIFZabJTkmtbd
zjKp*0_CVr!ar+|i{JDc*yrR^cL{R7z7bO>qgBU>|0t^@fOO$^xuH1O^C}R*qa&Aga
z22j{L&6k;lm5rT)lbeTEfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`
pn6!+nf})a&s=B6@j=q7Rv8lP7iIlF9i>R2pS1^;CpQ8t(4*)P&P+|Z8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp.756B8C951E2AA179.idx b/.cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp.756B8C951E2AA179.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c0a57a4b0cbfc970e3839d41deec03ed698d6cc8
GIT binary patch
literal 1388
zcmWIYbaP8#Wngel@vO*AElFfyU|<jg;^LB`%miiz1_e$AhKe~?Lvr(PTkw4QFDYS~
zJ^eP*6X)5g(FNJkZ&?3kToB6To~SxsMg8pGUx9^t*4#9>)5Xm9{ixW_jwgrBQyC;1
zw7nk(FjhsLU-jwZS=O`<W&hIm?byBV=Y|vc(;lCk`RJ=~<xacZ8)D-3*#!HnJ162F
zyXKvp`q8_$kDi@gUi0Hq=EU{(_w+x%{+eZ$<~nbG{Myf7+h3nQ{kwnr&#$`Kb!>a)
zpRZ0(&AL8e1+x}Um3Gv2<@t+UH}hL@yqtB7Rfp%P^rvY(8dkYeE`-hUyJ-Aoetu@j
z!;A0z<=*KVS2!g9IK5qZ(?tK2sd@(!9Mx1VRzJU7?_DRKox57a#_jysYK^)129lYp
zm065b=dW#UQq5KVV7tlD`u6Il-QxLi&PrtlR+k;TxpJ?3;;=lexZk);XHwIK%YoMx
ztT=K`=B`GTt!EF9_9Ql2w{6_+7fNGTGWKT2{CU<VH#7MP??%^-eUmT#G;o+Bn|}F4
zv7}Q@pW<bm<MJM1+nBeUdnbLJ^~K+L0*iGWo9F#54EpDL@1WRTx239!C%JwJoqnrO
z<;n|z9UlU+jJSf<no6B*OSf1q5qoEQ$(87HJ>HWz4wh$KUuM*>Zq4JSZ@a{%TFrHC
z5ZWp!GLPw}ZuO-u&&*{H*%iv}70u(4%Ht3AmX+<VHo0=H_<v#%>qW<?tIg9M{|neY
z-*oF{>DOgqWiN71ty^xJ!06qyNZb3<1=ckg47(1vT((#oT6?rT%WkLO%1mB=pY?YH
z^Ae;lZ(E`y``ODhS@S}?Y~)GhCtnVpHe{8Hsz1R0<fp*ibvD~di!Mg`J^Nt))mt{j
zU*ok%XY3N)nirC(2b$y8g?~O=e(?U4RgY(%jJveXF>%fv1CtHyOi$aEPd+Jm_l3po
z69-Z+9GkUxgZH~RuWc54{hs-v$-n==6R!__H@te*yNN5;1?{Zgd}X`6QN@{g-{<k>
zdw&R8GPm5WT=#4BjBIJ8B1YG%WrvoTH?O|veAMr$Xwl|RO-ldDKGdtP3D|OVQc-ht
z*=h0kRhy+NE$7vIHlLI`n^VsH;oO`DS6>vBD01bxu8r2*{8KkC<K%5&w*C9NYCoOQ
ztzw%v@pMJ$=ljMyGw=8;f6hAFuUh}|%YuBXvtE8FT1WmbH~g>ru>Q&`8(aT#eOb#+
z8&9rfo&VZ1thh2aNgkNr4sXxzTwImX&comzz!1pA%4Gp2fHDwJl$uto%m~bz>pP`N
z*S$K-$`i=#!0imz7Rc?&?SaJe;`T-2`Ev&$@j|%6VLW^L<DUyJ8izCT1PUq(2Ec?J
z+utz!{8Dm@i6xL<nIEp21w~PapOE^Ci8ic2z1*DK{4l)}g}>RU%!xV*6h_v&&-nQV
z7QIMskZJ{Pdzk7&A$Bk0<*y%S;tAwq;ZuSM^8^a82*7zTQARFsp!6R4y<p8lYlWu<
ztY6sq#k)a9igL@tR528#<|Kkbt+*(;SRBL%0uf-q7-*d6VxY0T$c8b9AvrfCCj%(#
zo#xBT!pg?Z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%
mRzXooMO9r>OGn?p(Ad;m&O}Pr$VF7l-7A>M&Ck(;(FXuheol7)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp.4E9CDD5C09C49289.idx b/.cache/clangd/index/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp.4E9CDD5C09C49289.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0e6743a5b62591fe3a3e54c4691874b57dce5b8d
GIT binary patch
literal 1396
zcmWIYbaTsLWngel@vO*AElFfyU|<jg;^LB`%miiz1_e$AhKe~?Lvr(PTkw4QFDYS~
zdp(!&)55cpba!m!EtCFrae?YBF(q$1Prq4ze+3rqS#z`CMi*oI@1*V;mroC88#VA8
zn7XuZ1;gGo%R_(uc*bK~Vf|Zs{{7hdH8Y-@cYnOv`RJ=~<wmQW8zN%o%-2|_=P%+P
ztMlGY_2}K(KhHidule!kyG?xUp7<x%S3fN|(`8e?Kj!Ib{_FF_zRTx-ejR=74^M^t
z?5_ucuB=lGWDk+};<hEv+s?f^O}s?llIdFR2#HhIpXf>kmfX@<xYlIuqVyZ~_b<IT
zwD`WP?Vb3P4-LmYOwTt@QJ#N7E4JxilaI%u@8_2PZ~LQn?N*p)MbEt1-vg}fCm3A{
zbLU9$vWsSC^}glypgN_wY+Lwgas7LH+FWw-OBOXw6S=kGDPz$z=eqPa5h|>Ss~4|v
z4D6h1xg+pO)g%eYFcsdao*Z$$1+RB-T&TLf<HuQ9TchJEq?5V@YSkBiPHeESKDX@R
za-)_TGEU1P`|KyI$zji!d)IUw_l4iKN^a3jY_{JYulVD1?;zV=x239!C%JwJoqnrO
z;mQkv9UlU+jJSf<no6B*Q@2<y5qoEA$%W{1J>HWz4wh$KUuM*>Zq4JSZ@a{%TFrBA
z5ZWp!GLPveZ}p`v&&*{H*%iv}70u(4%Ht3AmX+<VHo0=H`2WLKLW`QWtg6;M{&!)X
zeRfv*^=rAhITvqFiS@5Oz%-4;HGJBWh1?Mr8A=*Emle9L{?)~QwWdfp=(5zj8F4#R
zV-Bk=%k^}%dNMWZSkQuct4)*KPrTea-I>c~OI`o_lb;o<Vk>iAKUuVS&Y6GpZ>HZm
zJul#zrqFJW=pPr1Pc^XbkI_y4{eI5=m7(qCllLx(Z9ZhNGc6;5pY@cfzxrh3T^9?h
zrZt{k&}-_RIQ@?0^-8y?--0i)&X;Rk?S0TXN!xIr%-r4|o=^Yz2H%fM{V>D!y{+`U
zX%CiqSii4%7xn79!F4m2Crn+d-nRJWvxP5i?^=6G`$^hUR@XmoAJnU@3D|ORQc<&Y
z*=h0fS5Nlt*<<nlS-#3`Ghv&)L)JGBoW6MUg^$RsuBh!nX`7>CE;nsg=c)Uz`|YCk
z_7{>aZhDWeo~=J`kWyqjbB<V!(aXI)YKQlpRMEDa*eURT{;@yXTk2O{t*n|iNA{}E
z^mMh~T$3Jq`W08^CdmWy+h-dIzf)gV+Ve2@2QUP3v2s~}37`xF6s4vW>o77fNN)eK
zqDnBhpOGa{P+2elrp>Ya4a3hbCAWaW{L1{{FyRnCA@vs%ZCDw30=YT4`C-Bx)6S(n
zEc}!N6h>CPJ@=2}ruKENKw)IF_ZdI`z@iuF4KhT5+a6}fp%A;5@$%P?Gw}rSvG6Iu
zgn0r5SOnm_rv|KF*!jh~L3%~G<zb3oG$R+-m%WF6FIe-?TH&ui*rvPrN1m|)RXcDy
z!&CuzuG}6-JTGovB%VKa5E3th8}9p})SN_6=oJ?w7mI@!K_CJQ7z69eLN9!Ib}*bV
zh#@&QB_{(Y?49Pz%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAEC#r3w
vY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)P$<5EvgV6^7Z{Aw?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_fwd_f16_instance.cpp.572C171AE3FAAC3B.idx b/.cache/clangd/index/device_layernorm2d_fwd_f16_instance.cpp.572C171AE3FAAC3B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f2b90c03cefc287e576d56c1c60e56c34138e8f1
GIT binary patch
literal 1574
zcmWIYbaRtqV_<Mj@vO*AElFfyU|<jg;^LB`%w%Q;hC>_-3>9<EhTPA;Z6<K-fAAbh
z^_kbXJ~%(RwR)qP;V!XCk&I6&g3CPKdffc@@0XnJ>byB(H#W7ub$xXA`0n5D#Elzx
z4oqMAHK<u7)PMD-mv>vwJox+Ods1C}@$bU8J9@_^8y~%${wb-nEOFzmT{gi!>&}au
zkJWi^uYUCJ@1Jj{-~aL9X@BL$Utcz8<b|DnIYZR-W8zM&F!_8teL1_kqUPmmehPj_
zG<)5+bVY6RBEF#03ELKZwO-<6mHGW3%fgz=zZ9;Eb@NVg*1k4(W0b`5oaG-yzpdZC
zCCzuzH+6pga7i1l?4PH%OP5Uaf3ZdL;3?Bv6I_)1U+#Ckb^O*2El&T5J1YB^E_YK?
ze(clIu<gm_&l}QC?%=55k2SdWHS6oMv%9xX=#cF>XL4cDG(D-(mOYOr+?T8t3v|9Q
zb4AvLh*Rg}?q=k0PVM2z4hfsolOx`@;QbDc3zPC|{ygKGKO@<Mcav+!m6b0vc?8mr
z-!yp^(<U%G`Gm>okE|z8l`70F|9z~Sv1-5i5vleI9*^qgOsQ9_yy0aOBz<&6Qee>4
zrDZalrfMvDHdENvwuNLXuT4B`JSVuv?p?IaE7|m8s-7(mtheR|OCAV0SL|3FtG(53
zu5&|>l(hy2^QCFFl4qCjTwcJ~A!~Uq*!l3=#H^Kz9~btw-4gyTuXbI)c1fshhUL5N
zU4K`soO|11fANONPlC#9-!gJ<a)~HU?PQ6aAj;<;J|jaa{bfn~$Az=b=eguWoff{V
zU9Mrip*J%+XkuUG(peitKAh*9`$+K9vj<OSF!AfkeqgS&b1@P<xy?_{>i9{Sf6MD$
zHv4FV&Q;pjy`{C^b}>)IhtpS!%6^>tIXkMf#`vc9Ez^kT^=3T0ZxVH{wk@B0GV<~a
zi@hffq+d8TYw-r}cXQr%1m<RF?Q(c2I3tGpo?v+A$FDDvj`(YR&!6?{TQ-~j&DZxf
zKg{J^;&Q9fzcqj3{lkZL2&jq&gygEnR2lD(vJH9n#lG){eiCPI^x_R0iq7wo*n7n@
zcFN<E=BFoS-S+vgh4a&@#(1~z#oQas{H`nu)tY_xB(GkD<)H^}T&GOd*86>**YeZN
zE^+<Gt~zzEW-PfaGX3a}nWZPLsGUErJ@NV8={EB3o>qn#y^rbmpM3N`bC!L6%tSeX
z_im=bkGG}AGZyD5Z)Yg3%uTWb=D9+~*T1tT-!$Q22nb{d;$h{n024qNFxZmyZ+2hQ
zHdAgUf1orM8y6RtfJqmnrWNmD1m@oLol>RiUL9uT3F0;9wSpTO#B0NAkHmB2bwT2}
z@p>Zhym|d#JjeDo3_rh=++t=45>yuSfeEt&3F!$1Ab29mB5*mURS9z6e|$X8$P<Jt
z>}x(LUre`W0#KOCkjn<<oe)1E^%oOuSb@U4oV@%nBVjZn7dQ}l5B*-S=ApGh;Z>L7
zj2Am>fvS-mP&`Rmv2)I`1t3Edc+Fs%VRo=Wf(m8`P*DJH2uu-Akf5(%Ac7|pC=`m|
ziTH}Z<@T@r5VMa{N(X2Qa#$P+v3nUWfBiT!PmlnMfD%Hl5Q`8JPnbm*&U<RW`h}fe
zyc^_RQC@kNmZH?0L{NMd7bO>qgBZae0t}dgwyy|Z|M_6K6k{+$a&Aga22j{L&6k;l
zm5rT)lbeTEfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`n6!+nf})a&
hs=B6@j=q7Rv8lP7iIlF9i>R2pR|u1vpQ8t(4*-ezn%@8b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm2d_fwd_f32_instance.cpp.C34D1A2542CE24BA.idx b/.cache/clangd/index/device_layernorm2d_fwd_f32_instance.cpp.C34D1A2542CE24BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e959ff7210f0bb9d09b44c72788157d8a62680b2
GIT binary patch
literal 1574
zcmWIYbaRtqV_<Mj@vO*AElFfyU|<jg;^LB`%w%Q;hC>_-3>9<EhTJZ?ZNYQ(e@VhI
zPqW)u$M`$5^KEad%!^a&mzve0(x(~39aOd_KK{a{SI08bcin9N=BiY{@!-KR9tGwP
zshNAPD0p6-ysD<ASk1V?`tR-;cVqAGnJ;So`Qz2jM_+|2_u1?;h>x2yUt^t~zleXV
z&O1BRqjzuLJo|kAj}Jc|uZgJJF<<=p>7yxUx)xPzj}SZk{rdc8-{teVUq@f7;jb{C
z`|W_!m3@wZ>?>0Zq%Z!`4N{!DWp_eTz`o4d3EJI9rA{ssU41S>w<mbp<sVGHr$<L^
zR`T4vxczrGkKB^2f1cjv-E~5}<kpmgr(C%nN)wlVxv%gxDK}<n%W;o8d)hArD^H$S
zsOIDl{pn_9#Kx2umVNDd67Q?FR(*YSHrwOKH#c>uj6iSkUMq)tg(vcRY`a$|=J<w~
zX6U4*^XHjtYw>bxGu6`eblb+wexW>uCF7*IU0rE={~5zsY&Vr$UWI%SVe8oZI7g~j
z%emvMV#=(iKYCK0T1_~&JNEHzhQ0TdlX{&qR0{3;PSsDe$x)G9*_#xyQDfz+AZxxB
zDPbmYIj`n0=T%!LUfb}H^PHxee7UaNmp<dilT@5O^xoROk|#my`OXD<*G-L*Kd<PZ
z*()p3!kBq@UQb$3Tjm9ZBYiT@S1x?`&0y=5iywEmJHPGx%^yBpz;;RK-E7Nun|IZ&
zUpY70dVlGLsZWB+Y~M0+Z*qw!P3>ffogm8RAU-2QD*a_i{Kti}{PSFLqD~86*50mh
zeM4tvbkIb;%B8b5ihP*QH}{d?r?mG^Cou8rzWr!hY3E`jdUBhep4IV_Z~h;zd%4#~
zBXq9P#_lbx{I-jEDn6WknzZZxxgWEmN-K<SdfzgQh+Kb-hxbjQ>eX4xC!dVGJj25B
z^nvsX$7U_w;QemS`;Nfe45eKTF9mn(64|4?PU!LJi#I#whrGM5^7ZX`9=Xla_og4d
zE#%R;<?}qQ`$_xTTMHGv^cJkR<+tNYT7hZRiaW3V%Y3XqA|z$*mYDdouhy_?-Q-;w
z`semX^@ZF{Yg+C*A)NP*;Vy}}ol^@JnXL`F`6X;!;kna-Z0YA0$;1}ktT%gdBHcS~
zPmlP<9}ANw<;Bhsf2@8@G5EEO{aTlK-{U^j?3n&M^5mXk#k%8N|CzGv^I|5-3A}eR
z6@I)eJ)R+blk$3o;>z44J7Avk=I-S?9`f%74?{p8Ll6%uj|G?j%7DR^tbeonqPCfG
zGx-Chx!Ab4zywUXC^fBk4<iEu<KOjn0{c_`F!BTmDhv9+3{|YqS+rlnP8ldHqAUUz
zc5Hvc@bgQ_EoK%J)ht0mdO`s(y(}nNV4{p%VCVH7`n_PyLu&<J^GW$)x;+zs7H}DI
z*}z;I;wPm3VxkQzP?(pKmmj9Nc#^bY=bU2;K*9>VW-wuxE>?)&U}ggq1@MNz6ahWv
zD;S922?Yv;B6uRcB5=7wA$Bk0<*y%S<_QvD5l})X5@HcT;t8_|!+AFua)ZyrX}<$H
z3^@p%8nAw0=NIn=xmT1|9%j;Cfv`<?^N&1Z1qz$<TET>YJR4qnB%UL$3lh(b*At27
z&Fcr_6{Y4Rg5tBdD7jc1#0UlvV89%d&7A+>YbS#TV=zN<Zc0uDP}n=omzjl?jh%y&
zn}=6`pU+d!)mcrD%hO3nTv$Xx(!$c(#@5c>K~GfMO4&?99-F9`w2Z8RqLPZLx~7(n
czJZ~!skxkql&+DBsF=G~2$P$iqX(l80EZEVH~;_u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm4d_fwd_f16_instance.cpp.4D368D0AFAADE23F.idx b/.cache/clangd/index/device_layernorm4d_fwd_f16_instance.cpp.4D368D0AFAADE23F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..737cacba4620bb6b8571c153d18add907a5c5868
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%p_(8hC>_-3>9<EhTPA;Z6<K-fAAbh
z{dw29J~%(RwR)qP;V!XCk&I6&g3CPKdffc@@0YCZ>byB(H#W7ub$xXA`0n5D#Elzx
z4oqMAHK<u7^!%z%FYiu0^Wg8F?Ma38#lH&!@8})B%zgBB`lq7W!o;2X{=IZEoExJt
zZ+FDqdcUsU-@9+?@Bj4h^!x6lzb_Lb@2m;mY^dG!=x|o-8vA=SadtJkwDaCYe71dX
zxaJ<ScHnO|SJ@@g4!&`D{oSLr<dR$yr_;}6Umf#w#G+N&!*Z;Ww;K4}^n0xRCO$v&
zY@3SgIXSs?#uZbpeVU$c{z7T~h0NflsnS~%J6+~os&6mp*jgAWG*7AUvz)houaDcY
znH-F{C)1xN9-dSv_(N`2!k*VxU!9#@p06lqEft%wz*RTa<P}ThamD?H-*gtWZ7>hK
zx?tm!xwbnm-VoA~l)ScLjf&(=y*Uo|3k4Ub-v9CAs;r&i@eHYyZh@7-7lS1f&h>51
zJhPNX(QNsI%xRCgCry3rVD+A__dU~>y7L{Td<!QY`D>y5-{sSW$tg?Bx&n_bTC&Rf
zt%YEg4@XRe25%Ja%4=>>hug$0mP^#yY^k_peXhrQ630RFtozH18dlDE-t=u(SXQlF
zJHt|wvH$_LC9>5<vpgl2IWh}a7tLAT)^_{wl_2-ykL7r`s=u@MQ5UdX5^9@e`L26c
z?fR8-Z(Hpz-!T12P?_ypM(#~6H%d}FSz;%M@;Qjl$dF2ZSrY$o;jHs{E;&)Bg)eJ=
z*Vw+HJ2N_HqF?3GSsO(@^z+SqB>3ssi>FhV_;qDJDp%UM7>S<T=BH<M{G`nP=XEa^
z`)GvDRod9SrM2I7F;B&Z(_gdd{^<Rf9aUOoeAD}u>5Yi>+jw~2Br0E>wtVu*$jdV<
zcAq+se&N`x#T&fe&3WGun46)t%i*Qq3_q6l9jlL2810h!IQi7y*-Q8Ropdnx+ph09
z712zZ6U*)eKa76!eS?9X%Sl#|(A%8T@A0g2<DXjnw_o{RcZAd9a~e8pY&Ph-savh=
zF0%QXWf_z^n{%F-W1Q)S&o8P<RLzcc-rC}o`q_HtQB{3+i8*^aipq|r{@<<WZWsOe
zQR<$Kqx`ayOKv9x#n^A=n|bTm&s5b<d%jP4et56Vd9P=4xj)vg`BCri_IR)MCFVzW
zSEMD~`SwP?DNOlbab<3j9Wcjj$@(|DFKU}9H<N!LLl74m7Z;cS%7DS$#EI<2zkRKE
z7y^LOJghtxU;-vxl$ut&hY^@}*LO;lu6uQul_!YToYxAjD~Q*I*B*)I$m@c{bK~_y
z;(7D>!FZ1CZy0`lDY?bW5+tZB=mQgG2@=v13PA8glttijPOB2+zW?}mo{=XAS=iTn
zQofjO&jg?_mm!x8%(o$aLh3Ij+OPtJc{zFcVMZ2Sbt%qxvBMT9jBHc!Bx%LYImZ@&
zR4eeB!BoQxV}%4+@1frd);zRU04fUL4S^}*2@>=b3`FpR0);{mJP}_JxZM7=A7b`#
zO6dS?LH6{a5WAQ0^4E_u^8^X72q+=+3b6<w@q}4~;XIfcMlNt<!5sF~fb|PIzj!yu
z8=}1OFjYmVIf<b7EG|ke76&ncK?E2u2Tf#LQZ0~Soy-`_ker*6lK~X=PV;4EVP#|I
z;N<4v72xOd6m)e~6Xf!A5)u~{k&v{ow6?Levv<%F)wWVL(~!p|Dkd!>tDvZ)qN=W`
erK4|PXl!aOXCkF*<RU8O?iIr1=I7|a=mP+=W`vLc

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_layernorm4d_fwd_f32_instance.cpp.82CA4512D672E7A2.idx b/.cache/clangd/index/device_layernorm4d_fwd_f32_instance.cpp.82CA4512D672E7A2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..be7f2f597be2b41f80c0f7f42418ec1e1c8cb9ae
GIT binary patch
literal 1572
zcmWIYbaRtoV_<Mj@vO*AElFfyU|<jg;^LB`%p_(8hC>_-3>9<EhTJZ?ZNYQ(e@Viz
zt>(G5kK2!I-99f@$X|c5JFlyg*m04SteV#Eu8VJ3v#apdOtab>wG%mHm@IY}GM`|m
ziHy=W^$5CJ5^J|-=LE?;bN^00pTB<H{qvt+AN}yj=i{%5dm`q>MyzquKQA)<vGLJl
z@zb~cJqv%A)&2dMZ&&lr;(WOLJF%Zvvn*vdDV@Ks8~*dv->b`y-ff@#=T)1jJ=49j
zX|fwcrmhbRam<oD6J}D^KE+%8t=$ZTDeI->QxXmDD_Cw>l+k<jkVR^A@A(b&`<9xr
zdfu9EbEns6$MkETrsrROp)~(OW^mI~)>{)?l>A@rcfED|)Q(aiIi<qS@4WrD`nVmN
z$-$U=GX43&gp)fsYWQOf?v-YJeROp8^$8uaJ?Bg=ESRPzRob%W@r3)5)nb9pH)gKT
zx)5>dn%v!t9FD0yJlP>(lX`N*`xfNy<hU?leeIuTeDX7rO-whrc3fHYLi3A2`th4r
zp2f5YyiPu0a{43Z$*tLrR`0)U&1WvD^Y1X_TR82=UklCs3Y9m!jDn<(u1E^Jw9EIc
zh0rQrj+hD!-YDLc*W98Gx20Pwm#DpCUU9|xT#xr8j)Ub{_m>zoL|UI`eOuCd<yQ$G
zqgQ6mdIcVj=x-?|lZ||us}#09vGC_L-QK=(X&+y^tdy4jo%$Uz&R-^keO>T$hySZ@
zi<fQ7d;I6zLBC1fH%qrv>zoi;m~mQ&^OnN(8_c;0i!$ddy&QY=cb45Q!Hs!3{yxj^
zEWC5TY+0_StJSlqS;vA7)JK<exu2NbJUyAqW=qKd{wF^LPOkAZ3oW{6<@fA^{a0`K
z6sMJ|L^@-a=<azhk$Rvxe%-Xqhs)31zq0D_?2~bq);T8by=8D^bGy?jUVrt;o0pjt
zR!wX?zo1jpJ#qRS%lkr$Z(Ru4(s)U^z*20Fa-MG6<AX;V=7+qyZ}RnRC>#IH&-XU}
z%jI0+a;wsRXa0uUhwtnVP!$hYGdpa@m$U*?s}*-%-tYUtzOm)l+Q1pnJJaKP`bvYI
zKYj7k{B&d1ET0csI6tjw+t0T6V5>%t@YKvHr_#)enz`fpE?8F!J$=%3`rm3rcDv}0
zk5cz^{PdTdTr#_6PQP4ayHBp#<j-C{YICQbT(aMMW{T?Pc^n_>*Zin=c>BEf^kwEp
zcNZij-TC%LyKI*7&f?14Bs*Y^t2ui5$I%t8Yk3#~0vUpMSa~eK1W*PHwq*UA-50gZ
zl$*&PD9y#j#RVo{(nYCh#d{bT7#RPqzZ2M>@`sTpNKjeO2WF^Zh0db=8g|M+VG(5!
zxUggU8-|}>N^UW;pr~dE64Db2fazsH(c)`9DPK&tX9Cb9E<<3XLY)=jC#3#jq75rh
zn3t26AEvi>lC)yyoMQ_>!V0`*Fku+Y$OZOe@1frd);zRUfEmUL@f%DSs5gK&1f~b*
zF<-$z1WzbXC=|gH@fCr~9SX5~883hRI5SU>0E>VULXi-Q5E4(AMHtSz$&ed-CQkbu
z(7nh(@YI0y3p>AfH^>{Jyz(%U{tAR`x|@IG87olOoYx8_4CLAH+9UBCd0mirZoHmI
zJa1k<7_TTbClM5%#YM@*;vhyahyVlTpyyNm3C8#KKVl4KNX|{k$p8v_r};9ou(Gjp
zaB}nT3h?uJ3c5P0337Ql35g4fNJv^(THDy#**oZoYFjCrX~<&}6_b{cRZvt?QB~K}
e($P0CG&VJtGm+9YauF4C_X=Th^K<lI^Z@`?ceE@3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp.59F686652075C1F3.idx b/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp.59F686652075C1F3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..93072285a19567fa72d2f4e9f773acc4bd9aef65
GIT binary patch
literal 1838
zcmWIYbaPW<XJBwn@vO*AElFfyU|<jg;^LB`%nW7*hHD%Q3>9<EhTP7(Z6R>=zew`Y
z*=x3&K4?F(b^9^1&WX7b-J6u$#S*)&e0p@c?k}INrr(r=#%Bv(bRU`daE@hhGTQ{k
z8=J07^M2qLve~!tN6A6sA2$EHb>FSa-=}xuY>|zmhTXmi(_bHKefsOu_Qs?l(||oK
z5qE37yLO-Np8dZ5(!<~PxwqBrh~M=3YnGWBYjS*8_O^dje{(DU7o1*~@A-M!!_zhQ
zq{1pwv(IRsSN)zf-Mp;7uki5Ki`w(Kqs^~<c2d5*$auwPB`w|J)RsP>`w#!u`LE?V
ze)nTfLH*2w%GYP@E7|QP{G!Eem5_;kz;6|+^w?kTFUUGCJ)_3i{$NFA|I$)Zf4@^J
z801Pl(~}n}c+O-f;qMK<wk7N1!^6w_Cv?o6!zPsxP$WF%h0nEs6Y1xsxlDQRU8{pj
zWtz<^KDVUD0UsRFI-QD-GynW~E<;(AQ7S!KdftN%4l>QFCfd{#1hTxY$zSwFV|Em8
z>W78zoR&rMt@p_DVaZ!#dCjaUqwu`gS{|pfvR@ys`6E8JRkpm#TlIvN$CEtaU~4O<
zt)J#`@h^4KRlFKixb{S0!g7&j-s5c6dl#Nrq0%W}6MS23D&v)75n36mvv*8+zWIAV
z(bDQx=l(o*4=*82SCg20j*F6~f}MQdDrT)*{J1dt>(tUK^)q4|H%tiox@75&^{>8#
z1uwoSwcC8d^K(nyMBUnZ_OsB!j?_-RMbD+*@Lm+YXmak(iEWQ+P2=xyM3~P~TNe6m
zVt%6drEN=;N-LMnn)u}7{0_GZ9Fe6?FM}2=ELH#AX(QVdEPZ^omZOXF$vOX?*PWdF
zW<s}?t5V07)_&c^JQ*KOf6c1<rT1cX)YZK+TPBt9a^)VK=3`r}vE#<25+TdAS)ys;
zk~er}1q6yTEjf4d#R;vcv!7nv{e{D4UC}MU%_onCFYXq!U3q1H<g45Ak}{jV?@j-E
z`_bGJm(N8X&uM-uXOvp=Aml*JWf!I`zk)Uu>DOCascw2ARPLp)`e)<Zt!GYT=|=?i
zsr&aSUo?sI3!Y}Jo%N%(#B$dO?$l3SrcNoJqxT(E)op#RIRBt+V!+&boku6q_%#+E
z_?@O|e9P&Z)L-84|NFy5dG9B_P}p|)<%X=wncRkRzkOeJdCd})<(2=q%buqjc^?<w
zoBA?vV_Jxypj77@d+8M#<;9h`NwL6u$F28&=TC)`^1KZGK@7oMVq6Mf0w@CpZAaV}
zEXc6;;b8~>O7pPt=z|HE^rJb~d5^q(-ps=g2$U9O6*UDDFlmOO)U@ISObiUn4p~cu
z4t}g><O$|7=dyxX3Zq%Mz?SwN`n_PyLu&<|U_NiYP`K5>LOw!a2%fmNcqrW7U}+!e
zFc?pBMeMw9(>5?N0S%DgQ$^AuA+8EjWOo0A?}?v(53&Fi+44Cd6!E$6c_Mh?w&G4O
zEj+>EF5;dro@4tPhM!+bZZWY03keIEz_hRgOAAYzz!Zh}38}xBXu}G03NI%wKTHu%
zuo$NpKTJ_`f|T28L*p_opi^}C3=oR=jQGqDJRwsdOC+APkUbL5O~?zu6W0+pfEf;S
zw2`<Of+uY%ZHdIQmbORYxk-D$c&&3iE_CMJngVn=viEpUya%(Akqew0U=g=-;fh+T
zY4_DZVQax>4O0a)$x+A!!4tO-w}x2<G}lqu1;#rRV)rs${`zsC<$^4NN(e1dEK*7^
zMKBs@IVi=!EPrai`h}feyc-nzynNyaz2dy$;xIigv#&IU%!soAB}8>DQ&<8lO3g_G
zCEDVm<YI9k%^1w_eA3;`jBO5#Aq*k#<jjzqo05|O<a($1GPAI<v2$>8^Y9Ar^LYxo
zI;#nCc{&M+3yVldT3A}!*xK1U=!t4uDVu4?V-ppVmXTFZR8mn@*VNL{H!w6dHJ3Ay
U(lv4s6?69rWpeX#^kDP>003<RLI3~&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_f16_instance.cpp.075B2EC4AC466F2A.idx b/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_f16_instance.cpp.075B2EC4AC466F2A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ade6870ab94c81397fffa6bf7a346ca232b5f91b
GIT binary patch
literal 1838
zcmWIYbaPW<XJBwn@vO*AElFfyU|<jg;^LB`%nW7*hKn2w3>9<EhTPA;Z6<K-fAAbp
z{h8OLJ~%(RwR)qP;U=+4m5fg*g3CPKdfYnq@0YBuw%?p2htCFYTp!*2xbOEnaqb4D
z1L0G@1~sdM`d_>H?{cfogU?_14R`O}_tQ$){PV*XJ|90$wTYh>J40W;-ZHpn-8qr<
zielyWH$8gy`OUN6?`wYid+wtn|L*yts+}`<A{CEJKN_U7KdyfEefv9IucNR1Iq@OY
z>Mc+Arf<B{N+(3IZPtz6?<{o5^S6kP#?<^M`6s6q-B7UH@>EN5wfEyGzEbxe{;%`T
z4pVvgvZbIt^1$LP)ArZwb`<~F;y0UXP0T{xNR>}h{zora9wEK`+#x$5^Y<G+T{-b4
zWlpESg|M13xiB3`@y7QHxsUDkTWeSI<BOot#^%d2*`lsoa$Ve~_TJ(NclvpyEDmeo
zP+lqRnQsK$lD$`?q-=IeRXus4?qSB4&Ta<XpT~};Tih`|o*|XeFQB>dh2|HvbA6jL
z&+Oz;G&??_a@u3*N!imKtlrO)n$J|SL%rLSZ{frve=RisyM5X)HD!re*V-;u&(~9K
z7ph#DA+X~^faDgbqP4wS+GNkanUnP8<I1Cp?-i!SO!DD~TTwgvG(*~38}Tidw+6kw
zY{{}>$;Q1+1`4^7Co|MiXRv*3n{lFWo|IG`fADEpSw8FR6}E-{6H8bxI>%jYp7wa(
z%G>?3X5BgVb8hoomFb&bXMEqLpxqL>N=wDVaotAqB?nwSTX-+u-Sz!yO{sFy<*0cx
z{B}C;Y|K4#ds27F8SU#mOB?D{OS;@o%w|4i$SN0Qb%X!OPl1zbJhzp$dGt-T{Nw-k
z68p@6RaWjv9eY~)WfwD5d|3T7sp`kMAG4!NYm6mT-%G7o6E1D`^-ZGc)wbo6Pexvz
zVX;%=fcS-Dvlbhul%G>iT5V>c8NA>RN6$L1dxG2b+QXZ>8SPiz+3)%8c0TKy?(cik
z|K1kzm|FSVmh*nn_Vz<N1XRU0#BA?m+`IS6qFXQL^L?GCu=Lp4Kn<y#;_7P)cHPqq
z<aIw^$9#30>q+OZN#VSItaix6OwoRP>ed!7(aX_04_dx=c_A3Tz&5#HX1&&f1LFKE
zCLj7;wvhXt<43VSYtw&RpWUXOUtboVd+hP&uHNIahbr#em+tTGGM1VC^#6L}|Ee2n
zYOaP)yf6Ho%~bedp~gSPU{AH);>z5lSYW=BoEpOsa^;RF4?{o@Log33j~<u+%78(K
z?6<RB7s}W2GWY|fxx}~>zywVC)sr~IqBB*yco+hK(xR-QCSU?4%}|t@R=j|Tfq~f}
zYpKw|kM)c^!CdBCRxncq*LO;lu6uQuizk>*hfg2wmS8?ZJ`)5_$W+JziDxBbi^Ovg
za!2sQb;R}IZVDDR6gNTeq)nwQka$+own#h|X?GY8W*Q?GI3#)x{a&!<p|yhMir9JI
zrfpzk0=iX#PZg$$Cs<rUTotC*?EVSg6F>hRWC1F&=5s(O;&b9dcBZws1569h04H&{
z<&N!d7=C^!xy8g1EF>%h4+EB9X<=!&UqbwZ)L%@rVFfycmy?$tVT%~27(dM1*0~-R
zI&*JL0V+Z^i3i0bSVXXLfuj}XSD@=X`GOJV3V8}4@x(pFgJHG+jq{X-^L8#=QEN5r
zzB<VJW_*@N2G|QZA$a0u;+9AT*h@RXc!xslUdGE`KMr)XAd8?9LW>lOloCwQQv=p7
z?EK>0pqSw06Gtc#=M@)+DSDZGr7>hioDC?&sdJgaJPMNrMlL9|!aQ7*nv)1hw8cfq
z#o|DkG5G9`Q${y?pVTsjFoeL9GedH2N=^ol>z(Gy%)-jX&cVse!z;kg=PBswtR~3i
z=_DjBEFvLkVQFn+YiIAEC#r3wY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw
P%-t)L$<5EvgV6^7gIn}&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_f32_instance.cpp.2CC50FB6EE948F2B.idx b/.cache/clangd/index/device_max_pool3d_fwd_ndhwc_f32_instance.cpp.2CC50FB6EE948F2B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4516486081b190fe14b65e1692da4eea13e13990
GIT binary patch
literal 1838
zcmWIYbaPW<XJBwn@vO*AElFfyU|<jg;^LB`%nW7*hKn2w3>9<EhTP7(Z6R>=zew`Y
zQnTAqADkcEicX$=<ixfU%8rwjI}Kd3DnF*~-{<VzCA>6ZL8ZnorNT6e^mBJ+Fs3kU
z)7bJ<Xov9CH9CCq`W0;RpVe<ZdiU$u>8!gW`_{8{@#{Z%T2+v^bKV|#jdgnLUDE|t
z?_Td&_&u}m^ILtZn&0UMuknBPF54S-_SlkyH=nc4UW;G1U+=wq-p#MNnRO{YZ02uo
z3yr%e_4!os$-TFp_TF8be5auH%c<vwuJwMcSv)cKqU4pIYFfI*sV#j%_aFQ(^IvP$
zxBIc=!@5Zam9Nj*SF+np_(hA`DkhWr3%+@loZI!~{zB_EubD=i?GIK|_P3Ur`um+)
z!5~-an4X;IJjsahg`Cv-HJMi*-QC?UuOMh`$!l_9!4ox&i!;|OoN#`QZl}h<_n`u!
z9y%3QWP6SrTlk>yj8N;dKK4(a=U#NvVlp{*&CKTD!$u3XQ02-W4;FD;`+3jhW}w+t
zsZ$S}@3#7EmWi9F<I9q_#`2n3RYu`?v9&x-XJx-WUh{{4ZmVp0m$&K(EsrPbdREHH
zE{v-5Z*6y4IBi1K+8tp}EH(sp9c)W(o_jAKEo71l$GnxflTR_MGMUD8>y?z=)1Pl_
zMPgpYJ-o2^w&0{mE?$DX*QK3G+CEJQ2(IpkdiA2>PE>8G>{t19dWsPqtE(;r-3c%K
z9T>d$rqpip4bRUlc@uSO@7d2xiNf7XvM%Sn-|$`(zG!mp&WUZ0YE9$sa738TQd<`K
zZeo6-_@!-2lu9d?&YJk-<NXe|3mlQ9PA`KNEG$+3-DxA+6uk7Y?^K0F3se057w>=K
zS$4v8s*<3~twYDB1+tm^I9g?DUn~B_ckQctK2Dyy*`#xiPV=#?zEQAo=}VQOsH<9M
z^o%x0nk-zT!Qwf0^TmmwTIQ!0m%kL;vFl8xyU)yieRpx?s-Ttio3CuQH?~O0uRi~4
z`%&wO%ja(IyUBLjHs#ch11lRUZ@MsTsZxsEINe_Mm9Aq^r?u*YtA7rhi%Lt`D!yjM
zp88MHoy$%ZNo~3l8F*NJ@yQoT_cbPmuM$+6egBiLUZ#3Y{|}zGoXw~H1T(suRXa$s
z-ZR&pd693^^#|*1|Ng%hVp7?Duy*5Izjo2te$g#2x9m5co5q@IarX25^~U=rMOf{7
zq3!*?({yoH#|4X&`;0D;YR$!!xk<6Wd}m$tXWp?bVP|+40)iNVd02V$zyweR3@!&I
z%NxEu(#^vV2$U9O6*U19FzMI53(m^2`sDF4_yeW6#JCi|1WcNtC^fBk0TTlQvqRQW
zp@Sdm8F_-a%(<*!+B8?h&igiP10xepFrNgUD%>r>;u7MjFsse(pYT2L^Y1|xpdxEN
z2ZSO%Cq8!sPuyDE0j7l~Slmh69ma#HVdMgPt@qIH1#2E!D>$~lVfgu_<Q5Z4u#m8j
z2}~7Bu(YtW2~2N@pOE^Ci8ic27xHrQ^1~GI1dDNs@xv6g&h@y^nR{yrP!Y08JSZmZ
zT)3jvYTA8ukT=ZuED<IN*$X)#c;aT_mN0XH2G~nG!FY#4>|VypUq24CMUX{M386)b
zMM?>#2u1__3W_Y4U!NMVeqrYq?*;`lFP}I<uQ;!`I84vW>?@5SGvaJOk)+OL3X3Ed
z&B_IiTbPkRlRWu?5he+F3L){tJ;j4z<^sLpDGlfS6$smOH~+{pE}#KAeEJ9j_zd~r
zkpK)xQy~k4oRF1}EfUW~2sz?)#PtzYi5rT;BM)e?sk8+`PTES^7K!H~4NoFPsX2+D
zL|a^xTr3V^gn$SzU<{s@{39`O&yq~W5QgO3l$;Epuy>j-GYcylI|nB>53c|}pQoU!
zvzj26r<0Jlu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!
TT_YD!F?X*}CO1Dv4@MsVM0oxi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool_bwd.hpp.67E96E50742F2EC4.idx b/.cache/clangd/index/device_max_pool_bwd.hpp.67E96E50742F2EC4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e869eceeb8b6b03211d62c539f15d1e1a3f42b9
GIT binary patch
literal 844
zcmWIYbaQiIW?*nm@vO*AElFfyU|<jg;^LB`OcO>1hAB)83>9<sPV~+`Y#`9`H@L(j
zzi;{yb3rEFg%g}J`1VS>-n(^ZTiA<4^_G9%Pi4*$Z@kCEr2BcJvGSBNZta}ktDB>?
zoR}tmP`_1te&7$kn=cIZZkVc9P|h5`UM&1q&$E+n7ABsOtB!pcZuU8)bMB-aV!L?*
zH^1n9GFQjAFl<*!?b4K^O^2IS9`wnPPD#1Iwcca-nbwZh<mF$dteq;?rgQqj=@i|p
zn$4*hfxme*-(Hm4I`!QOx8Ld~XBGF~nt1m7y*c8?lji(ye*gP;tNv95_J*#gmQ$Th
z7FQb=%H`^9$@u1HIpK;_(c$J#5w6M1UZ>SWI5xHYw$SixxPRpG)cn@NYmA;*X0E;Q
z^zia0xA*PvRDFCfEWpe1+eGUxk57u|Ol*JfM)hgxt`2sk*8x?VPVyVDe(Wi(%uR|1
zhM|I%ulMOXk5o1$W_AWHes+FwFu}md#0U&#MlMF{u=bKa+rp&;m{>W0>Vbeuh($=8
z0m`Xb9`;0Oe-D!Y6B|rgT2tB@Ce2WknpSKGj3bymj9g&zJypuOC+>2tgPG3A2{sj`
zD#TAn{l!EZR%RY9aanO?m?|DF30Vmwo}{cKoEN*@GABk&MVXO@OHf8o1EwXlk5#on
zsMrZ8%&*9=3$p;`Y@od$=fZ@$y_j3x_W!qF<>3;Q5S51+!owvdA%?^g*AUl7$Vq5O
zAn_zMB;mZZ##RAG=O#=CS}rXktq!yFsR8R3U_f^>0i7x+D2*^lT2NXVrU)2vIf+WZ
y;D))4q3NOjugCMk*kE3Q%M=$S7b}BfCOLociO6&HElk{;44fbl2;hb>7#IMF@+NEm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool_bwd_bf16_instance.cpp.142DD51C4CE34BA2.idx b/.cache/clangd/index/device_max_pool_bwd_bf16_instance.cpp.142DD51C4CE34BA2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..975b9b121b3fae626898c1ca91874df53e7746ef
GIT binary patch
literal 1344
zcmWIYbaS&{Wngel@vO*AElFfyU|<jg;^LB`%m8KvhHef9hKf07L%fS`n+ZJoe{xO2
z)pfH9+}NWhO8F*7PgZ-Z%+4}R$xY#ao`dbL*P8`YPluIc)Sh_al(hfoySm>_Y!euN
ztlAWpHDSrCkDG4Q?U*E4Gv`mJk9^*HJOA$8Hw#MG_SCDM7fTmU>wbOzT*Jk)k{@mz
z{qr!JNzTsi{l32^+w<%FyN^yk9$C11Mt?hx$kF9L7e&<U`)gJGrzrl;J>w$JkC{d<
zMT-M}v$@JHnRalE>uYn5){;x_nmC<)F8k^@O{`mVlC$=ya}l~d!P_qXVEes2Ix2Y5
zlDj7lKQ3kSzZCWF(_6OKC&E^_S{pXGNKf*3qWSl_$L<K}?LLRhgv{Ts{B-3+O3Iv0
zfeT?ZWpZBAdb$sMzrgx9mVb5md%HdDE;;!nZ4TX0yE3jgK9y~+NqiIO!kd`9Xq98K
zs&#e2;*F{yMw_A*Z<>^nvb!MQS6?@S?$2Y#)5~|H_FXhNA*L9#e9`iPkixP#zRBlO
zIW}6h_-vM`SJ+u1>+$^Vx{t99_uhLZ_c~{&6x#be*zX~CZBow_w~b*NH8O85EtBCe
zRb$b!31M5?7Lu*JHu12r?Q)6QJ6kIb$fh4t^=x^-`sQ{Bn}PT9oeONQd9C$-F0f$I
zF*(LV4O`Ou-HbI4o4i=jA!~Uq*!l3=#H^Kz4=1kY-Kzf1z9&zi%478|cguJDrMs_O
zIk$GseTy3&MVh<keG3#%X$>d~ZDom_;BBL|>wwE;3-9H>yX05@eC2jz*;X4v+Z~QO
zn@W9dPwFZ;p?$SyX~X}L7b5)=%&Ya=MJh5f*WaJ~+*xI_%FSyhU7uUb{QLik{@&@F
zOIHP}lug|BVP*OhM*h03GyK2Z*Q^g-&6lrM?YWz6>DR2ItCXuXmPUOnF|y2-5<VU-
z`Ge<HM6pQO$+@3pH-<?|b?XIGvACHxl^@YP`XQ~r^nv=+-{yX`zm=PFzis>e=EpXc
zAf>l^mLJ-F<9%XcOvfZ{jcK{+Qag2{-scA9efiJ#ZJUA9<69a!aY?R`yzb}Kn6GY|
z`G8aSNhtGv{&no(i%k5kEDP0|edo03xeCQYuHon7cAWaJd+k8_o`Rz<-#uNe7`)oX
z-rwYW^==#cd#69INh;eZus_7HUT=A~uI4502X_~wS=~AP##&vEcWrTHZjwAOf9=sX
z(OC6^Es%%7KY$^Si<L_YOaNsdpeQx1n4OV<K|#yc`*fX0DkD!Iw*t2TOq*l-8-|}>
zN^UW-1oG<fdclN4{DjnBOtfJIspjP7hY2?)NV%;xG%jNS3LA2p!i0f58*T>#kJp>m
zABh*n8wKO-T)3jvYTA8uCZHB`ZY!8DPatm~ZwQPBqZzrtUg<sbd%>EA)(U&)NaR2M
zu;4e)O~_6?6k_)>UjF)VkT46c5=`?`1J*C>{NmjpABb|x!-N@%QgaePAyZtGTr3V^
z1c3-JU<|w&`S1QzMKL|bAco}Jl$;Epuy>j-GYcylI|nB>53c|}pQoU!vzj26r<0Jl
zu!w}Dg{8HPt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X+E
LCO1Dv4@MsVMD`~F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool_bwd_f16_instance.cpp.DCBDF5FDAE579133.idx b/.cache/clangd/index/device_max_pool_bwd_f16_instance.cpp.DCBDF5FDAE579133.idx
new file mode 100755
index 0000000000000000000000000000000000000000..64d336ead6583e588687554c3b001816b478d591
GIT binary patch
literal 1344
zcmWIYbaS&{Wngel@vO*AElFfyU|<jg;^LB`On+tuhIS4HhKf07L%fS`n+ZJoe{xO2
z)pfH9+}NWhO8F*7PgZ-Z%+4}R$xY#ao`dbL*P8`YPluIc)Sh_al(hfoySm>_Y!euN
ztlAWpHDSrCkDG4Q?U*E4Gv`mJk9^*HJOA$8Hw#MG_SCDM7fTmU>wbOzT*Jk)k{@mz
z{qr!JNzTsi{l32^+w<%FyN^yk9$C11Mt?hx$kF9L7e&<U`)gJGrzrl;J>w$JkC{d<
zMT-M}v$@JHnRalE>uYn5){;x_nmC<)F8k^@O{`mVlC$=ya}l~d!P_qXVEes2Ix2Y5
zlDj7lKQ3kSzZCWF(_6OKC&E^_S{pXGNKf*3qWSl_$L<K}?LLRhgv{Ts{B-3+O3Iv0
zfeT?ZWpZBAdb$sMzrgx9mVb5md%HdDE;;!nZ4TX0yE3jgK9y~+NqiIO!kd`9Xq98K
zs&#e2;*F{yMw_A*Z<>^nvb!MQS6?@S?$2Y#)5~|H_FXhNA*L9#e9`iPkixP#zRBlO
zIW}6h_-vM`SJ+u1>+$^Vx{t99_uhLZ_c~{&6x#be*zX~CZBow_w~b*NH8O85EtBCe
zRb$b!31M5?7Lu*JHu12r?Q)6QJ6kIb$fh4t^=x^-`sQ{Bn}PT9oeONQd9C$-F0f$I
zF*(LV4O`Ou-HbI4o4i=jA!~Uq*!l3=#H^Kz4=1kY-Kzf1z9&zi%478|cguJDrMs_O
zIk$GseTy3&MVh<keG3#%X$>d~ZDom_;BBL|>wwE;3-9H>yX05@eC2jz*;X4v+Z~QO
zn@W9dPwFZ;p?$SyX~X}L7b5)=%&Ya=MJh5f*WaJ~+*xI_%FSyhU7uUb{QLik{@&@F
zOIHP}lug|BVP*OhM*h03GyK2Z*Q^g-&6lrM?YWz6>DR2ItCXuXmPUOnF|y2-5<VU-
z`Ge<HM6pQO$+@3pH-<?|b?XIGvACHxl^@YP`XQ~r^nv=+-{yX`zm=PFzis>e=EpXc
zAf>l^mLJ-F<9%XcOvfZ{jcK{+Qag2{-scA9efiJ#ZJUA9<69a!aY?SGGPcn?A-B_-
z*nK91^Zv2hAs5qm$s%xe*vgzD{cwxNwH)HMw$%@n{x1(;p8s)|=<?O#`E5RHKmD;e
zZS(%@r#}_?=QnrleCqVa>*}AKQ}n(q_E_Ig-aPANf%4|}pWpJRG89+lCdmWym(mZv
zmc=uIIC&WS0~i9iSh+O81W*P7ic-^x*%^U(aeb#$>AF{kS$G1u4Y^I=+5)+)x$O`<
zUT<DMBwi?Q1dOMk<?DUA&Lfo(Xn+E@0ZbT1Gjf6L?LG8+!J3EG3Xbh>7=C^!xy8s5
z$g9We1ydE`C#3#jq75s^5KeA>nDEYpD{8H#-B)J<3Y&3T!i0GOc>{QZVZ6O_B=R4B
zSnwOD7um=|A$Bk0<*y$H3A6Aj!BjsrVEw|*FWwEZNt9b2Cd^Qjnv)0$nc|}4VsRkN
z7&uXExBNHu=_-st3_*}E1ChzODLENH+B?mcnT3^&or9B`hgX1~&r{IVSxu13(@98N
zSVTh7!qVEt*3RBRPgL7V*-S$oo2ZzyjI4s9l8UOjrk0MrfuXUfxtxiVu91tVn7dan
LlbfHT2cr)FxcV!t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool_bwd_f32_instance.cpp.EA3D33FB1230F655.idx b/.cache/clangd/index/device_max_pool_bwd_f32_instance.cpp.EA3D33FB1230F655.idx
new file mode 100755
index 0000000000000000000000000000000000000000..002db4dc9de642d868d97b182163ef612d0260fb
GIT binary patch
literal 1344
zcmWIYbaS&{Wngel@vO*AElFfyU|<jg;^LB`On+tuhIS4HhKf07L%fS`Tku@_|MEc3
z?3`Ow$J&o<-9FE(bNOw*KB=x2oj%PVR!#d~uf?~lu_&y)adzS5ZkHN`27diC2FV6}
z@7GJ&JXcBSpMKlUJ@er4lh13a>WhD;&boiD?YjDb-{PNg3X2lA?Wn0{iiw}gpdWkY
z{eIO4{~mw&_WOR#kALS);%jPWJT>qBS#iQ)Ui~cnPe-TE&fna=zWnD|S>@*<^VKH%
z8H!I`9~k18C3z;wq^^02xB6SX846R@OUtJu8pbPHZdsJkd-ae-YIN`V4fT7MnzDM{
zlC`yoN&BF&_No5;br)Uj7G7S)qRqWUvD0PVrTX42-BSxstK=vbewOq0-|FLbY$gX|
z?#cA$3lmQ6;Hcq`HMm#0q3ZLq!_^)~zPYJ$n5<aZ?Is&?ukb|rv3agn1kBV{ahXip
zw3XdlO4><Pd9|sQwx`=RZuX`5J2@^)SYKOH)E@mz!nZd^P-&}H71!Rb&5v`kinW|O
zzAC27disO=<kD<MtM_GF^O;NP`lYgEoOC;X7qZkVRYrIl1xX)`I=W~{>g-#E3L&!v
zc6<ns+9DOSws%V#um6`hNnZ-DKVU4MBRO3~uqk~*evsq=ow?7O$|97v+Rbxr2$H(Z
z+`^c7c3w|fP+Mle`VL*obHUDM-zH|QTzoijIqy{UclI&yj$bB(mG&3iiGNkLZspo&
z%l(BLCO!!&lYJW}p3)hxCAE_!cY^mD(cFYZnRAw2jy?K2Yu^sRjd?o$KJ0k`>k?!y
zZ(E`y`^n2RS@S`^*xW~gpVHnxoxsGe`}S+~r<#tGX)3d|c4o+`Z~ia;d+B$hg(0C`
zNAxa<*_HJeJvgW@{`urT>!<lyS0A2D(a*fTU_;#6h^w31omPqZt54p%%&f3xV&nM*
zouclE)9+Z`7s}qcaAOAZQuhbHygu~pSSk_6tH<_dNzwmfEA!tQ-AG$seg4;W)d@nG
z&uyRXJDA5`{MczqjKjv9sKC>?_bT!~CoTS8zTtYSpwxA@#D6lUPoLiIFUEY;Waa}#
z;U}TI^}KrQ;)@LYLY4(;&AxL|^jwAfp-6xKxE&|{YhF2!z9;YJ$5%IJD+aH=_`>A$
z#TCZVy}zm#2ZkAKpD0itWLmGkbo1X6m$e_<U65pT=fs=%mx4U948@hXN%Fw__4xB_
z2Gy|iNjwbx0Stj$tX!I40w@CkMX71U?2HTy3R=G2r|UdY8F>P^6}Syx+8o>8F#P;d
za*L5AkXMh_3nm=mC#3#jq75rZH77SeOnB$Q6}48=?yEBah0VAvVZuCtyaBwyFy7ud
z68VomEcgx7i)`ef5WAQ0^4E`pgjsl%V5(s>BNy1+y@!4;So6?Y;i&=Z7j}N}Zjb{+
zx#eN1{tAR`x|@IG84FO@klO?%4CGmJ+aY+o-n@QDyinc<7>}VSH75}iGQ~y7#o|Dk
zG4PwhTZ>omyALx4F$6)v3`8d9rsQM*Y40>&W)@a9b`DN%9$o={K2JecXEi}CPbVR9
zVG#*Q3rlMoTRVFPJyC5dWit(VY@%Y)GO`MaN-C=Anp!&g28PC_=5i)dx<)RdV(wnS
LOm2RT9*jN!IKwT-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_max_pool_bwd_impl.hpp.85737AA1C70FEFBE.idx b/.cache/clangd/index/device_max_pool_bwd_impl.hpp.85737AA1C70FEFBE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7d0381ceed1c5fe0e5589d825ae0de57abb4e17f
GIT binary patch
literal 7006
zcmc&(dt6P~9zSbIa?b8_;#8+Pr^h~}qsOVvN$;Wum7X&4sPvE#cQBQaJVImWVTf?i
zgo~*e)F7d#uIQn0hl>oCJBDGdLFwk1%$>W==4b8e{^3u3>bLiI|JM4gwSMdOTC4bo
zapMv!1%i-_h(&W}E=bc82n5FDFK0pa><BCnC=CRHMPD^ex*2=jP4aB##Uao1A9p^J
zy6+8kxI84ZPV>wyddr_~4%@$OZhu(Fnin^7V}@r=sEar-w)pmorh%T?^Zye0_$`*1
zX5Ia3uZ8|E<<fTlzC;gyE2DK*F*^?xZTPf#aQ+ROm~*C2zLk{hj@#D4#=G|@3NwF?
z`h9hKPWI7>eTxnz&OO~^yTJcNNt=&(zosjdy|(M%gyN$K>2`N7pSCwDY>6Lw{MY9b
z|MYaM(xL9s9pUZV2WHy#!FMw6bUI!8=0g0Ykvo6AByG;@(Ci896@Q|?Au6zRVUf9_
z(_L@n<FAufIFz|X{xH=#DA48QcGq1yqpCH=zaMrP-k5zna^XKM8f+TRZ~8vk%yUzw
z@Yc~j&&?+N+nN54kEbM0tJSkUo0NE{BQe8nmbPGMN3Ga@edqSwrE~L(Z(H})cr<Oi
zJbz&G6SZx}mQFXr%C#nsPo{<bLwby9-n6<Pe%dvQOq1hxf?U3>HC)tY_~ZUv*yQL!
z&)$0d)oF&A`_^8Fz$T)M)gAF4O=~qt{;tGl{yzsg>zA!bIUV&lBlI(am6CwohMheT
zfAI^vW@2ykb7a}m`KngO3sWKnCN13;J+0yL=AabAQO&c&56b>#*{{D(_~@_d*;`hI
z+&|sW^;G%2qW9X}(oXw@m!EZwU-+{7dG9Xsn^!)&n)tG@`9hsKKIQN6g=K~9#x^^c
z*l&+{UX;G7{o%wV(Na}<XU(JA%U2mB>pjdne#H7k`I(IdqlfGq(z$D;_m=B3tg<Jc
zIrKPn?<v>oH-3H;l+)f}o!KiU7L_+QJ#mmgAk&(x9P_Kk5D7Lk6R_oYXgQX@c?gKH
z>3*y|Qlpgt6A~m?LYkb!wkNUn%>$Y!zm%@H*ZJ>LV(cUnu=5r39XJAPml;#X9ol`c
zQiMfBg0IlmoFl;Y%ATBl_f|rh7+a9u*^+E)jsQC-*~rQFY~3GXY(m&Mk{k<;06X;P
z@)?STe?$$##zeSEqGCA$?D$gU=RNZ3ED;_?*sPE>;RvulnsPAiw=I@yL|97LN}<w}
zBfy?oXY1$id;N4%Jcx`v*<iASBfw62k=)Q%mC+@^!_5V3o;c5#Bfw6+9k}+7lHzy~
zHXy=7g`xHw0rm{RnBOY=)2F`fy#|ZdU}N_6<8709ML*>>rrZ?aAw)uyFv^@Gz|Lwc
zQ2nRB!@>j$3ESSnUd|C<FX{H)p1JU4gcw_r*=I?zY&Zh!f?KgKf7y|sua5_t3Rue_
zmX;h548=$#5Fi$<GwI!Nu%s+og7t~`by&0xTfTXKjMDl9<FDs@^N~ItM4C(zO&ZP-
zV1Ktc_UWP+ml!d2CNo%IxWJMlZ`uFewrAzhnd63GBhuSL;$g!PU^kv<kWO^{xK@lu
zlHQ-1d}_%NV4qT(8ErPUsnN#_>76Ue9mWx0w=Qz7`s&D+wIV!}u>FPp)*JzLr?Dvi
zOZP*AL|Bio6+(prM}U1luc!P)v~RZ<JCd>I$>%9K0_?ut#q;FLN7qWQnDnl~dR5r$
z%>$YUvS)srvsX$y$mM$<oY1J&W7q;@Q-IW9HVk_l$&RBD6yHO3_mGC-C&=yz#rh{k
z+4zsGX+#X`!ninq<r}&cZlAN+{W|$<Z*6Z6Mq$_;m3mYuOdZUy8<A!s@}y!bk*1Q5
z&m-A+B&YoANOql%2aslfwuhL3gl{4R7?SMALk6#^j%f_^*rlFJy+9s8VN1g)T7^PZ
zp-_q&kY5Axr}#b!yN`J7pP{g4d|Zspim?p5M~~f#9k*g9Fji~vN}nQwbGL{#>X5Py
zxq)DaV`SQ|gBVQP#xXkb`~2{kq=%)w<tX@U+mN04aE1SBq(=l<)gfCD#Ew2v-{*Zn
z_(7KmAr~`jsyI~!fl9~)Uhql?yoN$9;KIrn*5BJd4E%#(LjyyjAyzSLluuOnj5l$I
zVWWMbd1x84%o5bVu=z|r4;7=}Vl;trcc9=MeB6YBoA|g5d9<Mr+P)omwDWNf3htrp
zT{2$zDe$ruvKvQ`>Iia#j_TNqFwOql60FDWz}g*n6qu(TyAylt#GZgP;dhGl&8K>D
z4Q+Q|78tsQ<W~vUBjAkiYTh&{QSB_rEK^V|Lqeb4IxGZ3^Jw(h)3x1c88I7WfHDFy
z31l<Udxm(10(uL%oCX`?ZJ#bkGfMf4ti~zSX)JiMzc}Db$~e0`BGOatIr_Ud;Sn+*
z5t$**;O(_88G=ZVfrVWDqkHXT<Yfzz!Y+N~S3YfQjy}VxRca0NVOUp{D-XG;+<3@y
zjOTa|z_33d`A__ibrZ>NQapZJ^ZFk?e`shx>_{8s&9ftIG!MmSV|XZ58_Pp&NZp3C
zu)Pf1j@0dZ+=0{`e0&zE&r%%iwQ;_w%i2<+LN<U6hx8vE>t)t4cIyQ4IcQ=~0@yyV
zMMADBph_F#JthPiF>H8Xcm#Ejbtrfpil!6XhBVty0L8me@NPc88fmKe{1z14!slN>
zvMWeQ`}d*XK0dzyYYOmN`5(m$bGWy`e=ym?BzY1G_QkNt@?;*GE}zar2e9G*RzlKZ
z*lMh(=HnWysG%54Sx?9nLsTnq+PKklOFNOe6Zz74-$t_A$PREq!@RXd*Yr;ljawO5
znLxUQ`I0ECE2qGEp%?kAE2AK_Cm-|qBx!`5lH{;Sib*4(HABiI*&1vIIY8%{<W~)@
z$&kdPEXS_k(F|LOrIpwQ6aafUNXQjrbb&=ExCkY~WEi#)1vgR*!;wk8F4JI|pcCSI
zx*CnjYeni-y7e<}bw<`!1}-IqKoN2-LP2z(3gleD$BoFjk&nBPb2r7Gv{a2OuXvjH
zPWXV#Po|-(1?e;#9UYy)*K_t=Jh`zjRTs8qxz8E}`vJ8gi81NobSwBQ<SItcd%>Ri
z!^h&M7?8fLNY;v6U;!``(TuK+jIdi+vY{a)|MNe4E6$<(<)nSY)QEIY`>H+57e{~7
z+D=T^%GAmlOoJdVt1%R<K;|nD3qvsM3go+jpCHPxMHyDoJukz4WjGXY$@oF5y4U~V
zJ&a*h8kG+$pJ6?0JybNDdANIcQ#8$Z+S~BTu!+WrJe1^+qz1JRZ*xfIq3I6Oc__;u
z%Lw$zu-OLLMvxR3Hpd{x2;w=zE@74orwd%lEajnnShf#agUK@Nek|M1#|N<L0CuJP
zYOJc};~K20p}6FJ=#|WmvtPY8`>pM(KlLcY`tE!4DJYuk`Rk(r7y($zP$72~fX~eT
zI6TAV$?|OI+KaJMG1h{IF>EOomts>GLx;7c*pK20EUv(&)WIvTrULs?d>Gpt#wyUx
zhKu)lSDsi>^Nz!8ex90KU?lG%4w>hY=R;?c7nwJiqCzYy<OiEV>{*C|VXj-whF4u$
zo$pLM=P;HY=5J9GmNsE)C_J~H{-LO4(~*aR7<RVPY!_-vvm<6lQZ&bFjt`*iJxfHD
zi+@=_9B7Hlk_ebmMgMnSTs13Ekk4wZIudMqcYewGb7HTt@2qlf$aO_<%ggfjPVm5@
z6NBUC?jA?f;i`4@1F;P2rgh_?cx}8NL_CI_sGay0Rcn3YS}RijOh(F;=Q==mWJtj+
zJxLV-r$VBLnc|spSS-WNk<8&m9^G*fYzbVHq$J%jk=m56z3(U}a~XD-%d)qLlGLXz
z`ESwT>lP+iXWFK4hcHP2vIbo;>@g%g#?Qa^kn|oOKS9zb6xTgiTxV9?x9@{i|0k!!
z%nG{Zo=X(%&-#1ON+nbk8USJTWJlN3YQM*KKj=*13{Pg7CQg&WUJJRCJnU|B_Afn2
zuNKD<!@Y`ZuTo`$1(8IpJKxi7Z^AZB*a19?VVm&CChP{d#UbC_>&W=8-*JS~-^9%`
z9A~iVy^a9-B*W_N4d}2HD7XUgB11O{?xq-07YTg2`vlr8EHEq)c7<VMM#m&k|7k#>
z4JeIT<0%w;3Z+oofqXmY5DeRe?7C<rz_33fyPx@ZE7oqsaS$la`;S-}A1mMTL2cm2
zyM+mrmqsEIuAHo$MnPP;`o)hj{m(3}yf@QaGv9Er;-FCn@!c;_z!xZtYUczRbpnm1
z_yG!ffFdb=j)I<3e9QZM$nyg-^g*?8NA|UI>90m=;s4*ulp#$Sy=Ng+>pS-}RY+6C
z=f8Vza{<XN@b7Kjy~nwMWH<Qy|Gn2?*jGsNipsz1Hu%X9BeN3XE|qv_B_0mhR2R>O
zp2Wk!-TviVU#(T_-%YCYQe;_*+^PQeBFnved=8B~$KS83$nGlt4AGA)`)PZ)gd;OZ
z5vQ2Y;Gw&ZgW%EIY`WHQ!MHKRx}xM!8t|;Xuj-p!58B#&(7gtYks9gs9U;Vo$CBS3
zeqI#o@!oRm*|N5L(}7b)?``-(n{_(8rftms#n6F)%4XLaTb>Xl26+cffY?S#^1z76
z5S6lLev%eSt}5X!OaK!NJs`R*eYAWsf^r9rWCGAXBz*w{>zwQvIqSLq9O%gFA1inf
z2o&_yp1#`BR|opac890j2oO5X6^(GPwz0LhQ_2-q0y%A<q_1}Lbp(C2p|Af1FXvF6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_memory.cpp.A744015D851F9812.idx b/.cache/clangd/index/device_memory.cpp.A744015D851F9812.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7bbf88908e478259ffab0dd91d0ecb83fc6bf13
GIT binary patch
literal 2700
zcmZuy2~<;O7Jdl{N#GG4Nl2cMjh#S%utSg~V9>CGB6(VMWRXRnV}cY>ELE&yGmfZG
ztRgBNgqEch(Bd*w3n<D`mq8J<Qi_Te96%3XU6>bpeqYa=zH^e7yMOn)|NWDo3J)(a
zf}pTeRY6W#-g*iIL1y?zo2SW4Hh>^IIs_F|Hm(@X9ri;87C0$gN4$m&r@}@D8&5|k
z+e*kUmB+jHD-T%@SfyPp8vp*2x}Ni!c91tt=sNv(Yeg~SQmCz37UPK6_FVg>;9Y?=
ztF7wY!M<F#Z`*sD>tD^kf##<|tK3Xodn;-`6+lnYX_q3|y5Xzz`)hw{YTrKEad$U%
z_Tmpm(kptpQhP(bELtmVc~sGJ_0@y7(fM0Cx>t8TdD?2j&3x(<73o$sRe11DuC9#X
z+}pqLSVYo5zwZ9u&W1N<+`4!(yk;okw&UW^g@WjlT2|}u>b8LB5f$38u6@l3lWn2(
zd8F{MjBT7f{H_JVOkPAzQhWikvu~>tKhfGLv)-V<i+ZeaXQf?^!Bk-D+JNE9GY>B8
z6NwG)I0?jYCs{rj_m=lJ_dF~U=;|(gyt&lQQMx3h(9>vRsDymmsK(5e{M)OlJQE%%
zOH*#VaJVD=U+&{ZSO+aD_3vY}FIs-hb}Dk*X`1+7kA6*o`Pg+IaYM_6>)t~7g{@aI
zba9N)?Gu+p%a}epTH{uEIAfgFtmJ^-i=8F|c|xPT;fjVc-0+rB(oYgYUcUEi+oRq#
zTX*riN&aT3!Fc)Rk8UvQcO|@TnTw<67#82K$r(Jbl3(^%mLU!e`|`z#xJ_IAS_f}V
zx1|(4{l{z5C+P^kvv#O@VlMH)&skY1ckkr5WK-`QTEm;Yv?X3V79B5E{MuzitDfd(
z@YB~PCQY9>?3PPu$m(F<{%kd8<6z1Bs^tB@Urg>(&&>Z@;k>1!GPm)iWa>sj&-a&K
zSe35bU)PgSeY!I0^~1TyGae^O(_VJkk2-IAW*t#X`$6q`YVwmS4+p>5p59zOe@E;a
z)AW@^_3n?ly<;6{>3KCBW45G<5JAN3kM`%ZO;#JqPpO6qzK9%oBYotXob+@)JI`aZ
z{QBD?o$YUrUeapr2U+Ip9KP~VSPJqN?%lHVN!VrS;$Huqrq%M=`Iv;Non@5`BX7s^
z@(w&6I#ttD6Lrcrd9PNuI=F71-{f8UaMyFWr!5<1YjwKRe&_0g2V>Eup<fC|+Ld2-
z1W3;<{rPZ=y(!v0{UTbp=G&Fu&FQ=XHMO;xzQaMOVY8J<_K{`3HLU44dEiCxt**;6
z`<e{LCs!?)w!gc-EXMWKxsN;Djw(k=v|DnLeMk_5_PG~ct5f_3A;-r-m=uv(s09oG
z8^RY*{{&ACJa4M2%16j6akB+tfvPPTmS(^ledR{{fXh@jmmG_mX~r~;nn&T80COr`
zS5xd>wdarIeu$riTF4N7NWP)$(YnfKgP-%rtMO>1u@Y8WQ>>{VI$tobXy#eKg+FRL
zAP%US&)|Qky`wQIC-Labs|a~H9?do5iq(sZ7A*qNhh_#I%b$PELdYv{GmOAIHOfE{
zU_L(K6xYF<`w<~0;^sgkkf#<gL?6tP%T`>@50N_}<OJM|@=#kf-n;-rYtlAo=kQAl
zutzd_PhF-)n;$!08zPKjU+1siyMqK{3RGbQYKfSi`R}3k{z8T^8@7!Ya0_EzY%eiL
zWp!SYF;Tqub20_9VcM7hCkp1p^fCiZK#SMa&ji#BF+5{lE}Sl+z?cXYnFB8vlW-&g
zKqrjJP?;6T1Y>fJTmS~OWOL<duQ&g+GlemRCnE@q7si-0CJQhMV~7}WCM32eJBX0D
z9^61M8W`i#_$&g3h%2%pB#BTWAwap(Tv-GtnVn2VNOCK=6hMTrU|KK>P>K_5t#Bcv
zrL?6iaI9cVC002DDg5HD=h+h*j!|KZW<e7Zt-%7=hJcS!QDZQC7<2G+@FyVpx%qhm
z7$6>>fd1Hn2_#e8+e?mcy73V+I1B;sf-yFSEdW(O4m?)B6Cf5KC6V9n0#JiDj01;g
z3?k*9>%tgJg_)oLjByRQro<Fb7&RxFvY=Wpi77Zx9hgLmQmT|m@a{+TV*;=+7Dx?b
z5=<*0B?BOaF(s@t2S7nKygmKW0(IfWcLUc7ib8N|5EKy+SXHP>1Tqz#^{=a_IG%?)
z$rj7v2`F-3c@%)M^^mva@`dMSbes_vnm54+L%~oH5{trui7IRbTSZ8y4;n-;BJvk0
z2u5TcGC5IL;h~U&t^uQXXZrgFbaue_F1#{%A{GZN3KOY`Op*x6I=}{~jmPRQBM=Kx
z<7)jS1k}yL7Ur8euc#ULIifJwL{kVI;ek~6siLsige0U3d4DQSk+4B=oK3onM>Hj;
z$k~8Re1+%=9>9Sw4PC_p;{X)l==66WKnH2kve)bX9<-Vit&RRilK$)cL&D(3zFf)v
zl3a3<8d3u_1f}GpW|!h(<X%#=0flNvqce<Q6U3CMGE;`xn;EG>Sv)oeUF7E<usAR%
zIK<W5S#IOwDEccc*OD&~TUp!MIY^wOZZdaIZ(pI8g^R}$bMF7S>vQ~-^U{b_<nS29
IGE&t402U1#<^TWy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_memory.hpp.C2F2D98A33C55529.idx b/.cache/clangd/index/device_memory.hpp.C2F2D98A33C55529.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3368d3eba49ad255b0abdc4c3b52427d020342af
GIT binary patch
literal 1700
zcmYjQ3rt&O6#jd;h2HD!&*jw@ErmiAR?3Qifx=jqZ3qivuor4{JVwdbAlX`|Q#A4j
zL^h&qV#Xp26t^*T#u-DF0ZxWaFm6$#1mhxW!iL+R5+E|Ci~jfGsodnAd(Z3p&pH2h
zcIISdb^0T;xh&^smA$%zMhFFw*IDhTNRl8lMc$*GujdcsjoA1>Ytf~x$7r>l?>XFr
zPO!`g+ot_p^C|o8?7Na4o*Vk5ZlmFRZ)$}r**TC@>HqTS=y=cYZ;{!(jpL~c!>Qxl
z`EN$nC4W-%YtL8B4!Y^P%dThq`a6DK|Bkuy=gsue<2?aa1HRaHaNP5URo>>dw<K(B
zVQaECJic8rrWlHA$kPwMF>b8Yx-xsM@5hwvxY>84{1CU&c;u&he^typ`_|A0ZR0bQ
zjh>fAHMI-%H$S_W_GszD5@Swo>g>A(qo(z&=YnVQwVGL(x$ME%(yiD-gQGF$(yKMb
zQ|)P<7xJv9YU(qlx9`?v?>zA5bI)7%Dhem_23qB}PQTbS{q@HWOCR+bez`i~wzgfn
zT>8=RjIv#>W7jj9m!}GoW)5;;b1M^*`uF0`K8c*KRFy81xv6d>zB<;oX@aNxrDUY!
z!8%44f@G^o%DN4+(~DagI224gGF}!b3dHzPC`2y7l8y6j@27odTs*}=emD~z1o>NP
zIxi*v-D<>CkQCt<HYQ9I;NE`yhXS{GVN^lUkgcSZa&Yghd3&kbn@Njqte_QgQGmNl
zcKmEz;+d8;wMQ@!zS_&%?-X_3pK4Z8p#W~64N}Oia5!B%9-g=PQ-n`AG$5201!52)
zb0Mv{bZ+(H{q%t~t!WvpueI8al<-~uc!My-lh%y9Q6&n1Z@8FSwWsUHZ+S`vr>j}D
z8r&U=?z>j+g<zgyz^!6bJh*%2*5!{VSFZ8Y6X4b}dY^m5Wk2aMnMmg>IXjMzQ{gBk
z%2(Qa_U>Vq%^J&70g$cXH97!y*vp+UehA6ubBp_<*<y;u7L6qwK#8WaM1ckbkuV}L
zU}}0?pVlJ#-Xr0#Im8?Z(fWH&^)s&98;MzlWb7)55Q@6uI51b2(JY~{8JmM3I^i3W
z=G<NHNs2_r%aVO$!>o+uL4z;Fu2SzmOb%OUODG@#4M~*({h&&qpC5-6!3rJli4{o*
zK?EtqVgsbC!6t9zTzR)Yhm|2pH87OJDuqf5x5i-&qX~pDb66X%eH!S(;fUCX1mGrz
zjf^o6xB_UTOo8W60n{a7MSg=BFi5n(YcK<^B@!&~8O-pNva3Ay%}LlBaFV9hoXcT@
z)PVo@#Yy8Z%;9k1wqmoh{Bv@Xtc}fvTYxS}O~NujE;LA!LB}dA&=4!G1O#!|6krO3
z1IXC3EP0^8x)5ApO&|pZPRzo3fEm_<n1!_fGk{TyxDH@u#SH=NfjzS|x;-&VXe^A4
pfX3i538b(q5bdy6mI#w{I`%u2!cU+eRv&yLYPq0?8L~(M(Z7Ht1t|ak

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_multiple_reduce.hpp.188F02C49C64E773.idx b/.cache/clangd/index/device_multiple_reduce.hpp.188F02C49C64E773.idx
new file mode 100755
index 0000000000000000000000000000000000000000..334a2253ffe79aa327963c3ba0de2279a1686d3a
GIT binary patch
literal 1302
zcmWIYbaUflWngel@vO*AElFfyU|<jg;^LB`%uGfGhB}}?#hks9;<FDKh#dVL_-n)7
zmrZ%qoJtmVn)&@E?wvm4US`H5=}EJ#Cf0txy6Up;clIfc3(i`qKmR8+t7Ki6?z$Yl
z^^!}46T7#&Rc+0x_wc{yqtGw(>(rHNZzsBP<>uZkjI!Er$iZquzZ6eI)CvY=_P8~{
z=PzciSp5BWhX1Meo9`|u=D6<RUnsl!b(7mK@mbAR{1xsl;a@t*pl4Rd*86AdH2Y4f
zI|n+PKmTY~PR}W!oGp#I8yWUb*1FSJI&oEg#QuJzT=Pv2<{V)ABKgk3;{5B(IscR9
zH?88^tHC?R(9E4VN6tmBNo(d|)*n6*3#U7>s$I0;Nm`}2+Cn#D*_tbc>}uUvchAh-
z?Qk)?%&d4~?uYIDIa*AA?wf4hyXMUG?WGG3PMi|9C8h4WcUr<?&iS`43Vtp&o4m=h
z!-~N)I`^4i=g|k>bFaT`&)RrO_+Ix@$>*YMpA^nWin8(iDwgn?&|+7~{z!0XAVYCw
zZqhzr$m_|bg@vA!IVsA-&c(nhDJ!V~CKxz?azGHVV^)|SpX($xCT30sUU7DDX)wXS
z$pnlo21Z`S%xAUrIq!m1M432%`lV#0G{FQ+e{EgftNJ3r#{x{OTwntkcxAL?Y+=$Z
zNB9@~*rUtC&BOvU-jK~u7EB<FU;mJ)j9uJsnE(?TOuwe4rZr4|QEFPT4imx`%v@mC
z!(7e7E2|`{3v&Yxubh$`oCj0G#0B<|r%GA(#9hvHJiLYmhAuF7#csFEiBVHgX5`_O
zk&)4WnUvbcs@fn_>;x1RR~0ve83hXqpesND0Tbrol~<D2gX!hrRnS$igz+j4KYUWf
za`7xD(5b?*5-?$)Q-$S_c=E#XNIV5$1!SI_f-*v%qOc;8oRY8-5>HuJ8HuMVtP1DB
z^fPjTqX*`4STHbhfujW$M%`Y_EpPk(Td?u)N{UL#z>*LTuau~i3@i|U9+H$*fEf(*
zkfa<EPhL_UiKn2Vpbb;G*4QfG=-h<qz_8Vn(Nu>e02mE)At(XBT=>+0^$R<{csDaJ
zf&^uxVTJ(p3TjHj^Z?^j)leVinWEI3L?xu8%}_Vz`c_T*!#`oM3QzRKMajkSK$T3q
v+y3v_8<up}lU+!VUx1s3kAV-8m_a0{{DJ~*7=wX<AHn27Foh6I0R$5OrcI~w

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_multiple_reduce_multiblock.hpp.BAAA7F55A7B0C61A.idx b/.cache/clangd/index/device_multiple_reduce_multiblock.hpp.BAAA7F55A7B0C61A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..69f1d2e19cdcbd98c1f6d7cca26edbc65e0e24c1
GIT binary patch
literal 12272
zcmd5?cT`o^(r4xz6|aZOl_JH9fWV=Hasd$$5iB5>*n5o{V{A_ppIyNYf{{cqc2Th*
z*cC|>0ntPfQ4E%V0YxkkMPo1V+rS<7ak<v_{>b~|vEc0cW9IHMd)l6TdPQ{Ua@X6y
zAZ&ERjEQ5WjT8+G44SgPsnepzO%e<Y8XFlH%-HZ_zrmA>Tbcd(`90wWqtNp)O+&9+
zM*Zdz^=JOx+ks;f&CmLuG*jI?HTCB9EI*sj=HKp-ga63dd|`U(v6b_>n|k$r*|D_p
zkavdNv%)D(^KHAgc(=}J(%z1jP@X!oi_MP(V?s6^JJ)zr`@5Z+%zZqm(a_nRZg0n!
zK8pW2XxXsp9R_26Ff97BWN?Ey7e{>cZN>UYOKe3myRv^z8obZ{;S}%NzsKynkeoFn
zWNYD3k1ZEp+Rg3R&ky@fFCM+)yR}2CcidYS_%dqzh?4JSxS5?#c{|qT-b}$GLnrvo
z_5LAd#Gpir#DX)Bm=b0idD`mm%e`Bd!rv!O--#Y~wf~6EzAqEgCKnWyygc4v@}Gki
zJI#6X&TCe?g3N6{86FADRqqYFWWO~@9_n>uuUBGVmC*26bU^sBdA*M9dgneWNuE7q
zocmV;F2z58b@7Wgg;x)a|MKKmzqqF}Te)RQlh(h>zMG#u_gco?xeYxYC3*!eG+Hur
zMn!<v-iwR+y0i~$H1K)jK!;O(hI|pTdS;^zZGtwYjU93A4?CZ*qSYtY{Q9q~q^g6L
zU8Wqgvp8P5OB>~TEHd+mkNQ&Dq9!IL*;Or{x($-9_%7WTy>Z+(Bm3`>q$^j~x4zbO
z%xCX@F;c6C<op?QV$%LchccRPe>vpq32N^>FO2*){9I-|KJ#>#QF?aoEAx6R@yveW
zle(hmS6QEbCwt{IH4of2I&5*H$W-&=zlK)c=`z#ib>0)lDR%pOz8W$))-gh8l3eLs
zQ$0HC(5p$4{yn{qvHD%9op*9TlCE#ahJovPoo{oX-yf$nZo97znlvg+wa;VdNZX6u
z;+AEN&fTdwr1jkAaAV5pGXIPbEqA4O%w9Kct}!ebQ!u{z)}BkRJNCPro7dU>{z(&4
z>o*3mGq;|Jz5Cm&zP~<;HQjw{TW07_@2zGh&Tjv9v1j|pMxHLuTI_1_F3RJ$|LVR0
zR(o}~H}$$YanI48-I|Yi*ml9<$t@}u{M0S`>gdPo4mhhj80^35Ts&~XchLuqrN#U;
zV|x54%hpNGFzoUfn;}=Phh4k%FYTjY%ltJ{UNx?~K4!+3QwI)9bY0XxZiRH=@iyDo
zOGaN;%<cJ&w!t&cP8*`lR_-nL%bmQXDq}<TtcR<fWGs(=6m!*d)LnnW(3-DbukPYK
zAbjMI`2PKlEdFef%jBF{9iH`iay0w&whlIZ%im3%Iq{1Q4Vn4*nRb8is*|?{P3#Qh
zEYUto)ae@oLq=vljYfUlCCWb5rwOVUODU*I!BXE4i_xixr&f;M(L;^R?G5Co==2nW
z^$oF@|82Yd*yIY|hA6OJmSdCUXsU0B#j<gvyz7VcJx#D7V;QR%>s|LCvg75v8q<r3
zMrv%u2Dyc<w@{~VKCo04g~U(3|DYKPj3ogL6HudXh{eg*^TLA_Ca=_}V=RB6<6ju0
zZ-~X&Y1v5ONuMX?*o@6%fXx8cx(AV60u5e=u572%pmj?FxmZva3r+P6v3PuaVeIx-
z(?ZQLnDsE!dZ<g?gUGteuYz5N2h3DsJ2r>M*y1tz=^J9{IA)+j%&R4yYV=|(Z_xM+
zTId^M8SHOx&t?0T!iLy{4YELJvOqA=H^ef$B>08ffxt#;3}n63ptc5E>l<SEVt!-I
zsY6{4n4=l%rN2#om%0a$zchX{-~0BUn+>snt%2;;$jzzlVbB<XsRqa}y}|TfsvLrc
zsL_TEaUWgpv$fPW)W`H512z|zU;MfWTCpLPqG~DH>KkI2eQJcyzb*~Es77zr%R4lF
zht2g3u`EdUI$M?d=Li!tWWBrv&0Fx$H^j1VWKN6y1#vg)cP36y$FZHMZ-`~_%c<we
z-Zk9R1T9!E(W+?sx(AVy=BKYX@ME7jCa7WL!=O0~4*KQ;%h`&Pj_N8OHTtosK1cKC
z=&o;wW#zmdyX>}ds#IfJ#!`)T)!0hk5X-792^zN-{maxCz*z2Jr#sk3-w?|hv%gK^
zM<v;6P-1hrE;PO_X!XqpmMP-9hX?1JHAi#C($}W1N8RHCxu)CL8P0Z1o1i5l-vHGO
zu+cZfveVuv^h-Hvi3yrAmT8)44s{PA?=tj1kdR=#N{w1J&3ov25B>EGu^c?OZ|K}B
zlML0^p0T_@^B3r%Z-^z!f6bmdd0tO6*q9A+Qq-IjZS)PXoX8kE^Mu#6Of|M<y;P!W
zB{S~&hFE_5%c^ui+g8OUD6(EUS$FcNdk{JI{IWD(^EgivY|6;xpeYA$ee;3kVYAKw
zEoP2aqc0oe5n4P#XMIB~`GbPyUfc4@)EtdiFJo-Rc+@?JTu@rGDC94vY7=b4$QMC#
z5#00*v7FlxZRyn6n7IdI#<E##xmonnH^g$q{^qLhP5T~IV+%HqS17$gOMOEu*G?9^
z@!M^qHpd{=%RuXaa@~W-747b{EZLpD)*RK0JlbZoN8N+SH?w;`$iCF$3pM(%0Un_B
z1GLvS#PZg6`@>;v?lW1mVJubXR>kbLz9E+P-<E5iJ&Z3^qd#MLjnZq@OMOEu2GL`_
zoEqrFT!B80fJz@dJ~)nSKf9pL89y3I@<Q-f2we-tFRv(U9uAVc94wZD4{4wze-9Sl
zEAbw1-UFeWeiCd?LLix_B>x2Lw<E{*!S+4`k{(F%GmxHvGhqs2qDmjUXoeC~0mDbw
zELX1`Y+^$D<w7KR6G`4I_-qzJ2&)DSUe)Q`>_uz^`vsT%LO2nOS}j}m_3GBYn@RFy
zX|f50GfAG+aaLEpj5!_WC{P0QPJn@&x*DQZLoY6*6o^QH&pA$okW}cx`L{vHHYGj|
zp~oSd3*`iao>1a^=#UScIsF<)*T97IT9PXvv=YKeXp&q7p;by;1O7D-#OaGrT7)L#
zoF#c5YW86>M|~ic<jbhJti;z*b6tru1hWjm`hy7di3BojK=t?NOdYB8=>(+_#+H{u
zTNU3J52Cyp!d62kl8|P}s9mRCUvJSsl9zzX66in#NnQostDp<V$>5O;VH}?ZyVKBy
zi>4avs-X=nl}2u)(x)w4f9ewrVvtzKv5XV!Xibt_BDj|bI?5X)xl~Y>3eCvB*J7Vi
zA&}#A(JozdAxA=bgsQr9hz}d>868FO%<{?s!?!z+`7n8u2o@!RC5>kJcxmTBDI@ck
zwyqO2>x9-cE$bosc75ko8OKC@4-N035&3$Fg>S?AXhe~{WrxFAPXdZmpL9(LnJCHY
zQChFu)%~d3kL_vmB)JG3i_nX(N}slo#OVLbd3-ENO4isM*P?nY%EYPCr)<;%rEHd}
zx>SwJ=r9ByhHiXMk3jGdB|Z)j$DuE$pMZ!HN_-mHorbQQUIHCTAd;KQ73gq9iOZo|
zIrJl)k>m>KR-wd|5K#$zIlT%Zs+71I+EqhWPLIXNSnSWuD-I*$lsFOH6R{16MUs=y
zKM6xQu0m}Ux{$M%Sc>6%2X&NvNb-BZ`@PVPtKLG<exc~eak6Niti-8e%T&>eR8x|-
ziJsfUwuDLP*hbdp#gwB`Lfk-Amk9IyIfzyVxs^+DnqZYCIFpV_@(scKhTy>G6EB*_
ziyyG(=(p``aBglxQ4X*Tu%~WCxvfK6CyoLg0u`v8Lpucui3<6Qb0&!`lSD@v&g=8h
zU48ccJE5^CR|u{Zf;$mJxmsvj&E+P_H9};K(4Ett3vSN^4~}06L9c{hj^7I%-wR<J
zFA&`phzkCNVvmJlFHTPuJ12`>3Ae7=l+k>|naRu~*NU!dMK>b&gjdElv>NE&K$Jr|
zgoN;g)d+2CxQh|xSAyFs<$4y1x`m=IbtB5jqAr<ZI<PQ1EJ3LR9Z11c`eO^(p>O+q
z)zz7U%$Q5qgqlrCm+%G+-=Lb#H4@A24d0=f5`Rf97c}MEDF!%wVKCBoc#J5@J0Wr>
zL{Y>O<z3KW7ld$p79!6=6q%?fKY@rR(2YzeAT%ZP&})0+Pec`PeM5tZ4=rqK!#Ha8
zV+&eSV2j5;k8e2gv09Y#L@7^HQ~FwqH$tAc8TfzlMJZp&kuSREi=Ik)*L*R8<1?ak
zM#+ChbU&lyKO=TMqvS6Tr2-{?fv77G{dj)`;?M$dILBv2>8z6foEULVjO6s6MCm6^
z51HIvSX*#u`6n`<ZDJc<ic%?>k#JP{Lnv){*WwNB)(`4D)k=~VgV|ze!NchiFk7O;
zOTlcZ632rs9@^7BN^(5(jfdeJuK?W&B|Q;ziAua4%+^B-&c6}NHY)KZFx#ZW=^&*;
zGtQp@Qic*|f|RMmIUwagGtw(bJ`K`oCC&vYSBcMqbY6)IK`K<@UqJdri64OU0Gg4W
zN^&(w)k^#jq=!oU99*7*59faYE-#e$CAhrgxcldAEY5Xa+mmf=p4d503@07v-MI6Q
z6US(h7~!PY_@rn|gnlDe4@}?sdmtkeil&8PGa?KyJ==OjhNY7r$@8@H0=NW!0E-_$
z$0c+ZEbfAiW;o#Zx_+OjJ14XL+`Qew$QKQGv$l0r_CjA~kRNAm6f0PsQ(w%WZ%A^s
zXqPQI(-~t((0bF<26rcYV$ws8wVg0Tc64CgXF1v|M_1BPiG?G}V!@H(lf=%^4%Y>5
z@^>u77tJ<{b`-~`Zo<_65=xiQo~wU-5rsOY>WNxaS2f`?o)$bUf(t4pI7WeHh0p5E
zX>-EoD9{oJTLPbPBPs)J8939~MUu-PqzocCPC)kr^rz`b@(Ofcp~SaPdyB<aed&y)
z_o%HzXF6?5@)oh>7SWT|Ey-KOmRmWd^{{weTl$G`oSv}iMA*D>4!LG_M(UkRJkQba
zIhrNuYa$Xm_#3`JGpY$p$ld9>e&e&j|JaLDF3LvSikhvcr4_66wI<pSsx=8JeT9jZ
zAhAeg_!w;{l+_;IBK8ZzDAky8!s}R?W$Jr2!g{n?k3KYlByT{g4IIzNtEikMUi`&D
zlBao3^W_^E;~k?whr#<W_;TAg0^UcIxCC02fS)pvY;{G6%fP4%O!!8ZLF+Q`=eQi&
zl|u;GT5Y1)u0n}pF(?*0Dig<`I3-R%Z35TpTE)?>P~t??CMt0fYLh<3&PnLOg|rg2
zE0y$bQ2UJ%uR`rACH@w*-zxDfG`fW*N<}rQL=&#4RcKU&Cdx$C=nk6jMD`9^-a#9V
zw}_@&L`$V&n{HL&RM9k5iMNTS+djrEwuuh>``bm+?MnI%(R7Crr-`O%O1w)n-NiAb
zIZV6$K<N*(B*CflRZh}h+A!t>>r0$mf0v^BQtV0!9+z{daYgrWQ<!(%gKhR;*T#bv
zoN}F--HjBQ*=&93ll<z!lmMGTm^Yt&V_5L8u9T9~39_N8t`f?}+l{Su^KMyuLc?mb
zrP`(>FBRG^75dO{lDtvy*(gMCd`O5mB#a=u>|XnX3(d^GVxvb5jv7J5yo7GTqUx3J
z18gL@uX$hX8ht@ilKY$Y*OGUY<N@Xbv~<Rk<k9A%wNwF<<T2)Bv{bfZHt9DlNMVzH
zF$#1T{0>8q(kA_mC~*n+mq2@^P5NI^;xaHT18Z)BWzeP!f;ldSj^)Z)Nd<JQP~un&
zjl~GsVpb=>&^RSdz?KQ<N~%$7lPy;$aU!-%RN^G`OhRAIe+x}-p|#S6Oe@h^X+x$}
zXsxuN7I)A|X){(^co3`&23A{@I90StRpM=;<2KQqY$V}+r;@L}jDGuHK6|Wql6Cq~
znC<QVxkDJ~5eqf-7X{pOO2DE7+_)kYqf;>k&_+?TWNAt!nrEUr5o(id-%{+xaXEUH
zV;EtT{+@xH1=%9=DrqRCDWhWjr2{e27Bj{aRHyK>j!J*$@PYAI%zV#%X2#4)7Ymxj
zLJ*nns!MlkVzXyw{NqTzy233d<mC4IY=j)qGDmczEOYgX5kupQwZAc`#0j2pLR%WT
z7DvVj{W-oYSYB2pbC(7G%R(q6ca+pI2Uve;Li<9_rHQKUmW2Mzyo=|6AnsiTL<|_f
z(Lm3EK~#*dEg^PW1p|3{mJFSep*P2?LAM%0IZlBdDKMPlJA&06!G&hXe5&ABE$BF|
z5gcncUh`A<&f+EWJlM9Si)QJfEty(vMr@WX+EPZmMt{-6YTNa<JbZ06Xjp^IDWPSj
zETMh8(1Xizi{P_G=uCLSM)#X@a%0+jlALdRmAUw;MS>06opi8B2XF3U_kqPeB|Z-p
z=fRt-gQbX|dk&G5scZ`UBdA5sNFz4FIkCq%u@~)s>b!)Nh3Yn)|FMQ`5BuFFH1{<t
zN&W|^(Gr*e*Is;anU#WTDFku9UWA$=Uh|+jDeHd=YPO&aH_!SDGBT#!{VxTpj*a(Z
zsvZx)@z93`OrMe6pvT12F2*c~g6&dp<_fX`f>$WZ`H2vmsKlGVb`v;LN36;MUYo$5
z;|#FK09z%NGr)`EOt8pQ(lbHMRMK<7A{T7=<`sfPp%VWB7QZO*L$G)VwzRn{{|DU*
zXvgtO(7oh%pDwV+nT;b-nCKP=+696$<t7rlH8RWM3QY$HR>jXm&rA#<88b$-%2Fnt
zd1#l1I!-S{^FnkXtkPda(w-dobJh`y_$Paq`h8p{Kk9#Q?d*#W!Wkh)G|l0sU(!$J
zPU>$iNl&u=^qk#Y+aaayrlHRg7{GUN6%1bmqc~26;mJyT8oHl`o?Ob+(7hUZlGIM9
z4T}l}RCfQwL@1K8gZKuNHYlrWMcAqc1No}!uTE*z*#^sd6{o(bbtM1);U+uCzG>yM
z=ZS2DaKmsD%F-nHZ^7_y!Gt@ibE4rn(S&64!@1uE?rZng?N6ji9k6t^_QIA#C&|4`
zdN~q06@BaBkhimXpeP>|td0uJ$s4h)5X^G~3&J_-z(b|2c8q2Qm5EI=c~vUcd0y)_
zhr8|g$G(HMfVqVF8)TZJB*&sfEc)~0?J73E$|X*hhfKm3P`aQD#r5~hv=w=~dQ39v
zV#uz|{)2a-l-5vHcLU8QvICrVC`~>KTV-Jf+ImSofmSED0kHB9w#maToSu(%`RL7Y
zA-Wgxg|Oleh8AK^qEiyXcCh}sn@{8hxZZ#;ialp9T|9sO{?i<$10UCQ&fZ(MVB=l?
zi;ZgQr{A#pY4S()Q}cY$g7ip~^F>|0vc7pnG(V%HpAmIul=K48zCd)~-!Blm6^K!U
zDPA($t-pw;$)Atj-Sf?@iBmr@u?s=+oL--6*L<SM7mnxO|JB8?fa%^|!F{jLo0Rfm
zNlbrtyMp=54Mf;Qc*U-n@Z#u$rdK6-y#09Z-~g{1@}p^wBzZSl?napoX|;zomp$mO
zz=FpfA@E|rzs4>tvd{X>`@!#r`i}UPMTl|6<80^(kX<E9v#hx{pCiptpzpxzI|$-l
zCk4z>pe4u0K{^iB-0PeG>4Xv&gK;s~(1zDu1xuA+O(s!$6)aUL@!w$lH`s8CNJHB+
z^d!ZpEuA@D;&qeSs{{6{IIipxoOcO+Bu9EFg2^!tHF?~EB>895{H(-97*vGee4sLP
zEki%TD*dYzWN}w+FYA|9><X;oonW#PT)6zsg2`EM;SzZQwoky1uecO#N|mvt9GjK%
z*mHI970ZS()+PUQCzGpBH}v{yZT!+tOpy{<X3Vv(l#n$r=LS}>;6#G?{a$+~)jgZ`
zOxGhFBcsTPR|H%Ldy#GlpV$VS#aRiV_Ei+#?M?{X35v?+SqMA}5nSG-7*&dcY2#RR
z6D`X4C9lM8q5P7?P84<Mtrq4wx1+QjP3em&{ktu+<F~Y9XUP|GSpi)RHVOW4Em&!}
z!{eJOV=tD&9v5`Sg*K#mEU6H5SxTHMc;^Z_PR|p(^EkfUs=?Dn#uf=|BAdm=n?+Od
zGqq_<<1EpX^71<#S^IQ0ds7+XQ9*lD@Zmk?2$ng57vU!xvQByLw|D$Rw0~*tJy^K>
zbKn2+q5ApUf_cr}oHqV%ce^CwxZ}?UpRr|pd@uRpO=h~q{fUXp1{Z_vV$gA~v>t5N
zgN{^>Wj0{DL5Xv~A_v^L&O88%2jEUN`cgL~FFd0l<Ui&PZ{DPx@-AKTlx@W@hhaff
z{i?kkamy3KC}L1?gt4tb^%}I|I#&O36zLef9L1EW05t{ZM2wPr9yRBcxCpxy;c%j>
z^si7+N6}*@jO@f-pz4&~<fhr9`Tc>;v2=f8K=wq}8wT`MN~aAdjh-4kYU)DPmqG0t
zS^6Ir`e0z-_TeMS6E;1c@Y=U*zskeY-ObU-*~Qht!2JW=z>$CS<R2aQM_2yQgMW19
GAO8=HGlO;j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_multiple_reduce_threadwise.hpp.956BE9BDF24D2435.idx b/.cache/clangd/index/device_multiple_reduce_threadwise.hpp.956BE9BDF24D2435.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6104a1590113f8fd59ed067be147e2c243dd0e65
GIT binary patch
literal 9686
zcmcgy2~<=^*6v#mS(-=Nve<x)@Yq)o8c>maQ4x)Vpd#v^QR0F`g+X!Gh$xFj!2kk^
zVKlD9I022|f*_)CiMT~WG^64kj35Y##=n5Rsh9Dbe@@PvGkw}u@8f=5+g<8bm62h?
zhF!FgNP?$_r6k1teX52;BI!f_CH*~dmV}c??nxz*l%iu}8$RggCcic2Y;dP$vcZ6a
z@!AgdMj10Dn|3cNQ~tIwy7{Nmc@Y^M?Te2o=H_nn3)och>*9+k7ta?R95m$B^BZHf
zocY(FL&w7NehGN6J1)|vsm;$v7TLKyV54&JG&jrK^KaZOwECDtR)E>CN8jHxEf`W8
z{$}@`O5tOpQFT8=uIzJ;Yqc5Xw92VhfrrWRgc%)MgJspXRR+f_e>=Q>^TN)T{=1SK
z7X20z9@_U*wT;88*9Rxgm~{Et=C^wm@TODTC(gFGxmhLc<3ICxW@*RN)t}c!T`3MI
zy_&3eJ-p%4cdt^ueRz7<T8E!~x7yk@c-(%Pu<3Bz9EXN~E(pE#w7e~A-=WQ=O=d=e
z5>r0hd|}6e?|eUUd2r>C(dfxLEiYfQw=8@zr#!FU0*Ax`zOC9Q)Y>X~$%gcMxv!+(
zKik}RV(GH8L)Mi9%}jn9<32TTZAL@KkxwT0_gnR)=XBlnJ=aD#XCJhGPHGL7-pPwU
zLbhgIxawe2zGi>eW3u*d7mRzC+vPkf*uQ<j*V#H=v2kY)Y-uUioLGBh!HZ?4Ww+jJ
z&sx*ccJS!ZSc41ecK>+%pVFm!w8l1#P1xCVIz2TaL%1l-l{S!5I(KG=JujH2rD>P+
z&HTXevt|bx?>V-@bJDDWUSC{pHGlr8TkNr8pT6G5josTZ`0(GZ8VY5bhiu;|n7GYZ
zzxk-v`gDsoCkL+Hef|rN>CYxUT&VM}bvoZq)?D}Xn-8WP{5DI~xpJg&(p`<o%VQkd
zG&OFy*9Q%W8W-dHyvR7k|IpDVPS?LY_*=mF#M;;?bzYBR_>Q4=k<FdaM-1QPEV<Kr
zVTbOnl#R(T5go^tDW0TnFpP_gJmI!#)ljVx)v(A3JMv}*K0e-7?7=NK;U5q)GP9`O
z?R?$IE$@WS-B$3+bX&%!4jh|Wa_vsZjJsOv)=YjKzbXCh(ComJ+U2A=sk7<HD$CUN
z7DeKQ@wX!;KfW<C{7sv@#P;${+v6(^Xs5;Gtev9mk+g7$Zbj=2?S5yn6SdbCDa_Zs
z4k-)!oqT+u&qqm%5<cd8NhDf>hL~wI?zkZ*ea$37CXr>5KI%fEP5p&_w6$00$6u|U
zrB8bGlL*7*!(F={D3*SbW^>`dt2!g1OU0pPq0Ze86zggps5#`*8mLb+srUfM4}hb(
z_=B=<J~8|$a+DG2Ln#4f0ruSw6!#ftr)&LsWSl19sQ5VO9tQ(;ffUd0T=wnWQcx%(
z4oZp83=YkpQWr=G&g}EaKZcy$B`3XUEm<7>s?-Hih6T<Zv!+ksVmZ;Hlon#yLOj(4
zQo^ztO`KwLj>w1yZ77u(r4oB}ft09}RfT64=wuoZ16oU<S)fDr1I1I1mHcb1W7E$@
zM4yWN%>10YA1I!tTkeq_y2DeK7*cVhG*ak(pm<i4@vqn4R=LQC6K(%Bn7#%<T_7cX
z?)+^^$IoK*iJaCDC6BV|exP`PPg<Gj%@;f6L`ucAM5mT0)df<P7yVf7;9v7flW3Sq
z1Y0dzo9>51hX6GP5MZTa`*$PHZjaC>GFnGA$g{ypU0@wqKa|$ZEbseNM(k-v(@5_$
zqEHt|$*UZ-#`k={P#JNfluoedgn{Y;DPNvwzxm9w*j7edDCI3E-h!RFK+693PJ;)M
zIsJMJ_&AXrrv}_pAmvc{r<Iz|J(uYcBih}gAUz6Z>H;a>9SC^&&Edk0GNPajts*w7
zh@ZMZ%8^-n-%X6$vsIU9(^@J(S^>uD0x72|kDgzjaW+Fvv?yg0C*Q=GsSBjk9$ID{
z;<GEEFX3oIRiIr3a&>`}Go!PgrWlQzEhDa!(h82P;HfT<a(+k0y)04?D<h7S(gF@G
z)ZTjvq+A_QYJF@(gj`0PDP<+GUP*k^1yX)?K2#TZpkjoKxKqmSp!^-&)CE!=IMzt@
z-J|cz$UsVIgOE0eQWt+vqE2kyW^TXAi12jKA!Z@Y-47JEoN6{&7_oT1j0~VPtR_aQ
ziIciOO8cPL@p`ZCg&7eQts%%X$hP}|;*KYJO9njC+NMc#s5qVIrW1X2fs{9ub@3yf
z6&A^e32meU96IQf^%O{X_xfbeBkhBpGGa_A?cmT3Ug`oViE*DKT~taWnw6dwKA~S9
z)8vJz&QpgkydJ*ZYX03@yf8yC!yVm=7ZxfOiqJZ+TL(c*S_I}r;EkTb3%kH}7X%{4
zunSUku*E@QuqD#&Ad9ZZ3z067qcAvN$fcyNc*{s>FenX%VJ9&3Qd&>kWyGW4a1?_7
za13<-mPxw<FQY~IMTKE)yfD^ptO!Mqi2jJlVn)P>(DdNxAFyewfDbER68aA>RKloA
z7|(DOI8}icdIm3?1gDc?+ziTQu)`S7Q$ts_fF1e(FT5a@FNh09Z(e96maPoy4O~|+
z`wO=Q883|C$LQk(^FnlB^l;Gt0%JsI)<?4@Fj>;5q)0ZL<gny146THKm2Ap+VHE_d
z660J5%7qcAcV5VcpnNed0sj&RW%9$o9|nC?9WUGk`CaJCCg&;0pEBIr_-fF=E4Sl8
zLr4lv3c+s4)28h{@%pF!8oV&we!4&YnHT1|&K<_kJm-1A3>AY#F$}_<;Hs-7Rj;+o
zDte+85_oFC#+k$dO`7VDIGrRa49vVxM`U%RFD^V?*u9Q;GyINNyd(CgKjR+{YQNca
z*qsjPBvG6s##oW*q|)7+9UDI*JoQ`ccF;qY=BcY|cYq#70Zad~BQ1k3N#Co%W%`?;
zW@K4)FAc%n++DyCX$UGSl?_84RvsecY2_(G-lM!fK<0ok(}%n7TbQY(A=Hy0^<*dx
zNkh0w!mpCi4F5zNe<IEd-y%M@h#$j^B(RYLGkl*o-WT&9k%&hm66NkA8Z$IaCU|LR
z2;TnQ0ob3eW`|s5GpixoB96C+6aGv?ctli>hzDX1-&4{Xi((AvgkAyp6=qdDU8%Lx
zhzu8TbUkWLmw>+{Cf*0teeh-G`4ChO8J4P(FtnAyHXlo-_Kr>FG=x+LPlZwFV;Vvl
zj7Wn&!i$Em6T)}GD741GKNa;#c&0E1UbqkP`_La-^zPsMeZ1z>W|@YN%kjCK4A<Z;
z?3c?0Fr3Hnd184UXPL)Yi{*ZKVtziy=ZpFIoM}F%Wc+;2E1w(8@D`5WBIa-5R9iSt
zv3&9tE|%c}jxP}N3%RgDE}Y4WIKGI<1Lpd2`KK?Zzc&Mz*wUq>C#ywsK!>7&o?I3M
z!><$#EsP0GGT?=2hSU192r%7nx(LM?#);5O)l6U17B9>kJ2RG{*{azhS-dJ<gccbt
z>d%-<43~(|Qp2SplmUDO^kv*E;IqVdJ@D(rcpLEBpf6eoFBAh`EXLb`-!8^`fZrp=
zdx768#`}QZ$1rC4)Hxc#r4a(q;-u<)A8mT{m_dp&q4^(D#pIFbJTegj;YT{5#}eY?
z8>nCtr?ZLcjjF{Rf?V3YGeEWQG==udhLLPUIpCKg#<>ud3uBo)AHwp*xEO{O!*C`q
z1OGAzXKP70_?L@uH8@qnARHJk)PPqF1TowS%2u$)e3Yk&pIsZMFejuFLtGnK)Z=M*
zp<fpaFO*L#+sFX4Ltc1ETwanPh{sL+V#59}8@%cK6mtf}Twk1@XoEVpDPJ4g(5auK
zoHUqOL;)xYK!xLr{;pu$2eQy?`g>bf+hCl==+1n%2bHTm{%uK|e0a#5Nw(HjG=?1q
zv*X}^`lFZ_A1Bth%<yzKWK~OC*)Up&VGFUqZsP_{SK7UC!)AFQ&?V3ZRZCG=U^GKP
zE<rw+=F(Mo<lG6&>*j^b`-q_=RZ<W`$q~u144oq;=ZHP)SUs_>Cn|=o66>oBW472!
z+P#Hj9lZ?$Z^Q6WVf}`A@7}wDr+Y5#PB6#q7%yxjz8gstnlUfzAnrRz7~)x0)rJ;&
zvRza=tBLk%IydSJpQnzYok|SRb8xRpi8q122?l6bQuU4%HGm4{q}@AJ*3k-Zr~n_9
zC_V)9hv1IY%&FexP`I^i;$L1&aKdO{?AiZfizyWY4u54_@{acS5|~_KP4PlKnA9_z
zQgf$qu}1v`b6$vYjq<<&&=lEqtO#X;Yc_bGz4Ag1xaNp)84N4~PqBXvEEnSn2&#bL
zZ2XnrTM0o7SAnt$1e}kq#94V#jH^LeEygvVtobvxs{v>B`%|DiC6?EMvQ~^wgYvW(
zp8@3=F>VIEX3)oMp(|0>YXN<hsJDV%E9i^kiC!D%vv|@5#%*B6@C#z_f*7OYbtU!&
ztzz6p4BEu_B{6vUXWaiKv0}e(CkE|ec?U7*5aU<G;FTD65`#{LmrN~x-CP#^4;pVa
z5cvk;gC2qAL+z>Suq4cSLFT54;N_@yYSukRB*^21GEkI(BUAhnuz3Q5Q4yb2eXDb4
z^sM=`p<iIoFU(alM{=o6pT8Sy!V7*zeoBl)ybxd%pwv}w;%MkG3RJSt6Jivie4yT=
z^FpXms1mo2)ObCkd_>0UIaY+S!803ta1dR_>zO0QW#Cl?zHAKT;8iZh6)>U##xP^8
zgy2dT$#4}|R)GVKxyyJhPl|ChSXPU14Y<^R2je$`K{J?&jm@A1OvT1#&<dtvW9#1r
zHe%y4dBLK3mvNc2ig6n;X%pj@#QG(1LSxE&5^{a!Cy8(V%M=_lY#MQzfNugr9G_Hu
zScWk&tL4ba--6BW|Ihatn=U_kzN&J`*MB+6esRaKD&Y8zC)A2_IO80aX@Ai^W%6ge
zmDlJ>oJm|V$q@9TE*zdo#xs0^7@rWY%qNJ~2@-@Wb?(^_yRM`ycczo~ju^fp=BOO%
zpq$|b&Ky%?T-qqH0r&<Q)zqhOQuTQw&T`RKr@Kq5lL!4<?y&XMfi*W2nPwa>WPl<A
zT$#^gf+ACl_kdy#xT4o=3;M;U|A*mvw7x<vqL7Qk6bBDBX^$QP|4`hco>NZjQOWYm
z&(F&Sh5qH~-x1tk7F7MZ@IOw@p7EFcq=(~OsGer}&5XjK;Uv?b)^nPQCsg}WDtkpF
zlnmo}A<jI`j%6pa{bv7h^u!DCe(@r-)O@KO*31h_2Q2ksC<_!>U@peCSulX%^`KZU
zmahlf^<w#UP;3WtoN`{+3yQsBybl!n7%o+LN7Qbel1JwyohZ|Z9gd5q+b6>;W;Eqb
z7ndtCTK2woig3v5pZQDbe;q*uIh;Wb%Ukg{O+%_aU1kCQchLSF`r&~q^$6mdLn4@=
z?;`HINGM`Fji!CP2(pV{$LzJ|Xd0~pqbY5%=RlgZ_!hL^LSM8qy0;*{*(4lW;DxV<
z`&T3Y@sX!4OGYdG^STa<qNYbPrOu6=JB^`vvGYWz7)BSvhdA*SGVM#JqZ>!RXX}-c
z^vEkEon3DF|7TT|K0=?ybuZUb;yt2$kLY8f##8UnF68vldybyDHlcLri~H}X=0xJX
zXOrG6j-Y~4VpK{LIKC>G_qRU}+&P^Zc_Z{`WO26IZn@i_vXMLgvVYa=j`&n>7|nP8
zQ}Dq~Q%CB#5kMWGQmNq2Cu((<>(8Bg(pp7FAwJZrdt9(`>m#pvx&-8MZn>NaYt|6*
zIHNpHA;yk*oHN7uoKe15p3hn4vw*B2Y~kqFit)E_Be!s)5aSw3hk6cV=fDJeE>&Mk
zpgEpP-1A}Qy@aIqOaND98bIrTuNNO6Jp})U5XNfkxy8U5fAFzg8aYOp`<|xjuP2Xp
zvN*kh))+2?yY9DsYnw1QGnf}{gYj*!Wt*E{!2TETLX5W}bSM`<b^&Zz!#!6d*zh&i
zUmLz1J0(ZT3t_#(%tWWge6%TxO)-3o2qg|k^kL4DY?y4xP%iMfV2aMtbq>tui}69|
zeGtr8I@JK?4d8;^<Aujy{g`b)e!RbK+*el|fck4H=%<1`>rxu%r-|`S(BBF6tl=gw
zYXTKkd}Gaz#+u2d*Z=2v#f`@S&Ik6)pwl3HN#tJ=H+0n-twoVD^E1}|oAqPtq?&sS
z{A0H5kgBh&P;++%{}9~1-x%IAnS06^i*0|XenwjytUvgV)89tpoz5E{^d3MfK0sU#
z5EXWW7rr5`-!OcCV6RrK-il00Y#}-=#DJw6S)5K5XMpJj=9rXt9%Sd)VWm`kb%tHV
zw4CnHyY9^}Fm&nEJCz-V^7Js&dnbg6bvUXCCZY~`;Tb5NF<XA<e56!mwy%T^w3HZ^
zk^wlPmZBpkT)z>l-<ulT!c%W*1fxc9VpTVRM-z-hELC6K;Rs&HAOE^KeT>I{nFhaH
zc6xc=&Xc|W%WCk1g<g1d-6FF2j5Ue1K@Vx~Jg{H!Bw;;uooVLNRLnycnJ-cyi5C`|
zFBYL~px6dZsDKXD{F)J8pAMk&_~&H?a|zmko_k8xg$<m=2F?ymq_gv@6Rtn!wo(=4
z5#KyA4poGA7}R#IgZ#QUG3>d|#F@v-4@$fS@@rtmRMd01iNwUXIa5REy?w}j-Mfar
ztr!oO+2Q$_1UrdGWD@)-=1LMAS5o4%q=mFDBBX^3xb)e^cSCh-?QE>gEiA42ODw2B
h{jtUuiG=<iEc}?`i$r3_Ui-6GEB0#3UM<<{{{jDW!JhyC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_bwd_data.hpp.1BAA9B467AEFC28F.idx b/.cache/clangd/index/device_normalization_bwd_data.hpp.1BAA9B467AEFC28F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1168b13c40bdb05f04429461c9e2cdb5a3f2218e
GIT binary patch
literal 1082
zcmWIYbaOLeVPJ4h@vO*AElC7283cj2xTGjEgpq-vfti7!V$R<FSntCIB5l8es|??p
z9eyU>Vpt$@q`TK}?}gPx)8wvSP+P8&`CR7r`)Pi;um0>1XmL9Ed&~RZDiYtzoZoF@
z%?glk>ld`xvA+Lz-HqG5QJ*d_rt9utT)Keye6)r0h3&g`6?;DK=FZ<3U2}4&-v`UG
zhe9RST*Boq?g{bs@oJnH^4rtkpVMx6IqB+1_iLy7ZEYngefPyye>$q>*fLebLVdBq
zbo~b^FJCUx{Ku#}hcj)$!>y6ixwb}6xWdODnd>Gw>$^wF9P2C7e0*4EEo=R}WaYCz
z;(zu!{(U|9%tYg3K|v>ixH%&tH#V7NBwd)0zjlky1g#T?w2tqXc4>z3O4m5G6E4T>
z_&1jZZaVz!wsu_O(aT#j{-yi=Ht!9et>X7;ciFd<1!tDz=pH#B_*Cp}f<;o>iH(<L
zmAy|fSt}oA?0HR+<!gHvSF5t{9;Q>}Zr#O|xk(McFitomyLgf7taxrF7Ip?M12zL?
zFu}md1PpivMlQxwrbEa3Zx;x&F);%bh_H(&f(e9ztGlG+%h<CD1(?`4fLefnOHEVF
zngPmraHV;NwXlT0023=rTAW2(941|qnpS)TXjO=xkot>>Hmn>xT;j^&nlN2FToTF>
zNIXepNhF?>vJ?_eT3H&2C!;Kb#FJH)MdHaR%fWfE+bwfq)KruidAP)6#57<Q!@`A;
z3mhJvDrMahcRAO=LW7A5911XD9xekjg8-PG)IL_#2BBgnpphclBIYnX-CoQsZ~OmS
zfV`+BZV0mp=tV6FB%Y*}Boa?bOA3i6ttE}blhKkv;>l{sBJt$3<lwxu##RAG=O#=C
zx<^e$O&u0tFq)AQ9APkL!rTQE21OK1_^AQw7j}N}Zf0OO35rR>Q~`|?RFj4o1q=sm
z15;Qy6s6`QDgo02%m{`j*GzXtth8;01rS`OxG1?;ACz=BK?DSF9nZ>TZIra)X6EMM
NWZ;1bGca%?m;gZ2Z8ZP@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_bwd_data_impl.hpp.A627060CAA69A19F.idx b/.cache/clangd/index/device_normalization_bwd_data_impl.hpp.A627060CAA69A19F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f4e31374887b6370d16182a695d1a8e7f0ff5a23
GIT binary patch
literal 9770
zcmcgycU)9g)7~=&1lG%nvI{7?(nQKdim3D$5EU>{G)9r4*keJB#E1<ML_us25fC&|
z6;ymxECdy$8GZd;3HA~Th_AiGu4ulq><;(-KL5ynAHU6=eLT<3nbT(G%+09aAw#Tt
zOC*D)1jofp{b{0(L?Y?R{)_!-{)|Wl26vN4;_{A-xjyH*r%~&Ucln(w(kjZf4vWaw
zZ@h5jgzJr{I{!PqT`TGi8VoD?dgSqn`_<1j?hgzph|)XV>3LHB*n9Il-F<s~XnQz)
z;OMHg^D~-mKA#vQuUvg@!t`Tf2kKpNpXK?VJuRDy9A<dB?2ZjEuz2x$biH9+@!+M3
z8Ds3P&70t5(Esu$qrN@Q_Zf4{r)F%!t9=8O@l8HGOuMS~UpVsB#q!+1z+Da>NA7=h
z>+<^3pGJ-9w*1>Qkt1hx{nW$ls|2$OuM&f_ov(jA+&%nr+C1mLCiB+X)z$i;|CKg5
zzuYjaZtDNIF3mo3-)G&XdEJhb=}*&Zj@dVD+w%?IADMBRNF)4{9yW0|$M!pPF}{37
zpVH*T`g#kZ4(X-Lo;uNceydSr;XTh=AtrOfhF2TpWUTyTxyB&=K<e_S>PxjMgKa$R
z(d1iW?EAFd=V-!Q<*Y$##(tdf*79)A+sR#yG-kd`a9MU+9@27h{QG~c-TPztq%Lhq
z3uhf)`{cnL-7C){E#$ky0^E{*^w7QjYjH-CW8kxMO^N$f_6{~F9lvV9r*3X`!3!Eg
z&OZL3@le<Edova$*P9>SyW2S8b;T~9@rnDd4R3qj()>u~DL1|I{N0?E#Z_|~r*HUc
z+w0#79$o1%^~cA5Zc<GjHdS@5aa`@iF&|sclpVaZX6Cn1eGVQkU-NVG%w>-Qy|&eE
znNb)%bDZ9S1x*cS+pOYaDkesKzkASz`}qeJRCl|!?C1WCSEqmXseJ1U?_NtU?;X*2
z^uWz$t^Y}TVZYV%vck*d?5?@vr$k(vy-v0(eDF_Y+rNKzzG(OGmm{sm4E7s;yl3~T
z^6~>WOEX?~T9%$>T^wmO{X?{8dhGqPc2TuCF&lmzKV{Q2gPGgRRr7{rj*piP`%8Cz
zRhrE3z_hC3ynu|MZs!e)uio-4(08;MGrC~pSAF(>6EwJG@xa2P!-xD@^6W&n!WXKm
z(ft~RJb5}XzM}KG(!2d~n(B16r;=~Ow;PQ6W5v~#Q*u_U$Pa&Fanh43*!Sd@vTvh;
z8($MeCq8}cygr@om*4*3($|4^dJNv>I(pN8116<E(evXbpI__Rw(euxxoyhvPd=Rv
z{Fq2a9KHAwcJwSfq?0gp`qrvLFDLx)ote(bQAPd2?)B>J6`_}NB{bE*?zHptMMfw0
zILz;>cQq_$v)<aF!ClW=hWk$`pGUUTTyVDk?o^KaYSq-uejdvnG#EI<RV*#hee-(5
zTPhtqCwJx2lf}<7db>)`RZa4XJZa6p`L=9v{h#k^UK)O*_sx~gyU#_ZY_MC=|8BB-
z`n?4!R`80pzat|o>bCEjJGrTOS;&aM5~$=%M!p(z(%fOk!<gkpK~|$f?>3cxK6SfO
zJb7o9*DG#t$%yS?Spjp)ABK<i-K%n~*|B-~+OX0)ifbRUN$i|CcI2?}kdKp}?u8Kh
z2|H~=AIz%0_vCo7YvRxm%S_jmy_F5Q*K_E-;H0vJdVS}TpwWN3M+Oa?6zV0LkP&#x
zX%-oqpLuP4(`UUh=M(<1i()2~b&^PQ8*f>>jb7;`120PnUj&LGu-6KSKJ!KFPj9hJ
z)~)i}J&ZtVA>ofxqvKSi6-eQ1y6)IIZeN}ZteKWvQ00Q1R&-F}nx%2m^v=t`l~J-m
zl?{$sffTa{x8otV_ViLfch*saD#E?}fz9SIfA}4E?D(4ux-t!AU{nSQtw2g|i}dV#
z>p`131IIL!5osCeqZLTu$0hu6r*J`w0u+oAr;2lGe_*rAI-k&YF_F%kWI#^|uaqcN
z?GK584oCu2geQFu=WaXa>Q7`akXexq*6HA&6{w^C<SE%}9>%rEz@1UHfNBfyT7eXA
zkA!)7PVfGsfF8_>g{p;4?GJ1o*m=^ys^kw{JA)x>t|ii1;-VEu39uQz$;!}tu?%{#
zPBw#TGn;aa=%766`FGg0;o%DC#I!`JBKh_QHV-p7H>EXtPKE;fS@V(}OM16IusQtY
z<<E29z+odWx03MJKz0qfY6VgvBBQJNZMfVZgMQ40Vo(&bfocR&Cf@Z6JJa*TEgASQ
z$}b@Q1-femQWhLd>XGA6xKGnj4ajO(M;d{Y#j%Mcv~YQm47{0^t-x&sYpp=a5{crT
zl~w*<-N1y6Cf+FCvi*V0OAi>$@bo))PX?Y$Lk_5N*s^H^QdT&{%nh6#d_e{tjFJhe
zOt94oq$EDb&o+ts*FqV%F-jpQ3Yiv-K+0;T8=oidToWXNuNWl*tTWi0XarJHq+|1g
z7tDOz89FmlcM|DNVxkpDS(~sn^4Q_y0Wz>*T1r4s0==~YDeKH08fERMAEy94rX@@j
z#<xGPIsL~Fzv%COml=T(Yi<UEW>9GbQa1g2m*e7+Q68N@#wdHB^B(A?6-de4?>1|3
z>+jc%pbMiUP=f@j)C#0*wU7E7{<R+4?Y_*^_0VHISZD=ON@m|0d+x}MB}QP%v}982
zOxj;7ka8f3m;OEGVwnPzj51C&&awT0%@wmN9=utc#&+5u)=3tqve<lR1X8M_?fOI4
z6TA%Mj8Y0lrOZ8OL<eQ%#c6up7b0X}%P9Gv$Y)wK0x5L?>!lmbL$}MIKcnmf!<}ru
zY6Mam_n$fyHmS5w27MW28+6$QW?F%irc2w-r$t^$Gy*d=no8<WNqw{eDc2u8jHw)P
z!AAyxOiKYM3fOMb2&CM-)W0&U)Gw(M=rdmwsuyb8{*V|@!rU_RMfVGJhxe|jxMTvl
zOvg>4f0Gz#1?qTEel)kie?oXC&}Ej~Af0azW352SV`Xrg*|zomGVo(sHiBv+vqU41
z@?!Vdl;bPkvh(*fqilfg8`x&j2&BA;*zUF>zx<p6y0MNHsTMi4Kd||2tyOb?yJLj{
z3|MoxD%`pKfz6-)=ihrjWpzz50%PWE_JZMFFxLvC%%3_f*1m@jv4G=O0k;Y)a2BN6
zxYaS_s{WR~gUgRG;znY;kvOA^;P_t&_bahOi~){FwP6h>9>W^t_z#5pKuoYJj!y-j
zR2Ym{sttD7HKz@B?1yg%_l9&wBF85K`*<RjYC|CY(9kty%gju->$)7DNc<DYP_&Wb
zlZbzk7-tc`EHX%F&nAA^Vw?sZY2b$)8LgVUv--oeo@_Ew2>S#?M&3BQ`;G1hHWb1>
z9T<bKp0qvqc1#67$6$_yfaV&`wGq%U;(Uw*V)W(sOT_sS3B)MI@oCgHjryQ7;rMmb
zcAbFDF7MNC+i|!r(c%9D(?7usLk~uLHjQmW*+#5zG#vk#4E#(25lgjs08XQYY=u{<
z{@FW(<9ATg9kdsEAd3ffhYWtcJ&-k+OU$M1`3-wm{WkxH#|gfxm<2m%w;fc0eCvS~
zm4A6QO204-#|o_L_B2D7j7spRgmB!^m~T*Nd!8ZCa*1~?2^Q8b0lX3*2uI{J#q`6H
zl;2LX^;%1e))IGNz19<>^#aC(h3QvkE5W^fPD}r7P&~))hwl4f0Gh$^mC(IXz?jA`
z{m+Q(8Bt)zQf)GWQ`yh2PI_fgxSY+{HqhS&GIW2K<*>zhN4R&S2WntD4A{pDu~eJ;
zphi#INs@_OCM*CQK9>y3CEws2=<s=DP#y_H?0GFuH|D9S1?#$v*tQXO)ac#w#hF>1
zCccpA@U@hyr7}FhYCN!(`U|*@a&_YNI_h3W`-|H{>*y!}*Hf-u%&(^|_0&z=9#T)g
z5pV<L8pQkt+OL5Q5VsF+prK+suYt}N<6P1}w-VuBPEqcZpb-;Ab{y4NqOg6II4<$R
z3{HKfhoq42g*}o=2Bnfw0)7PDAAvpQ1RVbqOrJt;0lxs#7XrqE%Qm+r_Y}I_4<7p=
z5_i?tbNy&`LsRM(D}>o8+Z?Y5_e$(J6Tv4@>^U_#D{hVeZO+<B+McTl9mmRJtp)Ux
z{3j8rCeGEw7mp;zUm?y{h%b(UxqoWAk-DS%8G1g?CpUa>Y$uLSCobv43-=qx|3X}T
z5#vq7Ws?|hC-&P(Ux8mh><h%Wkk}Ur7?+1lqdL15tY2!pw8NDSvui!sbUCDi9mHKk
ze;4T^45N<d*O5LrjA4bD@tIrc8AePb)``SJFe8arCkgmFgTcpQrZOMFv2bQ^oOZ#z
zHF|hAyZh<By5<}o?-VZ@#&W0SB6O1|ZxU-fYaD-zC~t}JZKAv_#&?MFju<x+WwRK!
z5M_%P-zCbsV%$oUtzw)BhM8cD**M28HHMqP7<UxM(mum1Fvg6KV-eIa8;mjJaxAMi
z%mHIe_&Mf%47Y$Wx+{)_YQtQ{*IrpUmJ~ZQgNtw$TEL-2j5Dc!CLM|qf!Ra-H;Zu=
z_0JOHZ0esa#yQkKM~t^n|1DyiOZ{`jIFI`033$xJ%yGN(uY0rAJVTAoP!n9uFym8h
z(dA~gtRb27NG1btKVbA^>5S&qhIu|-*VqnP4cuzzSYxU7(uRjHYRls%LycY+Fyd#z
zK3@r&Apv?OfE(hd!(Ou^d-t-@;aE<ox<MRJ1IM~Hy9T|`7LHvSY!avoo<WYK|ME;~
ziL;8=MAqXAB6~rMu?MO4vWSYOLd~fS*XJ3tQHQ#O2IHBUU9-a>ckAnje~c-yWlcvj
zQlgjXaeNUqE~2Jrff~CMQFk#8DWby#TujZ1sZxvw6jN^jm(cDdR3*m!OQ;uOyydc0
z(cEouRpNegK3rl}@HeyQB~iX4{c-4O*QtCZ#;=L;wHUu4$~R*CmMGtf@q1$Qo;V2l
zK9KGoNMFQwsn%DoGWS`01QTIsB19pP<(@Dk2_glY4Bp8QD&SQxY!yrta0>XOz(@h7
z!q8MfpHzEw$2G-`t1oT8vg4X^>`rXGlk~;wGV^<2nFy|e+9a?{660hrN(M`zeHB=*
zf{yX3^Ae*JaeFFQrV77DkHKQrTf)61W@1Hts~{M$R9iSe`_WYys8^7CN#saW=Pf~P
zL^M%};Fc(4M(T-mOA_N`uu2Beimd|oRp2A&O987CaeFGbr3$~7YReU97?vy8sagr#
zN@3_yZP5aWShQfo=S23Lm|~AoZRrAuxE7tI+Jc3!So?_cKH@DzuQSB?4DrV3#j$&~
zem=;nwO4SC-F5W~K#td7j#;E%2y!%;9Y@eF0y)+vIp(7Di$RV)jAPXh{SuI4p@id)
zfX5LSgqaD)SA$2jfYU<g$|oEC_Uz5^);+ClF>B`dFh!UWODP;bPBBi2r3j7>SA;9^
z;>z(6iU=j%jyOJ25vjyvh2s}07Ai4S<M=p5oD%bXj$fo$q{Q1Xi@fgf-nij8e!2T{
z5xPliZW2eFdv)Zsxh2N8iOp>>zC&#8h;cKqX%^!aV$&kVcZtniF>WO`tzw)B@=P$t
z?W~Tx^37l_j=b_LFc(K&c{Z4fBd<IM%<&wlBd>f5n2RH?JeTo1B5x_^mqHgT+;Hp?
zr(XtL@Ybh}%&yJgA&$(hEn=KWgEHv|^sws49JE=CvuIG37-!R<Y%$KEK{;Z)g$8XA
z<6Ih)E5>;=C{Mt+_1K|F1uhj-xNuT!r4A1bZZDS8s>^i3k%^W^i?Y~>@`)mJm^dF6
z=d_K)xlx?cCR5vF+L6=Z9I@!4DfM9|=}q^auc%s5{*QqX!x2kXHI+cr!0}bovWnVa
zO$$Rf>q=7(L}KP0#}i*4FM7yA^9gEpg7(5pCG%}`RALw9MV2MhK%W}$!dR%r!)jox
zfUkqub#c~k9R^&70L&sbpAOn_epS2^oAFXAFQpd3(pOP=6}3Pcu_nogkBRKDIGxZG
zC~;}FzTcaC#WYRDh}((#b}~#jn)$>%U%&<NX&cVUJO=;cJQfYQt_rC6ko%7|qU~%R
zH5E~u$K7#Nx}h-zL)iUf6Y<_e0tA~jk>E{aq<{;Edjatg<A4GZBH%*eUMOxaBmsp&
z`yO}i(E6>>S<KWupxgs?xCdCVfGR76B}b3QHn6%fi-vOi6!{b@0Zo-p6`@MvTuFR{
zc+f<gn~0AP4^~s#)wD0}CUhz+_`W6_dkct%NgX47l0a1I(Ug601rI)2e9&~wUu9oZ
zXbuh7pS|Rd$3d*&1T{V(B+iwO`YicYIZ!?96cUj_CJ5K$R5Cu5OhQ~GNgZ{*;G@(3
zduDd@_lx89_`GcWVoc~N*(|jYMVq+hk^p@ZzyoJZs;zwEEMeGXKTHHJQSeo)HV5Wu
z;3WJ|QxC@<o{)JTxZzqng&pcP(zQ*<ZEEb|-8?EnivRII)CNAC!o5D3tfRxK4koPd
zfWAtHe*q?}@F3tfVDd(c--F3}0oU#fontsepJg)taQd-^&xmQjr9np`$C8%zf<6xJ
z)PpSvTkj9}L#nPD^{u7-&|~TFbyQVHEyUQnj`|9?o|@NFl^74Gr=9|CpjHjkMvQ$M
z=vQJK-asS8czFX|Au0nk&>~hCD(U!_bH2TojvArd{KcNYdK?>V0&oe?adM>Eml8Ny
zbdJp7Xx>QR$-Uf~GN$Cb17T~NOS<I}dvqX}oH5l&z$JlF*d>}b6{z}-=ea>|cbkzf
zjJs=A?6mP-7iXs7cX0e2oCQ~Z8XQjx7;Rz3sb6$p8MD<@_my53Q?`uENg8BQWAvJ8
z7uB_rn&Y7U>r}bNUA3!#jiCXI8^8fOLX(+d%?lE=86~T%X<mz<3Ecl7s-u)y4OpT^
zj#a7I#~q^*J5p4ZNKMd{sd2wV>MP(<YFSF{(R6&@!p8KJ$es$dF{$>g3tEP6T^R8R
zkv$>Za4b^oyB8$>l~-Bq@-M4iUraxyojTI?modT!^N4vKam002UjT=ECf}oe^|=|E
z24fIo4$S<&`pu1?W`=ZzxqxO$XNpi2ajqf*1^<76IA0(G1^=H)ZByw0oTRrO_wF`*
z7_*LzHJx~;li`AG8N@q7z?k5(JVo=GNC;YWU|R?7xJ>h>&Yl>^YWsn^J2%UpEgws8
zy|8E@!Rf=Af&_<xfl-3GW9LtfU8sIpV6T2?{+#fOMAAq2YA1Yk6uz?m$qWDZTA%JZ
ed8PF_2F@;yee4~adfQ1{uu&p$7QQ+NU;h{K6C8m6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_bwd_gamma_beta.hpp.78751175103676B4.idx b/.cache/clangd/index/device_normalization_bwd_gamma_beta.hpp.78751175103676B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9ae7a53839ada48944d522e595040fbfb7dc58af
GIT binary patch
literal 1118
zcmWIYbaM-1VPJ4h@vO*AElC7283cj2xTGjEhLM3`E;9o|#hktUvDt?VMB07_S8cre
zvgsXn6UXF3H(HhQ#BYaMEK4`ra#GWCRq><x_^Y|T&kuATR0#NFceZ+-Fz-1t+jrMh
zr7lRQ>vL9AtXhBDuV&jz^Cvr5^`s5l4EWQ2zPNl%KHTx8Z3eGymFxw<UuJtAo?7xl
zt;U9V^83ZN-)989lw3CTh|9~j22y_(-RtaM+x#ZNq9OOXK+CMT7A84)3-1_s6|7Y8
zJCUVv#k#^ZZHw1;=Cah~Kd#));IX`S({<^^ADWtvKTf#%Z_47oT`A{wx<5X_rFHS$
zY777Gf9K3Upl<T&`}$Q5=Oj*hD4&*D^5Dv(sau@14>)WrJ7ws(ONIS2lWT6;hJ;fm
z7I{rPIp^=3b7HAS`IzR*8?v{HM@IfTZz8--``PLcoigt3%DsiBo4n5F91{yFvkiQB
zp}e{_t<z@poZMyezPo=-G!1^T$Mr(}iikyDQWo>i_#Bv5T$!8H01WHs)7!LIgL%9K
zm{>W0Tp-{QXAu`?fO57ys`}gH7+NL3#0HaA*HpKLN$)d${((g=(wmKmnVo@4gk3}p
zOfYaV0i%zBk&E&6eXVQH-KX5<W?}&<Fk~~-022rWMX71UAApKO{DjnBOtfL;<lz!m
z7T1J1i-$`>SptbCsVs@alTwyK;z=t@Bk^REWsrEX%Cbm2Ib}H{p1iU=oCmX)krNyY
zF#pAFx6FxAQ&DE*;S!S((}1Z;?PFDK5Gr;83X2$s*uYeGdoj1X?f-89@{5+ZAxt&U
zFIo~vJV`A{B%YL(6cSHbOB#tMqa}mHlhu+%;>l^rA@St3<l(%v##RAG=O#=Cx=vk2
zT^$ycFdFDZP*lPk4WpU3z)|R_Qr10jmvbEtmyw8(F)SKkG$R){8exi_8nAw0=NIp0
z1_qCym^8w}>VoRhF#CWeaT*E3OejjtNmK%+5m*p1!0d<73}X57ZXNr>_@=lhxmX{R
fFnK@(1aQI_z~s8}W4>VGXW0^FZXQksZiEB?15$R2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_bwd_gamma_beta_impl.hpp.A5BD375EAE722F5E.idx b/.cache/clangd/index/device_normalization_bwd_gamma_beta_impl.hpp.A5BD375EAE722F5E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1e64a0d39b186ff362132f21dd7b1743964765ea
GIT binary patch
literal 10178
zcmb_icUV-{wx7K=AdH7$gn>b4Mx;tP3W^{iQbZBKUZUbn?Dratnuxt%kRrt(7Q{kP
z$`uvRK@8X4yGAKiVi!e&U1LcIZx1tzbKdoj_vQPZpE+xO?%&yE?X}lhdz}pH9~QR1
zqeK!qx&MM0Q|3%Cl1L<N*niP;W`En8fg21g$U5{@?aW#~dEK`9t;V+!4n+0LoR{Db
zJbY4Ihq>CMxZ?{QrY29fS|1*pmsQZWY3YiZgmtdle6OjFT3oxh@&5OZU!~VS+-dhP
zyjR`2wJS}k|8N`nNEWy(ZTYAmmuq|K{RS0J3Nt(MxOQKJf70;Ok9^s|5#<j9qpuxw
z7!&4yC;H6d{v)q{3Y-|Qb=7F&iuhS8pN3BV9QDtNZf=wOr)>P8@#g-{!`sYy+^{%f
z%Afsy+i<(Y+4<a)a$b>P(`9Cx?mgveW6~C$pE#|tV<XqA>DzLJgL!G{)5dW|!<5qV
zrOyVtpB%G!!lIze*WKHB+*f>m$7ys!QpKexE`Hh6<Im;uA8BWl4+*+|(syspw5^}z
zwO6N=Yfc(Jd-FP?>`|<BO-<g0dqJLerbeA!t~|eBTY1uJ?yk`}t8b2VA5?CCV)LeL
z^z0dO@2T7?u0%Hea^milvwqmRhJ4)D%3@8%%(T>!jn5r3uQ*yv{q#%ZXcJGLz1oRx
zr#RDy0o&YvAJz8xK3U98-Zs*9&t$E1cJboO6zBe%^M_8bxopyF-o(>$S9-RJJ2rB9
z_Q!J*Ue7p?t>`#(R77G$RK=a-o<I3Jc&op<v^8Z>YM{*ZvFoc-RoYLg=&WAJa{_L6
z+9#`cJoMhQyvEseSE9x`m%EIkC6`A9d6r)F@jU;!;hMkd>hOKOXRmix?Xi_D46Dxi
zyyEIbx3kx~oO$YYBO%m$@WIY^8n%wjzju9&PtxnF)8xLcBLZg!=O+d)U08PLX32no
zIZOL)KlZX?Kl9Ju+xvT!d4%NO8}hoX@rrZ5jD6R?^!&=%30cFTAZ+;lgwWqdo?Thc
zaNz0dy3mQ|D?R@4uZ1b4MO8a5+Ix?)@7ep&8tSv|_QuvBlTIc^7(KaB*VsL{<M=th
zJ-cxsKsl_k;P#12ofaOgH&J!jPnS>i+tl{A;#tnBTSk9%vTfzkR<+K{_f+Ki6Pr$d
zrK~@`rKo?a$99Wuhsh&%Tb?t$9u+$?GR;QQM?GWc!DNTN3+<h37GA1%jksu<SFe20
z%VlnsdSFG`+L9M;PTuh~PRk#rt)BnqvB|D8+vKGD`25`d#Sd0g26?@8<xcE9JM}-E
z!+T4AHjjQC?Cg?NFxc4RxNT|V&~)b_>AF+jw(Ht{+mzlX7SPoR$DPAv!v{~;J1{G9
zb<^Qr-cH)K>sLQ#=T+DHMQzC$Iqa>qNmX~VPs=0k`wWcrJ@d8JdE$T0Usx1=srWw^
z`W{>&N$^ou{bO;gyRBDD?y%h^hrjPt6)?{0(Rrh$pzGIGbZuz3v*F#<-AzmSl=j{y
z8}-AVgO05&-tz4B0?P>d5u<IYuDx0`+w6#Z3HOUlK)XFn=@X})Z1;59`>U~S9*%Pz
zw|(x`rgoQRo&47;=lG#Z|D9XwQnf45cU9A?<EGyPJ-42BJ79^IW#P1)-_D$0bGUW-
z!gn3smjBrMeb8Ff)9dCt5}Vuxz8_d_{e4N`&&~nAUvE-Ix8JZWdf|+Tt4t&k>bk_o
zX8)nRGU#R};S0gK5IhV*V#+)Y`(x6M%nzBj8f4IuQPzO%8n8Er7Rsr~rQc*!Eo}{E
z?IrwiB0WwN27#0|N4Ly>Sa?2G29~Usogm){)&_wT#rPG_<X5YegE{NvBba>zjX@yA
zW>(Iv(|?i%6KJK8@IzV+ac_P|%qU?qB|Nc7AY+T>humxpQr5>YB0WZI3<CRT7yYX?
z^pzyS1UN?71+901r9mLY{?4|-SKsZgk%2Rt?0V3whmHo(La8&7uC6<DMh1b5k`4~(
z;ARj==`<=^TRbe}pVrWp4HT`3=9?eb>>2fD(B8v;Es;Sl*28LWSj|+g7cG>hVH+18
zS=LVmT^J=7EONotAdun}JFu)|dyJ(F0vII?+NFV=K_JEVR+W3}`urPmkg<8(1hbpa
z!61<0A26nFXJg<VIan~t3Tn23stp1uy`txC84@_!N(Q~zK$)P)1Q&xqN*`m>AKR*a
z9%&72STA!mbNS{6Hb<PedDODg3{DQFtoZ_!U!WR;K+1?Knv&GHSJueDmkqK3G#i*!
z^#Unhxooj({aeisGU&=E8KB8vd!rXf8GFmi$}xY{*VbUm1{$jw>)rgo<_U-2IUc<H
za+n<2b&&8{kZGZfK_Df1z_sxwdLDDO1}oOX9L*fQ`GL&~cK&#7XS*3pqYBoXO&zkS
zk3k@1$tTEOWCt}SV9Y4Bq;)N^G6<x^*uGiay8ei-4D8uTvOtps?gr69S#)ZM(|q?A
za<E~&oThH4X-|VdO6<^-$d79_ca}jA>tz+RUBym~UbIl=MNf~+p6F@~D%MMsCd$3}
zfz8RaznpwBGW(?r99R$A!D>4=8w66)=01<ork<^oK@UdR1e#6YY7j_SJ1APaI`PtH
zOR#2x6cBj<aWn{|tova0)}p_Qy<YK^Agg4G*9)Yqmkp0{+CT3zC-GqiyuGBoWAj5|
zW(1NDH9O#&BBikdlK*9H1IBD7w~6U(A~y)^V{?z#xt`81m&?GBQ8t5SGdoy%ft36|
za~|zI{8l0ZZ${Y&nvIO27f30-_FZP&^saF-=*}nwU|qnLtQSbp?hYMz`N@e>GH_y)
zBCslAM?x=<Quf`I`pnf?<uZ^nN-@ZbncLF~r2M>Q^x*B<z$Y^BW0V}=a=^hLkaC)v
zSnT@#sGST#86_XA^VtsR1yat<J6LyOd#+h)XwB4rl1NVy4}(BT<=xi_$>;Jzts$6E
z=C_;g()_^YtJT{_<UA|CBLjan$XZaYWu8PYkn&J-)Q|HlSZWDM*25zCBFE+jHa~g)
zE+Veg)=k!6!kWiu#&|S8u=%O+tzS3a>9I-%E^L52u+C%Z(hH<WW>5KN^gERhmw@9}
z0Jj1haNnhd=+tt)4AF@ZUlG|Wq7;bwP}M>_;L{<XZ|NZ;j$cWvR+7%R7>?gWI&LDp
z1zbw(N{K%jG{^5FF8fFbViQBKGnN`c90nzhzeu`XBmuY;9J2+}BCtd=;n<-v-42#`
z-Z=gMbUOfn=#R|3k~U4t=u+E?<72{N1_<+67Pd@;){@}0Bur>sM}pUhaS`cVM8a{l
z9KVC~-XX^2B&eM9#ddk(q@p9AZ?&f!A5T2uN$<NiCXBTlKgpNl6NpEG7;8xfE$NOg
zVnAmvHWKSb(he2R@v+b~7P=#r8p1qo1crGcH3WCmA4=Asi|qZK=B2LQ9~y%py_M7u
z$dSkK7f9C&Bv8<I9+>8V0_}`rN5nKA6nJ_#c6?0>K!GQlW2R(U2nx)BIDS8L+YiCG
zAGX)Erhjg$=t_+E8c^1NE&4CaAV6x!8-y91r5(=Fjxh$eZ6}NEk!*VLJ~Pf;RJn^f
zVc_8SJyf|zz;;hN{%khyUJ&7UjYK1D&b!#de*f&L4L=4Xf)RV)au>D6_l?;5mU{$@
z2@Bhd4}|+bY*6_e9|ys45P?`~NMmp_><wuQo8b#0dqJ!PqCTfVqC-&mzBhhXq+ck5
zpRD}6IBEYm)<Y7JClPPKfKrG&MZlO2SxB4HAVFvI-RJi}-=wp`uy`_3=p=!JCXitQ
zE+YMl$Y``eoq6})A;wzLQ%fR*_6N}R0k~og#PM~YtOFMT*MqWNjQ;|wzrX>Flo>O%
zilri~r5&`?6E}k6wKP&oqXb+=9m;4I#F*qU?Ytzimtyy%PxVm0o`zJ95o3Xi1<^+6
z^FAbE!pDdWMAks8gdNjoe@N`s>!kF?!YBo^sa!D41sM*3YiH+7pA6zw_&HCMOt2U?
zmm2YzBr=n<I2|K?1L?a#a56@G1&OR615nF;x3Y|9yinS)p&E&EBk{&L`L}<4e!9to
z*XBn270O+qGRy&WIP?nbE8t4XRf^jysdpvy5w{Piq{9SUMY$?5zlwTR(Jtclh$=cj
zz}1wi7W1oV*J|2LXs@AMjkvvrdezXb;`U)RbOd5dn3)ds*)r-7(`0rMKN9YvaG;oF
zLvTFwN7YCT$un+O@T^{R!|AHzFLa1mHI+7J(>M^v_C&LR@IoMHB4JHrB=+C0I;-dA
zsJ_u0$Nxk;e<D2u;ws{~N{rKqXSx_?5iX0^2>cwv<%n@E;c~@zH{o^@8#D%XBnYP!
z<1)gPiSYr#9T4M#ggYq4hX{8_z^HPjYW<x7H4vGf5tETk&$E*d;~e8J%y*gdvKBPl
ztnZ;2Q{-t(trw`-1!{rajTmIvv3_-cu>;31@?7L2yuH|Su?Q6q&jR8jtgVoE7K-t0
zqP|U9vJ;NKL)3S~_%2c372|tEeNT+<6ZL&Den8X@#P}gmKNRCfMEyvN>xjBej5EMI
z1FZ0p#4&eoz8<VFK5)$F%rn6Xt&|xNm~Q|pymxUd2{zvdR+t~N+z8A!ffeTe9LuB4
zH-i<X*c`jBm}h|%US&AuQP|(E=wdkj7O=k}{1Mzff~VkKKY`mPG0vtT*|Z<JPma%_
zAvt24OG9$Scnb~LBF0;3$W}4Vqak@>oKHjY#khcm6o_#l4Ji~b>V=)=PlWqKZ1Hq)
zd?I)yLI7f^;g*W088?9Ca=#PqcX2>&B;k!@7`~_r(B0x7P{8rfEnbWh!6^|s3++kZ
zlq6uO;qKcq@Pv#p#anCrnY~}9R@bQw+WTm$GhN1)+B)k33{l4tKf&UQiL#h@qmAS3
z*+klWYsS;Y@l(yF+6m}e^KTsmbd7YqM#6<mES<Wh(?Il+9KVLTt`YFpn;M??lfTbn
zb7>+TO+>W%SZE&$U2wHuSM>PqA1+RgENQPLnp)z9%jDR-(Y6wtFePKx1nRhgdI}?E
zQ$;rIfQE!Q9djc33qR&?-yPrPwmJL#*uNd~utZ=aZ7vN6s>me%nWVp<-3sDgLHeV1
zb>Y27EQI3<qjyvtT$40@sVT>oQ0o%f7H6!(T}r5r7!N3+Lj+t(m8DcA#@$P4PXTMG
zjh40(;}9+FE8sF}TSnE0Q8Eqb3mUkh1^YW!{M$CK$yjH~s^3XBoVqU8RDBTRk3{uR
zj6V_8Co%qm$p0X2_&)9>OYHUa5gd^-T<DR&A_~}Qp#yFr$MSKH8t@i2Kbgvssg1Dt
z$+TNC4G{2N+F>ts#U-M*VyF5Yk-Z}-oDs*TLTD-sK`b?tV{jsv*E1dIYcV)uEW$8@
zeokc1iKWnkz7&H!#2QL5?DD9q!{GZIKS?pkUO<x-lSQb8bgdyl!r@Azu4&W<SBKY0
zHg+m-snA{+TVKn;u~Q>x{PR^!ejXg}<mBYhQflGHSdUTT-JRp3tfSO8O^zRHJywk#
zf#ajCqt%$lbNn3ZIchux96#53t{QU{7E^r}^%Bm&V&BCgR6u+SNH5_;6cXP;F}_V4
zZxeT1hAyT$-Vx)w#PO~e-y@Fq#P~jOyf4NNh~on>en=c2it!`j_(+WFh+~}?XMiFD
zROr%mF;%f1RN|Pb$OM%*rYbgo3S);ZrYbgqN*q%an?NOwsfx{@630|U7O2olId*?G
z-3b<W;pSNV$g~(Nu(YI$x!yOyUmSD2Z;A0o@cIbd#WC0GlNe{y@N7C5Plqn%hUbWJ
zE)CBW<1I9Nix_XE;akNxkA~-oaXt;t7vlmNULeMWG`vv2s2Aq*-w@dw(pE5p6!1-f
z5CNxxZ>oT$hKeNG`|7ItSFr~cm;P<nj+V(>yS|Po9L(dy;W+WZ5)j9yfI|xSh-c)G
zDqwU_%wOosoEUX7?z<liK3wpOMd?e__7Zi%D4p?P{MXAZ)K}T!wa`fmJp`Ypg-9)o
z5^xpRR*3^(75G*`ZyXX&H>*+Ts;F2@;rMcpmxDKUv;JK8w(7(fo;i=bRI!)Z3082H
zD$Y_nba}W$wiJEQ6}60J%be+3B6~~J$kZt(GzEqtmKtiX_+s{_T?sdosT%fT7Exvq
zPr+Se6J@p-A0WyD#1pqAFD7->MOpX0e|ul?1hNr76YjIPWR(P7NkY;kHPnJ}ss*97
zn%>7hZ~i~ERJ`DT@sJsLVP+hkPW;owMa6W|Kb;H})|f+-ImAwkU2?=l$y}n$6}RUS
zmt1js8BvxIJJbosA0*0yVtj}w4+$7+UTg&kz$Jj4uq*nSHZHZq+keQZ?D3h5xC_*~
zpd*?r$3mX^v{2i^swN|*0GA@}QC|Yb9??%Rtv3>NBeBP!b;(~)6X}CkYAB53i+g`t
zv{#w(X2%yLr=NYxHY~1e2xkqKsMRGQ$Hh{*iLPs|NqjGgg8^|c5r@)U#v|ilf`Aht
zFcC%wI0=H2;48#ZLqQ&=w(qx2iwCNMGTEBq$=G-@PN+{MkZ%&mL;-8b5G@&j`*c8(
zJghqJ5B~o>nL2p(>X?$AZ|lBL)?x1_uZ%j6iDa&3sQFMHUD;8Y|DIpH3MaFFZX|6Q
zi2|2;ykkt4?&T4MfBSTu=<{MS_x@Odi4m)@*|0Jj1|%cC9&A{dO~B8==D8TZ0-IN2
z{1;fVG8_ghBUXd6WMw$SCyIv5G!HY4|36P77JeC#k>jDoHAxL0PjEZWJl*tTcD-+<
znXWA5dxiR;J2v7gsiu<JiLrkr4G?e@RaH@q82eU<->;^2)zn^SucrRh;`eK4+Zw78
z+G}V)4Glz$IUn=l`cFEzeL6qhtsF+8A3v{J@W5<=%$V)k{|=f)|K*j9{ww9Z9N@BK
z*kJYgFYE>@vTSJGry^nC1y&1Mu7OuQ8*2BA9KV)nxQK8?q#fG1t}e^%5HNmRV#L=(
z_FA|POAVixa4W9YC4KepYB$1&nZz=axC$P!f>>4%R~+HK-{sH`#kS;&iQ^0G{;K;z
zB~1JnaX3bN@N<Ih;@u$*T1u5V8|e@)#)%+L1V=m<th^2KBmqCNE#%W5TXbd<Jp%4W
zpfhd-#~%gvqXMq;GJe;}LdkM-cCj+crdC3TJx$x3rYa1vPxyT$-kP0x>_{F0t3%+1
zJz{RoE)BYmYQl=fDaJVnXuM*)2vw1;RU{Y<h<)^;t}Cf8UK;Ql37b?LaB=K|uHn0?
z)bK?P_v200z9XKy?VY|*^LwS-c=L*hm)LAi5z|wo3l8<+k6p#eXET0czVRpG{}UO6
z8^ZBv#6L}pcN67q;(<Fld&+kc`mnDaeTrK@P`q3=ieN5-RR9T23WJmcgA+PA32GyH
z_N3_fy3Z&cx`+Oc8~%_;+=ZW=grBa$&(6Y6_W$AGA20kNk$4I}`60J+);c}0B4(am
N9-ZCXJ9TuG{11_$7%Tt)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_fwd.hpp.DBDBAF89813C337F.idx b/.cache/clangd/index/device_normalization_fwd.hpp.DBDBAF89813C337F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d1f37b0c5860befa7793c1e4461209634e7c1972
GIT binary patch
literal 1086
zcmWIYbaOLfVPJ4h@vO*AElC7283cj2xTGjEosofIA2S0(#hktU{{Dv?1djd=tU4gQ
zgYkcwvxmxxj|@T|6qcvm;M<T{zWnB0v56n<u2tA}Rf=CtZFbJHziH3Te%{raJA1cp
zla$D>^UMVwriAZacK=2hR{_75BkMfT4gKp*o#QrJ-O$bCqj*@csqVtfE3H%4oxb!V
zw})$Mw0Czur&PhDFZ*70dCNZXe(;C$MTUNCz1QLpUjrXjz9qN5?s@mU!FRu1w5R@Y
zt&^Lj*4p`0uF!M8tnaR|>eE--{Dzq#7D|g1rmg>daChmCW}C8O(oQYrX*2iz?)A)S
zXV<mwQ!;V)beUs)Wtxw}@oo$L@7b3BS+DPYp3E(&oZhq6Loqs%<KU?t)h%wq4Q&<4
zOH-RdoW!0Pvi)?FN><z2aamwy$BppdZ9YL|AO8MmvW&2aRXzEC$IGHq2^;yeuT8N?
z3(L`TIW2fpEOzzX#I_R~ugp61YRTpE@~3C2yyj{8b-0VCRr&E=#*@i$=NO7BbCVi?
z;Vvh&@BOBVKyv{mHVz;c2)HygHLV$-97_*%=dT(emjswtVbT&T65=rF;z`nqopX*Y
zU}IutXW$ZL7nKDQ44h2BsAOQ|V!WCA?D)HmBq?qt7N7zXHWPU;flyGCnpS)Ts2JuB
zMozH(FoR+C1BF5M!h}Qogw$V5v|;7o;gV36(1fYt;gVFAMB+&)OCj;3m8Fq*GRiVY
zJXvK~B%YkI91>4nSsu=d-ENr^qo$(F$ipQrBd!6nIJJ*ewLz%Z2`DV8E@}*OCX8m{
z0tbtyN?G^BUCwnpTqZ^)9<ZS4_F`^%+yCDJ<QFXoLzt03zi3G!@uak*ka*Hs(nvfR
zEg2-9td=YiPfkk?i6^fm59h5lwhA~pH(@%^(V8-v>aZAqxr>nt991w!KQ&<e!p<+=
z%?$Lpptv-`nVN!{(lEV1pQ)P|!b~Vi%}E61ZCL0qz`}!}xG1?;ADr0Egx^nDF1MhY
VnVW}`ffFPI0X#4U0|PgL2>`P?YP0|V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_fwd_impl.hpp.2930CBBD3A2BF4E9.idx b/.cache/clangd/index/device_normalization_fwd_impl.hpp.2930CBBD3A2BF4E9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cba7f05b52d4e452d9ed10e85c8ff1da7e9f9e4e
GIT binary patch
literal 10942
zcmd^FcT|)|*WWvXi0i|$$^xq_?O9PoI*1C2T|kTl3u<hL5={gnSTUMdFcgi5qDDl8
z*pg5r7Ho+XY*f*pqS%$#V;6b#o0aD}&wKpiJvrz5<2~o)BzyTeznyaD&b@c;5E|U4
z&mL!mqUVI**-;axhZ`vrie~Kp*y%BotEfWZXR1)lUjO^BsHxSy>N|JqmGf)An>p~;
zOY<_yEhjy2HSPRG@Aa;pOBRgVoc!5Zu~BN$P0P2B{exFKoqGAwVT9s@<>P&?<{oeQ
z=E<$M_A%*#X260m%l6&0nqB9=X}}+5UA{hAxoXzoFx$??RfS{bPT5@PGOK*S)!DOa
zpeAly(bRC|sEl8#e+=;(`FP}=t|39)_wC;C<a6UC{+UC6ejd==d4d0(+D>7=hCUyx
z_d8P=8}w9LB)VoEIb-a<+HXwB@dD>R$CWE8lefTzr4yoUw0WLWd|Eg58S`RZ@#cgh
zH}l@w#=TBAKD4f^Uwz}_{nJ~&D6oy{HuY_8_cP5Ntvl3p)3wpB|FBq8^S8%0`|PIb
zgn4FnBktF2e>`^T9h>t5ZXI(Aq5C&(a-EZ&?mBJxYS#&W1n$|mH0q~$orOv13*ViK
z+_=W~;IpjJMRTUC`O1H$Y4XhMm-eG0ci+mG6Zb}yuIfH|XTTzr{pg?EzB*Vx_I2~}
zuW}s!b>*nzN$aT80k?LJxK+6Hq&7Og-a9JPt*{_w$f4|&l|CzP8IAJ1_@j3Z*OQ6O
zPHCqUZ_Nqxb<hsTJDf7u|IE&Q){TR2eiKt#wY1;xPNhv9&iJkz+uil@SuZcw>yynk
zJheDy`^T4=UuSKZV!hq<-ub0%TdWALSR2tKefEXa=C@jY8Kd7Db0PX^+ULo-@;Bpy
zOJjTAd9Hsr{l;HI2mT;#FM9o8Qp<`Wnl{lUaAm?Up+;F-*=)@;w|#Wo_?T6{E*|E)
z{n(aaP5m+x57)UA&7NtLY_;oNlcM*}TsHmm-TWrCx5q;Jw1^>Yn?}x`o^SL`=GNZ(
zp{_0{@OfZeqxWv7W*pJ2A5Jb@yK&fZ_PAir+*LW+#IQlPJbY#bSGj+oadb4zs_lC1
z*pz*1CU<!>x6{R^bY8c}Mtgo-VSQ_eX36!?ZE82y0edY9!nan9(}s<T&fXEev|!_&
z`jMn(Y$Gp?^XfSy{9uoe`@<hC)<l~R8?mFf_CcZ78QV+q8hM|JwQ#p{>UJpYx_G8w
z#<|<uXLh();y34DxyA2x4jZ~H(Pgi(sMWU_G_<rl?3eTJTO64E_E`T@Tee>)`Mv+}
zl-otKH%<9+*qXG}@XPTyn0}|nPQhwg{pIVazg^p4wj{ISQvAO5HP_9KMbs}&X!6Y$
zUr*ik>g(_VTg%%qxl_+PJ$J}0IU;|xw%LoqvQ~TGYW2PSfrm@~RQy<UKciXNwh}-8
z8P&lasrgUcimo~<1BQ<pbN<FJ0}7qnZFroNXR5ad4hR|H(k`#2)u`&}M#8fRmKl4R
z{+u%2dPd=_Wq*x6^yD)drG5Qv?bk{1o43`Q#onpjrih&rHQvKmp#Wj}-Wy@3=c!4f
zmJ0DCRiC6bl2Dj1Kfrz(Z@t!O`im!<Rm9U;A^r{ae?vP-AjRym@%g)(|MF}~G_02(
zWst>(4{X-luCnmizGRt-bY?x&gKa&yO9Ckt2OsP_LFRW=5rI*jgUxerlmt>NXBE)k
zVzt(S_^?6xx9s2g!v{9od>7(puQvWmMf_P0ufgUuw3Y-?+A33)HL843tRlXQ@&;_)
zfU6{s(r#FO+qe6MT~v|wj8X?Sb!@5)0x4c$0gsapJhW1iCTt$Xq)9Q+NdhU}VfjCA
zn$-8Qig>eLo`Criu#yB)d`mZ;THo_)7ZtH$lviN$3ff8nDIEscPulV>BSuBqG0IY+
zUP_!Kfs{^d)ML}br?5r)G0FmBv4A*A0x5pu-6yIy(5ouq!YJPp)9*<ONg%~PcXRma
ze6L-WgfdD#nB;?65=aSf*fDQC{B+ou7`0G{9*sR5K71&e5U^xcBtn1H+Kc<GaxNPa
z6Q-5rr0H_fToTyFfanLB4VlTzkh-v8788rbY_bgkDG__7y&Cke-FOwzGRh)iv4}0X
zK_DeA$);D%N*iw#v0;?^puG?Fl0eG0tJYqzYado%LCo1)`r7n$`S5|w37Qj+pKSm9
zfQmS?9^Ql5d(cV(DT%iRZ+^4HznO}3V3Z}qd<k)p1X7koWko*MbeXCqYBtC*>TryD
zOX4GCeeJP3M>;c$RWVA4O^C~f4{To5di;{$+B#1aabtr#2Hj)OO9CmYwRSGGk5{Tx
zq$8um6Z3e|RuV|r&}?6S_sIAzD&oN?Pr?2vGwKF`l$;)^gL3<>UD2F0GwcnhmxG%m
zkWvtR`O@_-wRS4fmi4lbSS(}%H3+2aUliDG&#d}kO^Gq<B?pu_ppgVp4*WcI?fo>D
z02T3Ky*z{#55YzfNGWWz=5TqK_s=Tg&M1$-<`FyS4dNqZ#NNbRQ!+eM#FSAIi8_&3
zNCGJ*j_!ybV%H~DMVuJrCD^<KXGtKXY@x@4qPW3zD$<EjUVzOD<_{VKQqE;ZW`~`A
zqE-_#wxV=uluoTAfs{)XX+}}!$Bt}D8Z$~ZD6_#*5=g1u9#=c1q(jH%q&cGuQV-I9
z_`v2H%RHy3JFYFUAWd0wKbw9oA3m`8w(B)d@5D8rB1$&RB%)4Yi)|1{dFv9gc}&@Y
zooZsCQ-~{wY6WQ~38d8L(ngPFp8m|3P&P;&G|dBZNg(BYbk>9CYaqX&mqMZ{WW6+q
zkCgar+sCB-YoUs`vbnqi;T?#QKuXNS$k_dMJZ^weiWVOo#DVtqoqOm0ZY+pnUB(9X
zJ$0&gmDVgv5GPtq^g>k$;tZ=9G87MO;-Lp`T?fKCXoF@eh(7`QuM1)fNJJ0|K~)GX
z(0?hVu!Pz$^+;IvWtwMoV?hk-6V{LSKfKRy8Tw-I7bAGv{BHC6aP)1rZ)IpB_-};n
zI1fR*3GHrzCk~HMkm<w_0#PICB@x>X)m^eR62$RN<GXT=M7c-x;V9ZAx+h1e&?XfE
za3l<%Y{e%)bpq^gNTn1;am5%$m4bK%gfpN)q99g6P$dlH_#${;1b>dJp;a}waC`+?
zU6JD&@UD^buR`Ee7{K}0An=+T*Fw8m@Wytf6wGm=7|f~iLl8$@3E}|r0d^dXZZX;s
zH$z*MZ&JUlzzdATlf>#I(V|6T<fnpI0zwH`;Glw716^vM4`QX158x#1R0}#KniO9L
zLCm98d9)Qe1c&`GlOBB6c?laTk2cSv7T6$m-L>y&$8zN-b7=GE$cE}S52mrMRua3F
z#19WG<{WJ6%0Vax9iL1?0)msl%!3gtK&Sv)PHf0OkmxeOs%~z={!`4JrfH^i;5HMh
ziIt&M;J6Bc(1HXp4II-rb{jAu(CgP($!1341>$#sbjGa4NUSEkt4ScoHKccq9N!}C
zZV^}BewX;)C0#haNBr-}@iXH0Oy2)X()A_j&iQYN`&;74@jKG>oxDAsc8aI|ygh;T
zN}zo>UPj%PQ4fw&Xg~_>#_=lJcNOi2*emE(l5vX>zD*1Z^{FQ9vB5~JAwD&7e3v-i
zCGB|oJ>q;%j$abzmva1$IKPwQ1nQAMWeZE8&M8#Z-zw^|ihfiij>Ei1Lr#gO0Fy~(
z7{@_%9GY{(Xvi$lFnqckS6-MiQp0TUJ(#=)6(%~EYqA4U20|HVu?IoC0&Z8p53y3p
zK(R;6K-m}Xfcg$NAyE)(L0!wSQp!v5iya-uE5gmj&Y(u(LI_z1AHA)SxCnYLf{#Ac
zNX&$gOz4Mp+VRr*Mp2Kf?ARJA!LgD%5X^MhR4#&Wk)I=j7gpZ{dz_k5%71Yxezt#~
z{>nJ~Z)WKwR4Ab;yz&^ZUkUBZaVZr_<?W@^yOg$<w+}3(p&XY{p-j#%qwZzYQ{LXU
zjP~QWoC@V~emV6ir@p+sf(jM#_6q7zLA~Vdp%rw99LH49>2kajD(DY@|AMJMvz!|s
z+~E7Jlr8|MBh($Uy@u=`8_)(>gCXY+ZZ$J`#r~f2yG`vv*A|^k{A8B`E7$v{kL(?5
zE{Nfpa9i%1#%spQP=qEzh9bQqyP%U7#K@tMV>z1SJxSI!*?Y1K&C<-W<;*#nIWiQd
ziIbu4LHHglIhRd5QI3;9NRs1pa7qU+v`Il+4^HdlI0KwA<ai4RTfma@w}P-$j<<oZ
zjpLA?(&ndS(9>*N!W_f=__T&Q4*!Tg>ySFt(sgMHQ*ltopq`jp;5CaVrOOrW5X*(A
zeIadw*Al#Ju|aA;sNv^DDP6g6iw8I9{d?3z^$%>IleEc6YKGP`q<`}^wJSO|vKGWa
z?t|L%X@<Io%FuG~TMpg$+Ec(UMUL}9mk%OZm>~WJy5Hn@C+K#{@h;HqlH=W=+bzd?
zK(|MZ3qV&O$9qAySB{^9>2om0#7JOQ0Mi#>j<*AW-5N|^f;nC&1m*!vUx7KgOo62g
zrmw*qa}0si3rydDIfiwC6$eb~z#OBu!0r$FblMgZ24*`{zg~_rs6IoEH&FcsIo?S1
z8|8Qt)o+sHOsdb6<1DJr;&@ni+KBATYCpD{=cxHPYKgl!-28;+xLmCxvxkwoksbMN
zu7Q?oz!&##^tuNR1Jy6G80|c?I}ZVT<Ep^6iYL3H3p!01)w-3QWg&TBlLu|lo|qvL
zZ6Rre*_pt?nZB5~V|)|DbyTyC+M|Z>%Fjl;48mo3ZrgD2$Bw5Q-PSfe<HhIyIHts=
zT011JFWSU(;?d5dJD!->U(Oy(O&B+i-43#;c{a7e85yu!HuaL@&$8(tj&rD04z-qJ
zuN>;jaW1varFL@MH<$KDJbS-e!B%b7`$mQ&U3UiBa~&6hu2_ysKvyEirJyU7<1)~d
zag3!Arpi+woZ{KPQYw(3DzQLfs{BwGL5&Dvf9L+4Ff|v%sJ>AnxC@W&6+M`vROpck
zgV1aQ@iw%$&GP|6e&ACJJ^9Mck``x)6E}hj#PI_0;rJPGd?v^7v{gLy=Iy!EF_-$F
z3I#Ej4#}kvh%wV)E3X2fieCYgQeg*Ij)fgYyavKGaKsrZrSc9Ell#)e50}61C<<a<
zi@rM4m>`B&gy@z@g&jfcXVFiGH+<$NeFw?nQ>bsK3=M5D)SjQZ<<NaO&&CBY1-hrm
zaX#qt!5JrK@S6JH<aj6OcgpcD(C?Dt-Jsts$9q7(M~(|XUm(YOLBCgypM&N(Sj)Yp
z<^@>Gy{6_RSj)Yp<`r1Wy{6_hSj)Yp<_%cOy{4uPthv`rBqoW(tWc`_2&@8Pl0?j~
zI3lq0(qt(y!+g}>Wu4QhCvLyN%Q~-@;|%JYA;%l2^9DKINS!yz@h0lLNscqAbEX_;
zQRgg<lkfHVb5dkX{eRshV(HU;M;`aIyZt|Rt;4j3C8`ajSoCRlRE?A$YOnau0=t*^
zEreixA~V4+6N0%ztAb8d5QHai#o5{?3l7Z9{l~!vi@eO?90S!cxo2vq_@cu}yY76r
z+aH_nu_$$#YERQv7`fA4jT^ntOn06|+7-lg1?hyP1p^LQLBcpLBHAK(<Sin;MWhEt
zVk{=Jsayr&Do>J>Qb`%7vielew#p^*T^KQ+YVzfgtdMF7sU2!N<4s=t1*>Emb}9QF
ztiA_#uFeFoN|56%V6_F@QN38SW+N1VstBx6&q}FujZ?{*pS-+S)uH!4*8Aq3)iynj
z)}{Z)HEa!SvdS{>-Velx6=>|r`9;!*)nV)^IQ}{8Y=E-hD_6EbkpYoIaH-j|4>TGW
zmDNWO#0a~HHatR33Yhfqnk$Ht118H*oLyWS>_8CX+~RyWN(8G!u#@A~iQvX@5?Ce4
z+moPmlDs_wyfeTbH7$r+!D=hmal8$zwsD-}?KiM6W8501r{zSqoV39k2D4YHImY)I
zJ75Q?;RS&or5s@I0NwC?z+hcYS9#UV5NVyRalHF}+-|F+r@wqsL)+u~+uFGm4}w_3
zX=;9&=dSoT!RB}ggiCx&l+qIgTFbusu5&-rbzNg9wT6x##`hv>Y*YkC(PN`!C>6p|
z;d5Lbu7WMI;faGUGoF~lljeAe1eP$EEFsM?qhK+K_@|TJcp48V5<@Gq-n#t1XL<O@
z!nAP&R7d3XpoSM0)W*S+=SSrDKfC>jq7TK~d^D;)ZXj#OAr?8r3g>)S<+t;)$Ce4q
z#4Dj$B~KuawwdqQVOQTR|M(al+deSbw2#T+|9A}XU5F8ng6b$(a*xpPI)tmj9D(_q
zO0cMemPj<DeqL430kKkgGr})ovckSt3PLG=HW0*Wc7AwmTPeLHef;8I>%PmV^{r4E
zDn_m))EzyukyuJ?N+~N=H9VpluwSXXqEtq$%cu=+FQdL?v?Ism)S;Z}Ij*3#71Umi
zJ6F&EIUZd>$I5YB1^otbS?AW@gbvWH{$#^3Z(~a@2BDY-Aww~(%SC=RmC_3tF1_N;
z(W(2pp1aF>WW{`z)1z`R^J4L~;oS_6N!9IyVOuNP0^7%h(0m~{amSSj%`?FX588$J
zDy{L1mVf@wr7HNOXsnbT5>a&mt6Z8aAQ~)nG7S=wg+znyPhc<MCX0v$Z_EO#<gk;a
z!BP*qSrh+tBoI$CYMu>bc$MVCWCI(yJ<h{m19lg|3%AyA9MsjI=lBY!ugGx?*wt|U
z^^RpdUuA3I6RX2}5qsk?JbUu-7lF+paKY&rDh++k!%!UWmUfFv${jN|W<hzxJ&$<f
zdIWI?ao@o)-VQ!^0$n9IU=M~KyPk($h?UZVDq8hzk4DcMn^`5Z5!TTr>sXycdI&be
zye7w}Ip#NaUH0X8+x(iv{9z8!<`6fW%G34x4!Q4eX!XgOpQ-<PuqbhW&wt&`;N`rc
z`Ic{wnElsH?A6=6Y^!@w%UH|~w+r{?QF@l$EN^Zjv+ZWfP&#<0Lm;}6*WS~P1?3iX
zW4rP1@eH?%d8vkHb4+f~5*jO|S8{Hg|6Tw$<k;F3#2$NMMN-o4B=N?an#B^LDkYY@
zP*F;JO8Ik_A%nKhr%q^L_&qkGhPQW|(R=CboxPi%2jM(V&JD*i=pu|kV>TSmF4ZuY
z<15hl3WRc813@*MuasW!amg_gr-b)nf7aD2yXg(h<8Q+Nr-T(f1v0U2ra<LjuBgCC
z$Hqj&&NBRsSS!Ov!;d}wqEI+~{FR8eH$JjlytZ+#Cav2z+B@i5+1e=_J~k`b@Lz5D
Suk7y(@TWEYqEP7hum1--qwqQa

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalization_fwd_splitk_impl.hpp.6FD53387232480CC.idx b/.cache/clangd/index/device_normalization_fwd_splitk_impl.hpp.6FD53387232480CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5de48554c525179b3e87ddc02671317aa2a146cb
GIT binary patch
literal 15510
zcmd^lcT`lz`}fYwy$Z{9S!oNh6zL#R+>6pwY+%9OThyqD#+c-*F|iw?0tzZB5=B78
z0-_WZRFtlQEjFasP!J1tv7p%AA-j)re<$a>@5#yg$M>9{XZ6{S^V#Y1%u{AapWxs%
zfihW-DShV7oci-5P9~GJrhjMuJZt)}7BX3HE17I={GX#f|5)a0(s<#0yAQ;9rb~48
zWMa38yMA-i{JR(ZQF;B8c`K*yaQaV-?-P0LlCIZRKJo2;?fb{OH<V=MHg>u2=Ka03
z=*Pql=5xPKSkIMmUrm{mDWBc2qUWt?W(#+Yj@t1&;G4Gn7biaL`p)~+!nitNjsAcx
zcS5VSZnzOLS)6Y(cnO-E9@H;>yT!=uK@YypI@0>+R>i`GJ1Mh%3XWVpvv^V4r^Wdx
znG=##zH44ADE{dt$<1DBG&`4Qj4zwrIIXzRN|jak)zjb7#(HJAo#Q-LMm(B&AU?CO
zaN>E>-YpKbh`m^uw{6mao_+=$FD~oh8H#5`nup5sEq87hxx4KYhjv3f(^|HgvDy4Y
zafH}D{MxL+9&zEf`t0#<I6p1kEGm9-SjN1Y-PYfmv@_c6<b-L5P7Qs$A@yR`>#`wR
zN4$u5&}vQJ2mkrerR$^!<E^89+W6dSQb>>d*)#Lp7HnVD$0}*V&2e~pRGquQqNN4d
zUT#+p1RknQJ+p7p@I$?Z&s-6Ea(eXf9zPy6%{;WPW#agOW4CA5?#nT+zPiHWa-7P?
zeRn~KgOk4Tj{1x&H=pW$Cn8$+Ut;}RTG_r=)35l?ET4FJRk}PiEc(!lVF%RF^Q#}n
zhdLMd_nq+ON$JR6Gs-Hf(>or>Obs~^TRtu44^kfYP3@?gRjbym&RMrG;8=Fn*o4Z^
zMS0o3>^heFRo49PTn9uY<r$}%w3$10^~n06^YV*NC-q%xw8*`-osmfwKf_L8N8NX<
zw9flx@!E|07DYv)ht=f{idwPx`VK|pj@@hfm^7VuFhT4rzA7E_c}~;jw^85BzvUe4
zbNa)O>A&r+n))egMOeW@FE5YKi#=mI1qz1OUtcS_$o(`w@!;+XE@RzKXJ47**2fK0
z+;Fds+qZY?cTTJ2oo_z#j#SV6>|IxQ)42D$XPp+tCsh7=|3;4Uu>R`gc`tvx5V!29
zdQ8;llb;uF&r=Poxb!q+-LIPuXWkhf-T&)cgZ)2rDzLPA8vfm+;y>Fwtu6fcUEYdD
z|3426iyAre-GFh2yQUP)GR;&K<jp9n4li@NGQMgbdH-Tm@}0|1PELHLIFz&3e^l3s
zF*Z&!C*P8P*Z(WSv5lsK%0kCZdD1d%*|qZ%&x}(Vy&QhuE;w~`MC+LihC8lLv9|R6
z)3xKE_Hk`4r0Azzzj><DkzRkC3G;c^GrXk5Rdw;H!<pejpL=XEwX?Y$@BVIZ+M|1G
zhkNze5xv^uj)+yuSB@Jnce|73g`-nR&rqR8UwoQ&@9<soan`?|MYk=J4Aw6_yKu-(
z*FVkmU;Pq1a_)*kz0*@Byxi`hvRLQd_R_XTpZ2~w{w83+^ngoC)=w$9=Z7cfygjo2
zt+{8!wOh_-O091ljUHZ_64|Eh{k>BroteC*WS9TTaGyOf_3zqsnh?Hi-OOz*vSt=9
zSdrlrBHl42U1|dF`-DbbRgW{(e{i(4aA$0vIW6*w=FbhPyer%NynWV?W1q{{`8<qy
zy(qKOn{Cyqs42@EgRR})A2llcs8PS`wcKKzNsC&C-yR-ax@p2pjqmT>PT6mL>7#ga
z>am~u!IPDr-`f|owtBntW%d&Frj2WI_78G>SN%Hvxwppb`kW&fho1{wR@l5P@C+K)
zGAZ@l>(zx$*MkZhV+$wDP5f~3t@Yefd*a3?s!pB1x~IB)qT%hJ;rGdr)Kzf{Qcv;t
z)?E9!>5tA&54|yO4T%r$*J)R~^=;#CPfG~TdGzwp@Auw@9o)3y`%axgKcwhIJo5DI
z@NLR%6Xz*24&N~7tBLEfHMzdWG0$?B;$F2EY^E*!Cg|qqF~jZ;D(&;oyVtD|MR(ed
zAMwe4pQ=G#ALx0emudTGulaR(G3a)#X3g{ly&0vO2Cg0(Q)_E`bxBU*rjT^g>8|US
zUl<T%`k~wTPb1%3I2o3=OwF5`e%I)3MM{D!u0y9S$27N=l(?*(Rh-x=?1y02*^z-7
z#S+W>`&~UMS1g?SQ;F*H?~%C;wl<DU!PDn8%xe4DWG7xaX#5(N+DW&`zFec$qsC3!
z(m^0ucjsk9g{3@l{Ly{>k^N<(z7ld%4o#c43N1PH-I<^FRA=^%Xjg5U)4%`FypzAp
zKee>s<Iq_quN^aMpKY{jTi&|;;M_?L3cV9sj3T~OyM!M<=3-QL$~~!U=ZFK`%xCZW
zmVNk)xmZysOe?w7adv)1>b(v9d$%e1jORvP+dq5W%&(pGWitM;!-AZE&x;JOt-DN2
zLh>YJsuMCjdI!-zh6Q@he7xJn8DLXN*^1>`v8zr1MR}oR@mANQAqqUQy-ds_R(Zrp
zCxBw)*3^2}Z)QVVU_)BV4PtPE80o|p%8-whPkV(ARpPcDGO?N{s!3~|0E+R{xV1l|
z;BKb46|H3_YO@ombpj}+l?yVCeQWSX3*3@YE@Fd=*jOikVrG#jeE(^{5+iIyDUC?k
zh`e+HD3&8U7wm9#^-$nYI-;XQag^BV1W;^KKdyLcIB13uwx+c_L&|5UgH8aYt^2(j
z4cnXw46p;GY{Bv^*g_|O;%a|)l;^JSDGKaD8~T%&|4G{E1W?>#!mR4Uwkj34HKkM%
zWhF7y37~jfO4#1|s`aWCSg)N-j6f|Tkh4z6^ji=onGA`jdrsJ#b>1N}%y3Iu&t<H4
z87p-H>gijlw{7|_TW2Zoa7wvA+Fl@?bOI;?T+H4LzQ)-paJS}Kh|6UXsuMsNcs|tT
zKeCZy6nGS+oF(RGiHlAEWzd`r`(0(Km!{a9&P)ojN<r>A0hA%{?NTRN{2Hdf18FS>
zh{XY7r4v9IYSAI@$o#|(3fzTKP7}k^#9SwUGR7=1u4nAN4GQc`DFwu;fVk=eP^S8<
z__^>@->(#S45geP)+dOkP5|Yboh}nnx&-T+Vk6q!J;-Vg?XFe;WyYXWuW#hOk5S;E
zw3Z@bRYcnB1W<l0`}`p8;x8){*o;zgiB&Ff)Cr)>+mM=g{!XHq0uP~-Vq#TH`=S*<
zS(uaWFnWUMuD}B*C5I?-h_y}tWpT#VGOKr%?n-PyyPM8)>AXrOfU=}8+1$-{?kFX0
zLn#$}+X`Nz6F~XhdE%V0$E@`gIFL4!PptBZn@)V8^sM)8*QREv0tZvdDI%OAb~*u+
z<zu1}%~zhMyLT+5oG0ez>1NOhphR>}D$TpHe2gg`O#8AIwb_fDbOI=A>-!E!F1fW|
zfrrsrE|RtviN8(&WnJNsIWudcdmCUAT1yg^Ct+uu07~N4ewq*F_opjxcUsF8;&g?C
z=mb!<ZMVO4q;S4mfxA-5RieB~yQ>vI*?!U>^}bQ(Y$aCGhN?OJYEG>aKuPiqu`^xt
z(+36iq_t!c^GvEfv;ru}W3J~6+ck2%0{5kq!^HM5-7i`JlzmSh{^5|YGgg5IQA##3
z&L%3I07}~IOrKvXzwTpz1=^P^EYHGDIsud{6!iFN#=|WJ*p5<mVfij>rxQRqMnW}S
zC0sWJ?nfIsL~IXHC8QNV$t_haj*1XNlvqw{$>0n!I7^)X%JJN^D0knuVg(*eYbhjF
zh4dEC3ZR_%cxv(qzVQbI?o27?i1|6<t`k5x`(*ym)$_g?sl@iQq2rwXaXPzN0hIH<
zmCv8oswCDF+fYg}vP!0>Q7eE_PQvE4`>}Yl0teBCN{CQG%yi-lrN^0H0$sZ`7+^<Q
z%Q_@qhwO9$DA((k8=P`dFIV9Hw3Z`8afI0D1W>BS+x6}<>C;3(*3(TUwvn~5`s*K=
zK8Ix^T&dQr{e8gaSwTsA6u1Yiql~zeQGKZuP)BW^(~TBe*QO|NZ%WA`>MUZf6F|9p
zZsLXXrRuK*Sufg*pUltduYY9vJkmDvpigxD%e8f96u1wq;~=p*NOg-=Kppkpbt_bM
zuX&-sDoUv!$_jdqXa!InkN%XLJA4Vfu#D-P#AA6pz3a6CC=KZ!2WE_JP#9o?=E=wM
zWO}D*1yEkS&zt@!@~huXv4T=|p*FkdEv*$md2PLFQDVQSMcTVQg;S((twpT>N@Mc5
zKjt@W$TY&1w3bgu`3ZUG#1~55ytl*Ll}L&8Ddjq^yv|$c1W?|pA5Mv_8>pwiA#~8C
zL@1^6q7^`CGN@fVHTjnSBdn&iq$6cIveyZq$YxFbW;X9GlaW~q0{o8O>&gqFt+lP=
zm(>?T975WCStK!xQ-^W(MmmE=5aT(!c+QpKe9kAI3xqQTMi`RoOs_B3`7s~cz+UNj
zEd((Sx#yvf9plHonz`zkhaetDs^h2wbX^e3kg80I>yUFD>dwknVqqnAhB+0)_1JDb
z?hLq<S47-*8@;dV3Dlah*@+y%&=Tmn+3Z4&u%v>Rje@gLD69Ao3O*#oC(+Q8Xf!Lo
zfr4(JV1{p^pqo;>1_!Lc-Jm0e=2yCVl~vC`oS4sb$>+Ml9Dx<e$#u3T%n_X$?3{;#
zK@-xcAbMrQ8P1&`z9N3F$S9b6W!U7Df~ITMoFEn=mqOH)&FNX>c@_mTd=9yvLp>N?
zik+9@F3?pldWj%D$HH@L3ig#CMiY-{(h0C!X9|O53Z^iY>kMEvlq%k~ig$*)8M;jO
zh1ME|U&89p8niYrtD%X`E$R&xE;MN%*O|bKsI~uqAkY@ftBEZ%DTsTyc6&JuoKr#E
z&nfnEX0RO1ShJrCV7QdiE9I1+y#=w9b1vmP8NSUaZgXaEcM9Tdj(!3eKEd0b;9bCS
zfdLS?&M;>k-+^22z}|4%3*v5~+)cb-vdt^^>wQQ*+=1ih2vmDHa~J_9W^*ptoGbLW
zd4sv<bL|;E%ekE8T%pI{0boIVh!qd9Ewmul`2jF>Hj1TwEA`IRP+~2ytR*gN4vTqJ
zG4B9lvZ;4Ir2ot9u9RpYvyf}OE166XePzC15O@{5T<1%$fhXYB30MOvS`fG4*4wZK
z4u&8m5M=`C0NoVC?L@hqVem!t<vI_9Rh)_gQgL5U?1ET;oeOYBcA(A@y|Y9GR=q$k
zY`t?t1$VhX-4(q*i3&U!fm(NZ=ZOjoKS6vz{2q|rY+BwBzc&oGJvwXJ<KF_qFh}Rl
ztd>{+krRt~qhj6^=FhHcR!)QOMfty8TKj9t?_a-P5re2U!+P1+35E(opewY22pfnE
zOr~7tlEIq0Pf;~3h&XzhDv=Aw<O1r%l*A=ua!HDpVaH{-Hz*B3T#g-=GYq~Rt*8;J
z8*y7`Q4l|3kB?Zxa2U~q5nsS^od*a#YCq*efE8X;Oz5#9j<<<3o5kCt_ifUb;Tqyp
z!*;31z$qae_Wl}a$cYa*|A*{Ia^e%N*Ap%TjwmNS=XyPt;x}BoH(YyG{*DWH$8~1-
z0~hc?ikI>J%lI!xo)g3Pu3>yP#*gGZBYAIz*YI7}NXs|!ejE7!R=$bvxrq;EIDz*_
z;Jp~$#&_Aq2Qi$?_f6(Q0mJ^M+Ewc?_i^|oa^2A$i#znLTh+p1bVvQ>8K55@_(k&!
z(2oSa<J{hH?O6E-&h3K~hw*M<yrg^$@3uyYH}PJZq$A$OyKR$}C-d&f{FfOPJD-qO
z%^PDvZ=BOa?=(?>|0_^SoXg<kpvwgDAg6bbQvmk4^6k;FMvL~)4WEqF$=Dyxy07Ec
zvPp&$=FnYJgZkC5dz_v{)VUUQfqR=1qjA4z918o`_gXwZ^Qp>)*4T(08`(bfw|;SE
zMvF-==ytlz3%7Yiqs~Ta#zD9F?hM!PLXEV%hS${aKGO0*HT)2UYk8qo%CF@;Yk6;J
zdEZ(-l;JzPa7W6&!+YQ1HLSdj7wV+tb-Y&{-$7bFq>djZ#k1=8pQSho)$wZ({7=x2
zoyn$>%PapjCvaxzPWzv?Zb(^tz?Lz+W?N95i%bfTHB%#n$fQt;i;zi?6rVtv6R0yR
zk072v!%v_I43{8HiL|^FX-cKI9GR3OYsRlcCY4fLg-oiX_#P7OAv4B*fP@E9T#tl$
zDQ-Zn4M+pGnIJZxfCkiy;g_hxOBBd(6LM`r8o2fau?Yn<p<WEXM;+dyK!#UgmsPj}
zY)3(i#4eFi9EDw?q&Nl(G1!dpH)3I<6vtvAmSKp4(TiXc5jGJ!SVFlj8U|{&-;Rxo
zHztrXl$ebivynfWp+m^=5X1dDt^d)ieN;5v97)JG3H1P@AL41WhIk^xGfReUmcN2)
z&=lOil(?1%Yl)imNgHAVBE;5O%5@Pn&~ySl<o-3N59?zz_K(JW*sO#T|8UX=DsDC@
z`X-Rk3?~sy66wxxGVx9(L4YA*M|)U{6}8w13go)D9rSRhey_ZlQ<ZDzWL)R<uk(g*
zgg~Lu8p4SXPQKXGa$Q&tY8W=a$oav#K;BZIo|<Pm@?m4jK%N;={0aGgLP2bdpOOD(
zDZY%=m+_YcrtSk)mr3yztiB?}<yc)V#T8gxA;pzgT`9#?SY0K>SF!r46z7swxx^S^
zgVdcNt@4O5cozbVxV6eB#t@97nu4?{AjS~aq+SeZRY;5>_#)7lW2+)!4B<$DMh{yR
z6JrQ0()r@;(|K1gs_50h+h<5|CU2i9#rt{t{Zf2@w?81o2YLI0Qk=!xXGw83Z=cQZ
zh)Ekq@7sRepDz7<-uOOm21`H6xYT=My165@gpMHf5#-BOv;heX$VI9vgeK&|bU_$y
z6NWp&3Xgif$@|{1W#ZplNSHEueZIuPOKc9?QxIc_M-1@?EY}5yVY7{od;BCs`8tVq
z?IUjU5w{1TAchf}FlJ*n?^>I1DPB#?Rufm?k3a19{WvEFd+N_$#8wxvE3_cc*oS!v
z=K#)1^M!7g!Fj+1PW^q(>Ns;wsJq75S96|Vc?e<}Z<5B_!m2_@o1UXbSn&u;3|Vb_
z8ybedHg&7D;cTWWZXxGes3Tj2kJ#rU?g>L_9=d-R(-pAaXp66~@Cr+`lQtF(Eq-5|
z?7Ai4_1J%Ga?D<D*>+`o$u_!kB9M0k3W2W9PQ9GFK4Rjcmh^1!#s$0zW}z8-6!0CS
zxPJjZnBhWRRmfXPafd>_Bf~|!Wf5;9#eIwT0f6Dm$_4Qm7M@`fHaU&d#K(gejwTM#
z#GT>w#9_S@#}J1YDUKx$u~NK=glr-M+3z=#kj+w@M%ty3FV2EomsNl{pL^8f@NV;@
z&opZF46C1EA6C(GtbQ)VFR=QB6gOaXgA~8S>X!_IR-vo@8Vj$nHB>CuB`u%_khGx0
zdaS6I615o&Ai~^ITS=R@fK`T|2rbZNE1&>0BR$p)SZKf&&<7giA|C5W0ART;c>zuR
zu1j9%3*tL0yp!tWk2vHb9uGeh#AU>28F67ajD&`fuNjUey`srjhS!rm>&XO$V@U59
z@)g6eq*pBC%XKLe=oom;bh0)QVWae>*CtMY2#FIKCDCS0U`_;aA2QyDY#2U<ERG>Z
zIL(oL$?_*_Kl`>9M76bA62|Oi-OZ7OF}quLm!KZjJtQa@nI<DOgyN|JL#8Q64Yr3s
z)rsjIqz3)cd`)*uml~iMsAGl{A4C=hku7wPZd9ashWwe*e}=j|lNykX$h#5wvP+;5
z1va7{41Yr1J|T(U{TX%pEX9|x{blUNmZS{Zmr3yzY=1?H%dvgA6jxyT3MsC{_LWjx
zh3%`P_$s!)D#fda)hgn^+Rr5>xx^AwO0yYYl1D71W`IdPv4qXuYzCMV5KFkzn#}-{
zLSiX315ApDC8*_QGr**nSi*r7Xy9M3f*3+hK%n6Sy-H#T@d$zDy!5UTLr90w3yagc
z$QeR(jQ*a}tB~S!-YuQ?hU438R=8zIaVGDUDaHGFxBXIlfOk6}#Rqw}gHoKuyJbmn
zHt&|raOA_D71O?%_2IwHXOC*gDSFn!=I;Nx@CsyXsS?{vgw4bewmJ<(|CPQ4B?ggP
zYJzB!woGZ|W9NL_6QXkhP1WjMA&!v6rFRbTdrd|_WNY2!2hW!tTaf;bEsKpcZiPYl
zJ0DWZ;wEo?lXn1{bYtVh@xL3Y%c;FqP1;uzKWM5M_p2r&8Q#O0?~&SXd$^8!xbDoJ
zJIUFfl!n7ka+;G|S2#|wr+X)tg)MfclXZeOIUzMUs(6zs-Ue>J_=I+~3ybG;`kQf^
z@GfITl`7JT-h%g#>K^iBn)E(W-IwARtct;&U@pUK(x!J2VHf#g+{$&Caabo%1+;~|
zM7dXb9i|aw8pCp3z8rqI{fF6u@yPOc+QLg@_!8MMrT!inzDIVTw85ICSG+c>4m$zT
z>9mF?Sa`w=Ex9hK4oxL3j$C_A(W%!z-U2&&T(|08+_dHYxd6L|=*O6aJX}Mg)vvJ4
zE0z}Es0C#6nqkQA(;Dw%#eHna_P;jE4}IP<@^TmX;Ti4e?mL0HoIt~1<@U`z#t)jA
z6l_F&I%HFXoLO|J1a&EqCecb!mr^ONLN-;%87dOQD&$cmwdEfm)dOV1${!%72grlr
zdZenCme(VvdTDtS)<od|*h17y!K#hehT&MOik0F7EGFPUY3hPzF9KQWVh8T91N(y}
zXpTE*5}4uEoXXH_m*U;TdpGfe8HINebgJ)S#a(IKLHlR|raDdIKj>t_#6KwU22tOj
zL2})zG<u!!CKc?Mfy$u?Htm}V*ooAiB7Tp^a5(tzxPqo@v@a_du@E~K;;v8=O%xEl
z3&a*8OY}k|dKZZ;s7V?}C4P-$FjSoVXnwY8RYU5(PIesXcwp0lOOJa~jd7DVzR4W!
zoJZ{!^jCNNgL<`vXh<O%!M4{~G~q0o%<wrh;hYpN#Uq#Ev9R37WGjc1CB1k5|L$GD
zvlO~q_pss~v+U)%$0@L`;3iYWp?#SGtB`xDd~{mC%e()277J$Gc{=XH{6SQ)G!fG#
zqJnke#0btbf|KBUPL<DDK$i*?{s(V(?V3VQcO!1y$W&Lc^J4E#S$%i?V^wuLWzY|;
zg7ucuS^2-;6<`Zeh5Z5xFQf_TSmF^YO;BrJ>A-ZBUcPkp?88@uf4gK#>*Ib)c+l~Z
zyg8V!x&1$b^;R{!4P2ZwoxuCo@PQ21@|Lx{l@xcZ<$W2x!?(S|+cR9pTi5ZnQXE*v
zcah@pb^HV=o?pi=1Pr-JI!Eh>u#QNQx7tTXFjTmJ>CVx<F@l`~j!i4M?qv}a(Ay8W
z?L~gzdkNxM)b}hJEj=wg&!M3VFU8%L;=yoQ1TmFZq_W$RhHeP`^Z;chh;^Jr9p?g<
zAx%GXjt@DJ;bpwzGG2l=@(vq$4J$9=9gBF8S>i?fup)jk;0u9H3x^C;$NuXhap_+1
zk4L*+xleZq4L#67zzE>&l|gfiNC+clP|zG6>mN==GQ66&t!6)LemLP5EuD(>r1N?b
z!uT;HFopy(983ISrTj!<naC_nLEJ$`?jVyGKZOiWVSc$>_aY27L%Hrn7+rJib1=5%
zdvL%WJQUJq&3R+zY^id;NBr(F<t>PhiQi)>Ue2kPb0R1WL0rM9S4eRjXBWr$v+^yR
z-4-c6!g(FxdV^9C#ABS-F)6;tb-c(8fZg2eo%vUALjYIajTo`J%oR~pUxSQlkPEx%
zqp?vmc43?I66bV@3xpq1w~%wI<bnWKt*A6_F{e$%|2l7`>Q(%ZX`5Ds|4k*r6G5tJ
z9$>`-X{1&AUJy1HJQbrRMe{4c%*1vDQ;)7YM2|bMtRud#^fkU0dNl4cNB{bG!uu!M
zv28@yCe15pUmrrp;PoLT)?q~*yC$gZ^Vd^EAl8}ha*w`e=+PVoC(4h+1o|w9pNR64
z6i0H(NKR57#VMm0zU#%mYH6s7q%EfL`f0o|=%!{fOuvFR25$!<QFP9lpFFZTdqgZB
zu{pc%o|C4r+M7h{Ya-@N#DjG#f-{fczF>GHNr`u`;tu|zndG{sk}y;Vi%`d}5t}yR
zHcW+t6HPdKh9QWF#3YeOVvak=pdDm5<EM}TDP$;M8od1L-6b>xCNXt>wa+Y>%Grwp
z_Tn*2<y<Cumx(i^3uzpc_|>y#Z4L1`1)iC09sc#cfd=WEYhQJ;9@P=QI`(eorLz9<
zvXukH|GJP1yj7%8QSGBsHi8D!rvde2{$~^F+k`^F{{(NH8Uxy=sNf#FSyZyv?0u2p
zf89_)Z`19geIUz*x(j>k!lB@wH0QBhHc79}&BSFh!(cekp=+PfLPd@5Gxn(-&Rk6o
zeF3s5KpJM(lp~vRq+wS{1+uA-;#F9)3Wq=!;UO+Pdr?G)Vz;zh_Z%0F#9PgOj`vP4
z3HryTHuR(gO&$F8h!@5n&=+rxTd*%Yforxil+nZ<4rOzqK)GIuV~8?Fierf~mSOrD
z`1?sc>HEE&`wVJLo-G-PVCO)BPX>d5q?HW%0xqZw_SWoKlV`7Oehb(U9t6r{9oSbt
z_Ep2a`m(Ry>}vq~>czfx`tlWv9l!lLc%du5+rU4-uVV*|PbXh5Z<+rWg3O0~{XaV)
BkG}u_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_normalize_instance.cpp.FD40702354DAF5AA.idx b/.cache/clangd/index/device_normalize_instance.cpp.FD40702354DAF5AA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..90d0ef9eaa1f4f26c2c907363f27e2ce71b869bc
GIT binary patch
literal 3074
zcmZvc3p7+~8^>qHm@p$Vh8bp<z0=q<JEn0jNs~-SawpMfQc<MbKBXL@)arDRB92nw
zJ8D!XDZ0pM%B`E+Khh~3ry`+JryG5H=38&G&syKCHP3qg&u>5beV_k+-`C$$BKblk
z5Zt0XQ{&bqttJx)1P1m`OiGCPqCp_UU@Wz`E;uf+&x!uA<m&R{?lW4pSlMpZB8R;w
z%8u!1S#t1D&lSpwgb$=s^;5wPyR!%91+)**jAEdezk5US2z&z4dUU{eGkVuA*EQYg
zPwnvKQ_Fs>3Hh_NBWeC?y#wU2k~i!3m1<mSX!x8Pxp^x;^}H-S4brUX>3Canu=>--
z<Ra^v{SVsZ4cj{{#6{^v!uE!j@~^#53VxXw-p#*2EJ5;E$1b%8HhVSMjt&-et;=%w
zt4Mw>?|^ndSL54qMbfU|QrSH?sIiPCh)PvVcEk&=R`+kx<BVGCr#|y)%kv4Ta5?9@
z-ey`uK}C3f*v8mz5v5PQ_z#<H;?<Q8x(qf3&tG*oY}@9Fgs_boOZz<EWIjBV(*7!V
z_8)a$$9ZNKU$p0vH{7rN=syCNwhZRo+kNF&=3@_`SM!{g;y0a9p(_pgnZ>7k`xBP`
z-9Ozfb0^`l{n58+$)7w->M9Rya*EFyZ@f0$Y^?bgFW%U#&l}d;bXWZ(u6jhQb=bDD
z@Y%_(w1r1HwG3;H9LyS@7N)UrlX+_y9QY{zvDuxZ5v`#1xjOR)^@Av)?!%}j`{3q;
zHg5g7Efzh4#=H3XJ1t}ZO>;74SD(tbQdP9Rcl8^S!SJ&_k!`0U@{mHtEKj@GX`dqP
zKNpJKhh}ULl}@JbyVf3h*Og{{i@WDspGV8+zS_F{&XXa1v39YwU2WI%Q*!2>m^s_+
zMCq<)shOs96UP3(?k9IkZtQz;iOG|P`88B4bl(|P?ik58DmyL}t*cJ_<Z_Y~xJcp<
zmSXMO)<=nN?og~;5xT0t&;8M$_xrTwknRzy1V72rGP9Hcw|kaE!ME+A61&z}E^E%F
z*ydH=S(R08BXjp{;^l`|R?OR5eE8@@=_Rr^>9Vjz=S<oC!bt50Wu_OO#hQ?^mP%Ys
zQ@iU&j(><Q^PiH4;=l99j}<$#Uh92c8F}&L{$<Pd$3g%0cW#}PQCXqkwo7Ep9U3o)
zJ07)bZEZd&z%(bw`}S1S^3G&M#rb3Qr+3z;vFmKF94R_l#WFK4ec99h{%a^Dk9$*L
z{2}}Ggo|j)BabapSL;rAK2O^uc{83HkUOf|5BEHHe&u?>_DNoa;v?a5^p{2d@>yeb
z?v=-IQO5(DF}?P0&u>gN+V61Iw`Rx(#+|m^UM=71H1be4r)8*7r!%_zKinNzXEnx(
zPrF8q++ib3^vi^o*XPHEra2eRT<u;G92r!8Ueac_L*qM9ewd<=AN73a8kwFm>!$s!
z`i4K+N{&R=FeiU%avl1);MG{<d*^9BOUetK&C%?WJ*9rrT_##%^o^g=hIbIA=0P1f
z3?$l@S(@D#V3usU-Iitf?8!UV;XRN%aci6`iAW%5{#G};Jc^-7CHaI9q-+XXUlp)B
zLiN+O$d-vS{hv}v0T?!*7))0Mfaz{cPs;{PximsBh6S1eZ&d)8abYhf`TX#0D#;ha
zbP8Qt6#zD0VNg?cn-)tY`D2(xVKG$!VBwJC?+flWNvWhD3>#7m4O9VOcgsu8St)!I
zDrp6VxfHIaDgYegD(~IX)$T|ot--Jn#fYv70B`l)`RWYUw2Mj##4v}#(NhJ0t4iLX
z=^czoDoKW6CWWb|3ILBb^J@n}g3eM&p%~^<_`a$D@FYFB>Ph^a7gUlTh8YxwwkiOe
zuy$SIAx*XSL>-mq`xReLRt`ABz;i6EG#sRX*H~I<G)M!FjdfJsf;3IaWHULS4Vsk2
zW^sVuG%1_S=729xlNzy&IKVENln3)>0&g)T<k{eq9nTS`#Jt671;G%m4iZARI>-*f
z)j?tiR|mr|t_~K$xH{Ml#?`@M7*~hD$V~7ZXi`Lo*x-~M;)qjXWU*SIC~6J7q)E|v
zs0&WHqKj~9De40h@D6idxlw`5V2@&H<pu?5(Z=7t+g2PuL!?PL4jfmo*~B>`gwGR6
zeI&9}$P`9^cax<yOq*b!fTJWG)qV!Ar2XR|ufqC+53@#eb+9kQ;X$AVQ_j4FY6Wp1
z9D|%89D_M9j=|0_jv*WbQagq?BMa3EHAe+tt(brAs25OyKKd4~a_4JeU8CBl9!P_&
zg?XerhM*x}3PneC@<1AE!Gq_AYrzA;DcA$XDZ~Tu!!67Y4FhxLjJuo6ZT*>pErm{}
zv%n$Qm)@(0So8oTW3!NX#<*F?5Kh5l7^e_2Vhm;}-kX-`NEz)=wn#_$U^@BhtgR~s
ziVu)zQm!92Sp8mpC_e(!R;FIp^o=_$(ZVK|@iwR*;$;wddx&Zc$sinsGKdhK!ZH{~
z;S3^-r-%&M04BsvMFtVsLj)S!2H1yC-e|yaMc{%VDb&-1E#8`JZ3cRxNkwGQ4+Y+V
z?Mmf!2Wn~m*vR1C)74Skcj|`>*ajSMt<j{0Y(ow>IG7jaJWF-&JOR&2tsrv<#~=ZO
zW3V}lW3T|mF~l6P1S`f|5Fl1+h0Z~3)jgv2s1r~LYhza{@94yYHHiVrLAWyf@3#^R
z2!sgrFibsM^<$_ZeL{@Cp><JzJtZ<MVrBR$LgWudSQ8f&8;!}O>y~MdH7V1yw5c>*
zy54k8hPxZW(D7WP&t@7J89U8&p6{~2b)mH(I&Y?fCG`KYEUt+uJj0A{ZZS(RTR2B#
ZYwyUj<5<}&He@aFT1}EH7kd!B{|or^W)1)V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_operation_instance_factory.hpp.115B69A1EE76F903.idx b/.cache/clangd/index/device_operation_instance_factory.hpp.115B69A1EE76F903.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88199850a1c4d019518e83b9fa496bb711dd0699
GIT binary patch
literal 6784
zcmYk92Ut{B*M*0x9-TQez))pyU|<9Rff0rh7%9>bQ4vrWm9BsaD2Sqn#u5}uQ1jRl
z3s{mb-)BWMHZ)@SjY-tljj<&X3yK;|6g2wpb7Xt<d9uPRUN)Co_FZdOd`wJ*N+OBO
zi?1jeGda7fL?Thbe`S+P3vC67q@|lgQn7YVO49i2{md^9d3Ma<$EbrzbL>9zFFJ32
zeo4U2VqL@@-L3_m(uBeNa^F=2|FH0a=XHn6v)kG(Ouct7a;mib>7TEswFP%scx7po
z=h~~KeHuF^MQ0o+&gv_Bm3MBE^V_9~_4(J<8~^&-+HT{ujVo*?cup(&{A$!nmxW=L
z#W$?-5<cHqTi$%i^jYND`o~9RSS`yr;@+|7-uKGi-`)**d}#gJOmbr7wq>{PEEq-I
zt)5md-t%JP-ZgjE6|3V*x+ccylP^z7{eGV9;6U~MZA)4^`meUxZMw$n+(Bwp-H>l_
zU`FloVr|2$!w(+kEK{vHv*Bb$m7-&$-%lG}U6Qnwv>jf&X-%c2PT76If@w$J=x<$m
z6g^_M$&xn@d{(XxOKfUBkl`c`srzS@UtX_q6C{5(29{YIl>Yo%Z0Pyf(Y_OB|MgRT
z@#BVZlMil>OIy>uDKgt_*Cvw?oBoAc-|n;A>zuH?t;ec4BkM2cUmF;hUf*i$du`0N
z&wTst|0BCKe%B9W-A_NPe&E;AUw7>4^EbNUUv|tj8zq$nZD~C3f9~{O=dauxwrT3Q
ze%tpw+Pd>%V0L56S&OSg)or;f_0-<3#tHW)*&PV*TC!rfuC4j{>DBL2&3a#~8`bbr
zSRc95j+Fi7Z}0sXyf8WIlg~z+X`QHl>++4&Sw+c#W53k|&-7n8rOxq^)NJI5S&Pmk
zIoJFW5j5lGgi9VTzSxlaKeH8*vS~#*&!iHG5a^U$+U@3BD)^*IjN{3885aqB3;vlH
zO^D;Z&m-#tHB=Y`fID%Qa$!Jk4oE{EN9=81c##TT0K}45%0-+Y>!`aqH^D!d3IhQc
zO-3uZFklp)pxjZTD&PL_7!_0iSP%;-7jc66uykCVr`;nem;q2m${e^zq+LwjAEZOe
zuWLv7JxZfO7yyYRQO-r2;P{1|X~EEkUQ{p#V2%lyW1{W+pE$*_`N*r<GuJ(+kN{8y
z$&hgor#LT>-ZHOlJ3xg10D6<&rd&jzw`oC$_2ENzsW2D-PvU9DMFjdR_c>70JwJ~M
z(EtRKU>O$?@OEyhod3`9P%7vEh#@gDE+P<c>r~>Y_J5C3Art^T(NiuWVDMQ)-dM-E
zP@y{jMWjf|MFjfSzkT=PiX*vH2n4{N_)EEnKuCMHv?rAoJE`CXfRPxLTtpx=_F~bD
zs%h(}&<%h}QfbdcJd=ZMr>@=f)pfWYegHTVXEQG1guGF@eXfS@Zc`x$fB+I8<MI*k
z=2c61&GXe%2mzpw6e_uhK!JI`rrNH@;hvBIP(TWlTs{I$uCUk7N^PaW5CF1B7Ud!W
z;|vv5b;qMFQNa=bdtz_SMFfia?q4;sXUbqISObtvvQ4>s1WX;88+z|>5EUE&a3W4-
zTtr|}4?{`*^z+xLFbsf9k}2aN0+Tz`%eDrTO{PL50E5UN%0&dG9&H&rePPW(D)a{+
zi6qInh`^`8n<u@U7*b9J0>DHvQN=~P7IU*yG5_<?PoaVZ0G7m3#zma4;79e7VnyCs
zD!2j=Ng^p15m;PZYk%=+06Z7|09X;L9$Z9VN%QNvt#=F>K^U3B_lCr%kSQ!W0m-uL
z%on@$8%|Im3VuF?q);y6&#$eT6<}=b_?8Mz0BDHDjEe}YOS9F~luv^#WeUJdGSiWZ
z_(!hW@m<fChpnop&<B7h5+&y%PS~(YRch&2UQY!V03t|)jEe|toD-vub{cS=3Jw6c
z5?4zuBCuJVH6x~C4LsYP0ECe+%0&dW%u}Ckc(tyH3cUbO6SWzakAP!$6eed}+o;e3
zfE<!z#YF_ZzJ4-DwqZsT6%qmPB3_nUL||+Dr!&3&3WvKP1^^%8Bjv&XTfRzn7%QCJ
zeGe_Bf)xO^#8%2hoUn7b^gCj|V+|FY0k9!9J-CR#{`&t^1+To-hYG#`SQBe07ZErR
z`H7d`@Hy~Yc>@qiLS<Y;pvi3S!xN7Z(y0&ufG_ctarp?CnOJRkrR+;87yt+)fs~5~
zd>=o>==EjTODcQ<zz{M-#zh2<1x=dYczYk5s3ZU;kO`EF2pm`K{#g;#p`=1z0OClT
zoQnvYY*%_-wi$np3UUC(lCkDoMBvo5=XnDJSr;nU1K>a$EV+ol55g(WO6k}0sSpc5
z0!ff@5rLn~6xr4%e&0cbcmRS(kd%uETs(REg6Z!MV9zQ57(>P=xQM_tg)ueqK_pzP
zZ~#(Cs+@}m{JQ68>isu8i>WXafN^A;DHjpAk+=A(hb=}46$SucAO^}s1nxBd^Sism
zY%CQh0C^-&!9@h_E{#$@ubL7`1$O|_NxCT)5xDp4#;6S|rXQk$2LRzDT+T%V?t3kb
ztC9DDhgS)}R5I0>i}>(93>~}SX>$TRdcgn;B||9}al*rF%>ggJ$9gK*0N_D9q+CSc
zg-6Hdk2ddHMFlAUxg=M~<s)F{vxbqge{rXR9RPO3PRT_C+IJ6`dAdWEO$9vwTB5b&
z@)6MJZ~4r#c{&xe0OXT=Q!XC?SH3B{(C5}ID#QVhOp;|>M4)5us-xb9A9oX>XKLq`
ztTyRhZQ>%4c7>$Q!q&g$!y9gt9RHIFYM7xUN+}m{#@iNY-H?im{Z!BZU{1`XTo{nd
zC%pT=M;m56C<&uNKLFB5nv{z;K~g$qY+1VklAqU=Ly4yfwwTC`Lv%y*_{?D-G<hEc
zakY0TtuzE3`!ft^YlXEHZV60>U1dBKA}u$1YCN@=hU7*sjh7ae8+>x{a>GyE>5+LQ
zd;Rq=ce2~$a7=m_D&Z(tqQtovCBbGn5^dVHXyx#ic_(0qzqh{+1|0L(JM3%wdu#&R
z*W1?@_mVR8e8I%c@||GE>EirxRbUE*KZQk5oGmx1HEJ#HcDd0(<DkXe0H0hO-0+jt
zld{aT|E7$FxwG764ZwF6{z-7mEV<%bjI3a@yoxpkRj}Wu)DPe245#o7mQ2ysR7b`;
zZ)+)rUz(^(^v9`A$xe6TzM5GFcCH~e2A>b71(omC3|!e8>>NW*{Fcra66RW%%|b2C
z#_I<*i?V3<iE(JlsQlwB44Gh<5RSVTqb<-Z*CO4wHN$*ZZ1_x|V|B59cwsT;0?k4$
z(thP5&b%wi-vD%kE<%rQZ4A9Yv&4&ZU~%7h$z|uHKnHpU>YF;>D}kFrdyGwLuLV2H
zkQH9g83ZvH!w{B=afso~{+y6Q>)e5!IACHb<|(|1fMyXHY0So8vtW!iCStHz97a3B
z=Eh8&hn*2-7a9s<@;XCp)E6)ACzV>>1Up<8?vEQZ+Ul{g-m=Sdu)}m=ei%5S-yQ!=
ztsZtB>;zqc9^X3y*Bq%mGxw1~AvfCF+V{dV3-_x^?fB9Cs&a6|J5+Aes5Fk4w&X@9
zm6KypXNHv<omI|`7~JJXtxD^NOHpp@t?KQFcZuBSs&aM2yabQB%F_{V6?nK+UXJ*1
z%Z=$h)7AI}ksFOYjcPoBiGj;YY`hjPg}-XFVRQ^W*?0hOB3W&~6PcoE9zWRZ*KdFx
zV;B>GZ_d=<q)AOBZFj*=^GgfF<u|-A?SAI^+09@Fc?bF6i#V-e&$*D*1D}8$steU)
zkjboATawju=nt?hm6km*FXy%_?X%c((Q~kqbV>dg#Id=+0kYbH2Z)yzY*tp#9+Nnr
zY4)zKb6|*-!pahpG?o{jS#v=e?^Li^UqQRDH0Id5`HjEAkUT?P1irHIiUZBc3)0wH
zz-E=@gWY+Z!DdAT?UELke-BTZvkqp*=wkHvSYmMjnza_Bv95s4x(eD;hB|i(z4}KG
z49PHLgkg`yQwB6EE=c1O1~#iJXybVo-uDx<;qF%3JK^6tuW`!U!ZC?<@S6<Y20g}`
z`A?&J1)W{20^34qVT-FAn+^OXR%>wNqVC_<FK90P6llAicIp?M*#IvN902P$NW;yq
z>Fta+KXw@y$%+g@U&u0yUUXO{jI400h{DM8rBCCxQ|9`xSIr>ZAU`~PYsV`!BTtM>
z1Uu7^8HOhtA556dIuOpr7YW#``FyZDD-YPL@1VWjYIAu*#g;`dJ6V^kZ|_VcxJiL#
z<p=4y|K>MeJX-z>&_#x#P+XDta0AVn5Yo7jz-IjiZCo~Bv)1#$?tHs|&AQG9`}&g{
z-=23|;a5iLBK`0h;m!b>6(6Lr@qo?h4%*mrz-A=}?VVvsA$wYP?Sk234P&D*<YVap
znl&Dz8~aF8m;PM;0B9SfjWtr=&DXaURw>v;tBj}|fvbK0+VJhG&nrE^9;zGahoKdd
zP!}2RNjMwUsm580+YHvJMythjii-yhht((i5zHrGv(ogz?tH6(&8iaGhy8}-b&KBE
z3WwwF;_ih(3}YP7tT!QzFKMt@TSEJ|NyDdKU8#8qLwvk_eDGvrBm$aMC#3Oog3SsO
z+W2+>o7E(=@jV4LD@SPKG6tJfBD60gNK04u&taQ9N*CpieI6G*(5xRJeJM6qy13T7
z8t4G;0AJk6m*#()w<+M@NwDXdxXm?*#x!+l&wsT33)kI&Z)6*?BXLDue%x=>>$&wT
zH;yxmi^Mi{g`RLvu366B+Inxj9>WK=F*p`h#qe10r@&?f3~gLZV6)nV_N}J)Yl??i
zH849xm*R)(<hG(#(cJg!Ct&9r@*~?j%LcA`7{aO<4!QHHc-!mnRJQI?b*cV%VECd1
zn)NiKTUy2x>>6J44rp7Ytqndpm^y%Fr44CZU0}1ShBn^rV6%eu!5&6ciBX01Fgw^g
z*cUS(7B-++TSL0F$JSaQcqe=3d$@S?LkiO=(5$*4je8kvR@l(~`|3jJvW{FO3>j}2
zABKJ9Wy;m4*gfl5kti?}#NZnipE(%9N*xZt6AU)1av$u@Z3#9jaA;$5Gm#qKH#YnR
z_O)&@-qX;AWTkP^ipfgrqIJUqf%K$t)#8GMPcE)*_zBw`yn$G?6US*Z8ZEvZFyuf_
zVf9Xg(jC&B@D_lq<zRHetOcL6Mpyg<4RAs>E1+eixn&0S#c(F?|E1#>5=ol)<p}Xh
UV{!bK_ebB=CU+e^A}w9=|C2NULjV8(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute.hpp.47CCE84C97E4E9FC.idx b/.cache/clangd/index/device_permute.hpp.47CCE84C97E4E9FC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..37c5147034cfa02256427b79a51dd61bd736823b
GIT binary patch
literal 962
zcmWIYbaUIq%)sEB;#rZKT9U}Zz`!5~#Kk2=nZ}F^49q~eiaC2HIc75%3bcM-{^sDl
zEzR4OE3}GL3*6G$@?AA;`R+vXnFcejcKx`&JIi%Dhb>2|!;J4~_sknr9^Xq_vF7|o
z?a5!(77D-K_IJy6NzF+1FSin{z6$I9TCDs1+})EiO)qVH=YOrZmv8TyoFkDA*OE@=
z%TFp4zv{0aaPEllftgjBzZGXbxFj`4dAFNErDME@<mB+!hsxq-^Ah~7aes=N{Yy3U
zg!{|b8_y(!gEduOC>OdN&sh4Yc1uwGjk&zH-uXu_W@?xu<0M(SZgH&B0@0#nd)HmR
z=w|SQ$#w4v!|d-LBbvW1%PV3DKA0~ivapf)-SbmoR*!Ne5B<Mz=wYF(0xP55LW48~
zmr@r=9`^$-OukQqAEbS`B9~!%cFpzhGu9<XuCu(^7o@gk?U5(ap7+Ju8Hy`&ljZ`$
z(4twm@&j8rBO4Pl2LrbNyMPFoVBll|1}p<3H)Hl3W_6jX2mc8$v2g-LfPh<0Q_h-!
z1H#$$;g77IOQVhe6Dv$wj73ZwCcRIz-gi}tN-Q@M3nv4&5SNf3n1D(DP`m5#?$qD&
z+)V61X<;s5K`;T6W++NcD-HxY1m;RcF0ku7Rm!?2?sBe+-ENr^qo$(F$ipovBdP&&
zOKKmhYJ*U*6Hr({TtFG78Rip4POuMP!WD-fKB;24cowKyNK!}+=HhNI=9ahp|1Chm
z@<M7b;kCwA0Y~R1Oa}_f$;he0%!FADG#li7nAtF0EL>o3!-RRbg++v=VV>vV7LyZG
zg_*#^Eg>fX=RGxG{ld;K-pvH`p`fTVLXn)HoHWdGm_f{3;Lt<ZA}%KmH;IQ^Qce=i
z1BPr)q7pC)VTu{DT>tqoNzW;R1s+_cxG1^U8Wh=lkQfJ%ya*;2g2@ACGI2lYIo<j^
Ny{&+ShnJ6w0RRDgIxGMH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_impl.hpp.BEE7C34D1958CA2F.idx b/.cache/clangd/index/device_permute_impl.hpp.BEE7C34D1958CA2F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9e44096662d42981d4df7088797bab1d8193b5e2
GIT binary patch
literal 5400
zcmbtYcUV-{65n$u!gASNm$Gy&RTP$vpo;|&1cH=RsuBxgUlp-X6bph%37~?sJeA02
z0u~@agNhhHRO~O$7KDh3SVAJGfX4F9mf`Zg_ecKFZ#{G8o;h=7=FI%&T*UsqzC0Zk
z%X^W3q9ktJ0%aD9HJbcSShseWwi1gKqrqY&mNkU82R(M=y>*$FKWaczO|yExni=x>
z%4*eb;a8Lv3inbuw_66@i=UN$9Q@WpVy~jEY!_CLe0?afHK8@=@}Dof9iEz}rz}r%
ziFlLLw^10mdDlYGEIJbXGWAVyoP}${lB&KkQG2|^{rGoAY~H{(h8__Ii@U``6GZEO
zN}FrF*X;AG_$lsDbGHsm&fW4u-($|Fx>g6#&?y5`wcLLle)_C)`G)5m4c7)*KkUd~
z-1K0(Nt{7sa*-(3>%dmM_N%kYI<I*AZE(w}?4|ZPpX?WlGA6!L-(8R#`mi&+_od;P
zC+(|VXQ*E(Zk-vvw^Cmu8R*ZRRNG##PMA3Dn(jS*ZTjSjuIN{O=Ts{TRvdhM@n(1I
z(x}Ryb2l>IG~JgZmE4ppF!)L6zvNHx529?V-%ak{jg1J;Gjddli{E|2-k{GYsMvg`
zIMA!zx#qs#H)U<x-YjU-I`6|?vLNL9vQ2kAg>FT$i|ByET<c%DW(Q5(yG?&$#o?2A
zcbs}Im>pS(-R_OIw-iL`<`z^fJMOq5**Euj#VA45sa6fn_`q*lxMMEeELX{({wiXx
zUA&T~GWMd4YD(+THHXSH16#FQehEviOKQ+P95Ly0_N)4unov&F;+Wr0uhI`~e-<<E
zg!1aYKP;bY`|tY^J-1(4zpFeOU7uH4>#}sZa6na9RBl<^QgeY_ZoM}C!y}KkA4gxj
z@v~{&trdD%K}nh$J01PfpBi}CTb=s3*@J2h)a5PCsJQxS+4oP3m$oTiS&~uen4i8U
zqhr3)8Xc5$$!E_Q?WRTB17m+^4u}_N*t%aU*784@KlUvLJr&LDIf4gIQ(cZ}Mz($D
zM!)<mBeuJ#^6YEV<h1>lF8a(WC^F^c#WvdcR9Fq5LmC-_cP53|&HQ{~PwZKfkIe_R
z4<74%L_b{IS@l`)mBCQ&RLhRQirPcho8tEzoijE;X+wh`JvzW};&-KSlYK@FY2Lcn
zCU{^tq@pcujNA4uRZWn%^~jsnjz<TSqDv~QKZT^&p6D!2U0E4nU^P_utUBO~kL|UR
z2m`l1(|`ZUJ(!RrS;!@YqH-i|&4k$>QdKcqk45i88v77mL0D>}hR8o&)Qp6sf?1ag
zFr_DdL)#nMTf#Iyz~8O6_MCu&b@f?vKGMxcV-y5LwES<~QE9em;9(Bgq#r5wBY}bp
zi#*oVFS}fr%g3rj#GU7EAsb*?*IuXb#ym%%0oDc^9E=@?Z7_L<>>oUvG?s_83|MqE
z;#8y23IZDFHvM<;J#!Z6VkKf`yjna@Ho)}M7qu#n4G!jVuq~O+LcA<ws30Knm4NSc
z@Ao$eJUp71n1;>Lu#19#i1@cnt~+ej%;#fGB9eubvoKFVKxDbkryy)twm}t-60m67
zQMP)rfu)WSfiOg)wM~yV_d9B)aIgchBnR<wkdcCbj`jSYGabhZPxCR4Y?F_b^D$RJ
zhDD5@Z_IAG{8<$%5s}pztF>eUOebDdxuY#QSHj0!GCi3$*-|#ZbTVgXeR$x+DI9D^
zY{)=58OTyWKxD`Kk5TCZpKLgI91+Pxyi5`>IRTNZ0Li$Lu+>L2urk?1$QJTt15D>s
zyDca<dHs|I=9B3tc9f26fawc<MbYb`dfYg8EZJZu(%Ok=1p$%PMOXYgmL4e6!0JRp
zBs)@DHo){%FURZJVM{h}uszX`jd<B4l5zqf?TMduT&wJ><zo$^p#kw4kcEPP$iwO8
zp^lIEA2qNF5s|PZ0@(o5&;B?R=!Z3qaj*^9Bpn%~BO3()k>0x0c~5J%wQ;bJi0nez
zyO6npfJmQJsfyQvlM)TAMmCwlo}(`tV7j0EpQES8)$QTS^1Be}6e2xZPC#TJc*FLv
zVtyP?=1DhF?<UU535X1aOT7QITK}Ab4T&AO$T%0-DF}$L*2XPPcu0-F1-1eWAa5x;
z9h2V}s0ND8!=8EA8*)<0`{v=944;p^^Kk&f@5S7`SQmKJGj6n}&+!IC(E)UT2oh7%
zaaYOm0;hIN(W0p$Uo8bZP;~H&;F%x<$b{5ohy*4AB9XP*5N8`Qh1zE;aB01wH~^Or
zTY*NP0k%_S%TNejQ?&P3Z%^2ra5KE487|N+a0+mMSHv<YPJseJ)1W{}_?InlHdqjB
z3(|UbYt%j*IpwUZO#9J(X7E~>_NV<vaP>$~kMscRl<8h%(TgnM;Eh&|KVKf|<4SbT
zF`8ow83p3RM>#G6;zLl#gcJ=SCgxLqs_*x>$^If+EQAkk++Fn~gxHdW9kQ@L0Hf{s
z<yYeMv&WO8$;7-&YzPz5wBhApLr4p^Q>^`&YTTcR20v@RaS#%ZG_9mpsv8kmZ)<34
z197Ejp^nfT@MeNbhPbo2>MaqPDVXUDYbiQN5H!pI;1i+EoXz?W4YmUGAsjPCzxBQM
zdiG*sM;<oI!y+(n#;;|f(rBLqWr_}@0(F^9f~nvUoR|`i;KKC6%s~<gh{RvsOTZZj
z37L?05SWP6y680jxSkcU#hdcB0Yxz-fAr4ay)7rN&m;Zw$PPd)=G26hbFD2ch>5F>
zSC0jkD3S<z)yNVugCa0t(TyA+ktn(kY4;%-s$(JB`viB{hD}6&2_ioyKqIh=7%J^!
z{+1iE@=trGLIWd_-NM;oz(iKw%fJi>Bv~!*V4w~JG}yA11+S_9NmGizpJ^c)1GZ4~
zkI3^!G@Ws(0gY=w&JYWV?m-?sXe#66dqlrSHsGW*A6+xCCj?*0`(@$~hR?zQSy;^Q
zIhdP+1;8gZcRfi~zIBZR$=$)-6#xXzhyaPS$uhn~nnk)n`=aP%qhv>j6Gf*Or#JzZ
z)OM(2P2ItGqBj$pWnvfD5Lz^eauO%W806WCHVr}CBv9I1tn_Jstk1#5IoJ}`L;oi0
zH*q$N$bWechxNN3XN_86Qr$t;pG1O_$UuGrWU;aWly>yye)do&4!gN~g2SX-Iet3e
zHbwh${6}y_h+l;CnKNrdc8$n|Iip78*@(Ouz6EhxknTv{v;|p@<Q-bjc;KP_mDsX-
z1`}+1Eqg<TbI@`a=HL!O&PRFyVT5ATVzdB4DblRfb|Ee3;}l(noXgN8@S|XGe#~~R
z(QQ&^g~+uKO@#?5FD^u1Gkguws~K5&HE4Vdng&T*c-imBuNkR!<V@Zp?f1x($<B1F
zosLZbGY+jv2*nY%%gCwB=FT35l*8Wb#?zYKm;U4Wqlsr!LL}wBsuUd}2r*@jD^w6V
zf)fkGBROlaJtHZ06Hj8eFrTpDu6)d8V#tMW=kF(0>_a;H5DguNB0#5|#&~gJP5K<K
z^AmRyFC?xKKe%Ysy-cnXWcF8nQ3pDE>2}mH(d+InQquf=$SO@=wH^N$&kMe{T5h(7
zEd0MJgG;Ousc}xSAyldKViK?oMF7uMTx_B002x%+vI{MfmQi|9=8&@#Jy&fmpGnf4
zNNp$L!;TbPikwT4C*&SnhKVn!oYWDhEWZ%L&ez&JAMq1(Sj3ZjOn%b<rf+y!ZxRO^
z6@F1bZ#(?a{$tquy+pf+D>4SqfC471Kz;=Whan|MMN2Q<@S50$)bf!IT=XdVAaXv4
zrhu<q9yh!PkLsZ>vh|5xg<alLu009ZVdQa`i6li=Add=$?{-vqGm0z7B7%KLy$@+I
zRhx#@)36p)ZI503F&E=1X`?PjX648o?1gU#q9Ku!XvAcN{Br`%2JnZ7rEpS4h{-=G
zKnya1B#-n%f(e?$LgEY&m1ZFMIlx}h-tJt2yuc)i?n2xyWCBM6@5vF#zdN9)`^)MZ
z9jeVNzbMwP^P|63b{bbjBGZM)PYByWyC&Ppzh%H1_?98D*ort?(TD~p|FD5*4E}xM
zsNjVpho}w|1P))h*T$_}FrD-=xKOe{@${n_kF@{V5`h4e;TFXLjeu{%e>+K78=H_T
z?Tg0H99fvYq}%BlUjHmc!^(P$g{6t9nYpnM%WPPHWz9UAGml2hqb2ib!91EUk5<g%
K80OKGdHgS4b5TtI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_1d_instances.cpp.7EB32958B0512D73.idx b/.cache/clangd/index/device_permute_scale_1d_instances.cpp.7EB32958B0512D73.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f245650045cb9e5d9dd5c4875bb1d6be0c272a80
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RLvI(|HWPUIKP7BK
zQn;_hV)p2XQohO2lhq!(^Kt07_Bb>gb=~yqYi!4o)YE?IrI+>v+}JEWXRo8A16#ss
zuV}qS>p=hOSO5Mto_DbQ=KG_6o_%_CdC!ih9(DBtyVXA>l_qV7n7iMWQB7R^z-58c
zv-LT5{507A>zB0bzJG@LVe|9Fk3P+^lobvTk)OUg>Fdi+ryqU$Tv}M^d&0Y`e{<yW
zh)uW5PGp@|{hoC?YPWNE!Qrp>w9mKXM!l|Fx$(vV;}wyTQ>XbzE?j&h{Kx)#!K*J7
zZnt;aVSD~TU$$@jro0O*RtqPtVmj-6#jf*Q<n+4jFTO1dOPkyB`@xFJ@~Ne~{(h%c
zFvyj<rYA2{@tnz!(tb=k+bUvD<>!C0DoO0ig;}De++k6znrd3|R5{&S$?Jo)s0*)<
zwp^*aThi-*4-Q$KPQ}NWfBuZiR2F57N|%<FJM_@DkY|<Srw<2R1=oCT?%K3CN2*k-
zncufQb=K1lJx^}AJUBM1HhFQo-n+oXQynvQEUuF|xc@@rjiTi{LYlTTxxBk(xOK7m
z6^-{RKLluA5eix>dc8}=|I3u54T<-Y`lV0xcunFssC;AFN}dJRjs-bduX!3IcmAI9
z(W!C_?Q(agDQS8x5xH2<tmKle;m^stwSC3XzP{sguU)?u_&>3T@t{-GRc5d6cB^lD
z`}U?Kzd3iH&wBEv*BSg)=e1jm!bGh)=B+ObaPfD!Y*DPdyNmy7O_Jb-OjBPU_IDHV
z6QnO~TcY%}a_OvzPd=XSaJ#?}S=w=Fxr6g-zF*9hRxHb}^=&WZ6694a{$cygTP~$N
z=E_Sou8G?|tV^E4$XmBH>KxyF%i7@8zsx5uS`ynF6unXL>H#?~Ju{QJE`F=sT3<%8
znKheTDdBEia_;5}5ACV5H@%3hVo?+CDLb;}=8vQjvqkDtf0_H%{x)s?{igK&%@5lu
z)FW<HzUI1}w7tF6$gV(3A$t2n!COZfTE$}?@89=!>UV?AXw?lBOk&ADA3Snjlxyr}
zd~A*=?;j&8`MDc|B*RvRb>x(tHQ%vd>IeOHsYg94F5EA_c_M8~@zL!$t3N);_<w%d
z^L+WpF5h|c#njGQ-?x!}^T}ql#>_bz<^KpLCVgh8-7Vk$!v1<+)~Y}MPd>9={`kh)
z`3@iU&kUb+_Dy~iYy4k^;>z44J7B)M#$CWKrqQj$!{8st5W>yMtp_H6GGOps=fF%y
zu9hesh5(?n5UY?Ln1D$arKT10GXZn(`cA3Rb*~OH@r3ZI@T$WN4dFH6HHYyW+utz!
z{8Dm@i6w+zk3Rq=%n~A|Cl&y+$ZO{Q{&UT5cv*lZ$?+<~6ajg9yv7Kgh@6NrObbtl
zh@OZsjOU&4<ITIHPM1MW;oy}-(!wDk2~!l}C#3#jq75rh5f3L1KSGf(r!YTEQO2>U
z(#G|nuR$hx^7<hZiFk_m!4%EkXS?N2e|s(?4~jQdJD+dw{(t!bP#D>-TeAMm?u*)H
z3Nq4)*AB@@D-k=Gkq0mH#Jti7ivlTf;B`k*<RIb>Q*<cA?q$6E_2WQA{4D%R2t{Hn
zVoESYFi}P>aHRGg`n_PyLu-Yn2CQG$`Ng||dU-^7<YB6KLWD(y<zaeZG*B-n+F^SC
z3WROCn}6gPP_KxJh&oIakY^%d4&yNtrRF4p(nxVpa<MXq5ey>0fH{O^?WBV{TTVGM
z1qU$%!GswYl5<mXGJss~G+$;GRyKAHPHrAv0e(JDL04xrK`u`xA#q_52}uh}Ya3fT
zdj~yHZ7XFn4S8&$V$w3Q3W`c9s_L3rI{F5N#-`?SCQ`aaE}~-YUZG5GevTfDJ^&9)
Bj(`9F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_2d_instances.cpp.2B80C1D2C455DED0.idx b/.cache/clangd/index/device_permute_scale_2d_instances.cpp.2B80C1D2C455DED0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..372d2d9f3a9c3f199806850ad088fbd63e44fdf0
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RLvI(|HWPUIKP7BK
zPPnhdV)p2XQohO2lhq!(^Kt07_Bb>gb=~yqYi!4o)YE?IrTY3J-`=pM=LfJ&V7#&F
z(lqY}epfD+e*N&UC;da=w`{}xdw2b_(O)P2_$23te>3+)%$3jy6MwFL;K?VJ3b~}p
z-=>V)<)2@_U$^k^_WRF{zB+w((T*tH<;hH4k1iL5uCa@+jlEO5Q*&J&N2Tw_OtaS_
z-eI@dBH!+O5|<m9UH*bAZ^ymrr!zPBX4~qUib<ESnQAY-G)YsrWXcb=|J$pxLap9?
zJI+xqE_tVT?Mv@#yt5qCO@cxfRCeA{zi9dT(tEMHc2`cVnfyUO#`OKZP3s<>^pq4m
zuz1~z9GxjDn!*ilE>xuYvUiuix7*Y1^5#KFlSB8U*$Z6OZ)#gt^C-vDg|lvX(+bC=
zql@1i_Q;XtlJn3F%y?I!V1HjbcZrh2>zw4n$rd#VeU7Fca&|j3no8>e0?VfOF7`|P
zu<+%2pUpCH6Jlm_+=<wnpY!FyQQNF7MjX85SNfHIuv>9QzZ046H6d)miMv+I%D%Kn
zJ+&2;(`t=!TDA32)WpY~^MZKH-bBACbe<XLDWq7zblYqy<CR+xS{dzIV@^H3S-oP%
zCFTbg7T*q1QSF?fB$XG?@`C57=E8;F1h!te_;JVOU$?AY+56}zMtH2Qa(H=9ytI1l
z%8NNYcO`EWrU#YDzIA+F*&1+cl@t5L&!&7`FFG@%((_WnKQ5egUWdtS?Xu372V_6-
z<~ZM)7P7={eTeZSn}2c_IiwU<^9BAA$;i0tSg+vU{@}{3kK6nd7cETT+kd?7Wb>N|
z-Bx#96fcS0EAKH{aBzLt>6;I&m*;1N{!a^<vL!npOWSOT)D7p8N86T7K9PBKg@Ntq
z18Eo9X5}rMaK%zuHgjoFu-IPj{sT`YUF@4-DjnZ<&h5*Rp#RHO=G&)#n6US)t@J&y
z_blGl`LT)7Z@z5^IL7YBn)2Gr$Z593h7V2Ie{OHz`%zA<a5anA`2&w)>~{T_m~p%3
zVvk|EE3<vCo?H5wkhZI<M3rXW{lKf2q59|b!Zr(KtuOttW}776#(cEBon>Qr>;KlL
zM}G=?Z#?<=@}Vb{&u2exj-8h-syf;D8~=Zo88>X4?!P}S{7Zl9;#pe%>o@&8_|l^6
zb-col`)9PzrhQXSPEBB9D6Y&+vIFKjnM<`dt{mQfi-#c~kRe2fRY(s^0A;{n^~y$h
z58h}$9tMA)G&d`^9+-eh7p0~Z^D_Z+@cK@v(si#6Gx3D*s_?4Ab%pSn@S4MTFsm53
zz*h7g`n_PyLu&=c_BRYazm(i!VhQ2b;}3wTW(g6~6AOUp@tV26|6KDMUKXGQa=gkg
zML?b&uQ7rrA}68@)4~%Xq9<Yu<9TQNc=PV4(`AqgId~<Jv~Y+>!W4!038}xBXu}Fr
z#KXzMk5DAcDa;R3lyPjTv~hjtYmiBvynYBpBAz0CFh%qC*>1Vh-=53JgW`?V&ga{^
z|6jfU6h`(d%nV@2fPxqnGF!6#&F+iZW(u;&iq{UN3g|H_5j&Vo2QTx)ywV7Z0x5Ff
zbw^U<AmR>FbST8`WxV|L<3L6HEc{9cMPe*sN-#xF4OqXh^NV)_74eAj$RiX9iweub
z6#W$l+jKYo$TOfK5fu@2m@trMB4Q5XF%+fdB!bdNaZz%yGKdijBEWz-<h7N{-}oi>
z?3jXs7=mEJ3=GM+DLENHu6LR*GYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HP
zt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X*}CO1Dv4@MsV
D^nJIn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_3d_instances.cpp.07A2A609833FE4D0.idx b/.cache/clangd/index/device_permute_scale_3d_instances.cpp.07A2A609833FE4D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4a8461f376a96018f4b412728f7213382c5a3992
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RL!*mtS%`f5|58T8
zYudJX4;2?~NZsZ+BhR|>3&+bF*A%?tVimqze=m4UQSqgp`QAMebKl2&nLGOggAqgY
zv|B}-cRH(rrk1}ySHbpvdA;c7*WIW4H$N(}lbO$uFMsm$)6Ro6mQ}S(J?H%!jw^~?
zpD$eSx328xQ-7<PpVeVH_IJ;8cRx)sQdyv3r!(#7r^iR1|2(^V^`l2!6P}C6FT3ow
zF=1=&gsbzs-(79nTGqYpLEG!O;qy3eZ@Kn)>anEWa|_Z|uh^1uibGa)-+}*6=S8iR
zDc_ykGCPoMdrbDPNu|wd7dU%ER5SKDy`B7GPU*|?<~;GN)Kw~XSl!R=)?J%jQ559H
z`eB8)l{B~Zq$!*ZWf^-mEo(lyI{f`TC#7xWR*nl?Px=NtTCk~YUCo0WPZ!R*)lDlL
zlTI$SYngb%O2lSD)S?S_K04I>Td-wf7sJ}ky}aKae(1AcyQ)z6<G~`1Yd_;$Zw8w6
zT1`FJ{;VK%*3%CyDXB^Y$!FJXyyz_ceZ|F73MO{~?fV{tmx!6&(bSu&uvTH>Zmnfy
zG90F-s<mvUaBXb~$v(PO>A0}%Qi+@!TPqImrn;$mwme|WxgEk5Fm+?d1-|T^Q;u)`
z9#FKj`oKc>?F&7;gfv}EV%Bq9lspye<os4KYvtm@iOaukExl4NG2eN^gs{^7OMCra
zeTxZR94Woqbi>ngOWs)B+I#M^P@?E*Az7F6+HXWJ3STricjv^qN42K%kt{h;zQUIq
zqn~8oXxXB>(xZJ(&{>a9Km1*kGg{W%a<^RK;QadAFWX8hmgQQ<Z(Fr?v`xDCU+%Z+
z_f5`vuS$elJ)$da8=h)l+aII1?DzYgeJexT%_lEf65Bj=TaxdpdA4hIL|iTrvRo@A
zs+KN!gXdPjWnGa;R_TkC*M{UBT2%g0P-4FEEy3W84}U($aq!prnm_5+H?`*7Z&Ke!
z{@+%i9&xMkHP`K=?d`2bb_H4rt8bm)$W1zMh<n|~`upDwRnIumD->bF)ID)e`A6o9
z+awoDCZFqK`zO)EmY%beXLYEyi`m{s*>f&?{dn&DwU903#rxek3C6c}K8oJ<w4i3o
z|NPSzU$ak3n)`ft>&egO%+FWvwmI+RrI!AN{lCkMi2Kg>-yLWE<-fIgmezmkn>Ei%
zWOjcQ=J+?Oa_P-8|5$U5Ca^OUSLP<!0rQ=mM*pS)wzere4E})(A>6FodSC)50|tvl
zt#^j{Th;O~1OTOlScUY!1WdXpHLaMR37CV|cS@D6dv%zJCxlmpR~>F>2(JmRIgAIR
z8M(k#^d9=XV9i5o1;_R`3_rh=++t!0;n(92fT>~$5z`Y3fZ6XgbASK2<~O`7Knvt}
zm0^m2JUw1x1W!axL>Z=qCqzU~#2Ci&&iL`>-BG8@AQy7*N+N0D5Rrr_3h@(Ce=*U9
z6{v`ZlZPLnNSITYAEqed*i>ob`q0-PlRSC-5Q;=RMf_lj=I^uJa;LvNmyrj>8!%~L
z5P-rG76hxE&$oB~zkC6x8aW8IWc{1n7q!h4WRn%I9g<B}B6cvN4qoPod8H8+1yba|
z>yD(zLBt)V=un8=%Xs<g$AOCYS@@L@io{sNlwgXU8nAw0=NIn=D&i64kw+*J78RC<
zDf%lAw&`yEk!L_fA}S*4Fkv9iM8q7%V<<|^Nd%>l;-chYWe_76M1TQv$ojL(yP~qC
zSeSx?7=mEJ3=GM+DLENHu6LR*GYcylI|nB>53c|}pQoU!vzj26r<0Jlu!w}Dg{8HP
zt)0Dto~X8!vYCcFHc>HY8CeBIB^6b5O)VXL14Cm|b2$?!T_YD!F?X*}CO1Dv4@MsV
D<)5KO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_4d_instances.cpp.18341306CF0F5EFD.idx b/.cache/clangd/index/device_permute_scale_4d_instances.cpp.18341306CF0F5EFD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bd80d15cc7faad3e81b23fb3674182d97c0e8473
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RL!*mtS%`f5|58Ro
zYTC9lA0`CcFuJYcvwf~jRm#g7*A%?tVimqze=m4UQSqgp`QAMebKl2&nLGOggAqgY
zv|B}-cRH)0rk1}ySHbo^z5eOv*WIW4KmRE*l$p<vFMqQ8YUjbE(tp2LB>Mdtjw?>P
zK3}-tZ(ZHbr~XzoKda}hx4WzJsrzb?m5Rc7d!6u~uO1(Le)R0})qfs!O?WOMpSIa=
zW6IXt30LQNzq{JDwX9q3dE4u`;qy3eZ@Ko_DS4k;dcbBSE!~+$PVSxhAL@554{Lp#
zSKd=lJnx|H^;!E;?k?agaZ(QDoF%@ZR%q^Ky}x-E-a4)IncA}Z!HUZ5OG|nE{Z6f5
zkSp~}PhP0tIg=r!{g`&PRm7gEukUnKlB$;rvqVj~!=lQkmHpxr`?+Z@Qyy&B>X1^I
zX7kEfqI0eDLFSdJTxVtZpFHRHl3vMi!Rq#n0(rjo9mxT_liK<9x)SET51k^LdU-{#
zWK9NN__9boe~&OVCi81=q@-<23j3wj@;Duq{rZ@r&NV&p@Z3j1oR_$ScCJp!I^Gzt
zc+c`f&MO1FJ+_9$XcyUB4DLGE7Te8!Cm=0kk_(62fo;C2455<VtZz3=^RBe1=2s2h
z#V%7-BHgt_Wnz%a;jdi+a~ImEnp}}vG|kFNrYtHx+Wl|)M~<e}Evu@hysHh`{(P3y
z*;_Z(HtLqErCz`Iu5{jX4p$xT5&_%w8^T@uT`pS`EAQ_5ezhh=a6_i4uMd0P#Px~7
zm$ofYlKtdms(kXXy<qPG!A)86PbV<(uD$*G`;(Ufb5_aZzZOxF@=E(({bu@|(|kKu
zT=o@F%6*u7T!UHaPtn#nZ}%7fS{5dI&!=lr8827v(ODt3)fzh@E|&;du9Xs9mM(dN
z=T^XFU6Dyv>5G-uhUA@ERQ^&>qF>@|$I2T8CKl2cCVTyzy=>p!lMnQ&Z&qL1uX~Sk
zTG_o_cedSlo4Brre-77)Rc1y`vn33^H)a31zkTlqIkm#YEYJE6)Tq3V{o}AD_gKcU
z8Jib1+Os`xHV)G`oOP9J;@NdSq{O!f{reeEJBLy9*Ka-P8$D%u|EJ|z%E-U{fBVy)
zugu;X=lS??8_z7+H`z1ad}fNOs&V2Q_Wv$3BJMlie|Mbum;cu0Sz7<CZ`M37k=gxK
znB(89%B45Y{9`RsP2gcDuFOrc1LnJ1-}YZWRqy_YhrvIPA%vTiTMtYCWx(KT;DSxZ
zQ+00gFa!Xlg;<63zywUXC^fB^p9z?Q*LO;lu6uQui6?|tg;yPJXb7(fuQ`n8*#3s$
z=a-UOOe`V%di((}VU`dvJ+T0oMP4)a_n&Ki!^;9RNsd<;rU=N><26R`MC3%2VOn@X
zMD#?AVLb1QA8+0rb-E043J0$wk`@jTNtmJ#KOyxO6Kz<5ig-AA_z{YPIfeOQiZYH(
zl{T&qeGM|nlh+TSNW@db52k4TKHDvK`rC6Ec~HEu+WCBY_y5ZmfWpXr-IDchc3;#s
zQ;?BXymm-NT8Y@fj68UmC+3w#SQJQ+1Ft)hA_oz7n4&`=b}!@QuOA00;%DJkLMRer
z5mSOGf{8M6fg`o|(C-Cn9$G6rHDLY1&M)2#)XO8vBM(!>6Cx}sEDzHQqk(!s(GJu5
zS0HTD-TWiZfO<t#MATubfIJfsa~O}IC^aV$ltzk+l8co=j9?G}2FxL=+5Rqh;yhQ2
zDL9BB2qw(Hker*6lL6#<r};9ou(GjpaB}nT3h?uJ3c5P0337Ql35g4fNJv^(THDy#
z**oZoYFjCrX~<&}6_b{cRZvt?QB~K}($P0CG&VJtGm+9YauF4C_X=fl^K<lI^Z@{)
C;mZ;L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_5d_instances.cpp.81367B36B3C16096.idx b/.cache/clangd/index/device_permute_scale_5d_instances.cpp.81367B36B3C16096.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd8b9d1bd5f6d103df5c1d0f8d5255815fdd9f0f
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RLvI(|HWPUIKP7BK
zlDn_PV)p2XQohO2lhq!(^Kt07_Bb>gb=~yqYh=fg)YE?IrTY3J-`=pM=LfJ&V7#&F
z(lqY}epfD+mi~X(lm4ObTeRW+oqK-S=&uude3J9Szs@}ob0u`bp6B~FoIJ|;!Q{xN
zXW2}7cIVdb`|H)7zyEyq)#<lg3wK8O9cK|ay8P*?h}wOBcJ26E6dMyK{Atpo%Q@FH
zr>)t>yXjW($-TEWU3+&y^iILv@289tXJ4y|%hEA>!IKr{wp2CP?S;li(SPf|Us+Xh
z=WV}W-g)C4kE1S6Un6bO=$El%mBVMDt^SLOuPwc=yQ_A^)JXM*b{1LpVpC%ePo8X~
z)!-g`@n(dEXOJ4h&4nLN&EgZwzgJVi-+ANUOO{44RddJAxK!TQpC@ik>Ja+l#~Rpt
zq|5zo`@|boA~qAE7G1dW(V=c%_$?3DhHE#EwY3*i_{%hBO|Ypc2xNKvXThR38nfNz
zo_gSXIeuoErH%4VGr=7j)9&ARv9P=PN=B-H)Vme(?jQV1#J1hh)SIfXR$=1qQlGak
zI89GgYuSWwZEXq3KDt%uxUlU~iJTkTZ#-@@T0BX`=>x~MTrK7mTQ`PW;LF}Q<@o0B
z0fkH14=i-wzSP4@NYmBi&H|2$lBa^5oZl*Dtz7)LaQWA*rB~`_#5io25caj<(%$p0
zz6Av@zA3TWc*Em!OWs)BYCiXwYhh1lE8n8$*)k#*g)f?%yK`dQBWKh35iB`TzQUIq
zt)EEWXxpN@(xWzRWtvLmALT`YCQf0pi@s=HxUj4FpTj)9gDbZ_&YkPz;(Su3uK({O
z_M3|0rMo(vmgw$zEs^Tb9KUwj=ELQF_phw_cXo+J=Jf?v!g4%JH?&Xc=Ji#dczIP|
zV%4<9vkQ1l?>Z^2EH<;c<n?r!PStd|#*?gzWeu~<_DRg`ec`$EpKtK}x|0v=s_#~t
z?a{r*Ij!v8u0z{yyiHu#!#{`X#I@X%7Bj<yhpg9sY|pQLYU}eTj8kV`!%^#>4<5NM
z$~AT~J~l^`caQYk{&Nv4c~*yNyO`~MCLMFZ>&J9w-a@yK7xQ=Lq!{1Y`N)0S)rz7m
z|Fchj-rlcsRPDS!_vGhu^XKuGRn7~Y;&bkX{U71Pq|Xe0cgXj@u)p4ywd&9Rlh4$b
zKfbZHzVSi5@p_ZlH_Pi=69gHGD|3_VfcehCI(Fym6bEY_2LC{Y5N=j(Jum^30fWCH
zY`U9%O0@7W1OTOlScUY!1WdXpHLaMR37CV|cS@D6dv%zJCxlmpR~>F>2(JmRIgAIR
z8M(k#^d9=XV9i5o1;_R`3_rh=++t!0;n(92fT>~$5z`Y3fZ6XgbASK2<~O`7Knvt}
zm0^m2JUw1x1W!axL>Z=qCqzU~#2Ci&&iL`>-BG8@AQy7*N+N0D5Rrr_3h@(Ce=*U9
z6{v`ZlZPLnNSITYAEqed*i>ob`q0-PlRSC-5Q;=RMf_lj=I^uJa;LvNmyrj>8>^kq
zw|D=)d;uto?AI+>|7Q0^Z8HTKX~k=YWTcgd9n8psmw94dX@o_A6glv^BPntaafc~7
z6k_)>UjF)Vpdx-2ekFt=F%~f;n4+f!tY6sq#k+xuctm;R5sHLGh2>$2{tAR`x|@IG
z8BmdkiikQ)7|1gbF^BPBG%!j*X#*Ce3`MCqiJ&x6T$Eg_3}OU>2rys{`MJJ9=s?6L
z7N+1Jh9H<Q14D9dN=^ol>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+
zYiIAEC#r3wY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)L$<5EvgV6^7
DBMzjC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_6d_instances.cpp.A19C7755BA5CCAF6.idx b/.cache/clangd/index/device_permute_scale_6d_instances.cpp.A19C7755BA5CCAF6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..965f8362faddc5250e985e6f0ff77a26716a9a64
GIT binary patch
literal 1674
zcmWIYbaQKBV_<Mj@vO*AElFfyU|<jg;^LB`%yec3hFDGphKf0RLvI(|HWPUIKP7BK
zQn;_hV)p2XQohO2lhq!(^Kt07_Bb>gb=~yqYi!4o)YE?IrI+>v+}JEWXRo8A16#ss
zuV}qS>p=hOSO5Mto_DbQ=KG_6o_%_CdC!ih9(DBtyVXA>l_qV7n7iMWQB7R^z-58c
zv-LT5{507A>zB0bzJG@LVe|9Fk3P+^lobvTk)OUg>Fdi+ryqU$Tv}M^d&0Y`e{<yW
zh)uW5PGp@|{hoC?YPWNE!Qrp>w9mKXM!l|Fx$(vV;}wyTQ>XbzE?j&h{Kx)#!K*J7
zZnt;aVSD~TU$$@jro0O*RtqPtVmj-6#jf*Q<n+4jFTO1dOPkyB`@xFJ@~Ne~{(h%c
zFvyj<rYA2{@tnz!(tb=k+bUvD<>!C0DoO0ig;}De++k6znrd3|R5{&S$?Jo)s0*)<
zwp^*aThi-*4-Q$KPQ}NWfBuZiR2F57N|%<FJM_@DkY|<Srw<2R1=oCT?%K3CN2*k-
zncufQb=K1lJx^}AJUBM1HhFQo-n+oXQynvQEUuF|xc@@rjiTi{LYlTTxxBk(xOK7m
z6^-{RKLluA5eix>dc8}=|I3u54T<-Y`lV0xcunFssC;AFN}dJRjs-bduX!3IcmAI9
z(W!C_?Q(agDQS8x5xH2<tmKle;m^stwSC3XzP{sguU)?u_&>3T@t{-GRc5d6cB^lD
z`}U?Kzd3iH&wBEv*BSg)=e1jm!bGh)=B+ObaPfD!Y*DPdyNmy7O_Jb-OjBPU_IDHV
z6QnO~TcY%}a_OvzPd=XSaJ#?}S=w=Fxr6g-zF*9hRxHb}^=&WZ6694a{$cygTP~$N
z=E_Sou8G?|tV^E4$XmBH>KxyF%i7@8zsx5uS`ynF6unXL>H#?~Ju{QJE`F=sT3<%8
znKheTDdBEia_;5}5ACV5H@%3hVo?+CDLb;}=8vQjvqkDtf0_H%{x)s?{igK&%@5lu
z)FW<HzUI1}w7tF6$gV(3A$t2n!COZfTE$}?@89=!>UV?AXw?lBOk%=6A3Snjlxyr}
zd~A*=?;j&8`MDc|B*RvRb>x(tHQ%vd>IeOHsYg94F5EA_c_M8~@zL!$t3N);_<w%d
z^L+WpF5h|c#njGQ-?x!}^T}ql#>_bz<^KpLCVgh8-7Vk$!v1<+)~Y}MPd>9={`kh)
z`3@iU&kUb+_Dz1r(*!Yw;>z44J7B(32={iX6ndA=!{8st5W>yMtp_H6GGNdpnpSh$
z^Vv5Zh5(?n5UY?Ln1D$arKT10GXZn(`cA3Rb*~OH@r3ZI@T$WN4dFH6HHYzFG$R+-
zirz!N7p!?`t>D=HhT-Rzl3PqHA^dv$0Wei8A!2%B0WkZ$X72Al*ZhW;1!#dBuQE&#
zkf+CMjNpmLi73Og@Pvryi5SCp-WfmMygTZ28RS9^UP&Y^93ql1MInAd>MthRumTnF
zaPsgY6bW+*^TQNn9Gfa_Tp#)xWRfSZA3~9cr-&a+5sU`<6%=SNzs}!hyX8)QdoCjn
zitkoCpKtH}fB6DX7&!>GWc{1n7q!h4WRn%I9g<B}B6ctv4qoPod8H8+1yba|>yD(z
zLBt)V=un8=%Xs<g$AOCYS@@L@io{sNlwgXU8nAw0=NIn=D&i64kw+*J78RC<Df%lA
zw&`yEk!L_fA}S*4Fkv9iM8q7%V<<|^Nd%>l;-chYWe_76M1TQvh<)_`Sf9qM3Z~#7
zh9H<Q14D9dN=^ol>z(Gy%)-jX&cVse!z;kg=PBswtR~3i=_DjBEFvLkVQFn+YiIAE
zC#r3wY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T_Xw%-t)L$<5EvgV6^7bc&;j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_permute_scale_instances.hpp.6002ED9F3C6C6BA1.idx b/.cache/clangd/index/device_permute_scale_instances.hpp.6002ED9F3C6C6BA1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c7a1340fb6065ae3484dab41e49ce2dd3dd90d4a
GIT binary patch
literal 2070
zcmYk74@?tR9LJASy8gMgR~LGBPw3q}U@LBfjbdoWqzO=*a|NuEL;@icVJaXA?Pif>
z5}m1llMOcAhWm%f%qDDPLpFuX%o!CUih)d*Y*8|^WGIVsOEza1d+qXiziay1&wZ}%
zzQ6bGy$05mmRj@-Q@mweq@^KTx13=Z8+nJqZQHC#3^T$qOytOLz~96^$xoI4KHXF>
zv?r>(-+g|pVwdIT--Zg_&dJ@p5cucD_^Fx=+kXA1$TE4UBj=BMgZ7>BzQ)?C<CFcz
z*1mM_<nui}X9C0BC#$ADo=^78Hgz4E9XLO`X6>dOCobR0%y<}QUP_(#K4szSjNzrN
zUmia23RCsxz=NkUqU|f@tIXm|fAzrRg}KU&Ldey3OSsUyOZUkSeY(y^+<Qt#issVJ
zjvU)EeRSc~<cl|heDz;<vNOx=*AI^k>o2yPcDyULD&a+Lq$``Vch-#F=qNsvdVk^t
zDLd^{J{p{A_Ux>vEA1bDX6fpQiL%8U_ceL_wZ`%{Qfq26bl!83?z`@rkw9%F!|rE7
z?JdE-V`KDr7kcrt`R`2ZG6TbNI?fPp7z1(S*m~}*m9E{JZkt#g(KemU6mO`0?aYG@
zhvo_tmbH<wj~%?xXk0<{2j8q2{joRNVq!V+YlqI^j5mydB{RY>$BVB-dzNOVO{|IN
z6kUol-o&(#=+W)}4jil+TWw;KiMHyj#&|<J|8|C}o_AJOnb?Pj=5@R|-cY^BSAOwR
z>)BsS>~f;bI<q0(P~Fzh7<w;RtKVRZ*G&uJc)=+=MgtOZ3k4d9P7yfKEdnPwCEz5t
z1RQcA;E)?79RBqB@s)u|QBQ8A7t+A3^dcZhF9AY&1V}c>4m#i*uNV|SszC*W4LCvh
zM+^0h8}9ceal9?h=ArZLj_lv@X6LFwa+=TMy>zUfedl!V<-K>zJc)^gN41zju8^;h
zh(+MUToE`4OTbCF5^xA3;7~3CPEMDdT6wZdc59@hE5Io(1voWb1x|ITz~OWZ9CqP^
z^XK=q+OrS5PU^SjS-td8q{Bm=me^VVPm5m&fcpI+AjvNQLVg5DE|WKcI?5D4YMBZM
zm*E6O@A^_bR}LVOj=b!24bzU2jPYeg)6rbybZm_!PT$_!zIpD*L6+m~2D^n`#qo|O
z9II(pLy__1ik89B*n|g#?b<~`Na)Z=42r;sArUw!C;=yhB;ZgG0f#~eIJsVKqA@sL
z4$CiVq|__GDPaXTwO$2I4XeQ6dJG&6<AkF#OlCH=)6hX8WfuxXpwL1QkW?rELWKyB
z?3If_V6Os5^{Rld7bl1=G!i4Wqk$MD!g>&+L<A(2NPtiY0wnw7=Rgdf0!a0#fUplI
zh;E)FMr`NQV#rzY<63vJM}Ar(B})NL@hHHlSt@X<M+FXNVc@U_Cmhq(&|DY$rwFyx
ihjPd#%YPM}-Mwm`)Vj=Uu`=eE`~R@eFAQVVzWxU<R2RDd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_pool2d_fwd_nhwc_nhwc.hpp.8F42F3A58241D669.idx b/.cache/clangd/index/device_pool2d_fwd_nhwc_nhwc.hpp.8F42F3A58241D669.idx
new file mode 100755
index 0000000000000000000000000000000000000000..416eadca25460fd7ca32350d265cbfb2711b11fc
GIT binary patch
literal 1690
zcmWIYbaR`;#=zj5;#rZKT9U}Zz`!5~#Kk2=neB`W4AX!D6?66mp3XaDAaG{+lLLmU
zxW8m7E)fe64HXFqsP5_6E1=z%tI9WJ(ZA|uXT`>rdsQ1ZS#LI;9l8AS4U>(L*%MTl
z$`yVenUEP)SDlx9*-ByYl_?9#3Of#_bMCQfuoZoI#es9*F_R-dcq6<Lxr_oodhD*W
zyu_SlX<o5x%8grTkFOovmj71c{<VFD@~w*=w$wk}RIaLXUD~f%g5!0kiJ`#zmET3L
zOwY>tyeY0s$)c5cbL9G}BRNxNFk5!MG<$Gl53{M3`FpFmUArstu6lCay=2Pucd>KD
zyPMJ#?+X9?JmC51X<6)x&^4F$oX>ds*?Muv;gvqyUw=uQdGYwo_-!{Xcl}&{Uf%Y%
z$b_={l9L4F9=S?2nj{}h-L`6TsJWSOd}c@IL$9PWZzPV@AK1F@=94+x2RIfwxHdGk
z7)<)0dMS%_qrZya>lXEnwDu6rnVcV-J5)K2sETSbyEvD;=@zqc-D&!0qS)^I_~`oj
zdAbVCqFc|qF8FG{{=&Vr2j1|#DesiOxM<ByhT_WHq<CPwR5jI_?L1W#%gw~X&cLC<
zr6L9<7&w4(K(IlvdXw@Gv1T?VW;O;6advS9Fu}md1dK}tMh?c?tCiR01V5TDz{CpF
zZ^U9G4klpw8H!TViamgFF3;C!*~xL{ArlXWhq#9dOdAh}x41W)r?{&q<4uE28xsqM
zhmeOTOc4u*w~#lSr}e(C_sL9C6-E{gF9|Oxm?Hmr8=3mJ3<WM84nHwJ8JI8+hrhVL
zEP@vx8X%3}1&ReCsSFYiLQ)wj9*V>Z6Awe;g-eFZBTR{qh>%0@A|)f?av^>~>MthR
zu=4P5n97<eBNUm*nj!JbWzCUz7P1ydJWE+iB%YP56%x-{)*6XtBWr`ivz4_);@QdC
zA@S^G?U8t13SMe3m&I<k%!yG`QD)@fFcdLVfC(2?*qxU-b?GlC%%#2Ho-JN=Y0spN
zgj#kU4i`xmr0{XEcX5X4<>7FZbVZVLwRc65bCYp{+rq=)=Hljo(C04W4wr+aQzkBO
z`h=xR9*$W4Sa^hXdoj1X?f-AV2MlfvQ+<R<W*TNlJaY|mB%Xza1rpCv!xD*SrD27{
zv(~Ui;@N1}An|N9Y>{|&8g@uLdkuReo};8A+*dptPLfVYJTGl8Bbcif+~-tZ;2aqR
z4D)EwXfc>D%;$_;;GF2GQr10jmvh~w>XTnA7paN?^Ov}mxGPK*4~L4RiU}<6b{s7+
zi=3nJ6_iJ`{S0760`rKjKipm(jsVR7J%n7Kb|8|<Al)D&m7%(!NW3uJFeF~MVYnH>
zlnDI@69g~PFcK~Y69sw*RD8f({?vf=3p>AfHz*}?8VbWy6{Y4Rg6b2Pf8iw*Lvc}Z
Zu{fyA;sg;8z`-5o^Q0r9kdu*<0RY({l_vlI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_pool3d_fwd_ndhwc_ndhwc.hpp.821043598893E6B4.idx b/.cache/clangd/index/device_pool3d_fwd_ndhwc_ndhwc.hpp.821043598893E6B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cdefc863784555c9a47e36a569fec37760705cca
GIT binary patch
literal 10036
zcmb7~c~}%j)4;cSSHxu;IfO+7m*pJfh6O=DklT2G5Fw~&P*Fp~1B{w@f&}FT4PHUN
zprA%UO~e>Qh|zdo!P{uOC4dGZUU(#ay|CqUl0WjiJUYEUs<yhOr>DBAo0uWt;ZF7v
zNywxj^HP&%C&?rdi2?neK0EEB+e9MSEtg2<mDR;vpWauc{~-2FalNnYcPq!mjkNKb
zeqq!XcDdD!1_qYnCYk8(n)U0}w>J~!EKHd4<sh3qlT>RwJ*W6AO=@WVG1j%mP)GHd
zF#Agr&6;|z`k{9H(k*}K>~jg0noa9I>+x>rH{)RX%Ij9kbr!7H_34Tm@^NnN#=rEP
zo9cI|{g=)DuNFkM7L2rY&WgyM`j=@)@;<wA&zS5p_U-Yd&ts*v>n4wgUbI85%o^jj
zaJo<8cYe<be~B6X#iR-ME_{5d=H7)F9TlJ5{lKB3SD;L9vz33y7Ux%U{Zh}*Na(w5
z#<jV!wmIW2?!SJmbf0|Q=9w4XjJ;v^*VrqUJNAsOZdg2c%qahm!wsLWT;SyYWa{_p
zRyd~RWalp&X{8wGRIw~kFJh_Mtmm&)LiN;|SrHQ->sAf0Jh^0X;>`AiZ+>q*VAk-t
zZe_*F+Z7wipXycyeDQqbz@TRL_oAY=jrlaD=cSu50n4s!TC2A#CeWeE%SE@|<?f2m
z>Mb!x0yb^`)oXp;nSSeQcBOkb4Tubno}2Y>?11o;;(%XLPL_Uj>E)=@y`P8A{i-VA
zC*rUc;7+B9?^>hKKGqw5>FvJitJm4m{;v)1eobo5uC?(`S!45d$@t<|pM2wJF>ZBb
z9KXUpsFz{>tjTE^o>9Kf+6~6+)-Q}6e(C(sqCRih^0&NrPk;IO;H0nD_W%CiG`?g|
zu~nq)VZ-=GKMkCCYs%&NPhQFGa!uAH&j^hkdi>$a?pCeq@~>MPpUBFbUhjPG)q>k=
zom=dx=XZ<9@^|X!XPP|ylK1aRR{s?GWZIA0W}OTBeqOTv@zTC^-CK{xn|*BhCN0Hz
z?7sPyJKe)pZ`OUga_52Yd(rjA73WS&zOJfYR9QE2$v05Toip|0X1W~ET~;?}&QaOe
zd3#k0t|))(CAGUVddd4Q9b-BMF1=VVB`Bq1tmV2NXW6{hlyH2(1>K|-dYQxC?7I4F
z+6VDFR!Mr)l$T{3e%`yEFtu>!u*H({@lV^*KTVy`K_|W3eYSi32TRBB#6l$zo`U65
zaMd;vUHU%IA9;nNx?-{QH~PfXQX*tYRasI`ZG$R(boRE)^pPJmCcUVNyP3Or=LN-P
zg&lJX-VAm(COj3pnYo#FUQldfR_^z{VYfUz(v6D!<bJBo3yPg@^%|NF+amRenUzF%
z3;bI!&^D;zF|nREc8EExNA#%$AGwcZ=LN;;L$~KGeeYX8W5Q9fi<ygg=LN-HPGM)|
zI(ya`6JshKq#ESbc|mc&{<BKM-dinr;zoz)1*l$tm$pHbNLl;H(<a%WI>eA#_!IR0
z1Ve3uD#Mc=x*isCTlGkHs!|W~dN9^Ds50W<qMeVXer?1PPukHdP`#oZbv39GV{<Oh
z0(Q7}jYTctYl(r-)x1*?ULWE=?RD9h$Z1DTW=`gv7Zk@Det#m!HG3CNJZL8!py~i`
zZG$QaH5cnMf2gDu`ceyO3$=ac1;vR@A56_(W-^v1?$p9dP`w1Twn3E-AHKZR`+KAP
zJTa##e}Ks!V5e<RW%9GHZyg>GAQ%!8+R0%3!7iN_6wmf8IsMV^vyyn?MJ=>~aVsdb
z4XVs32=FV*51^CIfLidh@U-i^pm=^+$VtbTysO5fHx&n}0^K?<C|(%gqr88y><v%6
zX$NhfXafgrgDP2zbN(-W^Ory96C-Nj5b1G<m}?tUS#vi#VBwe3_v@1$R3%5MpCh%<
zHmI^OdG6RgmanNFdQ`>P%-Ov2g5q!fC|_~RWZfNoqM+hyM0SlRv<<3!JO9K}-;K5=
zIz(3~5v;mfS#@3{JqXa(69o9~dCFJMW{f||6E(e;-@)v6u+}zc<5d0r%YAA?`<W0O
zs&WW)4}re6L6x5iYayVqvQ3Z3s3&zGuLFg)L6wGWU(AhssgBYkQmV2a<ooF)?rKow
z!v3t+%Rhd<mM5;XqbHzz0?yh7Rc;QwFyT9s8Ie5cM^zp}@5f-JZBV7D&{a}<%Am%W
z^rRiRnz@>HUQpcHG-bdj{jaL@NFOTpmV5W<yrB4b_6%q5>*E&b5FIK$2zm#>NZX)F
zTJq#{uYM8<DIxAvcYX<$a>8=ZUk)}nFF7Fx^mACw1^rx+rD^m1d=EDU0w**A*9b;9
z!aB8Hmi}Qo>U22aGmp=Pi}td>D+{7UUI{)cAy(vd;IR%E{q5kg9fpYVZ-M(3?2$2W
z6HaIb+h$N>|8i~k#zDaF4RUSJ#_6PI+OmCr&FdcY4t#@r<1mKybX&eIrPTd8_1{V9
zWE-OmC!FA{bhgFG%n2?^7h8<0oZzZ-wZ&(~32sU^Ta5Xf;I4GH#RWpKby~=}z>NE}
zPyZnQc#Qm}SH97Gvt_Rb$b@>Ls3)cv=kR;fdCjVVuA_1KYt3qczgF@YzB#(5@6d<&
zh$tQr8+^Gqp^YfoSZ*hZc9B*4)22RJ;-4**(JvZ4BF6Yd8U3PR8_Vs)uw7)Uz?yyS
zK0nL<<v8A%Oy0Scp@$rOIJ!?kiBFFc3W--C8HQ^dCsYuJ3KE3uur+bR^HsZmj+d2_
zl`op<-&R;~f324dkqNEDw3V2N{yrwk$3%tfJn4;}1zC8ADxM}zr-=*3G@0--asHXO
zB4dYiM7vg4T+&>kCPjE`pPNUA<`|*>`r|ib!Wj~FhD3<mK*AeHB(g{3{al?s3Es5p
zGsO7}+jRqRX%M@{iARS@vvT7-ax`k_K7oE1P&lE0comSL=qV>`A`Y8KATmxydIy>n
z9?jt5E|qsK>iBC;NC`@b!k38?wi3s!WFXq|^~{dBd3gIX87C}IEbtK3mVjajFuV-(
zmw}@w&jbBDkq7pBeDY(Rq{niZP$=aJr93Vd8ulxc28diF<%(E&k<_tB>deX`iljqC
zE|zk|tbVc7v6$5_mPQn_`Xy4Xgw-#R29-#I(H^EaG?-ij{vue3!K5qS!C(?`q0H~g
z(P8P9oG_Q0Ya-tGJZ>IC3%CUgWdN4}3}*tD$#OPu*(@&y+vVUY+FlMm%OP0g9I(w{
z<vHM!!^(5PHkXy>f=@0h&j&7_?XLj10+vgGEB!k=m4dg}|0-}=#qw(4R<rtRfLp_I
z8E|DRuLo{D%Nu~(!17k$wzB*+a9^`r4P5o#*{PcKX9sXQSou!icCx%1gx%~Ut^uKj
z<vk$mVR<j~+6#T~{xLAn7pg1K6<?`rsb#j*9v2)=$d)=}OWlx%M-5OM4_h;m#<C(3
zT|{E>ZNt=<27s=-ShV;UbRUBeCJdaAN4)aH=*S5hiNi*sMnjlD(`kGK_$%x*?n<L^
z8b|ZT2lklVwvf8Amguh~0$RXzhX#?ZtQzkYmqF^ICZ!fVtb?FB2*%ur6YhcSJ+R03
zhZ8mvzs)27?Z@^At4mGRFQgskNPFZ+d*K_1X*qTM67ZM68ZF4R$vL|I{s=?+rV{^d
z)|@a{J$N8K6;23N2Qw6+4q+%%9m-IcI*g%kbvQ#2>IjA+)sYMhQ4eA02$&oJb99pv
zj)KWimXCqSF_w>m$#IrXfXNA#PlCxwmQR7nDV9%z$!V6)fXNw_&w|NWmR~^67hr^u
zjibZZvjdE9oH<&t^n3|MSj=#=(CYaLjBpBZ!VO|^gIME2$_c*`i(gs3Ni1%%e2ZAz
zV)-_)xXtn%VsVG%yTsxy%lC-IJ(lkii~B4;AQlf;ZX_0sEVmNNR$_<##wC>|RFz*J
zNGIACQll@VCOFYPFgoV`;SN(<I??91&KWAspDp0F1%@GG4X7jUto_9K-b7-XNFsJa
zlTQ-c%yJ9q-$K-)yoJQIkPk(ENcuk%<v9G*o6Epo1{KaHxwfLjX`i^)ciLDx8*92E
z9|Vho;D7>-zNn_P#0KA7j#g;ahs5lH6Rr~Tt3*H#ur#GP;(6fDLqF_M^XhW<#2F;l
z7OdD2jtq@^nrc-Xky<dQ1y?MMIN=&7uR#Esoi%UJIrEd}|D+jWE9u)x`k^D~+s+?Y
zvHZh@-8rF&D4W<v*+hmkk>RMQ8DnKLD{m%4npt@Zv1}2OSx#sn!&*qR`2B~(@}Vfl
z(wdI>&%pl-rs%3%TVUfIV}VU$o~F7M@AxF?dy?2-qNnlSp^12i{WTMZW|6TFC-Tk$
zTr`s!l=?0Mo8M7T`(Tc#gw9uQTW>!s@UTj!zBT~g0G6V!T_rpEiq6o$-BqY#H_OKR
zmu}4798I6MuYs=}CY!YWFz~m-ES(M+=(m6^TF-k=n)Q3pTc0RS@HF-`$ARDkwXxb9
zUky$e)Mt<-=9ioh*eB2u@0AAUz`@}-WoU2?3}z@KFodDdz)*(50>c;z4-980A~1rX
z$iPU3h6D~_Xed9_SnPcmKa8Q_{BVXw@FN(C;-eTE$&X|xhL2$=mXBp<6hDfgI6jV{
zcs`z?1U`YG(fnwJ#_(epO5_t68pn@gXgoijp(H+up$Ys1h9>e88A|4p8Jf&bW@rjO
zg`to5j~GhfQy7}YPh%*RPi1H(Ka-(Z{49pj_%w#n`E-V6^RpQ$2fiE_-U$3gmNx;v
ziRB95D_E`szLMo5pgIBqPH)X(tvbr`F;E?2`8cSKvwQ+nCs;lSs*^0AqLG{B)1W%d
z@)=N_Vfiem&a(UhOkaSKU93$zK*=uFrY}K>tA%E<Hhl$3aj{-XES3@*%(66Dm;Pp=
z5P2&x+)Db0e1p*S-BpZOzY@B>GklW>H(9<#gj+1%Cc<r&?-1b*%Xf)zm*smzxX1E+
zBHU;B0TCXs+(?8*mRpHqs~A!9T0$>Rot*aOKTGTU_I*FLg(x5V&&7M`o&E=d>ML4k
zIL?!r=1FZZeJFkLVPa-4^GkFcehbcT!B-s9w=ncA#EX1@m>yu)<pad~013u*dew=@
zEf=yEInl`ejP!m+l;Xq4miEq;DzR+9o@r$7+F!-UzW(LT&#x--ENEme14S8I@s&Y9
z8H^GG;BN5T4exRt+{B_Qxn?^S?QyjF2&#u*EJrxuKG@zDbNPyp>lVR>UVZtW9#_q)
z?G~N7F`O338KB4jrRYco*kynV%U&4}z;bAYn1pG*ADaPVM0qACGFf|>V3*0-%LK1X
z)?OxrX0rA&AvTk>mk)}3)?Pl?<+Jwk!7HD&mk*)&ti61Q&1dZufTDo4R|twimWx19
z#BwnxidilJMG4ERL9zPp?6De{B4iCH*0Az5V7G?#e?2JHv-0&|x1N=60L2DYz5(nu
zu<{+C*ulzof?_AjHQ-qTe&X}l1D<<CuJ#-lePsQItLcpS6U_evd#olox)s|yOPu34
z_vsw(Isw2ri7gdq_Wg(0%*iX+AF^oA!_W38a@)84YuoWV?{C-4Z+H|*1$k1VJTdLZ
zBLeEmIpEKM0@nh$_MiY0sRPF^B~<%8dGIgO2l;`!Z@5)Xq9LIf44R=gS~$EldV0@r
z-K_t7FOD^@Te7~%yFo{@{pC_9b;DO%CKO4Hi=+ycJ&UAXA{R@Ii&=TG)U%kCmq_hP
zqz<TmvhTdRJ?8P<WU_y3XUYCiT+9AZb;$m)Z-BR`$tpXKOmLLW`P(gsi85dHuTjE{
z0Gcy*ouJ?o<Y+me`v(|f{lN*j#4DEsi}A9YIFu7_WUMD>km@>M5e?OWM;!!UPnxjh
zQ4awawi+&f67OhvbTNG#bzoKp0;_0N55hae#zQmq2VVTmMkWk(8TxL&O(qO;8OG3X
zm*EVJa2df+luHytBV9+j;j17MVq9Ywigk@;Xq0-Cm#7)1j$<fZ9nVmLPlB(gIofA5
zLt}i#FqG(@_%4vjgmM1k7#i<Co}nb~qyeJM3EmSJn&>@|p=6(AriaNslNp-gGlij<
z>Y0q!v(&Q~N>isXl&(%^XtsJbL*?LAF7D*XgpJ^}k>yR`wTa~l@Ty?B61*x|Zlni#
zFbHR{jGj~_Ho`cOe}jPEL^<xX)0xu@#?9;=RM$Bk&i;%0QVW-qezb&!f^4FnEpEVF
z>Ns`J;NcE$x+E`DEOfv@=7dFxMGP%gEPh8<^o}^h=ByY^JGGCqA0y6kAx?NlR~{^n
z+kC+Rs8Wh`%DZ9bgjDO)cXajhE2cW1m|gy#lO5b%q+VYD{(`s#pveHZ`(TCxCfA-)
z;yB##IT!Mx$`t+;1@4<xIu$qea-sfKk^WW01>b5;s3!fZMSkGc?P>R3iae@#l^9$l
zeQ_nwtPnl2q(+zp{pPf<+S6kDMmpr%h|xCUfEMx8jfN=AF)5B87SOcl)*O%GuQ?%~
zc;%B&{C0cUz8_q7scimr2T$~yA1%#^QvYX)`F!EgMJBI*?DgM@+C$mya_v#9=+ob0
zt0uEFeG8W8BMorGv7UHh=Ew;@5&kD)f+fF(J$@nsMSeyspNTs@FJAAgQnaKN(8pT}
zic;}-h7(qSVin8Ppr{u4_3O<)xZYUxl#Y23@hT!C&?8A&^0cH;^z0pv=OlP9p;dZ!
zd$%o)2Xq$qtt2>SFei{;E`VQ^>{8AD*1E%FsU$sZVtSG0FK=u$m#&}w_(39Z5<fbK
z9|ePYoxwh)Zm0E}U2N?f``bG>S=&e)MFslDZunz^A0!eN@uRi)(OLXxCw}~YfW3A3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_pool_fwd.hpp.048422E56031C936.idx b/.cache/clangd/index/device_pool_fwd.hpp.048422E56031C936.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f68847155416e94b2cac6858217c34bc72c3f0a
GIT binary patch
literal 1110
zcmWIYbaV4#VPJ4h@vO*AElC7283cj2xTGjEmyv;CEi(f{#hkqp<Gl|Xh#dVL_-jG_
z<<rmp8?d;xG#W`V%X?3|m%A`G&m+&Q{m0$2R_4*me<(REJp1$MykdUeU+n8+7y9<;
zOD<<N=CPTytm5?&_Vs(`{*$rK*vYKk6fV}g#=!56=d{NaYfipY@AT6xxhB)KO}FA+
z3xAPoWA@)gPp>mfPim<QuyA{Pk~epkCEr_x^*^W15SaQ}QF_O@>fL_7s%@=b)*bcx
z<-5D>hQ_sBw_aSiWFEbpciXLgRjZfR|E`~G9h1@OAGkmM*~!H^AC8@MYHgeEA?+5P
z&AfR5)7`M3R)Lt`=_waIucfq{$ktY@JGk~pMfECg5e9?hT#4^e_4dX!ELGj*BNe%B
z&$k6DHa&T|<cQ9@#0@ODA(BNbp$D$(%(%%gdq;n`*+jYi=hy!E&+2`;>7~*C#YYb8
zy0V99`@hMrPw@Dzs5w1Xr+M~apOwmeFXeu&xXtW-L}1VF66-g2kFhI7FGwmnX|2`~
zxsjo`GB+t682U^sSJvJ?Jn@786AL?#3j~}}EK=eO91xD?ir9JIrfpzkV`64w;1p*U
z7XcFtoJ_zNV_@WDoV>D5#!gOsn*b9lOuwe4rZr4ILs4p4@mHW)m_3Y~VDn+-!e~Y=
zu(_TpW!)2ZIoE~w38}xBXv50Q!zn8(s|?f3!zm{#hs2YYl}F+!$SNT56lE2WcuKNL
zNIYd(Wh9=8tO^oORaO;=rzWe0#8a15hx1~$Tjs>5sVFn@a7xNZYQUVB+Q+KeAXMxG
z6c(2g*Mj-4+l#s7ZU27@kcTv6^<k=k9@3CQ;>l~sBk>e86p(m|8j46fB@HDcp0b8A
z5>G`#1&ODsp^C&)(@;a=scWdid9auS#t<k5VJ=^5Y!z^HZo+h+hcsn0)nTfh8nAw0
z=NIp00(xIiQX1hQO+ig*m?B{8=Oiit6AH{QhE3HczgR9(6@vv6T&B1vxmX{R=D0uv
d1aQ8eRk&eJ$E(}S+#Fo+6a|#yhN)m+008B2oS6Us

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_prop.hpp.51F1A892B5410B39.idx b/.cache/clangd/index/device_prop.hpp.51F1A892B5410B39.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bdae4b0f91741574208df65e3d7302852f21f55e
GIT binary patch
literal 2020
zcmYk74@?tR9LKM4Jg@Zd+5!bCJvb@_5h$>BE&tX!=#}PLC+%h;bAM2)Za`cpB#;nX
z<~DUK+cZExo!}NT7BdCL(2x<gsKdG87?B0oh-_@bh)fdHh_XMgUEbZh<bJ)+@B6)b
z_kO?ku3i_<|FVW4N_M##YW#JTBtZ}=cn0cfUx|?s#3Oh%oH%{t>&=&P=(m@%!{uwM
zZW*$RyCeVF{(vaI-7!(2b@mkwo==P3nX`DP_2JUL|JbGuH#V5U1E#T=`*(-;_Dtkk
z6Q>7n-a8N%mR6Bo^O9y?OV>#c^RaPrOIlmirID+%Z9?0|`kdZ{g5^!B!m7oq(bMHk
zJ+s53tB0?w{61G+mbuirZB_Z>V(<EQ+truzyWe`j^Ulz=4M*9UuZm?u+pU(J*W-%z
zeBmE>{`J#8eMZ(#ovp6iS}3esnK}33Lf^#k)&1i)je)3_JKBF39cmmR`D>pnTszTf
zsOCR+)X(dVc8z^-A^vhi26a2-@$9baoh>`8o!{R5bhvHn`SC{;1#dod_SkOJ&v#xV
z3M%@~1sZF1d;@Dlem`=2y6S!;O+-Zyz(u>*P$1+o84*UxfGqL1l|QRx<pND8prME^
zN)H9eP*ikJ(I5TyGc*wi4T@+*N+^U3%5eCE)=|}=gIV)uXo7(TGi^=`1?dSrndMD)
zy4#m&Vl6amp|`N1AQ^r&PL`RDF&s_M&`?5`u%RFsYW=$d$0DTJ$Rb0t=smgdqJQYf
zpkiq>;iDaK4jBz}N?kO1V>C7B4$cd?)A#O{KT9hQKzFVtHxaqfZ(L+(e#lJ&3u9r$
zIRcBxQX-L+v0|FlWW}_6CLhymjEzMvL<5`2Rw9v|v16LuWXCiI<6y=50EgL8DiP1{
zm?kg+E4CX5W}#FfH{-@Mx7m$p9>#-d9<vA2yo{F>+XuX6Z>dB+#)oM>vrp7~a>sCu
zU-=%Ofr3`3QDteMj!|nRijRrMNXuwrB~6#0OTx&Y%S5zCC>|jB-UQgbth}rzQTo+D
zsXt=Ze6XuV&X|hQb^W(C?mMI32Gfk1VNts2%Lmm<^G_RKIwv_tkJ4|1O+NLwzz`cn
z0+wRsC{2PSDhZ=xDj6e=;xI~~QZUj}dW_an>oH2DQZX`628`0EG>p=zbc{Ao8!*bC
zGDLcRX7!}zIo1{|O;VOv1GHS>J_znkw6fdc4r(?v$25nLK+UMdE|DOfi5CgA8{Qk-
zxai)V&My6T=fYHQByFlT2Nh~yzi}>kUF{PZ5?DBk9t{kMx0Fd_<*b-y<*k^O&*fv9
zjkD>|J|TgPx0Oj`=j@nf=k1v0;2e6iPe|Z!I$RR*9FJ)NC+N{WA%Wl&ToSoCH>SCr
zZcOuV9!&E%J(%X@yn0kP5_p|nmqb3!hiN{iPt-2(Z4GDb^OG_XXf;}Ksb0+Xb>;oh
zVux+YWpZOssY#Hh$YT(Vj0j_Mj-CnFkJzYKaXpXU7+>v<=tvKvfl{Shhuk#Ks5Bx?
zwVs;mojTPF$JN<&n~(_=40b;FMMj;!6Z^C5=xls494U&762G;z{_0BjcS8hfcLt2X
f3o01?Kji3wAmG0^IzV%HZC}P~*&l{eYvja#f26WE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_put_element.hpp.50E2F2788BD7B2E7.idx b/.cache/clangd/index/device_put_element.hpp.50E2F2788BD7B2E7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eeec123aeb1573006974fd4fc27349a9bc95b676
GIT binary patch
literal 788
zcmWIYbaUfjW?*nm@vO*AElFfyU|<jg;^LB`Og}~j20Nf!#hkqp{j-@21&)4S{N~`j
zEzR5ZvP5TGY-YdcoWE4Z(E93wJT;ZcTf2YUmHvM<bjjW+P6|){8ol2;Vaffda@Ol&
zy*h@nGaFNWO7+h!yA}0)mdf$-#uepFlMkf7_}9|LIxVAOsjZ)m)yY|B-kocfm}xle
z$*Z?7%%x`qEEC>c>Gdl-YU}0Ktx{Esy`A@lEuQ=D$^%Wwgr%Z>%<^rMH^w~LJiGmE
zqSq~hU31NhM0GF9s%{l&x>!*pDeK#MqCC&(Pm{j$;{)w_0SpJ4yjohP@BFfh$y>W*
zrpcxid%rDMvE|9rH3xLU3?mxMR&iK1MIB0&Ig{H^`zY~r+NH;D58e9jxw+B!j{S#z
zC+62q#k_o<7i?Q8d940R`nwd#J50~KN@uM2aAIyAzl@;shj&lPE?0MQE3FPN+}Xl(
z{8;|2CmZ+F`^yLCE}Broxcs<DKyhVmQamtpm2y@Xe@f@oWn*GyW8f5F7ZC>&44h1i
zJPZtsoQydF%)h$wd?yPqv9JTx0|BQvi?}!g2ZXaGWO?w;34f;vFtNg<<uv82VbTmm
zscFSmfJ$NZFmi#-_f#qCp18}oE_S<RPK=t0G9wSCn2eYP%(B!zR@DZfVke-mh@yxt
zOfyWHkrV6&m~gijbIaTQ{}xO<oKg}}@-S6AoYE4~a30K|K)oOb!t|~+whA~pH(@%^
zCOH{7b(pHB2CQG$`Ng||Mv4iFNh6Gu6O@yNDJn|MNmN4e8v~oF^8A(YZ@Xa5hx@v?
mD7jc46kuE+0s^1`2;y)<g@6?2@%I~f<!;30FmrQoF#rI~QW(zw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_put_element_impl.hpp.CF1E04C6FC237429.idx b/.cache/clangd/index/device_put_element_impl.hpp.CF1E04C6FC237429.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dabf15e66aa4c8db14fa26b919d2626e4be382f0
GIT binary patch
literal 3456
zcmai0c~nzp7Ju)F5W*t~5FtF0CV>zFgFsk9R75d=fXE_@L}YV`*n-t2ARvNTP${vc
zIxMFur<J3Cg8PDi*1e}#M;#czZM0IUb)i~pi&P!w%d?l4!ylcTllv|Ae)qe}?|%2q
zi;j#ew8b#Z(&)k!y1Yde7>3!QzadZm{zL-96jm5kxbt*!dd_V>esBCQnvvQ#*{edo
z^hIvEmr_~l65l1Pzjqin>Y}4kx{h?e4ek2KK6@#@<5<>~ysM{2cIKoT)6(x$&AFrY
zJGHUFKk2jAIk!T$R2pA(eZKYilZv$l$wRxU<~FOxks2Gzey_84Hau)@eYE&)P0bf!
zId9}mKexL$+;+Wh=#Mz<8<AA&WekXwf4yPmla;+I3s1^ZJl0lK_S9OvJUfI<*X1RL
z41ARQ`n$V9qd(OtQ)Ee6Puy5>I=px4+|Xk8qkXu%rt5ocoV4QDf}55hd(W@PySm1r
z-B<mC?NgOa3)y&b%V7AyH2v9SqU%mKFL|F7ZaQ>Wmf0XJxK$}HUM4Fj%bL|Nbl~=p
z>huY^QC7*mYaa5Uj^f2@?iF0#bYX8}USz44t^1y1CnjH?x~i)zH8H#Yd*|S^l^?T{
zZ}i<f`z-KK@}>2SmsukBgn#Y6vfjI)n;o2&Q#NJgf1ccGH`Lwhj(xE_{Ed*nr$1la
zHvJ#7)}<a1AJ5BJr|CQM6~85{>D-iq1Eu^&-Iq$q-+Ople>02_4PJ9!TkC3nd03s<
z*txGb|7~AFk#|+ze$C^Y4*r1s{$BjKK4jP3)ZNOLWmj9entr{0(X+GuySECvKiHR-
zW>Z(}bABA%Zs(#5yVp}+|F!mI;#x(AU_i?~{NWcp-HToR`K`Fq_rMv<P_$xk3POXW
zyW5w0GlwcI2@5VpxmdZl&;+w4a50AA6wd8G@n73TpKj(6Hav{lhD*2M0e=n{3&9Hg
zxX*f?_WXO-Y7XIQhfxWW65gW;uqDIk%EI7!4<2ESG?e46<#^Da1BkfKbt~JYvwvVi
zSR#=?cAz6ofIT@^u<8?U*;XFGLhK^zA|XwHJvV#b?z+n6L3@IY*giZT2~B{VmLAT!
zu%qcK4k1D2#f#!SXael_T}lRSec3%^Lrg@cRcw`zCcrkdJ!#P?kH$I@6VU0IteHHT
z0DH~GwjHr`Go3kv6tUx|I9Hkg`?G|CQ}beSf-MOvq<p^3e0!Pzd-n%Q#l<^vDTm;o
z(@DG}iYCBrmN@4WT-DZC5);trP@7N(ngF}qcO-1u!KzUnVQY_3T8`F^Ccr-XP1gD3
zNY@!0!T~9tCz|I?6JXz1xLP{#zFcijkVv_hCw8R?uy1~<Th&ygKENT~L#GpYiBg&X
z`_1}2of+%;Ti@NS=jiPy69G1+*DW*15k3H4IGe^*v~0C;Ylo)PvNK2$5^>-(PEsDE
zhX>eV>sPV;gBs;+v@I1%UB|YEVzW-Tv9r*2wmLfm{03+u-S2P-mjNINlA6t#?F1=g
z)3Ah~3i=PTUN^P*;}%rDQ@#Kk0<<Ef@1P1wU1uC?A68#ykCS)JZ4MaDww;V@6WR(z
zAPtDNVADv3ZIV(2slotqB(;!SC;)dzDpQp?3s!)6WQPgxpui`7>1_X&0wc1b#J?mE
zr2L~ER9ZSOP?@&fN@TFDc3+Gw;%4TfZTGqEaS+cC&h(3!k+LFWv!Bp!2Gq~&o}JTn
zq(&GlNJ^nl_=7Y_c`7{3R$dA(vz6aWzZg&e;{-NcfM7F81-b_MLqtd_jh7|?H=#ID
zsv<d&0wz_aI)zlFw@t}Slgr$YVWFZ>PuM0Yy+AL6#F12yTahn#PNMAb%W(;$Q+NOJ
zoTH7QYmv(tl8hj@9x_3&>7r*W&QN7Ufl%)GCeI!9zbzaWMX(iR$n-3LRq!Gpk0Lk`
zjv&RR-vtmd#!M(5s(Py?NlNT5p2CD$94?;2tYjS7BxdC!^-(~GNy<m*qc&S<e8$p1
zQW`tWSSU#<k(J1Wawe%HRuUK99g<2VQysuhlG==`HsjMFW+YXCt16ha@o{+H@@4wh
z<3P4$xc%HSjd;+sZ3anr`VbgQ?;aD{YFtu{dqDA$R5k8djr+lRYxhsjH*^*s7#HlC
zk*xHOxx!uq0qtb%6vkxj9Bm4-G6)Q2fH0^Gvj85oT~SRvn@YV=T8uc~h>IC7t8soc
zE(Qa(|6FX-xaNmEgp^cq>NL<@KYDQE4JV@rZ7Tv4V>xOlHCA2W1V@ZBcem!YXx8~Z
zcKqMlAQMP1h7&WEG&bD_04*?DvFZ1p38gviP+$bwvAQCu{J-S~fb%3(%qezcYNOb{
zcq+3p;u0gSfOxbPc1}!K(GZC&@(l2t#=y@jz-w%^KVWLYrTVNoBs5<*-ve?@qKBBj
zjfv&)Cu@#7RSxe%(J_BxA*QHuOkE33pVG!<huu^bpb)FXsz9jCQ{VP3*gx%s318+)
zpuHgzqB+qbrp`>A4kC8mqf{AuEN0KR^jv68_|P_zRgTX7H5jgi$k)z`7uJruyMM4^
zeq)a(jx2RoxQ9ZFU_`Hpsh=`=QgW42sMQ-A`vsQyR3B1lUM^6e$fPJvl$deH)Ox|4
z`|4|&kyZgd&YTB=J>H-H$n`;I&HE!sEmtoe>z0Fl6MwbhIvJ5rIc{B!J3zusNweOD
zJ3s+E^lslDAZlqq&grDOU`W>R&h|6DhorJ`cfa60dRq2Le4xo4v(EE-$v35rug-DD
zZ37@t=1rdqka@i>cTqU{l7wbD_Sq2LYBRupwk`#SgU|{v2niq&16>CF5<`LMv((8%
w-u)=m_IB>UYH!?HC~^`wyYMGDV$Q(Cuu06+g_(+&DW91-GE*Tl6)@BP0TY#<XaE2J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce.hpp.17BAD15A3D93B7FE.idx b/.cache/clangd/index/device_reduce.hpp.17BAD15A3D93B7FE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b56af31419da9894c1360ec610adee2adcabb6b
GIT binary patch
literal 1132
zcmWIYbaP8#VPJ4h@vO*AElC7283cj2xTGkvijjfA4Jc4CXK&zXzrzLsXO=%<*xr;`
z8NDJX(DDnzT7mBwuXQ@7{!qNF8M5Mk-)xabQxy0VRJ^a9TT`BX?Bdyjw<qk{>c+|R
zc;=j@X7%_ywvPvIuSiwvDP3`S+N|((&T|9xf@^$Z{%~wsvGBA1i-kh{Gs1Y5$6UGM
zYP~}7nG~<Zx)k5|wny*0L^SSlZ?BFGmtA;&TkPefL8*HTyqwwi<QE)lSLb~=_r<BR
z+k(s=zgVU|&GF83j|R3xo<|1l+qjfD;@<7#pHcTXR`aiH{F`?R4!p3K^{C4@=URDJ
ztMNr!5evh}t+M+9VytpD%#2ydbe8>zZSLG+!HrEp?U%GBui#glvg~VwSN(#+Z}<J2
zIo*R(ai@?+5w~W<KBcI=3p^d1c5b^Pw5%meY?`WC+=L^O3QnHVay`UV-&ef$f5L@*
zX_H=dnQST#-^9}WfBV@hilyH)-*%kzJa%K^=lwYsWqI9?Xgu{u^|k(AQ1Y7pql^Ue
z>-at58dD<O-%ozu?@(Nsn=~I76S998`rcmCh!S98<pgqpfJcr+PMm=Q!Z9rsGL}vV
zTP4cG&dI<dEi0`CCScOeDHm=UTr0|FV`AoD;E`mP6b2IvoJ_#TWnkoCoSlAtNz{t0
z83IgfFynPJb*y3fpI!KDyRm+;6E_nJP`@Rcr6`y{=w~QOO)K6EGz{iWMozH%U>3vN
z$IJzG7fhIkN5M$J8m5PbN6|<T&V#98;sU$XQ>CnX;x6Yp9v({#OB0xbVz*o7#Hgt#
zGxG4r%E)TKOiJxzRc#O|b^;1Z%1Ua&jH)>N@JSWR#j~tDJaWo%aMe6K^2+i^JOyP1
zB%Y$OA`(wYSqaYT_F`^%+yCDJXq>dDv<%FTK*KfV3}8+H8m=ji#8c2zK;kKCD#CfN
zH~<D6DC}X5UTbU>aCC0Mbf8~!WOUSFs-7CKeqrYq?`8)2Tu@dTVV#bkjx<aWFtB7T
z)nS1JqZzrtu?GuLVBF;-Dgo00%tnS|^9`A_zt8>w3r4t1aZz%yH7LdLfe0{Q;yL#u
b&uUT7w$m)Ud|ccNTrdd+23`b{8^Hts3#N?f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_common.hpp.CAE83A095739383D.idx b/.cache/clangd/index/device_reduce_common.hpp.CAE83A095739383D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc44ece3032709ad4f5a9e90f189e6436e4aa682
GIT binary patch
literal 2026
zcmZ8g3p7+|9RG$H%*>^^#$$+~iODm`C|%RKiBV&RMNPCvd5k78>oD@z%J!gGWs@Yu
zZY$CzC)PS@v=y}!C+pE}8yh7jYaOrjklO5>>3pu!J?H+<_xpdp|Kt1pzb|;BNHnMm
zz<SxnG(}8uBozRF62CGzDK=UQfEi*ht)%{B(^uxMti$`pLqZQqeR-ijuB~QyZ+0o<
z89WW$e5zKS>U{lm3%92`JN3$=aPxD=1rmqE$vpXuxAGS4^nsf<JId=Qv3Oxj^Pbw`
zomEeCTbKN5oZ?<FQS$IN{lB!@Yw#<^nPbOaPqYs6LLLaE4d==-e^-gSGQKhBqpXo!
z)ye6YaN{TRJRdQ%d0_ay{D8~Lz|)^u)$VW9lQzq^qVqx8Ge-^{d0AY~t-MxVOR?B8
z({NzfZvDHmheaK=_T_Q6`jhQ;drHFJPwq5zLH>L8$XZxB9pf5EU*m-;Je^&l=7NUM
z*Q_<=!@aBaB?|(tpJCp|3TE=l1Cxi-Yud*+EVK(PekGp{dr|7VqrNkvw{hgLMc9{j
zetA>zCW77KVZYTazshNSO;h^#u1T%>y08lo)wVLTd-Vs7x=K}yP_s<Kbyvd1ZX6nJ
zX4KKK6(g5=X|;k|Tb?*`E>sJGC<Trw47smmSi4+xNs{_p(V8B$xg)Plm67!~THkdw
z()nbOwb9I$9jjHTS&22ZjJad|t2!;)Q{7i?|Jl3EF}Z_!t7_DBx?@RhlNT7s^Qy4&
zPYuol-0k?zAB^n5+t8`;xxAARC-5_VeRuT*Itr%BWE-=>)^|gliQc9eo-`uSbTG1k
z^7!nB*<=6WSyQp%#pZ3Nhjonf^P=pJr{r9057}@?Qgy!jAfMM6P+8`GH+;6Go0e9v
zw@?0u#|iP}{n3|0T9p$H{Z6ClapnbsV}rD-OH0lB*FN|0KN4O%H27!3>c;94j-2`7
z-RsG!X)G-H1yctb@BqqmMN|uM=bU7wvwN~t_Ykm{1yGEJ*=Y=*>u7WI0R^Ra^p?;R
z<EB;!FeM0lK3||QfX>tdTmyoDC#=;z-O`=&k-+A&1*qBpIzl(o5d6;<W)zlcw;{ls
z$bi#ubBzIXnikNMNK-9uxWFl!Lx2TAhzIdz8UyH53b0i7&<JvZd&=H9eI$sD#miB(
z0d$5guv$bAl45o!R};@jE79QzEt{IFg`nZw@MZIHsEjL{NBS$W&+IC6?V}(lj^iTO
z9_w)Q#hbIeT*B>R=i>$44o@pmrtbI|5bi)~AcK%h^FiIT7#h7c+N9*>tJY^AsE8?A
z4r8Hq6oy7KNP>zHZnaX-?blP{d3LL>lBgtP3XNd}n2cbnwS(+U0fx)|$rsXe5tNOy
zMKC9VVgd|<vBU}axD(8PpkkVsNs1!I#5jrkG5-ZfLYFLpd$CKKjjZ(zD)op`47Leb
zip^o$kjRv3$|4bm%2|N;4893e13~#5z72^Ss17U=IZ_?xQDS9qoX&L205MKBjZG($
z^J#oK41u?Y&_{iBpgw80pALOZ?Yd7R%He!p$VSjOS{wt$AZWa8yaS08wu%KPfu3Lh
z_j1MBnPcD0c`%7q{MZ2tS`qOjWOB?C^Cra?@`Uyz!kux!f>9u1HWZi`%sTM~5=mGR
zQxe57;|yUNM8l5pD@l~FD8U4xqvfX_C+r>Rqfk+anZyAm$h&dpTgB%oHH3&f7SD|A
z+@0la20Q=WdqBty^4BL~Q#dK+Fy-a6M%ovOG7)hlu>ed8&vI?ld5^@xt4RG+f_<d$
zP$~nx&X&+2$tOu*BgBBzuNDsIw$Hc>dDWDqLq#p^EbU=mRMgth+KEKMl|naImx@Yl
zrF@u=iiTt1o+OfCvU${=^i$BgXA0#ftxS#V)=G`;bR*nyrrZL?f~1)ENcFQ;CPgd#
m)C)tk`S`Kn0std&X+kd1l5Nz?BIRf~&BWM{Z3K+Tt^WZaaOsx-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance.hpp.15BDAE27A6CA281A.idx b/.cache/clangd/index/device_reduce_instance.hpp.15BDAE27A6CA281A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7b8f7d676c221b3ff7c57872f6661000ef8dfa0
GIT binary patch
literal 1438
zcmYM!eJs>*90%~PhvUxEc{p_E;qu(!<RQ-Ec|7irNgj?!gp!VSXOWRc={aqA$U`GV
z-P~k)GG^u)i6UE9&XF^vY}ITit+MXM&*!~;_j!G{&%XbB_xZ$e!o%xu2!&>Gih0?E
z8R`fj1H};(@^hss2z_x!sJK}oT{uzcO2WyRQ{uh@BZuwpQ}T{S+fJP>`y&+a8I+YT
zKCx4En;X5e=GT9{^)3jkjpKb2op>zVf0`kA*}s2AeUU27P8ja0mNpMalA0>k$NB~C
z%Xd>B{^cG@NPFI*VaAX4{pr;(G~AjtT1lu)s6QM3>iuG2vaG37wE6qlzs2_Nf7EK&
zEoGvvqWTA&A8zG;yfH3)=UM&KBi(;mc41^J{=)>0l6Ohy7n0YM+;6!u9+fE+$>}l?
zU8X-v+~f|~B&HDBj?{H&mYTAiN=a<zl5s7kgJ}ykiCj*bd~L!AH<(x@d747#+i5z*
z!I}#2;wx0#kI;VF;|`+|eLQVrtg(GulljPtMkMjqi|$cp+ax)c?sP6mo(#B(%cf@*
zs#5bN=X7%i%-)y07)+X8$mwLPE%7ck2|upL?px%KmJoS89ronpy6)SP(RTEM)Yt@*
z)yq%LvF2`SxeUpBT0e{1p3^tK^_*u_r&KMse-amXL~{QIotq~ZcgNQbSEp2vm&#q%
zn7!xQxta2#(XTGdx7K%W{2pG@Qz@@W7K`V}_^O7AtIRj0F|72%Usqori??2GRGA-&
zxt4_lXLy;4k4F6L9KlvSnW3x#{!u}tngV9Z+g2zT5HbPU0d01us*)XC;UilK2#eGr
z;I(yhZSD4XxO;jr^z@Ak4TvOT6MKg}G;iMke?K2)V2~-<+>Elz(J44IoD~*gwVP^T
z85PNiV5_QYYT(qYZJb@)T=&wE>dyPf3g`md4b%f_0eyi4pf->U^a45p=|CT#Vm`3Y
zFd)PPvVqn>7SI+L1`Gj)11*8hKnI{Va39bV=mGQu1^_ugb07m43e*59b_f<@3<wzl
z_W+d(W;;X-2t@#?Kuw?<5Dzp0>H?#Hk-%L*3!rl6ZATviLdp@?jy(p1f`B?eCtxtp
P6-WZ^1qK5BfquaMU4$SQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise.hpp.C5FE82438250FD9A.idx b/.cache/clangd/index/device_reduce_instance_blockwise.hpp.C5FE82438250FD9A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..11638553188b369d4898e5965b76413e61b6a3c0
GIT binary patch
literal 2296
zcmWIYbaVT_!NB00;#rZKT9U}Zz`!5~#Kk2=nJJ764BvnP6?3jmKACmcK%nifhTW7&
z4l6wRl~}l5Do$5vxp&rOmc^l47am=ivafn?>PpTp0a3SRoBTiXM}AhI(XQ&*LGl($
zRW~Kv$!T1gu)|eOGOua2?pG$JJt1Q1FNKq*pSEf{oyxsy#*)=5b{%RBC|zT_PS=X#
z&<oF9SI!HGzA6ye*SGyI+r(aDu_BHgOn0MhNKBjc_;t;^d!NO0jm?B_v0vc%``beD
zZ~U)9ra9?*A4DDAJmc+>MAyrfiCcUk95Wr>_O7=6;xnae>6Ce08ouA0lvG<}dYp<r
z95~oLDWceAdLvt8<wH*6Q&Rlaoso@xi&dPAB`t-*CM@JVnX58aSn!nBC&>s`Nf++l
z1slG#|BXJccH?yEJf`f{bNiF^j-G2=y-v|@q2ifJyC*j17fyP+@aEB1*E?sGziM;Z
zcWb5niEC@Z!ggyej@}iQ?)KhQ?swhg)6aMQQZNse&tEpp?ZXr1ojV`xD6Y&+N(P2`
zPV2`tt5kj|b2D+XGH?p82^fG022P+H0|O%`W4Gs=aQh$PsoYFFKn3z_@&;f6p<t6i
zW7uW|{azjhW}pIJR$omp!NAGH$iT}2G=ZTgHLZ9H8v}z-vAS_np`bAn4`;Mlv}Na#
z?4R?lub#-mnP8XTc-pS`+!_&|1j$zm@~3T*KF-0zDa;})19Kk_rwEG(5>J#x6p1Is
zB8J2hXAwu@Nw7#D@g!L!k$6%pQgEJ3V95LrpFYmS>pIzV!8zB@X9(g|q*Nv!ySs6n
zH(o{R$2N%72*k?cRirW9@Y#j!JTver(pvP#olWhZ1YSiJQ)@~&CLBt{tH?3uwUFbV
z^_q-4oO0}P65u2NBpln{F#P;da*K(DGe#iB3KnlHoQV>NPOx}$HB-3$Xy--zuJH9Y
z7eC=TBM@joj#rK^OmE<ouS-}XJ}toOxDY=f^%oOuSb=)|IQ{rx$u2aVvv0?AqY%7$
zV_x)axY@NT39q7zy{>GF1f9I_DuSg)W-f3(={@v&!J3EG3Ot;VN|8P=H}Rl^Q?5yF
zM(FH&*?7&Z`tqeu@F2%oyo%bZs!W@cmUiM*1k12M7lLvt!i8~KaUlp7BF9}%qvfpC
zh2OG(=0^HQA}NXskBfn6nRe1+`Q^4c9bouZ$yRB=gy%#=-OT4W_l=c@Q;9`M24*`C
zr!tE&5>JIi1&OE1qKd>*V^M?i7XIC}rhezi9K2p!;(S}y>Cms)Xiix&+0^lnGVd#p
zH^d_qVJ_t1jFXSkgz;9tZmx`cIHwP&NWenC78aj8obnd(wlLSj6f<&x%Q9FPZm~bd
zsJ>D~5wGKRRr6=^>77~(%q@{Dk$f<tSUBT&;>2LQ-8PDXPddK!f}Dz^2<TKKMSF!5
z8(b$yU&m|R{@9jUixU>?coiL95s=60#?gWnV8`YgGHHLG{R1eRB9Wp8%Z?|nL`y_i
z9%saB<msx$trMHJD&tjj@mguPw!*P`kRr)wby&#(Eb$c+^kLTBxtYFVNdxmDye2)n
z@Y!}_{bDDeQ~d<|<YD3XmjAG6qJ{Q(yn27^Th{tw`46xwaMnA&e=pmwe`nekpanVl
zIUX>_6{Y4Rg37_-qU2(05Q7~=KmZqv!N35l-CzLNuz><5&a|$^Q|#wTyja-TIJg)9
D{{wIW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.cpp.A1938EEEABE8B42A.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.cpp.A1938EEEABE8B42A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..766a32987ac7db0b5baa4257092b6167dc5cede2
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TL2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$4!mC|1&kXwnw~dk(Po1q*-z
zz*UuTQiJrWo^Gn=w^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFI6lb
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(|u@tpy
zP}D(c8E7Y3w?|(}E6pKL5Xv>J3vz9++@iN=#5$4Df~+!0;|2FGziDXgr9YfBJYE;@
zB#q%OpKj1b-l5f0fst!8A&cHSpVuAwL#)YevhW-|7w@TZzLEJ7S&9(QRJ^wi)+~{Y
zz~|5ga<+%nt5<K|Kz6LVbcR5STG~kJNIA<Uzq#yKL=9n2QPrTalF1y^Xly+y8JX6o
zjPx?{y>KL5;v_)pUE|_cb@<FaMrx36gy@rs&WC~G=s22&&SeO(IhxG99KUmNV-ra0
z?4fZfK(F`27_tbB^jwb<_%C_~JLySkQp_qArQf7nj-x{jg}{l2iqL4txh83dFP`xn
zVs!D);t{tT8K)#W1mD1}&1nGISjP$Nu~uq*+J^v@Br^aW1Nt3nBF#<_yWs~SCT}@T
z4GouvUCR|6k9#)M&U5`?as$yp9@P$IyWi#QM@3C`Yu(<$UV59O>)jr>AS;A*sw(u5
zg52Ohq`;2mc9vH686UFw;zY!zHQX^hxJS8#UCh&xlXQ$L=ZvL*|0J}ETgXwT<ro4t
zt?&5GdEP`%8!y{Bt4NlE>DE)dbSHb1s?L^JMV`W^-lGDz@vGHx)_v#v&80fNnXDzx
z8mx^|Tc!r@VJ+2`yOW(W)KHvW72pbJW6f4rC0EWR&#Wdt6)E`q2uyAS&xqZK@y+OJ
z#7`>MZ<_1&djZeIW$)SWwgj(4o_Xg>*J3}1CD{-qCs1tKB)H|0!k;$tvj&959Vsk(
z>FbHI%ix{@@5Ym|x_;fQ-2EffwuRs(6W*x?ej8nnuEd(GsHs(|nBB`QE_f!we`p?Y
ziZ3!caz({@{}AHBUx83tks=w2hd%f&Cuma>Ml}a@q4u-iMYb|PzoY=a!ngmo8@Tet
z7eG55^>cY`VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002+`000000000008an_
z0000000003QB^^EyPXteW&%%80Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOpvPCf&c&j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.hpp.20070C9325736FAE.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_add.hpp.20070C9325736FAE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fd2943c97139c1798230e5e30a1a6ceec3047c3e
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er}w?n4GVE#JF&w0=C8
z&zUJ58*zQgvb!(#i<^nuU+({~@A13SA10hAXOJjOljvodsWYMUY10&qGGXTR#bL_g
zlInV$${wFKva;Dbt`=Ha#u37qeZEGjSg!7Rw1r8$t8h`)X4%aCM_g0??ftY(OZ{@w
zteqFk)J|%tp8s$rH$8!4Pe=NltLlw<tLBGobe{izHt*e^ZT+uPcZ(c4_KQg=@<>*3
zWo{DCxeP_AX~hsaClN#!7bO?#18F7})s@ALLOfcc%#2LT3``(72mpEz2AE+C1_l5<
CyIq6;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp.3CFF0C60C7399DEC.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp.3CFF0C60C7399DEC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c11e67eac4607d3a25eeb9362ccc75f9f1d676d7
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TO2mk<hoV8V1Z`(K!
z*8eHSjGL@`0DeeZduwcy7(I|jfhI@Rz!a(F;iJF4L&>fixBF0|`yz6_kDPm`0-ykJ
zRb`yiAib)`t7`Y=N*3^L@=rN^xV`%@E$?Q-(HZdinLJyZ9hfbyU!DP@n=|0KiuuFm
z3_QG>-hKI-H(x$Y%f-jXS$VsdkB1DD<@@=fTzt6wJbV0jU*6s`@+;!~WYQCJQM(33
z9i*0lHllS~^rf`Y>;na%T+_NBmj=r<dW%LZ6B#YYDw8x`aR2g~hL&FX{Yk^)bpcP(
z82<9<25saWT1^!gxkMAP=)LoK-Jsvcn(QhIyXd)iPnGkH%$LYggn*{vy>+l;iEIQu
zhc=M2J*-~5c=H;vW8I}Q1X|S6Mp6gLSvL8@Wyd0F2wRG(291?W=AcGn>ru(bv_@s5
zmyz#<1L+bc0b1`G7r(2+9s3xmLB0{9PbxYejuc17(KK`}Lx|16WbWnolani(Kw4+_
zjY|P~y(h+yMQEhwdYr(o=pAgNC#gv>t5}qNm2x=_4mA`4Cmt$7qao*-q#?d|#&d|#
z#eIth+;U``lI#%t0J}D)0cd3%C$z^}sr6|e0#uUB5%3t$Z&(v)c8b^ye<Wh^mgCgW
zaCz9ZT+!jMXG85g*B?eV5FO-E?NGM+Q{H}6)MVGzZ7po2w>i4r?IRatg|JLjg&tCn
zD?Ab@u%o%1rPY1LkJ)^7B4X1T?wB6jqFlo!=4r`EI>eQ8#!|q46I#VB<fzkf2!Wf{
zcYNnOZ=$D_mu;O@B+J2c>#1J4lRZjRXG^RiPvKMVQ32fe)nY#BzEl3@QXSq*){<up
zmd2?yQ%CP%Db<?0lbtivP#j+s;0kDC%~n_?SI#BRtR_DeDfs*djBW(ah+T={&ERUl
zPb$}Moa^?gfL(Ffdv<(Vf>$EXyz`}Nv7f_|Y>1K*C{}F}+;U0bZ=3mD148456qdd8
z{Y2Sia7%%A!%<mXzwB1o$lXA-Z6UbJgm<cezXsQXE3qUiYHF1#CiilUGaicYAF2nO
z<IAiLTv4&&w?bU_GZ1PmQY1rh--q921#N1=pysG9)PDBE$X1TfA1T3~@a=!@2d;hb
z4GAqDZ*zHVVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015vlS
z?QPJo!b1X2PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOrze`+5i9m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp.8C9F3AF79C3259FF.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp.8C9F3AF79C3259FF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..769ad2f0b5d0eefa84e96a1ff32a917c042e9eab
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%#}czfr){kV$RlnZ@xnYJTBErIu1z(
zQ+3n}S6=-iqOQBsHb`6G_n}j>E@nKdlsNiXmSLHQgqo_m7Wb1aMUGuhc6T-TDypkK
zmQ2^1XzQ_MpCy~M^Xf-S%L+m`v(MMeDw3<a9(}_!TGjQ*jV8Iw?@3%!|Ly&>P0QVv
zHTap+?HMj>MRX4<Z=c(E@T1bS!tnVlQuE*J>5}{P=YCGcIdhry)87Uyc+U2l=|#Yi
z7YxOfxk(@g7p0~ZL+G4D5M5l9T&xdbFoOsP0D1}rfPRMpCYDtOuJg^V$)9ItWMXCj
E0OUMgz5oCK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp.39D9A35BA78EDFCA.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp.39D9A35BA78EDFCA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f6dee36658b1178086720fd70e5c7737f6aea5b3
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TL2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk(Po1q*-z
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(j3;Nnn-}t8b#`E}x_*5IOm5DAmnxPI
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)p9y!pe)}nSLN!%?We`#$NTd3o{=9B?`N}tSc=BA
zD4HO(473%U+o3O|m1Z9(2<6(&1-Uj@ZqQpaVx7onK~|Zh@q+u8-!!!L((g|i9<K{{
zlE(0tPq%0z@6c+hz{oY4kVWsE&+8WbJ~m|QEIdcg#e1rpZ)LtjmLdc+74NNsHA`e8
z@Hw=BoE>2G>ebsfkR9tUogvVnmNt?)P|mW+Z!S9)QA5~KR1IjXWHJXe8at0lMy53?
zBfX4#FC0jhI0?{s*Sh#s9X_*<ky_*%A^N1E^WjKwbR11f=Q4!Y9!%z5j^8<1+XT`&
zyKh|zFz7uohAcuWJ=fy|{)^thR(g_}6tjv&>Fbosad4=m5IFHr5gH9S*CY+`#WS8m
zj4tjwJm8ii<CJ8F;2YR=ISoLyb)3)v>!jAFeF#uVGDpB;K)+>8q}eHAH~f)^$y<(7
zL(Ao1*K$S2<ADv0^IU(J+(2}YN3}=U?ss|nQBjlKShusVliud&2Dgt~kQKr@RTX+j
zLDqOAQea1OJ4>sFj32Z4;zY!@Gu$x)xI?*tZOqe>lXQqH=ZvL*|0J}ETgXwb<q!fl
zt?&8XdEP`%wU=F!RV2&7^y{f!x|cml)nrSoB2VGd;86kG_|<AT>%VjU=29KrOxBWT
z4c5k~4O2(&VJ+2$yOW(W)KZ*Y72pc!V#8KgC0EWR&#Wdt6)E`q2uyAS&xozX_-1r9
z;wP2sH_dhXy@2Q9viI!xwgj(4o_Xg>*WoaSCD{-qCs5Q~65Mi0;ZK|SSp!1jmK2tQ
z^z}s9XK+V>cjHM}UBB*ExvfX4>k7e5CcINE{5HBCU5PbW(NL#UF}s%=T<}bU|Ij?*
z6klX?<cf;*{vpJLzXG8)B1JM3_e1b~PSB<{j2aH=LhWb2i)`fx{gMLw3g7<UZs5un
zUjUDq9y9=Rd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z%APzG<JX5#0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8YMKGsnH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp.D3804A89BDDBD68A.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp.D3804A89BDDBD68A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87dc5e1005df22ba7c4e9cf1d7dc82ab3ca4e29f
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er~eT!#&KoPWC--%e=%
z+fyjC-PHH7fV%GA-;Sq`y}Rq@=HzicR)6Ar27^~;45V0%A{AbrV$}@1<HNJ(Tx4%d
z%A6f4eF`U!i%I-pS?e`rm*WDji0x;l8$N$mzPHu+bEt{pa<khrziF%sxVbO<Uf9!=
zr~qT}TgO!*mp+kfyTixUcA&fBm~*Yri$%XOMXkU7+<)^zto)7p`M;N@-iVA>dU>S2
zxH2~h=v;=P)U;v<os$Tni;I$r^+60K5CH~EEN5mL-B(%0)x^xm#LNKnAPg|W7z_*m
DZ%Aa{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.cpp.26EBD8A22938C32C.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.cpp.26EBD8A22938C32C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94149a0e5e5ae43223fbedbd77c4177466c81bc1
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TL2mk<hoV8V1Z`(K!
z*8eHSjGL@`0DeeZduwcy7(I|jfhI@Rz!a(F;iJF4L&>fixBF0|`yz6_kDPm`0-ykJ
zRb`yiAib)`t7`Y=N*3^L@=rN^xV`%@E$?Q-(HZdinLJyZ9hfbyU!DP@n=|0KiuuFm
z3_QG>-hKI-H(x$Y%f-jXS$VsdkB1DD<@@=fTzt6wJbV0jU*6s`@+;!~WYQCJQM(33
z9i*0lHllS~^rf`Y>;na%T+_NBmj=r<dW%LZ6B#YYDw8x`aR2g~hL&FX{Yk^)bpcP(
z82<9<25saWT1^!gxkMAP=)LoK-Jsvcn(QhIyXd)iPnGkH%$LYggn*{vy>+l;iEIQu
zhc=M2J*-~5c=H;vW8I}Q1X|S6Mp6gLSvL8@Wyd0F2wRG(291?W=AcGn>ru(bv_@s5
zmyz#<1L+bc0b1`G7r(2+9s3xmLB0{9PbxYejuc17(KK`}Lx|16WbWnolani(Kw4+_
zjY|P~y(h+yMQEhwdYr(o=pAgNC#gv>t5}qNm2x=_4mA`4Cmt$7qao*-q#?d|#&d|#
z#eIth+;U``lI#%t0J}D)0cd3%C$z^}sr6|e0#uUB5%3t$Z&(v)c8b^ye<Wh^mgCgW
zaCz9ZT+!jMXG85g*B?eV5FO-E?NGM+Q{H}6)MVGzZ7po2w>i4r?IRatg|JLjg&tCn
zD?Ab@u%o%1rPY1LkJ)^7B4X1T?wB6jqFlo!=4r`EI>eQ8#!|q46I#VB<fzkf2!Wf{
zcYNnOZ=$D_mu;O@B+J2c>#1J4lRZjRXG^RiPvKMVQ32fe)nY#BzEl3@QXSq*){<up
zmd2?yQ%CP%Db<?0lbtivP#j+s;0kDC%~n_?SI#BRtR_DeDfs*djBW(ah+T={&ERUl
zPb$}Moa^?gfL(Ffdv<(Vf>$EXyz`}Nv7f_|Y>1K*C{}F}+;U0bZ=3mD148456qdd8
z{Y2Sia7%%A!%<mXzwFjV?gpxD3&Bk$yi*PQHMkyJi6vQ4Q>#=lxtD94@l1sO&^+K2
zUu1OPii-8V72?8QflzCaA{mPNKKL#tXj2mgH3xN}_OstbwsM62NCEzYZ~t#MaOI0{
zt;-%v0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015pMe
zQ!o~BrZ@smPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOjC8y?*IS*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.hpp.EFFF67751118917D.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_max.hpp.EFFF67751118917D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..136e14fecd92a4422de45653cfd8de3173eb77ea
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&enc!zC#8)F4amp4oL@7
zb<_)2Ui~AYuDjDVNL%3dp;NOiW<0BuIQm(ZVVQ`8nyR}N_meF}j$Kc7cQyGcs;fSh
zOxK%e>#=2@C7ZSL>PJh<3PL!u&)3W<lB>HOeZw?b)%D4ZCb`V-NnBI^?ftY(%iWhX
z_?gq~87^x@bPp?UpWArwqtdj(@cArK^WW_0lKb`Neon?YbD8zi-?CNZt<hJOInrKS
znVST3E<;gjS}}yqNd(cwMajkbAO;hNfB>MEpn!=bX7YiVv*gWOnHiaw8JJ<>3=9B_
COI;8E

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.cpp.82123925B2E47DE5.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.cpp.82123925B2E47DE5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8cb9a23cf80fc363ee8c112897db19af8ed2016a
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TL2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$4!mC|1&kXwnw~dk(Po1q*-z
zz*UuTQiJrWo^Gn=w^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFI6lb
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(|u@tpy
zP}D(c8E7Y3w?|(}E6pKL5Xv>J3vz9++@iN=#5$4Df~+!0;|2FGziDXgr9YfBJYE;@
zB#q%OpKj1b-l5f0fst!8A&cHSpVuAwL#)YevhW-|7w@TZzLEJ7S&9(QRJ^wi)+~{Y
zz~|5ga<+%nt5<K|Kz6LVbcR5STG~kJNIA<Uzq#yKL=9n2QPrTalF1y^Xly+y8JX6o
zjPx?{y>KL5;v_)pUE|_cb@<FaMrx36gy@rs&WC~G=s22&&SeO(IhxG99KUmNV-ra0
z?4fZfK(F`27_tbB^jwb<_%C_~JLySkQp_qArQf7nj-x{jg}{l2iqL4txh83dFP`xn
zVs!D);t{tT8K)#W1mD1}&1nGISjP$Nu~uq*+J^v@Br^aW1Nt3nBF#<_yWs~SCT}@T
z4GouvUCR|6k9#)M&U5`?as$yp9@P$IyWi#QM@3C`Yu(<$UV59O>)jr>AS;A*sw(u5
zg52Ohq`;2mc9vH686UFw;zY!zHQX^hxJS8#UCh&xlXQ$L=ZvL*|0J}ETgXwT<ro4t
zt?&5GdEP`%8!y{Bt4NlE>DE)dbSHb1s?L^JMV`W^-lGDz@vGHx)_v#v&80fNnXDzx
z8mx^|Tc!r@VJ+2`yOW(W)KHvW72pbJW6f4rC0EWR&#Wdt6)E`q2uyAS&xqZK@y+OJ
z#7`>MZ<_1&djZeIW$)SWwgj(4o_Xg>*J3}1CD{-qCs1tKB)H|0!k;$tvj&959Vsk(
z>FbHI%ix{@@5Ym|x_;fQ4cU=u+d^=Y3GY+`zm2X(S7J?8)YK|f%<knD7d#W;KQxax
z#TOYJxuRmde+Y5muRy4+NRbT1LmzyX6SS!bqnd-dQ2W{MB3l`tUs8Zy;oJY)4P5!+
z3#M5fMss;>VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015uu?
zPPIEx7MlW3PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOorvyp#T5?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.hpp.1AFBE8401A0BB727.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_min.hpp.1AFBE8401A0BB727.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d23c3412ef4b5779bfc472861251b984fe49f9d0
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er}w?n4GVE#JF&w0=C8
z&zUJ58!>y@vb#6-i<^nuU+({~@A13SA10hAXOJjPlIUfcc_!e~AyH4wb!u#O%O~~5
zq|Di&<u34bBRiYDW4_SRwE`;yw{889sgwF6Hvj2`+R1B9+?p(#+5brE>euS0xu#*K
zMT;ML-JY?+H0b;Xo!ayS4x8@uJ6F{k_0sHhliKb6n@jKd%qxHWblKz=$9^#>OB`t}
zuFOpWI+vj+HLVyz=OlvY;-chYeGr2QL_h%0OHjbXa&~p6)|)5Co-#8sF*7j3#2FX>
Dz>8!(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp.4AE84C1155A0BD16.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp.4AE84C1155A0BD16.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2700fb7e29b1e4810a638d74f6205874d35c7edd
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TR2mk<hoV8V3bJIEy
z_WxlvYT6Ru%;cdB31pych6B?_k0Pz(oyd~WC29Eetn7r)z<D6Ud9k$LC+&UZ0-ykJ
zRb`yiAib)`tLo{^g)HFh<U={Vzq$Q5EpKPT(FySSnLJyZ9GESxE>D2b^$GA?#r*zr
z2JYWZZ@;|b&6j_t<>J%Bth`yw$3q6n^8I{KE<WCTo;`fJD{t-?`5p0oGU<uAs9l4i
z4pPfN8_~Kg`chhHc7cLWu4!G6OM~Say+tFIiHsIxl}Q@UxPSglLrX9H?x^A6s(?pn
z41f7_gEsOGt)>c$T%rkC^xpZrZqV;yO?H)qr|7wOkCpR{%$LYggn*{vy>+l;iEIQu
zhc=M2J*-~5c=H;vW8I|_1X|S6MpApqSvL98Wyd0F2wRG(291?WX0JwL>ru(bv_@s5
zmyz#<J?RoB0b1`G7k{Y3C-yN?gM1@IpHy@{94L;CqiN_=h7g;*$=u8F7bjOXfwa!<
z8kYj}dXJ1Di_l2V^*Dmx&^y>jPg0X&R<S7kD&=zQ9cm~9PCQhEMnlduNke?`gy#^W
zi@O&0xaG(=CD|eP33hEx1JKGkPH2y{QtRVB1gIpL1K=^B->@dq>=dyZ{y@a!Eyt;$
z;qtI+xuU~i&xYE0u0M>fAv(yT+M#Usm%RP1sL8Ia+gjL4Z*z3L+XpVl3SpV53O%GC
zS9l;&U`KO1ORM{gAF}!EM8u{w+%Y}4MY)De%+r#Sw2v$2jHQ5I6I#VB<fzlK4}qK3
zcYNnOZ=$D_mu;O@B+J2c>#<(ClRZjRXG^RiPvK+lQ32fe)nY#BzEl3@QtjVN){<up
zmd2?yQwQ&1Db<?0lbtivP#j+r;0kDC%~n_?SI#BRtR_DeDfs*djIIUGh+T={_26Q_
zPb$}Moa^?gfT!ZT_w4Yt1g}J%dFM;lVn2r^*$^csP^{V{xaE?<KQ{BH286~9DJ*;G
z$BDAb;FbdKhNH5&y6je-gywRf+O`nfX2Ltwz{|na;6g0Pike!bipiZ^<BX>w{D<xV
zC;3dX16Ne6`5z(9{2d6j7AcaUxa$+p<pphO!l34`&eVSL%g9y^&|fLS-|+4K?g*}a
z@f`-IAVdIjd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zY`sFFhqUql0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8ZC&*~@u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp.E3EB62B3883D54D2.idx b/.cache/clangd/index/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp.E3EB62B3883D54D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..724964577d9d77767c9fb509831fa3f6e2c6eb5f
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%#}czk%@t!V$Rn7gItFkcwBzFvT0X*
zNO#&Lo%_Np(rowCe|s-1s#rho5YOQ=-sRJem@`P^rb*0I+Nm?)?MYTo&9!=b2cPBm
zIVa88p`k9oyVAAIz;X4YCF=xM3SBe(oOx#F@7Vl@H=IO+Q*t-UXZjn7XubY^>h`KN
zd91p{3v+$0XlEu=JlSS!;1Z`|@b1;*1B#I|<~?e+`)@AI_gwp){@=2zr|#r$(O-M!
zs6In+Wo{DC!3;&IX~hsaClN#!7bO?#gBU=cKmij=9^0z`70F|}m>HRv8Gzo00cIG3
GfdK$@d0>11

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp.116108BDF1BABC19.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp.116108BDF1BABC19.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ee74385df8375b41bddd136039519be166dff78d
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TO2mk<hoV8WkZreBz
z)&CS@#!c337w|*k+FN6r#OMWi6lii}4NQ?*UVQ1-?@+Si#_c}T=srl|oI}oir~;q>
za8+fT)F8d8$E#}h?n)N$e)4xYeYm~*I4$pH!_gV=_JurKoNbsbu3w)4qnk6}rHc8(
zmkd0-pWc1_ix*!%Ps_!p$60y1n2(1Hl;wx{qFj8u{W5#}bYI@yGx9Uy!(`GEb5Xkn
zMIEG;fi|LbTlA&0((D5Tp<L6tAeRQqHF}FiEE5?m$SRXGUU2^MlZKXF`u$16<8=W~
z(ind8?gnk-6<SRd7`a3fvgp0@`MW{Ck2Tp<7Ix8d@t!K@8<`K0r3e8{#cS(e$r9NJ
zybo<4XM0$^diCxtWXHNgX9%>YrH!Nxl(TH|yTgt})DX55RSg;|nan|r#@3^fk!g*}
zNG~Iw3kT9AP6D*vH7-7@!yWq=sX;ywqE9M1AC449$I&!&E<=dT!DQ~`_=A%xn?PD;
z_l-*ddc7yckVR;u=X#vLZ|EItq$jCKF{@aVewA`L4h}UG0w*3SLZcz)nxr8<c*b*x
z(Zzj>2i$UGoRaJid<VNWrvYeX9VfKMTB-GE9RgI6%n|Sy&~I21X?BX(4SytJ@|NS&
z&~SO!wOrBRuxCT<Jl7vaHxM1<QSDH+`$OJ-R@7wI)@?0prMEe{-t8k7WQDLyRfQf>
zkSjbADX^otou$=%#*f*2u_I#B8t#}L+@f5=Cgy3$Njk)pbH-A@e-c{7E##=vatMK&
z)^~j8JTIcBm6vUuRV2&7bl+3GbSHb1s?L^JMV`W^-lGDz@vFsr(!Hnr&ZRosnXDzx
z8Z3=dYo?Cg!&0g>cPBe%sG&H%D!>)c#+t3LO0Jwso>@(PEK>0K5*XbGo)Nne!<)g?
zfUi`p-#FLpO#!>&vit0Kw*;?5o_Xg>*J3}1CD{-qCs3@~B)H|0!e2J?s|JL|4Jj;p
z>3K)lWpGOY%W61!eq<we1J$;L;4Ty1sRsTSTo10qlB}qyRjQcW%Qen;D8hfJ9&nB?
zvpR4^#ftwF;=-SSP-~GQ8H)Qp{4OhKQxgU?M|Gk0vmZvba)f?M34Vuf|8qZZ?Tdc_
zFf|{I0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015y4X
zwL;4C)>Z;fPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkO!$J%82|tP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp.9716CDC40BFB2514.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp.9716CDC40BFB2514.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1f83b89cfbedbb5b5e386ca3ca10e7c5f28c294a
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5Dfr){kV$Rn7jeLg;cwBxe>4>q_
zsTB!NkMez^;TQGkyT{rUH@KVIt>?a*r?2c@#c*IrXMzz=#vX@jQ@flG-st7svusV9
zvx<4JVAp|1$=ob;ZDE#0Zv{3iT5Wr|^wiH!yYD$I{CP^NZ;7h)Ww}a?h+p4N9bcug
z>}ck)rrQ$Tp(6et&g7;maBSJ2zJ0aw!?qQ5aX0;H|K*#R-+OmF@z8a)-)t|w9QjmS
znVST1Y*A`jF@(-Z1kuGs$;J902GAEU0Q4*rFtO~JbkeHy(k2IHMkZzkW|%kw0|3C2
BW2*oF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.cpp.F4D8C3EC38AC4A06.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.cpp.F4D8C3EC38AC4A06.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b582a00cfd4a05da48c6de07b7d411f142b8768
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8WkZreBz
z)&CS@#!c337w|*k+FN6r#OMWi6lii}4NQ?*UVQ1-?@+Si#_c}T=srl|oI}oir~;q>
za8+fT)F8d8$E#}h?n)N$e)4xYeYm~*I4$pH!_gV=_JurKoNbsbu3w)4qnk6}rHc8(
zmkd0-pWc1_ix*!%Ps_!p$60y1n2(1Hl;wx{qFj8u{W5#}bYI@yGx9Uy!(`GEb5Xkn
zMIEG;fi|LbTlA&0((D5Tp<L6tAeRQqHF}FiEE5?m$SRXGUU2^MlZKXF`u$16<8=W~
z(ind8?gnk-6<SRd7`a3fvgp0@`MW{Ck2Tp<7Ix8d@t!K@8<`K0r3e8{#cS(e$r9NJ
zybo<4XM0$^diCxtWXHNgX9%>YrH!Nxl(TH|yTgt})DX55RSg;|nan|r#@3^fk!g*}
zNG~Iw3kT9AP6D*vH7-7@!yWq=sX;ywqE9M1AC449$I&!&E<=dT!DQ~`_=A%xn?PD;
z_l-*ddc7yckVR;u=X#vLZ|EItq$jCKF{@aVewA`L4h}UG0w*3SLZcz)nxr8<c*b*x
z(Zzj>2i$UGoRaJid<VNWrvYeX9VfKMTB-GE9RgI6%n|Sy&~I21X?BX(4SytJ@|NS&
z&~SO!wOrBRuxCT<Jl7vaHxM1<QSDH+`$OJ-R@7wI)@?0prMEe{-t8k7WQDLyRfQf>
zkSjbADX^otou$=%#*f*2u_I#B8t#}L+@f5=Cgy3$Njk)pbH-A@e-c{7E##=vatMK&
z)^~j8JTIcBm6vUuRV2&7bl+3GbSHb1s?L^JMV`W^-lGDz@vFsr(!Hnr&ZRosnXDzx
z8Z3=dYo?Cg!&0g>cPBe%sG&H%D!>)c#+t3LO0Jwso>@(PEK>0K5*XbGo)Nne!<)g?
zfUi`p-#FLpO#!>&vit0Kw*;?5o_Xg>*J3}1CD{-qCs3@~B)H|0!e2J?s|JL|4Jj;p
z>3K)lWpGOY%W61!ej2$OsJ1NxH<|EGHSovadT=F{WJOJ_QpMz6u5rdQ5&lE-fKz;t
z(Sa)}*88Ur7yb%_T8k9PP~7*ycR4|unlPw2s0+29{VuYVBlKGe@H>3_f4hMzU;GQy
zmmX?!d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~p(Y<C
zqRmuP0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y>!p@-p

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.hpp.E4FC0C2AEABCC117.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_max.hpp.E4FC0C2AEABCC117.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd84c05a4885986d158809909c029aa7b869540c
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oe1{BpTz)F)h_ThF
z6$wv|@_nP>7xn48$J!M)xSQLp=f0b#uk2pMa9~Mif)P)~9*1jFyPOZ+=;htBY)zZ9
zig~bL*MUdL+$?o%VU|U21vV^NZF{-&)Xz`5?>Q~}c}lBqiK_Kwxk`<QU*At1U!}6_
zXy&q}+Y;TOBK{xF<fbcdY}uf`eYNt#wiR`8H~nh=<(rw`dv{#5CU;G~^0p%r8Hy`&
zlYow8C`wH$hR`{QAiB6HxmX{>U;+_fz{H|)rsWlfs_+(OMkZzkp!Z;a8OC5>000C|
BVK@K)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.cpp.8D16D62A0EBE3E6E.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.cpp.8D16D62A0EBE3E6E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ee8bac987635a682cfabf531b097d807206af0e6
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8Wka@#f#
z<^N<>uxeLLI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$@ODD^}8nXwnCg*mHorFIWH+
z0IsTxlNzK~^>kA`zq^tJyr2D9&L3{?KF-U##dvZCynQ7vR%aU)tLtCRfXU4n@LI+4
z;d2Ha-p}v8{K1Paf6dF)r^iKkyIM}i43y=E<*HnLy#2g*{B&R5-ZSz&;=^p#6H8IM
z21OmDmVtJnb$j%sw9*^`1)*Hix**pE%Po3~MywMVEyya9G+uE2@{@+vUi!mH!{c=U
zPtq8E^X>+1<P};?6&Sfj6SC;N^ZC0&e~2~NO%|S`=i)t8&Nnh2B1;hhnu^!f!I~wq
z5qKZkK+g8Cdh_PpTgZ-eht3daQA-<19Vuto<X49si>M*&DXJPYRx+8R8jY<-B_q=s
zm62XXJ{OLpOPmB~y=z>2QHRg$W26T8M2J4A=zJI`j*g>g=v;;no1@9x%kdj0H#UK^
z&K??<0`z)Mj3JBANYC{+f&ZX)u#=voCdI5`QTk2F<v2RjPzaoOs0fXQoNJPX_~04O
zAx0MuEgo^pk#S10L+};s+MEWUjdh&R9&4r6r*#NWNiqZAF`(bECerK_u^WCMV)B;b
z)X;Ew*tJ~I@wjJ0?L5~XCN~fr<WcQVw);)qzE{*_x7O_~?4`Fky58-93$j93r>a5^
zDaZ{DL<;O^Zf9wApYb7^uXaRiTEiXFgL{-)*u^|8IZ4O3a?V%^_;*69xP=^bT8<%b
z)B29@oaaUKwDGd7vx;OnnC^S3m+oYbQq|cKtH@LM)O%C_H-5ER&bs%U-?>!BJCn8K
zS%bB4YRlB%J*=hLa(A+Gh8l{~s{&jBZLHY}tK`bL<eAmvry>QPFM-L8;2E(SF}@jH
zjrdCC`b~4)ek<U)xa>X~?v~({$TRPJ>00dPup}Fz<OGUMn*_I9QuxDWe$;@_xFdyS
zFTLz2yA19rU|EePFHb{uq}sL+++@N#)xfW#>(P~1lNB|!N)@wvxy1#~MEDQQBTn%}
zMn|rwSnuybT=**xYAaGCL-EiD-{k~tYQm`Ipf1#Y_PfYd2I!|0;Ai;u|8@gczW5i-
zB_3pRd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~SpbpD
zLq}T$0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Yj%GJpL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.hpp.4404F6E1807E5BD4.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f16_f16_min.hpp.4404F6E1807E5BD4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..03b13083da87661d997426688a2ba1a80cd98c7a
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oT!#&KTt2&s##((z
zZ%NYb?a-Az?f2onxRBJY<+3Js3h&tRPG+fNIxt7+aL$vHDXnwN^gO&4&X(9AQfwr}
z{P>WoWy1>Yv&K7EqoyisZC>sbQT^=poX;P2-#fH#r<Ui$t**N-$))Io{Q7?C_^LJ2
zxR#x1{bm@rdP(2IskaXvcu*;NykdoMW3ory%_DlX|MJb|yzBqG=*78TY|5LC^fMG!
z<|Y9h%TSb>Rt%wY5<zrvQF5_9h`|IRz<`M*FJ}LWo#JwK%#2LT3_$O}05goizyJUj
C_+&Z&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.cpp.6312AF648C08A121.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.cpp.6312AF648C08A121.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a401a5d92391659418693044e8b887e3eaf488fd
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TL2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$4!mC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWo^Gn=w^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(|u@tpy
zP}D(c8E7Y3w?|(}E6pKL5Xv>J3vz9++@iN=#5$4Df~+!0;|2PcYZ_X6=?^ChkJkk}
zNn`lSryI19cW5<LVB{K2$fEbo=XHnv5NooVEIdcg&3mexZ)8>?OA!K^iucyRnkBLk
z_#E0m&i1f+_3G^#$bofAX9%>YrH!PHl(X#e8)e5LY6yFZss@dfOy+1tW9w1L$h1af
zq?eKP!jW`|ivX>6jf-E+;WNh=sX^8V(I*w14+F*7aW)N|%MfC7beU&4ei!7%CXm+I
zL*r6_-tLJpWDy$axgRI+U-S-k(v#Grm{lxFze%|r$AB6NfeQ~6q0x|QP0|o6p7ESw
zbn(#Q5x1Nfmm~)S-@vZTWdPb(#|7=NR%(6PhX9o%GXNd~`W;&$%|Q`|;RhlnZ#hp5
z4Y!9w%N-q$dp6Y0bN^v-1JOagsvXLHzpLAinwsp^y1j+H^fqVLhdppXRtW1<Rp=oF
zxxs-*fdkFsEUoS<K2-C?iHJ>Wcw%~Rk8%sUm~Tri(lM`GGnNAWlh7(2AxE8+V+uU9
zzT>;#c@sTtylm@iB3VwRTTlJco$OJnI(uRh`4&F)5f#9LU#*t2?mOpiZq@P4WGnfu
z!P+>rWon2X)>3VGIypE)4aMnI0j_{H*6f8%a_3y~%xdyek%G^Uz~n~o9kClRz8PJO
z_(|pdO>^IVFW|Yj>^&Ramf)4hci#EZwb<Xol5B{Q3n(^i5<GHA;ZM8ySqDPnjue)?
z^z}s9Rd7!M%W6C+tLxX@%F{nmZCeN)GU1(S;J4BB=t``~ike!birKx~;)3r)_z%k?
zF7ZW9N3N*Y?jJ&2_$v@<D^esw@z5vV)dX#7!l>q?F0_93yU1P!=$AC$SH$-JP6Kzo
z_yS*39t{9<d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zxx+&;ee<G|0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8arRn^V_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.hpp.70C411CE7496D684.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_add.hpp.70C411CE7496D684.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4179663f06e97cb992a8f9827bb3adeae173eca4
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er~eT!$QZTz<Q<iQ4=y
zR=n$*n=vcaZ1>cEdoL`iSU>L&&*3xP<{h788IGA4_xMeW5j*knB<o3!aPjsD&*sdj
zN}aPqLc4=CccP-tg;guAh$@91+LdP4YgxYkwXRI`ib*^-Z@qE975d95^>18d^whJ#
z2dC`3V5atLs*Zt8(QQcyC4C`@@~@r=>otA+KRVC<KbyDxGjIFr)ZHSFkNsj&x_P9i
zxH2~h=v;=P)U;v<os$Tni;I$r^+60K5CH)|FF^qli|_6i(<NmRE;BPSF*7j3#2FX>
DE97DW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp.BB89842592147883.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp.BB89842592147883.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8a8042ffc36b84c877abb127677d3a581e5853db
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TL2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(j3;Nnn-}t8b#`E}x_*5IOm5DAmnN1E
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)p9y!pe)}nSLN!%?We`#$NTd3o{=9B?`N}tSc=BA
zD4HO(473%U+o3O|m1Z9(2<6(&1-Uj@ZqQpaVx7onK~|Zh@dEwJH4Ux3^!t;A$Lj)~
zq%r*E(=FP_JG7cAFmjD1WYK%)^SVX9j}6&63(wJW^PVc_TbY%}QiOn};=OgSW{GSB
zK8H4tvjePNy?Xlwa$vpE83G+@X(Oov<t)4WM%l558p4jEYCvNplR22t*m+bkGObY=
z>1AZSa3EdcB0%R|>*7~)_{=dzYLPWU^hrhM!;#|bIGdKvWeBl7xXiO0zYDUq38Zy)
z-?|iFuzO+*S%g-4?#BuI7rle6^dvPYW)+Ll*D1H-5Kv1YaN(gMG#YZPNg86sGoDk7
zF77)#;FdGvlH`Ek8`yQZ3_!JYT+jjQq}Hc>2vA8fN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35D};5bD)f+o
ztno;sz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sU=UUp43kt`?Eucv<LUiK(elRdGCd<&n3hzj7ruU5-h|DE$Ux9ad_vXy+-
zU~Qb*Fm;R`)>3VFIypE)Eyd|o0j_{9HtdB>a_3y~%xdyek%G^Uz~n~o9kI0--;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n=O?2_Ctm@TcAUtOKEOOA5<D
z`g)@5E4ZV8Wi_6Z)%ELsmD_rxx~>pBWWqbu!f&JN(Un+}6%BPt6|;M}!3E!m@E?{(
zT;hwIj$Bc(-9Lo5@K+$zMx;oF;(kcJuL;`JhEc;wU1<I6cagmup<mK~UlH5?I}P0V
z;tQ^_9!vmpd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zs`K-ds401t0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y-!{Frr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp.08DE17CE7004DBDC.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp.08DE17CE7004DBDC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..61651ed42a3a6de18bbdba399df6d174dfa6b062
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er~eT!#&KoPWC--%e=%
z+fyjK-L!PFfV%F_-;Sq`y}Rq@=HzkyuEHsQh8d3z8Aw(dMJgWcH4ak9Rb$iB)0Fyp
z%A%NWjl#jquRM8eQ7%hOK7??1d)LeidR2FQ_Ra8c6}Be<;?|efT}nUow)$zVD4&?|
zu_H6T8FGecShN{lZEigHS*!13z<ic(Qg`=s&i(c0evU<qyv3dT-%CsHM8+#;9%(PG
z%uND1m!T*%tr$Y*B!cMTqU2(I5Q7OsKmgE7P{71uInjtExY)3QnURT^ff**wzyJWp
CUtc`{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp.15C86B2B21F27130.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp.15C86B2B21F27130.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc4534e6624d755770ae717db5ac248bb0176034
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TR2mk<hoV8V3bJIEy
z_WxlvYT6Ru%;cdB31pych6B?_k0Pz(oyd~WC29Eetn7r)z<D6Ud9k$LC+&UZ0-ykJ
zRb`yiAib)`tLo{^g)HFh<U={Vzq$Q5EpKPT(FySSnLJyZ9GESxE>D2b^$GA?#r*zr
z2JYWZZ@;|b&6j_t<>J%Bth`yw$3q6n^8I{KE<WCTo;`fJD{t-?`5p0oGU<uAs9l4i
z4pPfN8_~Kg`chhHc7cLWu4!G6OM~Say+tFIiHsIxl}Q@UxPSglLrX9H?x^A6s(?pn
z41f7_gEsOGt)>c$T%rkC^xpZrZqV;yO?H)qr|7wOkCpR{%$LYggn*{vy>+l;iEIQu
zhc=M2J*-~5c=H;vW8I|_1X|S6MpApqSvL98Wyd0F2wRG(291?WX0JwL>ru(bv_@s5
zmyz#<J?RoB0b1`G7k{Y3C-yN?gM1@IpHy@{94L;CqiN_=h7g;*$=u8F7bjOXfwa!<
z8kYj}dXJ1Di_l2V^*Dmx&^y>jPg0X&R<S7kD&=zQ9cm~9PCQhEMnlduNke?`gy#^W
zi@O&0xaG(=CD|eP33hEx1JKGkPH2y{QtRVB1gIpL1K=^B->@dq>=dyZ{y@a!Eyt;$
z;qtI+xuU~i&xYE0u0M>fAv(yT+M#Usm%RP1sL8Ia+gjL4Z*z3L+XpVl3SpV53O%GC
zS9l;&U`KO1ORM{gAF}!EM8u{w+%Y}4MY)De%+r#Sw2v$2jHQ5I6I#VB<fzlK4}qK3
zcYNnOZ=$D_mu;O@B+J2c>#<(ClRZjRXG^RiPvK+lQ32fe)nY#BzEl3@QtjVN){<up
zmd2?yQwQ&1Db<?0lbtivP#j+r;0kDC%~n_?SI#BRtR_DeDfs*djIIUGh+T={_26Q_
zPb$}Moa^?gfT!ZT_w4Yt1g}J%dFM;lVn2r^*$^csP^{V{xaE?<KQ{BH286~9DJ*;G
z$BDAb;FbcG)o@f+SC`$&lh9laRNEGU+e~<;8hAOl8eE7aSy5A~R57`eYn<^^g#XYz
z;3S`EcHoMNHUA^TnZE;})*?kR6nA|By1bxGO&HW1)|uK*ei_-y0s1RN_#3|c-yOl#
zFTMjNA0S+Fd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z0j|;;M<Hjv0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Ztbl^?^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp.152CF392CC9366CD.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp.152CF392CC9366CD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d004a4b8a42a161bbb03092bb0a0f22a3cc56cc3
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%#}czk%@t!V$Rn7K<+~ZJS~4+*+gqT
zoNrlkHYQ^Blx3mc-na4|{r)m|$KxHpizHM(@-nEM=9wI<GM!zdy2v%itGnBIL(G{>
zj~?Ujr!0#Uzn&0Cd*M523#*dpp=}~|eoJNQu18ybyKqD#vih6tThBJH(w}ja(Y;sK
zE)re0rLVI}nV0ALC3Cj66a1V#vwx`+c!)h){$t(w|7Y7CzwCPabWhyYS?9Iia<7Xy
zYR*txnVST3FhfylS}}yqNd(cwMajkbAO_GUFaY#06fm(Q{wudS!FMx?nURT^ff**w
GzyJV6{$!8<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.cpp.07FADC8850641162.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.cpp.07FADC8850641162.idx
new file mode 100755
index 0000000000000000000000000000000000000000..70fcd9199efb675dc15e638105d9ae8c89f37601
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@H84eLd9maD`VJ*IZrtuejqZaa&N<}VhbjOH
z09RGUNe$Agda|yb-d)H7-cSE3XZJU^pJwIle0+TZynP|hmnR$M%gfg%!1dJ$@KVL%
z{!0e#-_LHpe&EH|zh~w0^TWKnSuQ4H2Fmi|Vp%Rf-F%rpe7-Ah?il$Q@o_rsiG`?L
zgQ5;n%RpPvx*hsbT50xyf>5q$U63n-<p#Y)BUXuw7G#x48qYX?{z*eCFa7?g;o-7?
zM`;Ydd3S?0@(Qh{3XEK#30d^s`TX6Y-^ZHlItx$HbMYQ4=Np+1k);R$O~q^LV8s&I
z2)qw%AZL46y?XWTEo8^KLnjEdsHKgh2Fh7B`Q2g1B5DXbimC>Ul}u((qp|g<WMo>S
zGSbV)=fXg`#7ThGyT-+5b@;?SMrx2xgy@rs&W8iV(Qz~loyrhmGnmZ19Di_fZ4*fA
z?7nd+K(F`67_tbB^jwc4_%C_~Tj@z^Qp_qArC+C9j=`aZLg2(hMQAkST$41!2Tyno
zF}k>KalkD{#wp1T!FRB0a~gow)^S36td&|H*C9Y9$s7QW0sWRWk!Gie-S7t@CT}@T
z4GouvUCR|6k9#)M&U5|Y`U;|hJgObac7Mp*&x)Gt#=4z_o%A+G*Sme-f~*i$sjARJ
z3UZAHA_aCdx3jdm&-fvmFLp$1TEiXFgFBQP*v339IY~oYIcF>d{3oGR+(M2zEkg+0
zw7%mz=Xnu5t-WmPtRh(sru!c2r90W9RCTt*D)JOQ_8t|$jbAMn)9yXvcP`a%XR?+&
zYp^m-ZJ0WE4=brQ+@0*4p@!n*q5xMw8*8@0D!Fnld1f{FiAcfcOW^uS@Qm2C7+;Mp
zMtr4m{U*6?e-!XkoOhoc?v~({$TRPJ>00dPup}Fz<OGU!n*_I9Quxbee${}`xFv;U
zFFo%lyA19qU|C(hetx+7N2+ZL!A&N-Qw_WsU5+lqima%qRjQcY$qmkVCc=Mc9&w7#
zGCFca#d`k~;>=%xP#cjV8H)Qp_%0`CQxirt2X&_Qlix+Qa)5qI0e**X|8F;N<%@4m
z>mC7fd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~IANm5
z3D-<r0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8ZkIL%`K

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.hpp.B81990FE8C5D8509.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_add.hpp.B81990FE8C5D8509.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eaacfdf4587ac82d8a923d8bd88f6d1899016c41
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%q2kjKTxh>&er~ee9Z<t&OaBq-pczu
zQ)5e_OUtsXD1mx=PLnOxw-wo=r~SReyXX%~!^O0=i<*;8Gu;f+^XM!oF61yvFMc((
z@X@3Q$Dj*hl80E<c6r@Z2vS`W{p|LP;_tif9opEcHElxh?7v>Sz2aBo+}D<0T~wr<
zF_Zgk@}xaMTs-UV@UgX>=w><QJeTv$)k9ml;(z_QpJQS3?s;I3^=oORrAP9LD|3^8
zPGu-cO)G}bIf)>;xG1?;AH-k+5n#Z?GDmFN{KwDN8Z$F8F*5)?2LsG71_J{Co+)9p

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp.562409E747DBF8A8.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp.562409E747DBF8A8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c6b48fa578ef51882b9f693fc27158c25bb2d498
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TO2mk<hoV8V3bJIEy
z_WxlvYG^5NX7bR61TxSz!-46eN0HX?PGrgGk_*3{m7O#+a308T9&BsBPuly+1waAd
zs>(R2L3&k>meuy{g)HFR_+vS_zq$Q1DQ~C4>l5J3GkH2c*)W}7zB~c0uTFsHDrWaz
zGI0NHa{Ki?FTVaeDd(Rbrsd6iHX1TemLF#Ga{lS&%k<&%U3qiI$nS^`<8e>SMC}?B
zb&y&HT8q|g(3jFmvkMf2a!u=kTo^1@=q(ztNMy7it4z{(#`*J48d`YicSj8mmjygZ
zWBARx8?=#EXf;(}<N{5|qW8|{?;8Cs)?}Ai*hbIAd#s#qWIjZeA_O!QudRaxOJpPP
zKD2?H?P2xe#oITK9qSIAAkd<gHj>&?&a%m$4m%c6L)cJMHE674GJ7=|TaQXcrZp-f
zy^MS=>`9k63DA1ixcH<Fx9nr22KhvYKB?$@I8YoNN7K-$3?Vjqlew4UFHSCP0%@Jy
zH7*6{^&S~R7NL=z>v06Xqj#{Do}?zltYT66Wy<B)JJe7JoOq}RjfR|Sl7{%;3C|%$
z7k4f0am$f$O0q-n6YScY2B4*NoX{R?rPjxF2vA8f2f$-Mzh+IO*(qW-{DFwcTaHsh
z!{uSuaz%&3o(;A0Tz|N}g6JTRYKOAjU-I_5q9(htZew91z0J|}ZXdWHD}+U=D)f+o
zT;hR9fgR25EUoS{e#qvt9TA(>aL4rE2IUIYF;7cQ(mt-7GnN8=OK26hkfTn^J_K%B
z-|?OEyojEbUbb~skt_$(eUJ6ho$OJnI$L5Dc?utUj|$+%ujaFH_nz=Omui1!vX(q+
zurN-om^ye53#nGzo$Q>UhT`a=09Qa8Yqr8FxpFRfW;OYdNWtez;QC7NjM$|ZUJWh=
ze5G>zM!9aU3)mLt-DiipC3q$B%sXGY7W+9Y$%ZI7fnwPv!7Y~*{;`=qH6S#uNnzPb
zPdmykgBuE1R+lfI9@)t4K(%clxXXlhs)1L7%fW?MkQFtxN)_Wfxxy(AMfeZZ1J3bz
zRtK)ASn)qXocS{lY9&%6Lvh!K-(>}DYQmuAsLs@W^25kh4$xmI!Qb%hf9?maeenYr
z^&fk4d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003QAQjcJqhwH5du$808an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8X%@y!+h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp.DF4FA5048878CC52.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp.DF4FA5048878CC52.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b9aa906233c716da2e68a9fc47dbdedb00ecef22
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@fr){kV$Rn7gM5b!c$|MKDIM&c
zzjBjV^2J+AvTpVMv%RFI_O98PZ=25Vr$$9f*cp6gai|?`IL4h4yV63W=kARI*{8BB
zT#AfE7jo#>ZF6{#o-tMVIv=C<u3N^j%jf9tf2}JrYyFx>9#NZ&-%f0AT3WxpW{p=;
z)SNV}9laYPTtBj$zf;E6<iK5BwA)H3ujAG8sc)}5|DRU;_-~nvT21bT{1*$3ykRJ=
z%uND0wkS2N7((YHg6QI+<YIjg1LzAVU}E8&!6sRGb{{`8BNH<N(CaY33}Y}b006B{
BU}pdT

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp.F349F4E1672B2959.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp.F349F4E1672B2959.idx
new file mode 100755
index 0000000000000000000000000000000000000000..abc24bb695f3882e7310355d0f5f5ff509df59ca
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TL2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@G%!VKd9maD`VJ*KZrtuejqZaa&N<}VhbjOH
z09RGUNe$AgW>QyA?=EBk@27v2v-_LdPqXrNKE6Hy-oB9M%aaZB<>l)W;QHzWc&TD>
z|0M(W?`OAPKk(w~-?MW0`C(q(EEkh817-Pfu`HLLZobSPKHrr$cZ~du_&A*o#6mQ#
zMbQMQWuT4d+!lQ)tu(tpK`7UDF36R^a*f`i5vxQ-3$n^2jc1%c|D>UnmwtED@Nikc
zqcn!!yt_pkd4*O}1xBvWge-dReEx3G?_xu?&cajlT)fB1`BvsbWGO;GQ}Nn5Sg}Ml
z0`Ef`$k_o_uU@@-3)!*$&<O$^YH1^>J>@K${O+)05jBJ@Mb&`DN+z>cqp|a-WMo>S
zGSbV)=fa+JiIV`Gcdd)h>hOtujMO5Z2+=1Moeu|!qvL2=I+Y>Bc5gEGa{R%`+9r_J
z*<I^WfI;t(F=P>1>A4<9@L%)}Hqw*Sq?lDKN?)g3j=e)Ig}{l2iqL4txh83d51#NG
zVsvrW;U2df8K)#W1mD4~%V_|rt>c6aSSPhUu0w!Ik~sh#1NseXBF#<_yWtN+Ox|*w
z8d@$7yOt|D9uI72oag$(^%X=1c~pCp?f#IrpA|LPwRKwyTj_0%ZgBg+1z91iQdOac
z6l9GDA_aCdx3jc*$oL_fFLp$1JHs6_fLoMn*u*?7IZ6Ata?V%^_)kKsxP=_`TJ|Au
z)B2w8o##dLRD0PqSw*rOO#eOBOZT!zshVtwRpcpr96Tz38^2mEru}=y?_8?=oyl7A
ztij4SwPxzzJ*=c!b9b_HhFXe~ivnB$U2ND2tK`bL<eAmvCn5!(FM;bT!82lOF}@mI
zjQC3B`b~1({wUz7IPX3?+%3T?k!Rld(sek@VM#Vb$q5v7mjt(5Quxbee${}`xFLn*
zAU*FW`wVU=U|C(hetzVp9;vP?1UH%RPPOo6bUC^ZE3%@YPN`ygC)YUVnF#-(dBiC`
z%jn1z73=*|h%<i$LajxLWGL>2;QO4QO>G!89MqZGPktBK$^rT<1^6Al{lDG7l`p;l
zpGqD;0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015x?S
z=ZC~kKx6_>PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOw<O`kpKVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp.1B257007DC8DF36B.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp.1B257007DC8DF36B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a855882109160d2d9086cabf635518527a9e5ae7
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%q2kjKTxh>&er~ee9Z<t&OaBq-pczu
zQ)3IGl2fp$ZpZ%lCMH{~Z!5A#Py2g`chMh~hLd7#Cp9OX=DKO7=h0bGUdXX6rTEp<
z!e^5r9D^>1Ngm=^+vRmvF-Uby^|RYkiofr^cW7g)*0c%1(SN*ld&RGKb6;D2by1Oa
z#!T+L$5kqqibx#0!^hTkVl~S#=eeA3t{&Rc8UO3g{TvIMch4_c?Td_8Ha*f@T$!5$
zbSgtpYFaUb&PfE(#YM@*`XB}qh=2f~cc6fYB_#IO!!E_3Rm_Y`%nZyhaRvqedL3b;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.cpp.58C6C0B25D770B51.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.cpp.58C6C0B25D770B51.idx
new file mode 100755
index 0000000000000000000000000000000000000000..503ed8527916df65ed640656564228bfb6d2899f
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TL2mk<hoV8V3bJIEy
z_WxlvYG^5NX7bR61TxSz!-46eN0HX?PGrgGk_*3{m7O#+a308T9&BsBPuly+1waAd
zs>(R2L3&k>meuy{g)HFR_+vS_zq$Q1DQ~C4>l5J3GkH2c*)W}7zB~c0uTFsHDrWaz
zGI0NHa{Ki?FTVaeDd(Rbrsd6iHX1TemLF#Ga{lS&%k<&%U3qiI$nS^`<8e>SMC}?B
zb&y&HT8q|g(3jFmvkMf2a!u=kTo^1@=q(ztNMy7it4z{(#`*J48d`YicSj8mmjygZ
zWBARx8?=#EXf;(}<N{5|qW8|{?;8Cs)?}Ai*hbIAd#s#qWIjZeA_O!QudRaxOJpPP
zKD2?H?P2xe#oITK9qSIAAkd<gHj>&?&a%m$4m%c6L)cJMHE674GJ7=|TaQXcrZp-f
zy^MS=>`9k63DA1ixcH<Fx9nr22KhvYKB?$@I8YoNN7K-$3?Vjqlew4UFHSCP0%@Jy
zH7*6{^&S~R7NL=z>v06Xqj#{Do}?zltYT66Wy<B)JJe7JoOq}RjfR|Sl7{%;3C|%$
z7k4f0am$f$O0q-n6YScY2B4*NoX{R?rPjxF2vA8f2f$-Mzh+IO*(qW-{DFwcTaHsh
z!{uSuaz%&3o(;A0Tz|N}g6JTRYKOAjU-I_5q9(htZew91z0J|}ZXdWHD}+U=D)f+o
zT;hR9fgR25EUoS{e#qvt9TA(>aL4rE2IUIYF;7cQ(mt-7GnN8=OK26hkfTn^J_K%B
z-|?OEyojEbUbb~skt_$(eUJ6ho$OJnI$L5Dc?utUj|$+%ujaFH_nz=Omui1!vX(q+
zurN-om^ye53#nGzo$Q>UhT`a=09Qa8Yqr8FxpFRfW;OYdNWtez;QC7NjM$|ZUJWh=
ze5G>zM!9aU3)mLt-DiipC3q$B%sXGY7W+9Y$%ZI7fnwPv!7Y~*{;`=qH6S#uNnzPb
zPdmykgBuE1R+lfIo<?p5s%;CwO(wik4ZIp$4lcxktf;9~su<tN6;633!hdKUaEi||
zI&ekBdjApP%wK^}E0H1@in~7eE+=SH69zR0b*A={-$k}^fc{DW{)TV=Z#Qt|iyysv
z9#{Z#d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003QDcHgKA#Jp(*jRW08an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Z`^T*Nv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.hpp.9E23A16BC92C20DB.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_max.hpp.9E23A16BC92C20DB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8774b23bacdd3b1d338fd6eecc1adcb8ddd05ff9
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~ee1{BpoPR1Q9qgUI
za+6u|#am0VZuS1Ny`-l0uGyJyo6hg2Mny~58GL4Os2y%N#+?$o(n6%??u`T4r?M<u
zii|}Ua_HD?b9j)RF;)3GAEWlJTgI`==jiW$tt&EX{hCJ}QJak4PHb;lTED(#jaO3C
zoHVT+y&EH3KeC*^Q^wZhz+GLm+e#>}<JI%2Z?8Q6pH}?%Z&}Qt>&JevDQ6t%W+<-A
zO#(WWp(r)27((YHg6QI+<YIjgg9$`{0Tau+te#&|s&;dk8JU<FfZl@vW*CEk0RY%8
BVwnH{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.cpp.0B95E75A805DE595.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.cpp.0B95E75A805DE595.idx
new file mode 100755
index 0000000000000000000000000000000000000000..70a299bdc43531d891f7f6ab0f0c43f22a443b3d
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@H84eLd9maD`VJ*IZrtuejqZaa&N<}VhbjOH
z09RGUNe$Agda|yb-d)H7-cSE3XZJU^pJwIle0+TZynP|hmnR$M%gfg%!1dJ$@KVL%
z{!0e#-_LHpe&EH|zh~w0^TWKnSuQ4H2Fmi|Vp%Rf-F%rpe7-Ah?il$Q@o_rsiG`?L
zgQ5;n%RpPvx*hsbT50xyf>5q$U63n-<p#Y)BUXuw7G#x48qYX?{z*eCFa7?g;o-7?
zM`;Ydd3S?0@(Qh{3XEK#30d^s`TX6Y-^ZHlItx$HbMYQ4=Np+1k);R$O~q^LV8s&I
z2)qw%AZL46y?XWTEo8^KLnjEdsHKgh2Fh7B`Q2g1B5DXbimC>Ul}u((qp|g<WMo>S
zGSbV)=fXg`#7ThGyT-+5b@;?SMrx2xgy@rs&W8iV(Qz~loyrhmGnmZ19Di_fZ4*fA
z?7nd+K(F`67_tbB^jwc4_%C_~Tj@z^Qp_qArC+C9j=`aZLg2(hMQAkST$41!2Tyno
zF}k>KalkD{#wp1T!FRB0a~gow)^S36td&|H*C9Y9$s7QW0sWRWk!Gie-S7t@CT}@T
z4GouvUCR|6k9#)M&U5|Y`U;|hJgObac7Mp*&x)Gt#=4z_o%A+G*Sme-f~*i$sjARJ
z3UZAHA_aCdx3jdm&-fvmFLp$1TEiXFgFBQP*v339IY~oYIcF>d{3oGR+(M2zEkg+0
zw7%mz=Xnu5t-WmPtRh(sru!c2r90W9RCTt*D)JOQ_8t|$jbAMn)9yXvcP`a%XR?+&
zYp^m-ZJ0WE4=brQ+@0*4p@!n*q5xMw8*8@0D!Fnld1f{FiAcfcOW^uS@Qm2C7+;Mp
zMtr4m{U*6?e-!XkoOhoc?v~({$TRPJ>00dPup}Fz<OGU!n*_I9Quxbee${}`xFv;U
zFFo%lyA19qU|C(hetsIVBh|Kr;3gB^sRrJRE=Lz)MOM_*DpgGH<Ob(F6X8EJk2u9=
z86CNzV!eL~aptc;sEtUG48?sPe3uinsR^T+gE~|D$?qatIY7Uq0Kdbx|F;{s^2Il{
z2_8~&d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~(S*=F
z9l~!j0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8au0?t<e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.hpp.B0727F2EA98E1F00.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_min.hpp.B0727F2EA98E1F00.idx
new file mode 100755
index 0000000000000000000000000000000000000000..51496e3be05758720b4d9591f333984b1e825b52
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%q2kjKTxh>&er~eT!#&KoPWC-=WPGm
zQz%?+>N`h3UH9j2$J3|Y-Q~0M;CNX%L*;`o!!r@dG-ZD^?I&9b9fP>;9yzdd#^j3<
z65@KN7+qpB<{i|%u!?1+tWD^lTgLaF&1t{?RqWf@P|vP4y0@l&Q%GKTX@9s~SnAdg
zgH*jO{hd*&r9A8J@UgL-Uu|*Bc`m1nbJ1zFvX`IhH$VKcW4^d+`89Lplp~$RmAOek
zr!o|!rWHfzoJ0^^T$Eg_52Tq`s=r27tb7+z!OY0S%)kVag8-oCV1OCMU|;|MTH;>i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp.FA19AE81B0ABEEED.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp.FA19AE81B0ABEEED.idx
new file mode 100755
index 0000000000000000000000000000000000000000..635672651ed9d212d4685e32ac243ed10fde979c
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TR2mk<hoV8Wka@#f#
z<^N<>ux!^&I->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$4y!D^}8nXwnCg*ge2r&hCN*
zKmp*Y$~dV(dR0%>)$_XxS-|`0U*+uK_U_}Xyqk}2P84rn$@AsOhWYa9@<efSeWG}6
zV)5`fD<0m@?!NrVi!XoA%H^lWd3n2BOvbDz%MXiXx%_zhdH(q6zP!C>&5wW&(`gSZ
zMC}?Bb&y&H+KSfg(3jFmvkw%6a!u=kTp27k=q(ztN@TPkt4z{(#`*J48d`bj_eTql
zR|PytWBARx8?=#EXf;(}<O)s5qW8|{?-u<&)@0XNc#fWn_f$FG$b5(_MF?mrURwt%
zmdHlneP{zY+pE=^H}Bp;4y-$LqJb8*w2{;RIm<4;IqX<O4Pi%7)u6GG$qZ&RwjPy?
z3~N+IdKvj#7(kbp1Zcf$TzoNy&m3c<2KhvYKB?$@I6#~oXVcKB3?Vjy%iPQHyCBy#
zfwa!<8<zs~c8`c5i_l2V^*Ad3MekrMJxNW9QN^P4>y*ne1k_LnOgvPCMnmSBq#-_d
zf^&+|#eIu|-g0J4Ne&3UQM)$N0JOG_3GH>Q)cUv%0V+x6pzs*bZ`l%Q4vIJoe*j|g
zmh;rmaCtbiT+#8kM?>vA*B@@KAv(yT+G*MEcf9?msL5`u+gaF2Z*z8i*as}g3SpJ1
z3O%GC*LVO@;6QUbORIav4{W~L5wU3vcTBI`q1?bW=4r_!4S8kGSPJ-0LaVri9CcQP
z6u4=9hj+pAB6?bT+1A-avYbr!J=RNivPY@v?1@d}DSYZ9Du5fmS}vyDd&civs^QLL
zD|yynWt`eDbch~SQf;_9IXFWN#mPkhu7Eby?1fEo<y`WNYVs42g3p)0&9&eev1>8D
z9$k$1O6B@Za@~F};JG;OK0Dkk!7Gtx-ucqC*w0}}Hblt;iglX=w_H;A({6s&fzY@m
zg=H_j>?k`1cNDO!t}b65o`mLdq}sL++-AZ%)xd9~tI>s6krg$yN)^+4xxqP4MfeZf
zBPRJww<A|nZ21o%&iow+wGk<jp}6k}=y*Y!nlP$4tuw8k{4%nagX)(Q;a9}=e|H2|
zzxWCkhagaMd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zD~3h@S#n_C0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8ZaG}P_@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp.7E4B9D2EA0AA55ED.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp.7E4B9D2EA0AA55ED.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e8c4241da32e4dec6dad693536358b2028293d7c
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5Dk%@t!V$Rn7gItFVc$~kx8kcPU
z>r)_HZtD9+KwbCeZ^zT8-reQ1^Wb>-)2wibI>VV+9BGFej%ladU11^8bN9x9>{D45
zZbjUp3psS`wmCcy&zP!wosUs_*Dd3_&*$Xtf2}JrYyFx>9#NaL-%dQ<w6uPG%^I(y
zs5xm`JNh?9xPD|=e}|8)$w9lisNG69ujkeD-nW;Y|4%EvoNVF0Wl7@lTIO{MM~#as
zbCZCMWhhEbD~8ZHi6FYTD7jc4!~psN27sQ00w$KgFhQR<?)eqWj7-c7%rJ2V1^@$>
BUYY;^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.cpp.B429C5C7F2FA95D4.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.cpp.B429C5C7F2FA95D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6d8f0986c35cafca3bcca695c74c42acc641ba48
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TL2mk<hoV8WklG->B
z_5Y+A4H=ePs(6?ou%R|HL*-KQ=n_&JcOgq&U0}%9r)5LPknBUK?2Dy7M}2R}1waAd
zs>(R2L3&kB*45LSD_Ow1=|AP{{^s_>th}8M$7jInXYzb`c3{4|et8CrM`yru6N~%L
z8MuEpyZ!PvZ@zq*mCKJ0^YUi7m<$;x%lC_Ax%_bRdH(S6uDrQp<X6P|>9i*nqIM06
zI!G-8ZAI&L=u2s(*#`<jxu$hNt_+qN^cIa+B{Eu&RVHb?K>u=0Ln|-+{$%0dx`0P%
z41f7_gEsOGt)>c$T%ieB^xpZrZqe^!O?I7yr|7wPkCpR{%t~Y_LO@gT-a1&ZL^cAS
zLmSB19#$`2ym<{duuka=fflv2k<@{5mR<g!>{vt%VMkHbps|w49L#8JJt`TQ)~Jm1
zGO}JckS=i%p!KeC@w+*E;us?}$QmK~q@wfTNO5+YO+)80gxDNh=2?zE1-Z5fq;+=R
zxD=qbdtwY(ghqPq#|iw3-oaLSlA08=ibd(yDYxSgP(vYb;h`cl8gi{k8e+vWo>Po2
z?pr+ImNVm$<bdD@*tNL~Kx^x`pgq<~txx+9pps;cfX9G-%a%xUP{d*QBN3CgoTr9{
z+ry#djt+-C8*1md|1cgwbdaxVhqB+F>h`OqCcCk2XJIG3&Dr%~AGshagjK34^pJvF
z<B>>#1I^<st?ny+tmd;55u4WV#Pr|}<p#Df-<DjYLteRNECu{Gp;bIWjyfxc6nJQT
z$9KW=CVE<X+1A-avYbq}p8BOb*`ri-_QWRgEqv@FDu4&SS}vyDcgEk`s>7SfR`Okg
zm2qmr)G>NkNwwkW<lqc76em{&xB}W(vllkWopZ@EtI1D93O+vq<B{MyV%K6g8e9$d
zN#*`ca^GGR@Kjv(o*mzo;FZXC-ucqC*x$pFY>1KzDAsKfJaS3lZ@c+j2SVeP6qdd8
z{Y2SSa7O{l>iT6_jYr+e(?3vcTL>O9;hk#Wufg@;O039=np&lb>7Cr*obN>V56c5C
z@kLGtuBh1VTOls|6$rHvDUzYM@00Ipf;Kf_P;*ijT0i?;WG_ePk2K&<#P<JA19!go
z24h7Y5OaBLVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015qWp
zei1|*>1+Z|PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOrAd4tN;K2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.hpp.5BB1B8BA3E60EF71.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_add.hpp.5BB1B8BA3E60EF71.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2bf92945dacdfe77788d58f94f750cb6faa2dd5e
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er~eT!$QZTz<Q<h1mQs
zR=n$*n=vcaZ1>cEdoL`iSU>L&&*3xP<{h788IGA4_xMeW5j*knB<o3!aPjsD&*sdj
zN}aPqLc4=CccP-tg;guAh$@91+LdP4YgxYkwXRI`ib*^-Z@qE975d95^>18d^whJ#
z2dC`3V5atLs*Zt8(QQcyC4C`@@~@r=>otA+KRVC<KbyDxv}F73(%mADkNsj&dUvF_
zxH2~h=v;=P)U;v<os$Tni;I$r^+60K5CH)|FF^ql%ekz7Q^SHPn3x%vm>HO1;tUJ`
DWPM^f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp.CFF7599DE0E79FBB.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp.CFF7599DE0E79FBB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3c55d86c0b66dfc5837aa2d520f5f7255351aae2
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TL2mk<hoV8Wka@#f#
z<^N<>uwzwCI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(Tu;t`H!tMH>g>Q`b@TcRn2gVWmnN1E
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)pB~xKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3vz9++@QB;#5$4Df~+!0;|2PcYZ_X6>GvlKk2eK8
zNn`lSr(3j<cW5<LVB{K2$fEbo=XHyI9~-iD7M`Q$<~>!;w=ye{r3e8{#e3^u%@WxN
zd=70OX9rlldiC}V<iL8RGXy%+(neAT%2{^#jk04AHG~~S)quuICUY>OvGb^8WLl##
z(#yzt;Xt~?MS#w`*2S;p@R?(b)FNwy=#z@hha<(=aW*ZT%MfCFaG7U0eivkI6G-dq
zzI7?UVE4orvIwp8+>aCZFM0=C=}Br*%qkY8uTyTvA)uB*;KD;iXf)(nlQhJNXFR7E
zUEFtgz%6ISCCLH7H?Zq+8Gvf*xS#{pNv%)&5TKG|j)2F2e#@3fb5O)#_#+XMx16Vj
zmfORj<&Iuo4{T_h=l;WF4ADWpsy)hnzpLAinwsp!x}Alc^fqTVgni_KtPs|zs?b9U
zvc@Bk0tcGMSz0|*{8-HwCnC0;;fWc*9m)-CW4<l9NQb;~%~%TfPeQABgdFu&4k_@^
z`kwED=S}ofd)YPFM6#Srzn=P~d)cE@P4>hl@-2KCA}W9fzgjJ4{ddma+^WNy$yV}R
zgSByL!_+Z)SWC6x>Ez%HwG^jU1-Jsb*svEi$(?h_Gpor@MG8JY0+X@eJ7Q~bJsw?+
z_(|pdO>^IVFW|Yj>^(caEx{|1@4WM+>u|V-CD{-q7f{q)5<GHA;ZM8ySqDPnmK2tQ
z^z}s9S8zuG%j)KJSxv_MD!27WbzLEN$b@&Qh2KUuqbso{D;nyQDrWa`gA2YB;Xf>o
zxWpGZ9l4@nyMG9A;jcibjYyFU#r=?cUlX*c4WovWy3qRB?;?9SLcgQ|zaqB(cN)0!
z#TT-q9!~&sd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z9}o^)n2&k!0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y=*w-om

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp.46A8C8C2BE2BA19E.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp.46A8C8C2BE2BA19E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8d08b133051bb48c69af5755b76e4851942a3cc9
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRZTKTxh>&er~eT!#&KoPWC-=OnlP
z?I{%AZdy87KwbCeZ^zTe-re<cbMiQUSK*XD!;D9V3?wU!A{CGJ8V4!ls<G+mX-a)P
zWl_wxM&V%QSDw7KD3_%sA3`|1y=!I$y{fxD`(}8!3fq$aaqG+LE~TG(Tm3Xwluyj~
z*pZpv3^~IzEZPjOHa8yptkw51U_Q$?sk?hR=l=S0KgZ(EIt!b3zn7NYiHujSJJL~H
znVST3E<;gjS}}yqNd(cwMajkbAO<stfB+^KgMk6)J1Agc@lPl%ZscM-%go5c%m4tc
ConTe~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp.384AFFA95058603D.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp.384AFFA95058603D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ef168aa65171a42f2259acaf45cb1eda2c5c62aa
GIT binary patch
literal 1066
zcmV+_1l9XeNk&E@1ONa-WJ!2!WprT#0000H0001UbaH7`0{{TR2mk<hoV8WkavC`h
zo&QNSGB#n8t>VEZkWd@jR5mF;P6?@p=|UsT>H=oJKCKxDG08rJ%D#-$=SY2TO#x5<
zxT-QvYLH&llXdm{_DUA;Zu(a_d$_&(I4kew!|@sL=7l_8o*kGkZ?4aP@#qYAsbcZ)
zIRg*xW_MrS^XALnvvT?AabDgo7n30aW%*&TESDc|KhGaO-IurbjQouFFrD_qLe#E7
zQ3t7Ipsi@#4t*)DH2Xk7DA%+u$d$ozgWjSMt3*Z%vdScl7u>)6rlFOWet**NcvHZW
zG={%?x<MOxhgMSsMy}9=EPC&JUbpD?u_n9D!gKUoyr;_fM&?UoDMCO~@!mREu|zfk
zpF<nS*&bG}UcG$-*|F}@83HY8X(Oov<t&^0?y_SMHG~~SRfEP#CUa1uvGu5AWLl##
z(#y#A!hv*&lK`!Ejf>yZ;WPUfsX@LGqE9M1AC449$I&!&E<=dT!DQ~`_=A&cn?PD;
z_l-*ddc7yckVR;u=X#vLf6+VGN>5UgVpg#z{W|4x92{yW1Wr6ughoToHAzE!@r>sX
zql^0%54h#XI3?L3_zre$P6N=|I!<VhwNmTTJ_M*FnIqsapx?44((Dwm8~#Yd<Soak
zq2cndYq_GsVb6xzd9FW<M-Uz4QSDH+`$OJ-R@7uS*6l3pq_;V`-t8k7WQDLwRfQf>
zkZU{=DX^otou$=%#*f*2aUx>V8t#}L+@ajSHs)!`Njk)pbH-A@e-c{7E##=vatMK&
z)^~j8Ja3|>wU=$3RV2&7bnB^Jx|2OhRcA}AB2VE{?@<BV_|<YT?Y=Yq=29KrOxBWT
z4OYgf4O2(&VI|duyOW(W)KHvU72pbJW6f4rC0EWR&#Wdt5h?im2#iO9XT+|>a5T6Y
z@RQ2*o8-FvQNVL?*?V?;TY^_2&%E=cYq6ihl5B{Q6DZbg65Mi0;V+x{RRcofmK2t~
z^zB61WpGCU%j)L3tj42m<w<C+2dZrg!EGkIQw_Wx+zhV7ima%qRjQcY%MH$XD#Cy0
z9&nN`G&^ub#hU*V;=<p7P#cjV8H)Qp0bO3urX~z(4(megXTOYW<p}+jBK!{D{_l?9
z>K9)DDjgtS0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_0000000000
z15w3uP-QrXr?CQ0PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3
zLnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KL
kAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOqagd1ONa4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp.54BEF21DD1A84A95.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp.54BEF21DD1A84A95.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cfc3cd50fde7d372c176ae5fbb1de26226338c5c
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%#}czk%@t!V$Rn7K<+~ZJS~4+*+ObQ
zoNrlkHYQ^Blx25s>=#$-djBeT$KxHpizHM(@-nEM=9wI<GM%++ZOp_em!3Tm&=D_|
z5_)_l)Uq+d!`7%pvTCwQZnH=4hTPWvV9oaXU&H!rO$=SH&HFq1n~L+&UA61>tU2~7
z>>^kAt>Z3xC$hCwzw~EoJJHQ@Y<ZP>frr?$&_Aos|3BOI_-D)Gubbkx&hoc@%e=1m
zs0BlDWo{DC!3;&IX~hsaClN#!7bO?#gBZ*p0s?@Zf&rl4p@50yqIZLoK`N&OGb0l-
F0|5Q3WPJbt

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.cpp.213C81B0FE4954D8.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.cpp.213C81B0FE4954D8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8bd404ad4fe3a34941ddaa83ee0d66900a27dafb
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8WklG->B
z_5Y+A4H=dsTgAf+fep2p87h~WN0*S=xC>eG>H<T)ep@yS8IpYnm3^?SK1b?%%LPCI
z;Ht_vsX=;GPuA7byDM41`|021?EdEV<E*@$565S~+ZXbDdA4D`yncNKj7Mj{OBIXz
zFB!OhKfC?<7cahko|Vf_5A*V7xtI(YD9aCvWx4!#^JV_<>8`xFW8`PVhv~E@7NT|y
ziaJOw18qg?cIZoKrP&7xLb;}OL9Psz8}t^9SS2!AkX0sWyx{!hCk?H<^!t;BhwB0!
zr7`^G-3{8vE3}#_Fmi<^WYK%)^LLAWA8WGfEIdWe#e1xrZ)841mLdc+6|b#>6-#6z
z@IJJGob6%t>eaiqkR9s|ogvVomNt?)P|mW+?+!Z_QA5~KR5fU<WHJXe8e5M_My53?
zBfX4#E*wagI0?{t*SPqs4xiY^NDcCd5Pee7`EaB-I*z8Ha~VQx4kmLi#~++r+XT`&
zyKh_y(Ca-hhAcuOJ=fy|enanID?Ld~idn^?^y`$%ad4=i5IFHr5gH9S*CY+`!84vi
zj4tk5Jm8ii<CJ8F;5*p0ISoK->o}o3)=I5U>ky!lWR8HxfPTxGNV8MKZulb+leZkF
zhK9?-uH}jjhdmo==ehnc9zk@FN3}!Q?hkqUSy7YSShusVliud&dbf{UkQKrzRTX+j
zL9X#gq`;2mc9vH689!$8#g2$gYq(>2aEEdO+nA>%C+QGZ&KXMq|4C>Sw~(Vw%OM1A
zTHo=V^Sp?j)?T)CR*@_R(|u3%(w*#4sybU@6?qCDdyfj>#;=x(Y4@J-JD2KkXR?+&
zYp^m-ZJ0WG4=brQ+@0*4p@!n*ssL9&8*8@0D!Fnld1f{FiAcfcOJF<_JR^23hNHpN
zfUi`p-z3-VO#x5EW%t?fZV6tAJoC<%uEl;1OR^zKPM}z~NpQ<0g}-d(R}Bb_TT)o|
z(({h8%ixXzmeqLl{BZXVRNEGUn@o788u(*yJ-8ApvZAI|sbYF3H#p~+2>+pZz$w1S
z=)e^f>-|%R3x5SdZA6M>DDL~<yPTj+O&HW1)P>s5eizxw5&A6!_#M9ezumx<Fa8CI
z`5rEFd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~dOUE|
z<}^&F0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8YJQOt7y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.hpp.496DCAD6F6453198.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_add.hpp.496DCAD6F6453198.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ce0b637cf256eb4633d69edc0fae43dde67919e
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oe9Z<tE<YDlt+e`*
zt}}T?QsN^GKZifHssVxJw+<ej>sNmMuEHsQ28rA>iMdKUbtdFKZJOd?F5aF{Y_>d3
zQeCf8RE0HnqN2}*RV%LWDuo`}mAbFjvV8w*U76?=Nt~8jZ?506`b+DnZ+ky&)KcHx
zG;QYvJGGPAOAmZF^)?})V~<CI&6Sx4;&gph|5$qd|Jk;~kL%=)?%2GKQMutrEkki-
zZW7S33`MDF#Sl6t5kwakB^T?17)&4n447Ev{@*7Zbk;|SnURT^0q8v#V1_Xm7y!S2
BV^RPB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp.13D835E23741A152.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp.13D835E23741A152.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5c08ed1331d6997f95aa9982ec05e7294109fab9
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TO2mk<hoV8Wka@#f#
z<^N<>uxeFJI->_$ks3{7TazZ^M-Bum#T|hFi;FJ(`YuSxik0*sn)HDr_8egE3l;zc
zfU7Fwqz36#Jy}=#cQ>+t_tQVi+2h^)$60wlAC518x3A>+@?yh$dHc%+Fdkh1uT?A_
zKWE_a{p|kBAH4YT*Q{KAdYYGa%f)2KKv{lREX(D`yU+8dPY>nY10&xfK1`=Qu@JRu
zP}D(c8E7k7w?kh_E6pKL5Xv>J3vy+!+@QB;#43@|f~+!0;}z$xKWS*?r9YfCJlz)X
zEREqe?{3gWUZK@gfsrdTA&cHSpTAr5hgg$cXJH>b7w@@pzLEJ5S&9(QRJ^thRxFW?
z!28e!a<+%nn>X*?LUyb>bb&yNTG~kJNIA<UzdGz#L=9m_QPrTalF1y^Xly+y8JX6o
zjPx?{xo{+1;v_)pUE|`5I^465ks9O^A^N1E^Wj8sbR11XmokLd98Kn4j^8-Bwh5$l
z_RzQ#px1k53|WLmdalPA{0F^*t@I={DP|Rm(yvo4$I+pNLg2(hMQAkST$41!2QPRI
zF}iqY@rYZFj8l>wg0EoL<}?7Ut>c9DSSz(YuS0-Jk~sk$1Ntp%BF#<_yWvknOx|*w
z8X7JSyOt|D9QJIeo#*<)cm&Zw9@P$IyWiyPdqqulW8KcePI{Z8>)k$aK~@N>R8{C9
z1-Zr(kperK+gV!OXZ)1SS34p$t>KR8!5zvCY-665oTOu1IcF>d{5zpl+(M2zEyobJ
zX?@3c&hsLAT6@{nSw*rOO!qz4OLwwIsp@QrRpcpr?ma4i8^2mErrmqS?_8?moyl7A
ztij4SwPEVyJ*=eKaCfqEh8l{Kn*v+`ZLHY}tK`bL<eAmvCn5!(FM;t$@Qm2C7>))v
z1HMwZev@3c-wN0l*WG8QyCrxf^2|G5x)%F6EXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zOD{XhE`vJ?SXSfF%Oe}PAE>r11b3P6PBrlB;C65$R%As@ty0DGL2hu)LlOQ%^?-AH
zoz;OWDpvfv5Lf;TgxZJ{$xuA>;dfa<o0>4FIjSqQU;Hq#l@s(+O7Js$`=9%PYhU~e
zQ(YgZ0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%00U9c{x(WID@48mPf!3)
z0000000000002+`00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOl2g=>Hq)$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp.2472648101515DFA.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp.2472648101515DFA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7f1a31c7dffe588265e4a9cab454f4efd1b7a599
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5Dfr){kV$Rn7jeLg;cwBxe?NMN>
zQ!5gl9_9N+!!PR7caOCzZg4lZThDzrPhZ)+is8VL&IBW#j6Dw5rgk|WywS_MXW5!I
zXBG2c!L9?3lDS#x+QKZ0-U@73wA%J^>8YQecHeVa`16!j-x5{p%W{<(5x>5lI=)I}
z+0o2pO}8bwLq+^QoXJgB;MlT3efw(VhixnB;%@rY{>wKrzn6bJ@z8a)-)t|tj{GjJ
z%uND0wkS2N7((YHg6QI+<YIjg1LzAVU}6zJK4FufzE>qPBNH<N(CaY33}Y}b000C8
BVgCRC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp.11BCD7B99E3C6DB4.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp.11BCD7B99E3C6DB4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8ee9cb2f225cccdbcfca2c23df550a690691c510
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TL2mk<hoV8Wka@#f#
z<^N<>uxnLKI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$@N&D^}8nXwnCg*mHorFIWH+
z0IsTxlNzK~&9ttb-(AT9-p~Fl=MT4cALr%W;(BrhynQ7vR%aU)tD9fWfXVm_c&%dj
z@Hqny@8@@4{@}%zzvkuY)8nGNT`i~A43y=E<*HnLy#2g*{B&R5-ZSz&;=^n<5KGax
z7DW@JmVvgSb362<w9@PY1)*HqxgggD%ME&qMywMVEyya9G+uE2@{@+vUi$q>!{bc>
zPtq8E^X?XH<P};?6&Sfj6SC;N^ZC0)zmE;sIt$OybMc-k=UbT%k);R$O~q^LV9gTQ
z2)qw%AZG_yy?OKQEo8^~LuUwdsHKgh4wSQO@~gv+Mbr>>6jcKnE1AqejmFNSl96eR
z%1AFGp9=@lB~Ai#-nA~isKaOWF;a_sB1E55bUqv@j*g>g>0E{o+k?s6%kdj0Ynwn?
zXZNj30S3J%#*js5rRRE_z<<y?*h)`QlVVn}D1DuBISvlB6apt6Dng?n=bEG;K6u7+
zh|$G;hX>qpWSo-h5PSu@E~f#gwvH1zV4c+Zv<?9(N#+Q64CuG4i8MP!?1n!QF?q{z
zYG}DU>{_nq_4UAp#(AzkOvVr$<WcQWw);)qzE{*_H`eVe?4-9jy20%u7i5L7PE~~-
zQjj$si4@q;+|JVKA>+qvzS<G7?F@I!0Paw3U>oza<Rl&9$~j{x;NJ<Y;udn$YdM6#
zP3wEUcb*r~Q|)EfWEIJBF#Y#bFWt)?rE0PzR*|RhY4E53Zv1Msob~THzjLV$cP4Ae
zvj%J9)P|{}_pp{~!`;cw8EPp`uL^Jlbg^M8tdc9|l4n+vpNbTGz62&?!82lOaXlVg
zjrdCC`b~4)ek<U)xa>YV-Yvl^k!Rld(sek@VM#Vb$q5v7mjt(5QuxDWe$;@_xFv<<
zAieA;`wZ?VU|CJZFOS^TBh__<;3gB^sTO`6-HfipnyhH3Q>vKV%MC7gCc=Mc9&w5<
zGCFca#d?1i;=*5nP#cjV8H)QM_&z6SQyWGN2X&$Lv)@Iwa)f?L0e*&W|8F;N<%@p-
z*;pQJ0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015qhq
zPZc)Z82SQFPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOf|O3<^TWy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp.E5A6055D6ADB74AC.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp.E5A6055D6ADB74AC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ce4c28b93633a54e30986fe62e455b3f23d9a3c3
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oe9Z<tE<YDlt+e`*
zt}|I;gF&H)y28JGL4kqgw+<ej>sNmMuEHsQh8a1DGv+AmTr(l}anlr!a`E;AZ?olb
zlIeP#qAIMp6BW-~SheDctWxNqUFrLLEz9@6)|H7~k;G}a_2~L7tG~3K`nUJfMlJR2
zP1AN>sP>r@?$wxh=50bk#~!Z)n=3O9#Oa=y{bT9*|7Y6{KdzHo`*`+ORuzFGlNgFC
zbCZCMWhhEbD~8ZHi6FYTD7jc4#9#su5CHTL6fm)Dxy8R@yTh9E%#2LT49qZb1_l60
CWMu6C

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.cpp.3646CB646CAA9902.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.cpp.3646CB646CAA9902.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9cd1a3023851451126c71d28ef14440e9246a549
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8Wka@#f#
z<^N<>uxeFJI->_$ks3{7TazZ^M-Bum#T|hFi;FJ(`YuSxik0*sn)HDr_8egE3l;zc
zfU7Fwqz36#Jy}=#cQ>+t_tQVi+2h^)$60wlAC518x3A>+@?yh$dHc%+Fdkh1uT?A_
zKWE_a{p|kBAH4YT*Q{KAdYYGa%f)2KKv{lREX(D`yU+8dPY>nY10&xfK1`=Qu@JRu
zP}D(c8E7k7w?kh_E6pKL5Xv>J3vy+!+@QB;#43@|f~+!0;}z$xKWS*?r9YfCJlz)X
zEREqe?{3gWUZK@gfsrdTA&cHSpTAr5hgg$cXJH>b7w@@pzLEJ5S&9(QRJ^thRxFW?
z!28e!a<+%nn>X*?LUyb>bb&yNTG~kJNIA<UzdGz#L=9m_QPrTalF1y^Xly+y8JX6o
zjPx?{xo{+1;v_)pUE|`5I^465ks9O^A^N1E^Wj8sbR11XmokLd98Kn4j^8-Bwh5$l
z_RzQ#px1k53|WLmdalPA{0F^*t@I={DP|Rm(yvo4$I+pNLg2(hMQAkST$41!2QPRI
zF}iqY@rYZFj8l>wg0EoL<}?7Ut>c9DSSz(YuS0-Jk~sk$1Ntp%BF#<_yWvknOx|*w
z8X7JSyOt|D9QJIeo#*<)cm&Zw9@P$IyWiyPdqqulW8KcePI{Z8>)k$aK~@N>R8{C9
z1-Zr(kperK+gV!OXZ)1SS34p$t>KR8!5zvCY-665oTOu1IcF>d{5zpl+(M2zEyobJ
zX?@3c&hsLAT6@{nSw*rOO!qz4OLwwIsp@QrRpcpr?ma4i8^2mErrmqS?_8?moyl7A
ztij4SwPEVyJ*=eKaCfqEh8l{Kn*v+`ZLHY}tK`bL<eAmvCn5!(FM;t$@Qm2C7>))v
z1HMwZev@3c-wN0l*WG8QyCrxf^2|G5x)%F6EXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zOD{XhE`vJ?SXSfF%hSmHK(%clxXFZfs)1hzw}Tt8A}eZYl`5tWa)Wc8iSQqq2b|)o
zj1F8;vEJW>xbjyZ)JCL8hT@?QzRL;P)PzCJL0zf+;&+j)oS>glfS=*p|Jw~*`Ql&q
zi5`V>d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000Pyhe`000000000_00000
z0000000U7oaBDmaHyXkMPf!43ZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Yi+01zW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.hpp.58F762D7856C2DD8.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_max.hpp.58F762D7856C2DD8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bd26efe70cdb014c0483c396f8b2fe2861c99b06
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oe1{BpTz)F;QDCc6
zD-xa_<@-j%FY42GkF_gqa5uMG&wV#fU)jBi;lPs41S6h|Jr38Vb~zus(aXDM*_t+I
z74u-ht^<#fxmoJk!Yqs43T#-k+V*nksh^*A-*Z~{^ORQK5>@NVa+MkpzrLS3zDi};
z(adE{w<WqmMf^XU$xT<_*s?)=`)cKfZ7b^HZu-^!%QrK>mw#NfCU;G~it~|K48@hX
zNkGRk6s4vWL+G4D5M5l9T&xeInOL43X#U8yaCZ(fBNH<N6G#pMfZl@vW*CEk0RX3q
BVtfDq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.cpp.1EAEE023D4C34351.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.cpp.1EAEE023D4C34351.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f3ece7947de6a8acd96efd6361ad69571b81136a
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TL2mk<hoV8WklG->B
z_5Y+A4H=dsTgAf+fep2p87h~WN0*S=xC>eG>H<T)ep@yS8IpYnm3^?SK1b?%%LPCI
z;Ht_vsX=;GPuA7byDM41`|021?EdEV<E*@$565S~+ZXbDdA4D`yncNKj7Mj{OBIXz
zFB!OhKfC?<7cahko|Vf_5A*V7xtI(YD9aCvWx4!#^JV_<>8`xFW8`PVhv~E@7NT|y
ziaJOw18qg?cIZoKrP&7xLb;}OL9Psz8}t^9SS2!AkX0sWyx{!hCk?H<^!t;BhwB0!
zr7`^G-3{8vE3}#_Fmi<^WYK%)^LLAWA8WGfEIdWe#e1xrZ)841mLdc+6|b#>6-#6z
z@IJJGob6%t>eaiqkR9s|ogvVomNt?)P|mW+?+!Z_QA5~KR5fU<WHJXe8e5M_My53?
zBfX4#E*wagI0?{t*SPqs4xiY^NDcCd5Pee7`EaB-I*z8Ha~VQx4kmLi#~++r+XT`&
zyKh_y(Ca-hhAcuOJ=fy|enanID?Ld~idn^?^y`$%ad4=i5IFHr5gH9S*CY+`!84vi
zj4tk5Jm8ii<CJ8F;5*p0ISoK->o}o3)=I5U>ky!lWR8HxfPTxGNV8MKZulb+leZkF
zhK9?-uH}jjhdmo==ehnc9zk@FN3}!Q?hkqUSy7YSShusVliud&dbf{UkQKrzRTX+j
zL9X#gq`;2mc9vH689!$8#g2$gYq(>2aEEdO+nA>%C+QGZ&KXMq|4C>Sw~(Vw%OM1A
zTHo=V^Sp?j)?T)CR*@_R(|u3%(w*#4sybU@6?qCDdyfj>#;=x(Y4@J-JD2KkXR?+&
zYp^m-ZJ0WG4=brQ+@0*4p@!n*ssL9&8*8@0D!Fnld1f{FiAcfcOJF<_JR^23hNHpN
zfUi`p-z3-VO#x5EW%t?fZV6tAJoC<%uEl;1OR^zKPM}z~NpQ<0g}-d(R}Bb_TT)o|
z(({h8%ixXzmeqLl{4`_-s%;CwO(wik4g4{<9$bkPSy5A~R587i8=Uh@g#XYy;1pkE
zbl{4L_5LZug}(x!HX=nb6!(4bT~5%ZCJbs0>O$>jzl&_;2>q4<{0`s#-)`W_7yknA
z7ao6ed2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~ybtk8
z8HM}F0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8YFpvsW|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.hpp.94217680DDE41BC8.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_min.hpp.94217680DDE41BC8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..13f5d3bc6b8b59b40b8a814f981b0812fcc04fbf
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wp3KTxh>&er~oT!#&KTt2&A4YK-@
z-jbx<+o3Cc+V8`CaUrQ)%VkaO6yCAroy=0lbYPCs;hZNYQ(EVk>3Mi9oGr0Kq}WJ`
z`SBrF%Z3%)XN`BTMom@N+PvH=qWanGIiEl5zISNfPA$)gTU~cwl1tGE`StzO@l|W4
zaV<O3`pqzK^^(4aQ*R$U@Ssxkc*P3i#$=DYn@99&|K*#_k^BC6(Tj7x*i`I~%wQ<4
z%uND1mZ2y$tr$Y*B!cMTqU2(I5Q7;+KmZer!N36Y8x%0HXuJ(wT08k%GBYC+GXnrE
C0c8;Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp.5276760A8102F515.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp.5276760A8102F515.idx
new file mode 100755
index 0000000000000000000000000000000000000000..868f24e5f3da06d36f2426c5aab0269ff191eba6
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TR2mk<hoV8VLbJ{o%
z_W$WD3JvAjo8d!3p=GW~m|i>i=n>Ky?;uM)y+F8M-<1uap}h}fdLL|S_mTGH*<HB+
zC;(hl87DPJuj<LVdU|&y3wS^MtDN26+<uysxAWonO!4-GJYSw|m@ltipDD(pGsR02
zi~BEGasPgH`}G4azWzNcm!BWz<;`+28M2}*KQ5N#^3%<i`NQYC^5%{;KLb8ar#-L`
zwQEq+L24OjD_XZhUrH;@K2Q+KHLVMBWw6|!w`jyFk<o&zGD+hF=Py5LXyv8fpDa9F
z7w{;J;WzJY&_-UN)l`9zD>NaC-aDVaTlD)_lU--wDS9s6W957!^C7YnA)u*vZ5^yw
zA{&ACp$+70uU4;Ky?YBeu<p>A23pk8Mp6gJS$6r|VaFnB2s?_Z291?W=3qu+>ru(b
zutsI1myyqf1LzWy0Ihe8i_hlpiDQh^AfE`)Cl#F!M~JiIY#KV3A;jk3GWT-)A;`5&
zAg#0e#-#wg-4kNSA~e!-Jx<Dh(L2~mPg0X&RIw=iI^}X40%|A(CLSt6qakxm(hwg!
z!#Tz1;=aX$-g0J4Ne&3UQ@b|P0JOG_3GH>Q)cUjz0V+x6sPGujZ`l%Q4vIJoe*|Lk
zmh;rmaCtbiT+!jMM?>vA*B{0shz|0oc3SrP18+YoYO)*ab{2Nh+nik=_7MxRLRh7$
zLJuj(H6DQ!IMCeA((0b^BbzUFL~L5a9n&j!C^xW;d0H|_hrBXpECu{0p;g>Mjyfxc
z6u4=9hj+pAB6?bT+1A-avYbr!J=IHhvPY@v?1@d}DSYfBDu5fmS}vyDd&civs>7Yh
zR`RUD$~d)Q=ome$q}p(Ia&U$kij%7XTmfyY*$bQG%DLnj)#N841)nd0@ksED*tHmr
z23G^VQn`MUT(>_8cq%Tt&yIIX@Ji&FcfNEj_H$U04N)?IV%;XeEteGjvYTIZAT(}C
zVcAR1JIap19R)0_@#y*CNoZaVRNEGU+e~<;8hA6f9$bkPSy5A~R587i8=Ui0g#WNT
zV3IF%J8(tCmj4vu!ry^V8<8Rziu;~`ju*7434@x`y3qRBFC%+7s(women)KocSmsb
zi*Hd$AbkLHd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zN}MKq-mX+>0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8at)5shE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp.52EC3DCBFC6F1222.idx b/.cache/clangd/index/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp.52EC3DCBFC6F1222.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0034e55c5ebb7968baaa9abebb038841d3470ca5
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%;i9uk%@t!V$Rn7gItFVc$~kx8ux5G
zT&K39bN4K@vX05q{`_5#n)*GDU0$W-OO2%EC1D0NUzW*<4v$$ktqYOiI+mC7z%<oV
zMrr5Ft^gJ>`Dld?##>HJ%x-TGjeR?FUGTZ-_rHp9oed8wRMEXT^V<pMgO~P)+l8qb
z>iQXr-Z>VrW>G~`c%C)0<APS(ow2f<<}P19pDKIxxqkD`kA?PrTbA5*`^~mq@~9I-
zab<22(76mnscFR!IwuiC7Z)WL>w_3Tf4~6HyHLQy@@?J>!GpS=^OzZ#m>HO1;tUJ`
DLmgh1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.cpp.BC429E8BCB688FC0.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.cpp.BC429E8BCB688FC0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b46a273e0e0b22e341ab42380847d6864be56864
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TF2mk<hoV8Wka@#lz
z)&F#0#!Z?{cg7EiYqy?l5>L0CJZfkXvM{DdEiZPwUtdtN>c;Ip)YE+ti38$x08#}&
z0pP02IH^H;RZrH{^V<tqz`N;R<?P}1?&GYyn~$$gfSVWce0g$UzPx;W0$g96054rE
z9zJK_;oa=+%X{8@`FmC_KRwRN+vQ?1W}qxTESBZ+<L&49<EQ)b_MVZSDLzc6y<#D1
z*Py6_)H2Xkv~GvKlvbL3pdgfMS{LNXV7WnW(TG(dqXk)IlEyRa&$l$R^3v~*9v&|X
zc#_8OmrplnBk#~^s=&w<nvg~BozLqQ{XW)Y*I9Uuo}2enIp4@^M3y21G!^fygB43;
zBk(!2ft>AO_3G8z8_0=uMkffgsHKgh4k~9k<afr7Mbr>>6jcoxE1Ar}jmFlal9AOK
zm62XXwhITPOH2Z^-Zd_MbBE8IW26SzB1E55bUq9!u8yl|=v0Ofn?uMv%khUK*EWH)
z&h8tR0`z{58bcPLk)Hc;1ph_vU@JXIO^UUOMd{Zmx8smdLm@EnP!SppnQM}U*ziQp
zB}NzbEgo>ol`$nbA@~k<ZKeTeZ5<QZW3ANsxDNp;NoD{%2J~C@M4FQ#PQwq1n7rjW
zH8k8FPAzwIJnprjcAonW*H;i7<g41D9QOy_e%91vH`eVe?4-B3x<2he7i5L7N>zm(
zQjlvL6e)0`d7P!yJ>vtLFHS^kTEi35gFBQP*v5QYGD(NJGG{CW{3oGRJVK5-FNYF%
zXnm*elIKnIwDz*Cvx{W8m~K7xOLwwIsp=ewUF2K%)Mr!x4}P^=OuO%lzqwV1H<P{O
zy9O)c)P_|<_OOy_!_&#h8EPm_E(&l3w6W$W?2<d@l4q?ZKM^VT{0LlM3BDtCEyh=)
zixEGm+`mcg+aCoy7w5fa!`l+P68X+MU%D3idsvbUQ8IyI-6p{!mlXbTm|qPbG;T>@
z*-PI}lpTXR3Jl$pbon|jJo_WnwuRsk6W*x?-i$6s7h*+L)YK|fOz-6e=X@i=fA}3S
z!DoaVxuRmPe+qHtZ$PMxNRbT1eP4LT2inwxQOzZt>HXwak)sUIZz2D8<o16~0(ZUm
z3Q{N?4s&^JVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015t5-
zX`}zuxy%AjPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOgUD`mH+?%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.hpp.29384EA1BE892314.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_add.hpp.29384EA1BE892314.idx
new file mode 100755
index 0000000000000000000000000000000000000000..205092d6d92d598b93cc231f5f355bae6548a72f
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<6OFHo*x&er~eT!$QZTz<Q<eNp=7
z<59kt*Vp`X+V-XO`r9<_tzK?$*x>V}yRJXk7@B=~nuArQOK#GQnIQD?)gytNQ_rqw
zKR%(lv1!QyU85GkDAB3AoR)Y+TtBm1@%g**y{*aIp*o7o&3;ZSo4PL`a)0=}u%}PL
zE*Od5I;K*&R765<=UX10iRqj@yML){2<LvRRQKX@{pN)+fqKucFO)F9q&`pOh-`6X
zZW7S33`MDF#Sl6t5kwakB^T?17)&4n447D2v|=q)^b9$f8JU<FfZl@vW*CEk0RThO
BUf%!!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp.C4B64A596D6DF817.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp.C4B64A596D6DF817.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa5be541b4f1fa6953b75d9e54b59e859ec843cc
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TF2mk<hoV8Wka@#lz
z)&F#0#!Z?{cg7EiYqy?l5>L0CJZfkXk}#%7EiZPwUtdtN>&ERq)YE+ti38$x08#}&
z0pP02IH^H;)lBN@`R#=);NA4Ea`te0_i<L<&Bxa#z|9MJzC1ZFUtYdG0j{r3fR`>7
z51%ve@NRbZ<vnk{{5>m|pC0Gs?Q$_0Gf<Wv7Rz$^@%Hol@zZ^Id(X(v6d$J3L9q~x
zYf&^oY8hxFI=4k%N-NDSP!P(soeOehuw0|JXv8X!(SocpN#hyz=UW<DdFgjY5094x
zJV|5t%conkk#}e{Rbb=_O~|76&gXT5eis|EbrznZ=jJ_C&bKlfk);R$O~rfbV8s&I
z2z(B0AZG_yy?XWb26AG((Fp<_YH1^>y~<e*`JJ(25jBJ@Mb&`DN+z>+qp|a-WMs8Q
zWu%vp?ZRH^5|aR(cdd)x+~G6l7^y|J2+=1Moeu{USI5<~bSguL?LK6l<@iIAwM`(c
zv%A)%0E6G7#*js5rRRPe!GF;^*ho)OlVYu6QTjUNcI*>sDFh}SDng?nb4}6^8=mO7
z#OUI#!#!@fGNvRa1mD4~%QOJh)-j<2)=908`w*a#WDbDGfPTZCNOMxeY50R8CU3b;
z4K25aQ_CG4j|XjNoag?-^%X=1`KtCP$Nhn~pEWhvwRKwyTj_1CZb<u}3$j93rK&;?
zDaaZRiWE4}JkHYUf$;;IFHS^kJHrz*fLoMn*u;EWGD-WoGG{CW{3oGRJVK6oFZ&XB
zXnn8mljlwJRD0Pq*+sHkOurucrF+?<R85Y=F7hpW8Zs(?2ftb_ru}!u-`uMGo5^1C
zU4xZzYR#%c_OOy_&C|)r8EPp`E(&l3bg|(m?2<d@l4q?ZKM^VT{0LlM3BDt?7UQeY
z#fYC&?%yQ$?T-SUi}T*I!`l+P68X+MU%C#5dsvbUQ8Iy|?vmh<OA3EE%&!Iz8aJe{
z9Heh2%AUb31%_@)x_q4%xv590>k7dmCcINEycu1NF2stgXsA=FnBL1Z&iO`!|L{9v
zg3ky!az({n{}kfP-+)kSks=w2yP@!&545QbqlQa5)BDM<B1btuzlHqYk=y?{3EcJK
zE1?M;O>=o|VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015wwp
zD8vejw{8MYPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkO!(f*$^ZZW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp.615A6DAAC49A37A5.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp.615A6DAAC49A37A5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d0a3c592eb20d7c7450414ab67597006b708f8f0
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%;iA(FHo*x&er~eT!$QZTz<Q<eNp=7
z<59kt_paLMwCzjl^|xu<TfN-iu)*g`cU3>~GO*2-U|a5?XPgqfLqT)t)uW0xrkq{r
z-=i%4lx30P))NA58KIM|h$@91+LdO<YgxYkwXRI`ilnYJxu2BZdj4`s{To*qJ@x9^
z1*v*ldONFBd3pNpl(Ds)*vxV)wpQqdI`?Cxx)-18H!q9{)O&t?;m?f|?HK26Iby_6
zT$!5$bS^_tYFaUb&PfE(#YM@*`XB}qh=2f~m!N=&<y>x#m)WWR;>?Up%nZyhaRvqe
D5U^l>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp.9FC33F8197771FF0.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp.9FC33F8197771FF0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5d1ecb69618fded34da9b108a9f95ca85a1bae9f
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TF2mk<hoV8VLZ`(Ey
z*8eHajF&7+2lzwc+N-f9F}fD{DA44|I+!9gzW8Ooen&~J8@KyVqXPtq_dMjiA5{Pp
z0IsTxlNzK~^<-V`-(JZA-cA23XHWMJA7<sle0+Nb+`W?L%d-vh<@GOT!0pW$@LI*<
z=~D)t-pw99|G|sTf6dC}$LD!@zg$el43y>j#j;#}xc@YN{`gqlKQi(?;{9~m6AMwh
z21OmDmVvgSbvyK>w9*^`1)*Hix*%5u%ME&qMywJUEyya9G+uE2@{@*EUi!mH!}E0k
zFVYx(^X>+1<P};?6&Sff6SC;N^ZC0)e~2~Nbr$x~bMam(=Np+1k);R$O~q^LV8s&I
z2)qw%AZL46y?OKY4zgq2p)&+p)Y3*$1LZ86{OYh{5jBJzMOA~wN+vU?(b#%aGBT}E
z8R=!@b73G|;v_)pUE|^xb+~6ABQ?k;Li9;R=fjcW=s22&&SeO(8BFG0j^8-Bwh5$l
z_RzQ#px1k13|WLmdalO_{0F^*t@I={DP|Rm(yvo4$KX&yA#mcMA~YIuu1Ol=gJ(R4
z7+pNHIN+8e<CJ8F;49d*ISoK->o}o3)=I5U>ky!lWR8HxfPTxGNV8MKZulb+leZkF
zhK9?-uH}l3$2}Wr=ehoHdjruy9@P$IyWiyPdqqulW8KcePI{Z8>)k$bK~@N>R8{C9
z1-Zr}kperK+gV!OXZ)DWS34p$t>KR8!5zvCY-665oTMSHoHLdJ{+-Y&ZXrjVmLUXg
zTHo=V^Sp?j)?T)CR*@_R(|u3%(w*#4sybU@6?qC@dXEa=#;=x(Y4@J-JC|y>Gg(WX
zHCP#^HcTD8hm}+t?oM{jP(yKYRe&p?jWt_gm0UTOJhPhoM5N&JC2)Hqct-46jBiF)
zBfe6(ev@3c-wN0lm)&Q_yCrxf^2|G5x)%F6EXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zOJ8=BT?Tg)7`huC*~tA!wQV7|#e{dNfnP`0qbso@D{5+$DyENegL9sU@E>|doZt&g
zj$Bc(*58G=@HZgTMx;oF;-L?`%Lm%jgi*~QU8w!+SCOq8p`VicpW)m8*$G_r;$MZy
z9svMzd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~FY|`q
zynHD%0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8X%;?0Nv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp.CF7CEDDC624CFFAF.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp.CF7CEDDC624CFFAF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9e41d8b46c4cd625ed62243215bf5d52f7128042
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%*8<ZFHo*x&er~ee1{BpoPR16x&HB)
zvbE9G`|_$x^Kaihq!r#VYl=rt`&+cD^OGpUu{jcM%RTgrr|c<kHM(^5sN$25O_vw+
z2!|)LZd_n%)WTUMJawDX61N-I4n5yE<?{O~?c;K@Qk=KV`{S|uRKUu0_qFp^Z>(H5
zgFF1zu_JpXvbI&fv}SgG63O&<ogJr)NqeTC{g<EnZwBbCU+mJ~`-O91_K`e>;>z44
zpi>!&QqzhdbWS3OE-p$g)(0_|Km-^tvE1qY^Gko<ns>~MOw0^G&%ppQjKRPF0D_Za
AZ2$lO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.cpp.20716445BE6CBC5D.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.cpp.20716445BE6CBC5D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..97119d91a68d377f08ea6c7c25f91d829125f6e2
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TC2mk<hoV8WkZreBz
z)&CS@CYx+FUBC~CYj2Hh5~JH9j{;4Otbr*~%Zo4j^*fa0x^dfw7%d=3oO8&z4^;pZ
z0IsTxlNzK~^>|h7-d@N8-c9~4r;m5{AExE~?DFOWxP2kd7AG5Ki>u#GfSc<R;H8TB
z<EIQfzMI~E{)-o%|CyGHk59AmZZRKUGEkQ9=ZkXj;qKGy>ElCr_rS>ai1(98Ps~N_
z8WeSqS_ayP)@{+3(n_-r6ohh3>w;VwEZ68Q8nH}dv>>ZY(s;)C^G_ODdg=E^4Nq4E
zJWFHv&AS`4kymInRbb>2O~|76&gbt2{XW)YS6SFa&&7MLoNr`4M3y21G!?I{gC$F3
zBk(@7ft>AO_2$jnTgZ-ehfWY^QA-<14V1HN@~gv+Mbr?s6jcoxE1AroMq}$y$;h-u
zWu%vp&xL_>iIV`Wca4i*)ZvbOjMN~X2+=1Moeu|!qvL2AI+Y>BW-yt1Iez2h$|jK3
z*?r?ufL`yBF=P=M>A4<9@IUkpHqw*Sq?lDKO210E9D_p*g}{l2iqL4txh83d51#NG
zVsvre;(%L@j8l>wg0EoL<}?7UtmB0CSSz(Yu0w!Ik~sh#1NseXBF#<_yWtN+Ox|*w
z8X7JSyOt|@dD*j}cAo1GH`fpy<WcQVw);)qzE{*_*Vb(<Y^Aq3y58*r7i5L7OjU&*
zQjjY=5Gk;uxt*oeeZ~*je6b^9(;DuW9^9f_!zSiw$w?aG$~j{x;J*p2;udn$X&FM`
zru7})InRseY2{^GXBEkEFx~f9FWt!=rK+<fR*|Rhx%a35Zv1L7pLFjjzjLXEJCn8K
zS%al<YR%Nadss@f=I&(Y3^f$T7X`Qi+E}v{R>_rf$uq0Tk3|YTUjjGRf@j38#O3wq
zV#HS}*KeHb_D2D`;=KFpaJK}nM4ox)OV?sQhb7q%B_~j<+9bH;lEPP;d94AVaYG8r
zUiz}5>@v8ez|igRG;%jmZCeQLFyWnQ;P=th=t3;Xike!biphgq<BSI){6Ou9^ZR;>
zBUe<c^iLtq{0RuP7AcaUxbMU6vVk@=VN`QOXKFwBQDiFz=%=LqXZZFD`+#d+{0}g@
z9NqwPd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~;5v%9
zx7liL0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8aJ9M^gP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.hpp.A0110584C215CD9C.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_max.hpp.A0110584C215CD9C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e9f0ca96d054951d80b881e0d5d3de93b30ae0f8
GIT binary patch
literal 274
zcmWIYbaUfkWMFVk@vO*AElFfyU|<jg;^LB`%tb)@4^XaR&er~ee1{BpoPR16x&HB)
zvbE9G`|_$x^Kaihq!r#VYl=rt`&+cD^OGpUu{jcM%RTgrr|c<kHM(^5sN$25O_vw+
z2!|)LZd_n%)WTUMJawDX61N-I4n5yE<?{O~?c;K@Qk=KV`{S|uRKUu0_qFp^Z>(H5
zgFF1zu_JpXvbI&fv}SgG63O&<ogJr)NqeTC{g<EnZwBbC_x^IjVg8*(M|g`XbCZA$
yWhhEbD~8ZHi6FYTD7jc4#9#suV8F!kVaB3rInMLy%#2LT3_!2J05goizyJW3-eFDv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.cpp.D8BEF9441B9F4BC8.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.cpp.D8BEF9441B9F4BC8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d9101b8dae05f7ae2e1266e5507ee0bcbb824021
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TC2mk<hoV8VLZ`(Ey
z*8eHajF&7+2lzwc+N-f9F}fD{DA44|I+!9gzS!}8{f?4cH*WhOMhgfM?|I03KdJyI
z09;iWCpAc~>dCr#etRJccsKpKoITv$eVCPZ^YQfwaPvZ*FHbhimzTet0M}P1z)Ka2
zhff)JcsINI{1-1i|1&F>A0OxC?Q$_0Gf<ZA7t3<_;r7$~@#B4Yd(X)Ci1*WJPb@_3
z8WeSqS_ayR*6q-j(n_-r6ohh3>w;VvEH~&a8nH@bv>>ZY(s;)C^G_ODdFl5@4Ud-v
zJV|5t&AS`4kymInRbb=_O~|76&gbtI{XW)Y*I9Uuo{RTXIp4^9h%7}2XewS?2P>Ay
zM&NyD13BBn>dl+CH;^6c4xJ#-qLwz28YpMk<X49si>M*&D5@GXRx+7EjmFlal96eR
z%1AFGp9=%&5+?y#?;01ssKaOWF;atkB1E55bUqv?j*g>g=v0Ofo55u6<@k-0Ynwn?
zXZMXu0eZbh#*js5r004Z!T-=Z*h)`QlVVn}DE&I+atsbN6apt6Dng?n=bEG;K6t`&
zh|$G;ivw;sGEPZ$2)=?{o6`WawvH3pW3ANsxDEj-N#+1}4CuG4i8MP!?1n!OF?q{z
zYG}AT>{_nqc-*t0cAo1G*H;i7<WcQVw);)qzE{*_H`eVe?4-9jy58*r7i5L7N>zm(
zQjlvr5Gk;uxt*oeeZ~*je6b^9(;DuW9^9eaz&7S-$w?aG$~j{x;J*p2;udn$X&FM`
zru7})InRseY3*fOXBEkEFx~f9FWt!=rK+<fR*|RhsrRS=Zv1Mwn0D_OzjLXEJCn8K
zS%Z~vYQxmQdss=e;qGMT3^f!d7X`Qi+E}v{R>_rf$uq0TPecknUjo-xf@j38#rSG;
zG2$zg>o>`D`=fy8;=KFpaJK}nM4ox)OV?sQhb7q%B_~j<+a$Q<lEPP;d94AVaZ3uz
zUiz}5>@v8cz|hU`G-OArZ41F2CcIM({64xIU5FJ~QB$i_F};@?oby10AE+I1eqV2K
z<cf-w{wc(nKLMdOB1JM3_kGx1HqfRfjB1YPOzkH>ifrWo{gl-I4BviXA8^f!{{bmm
z9M}MJd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003Q9$(p6%=T3l>$#t08an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8X*(8%Qg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.hpp.8F13163D01C8AE8F.idx b/.cache/clangd/index/device_reduce_instance_blockwise_i8_i8_i8_min.hpp.8F13163D01C8AE8F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..30a528ffa35cf83f16e1714010a703b147df67a1
GIT binary patch
literal 274
zcmWIYbaUfkWMFVk@vO*AElFfyU|<jg;^LB`%tb)@4^XaR&er~eT!#&KoPWE1E`5{U
zGrMASW>ULF=F9v1w{qXy<E@z}aOrcw^dsgBGjxo7{3gbTX>2QU4f2}Z?Ytr8Or}qd
zarje~MT%cf2&BF6owS8l$@I`R(S1HkdG^0vD^s13)V0R;_snl9$xC<DuHUov*sHLM
zT;aElyX>9F*5>|_o!R+`E7RkJcAPSkTPsEFzy92RGhp5N(~GQM2;Z4^guA#hHwoxa
xhN9H8VhEj+2%?LNl8g003?>i(223o=T<UKA@(mARW@KV!0D27um|+YC1^{}wVE_OC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_impl_common.hpp.F2201C19BAE9FFE8.idx b/.cache/clangd/index/device_reduce_instance_impl_common.hpp.F2201C19BAE9FFE8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c3b24698cfe43289e475d85e1a2e935b6719fe4b
GIT binary patch
literal 1688
zcmYk54^WeJ7{`BmaW8|7{evMdOy4mWjCUJoQmjyfI1#uIONS7`1qetoBt`-z5Q2oV
ztOFqru~6@d0)f!01|djCgcJ!2gOr#c5F)85bR{1B-Zy{Z?)Kg8`}@An^Ssadd0z_)
z3HcBQAb3w$u0A<CVHE&C1iyytth5pq048YU)}DVgF0y-_!&v0aqOFNmxBC-Ah4DsL
zk8Sz-@SFs{x!*<)$i&O-Ny*7%%}I+n<7bmUsn7J=Xbe7aAk)Vo(ru}vdGL+glRHQL
zQVe!Yn3P8LeBa!$Lw738UHzMW2z}sZiJE1n*!_eLiVjhCx<~kLn|WfJTb9qSrx-Wr
zO7|`hb_Gon_X^6*%kx-<sdB7nY(%q}Y?zuJOtT~n)y<E*)PAHO>Vv#3&JRx=Z2WRK
zBBD68jLW~EcvRo34TwlhKROiQSXHYseIIdV^nPG(r>N4Rd#>Y?^F|lPZ+e_=NIUk$
zMzYPqd%iqP=s92Ai`paWj0v$Gf_`hi>b}W`Gd<%&@!spb6xWn8_Y!x?s=~=}Lvd_H
z!<xNU-0j~fF3qf(y3$ewY7SS+IN{#MoH&tTeqi(#o5dnRV+&{bcD1{jDmE@#&vc&+
zZ1-$y%E{XPuh^<c6&H|dt}kuVrRDeNW++^k^Z1w3OL&iZ^~LP|Z3~^O_>8VNL!LhI
zWefmLOTuS4*WE61G0AEmRpBZDBftheF929l7SA*-{WGjBT7p><KuQv%7bAef1ORd7
zhH^4=oGrmPPC)8FIH(zUdVITXe|4)%FOG>|dA>EDWCYk)K@RBh%F)?_tx8KA7ju9C
zHLeyhveJLon;+}QPn-O?m>mo#aix%vmB4PYMM(08=C~Li1{AnLz{oR!iLI6&*Z&?U
z!Psy_F(Gzi1dVv~4|9y^qnHK>W&>|fPRJ=n(DCuY&i1d$g%2bc7aqG3t`sBaxHkIr
z(fyo(MH~~u?ow+h$q2An1(!?TV10|<PSKi}TrMVn0Z-ghz{tvY1Kzf$?l$kRmSDDU
zJbS`k%?N5RtPXhZu9cc6!FaI0jF3@`pyOXBgJt0(d-}MTBP>thl!%d)^3xMnX9w~p
z6kN<61~j;a&&Wz(e$x(0SLnY_g5j{iT4Jq|5!B$Hs44yCSEx!ZCWO~};2r`-R<6%V
zPBG}8A&F(n5cGccqq@eM{xZZ5v}t(Iz7RLirlBF8a6;7oo%h0}(6cO_R7on`(8EW^
zFikUrY9n}{O_PIcZEjh{iFYZw2&JWh(Tal7=fpuv&0a9=!|@^h=NEtw-}{yFbE4BQ
zt=6i8QBBkurfKL<ZA2TiY1E$DPZ)zX4H%j?URd+3?d9_CU~Lc5LymU&*6^jTk1w>w
zLEDq`lp{nCiLeNbAzBq85ZW|;$VR`=rm;hlpkmObVMBI*wV^}f9$f@ms<f&g^x7l%
zV45Zm&5fXgHccFAb<MMRd)XH6FJKXkRuhQOLJNjz8bH)?{Q2DzMHi+pm{yPqIoce$
z3Z`iqQ5s<gZJI&U60Ho{G<zrwvXavi=qG5%N;0_7e_Uu`KA6<F`1tDL<wgPK;tZUT
G4Zwe-ryjik

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add.hpp.F65D86E6C84EDA60.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add.hpp.F65D86E6C84EDA60.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a457c7c5b8a96923a4ba21c1c79a09e395fc6fae
GIT binary patch
literal 2304
zcmWIYbaVT`!NB00;#rZKT9U}Zz`!5~#Kk2=nK_IM3~H<l3>9;(PCo6$Y$($9JTOi1
zl9S<;R1MaJdb={c3+4SLU(TFy$HOR?{o~!a(b2M6dM+1ZKg~b=e`4{Zp1inaCN7Sd
zoi<VS(FvNCd4VD_<`>UTOJfwAr`10BOXKI$sj){+r*f;!U$WY%yp?-p_FCI@vQ-_I
zW_j*1>E|+?RWW7ZwcB;fsgKt-^C+!5utjWLhqv;FN<062*Y0C$k}tJzv?W)_sxSWU
zZSUv6agDtuyMoUsc7;aysZVnpw?q|%sjJVK#(7?|H>2rm>Gpa3m)7wvUCuT|t$O7I
zAt$~tWlI(TffUa<wu{me96}x04y`RSJK(2y`l57*i_|)<#+K(P8O}!G6V7#dycPM$
zQJ=Wa+T#1+e{0i~zdf|_Z!W%+a^En#==s6$_lka(1fJPw7tMQXaOm5GH>qD=yPUNy
zbyl4JCfwLAY<Fh%KZUzj-{xEX_Q^XNm*#rEvg&8Z{F*tdQ@_Vgn;O>jg!hT<H<9AX
z+@xe+*cXV3ELP$2@#JA(W@X^=W%V@$6AYY8j10Uij9iSX*9ABm_UazyX5t1a5MUFq
z0TT?IK+OycKn0gqKKkGL$HJ7Gi3g}ao=x5cOdu366s4vWZ((C#5GqzTZYmTsX5!(B
zF^VyDn6=V%?)xuhJX}d;NtTBt`-9m;PrFIJT97|&lk{;89xh=PVHue9JX|6yB1k+@
z7EvUg7>gJZPn<;@i6_A#fy9$!kwoH2u}HyrGJzrUKYaQ)6R+!J(*@^TKc69pSCLYg
zeC+PVb>4UtsUO=QRwEEAk5`e#bi-#Cw)4!ut4M3nA9psje-d~VSxl`d<(P0N5w9Z0
zoYz8*f7WX<@^Hzq%SnJ^7)Us_zhU_KrQ{Y93s)>ptP?EWSh$h}lHFkO=4z&J{n5^g
z_+1fr<?9mGh))afx+26+Nd3h`8&;qz{5bviVJS2;owIMpbfXZwdShPnZMfOBDhaQm
zjJ>XGiv*p#@G8nR$;}9zeJ>lYBADA4xxl%m_t5VJYaUuFRDJo<CwP$KEMC3M(=BAr
zFxJ%ry_p-Hn*a;P_Npq==A@;ac=h%)TFzQs_$>>lC@LZ<7Um2du7sF`L>O<{Nsr~1
z+v;?H7F6k1S;2(oL`2=p=Q#I`m4{1-MM(yheR;T)S(K4@Dl95UJXID|B%T_J8l1QA
z@2)lVJ5T1|b?*}A+p<oFe$7U6%96>Zj)#<aUxB<K5+w(7A<!EV2}&>?tOQ`@0v9~6
zFa~Ooj|zn;;z3E7t6w))Mn0U=2Q*i}MZgng0uPtGi@Yby!&~gnF{-arQN-);UDf=V
ze0rxA1CxFvOC%r6Iu@>Yj(8y$Z?}!2;FFGTy+D(Y74e`b+AE~k;5tG2I$rDc$F|g3
zoUmZWtLX5GfIMC|juy1QJT~8uN&EZkA3)($)l_R(dOCR}S|Y;oI3r#oPggZ=o!GQh
z8Ly&?*Gj{+6^_+|6p6+t!-@f5@hzF84zmsxy1>{1RRxIHN>EOSgasH6O388OX8MXH
z4a|%1TK??9XWNbSi=BYJ^Aqrshne)2|FCJIh4y*8dVlO&*7{=k53s{=);qs{FWaww
zXWAE_1-Ul4p)mgxrRF4p%E98I<YH?OgB?VG0TY+A*<`czca9dZu(NS+LMuG56tJ>`
J0bqLv1^^~UDZ>B&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp.5E6D995E61ADA8BC.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp.5E6D995E61ADA8BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24fe3dcfc0343262d89188bfdc0734cd07019bd8
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH830{{T*2mk<hoV``sa@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$@POC|1&ksF^;H#GVCq&)&cS
zpa5`nZJg8~y=tcG`uXjZEa2Vjk8=KSd-q{p-Yv$HGvLh&d9ga%uvlHcJ_9B<XTVDn
z%ZE=Hcz8Fz`}{jEKL0r{S05i2<?U)Y9Wzjt@0Y7`_2Kr@;_>5sd3(>u_Z07Evq7;G
zjcZXfL24OjD>}DBUrH;@K2Q+KwVeyHGFWcVTQs6dWV9fwOwxG4`O8lls=W03lZD6Y
z0-mHX{N~*)+Q=)knrbkzLKCv+z4Q6IMZb>?*>x74qvzs1)y}svA0kT;0-BoF)<MM*
z*$BK3Z6Id{SiO4n_6=lW{h>1iI@HoeQU{f@?DDI_jz!cEb`(_u8f%%%!HmYvqmq%;
z8kLb=Mm`q~N|!hZ(0SLo_{AJPGsj3R@`(_AQqlQvRB?11O-tu8gxDTj=3b89kX+jY
z(mK0uT?#PRJ!uSCgjRa4#|iu!y@RdvBsD44Di)<*r(BK$p_W46#6wMJG~`^9G{gtb
z^c-Szao^zqw;UO#Bol(KVAtg|0IjX#gbr9IwLYywfLf9{0v-eUEn6bZq=;$wqar46
zIZh2Nmxrn4ijK#FHZ;z2{b6zg(Lo;79%aAZ<n4P!O?G45&caT5o1+_OA9X?22vw?U
z^pJvF<57_U6V2@`tsXLd%;t+75!=pi#|+>O<p#DfPfJeHA+DS=mID5j&?;^rN4=Fp
z2;8*3*Z1Uk5k0NF?3!#MSq`TEp6aE0*`riV_QWRg6g~}%3gE`CSIb%dp7T4G>TqYW
zl{{-u8K*X^I<kjKsttE1lQYy(oL&{+3g}|PUf3j8&Lz)UO@1m;@c9y$+z6f#yB6b{
z(bb5rRIcAN*X`E=o{P)wv*X<oyb^iloiAO7!yJ}mLzJ9AvF?)KmP-nM*v*eR5E{3n
zupFc>JIa>(i~2N{GP;&j$(>(T<4IXxzwRHFN2==z!3`(8Q!V^5x*lDLimYg;Q>vKV
z%MC7gKEl6v8F9M*Gl7vSYWDG)5EuR;gxZJ{$xz%6N$GQiHnm~YaDW$DKl_bjFGuL7
z7v^X1{Qo+JyQBCAII1yM0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_
z000000000008jt`0000000003QP_mI+xUVp%mPnP0Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOd#CRuK)l5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp.510427928F98A02B.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp.510427928F98A02B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e1e0b187bceb6363df8dddbec032d32f3fc53dac
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%r!t-oQZ*<V$RjRgM5b!c$~kxx-=YV
zoZq?0EcxQ4Wm%UO{<Rg#=&+8l_>pjW(|v=^PofNI-aN&NJ$%D7%AR*k@!YP)+q5*P
z*DW<>M_SkfnJw;Z8yw05Kk_PBF3fM`FV;N#zD)bLT#SzL^Le+-Uuz{V-Br8q&BWro
zu!~&bRr5Ua{45Rbc-0#kxWuU%?7OqOapmgME5yW9wyQ8|9;<(N<YUFbmo-&c&+8b1
zWX?$ySLP-Goy<^_npO;<a}q&xaZz%yK8OMI3k)#B7z_+R-$MZti*MUC`NqflJee7p
Gm>B@#%wVqo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp.54085188C8447423.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp.54085188C8447423.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7695580feab1bbf91b84a8f45d3e1cb99ab50ebb
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH840{{T*2mk<hoV``sa@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$@POC|1&ksF^;H#GVCq&)&cS
zpa5`nZJg8~y=tcG`uXjZEa2Vjk8=KSd-q{p-Yv$HGvLh&d9ga%uvlHcJ_9B<XTVDn
z%ZE=Hcz8Fz`}{jEKL0r{S05i2<?U)Y9Wzjt@0Y7`_2Kr@;_>5sd3(>u_Z07Evq7;G
zjcZXfL24OjD>}DBUrH;@K2Q+KwVeyHGFWcVTQs6dWV9fwOwxG4`O8lls=W03lZD6Y
z0-mHX{N~*)+Q=)knrbkzLKCv+z4Q6IMZb>?*>x74qvzs1)y}svA0kT;0-BoF)<MM*
z*$BK3Z6Id{SiO4n_6=lW{h>1iI@HoeQU{f@?DDI_jz!cEb`(_u8f%%%!HmYvqmq%;
z8kLb=Mm`q~N|!hZ(0SLo_{AJPGsj3R@`(_AQqlQvRB?11O-tu8gxDTj=3b89kX+jY
z(mK0uT?#PRJ!uSCgjRa4#|iu!y@RdvBsD44Di)<*r(BK$p_W46#6wMJG~`^9G{gtb
z^c-Szao^zqw;UO#Bol(KVAtg|0IjX#gbr9IwLYywfLf9{0v-eUEn6bZq=;$wqar46
zIZh2Nmxrn4ijK#FHZ;z2{b6zg(Lo;79%aAZ<n4P!O?G45&caT5o1+_OA9X?22vw?U
z^pJvF<57_U6V2@`tsXLd%;t+75!=pi#|+>O<p#DfPfJeHA+DS=mID5j&?;^rN4=Fp
z2;8*3*Z1Uk5k0NF?3!#MSq`TEp6aE0*`riV_QWRg6g~}%3gE`CSIb%dp7T4G>TqYW
zl{{-u8K*X^I<kjKsttE1lQYy(oL&{+3g}|PUf3j8&Lz)UO@1m;@c9y$+z6f#yB6b{
z(bb5rRIcAN*X`E=o{P)wv*X<oyb^iloiAO7!yJ}mLzJ9AvF?)KmP-nM*v*eR5E{3n
zupFc>JIa>(i~2N{GP;&j$(>(T<4IXxzwRHoU5`}P6@nX1c&A$UWpq8d5*1m|P^VNe
zyO$eW@O*@S@iO9c|7QXtSJdp|Hz6+kMF_PKDUzYMACl7N3T<k`sNn!Fw0`y*$zG1o
zPcO{R;Q9Y`40lKI4~+FOmUDS+VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_
z000000000008jt`0000000003QBFy{DGw)GfC5iY0Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOsn?K00000

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp.0ADD6329411B78B1.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp.0ADD6329411B78B1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a017998f92fe5d9079b029700c98feb5cdb2183
GIT binary patch
literal 286
zcmWIYbaNA9WMFVk@vO*AElFfyU|<jg;^LB`%ymFooQZ*<V$RjRgM5b!c$~kxx-=YV
zoZq?0EcxQ4Wm%UO{<Xa%&|w{8@gw2%ruzn+pF|nbym^Wjd-#TFls)g7;<;Uow`plo
zuUl%&j<m1|GF#l+HaL_Ae&kiMT$ta=U#xlfeVO)gxfmVg=ksovzt&1#x~q2I9;N4Z
z)+`cTw`IO(o}Z<`9j|&L1D7~e1G~Gs8&|GAy+TY(WxEQa=CS&RM?O{@d|7kL^>yy^
zI)<P%=k$v!bCZCMW++NcD~8ZHi6FYTD7jc4!~psR2AE+C1_q%2p@4~Hl1}+^C#7m(
JW=1Av1_1AVVz&SQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp.4393525F2F4E8DA5.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp.4393525F2F4E8DA5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..26590678654249a2db9f82c01833752523cf9307
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH830{{T*2mk<hoV``sa@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$@PuC|1&ksF^;H#GVCq&)&cS
zpa5`HWt`L?y{f01>iO-JEa2Vjk8=KSd-q{p-Yv$HGvLh&d9ga%uvlHcJ_9B<XTVDn
z%ZE=Hcz8Fz`}{jEKL0r{S05i2<?U)Y9Wzjt@0Y7`_2Kr@;_>5sd3(>u_Z07EvtF?j
zwQEq+L24OjCt9~hUrH;@Ay5#?HLVMBZLr*;w`jyVk<o&zGD+hF=Py5LXzisxoGd(E
z7w{yF;WzJY&_-UN)l`9zYcwH?-aDVaJM@QGlig(DIeISMQ{{Xk^C7YnA)u*vZ5^yx
zA{&ACp$+70535(N-oAlMtUGjuK#N-1Nb0C^mR){z*s+Kj!k(h4L1QJ8IhxVfdQ>v9
zTB9=3%gE=#QRxyV0b1`G7r&UpXXY5GK|T?pPbxYe1{Fug(KK`}Lx|1MW$xwp4atp7
zAg!~9#-#wg-IK<UMQEhwdYr(&(L2~lPg0X&tzuF7P0Hms5^5*}PCQhEMnlduNke?_
zOwS=k7Y{8Sam$f$N-`n%3U+Ny1JK4gPH2y{QtQ(?1gIpL0q_{m@7NM)CPhrc4~m$)
z<v2AoTpp&DD>@$c+E6>s^@qs~L<f0PJCyx?leg~`HQB9odkcH%ZH}&|J?Mh05Z0-x
z&_fDxgM%UkCYsw>THR-S$mWY35u4U<$MoPH<ra1^PfJeHF|M35mID5j&?;^rN1c^p
z2;8*3(|6=~5j}0ZZ0l?ySq`TAp6aDL*`ri-_QWRg6h8Hg3gE`CR?AuUp7T4G>Ud|e
zl{{;(Hco9>HL!=ZR9o&&CTFOjIK3*s70||-y|78HoJ*dyn*3Cx;PWLgxe+`gb|c0&
zqpJ~Lsa(HluG_B#JQtVUXT#kRyb^iloiAOB{T!BLLzJ9Av1yawmP-nM*v*eR5E^%+
zu<WHTJIaRpi@G$HGP;IT$(>(T<4IXvzwREEN2+ZL!3`(8Qw{twx*lDLHCa(pt5h+&
zms?!$e1w1TGU9aqX96QvRP5t7Aujwy2(=X{lA(C$lhWl1ZEC`(<^V6We)b#5UIyr=
z7v^X1{Qo+JyQBCANmemq0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_
z000000000008jt`0000000003Q7y^nTnjs=tO8F^0Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOxkA5Z~y=R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp.21E4CBB422BF8D82.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp.21E4CBB422BF8D82.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3149ea0aa424b85effaf07e00489c3868313d9c1
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%r!t-oQZ*<V$RjRgM5b!c$~kxx-=YV
zoZq?0EcxQ4Wm%UO{<Xa%&|w{8@gw2%ruzn+pF|nbym^Wjd-#TFls)g7;<;Uow`plo
zuUl%&j<m1|GF#l+HaL_Ae&kiMT$ta=U#xlfeVO)gxfmVg=ksovzt&1#x~q2I9;N4Z
z)+`cTw`IO(o}Z<`9j|&L1D7~e1G~Gs8&|GAy+TY(WxEQa=CS&RM_yJOd|7kr>!Uh`
zpqz7Z#g(~9KqoU4rKS}_=$u3lU0jr0tPf%U{Q?8bFa`qy(DzWl#ByY&$h~iT5!0C&
HnV1;>^z35c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp.4BD601E4D7DE370C.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp.4BD601E4D7DE370C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..95fe520793d28444b817e58930e0d7510884d5d5
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH840{{T*2mk<hoV``sa@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$@POC|1&ksF^;H#GVCq&)&cS
zpa5`HWt`L?y=ta)_5Aiq7VvKNM>&7Ez56gP?-t|98Sv(XyjY!WSgfvJp8=DbGvK9(
z<-?~8JiMFVeg2&npZ}bftB;S1@^-bHju|M+_sdne`f&Sc@%Zt+yuD}Sdy4n7*`Qd8
z#<eJ#Ahis%6`k9mFQt`cA1Dar+Rg>JHdt=ZTQp*w$Y?=UnWXW8^Ov7AwD!{PPZl1p
z3wV;o@SAtHXd|!CYO27<HJXq`@14)zE&6?I$kthSj-HG6R5{<ue26SX2xux^TL){F
z$VT9OXahMr!0Oejw{IX5>kpkF(4m$#k~*lIWtU$ab}XWXu%oCN&{)Z24rVlV9+ix&
z)~Jm1GV-}_P`boPfX=(t#V_XYnK?#kkxzu^lZwuVql%;BXj(d#A;k9JGWT-)hGcCM
zNbBsrbt%AL_oOjo5nAcF9w+c`^bWStlhmYGt5}r2PPrThLM?^BiHC~NXvn!HX^0P=
z={dyc;=aQJZaFeeNhSne!LG|`0IIFygbr9IwLYywfJ%}%0v-eUEn6bZq=;$wqar46
zIZh2Nmxrn4ijK#FHZ;z2{b6zg(Lo;79%aAZ<n4P!O?G45&caT5o1+_OA9X=i2<uc;
z=phAJ<57_U6V2@`tsXLd%;t+75!=pi#|+>O<p#DfPfJeHA+DS=mID5j&?;^rN4=Fp
z2;8*3*Z1Uk5k1vjc1<>sEC<tnPxaEh>`|&FdtwuL3ZDi>1#sh6tL3bJ&-tB8b+|Lx
zN}e@X8>cp`I<kkgR2%M2CTFOnIK3*s70|_oy|78HoJ*dyn*3Cx;PWLgxe+`gwie@?
z(bb5rRIcAN*X`E=o{P)wv*X<oyb^iloiAO7!yJ}mLzJ9AQFlpj%O!<B?B+)u2#s4(
zSPs&c9c9b?MSU7e8C^@N<jyav@uaM-U-ysP)+5z*h2Vx0-l-OT8C{RA#G0&Vs8gz#
z-OCLwcs|0vco}iJ|1*J+D=PN!n-CZNB81wA6v<HB4@v2Bg*LTe)Np_oT0i@ZWG_eP
zrx)gD@cjQehP$Kq2c4lYqH}p|VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%00U77
zegrxps|&yaPf!3)0000000000002+`00000000000Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOs%%um;e9(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp.619ED16F95775BEA.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp.619ED16F95775BEA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3fc5753205f0fb65114d8157b2d2ce85397e1490
GIT binary patch
literal 286
zcmWIYbaNA9WMFVk@vO*AElFfyU|<jg;^LB`%ymFooQZ*<V$RjRgM5b!c$|MKDLFhk
zaNZ?ycJ9SX%d##n{A+tjpu;-G;zz>iP4^Q;o=P*EnI`efZQ`so0dLNUYEI7cm16Ne
zD#d$hN5R=O3YHoDJc*6(7Cn-7DRR2cCI38#Z~yJJGSvked*)PU-CwhsXX>kOKR<QN
zDc;-_w&PNr+D^--2MnhEOG)V1<CRdetDHGFe7d8qjz^v+Q&7+UL!FObc0GQ&r~6vw
z^E!s0Kj(~#D|3^8j%FxIO)G}bIf)>;xG1?;AH)Fq1`3#1N>AqfS-J7>TxLclW(H=M
LI0FMPG++P#XEJI&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp.873C6FA8487AF602.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp.873C6FA8487AF602.idx
new file mode 100755
index 0000000000000000000000000000000000000000..973d3cf99414a0407ad6519d405b88db513bd42e
GIT binary patch
literal 1070
zcmV+}1kw9aNk&E{1ONa-WJ!2!WprT#0000H0001UbaH800{{T*2mk<hoV``sa@#f#
z<^N<>uwvIuI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$@N&D^}8nD49Hv#GVCq&)&cS
zpa5`nZJg8~y=o?_`swY3Ea2VruX1*ObNgXd-p<F@C%~H*@_c!+VZOY4eF9uxod7RQ
zEbc#L;Qrn0_Vb^-`26>*Tz-6*mp9ABWXwQWzF#cM<%gS3^M{Xj<;@)<-%`AvP6x$8
zG_FO_1gT}9jp*DKeJQOpyFfuG*LE(*%3!%hZ_$V<k<o&zGD+hZ=g&WBsPfY9jusv+
z3wV^q@SAtHXd|!CYO2A=3QfqO_s-|<2K_EJWLH^uik^%2SUca!e26SX2xw|vTL%?O
zWFzoCw1J!*VD;+N+c%Jj^@mOn=uk@=N$pk6vdb?HI~Gwx*iuvtXsl&2dovn4k4i>X
zYg9&h8Tnk;D_!CwK<8cS;<Gt?Vvdno<P#zKq@wfTpyKE_nwCyw2(jI}%)K1HBDt~&
zq;+=Jx)flrd(;@R2(9#7k0bamdIuZnNorE8RV+%sO1T_+LM?^BiHDldXvn!HX^0P=
z=sCpb;;zFzZaFeeNhSnez^=<_09skc2_3LbYJFUX0JS7@06Yft8@5E6NfFcV2SrTY
za-14kE)P@76&;TUZD^e5`or}VL<f0PdzAftmA7vdHQBXwTMJw1ZH{iBeb5D2BUGub
z(L)Myg$G3nOf<K%w0g++A)7CDL~J|59W#Jilxx_;JS{m%`?zw>SPJ-0LaVri9Q9WA
zA#l_BUf+}FMf9}tvTL%5WI34rd#so4Wsg!d*%O<{Q}{SADu5fmUM{Bnd&civs{NhG
zR`RStWt>{G>cAc<sn*<`OwLeCadJ_BE1-)FdtsAYIhQ<ZHTj81!RJfh`bzMO*p(Px
zjV?xfrE>iyxo*D~@Kl_4pB?U&;FZWT?|kVx9Okej8=~X{idB~cw_H;A-EO|ufzY@i
zh2<bU?<iaDFY41+%II2BC3k*VU%q~RSRSdaD+D*3@J_Yx+vsw1Au6(>p-!n{dMDR7
z=lKZ#<7LF@{$T<mSJdp|4<XL{MF_PPDUzYM8<NuJ3T<k`sNn$5w0`m%$zBf7PcO{R
z;Q8x1hP$Kq4?z+!U2}PDVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_00000
z0000008jt`0000000003QM+t@q6@6p?*dOy0Ap=rY-j)g083^~0|W&I2M7rZ3=I$u
z4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K
o78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOfkyU761SM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp.DA5C10F93175C08B.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp.DA5C10F93175C08B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ffba7fd45cfb6c1d5cff7175e8fe57f998061b98
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRXdoQZ*<V$RjRgM5b!c$~kxx-?1K
z3#aBDTeQ^o>N4lw)hf$;?>szokUe_Z-!{)F?->lDlV(V}W^HoXs2;w=XP0TFfR1Rk
ziPqyorkk6dEU-0dVcpf;l`n9TvsSxirnh9>_Grs*50WG;zhBuG87`@P_3h`U9dn8|
zcZKb^^xLPX_*CP=z+dMM2vklzuw(b;#+9p2uMiVc*{;H<d9427k(Xt*r3L#Lg1pYj
z6j$aZ0iDZGl$uryp>q;Jba7E~u|9|a^al(u!x#(<K%YYa6N_@X-{lhT<QL40Ow0@b
DOT%H;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp.A62B76883CD2E814.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp.A62B76883CD2E814.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b92535801bc151dc087ac8deb744f5834fd97925
GIT binary patch
literal 1072
zcmV-01kd|YNk&E}1ONa-WJ!2!WprT#0000H0001UbaH810{{T*2mk<hoV``sa@#f#
z<^N<>uwvIuI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$@N&D^}8nD49Hv#GVCq&)&cS
zpa5`HWt`L?y=o?P_4M{a7VvKRS2?@Cx&1IJZ|CFd6X4AYdA>Z^FkfE2J^`+;PJovt
z7WbbraQ|+0`}t2^eExe@E<ZlZ%bVq5GG?GG-!GQs^25!i`NPM%^5%|_Zz<kSr-Nc4
z8rPy|g48n5Ms#kAzLZv)U7#S8YdaU@%3!%hZ_$WVBBKRaWs=4-&YyqM(8^1{J6d?S
zEZ|WZ!*AZ*qK&*ltEmDbS7<^Oy>~u;H|Te<AzNqRDS9s6W957+^C7YnA)u*vZ5^yw
zA{&ACp$+8h0IOH8-oAlMtUq*uK!;k|NNTTgmR){v*s+Kj!j__HKw~A7*_+YWc~mm8
zTB9=3%gE=#Ug;7i0Xpwm7oW}H6LXBzBA*D+Cl#F!2Ng%h(X@0bLx}C(W$xwp70KEr
zkk;8<>r#Nh?ongNBDB(TJ&xeN=pAgNC#gxXR<S63opL$$gjx!L6Au-k(U5aZ(hwg!
z(Q}B=#a)Mc+;U``l1vD`fL)i<090GY2_3LbYJFUX0F@+j06Yft8@5E6NfFcV2SrTY
za-14kE)P@76&;TUZD^e5`or}VL<f0PdzAftmA7vdHQBXwTMJw1ZH{iBeb5D2A*@nW
zp@$S?jR!>vOf<K%w0g++A)7CDL~J|59W#Jilxx_;JS{m%`?zw>SPJ-0LaVri9Q9WA
zA#l_BUf+}FMf6mA*)`cjvK&nRJ=RP2vPY?!?1@d}DSR9l6~K*OEf>@NJ>z#S)&9<8
zD|yynWt>{G>cAdWQmwf=nVg}P;^d+LS3nmV_QEE)axQt+YVs42g3p)0^_Acmv9%aq
zjV?xfrE>iyxo*D~@Kl_4pB?U&;FZWT?|kVx9Okej8=~X{in>dJTP`X5Za3fSKxo{M
z!g7$Fca$yn7xif@Wppj6k~_bwE?+-Ca#N2~*A;>rPI#wU_-%AKx)3X}qM=TyVtOan
zIOq8Y|KnxE>Hc8?BUe=H;}0Ru{6z@07AcaUxEqqv=L&6V!>Hi^&$NE>8_8Y{&`&SS
z&*1s%I)=NW_z#s`F`NK%d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_
z000000000`000000000000U9%_@e>f0V^>APf!43ZDed{0000>W=#VG1qKHQ2@4Dj
z5DyMX5k*5L5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}
qBQ_Qr933DbA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Z=BhS46

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp.82487B9F10609357.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp.82487B9F10609357.idx
new file mode 100755
index 0000000000000000000000000000000000000000..06f0becd2223c42169b24ecdf49ca709b3464798
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%vC^IoQZ*<V$RjRgM5b!c$|MKDIM&c
z-?_=`_M)Y}SC={eu3qBnbLZiqgY41M{^~8(uxCivcErHE>q<&%lK(o7nI+km6e6^)
zWds~+%}!%I=~$h@!CfLQdS79Z@UO5RR?~}q-pzY@qq#fv$(`WeX|bs*LbjiuUjAgx
z?9eTTrhYR%c~+CXZRMBdM#WEBjUNlrnS;ZpJL>9q<asg$_5457`S@>H*R{;&b&Nr`
z&KVY0<|Y9h%utk?Rt%wY5<zrvQF5_9kY-}pyyBdS&Xf8T%#2LT3_yRu05goizyJ&h
G7ytk+@MeMl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp.979C0C60ED1AF7B3.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp.979C0C60ED1AF7B3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24d95cc28b6444c01ebc48bdbe8d0825e828886a
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH830{{T*2mk<hoV``wliD~8
z&i~U9Ub>XFH^aj&g_gPQF4JpgA2Y^Qhz7^b`3Efh>rYOgbZPIQT<2b5OR}U-pJb;1
zC;(hl87DPJuj<LVdU|^$3wSsEvz*=E+<u&uxAWonO!4NKJYSw2m@ltio+-woGsW`|
zi~G-6asO_1`{g}vzWg;Sm!BTy<;`+28M2}*KP;Bz^5f0t`NOBX^5%{;KQnxoPJ6>b
z)UH8M2dQPCt!Uj2eJQOp`#?b`*R(FkmBDg@-l7q!L`Dm;$|Q{!_+Rd6Xyv8fpCUY5
z7w{;J;V++V&_>>&)l`9zD>NaC-aDVyE&6?|$*!~T6g@ZZv2wnV*@-Mg2xuzaTL&wa
z$VT9EXahOhtJRAaZ{I)$);XPNphYchBy})3%PGHeb}XWXu%oDI&{)Z24q-I59+ix2
z)~Jm1GO}Md7+qo#p!KeC@mn~2VvLa*WRDPiQqlQvG;wuYO+)80gxDNX=2?zEfLz-I
z(mK0uTnf;~Jy{G{ghqPq$4U7Wy@RdvBsD3vDi)<*r`(PMpoT(V;-Mlm8Zy@;4YA{y
zolA@^?pr+QEmy{rWI*tp+O?SmptW^OXs>If)~9_4P)RaJg~xz?%aKSkC}J4?Xo$&M
zu2Vz9?O|xSqr+iu4Yl*!e;AJ-I>=YG({kP)c>7sXligUiv#^ui=IVObM_Z5;!YWl2
zdPqU8@n}eaf#z|RR`-k_*?e{)V$&L)m|nR<xq)rWw<VKws4H{EQoz3xTE!#esEcwa
zfrr+2_6|I6qNlZ&ZJk3T%f)o-sb9L2JxW#QOdKNL!p9y_0X+ECaxv|`GydjQ9o|fi
zlJ6R<j8hvn9nr%|str#kgEQ1noLm*)3TR`^SvVwj&Lz)QO@1O$@c9uKj|ATlyB5RI
z;A+56D)(=a`}Vqkr{c2r?D)0>uSCA{&X=ym{vMWOLzGOQShq>=$R&lpoaR>(2#s4(
zSoYGl6J^8mMIDW$jIJS7^5mD*^~<svkGhrRfoj`A@W2W0R0FRD*MlpuA}eZYl`5up
za)WceAK_oJ44Cf!3}E1jigWxS#D%{Ip*A8#G8FecDIHg6QxgU?7kFXxv)@S0a#a2H
z#{3SP|KG&$bQJ#pRhThs0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_
z000000000008jt`0000000003QTxI<fT?iMU;<B20Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOtkgdI{*Lx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp.16AAA28D16A62BE9.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp.16AAA28D16A62BE9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5c5c8da3ab6288be84f5c5bdc4f19a953079b3d6
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%+)|zoQZ*<V$RjRgM5b!c$~kxx-cDS
zoZq?0EcxQ4Wm%UO{<Xa%&|w{85xGt0H}7PPc*cZnM-0rnuB5akm9O)dS(1H8Awug~
zM#!<&>@?Psj@2m~+$G|o_Z21y{|ftIFx}`+-u9w52ag&Ry<1tkd7V+%s@u;`YZsk6
zs}=Ro>-Wqj&!#XRUi(Fz+4)H*)8or~rDyEg)svOzSmZ4<A?c5}-yA>pIeO1y<L~eX
zO+Kf{P+Xat1avY(QEFN-gw9C>(Zxl{#ri;+iRJLA*-qEWFCAiLWMXCj`V0n`VGITa
IU|7Ha08&0=<^TWy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp.3DCB7674F6E295A2.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp.3DCB7674F6E295A2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..841c9cb2372de1f3b973848515372616bfe28e1a
GIT binary patch
literal 1074
zcmV-21kL+WNk&F01ONa-WJ!2!WprT#0000H0001UbaH840{{T*2mk<hoV``sa@#f#
z<^N<>uwzwCI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$@POC|1&ksF}VH*t6K(vloB`
zKmp*Y$~dV(deuzp>iO-JEa2Vjk8=KSd-q{p-Yu>tXTX~m@?v#%V6nP+eFjX%XTVDr
z%ZE=Hcz8Fz`}{j^KL0r{S05i2<?U)Yy=I^+-!E6?>cj1)#pB2O^7fvQ?<wBTW`klW
z8rPy|g48n5R&;KMzLZv)eV`zeYdaU@+F-dsZ_$W#BBKRaWs=4V>@T-8wD!{PPaYm`
z3V4#n@Rv`wXd~~?YO27<HJXq`@14)<7X3apWa}(EN6*cBs+@0SHX=(A0-B2V*1?)3
zvJv<k+Ca_@uzL0C?HkC%dZRN0I@HoeQU{f@9P%q;$0BM7JBq3Sjg?I1;6`KTQOU?^
zjmk(bBin_8(j_hebl$ZtesPD-%rR1nY!RYQDmovID$b6xY3W>s5Zgn@Jj?MLlC@1B
zt+V^qr2vE9lg5xmXr<?VoWQ@)JJ?E3Qj=n>Vo~}!<#rqhwG;vu9x6hkA=jFuAvQeI
zbBfW$eTN6!a%NnTObEV$U6;!MR9nXd9k5PnecFcrl_YZnJO=bz_C%UV5!3KTMNHmu
zo*G(i4^zt>y}ll_p>dx350f!O2l=Y@D98P#Zr^KavK#Am7IxCxoZUeCs0*?}Sf{E&
z4=KnRkBStSXdY*2^-%F+HD8>F*mi~|W&n36H?WQQw&Wrm^2#-1Dd1lTt>O`K)O$Ik
zz(ebMeNUb@(NpbZ*JKyTax(pT>X+_ik5V-`61&K^@M&OF01tk(TF(0KoWHqMhc}bG
z<hurI<J5*#NA|FmYQxjX<P5bGr&k5I0=n366n4p-bIG$-lb?zde0~HbW5IXC*5Z0R
zx*GA5%Ke+>zWrLjb8*>wc6?icS0dke=S$b&a1TqeAxbWwsJkS1<dVW44)db{gvKo?
zEC=b!iL&MSqP~o!jIJeB^5mD*&Fiw7jQdq?>yhfZLh!%|?^FxFjBZ9(Vog>w)G1ZW
z?&Ss-d_Tg!L>Y0p|1*J+D=Lohn-CZNB81wA6v<HB4@K!~g*LTe)Nq0qdO!P(<S0kz
zr#I$j@cjQKhNq+W2dU#Rt8;m6VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_
z000000000015pZ#ow&UOs3ihVPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ
z3=I$u4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-O
sD>@@K78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOy|$rQvd(}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp.3C8AE5F8E636FD38.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp.3C8AE5F8E636FD38.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e7ed5b597524a3915820cffb2a3de29bdc6f923e
GIT binary patch
literal 286
zcmWIYbaNA9WMFVk@vO*AElFfyU|<jg;^LB`%(XyToQZ*<V$RjRgM5b!c$|MKDKS1e
zaNZ?ycJ9SX%d##n{A+tjpu;-GB66F~Z&|m%eT)ZkIur9mR-WWK`YvXo(aUQ~95;ll
zxwxQ*`}!HKNzLC*2#CGV)7s~#qW)#=heExlA9mk8wUJ%?^ogCzzn$H6Dqv;aeEsz&
zEX`MEv}(UipENs|k0<!0KU>?0)hx%BSDPD@mP%eZ)bwPUio($!`g1Ji`CG*4@4jU9
z_6~or@;Nhx;>z44praXzQqzhdbWS3OE-p$g)(0_wzJUTJmWzRxFU@r<Ph)0eVrF25
Li8C+&Ljwi?A#i7j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp.EC103046C3FCC222.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp.EC103046C3FCC222.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4101a3702af012a2bb4b9f686e402ffa9dc1f76a
GIT binary patch
literal 1070
zcmV+}1kw9aNk&E{1ONa-WJ!2!WprT#0000H0001UbaH800{{T*2mk<hoV`_TbJ{o%
z_W$WD3Kx!RZ-x&Eg)4JS!t~n7M~{%!cn4YX=>@|5`mSsU4efm>?ev3f?X%MEvu|<%
zPyo29GEQocUe%Lz_4MXS7VviZcR9PixqUw?Z|B4D8SwgrJYSw|m@lt?I|Ig}GvK9(
z#r?+&+`paOe)@|SpZ=MZ%MTCp@@Bc13>hfPcZ+4Ye1G$C{_x?hyt!lKTZ(tnX|Gs_
z+BGQZAhis%6|LK$FQt`cA1Darn$`umGFWcVTQp*o$Y?=UnWXW8^Ov7AwDQvLPZl1o
z3wV^q@SArxXd|!CYO27<6`GJm@14)zE&6?|$*!~T6g?O3v2wnV`4Cx(5YSY-whmS-
zk&VFn&<1k0ht;cBZ(c(t)*U)SphYchBy~_Z%Pzk->{vt%VMkHbps|w49L#8JJt`Sl
ztx*~2W#n_=pmd3o0Ihe8i_hlpi8)4UkWYl@lZwuVql%;BXc{_~A;jk3GWT-)isafR
zkk;9K<5Ga$?nz_FA~e!-Jx<_%=pAgOC#gxXR<S7kI^}X42sIP}Cmt$7qao*-q#-_d
zrsoi&i~AN2xaG(=C7BR>0lPM*0cdR<C$z^}sr6|c0#uUB5%3t$Z`l%QCPhrc9~Ciq
z%W-OGxI9cPS9Cb+wV`&N>ks1*L<f0PJCyx?mA7vdHQ9}II}1DMZH}&|ebfb6A*@nW
zp@$UY8jp$;m}qWiX?36RV>Vyxh}g7-JEjMBC^xW;d0KLk4sqq2u@vy%gjR72IqIw&
zLg1$LoxUT_i|A?XWm{(x$#O8=_f#+4$sVPuvnMu@r|_|7Q~)=AwOmZQ_l)1UREIm0
zt>jsQm2qmrsv~<?NwwkbWO9ZYij%7XTmfyY*$bQG%DLoOtI1D93O-)~<B{MQv1>6L
z4Xy@!rE>iyxo&?H@KjuOpB?X(;FZWT?|kW6?B}o~8=~X{iglX=w_H;A-EO|ufzY@m
zg=H^2?<gDYFY3}*%IF$WC3k*VjYrQ9%LCQ6h2Vx0-l+zDA6yTv#EPt_sa2|&-pLKl
zc|OAbco}fIf0)3)6&3sVQ-}+H5khT5iexD6`=oTaLYtZ}s5!t3t)KlyvX>+D(+l%6
zc>cPM;qE9t1Ae42igS5wVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_00000
z0000008jt`0000000003QDvj4%Me74!2(ZE0Ap=rY-j)g083^~0|W&I2M7rZ3=I$u
z4oMM3LnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K
o78@KLAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOv9Yly8r+H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp.EC5F0E8B3A5B73CB.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp.EC5F0E8B3A5B73CB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bcee92fb44a4c6cdfb9fddd0190ee7d096929431
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRXdoQZ*<V$RjRjeLg;cwBzFRygwf
z>3OQAxAoPLz(*oq-%nf-wej$wg2QwD%JXNq{N!WUtlG9SaEY3>NOh6xsU^{q*$xJ6
zbX#ns7h|mL!FNm9Iby-?6+hS}&bhGt(Da>C9+vN(n#`{^ZDO(i+u5a4HwN9ci|;#O
zX}&t6ReSGw74zjX5^_u5^YBaz=j_>SEj?q`uAZz!$0Bc`2}ys%{pRFd4`%<)9jtmz
zySOqp3FusgqSUlv2%VD%qKk`?i}gVapg&-M8OC5>0QwvXm{?9-_WkMI#j3^3$i&P5
E0N1cyy#N3J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp.F69FD80D45178C79.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp.F69FD80D45178C79.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b248cc730c55ee849a50b9f2dae840ce5fb9fe82
GIT binary patch
literal 1072
zcmV-01kd|YNk&E}1ONa-WJ!2!WprT#0000H0001UbaH810{{T*2mk<hoV`_TbJ{o%
z_W$WD3Kx!RZ-x&Eg)4JS!t~n7M~{%!cn4YX=>@|5`mSsU4efm>?ev3f?X%MEvu|<%
zPyo29GEQocUe%Lz_4MXS7VviZcR9PixqUw?Z|B4D8SwgrJYSw|m@lt?I|Ig}GvK9(
z#r?+&+`paOe)@|SpZ=MZ%MTCp@@Bc13>hfPcZ+4Ye1G$C{_x?hyt!lKTZ(tnX|Gs_
z+BGQZAhis%6|LK$FQt`cA1Darn$`umGFWcVTQp*o$Y?=UnWXW8^Ov7AwDQvLPZl1o
z3wV^q@SArxXd|!CYO27<6`GJm@14)zE&6?|$*!~T6g?O3v2wnV`4Cx(5YSY-whmS-
zk&VFn&<1k0ht;cBZ(c(t)*U)SphYchBy~_Z%Pzk->{vt%VMkHbps|w49L#8JJt`Sl
ztx*~2W#n_=pmd3o0Ihe8i_hlpi8)4UkWYl@lZwuVql%;BXc{_~A;jk3GWT-)isafR
zkk;9K<5Ga$?nz_FA~e!-Jx<_%=pAgOC#gxXR<S7kI^}X42sIP}Cmt$7qao*-q#-_d
zrsoi&i~AN2xaG(=C7BR>0lPM*0cdR<C$z^}sr6|c0#uUB5%3t$Z`l%QCPhrc9~Ciq
z%W-OGxI9cPS9Cb+wV`&N>ks1*L<f0PJCyx?mA7vdHQ9}II}1DMZH}&|ebfb6A*@nW
zp@$UY8jp$;m}qWiX?36RV>Vyxh}g7-JEjMBC^xW;d0KLk4sqq2u@vy%gjR72IqIw&
zLg1$LoxUT_i|A?XWm{(x$#O8=_f#+4$sVPuvnMu@r|_|7Q~)=AwOmZQ_l)1UREIm0
zt>jsQm2qmrsv~<?NwwkbWO9ZYij%7XTmfyY*$bQG%DLoOtI1D93O-)~<B{MQv1>6L
z4Xy@!rE>iyxo&?H@KjuOpB?X(;FZWT?|kW6?B}o~8=~X{iglX=w_H;A-EO|ufzY@m
zg=H^2?<gDYFY3}*%IF$WC3k*VjYrRq+^z?zZ41E-C%jV){64rIT!|G~QB$i_F};%;
zob!By|M4>5bpJ4cfh#KZ@uv_M{vw3hh!n|C-1kZ8a)mZEVNi2`7g|62jbtxJ=%*Lv
zXYl-W9mCyGd<N3*G06aPd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_
z000000000`000000000000U7Gx-xWREdtL1Pf!43ZDed{0000>W=#VG1qKHQ2@4Dj
z5DyMX5k*5L5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}
qBQ_Qr933DbA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8YkrqxgY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp.F6C8C49B5460F242.idx b/.cache/clangd/index/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp.F6C8C49B5460F242.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2f8bfb763897ec39edbabdb6d8cf4f62eccb7051
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%vC^IoQZ*<V$RjRjeLg;cwBxe?NMN>
zJ60sTee0_ufsaJKzMmKsxbg6zg2QwD%HJDxePU)v^X4gD?BN@x@$E_16wmEyyiH4!
zdf8HAcBF+(klEtiw!xu1@FTC1<-+{d^~IXg@4pS}v)vKlcFuNd{<Y9HuhQQ&mG0+u
zM!IUpWZF-@Gp8tFhL^pBgwlE;3I1|3gVIvTD~Fn%OjA)f`a^$?#l8IH8FS0ti3e{z
z=TKakn*?+)Ls4p4F@(-Z1kuGs$;J9Wnu+CN+m!jUw!2MZW@KV!0Qw6Cm|+YC24G0Q
F008q!V2=O*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise.hpp.C398C657987C23EC.idx b/.cache/clangd/index/device_reduce_instance_threadwise.hpp.C398C657987C23EC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00f79324f4069cec64f2b7348a5547ad949948e4
GIT binary patch
literal 1634
zcmWIYbaRVhV_<Mj@vO*AElFfyU|<jg;^LB`%os)nhI2rHiaA#&pUygLAmI8}Gj7Tx
zhZPyql|;B+D$Q1Dd3V%h){R3|iYNViU+%xY{>aLOt1n%8x2D>*n15EF(eBFK%N4o2
zw4Yr3^^{BdWMTiJCp)_9cA9B-d<e3RS>ivZzdcv-_+dGpjUMJp-*JZrnMZx!S^7fJ
zD_iX4EgS7C8IL^M+}{84oicYjpHasS#>=idjMnuYxPJWjm#~tIGgfN89SIg~-z$xN
z@Bi|dr6r;H;qC+7CzIX@y4<#UxW%i&G1EnOO@0(lz?5%Wr_5We;oBFasM_)*O!*`;
z2gfF*bMr138!ULG!2B@GTKd6)38F7}wUl_nmpUBcHtJQJsqH!6Wl33AjhdX}`=aM{
z7W+@<96gx#@A88=CO69uO|_}`@YVE4<g7z6Y%Y4tbFJQRyfwPE-|x!3V0)2IM}xLL
zPWGL(?HT_iU$e{&J5AkZPIcbh{_$sw{lrsV6Kfc{c3iqxT$!8X1Pp8a)`-Y9X6}tV
z49u(yT=uN?T3~{KlZlamhlP=gv6gR}tU!HyCN~o|P=O$updpxG-~?)BU;rv8N=+;F
z0LB#0zV8xmFK<$2<lzcZ4YF=-+$ZE>X%Zt;tZv*?C}<26_BZplgb7Q&T97|&lk{<*
zu(zPMCd?T!fg$rheEK*OO|@*g;GFB{GX&9umCEE}cQ>x{MiW*)wn3~$AXXktSYx{3
zvkTjKW}pdcE&AimruI()O&I1PMlNuG^d9=XV9i5o1&gUQr5qCuC8DWz%y}*3_-DN)
zP*{LnKmr^#K*F*84a3hbCASz^xB>(M;9=rwrf~hy&Wq?)244BPgf-&R0yMKj{DjnB
zOtfJI3OjH*@WUJyn$Fp`W4ciYn(CNG?hnow-g5v7$2rCMz`{G`Mc;;-U8|DNRA=mU
zWm_cZ<b@`jYm%E0I{RKWnsC*ZFMWasInJU9!;%Oy7dUCaf{KUBSIO5G=3E|>5NNNe
zGHp&;+KFaTPow3m)rH@(fWp23zM(L^(@uIUzuZ=*0~AixPql&x&xwe-na^?V8!Hc&
z2#bge%o#jfqAa3FJTVqAB%U~nI1*2SMFP%S_;=Tu`kg0p(A>1d`L?Xnp<lDnEMGF&
z)bWrq?<<f~#eEfFZd(1ixia$MoIaqiprxQ4%uQSD&oQd6R8d4Tdsj7oCZFD^#lZC8
z%i_xiGi0}oqTrK`Z@nP1k%adODK@xHkiL#)$o|-tT8k4F>}bM=R|Mqox^c7sg^>g8
z<dtZN2+QM)XsS<FHEx~Qv=x+lk+bQ=Yo+1Z3dia}!jk^#u+((tX8MXH4a|$sR6o1$
z*>+?7Vke-mgMfoP%sX%S51S@hXrD(@{bS#<))&ivfXzlK8H!SK5<&T=xG1^U8pL1+
j5fA{ZEMNdq{(yNLFfj%uF0stal3oo_NfvfC4o(IDk_|7e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.cpp.7EFDD1FE32B6F3E3.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.cpp.7EFDD1FE32B6F3E3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0ae241139573a507d0abccd82d4ac9d5568414cb
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz}2;JQiJrWnXc>Sw^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3$ij;ZqQpaqDo}6AgfH$c!B=qnuaPb{r+U(@w$K~
zX$*h)bc;6f4y~pdjI7XvEPC&JUbpD?u_3$8!gKW8yr<gvR%Ru#6d|Cgd2bz5ERl`C
z=g<anc7WBZS8v}y4y;!?L!d(~Z6tM|oMo5aC_5HWL)cMN4QQ-oG6ypnJC906rZp-f
zy^O3E4x~$51n9hLUHobepE<@zEwV<4KB?$@I8vM)XVcQT3?a4$mwA@scR{Xg0%@Jy
zw=M-3?4B4y7NM1%`*8yQMekrMJxNW9S;eCC>y+DZ2&knHxbRRD8V$MDBn`3R8P6$3
z7xx_=aLbu-Npe8&4eYvH2B5WdT+jjQq}Hc>2vAEhN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35YlJG*HF`)v
zuJK5uz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sUYz3iH7B3VwRUr+tgz3fq{CVOHN`4&D65f#9LU$2(4{yXPyZq?z<WGnfu
zL1moUFm;R`DycR+ogAE@mg4lP09Qa48}`B`xpOXgW;OY#NWtevU~(h)j@Y#r-;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n<oI5<GHA;ZM8ySqDPnmK2tQ
z^z}qJRAxs3%W6C+>+9G3%Huy$T~`R6GU1(S;kVKC=t@*%MMIrZ#q3^gaKV=%{D<ig
z*Z4B8BUjXH_zxj2{22(f5h;?PxF6E*tAaMQVbpL|7g|62VPr2y=-0I1R|NMz4+J;A
z_yYFGAE^Lyd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
znn@}5EG&fr0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8ZC`_Cc(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.hpp.2D7F2A668C85C670.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_add.hpp.2D7F2A668C85C670.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ad4959b9df592cef35e8edc23cf7146f87e5b2ed
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>S<Mc
zNS~H#bZcp*&?Zms|8W_Cd&2!s^Bg{N``tpuawVxtEK)vuW}KUHx#^IQ>Y8P4dcKx3
zzcAV?a$&jj!s)YudC+R9E$f~v<+^VA*yPEbzw5RiFb_9%b<3G7o5_DmT=Vt!Q@2-j
zOygSKqjY<Q+iD@%7c+04WB4@VYGe6V&jMMlBl=nGasRDlzCV|?UpM`2*LUeR+}oEN
zbt$gQO#(WXp(r)27((YHg6QI+<YIjg1LzMJ0D2b&m|+YCCYI{LH;W#d9b3rE$i&P5
E0R7}-HUIzs

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp.DCF9F78C2573F20F.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp.DCF9F78C2573F20F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..42b90a2dc420591e77909bbb1b1c1f5e9b25e02c
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TS2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz}2;JQiJrWnXc>Sw^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3$ij;ZqQpaqDo}6AgfH$c!B=qnuaPb{r+U(@w$K~
zX$*h)bc;6f4y~pdjI7XvEPC&JUbpD?u_3$8!gKW8yr<gvR%Ru#6d|Cgd2bz5ERl`C
z=g<anc7WBZS8v}y4y;!?L!d(~Z6tM|oMo5aC_5HWL)cMN4QQ-oG6ypnJC906rZp-f
zy^O3E4x~$51n9hLUHobepE<@zEwV<4KB?$@I8vM)XVcQT3?a4$mwA@scR{Xg0%@Jy
zw=M-3?4B4y7NM1%`*8yQMekrMJxNW9S;eCC>y+DZ2&knHxbRRD8V$MDBn`3R8P6$3
z7xx_=aLbu-Npe8&4eYvH2B5WdT+jjQq}Hc>2vAEhN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35YlJG*HF`)v
zuJK5uz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sUYz3iH7B3VwRUr+tgz3fq{CVOHN`4&D65f#9LU$2(4{yXPyZq?z<WGnfu
zL1moUFm;R`DycR+ogAE@mg4lP09Qa48}`B`xpOXgW;OY#NWtevU~(h)j@Y#r-;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n<oI5<GHA;ZM8ySqDPnmK2tQ
z^z}qJRAxs3%W6C+>+9G3DqHzHQe9UF9y8&cYT>uh_2^1eWJN=rQpN0EZg9c3BK(K#
z5f}MFw<A~7Z21o%F8mz`wGk<jp|~Fk(ANcRYQw1Ev@W!M_RGj#j?k}Z!mkMK{~if$
zfAIyyD<LCud2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z1<1szkPdu*0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Yq|Ie@h

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp.E64E06F9E993015F.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp.E64E06F9E993015F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2016d379274ec1e8841b2342d73291967bcb9b4
GIT binary patch
literal 282
zcmWIYbaNA8WMFVk@vO*AElFfyU|<jg;^LB`%#}cziHU)sV$Rn7gM5b!c$~kxx-=YV
zoZq?WTC!klw5jfe{qDC|9`Vglik|k@$VaL`ZIPw!f}2OWU-lO5TD-x9^X`=sOI6h;
z8#b)h5NtHrb;0ID?8;EyDC^3Vhqk7_c{Zp0{#UUd&sHrv(UH3`|5kIKtJnXS%1F_R
zK@X<{rpR8-2wlV+TXb8(LC(DNf$v`7nCDC%%eStN`)@7N|6ADq_~EM5iO0Quv(&X5
zJ;YF4nVST3FhfylS}}yqNd(cwMajkbAO_GUFaY#06fm*;%Ubq?`vE@(Gb0l-12asV
GfdK%#Ok{Ha

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp.71BA7AE5BF315EB4.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp.71BA7AE5BF315EB4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5369d9d439198dda6bb89834cca27f1849108817
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(j3;Nnn-}t8b#`E}x_*5IOm5DAmnN1E
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)p9y!pe)}nSLN!%?We`#$NTd3o{=9B?`N}tSc=BA
zD4HO(473%U+o3O|m1Z9(2<6(&1-Uj@ZqQpaVx7onK~|Zh@dEwJH4Ux3^!t;A$Lj)~
zq%r*E(=FP_JG7cAFmjD1WYK%)^SVX9j}6&63(wJW^PVc_TbY%}QiOn};=OgSW{GSB
zK8H4tvjePNy?Xlwa$vpE83G+@X(Oov<t)4WM%l558p4jEYCvNplR22t*m+bkGObY=
z>1AZSa3EdcB0%R|>*7~)_{=dzYLPWU^hrhM!;#|bIGdKvWeBl7xXiO0zYDUq38Zy)
z-?|iFuzO+*S%g-4?#BuI7rle6^dvPYW)+Ll*D1H-5Kv1YaN(gMG#YZPNg86sGoDk7
zF77)#;FdGvlH`Ek8`yQZ3_!JYT+jjQq}Hc>2vA8fN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35D};5bD)f+o
ztno;sz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sU=UUp43kt`?Eucv<LUiK(elRdGCd<&n3hzj7ruU5-h|DE$Ux9ad_vXy+-
zU~Qb*Fm;R`)>3VFIypE)Eyd|o0j_{9HtdB>a_3y~%xdyek%G^Uz~n~o9kI0--;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n=O?2_Ctm@TcAUtOKEOOA5<D
z`g)=qDzl@2<$63RtLxYOD!27WbzLEN%7k~Sh2KWkqbso{D;nyQDrWa`gA2YC;Xh1|
zxW<=x9l4@n!+!{I;m<&*jYyFU#r=?eUlp{e4Wov$y3qRB4<mayLcgX3zaqH*c_6s?
z#TQ1OALsycd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003Q9@TG2df4<A_7lP
z08an_0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8X)quR{?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp.5814F92CA6CF0198.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp.5814F92CA6CF0198.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0101b167611352a4aa412943cdcd17254f75d4f3
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>S<Mc
zNS~H#w9WgnO3Gxf|8W_Cd&2!s^Bg{N``tpuawVzDJW{@UX6Q}1EZnN<72(?(GppF>
zC3A&SC&!YDt<N0u7KNE+#GdpPU6*|<^UTiQvH6Yp>$19fZlqaXlG~ymbnX4r?V*CY
zqJEOD+YEcdRIM&dxP6Y{Q^b|V@~@r+vfM}Puk!8vS8nm{ym?)W{_Ux$HxlA^m>i8P
zuFOpWI+vj+HLVyz=OlvY;-chYeGmia4=7+_;XS8deksgel$nu<nE~i|7+{7m7#IK$
Cv|c&@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.cpp.D9E364199F88743A.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.cpp.D9E364199F88743A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..32f80edd744719ab025577ae62d972aecd4ae2a2
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz}2;JQiJrWnXc>Sw^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3$ij;ZqQpaqDo}6AgfH$c!B=qnuaPb{r+U(@w$K~
zX$*h)bc;6f4y~pdjI7XvEPC&JUbpD?u_3$8!gKW8yr<gvR%Ru#6d|Cgd2bz5ERl`C
z=g<anc7WBZS8v}y4y;!?L!d(~Z6tM|oMo5aC_5HWL)cMN4QQ-oG6ypnJC906rZp-f
zy^O3E4x~$51n9hLUHobepE<@zEwV<4KB?$@I8vM)XVcQT3?a4$mwA@scR{Xg0%@Jy
zw=M-3?4B4y7NM1%`*8yQMekrMJxNW9S;eCC>y+DZ2&knHxbRRD8V$MDBn`3R8P6$3
z7xx_=aLbu-Npe8&4eYvH2B5WdT+jjQq}Hc>2vAEhN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35YlJG*HF`)v
zuJK5uz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sUYz3iH7B3VwRUr+tgz3fq{CVOHN`4&D65f#9LU$2(4{yXPyZq?z<WGnfu
zL1moUFm;R`DycR+ogAE@mg4lP09Qa48}`B`xpOXgW;OY#NWtevU~(h)j@Y#r-;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n<oI5<GHA;ZM8ySqDPnmK2tQ
z^z}qJRAxs3%W6C+>+9G3+REpV>bgSklnL)t3%`x7M^~aED;nyQDrWa`gA2YC;Xh1|
zxW<=x9l4@r!+!{I;m<&*jYyFU#r=?eUlp{e4Wov$y3qRB4<mayLcgX3zaqH*c_6s?
z#TR7wAOHY!d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zkJ?HC4fVph0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8afGSO84

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.hpp.B8A4A04F978ABC79.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_max.hpp.B8A4A04F978ABC79.idx
new file mode 100755
index 0000000000000000000000000000000000000000..afb82de642d93e6cb664d7da2622952c60b48277
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gM5b!c$~kxx-=YV
zoZq?WTC!klw5jfe{qDC|9`Vglik|k@$VaL`ZIPw!f}2OWU-lO5TD-x9^X`=sOI6h;
z8#b)h5NtHrb;0ID?8;EyDC^3Vhqk7_c{Zp0{#UUd&sHrv(UH3`|5kIKtJnXS%1F_R
zK@X<{rpR8-2wlV+TXb8(LC(DNf$v`7nCDC%%eStN`)@7N|6ADq_~9z?$~7DOccdRp
zEUwH=0y>wWC^fAZLgyrc=;ET}Vto(;=nohGdKU_qSk}etOW*lWFoT(qiJ5^JCeFYB
E0I@-1t^fc4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.cpp.B60965F1C97196EF.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.cpp.B60965F1C97196EF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..694c861797c0d8ae21a1e97fb7c43b9ea1fda0a1
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-B)q#T|hFi;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz}2;JQiJrWnXc>Sw^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3$ij;ZqQpaqDo}6AgfH$c!B=qnuaPb{r+U(@w$K~
zX$*h)bc;6f4y~pdjI7XvEPC&JUbpD?u_3$8!gKW8yr<gvR%Ru#6d|Cgd2bz5ERl`C
z=g<anc7WBZS8v}y4y;!?L!d(~Z6tM|oMo5aC_5HWL)cMN4QQ-oG6ypnJC906rZp-f
zy^O3E4x~$51n9hLUHobepE<@zEwV<4KB?$@I8vM)XVcQT3?a4$mwA@scR{Xg0%@Jy
zw=M-3?4B4y7NM1%`*8yQMekrMJxNW9S;eCC>y+DZ2&knHxbRRD8V$MDBn`3R8P6$3
z7xx_=aLbu-Npe8&4eYvH2B5WdT+jjQq}Hc>2vAEhN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35YlJG*HF`)v
zuJK5uz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sUYz3iH7B3VwRUr+tgz3fq{CVOHN`4&D65f#9LU$2(4{yXPyZq?z<WGnfu
zL1moUFm;R`DycR+ogAE@mg4lP09Qa48}`B`xpOXgW;OY#NWtevU~(h)j@Y#r-;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n<oI5<GHA;ZM8ySqDPnmK2tQ
z^z}qJRAxs3%W6C+>+9G3+L9fqt}6sjnea}v@Z0EmbR{aXqM=TyVs<Y#xZq0>{=@W$
zYkZm4kt=F8{D%-1{tSfLh!n|C+z;vZRY9BDFlsof3$361FtV2;^lMu1D}wu<2ZEbl
zd;wj2ANl}ud2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zlZm#MT+0vI0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Z&S<yQH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.hpp.97362179A977A149.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_min.hpp.97362179A977A149.idx
new file mode 100755
index 0000000000000000000000000000000000000000..421d4c6defa7e2344b2d0e8e1c2ce3881cf26070
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gIr98JT8Bgnz`iu
z%$$(xZhphp`?A*W?}zldzF!TVkvOCBW~}K$Czs@T4vW1-uCq`5<C1FVxp{T)iQtoZ
z&P?|Mc{{H-y%H38cV)I(uCZ(8p<KNi{Y&4}U5^&5&Y0DkSo=x$t!KMd>A$$jXx?W#
zi}Y4(>RDMS%)KD`r8cw1qlJunR{1yGj!^u#a^K6Z`wxFCu=xD7B<9_cZ3p%->_|Qu
zUtF1+1avM#QEFN-gw9C>(Zxl{#rhxy&>t`W^ez-IvD{pJylQs;IZ<XtCT0d^m^cFi
E08Kw-2><{9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp.C510348DB0E9EAF4.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp.C510348DB0E9EAF4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8d4f2c8f676e9a557e322f39d260b87ca3ecbbbe
GIT binary patch
literal 1066
zcmV+_1l9XeNk&E@1ONa-WJ!2!WprT#0000H0001UbaH7`0{{TV2mk<hoV8WkZreBz
z)&CS@#!2dJ7w|*k+HJ5+Vsu;NQJ~0?H84eLd9maD`VJ+#ZrtuejqZ!cIftD4Pz68%
z;Og2qsX=<xOjq^u+Y4F1yV+mm{Nd*I<Gj3Gj3+0+n-}t8d2(Q}yt+IACf6sxOBHty
zpEL0AZhrgaJ#W7JJujD^9v9`!@@_h2pe#S!Ez9M{o6n2KPxs}`JtIFOKFnqVaVHwr
zqG*EDGSEhJZi~K@R+?R)Ae3u67i49yT%)&WM3u;BK~|Zh@r?WD-!xQt>32sBk5>ge
zNn`lSr(3j<cW5=$U}S|RWYK%)^SVL5iw)UT7M`Q$;yu;Qw=!QMOA!K^n)lX0#S+;F
zd=70OX9rlldiC}VWXJkTCkS+?rH!Qal(TH|yUUJ6)DX55RRbDpnao~|#?GUXk!g*}
zNG~Jb3wzQfP6Bk^wJv^BhtKR|q!#%`h(4+4d^k`X9Y@pBsSF{udy~1B;}1@*YyxSW
z-L)<S81x<)Ll&Wxp6hW0|3&X$BRxq?idn^?^sAK1v3IDY5IFHr6B-RU*CY+`#S@-G
zj4tjv+~bxb<CJ8F;5*oLISoK7>o}nU)=908`w*a(WDbDGfPTZ8NV8MKZukQcleZkF
zhL+32uH}l3#{(N0=ehncxrXQ<k7|#y-5>Jyv!W)uwr*=-E4|Ip4Q?N}AZvsw)iruZ
zL9XyXq`;2mc9vET89!w6#fgY*XSibqaEo#co0z91Cutv7&KXMq|4C>Sw~(V=%RU5d
zTHo`%^Sp_kR$g{ZR*@_R)33*R>0b6IRg*2Tiado+gGU8$<JZf(S^u5$H<xPvX0nz%
zYfu@d)=V9|hf1n7cPBe%sHHf)D8Loa#fGi0O0Jwso>@(PDpK(I5tv*Ho)Nne<LlAI
zh@Vuh-!#|lj{=^H^WL+=+Y-DIdFGujU5CRQmSjVeoItValHit03V+$muNn{<H>9u}
zq;Dt6Av0SFSXSdnSzlfDE002ZIZ|C$2<|iCooeCr=xTH!Dzc)XPN`ycFW0!>u?YX6
ze8gEk*Y3y_HLLzph%<i(LajxLWGL>26!e)vo7ymHIIc6bpZqkkl>_v9%J4g!``?>_
z8=&|K#{(hw0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%00U7Q#>%yMHcCDM
zPf!3)0000000000002+`00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3
zLnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KL
kAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOj<P31poj5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp.B82C759E6266A9D7.idx b/.cache/clangd/index/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp.B82C759E6266A9D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..434021860bc8d3f12742c51f2789a45ec4251a64
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%+)}enTdg+V$Rn7gIUamJT9MIMQ=Ls
z|CL<0RXtX4eU6HC;{Sac7Ak)C^-pR`eD<>BVpG0S)GMwir%f&QR$R57b9710LdTWs
z*0!Btt!b6!TDjE7NFz40ME92YsY|WDuYK6KrK~D`y>kDRtx|`NNx9$ZKBu>D@47#c
zvSL?vn(cU#oOSW2(2M20?93XEmvR0HIM4F^sN&Ccg)hF=w?F>*=jinFnft?P=a$`x
zce!<plcBgWHwoxuhN9H8VhEj+2%?LNl8g0044_}2fQjYA#~C8CHs#G=W@KV!00saI
JFvA!O3;?|}YJ&g(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp.6F6EC6F8F6066B26.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp.6F6EC6F8F6066B26.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b3221458b780c6c925b7d6fab1ae4baddaf23e98
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TS2mk<hoV8Wka@#lz
z)&F#0#!c33cg7EiYi~W<B%WR-j~beUEQ~2q%Zo4l`UNd3ZrtueJ>3UM1P+PI0YDW1
z1%Rt6<D>@ZRXtu+yLVTzfcKNX%jv`I-N$KpHye)56mMV1v&Gql+2Z>3nPPNvrg&*$
z{_rI$9^OyyzW&9Fub-#o;?v`-yj{%4LspdKhxwvhe7yZKd;D}?-rlq3XTXQaqyy%n
zb`6R;NG$_xMC-QbOKGLq2MR*DrgcFs4VG*47L8aYGFp&TCTYCj{N*PNExq*nlZD6Y
z0-mHX{N~*a+Q=)knkq1Ii6&&xd*}0agMJ@tva2lYqUYj0Rn9juA0kT;0-B1~*1?h`
zvJrS6+Ca{BYW3>XySI=7I~+RGK#N-1Na_eV%Pzk=>{vt%VM|ffps|w49L;EKJt`R)
z)~Jm1GV-}_1YKehp!KeC@!1^iIL1f~@`(_AQqlR)L!2FF)6lsLAvQ;sxtHS)L9T29
zX`S6SE(Pf9o)AM8p^={JaZ-Lm?_eW6Nll7T#iI19l*@4psG$&;c&G@ChRiidLwxWI
z=M<xh`xcLS%b77HIUx8>?b=KO(8@X{w9~ax>(e>}s3e(Q;W41!uqDzQ6mb~72V(M;
z^VHCAc{sFO(c!Q|L+w1*A4WG29pq6xXxZ-%y#1`G$*!&2TG&c&b9P<WJr-nzuuN5j
z9#W7i?12<G(A>__>W=ZA%@;c&Hm%`~>6BZPYuLm*Et#ZaUYRqN0{)ZGDsCZ12P?-E
zxM}?XKLpQ<=xOC;TW1r=ax#bSsb2aZdz7lqp4dd5!ly2x0=V(3#e8yjPx+lob-Xj#
zN}e@X8mHC__0hvpsx@~f2WP0EIKC>t70||-y|78HoJ*cjO@1s=@c9xL-3Xo$yAs2j
z!PS7TRIcAR*X>OKyW+C@tiM}=S0c~6^QCLCo5PZ9h>{5ut2PO4xuo!y-TbNpp>abB
z%T9XUQFhF1DPUO*N6(LJ<Zhtawh-KA!aLQ#AA{?`l~|G$HML3=lY6<w8Bay{58DGK
z`9ik?S5$2IPa!V+9SF4+DUzYM?+7^Xf;Kf_P;*)rT0i?`WG}ty_Y~oG1owZp1ed?~
z7wlyrL34R+VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015uQx
zue<T58ioQ-PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkO#S83oB#j-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp.2B3A1DBDB134572E.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp.2B3A1DBDB134572E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b90a662bea2069ee080b11bcf42aa351e067adf
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5DiHU)sV$Rn7vwVjPc$&YvxWrg|
zFbd7E6<b@Mb*p!u?Iks}cg@ax+jM@PHt0G&w_}-z!{klLQ(KFo0(soO#k!tcHevFo
z!-vmk2+mW9x-WC$+?7=oE9c6Dn{7>xSvIHr{#UUbX{(oH2t;2nej6v(>is|NXM|{A
zu%YJS4`ynC+V0#{&$byi$cSw;*li`W%#Za=*u~)W|7FkhT)$l?^JDA9eT?-QN6!>j
z<|Y9h%TSb>Rt%wY5<zrvQF5_9hynBk6fm(AZR?o5cX?nAGb0l-1JLU*zzkzBFaQA9
C>tO`|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.cpp.501586DF17B9409B.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.cpp.501586DF17B9409B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5d714d2d3758bd7a305375a17f0ea8b8ba5c60be
GIT binary patch
literal 1060
zcmWIYbaRtoVPJ4h@vO*AElC7283cj2xTGjEfSG~e00#p@#hk67xASgW2weRyl6=hf
zTCVs9=SR1qlV=}!!CEPjQIx{#GchWz=H>PG7dAaQ_SWvo3p<ZTbDYoBu3%cgP!O6r
z=ZeDKRmWFneShaA_Mz@mxn=E+y}vDV-=`hEDEZ-At4;Kr+!?y-;^#9io^AYKY1hxg
z(|K(6%)eLrw_7~k-v9K`<;No{chBf==b7@f{pX^XnmvDws{fS4-@9l0$@tOXluLre
zfxFpUWtU8AG+R^}{blKvOK+H27sg!Pr7)fCv}skhn~t!g?g_2QCUFJ-K22_$7WeC~
z=i$etjQ*D->OQ|?%Y7mm<tr+&(?xjGi@l$J?snYYZ+bIStn6a($9Y=MrDyU_T*|@d
zdo=rX!bzc<9DDe471FIE_Evp;r<-z|ul<}rMxeK}r>tiB%_rsUvfV2bb2L{;nS>e5
zTrAbQ)?HQkveeXxac*|veM`LOuw?Abj;RT2xBg}r)Lf(FGC!om?u&@~@rx!Udd?kZ
zFQ&YE`lETW_H2i!^zXgij8=Qqk4<pRSn|AT&Vl?FvN=oCS9%(*j+jz6waBuM&s2m-
zT~5nsvEuVtOBC%Yjum%YJX+iBey1p9&ZIX2HSD|1mNRCg%BpWU9JTUmrYuv)Bp?0?
zo)_k8I!&xy-?Y8&oU!Mrg{eu5T$yfoiN%{60oLcgwqN9RVXq9~*SP&xG2dQ0Yg6{?
zZDQLl-Ij@cUUr@}Qc)*3bYZ*56Rv6{)-sdro44#bT`~WxxV4hm+J}8F59ppU%}JU!
zZO)S97|V1W)<5bCL+?2C@kRA17F_<tTwgFx{Nm+Pr?;Kcn)Eb*?|$&!FOJ_dLg(sk
z?6PU~pX<(A@Ho8t=aUcfK3PwHRgu1Fnoafw6Th`HU+ryQ8j}6auk2iab74(rgZ6@6
z)9A#hF_zP9HqZJKvE;#LRfAoEHp*qHZRX9Z8RJ*R>@Qq*dp1+q(bVse_1jt2yUeS6
z?tMMUyj}ICfU0`Hnp-M6o=!RxZ}xdLTgx$CO&#l(mF_oF)4Ul^XIWI$N$r$gu}zny
zeC7u~(+|fVR28UBOL6sG8+7zZ+`7VJ&Vv7(zcv)t%)HvDtrhlXxq$dTo|dR?k&^*;
z${)&aYBDoT3@8o}&$9oWn4)d_rP}L*^_<`HrJK6r|8RGIQ;;jJ%uNDjFNUJjv|<RI
zlL(@Vi;|0#K@5Ko0S3&04L3E^tbR^h&*UG#5C9WqU`WnQ$;kk6z0-V|Sy<WFIXJm_
zcm??RJOy2y)daaborJ`NMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f$SNo*si>-JYU$`3
b7#f?J%b7^&8o7vyxqAgMx%oMIF!}%hIxD}6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.hpp.EA77360039638CEE.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_max.hpp.EA77360039638CEE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0732c7d85383c93632632b9b21c4a860b5fd93df
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7vwVjPc$&YvxWrg|
zFbd7E6<b@Mb*p!u?Iks}cg@ax+jM@PHt0G&w_}-z!{klLQ(KFo0(soO#k!tcHevFo
z!-vmk2+mW9x-WC$+?7=oE9c6Dn{7>xSvIHr{#UUbX{(oH2t;2nej6v(>is|NXM|{A
zu%YJS4`ynC+V0#{&$byi$cSw;*li`W%#Za=*u~)W|7FkhT)$n|7%%-sdMD%2Y=+{>
z+$5l58H!TViXn7PB8V<7N-ow1F@U~+0w$KLl8n;pC6zxiGcqwV0KE<a%rFK60{|LU
BUDyBs

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.cpp.F85FE38D0DCF8C44.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.cpp.F85FE38D0DCF8C44.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a2b08f506df5de5034832daa96bdb335a2169f3f
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8Wka@#f#
z<^N<>uxeLLI->_$ks3{7TazZ^M-B)q$sK_Ji;HCS>$?yoD^}8nXwnCg*mHorFIWH+
z0Isf$lNzK~?X;<%-(AT9-p~Fl=MT4cALr%WVmvtm-oBC-tFsM@)%7oDz~tr(c&%dj
z@Hqny@8@@4{@}%zzvkuY)8nGNT`i|$2Fmioa#gNA-hN&@e!4Gj?-}_X@nJR_h^1&<
zhoTKq%Rn2^yDj=sT4{EHf>5sOU67T*a*f`i5mh3i1zBa1#tY70e$r6orQe-2JYE;@
zB#q%W@9xk>UZK@ggOL@QkVWsE&)*IDU2MrVS$K|~i}zGJ-^qN4EJX-tYF=9h6-#6z
z@IJJGoE>2G=FPjekR96}IzymGEo~%qpqyosUmbQVqK2@gs9Mli%VZ8}H1-~qj7)1(
zMtT|fTsV*}aT1{Su5<B49X_*<kvilPA^N1E^WjKwbR11b=Q4!Y9Zcq4j^8-h*aXr#
zyX#yEFz7uohAcuSJ=fy|{)67ZMtYK(6tjv&>6?_xad4=k5IFHr6B-RU*CY+`!84vi
zj4tkaJm8ii<CJ8F;49emISoLKb)3)v>!sGGbqG*PGDpB;K)+#4q}eHAH~f)^$y<(7
zL&xP|*K$S2<ADvW^IU(J+(2}YM|F>~-EZ>ty`m<&wr*=-E4|Ip4Q?N~AZvsw)iruZ
zK{j|KQea1OJ4>sFj32Z4YDdJbH{3A;xJ9{!P0Z7hlXQqH=ZvL*e<!qxTgcH~%OM1A
zTEFM_&hsLAYP{^*tRh(sX8%3aOYdcmQnlF<tH@LMG<Z}1H-5cZ&i3y)zjLV$cP4Ae
zvj&xMYR%Nqd#I#Zb9b_HhB}JVs{&jBeQenZtK`bL<eAmvry>QPFM-L8;2E)v7~hPp
zMtr4m{ieBYzZLLYTy~!w@0Q?|$TRPJ>3SUIup}Fz<OGVQPl8)6Dg0qGKWac|+>pX@
zkY09_LuR%Vu&l?Em!~5;Qhi?t?lR$>>fqPW_2^1eWJOE8QpN0Eu5rOb5&lE<h;w|I
z)sZV|R{XmV7yb-{T8k9PP}~jS@3Vq7bz#(UR2OPL`(b1&N9gC2;Ac4ZKQ{yyzxWq~
zdLS5cd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003Q7*d!m|bShUjk2108an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y42+m#r

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.hpp.D976B4CDF936A58A.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f16_f16_min.hpp.D976B4CDF936A58A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a4c330698dde5cf1ccef220d2117371d835ebcf
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7gM7^fJkCECx$b_m
zVg5>|Y&RFr)yr5vexK~%ynC~3(T3EU`gaus{gtFH^GNkF&(@i8x$9AcWuBUw-C0Y?
z?}B+wfi6N{H?mHczaq3V$eb(Y=2hLtX3y^YUAJAqBwlr8^R}Dowz$9Kp8mG_X>#a_
zrH_K#9@hF?(OS*Ex5!$;fw8+{hx1&{@Z(H;t8Xm5{-5u8+waGFB1_t9nRom>n#oXH
znVST3EJIOhS}}yqNd(cwMajkbAO_GEP{73U@~nO8d~e$a%#2LT3_!2L05goizyJUR
C!C+tj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.cpp.630A4073E2B46FEA.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.cpp.630A4073E2B46FEA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..92ab9dd70500765a54bb844174be48f450a51454
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum#T|hFi;HCS>$4!mC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWo^Gn=w^y=&ceB6B`NQqqhk1Fo7*EcCH!tMH>g>Q`b^ZDbnB1HJFHI~T
zK4swH-TdzJpS=0}_q<$vd|Z^btL1deKv}+DuFBPi+fR$fkN4&6JtIFN-p^(|u@tpy
zP}D(c8E7Y3w?|(}E6pKL5Xv>J3vz9++@iN=#5$4Df~+!0;|2PcYZ_X6=?^ChkJkk}
zNn`lSryI19cW5<LVB{K2$fEbo=XHnv5NooVEIdcg&3mexZ)8>?OA!K^iucyRnkBLk
z_#E0m&i1f+_3G^#$bofAX9%>YrH!PHl(X#e8)e5LY6yFZss@dfOy+1tW9w1L$h1af
zq?eKP!jW`|ivX>6jf-E+;WNh=sX^8V(I*w14+F*7aW)N|%MfC7beU&4ei!7%CXm+I
zL*r6_-tLJpWDy$axgRI+U-S-k(v#Grm{lxFze%|r$AB6NfeQ~6q0x|QP0|o6p7ESw
zbn(#Q5x1Nfmm~)S-@vZTWdPb(#|7=NR%(6PhX9o%GXNd~`W;&$%|Q`|;RhlnZ#hp5
z4Y!9w%N-q$dp6Y0bN^v-1JOagsvXLHzpLAinwsp^y1j+H^fqVLhdppXRtW1<Rp=oF
zxxs-*fdkFsEUoS<K2-C?iHJ>Wcw%~Rk8%sUm~Tri(lM`GGnNAWlh7(2AxE8+V+uU9
zzT>;#c@sTtylm@iB3VwRTTlJco$OJnI(uRh`4&F)5f#9LU#*t2?mOpiZq@P4WGnfu
z!P+>rWon2X)>3VGIypE)4aMnI0j_{H*6f8%a_3y~%xdyek%G^Uz~n~o9kClRz8PJO
z_(|pdO>^IVFW|Yj>^&Ramf)4hci#EZwb<Xol5B{Q3n(^i5<GHA;ZM8ySqDPnjue)?
z^z}s9S7uKE%W6C+tLxX@%Huy$ZCePQGU1(S;J4BB=t``~ike!birKx~;({+l_z%+~
zuJL7FN3N*Y@E<~4_%je{D^esw@zAH=RRwKo!l>q~F0_93!^mC+=-0I1R|NMz4+J;A
z_yQIeAHe`~d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z%Su+91&ezd0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y9=hOKB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.hpp.7BFEE24C60F74EC6.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_add.hpp.7BFEE24C60F74EC6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..574fac2616f4d1e6ec5763b22fc33e9e4bae2088
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>S<Mc
zNS~H#wAK5vO3Gxf|8W_Cd&2!s^Bg{NJMTh+wP0^1Td!K2kNByVori=}*DQ0>J8L=f
z3!}|K7nVyeoIWd<2d<XdvhK-JuIr|cO`hHPyKegd^KesFx17zgnf$lJHD7-}b$eCE
zG_K`6O1EdYtrn7fG4b{}hEEY!8q2?W7RYiRvA@c<_g}e%+&Oc(b<^K=eV2a2y+h@w
zPjO{#641E}MX71U5IQFjL>Ct&7wdx<Kz~316U$%SWUg&1{2nkfGBGm%Jr4uSFa`qy
E0J>RXB>(^b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp.BB154BAE452FBECD.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp.BB154BAE452FBECD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b9af3c032b98c32f1e012b8640b547e4ebda316
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwqwDI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(j3;Nnn-}t8b#`E}x_*5IOm5DAmnN1E
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)p9y!pe)}nSLN!%?We`#$NTd3o{=9B?`N}tSc=BA
zD4HO(473%U+o3O|m1Z9(2<6(&1-Uj@ZqQpaVx7onK~|Zh@dEwJH4Ux3^!t;A$Lj)~
zq%r*E(=FP_JG7cAFmjD1WYK%)^SVX9j}6&63(wJW^PVc_TbY%}QiOn};=OgSW{GSB
zK8H4tvjePNy?Xlwa$vpE83G+@X(Oov<t)4WM%l558p4jEYCvNplR22t*m+bkGObY=
z>1AZSa3EdcB0%R|>*7~)_{=dzYLPWU^hrhM!;#|bIGdKvWeBl7xXiO0zYDUq38Zy)
z-?|iFuzO+*S%g-4?#BuI7rle6^dvPYW)+Ll*D1H-5Kv1YaN(gMG#YZPNg86sGoDk7
zF77)#;FdGvlH`Ek8`yQZ3_!JYT+jjQq}Hc>2vA8fN5Er1zhz6LIVj>V{E>*sTh3EM
z%kAONa!1GGfenrG+<%zdKy;9=YLBwt@9OrWrY5_wZf9X9z0KJTVIR35D};5bD)f+o
ztno;sz=7s*mR1iHKUVX_iHL1ycwz={hjIhkm~Tri(jl*0GnNAWlh7(2AxFKHLkc{!
zzUTYkc@sU=UUp43kt`?Eucv<LUiK(elRdGCd<&n3hzj7ruU5-h|DE$Ux9ad_vXy+-
zU~Qb*Fm;R`)>3VFIypE)Eyd|o0j_{9HtdB>a_3y~%xdyek%G^Uz~n~o9kI0--;AzC
z{G@XKrnzsw7w}wM_MRQzmf)4hci#EZbvWF^l5B{Q3n=O?2_Ctm@TcAUtOKEOOA5<D
z`g)=qDzl@2Wi_6Z)%ELsmD_rxx~>pBWx_kv!f&JN(Un+}6%BPt6|;M}!3AH6@E@i}
zT;t2Uj$Bc(;Xj19@Mj>@Mx;oF;(kcKuL|1KhEc;=U1<I6hmpM;p<mO2UlH8@JP_Rc
z;tOTpAOHY!d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z6wB}-NqkM~0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8a6m)^<%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp.442706BD47327B2E.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp.442706BD47327B2E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7e2046bb1688cf157d43953d2bf3027d1ab9567d
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>S<Mc
zNS~H#lskQy=Lt34ztsyDRfPMW<~e-kcHV^sYeC;kMqjl!pYT(e&WW8uWx>kRr_J&C
z#V{{Wi7B&WVdaF`D?@ps%0C6QzMfY&tLR<b_2>t)SI<&X-gb0fCjTvQ&EL;YonF;3
zjca+2((xH?tA%7U&b&Rz@G0VI<NQ~1A6SWY+`qD~>hJpkn>qP5G5W`+mf9r9@7QuQ
zvA8lf3FusgqSUlv2%VD%qKk`?i}gVapg&*$=v^paVwpZ~ca2%+G8Se=CT0d^m^cFi
E0FDG=!2kdN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp.01982D48C410CB47.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp.01982D48C410CB47.idx
new file mode 100755
index 0000000000000000000000000000000000000000..261169ee219fa00e7d474ad5a306d6b89d256da2
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TV2mk<hoV8V3bJIEy
z_WxlvYT6Ru%;cdB31pych6B?_k0Pz(oyd~WC29Eetn7r)z<D6Ud9k$LC+&UZ0-ykJ
zRb`yiAib)`tLo{^g)HFh<U={Vzq$Q5EpKPT(FySSnLJyZ9GESxE>D2b^$GA?#r*zr
z2JYWZZ@;|b&6j_t<>J%Bth`yw$3q6n^8I{KE<WCTo;`fJD{t-?`5p0oGU<uAs9l4i
z4pPfN8_~Kg`chhHc7cLWu4!G6OM~Say+tFIiHsIxl}Q@UxPSglLrX9H?x^A6s(?pn
z41f7_gEsOGt)>c$T%rkC^xpZrZqV;yO?H)qr|7wOkCpR{%$LYggn*{vy>+l;iEIQu
zhc=M2J*-~5c=H;vW8I|_1X|S6MpApqSvL98Wyd0F2wRG(291?WX0JwL>ru(bv_@s5
zmyz#<J?RoB0b1`G7k{Y3C-yN?gM1@IpHy@{94L;CqiN_=h7g;*$=u8F7bjOXfwa!<
z8kYj}dXJ1Di_l2V^*Dmx&^y>jPg0X&R<S7kD&=zQ9cm~9PCQhEMnlduNke?`gy#^W
zi@O&0xaG(=CD|eP33hEx1JKGkPH2y{QtRVB1gIpL1K=^B->@dq>=dyZ{y@a!Eyt;$
z;qtI+xuU~i&xYE0u0M>fAv(yT+M#Usm%RP1sL8Ia+gjL4Z*z3L+XpVl3SpV53O%GC
zS9l;&U`KO1ORM{gAF}!EM8u{w+%Y}4MY)De%+r#Sw2v$2jHQ5I6I#VB<fzlK4}qK3
zcYNnOZ=$D_mu;O@B+J2c>#<(ClRZjRXG^RiPvK+lQ32fe)nY#BzEl3@QtjVN){<up
zmd2?yQwQ&1Db<?0lbtivP#j+r;0kDC%~n_?SI#BRtR_DeDfs*djIIUGh+T={_26Q_
zPb$}Moa^?gfT!ZT_w4Yt1g}J%dFM;lVn2r^*$^csP^{V{xaE?<KQ{BH286~9DJ*;G
z$BDAf%$5R{)o@f+SC`$&qtILqRNEGU`%HMJ8hAOl8eE7aSy5A~R57`eYn<^|g#S=J
z;4GhOci@VORsSQznLh-f)*?kR6nA|Jy3C+WO&HW1*O}T+ej3@z0s1>-_#4js?@hrC
zP<#jGqaqM<d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z9C&$eV;Kfm0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8YOJKui*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp.D90C9C6069808C78.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp.D90C9C6069808C78.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c66bbd1d4b79192ae5fc0597e507883e25ebde5c
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%+)}enTdg+V$Rn7gIUamJT9MIMQ=Ls
z|CL;LV{@$F`WzMO#Q*y?EL8mN>z~w?`0S-g_Ca$&-BMOv#Ym_8E3ahFB?WD}ups1m
zn6n;hO{+B5%B4m|8nKxry0^?vU26S(?Zd_`WmWO(mHV%3l{$P(%KcV%Tue=M?9a_s
zI;)Cv3vM32vZ!12g1<CBTfi}2;U5d<alY$z{uKN8!t4Kh$3Oq<ettf4e^~9@vODpv
zS;qt!iYs%IfKFy8N=++<&^d`9y0|F0SRcdy`UM7nUWNiDmIj|q7u~j1=P@%fF*7j3
H#2FX>PkCh6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.cpp.30899F9F46343B0E.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.cpp.30899F9F46343B0E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..167020fde34f36e1d178a516b32b14d4e7df6f75
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@G%!VKd9maD`VK8CZrtuejqZaa&N<}VhbjOH
z09RGUNe$AgW>QyA?=EBk@27v2v-_LdPqXrNKE6Hy-oB9M%aaZB<>l)W;QHzWc&TD>
z|0M(W?`OAPKk(w~-?MW0`C(q(EEkh817-Pfu`HLLZobSPKHrr$cZ~du_&A;R#6mQ#
zMbQMQWuT4d+!lQ)tu(tpK`7UDF36R^a*f`i5vxQ-3$n^2jc1%c|D>UnmwtED@Nikc
zqcn!!yt_pkd4*O}1xBvWge-dReEx3G?_xu?&cajlT)fB1`BvsbWGO;GQ}Nn5Sg}Ml
z0`Ef`$k`rNuU@@-3)!*#p%Vl;)Y3*$2g+GC`Q2g1B5DX*imCyPl}zTKMq}qu$;h-u
zWu%vp&xHf&5+?yV?^+k1)!`HS7^y`*5u#5jIv)m#qvL2=I+Y>B_Fyvia{R%`+9r_J
z*<I^WfL`yBF=P>1>A4<9@L%)}Hqw*Sq?lDKN?)g3j)Oxjg}{l2iqL4txh83d51#NG
zVsvrW;Q_ZC8K)#W1mD4~%V_|rt>c9DSSPhUu0w!Ik{JMx0sV$Gk!Gie-S7huleZkF
zhL+32uH}l3$2}Vw=ehoHeFf1$9@Ra{c7Mp*&x)Gt+PbZUt@JiW*SkG%K~@N>R8{C9
z1zF=jq`;2mc9vH686UFwVn@WbGu$yfxJ9{!P0Z7hlXQqH=ZvL*|0J}ETgcH~%OM1A
zTEFM_&hsLAs=e%*tRh(sX8%3bOYdcmQZ?BUtH@LM*n3m}H-5ETO!w~@zjLV$cP4Ae
zvj!{U)S9Wmdss=e=I&(Y47C&|7X`Qiy4bK4R>_rf$uq0TPecknUjo-xf@j3mVth5a
z81a?L^_%3n{ZYVEao&A4+%3T?k!Rld(skI+VM#Vb$q5v7mjt(5Quxbee${}`xFLmQ
zFFo%l`^;=9U|C(hetx+9N2==z!CfZ2Q!Tt1U5+lqimYg;Q>vKW$u-V-D8hfJ9&wJ(
zvpRA`#ftwF;>@3cP-~GQ8H&3;{C!r?rZ$Wkj_OS9CqImAWq^K934Vuj|8qld@r!Q;
zn;*4vd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003QCbs>pei$j0s>D^08an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Z0eAmwa

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.hpp.3C1359A00C377B18.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_add.hpp.3C1359A00C377B18.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cb085c1db7b7335fa0e5fc13df265b3682da9170
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%q2jYk%@t!V$Rn7gItFVc$|N`8s}{P
z>r)_HYUn#zKwbCeZ^zT8-reQ1^Wb<{IYZ@vr;DAg!umH6sY~aWi5#Bj93HqS^Q31b
zkAbn5i$bTF3Hv7Fty+#*>c^-0zLhke|M|o2dra4Jri$<;MSt<Cp0Kgt<$Z1W)k;}f
zYN;akj;mxn&2U_QhmUQ-VJ)#|0qKWA;t%8r@BQ-Y{=+k0cFaG1?(K&B?H`VM6j$aZ
z0iDWFl$uryp>q;Jba7E~u|9|a^aB(yu~@N3DaPtNe#6Yj#LNKnI1Dhu7z_*msDE9f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp.F787647EC7B1DE3C.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp.F787647EC7B1DE3C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6aabacf3732f54e8b55973c298464a3e0135d280
GIT binary patch
literal 1062
zcmWIYbaRtqVPJ4h@vO*AElC7283cj2xTGjEh?#-mFb4xe#hk67(Z!do1it-GG|S4+
zEJ*uwp^e4pg4yB5my6veUi;{ma`MF_?)~$o&NVS9kULQzmy`W{^0!ZK*Tpc1FsxhE
zGg)d$_}7fc-1~7SnC|u3Pn-OE_HF;kpUSNE*ahyXRP%qXF1_sY>FMfSuU88nEf(JV
zTR*b!e_qk=ueSR8{#k22pD(}c>Dd}HsZEP??9V)V`sec5=`RmI-~H)XpXq*PyPo5=
z2Etpj1Fkq`NoMFqZFh|K(mtyfAaKfbEq77khM0xhGma=tyx6qH$n&w}^YiwZ84s7<
zm({h25&NLg`_TM--UYX~g;zrwCvux8cHgs}zd!rS?uf~;X^GOd;_o*W?VI>hAhJu~
zLdc$Nd}2Bidk_4+z*e*^e6_g#y?u=`5#=vjn02!<l~>w6(b@gy<%VQI(-vOW(B@-a
zYMad$Z~VF<B_(39(WTQTo*$Ac>GW*4w)@yo_2L~*C6rB1s5^b}bbWr9@zEQL*=>HQ
z92;-$KC@S*p0g-Z*yHt^b%mM^`QJMvHgWWRiRoYb$-0U?`=w`AP>+__()eXDeYFm}
ztp|=cyJncAtn-w5)U0l?Tq5?))*aWQ&-G}YWIX7fbv;e0VcIprg?racjgr5v=%9L1
zMnsPJmua-pSx3S4hx`sc_lo98N$K&oPt&%N`IfO_?(<)VzlyrBSB6;Y{{Gv#{l2K_
z&8el*+|n;|&#mMC-M|{Ds1p>ruwCSd5WfSLOh#UMX6gAKjc3!XmCV*Y?0Y#?dWv?$
zaU0#4OKRh08c$=Y^Ue_6K4I~>Q+BL*QTHFle%RA7Z(7{8QnQIV>X&Q#zh0XACUBKi
zb&_x;m;CPzt_R1~KigdReclJ_=~?yZo2J@iZ!qv%D`Dlr&%HA2&Rm%}3)>#OSi}(S
zAf5f~klM~7y-GLJpA$U}KKFj`%j-kWjinNCyn1XkON;)x?aW)xeM9W|-Sqa|suilA
z&RECq68mkbbFxiiro+aZ$q%M^bN#<@cD|K>)b%AB-Yi`%n|l6C8`HF_Pfss*H!HFZ
z$c^U6Gk!QX>%r9rMFomtDXzY2gO2|2TUT%@Sn%KBZw;Pym){-`6^;Jq-eFwN?zGmW
zD@Eg76_Zumg=K0yS$?m5UaRl#dGmC^FKgEy(g*+E*FE@DzmDN5yP|$^Wo{BMgE17P
zrWHfzoJ0^^T$Eg_3}X0$2nYy(F&LNw8BYDmFuwOKmB~MVAvrfCCj%(wo#xBT!pg?Z
z!O6|TE5Og^Dd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%RzXooMO9r>
eOGn?p(Ad;m&O}Pr$VF7l-7AR6&Ck(;(FXut!o~9d

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp.2ABD3DA274EED7C0.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp.2ABD3DA274EED7C0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f818c5f0ed02c3752da922226c5c41fc4efeae8d
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@iHU)sV$Rn7vwVjPc$&YvxHL)Y
z2e)2JzW8hTvb!&CE9WLXva@)SaC*~zp2Zb+Iy1Z&O*E~;A~lwa9xX_@A37z6%Tm+c
z=1h;!L(fHD&o;jCS=%-J*2%)u;Ob+krz(HM<~L?<^Xg*T$ZC7}-P6DwU*1pc4;3^O
zoovLFpDeVsXv>21JZt6?iGtjlS1TvJ3OSG`YWwxq{fB2NcFP@ajF)~ZU3cZ^F^1yG
z+$5l58H!TViXn7PB8V<7N-ow1(o8Hby_J-87pp5VGcqwV0R035K(E69GmOE&006c<
BV0-`o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp.0D0748D60784A612.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp.0D0748D60784A612.idx
new file mode 100755
index 0000000000000000000000000000000000000000..662b3b08bdcdc30e2ff3c9e963e3befcad695dcd
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7?0{{TP2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@G%!VKd9maD`VK8QZrtuejqZaa&N<}VhbjOH
z09RGUNe$AgW>QyA?=EBk@27v2v-_LdPqXrNKE6Hy-oB9M%aaZB<>l)W;QHzWc&TD>
z|0M(W?`OAPKk(w~-?MW0`C(q(EEkh817-Pfu`HLLZobSPKHrr$cZ~du_&A;R#6mQ#
zMbQMQWuT4d+!lQ)tu(tpK`7UDF36R^a*f`i5vxQ-3$n^2jc1%c|D>UnmwtED@Nikc
zqcn!!yt_pkd4*O}1xBvWge-dReEx3G?_xu?&cajlT)fB1`BvsbWGO;GQ}Nn5Sg}Ml
z0`Ef`$k`rNuU@@-3)!*#p%Vl;)Y3*$1LZ86{O+)05jBJ@Mb&`DN+vU?(b#!ZGBT}E
z8R=!@b73G|;v_)lUF+hrI(%XuBelpULi9;R=fi>G=s22|PGtzO9Zcq4jz2hA+XT`&
zyK7wv(Ca-ihAcuWJ=fz1{)^thMtYK(6tjv&>FbosF*wvx2%LDR2#toEYm$ce;0ezm
zMi+M-4!GsWI3?L3_zremP6JSF9VfKMI;r(>9RgI6%mMHi&~I21X?BX(4SyhF@|NS&
z&~ka$wOrBhxMxG-Jl7wtuOK?eqq;}g?hkqUSy7W+Ter2amEPv)dbbZ;kQKrzRTX+j
zLDqO6Qea1OJ4>tkj32W3Vn@WbGu$yfxJ9{!P0Z7hlQhJYbH-A@e-c{7E#zpgWe9<r
z*6;bf^Sp?jYA?Gct4NlE*?*7q(tFvXR86+TD)JOQ_8t|$jbAMn)BStK?_8?k&SWil
z)?j6vS~GR<9#&GVxjWf8LoLP0MFFmWE;ek1RdVHA^2}=T6On??m%#Ov;2E*C7+;Mp
zMtr4m{U*6?e-!XkoOhoc?v~({$TRPJ={oG^up}Fz<OGVkOM+W2Dg0$KziL2e+>pYu
zm!5Z&eP*^4u&gd$KR<F)k5tzcg1bz3r&@S3x*T1I6<N_xr&KY$lWUywP=x<bJ>ndn
zXLaO?iWUDU#F;+>q1GZrG8A`x`1`D&O>G!89MzfHPktEL$^rU4CHNiA{m%`-#V@`A
zR%0LZ0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015xuU
zTZB*dgH!@fPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOz_mzi~s-t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp.F51F72472293693C.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp.F51F72472293693C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ac69dcedad0fa4ccafc8908ca31d6907d1c6c8f
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7gItFVc$|N`8s}{P
z>r)`S)zEjcfV%F_-v>{fdUu!4&V%D+<qVYvo-TH471qCrNL@UqOyuxH=kUNynP)vK
zc?`I{TogLZOxQPZZ`E?lQa?V`_pPM){LdeD-($L-GgX8)Df+8d^@NQDFYjy1uU5*^
zQcD%7>-W5J>SEKgJA7;t4y%bh3rIf{5`Q31eD9ZE_aC15vSa@7+ke_>8Fw@sjb$jV
z%uND1mZ2y$tr$Y*B!cMTqU2(I5CiB7C}3hykFt*Od;jbfGb0l-1JLU*zzkzBFaQ9j
CtzbO>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.cpp.35DDEC8F63798EAB.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.cpp.35DDEC8F63798EAB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac553b8d2d0a75cba263b82c4cb959df701c71ae
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TP2mk<hoV8V3bJIEy
z_WxlvYG^5NX7bR61TxSz!-46eN0HX?PGrgGk_*3{m7O#+a308T9&BsBPuly+1waAd
zs>(R2L3&k>meuy{g)HFR_+vS_zq$Q1DQ~C4>l5J3GkH2c*)W}7zB~c0uTFsHDrWaz
zGI0NHa{Ki?FTVaeDd(Rbrsd6iHX1TemLF#Ga{lS&%k<&%U3qiI$nS^`<8e>SMC}?B
zb&y&HT8q|g(3jFmvkMf2a!u=kTo^1@=q(ztNMy7it4z{(#`*J48d`YicSj8mmjygZ
zWBARx8?=#EXf;(}<N{5|qW8|{?;8Cs)?}Ai*hbIAd#s#qWIjZeA_O!QudRaxOJpPP
zKD2?H?P2xe#oITK9qSIAAkd<gHj>&?&a%m$4m%c6L)cJMHE674GJ7=|TaQXcrZp-f
zy^MS=>`9k63DA1ixcH<Fx9nr22KhvYKB?$@I8YoNN7K-$3?Vjqlew4UFHSCP0%@Jy
zH7*6{^&S~R7NL=z>v06Xqj#{Do}?zltYT66Wy<B)JJe7JoOq}RjfR|Sl7{%;3C|%$
z7k4f0am$f$O0q-n6YScY2B4*NoX{R?rPjxF2vA8f2f$-Mzh+IO*(qW-{DFwcTaHsh
z!{uSuaz%&3o(;A0Tz|N}g6JTRYKOAjU-I_5q9(htZew91z0J|}ZXdWHD}+U=D)f+o
zT;hR9fgR25EUoS{e#qvt9TA(>aL4rE2IUIYF;7cQ(mt-7GnN8=OK26hkfTn^J_K%B
z-|?OEyojEbUbb~skt_$(eUJ6ho$OJnI$L5Dc?utUj|$+%ujaFH_nz=Omui1!vX(q+
zurN-om^ye53#nGzo$Q>UhT`a=09Qa8Yqr8FxpFRfW;OYdNWtez;QC7NjM$|ZUJWh=
ze5G>zM!9aU3)mLt-DiipC3q$B%sXGY7W+9Y$%ZI7fnwPv!7Y~*{;`=qH6S#uNnzPb
zPdmy!GaCw6R+lfIo<?p5s%;CwT_(I!4ZIp$4lcxktf;9~su<tN6;635!hfh9aE{Nj
zI&ekBivJPf%%6c!E0H1@in~7iE-PqL69zR$b*A={A4axvfc{Pi{)Th^b3<_Piyw5n
zAPfL=d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000000`00000
z0000000U8HX>Fy_k9Uj$Pf!43ZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y9W60kC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.hpp.4E3093BB6320EF1D.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_max.hpp.4E3093BB6320EF1D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e78e942fb045b23ea43014d78f7b469770283c4e
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7vwVjPc$&YvxHL)Y
z2e)2JzW8hTvb!&CE9WLXva@)SaC*~zp2Zb+Iy1Z&O*E~;A~lwa9xX_@A37z6%Tm+c
z=1h;!L(fHD&o;jCS=%-J*2%)u;Ob+krz(HM<~L?<^Xg*T$ZC7}-P6DwU*1pc4;3^O
zoovLFpDeVsXv>21JZt6?iGtjlS1TvJ3OSG`YWwxq{fB2NcFT$X*m`gu<BrWoQyGdY
zbCZCMWhhEbD~8ZHi6FYTD7jc4!~psN27sQ00cIG3fr*9f`$WrT*<HoVj7-c70C=cj
AuK)l5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.cpp.1CE993853CF695D0.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.cpp.1CE993853CF695D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6e2a4c15157147207d56b74a6c0c8bd4a38a19d9
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8WkZreBz
z)&CS@#!Z@S7w|*k+O4roVsu;NQJ~3@G%!VKd9maD`VK8CZrtuejqZaa&N<}VhbjOH
z09RGUNe$AgW>QyA?=EBk@27v2v-_LdPqXrNKE6Hy-oB9M%aaZB<>l)W;QHzWc&TD>
z|0M(W?`OAPKk(w~-?MW0`C(q(EEkh817-Pfu`HLLZobSPKHrr$cZ~du_&A;R#6mQ#
zMbQMQWuT4d+!lQ)tu(tpK`7UDF36R^a*f`i5vxQ-3$n^2jc1%c|D>UnmwtED@Nikc
zqcn!!yt_pkd4*O}1xBvWge-dReEx3G?_xu?&cajlT)fB1`BvsbWGO;GQ}Nn5Sg}Ml
z0`Ef`$k`rNuU@@-3)!*#p%Vl;)Y3*$2g+GC`Q2g1B5DX*imCyPl}zTKMq}qu$;h-u
zWu%vp&xHf&5+?yV?^+k1)!`HS7^y`*5u#5jIv)m#qvL2=I+Y>B_Fyvia{R%`+9r_J
z*<I^WfL`yBF=P>1>A4<9@L%)}Hqw*Sq?lDKN?)g3j)Oxjg}{l2iqL4txh83d51#NG
zVsvrW;Q_ZC8K)#W1mD4~%V_|rt>c9DSSPhUu0w!Ik{JMx0sV$Gk!Gie-S7huleZkF
zhL+32uH}l3$2}Vw=ehoHeFf1$9@Ra{c7Mp*&x)Gt+PbZUt@JiW*SkG%K~@N>R8{C9
z1zF=jq`;2mc9vH686UFwVn@WbGu$yfxJ9{!P0Z7hlXQqH=ZvL*|0J}ETgcH~%OM1A
zTEFM_&hsLAs=e%*tRh(sX8%3bOYdcmQZ?BUtH@LM*n3m}H-5ETO!w~@zjLV$cP4Ae
zvj!{U)S9Wmdss=e=I&(Y47C&|7X`Qiy4bK4R>_rf$uq0TPecknUjo-xf@j3mVth5a
z81a?L^_%3n{ZYVEao&A4+%3T?k!Rld(skI+VM#Vb$q5v7mjt(5Quxbee${}`xFLmQ
zFFo%l`^;=9U|C(hetufABh__<;4Ty1sTSUhE=Lz)MOHM_DOF7G<QnHZ6yZNqk2uHY
zSsl5eV#R+7apuoJsI^Fu48>g^{yr;cQyWGNM|Gz5lOIO5GC;ql1i!<%|G6Q!_{BGE
zKp+Nld2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~K#|<R
zqrz<|0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8aS<kl+y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.hpp.448A5294C15B52AE.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_min.hpp.448A5294C15B52AE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..126304571130584b1617f16d4d618307fd5c8f47
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%q2jYk%@t!V$Rn7gM7^fJkCECx!%hA
zJyUB-n#+_eSy2M@_M9eLtZyr_M^F2EiC5^O$|A$F3ocr0IO$}$R-<p~<FuqFm*-3_
zlz*ru;;IrTAI<QyGis_v>EZ)ZWA7T}C7-+dzKYxFbLb?ijM;C!s;BHLxVbOfF03_5
zOD|RA-f@?;o2o8oZ*OK4+0fB$ePw1rsm_B9U9nYv-xuUn+&Mqnwfu(pj`c^Aiz{=J
yfKFv7N=++<&^d`9y0|F0SRcdy`T+)jUWEZ>7=wX{rQU9S`K^@x`OJ(=%nSf<U0~J#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp.F71A95723F67A7AA.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp.F71A95723F67A7AA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b88cc6233419f2e3b09e95a2adffccace55ac0f6
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TV2mk<hoV8Wka@#f#
z<^N<>ux!^&I->_$ks3{7TazZ^M-B)q$sK_Ji;HCS>$4CgD^}8nXwnCg*ge2r&hCN*
zKmp+D+Bm5}deu&v`uW|3Ea3h0uX6Tqd-rix-p$81CyKYP<oWVs!+d#ld7`+vK2f|j
zv3U5L6%X%ccVGVG#h1Tl<?_?xyu4j5CSz8V<%h+xTz<U$Jb(OjU*6uc=10JX>2v@V
zqIDgLHb^Z4twrxP=u2s(*#-(ixvqCXRtC!zdW%L>iHsIxl}Q@UIDh_0LzS0)d$jO)
zRlt)phTpuqLmPR8R#Oc|R%k*Ny>~u;*XXygCEH}-IeISMQ|){w^C7YnA)u*wZ5>oB
zk&VFn&<1jLP^&j@-o1q!*zVAY271)eMpApoS$6r&VaFnB2pfv31&y^#W^YDg?@`If
zutsI1myyqfJ?IjX0KIpei!bKznPZI9A)g4*Cl#F!2Z*!dY&trXA;fO)GWT-)F383v
zkk;94=Td;d?h!F$5jyF)9!KTB=pC%3C#gv>s#ui1Nx2;RfI13+iHDldXvkcXG{gr_
za85D0xb1PTx11SMk^_Qo)UMAo05#Szp@XiMS|8UTKrP7}6dnWmHCrOhK@o@H4?s-b
za-JGGE)R#6D>@zzXlR}1`oqmNL<f0PcUt!Q9dAD>YO*WqHWoJ0+nn7H_5lmBMyOI<
zqlXk^g9ji54m7v3w0dCtz~-wR5xd@S#|+91$`!0*o|a6~KCjFfO9B5$Xcf1Rqn(v~
z3f#1Qhwp;tMfB8o*|*t5vYgEBd#sn<$sVO@vnMu@r|@Zrr~q#KdbybH-ZOsZQtj_d
zwvuNJD&y3Op+odgNwwnc<lqc-6ekx2xB~jvvKKbVm2=56s>x483O-)~H`jt^#5Q7l
zJ-Qh2mCE&-<huP{z;kileRjB8f>$EXyz`~&ahSuBY>1Kx6iuH5w_H;A({6s&fzY@n
zh2<c<>?j9jHWaX|uP$F69)<35r24)P+-Jf&)xmG0tI>t1$cmPFrHbjjT;ZI@BK(K(
z5wm=*-;pb7HvNYXXZ{d`T8R|NP}~j_?3h8Dx-e=vuQRQm{4}zcgX-6m;a3Frzc&Rp
zK=BpU_aX#yd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
z99fr;ZXwfp0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8aas?$>d

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp.1D44113CF201458F.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp.1D44113CF201458F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..33a9defdc944f8650a6d492b8316830a5861eb83
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5DnTdg+V$Rn7gM5b!c$|MKDIM&c
zzjD*H4IO%GCzW0PCzsjL`TN0yX5G_umUmSjySXIKbXZ)RG<{}PWJ2TA&#xYdoDS*!
zJV9`tfveiZ8P!Zj)xWAm-ty{{Zr>I*=XjL={@12U<YtTTUa<Q)@taQb#<lmg<ySlI
zjnEK1c6-JO^J^ScSN^3jsGboLyqD$QVkTvw`=~wczje%xc>Uj*8((gExR1eA?-*}!
zWo{DCu?$72X~hsaClN#!7bO?#gBU<xKmikr-G=iO_LJm`m>HRv8Gv4g0cIG3fdK%h
CpJDd^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.cpp.DBE748926DFEDB8B.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.cpp.DBE748926DFEDB8B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8621f2942752baf4fe16add8105b1fbb67ccee5b
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TP2mk<hoV8WklG->B
z_5Y+A4H=ePs(6?ou%R|HL*-KQ=n_&JcOgq&U0}%9r)5LPknBUK?2Dy7M}2R}1waAd
zs>(R2L3&kB*45LSD_Ow1=|AP{{^s_>th}8M$7jInXYzb`c3{4|et8CrM`yru6N~%L
z8MuEpyZ!PvZ@zq*mCKJ0^YUi7m<$;x%lC_Ax%_bRdH(S6uDrQp<X6P|>9i*nqIM06
zI!G-8ZAI&L=u2s(*#`<jxu$hNt_+qN^cIa+B{Eu&RVHb?K>u=0Ln|-+{$%0dx`0P%
z41f7_gEsOGt)>c$T%ieB^xpZrZqe^!O?I7yr|7wPkCpR{%t~Y_LO@gT-a1&ZL^cAS
zLmSB19#$`2ym<{duuka=fflv2k<@{5mR<g!>{vt%VMkHbps|w49L#8JJt`TQ)~Jm1
zGO}JckS=i%p!KeC@w+*E;us?}$QmK~q@wfTNO5+YO+)80gxDNh=2?zE1-Z5fq;+=R
zxD=qbdtwY(ghqPq#|iw3-oaLSlA08=ibd(yDYxSgP(vYb;h`cl8gi{k8e+vWo>Po2
z?pr+ImNVm$<bdD@*tNL~Kx^x`pgq<~txx+9pps;cfX9G-%a%xUP{d*QBN3CgoTr9{
z+ry#djt+-C8*1md|1cgwbdaxVhqB+F>h`OqCcCk2XJIG3&Dr%~AGshagjK34^pJvF
z<B>>#1I^<st?ny+tmd;55u4WV#Pr|}<p#Df-<DjYLteRNECu{Gp;bIWjyfxc6nJQT
z$9KW=CVE<X+1A-avYbq}p8BOb*`ri-_QWRgEqv@FDu4&SS}vyDcgEk`s>7SfR`Okg
zm2qmr)G>NkNwwkW<lqc76em{&xB}W(vllkWopZ@EtI1D93O+vq<B{MyV%K6g8e9$d
zN#*`ca^GGR@Kjv(o*mzo;FZXC-ucqC*x$pFY>1KzDAsKfJaS3lZ@c+j2SVeP6qdd8
z{Y2SUW=8?b>iT6_jYr+e<3CVsTL_*q;hk#Wufg@;O039=np&lb>7Cr*oG(TA57Ps#
z@nv2IuBh1XTOls|83?rzDUzYM@6+$9f;Kf_P;*unT0i??WG_eP&$QrA1ouA=1UJ9<
z1{w|@#B+IVVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015sww
zjH!H1V7LNLPykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOn~3p=Kufz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.hpp.55EEA55DB1346E7B.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_add.hpp.55EEA55DB1346E7B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a7dbc3480eb4959ee371f7a02ad3103620b5fc99
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>aD2w
zkUlNfXsh>Sm6XX||Kl<O_k{bO<~e-kcHV^sYr)=3wqCV3AMsN!I}Zt|u36@$ch+*|
z7e<?fE-aT`IDJ+y4_qy^W!;mdT-Qw>n>@Socir{_=HaHUZaJG}Gx=|cYrg(|>h`LR
zX<W;Dly1*(TP-B}V&d&{44)#dG?st$ERf|sVt<uy@4s>jIlXkbyz6hfzDvL1-eGyv
zx41Gl3FusgqSUlv2%VD%qKk`?i}gVapg*91iRIi<yO!m5-a9ZeGBGm%Jr4uSFa`qy
E03tkN4FCWD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp.E1314404E465BA33.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp.E1314404E465BA33.idx
new file mode 100755
index 0000000000000000000000000000000000000000..01f696025a8d50e9ecbe17386bca2040fe062e2e
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TP2mk<hoV8Wka@#f#
z<^N<>uwzwCI->_$ks3{7TazZ^M-Bum$sK_Ji;HCS>$4!GC|1&kXwnw~dk*%#01JQu
zz*UuTQiJrWnby_w+bdbXyV+mm{NeWQ!@Rs(Tu;t`H!tMH>g>Q`b@TcRn2gVWmnN1E
zpEB_9ZhrUqPu_g~dtR<SJ}%1J)pB~xKv}+DuFBPi+fR$fkN4&6JtIFN-p^(Ou@sGK
zQ8Ynn8E7jyw?kh_E6qMo5X!Zk3vz9++@QB;#5$4Df~+!0;|2PcYZ_X6>GvlKk2eK8
zNn`lSr(3j<cW5<LVB{K2$fEbo=XHyI9~-iD7M`Q$<~>!;w=ye{r3e8{#e3^u%@WxN
zd=70OX9rlldiC}V<iL8RGXy%+(neAT%2{^#jk04AHG~~S)quuICUY>OvGb^8WLl##
z(#yzt;Xt~?MS#w`*2S;p@R?(b)FNwy=#z@hha<(=aW*ZT%MfCFaG7U0eivkI6G-dq
zzI7?UVE4orvIwp8+>aCZFM0=C=}Br*%qkY8uTyTvA)uB*;KD;iXf)(nlQhJNXFR7E
zUEFtgz%6ISCCLH7H?Zq+8Gvf*xS#{pNv%)&5TKG|j)2F2e#@3fb5O)#_#+XMx16Vj
zmfORj<&Iuo4{T_h=l;WF4ADWpsy)hnzpLAinwsp!x}Alc^fqTVgni_KtPs|zs?b9U
zvc@Bk0tcGMSz0|*{8-HwCnC0;;fWc*9m)-CW4<l9NQb;~%~%TfPeQABgdFu&4k_@^
z`kwED=S}ofd)YPFM6#Srzn=P~d)cE@P4>hl@-2KCA}W9fzgjJ4{ddma+^WNy$yV}R
zgSByL!_+Z)SWC6x>Ez%HwG^jU1-Jsb*svEi$(?h_Gpor@MG8JY0+X@eJ7Q~bJsw?+
z_(|pdO>^IVFW|Yj>^(caEx{|1@4WM+>u|V-CD{-q7f{q)5<GHA;ZM8ySqDPnmK2tQ
z^z}qJRAxs3%j)KJSxv_MD!27WbzLEN%7k~Sh2KUuqbso{D;nyQDrWa`gA2YC;Xh1|
zxW<=x9l4@n!+!{I;m<&*jYyFU#r=?eUlp{e4Wov$y3qRB4<mayLcgX3zaqH*c_6s?
z#TRMTAOiq%d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zJ3<pZM<={_0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8ZL{MF6?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp.D9B461C00E68A892.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp.D9B461C00E68A892.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fe519a6aad01205ad91e23775551550b233578cb
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%oRYIk%@t!V$Rn7gItFU1YCZ*>aD2w
zkUlNfD0liY&l768f2$WRstET#&2#w7?Ys*O)`Gs7jJ|4dKH;Y_ofA8S%7T@rPn+ZO
zi(y`%5>sZ$!paG=SBCONm46CqeLb&mR?)k<>(LKpub!o(yzS_|O#WNqn!lf)I=!l6
z8rSk3rQ<W)Rtw2yoOyea;Zwxb#`&-2KClw)xPN6|)!+98HZkXH?yNgLwbUj-e#gn9
zNyU}9NkHc^6s4vWL+G4D5M5l9T&xdb0Q~_2K<`2U6N|ut1HZoJsHibBGBGnS!^9aF
E0C8Dlvj6}9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp.C0DCB6010E545A4E.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp.C0DCB6010E545A4E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e5cff3fd009784ebc311f9eea7197b98d326f403
GIT binary patch
literal 1066
zcmV+_1l9XeNk&E@1ONa-WJ!2!WprT#0000H0001UbaH7`0{{TV2mk<hoV8WkavC`h
zo&QNSGB#n8t>VEZkWd@jR5mF;P6?@p=|UsT>H=oJKCKxDG08rJ%D#-$=SY2TO#x5<
zxT-QvYLH&llXdm{_DUA;Zu(a_d$_&(I4kew!|@sL=7l_8o*kGkZ?4aP@#qYAsbcZ)
zIRg*xW_MrS^XALnvvT?AabDgo7n30aW%*&TESDc|KhGaO-IurbjQouFFrD_qLe#E7
zQ3t7Ipsi@#4t*)DH2Xk7DA%+u$d$ozgWjSMt3*Z%vdScl7u>)6rlFOWet**NcvHZW
zG={%?x<MOxhgMSsMy}9=EPC&JUbpD?u_n9D!gKUoyr;_fM&?UoDMCO~@!mREu|zfk
zpF<nS*&bG}UcG$-*|F}@83HY8X(Oov<t&^0?y_SMHG~~SRfEP#CUa1uvGu5AWLl##
z(#y#A!hv*&lK`!Ejf>yZ;WPUfsX@LGqE9M1AC449$I&!&E<=dT!DQ~`_=A&cn?PD;
z_l-*ddc7yckVR;u=X#vLf6+VGN>5UgVpg#z{W|4x92{yW1Wr6ughoToHAzE!@r>sX
zql^0%54h#XI3?L3_zre$P6N=|I!<VhwNmTTJ_M*FnIqsapx?44((Dwm8~#Yd<Soak
zq2cndYq_GsVb6xzd9FW<M-Uz4QSDH+`$OJ-R@7uS*6l3pq_;V`-t8k7WQDLwRfQf>
zkZU{=DX^otou$=%#*f*2aUx>V8t#}L+@ajSHs)!`Njk)pbH-A@e-c{7E##=vatMK&
z)^~j8Ja3|>wU=$3RV2&7bnB^Jx|2OhRcA}AB2VE{?@<BV_|<YT?Y=Yq=29KrOxBWT
z4OYgf4O2(&VI|duyOW(W)KHvU72pbJW6f4rC0EWR&#Wdt5h?im2#iO9XT+|>a5T6Y
z@RQ2*o8-FvQNVL?*?V?;TY^_2&%E=cYq6ihl5B{Q6DZbg65Mi0;V+x{RRcofmK2t~
z^zB61XJ$tM%j)L3tj42m<xyy^2dZrg!F?vYQw_Wx+zhV7ima%qRjQcY%MH$XEW&>%
zA8?j0wL5S{#j5`l;=&(-P#cjV8H)Qp1zl#)rX~z(j_X40XFrW><p}+rGW-tb{`aQf
z1}MG)>ZT$S0CRb5VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_0000000000
z15uteY;tf&vq}O_PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3
zLnaXlNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KL
kAR!_oB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOf{j{YXATM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp.082457B5EE500DB0.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp.082457B5EE500DB0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fdd7fad76fc5fdee8014684fddaf6df22a383f48
GIT binary patch
literal 284
zcmWIYbaNA7WMFVk@vO*AElFfyU|<jg;^LB`%+)}enTdg+V$Rn7gIUamJT9MI1yz=|
z|LX~Ob8}t9`WzMO#Q*y?EL8lS>z~w?`0S-g=0S16v!$$O6(gPAUwOrRE-7f+g#{tk
z!<_S2_Z;ADy%HoTIYmq=w)<P}rjm#IOy@km=2w6H^$F$DwZ}GWJT~#$B6IQk_rmMe
z@^!z8k&gLhxb;O6=a;+3CNoYcyv|a$D!l3Un+bm|SbX`l|M16>y&t~(+!()lUtHN8
zdDq5cLJY-~xk*4LGZdw!6+`HpL=atElw7P2VgUUD13)iB0TWB|-#*p7hZjy@W@KV!
KV1|h^FaQ8Hu5By;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.cpp.FF7D2E3E29E0477B.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.cpp.FF7D2E3E29E0477B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..794d9a1782d07b85bb219ec23a20058c32677b6a
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8WklG->B
z_5Y+A4H=dsTgAf+fep2p87h~WN0*S=xC>eG>H<T)ep?t18IpYnm3^?SK1b?%%LPCI
z;Ht_vsX=;GPuA7byDM41`|021?EdEV<E*@$565S~+ZXbDdA4D`yncNKj7Mj{OBIXz
zFB!OhKfC?<7cahko|Vf_5A*V7xtI(YD9aCvWx4!#^JV_<>8`xFW8`PVhv~E<7NT|y
ziaJOw18qg?cIZoKrP&7xLb;}OL9Psz8}t^9SS2!AkX0sWyx{!hCk?H<^!t;BhwB0!
zr7`^G-3{8vE3}#_Fmi<^WYK%)^LLAWA8WGfEIdWe#e1xrZ)841mLdc+6|b#>6-#6z
z@IJJGob6!s>eaiqkR3Z5IzymEEo~%qq?~1w-yL==qK2@esA|wy$z+adG`1d<j7)1(
zMtT|fTsV?0aT1{Qu5s~M9X_#-ks9O^A^N1E^P#6WI*z8Ha~VQxjwW+2#~++r+XT`&
zyKh_y(CIxfhAcuOJ=fy|enanID?Ld~idn^?^y`$%adfDm5IFHr5gH9S*CY+`!84vi
zj4tk5JmQuk<CJ8F;5*p0ISoK->o}nu)=I5U>ky!lWO~43K)+>8q}eHAH+)aT<Soak
zq2cndYq_GsVaJBrd9FW<M-Uz4Q9Yn+_lLawtf<LutlL@GNpEv>o!dPZWQDLwRfQf>
zkZbIT6xh++&eG~G<9#+??1<R3hC8MMcPKZojd@yfl8$lZoUs(}pM+L%3pqMyIflSZ
z>ks_Fd0s?MYcJb6t4NlEIebs`(g)e2RCTt*D)JOQb{-YLjbAMn)5Ckl?_8?moyl7A
ztij4SwPC9F9#&FqxI5W7Lk-2rRROMmHr8x~RdVHA^2}=T6On??m%w-=ct-463`c{j
z0bi+Hze%p!n*yGS%kH!OZV6tAJoC<%uElN+OR^zKPM}z~NpQ<0g}-d(R}Bb_TT)nd
z(({h8%gl}fmeqLl{BZjZRNEGUyG(ee8u(*yJ-8ApvZAI|sbYF3H#p~^2>+paz&XCm
z>cABhEB;f63x5VeZA6M>DDJ!P4_QH*nlPw2stdKB{V=kX9{N2c_#Mvu&ke!FFa8BU
z^dI7Ld2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000000`00000
z0000000U7tnd87vk0kj5Pf!43ZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y64$fTw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.hpp.774FB922ABF22CDD.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_add.hpp.774FB922ABF22CDD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5f16201beac8847c0854235eea9c5139ff4cb6e3
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7gM5b-c$|N`o=&}$
zXx|)E;2y%&;h_DmE@*Y=?sDrV38y#3w|O$nb9r)^>C>sc(6CdP&W19UC8Dacr_Gsi
zoWXdp3(KSo=VArx;MGD)%8FucUYS?;*^+<%Yu$pgVV)Niwz;m~!Y;`@{omeC8?_cK
zOA2y(SnIPWtWWIAskaFXmTQs{`AXG%zOueCyXd|Czw9}m{@;~v?(qF)+j076I74w|
zZW7S33`MDF#Sl6t5kwakB^T=hX(pB#(`KtW+|7K-%*e#d0Q3_K0KE<a%rFK60{~NG
BUjzUE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp.095F5D6B26D2F29A.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp.095F5D6B26D2F29A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1b63a378e8878b8931211e10238938fe0556c9f9
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7^0{{TS2mk<hoV8WklG`>6
z)&FE*uDh<AbjA<6j@@{g-Swo&?4yPzAq!)QRC)2GUtiF&)>$Wgh$nrJMBtFP8~{`S
zPyo29GEQocUe%LzwSRXb3wS^Mvz$HN-G7{w_w(WSLh<&MJYQaHm@jXCxloKp7mC*=
z7LT8^;_?0L{>vY{`104RTz-0*mv_s>WXOuL{IFP-%a3=T=TDy=%DV^Fd=L0Aop!)N
z)UH8M2dQPCt!Uj2eJQOphd@Cn*R(FkmBDg@-l7q!L`Dm;$|Q|foWK60p_P~ZaJKMt
zTfnn4hTpuqK^u96R#OE=uF!-mdhdMxZqXlNO?I7yee_(s=gRp;=0jvDLO@gT+B#UV
zL^cBNLmSB1POaX&dG{7_V8=rj8fZ~V8%doYXW8XfhaHQkA?zrs8Z=finUfiftw$vz
z!y1*5UPe9_PM}Ln0<_*WF20z<J;xZSK|T?pPbxYedWf^*Y#O?hA;jk7GWT-)Cdjo-
zAg!~9#-#wA-7{jyA~e!-J<iI1&^y>lPg0X&RIw=iI^}Yl0%|A(CLSt6qakxm(hwiK
zz&XX};-STp-g0J4Ne&3UQoA<O0JOG_3GH;P)cU**0V+wRS9lEQw`_?t2Spr)?}3=S
z<vcYsTpkWBS9Cb+&`>+i^@s5YqJuoDM=ksPhPUq(HQ9}II}1DMZO*O>yT^j85LT(G
z&_fDxjXjV82b$YiTHP_;v-xUA#HKaeF`aUUas%6#rzMkg$}4lmQoz3xTE#8o=xF7X
z0ynKc;>X~55k0NFZ0l?ySx)BoJ=aSgWsg$T*%O<{Q~2CPQ~)=AwOmY(?-{>ysZMt$
zTgkHqE92CLp+0(8Nwwkb<lqc76el+YxB}W(vllkWm2=56s>x483O-)~<B{MQv1>6L
z4Q>W}rE>iyxo*D|urIE=&-%M1cqQ`8J72mMyE!b$hA5dpv2K&#mP-nM*v*eR5E{3n
zu<WFl9c9PNjslj|c=YnfM(ziyZ41F|CcIM({5rTD+=vxfQB$i_F@2C5obyzK|FAt^
zlCN|-a7D$Ie;4A)-+@pYks=w2hmL?FFKAN}1~sR3rS*$nM)uOHeohg7MsWXkOK|y%
ze*q84AzT1+d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000000`
z000000000000U70u=kiwXX^X{Pf!43ZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8awJ;_r5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp.5FCEC88A65C41974.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp.5FCEC88A65C41974.idx
new file mode 100755
index 0000000000000000000000000000000000000000..479067d58cba32bb2311f53e1ded648016c3b166
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<5DiHU)sV$Rn7vwVjPc$&YvJZZA{
zU=*5RE4H>i>sIeR+e>O{@0y+Yw(0ynZP0amZpShahsm3gr?wVF1@gFmi*-G@Y{KME
zhYz385S*tHbzkPhxhtzGR?d|PH`|&XvusZL{jXv>(pE3Y5Qx5D{5DRm)%$<k&j``L
zU_;HtAI#JOwcWX^o^3O3kP+Kxu-i&#nIG$$u#3U#|I421+5Wmv=Ev5H`xxu@9lcsy
znVST3EJIOhS}}yqNd(cwMajkbAO_GEFaY!{3^2nO3`{KYInP?RXL4<4W@KV!002)T
BVZ{Id

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp.BB5B12454481FF04.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp.BB5B12454481FF04.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a1e80fa3e1a3ae7b4b74cf3564c06012b032258b
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8WklG->B
z_5Y+A4H=dsTgAf+fep2p87h~WN0*S=b{Deb)dhxp{kAX+8IpYnm3^?SK1b?%%LPCI
z;Ht_vsX=<xOzP_C-IXlh{q*m0c7JpGaaP{WhvPHg?F)InJlil|UcWvA#-lUfrHaM<
zmkiv$pWS}_ix*!%&&uVehk1FkTug=xl;wxTvRr<=`7(d_bXVToG4eCw!*tpc3(>e1
zMH8f!fi|LZTlA&0((D2Sp<LU!AXf&<HF}FitP&Y5$SRXGUU2^MlZIAa`rS#x!*v0V
z(ind8?iOw26<SRd7`Z|dvgp0@`MW{Ciw)U23s2E=@g6JZTbU1$r3e8{#cS(e#S+;F
zybo<4XM0$^diCxtWXJZ0&JgHOOB+cYC}-K^cZVH|s3B}Ass=PxGMR%Kjh#m&Bhwm{
zkzPhV7Y?LLoCN5+Yh8R+hfnNdq!#%^h(4+4d^l1Z9Y@pBxeOt;2a~y%;}1^OHi5Lx
z?pl`u^m<Q>A&byT&-FNg-_SeQNKaCeVpg#zeVuYS4i2>x0w*3SLZcz)nxr8<c*b*x
z(ZyYd2i$UGoRaJid<VNOrva$8juYBroz(iY4go4j<_LHU=r^p1G&@D?hCdQ9dCPHX
zXt_M>TCV7D*t4N=p6d_e5kv=hRQD*`{UL8ZD{8W9>$VoQ(%T$e@Ai=kvO-v;szMJb
z$QqAC3hZcZXK8hx@nbe$?1<QQhC8MQw<y=JiFsObk`8g@oUs(}pM+L%3pv_rIfTGX
z>-YTLd0s?MwU=F!RV2&7?7yda>AmbxswP`v6?qCDdyfj>#;=x(>Ha<AcP`c8&SWil
z)?j6vS~GR@9#&GVxjWf8LoLP0RROMmE;ek1RdVHA^2}=T6On??m%w-=ct&h3hNHpN
zfUi`p-z3-VO#x5EW%t?fZV6tAJoC<%uETx~OR^zKPN1l}B)H|0!e2J?s|JL|4Jj;p
z>3K)lXJ$(Q%W6D&e&nVesIDsncbV``weZK_dT=FHWJN=rQpNO6u5r#o5&lE<fOCAA
z)qyK2R{W<B7yb-{T8k9PP~7$5@3Vq7wPDb3R2OPL`(b1&N9gyI;CDFpKQ{yyzxWrB
zydWWSd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~&vxJW
ze$dJ}0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y!Aj<gw

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp.7B5AB5E79B93DFFB.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp.7B5AB5E79B93DFFB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ae3487e2bbdb03f4c226df304bce5f9afbe45d73
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7gM5b-c$|N`o=&}$
zXx|)E;2px%;h_Dm?$WBz-R0I#5>9W5Z}ViD=knw-*QZl`p=PHt-3?_dOGH&?Uz;=K
zIK$=zE-aHWoQoB92doxaQdSgm^UA!!&zAiAU+Wf}4fDLHuq|}`7IsPQ>HqeA+NiZ?
zSyGVO!|yX(*2?H+oO_$VV40JY$XBZ7^Of~Y*+uX5|7Fkl^#88hb2k5m_|E8~ISj>>
zxk*6BG8Cnz6+`HpL=atElw7P2VgP*s1xzd!DlgUEaH`&8W@KV!0D2t;m|+YC1^^c@
BUs(VE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.cpp.BAF3E66A0EDF3C75.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.cpp.BAF3E66A0EDF3C75.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6c63edcd6e14f289248d7d5fa92ad7bd712e72b7
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8Wka@#f#
z<^N<>uxeFJI->_$ks3{7TazZ^M-Bum#T|hFi;FJ(`YuG#ik0*sn)HDr_8egE3l;zc
zfU7Fwqz36#Jy}=#cQ>+t_tQVi+2h^)$60wlAC518x3A>+@?yh$dHc%+Fdkh1uT?A_
zKWE_a{p|kBAH4YT*Q{KAdYYGa%f)2KKv{lREX(D`yU+8dPY>nY10&xfK1`<_u@JRu
zP}D(c8E7k7w?kh_E6pKL5Xv>J3vy+!+@QB;#43@|f~+!0;}z$xKWS*?r9YfCJlz)X
zEREqe?{3gWUZK@gfsrdTA&cHSpTAr5hgg$cXJH>b7w@@pzLEJ5S&9(QRJ^thRxFW?
z!28e!a<+rjn>X*?LU!zU=mLQjwX~7ciE@@res$Qfh#JC<qN+h-C6hU+(b#%aGBT}E
z8R=!@bKyj~#7ThGyT-*Ab+~6ABQ?k;Li9;R=R;3%bR11XmokLdoJ{6kj^8-Bwh5$l
z_RzQ#pwoM13|WLmdalPA{0F^*t@I={DP|Rm(yvo4$H}3FLg2(hMQAkST$41!2QPRI
zF}iqY@q}BBj8l>wg0EoL<}?7Ut>c7tSSz(YuS0-JlIa1D0sWRWk!Gie-S9mTleZkF
zhK9?-uH}jjhaDSg=ehnc9zk@FNA-xZ-EZ>ty`m<&v2JH!C%w(lb#C`ukQKrzRTX+j
zL9VeUQea1OJ4>s(jQ81mwIgEF8t#}5+@ajSHs)!`Njk-qbH-A@zY|);E#&B^<rD%p
ztv~Wd=Xnu5t-WmPtRh(s=J-9=OCM#AQq|cKtH@LM+<8<0H-5ETOposwzjLWhcP4Ae
zvj!{U)P||vdss=e;qGMT3^f!dHwCx?+E}v{R>_rf$uq0TPecknUjpNi;2E)NF&qtU
z27IM*{U*6?zZI}AuDj3ryCrxf^2|G5x)!@REXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zNiRFfE;Bm{SXSfF%hSmHK(%clxXXlhs)1hzw}Tt8A}eZYl`5tWa)Waoitrz*2b|;U
ztPWgJvEtu_xbkNp)JCL8hT@?M|CklDsR@Icqq<W2#SbG}>7k!ff}i2s|J)E<{Ni7)
z6(Bitd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~<*xwE
zKo_{z0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Z=f6B1{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.hpp.390E8CBCF4E86ABF.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_max.hpp.390E8CBCF4E86ABF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b5132a9be50456c9ce0a9085263c31b0d2d5d225
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7vwVjPc$&YvJZZA{
zU=*5RE4H>i>sIeR+e>O{@0y+Yw(0ynZP0amZpShahsm3gr?wVF1@gFmi*-G@Y{KME
zhYz385S*tHbzkPhxhtzGR?d|PH`|&XvusZL{jXv>(pE3Y5Qx5D{5DRm)%$<k&j``L
zU_;HtAI#JOwcWX^o^3O3kP+Kxu-i&#nIG$$u#3U#|I421+5Wn)F<$zO^v=GcB@D%t
zxk*6BG8Cnz6+`HpL=atElw7P2VgP*s13=Hh05goiz{JuP6kT{|%a2vej7-c70LBww
A`Tzg`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.cpp.778D73C35496348B.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.cpp.778D73C35496348B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6714bc73449fe6abb00099504af176f58dcaf1f0
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7>0{{TP2mk<hoV8WklG->B
z_5Y+A4H=dsTgAf+fep2p87h~WN0*S=xC>eG>H<T)ep?t18IpYnm3^?SK1b?%%LPCI
z;Ht_vsX=;GPuA7byDM41`|021?EdEV<E*@$565S~+ZXbDdA4D`yncNKj7Mj{OBIXz
zFB!OhKfC?<7cahko|Vf_5A*V7xtI(YD9aCvWx4!#^JV_<>8`xFW8`PVhv~E<7NT|y
ziaJOw18qg?cIZoKrP&7xLb;}OL9Psz8}t^9SS2!AkX0sWyx{!hCk?H<^!t;BhwB0!
zr7`^G-3{8vE3}#_Fmi<^WYK%)^LLAWA8WGfEIdWe#e1xrZ)841mLdc+6|b#>6-#6z
z@IJJGob6!s>eaiqkR3Z5IzymEEo~%qq?~1w-yL==qK2@esA|wy$z+adG`1d<j7)1(
zMtT|fTsV?0aT1{Qu5s~M9X_#-ks9O^A^N1E^P#6WI*z8Ha~VQxjwW+2#~++r+XT`&
zyKh_y(CIxfhAcuOJ=fy|enanID?Ld~idn^?^y`$%adfDm5IFHr5gH9S*CY+`!84vi
zj4tk5JmQuk<CJ8F;5*p0ISoK->o}nu)=I5U>ky!lWO~43K)+>8q}eHAH+)aT<Soak
zq2cndYq_GsVaJBrd9FW<M-Uz4Q9Yn+_lLawtf<LutlL@GNpEv>o!dPZWQDLwRfQf>
zkZbIT6xh++&eG~G<9#+??1<R3hC8MMcPKZojd@yfl8$lZoUs(}pM+L%3pqMyIflSZ
z>ks_Fd0s?MYcJb6t4NlEIebs`(g)e2RCTt*D)JOQb{-YLjbAMn)5Ckl?_8?moyl7A
ztij4SwPC9F9#&FqxI5W7Lk-2rRROMmHr8x~RdVHA^2}=T6On??m%w-=ct-463`c{j
z0bi+Hze%p!n*yGS%kH!OZV6tAJoC<%uElN+OR^zKPM}z~NpQ<0g}-d(R}Bb_TT)nd
z(({h8%gl}fmeqLl{4`_-s%;CwT_(I!4g4{<9$bkPSy5A~R587i8=Uh{g#S=I;2d9O
zb>NDM75^#3g+BwKHX=nb6!%^DhpeDYO&HW1)rH#6ei+$G5B;7J{0`^-=Z4_o7ykmM
znIJQBd2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~=9iJc
zM3C*U0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8X&md%;~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.hpp.2C7338E3132300C6.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_min.hpp.2C7338E3132300C6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2307d370468e6ca5c1f2a5b12536ef5c165d3d27
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%%wn@k%@t!V$Rn7gM7^fJkCECJvDl}
zVg5>|Y&RFr)yr5vexK~%ynC~3(T3EU`gaus{gtFH^GNkF&(@i8x$9AcWuBUw-C0Y?
z?}B+wfi6N{H?mHczaq3V$eb(Y=2hLtX3y^YUAJAqBwlr8^R}Dowz$9Kp8mG_X>#a_
zrH_K#9@hF?(OS*Ex5!$;fw8+{hx1&{@Z(H;t8Xm5{-5u8+ws49B1_t9nRj*_En+CH
z%uND1mZ2y$tr$Y*B!cMTqU2(I5CiB77yx<}3Yb`!7j$?8PBA#c%*e#dzzh>-U;qH7
CRbJ=-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp.59008E97BEF36435.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp.59008E97BEF36435.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e0a9743cf2fa00ff00eebde6bcf74fb66c4ec618
GIT binary patch
literal 1064
zcmV+@1lRjgNk&E>1ONa-WJ!2!WprT#0000H0001UbaH7_0{{TV2mk<hoV8VLbJ{o%
z_W$WD3JvAjo8d!3p=GW~m|i>i=n>Ky?;uM)y+F8M-<1uap}h}fdLL|S_mTGH*<HB+
zC;(hl87DPJuj<LVdU|&y3wS^MtDN26+<uysxAWonO!4-GJYSw|m@ltipDD(pGsR02
zi~BEGasPgH`}G4azWzNcm!BWz<;`+28M2}*KQ5N#^3%<i`NQYC^5%{;KLb8ar#-L`
zwQEq+L24OjD_XZhUrH;@K2Q+KHLVMBWw6|!w`jyFk<o&zGD+hF=Py5LXyv8fpDa9F
z7w{;J;WzJY&_-UN)l`9zD>NaC-aDVaTlD)_lU--wDS9s6W957!^C7YnA)u*vZ5^yw
zA{&ACp$+70uU4;Ky?YBeu<p>A23pk8Mp6gJS$6r|VaFnB2s?_Z291?W=3qu+>ru(b
zutsI1myyqf1LzWy0Ihe8i_hlpiDQh^AfE`)Cl#F!M~JiIY#KV3A;jk3GWT-)A;`5&
zAg#0e#-#wg-4kNSA~e!-Jx<Dh(L2~mPg0X&RIw=iI^}X40%|A(CLSt6qakxm(hwg!
z!#Tz1;=aX$-g0J4Ne&3UQ@b|P0JOG_3GH>Q)cUjz0V+x6sPGujZ`l%Q4vIJoe*|Lk
zmh;rmaCtbiT+!jMM?>vA*B{0shz|0oc3SrP18+YoYO)*ab{2Nh+nik=_7MxRLRh7$
zLJuj(H6DQ!IMCeA((0b^BbzUFL~L5a9n&j!C^xW;d0H|_hrBXpECu{0p;g>Mjyfxc
z6u4=9hj+pAB6?bT+1A-avYbr!J=IHhvPY@v?1@d}DSYfBDu5fmS}vyDd&civs>7Yh
zR`RUD$~d)Q=ome$q}p(Ia&U$kij%7XTmfyY*$bQG%DLnj)#N841)nd0@ksED*tHmr
z23G^VQn`MUT(>_8cq%Tt&yIIX@Ji&FcfNEj_H$U04N)?IV%;XeEteGjvYTIZAT(}C
zVcAR1JIbD!9R)0_@#y*CQD|NdRNEGU`%HMJ8hA6f9$bkPSy5A~R587i8=Ui4g#R!;
zV3sfSJ8(tCrvDV;!XJWA8<8Rziu;~|jv2J634@yRy3qRBPa}Ids(w!yen)WsdsA=&
z6yE?66CyEld2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000009~
zBnPfr65D{$0#8r?Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L
z5ei8{5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933Db
iA|xdzDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8Y?b;a2L

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp.4329A8B0483B68A4.idx b/.cache/clangd/index/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp.4329A8B0483B68A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d60fd8507f012b99f23393f02351f79e8b236e41
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%;i9unTdg+V$Rn7gM5b!c$|MKDJdRl
zoWIiP^~Q|YC{^9d|KvJ5JFN>&9GsTAUxw}E4xx(?X%}A1x$)F-tImc4rz%SeyHeL3
z{W(ogZHCh1j5D^4k6iY8PTLmr=-9%wqQ^I0JN&+icja+4SLQ9}|M^x=o4s-EeC_<z
zj(a0CM31dE49s66ShBu0nL*V?MJXrCzr{?-LibVp-hbsg3}Vjx&fNI(%b{8(H}zvO
z48@hXNkHc^6s4vWL+G4D5M5l9T&xdb0Q~_4Oe|d9tHLhd{b0w;$i&P5^gIkO!x#(<
E0FaPlXaE2J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.cpp.509D27379633375F.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.cpp.509D27379633375F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e26d1ea370581c908684f4c956bca140c399daee
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TJ2mk<hoV8Wka@#lz
z)&F#0#!Z?{cg7EiYqy?l5>L0CJZdNsvM{DdEiZPwUtdtN>&ERq)YE+ti38$x08#}&
z0pRM|IH^H;)l63P^V<tqz`N;R<?P}1?&GYyn~$$gfSVWce0g$UzPx;W0$g96054rE
z9zJK_;oa=+%X{8@`FmC_KRwRN+vQ?1W}qxTESBZ+<L&49<EQ)b_MVZSDLzc6gJK~X
z*P>{G)H2XUbZ(2jlvbKupdgfMI~QbSuw0|JXhfCBXhBw)r11><^DPZkUi#h9!{cQE
zPtq9v^63_B<Q-Z~H5gf;30d^s`Mhq>?_xuCm4)Z%xp_~u^R3KAWGO;GQ}f<Bs8}K!
zfzP20<m>>eSFhgQKu)YTIzgaAEo~&VS2@cezcY3$qK2@gs2b2%%VhR$G<F`9jI7qE
zjPx?HUDzvKViKV9u66O7JACFGBelpDA^N1E^WmW4>bRPgPGtzO-G|Jx9Dhi1WfMs2
z?5=ewz~J|&F=P>1>A4?A@L%)}Hqw*Sq*$w1lzx?RJN5~+6ao_uHKEavxh83d4Nvr3
zVsvrW;U2eK8B>xIg70A0Wg381)-j<2)=908`w*a(WDbDGfPTZCNOMxeY50R8CU3b;
z4K25aQ_CG4j|XjNoag?-^%X=1`KtCP$Nhn~pEWhvwRKwyTj_1CZb<u}3$jM2QeC5m
z6yypIiWE4}JkHYUf$;;IFHS^kJHrz*fLoMn*u;EWGD-WoGG{CW{3oGRJVK6oFZ&XB
zXnn8mljlwJwDPiRvWsN7n0`I>OZT!zshS*#UF2K%G-Ol&4}QH|O#APQzqwWWH<P{O
zy9SkUYR#%c_E1T+=IP|*47C&|7X`Qiy4Y|OcFCP{$+K3IpNJHEegv+s1m6+66634U
z#fYC&?%yQ$?T-SUi}T*I!`l+P68X+MU%C#5dsvbUQ8IyI)g{3rmlXbTm|qPbG;T;?
zIY{45lmjzc3XI*9bon|jJp3cob%o#=6W*y7-i$6s7os978tRlPruTA<bG{PcKOB#k
z;d5e+Tv4;zKZQ8+M<CQ%q)3M1ZYaHH1Z`@=sNtH<^nUWQ$Wac^@1g&9B=^4#12?|-
z3IR<X!gG0TVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000008jt`
z0000000003QF-vW!VwZ~Gy+dh0Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOqa{k<p2Nx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.hpp.7F575A4F4970CC92.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_add.hpp.7F575A4F4970CC92.idx
new file mode 100755
index 0000000000000000000000000000000000000000..96731ede63f35a1ba63f0eb12a7af674bb49e871
GIT binary patch
literal 278
zcmWIYbaUflWMFVk@vO*AElFfyU|<jg;^LB`%w<6OKTxh>&er~ee1{BpoWHxeG;u#l
zU$n+lC+d2b@9xa{{u?Zh>?~Zj>HJ>0JNrOyN1I6_o3Q>e<EL8|Y;fURy>g;=u%B8j
zL%)V#qsgudHYa+otYTZMUAgj5=~)Y2%kurNbw8X9^_?hqJH`E0^EcO1zxICGrnPA4
zA<e}n%$8kT<-%U|WSen=O<L)L*jk~OcBYTrw-#Ri&v(2kb?=;b_dWLm<{O?m^1Zk+
zHwoxihN9H8VhEj+2%?LNl8g0$G!si3$N8(}SKITM8JU<Fm_Tw60Q4RVFvA!O3;>t-
BWSRf~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp.BBCEE5A5D01A00A5.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp.BBCEE5A5D01A00A5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..de1ee93dc10c407c9bba1071e642be945fea160c
GIT binary patch
literal 1062
zcmV+>1ljviNk&E<1ONa-WJ!2!WprT#0000H0001UbaH7@0{{TJ2mk<hoV8Wka@#lz
z)&F#0#!Z?{cg7EiYqy?l5>L0CJZfkXk}#%7EiZPwUtiF&>c;Ip)YE+ti38$x08#}&
z0pP02IH^H;)lBN@`R#=);NA4Ea`te0_i<L<&Bxa#z|9MJzC1ZFUtYdG0j{r3fR`>7
z51%ve@NRbZ<vnk{{5>m|pC0Gs?Q$_0Gf<Wv7Rz$^@%Hol@zZ^Id(X(v6d$J3Ua=63
zYf&^oY8hxFI=4k%N-NDSP!P(soeOehuw0|JXv8X!(SocpN#hyz=UW<DdFgjY5094x
zJV|5t%conkk#}e{Rbb=_O~|76&gXT5eis|EbrznZ=jJ_C&bKlfk);R$O~rfbV8s&I
z2z(B0AZL46y?XWb26AG1qZ0%=)Y3*$2bHrN@;hV4B5DX*imCyPl}zT~Mq}qu$;fJr
z%1AFG+l7PDB_;tn?^+kXxx;79F;a_c5u#5jIv)lVSI5<~bSguL?IC2I<@iIAwM`(c
zv%A)%0KMO%#*js5rRRPe!GF;^*ho)OlVYu6QTjUNb{rCFDFh}SDng?nb4}6^8=mO7
z#OUI#!vk))GNvRa1mD4~%QOJh)-j<y)=908`w*a#WCp-vK)+#6q&X?#H2k25$y=^d
zL(A>q)N)71<6avY=ehrIeFf1$zN&ka<Nm<g&zhR-+PbZUt@Jim*QY(`f~*i$sjARJ
z3bMvQkpd^0$5~q4Gd{5S;zY!@GdwXpxJ9{!P0Y6?lXR#nbH-A@e-c{7BjjlB<xm0-
zt>5eS$@3<9s=e%*>>^n%X1^Z$rT4N&shS*#UF2K%)Mr!x4}P^=O!waze{-u2Zzg-m
zcMVp?sWq#H>|rI<nx~VKGt^R?Tom96=wicB*d=$)CC^$-ej-xv`4PCj5`0H&Eyh=)
zixEGm+`mcg+aCoy7w5fa!`l+P68X+MU%C$adsvbUQ8Iy|?vmh<OA3EE%&!Iz8aJe{
z?4@rf%AT1m1;%blx_q4%xv590>k7d$CcINEycu1NF2stgXsA=FnBL1Z&iP7&|8P8F
zhR=yPaz({%{}kfPAAwM7ks=w2yT0^2BWP0_Mh(|=ruUPdMUFB+zlZ+ck=*|}4BYtQ
zD^Qmo0dsk6VgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000015t(8
z+|buMD>nj9PykQ>00000000000Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXl
zNkS4A6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_o
gB_}B=EH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOfoCZPyhe`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp.FDD16FEA381DDE62.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp.FDD16FEA381DDE62.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ba1a25667ad2c40c83426eff545e96f751068f3f
GIT binary patch
literal 280
zcmWIYbaNA6WMFVk@vO*AElFfyU|<jg;^LB`%;iA(KTxh>&er~ee1{BpoWHxeG;u#l
zU$n+lC+d2b@9xa{{u?Zh>?~Zj>HJ>0JNrOyN1I6_o3Q>e<EL8|Y;fURy>g;=u%B8j
zL%)V#qsgudHYa+otYTZMUAgj5=~)Y2%kurNbw8X9^_?hqJH`E0^EcO1zxICGrnPA4
zA<e}n%$8kT<-%U|WSen=O<L)L*jk~OcBYTrw-#Ri&v(2kb?=;b_h-l3ez6!u9pzyt
zuFOpWI+vj+HLVyz=OlvY;-chYeGr2QM1TPk%dGaGrjlD{o-#8sF*5)?2m{P81_J{C
D@dIRQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp.C6B6423A4D3C4156.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp.C6B6423A4D3C4156.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e87b3c98f73d8d64cde58f5dac71166c19dd6150
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TJ2mk<hoV8VLZ`(Ey
z*8eHajF&7+2lzwc+N-f9F}fD{DA44|I+!9gzW8Ooen&~J8@KyVqXPtq_dMjiA5{Pp
z0IsTxlNzK~^<-V`-(JZA-cA23XHWMJA7<sle0+Nb+`W?L%d-vh<@GOT!0pW$@LI*<
z=~D)t-pw99|G|sTf6dC}$LD!@zg$el43y>j#j;#}xc@YN{`gqlKQi(?;{9~m6AMwh
z21OmDmVvgSbvyK>w9*^`1)*Hix*%5u%ME&qMywJUEyya9G+uE2@{@*EUi!mH!}E0k
zFVYx(^X>+1<P};?6&Sff6SC;N^ZC0)e~2~Nbr$x~bMam(=Np+1k);R$O~q^LV8s&I
z2)qw%AZL46y?OKY4zgq2p)&+p)Y3*$1LZ86{OYh{5jBJzMOA~wN+vU?(b#%aGBT}E
z8R=!@b73G|;v_)pUE|^xb+~6ABQ?k;Li9;R=fjcW=s22&&SeO(8BFG0j^8-Bwh5$l
z_RzQ#px1k13|WLmdalO_{0F^*t@I={DP|Rm(yvo4$KX&yA#mcMA~YIuu1Ol=gJ(R4
z7+pNHIN+8e<CJ8F;49d*ISoK->o}o3)=I5U>ky!lWR8HxfPTxGNV8MKZulb+leZkF
zhK9?-uH}l3$2}Wr=ehoHdjruy9@P$IyWiyPdqqulW8KcePI{Z8>)k$bK~@N>R8{C9
z1-Zr}kperK+gV!OXZ)DWS34p$t>KR8!5zvCY-665oTMSHoHLdJ{+-Y&ZXrjVmLUXg
zTHo=V^Sp?j)?T)CR*@_R(|u3%(w*#4sybU@6?qC@dXEa=#;=x(Y4@J-JC|y>Gg(WX
zHCP#^HcTD8hm}+t?oM{jP(yKYRe&p?jWt_gm0UTOJhPhoM5N&JC2)Hqct-46jBiF)
zBfe6(ev@3c-wN0lm)&Q_yCrxf^2|G5x)%F6EXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zOJ8=BeP(tP7`q!D*~tA!wQV7|$AovPfnP`0qbso@D{5+$DyENegL58<@E?jtoZ(As
zj$Bc(+TVq^@JArjMx;oF;-L?{%Lv-kgi*~gU8w!+XOXQOp`VlfpW)p9+6-Ly;$Jp{
zAGH8;d2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00003QC#{UMyJCdgaS`c08an_
z0000000000Pyhe`000000001EZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8aWZ_W<@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp.F8ED464DDA5BD732.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp.F8ED464DDA5BD732.idx
new file mode 100755
index 0000000000000000000000000000000000000000..408cd306300fd35cc88bfa05b89b7162a10f2e63
GIT binary patch
literal 276
zcmWIYbaUfjWMFVk@vO*AElFfyU|<jg;^LB`%*8<ZKTxh>&es03e1{Bpnk$uxT=)1)
z-P-2%dj71k#rgeP7$4b9DB75MGv3dfbBEBx^MX&N%C6Q4`P7wVQ7ED1Wp;eh)a4Au
zi`7^rr8t)>SO@0|E#WP?b7M<j!MCMz>aIsW+?6|Nvg513{w?e?wXT2M`)Qlhq-R-0
zj!$!aCYkw(mFU-|Gg#^*C6;~lobgQi#@a>gasRF7Og#PC;#0D7Et9F>Q4faV%G@NN
zQyGd<(~2Q<P9lgdE=n%e2QipH1Q;-}9NJlyW41VaHZvm=GXv0bFu)9BFfafBM{r&(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.cpp.C1E70E7984041F16.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.cpp.C1E70E7984041F16.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9042dbabf4119f2a462340bf8e4513bdce1c01c2
GIT binary patch
literal 1060
zcmV+<1l#*kNk&E-1ONa-WJ!2!WprT#0000H0001UbaH7=0{{TG2mk<hoV8VLZ`(Ey
z*8eHajF&7+2lzwc+N-f9F}fD{DA44|I+!9gzW8Ooen&~J8@KyVqXPtq_dMjiA5{Pp
z0IsTxlNzK~^<-V`-(JZA-cA23XHWMJA7<sle0+Nb+`W?L%d-vh<@GOT!0pW$@LI*<
z=~D)t-pw99|G|sTf6dC}$LD!@zg$el43y>j#j;#}xc@YN{`gqlKQi(?;{9~m6AMwh
z21OmDmVvgSbvyK>w9*^`1)*Hix*%5u%ME&qMywJUEyya9G+uE2@{@*EUi!mH!}E0k
zFVYx(^X>+1<P};?6&Sff6SC;N^ZC0)e~2~Nbr$x~bMam(=Np+1k);R$O~q^LV8s&I
z2)qw%AZL46y?OKY4zgq2p)&+p)Y3*$1LZ86{OYh{5jBJzMOA~wN+vU?(b#%aGBT}E
z8R=!@b73G|;v_)pUE|^xb+~6ABQ?k;Li9;R=fjcW=s22&&SeO(8BFG0j^8-Bwh5$l
z_RzQ#px1k13|WLmdalO_{0F^*t@I={DP|Rm(yvo4$KX&yA#mcMA~YIuu1Ol=gJ(R4
z7+pNHIN+8e<CJ8F;49d*ISoK->o}o3)=I5U>ky!lWR8HxfPTxGNV8MKZulb+leZkF
zhK9?-uH}l3$2}Wr=ehoHdjruy9@P$IyWiyPdqqulW8KcePI{Z8>)k$bK~@N>R8{C9
z1-Zr}kperK+gV!OXZ)DWS34p$t>KR8!5zvCY-665oTMSHoHLdJ{+-Y&ZXrjVmLUXg
zTHo=V^Sp?j)?T)CR*@_R(|u3%(w*#4sybU@6?qC@dXEa=#;=x(Y4@J-JC|y>Gg(WX
zHCP#^HcTD8hm}+t?oM{jP(yKYRe&p?jWt_gm0UTOJhPhoM5N&JC2)Hqct-46jBiF)
zBfe6(ev@3c-wN0lm)&Q_yCrxf^2|G5x)%F6EXjr_Ie}u`Cc!P26#lT8A2lE}Zb@O;
zOJ8=BeP(tP7`q#uM(#(dZ41FICcIM({5rZGU5OQ0QB$i_F@2O9obyD4|Ij<)1Ycls
z<cf;5{w~CYzX72(B1JM34}IWWKG3EnjA{<)LhWb2ifrWw{hZ|g4CnsOR^YN1{{q~G
z9+Ciad2M0<0001TWoB~#0001TWo%&p0001Ua$|EN00000PXGV_000000000`00000
z0000000U7AC(SS2LO`zqPf!43ZDed{0000>W=#VG1qKHQ2@4Dj5DyMX5k*5L5ei8{
z5*HH`7#TP@J3Ku;KR_-OD>@@KC?66R6BHO3I5|5!Jw87`E)^>}BQ_Qr933DbA|xdz
eDJm>4FflVVHy$(^EiyzE7Dq}_0!B_jNC8aTFv!#Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.hpp.1F0CCA2641B90428.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_max.hpp.1F0CCA2641B90428.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f90de1f4f13d29eb5e554c3818126d85df4dd32b
GIT binary patch
literal 274
zcmWIYbaUfkWMFVk@vO*AElFfyU|<jg;^LB`%tb)@A5gAh&es03e1{Bpnk$uxT=)1)
z-P-2%dj71k#rgeP7$4b9DB75MGv3dfbBEBx^MX&N%C6Q4`P7wVQ7ED1Wp;eh)a4Au
zi`7^rr8t)>SO@0|E#WP?b7M<j!MCMz>aIsW+?6|Nvg513{w?e?wXT2M`)Qlhq-R-0
zj!$!aCYkw(mFU-|Gg#^*C6;~lobgQi#@a>gasRF7Og#Ns$=>XedVb!K=f#z|NkE4(
x6s4vWL+G4D5M5l9T&xeInOLkuS6vMBV?D#n$i&RR1d@XQpx0o48OC5>006|LU9kWF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.cpp.05E7FDBEF4E2FEA1.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.cpp.05E7FDBEF4E2FEA1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d640fb62c9cd3345aa5df9014a82c6bce11a71b7
GIT binary patch
literal 1058
zcmV+-1l{{mNk&E*1ONa-WJ!2!WprT#0000H0001UbaH7<0{{TG2mk<hoV8WkZreBz
z)&CS@#!WU&7w|*k+FN6r#OMWi6lii}4NQ?*UhKGEze7o`8@KyVqYDTU=NxkGLlpo8
zfU7Fwqz36#Jy}=JZ!cs4@1}p3vxnQe53}-aKE6HyZeGaq<;jNm^78cwaD8<Gyi~Dx
z_?&@<ceA@MfAQkWr&+oD_&6_bmy5}mfwFwRSeDBVx1Z;aAMeZCdq#dnyq`{cVj*hR
zps0h?GSF7EZil{<R+@dFAe3ub7v#!dxj}Eyh*ct^1zBa1#xu^Jf6~y(OTRyAc)Tp&
zNgBg%-rb;$yh5v~0wY&wLKeMuK7Y69_pv6s&cbu_T)d~s`9|hLWGO;GQ}Nn5Sg}Ml
z0`Ef`$k`rNuU@^qf$Ug!=mdclwX~7cKsn1MzdP($L=9m_QPrTalF1BeG`1d<j7)1(
zMtT|fTo_1~I0?{t*SPpi9X_*<ks9O^A^N1E^Wi{obR11Xr!s`t3?_3g#~++r+XT`&
zyKh_y(Ca-ihAcuOJ=fz1enszKD?Ld~idn^?^y`$%F*wvv2%LDR2#toEYm$ce;0ezm
zMi=)j4!GsWI3?L3_zre$P6N=|I!<VhwNmTjIs~XBnFHW4px?44((Dwm8~#AV<Soak
zq2cndYq_H1anFX@d9FWPUqN(`N3}!Q?hkqUSy7YSShusVliud&dbbZ;kQKrzRTX+j
zL9X#Yq`;2mc9vH689!w6#g2$gYq(>2aEEdO+nA>%CuxW)=ZvL*|0J}ETgXwTWe9<r
z)^~j8JTIcBwU=$3RV2&7bl+pWbSHb1s?L^JMV`W^-lGDz@vG%x+P!D|&ZQdeOxBWT
z4OYgf4O0j2VI|duyOW(W)KHvU6yOSIW6f4rC0EWR&#Wdt5h?h530z+Zo)Nni<Ezod
zh_6(x-z3-VO##ovdH31jZV6tAJoC<%uEl;1OR^zKPM}z~NpQ<0g}-d(R}Bb_TT)o|
z(zhLDpP3y6#%_kEAv;oSTL^A3;hk#WkJ07mLafM&np&lb>Al?GoF^juhu#q<_zaUH
zS5&O^Pa)3y4G6UnDUzYM?*s4hfi^W^RC7pYYCri^WGe^g_ay&!IQM_H0++q`7u6gd
zjB|NyVgLXD0CHt!a{vGU0CHt)VE_OC0CRF<b0Yu%002(_000000000008jt`00000
z00003Q5=VF9h2-%ssc|?0Ap=rY-j)g083^~0|W&I2M7rZ3=I$u4oMM3LnaXlNkS4A
z6BHO3I5|5!Jw87`E)^>}BQ_`>5*HH`7#TP@J3Ku;KR_-OD>@@K78@KLAR!_oB_}B=
cEH5xIGc`9JG#V{3L=_fCN>TzwPC-ZkOm!O1egFUf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.hpp.078E8B6FD22E7DA4.idx b/.cache/clangd/index/device_reduce_instance_threadwise_i8_i8_i8_min.hpp.078E8B6FD22E7DA4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..759a6f7d178b65f0b697f386cdf635d2c5a8490b
GIT binary patch
literal 274
zcmWIYbaUfkWMFVk@vO*AElFfyU|<jg;^LB`%tb)@A5gAh&er~eT!#&KoPWE1E`5{U
zG&^H;=A#yi%$N81Z{@zZ$6GT|;L_)U=^fI7zLt$@j}NGEPu;V~Dy(mI=fv|XQyv>0
zI=iCFW0BoAhEK*{)dFw1HA?T^CK|IWsO|pOu$FhTQ%Wz)d*f3*Wn<CZec|`Qnxod~
ziHF=f?y`1c)rG}hni)klbhKMvQGa+V<Y1zx_19nbAI^!1pYBqANj*R7$kXD=+$5kw
x8H!TViXn7PB8V<7N-ow1F_=IE1OPn)1I#c60~5=vZ+#12%{}#*nURT^0RYKzXxjh)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_multiblock.hpp.C0B69E5EA2AC7CE7.idx b/.cache/clangd/index/device_reduce_multiblock.hpp.C0B69E5EA2AC7CE7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e7dd55f72e311fbfccc807785c364b291874fbbb
GIT binary patch
literal 12110
zcmd5?2~<?a(w^#j6&$a_s0;%LvRrmhgb|cQKz4A)715}uc^bvth`|syqJSbUh;iT<
zO;n6RTrh$V7eq7|1vf<H`Rh|$F-Aa)ar+yXqVGGN^WJ&Kb8;SMxP8B=n!eT5Rn^ti
zH!31DG}Kil3z`xUA2)TuBwi+yb*BIG7tEU;%E@Gl1eq*8>-4DT*<Ia~x3+E2i+6U3
z@EV;HHNeZ-A$M!a=yT5p9~xEsQ>~~}zRy0e=+I~*<<6mwk!20%Yrm=Zd_(5TV}mqD
zu3Ap-u(8m0XWA!s>R+F{ck4)Ws+p#0g(Bu*O;Pv8zIWYMxt0#<xo2G1)s3MqE@oz?
zC$!&Pyr%N9=g@XpGkRT_Yb(Cl7gI9>Y{OrEb933%hEKNpdnBwsUu&|dd}h$ePEBL~
zbgzr$Q)XQbyB_7}R@f`@vTEV)>#I)WlgiNveGFpuH6Dt5pEk?#vj^_CuO)oGYS4ti
zD`TIudq4K@oEuHQ*Zt5dpkZ`LjNY!p(_;R7XKyrQm7CFW&9vu#JEhOPS-hg6wq*a1
z6It#>2L?@8vwdwt-7rq?i!s+uuG^3)uJqZmW^V3xPb=#tUYa_l=<$ZVcfD?p)HB&V
z;`pYyss6T?PEJ`Dc=(5)l(7vRZ{#P$>0kb#Y0stD*EZK1W|SR&;IX{X(d~z4KEm`l
zo6NKs2HArKPg?POuVeP&JMlF!^+w~2n@UW2j*d9)7-6_*$%(6>xhHMr7yUlg`L5;c
z9mnn6YE<gE9e=q#XVA6%nqx!9#U|`f?9INvd*U~DOlNez{PVE;r4uGh_CFa?ocCR~
z_BY+0eBD1S*03%*z4Gs-gKKB4E$e@@(S7@x&eQgUZL;g2I{K|m^zx4X7yGu#;m5mn
z)!ygK!uJZMRh!No^!4;Sm&mPq`Mzvc(Wkfi2K*zW@pzfl)ahG|eps2mzxVFVv!inT
zR#=UQI&ipV*T9Ui(MJsfDhDhNFPu{HApM}n7h}VU?=CNj8*6K4YH;UN;FFt*2`7Vx
zx<=0((BbiKwI0j%e-)JB{@zXH7;xKicAv>XYZvVNvfuar*f(&>nXk^C8PMa;hz?~=
z8O7R~seKL3&FTJb>bs0Pi{GS-{?+BjgL;1qmhZoP*8P*5DM3lk){UQ@xz&2SThoa<
z&dEO+>^gMCp*CvvAnl`fb-cf0>=omqZ##aKwyN8<<MpnA)!u>CvBf!y$R~m6b6&x{
za@FQ<jc<ARPU~OF9Z5cZN>4WOYQorVSANv%+*hx7b?}<1i$1kE)q9>-L^RZ&AKx!2
zc=0vEW%Gu&D<8Spa>0gCZ+wlceMfjizkW1hs_T&><0I|;Wn)6W+ckf2T<no{GMW7E
zOM6coWM*YVxNb7B3iPW$scU5V)a&S9$M+pb$@hJu6r?X@`GqKdAvU@OS(I;HI&CYT
zxlcv9n99UJ^FWuD3zF5`jUV6s7W2E37@En%Tuzb8b=Ea)EOmb<$L98&qaaT7As30k
zMbb^zAd5-vj7?KVdChC>rHUx3NN2INK^F7e-c}>mE6=D%M=Hk{^D!<h7bIK$<5bk%
z!U1c#5G5tYC}YHy3zEgoWA&<&S5E9gI#IF~lv)sV4YD{T<Sp6yt2$aiJgJzci24+<
z)iub{b3>;Q@#P_1RirEJCD=UJuH}Mc&!r@<!FBQk6_HbNw0X2k%cYInurP0z&*!_8
zM4;pbj&I=9x&~SNDsmq5a~n27McPxA0Mh`6mJ5=D-z6WP;`%Aofj1>bnnpUbT#y`5
zr^t^EJomMN^rj;yBKAeZUDqJXK#zW-U)ILU6vUOXoFwKaiIc8rV>vpf-Lh}LE>RF?
z%2GhA3y6!ZL6$*N2Yde(VtztFj3~>`#PDZgs%wzt6Q|=}<VKzgQ<2V8)DZI!yOs-*
zV<(ZZr6cN3Du{;maFo~|rJ}Yr$TB4-dEkYEWyuPnrYy&Z;W1*OYmnt{sfV8beBb-5
zf_PDub3}EHIOrN=nNMo_4j<BGn1XmvmJ*^WA?CUUSrTTrFSML~c7}rVpe!ed=?P+`
zYmjA0yp?)CH{y+g^rS4OiN$H+rfZNTu}7zMUf=RA3ZkMcWyG?KXmky-EbF$!K|7j!
z)1K&4om?RuuaGXfrj4a=^0Yw@<>wW|nD%m>2<M5hu0fWxr~yUpd8T1qh=TS~0LlVz
z)iubHZZ`ew%XzW06vUDCQb7zVh+5YmOZMD5oowfX_EVB}w3iHCnZcXu8f4iue$$1F
zH+c^fq!;bwEK!~%_PPdHcJH~n|3?F-FI1!t?IqMS)UM@%<iq#3UH$%-`%e_ah4xTL
ztP80rwKm9dtj0U@!suJ;I}$zGgP+__-Eu+lX|ZBY2mfn36vUnOaDfOH=-jk6$Wpe=
zjX1y5{?VTBCNj}U&&j;yBI`haMmP}Ry8o$Lf$cJD+Y^pH;tF)U0$p_t_Hlz-67O?q
zA?ZRoP?mg7na_#323c-)S~}A9arqi0F`z8zyj41{(KX0&FW&vRt@axaWlJwlxeiY`
zGqJTnmd2mg_|5<Fz-t9@qvI(js&eA2Ymh}YZ|bM>eQjkjZeD_q*QxvdL=gQ<{j5+^
zg7}vBy(J+SGx?;eoLkn*fAWIZ2)!GjAI5z_d;!icpf|%!;L-#h3@;_#OGyyo_C+3+
zz9D%(>kHy%E}w-xbhr4$=B!19AkI<GamV;7h)Y4e6eM^(*sq5`M&Ah*JHZ<Rt|0CQ
zyZz9g;S#VY0dKToK|Bw_d9X)}!hl@2RN?I5Qbo0~fe0IjC34Dj3l=hV>a}|7XBloc
zI|<^X{*%I(V3Ye#mY{jV=S4Hxl7UO2*nkrw5{EF93<1dy%6L;CAVrF|LEttBX9L_0
zf!n2cFZl0;5JoSCz+wo;0Sn?W2s|dmXTkF<_#)lF_)1`}%eQ9e@#2!8CBY~bPba%a
zBlJWqG~=O-FpS|BVDkcOr8w*bgfrX(woPEi@KT~#O4^J?`Ss+S<A2`h-cAtX9OC@(
zv4R-y8XwBgVwc514CR1j4)kFoJq?zpp%0D(GYLW7l0l%~3Sx*;h!?IG6orP2WGK`r
z)C;2weRyPiG<)3Q;KjokTB2DJ$WY?Y#0d<oB_?Z$1CCV?(}_nq31;{hXL5{lz$hSy
zrJPMEr$MZIJ8{(a^M<b06T~{Os$)uR#{KG`KkMizI6Y;P((IApjWCeWUx5D$X!AHh
zY=Ztx5W?uM!2K0?BF5y3Izww_g)4|!mr_w@$R|QRF~(oWb%_-^gYmTk`fqk0aR*-9
z&(S5!7`w#i(d=g}=@xd4c~*OQOMb;>DUCWUL@+nV1MfVDV3U1-3_L($P!K^pNYn>O
zn|%r5ZO-*J=Z8A&9<yLkoma{l!i&$r{W)mxhrIX>0^czmStyoyCF}oshYs}_bbAJp
zS$qo?Z@~qBh>{U`OXkM}uO@-jWF!vQ+CRUrzHg=c1C8KF=<i$eL)ODh;&7ArU=Nrm
zQi&dd;vtx!1ae)vh|2FV<&BRy`LcvQs)n<!;cRg5J%>yQbKkotwF6HL$gmEKu!m;c
zs}9^5ehOw!p*zE+oL4F5gB|hW87}Ay7sBv)&gneo!f++$TgkM}i&r_{t5STM^S{mY
zM}ACjsScX6Tc&}yptynPPVV9NQreqDd-BascRRzm<F9$Kl5?u$oEg5#IbP+Q7{1N*
zxXtxM+`C_?yn69iL#oIMqF+H2n70VjiuErN1)4b~*81|6gqn%92JF{>C(gVeYDq6G
z@keb5ViqyUlKTBl(q|_LW%OOdau;!BIFC5z5pO9D%p+|v)l-+iQ=e(g-cgXgwz0BF
z1``%ijikYVG#H2y@!~r0Uk3pU=fQwH7>KsUi{%he4v`ol``*ZE7x$am0t7LYC{sz>
z#CvytUOuD!q~`)JZsmooyaMyLX1o@*^6P+oqvZvyltarqYxy2hdYG1vU^t5xvZVZ3
zymOY6KZ_5`lJaNsLbjAYo7ZIX9<0A?etb4RL5h<hn@??%5O+D$%9^)0s6|11Mf_fo
zf#_!ev-@#d$}S~;pp2j?mDhV!&6`aLPWDb3{5|fD^yDqOBK%qqhjkrhB|&z>dNZ_%
zm~0~UY{ru~lO)a_MG(ZFIMq*_h!!h|)tsuD;n3ow-)7c*x{XTm3RJJ48+H_SIm<^o
zHe|kuAkH+L*&SVnE~bWY5;V(jmITe!%<YGkB8YQG&Yi%}e9e3bZGmQi1T6={<<On+
zt^~uCQk(>aNm9HKgpFXt_%lGrkmAiCY?k8PAnXPsv{^yS0U<|<_kgfRiVuQtP>K(M
za7c>tLCBZlQV>eP2z@{hPlIq;ipxMKli~~DZ~@#Ie+4*HNO2`NR5FaaVj5-cf#M#R
z;2g<ydt(gOxHqPZ_d#)=O{QG8LvCZ#?T~41w2}xbNn7}C_F7FUk!0ZM#3`L6Rn4g@
z{q$fGv~I7l<3U6EsLzIM8Ai=%D<86z4@V!u4L22S6%kgEF06;v-8c3yyhF&(aZ{CP
zw1*77Lk4evJ&c-^5xsBwO>av0mhbv4Z-j&~UC+2p%rmy-1aYd_)V|Crz614lpkcFq
z0@No!gIXW+dW+lrq7}XW;<R-95wDr!>}*83*|`GdSHKYmF3=3gIG3}*{6?S~FZyxC
zJ((cZbB6VtCB|^fFgwdz>Jls<Xu(elVQko0;Fl%E+aO{aL^1kyh}bU0IpC86!Hix2
zkp(au)h39A5Lqb2#Sl>pQH*{JB92M%Y4AA>!HixF&ehUA-F0xjF2zTQ*->JP^%a5I
zxLFagMJJ$pLeBjK=YwjX87Sx8#PBDvid|mzK{G#kGl>FFi@ghqyHc-fE%V^SVVQ?A
zt|r22(wUidYt084vDibsE`{_+ApywPyzHB3Nn0W)h&zepPSTc93Su5n<*_tCt}78?
z$J50UL(5Kz-Bf~dqF+u7ag#1kM*VZd5T87AW2}AfoybKkUZ`3{te6C=iE1^&GfQ?m
zYO`O&d~`m5)hqhM2cUSs5=FVLhJ|B@Kic!iF5?~VDdS3FvXZoAkb;;*Op+LmKXvEc
z7yQNZhJqO58q=G(`#9He5~KxJE%atqoCU5~Qd|H&1rWfjvJiX<rML>rsz9QjYhZRw
zimSn_T8giO*>x$t0cJO(xPWvjAYBIP0x$I<(y5Si!TORwONX6~k}ha;%~s#3h;(6A
z-@wt7&4F2cBPTXW@e7VRhJ@b4QOA(rmz?;LVXPQZZP$XL7F1{ma$VgJrv>YVl<^TL
z9)YB2(ON`AMywuE#>b#|%!V%4)ew<!>HFE^uH+<tMdSEgLO)@cs^I>OCRMGKMHbRF
z@Rl2Rdn{lr)0G&hi|-`LouqC1B8dBla-S3*B#sA3TfIoGD@39wD>D0<{|brCq73jG
z5Qga&HNMDl&)d)F*Ds0fOVSS|T>YqDYK4(%8YK)KIe09}mm0~fe7f#^Uz#cfsshd0
zssv)ND%cD+B(yA`3Ngc=Er?@OW6W@;Du~glXfxa+P`mbs@s-%M$2bYnf`=A-rFQL+
zCB+5Mzd%}xDTMxoQd|YrRp5xzEr{2^`kEA1gLSnOUkB^!Qd~f~77$aZ9d|7xrcyiZ
zdX$(-?YL_ZF=cjqmgt`)23VODXz5G;7h-_5F@YMnV*}SqYUhrPQv8B*d?Ce6oMRKi
zsgHxJrhPi^%}4L-*3=iCd=_L;`#-nAUHkCh+~NJ7tfPLmiu}6}j-idZQ)`JG%~cR<
zNKg%lMXSa=EA8<B5e^VTG!D70;*RdJ{?fzWmlY-Cee|NRA$7$xlixp2`RFxzllt#}
zOLSc@m`YR+j`iS#qBrBHdKk`d0|Ym)Iu$Rz2mklbwgP9|aXJxj?EA;mvom?)Ox^~Q
z%Z!&3$1gW9yGli^B0Z{z7iz5;52zxe8P4O3^Q7rp9_N|I1!6k4`IqqBm%mPQp;}(X
z8!qE5n2B%T4L9%>7~Qh6o&UL{WMQAbI4{Kk)6zieSqhp<_NyP0s@19HR0A79y%Ahl
zWZDGkO;UUi)CcKNj_#yV5HEn{0<9+Kj(fHj*?wj>n7jVbv%VwFv%{doNwu`{Ru4Y)
z;47VF+Xhw=-jSHP?pH;hkdHRl?}Kid2bTP){g0D}agFXp(ugumx-U;B%5;Y1x^pP>
zEvyhzZ*4w=LgfqMG?QtzsPR3cFZs!j%;-U_bgtjr5%~4K_@Z`0<937^2;wY@S@vwJ
zKHqP?r0T!GZ-E3QfkhJ7V;3~Zf}TmzLR1E*Gr)q;Gr%qbdNRBj)SIRB&0x1#O5X$O
zJz&9X;Si_~NpU`?^QE{9)Ma3SwnFP^ps4^KhATl+$#Aa5d+6!xiJ5dlt`M^;#2%9e
zS{dgJ)0sEn!bBO9h>*l~x^mqy9!|jFCtn;^r_}%OK~?o9&x2bMF4l!pLMGoelkGbT
zpY%u=WagjQe83PDH;Sn}enR|Yh89m)EI~OCnFE8FT24dcX&8*dERrQhUEcA!IY|8b
zB)ItO)#zNGhT0FL#WOZqXly=;WOj55)VII}J%tvtq4RB4mKA7$Kyi|E!wP~x3&HxQ
zhzcvkw4gx3@3R`TASUy^$$SJ0VyV1mDj$US_`JKnjeGOOQ0k&-oJktj4Z|caZsAm0
zI5pxE3hzTTy>?EaCY?e$r?ADR#6HojPk!XikM`YDIYVc43e~4|j{gvN@uZgeZVplA
zkhYU2x-%!rT!!Vk^IJ3>R5CAbIk%0l`8;*~(V#?HY9mh7Bmjphh&PGyCb46<hB()d
z{)o?9x_Iu~<Hi$zao|7mXzSAKhn^Spn+vf^X(4ubTbYv=wY*Bpt8r~;#u}}(P@2W7
zvZVAZUXvxIXY*FsyvTl^%|~YQ1EqLjHopjQh0(OAL)rdQs7L*`2f^>b{XO_F^ZwU^
z;R;{7B~e4nHh<tQ6$Pb_FF5Ga`kc6$vs}%E<78YeEimg`Zv1~Ay;xm!t-E<><iw5i
zOnQ<q$&>}q$--m_O6Zy3%VI{NVWKrdv^cvBOj*p>4#IXRE(W0(OfiNuAIA#E7`|4!
zdepAVjzE=~23^v?fr*d@UGl(zbzBY><>0|OIzY@1NUIQsiP2$JrKnE2X54<EY1RK+
zid@~86*X;3%BsKEkuV(4>V4}g8W<O7s@U@C8ZtDej>c8Q5^asXVZb@>=QNlq(j!sM
zDW9E6(Q`x2^BL#IDuK&*%Vm7q<{oSNG<9oz@dF=>DJ3PezV6XRc+}5U{>Z!^HTyx`
z2ZiF-cluTYz1(LEAGpaq<DD+)_YHbdZ%F4f>0BQs`g%^YUW&JJu3I?`qiZ==EyFkj
zDpBieCHUytUhNw74AiMKO4oB8>bb7Ck~A;q9n$%(IF<)4g}EB@y*nsl8fTWqxwA)Y
z;f%L%JrO_7Jiba^67ELFxP~)a!#N}2_kz+$Kf5_kKH9bGvkH&9?zgh}KwS;WM|CTb
zhxDdv*T<bNVjYi)nnF}5#F+)S1Elu>63Osk(&I3D6GpCkB?nEe>C3W2qt_=5K3bxe
zuYcU9ejK-!Mw6KqGc_zoEeDI`pkc9P1z4<*;@zO$4bCjql!CexoH5q?p_zLsJg+R^
zKjsU6d?SBX?_cogM^E0X{yQIArVgG-Ux}y(lX|dbs;q~=dKkcP1GqFuUpo44*QS46
zi(bFpb=<XPb0an20pb8(EIKtGm^f|agK#aJH+ANuVEXzO9u>;aQ-b$({HNiwuSNrg
z3p`Pf;bdcSEW@OD{=CWa*EPQ_YtejZ{j<g!nT)=vg@0CfBa?M!w>Iq7f!*4*-4gNb
j=-ac8Jc`QdXzyTWVQFP;V{6`BX5PjivuC%K?DqcwRn6s-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_reduce_threadwise.hpp.71DBCBAA668DE14B.idx b/.cache/clangd/index/device_reduce_threadwise.hpp.71DBCBAA668DE14B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7ff39115d190d87936f9c3614e8668da24c048ca
GIT binary patch
literal 8606
zcmb_i2~?Cv5}s-fVf-0Jm~ns%{@k~mf?Nv7p%CH$0Tl5<i5dhWh~fn%K|{PDDjFbJ
z55QH#h#FAhfkzOdf*@ej=<2E%kM#)gMA(Lzvj2_m?LOaI9{xQa_5FQRRaaH_7!@8C
zw$WT93Z538nUXYjil#^;>PP>j&rO^8jEF>~Vv#7b=v2&=ksr88e;ZY=vE<7`x+z;6
zb2KADEzg8neaa84ORkLl>Bpo+!O<79AC(`TzcH4?Z?`jU*u1j+Y^H(f!>enjR<%Ut
zuUh>{T7+`ku_tZ+n$e!sWM;d5-rwG6A38Ik&q6+Z^AG%`)v37&i(0;l&)7VDuv5K#
z;nTChd|@C|H=UK%ZRmI$^g8W(tzS0HoqzRNxx`Pg<l=+qpFWOQRk6BhTBx$dDkvaJ
zA=(kro*J9rdABa;sB6gO@k#Fb#eNw!7xQx`d+2&iI=y^Brp-#DD*eQVZOS5XXj^)h
zN26!zi$i0#FVCCi?46`{ZQcB#x!L7UHWz*HYW0}1$k0bGChMdW=_gmu*p@M{)7#g?
zXw`2Q8YB1b-#zD&=>gj<J_o+~+wEWPXZJNuUGm)FM1Osgi96#T+Iy}YbTFj5@mguc
zp}nhjUT@O=eek}ZLn+FO^G#pv_FFUZtb@y?+|pSZPc-(K-<WS^=R37xM&ngu=eRLd
z)u&g+=-=2sCi#=Cn~U4y9sfBtYIs21ky~1dBRO}sX4@pk8c%X`Ha%%S_p>W=+><t%
zRNOc}G49UH!k~lS{AheuFO%3V(|mUFGmBjwMvA77oGpjhe{#-ya><;KoRPX~HlD2=
zJfUvy)Fuawu|AG-iZUy1zHoW$S~6<ht``f^mgajuKB{$XYR>4L?#rV-G)R1@Ju{<@
zv&qlheY^jd4KtJKqg=U(`MM{q<qM8WCwy6!;XOa$v#L;Yx7NoCua(9ho%=#GH7G5-
zSrO4VU4Qr7iLGNcO!fb~+njIMf7j-e)G_*LMw(xM`>91Q3SP!7x9$2R;HPkdN`>qB
zv>PufFX``@`}uIA(U;A?*(xttB#Awolsel#D9rw=wB&DdX4{lK4qDo8>gi|3UCQqu
z!Fo}c&iF7phd+01{AEvMMZ%|Uxu+(M+N5xc{Ji_2OUT5BqnwteXBHfPUDYwt$HL*j
z*?kME8WQR=x2MIGX&pS@H*G|M&W+h;_!Caw>5M7Am*!gY`>vTO?Ol4(iBEn_Uyzdc
zJKYMJn+r7bD%(vZ#KTy`UkCYhuv0f89lA^Ck9MCB4gblHzVA;oC}kVyYy+vfL5j}L
zQ5l8ugfFCoGZOI^iRMM3P&aQW7wWG+yZv#8ob;oVA*Mr|dLJm}jyyMd*j#C;O*9oE
z-lmU@Y41a%OMp%Rcrf*J`!gl#(+DZirSF#$-EyK-H`s^ecL{@P4-{;akp6U{%^+(A
zV|9ZR>&*M76FTO^OUOV<X#m>>a91}-vAx{ly#CIs4hgZRlwX1Vx~Uta*zJAPFGkt7
zOF~>Ir4{6@V5)AA;?*$bT3(2Gn}o<Ir3qwBV4`l`QW{;u`s|c^CMVi-MSdoJcD)Z2
z`yPyHl<kaMYe4j=I7}L5+513oK$^nf`(@qDeMuiG_7VFidLJkboMvp^;GefsLTu^y
z=fUDUIH?<?jEa|J`^qMl$%z5&A;2`iruTv37@vvtA(FaA32~%7Tm$(va8x%)`Lxad
zLHkAP{!${LJv=4mPl<=RLCUP-w^m(k%=$=53{6CQJ4o80pSnTH;z5Q97EX4v<wQ(-
z2{8?E>V2R%Co!fhaM7-j5@Jo~aTD}!f<oOO<*R?5D9^8b_C`XSDCGulH^5xoAZ4|l
z(VjoDc2L7RP|8)1UIklqgOu-T>!L*+(W6yUEg_N;(vMd)NZC5p(&g8>SEsaz2A%2_
z=(`0B)eTa%*}0`x=TtdJh$|ha4HRu)t!|LA-F`-q-`s@V65>uNzk&2Ouu?ZjDP44F
zz{d+lO)?;Qbf6rN=76QTLCUU>&bai++cV|FjZ%V4f^2#pDBgWQyCvwlWSxXK(?J@+
zwvl?eszFL+$fn?{s}?Si5-r-pb0U3COw|oiYW$vGalCmrKu-EnO0a3LQ||-CN7p`B
z`h(vKFA1@ugS3E23st<TLCU$0)*PRZv2=wt(V;!uhQ7DKK;0naA8u3iEDv5@A|W=k
zmvdloj{2&qLCVc#%VR5AS0qV@j8ZOx*=6uhH%PfZV%F9jvqKBzL`nw=G!3-reV|yB
zmh?$_t&vEik+#Us=fs@=;CLSsA9LJC7zi~u{xA$Y41pN-IKB?7>%beaSREeU`hq$<
zf><3C-wJ9x&HO_v4r_CKv{Ur(EW1+gjT$3A<oI!laqhUgIew;MrT{IpT{w)%)&Rc-
ze9;*=elyr?h9JZkb?L~ekc&Z7M;&W9ac_*pKXZJbL!b}piK4K;7>0rzf_%`YIDSgR
zlyPjvi6M#63?+Fa4P$8f=;@OfItIqazz%iA@h8CZ1cWeLNsKFr9R_obuO?R2#N#bi
zX8{dyZ}bnhnzu6VP;eH<-vjAAu)$!$@ee@yfMGq?+;2bK;C4mK@ex50!<qgfgCYfJ
zeDwHnOqLXx6vgH}J$(8ohSCDl!Wc6>FkOIhVOTDVKpSxU3K+IRh}T2FdI&|izR|^D
zuFcnyL6c7kP71*xq|2x7IQr`H08NgMwTlhFKU2NCObKHs(J3*Qp*%3lgTdIdQQKaf
z&ax_Z+Vc*g*g=diexWj{Td2|`+TS8z|G}r;jpBFSz{E*7z6Q7&=!f<atJ5dCG^S5V
z+yjz5V9bcBT#7_YqLjEFB>TaP5mgx#iI_?$u@WSe(4P@ic@>Gyqoxga-!*@wmL~NX
zpK9WV%hcp+NN^1aWw@3&)DkC#PZIx=WGKV+#J`?lw>#fed@%XcY^s%W#QGeuL8AMJ
zTPw7U$9ZY=jNn#7+)<#(pCk?^i6g`H#J-+5ARaX2r1;W;co~f_H$mqnNHB$AQ9u`a
z2)ILFz!s~j5OA?x>e7ISdrJZ|AjjttpIkB=_kp){qG*c#qzpR4c^}S;M32?vGe=~G
zGgJT{7C<EWjrXvEVf$X0>%X&J@5@D6DgRRNblOKidISdG?0g42Ig_kC<q4<BcM`6X
zNHB5qV6RT%&F~w-y%Eaa5Q{g&N+=)th6FO)MYt{@zl)f35v5S>)g|n|n{eGiemC*x
zCZ3GnO(t}cNkaTJbd!~Ue?g^C#r9NGOtJUC?jHD|VlgRbi+hVIMk|MY<xDXgP3$^t
zAjh1~@z22X84Sa2IDQrJSw+Gaen`9?l3)znSfo*Zs0OZD==`dR4gDc-_7IItKR12-
zj%k9YHU8?_nl_sXY;0{jaH7MTi~I`XL(>g8K1LQZfQ67)S*!rX$>Ibk-Xnep>WAaw
zW8x<<l;DvdkWKKIAVA5o<N=JCB1;jVR9UJ3E#?;YXUr@vOMtSuYynya+&buwF3a(G
zz~u>XK5+R$TnJpD5ElVgB*ewQ6*K(NsKJWUp_|4~t#!iaPKZVWVM$Egqo*E5b<!Xm
z92<#R94$HRt2_7668iQsVz!KU<J-}?p@&nFq-&_)InjMi^wI0El<uu9<cGl#cNWLj
z!=QSIKrB`l*l73IDI3O>Z@J=2d;CJf@C%Lp*yC8kI=9IcM%L7GMk+^oGo`Kq#VYVX
z#f^Qv$?Z<fmxJhD7iOOX|3bs9T6a_52>}zTd}bxsSXt7vv;|DJfIUuxqlK_h39-T)
z!_m}YSx#Iq<Z*m6ku?)D^z50I$7N=Ek}^8KYoK!tWT*m;R!=(DL5AfP9wBJ5Qk@@A
z=d->^v<<ltF~2uf$`64#>Zm72$!i#%b-c{Ju=vmM?-c~mWOP{7nFPz!%tOvqrAAw6
zoZ+j$xC-3au+?B(&2Z+48@Co}UN~pK@nJ4ugO~+|y9^hg9B|2jL1+)^G2pUHh_``q
z8(Ry<ZwKXeAua)Bi4d29vQ&u6Kv^cl=Rofq7+^)s(IcPUc`(30!_l*@UIQ3l9_Q$Z
zK(7%D@UXzqLelagu|qR*d<(H`5#m;2*($_s#Ij9@FA>X24CB#GQ`~!`L)&tE=HN_0
zNmU1ab!-##6x>mDFa<S_*^`>29=Ljtp+KxY2BMSVF_04LfU9F_5vvb_NL(4Ak@IkE
zx3_}hhscL0-<HdKfIL8nF@V;`@<1i4kAvmGN<0p6e26?miL0X?;2GvG@Bq)@0+a)u
zIp8n!0MBJYybUb2fjzFN#{(?33vmfpln8MtSd<F!IWRm2CPIrFo(B`5#SI(4L}+ot
zMlfL(zY026K_Ab@96gBY{091XM(yzk`-{X?=n?iULflI1TZOod*tapf@?J>mj8D=!
z-@Axf_4HuvqhS8_|GnPBmC|5b3({J!WvV{`(i04e)%OGFHLLGd&WR4XzxIDdq`aoP
zk3X+jRPkOBX~PYNV}VDu-=m@aB{6!*ilz<ECr|i7U)e%qaytxc2Oo@GJ@~_Rh-0{v
z7?lcRbSd#FCBragZ>%3t*1SB+iO&BPk=-IZvuOvBbr2q7aZ$13&Bezv2LB~%6u;cF
z?2=-oDOKEJ#bOuc?@JU*1gH=ch2Vnzj)T)ctGY_TK#N-KBUWFau;8G&VnjmEGL727
zj4g8qaoj;XuxF01Ad(8wpQWM-;$A^~8NNl#Zn2oOHEZSCvy#D~?`0jV@Tl(gfn*;j
z(P_l$yB(Y@?i)%x1>7m&!AEt`^OhKV#dO&5SB3x68mLKHlc?^2n4%tl-2+y%{ynZ~
zsJJS5AoZa5q46K1eeg_A-|x93VsDQz9%GH9^2{2o(J5QQIF66w<7~0i;rN8134%a9
zVdw+_O65~+u>+1zbx!qSC`*yW3y@8gGeg;mY=JD>CR-rO2Sq;cxMq$o0!5J!7lWdh
z;oTm-qfZr2E}%-h4a(bKi{8c2YC`suxip45O8fyNKL`)5syi_>3l><^UsV@ksILR}
z7amZoc)F8rivK)r|L9e@Y0-rTBdFjdF?>nh)@TRs4_x$-GN7R6j2{&n%XB(r(v+zT
zB~D5dpgf4kgORxY8qwEL&0AmhY}x-_jbLKZ68GNE;iNf!p6|T3Cqa%c0N(-#$KHGL
zX=FW&M_=anGoUyFHi&;Z-7;>s-ygT%Su$Sx(P`}k$zG7-8)Eh4A8PZM#CP9S*OF<}
zB@aWt!z^_lx6N`Jyg#Dky*A;A9ivnA!gO-zu>U=P)iquC<(GRM$Nmz)>VE%rY4KgJ
z3pzdLR{Kukh{o3B-w^p5q7Y(-H^hnIE+X#|%Dad~7Yosvd^a)gCYFrfO(MEUq!4Fx
zllh3zOX-$WU5GM!br7=-VvCWu!QW<a)F|b~cPw!6V9J`M8)h!0iv0tmf50G&jx8<^
zuI!7M{0+TyjNl?n*rto*A_Zug^E7|vg41QwEf~rLE*DIgldJ%41;cH(b7M=J?SXpr
zJcD_5Y=BIIOaUqYg95N)oz#PAJ$Rs>Tw2y<q@7{X`hQREmpY1~W^7vV)n6=+)sa@u
zKH{p_LA&Rw!{Z1H!HM-;SUQ}5K@6V)%Tq8Au~>aCjShapyCL{_xe>f$AiS`qzIp&8
z2ZXh#>h>CYxMfu0wC;|+GbOGk&g+RgT9V^85a$gH-*(mN=%cS#Nr}xwx0x8S^!=3R
zJ|%|eH0@3YcYB!b+Dd&RpD6OlK<x2tEi3N58D~zH14DCQI5XBV7`jY|b0ItzqL_RI
zgs%|dJn+kd5GF4H#}c6qOTn>Jh#NrJ0M^2iRM`mD7^6A9mAJPOKeoO$;@-ya)1re%
zU4AgPdPnDfNPmCu<=3MIy?3qoY{}1A{a@GWzyFKN!>g^%RQ^e)@}GBy&tLB;SKLck
zO?OEQALGF`S2CaM!S>87ewF~O1I0S3DfL%BuV0rQb-B8+gPO4uJUd||nvs5!F>6W)
z{Z<q&e?(}>kn+AaWskDs0Ba*W&x_F8Fky-?1f{1<O`qTMEw5S6L-j-dvxNGG@U;Q5
qk(SnfZ`QT6viLx0I>5|aq(rGmWXe8Tv5yw);|J`cCHpvlef&4#BB4qE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax.hpp.497B752134CE80EA.idx b/.cache/clangd/index/device_softmax.hpp.497B752134CE80EA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2799d7a00b7d5cb5552361ba7edd9b3cfec63392
GIT binary patch
literal 928
zcmWIYbaR`*%)sEB;#rZKT9U}Zz`!5~#Kk2=neL1X4AX&f6?67Zh|NA^Aky|bxN76w
zmwdbQIg~8!w2IGuR2H99wPm5N-wF57BmejLZQojYNY9L|vEZKip67}eA0%B*uFDQ6
zZvE)ezw`O}t2^Ew5#L?s>*loOLTT@^gWE6oZhpVfq4n1KrH5BqKDZtBvb^s262HR9
zKPt|Z)^)twzv%gbXEWFulUIH_oKbLWkMR7-c}sZjyHzc^m1pVbd$1~N7w=-jX+aOT
zb@x7+{mtaFS=A%UzBgKXuAlSB>OY>d$otgAIya^CGmFbogpD$GPG2JF&7a%zw;|el
z^>yz^1?Gkx5vOMrr{?M$ydaf($W(uupOwH89?N|i1}kk$#1yt_ICUv}dKHs;?2qHc
zl_qO^8Ww*)K2>MO;s3Xv_3ghB<#DX}Y=WxGbzxDy+U?iby`LyrMC<Q(V;{n<wJc!b
z-G7S~xI8@RmhZN}UxuN$GB>FK7_MHAI*!dcFO$d3#KO+NZOdjW1|}FdnSg=Kz{t%Q
z#y7?OTfOBuHYR4E0x@<m5io&JkoaIRZ<EWie*#Rb96&8Vz^%!mDb4`pY}>A3x0ZM5
zN&zM|n6$B`u{BJ(C^fCv3TPF~9gJLH`#n|4x+m^(t_$%KQhzbghLxFzTT@w66J{3=
zx0bRN5>H!M8_t6{n28JQP?$+P+_svwW-x1Fw_E1KsHrG3@^EX&XlTI9h0%<hU_Zc&
zO6_A+Z4fGU0;(316VrkTcY86nyzT#Q!31=imZl+0GthBbT5#T4W2=Cpa}%Zm^%~0<
ztHXi=MgwgE1qaN?rv|KF*!jh~f&SGH)R0CPX)I_g4bubkikz(m%qvByIf+WZP=^`8
zP&9c*_tS}IQ(&%!%M=$S7h8iOhzCSK04I#Wz`%uI@*<er>>|m#>xw68vheV7axnk^
DUZhJ!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.cpp.6DE6521393D18F5C.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.cpp.6DE6521393D18F5C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2d715ac5c2eed19ccadb0d8b96fac9635a865995
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%va0|4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY<1?>wl{W6#$$L%3ZMKp;>nt137R?}KP4#;l{_k6ya<ubujmI`I=9JFz<JNZH
zz0~#Ver3=2rEs~VQz_#6JBjID55#A;cilhJRBbNX7ysVJz4}YnYRT{aHkEACJf)lT
zr7Hc$zrW}7PyIO1>+t{k=_g0ajs6-jE;#;d=e8K8@<nq*exAJ@JMX~9i|=oq{CV)<
z^2u)|yvZ}XcW3eP&1-ot-^|GGHoSK~>5h3}Rh^{d^W*C6@%9%Uy`DZf=iASuS+ZyR
z)SZ)3ZXUlJY4*FMz;f@8nrX6g1fNGY^}D;zeK2L}Hm9&Ey`ACSVYjnOB+X;i@A>3&
zW7W2K=S|bNOV~_jn|paL>$%W-B>jbbeDLa>kMp)}EHLzX(3S0*e=Bw&=N2dBP`+8`
z1Alv1rN{o!ySUbArH^3e><1c^;X6x3`@K%BV2~^I%T7-8n_|S6AuqLl&C45)uCCsm
zufUW&b(=-N;wNev8^t1>-R}i%b6LbvzqU!kG3lK2UNhy5Wf~O<8(kLeeBAiuPgI69
z7t@t<VP<z`9hN>M5z@NkVRLrU#ouu*_a>T8y|wIMdsc1Anx`KaQWB*e9GexID$PIl
zZm)}Wb3mc{&l?8+UD6Xz&V3ZbIf+YYr}nHFeWnY%D*YSVV-`*mNM5@mY{?E)b@682
z)l%1MoaY64C@JhXyseandBvd!u8iZWcTSnT_4|b4rPXcD{du0#rl<sTCG3b3kV-${
z&nddJeZkQlzV^9Sre6#Em$--jpi|D(Y2MH6I&N>DbZbWI#_MeBb{b#Y%~z)X*~Wc_
zd1W7;)OqeVti9?BHqO~OG4{ysjC;2QH>@%Z^*O%BZC^ub$6cZ3E#I_{^=S0hn|L(y
zWUY?U&K6&h5v!el@UoIcs72iLi!RA0W`5g!a(Yf?d(4!NDn^34Kdf?|!XR2_XKkBx
zU;Rtp%Bb@`n}n))pWW(W?X3D@@I1-!s^^0(t3qbHQA-eB;q1A5g6F9@xiX$hi-fOL
zh_@f8m><|bqlx!B`)+2bl~eY6p1GaRc}L~^?DW4mAN@Am_<SvO+p%^0oBydLxZX9m
zAFJta_xz-~!OghDo3qY*dbs27*J|VAXOpG`a+@%<GY4MhTr8D*u8XTqwufzZ&N?3L
zP+gazl4If)OQ(IfomO4U%=Phdzs!vmEB$Jh_xHX<YugFFG!pDKx!dPFtE}zPmix1p
zdHSsgKKAcv?!hCErhTpW@&3oB04oj6BlbdP7kg})bG&l@uZJD;;^+NOUsCAzPn=<Y
zdiM7VxA^oU6<K1v!m4=;?kt|SeW7gscEewl?<dH)ZF<LF=v1r_8KpY8o*_^|>1T0e
zZjvG}=U@HaWPPH|@i7lWU@$`jKP$g^7>HotWMX894q^;rWGG5aE0$wqVBq?-;9j%0
zX)qH{gn)s7HC$tafUSTBj3>ChQ>t{`tHU5giUKMyVIWUiKo7<fo2=iFTHh1H$P*!~
zC+r6kc5Hvc@bgQ_Ehd%-Aw8i`m@o^9BCkgs$7Y?E$ph*Y5D-v+DGKouQhzbgh7}~t
zDZmdC-o8O{+aVdr)j(lnBM*hxy^NQ?ejMmZAr>Jem};PNScKs`m?$F`I9PfQ{a&!<
zp|!$afv`<?^N&0O*`y|*2~$;+nv)0&PKM&5<YIRaBOF9PKq!pCzz_mwGDkdbD74-2
zq9cbjG9)xCoFNh}#Sn&ICg-N)WB|FoY5uHi>>Qk2+`N4JLV^O`!tSmb!aUwCB9fwF
zQqoq|ws!Uoj!p*Rx;Cm7T8h|2C1mB~m6TP~)HStr^bC!RP0cM8%w+UU+{7h3eWF-A
J0-U{={Qw3#q9*_V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.hpp.B94CF584D85EABE4.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce1.hpp.B94CF584D85EABE4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..db413c0392b58e5f9d940e0f67f19b8ed8b044fc
GIT binary patch
literal 614
zcmWIYbaRViVqkDi@vO*AElFfyU|<jg;^LB`Og%;h23}?chKf0RgAQgf848?TUXjt=
zR2IR??`SjSQDnw@MYYX&H)e27ezPs;*Za`zx3XTkW+-xcrhU5icyGtQv(bvd+k&=b
z7=3&r7@hv1^VFV;>rbEA>}n|cO>FYVyqoW*SuK6uwK6hXy#3^c@H5Z%T8jr=JD5-<
z-|<i0k-ueI*iGgIUb2g{x?d)4;`{1ic<WfgLcjc$iE9nEPLXt03KN-FS|+mOO~VqK
zZ(f>S&jXnpT%}vIGFMN|R=mW$>5}8`+#*h%TqV`LpH^f&X^@LC*td~mS(~?1#Bnw&
zHH+}N_@foN3(K{y?fL6hvuBq`!KOzZ|2Hr4+S{M?`+e89do{Zl_aCXcR59)S%%CpG
z9Zg#e7yCRBeimrRrgrLr%D<oz>$)4fH?HnluNhSoR>Kf+OJ-GZWp0upF!Zi|Z?Zno
z=J=S0fteM^1p*F!R(^8^P9{bME*3@(Muwu)wBltzajstr?lpUx1~c(+2p9-h!!>dU
z*a~>Sc!KLYrApVmIt)^zD4+rp2J*B8^k6)%M;*szotMdD1PTiXD8Ph6{DjnBOtfJI
z33CeY!-QcpBPZBpy@!4;So6?Y;jcj0rn~t^o`F=W324Gp6{Y4Rf_zb2lw7P2Vz7Y-
b2w;aX7#KJaOb#jMxeA@>CLPS2>}(7GQ@`(^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.cpp.6002EE43735BD2AD.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.cpp.6002EE43735BD2AD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..abc04700944b02f27ad99f93a30577f0af051169
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%va0|4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY+mY9}N*nj=<h`ciHd{%ab(W21i)N6r=J9(Q{_k6y^7Lq?ol0~!<EA6KAIr|K
zUOHLa{;#P|?S#x-E`n>S%X_?4KXjjUJ{rFH!9Cyh-|NelJr>;iDC^kX{W^1VyNbFc
z_TQ7-TfhHvw~<|e@`e9@jVeFR_1W*k@M2@-I^F39=4z<B?yK5;TG+zwm#wv>y~UrB
zNwSmf>K>b2{iS5v)<i4m{%sP!tD}2<tNESRSAP8G$CD4Q*`sRy7EL;97w@yV`R2<M
zj~<=h6;`Ly=JV+-7oUHAQ?v1-8t)GuU#t+Gv*}XU39s|2-?L6;Rrl&WZhO5od>&`?
z)oY)39?NMww;=7aTjt3rg1aXE;Q9MJJIl2E&f0w4m>%IBkD@NmE(_0KT6@7`YV)(=
zD|VC4MT+nDE_ofG73`8EDx>;dYx}xKfu5402Ns{p%+)#58KlPW*7@;NQ`zV}RbRjH
zx+JW<B*vaK<qnIgnNs$P6V|7uc}#h*K2$)}<4omZ^E8!fe60K`(=<%VZ6?gS&ynlw
z*l_FSsW$229ZEfpp&n9pJJiJXzIS~hTN%7ESE^=(nftLQK7I$WWsG@iBEz?Ryl`{x
z)QnUCt#=yo+t1ivjEpE+z9Xb*Nt4IBF12OJ%qxQK1wT*>U!i&=LVKO4X5Pt@M-vY3
zI>vi{Lb}K#4vu#la%B%QtTLF!bn9c5-qW|Ys!!~^WP5m_`*sy?FCh(AgBX2|UgIar
zTe>beUr2ImKkWC)`?bdZ4g1*-I^|rQ=Kb8R<M#GRw`R0%yw0|6r}4$zd}aEdZQN&=
zSN8Eqo#%eT+N-``<D8unV~_mKxOZD{!z$BIpW}Pn_BEt-+!box@=^O(k4AsJiAOU}
z*6Jv2Y4H^qvD*0uFDqGuTEtDi=#qS5=C|tE`Zq7}?bLYenW9woFtlBRQR`1lS=E(&
z{x4*Mx6YiIqWVqh%vMoup|39z&mC=AHTghh=t_f|z6aC-+9&xbPM%_U+hUT}Q?)f8
z_4pe9+Pc{$ahctbe#4isB<P=-Vg7#A4PNui&;PplXl}yB=W9>r^v24k{q;T2wJUMo
z?x1-!=O+6nY}$Kh6YtEYhyQ-N{{F0td3T|k?gb7RUYA()o~#~AZLJ@r26AsU?KD~w
zvQ^;ZMalRFQ=$&#&whKBSL2AkeNlq&t37W7_g3G|U-QFd$!QmzOC{EPnYSdpF4yJH
zQnOs>C;5B&c2?nT>1&T4?tgsR`Q?Ho0{=Zs`jt)=$vyx3t6jOOuJYgfiN`E|#xwkx
zxAv}cw)LKrR)G@D$hU?E3S4aSovr2alfHa@F+t94(>wk`r(%W3DAmdJ3?-$?48@hX
zNs7RnZ~UO*#4W$2>v$LfgBc?DS^3SwKm-FP6C*=(5Mvl4Ls4p4u^b}<1J|zw_nN&;
zgPC|D1PlbM;Tj_ZYy~`EJi+yyQl;x&9R?{<6i|T)19{p4dN7{YWc`lR`koj@o(N$*
zVLzBKjArBlyQBBe?*(ffS}QoVzhU_KrQ{Y9ON5Y~P$*0l3yNN^M;*szotMc28Yv(k
zpa4@8;wPm3VxkQzNSITAA11tggXFeDGLox-!pKG*3bA_`FMs_w(D6bnLP{{zKv%K|
z!+C!N!ZzK_Kk^J@q?&*xOi@v4P9iWk8H$UNi`_wtFc1L&;V=dRLnwk7f?!6%namL^
zB_@*Ff6d@$jSLA53uj2qP07gsO8chyv$C;saB^|;^6?7^3U~{<yJ`sYc)N&5ii$}|
zTUp!M**iEo8HnrJs9I<#ViT2+m6KOeR#8*e)Yj26G%_|dw^T5b(Km4um+<t7V(|!Y
I_G0z}08@;YO#lD@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.hpp.453CC2F7EB0502D0.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce2.hpp.453CC2F7EB0502D0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..999ca4064e4555d4964d69eecab105dfe7ef7c97
GIT binary patch
literal 614
zcmWIYbaRViVqkDi@vO*AElFfyU|<jg;^LB`Og%;h23}?chKf0RgARHz8w#9V{$;@>
zr`H>zKL~j7E7^*Fk-L<#P1@{ii)Y!^j{kYLw_Xj~!@=e0VKn)>+<RA^^>N%YH%;B-
zJh?NS|Au9Af5_9`y2!M7x?Q_JCK&3LpRU_}Mbq}G&)Z%5EPReP7um_**)d@gAKT^I
z9z6d!s#LaYJtfBw`bsb~^p33B>kB@ji*tFWz53!Nag+11$4o<&Bv--6ba%&7avrCW
z*H2kgB&W)7WsBpA$=j@U@+sLWYs+rAuQvU_1+}0{+Wo5>`Cc4keH`7LAhlyjV`<L9
z^rqsc+e>@)tA5ElI%|7;XMgx~CciM5m;1v*mVOV+{U3Yr-fVvJ2kjC5Qs+O}274JY
zU*BMsHmS4bkd3BS!XlpBNB*wYpZ?7|u<lyD@5PCmAIT;Z%d9J|%uP}RhMw_*iW9f|
zmagMrU}gn!fq;XbmEW9!lZlami-nPck)bFxt#}zwoa@(ud(Gaa!Av|H0tN!saE%-S
zwgMh7p5XdUsnT_?4ucda3aG$@fjn&iJs1y0Gjf9M>OJ&(!J3EG3SN&oj?FqRlg9{D
zEg+x(Qx)PTr2b-}4J$~PQ-B{P{8u1s)7|_d&p^Uz0-7-4qSTy3kS~ghl8g00435t7
brfa56=aiW_+1VJ_K|&C~31cuYupyWLpwH>t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.cpp.62DA65F400167AF3.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.cpp.62DA65F400167AF3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..36f8af3325a5c03d8879e92ddf9ddc4e26a0865e
GIT binary patch
literal 1568
zcmWIYbaRtnV_<Mj@vO*AElFfyU|<jg;^LB`%oofI4DH+u3>9<EhUMnpHWRq!|FS(d
z<KkS&<B`|6d>fB!yq@5jv_<n%;DVErU1lhmPWp7Z?eBM4Rc(F0&x$vrcqXXclYRf=
zj`rL=e;%$+eAVo8S+y(S?(X0it;YGea$5W5v3}39vibA#Xvr4;^*d%2zF+T^K6}v}
zMYk_|>i<7dPpx_2w&4Ho&`+PFXVy(+xRCrgCR(r2I#5pZ=dZVN{trt2RGaOzu={Co
z((lAOzQ<>6tE_I{G??4_tz4p=-MZ(u+Bv;-N{=6v6n%PX9#!+x=49G@z2!GQzIl`3
z(WCRb!0dFIeE<4o)90Vt{J8j0jqis<n|CbfYqp4Px>S5}Z`7t+>n^0;DcJk*lyTzf
zTUB<~XXsv#ypnTILwBZ;ley6Rh5u`O*FNjpeRgf@=|HCIG1*m@-*Nda5S3b4bfKp2
zmir~k*N@6O^Mt)pyFAu3FMbmH{934umgZuniYu0Dds{67Pq8eh&DgVPS#tOF_4oHR
za;yox_0XYr((DCC!cur;E8TAjbqf9QVhwCMGRM3s+b!vBzz2t<PN(AI%rAegy<jH7
zvSRMqoE?Vk*@q-TT9-U*&Q7}cJI>|aMDwY)mL6=+s!mz+^aDdmqSS+9vwTyf`RCq^
z?hIpF_^9tw;psoZb6Zu*yS!CBv^t)w%Sf`?)}Zy&mQ}8(HOi@H>!YZNkNW1s@R;Rf
z-TA~@>89A(@PP03HA9{ST*;aXWY<3ny0>m``yB1N#xhkU-kX-FGzPI8{@QfG!nw|C
zg@!GQZd!@On<e|U^nK!QXBW^p5|+EO<X!WQTHTeskrKO&Hat4#^4{vk-ZP(N67M~+
zw92s2-ym)>ui;qHGq*Cq_lqm@-4Cp~7Gfw{+*-qUT41*-TjuNV9?3wt{}~h5B(H>R
z4cn#{aB)}Iy{6@^1*;17=q>I%Ho^GqceD6SOJ$1!k4-+|^5)QMzCfmsA3xvZ+}Wr8
zrEle}b3U7ds(GK?>SFDz`eN`r$?>Y^gDtB<EN*!pPzz|E<fk}!isfyKNnTI;A|A)e
zFtxMyR2>n?sW7-9eZgbtAK9aO{?0mJRe8_mb$WMs)6sd`^){Q9)z0{8&ZzyW?dLnM
z%1@U5c8w>$ah=|lGL4`AcKm+x;yXHSl3N{=3rqw`=Nhase70uAgXsq|HXMI;YD3Uf
zCzZ#=`wnPsZMkoHd$x2y$Nah{2h^@q-c+vozU_X*$4-ywoe@i3=E+>zVmxKp-+Ly$
zMNV^#zUk+4tBIMfIexhQ`1JNmj-HDD=4HsaPI_u{?)TYp_b-1w{jpc>E&jBh;fHSY
z?)GbC6(_kAUj%KsmE7>q?&}`jw-tLjFZTad|6(wyw7x@l7W2tVS~LGJPrszhUR;@*
zqzKISOjV{|qJkyf@-PGjGeq#S@|%Z&2nJ3jMuzAh#xO>PqSUlvIYxv^MlP_j-b23^
zta)gy!1Zgvy=HIIU?!dj0RsVRxOou*wgMh7p5XdUsnT_?4ucda3aG$@fjn&iJs3}H
zvVKQueNPM{PlT|Zupdm=vHcCh&o3pnm{=l&^n^lT!Yn9?ydHHNn{{3$52#l_KtKVe
zD8x@l{l!EZR**2K06$E4`v%Evhh!vI1BH=|JQQO0GG6}raiA-OScH^ds)5d75r*^r
z3WROCn}6gP$VfE-O_-vh)SN_Ma55AZB^SGc7-1j+0>WVo28K`sGX%kmgfp2Vx)ikB
z7`NK>vPOo4hJ`aE=ceRj0HuA?{8`!9IXJnvdHMK-1O>c>-CZ?=dAwajBt^xfq^+!N
z?d%;KoeacvZB#9^6tRg)$jZqpDXXZdYijH085$Xznp-NE$>^K7iA#9;M6q}TID0Ys
F0RX_CmHz+$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.hpp.3F8E078976E83385.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank3_reduce3.hpp.3F8E078976E83385.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c8754b520921ffc882f323ab6c0a4858d829238
GIT binary patch
literal 614
zcmWIYbaRViVqkDi@vO*AElFfyU|<jg;^LB`OkG9>23}?chKf0RgARHz8w#9V{$;_X
zrqUaFUj)4Pm2Abo$X!a=CT%u5QtjqdrGL9`uUotAk!!$=7M00+zE|ic)qI_Ga>kk|
zva?+FR2(=bS+hjx>aWzf6&p`I+rBu|?)CD&S!<JPv}WGQt=xF>ca)yHx!#c_%?E#&
z2|l)On#Ym5b`!gUD7#B&%ICugva4bYZ(U1R=a=6yajwDCDV)wKVOJDhKhf~8Wbmxq
zqrFsPjyorV>Q#=^D`D=}9i7jg@M3>ke$mw+#9ZX-vBj$<Fjo~E{Bccb2A5uD!tE^v
z1_yV3+Px{W&ilpPHq-a}gyj46Sm&&<xK#gsh12`RxBl*4wAWld?*RYCc_#CoRxg_}
zizVjJwHcX8p9C+P%w`l;%}}n_+BNU)ch=hcAF~!IrysS>;E~zHP+XatqzDW>rYh4f
zQNa>#c^H^kfm|Tq;AiDGXW(RFWZ+_9<X~hdN=++X1{8;>WaI>^=som%!J3EG3S7Sy
z+-vqW4QAru5HJw1hN%MbYy~`EJi+yyQl;x&9R?{<6i|T)19{p4dN7{XqmE;<&dcO6
z0)+(x6kx(3enRRmCfcxqggFKHVZwg}!ZzK_Kk^JDtR|od6D~^4Nd)<#xG1?;AH?7&
dJlZ<5d{aFSGbcM613O3v0ytp|1_m|+695E+;NJiM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.cpp.22BB990A62BE50C5.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.cpp.22BB990A62BE50C5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f588676bd65f474895ce4713a432ca01c0baf962
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%va0|4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjF?^?BF1N*nj=<h`ciHd{%ab(W21i)N6rruw}N|Mx9UIof%-#$%fpb4q9VacjHp
zUh4XFzp`iiQn*~wsTA@3oy2sn2ja8ayY8Q9sy3JHi+}I){rd~iaN~D>Q(xu;Pl-PA
z;_JB&e}B)7pYpL$y7Axp=_kA2r~FD~a_B!>oV%0dor|UBr`g+g+cZ92e1G%g&&`MZ
z)o&`^yqmOVr@Mc8wDhvgfA{JhtJrdMNB*NPe~gUI_51VfuUl~Ry1x3&H=mE1SkIX2
z*LLK@rvBxdbH2TNP*nBtr;fFS^11CS^8I~Q2Q{^GTi1MvDOx{m&9>_=jPiEGRX&}$
zAvD)^e)bu$7d%;J`BSF(NG_D_Jb$5n-?Fgc<9D-?A12v1YOgchn_1qfc7dxm#4^Lq
z{jKuLIi(+WyKLuLYUHYt*R=Ra?9*$ZHd>mCnJTVWuI+8L3_QiMpf+RAre(?9;p^}1
zYvfoHdh3xx@1)rajzp#K%2vAH6zUZEvz2QROQ&W2ms>qYZY_M!ctoi6Ss(kQ&)Exe
zH8=vSqi+^u^W}3S2Z&B;<6qb5`o3DQ(*0gY)@(NURbMA=oOYObLThG&bavR(T$_?I
zbD_07PDf>*CLjJIJhxS~yvtkFL#yM-x{M^NYYkdYZCT}tTBDqLwmyoQ_-NXk7#_1V
zSuvk^E8P@38y@i8zGld?fGb&Zf$aKcL3!8iIiH*QPLj`dmuh6tBnQo=hgFUla}@1W
zL#D_zi5XkTlm*2{x&Lhc$kEh#W0mgobF~87^HsMRawo6njVVrFR4)4_?pbA@LH;Kh
zS(AC<H@Kzz9g;1Jm3Mc1zqn_+azbeKsu_Kiy)}%d1$L{lWxfsXkqngkpD}?=@=Dm&
zuxok&7k7o-Yg+DFu&Q8>-r~+<6O7-MPu9PAiEpRIW6u<&vWKDV8jM<hYRamv?DKyi
z8@zSi%oNpcQfIb`atnQZk$CQC)2hh_GDBAy-1I%57SKM)PjT`T%i9){yq>D9`KZU&
z_^~6fe?}9pd-Gk!-jGxGRi3@w&XPCj`Pt2N+bYx}-c-IeicY@X9$7EEVNu?U_v@yp
z&;R^Hct+0p4LQDPKMU^s`?YuG<21ulD_Er(4>zuO-4fVqnC{BDpYI#{?woZz+M&8G
zMJ30?EtXFEa67HKn3?P2<$jqPEmr!~F7NMsi`KRid}$=uZF0BIc~)85r7iboFZ1+U
z5q#|5)7*nc9!>jN@#FoEPXSgMnn&z~&Mx-YH0OBb{$CF}=EcwZpT4Bf@1Hos{`Bnc
z7jE(CM=G+!dWBW<7~EMrZ~H>o{_TdpD&J3#bKCTeztE{zAu>vJay>)IVr7Qn%G@MH
zV9r;Y_jK+_qYI%t41vK65&W$D=3yX$fs=`mAv%aLjFF)zHLX~Vk%58h*MfV^-loA!
zJP`s00@iSi5dyXX9x$HZ`cA3Rb*~PC6e$X*z=VN3Z2>(PPi(S&M{0df3?om3u%56V
zOc+Kpa)I5^d+7IqH4m*79NXV8{QOdKi-{#dNKYsfriukcuh*lFW3$f7<N=Ko5D-v+
zDGKouQhzbgh7}~tDZmdC-o8O{+aVdr)j(lnBM*hxy^NQ?ejMm{Ar>Jem};OaS%l%d
zzXD;K?&cqP1~O7jKoh2@C^aV$7@Q2nMajkPAVwI7fPioqgMlFw!3;q#BjHTuh|gEK
zd#*4_iL*wAgocGPB<H5&WB{do)BIW4*f}`4xOw^bg#-n>h232>gn7JOL?lJUq@=B^
zZSCwG9Gwirb!}8Fv=p(4O32E|D=Dj}scUNM=ouOro0?lHn91myxQR=6`b4pK1UP#!
G`vCxs-ICn^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.hpp.7F7D38DA481813D7.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce1.hpp.7F7D38DA481813D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b01e47c715e75f1f3dac03742edeaf69f3cf2b79
GIT binary patch
literal 614
zcmWIYbaRViVqkDi@vO*AElFfyU|<jg;^LB`Og%;h23}?chKf0RgAQgf848?TUXjt=
zR2HEv*Hoc#bn}IK&OT{(Hy8@5-OOG3<^JlttyeF3FKks&nR&9h|GUz!@b#@eH@$8y
zI(cL>d(Pa0?pmMS_Ujm*7du*VBl=En`RTgtS2S%yea`OMW#QAdH_7h(-5nD)@v&XL
z?eOJ4N0q{st*7J|LSG4lM*8qiE?uGCb<vECyR=HVC#QAg(lZi58y9t~*}PaGm0jiO
z#&EAdt9~H{ty>Bqp3!q-xF^n=IF;{K{$%fj3}MZeQ{6+A*}pt!`q=AgU>bAjfR)(=
z!-G3N-PSsGU#Mz((%IYVj~rj^&8V*3_vOC!sw=-&Z2P}1<Neuo=?~6p)O*u^&eL2v
zqft9T+W3@<U80<5a>61W-$(wg*B}4B+Yr0<_pC*#=|`;-I%U=sSLP-u0z*%2-qX1!
zjV^@pFfg+Mxj?|d&&qGkz{$kOz{SGI!N^dQnpV6FD9-h3!M$d0(_khZ4gmuJYq&-Z
z0b2nN7*BA0r&Q^>SBF806a`dZ!a$z3fF6toqZv8DcJ&_my<p8lYXz@I9mi&!m&s!U
zsumDXfT;@c6H<RM(S{Wy%qhSR6aFg@w&`yEk!K)bH33bSa8YVbBFGoTMajkbAO^>$
dSLdfS_Rc!W%*oEizz!0E08SW#fq@Od1OV_B?A8DP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.cpp.AC6F1FB653149557.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.cpp.AC6F1FB653149557.idx
new file mode 100755
index 0000000000000000000000000000000000000000..36fcb19c6edafc7d2c090ee4759406ecadc1a677
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%va0|4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjF?>$q!NrHy-b@?KMMo2?|zI?KkhMKefQ^Z30D|Mx9Ud3rR{P9?gVanq6Ak7egq
zFP$uI|JT%~c0%SZ7r`~v<vrf2AG*&v9}VC9;GS>$@Ac)&s+FsbUhS!_kF>fi@>I;Q
zZm)6G|GMYmsWlJW7X1I6`stDN%(|Hj7m`26MC&zL2l{o@d@Y-<UQqL;x~!<K;Kxf9
zYn9#6J?7tEyv)fyT(ZslR??gP?GoR8=gf<DJNEPA$%ohZxBU3^RAqMUz8Pukn=hZ7
zAQAcPW689$cCs;k`gZeE+0&2woc7@H#Xxn7)TL`DOr7Wb?&_qg-=t!X@m|kfZzH^I
z)!OI9y*HWWI-Y&jdufuUa*6Us$zSKMU&(s6Bl>>SjvL+&y0U%qZ^bU;+~TAh$~WtL
z;BODB^w>Xo7uPzk^bza~Ets-r^_^W^?y9LF4E<Kiw{A30_nOJDg}*oaT1n2wS66SF
z3p8%^ik4XsSR_2<M#q|k%I`Izl>(dghdEABFnr#7Z^pz3YZ03X5sNO|`RGtpr@V!i
zl`$(_TRQK|!!851P{qt24`xMt*(W}+{E6?<ZJ8ey-jb2bwzOf0oWZtZV_NLZW6tT{
z!xl|-T(M(u%^jovF6oIU=ROMJoW!NHQ+w8oKHddhmHv(GF$<>&B(L2Owq%E@x_C41
zYbo;@=Xrr1N(ws;Z!6_tUU4XbE93a;ol~}FzgK+j^_{0(?(QTVP0s})2MhKpn4Ft1
zUr5V~-?3X#md|>n{@TTV4*!#AYQ3>aclx<nf$jOKTMfCB*Ym~{r!Oj(eG~Vrvd<v@
zlZ>p%Jn<XcQvMFfmc`1uJHB7svt2nMG<(&IzRKPj#?u13RoOBhhxbSZ%Kgunz$SSm
zY-^aAUckj&VfUJryB4e}*rT_&^VkIAx8Kd;H!YPd3OqLXgv*;ltN8+%LVo;w^JPWt
z{0r90vW$&Sdc857k*zJJ^6KK@Io+(G>W!CI1to5p-Qc@`Pj#-7nr88~LY1kfd?Oym
z$}s&-cC$_5GLzviV>b!Xte-q%`+cDZ)$_B{|K@!3+i>IawbX6L*70xtr<UM)*WiAv
zroY|ulj;UH;}UPSo%!^z?)U5OX??SgS}YdJVCiF9v`*N~^w=Cx(LYu<q~C1XX|yI}
ztH8;NlJO6wL><bX{q`)c#u0z}q6Fbrd)^4{t-hVV=7-CY(=IxfO04-ZZ%KMxuFIdL
zX1UN$^7r)Ztis*W*B(FI|M;}?%LPjW{(G48E1fKod;a%VyK+@s<-hq8k6HeVXZSO3
z?Oo?=>pdy00wtP}Zw(I=xY*`9Tg&Aqefj)if}Go?cl?D;#R`#8s*~#(W-utT7gy#c
zDFSnT)6`tfH^Np2c^CqN86x;u`OU*X1Oq1%BSUl$V;CbtQEFPT93uk**RKWln!Qbf
znRp@u3<Rv<8Y2X31w3Fp!S$U|rR!cD1}Rb$P=N^pdD;ScFrL_C{f^Z7o)|`+2w^>8
zKbWv%`x}OzUrKH<u|x>z35CLhSx^*tJ?c0%>%2@JP_KZ1fC5ZWh@X)9i-|U@AYo1c
zewZ+fX5<2Ut@qIH1#2E!D{S8&x$TgQ<Z7U5WSb6!*u9LGzkVF(cp(-cC75QQD_Ml$
zyuSiro9^Zxc?PmcO+XW-s3<ij5g42d#YM@*?jS}Oh=71_7=wW!6u}HZFeBki<_O`>
zhd0jjEp%m#3<(VjXGqRX$;kjp`=<G`vaxe;a&hzW@e2tGcniC`Y6$aqyNF1Nib+XZ
zS=-v#J2*NSi0j&@T4*U^6P1valUGt!QB&8{*3mOGGB!20R4|j#H*ph}@brmd@d$AC
HV)g?75{7^)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.hpp.44F668BC92D2FB11.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce2.hpp.44F668BC92D2FB11.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b0c203b8b8099bf7da78d630c76a658ba96948e
GIT binary patch
literal 616
zcmWIYbaP8!VqkDi@vO*AElFfyU|<jg;^LB`OnpWM23}?chKf0RgARHz8w#9V{$;@>
zr`H>@Un)+KbE(pMVY}pHj(LvR=9EpVUH_EdUblAJBi8^Ar%5ySe6P??s`xs~(=Y54
zpYI}z`UB@AYnCWo{gpcPQO(S=$C{tU?u-7mdum#pmh#rwpB@>=@_#y-RPdNdU6$w5
zZ4aLR991e?ww{t>5dGe<>O@+*Wa)}!T8nadpS}9xCUKK9*mI_VfvYR8+guq|&1wbB
z{=Gq6hbpHq1O)X4Y}z&}mYvr&Im9Ysf9mms6?q+BZ!KPB!1m=q)5o>02BI;S4lK)D
znBG+UbbIRMKhrMkF1=Ftd+LvhQq2b`M<@QPb``CXzxwUI=-a(N3z+LVzApKwbKiLB
z6_pOvoSvB~O6O;&Me<G5a6J7n_+n|@e(4RXzs3ibJXur0o}nbOiJ`bMH%So~eoa$z
zIo}9d9pqtPW(9JAfP<fv-<*MyiIIVeg^`1ip(r)2co|Tf>(_#N&EBTLOgtO{1_IV_
zjT{2D0v<4);QCIf(si#6gA^$WsKA7QJZ%9z7|-ib$FW)GW%3w-!U6&cFyRnCA@vs%
zZCF9VoC5qXVHnNG33gfUq2CMEJhWE$D-gEnZvK&HAk}IDnlM#GsX2)tUlbQ57wdx<
f9H|fO-|xvj(8A2g&c?tF5`q9u7=wX<4Z#EeZhYq}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.cpp.2BDC7BD673407442.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.cpp.2BDC7BD673407442.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3887f062475efce3e35684866b25d0d6767845ff
GIT binary patch
literal 1570
zcmWIYbaRtpV_<Mj@vO*AElFfyU|<jg;^LB`%va0|4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY<MXa_l{W6#$$L%3ZMKp;>nt137R?}KP4#;l{_k6y^7Lq?ol0~!<EA6KAIr|K
zUOHLa{;#P|?S#x-PJ(Ny%UiTnKXjjUKDz$-!M(G2)z_D&Ew-&XdUZ~9eW2BCk*8vY
zb$gAg{?|PhPpx_2w&4Ho)K8DBXV%SRxRCrgCR(r2I?%8C&#$+7{tsS!`2OZc&i@CO
z3(FkJwoBx{ySVx1wYHlzTecsa(X(CRyYHMiu`b7MUi^6QI{%g*zn)smuD6??`up?b
zmlGr+zkN)Zch>%G{F!t1^V5DWKk{?hgU1gO)h!a2uAMM-p7*<}U01(x#UA6mp1s~i
zc-yMA&$)YV3e9yqE7p5ylBRNr@<+*E=dWMMdblI{ew2m0YQf2^%go=dyTB5)aH1Cb
zS@9LMDswmM)rG&fwlHK_$C1*4DSJ-e+12H)ni|5;Z?$~u#tjNyGa0t<_l939$@%!|
z>TPp@#;sn_GAja$gs0r-ShG<1y+*WBVAK8}$0-Vi&wKC9m>6L#VlyFP(S<u79jfY-
zxA3wuW~FOO=bd?In#i-t>C%UTrW;?>>MOlJG21Km@&o6smPXf#Di}5y@D?PW-L<K=
z{oK2?F51ljh3-Fhr2gwXcc{zy=u#mSQJ11{6GK^PhpC_DG4bzo(p5YbRk+rp(A!Us
zP5PQ?-j6n$#S>f{3fgmDORxpBZWOuD7hbHHfBl~GxvB3Y`D}NIMg~oC&}@2G<(M%?
z(OxxVid>VJv6W0&P<)j8&-RlNO|3Up=}tdaE3iFZb*rIw@_L?_!t_PuvTx#^RrYPj
z_#`81GEe*lx0Js_vSqRI?vC#l_iR^A2+dwKqpz~HhI6&RZdJC-$KgGafpY&dCa_6f
z3ELWGrWbH=SJ=I#<*o&*3ijwN?mRZZ`0aPI_)SY?ivo{LKH>7_&}qIvrjQ>$|2)}H
zD}UK~S(dT!Nv}7i5gFRkR9;;?+|;cas@`~cRZ!xt*$uu6_*CaQsc9B(D^!_!$~W?H
ztPIm+!^LujtkQk#yP2g{PT8M%=5{`3jQaW6>3?%R`fa%J`C96>W9#@g|5HnFy=!nk
zR@2|^`AKzyn{kOZ+s=G?xa9BGYUAT)i{=D!n=rXE2VUn~ER}q&i>pqyhi!MxLLTi<
zU6-PYQ`0RLPy28?t-6?*>*M8qnHw!u`qeJ)?|qZjwiA4*C3xE8ZlCk4vbIZG?$2K4
z>9->I*uST_2akMe`&#ki{f|!pRvMZ|?1j!Q_SiJ%c;)_o4?E_?&-<Ugq|oo5IK%$*
z?C%$D@##k@vc!6YRr46!Sv+t1LfL+E!(Wx}8|2(Jz2h%*DprV$RGnPUusc?nuedTd
zNfDUyHLiSJwAfs$fQKP4m?46nmESxJL@;nNF)~C4F@`ZR6s4vW%P}%AaQ#|vui4u)
zn29Gsz(Bwnt}#NuR=@+s6I|aZRl4rgVUQw40Tq}qkf$x62jhuN*6&EI?}=gLi4fKk
z_Jaw-XhtruJ9-cOUa;n&wSr^&8-|}>N^UW+L<s2#g~C*^py>5_)NyRqd6_(*kpcn&
z3NS?>enRRmCfcxqggFKHVZz%tNNzhMBe@zVjBMng5WAQ0^4E_89WTTpqy$q9bR~;0
zocC8CY}4KRBhNrastIVq6cwfBBm#qzp|~iy*d0hSM_9}Z=u?`aV8<F65*ilH5C)Qk
zfN&Utfgu#Z3_&m>5zOS=l$;D8*Eh|dm5rT)lZ%^|k6%boz+2eeRYRD^+eJiDR7^_R
z%G%b>-oeqyKwQ^G)j~@Vo2Z1WoV=2<ikiBnwvL{mk+G?{rGlA^zKNT-gr`pwi${R7
H7qcG#8BL4k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.hpp.1A5543D9F44A4D1E.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce3.hpp.1A5543D9F44A4D1E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a79cb2efcb55bee4447d22e1b0ef384ceda26693
GIT binary patch
literal 616
zcmWIYbaP8!VqkDi@vO*AElFfyU|<jg;^LB`OnpWM23}?chKf0RgAQgf848?TUXjtw
z`gVhsTvLU{(ajg`Is2sD-C!tebTfD9m;0;pwqCvDy|7h9W#-B1{_je^!q>Mx$yu6n
zF~w^$d(Pa0?pmMS_UjbS((QV=V7o+r+3C9NSI+o{E=r4yl~HTHS7G=5?v4YI?9G45
zw13sJ)M{_pdP<H#w7g@L&a(EIB`egm7UuHwU47YWut{{;G^6AMuP!CExw5R9)(V>a
zdxN?zRZd|DxX89JDc5wjn38RBkX6S1)Z+;m!ku4FEna2F`sG2>$6i+h)0j&KtjsPL
z9^Cn9x7M-yLRH(ntaA4Y$*<RAowLT`QvLcBLGKse`nP-0UUT`p1N<B3naq1yy=+QS
zTgbs%GZNK4^;AamC<vxzSk!9mlDqq!?M>F*>ms_A+V>dqcgt*JD6Y&+QUr#d#+9#&
z7Mp7o@Gvm50=YoI!OzNX&cMmU$iT(I$ic`^l$ut&3@Fa^Yr(x{Z_{8V9u5Hm0c*HM
z4gp&M4;W8yeWz6Ex>tumiWCJ@V8TG2wtyat2csD|!FKf?`n_PyLu&=EM;*szotMdD
z1gaJgP=Kil@e@*iG0}z<B+Mzm4-@_?5Vq-V{*h-OVKo6wm~c^QP9n$`#YM@*`XC0!
d9~*lQ^-@tiW=?iC26m7T1aQI_3=C`tCIA-o<N*Kx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.cpp.E94DE85158A7059C.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.cpp.E94DE85158A7059C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8d68770e0ad4848883e2635d1a6c714ca3fa5bcd
GIT binary patch
literal 1568
zcmWIYbaRtnV_<Mj@vO*AElFfyU|<jg;^LB`%oofI4DH+u3>9<EhUMnpHWRq!|FS(d
z<KkRN^|)(XzKusVUQh5%+M@X>aKXvRE;E!&Cw)5I_V>H2s<yu0XT_UQJQGy!$-e(_
zM|<v`KM&U@zH0WltlE`ucXx1%R^$9!Ijw#3Sik34+5Gu=^h?70xI)v%_u{9XGjrYP
z-1DOH|G%UDr+yshb@=~%^^>ROMt`*!7aV`Kb6X5k`63&wPhW4_&O7k($M>9~!kSNo
zljcsiD|>8q^_P-un-i_9Z`@Cm=_`}??mH(o#`W0ImrtIY&foImQ{|+ycCmh&AK$!r
zae_qTw}&~?&f3Yv`|8`zO+Wtp$j@014p-daJQtCvowD@#r0T6HTVofV-tn;N@znG~
zVOzh}tTT*WV7%hyp1>%h)RsIIJLmsDXGNWrEuS6DJ)NC1rg-g3-`(0KjbRx}PC0(I
z%#!zfzNY)W>Mmc6shUa~`T8c6&y5cMxME2U*Mq?4Tg=3sxo8SEd|UYO)T}<;xczm%
z*_1YfWwtY$dgV0<MNc#>c_N$Yt>X1TThxV9sQBKCwUV7%oewg1s&bu`<zMoAyVLCe
zfdysTHa$EhvyL-4Ky*?Y|GG}s_tk=x?)O5nrnAYfdOLCBw8P93S~DA@v(2XF+LV;#
zt42w-bXlJ|uKU@)L~Pm}O}(iKD-|Z~Uex*W2Ggo3-$g1;h-PsbUhB$s?XoT^GQ6>2
zW#MV_Cz7oyOik7~(Z>xMw0f2{l<k@2UAg<4T+zCc^93(1>Yki5iN#Bh=Q_93W1gRy
z3m1J6*fQhd!wr*vU9vi5FVm+G(XsYc(aSsRg}<YMq&FLsr6wGm+j*~aL-ow(mWTG7
zEH1fF5tpE!VZ+$_^i1y?rF-t5@AWl=u3edA^^EH$(=^30Z=Orn)=L;KviWyWf!Aoo
z+N`xXu?rTJuHD1x-~AxuVMVN)aIa$et@pY6QoOC7Eb398*m<M%y6hsB6(2s|yt$*+
z|AlPu*10oNRKH1`*(%B{^z}vJxuZ?1CLhQQU0JYYdV}u*KGnHSYMRB{3RR|_lHGW0
zw*?EIdr#F7ksJ%DH@q23g8ua--K+CWm}`?i@9XBH)(4Z$M^C>gwcGB@uh|WvrH5<E
zm)iWCqdxyYN_A^$w9)DI_P6Wr&$i4HTig@1K+r;}W0#-ARf*!TkPoLHBy2eT?9_&!
zt4=D9i}xMS+}d*A^!9A&fR6cfPY$SEsl2IN^?lp@h>x8f(>o)Uyv&oiw8ePJvcLCC
ze2bjs8hz8x=T;LlUvvC${qgDTmmED6|IN#gbDi|m=G^bI<?dhpeEMUr+*|x<Kf@2*
z=-utt$|_EBDZU8WbSt^xq21R#yl*S^bYATLt^UPeQfYmM@GRz&m$YX7VZOIZS+uw^
zH%Sqg@1LlLIj8RHHQ`|h3}%SnXXQ5!0}%|IOpFZCL5yLH3`MDF#d3@c3|zk!+-vqW
z4QAqr5HJw1hHH!vuodut@dVd*N|mmAbr_^bQ9uPI4CHAG=)rhmll41N>w97tc_M`M
zg#BQ`j_q$4ets#r#l#XJq$d;#6J|kC<n^fI*sSw1c|g4a0s;y!MInAd>MthRu!4j+
z1^8jY+c!vVJ0v5y8Yqlx<e?C|m+|t~j{{vP#3G~wQw?+ui!huA6J_KA2TSjv-wW0}
zv{v{l5Vq-V{*h-Oo74m}VXBH!a}t5U$xvLBT<i{{nIm?(y$T5M`1qSOG9)xCoFNP(
z3jyIU1_MJVf*FEfMk1KWxhXjrK(23^KPwwM2PYRdFCV{<pn$iqyQ_vUkGG46q^Ovb
zw3W53oxOvjlYzLdjjDx~A~sP8Svh$nWfe7bO>G@LLnC8Tb4vv?8GREsaS2bKC>D<Z
IXD?<y09D?YFaQ7m

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.hpp.DFA0CA7449260515.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_rank4_reduce4.hpp.DFA0CA7449260515.idx
new file mode 100755
index 0000000000000000000000000000000000000000..96e72e1a473508fa605b0f48256e1c863cdf17b7
GIT binary patch
literal 614
zcmWIYbaRViVqkDi@vO*AElFfyU|<jg;^LB`OkG9>23}?chKf0RgARHz8w#9V{$)YX
zL8}~H+m=hs9_Ly2wwug++vaPWI4M^5(f-|UZ)Iu6v9x-67@e%{|E|;~zP{Dxrq|6y
zCy#7q&zXDBUF);kejVfUVn@HEZP+1QemQUJ6;0bvpR>EHZu{(-b)G$3Z^@eGgI~&Y
z+v-_rb+>Fi^^P(0m0)P3&-doi73y6V&Dhjm{cw}G$rbE9(@-VJRWLH$-SL#1$Em}y
znyxQxyckwwGzX~WmKN!{RJny&W$gFWYh38NWJ$POXs^tLX6|F=Vh67j_%Xk};c%X1
z`KkQX67}9M?sl2x?-!EqJI5-s*5X3_`xQa&7v1`|d(m8T`Md-C8|6&rJ*i$aC5dfa
zLhcz2;h!y)Aw7+qMkby1qOpDNcQ>uOUhjKx(&k69<smY=7>X-%lN5oW_e4F+IdxyJ
z2@eA^E07BW9Q>^O<_w%nj0{{Xj2w&%MX71U%YfoszZTqU_BIV>;^7c55U_@8<PfkG
z@PP3I*LO;lu6uPDq)1Uf1ttvSX$$DVcwUb>j?FqRlg9`Y77$Q?35WOzslS+L!wM4S
z6yS#m!)QiMu*-T6{a&!<p|!$afv`<?^N&0Osa6xvgsCb@%}E6LqPQrzSRcgTc;Nf}
a*~Z6ax0yNF*%;VCLJ+_SV=yqVA(#Nm&fYNq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f16_f16_instance_type.hpp.D6CB9472D3A0B40A.idx b/.cache/clangd/index/device_softmax_f16_f16_instance_type.hpp.D6CB9472D3A0B40A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2dfe73499163e1b2af9ab097daf7971ffe2ce9b
GIT binary patch
literal 1588
zcmYk-eQXP19Ki8MceY+vYWIG}bD1<pTs9<Lj(I!ArYhJiBo>pf6|;$Gx3I*^=)BZu
zyu?eGY=$7j=4~VR!&dP+AtVw^5d5LR!ZKzduH*D~NuS&|pS$LsU;1Krva-?=5y~v|
ztXbhJT|^*+Qp{JN)L)d+3!$$zgw|Bw$eFynWq7}*xofAd^zhw}g7&c;WLL$UcbC)C
ziZ@@HRWRlE!fU4wM^2pF-|UDyXgDN>K7FkAM!n;QJnVelbfBy;gV*f$U%mDXZv0+X
z)_LbkWzmVkj!<1)Mb1TKm@kxFJSY5uDxTC9VU3Xbxx+KJ>;CQ2j|(nkwce{K9Z@sY
zhi9N=r!trA8CrHB+upGyr>MSdSBWh%zxKGR@7$`##}cADqh8`k_}k0M_N1!)H;?w6
zuppu0)286GgJgH>lZSF$xaG!XF)y!6E$M7OlCy90%vC>DXXkor_NDdDTi^6{;OpRI
zhx7c#=Lh@HS1YgIP2M)uxo*4Wo_#AiX1M=l($=>){7gE90_7_fPc$dQ?H#tww_(SE
zWLxh9M7k1PgL)Q3+RU*b)YKq0HV+WbB-;p+-HGn>p2cFnZ%JUOIc&$z+E+pEs9eEG
zQaa_e-qb9E%8wVV)4*th2FB<NFvehju{sNkHCSMr&H>{L4j8ZVz<7fPCg=h%!4QCn
zx(G}(L|~FG0h0^~7}hZ`Y+zurE(4Pd8JMCgUTe}$QZbZ#t5WJiQrWqWEwRC3Dm6Yz
zN~1F3MPnKmt<k_3%m8CF1{jN3V64Uh<1hz|(>P!}=7I4V4@|%UFhLW5iC6?CY9cTR
zOTZ*e0){aLhBXXK#xgKjlYuE%Nwr344pvh#tg>d;%hB_B-`U(<d-lz^RR7*9RquDV
z7COy&6BKJ*<s@l>2E`BzD3)MBaRdj7CwNc-A%GGI5tKwopfG_!$%G6_Ar#iy$C+rb
zeEys2u+2%jhPp<?AJ{$0{eQF8#pY&XUtw#rzxvI)(6%!2A+}kSN{`=5t28J^Wk9hi
z3yM=YP`t{65>x?{sEVK@RRV=o3`$mIP>QOgThEXDzLG_;pQV7mATTa=k`!D19uDgS
dA@iT$-(e#QFBVlV9_hSAI1=qi$Zi2be*iO$>*D|b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.cpp.7A9B6FF1C1DE5769.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.cpp.7A9B6FF1C1DE5769.idx
new file mode 100755
index 0000000000000000000000000000000000000000..13642983a81caea3edde571f0c43c369bbfe0c3a
GIT binary patch
literal 1560
zcmWIYbaNA6V_<Mj@vO*AElFfyU|<jg;^LB`%;(Gu4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY<B`|6N*nj=<h`ciHd{%ab(W21i)N6r=J9(Q{_k6ya<ubujmI`I=9JFz<JNZH
zz0~#Ver3=2rEs~VQ7Pj4I~H-T2ja8ayY!bI+&jDHTipAcxZrnxLto|uPl-PA;_JB&
ze}B)7pYpL$y7Axp=_kA2r~FD~a_B!>oV%0dor|U3r>D1L=N&lt@%_!0M;~5Xe)&y?
zk<E>yJv-g~mqkmfUA}lv>~TVUcj5gbH;;y<#LSr^SNqqYdwsm$=9kY;dzM+)O7nM4
zOqD;s?B<ut53fG?^trOE!f#GKr=6U9d6RM2Ew0FyJD<eqM&#ySNWD|A_tPok#MQT|
z>@LqZdr|UA&NU6)nMO|LLh~2?ukl@**0=lWu|rQa8nf4#?#(Q3RlC5|8)BJZ=loXr
z<($%wyB)T3Ej4mgnRYPn)4Jl<T5?mT1TxyZnse>gp&pG?rVDl^agmod9$g)tf8T-S
znpSS%0=JXC0ZF=>+WPh^&gpdF{O86R*mPu$c~!1k(%XO!4oRI(#mAXn{@i-OOoU~{
z+_gD7OxwK=Nrbd6dDxttbn$nb%e{%_Q*SLh*q&9JvgYXrhLl992ghcGrb_eAz1!=e
z-5gNp{_}>xf0y*cgL5ARaZch=+NnKjMxW^duS);M_Lzm!1d`Y82wSp4Rb9N9ceT{@
z8s~X|9!d&34sR>vVP0`4f-B?r>YY<&Z~Z=@cxiQ;bAO)Wv?(e9T?sql1f<eW_;ZRb
zZC`M-hp&C^mFd?4|2<5QZ)&}<N_YCXT7m8Ps#^_<lk<7*6r?XImwglWtg_D_`;&~U
z$vp8J+*1Ay$(F^+yF0#L+_POdAvAl{jK0d|8oq9U-KuPvZ^L^e1Lgi_Okk6|61Fw$
znqI)gU19f{mb(_LD%hj9xbxTq<G1CL^>1F{+o|!`GexQFVQIStqt>6As<J!#{9noj
zZ=E+YMfIE1nJuElLSJ7b9y>Z`)#L-2p(_pU`5sUUXrJV#IC+ZYZHq}>PmitnsK?jX
zT<d0=#AS9z{taKolAwRb4D<J^Zt$6Be*V|ZM{^T4K3{t}r#DtU?XT~Fu3d@yb_dO?
zIXBrqVbk72n@nduUA*_(_4j9G%s&UZ>0aQF;d6;q@5$=1)Yke@Y9RGy(@di^AzKAb
zezctbxFzaP{_MABc{Pss+ZQDWzuNOgaBubP{Ix$^mUO$Exm;q+mw8Ll>vCQGEH%r8
zev-eZZ)X+$9DVKa!~Ks>JHK48MBu-NNx#y`BDv>(U$rY&)m8qRKk=C5&v=GE^VZ&V
z&bHo@(kf7*8Tr=mKtX@iy|yx&d&gcp->Y0DadMaa5ze!XDTz~k{xR;ksC1K|xH30M
z5t#3<>9sG>pS(znhaoVSA%dTk-#iRNFmN(4GDHV4hA}b}rKS~&GBPl*PvqmTWt+W?
zktae}PuLHpk?Yrjd(Gaa!Av|60tN!sFkv9iR=@+sb8LUZ@bgQ_Ehd%-Aw8i`m@o^9
zNnVdSj?FqRlLxdwKtMnNrYOWuNd3h`8&;4orvN`pIOW{Kss64Jtw3R9BM*hxy^NQ?
zejMl=Ar>Jem};P#S%l#{m?$F`*vGwxelJ+_&|2ZIK-i|c`A42H18q_iP=Toe^3()0
zk$BnydN3X^baE1b!N^ctlw9l%VuXVTFkp@lnam-h-Bl#S8W|EA7S0d~lVM;8K`<i`
z%rFEqIX5LI1IYDF^Jisa=iubx=H=rT5)|+jc6Zef=J9qBkrWk^lD4w8wX=6{bTSav
xwNbUuQp6@IAuA`Zq^zQ*uBok~XJ}+>YHq1uCZliSCNAOW6UE{Y;Oxch2LQOHjX?ka

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.hpp.901763453949395E.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce1.hpp.901763453949395E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..599375e3d9934de49b72db608a37cc78ea251d31
GIT binary patch
literal 602
zcmWIYbaM-0VqkDi@vO*AElFfyU|<jg;^LB`Oie}x23}?chKf0RCpmgE8w#|34z3DV
zdBN(<JKZUAE;iy{<SwOblQuis;#sz}<Nw`l+0oG-l{8cuQzm`4yXVrg{vOxNO;b0y
zPwq_TzhT|%AM$)_U2Muc+pwo8B~fBlOP{x`j5ZaYon@r;?DNH5Zq3&T27B*!eU$gS
z&y~A&Q?=t%HrLSbr*)h7a+gZvrf&#1Zhq)Q+>EeOJ&J<bT_>!pxq|qdg5=(LX=XhS
zWO8uMZc)ozJ^8ibCGJg^9Dn;3aq_HHR^9t$Mbwi9xfp|en>m)XdP_$f<zrK~P~V@O
z^k?nDa_?(<erDBJ>|(KSE8OruE^zC;ty!PnyX?-fyWAjut189j^?OOpqda#GoSKoT
z_KEZJ7LCSEqmoX4m+MdemL1q;R^xkd^5#d(7Q19-F%(zkCMg0#?wVfv68*`G#CRB(
zS%F+2;NWNFH)r5vVr1ZAVdP+BC`wH$o(mM``nBL*v$tt56Ay=gfq*q!BZq*kfCr4{
z^{C_6tn)H?j6g*K0tztU5I-UH7ZYt*LBgB@{4ik{&BzJ1wD-{O1#2E!EBqA*+jKYo
z$TMc3YDEDRm?|JoO+XWgr!Ald;{jcnlL&HsaZz%yK8V3lyscAW$3aJbW=?iC26m7T
N1aQI_3=C`tCIDmE*i!%i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.cpp.B1019C659E1C387F.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.cpp.B1019C659E1C387F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..320b79c5e3f4b726555bf7a2e0874edf6c563add
GIT binary patch
literal 1560
zcmWIYbaNA6V_<Mj@vO*AElFfyU|<jg;^LB`%;(Gu4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY+mY9}d>i-d<h`ciHd{%ab(W21i)N6rruDrI|Mx9UIof%-#$%fpb4q9VacjHp
zUh4XFzp`iiQn*~ws1))29gDcv1M%7IUH5&OtIZ|*<KE}Q1;6_n`Z6bYO7xK*U(dbx
z`+IKul#h+ljsM<HKiU00<yR_`L;u<0+?_1%Tq*;enCF*QGN~WmpT1nY-F?2_=EWxq
z>bpNale3+fmpjAf_+s5-|L%!B-q*Rgd-aJO7M8ZZel?24@1K))`TX?BZwji)3{>2v
z)!NKXyX@Z{KKb<d=Qkf(7VcC1QBiY;^ISxxcFNM{ld88SY>i!XddI`6r&H4pg>C&>
zv&?YzV&fGz*91lxrMBd$*g60IIV<X{O!?~G*3*Sd*JHA)F2CdQT_7s8vgkrh+b#D?
zmaiX`H|B|YrFMDTVR1jX``p^q6(K=xtRGesM@e%VFPg&X@cTkVs_${J^>KS^nFS+O
zZ#mj1rE2cjx%GsU)u+BqszRzimU1m(>9oxMa;xXat%VO7j|jCs>tnz4IeTHQ21kH(
z^v!~7zI=}40MSWp{OdYh-&YG(y59@Qn$0G^>g&Xf(+)FFXw7Vp&JLTJYg1BYF0_`%
z>8R|}<imf2=eDwzcX_LNXmvbUmyu+3twHOlEvsBnYm`&Z)<;njA5EJR!(+B4E9NtA
zrJG`B!vntC*9>_Ua3yOlkX`>QDDT=m=W|ovN%Gn5VvP)%<e=H~u*xxGj-tJ4$P~FI
zF=H#4vY_}V_n+;5Z8^k}R&R^3`tEpV-_$F|&h*AgMHHnkDwll|_pGwdAp4VytjRp_
z8{AU<4#}3q%DX$hU)-}@IUzKA)r`KKE_Us%9e0JAw|vt+)}zs1Z{pF+leIcZJ6n83
zMyz)J!OKb(p%!t|FS;b3nE7q@$>}+n?J-k6su&6G{;<n=3WI2!ovn4={pDZ!Rz{uo
z*(6lWn|7<qva{-o!Q&*qtDX<GtO~KX;d?+WpnZ~`;^Zlow=E`lJ>`pd94o{0ySk_9
zh)B*3^Bb}kJeK~EJ+kNTv;)4M_EcUwFZQ0L+crNoE&I(cgV*^?VOMxR-<|sDN%6cI
zrpa$br>#AyBj0~3e!qF~oioQJw>l^nm<W{4HCSc%Y|V-X*AG@~IGuKCL(o+xmE+HA
z`h>GM_vdaiH(wxN|Mx_r?~2cx+`qib-M{sb@I*Z!L;uTfEtX`aYxw@&n>p)Ai&g5I
z_<LeLXYwQa+yD3J^Dk|l<n%}RqD_~|>B>3ZSIhUk`1$l_y_?kYC-saUs<)N#uYLPr
zlBm;#r72sFGaMHDx`+2|#h%`a=NBtiNu1nee}waFV@l#wpMQ)G^pqYj6j$aZDFXBT
z$tmeeu5iwO$HNd9%n-rP%5NS9A{aQC7#X627{eGDic-^xMHv|w*eCMw*Rsvt#>f*P
ztS9UT)5!H}!M$d0(_ki^2mu2DYnU*QXDi?V<2km!Vfgu_<Q5Z4gpi(4C`_0I#U!ss
z9mi&!m&pTKARr*108<p=C#3#jq75rZm{WisCY*Ba;Z%Ruh*qF5vXO^E>|VypUq23X
zju4BG5==GF%`C!j9!!*x3+&_GL%$cSd1$ThS0HTD-TWiZn1MDa3aG$T0eNZynn*lt
z0X-NG7&<wLz+hx3E=n$T2Qk7y1Q;+!6q~IHT+A2i&l(vL8WzqF3X@@A2thC-5zH_I
zGdVXUCj-d!P4j1EW9Q)H;^yV!7ZMck7It^l5a#iA5s?%XlajWwwzac&aC9;d*R@f#
u&{D)EDj_Q;ucWM^rmm^2qi1MjY-(<)U?!t);wCQP=@Z4`5#a2_><0k4r;~{Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.hpp.E7E5CB6E00C1ACF7.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce2.hpp.E7E5CB6E00C1ACF7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..737214420ac7bc771ba6785af1e8a18493d0bdec
GIT binary patch
literal 602
zcmWIYbaM-0VqkDi@vO*AElFfyU|<jg;^LB`Oie}x23}?chKf0RCpl&_849$14z}6i
z>S(>An)}hhDf3#t$t4-RZtI<Ca57JO>Hod2!q!fUYZ05|V05xtK3}QL{C)7uO;a~H
zPwq_bzhT*YKIG}vy6BX7wo{K*6hw(xEuB7VW$ab)*+n)Q&pscV%dPo3!JxM1$cO$V
z?oQFG-^2whjSCD--8g@ZbFh}%Z7H2AAA1cp2``&yl<aY&OKH=&zNRUACroMI9n^KI
zatcGhWvPXJw_ZK5>iE*NrsP81<QOLB*B+kwwyW4I7qE-><m)zNKDfjdwz2gd=ed*b
zUrVev{gQWbmihf&e)s7N$2IuA*=w)5v^y;K-`tGpXTzmFct`j5r2n;5UOJ;OJ3`X<
zq>EkRJRk0aMLfBWgfHfP+#tJQ?brC=k|%3^us8(E%ws67%uP}RhTO?1=}WF~&VR?l
zz|0Ed0s#j<E5A7dClezB7Yic?BSTSYTJc<<IM=TQ_nN&;gPC|Z1PlbM;TkyvYy~`E
zJg-L`$7Y?E$zud65)e>;35WOzslS+L!wM4S6yS#m!)QiMu%*3+elJ+_&|2ZIK-i|c
z`A42H163;usK8VKd1?ZhNIY!;Js1z@%A7=y>x+w$i}gVab`Svp9PR66iXX5@USQ^A
OXJg=m2{SOTA(#Nf^yWJN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.cpp.D8971A21BC00D0FF.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.cpp.D8971A21BC00D0FF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5f011f2eec86bc2d2313772fd963a1bcf9ab882
GIT binary patch
literal 1558
zcmWIYbaUflV_<Mj@vO*AElFfyU|<jg;^LB`%%{u@4DH+u3>9<EhQ;UKHWRq+{_?Qd
zmKW!G9_wVc?s9l^BzuFJ;VqF$k&I6&g3CO5m%OPsyzh74BvEz0&l7TV*-i+(?<@av
zH{Po9=i&InSJ-DR^A<U{t6Y6Y2$TJ7n~*<gVt01G%=rIY{Py*p>iSTt+agc-3>Q8B
z`&-;z^D(nI)8G5+Cr!V9;>&852D#bKZxwOg=`4zQVxIrKf=PYy{mqxV7eDS_e)D5V
z#g4-j#r^Znm6`dZFa9fg?BPMa$NM@rcdtIN!@|<`-zRC^eKmGx)8^|fzj^UZj*W`j
zv|5YVshj)V!!Muqe}40#W#K-x9|ki^nvBD4aYeq|`6Ny^GB^B%QQnTY>Zda|gy!1L
z&pspef+x!?f66o;$%WFL=P%UnTNYM){BBkD!=fn%Os~(nw`I4Ja7L@!DjAdci@!}Q
zIk)Ra?4s!QpqWZS<_7~mtt)=5B{y|SAfwHzIoFOInxm1*bivLfF7oomqpz>$-*;fS
zruDX9f!j&nfF$KjZGC$d=XAPo)-7$)a7;QUz1L8AW0^*U!bX>cJ0CZG`Qw%$&Bb)(
zT$tINS%;+$Nrbd6dDxttbn$nb%e{%_Q*SLj*q&9KvgYXrhLl992ghcWrb_eAy}Q;$
zyE&lH{pSs%|1Rl?C+9v2;+({#v{QT5j6U84UX}ih?J*0d2_&!G5w>K9vbuOP?`tXZ
z8s~X|9!d&34sR>vVP0`4f-B?r>YY<=Z~Z=@cxiQ;bAO)av?(e9T?sql1f<eW_;ZRb
zZC`M-hp&C^mFd?4|2_P`(A0WkmG1O&wF2AMt8O*qP7Y@ayEF4eEc@>1mGhL(%(ii7
z?|shtt?8KXg@|)^PUt<VHF^J*BjV~Tt!2sUCg>+fCT%~Wko%|0Q2EqjdEQG3&9_!v
zJ9UgzE9&~G=nq>s{6zb%r<z=pC|dS6@3U(3Ek*HEn@*o2dVAgps5-RHKYvc`?e8h~
z3`1wn^bAW{e=sI`qe5W6+_M;+6;_;c!lrUP-pi81RI2e(#K?2z=3eEMA$ez{-aits
zut~V0ye#2x^TB%!Zdy<CWj=q)W->qd`B|iW^d4cIZ+l8-uHE?h@S6Rc5rO7s%EM1h
zK41BX(`Q?F#5TpvHCyig`?YuG<21ulD_Er(7dNhW-4fVqnC{BDpU17))-1ej>MAj%
znR7m~`mI#`v+m}-I}J{Mz9_SsEj)Kx?&9w4f3J1<yX^8gqU2ltc*3%|3r*gBUzRyZ
z_|=q+^_9^E8+WAI%Kfw6XETM*i|gZem(LoLbe<R9Tfg6W(evrg|4U}hnf#yiK>p2J
zwp+Fx_73B4oIdHQ95c`NmCtiCjz7-rsk>vptbO8|-_CO`$aqW(JNZPOd6v?1hT_WH
zBt>AZpBT0#MVh(mBo9MiFhc}CE5CUdh+yDkVq}O8Vhm$sC`wH$7G-2$V4ujxU&}Uo
z8zWDIu%56VOe5E?1^1f0O@o<uA_NQstYN}Ho~?ifjOW<?hT-Rzl3PqH5kh)Gp)g?<
z6qCFjbsU>@UM3G{fq;O30!&eepOE^Ci8ibtVNL;lm~hIuhg1DsBU*vN$VS4X8M(lo
z={@v&!J3EG3Wq}MUdGE`KMr)I5Q~r!Ocl^8EW&W!UxBbqck_=tV+Pu!D4>E+B%mgs
ziNw<u(1Y=Sp_7vc3`U0HqU2(C5F-pkKtMQ*!N45x>&Wd584{kBtdSw1Vc`rRa4CjJ
z1Tz%DOwLWo$pCVF)BIW4*f}`4xOw^bg#-n>h232>gn7JOL?lJUq@=B^ZSCwG9Gwir
yb!}8Fv=p(4O32E|D=Dj}scUNM=ouOro0?lHn91myxQR=6`b4pK1UP#!`vCym+KWm6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.hpp.5132DD2F1F4E6659.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank3_reduce3.hpp.5132DD2F1F4E6659.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9bd481892b41fe64f386bb5a37bb79238a7c219b
GIT binary patch
literal 600
zcmWIYbaM+}VqkDi@vO*AElFfyU|<jg;^LB`Om#*E23}?chKf0RgARHz8w#9V{$;_X
zrqUPhbf?I<*oc3TyOgp`+H5xG<TqJ@|L<<gj*kAQq~XJKGUZ<N9&y8cRkKe{Sv$pc
zmXA&SfqRlQi?p`?DxF#RFG6hl;>+8vXxd))xw~th#gyAoI&S8AOV%_W{9-5g*uH5V
zNAB89)eckHTtcTFt=q`wyHX-IeM87`$;1?WALZ0z6FNj)Q|8LDYO*V6^5-uVnfh6S
zaltaNg?_hQJ*n#W(zK@JLfzyTCg<EGOTy(sn`JIEa~~@gJ7`ki$^81p!2>MMPrZL7
zQSbfYZkK8PJ|X#adaN>QEiTl*U*YwB(XBtb7tJ-FpLc+Nqnzoyr`5}*B(bea%spcw
z{IjLfO1q)csHD@sYhCy6+=JW9YkV(G-u#H!{f*2bhT_WHBt>AjO$=L;BF)@&l81qr
z703kw4t`dCa|TW(Mg}exMh-@XqSUnFxj=EQUkmOvdz%I`@o)$j2w1~4atPQ8c))mG
zk2;RcIxmyQ2vj5>pa2sN@e@*iG0}z<B+Mzm4-<ybjGSOgdk_6yu;!t)!e4=~O?UH;
zJYxo`RuoWysRHuU1T>L&+5&no9?+FJi6GY(7bO?#gBa`}0s=U`SLH{qN#NYh%*oEi
MzzGv(U|>Tq0lhERH2?qr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.cpp.70DAADE138E1FD1C.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.cpp.70DAADE138E1FD1C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0f68683264140ef36bafde8bed8219f0275dccab
GIT binary patch
literal 1560
zcmZuw2{e>z82*j1X89S*s4&QqE#qbqGqUeXGR82vQkG#1g~Dm4=!!CvHN#bwBx?#u
zmS}`*5x2Qa7?C|jsZe)x&RpGd?sv}jp67Yq?R(#Iz9V+l)*wFsSoqmRll>!op#T6#
zaBf6oc+fO203_@HfM{Ab&L`||6gX_t(TaLR>`j90BVDaGooFLhsv#LtF;c{pJrYH>
z$py6N*2eW4RB3f*c1nm2|4YfG8&MmhK1MOi13s;T0$D{4QZ4mS?R8j4(_d)pMu5!7
z)0ywgxtwki{aKTC4mmOWI{X<^uJK|M`?|i#NtM1IoBA=!IgU+^z=x`sa-EFk9F2CX
z3+yu)>L<}(F~$cEp}$2^<((1)2uTOlKFoxWS|e+kYF`+_$;icQLrhar5}S1^ze16?
z{He<hwSErSEf#o>rG$Jq&zjNBK5x`O`~93rmssyLwoXfXM*)oX@HOtGeXr_Ta7Xaj
zWL;MLT$)F&5aG=b`w2bda!+0W+d8NtRid8qMPzfD6l@SRQihVz{S2#{@EXWI<3WU!
z_1UJyO+^fvP<uROn;aq^9M*KRoKQ2gXS#<TeUVSq*SuttGr+)S!U`P$Mgm=nWK<LV
zX?9`PwPvqE9kD@;Cqnk6RzrQ1s8k(_MvKn_gnddT+q$CLwp3nN!B*RqxEPJVG|4dH
zV}ZxjX`;1LALI4--IapAjNQhsHp)_>CY=gvf<E5yQj{YdL<63g{B>0U?r-T9fl(^`
zoe1oRs|o7Wp?*(%_`N!8Y(cE;Vxn5n3DM#sOt&#5k7Cs}yjHpN-q4PY7spyl=mzgn
z0;Q;ef{T@8`BtDrFCD^I4%U4iuz;ELp2;2W%O_A(RfSTO;pMVd=yEG==2kOGL}Y%e
zullFL)2SB?;bxT};qm*|PsAramldD_eCHouCM`OymdA~Dp?f8!oO1Ky=xoSYGnNoh
zJ22EwcG=|i+8JnRu#@xvvVyYjnB4K&HbmKGB2uN3r6~Fu5yvPQ>SXX~dJT6HK0Xq3
zmcB9UzrP<IUbGpyqDo*W$@-)JaBS0Um<3f$um0KVtiG^*g^6^Z%d*2AuW0cmR3Th%
zDolB67aIxDaB0G<QD_L{w8o5-zFk&zvdVpI=&PMe6QCg+IV<5-IZmuJ&ZF3W-PPW(
zncnQQ(hxJ$E3*W>jtM&&K$`!iHx~xM4T`cy(^e)Ua?e3-El8)Ab!fkPN4mJle08cC
z?mq}gZXP!sJpH_wo>_f2)<<AH?kLJ5KCMJXIg7T;=X_svqvkDh<dXTus)_*WPEVIY
z|I^ydp+pzUddD_pr&?A@QO=$H3u{F|RM>24)x{WsUR9kxM&Y}0BRW;gUU>PYC0m24
zJ$;H<al!EJbjI`#xggrDAAHSAZyA_JHLVUfL9-0%8wFmGwc=?qv7JRuy1+EPn?NaD
zx1jaywoS)%Oe&6fD~|yYG34V)oPS?`#r(cQI3rmEaCHOl;(X!<a9aitg75%@BRn`B
zAlyG7LYfBv1Sn$S=lQd1ctr3pU6?aBQTSUPGtR-l4I+XE^+02;%wZVNmdnf%=7F!P
zkqihFFR3f(&XuA6sboKKjefg#cPOU;2!aT15Y|~z{R0Kfw<QaK;#}FUXFSct3V)R&
z|5}pk<`MomO!D*fywY)F#B;34jE~}QJLi>>P)TKO6lWAD>|eGC;vSBMwN3GIIw^sB
zwu3-5aPKcn3)JN@&UQk4Ico$W!cRolYypog^YifHodsQRC)1bv_&l-hI1j*`>j41l
zHuT(vxNUfX>=#1hP^SPFK7Ii~A>kdOV&ak#paaatN&_b1U@0XhEwgi%kuk>Pkg1uu
zo~#a9)lgIE|5UiVf}*m@9yRs7TH3lu{of1@97LdY@7r%J3%7Ihg4!Oou;+FD2a#@_
ANdN!<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.hpp.D21E6E14C4987978.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce1.hpp.D21E6E14C4987978.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0b2c120f27d5569119f7a65f7850fb86d1eef493
GIT binary patch
literal 602
zcmWIYbaM-0VqkDi@vO*AElFfyU|<jg;^LB`Of5zR23}?chKf0RCpmgE8w#|34z3Ew
zTv&Qz?ec&sfsZj4sy!y(%)4PCtakIN(!cWC>vD6Cb}yK~GRgDa?+RhVJyqH%r@~J0
z`KrycO?b{@7dZ9$uhNN?-z;>GH(%a#LDTlS&)Ze|Zrtf*KU!nAqfjwbmS@@bi4y-N
ze(}i2o_d#i)fMekSyAuQUthSR>2_O6=gP-kgH6KACK@Gs9O+Wpbgr*y%H9c6+II(a
zovNI|5O7&)q2H}nPpmq=G_5JQP&YY-$@#U1r@rkfHp>O<;ywAgO_>iav4w4Hy~la(
zWd3W3`s^2XCz<B&lalY#W9$o*d9yz}WNG=@TYs!C>diiH@}Pa2oaFgm)oxyf%-1&<
zrA_LrIb@^hm9U5>_mS|$(#rj^8`ge}4=#DK<_CjArp&zJ%G@MHV94!iaa!qJbgP$#
zfteM^1p*F!R(^8^P9{bME*3@(Muwu)wBoryajstr?lpUx1~c(+2p9-h!!>dU*a~>S
zcwUb>j?FqRlg9{DBp{#w6AtkcQhzbgh7}~tDZmdChS7|iU`u-s{a&!<p|!$afv`<?
z^N&1Z2C7yRP=Toe^3()0k$BnydN3Z)l{twZ*B2Kh7wdx<>>vUHI4a87c7Lg=F=pmu
OXJg=m2{SOTA(#MV59Hzi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.cpp.C8298BB69F4C257A.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.cpp.C8298BB69F4C257A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b38bfaee335ee353f58f9f4c2bb1217de72368c
GIT binary patch
literal 1560
zcmWIYbaNA6V_<Mj@vO*AElFfyU|<jg;^LB`%;(Gu4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY{W{sL%N*|AVapa;>^rfa$#<U0A(536rwHfYsJ|b$>Cur)JC*2e#!W|dKbD<e
zy>zm;{a;g`+6kGvTm;utm-l$9e&{~yd^A7p;NIDN?(56V)=Pf>x2a^C<|*BzPgUt(
z{{20tf9l79UWfnRPd_<YZuHlPal!FtJGaF!l`p!Z|7rI2-8PMn7vJAJ`E&DOfAyOW
zOLlBOT2b6T@7y*spY+EU#gg+k9o>=t=*u4?qjUZKeEaJb9KEiue)G-eqbAlf=Dy?a
zoR})_pMLYr<%3VZJo+42R^T=#pVLlGwzMfY>=sw#%bic+bR%=uUrfDIu=mp`<HXgs
zs_d@M(7hmeCFh=o?o1;mbD{YQ|JV4geb%@8>ajydEgH4gneNRjZ&ka%)f-}&Vdwr<
z`Q@C_kGoyAb1gM;Rmp2w{3Q12wNM)^&BaU=S1i}|wps?BVp&j|v1ik=<nHkG_x3e%
ztO>pK$f0-A>;*@nQg~%6-ERtY3jNv2wTPwDGXKl1o+GywK4?55)cUNC{nF>`g}E9W
z0oKts3$pq0Ig$fJC$;gf>vVlzEm-M(FC=R=oBXP;6E{vf%sinrvq3sLY-+AeNtwCO
zS{|pPvQLu_{}G<ss#@OVt?HrG@nl^_lGU{at*5rEaz(9CPCZ*6MNND(ZB7i2*_y1F
z&%Bjxik%G)_-<b_<XOO#thqpT{j;FFYxkVbO?@ZHXS+)^GH8;6X4Ate$Ba3O_NpOM
z<eJ2ctz^oA;-lPuw(sR+;ocY;tvC0#Lf(C^tYoocand`UopX6_bz|?D&oYVko>*FC
z*ywK%H<{OPtmv6rnc(}ymHF-mR$U7*l>IFAlWCe_nK#d+Tk9o^7uo!~sK9HqVr|yi
zHL(j8m9E{x>fik!<Y7gun{cmU`mOhB@tc;)76l%ge8T0;q1AkWOd&shzWK7EcK!wH
zWm&eyC%xX7&dAmlQ+ajq@SJYeQ1!;ktAY~i%^7Dq$a-0|`UE}8dF-Jz^>%u1c_GIe
zUx}|Bnl}oJZpdEnSo-Jf@jZX19q|3Mr}El)vG*+9w)wGX*>8Rsyv}C|yTbeV?$l3D
zis#iZO@1pnZS6@N`Tkq+`^}5*=%`6<bx<xa5h$H&u*&e+niUVOAFSAL`q`-sL06qr
z+|z3sgR?mI=Wa7MUm#%r_e7)biqD(ezr4%czx9#uL_Oh+9+%%*EXhpQ@cq3vbJmj<
ztJF8~_r!d3@*?}&|M%(hFKwRW^hf!kO_$2)$~oUx%lEzb`SfSKo7D3s^^6~?x0Ug)
zefwdOsMCd|DO-;-92We#hxcv8p5BY+7dux;oZMx9g!61;O6F9be~ip0l%6vbSLP-u
z0`vXD_4eWN#py*n41vK65&W$D=3yX$fs=`mAv%aLjFF)zHLX~bk%57IA|HP(+w5(O
zJQ2cr!hSG~T)!6FYxXt`X5xtuFc7eY2?Key0v<4)WBVJ1pI=IDF|kAl=?R6xgjrBb
z@_N*9Y}R?1JfH;v0s;y!MInAd>MthRu!4j+1^8jYDd!$e^>>YE1qve@c__s0WxV|L
z<3Q&Ku?Q)_R0G}2A`IujL>aljKJGpAd%>EA)(U?G!ZzK_Kk|$jXp^FV3QQG{rzW6@
z#M2hggYkf&lamMxMuy^|<YIRaBMd}9Ksb!Sz#OqA-K@Es``|j($dJ&maE1`L6hkC}
z8H!*g=ceRj0J*+t{;X{59GqO-ynOsZf&$*c?yefbJl-xMlA>Z#(pJ{CcJ>aAP6pz-
wHmVj{ir7RYWaZ?QlvUK!HMMp042_IU%`FwoWb{qk#3ej^qF6ivoV}R+0C#kVApigX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.hpp.E0AE02E34E90B0FC.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce2.hpp.E0AE02E34E90B0FC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..39c9db3eab2a87f361095c16b9d9898c08af3ce9
GIT binary patch
literal 602
zcmWIYbaM-0VqkDi@vO*AElFfyU|<jg;^LB`Oie}x23}?chKf0RCpl&_849$14z}6i
z>S(>AitEwBDf3#t$t4-RZtI<Ca57JO>Hod2!q!fUYZ05|V05xtK40mV`TOLVo2G7Z
zo!pt;f5WQze8{t{b<rvFY<;^D-*3?=@zNI!-n>fB>~iAz8Rx6Z^_Hw@KKR8@;II88
zIgZ@5o2nhBTDgY$b^T3|eYG-aOYcD^+4~&I(TQ1_#%(TZG?ZSy(eSv*=~=l$d#T1;
zcTNV;wSudbto2*h>aqTW7yDcDld1+=`lfujw0PA-=Bk2&KaMHQkkZRd5S4y0=fRvm
z`?@|xIlrH_rt<UM9|a|x1wD@v{%>C7wKx0fv-_Q8H*1zM?%(qDM8&oH#z9?@JDR2%
zF7|mM{OsZm2DMWcRO|%rit*oOydCx3)Ybc3m%j&-%p!*3%G@MHV8}gOZyzpSoL<Dk
zz|0Ed0s#j<E5A7dClezB7Yic?BSTSYTJc<<IM=TQ_nN&;gPC|Z1PlbM;TkyvYy~`E
zJg-L`$7Y?E$zud65)e>;35WOzslS+L!wM4S6yS#m!)QiMu%*3+elJ+_&|2ZIK-i|c
z`A42H163;usK8VKd1?ZhNIY!;Js1z@%A7=y>x+w$i}gVab`Svp94u0HxnT)QHJLft
O*%&xs!VC;-2qpkFu-cRW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.cpp.25E3FD604E588BD0.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.cpp.25E3FD604E588BD0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8e8ad0775e1fee761d3aad40b44bec0a26a14849
GIT binary patch
literal 1560
zcmWIYbaNA6V_<Mj@vO*AElFfyU|<jg;^LB`%;(Gu4DH+u3>9<EhUMnpHWRq!f3iI{
z<KjHY<8jxxd>i-d<h`ciHd{%ab(W21i)N6rruDrI|Mx9UIof%-#$%fpb4q9VacjHp
zUh4XFzp`iiQn*~ws1))29gDcv1M%7IUH5&OtIZ|f?!9+&&$2teS6{lhOk;b;$JcW&
z{{C*gU*j>eIn&?!`jf@)pZs!~r9p1?^IJuncRD{UIFWzv-6s~mzWVci`h5L%bJN@>
zKl~^5_^fS}(cN1HGyA@6llc8u_gJlPx_J1+!os4eFJGB;_tjaRT{d4|{pQ0jZxTFu
zbbeQu0r`CE)uzuszxi<SqZ;oIAAjr+p0n{%*a@%ms^7B?X6<&FUU>NHDedzuYqMVe
z^hq{zPY>7}xp+&;DUQ1yHEsVsn{J(D`7TVFd-`IQbvv(pS$0=+mV>rQ(9;D!dA9a1
zDZX~}J=0ySDW^mx?%?d3R6aL4{NsuxJzNh0pKmb}OLNf_ZuqwF<EdGFx^a7}ezPfU
z2+QndHucJD63U)vTJl6T)mz2ugSV&)r%>^|7h5Gew>lqW?o{PEE6cy+`E{q;0Rjuk
zwrzTNP3E4!u?1Qxyz()^-S@sLed^n@@`@R+UC3+YWZial1@6lX=GWF}-L81?CQl_w
zvZc%V)bX~T>Q>y`??k41O$ZS<8GG?a$r{GcQ`KBHPq?<WNMs+~s&rJ>a;HR2#MK?o
zq(AjI2{9g&-L^Ja!a?*{kb`yH*`;^Z>}{W;eb-o~s)T#f5|zdvmcw70E?79%d9Bc}
zWzkJ5v3Rp&-<G~l{J)bq#FAETi?RCdcxT_#E62|C#!5vLr7tR%eG~Vrvd<v<lZ>p%
zJn<XcQvMFfmc`1uJHB7svt2nMG<(&IzMU?1?XDeng_^f~(>~Ut(O+-k(ae*zI!ZfR
zd__jAcK*T3N*19OanmolB%hf1ZTHFPIhpM-Q$DH~3GV)|%XtceXq}y{b>98uU;0)?
zo%h)!RLz@qtIM*p>WjhSB)_Yk54NldvAE%TKrNtslAq$_DVDb_CV4&Oi+CI>!}L4Z
z%{Gb4OoqRV-6Tk}p4~8iziPo`oBVlSHy^b=m~=jR`c0|bc4vOgZV)X!TvNW(=I0#s
z`3F*}TT`=*PP<pXU4MVJW!|&LJy8n;Eu=bj`AJ-rC=Ltx@H)ZdMyhdY#L6tiNso)|
zj&p8pxo>uRwsb(p{JJLx)UH(CRId8I?S9n9P7m?U8OvXm$z0lEJZ0J6dnUd`PIHaE
z>F0B+J=?zK_~H8F)7vjOdMf^#m?7sn>8Z`R-&gtFzx?_1$6mR&_|txdAGXoE+pm>X
zoa9n`5wz)6a>GNpuX}jkR_y7$c)r^Gi@~JQ`VQe)%qI_N&HTfxuB!Bop|~<PNfDUu
zh0@oS>3omL=V1s8W{BWt<u?xl5e%G6j1191jA4unMX71UqKpg->=XI;YuRRRW8{et
z))V%FY2^B~;9j%0X)qH{gn)s7HB1=DvlZ}w@nAF~7ubs4L%$cSd1$TR*#3s$=a-UO
zOe_&XdP1QvRV*kLcs=SkHtW1h9?)I^0RaV=q7XkJ^%oOuSV6*^0{k%HlyeWK`nyK7
z0)>%{JQQO0GG6}raiA-OScH^ds)1f%5r*^r3WROCn}6gPGtfvy0TqNI0W|?lB%Zc_
z9*hSJot#8qFftSuB^SGc7~voS0zzR72Ih$P-bRyI36rH*BSS*N!WkmrQVd}TW(b0r
zoSTx90p$9o`LnXIb8vET^YZZv2?}@%ySr)#^LV?6NQ#O{Nn2Ul+SxleIvI%T+NfG+
tDPj|qkd>2HQdUt@*VNY0Gc+<bHMdkSlhHSE6PNJxiDK~xaQ0&M0|5RZf8GE9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.hpp.33AFA8695A15063D.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce3.hpp.33AFA8695A15063D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e295fdfaa31b8f3a10458f859f2a4658ceaabc6b
GIT binary patch
literal 602
zcmWIYbaM-0VqkDi@vO*AElFfyU|<jg;^LB`Of5zR23}?chKf0RCpmgE8w#|34z3Ew
zTv+;I-SU7cfsZj4sy!y(%)4PCtakIN(!cWC>vD6Cb}yK~GRgDa?+RhVJyo+$O<6ld
zc9xG#^?`GeHH);?e|@d=`OQY1W6hU0UC^|>?(=rlz8iOX*`L<f?I=`CmE~FXeWJv_
ziC;W2vZvnVUUfx#)heBDlS?z+i7d|Lo%ZU7o5W4d%N{cgRgzo<Bh%d-PsMqhIvlI%
z`qajYVMS(dfO>9ek*rIVW0+OOeowu|g{4cDOqUC7mbuW(eQdkfL6d?^=G8Zx-m^S^
zvc0rtzv-8}le5g@d->g`GaT38`*vS@)ur8Gx&P*7Oh210^}#v1zbE~#t@6?tjoA^B
z#wT6u66g7FCoJO0eI$Ibv~a)dhP7YggG-*Q`N7~nTV`2tWp0upFyw^N*OuvgkICm@
zU}gn!fq;XbmEW9!lZlami-nPck)bFxt#~d_oa@(ud(Gaa!Av|H0tN!saE%-SwgMh7
z9*kz>1RK+P==Xv(53Lov9(5d>bzUZq5vW%{Kmn#I#7{{5#Y7uckT9nJKTP<qK-i|c
z`A42H1BDd@RA9nDo|=Fr5>Hz|55@z!GA9w_`r@MGVto*U9YjC?CyW8~+=HD5Ihc~U
Ne=u{hvoWwCBmjg^+p7Qo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.cpp.86DA1C4D6D7DC8E7.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.cpp.86DA1C4D6D7DC8E7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7e9f9bfd7c52bf7b32659fb6fb75dee2e6e208e3
GIT binary patch
literal 1558
zcmWIYbaUflV_<Mj@vO*AElFfyU|<jg;^LB`%%{u@4DH+u3>9<EhUMnpHWRq!|FS(d
z<KkS&<8jxxN*ni7nqE_Jo2?|zI?KkhMKefQQ~lnC|N9oF9PPYZ<FQSQIi<7wxV7DP
zFLnL;AK5d0DO@gTRf_ojjwRgdLHg|WuKQ=2tIcKK{{NHodsgv#eed+yi|!~M`?9D0
z|0DI(ng?zR{{IgB^htVV-BgAP$)97Q^%|`M{kr}<dn@Pv;K`5gZ@&Eb@Zxc1nTf32
z8-sgy7BBxim-q6|jQaLMgL{d0%nhyg_w;<8e0j0{{)>-JKfj#w?`6^~*>mZW6%9AN
zIh-@^ti4>EpT6DP)b{ivKW9BS{IO7d#>PuwC%n$9e$P6ZRo$!ixb5}U@OhlkSFe3;
zJ(kmXZb4eNTjt3rg1aXE;Q9MJJ8NsfoweJ=^xQaO3fI2$-K}lX7?!c*l;daCEP2o8
zYr5}??kd%os;N{cETj5fYx}xKfu5402Ns{p%+=B92vTEs>->1CsciI~y1(CeT@uz_
z;$zR6a)(9LOey=t3F}kSJf=KYA1a{gai(&yd78>KK30B}X&NTwHWTLE=g4(-Y`Asv
zR9kd$g|kHSstG1F1-?2}_q$J6e_9r_+vG<;ZePz<8My`>ALcs|o7d$$UbuPp)QnUC
zt#=yo(a+|;&{?x1I7Z7c$Wf)-RrvBo#t_Z-nja=;UlB@LBYM3{r2NT~qz#F&kK6K3
z7;|~Du$14})|bc-D&fuew!(CJ(e7_@Me9n=7reN*_Qa$~EM9^<*SVb@^Ze9Yxaf<(
zmKhfxZkYV*lGQ1DnI6T6j<vUnUfy9Z{2dh}z1ig3tpldzYE{<{zDu3=oTFRq{6)vd
zm5qBBd}J}1W^BHxyF&hKwKY@N)@5BUXIwuqeFN9V=%fi|bw_6iKK<CwmN~&OH#9r-
zF_Y-p?5DaFQ7p^3j=wgVmBC|qxo-Q<NxE+*bQ{fcQA-k!EA0?kaOnK=&5ys;dcB{q
z>g*X6?TyzH7Ost$w8EXATW?y(To%98UacqNn9Lfjrj&5aTyo~-v56sC=83)Kg&Z<H
zl3zPCZxq<vkiOut^p9`Sy*l3qYL)kFUZ;1LHyxd~U2n5#S?!Fk=8W2}+J3(Cs{CZ>
zZ`XM88`o*Glxh9nZpZI8FTSI5Tym>}a)FIN>0E<VhR@clcrg87#)ji!r#1v#by7K=
zUeg(z#koIwo4NS{0sFrv8huxM-sJw}UGDy^kAx?#6WZu=`L4y1%ybRk-+MD>J!!E@
zeG`9A%tt3bvcLU*pFaQ6=1ER}Y%kh$shqBy^ZoVvz862A{;YSCdj6!I@k8{sGXAx1
zKTHyJy0A25>v4v|wXZ6rZ+)zgUM&B;|3#w8>wg03ChU`@h8q23_i9r5%1~UHo1_TL
z^;?e_?Fw5}S;4~)7|amC&&qEe1|k?ZnHU+OgBZgY8H!TVibWY27}zKB@z=7=-p0ri
zA*?6t2h+&)Yr(x{Z_{8Vo(KU00c)5rkY_940pmHgzhU_KrQ{Y9ON5Y~P$*281;r$<
zM;*szotMc2S|A`Gpa4@8;wPm3VxkQzNSITAA10h~?%`B_*N9f2FtU*_X+|!vXL=9)
zUa;n&wZfqgyO;6u*N+2TDa0bA1XBg{3X3qD_g5fn)7|_d&zONWDGI0{6bYyaXd>~n
z1@vG%VCdu|0)vsExG1^U9mEI+5fBgxV=yp;z?sYuiA{NrIXlz%SR+G1!@?ON;Zh7?
z2xf9_N=^ol>zn4!%Er#Y$;Hjf$1fx(;4SR#sv*qd?II#6DkdduWo>I`@8IZUAg*hp
vYN4fwO;kcwPF_h_MNM5(TSw2($k^1}Qo&3{-^5K^!qX>;#UsGki`fqV)l-$M

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.hpp.35C1B364BF55F9FD.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_rank4_reduce4.hpp.35C1B364BF55F9FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7d193ff70367f53bcd9be7fc3c676f2e397456d1
GIT binary patch
literal 600
zcmWIYbaM+}VqkDi@vO*AElFfyU|<jg;^LB`Obtc`23}?chKf0RgARHz8w#9V{$;_X
zrqUPNmIqV`e2lqJ?J@ag-i;hJpG~V>|CHZemz#UEdw~I`%H+!TALEXEc$KRZye+6K
z!|3B5M(gwsou}@+xSsm#=i@CqWan&K>i2SIQr7jWM-#WLJ(^VT*y)s&gwK0r!#~O|
zCSJIvwVOR;rF+QJEjRS9Eu5$6c3W!B%Evthn}nB5GfMV2(xtTNTwl`^+X++Jw+D4y
zs+__Q5ZD}`np;{V>r&+yW|gtuQ?GHMTF@o!{#9;#FAlOkj_yv#+OeduG-qLYQ}NUC
z(4PIOU-FL5GLP@%51-D|7bf#!e|gBt^1$4Gu@}$H_BVgf9?>s#{*!I6mmzcb2D7v|
zoi&GS=7}6|k-T+8-Szs>zuOXSzuv#>#fiv|jNv9StBNaglN5pBw)L3NuCP^=6+8^g
ztUxXhaPYJ8n=^1SF*0zmFmf<56s4vW&jpHe{aSFZ+1oUjiHAeLK)@QVkwd^%zyrqf
zdem`j)_IvcMxY`A0R@<Fh@X)9i-|U@AYo1cewZ+fX5<80+I#5tf;A7V75)l@ZMvI(
z<QX$iwW5FuOcju)CZLJL(-zQ!@qn((Nd&pRxG1?;AH-k-5fH!*V=ypqBA6UHG1>n_
NcbCj!=45AM000ZO>23f3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_f32_f32_instance_type.hpp.7F3C1693216AB251.idx b/.cache/clangd/index/device_softmax_f32_f32_instance_type.hpp.7F3C1693216AB251.idx
new file mode 100755
index 0000000000000000000000000000000000000000..60f68634bb30cfbc85b2372b166ef5eab6ff3b2c
GIT binary patch
literal 1584
zcmYk7c}Nsd9DtwOxa)39Yje-Qn>MzInA}clY%QCTq|&mchztrXjiM~twQ?&Vu@Eb*
zu#1wsqJvVTTT9b0k`A-7i>S0a`lq6cRBALve*^n=KE7|4+20#>^cit+!>kC!rq0Mr
zPs>Ol5JCfuS0KZ`k_$no&4N&7(S^Cs*IXavynEt%c5OpM&84ufRf%o3vOP8Ft&s(<
z>677q>-?V_`+uIF<v;fOcT>zw_Q;K(w_@a%WBSF1b=48)ZcHe4NBD9MRy>_pc;MNp
ztl<fsw5j)KCUloC>6^}GR$1S*<QM*YQhYgf$Ek#hm{Z?8*J>S;;#Rdbb>^v$llDC>
zCAJ=3wC(Wv(r1#pWnlS(C0CAmb4tgDf3A&vl>Gf;{iZiHmWIgKt4{`ZPARWmcXFz2
zW549*ZOO~u@96rlBI9XUVoMNj{?XKWf6VO0S9dmU7?j(5&lk6%@<T_mnzB3e&acsv
zD`#fabJ;Z6C@RYq<TZ^x{>Ya*J@U}oD0}es<lL>L2Wr~pYzc{xsOcyWOkWyf%*nQQ
zfGf*-<YIqIXg_3kSzSYW7G$>=!$T;wxqW})w1wOITL?pYtR8RAVrqZd^1u>f<mR`8
zIC?&vf4!sROnbE3@u%cvPW-reywgr<WV~4!q?3#Lh|(zFD4hb1)@a~pod%B47~mM4
z0glyJ;8>jnj?*~cIGqEI*LdJ~od-_P1mFZ+08Z3I;6z;n4r>@VtYhFLO#)8RCE#RD
zjyE^JX_s|*u}S8;G4Jj@G3V_syu9<>-n*$zW5NW<fC&>6Aex{7F$4pMC0IZl!2#k4
z9*{r?fJ8zBgb55tA|yaEAv5OnpZ=T^OlmJGusH3m5w5X)7wj45@tILq_=2-9);7K$
zV*Dyhy89whA-#Z9($_~6ra(z46euZ`0ws-UP|^wwN?N5s$zTSQjKYADQ5jIOm<1)P
zu%KjB7L*+3K*=c_C^?k_C69Sf@(K@1Ugbe4U;&hZB7jm*1yG7u1f{5mpcGXRlo(@B
zVg-W|s~D6LmOv>f5-25A0;P;)x4Fs2{jJDeK&sq(1NhTcr*!`r1^mkb(cLGu?&bgC
eFi()h9-a90{;#O{w+M&T7KXxl4Kkh&=J5{&e#S@u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_impl.hpp.B9E8182EF8FC3442.idx b/.cache/clangd/index/device_softmax_impl.hpp.B9E8182EF8FC3442.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1b723e0dd8e430a81919aba720ad4efe02b0e999
GIT binary patch
literal 8366
zcmcIpcUV-%7T+^>MOd!O3X3el(k{KNEJa1yQU$TCB2Tel!vYd(G%?18iW*d!B^tX?
zDf)~NvBee)>JtTf%Tq%vB#J$j=$nN-&V6~`r+?&oet75pIOm?3IdkfuPmGV>>?D!I
z3`?AnHGE>aR3eeIrvE2Q96zdzNF=MwB$6qs4y0V|W*4A%_U$h-59+Q<?+(vdvUcnG
z`L(^J&(3RaIFBj2xO>`7{kbo<U%dReK4tx=jU$KFtXOfo<L%XP87(!pkA;s}vTMfK
z{@u22s7Qa^@X5qR?Wm>ZRiB+1XYRMZLFaLO+90XY>AFy{<4o&_-`%bs``M%a=Em4-
zcjul6TNx3wIs4$AndbtRKX!?Bw7aYKEY`fNdFC7u{Ji1nDM{Bz^KA9<s`NMIkM_Kp
zeSFIDp<W$kw^_pn@{PAm)|}jw9yz1q@cXkDci+7(t8o;!>{Y*uhilGWF3wZu?G7#(
z`uo$ZL)E*stQjtuWLNiKZ}_OPuATbTjqvD~`CUu9abNAV>zTIX;H|XZ|D3Vz#GSiG
zN~iS=UwI^E#h~l^W<~tV-0uwE-|Kd*t$*ss5mQ22oct-ca>3X7?I*r4vE(yio;+?H
zrE%qgmRWeWoc-(Md*An|(p^yuTDs+ypUpo*Cx&0Evu}In!LR#nX@62>1?w`d=ghV*
znDW`tNuKJQu_o$SBkH~X=yh~!>5Ida{jX2S^&Bd_(x5AJpD$b8bJl}HMctRyEP5RJ
zO<{fFt*Cy(X9=qNd-4<W<I@X%(KOC|yFGWyrm0;s-O@}CSk!u`ZYNvM2<*~p^h@r-
z{nq@t`~?{&{YGp%bE38rxW*<bmUnTVCR<+bTpM~zZF4RBcCg3U_<I#AF6`^}!`$4p
z19xU7NMmj`CcWO4w|7;i(dl1iPQJc1T=Do5+tGcjJRHm_Wo|D{4M<5&y83S8z?>D<
zj*nXWTz2}I^;e<Ai>xcO;bfCK#CBl7<X;Lc+gv+4W!6`E`R8M<%*l3Nc5@zAb<|vO
z(Co*^jXP52?tI$ez@9g<@%7p+M|Rr3@;c^b)++biQQc1sm5H%uT=y?^sj_nY;_RZv
z%93O8C-zO+_N;8LV*RR}Kfv`vOME30re+P<Kus7jUZ+_O@>*<7EbSzMzuMou=^-(t
zmQDXDbl$ha1HI-ch=P*manAF&K;wWcmfCTllW+EjuqIZN#YgSq+w?%PjniJWU4_*!
zb7Dft6J-<IG(C{4>JoReSAs6Zl2}l34yVZB?2H4l*al2e^>l0qvm~u4OOPT+XnG)7
zttC(LvYw=xla`b`SvJ|J>49X&tEKAJlkFPi#M@pX)PcGVJd6Xf2;&xbZg4z1){?ZM
z5~_)!n%El$WC_V@Y4&t*@)$YsrBbef>MD2}2V{v+{W#8jM@G7w_)(S{ptu2^#sOIp
z{@mnyu5q)(oS4#y=74MtxEKdy>0Yxkv!utA5tc+oS&E6Gm<YxJS$f_1^OO9tyq~R!
zlCt=z{oI=#NFK1!GTFzc;gFoD=sGTd@&ecz2V@yJ{%p?k5l<bgNjoY*(^li&^gwd<
zqS4zGz5ghZ6Mrh<Ch#}G**GA}_<}tP>zDp(m<f@#l?XyhLDlq-m=T~>0|I;+Fh5y#
zWX>K1k<%#}iAy61HV!CbdfL7)UHJ4_a^gv+xdj%tz{WTr%PbGoR-f-5WGR}oag3N9
zqbJimAj|yiWw%>g^Jy?4W^|%@Fs}zo<A5w*r9L`7X=Jv8oOscRYC&B~O`v%|mUR!L
zA?E!b=PHP`S|Z#5`5kC&9FS$J=hDOZt%iD86LUI|R;~4IdLX&Z{j)U{=3Pfw6P}U-
z)B(Ot4<ujw{n|s{87K736D%h3#iX^+JRr;EAq#x<Lu%g2i4R@oC1`tzs=awYmih#b
z@B5sqm?kH-l;z*R{~J`s0a+T94O1r`(8v{}1zpD?PO*qn8wX^0cxgnJih`<lZ3suD
z6oP3XD2xNLG@k6EeY3_}Cnq+P<vb|QLtEp3EKi;uimtXAvqnzZP?l?8dkuVz1G2oh
z7aHFww8IKH(NLDlV0{_X#sOJgzm&Fo>-cVi3E}8+^T9kHtc(M)NX8EzF=3F3#VU|B
z#i;i-xG&f#EUx^3low)rV-g&$CLY~2=)y~0NVG}RV({RFzBYYDXtc-ZXr#4jH-F{m
zqJXQWyb!7jO~7CCLYOX0gu3+V(wEUvl2i0-s?@~P?hFly8WPVqheiz*p@-n}5CYKN
zEL7*B+tu9|15zO+CM6a{VA!QQZH~FPSH;|+TYh+T#ew66c+dDq<mZK+nx63t>3#Jv
zh%i34k~PI;l*9}E-u@ji{8JR~pTdyFThkGv1QnjtvoDI{1%0f(7egt5DbWn2_DCJX
zP!-r#fhU?9FH}QFHN-KzlGv^!o*43YVJ&f4O9B|KAm$aMEn>_hEo4m@gl%MM*woIL
z1$bc|M9hOE47I%Q40O++D=M59UV-iv!<c`lrkj%wid9|wr|Fw-E3`x^<P)2GqCyXr
z3RQ&uxMFO=gO<vg(h`2p3lm)?2IHD};U0wFgC1-d4G_@)JrU#biL5DkF(J=E_Z(7@
z%^_{#r@wVvu!t^d8)&ydAnH;oR6ukE#2|K*&+M3Mdg><Og)LyWMSKXm!ErbEzGpP1
zN6L5*_=C^}2{L1nL<e<=+*SFw!x`B}^?})xYP30(;%5B24@=ONe7ev?N)NjITG9~@
zRw`^H(HluD!==Qtlz20|gM{uNVGLIgeFf>w@P5*MKa;2J{L2Cp+rA-EgWz@>nMFy3
z9mI17@nRoVkjM%W#qfUOx}Ufq4hlaeyEL_*5_p=IOmBf4Qx=v3RBLNMz6P9_)|zVr
z)LI9(A(C{9L6fM~GP-0Wzqb`(WaP*wL>;cIYLWH8#*t2Q3>3$}0i}gHJUTJPB>j<G
zDlFyrrJNjtJGDtpna{NmW1oD^PmCk;IUU0V9A6;*zJT*75c3ys2?bmt<1gg+LNR|K
z=Tj)=FXR#m#r#DaUnJ%)=J;YUUdHju{u_5##)aa1QQ6_#@~Yb0|Iq*zRW#lm2gPx*
z9an>*nqir-)I!C_@6UOV`P+y!RN_vs-U)W7(}cQJoxbiDH=z|T#3|z(Sg?v$#*0vb
zGC_or0+Yh=w0I#YC20^t$$`ltTGzm?B9x{~b6}hUlmkR)pmLxHW%8MA80T1itO#ZE
z+3yj?H5w?I>qqo1UU&q$N6;5TMdx!RZL|KUOQV@(H?-XiYE%FkH}%iv@)9NVGK)Kq
zHC&NLRknyTTf|x5>0rT0=QLEC$WK#HOnfpbuy;30kGo65T2Q?PX@Y{;UV=5jB2)<W
zh2Vl~;e{fwFB0QourC(lT4+@ZR#?IDG$^&I11l`*X-XlU)x;lT4lkS_o+rfkB=I~c
z#*IYRND{C*rAi~ZM`HY#=pHlNM_w5D{L7A~sY(PFL4!N*lU~%fblufZ`s<aP)k@}v
z|FAk9Ft}XhMhOx25gkwo|9HJT;MaY#gZ`4`u~Mf|WCQRUpf##RW~|rIm<LpJ%<S*(
zD$oiiAJqBag#@0KkgCPR6>}2}AcX#CFgNkE<W@EkC)_Wlbb6A8svq~m3p!h!8?JVg
za75|UO8zZf)*;{zfeM$!3x|O}%y8zBZ@s=={VeT+!3~=UD(w{Tr@$Gd$&4)p?gbaa
z$(mY;MAyj+m+p<3ZpfV8_n;HnH7_)PPXmM@p0KTUfByWzU$o?fP2ju<+)%y&hi?KM
z!<)f-Gx#%H2A#_wnc;GXEN3Z{7q)|LySQB03EG_yg8Zmc>POA(4Q^n{UjK^Cs`c;a
zjt_wC0np+Xh7@aikYUUwl#yE4X;31g%-9B@YVbJeRL#u~PKC`6)md}rg9M&dGafmd
zKbAMVkjpvcavoS%E==O){J!{Iht|B{W9?&)H6kzgS^L?eztf`7+TWfPg#p$9_Soj|
zg4SAVk3A9f`H&z{+zSo~7NJ6*-Gw)9&fxR(M}&((C>G;dP}YK-*yojXV8?v^2AJLe
z3rr1$7@(;pA>tUIIU&Z4q;n(bj^}0Y`Oc3RUU)zD!pITh-~8=rVbSBt!w+K|>;Lcd
z4Qhslg9E@HU^Sbeeh54U;Y`)l;8P93h-JpE7wu|k?cK(i`#vxKpaU1$e3LO<GqD@p
zRtb2Ou!AySeF^kJ=~AH-VoRCBNQE-!Q3gF3y&MwCp)<pKAaV~xAuhV<eJJYS#``pQ
z=5eY#&K2`d(Tl+YW?9&ur~ADFes`cF`n~~oxx>n1URXj@OT^)D2?<$3qA@%cS9kul
zZtiqn>Ts`!@)dDpHj&FIb2&$Jk5wk?B4&+<uBD8Vg~^?;I_GK5b(;g>n1olY_PI6f
z$fWjviR-IhZqK=7vrtXH&a}zYU|i*eESs$N2=_pJy7{^ReR+LWh*{U^=>{_@fleh5
zhWat+*R2%X87>3AG8WLJLOHmWvj8Zi0nBZ?n7<MND<KHKUq5}};#2bWaUVSWjWJi%
z(MMi={ZW-+D4=bS;ckQlqTwFHLzoIqL8nuYgiA2w1pO&UN4zz)HbPc0#*dmnQbba3
z{H%1!z7{>Q*2kOiLV{z02bLVXkQ|XLYOT9QbQPh2jsrbV051&mA1GSISevnqj5gMN
ztUp89HrXOtwtMz_T3KLdj~`bLUPg6M5BBxofk#H0PEI+Gna}qPpM7t$;PE!9*$?-Q
zyF+#^|NQK4ov8ukaaMV(kf?0%`@E}tB(0i-czt{yw&R|IdJbhsKS=)`?UUr{>(;+E
z825+M&i<dy_bu!6r2eC;`(ga;2mRlC)`PC@Em*$=8?@O&^3d<Dwc9j|9@9~1eUv4x
zBOcQO+V4u*^g)w48v9@v|LkxMClz*qdKYVV3|PBM+|*V=w@T=a)^T)8kFl-dP3L^n
z7LV7SJ$dr}(?fsp%j5S}&RBgn<gColKJoeDb|jy(%jeV>a13;9J{QDr0cTyn*@$sq
z0T;}0A!l7ErWbO7g<^UU=Ul`IjK7!@ia9rimvOGkICsP~ZAR+9TOB!!2D~x|Dia6v
za_}i<E^Me#e7Caz&kK7%y9YX<#%jXcr|G-f7k~7q&;EKOdvD~qJ5*ydU(kCRlu&2-
zr)*Nz7#-F5BhWsAD6|1ycn;d<3}azH{oio;$^1VfG$WCD*J(&T4QVKy7iz$!hJ~o}
zn%}N|mom73ZakC^ZOc9j<HJO#w}0<2<^-wAR2M`S>*x2`T;~OJxfzxjo=m`LP@iU&
zg$aQM-{w15^yEu97gZ*c+Ftm-=kQBUSLsJCUoh`4CW?2!bp5MAz8b7iQ!?XCFsAOC
z9cp4;l&avPC;N+PlkbvWE&Qp(8ALOK1hNC1Ni;JVu5Z`kX-f;6g_N<8m^BhBmK^7C
zX7e~JOpbU7O=HtZ;7_uCz;K#gHOv}i#;a*u^Bv#HvOsmodKxy~g6b{!<5JM>>2Wk)
zSflme?kS=9ni@sn{WM)H6}KKNkyB>8rAEfbt15rc>~MDdsGgoE?)^49w|mgvE(7si
zoGxuG$k&P+p5{w(T-x(5s;0Mjec0k}&xqITbjEK${tc@KWyZ^Pobko$?WH#Nv%aDy
z7v~rkh)%-`X^v@u%vbt5_J5E52po5`b9q(N|Kxy|x=r_;7Iw>|u37@NCE}jF1folz
z3!Vs1%NE~KW)-|p389s&&%(u-%8d8^c=-HZ&fHM#sqTOiS`BT6iBCQWL}%vdeTH{2
z(c%RMPj6}4SCL5Up6KNPv8W?9tWkJHoL&)6v~&7*hRpO>`sb_IQuABNhhL@wO2%sd
z2`&vAatSUTO9Kf8k_qF7PWaUD7gRyH_=;3g>vGx5-NVhv*~QhtQQ~SKH2-tK7m37;
Wy*jd2C-&;WULDx0J9~9zum1r>;|@gt

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_softmax_instance.hpp.E665F0AD026D2FE2.idx b/.cache/clangd/index/device_softmax_instance.hpp.E665F0AD026D2FE2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b22cfc111c7c7d1d183bb3e821576c2920fc94cf
GIT binary patch
literal 444
zcmWIYbaUIn$iU#7;#rZKT9U}Zz`!5~#Kk2=nJ0ksLrw;UiaA%W*zz4#;9-3*XGK`u
zm4(wC`7C(3HVRJIva@#SikBgVhaWILc{q9A*I6g6su&LFXdjfhn6SmVqB1bx#EZTR
z<&4O^E;6z!TbU*^ryp;4$Z+~rZkvPt{`D?0su~|f3?dU#E!9$dKYWpPnR%%z-(X9$
z@2lS-!PWN9Z-p{ez6@2^oSBus`MWRgrFBx7@q0>EsQq0y>r4BqOY&WlzTLGH-4fy#
zb#3mmC8>8`+GozxojQ5i{5lXNb~Whwxy)Y-XPeH-7gy#c0lmXel$uryp>q;Jba7E~
z@n#T%2}D2uf5$Pm)1r5k>iHO%m|0la*f}`4xOsT_7?@!y7#Kj|1O==x1_J{dg2|3x
f@*tSJ2qqtb$$?;UBbdODhXF1agMk4US1<qo*hqD4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_sparse_embeddings_forward_layernorm.hpp.469278C9F8399B0C.idx b/.cache/clangd/index/device_sparse_embeddings_forward_layernorm.hpp.469278C9F8399B0C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..90a8972502ae756c5548e91baa0ebe228ab9ccb8
GIT binary patch
literal 3838
zcmai0c~}%j7Vm0C7>1^qfdR&5xCfCNLC%qtLoP)@6o{j!gd@y~7qH?$P`p;+6*N%<
zVI?SP5JX%Dg!Mo`jCYi%tmIQM@jk!<aiheoX?!-{#~+#Rn^*Pv_g>YjcU8qkhK0SQ
z2_hsVa{0p4>?9sR5F-4~%Fa}-BMIUwU4mF%b|P*;bho>>Kl-uGTm1*>77I~urPINp
zu0ZkggZIZ2w^z@5wdPRBhXFt3+-W7=O7qz3vVmXo4&`5-y6H0gua0lJ(~D-dJc@Ph
zZ5w>u^XRVf+Pb<F?TL!J$Fv^I<ooa0(r`CYXWHpAUCQ-|dzHKsy%lD?xBAwnD~qMP
zh{iONMdKw;c5aKibyk*Wr|Xmu`D^<5xMs!S-Lp4Di@cNXc@ErIr8Mg3t=uG6-##ZX
z@LP0xW@1BMN`$cOY}=Fft;8k&KfX&E%!w#}(m!#t;YpRKy}&uyqw&haM;YmNdeUz;
z&2{#uHSK+WU2mY~N$7&&+^u&c*}l0~clk=Y4U_Aw<^R}xLmuj;we^GCtGsWVIC8}n
z*Qgg4Hdl_Ga71iU%IGDwnx@|PCFN$eR+yu0jNZBzyh}f9ThtR(SkaU;%b}})7k4>v
z&Ber^*0w#0RY3{Mdere9K{pQdI3)(XUUG0J6&zRJyw6gwPuZI`e@0t&*VyLQhTGp~
zTh)3@^4ez?8UNfl@=xMpIgz(V`L@ZYRk624=gY{7U3UT}3)3$g+r|rUr*%7y#gI#q
zlwmG;?TXsdLPv)?U!*OqUK_u|)vqYfVO8>&8819r9y`udgda~XYtp}cXj}T8H{HQc
z^E($Owf<PFtlzJ#{$*BxJm^_cL`gzMsNnAD{=sD#5BU#o-5ldJ*X74`36DC~KA1XX
z_Efi3%f3F>wrs_kfA<p~uNdF&X`4A1SJ(E`Wnam?B3bKy64oy6Z%MB*_a1e*^UY5V
zXIzf$dC4=YlD@r>J!kQl&b~I3+|$pu%ztShZb%zzZ0%hp<SCC(;b|)UBVDgbi^CVF
zW|&LkB2?i&yY0TXqK?Q?E&TEh&PLtI`}$Vr#})}mZ6U$bAz>Xd|1=PMoVEBj`ouBI
z!-KN(dZZ2(Nfo48vjku>$8NWx;JP?%QWwKHdO3zH0ocNQUfZ{YiF`v+fZ+-xtU$J(
z1`v^TUEH)_RG+ansf9%x^&E{@0<gQ^7eVm_q6ai-BO;hk^H4jM06e|=yB%+?ojoQb
zX+wf3L39ZceHuU{Ui<Nu$G!_Zg{%!#NVf_}nV|th6xAy`b;ELQ4y~f1RU$@10RHOm
z1^ww;PTi+TBW%NDtI3`$0r>B6%OaCR|LvejOAJRDM2%(%z?s!C<y~a^93e^J16Cu4
zY9#+OfXI3;=0?f-nrVilJ{GYP+R0b~aG~3H7sbK3nc5`JfM9HOY+YDF=#mIW2r=mU
zS$ZuM#~BBjlwk{kMZu;l0sY+}wDXdrjwqV6!f=>0%$6kpABsA8>d%+IInm@O42O%u
ztyu!_k^gPF*1msbCQb4&93hUdVhO+vE$2H^F6TO#kXjga<~s{n0`OH{LG!6O<+p{T
z9zI$OFUFK506#4%&HZNdl3bd!#oCixCV8_2;DLsBf$#EWe@&Ct7!Gj?@n8wSAAP!(
zzpp+%NRwg=M~Wk*ECHCvOijzm#fc1gjs%~w8$4>1_q4I8GA0I5OujJRAM!wQv!4H9
zf1?NDF=fcK3`ro(pc+U4TLrKT)|ThTf;Ka!_Q3a!zpvJ&7#A}aD{v6SxSF{R<3gwq
z5uA@=CQ*|_;0%hHZZq8;JV!CHHnE>MC^&3^rrJ0QYLHb8vWH}Y!ow#XsyYy~nRvA6
zmG94j5l4eM)Df`t2l5mfO#mrIVWk+$S*DVxtl-!b_LWl^@&sQ|cxT57<N>Bo%zos(
zAN>vL(Iw@B+?>pB#g<eeX(h4%K`42WVti>|6Y!saEqt(Kd_cRPOQjIw$VKs^XsAJd
zi<jpYXeYhY;W0-MbrjJMdJP|X6oqrV8kwn)`EY)`8hLYk1Clf#W;pNKfZRF05iyNO
z3j72*pnYDLOoi>SH?a4BiVp8AlkZZ5W@$cq+6TaA@DUErP|t#}&d?1#u`pG*1<ThV
z-8y6d>x0*@#!&Bq0L3S5pH=fiw?EcsMn=uZ1T-KnSVPmyV28v;zl=F!WVSqtaWr>y
z0Dn?Uv^d&|i);y!lpq<{N#RXxtB@O5k-BA+U4da%Fy7SG$2J6PfkufV7eEJCbCJ{Z
zOxOuRg@q<k6NiOY(>Xy1@)#RG)G=WT@Sv7}ZISSqVlw<Qra^g9OrA8)9UK5Mcr8uu
zjFc_gxV1TC#m326YoCAn(U(s#c20JmBZj$fyJ@fu#f&o>X9EYJ7!Na#VO%&BZUo(v
zVj`#rBWRWs6DN(c<=h!BjsMI+x5f_eqP@hhxPa~1;DGHtjxB*F_ddZMu1Au3B!jSS
zADExG+CX|8$4}*=@`FQSZi*^_<4TdFba?zqk#{KyhIsBc7qP$dpDUcP2@0{olC!f0
ziEEH0cpokfEapf18FQ9u?hvqaZ%DU!aO)qt{$j0KPyd)R<@1ZNSdKKu9@3G*@s(C{
zjSJj~#r$c1Gj2=Gy#!l=m$8tXmNT4?=5_)h_0wFn>4sO;BRXA!_mQ1>SD!^-pdK03
za}@*Y;6%~fW3bTipRUiUmA}3}Vxd!+zrC3AAvXpCg~+%NnS%S<4#X_d596=>3zwXI
zTv|~6$oneWtbC6me~7WBSw$U1(>Pv@Y}8y1@EA3ct5E=lHz3OfWIdesYCt|5--xUm
zkuC5Sy!dZ?$F1Vy<VxeGxkK4pnK0(-*l6jF5eJ3*z>d+};80LKzE?s9Y9wfc{O^(~
z=b{G&uK4^LklP02367zdjmT{y$KQ7!^+w0QbR!mw){QoTeKbiSszzpzhYy{bYCX*l
zR$_Y#k)#m0!g_FHVtpUl#}oo6U>^|>zgJ~Vr`!%(%SM<9w=dTAroD%yx8|V%+V4D<
zwnZH32pkb)czj^(3G{?veN6KX0qxAxjHERDNSjoncTe2AG66vaPRX-8x$08TQUom(
z`Z58k&|?X3cbX)%G%a*sJ%TVoNfA&?S((XMIht3ysfG;wBrqWe6K*=1n=*s5iVwVa
fwxM3v+``yowA55$Mp$sLF*h~irsmvK%1!?d1LjZo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_splitk_contraction_multiple_d.hpp.B9F979311F947532.idx b/.cache/clangd/index/device_splitk_contraction_multiple_d.hpp.B9F979311F947532.idx
new file mode 100755
index 0000000000000000000000000000000000000000..498ffaad97a47f8dc6a00f1f650833c6e6414887
GIT binary patch
literal 1342
zcmWIYbaOLfWngel@vO*AElFfyU|<jg;^LB`%-@U*4C{b$6?67Z&dt7UAaMNeuK7;-
zRuNs&rjg%Ob~VY)Ht5-V@K+ws;uG6lj!bJ&S+1LUzjo8y#N*!L#}X1Iy!*bN{jTh{
zymRx@TDqC!in=%L>Z;iK=i%bd%YSw()2%gK6zRR*-6Zp6c9Og(M-(HMQr644o1-53
z>bI~Tx*gYXrt#Cd@JGxK|1R)|abi7mub6qhNJiF9k+qlKWCeVzTp#&<YsaI-b=&Wq
z&4{p*I-AL}$<U+b%kKm>9h*bH3V9N0T4ZGp8T{n@JM&onUXHRk2gO5+X6r23Fl&>W
z$LmW)p-xYhmT))5UH5&HezKMI?xy}E_3aDl7pAe5EA(86D_`JfR@8Q9LrAWm<G}}V
zeD5w!?O!yrf7#`gi#KhP?VG#NL0;hO(ke-x3#?J5+Y265ys^+;uu=S<!NDgw2FLd9
zSSei-Yn6DD)6Q$|L@VXF?;RgGJaV|yn6IZ|!*g=ZL`&u0&K@O;PHDIc3;j>+IKHjr
zM%lJ?kN+}pYTjZMDsr;hnm+x3wSHqlkNGC1J%(Erw7uUj>HC6Wt&E=;I}ZAUc}eZk
zQY^nM9+0FL5h>nqQ>_2=<E^U-zc>AzxPC?BOShH(gU_1$%`3iUJ}3Nl-{(gUB!eQ?
zm{**=v2^MtlLY<r;x$XR>@eG<|2s);7r%9{ce3e)SMot`8X~_R{=?$y_h30gab<2&
z9WX60^)GFi_QFw4l!=X#fm>ZxT^&p?Z~*0iz_{j{?=s;Rw*{D3IUyq4)|%GVFzI~^
zTuQyg-`xb5SYXmxEL!4l>8X)EtIn77urV=nFmS7~tEz(u22LhMb_NDUZpNb2v|>Xh
zgezFMz_#y$>f_<oG0?GsS;E6@WN+k&;2GN+!+9`wGIE052QwhVPe}d6L>pEP9&SBl
zJxzpOePw+lo`JFf63<ZC5Q%4`XJmrVXRK$8#4}MgK~ia|YzpVaZnw;dQBzT7<l)wo
z(bRw$klM$p+8|Wy1Qb>^RJDaUv*Pf>CsiyL&oTo&EUTjgQw{X6vJqS_54W+hF`U=!
z#oY3?|Gxzn54XCgx(qBDf&SFeGej7muceQ~Gte?X;u&fgBJqq&jcgG5j7^P^cqUpV
zNGeUWOp$n&5|;8XeQS-a0*=m2m=5%wwT!hoEK*?7z{mlm4w$cDG$R){Rd}kDbx+*o
zT=&#~^$R<{csCO;=ma&TVVZ$P3R+9UjASTE%}G=Orbt+bGQdoO(F`tg+t1fNKH61W
klw7P2RLR6GSasFBE-HN`GY>Bp0~aJYgGgQklLx^B0AnHM2LJ#7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp.43569A70DF974D2E.idx b/.cache/clangd/index/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp.43569A70DF974D2E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..db05bd26a7a51dcc37a2f702ad48d9a161709eff
GIT binary patch
literal 22282
zcmeHvcT`nJ*EeU*nRCa>@p9?CC{4Q3K_dvLh#jv5qp>HNM3Yz&6QjWbqGFdO*bDYp
zu!DjXL=<Zz2r8oVDpnA`A@`0m?_cj9&w8KdTkCyYE6={aXZFmQGqY#6*@wsh{rZ8?
zT3Vq~2h5*2ZSEvNOG`@^|2Jpu>>1N!T3V}ZwY27^WsNDGRqU^PFy~^3&iL(4k01S%
zHgNSU&gtQA{9i+kDt^{>H4@zJoSIl-b7#igou4k)uHO_swt3Is&KalXe;pgzxnj1(
z^UF0qe6uj9Iahz@%c+)%wjbYRO}o(k@Im2Ei?@DPxjT4)S)g;_s@@BSgf9!7JeD8S
z#2=F9OtU{~n3FK|*}0XK=k!VjCDpv|WwWwvp}qCgmF0&<tgsk4+u{8jptm7zQIpr~
zq;&=(g4QID8XY+6#KW_5vlrhx^jD3}0Fxp5Wgq+-FB<8M2p;re=cnMEu0wq658gd<
z=&z$Ls}KIRdhv+;<MK}@ZksrF+lGK0^|K?ImMpJaxPPkb*!j&LCZudHc7o3rY%mIO
zSa)}Tp~}=J5{`_F+T|sVG@Yhh8S>M+UmI8CB&|HJci})r&;l;JJ!frO`P~IIU7c1P
zDvQ=}T5~*LfOVo%PU+yIx4I1oT0gx{lIih&mfro!ZdPo`ia%^n2+t1uVuR=T6AtHR
zREL<ZbKSV<pv|nx_QB)K(uXY?^eQuIPIZ}9dgIHR%c|a*wa82#v>Y6Lq4IW_|5I^W
z=FMMsI=$Qya-_NBSVosbeLMHowr@S27mWGJujs+m$QyNa$7<(o_Nkd%)bqQ-KUeIT
z|Kh8pt9M6D&TIW{UeeC@1EM-lu$X!ytMiKiCz@mKp4xQzTR(62W%~~O?EP#)M31)z
ze!Ktn-jhtvw&?~__Vvo^{yHmj-pj!&baq9)>93sfZPD*$`x|Y}ul*R;+*&sy@|!8n
zK0jCGFC5$cV3Fm7hygBt9OLrF?X&svJL|}rV*`E6@87?%F1^S`f93vd<+>#eE6a~n
zv^`SXUHJIQq~m+bx9<Jr)vx_3OCDeA9^xB2c~Z$dtKW0Sr^Q8$+1<9y-+W4t*|D2z
zpO?2@_q%E`_j$KnW!DN$qdj)J_BTX&^)YdKS~=Nb*1fa=n@2D786I3aZA07UJj<Ed
zUH4~J+zQ)#yZNuavohE0OO7A3!#8VCJX$(DEpBmM%XeM{X)`N-IlCzIZpxAVzTe-v
zno{jmIQ350SME6nP1_o_D|_evbaU;PZ7DAYd2U<$Ub<=f&b7avS5U7>gBSJuzUIeK
z;i2<KzA<R>;7~;NW6QM14grsg63(pt5ZHZu?5<^dySdvn#(X_(wU*_xV<Y8PrrtUi
zH%x07vY#{g_rkc}^6&efQw*9r@xG_ufV0bre3HYi+N-{Kwqak_r+St5>&7KrY@YFR
zuDSE}X|r}bayyYS&M&vt&9mRWgzi5?Rz-~4P$_r5H~jCZ`erjjhMm59t$B(5)On$B
zUFFi<_Dh4pmc8}bIOEG*F%ge;pPse*+b`#OSKrwkHFVyWUt8uS+-uSL?fkZ=&}FBF
zk99J96IbYNR*V}b{<_S&dTiaXgp64S9^4&yvi6Ua`y#B~?%C(&RbfBkmjRjcv@&*#
z<K`XSbNKF00WNpeZ&+FhDoiYA8#^W}ysxux&ok@58831g^-pyU9o;+AbL6i(XDM&3
z{o&^k1&L=<r=Gg$@LQh!>noApj}LW<neLZju=)Iq-Gz_W1}8fDZwgG51#a?|U!E|;
zIA+&`K?eErt%vC5o7CGZckFlhb@)4z)soBjL$4P2n7Q}8<hcFFIUnB%6Xh#vb^n~N
zGbPw+p>cCu(4dXk>AO<>mrQ#4a)r-Vvin1$J-beC44=DT*SW8~mxT_QYBM2l^P}Or
zLqiTX2fx2DYSgfm&3A{73a~7AVl&occEczmqt&5LPWHNerq1N_H`iwVb><7dS-Y;!
z?;H8tH}%LyyCF_ZHwxD@tsL`Pzdqfy#|<yjE_IdPd?-0S{&APl7}sz6hHOuNeRO!&
zADhPBJ>fsDy>)SqiO;RN_FH}N;rbk{5$B$_yXgOIGId3fpIcQ?;#S#~#DHk$pfA3F
z-d~LndmcU0<B#RBdX7IF(tqVP7tKFDRO=V5+r@f^ei+*<I{tacb!oSkiw^&>+{VPS
zPrk+F$<nHyl6YIK=CeMNv;UYl`JjXTm_ZYJ6ixPg^sH~Ot=~@y?;hO`F4|=HB1SrU
zyW(K%nu9AtGHxt>FnUJn(XltZt-Y@HpE74u-{JLB99oligpbX&a-BKz^4Nwcxj(k9
zjSY;}`to(_Rhe6J=kC1^d|3GBbf2DAW%G6?l@7lDeqh7-EX#`-ex*9iGgB)JJvVjA
zjJAvSJ|q_T1ZyvJI-I*#d*=fG=^+)>J2Eepf3MhI^uy`me1E&TX|Z3Vy<b~r*B%nN
z%+{jUH;W$L3f22&z8EF-SQ%dU#@27lg57mTvORZA@9b26uflg#?})qp0plI=2L5pS
zO1JEq-0>II@moWzwi&Op4KzsaKQ|`PE_&61sI}7;W*Fyfp0MG}L06CUhK)_f`wh!o
zCH&go{~DN%EJE7P-_y<fFiEf0kZTbZUMd)~&HqwPh?MYtnO^O${qumkcB9v@gy-*j
zmc~@9KK?E@z@w~4G*+DW_*d!i>0`GWghaS?diaC)km)_gpE5quy5^F{FD5<>x6ajh
zubHu6%bo{cU7R=b-CtsgPDyur|F#uwo)`FrJqxC0?I_>frG09&)U<qY+kl(BLfdNl
z$D$KWCaqhWf^Kmo_qFEyI&*S?TuX~nCH|1@el1)9`i5w!mJ6oK1y>ExLf9(fe<D{H
zpW!}stsyWC)>2*J%&u@28uFRtyF;7LaxFs@U_c)&RkUalEjnumVL`<sZ?4<<`wJzo
z!lSH0omL?W4IwPrzbyJav#^huHZaC4@q#>FFw&6EEHUz|`RClr41on^xz3qg$8%8=
z!lGEVbVk5__YHc$wu6ONUgw=Pgs|wE)#veOXegd8e(5;DBu;SF5W-^kz-xPpi%*>r
z7~xUMkg^O}Y6xL5w_UvR?Y5<L3eX+D^ap740Xk^NXO_Ra#)d1C@EbuqN{t||5xQsy
zVX>?4Vw(8<iH!=-1CNp*DicHt4IwPuPA#54%~(3603MiSsc5rQ4Ac<9;<C7U$!|5<
zc6z`bKj<}3z6MSjLRj1)l?S9%dZ!g&0v_cxG<^-bYRG4n4Kqd^i`xBA0m3m$ve+$I
z^wAK);t{yFK;F{-7bVccOIOKzRr0|aLRfn2oxJ73cYSXwKo}lno9MVr4Al_A;`_&s
z``x2P6e@rhX4xVtw}{pnLRf;w&NcaN?8jpYFbK1xh{_baE^0zpLJdj}ST31zQUL~H
zmN-$0!|S3ZpIPoCcZ%Ru4OqM}%O`013A$<sVF}x^{Px1<S67OlbC8zm1SdPed25JP
zC!TZdI3ryD<h~1@JotO8-apcNM5a6<vsI}H87F*zyy)6JzDXOH;4R1$<e7rCh7guv
zslrN++mr1L!FbG)$C>4Ejv7K(M)9U|thTMKF$CtA<r-&p4GXxM5SFL~<Lvq@++(N>
z6qw}^FMq_FX$WB%`^A=t=+KIT+Mp9=xzEe*^TrxNSSAcwJ^$z9KR;B0ZkXjU>hc(Q
zX$WEYdfuQ4?FI7A+CUbnrAikS>7tQ_d}gV5zk~0a`bY`%u{hN6UUhu1h7gwNQP!bX
zPZ@hCfFIt1BvF}!H&RVLv%Eeg4JZt))dLQAE?GdC1spYmu*@ji5dOSimO>lIF-x-`
zZx&26gs^=3XJOJD&m|9(z!^WN7<Mj(dKyAlW)|eXU+}W)w@T0jk8%bHXOOXm5SGPD
ze0<G*u$iC${qQJl(6|kD(-6Y4Y*|j<Me8T0>j6hRN(NG9Aae~NEU{5V%C3Xk2kQYF
z%(5N{>yc7J2uoaKUVeUr$1e&Hh?g!_P{v|;QIpRs>&913-ZSZd0(j!NY!a25@Y1Qt
zXO`WsXX-4Td`AH$VV1X0|1Gr85W=$hbVbpg(7xLgAOJroMo`A!2dN2RS!ejF#wh4h
zwH~m>(=FhX1z0fEgs`lScfR`Vz~(O6KnJrt;N=f^6Ad9ONkv!9qI_Kg6u=+PB~er+
z;_0dhVM(2!@H=1OW2y~w@hA^@`9mz2YC>3cm8fEJMwhhefv%V(1}I~&FsKP(Nk?Ol
z<m)p=8;F?Y9xuO#M^O{Pa%jhMlk{KqK9RJ>_102#L)|?8`5&!L0??Y$3!mu2zG(K5
z;*4?;XyY-8dF^7}NkhmOhsXVtg*EC85y&vhYeD<9psyi><w*9R@*l0bZ&rXW@GGwq
zZPtk%8bVl(SqA71xVkQ08z4MNy&$j0Hb6}X%W=J7w+ErLrwR~(NBIbiK4MFvCWPgr
z*|xPS<AF>GTsqF8Opz|r)eypxIiR|!{gsxr0!+lCyn$|Spr3|(W{JQ4%iz{YlN6vQ
zW?3Z~uM(X!gs@z7G}M`%d#+XizQQam(6j}c2{j=s*8>vko;@h)p#bifWsPXHMs(2-
z!g2#WcwoPE<4+3Ui&<KsQ7e8UH6bkd`$Hei7yEbB1`=MEY)+od>1haIx%t@2zqY7&
zjRN$?qr{3*tk_dS2us0Ae)7OElg$;t8?&qxt=8hXsL5xR+NVEn<wPe1@WCw0M4M&U
zOsENADYOsIX}kTCy*5zdxfJp8BHmO(2+LhJ`O5_Zo8ydt43F|wKyL-5h7gwW)>Abt
zx=o{%zz`p_cHXO<57rRE@?!MQwo@bXw;KU%%u>Rm65c^WKC`@eYgTl;`$Pp8j32aF
zRBpznNKFV!-K*ZI+Nr0SL|}=hyGPXCBkE}gVQF|&X8LJotg`~R;Zf2=Wg51yYC>2V
z69S7m#aRzii`0Hmv0v0xsR?0e8Dy|)L`CduB{0C#ZRTB?`2Y?1%yKg7=A~l$XhYzH
z=W>a&yo6PUnh=)P`*wW$%USbF3NRVZWrwKTA)06iVR`d==@`Rv8+RK5Gd#*w&g?2S
zdul>h+6G3wI5U2<w;r&=Ea!RUdA^&55EiZ3)4rawKNS0{ZutRVRF*p@sXmBZKZxDQ
zyApSomsHPr#dE$3*$_$fj<<ftdy+SnYn))BCirh&MfhIcnt(?<DKk4Mb0i)C@s5#P
z;}a8qLQ<8$&=S~}m?cT|1cp9gaT$y%gJ072moVxji)-MZ8aRZu*Tc|yw*N^n=%m<(
zyqc)-!SSnYf{IPhhDfZWIsgY8fL{_U*Ldci)!TU=n_hj7WC?h&IFAop1hEE^>NxN?
z4kCy~mQ>ZiqZ&j^*O+Tbb%gJ6gdaqlPvYWpa*bn8mV!9;hH{NNPZ}gutzc6txSDJH
zWJ%Q^*fa>PL}yCa#M?Cs-HG*=REf|j5&Do<kyOdhDVfDvpxzc}McXr>P9`*`_%bxO
z%r4D!XmFjyZ=miQXh~G5q>2%BV?;}$h$Pi=QFpm$Njy(Ubx>3u6fMXeNU9vsAV;)0
zTjl&LBx&?|Np)GYyev9Vd_%On!Qy<;GM~i-qGbV#6J<JyGIKiqNttv~W<rEdTN4gp
z;Zp}ebYWk>-Y;M<6`m^CyNbo<#jx{YKRR%sEU-}4iwrEO%4LD&EN+xVHp*1A{gZ6q
zC)r?vJ9m%U{B1&yVpj>D0_!WBJvkN<KE>8oIeT)zCG5;vU*qhFpDW>sTVLnw$(u;3
z`&`6*ZWxupN1WXw&Y8u2kGL-={+qM=n{EG_^ZT1^ui?UKSedTn!fILEz_~PVJ?MCi
zoJ%8%<3U6`7)H;)I$*aBIJ4Mq9r%Lc4Zv;#+r9z#ZD8AXfv{cdnc5A)cC$DgxTJ#~
zbi94QWgm;LfQTz#7(Iizz%Ca!v)C^ed_nOIV0VLUzXAMiu<gYlte8F1k3raD7C!+l
zPe2bk-c#W6l*Mfzq74iqIzUoI^KQ|+7mEX<`Cy8d@NP@k_9cAa61F{+4@>1ED1RCs
zmd4^OeApHiZ|B{&^WKzy2k*Xv#XEWToh-h>OILV3VofAfAukp3dL(Pm`Q*Nf-QbTM
z%CM40mAtYe1JdzN!eI@nV{tQ&npxb=qjrkPfIL2$|KuR((R{@FJmQCvFiui!gccj2
zFVRstCig?S-z)tN398|~!~KXOE2u{F9MOv+RZkUzM)e%Up!-~(`&<YSX+aea`ox0}
zB2t3t3g~kMgpe&2RBfP78wepwsEnPm@8XBQtvX=~#yOX8eaJwP>KSMFjB}+p0XQds
zK4f4?l>!V?fGfoXz_|eQArF#NcYxs?;7V}}?;OMTArF>RD|o{dyeq}=ymLHlCutjO
z;mUbMId4Oa_hLbRv0zUeFu5j~LyR{`=3v95&g782F!}2kYD+)uDHBw0`QW#FA7cIm
zRkYw8E%=ZT1=SgG=oxVs(U*eioEUyi96<3$ar{T|D~b=uj1S06h}$lxRwLWh$c~su
zQ`BatuX}0^ew7%}Iz}8$EzAwk<%T$(T8vks$ty8}U{jr5dBJ<LuVOzWI3PHj$OFkm
zK}od=Dps);FbV1<!LCHQI=qr@CD5N5hEix+3Y`d+Ycg2GIWYIDe3-E+b_N!P8qTnW
zv!|P1%Nf?Pco#6-1?=hO?*@juSzHVZi-A4e{KvrXF^f}q!&Kg$ZhjhXn8xBQyx|s#
ziOV3!H3>DcrxH$$p{1fV$(l*3T+u35^dT!KsTxJAMvAQ!ivw1mo6qoy9uv)viQUK^
zT30!q>h$x|AlzW5WruU6>VI9%&7NKx-7A(8Fu&0;(UkCa4azv%>|ZGVw+j<VA!1?L
z2^Bk`8QtsxC>1~xg5{dTBH3KBM6wRQN|6Y=5k^Us30*Q_Uy941cNy!8K8L2yp%;-F
zd?2A|Ig4LF(-#!WHR($7VCQFher~$bUxBr3A%_Y%1zB846$elpU~o1-*(`ns&^wBW
zv|uSvCq1bY6!2XN_&}nDB^=H<6!AeM(vwsP&?13)jwCw@@C=|L4jPkJ#{>V9DkW2Q
zo7&dX5-hwV394(duGeIC#03&m*JYEg%cf9#Th{e9+g>P}RLHjfE$jL>+g>c2RLr(N
zmvw#4wwKE$m9y;)GKU6PH~PFLS!k22FU9RL>vmZeilb3*GzuZuO_S`!QnLpt_OMR!
zJ*d0~E$IrVGr?p9ND3HhQ+58As9BQu#cQ}9D)vKXI)XY2Oh)hwFJCE|M)`^zYu5X^
zEQriWP!-C07qYlm*1MR+<+9%8EN+yIYLtzp{XfZuev%C%*r(RF==;3JVfY9;2o8os
z%1HVdzxqC?*hgQrBUG|I06po%)wyXhaew>CT9b4p%o7CcrjESNjUw8x!^Vz`2cwAg
z>#(sSuYgfR+jZF3k!@fUk=H=0rW@ahlbQ?!tdEUKpb0ri0*(@mp0KzK8kMp5B{X`;
z;u>gH11)I(dT3P7wpSpj0x5_I6jYT+s$_8$GOj|Vw7nV`SF^YV8P~A578%#FxDJ`t
zA#>Wl9+}p&xB;0qu(%PKHnO-0={F%m+P@j;H?#N^a(;zeY5Qv=y{7Z;IVV-;v~Z;a
zZ{-_k{RX<y(CxD5a#{4JYQIMGt`YkXOzs!36svC)$Q}phyLgT(=(Q2cOs(KvD~uwR
zh6H{1QK?Xo3VRVLkyMwU@)GP#Fm^xwxr-nphH5S$aK{Z$u>p1=j2+_UeE{|+SgyI6
zARWUZ&3FI)28H1R^^WiNjvqzb6G^pL@LeqQBkl>&dO)tZxfo#g^Q!gdLKjImqh^@R
zg^*-hSYBea>VvH>R=g=PrxaNrF<$tcWZtQ=z7%hlIc;a#cgnnX()O_Y;J7P&ynEqQ
z3UUncCVmJxcz9LSw<Ba`{qLpqOdZ>Ij)A1A=aluF896yP<mZ$PEN<hJZ7lxCDL=Bf
zol~~6crQ@y1!m;rNvd?9OlNThP-d|BBv77Y@hPA@#o|(+ECpud<l?jqP(ERCEl}37
zxDF`mSX>X3^(@}TE4T4x<m5}L?Ywe3i;wckqbxqbD^IXEgI8uyOs<(kx#n_-tgxi2
z5)G=@+(C_KStEK9fycQN(XyGvEuv)$i{FZtZz&#lu_9=B!PXAl`9V~E5W5i37(CR^
zIJa-|aNLj}*d_=LB)vnfw`6k7r52r96Q|$A+0Z?Y2Kv!}!TW*!eqclQ@C?vD!{TQ^
z{~55Md(sH>8z~;qsc+WIY05QtMrEi|8PXxzLN!Fsv2uMcXeZ&@46gz%fX?6+=XHz4
zF~BPZ1kf480<TyWX92G)5I|>e0eD?t@oV7o8uX;Q{08{EVR0sJmC4)DurrId%3^UY
zZ<Wj2QFkSex5{I2t6<S8Fm8OCVA00nkAlTV7Pku)?JWKzIDQh`h^E0|I&_SN3|<T^
z7PDc(5@@l6#R<qF0d*z!TG(Mk7KtoQLKaCZPDU2VEKWg=DaeiX-;5kLvp5x5q_X^J
z$Rdr!naDO1IgtxINtK0cvsioq*<N7rMPz%C#o5R`8+p_IImkVS#g~xnC6@m(vb{{P
zD%rj{)TvwkKVgt>4HrxWsh0DtW$`ZHy9)$ULD~&`ceA({_!fg;DoBrk?_(CH^1i8T
zJ|vCzO=IyE-ggVdB##E=np<Le=E`C3ayH{v342$v__{3cI-B7umIW45JZe%x)S(^4
zf%r_lL;CNK5jj(1^so1rc*MjW;fx69QO6CWOYxRdzGc$`dw|CtFpw_ApFsI1n;s|u
z9wlHPU5W~ztYFgvTlpSa`N8x|9pIG*XnJ5wYjTg5=a+i_#}XwZgM)()^({PI`zO5P
z6W)*JCzcCnxnM$WA0%w&(I;p^90PJ)k2g_$S5F5@5>S$0NHi?YmI~;)U`UJrj%5X@
zQ81<X@>Ymip&_|`8-F_B`*99-w%FXi7A#*2PQ<goR#Ik?F0&)~37j@SwyTf_$xleC
z)5zd7vLWY%R109ct}YcIGn~B0t<&F|%5|}r*bDY1#KR_~133Got{Wg%RyZ^eP4<g6
z#F$_gNak`t=0zs*wWgo|KS*6oK!OWWMu45D-B7Wc4RUWo={6hWs*4NAs}k|Ti>t0I
zAXCT5dcI2z?@x8|L*DKo-;>-?ORDwIVm<VtI`cH_aT<mbJj3>qflX&cx<XQg$U}^X
z3nZyR4MJ@w3X_L1ZG8>;GHw0j{g}4?2K||~O5UN8_awm<zI)*vt9dVq>v)Gc-jl2s
zcB6R52HuO}7T&mpx1}EATi)d@??-Vv@6gVB5)Tp^1i?C5a3Pr7xL}V`U2H*DZIy^t
ziAI^4tK@H;dYxZ=4erH0YP?_?FZ7|ZnkjV66e5VMk_ZJeK7!ICs7uySt|`(W8}|LB
zbf?rUufP14gDfJvSeDeK9^@T<^l$iS#J#!3SZI&I9>-vC`*(Yr%s1GU&%gYv{5RoF
zfeuc9@cNm&36+g3UO$V)xx9WZZ^GJZ{X7=8@VYI$5nb;0g3)_{xrq8CSbP$gNFW+o
zM6;1Wve-FU)FVMUP7))djmVtZ&lF^lg1S(=8Ch&*@i}C94s|6<fa3sUn~l0roP%t0
zC?-`eI1{uMD%L`4vN>20K;>z6Vi%yx1vd7{g(kVsiSpM#w;I@+>^2Ufpk*y|Cs?kj
zr6HTW<Xoi7%EzY{;jM7+a0ns#V`-AV<xOtUANc>iD%f5XT!|Es+8ey5>cShcr(`8?
zgSy;?G>|$QtT@VeMHz2NwQ>XR)4+$3BO}*T-w=%xr>VZd9d|&*4p!sbgwjn`<EU$L
z$UDVpYI3louY`(~Y%WP%phFl}Mexh&)_(F2lT>bcZl+Xc^p^M5r6^b)%%H)BgDq&=
zCC=~?XVId$gqKt|IO`jn1M$Q=aJL(r2a9{$;QT2r-~tQS4F4@I@D__>Kwu1;;g1D@
zu`JF4fmv*Z{{jfSz~a}S*J}_;P52wo>kW%DdHYP>l|5$mSuD=w?Q?lo_L$k{vABeH
zEaBbhmX`94r7W)G`&9Dbtk&#P$A^<G$1xEf*i6&@5-!*1)6RzzEr_%80*VoINC}mM
zMFhnPIz*-<RjXjzDtu0KNvbx%wvELf1>27-ZWnCZS^P<G|0FQ?+0oEFnr0<&oDXdm
z(?pu2S^{mCus9a##!^YdH!M(jhYB1n1rX6DQHO{!j@U$$AnK5eF?Qoclql+u$}I^O
z=Ak4}hZtLYzb@)0i)K^}rHCj+)FH_e2^XfI&7#f=P0fmgOH5Fjs6(nxBwW;qwum|;
z;f7C=h<1oNq#^_tQ;2#eX$l={8&U5xixZG-0&*sYyCdwdO=NKrvQ1)fGO|r(aSC!z
zK|ZwqX5_w^#i__PmE})EwrMQhj&!#p6WTu$IcK6CRE=gK=PVXqK+YFfd=WWcWN|j~
z$ws|s{~Y9#!{STG`4Y>28985O@pWW!9a)q9tE$5uPXBs#+y8hXl^g)93!m|dXRH#f
z=6$O9Fml4=n#w|=MAsJm>fc!_qpm~08#%ZU{Z3*A>>sMj42iaua2=oBJ-#O?@xxvP
zw1}rp7%4-<KDW9Ck#<ZLP_ke|-H%)W<qAf`aNwf}Q5!TOmTvvs%Era#7aaMQy^IYr
zvp!9`GyV(whBr8LgEJt?byI<E+;X>ygSg?SOnFphLy}_&&z!S^FYK<y4duwB9NCf3
zD&g(K@k?|p9$<HUzv#AK44_`}esSP_aWuvIWG4Gq_kEwtcb_bTc+rV@;pz8c7r9~!
zl!pxRsK7|7Qe;qytcgEQ4nE%3T~M)$wQ;wgbc<>hxu%elm@eWZ<9$$<Z_<585l{*n
zvg8RUj}2KqLi7<DlA6b~EiPrhUz+FlAB*C)hCJzIHGE)an~qexeF=0Vu1W{)UjheE
zTndd!+0=R|^e=^hwEYRRf5P^E0{cH<`#*(7Puc!Yq5o61e;MpiMla0-RT&&#1}D(*
zpF^YPZ2#vl<T(tZ?d8y@oNX_MA?0lQ3uyF$ZGQnnUa;*iVUL&W^IyX8FWKi;K%)w_
ze+3MwVEb1>qe`~D5{6W=?N!jIifylgAysU94Ro%7uJnCs;IJAPNpUSSt7We`YGH6K
z4597y(7v8s|9aTJo?ZV2Xw<+yzXAF;(DmQBXw}-A3ctSpvPK{Y53JnOb<6Z@+~@7?
z^F2wtFR9i*i#1fa?G7!r47v0n^}ju~L_6Y`R$cZ?pO7e^ME3Mu6VNsG^wkMconT7t
zG;nSgq7TrJ)DIk*e;y8-xw9Y6bG38k?VK%9)E(Hlo%5vl6fi#pY{|Pys#C!E6!4_D
z9+=kyTM~Wav_EjJ2c8sX@a7r3E%gjDc;^h>lf?lUwDyB|ym(DJin)(&b-6h0-1<{E
zt`Y6B#UV7wT`AgCQhXSO|J4*eBLHjED$c5ko&OGCwS%2+5wI=-UUYsbymbn#Q#$Mw
zIPB_{iAi{C_J|gHL`Ry?c#I4lQ;kXtAKn^uQ90cjb-6ELknlG#Xq)IjZcQcCP0`>c
z&A;OdEz#f&#YFSs?7F%Xoer$7MkWk6w<Fr7i!LMwgUd8!&PQaP#QaFO7R7Hh3Zl85
zok(XVGNw*W9<s@!u_38($FIH}Dz?+xQn{wmojjOip|DL**SFIdE)&r*(YioWiH)t2
zXq87xFeFuzXw^jVv2l+)cgE}Ji5=A{x+Tz%9Py4CTHPlsE`z#dEPe@fU$VFc8rDEl
z+P@y^*3))kgmBoq4l35MWiwk~;1(D`!`h3`<s$4!@fB!tg)Nw<hu!MgvYG~H+Q638
zsLT1uQe`BLe)cMAB*ZR2gW%pEj3R4CEI1Bh)kXe9&Q8|+cG7rd^MQYRb;&8;<H-x|
z)Q01FmvW?EP7}drzJ7W$On&1TK3kQ%ZY6I{oV$)vIo&!IH}ks9EN<s@+bKR<<F;V1
zMNrZ|1!s{Z^j5v2;B4?SZa7ss2_Q0o-m-TToDEI^!>Q6K0Feds*1e<PZ15d0oF3{J
zJ~D>?d<)-Ea5i`aKb$I^HGID{w7Bq`0u^2xQ$G0rTm#krZr9*^uP(afac@@R^15Nd
zFa<G2SVi;F8#c>#6iP>-5!DjaP^yMTjQWwGdYK;4#H3aMd*3UdVg>cKB~=DgW>9Zj
zuKA9Ks1K5S#U0b2B8@Fz&xS79@Cz~nT#5hBM?RlB=01oz-1F6of4dgBkPtQK!0oy5
zSm9-GoijK$VjC_wF6!Zz5wYiA_PQ>I)lZdP=fvPwONZUlVGptf0xt6DzK_K>q1#R9
zPTLEhTLFu2LAP5hz75@Ov-lx&dr0x+eZyvT?uVBBw_e+oH(!tZb9VN2yavZPyW?CC
zSp)LT6&_(ZR4k|Y8@cAoE3$p$%PY*d6)LvUnk~8J`>W54ckkrqSGQ#U$D*q1Rhy%?
zRQle*WpY(Ws-guuf~p!x)hw<-QVol1kyOj#IwaMxxE@LMEN(zj1B)Ax)X3r{BsH<P
z8A;77eubo0EN($k3&rFr4Ilm0P_deI5mKOD3VTKV1a^DEx&Tk1=~LDPP=EhMw&;f0
z^=|JMj`YM<ubi_gXI;|@&Z>gNX}~Itbyc?kt8Fa453KI9uJvEQ>Ms_r=dISWuJtC~
zY7>i-c&j9e^JniL-uiszod2_W|I4i&DH>DAHQ!|tebsSuYMjGc)3kj7Z(P7z(=>fN
z)QhL)t0T_UTf^e@P;WhpH$uIQEKY!W2`oMh9Zy3KqOm(79Mg2M3ypAcMT1<?f<}RP
zqCp;u8%2Xg7B`6oO%#*jdu&cOL&aulKJnK|P<agokq6@|qJKU@C1?0%#o?UVTlWV2
z6JB~(ae>rR-2uFJfI#Y*76G3k(35(iDZEb#|G8d>e8!3u`!=Z9#wvF8$E@UuB!t2d
zOEOd>vk^!Zbje~%wqHW;mu$&)1vIUo*&Xb(L(@tYS3%P%ishP5WXXeX*)E7%(tCss
zj>$@4a4AjONUEnW_$iBDz=#*@gR5Xf6^m<Ocr6@6ouUR9+(7%^?&a`%<WP&mf4hr9
zO2cvZs{R6)iu+UE@hR_1Z<*o*6ek#y?@{pGxrm}gWAX)N;plHo-u=AA8r%6^La)!Y
z>~|i{8|`9q?GJ289`NQ5cqfv}#uwte<s;sO)RS}+o}ya8h$d+o1k@lHkr=V#*1GF4
z*qz*36Y0b;<A2QrJn*P3J~49QM!aYDIsN-w7b>3dKtCRIq2hT3=wAU{sCc#k{Wj2r
z21XZoor}CBJ@EOwPCjo*4*a8*w8-hnE0_Pr+>IjB^>?=%)yc(6{z>fmNpvNd@@M|H
zL*E`U;r{K?RBp1zb=^xHcdQ!jitcwsPckEX!y&rgWARd%`%)RxzD(x6jAC*ZiFa51
zNjXU%z32{GWID#Hu+BJ*I-N%P<W@++H9DOhBYl$OCvuBN_^&A_a&3q8-d?EK%PQQ5
zP<lw?5xM3AbfOF@T+bf%vOKsGKe|vfE)?D9%Pp1}FP44AwGlB<(w9Lg|KB+|B&~uE
zyZZZide}3NV+L|3g$5lrCN`zWkt}8hb}B{P2_{#acn#B`BAsT(I*PmX4$#U8x#shG
zvW8^I@UEyox+ha=?)Jww;$dMQ9&x#7y<8kd3eP1~zUY!Kj-zJowP^BM96)e$+S!X9
zN4wbl+hzTgvZ{9JilN^B<JXX+oCdF%`i~Kivp@<UjOCiYJwUfJllREv1E}W8;ytoh
ze1+G)!fLi$-XoU}Aes%QDtV7Q78mmRg;euNsute9h4&yc>qrThz2WUBelIw^7nmZl
zPlEp^A%y7cjzV1jXvpBB(CjF*qwRN~?H#sIw;GyNLpy3clSPYU(ShF3*UOCR>0M(-
z;gHit<VBOaDabzsg;2a1`EO?NIn@0e3Z%Lu8wF*f0Tkz;pd5<HO+MBH2chC1&3(u<
zf8l|MJNW<{f2ywjyAE`RN}+ct^rdg~6nZ~p@pIVwIa|0|0ee@lxEA)RWh;&vpmzh?
z|C|_hj(+O?X5qC(My)wI|Kq~?)?x6%`Y$(*!RNSH@N5?PlbSkQj4rwqh{5FCN~*V_
z_gi|7-?sjF$hdmuYOFNdIP*5ngPQF&u4fxJl;R9vo&h|l8P5PcGr&-aYk_$!@StYB
z7WAwILn%JWn;+#psF^>;n;&EGao+qm#jUM-FM2#qY{8qIAfN=H3%zW=Dxj-E7joH7
zP8+TsRR5t4y4eqSuLt}<;`-pf*zx_OFq}9JJEUD^(D27-jwHZI>8chHAaVnvMf4;|
zDQS_x=ggimC%q#HJ&a_cwY0|2pON(EDEbrsC0g=t2>GF<6-9rJpg%{_pF`=-FF*eT
mssu;v@A;Sd3p$VeGHTfHQIR7?j~p{(sMgreELy|q&;JKr+#)Ri

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_transpose_instance.hpp.9CF832E322CA1530.idx b/.cache/clangd/index/device_transpose_instance.hpp.9CF832E322CA1530.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0d8ef17f0653eb3e5e9c472ca0443cecd0abd69f
GIT binary patch
literal 1608
zcmYk6e{2(F7{_n7+jZ^TZmX1fcQ2-Q_tL_ok}lZFvLmPs!)X?A1R^G;K`Dd5q-jCg
z88xOb1JuF%5hZJ2njl6HT7f@c;zT6-V}|($0{a7!4Kq||bpD8uDfsrx+j}m#r|<VZ
zec$(azR$ha(bn4fD9<p_b#3YH_|~-?!!T}gr?w_H&X8ZenqkuUvmK*d=R*yb4*&68
z>i&^^+1h5lXTizs#r{ZtA?W-pe`&`*uXIoDFncyRe=K~vA-w3w>^)ba{%!b>(EH`8
zOy=f|j^IoEhn`)PDt_8IS-pRFbKkYgyPs~l{-9^^`q{JM!G(NdT`B)@@u1LKyl>tU
z@8#ZDxvpcN>)1u}-s4wJEa<smwlRgGzmVg{OWK7c@tsw#j`ck@^74iYr{;aXLi*uc
zO{pb1I56?W_`7!}KAbsOXn!DoJG<$fgI&uGJm-r4(i=MU4E~dI%(_|J`~2T~4s5@b
zJ3V}PN%XHf@4wDI+}b|v8$Udxj3n30PhTyd*wQ%;|86Fg>3-o1@xXOw@Ot*C@JfNL
zaWImbce|~?I0++{`}i((+p(X|3v4yP5-&NeK=IO<>Cz`tnd1V>6YSwV4l6KDwyHck
zxodQ6e$AqL0_!B$$NT15fzB>{RsGt+t3!7Mww7Qo@11J}ivMj`{rR;m-%JTCM{qq~
z@3aEN$@uz|m$Z~tQ0uZXf9va+!J`u`nkdPpyqs<!N{XofrJ5>G$V5QtroNnJE=q=J
z0EJClQPdl!%kl(l9FQCBjpc~kY%3+80H#C~z|?>Wm>N+5LjeRBil7RkJ*aY7-qSRm
z+&{ecV)k~OM2yg6zny^`lEb!AGzBmvqyVOBDqw0z1q^8jFcd<7>6-4h4b($=*j9#S
z0A_>?z_5k^!y#N@?T@^@#rsfhUzI4i7rUdh<#bt6dimv|rN8;>vN_FTIhFidYLL9N
zPe*QLGixUDeXJ;XgWe|F17FbBM3*|69;#W<{ZVVZD0w80mzqmVjLDt$6geTMZKcE%
zz?6gnm>N?7Qxhs+D24z-2?UrP(>rNWqNFGEw5^Po0hp060K+j17*5~{oBDn9T==K6
zI~}ALPDUVoP60~gRG<)tfYLb~fiyS+D9quCqHB}Jl#exS3_Z~fmvxfqV^Ur~Z?xCd
zBYLx~jDP`{5itP60Sp+9;0jB|H?J*!f>OyBQ!V9zSPuVR0v(W#M`g%L8-GjV$q%Ln
NIF~5YF?AKf{0F=AL(TvI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/device_transpose_instances_3d.cpp.5DC4547776B7C3A0.idx b/.cache/clangd/index/device_transpose_instances_3d.cpp.5DC4547776B7C3A0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4483a90c0482bd9e231b8f747c598bc10fab3ca3
GIT binary patch
literal 1728
zcmWIYbaUIm#=zj5;#rZKT9U}Zz`!5~#Kk2=nH9_o439Y(7%JxM4b9EJZ6<K-zvj7~
z*JrX>e>gw7wR)qP;V!YC8W*0NRPvc975}DUf1Ug3F5#sI!?tTAoY+&{SAKr4r=$a0
z!s(!By+-Rm-)m0~|K^@|u>Ixyqi?>w`gM8Dj(PmYZIA9w|5S6+c!s`ves$N!RR526
zjz0S49dxH|Mr{4xX1@6UvqVpyFaP`~Yv&9e&p;7*@zY5^U%op1=iO)D!cym1`Hb?*
zFN;NNICbrW)_K+MS*N9TJBJq>{(4XQe9Kzt*FU|IH#(*VYz_^HGD>af6FR@6{-5vK
zu)f`YPaeLj(OkXG)b`{qH{lm8ZmWb$;un7NEIGI9%YN7G{8NoytJtu(pWUs$c6H^h
zB|TgZ0;iXliG`_X3OBsDP?74pS}gzGo_hAq8xMp0nR&hPnuOd>NOApqvPo4?^~YSU
zMNFNl{l8wgB|Q%K;IP4E;m*g6?|v$0D2p;mrDsda-F?Vh$g|3A*Td#+rHJR&9h(;C
zWL=$pu>F^W(b=L3$D&-Ng5<ONwq107{k<h4RUqr#3VHRX_AhjDb_U02Ij&HcxVzNn
zt%cC4soz5?G(@vFS6=JNb``I-SSpco!@S~(^r;@NNgM}-x8?>(I;`zo+O%y~n|GyE
zHNR^3u3(EVFSb4L^i*2f(e}DQVD3U2)hk!z7EQCVlG(Om{hG!94*w8pYTdGmUH52h
z;P!k|(^<V6t(mRg`J7s}^lsw3=^U#groB+8+INhvV_wIy!e?&VMBlIac}6MWZpN$`
zVLP4oHRgKUp43%xM*C`yN`Jk(Nt4@&Rn60rxlFdaYmh&2S<zx^*`Ae?x-&Fp9{#^F
z{*!|BlSN(=Pq^MX^m|$$o5_!-MTWLE>0f--TJ1AF>3z#IB2(LprSwhW(W7n4CZEU*
zU2U-Q#DTO6EnSNyc)yy{+ZV9%PK)dNOWp}JCo|m7%;ME&7iX?o8uVXnW&Zy34-@8o
zw|jk0_a5i8vU|G@ncaAoxUh$R4%dlmz9~npReWl^FZ;4b!sk&Kr;eOxf7w2XlZ_d(
zdM?Horn~a)=XGo5-WJprx=K)K_Pr0h=Q4T!yv}0_x$&RRXoKY2@Q;Vq$@KYeTR)@c
zo}HoKY@7GX=RJSB)aL!0Pd3_}W#?|plowYs`yw3@GEctz%Jb?M0p2?go7}%GQKu63
z{`Ry-Wvi~cKlMBqzi9uQv;Sr8%s=@0d7eP7_c#5#vy(H5D|3_VfcdcA|8s?YsLgU7
zhJZkZFbP%(bua;x0fWzXPWabqbS&jz@CQl@vI?q$37B+IYFe=@69WU+uLbv-y-kCe
zxx)C__yv0p{a&!<p|t`}m{5>VEZmGR$soyCn1zDtJEcn3y*dok!pp|XiKN9w$R43Z
z$VbQ@#&c|c!|?M<$t@<9Fi|~GADA#pn2er`56oPznfv?CHNW9y0a|V*WDQdU<hcoX
zBY2W#lGZRSJYkY<lHM?$ulb~WG2NaCj67ld7W~dIuZ8#tslS+L!^#9SiBpIl$s|rm
zewf~(&eP{LSdIS!^&-2eLUFx##<>(XkX{8LLnOTll7=w7Fq)AI90RZbY5(9VV|l(u
z3TPy+1g{({*bar*y^NQ?ejI3|D2u2P!blkw86}uePYqbVu=9&|gX|R*l1C_#6qS^R
zDSCT=xlZrK?tefFkV6%w80Z{Ow8Nb9S0HTD-TWiZfYHXs#)lMbk~Wg|Fe8Bh>Lck7
z<1rMa<|KkrOmR_iu{Drp4lB%@5!aXfYa2^YXmAKaFf2JSFa#l(p$KLOf|;C~l9K`C
zdZ+m^v#_$Ub8vF=@Cxwrc?!BZs|j*>Ithsji%3XXSX$fI+SxnkiE3LZn`y{n6BUz|
nkyTJsQc+dc)Y8#6Ff=wbmot&lHF6OZbN32oa`SWaVDte19u>@>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/dpp_gemm.hpp.CD386ABA025C1896.idx b/.cache/clangd/index/dpp_gemm.hpp.CD386ABA025C1896.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c000f5d7f1fc9c5caef0f5514cdc3a751ff8ed50
GIT binary patch
literal 19250
zcmdU0cUV<N)4#iW6vYD<fdg_q$14^tO~3+zV((Go2{te$#u`cN7+WmZupyuX)YyqF
ziU^7=Xf$eKL+rgoqY_197i;93d-u#eH{U<+UwNMQ^+|TzJ9D<q&d$znjOoy}ZD2=(
zq1Av6<Ax3#-B&Oe43**kgwdnoBcQU<)nFL6Y<th4!>>lT6@OXgJlQ|j<w@4tfm3gm
zBPDAp2^T*;<$0;y)XbuYzH`pC-?-z`7yBN!-y6H&YfDkRnt4ZVKi&Oj%AV(QM&JA4
zlko#amFw1GUGwQ-y@T)nb&&0S7r8v<Qp@kE&b~jfozMAQnTG4RTXNNy#q&O%HF?14
zxh1BgN<*tm3p==WReX-id^LS!ixbZGTKdho7VA;TwL^A0<!!;iO5tVYBfZ8qa%yvz
zo|kfCZ>KMrdw4~w6Y8K<%kn}_d{eY=K)YM!i7_=d|LQvFZPcN5gN8glvocdXAOvR=
zZo8d5d3&R`JEtxGyzb~GgOYq_1x}7|j|dELtyg?6w`9ltUB-s{2A6zVqw?7=Hs755
zd)*GLMu+A<9P!7I+}2rB1Llu+Zu#kmb6u8vGO^=u=SIEHb~;)i|F`K0kG4-6vp>r<
ztA1cgmF0i0Y+I?zgCpj#uTJ%PKJ0?QrAta;(dQL=h1Cf;aI|Zl_l*(P%z3qrd0Z&^
zfo8RM@#W%eJ^SY<gO+W-w&MMvq2)$aQ=)cWh}vCHt)4ux^79H62mcgDp2du20|US7
zd7EWyG^O5`3(Gg(HMpd0R=1<8hej>wFg@Di&o678jq>iaI=*wA?4oU*0v~KmTN>i!
z(dLKV0l~fQkwI0*US9WO`4NWLxnipSLDs$P*uil#hHZOT*?o3$x5eXjL|=~1D9LJd
z^nAhNy_Exnh;Qe~A4Y!G{zM4-C1zFs!ZV?fZ%(g${PU}JRh%v*C4Fpee(Ph4a_&rA
zt47&t)1ACF|6aaIixtm?R%lx@dUCVgFG3F9xa(E+ww>3`=^yg8f0#4rZ;y8)er>iR
zB6@06^wj$+Fa6o`o$G<iYd5!lRBqdj{Yv43<~beKb{s|vx;0JOd1u{67mR&+PrecK
z^PNAuri3kP8(DGvx67Z_UGCE&Xl&ZL^UuHPlH=F?+inwL+RSY*Y*NBRdi(opQ+?Rh
zU41;4{dV-qv0>}BP54mp(3$Lf>2~I`kt@qRy^vzrOOK^A>wDw(%YmjtQ(sq|@>$^O
z<rn95?z{GdPiAWD{!(>}@9YjwRtQbpUvvno{!5xLCGF_Dm>(nFM|s!kmD={vrsS)W
zdwibtRj+b))lt%ijP`>jEc)!qgEKA4-saDJ*KG6IDGQDzWR~A>du0!wro#un-s4*~
z-gDoaY2$sa)cR?H>HhQyFJfNSDCo1{db3IeQ@qB+UHp7QljwOzp9K4Th`Tg-{`?we
znmV1C+aUheh51=+dOfXPC$+<I%b9+id+m7`J-S(+HKS$(=brQ#zhdXYfN?=5`mKA@
zGc>*GkgpeiarEPazxt-9bvRL6amLpwbx)n=f3*LTtnFXso?H-^-MGr|C3kNm{aCoI
zZDy=<NRKMXWyAaruSj0H{P3Aezg*ezVE*b)0y1j!?t8i1%ik_t`eCy?;ClUq>mN5?
zbACnoAdgO^!vofwADKR=f5*~LuhxCzF67K;{_tVOq}o6BDSvd|ZI36962=eh=iy~A
zuq7XwUH$2%!HxEgHmEO&=S$+P5km#AGVn*(d}Zt9R$?%SG&86zJX?6%3OhS@&9m{#
z|7<PMp3M#FDWaSr0UAM(auFRbZmPf*Nc7XD26YZq=1_}9P^5fd<l)`k>46fBg;olQ
zQb;Tsv5UO;dXdxqsgugnE+G}w8PsV8HEIM!oZ_2D9(vexn?!4e8q|%<w2{dgv5Tag
zJbkoM{CtV_YGP0~QDqYi&<Kh+pXmDWyU}t647Y_r-9nTtFcwZw#HA#9{^0DjArk$(
zl|lW~^r^3{K(_0vYtsYQ&ib(utqc=Rqt0p6T_Y%B998z)&Q5!$NVErt%%#d)8mti%
zadZ1(Zr8&L6n9!3W^;*Ey~IK_f+CftmiF1Zzh!DA>H#8Kne$fWt`WOPapIj$M_MGn
zSQ;DD#l(3ran}fnxGx&G>+T;H-b%E4OM^O685v|NknK5T_dzAMdq;O#6FS+={CBfR
zji87`_QhT4v}S`lRia?~sHu=v*9eMK`E$AJyCp+9NVFStGEN!iXDg8Hv-yXJEtQ^p
zfEMZ*)Em@v16tq&MdW=yNX3oD?*m1IB}=EC>C{^zD57-ms$a0<m$?%4hoNp}rp=%L
zoY+MsZ~le8pVrD~_$bVvx(RL`wgTDaYh%t|KQ?#~EEi~dQ|G4cwgTDJJNTT8c`$qu
zEESAl7FA}^AdT2XrleLp>+W?0T7V<<j5t4o1?L1sYDEl*E(n_<gNA{~1EM?t4dcWv
z($%xm!YMzD0}X=%ah90Q5?_tjMFzaz_%im{G>OK+Smse>9;gr}c9C7vgUkB64T4pz
zZ&1%r(-~M*PEaK1@`IQ0&%PRGG_-~Fa1mTQYz4A|4}P7vEu?P)=ma#qr_wXnRv^31
zqvat5JwsPp52cGb%xRd>Rv<fMaB$$yM+2%u|KSGp4mI6@{y9OBP}ez9(=p3nEqXx*
zaY|f(tw46Y_gjw~z25!=oD*R8P<jO03S>9v|H-DIM^ph;60|BoNwC-oJ3FzO$C<-R
zYryD%{f+WXu&qFL_>H@bDxID)9Tpjk!DedOOno(iA`##2yrI0X)P&(d3uBeB0k#6!
z4YMjHM0Q;gVKj7r)sqCt!&cbY2Y23YldO1yy1?N}B}ys@(g=z~jt!~l{Oix9ZnO@l
z?0zQfXHFWii)`B7>&w2^GGG<!a+|Ckq^5(kItoOcztz5HZKp$EgF$H<DUAYb1+t@C
z_qot5ru^T)hG}l3%0{qxoY+O$`um>xamhnCoiNZ$qGS?3jo3xHKi$;q!ejy_8d_OL
zm33gnIYE&o<(hr!daiJrJN1RRoMlzdg2m#*E)u`(bovcxv^zD|HK=QtvWEF<#4fV!
zizS&g1Ny*p8yVE!sOdLQ8ctB8*>Uff*bmLYM90DbDj`Y<93W0mq(xMf`FGD2e+#WN
z01;{`fmS%Nixgb?<>i=`0Y*c6*gT_P^sp6n_L0Id7h={A0PTSDl}b&ipdFl`NV_7-
z{YfjbURl-ORaKo-qpd)8`)b#&4)N`_-O4VnCKIP*V$_ITB+R^cOVe4kU?#BeX+%kb
znQ($4AC1fTCczk*?M`dJsur@Ug)C4bc9FMTrcWBQZ>$$J!K(gEYy3^?Xaq$%xp`F2
zyu5S>C^F1uK2_#}5^#bdU8+7`vOimV=}xObN7q@^>&#Chc9E!}vcGobjD=QU;U5#_
zG2Hf?pvWh_<Nm(=`>9u;ZE$^VCCXM1;RHpxo|-!5!caC_qJGfPCT7|MW||We=_XFO
z`=(V)8)yYOdPI~*&<ZCg(qlmDt!qwofJu8pEBUNOJ`2zYio~k1*M1(;A+Hj32a!DH
zoX6ZWVi)N?y0-Ug`2ieJQ0ui+Sqm1O6BLQ7-K2Bw?hF@+20|-am}v_PmlG7}?b-E3
zg&m3n+7<zKH8ouVZR5l)(xc1Q8527UgBgJ(i&bI+Yz4CWwQg{*|JI<RVCZ1ka*2{l
zf;D0n`F@DYcJF&dV9nr8I7yU~Ai@cX^k4qKufXZnIZ!8PC5I?EpiZ2i$bbiKi|;RN
z9u7wx#?edZ<!399J<vO#Bzjj2?9Jz(kc)_E5nMc+*hRXX+<)rVAx~h*poOkV*C1Pg
z>_PpK=AZcddK}CIOy+21bg->J_Lt!&F0@L2GSGS`oYi_x^^CRx+3}MXj%?EOaa))q
z?8lE(`4Mh4PV6Fc?|lDrU&FZa)F%W?CUx2m?iDA7if+~iGEvEpPiEw1-pD*;?M5@2
zp-!PjL&b8GaDo20IT;K4Yyo!ymTEWg+zqZTC)m-jUmv~6Ez8;oyBcXwOR1?8c9jzp
z8QwY4W7o7l;C_mQV>O#9v%xBIViy^eeSXaC7b9T7VMf_h$)<i9L6H$}j@CT(tlTtr
z>IoY5l+}95+GzwuMwVN<yYcNM3D61*H=ii^V4*p&i)8L8EH0BJ*cZ@`<3u?Qw+$yK
zGU~<TtCPf#&m`Ih#<GSgYhWy#ph!Y$$;e5S5*oqw!O43_l!u^YoS?|q+as=peZ4jS
z#sVFkA<7vT3nwTt{^Nlg%q5q5fEvMYw-eKLP$N$4A`^~8uFU<pKP(;S$tY!%#a7ta
zb3Bc`4otlPLj^61QDXdTg`K^9(W+1SU4y&34~*d$QI5eFII)YI7;tV~>Z$7z{TNm?
zMKGlZH8p}F6TENUxL|3z2K)zDkM&ep58KCyU8H!y8Q<x*Lt&V3xke}v7F&VrN#Bp2
z>Fv~RfIF3-lQ*o^8`wHdP$aP}jmuu$&kHmShI)~BUIb0!#4a*yPHO3{`=7c~8Cppf
zLehoS8nKJaY?zc$^>mh1p)06a)F}(zKb)Y*w1Z;`h8(#Q3mw6VPV$`OZ7YyHGx6}=
z{mT=q&k|ryaGv09E9~r|*CmbT^vs8ihm*gMC<|fZIYE(Gx17^<q&1o&Q40(+o0+n~
z5^{nf$+_PwYEXM8C}A8dSCkSJY%7qR@@dfO??;9YDrcwyXSb4A$<tOCD%#E{oZFO4
z;j1y_5B&i93g0(S8u-}?{64kjn4f-#i}VL83WNNPDBr=ozzK>h_>gsKN90y;*?Pco
z7ZIfh=D`VyEL^s4-{kZfU0_c^x4J3aEVcsKnVBKM@vFyd1?7b!@RXRIg7R{LB8ynU
zy=Lnl?SyH9e&-P-4?5w*E>g7fcv<}8$}kgX<sMP)frfH|BHur;+=v+02h4R(*pV^H
zm;hUW?4=FQ#a}3Oe-5`Tbh3>o+hDSs*hTt%=-ufv5zbjRus<nONr3~+35qPwO*F3D
zGkO7N1uWT8;<*&mj}u+w$H4u$KJL&84D|_7p1@E!L6IN6{VQQt!b0mwfqiMHG_=?X
zWUmie9ACcWgFX^%4VsZGn34rWBPg<==t8s2ZNB?SqMw12@{XF`!B9Ctk&PZ%ONVZ4
z8xI-^$1#UG=YVDB1VuKjTXg(U@Aea6$za0CR7nOuh7%OY8T-XoKdnDs0Tv!6oI;co
zm<=Z=vSt6f6(yBk9fmc5j`FBy9=s(vL6O{rxgXL)e(whphE^63WdU5JoS;ZvVq(M6
z_`>SeBPpsIiPJ{#?>Iq`{FXf*)(?+8CDCA5mj!}pflxyuc9C^n69*4|8U&LDeLY8<
z&%uL{6BOAUP$zTep1BY~=>wBaqDm4h9VaMKaP7(5IliAR0i6X^8gCl!Yb%hwx6J!i
z$YkRVxXK#B4pY+;xXL&|k$rXQr}=J71d&=W)YGi$X;w=kD6(Jfvt?eRoaaVEJMfrY
z1y>JSf$W2szsy`^?A;ue431TIrMtyeAp2KR=z8I)*KRm7pk;p%<u5ohoS?{WGcUPc
zy|q@guBek*Oq_~gML9u{!ydgO^DA#&XEd~dF?b4|p0)zn$JWKX_c}b+6;40w>;|fA
z0RMp#6e(=Xa_$~`)CD#lhS|rnkGHKr_K7bmoftm45~xpiICYzevKg*XPEe%i{lSWV
z&yImWQx7-{kxFE+tw8q0hc_3dz1&pUXy^z-tSnUauocL@BsTdX@<i(v?zAc_bQx31
zSeQoaBEMgZ>s~9p3=Sl85UoV}*$QM|om}_&rG%2bMnfm)z+G_nuocL@u`K1!!r&C}
zC3}JDW)LL<>;@+&a&w)3OzG&b<<^C+q#hwoM_{2jL6PEiXGX>aHauiBbcRkm1P>2e
zVQ23*WnCEDw5mh{U=}&dlmmAGCn)mgO4aiFvOD`=!(klhL`jDY=LAJczU|U@#gU~x
zpfaHRCx~(aP8laC^3biunO9X3zJ~n(wYg7}`>-FJpvdF!x$S(_4{OU)FSu1UFsBX7
zRU?LqM#eqD$6uQ2{(h$p*q=VIuh~S&2D8A4y`766>NOqk;2A6&jOH{^PQ$`+f+8=s
zZwSec+cphOD=1HI=ictN0@<a9JLa9*O57wG1gnxJn9|_9bAlqTKR@NWDq~es=mhq*
zvC=reRv`OrY|Bx79yOrQ0St2oaoz#ufD;sX_j*DL=~UvMup-bxxDp;@E0Fy@F}BT#
zH!lvsM8On{P)7LK3S|FXr_Ph3{-H@=Dd6iSrY63&0@;R90|zCH3)5r6#9507<Ke`J
zFfgp)V07pOEfP$fwfHaUn5bSR=F21q{UD47!}mNGTpmNaS`-)<g|-kdwtxcAG!Ojd
zt_a#U@TUbkM2rA)7vF#pVCJmFeNi(+^#QXyU;+5X71ilNo#{d-o<%1uE(|^Jpm3t*
zZi#9>HRsbHyfH933}f743k~BX;1V|b`>!(UUabH(0W)u8wQ(#WJZ;UJSZ#duir_Js
zvsrC)rbLKPnm4oB7`znW6q$2a?PBeKiSX7mZ(+6Z87it7f+<6&o~ZeZ)_Qb?&=Tt|
z+c)g<FE?UuIT1nt!rxTX*2Vx+VM0E0s6S=GQ^)!nCcJU1rwc;5V3)_JH>{C0<c+}(
zQ5|d^?1x4bW8Wae1K+rJqPm!x7gGy{ag9~ZwG2CZ<7+CYFIn@KtOY)T1T{%$og}ou
zt-(+^tT>O8V~ZFdhpu>ZT;J^9neXo`95)-o<j?|-k)vQ`EjSJ>@W?nuY(#YxF|Q&P
z^qnw74lP*Y;~4hD&^Y|e<Koyb2E>8Eqv778YVE%H>gs_i7xOB-U%#_nd6*?@lf>G7
zq=jw-bv6r_%>woMY8JMd)z#}SS@=s9;kZivNkRawlAz8K{AcNPjJLzotig61v8YZZ
z=BdOVp9&avhk^2FyLY>l_e(4%N6PE=bxrSn(1i%<9Kkq8kgy}%Klp;j;Bh-pe}Tb6
z?zk9=7IKF}X^pqz6F`Lf#&m=P=<4~5`aGll=&E2S9%f*T!=srI)wbrg7PLVakcUor
zG#*D&T?@m**F5-+)=O0167ySPLFlVSv;5u9B2PL0>pOx`e;6u{_2Y<YX|aA7Di8Fd
z*%Q^9#C((Zq2a8hMf{-!9_Po&*3#nqz~IsS)|Icn`eSioM~v>n(py9O`qCE@Q?cXH
z-=sb_sXuB1M*X2ZYseo@dYu998!L3u0armSqn*oW7hO3MS?5INP@l%?PGj|T{tVV}
z2J5WXQ&__k7NyryS=Ut7U9YFHu4#_-4AwQnv7X7gW;)gvvxbWu+griHR<L?{dn;M}
zl`LGZuVVFAIo5ZudOKJH3@PGifgQ188v1Qxt)admxB-jAA&p`C&YWD)^hy~*8t}Ch
z&%o6a)qduFL1-uI7O%*u`o*uqodtC!Ydw>-(Z`m?8m6&Gy`IS$W;)gjm~R2Afwo6b
z4>S2NQxHfHz-RApntg=syr5oTzE@Zcw3dQ;o3**k+M;n3)H|%r9mo0$R{I6>Lqk)q
z{hhhxec~c2SvNhTjE3q1OJpI5%%MJw)tSaZbp8xhXNF@vl{HLtoJa<1nBiF8!Q>rG
z!A%g<BIaMj0x<q9sOOpgc@}^ugP@kOnx(834pdOzvs&+2ZH&kubOifj#f@+q5Hf<F
zt#}b`bps75f}gE85&n!25$q+$hjhQ-Cu;tQ`k`NdKoJO9Q6l_aR1=9gkp!W=MZgFS
zq!lgF4`e-4JqJs8h5B5f{uotmsKH1uwp1I2gx6GL%SA1>z4bNz*DDb!C(M9DTzEG`
zYS<R7s9{?1T5wl{&_KO+G!ml*h;hOCS^+NoHfUvT6@<3|B3&TF@hyF~wo&so-HVIX
z@GcPKU>B~asIDaDmBb$}{%8&Lf`N0a3tL1jgpK3~7tTlocc$e_)&td21b3!wlF$q9
z&n6lI1}$><3uoL!!@YpP(Jqsd!E@DuvkwvN0tUyp@Zvy>%gV3;T=B_+p8H+*T+~8C
zwqS5pSWDNk8Pqg`I_Oz3^(m(QXvtb=h!?C1$GPx+5!GeXyo~zc-QPN<>Y9nA6<QNf
zz0bt^Oh#jZhYVV<qG7t>wARpYwmZA5@zP#x62MdbgNc7IFZ@~r(^UI53)Sm)nRwTc
zU(Cc}$ND2?e8hZo`BG*qWj^}Szh}nx%m*!k)l)WR2tMd3<EF!CIXI_}m)CK`?KPAR
z7Mf#o_$AsVXn_NBx)xkv#w)BET3s}Xglf1P?mFUf&;m!}^fk|<=1dxdYmSoyK@QF7
zT}>6LPZfet!MbP|9SA<Q0d%M?qWYScUy~s8d=RMv!L2r!4xcijT0qPNBxr~Rk>JY(
z)QO{Xde;w`<)P!NYKBl}hW@HT6c2iMWP|YZp}ZpISNc%U=EL%EAP;*$1P{9A@EtNl
zh?sblF&Sqq{y#T0R>SpR%Q%{c2RHWJ;=08J(;NK9cL%<=pal;0Vb^Gufx#iZW6g$)
zxJGAG=pDV(BdgrKVgL2!xVMHy!oFC6NIbQo+SlCI?xFM<w`}+qd-vRiLw=CKe?8HN
zBUB0V=Qt(mO+O8w1O|sC(UC=15*Rr)OcK99FcJo0MI!Zq^f&kSM^_0UNf6>dB%VG|
zT|>-kh#wm3ft_3Iotn3<AB^}A3qHi^;fM!nh$ZxIg)6ZeL@MEDD@=)ZHr_4JpB0|e
zje@m?CIN9u_?DxSs9$)nL1Q^4sejLbc@z*KK3+Xp2*S|$kSf(oe#hghQV~>L%425K
zeGy)*(pF~1XDmXSaA#T}O}+hhG~gYri#H~oUKk;VHF41q)&vFzGmpEx@3Ma7?Y;<R
z0)s=CXp4qx5GJe$$1icRqPm@$w^M)gg3$6p3mnD7GlTFYw7@}29E_;$CFZ@vA4fY<
z12Lfm4qc)SA#@204p`zC5U>OWhbhs>AWR8;a&Qux64k53e3kg)y9&Wd5abZ0KDYzK
ze1Q1j;Kpl^C3I!QD^X#_YfvTpY{e;2y~b-0CH%~RN$h_-0+XV;gH_wXENK3~@Uof(
zEEs)z5q$e<hnWTS7@o{bJi;WsUc>^6;GLzVh=}0(cRbCyq3<t(%N%%~)kkMVgttJ=
zD=ZjqCF_M7cAG`v-D0h`yUkR+eustKapd1&?d~}8UoguH7KYny^?j<nWEQ<%%4(Og
z5WW7M`MqbM`h}b%1SSdfbt67Y2%M$Yoi(@?wa9vb`(_A%c!68%tulm8SVuS)R-A*l
zs3Q|J5EmF$1Q$;@jst$?kS%tEngI+B)Z*4*Pr%?Pt^ErvUI;qGL0SD+Y9Os<Iq)R_
zEm#p+)FA|C;d?7Gi$9C%G-94cEcn$-4WxyyIW&v55Tl8};D9U+Y}S)Da@5IWlZU7_
z4Q(2ZQ58|`9ojn_quHW5A#_4GjsZe-By=$e$M6`aDGA*{!Z9>00z4IZj)doH$dw2o
zme6M;9Ahpbg#JR)XgJ0KA)-k`b7(lidm=cvzQ0imdb}b8$$Sf`1*2vnL@|8NPz#1N
zM7U>tFHs9#yduP^d~Z+-hATvPfcxH|7W899aEE<Ms0ICF5s)t5C)9$GB@yhNZz;8)
zA+%m_!3Xv0V%|q=#$Tyrk+GuMKx!Z(BrU4pQn-x4cTtUyB4ku7QEezSlo2Eq)krB)
zMwCcYqogPqfpJl7BsG%pIU}l#rN%Ns0-_o%Ma!2o@JCc*q!<~Ei>P*$y2^NyiE1~g
zn~XTNsCJjS%Lu=TY7eQ0jL%<D?J4z?F`^@?u~Mvz_rIvdNpUj9D@3)I)Jw)ISyV?z
zBV@!9M0KPzQpPZ-2<h~rWVD8&njj^}_^uY!(b8xcuU1k0M*2p^$F-=Ak;ce~dW-5<
zX{?M7XHgv|jgv8KDFR?Cr4SjhToGQrQYw-0DGVw`q%<O<XD$LXD=i>0qP8L&4{0Hh
z@w|wDJ4xw8Mo3D8qygzWB4a)QSP~*-5E-E}5#EJTCXw->2tj)yWfK`eBna9QX)}=#
zECdxNQVx;P(}AEpk+u*S;Q|p{B`KH4=sUq3Poz8|V;%wA@kH86Wc0~J!1<+ZM8*^V
zh}{z@pU8NRig2Y!$B2w~j0m@-bR78F^$ACcNQFek#Cj1xHR%MA5pNSABSI=7GA2ui
z;IT<3VSJhkCjy2eogy-VI&iZP=`_q=L&)HW66p+)@s1TCvq8EC^U;zX;9*Fl`$R_Y
z4vr*|N{EbaS-@k6^nl116cyp_mL3urFLM!qVd*a-BW5hZ144R4WPDMG@V1d26B!>n
zaKwr9gviTtc_stAx~Y^zWkk+J2oFoMsEqMq5j-4eHkHxkf#(C2=1>{2DG{&+X)cu!
zd=MexD$S!ZhG4;Yqf#=JF&#mK*PAq-%J`fE{iRY0m6vMy8Q`i?X$_UpixJ^fBCVw|
zK2t;pFG%aCjPV8$E--05mGMRs0mGIyP#NKE5xh?6M=GNm3ZN5}Hp2H>t}lR2RN6#k
z#I-<wsgw=#)jT;7qK@t}nGe28VZE4pl4E@qbD!l{U&Ev|OvdPi2!KcR-OQqUDoH|x
zB*Ex-lUJCf*XNbAzA$*usQ3SKQA9M`aIadEnb32<kcx-1Eq_VB?CZJJtAcL?QB~bk
zhuSc=VGgxq;+9Mt>KlpMM#uUQ;&#NbUQFDI9qTiw+YIXPeHL}ga;zVr#v{~QzhjH2
zv50!>ckDT8JV(9tJN61SUZLLl9ea}+Z&GjljxDCfV(P8mu@9;7A@$bp*k{!EjC$*L
z>}zU#O}+7s#Y+^{(VEnS>xlLj+#a6Ah5n4FE+^*YBnaPjXsF=(B3nuqdSrM90>hff
zh0ZtTa6!X7a|>SuqWXcDKM+5Rf}jP4@AGVFT<H9xDFX)2+QJs_{(`2hDO<X$HpM)}
z9_U5G4MIF)3!4?yJ;c0653Znv070I&g^w;#T|~``s6PgPmMstZW8#5t8ve(?1uhv3
zoTqu=nxb)pu6S}6&PY_>5c3=2k884W-0AwiocSoeqNsLs>F9>xRZ)$1iFd<iIha)9
zGK;vOYYHZfxU3~^*gbev#N{Ay!^e}T-Xboyh#UIW5WyiXWyB3Tg&n6Z%c&dA7al{@
zrGUEO+Y+5Cn3gsD3^fR!&+y*nX=mv8f(aJ<QiK3}ZGnp*)LtO?V@ejjlWRL`S!nvh
zyo`F4(HgpKETds%G+M7GGOtADz)xghiOhjNjd@LT<j-JUGyYi*oxvLG?agF$X6opb
z)i$@C$zpVV604Ks$WLPJl2})rKbwiOSrx~6&}`OFuO~ApnfdDV6lO|cevb7<DLNDl
zz7{j3I`UIlRH}|pf$zmkX^#9f7M146U%*TY9Qhf{l=07cREDE`CNpI^@-tafrXzn5
zGc9uDFJ`92|ExD!>^T1A%xk&hdaqz!D;(=9nb*pH)+1K3mil_EVqU8p`Ky`N>VMYj
zt!7Pi`8BM;8rB@|MttVLG2^+5_zH!06gBUl7Sw2T9zc+1FzQD?gPJq+*l+f@f^zE6
zthNvuy34%pGPC|bFJ|7wdL7+mI3qma5s!Y(wPgoaHQ4_(gvM^M=C{}v*aD6PzTo+f
zI9T+gfx)vLu_-iUFlwIbs9%0s?xS}9<y$ItPdl;tzAYRF5xH&{w!jl;ZQ7O*iANS6
zl+ZL!fJ7rNs@I758nK}5*#2O`cJH~*Hva3N{jNbb3tKxMyX7USErb^EZ|(UXQSBr2
zai~oaCOOoW5@9KEsP87iZpZpXB3yK=za+v-$9g&y(y7Duc~r=Atp812{-)mgSn`=m
zK69w=X2Ndf@clj}>~pLaGNI71ewqoV9qVVAaMrPYi3yh+>(`lZ-Ld|Zxjc33{|$3_
z<5(|aLYbp{y5N#7*vE@5I%oq=X~k1e5bay}&g|0ppg)IMg~QAp^#^lZK~1cQuKER>
zM9fLV4=<!WN6yD>Zv6c2zpf<c^S}&vO6`HTYRP*(6nu$sIAHL6T2ypQq=ie%nnjD=
zqzJJiQ<ft@_mKKLr2gn+;R_$yvu4+#uY->y`1vneRxRG^NMWIP60LqzwWL~2i7=@a
z#$(N>)yI=eOvxlrSG9B0=Nt_{r}6Na)5njOyx9LAT@rk@zzBIRE?OCMCV*j0#Kjrm
zvl@QpnYU<kFs1?wYuYV30VkvVCdPD;GynCi#?)R|C~ID?zEB&9X`^F8&1>rOn);*n
zi_an$HqZA(MZj1O?2t9f7w03YZOm=_(eJwSVp+`K#dBx<$7RO6U>Kb>DOexfFw-yx
z)VGNG6j6VC31G+rhQM=#(a?+P3SwTNyWnVyL6Bz(>)NDc4r_XHEAnKkSJ_p_zrGuo
zq73u3<|pg(RZXhHe2-9{Bh(*XLYP+xd%@F{aW6171Pq>}j0=H56JYSvWLyc1ZUBQP
zB;$em>s{~HotGvxEHA1DnAZV4JI7jYbbz%$Q)Uf4nhr8w$9mI)tc_kj#7u`A`G;83
zLpmSF4tv7Wp7lNXCy<JPD;O%zdBzik5gllOXFB8L04yNpD|*`Qvt_&Yh35L!`qy{$
zx!e7R)8}+)@IMzTz-T%wr8O}ejhUzpFb}Yxr7NBE+hnhIzZn1LW>=cjVOOnVTVMXy
zquZE+530^H@9~hldbcIpyJV=<p}5W5ZnIkG1Y`UiHqM%luUp~P=GOM)qBrNKG&ub%
z?`O!XE@H+aR@)JE3p>r);%#q5-Hhj1Z5?%c&Z<3U4RjRj9h2X&dWb$DtOC9BJbyfz
zqWUv2|4c0Sf=1{8CT30g$6>zj<d>Ly%CZk$>#bZ{Rl{2Y(0tdv)llt3b&BhhYUmnS
zqbsh<Nj3egUO-$6NHy%qnzHYDnN&l6TU1{W*H@&P{#MVRt{JqN{#M^cUANI{_?kf2
z07h#4_Xp^q;XgfqnOgtd0Zu?vKQe#hhvObKa7bV4MUpV8e?llws=D9&?Q)NM@e|o6
TF<lLxXqOC-uKHhN^uPWezqqcr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/dual_reduce_common.hpp.746AA107CF079C11.idx b/.cache/clangd/index/dual_reduce_common.hpp.746AA107CF079C11.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1fa3488651d58531a435691b37a02cdf47cf3ae7
GIT binary patch
literal 7180
zcmb_hcU)A*x4&n&;x1g4WiLxv+Ac-PvMNmwR1g6hf+#kEM2QVTiGT<iBZ>u3Llgah
z*pMiK1q4wv#)7|SP(h3u+cU92G=P%$q7l*GEHZoVdw)Ft@a<=rx!;{~Pn&bj9Hxv9
z3Zl&^YRtUxY4P(@=F${J8Q{O<l%yDu7DWZ<QdC;$sp;{dcZZ7q4((o_u-fsut;!;x
zVs-WkzdP&bueS}l<TW<hqsHxS%v16E6$w#y=n1XU{bKEQ?))pcz5P&o*`xcXtnQ?X
z3g&0D%yIH6l{%kVQ}Nr2-yYmwQGKrEiD_-eonkxJnF`~zoqmndl&187-h#DtuM2+q
zsWENeQeT}-ueGj6&nnw#=H_^%@yF7lwy-b$y!P@{quF;;%pJW4s1gP&oL-T0@zo_S
zf1QbL)!8|fox8lF1GQ67gnvY*zW<@}np?LVnBtpp-*hxDe?2yTT6Af4+{Puf?)oPy
zl=p2GuRi+lwb>foVUEZ0p6v?hmX(LSNt1rH(KqMW%{|rHcU>$>p)kBL{$TgieY;X!
zPwVcT`*hi@gUw;byhGlHq?fi`dgGP8qf@DWIPBoj+f7wR?-n#RPSW4Bw1ggc?{R(P
zl`YDy2TQh3>8@K=p8DmI_3Pgs9CGK?H&2t^z3H6O``rbfZB5S~Tq*i|y|3!clxpM0
z_W}pk9Cn;1?+i9MesT2HoG9aTot9NEPwLhi9B>`o{%3mEf6oj&1y#jWB`QUVo$J_T
zCGGpXhs{<p24hFOz3$WC9hUy|N@Mb??PX)**CYh$C=|9~rg`V9vrJv=8cJ5}4L*{R
zutpI%zrnR*P<+IOGZvqnx5<PnIoEZL?VV<yf3bJwGmj+$-RyeC?|7dmSENS-_`N>A
zA}YJ0+FF_#V*N{SO6%dnmp_a8!P)AfwcFk5eGh}&MkbDpsV`aNys5}&*iZMpM4=g>
z4vTt!TwIp)_{9D#jw4q-nwjldpD^Xbt>w92)tfJPZk+w2+}Q0=Sg@yYR>-)^|Lj<z
zzaCDu4-8PwA3yhHPS2PZ|6Y;R*knI%!(aDC1`KXdnXEZkq`f76w9A05%`+u+r=qLB
zJ-jo;Z|jRi<I<Yu1;3rU^0aS6jgi3~w<}4W1y`Oe%#zqX_ROo9#Kexgnh+mWIWM>>
zKW_3rXO2eI{QbCYP6xX_a9V!AjdZ7Z^iqdkaxz_>*vyf5>{&b4MQ*=&+vwv-y~E$G
zjabQC4bK)_x%zoxt{F;BkB@$eC~Cll;}houxj)oJ11XBF1MNC+=<g^YK0Wv&7EC>3
zRZ?+P7wO<BwV+)K_Wd21Vxs*<91#_JS{DiMlzPyv2StBJrdSRi7vB8GuiuD}tsce7
zP2`HchQbTfA20VZ<&DbvAtL02hZVXCrmrExW1PZtx3x#v=~4Fh1yZS0(bx2CK3Ja$
zoLyWw=+9kaMM#NPTL>%+`x>%(_POb~GUHhf1|kNpmTAk_zJ`#NdOojpLeGC;>b3E3
z4QSVZb$>^u<ZA6VyZmdyOI<VoPdN_S$MKX89hp*WZ57wNzdJ~bRG5aL!l7(mLufcq
z9wplo6fi&+VX?B+pk0kc^r0g=9382ECZo&|BIKf$fq_EO*N`1fgwvK2Md|yE5QB#u
zXa|G7h7A8`xO4IHm`7iVkd1n^NTJ`?kl|}5;G5se?0;b>NBsU&CRN_oP(m6}Lk*~(
z58X;FYuP9hAqie>BeXH;Ysl(5gO{v*dFH}h5fZ64H&jUb8ZvzEn~g90?mn2Ji?EYr
z4}ta}koI?kxhJFN(f6~2X$<9nr<?`)EZFvUlu!$(L1L<VOtRn1?S|PRGz7oIPH1P?
z*N~TVM=Z!rj=aA|gs>T~wgOwWuOY*)#?Mx5$o6N&$Qcih5{zQ|8Zt~J&0mnrO0lVZ
zetLAXo1;Er*g9lghm;!1*rUk$D2L&BaERa8fTuLgZUWUN7)(S+vmM~j0ghxL&31ya
z6P!8R1uk9SO0brT?*5&Zqc;v<*c)Ja11!k^!`=k5n;_?K2dFx@g$&-<wG%wK@h%wE
z1!D--N?PtSys^y}vr>pW3egBMz_6Q<&t~LLFgfK~I^=Bk9Xg6)*amQJfDmHe>V?c^
zG?*;Zu6{LhrT&)(K3N9Z&8Pq9aH`e;hHV4;Ht^EC;eJlu!k3bvG{XiP2Fr*R88&o8
zs6U6M8&2nq%^Ec;fQ$`XlAJatpl7NU!!FQWU_gFm*ciJQC5K`?V@DFCw@~I8Q)PYz
zXtoLkSD_F>3~_Z>P9OYSznHL`j)C$R1QSQcu=!|MKANFf7;;(>-;rs9*?R)^PdN6(
z4VLk4MdMm9bx#0)2ay$OIp8l3w}W9jUfX|8j8?BI_#&h0GeolmAS(bf&iV>LUI+?;
z2}z(sJZ?YHX2YH@-Da0t{82+Eu31;7&m|LY$*@k0lazQJhE+039&%=!c}T^mcqmjH
zYD#3yu#?%z<{XM+;-nmkXX1G%fl1(@L?%&#ESK)`y8yOr*h(^xI0G3HF){2aBwoeG
ztC4s$A7>(QCWoy=tA=L_E<eB<O|qKg%uy0%6~;p|tY+}gLgR(DWTxUo(xQ%x5$o`Y
zm?WEIL4Ky$Fj*K6&5+IDp@ou#8dhu{H(bzhj|%^&TAZ&{OcX>PcK)aZ3<=%y%xJOl
zz(;p<L5uZsQP~J=_myB+2~uK1?jz6Z+)9ry6f>+`EH@$sfnzYO&>^82A4&I6Z^AIc
zrYq9z$p#F&OtFlImMfO?(ASEud1yDd?*?yf<yo*h3p_7;4(!hHaU<9@^6`1FJI}{W
zVAKTWoIx~$xEYK%d<*DXpu^$YK;P!D`@SqTa_yZJSa_u%E(I%&uu3qh1UrI>7_e_p
zrzgBqB5)FNrzB1gCnPfRO3O;voZ(c4y+SR7)<OjFFW#2jO<!uy?be}LSBa}3NgXg!
zxY}^YP2$EI^N@J(ke9@ZhkPVHJcN&YF%l7u)lN=Sg~WV3s0umov2PU`O|XArh04;m
z;nRJ-|E`tR-CK7|#q1ZukYX4^L`kwaEJ}4M$2r*`w;*34Jcf<5iqtr8lHuVc>hzA2
zevp4q2ycmVq!&R!?%jpfdknT?&DOy98VDzO9uE8vS_cs%Dh7SOwkxh<!A@O<T>}Pd
zz&Kr|J+bHa@Op+_2hw$5#^Ee5%>pY97lW`E^a)=KTLQuoKHdR(JAmcJcY)q6KHd#_
zyZN{rq~&0Cecr@D_6;Sq40|7h_d%a%hGF+2;a;S{V}CC7*%lF)j5*9lOg`cneF0($
zI6R^FbZ&EIhaR4I3<SqOj|2^M{1-HU9*O^xJPQ(yR2%cK%Bo>}HH4G2b*eD%RQ!DL
zw|Iw0lSoS<V}{*<ggZEj!b2qXzvqt_fZeNHDA(s4Nht_R`GRQz{U(rdj5UL>nPW8k
zUett}CBeR*oOTd9K{_8Yeofk=PM^tOOj2@c=<9ucI7W8>-2r+WS)D+4@^Kf?T^x>^
z^4%SIH+}Y#$&4$EwhwqPVJXfKGr%qbTsR9`1$L|WI0syEU?ewQ1d1Y1aTc}(6k9mF
zIBismdDFFCEt-{E%Wa50XtoKQnm8v)v&~@A4AzA4#Yb;X`r9&d`zJG&TzT#Lf~+|!
zusuz-n5-aXlTx&OM$_}^`j3V-a%Azb|8Otkd_@C@8o-Gdt~vu|3OT2i*7)G@GP?B&
zmgq$odJ#MbYYf{1!XD6gM}~b1!nb_f3&LJLeh0#L9R7OI^dbAdZ+(P0+6}h5K@-?k
z9}JG4w!QTNwxV2w!z?e$&qI<tq~IvXL+*J<lf`eCptYtm_q~@k!@dT|YcS;|{sWSK
z__zl!^)hb!Enw<-xEI)7K7I%6I}UI9{m9~}qyNbLf44HC1e}8Q---LxRP<`j?A7|_
zH?XkEkw-ZfV@Qm{dR5<)$w_3`D8ne8bBUHlTM?Vt9XFAl5Vr2&KpZ6<T%E`cH0v&L
z=fyrxi6;+vOT2l=XSk0Kx2^;+CHyFP0ErH8QL^m0)ba5%n>SdUFTnT(KfySH6h{y*
zaj!+HTBNz{m1oT?ezN2TbIfx)IJCoXqOS6nMtRbSJ$vzvvxd*oTt@Ik#Vkg{CBw$5
zVqFN&6)(o^j>_DB2k%u3_Ql+}CT<zaufK+%@odLHbPOCh2cM6O^Ev7cJGM>xtUFAd
zR8}Eb71HEQYV1;l+z7Qbk+)pSHtI_TGOVkfs{s)k!@B9Y@sNj}2M>AadGU~so(~WG
z3+ep}8IbK6{5oM3(zx2Xl|vrM&63=3Y8vSmsk!)(k_hHgT@T^x#983XOS%#S37Vwq
zcu-N=A-{H<^|6s`q!AHigWl~x!LuVuEXp8*AkBrOLG-j`-B+;`j%)vSHGpJz*v9#>
z<Lu4v+^FUsVm{eZHntu%wpi*t784@%kOpyn&dAkh3_+yEqZS7)@NnW#3Y0u?#0%mz
z;`qtbH>0p}<wY!xP+6!M5l7S68KWv1rp>_Oi1dyeMSgC!S)R3e)Kq=7)`z;fb6R(o
zxm$9`Q|8G--ZF0<@)_+D$jvN4h9$^^)A|8qbO32QR?GUWJu5f=`zEGBDv}zKys&j#
z?xhV!E}4C@P+Qy1C7$rT-Y%qBC#jPW(IL$$rAi)hmOAs0N~+?aIBA^5hG;fk8qY%s
z(gYq#lqT{}CFoazn8an$<2*!NdL#N|*q0!9$@vyig~QX-B{)vMv&6GBF5;%;$pN>O
z^2;AJSyHr9TMMq(al5GmDxS3@3KBKeLJE1<%0E=`2&KPzL_S<q`vhT49Rk}!U{7W~
zh)Vg*%RgWDlXdmzgsF>jw|yq|O#d^K)Bmd`MT`m4_@RtQXl(m6;M=v0Pu5{mF9k&@
zI1s7-QF!R@?aG#&0)|bsPvs>`OBG8s@ta}O6lpwk8R*O0HIZSj0)3T_TY+xn<7+@)
z<Kv%!{+Yw1wuv>XE^KnTh!w=L<n)mWC!VA(Z*r4j1TmUPe;w}J+PeH)3f6@2P-7p$
zJlO+FMP2(OYkRZK%+@+mlZuVS&&qEYnT7B2=Amu`;o4~2T=hoDNna$N#yw(HZAoj&
zsm?%sD@(bBwZhOyYHVU^W-g<w$qOk8|FeyJ49NpU$+$-o?$MZgH02)U+@t0eNKr=I
Xuj+bj|HpznP?Q<>XvIBRa*zK3QlCvP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/dual_reduce_multiblock.cpp.C5E150B4E9B1E7E8.idx b/.cache/clangd/index/dual_reduce_multiblock.cpp.C5E150B4E9B1E7E8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..40f1298f14ae9b80a15b264b97625155fa1a7e4a
GIT binary patch
literal 2770
zcmZux2~?747ydvrMQ}j{LD3>tL`gEmR761pH#9{N+;Xc_Y7$M;vWBL`WwLCu#U*tz
zP201~N*miP&1oi6TP@qmEJtmoWj^V=@O#eBIb7~@pUb`Xd!PG$3j{pg4KxICk_EZ4
zl&pm)2!agZuRJSri5?6->OxTNzVneX;ZvT$c-5==JE=OMc~OP>Y(D)p!f^PKfBJEc
zmZ*M0PCT+T{!#SctXcO(RP!A*xjB^;i{f(LDvl|fqeKHgCqMR+OQK4v^r{oK6?=De
zy*crTzH9h%)~e&Slam?V6(i*G?oz!8l<n1g#S7yTZ`x>QDq||wd|ii+|9+PC;<VeH
zj~Ofv^0sQ4chB?TFJ1i)#-5KidbUxnPps`B5WMC*{dMlOy%0_z8NKUOO~0s!n_+V1
zrvuptJgLh%uXqo=?2j@Bhpe(m^w+-hv{N4*_enSGtyMf&cXv!!j~(DUoLnFZ>O^K=
zla5sM^qBmUow2?7({txM>ciE;=YDRgYbv5uA6_gYOK*qtUwdgjP`&Tn<)O}vLd)nC
zP2mIEj^eHF9lQj&2ioLpC(FW)f2gimTz&~%X0h0+#X^&-utsAoD6iHdvI@&OBqJro
z!mLR?i<h#+#dl1=ebK2^^U{t~J3c|<&Mm)r-aMWeJ$R{Oew(Y4=G{+A4V!5Ujmp=@
zObR$7_t58%4XFK!md9_lcR4eYBeb{YM%w>bHXP${yU2LrC$#g~GfS6mkrtIKC?h9&
zB=$XbnIn?=>%2>G<hRCd?4X9`&Hi9|`^YANATwaAS%Jo-XO5TccklnWCB3!g>GIMI
zS5`K9d$_0e9$O{yu<u{J_4J6zy#C*^^>=!$J?(z>;X5q8+P8%-T3`7aG;irz<7>V5
zyAx8XMuT0eDmE>a1e5RRS{m%kF!?0AE+7rvaIcN+Xg8AhI6jNG%J_vH{Kn&)yb|Ag
zXn-@AT$F5I{=T(&*V}OmAD4by*fVMFqx^_DX<yP<ZFS#c>K01M3NEB%?5_~xPLA$%
zWYZex*BzTnit1-*$~DuIIA6HDdGhN<)E(Z#_LHl=3OBny(35@|9etU=?c7ey+2b%S
zy=YO~7tC0N`l$W6H*Cql6-9TgPl<4oRXS^eE~g;9n(~{vuJTXHE;cxy$@%?gyuW{~
z#)Z(GZjB~2Oru1k;rmP?{cu8I`0H0kox{Ja-G&>dKj$4Te;lB*@AlR~R`>4J40%Nm
zeV<+U{kloi+olm*bdSQprTfT<$pe=QQcdI|l7kH$eF??Vqg^$W`ZX`URFv&rR&=6v
z1Mc&M4dadD0ZE;0#l_ZN9}fMk7gpf!9)C6%oBP1Vsg0svnu$PspOIdKYjtkj)%D=Z
z`S#frw`k@G^A}aimSDbHJ+-Gc5A6v|`^M<$ue#ZWg=CB;#C$h8DNfgC#JXOC=gO+n
z8A^1st-4p>?pbeXi#fOPZ?iGYThXNUWeK61SvCeY&Fmt}+m}~ruMqxaTDa^_PmTBs
z^p+-p?`QUr`DvRweEPZ%-78mIS-OzjKV(^z&wD3dDNA%kKv4K7YEG3Ts}c)EK#-VX
zMVYTO@Owk~Ssat!S4;TynJzLIrb${PL!|+9dSSx1le512VWCKvChL)LN(1OkJ@%*g
zb~`R&pj4P<qL|uB1L!henw^K$T_-FQ2GbUL79^zsbWPD$mXt-#(?u?TX;UpzL!|+9
zora>@X#$gtg+wqtQ-7wJ(g3>6n#ULkAG@xL6vDKXmK9!U0Nu-O)Up}0J*bNeg=vBo
z!9ZyM{o>I3H=G?_LS1AyOcS+;c%=b!W=g7j9{f%<)%wRk5|`GFz=$1~4&k6D)1}7Y
zuf`TS91h?H_<>^x*xg>*z3wd*4ikw)3(zB2=tj8aR~`n3DPoEgC<1<w=xK^z#EwkI
z5YQ?1_sCKbA^HjYUzRq@1awN(A6Y{mmJGpRE25Pf$OdYn(9=}Jh;5iQ9FT+&^BH_E
zPzfWJFeEAj^o9#ei4BeeBerAOd8!Kt7y?&y0VzYOLb;t!p5>u>Zy_;a4wi#cCyB5k
z725d0&)mJIgal7dAQNT-sX#IUJ#C9hH5$xrK3X~kXVaLp2rva$E-m!5rNW5qnf6?e
z1W#egu>yr*wHN^`H4>SM{)dZU#Hcv1nwscoD+%jrNu&Z@VMHUksgXa!Uv-ceaWo@Z
zg}^GqQcvx%`dFQqPO4d(a@#aRWi>qbIbA(nRpC=!vk*7%>pbCb5I0D*YiA=+7J+&z
z8xSZl3(q2ghXN0-fu437Fk(N3pD#$lh@%)$Ds;uTYI**Wr|EEmKyIL_0Z;{Ag{j-4
zp4pmd?F-sq#5@MiQ(afg5UbFg!}nIy{7^ZB8xjtZ2As0jk9AxfgxdvkgH`SB7nFQp
z4?bA{hp9}eKWH%QlZ<+}NhcTHCqF|!XHW!w0>da30usQ-J9U%QquMfU`JmD$N6r>k
z7#G3~Lb;(rP-#?>S}2csSPO@RT%jsFA&7M>F7!2m!)$#v31t8K?}2vu{hed*I21Hx
zHn_qt@}fin1ZAdVESwrbo|z=CpK_w6%#Tkpa3CmJ<!Xpas3F?iwGjgB!i2>5<fQ1h
zScx<y3QAI?K#6KMPVL61U8&lYsNHzA8>@De<NbgM>S&TIIb#tF15$!CP?}mZv~_eb
zSOa}Mz9HAw#t_ZN5sZvYiA;A7PcN1?+sVXX{%lt}%l{X}o0*em&Z1aZ+uBoU^f}IR
Z=eb$9kQ^Bt6Fg6lg7gdV@kazg{{gvL<Z=K2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/dual_reduce_threadwise.cpp.7BA56606324C3914.idx b/.cache/clangd/index/dual_reduce_threadwise.cpp.7BA56606324C3914.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0e3349227d0f3e8a38734b984eedf59c1dc4cc72
GIT binary patch
literal 2634
zcmZux2{@bS7XC#9S!^W{8q~g5LL(|PA|YteSbn=A_Nt05L9Z@TQ|TzZXlGhWmFZNg
zMW>h=Yik>Ahi*!1tJ1}#o3w*YTlbfF{^WX|`#(?4bKdiw@4Vl4&i94u;pWCf1Ar^f
zLy!`?CkhDwfGYG$-@{LogTN>T00gCtfhj%)_9_!w9{QN{Ax+zAy5;qLCVW!Tzb!1N
z5jwRb_iPO9F#NfreR>?HI91D?I$7CW-Y2?S(`Tk>p_v|Vfy6;Y*uC2OX=zpgO>O>s
zy|umh#ut8ggm+)O{E>pezaB2%JUshlW{I(TU#6bGo7~boc%-EN(~IA>Bz$;NH^<N3
z?`XC0<#iER@%DmfXT(59*qBf1yY|zBNRI`lXQH71<-3Au?1J&3qZNUzdy_vJ1>GyD
zKoABRtUH-8&8*bzA~rXKADiQ?7I}UpbK6zC&X>5p-C5jv`Fm&EXB%9!6LSJDA5RRb
zP|BEV6D{x0$zG=&JJ*_#SAKla*CAEU!mLIgQ*99v+J|rOqf#~>qngHNU0ZrT_0at^
zX|7nw@CsXzYEEVh6!tsCx`!59UI|dg6$QCGkUi1;xatt6o^}4Ft6`hXA?xjGL33qk
zAB0M}{e9KgLbqIJ;p+n<Omz}F=FwDnP6Dc^`NwIDNpgv1P`rKbh0e?gQspUjT&JJY
zzFP<HIW{5c80Ie36;?}|U+}R?L=A*dc>ubR=P-<J(C?8gW#(D$3RypMTED88cWk75
zq|pJlb?EVSW>zqU`}XmMdUWvwnvVKt5zQ?u<<5LRi6CcqrFq6x(`Q0zUkk>n&Zm|C
zXK=tfc#mb|khkZ)$eA};>}~rJ$DjVXNl;QXH$D+9W6t}*Kj3vR+IJeyIlTVQH`u_x
zh?8-mC&Juch8~gIcAu!{)%=v|<^4bf93Z|yd@H+$PTFvB^XU!9q>#y}GEqZr#E{NY
zV?RT$n|nH1GW&PEQ)Lxdsj}O1l-u0ps(R~QUQhl!z}EW3X!ccTFaD{}$R)Jn38K63
zbK%&7+tq?>e;cr#O6~UH`P=?TKhIrGFF!IkyfR<>@J;<D-E;Z*ZgyKQ@n1DZcWBIC
z6O0>nx$702JXc!$ebU8*aBis4;gyp2eg0v__qu+ccMYpKWN^Pp?qct=3B|5#QdwU9
z5;J(BIxzP+GgpyyV=%V$Guo?^>vQbtnFaaM;`$XEdwvfR^n7+Sf9zA*z<kk}5?<yM
z>&$a%kI9bIf1Udyle}1)bJjdVw&{j!i>3L(+x-KRX35K*G{+><n0LvqE9iB0R5~}z
za#n<|Tg+XEnK>{UGOw!;@tg2Y?b$A$W`O_q*Q;ipGmmFntQSR#-}3iRukSuITY9+Z
zLdM@kc^9i2j%!<ATE0W~JxF<H4K(y0jh`Riksz=_Ix)QE^L8HCT6L>um!9*rulxfZ
zy55=s<r4eX7_Dl%_{q+S^<R@2HmMasQQvGutonnOdpQTHUJ{oUqDHP6E50~8VV9KO
z=`uV}GxF1sm&s9#C(k26l9LEX&q#?T$N+%tI)|3r#G-H<upIzE9VHztv4Q?C#2<6C
zM1A_sT_z3)fM}8uNlk2EIw34;xQcLM27}lJ(OM`iRk4BT+(ENOj_IkJSRfXnZIQML
zVgu8a$G<v<I+HRn2rfixqBK>-2Buq#+$it-KK5e}J`k;s(#MMpOy6f*Mj4NqRAUgn
z5KTZ4RKy0RU)8*wb3NtcgF*N~G!aF_iw#V#{5v3_au8>M1N<RcPgze}Y+xGT$Ht{U
zk_7-w)Mu5Dh@=h~ENDhE^Mm&Uw+)3}(=IG%LNj5(Nm$TX)7enc1dvT(J4=zTrtkl7
zAO)0gG`mWRq03DhAx$=m&4!o2vp}NP%m51-(~Rj-O%{bkg_mH#KuVwld7kMg82D};
z2h}AKi8^qH!fyo@Xc2;^ps<62g9f}7o-HJ|I#{?fSkRDW=qi<SrMNo5OR!)7B|w4%
zT~jle$oqdHpg!Q7)uc(jIA00ME*)+(+tRfRiR%;f?cut|-m}*`4CZM;VFHO@0M8n(
zA%k9XkqTrlY2TgP^%2T8r<wc1Md14-hhB59pwsIT&EeBS$dYI;MNA4)asn*qPw|(a
zD}oj|-;_!>NQF!%OQziDv_a-J{yRDpc4xUuruDl&QpZav<B|*#big|h;hn>KlSQw&
z%vg{?VKCt&EXbvBC8$^ZTv}G*bTTC1#&VMgz&(NHvwHcZ^D&?qFku-i=t6O!OY8bk
z{3PgM?a02$e|qe}^6-O`o?7qC`PkVv3dwO;9EsdmcHVo&=oA+Uo6=0(VSyJeJmf^K
zq5v8)LzO{>i$Et(<w_9TT?Bf~8^nS}G$RgN>4R%JBdAPW2nl$xynNtFA41}U>0uL9
zP}q~@DG4ungj@X}WUhn4j>?W&a5ns`A*IzfEp3Vg%|bFUe(cVuRVJMulYVj430yVe
zPZmt=Z#|h$yV8u?5g8Q`6CD&D790{92*kh{01zZ~qouC6;y*A%8V#4ap;9+k>PAT2
zFsZvk>PAZ4m=xa5cnEsNZk0vKq2v`5F<6|6vJyv?#WYq$bJPgy)HR7T2S>V-^A?7+
zhQ($Bs;Tb(i{iD{lk_&|8ycCInOknOA=_@U*Rj*GqPS|{-8^^*wzrGBj2G}<ip|LQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/dynamic_buffer.hpp.9B076D16FB6FDAA3.idx b/.cache/clangd/index/dynamic_buffer.hpp.9B076D16FB6FDAA3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..248bc9f465e9d992e1e2debbdf9e21ed75733106
GIT binary patch
literal 10820
zcmeI2dstM}`p0Lzd%S=fMdZGjfZ^&)qc~20Ff*Va25H_-c}btrv@l6i%Pft;6ckh(
zEKox@fPrv+B2aRX1j`IVj+ZR8qg0Sw)Dt42prD;K;9|Xh`H#=<KYes<KYZ8TdtKiB
zu6MnQS&<PDdX-ETwj^@h+e_bFB$vr#Bk8|2@2*}p6l5~_5SeUU`r&^pTz=Wl_LtkS
zo1eMAQyl+Qvd?Uo_e#WTmtVLt(?7o4w9<LxNA5rV`pr*&8-HATd4J@$!iX4;g}<*q
zUC~nO-SX!9E0aEWX+mf0vbaA#FSL6w_t3@GyT`kn43+EWZ2BR4@aC^$gqxW#$HwmZ
zS?{qYG7J9*E7nH*wE4F<+m|v||8;Bm!Lt^_JA#95TP@k%HC{WnHYR2NXPt-tnyhch
zy)<a|^pVZG(yx8|!_EI3&C0&<xnci*|I?=Iku`5K-m5H&eKGS~R^c1(78dMi`7Y*c
zR$Ilr<1751zZ8>l<>Zdf()||C&HG!n`<%nneJLOKMYV6ae5mT!%GIM?I^J0nGkfou
z&wbLSb<Svi@w<k$-k;^Quh0C|?_XW|2b<jMdrmHC81dz^z1q-qSH`>^8EL0`ck|Cv
zmb~+G;E|+%J9b@fJ`vzGXyN1}|DvRo=T}{u{nM?j?_Ft?O<G^@meZ!sZ$+DyzcMit
z-t*hEGAa8bmjg+~e@6P1*xMxO&jds)s=75J&Y|+k-d9^4qjzq1Ft$X@`*V?(>FsmN
zKC3vp;DTNA2I)|Mw0w%J;M1K--Gl>)2j~A=4658*rP%Z7c-OtV7i~Z5{_>jkMN1L`
ztzTQV<n=$-&i+f|?4*B~F>B-8<ZbJdf<9U}HA`D~ZMxgnnd?ry+wuOfEgLl@$K$sa
zw=Vwji>a<@A+a{6=2u5*(|ZTK)tfeM%?EEUPNDW7%*v~@cx{EhHMrZ!B&}LIzORrC
zr8YzV42${2q}kkZ+8RVE2l)i4`U)nGSZ<Ma`n?Z!Tf<N)CkToJ!TZsLRk6-W3IFZG
zQzva;1g(-MjL8%9k1ni=v1#=y*MnAs3<rB!#l_CWx36GwsF&~g1!HDS9R?m$cCvID
z-d8YrV&bWww%^z66J@TpGO0-LC=&c1U1S!6KsHK7Br@@bl|?m!Umal$*7O-QL{UTB
z9$naTCXLJ(+4E75?QnR8R#6OAxbziFj(k3GbNaCnH*LU*%9<ceWM9GLSLeBJ?L9tj
zur;_*S?8tG^c74#uwm6_zx{Uj3u_oc<qM#=0N#%-tjbp>Vtvn^%gVQhVYG^;(sOKI
z!Q`s3_a+p(FZ`xoc2&BLmCOZ`f6TsOdqQ$qWepbe?joWnBF{d$uqvl4!oCejDjH$~
z!)cX#VNAYYcywV^PJjO9$yS$@o2|iyR*6-_dh``cKEL<omm^krmktI?I@AtB9Blgv
znT4ECA47-v+GI`o#$Dc*hl7&dFAb4~_Z95^j_VDHO(z=i`lqT&kW~qzr2dP{VzAIZ
zRjZf2vBu7eTPA@%i)8N>R<949aJ)HWkSL{sDiy}DsTHL((4|2z=R!~wV*L>44q?3+
z{F)(vO${@3V%cZn?0sU~wKuc&Tq+$bN_U9y4vFB;zfX+!G4~K-59YmK*b9?+{bDc_
zV=e(h31_Rr2daGsy>oPgD7ngA#s2S4CKDx3xhEoxT!Tm}*CNu%b%=uGL5O1Iv51O@
zyoeyKA@Ul`7eIaike(pO69mM0f;<m%z97%XTqMYfghxKa+=rur>?r3Ejz3TGJOA5q
zxm-#kA!#I(*_~X<Ca+|Z1&6<Q>$YiMR-|0oPXhLnK&}@MLjf@|X8woHS-<mPv&Os*
z75ZHc6YLiq4%t(b6+uL4mmuFI__A?!yjVQ6H?K@<BbTBbq8*tv$fbC&jR!k6i*hLe
z923Ba^A51xf%Rl?OvZXD*rsAV4II<3o&mNQSkDB<OswaEZ7$aHz%dW&g<xBV^+Vuz
z2<xR_TZ;8Ea4f@m71&l`y&4>=v0ew(bug0k*ZF$gv-?QsemZ_vL4Fkm#!ngYbwx+u
zS<9#6_t^2%Ea?qaz(?N^xl~TZmy_q1;~ryQw<_YuR}*ci{RKn^V*48(4>~sfX8SWF
zfRXbK(CvWdI46TK8A3Rxf-V*7X<$sldItDsKmf0w389%Vk#jEi=VCn%Li4a*2)aW2
z{6k<ogr8pu{-yZ-G6*fh_g8^`71paEv>NMmpsE8kZ_h&rcnE>aYs;k^VSJ9DW1b**
z<(4<UytZZ4P*Kv1)=XflwkSoZqJr3<ic%?2my%#M_9C4N|8f$+xt^%%Nf3J}wRJM~
zDj675QECKrBLuTXh>}TAn}lH22>sA+UT8RQV65565XCN{X8uY)e9+5lhu!d^!VHHQ
zlURFL0Ayj=7x*w=B}#6R+ZeX+i;}OUuYyGhqBL7Odop`~D9zW-M|6znkCExTJB?&q
zBMIl->4tIL5YC&QDU8b$!ddfKz(t*>Ioe{)VSyI?aetIml%cL(dc;)9o=b)P$co*;
zLMpjsUns?XEJ~v-N4xQ!%%3uUA@9i=#hUSq!e_iRw&KedSJUyTBtt8S7wgjGub1yi
zOAa{<qLe|LGl&-(VNuE?&Y7GeEZ|o~QhpVELIV*Shzo0M#Pt7O-06_M-d2>N9iyLR
zZxpG&w~L3-teQx_H>U*9aHh_F2iB>xPsTcR_NiE>&OQz6)Y)fXojUtWtW#&7i*@Sk
z^RP~xeIeGVvp<A&>g-FgPMv)j)~U0v!a8;K)mW#_z7D*2e;>j#4?)XqB}Z`15ymiE
ziHum~K5sXa(1|D~&zF-I*bJG!nTc1q(XlUah-F{g%xxjcF3OXSekrjp<<l%m^~A28
zc=Fa<CC*m~nw>_lZv-#iI+I{;61-UJrcDoUI5~OuOgj6Y5dTle1ZF04dv=d&uKZN@
z#1Y8+^Od8O|5)Ke`+tiVZ{hGS5e$huyjw8!+f(JoQeUMNcYtaKj!Tn4m(1f*wv(W7
zuX!862I|#y=__A5`puuTD<+~b5gqHw64TcB4@ZCU<I{!Xi58Bf(<UB`E6!a^!|x_C
zvx&^%HrGai+DItpn<VHa=5}Ii=Y}py9mLpyxsw<>G2bP|yO`r3FpjV2qO=hLH)4(l
zRXq6d_a}fZ0np;K6%1P;jO!*am@w}F!ye38V94UUtYpaFF};2ps9`+>>xbaXKBkym
zi+#cyqTZ)FpETl_Mm(9tnb|juTcDXIrjcn}&m!JgM9ZRhkw*0{*~FJ~4sprh{PtY=
z+SSoz7ikOTILvvGIfE4=duohx(#)=9k$jOK3)Dra2IMs`!0(sWR&A)5*-h_0N95;-
z#45hu;#ogh|KW5hT!m+^!i!8;7yMT3$L@6sDr_Uc+xT|kgV%?7=TEACD2S5%Wc!!c
zrzc7Wg~<nnm)HVGU7Zk8E=*_J*-z5~6?G)`WBR;JBzO}EX9KY*=i?8j9XL@VH-EEE
zCZc45C^duF40dd&MX3eE7R;?+)e81ZPiUX)GU3cNXZnU#7}v@-t%;$BALf+xtT4A&
zV;_cFtT`xz;THRp3kKHWq$pw2U&((5+KH0ZUu$4VAW_o!=>nJo5~Uz@ke1DcC>d-F
zPMkt)LJ&>0pQ_+mtY_>%$RkR{L{m(3d|fCZni9-4#Jz_2a`UYvHnn76gC$CJ#HJ4O
z@5KIhqTux}fX@XO!-h(fE`!ZwaN=AKHuaeAfc+g%a6LirN#F~bC?yIuiGmYbLa2=h
zHrp_#3-;-Pg6Z2^OIFREcq94${VPe|r*7Qv-FpE~ykw=fTa30n{>(ILOgVy>BiOU9
z<Xn8v)%5maOKPqepvr*pe2g<emk9>WrJyRsdKu`-xX#9w288{oE@qeix}k|$bt27y
zzIqD@-on3>gZEw*cQ&o^prhDDj9p|3e^oCr_F{fOj1MqpgCQGGR;B_B6_}5J;Rt6I
zb<)+<oMK}QJAC8A!w!iZ1y6UckKMhNs@-?GovPQ+`Rf9)3*6aEm%s5xUASd!gE@I+
z<?6sZElploc_PwSX%K0xw1{+8Iz&NML5O0lVi6S)t0IE9hFH~Lz5rGi0O<*WRf2#x
zPq50voG)1AV=fXbiiA<T4@U)yqns=C?jx^mxx1Z?!9R)ZKgl!9seg0wmw5++Z#O-)
zp~u<~+t1Q3(44?#U*PvYrQd)BM!HFpG6XS0aA#hr=BIO~PF?Rj_QZaVtw84eYhQkx
zy_zM)snDMlXToXUlwQUC?;cREbekA&<KVuV7`rjwBgT7}Q^1hIgMU%V2SYyQ0x%SC
zKD}i8cJ0!pK-&3>M1GNInej2(qD?br<5|Pb=)E`0n(mVN)YKkpYUeJzH%IOE!^b)r
zyyy*4JU|K2haj2}He)8&=1iZ1vJuHdo6PgmqO^|~_K_FZw1`qI(btkFzFO50eI4co
zqG%uk*(>wjBA|(&Gl4eZ(?$leccOHY_}s+YPSovW0)PH3qP~T>gQz<&-zMtYm^+EO
z6Z0LSzJvKLQQyVfMburGyNS9R^F5-zhxtBH-^biT)IFGciMki_1EPL_IS#zya3HY}
zyf$Lq3aYIT$oroNszl5tP?<2NfGP#^9#HMUycbk^F=v4)3v)K8vN7j_Dj#zJs0uI_
zgQ^&F38+diSAePl^AS)T!F&nETmn5?M$N&i{xU>yt_OWR=4P;O22ZZHfPD+*R&Z#=
z`H(J9c7Y$SpD5@Pg(&7*&B3jH8|Dl_nIZTwo%x$VmVF0(Ea68NdE~k2B99^|Q7RFJ
zl?aY}>v2pNc1&<&TaQbf>9eADCnh|xJFlM-aK@lY{GLXuxuD9$Mw17+JRXgTQWdDG
zuwD(iYOXf~o(b#vKS%P^nLgH;Ui~95X2JJD=F>dW6XluM5xIk9-_bbx!f3AF(rqM*
zjYP@E(q|;wtI$pQ4DB4WNe|b(z&Gi$ZD&h-^_p)xA5nB*bO_ffiN2DAu`q}pVvtdd
z#F_h6ddw8h^N{8|<#Rhg&vkl?l!}=i8)e`;6+I@(#Y~Td3UPgU3{;ABdhAn$nI7}h
zftJ6o8%A~Wbd)(g75ETBxG&2TMr8`l%vYIrv$`C92y^54?&-Ui?YT>TQ6NemHE`Bq
zKKn70;$f&K`g$^nzqSz!jS$9PV-gG|A&kA|ubu_n)4qrga5t~z&P~Lf-`GZ6+lV*k
zo5b}d=62%IPE=gKMLcd{?jRl=m~Ru0+n76vM<?bx#N!U;yTs!z<}TvVg}Iw}bYs3p
zJnmt>Pdx5p?jasMn0tvwFXjit;{oP4P{aYs)o%pFM$B75xfP#0B!V&#vk8<Y%qgHu
z!Mq2Qdob?><zCELpv=OY4a#iH`Jl|lTmZ@f%*CKA##{o*63i8#tiXH(lt(zXxuyiR
z&v1Q{#_-|taAy|5ccfQTk1cZde7d?kc6DLl0o~~I|3VOpHSVsRH?grYdaX!fct1lh
z_hun}A&8QRcQP5r<MMrE+&-e`Tuc0F3HsfjI^tJ{xrwNoh(G@}FM;wBXt)!&41SmK
zfp0zd)niT+{1UnEF~?PY+c@|1<W`TplG;gQl}`x$J8*bo&KSFG;^$K4yqMb-%-e|D
zHqNruOJf$9O>52S*VoK8|Hs8#^#2)1amZu?<|mUmF<B;)l8R67JXv?Xda#4Di<6^5
L=EyY#|L^|*-;`cl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/element_wise_operation.hpp.0715A40D7E0DC380.idx b/.cache/clangd/index/element_wise_operation.hpp.0715A40D7E0DC380.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21141e524593a6b8f711bec4cc3e201bda9ea30e
GIT binary patch
literal 10376
zcmbuF2~<_p*TCP|=L!$H7aR~D2r|UwAs~~aprDAzqzq1Fh830;C8bgsYNi#4hM)<G
z`IOS4fZ$&wl9CqTP|8y=1abf|B?^_)(!%`w&f~uAJ%3&6TkBihwY;0(JNuqtpJ|_c
zoE8@yz0ymgnfPqn>*@1eNfk62%>ec<<CVo}$3UZ@RvOLghJ!PoNIDZ>cPwMY!ORFR
z=W{{5I^DpU^swBxp&QTtuz2J2xgD)_uf@~f-fUmwm!7)S?nG-@*|?i~Zl2x$^qbwO
zqZhrpX6+E|V@CrQgimR<-F)=zr^oM$doS(SrQAJ%4Ig#p4cnR_4h>tHGSspD^q~n0
zi{|}tB`^Q%-Xt%dH}5ng%<p-}IQHh|_%TzDoc^&r_Rx5jddrY^MlZD6K5(#ey^Y~(
zmpVv%t@!uZo9_Ny_~ooc|Ln_mJDu$Ch*Qra7pfZm_Dfs#TE><W^W>J8Uw4j;Y4|oR
zVxG_2>wnu}H&8ljH9YlDmJmocKYZ<-ejk5$-p`n|H@l#G%Cl!XbnjHzWLhlU9n^oe
zTR(68i{A(5>?m*_T=|3d51)oQ`s}i<|FNHbm}lzq+e)Wdo~tiC@K)3Cn!l<=m&gzD
zCtg@zKmJtdfVsXc&4JNT*%=$$H(Xo~DFqgtv!7j8<kz_HlZ4>4mAemD2q}59=lU<d
z+>^EG*6osRN8cBYr+>O^^_}`18E19R-@P1^a<$(tr|nYKE=!sI+;3BiCxrItejB`7
zi}(LtXL~T=jR_8+FHO7r*6u;GmL<+QfBW3jit5yK!;E)4>PG$4@Z8uq3#;lcfA{U!
zpSId3J+Eb9#r}P(UwmcfnB>%FmRr2=+1gtve^<8BQ!82q(wW7h62EcrnRl>y$KceA
z^s{x7E5b+p_;PAhwtU5^VphP;k2f4xWax&U&tA`cA~1jQ8>_Py9h{e*weHU1#!X3b
z)zDz;dA>8e%f+An@Lb_IWBHs%7Y%tVZ16ab-430zug`gM(|?@5y?bN%6N@e#c%%NW
zn(gH+$1^W!=chOivRTym;d%cv@~)#lEk8MHj{nUMs+x}e=$IF#>rIojAB9dToI7if
z^;0)?oEP>qUPxb+vd-0F*^4Pn)whoshStTs9+eX|>2UqKPIY%z+q+uD>~rk@U8MhS
z9W{%Nj&LgZ)ynF&)k7<wv_1G>MrQilcPup;XnSellJNRwdl=-Zk@tbH4|Has(X!Fc
z9=H(_SQ!28*ZslUweJUwd;o+4Aff_YPU;)%LYLUU!^1RkEq$n#4l)Z=u^9P7{$tDY
zV(np|i$<=eLOr!J3skY5{AbG8n2V9D7)Onqsn7K9dttH^`bqSkmqVYn0ee=_ThRX&
z3^5B-vH#$!yeOlCk>xPVT+x+Ozmoc+f?LkAyIbNnd02qO5RH6;Sl%Fm%|gRwx1`Y!
znYeElF+cl>w$&^bCvz?ydXIZ~cqMo|^X{)-M6+Z!joe9{JE?3ISd90aori90Ugl^4
zc0;Iqky>7)L(BqI{JYoBda15zkJ8u$`UU>7>4nLGFMfV^<fq$8EnonvXfIgyLVvR`
zRg^ufq9dP`Jl5^`I)&xoY_8}L;(Ub2m`m{Pi$lHN6y2G^P>ozloJ)yp7N`>Oe9S5L
zVFxsQdB{fMY-B2mK$S@Qb1VG49cEjAEvw66VtJShGz(K*#J;-BSZ4V{a8Oq-%g0G0
zhv~!o`(BvG%xB_1ctTg)q12?7IM=db6md_bU~%sETY_e@LBvMLIr`8!Hi#5)Pi0H)
zU-K7wPFD(v)<^rxrWYnZwqM`;UeC;PCbNTlg<4)=W>*BNJgIp(KJOK`vq}znz209o
zy)gNyx>urRUt48Lb~IZwO`moz`PtzUp6vO^r$lK*6>+X&txyE!@Z3M^AKPOby4zF-
za}FKUvV#uMSP4Yww)9uuKhrsPo5XT)Fss}m!Yy)dI=%GibH|I)cGfZlHqsA+{xJBP
z1s1d9k+&;%o~!a_Ds0A8P<;jUHw#o*@lB@Z>Ls5%-dEYaa^w=?T%wMS)zhcg&v7X?
zbr~zB5%i6$7)4+n@5T)JXX}@}r)^*`Q@IKKZ?ZY3h<hp(dmDeV8sN!t(P`v0#CZ*~
zi6T%XZ)08i^Rq`!Vk*uWxr;b=F%?CiO5x0dUvA2fh@#R?oZFd-B2Z=XcWyuVz5iCe
z1q@`vu#Q^RvEEgLX&CzS4a1K<{^acxj|m4@KFlt~RA0>Oq6p0Av)Y8e5>70#Wh!i>
z2IvF)`(Bv*pN<~iy!UoJ$C6o-yFuU0tfB~1`F#A;@apa_zOVo*RueZ%w*h@GQ!V=U
z)nfPSy&nayT$p45gP6izu-pqn&BCN))2DQB)oX9fxKMXODX0zfZLA<gU_lLolkJ+)
zm&~*POP0@k%lY<wFH;`uVJm>~{Sj|8BnQbX4`$hB;@r&4rU=ZVY0yY%PUeQqOvPOz
z_fY2^wp=L!RhqBV#|7DbDliphkz{?cf8Ps}TdW;y|8R?PV##b`9-{g~%%+M!mD3Nq
zrKZlOmzj!#M$Q6#7Hf+lQ02_(vm@U-^x;p6N*U<On2I7$<?OeXa$@bH8TK%Mb>1Bi
z?to+#sM0g4V9@PjL!vC8AG1g{vCJltS(vQyP@h$PKil%q$NQ`gD@EmjK93cp2+ZfY
z$E;f~wEdB&sPqzjFY%WZfhxCVv_{2LZyRX=T2|CfV!4wodx|g>W!+cQU(3cUwR~i6
zgOZN{^afUxA~2uZ&CXjgS8kofR9LHWs6L08PZ6lHc;5Vs3busfUQDdbJ1`C;+<^%(
z@4T2m7cUzdTHiUopD0h)PPfIL6y+J(863^j&g7^J#+1PX>=02d2ZwTSMaIpXWqAkm
zcfcRJMU;zZWD$+W28nVh9al=Dan6Zy4GpQG5x6#4n_Yl=dzsk<V69X3f1J^{@e}CV
z@Nv|OZ-;OFXZxc|1X0eXx_s)S7F<GgCDchRco)^}qE1+_C^u4FBXz=oi=KmpDDQ`%
z`@tRUEXobw*#IMO!4>5u7~2GqxH5`z2aN52NSvyod=<uCg-Bc@tj+!h_J&mb={H{e
z7VeWp*+cMP8>8}wvR=@06d(j}6efgm6fHz^lq@83lqRHcv_M$EQKpc|(HbJGVf)Fw
z@=J(N!gDDRN_nm#LKV-oM5yKY2oa9(Y$Sq_=Vl@_^W08^cAmS4(8cpjBHZM;mk7N)
zXMvE#+{=6Q&jTTkX9EZZp36We<GCJ$dY%u1aG2*t5E^-I1EG!QZV<Y8UP*<Ol&en;
z6>@kkrb02#6;!C;`4ANj@!Uy;PM&+H(4(@gt|>C$$F>&*K{gP30~zwJd5<-*uYrtE
zxrGL_&_EpMg4{}dTj>a7^oQB9q4>c#6?H9NX#e@7CWr`f3r%XF$@m32$ZU=(-tiBg
zP2N(zu1t0xziZF+kQ3Gqe1YCFE5qa^<KBo~GMm$ii;O-oO0YJ2#8`re9<sUD(L-iA
zm^@=#6|f2Hd&MnQzgKn)R0!<cUODvXgD>~h*tLbaVH?qHW;IsaWpr@SU1lXJzA|nB
z!6#xbb`_3rWD7_!xD<mAF8d}9Erv-dcYv+~Jk$lF115FARFyA)?gF3QMf-PAC$v_G
zJXe!y`_!v_b=I{|2dr~Q<mSjPd);gwXbk%Atj>zpj*UUddQx%O(N5^8voB0;I(8U(
z=<H|3FUQ8uH2dXLWXrZ+JspKdIgyQ>i7j+8*3sJRt>dyj)9kIY98F$27KC0p>p;af
z$BK#QtD6VFtlLF_owaWV|LbXx61H{Fj(%M)JzzXM4X_%TP61fMr(QR_{Pe!Be|8e(
z=%DCOYzL#r=s1ohM^95JMjK<xkxxtzN8@59sbqj41B9wW#sHB9h~s&(0j8<E4Yb?9
zmS^v65Xf`*Hi%NW1RP4hRpmAaYJ*U92)4o4HhxcQgQzx`%%>-`!3>q#LE8?tJUh07
z8_&M&Fp}r-?GUZ<FW~SCxT<`SK75gStBZRl9o0$0aSDiX7xn6*{wiOgPFLu|s(ugk
z?V%&luNCEM)a@GeLdGK;o86|P8!j7obYoqu9N2Jp&r6uFZ`BtibL~VqIVd?q^`Mf+
zBu8)*nH<Yed~&i%Q?*lVIC7Zk!jaF^K#szvPE@H3g32HS{VtJt!xPIOj_1i`FjeJp
z(3XP@&pzc4$a8o(OjNlV9IC-Z<!%V-h7fdKM0Tnf(+v?SUk2@Eu;JO^GPtOG0~~Is
zj3-@UB`YUfoX0Otd&=+R^{0A3l#O(#k-FmoVd8KjjY3|qajrh9BYw#ck?m_b1M$Fp
z&BT5NGDhWX#BLjLRk?;Z)(}tZR8c-d9L^9Kk3-B;1+fTpcn}cfN)RhSr<$S?*iW44
zDq(yjOjgsYL9FKWn?TnD9;#p11e2Ozs>-K8cZyFx4dQ9gsp(ficNIK%Q;Ii0r}9SX
zxRH9Q4sHQ;Ea172Iu`O=ME#2B7*&5O6}M8I%B9r1l#Wsz<P&t@2|Dba;lKN2_||QI
zKJmadgP{pFd6bX@`UB}7zkXndXY>X3#af713$b|OL)V$jD<x2YzVO1m)gc3C70zL+
zw-Izka79h#q=2p!T+uj!d<k@y)CEG2uYv9wxT0}hOLp%!uJz|ob__6r)(9eIB(jXO
ztsr7XBFjj72}E_Jxdz&6AYw*%JYZE-0w(AwV5Ee#T8WV0+JHH;FO;|lHg5Irlbk}2
z=Q_T?T>Bt>2<{_{#^@tBiqyw)6t7QK$xG{H!;yoR3r9X)fgFW<O;pKS>&<=VqxIn^
zEGR4lw|7wvn-CVmQCwJ(O5xgY8;-og12_r`k5nl$C^G~PXri1sCNqMg$jn%d;xm&~
zTB==Y!;!;M7mj?E267a>bfQYR#4VTT)tSDYxUJ{8fVdYBFEzc0xEJwUO1w&mAG&hv
zj7kDZ$ta#fO365$BT7j$&#|Q>QROm{SVm^3Tuu_pdH#$fe#Ubpv8g0>s(uymsv>^s
zN>xQhSCMd@$5)X_JjYa#1eJG@gk5C1%GD&Hn&%o~Q$y_3^cteCAtO|-C8KId7<zXi
zbMD5~k|>q;ktzGgG?n+0Df@Xg5-%h1LpPJnK@wmjqj(N6l5spo7)dnGu||@p@-dQl
zjLcB^I7vLt^9ho8g6EUO<|MIG^_z)TGx1Yb`erh^nS}E^zL`woIi{H;sN6~tTFG>k
ze<BG#@qC8ZoFR5<`WfPRhWM)7PDZtpFpO7<EJ!)7okXeJL8f$&X)0eJQ!en_MZCI*
zABIGkw?ig%k;y7&K~NTi;PNW6DB74Th~POgOI?0V>G4^Rtfps!HXCescE|=7p1rdn
zfakDmh*UWTY;wR(<+U(*EllNk`dWBg<va+=gAjG6&4USf5W{m^9we!p588aN;dyvI
z`13qAA10_=1P(>uqH-eyH9`ok%q-LcV;UiX=g3Bg<vG3)l2twl+LK_zv%^Vn;o18n
z1n?Yol8*?T28YwUehv-Fp&{zdn?on$&={WMa%hsuYiQCMI#cCbnv~1)I-0bO=dILX
zD|JDA49NF|vaI*R06vuE)Bq2wn`09MHt~k)fWQt2Q8&k{5O`Jf2CdCueca)2fn{5t
z63oZQ>$<M0MbjsA=d<nj|Le5<w-bUxo|T~71>lKw!^y7SSJf`7e84-wNyWZUt_JXn
zO~am%4y;-93$x1ok-{Q6tcbd!`9%5t$YCjUETzL#N9q5LB;wJQon4ii2lRPOXJgZq
zCMp=V`CM53-K(Sj_1SfMcio9EosML%Xzcw>yT2*-SG|UM)lgsbdhQR3?pX0|euK^E
zNe@^VZNaRpT%4ergcm0)!E|ear)6}n*w4x}2@WAtU=aw@9SNGMx_fE$khRy#`U>V_
zPxu`(u5x99Ux@cb!@rt;z9Gu`NQWuRDtVZ%+1N#-te5m01xNuLg-Kx?MN82fB}>U1
zrAcWVEsz#)lqqF$w1!A)2*)KvD&e`5NTobi5vhvjS|Zi*e1u3xcs3Hr$a6E1nt5&~
zQajIGMC#)CCXsIP+)Jchp0hy80<QjfAm#CF0Lj2}8AxS3*Mn5g^I?z<^V|qhBhPIh
zwej2yQa8^lskD;al$sxr-I7Q-JQq``nCA*ARq%X>N{4vvq*5o(JyhyZ`RngbJzg>P
z#-#@a;&=KLfuncDfBeAIz{Qgdb>*TChwR~#jYp6E_Q$vXdb(q{g3TGzbsx?EbU#?Z
z%IzMme8<B)mrqM_+VnshFdV|ZP%is0>~_lPR?>$y?isAAtH5a$c;Ho-AQwVJA;jWV
zE66(`Vkg96AW4v`AfgIl@sNyh3|4~a$`EHeo-x_a${istiLHMarafP@cJ;ru5&9mi
z1m%hm52;u)mY`fLVglY^F{}P}jQG5iFMa28JcO;Tt3bO7`lB&KcFgr%3&B{H$c*D#
z2*GF^kr~H#Cj_H$L}nb{DhS3Zbc7tAc<VDAd7xje+)nu`VQcniQIvCueJ*jvjn>2?
zbNQIZI^wvFxT)#uiT!$Be?1wwp4Z<%95?X#1;oC9*DoN!1!Sz6Um<ZU<n@b)eG#u;
zM1qTW{Vl|Ci>i<Q0JEub4U1zHO~iUZxpT!{K-;ne<)-ysf?>}${-3*fJYcMw*>PWa
z9I5)VFM$u?;{lx&uH49C;n*Ln4CNLU%Rmo+B`9~VxU2n7K*jjQ{*UQdc)J9|63}55
zqWlYpzkm)Wy(!Sx{~C3|nWb4gZ$YYQsmfUVOh%09554K7JnX*_;v;_>a4cr~_1Uid
VyoT#NJT)GepwSFh|MXP<{9lRV<G}y`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_add_1d.cpp.CD0512697217A13B.idx b/.cache/clangd/index/elementwise_add_1d.cpp.CD0512697217A13B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..08d7ee996a273833614a112815957d5e1cac961d
GIT binary patch
literal 3570
zcmZvd30PCd7RN&fX+kau5J<u*YpU!Z1V~~+5=co{V&Dc4j1eU!Ae1LiiKr-G6)L!Z
z*r!!NXax&`wo(_WSglIk00k@5K8j*pKo(c@J?p#Sy}60+d-*<w`JMlqGk0droJo*~
z#ZN3S7*U2qm!mF7$6+uSE95V~ATRTVDF$Oo!eDg!&M0zn@3Wv+b%RZ&jfb9|9@|*l
zJ?3|(!{XBgFB<EkkNZaM=1SuCzI!YG;a2dePtO&NnUuI!TfdodjZ$OmHeSuYzX>~{
zUXwT~`&Hvs_IR|A*XdBj@0ro@eq_tnajPF*t+9Ci;NqK#drzK4t{Pvq>h%tr!drtg
zW4Xs?-d?Ws@4DObMtka{U3KYm*;Y?gvGam!lU;Y8_4#HlnZIstEc-%fZJRI&e=xGw
zA*s{1$gee^v);6(?OUa#ciFnCe>9}Gh#f8-uRSE6+59}0pZVensf(jK+134roFP&T
zDGyd17q%3T#jU-YJc=gQy&bL}zL@Zv8&TGmIPIMO(0eC2v&9`39Dd|TnXP0<T)$ph
zw)%&}tCTryd?EI>GrKRMN_wYRotw8?cs6W2I?^n?rO`)_?q`3iQ=YPRms|CsgTd#w
zMn4qh7Fg`WbQIQZ?U|UW2oNh>_Y?X;gV=LkF|Ta<?(mDlkZaQV?*6OO4m&>2iAd2D
z)~>`B6jNx8zCus#!2SAvk<OG=O{LED{l!aSb5@%WUKCMFE-Bx#xKY02<n8lXwr*LI
zP2<*1`?Y7jxmI3{BP?lF?$=_JDW7aweL8jSXDM4Sr+QbkPceFPyK9ngQYzFH`tzl<
zD~W=r2d>(~4Hwci7T-LFH=I1Z>5+?aASq^5*T~b32G;7F-*2B!(tY*a*wV=y!6>1H
z+0!SYMwYA<Q3={CO}gUp%z^8TPF5{L3ojR3@%<tsd)aB<im!R?Kif%6-e>npvs|0f
z{MVcJA3xi1;*U(LGk!ldUma~Ywk!6Z8=Ak&Qg0XTo@ae3`oz>R(Jqsa;<JmdJ^Jd4
zepw8=?o+$g%9@oyPgh-#c89$5{3+=4&FMdVOurO3#XOVUXkO7aouL>SKFXXac-QV{
z_d7Z5=*WXVvLi0+@+z$HdKH704gSWo;}M0MvUS&!hi<ML(FXLL88CVOeb0uMmzK1K
zhxp9-wxo5eQob@^GQ*5d@xyB><G)MzHDN{LG#oj-<4IV8*H_$_nzUn=!aH<#wv4ui
zeLV4(>!0V;iFdUX-<R!i^T{qRP4$e)9o(17YqIPLdL`28mbd;DzBDhcOYo*EOuBq{
z%|zC?IMDma=EJ)K4$b?%c>T|7Zgy#Acr*@k*NJ0wg~gPOHLpWV23z(vS8d}kV|J~K
z&wawV{P^*aqj|-zZyb}--6k3DSMAi5uxiIWR&WT?q-cq3+k6vus3$Y{<ehEjG{gM;
z+aF)<W@h3N_DxRJ6K${G^=jN%(_yJ9c=@-z3gK_#x#a=RC-T?iESo`Yx7l`W@Aj+%
zItZh}U|<(B7y4|2D6>DB%4N-%m0J@D8;{5~cpKtugUZzsF7EHVQW92%$kupk%4`GV
z0D3>;{zTp)2(uKC{c-+;*#^i_RnMlg+WwV7!X_Xx8BZq8Hb9nme|gU_o!LRc#v!s5
z-pX>e0djRhQ~Rq&Czg}2iHJ<WQ;4$-kZ-brVh#?M-z8z?h)l!N+-4gf=c!lZn-Y+!
z7Dzf=cWqgvDFh2x0uK7cu#hD*B2kEl4K&=Fr@IgNNIOj+IKeX^44A=AV$DC;K?v3m
zH1=SS<tgW!`L)qKSP143cn-h}!F&QAC9vsc<{xY}1lxt#alpDDI8l-qYY3o8(kwEN
z)?Ld4MEma{;mtod90+Cz88L=%ToTs?xInPTL*#29HAT%ZkVdLW1jM)=SoBM?pdZ<c
z40?tuFwY^EgjW!*jUeV^YO)J3yPnR|j&I~{Ho?Kk<YX&go@;-=vT-1AB@PF(>1=ef
za4?6?F(NLFYX@9#u-abj1_=B&G5;XJ*v`{eR^rpAu{c;oR#}2k2r*T20r8&>Otdv0
zSdAP>6g3Jx1PB&W#VAF#lJ40|Zb(MfqjXbx0$-59rsf}HE(BXqt!N;BA=sK~O#`PO
z{_*pP?C}+K$aE|}RxpT!5T6wV2)L5WKS+59rVHt@h9jfL(^Cw@Cb4Zn00`!gI7Y<t
z<DrR)B%2@Khy)%2KZBppLuf=AH;s=$)37uVfTVl6w%;flw?NiG=h6j+&tPyFD8*gs
zl@It%b(`W~0a@S(d~vXlEHok!S%gw%UjAy?OjCsZLgEA#zy*SZIH3`Va3Yj+7e+>l
zaDCVD5FAa6CL3I&M5z(Q5@U@hju>Y|@x*u|k`ZMnmE5}PWXBkB*XL#kn+1X+xf#MS
zB5p7@#Gt7csCj@&kHqFE_V-;z78D&CEe2)?mV`=-NE#|dY0FfUV|Y)cJ(6u?OEL{)
z*;c9Prl;HAhnYceFd>*?h!jhWMXw12$5G>qD4rT`L^7%jrLDz-=as_i-pDL;H+m>I
zK~PtbyQD9wpkUU%UaBmmI1cD<081$Zlh=Q%Ip6=&XMKp-CctJsC|w7$V{lUW=6lFR
zSNW>~!RUrb+j*l;+}|R55#<tvUL)kLxrmJ@gPGwCT)r+(+4JDDvyS?T>BuvxNqk2>
zO0Bw1vxPYgVq}mb*D(?-1%jQpPADBAwRDudo_U9SYFc2L2u#pE?@vYAwA@GksrD)~
z58Ly$WbIi1zd%O2)7<BR>V9;XSGb2E)9TZcE0BR8NP2RGMkL~h;(>N-@tvTCN{TfS
zNkvqlS$k4kt80sRcn6s`D>!Q*7=5K8@9oQ5{Y#P23R?wQrEa*dEidIIQuIMWRU9xH
ztQ2{4^iLc(EC`k}<Y;ZV?c8ZL=<RY1@y&9`@&LX=w64F3_%22I!(c{v0W)$KOtBGV
z_+$hCZ7i!z5l$6eL?VS!LfuhwP1~#e-aU06K`<M~HV0;;2;n$J#Km!qNKH_qmkfC%
z9W?0Ufsnrjy-f&-G!hW=ZN;-~6MjQ7WY?4~O7sDE7gzVF;CAXIBmopb3L{^ozcL8;
zzB_k=xtFWk_dj2c$x~;i>vt|cZ+U*9zOzW*{O6YhIv7l<p_^vtDh%Bev}+1GxyvbQ
zJ?$?Nm5UWg$tjCc(=f@Xj8Pi8iw)i7IT_g+1f}X&Q=A$8BLa~GSy)oY(N>W{rj>cL
zHJxh1unP<4a2N3S5q`GbY<Je&IseZ}b8vKWadq?X{CJ+1kAFbm{E$%RV0+&nku6Ou
OS%QsPDEP!A2J>(3pDZ8%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_add_4d.cpp.6D6ECABFC0DFD709.idx b/.cache/clangd/index/elementwise_add_4d.cpp.6D6ECABFC0DFD709.idx
new file mode 100755
index 0000000000000000000000000000000000000000..59c76f0ad357d82e5ee11199c19180690a73f9fa
GIT binary patch
literal 3870
zcmZvf2~-nT7l1<;B9H+R0tqBx3m{vlfNWw2n2@mVLkP&K>>7;V0*YF|1p#eE)Z&Jq
zRg}fzk3~e(%2p6XRIF01t%6Ds6|taL^&j}pn~dlDIfs|~-TU5sciz4CO$hT9i|JYf
zf-uQ9N0z)M5h4%>RP--<O;-AE$^^m+GJ%kDxGm!M%%{$j!TxvY_9=%ijGpu~%o}Gu
zGV5|K;i)?=DpQ@TE-N>f9$Q>K>^gB$<)Dmimvobz^{j7OJ8Y~h7gZS?xY}0JaQ(@|
z-=(Q`hdv+eU)8<0e=_f!UffYj{^jo(wR<&620K6Oe3G@MIwyTu?#4=;=Kp*c9$VG)
zex{ojzo}*E@YBJ!8+?j+Uz%b@YWtsl?0j_KP+p$!;BVgs-$|dxjT(W^S<8D0ht3U+
z-+KNqt4+(=?yrDxV*H9emlw7fKWCQj4~VK_?Of&hWomD9z5DdX<J($GV8aXd-u?K9
zBs-xSeOmoi-Yz$@fsN1ae4&5hOM3VF=9C^So(^$cmE(4zu9y<SX!=?6o|zRZf2vBO
z_d$>?Kkvef>Q3p#7hcLvewoL$-G{j?dF$@)Dbl_cBnr*%C~&#xV%j)mnPVoczV&d@
z=ea=Z;`TDa`So3$7TNE8*vB1wECla%72VvS>@^VTC_d6lsUD3z8MpiK_gtCsL$_9!
z2DMuc+&R%vP`mr!)0%(DZH2be#s221x@KAN&ZpV#CdUQGI)qgr{yr^|7#p{og7Bh&
z&P}`8{s>DdFY{78dOG&VZNizR-P%K=8Q+YoHX!Q6j22#7?s>)B-%s00<oK^r(XM8l
zQ;=}PB;#~-d*Y7my}4&RVs;z5`9z){EPK4_?NoFqtF<vRHL<%baA0kdmh?}RYZtB5
z-`#@Zo+zaese$8t^)DSZ|N5-NfErkTdb!((<_FTx_K)_LpWru?FDxT|esE0pt{~{V
zWQJ0czNw^nIGx&N?S8g4sk%`o;NzCM11poii$Qv56+X>xs+N~4Ic;u?zBziR_-?Sj
zTV<wx-FX_TK=V@Kjn3^eWohZj$R+;utH~kV#V_~TpQ%{Es|x$D;vQ7l@MWZEVEiX@
z)^1X&GV>(TGKMt9JLL5J;oS3P56v{Gy*96&d*cAx>ExQ%$ki_^C*$iIQ^yW7osD|(
zYsN}LSJ=JYnqN7OQMgcS5%pt+$+?N7F4HT)YuWx4ft4plH6H(25Y1!9JUgp*>fxut
z{`~7H7b1wyy?@|wWkrXh&HYzBD2Tpn%(&t(CA|Gc9RDf%YRC4?g@13cFYmfM*;3oc
z>a43{&HsI?y=c!rV~j<?Gt+Axe_XbJcCh(!<+{yU`Fp3GsEcY2)-)Aw>{>f|sd!&*
zsz<e;v1fFAbE?NCeGTUG>DMo7vrL}FCT}@>>XO%?=%*d2i{;i5^(DcF?kJhU{pnE4
z{oO+rzk3`Be$lN;bIj=JUoQ>mC{AhZNoz{B?M!V>b-WHoI)7wc|IB(%$j+55dyn=x
zJ>EEXjLJDhCI%4*2%SV%n;THe+=pJ~%Y%O47m$g;s9A@kqct}GvmN8E-qTlEQ7|D9
zHM>AAs&fM{i*~)8$Y}UiN+yP(W(tX-Ha7sXug!rc?8N!!$;3d^Oe4`Ka|1A!g`GM(
z^`dz>nHY|msU)h}+yKn`+H3a&#tFt?LIP^$sBjo_12ErqcJQxyT=bAk3`Na)Bt6dD
z0L)p*DcMXl0zuiHe7dr4St$udL=GYs?12!mgBVBN_TJ74n}i`@D=3l?i6bdPiV^f-
zeE&(VU#k*~EHz*13YJh>3}p-V`ScTEL_iYg12v2YNkWW3VpP<>N(hXwTv?03!@)?5
zZ%lx~fm|YwQqXGS)uw<zB1!6BB@jlK0;a#hohR8-2P}aRZ;rQ(f-<!-^%NxclZOLh
zbQG6zUgzdQFcM)A!3Am<kvK`*6cp(ci6g0#6eGPG_8WR?mfS@TN~S9_0!s|nAM@;1
zGJ7`bmlA;sOhS?1D3TJ1BPm6Sk&$NpqMfRJf1!R8XbHx^ugS%%w3pxVwkttMf@T60
zsKK|A`d3x#W;oHif|NKxgpf4NGzyr7()=__Kpx|RuNrEOu0wO`OZCOF2qS(}Ka7Op
z$)WNK69W*0@ELsUFd@W);ejJhx+eqpg^)~UCKk9?9r=mQ>n)M!^Q3T6%)unkXUghd
z^%soLXf!(LG8oaJ>CizPh%aq>wen?3C0h693(TFcW<snrtZ53eq1rGM#MR}p6lAM!
zYos8aE{_G`0b7#QziMb0(HH0kD2h!#NMEWTKAEoz9KeVN*#k!ct^n75Laq=;B8~`G
zabk`bM{<rFPtI~ru!CR4mG5pAzSKbTrO(q}q6mY@V`3E8bt`nx_RX)#5F*kPF~M30
z5o?NZ<gMwA5fY?r@nR281@&ugYHo!)A2ps9aFCvry)O9uDNkke(t$)8Kn){eNQ@(I
z$Qz^8IsATh@9j@$$@0v2Smtxu{~F!^4ctJ}7@!uQsqhq}7KEb^wGbSIs)gbxOf3vY
z;cDR+71RgFB907nqhSU(1$Zf(2Ra4fD99-YqaANV>_z>hOte5L6bc;_W{IEj*75JY
zxT1|5LJPqT3r0d|p*RYoh2bci7LHNL<_Fhe1vhL^zxo_~7mzOSia<Md_Wb}ZcW~-N
z^{-wDFd|o#V=nhiC&U(Lu?J^+$vH&;p6=M0)N6NtSv!EL>2~ym;2N%4=@0pZZhwNF
zf7*gHdoa8DwXW6h-=;I@5&5!xO@JD0d$u2rWE`0Vu$?H9R*bB@%$~jdnF8GXDr5>V
zs>^9r36WKc(Mg6pLoe{WFk-|r!l<5Hd%ke`{U<ckB>N;G*x;<yvxub=q1@R}nQ7R|
z=xo-Gj0$(;9Xi{RZfOb5=%>g0^%Z)-X|uKHD)I+T&{}jA<H*~^+Yi_pW9~atmuhLF
zp3>CPuoAr>-j~zh`TRb*Z?;`_0GQqG^(qPe=a+5hrDttsZ4UN@5F1MytUn-x%jMcC
zY_@i`_JDq+ThujLcPv9Alo&{`bA*vdgGd}n4Wt<L?#!C`=g!>(G{QLDIP8+XX}Ynf
zfFG_k8@Nl_Vh|xXU9?<ge<y${f{|#wXh$&Vo?)xX0~_N$bZs^(+YG3O=vO{(@G#ym
zTL2v2`9O^p0LKqUGD{hD5smF_m^xz9UpWs(_z+)RvBU%Nz>z29iK9%_Osp@_I%dgn
z)g@mbcT#wf3+29mW_G^a{py>cU^FEu#wi@2{uEgGV$HqiF0{Jh9O7_mQx>E+fJs@&
zE72cy=zoG)%d-n-h1l8Qe;)}D2!v>b7^4uED#Qqd7^M)U3Ncb4#$lo|a-C{BInYV|
zpdJ?+5h;;IMK6sZESDv%OhxDTB?l-&Dx`U;YGhc0qNV9W^%Bgds{3d&XgYfOu8S7)
z-1#1!-{{)7nL1k<{XZ?8#WpZD;h33QSXpxy*x4_1axrvd+Byhz>0;k_q9|mEw~{~M
EzZ^8GqyPW_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_layernorm_blockwise.cpp.B1B3F8F277F6314A.idx b/.cache/clangd/index/elementwise_layernorm_blockwise.cpp.B1B3F8F277F6314A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7c44c49b1d4ba1cb6b3f827dc9dd0f6326247769
GIT binary patch
literal 6260
zcmaJ_2~<<Z-p{}#269ORB`gVwKp<h0MNkNWAOT!}Qbggh$SNRVF`ytcC~-k8;zk7|
zDgq)Ps6hR!){R93Tu^YU^@-c7eYFL(YSmZeoABP;dp+knU(R8e-~49z&&+?D`}G2U
zf0+@3F+XWRenxU`JjP%!^x$7kt~^Zw!S}ihM!xFO(x@>0?5PjEhmVxn;Tx^?Kkkf^
zyqdMXWcQ4<_ucOt_{Z&tiq?A9A1<5^C~Vt)Xn3l|{#PxB|6DL~deK6oMMfQKkJ%;R
zDpAO{e}8@oEM?z(F7!Ap*srLnPZ@sPf8bH~h7Y;tPsMq7dpww;5{zk2`}(RwR_!CX
zr~AJ@Ze=bXudM&`#nCj6eS^UzmVv{A9cw+tozGOo#w@R|WAFCV`o_yUtJrvEtAA`k
z-^D8i$A33HGZk+<S#f}U)b>UEA;*|nm&W(Cdu^T$vBFc{mwp-?X?oUG9F=8!tW19m
zd*RKBrRlSZ{ZpN*T(sCryn^?4ID9^SqjkPGNYDDO&d}GHIWK1Q1P%3F^lVo3_O@$X
ztJf~QbLiF3zNK~#&;9#-v9x<%P@(?Pc&1h0O`khE2lijv+feDW&U%9=a<lub=DPk9
za*^3&)uFE*nceUjDgK`Zvl9c68Qq&XWBt8mu0x#Z(e}#K=S%t|ul=ula{tpO&nD5n
zB*357up+J{GCivB&g)D)p{}!T+3Plg&zr8d8_dzJnz~VEpVIcX%ChTb_UDdV?1-)%
zh$uaY3DcfV3wB`g?-;C+&p!|vSnDGTkR3H*TCo$wHLWd5yPwO>rML&)W(x{(Z4-7I
z8w>mKyPAjAZ2tRdWy71DX-;+B@!x(N@)+=zxUcAJzt_CI{EgSPpTickUK_e6=^njU
z^=nM}njLHVcG>^t8E|<eSL@xpKl*P(T#svad!m>%y@l0iD>WF_8JoOUW{_~I$82@c
zrOKOEpB{8I@ew8^cMqMOCw*aZ*69GVv%kwH+<)8soi5kL-IhnX<gsc?_iZs>^Ee|t
zw5##upC4qtBbD`|-`9U?yK>JrvdyybRJHv)yC#QzJK;(E$RzC!?W4X6{`T{C?>K2O
zYtLx;-G`UndAt8|%e3K9_tVbK^_;MQ2WeOD9oPOb<Uro;$8&c|`*w^RX>{!U5L~c0
z$kU|nUky{nUEc24`8vk0bKe?$m0R()_Tt9ojbXR5Q#P#_n-s7?^Zd@u747=w#@lCH
z47&O<aLcv#GD#Psy8nZg-KvGAj|+!~)(r=*8rNe_ax;FqP)B7veSNdt2W-iY{A(2_
z&-b?Q=Ir%2&FolwKI25z9{cg|O@B01d^7TB<X}&Z{ORoL<EP3!!`9`cuWG9wsrg&%
z_i9bM!%2lKKlHX|`cS8E?a$lCe|kLYSb=6@v|X`(JbTzfH(Yx8Nyyi+lbVtu#zekh
z>ErtE&OP_PHT11usLgc8tv9=dm8yqb)g}eP`30w>TNA2=qMm>EHur6PXw<2q<e0;6
zXKXi|A#@pd=o?dg{nT(dFZF&=eocq_n?<oKpPAgFo=i8-`t*yNf3bRyJuc7p_YSVC
zsoweK?E3iUk$Y|a)xD^(&*#jgw`@`4+M74S&ZOLnSsG{7<x_P-S~;upf%^#O@>XWr
z?hD}V<Bta!IcqZ#+5v+Bc3E34KXg{g*7%yiAS{`dQzr`OWhQ<Oi7)fGH%<n!HNJwh
zG1FLoqCm8x;Vq-5qw+|cA%k=X7Q&h+5bfDKc%bCWo+h?N2&8#T9(STZbV^pfXp`J8
z181Z|+EdHZaH2qTYINb<I>YVH*c!o*<}>+v69q}f<5y~a@#|r0L_pevX~LZ-5S{U)
zct(cCp#fb+IHad@rt41>h?WoBOWf$>d55hL329TNDSx6s^e!KN5Pj=&0$U>p(gsWe
z?nHs;E5-XB*cWcU%+^>6X+x$VccMV_)43(S#=)OT*&1PxHewo0nJ5rFI;gT~9en1-
z)>sB<0aL)AC=mVeLyy+3URISZV+o|qIA$gj1)}B2D{^=+44|9S(D~xVaAGZgtr%U5
zrW3pA;MIguHcrTFWHV7SPGs0*&`_pZri9|MT(ag+DBC8R)>C1vu%kGIRI!jkMb<?$
zu1H!$<CMTc2}D$1CGb%Ke}tLWFTD8O=7bK06DcMswx}5=QcY56C~aCAO_Y_UD;-eJ
z$)~4fxO^8lsGfkCVvL&M1k6+#O4CU*rMQ*aD+LG{U2`hvZ}hnZaH0ordVmokL?#C&
z*WGqXpM(+dym&*@j1dXE1R6>-Nt{k`NybU!IvA16OQ!X#1tztCPxWmCCXK)hVR1w9
z+^TgQQ5qQ01O!dM1`!z13<S-<9$|6!!vMGUxqX^Aq3}?6qh_2a@+hJqC2&&$@@|ni
zPty6qjB&ymv!09;3MZ1-Nu<Cyv2x~0asr&lnwCY%gcFrOSP7h{#g75WF%W<-`VVxz
za0?#OUJax(<6ki`q73MlfvKn;MpOe%HJD7tj@7`K!YaT~(RmecR)JX*-w$~G)ZHQf
z1GlSoKFES_LTD^Br?#8GP9Q~q8e^C0iYA11!q2dz&|V<j3xX+J56tTUsi59}+#WsF
zcwh}wLb;AypX!#QlS4zfoLnBot=3&_fRJAc`^nU?b<3a<mIKRjV2ud>vzta`zppr`
zixb{6y<JJiN%-mdO`(ufPijb^We&^8lfj8)-pizDI-H1bjiB9Wq-!J%MLR^(#zlKa
z)5b-A6}^<2CB`*|HZImRmWFD8aSb3%Vqw*V9oIMft_yc2WC?ZA*$2(uoyD^+-3gmb
zeNBUqQx59A5qjJHWuF#C$T+eoXhDp~;AGHHCMT0rQ;f*sWRb^;5!sw<vgeXued}L$
zd1`6H)3fjvQzsN5j-dIINO2?$#fW2QC{`RxLkc|wxiB<1y&@V?0$n8_m&S=2pj$)X
zxY%&(y62s)YO9S<M4_F*Ng|49NC}oG!7{|gC9S{ft+QO90q>0)VS$=)B8D47BD5<Q
z4%Cs9()BvPsskoy5DKbHoiBnZPKYoO7hMYueGj*Z8iYgN!)>PH6F_o;a^E=71|)41
zPW^9ahIXfp%akl)OibQAP9(FFX?B;*&!$;g1rS$IhL01CfZIqJHclJ|lH<S+?J7;4
zvnJ&Ik$GA$%waNJ)C|oR%b=l5ER%+^uq+zN#<EF-atBSFFLDN)u*2+F)FE{MNe3W(
z?W%mQLBYV#CpJbzSVoYhixH8Qku(%z8AC&{ma#Nc$STxEvSY-0z*-MzxEQdC>39QR
zZJ^^CU{yo842(Dr?9T%-LSn=PV1I$a$kw!UzDRs<!WOeNruMiEu(km*mf=J@khFuZ
zkdQcW4%nXqz6j@cKYmt#-Mj|vIZr!}gPMzq<9=NH>rXkbc{i}$4crl-Y+alYzW?TB
zcyD<Qd2>-SL;-m;KeHL=Zw50cwgl*vP$3g1wgAo+O30FzQi1zmsR<J&49pEgD3!wr
zo;gp1LIO@0nj4Bx9)}Y~=0+lvw&8@axv>bvN}Mn;HxVI~{&u0}#y@t9&w-PNvBJ2N
zOiNfxXegW&PD2WTg0xd;>VhH~QUW6-<tNJqUOyNf9s3MV>J%_L1;~5~?E;#VIvb@-
zO2`o!Qr8vFj;EPaid_mVl*#xqnt|;F;=ObOs{`hBR4DuIuecZ`S8yB3Q4T!IK_J>_
z#j3aa6v027fP|&1<ubH#O>;`ut<KdqVY9$jAVnGOA=x5qVd&;3m^#wWE29)e9qH#*
zQ^tc4D&VC8^UyLF(GP6<scffq%s~9?H;a#xU}!avlimy`atv~4DAzRC98G}}tNE+R
zyh@$c@q6e>(+ezm=^1G+aOwqaX!LRMFX2(ILhWE4Rt+qwfe;aDNv%}Wz~l0gl~r?>
zyn#FS<@t(`8$L1Ol~-jHlB)Jdj<We^H1tT088npXnHfTHS?*bjDU|J)P4;YyzUN$9
z%GHOdMH$d71A0`qYM@;WNY{l!@w8QR+z<5n0hwa_;Qw9zv3bw$0gUJcb9zBInyBk}
zE799@hS0<a1+E~K7bA-BA{tTxmJ+bhpl-j>B>ZjXCU{DImVQFCd5n;Xq-31!*(2<~
z`*?#1JfRjKY5^W-bRW;UrOl=%0osznUNIXrL$pXi$_4GkzKF&ZEh?gMN+4DO7iy#u
z1Sml$!hHhQxA$5Oi{Q$oz^If`@4nvbh`*wbo`pmwFzciU<Pb2OFCKI_Q46@W^!Rfa
zm>;Hwsdpx6q2keXi^45RuQ2Y!bp@6I+O1XtX*F1YBvfCQzY5SEO8vm4ACUVU=B;~~
z+IJ-v9&n;ZBAH%cMADoj(u>}i_M_HMwq{r1ZhC=XFQw!^jh9y*+;6=Y5?x?w7cfN2
z-sfHTT;Oe10JT%jmy=l<G$(!z4dq(olF<p8(~Q-m6{yW=Mh|5b_az;J2ZI-1f*D)9
zb3Az=@Rj78KtsvS$uv|8+-j+<Vlc)_8i6O05qV?i@zgJ9%HvsKR(i<q;6x%jk$kJ+
zM4Dq7EsA7L&m<ivPE-J&3VQP13M8$RV|%gZ*t`9<g9mY(h{xh}kQ6wPfF;mSB9=%)
zNmvpMC1c4nR0~+OfDC~+Q3v>S^dMFb`1N#r81N6%aU)<g(#JOeRudgJ16DJIk;jGM
zT>U|TIxQQ_h8E7Z0@hY4QsP7d;5PtAYO@W%y8$dg_}%-L*<)k=|AhKeU|ZmYnm=xR
zT{h})KLq-aa$pW$DkSmWFBdRxQh&Cfgs{i#Y0t~=f#iGYk@zX-;IFxNh+df8mIC)u
zN&(cN$F&@gpCUe8yl%D6Ghg-pror-LS-g7P9C>2S1$8A!UHs>_990;M1gaWERpY2?
z8daT`?>{h=YK^6;F;q30s;(fbnuNm^$7@Y32HM<|)D_8TQL)i6aq$U>Ny{0DWG6-n
zRZXU<i5W>UHM1~zu_mU)oW#;*<GPc%oIpK4Un@PGKz+k0Q;m%0xJx|edU?-tFtGKo
xa1%}czgl^GfvMRHb4#JM*v4+Aqmzs4>}j)1?49Qu@cb97)R2by1b{`1{{_?};OhVY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_normalization.hpp.B86CF9747F7CBAA9.idx b/.cache/clangd/index/elementwise_normalization.hpp.B86CF9747F7CBAA9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7f03b779edb8cd08942eb976092665070fa4b871
GIT binary patch
literal 1404
zcmWIYbaN|VWngel@vO*AElFfyU|<jg;^LB`%twq24FA~}7%JwR4Gqk`Z6NT?|Knk^
zs13`lT5UQKB2EVVx-@mtblc9=HJz?0o|99=^WW_E-}?ASwsY_eF}6n*cjvy}_xoMT
z;?3puMz3ldGn?;ZTfhDi)tPT<^!RN~U+;q-kx8wm?77;v-wIPWTK>UlwdDPPU2A%e
z+*F#aRwz)Z&R%GFGwVirg1@H5-w+Yc8uoL#-m71i&OD;1clxu<zYm{ntbVU&nCl)J
z)PE}T^cKh4p^LV>m=I9Ka<O+-b9-v`hQm{jTn`8}xH0Jm*WVY{82{|rFRxe2@F$w7
zKKp~*6}8U%*RSNdXFZMysg(P=x5m&qb@od)Oa89MO!L|<xpnW(k&0a`|3<HAQG%^9
z&-V}m(`QXL119@M91FWSrStCgIc67M%kXcUvd{9vSH8Zt6ZzjPXVVFJadDaWOIM-&
z8&cVG5^J<q9+WSLomVxdaqZLz9-N9x*G-th$5pUM(anBQ>H>#teinyHQ@1p;cjh#Q
zEmUzTj1>3yp*E*~sbRh1#m{9YZe3Y2U)pk;rFz4LubvsDZ~v$6&7GWLbEfvh{E%sn
z4mRZ;=U}^Btg`#MdBFcmhr|8`TUss>^mI3Kl=jc@yf$ZfT=vTg%Wt+sthHS^_ilKt
z{Z6iS7J2^vlN%JWwrDJhcX>V0xVSPmsSX(9;%E9?r>-z-6=GuNW)QRzu#y223>-i?
zAgG?QWA!n|EDJU!RxV~iRd!WXFaeXE6eJ+${k&V6n~4Re+kneJ2u#4FkIk&?yCEUR
z&BMUK1u;)hgjK|Wfs=`mO_-BWkg+H=t+<1Ufr0DSf_u&0rooIng3|WVUNC)v>pP`N
z*S$K-!6PW@De40g2J-ww0}(tidod>jPuyM{i6?0<iNv$hvonPGm`Bjg&n^(bv)8iM
zh4ElCBNsR{V4>hObASK2<~O`RXNo$Ddcf@THJ_9(rrR@tiAPXJL&q56Qw9beL9+<6
zWLOY|_z9`Mm}tWaR3yqN$`8}aBWS~GBMh?-W)aX{P+0dK`n_PyLu-YW#r0LbD~=v#
z2D(|)P7bCT$TPFENAT=L?cs7TYniyfF#<CfXq}3U4ou5}Rq2PW94x&EG*`-3Dgfao
zZ4GU0gf~nhOrv4u9oogY^6kF+TtF?N3ZixhEvhc6NeC?lq6WrrEsuMTKGyDU0frhc
zJFh5A5lkx+Cpb~S{K_NfEa2=7bKz41)-UY*;@zN76BU(*=>d8}(ncAk=-0OC2lcE^
zZw6Wb3N@Goe+9xe-OWGp4Cqt|dkLgepvj`i4pYP<Xl!C^4s&WzYEB|Zrno4%I2gnb
pl*nMQtz3U*E}IY!A1}WE120Gv0{CDIU{oWR{0OE1f+>Vx0sy_D;P?Ol

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute.cpp.780590C9EF3E30CE.idx b/.cache/clangd/index/elementwise_permute.cpp.780590C9EF3E30CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bb27fbc779d6f883f08e63c1f40c19f66d5983db
GIT binary patch
literal 4106
zcmZvf30PBC7J&0YNC}CAH9)|CKnSRTViE`e3=m|M{l&0|Y$A(5BoQTy9a#iwrA0ur
z(juEw1VIED5frSgLLC*csHnBA3xZVyEXvZNb7AJaJil-9ecastoO93J&N(-P_;`Ba
zHE}r6R-df2gq^Vj91f=i{^UDl$-k=LaG#NJxU6q4iqoVI?KSUK-#_u>I)A!-SB`Mx
z%tiWk?Rvk`61;zv#pOLib8-EDy>;@HecP3s+u#R@o<9=(>N0ydo_G)%X@?K%U#e>C
z7tc>UE>8TkXMt1w+s5yzpU>~KyWkvrVXvj_CYBMc`M{q#S4PJ@rhobI&2t;;^lg8>
zI3i4%9zCt-UD(qzz-k`N^8Uj4L)Yff=D_lvD-&Zc_sIM!kK~j{BrDstcHS9C86>|R
zJ<QuxWHVOSeDTOJQnS&^`H0wFx5S_`nTMKUZma2$_gtC3R}IT1x*nRaf7!Bi$6!UB
z?Zn+Ae9JA}cEXi=jeE4KI&2oDLlYvlf5pC{-y5@L8!{_Th;9$y7-FC8@v;78Tqwze
zZrk!=_{pOl>f|b|?bNEIn(y<zG2K877ui%B|H~o{`X?*DvG!7#*PU;y+cK=OOuJlb
zyPoFWoM)UZEyI1;@u1L4KIdwjKkUugvm==B@^7ea+=|W)Oqg<;wF$K?I_<s567O*L
zRp*xsL64V*Pkzn$`fTHh4>cWaqXBQ+)!kM1|01@lsAnB-m_3wb#caZdv@#`S8pnRu
zjcbejE#|a?u3>|PRvs(m;c;%xsp5uxySfYYuVhTP9NOHJf4SD{9MiAeG^8x4OGUGe
zqnij-UK0(6Y$*Cc<85aqytm8Pt?giIN;9)BdUT?p&a=aAS7J<mmCsP-S?bqg%*%qq
zOu^^JBw4sRR^7>}^VZ$PChha7UapU<{`flJyr?%f&HEgad9q~9Xt71fUN3H5#K5(?
zz9x8H%l7HR&g{JMil_cn8$CqdpZ(CEt=IDE-~oZcOy>pd*WJbIYOdN!{_Xf#Xz4<-
z&)CoG{Lc0`(XCvEQ|Cv5qP5w~PRi2`N-EL6@u&5M^_BO>)!4_4I>w{A-ybpeJof{f
z%BO$3BI)ap(W@up`VyuNUQ920titjf@T@o1`_6gtc>e{r6Xp%Gee}g%$9H+@rj?fQ
z!z-_c*FL`Y{gvN~TN+|JQ}+#6au{<_P0waBGBa;Dbkz`NKiO;RXcY8E@92)04#V_@
zOY>UMnNzptE}e{OpE<^EoRI9Sew~sSFHU4SE>@kXK2#yeercj+X+L8${NX_DjAq!8
z$hvzPiJN$(dTHuMkB=<8?U*Wi&we!98`5#BCu>L5xn7gq@6~e4^&PYv8{-c87P>rb
zY7Z!k=sRN)TtD(`rn2u$b#F1fGxINpcj8EsH?8f$!O$(WJ}ntfI;w=c=6yB&H&>TH
zg=5^()TE-x`ibb9CuZXJk?6h8#y(l=9k=1<Es5M0?T4R~Kjc&<RQi=Kk{wnv-^mqe
zadRNsP59f?i0_+NWPA_~2kR5{EteahS^j*gKWy;ywJe;1iv?;Mf{og8gQ%UG29FhW
zm!Bl#1AtnWs7qaL5cP&H-%O{p{vAff2LrVxQIoXXAnLT<lC$di-DMP93{bC9U8TR=
zAnH;#PiXV-`#3V*AE<SRI@IL`QJ00BIQ?$?Ts#>c0@PYWEz)v>sM`o{)2{OdCdv3v
zpr#RNTFVWhmL+VH<J56D6+80D>g#dE1PUy$6WF8sg`Mo2FqCYOY^~I#Sfrptn4dWO
z-CExshyus3V_Xp<&GU>jPKEa%5c7$ALqx3q;}^wdlL`e8N3o*TB4UFZcG*8Qx(x$y
z4y2a@nIH<r?i$yA$m02uy=IgyA~w;;b=<FZ{V@<H=p>k;;{RPHUj3s>q`=X}(awm`
zs%}}*s~qP-hyW*OBxoUG<Qr1|=!ON1>1*4GvC~8XoUD<oiC7Wf6pa)NDKr#VvQ$_p
zEfr1*7Yv0dA`n8thp7Hh*ic|YcSAoUGAXb<*&f|F1$HuXT7x*Cz;deGKuNg}H5Woz
zIC1B7YvZw9;FMgnT+llZdH(s}P}lU33IXQPc>2g_0&GvS$51Xr&xO#V^*BX-qC2@)
z0(_-grlYryM7WCjN0Ci|b+mQ#kN~E@y4t#W$Ywn+6ueG-wXIs60-LWe*Flp2w$QTB
zR}xc?X{01eLrW7Su~xFomBi8G7$KXWz+62phNL7ZRjHF2OU;y&LCQdNzDN5?ZsxyQ
zfv3O*&IWGEvo~}$M9JTwOiE)H9|UJW<I_Y)hWTH;6?}(1HGn6;0u2EjnScO0X*i(-
z2WYd#OPy801RN3vofoW^#6>AtCf^k}cfu7I2?&A}h$nF81Sgb`je{UxDj7&ZQ{X&E
zEe}E`%*a|lXngJVB3KoN$+1Ly5nwKpi&EAPkDp}|hHj`*U?EAUff!L>SCT7++)3^j
z3L*tzNJ5gJl-IH`O<Xn957sZ_2|Y0Gc&;cFO>HoC94w}TQ`FSdq#+S>z*lA8M9Bxf
z3I%r8aYwg7frE5{FeK5DpmgA$_b*4e-&g}Y8JHP3AeV-O9muMsqNzL}19t;VR`JL@
zOjhyN@_EYj3d{tU9FtLHXigzj7nF&ml8ZR6n2m|bqZ;2?N)JK0C*r9#)sNsCT=)c}
zTJq{-wA_K}u{s4q*&5ks^{AWFvwrr6`2r}&F2*ja5Kl{m(JTjIDY*bLFM!Z8LxBq*
zy8;L;*~dQ*J35-#YYY~|r}NRxoXYA^4N9x?1UIzOd8G$(WE7a;%s{Dy+}xQzJGThN
zraGqjAyb~VelCuj4qgj3qo6B{5iuY;1qVaEij4>%DMC>H=<%VzbUt0AoIKVg)*BhJ
z6vvi%kR8HpzpdYW)F3cvDJL8RzK9s)go6`GXEu+lt1G5zgWyRfC8McAfm28+7*dcF
zD0O*OWVO1EjeuvM0OA)wFyi^ze%Zp?pNC_>?&A&P(ObG{{*TY|)`wD;rdXGZo->$Y
zT?&S>Y1wFe4n4mt?G_G?nG@h(MlkxO0kRHWjZ6oK8;ph?AYm~2h!WrsMhIqHh;;~N
zTqq+HqZ2d4Sh5ylEX7<gW`;0E7{)TpIt*hO#tp+-3d1mK31@_3#)VslW5$JZ!!hF`
z7!erD2<r%pWh5gKqm#0wHp-Kb@}-U#5=qgbL4c(J(m;$ZTpFPynVHNSLu?twr;IPd
z_>_rc7@x8L8OEpFOpfs)XUj2O<a{~Ci(Dkfc##LlQ7?BHJ*xNDm|h3LQvm4~DBsTq
zw7&P*Lep#z#x7<qXaQbIYpWaxjk1^Jt?#PV3|50E;}Ll3$fm$=0D?V+av`-`2rat{
z``>)^+U7wZ2<3E(bkz5vfA#pzyAfBxl=1fQ=<-2@;-sSyD@#ZPzih#Oy|Q?D{!%Am
zsrm0`6Vk!q!j#=eWp}f(D^_;H(XI*{SH}JE;mlV)>RY13l1*XZ5s{m55>$`dqU=T~
zyYXpTQxgIDCTvt8s1iR>Bata9G^rZiS|09}TI$}~`Z~G>hI~gSXP5OtR~x-G0&{yS
tlm92B85yrMU1es$WLd9XYx}7khv&fH(%I`odNfa;X#9piH!sK!_a8{l`UC&~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_3d.cpp.2AC2BFBCE3237215.idx b/.cache/clangd/index/elementwise_permute_3d.cpp.2AC2BFBCE3237215.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6c32cf5ca34f3246e125c994549ca9d33625ea8a
GIT binary patch
literal 3762
zcmZvf3s_TE5`aSpk&u`GAt4DAB@muMc`OiwBs`J;fkeQ|E8%HEtAVHlMMTyj^#y{6
zh>Fs>B35uMidqp`L1a;i$fH%;s?ds3v?@rkC{SS!e0y)M-?#a`oSc8=+?jLc%uJH#
zFtPY78G{K{h2?6}H>Ke)7>o&cXK&K3d#aDYln^nP-0EvF57%4x7!RI!zS^xO{om`O
zuTxq>OZ$JOM=#l)82R+2t99~~??PkSS~~_S)1Acy+$>C5XH{@-b@Rq}j2Wik?Byq>
zSIzGvB=PdO!v={-b4judy&v<XxA*<>?oGhNT$R7VGV`d5Y(K^9Ooe`DHg)p&x96@l
zA4|@k-bYWFt2(@|uj$gC`R)RXikcOIUh|RRAEyUjPEBs!x?OtgNI}_=BRscHOFg@e
zUFT5LcjQfhDcakAdDmUv>G?G=zrHc?WK-QQ9*!<m7lZ7J&U5EyD&#J_w-2xFdRT!S
zYrg;H^dMeSOOw|-_2(C@R1It!ot$UA_xA3-pOCxv%i=95z8iA+wdabBBjYY#As@DN
zwAe9qa%uN3$u2|JW?x%*G>-M}z{?Lk5_9lf3EJtW7i>&8J4=ge^onS!^4x41|8mH+
zjj!o{G#NU21Jfc5E4c6XWcZYC{__1%%v!4TjBU+i+LEYEW~EVkJa(?dA89`FB+fSD
zc!A>6rkkfR*9Lhtl_IxCt9eu1^ZkDD+2^l9g^_{mx0}017=}z5mSg3#OCHqw&9Ho+
zUB5c0#C5&G^37rPsY9xu*TPp1s_~KEY(CVV{FB!4?F-|74NJz&tn_ER5~2@=I@Ubi
z5)d&IU2RIL+*3DFy#E%D#Tu=j+qiPav*juGWZ~P}x*vv%q&ew>H*f6TcKp<fb>kPV
zJMT<C>-6RT_I%co+5lm;U$*mX$@s<T%n;#^>OYPqHwU&BZtiK~*st5`JX%s)FmHFc
zP~KJF+ii^X`g;8Jet(ymxNk-^6;j*q_D%D%dEd*Xx<nPc@}hj=_%^XY<de94(}!A{
zLpv_lPEH?gkM?V1uf1h)IPqNF$3=%@#hsT<VH5Q<nO%J5sbkGg36<R^8qM-s=gv^Y
zO&#|${h1$?u00w((yQ&9bNKCK?x1bQ5?RRssU1h%vvsJhAbsDEtz`Ty>-`nqslo>T
za+`rG2B+|8Lj?V{GMP9evwlSe<SfkE-f7czyYsu;#H8d&T5sL0j^2^*xr-XfZ$6Lq
zCuZ%+;B0L9_2RQFJ(|7MGhcO49+u2}7HIyWVZ!yEG^+fCfH}MQ`PS1Nj~i3hh}xni
zSCd9U+NXao=~y3msYM!*lb|ZTVR65Y?O&e0d+1>MZ>~OMUWMOy`siHNXqhoLx4^Uf
z3n$U(m`5HB8KW_oJr=VjON(Q|%4RKQnNI8axo!Vcrj!yJUpmje@UQfG@QAg9|82tE
z3v$D}*=jN)EkEMN?>pDi1rO$4y(`>a;ud)6=-+=A7>LfTeRsU@tK@;GeC-aqxLuFC
z>GS%-?N0y6-l|#i4%GZI?k9%$xwG~}Y$OH)nd8mv76+hNe7Tw5Wjq<vCXg^`K<$C^
zFkBp9wg1_^FZZ-m)Dy8Xpf<ysQ5FYSEh_zULf!CpJP{iO)W&#Y!r}m{HLYcrjm%pP
zk}#=2%`xDZFAlJ}JW#Bccz13M5i131DxOMN9ANdq=riZ1o;4|n*l3_O!J7~k2UvXp
z_m`%}t81KyT@BQ9Jl$k*fYsXcPqJ$aF&KS!V*QDpH6=I_By<;gBj<%g?jjV)ZRL(S
zox)aukYS$p*d;D)NRI?1`6h+Ki;SDCG@g~AeLx(94`RY%#vS*(@9P5Z0&x~CYZ)wF
zM&99DWZ3f<h-0a-%MtPaE(dG$!KIL(NXtlnc#(CBcJ0do|J`~xD3%;+0*m2KJ);k9
zY^BvicLqLf0*`~_WVtcC3I{333KXfyYBbrY9ds+zerg_y;?;7P;B>$neUJtuh$&=B
z;Y=ez0mJ|^_z@B$VvC%0q@t)8I?__K2nk-@8*8ZhG6%#c#3TfXXs~$l!J(@Y1Nt~f
zKo^+9&*Go}dH{;FEG-gN;W^^4+4yI1z*34`3UZThwt|Q)q$->#;_r5iWxo96gb@j{
zUusW9k`{6>aWL1BBhAr5M@~$qWjb>H$k|3mJQ~je-X;m+)A%S#C8Sbxx>U<lwvN^i
z)*w1bRa;r_=P#FHNf3k22-d}x$!8)Y^*OkK?ox69oRiL_2g4O1z1bgi&*gO&76%E*
zLKf@*2Z_icgdmxz{j+j^1K@y1;30m&ehGYp)@ieIR=qnD02T>x!lke!aHlvCLh!*s
z5-*fE999yPZJ3R?%gW{VS$6;M9_)$7;n~5Ka1fuvM<}=B@zXrqz#Rh;6h;Um!;46e
zgdjmt1R(-NQi2pkGJ*`D!mlGVF<%Va0=p0M3KOEtc}Wo3^IBx-+gHK@F*G(dro%b1
zPolr`RoOgOp9DovBamZ|ASqRfA{kYN5Zul{aV+#QT^txf2C6s&L;<Qef>s21={6>0
z3sI$_q$rW%fSaMd(FaYC1W`??G}to<GBY)!!9`m6HZ`e`VtH?&3Rb3wVJ{~$!*P<R
z-NT@s<WBMxuoyH5Ck2W!$Qek3s2Qg@KK;XH2K3Mn%Mfe$5zyk;S@t^0vB^QY2not@
z&q4ZY?cVsRk<G1^U@JkaAmoVWa=$f*)YOQ<1zGd0#kwSC^VtZUcl<p@F%h)_>{rcF
zTf*)DxvM=<BvD7e1Q!C%=!35W62w}`66w4|^P)rHM?kjPWxK<4;aC2aD#og{3wiG&
zl)z$;_dX(o8WVr^s41bCf)tSx<VZP?AO%5zA~ivc&{c6|ZbQK6&!Cm(aC0EoZZr9L
z-|mfS3<yt-c@EOXI&E`2f4w3q2I44}r~r6%ug$i-h5Xf&1-~A0B*wt6hXO?z^b905
z2Y$b?p+#_al!Aj&T~b|POMtkkz9<S#Md||RWT~rAx`b4@j+AVr4T@ZpD4R;I5@k~v
ztVG#V%9JRZDz*w`MdhMGSy6FSC@ZR970QZArb4VdptsHC36|%9s|sO<Agu)Crb8Bz
z&c9FIZufeAZ5N0ngU#@TQ{Y8K!_*hfeJ4O91vr5bd?4@zh6_NE)=-P|l((6CA2*))
z=@_sCF(JgkOwphF$6Oz*0vDTNn}Y1(z4XMhP50$D!R8=u2ssZZNY51Hs<i2u;I|j}
zccxWlf4(r0FAR$>edyNn7KxQPM`Dzm6cZO4pCDHxVid3%gGtm)V|CLQ-84Zrjn_@%
zbkk(rG)XsAYE+q@0+ggj=;I9VOAHA_(o$mzIn+cLWM^U&YHCh3V=%eCe*QeZAi$O8
zywb+oVcGwa(k(1MT5ip@<=8tqt@zl@-P6m*>K`l@k6;>I9F~F=tqKg$3&;EyLehPV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp16.cpp.267AD0382A416744.idx b/.cache/clangd/index/elementwise_permute_4D_fp16.cpp.267AD0382A416744.idx
new file mode 100755
index 0000000000000000000000000000000000000000..26277d5cfc1c9e49f257f7ca6c0e9ab60dac74be
GIT binary patch
literal 3922
zcmZvf30PCd7Jx$tk&w$8NLT{OE_(xnkeEUeFondhgi8~UNDvJu1O$0D1lp%)K@@61
zrQ#D8aNnp)TPtEKsGx!r3Ou1wSCqC0#)ZYIyn*-b&Gq|UzK>!4bIzQZIdjelNTY>9
zkvR?*u_9WdURjVvz~OL~;Ad??{;DU2I9weWhtuq7PE_aq7GlxAZ{Wzoj!@mDb-b{v
zXPWJ^tQy2s+whV)x0d4WspV~dzg{FdTX%uK^^n*g;>nP1Y;Vy@;@1Xu&cZdcpX-hW
zoOq#Ywd+!9-(`d~HSaFoJ-mH-O>v86rebeEtbfNhw5OgGx{*z~yH_7hgfViCoV)Q&
z_{j6&#-z({i&<TbHCt~vGhV!Xw7fPTTrt!&9(M1sjCHjz`nm4SMT)On%{)8!vZUz8
zw929STR%Qn{q#^q%b(Y^m!-QJyY1j?n?v5af(6&`sRw4X5_j2GO+(!+LyL*#DPC`S
z)Ri8;kR;CHxZ*EwdiJpY>X{)9`S<p~<r?U#Et7G{M+&9YLgMqIB8~&ur~V*`H~EWZ
zm86Y2{`=<|bCUc-LCXdrIj6+Bx!pd0I&#3O+-T|)R9}i0J0<2G{`vZW`x~wrnOt);
zKlrrhK;;fX=AOO~ddCCTG)U^YYU+NyasPDl=K6-r^V~yM<_$G1X{KMsC)FC?m7n`H
zL_0F-<Sw!&bUOAuYJWSg?n~3ps%g@%VT&6In*_F&m#mE~FO$}LFBneFcaPRZ#yArl
z)|pNeM^wd&tQKfmzL$s$+i%K6nQd8`czW0G5jPGE2GO+z4rkrScz06iZ)8D1=|f*$
zLZNZ;@sTxmqCZzV_RYWUYE11p`hCiazY~97=Dj*NEBy!J_jhZp74wP$US5inUppH)
zuluFRWXT9Ab9F|}7XL?oXNNUe#<8|=4yWf-HKtK|&L4Dpop)%D<*;sivoMaFG?4s9
z#QwwoICaS=w(S0jVNHT3r91AUXESG(Efz0&TpzQyJ#u$TnOQuscWOAF(U9vc{mCyn
zTw-@}hfS8%v%J<VTbZxjuSli0))dxk=-f|Y4tS=+K@N`5<UjnAUinwfXs%{*mRu`s
zxH7Q5#w&3wFmdz;$m!g_)q%y2HyICP?r3P)@$}TIU96FUgRkDp8t#R(ExSM5uxeRI
ztBw~exqei<>ztsqapTr2c1xpMR}UChDI|sWCy&){Z~fy<X=7eY<=EsGsrlM9ykl>j
z)^F%Me4z@T(N=V8Y_MhO=VO&;U)}CpMXw3YnEdYCnd9fGd2esuiM`$L(>!5rOv!p<
z`sd8GIPJodxBRYTf5SSlyM0T@&Lz@{chgC`%4cfEic*i2j||;yy1n#X*}uZ^_a85g
z;Z2;hJy@OAcH#KR?FagEM?dLZ*w1dB$Q?Zr&>&O~ja7c$*I#z&#Aw&c9kNVm)}E>C
zNol9I^9bYo&GhvT*lKTw!QtX>-`$$t^hL#rlA56DX#;X~TzA&~puzP?WZjKd6^Cn|
z^sJb{Q*r}Z))uOlPl2kRAF47YzCZ3s#>e4sFpWs_m>Ym*?(-3?*Y?p^eiDRB18P5l
zpYhy)sJX|x_Lg1TevFKd0cvZawfWqDs25kh(5*iGR!+vp1GNRwf;2ZEYW3x9=S^rA
zt09~csGW_RX>$XjuHp*~QhMJnC*vhRO(9au=LSSwEj`jWG4ew?87~EDOQI!dZa~!E
z5?-r2nC;_ad;(BYiB!wE0a53#RITNaa5zH-`PjaW<rM@7W;56!=zU=hgM%TNt8AfO
zCwG;jM3{b7bJ$nZWB|c2ela}6$l^yQb<obJE+A$RS$2q+c6CdI%XncS5Q{y;K8V=1
zlcBwFl-~=)**4kp5wX)}`+02!Zck>dEyUJ{c)r;NcDZrK10YVIBrHJ1|9eWj$p=pe
z!4kTJix@dy%+DNM&n-0|zzJpvmWUYnZD8`j`*NIgH?xRoIwAp<naL~=D*`MxlVeC_
zroxhkN*^6qwMg}`ew|9LN=67t9-_$yNdv)l0y{C1mk`V(GtrwvFvo?1Mh$|Mu1a6T
z9R#b*)wT$QKkFMieRS_S5F@T77ma^}@a5o+3%YJY0vt>YrXkh@m`P<~NNul1!y0*#
z{Lp&*pEBSn#UsTFS%T!Yp~(k14Z##EiVc#w5NvH_ZG&7?c<IZr>`~P|5(K-Nx>C^e
zhutjQXnJzDad*&@hn>fKJq6kXIv^*2;2@hI3@JzobG=SMSGedYm6VF=M72L{yS;JL
z6c53+TwA_A(so=slq9U`T&iEi*WlLFP-+BHO_J-s#P|6<Z^sj0wi(+V*?<6Z%s42)
zF;<=<Ra_&mK_Dp*-4~pf6ok^M{I%<1r;hM|5u3m^MLdBQCU8(fE)G(9wml%l3&A<Y
zIp}tIny@bV&7ODQOo48J?uaJ>9OM>+lBV^+V=bY((+GmYNMUA(5d?>m!ZE}r@i8PI
z2{0ri2~jFJwM3m**WCtAAI1!e#JFRIqg3{MF`eC2VGm+xVPQc<GG?pDaMQDGGogkM
z%%||tYe2AoBEXQ4B1CEH7XvK{L8lM!Wb0zfLY@psGf+XZ?NJ{dTY)X6f`ZAxm<nR~
zvY7hwvR&AidP;RoMe7Oaub?H&Hdn;?&Pk=B#GKwYTMnEQA>!#kwwNG_FMS9KOXej*
zU4l;GCC5;PSq9oE>c?%o9{=h32DDx-o$HKv0xi%b+fz?@u6bxXf#5tw9@>@<ZkE>$
z|Kl<p9ED}iLU%c-X)=ma*9*bhIC32$kzhfv6W0l)Q{)rhZJeBX2i7LDlf}q(jb2X@
zm+Inu!8KI&Dmo$t#83rcNTgbV5K<ZhlMlXNAlN?CK0?2F?8mWD$P$ph9yttzzP%Ur
zS*>ks=4>Uf*djy>DuKm8=}gL<!1@YvE08oYk_;^n2$qxN7*df`C|wZl)STu$y#rc(
zUT7YS_`O-4|K|0ty-INNc)NJ?xo^7`Z7vB*FrVFkA;V;doFT_h1~miC%I+sExfjEG
zpV|^&0bd|MHUK0Rpy2_ODEL&bi{M9KEF;7b7|Te0Bt|FX3$Y{?Vl0J;LX6KSeiX(s
zN*sl;jOIsUbP8XEpZ+Y05CscEJOv*^A_bZi1UQwMil!Z)@Kp4R9#C{D`b>ZlQqfQo
zU?o$DS*HwFV%906m6&zP1SPtz->J!Hz-K`Rc%nR79{No*Nd4)(Hhh5=BnsDs>w%mV
z6wJbGUxfbJetM$Lr)wX0$6!LR2_gpHqXZ^~)W&MGnY}50@zt3BU@S<K6t@)A?>ou9
zk%Io@>tMV1ka)J<Q%X<@da3-C+2Aij@ZV#8`r3`N6UFTC-_NJW1cyu3PgC^MrTS^2
zeww78%Joy3e!2{u8p3kxnO|vKZKugHMWXCec~bJy6kNJ`MRo?T5v^QeNH8MKGbWKC
zQwwvmD9cEJhowoB6^&wTYZuC1#Qiudoab-j!*LB+IRF1isSb3<1<o#R?w(%Wz5yRG
af|x9)V0*v72pg&}Iu*Y-mj8)?822AY^uZhe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp16_2d.cpp.73F9F8FFA71E623F.idx b/.cache/clangd/index/elementwise_permute_4D_fp16_2d.cpp.73F9F8FFA71E623F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c5e4e7d9c29d2faaea391b2dfcc2d02e3388886
GIT binary patch
literal 3800
zcmZvd30PCd7Jx&zAR!?kfrNbx1|o}~?2r%=5+Ee9ToH(@iKwwqG{FdHZ54MEmF2ku
zg0@PLqMubkTne?KQk06mid9q;s3I6qR4V!g-n;k4?|b<^hWXE#Gjrz5Ip-!si$u4`
zIGkWnbU|)*einqo;i%wGov%_p(#PR=L>#W*;DyB8y!-y9cdPIH@T4<PdrHIOx}3aV
z{griH%<i3#w92K;{J8c?>AM%JDxL<eD9(!6gWuG7esk;Uk&RdJHMr&3%7bx^tqrs5
zUq5c&>XunF8B+bt#nuVm{-iPOhuu#m=N4WJ?e)&6I%Lo8a*f<J$YNBFo_aEtuX+16
zK(=O5@!Ot^uLnP>n$KC5uOHi4GBaasOY4EYwat(B1RN<Oyf66Xfj)7;dBYUD4YcsQ
z<QHd#nw#2&*OMhJ->(dBAief0D%o_R%zvBGU7GNI+t^QG$~V`EmqL=yaoiuJTM9CI
zAJptP&Hpi<AZoa^!J}w&&8NPazALi(vxt;#SUVCkXtSH5{E-Ry&OC9Vm>%6Ls`+MV
zvF1$Dja8%E*3FWJko+U-7?n-&H}q8hO63L(M@1N9{aDu!p2aS+>U7KDh^d4HZaZE%
z|IzIIomV`q=56B&ZMQAg^y)F2BW6w$E_GvBl=yseY2D6jqmx#t1L>uGyBc!vS%)6@
zc^M47v^TBu<TFF>^w)eOUMSx2O6KNCsB|&TIGS3zF0M{c^Q+)^w_8W^-L|E{1smWk
z-B;2|iXSE{swk5fRUJz^a2wZhs+89ILUy<(%MM?`KX9Q<!Mo%XBcrmVz8~$kTwAye
z4i`LgTGm|KzhdX6-opB@gRzdm9Z5gmF6+yCJ)RI>p0@Dp&rS4)14Bi<t39^RvMR6B
zI7b|3{iV+~?f$E)t)AQePY(A!^`z2drR1l-YLnYp1NY0yeSfV~ixMt}naQF(pXM*g
zHd?p;?y7>wDn|E<>m_HpSY<~uTRsae%y(UFeyXdgBjV!5cbq!>G3||Ka#3IEku2FC
zAJ*K+%s!ZEc6!jn-Ra*wdm?TOiLREE<5Tr=8?JG!>h?E3B_#G{G?}lt_-W8gbhdp#
zPN>zP?MwG2Jo8drS<8NOxZr_H`;74IyQC~zlY!8@mAr?wd9<VNE?G}5Q4br~RZ9x*
zmEQbt-1^!XPfI`Jv3dTl#)fk8+ZObmlATRp?18CQ(!8GE>3=>UYFIJvg8g`7#_NB2
zX3Tt(am%t<zRkvSh33ze;Z<HaTh>n;@1Q&^^SYn*GOhl3t&8=gIph8CxcF+jpzZ1Y
zwRIEU`cycO2Ai4RKQLXAcwq1IIoEtQ1~pU*i|PwDJuLg>kyn5j=kV&$R`us|W2=p9
zIbHiY(kpBIgGcr@#5V?UI`W2>bSy94F?K#+Jn_6~VYSv_aBI?1^|jJ98g<>`%8EDX
zL1x}<?<-haOCEe&b9(aj-}Fh#tg->u4<E?H{mI>zzc$_<CydlaD`LMFGEaMcGAzs{
zzEKzE%I|@CcL-c;WcaDxm57hS;b6KUoi#lG&GgUPqQ|m-M3q9qWdOA|<ZU!PAnMTi
zn^l`z%8wKA^MIOWNTW;-h+4SgwRTzK$7CWt9;hjX6vFg?sEfb;TbNTtm`%c^1GSTZ
zlg0FasCS2p^iq31$%%L=P*V-5l<5IcH$iW6JN-IFiTDJdHa9e<P7jD0r^;Te9yh|_
z^nHlOt2^b}Arj2@;rpXEf(1ST9ZF%QxM4aOQ-+dhvwiN|%BY)qB$#K&vqBbG{_azB
z?O1pZ5HF`McR<7r<aI$?jXEC#aiUqGBP#x1obkqA#Fqre+Qx<=i=0|iONYM;-J}P>
ziR45oB1Zn}8GjM{V0-QLuMD%ah7g=WPBBGRL9mQ0(;+2UsY@&+dNBxAhA26hoibS|
zLkM}<8OC3{C<(TRu#h52LxQ<Pt~p{tf(6b3cZ?KFg$G6|ipmlrwV#@cP}s{mBaO$Z
zG$3L^st}E@Ks44{aZcN<55XJ;#{yXm!CVGchg8-oG{_MR#3$yX-zI^b(pl+j!~{u0
zed90Ej|5X`R63GnB$!5{(UEJ5S~rg53@@%WCc(_vOg}VJVU`KY3?o;ns|7}GbT=D}
z*j8)@jDqRGHi$nYID{UeLz#q33Z~1n&2+{{PLQKI@xDttyNZWR@FbYQWeBiWu;5yt
zBn>FbV|Z@g1!7<xXf8miMEcvU_&c7jI`9z8C-be510YyH7N8XCp4W7)Ji#5<!d5W)
zc=PAcT!(gt8t8*yFqv$fOA&%m=fI^1At)(TYEA6>A9%o3KEyXcjsqzI2~b*5z`bdE
z{f|lDQ?N_$Ok@!Rhq#2GRM7tTX%W=@y8#J~Bt(*tMI<<i5T!$6f>?(n1c?qw2~w2S
zpOEAx9_qdd9E<dejL_N6FA62(hM-udx)Rb-NU+*SjRtG;E1_-B&F$78SQJwV14*(H
zvHpgaJ3j}4V2RDd=xP!yF_Y+!)J%#Jl7gVtrn($<93z6EQ)?U|N2k_!9y~wHd%iPY
zr}E@XIr@0OX=gR6t9+N4ww_{pXR1Pi(?!Szd+8fLcQFbmK%5$s8iR-r=fptb_)Skh
z38t`9&`m&xVas%A8F?AnVQNR|?5EF}6QJ7)ZG}$APM}9RFL%XA!_=T{hy-hVG-xM2
zvL$)nvy~TZfh#;~9(qbc!FhwY+*%QMH#@GK2+MtYu02XWu?G?tYU4eCze;PREpi+n
zALU#f5-TMLA@yNk{KeOU1X~7M3b9k-Iq^}535aO88r^^d=iB9@cWLV5uG(iAyL2k=
z1NdUhL=YfA>167kb85F!Xdp#W2q|bekYE`>rb9}C5~XvZ%7RAT;Gf{jpb6B>NA|t^
zK5&nys$UL*73UepL&WV7Ba29Hf7u8@FpJHag`_hCyL!5MVdUoHHU}fNAKM?HF6Qbj
z>$xyx%0r(NojjBI$aH8KV;P#n-2-iTEnz)_BnZy*$n->90py<<phLXOa2*n7qR9rq
zaz8nm27tok=!F19%hC7&g5_uoAXwq2(6Lj5DRk@<(Fz?q1*|~r?%JO>xaaQJ31X+Q
z(4gOtUdE+QMPZIb;01-wLKfmHNMN_+Xi>c=ZyZ15ezO`hB@V<f#ujlQt`4b;RA}P?
zS!1KtsRy-}+843Y3U52qR(+HJR~XKWASCT3w!SVM@#>8QkDl(5j=C}_tsct1EBg(2
z2>Zi9$R?no=cc1^P-W+2O*Im=N}*mgHA$Wt{`*ToCOCcA^_cC?M`!zgAfzswzaTLw
zIYov`MAf+Y*i?p1ld$OmY?_Kq7h+RI?xLI}0L9r7eaOIYh7o~CGBKr)qo@&l7S%Y4
zW?^P-X%!e063XF*@x18n!AyTQhyN#K*x1@RIyt+{bY;)-nC<N|*Dt`{*V=QAfX)y_
Or{RUM;gNbVxc>l19*r&l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp16_col.cpp.427D8FEA072CC243.idx b/.cache/clangd/index/elementwise_permute_4D_fp16_col.cpp.427D8FEA072CC243.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e489d36c0d2cf308284bafa61c9c72a66229288c
GIT binary patch
literal 4752
zcmZuz2Ut_r9?uPzNXTUb2zw?B*+gXTBp^w^2qYIlP&S2PQ4l3jiBQo{ks;BFiUVo|
zcdes!QMC#pf}kkHCyG`;tAK(OeQMtg@7>||y?kHJcYgnM&-u?2%a4dq=#fa=41V$Q
z%vGr%iA16!|H4%TIiJ-?Br-%I74NwYFVAQC=njV56&<=_bKpht!F3vdtoeKJkzL1v
zTgBOHTtA*Vb4hpX!J3N~HvSea***W*#@{@j)k}x_GDs#HoHbwbw6ETJ+2~>!(x8Cs
z<69?R+<rcDOIe+LZIX4PO-**z_q@lFJq?!Zt_2G!x%!j4FZ}*XeN@}utvy4ZA1d81
z4|dHwPb*>DQbzWT&@RsuNPRARlHGPcQ8Bsg$z-3zVdu{sCK(dlUo~=X=bj7eE@#Nh
z#3{Y?1L{BBYmXmw%{`l^?5aN%>*IHs^Uu3oiLRF>9$qSc_%md=@%~iwIjw<6htt10
zg>JEXSra+lAZq=3>fvI?CH7C*GN(cA&_z0*pTMx$?>3re;ZhOnO@72RQ~qqRq-18|
z1<N+&STB7}ZfE&c!89}b=5x=?1-8Sw-cB8w1und%wVr<WEe|SeyJB*O!fh|zp(S(q
z8wLWNJ+1RwXImL#V!Ks$J*#Ol)uC~8u4z+Mjz-UvTbz5FRp^ehKx)Gy(+A+_t1-L8
z1dtWjm)`xcYVb;8v0Z^3->Jgja$c;TEHt7#v|=;oK$p$)MC!4~H<1f%B8<*?O|KKy
z#t8JRi*Kwh5~wxz`U}2lI}*NkyxS(!Mmfk1q)dH%>W2dIL5)q1scFw#uh)Nj;H~ay
zanI1nx=kI@!y$XBt^FgL1`2MD>qytOG;d8#IQP>@`rwbF{`~<7y6RU?IciSb0b`yn
zyaSSV4J7nmyfx#tdyWrtS7Kb-s!px%7)R50?NI&@d|c_-NdDZjfBQYo!ZYv(-R7M2
z&iPmF+Q=lbnE0E#bjO|3YnGk3crZva!A_LHj2_}{o~xGoF^#>ica2kQt1CYw+Qj8d
zowMHS71v_h8Sm__^ma5CuiRF8Ykw8v4~KNI=Y!&H!yU^`+Rx1YJi>~&n=14+*IGz9
z9o14`)RNYkF|qT~H+4fz*1YxzR+VYoULPmx6Fs{(_5NG5_466OX>xV7{l(_ZBT4>2
zHAM~Vmj=TP^Cu>s<gK{6qvMpIFE;CYefss}!!KuEtSWMaeVej$0v@~j1+U;ZehLVs
zxb1$D$lo(I?soo2p;_bp9UZ;9eVTd}-|2iL8T88BzP^r=zB)fA{_*{UusbE0ujK8|
zoP+-ErqwxrfDX&AFOo&<8L8TT_DI9_qHAxxe~pUS^Xa|yy_(+kx~ZVBDa*LJm`kfk
z#ao`d7+yj$zcfH?uj=C+l>BSn0pW<`XlKB`YZ)DkLvnEZ)gkvE!=8PdZdfmI;|)rt
zTPJ@Po^J7~T6cEsk+g%#sT6+;kE;_(@A@sT$?{X5?!MR=T$!=W(D%LcTlMhf(d!=h
z9z!n<7TGs3cAsGfn_ca(_2m5VO7xdJwyRLOJk1**k@Q-VZx5Aj)ufRnBob^)G1i)0
zkUlZ{v+~W-pnN{;KqD_g*rpUy``HC$dl=m_el}K+0Fg2fb`Th(F}tAbz{A~pH=nO-
zp^+CNYzBp)GrOSd@U3s(<R1UHm`094*mMe=I=i6k+OP;9>E7ow8aWza8&C{%W*3yb
zE%wlf>CtoPG;%D$)}!cAXBU)x5`4G3)2n@)Mvg<+Mie9Z?1Hkd9T4n1a8%b5BBdj2
zdv$xG*#%`kn6vlC?6Wp<8d;36O(-U`*#%`6WM&nbBmGm2O>5cLnO3O*!2%nBJ?0Kr
zXeV^SQIunp8;+t~qvzpB%ocm&NWzv7kd!5zi}PeG8HUiXftp{03xX|uEkn?tLGTK{
z6`{C+TEMFo2t={b@84B2$E8vLg0t<jol&y!ZMN)YbJ#tETn#X)0V|YYp8xjHrpE(y
z2qTgjsf&^=*X$2%18j#?W^}U+P_mVFsb7Ug=MX|J(l4^c$p6bHvgQ|g1i?AxIoQc;
z&KG?3YD3^A0ECOQi|D8>^KkzWv)G%)h)FZ1nGbqPH+gNDf6MzaM4Iow$Gnbe0h(W=
zq)$_@Mpn%GYQ%_%uSpOZXWwGQH_8o8jR@J;*O-gC*WdC@$Jdn9Hxv+-YD=*@gRo3n
zMnDQcTLFB5?G*qPD-c!y0SX`pJ&_78Rse}8MyrIP`9*a=usO$Eh~5%{!)f7|cOh89
zlwcu);4*+!2DqRGA-D$6sR4{p9Fk<zr{Oe_gIsQ%-#Q^ie%&{Aym{|xgd9(gH$lnV
zh{*xvFK>F(K$vI9GeXHA9Bvp+KsA7A4S>a!ca-+jU|b%L*ebCvaYU8S3RlzoqOu{_
zK+k}I);9z*^cW1(3lSH~$MRlf?W04m3)Mvjt5evO>PkRvR5t>0r@9l6hn9yvuE(ac
zjc_!NG0zM~o+h4FIN~rkX6VTwn9JZ2kU&eIkMo3dA?6AQj$%X+B%@8Dv6_KkIaRKM
z>&eaKOdOR0)KUQB3hG;{yS{y;MMly)&@2oOym_EGhN6A8<r}(I)*ya1@-^b3O)mPz
zotQq?iFPswi?zk3=qEr}qAkJD!l1EjX`Q;|i0(?imBFZxc%h!dXl<Z6@?8#<gB=HP
z0F{fOoPxsDqM1XXh>#c*YoWSG9)c1Ktt<}iHt*>Dgq(t7%dta+K$vUG#Zd9Jp+99{
z&uw)GPM{`eqe2jzNKGW5Bx(`?Eu}6cpmb_FhRTjbEtfR(v>^ve@Jiqj%y}hZX!At4
zxnFmsDROhVF5M8#$}Ix5jjyZ!@l}K1B>g1JPY}FRe<=Z_>!)J~%~_;YRRal~2S~^`
zW`tVxGVrp%^NL5~5h_>A6cZ}96kwI&m5a_aq-m)J8ft#~-zmvuI_7<<>hdp%Krh*F
zPPdTx(NT<azDRl`_K5>|LQoL1=_yjcQb#HFEu?8V$_S_e(5}FnR^vFs@z0m64@g)<
zh9W{b$D7C7pdVC?1ZFkhfMX@VQsSKsf|USU31AcKXz>O0Mavr_kl-)_nLIqDS%EAJ
zop2nMEO`^-f*e-?m?{8s)Vd0@6@VuZ3lu;UiqYgpW}s@kqt1e03ttN^{;hn!d@K<l
zxE8Rl1=yI0$H8mrjYVIn(%MHXK*>m2`$#bKQ&RuD#!4MMBu%7LDVFUJEThT@NC8k4
z0ET~w*j{`*bfh1NjS}!x0%Fwmg{?Y!H|W+}29YO&g8(}T1Ph&oUZ@b#Q{1Eca1`ws
zjg60Mys-@E-Ia0>gaw|0`KS;G3w?yXIEo60!Uie`M~6gX4*@hWkr}Ld_uzi31vF{_
zEY&-QlPp^LXpfN_#LSPuj@HFmt0)VO(@}YYEyc))H`p=)ssIcta9cgYSM$$@+#9h1
zVUd}L5Mhy>2zxalJa>@?egl!8$e)1tA}k_^WSj`|C<rgITSTz3ENmGE*GmaYAtGT4
zmOmhz8kS1nr3zDtJV`2yTgeE^AV_8iGYFCyk_>`mW>_XcGAk^Lh=f@LJ-LhA75{?V
zTkb<Zq4F>S63DTs4Z@{>S1Dd&DjZS@V9`}!eknlkRVffxisv7QJQlpli27wfNEuPT
z4B(d$^~->`GF<<@<puRWoUJ>NUZez!l=u_izFj|;)2}s+FmhdUvGbrWEaVBGdPd{f
zX60)YgjXqXA`KX}ewKw;%aE#!XU1cZRdt`a)p)b~tM2%8gLC&jBolccPZM<)@=^lB
z38)6pr~$BfG47vUZmPUJf}FC@q|gOb`mi8)r8Hb$H6i#^KGyI)RJ^Ska~lvLHBe$(
zf}P}3^uEzm_Y-d*2Q2h1B-Ck%X9?E)3NrIjRktlHNH45bZIV>We}7BR4T+S9Z<F!u
zQhXbaZx`d+S<|yn4G!8_o0QmpsEwAgEHyr1apIDs<fWtpTs8&YF2lF!%QNz_5h%!v
zQUlc~b2O+lNK02oJCe@h*wZy5^^Eilj7@y~0s@1ALqgpc&i*X#xmN#=YG`I|VQs^-
iwR3QEa&h-ydwTg;&Np?P$7L8s@YBfQqOb*kkn|rD!O6Y=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp16_row.cpp.70A9A9EF321166D7.idx b/.cache/clangd/index/elementwise_permute_4D_fp16_row.cpp.70A9A9EF321166D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8641b5b73030408ec5a00943b20fa046e7c02552
GIT binary patch
literal 4210
zcmZuz3sh6b7ERz12ni5=35Ea?Ac}yB0!e5{2ni`6i9mo0Qb8eNA|?hw18NZs{uON%
zsuq-jwj$uSRxS8h^@mcFLd6elwJHk!K=3{P+Bfjt++5dsSu3+<pMB=cnVCIvu39dW
z<<STP$!vMCA$MUmLLd;F!M|`}!Q9u@1i~*A0-?D20%|b+5kb3Nb8GLtcHYoQ({hpj
z;Zw|cF1uA<ZAMf({m;7X8@jOR-IJn<UfzPTY*m%RWc_Wg!GWLCBCRhG^hvs)0@^fm
z>yhHu`Rpm%vp!zd-sof0%z2+9Wch^N7VYl)bvNb3gIkZBHG;?-0r&8U)2fY69zDG`
z;1v01$&R51eN}wX%!wRHPX5OQhkJib=sW(Rsw3_F`}Mp@WMNj8$!yc8FK-%VGqMcn
zn&bT^CdUk(99H%;AI@qD8xD^w`*wd_7e86{c>DdJf6Y9&ue*(W{Im2E{9pgh(y8t&
zY>~P4ujHiJ=nls`bA433=uuJGe8z;_HmLyl{6fp5rp8#cT4wXSo)YOv3973gN5A^5
zI9JvB&EV~lVkdHX<HGYRNQK$vEvGIq3*Ja?xvb;A-fw@z%%V6iGIClk*1D}Y&yNqO
zw%gOUw6?qg(N*s^6o;*GS{Yi;l%KO@<Q}>@80+R0o7|{jep$bC+rDk*xMLNrVvl*I
zpUS->TWr^%q+0r>#|vgHi`8cPRlEGW#lKQ&pLQ{2lXv3&lY7rgY7_guD%igg-Bawp
zmsaM`LRxyd_R{zOGEI>E2j_cY%G}_9qxZwv9ZuT8T3Y<z7{{n)hv{nm3N>=>nT>Cx
zWWmG4<i&Ow4gL8y<t4dZ-98un?C4$f$7T$_Q{B#*kQX*jbDVVS><+g~+oh9U)J3OX
zsEfAk+~D%c+)vi!X|snX-hZbTwm2V}AGS!@`e$AGInJKBa{1%?@qE9w`!k)(_t1|t
z{M5B3TJ<DjF=dzIFZ;UI{SeQotDV(zM{;WT8`JpNm4dI=g{^w!RY$EK+VCJpc4uGQ
zhhzR9!g3FFp$FtgTOERYZ(Q4kC>k?XU6zv~$~Y;rYc?ib(J%DBQdxDVKa=^-(Av(0
z+qXD;UbdAOlHBm8f9liYRPQ4<3{$?Ot+czPJD;R!dx0+Qb!d)MFFA9gUsR=;^!&B1
zZNnVi!>jXckIbDS+j+LSGW>lc&$Iea<&-9MdBv?aDaeAVh9yJA%_m*vRVV-W?$qEm
zmZ&f$@B=rR6j60gCa?AynAp-txnK6(H*H;2+`XOYUG0CNS1%rEI~`XW+A{TMqw(9O
zvAQ!;Uy6E%DtEo>TXi|Sgw&l7TCvOd$%%maXL;IO*Xg%8+XC1BC=yY+_l6Domndkb
zTCdGcIT%H_%`QIq#*eTm>lre;W4(B;yGhyK73tpaw*12-`v>3tE;4gp{?7REVo&p5
zp6?o(oK7}AE$v&kO{bp6dZT}ZUWgBT^vls6&GNf}hAEysHXn8Gw*5|dJ9&#d<L}-+
z+u-prv6EywJ6E@7tWB7s3%in6t9;4*UK7<)xY&>rK_n1d&SYM>yQ18lVwFlDpbQej
zVYC4+#OTk5H^+|jv0)s=DiP3QNMkso4WuVCu6f)aD40$qXaLPcxOSrrqy-HfyVf*s
z-AA#S252{uo6~3m>Dcn&p}Yg{(<xR-fTodX<k1GwUyEhL8P`7MP^?sdb|twwjW&?p
zqTYM(W&d#v#YzomXOc5{w1M;?<llz&@JoXft7JgaNp$DY2GZwh<G0r~(WX)fIe-qZ
z31EyikiKEN>j&*I-%^TI3ZOkm9+c4r(gnG53Z3i;1nW@BzMA%&jR+N$gi0c??LeiW
zQXJCyX~$bQou3Xvw4c2VQ!C>-h*VTe5|4pS@Kg}{Pfml1dhxtO@Jy&^7C$Q*4xyr@
z;iUqI80Xh-WDhQ04C1D8QiGt`<4Wj~%k|=Gz+B>9;sedzMc+lY68#^Jc+*ndpxMW9
z1%I7g`(0qxxN62?=KpEJ%Kj5AP|-A28kT@>bAj&Za=~gMf@&N!&TuTd;bx<!`mzTI
z$z(IRa4C=`+JB-S?rx#ooTN9Kfwu?GBMKf0uXyeJiscjPfSJK#NZ@8gdJnbDBV`Yf
z5LD}^rNNj8s&mxgkjc@6M;sG&7zk?m%oJrAXG%9^Km?^svj0TFRMa!ZQvsJjMMV@5
zwtZAo%9dhVPet?n@`GVGD!R;RnFmCo8H{eb3D4((1?BPc6qxy0_rQVrT_#|bIm<9{
zNMtX1DozY_S|g~CE@VJ&1QpRmIJAtp3_CJu6Xl-UU}+kNr03`Z;ShKMt?fU#cvRHY
z#nl~NKPu|x;^q$5FFUzvK>u`34Vj7tIRv?4*9Z-E4rW+Th<k{q1x*|?(Z_-W?gCGk
z0TqpMkHR4X+2Ca13@iiNf{bJ%#>MYAv-xV-QwJ-cBLP!vIRH<ACx%qqEk=6C#;?Gt
z={&jwz6h#|za@2tJilaxpb|$36Hb7jQb#F<rbP{G$!Vvp0%K?LX9}TH$~2d8{a*`g
zz`O#o080lfgB*pSxdnx$gx7ncfs+K0IKZ($4~P^)3yOsuthTNXAPa%NAOJccXq0~x
zhKkSK?OlR&Ua_H~apX8h=tM>3WH}Bg$O;@%l9f1=NKV9%dB4(-y0fztBo-GQC&h&e
zmt$zn^H>(YV<Qu+oJOP3p++{uTd#b!`5&G&6;-$@uzjMUN>?QgCAubJ2r9_d{*#9r
z99aw#cV{NMP4=>=Lr4+gZjd-w9BxTd!j|CflhMzJ-6vQKzymT;3E(?j`D#|C*@@LX
zqWeg_3`VNeE0Fl4)%U;+(gtd=kl>{V)ZtK(V-faJ)D5}^_CEG|3toB|OXdqBfp?T$
zGR}g`L_aeTdr+vTnHXv&VwGUe+VmX{7M*5+bi_<CHd9k^i%p`TP6m#}SKuqPh?y;5
zW9VSu!_+T^l7d0bCZ>r6rvxa}G!=*9O-hKM;6W7`sef=&s3?oal2~TR<Y!`s52^^K
zG!zSPM_9DOBSAMJWUeG0nt_nHQVboQadS%DMkf~_3@urU>6?n`$T}P{kxdvnA*(Dt
z5Z!kZ6bCbrXC|h=$WR8Lw2YKwOU9*krI@U)`Yl~+7A8B5Oc<)g$wL>a!=WO25jLdr
z;aOv|=vto#f-1$z7&sk33MIB(0HrEX3#SsRaGolK3g<aZJPpSsiWBijk%;q5Oijea
zREyO(Pqjjg^Gp^e<2XaGA;gl4A;Q4Lp=g5`hvE$i3o?cqF})$EQDnrv5CD=Jv10=$
z*@$gCg64<k<HqHS@^Rzx<@vaA`N{d%xa;06Hn)PtwgW|*i3~Flt2oyKZc1WqI=lo#
zdPq7J3zl$DaYjC>EC=k(^NzAe6944((9e8aA{x5{h)~9sao|e9%^P154AGOV2VU+B
z>Zk#l5F$c*Xa)rt5#i7>yJgs8I2fr}#jLo}2QthalZ{1syKZ>Xz{H*epiRBM9-Hoi
zs-}P8^^A)ki|mMOzGVXaR6VwDgo0ds_Q;wF3p9mgBb^x|&3}KHutOkZVO?vq_m9?j
zZ;t&yu%9(EZF+jf7c(-m2x%BkNVjyquyivl-RYL@EK7H$rK>T_)@uQb&sAC@HY8g+
zGKK0ub8?Jxmd0?L?c-b+u5KP<c>K=<Q9@DlME9UbzliZZ|IbSIWO<GCW%~z=3!D%<
dX>#b)aISY4Gh~Xyoi3ASTg4`bKPM^({{dfP04M+e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp32_col.cpp.645D32DD67217BF3.idx b/.cache/clangd/index/elementwise_permute_4D_fp32_col.cpp.645D32DD67217BF3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d05a298ddb1fc4de2b2512f27bb3205033ced9d
GIT binary patch
literal 4734
zcmZuz2~<<Z8qS4lBqW4{KmrMftcole)&#>60!R{KSdz;kB1_m51X+}z5)(y31pymt
ztt(&^#1*Y|_gMv(;(`yXOI<1+5v?F-)wW{a4e#B>b6(ELocaE5{<$;%Itk+F=%rQ!
zf=DH<EXpa*1PKHJ9sZY<mlk|9B@ix<353dRT~JZ+gFxDyo%i-W?ct4fE?pJodgLs7
zkyWd#u>q8Acm2t_eY9)C->;S#p7EBh&XjGnm{s`5VPyDZQn2Y2LScNqXDMwiG|*Z3
zsl?g$+qChkdxKb;ckcOA8Yl?MIuo`lZ+$EIXxF_V`oN0vOFVz>`IFMRtykZTeYOn#
zRkLNRYw&^H`Q)C<qO6k7OU^x6FJ~<~`rzg-pPqiln@I{ypI@QlcNEs|A9H4<A4tg#
zdU0Y_MDf|N*!!oBq<aO7*#xit&b#$CKQa2{w*#K<7948tzfJn_;k*+eKR*%X%O01X
zj%K`A<DN7r|48I(#?ZEkp=I@pIWuyuMGA;>x|UDh-w~CN5IyNl8#$OmmbMs3;U8{Q
z=EyF8GxBI<r6no3qx_PVSeB{Vc=igr^v~FPR<-<(_2!*AdosPE*!^;Mv$ggTU*f&Z
zZ1>>u=6VB|zwJO#rSH0iHQY9~_~K+%&fyy)k=71Ti5=PO6>ZBmw{N}}G*!wF+Acc&
zL;aNC)5~iiVl#8Y%OiQ!QF)o(+pJD*blntbp46?_FeUE5nZ1`p&2fW`r3cnPyDMGy
z(pFnMATB@Gd}X>Di6+Q?;Ql?JDDZMSHW1)^^F!W9GfgsLOAR>{kdnaHCV&@TgHFMs
zCC}p$)n=*tUX<Jwugr1icf2u&=CwP%=;X5xThE+T^yNd{iz$U`{5(!k-_Ly>r9E61
zrKPs8+=jtk#;v{^KRuh@z3iUg1pV-0|BBelzwS!D5WKrUEPgp4;k#TLSU|7e&FtLw
z!|iq9vRA2U@(${S_S?04B<?NE8TTKrJUjNy(&?&Af~H#k8cT;3O55oAt`(BU?J?uW
zUB~@%4);L^V~$<4kUAyxZ3d+ssWqcwV&H0bg=%MA{Pn_e*Xx@&M_<fm|MQmKIyB3T
zvw=9xG<A>S<dS`Ni|quthaRXtPB)`>_Uc=^4TJS#S2*PryWdqG>eU2ZPJc4myM}fB
zp2ud$QuS|DU2qNGbJ53&l0aUjG$;9cw#U)+wPsI!lz)s3l~;H}yw=<gvmejohn7Tm
zj0c4i1Ghd&6>l3J@j2B|Hn4ivH`i`&4QlO8zSZ+kd851Y+PRo!?&+Xo9mQMsPt8B?
z_de{|=%yWi57u0@SxM}da}7J_uTHqNy<V7?!$^7LzS-leqhXF~UM8~D8_&7vA7wA}
z-S-QXP)rv*{6r_rIkeQIGrMcoG4pxu<ptZ#mE+yjs+!WSA;$}xzlYYRWEVYrU%yg$
zXYwQE`}1!vN&VXcHmo?lyd&#-!@G>s*W7=-F$8Y0cx|HT9RGR9ZKds1SMQRsx28<%
z<`!1#jeJqA_oX~P|L;^k#JT^?sdEUCmZ^)f0s#WS>iqocziI2u$tFqy0b&tZ783(p
zAtru~yj(Nl=Y7-2CUG!rN3?UF7!Z0EtIu{|xHN@A$cAYFC@`BC5L&SB=8kozzHTR*
z%!O%dqP6A3fY4F(W1|ZXeo7{r#KSa=NFz-Q2;C%%2GaUIXOT^0FwGz`EGGtp-k7lW
z(EArZW|K`4V46;(lO_g)J`Da*)Dv)Jglv)s(@Y|hJ~1Hl#b(L3&HHJ76hanEyG?Rq
zO$-SA%j6w<@{T)c$R-Mywk6t<CkBKr&B-mZrw|CHTyp!)o~$}E3M6roxTEHPq;Ap~
z7>f0X^}&$LTjq-)1y>P>AthIdLuwcGbd0NVsX`IrHqiVFcTpe*o<kVoGzwJ0FA2vs
zXaWM700Dwom%gfV9#N|S3Y6!b=ZT1IuXC%ewh8-SaRXq}05~Fqeeu}dwU6(A4HGe>
z7#bp;vTRrQWx(~haT{8mH6nJTYWcNhJ-@-?3P#0LRQ$hqGBN)mjwn!peF1thr&Fc*
zzpoOk1wg2RT0uv2o%h}C;3QnNg|}oovxAUZ`e>ST!R>$Q;ML-3V$|x07GVBGNCvfr
zn&rm-*#K{1%d-tZ{LHJ=FE*^2(E^KEJeCMCZ}61SYm11PqeKu=Q`P95L8yvag+qFP
zst3Nn?s@?A6$t5p*?J%ZIgt@3>w#1RBUwT;{~|jm5GR5oMQ%xfBFT}cbt#b2S&6!g
z0_gx39pHs*NP!vw%SON!!C`5vezO^G3gFAF;IELP;@ADd2itZmg~dtqBs)Ybihg_F
zaALIA6of=f5epH6P$V-FhZ+HPBY^rfazFX0^@t`3-mBWZ+5=HSG8}e-F};!Wi9S;^
zT=;wLPC5nhBzan*xd?fYyl}{y<c&i<Bp)1_X)%+5>CJ+J1VdaKE(b%tcD{}n650qk
z$mu9hgiQnvNh~A`jFZx(=yfSjtW7L#rOZx-<^=`PkTjN<p2l9|j3F&R(gG+e*>b+&
z#_HcKOyCd}u!Wd&IRXxfWI-E?ncj7cuq~N9rU)ryvhG{){oZe`n1GOis$e6Z03jt+
ziK4k7!yB`DXf?3z68@4<WRYU7)wCB)f=Tdqg(M+*9M}M)2ox2RmMxWk+#3!rQh*8z
zL>G=VP>G@?m7zE7uiYMpPa$*_x*>}|D8e-YMU@wSdsYSZUY|sPQb;LOWDx~QC8gp}
z8YvBj(n;w!lugP;k?ufjk#c+QW%ytz0V$EV?E+F!wC+umJ^yAM8@@S>Mq?t8w_akp
z=5@nAJW~pk#z;f$M1j&7={S_l$VL&;ci?<9HXS5rVV6a)arx+P?eBoam540DWv9Yf
zfy+)U;G)H{6X{!~=3n$H3dFEt*dX7gK-N~)Hb`aI^dV!u&eFc$m|elu(TD-I|41uy
z{^X&6D_;yf2K{_<p%j$FuYC$9u-Ze7o&zpg9#uG03s7sZqSZ2D<MHgJ%U`g|<V-m(
z)RXL!oRAM1TY_@~Fb%^7z{P-7ItpX}xCQ|2VEZdiPl_vQiH05LByftv!rfWmjG{vx
z&z18=<GtYH>H)SMut#<`Vy+(W!()jah($1x7;q0Xws*`=_B?wL_N`)mF&ZQks0nay
z0=VcFk3*Mjv6bf=Lq13$LBw##2Psi>B<-$mOP!?^971Z68qEX>R7I-7Aw59S11LTb
zy{YnG_~2dGHwJ)b02IjHC#gd>*On|Kz`is9ECW_TPuE-SSVe0*2f}ZOX%h5^6iDhR
z4M4QvLNqg$k0F_#4DBlyBZpO#KQ7dOASCgV_#=xzNE#&NVJLQXEZS*7NERkTUkFGv
z!_C?FAi@mV1ei4eXnyoOPjhJRCqIU-7VjUA9_@z9a=k7z(b8xpt{N4?X5v=iP%Xf$
z#rEob{!{U(u)aY@5R!A`IM>VF<mht>=4Q%gVK<QT<-s^4mZPqMS0>6)Q-e^PTO4ke
z1;Papm|li31COK`XkdX*rZ5x7WlA&goH7%_c2NmcxRolY3b#_FRN+?U2y<{NbA`Eh
zB+bR?X}mPv*cUW`njjns*9dV)qCwj^2x)-;EtY&n9Hs?O?;5dK3*c;}1roJbV1n=~
zB0z`N*8yQVyuJ<)>+t$IAW?_u-*LMuinwd>9uDk<-V4#sA}=hs+!>$ISWr#jO$ouW
z6yL4xGtYmn3Y%I5XJwLe66!!>^_kv)71CE<AAG;v^X5)C;6$Lv9I*=gQUW7!s1Y!0
z1kheG5}aMbHe4TsPg!PH=7lKzH79h5I*MWJ9%7OhEqH&`j%^tBxi5!9z1p=JJ;}Ii
z=Zo??son4a%L2=A=~(SojTXPsoWe}wZOclt%T^mFX~yBdzxl|7KuE`?N!T<Qoti>z
z5!Y(G*EPH|&sdO@lAJm(ZGJjo9?BC^v1tl6&A_G$uxWOYsxS|RlAKsma1wE{8Hr4>
zpjlF5=#dfbbn_T17Q@=sj>n%Z2nh`f_p$K|b_tyB`2VC#j=jTFCudi;X&y7YX3paJ
Z1q4m;XM6jKY?#sFOp_?Na1J0P{0DY6xGw+z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/elementwise_permute_4D_fp32_row.cpp.51034D29A216280F.idx b/.cache/clangd/index/elementwise_permute_4D_fp32_row.cpp.51034D29A216280F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..73954ef1c93c1f688635207a3e21c9bc5174982c
GIT binary patch
literal 4210
zcmZuz2~<<Z8cyIE2?<#U5J*@gECPax2qZKRVnPH$G;9|{LD@{u#2`Yz4K!fY3hrpF
z3xZTcabN0gtpaLUga?+Tb;GA<6@*%~t;-vDZ*H#VyquFc^Znod&-^q0GB+kdCR@ND
z5F{B91=`HqRD?hvFu{LGZcf&FQv%^Sg+M6SQ;%wOy#b7S72P$DoA^Vg^s7X!C$Df8
zS{_hr-h?Rjy8dKUJ5;~%pWpMhKjG_(Qx&@{CTs3H3=UjK2r|7z&_t^z<S=HSeWwaO
zWb>wcJLlu=nn3o>ii00=0)(Qp%cA|N4F@Qv>${&b`&Q)M<WKdzc0pdc`}UjRPxPQ)
zi*^myKkK!-me|xNNz4AU`0B$AGuip4ds~0`@c29aByvbfa-KnOL{nBb%wwk<os}Ng
ze}3{b-Id{}o{J|_y#0r*gNncNInXYMmA&}3Z^E0o$7?&=$v^xy`+V@v55;Q5gWO9p
zR)2|m!Wi{Q=_}UrJ$cXb%NB9HGVe%*r0Mm`J?f5x#l*<Qygoz;;!@;Q+sUEt?iOS!
z8g~rdUs*sWCmzYYxtg>j)v)!-El$o~QQem71@FtuPZ{hf%sicYV?(9Y>YIW{pFL*P
z&z4t~ZAa95j%o|0tld=Ndx#TpV=O!KM9ZMm${{TFNIGZ5q2)VkcispbCufOm7oIy?
z<`{Ho`FfO8X|wUgNmXH(D%EF?<%O-T<x=y6hPaK6%A=QSZb~YZ&o<{AEkUabTx%G`
z7QLk9S1WIMxRDvc^j`PxiE&xpZfE-Zd9CkMgO!ZPL0ekzMgLhbg4Hp|jaP_skYw>w
zW$aS3q{IE$T@foY9Xg#_rZc>&qqP^FyxV>GN?eB;ZCFIrl=yjGq`jT-G;H;W(y-OE
zDz@7I(#E>m(el%i<c9oi;d$nXMN{*l8h_oNcs-~(D<a}WU!=h0PTyQ+S+&in!)M#q
zhAMtfT1weRyI$MA{$QkgRpq>%2P>}(@6dZ>lnb}4pSp(bP(?j7w4r`Q<b&GqkLO%J
zPR%^gfgTG#bHhUJoY1ioksnD~GZaAzD0Yv_s3?tY*5tZ2mvc||Cv!gk!94JMvKx0J
zX}oFD!MF>H4|nP8g!7O0X1w<>W1ebTcc5YWv$Ek^+}ynCH-*RBiUJx_9uBpYu$#L*
z^Ywn472AAYSUs()d-Tio?uR8l<*DThs+%Ud^VVI8>-@RqU!7+YAC-Rh+u(|8Pu~AX
z{{GDaf?b!!u+VMs#cP-PkF}K14M|>d-ThKoy4Wv2LBRdl>mS~lhL`v69=obNW!V3*
zrR($N#)M@96>Rr}H+FHNdgU@%UEhxV=Z@?xS7a>RJ?EfD%cm8L!!xs<MEt#D+_Fuz
zw-2R!MYtUHI`@z11p6yJ<Xao9i`%_4yY7CU=N0!S_n$4+cdToQ5Z%j~N%PCj=Jf`?
zji2RIkkI|Pg*B<I`W}C{d*J9d<uUEWS!WvwH)4A-0?trA+<rT$X6Kt!*KKEhyxmGq
zRF-^`QW|u?V%J$lMe|+t*3N0WUg`?p1)pBBRGSt+BoHjGB{x4>U1m-(i6;<HHi>O9
zI)Dx_`g7vXHzR#IJf32r1hgH=&V6)%^kjC2ZQnr7EGi)#(1A#x+2{ai;o;VOYcFoA
zrI^eBv=zyUJ~}`;tZaB_!LbjC6q9H`Ge`{b=m6<0Vi|Ex$EP%ki2~3p5{o`MKzeIT
z&GEPWKcrJkVgSt~G0CF?q)#A!YMcCT4N^>E0c}IFVU7-vzEK(ZZDk$9k4i`bwA&ar
z_UHiVU&ih`s5<9dL@|j2v@OY&GCDvyCv*N1x;cSh>PxAuXi6(ZsHntO5`bL?D)p7(
zkjh2nVdT^<Y78+hIBNaM!&`||R7?`vK_wU!#QckBP*DfIg9yfiilzusLZJy2E%GlC
zLd3rLO(}11=~A$6ynFlvD7J0(U3UABxC4k+TCa41V#oabp^Ze>rz6^ocq=G&qOBIJ
zH*0zX#ObW`ahUjjIbmY{MJ`a$1p5Tc0q2W3>X)m8UlS2DotDmow!Fh#N4PP!ZNZWp
z9w!hU1=2+GFY+hwK#1A==)X1rZCk!=FboR6VBMnatGud!n9XNP;AI6l4&7NuN*y90
zsEVdyz%>z6O;h8Lo~FkmcRdyi1l50~4>sEA6ZJ_D!Bi%hf04pelsk<phsU6zB8mvR
zJ}N5ZNwKS^qS-Fl-f%f8T1YRng-A4q-D&3aItv_Vfnb3g6Tj*lICf~C9*AX38I~Lp
z*&jXI&kwbkB4~(B2pei6sK`cyLxr3|EM(F;%44g+q6Dy##$DqHO<)0nYagjTnBB6=
zYX&qg=U0%a=md)iEUZ*eZ>Bffh<vPlxJES5ZlaSB39W@(xC<2xwhqQ2Em=!9a$0*W
z&xmwn9mYlOy0)pM_@#vj$Q2<+Y`kQykc%Nj;8vZDPw8fGG8?{)1XjJG;cj%N&+A(z
z2r8jTIB)|5mC~dbnh`v(HLZ!U2H4IO%ngA`aWgE(_iqu70s9KcLd+d-4014rvT~N_
zXTGlq1xgY`VgYSI%n>Pu78iuH+TUsa2)q!w3f-U*f(E+=W2oT9qbJLdw&pQZG@Kkx
zgGy91f*gTEa<UwUqR3G=q$DdbWH=h7jo;hW2pkLd50~PW^N+yL+Sg(Bg4R+FI5~sC
zuz`87A=0$u)uzvUQz|NF$+3H)qEW0U98$8B7=lfDtoaws8H8*a2Ui!9ttLAdvm=BO
zf~yj-x7gq4ri3TK)r`(Xht&*h_@?GxG<qt^vSe99Kd7jcrIj_T*7A4rk_~kG&XK$i
z(#YU0l^Qt`8U6KRP(dnB6}BDd6rO4v%BSUHouX>c+VjZ^m%l-$m)Xml;Yy&7@>Y&F
zA_LLIK*Smf6*Ulj4MglEs9u-2>)&}-?SVUDju_jiuHe!brM5~2Lgp-VmKqbCC*)!1
zxaZUO*+bFZz-K*2Zx6Qw$XD-&Ly`I@h+r{*S7hY<gXK*{?fLc+<1Wd9WGpejE8<?{
zi!JaVB!8FfO!Y{T2TCHL7$kY16hkNHbWN!$rCWleQIS<x?o&}US&c(_vK~X{W#t9O
zLZ5Yk$H7448;Egm<qNdutzTy^Ab<-p5ZMMI*1%xyfH!C4tu`K|$gBv<#TJ+4>kJ{W
z^bt>dRk)l|`>Js$-zFb3X?yyU?xLvUnJt1wiKC`LKLC<PVfP47d=zTr6k-KVQz2L2
zG-rrs;5emNiAQoJPE#4L#H|@4j=^ch$YXGtvEo=9r}ftQ7=6(OXajL5R4c}zNUhw6
zbpARl5fM};(qW$t07dAqcmfow!y<y9+5Xu$yKGT5&MrG58)ugtn~mAsbL)~!>#}$Y
zBAMuuh?&4A9Mtg0N0l*-j(o>pV@l%NagWWlPs>E(mVqRc@nr7sK%nwGR(eD9`?h0m
z_fBZ70HF^-Ld>BUJjjR$hYHOKu|_!<l)i?uz4;mNFx4&<Tj}rh!y5-C_RIvisd3d{
z+kI42^yl77Y5-oO2BZp%8)*DA*!2){GBv3qXIhezzNC0$GG}D??=Kln2&U+p)Y)jO
zhHh&<Z*Ib@#H87Cl2ZsN7*9wrP7{sO*~V#-aXQO5ooAfRHBQsD85$LUk(p7Z$QaUC
zGctv0!JyN^nbK+QO!II{Hp|M^jxYF17#t!BooGEF$R)tT>HkS>xb_a?oO!Np<2}8+
eCr$SC^AB{K%JG>Zv9^&#q?&}y6i+A03I72Np$EqR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/enable_if.hpp.10016CB3379BC2AC.idx b/.cache/clangd/index/enable_if.hpp.10016CB3379BC2AC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..882fbc2bad6afa79fc3861b7ba00ed418cd38518
GIT binary patch
literal 356
zcmWIYbaRVgWMFVk@vO*AElFfyU|<jg;^LB`%qAc`9f&LDxb`@59WoGL_^xSLAtJth
z;g<)}d)iIzPEwtnlM%GEqdr{Zi&^oE#)ls)ZhMBFd1`s+!_vhstCv19p81+XF_Jrf
zc8iRDeUYq~*rpB%Q{iLJ((nIh3w^)EDt^zKUTL<op|{n&wulKm5lWfO>3nGI=7Rel
z%WaQFJ!BO)X_`}9nVaMUbVb&7cJc7%J7l?;SeO~uIM_J&!2|;*6VUk#jBJd{im$CW
zx^T}!ZYEZs0&X^LK`?<(P?VZh91K(|A2IR!3zesbnRwVZ#5ojT2J^6SD{*VXOo53q
za)GVcFi(4b>aK~~VA4QgkPR?lhN9G*M4-_O#YM@*+#m*<)`XX~uOc7%F)#oC@OF9$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/f8_utils.hpp.AB5D9F32110A912A.idx b/.cache/clangd/index/f8_utils.hpp.AB5D9F32110A912A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3210bc0553b31e5f8154705f2620d7476fc1ab0d
GIT binary patch
literal 3088
zcmds(Yfw{16o7X(0ZF(?xVexB2?7ZU2@oLxd6$BUI991x`xv!|Jf<o*@)!plA23sk
zC^o~(M+8c>rB-CBC@NU6q7^}eAkOfRI%u(!YUDANQngNRf|tGj{O1pMCVTVE*>m=s
z-E;3baZ974Z3qAsCN0fQOU_E*0su(*mzkBZF%<%!;{cFdR#SL;OS3Q0l6imC&@b@2
zO<g%@Ck>ZI1%vi`?|x5il+RVkDw|#uefW{TNt<@~rptKeo_k3l^<kyu+62Gh+})Ep
zt#hN}_GKemYu7Kb&FYqBWgOM#rU|{<oR3#oPby9-$8<qM(MQVGeqJ2sb?m?4>QPQx
zRea0j{73U9m;czaa-Z(StDk<Ob{>6D=pV-qLHy^LCcD_<vviNw#VTtm|55Kb*iqov
zC;qI7Qtce9yLPCjx|;GmmRH*z+xSozx=M9!*ym&4FE3Y~kG=3-@o0QZjekkOS?e|H
z4@yhN{<`VCJLmUin_Uxay8N2O^S3!o!7_Neq_Cknr<JoT+hN6Napai)oy0^_M8=7s
zu&WN6KY6wQEayz+bu3(a$+vW?LrvGge{<4LH0O%adj!Uir-XblXW`19;|SH!TTcgM
z-&P8~cud8}SFW}%I}owAWZk`yY_-38#o>Ye;=;(<lnE>2E4RkX?P(iE=(A6)JF_OG
ztUX2q90)*)xFVrt0yeaa9tWuI@aq>1+pU1e7LbH3A+t<?2Y?s=NWvpgWFyb5K7{~$
z0U-5AqDQW;AAo1Y0kh1Du9bdJ<>Qff03nI(v|Umvlu~9Z=3)GM%~)O5E;j=3Xd4D3
zF(Ae32jE$Ar)>;aoq9B<+tyJ8>}f}ZTp?kZ7)S4adjI{&s&I+`LfTQWy;y3Q01rVx
z+0Z_KjO3Ke4fKxu(_^>846pv+6Qp>y*mXvcq?{-PCb?R!eO_A<9QEg!7=rZRdJve@
zLbU`_AUBXD3*rW`s26g3AyN0l`}wvKfq@|VAh(ardNfauG|V9i>P}wqwA$KeMY{&X
zude_Af|LU}iy9zkfGpM{phqk=AYi~eyXUOU%O5UkXK_i5lSan)%_TdapaTjS-?(Ha
z6m+uL2n9wqcR@iHo5!JO9NJ<1d}N!CW@1%DoUfk<GV<T@wmwNOL!HsCW*aUUt_XL>
z(MBnvShR#(LSflm$h!+|aU>6+;2{)Y9)^NpHk;ra6LiIL6O1syNQO0%_Vnn6bCE*&
zz=Jg*cpE=QKPOCDM=gtVZaM{)g-gPvn4%<6Eb4&j4(P>fX1+FcC!38>ZDey7RClqt
z4?6Wh8P@NIvVJJXJPy6bp%3PK<e86D4EuRS7Olwh4saky7iSk0<0L`4I=iw+DOY-8
znY-MbMQVjwgJnKSpZS>lWPTo)v@$J=bV?m-Wq>GPCYK4a5y~5(C*veRHbHq4o3BFW
zt8gX=f^3G)&1`OgE-m<w2(lHbT44aT*9KK>Y`z6mx7gecRqbpxK^qga$NDCyGGY1Z
zc+V$iRjV%}F4+s6dfBJn50(9JE~A)KnE%xdRZi<WnM#}E;x%_B;zzjzS%j2DNX>i(
z39=X|i`iU`l;uc0Qzys@q^!Wq)G)}m)GnrG<&7<t2X_`YA+vo=5%S0O%aNuW`7`zy
z1zO)+;m%ZeT}f^zyikxEY^YA3++^`Dun2`={bIDB7=>akM}g%ijIl@1{w=6Lp$s$Y
zX#3LzlBtIk)>ZXiZ`o4z#yqHdyrWLKf8x|z{+M1$8Ejv~Z+k%3BMMOb>1u7MM?5)&
zsR8mDAdB^gr$;O{Af5s9<$Za)8^0!uyfNs`uz5S<V&#Y4a?s4J(YH0-iZHjubRe|A
z+<q`w!?X+>zNO<}vWHJn63USwf}Da5Q*bu(Vh;@lrLMis`R1K`u``lW6U^C>nUR<&
WHUF_QPt${R7<iVVzERZ;?ZE%u^<)|V

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/fill.hpp.A302158AF84698F8.idx b/.cache/clangd/index/fill.hpp.A302158AF84698F8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..140a1fc80d0ac5b7147630851fb1590557b6cb94
GIT binary patch
literal 4772
zcmd^DYfuzd7Vd7wff<_NHcSuWFyahjGe(hz5G7zhL>>bQlR=OWgYuL|284+bSi`u?
z3IXCHYTOWzB%-)J*23VUf=1%AvWwBMMse9F>spBqDzSnwDvIvyA!pmlkFDC`$Nbo?
z;?QT#J@=e@?sv{Tw@rq)xK<B>h{-cln)Bbx77+x&;6KZo6-9581aaAkASzpTnvzPs
z4MLYw`aW6v3)<^FPgd|3?F(zw)rkW>XENHZocgOOKlIvHrjfytn|&({t~*`B3%~ze
zu&()j%iX=14<>rjg=N+yAKd?>JLsdSXQwB}YH#QM)a*H3FP&Za?8?JUx7O!BOo(0k
z;VTUn2S@r12Vx#c%je%)m+y2kxA=$CT{k?hrkVx=Lf>laI@*x5yUXYwnMXI8^UKh3
zG$ZubnPT%7EoWOgcMlmRCcPebw>LUV*SkBb-KzZVqlDfn{fSAj`?gtLcE2@3&5#dt
zpNo*YOa67-SpNI^a}z#%r^dD^C_GWSenxj^xkY{Z+M$pCd2{)mk=AGZvrRh9<DJiv
zhPdSw>w98Xr0YK)niuxU_Zh~N^j*mxTv`_^OD{_a+;zjxpZ<81sW!=tTXp^-Q|~E8
ze6Qnk>Bg<PC2fBk>CNs~=o%Ge(ayW`?VWui#%+_g<hs`AH-;pZqX@;{YJ#>TjK1~M
zbZtdV=J}%!I;IwswLOj^kGf__2uqbYCygQqsypxciYfPPGD;#NP_#6fwiChy?-YK-
zbz2n09VTZ)DQN-~N(vb}0Wq4l{Xn3m`ztJVBal*}Wb6dQo^|z4-KleJSd1rMC|StZ
z35eO&Tzxxx{4}<WwOJ`k+X;w+kJoqikL<WAqb6e8uZdrCwG$A}K374zL<YPnrbJQ#
zWjJSe*$E+`Ft`X3k>%a@!mez&)Fh+4uzsdEld%)fw>13S%ub{Bjx8-qlEomNfOvyc
zaksIFGs-9z?E75lT-r`Ryt|-cQskFc2F26_Y~91z!^chp>w9XWBYrfSys$oAoLnlW
z?F95Y7u&pRI?{&4loQssI9ojJM9@FjUEf`})XIq|F&+@<968NS_yNxL<(<z=4Y=6X
zvve%D^Cj~c#1k-Zvp#LYf)h7&V_Tz<Xc)v35RbN;Bxm>99?2+`D}h#uSJHO!A90sS
z)-;eM#@bk1BdKBR1jJ9ZQRi(t5>AeZD<zc-;t7a}iu?jgkW4@qCAFgp)<mQDS@B`8
z0KgbY*|7!tD!Dvu@rKgQI*LZ#Zr*+XHLU9v?+3c@E7s-F2fO0Y^J#0@qd65K8flSM
z0kNP_GD=neDrr=NiWFd-Mm1qIk)yw^8^+*ydDw#EfLOd99<E>xdpOHG418gKZqoE`
zz3uEmBVWe%IpZA5sRUP2n3U&SF|j5U7^P7$Q~aD`X4b3{+_Z*UqrgoX)sk#2sRCZv
zf57F*t|K!=BE(5JDHszWlf)#2=|#vaF-u`W5uQR)OG;r1b=mN)y1{|XIP5Anl^hmD
zqd2!XIfUIe>fmR?VP8wfVTapTJj-}qw#tb{YBx12*hz3pV8ITJmboosAycPbC~jf@
z^oKqN1UC4fL>gD6`i7)YwOcg<_=ggO@xa$7fG&-4m>dPT{o<y9*vpIGi+6zA;1KqY
zFOa9+U(hi_pWB3yA5s(p2ng%aPI^7B{xSAX$LQn$8XEbreoBEeGDbN-h(@Vws#4%e
zm=ZZmN25}<R4H(4$nZ7f%K$GL*+|w#DuI8=CwzL{_UP>nF`nidGr94I=Bi_izh8AD
z;JrD4ln8kTcn3oQh){e$d@v-y;`We3xvXEB1F(}{HFZbTiB8YoK)%E^^XG3OV{HzJ
z*&Ca4^z*MLuENDh9i>i!n9<1aq9Gb$2IoEoJ%8v!4&Yicec3EXO8AcpJb&uLpw(VZ
zZ#O$V{RQFLt@hOk=+~)qT7ly<oL=DkeEc|pOHrrjgv>Cij9P&+X-s;7OZ7?R1g=<J
ztP`@PL{*{{IJ3sA7r0WNQcmEkvDO8UO*C3V#;hR|1-_Ql)RKCEw~;<Jk^_F-txwBS
zqlY&9*LXrFgyT6Ew?S|g;)8f*d?*Bc<6-aXix%#;I0P2W4L%tMZtuOwANQQrU)YOe
zF4<p*a1mGJ|6JS@)=3bDmIz<>=9l6QJH-2KN07z$$qS2cAyBa@C4gFl;@CJP#1~o(
zya{}>DO7T8h&B-x3_T|v#5bIvgWHpt-k~rrbe?z+-*yT!S*KWoARw^Gco5%u3WJhE
zlG7oghlZ*Tc{L928Algztc_c~YV`0ZUc1_)P6Z!D$dGDS1ZzL`<)v5u9C7cvaoFjp
zxbfqXC)Gw$gmPp#2(tf+27{TLTzkR+<@HQl-|4UaVa)kIsWh;cxZd)&6u{SoIex26
zDeA4`*xHNPOVe8lej38(lp)cO0*0=KU5I(GPf3nrW=JdSvj$(CAf4_+xBqbG{^n~A
z0L^ggXFY+P+9~5$)WeYSQwjUe&izd$jB>YP^XRtr|Dhd+D<rNF{QVK?3ht06S98!I
zLqWkY<ZqEM2*mhj{^AHfSLDBu&3|^VRODKg@-K3pj6Di?Aqf0c2_F87(yxy@yby$^
Z@QU8eV^*uO)Au_o7!OZ5>rSwO@ZS&<@3sH{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/functional.hpp.AB02FF3D06BB8256.idx b/.cache/clangd/index/functional.hpp.AB02FF3D06BB8256.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eec0504ad873c511f2a094a91bb994e00fa48bf6
GIT binary patch
literal 2350
zcmaKsc~H|w6vsEegv6YLBoa<nxdR^96ech_pa>Ny8maz7YjHRdu|<%eKpnx(v?x<c
z!D6vit3wz(N)@WbDpnDX7D0!rij39Lj`)WZtJY)G4s>_%h5Xgu3}5oz+uiqe_kA`c
zF(DyHicnluV)1&*#&i)v$c_9JY_zSb6d*JpM5uV*(Rb$N4Tmf4F6lArq*ZgCW*YAl
z)byWD+Gdrdu3Y>{h<$3$rr-R=&K}R~U$?hDR-aUK&sVx%xwY|5?uoU2VOzS(mbSk%
zX;<VG?K$pISA4&_q>}%|XQttCVneZ}zkmC=(k~xe%a7(=7v=s`c&a~R;Ee+%R(oFb
zYX0{{j}!YXK`r9GzTsud?G=d!b~Om<tK%ZRn2}vsDX<kUv^;ET%(1y2)P1@-tK^}e
zrg97$-L6_Y<x3^Dy)M(MHBA{@bLa5~J4I#bajKE<`NJ{KJyPU1(o2o4r_bNMr&lcQ
z8-06C%Hi3%=qEROgGR0#>zTa9<E?+oYUk-Z2M#wIUR^G!J9Ax`Q*k5ZwmtvZKd)Ne
z9gJ!-do>shLp9G<<@&pOT-)(v*>F`t{H+Drdi&W!<z0MZQM74&eSQaDR`Y&vSmccj
z{+%61sz(kSecqa#<1?JRu&)pcqDL)bkr%6n)SmmwT-MHR^H5jy3-(;dlHvSVEcl`3
zVd!W@>-N<0kJ}dr|K4WjGf=^Y>oZK`PK0^20p+6IX>!h>ML1WMt73vCk-H<Gx81Pf
z(Fd&-H5cT8aG*BOn+b%Y($ZhJ)D<>qIHecD@d5E6OduS0$!GJun_c(axe#xJ^?rH-
z6A1Gyxu3ZXh%d^yDH?=rVw;u;lDd$FK>{q`f0k@G;}&^c%mt7E28lt#1W6UNy~u!6
z5m&o%M_#&^3nT;lCH`(qCI)1!Z9LCEJ$P5l`Do~V5W<;~OwB}!b$Pw@J(*sZY*3GI
zm^RFt36!`sJ!YDxXVeKX7eW%~BsvunL^_>(q=Jg&ONK&^el$Q6Ou{*m91RngUmf{!
zW_jSiH3jk|^9x-IF%u*ebD?f@es$kT{__*NOXZx}6X9f8vWf}J-xbk2y}{U&Ov)o+
zbFg_969|9(ZIR|!=$n3IHaY(YS%iuSgzdpqHMRIrs+jX9*?lFx8YUCRb7cQl(H(~d
zv6u^n;|Z69tC3V9RI%-y-16&4vv1TDIzWz%5~G?4Bsf)P*s;iWa1*(IQXU_`hcJQg
zITuZ*HRHYSNcHXrV=eY(0^zo~ZMMZ?nV*>RBjw2@ay1irxUMYwg;X?>SFW2ynp1~x
zwk%u41m?H@HKTjg_f>;rd9r&`uxS<(2w$BObLUpnH(j)LDNd3kshL3d$w;$ZJJ#Dp
zW=}>~6{MQU1j3{B?dH2fF;nQp%CJ>p)i8mu&0;O!NEZS>0vY25+CkZTxHZIO!&$Kc
zt3)aV>lPvktQYB>ER)FOWW|f(oh-A+>}2sG-eCch0*=855CDv%6iOmo9Y9FjltAcc
zfRDH-c}@GnS}iR*PXcnpO?d-1<RETJ8Mpy2k(6NtZpHXfz23EAoCt?y!?sY!4)~F+
zQ1&4E7QL`!k8t!Er@)(ao4p+r;Zj|x!-Ab~GREXYa_4Qrg^<E2oG=JhByI|&!~L2@
zvLr<jI!J+2TvN2bBK#*cQy^gwut?@o5aIY0c%JJ#Ewm~SN%B%0Aun(!kYdUoxM2;U
zj8O+}#W;hG1C4Nt@Ph28tb<n<$U6R@ji*%UUyq@Ur_?)HCZ);AidV)vS!Si#$>No~
z!vfq%^^|qkIbco>gfb3$1iVT0l<kS?V`Z_915w~KS(+9W2H;5oN;@!kIicrL_bqeA
z|F#1hF{y$+1#osyB&mWv0TWeFYvOp4-IBfFEryqYtVEv$*auWYilff~tYov~q|<c@
gY?%fA^al=4{qfH};fojC)M;oc^^AYh;0Hqg0riEmK>z>%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/functional2.hpp.4855B705916A8A54.idx b/.cache/clangd/index/functional2.hpp.4855B705916A8A54.idx
new file mode 100755
index 0000000000000000000000000000000000000000..530c9333b0fbf06932c41976d9652c130ee0c5ad
GIT binary patch
literal 902
zcmWIYbaSg?W?*nm@vO*AElFfyU|<jg;^LB`%&S294<iFZ#hk4Hj#-Bd1g-=ZoltCj
zp!)Kx#Qo}b%q>dmKR(*JB;)Be#V79t4_#S%{%vY``i!}afoK2Kt<sgA=)Rgu;rher
ze*KU0>h1qKt-r6~egC7?;>|bqc!q}NrgeGW>wdB0w%poH$+6QfCK?@)Rye7$Y1j9_
z%OVxM8jRf)qN$72k4Z7G*(>rSSt);EJXoJ}<(T9?PR<B>R{dv(R=D@@NuG$DXU$@L
zXikUb-39aK`ab#WZ`$%f|JZZ;vS~}_CB0^Ln_OeFXT{cq^68gX`R;wid9P;sREFZp
z+@yU#-!;x&wB%q@mYDz(F9!p+u)455m|)-l$^k+Dtl%HsJNA8MV-jFz<`!TV5Cao1
z=|!x&3QkFCp9SjXWZ>3R*VG3SFzH>@{F!`urxuGb@dKshmE?861WbCjjiTU_j&Hqe
zOnht%+_LPlf?$Gyla&c*H6u6UQGqS{A4kY}fNTdlhk=`yomY&3lLZ(jAO%IKX~n%j
zHm^a)tgNYqw;6f3MTJErVV>gMx?;ZuYr$S-9&TY4VIGJP3=BNn@+|U5JWUo&I8UH0
zdBy%;x8#5(X=`dbz&u!=Xs;+KX}K9>0f#6*%mNtA$OH~GSZKj$MlNt@t<K0^d0+o)
zJIpYkYEV$YyuYSmH)}xdA~&Gf0^$PlFjwwKdvL~u>(oP#<9UU}VZuPi^U5RfG<h}Q
zyq&Lgt9@ncr-K|HtsM`u4;BqTdqHskvlm7)ae;$<!#wT%sk<g_16qKj2o?p*oZz5`
zc?IZHU|K-9Nn1c$9%dawQEE;iNUFFfxmX#*-~tg4z+D}%p^Yu`Lnsp$4+A{Z85jUX
Cq6=RD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/functional3.hpp.730E80629C119B9D.idx b/.cache/clangd/index/functional3.hpp.730E80629C119B9D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cce5b119a5f42787dbb37b04ad7d73362bf6239b
GIT binary patch
literal 2212
zcmZ8j3rrMO6n!%cEUYsE%VYVtd|cxK*^*UJfvT()N?|R6C^nE5UHJqTA-Z5$!41Z>
zBB+2+zeNO*h^VRc11KUYSfd51#9D#0RjUMTsKi#RKtkUPUYR)Ay*uZfnfLBH_nw!H
zjSLU3utg{|IntDolC5VDLXPB{m2F)24TaDnYlKV{o!{OwSZ3iJEyI4sd%L*$=a(1k
z4Y~esO&+@aWTU5h`A3C=x}IiDLi(8uiA8QBD;ux)=g5q=k)_KT?Co4S5?(Q{+V&0S
zIkN3v)Vba<r|DFEn_V`~4n8=4lnY5ORdnCHu%gA(UQwox@>;a5B<tftRWr?&pklLq
z_MQQq_p#3D_1@P^k;c=D6IPbtJKKL=o}aqSwl=#dX}oA<lrD7PwV1L^g_SQ)2FKS7
zq#2$CeD|aN=&09;hT`C-@c~;N+DCVB`Pp8#H#TM#?R6hYDZ8+AYH_uz$Kgu1RU^kL
z=fAkC&8(T;-Emf`IG85yl-KlLe_0V)(w0AP_kMOCbyICEjqt6lt>3zT|Fe!KMTUzu
z<+JmnrtR8qj9Zt|)4X-PUHG;t-%E8_Ra>67Hr*eQR`pgdZmj;)|5ExQW5Cc@>eE~5
z@>2;x9Wnhwn^MjV?mpj^8+c)9uQq=4R()fi{y=eJqq)9&&A}D@e}}j>t#8uaEQvBJ
zzcdV6IcYpUL0OwK){c>C%)Gp{Yvi{-bq@427U5{0XpImMDIk9>&wLOuBY^d>p%r!r
zua>Nq2?5zLlmsCPQw337?Gdw-zuVBBQY);Is9b~q@9$n6Uix|2JQbOqLs;RX&<FwO
zV)giub+WZz6M8DbY9F;m2tXfI%`wf*nL8q<UF{LhRpzRM0QC3$w&gve4&CH!2ZRki
z28|GaK9_p+^#;?420~9kSnH$J2m$DZ-KC*(+r3ymbA)9QnTrsRoi$B1Pqx-X4bIPL
zi<8*UZsdV^l6ek7n0(;M_OE_={>$+*WC4<%h4Qgbns-02NY_r!+L;4%sGRm9`^QQ=
zM+iXo#k;q^Xgk$oL%R@pO%l@-A@K6r2su2tdjHqXpo)Eb4pvwv(K!eK-hZ*dH}i2`
z@vxkBCtD0u2C9Vs^sA7XH$i6}jLB&aB9LCG_ZI@t#+1~o!^EE4#ky8s>kTbZ7Gnmp
z6O~7Z#R^6tLTW}WLV--62x%Fu2t_l|BBW>ZB4l6;B9zPIijak3EYy4Kup_Cib)?sm
zW^tl2F&J8r#W~8HU{DVFm!gHf1-<-}#};_d<{LDFll_v_AOwTWer7eaRaoq_<6DMp
z!zdOPQOk;`L?~!X%c?=u%Ev*32y+TkK_fv1v{W#`*-CY5%H_QctRO0oNq*;#`cA1k
zMBa^+#`-~i#v4|3o8CTBelTp3J9`u_9P0mIC2$ob72I?xE&b#ln%4!pbXMC3OU8qU
zB1_mMPLLOi3n;dLa)R{1bVL*0anR^sHZsgR4b}iP$T06P5VrdD71k89HsXVXPn(*p
zV=|6~Td`Q>s`8%f6`iBbZ4zZfW-JGL!Y^5X_b3Pj*^^=3o8UIc)><mq5w=cRlb8oc
z7JIQ?3fLz}g`7L0#T_!dVNsZtOvgKTays5DK(aWL4RwZu*a?fhn~%kX6#007xJ{{#
zI>MxP9`k15uxM2Urw}1Erxu|=E>MKDoK}RQxo8p6b9xana0U^|<#I*HLU9)AJ-+mp
ze_oVAdw@4fTw*+kOw1CO6Av|QUO3(H>CWLjq^6#>o-)u6wni#3afre5aC#9!tUl1h
z2mFyW=>u{9fD=gz^T!ql4<{CxQd|KrC>iFDDM;U#vR=>EA<LMQ^%;L7<HN+|1UCrb
fAKs+zaQpA!zt)a2C%H3{Pg49pk9U_d+#vKnQwp5O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/functional4.hpp.579A0D5EFC68EB03.idx b/.cache/clangd/index/functional4.hpp.579A0D5EFC68EB03.idx
new file mode 100755
index 0000000000000000000000000000000000000000..471e2ef40dba41433ec05837e32ec246c75319fc
GIT binary patch
literal 1644
zcmWIYbaP8#V_<Mj@vO*AElFfyU|<jg;^LB`OixAzhHz#EhKf0BgM9rD8}OX<|H$Y&
zJH#}npljJ&flE8yFtC@++H@s!@iNa7>;b;N-<#ZJS=w`3Z{y!Te`Jm;UY>n;j_sk8
zAUWe5mO3}GpV!N!Kbi0H>HUpupKrUo;*|WdGjN9E`B@WFgSghZo|SIds<*_UKBo6W
zdd+Uza;c=xE^l&NbPINVtDLvq=m*=pst5W@osY@32V_k<xZY0t)27NB1{=;rl&oUC
z8PUr7iXkvsFN;HzU!bB_ZbzQu@w9$EAI?KB4lH$h&e?xZAkg@lJZsqwDOUBZNA+I*
zkEu4f{>|j*(M#M@J0rYWdxD~7vSmK4Jh-=J`L&t_w+qjPzFxZgkjb@*ec?M+U(5^%
z$y%7y7;?pgZDmS^W_jQC%G9@_rl(i9PFS{j=l)Hb?Cp<CO?7^A?SA&M{|A~2FH2hL
zKFASE+2Lgr!S|CX)ku|{p|~<Pi2)d*teJu9=PqR4!^R}R$1bG6t{@C17&w5T4g^N`
z_H1UZWtL!LlH`Mk2=TM?3o&qVvI4n`LX0tsQtrBK_*KfoAjZoeq^qo}2__ggnHd>)
zKnikggdg}>Ce9XM5*1(&5>*$~2NN(0rZ7p)+}vR20J4BXNQhks2!I4kdd^lRb_o-M
z(;(+T?U!Mf5oX}z0;Up>@ry30otyJ+z7Y?DIM9AGWiu@>!NAGF#0|Fp{PZZP{iYKY
zfcEn<2q~*8>w^iH1x2Z8#hZYs3KkBGTwwpN&d6SQU;k@6uR+JGtf_{#nR$f7EyV3$
zp5YNvwNSN#8O<Z4W}yb>@ortQUxT$^FAI;5D2pf$#MKN8JVMGW%1At223=?vgXGK@
z%-}qMw&WH2f8CM;+G(I+U=I&rAuCNQ2e^NQtTnCSJXka_bAqD>=6W7TSVEluG)~w+
z4rUyT26`A2Ya8Zi?@!$|aU0Bsj7;EYf<;e_ICJex=6~O~fZ-q_E)7!!<f)76A$Y1H
zs?vy1QB_yfL-5qp)sT3GyoMq$C-4XvDjI4acs9H?A_#xlDB5Ts@l0)OU_6-PfL;Kl
z8JMqOX$BY=ptJ%L=7A(^cwkrwTgk&rg3&;eKuHJ|7;7qavj*fYas$SMkg$+6A}SR`
z6l9QistUR=-uj~UteLhUUw~Tl`SleMp0eP#P=r~-BV@^M3FmDPPT0cRez+ZIISU7i
z6wIU@X%Ei0aGiPxN=Lk+;xJ)gI^tDE;^}hg3M1sqIL+X^ov(GPeP!&YgVIs3K{&!G
zRv}gq2&Y(wSi^Zo-7X%va62GrP(?E=2?3KNDEq>4YEf!VBFN*#MajkbAcoLVi51hT
ZX0>@R3yTOcz=MH-K?EcV0ienR3IH#?^}YZA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm.hpp.344713A6C512A32C.idx b/.cache/clangd/index/gemm.hpp.344713A6C512A32C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d4949b16bf2b75684bffd22d3c8e55f2b3fdb338
GIT binary patch
literal 5262
zcmd6rdss~AAHe5*XR@dE(5X@8C(%v!p?*#mjbx%pH`Qd_%EL@$r;DVrE?GrWD7Df3
zrZ$$s7Lw3XgxHj>L~6?txmB)-)$g!Q=iPmtpTEuX%$(<QzTfwIdEf8%J?|N@lSmXd
zp2eE$=d@w9e}sg?VzE@HZ+Jvlz)3cXwM>o0+K~HniT9cov+)h#QI|GmHX5hJt>wlO
z&1cTb3)(np1JUD8@9bFSCiNBY{clG-Xg~dfljBzLVB;aN@4Gv*qiR<ivC$l|*}ZE)
zkG;6iXx=O>6->vXWtV|g5y!T+oH*cfK|z*nIPINggsZiU56iuLzQyKjMa4Z0navNm
zyDb`O`V#6&Zf@Of_Fh$VbKw?!m-u~yx%PW5bIE&NymSBA%XUdk4_LX!I5_~OSQw^X
z(@`?{^P(#6hwf(ve&#p4)PL8xZFNf&dByVf%nq~bFNVFHbriB|vJ}-l?#w*Es#^6p
zu%~R|$TgG1Qa$%r`weWpAjxxgU$LY7`h4NGS?Z?pg^lt!rBKi70o<jpi)%feeDhS@
z=nwrXM{PW2C-m!%Q--U9bKZG$>^vy?vq#6=N7e0eknsF7zTBFV=-XBr+)&#t?JwK^
z;O*7AgrcQm0~f_ST{I@Iv(0nk_~S9Zx>mSf3#)%0e|d0;>H)QTb~3vfr=iKB>SH52
z&-<%59rDWZgT#1+%-(Y0Vpq?{JDc+Ux>|2~mm857aP9tZcUU-)c8kcUAGqB4Jbn9&
zWt^H{!%ju+YP!L5?3t-_eME28fJHdA*+l1D-$J|Tg^goVy8pUgYd<wRyLD@e<oq0E
z(Iwvsx#!i6bw=XKz)7OH2}4Ovak^K14L!15CT2VD^LW+6{w`p@<WeD~e<ig1x3%XV
z=}u2S^P9!pg#jo2h<;JBt!Apuk8g^W9nQ&Hw^dJD+&(e1RsCp;iWNJ_5>D>lUi+@c
z^T0^C>*8jc299FRyySs3t*Ix9`rhUj_r<-xb9i=}K-WatEWX%n6)#IWvchoDz8%$$
z#{z%zn7w>?RQukJD<gF&IWw-Cr6p&klU=X7@AV$5-IKiK&ALmuc~Rn9i{+i?Y-~?&
ze56n`XGpaHJ2`f+AXZnScWKD1B8`eCzgTQ}w!F3ek!(%nn;63fK{w@<)??PE?+w}d
zA=u=lXsg?#5n<jEd)*(ECz@|+y!+u%gjHDNQH@gDcISY|tkoI|b2MAOUmqKB%C#!+
zr;@qi`n>1z;a>T0QRISde%{KSjV4YU=Yj__y&_)t+{}BCc_!S!zOcPnZL`BK?jE;q
z#v566AM1{HHLHON`J!p(lJ=4NLypdI|K+doEi=ykF!zifcoloE&~AcMxo5`wEyeB6
zT=$p6_F96lpI@h1{-IrQ?Xb_`@OOzN{C)g6?}PoX_u(#w+C2qF6{W`uwz!OS2ud3f
z-|;<t(WRv9o@vAuDA+5Ihi_ah4Whc(=o@i|t7bwH&f<<`5k~>*C@}hbVJRvqKqVF+
zz{7Ie!$)*IQ*hR|l#(~$t@G)JrT7hpt%g$yFEkome!jYF0nQpvDKtRc1}OY|VJTu`
zz$YuLF#lEI{mLvBJ1kN#v!>T>3{J?rWztdWM1UZ`o}nt>T?M400TR_fPz~%E?u6c*
zP&(?G=z@YSXiu|(*WR^F6D0Z&C-e;TOzNLo&s)Cf>3p2f*Vdm()0j#NtyvBwXDKB^
zV?zh}Hcsq-6L!G=(!R!;?Z^&HGw;x+Y-#D~GirK{dTtEqPt@09$XICX$j}PjipflB
z#{us+m`t0;i88<|L-_>Yoj|!9@XAp>33w+_t^m9WlurTPDU?qG-f5KM;Ji5K%=k}$
zWCFBdI1!SGC?`QO3FTx+CZoI?lDkn(fn*BGsgO)XIUSPeC}%)21LZ78W}%!7$!wH!
zA(@ME9whTn&WB_^$_0=tK)Dc-g(x>dUL%}Lmkv%eL0%Ke&5+lOatq|Opxg?1tthua
zUK`3cAnyjsHzDsP!(+>fstgrEKUQHWV3h)l>dQ|boa!M~0~}>#UaP%XOIMH*hv)+q
zeZZ1lF%IzrSUf>F5egIGXtQvLWGGBVIRy$+P)>)!bd<B8Fbm~eD9lAU9}4qPE`-8D
zhBeEF?k9HSpJ(GlfOfz%rtqtQb~TtrhmBJ`tK9{s(G#!R?Uh&IwqFawi3ErzKrKcw
z3F1j8?}qqplv5#|igE_TGf>Wkcs9y;5YI!o0OAD<(|&={=e;mu{VL~}@a@}8Hac8Q
zT?E%~a6%%GI5CMy1yY231-=NyvTb77^XVk{e(wdE@W%^^9719(v7t?H2&uIcAzy1>
zghFgX>=?~Zmryr`DuHb!u%kof5Ep>U1>nZ;EogfS+R^SgL^pKlhHf;Q_wFsxSb1p;
z;1GHadLnwqA`=JBB<$&U3UJ~!(7X-i)AlS=yfxBuBS}i|29j^!O4^XMeTIG6u(tBQ
z49%PoC+?~_{EE^y;hIdN?avcy?BDzvn@80=C3*!W(e`oT1HeBp>=5Z>UR4(hsqwLF
zUM$;`_UCXSC?(@3yK+h(XY0w?(`kWYORgYinM1fTP8egxs<cOnjEzZ#OfgeL#>e;w
z^#iOQs4^k-1H*nW`Z&Oe0e}sl@<D(NqC5nQhQJJF{19*&0$vQi1x9aC`7kgVMmZX4
zM?(`vKOSQ75Q%39P8otmbiz)ewc1NF;c;p%B`%T$bY<X#)I~Z<p2@!|i@r_iq?G>u
z^H99XT}w8gE&ig?qB9}8ly8lxp&sAiShyPFZauahp^E^!2-KJyt^n)`%6B1l7pl=Y
z;6yLPdKq5wF4MfXGKT!>w>qC$Iov{phavwedmq>Hb*HM+S5P~kOd$KmsYX!s&r^(`
zlR34l=t{HvQ{>|J#RZVoowN_RKb=h}`MUVJGYb&Qc8+B)pr_Q|O?Y<eu>;bNAxx6#
zFltp`QWel)R;Ck9>V!J<u<!jH*A9<6SW5L(FVO4-vuMTDg;{g0GDW&yjBQ&?`Tdiy
z_&Mb?$SlZ`aT;tEj8F{3V&FL1vV33ECbyzXwcp}|pQ@h*Gss`nA0ZhgQ(<H>1DP2@
zjxrC1ssL66RA}=!Q3VXDfEmh;RltMcPKb3v6?z76q7xc)LNk;dJD~^7ac2$tZiI%e
zp=PKUPAY~5w4nrNw*2W&Bi2;%4}sD{G=j%K=`qSN>~S&d(dtjAjgrI^AAdy+jO0XG
zF^SXztR4tx`#A9!VvnJK4kRn6vH0v9XS*-1NN%3V-OU%)S$*}LRJ^!(wo*k9pPE)<
zuCWolQy+3aSl)GCouF>Vux(?Q2kDUwl?ooK3q^`JVahetp!F%zF*Rn0&*dXB{!~7n
zkqrRu0MKBjcMxy~QGN@^w@d@z#4sR-QI3PAage0Pm)yy%DIlwZsO5MDgwKG47U0AX
zz=nV!&81=61G8)wzN8cnLgj-{n-<_iG+Qs4ZAmZisa4(A<|+N&LmfA!u%;+6=R$9m
zH$tUAu@rC_X#-SjfTLX7+*DWh=w<a+UjR)v)-5*Cx*SDKOAs$epXq{NUhpWj`2BqD
z$&{4;QsuA>T5N+W>A7p0{-mSqyOk-FAP!=2P@NXEzswcOGUK;@_4(5YQ>V_4@7VO2
zvv2zBu3Np9g<tIQr`0d~uWM~;`&3fcL1u6LSW=qNY+bsf=qu&q@Aq`t&FeU)Q+mH2
zfAy~i*_-=#n|Y$sa%x(4fQdVRCLQIQccn)rJX)PddF%(=elWTVeynzKh&eG*@BP7&
tkHr}t<{K`cu8`nM%8p&O>9FOfs*F=+snV7E_aBR&X&iHx`p<y={10(Qaclqp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm.hpp.C45C6ACD321C3642.idx b/.cache/clangd/index/gemm.hpp.C45C6ACD321C3642.idx
new file mode 100755
index 0000000000000000000000000000000000000000..95e0d066f1ee4a682ba980611893e4b84aba9271
GIT binary patch
literal 17426
zcmY+M4_H)H`p4&e&%H5Iw`O)`XBwG?nI;jNCUBHWlxVFK2+}nZvvt?X{+C(T{`ad@
zWI>ED!sri`l;R&u#}o}L1Q*x>6-q)OWfKY!l?oCFCHh@<I`bZVo*wu-@0WAuo_oIU
zdC$3Lx+8hUjEis8>2A3{`T4~UJbkaE)9L=C{kQaK^^sSYPB;4+o$mSf4!*tkv7-}A
zU;H+^?>>hkcFy<DUir4^By+CV_n(t5efaHdzWe_6!Pu#nEc)){RlnTxMcune>iXqd
zOLokv%%Azsg@ON`zo9Y=E}v~$Gx^VJUfc3YUfx<cv;3a#6BoQ;`saThnes|e-Q2*V
zuY6TI_otCZCLCK*c~!&OMKfNVKm60earwzVIqS;)GJoaJTH^=HFR|`=;><5gh96wC
z_NLDlUq3Lk?%eLnqTaDjd+OJwH?Dfq_{6{d`q}Xr`b+m(u8jS4X2w^?k6wQBrO!oW
zT>R7ZSCuZ<9-XmnIB(MDi=DfNc0bc{;&09OUh?BzSx;E**e~BO<?&yxJ^j_a36KBs
z{>iWACOp1x>fsNoPbC#T5q;CkrM+JtNp!4z@#WIi<ws6D`}|?+%O@wdJ>0a_)gN8i
zckbb%i4{ASe0Kejy4jVt{%hUspT6+TXWJe-TW6bo_=W4%t==8EV}J96RSA>#N4-BU
zI`iR|JWKQIro_xW`*qs%XFp86{<5xV?kE3q^@`Xp6vNFs-7CE>nqogMJHM)+^!4}7
zFYdbcX!%RG;*Bv&eXhf8S$kf(BHi*{M{1SaZr;_^TYTv_%aV%HB}3bGesXR{>w8h9
ztM(r{pHutn%9qBkI{xC4<Ec+vykPy+=}r5de6DfM<i>m6o%YfF`3J3w50y_z{Pw+$
z;Z<h)%@LP3J(B*RvGb)#k6&N6zrN&&hL)l^12cvvUH{{nitvTKPk;XC$FtJ@v3vG`
zYpV0$`So=bz0ZAm?jJunPUknj_4$&IlOh&J_sci<u6yUR;sQ2w?$6fj35B<gn{>do
zspjvq-b#7wigyk^n2_~d?T-JmBz}Xp&nQe8dh)t!uY0F{%A<Fe`14m*ubeQn(VjYf
z)q%aQgyp~Ue{arAJ^cPd^`E_b-;Sm6O<%qiK4$h+#z$)wUR59d$u}$Cp1C0N)6dQy
z+jXXJ(;W?e-@iAx(0ozx;EU6CzI=ne;g#P0b4Bwv8RnlkyR_!7s~-5ad*NUp@M74a
z6;~}9$@`@6szskB{wLwLw>sZE_TAOVPwI3lA6Yf+!0e34S5KRH`P<Xhe0YDIA?(WA
zm)3v%>hK2#{QIB8B-^;@8Pg`eb>xf5jZ@#=)HCJnPWi6H@Yvk9<!?{SzhiF0TR$wm
zkG<<JS?Iqq>W6<_x%=@Kw>uwwX-EEvXWufet?Irz@$ijzFFrKwt4E@BukQ0H=hF{m
zEc)8@>$2J2zwDcT>GP+*+keTI4_!TT-tUoLt-a^B-(7+AWBW$1@UFYoEq!6}eOH9(
zbeCK{+;rECX&)-O8z0lmsr^&SKrVx1onicCx@R5*6^w5@^P}UT!&QpzCNB9I<e%YI
zoxuS?iL+~R8&1l9T+#iNOHQMF8t>2<CXES7TrC~fJiBSqC5rA5E;)$uLA*g{xaLnm
ziMw}!{tMfsixu77T(Sw}O*mU;xXKWec=lbg@Wscsy|3t!xMUg1%WxV=#<pCxW!0wc
z8b$YaF8LJYPe)tvnu4ud+gCCxwt4GYif#&*WHC95P1YH%z@Q{%>#2kH)XjcS(cQ@<
zzo7gJCg==4b5Jtrx~Fd|$l3XYqWcS%^n=_#+Aq<6{zLNp#D!mudDx=p{=p?*gZwpU
zH|@~}C9%ewu_;NhM-<&mE;)+wQPds)@4DEhR(n3ITlIjVdyq>iP_7v5y0I4rTbWwC
z?aZSUf4<=&yy3Cy)c07}do1#gA3A;Tx!(9s{e7>e4A~UjeBMeY$ep0|3xCqnN}ief
z?%s+?if$g4Yz28MXg3`j9=y=ao7upNTi@3P!wfDNM0pVZqBBge1trrv9!xyo+<(2I
zo603QOwJkY-56_7k_c4`5;w{1itctUIfC+$(cX2*K}p&(OA5Z~`1of<7sn;(Oimwd
zg?HV&6_-w$H!m<n(Is-pFv`QDpXF*}u$8;7y=zz31AV!QZZ4Os19{!(9gS1|kZe>_
z?p2)qitc_c`2gh)Mtj$Pc~G*TclD7MvObDebaS}mCzOA}Bzi^*pSyhVgl7F|MVHJa
z|3dj+qXW~)g00;1<imxxoc_~zMVG=Qt5IG(D&YfjkumGCmb&}u72U&J@+Hb&j((Pj
z_FyXy%v%zkboS=w6kQsZYyx@H=pDH)`a|;7&O>E})!Oi0z$H6S-ZAO`en$_jUf4V4
z-3f0ix`()=8s+NI3q?r5R$l!3=~rL<Fu6d{-Nz+sP+l`C;ck2>ciwaHOQwIM=oWIx
zVUQ1x4sY(G%v)wXa#45mr;09tOHPA)dekIWh6OK_<&$pf$ga``!!#~A2lBa5Nz|o5
z$vY;?RhQ`&U4IE)|F}-w$S&T<%zymg!~4DWjvws#k0(me&E~Ciq1+`tqxY{qG}Z?n
zq?mEC*2*W$^a;D-j~~31wXItpd*yZvD7wG#R<@$Nb+qgFV=US*CEZo(XjOC%aLGE9
z*NyfIe~jx_G~9AszHhmryN64Tf_!weU$_Idg)iSX*zI{m(cR1?6(CoDHW>Ku-sRXZ
z^REjp(^B&+F8L1S?{E%{y1f_W-M93>y=96nl}nmXZW{Gb<R!tbtD3w2yTN(Gaf<FX
zE?I^0s?kx$2gBj-%Qu`@-4~|l?&Xr>ARiz7xI8noR1fIlGnzLky6IeU0OSLsAD2sx
zC8cfu_KBQAMYo7c-a+}D(SG4wcjjEm3hzar<+^w-$zyWfsHb@={X^}K|M$>SZz{T5
zx#SCwzZks`@47+bLrP7l0~OsZT=H*_|2-<<qi!g%;;D<n9(F0ZySQWr$U8<&a;+uU
zFTaIP-T1?2FhkMZ!6o0L{JofW)dwD2+A>+EyU10Rd&`G6xZeatEe3lrI6`SfV`nkA
ziA%s)0`8DW)c{xmU<)NyjqL$&i0ll2o46D#rC<wvezg>wrQjys1C~8t3(eZ<9&qjf
zH*qD{E5Q*OU1}vbE5S`%1I`+7hq_R$14|v)LcOlBy$&2AJL|wr+yIsau!V+$+5pZ5
za1%Fyy$KwX_`Ih!fwKwR#4X@#0e2{IsBK_r16#;28r$2zA+oa#+{7JV=>S`3#Ht<O
z>;N}$7g)N$7MknSE^u~%o45z;J>UpsCA9~fJ>Vwp17{z&LnV(o2$n&xg{FFq?StC=
z@u^>9=b(0fAsz<HFxWyVO&td3Ft~}+(VmWuP#L49qca`d#98RfLU(93QFG9egSL=~
zHMZxVLu6+Tx{349l83g?RHNphGY{Rw1!yloM<}<e1?Vh5H*q043(+0&vRZ_eBDCGj
z=SYq1Md%ROS%hxlVzd;aEtDhGVssXxo45q+CFlrMdTI$eOVCXmKxY8mp){<PqNNmV
z5AYdHV|yt&M0S><o46b;<!B4lfND89%h63-h4w0Rgeoev3Y}HxCay(iExJR~qgs!a
zdbEY+WsU9i=n&aik8a{dv^1hERP3sa=xjtcaWmSR(GkkqYBM^U(M{Zn&Q^4X>Laxc
zEp2EE6^R<#+qBQmi$;x|ZQAD#aXVVt(H5FI)ONJDqeEn8JGzNG(b9>wP%>0I(cXy;
zk)56BChkT{H`+pzt=f(DZghz3>_#_n4_bQA7Fr*uJ!tPihse$zbQAZYr5A0Xf=lg1
zdoMafcJ`v1xF0S3XbV;AYCk&r(M>#rmLar-s#bLf?L+7g**S!6;t{lrpe<Bcsw3ze
zK{s&*vu7|zs5n(Km@|X9iL;qAo4G>^3^kWoa+xhu5^HSFWe$;@xy((R&n)@O7Mdv4
zeCEt&ZsIcLD`PRdl2Oz$mRQDaBQ9sYaxq@b{M9U$;x)`)BXR>vZD6w~-pKroV*Fd?
z|5oH~=Ia*Q?`DbJ;`L84-zhOZ!2AQ^-w!hXpvak0YNjMyKTGmwNy6t{Bl*{eyg~A9
zkc9i+ASG^)g!|tp`8JC2Es}qW`1c=6{*Oi8E~Rc4um6eU|3r-Mmi)U#u9AFJV*6E6
zVwHIPeUfjV7_XE3b>iRGOa6M14@s$q#OpUn{w6VgO!6NS`IO{4CANP`N<1ZA|Fq;g
zEyf2V|A6@SgOYzx<nvPMdGY$gl7Co?FPHtxW#M^b%f4(`=>KdvF<Tb;f0gW8CB}1Q
zf3EoVt7ZRc;)~A}{dYymb#wCnS6NLdhUrvE6+>b%+#&Lv#judL1g4b0bgHyUAh87Q
z5c$p$SV$azDFK*Hl~w=}18|4PcLrb~aVbnGh3Qmzl|o`E+#&LvrLd5A5BT>$JXK<Q
zAYl(Ai=46t(uga;UkUM4nN>nUB_xZSQVD6qHQ=v-c&glLAfX15MNX-KG~zn&*Fij0
zZgr4Q2gxF*)Il0?1Na*to+`NpNN9j$ky9EVjkpQ?O%P9&T@xfUL9)mxO^`<10{#|=
zr^>Gd5?UZx<dhajBW?qK8^lxP*9Hk~kSuaa8>A6;fWHIcsS@mfgbqj+Ii&;Ah`Yew
z1@TlFc0ocHB#WHV1!=@R5ZePcQKi@e(|RCP<e5D%pSTZV``{+3B>P}mAEb&rvk&GI
z4?^r9+(eb+AWR#CRFP*6!hGUkh#iKTsInY}X~U2z^2}kFPn?eabd0A;GaVDsF<Io4
zbW9`8LVp&<Qze>(30asda!MAa5$B*k2ji&{&B25mOcpsM2h)i2(4U9#RGH>sLLMfI
zoRWuW#0BUtz<8=u3oxMolSNJ`z%=4Q^cP}0RkDScP>9JQrxaouaS{59FrF&eB1|a4
zWRX*fFpanv{lyqhm2EL56l1c;DaDvZT!Q`*jHgPs1QSXyS>%)wOd}4UKY;O62?sDC
zfXO1K1Tc-b6#b<bPnB>fCX`~b$SI|mMqG~ma*U_SxEvG8F<Io4a!ezxLVp#;Q>9#m
z300UZa!M7Z5!a%>7UQXsuEm5}Ocpt%7So99(O-}8R7uxkLOmvnoKlZz#Es~0#CWQ#
z8!@2~lSNKx#5Ce&jBds_s<fMNYBSCdIjI@v5VvA<E5=bJ-ilLOafZlAtvH9c4WrvI
zjw<msoZ5ynL{4hMImGQ4-HvfoiMQj_cAOz{QajEe?!@R$jH61t6Q_3K43U#MaSm}e
zMt5TzRpQ+^wHs%MoYakTh<h-)2ji#`@4=}(I78&59-Kqmi_yIpN0oRlPVL1RA}96Y
z9O8bA?#DQ)%=>X_Kh6+2sUPPM58;F%oJy7X5Z*Y1Nh04igmZ~UaKZ>qrAmDSZydoS
zk#8Hpxx^XFpTXj(a?fB187x`klnj<eoXz~%ES@U+Y?hGCl0{C*W@*H^%%98Rsj|;y
z3Arp;<dj^NMx4+5`7E9){d|^?&yqz>$!BStc^!d<Ki=28_V4pWAGc3+E9!E|v|PHA
zm-Kvs(S*UbHY6;QOl4I2{V9+g{>-*h)zON!spQ!Mk^B;)9JL3$BK!703~?oRDj||e
zl}d=J1h2@xN{At@0Z$D?QmIk{Q8nNd*;fNG#C71QgGee>>L98PydwMRAcnXBJPi;@
zrAh-tHGo%SUjxJtH-V=KBB@kqf~Y3&itKBG7~&T2v_K@4DlHJz0$!1QEf7Q82A(#E
zq*A2~qT0YKvabzdh&#a30g+UybU;)Gct!SgKn!sgc)B2xN|i2%>H@FGzAlI%9zxF$
zMv`@g@R}ivCLTf02u6}sM(~;uj3&-to(vXA*2rK{8O$rPFN4JpXERSWizI7gv#4z5
z71@`~Vu*8@CznN%HF8;0F7t})%Vjad`OK5gBDpmbZP6T+&%7f0@>vXLp29Ro1Yc5;
zBYu=jKT3b+*5Pvp7}VgaN-pG+wAMoK1tkwC>S1O&%ogw#cr~R7gRdz`*vw4LER_rS
zG^+`NFDps7pPBYki{s?oYp!_Ac_%9B{mT8}!B=}aoud8@VZVcwGw(~Sz2KWoYVQNd
z^nr8>zcA0YnlSi&lY~Q(X-G=uLSA@k!r)s@5^k1En<<ZQn`pfjeAmf^eCpF$2)^Ou
z0Y&YVOudw{cnzVo5PZ+ch5RiiT>QtoO&;JwTWdP_PLl@|HC-~LOQ8qCmv@>l_-2!Y
z)sm^2ZkL-u69(UJk}yX$<;W?#U-&50gu%C*Bs|MZXIZGHd8MEUgYP;?xI#9qkV8Fv
z`wed_v0a<<`bCPmj?G@j7Vw9vsO#D6^&%Iu*~KDnWw&l+v-sOGMcu}3-6nE?-5Mai
zbEY}!WL~^HQEM0s>%mBi>&;-;3`SZ!Zv(?NF!J70)SY112}WA{mV==jj5H(c1;bu2
zQpT?aLp2yFqkjp8FTqF|`ydz&f{`-vAut>QBOgkNdIStdz(^VO7#NO$kuqjG7}~)|
z8Sw-dPJodQ3q?HxhBIKK#aS;Hdcnx&MnxR}!+>_Z&^G>gFq{V?EuNO6VL2MPofS0`
z4Vh@9#m}o~comJbxOp87ucMI`FK?pZO*GQt<Xtqpi$+>}tVP3GG}7W?JsQ@dkroe|
z(XbhfJnv`^77g3bNX7q7H0(qp758OmC_^L9DvG)n4SUhZjisnHXsAIW72kDes6!(a
z*9~ZBKqF6piuw&2zCj}u$1P}RK_eg1ih2wU$F%lC+tA0+a2$<Py#9cOAJ9m}=}9!4
zL?acSXV7p4jXX0e>d$ES8I4pto<+l1H1cGnsK27&S2S{yD(Y`&_zjITL#<$j70gI8
z&`M@l$&5TJD(Y*@@ES96?<(pW%<u*?@)<%K-|W(z%)&j++tI8WJcB^i&67-dluh{u
z(1gL02!c>H<x$q)PSb?J^9UqtmQ2l*hxm}ta$)dP0vGbW(OL+eMc@Ig^CZ*H(j5Nb
z`5@JV!4nB2JkCtVX%1TW{Ci98srdYRrsW-J(mH7hZ()@0S|=?eUN22rPjNoXwJQV<
zG|(05nW>(xz{i(%Gr?mGT)3$3(CyDy7EaP;wL=(o2>(ne?}QX~LJIK%w@<5B{>g^>
zjH7R2J;gAdcjPF0i(!Ju6N_OoaS3=zU_7O?67ZJ51d%6}z+~b8cmgn<QeFVO0hl22
z!~jetE(K31jHi@X3f@wfAo9dgm`q$D#Z*YwQJ+>wvnr&y#CxTfy%gu4OY7s{aSz^g
ziuzy4^j{h<JO^mP;Gqu^Rxwi*RTA9%+QSGQ{oukSDZA%wKfhqPHV$4;UJy!*k09(L
zDlruGb1Cd|DO7SiJN4Rk?tlJ~$=X)#=NR@mh5|h6X%`9}AmJC{>0Ap0kB#uaihIIm
zZ1GQeR#CL~vz{IpM~S-!#`l0v<meulNZbdWJ{U)dy${CsfluV<KA1>62%bS0M~QzB
z#t(u|<mf?|NIVRlVHigf!!V2=2A{~$!!VI^#=3vLc>AW>FYZ*-UCg?RMbP8j#bS4{
z8;N%_>uxb##>SK}55>z^Y?-JpXJg96cr}ZtW>KWCW^vW*X5tzaQ6t71SVY5x9M>S~
z8(Bo77;j|}trv1!tEm5$MSLs9Pq6S4>?*qc2{z>f6+?>J#lpM9csCo<E#6Nzi|rQg
z_Y@m*N{kP%hyn3>11xSp{5*p!Vo;1{N)efo&_9_{T&5)SPnHysCC2lmi2MsVE??BI
zks{WJ@pV%8I!U<Sb<$1iq$Ilk^-}nHF}^_>vq2J`?*=J$gCvZHjnbHnVtk7fu|>Sz
z7AbCv_<25-B0d)5+og!@7joQoQU8e)@rfAUB}MGIkmGiV`rT5*ZZTdVg;$98Qz1>M
zpovhcIHd5sV!TQkQzhO{l@wbg-tRtX%sw$*Cq>kW*Q=A_>cr1eFGbXg@k3I?p$j?g
zkf?8xBAUebQ7Pi+g&cQO)E|=~j*0OuDZER(pDroBOPWs4_oNhlQjDLH#+(xG=adwC
zO1$6G(wNg?d_amA5U)2N#SMs`XHbe56yxWmi1Qb6+<8$yEJX~9@nv$vGFce^%jCFa
zvM~Oa%Mr`Pc%~elDGTE{Q%=Z~6X|}l<nSypo-L2bmWA=0Eyrfd!gyXKk69(gbLEI!
z@p`#(T(0<eR?88qDb5qTHme1X5Avx(E7c^^PFj8P%&Q55hX_fyO)_m0gg;NeBCatf
zTvpT*Ean6gX5tfU)(Ixe#9b_=ON?jAlQQKvu`4EJ5%bj-Ui`-iMZRhstt!XvWIq0f
zg<aP>XX<{e(~6!`5OzwOSwE7(KBAeGm&w{a1&=h+J@rVY9xC;C%i8)lc)F2anlBWz
z7J|nasfBiCYNw)ruQ;?RIe4g%3;9w@3j~ic@&I4sX-x+YF!F$+zAl+wmqJ?{+utyJ
zdXXdWj#JSNy?B~&EUhM*aa=Qw7unm4KH^sNwBlG=ZM5RJRva&~w-tTFZRlykv9ua#
z!*OjmUSw|@`iR@n(~e_lwbG8`+Ht(d-gfj6ccP~g$I@!16UTMpc#*xG=p*h%PdAR`
zk5SPM&5Y~D@gjS>(MQ~ao*o=atDzno*Ms9l_V%EUxEDRWIF_#nwCxle*Nfvt_V%KW
zxF0?JIF?pZ{Wz{4$BXRkM<4MJMh)TBwAvcMs1dxHFLp|!40B&U_+huCP3qF5F6jn7
z3`coxmo%UFq%`TI7|)dFWy*i!ao)k&PzoM&rJ>X(nfj!Ea3LQVnlO0Wm4qXbX@rVl
zz983x!9%YktYoH2TABR&``2o}zh-rYR`XgEi%|1^3Spm$&g^7ioh($^9)7Fpq}M$4
zwMnDP4}CG1#pQ;+1kAi_SG0ql`T&^4<$}Hx%(NWX1NuE+rX7?@&{u+)yGvWKgT4mL
zw1ZLy`Z_Sv4oU;)8^BCEC{3Vm0yFKPw1B<^%)I1Qw1bxVHZXHrE80OyeFvCn2c--2
zU0|jilpfIcfSKoEMePH9ADG2LT0f{=FSLU)4EkX(^K`83!=XMM&9sA(h59Ts^Tny6
z9c<L+pqW1~MLW2t&qFg`S}NK>MSTI9`C39z3sGN)X4*k1LVXdM`Epr%9;h!yGtXy=
zcF<5?f@a!537|fJW}X-o?O>t46wQ3~q-X~R_2p=$9h556SD~49P-;<Mi)OxvQnZ7C
z`g%0;6|SNk{L?q0nRZZ`QQwSaK5`Yc74@xX7IT)qO=~~2gVK)rb~JNKDB8h1eJ7g5
zoTBeWvzRmVJ!lqlg1#5cqUZJfXcj%KA40R}Y5fSAxu+HFV4FUJnfcC#qGmIFHZ#)>
zN-opqGBfR<<THIfGt&;r`&zNie6)_=$gCTgi%*VxoN3cZ@TfUACLh3>V}l3Gc|cK@
zNv35qo$x2GwGcdJ&V`*5<g;PM%U}K9uBLT8xX6>6!MYh-<jHMd-3Bi5<W8{e1Q&U-
z9IWNwB2VrG>t1k?C#%6)4KDKJmtg%8T;$1vU_A&feoNZk23QY)i#&M*tVh5_o;(KD
zW8flBwu7}DT)eNeLItcRz(t-s1J*O(B2V^$wHI9E$pNqqXx9sQ@;q42gNr=59IeaI
zMV`z=YbLtLldq!nRdkUjUq|cf=ps+PiPksKMV@>Yt?#0XKN&?`i`KR1B2TVI>w0vN
zCpV*YGrGu=+t9iVUF6A~Xx)h}@?;rW%g{xh+>6$|=ps+nptS~F<jFd;)}f0$*?`st
zbde{&LF+f@B2TuUwFO<|$zy0ernMjP<Z-kfM;CeW2ekfxF7o6_w4Ov4dGZWe&!CGu
z`7>I7Mi+VVELzW^i#+)&T7N|sdGa^3{)R5{<O*h8!Cd6YmCU-5xyX~RG3#s0MV@?v
zS>IqT^5jl7W+$z=Pu=U@b8T$Tq`!<#UY268kx@#(QUW$IN(tCYz(HIJ_EK<=QTBj+
z4>-sum0+m^8yTerEHz*wqtt-C1{}l<U~d2i8Kns<P1>pZ(6<O$z|sOXGD-{BTfjlw
z0rn1XkWsq8-USXaN)K3iz(z*t14|#+$S8ea?*j+%FxZE|K}JbOOFG)fC|PL9LK_(+
z3+-9xAkIU39y-V<1!yTi8yTe#Ern<!qZFdO5FNzDXfH+w8Knd*C1@j~1ke&d8yO{l
z_5eDF%h6tr4l+s=TB@{D<otVtYAsr7(MCq8MSCqeh#S$~hz>GJGg_L_Mn-8xODo#Q
zD6MF3MF(*|+WXN#1A7GRBj})koWU#^%tiw_n_04%jRtZyvu860aXz!>GY1XiYL;3p
zo?32T{s!?>aVzt;UdXAfqJDs-4v3s7`7^~+ulbTc|3Xg9r=_E!Zjn;Ah`e3$Zx{8u
zB>%1pIdzw)uai>iL_Q?>4~hDtlK<$1oO)E$4@jv4BA=K1=SBT8*}sfV*=kd#oVrXF
zPFd&5sky|wU$r?m_;oBkm1<RhWXhmj9X_*Z!r+&(Ncamg{X*q5ceW-Bel?4P?@Oll
zX%^=xL=y(Tphdz1lIZ|d^5=HHwJ&<={+-(Ha`4-&yt4Ssl(wU7NzI9|qa~&%9V5x5
z=@^xcUXgw27(<+eo-B+c*Jfc<7J5bYWnm0)4tjDhlBUBPjLJc;$i5tmA<jcj9!63Y
z$it{S^os1u!x-WM^b}wuWr+feDnPHuz5<LPE<{fuMpD)&#Hd2_itH=I7~&%I6k#My
z$3+-bgkF(-MHoX|jGkhQq%2g7QN`#L*;kA)#3kq{!AQzdB^XtLUXgtz7(*ODPXHq+
zYXvYWfL@V(0gNFoMNcV4QaMnHQKje=*;k4&#O3HI$4JV8<rr0tUXgv}7{i%Q<JwUC
z{(oNtq@j3LGM%O42_v0`qFK-Xqg;DCQH7o=jHJ9;g;7=L71>vXF~qg#sl`ajv$Ysi
zi(Zj^wHQNOkDhvrqynZMqw3KsvacRvh#S$<h>?`H8!@U8y(0S>F^04Hz~lFh+Gwe|
zXz4=jC#n(eZv6dM=XVpsmOj0B$rDdLrCZ8vrqew^zb>I)7t^oW|M7tTdy4;}(>+bU
F{y#D<&1nDt

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add.hpp.0A05FF4E11D910E0.idx b/.cache/clangd/index/gemm_add.hpp.0A05FF4E11D910E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..92a4be2c7cee2a8eac2889e0f42012dfa62bca56
GIT binary patch
literal 2284
zcmYjR2~bm46#WTn-~;(tkc1@x*$w+oK(@#tLZT5#0MQN!iE%233!AJuEMhHD6x7y=
zfMP9FN38<V5o`fL5zwkw3skFRks?-QT)~e0!Fl=r&*bHuci%bp-gn=-aWPR*<Hisa
zu^}cmU7eMp4?z%u{W7yMHkpzkh>F>AYX;*}Drr8{hjkY@L%EbR(wla)tjzS$1Z`jM
z%MQ2azY3a1M{>fHA7A8L=54#MBfq5g)bR0J&0CY@9P#Yz@H@MlT2;ILWQ*<4_N1@Q
zJ5k4;D`0WwW6MKxYeim~vR98QlC$qlhB>RFw@g0p@93B0?_f5ZW?ZUk%B=o#I-{1e
zGOEU6PxtGd%8K!U!q%*}e*_l~d~z5_YO1O4Tkfkc)r2r4>ra}ko!;%~J;~S8jDD*L
z{jJ5bC?K6Kv&tUpY_rbnmdssk+TS~LuWv*)H`DcB$g0+nmi4bp=VHd+_H8aTo4?YP
zx6Zrm#NoDQ<CcZh8I*~dYuyL)0+>ycf3-Pk+S+Rd9S{7&yWx87?<+eS4U!m<_alNy
zoAhKIcI7F#M}v174X)GdDtI#Xb@N=m(V$<1NBrZ?jStU`##Cl>NZ2oC8t-yidAlA~
z72K;TnCH%@t!J)G)vD~5m6UW<AKn&LR&eCQ66;h>dD>A5#mfFrnt|(%LsFL`H>Q|t
z!%f{ENCK`k`;Oe%<lRf&Orq^A>8sLHR(K~*^wyB34)}g|jJCYAzUcCggFW<}B9+_e
zBO`A@i=DHhH>7<i@7=z8;H2==(gsn(q<`W0p9OZqHiNAG_TTRo-_9@E(DJr1E9b#@
z-lssAk;*IVkLadT8I99t-aLg~t#(RE?X0f*q5csJW#*+T>#)0_8C_$Un_pN9LlhW;
z*oSZoI1-4WM}nB9B!pBwVT`0Z_Nz_FhGq~FZW2yg7$EE}X}^Zy7w-mmm!83p3E<NV
zXkv(>PyWpBnq{kQ3TCSf$wsgN!Z+eOEewD^{9YKf<<y<)G_o-!FQH0o76xF3GSnM0
z&l*6GA*XGB#0h_{IgHRzy1&jX!ax}WWQkf55Ete8>k>rlQF{Vfj4mdi1YUv{2n9wG
zRwRgUl<1M@O+e9!VjL-XN&-*0LP_AMJX8dpDq5xEnatHcDi!mk#xP=yS_|=HY)~5l
zaz$N*c=&Fp8v%t1Lc=~+0D@pdC=e2m$W4TTvcZVhO{_!6`MvDybJN{q7+HiaBAgT(
zW$Tcm`IjNgtL@|1CML=Z!J~6P9YR1nK|U3RADPik`Y>XP+UmG4;*Po#&UKk!*-|{r
zm2N9}ATZI<Un*~v&X{2W0f7NRd{b~>Fabq+MEU{>j6{1x>yS)BbD26AY@mnT2nyW-
z50gvbT7WQNB-SjJ4(cV(t=X#RXdlBUa!)xQEW?PxOCca2nPNRa>Q!MK#Xp@^Vuv9W
zNn^n>Ca_u_4^V1QZc+WU$onK1aYmg(f%C26*2D|*Ule29I5>`tr<e>U>(H*~x#TKu
zuXyYn=v=xVAq_6q4=BNiz(K&r1;Ux(uHaNk73`=pK7Kiv7e~}F0H?U1E;>}!uq%I6
zW6yOemIhO52Li_sUmA)d8B<2E$pkVXZcAV#I1mt;5Q3vbRwBWc$WJ6_%9(P4raVv{
zf%6nh1;M5WQbgi5B}+-rRH90PKNU+wu&MYeoxeRVm)k78tzly&mbyy`?@zWw=7WpN
z#IiL2Y04|Dhh7TYu>-V1tps=loKPnLI8N|SrY!WweO>eQXP*RdVMI+;Cjli;60*rc
zj|8m`yM&IhJh185Q-Q4n`nM0eEXgj(x}pNNm;-GC#T_J3lM(>;`!d^`7vCT6z_=`w
zrF)uPWBYo>mQ<Eu%W*DoI!WN4WU|n5fg}*ambLvB?3(a4cx_~{s6zXmm~;+^2S&I|
zE(fHKrNW9`3>4;Czv=W0>c52f)i!o;KQq*uQ?&EU3{|F7JF(CXpFcX7K#(~;Mb4B@
ZyG}Hpq#M#0baSc&M8$0kd`iQo{{uDLYnuQ7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu.hpp.C1858AFB6E6189D2.idx b/.cache/clangd/index/gemm_add_add_fastgelu.hpp.C1858AFB6E6189D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3cd773bc4ac72b6d124a6eab1e86b1424f5098ef
GIT binary patch
literal 3062
zcmY*b3sh5A7JWAmAn+pJgqHwGKnREl5^^CXkzYWjARoa$aYd|%6hnduIwaJRfR-?V
zT9=>tE4r*Lwq4@4YPI#>I#qB@Tcx!fY-(Afw*J7{8DtjofO)ysw^m;6Ui+N=&O7&<
z^KMywUf$yXg2<`NZ>U{fzto!`h+z6>t+%|9!6JymBtbOn>6u@<>Dy%TamfSUEXnH1
zsJbjzbSGi==IK$om!g4kbH^?iov)H!X^OAy`%$mnzv^RY+_s_bk9KO-ls_3bz9c#=
z?P3G_=r;Spzt8M_D=cg4t;>G%Wp4}oR_W`mhqYG!JAd#(x9IM`uB}$X^qD8gkvAHS
zH*U6=PV!!K2$dx#W!oS8<MlnUroux<l|NsthJ#B+-Wu3jawNuj(0i(=Z_h=kZEw)I
zJHy%?`tUYwi+;ttzb2hB4W_j#+vUIJ)x_|sYj)mH6erH7P92SCIvBC&`wN{=_u#-U
z$?d~j{`!;9f8-`NZDZHLp$^O9c<rjDKPPou%1cU*Ebkm-PhweS-M%sXwr}^^cer1j
zdHyi>pND*xo_#jUB<orB?l*(aRsE#VOR!1GIixJwo^<v7#XGF-k)*v#29@qusZ`Dn
zm{jTBINVUxT&WK3sVMnfdo$@*edUgOkKQ~rwDH;3jpzC!^zEZRJbmV6eq{bF@yv%^
zAFyQi4d?p;M?z)f$Y)cFVGI8^)dBv~+JF~_*=PT=vg=RzZSfu5+H<#i>*LQ<L?-JJ
zwy|c{RO`6<=(xUpoSrFnsjAts^;_QVKfCMetbbhyzP`*5>-gnHN8h2clArItn_m7A
z>sIRe7atppE7!+HdG|kTSUqdBWJd4f)LXm%dzxDxIl6!9?!Js$r?+>PcUkk6*OWhb
z&r4di=>0h-+V8!$ZA-+F&+8l~w(@y-Phv8}g+FB#$7gNrbPPp(cRr}U;MkRyUZzt`
zRaRT=>u$Q5c&g4h+J@D9l9&`iL;9lBY;^^}VFMx}7*HTz=w5J7f2^A%0vUr4b_i|Y
z=RILibyj%o!UKIz0Lwp=pbY+okntPgHz8--RaZ8XoM#JRku+TsR1-OVW9TPexbJ>`
zr8|=(xQx7vEu-aqSQGMpq_-5Q|Hj$l!wLxDP^G@5apO0}f5h4B9xZ##ND@JeznCqi
z{e8VA{0YnQDr+_!jF0&A#+>~s2}M$&FjVDvFDeX%VTcd;7?MDV%JT?nGMtQ|B$%Y~
zgij^IWDKRlbd@KyltpHl#^iydEPt^SAY_$WlQEQUEkvkZR*(7F6gG%Bdx||3L&f$|
zgc=o%n56?0IzS0xogl>tFkdGqb^^@187P_o=I;V2F5KP)ie0#U8&I?%d;j5m{VlUn
z=lGFS1dPZ;X^w=E7)pQ%nJBSS;1mp{!qiNZ*ff}iAsy6ZqQvIF91jZaZWFFP{O~eK
zQsFQh+d~Rc7!pDuhGJnXwxe+{4ny%U9zzPK!1gx<reFv{h@lLaf$e=J%*0SO%=RGW
zHCUYSmg~6ma^ELzt%RgDfS?UvC9`4CuEx=w1>0j;B(*qUaoR*z5d=w97gg(-=3~-K
zYmId$(pnR=#-miQ^ZcgXmfyMFBsJG>E*B{+^jnCb3VB5eqf1hi@=6aXQf)Ew<?Wkk
zf%1fMB@)ml=rL4PRJ9NZtSDOHL34{k;_tsRgUzOUkP{lnL`_l>jwBG3gPbD|L<OMX
zXabpBk(7~Z3}sToJcQ*xK602*K^w&yWsDK?5B$dm4`bBy+OT0u^0awAbSRQgB^HWA
zqQpXx$dy<q5{(jtQkq-30Oi2IGbAtptvk?UFY3Dbm4Kv7Jd=nCfkvsO41~-)Gsc<I
z%$bO*<JDoDMPP|RoCR9a5wZ%b7-vnjVo#_S)MK1YWQ#|BHnmNMkX>ZQIJ?H4gSbXf
zBgQ#^$N^r#Y6rlSoIvEnc_+{~0jAarM9nzg3Pi21Vs$I7=K>-Z&bxrdh3mBeQ5)i0
zpU#X_-)tAsC1y%AVWZVlV496y$voApz)+#N457NjI*;GM%Wf{{l|H6RJ_<(ZP?;pc
zL>(#<6;$a^ndHJ;9aAPB{P?BXzw^ypT7U<6*+?J;#$+RbX>eLL5|{yJWHSQSei*9E
zX8C*30s<(&1f)=k2`Hfw6Uc&D9)W>m_H*CB@C{>4#|sss2qZj-7)2oG$;BuF4NoIx
zBB1+8V4Te4=iySxXL9&RNe}pF7>!LxIh12lPy;odNk>vzsKt=2(1v}F96;ef-w%>%
z1&Y>JabYX^ela1^lWy#wfF|8K5VQ_WO{UH1t{*!yFlNJ_?zotwy?z{!6Q_v(w=a_g
zT6FC6z*v*i+Mu<b@kyT~j7!HHtd?cgIb#dn*lpqxFbjg<qh*+9$)c`xw_nPzE&9Vp
QC<x>8!ijJXM+njKe<(y83;+NC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu_xdl_bf16.cpp.D697A303BBCD00E1.idx b/.cache/clangd/index/gemm_add_add_fastgelu_xdl_bf16.cpp.D697A303BBCD00E1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1c7a5191101e4b522788e54e149451d448e102c6
GIT binary patch
literal 3228
zcmZve2~-nT7stZ{CL}X~Y$WXJVAvu%vIGQ0gix_+*aUGwCZdQaASy~**(-`jC@AR1
zr>1qmulUL0f)zz9ZWP3#MMOjaLBWF}QV^HE<n&>BPIFHFhne5K@4dP6zn4e@{QQ2v
zDU|tS!1j%@Em4{j3WW{6Nm~-vAD~hw9&`$2`{A09@cc(EtOuXzjg^-w4)I=Rv|JxB
zZCso24bQJlLIf6A2TTl{8NOTm^NN@~$M-z1=ky=ScoLbY$<jFJ8HPH|72QrK81{bW
zkUlY#PbPSmblJ*ZB-+;^UUl*VgQASS_s<WARy+TKtjt;V*juan+q#3_PB!IiS|zyk
z{!LU#;Dgg|cBRdnsEn`}e01ygk*9BLOFA7!il5zy+baFvvwOk313FGy5{>>?>+doz
z?`8Lod$LZ&_SlE~5O*_Hn$RIB`c;bECr5QY|GCq;B%(UU_FBUt&6blq?}xR-s|jCm
zp}S>Zz{Tki8{)Qax@DSbmCVk(e7^IS{?2wrMq%5%46~xQ;YN~hc?b1T!Pa{N+*w|!
z<D+cktm1b)WVY<&W!}%X*VT<=RYeszZuWSwfNs#&e$pk-uSS?wcHe-W#h){IF?7A5
zEz<TNxN+6B)-v}U;q0Amn~a|3CCBK++;|u~5`p%g9=mOS=aYv#E1pu*nseJ(=j#NI
zJbCF6pF1B1u3oum*nX_&U_gCDJ<)ND{%RdP<es~&XRkaMJ^FxCMBWh1IWdsu-PZQ3
z)Tu|V>GK+KNw%ck^f9s>pr3ZQQ(Kj^UlAO0Rc>~vPvRIYm9KBFy5<_1RAvy|X?mP5
z7V96bu5d1%)K$#bPK^r5w|p|W;&E4OtirWp$a(1Q%l(gckA?T!TSyLhT3~q%>CL%~
z4z2dhC#!;fZMmN+dwQx|Z*dd7lO4Z$u;E-LQInEYdbW3<CA+$t?Kq<U)1tJT2d3}*
zM^ayh^7(^~nd_V_l2W&QyuGhd$dg^zJwMYX)ypAw7tMRs8U6=+QS6oCp=sxEsqA`o
zNKDuMW9_da9c_f77dJj?2Wr;XxqH!i%1Rn5T?)4vyEbRPo5qap(phmmBkIhPo_woq
z@y6n@<(iWN->h=%FmpO~!8?#qUY<#NF@C3YB0c!+e;Yn6zdF9i`C*K8z}xzh7B9uQ
zDa+Pu@YT997Qf@8TanmSX5;HzqPTJ5Zu>2trTTkv#HF3j>YcqxwG);uz7!PWwvie0
zb^)!~wz$fp^R8TQ<3Cl~J+k)A+SX4Sc~P=%tfn_AY|F(BUEd|dt^bW?;^G!vU6gM6
zvS_PmO0MGK;NHj$)feXO;ZHRE`23aFi?b=beHUfrv);q~!wDa*oOd0nXuT|dG+tFP
zo<@Hk9`?xZb3<Tt#hgz&_iYbVFn-$i3H8sNf0>e$vT;p0LZRrb6><_ayc#iN4TVAq
zQK6AaKnJTne0hMU=0UG?3<&|4j?#5i0<js%61Dg?C1Xf9z-*MQqY{Wk$;rddM;97l
zNHD<qs6Iy}5ZnEka;a`iQi3601I$PHER{g)S>95XQ(JfnLzV)JqNuJ)AogdZ?eAN9
zJs(5D0M<se`6_`}mcGP(H22mphO7jbi*j{T0`a=Yoqs3<dA%621YjCUW2*$>%~ZW6
zZFHpqL!<x;PytIN5T{FPzduX9XJg1RfEg%*r4ooU?v~ez7GIc)ArSy$D8^C=#NW4A
z)krLh>Tya8z;>E;+A4v#V4j~wWZS1T7_tapEmTWaB@h>vp0J~NUOtN<kpP>ZCVZ7Z
ztT^8@qyO#b2@F{QFplCZl|Wo4rN`Rs3}ax(Vt{o~U8YJP?&u3(rdsZ(#*ozj^H3gB
zB@lN7{C3H`x#1j!tOA&cGIdolg}W`Ho6JkPF(e3J7Rq9(1ma;_toWdf$}wa)z#NpL
zqY{YUOl}$*?zm>kq(lL1$Tys!5{MIJYm>6IC=||U(}tc<1*U<M`g;0i&|}cZ8jOE6
zFv9qUadFZ@VzCtFKpTS~rHA2;ILS5Qnn5GMjX;pn$zOtqA`RlG27E2h)Ktbl%#4#n
zDv>=Eg_BkiD=FL`IvwOFZLYSon3xY2#z}LD`P7l2`@w0H)>liB5^G?}S0z{Sn=^)(
z2u|Vx+!$H|JeMBV5I$Gnp<WxWcLz>`dYZVJgu&d%)`s}H3cttTHjB;Tb^ZNnVHVg|
z*;8u!8WS^NE^IFFJj%{e3tB)dfk8NFDX|QKC4dbFQk31Mma<!eC{(MPwZwX=Zd(`n
zHi?$GECFXU^f3&E21@(pLS}vU(~HX6p@u8qZP>Kn=*o`$vITB}Tfo}D)&)VzzEumd
z!ku6cY+&FuW%t6My_-1|t~I{xS~#g?pk)I47$;E!)C9g`oMdnr0(cgj#JHFMZi<sk
zE>i%T4CaH9D2D*t1vtr$aQjk4p+q?43c2*enX!TM(NvtI8PJU3d%#I;18rky8JwhZ
z>3lfZ;X>e9l(PixQkK?d9<9?*sfm-iY+W848#u{kvw5&zaFWC3@L*fxB#+JG!3hFu
z1q!JgLU3JhdXk8ToBm((-NvTbH;2J_`Fi~6uy=4$peL9PErF9FiD=3w@FZYe<@8cp
zmqD=My6_MnNIApQg3JjU7zFDKf|Qd?Eod{5G8Oc<??I4qo~Z?K2}2k(v}F20PPXR}
zU^qLGT_CgvEDXq14!bY8Tp`y6Hqx8Q?}D`39gIM(P%QL^{Qyl1a+Q<tOD=I-gE%|Y
z595Pf1-8#u6EslQ&l<!o1Re=o??;FN_3I-Daa8?!5Fv)tuRkEfNA>HyR3eY6w%&Ux
zF*$XeD2oGA8vM>rj7~CB{t>Md)rU%&6jn^D4N9CuqGc2rybVG&l5yZa9_1@pAXQTf
zrD@YKoXOJFk+6Mz%-IYHN5JLj8@Rc9c+T<io@=Wo5}UbLPW!jJd?RC%=`)C#LJKSF
fS$6ggPP1K2oegaq=j-wP0@fl6Lg)EwETsGegg`7F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu_xdl_fp16.cpp.BE72E2A90F751444.idx b/.cache/clangd/index/gemm_add_add_fastgelu_xdl_fp16.cpp.BE72E2A90F751444.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fa5e345e3ee6e0795d38528fd61b5e1d0d03dd75
GIT binary patch
literal 3226
zcmZuz3pkWnAD?-LcgDQ$FgGJ(M6U*8s!>Y0q#I+DTn4#rZM-9^v_<-etxaEQRdP%5
zm6DAm`67~9m29O}>0)&yt<>&^qG-EleP`Zhk7l1|=XvJ$IOq5KpYuQG{QiHV_EIWy
za2jn9>6H|x*|1WNMx*h-H(^8kSNRMY&BcI5OFCE=6q@@`$?X_hKPzp@q>U9x-L>WY
z3r^gt+glOczA#{Sbe_|k*L`;x$iS~lFF(kz$SS?y_PlS&WSYUZGPcH^S$kAcwd+&a
zG;8}qquDoJdiFdDbu9V=4Ze6a_;qTs#(qWDXuLlDA^2h+ch|=<)??pQJzpPwKRD1g
zCEU%e_0;Y7>HX8QG%G|uuSs^<R5;JIOH<;a>{6!|c>82{gcWAoxwA6UXYJjG{VmZy
z_?sWxP;MU-C#e!@%xVe)&$89M<?Rc%AMAQMw%J!2EE*`@Gq^7IqJ!OK-_up96u0b~
z8ag}p`v~EQS&rxHNu%?5o6NU2)<6F7#ije~lHB$?+pK~HL(M{}C)Y7gTMc}F+a@IS
z-Qe$5?vb8la=qB={wXr?zT2t2*s5ltw|nEAUHo`E`yZ^med?rJi+*u3*e;s;u|fG2
zVnUByZP%lC8q+3etlT;GVEuxmsLs&K57s;G`!T=ieMQ5Jn>AyZ&%&7V<A#f8G&RZA
z9`{!Z0%IC1roJ-?WrWxA6J#Os5chM2)#jYy#m-l^?^u{~eNso*(VYIIV=sh_Lput^
z5ed-|IV#rji=~+wT0=)Vw&_xA>zlXz5~1xcrYE1ZtaZw1EyyFqZz~6zVsAR^$avJ;
z9_g9VnA@@MO8nT$Tg<O5tNzF9O}TLVQTS5Mrji>z&4wp39~_D~oM?!p1=+G|xRR>h
zD_NP}U0JU@bl7fl)pI%K8+0Nm(x*Ff9V6VLUMecTChyoS_1uEqN%L%JiORFqpBR?X
zZgK5RQ+UJakp0u$ZyjtM>U=AocRk-PZ^+t^S3mgvCtLc&(a44aLk^DR@55@gJf0CE
zvvRC-ev@vT!D*A;RoCQfzQ6S1WWSF&Jqh+N3k>q>8fu$o&M~{U<(6-<%dqpnh)bv8
z7Vn0H=KIx8`1fsIeayK%v4D51bMxuait9;PnHP{lD{|RmPW>13y`Q+B`lhPp=u3|C
z(W>0M9nv~w;Na>PrFWMGtILj6ZdWM-hTjRzTnc3wQA2(<CE>0Ix^uI+CL?F&)}{Ac
zSvTWQUH7v{r-o+EjpX%@op#j+#9WL2*JjJZnO@c_mQ38}vHx6Yr^(gyj@o;XjSk9i
z*>A<mV_syJ%_%tB(K1VU3Y)n)tIa$1VP%iiQ2u4k&TgqsQ}UlhrDG{6Juk(bvW7kL
z`Z6pI966j=$N69peRTc6;FE+)rwRjY6JO6h*}Z5aAvrFx0-@22qow?Ky4w{DiKNj;
zDJnIi0Msw_5#@Nf>K$-P!;m0A8=wY86hPXF<jULJTN5!P6wo}BXGj60<%x-}o)0fE
z!;nBgPe3Q|DS))%kK}Xb->M2Q#0StKRK%qK(ykS4MLW*smSc!NpivYxq5#sK?5$b7
z{y*)-kYGTwP?m@SNNdvk?1wY1y~2>?fEJ(vLkb`r6OnSS#AMGi4Dkar6J_!!fb=?s
zaVra5UV<TNK%1Z@TnZqararr~n*6}SkPtw#Q8t$XNN>AUaaQhoW<G|50UARwE(MU@
z*(R$~+3st?X{!LO&{MD|fOPf(B|W13&qxee3TS;)--rT8=NFbKn68c07!m<!bJSc!
z0i;XndnUbjGhButUjQ0MaV`arKCd>2Rip&7F~k?pMyL^o0!TmX_2Q)3ZmPwQF99t?
zg&YbX-Q{)ToYT$b(-^V>&>WOwM8P=SZ4=dMUC@mo0f6SBTn+`0?i=xFD%$vIfJ0je
zXt78<i2_K!!ktS-S!fQ1gaew7@(n3~bi5`yAzhzF;}5s4?paoX(Q$Hu@dPXAF=%8u
z`yUOAu>WRUoU~Ed_`?`zV<6Ic819IZ0yBXXG!onhh_p`rEF$*NiDEi@7HDb)`)_8(
zNg|a<A6Ma|OeIsp{h`xAjMnBlOFI*b;KDd*t+F0Z*<;V^q1dX?5Ph7~H`SLw&)_6#
zib~*mILQ{UMbNA`i3u<f^d?Sn1RN23P*_ahaN0`J$v{m+!VDG{T-bYaTR#WEN!$dN
zK&OFgG8Ty8HF^K%*;KcY{l~%cn9nv3hOrU1o7bMNQg(uAE{`kh`pYRW3hb*bQJsAy
z#AFx?YZY9Nwor9M9)uqh;iRp~HUQoztX~kKEnc0F@90FX&NJAl?8cvA;}Z8)d5EJQ
z$Wgpd90={Y^_w&QZt8y8pq*YO{sK<JYXzxm8^Gr+aC6)S-VMBMAkwx09gz%ofFjrr
zfOWOa0E*z7f-`8F0GtL+QV?@KA6u%Dj=Lg9UH0F%ztu-EaFS`tl)#I_NtP)~0^b-X
z4Fm=v*xBJiU~_FRfx8rK?X`|FY_8G6Nh6+-5Vj4R<neeyST8uq=kbNGZgEn`6AEDm
zfrkSp*ES)zF36rF*1}Cs>W##;&T4rD@)a42EMe{7q=~VKC457il&j?9FAFmP>uS4~
z&bn-Z2iJv30Fkzb>4>a}sZa#H4@BBdrXyNMB#(>!GAa;h`<aeNK!~9TS^|i)T}?++
zN++tv-PZ4Cd4QkpdIT6rAy;_AXNUI(qP0!$bF@Gzm<kJIsAfw5>;7~z5G!?-dP1MS
z>+8y%_WVnN3GQ`_PJD~Nl)yL(A+mMHM-ZY|cie{%{kr2(g!rU8-pwHPFm%@Yz#u-3
zkK;9KKobYQ^5dftth9ebX@UAM$hltp4{NTrD3B-(O#`QaB#vAI{^QXWYrGkH`Y4lS
zfZ-gjk)ev`zR;S-R`E>)!U?8xoLpSzy3LzE%~<Yi<!Ed1PqQL3iMiz@VzSgmW;aD)
dKiy&0?1?kQQ)etP7Ad`=5szgHJn2hl{|A#jDL()J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu_xdl_fp32.cpp.1D688E059DA5C6D3.idx b/.cache/clangd/index/gemm_add_add_fastgelu_xdl_fp32.cpp.1D688E059DA5C6D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c1d7f6846bdda0faf595a15d923e5c5d40137b7e
GIT binary patch
literal 3220
zcmZuz3piBiA3hGwj5)_#&CCu{IbqyFTajDJB}Qt~<W?w|GfI?l3$clRC4aH9p`sh=
zPyIzvy4jUdM5=A=qS{JSE2*up-Rw8>?5El1*?AuCGv9gN-*@?)_x*;<+tV|LKoEDz
zJ87LFew7x2ARO=)7atp0q=_I-ItY?fQt!Vkrq`a`HE!N=${^-U=#MOyV0qkEk=f1+
z`C5u|RBFYV@?p@aPGPz6^csWcDbqGKI*fIgm?Apq<_yJjTHO)xiR}B=tR>S$<E4v+
zx0mgYwA=Tt+9sW4^Q8On<*+jiDO)r`Ms3oI$bn6HoQC}iy;jbB{b;Z#@=Q|0K}GY`
z)LU}H8!nx~?#B^h_7}D^bZkGHbMfey^W-V~H+8f46&20m(tG9g-Mc${%X*xi%T`tO
z*oRoOby2pJ@uKp!rIp(l4wst7hYl>Tw(3bUc=Xv~tEpf1!RXNZk<_WBH&8um?{j9$
zqoPw|uD?t(e`mpPiaD3Ocv~&s)3maJ=2i9Zv|SD5y_Oa2)#F{3SCcgQvX7Q^h53z;
ze)s#b5!oz@vOn8a-`qaGt!Z6_(Bz3<nD3(ElC?hQoa1p^hE>W)+^#!X*67_Xzm|i|
zw{sm*mT@-Q$7H^0IbV={x-Z%y**x=&f1IL3F+Qnv4>P^zS(+jvQ+_X?tX%THTd{MO
zzgl?;vs+^_Agc+e$x7shJL$XaiO<DZ+l=?PpQ!gR>3;3jX`a$IYHQmt(B5F2&vwpS
zH0cHsTcIm_oAq<?)1ht$Gv6v+SZn_M4aR$Cnzzn+;r%G{=aSV-{@ERsS93>Fy#8;w
ztN(Lz!jy|!8>(pE4}Yp&f6~_WS-oAbZg8Dr#HRAJI}X$6g@?_qyKQJVQZUtCqJK@4
z@a>7Q%6&G|oc-$`?{vrrzHA$vdap33TyipKW})~KPl`__jaW<e=_Rcha@}IsNsH-k
z3lD8!yT;Afv8nguN5f#ZB`y2yJ?>pTW>lt?^y<G?Q`<U!Egy(#{k-u;SY=Mo>EEvy
z6q<E87cKN@UGUy@wno9cA)nmNriRs$pss;C%sUUGL&gWcZR<T87<uxy=ZRzTwo-#D
zxu5p_Xr0_!l#_Tp$82ciG~eI9B{P1E|LkM7-&lXmTt;zYcEp7c#r(8)jYhvadM5|T
zqCB*>#8*4sk2g<P-nKl8vUknPtF3LWJYp{xO_*`9>5`6%h_NvsuV-nL{aWH$K_s5r
zS=bjG@>7oRTIs2zIok@SZF*|@$Ma(9W5ttIf$@z|chVB0*DOl0>T+6Dm!I)+@Lb~L
z<m`f-QM*Fqv*x(}g#CG?T+-p_!j4&XFP(m%Pp{<Z%E<RWoSpr$`gZB=-mweSV>@-;
zhX(c@8kbA-e_@W#F?Bxfy-jBN8*=!i`w$YBye_;7MG$_32{%^5r3FXB5ri_vj15!(
z60G_Nc6&Q(6}x2Ms6U|T7@eg8NE=aX%TAB>1RPxkXg0=XssPfK2?@gkZ+#4K)DO@?
zOvqIMq-{o%FEoxw_u}X`faYU-whAEa<j?EATQM@Ki>v~)NFX9r0BPr{&I3En7gXZt
z5<qKX+AI}7+KaLEyZ$9t@^CZ|&@_xDPywVB8NSosX15OG=n6n{F)mXDkX{qA`Ij=`
zt|vI^3up|(I4Xej22FlD9a~X`qcT7XFacWykj{{u&#I-~b8s{W&^nk7TLqBLyj6AH
za&f~v99<4*2F75k0Mc2V=JisuJxv6%8qn5S)^rs>I^We(BjoPqa2#C(Xe~^Or2<G7
z?mK2fbG}rIqalD6V`6~{AYFF0k9_*(?J*n;1vHM~Y!yJdQKqxjW^*6|M;8N{i7|Cm
z0O_87Z{1Y0lsX&@12hlg>8b$I54<}rINWSLgQF_}O<)8|1rzi`Q+c~&??W723TPI_
z(p3SZhY82Bk92G|js^pogK?NDfOM=PA}&oELAY<*qxu5Ma1DYI@`XmwW6;PNjDIyS
z%J_$I2}&ZBEP*x9#z3U>FkF$KxCUG!=oGjR5GkGfMMUn=AP;N6zC%-MGX7y^f+ADN
z?Gq}3GLxFg;QG+%U}~k!)uukGLDo(bLC1rsm3CK?Ig;*heS$KPnoOJwrU1;WjDT8+
zj0}e*3-<nycQf;aE=o|i5Eny>faCMI`tbM*_dl_6c~^8Cv|T(~90+SeZa1%QtoFPQ
z9<w+sy$654Vps*%Rc1?VT`_44YhlR1d6WrL6U`@mp@^VNrKU?^r(p0vi86I+B|A09
zf{9KMl)2P=qP>Ye9_^MvcD`VbA~%sA?Dp2QhW|7@d~{CvI6xE%AH(p1tt*rJWeI!|
zZVGz?BMd~!{Hlq}@tIHrgA6RGOfeL}0E06q(+eL1huDzzUyf}oHJ*5lyJg2tjyyXn
zrxTQxNJ|W7ASi8-wiw!ipfC|8hD{|XnusQb@1LOPJh}kh<^-j~(-FXf5)^~S5Wo`=
z6u~0|Fv)N`;0VgA1n%{~)_#ec*?dZipqLz{9=tUOip^o`!88#Rhr`i>sU|2Mho=WG
z5ZEHnKjj?+=LJ2W$n|j18m)Kh+HJ27YZ4Tn&o_k0A}9f0U<jQ+P!>{)34g$yfO(ad
zm)g7xk^|?3n*bu^6{aSVkXBFxI{-w=i%d<lflQtd{cU$3QeJ0jA|9y^MK62}i@DpK
z4*{Q9TUsxKu7HgJwaUBhOD)%gYX$T3`qY-C^k2U=0JX-B#$GT6V}lR!tp~!U(m=z;
zHOO=nt_8;5p=7@L_z+4SRv!<d<O}ujN0j`eKF-l3cWJ84_g<6yFfm4A717|=82o=9
zD~~f$evvCd_0gn0q)SDAtSIR~<qAXr9|OrcDjNKhAqx}>HMO)c8eIn`blEJXl;hzh
z;V`6JAy4lsk-dYH^BkAC^Q`%njz)H7Q~vF#z(737keq64Vrp(N&1U-7Gi_&2o~3U!
T!<{ei^o~I12e^7^_#l4)6lxZG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu_xdl_int8.cpp.7C822BDA45FD8B5C.idx b/.cache/clangd/index/gemm_add_add_fastgelu_xdl_int8.cpp.7C822BDA45FD8B5C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6ac8ee6e62621b5d13ef810009d60fcce7b2d083
GIT binary patch
literal 3226
zcmZuz3pkW%8=m>_VZLvg8FO-)mXE=hq!5Kv6f@#?Du?K>e4`S|saR1_PMwxeq!j63
z*>td3{fTU;R>tmMZKW-RY>Sl*`AhlVnSWo+Uf0fbxvzQO`+lDHbU)AVm%6(b;1r4*
zDNT-%C4Q++p-@=hFCj60W3d{A;-pQXB$w5%3`yv+XTF-cQ)BU<dLYrA;mGz*+hDkU
z@^%PyIm?8(j_T{NBaN>ke3IuRb-tG=8N8y|bnUFV5w$MmTPx4S=6Nla&(j9OkwYVE
zACIQbZ#))R88ZAHX-E;ud3^NJ%?F$L%}G;8vLHNtinHs(oy5^G_7_d*C6RmEvP0~{
z+BFxV7YE;N&$o)r_PLbZb~QS=?RIFMyl2F3t`8RKzbd)#w+CjmgT-N4=-4xHIW08s
zVUK^ncV_D)XMS9@vBc6O(dxC})9}0dQLj6Py0jQU+Rl{^C&_^VuWq*9x%kco)1k`W
z%92K|8DC-SvH8YgFr(+y!v2Rvy2lFrsLrp;<hO1V2xUl}K<|9bf*UVevky8ja|LCy
znP(*x+Mafnn<nn^JBpVTbYHQ|K}`FC2h8@Jx|`qJ93^$FY;8_&@_SOcmr``8ir&BF
za>RipML*1RX;`H`Fe$zHIL^MRv9Klk;`cTk@e^6Sckl1KRhqVO6Qix6)?Mu8ThWt#
zxqDIXtmsZrp=VuST~6%Ewyk!%J^L2&XT~pE<yuqkTKj0kRW3pe#<}wvUbHuu<}?4j
zGlu&kB}z7nKbo`BLY6u{Gk(Q;_4$AH+8auxW>rVBoJ6iec)#dm@A#bhM?b9@K0CH2
z>FJJ!Kiuk=L1R1ry!QP5Q0?i5kuif0n|a4_9}6~?C(caBkR9vqJu%);aj0V4%kF@!
zeoD>6xzh8qkfNejxeGEvZ_JB!eOerR%ItiwZL#ym)rVNlu7QFIJJFlY#>`KPk26}%
z6}4997)Q0<F?<$ty?4ph-4!_waXAfp<R0$YZ@T;U{<pCFg7n@m1s{8gTl_;WL_WxM
znB`ZKGP|TWu!@!$i`eh7<QSG`#>u0Hg2hD!r|P%GeLVj`v^hJZ>RS4lt``SI+hUnc
z;~vP>-n6wgPmFCVE_cZTR}||petvPdk|F+i%qRQfn&we&yDsUmxYSt3it2*GY++Jh
z;hWU{>i>8jzI$p<^=?TP8=v49IF&`^MU40gTSK3AIhwi|cl_>`8b$Zm|03BYre)-D
z{Pyfkr6VsV0(hIxbQbjeF}c`9aCF~wj`Lt?$2RXq>jU&!>G_*qneVgr>~x#w;9}$x
zXERcpzj@#3^&6fTl^?l(Dx}ThrPW>oW=ld@+1p85@2Pj=w|=cZ9k9h<#{=%7IZLkp
z9u;(S>PYUb`qmwCZF!V`h0UGoO<v3B2`Mq*)d+>cjTGv{Q(bOhNH~Q;3Q?hf5<vYb
zKfD5|v-%O23=CNbXl+!Rp#(^qkW8_BX?qfegaDd_vSug&(&D6~;a8)}4KQQ{pmkAQ
z9VI~8`d!M^rg2FLhWrcAJe0>&0;HX*<;Sxd51qphKR}}>%1{ENJ!reW?(_TMAch11
zS_{?UDFM>54Bt7Ud3S~}WDTI%D0_wyAiW_h^_NQi{$3361+*rr$x;HOqsQKj4Mp+#
z>69-4t<TdpRsy8AsBzo1&^479;tyy(%4aG8(i#4ZITy(HEDQ+-G!3OOl>q6?&gw?7
zPs3siSp{ee#h6NfbdFq9FR>_U!71wjZLMytr36UlFL9@a-TxGhA>M%2Ks6XjfOK)`
z8EZ}FYZovi4A25pz*7RGE3fnz_rD!IgCVN{jiWeI36O5`*N(DI4WeO)51<(+gRTTf
zKk1Xww_AL58AH|qnuBuaN`Uk;>AkB9I$FQOkhOrOqjZK6rs;0;h<3A*ZVU+kG!tdg
zl>q5s+_7>}3oXEqP(bUTIy002>3CUWLY4-FqBGjQsVA@!qvE74SJwo33>ukA`$q#K
zw7(e_C(R}1elP~w7>E=ehAZME+kkBXjRY40B88Jbi-;mBagqw}9hzE=_BS)*B(a^y
znpWYYNFwrw>qDo57=_JMQaciEFfmS=NzA6V4BZcQqp-e8h(8eyLp)2aAMD5+q$4<q
z^KnD?_TXGxwmv+U*U?@pm$Ad&fcq3U2!dd2a9GvziA^KX8aSz;r(pnHhm)utY5+Hk
zlQcGs3+;}R7#rh4cjF|TP3J;Og#FYS+w`sbuV57>i^+NRmlamu%-g_jaVuBDNliUX
zLs*qKsimi72+xR<+H7qeY(;S2;N*&`RynyLVG4H-6M-95)V7Lf8Q}{>IB6lV2!M|O
zOCID^)VoT^X)1AH`sHBhgUpJ$SCQpXi9^#eoV1i!PCxM0<xAVe!S=r3^!l#)E8tSQ
z(i(QRbiepsF&(JC8cxIe1Upgmj?YWr0^A%vBD`iGQgjd%kqEbiBG^hmQ7HNd6v4*?
zXHfJHI1M~%P0as%J3@(Y+HY_cWK#4L7;=19pIO9=){E*m$zU-!uzlbpi^bx=vcgFn
zmJSD&H%@X`91iRxaHrrHie?1!f~}B5EL^lseJrMZ-tA%F5+0Xl1WODj`CPscd=Q)z
zOT^Q!33meWD!Q0TUK+uIdEq92NYTqwL}r8)6hQ|Ak)orih_(<Z)1trp3`B~)rXpe!
z`cO3JYji{>%lQO&Zfmi%7knIeF(6ja^ghS3g={P6_K}M_0<<2@H2|?fN1+Ea?Zm*d
z{MoP85t`t_rl`bD1g-_fUn4}m>i7ggoKzhTAjF{RcoHE#s*d-m5&P9t^1W9hK1`3}
zWzpc55&V}QACX|9_!6Oj|9@og3q_6eXv>sc((|lAB4m_^7&03Cq)|L&UTW$ZsHT=S
zhSQmh84}h~S2Gq(qQhr%boJ&haB^PgvS{&auGrDU-r|dYn&lZ73XF^iQ=z%Ya+dX+
cxwi8hX4>gn*|>3e?$StPS>O^6>T=5e0MCji<p2Nx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp.DFA0EFAE6F444F2C.idx b/.cache/clangd/index/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp.DFA0EFAE6F444F2C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2193d2333a475509e78035a1305ce83e465d4bfe
GIT binary patch
literal 3256
zcmZve2{={x8^`ZGIIernz1)?1T}h5T31!P)Ww{7tOB>maZKRr_5-qe$i@l+k!i>sN
z2(22;U)r-&RHO_O5*aNDP4WNT`=6igJkQ+c@qNxYpZB-Dzwi4#yj{iOb2x?KK)P;=
z35yMur%)&?@V7a3Q`7@F3WZOnP_`Yv;J0c+r<KZ+DTByo9u}Q*G8At+N3UaQdt^EX
zyQdM?xutzQ?SBnz)xW#tkJBs9R_!>~G|}H}d9*jBD#dgS_gOO8?HCup?`(0n(;oXi
z!tSc(@6Vgx|1>c$C_g?b&No?x_uoI9uJ_-^eM)aFZu~K6KpkuCXc{w#9iNQJG11iQ
zpV)Tg@(fnSt3cKJiSY||<}I}C>U+7szRM@^guBhn<P{giYvVJzAMdyPq!7SFN1KgW
z_LZv-1)nkvEHW>dFf=s2VrK7rFk-x5r1bH<q)rWOTYvJ6;QZ+g)fErrZGO%y3ljR3
z>|md$ne%o1gJJtq?zwwoHn#ox>Uym(>xh}*-1OSQnyPh9S!PL@8p6u6uY=z8G`w0i
zn-zUxsmhZ=dCrdm%bz^Rt~_phP(OE<*J5Ay5UST$Fk!(|koR`u3PtkoKh!-E8zRxi
z%VVteHW+`8mx<rMt#@Lb{gA*5bM9yx^bNOb`<fk<)1XYaaRh47TP~#L$9<nJEXS*b
zqK_|Wlc~+s3@Zh1rAKsjeYWX^B0XQDk<zzkJxg@lF8CGltn&v9ZU!|})X%W|QIJir
zLMnqgPnykLE_72EK8K#P&9944(63VzZI4wybdt#Qc5E2tY1uf$FY)t9IIwzfAhIoL
zN0`RPf`EYPpjE3M^G_x0D(Ot{NDYs9b2!+f+x=ceQQgyc{fzPuufouW4hIvQMjhv+
znyYmzPYi6b<Wq;ltv6!3vFmv;_Rn&BPuf-bnrGQg@*dKJT4$c=@1V~+U-9@-?`xIg
zh8pvh(^(9i*2x)v>8$C{I$-$1QR@-kZflB}cu(MLRrAkFJq8PMdJ8V5l%MBCdk@Fg
z@D9vx(s}r(%&%wrG1H%_bbYx6$zCn%Vozq@3-~Y)uzKo@+tSY9C3dwUTyb>nspf#<
zw6FP9F<ZKh;FlJqWGI__U=P_8FaB<!!no%zX;ap3epBcj`DMay(BG}V=bFBK{-L9)
zO*d~RMdqcnzTFt%)U;FGAK#yN3a|9lcC^U1-(B1`@XkF{c;p9N_1~{I-^xhcXH~Kx
zt+=SiV}n(0b6R(AkoA>4^P8hGzu27_^Z9HYk2K%e8dP)l(xDuOmr3_y#|PV6^V8DW
zBh!1tN{NyDTZ)CHEvF|VdlQN}sCggnsLrw;emcgAGHSg=yj3YKwE0p$F=fY$dRi{<
zZR-z=yD8v|ohi8-xzo&{ZfN(~kyl+^Q|8?vO|6GJYoA-R*SW`cIL_(p=sxu4>lTfp
zd9@XfV((wnzrM9*>*^C1ozHJs#kfaKQSOQIY&XZpgp?r^3NKuTy@_gf3qwLE6jBe>
zQ<n;;eCdy$<!UQ`%q|&2mII8Tn6gwL)+SYqS{xhWFk}_L9F)V93dBZnaU=bsZt57~
z3$PFsvZVsC$yof=YoA<lF=Q#gYN(ouR3Nr3Ysvqi_DCs)cmYg9Y06T8*qNS`(dTs|
z9Ya<COh@T_sX!c->^XPz(A^OXSqm@^<uRoKaio93(<0T>UJUUBSP4~PNd@AKa=b<w
zy0!>Iya5)V0u`x1oa|kDpo;v=!jP2!<0!5o6^QpdD62JEc6kwotOl5YGE}4j@qreD
z3oiQE^*ALGU{iThnp7apTr8&gw@rp%hzG!msG_n|AkN7vHc_;#tHKa}fHhG~zEmJC
zy7W@()rZky3<&_3i858B0`WC(rSD7<R?snI8Ne)*#gGcbPr}tYndW|b8I)jvMSPK#
zR3LuV=gLUb-+mrLf&k{De1=pY?s9FqYTev$5kmq2Rz{VTr814X^+FqUbGtFb2VgeJ
zW=I9%5!|Ne3k}V}kTn2vQ7%&|5N`?#-yEPop|D3AqhI<JVN{$H@`T#ZW6;P{`o9_&
zq5s3UIBDQw;04D(8-pN;hvAAi$y4WPLnFb3K#;`A--3v2D)BQFUJEp}9Q_|=#z`WP
z*gqYGlSVE^-f(^BbTCF@bD61a2nRSZPU^epPj4BzAM8eAeVHNNL<k(RB=@)U<~?s1
z2u?CpnIdQna4sHC2+!sJuwnhR3h@h&Pd1Cq?fUzYAO;IaibrMv5kbHOU=@LlOG-#4
z$cgZTK{#pXV(0_U3M&f?k(8Cpkb_j>h)jZJyUd<W&=xnxMx&J$o?wq6dyy}6Y0{3%
zd+WP<u1G%ni2~qfSe;<&lG6OP1XF{l2a^D66a+~MR3^xPX%2&6r2<n)$`l5{3I%77
zlqdWQ?%0H|`gUwR7rp8G%JMF*{4{VWl!lXv0!0y=0Vin!nh4%MPAUnML@=p1Nf*#X
z@S<@N<6wN)k8u*`;Cy&coMdnqe0U<9ROTr2VJ*V#fFnqH1>9?XQlD-pv!PlZCs`~O
z7q$hQ<ghqgSPwYKWpTN%E^(62;&WjOC^9ZReLsKSe2{xhOU*F249pxztEBh9<v?mk
zVm%BxC;u^~asHhVIh<7Esm+8XgOdWDU?wyOP8z!yPum7=5X>uSS~By}2^O3e8UO@I
zT9-`FX(~}Q?Jal&5GHA1GGV%eF`NtL4+KdXnM}||B7QpP?_vQ#l6EE&#36(*=#A&h
zW9<F5g&<R=MyBr2Q!uSytfbd{8_Uz-8N=#(SG~)J)^4f}#_HMVIm03vAL`1S(;qZT
z0pxWzLL?z@E%2U!5Sg;?jR?^s`~DmuUdX;*l_P$YgX1Zi!q$P#5B~4p6uQ|&@(?PK
z|Np|_1BDVc{U}G?tht^4=bgjDa-rd2l+YM*T?Bwj!rbNL6;MT*5{5HWl$kCpM|)is
z-G!~n;i?I&tQXk+$8O=GIXoj9Z3}(PfBVW;7irAYnnmd785qtsnQLl3-*U!0p|P0*
Rk1uwOK%D#*J5${#{{wDWE$9FM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_add_mean_meansquare_xdl_fp16.cpp.784D28B0B387AEB3.idx b/.cache/clangd/index/gemm_add_add_mean_meansquare_xdl_fp16.cpp.784D28B0B387AEB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e11eb70ed7e87be3b51950fd6892d927ba687a7f
GIT binary patch
literal 9570
zcmb7Kdt8m#*FWp&bh@0UyG}Z%)BRMsr;BbXBwaYEh`|_giz0*`x1rL-(CMZm-6j-<
zkQtW@X5=~=#zY#%T_Zmv$z>?>+vmKq+uQqj|9Sg-Ea!W^`+4@-Yp=c5+N)wC!osG=
zc)U@wBe;cgmd)hzcsv92FLl|{`9CADT93!$D$Y$>8289WzpeUdz2uz7Z@mqkYc4u%
z9bz{Bes;e3=qb0$FBU}?8U6m%y^<Gat)|U$802r1Jl|seyFZ3S=|^hqvKv=2Gh^uG
z*O#5w_WeF@gZ%tD+t!rkYj3<ve|`A##`~{}|4v@gIOB+KaDV~pD(6Q}Y&m^L<L9H7
zmi6{)YutNTcPU|IKvedozvpbuu>ITrE89)a9`rw7v9g#=8Q<f1{~7lu@3q4)_0Z$9
zGW?Ejj6AHq-Tl|=S0YT7#BDi$s3c>yPJHG!@g)<ii^{)?{K@I<>96aHoGw3aXkENo
z-l-O+U45*H_hJcu%Ec{?N0^Q?*Qd;CUahv`N$Q@%?kjsm#qSElQx6`DSnza@TFJx7
zlBP4;u2((u|4lBwllyml^jATxA^Ywwv;J15S$FcW%WSXmz+1_SwKJOb_XgH8vGIH2
z^Va{)%rl<}(>yK*cBQ%`I<kz*{vn%#cV~`y+&Kr1HLskyBxT?2QJvCNaTi<tmY+%-
zmp$d})$qAZwPPhGoyK+~?l?MX%k>c9Tn~%NZnuNki^@vIhASK{3_1LT@$c6R9e1V<
z=E@_-XJk)~s<T}b`17^VDSZta;v24cvJsY{`Dg3cnKeUO1!}i0M%slATcM~MyZ7Wz
z8EPj+jEl7#veE0gd3VW;&iWelI*mP9o6j%FZ2EI@OR#j>_>{!N8&kenQGM>kyomk1
zQ?~D~n>u%Scf+?QFPKH8H|+INk8-*&v|X)aRPDDF+eG8ezPq-j@yw_g$9-jC$JCG1
zNi3RV{5@}+&a(4#zn)OzS6Wc`Lus43^OUKNyjPlWPM(*43EaKg%Rlb)i8UKnElPhC
zIC-S;w?YeX*oK4$^_%`4RBaXU;-=xd9M>IAbKZ-;9iGyZuqJQ$#_zW{?>zn7Xo057
zTE&ML)4y6Kh+?c>%$eXPxwUSw!OXNy=WpCr^_zb^)+2SR)vVJ7AHt`<lVy&{-fa1e
zMucE=;h?gRZMkpr)Z>?IetP;wzxDnYotYhN>(t%94E{4;+o(8DyJY8_z|L2Vwg&Oh
zvvu4$51f2GJ$Oolw4?EH1#91O_+!D-2NOo@F1}sF9{A(Rgeo@cCks<)+vX<k9V>q2
zhQ8SFNOC0G^oZj2iGm+D+txkp=;~!n?{!S>)7EV{A66Lgk9+0WG0mk<`94L>3;k|h
z@pMicS|}O1<j$GB4kk0=h7QlId(pPo=jy?%0|gHYUaOh^ReU7L<o%svSNHE}tmu2%
zROTEIfB93%P^Y@m<xJA^Q?Bju-(F2WU%%;5>$x}g%dQ%BXX=-G>j%6)?lraXdEYei
z;?(-EKQ|TzES)rO{mzI*713u3=O$-d5FBkhJIhA&k7lLShI8Gk2le06>5de?_|ocO
zQs0}Ne?ns<vbd3{@-bh0F}k+l*Yj4Bm;<NFy!#+tmfa+^7#-kspfmf3u0z_mduw^$
ziQ<-PoT?w5*fjGW`H|m89DVFk8r@Kl{`2(N?<&g!ADu0X)5zKIqfyQ;QL=v5hko75
zGbY_hJJMP@8L|$YNISB;cf<$Zs4{11<NGe3SLrLZtWKy4di2`ik(paXm|wS7P;=b9
z_aB&L<KJEWLvx)$f|*;$?hiU*#iB<Wi>{m=ZG7G!y6T~hqu2ChE}t~ACP`x5_guPn
zA!*wfUh2w)3I6Il9)H-RlLuD)cw2{;h(g*((MUU$K(S8sV?0-8xD<lU3z<cTjuFI2
zR07lHkB7GxX%%i0GSd+4Bk<8u2~1mrO+J*LIPn_Gn~vxi>N9jy0@GHKiY=zs$NeZ|
z<{{ca;GnM(1GGK=ZnI$GP$9Dz(T)N~eU-qp&Ev*t>q74i5i-e$4i*H9R07jZ%a_0W
zt2as_WF{cmPvB>$5}0;-xAMfPx8X>0IiiONhUu#Wrrk}h4}R3O^b3}^5fGma{B+R#
z_dh%)aPQTsoTlPELM8?cu@qS9sRX7&WM{t{+BE{{Hy+Vm0xwo2Fg-f2=hcFSzR5yn
zCZY!m25YGVrpIWf7j(p&%@;E9h_)73i&O&Blg4-7e`xwiqQgr<^eE9NJC(rnv}?_a
zPBnzx5i*kyZ6+|&QVC4Yh~%QX_nglbGRqO|E%0Vl0@F!*YZ9$npPdylvk+}1urgK&
zOegzhy!AaZ2+>Os9V!SFsRX7Ms~NXw3a0N6GO>vE75M6_1g0}$n{p4${$L<vCL`KL
zV56@Rn9lrdZ<AB>u@E8i1)^;Qw)!f8>HPI$67Gz3J|bk6AUZ%0V4xD1E*udC)2@F?
z5HjNsZ6UDGQwdBL#nvBq+j(@BkV!zaslZfAB?jnTJz3h)jT3ZuX^2i$Pqk19OqUqG
z__#4A+gyjY0?{ezDHbY$>C*CgH;urPhlI>jL`MiBL@I&ls`L)KS-Q=QLM8>#5`jcZ
zB`{sP@6-7GHMdwHGZ)eR0)GRQ!1RuzkF1`&?yVOxUm{v2km;)grW?3PY3GaAHwc*p
zh#n#sqOTH|KFEKwu*Li0OCd7{(VhZN1C_w^;eabgjZSN%=<sGEI!qL1r4pDv6|1$t
zZS^E=Arp;g7lDh8N?^MA+!;mJSi4<9W<H{W1VJK|!1T2r9QTEkyp0et3lVKEus2Z&
zOy9JKtqHl6uu90JB05A6B2o!V-|C3aS>v$kdm%Fe(e46w9hJcJ-H2aL_+4p6?(h|&
z?F4puDltIcx1ZHwx9z@=i9@ujz*R>jF#VGC-_fTjC=xPL5bY*#(^UyfzkhbO(Dknw
z)<R|(qD2CchDu<1>72Q#9i}{<nul;tbxXo#ZI&ws?{e@Th#lN^@ZL`23h=3b0K!*-
zZzT*TxC*?h=<(IyQ%#Sr0iPNgw}Dd|xRd8^g5ynaBe)$L+X=>@p3(j+#IsyzaA-6>
z)H?3D<qJwYZZj;`3ZhmpBX7G7qU!|fA2(m<SrUGg&vGT8T>?@vuoSdQX}lG*wvq=~
zt_*}_V2;P*6ob@K<`-;Umb(R#TV!>{JDtx8XYU7KxjP`c10MMK!M}Q}y0m-5bv2et
zXN=Ps3rw(F1|!N~ED6qJM41GOw>RIK9@_8%y)j}~#0-3X&FW6a6;Eu|0Y3K~h`$3F
zejX<~HSN!m9Loh*1X$z1gph53Jw_I~>3$nETOOb(&VkW67|;c0Mzm;US|p2BFq#TR
zA5UqioxeD1uwf~~=T-)-98vUdfZ}s&eb)M6gwv>o_Gjsn<s5AtU9J92qAXVg!Xhxn
zZ)Uk_u&D+Yf^UM}O)w?6ADsKahhPO0sbD5xjNgkEpv=r<0rFgs7Xdb=RnyYM=j#Kr
zkulGK?l~|Rc>eCvat-qj7m;P|!=U?Mf(e|~k(HG>oxC{2J7hGD7A*G&#E-xcKV-kJ
z#m1=FSAyPl04xp=!a-c6zvG5Sk~W{q15qAG@DqGa0U`yB^T8ya#sy$fK)jgG6@sXc
z=8M3jh{nZWQcUAE5VwIXneQe@Zqm3NB<%zbiTL8M=l3}>bnc#j;}bdxbb@0i!MLtL
z_E8o!#6I!{kT(q&p5^v{cn{d&`Ek9&*Z!=|VL5pT$g8QxCWcRp#Yd9m>Oovjl*BIs
z?f+E@`Et37wybxoMfUgh^A4TyZ++o!{_2+`!OT~DK9{7Mq>r5*A-g083I!w$Cn#Ar
zd0-?iv5;#k>nm!p?s7luV=NaN9vee+R)A0egRrhFR|@W>;79Owu-*<%1S=Rb1v7BU
z16wDw_FUN<Bj9sg5ZMJ$cp#s93861xG{M~v)(v9_?g4QR?E(E@+z%qce*)uA1P6D%
zu#fEQ2}U+M1crx5Kn#9yqpM+e)e00@Yr&wFK9f7ZU<ZwNg5gdO;~7|P7Z~m$7+0X^
zj4I1dVm&`&KMD3DT!#v@KdV1k&O&S<#q|`+S&A*CcsE$iMr<R+g&fP-ifyI1Zeuw+
zv7Hp><uI>T#nw9vD$yE!{Cq-j1P%K!t1J2W+-fbBOVCQt!+VdAEWwFFBNM_2nx!?1
z8kypl;*JY4mP-jp89~r8c9|g#jx4uKy3B?`uFJeI!ZkLskFvlf_K}a3Ps837x$SiR
zmCWZl>MSQSm)YP#3`JLSM;jakS<c_o-yGL)C<mAZnByG4av{<XYaF@84J%3!xoj>#
z7V-4+48@h%xbwfo+;Hi>sKG}?gl;+6Lp~}Zbhp#E0(2{AFRKLIN*Y&zZWTSg8g#2^
zz6Nw_2*wu)WCP_bg4iH5CUhn)3h;FT4N~4E$RJk(*FksxEwX0(d+5<1<&A<2@^|(h
zicOPf-L^*d{IC!$zZndhNphXPG<8Mv`}z<KWfjFQ2OZ+n+kwBG#udP?pm8PeD`{K>
z{3;q(1HYQaHNdYSI4Slk*WY`-dWkej(@!(OzJc#R=vXK(Kx8%Y0C^<#4VIe_K4B7x
zJYC?{1wlmfm*Do2#@*o7O)$1DGMMrbMNQEO^7jKFIk`B&KJ-Ck8nR3mSa*Rd{eJ70
zH0}oLZW{N1eGhn&@x9>C3tj~GgGE0b2|t0wCxY=s4y6p`EsoN!7UcT|^kcdB@OUa^
z>;~~}k}{U~YVXaO!aJ+R=iY$y4T*Am?jx9fq#d;nO#2945f$3vG}&i7+Ly4AVH5Bj
zC2iKC)Z8rhdQ=g81nEcmeSIM9BY5rcv({h68HA(fZ-eb^QU<55J+|TJ`wxyQUmPFy
zC4O;+_>iW-p(ua!;&&i^M@|Tzdk^CG1mn1eJVSYVC1#P?%3Mi!$OmCQU2bg!x2<#*
zssNh`y8Kcw_6la8sLj|vcHzXzYZuU_9|Gq?bkW}k&W$ua46cXiOT=M_J`B^zCVvmk
z-_zrN0M{Sr@jpQH4`h7iR$plge7z3UAtS>_#^5^x%Z&;j6@!BX4)(|d${RXm0(Y4|
zF+l+c3uqIRfkzpwNhR1+er^I!1v79R%IO|$=67|o812(w{lOxfr*b<qg|{1>AEU2B
zrJ+Ii5N(Q7+wda)A74$D%LL6#&?5%e0Gb<UoCTU$G|mCd92)0>W-g64g62jVe*>D|
z(0CJQZX$To{-$gH(K4_@74IuBe+2{Ak<ICT<=-uNy8tPkCQpmQfhKRQX-@X^OC4zZ
z8?br<)>v_zKTx5k>~jz!%P)caI(7U$zop=mjY=kE|ATr=CU+%)D<(XKDw*;?m`9gP
zCE#8{J8L;umy^K2atg*k!3@+LxUGWRLD^FwQ(B3wq&SeWoVD0mic>Dj*^BL^xHrIZ
z4q^u>zQ?fK5b+QxZfO)<k2^m1W@{#L$QEE*h@n|7kFm;Q1}bc9P<8Fkx;AJSnB{$T
zkpVLWjSIlM0DQ68Sybqo6%rno6(|ZSTRvoQ`Cyh$hAHn?W<@kE2D4&<OWw`+GDFYo
zhH4q&7E+iim*JMj40HwX7NM-5>>-i)8o;=L<jc|zC%xu}mKCG8inPdwVOW1YQpBf-
z#z>1|f^i=Txs|dBMc(gX;4%oedRWfQz-<ujkJLPO9N%p6?hMlX5*S{h3(#M{@E3x=
zO^)J6#$-Q4zML#cmg3qCCx0}hvav;`l>Z3w3zR4C1@T_8#M+m}LmvOf<^zg>wP0CG
z2$tIcmOE&?6D)Vqco$gj0()!{Wn;iDouNwp48|^l#+i&=Cc!v@p)F8$+9-Xe%M<bb
z?983Ktz-EqGgbe?yoBTz<%Pq%l*U^@x|L4LWnfW88XNn#6YAp^)`p?DAkr0yaMua@
zKRU<CejZW6(cRIX)Sbm3ET&^@4cOGsF|{4^+Q|m7+$V7UM7&aYlH(Q33~Z?bL;jjL
z^;L{B@~R|pk{RAWgglZwDHNGB79(82qQkFjED}A384t6>w@#MxHTJc{1qG^>E&Ls^
zSFqfq@JVr`h`I|S@4`5&(ZPE`Y1;-zC#vegfO3#yf<=XKz;+r}Kwt$#6TT9HDj|m8
zDhQ~e$5%sOH9fut0&569G_8GT?Pfhgv@J<mNjiABxLiQKp=^|r<;wd(p37kGI2KmI
zHH18AM~2K|hGj8vSPedx&G=-~IEV4cA^5oN)2rEw7x2(o%Vs3mB)qX)4kO7S_>@=d
zZ1vH#&(J`LOyYu%HA*g{(xPxI#OE2EJZ0aO(pFn;fOS<y$gn(Oa@^2G4=KC3<e?Co
z5KnxMVY!DOen=WEmqo(Bww`I0=&Fx`Z61j?%0`Qig2pJ=7LeqkY^?Yc68?(K>fDT=
zI6Y+dWSeArY<HY0keexc$V4gma(OltLY_dt6BvuF&2pU(&`IOxp#L1K2;U3dy%2_R
zoB0X#Kb$O4qCwjx&SGpZftQW8UD>0irf&oJUjy^s5}p3ED%f(BDrDK!67wiSmfa2-
z?*zM@G~NXcyTBRGrwm!n=?vA|&0w4}Xq?G7XA*qti1FT@J`Zmcc|Q-LxPgu~Q`tSI
z6gQLG5DzGXutMSi`1-D?{aHI5HH#L=Q^`Yd;c-+AQ3%39x`wC%>ngftXal`A7)%z}
z171CJ9jai06-+qh@vRc+t?b)VdK<~bL<I$eDTv<pyr(@(Kit8fGZ7>Ua=?zva>24-
z2i#bErnxVl@js}%ft$(9-0|J^*@A|NenunDp*Hej;bIG%_E~PF(Mo%Ko>*?J(b@s@
zthnKArPI}F)cMZ?ejaFJ<xyA#zJkX2z|W^~0q_fGTnPL^8W#b-h{nahFQ#!DXtsea
zneQfO-lTCmXtooK6FD-b^0yqsnC=Gdrr1MRt^gtmNWgsQv&HD+6_Ze8Coey*VBAc4
z)wnKBv)x14YK`%a8H3|CPFu)jl)ot<Q^?yG`OSg2h+OCiBk!c~J4XJV#%I*z&1z(a
z<*jP+>ok6(CVx!ht9<!yd_4Z+$GX#Qf0p(l4;bk(GK{*uxQq(J*O&gEuGp0aauxrN
zt>!JAlZ<{(i~a&=>8#YN$`2W!l>hUy8h_yNGRfC!+?s6lKK_H=EWK>Vfh^bnE7z=A
zy>>mMLk47`OBQbpd2l89xQ={WPhQ8%p+2u7L%t>-v#1ek$=B)B*V)wP4b<mZ3uh;z
zzX0Gx&Kax5R~Kk#Y6)2#eLdZ9gV2$72HN3<CPsq>oBH|%1P%`h4skVh^0)DEu=u~e
q6-msbmR8obGW#Kp&Tj4=p5DXEy~Hj<M;VL4B65KlJ#oYsi25%>sOR4R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_addsquare_xdl_int8.cpp.C81A82E4BF27D67D.idx b/.cache/clangd/index/gemm_add_addsquare_xdl_int8.cpp.C81A82E4BF27D67D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ce078c9e23e4b82b1f298f88b3555e9d022bb825
GIT binary patch
literal 7896
zcmb_hcT|%}*Ut<gCgcevfzXp65Fv!nL;;niQmlXqii%3HJc<Qf7uQw5iUm+aS5Q|0
zS4HWTwX7h`2I8WKf{IA7VMATr_uJw-$+yEi`<?F}&yjPu<d@%+J9qBfduIp>4hZ;F
z!eWh#4Ni=YTQP^tVzCU+@A4H13rrAL!e_A(_t%BrPq;rszvpwTEa-Hv*MX+>#py>X
zj+MKu?e@JCR?B6Z%!sw0^+-E-i|c{l$Zv0Xr`Em(D{r>@+HSDfuva;06R%o%bMwAM
zpUZ9vT@U=Zv+;Fc*ZpZ;Kk5~GZS9SC_kC-t^Q@>pZ-vg7xzey?yz=$Iyw0^==1CP*
zbx*HMJ3nJY!|kR77mGs!$NM<$dH!zYhrvlxAO4g-toeKX?}M|OFRgoWgq1T)zIWHK
z$Q<KR``hCz_piv>mp3@Q$msC&QZvJu0}9x?b56)QET29~3cGeX?&GhKiD!eZx=qqK
zc=iJ8)iUkKtK03)aC-9Q7nSv0)LQ-aT=PD=mlx|gN6fn{R!TA>zw7jkJKM6UFf_AG
z^k6_+aQ;H~&u<@hxo@dGKGITP@8|r=c8M&%si$?$`hncesKZY^4R(|rGIN7t`1@CF
z&|X@ne4p6peu%f19TqvsB*}Ha>h5Q8AhCLt`Prc@mR8$!^X+3J<c-E@_4n%k7F<iy
z-LN2EZzC_gdFAv`!GgoH#l206%60CsR|;eO6GnM%ZCLED3><lEN8_B%V7VVpxZE*2
zGqH4iOlte@BYA02F|m6ygHrbwj5odQ+Z|v&ZhnzNw*IR&&rFwsbRE9#OI^p{l*2Z+
zN&_zqIN>tEPe-X2w#~?rqMLgzNB?#0j&0wj+R&m?&pwTBub9$&=;yXy=D({hOmDh#
zDr0Yq>5gx%x|Z5_oJq0Pw-BbiYB+VwZEeGo6PA*#VRNoDUpJmTz5M;M$mq<Rv(b$q
z32$wapWO2g2oKNEFI^D7F`|3f<Dv&s!=E&b`qgCIZ;N|v6|Jyq)rSr8PDAgz+lA}W
z4mujUeSdkN`_HyNW=slbx&GGCKd?Ra#_!@kgbVDa1^K)jRBLnigJ)Y{%v3j#GS9!w
zO3-+@uk!q156kaH4*22XT0N&pg%6y=U;ez|<*rW;V@u~<<v#0BJoF7ctSt^XlD8^m
zy>`0iUq{b;-ZLfA@7}q-rxk2v;CW+3onq#v#A$NF)te$~`0p0J$-8@WVZj2`*v4H=
zS*t!8_72OjPkq=IBF~?`(z$9-m&jvr)iSLKM^1)&o{O2jx9jj(>#h&0_m_%&Iv==Q
z>vp+P5xz9qW#fVFw;`%i?<sSa?DL`_3nK(o<|~{#|2&-W?v!bv&&>KYcYNZf{k9})
zn%n8)?;__+oFYBF)$hWpUmp&CwzydQ)FM22x!F9sOwWyFbj{QQqUAFO6nAx$U7wyg
zr!(p3htJQ84!-$CtSIU8c02z=<)bC<n~r`zwrgnXz2>Ztq(RS*P?IN*2p_i6_U((y
z$2&e>I#HW7xc6jX?E>MeTP@j3e_Q&edH4Yz|D6v*<1*56DjwJGT*42`@!2x+ZCG%U
zZ)^UN9>vhacw6mVUS)nqpT6olGJ01fD{7^rarvtF=p$MzmeE{WVFK{J%!jco79Gur
zHrEIgK{P*-_6yb{_0AulSP6(;qrFD15tz0z-=AgLJpLCxoQY^5N2sq6{q!Jqdktrb
zD<8%o+K^+YuMwEG?k=B^G^*X652F!n%dr(}1g0In`R2`&4`a>wa1x@W9H~$vFzxhd
z)rE?WLD_sb0nwHmOMQ*N^zh?1f82b2R{<Y}Ae!P(dK!W0K%JBwJ)yN3d^iQsx*T1x
zMqqk+W?gge=+zYh)^tQG#Y!uU!1U}fi4)%zT;9Zo3lMF=v7j^p(+iIunm_P%Z!I52
zB3i%^7-<Bi7kRAv=uvHm=y`}%a1>&V!1PirqXsS~DxVKS5iQ|J^fdz0sW<IU2ifhe
zqFBoieGIaXLGC|)aFdMW!03nLlxO&GE^1=Mu`<vIOz-py05h8Zj^@Mhh~{(ndK!W0
zJvm361`NMg#)sjEcH%gRH3HLlg@1<>9crWaa0a4f9GQVeU^@TjE~}^SKOE)52t?~~
z^z=0X)5VDq-(KF8T+D~F5iR0~^fdz06=6KuX?27SA5KKHF~?Y-5tzOdwy$|sdSZuw
z6@_R!v7KBaFkMquo&I{9{3kw~g=ia&jaVZveI4v}9j`SZUcj1zXnV1}Tq7{ubXrm9
zyZd7>AI2csfMZ~+5tzO!4Lju97X2L`&PTMIBNuA~rrUag1>f6!cZv_EA=;E<D$oc_
zw+G+2;MrJ%wqz=z^*Q=_8qrU842o%xXLs;n7@|!$CIXGXbVA(R<yK;ZW>xw%y4mY#
zQS@eLwi#OYpHDgsnx{b<g43aSI>F%0xCy>hqn>NC=@Y>51W@6IY`PFQ7BX=Wa4cft
zV&GH^oJso<;8a4emJ7e&U_<nFE`<hmKLNaOLkbP-UdY5nz`clxi@}g$FpRV>0Ygfd
zxD&W_0uRD>0Yw*ZBDfnUx(Vi#ewv!3x3KW5uTbEennUk)xeqD&5fDED7DR&%AnqU-
zhf=8XuMkSn&!Or$9Es1Z*}j^#s8PkQ07W|ub()P&oB5D8-|n-`$o}H3#p<AG)$tj-
z3`Ms>sueooX{EX;o?EyLkI}TZfan&G;eIlmjHMe*gnQU*dI#{|0Y+njO=khGEHHxL
zY-pPe?U>jj8}{!YP6k>!|4ITBZD?btvic_#Q1k<+dH_AJTKIdUy*fV-?Y($jB5WJ0
z^9l1`CJ)~P*hqKXTfhLTfHMf{txg}Lx7=Mm8Xs7j?sHdk21iZ)V$0$L#Mb#&;-Tmf
zyb(rNH;SGgG=DPQdy4LYsxIh-m*+Gw|Aly&L*W<Cjk6n?Tb<g-%7*cV_1lG_=LXFU
zBXd87s>j5%9*L=ovyv5gXvZx)Ek@y-<t6F6zK}DgkH@Anp*RzoU}LiBTqw+ihD@x;
zg-Ry&$%TFd=Rsi}vwag3HbFzYcC=Ygc#DZ|L*Z>EZid2Uf^o)1OID|AvgE<M!9rs0
zn4p+Yvg9_XYJ)@Zl>SSOIZAxaef=O1@XN;6i(bw>IDn#4fhrXY#zs(MpHy;aC^`-3
zrvWkH(|}bP7{tV`Y0R*e4F%ayMED#i$RQZlDagasMG7&>5Z(||Y?LwCl^KoeUkcdh
ztVs;5@tLA%nW;>I^EE|VnOaG3X+Y6(OZh;2L?%Co8q<<9&gZKKPh3YK_0>fbQQx2E
zZ-~{O(NwdzqB!6oa#Rz26H)s=xo}oN!I=|bp7SV*z6hlkp*7xRimrgt3MQ_E(n=<-
zfzlc#u7%QCCa#0hIwrmX4X!{VGX7O)aFvO#L4#{dTn~-wp*d;a0F4`%_y#n-K`=gf
z$O!6Ekr+Xyl)2*VS&*=N<;35Q`f95aDf<Kv5RWJX>_R3k0_-9tE(YvkCN2T&5`q_o
z%~IWZH|q_uZaOeb2j+NST&1FTtu9fCTJAh|6TCMR9TyZAfzKPRT+vF@Wh?38$Mfs=
z#zkR2waNO}RbSni7VU~2H0tq~Z&5J)53u?VaKtO2=wv8Oh7JU$KxqoWxS~dpMO{u4
zwTJVDGlJ>7pm~g7+6h&iB$zHc-8%lGWc`7!zNnRBM>RN19TI}JN9nJej1Nr8nzI|K
zIv$@>e;uxziN9V~G%kM1fu<|SODce}f>|zA0%au=R{>=e6ITOOHA&_aT@5BygD7mW
z^*?x6Hh{HBsLHVmvJ1t9HbvV9*@xo789O}C`9Bwq)E+MXez~$6X{)eQcw;x&HsDfj
zYIQ;~`dUw-XN|80+s0^Zc$M+l10C!&4r^So3KYH8VJ!nCIV3SqGH^%+44eWSQkZxh
za9GF0C6HSJ^{~biT?)CSOk4)JWlVe$a!)exDabv=#HS(mG{M_bJabMg`$(gerGv0^
z5P=twR{6Rla%A%%M6^Ib3pB;fmiDZC*(AT-j4yUx9I??cRo~?Bcogl!^XU(v6x{<=
zJ<t=c7hmy^$E)vmq>~%Zjgd>EgQ6L^v=ypaNiN;}Y4*%@dKPz(N3;U7Rv^bqr0BiS
zYcCAI7>5cJ6x5eRGS@SxdIr6*N<XIb$YXSC%F%en`o<!B)1zopeNz#>Jso<f2&pmt
zRE-+70AUL-#gkC<T_C(maPFe9>@lI6I<?eQj7=u&*N#o+L1`Ye!X4T49Vom*stg=&
zkvY|OSE8*q&zoV+Zm8-e=FEFzWZ(V5`hU=|D+97JV2ekm=yD({XX5oxzMe!Lirxz4
zTbXzply4&#9|>eT^;MZ@Gn6;9KQa}}O|bO7dhZc3$`K$s0u*HSqd;_&V0^hoYg6B>
z$!LbohF+v9cnDPwNn|Ure==oyU#JpUG8am6VgJQIjoowMP&}I&PsoKK7?-kti*Inh
z`UcfRmPVE`T(MHL#7H8;A(x`1Mp7BBq)|m8l_~HRpo&ClrNI7;kkV=>K`(&+3osr}
zRyO0FYu<J}Av!*p&>#~UVS++h336fom4Kp~pr8pFV7EJUe)IYHvyyC)@$!Hu4-6q4
z^MNRz;Ijd{6N`PH-lN#4+_1R_?eGX}RBqT*FmWZcsbu0BXj8+)wa})PiR++E9TQ)H
z;w#XMjDHo1uQKs9D89zT_0Xao%1HYLXwksLH=xB0g3swbzqx5C&HCayN5$Z<Snbhy
zy_%pFm<}dlQ{l6U{7U^KL;T8_=d6#9KSj4gRXcRYYT^`vTwDE6LpmvZ6ccfVsA~Di
zjn!DO;H#%q{1k*{SHA<1+2z)9H(d7OyoMU7UxP>^MW|v9R<Y4~b?Q3rN$NM3OsLF+
z{ZAptq2QoAa+WC+Y0U0G2{u4eP)7Z>@_O^HzBc?&g@U{Km5QkC$@4TL+YuWSI}I0Z
z6#W#co<bk2+wJ_h_se4A^3kIBW_$@w7ZgpIQ4)L+qiB6IeF=WKYLQ&fZgr3~e6d?$
zYa{PbznT$seR;n9MI%K&fvP7YTi}F%I;r2%NGDgGE5qeFpsIto-2II`cPqNI%QeZD
z+X5-PT)YM}n)>ODjOM}fV5GzQP<5XqKzt8DozxF<q>~HJg`vVDsCq<Hc;r5#Gv#Ct
zMB&E9*2WIIAw}ET+S=jgipNnaIz0S$^1r%gJkf4%+)Gh+QJxZ*38c8fr)XU>T`9gw
zQnZ1YffPFtMGMV@Qv7y9(IPXE6z7p%x_FAG$iEJ4`4*tJ1u&i)cL3uZfKhN~0pl!!
zdv_IoJm7HiAP?0FfV~&!;7CHz*?^tR#5sVS!^Haldmj_;2kiY!+zGgyK$ncy1-M;I
z+zq(hOuP|tH^Tn*edS4$xc6Mlk%vTkMF-;J@q2IkPSumy15qlQz?+~$&d@~OL<S1s
zg)q<*-V_E3=Y=y+1TTVtqIgmL=#M{+R60G_^8q=Cougd<j*_U*cC-(`0qL(xjq;q~
ziTl6U@T`QmMd)!G{r@*1W_h6c!y>i#-=BE=fyG)(J}&A1s73Ey@veaHT7Ou3;o>Fn
zi&zW$DOOB;>>~8v3f7poaawF`&HyftPYLw(bb|~=`O6J-f`rB*LlaXE&tbzyc>DOO
zj2yhIhuBH~zi-9n7M3!rfi|{->=jBUXBRj3p_0L7j;<q(!~wxe;poYJfnY4_zkSD_
Ay8r+H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_fastgelu.hpp.7F025B27BB774C23.idx b/.cache/clangd/index/gemm_add_fastgelu.hpp.7F025B27BB774C23.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7b588fbed10591395d9871270ac663af6c2d74ae
GIT binary patch
literal 4170
zcmY*c30M<Z67B*)0u4lBl&c{s${FQ&BtsB`;1cALBOarmqOz+f9)K*~2n<)mV+3JP
zluS?obwNiEJZ97xb<|bARcD-Ibq_pNS5X-6b2~DfSDWu6>A$P~diAPa)hm)GMny%4
z876Yx#Qdy`-1Gqq!wlm8a&og5Ofh4aP&<an-*KrrtM#h?;BzmMo<~$F!KUfP_L9v7
zR$b*$bMyaE)pN3L`NFKw&5qw3yP0?H+NiGedXJ35_YQw}zj9^5gA;3e_vtEa{b#~g
z$uE9eTld7>Tb!|c9qW5}w)L{NBc<DXq5HkYd-~e<uSVaPw@o$oR@SSnJ@0x_R`&HO
zGym#$Jut0AQ<N@_YdV-BzCYXl$*gx@|8{xr1@Fb>J_9?urd6l6Bv)n^o-6%w>d~kV
z_5W2E+HHT#NIWvR;GS99qg$7mql<s8x5+(Iv2gJHyK}CG*WX;RW=Q&?VFAgK=J1q@
z8HLL>Z=Id0A#dZ9f$cr23kiF+m_I#Ne_ApB)vlPwOXKAAS=zky$9mq^U9CIt%ccCg
z0mln__x+N$GoiQEF;-$Jsk2zs9e--usP4bXWM_t*zgoX>UiW9$QdRMNLE_>;l0CZ5
zn7~<`FZ)XWt^V-cpmGO?dzHIJd{O@J!<U^aPI+V|xAl$)J#|rZ^S}<vdHHwhFF(m1
zefG-EeASiAeLo%Du-La{0Xro3*tMR+Zz~S`+NY~%%3oPjcYn#kMUOKVJ^tXcYL@34
zebc2(S$H^GK62sWRI#0PpO1}Q%iq(k6xy7$(kGrRh=?@Ld};07Wf?y|;>_8lMYS8M
zNp6pWhgp3=_}KK}w#>or+m-rcTPx0B*d3F!#cRHLP50gbvQr~nY{uQtE_n_y*8;v*
zJd54Xa_)zeoqIp*@X{siZnF1kJQQ6z|NOLHW4zg)@{W`qt{SpC>T%wxpIy32c5LeM
zy#8qQ`UrN}^G|k+Nj!e%R!rZ?H|7U?-(AjGku}E?7)D#xasAIK(Ib`_<j5a7{==nM
zq)cT*)&rPfLxF?EueaY^+p(8*FcUj5Tp@@H!S(ka#;TvO;q!)2$?bv#EF%$Ygjj@3
zW<&#_--iC%3%h0v{O4VkaS#;zEc|!{E3<xuBf0|V%!r{~HfDDGB6-$%4j;dS1<qeD
zi@Se!-T+H8dq-Pt!N3J0KYj_~ABzi%yaJx-SVk(Cceilo&5O+Y&EFpTrx(xD%%8E$
z5J4fzB1+4MSj&Ee$Cvubz-*2$%M2D294s7o3pN)03QTs!A9Iu}!&tg^Zj96`Wy4uc
zLL^EPksC|~BQ%5zK}beqN|Te^2r>ep(PXsJ6d%r?_#+fX!jz`qavGvhns&lPlW2ru
zNDM++qE(s_%S|Se5t>S-B2)o>6);A~G0RnerUGIp)`46H{uoE=AeQ19@T<Z68qm~W
zel5srF~1g~Ycbydeg@1pfX0CNjUaEt{6>gw#QbImXa++4X@S5N2*x<R1rjN4g}7Fj
zL~$DgwqgA?h;PIC9T3=o<sA^;LFLx38*UZF2B(Nw&V{(B>0xjst_XP%FEvfSH}OU&
zm;|e7`V~ZhPy~rk)AUD@NQ5SlNosl+5=kOLsU%fR)1OAt5Xx3$kMFNqhGDrJMGiu1
zz1I?<5?OAY_c{|A)LQPm{4Wo_wPU%VWGFh>j>HimXX1>I2k}7Xd?Xo(kSFm(D1Zc@
zQy)kI5h8>jG>(iz6`>|-gvOKc2*r|ER1paz!Gt{Q=7&kYK7QMh<)lO!Mzc1Y3=gAu
z8bii}(aek`W5Z}JCX$I^H2PD>lrSOQF<*SM>2lFS+W{=MNVP~!L%meB6rp^1ejp9?
zYWZpt(k3`~{%6Zri-CND5lQT+cQTR8o_Z%2$?d6kDv`>bdZ!U-?5THJk=9=Dj^&EM
zwiqNr0SNESL?63;<>@yK`~MXxW4SsI*TEb?ZC3t{C9_W)?>1+-D)6a-071ZV)!<tV
zff&bCLyCYi;>Vv@e{jznGnOlbm{M9fMnkbxkR&K(hUTwn_-?`vz;e^c^l^e;f*7oS
zzjB0|U`cGp)V4RN>v_W^z?OhB4fIy<+zP=0UKjm(=I8!nCR_30lu3e6IAyXR6i%5u
z2!&Io3PRzOX@XEVW!fOYJ(kNHo@rW&<?^I?UP4xjW;;ji3kb^NJ)B0SeL^WS$P5!I
zHmqA2-*o1jhcQ0VQg9PCY{bE(FplCfkd|S287Rs?P30BfQ~|>=CKV7yu@0PcSgwO$
z9kyQu(kg7f3KUh?el<v|vAi0T)euhY*ML(EwqFCJ2CrWWPPJHG3&FK`{RWUWVEYXa
z+<@hJkm|8q4+=f@&j3yaY~KLHfY)yXr$#Jqgy2SOzZu+`!ISQ%8A6&Ng5nl%Yr*ms
zP_;lLmA8UhE0(uHNGs;Ifm<7vw}Gk+^V`9#9n0Gxq#g4+z^wzzJ3!Sz`K3?(;2Lse
zgEQYXO2MZTUzug#TZSvP3}VY@;jvr;_%vYs28e0Eg{lW%J?85nR!{jGN;bJ%f1>c=
zk9i!4Q_wp3lzgfX4!uxIe48;2hP2Ieg18fA3qjm|Pr)THJ?Hr*uoc*?w5Wy3<oU+A
zk@63L_z>mqF6J&AdRc$Mg5{iu6Kd?E$SBm<gGdl+?BOIFHTHNCZ))s9C7W4)97cuK
zW9*?G#o#A+eY(5yOt<f*V&3Z#@GPONM<DRc%Q&(MUJFf!-;8le6<)L4eh}}cn-L<(
zs~Klis#Xc&N|+_Ae#hV9=|KCTI^%UAmAIq+CY8AhOR=0>Dt8y6$Z{&F%3a6_%W0$<
zcVRV_(@M4OLf(v9bt;At!eS57Wk==YQ4rruhmm3E0+$mxx};S^g)ZVy5{ghW#5Lmq
ztQ7)U@c`Ej0qy^f<JzgdusuE~#`&BEWe<q=(AYiodA;#rV`mNTCr3CmiTMy8G<zwD
z5=~nZ$OJTBC6OdFQOzVX(F~QV$VF3A0eBaHX;OOKQFCf^=*6%2(_l=xa1HVQKP*Z&
yf}HHRIVr{m7vop|hkN$kvG=4?^HobXnY**A3*#<G7{+w5F^nsHE2EEY^znaILMOoh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_multiply.hpp.A076E88A74402B66.idx b/.cache/clangd/index/gemm_add_multiply.hpp.A076E88A74402B66.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5f659ccf2b6e20b52311a6521427c02db24e7e45
GIT binary patch
literal 3096
zcmY+G3s6&68i4POBoMrjfF^KwBpITBkU&BX!~g*j0x0F-gGw!EWknGcF$oD;lZZ{I
zE8$@u+Ern7YiZG8bhLsmy1Fh!SF3jGjH9mO*byBzB1_e6)Vg#pFgNFTXL9F!^Pm5}
z|LgoGix=eO@hO7HsamkPc6EKFA3+eo^v_&xs>x#!#5^8BZ0@+Ww6^Z1f`7f_AO6{r
zuG4cB7mmdDx{BsY_@0KV=A*5G*wnt8i;Lbgbe#IHys*3cNP)H9F>3P;jgAa#eKnFe
zFgO@?tZ~`lOS>zE-jtDsW!kot--CPEd+Z6WM+Mn%e0J`Gvi1X}qAQxDjxXQmj9WJM
z+IN{!?*{&+Yuy|BzVPbMN@eTAn1MCPdtWAv_dhGHE>oT{SOq)X*Ou?^8^5!B&5h22
z9`EO!*&`X9*7)P9strFZZjle=t&0|{U3caOWm`f)Si_?Ij;@Ghw=W#%)&KiYw4&`}
zxv;4ha^KvLr?z%?J#RG`s>_WHUSaFSrHR^+a`AIE35NFcoH6!z@1OaCw{_<WkIY`?
z)qnfsosF_<pyK9<(TM+WHcSWYgUZ7<3-=_R|Ei+8wd7g4&=vgIEO!%eViP;FqbBdC
zj;f~*N`}=XTgUcy{cp_TIQJ^K|Cwmx=q=&H-&1X;FLcJ<IolT%ccg|Re6ey~7-u|K
z$Z^>H+TfuXPt>2y7_0UF_2{&7udBPgQMbfr_4fYDtA?+fe!x=b<30fg2Y&oh_+Rw}
zM|CszEjg~dCy7zb;eBV}?f&SxPV#5fqJhTuJcCaRhhsO@&ySM7yWR7#{@m01C)0Ml
zq_#(Nc8@;He3-IYf9}!UeHk$)9-P|ext8b@rd64?fs@><Tf1`YrW@D03xE2j<<up8
zyu1FRJuSSR&o4Gcxujopo$OY881eMg+Q%GI+GB^TVkm#}Wl6&MCEuiX|Guez_t0@~
zD0@Mlzj;gTAAY90irn}z-8+1bM-luG8tDf_v!g2ruKzS5DwqT$R}{arc;VPSiU?vh
z2-pI8LxA7Zh9uczkNKBhHcV#)&Llu(U}ebUiwKx9w{V%}j^|g)PKubp=x4AqXni(0
zrN7;LA|~nf-76f{^bjs6^)H<>`C|0{IO}}l@bUh`6v1Qklh{eLK8H1>|7H7+`H_;d
z-Vhc=f4`Pr8#(!6^a<1IHRc@J875K|cQV4(i@zQ0`8i!efku+wNLDjuIO5Z7Ie$@0
z!YCkwLbWe^5DG&v0uW#%g;F(Rp8_#31|ua@s(nEK6;xrA4YSq01Obc8lEk<{0ZYE6
z0FhO0RbaHhT8OAY)`02Rls1TD_GEhsM#c6;h#X1>w$n)#I>{0wb&<&~64P~&#V!)l
zZX%UUB&P2slim1!H(Bh)_qUMB7POzqD_xC=c0$$I!r$VnMd0g$_nq&gED4~12#T^8
z9SVq{7^4K3kcF}@gEEZf!FgFI`x>ahC>Q2tq3q|uJRb!g+Zk>->iIiGfp8d(6*>Y&
zU=#(TFp7n-SlQ!X97gdl9wP`L)&waq1tT?7W2A#RtU>fpkC6cye8e0=R^a5}W6l%>
z>Sou;n3@*sJ>P!aIl}X!Kv_T;eT_^#6etfU$7q#&RWf6b0@d<r9~G%TtrN;ycF_iw
z$1PVO11sWIU{q67^Byv=u4tW)m~7KkpBMtD>TM*y4V}T#;*i+kPcqnS`V4Y~L8$Ih
zt~3Z$TESHWq1tM>+91Y09XGFTCKLBszT&%fzkw0jZ3&cMjU$J0tX0xr8rB^7FyGe>
z+vdMs@u?!|T@LM!M5x02kw{gTKN5us^GBjpA%98@rDZ5zm4eDRCPPfE$iT_D$K=Fy
zXxoQH-QQmV6fg>mA|@7^QjD31Oac=oGo_lckjyMFV=@b{L?f96TC|9)z>3MNDOT(Y
z4WI#&*+jNjq-WFE^oZ;tJ0`QI*>jN0A#z|cCn<80Z?VQnVwPN_$c4*YWSWb_tTmCM
zCS2Z3ikjbMO*3xKO^V#O+)bvraeFPKs0GP)jO9jZZnT8c?PrWPVq@1>U|fhoG9{Rl
z7!{fpBQnREeR@nw&`m#ab1?;`zz&kX12sLPO!vXWCC0SH*Z&&S1Rk&%=;jTBVLH@%
zbKo2u>b+!`tYbbzfh?G%V|wrMo#$0LR^T++KqQRB49tadF$1YE6*G_nb9@FKDcCRk
zgF@XCr%@mjBM+nksTg^n5GcgR1Fb+SW;~!zBdC)y@p&pGXXWsziXQ5-;cRRwDxd<J
zd|IgWO*aZ;!c2^8g*NP>a*|3Xx>zXCOe&k-=E7!lv3W~x8}fTQRV)g85cdH#wbmD{
z_syx-eP8@7G5zYdBsKM4ggLpU)f*}&)@IWx^ZJQ}VB$6PL@*125c-xZ&}$IeFCDut
c3kZ*lm?ad1h7q%TQX&*B!_YDkEyL0B|8!n%n*aa+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_multiply_xdl_fp16.cpp.2190407D709E315C.idx b/.cache/clangd/index/gemm_add_multiply_xdl_fp16.cpp.2190407D709E315C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7129663e581ceec4354ad1950aac8867473d8427
GIT binary patch
literal 3128
zcmZux3p7+~8{V@=+nBwXiy6}-WxHgU49O)b%5~i35=EtLbi-dRC5KZ{=Tt(+gcCVh
zu_Q_4l0&6a()pbwoWJW*e>YOPQBqF)^?$QhAI@55)|zMU_j}&=`}X&K&+`R(y1QEv
z2;!pf+!+_0v|bHC5FYp^CM9fcVj{?14FuU)+!%5{=e{lX_E)v2-Zgi9*<)AkUVE`N
z>kT$8>FJjn_~e>-{`H@lp&n)RndDskLj!J4i=T8BH+yO<LnG^b!e!};S7eR69<J=x
zKHB}oOZ4f;;2VQqWa?$o$9;H8b8Aa&#G{EFGqQdR{molcHD1Ab&@H-Ib9``kbjYn`
zS6bM~BcofE=<EAlK67XA^N#loX*(}^n<VeMSUzTT%qCMPjXii~I3l5D$C<j+BvZSH
zxtHGi+^efA*=l_<JPEV8f6L4JWL8h(wd8x39|cxr_r1NBvh!EZ$AvMz{w-@tHIkeS
zL$IQ)0pk*fjcd||dh}dQ_{{miW!tTb&pTU`9){7u$YF=F`g66UXCacr$qv0^T{|}X
z<fg}gYkuybtYAsAhF61GjHKW0dVy=+<4<M>P@@eO?^Np^%y+rMtTbld(BmkchnH7Z
zvI6YGd$+~!SARM)_@Ls+-YU1+uHzwjdeK3rW5j8a`#;4Ov`LlCmFc=MIpx($t#>WX
zUWuLP&-wgk`q$F>+CIPdO0~^14nBCcJ=VkfRPe)~CuGnm<7=y>XNBUmgQo3~ZRf7)
z*yV}yjrbW`Bg5OR_WY4uo?#o-UjDl4q09ArN2b=&7QK`}rQEdiNmpjTVwV&zN##;k
z^BqZ+n;Kecb3Sv*NHsh@#9z|>eMnPN(duK>=Te!eZ+$w?p7{PwN(mYJa=wdt<8%8G
zVWUZDd~tKGZ+rLcw0OhFRhx^BVh;=-U6oynJLtA$B-6&QqtD4^)^bC|?GEo_sAl!z
zLI3_iZt)*_HJL9;ZSy(@|2yhkqSzlHSio;S5I*+6(rXXq6kenAA<K8m<pRarCc?Hi
z)4a0l;`+Asf&t~G%^{~P-<m{fpK@OQr9(X0c_*hKuz5k%r9-crW?A)W9jU)kdw1y1
z-}4;oa`eUx(_P-TT8QT?THtIuyWGAY+G}rK!7bO2!9M>B6D12jvOPJ06{~D!j=F{z
zXJPy7rCwu~l5f7M&$c=fkk9Q<bTM~l-YA}LeSO8wQ_2o=r(OpmZBq7fbMi}>-~B`N
zf0Z&fCZ`?pUd`J*Fj2J4<M4OG%#TZtb&mi1Vq}+YqT`F&E8@0a_+tUpS)AhQR)ynl
zm0K;$Wj&>CjzJSC7uFq?$7p*vWQd%^<s&9hUt_wT8mzu4J>_)y>i;<IPeb=5HKvr^
zuJ%0qrFtm4_e(+Hy}Gfx4YEd|?QW}rP@5watqa}adlTErp6#hzVr-GNe4Se<KgMt5
zcwR(dztgkdC?%4Z5=X_N2*NLGa2);DR!vQ09fD9|9WhA*a5()FZZPI2FdVPrXe6MG
zFe6<WzOjg*f-CKI?c9!|p@8OMTrC<PE#1C-XyAjlE{+BRT7-%CG(g(oQ_6+bkMcqs
z^#imJ6LM*Q^pexvCHq^BRO09gK&xZw92y`Uoe^O9A+uu$M<W2u$M{+_KzdU|>fJI?
zc0Y~=0Gfrdcr-wI8&lZH!q%4IXds|<FdZ%pkj@BfIdD$#nTMk*0j+^)aA|<_-k#Gf
zQvY9<;b<73*%+Hk1Eddhn>NZN`Bw;J1E6JUG8PSx&Ry=#h`9BI!ckv9V;IJv0n!CU
z6&C7CE}g^C)qtLj&DNm-(q&D3<nxga6*w9WXdJ`2G(fsFP-BZlYA72={Q<3oX=%~`
z>E36anrV{n8gO(iptUh=O&TEm(DTLx`>wY0IJySV1V(UZn4}+@M0L(9e2k+(faYKv
zO&TCQL^za9urMW#t^za<<7v?V>4fMFiPkfwG>5YPrZs{hs~F^Y2HZc?9)talngi7;
z!NJK1im%Hzgrb1T14$~`O(l^<4D#gU5>SC4O(nmnG>Kdi2-k-?1W76pP9+)3jaR@V
zx9s=tww#>^(;z5JfQjG{6O_6@T?A(!C=G#z2x^O<*aEf)N{OHd0U?6@2fgqkSHMny
zD1;1~7HuM9rbVgbE*N<gUOwEl_q8TUQ20z-4{8N8N64QAo8xn=-`sKZ=PCx6mqkio
zt<P~gm$|<048X8RDAI>5gjHbsDw9oZpL8IdVJ&nUll?cB5!5{Sydc;kf-;etOh(Xw
zU|yB|rsg%3n@;wAhqr5|G|V;te3RH&91L$O*g{aN^5#@6J_nb;O~cNDB$Y#_C7I#Y
zFbP@}h)m^Em;}!c)T;amhY89^ZZvu3%D{@6k1v{{K(ko_mI(F$bagGc9(D#k!2&8P
z!<kCbpUsQXYO7-rlon4*09{K^93DsT@SDvDip%2);3Z2?JRVN~Jw{O4JZ%BAGg#A{
zw1jJc$c9`rE%G3}U<5Y~cBZ;J;P-%%I5HU~)vArgb=v%UXa+$Eg+hI36N1ta>gdCR
z10{PH<e^D3fjFB%9-S7O$;~E@8}1PFL3Js?b-UHxFhp;+q2PdJQkfhc2izCPR^3rB
z8`=X5tL~=B;cxu{!>T(84iDU@yi}g%>#q{eO7S`<9@s(90o4@+vw=s9`Q}jV@9K63
zvF<I<1+_*FMjlYYW3L|O$_640)WK%=q2v!J>==miQL<oqT#k|_r^mx6`EGig&LlIK
z)9Q1Xq;h)vl}YYVgM31C?0VIgCMHBBTBu&=NcDm+sg19Xo_Dy}X31Qyh(Z+fAPrH|
zUgH$88$sw5?ZZ@?fvK}Ja6*&I(US9Ao#*k`a=u7gFjKt9e({o}j!w&DLaBqHtz_;$
sjq2#m*3%~qjEqf8&E{KJE?8(|H^+LGxs{7h$K5j;bqiVU!SF`@4?NWhbpQYW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_relu.hpp.2FD450825E8EDD32.idx b/.cache/clangd/index/gemm_add_relu.hpp.2FD450825E8EDD32.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2561e64e2b88603c771367fe3adaf53c392a798c
GIT binary patch
literal 2290
zcmYjR2~bm46#YcP`i1-~KmthE_f3AnzDNLL0vJh1#tjk@P^;0AO~kQG3xz6(VCzBw
zH>PS8SzJ&Qt<ZvEi@2boEesS&6pDZgD%NR#U|#<JGylyy@4j>Iedpfyl2*jUjT=D_
zPqiXXot~}Ghad>S{;cfGO;i#D1!A_mnx16!=0E%>vnyvx)b$ndD_U}5i=qTilKT7I
zKKjsKQcL-VT)MkE`U@JjbneU%URNpEqmOFezG%IEJNNbM<3^v{qslgD?DC{6x3biZ
zOc_xZ=$p0=S$sdYhVNTewCcYU-?f{<E4L#x6Yt(7yuWa8aDD7Kv)HK}<9`NSymlnr
zu+(1jo?Y55uekYfcV$i0qvD5Vb(Yk*V<y?s5Z5PT4lgeTcP+_0?ft;O>s_?Q>djtq
zSHhO!)E5HF2-SLP=eV4m1$}#NF7y1nZh490_{+TXyCLeHfq=_b!Y&WYl+h2|nW}zf
zH37fM)NHypRD38l#rew6=EFiajgt#M!#*IXL4C_Hvel{VYW<N!=ZbE$PqMCa1LvpU
zDoso>uZ_NSpEP30?X<kguz|{a->%-l57}xKL&GzCUcBb)twQyV=e5%O(TSFa{q1%9
zZTFi#?=GC;6;6%RH62c9I(G6<xv3tNK6=#a8`Ic++Wpkvrd8=1dnEi@MknGuqpIp^
zLYkxEeA}r%h+8u|>`xthP@ed8(3m`{f<GE>VMD1dneho2gQH3=J@}pOwpV}huJnBM
z$`9_ALjUHzhFs&m2A+fUv!`geql(3u>t!C~Jl|FI-fH7WWzYSx=9Z^xLZ&CTny)?|
z?ol%^`1aZkTLTk+i-<SFPiPeSnb*8o4K!2mo!6zI?nS%klfzj%)oI7D4%yW;nigQ&
z^pHw2S_~nHlti2P0m2$a`+eKmyCH}xdTc;4gvkiUkmEc*0Q25@{l@%BQ3wo~zzYO~
zN~Ve+V|^0DFx5l!CIwMvNlOz?o01GL_t7TN)cFDM|Gg7f`{l&R?J#5v_;zGFF=V_*
zZ-EbGre|c;lOf2!?rJ&jL?GJ?MrbH4P!|}XqjUnYKrIM}jj{uEArO1io`4)tM*@<1
zNqvD7Fd|(p72rtbE%PHFzD$TCrI(VxQ!ZB$cq(rdfv4iDbUc%}<Aa4lj@SrBtWoPo
z{4h4C4FS2KZjpHU?x;HfMQ|gQEhr8_FcQg)Bp`vi00qLsh|pcALrCKuR!+nCFC-YT
zL@fzQu~3!{Ihkz?XWVQX#WpceMmV0G1L_bC@)1=1y#HR|Ka2HYBq$^(5>F*8B#eM!
zyknN)sqnq|IwVP?y1u9kBQL@VYD_c7qp*$H<{%1;h!%@zK<?DMnr-qMZ6g>(?jz@b
zWf)QTD!2qBQLF_hJz?36ubNM#VP_Ew#3HbaIan!A2Iz~>yxqrtjd`pGBQB^*1PH!O
z*dlBCwC*d6n*=AZ@FS+esX7FF=$Xvl8*pC=_;V=-{Ctx!d1!1}0O53Ob^r(fBU}eA
z2lpGzh;{>F70Ow0=llb5vEWXqQxHyZMO}5MsIDMC;Y{ak3U)e%cnQb_LmY7gjwB2T
z!6xBKB5|9PDRm$qR2q&W8B<2E$v85CPdP(Q@F@?G^KhPmp&-~4p^6yXrerD!K9#7F
z5KqNa5o{`sN*AwWI@Tt*w~U2VUhE+z+$xEu#2<Gq5lU781S)|QM*D2->-coMemIm3
zDnT-yf2qJSjI5@t)^$x;@qX*RX|4w*z7=Z41soW0MxD8!6xAB8Z6frl8|#`>*Hm2@
ze=M@Sd*MUv4Q!c-GIh77V`N|Fi0A$yY&pp_Nw<5rWA-n1$UFn<6I;|a6p!SAdW3>V
zFcQU$3dj2--aXzA^vU0T<hO?AmZjQj#Gu*X4%iH~9qvHH6gdKA0F`7i{}uzGu)F3Y
z@&cRYS`N4Sg#LCB%U3HO2o5}A@@XGxT)qrSMhf6&rf*Sb30axSEU|WiG;0PE9H;B8
f4XAXQ8O0nj*RC#nG%$f6Gki+Hr*wQu#i#!RrL90{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_relu_add_layernorm.hpp.2503A6C1813DEBB3.idx b/.cache/clangd/index/gemm_add_relu_add_layernorm.hpp.2503A6C1813DEBB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..76a4369d51528ed287c1e36374bd86410de58596
GIT binary patch
literal 3418
zcmY+G30M<n8pr1ggmWk%kW4rlAfN&Ql5iQw$9*S&8pPVJP*7SqZ3FHKY7s&c5dw-y
ztV<}+bz4#Fu^ttxTT4Z^%L*Q~+6T0Cp;f71pVfLFJAr5Nea`bR%y0hxdGmhnyz|a1
z%S=mK=0p(5Yce-fmedqk69nNx|5ewhO7mC*F+gh@OjnIB_FoL+oP6@hJzL)UBmK&5
zkF=tL#~)ay*4}=w{qnUR&h2T=JSIK$;_9JQKmGGkZ&l3vJ=4F04s22lIR5cr++%+6
zYGvz<?XUG|5C4gOTW+7S61EnNaE}7tIODiv$oF}|t{%s}q2IeE8w$UknQY7(-BaT?
zRrPbBcExX0--m|v!<>iRed3*pVU=;ZCatv2_IOQqNU~e%^PP7_kGr0{+(X$Fh8L7g
zj`egqJ{R46wst0*{Zueh{nn%N<8QY-2>VWaWQm}!Ij^U$<R2#&A7~cGWyGmI3a<RQ
z{f%B#{#~bu_x87Zo3@Sh;>aakaoMdkedBSb%ipcO|GKC4z-`(165pGr*L?1IVytq-
zldo8>upH$4{~SG0nR@Fp#TD||x=>DPf$Njpsl?aL))flmE!Llw8Y<*>Y%2CkgxSgq
zUc%i2O26fWv6|N({#Dh%*=e{KXgfHj_E+x8AMmM7Oc}e9DdRuNgxj8e-#W57dUfNC
zYu7KVJyZO2*vi>^@_PB;vEws=`}zpSW0&vzVO5_?*}SJr*l0{VbCX)C{a=Z8X1R7|
z>J#~i8NsuU%lp1u6?=0$|6yU?p<#XC#1@yN-ELbX$;z@TUV#Hd_S(cp7sAxsKz+o$
zYbE6~Io^wlQwN9LmUMh|U#6V0Zu|53^9sY>?^YgKDK9p(_^-M=BDOPKE#J4RGvGiO
zb*oqSM$cx)-BCk#N1OMpo&G8Es!PYeMP0osJMQ)C$G1$sZ?)&|hxSI)otT;U=1Syb
z1uJF!;THqyUvp1&%8hmLM9JB)q3YVo;_Gye`Ty85A#G8V+prwD_Edpw!MxcWVMlB9
z7M%{P_$!uwL=sNixigfQEzT$GZCHed3!p$|-hV?ptk&sB;uXe_$WEjUY%pi2Q}tdv
z$h!Y4NjNhG7q$y+u(O^ss7KTK#Tw^JB*9?}zHDFGU~4sJIK`Fn<$~??jw}aPf~s<;
z;?C}v6mNW~e|r<hSIK3Ow7i+K*=u&k$b_oot!k4!LD+a7ZA<Qr_7RX27jhFUc~P#=
z6(b(xVI+d01WSU{JU9=da2TFosSgziBQZ*asR@?SP#O@d0gISKNt)9&APcb;glIt+
z&Y4=EK)e+MZ^f_I0kIB5AiWJp+CU`EIc;Dm;&vcz$NhSs(1Tp08-T=sUvB_82K;&>
z5F7F9O+aG8{U(rO!u?%9+=cuOk9)4E<&jE~q`aV41}glk@Kubwp*KdsFgODhJ_Lqf
zv<NQ3NCssY=r%wIF^Y%r7%8A41KpHtn2k{`%(alqXU!h#Kb`v4jimUHk97kDDU95q
zJ4PPR10zr9iP3yGAM2Y1Z~;bvFc7047=+P6xDcZ#7=?9MG>pbb3Z)puz!;2TVJt>z
zFb(U>beN7&Cd|Yr3ua*?hjI%sLjyR>4h!Ajlb-(Tu|CXQVA;<O3Fb6OeE>KgfFh>f
zm|H;G%%cL?x&fyf*|K(j-9Dne?PN_-MRrB>D{yWjsS>*qjNXZQCzhEIBvlhtW1;Nm
zj&=Ny*6lVVRT@wljslbglwnj6P=V2g{0&7YTy1_WMjP`tS|~4vyWr*y8JkT{6ML>R
zYHE?a$Qd=R#9rdel!c^{?30|C)ss}cQ@txw6XtrlCg$mgYO)P*wxOCZfoYq0Mk3n|
zz}bOpO@FJe%icS5#*?HP0IvZEnV&IpWCMsp+z5D$xZVh&8bLhL)qtnQbv2NxK@!rN
z0IvzxHNag1L`c_w2n~RUwZL7A>sk=0#q-$;+_&Po4hVH%0lHoXq&ocf+kmhQ*V{l$
z8%Rd}b|7rWbv+R3UuLNuk7ocv1FjoDi~)~l1VSUOcLHJO%Pj50<C%cagzF{{W5VNg
z0bv)?wZEr&Nk>~fY|S?+r~!W`G=QuIummlv5iD#3Vw@F?U>RaH2vwsQOj2sFR1H=j
zy$J+0q4*z<ot8RuZ%UybHaF;&#k3<y`9uFKrt3&50!Cz^c20pQSxh^B9={YG{_Tq|
zSR_>*P%c3aTU~zLo6G|>IH4PwSlHeKNXpUHk;j-AndLD1VrQlc{e<|}2QwL;e&iDT
z1Hb$nouLp4(@}<@Ff<)y7zg9hQHD7%C!NXgUq9YmlgM(gq5}w^09#@J48WGS7%s+^
zm;o~^OFT!s`~8moBWLM(!u8?GPzE9&kql)Z@sY?-21!0iGA09(Do81KlUX%MRq?9C
z%-VkmW8brNrdrJNgva&9g7ZY)Sa6=i8w<`$@<zey-Ru3BYBB%m32pu&VoEwy<Z~<}
zVmjQ8r2L>C_HL0t3HFXif=QNHPg2P+86zD~=<sg`JrL{hKI;VH&X-xyiTsa`?D{4&
zcK9IO0p@Q_rWUG_iX!uIwQ6;By7_=Q74#RQMx(J~y9?ZSd_rJ8Ik&kn2Lxg1AcEkd
Iv+n5d|KHMjZ~y=R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_add_silu.hpp.8120675903700DB9.idx b/.cache/clangd/index/gemm_add_silu.hpp.8120675903700DB9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c6538cc9ab3298a5610931f7318629853394863
GIT binary patch
literal 2292
zcmYjT2~ZPP7~Vt<0}II#f&_@f0EYW8n}9?Jmk<sil90$DByON!r3%QUEmXl;<q|6`
z42RI#s#6uEf?5PbL93uR0_}(`qeQHzwQ3zpQO7PYFS|RFx9@xZ`>*ePZzu7Q;o;Ag
zLQtqSvM4)!yT$;5AO!#AZO_eOQX$9}w-q(rOvv6k>dm+ncilKBW^+!mI!2jrTz(g&
z8SE}Om-2RELnnrbJy;W^*Y-8`W{ifLhIBOj)BJw+LN`D1)!bmaXF&^g5qi}5m#Uz)
z!rqXf%PuuA?Z0nzXf8$SRaw6NLt|>!ogF+K&+oUVPaB`7yY+PMS911GwvS!N$;kH^
zGPer7)n}Rr8<o8tf4??2Bi1%bT-g%&xYYG_;!M+q#)=~&kF0n8(%cnhdwR^Y85?%9
zvi3d0F8S7N<LrU_>_gV$)9=^w=ZsFi#yFhnV@Gv$E>F)23rEH#UyT)9W5!E)eO?*Z
z3BmcH=A*02Y?G{t>we$ezR|O->8pBf!=7R3g#t>=yN3t8r#d$I*9EN{mC18hhsTdR
z(zY-Y%wDFP7bg8zWL@!K?W>}QCq34ZVx6utOVLE<WZG^%PNjGKkyV&9Q`>$o3Nx<Z
z^#osebIz={(5&}g`sGwnf613g{+<R>*@2xMmY$ZHxS-PMOW}2GLiq}cL;0$cP<nOl
zlL0Eb#~?a0cKQZwZ@n^H>`$AjFS{Acs}&?}3)-^$YU6Od0OH&U7sy+kEzY_ncYYH1
z;Ao2b;6Qx$gp;h>eBTyr+NfXkfvmB|l8+R>X8xT%iM3Wmh(ZNdCsIG2xpb~OK->A|
z*tUY5qeGMV*3JVPW{3AC#|6n+Z_w`^UNZ1uUB~U+{60z=>$y3USDd}+Fs>2hy~0lo
z`tIjgP?wrPNGv1PZef6M$;>~+FO6wmmo$XJkO{oFf>;<@$RL^l6=IoE5X$GqAvG;N
zZKsT=#;_qGG!{B748Xs9N}MBkcj}TU)d&wCViIDpFaY!2^pOVpJ#$7dL<8aN4ejL+
zZ3$&DJd~TBk$1unf{g6XSBJLwa?N3ci86ilbs?6hB?(!fRwTqlxxV@o$O?1?2|1t+
zB&2j#dI2h6L>ZxsB9O{M<xN6TRWyMxcZ}qRMPMX9t%sK6r<H2;ekOCRlV#CD`BE6M
zMQtTSGj^yQ2|1(A5~6$;)P;lsMS;PK2LeGbA`wYQD9R-Y1;oQhv`e%eA?^EY@=rav
zLWPlK=rR&hHmHprIhtn%uzD{(#5b`}Rsc~Q59I}bdZHToYHpQ1Gc$k@KYu?7QOX+s
zH6#?~5w?mbMd~5dBSowQ|8eaa!zK8E(wKB23zx>FgDfy2Gm|j^-QZ0zS<onuURqJp
zR&~!s9Ue*TsTP7=Frx9&h)77GNd_o=U2xB?Ge2y?Ta-u1WndW(5T{N6XtTKJv(|yI
zaSDvAL{|oanzlxFsZPI5*@gSY!|^sm=V@@79)WhKw1xXpb@^@Ui3egXj3hD=^;MKg
z=TZ-PdnMo;GPz6v@PQF7mn#72U_`_d35hHrtPp3wC}035Wd7!XEM;n&@E?5y1$c#y
zsG}bdiI4L2sQkOq9qWz_USr@Cu;j}@(>M~!0|}&HDM&VjNFgC?O16?mLZ~u;Kq|J1
zWK#)Mq?l@!niNy*uMQ>rG%O9trV(qx2ph)6NHH-KBjwYwwIrKXsMY5KWKm5PK0Sal
zTw0DYcMyE<i?w#EM=EV_4ar^Qq-(2iQ}}>QFrtW7tOuy_iyGUp7a~`jWoy)0M4X-z
z>Ldc3H~uN&#6r)n;~Glir|FOJ`*O~Wj@jlOe3^~1^;hh&u4+){R#T2I$MfU$yT`8H
zA2pwvIgRUtgL1?~CRfx|3^Kt;kSHjC&`FF-j5lZ<?7`JI|9J&EfsqVGMk?rM+S~B8
zF@0IZ{DovO?TIX07S|p`z*%F<9DvJQ*Ejv1;-R1Ldgt{AlINyx(ag{Da<M%5{DeL~
mEIv#y!T+LE#O{x*`uTXjk%cAGoI!`o2^)i$S`t$WV){SFRwt(b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bf16.cpp.B8DB90A206756A9E.idx b/.cache/clangd/index/gemm_bf16.cpp.B8DB90A206756A9E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1de2a6c84c346bb4a21c4cefbd1f895ac223a98d
GIT binary patch
literal 1820
zcmWIYbaNA7XJBwn@vO*AElFfyU|<jg;^LB`OnD$DmXCp<V$R;M-2BI8Jm2QuT;9BB
z+1p*a8GlaLa&p~`t-N`*UluPsvBlR#e_q_AwEcBc-l|;RB3*EHtAjxI5)Zo?!KFPL
zdkmweo$6}hRDT^9YFB3^D7kyrALIG=`qbl>{Y)|2W06o(<8=NRyL8&;&6k;v9sSAp
zcxC77+rN+O*gMN_-(Sb>{CeNdKd<kOw3`>--*)80=kwP;*ZfPZvETi>YI1*m$LC28
zK0i%W-|=kg)dSi3Tn`J6O7CWy&aHgq*y87xULC927%)Nqpk#pkx`X0^p$9K|h;mo7
z{(EgIsyC-#wz7`NxoIm7y!E{^H}#F9hSw41-0KNUYY&J-7^#Xas8O1`S+6$yf!RX0
zVAV&47O!TsU$WAjBqYUirdca9S0`GaRn6cQ`?1BQve9)vKD<=+J`uH)nJ;^RfZBpL
zuGc1<em_NbVt~Q^V+z?}4w{QzofWxYF+pj<4u+L2!i`^Q^tVjhbU>pf_wYFx8=+$j
zQXYrd#Zxz&%Xbx#bzS^%LYm!+d#P3q4<%2Wa#8CsmCjw9KKJfkmxYYWS{FCHnfbqA
zbB?M^7n7=|>cm5#hEo@_FS%IlWuq+^axiFJpoYp(X3JKIoEs(=KJ&V|X?ix?D9^dQ
z#Kd4t&(fB%J#HdL>)zUGvwrnDU-01I+7m)zEL@yzx4E23+8*i5SoTR^OUT8C8z%p{
zWwpwFw?+EFO(ybDdOvHYZJ*vHb@tMYt%=zWeY(~)@10s1$P|5K6U*^M;-WdM7lkie
zId@1T_Sp8DHJg+h@<d8~j>j}dYzp-_uBy7SU;Al~27kS~!KA1YN!d&J8)seH`bo@2
zx9RYy&SidTixwu_dcQBu)!SNNli$1(E>^AlvWt2GA5Z`MTJ!I6rLW4fzFgK!=1zS&
z$tC(uj^dL^za}MUUXgH9m)tCw?d1FIq{P)TZtj<u?smApzr?cB;(3Xc&(Zeh&662@
z3s&9H2z|4?$2M)fbX()Bp6~@<`Cq>`xxK$8kWZg&Q>KyM)Mw|s&eqA!;r08$#Jhgu
zL?h|=PK9^Dp~r&WU7LLORnh)kg5EO%xqn~t-S^$$XT_`hgv}=6_0RuIy*%f%>R&cf
zNBesLQZueS-PV}b@o`n>o(55`7xuY2?;ou7os{pe?d<KY`v<f2gOiezE4=F$DDbwv
zHIHB1{&!lkjeN{i&V$pQHarWRrTWv@XOh9&qsOoPs5g+QI(PAhg3l@o#|r1;h6@gD
zc({pe=b4v@ACJ}jdc4RndV6JZ{MMz4ms7v}wF#U)Bjm{Ant#_N{#iZ~dNaTIoa^P@
zk4K~e+RD^!|6cTL>Xa`RHk@)!I2^poF(*TRU#wBR<=hMXlYK57b?>x2#1tg?V%b&q
z83tdp3M+gh6wSE$F2p-leS62V=a1J7?PE?)va23kc=91;#VISEt><3ko!jCt`7if|
zUk0WO#g(~9vw`J>*d$LImKk<U+)P0X4Dm8-GW=iyC<6ut;#VXdzblI5VaNnZE3znZ
zfC-qi&8qYB+ILM}$;}i9l$K_b<_8ln>5Q1_>x;#EpKvn;0HtNwWCg(lOnRcD8)M9s
ze@Wa-{y=FdHYt8E0h4AZN=+-?3*^Dv$jF7&Z9rj=!(iHAZUPE}90U`N^lO=au<XHC
zMxJ;@5k*Cq$KosI<R|Ri^O%_@UP?kr9_A#Tcxee~d6-c=@iG!J@-Sz>d<C=s<RzE|
zFs(pgkXK;BKwsn}f;h!R$;IzMj93r>0qHOX14Al;8IE9vAef;DW;BACf?%d0m@x=u
z6oMIsU?w1#kqBlSf*FBef?W^hGRJ?otrFv<8GTSVJt;XQH7z7OEHol2Iwm$QG9i&6
z8KEZ$!A#Cg$pI!F2A_04W)@a9b`DN%9$q0nK>>a*VK)~IVJ<HbNl`H=X)9}6J9`I5
zCj)U^8&wM}MQoxHvU2iD$|`E=n%X*ghDOGw=9UU(GWsU2;u0R-8BFfM&Yq0E3;@!=
Bm%ac1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp.ACF8DA35700D8D8F.idx b/.cache/clangd/index/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp.ACF8DA35700D8D8F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..38f07a5b480675be0f80a3194012a2c05c7eacb4
GIT binary patch
literal 10580
zcmai42{@Ho)INJTogC+Lj2+W4&k<>MA`wC)Y0#uJC`H$8{<yc$JW-iaH@DKDBI@3f
z>WYZl=#s0_a4Bg{k|dRXpHthnd;b6V^E}o$ueJBL$F<hG)@Q7*x3|A7k2hkr?~157
zOTzd(9!~@RE?%-I;vohhYCPVG-!4sf5oqA7{@<#mVdHOT?wvYiqZ(ViL36)Ho^hW%
z*K-RlCAr=;@ag(!&}<$b_c|u}-2t1^PX=eUUz3?yn3gpap7d95TET`kwi_pfcpOR|
zC7Y61K0SPii*cr*`Y;z+<;0#$>%W&<P{Yf2ac@G(X0isttIr}oy@J92Eev1M))q5*
z<grT~>cW(jp*c^VUwyi4<>$6diO09SmA+y7MZ|keo~#=2EMrFK!s;sz;<K*xJ=Ioy
z_}=p?F=@7Y7eouj?fQ4@#&P<`=Of26U2k{$g(nui3cc9o>rvGI;2Yz@fji@d_VsA;
zJi(+SjeL@``Gw%#pp<IumS|(CRdejZJ>e}o2J}-`>s|5J35(L?0r4irx{kRKmp|P<
zmt^cSAg6H2aMNDO;R%g~vsI7&5`E{P2ES!_zc)LS|DKd9bAsMIf9tWPNAxABXhr4n
zy}3@x2Nq4Zo3GWgecr5X`ft{$9d9a;{Ws#NVNBNJhKvt;ho+uLE7dX=1#5-p+?hWm
zw))C_b3c8PUC!4kGCwH&W@*r5p|!*OO3U9<gDRFR5kx$56=XJQrkFU@?y%jcqZYly
z>2`kZ-YkQ6%m__YGqt}@ZwfOx*;k@eSKZsHY-O|S<&1HzCGkBA+<nLH4br?G?W#V1
z?mvO<0}k~J*sNzLo>G%u8}{Y$9a~{q*@UR<>~`OKnWN@M)s{cdTU7De#s{au+jq`&
z-j;J?arA-rIdMkWuKM~{GxVyw&nGpRC-TmY4RPwzBEG#V(&J^C?tvS5y8T_)4}S<s
zPprv_Jg~%JM%7utPg^PiqEANW^v>E)x3~ZGvz6)dZisaqv#hi=&b^6U5af~nruE;M
z>$;qL@BdlWdE%<iA^Yj(HFpF0mEQ_UpO>~{^19ZnHMiG{+vnL~!z>P8us(a+YnR-=
zHosc6Y}d@?JIxF;#fgF;{!WXZIL*3X-8hWBW3f7Roy0xnQ>J9^!0Q&R+VgshzNPzM
zmEF7MO9w0endekf(KOrjK)XlPdHqb)72_>!Yezk|zt#A;es*+9c=dsUIYR=yAMI9&
z&Nx}TaMq@g{d=4ke*D<%vW`r9&wkGr%sp?O)P67d*~^lKO}Sx-(;mfIXSUYsEpQs)
z)_kkjDtWi&i_nrA`=6~GykP3?i{6XQ9G8^$T$kFpCYmjrY88Cp+~B*f9*$g*_*YE4
zVBkpy6VH0VW`nt17b7F3(?zR-Cysyl(+N@4#jf}{4puvZkL(hjzv!D|YO-p}^+YG5
z7<aYz$8LxEOs<XT-R4ygy6naH0*B;YHxA#ezH`<lIWJDhWYCkcUAI+(?EccNJ~r#D
z%f9r|RkJ_5duAOHb9L3G!d9!dwN{0`x@%8+XXn~vssx@oydW_7rN$1gvOkq|-YBnr
zExG8MX!+hYpsS!fX5W|B|89Rf@QdK_Dl?lRhn!Mn_s&a!n~#q63VxA0!pQDS#$T$R
zM1T5i`MKb*UTGyLUojrHsl0OcmMr%H?vIP29vA$Xb1ZA+!R6JZnU4qVQZq`aO2};d
zIA~vB<imM4jS?eb%nhEKq%wE4lZ$U@8=w3s=kTa=f~foNw=PYsUGMPch0$78SIg{+
z3P-QLS?l}7XH=`R<IRM{zXVKvY`U$%X=Q!XywJfzBQ6*wwWOPsw=7+~W4dWWN$~F8
z&q^0~hUvD<IW*mL|7O30yq%u?BX(TiO{iL%zL<@gxwHq5CvfeR{w2t%L;z7d9_!!J
z-&!HCFI4=9EVBcS9n2V`4s#ds*i&faDKzQZ2Wi6Gn)1cP@>!)OgySZEAoV|x!M6|6
zL^9okeW77?SOh_M5UD^aRETc%>8x8SHW$rRc?+=G#?Zz>A&6S<zEADYGrTt<n1a;<
z1q1aJf~aj(^PT(5JJzlaVR-zrNc}7_{Psbb*ezY!-1Kggfe6OqC-f2Y(NqYcc5G+Q
z6}|V#5J4bT+X?K{6@sYy{&U@Zr)pe`DlY=7Z4GTL6rx+*68cMOQ`cKDEdGhd&NiBD
zrw~NlKmScy!+*IYs=OJvx1+?-NFj*Y<!EK*`m@PLL=cSC#sXtCg&=CLt#4cAAO930
zf>5m173c~Tf~ZHT#BX~XRF))yiCAqaFqJ4ow>sU=<Lo--pct0nsfC$_*((H5PfT)r
zpF6(qGXc!RJ2zi6UsEB7dTLtJ`?#mQ52^Cz<A+-tT3aXtQO}zidn-$4YoiFpV6}<B
zL_;BnI$EjMHDy6)mI%gSwY|VzT_L*F&-}K}Yrf~M33Kq+*+@MbNxyxNCb9b$EPET}
zv_S-8qj_wkI8vezMEz^|?McI4ByUya1>gakBu-L=AnMiAKCXW35c)(6OYqufn9i_Q
z2%_FlX?x1YCbfj&ZA7?x1>#pA<!>LP&31QhG_~f-Oc9L6O^gIaY6{V<etdLt?(?ZS
z5*Uj|2vQDGR|ulsG4^oY`=<r7L@*uq))VLn6@sW!ELyX!960eo3@dQ&nFceh6oRPJ
zb{%%?>GIbJ5lqItT?MWZg&^w8+{S~W-sUY4z*xL}fucYyg&^vzf=8xLTHj%Vn~Hmz
z3(VCOf~XG}Sa!MDjXa<ZKjWQUC|;<c5Z&smo1^D{-q)rMv+%>SkT46M&DRgoB%j|F
zb#36^%_4}z0~iPlG!%lUPx@$%t#2H0R0JVdZ6+{NQwX9yZ|eQ{Xr0LoF)YVx4>t*S
zR0yKJYWKnMMX!kqHDNX$e-F~%g9d*4AWf=%b-i2|JLrTeZxSA$pQN8uA&B~c@AY%e
zH%d>7U>a6i2rSeTf~cRK>NM&-`}_nkEXCu`F`8qi5Jdgq#e?k*O*70yun6~-2qZlf
zf~Y^wdD@mWp=#&X-AhB_G^D|P{UA-gte&!RC;xelDsLWsf|a3_g+dTDZ_%9a#ioWl
z9%}X<J-o#88SvRSWEzLe$jR{845X8RbjessPNE~fUz@o_%&=X^p$qlzzQ1f7IK%;k
z<G~@G#t9%x05`6EGRTr?oC2~G8mEFRmB#5XARU~!{w*-L1zb4X3PW3AIEUN7tqnXm
z+zun!!H2^gFtP*uIot`OI$;dKBpxZLd>4lpHc~IrnRo<-tpkfXZV*9!d&nxa1vx+5
zM!}5>FJAV}S5acv4Pd<ihLB+yb|YACr12)O-o#-N2a(FR=tt%$sycc3!_Bjah+#J&
zW)rd^CrH8~sC)~CqzA(`fkhKIlA%@S>JN@MCN0L3F9gj(&?X8#TMgoBP$yWu&?u^Z
zs!ydrK6$MTU_})2$!l$-@g~sPMB}ZXwUx$+pp{7DB+yFYFbUUqvGQ0=Zmr~7nC7c|
z7oHhb<|hjxL7HL9U`QE#f9`_UUGBVfuRFzD&UUZC`&9!PHDEv%K^}njOCI9MpbYD(
z=h}yOL(}39rb8Q#<lx=~ASwVO(%WqF)){Z?uAf1C_6s<C;dYm>1H6{6)PNj=UH=Q0
z+`|q=;71jKVG)>+M-j4sZ=jqha1)UoL$U%y=E<=8L4QBDX@o@J-H|g0s@o1E`;F?h
z7xedn1sRQD8^Egp0tqH$13yO2G^ob}%O-PUSHO@8dXw*g*FA2N`+lP&z_63WlMS|i
z3sfo!VnRX(Atrdk<*bAn=>n2fbWeDIYnjSzcCt3UbuGZOfF_wK!?uEYD@Zup2HI_)
z&*65^ZRd`NVLL$D0cM=O6HGd}4V0;0+;b{p`4?=Ptw^~QiMfXFkn%ekw;|;=4$F*M
zt}GCQwW#pf3y8UZ#N^ER*jI|XP`3eLCkx^@&Y<{gJc#3IoB-kk8W)01A@n8_;j>jB
zuA=o-U{wV+w0>w6xY9VJ3MLawxSmkuJC4V&ACc%I>b85iIXiCz57G^L7I5A~<3w;y
zq;V2BC($?s%u>LP>wg8@uFx*{0eC%t(M0cZbKK3hH?{;R$!*8$3wRR+-aqi_qH&y(
zbDWZl=-poZ`}%lx<}!R@8K98?dPKpnyFg<Xjdz2_ZVr3=l-<Y3t?<YG+GFo3vzGMC
zU~Ht@k>PgaM7*o#+1LJOm;tK=e6|BgI#9Ru^6_R$KGL`oNjho#2}wR7dooo%`x!|-
zBL@!O2GiRxm|#M{RaCws-VAG`XCx&AkztMXjHTpw8P-(KR7$8h!<y-tNeLlhSPMN1
zDWNCWb%x5qIIEV;k<Fvf3Rx_LHX~VLw^e$dU;8@pW%%#dDvLo_3~FS)3|k7qQW}?o
zu$;!1L3sIlY<d~TLc)jHD<HhW>B%y%H<YIroR_!_YPZ3d>-Hab{RiU-9{DZvz#b?>
z9HPg0AbAo(Jdj_;g`^v+U#jBMH8(UjBcwH8@Q$A(eYR}F*UDgs!3aV%8P?9(&Xc%V
zhP8LL_asS1K>0sG|Mq$FH=oZYgE*OULVPv_#3?jR1#v2k(?FcY;b5WHUr}?^<>CA+
z5}rkB#Fva4t!e#qyGsuhhP{J?cM!GVbx2r8<9kSWkH+_r@IH;}k+7b|4M^BP<3~vN
zh{lab*hu3%7?20uM#!-FFd(19<L-s})$a0l`ym&a5SBVAXWv~JUTP`QDn$~qRFW3r
z5G2nDNx)<<kFov(IS^ulLohNaK-PpB$+2w#>Ps+59|bDkrH%N`yUTpIdtCw!B`}cO
z8##S&EQ3KDt^mgh8rOkW9T(}Q6*xyuwX!hBcO`IW;7D?_7|h29r6CVOpBbFKxM!h&
zL*Fu+X^+-~oC^PG2kur3y2W%jE(P6E8kd7^IfthoIdb~gSm(L8<11wR3Yl{geT|G?
z)3_NKH*<JSVh_tW&BY#gxBUbCr;tr0Ni+_^@(h|YuX^yR=U8#Xh_w&<{j*MtdmjRw
zLtGGzShRRq;D^J*`3%1J{A5t&=92>a6dI=jKb6L5z)$0F<k*9CCN2DuA6}ct<eAo<
z4Svh9WyhMv_8|_LVOL65+7LwS5?>*C=Fcrl_B)at>Gom_yASmDal6Z~55empM~0%m
z6$$X}e60|WZDZJ{VEGiB$ZCIHF|1O0rTVinA6uv49n?*(`D`6BtfTQgWO$Fp_mSa!
z8rLJkdKxz%!v-2ZLWYlM+=vVtIsEgfI|1*FR`2=YHZB|GdCe|l&=`E8j&6=Ygp92y
zZD@?)S6#>V>j>B$f$q4(uz6scN8^03&F3&ljj*4VS4xNtGwed$g(l=!aGXYV8K^%o
z3sOSCo#gcs>X@pK>=t!QvIZo3(v4TnzBRPw$(luYk|p3)0>0!J<kst32EiPz0N)A@
z$Nzd}O-X%2p?uN7j+4ov6LS5d#_y@Bz?b+iI31>w$!u`SrtuN*KLS&@=N|>9qnw`9
zeDK|zD4s|~9eMeO_=fdwyrj;RTlDcHbk%hwB*2qr;zZ$VF^KHOW<RAhuadfk=rXKT
zFDofY?HSg(m$j5QIfk|AWg{gS62sc|vXv4l&#>2#@^#vSS0d#~8vlco|Dka;Qm&@)
zO{9F2#x+Q}hQ_r>xt7LvkoFzaomt^*8EMzi_#V=}N8|fQ`#z29k#;?e8<2JbjUOTH
zM>K9k+Kn{M1GhYmHp(NETRx48!L}Isk{y>*7TZ!9mxFCNhqpdG`E#)Qi=_Y8;}9y=
zQ{}rF9ZoussDtL#8*nXy=E_^ac`J>R!7TZE?reSqy6Gar-UqMy9Cs$f2HzTa0g&7=
zQv1V(B`^GutHiJ-dL~lh^Km3xDEpa2LQ)UJtDGdBMBQC^X^=cOwV`+M@S+3H@!dU-
zbk8FT60=j=XG}{_lm3fMG9BrrBMV~Ua_o|hd^lVLc10Y&C7CsLvGO7#H`9ZleUS6a
zX_tKt-(0uS9uHiAv<r|iiSlG-*wf3akeq|2w*yyihDMGcE(*Icq?U(th<D$6bneBP
zlBHF6<<cS2p``l2uzj5SxR7jo@5|m}N_E;V;t6DeMkeTz+p{ln6yGms-NT-Iwiq$R
z-)DaVx==TjkY|4b;y5ZH&;AC)(>MW!CvaCmUXJlE<c?Zij<Klb^40y#y=)twn}5V!
zD+4SuxHM6o{aEgz@ouo(P2(CESi><k!h7)%$Sbwn5y*a(t?zca2S(gA^tkwO_y0P9
z0~<mz9xpAD;-lXR)?2xn6~iWibs~+Ez&eS;Br4!><t1P)WHf_zGw6^FCY%Xh@vk*t
za>m(jeD;N|K2nEgv>o-@jts~Hb5@+|F(zuiH&#f*5(z1%=JxQp)#4N3jL+H0)hUPs
zABOGc+AoL{baVU5#$4JnBON!IFPrbhP3jbqokCvZ+U40cji1sIWRG*!D$uV2DVePt
zJ5_P@PdWCh;s_(jCa`77tJd6RKLW2u9O1(B4*I>~Qu+M<hAq#~4V<UoS;LdLkjCHy
zV|WUUQ(<^2jnhGv4gq9-C#K%+w||S8CQjVCkgy9;ge?w)aU3CI@bwhN(>MWy2{bMQ
z)k4r94`%Q}g;kuMWSl)zzAL@sRLk7lo?JPG?L?wZ+C8R%St{-Ru7deh&i$PpbaH2G
z)k_~-W;>6x&T~1#8Sm5;$A>?;jnCs1vUr84n#gNp@tVfX$fB9UXD%%nH0d{+HU*b3
zDB;R*axP&|M&k-FsNnFqt!kNZ>f4GGbt%^t7|Cr!E4a79C=R!QR~vL!*ci4QeB0?0
z>VSX_n8>y7grH82WRRB?_(o0@Pc<YSUq$iCg1Enu+JpUhH;N*D$oon?U(aURO4j1S
zMGN@0fFBu_kIQKRt+cyogP=AT%jw%;d^=3wa0g84;EGFpwiBjxLMYLfW{+8@>aD!y
zfBB2Q*GzKQI6I&OD`d*DK_vcMF<9|XxI)|muOJWn^SB4`arpMn=P-FYgZ;tRS2M&P
zTzisR^76(_fBaylV5dpMzHgYvZb%pwtkx6fT1dMX>5`Zww*Z$eM7=tQ1D80C&B-mm
zC7#9!;FG|aGJ_rTxIzdaD`v2Rwy)-RLzPd`)y*fb8vO7~i0{FUQ2r)~^9iyRBx~>X
zz_mu_dfc%yxrW!bK}mmuQg`lm>-0$PBU^$OW5r?6ISk#3I?07EF>WR?iV$I!!BwT@
zA;NAKjdz3HZW`CX&>F75dw1>QTSbq0oWvu%N6POJL)J_xN!ZHeZ@IXse?p>9v|WEj
zqR%u=1GBX6?cD7uS0Zc(T~a^9W4q{wTlctck9R%QbT;FOokAL?kO6Vf<P{sXnknKb
zR0&%CwvAZL(=&dj!*yePaic=eEd)!lEcsDtSA!|RFOrYH-(y$#yBfm==>&D#Aj6K;
z8B3vHjbI&OpA0)gJVTpg1PmJ{4$~%8DTbY?JX1^<7lV^a^K4{G_y|tDkogfBA4P^o
zkvXUDK>QA*LMA7#-||1wxD)X^Y5WQCKhgLz;(w;`ZBV`q-Sut~FO^lkd+mtL^%BU+
zffy{twII7i<0g<jq45bN*-0g;{Y52NDUGX@WH)JifiElOllFu$Vmtr(a+Ad3FJYhk
zr79=}pLxIFev^syki-aIMfnR)?kdWL$lS<7<r$-82^Uc0Uf!~d#uZ>$!D0NSG%_q4
zKg%;m^iaEb)kH+5H7eG)DE_+p&xr67QVQac+eJz)-T&hft~+N}xL5mnL_8278$7bK
z#fxSwE|t%G4gRM9@{>S)k0F2L@m6tP*HbOxsPDgUO=7vPE4i<$sm5!m@4s?Q;<-O7
zv9pd9WyA84-r{v=Jz9&_pqSWSeq9xZ;!y%xjaKs3agBMiqGm_px6eGkIsQug9)h0A
zLJ^~?t|s=;@bs|KQ1Q{!(bCq{8{$0F#dVn5aED%YGV?(;#{bt}B?eMM6H_w_OKaQS
f_Ktm>`VSmzG(f*kzY)D8-o6_UcujC0iAMboOxJ_F

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp.83BB18F6B66F2289.idx b/.cache/clangd/index/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp.83BB18F6B66F2289.idx
new file mode 100755
index 0000000000000000000000000000000000000000..becf34f3ae7380a1ec5561372912054743eb9fe9
GIT binary patch
literal 10502
zcmai430#a@*q`$>O*75gw3@b=HZ78(n`K&PQ<9Rb*+~-WTH@Z`xY{g*3N3D2U3((C
zs8CXJZFTELUAdI9laLD0cV;-fuiy9mK7K#X%zyr8KhJrd=RBwJqrJVq^y6?mLq<nM
z%wI8=%i(Y|@&Cvb%NN#SP%GkaqW-ut`9*-CqsD`n$0H_HYVDpeJxQIeNz^jiU6-0!
zm;GP<!>ie^)b_`GPVU_5wy`yOY5V?uXP*qq?!4(_YHnKoxbT#p#^We@R*SlEYLLgl
zv@uT8Qz~YMu5dBVmTHV}ajcr6LcP0v?G0*r*&f?Au{@a872J3h-u)hi&0ex##mdOx
zIs0xuT%E}Ebsac)#rydyhLsc-?%ijz-$iA0w%@clTQetnPK&nuef!Qcr#c(^Irv3f
zY_~pFbSitk>QjS1P)2C}M6-<bN2Xm8ww81s+?F-(=kwc_{<dw)5<lziQ9r9$Hi|ZH
zaSR-w^r*So@x{X;bWZKSU@sRJ?uKihGN<QEy<nb9l_y+Zwq(GyHgkWIzf>y5ES~mL
zOlNDeik8}9JLxzd<@hWUT4icNf8&{Fqqy$f7pV*8uzd27v72MBqXlR#nqhwx?!S|U
zEv6v%kK03EmX|8s)!Waz(3Eh5K2h1J5xwDl7#(Ff;_2A=d5_u>wyw(+INS-(KRSHj
zqgG7|HH}-6$MXWZMU~ssBDV>yX+DYyId*8ClT*T(b|a;L<pSR=dW)~6mv2&$73qc7
zPyV}FEqc9a5^w9Mkm*+jX-w7Igd!I9(GD)@%x>{JoZWaSp<%bV_22!HLJN%e`3_HW
zObq@>ICy2VXoKzZF)s%7w7g8`jtM?8bYkr%&wJTp7De2txNo?;>W`%R`JtV^FL2zF
zR~fl<e@9-d(V^i61~;<wtGzFzKGsg*6patEw|!$&za!k^Wv1T#%Hw(tE_ByvUfIdD
zdExt4*v+Xf;w{;H)qm-!rFs4LCEnfbaI2_ldswANZ{R*l9nJG?(-#MNoM?M@J9vGM
z{pkPBlz%#T!{^|D*?nsp{C}#b3)&u*xoz6|cl*}XubsGO<cEG#Wa#1zhqko39KE>d
zRm?9tf`9$pOqwl9;mQ2$BcIsME3s-ZrT;ZgP+2c_kM7PE?;di?{GCpi^4L1P`!P1{
zudf`a`a8_N*6MN8@crr@)fWu1)uJX@Sl=1*d_Y~x=f;qw>7h0I59G-Lyqg<Tmu8*1
z)Ec~Tl!Nj~w_`^_%0FZe82Qul#S1R<P3^pw_UvVK)5fE7Q)V`=vdVtfsK40W)Ae=T
zCCju1trxROEB8KIGi>pUotM3r7aS8;sI1TUq}23wd7<CIHNP%7+i)edO|t4es4WVa
z>6UoORy?fo%esH-WyRtzXOA7Oe-Ny@cU_OtjTwG}a>ftQj#`%%<mvWBZ>MygOMAT0
zt0^ZhZaI_pe&X{@KVN)#=g7}9Ep7e?`1Qc)m4TMSVwFsWJ}KW(uNG)?QLpBRc9F}$
z?PW0`UG2}Tf}(H4Y%F|d`Sy-wp}*d`v)+e}_RCfc$j@0EkoHor&8z$jU$>2$&?>$>
zI>n;H+P~*qMf9F8t+%(n9rA_uD8{T`v0Yvn-~H2-fK7+TdQEt7)YHhOAnT%9xA2Vb
z=4Iz{^vkM%f8BV}#)_()oA<d7c7IeH@#x%{yd(S89O$em%YHOuhq_UEb$s?jmE4f}
zgL<6egI$TM=MA6v*O0ImXP4w1IWyw#p|vGrwTB)}vR)Ori)Xbftvf06@q}pym7jEN
zJAK{Lq-<Vy!QH<v4EUvHjk8hnK(EcsH^8FA<MO@*c9Uv0)FoaX^zevJiN{FY4DM`=
zq|h4k7!;N;K0cC;2wtho;c)f*^~a4mbCm}X91iWP;%lW4*!n5{@jewdq-eJH7r`$}
zIP`4O*#i`U&|;Gjp&$Ki9W`LV&m8(R8g&{?{r1sYWVQ+YLenfq2!Xf*1D=6EAqcJi
zpU)lV0#2I{reoTH=U|`^gf@6{^^#NDa6=J9;r{0s&aqSoLR((DTEi*YwoMG<F&)4U
z&`=0M+pJvq`f>XhLm^DUjoa{Sv=oBS13T&S#T`CbLI}XLHP2c@AqZ{nxu+Hdz0pwP
zOvUsd@gRvp5PGo1)TZv`cb}?p!Z2+vH8)oXLJwUHi^J6mthHc1UYdhQ;~<iJ`yfSJ
z4p(JwC`!u}!UWuclqXeJ2tvCp?H}Rr@{zj`#xLd2i$sgW3PI?RsVAe8mk%}2fMu9Y
zN1}A3`R#)g@k)OCX3??kAR)}c9f)~ifkF^^lxp0TM}g(3LYRVSW1g{CA$sYqlhv7@
zUU7I3jAsxg4AW8wLWj**Rku$!xkU)$aBU-=k)}fQ(z&{eN+NS()j0l`9w;6tQ3ygW
zRqAt-&zrSR2oo{gpVwbQAqX84a459(OyDyS{E9a=#3W>(LJ&H!$~xbtUq&g#Nkq8%
z6~w)Q_}@NAnH1ieDZ5uET~^~P!bL2k7Ul{;=&kPFXh!XqU?Ggfw1g*7R|rCH8=rH$
z<LSA1LYR$dU7oH$AqbuB(e&X}YQzl<n1|P94-)LbyZiNn6v<du;kvWK_+K^7V%&nI
z)Y4ob2%WhjXP}D9#gjsqhG}P>vsfVrz1!;1t^MMd01>Rf3oy%c)&PYd^uBY=rcd6r
z<5M*Q*EZ#uYA8f6UHI>6$I^N`4Va6ESAaAMko4OJDRP4QKH}z(zg`O=9Jip))7Mmp
zUfMtA<mh{$abLA7TokTJ%LSoN`Dl%AZ1FrSgdkkoglD3z5QILx>vT$$x30e$XAY)q
z#kNKYLFn@zo-gqbbZ_Fpcsx9Rp})355V~Yu(S=*j<8!ni1k*c_!A>;f+XpFf^I5VN
zug83a7KGv=Cy~ZUWccla6uI}i*>O>#{eCshLOcd5sg=1x5c>Y;Tjw1s%gzd6CZ^4J
zX6gz-=r?`KOf{QIZ;N0hUjKPU^K2A?&|NR?Z?$_o$4m&zacv!*j*3DM`jd~-ok`ZW
z{uDtZt{rR?Y@-l_<}9Bd8u?nn;h@)kW8F$eJ_9bj#&nI@iEoR}rMDy9?MROtaV2|U
z?w>b<H;X8`2if%?hhF`sW5F&KSU3*s;@CJIoZ`WiDW3*TX>6PhPU&o%0Ztigyd4H_
z2S=v<TNw5hTo~LA&h6mF;0|!@fRPOD0?#h+Vem&7^%49S+zn&8VI0B48!4%N=Y=SG
zk^Uk_Vh<_$E|}kCI^mt@42n@-ocDu8<W-iuc-jAisuD#ff>k2ONT(E?1Xf9Gyb-K6
zGMM;4py<h>$@;|Ss46@mQDtHEscT*T6r~{CI-(MhCD}{01qQ<w9+6zaqbmfhLeL@I
za_Jfn)qn=U8ihs?4jDdGJo)&v6Ty-Y^6_aWvGGRG-pIzupq<RdDWILg#;Ks4%3$KO
zk?J>BO;%m}%~fMU?x%^i>fN%BzIMpH7Ei4fG;6_-42<}0Q2owlQ?#?bvn{c2#DnA2
zn<ARRT0OZlC#z?Asms*4B{vUuaSnv%z=(8Vwkdf|o6W5P#HD+{t_KDZg6skwtK99A
zhHM;4uDH)Vun;#?4ANpSA>~OZfHg}V2avf@wCo6yokL_Q6ulP=_A+ZlLIkerccB5U
zwhhVtU{%`<2D`zW^hVJS!RsLe5KKY|yi)Qwg4O#QB<t*LhJ+M&kn(tfRZJ#xWrlbc
zyzVk*Xz=iI)p~lG4%VBuK)nSm(kVr^gGM`u8QcLn9bmxVF3{^@)|#R}g5)EZF?=_e
zbTex~A{so2ui*^Q8;X7g7SF(*RCLm~yemKJ*Dv^_y+izWNJI!cBE)ZJ<M)XFp21E=
zZ>}%q&3&WFr7t4tA`<QU=Ev~vBT*0PJu6&n_e8PG+2PW0Ac|w-co4<2aUt|8g#Kh8
zT)G-W)oi{REUTd(n|H2;;cOgK4buoFk()sEyNC_1AQFB=y+>SravhVvgH)sFjo`SE
zjZ?reg^g3eF_n$e!7LqYnEKbj^*Y;}J^-%=FqZHx|4jVnZRO@bCHZOb>H%*;;Pq1S
zidACaI3>q8B`3nWzPjCd>`?YEc*n9pGYj+yLD4%va|avm1kIfcCN>#stUN#+^(|zi
z=y1Jo6Ji1>?9ObqBM0KnJeC}?HF7Qdk-_%Z9b@)pd-eo;mQ#^571@*aNIUrKe+_#{
zw<C*M{R*hT2|O<S0f|4L-qX*;8zcV6#-EV*6B~CUaW@)3rp={4Bk^Zs$KZM}t%qR*
zlfYV4^}B$Yq9ytg35hHzTB<LV5LKaQV|`-@3AriSMBhY0q9}?s(>Iflhy)v#S+cpz
zahJ`Pg|SeSY!wS_Lb8-z`{8|IU2FKu&_D4uUIW23P$%P+W7BKc>yfybz7B%x3{QrI
zw^4o#U<^+^sMmurQ>_ua8etN_WZkiKlwS`RdvFq2okVsdNTcX{WRj1ph~trC>wGkb
zVBf5g)XIdHYIvhfrKV;iw)G#jZHd@+^A<cCeVM)|30NuG+R=I>G0YU*-?9Hla+~n4
z_&e~nZQEa5E}aIVG{)9(>2wgKvvCH9GT1m1M41epAn>{vF<(RO&I^#B0I3s8G;yq!
z)zhsm%BmDyhXi$qwZ!#EP|wB<NYKE>_mJQo8$UpT2W;Gk1dVLmj0DYW+=2uxY<wIB
zABWx}OwlJ`@CgP_yf@4D&JI7<A4Ely=VnaJ+tc8LCt8ZMOOcpNbmlq7@EMloeevr1
zJNu6!4wk}4ZfH94AZL)mw;uOB$e&z@W;U-4$`4(#4Oc4#suc8yu26It=#{ZuWCiF|
zFnD%u?%5;b9T(t=uaNO8)R(E)ii}&?_%$+q&EWYd$`-L&ksf#kz5%||i4v0QDfU+K
zyD6j6P2knUM4VyYuBlk_<#$x3n5-Jf>UvKWK9fyI_M{gt^j!Dwr@z;WaHj`B_aNgm
z7cP(dC7>(EjZ0w-<fefdGmUiMrn7Mda5LCA6S$cS4j+HuuE`s2=@0fVJT2I2WRvf&
zIJ}CMME4^mlA>3euC^tJ91na><u`5CB>zOR+}=syy##|j%$87eGk7&KQOvTa5ml1w
zHJ|xhtozbB)EkU(>3SrsXX6GWZD8YjNP3TrA0X)iHf}`HMmBCn(q=YpLDCimFU$Ye
zzr!eD*ALd_moX!6+5`<9hd0UA)i#h^N~6k}TB5ntx3H$=f^{zRT6u~-4%Wxn_yky=
zVDOrdI_KIaYnLl@dqOG9WcIckT+3lJgRg??RR)u62HvD;qG?hRyQ(U38sucf{Z_EH
zw6UpBKH3Sk)5vJ!kNQPS+EslOD^w2H=dkZuhrs?28|Q*wE=*?zaTx3mGyH}&FNtH-
zW&_;5SVJr(E{qh%&h%><g=ohnU!}FLQhQ{E6m8zeTtZ?~ini!uAtA{Xini=yDIuv2
zini)wB_S7YioSvPH`o^I7UJJx<4VM@WaBEtuVUkB#II)K8pN+*<66Y8W#c-eQ-^x5
zCluC7oq9HIKspU<d=Kf|W8()%=K&ixBArGyZbmxIY}|r$TG;qFxE^PMF1bZ^J;BD8
zfPaZ`x^fFCEoC3KGT@i7`7*FBgF$46D7pgp6>Po&tScBk`RS=;6Wm{<{=cSxY#dMZ
zyKES?Z|*W5Cax$2k5Wc|DY_gi%Aq$rqUft&ag~i5z^j3|l8^{UMfJO!8ik`W;Rkkj
zoQP8k>~J?39Fy5N4b0NM4}be!XLoxayzVpMFUi?qTP{!75k;nGBYh(Y@dtROKg*Vp
znG&~!wf}4Oj<^q^(%59mQ+G^TGGAFBsV!d@yK)%9?9e^%y2q?TMpOSwZpHhb<0Dsu
z^oo!<ajqGib7sb?OKxLJl7aLxkU8T{GmuLL@?mf>*c3B?C%IK)uO?3!GQ&CmItLh4
z%Dm>2^UwM<18~RrNGBf|lTeGK6|s&?5=~;=#^o7CVu^MiUU0d#bY(RzWFRweCTRhR
zwsEv^As6Z0FZ+)x)9t*ByT}I3Y{t>=2_M7t4P5_Fg-aJBs`&d$+~6KGl&qHg5;r&&
zSU3&_$FXrdxWzN-AWxw96*Ai`PoS9BF!$5FulrcHJn#DvTgoi3$YO4L@=J`x4mREi
z7CYIv7KYR^p%V!}@gmDJtIQ(Hev@tJ^_FBWvFdzHuoBhT@7W;rxcqVF|62V0i9uP9
zRu)U}t|x<4GLvDV=oGL@VdGS=N@eh&HlIDS5_0e27E@85RAfjTdtTIe<#7>vy)mI9
z(h-xa<WXgxx;H)x9q|SZ96m6R#5fc^X!xK&l8ZR%;56>at~uLrp)jW~FA~dBI4R_G
z8hH_oId1)U()2g@VOV}=HmC*(vwhWIU(Mti<-Bh-qx>XZ$EQ)AU1nCX1-x3A0F0tr
z;ip#Ss!gVb7hj%mW`@-UI&Gj!dOz`C#IIR;0bz<b%PkEi5CR9N;FiwD8Q_+|#@oSZ
zJNT3GCuh|Ew0E<*7EYh^AVCje1#Gb(h-CsY3g;LEacmq9f_OGA1hqoYBn?w|Jp|PZ
zPcD1Ps^6u`@tvoy=>Vc~6#WSaKe0_@2AE~AP2dgadxJ57Buc=mCC{=mi!3Wava7wi
zdUoik-&a+?^ucMgBBWi!MEM2Y8BxdFp48*reTB?lAyx*Y6`8lP@oQxMn!yEERt%l`
zN5A)qSje#yMljPZ2g7nE@h0~bhF2MUK3P3GR%1)CB3(87EsSC+wu5^+jA3vGcy&N;
z4vnI_V00IIBR_)wN0`Et?}or`CjKGT3|nyd3j;Zw)OYzpoLP2W<Vms^6nzw_A4R$(
z(?uR4@Dcy|3PBnvy>q~UQ(3(52LWi=$kq_*ed#)!k9Z5C--0h0D;KBc{M*^4tOEi&
zU_8Tj!K5yj%;1kO^&^vi;?mtPvm0g+zU<JrpVhqiYyX$Y`D^V|yQC2RQcTGBvZ2Ir
zT{nz+D2Ng%<MAH{zvE0pTpY9bonY|ICr3+PR{rCM>kphTa4NAR<Ytf6?(5qNSs03b
z0_IPcG;L+yRf+K<CaA06d@Rx_Mta0*%1@h14`RI>#DYsK6RpZmn@b!U$AeEiV_hih
z3?>#r5Lp2VJA(l=OfXdKlX_#*sT+nr*r7W{=av7pF}aE7m!xEnq||%stUEi(J9qPh
zNKE8_ZVvPkcf}=NqFqg*6<)?Bi<zq2%h>E-<DFo$lZ|V^xt4ibX;}BDuDDtG6z-t|
z@jLLN!MEfLiJ!6J%3rNmXL^Ec8reblv9i0$4n6(@um70!A;~0cZRBrTOe5V$*v+;t
zpONr08)t%9=J)o-^#+r1Y?`&AQRcB#_=9DAG)TE!MO}9jp4(}pc^VlKZ}POj_iU(M
z^l0pK3PG<BEXbtg`=?U_rUbu8JJzwwrs_|1d|q_}daq0rJx+HV3k7Nh>XPdeMb8q=
z(jgHJMb8$^)*){<6g`JOM?~Ht@TDE~J&3HyImdg9`sT3lAtXJ7`ZD|n#QlI&$>8L9
zUhYRW{)D)n*ti>UyV>|N;(li1df?YXZ(f`{HsV#8Dw=9YCZwu(awK+)ICsA4cTbJ@
zU|#`Q1u%*sy92U1Hhv7UCv1FDNp?z!RsOP)tc;Colw|*~aS2y;iA&0pr%gQ8AW;x&
ztnyb*GS)A1Kl@2kQ8eEFK@$c|BOwQV_Y9vKigj^hp75d0>zFBs#JOWu85hUha$kl>
zKu*$69MU97^Ys43N~qra9id*G8xS!mB%tAtm?3icyvTgH<|p7k5+MKLSl^?`j~vb#
z=IeS^i5TYR)y&sb%-48U;kB&qzcEE(nV*&DD|~0`f~VTSYU|K?v=$|x=vAwKTN4wD
z;t+0X9aEGuFCruy|78iscfOwzSDB~67YHde4Rw)^=131qO;sN)U2PpbeVL=P%kUAd
zZgzcaoca#!XZ(NtDmIizO-#+qEv&5j4;VPe-eJfvqrnEYKY8{Mdyn3Lz-zMmC^Y7O
E0KFeN$^ZZW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp.6BD99D80213AD3B0.idx b/.cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp.6BD99D80213AD3B0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..27e0077ede9a8d9746e3685614d6d6ccdf956506
GIT binary patch
literal 12012
zcmb7KcUV+MyJuz(D{u}=6Og(ruyhbmdKD2?#I^**uAoLi{UR2OnuGM-5k;dAyT)$p
zMk6G)AoiG`K_g&^-6&1bJ7>Sl&Ul~u&-HnHIluj#J!j^fciuLu93B<b?xCSEbms77
z^Jgud!Dwh`7@+@)7N^ebMqsGEhQ_jQe;I#A+2*hR?#i<p)rrt8=umuc`=v3zmd~4B
z7`@u|a8mrti<?J8bUr;gD`=tTAibX&#%-Oqv+U+Ky_=3y>R0L~9*i|}&^i=+X7a4E
zX`>P%&TT&PxP7VTsoJZ5{g8gmU?uNe*Br!d`{&S&()$li*)7}kptmzrXTkZZzq(v4
zSDsmawI|VcYgW&@SyIhv{>P0sdM@AR11GKAob?~49<x_V2JgyvI@nKh%kiYlfaAkQ
z9MT+7J3IN&mJ!FDW7>XKf+<<I>aWi#kJ+MJWXzmO{PZ-<<@gUxFDB|9I+Pvkx%E}S
z_v}^c==z+d{KOGsJVx#D@jdM~JuP(2a%0D`*2Fi~-bt*9-*ESkh4tQ+DMM@gt6GdE
z_iGuxImPSK+b6$%X;=`yKG0B^IlW~>Yk1V=i<>Wtx7+03iMg`K)+T9)Pcl<&;{T5p
zd%tkP!-aJz;YOZXTYlK)`LSMNU9owM=6%Vd`S)8Bj^<qrdb_Rs&wz!e=fs|y*m*^<
zcd<>Z#gDGBUv2sAj%`e`jpm`XW#hW_5{CLV#W>hUSGQIF5ug8Lj+XKD0515Vz0;7y
z-SRj?zolO-zGiX8^}C9cx7<)e?FzdC&PfCEC(q3SduGXvHdzHldpq2kI{j?MpY;yT
z<<*hqOZB>^`}8_1GM)@c_S)RXmz=np&=hvu-?6oP*%z+PE#H_Nt&{saKI}Vb<dLSo
zA5R&%qAZ{WOvrT;jmS)L{aticb}gYfwkYty`#GT#{`$_zcwDUeT{x+#q-?ufyz{l|
zIqpjd&s?<}ap6MDh&E`lXV0s`V9BY~uD@Oi-lb)}VQ1_7(~&+$e_Cek{$1fR9cglu
z_SuFj4{sZ-JZJCor(xImxTu;-uY4j_w9dTPZT{47uJZ&%`q!2ERpr+o?U1gX(Hy%!
zUuW*X8~5i0e>8pMyixLF%=<?_R$2STHQk-n^t)91>)5HcFW#?^*(pZ9U#vN7i18g+
zarLy_FW0QQo_wOZX)Nn9X7??VEoBj7FSm>5jO!e@{FGZ&{}lP*T+{Zh)(J1|r^IY{
zc(6U#thn~eF0XxA1(^xPhrG@1?%$vtYNmA{ystYZH*ZPf(m(f4@MEq>y5|JX9bbI<
zfKTk}_G3T#y_Mhf8b9#e;V&z%O(}g8K4iqj9Wz&^t}T5Z-FPm?Yo2}Ho!`$S&6=9>
z$gF0WUES(E;R`Acq!?#!Iu()YSE0Nc+HbwQN)$Tx?fL)o^e_8h+?cg2ysa5L8vU^C
zVUu)lb7q*EzrN*+!;IdQjJa12EOA^kZ(`SRpBbwi(@*-^75Pevs$6H)#JY!PKep3(
z@%@eKIxPzK-amiod+{Lot7pw{zr`=34Eh2iznXF6;--@O57vJB)O?|pUE0Mz%k=NW
zMlEU!F1Z(%<jfyBwpt$8YI-JnltF{?!=RC$&Qx{Bz{|$+j+77Y-h7H2Yb1?}_$EX#
zewJcb{L90W_A*m_{?_@Z@x5Wz-D2mP!w#&L%&AV_-{Yd8xpAo3ikdljM=rUfEh*~J
z7^k~!SLmCKo$K;n-upqPZ^p$*ZfPDZ6=$c;>~hZxZTq<*jy<1!$|&#t)=GZO8_Vu>
z4>Uh*cs^J@E_}wb7aQhVEomq_lzs8KSJBG_M$$Ra#~o+yIyq|NUjOqcN5)EmpVU2g
ze!lzfmb5Wlcb@2{J10heve{ADd}f$&6bMWB#o~)Or@o!^;dY+s4_kDSADK_^+*y?l
z-jqIhv*_#ji37wM8b;opoh>K&s&q9LYH0A$CeaQmfjp$@pYd$DVJZkc$HK*kju*#U
zs05}>TQ2O`RdljNSK}K*FVtKpRS8U6w}<>*sa;Xa!l{T36bEXn1g2#c-&UL59C?a`
z^AYVNcG6b~HSNf>Hj3lDS-2F@E@BsbmB6%J`@zZSk*&@wOhxoi@lX?$z_ja<B`+Vo
z9bv)3ctnSYLkv{{(;i8$()6w#R<Q63M7xRI3{?Wt1G~OHdAd`9Mm`GBLE<2NmB6&8
z>CFLcFH$FP8sz}-PXO}-X#M*K-xKh%#PUJ&y-h3}gKqH``|GO&rh|7~*_d~tY9|ZF
zBHBi5qo)#>9#-<^b@JX12`o%P^Z@YyZI!@uv`$8O$Jn3CSU4WhwqjcomB92E>G;Iw
z7kxu@HNHf2j7f}4B``htTH}J#d!znjVFIEp#g^JCf$3>yJziK$Skt4cu>jFgCQ%M5
zf$8Zn%SON2c`lcQD-azZ4&YP*(<!^^=h)tO`ZEh>BU&bw8LI@Q=lN%L`Zx4P^fE*%
z#0nFY!1Pxl<0dWf)XgkZB05+ctgjN7PEF<X{(e_5htrsj=<7i9IxzqD556VsS;-<v
zRhloSQ6{WSx@P(SmB4(a@<jfDnY{)qoPcOMv7Nq3U^?sSt`n}KkA<^vBBJfZ_WCM;
z>9SSPiGRkr9bsV_qQk^t1}cH+iXl;8^39KlEF6hwsaUF~5}2-3Zr|7W_-GOfXCm56
zY^JRen66>}necL|)p{1rMYM<5!&)WO^t$Z>`vw1afQ3^KJyJZ<L?tl2F{8sFNz!<b
zg-Z}^A-2#~2~2O={c+5m`WB9b^AH^>4mD5-Om9BgCVTYH+wClzhG@B1uCEf9-n(pk
z+PSJ#ds+AuqMgOg`YM6xeeFhF_Ep7iSeT6HDdH(6DuL<!%)9wbJ{MlHFa^<mVm||w
z!1QUQcJja#<8@d#8qr>2FI|<ubmK1#r7vO~wzBX)h>j3Pn5YD%FH9SMX=$iY2n#18
zdb)VJl}cdx+F_U7;j24`vv3ij9mS5ODuL<SQe}O3OX6}CevRlzaiob#V7jGaxb8})
z<v*}65z&LigLPE`)2+iVpA2YjMCLF9(GFq<J(W<?_Z^d(9BS{gFb>h)VsBlQz;s(k
zUWD1GzG4>6LbR*c)kq~U{gMmZ{6R}x$-+s99wZ(lQ3*`HyL8P<ua0YG;T%M}i`~st
z0@K}3TPr*sPPb*@azvYmP5P+>ra#y$)JX@*FLX6lB6_~&e5p!cI(63UMWbysG(=wP
z&bp?=VjYezfIbB<Q0@BoLg-US<02SX1j7hl4E>8?2*D-Lr-Z)06o!|=VT3P-e&sNj
z;0hR40V4^ngu#_?2*K6Rs~Y-Z%=A33+2$5=5^(%gXm}OMF`<9ldj6o*iYsD{&xAUe
z&<YbAp9OWYXq*k%Y-mmR9LVO-I2UrcP)hhb$mS6qhf=8XSqSBL#aP7*?DlX#MK=l2
z6sw!#tAJ@0aKp~ac#GT5?94p?&1MZSSOcsufg`S#&S#O9<4b{MDe%R&aeODR>;!7<
zaC{fA?4ofGu<il=WG=nHvX|ypL9JELmhc^r>wpdz54h~L{Nj!wH_`asLaVpX5fkR?
z8(XGE?tPAkk$xkm*Zk|UEVrGBtBP;b1q}ZkH2;pQ4o(pmozGGOjt`TD*<xplkbRgV
zK^4HP0$Ad?N+lTq1zJsiL5@EOjZQ)vJQj}s0~-B7u#HaHS2+U=Yaqj?g{BR`s$}>T
zzAFMS!bu5Iu#l0E1#)q9@sa(Tl5l*YByj+CJRDyJ*fL;@2gLD5f$LG=L-2dRz6Zu+
zT&2*d6nYa}B(g0MxnYdof^0;{ddLuz`9N6-uuf$sQs+Ed6`YHva}r8ULIeDL2F+XY
z2f?z|)OG!=doQ3hsRjDAzz7cvr%)oD&r&CjkMxNgh8-ivZw2~WferqW<L)Lqqeg!V
z^t7$eWh?Z-uXQrtsK5D^SBiu|#%NLsEbuoNq#x5VV#5qy4ou5|9pO8HNhh$t4`Pt9
znsm{)8<=(j^~PoR9$?x-<6dCWOW*$<n7#*UJ7D+^!1M!98!yB60rNf}Bk%hN%svA3
z#%B0WK=uiE5WXGC+o2!E&ci1h8uUY+92wjWV7~(l#uMTAoxpx4!8iv+#w{eE)I#nB
z$`k5^<oJAGmJj6QF?G<qj%+rZtM=3REJx+|S&~`i#0YbMLk{pJ5BUh#kH8cUn&b1J
zS03$;E1|p+dJ(=<WLqjyPa0>ynmV6lz8tSy4U~1%*e56^D6v=L_*`I?OIDWSH$d|Z
z&>p`8{{=e#D_QnR+5c(Kjtx7JQB4V%5{1(V|7H0Ls#m#eLLUbN1VrM@6Q|Z_g9v#w
z(QAlth!iV5(E9bI6!DDL{TTiMkURkTICm8AxCbDC;5IOz4Vcq7unhzg+yNvV^!GbJ
zXa@)<{6iplNO+vtBM%~^^<;EPPh|k!yc|DHF>Wl;TLy5*00YV5bpf^u^vC)VLX~G0
z3?R4=+7?1rg7ZYSc_K9qzVXG4H_gRkwHdw+gw}yD{9uN!2R8LUM(`dmZVwnw@KGQ+
z3iJp*2KpZ(u8rXvfKvl-CVT_%Z2*2Wj&1-7f*XNRBj}HD*yHDpF^}Jbp#|LrUAB?%
z9rpazi@iHGE=5tg1{%~rGcvU`&|nRX*TVj5p(WvKp<yjFCwLt+UPnju^|1eXaz8Gt
zv~@l!s5o9KliK6-o#Sm}HuiWSIo?iYXOHU!j<=WD+v75U;~iuU_BajXcvqRLJx=R5
z-c9Cak9RM}yUX0|ancqw_;r=-W`hk#5rF}LkvI%SoyvJJ@7e4+ly1#;neT~{G=zfY
z4<RU(OEts+i{n$RQtc?@kvf>5n?U;}(8H$6@i&3|CUB*3#7&?e_!iK<Mc>~7v|9+q
zWhdHrLd{8;cC2zTHtm?&GiA+L&vg57yj(7K!G#6KyU1N!aBSoFP_s~LT$7*xZx&`v
z0(_)Zq%Da*Gqh&t5;r$nYqlDV^sD@jiF<K5(xP90UnDLwMxMJm_Llpr3pxx+S|tU<
zI2e?)N(yOQ1SLhpSQx$-N{S&BiAtcPgucHNN=ivafg&4}lo3zO@D)%}K~h47uY{6H
z`uSB*Qbp1#T&E*_3Pn1l&px2chghRutmt=fu=nF$ls(l0^Li3laUqYsFI4g5``!lL
z{agPnK5#jYZW3yFa+4xVF&gKk*tMcmK&a@+O`&d~-ej5f0d5~K#JhQp?QbjP&knIj
z{l}rvaiWL0sf(75?%p1*sv4LAs7rQdA!G_^Tm+dS8W%&Rn8qcLDWP#GWJ+mV2AMJ%
zS3stM#+8t%q;VBwst8U|PV>0_X4*?MgC*umWO(1;9tH~BLIZ<nLm8xu!K=dY35taA
z#D&)Zr#j*Zv0{<dg(e5}h{r(Lt&X4bsuCR||B6XtIOM4<>VPBGEywQxUVA_Y!AF7h
zQQ$)GF<^a+VB9!C@mlDbQ19CWl)KgM<M@e+iB!DK17>-2yxs`SH<Eab{Q{cAe|s(D
zvAH5HSENp%7N-tv)`c%mp<s)ukW!)s1{ET{oj{#C3Avqb7mZOB(gW0yjp0!h(o6HG
z3Rwm1umLeVszN$Qy|HveWRq)x?-;azF%dCe;0=|Qv|v$wj^}oi8y^P7hk*%J8^a#~
z#zzRopQ9NHogy;7M^Nh#aRkd6|7u^(T)T{xsS#>7LJ1xZ$6tWj7ifG5YG0!9uTc9}
zg1<Tbv+a~P0|gpyHL$4$u6O_$-yF+6d;hQF!bHZ!Ou-Y$G(Vtaa9~6rn#cj5e*hTZ
z`x*Wq&_75p4o@ga6na`@B63%`5Ang7z#$WO;)ig2H(<Nz$|M_lWy2scyCNtrf&&RI
z5ZM-xN+oko?ELr**Usy5{B~%(ofr<vjG^%k8t;T=JE0SK&`ub*6AmYM7c|~Q-~T<d
z{GLv5zK4qM;dpX?BlKycy8}PNpr2t3;eUZcet}~MJ_iHNK?T9*q5pXpP4ER6askF-
zoVCW^stNGvD8-LbM2*ERl;fin(PMF;k8=Q|a-mO1soYOKgs40VIAqZ(?*VKNt@0e`
zl|!q%7}^y>cRUh~FBI7p5|!t@8fF=ArPv(V$pHNUCODnR@6clJ9CT|(A1kaBq1gKr
z#fY+>mwoa_j`$hSdIt2!W1j=9=QMr+v|iBo70`M`<JUmzHI3f@tv59O2Wb66<F`QT
zEsZZhqe~=`a{RB*=vRV^_&*w^L|k@7<DVWnJr0Na;*5ZG-!JS;MjtOpSTY6s=aO&C
z@^YtM>_7o=ALzf2<fk}=Kw(Vi&5|WiUIfaUlo$NG3xV@Kl(GvwTk30ar3dl%@rrmV
zWnTpxR?#VYCty42lsyxAX3{Bp0kkckEjm|Zn@du5+!{v4B=m*J5M(l$J&s2lZ!5F4
z$5|N1JIWmGaetHJon%h-xP8I#&N63v+_mL+51EHOZZB8fj5|L2_KhrL+^x{Gl`I0s
zXNaU3B6Zhib=UMMnR=GDkk@O4o~?9q?jH2NM{mCjk$;AWYSv|m!m~u`?S$Wq>OP^X
zjVFgSi=3~}&Bh;d{1zzPLc9)cY@<aLy4vJ6F9WasxOdF)gA4}s$1T(PXD(wJO}iS<
z6wg4@GbAJ9_@AKFPsGuDH*W+JGdA}D@~>?mx(y7+A&=o70*i;h3S*oDp+N|}b213!
zDWE)0SynwXuP6PjO)rg|+yA!fMaggtw5cHkYV1LqH8fre?bkwQvO#O1Z7m6O9KQ}a
ztfTE`J+xm>-~Sf+z9oJH`yZrRp_xyOcdBv@R{WOygxZcJr!7^zQR_^ia-ra~&Z2QP
zlx9OuGP)cn&7pBFw8^Er33*VONBG^#PHK*szbQ(@@g@=z6Wp8T_++hQeZ1K@K1C}<
zjj-*aJt-VFkO{c>xeOs?M3N+lYQ&WTww!najz0!mkI{{|4}kqZ``$9>Q3k0lXtBt)
znE2X#&JW`!y&mhPDYOhC9)L(}41%#p+DOSSV8=G#O!yAq)&V?dJiG&pz<6I+bIr5U
zk7bOI&=0C79U}pU*3<QR1Bhz?U*Lc()ZbGZz%-2a{dAehANbk|+4((aa*r;PvqaWe
zBtgR+6_n8mr!z#GUov0zCkt>2n4KckJ;&dI4!1~=j_WJ2&S%Fr93NyHWP=+<93N~P
zY=hHLROd-UUGV1N_(_UMaU^A414gd_6Y*sBhm8$6{Pe&9<m~+e{D<N$&;EO%X|)4J
z&ruZ<K?QIWF~UL^R7m3@7+wV92wx0Ci{V&;OJGn5eSay8DuojWUk*dc>7=0o4y%9(
zgs+4Vl`x**YUo!@l8ytDfA`*0tY?T8;{niq0Ce$kb9@KT?jZQUlQE}jUoD)6CfW=g
zn@K15Sk&red&3|7j^bJuly||A_*JI|D`#pB+w>GY+CpyOjy)oZiX)fA<EV&TA#zc|
z85&t*rH;}7Z+YQlAgU9npNtRJ(3gahHS(oMyU0PL2Hyhow~!t%HW=hwg!4A?rOVLj
zG8~LYeW@U|_uY+~)6qO{L&Mv|*)`j($j=Oo(?io~1EbqO9MRN6VD}I>VvK_uvN+-7
zj=V^@M43xD-5tPZ2Z+LkB=li@cG7qc2;Bol6Miq?_L7G5Z`LO@|8TWwLgRb~E#8sn
zfaj0qDVz*aum24wAFAKqVshr=@-Um_Dktwy16{BJ1SjvXhQ@25%UY;zvZ97ObgG5w
z&cEQ~-PRGe$?@x<%X<3$w=noE49CxJIbyu)%?B?Zp)v4T*ulq^D4P(@EUAH8D(#3V
zl!B;IFp{XH2Mq54RMOlFqI!v`aQrIhy^2n0I-qX{X(_k#Zc7>z-vm$*SO)ydKmdNQ
z(C`l}r@dM?@b9LJzV{&LJy4(h2@U%pA3y}2rqHkt?gQ$xM27ze0zU$EnR551e)3T#
z$1OrZ{BsA!{4WQF`*bL#2xqmFT9TEEh+4)g##4FT49N^rZ1BQ48(T{HnH(xt*<L#9
zD}lBpbd=5)+2)fh5BFryZWB&}sR0-%&B*||KvWme`#qn@4>AsQK(t#T<PlDI$8!8o
z`A{c(hJ%}qXvKwtWb)*hl9?7{_}RcA8w|n=!y(18eP94qH0s|&?_By2w*op;z(IsB
z5!sfA)b(K72{SLZ*N$1p{ysmGd#bg&3<mBO8grI%OHbUVdYZg9KENpA7nBApU>8V9
zLo>}N%@MmXj$dK4LXDnQ?d{y)dZo^o<I~O3rDWu*fOHjoW|9G<88prW(o7m>0cjSE
zvw<|5#yLQmL*r7wlmZ>{zB0g+(YPEi<uvXDOefGGn(6{f7md3C(@o<Z!1T}<b@zH{
z{2nmxY5V~&A86bMm_8bR1k6Voe*(-W8n;8OcG8Eyi8AtR!Ve3GxePWKY=*tZ>x1cW
zTEBY<M}6^u@zFSL;~qBJMZ&KR$aiN;X4?`|OSeg<PR7mx?kou)96DlSKanFhbp8!p
zs$eKlUpv%mCv6{e#455a6RFR$aRP~kApEj{3_<xDRNhuMPSI)ABdB~#<1VP|rg4Kv
z*(f4APkBS6yh-CWk+Pk}R~Y402H*ewH>=vOSFKM#_TK=^8h|-Am-p|tof&wy<}KQ?
zQSMPuRR6#|Itup>u=%4&2*1A|lPCk$$bsT#amahxBL_-wH?T$yl<4X`kRb<3@CV?C
z94Ntkz!o`Bf<FQq<Uld*8(H^w@oy8)p-p)khTJC6MaU_=?vZegiULtUmPq}ezMq;M
z)&(#7_W!Y)HBx8IL%)<kzrRjRT69OaC;)=|@4rmyx{$igqOQ~N%^DiH)W><$bqRHy
zL0zw+u5+mCBI-Juy3VJri^;W!5Bl*}t1v!!oL(^~0YxA^3#<Z}AOmEBJdh6xKq1Hh
zXmCmMXU;=E&C!UN6)R#i#r?FjSx#49Pogl0jBqf}Q5c#U^&ep79}pBA5*ilnVeA@c
x=j$Z>zdkjwu(Yy4mA~B4*~M+3r`I4KKkLEf?%qR<O`?XE0yu2kkZ3UCe*h>&dcyz!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp.AC32388E298FFA30.idx b/.cache/clangd/index/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp.AC32388E298FFA30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..123bb480382606005d78065014426e86110e882d
GIT binary patch
literal 8538
zcmb7KcU)A*x8HNQEXZCLmcFpE6n8006-1hiMJy;yR4mwyif}~`m0~m?ii#CAViGZ8
zY%EQsNLN7w35YGmo}eUAtk{wm^PAoGhPl4a`{O-7ACL3pJ9lQznVC}tBgY2^N80ds
z%7x?C#4lPokLK}sBJ?+DW#ZC21V-xfcx%3Go^~g)+sEK-XYaK=hw125UlgvnWD<E<
zrrrMfc;u0aqAkuTGGCj&-`+X(o%#88ty1?Xj(48eCFj?VpAa}MFniME__KrWZvN;T
zB5in^5-9I(xb-I0r~2DdU2zGi)*t?g3z#XHxZh>_4k3NaUMJ%>r=FkQM<fi-+g_jF
z-SWQ1apv9o?Y9y<Mp_<PU$*tn3>Vi;?@xboN8w!?*t%o=)>l=~<5j7B<!(}I-HdBj
z*+vhx>VE3u6QyI84xbY{SCPQ?ikUhg<Ydm3y&Hz9*8M%9Xy5jl_+!8H_}J*rT9*E6
zFEz2z%{}smpg!3W>86h9W)GGn+}WwfZ9cakVSnwW>(v=8W7>|73i#c)VVf*O;%KDX
zq1C;2=(gzeKVLsnY;JloM#nHR<B{B<EAZx_R{O0k!qJ8e5u+ak1-2>fo!k}fXAofR
zJbKCVHI2#j+VNSVt$#S3H83stWws>Ki8t-TLno1{&tJDrxm+*VM?S*1s5ScXFWRl`
zQ5!6mRb&+fEPb;odn$k2xj{VzMxO*7yN9mc5+XA&@p<H0INxtZVAg9Zt*}V#8N1Dw
zHJ6{y(($P_kGr++O54EHjW(M%Trw(iSzw;smeai6`_%4`pVy{^MHasaeb|^evp_D}
zwI|Sm)!EP{elvMDbuIJy{9v6-{@b(3``0C=24=|0rYv*3y;f8ym6Z=sEEu*iuFA8k
zX5%|yUx4;0)m3B3iGa?zI#E`eFDCW`dy5{7>H9f{H|9jye2ZWBAGWBJz5C3M+$=Ho
zl%#w(M%k5aP@OD4<2tvi+M=Yjdu)PJQf+&|mbOElZ)+}pt!fcgT`G`8iE3VcKW~qL
z^T(AQ`M-L~Puz1^v9@=}sq0f!2De|yLtLjue?2awFfi-%mw|DCZ95CTjN0gXdap9d
zX?3vT#r(hX4~F*&bGy!sFL`j`M`O>`rk?M{23qA-K3n+m(k{`J0=Rt0VSmre4ZB_!
zzWpc4sAyA2^YElYijC#Bzv_*MtvFZI;@5gLG2yn+qP%dG=@XU1r5APGv-4f|et*^}
zYurX7s}F12dlOFD+-#nd$GH7@bE=J4uj^do{li1kWd(H!*NepE%YqF)SV_Z<-phGq
zGUpug_oWx<3rhOvr1z;0(!9kF9^VpM_|2GVpZn=cpFb+^+GmD(-)fy(MBm+$)UDNW
zDokGg&7(WxuPGk(EPwuPTtMyOTiw4&{Fl28ANxAD;)L6~t6N=z%!=x@pZmT)dwR+~
z`^r_hcV&*3i}anXtZ#bUJ(4w9#$MI>bMo)HHaYL|_k3!*nr4^vF6!v~tcKa28XRKn
z7KSVd4LY#bWN-49mM!T64&VPK_JfhiiAnnVUQl)8)$lS;VYgHI<O}(uPByQM*&<!4
zf1vt7-Os@(zaOn_d42Al>yTTUB@e~oPPca%xuZe{N%~$D`t~k+n3)|ldZ|f8)FbPu
zj9aBN?wHG4O7CffH128Z(GMy6N@FKZe7xySlg{m~PkDKN?de=9r)PT<y)(G<P`7if
z<r78c?0d_{+I@^@I(YYB-qaDbV@-m=ZDzB7lzHubzmF}M({hgppPVx}`LJ#9zwk}?
z){nfTl=%5}$V1sJG*EN~wI>UxnLHlr!S@hq1oB~;KZ@k!7d>x6ECtk5MEmf43^fAN
zt_xnS*1u9dPC!jTw1TfN)Cf$wy-%sBe>W~qK!qaOo9}I)5tw#2yKdh7JaIb1TMvjo
z0Q3RS{qKKpAKw?-B)@h%I4qzdP!}m*YM>FA4mjC)a8q^e2>}&}Xgj{0zD8hrq-{*q
z1)rEV19^)O?QiODqY;=Mv+eb(*y4}T0%|^@&H3g7Gy>CO^}gEqIHGBXfQm-6j4w0Q
z2u#oFy1>>K2j3G=(-AG<O9p5JrWa*|yT94e`a(d>MYJd1Q>YP`UNS4~rvkHWj|5aW
zq6hN_i!=h$@u$27d~}gN5m56G?Z|gD)d)-{XqmL@@@E$asL6;P#vf*&5tv?|_bM|o
z+Cesu7mw%(+7qN2f$5CM>g=+Ge~ARt3`EQMas!RPbmo<l)rzpHQ37ffqV4(i1{#6s
zoEy;_Zk#&QK9CoO=wR(&sYYOW*XUrFb^X(P0W}HHwtQQCjllHo$YUk%ey?00pcWun
z%oh*P2u$xgaLi37psq|n%|`SH{s>cz!1Ted9?KU9FPsukv52<hTMp0&Oy|G6`pMp)
z1D(e&5bex&*4GG37gToJ{MPs8n1GstXa~N7fkt4ulztoEKJ*gWv89N%;#-L{0@LL;
zQ-96-*HK?UEk?92-`7ARFkK%xAl5Bynx23PL$n*;ZJ<VAy30aw;LPP7Ljkn}(O!J7
zff|A7`w#6Gw9E4z3aBVVyYt-#Y6Pachi?iLhyIfzpcW$9iSJ~j5t#0s{v^dJqN|4C
ztwVIQcC@)hVEP5)SMX7nzgs}XAbJRYh)^Ri{o(2TU9LTIZ3WaaM4R$Wbu<Fgyu?Lc
zBq_{zJS`8wiTw8YIXVnm4MVHJum1wGH88Y>i)+EV7W@fc2R?N$oZxyGSr0)3H$XrG
zj3&4dhBlJ%gmXHk(eqB}1H(QA;Zv}|9SzP}#e41@*U4wt)1Y@624R9>&w$<;E-r@w
z<scz^1qdp@ir`9MDnUwc6$q*bZ*s)BNwBa8&<Zj@lmUaV#m)ZOzTbc3;o*pQ1cr}5
zhDR{}(Id6>_~`41$fQg%ss5p*!moTc>$d-bh#oNN0c-q)wcd_|EOWzs6wQW@2p^5_
zFwI7JNBLr8Q=Pb^cYVM{pwZXD9uVQLZSE%?(6RdK5*l^`m}~%Zyh7~ZDZS4;JHvWf
zdfH(pC09JCo|SRg8ht%t*oe9K7CKlQG$^>?u}qI<QwF8j;vs1^)jYM|1T<R+riEaM
zS3<KYFjaAJ5ttQmaWR+`lf9wY60j&Cn@Y2#ATA|)L9=CGTE_jp4=nm%FrLrx=%V3y
z>rT%^c1tr)vmuL2H%}jo5e{Uu-e&=fVU<mwyaHHZhMgy#C&M2x>}9aH401e^BJbmY
zlifG|KvsGN%rnr(gmcXbVcdX;W3*_t2)v767$(%Z1(cAJuhz}4l-vv)Ky~y!3!G?K
zyFmFJXIZo5vxksntq5H)gP>9nl!6(al3}aBql!$)uniz<01wRLpscO;SujRZ7J{;z
zGvyrd96K_}vY=(51hs-iE7<`&H0b@0=<PMUN$*z53}a-lKY;!N^l(oc*3qF<$8_v!
z7}m$cM~e3f2X|x~bz~=<lzvJj*|8b&8E#}{OM;e6B1i=S6&T}L)n38<4EPdU1GY7w
zz&PMq)V0?gIT2_Zw}Ssx2*3l;>^8952K}drX19ZAI~Zaf=NU*Hb&|nZX*VdJ_uF*j
zvm4Kgj~`rx+|_vyod+?Vm|-t~=mHlvfngI^5WX1<o4NQRm|P@Rh+$j6xP`d4z@BY3
zirScHT75UY3%Tb}1>PzyE(V`sViEOS^DZMi&X5M^eU=$9tgYBqf+Ib{+KKHXc()i<
zE|yDhiovk<VtWbBC>YjJ>?pyx4Z}K#og_H4F|0zYkl+i!u+Cy<36AH%wOP;OdcVlm
zX4qh-V0WB>AmkrBnxGIS#1K15h7B1MV#h(QAzm2a#13tjI;-PsSEO=2K3Eg-8g_JK
z_70>N)?tu?EzTW~-!$>Gz<zV$uK3u9qDPZ82Ym1!;DvW=(y-mjO<i(!BErhYN{Lg8
zNf)m~+;Dk$NtZ?;R9H=Rltv*`Si{A&Agtv&y*d!qadABe>q&s3(JdD?kP|_(jUa5~
zj`swFPssN;c||5rXRgEq?jm<%oQz;CQ73i&N;)Ye%E8!9IFU8b`z(P){$z|Y9Cu>a
zMdC&M*N0)31}zQ4mxp0vgJL--DKu$1;Z}yO<e*eRswv^p1Zf<UE=cE~Y|zaHBhoh;
z?6ZO6Lbn3570d{~6&$yM7r{Be<Z$^MaLgh6(!``yVIPi-(p9HIbTtgb`_7<!q-(gi
z7U)_ot^>M`i|c`|=i&yS8@RX;=teGn0`wDtmqpHTz50623v|vxEkbSZ3CC9qIeB&M
zz}dM3<@$bq#IValm(M0TrwCFk3EB?w?cj|ut~k($>T-iKVihRQbB3HNp34avaY1pM
zu+at<Z6s{0STmwi(suP96u9?6=sxH_18O{VA56n@uL=ojS44Y<BPIO<{iCqAU3<34
zHfFMD9QyG(**bUJ8`}V_R9*Ryl@@{oc|)>ODv%&=NN_P&A#X@<8Ay;f#29-tzTRiW
z67mjq4lcwy%#_dUUkQV<0YMR%;2{{c9Nfx@w`JH`u&V_Z%x4r$jGyvdR|`5-UEthB
z(gcRR3eH!#_!_ug<Hq=F5Oxh_<AE6VI=Eklkp$lWw;M2m;G5ullRMs1@P7&uFpvE;
z(p6pl5feJNIrtJ?XUS)AbS(x!F;~|L@TlNUQ3rN)MAuC($4GoTb1cy2nH!j!;y|-?
zg4Tv-JO1@SCddcfe3D`@>><!S#Kqr(?)O|=0J;TSd<1llaPbeI`vVsr1>K`ud<=At
z5uEdt?}5S<?_$*#D>^g=A1WMdP#&qSe2Ha>K!Q>~Y(NI(eYPm|Be)c-QR;{B_O{n|
zuRVY935~uN-1m}91IJX9^QcQ_GQQGH=|jBfboq2n&a*gZF(>CK1VJG;=Q$1Tr@3dY
z8f>dc&V%c5Expg`aE2W$9xTDWond8SnFNQ;z3=D7Wavw7pv6A}i)SQ1#_L1zLS4C&
z#dm;72Putk5sxmYx`HR&+(d51xQu3450Qs4E?yXRh-ip0u45V2OXOvYD{7o1Ad{&p
zelnagAC$*Ao+;Wo+7UYmhD{7j>~}B>N_zxlT%$FDZ6h%f!}fuGADH9qJN&}L@zE2z
zzfdK79<0w3)zpcc^#v|&0_!F&ZU)<Cu*U;2>_w1Y<WAHAgImCn@R^i6lj_e<@!dlf
zRzF{GbX}-qvH6c=N9R3Gu9s-4Vyn~OMn*dWR%f`l90rwxE8#0ZS^@3^SAumVDIL_M
zg|v$BN43RaZ&ZI6qJz*0(oS%~AF9^WXotri4n~Bj(9{$^FsN!KeVBA*g+3}?w}Hhr
z=npok29d`}vxtAS7&gT&Wf<9{RI5}@%<cnOA34b-BRlr>*8gsU!bdjvWrIKdjAnBn
zAP0tHjFT_q&D4)N#Ax0o-qyGrV%T9O!>n<XV_07kUu*p0!?5u|@%_<B{R9w}0sRG7
z>BxxT<xk7Xkg`%AX&^3bN*|0^oo61lMDtqWUk#zyIO-I^zlMuzA)pq*2ww*y>L7yP
zdKgm=(F8X@U;|7cxDost$#{6x=z6IisffDD>!7^PxexQj^Er_^IW##MTa01b!J?f+
z>Z;(qYl=txb_;D>7Uh*iMd7)tZ=c*7x%PzmU8D;PyGTA-?-jXFd(7ddD0{Y)S-RjW
zkDTeG&?(rN;=7J~kNUlgvr;{!2%lyKl`z4DFbPlDlq+lf>6c7(;8^azyx;F#x|k6C
zepX>E`f)C}<$@p91-l6}iTc%zGl^X3LMADL;4(7FO{*I1pB2{aXtXTKEQ{)Yr1|;Z
zJ^S;&v!95FufgJL@(^^}wBb|gNb6LMOOT%@cR}qE<QKTO3FJ*&+zbxQ;EeTCy9DQp
z<dUgtJEs=zCC#LqGpYVdij#GupZeL5qo0w|g6OA$U=`8tVfXq`yF{gTkQ)9|{paG0
z3fE%Dm8#zuIb&sjatG(=#K>bjv4I&jAv9qoK_wt4A+C~PD`7|_Hx;P|c|CYyb)V|0
zc2It$>Ox5(lepu$<Y{d26knsjX5<mZ3dTxt*k#yoqi}nCei$~&D5@Vl%`JZSouV_}
zh+(6}(NeOY>C)+(C!HD685}fII+KHDNoR3Tj5LOWW=m&tPzs%*M`lQ+Q#mM&PUE0-
zI-P?Gfi5J^APjmcqg7m71auJ>7Xw|)#U((OaB(TnrCeMFbQu@-fo>lNv9a-rQTR|l
z5|jOX1G;ZWafstAI#TM#V$#i1=IMfqLWV7d@x{bVy*jmSvhFPpOSF-Ty%#I-RM_{U
z7X<aoG8s!*ODXFpazK?oQpyf4zDFtVbMbL4<q0iL|Fc@kYA$ZoQnqn%1+A>4IsNa^
z%KQEJpT-XPfq*sNq8RyqJzxKy*o5B9<Ipc|(EkKTT#(eDZmdv?|N4`H8$4brX(p29
za?(sD%~hninl#svW&&xhB+WF^Od`!Sq?z8|)M9rYirp#OS(l@qzBX-nV#11~)oW5x
zS0=CGEr?$jhkotBo3Lo27Ol<K(H$UQ1{&xK$BBXh<s!XthGs^_=3*aT|A65mMvihd
zQTW+;J4pYxpPE`q23gzK4wl(FIyt+!dw32VX60q!GDK-&8a)0>YRt6JV<Ck1A1xUQ
A<^TWy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_relu_xdl_fp16.cpp.74D4B1F74A45E199.idx b/.cache/clangd/index/gemm_bias_relu_xdl_fp16.cpp.74D4B1F74A45E199.idx
new file mode 100755
index 0000000000000000000000000000000000000000..276fd764f70ad7fb3188c203b6c2065fc6d959bb
GIT binary patch
literal 7158
zcmai(2UJwY{>S%sU6+NuEFfWND<CYWbdVxVfu$D_u%OuWswe~n6f7?h)S!YQBoI-t
zV5f;-ED*tj5+De+D4JN3pd^YIHGuMnzS;f%hZ)a#=RD5gcjm))?%cU^XMW`p6&w)I
zXUOBpmj$OKa4C!VJRVOA{U)a*#mz@Bq|W1|9lSDYPU7z_nokc;yuQR|(zKf?uHHtQ
z`wXrvmd)voJkpx$w=Xj!>fD2ckB-WI-dP=cBv{*f@8n4b3lnW~w?+A$`7!y$rQ^S3
zy)JIKyKv9sqrcxce6M?>{kI2~HN<`B-SX<4zE_cP{JxHeJsNr|PN`(~SS-IgxO1)7
zA-4YWrl!H#V3+zUh0C|T>n+a7w*I%vffa9l{O+H>pAJ|W&KuVgn2~ra(rVwioBE5m
z$9XFKEtVIV!)t%+w>W)u+oXKa=A+|cGcF#vS3GO3Be%W6{9E4-{rl(Icn%)8vi*Rz
zukocjZ<`+o;v02h&Low*V^V^S4|be-Y4$eUI^Sl@xjjZ%gL^#^)1sng`5SomT)MEw
zc5jqR)GsP76IR4tc>DLCKd#un#W_SpIyb%B#q!3^gZ^Jfd^I-DBrfL6gb>SdJ>T@T
z2c7m58bw<BtnAz1cjdb3{N2Gv@>83eiriw%noqMw0(R6%ZoI41S)8J`J2KNr{YHvy
zwA~N7CZgErPm3Pt%~rX8|3T>aDsT0{m>jP+iGS^xl63ac4CVwA+?VB^e{9sU8Nz*O
z`>n3$ot$S?p<`*(;?VfVZznVMBsw^T1v7q!E9d9m^q#qRti|~!Gv~Ga<IdloeBvUj
z0mEi<qsI75-1h6mjF?SXYeQC5dJ1i3o_~~h;%JJ|^N7xAaVbu_S2g(8+r1t*R(xlD
z+=I58I)g6$cfWtw#&o9JKJ!$Gcz#vXrS@!#ceF>c)z8_9)$ig1`O$H``wA{Ib;olY
z!b3x%bZ58eN5T7&@6_VQoiebRFYG#Bm+Sg?t=1Jw&$DgIiZAGey?Jmq)aAU2ZCkf$
z`0<IGS88eP+I_flh4$uCrt3Zry&Vx9pnHF(`K}9>{>;cSA29#^X4~nswSK3~&hIk6
zAGkDT((FxNg*>r~8N1owLv{2Qw-$@;Onou7|F0Fh@@s>1{=N0O^M(h9**j7#@NjQc
zv$u$yVC1M#^S*Q7vsF92##euHd(MLX&W*(-!iKs<4^`!#zbyKDPnYkUStTJtN4+;K
zFQe5}wAIgy`qa|#Qn%skKu66bb_K6pps%i{XW6AM)dx!&rmb^2)LVGWs&3oFj*zb3
zIzRhDV)ZI9_iw5GgZ!==2J0UNPW$rJ9)na~^Ixa=zs#6qX!=95?8^rQ!57VoPDD#D
zpZ()y1>^H{^6kMBsp}&1VlSGe$^x(VWdAEHf5D$+9wqx?e!FG$L0A4~P|}oSo%28N
ztxEQ{TAWA_!keGic(eDsvufJ+%YQy-Fsxc~JEW?$SeOvInfKMSlLu@!MtpvK+Vz10
z<ukVYQ*beD%B80=er$n?r|r$No}1ExO-^2!JoDEg``eLmI<8?~nH1<RIQ#Hk#NF9~
z>^BEe2WHl2*53Z?;CSnT@R^%a55M)D>h76*>);e~t8G_9_uK;07skJD=w6ciM_>Ct
zNuG_FN5dW;%{uvL(XXxQ_O~n>FYf!^^SABaef!<ErXTZvi1je4X<Wkvd(7zgFl6{n
zC3lZ%8ZUW6!jeB29?xpsx|dJ?4mA`pvv@qtL*Ow+X^<aM{@4v}sI40aDikr(knARK
z(^MKPdzN$_%B=bFjEIRwvaP^YLus(=zx(y8RTaZAB4#0y^#ppVN`vJ<we-DDA{%o=
zObn9k1@;o9!SdXr>Q^@on|)B{Ek?4J#LGx&usrWZ!>4r>0S`pX93)ExQdOnF^2&Le
z?jF<K-7R9KBiT-1r=>JlUNcIkT_{*^OvKDUvb(@tQ)#fAoY=dfW{parCX*V)<IMTy
ztkPh4op<Q*q4ER80wxk&ZK`4_RvIkt=$!aX(1iT&Sl)Uh$MfTbN`vKGp8%M5`~4CT
z6OLpXfsKaJU^y@9)Y*YQE-n=@i;%1@&{tI&EEnuQWvAkK<(!C_hvcb(sS>5Z@}cx6
z7E8qqRU&37k_`ohs!D_9N`8MryK~D+5wirzMgk)(rNMGl(3q&l-SQF<GZ)DY0tXGH
z!E#-c>MFZUv(-dQ1d^QuPU=d7<(5Ul)p@SpuNE<}NOlpps4ESY?>`Av-#lUCc@eV!
z$*uxdb)~`b!{BSRZXFF5M9h37I|>{%lx9SJJbr1rMd4!+GZV?q0%vul!SdVZ4|8pw
zE;1D{%aAM)NK}*t%SqhwWN&RAZ<K@R%;EMW*+Q191X(3a9C=sVIgp*Bxe8=eG*^Rt
zH8_+04dC1W?t~k`xsm23aBiZx8RX3ng!>CR>Yw*oR;rC+xyxXB875(Y<!ZpRhUQu@
zttG7f&FT6Hqt{#kmXq;iEPg=E%SH*4@`E}Bit)skVH`c4xQ6Ch5ZBUN2jV)K>p@&k
z^HmUEC9HGYx>2;O9MJuDF*>`LksIk<bJ*B=+UGVBhCpWsjPVSN#p!N4h3&s0p$D{k
zzyu2>YB_7R>y0U3_*^mg7K1+)_}ocwKM7u#$JHdQ?A_w|1@O@b;vUc%5gsP(S1}rD
zK|^OVpJX$-SeRgNNb}fF4yy%xt^_0{V2A}iR|=9+n#({^Msqnx%3(a|UjdQ|vJd#&
zFi3~Va$pBKO6?=Z$#P-xun2rsS*`;NI=})CYS*4$q0`#95Ka3TjD1E{1p8aQ+DCqt
z<wE2k;bgRSFlYxE9*yNPnCThJT+Hs8-xqyTnEIa2a`_<42Qd~{t^kAuG=B}kuW8-~
z!hM9@jb62_7A$@x<fFqPt^`dy5FZ^D@j05SKwL$0HHfQe?gw%I$86CL6Ucb|;L%Us
zH=pZ=n0^>}A{=LgY9GZHmh+eUhmmC|1yLz{f(KO`B&QlOD9hD@Q$5X%VA=>)q<<ew
z?IS0Y<vzgl4=@Mop0{S)dfky7If~EaG4^?k0~QqQmB&n>IV6t>#XR-JPkj}q53NT>
zy#=&dKpzV%*9uy#G`E3P8)2XHBZu9DXC5jyu~P|gK8j83R7!IhIF->{4&LS9gKxy=
zD!{h_<b;R8eHc72<3KG?`zS!OoQ=MX6o)I8v(>ki;#a|P_WJfx9Ku=7LEk}&qX)}5
z>N`rYKM(kJdtZF-^25k86@g|ENbp@1Jg$fwH3d5s!DOQMf@&{l&@Al*Q^I|q+DCLe
z8Ak1+Jc8u{<$=+7mxL77=X7lCRcBbvH_(^bJVAj$BNTR^+h$7t*QMxudYgLt;%mcQ
z^AaUC*?ZBSjV2q(aV!mQz8?9L&0j5iKGy)^1~9-A<a3Q6Zlt*h#7#6egSeS6&UnxR
zDAFD5+F34K9v+G(f%6{JNs;uBPR^Rn+7JIfNG?2c=0a(d+X|ND(!n?#%<z<uuK?o=
znzw@SR+=-xIFsfb$dS{$6O4D#oCR80phL#b2CZzGcZ1e$nsY!mhqzmo+Y7pTY0d@R
zT*5ew8?E+H`o?kr@_=c0Nm%5QHHOJLvfLJC`W7aZOe&q3o=!L}DS3Uw+f&}iK{f!t
z0n|yyM&LKn+ywk4nwx>&On7zFLff0K7rsOr;s|_lgiIT!m&lJRa!c}(Lgk^9AHN0$
z*JwY!m6^Vk`0+K{Z`T>FKQs}IehbFlBJL7rrwpUaN^v%-@Iyv<%t+kDe@8n-k)o1G
zc<6XYaok!T>ep@+<1!uH#A~WoB;JB)4Zn78<lnf4j!75Dx=8$Exkn&-MEGBq8%^g&
zYXza}GZ|SXV~5u=qdY8OX3>ow&?4Uf>pR5Ru-sj+zDsi#1a`qJd?Cx-1M7Q4&-}|@
z>eiWUFq-A`H1#CdSD}nt+gu+n1)>S&wiJ1p%L$DiGhPo9z742-l$2+MjN0}h=fh+b
zmP-~UQ$8m}m_pGy;W~;^g{c&!3DYRrAlyLFBd~o${Bw4?+y2tE1FMw#Ch7>x#*@On
z76m#*T|i!$pWJW6JFzIG5tYz^?lL%CrUP3YIMvbI0HzHe2RgGpu*NrKx%V*rJqdKU
z073gkQGJj}*y-Cz@qx^LQM38-Xn2ht;R%d+0_J$M{J}-@Gc=?>A>kQVJp)%P;M0nR
zRa7lx*c+g8gItdb7<6<L)eGt7r0JxM3lNrb(R9(q^#;qiYPxFU^5STGC_g0fi{C~o
zvKGrq@|=pS#qu1@RbW{~b2YeBlla5u20%POvX<hPIupB}nGK;G-2x^pWXg&-ZqiC~
z8<@1wJOoZd;D%S?_>P#uC+q5@C?Ti=(>jvdD^e8GdYZ3-=~cp|vp0HtbvN_?TH8A?
z<_;L)5zEqQM^8^E4nTrLERo=1=<LL&Gh<&xTBDSu9}N0Iiu)_rp`Q$_;E;X@!#WOa
z$n7d>MDh~cblgmElFD-KI_@Ue)v??*X8JbbWEJH>^e$$>$dHewq!#K$tW-K(j|P~A
z1r|j~k4BoCz@v%gW(aPEdAR?%d0mdh*&1VzKd1!NO5%-K^nO*((Od<pRWw(FYBkOM
zpxRFur!OcAQdGX=p@YJMX5rn-a%CVYqYqy#IMvdJ?J783rMU@Ao5<n26p)`*;oZ~4
z@=;`#r4y&lN0C{ULGxCSZKXLAWSKPYK%s-?ogmvua~4RlNKoW+*&xZLc{fOQ)0_j+
z98!w#QDm0xr8yU*xr8q*f86e;)~2ZHWa=_qyjpcmQOidA7e7aV!_eH&2A^g{fDAkW
z5%?wIgN({1#b*S3HOjZ--XUbIOTe!LMiOa!-jN4Ud|V(G`G)(>z!!CxZQ7ABC0awd
zJsc~+1>aJ!Jsi)`Tm_C*G*`ovYValf2S74F7U+&q?dW?}Chh2f3mAg}W+bP&FR6dO
zajMBirC+dU0ZTGMD_FGB+y)kHG!KFM5R4QK*n^=BsQ5ZUo=+~9Q;SjtQ_4uK%yQ*0
zy_|#)7UlX;6~vn>lB1Aeh{TindHa*Qb-$0ULXS}h!a`uNfb(?~wU55Luw1Y_n2J5+
zAS$P~e+7stXl@46X0qK`?j@MLBtBNbQ(uxI3n#2-A1FTQkVp2B`%sTu3Vx+zA3Xkj
zi+8S8<<IC5yrRA4<DBKWuq=mhuT-SxQgf+2PL!Xos+j4f?c0QQWHIO#({nosx+e)g
z|FUAB$f{FOsFVP|1k~`#u&CVPm(pAY{4$!$fnQE@1@J3q9tPnssjFXAeI70Baxg>_
zeGk6x$*I788-;(x7c8<GbL4aU@CIkOx$?Ob#mZwTS|ndY(PH^xik8TiP!uPR8=-%$
zc3AB9Ogs2L*C>gLM_*&nf2op|Cf6$(i3;<7KWlJ<$BQG)<)pcCq&bQ^ZS-eQgns;E
z4fOx6waepHtX!4wY2upI@x1tvPP_!tOd-t`q`8_jlSp$FX?{wY$)vfKG?ykUi&qR7
z!iA0Ej~1v1RYk12riM63%g@(BOD$-OuJ$K-`jg!}JiVq&^|sZqayN6CVEo_yDlwFf
nH5q3rGao<E(%R16VUn|}k&}Uqqg+Q45WJ4@pY0O}p}hY9+j<qH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bias_softmax_gemm_permute.cpp.A515B9EF146522D4.idx b/.cache/clangd/index/gemm_bias_softmax_gemm_permute.cpp.A515B9EF146522D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ec5340700df455db9e6530a6a1ea76e9a3f02c45
GIT binary patch
literal 10794
zcmbuFd0fre|Htn+-L6mfe%f!h?Y7-QskE<Jw5o*443T{syAN3+v`P_$LCpAChVUCZ
zk!({UsZ>lU)u?8WEwW^(_?>&dbM85QkKccKJe)pHuk+c?d4JyL{eGWf!b3x4gLu54
z8Q}>FW-gt|=ka*j=wJNOC3BJx*lF^33BTSL+qAIBQ>(7}#pr;0BhSpaG}LtUh)wRN
zw;a6n+tAULS&k0lx4xNqZSDH4^BYH=`enp@`ze|KNS8dW9C5%hP!<@uJL0-diCy)g
zGfUrItdKbyt=&}ip|AMDVxPQsCr0u2Zm!?<sXU{+_f2~+w`Q5lbBxoKBw=b#ZNlq5
z$-swZWnH~f-Lrb$J{_T{y0~|ycg~ZCpGR$Lygsq{T2T4CKX$gA66Cvj2xoW3mi!u9
zTXL-UW~9r-WhK!If0!zb3(MR&#VxGJdR^p^nh879F8|%zGb-Zrsmg_M>Srr5eOzq1
zvM%Y(&T*;93d@=y${lQX+|51P_Eg+n(VlfTZpKCa6+idbw6@%u65W2zH!ix*NU`1T
zTfAo99>*fdTTL^+Bw0%AvX0<DVOU;wsz+|^>^H5aiu+rY9$)vwE8KkMq4|GWMa8yO
zTK>x0S?QefWP-|WowT9Cc|Uh0E;|*ld6M-Xx_6qkCQfhB2z*sz+<)Z#(v@|2Y{84k
zxs^-9^W0bc9{Duv_=S*zHABT`^K!;~)Q=4sQz_iRY_=}ze|DVV-CC`f5--*#T=iCZ
zdaP)Gs;N(LZOEeD^U0wnYF*fH^I(nAd}gY-?4Cene1&j_)yrnDt<DxJph)9|&@GD}
zu6A(hie!i5q5irD9DT#=XPmqLd-P7@jh_bIA6sgyGrnN!ffoJ6iS|biO`3bQen!<X
zjlb&7t;m~hlN+)te@gh%Siz8eOT%1$cDWaB(lyaFXL&-@y3!3t*Ug(U{LGvEae=u;
z*8(fKN3GR5u4OvT2kzc2Ob(f6ds&vFyC`JVxGnwk4u$^oKc}n4#vHeI{7=8(M`TOe
z;;XD>u^*qV_P+U3Sj^k=rEm9F+zK?h8a%t~7dwN=(<aL7ZGL|3(UO^$DfS!IcZ7|f
z6|r%)?1j4h9`%<=YeG&BSsI!+cF)3olk@W~{*rq-N4qijRpPed3=7MBW6l0%f8L;<
zda+^GpSI_Z_E(MlaoWMJzkl(Vm;aSrkq{j$u2?_*y56TN@r_UKuB&m?SA7zlRO*u`
zJldJlKRs}lUi99#ojkqs2Exs)f7x8#t69|VUU%ib`x_FECp4IB(j3#{dtIf?B(8+t
zQ~Pq-8wc=o%H4-rKX1;;$nomVO~29@J$x<Sws`88$bz@*bAO>(hgx@+{j-&J$-`Du
z+Mjm&{nMJGp@A*#f8U>VdEU*QS-*zFufF8-sQ%}1roO!Mi)J6+cly$Wc&pAUV*mOt
zFH*`n)cjp;3QYUbKJ7~^>@cy*h~CiFn|62d_-(KLsabu^W^((<ZC@QWtk&DzV1DIU
zQ~22f2g85lRcSdGTzzDIv?ZbL@Y~Va2E86{vpll0Fa9p-9hf`MvHX1gyC>)Rz5ivZ
z$IW?d`&!;u?ih6<Go^Usj1?{V*IG-rj24+ael#|I&h7R{*9eD$uj4j!zL}Zr@yMlV
zj&O2{^YX-5aTga~&FZW^o9^~okgR&4uG#EiMV2mg>K2C$pCn9nyPGobis-V7!Q_^`
z-OtuuGHA__4aoJLbm33UBe4M=pJ=!JXcZIL(XmKNysm>6zjDE}2;_$(Tf=?$d48!P
z7|Y{v4gv>FMT0z*;>X5_nYLf5_gVx;BiUA9tEFg^a{YAm0He?(b>4I&dzg4wDH<$0
zd{}wuYIoQ!5gdtRi9n*IXt3<&@MZVAq{rJua5j>Y1W8gwgJti*kN;{I^V4h*9E)TF
zfq|}~!Ln~$%d}5Rb#g@TM<m+`?6edOmP0nZ>6&|@H&z5EA=z4Bt)^(OJWQC9`zX2;
zogIhd!GggOMT6yt4XL(CA)nIJc?*ziC9#qz8Z3_pOq{Ng=olq}bCH}01ew6}-(Siu
zLAgDZr_L=C!BMD-lfX%$Xs|qcLgM}3^f&z@f-y+87uahn8Z6I0XSF%I<nT=q`~k`O
z0)1^ogXIMtqe8-(Cx)o=rXtxwVqu|ZupA$G_tDa?zxk;1W+2(k#LY_4V0lG?x3-gY
z*duk`6eOEV%uN*ymQ!PjGEdL=tSy2QknAXM)KWB9&aAd62(#W^%<_@}(zgNrHo*M*
z3-{S{u6RWAvbXoud9zR#PZLioMT6xn0ij^RpI@el;0PoS5)9H*G+5pmbNFO;$A#%4
zI2p+{0vk0&gXJB&4?FbpxpG<rCnDKf;4M)!SpMas#n!~gC6?;EX-IZ8akWx3SUzas
zEU~r{9TLH4Bs&Y7C5i^i$MS9_9lPDst<GDB<o=TWGDU;sQ~dV}D&20q62YI4EEb5h
z6%CdPvKdS7PdYUsI0wnuK#&cj|Ng>VE=Eu2dOW#%r8;ju>S8Iev`{oyF74G^-7U4r
zSLZE4@&L&I3q^zFTgQrBiu%ffL~tCEjRnS9iU!LS@W8{{3%0WAyqQRLH*vR8G+3_u
zPeycKzQ!UEj6$-Zz))AwV7WPbU+~WTcKbze43domM%s!7%O9WrwPo<*DKZh9hh$TM
zX+K4S<s~y`#rK%;c;Hp!$f3o-Px*ZA9yGoOrPyHjTs@T5LleS{P})dyGn6(HR&f#?
z-CH>=o6m9&!Qh9$1)s=rt-!GrxDjpxj%_rz1Mha=k9EPR53xy_^NxM@6AA8Id;Vh3
zDWQPnI>5jV;EC_Ra!-NvQ{YAT8L)mv^K)SRoaQcI-9_^|Fz_9@9&;&X^p>5l1+ZKv
z^yq}jR}t0dul`i4WsQXG&~!T-gzu=aNN@P5k!u`~Pzdx2fgu*OMqaa6@FL9;vfKw?
z{eiw`AF%EttaZ_Jfy?%=YAyN8>2`nt_zLou6F&t52|okkXEZ+t;^#DX0^Lsf{8vEy
ziq>BP@oSp9fVhk1cR=?Yeg1nOeoyP&K-^992O$1H^Jk#@nLfV<h<j+g7l?al?gQdJ
z!jc4G*_!UP;b<sjGMNYVwRoje6MnNovRp6|tc`sV%WVcxo548j2w1Keh?;?zj9m*5
zwa}anW$DnC=)0hJ7x@NQ?h@3x1eLSNa(AHS9cY028&o;1DnAnNCwkU0s8<HfuwbT<
z;+4fzwjtpR)I9^uv0zcOWcG_>pA5i9g2oxBjRnh){e=V4_p<1W=g{ssR2sMCmLQqT
zs4WW#r=ZR$*dL!^9sSX-bi*IH$h6v^T^sbk0?U1dcAp93D1sG!6Fsn8AQPxVhAtUI
zC4=!eDqy)rAZi49<f2c3_frr|_!aPeMe}<Q@E!~&dJhQbp*aOwra*h5?}SD>p%vjG
zXjcT?3EzgAx5?1k<j?fom2@_)AD=6SM&-}~cjR-oq48~KNw@+USI}GurIk?mJ@L6J
zXjBDdr2idgbcZ|`pR0z()#QAee}gP2{3f_!Ie*4q6OZ1YWiN!?LZf2jEZ+m`_h2Z#
zI?Htf>u#ERfOQYey}-Jcu>G4YI~K;YEJU+i08I*@(r_HEdfv)YTkMWzyB*rK6Z>Gf
z9%$D?7)O1o!f&EImJ4UXso72gQEBvS=YptQFrIuko1yV$Vrz~Qpj1QXxeq;7C$#8<
zgYo;h`j!eCR*n~|BF6_!3+eeRgzkmV2X|z-R27|66$7(LW)~iEMN?TWO+_b-=5;DM
z>j+~zM8hr*P4O~^79S{UC2#8y{_Eb_U&|)+W4XIf_bxOd_p5{AI%rI|0g4-FZi3<_
znx8=3C-nI*pzaHrQ&h}SRFr1qtyQw8VAqN-=*2Gs<}%PAH@X6tD>PpP%vGAN0p=Rb
ze*oqW!rrD`x916_b}@Xk?rJ;)TKK|zwC-xO(%c3#+GuVE8tpWfK#dYeonHb6lt628
zy%IR21p3k(TLQ;p#tDFm@SDJ&<%ToE^|0ZyTrM=uB^SkM0rEBSoPbPQ5EDdM-zE^X
ziMF-}K=gq2HP3+eGupSj2Hvk}?gjzfv~TGJ0lhS*LCZAqNwZuYG|HoGvj~P4!O{3l
zS*`+VRuEs~XY6wzseR@=^sOZgNg9PskwvlOkZdpte<c1d>pD*y`Ok9n(Pu#I401lp
zt%o}6Y1_$!I+-+QL7gnZfs2mVoBCh;{s<{}diI2nGmf9oXPOPovPparT+|s^$hxKp
z_*?^!GytVr;UhP%-v}sg_yFiXpt%X?H__Y-B+bAM4;`Ot0g@J=^anWa69~V__gHR_
z;UFnaJXp@w&{nEkmN+{@JE_ueI0r)qDPHYa&e70Oic_M{%j-JlznHZb%}XIvD}<W(
zDOm0zRJ%wRr<cfi$a70#g<(vX0XAKnWTH;;43l(<Vxq*#PP?w<+(~<(jz-kU#7Q>d
z-?TEK+0$i+=7JSy8GjCqo|Azd;jwj*#6CM0#i(LWaS%>eSWe<C3Bj(J<@CMvLvWJG
za)#c9Avif;IU{eQ5S$CJoUylY2u?e2-ix02zez87<~Pvl4RN;kF;Pc(>PxzXFd@qK
zA91rLy3YRfEmc033^kI8Ir6y_sF6ZkBcDr!8mTm=L5(z;)1gK>Ie)a?D9^KBH$FsT
zmIchSfFpkHA4cj7?AYR?%CcN3P%8!0{M`U*H)y^E)NavS4%EtNt^{h8G*<z&Dw=D6
zS`E#$K&_VMI-piZb3IV2r}-c>I0#LNedR-ge8N98Pl~AD9qIqwDb1v%4bB~rQ!gtn
zWPXp1YU-gyJsqVqLW@S4o1sNB;mJpjp8q|@>nAj{hk@B)pbUpt?g%hDLi15zc9iDh
z!0b59CxF=rnok0=lQf?OI;Vkz49yv!bB5-#K<6yY=YY;RnhStV0nO)u&UwPqOfrUU
z<SlGP#`6-2Uy?Wsmuwi}H>Dbu8^#RN#h)w7rGTgu5Jy~gDu_y@ISV*t0UvCUatjU5
z0zVMW0Z}<14!d}ktAv`B#I)xuiC;eI<Kdxb)lY`}WT;MtAqDbNXikOvRGQNuKaJ*e
z$WJFcKP=}@b${zf^x8vYLt^mz%#S(rz`ToJ{N09zoef|3O(l!vBA5tj^wL38Iz3uB
z&^U*L&NvrCZbe@7B5oy-NmO2Bae`m9^!D985g&czCZ)gu4-ucc0ZeYtd<&S|0**v4
z2PWk-R|1ntnyY|G70orkq=vp;EikF2xel1r(f#XzNj=T);J9}r-dSAGIHKD$ZO?a)
zHLm}mVUidB$T0abzRHYY`S9RM+gSG~G~wp{=Fzw)h<~GV<&{Bl?nq{&7Pg8NUjuF>
zFF)#w&ixDqeg*+}2otWi{j-W+TcyHsyP^7S6m;NfA$Jg}AEY@Ss^=5NE38oXO|cOz
zdX`T1cr{_UCLn5}!?fqX`#Bw!bph`#nm+*V4>b1y?>?GSp=Bx^rd@(|m*^C*8fsRP
zuqyR<<bogf)Rv)-ZwGYQ0Rzd~?SyVSA(flvLAN}bcR~AIWO<a2j_q#fjnB_G*EQpx
z>k&TR?S<FB4v~6QXB(l3u{N=`#K9204-o#JLhXjIp0AIzQ*=@Px6tt|9D;jp?Dxn1
z%#tO^NN_fBmf@=D#t4=4mpNZOby&_-&sB<Rtt{uJ=O)Ew$8zp^?owQfWH}E#4=GM!
zS<X|>Q;O@u@?}_D3aA|E1`yw%`4$l0qPZN1%W19z;!2vUfVhh08X&HrxfY0PX|4m}
zI-2W&xSr;N(DEP&u2?P~TISPy0Wudz*tf|!ch(S1-v=lv`!*}VFO0q_c{QATRYRGf
z$}bW}ry9a<%Hb>*zywegrlhE(ab$VP0)AOw7+F5DK~y$fe`)}t2D;wV0lYc@Ww@Qd
ztCQw;!0R2&$xxOIsru9oXtaZ@i7ZzHwTj3W!lK%UW;v-(v8dLeSw$F^_EAbFuj`X>
zus5`q;suuF1{)5R;t+cKhba?NHKldP5(}Vh0jVlsXN0aL-vl7n+6eVFlFEU+oJ9x7
zzxe9ub1`;DzWZCl#*gMyzRN)F@5A^g=X9SR*tT*;MFo2LZYb?01-kw7hw~$%Gulz`
z`w+N31j@RTJj-@(CADBa*9P3%Xl@5X+sPF1x$98lI?1vRywbDz=b7baRFKGk{WEA|
zT@MGYhsw-}<uc*GOq#Rcz$}`}pjR36!!7~UHdTB#t5CL+qcS8%MOkV3ZCOb_y~$Va
zqgQ<g9p4dehn)u+DEWR18K_VuRGADM%#7XjXxUXMS~)twz)teMEcXgnzan*NmU|7X
zUlTr-a7i_4!GTb8TT=~F3Eri^8#(Bv@?9Kq(`Y7InOq%L4ZGhJHpdGk(3U}#WAWY$
z%UK0kjm7C1%k>ZHKNeq)<pu-|7>l=QPTD;FF|I4xPKD2vK%)|Rv!et$l|bd5g<Ov)
zfsuGHabAWVPri#p9<LX=^+M&Co}SR)d>~s>2dy{{f!aerg?p_)t(E3Bpw>onJ5Xz<
zxdf_}kZ=#jm+1cg-KHW>@CnL3k@{9)==OvYL!UHg@KIcEbr?|9n<K#L2+c=<)lr&{
z1FPdSp8!@TXg&$7PSSiD=$|IO7uBbM{u!Fj0{yczp9A{mXf6Qy1vH-r`sZmb1V)8G
znNlGatzwm{qTE`+laEG3zAHz@V-zz=8Cw?DA9CT{xjN^&Lo{qa=x*{&K61AhCPsO;
zo6XHv=k6R^j7-~nwD}|)n&E65h0^jJKyvN~W&~B@+6tn!(j~5K(0Chhi?@vv+SL*?
zR3V=`1ldExF`;>e>|w%q9Y;?p-%uoX^JDy!ZVnd@kS~^RE0Rv@K-4-gkt|nRpz#(`
zth-~Gn3?J~T2o=(J`YJ1TE4~S*Gilm+F<~GZ8W!ouy&yQoLJ-pCD%!sRMX_H_4f*?
z6Drg^f<}+%4V5p@=nLW6u$+pGXDW=pdqV3?FR3=#npdK^+N@%<Sw)%u*1vSz*^{>a
z1`-Us4FYjd^MT~rmsS4etI(@wK-&zc%u?jFJNxy}3402;o3_uSISbln(Od@o%U}@h
z-?Z-0{i}bgo<a931`M(tEZ~w3G8_5cE}5f9CQ@l~EVl(jZK3net<ZQY$-vqsEp7D-
z+#>q!@p``PME4%s>b>eLDhTo$fDk{XeBX!PNHYqj9?;wb_)Ro71Aa5jEr8!bb2((n
zAypo~4Vl|CS3stM=1Ry^(p&|ZDw^*=<_^u(kf|n&7a)=Fo9%Iyo5oC2!!ZlXO=qT4
zG=rHz(Hv$DMRS?C6wPDiQN%GEMT?om6fI$vP_&X+siZzn+jm!2_E$kBBJ+}k;H>(~
z)V@fm5Ui4~hT|-!V26lS!*NzqsMT;>|0&dJIBvic?96yeX3n1~x2yOi)8o&|4-)0g
ze}7iv10HWFIb2K*<H=zHIb2B&*D4QHxFvlTb?;pA+@`T+?PBydmCF)Vu1Z|Jd<AcX
zvJY<wIb22#SCPXt<ZwAToW5YjeDud1yoi~RDtuKzKSoW&s%vR#glPu{4$u~c>FDc<
z4GcZKhWPmU`41heXX|b0Zf*Ae{jbDWYGQ67v+6(4W{{nOqmzr9hpDTPy>pPBBs4q;
MhKvsw28Q$g7ZY+xkN^Mx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bilinear.hpp.092CAB87B366F08F.idx b/.cache/clangd/index/gemm_bilinear.hpp.092CAB87B366F08F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a174214513f8006d1b0dfd24ed1dbf192628dde1
GIT binary patch
literal 4822
zcmY*c30PBC7Ji2sA^4QAgq<WT57{B)LW~d*MA^ckY@*f$#A2(T2x`S?$1Z@jg9VY+
z-Lf>)0i+F3L0hd9RA@_Tp;X5%jzyg*)}`%~w(>d7%izmB=KJ8?|NGB(@45du_ndo|
zBrRHGv}2eBYm>HbEZnkY8pANQ{9oyolJ%P`7$%<QwjUXK>*-Bb!yGDq%`7pbwoWTu
zvphG?>hDz_v~;Z>=|59g`;;O7;=w!TE_FScx3AiFeo=y(qSpQ&7bkA-x_oh#@tfS8
zJC?f1$3)u~wZtw<_nHU<JGThuzE{eME#qED=~<)P_47YFZ+NI*d{~zfp0pwC#I|q0
z`9Ay4fE~jVHZ8yTWP~5IT)xM4EXbmzvaeyXa^clW15Z8t{gd?v?)>-k_Sf&9-nyi{
zu1SAvwBzXIPKO44*~zz0h>m?gzL@BY_*b6Wy^ckOLx%BFi3!HGMF!)oul(<x;J>%C
zCQHVzIE_}Q?#ETlj9(VuRgm#jn9G`<-VJlU9Q)l1`oM$Et$A5ryZWDBeWaq&<M)&=
z%T&#kX*q_4?r)BMTIbbstV^?c|BN|SyV<<(jXy^I_4;Dpx|5oN(raV8lIPu=r~dV=
z_U^4WzkS(1?rOu?^zhKZZx1ClpOWQS$G(=ZrRMW>?p7a$r1&K|+wOIoSfcma5T8_@
zs?5_bIPllWkGx1{!ut>0i$?e55AF3Ry0R}nVy$fDoa(o~&kwrrNPGY18|`P_snWjp
z-GP9Bb1#jr+!h*jH?8gR=FYLRBO~&!*`2HGCo}9PdvYgQ(k3z|?@B*Zt1f_dxrIL?
zI`kl`Q}p^xJBJ3_n}I{*mv?sSn==o~Jdaz=$=`d?US8vL^_P&{W7Eq|p87I6E#ocj
z_@N%JYcKxr_Uh{&Rlh5*D`?+;&$i`mw&n&GSQ?p~b)se3!lJ`Bu9rJi-kYe&+vc90
z))l#a?1!t08bfTr{L(LOyZWD+K62-`y|u&j@vSBL%MbSk4@`(9x6^O^D(ZN&#26QT
z?N7^IznAtfrDYrQHv_}WNVZD+qd4GAmT{5s>B#>$dy)NOMm*yw#?J#daJ<DFe7I|8
z5zDv=1`3gaHxR>=L1^&0H&2~;`3B243I>}*o6;DuwZ)Xd+DD&P4#mDZhh^LagSn!)
z{0uWiQwD`!T5V9ib4|)JQo$f!l+Vw=PMb1#?(VqbjM20Erd!y{Y`F@stN`xuhq0RS
zx+3gK)^BAG&$EoPFh{m1o1bIW^r<;&k3M?ko^@|2%Qy)JiK0Z_Kw>pz(0}+**RuE8
z&dji|lQUc`Sl5Ex<Bzb&`9-%bwO$!0v9oaCO^$)&7`Q+F2qw3=;2%4)FaCvP+=OL<
zMZx@LZ7ipjVM+?um1fRjnCbGKHx@Lg6@DxyB~rCH5S$ZnLdb{ssLe6pLP&_(yaX3Y
zLJ?9DB|=doN^Mr2n@8p$lthx$W<9uMl8lg!=n%>wIcl@CTpr0oXf0W*HiwQ|N7f-!
zN=nt{#NhN0q=!&JJ(kA_dPt_Y8G@QI-wX-Om~RDTD-dem2FkW4v91l$DQ^Is0pku(
zc3}Gs&~;$G2b4XS?*Uy8=KCP15Bt{#34PeVehBKvd_N@gV}1aX19<*HP!2wcb%S`l
zA<zwBYy_nd+Z#b=#QX>-M=(DEx)I7-KdAry)s%=#3Cp<?_ZXUJ9>fD7f8rlQ6D@!Q
zAQVX=V`$o`i5j5=WI+r~w0IJa&|<PUh9=q)vIL=(WMvFZtyN?dLeG)sVrXh@AR7?c
zPPWI;)G8xoW+d#)Lj3<4w&#<{`nKy)3YO~vNf)dT^27GM*Ic)@kN2@G=S-Z@QF9?K
z2)PkAggC;XW9UnK5%ME`2!)eybPQEQg-`^EKq!X9phG#I%tt7W#F>#06@Fi)(o6Ss
z8YHLbzP$E`uSqKJo}R{X=_Eax>b`=kh^9)fC+nlBw%f?IXhBV%Sw&ju*^h2mv794u
z)CzL4oHy~-3Qn+G7zxwTWY7|=mWDrtq-g2FOUY6#UAT}GYU#oyq(mz$EPVM5npy|l
z;3!B&Ax{Wg^1+X*jj#SRli%QOu-XkXr%FG=uv`^bRbgBQ!F3QWlysIm0Ko?+PF2@#
zatyAj;U{_n{N4Z+ov0f8sxdwUsfS=Go#-&69;P@e&2IL;_eP7vEaxD0u%U{1ial+p
zVj<!X8>(28ILd}9mLyKHp^D{*b8M($1>yo5s#vkO*hWx_<?3N(J=hD<2=nm@ntCf$
zupK1rG)08Tcr#OnrDm5va)}Bp^jolg++_vBje)23FL%yL<1<7?WYKiLfg~`R?spEE
z6HWKKkSvTA{A0OHl8I7tHCc^Pb0gV^Qga8{VNT6WDJOH^pIq^>1<NH^CP-<#J_gH=
z!Hjn%K3G{B7Mg9%$3x+uqW*a*JXO@c5JiZJ`WK~$Qc?es6iF%+4@Hg&#Y0h`5+r1~
zVnwlv#-jq_D<GFTRtwTv@DmcJylLl-)Z^zb^3A>xq>UgG0%yVzjWCbmCXhB^z6m0l
zFs}y}J$O-jJrF%;DQ*UrX3RGOX~ujj$Xa1GwQmJYE5uRU2C_EHw}GY&^9GO^FmC{j
z0rMRo>%e>mXgV<84YF>`cY~%I^F1Kz!F&&BdN6+ur01|c=V0zRSU~;n1!*tldtq)b
z=KH{<59`whqz~)U4=(+f?+4P4`2mm(VE+a{Gl2aY1lb_w2SGE4`5}-EVSWfSLzp*$
z%!qj-XpER22H7y?he0!p`4NzfV15KNBa|1CnXdq*`>}AaS?)8Cd`2t4uKNq!HD6S@
z@)fcX0vaKV>e~p(jgUcc69hD2z6p|>C|^}s?e^9E$N;|Crjpc1da}|;8bTQ)1EFPP
zS)|}7%Pl9%5qg?DjnFgXnMisfi%1bd&yr^m+CsKO($l(?Y(?mK^1K-d9`XX2-Un2m
z3m~~b1=5#$eA@BH(KEb2uEZ7f%GqQ#>Xn2L)GK32thrZyGPdXZSYUNIU(xI=?aUR7
z<r=`UfmSr3fLK^Rev1gT+0+4i>LEx7gz)8S_SC~fFdY5hxOc?sN3Zh>)qr0OgwYyx
z1X7Pcwy?y_FwuW3ZJhpQYGzUgIVzq~PdO@{(hxZ+p3*2eDxT6Lxga3R<w$enG>iq(
z0yzy;v9wrD!%+_|_24aRVSJ6^bTGN2;xDWY!~tExJc%c|Xn7GYgrZ5b`NG6<xg;08
zR0>D|dYcrJV)QoIOg1Cb3c6N$b7naMC=K{#-3`j_C$X*@+Yf+l0OKK04q^LYP!2zd
zb;HzNkd@Cp)1#5*-f@r|r(OKP$^E?{QJ-}JD?EaQkd_o~SYsNOmK2n(G7UUU|E7Mv
n!ho^h#(VqJrOCcOTF&-WczJsJFkWVoQP80`9s1DWY&!fOo5OYn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_bilinear_xdl_fp16.cpp.372A779C3E21DD0E.idx b/.cache/clangd/index/gemm_bilinear_xdl_fp16.cpp.372A779C3E21DD0E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4701b88e0503e52394ef0d01d0b86ad1f676806a
GIT binary patch
literal 7866
zcmai(2UJu?+sAo^WnFeJ%hDFXl|@95CQ_9mh>LU)Q0&<C+AxT{_e6z=peP_>i=v=d
zW9-s`peT?4h9ZJJDk{d;_!>>#+5O(x8Nc(L^Kp*Pncx2JojZ5RQ?7C0At7h2baYhH
z!<Qw_STseTqoZSl|1DgUJbN;R90MJlWx3Zzjs514x9sVDhuWz@9?`!n8Zg*u)oaT;
zQ*6gRjy+JB9-L(ppYg74(u0GxEoaV+KM-zGyu-)iarFT6?>5B^Iey1I@zM{iTOT^s
zG){^hc<@o({;J1I`d`hnEt~z%!!2F^UYB<_%{%HDpC6i<@<;EAgWH-qr%i9&ZeR9h
z(t@rpQ!=XGyg6DmIpE5R-o3-G-+5Z~dBxj-*RI?M7Y?ZWcY1rvta|ax6Q}!}E?ro4
zZQR*gMa6wuldg5$@|be_){4A2Ve)ov*6zlL$fK#XKW-YGyY@@?$+(;MXP)YKomy*O
z<h5qOQhmF2aq38m27kTI-oNKA?pmuq-$Xp)UD%0@Wse8VXw3AtiVM+u5yW5g-hDL8
z!KCkN-A9>TF5_2ydjG`Z`{Eo`sAAmu$98?|ww?&uStA{98x(gVF35D`kczqYkEh4@
z$ovnC*-+csweHy}`}1DH^`q|Y_I_1gV(BJ~M|p2c8f?Zq`^o0L^qND1+pyJr{dS(&
zT_)La>A1sy#(5FrS2n-6A9>z=aOCU$%9T$av^ysx=U*I&j-c?@?|hFQ)*b2i#_g@=
zhy#7TKQn5Alm9Z?Q1gujd&R!jgKGNQ8p&2yT!`m%<1d}*Ub}nr0*626zL{N-+2*cT
zmK0h4gYAYmQCXti!6m{kyUv%tw?3a$x8_FXr&8bQz-t5ZOMdySz`(ro`j(cojB$H6
zpK(%+Zq0t4AX(}3hre!4<5=PL#=K{z&FZbzS0q0TxnT4l<d4ar>(2f%#?57+-m0KI
zy_Q6BanDAlTEtoXT~H!&?p0{%vQIMRd{O#<_60_{w|bu1cI=S%^Qb>=)a;oztJ?(A
zI^%*Z2_NfiY)mDAdu!}Iejol~ex$$hO6t+(ZF<`-T%3LDXY|r|%Ix@pq{Tyy^*Jxr
zdl5P#%VWIXu81c^V>%P&{gblr+xt_J8w20$d;iw`=jw!?Q@;c~&$;^AZlhOdaB+0>
zPQ@+v@OaL5=|5xd*O;EsbsQVHZ+iSchwKZs^>-FOJ<<2~H`mU!=|^Tx*<7&6dZU}+
zXE(c~{&lin*N#pv@p|x^?~e7dm1l~?0h@}u#aYSr?3sSC#jDlHVo5~EqpCcg%N>%e
zskc9#R8|e2JT2hwpAYBtIs6^EJ+AEThsrtbZLyE;e(t_IqBz0JBOvKSw0l!x_M-6*
z2k)EuVD-?2hGru7n|IeZ=brpyLu-rKa%+Fz+#RL~(P7_ybhY?vOZl~5Mh>`~HKzU4
z=+Ca(9&S0??~K{`X&ax!-<fW5zfm{s;9Sd#Mr+r_Z|->2m>3*+)inNuLsM|<#|uN+
zOvLApT>CP7)UC2Lx5qmx2dQi}xZH+`&$<eRO`L4JWwz_~*yZ-z;^GYkJ?zA?6v6($
zH}mpuFVsJ>c<<5+@ne*iYuDuViQ2t&?5d@EKL+_VE3;+`gG-Woy)~Np<9CDq${*-H
zZ`QXDqwl-st3JE!(pPLd_Q}99?^><#$yLiz98-Hs-@h|(<O0?xt`{ummyzM5YCd}Q
z^}Pp`ftG?tf`!Wyr%u<^(NXO0)$FqAYLOhp#_8~Kfn26B*b`~~oE9&B^YndW4+$Ej
zqr(RZ1C2F?$gY1bzjFOUShfU3V|kEpkW6EU>|aom_ub{4$0cY4mfeN!h8jcUq2Isj
znp5<7j08=>vV+h;Ut@?IDqg?iNo;wV1dYM6r_fWbF+`5`a^1cCqEWRF#bVjCo2iM$
z5P3pf+5GE8A@?O{ES6n_F8Ufn<XIC|-aBmm{bLD=!Lo<Y!$@O@oTO`3EfP*VEI}i&
z93TvkX$+CK)HwVY);F_+(^-n;Bteo$V~Cs{6ao|K|DGyA!?5fobTiZ#B4@<qpZf6p
zr)d&28Ox4BM}3VU^6s2`*KYpT&PdP%EUSbnxyBGVEBwSy_qQi}kfCLG#T51mSB)WZ
zcA@c-f_q<vOVJE0+lXys8bjo~^-t`kNz2Yk&@|k+v(Q;zV~G4ix@~a0@l!(soheuj
zk_TC743SR@{*_qmRryAOrefJu=xU@fL_Qm49M}F>RUkp*u<RxDGSnC%UyswD<GOOR
zSb~OQd7yBhfyNNIa<XV{#(>hf5;PvmzCvFEjUn=_8(j}~JlZMJytn2@-ooC(mB{se
zX@AY^tT&aSnfSeXi+dYt43V3kgd42tyW+eAO~jr13H=N-hRAK<cdq!{ExRB=30U?P
zdK+qtR&KYSR&AHvE<y2F_7VCRXbh1*zHCeH_jIz21Wm`Xz0kg!#t=Dq#>|E3W;!~$
zZj$5stEZ-lIKBdGE5Jd!jrf~jdz0l#u&rdd3i?-p7rnj-yqdt5ax-`}v)lq+EiAW!
zsujY>^}@S1UcPcXE!O4u8nCGW4<c}UE!fnuTn9FFlns77cB8Le(lx~K76J>7ydcLb
z1PU!lzcGtE9lhFB!0|i5Y$sUI-gkl7E|!mg*%6kHg4t2ZvP)Kp9+_b^LiI4EHPDN8
zR14Btmg_)T$8tSL>sf99X#>lRAZ?^<_Jeb|Wco<}j^B^W_9N|sy5DhIQFSb+9>4E)
zWV9Xipd%<mW`#&en7C0S*0@jNn*{t5uzUh`<T9mnz0X!rbq9XXXE6B;)<m!trzLIf
zZoC@__=_;)A`B&hfWHd9S78w0UYC<+y;|r0EeLQA($8R|721+>x><dy#9bdmrU#KZ
z5&Bx@$qqNT&D9g|#UL*R?K=qg5|Ed$Tnh41mdik12KJ;Q0bdUCa=I%Wj?NgAz4qLA
zJd_O7D+6iYk^~K1@mIlu<6~4Y!-;p}{@2vFR%aHORh3V|<6n<@u1DIbAu&cE{wlU`
ze6%Wh813y5SUv(<GFFbyLNQrr9ARHs>AoMcm;R0SWC4f@KuQFTF9cB`%O^l|g5{GS
zI!W2rs_WKV;gl|s0H0xL1;|K80(^$0H(9O(X(h{5Agy9K1xZs7bAJl5OF?~Ue<{c>
z1!)&vz^9-wDfEp<Y!!*Wim4nQt_qK$(^3MG5-=s5s?WD)E!|BV-vFKsEH{HqGdR)f
zHzTLb$dhs=ipfM{3H$##^4E8FQ)6`nd_L-*kKBl$=0W)=faQpM6iGO+v*C5ovAiXC
zJ=(yijh=aqe*i`gSZ)WSc9uK9sDm=eA@syw<qjP0X5pqFahT)XE!-7k#yH-~!b?G7
z6~}v9cq@qi54pPe^}JUz_hXlH4rJ#*P6n*zUgzjxP_z3v7)bSvNPi<TWLdEh*-+ku
z^fytRBsX|ht1}zwNh4K}Bk1yFAj1smWjKB}irI}OlTIVDZ=~Jb_{spgr;yMPX4RuY
zqqG#Y_pwXBzxEX3gB5HOJcRT;Y(U00a+lN{_-EIFt|}7QhgH<ZHn_a46bSexkT!uO
zd6s~025B?PEg)@SxfP_Xl!*((PC%WT(HB(bm_+3GK(oMJ#9xq<4LebFzD95Hk@=Xk
z{Wk<pI)3~@VVn;y<M^$h+zP$P06BgeD7Uek0?HJY(?FTV@(xh$U^yL>=`3e~GK=MG
zP-e571IiqhbHOMV%;<Rbfzdve^S~&N<pW@TfVx}kx54}%%ZI@H5X*<b{4iyb!Qu&4
zr>}IjB2*F4WR*F74~p4?#*;zKPF}cV_{aRg_}n#tpa~3!z~SW;G_%|Sf)<urLC{Kh
zZrr4Pzr36D25+AWV0wWLYu?VO_Q4$yOYxpm%9IL{tzsIYj3z|VYrOl_xiua8aMf_e
zr#}YE$E;7!Mlspc114>*zusd>o&z4j@1Xo09Lc*Zi43lG8si;<g`mKoSd!B&E9-c?
zLQr=HkE|1HJ82-}`1fG@p7QET<u(Z;jKc8!xyUsaY4fZNC!-SM_to9Rv-c95UxFWb
zBFDc1=T|IuLTD$9qWWuaeogi7-VRmx)TCNsH>2=Y_>*bD6s{OfsVCRdm?5{G11Kqb
zDeWma$Q<Ov6_SU7_^YBMB|>*oXWBmlarU`Xw3KmF%S6i<S}t17&`QxthE|DIF|=B=
znxQqKH4MFne(!0%n!4U6r*OfCIhxH9cL7F|Arc3TgP6KXqWd6B6{dAy9L`%L#cU9(
z0nZvXXw`#fJ<CmC)AV%^Gu({2lZSHrE)=ti1~F2O;e({E#pobBEIbtClySVLg{Oje
z?aa>0t1jumJN!bQ!1xpNA$@25IXPj2p`rl`>yguXG=K=?Kw{sfuIT7Po`czQdOulF
zgfvAR@ssLsYL!m|0V(ou0xx%#yOZn_A2J7}E9pb5Df~zmr0~N>PF?lU0SuH4G$EB4
z$NS2BO-Lce@qRKt6H+-HoEIsGi2e2<&Lb<pu>zdQtEn>|$D1ryf@3AiRp4DkL#2RE
zMbcE7`5by<=J5DM?@#zJw1IUSoeXs@VEusQcCc<|xdZxj(CS2;r1#&Bm@=slc@`oc
zGSDBkjLCko___kmT0Pj*vo|(?O#{n~VADvMco4jJ>XMkwU92iLlB{gu=oNmu?nUm!
zOaBs#UxF2R?1^Pp^kNbZg<!#6YA+|n;wgux@#DK<opJ7zf-F;zf?Tg=w-lr;%G4Z@
zf})5{LN!i#)D<`#%^<Tu)+Afz_yDs2YvN})J_p6*&@iB`b4Kh(6SZAF2w0lkefTU*
zSnz9tXd<Y?f?qSsE#TL}aw~+l!US^tnF+tSA4)Ye)>D_b`V}-@aCnpG-(<NG^eb7e
z0{tqMQ;>cN(#9u}t_j6oeNy1~s9{l~$P#dTDM(6LPf-V+b*yJ-1kXm6TfnA;dWwr7
znahd>Kl{y8fO8Amt-vJX+rV}k%PC-+!g3nerm?&OY<IAn4z}qm?*z-8bc+c1U0}J3
z<t(txVz18z+iaF|z&3~FT#)C|<V(Qs1NlCd^FW@*@&QmBpyj6k=N^iKEFS{JA(jt=
z;xNlc!14$?-lJf7l=8)y?bX5JTOBymS{PWE(=DnnP-yA8XWVqXp@&}Li=6v7yATJZ
zPAEbzz;Lpqh;PE~S^dd~_7$uOj-V@F48g^qtuKfR!#AlvB+;8fhJ}ozH<dzgDZPn&
zVFJ*6St6eEZts;_HUx|?)a+^Z3h*W)Q}1c_n=Ds?dnL<N5Ksj}==G^co{F?7HTh(v
z_Tql>!HgI8R{BvduKt!q#?okYMX$-px*GeElgRQU(q_N+=KiJAZ@y@Trpos$rm<ID
z<$IQ}Tne70ESJIHG6*7rR9E>!%0WfAS$^a16@k_(G~U^+jb>SD?`-#g<#w=ZXSoAB
zJE+@NduQ+MG?~JcH}Wl{B_fG4I6$dC&(S3dRfRI^Tm~^^G=bpoXyeL3TadPFesb^n
zBfYbDeolet6maA<NP&m_hx+Rt?R}VP7?Wv~fuxM}ALSq^XSo$@TB*O__{}JEGm6nZ
z<i9lv-gWA4g!CapRYRFqD1qP-y6f5>ts9(fbh-r(IcP*s0*N&*MJLjb?`d^?=F-Qd
zKdFsg&MAubF&XkRJ~S7>{31J+t6+YW^2?n?AND!bs4LlG5EO%$43fheLQukTDF{kg
zE(1Xs%jF;_XE_6jGLZHQO4r%7BSgQs^}qw&g@)`x+Jj319Zv4lpG4_wB&ZUC=^31;
zn#j;3)g*?dtEMwFLp6h;nW~u#C8`p&^riT&U5@{<-2ZoubdqPx!{52_|GFkmTX;_0
zNK%{s`CC95Iy$qoO<jJ)nZ-|P!a_m1Xz{}2S+nQNo&QbJf_aHLbF|m!B+}-5+FVSV
zvuSf4Z7!tExwQEWZ7!nCWZImTIDH<EF=9rPu0T)NO{6d33}lAVFr(lhc1GebV{;SJ
z?iK@m{QL(61P<<J=H%PkyRY*9`c>XT(bKw@jcp%$2S;bu{%#&#1FSqPUEEb>@{sT)
MXz1vmP>9s|4~*9mjQ{`u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_dpp_fp16.cpp.D49004015318A845.idx b/.cache/clangd/index/gemm_dpp_fp16.cpp.D49004015318A845.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e21b0700763e3d206e47f404c41261927c5508a1
GIT binary patch
literal 2590
zcmZuz3pA8l8~)}qedh8pV}=Q%Po|D*<##6{IfNmSTPK5DqRHSNT~xXYUFa&e(!Z#X
z>k)sN4kbA*DGntkB}cg()t^k|bUDtR`PXjES?hdj&9nCYp8f3ozWaIKZ;hXan;VCQ
zATCiJ$)f138_5WQsDWS7R`KQ|N(fR<L6GG0l>vj|L0kOM+}7|X1rG<`9M??|ttdEV
zP_TxuvZQ+GtXrIODdpE{)7XxGzm`V#HMx^jNyPAeKkOuFps42@e>l{?xOQOQn;Cpl
z->sQHT9U;!-3G^+-*!$<-U+R!vR)EsvoFLh?RrqwJ@tF(<1dqQ?;AXC>KPeynn`&t
zZD?;RAB|m^<uE4+87;p#Ju@;r>8Shs%y`W0$CGvQ&Q;V})fy*u9C~6NvtuGDzQK{K
zky}%L?Lf8F7*~o}ZLsm=Upd~Bd3VR%TTXuE_2YjgK6smarkVaYzp(D68+-F%x2vdS
zSg~q{bm)_GsL1KAXP&v%=LwIWa_)S}FIgDvpSoZ2pZXtb^!a&|I)=Aze_j0aZ!P0K
zCo>its&W|i6|=J^7H*kn*FCs2&rt7`zq*F|lH}r>maGe_n-?{+4v!VCK1yE@%N3%w
z30t^feZLjjx&<t9O`YJXRWb}Rt`-Z0+e9uI+jd0dkItq=ZPp(CIp0N8fQW`Hin5(c
zZy2n2@L;ZI!+^ND!T+GIv;N2EGcPA7kBrGe2X$I~>ccIq^+mT$&U+rWCUmy%b;(}K
zXX(v0iv&WY(%xWU>{L=n=ge}lVMDJ$dfWcJ{JgV^{f@@YvYFh3(lu184wr|giw-#-
zOY7X#tDAP6x%5)`E@|(LgijwV)^e|zZI20WYxH<jU*{7%JnxRgh$ihJhmKHVQ8n)u
zU72zAdw(Bm&C00nvoxzrN>5Q5%#G&@{epsi*IG}S6ZKHE^~$r>OIWe7U)7vWm$*u=
zN}mQ8xki@ezW7{`<M4#IQQGTqmiO#V&bhUi<!r0at|W$4_W1|D6iIYVe_$MC`E;7q
zd~QC!=yE{iNa@7`_iwizeYeiMx68WFx9m%XRddOd(@)V`?Y%m%>y3L)I9|E*)X?Zd
z4EFieo-p&LNfD_PPVM}PsksZI#)X$xFS~G$Zln1jHA!pHXx{arv=fa5p%*K&H)e}h
zMZM5E@x`vpvT%!+wv!|-rH%35Bfkb`1SN|@FW9V+WcCFm?!rn8-WDb#I`7aspF5U=
zm)%OR`xs~4QpuV<eDX5&RNt4VN#z`!Yw=wdPhPR{?r*I+sQUTXk72JPC1}6YSU8-}
zIUAVw{Oo|u6_HR`I>C59^z>PG<`EVC95IzNRr6K1rq{}gh~9Yg!DN5+E@HRqq3lN?
zd9UoRbu!11q@ALOUq}c-YDzbuk+t5bAU`08fWzVF$pQ4N{L$F-9ZO7d{F#n!0JI*a
z$CU%5Ij21w$hnRO>8L-TX&8+m2S}TK+<CL<gL@twT?1$q#$w6=(hgU<3)3ZM%IN4?
zKvOUZLk^I3&3rd4tojm6N7n-y$8f3~ARQaB`*HF7%$Ic37tmCUswM|WCn{-lQLxZr
zI_d{#HpXVk0n!=WM)mH7+08hz5zzW%eTp0)eahL56!P%T2s-KwXk|>9AqPk|HW(S%
zMRtCqqrrgI!n8Ey0O=+_s?c<I5RH!d09qAORgnXv+bfzaBwssR=x6|-)iHIZ93VY3
zY!KO{pEpcL*8!S=F;wIL={JwJCGIIvv8AJdfM#M$6*)i}5l3%I8Uf#0=qyV6o1?gZ
z*h6HHU<S}&673(2RigdPP$1SLjG-16@C<ooa0h5MkjR{d@c>Bx5vCwn$@7E~SO7pG
zb6-KiA#@ca3wZia0*eB~E2{`7@up@O&i&DuN5TdACi?c!-H=~e;+v}6o`Kg4CW8%c
z7M2b;S5`N06A<bW!bag-3!Wo97nT!{$ZAPJ5<$c%NW6G~P_mu>eV0kFwJ$iFvx?&n
z{YZc7s$<uET$_vwlr@zJSZTO`qDdj(jo<>BCXImYg$r;d&Vi}J<iHik>K2A_qA1*l
zEw+9Kj?=Q!a)U9!{XyWeGKQ-`u5>~jE&@)ncw88G;ek>8NY$2F6c?y6RoSqbaRHOb
zWW(~SA-@-OS+)O92^X-~ECRZZ3)pNn0dETzsI%1xSnu#O5V)*mU?^~ZP^hw2fv-V?
zcEn1kfqQ^7WDTRRD1q1s7lAZ_c@Z!T=r-6_)<g=DRfGqWjNLD5yqx0gBQq(G69!F!
zSpu!B!xXf9!VhYvYxk|A4483&mPh8P!E<M)hE6RT3+G`VgI*HR_pR5!`~``4sW_h@
z5mLo@6H44d;r?KKH%jy<&Yz>i@VB`*Iu3liz`rYTWRj_D5GsTJKOt~{AYtEzN`mUw
zA?9tTnst;y!a|WqQB+(EAU&hKl*sciWeSy!t1ua=?rPt!(pRIov*xqaHMMN*>>XA(
zI;~u$VX}OYwW01mz2$Iq7A(}`E&k5H$asnAzsxMGYzRy3rRFXg95;_})HTR?HHnY>
E1*b%GhyVZp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_nn_instance.cpp.684811D08F35F1E0.idx b/.cache/clangd/index/gemm_f16_nn_instance.cpp.684811D08F35F1E0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..832c1389ec316d7356f749b5b45e5fe9b8d02fee
GIT binary patch
literal 3584
zcmZvd3piA1AHdH!GvhLb3aQ~RgB*i!lFK+U8B7ghnracEK|{LAB}*$78)nngDs5V+
zbb0E_O6u!wx3ZP4o3gcCD8{udrBzE>eP`_RcFsP}H_tQgd4KQ!od5Oxzwg43Idg`M
zAt*={lD#@9Gl2;~&?NGok(s`HItgJM2+FQ%j9i`i#Czfc#iNqdDGTnF?QWHAJQq;)
z%j$Y;#yN`-!QHM9yRh#pLq-lZjg4y7tIe)loT%^0kM8onplux7w|~*kV>8QkPf5Fg
z3~b(@+~VT;sgCc|xM6mZ$r@Qb=h6gT-bmbv-pZcmV;ig+3F+XvU;_~*uSh7wZ~hsw
zV@<pD@zU6;PnEIlRr1+N?Ea&*PtB@CyMFCyWStlpJnT4i&k%EqJhm;@xbTNfH~UX-
z5e~n9nB?&(-A{I>XVErWc6rzR-)2?ji$;m_f2Z%RircT^z3KmW0j^6~!`^##X`%1g
znuF0r@}0>m{y4DzsrkXQRr~nSg}?2sJ|>=05ndR2JHp{P^mmh2MQl{^Y4`g1G222}
zS<5=#3+=uPORyVQn*%@A8l2qfv|tr{W8;zkUM+sje)*tx;j)B=K?}bgH?LFi;*b^n
zYDPt8lvTsU%GT=V#NBz8ZKWeW_`IGM@}@TOMTpd}aJhSq5qI-q(R^<qOZ+rueYF|v
zbToYBoPgy8Gq==3byK}XFV0$cPEIa}AKp0lS5JJAXni!aLYR7VttM7KA#<wSb~K|i
zw`C32WNUt>>X6IjlPym!hig2uzAP=blwO){p>q6a@?h)a6Xk~2Cv^*xypIO@#+0!8
zQLOra>2|*D!*VC}xBh6*(A88~O1Gtb({^S_Gvl~ZkY(wulb62nZYVcAZsj+3VR7R<
z_H<#r%gJdh!lA|Z3H(!orPlnw#kYG`?yG4^C_48vs&+Pa&%{*e^OBfvM@844{8SUU
z{>bf^ZC)pc;D=}05;O+`EFyLlI#ev>W(IWMZMl0VY!jx9^XC2dX8v}z-=Rr$1+Ld`
zlxR)MTdz(|FP^F2e6+dxfM>(0d&OVNSFXH}WuxDI<5q*m1k8E!rPgP0jluq>FJ{|h
z1)6o3w|m)0m{zlQE8{v-1q)>@hk5h3(SoM$4c6stV`*hCWtp<;fgi3f2o4Jl-Lrp#
zXR+7J3Q4%sZ(qfU0`XNxb4AIWzBj71o&kE-ZpZ_#jFmQHL>b55n|ph&p!&!BV<G29
zk5oJKudPbiop~8{zge32?ugd~eyaYpBKvb&cl9)M%_4qImhDV9ac+*E!0=F6vC2NY
zd`HTwu`#Q7@3H*T1!|F9W=od6&E~F-dmVk5MX#6mE?K`}u<up&>$m)CKAnTkoi~Lh
z;mcPgzWB<@&tJo;omL_o$latB;MrmQ1%H`z%X0O6v7zqb4`cUgYg4r8i`AytZNpw4
z&9W-8Js+oS>%R*<?QS+@=UpCiJ@os7@UgiA`fOz1x@Rg^>z6w&sCwI<l|2_bzW6FV
zUzO7L()*R7<@@YE(c<Ati)WS9^wd-(S81;#Z)Yfq?NyH2l{JmOwx#ZO5}D7?1aHbI
zd*9M*!+tcEH+q=C#bdZmHJ_k2Ln&u6a#kl!fg#8^_CL`fH>D;<h$;$F^7Z)K@qs*5
z#(x%-JC0!(Bf|(uBE=SZ7Toaxh<T4cOdl9@+0KE!AjMLa6pgoK*`kmk0z=CdL4|#8
zA7fiKS8<>SQXIqzLgRTX-e+;e`PUt{+IBACK#`<a#1f(L)-3DK;yUAkD^HmQ@kYo}
zQtYYcIdyyhwwt044LF%3sW?zLDVDS3R`K>Mdn;i3_jzKM_WCKOjF3c9?5yX^9Us8g
zT_FwO$Xka*V`w2MHn%Xh9Up)gN>578sMd!dJ>D5j&^{kNlcPj^Py*zGqr`kL0wj;4
z#C>oASmY=PAA$hhvp?)@FOqqE!66l>&{RqRnug}mC=XS`6k;kYmC}x-VYxKQ!_+W^
zxC&3DOyg;IE{*bVHB2F)B2s~0q?c(#E{*aCH4G4FN7&=LVe{qDy|o>Fe2!A$CJ}?G
zPCBNxTvzv^X(Ab!2o+P#qY_j~qcBuKBPE(h5hlXKz!sTO2_~gc7^a|+5=*2A7vW+m
zEL?(1X%vPlXr#mwDI!FK7{o{>Ktf1q6h<g$q$Cmn0e{JC=sJMPMu1y@7;IwD3NU-T
z55P;aM+aXW?#VfZl1b`?(LP#@(P0=Jp>rvMsu5s}eE&Xzw$ix^LA&WZgrM)}+{r+-
z4C;HV7smQ%8wDGNu@O3#BA6PXl8W6&uvR*EAy_w^hY;)?ojVzrmO*`w_riD|ZMWgW
zFg`-(QUq5cR59@T2;NHPE(GtU^ALi+qjM(%*D|Q@iC&oK`@Chzat_192%Sq2LXChT
zlJffq(Msnogy^R85JJ48b0>q)GQjuZl9$lPQ08MMQ%Ns7^It8&RED~RiosT3Div;u
z0B}x)&og!G`)(GCJV``$A_3Sp9A%JQkO1@)M=7<F3V;HRQf@1E1PW9x`IR5qM(*j3
zA_J61E3^T|qm{t`=h2d6K=42aNg9^J*QObPIuCqB5}F6TK@yq=zETOz1ED+^c#tIv
zzU30-Js;*M%>&KD!5Lbx;mkLeo_CzpEk~Nifn}hP><Hb_4>|&vC6{$aK3G2bM^R36
zN0o+j&cfCLqoU*6@-ZqpXIp2Ciq4bo=>VbwPqIsOcLmVP`BxDxhl~1^yoc9FD*?;^
zJKzg>urVcez+sLulo@JAr9i<{*nt#)C+WTJ$^lk@OQdn#RRfkuO|<<062oGT-!;@r
zs0KzWX(s`ufw4+jr{WqIucUPnp@E6Y&pW*vXAS93n6gP%RyUE5iYdq~NJzybb(0FI
znB=Z<9}vrn1pYZ!&sQ@I$c^QI3aHeh?x+Vvm;)xD3W2#}9u(mYxB%pUybaxP4~hr}
zLI8?GcA-1r0qF0EkqvFBzx0rvqY-En%~S*yMN=MuN72+p5K%xudeW){-Ibk@zBFT{
zZW5y#{`tj&2?WJb(`afMM@<(~(<Ps$3?)8RKe@rz@|r<HbZpGxxFzw>xV3-YSh`xZ
zYB@Qbw`@L3Pv5|hZDedR!F-~b>7=m9(mBqPIbjx7Q>LQUv;1cVhy#OyJ-F^7M`F7D
r|68@>@olG(>2-2(b;D<RdinVJ*$HiCc+cfphRTzfAqyljh64H@MSnwj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_nn_instance.hpp.295DF91B6CFA2E94.idx b/.cache/clangd/index/gemm_f16_nn_instance.hpp.295DF91B6CFA2E94.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c989cbd72b1558f202af4cb908c2c4928887285f
GIT binary patch
literal 1396
zcmWIYbaTsLWngel@vO*AElFfyU|<jg;^LB`%qB(#h7~}8iaC2jV*QmBMV>DIvA|9K
zOIq4i!zC}9)zs|X9{W01;ZB){q+Zd?%-6^M?=!w!R=U-{PF3K9<pIm~7y<X+^;{=!
zdfi-fa>}v8dpO@MbQCR?iqASWUvTof6DL#NUtm3FCGli=a^aVCy~mU8bkEv3MK2;|
z)t>Uh{=$a;<Y#the|oe_qVu)<(jQA_zF*~Xc8$>XEdu*nw}qYEu<q&>iIoB&fpI#!
z_X*3--P?Hj^Tl&dZSR#Prf63=OEL!snWz+sEZK6ys(J6#QU%SUS1PA41O)a5Y}__0
zmV2Vuj46D#*55FcSf#vi-}e>IW-wJ19Q<)iX-4IBzts}<4;NgoOB9oih;CC<cQ^5?
z{%&#M{(@JpUYLLW>Z`vtylr)w)z=@qzogH-7B&5;{q;g}uVTLY)`Fh%-gb-L+^xNo
z`SsF4v8uN&xu=DGow^tEJNlb)o>7rp7U#dry)};3sv2n?qGx(6PrA%AoG+rPvWSOw
z`SS$!r>WN$Jofp%utd?-`&`%lSG%^aDXz>-G6Tj0oBWPG=Y{49+)SJT4C10}qMTrY
zffE>8K*05D!M$d0(_n5U4xqF&o3tR9fJqCk@02QC_v$b=6B|%kh)swSOu(e&Z~fAF
z_fdNa4+FOV#5{3DRz(>GP9{bMNqz?V-yh$ucviWdn~4jkTarzZ9ZWEAvM_^GH@(pN
z6LYb54i5t_Ot(C%JY4tG-R~csxjQ3{hk*|!Ez2qkmp&Z1?42fGY6=en4@_E_RaqKl
z_p3yOUAiW3jk%dvfp&|riE)Ann9KhPgl)Q;f8-fA6FX2^m`#`yOu(dzQqziMm=K}I
z$OR6y-b23^ta)gy01GvsFeu1i!m!W+3WI_QCJYNDMow_(z=UC;15^zP8kn%J`J{X?
z-JS`IJmS(8(#|leLi~i(Ure-NW#JK*Rg%?!2adSBlDq~yB*hh#6g6P>@rWxcDyzf7
zDYcJPwLz%Z31on!tSgcMmh!Gh23RV(A{k(z>;f|YrUvK@P&C2Z+y22-#`1iR6wtjw
z0zzW2Fo2m4R1JzXnCdkZyIBKr7r6mdOY%vI!NOplNl*%_>!Q~{VNq34U09-pSq(G<
z6t^%#o*J-zVdod`26;tK)&gb~kGQ;^yamjeJmQLaiWW#-(N#8wnFmW8z@P`E44ApE
zIo9+uS#1CXy_l+)0ZbJvtpHVn5(-RpQEE;iC>a+QB^Sqo7(yTd44A|P6)s1=Y52#$
fE+Hr)BrGb%AOe$v6qHb&C_+>i!4yL<B@j#i`~ct2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_nt_instance.cpp.CF459381653536CF.idx b/.cache/clangd/index/gemm_f16_nt_instance.cpp.CF459381653536CF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ccf9f4f0e35eed9f8ff6f0fa2c0206b3043c179b
GIT binary patch
literal 3584
zcmZve2~-o;8pmfQ31EyHn+8S_K!$(<BeH}+0|BCe5Th0p5DZ{I5HzxgXjuwWAQhuU
zu|*LP(4$Z5f<CuGE8@OYl<IRqg|OER3l_xUo6vKedFQ;Glbidy-<|LK-|v3+PL>4D
zpFcDKg8Ws1nQP+GW0(*GO(DN&>8XjcNC@UYP-aEL(lzR*9+Mv`9v7`iSbVQ!Z;Nd6
z1>cIxYffOZFPIMt?sWv(1b=T4IDDXKbVR#JV{+x<WJ4!@c*neYT|@u=nq|L^&MDbD
zHKiVTvo%|}ZKl)bV|<5(Y@fJ^Yh@=mjgxpe!%?fd%e!8VW>0S*r2Xpyj6|5cEG8en
z^=II&wa=%YERL-BTpsznLhhr)9z0(6%%noJ`?szJ)~VtCL-y154KTOKBU`g3<nP&X
ztM}|S;n0Uiaju_Jy=8a1mhG@+mv%h(eQtTKXoUFr@6^2&Q8g;wyWUUr@Ueuo?5guC
z^S#bj9|$jy{}{jekJ_4NrUz1z_w&Q^f3K=ME}mKzk{@&@)b1tpcawWrWLW%JmlKO3
zb_B69R<(Z++H44pv3av@Bm6{Xbb7nP;$-;d=EMKKTKJaz>S6biRWWt_=H6Y-?q{OK
zfvbBpjI#DH%esr@EtM~cdkZaEi--4kzFipj?&#9KK<T*rM3;@@xm%Zu7I_F+;%5<?
zDotRABOyuieG~KMY&!uRo8}?vJ8$l07M~YAw7LJUuIK{Mrf_JrP<>>bHqtOAeVW{Q
zB&|KGc`bM1_S|;W!I_s%H$S};qIJvIP+V#uZJcGUvi~&k;dZlArQ>c)=@iC!9P#st
zC}Q`bSY@s8PQLY{QU}d<^HBeRt7=(7r-g0PPG-?{#z_Z%i{jg-8^86aD;;;z(tE*@
z!iM|oS;7-DPtRZxcFhx>!oSp6=uB%bzOPE!U)>Z_aN$|lQ6KKU$!h7#qKNNCL^q!P
zQoVH3;X4sK+)oh!kIuEmXb<?Bhwjd|D_hA;_wBsbeD7}X7EBlA!TagmqMdB-gHw*>
zIo-Hfq%$sUxoVbLILGk%k?WPUZgpqw7k(p8N~+JWGJJmXcAe`a%yDaD%ZsRnfO%&x
zW?E(VnY5Wccej!-Eq(SXquSMiC9>v2yoKCwLDLUL>vMLnbh1~nblDBRk2e+v1P27|
ztI2jNbe~fu36Xm5FFTbdzG`o(D7xG8PPNX>*WlVsx!;x1;_Dbu!kOp6y;CKq{3-W%
z;Ljt6EA4vMRV3_7zXZG7DvljI>|W1T8(u50y|8_ES6#<k;@5cDk1?k%%=Z?IJ6KYv
zvJENSmGF9W)H2#*H1}+tMr4!ToMCIVwWIBRTTgny+ZA3bHf8tsyv}_4o`21=z2C9@
zmT+Q7VsdQX*OuP%w5+2uiiB^nw&(<SW^ixbUlTiJSq5I%Kxg5{(fdb_Cg@U^Ym75n
zhulAzWRzvPJxSTodk=cndEJ<ub7|D+;C~*5j4pU%$VT?Bf1z@k{%Tjfs{8qil9yup
zzOPerRS7+>JYFlBf5`k3EgWiGKDVU0tGXh-LU$#8Cqq$atFqT6RX6<Bs;+Smna<V*
zY}r`yq4~NM`|$$a$RP$7kKj5~e})dS5}~w>YhtIu5afD#MU2U-)Uff0DhyKc4fx!#
zhP+kAKBB}GgY4`?!FVK&G@BckbH^HB_PK0qR}*^3cmlM9G@F{6T8}lrEVX}Mb-<<4
zYCN)%G`ktNO&e=~Il6Fu{lB96ZjMJ{NwcGYBX_I;=F?|tk9#a%(8z&8NwYu8AC2a*
zcqlXufuU8)pvDUIU^(t$#(|cSW)Vw-Mo(u=|6+b&Z@RN-Pg^Yq`ieA5SyD9Gnq~dP
z+~boz!NqJ#HU|nJ&2pC9GTN49YYCh~sd4dXm4*;xz&oe)-|xvc;wVuU)D`4|Bu|u}
z2y?+)K`J>)%o8Ih!d-AzuvCr`_rwW`2p7T?WQn6BJP862`(xGf0-5_)95R#&Ra2qR
z6f}!QIj9Dv5L01lDl(RWWzi@H)4&wsDqKwo!Bg-o8s*>`m_kBDsDUn0WeSl+qZ~p5
z1DtYPW3m2NU(;k#v<MYbx=;x!rBN`dppg=dr3e#YVoD??!K5?_#uPMCVzCtAB3w)<
z#wECvM!~p(MoK)EB0@xnfqZgp5<*I&U_wD7B@qh<#6{R+$7SoKk*cF@-h7Ty;w%w^
z!Uc;WC+SZPHObf6R}9J*93d8a?C4OFP%Vs>)3;U&W99Vy)53T;eUr2>QT}Bod`9|T
zAL`n893^+68%BF*j?f_(9j0?Jf@%;TmYn|pL0jnDfuNmq9zf7RI=3@W9fNuv>xQu&
znow*A#)j!!j9?lBWRN^94-l+{&K(HWN#_9s8>DkP1Jg06=kabB@1Z3EAA<2=Iu|3j
z2BCHwe}LdEbnZa#PC5@D_#mCz8Mux?Jx_GQM9-H*lQlC06T@^aMhFc8DuuK^K!_GP
zcOXP3od*zNkk0K4LdO8li;G@C!vpD0m`o+TY?%L3X-s91bC4KZc}%6kS>X$AIN*q^
zLj4U)6&QU|$U<ZLp6}<fm>i|ZMkD~m$x-^-_zOTibCgmWsQ^UbDCO32dk_Vrhm4}X
z$f+oFVXB$A<sVyz?~_qbBa}xg86%8G+fXB%2mF!gGa`6k4RKnQ-8aT*etMZ+A_+~V
zm%oH2(@QF$$rQ>1fXob8!0nk~9*ba(($vp11RTD+>~r5XzHB?MA6{x21%{9OQLr(*
ztwKwNGq*O!D7}1ZK1S(vw06WOy>5IrJD?Xxg4tv5FOU_Wn;h1^x4`hF5K*s^_vjiK
z&X?(H17?wz3R7YOTyc~^%pe<Tbqc1!2CNQ@l796sJm9wzXAKxmnz}`wHD{5K^6T&H
zFQoiRouxwHf!tgtxhL=oeo4RjHzn}f7sJ2c<o0?F=@+#_1yp)4J4`^O2e-ooRC)+I
zLIBe9_vEE@t?J8NdIzCs7|lT_7DjUriigo0gc4!EL26ubjQ;OJTI$NQBz-4B-~985
z1|5i@+%RKG0=9EivQbQUWW@5Q712=S7b_G_b)%^6a;m$6>W)SG2Uo69B`1>Ig{u~^
z3=ECNvByuCILUOfiSd+RGwFOsGfuF%<<x2D^ttnVe8qnL0j^vZkv%cX_WzAq@c7m<
k$fP;UbaKY$xVd|JdD{rBW_v8)S_H}CnSqNXGKK>BFC^DayZ`_I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_nt_instance.hpp.CF5C609546474B1F.idx b/.cache/clangd/index/gemm_f16_nt_instance.hpp.CF5C609546474B1F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f07b2cbdc53e601f174f7cf58e00bba9489d092a
GIT binary patch
literal 1402
zcmWIYbaN|WWngel@vO*AElFfyU|<jg;^LB`%nn8d25z8S#hkq%zW$F51itz|WZ>X4
z@Df`!HS{ZkkkInjjdygD{iRp;9j}<PCG}|ieU;pW+;e_R%BlP^@9p<AlWi8a{zrS9
z3p*v|yUaxHk=l;mY@uf|YZgk@yZY=>R-e4rS={1v!UXyLqZfDc$#oa%XFUtrl~B0y
z(>+srzau~DQ<U^hcCSuSSywmZ(e|``YlTcVd)>L<^hfYkl<DDJVV9GZIEA_XEWUZx
z{Nu%&+B<CI_wL+#e79y#j%&U~a)MV^n~w~urtY?L3HhaSCrn9tW#h%LVujm^i0HHH
zT0OXZyx8A{zmed%Ix*sYWk|6PqwS7_nq<K<d!(0#_RKH1=)K>dn|F=w;R)*QCVti5
z4KCbY@aokI^N+hOdB2|eaMesd+q&j|ZJ(=LegApceu+pvaXVC3=JDrg{TkKf(@Ub>
zTHSEcpSvPA+Ufqy*_-yK**{WKe!g(}1JCU5thSyr)=$52wkfGC<lvbZlYO2@e)bG>
zQaW{^plaP+$I6SmFDx$qjf|Cb^*+~C&+_YTTybS?k{K{Q*yMNgIWII<;AY|!V2}`H
z6XgUG44l9?0|KsJ3+^?0n+9_;aR8;I*`x)*1Wa0ReWz6Ex>tv}nb?5RLTo~uU;-v>
z|NG<H70)Wyb2D)Pr6t)U*}((@Ckr#k@GaYxtT&5xKgh$tE5IP`%<3#7A<rr=!@$YJ
z$RNefaBxok*O><TqC5<IKxsQxI~fUCR#~|8eRYv3hi5-r!o$D~l=fitkdaViRfJ2w
zurJ}&7wxX)Vc-Eu`?C5<ODMA{OT!F*m8h^w*W|4+Hxnz+X<}?*oL~ax^1lLMo9^Zx
zdB)Ae4wM#V6XpaHFzKSyv|<@1MCdVcfkUnL(C-Cn9$G8FLJcSk3Nn~5EVO{aprC>Y
z!$Jut3<@Hcu&?=~d@<dg35+}v(iYOrFf&8^gw$V5v|(l8k&soA)qn?*guIfx20TP1
z6qOV;;2|WTtf;IG3y{=4R@DZfVkeLRma?u$23X3wA{k(*=!#^3g|Z9Gfc6irGM49i
zq<{_+5)cxD`D#tYZq|U@MQ%W0Nj^z2n6F?aGID|=7v{TtCP68zu8Up+Rg0>M>cWyC
z%r!vOpty#qhB*W%42olz@KXcUFYNr{-5{^%$y&fP^GL|+$y>lI=8;g;Q?!7Yz$2lo
zt85PQ3XBGt3rZX?b76@C7$cyR0TX`Bv8JENY6B=n#8kx$V5(qg1*jU7P++QyQgaeP
zsl2!-xi}ug5CRbpAOd4R3P%W!N#b%Q(;P2>hwSW<f+9k~qGAlf2z8<erWk@LiC_W%
Db{o#?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_tn_instance.cpp.A27FD0EC622C1968.idx b/.cache/clangd/index/gemm_f16_tn_instance.cpp.A27FD0EC622C1968.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e5ccebc655debe338a8fe39c59552eeced432bfa
GIT binary patch
literal 3584
zcmZve2~-nT7sqEN2}_KM2xwq50b~d$aZr{pXe2;15Mq=91_TE%pt!M0Q4tGNAQhuU
zv0@Pwm!n7Pf_`<Sb*cMO5!AY%kg&R=s2~=<2|dr5p3|I@m-l;j-hKaj-+gZq9x`=m
z&tM1&(uU+LPRdGPLl88K`etTjq>Q5=j0ZtEm9<feQy=+Q-BUj(UOa#1?b59c%GIX=
zD*sq~81p=3*DJc+65<s0qkTy4?hE~W`sKRe7tdOmdkCXjCZ93ZcJ0_X>-YYNrCUd&
zpFv)&T}iGR>+$8F(5-gml%yd`l!tleh6?g};}*16v^Djw992swx|Rl8NHJA;LIHmL
zUC5>-PevUniLLxn5&NW4HHF0LAEZ4VUMb!3XIm}jXm8hE*OA-0+3QrXjk$vheqM9E
z^Y}VR&xiX-6TW2lD{r;U+Th47YpK6GsbZD1k2w8z#@5QXom#=0&d+DygY%bgx1XF_
z;CHfmcXXj@bMk`cyLLXd*`2;<hcLR}?)IufvJvGG1)(=1U7Dc3FL;;7YLbtSKRhF5
zLntSEp5cSU>D#abr&npq;D<(wW9!{!E`qPE-uJ&tMX$Lp?zM-{OQ`9x^KTP-{~9j~
zS<tCtl^Zk;HD@ats+x$~)9o8edVltPJw4>jfv9I83d@3&@yiDB*UpyC@R4w2k7Jfs
z4Ts(KM=YEgkdi-f-C^k9NFV95lXhOV$@%d;tGhn7#TQDKM?(uFsr%FPvE~U`BUO%l
znTFiDCHx`lR~fW>#-2Y`_vn0t-Yfgtk}`Y6xp8({*Uv-lt+zc|W_fj3t0c*1f1qDX
zF}D-Ns&-j#6gu87bJP7W84c>bl&YNHYVUkuBfIzt>xf&BeaVeu=f3x;DYHD{;6E+A
zsP+zboaFG>V`DgkOWoi{@NYHtMw?w{-)>*HqxwQZ;i<=(15^0htWp(C#W6qhNv}To
ztvYJ?zMC-{ypIyW_fIq?=ywO$MQ$l@DWA*F3TVAucl%b@8q65yBlzXbjE!9XJ;M&>
zdtAL%Y_u+GxMZ7AG|~LZ{wr0xylQ^EQ}mr`;leZ7qs^aOyHPV?DCWNQT*K42+Th8@
z&*qHI4jkTW^Td0!ob51WD;Z}<6@@G7_6nx+qeT~fvRIn8fn!v@P-ZEw27bCaGdL_b
zblc9AUPay$%jFRY{~hH=^JSM@ZPdlLI^Jm0yaLQFUsDBM>@T^35v9DzKKz^8MODA7
zIuvrcZ(o&5XIf>>)~xgJ`0FK!@ArA15vH17E_6P%eoI?T%Ov9WWaZ|Bqo=0&i!Aq)
z7HOR$$~Mh^+28LF@6*5Pc)m{RlvS7QJbG<Q^PT37tisoG{N^lQ+12qf=k;6RWnV*=
zyWzTINJPq_#AhoU{3q)<2gVdjUgfSaitwDU&is!<T9vtGepq*F(Wm}92M)|PX3W-E
z=QQ?se;%G)p5yf}eM9GM=yB^6Yi{28evdtW-HYg-_R5@#>{$9#>oMxZrZd|1Cr?Y8
zWUkLvWUSK8?|9+!QeF2`&O5ZI=iKZ`rPXcKmC2RHi^&^V>LO>Yt8rm<?VpXQJKdx<
zp8DW5%Su1gT^Y@NFip_6m&M0p_-@r-pgpP-D0A83#1SwA3Dwgz$KEEj^Pq1u5UJoO
z(0E6VBMMm}Ff?x#G~O)j;ahC;mO+S?G7HUw{DFqLbOt_n`nVCB?ai77A&HdP-OQao
z&;WCgqo(=du;<}}kR-}%XJ*GAXn;Ay7D@?ur%f9Sg;Qo5I~&J=2AJ1rmcEJG-&W6q
zA}MnaCkTxfa0FkaRgzZWyKGlzcu)joR&i7g@y;A)2cZ1KW`pgDw7ZuFA#*9Smzmed
zfd)ALtL#Hl)RMsl9u!5Hr5q_5KZ-Nzt8*wLDLJ#s9D>XQC-gx(e1#S~5_LgEAR>x<
zQGzDS1rvc>@kq=UBWS{1a1lrpkHmd(f+oU+5CNq;lJF%6K-^E;pA;&+zu{3!Y0*@A
zDKs6;Wl$cfgK5OHSSr0TmX75zC=b)YG~!x3l@@}h<GBpV!*wu?gqBDJx+s<DL@tB!
z2ptS?*dd+$(u2<~SW%*-sEpQy%25S_!caAXNHmcqOp3{9k(eA)FenUDGl;|zX~Lzr
zj8=@xaRr0Ia5aNSJdq|sO2~kGDmFQxU{DyLW)Mjv0s_Yd<_?^hwdeb`A87U$0yzkG
z;Fy469w`^gWuTtH+SEg)6GJ}~Dh`za$5Z;cUhZvMb_k_1*#@KSj6Fs_!e|eZixIR0
zp)-iyLC|_88xgdH$?p)fo5{^A)WD+WW9=~3!6+202gZ7tT!LUa1cXXedp&|RFu4W6
zTAAF9VDFi1U|~iUJs)p}@eW3{;XN?k%j6OS*C8Nl)CsCb@CGKgAb2a2yAk|7lMO7~
z$fD;H?J&{twPdM1?}3S4CYK<D4got&+3OLafypfh(aPj*gm}+n1B)=Sz<gQp3#hj{
z>mi#>28aV>;ND^Xt0dSYbMM)tTC5HL_Z!$4s$5JrC|xd0ci<}Ja0jkX`XN*gqZN!|
z!Spaz!Prb(591Y#WD|OrsQ8+Ui0Rg@9Y0Ruuz94^Nh$*3d1R1NkO<T>k5o7*M4)ha
zq{>m{3dC#2uB!Od*n5X^fLfpe5DQf_7MOrh0v5P{u|XDu0HiNT&vE(AIy2Cuz)vb?
zDDVrCGZgqK<O~H8RWMMHtqi^~R^u}R=8-mmHW46E`72L+f3B(dq-i+HCJqb(kJPy~
zy>LJtfir5@^r8X7`=1vsi*ByeQ{wC#?Jznzp`#F^qjPt3$LQ$1gkCNnI`B*Ft?3N~
z{PO-)0Oup5ouuIYWonfGc7PL|0y$gm1O)TQP<E&jodPvm?F3YV8I*d{iw>;PigUWn
zhmKfdO08Hb0Z&nzD-M#-eidScg!ZfQQ27GC&k}^EJiJ~`q#U4P)RPV$6Jwrq__!GN
zq{An~geM65Z>y-9#?(LBOb#MZ4Z}esreQdU#5D{Dk%R^~$Vgh0V0xoxX3Wi8XzIk6
zntwj=paZeU(~TLv!8dPNS|mir#>|eJ6A#6HwL;N!H;(Skrn__K?!apQ;M~R9MJZHw
z`n(w&Gjj_|?x4X#hT2#Sw;mQ|tC;F;%L}t}7%>tZHEHsc09jyA@C5#NsVgzg`M*Z(
p1wzL$)L#g<u^wW4qL;U?pTCo2w5QKBzI~`FnH@4yu4Jj9{{x6~TfqPT

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_tn_instance.hpp.C57CA23FF5A33CBF.idx b/.cache/clangd/index/gemm_f16_tn_instance.hpp.C57CA23FF5A33CBF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..45c477f2b04ecbfdb3289c8604174468eee03b15
GIT binary patch
literal 1402
zcmWIYbaN|WWngel@vO*AElFfyU|<jg;^LB`%nn8d25z8S#hkq%zW$F51itz|WZ>X4
z@Df`!HS{ZkkkInjjdygD{iRp;9j}<PCG}|ieU;pW+;e_R%BlP^@9p<AlWi8a{zrS9
z3p*v|yUaxHk=l;mY@uf|YZgk@yZY=>R-e4rS={1v!UXyLqZfDc$#oa%XFUtrl~B0y
z(>+srzau~DQ<U^hcCSuSSywmZ(e|``YlTcVd)>L<^hfYkl<DDJVV9GZ6a~2cEWUZx
z{Nu%&+B<CI_wL+#e79y#j%&U~a)MV^n~w~urtY?L3HhaSCrn9tW#h%LVujm^i0HHH
zT0OXZyx8A{zmed%Ix*sYWk|6PqwS7_nq<K<d!(0#_RKH1=)K>dn|F=w;R)*QCVti5
z4KCbY@aokI^N+hOdB2|eaMesd+q&j|ZJ(=LegApceu+pvaXVC3=JDrg{TkKf(@Ub>
zTHSEcpSvPA+Ufqy*_-yK**{WKe!g(}1JCU5thSyr)=$52wkfGC<lvbZlYO2@e)bG>
zQaW{^plaP+$I6SmFDx$qjf|Cb^*+~C&nkB}uDCKc$qX1DZ1Ow$oEMraa5HfVFi429
ziE@Go22Nm{0Rh*q1^1f0O@q0aIDpd9Y|?^Y0wyiEzEi4n-K)dgOl&}DAvPgSFaeWR
z@b`<@@+tK>4+FOVgSZE)hm3?GtD+18Cle!s6hDLg?~iX+JgZ#K&BO)NEy*Uy4kj2l
zS(rg~_eYd|h+X~cG7keUP`5Lyvy6m1t2|uyoOpqtC8E_4JPbTQX<t@fX$fUkWoek<
zHzs|15SGXLh=+j>C~e1TCnF)tDht>BDp6sVuE|?tZYEZs-C}HFoL~ax^1lLMo9^Zx
zdB)Ae4wM#V6XpaHFzKSyv|<@1MCdVcfkUnL(C-Cn9$G8FLJcSk3Nn~5EVO{aprC>Y
z!$OIX6C64)VOZz@RfB>CChTiIDPK&tX96RSgtUdUGt8<GKOyxO6Kz;ocqC+%WHsQv
zmylPI*MNtlgrbt72FyMl31vlPbyzT__OYrq2o*bl46u}SMKZur-WACJOGQ^C11ywX
zU<R~*aFww<-y;Qdn2>;w7|gRUtw7&_Vh`pGm@c3&DAr)YYbtiL2IMYs0~#X9Cn*LC
zgMB7JDXgxGUIT?iRYi4Si5BKapdp~Rg&FeHfb|PIzj!yuD|)gPFjYJf@_O<XF!%6C
zDC#L%AbCYs*&OC5SmFQ%Jt$?s%ze$Vrk}}b11RXlRK*Nns$gjas2Y?|V5*B!a}q(x
zxVR{}I3C0h0uc}(0%Jf5M+lEe!sK4Xw5bgYf$WlkB0|EVVhq9vb)pES7=kH@U;+R&
Cd({U3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_tt_instance.cpp.5D254C2DB3D84C02.idx b/.cache/clangd/index/gemm_f16_tt_instance.cpp.5D254C2DB3D84C02.idx
new file mode 100755
index 0000000000000000000000000000000000000000..95bc6d2479758d0409c30ad8d7fc442ca4e58415
GIT binary patch
literal 3576
zcmZve2~-nT7sqEN31EndY!Vnv02u-bjK~rOjRc4WLPRYfs{<Gi*T^CT%ThprRE&zE
z#UdgupB|_S`c<XUy3~ECDAl@vN?5HM6)lM3H=*Y_<2hf>$;<oAo%i1V-Fx4=nIAN3
z*6Yy_6rc*qT9K5Iz=j}b9Q94lNLw<6f)E}AWmPprtWZDnGPxUizhp(qyjx{ETNLXW
z{i=RjaRi&%XgMgl)e&SDve`OlaNqgiA?+HC*@ZJEhAzUWjv1$QO?|ufEckVJTG`I=
zsi%>b8*<5wlU+U?5;`^I&P*D!QgMWLcB~+8Fm7pgW!ID8+zCyDyl-`&kpxp#BoyM;
z-UMx1**4*5X>8S}%GkCl<xCR0bAQz%vnt7s-@BSP#|HZjI9l!MXKz%-w&sj3{9*mI
z-V+<e1Mlu7xqnLYRov`au*H^J-f`!T>6Q7CA>!wcX*;Xp_NWA}dq13p52dW+)|^^Y
z=yR%eUsRFu`{bp6?cMXpd|&GF-NLBCKWeHEOUGA)6$al3cX$GQJnvZ%8<~8<?a18d
zEy0}3#qICJc3+1i*u7kp4L{Hs9pB_MZ#jH*-NFA}EPlm(ez$x6;)I4iOW!V6&!6I@
zK}&lztcvzXn}#!$E!9tmTXU>iO9y}Oel;iPb$!IMAo-}mC2rZq{0%XZxn5$9^ilMh
zYBSiWE^OH>za<6JHXeZvS$RpGowD?>NG^yUSl9PYSA3CVO%$|Ltgc(7jWtZjuu|F%
zrMKraujG%}l;5t}Kl$A8=7;CPv>usXmzG=0&rY#aIer*(caz1j@=;gDb&8X`>im78
zOSru#R=wABtI+mdxs&F*8E8QNMYSTO)7t+0R(8o{)={Sb>(cAT&#v`qC?9py#&`Do
z;-=f&DdHoOk5A$d4$Y$<!oM_F>&*9_`MYM>?%MMSMU9Ul>u2(JnW*JYN}|6Tl3aQC
zOKrrOgEykLcpf7H@11N-(C+iI4Bt`cP_c-g;n#Vq`PR*l^_VWsOYr0Cxm&ru`^OzB
zaJh1|L}yywa?v8Kc$(qmy35sjJsN(xUHq+b*|O7_6Ajz0UT<(8i#cyN+wwTBDR9P#
zGg%We{mq`5w|P#Kv2AAVB;(rEqWOyE1A;mHDAD=tMyvC-aCD02iVVdS|Myqs1%?C$
z@7j~=QS3RbLKY_X-Cc34Kzh;9JhbFy&ui5x4?lxTSC#%3hD$GFL>X^}7ym|$sQSnJ
z!$Cg}9jtcfT~(E}E8`sOcC9q=?Lp7eLbc(gBKyWoJGvS=rW3y=E51)S);P;oG-`iY
zvC2NId|S$k;bEJ2ui^X?1saK6Msue9#0?!!Z$ItHD0;QfXW^RMzMdCZul^QZ@^0^Q
zZoeiT6Sib|;<ImTd}nAm^^;1(FLT!GM0i$6Z^1ufIu$tvK3IQe@%!Q1_4O&bv>1(P
zR_lQ02eZtIERP4NTY7Ink2)`#a`Vm&yX^n-ZrJebmxf$q_v*(gmkH0eomO?XJuZ79
zb$s?sTD~f!=egI5(B|z~Z_wg_voX`lYP)KylB;wVlDD!#i|tj8x@EOZzqhLQI7!T>
zY6I72m%VGgJdt~UwqWQ0i;qY1ooYWpF$<T-)3aA3j)x)0v}fzu>c&5h8Y7D$A=1OZ
z!)m0VZk3Ub#bVyQpc~ve9uz^DB^(JFKY=p=g+?JTw0Hp|v+A3E^rHC&9u!8Il^mr_
zygkR>2AEUg9{P2_CJ{X7E6ObA$kBLPj_v2s*`Er7*1ii@8zV`S+0ww0Khl7;b7S7O
zMc>Ft84b;+%;uKnwj&KN=bdStR(s`*m<NSZ<^WCr8ZY1oK2NUO{2$-``_9S6h>9``
z4TSuW28{k#lGmfI+q}&fNu<oq2G0DE2AH9=q~!DpLkKbuoYV&F_7?JZB<g^QfMkl?
zP<IA-qXbQu1119N^GM7Mb7zn@M$m*i;36P}N8)a{JA=G&f+oU&5P@LvNWzV9XOK5R
z0OG!{X)98Aex(<zLe;b=G!@NZP#&s*X~a~RnpTCSVmS=T!!$6BxC&R(vhY+qhe3I`
z2Bwiv5o(}|(ws`<Fes1EzyQY`)>yAT^z6I|C0c??X<eucl`|*=4P_9CCenmSFexn(
zlVNfOg<zo!BC$l8a0xD@72`5o&Y%!HltCn(NE0C;q(DBkHyI&kPzVvqAd*M~1bz|j
z$T8b+Zm6dIsjrYnhPsCOfzt#I9n2j$ap0wyLwzp}bY&k#DJONqXb&TN=m3lkGPx8%
zH3;n-^bUfyFu4OkJDJ>%pl_Mn&O&u8dOX$*V?B%<zy@G!kjbS8ra@?@Vs{X%g~=TV
z*2(041bfTmb{3{%(c|%M81G>u5<URqgG??(a1BBS1HXgdElln}@J=T8Blue;x3h2^
ziylvO!$i;LY^1_D026~uE=33p0)j``?;u1AlRFTilga%E@s`Q$EJDWu<E16fp~3!)
z2W&R!=jtZ~;bxOESD6%CYiu&uHCPI6Q1%xI0uqHv5dEDC5+rPnsbkOP=^QqXl-Nl`
zU|t>>U>6_)^YTc!om>Rw<&jETr6ZV6H95cXee2+DY68><6)@s&gb5hYF~S85-x(1C
zu(KpB$KhMkbbq}9ABl{iz$ZY)P~an%F%*cEfj|LxO6^QvF2K$*6@k|$M|#bLd8E0&
zc^Ie+Kp=IR^wk5r&HKWw;AIYXq<qjXp;{QNWE>hy3uBdxq{g){Udc#0LJJd>pOXUY
zhEl4pJz&~`+>>k1K6!dduQb9u4lG{xS5bD<(<&|H7)x7A3<&3uLR%q5d)C?38KXVx
zA@pzno;?>P=_LjCE>RQsvHk35&&t>`JD`O}2D5|hfDfrlg&k@KJPQ_~e5EgjKrio$
zr0c{v{f1-5uh+-VRU)QW32+S%)2ql`<zjjjrHj%Vtn$$$qM=p&TNgC}8jeQN`^3Vr
zNFaeKbm4d;y-y;Xhy?otPbr!DqD;$Vo(8HqFHg{y>h!cl>C5z;czyHlCmM7hD3<O<
z(cQSuT^31fe`AzX<D6=g5EUC86SpuPiur7X7Si33)&9jrD^$yuP~AC;=W+}TjYe^e
zM~@k6ZenISF2q7U%h`e#Vretp3Y{>0#!Nq{e?Xu+-%a93OtJrePpt()+ey@)1*gd_
fuJ|+$Pj4SzJMqM+UbFeu!OCQI&^(!f6$<?i!jrc&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_f16_tt_instance.hpp.A9C0C844390863CE.idx b/.cache/clangd/index/gemm_f16_tt_instance.hpp.A9C0C844390863CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8eeedf5c0d31dfdf28d243c9997fea5c290f879a
GIT binary patch
literal 1162
zcmWIYbaQKBVPJ4h@vO*AElC7283cj2xTGjEh>?LI87NROXK%>Cyh8>&Z<oK>;9HOo
zwwl*h@X-POH^Pf1>@ZRC-DIJ=A@l#ffENx2GIKool73oNo}V^3r{Mog<F2laVn9<b
z{p`K^p0`N;aEcnoeH*rk2VQ=87Z}}j$ZN^^M6(}P4<@QjdBnXYJMLMsobvhZ>!nfR
zj~_qZT=3Co-;8~|%cpu?ODj4V;o_}VaJ}=wIiqD8t_n%M-fA#y>MaH4hErELL^)MG
zwm-Pb#Vj5+gEy1uW!8I-GrXTv*}B=bOkEbpQ8e?-%1{6J^{lRPbKgJUXH-%;$$e$(
zMS0m8=CIi6&EL!C#%Ui`4fTzb+Qax?&N%7utRM4hGB&tL|DL<v<e}nIskFt<9b3+B
zY_cm^)$%==^Y+Pm+yBb!b@O%Ux~p@L|N2MnT^<i>S^^F_oeC^DCvYq<&Y?ih$@yWx
z4QE4dHqOZ}^*vnE)?U{7&9?UaqDRG*xk)v^5Ef16x$b$EYc&r88wZdJ1pKnBvN8;u
zOpFYI><q3FA8j@s6Y1q)V1Y?1vnor&q%&e~J3L_&2;*U3hDj^3D#CTw9k^?}=;UV$
z9tKvJv^=XkT)HSVtyqSMfq_kaN1yXTa|K2&eo;12&fY`67p!?`t-$qb!M$d0(_o;m
zG@G;_OjvMzr&Q^>SBHVZLTo~C)i94Ra)LboQ*HnI<J%R_D%S&5OR`C_!-Qe}0;&f2
z2PW)mJ}F;Jw`T$)55Kg9v@^^FA$~&YFDBZsvheWBD#>cVtm5I9SCZF&dy-#KNl^nH
zBK*pV%IYv{Qu|m{8-$9TKn7UKx*{21DesD8fTf}<k^vUVE-(XNqCoe80vP6Am{y=L
zC}3g2?H^oaEYJ5y0lgt4AS4D0u{9ODSp#wxxdDYG`6R_)L9@>!D23H^(QBZvsH&(g
zEH<7Ruzq3Z7w-nSSx?pi$<2E57BGu>_!adOEnu$W;aApGHix<SRieT!U6Z%Qz@Qgn
z6N5+8YmPPjOja8}VJxO9W&l(5S0HTD-TWiZfU1SrgyE`-QgaePsj|2zxi}ug5C9Pn
rzy)J4Fz_IloCqc_g2|0w^1+!*{69itlPv{~7O)F&@^EqU@-YAaCNGQY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_fastgelu.hpp.BF1B6FA711DF924E.idx b/.cache/clangd/index/gemm_fastgelu.hpp.BF1B6FA711DF924E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..14a85bb252551ee561c3c8a46f49f6f33d4f7d3a
GIT binary patch
literal 2688
zcmYk72~ZPf6o8jN5(z5_fv^OSETCc_frR{+0m31dToDzyWUK-r29h!fG@O;H3_=T9
z%V1ki#8SWu1iVT;prT`~sFhl)19%K#EghwJfvt7}v%CM!Wat0(y?y)ctc;0@D&`<a
z`1+V)U3OuH1%e<p#IK;xpb;_<WSo!{@4eozCjXwyu4*#Ta6-Mu?#ZvZa~-^uLZ<rO
zpytuAVW0VlN9|MQ_Ud=f6`MoNTaPrxN4xsAbH`M{V{>DM&rc8a4~#k7*j%D3dlj(g
zy#kRBzpj#vp_lQ|)3{2>s5*CJ^?J>xJ@uDU*7KY8Hix`)i(5Lh=hd;3DPALO1@(Q;
zWM0PewXKs`yn`dpTk_xk7`Oe(VOMX?k>eF5Lx=x5a5^1(awMx};P>y`s|UuvCr*DH
za`BSCZrpn+!l`4h;FxS6&%Tx`AKKL1U>zLCzTTAd;rNRDZG*RCvYQs)^}c^1{XhRp
zS2<(DTKl{QK0lWncBp;Y*_GtW8j1EkSh6g6%(bhS=UX?mK`xqVu<EP!>MQj8EWSs%
zCqW$3R=uk7z&DW%9yvQl<2o)$dL-zL`j(cFOPT+UUu<C)+<Y87a3QO;nXkG#kTm8Q
z_on;fUEVgiBQK}fjN{ymBI}K!4-HRmd#?M?froDSXMOJ))y8X0Ia9rN-kkW5=#98n
za-?R0?N0_L{C!K!b}f&L96b{rA9rZpr`26Py`H>kuZ3aO)eRG+YaJ#THO8~IezE8_
zFiOW`&+eC9O%O}oZf!nQir$Rw!JYDz4(|ddR<zZ<9jVzBlpAzJ%FwAgnY-n8CRSUq
z&)+^5m7KR@OV|C__ICC}n0`n9>7_?Y!&sJvsg#oPEx+YSzRlb3QDJs+#X}=fP@>Cx
zNOX!hbNA=Qlbs3_nad}TxG)J*5{p=}7>JXN8D=KyxzKs>OW{>32FDh`(l}}SnHw@^
zR$HMPo8(kQ*|8W_d`nEtRxg;jk>c>SQMm>g?M4wUX`avI6Xxew%$na?9~e5lu2_vC
z)}(?z)1Odan$0TgyDPBBP;Ta<2#-_{F-3#|+k94`X?gdb9y#qoJ_98j%;se~&)i54
z5JUEcf)^}0QikPBgmr!I8jn5<#8Iq-SCT{;VTsOc4?n2Hc_?NJY?YMzm>sZ#$R5~3
zgacejxrw<0cZj@!x01>+CId2vf<TaxiXYQTv_7O8D5hPijipH^(Lpl3Tn}iOL2iJ^
z7;j9VWkvELNLDT{r-(D%IJ`B^FA+sCC*Tx92kH!*AzB0$h0vjiffyn`;1@!Nwiqmi
zC=7&!(4mEcaEOvpld@);2SHFQIW-xg3~>e^I}XJ%#hDb@wAVO(a{Sp(gu@uXpzJ#U
z2Z$VjBSiDTe5e)+zygRofCof!AcrdB3w$9006-KBf}yH~fDnk5fF%@>VK6u|149l9
zisj7Dk&q=KpPF%I#)fQ4vK&OOn#P5EjAHq`e5zMhE#rGU-Kk<SiC(p|<&y897;cGk
z>7q$3rChp90hR$=x-dC&b8JahDx#-9`by@L#3Ib$whG7`jw>K@xKsg|!vhqwIdz0O
zg|2LxAZ-zu1X70x$TZ$`K%01*sDLUDRg|h*`>$|OOe4^^kd6`LrwOJhSC9+Iv_h>L
zEz<&R5KTIv4wC6z^d7WK@2?M|$>3suWJVVwO#Vhlv&f|gl9joX!Q@{CX_mW`Q?hL@
zBAxwj)H<3I!FXz9^oKREn&mX*dgf9xGUY_dObd@rSs|~GN*Vbo(LbimN48Vn<GTLG
zUOOtyPz+_G0<tEfGUiJ*7yH80F#=l=RAK=xBEN@XQh`(il~@oUBFjfKuQ11*{OGd`
z{0AxcTQLX4T!AYzOj00)1||RmP(y)YfglhfeY_r8S4O#!T2LrfCNG02zKqg**)h?X
zzWJ7z*aZS0fYy}|2qAI<ZqT|~2o^%*1AL%$r2q<uR6qr-t56UMQ9^10w5}3U6Cp|$
zr&H?+#ny_~QZ&`=XFDP3D<x)Y9+;=16YmK;RdnK&K&he=9|0m%bmCJ|Q`V7nM6oP!
zmYkfe=~G|#dIw(XB8qML|0R_S+1d=#xxkQBkZ3wkZGa#`6VG0D<N;w|LYD7jz-ZP3
VmV=|v&R&4nQ!>PnK046H{{g+Hm0SP-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_fp16.cpp.3FD2C2DD621D7AD4.idx b/.cache/clangd/index/gemm_fp16.cpp.3FD2C2DD621D7AD4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..547689150871a103e6fb5b595919d61b5e1c0003
GIT binary patch
literal 1820
zcmWIYbaNA7XJBwn@vO*AElFfyU|<jg;^LB`OnD$DmXCp<V$R;M-2BI8Jm2QuT;9BB
z+1p*a8GlaLa&p~`t-N`*UluPsvBlR#e_q_AwEcBc-l|;RB3*EHtAjxI5)Zo?!KFPL
zdkmweo$6}hRDT^9YFB3^D7kyrALIG=`qbl>{Y)|2W06o(<8=NRyL8&;&6k;v9sSAp
zcxC77+rN+O*gMN_-(Sb>{CeNdKd<kOw3`>--*)80=kwP;*ZfPZvETi>YI1*m$LC28
zK0i%W-|=kg)dSi3Tn`J6O7CWy&aHgq*y87xULC927%)Nqpk#pkx`X0^p$9K|h;mo7
z{(EgIsyC-#wz7`NxoIm7y!E{^H}#F9hSw41-0KNUYY&J-7^#Xas8O1`S+6$yf!RX0
zVAV&47O!TsU$WAjBqYUirdca9S0`GaRn6cQ`?1BQve9)vKD<=+J`uH)nJ;^RfZBpL
zuGc1<em_NbVt~Q^V+z?}4w{QzofWxYF+pj<4u+L2!i`^Q^tVjhbU>pf_wYFx8=+$j
zQXYrd#Zxz&%Xbx#bzS^%LYm!+d#P3q4<%2Wa#8CsmCjw9KKJfkmxYYWS{FCHnfbqA
zbB?M^7n7=|>cm5#hEo@_FS%IlWuq+^axiFJpoYp(X3JKIoEs(=KJ&V|X?ix?D9^dQ
z#Kd4t&(fB%J#HdL>)zUGvwrnDU-01I+7m)zEL@yzx4E23+8*i5SoTR^OUT8C8z%p{
zWwpwFw?+EFO(ybDdOvHYZJ*vHb@tMYt%=zWeY(~)@10s1$P|5K6U*^M;-WdM7lkie
zId@1T_Sp8DHJg+h@<d8~j>j}dYzp-_uBy7SU;Al~27kS~!KA1YN!d&J8)seH`bo@2
zx9RYy&SidTixwu_dcQBu)!SNNli$1(E>^AlvWt2GA5Z`MTJ!I6rLW4fzFgK!=1zS&
z$tC(uj^dL^za}MUUXgH9m)tCw?d1FIq{P)TZtj<u?smApzr?cB;(3Xc&(Zeh&662@
z3s&9H2z|4?$2M)fbX()Bp6~@<`Cq>`xxIf+PMbLMO%uuGUd8E4)AsY7V^gnbY`cEL
zV`lGq7lH39wURZ<U!Tk?-5GzkL-ovx*4o!<_p2Aw*nBnL5IL)RfAPOlndhEPs%xIL
zKt5lC*XMQ7Z3pus6<1y24sfY{nQu0&{KGZ1ljaJypXDBX|6uEM&5auk_bjy!n9z3U
z+wAoh57(VGoX5ZJRZGI@B8SgfXD9uURP&ti?NM^}AA1Qt+vgcI0&1ai6!t89ED?}+
zqad<5=Gm7G70LF03okBMdwb8u^|vlfcxm*zZr%!SAFW4)cK@?^{>v0|emn2@e9_Be
z6-m4r&eoIP)&_p|@~X;+c&fPJq2^r$vn}HH^=9s$<M-mYr`pR$i;u`9GOpw)x%^7m
zN1|$~g^gN|fHbT7i}eb&)#Yq)b*gVflNCRi+7@IKRp^DJ%C<$NmzbwVEjam)m8;1h
zm!Y^aH)%Goybzn@X~QzZu8Er|h=CzqhE0YaOaNuTz(D+p#N&5Gkvt5UKxsu5MGi0l
zleSrPeqQ^o$t$^;0)f)fY|{K-0w$diQ+<80c<&Q#rU0O{ESsz#n1D%7baZ2kx$-ZG
zo5>$2EyX6q4<=yJ3`MDF#e0D~m>U_nu(}N>400Gu8_Z2WVUUAh!jXP0^ADCi_{zu=
zuPCCZ2=f@si9ppL_rX-Re{hwtJl`Y5%o8sqAteWMDNnq#gtQ#Y2A+5s2^l$<>tJet
zCV{*HGYRO6oJ0_(xG1^!J&2JGA|N0Y#$aFwM=(PW%uqO!Ilf|p6pMN~_X^?kq~w&;
zw2<(y(1@t$nAo_;ghYmDxI%_B1TzJ}j6pD?5X>+HGXcSjL@?tJ%m@S%>>&s@79pC9
zU?w4$$+;;xz{JDglkUgN!pg?Z!O6|TE5s)#!0#pO=At3Y<s~91DkdduWo>I`@8IZU
zAg*hpYN4fwO;kcwPF_h_MNM5(TSw2($k^1}Qo&3{-^5j1!oxd*$vxQFlhKy}0Bw+p
AMgRZ+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_fp32.cpp.6736B177964E2D4B.idx b/.cache/clangd/index/gemm_fp32.cpp.6736B177964E2D4B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..81a2995254a47cc3824a35548770e461b8194507
GIT binary patch
literal 1768
zcmWIYbaQ*a#=zj5;#rZKT9U}Zz`!5~#Kk2=nesqREFS|y#hkrix%rRHc)rcQxx9JN
zvbVc-Gya^g<>a~>TY2+rzbsyOVvDbf{=B$JY5VJ@yj8irMY`bZRtJIZB_4J)f=hcg
z_83M_JJr?1ss1`J)UM7-P;&RKKgRR#^{K}%`<Y_4$0DJo#_9YscImXwn=dmTJNlFH
z@ygEEw|^hmv3Hi=zQ2y$`Sre^e_r1mX*VyvzwOA0&*!gyuKAZ*W54@%)#U#Cj?a@G
ze14j$zT?@}s|T|6xgHiCmEO%Zom=_JvBl3Xy*gI4F<^rJLCFC7bqB=-Ll0i`5aq6D
z{rB2bRBukfY-JsjbJJEFc<Xy-Zt5FH4X-22xz`hz)*cXvFj5s;P@^<=vtDiZ1G9y0
z!K#l8EndxNzhtF5Nl1$4OtV&Iu1>T-tD3<r_G61pWuxnUe0ZtseIjZpGhg-s0ks8h
zT(3<y{eFt>!~lc+#}u-|95ffbIxBL)VuI3y9Skd5gd4xq=x>?0>3~K}?%{JXHbTc5
zq&yC@i>Gclm+vYf>$>>kgfzPs_foAK9!j1#<)YSODxJGHeeT`8E(;l#wJvUYGxLAL
z<{VX-E+$n^)rp5f4W}+<Uvjb9%SKx;<Y3UcKn<0n%$BVZIX6r$eCBm^)AVe(QJ!;q
ziHX6Qo~12id)!2h*1fgWX8r1QzTm;ZwI_tcShzUbZgV-6v^~<9vFww;mXM1NH%$I@
z%W9SVZj1DTn@r@R^nTV(+djQZ>g=T(TNAS%`gE;p-aEB2kSY4eCYIxi#6@#hFA86{
za_*2w?6K`PYc?r2<cXB}9FJ*^*c9q<Tvc^tzxLA}4gPv}gGo^*lCqcbH_p1Y^^=&5
zZqwmYoy+{x7A;J;^?qNRtGBhlCck+nT&!C8Wf%1XKA!&hwdUXDN?(;{eYvce%$@pl
zl1ucR9K|P-eoacyydvSIF1cAU+sXIaNr|gx+}tlQ-R*FHe~D$M#q$y?pQG*1n<q2+
z7Oc9X5&C9(k8Rp|>9)pMJ>d(!^1ps>a(n-toO6Ecn=+03ran98b+%4+4zJ%2Cf@ZM
zCmKn|cPhLK4m}q1?%L$LuZs5X67-%C$o>18@4oL2KPz74Cu}wmuYdk$>g747RsXV?
zI@;e0keYGr>9)qaj*qK4_cVxly|B;KdH-Oo@1%T(ZD((H-9MPEADon&T;W~6K!LaQ
zt$F<7_P^7TZRBIFavq%awBcFkEY+XJK9daI9zA~TN4<ee)wzp56ns`$I950xH(YRN
z!^2H%JI}mK{CKSH*W*Qw(c3GN<F_tVyqx;&uT9|e86ig=*ZjLK@z3&^(3|<q=UgxM
zemo)-&{n2)`}d+}Q>T2nu;G+*!r|aujyW0n`(lmiE$3e7pX_t#sC%dFA*LY77t5}?
z&oKC+RaoI8p=id{cOl-f>f1Y>J%7AzXdiQWl3n%S!jlg%D^6MQY(4iP@7xxL$$z<T
zni`}s6j$aZ%?6eiVv{^=SZ3HYaWe%mFvQES$?$^-pbQuoh+mO-{H`dHhanRvt;nLt
z0VZJ5HmlChYu`0_B{x$bP+FQznjcKSq%&fwuP+wweZtKY0F;(xlNAILFzJboZj3Qk
z{v~lU`2(e;*rfQu1WcNtC^fCP3&?}Hk&z3l+knC#hrzVL+yoQ`IS3{k>DMy<VA+GO
zj6Cs*B8rMIkHMS>R1I<;Of}4DKw*%(V8TVIIf)>4aZz&ddk`ZXL_k0)jKRPVj$no$
zn4xecb9|r37f-p;oo>SENy#axX(8cZp%GEhF|l!x35g8RaD@zM2xba`8G~R(A(&wZ
zW&(m4iD1Sdm=Opj*h3I*EJ8FH!AwFhlXFvYfC+)YC*6;kg_Vt+gOi(wSBOtgfZt2l
z%|%0)%S%L3R7^_R%G%b>-oeqyKwQ^G)j~@Vo2Z1WoV=2<ikiBnwvL{mk+G?{rGlA^
UzKN^2gok$qlY6kUC!;R|01RD*kpKVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_int8.cpp.C4070EF85C2974E6.idx b/.cache/clangd/index/gemm_int8.cpp.C4070EF85C2974E6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..627c6607c415779c645edd47007477e2afe5d38a
GIT binary patch
literal 1840
zcmZvcdo+}39LHZX!kTeyVj7eQA@@7u)}YClK_)cAFy`hGRt+n~baAf@4T?nfSR~3K
zMTjJ~Nl20<tPoq&bh2c3=)BE8J?Ff?^Lf7C=lwn3-#PDjNq8JiM-l*>S@^g}b_`t{
z000Hy6CJ||+a}xu$O1rIVWaQJualPYgZ$57STm0oF?GAQ75ZDv>@=yROnF@`<+<eV
z@Cls1PaAW_*IIe_1x>mGtrOjcf_tgftafsHz5yX}PfK!+Ak4?%I7WW!L+k&{ni`4<
zx%J)rJMG%GXR|zK%4d61%V+!E%1%zjr7aG$2^P4FH-Xg4+!g5WQ6~bX`zwzeSY{pL
zVk9*9jx^6tvw6?H4UP_7O5i|y35oo7?;6JohHBA)@m^AuN)0sxPQ9^HWy2Enb#YB#
zxB1`EoYc=*;U6j955_xP)nLa`*4|sx6xh7=jVFf|8Jyo)XL0yQ?p?`iymZNks1EVl
z9f&fPU$^Uu3nnDsU&D?HrwZcv^rz#`P2_I<mf`Bw`A4(etYSEG7@li-ioSho<?Y$w
zrPegh?G&e%x(kQMgcM`!Gb~Gndb}h8OKMpqZK6~sx+w^Pim$%pA;cyUZz_3+&n#m@
zHSRUbA7{3<tJ*`D`(&`KlFIC=-pW1cvVG<U^MW1Ble%I~J~lgd{xlI8?wgTuSrA*N
zSOLF_vqrGGR?`aEu(!E7xyjy%h&yLWRiRgLN3+_hNjyVxXN{sC_VC_J`6QY03lfj8
zELS|F?Gbat^46)+PO|2o?`p2=)P{ec=swqjw(_*=PHOo8xe$1rQt}LZLZ6(p1o)NB
z#JwSKBh}{k9+{LX{hqW;SQCHg3(4!oMU{u<5xRx=K0(!xsgWC_ntWl3{Q3Iz8uep<
z`+`2PFC?L;N81)+QGt8<xm)YBjel|sm6V-&pVYhi=)bnvyijyGM$cI))NkmS?MJeB
z$u8~Jn2YHWEiq$}jb}1rHCaV=EiRZ8Is5BpuPt`X3bfYt^HLRLzQuIDBKq5;g=Spp
zNiMu$Bt|I5rSvDA9X?xgEtQLO?um#wr_@=rG!syuuKiTCq}4ki!489@C*I(clroyw
z^Ct15#_&+nK(GGQsx@5oN*8Y=OJlfp>)u=S)%WO4nVC4rgkpH5j~?5yULO0Vq>Om;
z&27GA`KN_BT=%TB!Mv^2@=NH_;=ou(aG2p*@3LNrO<wdW&f@BaC><H%d(H^Su`O;8
z7x>%>k8%jm|1e%b$d1Ct_Kch!Yk9BU&tNm-2q$jp%A^RaXuCW|3|8X)sxRA|XwjuG
z?Ngw&Y(`YfubtqAuR0=Q?I+d0suRns)f=u&B6EG|PuT99%aioQ_T#COFZMA<=Q6m1
zO3uan`IVzidHLRL-xAh_kZU&6it|^JdiV5aEw-W1&MG;hT(3`VZg%FkzaU<HmA6Wl
z8GTZRpEgPh`FglR)6hXPRt_f3ZgwA;r#~L|nHX@OC%-!Q+W7=_cKD&m$0a>_ITFS*
zT5N8a`beGW#B7<{**aQi;dF7H){$wF5`8q1aY^{MP|C*JiRU{c!XP97ps7lz%54;(
zv$2eo2auEha42vfOlYIU(NY@)wC&3Hm(w2Pl)@lhLR&>bMQ)>jb|^J|=!SB}6b!OY
zXsb!6DQpzb=3#MS)PXey4B{!YQ4%P*jRM*LhaDU}AlyMY@wPUt(~e?rni@h)1?&n&
z7nA;JWTCvC7wiT`70T;5L7C!_e5ECJTp;X?MxfE)yli)D<EgvF3@Dt2LZDD!QaDWo
zp@ITi!D*@pRTLNlj3vBaeNylO@KvF_J|QS`*b($~j^;3<KdeK@dilQCAOOGt5qgWz
zM}+<&3>G0vgw!7hrG3fKF-qNGDcBssV6xc3-Uobr$bSCRfIvzR9q|2e5bzTrMT9{j
z3=|<*gc}2W-v)?OrU)4#WJa<ggr5{Z4Do=9Z<3Id`UM7;QIu7Xmvh^UbJ5u>?S|O8
zMG2*Xv9)(_+>LcIRyMHHvfhRMnTk}~wta_&=1y&0J$*wHQ!@)I8+A)nBlA7VNLNBA
MWbb}wyqG)iKgva(djJ3c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_layernorm_xdl_naive_fp16.cpp.B168DA32D72A32C6.idx b/.cache/clangd/index/gemm_layernorm_xdl_naive_fp16.cpp.B168DA32D72A32C6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6262963b68762adc2a85e851fdcd667ab6ac4d42
GIT binary patch
literal 10906
zcmb7KXINB8v!>c%2y+<nki!f)C`phcK}lj51@j7K1r*mTdJKRh0TZl#i)#X0Bch_>
zs%r$rbuohIVwR}j5>`RQaC_!%PanVM{<%C4#k}*LP+eVJRb5Tgu&}T^M<u19vxX%l
z&R#m5r=+B$jenOcT^#=cgEyK=N=Z8}OnAJY!AtAGop<rUPLcPQo(VEvearaPwPj^D
ztb(Jha$G|bR-R~ByCyTw+2`hh9NUpfLnSR+JmwBCy{9eK7W@9WZ;r$CjF$hjwZHyT
zE$7Cp&nml~3Bw;o`|hwY7<{E=?mvGOotakd?e|t~z->0Q#j*K}X7ED4f=w4=UcPwz
z<);DLS9thNsdmxT8~th7#l1OgFDu#?tQ?v-$o`{t=Q97&l$AT%(5ok9KTSHZarlLP
zLkngs{xN;y*#VJ>e?_9F$zv=Vwi&o@On-I%Kxg)sW6vF~-E=+KG4RQ*(8jI9Hb|$$
zBro<x+p=>b%29kptzqI-Rxqbyabex`?ncJ8!>wljwCVYyuWAe2D^A>>mF`h!dS%((
zlDmgR|48>-Ty2{&@$?}(vCX~_K37a4oCdBuHK{=_YQ~DIT|4(}adR>8bohzyy+rlE
zWgCy`mjREH>t||3#4jz3IU?M>rC3+_g|IERWt0h<eAn;YeyuM)CraZ+WK3S}5>dL%
zZbbZPms4l<)jbwR*;^`m{JeS82Lr>A<<+X&)i>E!Ie5&oxPMQ}G3~TrQl8ot2elKT
z5$aDKRon|b@bqF@Sm|94W|(Ejrpj~7bjuk>2l7vC-{w5>{E5@nGl!V-_|8A~PtbA5
z(}}1sRchHdBrrN^+Nv#+Z^Y%*9=vt>%IEx?C9ejmg`dtVdTOw!Z%ovMdix8}v#N5{
zpFKFG`SWa}E!*z6yH@^Cp3(0`e;tjc>aml<KWp8Z>tw3;Yi0JKbiZjPWy@ZkO0)D_
zoBs35>Z>a>vuCQskBN_;a=FM;KXz+XipPw^i3KY9v4tKF&i^@K@+}S5-73QmN_JkY
z+A${Nc<0G$tF!NH*|-0PS6>(2yb+mjF}U^eFC~W0r^QJ6JE*@LAe9tZMV&K!U+i98
z(8Iqs_UAtz{h+()+qIRs2ZkPPXk5L}{qmgbg7b_~Uw?`8UYoCxzuPQcSNxENqxHs?
z4Za%jF|e)g(Ebk}e+u|D&(XF#=%o8u!_KnY86}$@#l>!!sxo(dev-}DojG~`#CcEO
zG&pU9@PN(`vnSi`b?$h3dd0ni%Hd|-zR52p=4`gBS#a$`M&h&#l}hu7nsxiLhJ9FJ
z$M-dU-ZJsNj_=9IVfp*S15=K^SZK0if&GEAWt*nQcy*sI)6Gg8c;aZ@2A7*or`DR~
zX1*1;P8yil6<-{<I(cr=sAGn^cZ)RqGIs|4BHgjRKH>L&N@MQon9j_;u>L=ntB>6M
zu~>Qej?5=<mu?paCK^0%8g_o}%EdPCv;VN!7w{n~HS=(#U-13@qdw*2e|?p(u>H2q
z@$|F?v(E`uztx_P`zv(e$`QAFHa^<DwXgN-dCd#r;<M$OGB)q^(hD5;BumFkdVkN5
z1FF-fe=)DJSh+ejzU=I<%GK*H1%JI{pCBn3e(=((u5<4eO&*u@HmiTigfFL311}%G
zeB5iM`ZA;TG-cyPi}@GTgHr6IX}j~64~?F^?eV!atCf!M^T)L$N!gK=nm<lXzZlak
z`tw5H!f6Y&3Wi4Sb@5)*SbIe*?x?)tCr*D{<*#zFFiCUWT)&YU^hd{?TsQZJ@CzLc
z&eO}r-gtHXaivQA)}DfbH>r75^KN%}mHet4o)`UlS=@xDw<j;Vv~yFZfz|S*_YOo&
z>=lhORQ?>I|KsSwPGMr_9iwN5iyRMLcpQ5rE&D6q=-`c$OI9Y%IIpav#H?}rD(q_5
zsi8DaNr@dQ8fq&SI1<W#^yk><EQWyV0<;LzQT!+qxgfN8<KRbmYFi5g=m$&>;tx`j
z3qp%ccK&2kJMwn{O2D)&-&RX56to@hSp|Qby8tc0v_0QmOD+g))mS!VUC1*B0s0Zs
zLHr<*ToBr}dsV2-oPu98lvZMTzVdtvxgfODvSl63?}nQQ&{$0S@_lvWg3tqIcCOI8
zdpbgZqA~5rchr#!Lc8>?Jb$q#0uO8yroH*zT5>^XSA$x^hQAk2Vw5%l=Ieo1532wE
zA=mgHzOy^?T>c>e8jUZp<XdXW1))PXc6TL|ex4{m(=ctwH&l}gLWirTY-t)(xmke5
zV_L$Oh~$FMQ|?tPxL6wYRDdR6+LUjqCKrUB7MUdNJaj!nfRZrn#rI<5g3t*A3bq`K
zTCJ|3v<lM;l^0sb1)=91-Ww;WX{!{VnV1&y#rkqV==t8MJ>FOKFue@Zq5M#hTo8I;
zKmBS|er&M-jlr}J-$zR>2)#Kue8$re&gBBM6x05Ee{H!S^p<iLm23B_Cut}xz;uWx
z#8xf{y)`%trqq6&AwVNBZNazDlnX-VMIAWS(|Ts608PiV5#LBnE(o2U8ZxB8`O;_&
zrCFE`7ln)Eg3#Lq|4izLHTy+?=3v^H?`$p?3VO!@7Zv{tCj@9RrX%^0BDo;+?vy6m
znZk-P0a}b{6TXR>ToAgb<m>38d+QkiipR7c-%ndE2wi-pLEQ4sy8{9=1=Ci1D=oPo
zbZOFr71#5VO9dzq({_A2Ex91{u}0lq>-^kq0h){H$^6M8xghj$-uuLA&l?>A6o+X~
zzNfZa5c*=2T7nBZL0y2Pn0Dd2XvhVjE3RGL^!EtcUj=ABrUUtbBDo-R<)V>SnL_&s
zj8d*_Ev!{uYbY0lzcF>f&E)~Qg9T_3repXqW^zI3d#4>sg0}Sx6QG5dw&B|t$OWMv
zT14#)s-Lk+fR<xAm>(>X3qsd74bxa{zv`3#O~tf3-(5p42>m#IYS`n9UrQOK7))1!
zawQo5`-fceY}l>yKDR5dKgQxqtohcOa-pDK*v+iAEqEb7QJ8k+yK2Y<p&JHg4lx?_
zC0BrEVA_H2peq-I?qGb2KdbWd1ZX0r-S}=oxghlWoA=x__b|5wXf~#u_)f-hLFf-{
z&$bR|o+c5X6_^(BMJjSZXr;xo=PX%auB6n@O>k&W^^9CqhJ6Q~@4#0Pp4j)``JRh=
zz_W*odm*S7Lh1D%VbDkLr}z^Deu5B+`@p{sf+_w2ZePHQVBW{K+xI(1o(G2g4e9)b
zYzRTJ3{w9tyD)5MV5pSD6Ror6i5}Y`?(i741JpafOmR2c3F@6(+yzV*SWvzj1l?dx
zFiBeYZL*9-?8UH~!C*5u6NlB`=Uge6bri6+`JkN-W`r=j<+kdNgTb|!s7Ge?$c{YN
z_?L?M*pSk<n1~z{Ic@vD0l{?trRe-|6&k=}vygEX9X3giJoWF=BE$My_)AD2#K_v;
zj-riVv=K~+E*8QRpAD+jFA>8QA>ATmNpxY@D@gYW#g^)u7p5EPY)3qHdBE~u65@I6
zDz8;O1d)WSs{UOfW>|YCdr$Gd>6l@s3a1*9Wy`QRAjkoIq7cI#0jDG2NpUL(T0x)c
z)eH8$;7)NnlC&ddg2`L(<j68Uog8TvNb`W$RD5o6Tzj&A1|Cik5*8tCaz77`TX=<H
z@w3G{RLno!z;m(<w6=jRQH+!i{nWoJ8yGg&GdPrl6^1PUtpczlH<6-3Mg6;CLS_W>
z9PlQWF>Eea=RyF*tsrg%KcXbVz6QJ3+=cQ6ByYG2s|_sM=t8tBskYLs@HWAN$VHC1
z$df$C-gviG@dLMc43EtQQ8t*6H}bGX207Gcc~~QZT(F{i6Ns9?ggl7HHiM{{i(5d{
z!o{s%&<YCo;^7e(yaro({~IuV17eEXz^IL`K_1%<;&vE7`D$cSjRq0yFl^FEk5id8
z*pGLD^-dT_6lK_5V7-fCQn%qXBP-fC{!t3ja|-`p*mN*T2OIjB3}l>v6#A0-4o{@4
zxTBAmA)H}MeLEFwQ^B1+qzwdZU_g{+*luv^hQSo~flVK{QQUzf9Y`@~QZ*{8e^)Is
ztaKYl_i%I{A2B|PL@9<{4@T?h%rR^lGEPI*<R#=UsQ*`i={E29H;;q64q>O6Ja}>#
zsR6u`vKIW5>{x`q_VV!xA*G<NdDqQ({OMgPJhldeHK0XuqYRI)fr%8?f?+KfbFptN
z_)}a5!aDB#br4VoL6m<8!iSV6buf-OvLcwygw$2)LzX+kMn^=Cp(aTN+hlN|^U@50
zX3!(1U~ock?F1i+--F~mI8potN!}m@_OBUR(|tR43?6$Sj4y-<<iR`~%!Efklj0K4
zE1|)T$CiVAIXF<h9K6b55EqA+Lj=WVK=%yj5gge1)-JNOI}p#sM&!5=DKcf?+XsJ_
z9^Abg=elp8{SAz$>d8nu8F6q5(oaF0T$qaVQjsaSKG-9CP2BH4jpxe(f_D!1kXOqB
zLO?DAQjFtMD-A)i1s?nwhL8vGSR9|;P+#Y<I6k$}kVf_^ctOcFD|8wy#1_`1EN56t
zv86ScafY=PTU(O?i(wta4%Vb#VpvD9qcvHj4C^FzvL@wF*ubuQNwM}WY_&l?gF;9a
z4f{R)@A>U>_NX!Je8>5&BvfPMH$Rx7Ma&`{lJ6LHk=Y_E4jQm%AVt?f?K)@@zh&6#
zU~?UuD6RsvDlY#gsQpPX*`eV?CflA-gGfh6rx3r6EV#7!_WE`WK2EVV_Kqa^F|41F
zpE=nQ;0nOV-<(ze!Dhh{nxJA;V>M{NnyorpfkqC>`%%<CcMG<Yr;ld{sq#l&zdPnZ
z|IQn#JX{V6-+?im;`bnY&&53;?BU{G5cYC&zy}b1;O2l&ApAt@EFRkj!aiD;@z~EG
z{0tKEKC(&07Lx5ysfAp%UG<**TUe9$gwwQapGq(CbM|wm7Agj&7<7m;$4Tl|+q9n)
z;F0e|x_fEZk6*lGx%9(<AbBawdj}ds7Y3KYy!Tw(1H2wC?gd^i7k>cW2QK~uyiZ)*
z2fRKm{tUd&6caDQ%SE=oro+kx(`@h}17X-EFm0k0G=tN$X)_nMfN2XCzXjE|)Wi&X
z3o*Co_~%7U9dN&UY6mvVQsbp!q6OK}V<XG9^HdwDpEQ#AD#MP87&n2$U{aOi$&+pI
zIk&We^n)Uo&Cj1<7xFT4ML!0IR<}YJOh<ME%#VO0#U)@~LNRF*;9M!|5pbSg1kw`4
z^BHzR!~{;ROar4dZmvv6#_2RyE+2ezgIC{nKO7CQk+Zph8XLI@6y=Dl9P(=BVr=9V
zF2+WF3s%H^c-YRix9Cn_`S6fxr-@#pv7N(*gpVbSg%vXwEXhiDJ%Fp$L!f^ML}c_l
z_AuxlrZ}nMW#cN|y<0f`A3<tIkdVBHVUHuV<6K;Z)XKQ{BvLy`@tU)hl9*`i2&~8!
zu-pPp<c%q7eoMRZ;^kSHhU3Fyh=!@gCsegh4EMzv{t8;Zf;PFH$L<HM{S>Dj9g#S0
z*S#tM!)`+Qo2ZTPHVWx)=HjhLe=8T~A=5mXd>J+mMdYCgbc9Ec=Mlu&RF)vW5)?`K
zqbT?&8clI2@+n0T6dyz0$51%M$I;;9XavFQ|Kn{|4QtooN^@w$&@p7|$*>_2A!A5;
zDYG-w)aOpFF}9AOmZ69wx2z^r!Q(RLM*MZCS!e)R134;}3$v~+PR3tb*;qM}UQJG9
zzqGfTzj@<e^AuE{f+kUcVV{BOGcJA(s?WLjC8)mS;#Z*hii;aSwSkMvkZu`L1agKw
ziF8j=oU`_?t1&}vIbr=`0%D>`w#ZHK*>Pl1PXhjW>BOZmBp7U5W0aW@`$rSrt`>t{
zF|Bq;2EtK7)|cZf$v;3^%UP1&?{bMvaioy->*#HvV?w7B6Osx6SAnvw9lc0uBOO2^
zP*g+|rwUwWyN*)>Hi4jtTLX50YX`Rmd<T+u+#2vTlDwvFNLqt9Aj|rKQ~|MAY)#@L
z!%D;wYm(U+)<$e&O&U}TYb&<3COc?`wG-P}leQegI*XmHN&hgfHu~(Ghc)Z52VO@8
z*XcYlY$LL0M2ZAS7A3ZvtQkpll70bc79wWZ);ldGRnznV4i?vu>vhEG&r~7rD(+Hh
zMBa^D+>C;n>Czwyg8Kj2*u-kcx|&pt9AuG$T!>eb1}An8S$C6O=A!MQM|y}1>!$6d
zNBUlS+Z{(&81!DnmN<Y64j@H~hG7pPvx78f6wV*ciyV{j3Wudy2(N`<qzdD)55eRi
zm=R1W0<46rhf0-@{tnVAj*n&{<4oF#DeBO7XnbS!Nm*8JTYdxDnPQWXWiq{xVN;N8
z3UwQXO-0tJ$blF^))MlnM~a<3i3QkpvZgIZpIB)enZSKn6APM_T{M+<w#+*~VN%&f
z#=Mh@yTG!GyTR-Ri*9fwZ!bwYuRJ=jC=6#?kx(Qe4MK*ESB=*q%bsBqR1*}4xF}wD
zvbHf*(0-78Fs(<Y38!&dKDi*srE!5_OTejw+wyr0g4f)Z&j%Rr0lbNp4Eq*I-qM!O
zF^A@HKXi?8R+eowhtxm_nM~Pa3v20CQHJem!GZF1;9Lg-xOi9{3@7+l;O*`07hA>H
zYsw+I9LCaQRt~Y{FqPnAmv8a1T)ND#A66kz6?c2pjLe&nqOCzT(YX5{>!MR@&Ed__
zqXs?>M#t%{kYO(%+Y5AaM+$Mi`gd)1hV|C>wj@0QhV{|+u_RRx!}{v`T9V!e-X2)^
zITBZA*ohGnqe)$M{CU8N0z+w>yu$T+2cxLf--F+KF75%p9xm>MuwIx%um1>xKLRJu
zeuB_XFp=_oFr*L0QTzo4eE~(DJ~8Eydr_{Y4z_6xsMUZ5nH7eu1GPGePrM#|v7mF&
ze4K{LkzF}bv<80*+m=)s)ba=?q9@4Z35`=12S&|O4lQcKk2bL}=}!U=PGcjNjw5M|
zL>wFmWJeZss-)^tZL;!Y^>tVij3k2~XBqe=*+B-qDcCC5gYLP02CbiIQ;>KOUgNUU
z4SLfZq<RM_T5QBU@MW@74tm*XWOf=2Br4q8u=vyan%ZgD9G8&JB^r8eTd`TG0nwUx
z)U~pF8%+)R5Ud`89l<0cVgHhym(UkUmq{}?A#f*n?u0O+y{ysZxr>VnA)pYXl*cWj
zBHB@`Ge58Vmy=~R*7Gqkc}!z9F%urA>^z3^`Z|y{E8br(y7YBbpyeug$hG|jj>G^A
z4!O3;h=Ws*a|%)fIa$bcOr>2lS;+OTM?vKN`f~ll-Jjh&vD<tf<jG+Wwu|f}h@+UP
z)QWmUHiTuvNIL8m2y3Apf`b>hzvb2yw~*H@r04;XNx-B3-}5B;n%9VVO+(3(%PKab
z&W~S$gVXong7;tHf;32Qgq9sY(dS7Mq)VvUV<X0LN~2g|tO0R225)u**|h7#uq_~H
z;g(80AnD<jMr}yaMspn5isMBgJK*9dper?|3N%AlGquxm&nd4`PBtOzt${Xy_GDo(
zY>-WmJ?XuZZVxWHWM^aa$<u|?O{n@QV4DISbVlBQ;0+j(`DWNIaPNXZia&wvC-9*7
zZzTDf)+r6=jNF{N?ABxd`2H-VO|@h*@;xr=mzdg^x{~%pTSDnLAKf9>a8gYaBwEl;
z(sJG9b|l&`>?++=3e=Wg+Ox~)&K_-sU2C+~f@;3bVjZU?oeUPq+!l2TSfp@qDp;g)
zaW?R>L7hG=2Y5MLoC~~MF2*H%6R1-=Gy|`hi(7!#!o{t?YvtnCz<bTbZ-DoPi`#(L
z#>MTxYv<x>q*_h4qNM1<D?@hDNhjM~+uew)&91U_(W;N!WG5X_zER;MtC7|pUIVgY
zP<rod;cN+Yl(m*?IR^))fjLd1G=ond1nu-hg2CM=$IlQz%~OputC3<$k83X^>7-ek
z6jN9Q*-<N1L0X5T4;95YZsxZjX)71^BIyS%zS>V((T}bhX-z+AEf+WRlQwek9iH?q
zk6hnZXI8K>`Im{<d&|M79E^!O^!4q(<nm<uJG|hA_8%I?+1K_D2_t*3FC+J~F0Gq<
z9sBYnH24w?nzF*jt%_zL+-O2R&2%gA<??OY9sWr>{~wD0pFhvX|7L;zXL9k(B@gi5
zer#b++o|WiTfH?mKsIE->SS26Zfy#rK{{kWCZxi8P|CpnB|v`Gkq=5rnVheaIiFY4
zOYqH#pES<rY|iJkiYt`X(2uE{E3!DBQ#hZ~IiELhKF>^?H6Q=mgi_?}5&d||d=*tS
z0i&U%DU8q#8Dgug9-(8Pt7mBB?c?V^I3O@+fWDKjm6yH6|MjiN#MI0Z?=Nla92}ip
cT-`i82bm8v?(aTSUlcYh2T<s^;BXlJKmO?V9{>OV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_max_xdl_bf16.cpp.40A7D230DC3CAD48.idx b/.cache/clangd/index/gemm_max_xdl_bf16.cpp.40A7D230DC3CAD48.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc3b58db49e6420c3ef8bf2774576470b500bda9
GIT binary patch
literal 4450
zcmZu#d011&7SDtjb8~MXi3B7e!ElK{*pyvxK}86Nh=?E>_6v%DRTRr&3sewLS+s&h
zKv`52aifY_w+iyqil`O$y0q46U5i>=sQM;(9q#q}Uj8`W{mpOY%$c*zd<hNk^P3}L
zu_jXinQ4hj<2Wo93qsbCrRm9!3|OrBd=@LS_E`9>^!x6_l}}GEuBqw0{VYha{CxJt
z)lbd`E8K)wr1{J(fuUu`Vq!wp?K*XKa!v4?F3TQK-elI6+EG=bl3r=&RuHu|yFK20
zJU7HD)%^omcemX7;H#x=!^?W|F6?>Q_4sLF*W;HPTw<4IaE}`9JDR)Z#}BFF-+euK
z^^3cI{JG%6Ud`#95B{#mbN%pxcI(btefNKh`)Whk3tm?FTd%tIQ|nCIeWNog?p-O1
z=r@bvb6=mdyHMz0_e6R5yM>p`ddKdptkRk_q?B0v_32P^c(uWokI1pwkx#Mx+hVeh
z_zs(2=+{!4d?sXIzSq2{<<Bq5CpC1Bln&h6o;Ce&hRWAEF)UV*5uB0`=~v~Pe<^4F
z(Y9@cUCq;D#(!_qy4Y*->_Tx+NMloL!+NJ$hiaSZ-0+FULGgwMo_A<k&G|gvsm4-s
z&dXQw<P<e~XRSk;XJO~KWTM6Q<zFYSSk8V}Zv3ol+02$DHSa=}+1|G=^04&SmK57M
zW_^2qYw0bQIcAi~^Rd6tN1u5&3~SjtIQzZ{T2$lJ-ZU&s{bI*o_7U4vs}F?NOU7*K
z+us%2+0bF=vEM}<V0<e2$M#i;u?rIHCHuS76>Tv&7t%xC^E_WaT4o#&Al|B3=`C(p
zwzA(gGT@mqd0$g!+`0E1UXLfpqV~l`j%^VHX5H<5c5K_#q?D)M$6dG=^h=c8z?rNw
zYZEgzH6<J<5-3V;etXfgCzhJqm*5t2OtiPvby^TUo=m;rc<Qovb?-B~t847bbFa+j
zlvd1|wtgY*x_|etvcCPi;575FDZ~uf&7fCL2UN0K%3{l;Mq!`v_yeoNek(Tbsao>B
zblT&EQ(u-huSvYQ%3+58ZF_@#TMt`B&Z$&uW30MIw(^H`HOKb4%=M1nntttQzoX*Q
zB}wz;IQQk5Zij|nObo0$QYIg8h#UAl!Ncr%zOmAx>}Nq$;+#-}S?VL!g4qt+w|#Q&
zpRnqux%~b=qaVATD{9K$-coRR%c6Nx&VDV|8aiffBb>rizRsH_=AFKNrB9nsTT-Mp
zyW4T{s<OKFOu+PDuXL}PH}_r}wjZj0^=rA;<%5CSpTBq_yH35|Re$zM=1-oTb>g=n
z4y8suO;z3hl+?W08Qd)>l;ev#n#MgT>)Cw#+Tr7#$J!RE`+rU@UApwQ+5RtP-pSbf
z<p<|=CoZ`g++LZScRJ}}uut%Vw{@>J-+1_GLSl+*<o(k4^Yw2t?>zthTvyfUJ&hSV
zcJ2KT+Ba?1k{jJVHP<6-HhMQT<W35oqWbfyQ;XPoZvWJ4;T;D%2V%nNK1fO)msNQ8
zB)<3CGve4dd7``YQL44d=G%V+#??O@UXqm-f0oT+Ns?6JbTIxuI6H>LqO@GCOfO(Z
z(EpUO+Vy6=(~jcoC`gOABB5Rk($<_io!nW@I6D^7Vy;-I7l^jY$awL`>!~uFJrmOA
zTywErAlm6o){*1=nrfUK3~4D>D%1-^k8QcIspyAqci`+0NaI}GNG}lW&(E)TFymA?
z&YlHn0asw67l@8|RE6^<KAI+A&4skeM5WXVM5naW&$qn%<P^@H18IUwNb~~Hsh;2T
zdv+K@Iv&!NTuT$ZK=dL5NjHxh-GsAeKw8F?3H1We#TRT3XlyD^608M~rZ^N&FA!ax
z=O2GN$i5wC$3a@bRfzNg(Q7CA0d4Qcc$_^A(tIx8NG}jwwXxZ0=-6ZXaCSJPZMZfj
zdV%Q1){h~(>TeP_TMKD9S1!^EL^mD0uYCCGbu-S6fV3gkP^cG(J|2oKaLS3`<Lv2>
zmT)Bky+HJ(1L{_v%KiYH9SvzAS7@pih`weKTJLi+J{xBzKw8OFn&<_hZ$1bREVs$t
zkFz5oZOSzj=mnzh1e`xIrl%9SWj3UZxJE{LF-YIDPUs$4eGg}cLfV9DBG3y&v(gij
zmULotMzi^!Wiml+0pu3oiDDzI>AV^3yZad+C>sx(Smafd@CN+P5}lyLR$@oxzv)g;
z-vRPF;DB_Yak!<f29QYv<$}2kLz=Djsv0fdMccv+D*;{!EYWIIB^>@|)j?41n7ah6
zCa8E#d>E=$g4zVgO~47MK$Qel=*oyTz!h^HG=QLDHL;<zidsO{(r55e^!=2=jqBrZ
zs8S#<1;$7>hbjZ&GA1qu;&LXg0OAT7qq2pz=xUa>#Q}2=)5nR@M9rXWsRCpbut!^>
z(uV%j)itfc33D<-Dgvudl=rN7D&WA_D8wX+0YS;l<O);;1f?)jC{V!>lv1v=L|Ll6
z)Vc8ZZok_wASObSVR!z`ub3S>+INPInMV**mPHncY$B*^i);qUvB+VdT#H-=$^#a8
zfPwRYMLrX+0HPH@LTfGnq5>u^1foJFE&`@SKt|6m2ByVKTmno>XpA}^^q#Km(cbgM
zyp56fQ2&E@u4{huBuC6~IGRLIF`AeN6eZLXp<Z1>q?dSMUV~9Voe@S~*B0qXZkQV*
z6XP^-j7+QpWF4J}i?19%SoU{Y6pV|~N$HHjwQQ<yx7}>_5IC%wqzXf)%>U|8!O45S
z9@Y(qtD?{_au(o!)_RK{@bFu0_P@i~B#*R3M_N7f$A;pL^gK9hs4yf^lUyBWQ1Gn$
zlP8y;Qg|swNGCz1@=_TnjhDtiU+}(Qphdhz478ZHco3~wB;M_D!1uZVLGh)01?u_)
zg-dY-Dk6e1k{T&cXDmJOe9xT8z4g${bwIEVm?8I~1c%YoT_<Rmpgb|p!H^SFJs|6W
zBRU4!5r#^4$<T}JF?&X#YBgF$p;iL2k}lMBq0M{xA0149HAqMaIZ7&m5=aGdRK)}(
zlnUkO>L4hQR3t}j7G)aDDcwy*+t38aCg6fh-IV`eWP)L58=IgcLdh_6OA(Z*&~zBO
zRvJ>Lasp?p{0(rZwLrX<?ntN^K=0`8LwXYr%!A>b&4AoYdk5tJv`lwP(u+o6qXrXs
z=M#s(#a$arb?Fp9hN5)Z%lSL4+wIJYAp|9rNJpUVOi(h3Yy_$hf-;wwk3h{Iwk->}
z8d(b4mW4u%qPkCe-MMj%kr?h$3IwG<gxqq-uQGGD&%^5khsv_bvO&WfD%&cXfpV;J
z7%0~&mx1ztRUTmAd|;K&#4CWw3LvF57XXt2CN2afg-l!o%!|MXdVVo5FJ|HrU|vFF
z6i65u-CKipvMpvSM7cpw)qt!9YNWJ>%qjk6{7fV0j<rC#7AO#bQVb@v?rB16bH<#d
zw6<tXbR_C&1l0)0M&OE6pk@ON(mh{j6=Se5gFTL*8UWcqXD3QPxJdV?p%>X<c8pWh
z0J4TY#jU~z*N)#G(x$Iv-a0@a6Hu{3ZMruQt!*?mno*V;0J(uKM09z<MY@*}y-1Cz
z876E5<VM<rzP8+%yz8T7&;+fQ)*t!g?UOre9sh{697<6Bm_MI(oCeb{PyiOdK*3lr
z15L-KGf)T?!axyN#2^|N*mlC{m#Wv$0@9K6NA&~Wa;flX4F8o#Pgvru`$*D>|NZ8n
z5BO)r;HLp~(oObz{b|imPU8G|Nvy;{hLw;;rRvrMCI%UBhH!`SFq{wwjSMv+-$^4y
ze2v(2nDKBk&oSO($Bm!h<0!H7BHe8){@<vH%v?S~X=$ahwpH6ZjT-IZ=Am$vIyg_3
RnD_-Gv!{ej^aoQ}{{_Hf;pPAU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_max_xdl_fp16.cpp.F50CF88BE688DE26.idx b/.cache/clangd/index/gemm_max_xdl_fp16.cpp.F50CF88BE688DE26.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8f3e6b1f0e1b136fc9c4c589e43fa3995c411a41
GIT binary patch
literal 4450
zcmZu#X;>3i7tRfF$RvX#fg}=0Fbv2p2qKFcDnbGZ0<tNaFf1x=pjN;N2m&q$XrZ!H
z5!^sUvAETWwWyyfYOz&8EtOK-1*+8D?<QX_Gd|CkALp6(zH`q#bMAZ2y-$MF-rjSB
zEY>tiy=qay@_8H<iv=MoZF%baM>;H4GLOYtRed_NGqrmh@$2VjH=3(%-G4qqvHDs@
z<;JJi{1q;ac!WvtF5jS*XSb!_P`1kkZtQ9Qx~VXY(<NY8TW+%0bl0w3Vy<I3P&ISq
zn_mj@W?E+L`o7@io5kgudvpI7*7vq-Vodapjxo_6Qvq?vA^Wycq8}h!wXN;K?@Q+Q
zf9$l2YE<+-PCj|UNKmpTMtWs=x@$(|<i;26Hzz7z3S%s{=7;y)lUz*G7ya1L|I3S;
zyJI~~PY*=isclyLCTq9%$gMnNyW1kn`;qTM-4Byr)c@L%Gk&E|aY!^s%xQNx9aV!j
zPbEVxj6D$eKG`EFYW0g7@@b7%N68Ah_N8C1STW4YGGTVCVukm@_`8*M3GBLmZmdd}
zsk1R)N6YHO!0_kGa}J9vw$*s~MNG4fahvND=yWD*%AYghu(HLiDmyuk>*Z%CGvN&U
zZ9m`EhHd+ueTrK_tIm9)!E>PRY=~KSZ-C+R%`1cV`BlB0)ok5uSKw;qx+gC7fV<?_
zz=8FhW9Ar9!`z<u=zj7{x;wm@UC(J+J7Z~;&9OT1Y^#^`eRje7EH)kutr5FN_cfo7
zZEZRyacy?CQ5&9%{^{7JgxCe~cH-u<z9mQIW?pxi^^W8A>d{I=wOX`Wl|Np%Gc|wE
zI%3Lm<@~0)+$pW^?s+_!Y#jZ4Y{Y~H7vJ>XZ#_S~Cp2;4vj>MKx*2?R`{JZ!3McBR
z&6OKXms_y_uO4->Ns{A{`Zx8JhaPalzo{Ix2mHgGdyL=Te9r9X`wcGSmF^g^Yzj$*
z$amjo>mA(xcFqL*wZX=^35sd6^T!2GKDFjj$4ahQafOxFgmu@8YV&S(C0#kcX-)jr
zj?OI~x~=0EE;?VnZ|mY?Skvy4<`HwmHuqi`BpQU~S3eAPd{S}Hq$%j6adf)0_;X)n
z`(sDrNhe;s&Rkv*)VnzDKWqH%DBs?hnXfx?`bFK=<+;;KCAU1wE)S}8Lh**jXNo=+
z*AC>bot9DJaqaTi!<Ltw{upb0FCy@8OlH{WNinvqX(7DQO?w1g)+<+xZ6A^JuxW<(
zS4*2iR2Py`@0E6!SI<b2yPYhsZ+pD@&DOS82Sp9<-@nTkSF!ra{8X3i&i>2xawe7Y
zyLE2tvk1=mEAf7VIL&XK&(!Mv<bck)zxTde-uSLJEOf>BXpe&>oBd6mKKi=t)RTe6
zXzzjJ9gS05=Pl@}S2R191iG*3{g5>H_H)495i5RMRTtiVW%BmpAD^W*_4Mygj#}WC
z(lUE>aA@Rj&Ns{lEjGMs+xK`!Nl2i&<MqadbC=5R|Kd{CcdhS}G(?&;z4ix#KmF4x
zT5j5^miYKo9K}33)!?#umd>S~ANteM7sXv-vsf4}sI?`wEl-a%m&KwiBo<1ofgM5n
z6DL}TQo*DPe0B_^Be6(htwFTXxcVETTQh&;v!fs_#DoH^8KNyY4_dJhCq6qC(jrVG
z&>BSBu2}K%uUCG?e0DITO)wLY)*#y9pY&5LgQ^NX+aJ<0OeWA8L{DhAzP;dwt@V8N
zEJ*V)zP{EV+J~1@(i3>Dn9mM@v>v7>(HcZAY^X^#Ykzu<&z=Km0wcs)gXk2uwS#Uy
z8$vn`(q@>ML~9VeR7ZT3i$&M**@2KY#*778gXqHR)<;!VWoHT20!ULFimNq<F3$Fe
zyFbJ37@s{4(h5u=)EY#WPVokjw?4)3*)t){!+82ygXr?gy$-q)PB-z{p^&!1tRz~4
z=-LCHW*w~QAo%P^NXs#~P-_rfcf4EK`}eE8e0CV5hhxJ9T7&48Abf#CW*Cpp4uG^6
z6YFUWqHi9xIpA3~sOGbyAuYfJQmsMsZPTC{&yKhZK06-LN=zxy8bo*WsP$G`Wi<2I
z5s;Q*Qa!Cf^aJ&^Q|@i8FfHMb*2nbqwPuL!vW&kvs-la}4uZ4<ljvy;qFJd4iD|7k
zt!Osy3ndfOcznDeN*>ZRd`Tl}<b<Gh0df~`LmMcoYVC;jYJ3g|%F5L$7R^So*WrC3
zcY+d`i$*K|6+J<11LQVfk9?tVxTK~M=p~(TX9MKf{IEr>+1qGqxL_IJmjN?08`TPj
z_eG5m)Hr;c7|kZAI91$i^lb#S9gy3B19E}t3A)e}6CJ=<eC$vF1Qn}_4WeCC1G1X_
z1`kF5<%QV1eh!B!0-_>dh<tOX%|Nu7iHm`#n2Ae(sD#F-cwsD>%B5qm$L&S*=R~QZ
z0_j-F0a*_0(2~=O2KpB?ehAkgD4vWbM^!^md>LPk{1cSEOka+ag?a+q0!>4pS9QQ0
zjL@pS6=#aua-ZpO;C3oRB-%EDk{ihtsJaMBVWd!?LMAAsTxo{XA9=HNaZ983eYl4b
zfyCg!zq*a7uRpdw$UTW5sC3hG5`6|iWte6#P^M`n17(?JF;F%z%?1pd159(6I2Q<W
zftdE32ZVV{ybcK0F>wKq764=V{X!ruWa9Nex}L_UWkS_!`X!yc(fH^QD0>7IqlyVb
zdxqL5Ot_|p(o;NekD-L4W(u{S>7?{aF1QOrm-AHf7`prpkl)d|T=r|r@y)M~M8T9R
z9h6RJCsz7-U9}A#HwzA1OtqMe&bpkhPvo8L`s1W#IMgBv4Wm2&-WT1vu!qa<JYx4J
ze4FHw*65Qq>i)E|@aNQQI6Pc2oJ4(fqpwch^Wx8L7(p%MF4RXp2`Yt~!a$3-ix_AL
zcL@V6<t}BQW!z;$sOZd#LvyCxs)1{60eV}25y}-3BHS9yae{^k$_;lL>UjiJ1IQXM
z8pVVv2IgIJ$k3DQa63i~MyetiHCP77GP(x01noUE_~>{%tQA5=$dN(_N>8RIN7Y79
z0+~RLjt+tn%7k*%L6K~rKr|;A9YY-;>wq%~b$d?FsQBTnN7w`<7KjbdDMe6Hfz$vU
zD?3yCIKF}T4*`cN1)@^A*`T6@+0mSb^dhdfD<eBQ0J($C4%$!{ndX$FCpqGdL#o{W
z)ZTa5`ER8fZK%m0q>aOz*Nd*YTzol<pk!j%NYr=<%2;eX63LyQOvEN5QD2Ac#Z+#C
zB8Bb6RAGa5wJGwBQ|%^w5nQDR=oJAWO3MlFvQ>?qy>|!>m2RGHg@!p)hIs}9WtwL)
zP?mWX17!pAY{0-dz&wYEbAcom$Y{@bK$6GA>wsh(6BhuJ0x*((zYv%dGVyv~vYy6h
zBjL_y?izHGt#NAs(gs0Q0I~wuAg66)X5rdN!TK;Ar9f5+6o^0)gM!vvO=xdUxRZ?b
z7OjeoKn;qZY5`db#v&J}bHIQ!*B9D_JMKQzun1}=Aa~O0L;{49G?yBBk}Yn__=+k(
zR?%P4xvuASOZTuN+FIsr0R##G9ahkr<_<)A8-tHwl;yX8{FW|6bjZL-nwt?l$p*J!
zgir~{N;-u8BU!=RJ7bJt2$3F<J}4(|o<1lY{a37+E<yR=K0G>cDqO`tYFy1g{<uE_
z1>gY;Gz*`_Kw)^;5PJW9&l!h%<*#4_<Y>|d)epSOrNFB({8u71K5e|_AyH%g_e(+#
z@Xv~&M;&T#%+;=q#Tzm?3CT%`tb`$k6~Bl|(aiBpn4!ZNhUs!~KA|VjAFdL5O&ulV
zsYFr(!x2Vq?&Bv+oHW^UwAj{z9A{<v|3)RoCi0O=Gjj_|Ya2TU$1%<>u8OfTd#7n)
QiMM(_d;07tKERLlUr{33bpQYW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_max_xdl_fp32.cpp.1AFB0F6127AC986E.idx b/.cache/clangd/index/gemm_max_xdl_fp32.cpp.1AFB0F6127AC986E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..205d2749f04183ff35c2ddfb292046c061aa717e
GIT binary patch
literal 4340
zcmZu#2~<<p7R?PY<|QwX1QR47KnP);WfD{zA;F13!72j614Th)s51ElLB$b}K^bNE
zf+H0Kt)QZ!*z%pKC>5(!)GAd}R6yIM{+s+>-gB*=wQ^SW-S3=t?wRgg2~c@^zv9to
zGYD03TvXCB28~AJ!*60zLd**-8f^ueMoTU~8+1S6i7S5R>#Lj1ZZWO*^}Rlp4ms=f
zlvPYx87^pYuTvKNJH5Y0uaK~ge|NL)_J+osM8+c?&C-0QV&^@Z+agmf^LLNUJktke
zT&yej{r0_%73prXue`r-ZOq$eMU%rq&p3pIZa7C<-C&>5BNOS&=O^#G-s!bEX7I`1
zHcJ}}pZBjgJzkroyuD^eRO=bZ<ZDL3BmFNb^G0&k6sTs{B(LT?Jv|l+9Y~HDc@f}H
zl^dF37wf-RXVtELQN^@tJw><nSj&?XgIZsVM!HuHoR_w-a)Yrn=CS_05AAL;7FXw6
zMT+kp+~8bs!fwj?7rSM_(Ym{4tnEJY{Aw$`rohR{A=jg}r7?DHu~X(RX)De&)n#6;
zpC8tK+~VlE)~SoO3Fi3K95`BCXkTtyVo~z@5#t0i$1LAd%S?;4$1rLXY_2x_)2HHQ
zp)j-Py2!P><8h`lbD?p{$46HDqwmePuk&fM>^^WU-J-WJzPLk@F9~#XxNE>!l5+Rb
z!1!Z$6Ww5_R_p8Z<@dD9>4zANo93)7vpQKTTww9$&|90pgNhw3K~=(up>LZnhj%ny
z61g=`uu_e?6ng&T&ZzK}kv78S3%+?xOVVyQ`TfOk|L|hnIF(9pK)J<}U!AaJ*z&t+
z1G1RL+N^0EfBok1r&rtQ=nGBB3sx1rJ9p@Q+T3GT@BVsrS=w5&*A4j_`X&)S-CbB*
z9a*%MTba{UaMS&II1xA)=n{68f84@l?i{Vx{FAYIeKFf+tsD`q@-DvOW$nDl`=`Sm
zzHMW~i-X^{Brj&vyJ~k$+Hw$&tNYiw&wE-1cry(r_~#_E_e*ABwHIy_@6i9;Y1VQ|
z@##un=iT1zw*KZ1V!08s6ID&kRtv@t>qhGaZ7J^ybojI6m_cK}Y1t2Fc3k=TN9pZ;
z2g#HlUk{}vl>|KB(iqup_9`#^m05>Yp2^d!vZSn8`J(QrMP0)xtsv}hfBUY_xi#;$
zY?_gp=W(s;$c@SV#aO-S@%8PSKHqBlbLi7z{mLT&IfV0O+Wf{NwI!YIy17wKty#6l
zraMh3c(U8z)KkLa<@ui{$Hvcl-myQaZcUNjgWr34PW=2dZC=5V!SeX+o{G1PA>)lQ
zylqSVn&9>_=-9p`Y)?VoPM@@1!wRqT`oue4?M;WRhi}DfPD<)q<nt!+;ljP^hn)7b
z-&x0P#R9eF4^>rU?(7<rJoh?r<Lkz*&Vwg9C6`M>zYLxkSiX0MZEo1CmjO#6-XG{I
zYwByN%rtAX_wQcsJ@;$&(e1vs8p0MR=AC?1+;Av+&6bh&gRLK}1$Qcp*O&h1!>-V2
zsh%EPM_q2-o5%Ju`y@H8-`by;5*P6^okkNzD+CE(N+*XNMxzlS%n*r2z=@#w$Ry>3
z`rUKSaOg`R&1dp?8Zk<nGahy@7dmn1;gA+E1w4&FwDtGjzxm_CYzc=R2x$YRfj}b=
zZ9kIIc5YZ%!lBQDw3sR8X#}Dt9llYywYA_7hwcYy4wIv!5s3C-XXZWgzm&_NFN8Fg
z$rWh?qE{WRT48eg<s}Y%F{E)OF4PD_$GUGCcE30d(h-n0VVa0E0?})=gjZS2&{_`N
zAJP)0gr^aR&beXPqO>TwfYVk&nqUwtjX-p6hEK$UIW{Lb^ktAXWE%1{0@3-?yg^9!
zmk17hE~ME^wvI+1y12C7e(a>PjU0Lqq%D{hB8@<F&CxG@$Etd94m||YQl^xz5s0om
z^+fjk{fBxEJs8s3Ol_V<Ao^SYw$eT=n9ZTjhqRC><Z1+>e`&EgI<;t6#i55nn#bhn
zX#}EgmkPM@bjJr=+7d`BL<*TkAiBpWplWJwL@I|K327NqCejE*_dZi`(=Ae)IrQ%!
zt;f{kY6PMms;;$7yxsw$wg}QXOdTDK7^NSXM_x56dBmXyKw88UaWw+bw1lYW#2}2U
zXgd3wO2&y=Ag=}PD0Skpj^5Cj4FdotEZi)@QP5GrYq7s6bes^F3hZS6lsiuB1@gVX
z7TH4U@RjNwKrZ3L1Z;vXvTS->QDgEc)Dph15O4~C3DQPg!eD>X9XR2NxeAdsPDCgp
z7NBm$i3%XE0QSfP>Lh4F-A1GX&Y1J412_?`3=beplmU4exd#v4#fp}a^<Nk`u?w(v
z0c~U(Cw2qYZYs_NtXwM21FSp}qqc><sC$<5#TK&_klR_RT<TByQUv5hzy`e%wKj~W
zy01wS_L#jsGT~d&o_js(6&E@zFcip94sb%MFEvC>f)j@NhK8sgaY81QnV>2S`K4p^
zxd!hCFd-tINcZ7C^%ZvL(8(hK6PM#Skz$l0M?T?1s!=Khr5UAB&_<(;6qEssG5`f<
z0;5bS&I0@_AS5kk1AaCYZwCC$RJ;}FZ3PnY`5d5^L&e*G-Zm1W$p=eVJ@&}x*<p6$
zQS@*kOc@r85`d;547hp}k}4jU$7sOOOoWM54@B}I7tDoHWXqJxC`DEd<mF_M#osx1
zYWGmnQW$cXz03(^V%_YSSFIPh`oU$zbj1QRWimJXn0?{V<J0QpAjML&jN$;;-wfTH
zXDrVBCYz`5X*o-7iMF(3?D^`PiwPNUS=&%sj>h8-U#;wc+^_CToLI$LrGsqZL@X<o
zg5p?l6tsr5hJx0z)>2SBD}EHAdV?8I4}UUJ?wI>HG6Va8d_S;5uSTZ=46ph~AeC$|
z8%p1VC_^ZHQwZdRWZ&!wsQ+R3#i_`#IKdXPrN|;qaKs!b>Pwu!#kdp|2~KduTq$Ze
zoX`>LNYNn137(iIMMnZo@Wp&7nvZA~uwv9F6zOdZkk^0-Xj7G$&&(pVJDTV?A>;{l
z(dmN|dOST{bRM8)h2c^kXQVAR%#9MRN+7Q!!-WzF9Z(-~q>=;XFj|dAUfTM`U*4;y
zt_hVq0M*2C#!%c<m!IB@!3nWYY>cKSPDq3jV^roiVIVXxM#CA76(gw?@)V90BSR~c
z<{!O_k{hNz{}pEtDW)kFXzvUn)ijlY(oEARXrt*y3d#Vc8GwQ_foUccX8}<b5R;a(
zfhe1bHv`dTD&7hVwgO}F`5a)7L&e*G!8Q`3w7>+YZv<p4Eip?T$}~>w1M+>q3K>Op
z0qa$LPau_?Fefpo6sipU4viU{s0Q+C;EYV5j)pl@-z7*Bo|xxo;?a<R&Z%z{<U<p&
ziKDX#C#rzFimW?ST2M)S&mfhoF>A_>N`btT+|m8b&w9>18Pf!-B_FWz0gn7c4FRpG
zuO*~4N6eAZn#Dj~Og1ArPoR?evO+3ZVOA6uN`SnCbYZY*V<78SM+tNx#3RH91@7a^
zhxvAYgqy&|_Q8DEWOS97l7dv2ih|~0^C)OOHlKq0Fh2?k#)3!D-+$M&+y7SlLESQT
zav#((2~n}|E(`xBNQg}IRR4%pi~oI=qaX0^gVCQ_gu%8rhq%`+$&9EK%cE&gqYN!F
zj)+z3_(sjqVvJ#qWnmng%hS<T@@Gyr<Fl0lJ>7BR_1!0WPMSQ$YpR{l+C%PYVf6o2
qMG^z4vCPC&VQy(<WAETN!NtwcS#0YxLn!iA#n5LhnC1g!)BX!Cv$%->

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_max_xdl_int8.cpp.A29BA4DC035DE52E.idx b/.cache/clangd/index/gemm_max_xdl_int8.cpp.A29BA4DC035DE52E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..188ac33dd649abeb2ceafe42b6e0647518698dd3
GIT binary patch
literal 4446
zcmZu#d0dQJAD=T#W1g93nng32nQEqH(q1Y{b)$r4(oG75M9Wi=h;-dbl%#sEv``VE
z<%-Bs$R4sUMYoWxTS(V>xnzCuo|(@*J&(`l^~cYA&-XdM^V`q)6y)pU6D*+7#u2{B
z3!;-E7&ICULfVp~#Cd<|(P(qoG+OeOlfm~AAC1Iset&V};O35&r9*j5F{4|~_s3Wk
zNH=EL<y!|?SYPbg=V)~~_~GM-QRT1JOxU8N<<K;45q1&RV#*p764#<R2@PNS#;D@m
zKiQT)D7bL=K$3?Xzn-x><b7Jo`{uNiH}@pNPI&0e&Y6DS)2DldyDa^9<@v+=4kR5-
z-MFK=_hG7+*QOOS?Am)@{<Zkq{$qaXU$aiA)5?C9MxI#`s<=K{^E&RL-<LcGeZ9}!
zC4J4}?fv_2ZCuzU=|~H+wm<6Z9U}DX{=PV1&n<)dm1|xvD%x*(W7CSKFX{djB5AF2
zO~l%}>UOv150~V;;wMIrZ0j7PZdrDv&Q{@WT}^+x(`BaY>Mg>0@Q{-^7i{NCr>A{>
z_vyC@5r^6(%@d2mGllli-0f*S@t4P6{_g>{d92W7XOhe(^66wR#muT3-HO6(3BI9$
zg6U7=BhFeIzw(WmtiLXYw^w+^J^lB&zaHG|U4HLcPL;nT*6oDgsFO%9s-<#&Nv!|<
z{I)9!+tY!b?yrAxzq+*BP-BrRA~NqytvA_I;hhq&I=0N)en#CP`kqT~@^;uWWd;_F
zM=Hw$eXA7j>5b8LdQ)UkwH4gp%;E=kmzA#kLw&kZeYjdPah=r!=leyh5|@Jkm)UVV
z-!&#Cxb#-vt-$YhUa!-?`Y>=$;FT{m3;!N*$T>7>8+REq_Tek(9d+xt?D()~iF<HL
zXLr#~4{Yrt+-Ic7Gx+UI8CVn4er|ZTOLOT<Zx8!=&C909mP`M93|Q@RgEeECdfs8L
zanIwL!~tu%J1&eZyT~u8Z0WeZcTPj2N#>=fp=?1w&y;Gbh~XdaBsx|9+4=kXM%DM^
zE1fe#>n)z`+`MOO{5%KW(%{wX+h4QpWR;lh_-L4y8e^-joBC=~NZ0vCo+lnBxnIlX
zIV}5X8dlSB^m%8!r?4?#%+mEerW(ONXNElp5Byz!#N_P_%m2ujmKpc?Kz+KEO@dYD
zj%Bq|p0!L1D2{HP7j)!6Rrsk74yntVg0uT}8JbJ{%}z(Xcs94jsrB5rvb=>o;kyg-
z0v;bxUbJ^BYjt0gI4E>W`={6H@~5E|Q$Kv@?~ix>w5k1ZoJaJQhm){0E29g8Dgym#
zzMX41HqG`)c5{A~YfszOMs;7ppHFA?2kfhR_S?y%nPw*CuD9C06#sg&H|%8I<Aj4@
zs}?qQoQ@hj*HTfP_}|N)_Pu!feOrUmv${7v^)(yPJrcg$>5aVlD5ULJM`>qT?Z=wB
zBXzMAzAyX?`^y8{Pbqd+tvmSKXxe|)+aKRP-==4C&sWD?t>1nd|BL#PNuTrb>X5>&
z{Cwy6+@!|!^kUy{`cp=qz?P&eh`dav(L^y8!bC9U0*5|_Mk7L*p)#F-6G8XSOtz(1
z(&2xCLl1*ApULOz!~ku{=sL@s=FFjoLt4NT@O1*ww!i%H_mg*iG7fzTq-9K*P$v-W
z_#x$Z(|b)Bhdv3?5~hT&6NvWMb$x5jvGuha`eaD6nQWd;AX?2{S=bwRu7E?I25Fp$
zi**9g^LJIuHE(}@jzgabX%3Sk(g{SzyJo(3Z8n5-B&5xlW@4Q{^dddcH5PMrHHRJu
zX(?06*9k<gymVlE?aSL&xwKG7tHdfZoj`Q%b?ZYKtCBM~Ee_IAj3|~)Ai5w!9r@sA
zyG9N@0@5Z-6M;@3x@fEqnAPzul0)~0G{(euI)UiYO*<S1d7P~0(1Rgu$+Q&f1fr|z
zzD?d!aTn*%XF=MSX)MqQL{}evWY+ig-3|^t1kzk4m#-6uZVJNU9G8W#IrIQX8!`>K
zI)Ui6LpF6@CGUMX^x2T+F?mKhf#}=vpbD?Mk*OSd6r>eQg;*yLeYe+_n`V`IkVBsV
zX)#mG)d@s*`L-N)Yds6YG9A(eOaq=y4A9+{QP+l+b#v%JkQOmTT%ABPEipP~$tjGi
zXgd3cO2&z5psWV2C^yp0XYbA)zxO4;3CmHI;V7!8;Pu!)6go}_R00RHf6E;wHUQ-Y
zV2^a6b$Fz<2ariP;fy&CMw(UoEvn2v&9;UI76VQ(Fh{#lmoV5rbO%n1z($DBZk&kF
zL`+59iWB8PSq>bL3e-tZg|>}I16(kd0RwPij%H2}sbVuwZYF(jH+^$uJ~QGCgF)m0
zVID9<x*5b;AY4nu`9PRY#RWiEKw{Lk&=zgalD61kc0$sfS(;gaq%9>tSpw|Pk*Kv{
zJhgpIs&K#@BuK@?vQq`Et6p#!FgK<GCCUL#7)y*zQ5WEZiNwSdH5^VTOcds*N@um5
zUD&kO=K)NJh%Xx4^>2O6sjY3?9pn~^<3zGNS&3}Ii4=JX1*OVUDQJa!1qG$c(<vwe
z$TI*1X97Vc5RrOU0l_LN&H{ogD$WK**+52qp973?s5loG<&qf9Jy^inp-0Bf9<v{U
zVuus6HM2ud_|Ytc;nog9a*G@0HV|$!5n)=j<B<GlI5wP8V&R%_N{MX&$}MDxExy@w
zIKRIk3<g}`sBlJkSn4<an(g$Fli{+;TQwEU)|JbTWS!~mIjUU_R)q~L|2ul2vD&#y
z+To758zTQuumJmq!JFI5;@oSndkj^Yv&^lL-)jb)uFP#t%z(>WQ?3$?$Tbu7vR)Q^
zcV*&494n58bmGK()_e+zXT?*{Le@eGN?;{W&?45N0krnit9>)abyUDJivYI>NRT6_
zR$>0MHv?M6iBZ_70snAf8&GZo4#*~SWx(WU?+bF1EoMvUxlm0grRR!)vY6~S6gn7J
z?VUoZr~=9=;DS!ty0Uj@l)>2sI!+k!4F{w90w=_L@nAIIE93nb69cm!0tQhCgoWg6
zKy3+~)80(vAtSMo6z8@A<yO);ln-c`_P!!FIbx0jmACu3{lvxRHyCMC>uVl_QoEng
zzu?-4OMmO*ghV7!pgD;XQjt`FN*X6*BAEh>VK_kK#x}@OI6&kkHYh3evpSut@_0gc
zN*>_m0Ralj5ucLey<UBtID<%5C0n6o29cslp`cV%Dg~`jt)QTERXPP_096K{;7lOS
z1X5D(Dj;4(#aTd{Ma9{`I2$O)?{k204i)DD<6IJ>G{U@S?|d@ER+tqZRR&IM1j>!T
z1}SY-F3ZgvGld6(Q3#}kz!VYal>n<z`$QnMIblvxQd^iNYz7)ZI8h0dm0&nh(Ow;M
zubzmm*24)*ikYJN#tF8RZHjIfoG_3Yn4)I`8Yj>P?aPBa)fICcm|{3l0hASFIire&
zo3!r{a+3{aLvef)P;Md}zqh*gcGDyM23_y6iU5wRMy5b*+UE<YZ5TF;(uSo#SxPn@
zx_96v?W2a=WQ|!<Oeh1&GSY-6`dzIXar`#C{5VpM91VS(z@@kxU7t9?lk(&!8#uw2
z^5y7#^QIwv3aise27L&14^^W)e0<(j<nScid=O6fU_NXzS!zs8K^ja$LH?LO1x>;x
zQBVLDKtaJ+@BsSqrRJ35{nB^P0&@p*HR@RSj*ExSV)z${#Hb}d!T<G$A-hZyk9_t%
z#E7078$*jh5u(vz$)9t{pV8#cs0Bp4_M?f>KkG5{nS)pu2j}v61{%S5@1X*=MrbtH
zaEQd!?I(}XV?4baM7HkAkyi5mw<?wyn<&i8RTh@kHg=A~oLokXG950pcOEAa`}oeI
MPnbGZ4g6^T13U5V{{R30

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_mean_meansquare_xdl_bf16.cpp.E1D95735FEB3AF62.idx b/.cache/clangd/index/gemm_mean_meansquare_xdl_bf16.cpp.E1D95735FEB3AF62.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9bcaab94261304a74fe91766c692e908538232fb
GIT binary patch
literal 5084
zcmZu#3s_8PAD?%cPMtYtYA%{;x=os<i!!5op+aiPr3(^rsg!h*luL_UrHhg-x(F#0
zCD|y#S`yh<!dk3W?v=~BwoB~#@VztNzMbQFzImSh&-~8segE(O{IBm67U=JPL&RcD
zSQwbLBzAcehs9zE;9u(Ul*JeXe{xx@v>iW%_oVc=@%z6ll&V@A<y8ZjO@(!B`&yj8
z>q~DBJEzMrid@(*|Gsu$o?F%8uvNF)io3_Qzxv(eLRN7Wzx2sP<+!c=KizcmSu1JD
z>RNPqZ~l)zzm3~GwW4-e@W4R%$;kI^4z~<_-VN)$RQ;$m@8;c@>31$0URT^Y{^8F<
zs#}#GW4&e_Z1Ufyx_-5B#fi3UV@AGqKlZFo?0TQ=eKn4Es3zMkLtKbI?yC-eY@B$}
zWm<rZS%2MuO)=ZnFV_jrteI6D6w>emmwj*9=TA1}T@HuOUpf0};`r*VVIR)jwT>?u
z@yqzfe$8ZAeobp<=X>)P*9^m!)+WCZ=f5vAnbVS{N_dv1Ro*+>q-3|O)#kYVky3y6
zgwGE<>WgH;M*Vj#d)Q}=RGiG-<X@rCGZ{9H|K4G&A~$`9_RZ~rkXaA9tQ^uu#BDpD
z+N3?DT~>T?`>!7kdwNI%6qDLTS9;n{HNWCSEjOBC_t9xtTjJc#d4o$;Nh_q2O}=%Q
zlDwz6$1*5EYJE7OXzJfO5fcWxg5<J*vf2wT+|H#O$qIUADzEMnmPtq5DzVGcS>uuO
zw7@6)mp@$nbL{f6E)cB^(UROd37Pnk<p!dfIimwo<TnGO&Sc!JwYDoR^OFp6J}hwU
zv{U8vk4tE-sSV9MaW>+BkMVrrQkCh=+4<ky?|V6CB5!s5<-_u;U4IzOo$xWsbBor`
z<Et~R&on#kNPkcqpSY-QV(Gv%xA`yMgoJZP*M)j(^=v*>bX{0_qthTYrzG<ra5*n_
zYB~CAPWlvzIw>zAmiGVFvOib8zH`AJt7a-fFNuv5H8w7S`Zo<xW&FhdtUWdRfwQ>b
zl;gnK9Zn^79jkP!(n>cCm<?M08rD+wIbkcmwD_W7yhTCO<;jO~xr-0RC%YAX`B3s`
z`BO);xq3NTCnq<j%xsQ6E$b^j`C6W-HNoFEpm#YDyKPH^@3xJD1w66G<F~JGf3|Eq
zl9B%GpNDOjgK9P3{_>yK(%KvCcXzwqkjyLDsbiRFd3jG=_5qJzdFtceV-CDoxLZ1S
z{@e1&N!`~tJ11X#a73%#(W6q~)&FA4%LRFdA5GnH{d8RRyAR&4GXmrEwO1Zn?_BEn
z!g|BUguPXZx(k1fd{z3x=-&LKo5EB6?xHJK|FON=BlviC#i+!bCytd5Tb+!jReE)}
z4o0k+UT?IbGT_tY+Ru+y50rFZ$<e(o6P=!fxXm@HtT-2WKjXP)+P>6^zKHbscArk4
zcyWK#ecjeuX~&Gm+TMG8>=W3vHlX+Y@7*`1|FYKk*-`HF8&$t~tTcG@&*1)tU)meb
zl^uC*I4dvJ`o!KPP7k*W{<b-=+GpbCyMeL6-*0bwcxm%eoq+AGIwxn9lmu4&2QM1+
za(e1FOQLJESS*7$YheoT?j+c;EEW}wMVn{@>=&AkdEeOX5`6PEf;|(`rkE+N5s0=l
z*->VEW6DW_JrB|XOu*NOA=-v>uLGOmM6lx^Ey6^6jX<<jUrXdVzk7BBI|kC$n6+3V
z5IthWir0_dPBJ0b(;#huSqL=((Texq{Ma_Af~HM{v>9f`*9b(9*?+Zq!->)cf*k^B
z0weS_0?`57jN<;#b43Ju2BdiyPplD$UJ#TP{HCEZhhQf_+5|HpH3HGp{$26IuRlLW
zu;)OU#7F~;Ky;$V`azGg!yp|GX?x6GtPzM#)-veQ#pcx#>`+LXW9EE~Ky>a^yH=HL
z`5BV64AKdl1YM0lbWwUh^xer0M+x>KNLyl-0*yfQ=5hWY^2X<Af;|P&T#T!)5s2Qh
zy<VX+=2SDm4u`Y?Q;0PJ(KQD?hkU>5Hc7DOLRyMR1sZ|q+GD+zPyTvaPp~5(t%vFH
zH3HF1X){-KmZmom>?lYJF(F?g5ZxArCn(m;<Pz*)NQ*HsPa_cB@zdGDSCi#?3HAa=
z%P^T(BM^P*<8bM<?brU|v7#YuC$^Jo1fs9C+8^*O9}Fbe3n9(N_=XyR=vx+HyL@j)
zuO`@wAuY$`VvRub?fyXCTHDn}2=;798(~H~jX-qIQ|x$0d-5C}Yd)lH#kO*dK=i%9
z%RhQv>3|`b1!;XuUtc4J=m$12UGi-Y2zD5x4KYKWMj(1%tG%|4YGfRb6$xnvv4f>X
zAUY*BE>%U)mCELRRiPx+2$YRrBFY*`WykG#eoZd{Nh#bF3sAgKA!~8JDqxbbw6GkG
zQVJrOr42%;xZxIcWg~k?N+c7FwEVZUNvawstAQ6f9zBOvs_UKJ<b`{QkTIz6;Uaa_
z(~IP;@`)%?sQfwHugadJ#^PfQkS3CfSH(|9O^l@W0_9#X1{s4|1XQ8!5A->s@zF!Z
zkW`#1E{s-D3zW6=`*@_~CY7bz*Fblgc$)a3zTjm(aFN2I1~?q56bMVfFmxn`Dg(kY
zCN2lUawgsagj;Bg+7vWV-J@s|U2qp6{gyGRm{8ip9YDDQc%mcyT?R^r*9xkjiKd>W
zekjtY!$Bvj+Z(Of4R<p}nuE5TF1nKWg2#bTH5XZ-=#iAvSZa=%0Za>HOLLSKBqcYK
z4@Xgoyw;J_*5rQ|9&X4t6y5tb)#o=f9BmBqTtt%8I*WBy$W)R_2Nvmofir+b1`}rk
zi)<#&0Tww-ya8BjVB$O=$O8tn{*6Gek%{wxAfJf~fMEeJq1P7z!$Kx50)|C2M&X7&
zRL=m~hf)Wr6LJq42Cz=llYn02in|&iN0U^nDt0CcAsP@+uX;MrTgKtzhGK)p1WXO}
zjGz~d!N)MNcadrlBYXD%<sLeFS5ETl8WG_Z0;A(F-eEfGG#TF>&pPwq;ScJ^;SO`r
z<ID;Vvo7%6IyOn!s%%41(<3Q6m0c)mzsPlf`_)v+?bjuGS{(YIN_#VVZ}fusI_<l2
z&!(iq$9m>^R%m|Y2We%!Ec)Vsk<>EXW%@`bNu}tfFi@&)Dg&+1UBN&rbyqUbD&18>
zXj8JV$-ULDTZ^Q)60SKKh$KZw2y;|#B&9FWH%E=T;Pk75vnSlx1<l+Fcw2!nayCkO
z7+dwEMUP3!2lp99$F?3Q>%nOB8t6nAIQ9IcH#y_ZjGB&CMKfx;5-2O_nl1^eKREdG
zSd2DFkrL7jr5H)^Bs?=z*(Al6@XgTRBPoGIV1}ADNeLxFGc=Hqr7+*rBcHZtKTz%m
zzR2+EjDC5HUPlW|05M-ILi3uWjQB<(G$MB=PT~ZG<~#x%ssacr=vId60s2aPai9<J
z#=RN7ssqY8+E*wepk?ao1ii@}cOOb}^cqmJ`cgqJ5-CMqbkEuelsoC%Kg4;rq|5cf
zYuMV%49uix7Lb&=fw>gb7D-tcSV+;Rfo;!1YLDKdIkMZSW|O`UCPgXWl>z~B({cau
zv?kvt-6V%vCtGKW9&=!GlcfU&&H%CuCe8-3Y$nbDvK%Jf0Aw4OI1h;PfP~h+5r{W3
zaXt{|GjRbhEdWw_eIYO{Wa1)VT0~<MUg%ZzU5xgs5?At3#*ox5pxg!Ak<v4FymAc(
zZIhr=Dgjvu6o^2#HyB9u)s5CBw~{-fX@QyuT%^9h(TnUu?W0i7xMH;?cfI#CeRxC#
zkW>J3bOg$DSXb)%9lhHV_mt2EFH|j@g&HkMH2`G;7>86KL*P5B?|`%lU)*=7@siX&
zpxj571<HK5NqsY<H#y->3`f-fWex49p6vcxZN1tpnwFxw6_CgT)K;Mn)E7xwo4t>H
zF!BLvOz=qcb&_7>fqO7oNEJ|4(M<ylU${wq$)q=p!bdUe+zphwX*=(`Mn1|o+|PzJ
zFSC}}qA-$_wY9Y^x~QQE0c}>_Pif8L@$o~YJ<zglnr`<f6Zpo_UeN(41RtK?+dT5|
zg5f$OH5H%ArQ;Ng2QyFz9>PE~@EHsgj)ybQOnfE-&BNynp^qQyPAl$gd8-~^j<){D
z({O+}+WDj0ONmW{zfR!)!%|{WN2|XqQH%e5lF%0xYw6HeEo$wA9^0*J-Zyg+mnJP?
zB@QvHm?aAn;jb`OQ0!zajy9&FixVV|udk;P_>GqfxGJHcXqb_)hv(=qW4(QRM;eUq
xvU0Pv_<v8uCZ=Xm%i%I>8#{Xkg>sa$tGl_2#L;Pjf!IGVnLTm(xBxJT^<OVQrxO4G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_mean_meansquare_xdl_fp16.cpp.80F76169F90BD8DC.idx b/.cache/clangd/index/gemm_mean_meansquare_xdl_fp16.cpp.80F76169F90BD8DC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6212ae8b7bee5792c7f68ba4e146020c7087f916
GIT binary patch
literal 5084
zcmZu#2UJtZ8qS1p;pW~*0umq)l1PXoqC}cBDJl>US49L=Y&4OJh#<<kx`KrwO$9_O
zh#;~y?A^tJUBGp1EQ(^mE{f|SzFqZA@^-k_b6(EjJNfQ6Gxwi={xTsnC?MdL35zu{
zHYj6G-29mw7K<f-e`)hmXJe3j;Idd5n=XXkOTFjH@B9`k)g0cX*xbFmzG(ZA+N&+S
zYxKZqfen8~%8lCT5A}ocTsO}SUD(#Jx^?)Ou0Jd<uUNf;UlMm!?YF-3bF-`G&#w-y
z2pM>~_F&G9-sE*f&zH{(?Cvh}nx5|36cOQ*-oNp6(c{B;7aAvod4<e+;&Aj>?+3Th
z5wAPF#CwFp^uo5ie11VSwY{+N*|ooxc$^4qKAL&GB#399kj$+*9<$8-cwMb~=9r53
zS5LRq-0xpIFR4b_o^`wGg+gQd%dIGV@Xqbq*OK65`-63N&V5kHu7@ZWKki~5x@4Ll
zbE<B|O7p7^U%dFq`puQ+p>u0eJ__?cZL*lMFGG{~GEcAU(PYa5enE+2QzQuk3+r$H
z)0i_>5+n~7ze>>H%Vs<LsAo|)WV)5B-7;yHbolkW%mV$}>jmQ`cQo19XAX+5xRh28
z#+;F_{(Zx_-h6kr0f7$08IzWKXMWw^Mb4aWK1O!Zab810)WvB%b2Z87(h$ofs*sdz
zNA3*>PL%2&j4TfSY!ErJrzuz=4=k;@{KoZss`Co&ODjcXtZ}Jy=-su-JOfX+>=%VT
zWxu^}3CLFFt+-4aR>p{O?<X$D_oUmJR8RE`bX44)H1lj$dyTDfb?GQk57|4z<)TuP
z)#;bGzk0b};fec^yL~LCnatH#-44$G`Qg*IQ~Y_0>YDvZ^bUFiMBMa_DT?!Tm0dft
zwxd<p;+vvzKH_w<x~12B<K6`scEKyf69yTbSrGf~$)WRdmw6X9ZGGB7v6}a<EZDu<
zSvKQl8+Y&JC+DUdKfJrCCa-Gf`d96DMu#2Hzxd}1e9U>};(7~Pj-}6(!STHX{)x$>
zy(D!7Nu`fGSG1hElHD!tvHd;t!0NBtxA03=|NcBdQaH0YWM3|K_S}RN*P?IlH$0yI
zVz4;s(|?Aa3fZ4JY=87=`P22M-YJ&bj12HKXq->@uMtN2R;+qf$P>Cf`}nTyt8CYy
z?97+19yMV0nnnFp%~#LvyL8QQYpcsG(X`cDev0_jyInITZBlAw!S<$ub&-#5g!sKF
zj;ko>YQ4hQQsB^WO#c$kt)gXe=hMo!Gl~vAZrOD6bo|P{dcEIO1}Tc&7aUzilqlcW
z=ENQUt;H$nQmpFJm0MHaPdb^H`=ENG;JMk7(z@o{FI{cscfMR(m;Nohf;Zu{s?oG3
zEIn+O`Js&l?aj-(o+Wm#Z4{+MKXLYVEFX7gj`{2I^HF!RUb|)NPA!j(%$%<La{9!Z
zhYP0@Wp^`<TC`t#`0nUuP_;Pl$)`VBZ<UT)>h$sm85XuV+;f5L`&T_XJHDOSb-vX3
zwb`V+G~2e?Jx-5a2|f?p?cw9UrhQV}^*tLJ9$i~Amk8WoYH(`O+GRnT|G^`C-xj3(
zI48PVkHs>Lw>3@$-WLgW9E(LoW6_p6fb~N6v3@$dbuGT8f?!8L*$T75b%11<<)%`L
zTVqZU>}gOIU;@4l`s9I}2aVW7M}i#>WfRPVuLC68JUtM#bkqYS!H$8lEoLj!0g?x$
zr@wplakM4Do&aSDCNb6lk{v$%c(S2K15Nt@%3@5+*8!5fc3!W{IZ?8MV2^_`fe}VJ
zKyn~AYjx-N^Th;vB9wU;PpAVV&j`*K`+mp8Y=WH#WlPMG)B%#?hfd_@bTrN9v8F@W
zPH3ml0g|bmRS5%bzCKT|r$Cv+NK+jkImvBVkK4KaP)>ld3R4MnfaDZC(<Vb~S`EP-
z4`pl2ny&*S=U!JH*4UMuC0X;JoXANu)B%!<GXtaBL+p<b>{(EjVKRXZki5n(07TvT
z8cndrK$(kijdXzIwY}t4^`M}mJXSQ6l|rRL2S{GGq0Yg;>(~7RI~>Xmn1fIUNUq-f
zb=;n+Hj-dZg|ZZr3Uq+vnxju-&;R^bN3bKI+z;!=*8!62Ga?pVEXk}V*fXJQj2ZKF
zfaLvmhkUuaF#b7@6$NE`p}kB8NNxzl6CHkv;1cYyP!?iBo(_=Qc;Q@8SBRpPV9$WE
z9Fq%mfaDv8Rl9x5dV&acER^{e-%JNczAFi>@@<P=M6hQ=S%E2pIzVz;XAp0(-J(MT
zdoq;GF>{^{ko+L1`J{VGBRrBxP&UGhjC9Z^cMObaQdD#h>`*A1VP-rXAUQQIK21Z=
znabvVm!TxJ3#fMie-sv?jg4*7M%BLsB;_#7VFo%el*oGA?-H1#WD?l`6f=;?WdjjI
z$qlDyGaK1MQYLbfA+mo{o1`j%x)OMx*Q4jqN^QQ=lRR(_6J!iZd^kv(_4FWxi^3oI
z8l^vn`(4_T)Np*bDO!o75;O^6sECnNEl}42FJufV5pW4@eW2gtiF@`LLsIdY_)vO@
z8lbMB_v4n4n_QZys)p{gbhjLZ>Vk)L_f-m;+0EflCBV1@^hdAcP^G}Ql*!A0aT$}Z
z1IFuUj>;4?QCp*E6P<BqV|te{nwasliJO3W6L3eb^e_C-oml@*q#j9eMO-n;aFQZK
zgcz+)Qbr;pF^a?h=kAgLHG<7>S61#;qfl_5q6p7LTM=nvTya+mWK3|y>Ef2<Z+ILS
z2-YSx=)g%zY9X~o6#zz(h0Gd75=kk<iUH_gqi!@NH`E8T!<U=!%}gHr8~)RG>^QP3
z)O{97QcEREZP4yWDicUD0fT1&Nfwi@1d^3Zo(&|~Or8TIIZU1h1bM)eUVjx3tYY$f
zAjoI(LSR-1Ea~w@z^sVLi-B1&%~6CvA8OkM?L(=()DgJ{^$$2m+dSw&Dj(HY)E-b5
zf%&IxA@m>@+{GN721&(f;v!IJps8?OZCjzI_~Cwi2Z;I$3^{Fcp$B>4UW}}mrJ2RZ
zifur>jn0Y%qenFjigX<ZkI;UEeHbduSxb(sINR~)xb|_l{Z#aLdAXao3H-d2O;UCm
zyYXoAB&F0S$D@ve{0q3>wX@t#L*m{6`=@Y8m00DCHkfa4W^3-b)J*uepLIVQ)LHU_
z^;Wzs{^o{})I7s^MrcivN;OPnur$Lo21_?gXRrl^3m9yn;le&vc)Dxv<dL_k;5*j?
z-g;nxJcq&@9+bA<qQ@lVgZuQS52_BR>wqV+38g7K1#K6mCpqCxjKq)DL^Be9BT#Ro
z6Mt=J-QJ!TM`QF!iWHGz6fY#j6Y<0-mr05*;)_v#BPoGMAV!s)q>M$zV$^rhNx<OI
zHh$V7)p*rR<ao3lI7r+6=|MYzdMEHjyQ|FVRK)aaJOGbe$QPQR=1fxNd~*}jF1IF)
z<^+$=ehfHNIWR7#iy}%B=uqt?f_{lN?#*!McA(x)I~2tLv`l-opeGH(hxLUu+ACaH
zd%>UwnW#-X=u)-?sJGD3zK`?QoF<pc?_k9fn~J5VtB{nnskIbk6G=%-B~sM1U^SFT
zRcIgkqgoxS3yh3m<dgtj2@oJRp$vgGYcDjkIV129eH%R%P?k~e`@EIpP)p@Y?a*8f
zER6C@z~EUxp2g%VfqW&CX9IaQlji_=4wL5rVIC0C>#qXBRZN}_g!xQf2&@W$lpbFM
ztcsYt7+4k49K`^1z4mrS?^TVf`6#SNstTy9z%aDv+4~;3W<7Sv&~qCBxe+)Z0o~`I
z=d_nOdNqZO!U=Vn7Mq`Pmw8Vxg0stks2o_M*(e=g@@emS^lW$BT}0awtBIY2N-Ih2
z0O}pU4=sV-0k@;Q57JBc;=X;QmZW|G>R;%zK#>n8X>W=2BuCtl;e%?RuBLr(Z)NA*
zh9~+5balsYJs^<@sGGypw3kYHH8<RiQ9?EY^=7(gpmqx<X|I~}q@nmwh6!7NdMj<h
zLzk$>SqD4WFrDSLay#UFlCrh6wL`bL?gL9F7`6_zgmy-IL<b_r^}c?vX2`P{0}M#&
z2mA*v?cK5XSOy!1k7KZj_(TQ^$HN&c0*_#@Y5250_V?f0PdnUS_fgwu2HOQN8qHv3
z04i^(aY^v^1^j<lYD}7^_RAbC{O^;DzOYzx`@ZT?4Mbfze}OEAlQcJZ4lAioVa3de
zO@cqOSix~2dK`Vszz`=$9^a^+MlfoGLcrA+o0;@Cw{Ua!^cwE%<2%H3kcW+{o#g*L
q6<S(}rLqBX+kr}zy@PtFlglt`XVGBCk*30cpcJ-$m|q|m&H69Cxu=8x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_mean_meansquare_xdl_fp32.cpp.DAB08B69CF521C93.idx b/.cache/clangd/index/gemm_mean_meansquare_xdl_fp32.cpp.DAB08B69CF521C93.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d61fd4efb2d576d12d596b5f9957fdb8ead206d8
GIT binary patch
literal 5076
zcmZu#3piA1AD?$j$IP5F%$N*@F@y<`LFBel$t~rQl(Jn0kwX`{*sX6xh1jUvqY`2z
zU39-Glx#_t)n;u}ThU!t)wi{M@0@2}=kz?^Jdgh~zw>+F|NB4x>pMdIeSA8^EY{>`
z|CGd-6$?2m7E1*Gl2;@x#vpjZXR%VsFNQ}X|LH93|N5)f7S-)(ZdIB)a%+z5x!n12
zs|J`VG7~OX_PG7PV0yK8S?9ZDwJ-2Zz1rPhhS;UTZ9-LJpn-yW*h9YL@Ph%*sqHJ?
zkZx=AHxDj)lHz&diKXY+Z3Fe48K0g$jq+Wd+#7Pn`A|wqc6R2IgwO9Kdh_M6PvXy(
z*H+z9xp+<4{!_%ho40x&FK&vwTU)iJ>ofPGLuuFf^!^&wc6UWd{+c=4HM$EH*G!-8
zc6nNe^pZ3Grfjw#dr5j&dcf70!dJO(2Pf1$Y;#x>Z)<Zv=S@dJ%jlP5z1BpgNgT|w
z6TMDa&LI=l4mR&xJNCUZ-*@SulxwAjM&x8x7wvQjHrRjE)Ucp>i1Q(B)y>Cqu1;9{
z@$)8c)vevR83iUKL)*r0x$*eT#f9rfasSe~`L>+fK0A1`0oXI5)1!rbLRIvG$AI)Y
zCy;H`VD<C4lJ#au?;2Si_s<CK7(VM1`efc1**Hc#kkc0PJmzc0!*I=(BR6VWIKJ_x
zZ7Zv>P-V{hxWc>5(#F_u2UFKdqbZAyn-*1>hWQOBUz_Jwa5kMR3iS8rZnpPc`nhg{
z&jEXfK!0=ZdB03E_|5BKxM5&i!<bUxiyqf)_H~_nT<f{9t$$jz`OSl|=QZo?r%&YD
zXou$L%82-`M{TMvzg~5EopB+g;mG3;)B1i2>#A7YV>HfHSs%5+>4jteY%E&bNZ9al
z6Ymq8Oz$XnN86`A?>hV`J~Oj2tlR$R8&-nvkmvNNfuZ7A9R>kfO(U)w&24YL93W;p
zH#EG;_85DnaPF1M9^1Ln^>uxTXT2@vF7U|PdEL9bWng>Aq=T;y-%@2<n_b<1?84jS
zm)m^|nkGN)C`-`4yKw#}TMPV2=9dj68wxyq|BfTbMd?$*{M@UX|8V(d{Oysolb1fZ
zck1@w{g~`R!?x@%<y)fOG?fv#gaMG<sh(?@edIj*a@e?~2?x*H*QVYcn%@{ea!d9b
zznptMz1ny8&3$M7+x;M>H2AqRUSQWAbhdupt|iqqCv9AVd}j3jF=OAWxEhn8OO1g(
ztDi@w2es)fyOaI^D-VoqNJ_f4N_i<_d2Hts`QeryKf4a@^jGA$t~#EM=UY9u$c#C8
zz0)z_QndAlU%S446>&N~tFN+5G@$pNg4(NDU;2CWfB$l&Xyw=NQo)R8)~&ijVJpMx
z^^cTk_Fi58vOoUywpQsf)q`<U9EyW~Pt<=_d?E67+B27weM!a9v)9eH{BrKp^Si6(
z<At|VjvMq|xjT6L6R240`{2VL-CYI2sg6&M5n*9F!rfO{ym~sc_ukj@^%n}pJ=2TW
zlx)$nr@`_5Q_-i<``ji>+1eWsbFE=})BP)3m*T$Lbv4gKY)kjw@mb5Q{l$jl9}-oS
z8Z4IXA`6`);Ms<=V^}Owg{cfF0s95@F@E@6_cq?vQk*>t(ngpOj}nMBH7qYM=$dv0
zXU~JQ2onh@F-(u<^tEC$9dPy{NQ*JCkP?VCd)OG6>fL9Bv!ftwfmui>f#@+SR}S{S
zn`(%&XFytx$#p1!XuA(ToNgNOgQiV~v<#C8DS>E@z1MbTp31Mo*};&;F<hGxi1y{D
zZFv-OArEKIgtPz?NGO5m1pz5Ruj<+|aCSVT4KYK45{M@ER>Y0^?b!vKJr~jhM(9!k
z(Frc;LoO{NAsq*4Ys^|g2}Cc`(Cy%2^Qv%m2&9cMV<9CFopsIXsGntFGr?L8>3B{&
zmlBB1Tj#6l4YWCivtuD`ikXTif#|IheL!T_KPsF(4bps!uT2R=7j3V#)ATrd2xo^w
z+77dmPy*4F`~C@TsOTYZ_V<uB!AwMyKy=md2c`pW-__#m*^t)4w1kvE^!}7ttK0I|
z?Z?>*A+3Yy2q}T+Cf1w!?itme1*}L&+emCoDKSik^5X5*%;MwhAV^CviGUJ_ZZ2*r
zkvZtr2v{meTS=@Glt6Ur#g^Qcfr>pidjX`)F>?te5PjpQ^**n{A%C154QU}J)T0EV
zZ_7g~yn0kW;_StcR$vMVB@o^7$X~G5^2Z}Mdk&=aF?|6g5Z&j0^|WheD-1~lq_r_^
zZAuK&_eMu`C`#|)>`+MSVR`~eAo^od)Zu|slQ{y`d`MeLEESYMbW+Ts<Ow)kscil?
z6-toxKv@r_psbOWwf4;O-v0s+q}_PC1t{LAkTv+<6fi-W%1uY1l!D0IbTmS!xZxIc
zWg~kCQfw|BYx-|#6XZ^y+zH&!@#s0UQeE%#CO4j&7#V{KA1+c?J-tZbq?m$2fy$r5
z|EBB-@;lylx=0g2#`(pCp(aL<dw_Be@Ic0(76DbL`vZNBJI{UC7&vwO7KPF(s(`YJ
zejk^VtR)5OtSg~A4P6brQD1N~etnt57QW_i$b6ua4@RORIb;FQDPZD4pi{`iML?&B
z#;8p}6V*M6HgOzpoDTh#QGQV&w29?FSq@y$kv`*I=Z~rq?SLj4xf*$+NTUu1ovdze
zv}R|XvjNf^P<k$}bNzDx2S(LcY=)vokR}Eu#;6&<v@kF=MrlEi3YlUQic;i_)+J5*
zeR|>HdO|&M-@mCozpn0BeW+_JL6E8PR5N5ML9PSxb%24>fIN+fHv;)aCe8rz3?|M5
z@=PY)1Vo#FF0DTsh_abD2Z(Z*cr(!33=HY@xj-+MiSvM79*t4Bp%2wFfcBw@jfn$t
z4;lurPSlftUgX4c(npRa$QZwvStx{PKtR3f=|FFp$eTDE8#E?hYN%%fy~u;-!N}fN
zzgR~0Rs&@<oxQ83dUuSO?HmlFV>8Jn3^m}i|D4#+eDD5A_2Y1x@6qG+#V)cA@KY+A
zAT9kYLr`BLNGm_95Y&E=>j3|osg(7Ii~reZ^AM`EmRWnE7tGN-zdNfXX&rp5Wvpd}
z=0{F|#)cPpUtKVQT+UstjdT)Z5;uu~lDWwYw354$fmU%>G0<x6>S46`+{*)VCU;f9
zGfRM=1Q;OCp$vwBQjbmam>?(cCXA#5RST50z#Z9y<|&MVdYICi9C?n60#*5`7zJ7e
zlx1{*ZVRnFF!cC%6qg|RQoamnBuHF}%TNO#NJ2`;P%aUqKq`=-5++D(skRJF8G;l_
zg)%hI2vQ^!$xyE*NFAw;49#KmPB6{Xlb*JHFHr6UUg#Bfrae+bX|*=OxJ!f*anrwZ
zm>~6q`eHQtb|*~b1cYS#1vq3e&?%-n7%BtkRrP&=KE#vf$?$3oP}b01MTr0{Q{N)!
zP2+jvhjSaf2Gp#+OVEqNO0gT=s&)b8E;{88a^5F)IJFPLjwaKUnV=CskjA>kCa9zc
zQm!jEK@$dcJGqH9dXGbq-42yMYwN)D$p?abAVO|B;ZvBh-)o?o;E<{2sg~$52lh1c
zb%24>fO#4dZv^HWnK%QOXE1RlFwbP-O+c~<NNN4qK$6YGIY5%b#G8T9W?(|E&jm)g
zOq>Ud@@R~z9lBn91*5$>i8pCDUML5k>(v)AdXbW+6r!{u$O@pW0OQfgs6{}{>gyT3
z$vVV(A?gUtcigh{hAfvr_m%;o4A`OF=mH0Kt8Z`gZiSh`5sjNpvo%@io-?%J?qVP<
z2F7SNDhF7m>dPIy+m+`kr45eui;h4|m>}zbvJOl{D$p75oz)jWT7?(SYq%*B<S#(^
z3tc;?Jm4nvwUFNAz;j?YsuCzGX-ECJ@zL$32O}D(4#h121Tq1YIMk-TMbg?_crJ`)
zvjZr1(Cr1yT)0Vnx1=}O^XwTW>;}r+v<Y{eBL7M|{D=)}-rU065;>nBEi5c7(f#ao
zW9kfUx4j{>Q{|@eMUF#*0Y*xFTcr=Qo?sn>;{5Sh-`27H3r1-Y<aFM2J{^J}UJwHX
z^MV;@CT}JKh4aE0Xcli41I^>j8%BTsU31RvPSHE{+_JUwLFU1^WozYwl9H7alK_8}
z!2g3KMJ2ndza**!_0b?NX{{b{I`*O=Ct>N5MAlM7uvkmzuL<<msKn?5`0I-m5EH1u
z8G&hXc{m{uYHRt4yeBC{d_NsM@ko6G7gu+W?>r}Xjny6FX69@u|9?*<hDI_I(^2LY
iqphrM?3DJ7PUDToNo^e_>q>n5m$9dWP4op*S^onZ8>Hs|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_multi_ABD_xdl_fp16.cpp.5859AE2CEFBAACD5.idx b/.cache/clangd/index/gemm_multi_ABD_xdl_fp16.cpp.5859AE2CEFBAACD5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..51b25e7bcabcac889e707d47069698ab81cb3bfb
GIT binary patch
literal 10330
zcmai)2Urxx-^ZD8Igh*Ls7F0%cT%KRf#ZM!q^MB@Q4}k7aTOF5TM(3Dq*#d<OB7>b
zBeX;X8=?j*sImSf8dNkYAOe=CsPEkUGdGj>d7tO?c^vk`cV~8YcJ|kXxR}VuI7cn5
z@aZub+>G?8jFy&`0s6lvec_yJ1b^siX=Us<Ii`KVV?Xvzi1(BRW!mpc{xdP_>&9NU
zL~CVBc!=Peh?cPFO`F?--rgy$+PN>Li+6aU(CpXfp)32Mx7%$o9LjB<duia45p83~
z)xQ7z$=af*%&Ml&DH)2X4Re0ZJp1>`h>?FeSnlubtu~7M@@Q?^;p8PB&CHKJZhG}`
z?1Yt>+bZ6Fy1zyDY4ss1FSCz+-&MbV)0(z?X=Vd&aO3-ptu3=}@@K}(7e+6<I`G}x
zty`Ap7Ns@3x$V50pS0z6Zw~KQW>%=f`k`}59i{05ABz8)oEdj0=kfd%*HV{`cxH6R
z;?HAsT1Khbkq5WTJ*CwteZFIHcO5*cd)#ZB%&&fG&#t(3ft8+98GQ47`~=HkrqNDD
zjxYN?-s-EHwDQC2C*FlcxAI2#*Dy_A#qJf4TVua_z^~ES`@TMs-FRT_ma(-7(mm-m
z8KITTp5u?f)%Rm&J2P1&%_e>wx9WMT*n?U@57!)usA%qrFdGmV|5N<-53x1&(_HgR
zN6g&6_m9MC*LhFol)baLeo@QySi?%mw#+rv>F0MoiqqL5kR88ta9h#sJz>j=&)gWD
zRpOI;<k*;uR}n?FJj-7P)~sB!(N=cd_Co=8iQg}ww30u&!TI=>ykFk=JQ><$CJH(x
zo*QSlGwxY#i%!@R_JY%#tmn!*$&o8gM!s0o?(wB2Q@Q4c%7hfRd#kcqiZ_if{h=}?
z{LUYxW)mO39r1LTAbr#iPX6PU?W!LUyDZ(O#Cvd+a%JAr>qS~a4pb&vJhk|HQ+ZfN
z2|vzq>^as+FEe2ZQ?eo9ml8k5W&GGnDSxFI?6@d#@BL}7m)Y=7ind|5Yx=qM%+^Ub
z=eBZgke#tGWam};zl&n#eksWaUSd9RwsVQ&z4u?4PH?GoyY6~KATF<6J3j2vhO)0m
z=LS@74L>mY<)8y6)_4ASI=VtWyyUf8;C<2O;hu}edQQvKueJU8<L3I@vZxKO>s;3C
zouX|YTUnplxBJbf(eVq1R~6oLab5AL$Z^=FDHCPQ-(-J&5YUyrImauee5h^Xl*|1x
z3j!p=ea;DvIXAU#U%-hQL-W@pUhH50W@_597+&q=_xdpEfk9T)!_RH6zR$Qj$g0qx
z|I=r=?o+?n*|zk0<?i#@%RU_SmzNY+Ej+wbcSXDOP*cXp!leEeD&pYFi>2LJ)Bdxm
z>FZ1T?i)IE8e8@_-&^6R+%WFj%gd)PiJbY@^rFr$+vG_z0*)U|=?TqXcW=~&XQIEk
zKEb}}H`%+b7w+kNn?EpkXJ6(hV~5hkmrJe<|5oOFDB^bFVA~9)>SgcvJ}^cM@0IVK
zYLQalKE1H))mP_F{OEnn_~Xk<w=XtKy_Dp#`;W1E`^!uE^>zK0>9ykF=Fr`uNBNiJ
z?_;%pNVscn*S~Ul#>nNf<lKIt<KdQ5)kpsLa4qa(T8YOv`<dPMxktjz*_qS#TS=E~
z(#bk1KBQpo$s6jE!xf5mFTM`2e6;!d;jOPDZY{X>s$<Y^H9Zeq59&wl=sRyY+U&%@
zfD>Ooel5FLZRl=hI=pnN*kE#8>Wt_cvu=O-VA=hqv9JGwRd4baE#s0`BTFRNZF|pk
z?TK;|7(ZHzOJEZ8)eW*$>VM9Q7k52<J<L)F6ScIsS-e?>>ITd1@0Qh_dK+CTgd>nV
zgEvE8-C#NB;MHAuwVMwJVFHqEcs6?K2FuE_=`$9EEhrJg>2X?|tA(pX-C%i0VNdtm
z^1ehNoQisz@=SHqjYfX(u+-(7zdniJYJj@O3S#xt4c4OsSsR{=I#(=&<571zo}G!h
z!SV=S_mZMx2A6p-9?9|j<Bik}mSdmY4~S}bStf?7(BLC%MtG?kEKj)JF#l9}<b5Gb
zLcJ|{mOAPN%d;n}xKn0Y*d~OrNVetK8mJpAr}h(H=JO_%3E>zd&*aV2S2tK*v}|+B
z;%EnS?OEvB9u^)Fb%W*nt4_Z}J8rFKwZ2BZw*h7w;Q#XjpRzIJM%>PtN@E@zg-&tq
z@2sb8G_o=hOt|?uSqQ&EvNg|IPu*a7Q{0a~zkPOenh+)<*^FnVqi(QVvi(Q*{y`@z
zg>Vv*Q+O#R>ITc(Vh$g@Uo`QpKFmZbD#;<qUEN^0^pN3_gLgiS6u}fEPZdnnS2tMR
zmG#7any8^l2<M=|EqE3>>ITbu?T>sMU2bqx3}>RLakp@ns2eO-FmJfazKvZ%I2-jI
zz#CwoZm?VxZ5Y?u7Jg6&Cm>nMlj^A(ET0G~chxu64%gM1hU7UWbF9=2mQTg$%ynNe
zRv?5Uk!;Vi*Ht%IZk)nT*);fUiV#jjvLnw?SKVOw;_2>38y;`wtEWZ%OD@qN(H+Y-
z<;5RocHT4=!Fgy}k_E|n>ITdAp2X;`bX;B|gj3MqPCO@Fb%W&xF<0sWnj3x<!pTUM
z@nm}HMkBX6OuKAf+A4%&k?g>8&{a2B{`33zRT*o;W{BYoG^uVDZW48a<(HFWdN*F&
zNkVsy<R~VJRX13EJ;-`qnEkW`|CPNNZ<bRvSpK{H!A7^IQ|yFrF1oguC+@Fqu)J`_
z%tgHew6ywp3J>hQoLnGaxh`Pe1)MdjoO=cAU(wtRWZl4t=siHzLvt_i=mow+&w+z;
zU=ZP4I5?N)Jm{MT1Bspwee-Ev2gBFFXsq*^Pq%jrst_<N_Z~>!10O7~+y@~2K=Vgn
z_Yrs#y${&+5!Us{-#&M%XA2t0RqE=5jUmgqN!^_AD`7czsk;+i^(^Nh^>D)GLHC!R
zPCII+o<sw9F<zPpXSqlwlA=T=k)k9fNkh6<mK%=>(<?^<cpG_J;1`L=$=idXQ12*0
z;RC{Dm_!T2!zxCsaz_Ky0dXBLC(n8Uh)>Xb1&FWE+ziCcg!PYGao$^_ukuxss`nn)
zk%2w{y$>{h1fq|?mgs#z)JO9dp!bD7e>K!w4XN?6pePGc{j;Gco3MDF%Q@lnBLJ}6
zT`0aw1hXri%bR{u-h?c75*nO@mgElFptucMW5x~@3e=7iyBUVt3@tW8dwiO;C@Wwc
z|MEk0p+nH<5VXO9ji5L+$IP$<GF&SNYXw8Fz;JCKunj0M+tx0e-LX0-7cgi9(IIG{
z5gshu-rwp&BO3Z1G`<H-vEXR2OTX-vX9|zuo&l3*K!OE^dk##V)BFOMya1YKWVm)<
z(oS;+FzEmeq<<$c=_KEv)BYKX(#+~4^bGaTww^pS!!<yQ1~OSVyyz$RS6IPv8BB&g
z_A~7NnmQbGpiZiY16vZ7Tf!_6kxm<-#YQN@w-~5@cIPjpOFyHf`a9tN4n$aBxf_6g
zgXWunf0O20fPaf{pjG$96yDTs9R?XHQ5Vq112M=@iC)p%4Mg2E_W)53&AmX>OLHw0
z)k12#T4-Mj9m)8$P+ki)@1NmnVPY*Dk9B$ScV@=P?qTS)mqNo*XoUqFh4BUdioRGb
zO_XLz=IaFzz5vF!6U%)BJ|D?yW4SEolSQu1a(U1$4^n2j20GV3${IGq*k(8mj~8@f
z^o^e8f>CJgH9@;3D8+(`J({2=&B0ADgm5eLXr=X5sAz@3v_7;I4#OPMdFy5QPrH_A
zv)p=UxE@+yf#nLIVFAsB(4Y`fHme946wzD^4T=e41FbDkN2xf(V!1$R;6Q8<Sx)XL
z_ogUbo<JzvF?<jv9NnVfRG}?d#n$H5l8yfaXe?)AZX>}jkmaQ2QV9-XSWadxlVHz3
z<ezAmhwgxt8{n(3oPt$Q6vBp3q-2#8#jr63*w+ixp)`KR$P+m)=XK27jRtQ4`YphO
z%xMd-Z6St0rF*pif1+1GohqnDv!seP$JJ1$n&>!YX9WL>)md&4v&fh%)q1E`PmBo5
zU4^k%;S_w4p{1vbo7Z&cqPv~PpQlT@aeR)V`TY45rSelLO5>+dl+I7r&{u;uEiiE{
z*nqyW0saHRalrZ2g=?d3xxQ-DX1E+E$|3H7;c}rUm*zYu$|I|g;qsv<pXPN?w2m<L
zAIR9Laul*mRH+&kS#FkimMu1>ICTL6b;^QY;;3Kd4bshfI}F|V0cdyt4#e|?a~b5Z
zREZ3^goD0=(Sv_N$H@l{{CX%ZV3t11RRZfup!t4Tt_oOJ(OeCzt7$$8tdG)s3|Jqd
z`8cpXPV*UHeTL?GU|mo1SzvvZ=5xT{91!EDVY&0b;5^M2fWZZt8-Zyf@ntO61WcP~
zz6eY&(tHV+ULripD0a`4t|Jj>8Y|#{3gWnN&WGNVD&Zs3xR6<>M;75V7<&yS;k%u)
zaM6;HfBzVYtW^$Va-c32ShQ}KT$=MBlSgwtWb$cV2bpz*Q{pDO{oXUV3*F`eVElmG
z|Gdq~4iOKBE=7UNB>hPeoUI`Wn=}H`{9S#i;DVO~^-dX_G7jH8&RfyvsY+VOBQ9l@
zQl57cwAe&EFLJmr_IKi^()<MnbH;0(Lk?8#A@{_V3Q?Rqo{+*rL1~IO1*L^}giu;Y
zTnMEpJ(QHDj8jrtxI?%jz5<S-(X*&BZ2Xn5T&OgZGUrN9r8n-va>{t6#+)x17IE1*
z(JvOw(VUPuqj2(=(eSWsIdlCAvOc9yRtl*^<vS?*j^-Uuwu9z9P_~EWy->E7<}xTN
zBfRSPIlGDR2GQvHx1jqilCfkT8O|l_y#5<nQ>8Gv6pkfpYCDYHPV;v#>N_}w=sRH4
z4w`pDmz_|K^}JU@Bmq|oEX-KWRNvGDdoY&s_3-s0ljP^&N0GmizdO+a%>%6o$sOdb
z6vfC#5(+X8qRvz}D5x_PF$(I;VDn(=%n*kV>dcUs5bDfO^HAzcrGt_>QyHVA&WtjO
zGRM9UrMr$%ZWIMaMG#8zOA5lD4a+6PB#k6Aft_GTk>`ZLgeKWea=?Tq7+LK9mUHlz
zweE3s@L+-7W|SAnC2~75JMv(81R;e)L1~I$1*L^ZLMSaHIE2!a5+$W6gOwU>{m_1E
zJBz>gqjY>X;O_>M2mc=Mzo&T*;P0Wi4Dicn-Us;mX#N54f1r6k;P0n-4|LmuGH_hv
z;`Tzfy)>6Ww=%*7SpnM*rM;c2wsUb0fM&7d4;pzjRs9RkD9g<d&lrGH2pngL1pg|4
zv0SPsl?shh%~S2L4P;U7Bz#5(%FluDInD2Z?|Yz$PgP|B-w!l@0lr^oUJZR$lXzC;
z2&6gCg-k#$l;+Z01NCa48POY{UIWcnVC)qVc;jLoa!aa89+?MQb6W}aSuAH~ZYROf
z*VcDaCT8nNZlO7?fX)?gFuoR^12m_qvL6|%1d2;YK!KMvgzCy4j#5xU3$0p7+=eTG
z{a7x+CBX|P&d3uvO>)O$2-MX<JOr)|BBQM;4U!wM*S9yq^*NSv)OR$(<v5me(swe#
z^|L+mhA~4&<vv1bNEdMF0&e)xRN0%$E1J8(pl+af4Gh-<2KCU~3;cRXq|0!1P*g{<
zH(atocdDvV;?>A<!O~#Ly@Ys%Q0^rpK1AbQ_I8P#+FnRM^jGCEcI%-g*;eH-b_FyS
zLYqP|`z%)kZHj0vhHk}>szaTG9w*5c#By!Wrws<+nb?<~So&n~DGADyzX0hMx}dch
zN>|e<UKX^=q6f}~cG-jvMx}4s;+E2a{7|4M(3qI3AW;xS!J=S_LPenzg&T&O;eH1{
z>m0rllKmXjJN(7|ruf%~GwQTsxxJBSv5G}v6I^aQT(@GDZcL1wy4K<Q4EW*4Q`I_r
zpVRyT_`D$XLRGE9r=8{w5ZVEhq<<#}>jdGLe|CDBkkmcOrJpJvx2T0wQKc4o){^N^
z=|gMbaNH564m`oXYSb(@Q#{iKM_Md*9mZZK@gk~Lz}Oo!--Uzk!Z19js%$m-E}V?H
z;?LlvrDh{%tK&;W4vfWuD!x?Y(wqn7d2j^L^O57Hc^!;d2Pa^?a>8A&y#;!PD9q^s
zI$eMYTwej5S2TA6oo<?YfKCt1y+Egz=31yzO9EM(?xJZ@73|3;pDs$LOmsUCw$nzq
z5BT)a)r)NClTBAI@=+2>su#y1w`P=w{&m-wLD`M05>Q2#Dj=((xf;l-X+8>MM`=C=
zWXEVe4rIq^t^*czKvSDxxD&wQ1kGoF><oQ=J&@JYd=|*g(tHk>oCBI72ZOR7lk+rR
z045h`ZUmA>ps6;Z>I;xG(R>j|F4BAnNG{QQ1z236@3$FPG!s5Hv-NU>;NnAMJiK(h
zOv&4d)QzMlQ8$sIB;6zp;V>Gx7}W+08DJT+jK%)vluz7r?IC;HQ4V7<(83ja1y!mO
z^#F{-YZN;KRAN!>){wF0iROv%Sg1%~I?syGGce>C(0rsUvKB+06Fwam^+vlwd;wK9
zeI>qD*a$J4pTti?*!3WntJ(%4*PSn#PhGbiM6?4<-3spy0ifO&!XB?#x*|V2I9^Zv
zo%wVThr=R=<@1W>Zs65T>>7(4mRAqWy&$+3P?n+&n$(d(1KxmAd1HA(DCLd)Cdr96
zR_#pTd%JB_r~Myio6E@bK7|%fNlowebETDNu2UWozJ=D`!a?|r+%q}-d3lJ<a<x-)
zTu(<e1<<j8=0a#+NII&Vntc(?#n7i1QiZ0I(C;M4fmBW{u#H4j*#9GUtlB&yuVewU
zKuBJG2Z-$;AtcLng1AngsbRL{Jh^k~v33=DtUG{z2eA0Da1jpKDAnE@8GSLcSW`J-
zxeg%gplwtq5O&hM4%)3FHi_kGU{npU(6|zVR+VZ8k6eTkan!1M0U}<IRfYGF&~vNy
z^+*^JBn_g%5QV3L3PTj}3QZXDXP=S%@a}wHv@CpWe4X)vXE{F`KZ^X#{bg8deY`q!
zqd`Rrnx;ANb0*?1x}ATx7zS3TDrc4hEj@76seNvFLV!`&dGsx}0@GG{qpuB^wh?aM
zT>f^a^VQub7kUPmXFz~wU-gYL&uM-Em=`p+1E!tk4#0HK+zFUYn(HCI9#!QuDM$n4
zHxS0T9|~_(o1u74RfTLx6C8lMbysD^^Y40EqUXH=!>&M$S;F%Kpn9Ve_kTTRz#5qT
z+fs(%0__5&_$-DCvJ0Xp*e;l&P`gkK;m`-w`BmGjWP#)WCI@JfRgSida%s*3Odibz
zfGME45HN)_7XhY-=B<F)O8C>6X8Y|y89V;p7pk>z#ys?w9q6y(7EW7KjehpyWVf`R
zxJOD?ik7cfmN9=pYFhf@B}*4B(prKaMoVicIh;=pQ_0~Xa+ppISCGR6<ZvN5Tt*Jl
z$l+pgxSSkjki%)*^m(%o4xKT)AEV9d&({&My83#eXoH9_djmnVp{bFvnYn*}JV+57
z66z*)4wU*iTK`|aHL;Wou(7q14RmlC<l^q(>Fqn%%E!XhD_m?68M6`&8K;Z_!?gYf
Dg3nbg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_multiply_add.hpp.6B6D21EE52049C9D.idx b/.cache/clangd/index/gemm_multiply_add.hpp.6B6D21EE52049C9D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..16c7745228f2273c99ae57311e6a9fa74e2003b6
GIT binary patch
literal 3304
zcmYjT3s6&68onnQAb4p2P2fNX!AN;#A;}2{DDOu=K?KD|L8O3-i$Ed4Bx(~x9Rf{&
zN)%8>K~@*2QXN@9yKARv5sQLrS<!XWMaxvtb+Op3YwOHjU~lfZGr4oV{IBmn=l}n6
zGm;Y%|49-=TtTw7q)?r=kRS*r>ZejyY)@wqgv@~;w8w5`m6UxS>^QRKhHbRy)WGwt
zOTOqH%Lwt>8{PR>5_Az|9b0}cEeXATT>j{8^Zvbe3)cOV(EQ_wX3f7xO3yTpzREt+
z+whvdLh{3zu?zd(4^{nf^GMdYLSc5Cz;I(=n_tb}Uw$=E`I1{5(f2o{*NdC3-)*>l
zxgtoz%}|y|&X=w0s`O3h<1}oUdi&oizx+6J@%@wWfa2htnPQj48AArq_==GWYD3=m
zs^=rBCpJrWR(vZQE}jl^H*VyzkLsKT-JHFbIr%BOTb058S(WIjk$$x2^u?jx{2=#J
z#<k8ZX9penDgsJg-Sc$casK3yo|!SnoL6J-A4aGrLv}e$-2Up8%kpcDMHR!Ayicc(
z+TQ;pUHL`#o$)tYx-_4rSWeg4Zs;xAU7EA6R8+Y3vCsO}8-rub*D946uKIV!wDAWv
zIesGA3;&W|E2e0a@?S_xoqu-Hq_E`8r`{y@^ZJgSt?9pewd3L2kAEAj=l1u{xE9NM
zJ1i+~`)tD88V2hhUdg`R^Fj2m?D_QN7EM)I__=<Mdy}4-e|<51eCMfMhQl?td!<_i
z!Lhzeq;2VIPx$vGbPceRi(+*>EwRUqneETat$(sBU<uA&d)mq!`t0T|kJH<^d+WP8
zQjfI%p)LJ-?}FDSMCBz<CrefoT)vcA^mR@|bEZtSQ`~j)7l>|4n|>W`^V2?vt)BWJ
zGOXjI^6Z0GRR@%cgI@p8K2k3FhWG6Ah?opJv&>s_|M;52+r08;Zf-bmV^8dU*C$VM
zr(T2ZQTO72ma_LY(&K+#snV5vJVDhLnSJ#nceE&%BpjV7NIh^Rd*w$2$A(3CH~|P8
z^5x!9Ph0Cr;saV?DSK%)!C`}Wg@$gOPRTyNBMC085X=syJlHLqS2&z|{@g~+`e+;0
zB4-ZFvd!|Idl7c?@s4!XUE<3cyPa9&Vgi;qmI>xww76#?XS}X6LQN75w0j<#N5!{Y
zFz<dspY=V`MYS(t*;DrU_W90pFWR1{DBPyva0tT2bD%!1JyawhAs_KWt$v{ka={Qp
z5JO@l4z=z8{1APJAt{oET9bz|B*PFw2t$!5GSs@!Fb2h7NE4_DqPKvAnj}pMgR~ND
zFou$~s~MyV)M0jdsU9&hgUq17P?}*agQ|eE3Shb>5NHCJClg3A0gQtgNX!6ZVFt-&
zfbpmYf%Uk36OcCH{w+Xe0T>$#NV5Qp&mkZ^#OT|<YkRbBl_JxQghC{YVh+g@d15F4
z1w=8&5{LpZ6o$g0m}3b?;TTFpiBZf!B%ve*{l5=boVofpU-`A|%tVBUgw4RQ8Em8X
zj}A(y&2|B!Rt+4h!S87`r!Q;WDto4`+LO=?xncF>j@&WifjltegM6^+@<qNF@<;v{
zQXmCZUCYsO48@{Ythy3Vf)&N`AL@hr;bs;I^L+E<^W8^H09cY*vXSnhl-92sM)tjQ
zSV+Q6cAKb6XZ|5!j$ICh3MGXy+JuDLCEKkiHS}N^Ut(&c9OU}uVo|pGZpBb>YOysY
z33sIKu%dNo&OXmRU%_Tmeahi;nGz6l#9Vs+Nhsw=xy&&~aw56R!InFeFQ$(o|ED{t
zm5X!bRMtUc5M&an0b~uZ;)37aPFB{=v{1A~NQ9M|1WB;sia-%qIVq9ST0jQnyDbNU
zgVx(nv_*V5Mq4D7W3)w5IYwI)DQ9SB#b<3`_~rBRed#UU-(lF5+WGBuAqiDHm5|Pj
zLJCzRgVa1VCR2x~qZnBwuM(4Kpk^r}(;&@C25F%dlW7%NEP@W|FqvMc_hIbxVft7G
z8H5H*W(YULF|sNktOA(S2zW+-u`>dN5nxek0k0O9*8)W?F0TW;I$Ul7&;&4^CV)%;
z%g_v<8JC-Z!i@W?2e2NOHvwT2&bJAKH32MF3lLgxxdnt<0G96|AUwp#Yo-$fVWTED
zTk81yRoG}%rKr-GO;QJ_r5IYR&R|fbe<fyT1pY>RMArfTItJ6#Mvbi5dzl$ojli*y
znQZhe4eaM`XIhPpN&P*0HPc!rs}to3UNLv5t_nq=F?0+P#-sQcx}r}%7!0#Nu_vDD
zDOcng!?c7q@{VCzLXPB^gJ=|ObwFEC)MxK}niC0kJMOm5+H(gv{Wk7jwo+6=P)G<v
zB@V@f&{TSTIO`j4wLb?WG=dZ(NTc%_x<6A8!?IsMSqqQ=<LHO{tQ_e!WZBPsTF{$7
z!t!P1fwZT|VD=kZuFGe$bBo9Kq`8w&%oBSu*-Cj*PljV8FVd6dNV$XMV%ptQp6HSU
z{Yy@5-ZHcd8~#x!$~w|Xs9&vLLz|GW39N1cR{XB7`J3NG3}2v<p8e{gM+@~uv?Fg8
zs4DVR#j^|E>}&pUrxyglw=P+5^UwX0vQEG6?A!$IOZYq&S7M1(O1LmfS7s?-mTt`Q
Ee;SZ@+W-In

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_quantization_common.hpp.A23B3AA4CB2F3E7A.idx b/.cache/clangd/index/gemm_quantization_common.hpp.A23B3AA4CB2F3E7A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..31cd900bb6f3cee6368646c73995735574e5c0dd
GIT binary patch
literal 1606
zcmYjR3rt&O6u!Ns7wBz=<(BrQq_5RZ#$z%D^4@q%CfgcntLv0kU<0S2!?eI44pLl`
z8OH|Va}OAhxJ_U}G!Af6T;ott350~lP+?odOvD6PfXNm;_v-O_ll%4mPrv{B&pF>|
zZP{5_+bM#`JeXbWEU75+BM3qPue-waqD4dy%VL74KKGf;QtpeEeV=oycex|q<V`Zh
zT9^RQYS^O%p>c)Z`gfmQnVlTn*p_j%!P1$#npjs?)i8AAe2{n2`;=*}WoXR3cBXB~
z9FyLdF<N7*=zOp0_+ZGT3d(YMIO}6)v)xmA^TvhMUrxMhW-`tMx7gl^^jdy)hDW?L
zJ?e3ltWF=ga;o%X!a#&d(lg`mW4bR+)W7ib^Wzz0RdeRnh}N`)PsCeJH|vjw4Qu80
zfV(lS9%Ia#sb4x58b6G_wp-$NAn!SOzjSY};mFi<YW1+pZfO#^yISOBU7m?Kd03>}
zkxk53uK(Z|`O$M|Dqu^#FaPoWvbV=yHa^sMu57)bIeWWsLM+kjO{u$EKqYAJTT0A*
zt@r2it}8U*?O)YiU)(o-*w4|?XJu0UYCqcgb#C#dmH68JhleF2MK}Lgs7snH%&eUx
z+%?YP%g|Ly-+wBzJ>Ec*as|OM6cfk;5eS_p2oWoi{@M1^u|08inv?=TLunK|2!!BQ
z_l><mc6~sTG9c(EostKE5aP2%KGt;c98LNG!AKcd9t48kH@Ma~G80adG!R0mP$>@r
zA-Zg>qwoCzAc%orpbQKT0%6Dc(&DB0njV@A0)m24`13#nl29U7#f6!nMCtjXJ85zQ
z5Y&{K;XzQiJ)4%CUO6b169GUlsZ0hQctY?apzv#v;nS$t#cebx0fLfJ2J#>%+^cWi
z<f!=OcbcStz)~#3!~Y1qOCROdk8P$&e;}wRm4XL>@Zj~a#a-pGl{6U$1TCcv;z1w~
zu98x>j}%-JOL@0W`qtaJZzUUKtlFS9VXd$;tHW^TPY-&_uf%O5Wvn5|V8fQmSmO?(
z11&<cL6>t{(A8SCxtJaZ*bN0JfcA@|yzO`d6(P(yj$yPKF5rB|u#DANb<dy_<h0+I
zhYENRC~?l?k>m#h#i`d?HbXtyWNnU+oGw|HgGx9TaF!biJiGh!VB^TK-{f#stx@Z-
zMKYGvuzK{-65m{|ifU*Q$yiN-CKFpzwP)vd5&NI8LfDjUvg6!TSvN|yPu)p`aA-nk
z3f6hT|BLq5d@c`!RT`E4-oGP&H$q<AEMi{W)mJ4}XGa#W6&n;P%>MIO+d;{b9Saa<
z%uE#iWasX9(%w+?&Oq4otf>$)!B5cU#u&A6exdW+@M0KhhDpboVP|H;aJLZ9-AE?&
zave~bmC>kxU&965l%o>X24QZtF^tAQn44$}yGqK6xPKV8>wtR~w@`BM&l7?Rf?$ND
ZN?5A@Tav7N*_%`y^U}P82~nw)#NUWF1DXH;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_reduce_fp16.cpp.EFAED30A7BA2F6AF.idx b/.cache/clangd/index/gemm_reduce_fp16.cpp.EFAED30A7BA2F6AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fa81f84619ce371316cd8147b61c42670505a3cb
GIT binary patch
literal 1554
zcmWIYbaUfkV_<Mj@vO*AElFfyU|<jg;^LB`%m8Kv1~m=_hKe~$LvH8Ywh*{_{!Or9
z+MJZVZ<*>mf~vD~CQIBEs}#{FQW1RSdb(To{q^?&lU^S=={%dM?dZ(ImUAqPQy2pn
zcI>*==lvjV%jacve+rIq|ETy}zH`U!UH|rIte<}QX6NI-#??FRZf=N)pEKVgZMu5j
zbCK1%*WWDsURe0|t$tL^@ASZR`|s$VUi~%8PSriKE^6)Rug_PnKmA*O`_rqsm2u2=
zckl2V=-PTtA)9%+%?jPuclc^g>R$i7fJyjViQWgV<eiTBf&Xh*R!_P#NmILI(vP+u
z+ugH5E8l(Jf8au8xOVS9kHY%E1I*V|jxgt%XDF||;1$~ZO!U<R-}iIR-`D+Bn=>_c
z(?<zg-!sl{_dK-ph!j;^ur4!KHY7z;y7AjW-cvUZ>fYP?`-5CjTJ42ScB?6O+*I98
zL|T3F+@LJ1+A_Cm5o@QZf7OzSH!4MJgmeNk-c>HyVbG*Ek;~z!wD|tZ50#%vOm*zy
zxTr2Tm!G|3)#9uktF4MI#&R?D18OFwBnub3esgb!NWueKrKm`j!?vbRUi{(T+pPP&
z%UhL0G^ObNk~@8dii@2T9xe=9p_;N*^m-SQe}R_at{v|SzC^F;Q8~$ZP<Z?6O*{*v
zUVB_Pc)d`w|Fyt^XC~T<4<|e9S5;Hh)O5}K!7MO&d&qJY=eG;56uB?I%*ea;^OgD;
zu}&ou?zi9be!lO*+pWvgc5lo}PdvW&>{Z@O`3(oMTePm4R(Uj<S1<%EaQSRetj&Iu
zKg;g6;Kh<zYH0=UCgdkbe=a?#)LWW*_5|0Dd?)RUmN~B?j|pXD)b@Wrd0DCIYQ?>+
zGnFE>mi?Fety+CsSv=RSQ{$6vjf9a_1Ka+%^<rP<&o;5UC^A1^GUQe#uVee3g^!Y)
zi#;<lQ$yy=7k$9>BFX!Yi<@onNnx$2^EZps)H$Y2@5yV=z44`}$ZVnb94qtlvcDHM
z9DkJCpSVOZnO7?Kl7VWYy>GGChmzQ9cT3C!W5W%y>_7WS{HQH5v`n75X-(){fum=2
z??0Hd%=Amh)9!+x2chRWwLdlgP*2|fYl@ug<42J~T~=n9E8EQHeM@GNJyAV3>KcdR
z_H!>A1eZNuGN=2mh>Vfg{<-m&`vsTD*oK_&b*vI=(~kbbzuo%&{^H8qBza)|G7!Ha
z@%UX)Bo9Lf14AG?3p)px0Lnl>QEFQ8X+{PHrNx(QXH1`w$-)yT!Y`r@*A^(rFNwsH
z<CjC?Df27Cc`zM}TqqV;$$dPP&6fC)i6u}&NJ9jsiX~82NEgm)|KKWPdA>)A2WYvD
zh!IQ?kY^!chs1Lc@j~z<btI8gN?J%Fsg!h)L{cfIBZs6?&O#1JrJRc#l1gPAWh9l#
z7RpE}m0gtKDi`K>vv>uzJp%d4Si=<ND4?&5b>X~?F*l4B_3#TZ@dRq{Y6!v<@dWDf
z>cV;ZOoCEaT^GG(<O$@|<TQjSI;ULS(PS6Q57KL`VGk1q>b2H|^ImhT>1VRqFafBS
zOOwk4rih^^H75}iyTwJx#mYdMF)-+TRmpC};54S70EQq)#DmBH1T#4|B_|We^-1?*
zW?^Mx=iub#;T7T&6yWy~c5~4X=JFDe6cv+_wz9Uhvv+WGG7#6bQMJ%g#3m{sD<`j{
ltfHo_sjZ`DXk=_^ZmD1<qi^CWF5%%F%;fIx?8)fM0058_X9EBL

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_reduce_xdl_common.hpp.BF32975909DCA917.idx b/.cache/clangd/index/gemm_reduce_xdl_common.hpp.BF32975909DCA917.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f5bb3522fdab70008c626df6327579e432f7f2ae
GIT binary patch
literal 6844
zcmd6rd0b7~9>Dindv}~+cbYWZ_GvhEI?d&Dl5{9HoRlGwtItD}<f<qYGK-hX6KOD|
zM23VEu30G=Qe8ulXu`!q(apQKh~7TAtF`}s-sj_dPOJ4jzqQx;4ePgt9ptZ8|7uDQ
z!z28Y=FUnCqX>e~WB(Eo<L9g;3BpdFAd+%kO=_Cg<f(hN;Q1cWzExAyWmTJw@pT7{
z(>voo#kLb#Z*;cGI+&3)^K{k4ZTheNl{L$@@$&gky5$W|CGnA?b}8e(tax9;tbJc8
z^|ub;)s^;7D;T%FdJ@lmT=`JdHT||jr>|}O^m_7nE49(>!ga?+COU>MiBSx-zHWRW
z{H#-0+o-Ve<<`kSC)IMMP_XY!^iOWKYRS&7O^z??Kb%pUM)$Zx<s9ptliR+%d2TK6
zo4D?ct_NA0Yq$QyfL}c>D;GYI8137>WQTBlecaIULBnfTtqJ!Yp3(iof+hRZZ;O<r
zbrbs;B^6Z|h;p7=d!H(2uKQE7v)`B=Oh52I`%cKj+(%|J?`#YyA;v|@E@zYkCrleS
z@|?`ZZPc)<PIYmCf$6g3hvrGbKf6A!l-3r`R;)M?@#(h=GxyQ&A_^kLhBXXqe^no|
zG%ZrK)V*nWz==}Zf$cu~?|(VruKKyI!+%X=^qZUxqi6T$g>CY<)LeM~wfS%FvIn;J
z^texnDe}pguy+3jW8MBIiwvhYJ&JO!fBa!d(Ql6zzi~AT&AJ!uaC1;~JM&cg(3_5%
zTR%q4z4svgbd7t-eDWv5*{h-t?|gJ>bC*%j;-;KNDtm>;jLPegDXQ^JrSeJdiz@D#
zMURdC<K>I=r*&JSKh&?tZ7hxs^hn+-k86(KDQS$4Zk?i|ocjLD=6@WleJkHQD7nM1
z>P33X{lr<5KQ1ha&AeCqbjKV&<6WK0!tV^enR&$1*LjWOg>K%mpZq|<AfNc!nu7RY
ziqHA=3thU8CM;flc-Nf;50_Quk2kIP_c7%xof{k1BwWvV%+HBSD)=Irt+u1}d=;aK
zE6^`wvsQH@`3F(3{V1hp7vX=ED%-nr;_#Bo{%Wvz#em6MwEV}}2=d?NZc7mc2Ys{&
zQM-?yufKIg>WPlLXEt+2{f9q3sdz<kf4fLu-dR~zm7!I~mV9BM)2@Ry1xmM-`?u={
zRV0<WJ=}ls){vSB!KYHR{c3J3tXj5P_28NL5M6M0^R|csPd3dgb8~DQw)1zvOnxW%
zu)BQmp9c;d-7=ueE6&Amhq>oS`}xbW``y^*5+Dc;nG+({OS?UU_8E0|8C9hd%FMqi
ztJW&M^!cuNp+(SCyNG?h7cG6e?n<|3<VE>s@jqmYAG7m=mb7ljgtjZU9xXQAK21?k
zldqij`=OKDj3+Go7+<?Oz3PK@zsZZ;5*E)5KLrRvR$`NH^&!-rCL9QYnWi;O(pwOM
zevmLG2)7AQM|S0IyGj#cR_vwaWz$;_0s%!hnGz#6zUhiO+8rW-I;@!CF@oNL5U}~q
zeu4)-Ru!EYeT^oxS+QJ8ZrWQA0-=`X>8ZCX=an5*w}>FkicNSXg5Cl>9V$N8Q*%Zy
zLj>97gE8ls3wjGAPL0z)I_TtxMiCUTVhK+o>@AQuBdDPG-Sg8kMNpR&oAJy9y#*5Q
zN`5MzDXu&&f`+Wvm}e~LEs(f`dONpvU{$9G_GiVWJX5{i0*Nb{z1A7Ovztc~j;#0q
z)H(p|{{BY@_#|QS15v*>vgugFti2*=z&5esSqXa!)Wb8a)*IVt`kn}~%OPXUvljIh
zNc>lOYr4y`8CD`_#EK1hhW&aALcn8VMiB9{A`>>Sni114{&+)aucJjX-N3vX*!G2!
zNrvoKhPjS3uOn^y6ld;}>{o{Q0m(eT3_8b*8&)ymg@9txfifMqqmC#h11K}NoC%bf
zT+RZnSzrKe-vpIS&>am#G0jlf%;grSY~gY%bZvzLaQims+J>1_hzj#+!&AjHtC%Vs
z3_}CatYWGRE@uK&CYQ5-M;1`w_H{s62Rt#?1BZIxin$4@n&2>`2+b<4YUXkaRJCxq
z6?(Kn6>i@KJ=(aOM=JA3Pb}X}I_xH0G4CNA_F(3fyboEa6Mf{ni-7ikb=gr{6ip<}
z+yX|ofCPIihcwC|CD@`|(kK@*<I%Zjr%l?9Z$4Qq#L0u#^J^bLnyCZ^mB12fltvn)
zk(Nj!OJQ1Eii!SKKE*gl9b{;Eq8M+fw+t<HEDiP^!l4n~ewY^dE*gotr<kQ)ONU}A
zhrZ=-BpQrjE<>-&a42Hwx%i(uRt;GTDOO8yCD6kUrjg<_QV%smzXRddrC5{j>__(G
z>@}f#Xz(|8pKIXWmqzm)k9_lv2TX+$vCwljpqVOQQ3afkWyqaS_;ok=Y9K)~Pf7k$
z(hfPo%e<>Pnitk3q!@)*p^L+%kSpvsq*4sU<R*6Gv{Z}Lx@a_tQOnhK98#%=V)7IF
zaayK8qZDX@^re_q(x{a*L5lgt9&|SMIs5(PCfbkLsiWD9|Mn}PzL4@drXw<sT{Im8
z4m!wLG~+06<dC<(TL(FWW_$!b9Ex>`?b`!rCe|mmFAHcUQ8&>Dxr!xeB4-;)bW7wc
zv1g#m8NAlh%o*r&24@V-ltSH7XoTFXVQDGkY+9wztrYexp<@SS#2PuLZf56~si*01
zv}Ysp*x98yav&$rOe*B3LNlakaC2xtL-rV-?{0xb$09a}nllm7A03C-pEYM8^z-zj
zoY)zq7k+1*ATg0xpi|MD_=JUle-#W9X!c1e9cUx7HTxu$!R1UqWpX(SP+44Vf>aaK
z#zAR@R5O=bAl1U<R!FsSxeZcnn4^QHyVSjz-pMZ33;Hi$uv%rXe;H)ZFfl><A6Rx#
z72jQSVlu;RM?4BxMA6I(VDSPtVhc`@(o>`}vVdkTk`@<9N6h7<MLFjAC-0AaXTB!)
zyR&e0(U`fDa&KQ_r%o{#Tnt8G3rm1!3Gl;w9C#k*@(JK|0{CP3C+PHvdj#r+PTgGo
z0-e5axd*!R;2o1@dSGA=3`HLN<&Tk)!8cP)*{RN$Fy?5n+Ymth^Ho}pCrvZcq|@y1
z=%Lb34o#O%=g<u43=V}!!#ET!4d>7S=zM^?W)(o^0xlnf&Ih?%2%QVLd<Z%p;__kW
zd>AvzS2n9PR~VGlsmX)056^!WrOA?@1(UH&WiY4=PDUQt`26_%i9fcd{m+vT-IiE2
zG*>00#+LUprmfbI+-3W(2By_OjuqJeT{l3^X?iPZx|Nh8#c1XU>H381I}M}j7Tdq(
zu7&#F-Qh64()j%)*33*&Ka>0cKc7YFXJOtGGnxvRu=X)UF>Z{T8Cv2g#*gtcLwgp*
zq(E5;>{C*+ZOdA*Qu&>R_D^0~e`F_iK9J=Ddu+{KAlu934j}K~ruHf*Uj;d-{tC)o
zL3^z2Ybblo<x8ae5>9oRX&~hdT)s!j?_n;Cj5qVHYHVZOoC*w5fdf`-BQV&=<pyZb
z03Gna_n^T&%tw;WYmJ|quVypEP;6*~Zjr_I&n8Xjn&8A^4^pNI3HF{sp-^%tKrsfB
z+EgvUo>eQ<N)836#~?!6F1y`o?#U>$G*dwO7La3+s>h}_y5*<p=(9E|1PX054$Y_q
zYHidP%|=i!1sY=eTS>iE(hzm>>$zXf{XBi84co~&V6YB&;D+mg!Fn#&L4!KzfgYrp
zdT3CO`K<WG&9!k+#5X_lmbnK-XpPKoXKzQwHpb3q-m_`!w;&K}TL%5gU?9@2V)N=)
z|6o~vHf&YEs|sIQD;oAIi5q2ezPmjFY0l>J*S8-i>^H2Ju3POrQHPC?o3)!Q4zr)N
zpDj1cPARbOB>~MQs#7c3cj>=9Dy?S2@fzm4GaJ3sVa?XO(LqKxn4j0W?`T>3&1ci2
z(8PzHzUiXx&JpPK73&DiyDQ{~cK(qx=vAUQk(=0<3_vH@_Nb$i1{?TZ;n7T-PMkGP
zgGI)RIE(61<E0!r3w6#yYc$blrW_iVL(XaHGBm!7xjpmfyIjYcdAc;Sl3FRmL#zVS
zD!}1nKqYf|HK0~=xeoB_fEYht5BT+5eg*iixLgIPDk#KiUWe3mE>}aUn#(nis^M}i
zq-wdGNAmMXG45|S$=}W8JtTh*mn%trC1$kZvvJY9G(>UvTz*5IJtQgT{}=%Fon%ay
zX1s*>nF%K}6&p?Q^<#l51R=pyDXv=Ms#RYVGFPh4_w)a=bDP-SR&FOTGnZOf%WNzy
b2>ZS^gc+{d;;JRCn&YYrSM6|Bj;sF#l!R11

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_specialization.hpp.F3D9282B71057FD0.idx b/.cache/clangd/index/gemm_specialization.hpp.F3D9282B71057FD0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4c4edce95d221c43179ebf5a3900882eda9b448e
GIT binary patch
literal 2058
zcmYk73rrMO6o&885xC3j&OU?zak`=Ve$C7-1#RtP4N4&vYqX_=S}MgQEkeLuHPB*#
z1`SnXs)nG3w%B0M7^^jg&=M#Hu+{~dCItbtQL2R|q^7Oaiu7@3k~?!r_9s98ckk?;
zJ?F47KQ}in5rCYn`L$Ky9i<ciz=&Sa9g%lU8UR+s0Z`k%=i||eLDA6sm#eAy@Mcf_
z_=(iwD;Y!nzcoYUeP8Yg=jTO^ci+G1H7y>ApWa{8KUG>dIC19O%N<!6eD{w#Pg-B?
z7<aA-kIrP&S5Gc5b%P~yiG2s>OUiq%tvKZgS9GWE{b%Tb*;jt?=<PJF>m%LZ$n(dX
z_uk>)zSm0bWEB3|YWjZG7mF=|)hFm5x>~#L{u!BcM=J+nbE}7Iyp0nn8O5K8gM}@5
zjw0c#wc&KAW#4*jeoMoh#}5wQJ@qskt1DYIQ_$#nvvy{z|47@GPy>~`a(ZkEBoth*
zSLa2;8y+TauY7bbRMPwDTbXsAJQeQwuD-IpENl4jwyW#>zTI!MY+E~@a5g(;j@DJZ
z9Yj^pR_9K{j$dtIfFT92Cd#Cf1)yR|KgNZQ;G@}^iwsCYaT=ATmj%Zu-qNve*Y_?k
zK#yV_rK4rRaZYE~Kp=BzCIieUPNtG+S#Vrl$X)o!aGha*3B`I!Ps@ViUF+72c#E<(
zECEz9U?B~klm$TfooQH2{ie(p$Im~LMT1lXSj-k)769WkAZ;lEns)azU%GYsD+WA)
z9zauck}UWEdp~r<y6(p}FhGam6e>k03y#}vE?#@Mc=;Ryj3{O)RwoON+lxn!ZTxp?
zI|G)WIF(A($%5n0%R-w6wl@9800zbBRJvXk9Dh?aeWoTf*~x%J6dNg{P8J+rzHy_!
zqV4rd41g$JPAz9-!SSu18$*YT=dUwBi()fn*2{w9Uq(;&ovbf-g#k2*4U|DA3y$Xx
zjQ`%bb8eae2`FAhEz`?_<KI?TMj~4}h8Pf!VuoU9S#TT)mq*`21BN>%PAhi|{<dh=
z&AA=;tHJ$)AgOs&NZ<q~hHw`lNNOP!vT-&$hHx(-NNOY%@^W4U+6343RfP=)H8g9X
zEhc;xHyS!Cb(wnB&e@f-_^&3d+-CW<UHcDxo$~>@Da-N_zKLe7(3-9~8?+J54()_<
zKnLNR&`CHKbP>)C-GuW%58=GfOE@3&5zY_&gySJkH~|WT6QM}B01OZ=2!o2lpG>2b
zM+p}J4-zV%G)1^ns2ZHrhR@(}LXb2|D&*sQN<-kGLXb33D&*(<3WP@sLDF2Q5YO=n
zga-^k(sZd%fD0(C)sx#=`$cGW(nzsZzcm}bRf@Ioww0>0^LD~Hcn9H}g7aCGbO|oP
zxdk`jJfi10mGtsn;ufFaBS^pCCmb*EgcC%8xJ492!Ue<t;evtS3+nUGo(*D$w+pI`
zv|7~KxHy+mK)hoRBrO;f3UWagz6oy{1W9W~g*==`sWRR-2$GhK3OPBa@;JP85G1V}
v6%sj7xe4zc1WAiWg&dqinbb(QvQ+vX5RH^YlhE%tb~x$%D}Nj~S`5Jd(^OWE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_splitk.hpp.749B3E67AC6B8B9C.idx b/.cache/clangd/index/gemm_splitk.hpp.749B3E67AC6B8B9C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ff7861103617ea8a723e187ecc1a34d9f71f5b64
GIT binary patch
literal 14040
zcmY+L4_s4K|Htq5&~&qFo7%07=3R3%2V=1DZ;)a#P)rvv6HIqi=t)XRrAR7yR1{D_
z5i?YHQXUmwNl%hWDm{rQsZ`=mF+;OI{;4b>(@M=$)Zg7R&%Q@ruY$MV&+a+*-0#1$
za~=Ei$dMg!9GCpU({Id}`dZEd9LM#O|0{Uy)oG3(j;jjcxHs1RU(JkJ@7Oh;-LSm5
z{h4>F50}jpT>Zwr^PkQgPrtM4lV&Nue`)LCxEGG?xIC=?pTm`n3upW=O7$xL`t=Vl
zUv7Wm#4YFi8C6R=uJq9>54HXJO5Mo41$=aC+x3r@ke(%P|1^H&pg(30-S9&Eb5rJh
zc`*O?>@#0QKBf63y<p^(Q$0J?>JKj8DtX;=NHIF-@ZGuFmaVut;#}~ej|#gVefHrG
zw*Pa>k5fvAb&k5Z(Z2C$aB0%q6P9Hc3I^UB)V;aa%1<mu8jqSwb#voF4|iR;_gBj<
z9Dn2X%VRHeY`$^&IG3X4`c#ao+!{VIlgk`9`rye)gZd7?%H_}0b#E#E5BJOJoXB;X
zV2Q{bu_XW2SRYp0Pxr467JXAR!Y11`d@;RYcS28gWy#l99!N`=x+1s6*zed^S3aHn
z!qFjvKDxbVb<CstTfZ-~pL%HK<+p~XxASY9*Y|%h?`yHofVF4G?Hh5Ry}0*xw;yxA
zwDLUJ`u9K2{3q*H;<tU{B0hQfuV10@%N;w!?O&(8@?e>9%dQ8H9GbguzKh(L{^9fS
zJ7*PMUcc;t%@NB6XgAxR&e?R<Sh;Fw$o8=0uB4So^KSq7>*>bkZJCvd+y}>H&zPy*
z|KZxl-@TBr__5Su&y{aDwB?0))7!#sOe?6L=6UkF&(FU&E4=mWA@S6@oRb576}v}^
z6E7WZK2^A8-0?AM4*z#?`O*(x{Pvaezn#z8`r7(`O-`G1d*_EmuaEm~!_-H==nNk+
zFI1dZSl)GL-lL-S(1?rk!X`KU_U_WS$j#lKCPutCShsS;e*5!N^ZKer<P_z6SJbEF
z(>Z%|r-%Le=HRFPo@09RyR%E1`VE}AWWxh%9vk{`@FaEb)PjFc{CQaO;Ym9WpRyGk
z8L{i<)Eeg}Q!e$y4Ej2VJ3MV+>)LwF9}S#tQg$y-#pf$i^;^Cfcj#DjtGzA!-~%bs
zhX1K4c|2l@;g<>Twt5N@vR9v9|0UOu*YxF&x`i_@y>(69H0;M2U-s(m`u3@f{cBT>
z^4q7b%O27ha%$c8zvUXvEI8i1e%-p4c0L_>zAl$+OQ=z_j>;Gv|A^^ZST+`3J@VHp
zr=NWN{Dalk`>)+Hukg^2hC5TX^$9LJF@HzatutJ~n=_tYso*$OR@zHb`-~mSbI)eV
z0VV&DCMYJPajM68Zpt{2;34?-(8!NYf5UTjDmV^`<1ms_86NaY`hBuyeNAPuhUcE5
zl9ixX35lFauk=g$d*{AK$_HlhToRSM4~q98l~bAf`z5-YBVPJz&X%`$E{;kDDF%(<
zR8REt3xazhS6!Uwu=1RZ3J!qcfNv(jz5Eh!*3H)6`~0lrxuH}tTrpg}2ECgn3SOE1
z`b!_>@tl(i912Gg4if&?N3Wgg9v{|=STjSVv(Wo2JaO*_r}QsSOk+XW**<e$;kjY-
z8sik>m_dfwvNsAbRp)tbEEQCOqSCh>CbfT}u5}AuobutnB6-e1B?A-#hH<LLdiw<<
z3cD+<dknwwoE-80`y*`w#Wonusf>N^N&ak_EuK1ZhUX%v<O@)I;ad*@{F0G>)ikQC
z!#49=B9){o(&gKw>+#G?UDaEAR>;pKk_wK1;)w4WAsYWR()#V{zvIf{TRay-C4&`%
zndg!@Zc}d6-kjffZZs9Vpm;&P7wrT7D`bAWX4Bw5{{2*c5}w6L*Fkd~^!I+yH66dd
zyywDi2M+PvGxQp-DPEHoOzh*o#sq0d$s3Bz$9T>|1xrA&#J3)nhy9WluPl9J&+!?R
zJU4<$_JU$Bq~MZGTe$t5`saI1;<*$mc^4G#`mRFVFh6tH@~E4m*6>^gm6U*@1V-aP
z`Rj(16#aKxU!EIIC2K&j#`k<54f4+=|C64}4_&oP=D9Q~c~kMGEYK<Zg2DqW8=Cff
zewXKBsUTbto{5e)7draqoqel|=aQ*l11L85=0T^r;Qi9P_YPm(%5(8l5~qk`p3L&u
z5zovK-wx+FHx<lO%#@$Zpojdk*!bAWTt`$?JkLc@fmWfF1&<KF;J~2G$FuHC|0mCl
zqJob>@iFU$pYQ+dzjL4aawgBEQprwG?DS2Pdg{@vZgs|;<Z(RrG?h$IOp%8fsPYf<
z)xgOm7U(kbTmlsgQ4C@H^vsms%j#Cmo62+1R4`C6kcr!KUFkb+KK%Lwo_mrCMkq!w
zLHW}U8?MxTzN?w%hETz1h3v3_P`L0!??<!p2LwMz!elSK2z@R>|9d~Emv+5u`1(NA
zkqn-5(QB*)#ah;Vy0c~(Bl>O!o=c~ad5U?AmtOf}_!EoY9u>`VW2j)7Vw$`bg9rH6
z;?=3S1+UpSu9xAX<;feIj~4?k38G*Oqzp+g31*~VQ?NPk93_L!pbspTWDpHThD-)C
zQiv`@AIP~<h!|pI$P{8m3e|<`16e={6+?{-nL^D-5xNL{z_X<YF~Z1@DZ-2tt&7$N
zlCcymMjIJ2MVpc0b@BRur%Ul-ypbVOycsD`m#7bTl9XslG$XlmE`6XnNG_AhjFhTN
z)d&1XN;RdLk=#1BKHxf%+vGMQd2}9qAj3)?lgEsdrOVO>3Wbzq$}%H)bzXg-?n+*h
z*Nl`S<`@IvBITHJ%t*PWTyr4VN_o0GeZZNdJTcG6kSWiMl&{Oz2U3BQZ^}0#6^ez%
zK=PFeO@(HpB2$q$kjSKB&=rF|;NLQf#b9LFR19Y1GSHQQK9EkOG7!ta$g-&n%*Yj>
zs{nl<&r1~`R)CRZQw5ljt3X!;`arFfs=!nQX5?xRtHBuPO{8itRf8G17EHDB_%xSG
zb)c&QeIUeT7VE&svZ)Tt$n~JB2Ynz-OZ6bugOO!ZJ(!UjK-U2JKn0Q-Kx_ab%cce}
zBR7Js5%ht^NooYK5sWOG8o`X*1iB{B2l@u72~163Ms5aOGw1`^T51MUGnkQEK-U8L
zK*E<=z|;a}<W|tNf<91WrB*Ptf*H9D#5OPnnlz~mOl@FBZU<AlJRjP$NFAW-0DYhV
zky-2jBg>`^Fe7(@t`qcuc2w#FQzw{_??Ct+h@`y+FWrIEJ1`pgE`;A@?IpxfLSoQf
zN*tvuFCp$FBpvPL#8J-LmlMZwmTO3O4PnNwA*nTl8NZf<*Ru9{;;3gozlk_DvAmVI
zx3c4JBaUsXeJ643WO*+M-^*TqFG=0Yj=ztD?_=$4#L>on{xjnEjOBLXZfD0oOdN+<
z`!V7;#_}l=eu};RDUy1M9se5={tatCM;zza&;LLiKd{_I++FPW-Nez&+AkBwWtOj#
z@aydLuang4?D&6^@V{Am4{`LcpTA8Uw^?2g<X#ZOJpYA3j)g&tzbp=NEJmgooT&b{
zf2#cS*7_~yhCI$ojYQK(CQvu(b94Rak}<K*&X#MsQ*dDw7gB_DhCG6V6zmLkVTBJ)
z2~KCo6D%PaoCX(GcteUIogt4wLJDz)xUj;9q=ck1<Oz|GLY<*5tni^Jq3H~HLM5aK
zXM_tYd_+n_Izyfa2`Snc?ZOHlof4hSkSAI~ig(7lu)@cu#HTaliI<QPorx~2_lYTq
z=?r-iB_x;A<-+RkN^zw#<Z($zsm@dv_5!IXsp$-PQY9p})9u2}z@6ewXUOB0kUUP0
z3;P65iYJ{Rk4Hkva%Q=(YsgB;N@vKEB_Vm8UKjO1Uh<}R(;4!3C8Qi@jte`AoRpk&
zhCDeEQm!-Ch22JON^Uwso?Hnj&za}KZX+)xFP$Mzo`jU|%y(h$k)M*E&X6ZxLMn_a
zOu|m2Fu5>|p^QQ=Qc+w{5_Ty?$wg@lWfXaliXpBTlCVQ5hU8*MV>zQ3yvSt`R|ZMg
zp_D;#8KkkCQ3hV*3W%$KB<xTsAh`n4Sk9;bFLD*cRY4MVELD(P1!*j2RDl<{8se%U
z2|Ji-NUnx7mNTlsi(CtFwetJ}9ZoGI*UIx}IHOjce}L;Ct`3s0!>NPhI!I$VqYk{t
z^$=GNN!a1kLvlT&v7Av4UgQRdYk(x|a2g=F0n%8`XaFyABg8dA5_UL^klYAqEN3)=
z7r6=Inji@~o+e0cf;5&hn!t<P3~|kngdI>bBsW7E%Nfn!MQ(w(7D&R5s0ET+AdTgW
z7VsjsLR>2(VTaTT$*qvaaz-n7k=wx81}^NF+90J3(pmPjfrQ)+&USENN7W7~?U2s0
zryV5Z4sdpW3p=U~Na=udmOULHA$Nka6I|F~bwWxfq_gbl1c@^JDOW)M7lc>??-I>j
z;-NLL-`4Wr*&lT8pCH#ji(th%C)fl#Ls3F3Qm`f1ighm77HnrIDmWI&U@=&+&KYb5
zJ3~>1Sfmh3h!yKxh%Lm<P*g}PQm7@=ighm37HVfGDl`@;!V+P{Iu~J!urm}D5sMUU
ziMC>$i?&7E8H$RIMT)n?Td~f?+v4pEMa9P=C0Y`#SmzRLiFSsf5@V5E7MB(4oXh62
zGZf{DMM|}#TCvWh+EVQdMWx0fxh-xh);YJ$ZD%OT9gF0#c&u3GJT{M=p(sx*QkEsl
zighl_mStxsDk~PrYw=pK&UtNKJ3~?4Sfs-6!bq$ig^t1)q@wVmNURt|j-nW(VhAsW
zNURUV;3$R|<T3~^gGj6gW#A}-7~~2FuYgF*{}te<fEeT|2(N-j%=cB`sDc>eYOqv;
z74v&FII1BAxfU$7V8wi13yxZOd?vqJ>cEQmyAB+65QAI~mU^&azODyHJ;We4fTaPf
zn4cTK(Eu^XjbLd6E9T=ya5O>;auZmZz>4{|2^>ujgWL?3X0T$uZ3ag(#2~kTr3I{*
zUt7S@0x`&~U}*&_=F?Viv_cGW8(7-Fiutn*Y;9m?IjRj}k=w!2F3&fxV{HdpJJ?x{
zYKK_l4zP5974v5Y*gC+@a#RPzB6otN6Rb3E^74+>)(Li&qdFm$GJVc+e)XTOV1B(w
zG#AMznqTRH%R>Jd3kuH=%^5P53Tb|mg}##(l&vM2wIqwm=z_>H|9J}vOM^6}L4m|h
zbA~MRow}fGGtq3u>(Rs}%Y0`qC|gf7>xr8VYgX*4*5?fzM|f#+@5zGy5QO7+=_d5P
z2?3@DHqhtZF%7*Q%?$D|zT+AC*SxfaXtv;NXn4so{~-+uzaW|~a9L?Bl7+sb8k98=
zO%r*VPMf;3Eb|}OpztT6`H74P2=DooEc74Sps+khQ;y3@lc6m19p0eqeWH0E*XtSQ
z%2|)tmaXW;OTWPQU*LIMu?sN%0?S=6zKi8QVf3GnNxuN+r7JM{3d>hv^i^c)!18?j
zXGJ)pn?!RH3q7rhve0*8gtB!+vksrYxU~L*FT6EWVd15RRS&D_I`fiFrBh>F7geGf
zYqv>0N~8YEOBR(yO-JJ;o64rfIvb^mQd5`XC8x@%#(FtaHB^l?F-euArasI|BUK~S
zSlh;^#;CD=WvDXL^s(?#rYciSy^WV9swS$j)?}-))wrjcs+y|CebaQ+bT#gcW~pYW
zasM-0HCv5UX^v`+8tc$})qFKpol;Phf*Pw#IjG7(ja6kisFs78x;Zbs4XU?6O|uCv
z)qtu-{(PWHtOM0LP-9it2&#>s#wxHGRGUGKS$_+twt$)@9bVcFs_merVZckfK(!0h
zn8kO4YB#7cYwrWqK2T$pJ_xFVpr%QOmkxpIkbJ#B7Cs89qoAfv$V*>=>MKywEXhkZ
zpzjSBK%<PVw(NtxLrL_(6-2WFqn66$EaE?(q-ljd1Nrm514;b+5u!PQ$$>7nEb|{r
zqVNFG9Kfjm$Ddz5Gkd_4XgRAD3@(Vl7~M?<?<N7J%OPLSe};+I`-NzJ!SBUhdhFRF
zdq&Nh87eo^N<oXk6RZr@V(=K01}z3ph%!Wr!4s+s)nf2OC?m8OJkiQ%Ee21#GG2?p
zlc-G8V(_?>E-eO6sxnoJ!Q)oCwHQ1erALdwlcmhkV(@sCUM&VsjxtA!!IP`Z)zZhw
zOL@vXEe21%GGB|qQ>ZM|(g(*&Mam*A22U|4i$ROwQwGX1&|>&hfU*L#^m+2~v4XM+
zv=~0spsWThhEFXhYvs=e!lw?Db)d!YsRv~}Xfb>mK-mCV44+0&Hi8zzrwNozpvCZM
z24yp7F??D;*#cS&pH@(|f|mLVFMsW>Yy&NZPdg~v<?98)rvsE7pvCa%1Z5{^F?`O!
zpmPv~et8jui|_=k4RqeJU-}Qu(JxOE&1v*=x}dVqe|(O@{Y0~$WYUPED<=#6hv+E$
zgJ}N15TlP!7W$6TQFfeYj-!nFt}OE%sH3coXzEBd9hN>2S>``hN8$HG^F7Xw`nxRj
z9j>En1JP{2B34>If6g-v2S3Z;rCT7}0+9~n<Ct5Jh};804{Ps%m>$-C2XuE}5bEzh
z%pKN#7j$=7dkGOs$Y9i$kc1NALM|m@DQjOs#3lE0!V*?rPQ-H7UP;8t`#GVK)h{RF
za@M|z2&)Kl|EoysDl!b`vziF2S^H`dyPCDv5M2#n?!Sh_)DUL<YKg9vwbv7|o*loQ
zB-FF>+eE}otbHpHx8BbQTUq@!B5q^tJBYaBeoolI>UR=xCu?sZLJNC;EhM^y48{B3
zLxer7eGiG<!`}a1qT9>f-(C{4m%aaeM7NK%w-K?89lwnvw6XL1jEJAH_I4t+-_Hr{
zto|?&53}|TB6i%*2_3Bd7!i-L_7g-n!QS5q5_^KZzmr5b$=Xkn*psaN6w#ex@9z|e
zImO=pH$?XhYd=TCbL{x%NWwXGem@ZL2iD$2#IE}}p^Mdb6S14MUn1hA`#IqftG`Ud
z%dGt>3BF29_`I)@C$ExZ<ZC4O8f(8sp1j7|uM^#M_WrJunCtBQ|4np%v-Tb$_ORpk
zkc1v~ez%Euo3$?p5*Gw9`C>s(!h#?sUn~p~7qa$6LE@tOIbjj2UmPSZM*HI5N9Yrd
zE(`9<ORGrODq_bQUqwc*A{of5N!V)EzM8mK<L;D~TF9stl7{*{By10>-$UGcSp5kS
zc7nB^Bw;7-XZJ}~f0aaDC9!z@t0es@CVpPJMk24FowgTpA@IFKg@s@d(JaDFlJ;D(
z%=ao4%9@F$8D-m6s+xN}wzc+QUMjQ{hB4iWr6`P+I$nOs!cq)jY`0=5gD_fLdHE#?
zO9h0n-HN3O!q{%bRt=GCw_>Z6zoKHg6<ZxdvfYZU9wKRR<mHzXYz+`etG;}i3ARRv
zWV;nx6GXDzime$U*>1(w0+DRDVrzv+wp*Fp2H9-4GPxbH*=}WW2V}F|%H&SSrrpYp
z!K%y^|F@wl$mc6zEhGY;<q0zS1nWE3i02wU+fRS2Jouxvs`LR~>fgJ6;JXC5C+7=-
zvG=Jay{idh-A(x2Bw(j^B|mBZJ8}4|e<qrru@|DtAq)L)#-Z>$(VWL=(Oy>;`reO2
z*;1leioQdeZdvAgOAcl463x5#g#hjMWts0?Ih4IaH1D8{w&}9W_r@H`8i=L=8ywod
z%QD}4b0~X{Xx_uV>bzZXwYU1wrM}I)@UQ{*BRWB6ps~SAq97XR{CLSEm<+fdu?Q9e
z?ni8b%|LrPUWyW;47eY03Qhy<&v<F5Fw{VohL@6rBm?e8MhYVh^hSAUj4;MP?}wK%
zgbV}jM>2&>1MWvA3KI>uAITQ74fGlD(o|up0rw-*h3N*|kIWKg8EC@brP;!414j59
zVU7VKe7-Q>fbm@lLMa$9zRN)<2Ls0UauAk-0pt5^5Z(p@#&-<}HS*^J@x2a&bzs2w
z-Uz}*FkpOd24OQ8Fuu2dumubl-`hdh4hD?xT_EfN19fO#+6}^PFwjWgrF|gm0|UnQ
zK@bjt0pt4+2#4hB1$HAxK{yHqjPI{N_zDa(zI~f{-3<uTQo0KveeS)ONq6Z!VaN>#
zr}0m3Saw$b3!3Pz-x1As7@BnDWufmiO_Y5~G@mlEoX0=5#rziSkayWSL62S?tP9qo
zR|kv1Mx;bB(TM)&5?x00$5b)Zi2mpn-A43BkLWR?KW2$pM)XIo=ry80=IC<t=#RO&
zTs`_@u9#~?$`|vE=#PcELOuFpk*-LO{#Yay8Ih_$tO6tYTQ%saL681c3%XkQCBJ~b
z)q+?nzvLI-CJ>vzi2l<IVlx=ge_B9n0VDcPD~PRNME_|6T^s1pf7(IUF2Cd!@Sk=N
z+vV#AxD&)qFrxpI5O)cFS;9+8h+_%+l20XZRNl|-N>*P_-1RJPC62AEeg|>vxS!oS
zSbZCDx3Sz#9PO;WgE%_wXLkpyKS$i>SneW@E>?euI4<4K?n|t`hq!xKUJ&G1fG;ia
z(xM>8qWjss2w!@l`Am)u-#f|}9SewN0oJ3NyVkTu3~B#Reo@)~4<={@y{*Y=`hM1d
zpLt%&vE*3MRdZ}Pc7~#IVv%w!xmI-CTwAW4p{U$gq&!QWl}=IqLdce9XDBK!7AfD7
zZ$<adx8>Uzipr0r^y<`?b9|F8c=aCzQ+&Tn^8L^MX9E32{wFebwzeF3HSDdd-jgSd
YpYYs7FE<{I&*85V@mDYYIvIcce@EP{^8f$<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_standalone_xdl_fp16.cpp.EE84FA2A4697A71B.idx b/.cache/clangd/index/gemm_standalone_xdl_fp16.cpp.EE84FA2A4697A71B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e712de5d3f495b3d1cf1258072d32e90e0501229
GIT binary patch
literal 6576
zcmZvh2~-rv*2kxYW)!BGfni4kv0)g7VRwOHQJ5K40U4BiUlhd#1q~RDB6$KDFp7$x
ze4sMXxUV76yoZ8O;}UU;_}l`<C@4=PF^WEwS2JI#yF4eJ<30WBe^qt$t-7~v59)}}
z(66mI97SG4Ii0^ao8WLbCh&jh;*xoP>Tx&%lEW!){#OEB`f!Tzt<e)EhS8Bj-W!F}
zd);sR)6-?@oBm|YHwPw09duOm9~v?o7ao3Rz$yCF>5h?qzfRw_=!$RI(jr@lAfv%j
z^<m^U!s?Xm=+@|>>b7SQQ?|Yu6IR*XJu)!ZziMFc<!-98cQE|Ly&r!5Pk=#@hjZId
z=9-ZYcOAFg|0ViZ-{8}ov6Zs_elq#eoSrjpUOxXI9`|7VqseE#9sX{1+M4ULB^Nn{
zl^JX99Jn?&t8+o6AmrrUPk;sQnA4|UJU(xdbFHB5AMMvC8U9kS(*Jt@>M^I5R<FME
zecHH}L!KsH%DAlA8|sWZBfU$?E{^C~?R~5Vq{lZ%jz>PZcbXd=zW?==-J|Ajqh;T?
ziT<z(th=vxzFglxb6{P-_6uv;*Ll?3m{RhofWCUAJ?G89jV;{)XWRA1N0iqYF1UF#
z-Y?3%wPO5{ugjC3NumyVR-aou$LyA!lTz)vi5H;kS{*ZBvsZ7oO@jOLpZch?R!dg@
zW9+r5-g5g-q!-s6-Sg6`Td`lhWgh?6FDCHsE<AjtNZLa@xD!}V+y0~Li<|YIzleRg
z?$GJNnJM)NB7R4K+W%yQbUfF*<%vz*!#%fbYQ?=9TGX}fMe_$AHF&qxYbr)ekL-)L
z{?2`WTV?xEn@#eo4lisAdgZImSFBJq9N)Qp((1mlwYSTomP#|?mPwA^ytd-h&sEo#
zKAmRYJltO{_fB<r^I6=tJzMYG{%T3#rJi2;;z6TjyZP&>%t=evwM+$u_guD4U$tuy
zZ{MA@Lwny0uBn+#7v!Y$*<@cx^m3aaRc}e$YrP1J7Oc0Z>WL{nu-N75)wc&$j0-Ti
z_N+^hE;g<EDzAOCdc_{MA(OnOfDLPoUlh{H+8e!3j_oseze~KNtmVbLOCw#rCmerv
zFY4f&(45d5u06Q;pT61myfV($4{r2&+Ef@YZ}Th5od!Ywx1+OZ`+p9v?r;4w=F^up
z+pnZ+*1hi<p_biM-M4SBKW87KetY5iY;{)i*}=Fb*`fb_-m<??{Pl2`fAVnavfckO
zJ#sZ+)5(T<BM*nJfXVZGy&6-NCm!pYydcf=-!C2<DDEwjXDD1=3)cy50`I32+5+<k
zFf?D|A9A_;OKW@2Uj;rLtDIuyC3a0|9W0LD5%ey%CGh;YWuv)sZ#x}4{==a?N5}mB
z>XnPn%C5_;^B<Qsbm)ov+6t5OZ&mq!GNtFN&*OK^KNa*n`Kl~B&f9R+S7kq)FfLxM
zFg#&4FzDZUf;N~}sEV5!LZ30+%KNVE?)*6OPn{;e9WC2-VdixMuMxY^(URKHanGV}
z4oPoa>aMs|^uRvv)*S!7>Jv5g&nK=}ZZVvDEa{A)(YeH9JI<V!UTt`<pL06lET?oS
zog;?P&+-4wdE9_s6`z~S;m|HT7m?0DLHCa*ZvEcB`H7Ox&4IEr&sn51NH*FsKy)13
zaaO=dD&Wvw23{haL9*~&keL7KxkkXrfU>uNw@7D@Y*8N(Kr{wb^0_H+a0{M=P-l?r
z@p|d^9k0Th_}n-s+wyFUbq2|v@*9><g4JtnxXFLrq8%I^Jdqr*x4WhK<i<UG?rhlG
zjAtg)86>OfejlEH@Ldw0n-66XPh_AoNS^)DXRO=xvR5|TIcV?+<`e9ZoObO@QOCj1
z`+ROD>@DO84Ri*{1!>EA+AQlH@wv00Y|b+`(HSH!)HA<4mY31S=O#dTJa4?Q&LDa9
zn?LG23YS+1IPuVHo(7&Gok4QV*N&?jehAF4;b#4Hi=OB>(VdaIUA_sIZ0;mEOW@#>
ziOFMi2FV+OLqXcj4>^49OemZ3Oocjw<d#*xiE{;Kj`F$lp)BAD40Hy`hlxMw%aeZ`
z;&bOg*@S0eqBBVDP#etmSe9tW=c=J>#j`Tf86^L>ztiXB`wI#_Hx<fuJUe5ZLGl%k
z-iyJ{^4$5{Y$!YNoJ2Z<<OjXZxtGOFy?kyWl&yKzMmmG!XQcH_wLmYF&&`9fE6-J=
zGe~~kvt;3CyN!JL+%zcL@NA592FW91Gq`8#2nQZ#E|lHJxSQ$>l1uXElrGU8ntw$L
z*YKk#A?Zfo+z5PX{_;JF8rV!5GH4W%t_RNbY_qW=9!*;-2IK@u*8!>yxT5|fy%A6w
zac%-s6V9!GYGoOPA<s}3e&}ULx*E7ugDLE=Yr&MYU>eKofqXqsu-pLT4LCOgzh)4E
zbQI3;wY0I!KBfk^)v%8-pG)>c{A9R~OVV+QxD?czq;nKG458Quh9AYe{hhb6=fkd>
zU|)qwk;e8-QYA4oAvPxI`0x2oV52ByR4UsjOPR%xtY+7#U*7$?7m#$BQJD#9gpvw;
zt~R5fr9sIA7<`mOeEej~ySBfNhEJ*nq17M^4MWm3Ahd>Mx!v%^LSFW;9zmCh%FIwR
zf-V=8W26BB4KPM3N>Okw+Wds(64WvB?H=QzZaDKAU|s`kQG1ouaKMpsW_55jvlO$E
z(QT5>SL8E<5}UE%M+uFjP0dU#QS_6vnVFd-dOk^8m|0k&6(MP{o!9{dBP2>J#z-P|
zWl3TuaX_1$q$QL@jFCj*%97Mh>VTq$q@|QpjFCj@%96}Z=755kq-B&$jFCj<iYR(&
zeX)gG?Rw}Y6JHYrN=>t3M9%#i0`&DrIzSvCW1SRa5F})&5*SqiQ^Y9U!{z_GG!NrS
zo8!>}k+ik3wJFL+&_hD8i1mrsOl--LL?{tqUNDndqBl*y_Tk>Z%?~kf{;|r~Tr`!;
zJ$u^ssr}}{MJ}-~q0nuTF10UZC_j1g%lQK*mcvF>Ahe3zhuVFbSPha;xBPYbP8Fik
z5cm-Ma6|``zTrm)lQs{@>cCXgTbqYu8*y#|vL>8cfvlC~g7mx79|7Y8IIVa^JQM3A
zou|lSsIb<dq&Dbbp(8;D6T$lESp*$IgkYo~6c{N9B}OVjg;6LGicuI5hEX^Xj!^^=
zfl(w8iBS|0h0zRR21e0DG)6H*3`Vg;EJiblnHa?paTuuyHAeA7JVvvKSr{b{2^b|3
zi5MjjNf;#)$rz;&DHx>^sTidZX&9vw=@?}Y85m^}nHXgeSr}y#*%;*zIT+;<xftaU
zc^Ksr`3#}*2c|u3?SrCfQF&0e?ZxZw0YL}F1jV9#vEuZ7pPd)}@P&Iytxz+&h@^8B
zxeQhIyT41jR@V$0g{VT9gMy?JRf!Ctx(Y_Kwy0u{uPWeH#U5WIy%yN41x&1w^m<^w
z9!SuHNV);oH2`Opn}J<3%hfMbwtn5UHZbHwRw6r;k5SnMQ_SC$YtR|mS`9g)R?~35
zc=^Cx=!{C>Q_0Tm%cLbQ()T^S0pB!I8OfY>B%P_uWC*2sIH<P5LxXPqeyH#A{l*N~
zH&_+S_=ThsR0#|<XBw+D=F5^`qZ!H>jE6}&Tba!eQzjaIR34JFm9dp6ssc&c#@NOb
z6<#n+T8nM5#Aqh5WD}#rT4KwRR45gpGzkw2YpE?uGNDX_jsSQ(n8_?LvX<E*+OZ&t
zh>WXxs6CN=%j9Ua;E6o7+>a#<P-uV(@g7Z62`N~jfNO`J1N)wRzBIY7c^$kRWFBOL
zTJJyFcEHnOdH}pVQ#msS-9{%7Jb|?pGV(r2+nU<iq56ZQ$D59~L!m&@cBXcAs6K|S
zaufTqt`f_|DvZL!YL+B!5?|JF61hZ$QJ6%{lGIJ=%eql2m#Q!dld4&gxygKy8=;Hk
zG8IN)GBu(jX}7(1)(S=N{WZWq1B{Rdj<;(LO-T*wgnKAd8Om%+l1^2oGW5+r92jia
z7780xfy65I%+i)Ui8U;L7rMFp;PfZAxddI}Si;01L6<s~GIa81@6)#XuZF|9L@T42
z)3jrfI!|A<^BKG?a1ywoeT%jz3=r*C3R<Gh=Hfq_JUz_elvGL;GgFdIQKm3-CVAI$
z7eRg=Y@}2vncF0ttV(7GEhT(__B#hX;L_7C6XS@^hj2R8z`U9*?XFPEYAS=`g|KQV
zn_i|wjwI>wY2^x*G$2X?nAC`N0j$ThUrXpABwY)fYuV3?Ymd8%Pfok}gdm{cR|~LM
zs0V)aI5z^nMx0xKUkl5(ww&&liu~*J$&XSP+^&RJfF=XWKtdwGwjCiAU|Wlj3DEZG
zxqVckve;Ju=T!&fb-*9JCjrY2`9_?ZfV>IkRv>R>`F8!Q-h?8fhj2wAlo8AtkaUJJ
zgCSI|!4=VdJ0fTO;%v6*%(gxa>>H*GV_K7Rnlg={M>U@9f_?FgaM@y%G0f)r%~Sv7
z7@_4Cuy25OKmeLP2~)OpC3}GMPg`u{VgAKuuvxe=oXMdiovuu0=-Jr)UvlM#wbiW9
zN$81wU_ictORW8*MYAU97U0~%E-^~7@a43hwy4>Q`+8TpD>{FNgH{3aDt5b}v(MS^
zqt9HJ-kgJ&uUL3WP{AJTTZK>|7%3<PMoLPFk&04b6iS6+6h?(%6i$U>6hTE`6iG#5
z6h%d0G=rLfQ8X2eQ4AG>Q7jdU(M)P4MsZXeMrulpQ9Koo(JX2fMhR2`Mu}7+MoCl>
zM#)q%Mk!PZMyXUPMrl+UM(I>KMj2EFMwwJ5Mp;xAM%h$0MmbatM!8fjMtM{oM)_1e
zMioF%0hnWMq+{3W{@o)hVQyOKwQ>@>Au51Z1;BU}@LI+4`{>sG#kW$<!CyCOL0~PI
z&UUN^f%P~yg1|<cTR>n7%bb$@1@IRk_!o<k+|okb1=`>oP7!-q%w8^JF7@bdt*`q~
zufsQwEh=14yl_!TDMvTP|F>AgG~_H`FH6|VTsm(7{3(nRT@b5B=<~*mHQ<v*B7upq
z&@{>{Oyy!`7-eo_VL8s$KOkt@^xzPMmz8^<*w1z1-`%ygv!6gYIyp<EGB*!zpGi}E
Y<qlKFdrl6uvW}Rsm>Zt1j07?N175ES9RL6T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_streamk.hpp.B73183768811D4AE.idx b/.cache/clangd/index/gemm_streamk.hpp.B73183768811D4AE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5d2164748ace7907d19a614ffb01b464d2802862
GIT binary patch
literal 1450
zcmWIYbaPw8%D~{9;#rZKT9U}Zz`!5~#Kk2=nMW8I7=+jv7%JwR4UNsdZ6I>h|Kwq_
zsE*}Uw|^-#EtKEV!yC43A9v`!&el}V$?N9M-1y(lbmQYKh2DwT<`a%y-8t#}$DQFb
zWcK`;xv9Z2nb%`W-M+r;#rv)wkp1=3(!}eKYRFadFAA&tKd&z{oy&9CwQ7%bY^T6F
z?)=xOA3rFa-gHtrm5EDP@sE`X`~QeTiZf47GGq~6^ZZ()_=+`&I`PHE+txpPvN<ES
z_wvb1`8i(?p4{nBc-N}qq@s=Vp+db20gaCPf)tJJ3I4mozNsf?--qtoY<E}o&vh4n
zw9P_yC(E0K4leUP_qx_d?tN%qboTJA1(WnIJ$@@Zed)ug52KHFrtntY`CO-;?0@ga
zj&BqF4n_1I;ndf1_;51tQ=Zk$!}t0V<lQzb7SxeFnD#%F-DS487xUsNV%wOaSSEHa
zkb3oo%SPPoh1m*+rEBN#K5}E$KPY`6L~`Y8nH4&ZE21sFi^rA3TgC30mB-~*ZoTgH
zHt~eFMOC#Y=TBb0TJY}Ls~evB&)B-_<Du=lxvF05oAlQA?cIClKOB7)wRgv&=*e<A
z`$|1(9j6DWWvu0E@U&WdFKS8Y2j+{fZ)VDvJWKiW&?V?n#MH_6&pw(fSzMW$R0oVx
z#cKD>YI`hZ2{Ey8F$fq77|VbO1`ePc5GZY|@6YIP59elL0ZMCfX$pY}nDkXY7mnca
znmuextengOa_n-dU;-xnW8>MV_4S9cco^6@A?67PvkKcYa56D62yrk9Fczhz6<=mz
zU|^Hq(dWF-T!D>8KvYUp7Oqb~R9;jGiKil}j^LH*l<LEr&m&OkUh0M5mFbkh<+y$=
zxYz7$8q5qd+(OhDVSuQsXaEu~NHiA46I|aZRl4rgVUQM0Q5~2tkf$eVh{Q7%HG}bB
zb~AE;0}vK6umEJ@0tXySm`A`=#Z(8T$JcyPzL;*$1fcg-HB^luLB+tpBcP_ChQw3X
zP)FkFd+Ue6Oy?0W@HRl=8G0MSc_Ds6>MthRurl!oh;WMV!`usW0k5er%mq<ltLq<@
zM|yxf%pod?P$cIf7m9F?rl_VG%+8j@^;Nzrjvfa(x=6H04rUUNXIf;3;1!D&BgvIY
zmMS6SN=-^_5xg?VGPqp(2Ui)(^F2~P_Y`Rr86XrDc@+g9c*R=9aJfT|dyhWW?r&k_
z5fEY*5`{S$7UIBo0j1C0L%$cSd1$Th)PVI1JHL20C{#p6<Pjb=l{8g`SpXAd;smE#
zm|h-%3V{lEV*9mi`awPG)0=_zf-Hd9%TSb>lL+Dz7bO=5gBZLZ0s;iiNSs$%5GiQP
cCdkdp!^h9S4U+-JI)cfEVDclFprRNG04N3hp8x;=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_util.hpp.1BFB1C62397B2C8C.idx b/.cache/clangd/index/gemm_util.hpp.1BFB1C62397B2C8C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7910701d358b151d2116f4e89427f6f215331b8c
GIT binary patch
literal 7886
zcmds+dstIP7QpWb0YZ9{kPw6f1g?Onr1DbarA0zeK|y?MeFPCThz~4+73<PPs7S50
zASe|PX@a)Js#UCu6e*&W$BII=7K>YL)%q1tM39nISoV@K%>1{XU;NAaC6oL6&SV~E
z&YW}R-g&da!s;9-O1Wy*mPE~_I4(s|V)mP|DS3?zhoY)@6t(5xjr_!Qw>+g;z4M=G
zw{x4m|EXf(-r&4~#NC@TS5lqbE6#kVkj;wpavglXe8rTqj@;l+Yn^r~ia+SQaxW!%
z@c!>U)fq>A?Gltfx?O17cj%|6uXrWr(zI&#4$s)RON#5Wtn#ickKdsg)z^K}tZ&WZ
zLmB05>YNSxSDyCHW;#F4{O!^4xN{m?^Naq+k2DrcvbbhC`j_yA;L6o;zr8Ko;jguh
z_d1rdve^8d>*!Q@{y(3(7jKfkcWaBv>bUFXF&}J!une_eZn}8|&+)@!UYmbdyXKu>
z)!ALq+xzbA?D~hd__oOPLZXAorWjS@f8wQ=sT{v|4lETcU81_1J(x0kqsO09m|L=<
z>+hfPzcsq3N4YX(W5W80qMaX`Uo%~K_T6(=;%*Mcp7snbT2)$-@KI+=HI#ID$~rn*
zo`l@Eo*Y>Gx2h*?-*)x?x_ke{FPGK#22A(qGE0Y}c{}{(^r!ATS{QP?cI}fXkI%Ip
zSX$jZ&Sh@M%%eU|Wtx=uOL_O7%Vsx8^N&~;Bo+VORvzl>?-sIgaZ^TD)?iwAPrrxy
zKF6iG{6WN_?#w2Wufy_>{h0M)=G5_$i?b_ZD*U~>cxGw*d#S}qcN)V^%ycNremKB;
zTXKjK`~C8!JsroM*2z}2?VWLA-}95L34<}NgXIf)N98QLKBy14_K(B6ds1c{-L$7Q
zPS9H2*%FY~_RKt??2`Qelvj9HQz`Eyt|$SD;^{pK4{U1>Fyn|vQB10Ns>o1K7N#bw
z6#)Zw+e6(dOGYiBDVnX|Z|W~G6qE%AD2WwY!N*^bcHy&uIYJKKhGL|aQn8_+Si>8-
zrhTWcD6`zqEaX_SVwt7P#!w)!X7RpfYuhW+gd83#wzjku8ww;&aCuR!OFDap!x>}!
zS9^?zFS4R6{(?g3dv)`yYdmO5zz)vY)Y;NdP!?uf%0<c!uIzAOGA(>pX~r>UD~Qd-
zVnZ?ff40Jz_GoxiP!uNQSg`$DSz3t=1?vCu!<@IHxwUtN91$zFwY0S{6iD29vwWOt
z)K>|X9A{gKDF=&k5D)*LDqX*rTy5S~+b!e>*%oXpZEOq$5_hXJ`BNKh+blV6uq~up
zq{|Hj5>v^V_>{GBcDAuK)35rB7n#z`AHaSlpss1=F|Z$kyMgH@xeu5=lKX+_$86&D
zMp;SIirr>3qnxY^7+G#6baE&`s>v#Xv@lK!Z=$i&j1I=>NG^p*rQna`W>;%E9!)sO
z2bw7Yw<7RE0%Ux^Hy9mSY)OinPv%Ck50_@@z`YJ8;<?s?dp*ew;NC!TBe*wWwz;sa
z;PmbN141sN1xGEbmtk!W@kA|nV%CA9j+E<QqK=f8f@3KuFNJ_o2*UrbfiX4Uf(FWE
zYG6_g_>vq}12Zw72g!MmlI(IG#**xN9{e!Zf~1zLUkej!!4t`CTTk)&zo_sw;WA1u
zr8incE)(h%N|4G+HA3?1mo`!lgEX(bZtQ%Yq!zj66u$DEgS5=y8$2SiUYfZB?sxFo
z*+(xZ>FR!3XtXttUM)V@+MFOZx*;ciPEwYwq=3U^vLP%RW??hl1=DuHbi`xc`lhnm
zuZ8~-W;JHBV@y@Xi3p<y>2klao<C{Xf>*w5ZgnR|rCG)}8a={spL#uLsi5~8-OvPG
z{|vLS|Ft!`A&;oihod&SAA99GLouGmH^g=nRcU69V$PdrAJB|isSd@oENs~fOtE3H
z1m%Hm9!x=YMl*ZCcQ45W;9G#%Z(rKGyxa?MMk^@5;pxBDnZ-S2ot{Rbk)qMDG(n>v
z=q=53O!1O<qUHw427<I8(Sj7sgUbd&iH_t_kd%^K1Ckojjkklm9qf@C=Q5oj?<Bbg
z<UJ(oL9Qox0OSLhQLe@18*(hXOwo#HA2gvMWe4wt8EDWnQvksQm{q<`t#>PbFy0!&
z&SiD3e-vNBW)^CJ+6wotc2&D$iV?&RvP_U>f*g-66I?UFon+rk9CxFC>^*F{sQWXc
z4P)VIi%-se>b=0|hSYz2a`NYoj;=7eq1Y`4-&tOL?jD;ob%AFWOhHDXnKQjQ<J9LX
zSV1o+dck`{Fq}%8^9?C9Uj0Z#B)Lf=)lr1@!(xbSxh!y55SHbFZ!X@1G?NF4Je=hv
zRIAK%;*fBoXSyaiWpkANi(nHr1LZ0Oxc^XYC_yT&YJ}F!i=B9_JN5}{T3KM51&-K2
z9)R=#*dd+JOb5t1zzK5~$ht`G1z9i2&q4MabK<1LgoQ=7*BL#@t<Ro)BCh3KF}k75
z<FgYN7XMJsYHL>HtQf4V@W^n2A|fM3D0}PH$sepZveW3kv%6IG{>{7Ptm$=tX9v!2
zcZHj5f1ESu$F7GJ&q{<2RWu{zNku4l(2O<DT7(XKG-JcF5uwu&&Dip6Md-9eGcuk`
zguL0lzU48QBKsR`uS=#a2}d(VXFoRm8anvl^s5c*+Hfj;DCAC%Lv)1Twa;_4>m#_~
z(L3&%bD4O@cm}D7%OpvXh=|n=f$cCA?T4X=6WB>|4+Qp*tcO57$pa8LfLS~6?XmRA
z#`Q*<mhOr5`1=o?pR;DY$Ys%ZG<up@?6R1kgs}-Dp#wQMQ@-I$j}0JTT)@aNhi1|}
z(*v-3r~-KvKBv)46?jyEKT^ln;~y+s`ZQX>CcGL4jT0ILOOrKI2#VCqMfCOQ<Oco0
zu?Jb}JOa`q;DGu%rM~H{+bxpNG@}t{#Hc>aY!GZ9NDBfjh|!;hj?w}h$)zADCAkI!
zH6TU<rkPq0)RNo|*6kp}@=mbsB)JEydq~!UwVvbwupYqt&0AN&t;~fndVP^gU^B>}
zhb(CAXlAxzHW{4!)c#mgF@Z5eIEoF7B`6mZxi}aNB{hn@Bo~090P~rSt!HdL>OT4s
zCz(oNd-IN@wbUz5roR>PTIv?vpYz&#q;m~779S7n`TZr7UM*p^sdDXFjkdS%Ce2%A
zs?v3^+cFCxvS2>iVKp(^-}h_qGdH@2#vhm3F7mGjqZ_)MTo(P}USi5?Zx5HZW}MlR
zdB)jzjSWAQHQ^hc#$snOS2343=j1_f9zNGJS%$X05IqWb?bUqM`R^ZQPMa@c15dnL
z{7AT@nIw6V19q-j@YMnl8Fk>RBe@hpN+Aqw8=5%})6PQ(4lK3cTMGf0+d<h5Gcb38
zvXkT<Q1+0l2c@3m0Z<NLZVnooHZRJ);3a&=PYzY4Kf60My>f?*<^-JKXZz^6^}l~!
z{pI0pCOV@H`ZvF3W#2Jdc!}9|tUa;N-zwx1hh`t**=2(VUa4JRw+nMe{)s<}-J45<
zG~;aJJQ}6VG~;69LXfN2)dtG~Wr2>^RDxtd1WmP?D#NlI;N^f2cfJRBdq~a)UOvhD
zfVYq2IuO?3Cn7Xc55jtq8-UjULbL=l(+Ipq%#TGEC&&4@HygbI=pe$nl%aPt*rh}%
zq6n9g3&FYAr93T9pU=DLWyc1Y78ut8lhH7dI@lK*L$7p@y>uU}tnoO>xV*$N>%aRv
z>yIz==|)d*;QZCG2ZFX7{J-{+tXUs7^mZ*Jd1cBv_Lsv{HSBvnXHrZRPbRk??cnHS
mM>!5v`0H{)HxxDWnrirxp&N>_$G4;LE&IwEU3Tb(qW%Z&6%!Qz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_wavelet_f16_tn_instance.cpp.DC996A5B769C800C.idx b/.cache/clangd/index/gemm_wavelet_f16_tn_instance.cpp.DC996A5B769C800C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0eba24c173ece16ccced696e5e9894858f520f3a
GIT binary patch
literal 4018
zcmZuz30M=?7M@HdK}e82Y+)5f3K$1ubE>SvmJpCd11eRPun4GtOVqw67TIy*iK2xT
zP(PH-mV!mWT7g34DTqs3(Na(p6s^=+5vle~(s!NvzVGGxINbmKbI&>VoOAEF5y1fg
z^;!txml&+bkZ*~@5CqYo|FXAarAE@=Yd(S~Dle{Dv+1sz{-dxv-EOURDZQyrPWeV!
zx_^E7cu!E|czE@d1x1Aa1{-;w>8oG1pO#eG#e7LdBX_I^`?4>s7{5KZd!L>~Vws~4
zHs$FuHGJA@Lr}x0r!Z6NsDa1+(Y^y-Z|h67?#-LteF;B%F8*Vw-FJOS$Ih*d_?5~$
znEdtg?H)A?`!3bh85RVt3%z;bvBKH%sn<_o$+_&CckceoifeJVT&AD7Ez?ZhaA95G
zFP>xf@BVmm%eJjcrCQ0!Y6&(cA5VN0L(XL>mS5~7!?Xquz0jYYecjF(h>nx2&zmUQ
z)s|QuyGL?<Z0ED`<gO#-e|+3*6`Ql9dUs{+wX{>VkpYvdy99sU&EDX;{a(}ORyA2}
zktX*7dLFLv-PGReeQ)$P-H>k$jLtaM&k9=)teI#md2ZXj)$Z_>2anu58Y@1%dZ8>3
zmq%5bZ^=(fcr4uBj}=dRtB^}2f=wfxztj!2<fnSq7zAznMAs})v%a%0J(E}Vrd*@8
zTG#obsnF8B{d-1Paq7O=jvtTzZkCc(iu%t;On>dUylB`cO8xt>+=c{;S##g#*FTNW
zoViud5@~E|x6Iku-SuJb@vW(8VQbEtQV#6a@%ef!R`1jYZrE-wme^_=v_BM`O(4l!
z@x$fZ&yqA+PdfcjoR(M07RK`38XYDlxy!o`Qn{1MTbkmHPWSKoF6UfpvE+QruCOKh
z+ivIx6`Li8PuiHAy3YTRU9*k*_Hb*@%j#686Av;=-gJ>yy{euLtxv7g(lq$Pv-oag
zaMQJ}pUP_-ehqoq?Ap<{Y%phF%8?zIHkduRXZS?TmzC!aW%SIYczWnmH@d6}k*sB{
zyDl$$rki_f)*vDOLoE+k=FS`a^#S4uRM3<*-K$uO)n)ZHSQnqxByV=Q&z4;qVHx`0
z+G`zeIMOQ6DbD-++Cb=UUicZ)q3<m^N(b5uczbttJgzL^{9t?h)XFnmm*>|Uxo)p-
zXOR~-kp4`&yx;7L;SLWUd>6l?dpy0<eb{&m>$_^!Q~EYwa?zcQMqEedBYrXOh`-Dh
zPCc(nI@<hN*JZDZY1obRH_U}a_164`oTU!pHydAWYwdp2|Ap^L--8dbzsdPP(6s3k
zg<olNJvbt+N-_|4bf^UymH*YeHSx-qw{1Quj&ytwrug$@!!6@i{aeS<s|?5co4=QR
zvhT2`?dHyo-6fZv-&e@rnRd=?d=dMq@WFw?#NZz5;QOO(qPA-?<dw86DB7MCd}G%U
z8=Da|^@KGe@ga`W#s(EfCU_>p|9(-Q)O01*(?MHm>N?F!S@~1S&lmP798ZXz*lueM
z&0D+~NmZ~C$dKRLn!5iwqjj<5tCrn$4iDq9$Cn;gTSp9i@|jEPc=KGe<BUk{z|chW
zSJ$R@SgoA7xS==pa+cNf{I+4g>IbpH95<IGZ$(EgJvATe*4Fd5>C}8O+^J%@pUJ?d
zb0`1V%TLX09zAF)X|S19I~{uH7?QmsBYud5AY6;~{eCs>mONDU5h53G1?nn6-#)6J
z#+&MqT#0KW4_!smLatCpB>=5+w#Z~_^OJ5K8cEZ7Ts?J_0Q6$+=HVB3#c>|Gfu`-b
z_Qom!Xd=_mu)>5p%0p9V+J)<4tP+6sn)~h9=|?-x^3Z6SHsPA6sRW??tfI%h)0v9k
zp=mU2$+a|A2|&jVy}S4L*1J$X5<}AwoCupF9gdC-qK>kVj98?x<wO%9@LtA8R@1bM
zBU_N9$I)8===O^H2D>)j?&G1UH0{atG*Stm{K9SLZ-tH5gZRi=nhxRw*(7OlG~X*9
zFC@;*Z-}hmBWq|{%8}Y6X>qjP(@)|oFD!C-wuFa9(6k}fP+cW}@~6MwwUOL$U#f|$
zq-k@3xs^%)`mOe=mY+BEJmR5IG_B9oS62x@BU$pK?2LI#X4N%R8Qs@4SKT=Fp(S6A
zi*P3pJQ{i9eheZ+gcHabUru-veheZ-q!UOPUru_Hehi{SloN;<Uru>begJ_A($r8@
z4&+7#u3!w~JMlag1PfSr8H9-pLcw@K>?HD75G-I3We_GaNCo2>xs%LeL9l>DmO+@x
zpcKGY+Vh=M9t(m6EUN512Gy>kp{jCdD}lHe@;wv}htPUF0YWKw8iR;HLJVyqln94V
z9FYj26e5j5WFRSqb`(m6Lnw|+gis2Z#vm$?5<{B{rNSW;M<qfig-Qbm*hMu|r<d7w
zm?+E#lsYg%f22If%tuk8C?9YZ!HG80P@QDvBfJE~OHn2;cqNKg!MG8{n_xVH;;&&`
z$i|D<(E18CehkL>7+!!eJW+xYr6|}KUFel4Q3c~hlxTwS3`)F)aUq*1VngdI*u*gy
z=VL?x#_(hbN|vG^M)cvSM9C@`H=<+{jAu~tHH-_{WDy%$U%@7i!8jiy3owSKN>Hj4
z1qr0fx)P<TVBCmOO)#E8sn;+rWK%_KXnh5nItJr>j4HqY?^Ey;nSHrs7{laZk=O@Z
zj+h)S5KInN0w#|U#dw2D7$k-+9_2mG)CH_g>Rz@<u&DGPi!T>?2>rml5*4buaPrVH
z4#t-U*#{B8IA1QcmlB{l`SJ++2m+KLUoQJVwip-(vn&l&mjK4)`#H8by4k);y*@!w
zNUx7n3hDKUkOB>~UT;|l(3@dd8(i_E_g}OE9K-Fv9_eDk2s@}WFwzc697frJux^u3
zR~@3=R?r(Z_cf0KZDG$p+V@@_9Xqd_US%E&rVFP1g-ZuZP3UPoZ9M_Fs~V<qcf^cU
z?x(%5w6eq*FYK-CamEW5D;J#c!qd{z7I@L9=-{l#s12a?n#?y5fFlfcNk>iDJVDgJ
zDms(OrU|C`|C4SIOM0!cb%Kw+afwLe_}5O+mIStfMc`%973KMOoyCuR^bW+B*cvS6
z%fqlRYbFs<Sd=xBh!`x!8n_6)q0K2<F);U#;Jj4RGAvOh*e%GL(JOV6dV@7|;kZS3
zGkRsNGB2PPv~iAxs(k~!bh_Pfk_j~7PLNEXNq3TD0!_J7Bna8myv744`GvB+4Q&kx
zpdI9wwxWCrKvs^j@LIS<9%T`=a2GtvB5UDBca%ldzVBt=6+x?2J|Q5rGJG}Uyo^{4
zIWHquL(a>n)xdfB)sY^je1c?WC1j^52T4lt_fJLqJzW3aDf(YLGhD|E6W$Nm^2LV!
zLu|FdKjtN`ON?6|pOAzkzGsnSW|)wXn4U_*(Bu^yuG&0x4W8zFEq!f0U7Zkvpa6RV
zeu%)-&`4;uY`K@WkFOu)Y`oZWA?di_|IM0MSX$ZG;&vZ6ED{k*TwL9kerWAs?&KC|
QY$BGXW5MD65_TB!A7(rDrT_o{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_wavelet_f16_tn_instance.hpp.6DF980223599227D.idx b/.cache/clangd/index/gemm_wavelet_f16_tn_instance.hpp.6DF980223599227D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..996f827df9971058e26522febb77cc8e21280f29
GIT binary patch
literal 806
zcmWIYbaRtqW?*nm@vO*AElFfyU|<jg;^LB`OesbNhPO-%3>9<sPP*vDY$)LRS<`OC
z7XG&zy2Ut`^8aPDS6p&(i}b9sN6x&tru1+3?Q4rVzBxLcn9z9m$-SK}z3bw&&OX}u
z?jq;ZpBvu?RlK^8qWO7gYR#F5Hw@z4t8Y(;$SFFpKH2ow)dvxg@6WQ`zV;@kU-0Of
z)iWoitorq^az%Gd*>~l2@6*MP^VeuREm?mjEc@V*5FS}yHm|7aH_gs#&&O>Ido2*i
z?BJ~Hbog2{pKZ#mL_@QiJDwkSu;T2IGSQ1U8&!G)CUwtsFNpr3#XS4YLC=?ejaDnJ
z&Yb)?|0+x5gAGP=<{mC*ob&L?-}zUbuhrn5!e(~Pw?)I~)RI*T1vpD$?2LA=oPM~m
zaqEqsoBtm+UVa{8%ewv^!>MQdF~ya+Nj1QbioN*n&dV$R{CF6cIT*NtSc6o!1z81E
z7&w_28Ti;4)@)k5)<Dd`hKGR#C>_okuFNgWDy$5X-gVXJfXtc;9y|<eKxq$F4;5}+
zR$jR7Cmjac_*Mtc;9+0|O8c_<s&MnO^24QzQqzj%fw2_gC#3#jq75qx4>zwCuL0ap
zZhkF(1GpL7f?9$GaOZLhYYOYbTn7_n<ODmu_t5VJYaUuFr1r6@HV73vfy{N}^+ht*
zk>3}|Tt`7)m<c@G!VbbdFmquv&|HwOVdlaN0t$nC3loN!2owhS5+?lAfb|PIzj!yu
z6=u8+FwH#N{AT<PFg-lnf@XpaNUksywujlyP?VaJ2#hL*;-chYeIU)mt-SuPN{;Cf
WA!Z(4P6i&390YK}7z_-&2qpl(ohms1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_bf16.cpp.3DCFB3ED2888403A.idx b/.cache/clangd/index/gemm_xdl_bf16.cpp.3DCFB3ED2888403A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cd629716cabe9091e17ee1372078e4893c315f7c
GIT binary patch
literal 2632
zcmZuz2~bm67XA5&*?vMuVgP{;Ms`snptL|l1Y{A?1CdQ6fCLmBK$^vd5foYryG2j|
zMHW#TWZ$%r0)~cxRtjwdMHZE1wAox{SVRG##d%59n@(4Cs&ez+JMY|gmvetb?CjuR
zLq`z1VCT5-keGuc1VPx~8yypMB!h?`zfut-F2BliC~9a2qyI}x;ETMzq3QG53E?~Q
zQVm1~7u0jl=%yQqe@zvve8haC{Hzk6*9kbuc1e^SH?j!9NWEd*1gC-9U%F}G`ik)M
z^A9r#t6K|(AI(1q+S(kI&{V5hQNc{@f28f5{_4M%$4u|Pe`R5zb!h&<qM&YRHN2pv
zKuh&g++9s(=cTED?23fguTJ_4tw*NDPQ3q}v1ECz{$ClM^$uR~TN=F&3_Qb|@-?bi
z{_*-@BR)NY>hny?)45fJy%|P2F<DFLON)+2?;an<yVq$QGmU%Y`!w_TU?^Jk*vT_#
z^6!1tIo<oSM-RlDbu-$hFV)ukVNvYkJ-@c-JpHA+YMr^tK(G-vY2oJF!`y+*)(K0$
z(>S-Kk1tcyvxZ4}mYwRB@6QIh=<l7i;@Pj^U|O-+4nbpXvo^K%Hyg*2e@mdea1teT
zJb#fg7}H>ozPsOvZQWq_UF<5!J=Q~1RUJNib2{DiC(Y4|n?$6DN{c@8e}v8#u&Wrh
zkMK9&9A8cCZogu6{}$&A-L%nIno+3ZVLVed=VsW)5!dhr&!IJu8%>@?hStt&l%My_
zpti9hhKkISHm2Q5sc%VfiT6LjTI0Pu9{Ku7xKU<?{a!ZnThult<!zQ*rqF(~k)>K)
zX-%ch=Yw%rn{89Mqu|;~@v&Xp?YU{Ut%z3BuI1IY65E9W<$>FVk9U7)zQ<MG$rg=u
zrG;6Yd^>qRJt?;ENRaQ8(kF5~f8<JEQFV2n*&X33_X$;<Uy;7Qi4z(+!rqlpvL~`U
zrnyXfF0}OO*h*!#@P$%iX|HqEhLQH{EAHfNOZSgjlpN-it@w%)nnG#^yuzrKb1B<Z
zHF8~kyyaQdUV1(G(G41Jq9nK|CoSErZ20><Grb?RM@^gX;^5Om{;iJM`@|7hYgf!p
zCoZ%I)=gWO*bZoq>kWMT(jCR?>DBqS%e9)+wpxQ+eQA*}+N*E&_P0;p{dCu~Qk<sz
zJg$Ry>g$bZQ|g%=7ewYQV-yWLbCutC1^M1nX_*mYA9(o%ZEtAJQyPAcC9?*_pY-hB
z&K&68nI~%X+;&YpAhRd3f^_m`mCG2hB>7>6$-6_Vv?EzhuS))7_<FomZKc1scqLF&
z^JBzn@%dvR%c377+rrOU-EeHJSe%?*jd2b0S!VCg9wHB2CmQE<j=OV;U*|_8s(3GX
zwMw6c4&TW&&iqV~jI5iBjt>udLO>ASAw5nM!MX!O4<LwSJ$b!`+<@7Ye^fG@g`}(2
z-(jc+fN5kJOKu=6Sd9N?%Ysu5hJFoTHkqv~HxL$<_Y|CM$}GcBcK}n#RF>R8*pZ%m
zal-xH1q}58Fh<5y<Oaf_ekTSCRnx{X)D6INGMz0q5RN4Bx~OFTLJSoHm_z0$%MFBM
zJES(6hFuIM;tOD&3Xd;05dOZ$=&qCD<u(R#5WspQJ*wP5_|k3%f?wZi5Qd5XOd(TP
zas%O?8;p!Bqz@M{bU%RA$ZBeG1K}1i?Wo{{HyuO&24E(c$&?!icUHEUHGO_yhoPPT
z=90O}as%Pfmj==<{hXH=x(~ogWF@BDKzMpEHZrk<xdTJJ0L&t@m~sPQBr4=kG!eXj
zVP^^Szw9WH{-;hdB)AaY1M^}?_=bEF=mB;aP-HEJHv!6F{2K*|4Sz4Bz|I4TtnCUE
z6WmgPg7ftu1*!q1|49Yl77siaQbe?L!!Ms6<`5VXJ!3s9*hRl5zeKcDJB)yBC1oWp
ztYAP~{;|NHo5x5D2?eL<!lp7LRGg{{Lu5#3I87G{ogu++Oc%xyY<wzbdE`?s@T+E_
z<_4pH$^_ZTWT}vyBQA!AKn>0K)-XG$RzQ(SRe=(NpH!f@^1UDhDi{<*CSL{0MFO5j
zfTZIt_Fcw4=58QHo-NM<vXT=U&a}OJ)hOHc;z{5(tQb&*O!UwxEDZ=xCU&?Dim#6g
z;7uSS4qpR?SCBlRFID>G4$6=)m6=?qd4`0g%;G|k){;JkcWv*SAu=RvE?WyKi6P-|
zIa;tL3<;OZ)q+z2Zvzp^1`CD){ex=A1`2M22r--kInW14LpD|lhmr|+>i<gy67k^&
z!9<{1fM?mTDUjCU8qhOL2pp1)oWda-*M*1X?ws03eP*HoYR)I|+3<p;ccYiKOa^YC
zfMDMc@V5$a;L8MjMR9!&#Zyt}51j8u@d3s4B#KWdt`|^zQE@#^#3zW558ee~;PnRn
zheb)F1>lz`xo5z+SKV%Qj_4QQkNCqh5XA3~$ACW`rQyM0hrx1hh%1q_hD@Q-Fa}eZ
zrR2o6x7BCUoj9u8wQB2jSXc>nTJN&i!ZWtiF*nryt6dchO?*9ngRY)|(Z)@JttMvM
ZztY;KzS-1{r{droi0<*;?MQGz{s%$8d!_&Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_bf16_rtn.cpp.52661E2CDB7295F6.idx b/.cache/clangd/index/gemm_xdl_bf16_rtn.cpp.52661E2CDB7295F6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..276a4284f6adc85d7625cc999a4aadff265e0349
GIT binary patch
literal 2658
zcmZux4K$SL9)IU$Co}IbV}_CBmAMQe^8LBg4&$>3S;hEPa(y+%t`db#ZM3O26`?C`
zC=^jcvc!g#WE$5;ip@v0?zV-HJ#n9zbDq1i=j@#0ci!juJ-`3Q^Zz}3-MQSoWCY<v
zxu?cQ3&Q0Q1W^Wm$pZd?>v9NUMM99&qMF^$`Oj@AgKL7Y(JR&WUZ^`Qj~=#C^Df+p
zTUwCZH~N)dR1004#rE=xYz~byn9wsh$4=U8Dj-B5cl8^l=|aX}rO{&OibM9u!qp=g
z)+J-T^F0w;Tau1sM4JStF_<@RIvH>5n8mtgRerwTH6^^#r{~lvn&?_O)OuJLIyE`S
zx7g%bax|!RZm@6p;K8+|D|v}M^Z5ICyZB7qn{nS)&js^u-L0;(7uZ?{YqiXI4L8@7
z+_Id}5D+bcT>q`>$9<#!LGFWAo;db5p?#70SZMV7&c^E|eYFGf8%odJK>CI%)6I(S
z7=OO^Myb^Iy4NE|8~X26w`*q%+|*Uu%8X;;jH-Kl*+?H}GT*W$=*Y^#w7E`1*S|2T
zg?<ysLvYrzVJbJwlkM@|Ue);vI)<lQ;zmsRyx(@)xwL6A^VN}cqi!z89*vBicr6g>
zo!LI<rtB!x&q`U9A9xpI-PqjFezG~p>?1#wQ*WUxtef$9{-(4qjoV<RVEJu%*T^R=
zJ1pp($|Vol=I*K5)9z=^_|&viJmi%6c)B$P1RIiD+79HlpWtK{k3W4<sdI|flJuON
zPw+orq}uS&GWhZQaGJhQq<4~&m*81&LB}^IM#!LDGjH%FS@v+m=dWdN&s}+JC(=4G
zL^CQcJG>xjPW*MTq;eqc^b7UjgZoAA_KxcH&ZFU)xu~Xdzo!FXfH5F`GS+?fRYh!T
zq4MnnixjWfG41g-tMX^(bHu+H4xTp3TjN?}?d`r1c`*SKH@y2!tLND!o85LV{S90!
z8VV-YsxLW=Dm7J#JT9<`A68u4MHDap6l+_NNUyrT#lW%HEbyO(R`SKE9fi4yA9V*r
zHv+=N{cZc69?zNcGZXc4T`T<)I<I~EcA0uhy;#YwzCGpEc-hYvE@<ffwE8&gYWjcK
z9bf8kyH~PXI-6I=_ZYw0voe)dHJ@U-e#ZE6NlfACb1S`jc6D5Ra5Lm;O@4TO?D6hL
z=J;}?Z>+=6WkG9V{Rc&-;q~=(wEp)~4tCx}0!Bsu!~$if_r#X*iL58tPIZ0b_2qUy
z_CMHv=j~FC-R{YGmigG`%+LcXqUE-#ruwMf=GCLyVrRwtTl-WxClViJG`78s3w(C7
z^2Uz_lYgB5e0|Ee@b0MQ)3VpIWli@l<(*RWr2df76n{J2C~(A?omaorWmYhG=jp?H
z8&*Aj38~h~7H!@b*c>^xXh-Q1iwF1g<@aQuqHj1Y$%o=21_%hEx=)wRCpbRB&>#fi
zX%n?IqyXNt^v68!?jV2JF$+We0Zk^7sZxNn$<m?wt&4627`hYCbRwN51xP#G=r1|h
z{6iIn?gBK4NTNys(yrvpv(vje^Ds0J&=ewtDFsOH$f{5aa=Z9O5!nl9CX=Zx1xUvP
zrwx^<eLsbvK7huEn6eZgog}B)t3V7X!%$y9Gl&eD6d?Uwzd@s$ettUzi2yWPo~<AS
zNT1uzB?OCCBQSIapcRM;R4G9EfzZIfHuCWjhVB8hI#GSS6d>K|ONuo~`<jfQ-hfsj
zDk(|<(%seV=FOjbI2gJc&?-a~niL@YazZb%mt8P{p?-j-5~+$(fb_eel%(_uMH>ta
z05pwAQ<MUv5q|W(<W=y07*3Wz{@aPlk^eGD3J-U{U0`1n9!sBP3L_XU9TUzmT(Srh
zo{F`K3oIWFA4HN&9u@#1?Z*$wMB2vkfRUgffJEX38Hp)wCnM2hv7rQ-1SS8)93YE7
z9t9;kNN4odK0Ph~6|xQ4_Hek+K|m|9k4(C=1pEpCs)GkQ5?X7yFMyg#qcY(Aho%-K
zhCOJRl%wzzG!=B9B`7?SCP@c&LgA4$$vW_vL>V3}Sb6bD1R|;1s{6nM(5j$qiBDzf
zaK(M$A#msBEJxTjv@eiIyelIS;2AO!FIE7QeB<fdYZz$d1ClepV*118Gt-5~+b4c$
zl57Vsd2k!v5wM5E{V)OS66BS*9&Ur{X5%I@dFd<-n71TznjNXsQIAr1N;D+~G(Uw$
zqtO`9t+(Yr#`juw{|*k(8FU<ahr(kp7&vSe#s@V?-V{s@BZ4tX-V)pfHDNd%YM?7Y
z;gT06b10L5=l*}z;Fbef5wLVPY7klS&SXR?xCV?2TL6b7uTAEVCawbyeW*X`r!Z`)
z0dAQk&r*g7mS<m{GoA@!tpmlrCE&lx)Pb)M@IPeNIVhfs!g%2PAd3G-c0GgQuVvSZ
zD83}So|40-<)EG)9TzUCG?^cnY%lo;k%06k$E(R!qbvkEoRSX?3k^YnA%P&Fe}0Dj
z`574>6}KNOJ)*tj<i8-UQy^g!MH*GfP1*S?wldj`uEtPVuWn;&@9?Fg(>7yOL%V-i
v>1+MnE>mNp<|b{Hj;@};W+RiWrskH`xW$GoW*k)}*F79{`Fgu6!4vr(n8TeC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_bias_relu_quantization_int8.cpp.FA91569EB26ACB71.idx b/.cache/clangd/index/gemm_xdl_bias_relu_quantization_int8.cpp.FA91569EB26ACB71.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c44995d421facc98c177f7ef4f2239e0d01cf613
GIT binary patch
literal 6876
zcmb7JcT`l@x}R^H8DZuyLme1;?`^08f{02Lkzzws1P+1}5yT*BKmv(^p-Bl!3t)?i
zHD1(+8Z1#M7Hq^|1oXzMpm;UWM9tkZ?{4<tt@qDkExxsX=XZAh>Y0cjfB$}M7Hd{Q
zP)-`XZY77sVyU42taX{mGB8KTV&&9dTJ#`&*iHRT-<xD#n|ZffA`L~$S8Y#}R#LAm
zZ}eBs`Nr7Q^a<VB`Bn7LMTz?Dwjh{g{L-w?`!ls?TaRv)?x3%|s#<IFx`aOYannT;
zH~o**bK9qOT@P}q|GiJh`DfqL+wU(FJo_;6&NpLyu3?T)dZtrU@O(Vs-Kc=~Xw3ZT
zhu;%-QcsVU_;7+=XS;<xt^MM!XFa0}oH`0dRuvy^4R8JC(kUy;UcInWJDzvA=p0kL
z7c*MF?x^FOWZe~(!3z^y_veHMRbP$}EZF^~_~YS6`$vAvJoB^3!}{4Tw%8O|EYxiC
z_HT=<lU~wni}@xxB{k>p)v-<XZ=L3z95mh_syoUR?P{xO^oWpL9yr2HjM(^Pc3XqX
zX_b<&CZDF`z6S@_8Ll+j9kbCYy85Q$`-I~csv@4c-72ZyT{Ry3y`xoi4d;*fdrj5<
z@i4wArFxH_^5W#sl^;xW(=Ob+UT~yq!&1dH2cJj15)P`r&giXoh~DvS%|}6H$TDgB
z*Mk~+G7nz(C8etWB0s&N?Ch(lOS<kp3~mgDL+7i0rZsNeRJCr)QkZ_-bbnRR;*fSz
zH~rrpx43`4D8lC6E6PZ*@V?qnt65s##H{`z-|blC-e;e`(7oOf)o{Mj<P5KJ?#$qV
zFGtlok7<9t=b&${uU*24M{7>lyNlniTD7-ZM;!fq&5vUx9UpD?S1(UKelxY}kZ{oT
zgjd6`<mjAgI$_C^ca@%Vp7^>{Ip0fG8@J!tw=l;#xTrX}P-n>I<nP}u^(&7)=hy9+
z_2#Ux-F!d)yOxIpmeEU}r@Yq6v8k=8-H^Y}rGEdTzc%fD`%qPL!x7Hb8$}|e)~(;Z
zI{rSjU{zqmU*BK+bc0U)cKhO)cg7P1_6zrg+=w_<{%OcE&OB~+-xiVMNOzQYAZnnb
zBh&4~-w%?vkKMT5kvHJ=ugCGl-OCEMe^fb|_U5#MB+Wxh#c+q2#+=t@`Jo?L$Ij;|
zy$_3BYxiJ}^^(9b)sXLU_k@+#|GUxhx~2dAi+8scd>r(8rgurOe>kSbZ}u0`KR)=h
zWoKOdY|rbPzFW`E%{kS6H2qrAJ4^kOAzwM3(TuxwEnxWS?C^b#sXEKXH9~KCwP%^X
zHyaGR?cjCa`Q~r-0i!m#BhF6N+P9<mf3mCU);G2$w2n89cMN-NP!Ej`y6=>}am|<h
z4GjTVn>b@xDc{lhzo?$JU3cF3x24CojkQqipCniQ4r^JF{DU_u{^*Z~YyY<JecbHq
zCN+C-ciXYz+nz>W&uBLd&*QCrJ9TljLWWm*T#VsaC-sErsf#-H6#C=`ABfEU{_I}w
z)qft@Xrc2g>D-82jmSQ1%S?-7%Ql)iOjFDlI50~*xNX*x-IbqfmegOgzOZ|=Ve~}y
zg83_7j5VY^x%}SkguyDstGA3wUZ$&w?uU=n)&0r3KD=AG!&CC)r{kZbr7pwITR*6b
z{bShZbLl#EfJgJA{8uB(>nxvrSlX-{5HqXT-8U)bf$b;3uMbQ4_StXTi)?LIv$8g%
z#r?)+u{09RR5QWnvVa}OV$rd@SRJ`QzDWLOlb!wY>6;K80eb<W4R{8sa)D{PzmksG
z$6tvPup<#|#xqlu3rss|_G=A~WiFvu>k*yANm7stOnY{GHu`w@f}?;Pg=ia|jj~)|
zdir;LjbEMKv0uOrN3=1|SR@yi4&ZMoe-eJNOu$}*Xj7i4SS~QVIA-^wcPZnk0(LB-
zwRl<za)IfXn_cOh?f$<A*ozUZ!_!fe3rsJ&RGg%CVOEoXy#mn^o<vbDFul^F^qaQv
zpH-At%MtA$b}*6)Ow$GP9p4mReJNlsMYJu?Rwx&kPKw!huT}Hw5dk|4(Iz|-6}iB4
z#Y}&Q>HmZdH5AcCJR^}@V7hV36Z3ds*D(P*9?{x7ZAH1jbSqMXfByT;UI9BA(dImJ
zWx2rg5zgDRZr7jC(pDl`kEf?17nnZVVP)kJf9($edkLbQcurb!f$7c&#nlcQ7x4w`
zd5E^+St-c{rU#w`DQ&XMJt1H(L$n>wPDw5>Js8w;#-q3Eq<|fRXbYZ&NG>MmUoGOh
z&Fg*@up<y{&9hdL3rs(*_J|%S2=o)M=ObF5r>`m(n11(su)_Xntg(QdfM_vKJXJ0*
zok=HV8LP5bQ>23ZP2F+Z6)3t2Osl|ZLPzOpFs)|d8ZfT`8^YIuV=cH6+yJf(;7M>J
zxHd9z6HIS{nS^hKS<MiL@s!MsULIXfX8=V@D2W<YI9~6$=P#^}@X_O4V6Y1u@%vK+
z=L*-Hbah6AfrmjjHcX{e+V05*bk75a?ge2lDC75)PwA&QR|ob@q3Al`*MS};D0(OG
zcQWxV5bOed%wvzh=1=kj6nz(T?vnLrwAo%1Bpd>?o&nGp03+OAt4Er9b>GZ>MBD(C
z8=!*;9k-XC?J_H?L&Q_iehMa-Fc5C>*rw3^7!jSI)(M7~FyxnI6ltl}usL)-gylm7
zCOC8v1QbCq#zv<zlV0Rc{}P=-CkQ)1Wr7&Y+&NYM&!5oH4?+DQXkx-ryHUCIjx>eG
zp-VwrN><6C%RpSl#N{9<2Ls%ZLsx*f0xSrw1aT$N8xGwMy8Yxh@I_7GPx71;J%c+#
z1)Cd1?*r|9V2%gHewE9g<WDKu+{D}gTR26RgP@#jNv1K<UEqxeq39+sYXT|84&Bx5
z8doo_KxfefdR^q0@e9D8?4I?<uch!-jw^7GsR*lxYH^UM2&<X6281<CTnoZlChh@Y
z&tz=g1D1GR4&4J@J!IiBytoIVF^}T}(nwj5z$Q)6wtBV}*w^A{F_k|lN>KDP?ld7e
z=aqpg!?A`_=&A)}ppNH}S%%|I@F2Jmj2oFQ;0oAW0T;|qza4e^b?^4@DI9tT9EZqB
z%dqzl_}~|1IAjRsVmxE?&RF}t#`VZHTR^1+B$%M+R#0hW;x<rk16|Df|4=lR`XaGO
zNYS~<xnlenA*0+0jYP;PcN*dHyz^!eln>tdz`%vzUC6{m;9bPT+rWMs*)m0MgBjak
z4jHc)T#Lbr;1X~xVd6`mcnL&=zXZCMz?k66pm>?^IH;qmmPK@O)t+2W4Sb8ib?3@@
zzj&d<M)sg*ZcOaKR?}7&>%-h#BX&M-YY!l8ck*yrg8PQLSFRD;Z7)ZU^*r=u<8T;y
z`9}C1yH`JPICKLD8$cT`kV7|uu#t(IK-fgQCx>nZVKd>Q=BZlzRxzE+qv$+oo;x0f
zLJmxt4-8xg(n2OK0%;Kw7lTzXIFSA&U{ykJRG&}3;e)C#ki(noIdIy;;ZI6;DB4J3
zq=PdHiZ+%Q>tKtaXbXvj4i3&ZF+w*W%ZkViSgTo!2mc$QakN6S$Z{j{l8MAb7pFHA
zZ7MO<#WzaPW)d@9?AT%x^6$;yuh2ynkf)i)2(pEsS;(}wVh|ORND=$Pkk>WQU^BYC
zUJ&;Z0w-i>)v|1iEXbGZtBmiS4xM>+%h5e?sJ9lSrHYRVYYgfuOVvnUSFWoP?u$2t
zJhCidBQJSyJtiig=(xbRNa9}N1LGN}A4L5`QIa#W)`z{@>&u~##&H@ziClFfa2lDo
z2{=tm+zgy%f>R<^*x!1+;w9SEX7$Yz(Ke%~zf)&@qZN955F`h|6pz0)$EQ#C+O3Z$
z^5xm)O=y@b#M<T)4dc*-U|Yz<MPOUR#KmA*%#4F2U|K@(`jFY(HjCZnqxl{D9HQ_A
zt)Fkq^|?RnfgbC7=zHU&Zu6;&#>*mA0@34pV0w>Cvt@Hf;e}rxpOOu;&|w)Krr^-r
zv;_?}ub|UA1kQ)R5ATkm+rhb=i4TMOVelvX5pX_2_>$l$g`;JEPn$~7=?du#Kea|-
z4FhE;WH3;sLM8)cDP%FwI)!x%l&z3Gfwph)*m+>>@2k;PHb-n;gjZDltg}Ap(9%DU
z>nR1dQsM(-nTT5%6IXy=1@X-sx)Nqr5?_q18%01_(Lhwxo$JnsfU$wGj0ji?f>LG#
ztOLh7W&~^i;|8WZUIEb+(86o49zDJ36c=8jE&TvWKaiB7`j6OU1tQ%$i0A^FF0u(+
zy&%saD_zJUt<<c<xV)ff8#Nm-4*L{sr)DR{wZ*roA)MgwFNe6Yti-wsZ1GxTS&4Nu
z6W4%k4cHUD7TjuKA~4DlX7gT>g|xiXup0T7$)8B2TEMUc%y36pP&RC3;x;gCBN?A8
z&vm>(DgzvkkUx=?Uu273TrUy!C%Z$7)<5V-H$x7*96ZZO#5}Ml*Q@GY$S!nphd^?O
zOnYF5U`_DYUuCGr5a<t)jPzj68SebF7Jt-3EEJ1zA$0JZiu60p@W0XHelYF_H@u(2
zR!<i!9S^rfX-f}i_mJ#DhNV5=j9--D;2x5;;9#l9pHwqav^UqA;RAiRJ`9u)n85IX
zbRf+@iGhg>v<I~JknK|RUeMml!~+mH0L$^>jy&@DtWGN|N$x+r8ekqKC={{18kx8W
z{F-0^;hQ0-8DcO#8gt*NWxGg~E2{_<tB7NzkU}X|GjR<l)-Z7`DAqD@4=DB!{QZZ)
ztC#n8d!sa^4)p7YPm|@D`a226I)M~MR&f)BIRNShNKw$~5|O~2)AAgRr){Qfhb;+t
zlYq@(csJODP`M#1&&gQMTxSYTL?KV(UkahPQ?JR!tpz@jBKZ+GRS{K@9f4Cd6W4%C
z4Y47Lt_7c3;!R~?*}fN?@W2D&bDwf&80N}t%De@vF(I=l^HwHq1M4<otTLN&yFsdH
zZ2ib>$UZBOZMkvX7~86VhzeLjCaHv|N?3|9PH}krN#7kPdOCMH!$DSnpn~ZVD?w1n
z#LZybJlQ1%T!Hx$ll(VDLKh(Wgh5v7!gXP+v<&>q$OYgqh&stWY>-X{5eCa~1t}{M
z-S&~gey(t^jO}?umdfgw={VxV?fH)O-y3ZDnotH>3Y=2l<CV$E9Zne&mjkDqi7S9p
z!Nirosbu1QQ0NCC9&h|uUZlc(sScVy)-yH$r#;yJDDWqJfT8HQ+_^lWfuY<`2Aap6
z$3T(XNCt}HMlsL=?g9pi;l@m$kLP>McTUf#N6Pm9$tnN0d(5O$(Wf!=TTEtr);U>Y
zjZFN{Uj}ZlSjnWBIMJL!H=Vriv#<DVkO=*hvo<k#RnqFT^fehNsjQ@lmsqK!nMRuH
zNOKiwrjTYPX|5*CbkfWs&9$T%pO%m+8!(ujJB7pLO;u17P)f=oVW7%vKXVm+psJ>t
zx|U>`hu3s(pBcXP8aAFLZk7iB*Hf{MuAZThv8kDbm9?#dqtw~eUEf99&S{p0*gq(n
MJ!j#}00?3I2b39nHvj+t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp16.cpp.FC8CAA78DACFED36.idx b/.cache/clangd/index/gemm_xdl_fp16.cpp.FC8CAA78DACFED36.idx
new file mode 100755
index 0000000000000000000000000000000000000000..45bde02093b0f0f4d1f619f726e0fb6313fcf72d
GIT binary patch
literal 3214
zcmZux2~-nV7OfO9oumo~0m2f4g+ZhUkk%Mvi69_52(sfwvBa$bHAIiNA=-*cqadJ-
zDAP393@YlNA~Hvm(ROPU7hFb^MggT&6x4<wYR|9C`B<LQIVZ2`|6A|=_ujwnMFj^0
zY{3!4FD-aghFY7<Mi9gV{Ij&0#gzsKlF37mRr~5<+coWO_}#y>)Wh<_ryg5*>mIIl
zwTRd|(?&UkmpC@MdbdT=;7e?F{XEa?|5nR|xqch++%}Z3(vSv+6FtI4(;h1JRkJpm
zQ#{~3{65*Osr*@0;?bNQ-P+_<QM`CTe!2INv7g#ym(<&b&u3{=VZT(3OIke8H8Leh
zHSNaRCn;XNMbg%@*DZ#hX6jUH);6`~4RvQ@I#-*i=k}F0_{ObvIUO&KYj<wiCq6PX
zKifV%c24Ua>vu|zO?#cIZto(c+7jIk-B93?liEjptyVTmC$D-k`%dx7_(f>l&7jx~
z&qCY0%5Eg?=@4s2A{~#1Ty$|0zBzF0$Wt=N%KTiF%c2Y?rz=tRNS|L<sj^_s+7Iu0
z6m}_jKgo>S98U8>P82G&-SU>*VI`fX9JdPXM5-6A1s5JxcIHyse6ubY{5mW;QM4U7
zdA!uHM%yUe;S(EV;??M|am8out-fB@(x!%%yo1vw|EXE!uThU*mU_Lp{+3_y&<~5$
z0>{@uJCDEf){0yokuIe~OyQNsj`r@eeM{d~@QpY7I@{izy}Zt$pI;;6hzFGo>;0xy
z?XvUj&u_n1i~kgNG0I1rbp3$p=$X8YvUOUw_r0zq!NKDXY~A9C+Isj6l|&Y)W;l}%
zUn>78uakbN%)wfGt?zEyb7NiXsZ%ow7G%7-cW$?5apwJ-UMUq8+}{$+dWqG||B(u&
znS}RUD@b=-fB4z2J2u4D)uviM7Y$px+J=WZMy)v-qo`MX7T@A7u{~sz_)r#kr^#SF
z5s}~7c6FdBjQqLsx3L`y&vdt}yzE~#^JlZg{3PoDtH>^k{Na*{;OfEef(I5~rE2RE
zOtsyHp)U8@_RI)96tdZ8szJGrB;;zul^}&!QW<rcmEUn-;LY!o_Zl5fInzG#l;61M
zsLVpM5!t4BgUxcQ*KU)1@AdV_4j=q$e~)9ufxzZs5BZDvXQeMZQma*2aWVZ1pZ{`4
z9jZ>KU1D4R<oj<v?~AdXskAQ`_cDFer6JcX=?)ioHv8kB7gVk6(y@26cIvnt@02$d
zi9QxBUVo+V#mL^QOAET*OjuEKG5Repv*g^-6Du;;cD6+>ZERYqTb=kpVc%N4f9J&N
zh6v~EqB>%Cj$&{Y>-MXJJ7n$Kvt-wI!8y?>7U5L^n6~%+V8^4rq@r!gP;cj#nmMWO
zq{)XoCTHK=OnTJJ`-k-t>B$f0{&l)PJYRC7X4_@QBVmK4p)GMbBr7{3#RWwW^92)x
z8kW~Z44sc4lncj2tT&*S^*_<BU={mEuZ<WQ17Mtk8|w{(<wM#3ZXOIO!_b)kHshEH
z^ajGJgRK>LO~thsIt#!&4$oL`Anaf8c3{boj|mt$7r;gwBc9$sIFO%P*fZ<GP7F-|
zFvh_|dIRA_bJyIjvMA`r&`1FDIeZhnf$%Z|^J`qr{3;BM0<by9T%b1)-r7o@400%K
z!4Wlp6>J4pZy;Rc6Tq6=_BjPZ!vW0YaE<i_!e<&u(mnO+5QZiK*oot0sW%XAj^Ztm
zuZiblXas--9D$+UK=`baJW9##vd7Rl0Jh`US?LXgFCT7EG<~|_hoP|m7I8!Zy@Bv`
zNt#glQGF0YX9HNw5nJgEggZJXre3oz>%`D#0Gn`34D|-Wo!Y0zYK!|H7$WlkY$>v|
z)f))EzQ1DGx*9__42=V@kRvqI8whLE3$kvF`Jx&Y|8Ld8=6|IUoFZ%qM;JFwkt)(3
z)&QsE4)RG*g-}m`fEE)x77)Y|NeqI8<Z=eVba@CQfQkbIwA>g3B4QkaU=r!UAdtxI
zApz7R$cz>wB!F5(`Cn-e<Pt-sL8frt4u`*nua>cJN+DBtLWxmNw=U%TPIQ26vW87I
zu;KS<vYfIHkB4sqz|O;N7Tkuy2075Gh9Q6=I1y7BIRuhX&;?XcA@_ngK<NVlTK5cs
z6e68Lpe5Hc2*TuXkYG6*DXaZ<<R%^QHa59?G~$(^Gp@-JTq6Ns>1!DS#gx0Q@%xs}
zC#UJ{ILmo(JFnk=tb6MZ)?;vrE8$9EPvI0#!jr<A!YNFGNnw3)3YXwgICXH!NMIy_
zR|4w^u7>Vcm{vj8z6j@>L^U03Bb#g*4HljU%0M?T{2tU1Bg`0OuqWk=G6>~jScdX2
zo@UGW`B{Kj>>&<>E)}^w_Nhka>kV*9AP|_sX2&TLfr%+>Kb#T@gr=~aj<NsDxTd`P
z2Dq4+n%O{^<CM9nxedH7oD!LeY+(1p5d*TN=L?J;o@T)RYNUXq+==NhLQn~e6vCw7
z6KI#7HINCXoWxEp@ECY3C>A|&zI;S1C9)a3xhyiD1sQ?t667hcT<|LpF+G(a1*`~&
zn4U&(8$|3&1VbMXu^B0byih>Eg`Q9hmvzKe7Q?5IMV6071B(s3=)uMC3MHZ#P7=}%
z9yCy&6U}`vNepgS&X$|NOmr_hid>#i;|)QO7g@w5#*xH*7V&_&e$FC#nd^F#IF5cf
zjW~}I7ntiiDAC4PlYg+t7o#bFGJIf>Bh2-7l*~tAya2z8lJ}VFXDIocxgJExA?CWv
zfb2Ga{-4jZKaFwii=%txM>bhG+AD}got{irEK8G`<w?KjSM&>lB>n$E+te?wDWk6j
z)E^q@QTP3CYS~FCsmVxcMq2tpfQG2U4A^5h-*9;tZYVG|3No4DYj4625?Yv=Sz5Wd
zd#a{+c~5sSmw8BC9mf6Lu1GvqLfDSCn=p}dbdrBNNuiu-GsW6@vY)vqAUF;6kM{{=
Hg(Ck0`Ro>p

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp16_fp8.cpp.44C0E7F4DAEEA470.idx b/.cache/clangd/index/gemm_xdl_fp16_fp8.cpp.44C0E7F4DAEEA470.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f911a3a75d9f584a370f682a0981a136d025e8c0
GIT binary patch
literal 2970
zcmZux2~<;88h&}nJVIVV5<(IrfRBbq!V)U7$P#ct2%usJAj%dAiwal>R&hbx5VZ!`
zibE9?5l1W)$5wF}&>S~JTmf+q6#*TTVyPAdZO6IEoQvr>%{lqryZ5`_z5oBe@4pct
zK|!TBg3L`0$<0hzy^Mk&2pjx!R%=!qCnLxoR0PRA@I%xi&7+z0u1~8~J^LHH9#07v
z?^*B34=-J1>)~#?#9Dczf`9#Wx9jp7ih;qp6^Zs}A;Qc?B;-~kL$Um(D9fSk-qDBo
zuiw&BJM!O5jO1!$cO1UajkafxwIwv%_H>Ts=O6Iiu`jKtTT;=q_eQqw{_vK(nyXEP
z>sroVj2#^4(zp}B)%o!!-gJ$PXJ@a?FWdW7>s#B=uYEE%i<~kG8h(q@G_Gklb7x4i
z>a2h6^U9M^nX<(6_NF2F$}>fwVMq7eRrH(m2iVqk^<=Lda_fyx9=&q<sKdyKyOgCh
zCC8Cl-N*9Ws(zF<uN^Y0iA*iqlb2=xM7b>fs<PP4adU9|-JY13vby<5PvJGEbgQQE
z-(J4u+9zE<G=teGK4%*GQ?bYDL8h)^etF-)D3(a>oZFdcwd?d;9a(2y@S@0M_gk0c
zLM~}#nv)>m{<G?tSd`to@xH^<2NRsYrrJl5uX?em$KSPf-h9Tmq1|;!>rvPi#9Z*i
z>|PrY*;`^;_14!1_kC@#FO=ojI=G5cX6;fFx-Pt+a86A{@yZ|0MqA|W=-_T+6k5ez
ztl6&!F}EAnWe!JTiv;TWL)yj{5o^+asp$(nXl*7tZ<7*HDwiC7ws6+Y+jnK6NPDIt
zY_}Q4cojA*&$mpNbFw&SY^W~hY3GPcQTuPJj~RWHzGbVr@A@yR=2M4bEQW~nEjkWM
z#$GVeuFY`Vbo|BlHl**DDzCfWP8*DUSm5qbFCS@+xYhb*JTXvstf0Xsbw%T>1;a(-
zC4Mejm4fTTZgo3%NOLx>U4Nr?r1%@RszAwo3xAJ2ZilY9M&I<7mB~lCOZz*@bxULa
zDH^&Wco3IS@4BlZ`deD<Cd_l2PkH&l(vnn9pE-3JA+I@XAmCrg?l!5Hv3<>(_WB<^
z(kqs{Ta11DS)J(Co0GJ?A)wVOCMuxnxpQ@0aLYa)*H^k?+r=_P_Y~n+<&THA25jIr
z`^P$64voS~v}t~$*N=4Oo|yh}<Knb|^i3zHC#&re`a+ssc`oeHcJQ-$wsdfEFE^b%
zv~JYl%tUNO<bU3ud!ON%C$3j|tSMYHCF#<@`W?QaQM^E^Wgf6{3A@%%Rgj)OY@&$R
z))1+!ulZE>>xJNzzju~5_S5>uz0?;SXMbm#zFySzhgZtR#LlPXz4JPT?xC^|x;*!_
zAHPaYx_Q6x=K0<S4}V|y@MzI|+p9^tZ?EW%ccMN$w$a(zBP}{R<<uuqp^x*doOPK=
zQWQbBsUmX?$^ROTCLoAPhzSJ-06lH^@k&GdD1Z0gildQ$HpNVt27t8d`*jyu-pMO*
zbRnSG7@K7PNc$bXQ(btztR6=dfTm(prU4)wOxsfYTybSDj>Z5Q$8eqjAiX?pLw7A-
zJBXu;08PVaYy&_#i_C4OVhOc48Ubi?%$#KaNathWulyT+_G2J%fadXdQw;#=op&Us
z<>GI3bR-$jQi_yn07z>syrN}s+shb8JfQhJzMTOeT{1g}6xaFRBph7;XcNqYX#hwc
zQ!eyRPwxt2AhCev^0;;efb<167niN+jl<C;fVRXe*#>}g^I3_+Tix~^N0$OxfC(%O
z0O^(p)0M6p7SnJv9MBAm!7u=%TN`vU&i`>^E{;Y4nuBpz27vUFJ_mLCw8}mlRRY=!
zGh-M4(yzO}&dU1-V<wJ91Dc6383ur~CM7keA6)v-H6+?!x`j;plS}9-!j}kyY0*_8
zxkv$bfDQr@y?x+8K$1#ieU>=L9U`Cvx(Z12mNJrjNqB#j*vsuFPk~NDX@7DUI7LaQ
zp=5z+k@(X@TP2CEk~&F!p+n;CU0&64BIq%gHe;D_U;$PAJJYbI#1k;h7xTTKVUnHl
zD?dD*xDTcU-hxGNZ|G$Zv);%?F$WO}xQMQrCZDE+X+nPkiQe2sk{lxMvqUTxPgZ74
zSYW$TjOQY7ICl;=650^90?1z92Sz&;P>3Q5+-pnT+3mW%XXo_lXl^`I7mm)e_P(>n
zib7XWZKy(6eY(oj##9LJ2f7Nk!G*BQbd{ypQVMMj3kl4k?=yHjtQd$&-(PSVe3N1d
zv41iu?ZAO>m;5Cu`Z{~_eW40=AgGSM{fyKiLJBjYtAujlqzS6GJfEgEyL<|zs~9W>
z2lf$N#bhx#(D*QG5S6}fjrQG3BEBQRoljE6GTS{`e<9OVY!2HNx|psq=a}2VFGp8#
zI2>EpkUu$?Mit+xc?l9Ru`#iO3BW3WMD#a-(ScPI;@D&v=qg9K<D@rW7@(y3`vHnz
zN+5Xsy#S{{NofQdYOcIXjcU|fIt1e2iFtFOIbfI|4*jKJq}@y+c26cwS4rfONyEWr
z00-)?5+f0Zu!0A|(t<_$3&m)WH6etHem#}1q&}Q308t5jgu$>0#)qGjNMESFFabFX
zk%*T@aS#(E;)C&c7fS3#;eKF!CrWe~kDsH&3*+$^O1v{350HsLGSq8QGM4Eh&C#fH
zT=g%8k^Tii6950uW76a+S;an)HppQ0-zFy}Eko3q$r)*Y3{9C&ru+>vp_<}!28(GX
zX9v!i#-_>5`5X(&DKou&{bc?Dv!q-nAEBps>R-+B1lBf$-52&E2Z^Jz>zC7KczD^m
WTe-N+<?@0;)abm$vx7-t$o~MQG1nac

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp16_v2.cpp.BBCC41C35D9D878E.idx b/.cache/clangd/index/gemm_xdl_fp16_v2.cpp.BBCC41C35D9D878E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7dbed18bddf92647e01b7060135daf0155ce9ee4
GIT binary patch
literal 2192
zcmZvdc~nzZ0>&SY@UpxFvJntj3K&2ZBOr)uqAVsr%8n6v2n`4_1R0hN42#9_NK}?e
zS)@fMs8~>oIzo)LbPz|dP@p3k5vyQj8JQ}yARsEe$(%d$Pv1HD&OP7n-FNT1-}g?!
zMF9bF3WE41igIO1*$E^BL0I6Im7RIuxHf_)0n4qw6g8GP=0?B2mL2~{{9x>5saaZY
zy4M#c&Wo)pjE*1KP`)v||Ab(5QO99hdi^(PodR~T?<ZupFRQc?kxu)LN%nQl;#u3J
z#uCrJeX;U%<yn#YP2&0ACsPV0mKScteJyacSM4?v82>zr_x5uytoA;?{@tE|qJpM>
zR=he<@a~s{x)9ZcyK>{a_c<MdD#g9t$sWbcg1_!+P<+z<TORqMM7BP$W4bcGBk9sl
zP5sJ=H{;dGw_GnRelvXP;PASUi}JvrmeS#{XCBX<nzcqvT=%GV_KZ!MyWP=Z{Y&~V
zDW>UU8|epp#Jg^APwkVv*(ITN5qznc@v)`I7}Xs2Y$<I#toE8Kk9M_+>a#kWn<Sp*
znRyqyCmnm}ICD$eAu#<g+bfY_vJ!Ne&*3khrZr>7_!T!oEU)C)<bEJ&9)3LU^XD$I
z?5vf~Y?`zu%l$E@_e-O*6th@vxBH>PU%9ungw?t(4=?O$*`wPojJ86zJ}@>~@0Fc5
zB#*|LCdwDwJl6b^6Z%g-3#abXse2sdc=tz6YSq$6u8n>wwN-WePyWRQZq<fqZr2{J
zGNJ+;L`TLA6tal*K2dkVeLeQ|pYv@0wxrUoFw<>ern5>UvN$I{>Y-)s?wb>sA~bET
z;hL1WkEkA;e80KtXuLUb|4i**;eCa|uWY~cKZE}$_xP-BHl=a%P|y*J8os-{VRtUt
zY~6gIUU`z}4f(0yie>!0)Xz^1Dt)_(WkLntrM2B|vDCg|&2P)$fW9_qO&7^1<}bG^
zo*!V<w;3Ott+Ft(om%#u<yUn2ZhDc;J||SnwMwhB7)8oBhnYIR7vI%z8`IZuA021x
zuznJnaivMQAtuJfHu&hehxA9G5k-LkpO4N^s=pUY7B&%U^Dbw`5^P;>n9|7$j{He0
z(UtAa50hW5R!%f<EE9Xe;u7w?owhfaxA*$>+mXD{@;fW9fA+a1yCW=^x|lLv(S1L@
zkCr646k~q%-weICj#L*SCD>{Ewb6wdvo!i>q&zwI<G9mX2j7(amYh+9O<T{EAIuP`
zx*PkZ+J)h@PZPZ_kG{<pOWWpG<O_{^|CO`6bM@exV}gzE_LaVSI(1H|*dH3vUleg?
z>*TUnNo*mk^W%LEwJc4&Jd_%oCV4Y0sU73pTthBA>f4ZB?b9+-UsT8oTmG;qGA`9b
z9Q`08>1yLLixrZUC;R9}Ed=3JiabfDybt4OB!Un)hU;kre6yNMu#|VR>$R{BM?VCZ
zi*cD6f!MQcsG;P>XRSCI1~3Jq=xGFE|FU1_Q@?&I!O=YcGcX2KBM|SR6@NMzcDn*c
zB>>Yf8c!n-r^M!~&KQ(E$I(!LsTh@|5r{Li^#>?e+!-7V2Us7|XKDmu`OwA=p-pu+
z9Z3M#iR47l2*f9M1Zc&My#EMC#Q>8rvYtjDt_(ll{(43s#nBjmbunG4Mj-CAC_nFd
zG+2j$>;>46XK1bwh`Yk6se=67G#m{9n29kN8iBa)Vz={+cXxepGzwrI#$##(;wR(Q
z(gCZwaU6{Rn1!(z8iDwwDkr1xJi`q~qXA}PY=%Z4MlzH3XGNg@PgaZehlv7P9VeUy
zlLWRpMrgwXfvt`Y+Auj_t7C(<m|AYLw(@<Q7M*aibMk-@z#zfh>J;JKFhF3d(}VWM
zL4E^v64y}h5W}5@k?^RC3$GvP9-mOE?Px;++9h-T#@mKUjdbV)*@SEfW2F-m6N)8#
zDLO$jp;^LB(+S)Jw}i!~6AUJU2j3HJ8SF`28Mr6-SYWHG0&Un5V5=(v?S|q>E2(bJ
zRc$)KWHLFh4RnIVWN~1+bb`%fb6}X4NsF=pm%dkE3YWt*g`XXr(C6rz!h@m{JPywk
zwg;X7SX^B|xHw!5*y`dz8_oi@x^U2jy8*VkXwZh+0=Bwf&<5Q~178#PKhKnA3Dkpq
zYWeRK2L}ie`@f+!5&MG?cKU^Lk9O?7I7BK-OiKnZFezA@q=S(uRGiLW>gftuemkvL
zG$Gr7!!=yzw#~!S%X_=eCVe}1OII7SKXQ5NjZHo<x7fh9-e_wt*zD-+vek5pk%N=3
RJ}*ENkNWT4u}dom`5ivY<of^s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp64.cpp.326BD4F598F31DB0.idx b/.cache/clangd/index/gemm_xdl_fp64.cpp.326BD4F598F31DB0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..41318b8d75351e3168da0504848fa1a48201c945
GIT binary patch
literal 2468
zcmZux3pA8zAAjdHO~%YHW5#{WOOr8f$vwBMOOsiH++s77kjt9bVHXLzRN5_Bl8t;4
zyRo_w3B?gBWXpuu?KU>IO>W6nU-|l;nR6cOIXmafZ{Gj&`(K~`?;T$rms?6i5Z4G^
ziZC)M438iP4SW-mL@_yX2vRGLAStDn0v?GTZ6ObROxp8@*6^d*<Jdu=<C$anXS|qJ
zmPFnsem@nd-FbS&G`i=@nb(m6t?qb598>a*FX0E=NJW2<hs4YOdEH3n>>_#MTHE5U
z%Mk}!x~uNCPxUVRbvyL(6>H;Rn@oW%x5Ma8<o83%exk?s)du2sM=#8r-PW3MFlgfB
zQmli{_Oi^7ON&c0V=?dMQ-1WG?OwEyOvgP$jduR7`T2O$Sm;bp`0oDk#|i24y%u59
zT?bmlr{l#MgK_jzMSP8Om7U3*ZBD)ya-Yv8+?!80*={jhQdZx>rAHs;)}WTp%Kz?3
zy4U>r={fsDpB0;HyqCD&oV?v!Ql%5=kK{F%H`HF$^(aQ_m2v|g*6&{!?U?lb;iz`@
zWrzOqF*Zl+{IEaY2_J0kla=Y??5|TR@kx%^Xsx(T3Gcxuv1fAbC70(8;-gD12~#|C
z(*9KU&t;5eAKMT#=}isTJ~a5^4+8dLw%hXVD5d9to20A6Ralv0J*Oe=)fs%8&y1SP
zoHBlS!%tx7T&;7m#hh(x=hE=1OjGHwi;3>it}icf<`t@xvne?Q`?hnv)q3AiKS-{7
z#>!Pc2#SrK{^CaAWW$>s2|s%)<}ja}FgM`+#1JRw>$pW#dl9XATn7s(vYn6px7Y3g
z#&wlfeGuSy;IGFv>F<lH#c{dmX7jiI!y9U-_dY$j?lxO>o!!y0A5+i@{fd~<rhMTY
z?<Y4$Xjk`Jdh(|aI=3DUac|VK`sISr(sARwk6g<vZg)+1(Pg~muH??Vu^3vZp>}*t
z;Ra2^iMLy)_42N}ZmLOA{@J~8zA3U0N6DxSHJK`Qd(m3NIJ;jXu3jZL%j;%X$FD^b
z1<^497dFpE?Xie<+4jyrBW7PE%iqvupSUGydAlpOv7<0<^UI~Q*JgU-UA$T+i;<|g
z_XVS6bXLT&bLifXk8{R>3&xJGZVjf2j~{UqTWYYzi?3B2$-8wX^sKNj>{RTbZgwK`
zt)Fj9UH@szjOe-^6Q>dNy6>o0=3n2mi!c3JE9#2mCHbD0MElvF1+!UBX1|$g_+~y6
z&Hf&D!d`Elq-*qCuIgP+f!Q4Q11&~0;2-jmGgaQE{(gvgexdPvpwD84Usq$zhx47A
zO0J*N?yqbcNJ~t1vSwd5UQWJvSC~E1_RT5Jh*un<Uy@g2#?Y$~Bh8wN6TK17aA+TO
z7m}DN+*^l35S4v;$|Bs>E)u#6K?E!X7A6Dm#>zgb1w04*>8)8L)F0490#Qi@kT!jj
z+S2;Bdoc<10yLFCrOE)(4mJ17a+**6LPC82El-eFk^!Wn1!=>TYPn-1bO)dn2nsYA
zKssKF$7!<UeL+D!2Q)*Kp(_JOC&;Pv$rC~=NvJQNX#^To29Q2_-=M*rbE=(;>;be1
z-b7vokj{7J;sk^L+e<<{0ll8EUP%U!zIDUEz&5=14G9ebv>HK8T?UYD^(DrdrUfdH
zP;Wq!31o^4Al-ej-J*H9$CZQz09u)#OqBtopGx$@``E=266yzN3V}kA0i<V#ze-52
zqHH0dI{~doP^8EJ(xS+HiQj{NQn)#s!oO}zPT?P$A`38G%n*)67U=Qy0$>KXMGzug
zAuI$!0<hhyA^Lp%oiGG$7KBLGx+X*w!&XB!@;9zj@g$z$e0<kij3Wz7j7;p|egxu<
z{jHa|PrxySN}<EMf!Y8Cq{3KJz!<YxEx_ip17S~4FQ~#lbOOdm$ByD)43rH>EY-}K
z#Cx$g=tLH9_?(qtzxHtJGYYca0a|3ZF#KViZ_{raYL|>QNso6jLf|nR4D3a!I#>V-
z4%C&Z4IYE;NEl<aE{o4vxv?{RqUcCRXd)<}W~;`9PG#SYv%?iT>QJ(POeNEy7Rdrd
zsv;eV=PG_l*k{!}Cr1`g=~O0MF<C&P)0prM$O2`$G83v5<_C>S1rIxf`N6eHg$|EF
z<F?qRFb3uT(@EvN#wi|4g-&FF0pDQdR^g^WUg;~aCd38f!4NnOaFV_VYn(hVU+A<@
zm+2=zVx|S|iM5sG0rl{9`f2{Ai9H(#pw%uMc4tKbz#hh7BWvwR95%Jqu0yeFD9jK1
zJ5a21tv!HZgDbWuGA>NoS)wRB(G>j33C=HCrLO0u1<M78h9DuZYXk{ieGFZF3>QYk
zMFI1(NWL6?9bvsZkwm6Yl@#4+ZZ2$^g1fRBouRI=#n#@zajVm(n^cVKSk@fu-%eGv
pKGD|E-LO$l-@wq=bhDX-l?~HU)5P3WMU~6jjcyBc-j4G?{tu9JPr?8I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp8.cpp.67C1EEC83062DA6B.idx b/.cache/clangd/index/gemm_xdl_fp8.cpp.67C1EEC83062DA6B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..005908ca90f232d6de4a5541b3b03d1b3ffd0be7
GIT binary patch
literal 3028
zcmZuy3p7+~8{V^L{%yvd3o~ZM7`Kg)5jj%smqP9mx#rd|Lg9!ysmPMQbU8vwNjXWn
zsHk&tk`ARDw{(=0Qi{->zfu&Z<9sv!`dV|=I(w~o*52RqKHq-dcR%kp0iJGdT{wcc
zNIf%><JL(i2!g1BU;4VV#A-5vOi>Xe<M6qlM`@3&>Gx*Wt?VyLvK$aNeeU06$@klv
zqP=hdElhLy@e+R9c%NB9hyUA;Rf*9CXdc2^i-^hXNRofTO;PHc&Y^|}S>qG*6UEsd
z9u6dATHI`#XuelEJl#@m)@|8!fS-NXre<EZ=1DPcgI~v5t0VbWvnqbS)E3s1S^v|E
z%F!5WUH9_rNRx@tTQ3v0%%+wVq;yQ^55AH1Ab4=Ua@ecOBT>m=$<en*AE!1>by`Ty
z^lxhFIr#l;L2nXwe~FKvqOyh6l5HPwX8&-?)42Mp?SrLjeEs?(D`*#;^@B)<YM0GO
zZQ~<1eSc@n<#d;=;L4!KZ!fMl>^i9%7l;%z9Id->UeCJ}xyZ^7xxXOg{j0k#)_Alq
z7sqR#@=xRJOCZ*~Y`1DI@!2;R9jGGmFv;jm*4SC+(oAk<Z+pH+Y1f2lf<BkDI)1J&
z>h91UYb=OZ@_FzW_5K%q*GrX;0$)GIbgC!Ydv6YDb=2*=mcMYO(~af(gn6e^H}G+h
z_K^uYTRePR?SPkxar;P$R<WjOh1W*EZF$ZWC55Zc)d#C(Y`@FRWz1-VU#%$f_hb_p
z!@sUKCZDRi5O}*aBwc)^ZfaR-RRJ?kd!S^#*z>sBom6*S*Z7y^Y~#Ior$kGXa<NZU
z>yJA{L^_`@bo(?|mHzbJnENreU(6?ezlLux9H^<tYOyy{3BF)-d-2E2e0SYVdWZT;
zg%1P&482vXMLM%$dE0_N_#?sx+ZULg_84mnxYhpQbF{1escmO%rHQqReBUrX=h&NW
zUB0I6&HSnz+vlZcY}nM1J65=R{t;JkkD8;^kMobVnw9JpI23z~_3eFjcVDwt_yyu+
zyD&H+>6F>dlHmQyH@0AwxwiZE{j|4em8Gq7RhkR0(R(<1eMf?A*CEEd&0n>%d^|5p
z0^Zire|Xr|B&{EAp@wA_)?TRIVtX^p-n$@k&ibg{J_r9kf&NhGFUPhzZRR&RhKp}_
z1>r^cC+y#yH0Vh<VeXk39bYMIx@0cZC!&VdUV3f0=TUxfNNP_`v3|x)E9;Z~@7#}l
zp)IZ+w`+AM^VRlV?VBIDx?uzT;Qh)=t=47Ff@yy`+wfYImS&ZRV*B@8&#(JbX>{7D
zq`I}Yd$#AJgYKuS*B!<6TQmN;^`O?caC_UKD9^fCU%}JF>9P+E2^uylj=Ep{wcvTm
z>@J_A(CNgm1CKPybZFsyMIE_IeeT$dHN2q=2`{r2RzB%6uJM_4`r2pQlfE%I#ur5p
z?kW*GjpW#hqtOT=F~AIjasV?ezj%8+?I{Nxx8i6ZplKM5B?n0B{81e%Ic{;3fkXnD
z$K&b90n%nu8?Q7?dX(bmQb4O>sw#4Tw0(7Vd0s>DX&m(jG!>(=<N#@R<($G%|CRzA
z4Fxog;XFA&Iw4|nUnM_(1V@(vS{YMTl>?+x$=og~7FCI(0f1&>Y!x{`I`?|!duip)
zH4G#Q(0m@BkOQQ5bc^dejP^IvkvKq$DPpP|AYHWBjTF)QR}7B&0$K@EV#xv0WpDNL
zm}yQF1`+{iE{{vd0n+8&RWsPLMtd9$1GGA(&XxnD8|%elo7m1N91RDwCZ?$_2S_&s
z&{msm4pGKYKR`1u215>zZa>>>(eSmy1xJGb&A~V-a)5MClX!5rrt}<+t^l+E6R_m~
z=_iA8V!I4W2XS;cpqUtxAqPm0_pMLOI?1rc(O^KcFcw1&kVewtR;AB?KLzM2lJdX0
ziLCq&7ttkzJ+TCqO_zv_MCNb@=rC0IAI<`j?G&Pb0!7eiU}u@lU_u})o>&Wmz|M<A
zUT|mVJ|K}<PeEc%*eXc0M21iT9SM?=*-=3fNJyc?m$t`f_Dg3ei7pY(72824MclcO
z(p2L%0G635Ob%@H%4a*`v8TFEz%t*6Zv|~Bots@cJ@BOmENj|mE`xhR&x4$0hF3_y
zod|$IpiB#qBg`4j0g%XsKtZyO*rFiu5d}lZ58kd_b3-kcftPcgxq;BlIa&4pZ5|xD
zEL#rdM#5#7GH8lyj1*o_L?I4;`$Bo|mT1~^|EK}fG%cDwJXm0RfKy~s2G#k)hy6^m
z!{R_R--d4o-vQ?i6eAlt7!5wNAz`La3|l0G#lSQ{da^-OkQ9=LG7=P(=ZqT0GH?8h
z(j^QP1_urqUBXgfaiHzt;DWtmQ>(C7IfbbHR_b}mr{u1M?f)awC8``%ZRkz9gw0`V
z!|y?ta5x-oIIS1w(1HqYRlK6mB}!UK1gw)Tp=wbH=v2By-ALU8CP9}7j07fdo?#uJ
zkFpa3rUvT(8Ou%zxC|PG6KtqynOqfAn|AFeXg1G?=K@C#<_4NAJ4O_=IV58Dx6T2(
zg^FTeyJ0@yg|fp%L3EZv)PHLcv_A-w9Wn}GYJ?Ex1=|jSWJirckQQM8gWms~y`1{M
zTnLgEQADaRkI!$O6wQ0SQcnq#`iexnR!EnaCJ{4=&%01!HwyOy@%K>TzT)$9lz5@|
zJc$xhiq9itVw4Q^;0#QX$Wl#Di%mC^-NeY?|JO=*gCMcrZpjk<`IjT!ZmGrOm9a5M
zY_c>d9*|yfK4i*&FeNGtr!!PoOb=C8XG2wG4>q5prY^9yv9ot@bXqizJJ(j<(n#mu
rR(V2AEnPy-Kr}~eY-0AExy3>&?FAa9^If<+H%}?LBxJEW$s73__j>kn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_fp8_bf8.cpp.AEAA96B6834BEA56.idx b/.cache/clangd/index/gemm_xdl_fp8_bf8.cpp.AEAA96B6834BEA56.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c625ea67d2b22a08e0f66cd7130ceb63ff185031
GIT binary patch
literal 3068
zcmZux2~bnX7Ja;gKzIou2}#(3JV`<VK|t9-T-c&00<wuLh7b@0MTm}|4uX^#8ANae
z{EQ2J1wjRukMKn;l~D(V!BOP9G791vL{Sz+e<xFonW{-uPQC8acj?>b-mrNA0mB%A
z_$JL;mmyb3i3ozw!9QD(m0C|g5Ir)2tUFL2@gVD&E9LIDH<!;=I(MvXqqVAXD~fDx
zx|Wz#6pD9<!uqf8n;0+HYrzkWuTT~F^&0i!hE{UKbd$Y7#s#9v&m)%e`P9a$D~k6~
zJ}u_OtCGjo`80WE$6jg}`1K6+L;v+vDLYR5I^&lI#zig1`kE?#Yn#>D_4Jo#;sl?E
zogccgoGf-)cV2SXJN7<#*!fakQ%B%``o{D}eaf9L7wkDdH!^qHg{Y;G54g<-%umwd
za|G!TOFN^?K2SY3SJWTAT_&<t>>gE)b_S-M%X-Ldri^e*y|e3I&TFxA&W~JPZ9Dq|
zEA8h619d@z_RYS+Jxgq4Q!JDtvC&aO#??O=^oDZIy0RmaM9fW}4h`L^XtQtKFqTVI
z_Bg$0C%DXCS-`MQqH@P<d&-oIFTuK@1=k7!AN{*D+RND6NF2IcweFz%lLFT;`oXp5
z8~@~URZ3Y1E3A+<ZTqr+&6=a|YF>Esw~Tj((n=Oww0K+^TdDsrI{SF~@b7(Zjp`|L
zuXA5^dE^~FbGkoxbq(=Q!2(sc)6FhM?e3BAbwcAba!vR475@1t?r}~%na3MGnm3#7
z;n54CRxUj@eN&8ikZ#5Xmwk;(*VA;P#``i~bY+NkmHEx5Q-46`Dz>OZho<IiiwRRW
zE)PkF2^_zAmruk}!%N~lN8%p&FkC8%8@vc!ss+t)zMl(o;;rfuLl16#F=n=C_OPGx
zxfyPaHKR2ZtJV_h(iU?6aR1?k+i7ZQOjzmt^+zj8`-gts77&6hS<`+yKEI-yJ7k>`
zLp;{K<DMKuk9$6<8ZHZbfOF)Y%CKtDhLtAr*~JNB&Wv0^>HSmu>&HZ!_nc4k{axu+
z*VGe|z_72kWo?<Dq?6VNMYH(U9qEd~`TbkOx%=9_y(;Tc>~}Gb{WLyve1BQi)#{!p
zrf<%4ZoHY<siZf4wMvjhObm!qhQu@9U%ZnOnlNzUX|v5MPZ`fAJ2L#;`oY8ZQtaf>
zHEDQb*KWP9y9>=aoCW2BuhSJ*X|7w+dDk#+Rpg_#)h(;vqrsk;S?H=U(txw!S3lRh
zz_Qve_=l<vK?(m{O`A%;b`E8K_mO*T?v+m?$zK9;*ll{jGxzaj_jg<`KKJp6uqJA6
zZQGp(d4KSHtU6}D>U3MJJ2bJY;oFG-Gv10^e(JlT=|3JHbh!2QaB6jGzU9O#pURer
zpy>_Lmo~|9D_5ldC{B+H6qLj``460ZebTOQ6>nB|@woGers@B?G&t-SjAZ9zBrZe|
zgq6%^Wa;>{VQ2z^NO>flxduRAYkusqdEUhRJ_Q&W4rnroOw$0Q9Y%6mT0cq3F*F#^
zbQ0Z21CaKv?X23`ysHL7LjkQ%(x+(v(t!r~JNrYgm0)NzpfM80)&Qh~4sOVvk{jnw
zMdAU?VRLW|Kze!X`g=z>#eEnW0%!x00bK)-&Lptf^-1wZFf<I%3=+dg1CZX<DLN++
z?zuui<bZZ0I_hfx(xr0(bYky(O~lZJfYu}F(KG<*s?MXINl!0&W9Sk<n~+Qx8h~_l
zu%&X7lY1Q%i3K!^&B8SR>5EMwk%#Qo2!<{Nv<1n+L<5j+4I`&HtdBCl&_#fzlBiS-
zK>9}Al^M+wH+?ZQ0?<qn(?|o5zTGP7eRi_E9z$aQZAvm_XaLfWdu?Rxg7RJrT?}YL
zk|9+CkbWYGyj=C)Yf35-2WU2%Jw*eMR^3~Zx$zj)9YZ4lO(W5$8h|vCB~Q+_(*3R~
z9fN;V6T#qbDxyenZ`>bNg(4M51ffuuP+`DCEiV`en55vVzncgp!Z2t8RR>JelG8GA
z!acv6@Fo1oC{T&0!QT`JqJ-l~&;;rc98!x>%Wo+WKR|@X7Ls=ezm450*P%!q?Hs+J
z&|<r;u53LS@Ca-h8W}R-J1l#;Ers;N>M_{n2sv&rMv`4#`S_!;+hE(m!y*LwhMEUg
zQwv@zg+M$E9->HXBsPm-HDDhA6Lk}4nJDldznh38qR9fT4f1QZi*^kG!C7-z;ZV)_
z8=DHR^!|B4y&cJlgWFr*`CE8&{$vKW$y~BEoEY%lKmzLSf^mzV9av<aw?qz3b38a+
z@V>C$K$hxegQvj*2yln*S@I;j$v3OY?-$4nuQs9-Db<L|gk3|C(u`<KC{);}z^}SX
zwft5Q@!H8Zqe$%}cHyvO@Mhql)xE1_v{MJK)Pd=oCVtFlceyc0ph)RVx+PR9Map0@
zEMeIxQYMpW348Xe4LM?G$Dx-XXFaYS4sT77ng~tAFjwdwR8Bn`V0d^zkd=BSz->?>
z3}--_YoC%MPGwy_2+GSAvVCDY!}y>K>VctUo3DcxP2PwiwUyXT$^cFhV5A-+T1It5
zyh*EeVu^UNc2L(KOZ9NkGBU=^VU{p0a7aC3v<__{;w6)w`}^g~8zPn8K(+L_`cq+F
zQ>3OsQ+t>ld}-iIJ&&|}ad93DIMkT8Sii@~9F&FU$qR%JJ^K1_>9ps|rs{zrzSO~A
zX%zw=*TFw)ueYOk5eof*^LJ6aTYLQ+#Rs(4pHO^6d)-IC`w7rKOP(%OXPTWQ%XUzI
zB&y-B-!k}tAhO9%g7nRXxBac%Te}F$WQm9@BPl%vkn`mW2t-|yo<13)P>pDY61v}9
z0o_2t;4qC%Ox-=ayl48%nmvtW=V|RKobpe*Y;y~)6+V^6w-MQj9UPrzxVTw5o7qqI
SWw8V1C87RNa{_gOkpBaJ&->>9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_int8.cpp.636EBC0B4897BF3A.idx b/.cache/clangd/index/gemm_xdl_int8.cpp.636EBC0B4897BF3A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..98767742f762a93909a1a9941f3278fc4b466f2f
GIT binary patch
literal 2654
zcmZuz3pA8zAAjc^&5U_xn2T{6FUDQ87Mt86${@LiY7H7DAr*$mB3iVcuZY-nNf#R`
z=^{d#t?aQC-wxgMRl?`Cg{UtpB;PahooCPNIs2Y7zw<uN@A>_o|MP!-|MyIQC!Zfo
zMUahAo~iL7@lFK<L73o|Bu<PeCL%~T1wm4)E(N_v9I>H2S(&<hz4EJnrF1EGyggV^
zU|f09F*bsA&9>H~^hd3m;o-f-Bh5_~-t%sH`zf;s`GnOW+HFp5LOpWJJbzL+-&Uxx
z#W-uMDZlUI-m=1p!<E>t)8*^K!}k1OaXItm{+FTQw!a0MYgd<y?6l`*t=uh`8Qit-
z{&hliVOhcOAw2DK^3|U257bO7j6BM4YLAvk4qYzs#LPFnUd<A1$aUH0HXU|<Vt3!v
zVI|l03n_$<y!Z4pFMfk&{E6P+3Vt8?KhD}QA0E^Up0aWb6V2vy&6XSpiPFD(Z?)ji
z9Zq@CmYI{jGiw$y%$o(qk!tG;<^ycQcTyh)yHld5VPBKYTCN|xy>yQ>y=SuQlv74W
z?xrb*csRR4<$1k`Y}f7nnKAXuVAXJt`}Ul(DEHAh|E=l8>KefY^n?G@QnT+}`%<C)
z*2U0oyxbke)sx+{m}mDziJi1~jmIvY|8S1Z3m!NgQ#RsI9QOj*5j|fThh9(fJ<?Zp
z+EsF}s%oX^z!kU4{xSFKqu9&so{8dbSL5h9ZK^ll<xMlc#%>m!VAT{XRoMSL_Mp}9
zJI1G>l;l|A@6E|Cd!L5X1k@!{{7&1r=htn>>@L4945*0tsL$Z;Yx7yf>F#`7eD<(Q
zLEgi(r#gEZdG;P5b3<cEgLBi-XKhj<2QKD@e6R`Xf9ND6U^zj7Dx+r<TWj@u6^tsn
zG%m$DWKNC$`<O>ie5!EAgxN5u#d4tR@sEv-kF8tK<$<D7M^_)6YhuIM-bG@juB71i
zx#y9BT9u<?LZz9Kus0FuuG?!%Y~9*(E!Nb~QfCXsgk2JYfcfUpkK*nKQPs;Y#iMky
z6WnUAT9NqN$dyys-KUQkPE15Kl*c*NY+gFEJIA+wWqil%^p!Zv*0*!%pDaaVd7f9D
zt)7W=mWun7%9-DN)(sB~TAnahd#AVl{Yfjk`NbCFNqcs~PqRM3v*|Mz`qj+h-Ok6E
zHub6KI~kYqtv7m({A$k~7=E?ZlIyuZD)hbhs4L+{!h4dB&(J94)kxm>H)?l}$7KFi
zw6s*6BoUSDGxV)Hb7M5{d8}u6>g3(LBaVLCt!f8)oSB7*noUjks#Blp7%Ml<zjm^H
zVb)gpv&l$$d8WT`@L5YI-a9h*#69UimOXEs-tv(uiNJeW;?gk5uRm<y`d_<&MT17#
zrOf==cda5$Y~?cc%Q~)0Qd+#QkANVYU52bgf>RfUevKdkU6L+W2B3Fkk7}`}qrwkP
zIT*SX(268QWf?%)Vm|F=#~&W$80rIPCW*<A0i+!-NUHPOOU`4cKcFcjin0tK?KXo+
zY+S;`bYushIjS6<3?S`J%`O`A|D_N^w*eX>VX88K^zP8~{u;HTqZsN3Xex=ylmVpo
z6FEH;QdkX!1^}8xVliX@=_3-;W)G9nJ2Ye`pbZraDKdccaTh)z^zmmQhI#{<Od>1G
z0MfTwO-=10AIxLuc0j9>)YWAG>5c%!Sc~)!Du((3nogqAWdP}W4R@^Dm+x=H&>%pw
zNo<A;ApL5{II_p6d<a7W0j)$*qRRl%Z~K$?XVlScF*F#^$|PmF3?PjpigqPwf)6k;
zG2OwgbKIFg6Hqm%y0BdE!U@#By=Wr!FBeP`;EuQ(JT6VZGvQgn2=MZOM0)cu0U(LS
z6XYZtc$=XF8Uje9Zjh5$;`VY9oM!|j&?G4JFXjNd6cO-B0u({BfMrsr$VJ|YN5Mtj
zipNY=79Ny?k{Fs9IzSJE_I2*<XygxoX(fgd8`eL{EUSERVBsN{R<~34gF!^zF1NC;
z)(RD90<s2K2Rel&plDEZ-~nj@MGZwA*gZ4>rh)0e6yZ&PB1l~;m#RA+0CxeUw&pp(
zB0w_(iPXt*5;2}BC)vUahLRMoO+9AYZ2Um>99Pa(SjOy()^G0&4Yx_BgE;~?4NC!B
zA?*dY0X!VoSK0$`8Wi6Mw~*VH#pA+#tFy<9B9%I?p)>)VL1)8Opb3;2%53P}D++V*
zJ!|jH5NQG?o2d<*L=&*sENysLnt;t_Yr|%NyMYX){RUHk@xi@I`wUKl3^AMqH82J^
zhP2z{7G)Fgf<HY3YQBvpga?5(0Fk9VC?`_ExiB(35LhJbM7c#cZU7g(yOtG5d1lE4
zHRmbtm~exSldq0{F&?2u2FboA;BVyez!wSllKl7|C|-cVcwqe#6n`o|9!K#B`SBkp
zJ}*BWCE{a5s81B_0Utf^?<_Gg$wK-PCWZe$q3{JE3SP8X^=g*HlZm0>VMzF&1QHn^
zwI>=3H;c9q6;_eR6h(|iXDBOqFgLjxF{vIbHMWYnhOM20<9a9O4PS7~?DcI-bpCEu
qm8*s8^7IT0jZIgZTYPD0y>^}USDI_AHgZ(?o)M^9h>JVH3;7>DYlr^;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp.4306E603B691D828.idx b/.cache/clangd/index/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp.4306E603B691D828.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ce5f001a6a069f02ee5f210ca012b58b0416663
GIT binary patch
literal 8404
zcmai(2UHZvy2s~hnvr3e0Y!)4NOC}jq>LyT5d;ST3FfRoD-sk?5W}iqj=1C?*;O&H
ziyjb=jDRTWDk{478P>elgu3gB*Ol8d=M`0+^UismV^Kf;Ro&H9^@S=ze0_YnO?W)7
zaNndDcHMHC$Kws<{;ypZziKfDn?yWb(vkML^Vj^#QRjJ)?3ev1#|pyF#cXrEd+5Gl
zew!lXX>f7*k5hBZs-(}#nzrnAK7FyvU0Ob~`LN^1xVHvBT4W5H9*{P+E$s4{OV@hB
zKUQ5AH0TfL22|`WyW(q~Cz|oM{?fu<d-@;j`s3^8h}5`^XFHQlh9>@Ctlz7bej{~F
z&F46Wz0c1(^h7uo6=)wG>iux#$A~RX^+gZ<XNh;4)zf-<vvF$FvJQE=+7(G^ZMxN*
zoQZe$L{(JJaJKMYqTi{fS!zB{YoE>Z@=%MaAeLEkc<Y6uX~HXSs(WJB`oC10t65az
zGUQF1x^tFb_GY-Oese)s+a~qIR};#6JfiE{AG=9Ap1W%J+PEbY4_g%FrKj+HBYUg%
z=ID)y3paoJ`j=4a&}`35k_8^!RoSJ<2Yk{~cW9;Q%sij9B;0=2j_i}cx?ant#HPkF
zo^e;)F4ywy%`K9}cb;~<yVb(~;ZOOHjMxWRts`G8&?sw*h_4Nt`Ql?o_)B)6yK7c<
zMpW7PTG^;Ge<kioGmN-qbEhoq8y)p>-@>jina&7Be12Am<e*pWI+1RmuE6q`X>Xd<
z-jq#_yIvlADE_r?qR%dwKy+hT^voUhLAjCJ1!p#>r#@NzMiNjn%cd<w?Sf~3+~>Ho
zZu!FI>^Wga?5<Y2zFctP<+7kM+X>=Vfq{W19PAH7wtHW={^3(duwTQi6=(j)TR5up
z)cw_Vow~*Q<-W}~Z@wSCpm^QOh=`RHPn@5`Oe%WacUVu;<=Bdf<Z*vo-n~lT8~Iq%
z$?I}UlX*n4O?dcEKM5fJ&YYxYuD11=<<GpuvoZ{~mp@qC@OWQ!v$xo~WZ(Fm+Evqb
zggzPb@g7}NqPW^S>{(mJjh*t)y+N-W7ROdzY@B*Nqx+}H?{p(IY&P%u5^Qj%J41Y9
z-fwH?IFIZY_;bha>t1GP&Di*l(H8&8_!lLiTXpM9*Ze1~Wmnm>=0p5o%?(DfMsNMa
zTdEodZeO&F*FAM;#46jlIy<W$%J!s%#{Q=3Jw^YM<~g6IfByU+?&s^`kaPZG&%?i~
z|NYhdR)S-&PTk{UCT)I=6Mq`xW$HX{&+4{_P9wL_Ng+?ZiAJqIwfbSr{q^nH<J-&Q
zV~*a4dfe*VI7!>ht7E;vwLK}<?8+{wpFX;6&57t^1uyRYCOy(HQ*xuZKH^xPIJU;~
z$-Bh+DVC=??-|rijEbK&?7)iyl@I6s+rjzQBgy{O?|ZLo*;^WW$6#bG_3w|nfAa2n
zY}uPIuYwo+gTV5$+Ue>3@#nUb_wG2A|3uAo)%e*5TTS0gaa;e!Wr5$2`&moFYyS-K
zZDNx)KZeweBZVD}oy`yavDS~AUp=b*<@{f%?Qb6)?^$QE=C?j?eyG9kTJM_^^F-H7
z9TGP!TbLWy==-5_e(TvkwtotY(lw5lRy$_J$u@adxlU5xgv-gtw~uypKlt-oEtiC%
zLq4Cw_PEuh_)K>mFq)mCUSIk2T~NvU#Fg*Ys1^1}UxcJ5>JNNQ|2pKv3C~5mwHsnW
zUs62Yh&}n1A+qC#`P6bAj~zE;oKRJ8en|CCoaXCF=eVZ`sX3f%!MD&>6-c(N*%)Qg
z*5@LmmT|Hl-%n3fAbH&18!j|{QsfG$8Jz6Mchpf8NOnDWy>NG3dbyAa;be2Zxwfi6
zviHFc?;|U|&J$9LIoXJBBv2Jdo~D_6_*wA9R3Wv1lVyCFSXCf-(T%p%jTJukh15b$
zHshNKR0WbFOlRE6J^s%GAvK$m$MDDKsER>((Wc%qy@OAMR1hbR<&PbzDv%uAlzKeh
z$3a7fS|7?|<)h^0sshQeY9qQe_)E%!)J(3o6W>WkRUmoKby<~SY-S6?OXTF)^lS}P
zf#m)B*jFtJ{yHI~f;rih?>bCXAo<5BKCtN4w@@J!z{yg+R9jUbIW6S)&!1kN4--;L
zIN6wQEKn6lE=+!A5hiLoE2P3W*_3Z8P!&jSsFBH>!Y=<Uq!w|q7vD=?RUo-BL=ZV{
z(_Bp<70Ag}d@C(gf#gf4TPD=~)8!?k=5n$--(5#lAo-5`-i*pEwc1*|rJU>`_Apfy
zNPh5WOjx%??js>Ji<7PS)>^6p$!~iTVt+WU<tU`)ak35HMoU#7`SYs>KiWQDW+tS<
zIa$J&3{e$Gj%Qb_9i`9Xfm`*T34vdS1H<}F^7BU<Sj|pYUeq0WK*X>GAT0pd;2N=o
zAT1=h2*ww|M64HsburjsE&<yTn1HzyY)eTl1M4!feLYO92WQ;A0VXz(+z4Kcpul<)
zI5mMg=4NneCb<P%TEGMAt>Dy(b^f)cSFbI9(d091Ihd7`13L+3CrPdXvnrCSVRSXv
z;|HII(dS970kaz1UTdYH!>Ut~W}w+GAo>D2sH0ZZiKemYu?^e`$Z0u)7D#7ClWENB
z?V~A%tp&?ka6#=EwhpZ8zzwlZtx?Q`OvQCZ`30hKFhv5xo&?cJlFxwX49OKBsvx-%
zM3p3;1<_fOt3Xsmay4jIgBgC`dC)#jat(-T$o91$t|hll9q8AA73z;%ubSp}E|+0l
zN4OdwC%~}oy6$?&u#h_zXnyC08CKz<n2KzdVFLmJ=AfO;u$dstgyHC|44VbQERt)%
zwHCaw-Uw!mc$5r#8|1gi<@gB3A7LzNuYc8UW7ml(x47jTfDr>=j09s*veRCT?%z4#
z1?av26C{{urpE5nAC^VY?0)dvk6%r*so<Q7SK72LKH~K@*InEedI6#rFcdvOD(Sd?
zrhCdx?$J9a^BvTn^+^Zj9=1!kx`I!$>0qCZ&6H*{z&?ZIOz_BrDd>qbH!Gf5;DtFG
zre>2v%LV6LJPTtD3w6qF+eN8fBu<BsxMK#0Gf2(^aVCtxdKQSY@TQ>IY|zif(@nFv
zpqGm`k?a`jmb>-bLN%Jr0lgeNi!_@DdU?1Z&E|t%KGxCtQ<~p}WQI+qMkZ6%XgnxH
z)0*FfW^PAqRqO@4rGf&2=AuV3>=iJ)0#fu0M~QUS(T6#4+;XH)(iCbO5}b8f3#xM0
zf1??8i^dia`ZdFD)!0hVHqh7x1Sf+=GG=F^_m`vi%ijxW&NYh)up^;4*DNX|xd=o>
zBo~9InB)==m5^KtqEeE}KvYI@KZyFjXN!Kk^J%spT>9~LRr0)kSb#p$#eeO^`0?u=
zans`E<TW47CCYj@8&;+~$p3LOK3?I38ZoSYfInfw86eCcZT2MCpTuuw*b1<(Ah`j|
z8t|rK*jpgKg?BH*egKybFcpo|&2C@#jK9F28|gF9dWJXA<Ttn9R-7nI<f6qz7<v&5
zP!EP}hoS8xcYsa@=wrPTv^zl$^CcL1iEQ5mI$gLu%Jqh5ewX7htkgg%K^YIjnj4r)
z(EKs1g@J_xc|?XCV=zX7vO0#98OS6kLuA-&$|;+2MT7CF-}yHB^@^e)47&?-c7YfP
zN;chvjaA9=UEqlIJ`nVQHp!AcG6M90Qy)ym`Uenv!0pk1IAc@htHjbTRjk2FpGKLb
zQF8o{Y-&t4<$yTAA#IJ==D=YtP&zs|dLjd6SSJT3Pn3BEbo?58+otc*5SpzAQ9T%<
z-Za|)q6U&1LDWcc6NsAdLeXq9h??<2(QFHdTFCaTAZo?!Ln50_INcm|pW8}3V9*2R
z=*1`>=lqc}87ID7Php6@eQ9}l^{Ei2m0WC(G>#mN`Z8>kaTGz(#?b`D8pje8XB<aR
zym351>xQlyfuET$G=ZSRp@{?~=_L)ibcWraw}GIIdK(7`ZMfl@-yI(~WA;%5p|4`t
z838j0@0kO_93Wnn3&LEI^FWwKay|(2Nv;C>DxA?VY&A@*CO6%An0TJ#8nCY++c$w}
z6X_h9!K4{1@x-)%SqltC2!?G1(^j(m9gyFFL2F{zUXb^ad>7<*Nxld2dnA8^$sfTN
zy)WW_k6L`q?~Yj96ox1k<0(|cDK=vY3<#Wu!%;Q}vq>wi1NS=of>rTr69YdV_u$OE
z9_V_|LT_d`Tc8_AZUnlK<R+k-NanKHW|CWgZXvl9=vK^8Axmv<epvcWjbZnK(Ozr|
z44Xz7rcp%ZnoSvJQx<4o(OT^n);q3s=ZxquHT*C&=tj}G6DZYr6J?Od>N&rnJZuuS
zbpRAuL=+DW2qyMnHyC!~D4ysywcBdmq#*949-bbv&|z);z?4hd63g8<cT)*wmE<~}
z1+%l5lYgk$(emi`T4nz^9*a@`lr#P@vkPu?a<kt9emyumDua^WEs}48-))lb!1OyX
zA5DsKpZWEY`d#q93k$J+5B%<7eP?E=_igJ#OStj8hEcC^P}|+-EpfVjz>pJ+bd1C(
z^q}W}=J%)Z>1sRPqz*W8A#jn#BEt1A)>ur?5{)GUEz?*=&~lCC1chpZ4id`pxrI}n
z^NEGa0mV^b;X(pJ=HvZ-6%4P!;BGmP?38yZ?o*^{n=RT4%h8u8x0!PqMTDblYH~I;
z81_+I<uVcFH3IQ=bA>IoHNOBqA`?jmVLF+Kl!JXanR%Q6`!gihgIPW4JbFOh^L-{_
z`~l}w==Op;6DhAQ@Q}tDjFliyq5LlUG#n8#|6aB@MO$)PwbE9vVK7vp%MR{o%9{@S
zG=)ep6n~(?L@^psNwgn5J$Tn+4P~}zS%BT3GTXE)B)JGC72(`~W{W{yjL*X~TLSha
zFbORN&6a|FDamCZFC*K30gEqqCQ9FpkUjnP=r5ezTm+Mgr1NM8lXj9jU~~uW$grI-
zx|8HfV19}8?_Dsui_{0eegK@%K#%U3m-{TCQNkJg88AD8PxQ(x)2ssX>7)zlK{2I1
z+_XzX5;3}f`C0aS_QLnU*4&Au9}N3Jf*!17yMEkX$<zD6AL+k5bYGvV9~hzX(=PQe
z0|`n$?b1MUBltGLBCI#T<R%Eg+zgYONp68DEien~tuVP2>y?Y{%1aMu57SUy{s;=N
z182B37Zj3Q1cD-xi$PFKatR1ZNG=6IDamCZC?mNa1pSz6d@_?NJYL)t(QKqNa`1>v
zvr*D0f}*9-1jR~Y35t`(5fm?tCup5`9gznnh!Y4(6ekjtBuN@PuG8!W$p(TpN;VRd
zNjYXxgEPw=N-4)Y%rz?>bx+m2th@t|YsvMH0U~4LazW)4Bl<PN-hkmZuw!eq4+&TI
zE`7xvA5G0oZICrHtd-2l4qZhutc}dZ4qXl^BW>_r95hfX?oG<uO8lmDn3@iLXh)#?
z70wSUZ!K{%#c@zn5`K7Qz)T{5UIW8xIDlR*kTyuQGoEuF`vt^bU@-B#HhR;Z6!)3h
zs{JJ|z&=2^zvP7^7lC^bc%or3oKJBsCR4W(a4jJ(_e#OFl;kpSE+gB20oyMyc$Bzf
zbV2=|l}R^e+8LB_1~s@h7&ezO&c%#^H#cp{Yg%GHszK2*I3K@>o4#$FY_d`1qbx3x
zKBgTk+DYyJSqIqQ59<V3C&`z<>Jp5@dKbvLNPPgD2k?jeYv;4x#wY4$xv5S8jTB%|
zM|6kG4P1GFj0c_pQ#0^-zbMs89UguA8rT0O7~I5b_)6n+D&_o(@^EElZZ;krS{T+^
zW^Ioi{wlI!wv(=BJJ)DG=<O$?TPo<KVt$og@u|S-dXa$PuA}I5&_u)J<{#(`k~4wM
zBsmM{ERwT<&L%kr=p2%BfzBm459mCS^MTICjLbz{^Se8A<y~P=Cd|dd%z}_CY)#xj
z8fImaDOWB`&xK&r9)$odp-^7H<JniFQi{VAGJ7tyFQybFB>zGwDoOqirT9wn0X4-z
zHKP9!HAMl*`)I{}n$XYFikd<Gr}>&io@>&P|98E3@oY5rXE5$>@bO`5o4B89tnHg^
zimm6?YHM?UmR}LGdQEKHs))#_=#{)!^mRO59Imdy)fKoJgR84?bsero<7x!1uEf<S
zT#d)oNL*cmt7~yJEGC>ADvvjv^;e_S`9m}WLPkqRo6CNtdRh$CR1DM89j<TS=;Y$+
z=04fOc7&Dl=t*OZ|F7SQM@mMSn3_q=$H*+L$Bnm}Fww!t-q1$wHA3v;n?QNbpE3>n
Gc>e{HU-m@+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_lds_direct_load_fp16.cpp.7BFA821D80385E72.idx b/.cache/clangd/index/gemm_xdl_lds_direct_load_fp16.cpp.7BFA821D80385E72.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0f980b1e260c12f22e77ed2a73b16d3d2b620951
GIT binary patch
literal 2806
zcmZux3s_9+8eX$ztW&e5%QQ39n42!L(oJf*UsaP38li1hLwA|(Marev3_I<u5JK82
zVG}z_=yEzJB`QTbc24Y4QK^m)#rbE>^PfGQXV3G@`>g+ezxVsU?_ck?#6BX?C>23G
z!hBL=!jm^p5Coxve^PSd#`|OhiBLt5)RUK1$HYCeqrE?UV+FbB?8sy#`IDZ{A+xGl
zN|0=aweJ4PJ@;}~<aUJBeL460#yn}&i!?Io-De;+6o{LVg~(cF!!PK&@|)&a4`fy?
zy%7^@U1|i`aivqKHq{P()vmdbS%=iF-n~1u|K;lm&xGNS9bXSKKiGDRAN{%d<HXi{
z3%(#?D$25sU#{Qk`aoyu<>q(R>8aKCy|-MPqK|D{IFxbYT%hk6>bV7j6;XfSQ(UwB
zm^SytYi-SYAFZs_k89c&{%f|@(<L4A({|51Gs_N6d-rU&OqN9}XxIHRF`6&!V|`bG
zFQ_tP<PO&SXK1JI4vtC8eM9M)$cX}cVAC?u)7L7Q%j!41S>KdfNlz7Y7oHmFoaniw
z@m6nErIkawn+G+fvs`wv-=NNNvuW<Gu`@*%-TX2y4fPPwPCf@{Cb^iis1kP`TxX72
zKYwgqd$UNbQ8&d`N<R}9p6oKL_wc8y=O@q8^q$P*oIm>K7Wcf$S=S!MZtB7Yr}>&i
zX71SqQ+7mM?v^W}%YGZ1&V|vpcN;l;%82HnI((|H$o)cvsHl64GQcd#dbT&MroW}i
z(0I;A(MY~O)>e`d*tWc@P%KMPSy^b4zNge}$Af}vFAJ?@FFv?QYx<p7oy#<2dAdAx
zM{ECwR$t>&RNLKWKR%V$I#U+ws$WrI@+R1{y70R3E%RY#?JIKvD!ttKPc!C@WZJ|R
zo<G%oq3q@7X#U7Pf7`u|XIIr~maXS?Tr1X}+**?V8*}h&-o-xc#I%8u6`3cVyd4V~
ztKJuUlz(FWjVQBuS3kRsm257bcX+kK{-9#xZKL|l8JxO+meOfo!^!J&pOqT#E=fqd
z+vXA*a3(N8k{BTBatJ;h-n%d=<nEVlcG9bLt>+$nS$D0zbiD2+(}`EfynVmEJHujC
z#HfSidvEN-aEvWG=AeOof=$cP3va#-_umOFUQz33WIMl)>(q3tbV;9=RrTz9XPYQn
zc8x86GWNMQB;F|R>5ofn#SulE10N3uaD3I~J{06jle50>YLFWCn>e{#ROd9#G&Gsm
z*=HPbPL-MDk4t~p-Zxd}-SXq<0qgkgzNEW7(!t3w&bm!Lg6`y456{?tn${e&;ujZJ
z4@LIpq*BD`{1H>Tf4T1YSQl1opSXEt!sDGeji)<4XQ@wI8aXzUNpek(o9wAOwy4ZA
zt?guMkVlV!LGxKwwD*y=ZOiLitoP7(mrGRztdy?yBScBxuaAzq(f1Dg@u&Hy=)>gn
zzaKVEGY)KQyo_%+l`s$~j6N|#E^rl#kfg0KAsQ%xup;JaC6Zhka5M-(B-WTUM*(2!
z6+i7<A1BH&mt8pO4`>#~(og`TEk10$Rr`MNaU5L^Xb#5FQ~;!%F0>ZNZXP<1qke$a
z!n8CL0BNu6iSg*N=>QyE2WWLnT~z^)_NHbY?DxC#GmZuTnu#&B6#!`;+;kf`s${Pk
z5(sF4wm?q-kd6vW?<&&C?!(cQfM#F}x&j~_PiD2KV8KN=Dh4zc<7z4Z()(JCD;5(+
z>S;(ApzSI4Dhhyfj=P8y*!d*{M|}ZJ$LJagfOM|7^z!?`>!CRMJ)ku(4OIm|x~9t5
z*gmxR1CD+NXameZR{@Z&6{|*Dq_3so=n6pVU^;3Ffb_kK^;S1$nmllHHK2JIPg4Pq
ze%?DTw8iLnFOIGPv@WKrrT|EfcBRB`FIBU{(KUc(V{A1AKpII5k4Q2DAAaZy67_HW
zK&F1<1Db?zBD`Q<G>MQ9TEYy_BS0cI2s{Z$B8dbgiJNc<ltAAAiQF_w5=+8CNx~Bv
zK?(E{h5C($Xc9V`&V{dwCedPRabb3vgvDlYp~IouQ0g}>17-RXVem}oIS?c_oKj&2
zNyG^f4D(g}iTFCze4Ip+*qhipLNf)n-;S*<6ZL>?T}@pEd{fY=pj)|3m6CcBVwe)#
zsg=+Lb_?ALBy#I2Ns@^jN|I&5HBhp7sb`BxfXzy9xL}dMA9k`ek>0uCP&*xL8`v8J
zLv7~vsvqimU)+#KuMq^nXxW$-*S__@Y_P4asjdw#9X1HeA|Dl44!m|yrF=x-Hn>tF
z!UCQI8q*STV3lxkfJ8nve@Pf@1{YowO~Pa|x$uP*WcC|{s^2b0X%Zbx9R{2pnuM*%
zW<a}Np-jfK2=Bcm(<B@QhYvkTlW-YaKD<qugva3V;e5dYK~wS(g{8p!zyk8ogxjDg
zoX~<Am;<ySA5f*BOcIgxf1iNf1B4;4cjyO@Sw6N(A|}CsnPEd9NItwuK|I0`291~R
zT&426B?r7FAw@`s2Yh_<JjeXidi`0T*dY?}TB#1=6N&h&yskirN)+Y;@eL@^sJt#F
z6D4HGgRh2kgS?ca#Ly&1`NMiS{O=bG9}q+lqgaw9+>mac-aD`MQ$m6_tVcrr3JQ%0
zlSYE&lJI3@$}H?3Dyle5O;baCG2L^K5uLhNONYVKW!u?1Iyt+zx|y>~91Ls-!Qbs_
qbGW=YdisWQ=NZp8wOC+jCA8&R&o*1=!O|A_grZ(+-MvXmk^cf3hr3|_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_lds_direct_load_fp32.cpp.39A0D92907A2D21F.idx b/.cache/clangd/index/gemm_xdl_lds_direct_load_fp32.cpp.39A0D92907A2D21F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e5af5fec80e800a22f53bf3142e1491206b61711
GIT binary patch
literal 2726
zcmZux2~<<p7JYf456A!sfiNUQLEwQ1C<IBEEDp?y3{qqWlY&4H8AK2jCxnVsApcvD
zNkK&Y0#+0eD@d_SpMK=y7``$H6$@HGWsvFLr0crkTFqM7>%P0sKJT7;_IX||j*i77
z1lh2|B_(!eQm7n)Aaw9gOiGAqCn89gB7&q8R&I-p>*Z0VO740P%PznBR7?D<<03Mt
zYmf`dJjB)bzBZ#R&m*sAN8{p^=TT>cbp!i}sMFvIFXh>sWGshJSGrnB?ZS^w_HGy2
z7lvY1qQ=R$3%d^3{MK4C@Z|e}h3-SOJpSr{z`_!ZDpL<itIqG!9_@S5Ews~T+p^6{
zDrN^KLeHLIei+qZP0sfm$Ti#T<WZB^dhb9|>)B<55(AOax$KT7de;t<jtR0`?&oxM
zm%hNXk0<SYjha%2jVI5Z@~agaUaaI+mRij;zU&z(cWSiTm{Ql)uld6OUQ*TWD9GJb
zj~sI|*NMHq&wio0JubXFHh*ow$o8M#{4%R064hF-U+z6Fw)UIc(yMUbwD8e?k|So{
zzs#J@$$D3~W!0mc^{m@tI+3#-afRA(E)FT(p1v2ZJb3ysRMoOr&+&ws%Q4Q1xeQK{
zi8?{+&uI1F=I#@G#~?=9@sR+{BIUHKn%BbUgY}uT(Z!C(|DLePD~}~q*7S;R6YA^a
zAMSG7S{7-i)KWd9eZ4!z$KQNqYIcKA#J0QWpZOpobG^tjC%UqGSF&vr|JRYh+KaDt
z^DRtWbmjWa1&KP4`dg>T;ul&zf80<akRC_GjjOg*^YeV&T`j1M@?oPUZ?0a5NImE8
zmBbr+w3XlWx%r9yM)@e8eZCFT_v#m%A3v?Sp7zuD(N~wpti@@oLL6=gvya-|BDjf(
zu_2{bi@L8>4bDY#in4sHPS{+wXrNVvX!qQ@r28qg{QQLJqdt@E?IuGXO<hY+tIECs
zo5`QE1V3^LR%Yxn!E5H$jg%*+c^LjHiIM5-J@dac9fw(sew`Kb?z*4i+@?IdF}K%4
zCokH?`IQF72b-Ie#b{S|neWl${xmqlK4qC$8R-+1ZX)jQnF$Ue6)Odt92o8wJ}~q4
zeQ#|v<%AWC#m1S7ms{w@ukCcZHoC+fei(eosln60YUPl7@s3V!{mIh1yk%{d+vWBh
z9pCbFe6BxikHJ~-Nmnbch%(lxnSc1OHdFO~TXtTUlpXM_MW{DyY-@jANxpTVsoj{z
zKP<9K$uvGcAv|vPVZKkZX5bB<8n|{gFy4x_|Ni`mr~dzSN-pZ9<+4^6Owjb3H?PCb
zJbYK;|6#<ZX>3t-aEX)~(e}dOSGA>0^mNmAX7z1%YQ4+lvHAA2n6bqeqaoKzXYxPX
zjkz-SAnRsShPYji+t*XAOx4*?@iyugr%f;SChEHy&04;7zo5{Z_&!d3U-;bqrMNfY
zi;69`UYn6ai{hU}T1FQv5T|B6bCJZ<*syT|f~ZC4t0WNYTgYf2f(R@y3zihX%u9dF
zJQrKJbM{Bcs1Kl3F;!(LK-z3N^<Kl2b3Pf}0%$EvizWp~+g|G`&a4+*C8M5zrekzv
zDL~qZbR=ik^WhmX>IY~B#$ZYT(z^oFo|UQpFhoYT0$K%Ap-TbMdx&bC3YegbjCui@
zg|TQ-fOJ-u(QRkLe>PE&a6t3qcnVU0bgqLVA@K2H7#ZCRXc|USmI9>nyeg`vM(%`@
zQGY-yVM>Zpfb{)3BO`uz$21w;4rpCWS3?SrZtzl!HcR`CL`FRT&BT~gDL}gIdXq){
zLi+|Xx((25j7^gQq+j*p;hhHg{bbY|(CU~vRSJ+EV<G`0k2@w*BmmGHCPzmKkbeJc
z@1BDdR2~`i1+)gHL6rie6Lv-<E`cvO>^p(<xBVuPzUnhYfZO7mVAT`>*O0pgW`G_5
z5{Uuei-05&kC%~r%XNhk=nEi`m_kOf2Dg@xXmbsq1bPJ}edQ7G79V^Elx$W!ZMZbs
zkx!rqc*Z;%Xo0})=G_fdjxWKsI!&Dc?-%+DoI+wQnNv97UN9x7)q-mePXRp#Bof2P
zNRseFG7@*LFO(#^ZR|AmTe}s!oU@+e1G|@+Kz|%6>ZXHjUB0dWYL6VOJKoeka90xT
z%L#(fuoBQMiId@bp^3q9Bo2n#;8F~5Gnf)opu%Or<3M9$q^}H35zsVgY?zNCplj0E
zuuBwyilz!1-WEl`&}6V-J17EGO;tAhN)#U%HV9X0zKK!<Od68`$ABWxplL9mF>lCy
zitV&)`;$l!Xfd=n&^{Cai^1Z+3#JIz3^oT23M>VbFPSx1C(I8#A(=6_4az6uDo_J+
zfIE^*noQ6U0-jBPVc?ehxMA>?phZAt$^6NPRB;xRz~+D;$rQ>2<q+_KFPUN8AWSli
zGGW@d9?UgyGu>O^`5G4JJ6DcNhv%3XeU)qadx-8gpr$_v_~@6U;Q2EFpObyAMe#c*
z%m?@u6mOM%zC^^!iI7JUb_zo!M@dWwPqdLNf+g_RF9<FWB;?DIC~*9^*xgHb-X|9t
z6cUU`)qk<!u{(s3VClNkohbJWCa<7Krci0hO3w6+>ka56XBBmZs)i<yZ)0m`zwTR8
zHDl`)YYmtE-7b^G)?TiotEZ0}tz2ccdX0sp6^E;3V!lC*>F5%MZu-u_iQtC(AL#_7
AO#lD@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_quantization_int8.cpp.A58C21992E5540BA.idx b/.cache/clangd/index/gemm_xdl_quantization_int8.cpp.A58C21992E5540BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d8bb10aaf4264ef05ffda9f9eefce970b9cc951b
GIT binary patch
literal 5932
zcmaJ_2~<<Z`p<+UKynF-5JJMfge3&Rs*yznWfv6`6;aR&3IZxYgs1Rq!YT+6Dkz{>
zp;mD%3Ka~ZC|aS2)#}q~#RYJwTdmfIYu_aQH@TN{{^#W!zB9l3n>+K(%s0R9yTpWt
zgxH%>D8b3$d6_Bem(wT|iXr^VUN23Hfq>4YQ1W)2j=Qtw&J=?iSDwet5Jn&Wb-m1g
z{5G$#qWarQEz@J|D#Wu6+K)w^=_#$r_%ZtG1wkW4KC|h;p!v|O<i16HJNrFT@&-Ch
z-rw+FvN`a9A)oyvJmlo*FMl~bRVYnE&yMxx`v=5}o0@i80@KH#XZLRDe}Plhl(u|1
z>CVmH#C7=PqcZDSWAxEel$z0J$qiA*HYrAw9~-~T)f!zWKUO%yZf9SYv(u2poUX5b
z>hu;K*S(pLxodr!cvhPE9_I*7vQKMXTzKup7}kP)Z<(KOrn^W^{I=#D=jD*?yT00U
z{6nm3mUh=#2X{X7igAlN;^UjYBYAS*j*XUr^tIL9`v-5ywH;dbt*V+|(&n2{Qh&W^
zxkZ{|<(j0|Pk(tF$tk-L5vn6}9=jcRwR~&IPpbwPu@Uu4!ri5HF&Bk58p2!K?K?}S
zIPC2lv)p*pq=7jjX`fH=YrpJvMyNYWUh?J#qv1s-O_t~0nr)+aJovsj$87blXRIYR
zO-61%nZ4ZQN1k%imYRk|y`n$cv^Q__$>={a+G@Ow8P&CWp^b%TZ_w1zvnQfzA{Li*
zj79{#4yzJv9GUFU(Rp#mpk{hmK;dk&b7w~r__X%kh^Fs8wq?7H#AJfVs^>;^$;=od
z@8@y%g@@X%Y&QwhKTy1NgR=7a*hSmFpFiC7kMlu?{^j(kr|izLn)s*1zQW6Kzy1Eb
zvgBRdqr8xexD>O+agxCON4}9-^9JglwQm{sZO|6_o9qj^;C;V*TrMhjQfKh(hD4i{
zJ0%~MRdbwd?9zoVICI;7sJl7p@kiW{$o)-H^R-)1pYJ{v9(i2Xd8yEGr_R#x>*8`Q
zb6R6`NM5OZ$L7c<rEa)aN*}dZwAOcIh3phf<iL7*EZTOZZQKX*cgrz}S8cI(?K_vF
zyYyac&a!Ord+YzQDK~oRc)e}*LK}SZtPG3OFXx1$vqNZK_wOpVeZQ<&@MT`^{ms7b
z{qU^H!?8JE3BvkQLbs<Fw@hD=KC}1B^IbX759_aIxkcJE4{zJ_g<JEh@)1Aw{Q<{5
z%Sp@J(wF3>T))@*dBfABDLiN9Us|v4m$;{H9a^{KVE@`hHgEq9I)yj)23NVnwP*W_
zU3k}Gm>;Njw`^!?8E@Wu=Lq*y-JMkJ%6R(&*d4cqZ^i^Wt{T7iv~x-9<u^~7es<qm
zryq5%alu=Ilw1G0m-VUIdDFh1-cc$pQ0F-Qoro0|4&3V7p1kDu6qEfqyoRNnH8byC
z>=@VTjwtG|m|QK|Hs375?MSdoLhKZ)>HBJ1PhA}T#mZ;*51$3Xr>SQ~MNJB~X|i6<
z@jKszEn8=FzumOdwR=+OK%~QG0Y|cBmOS#N-^a9P2fsOY>$2a+fyy_Tsndog)$9xA
z+`H?z?2ne;WQ`x321dFJIzMp7E}eMk%yt}k){?jCrQ(;sCa$Nfh5OyTqzAvA7SFr(
z{^ZUe^Q=mz;iRCEft0KrKkAyBB>euwQ9pRnKePXlx#A#|LNQtCU?c^B7g^LK3I$)r
zSjJZys6FbRFem5LqqmWK7Bv>qRtzg6wL!Gl^g4IsrF1bySr6&uwB<T#gJ}PQSDH#s
zecj5U&V#fq!&YBy5FN^tZ5x}{r(jVRK-!*R&rushFG|>VZ#?~V28+53(p&~tM{N+D
zaP|C}o}(dmS=2a4^BH_ywL$dK3+1VkdV}|{s7oPj#xT=W8$_qb=ZfE!_rGFM7em^G
z;lfrMM5iWfy4hk{Im)8WfwUdN&QNU-omE|btso+orbkJDw1^|JQ5!^8&j<kt*T1Y_
zQD;NihGC<xHi&MLjR|n}`Qt3=a!B(SJYBUxbj!&Ro5ye7?q^XKLE3@gV4yaL{+{+%
z=Ah3d=y@WfEf^MtYJ=$0ozBh@eBj?K>Ox418Dg&5Ai5_;ceVSbg-jMT8q!V-Cq1=6
z^zE^5y?m#Q?JVjNNDCQ4J+(pfaQJ1VWa#`4ENVQY1q^|{+Gyzej`*OU{yvKu18HZ5
zv!2=@`a!KEepDV7#G=lHv?asRNNo^3{%p9~?NO2~i@E~R90o^QZ4fO@S($wn=0PnF
zR_mU@6<Zk?UIOe(fU_nD@ls%4O2%bCTn2myz7qIUf~f?q0zOq_tN_6Z5Qcay=_WtP
z`A5?LhMQq#lTbkEt>R5iJ7j(q(C}`+?gj>^zd@H}rdMs)6)J|817<m}Km>+Y0A>Y&
z(Ve3*C)_m*KL_~dfEc}I@~vwhEBOe3^IrfC7k~#MxR*UP_8*vW9TGjjum|uF!S{WY
z+vuREhr|Hj4FEetSg~c2Ejoh_AaNK>It;84Va-&mE9M&2QE7PIoV*zHn1&aG7DOOq
zb4r@}{44(wxZ=ZreHa*Oh+%1iw&jOQaM(-0_!2NhgcGmHpyh@~I*o?Q07pi6NW<lT
zBPZiRz$pZd=tUY{1UN;+86y3mF(>E_h6m6C3{j|Kc;fuTB($p-UI%z}K!ApHAFMrU
z(%+W|oooXZZN%D92f&;lT2n1w52Q1ezt*O~_+pn3=Szd}#V#e|GQch)V<liKCt`sT
zIH7rHxDxm&K>!&qQi6Dd{XTWA2<YDJ3>~$SSj|RDM<S}toFJYU?o0P&leWMjQ9x{g
z9I)iT7>%q7Kyd|-5V#81RskV_JAkkQcq2UR#f_Io4>WCnkz5T7tAQCJFuVpB)(|+P
zyZB|s^Obv;7%nxCa)@qH8>!|@A?GQbLU6eOxxoa=3&<m({D6ECDhMbbp%Xy&1kfkm
zJORv409yjL1KoClN67_VGF4_lDuUrt=u=G4wna6LiUR-IaR_d?SR!7G9?za!vz8;;
zx(z<IkXXz_iaYz_&+~4G{<x$?!z%&167bO6G`tG1tH@XZ*a`yAk2Z3AQteNtVR)`b
z?qu{DhUa<Ykx;%zJ_!|g6lf4ih431v@*uGo8#5a|$~G8oYi7$w;ez1;GXWn7HcFOo
zH&mID7{J-k*?9QBvW*l@i#Z`xVz`}|ojJ;g7;bN7Z;p01Dfz3L)0m0oFj8_&b4dY}
zZ<?<`N!<_q2K0x&f!o;)INgLSrOcjjTGsa63P^A<t`S-#%FWPORbnQb^`d*}p)*4M
zz)q_4Omvz|pKO3SrAf0l%o*Q5T?>OJMymvR#MV~<S``^908K&Q^q54qYkwxbf}Z6V
z=V%0WUGeoE{)Q&i6>R`!4Zt3amz5WI#eCq}XDE2NuDR~$F;rF8JQB)x%_pG(*8&Yf
z6$M<5s+=H}6G#u#EN4UH%t7HI-??ylQIKdpdZ(bP&vt37VHkXT9@w7;nn7g+okhL(
zA9Shu$BCAr{_-PHnG1Gb{Sod;Bk*bjLC9SU-vzvOk#Q67Y9erj_SwB<=cHf3({`|M
zutkLmhP#-!@X)DML}(SgQ2adw=I2#9tMrMnQ*}~FC`~7ggwl1=Nhm`ngM>15GBs$c
zOww>D>)mR&qMVqVg=lbe1)xT%sx4yoeCR%8jWkN+6B;Q4EE!oN%YnF@tdW(#wsN9I
z>URJxnzHuAsr)WF_!D|^5a=Bw(n{^Wla|W$&2K=W4G7zS<`Sc_4_+@-%||%tWawm!
z>M{%$8VZe38MZegk`^(q<RJ{J65vt-T+!^RT;x(p#$~{_3}{qVRhtO9iQLot%EWo}
zce@Yp$g6>MHE=)!sTAG1hKzfFxQ8g3&>6$HR8^vc(-OKwA6>IU3pe_0zZuyGm)r=<
z8i~;k=PBuPGn+#o!C`YasKPtu{Aj_F*YjMVkSYOB3Cxi<D(s;IUSu4h1W^bhCDvh1
zsChBm(!|mlm6sUqNB7g*84UNQ`;$=O)Wlgt3?zmo#t>*b;B6<I$M9Oft0nOF_X2b4
zxpPw0V&Yc`qLEh^lz+b}GFE_a1xP@ATf!~R=B@fhQ2tARZV90>7(7MYQZg<Bx@Ban
z1iDHBcYGS|zt}n$psp}2%ZZDwO7NBy1V#q~Pfhg&K%81VFs>&`qaN>=WcsY;XK*~8
z15bnk6yB}SoH@vQ6qzt=Ri6vQSRQl_3{8Z=I&Y>7M5335>^7Cj17r2oOZF@QzKBpQ
z*|U_4%RpcmAu*~H;nod2QU9wQ*1;PpPQO%FlHM{RFjbYLx15X%!Sq5f1HGuKB!h~8
zrhd51IrC-XH0zD(8z86#E~umG1_)}%xCi+55T6C;F2dZa`W_<I?@jk6t-lb=Ed&}t
z!r<&NMPM<S1m$5UP^zycVvwoysbqn2O7tMYu@JBd$pTdbSVd&40Je&W0u|Z;=4z(=
zZ&rurruq^iru3wHlKhr~nR4RX?)4tZ5|{j3221^kRW!O>a5UA28!?(yv{gJRRaK?8
zZyj-%&vXtes8hdJ)lqy0z8DoD&%QqTZnyBt9vB=lK$8I`av6hXN|TduA)pnKaS@;u
z5%~4-&9OSSJos?@L|URD+BqaMP$g7fk%X@i^autKv61vh5{jZnk<dK)JQ9kb$B<Ag
zJ(h&x>G2x$`RtIO!9Q=;|Ez|jDH-syAO3Gwif5luH8NG^|9)#wgF;zFG?R(uN}{=%
zXr>cQDbY;PG_~*!Sy@=_j`i{SS<-cDlUJpzOij<oT$8q%lBVfISw}QeiDm}T%p#g=
zi6)+zoS||yA|*<TMrUa2=&~?91ATUw;mja`Av4U#bdoXGY>LFsKOk`0bT<>>R6AcM
utN-sQhi`6SZDVWi;OOk)>Mr*1@|kSu%@cVBn{YzHv#GP@&j<yPl>Y#2vlcS|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_skip_b_lds_fp16.cpp.86DD71A6DF5D193B.idx b/.cache/clangd/index/gemm_xdl_skip_b_lds_fp16.cpp.86DD71A6DF5D193B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c135648fa47a46f0065e2b865aa0e05fc55795f5
GIT binary patch
literal 5942
zcmai%4OooX`^V>=rfHga=B36=nNrF$sVFj0sicxj5-Fs5q#`2H3+aU|vPA3cM=xqD
z6-7n0k#?1KEwb%|ke6S{Vy)Me^%DPGerM+QKRu7@`d`<7u4~RUpZPxLoclTFKKFf}
zkQx5|jwTF-Z`_RZ#3iYVSquil2z^shlb4^+V=!_I8I1JO+OWIHcPEM-eU6*AD`tL>
zd$~cBI4JvSvu1F&9y8E*lqecz7jDRwS}(t?Dk;8`ZB?Ina(m;Xkr2RGsb2AmO_J^9
z8M`gLgzK5xKOB7WdX@Y4;YS*8EXz7H_(nx0I!>&aIN#D~*n?-hpI4YX`~G58bHteV
zd#zV~ySwO@)W4G2R7VfIjq}S_{?pP|ap1s{2TwoUpIFwn=lA$?Js*D1zghF0q~Tgh
z$JsmecV1qt{C86Jo68dxAHS2KUOw~mA4I}_7ggYq9ajp@PrB9C8*;pJ%{q_F<~CLT
zwiQ8(?nE6j{LW`on4l!SDtlO2qvNM_e|&Q&^vKp9tz7N<=B|qRxqP>?{a4e=y6-Hw
z6cD0dyzqUn&!u)@*5IoTV!s&MavuNs=re|@)LkyA4|%bH)~Z*T5<f%Zl)gVS1?TS^
z>d%SZ=96+!@Ae15cYGU2h<7kk_ukKOWrtaX5BAH9YX0){yE;d?ujOCuUmu(7{4RMg
zCvLH<|3XRZ5;I}#y$Of+O|5D!&nVsZ`P;^m4`!{YW*y4)dzr7;Up^|-?oYdcT@Kfc
zW9lS5`Givvcl@oSn)+_D{2d|2hV8tj+E|&l<U)u~MAWa99w*LzUFy*5V4jgW`E`YS
zcZlClZ@rh59&)t_4YfEdb#Rw8RXNp`fBb1smYe5#%yw+q|M>Q8n0Nic%7R|ATPg7|
zsbgPQ{N*wIQi;cOgM!EU@4O#uTedQORosCAuLCF6vT727Ms(Z1zjXQ}Z_UD(JuS^K
znt<1X>o)n%=7p!dzLw`5^laAPHjlBnbMqdg4UbZ8$#wgu>+K%f+g2t^Jhnxix64Wx
zZsT2WaP`f=j~A@dPTSnRXt!x*M7>SbWBzcHh^K*vE5j;(j4S#o=v#r!2iy4233pw6
zDwD-G(@idJ%=&ei`^<L#o=x|^a-4hY-|~6S-?IcAcUL)kB*koA>8j$@9Q*wI(U0n%
zgEMxyzl{hx8h)%Y_^%A5#nzTtrY?{&FZ9*j{)BOH#m+UMO`?v}cD;=|^ba_?Tu8px
z)A44DWukabwW#aW^B>$-m!_FL?&|-W&}6>azk1EI*!!8+p0w?$toiJ`CnreSJeS#a
z<w(;&&fHBl1NxCyzx8``_&}aPg>>hO_X+W{l%JxXpY8Z2CFa*}E3SVEw%8cnI!S%i
zbHU7&;aR~ofr$k#Ms(%3EPnm38ArAToZjHrndy^xvEJ$MM&YLICX=0?gayrBm^Z2;
z$MwlOQ(0ccZ<pn%Z(nxrbgt~VUG4jNM8wi)f|0U5F-FK_Foai>d!wa|l>{S>!63Z_
z-sZXnoowAlved>n8N4p=n2|{K;Cfi<8Z1lq%<y0pdFAq$p-8r4+X;0ImYv?O|GxEI
zKrxRw3&}=oBayDb@`SMuYa+UL?Bp{dkSvi(taJ^QJ&tv$zi!!E&123%@@V#Gp{~L5
zw1U?I33Y!j;4#CI%w=;8bq$u6MP~Mtn-o0cF=r#$o^5ZWYp|T8C+Xy{$#Nbu1j)v1
zW09`G^42c9rhw7=+6hK1l7m^n99@Iu-BbM`^7^M39y18ZwrpFWuEFx2kfSHw{c$>$
z$6SbH9-C*VYp{H_+0M=__VRlka{-cNY?-O9!E$SeVS-cUd@hd}jATBW&(}3rzEsmb
zq2-?rUmi0I$zrxxq-(I;*KZr!sVMH}G3O##z!vaz4VGW_q$On^<-78j^N=iL3;DVR
z%Z%hDOH=Yi42Ir#UiHDwm>nE~^tJL;47myEZ{u%IQ7Onv!ER`^$TE<Z(OC`S)Zjwt
z_25<yUKBTgTLYaNVR|FX#5!Bk_UyU+Nv<A2R)BQ{Ouzy`R)TdUoom6mmSTRqY|`@M
z=I59!QUk^s7=}Bt$UYGCfe15x6X1T~_M%yBR*BCyjRJzK1M50!V9^<i#0mQXy4YHa
z1QlR~2hv(3sHAfh2&(8@4T5Sq*Mgvy&UGNDqZl7NCija2N01M}>>-T9(@HAFoae>W
z0y6Y5*gS^uSimRKfcwRXBuLLOo-TNBf-DAJF||4bSpvKgIyZu%5uEU#1lbPG?J$|*
z9+=((voV|gIzFSLYU*`lO+QHb!4eB2xjU2cOpW)W{n!aUJ7F3YSma)C-wU3Yt@@Hn
z3@kogLL)qdVNbyX3!`OfQTdJWD_AVD5Tu1*h6NT`1kxfp7lX7IY$?42q$QM3j@f#o
z?cU!Hj-eSez}J8geh>#MJ?<Al3ayxzX8^Tg&0yLL*0|F+MTUFF70nVfS{n>+0~;*h
z*oO8`8`$vqB1lELViVv9p!7V?y|sSfFeW->9A5$UMuPMw{EacWvd$bUcaeOL=8(C`
zX5qf>q94neir0N&5o9iKazTKnCdf^|*+l2fz}ZaaJmBO}>~1k|c?Em%0Kp=C415gn
zv@CKOZyKK>6;DN@nFcdyPnLq9lsZ%_6b1xkbXJ2vO=k_*YhVmEm__!2pqH-qf}$5j
z)Ah-{l#^(mU(gE+v5pf3hx<i>Ks#sUX;1B35%7xWzFi5*N@`GotOjK@of|;WK==J~
z;C${&-*@N%B{pKpiyJTNs?=-IRyKj82}WRnAWwnh6va4ip^&3ZTG%uEf5>~e^7+z(
zD6-^(C?BL$|9r5@ryN?VJLiKdrQZO<8z7{!`3=fN3GyZw-lX*D3!@S*9Xz!cxl=0)
zZv{E-7<A#voEu|bUD9Kb^&qGR3oNk61`ssRxe)}76h|IX8uvwhHwV4Cx4t)lT?Uc0
zw;iU)t3R!3t@H0i4+%s9F-}wHKulHmVV_=9U480!h+8~CkgF_Ljlu>H<Z8>+G)l5e
zqEV`4Dvi=C(`dBTaxIP0jnXC5nCp$!(<s9zgGN~<Swj(qAhS)fX|%y)!w}&r18t_Z
z#K4=mJUL};@ZeEzl(_1FRS$e>)&^iT(76#<jTEm4S>$-_^`bw~<`lrN0xGN}PJC}~
z4%PdS&<!KHVI&?A7cwYpX{#7)0djtoXDH?TXTbCf<r8ZI{W={MOq`8IllVwNaj?oh
zbKW{?u2BH``w_?=QO>_nT)FG=nbs^MbOOJVnms4Gd2@UJ{WIEu=1HRPK)JOu6T`}`
zUPSNT2ElDWVrvQV90Z@Eb2|jLQ~dR-Y36QSJ7j1)u81qeX9agh%dD-%@Jnu2>1}?o
z^Rp{DGxIs~X(wO-X911EIpH*l;6%_Uk`qazC{EN6?Z|N}Ilksy0y3olLJD9$HUbxa
zT<#Y|A6jv#rzN%Gg}^JMhnoseR?x#s6)3CdTnoxtI@f`+j?VR<sHb*ByRgn}Uxp)L
z513*jaS5u={i5_l$IVn>hL0N}ONBL!6bd^UIVqem?Ry^`l`BMHMeBYJ-7pEyqt1Dx
zh!y<QfU*!@#24e*l^}&8p%@pNXdYL^WIPrkl_Ch!!L1&`frO7KXygEs05cpe4y_Di
zspf3?%>X4OfsY9P%_3FCs$mqBf_*7Hk(7ab8J*QIQB8$PZ8bPcL)9zVYEaQfC7Q#3
zO6(p!9rY2pZ4=0wsG>m|!{w*w+z-lraKo!mv2{W5qqJ6Y6x^!7x{9h7wOPfwn&K07
zzlTK(%o&5qm;lQFE9ymhLDmcAcsy-}8{bQH)3U0UDvVFE-X?asT>29^zLD6-0+%iX
zX)HFjz!8BU#bU7qzAm6u4^}M1tBxpA5l^G#id0M|?|H2&HVs~;3tm&|Ap~1Rpgb_8
zfzFK()Ckd7uaCUte0Yb@82Lge7?x5|jX(u~VHus(V5p{ZFBtYx-0Z(Ez0UjbEmV=L
zlCPrG0juS!X_O>SqEV_ml}2guG#agyuccACG+l;mW|8Zq>uHoB&7e`1dDc*%Ws%wD
z*)-Z<zF~;Y2%dCpNlIWKCw>FsH`GR+^>EtoIho&r976*#4LytZfwT{Xk`rEhR4-~T
zggC+x<XISY78KarR%J+>{<Oo-bU|r)0R#^T|9fdf{k1nns(&H)6+$3>?A+9Aa=CZ+
zAzc}*ETpbx+A>;MMCW4gE{3VNBMV&(d`iF<Gj=PKcC=SeYCg|-p3zhgYys02D(z@S
zWp2&&nkzIQNFS+>40}I8s$?n)ib}z`6ei-Cw9eyPMrSp6slkWRH84*@Rf7aN)Q)}N
zjP*Oxwoe&TMrP;^wS5yfVnKVT?N8CUAKd$?+vwfAM?I|%^c&DxZvoC0AaF-qyP&n!
z-mj^5F9N?JYOS9o)P=c;eSSuJzY|P$f&=bPpi7U*UW%XXt$SDI&~;EBRZLFaF4RI4
z0;>?XRJS5v716mESjBWM0agjcIQpV3(%$naGmBw%F_q4=m%fk^2*=Y5G;Ey9xi#Jl
zy|yREQ-b$$a4G8%WAesd07WigfDtxw@LAt($KRu^(dsJBfueorYJ>P1D7xwV8z>$O
z@jvG@iV~0X(*JjLl9#MpthGHQIW}do_D7Vq`R^wh|6nk5&*)H(yxF~7w*1LnF)M0O
z^kT*$>Y>=gxRvn;RZC{-vGmyn977(#7YPLcMt(jDBW{4PiFlala96j<9-dxPyd5PD
z?xQA-w*0^Tm71B47-?lKx3RUeALBG``~;Ut7E0MzXJ3ibe?|;*+We{0A&~JO-0yH1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_streamk.cpp.2BC4B9D993728CB3.idx b/.cache/clangd/index/gemm_xdl_streamk.cpp.2BC4B9D993728CB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..130ac4aba57c1f358f612f7402e540cb7fceafc2
GIT binary patch
literal 2616
zcmZvc2~bnl8pm&L(g3*$2}=?~;37LAu^^iui)@BP5J3fs5M&o53IQHiP{yq=MMY3h
zW)M)^Kns>d5sHP1ETSL+1qVT@Ahh~eECqR@ypy~+Z_=6GnS684`F;O$zVDv@iNB{%
zXhuO0*C@|)adb*J2|*AQ@JmjS#1#?|WI_=^(o25W?UOueLmgWCy|sQ*+}$ybc0%U?
zho{*_o<{C8Zx>dldS$`+$hjKRecef~Cu?7b^wB&-dja7QyO0pSb=j0>J(5O~yJ4R$
zwoOgCJGO96*IZvow{Ema$~+X^$1>u^<ktPXU6Va+bY__Mb`!ljDcM$QZM7q}Y<VCv
zDb23OJ$EIs^^E<RW7uTwt^SAamu5F>ht^FyOni!ZjwbsxeecHE5&n$p*KG8`HrTSc
z{P@}Gs?gSJmot6F3sic9T!Z?;>wb$zKjyh>v@Fh~byDxea_p~cT%8<azUboDda?H=
zdK&5vWK0iaxE^6|69;yW2>V_;uzCY89;+WxL5q(K|JqitL993LSnF&z+B9{;H|%F8
z-MjS73r;>1@nG%ok}16~lAX1ubNX|?qVk)q9Vv0@L#6t{@0gy)alOuy$|*E9A%4z;
zT{h5PWGf6+aX&G`<y@jG<Q(k0SGib8(+L>Mue|VvTfiShf@9MP_oMZP@^bo%%3b=t
z)x5M;kaowp#xJh>W|Z3JDo;tu2@@@pjZ=Jt&Q%S<3fp<oG+)S0vr0YEv*}HNzRUbC
z|Ndha`5FIRWBCs?!fR*VG0L69%F4b+SHH7}dvw)uU|wOp_4o$|<}Lw!=eF;a$s07h
zroYK*w8@@XfBKK{p6LgbqZwDM|0NzfchYFldaU}|;Z~GI>bY**<*JsrRiRwT+p@tT
zcqm|X$J0PN-Q5kpor$~Tdt(1cyXA&TkFgg2rz4cbTRQyvCrid-f3I%!dAIzr)ZX;U
zFh4J~*x$0qv@T2gFaL~Dwa4a{+Ii2_9q*nppXe}U{mI*M&SPx2;LX7Kw&ynYV`f^p
zMaIJ~FL?LNcy)$vAzpCS+Sc3G9b_?_E-YMf?I)FWti2AIPZ=p(SF-#chc6dvB@eHT
zyy7I(<sI(cm-{5gE$eK}&8jcSMd#A<@s)t$k{j7~f|sL0V~;&zk9SiXmwUfG{NrH2
z`KBUsZfHb%w*KJ0c;oKOJ+T%1m-h{0^mt{zEhMGBw3)TuUfR~JQli<qKV;xkr;cXV
z+k+mG!dKDb$uA}!Oj=tlX0{uM>#rXCDdT{v+w5(2^{2z{^8}x%b2-g+spGd=zOq<x
z&sp0uZK>P4<W)VZ{rF?GB+K~T!Gt}F$}<bT?<?*%ua%wMpMS)+ce?ue(gmG>RJ}nF
z_3}!|fzJVRyQ~ki-gXh!y^=KO-9VBLi6iC-2*QjpP?ZoIJ1{f^K}38qpDPRC{>mO!
zfu}v`qT?|P-32g4#^|yDvFY-m<_Ak2#TdFBV0E%OLlz*muj(t!ZaP<qp?&}>k`?K)
z0P)t-9~R=PSA#Hg55P1sO-U9YcBdRYJLT7&hoM0LQ^*vSEI_<3BxAUOeR>i@cL1zJ
zR#K4#h!cs-CyHcI1%~<q%p@}zvH)>TAFsi~xUh|iL;!3?GE<ZVi1VF=gpk265g6(N
zumV|uE(;JBy`2kDxANM8p}_!i$Q-sTK-|*E<Jm?&UdGUG0OpZ->tq4q2mVU&rWt`0
z4D|(=PNvgj0phOPZ5B<RyInDKH^3}1iy;dTk53pyJ~1qwz|frlGsp~@EI>RvoSJy}
z2F(UT0{~VftI}ivVo7vNazFT2hUZS8{Ns5ODSta>stDJ_4Pe()k-%865v~9|03=ca
zz>R=p2OcaZaT0hz3G@Y!NKGLpQNcCjBw7MPD1lx9^-7HbCD0>aT51qD?e{`i%Y@gw
zK%k1u_-3x~u%TtZzEZ!yY*4BZZZ5a4g}@Q+3vC1>QV+>VLh)ERiMJpCO0ws-YTEXl
z(jZYq3R((!&`MO1qL!i_Y@I5?v@kt*y{IA@gT{hA!g4`VQe(nB;X<GksUhJssGf||
z<w{W%aA7HM7Nz{npHvZ#%M(DCmL8olj8uMDOQeeE47wV0Cso8?Fw~&is3KK{sv0!c
z9nxRoCsti^K%%aujzbGjMNBm&4nG-H#8PA7(9N}fu4mei3_QS*aBMi<urgQ#=v8`8
zU{kONP`LDdz-iE{C2j|`aB1*Cq_;#aDIQOONx-Z@f(ZCQU^`%4>CKUoIN)wj0=ooB
z()%NqB*cAS(n4+4PQ{UpT+okzBv64He0(>a|MlCwx(cAwIs#r#fcaqDPQV|@j|U0(
zko>p|#miB-{2x)gR({-u;=jm`Yl!$ABFsl5(Fx(wQj;Z-$)?hmz0&Z%Ck(zINa+7x
zi6U&?K;#q@eVG`#Hw=jsM<v7p=oRfvBz;9zP*lRGGzMMSL&eR-P=(^5%2rcf$FZ@s
zvv+XZ;`B9>zgf@PSo@!5SzJx6^*Xxx21Yy+fvMR>3#(1IrN#zxS0+p78IEoZbap3f
GL;eeGNn?uu

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gemm_xdl_wavelet_fp16.cpp.5100C57848816EFC.idx b/.cache/clangd/index/gemm_xdl_wavelet_fp16.cpp.5100C57848816EFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a44a41fb78543365bdb7acba2d5722ff8b8aba7f
GIT binary patch
literal 2728
zcmZvc4>(lm7sszVV`j{~(+p#13~mi0Q%QtHq{1MkM4>RrpG8dh+g8#>mLd{MN!gXs
zA7K}5%CbpBDQ#?sipBPK^QS+uiQhZ(d)}Ve=h^!_bDsOYpL5Q8?s>mwJX}PgAL$5U
zAMCPA8WOjaj39_A_{GM>gcp+#WSWK`yH1vG?2GBM#Q&I$3wTu6)%WHwp?zz+ZDGa&
z@q!~-$5R$(2|Zjr8eVrBhqXlyonaP!aO6kR#t>?@>3;k==ey|hNd01`vcSo!N4S3L
zX6~K)W~i(%e4jZF@4$bMybjH%CRH~#6QQYtBL#yY2d5qfnfCjyt8Dmouz7Mv=ic!s
zyF0fsroyhL+RR#S8O|siA9_7AXhA%!8ZjDp75o5=U4Jdzf$P5YK5u=U(Ui8)iNyP>
zoW1{8GVs8?!C*(*!H|+nwMT0@4fmwYezMJ!?0NUeGv!kTUUYx%zKIv<k;9x#XR)QV
z`s@b}6P|S@*zY4;rQYGy8AZ(r3r|)qwF)oOBbzE$mLCv|)?{9P)=K$1G9sT?v3G`i
zsHgEoJ8CYhS$}6vOn&8bWAAvrdP;d#71>eQq7*AR|3Tlx^pdmIyScr_aX&`X(%cig
zuYT0d+SzXMGQCI7;c4shy_Ty_>V$f}cx#*D6YpIr$*Od&h$MxbdhA`Ha=*oPzQ{M3
zZu=_QbjI~bBc)z+R#1|2<+-;wr$leV3s&uzdzdNRqugBUZdRz}xw-!R-Cey#p)CTJ
zlmTjWnJ0H;<)xwIUw+Zbvo5#ak6yOYvQeJkWu4*VK5%Oyl6rRC*}d?*k8Wwcx(>eC
z?E8DwGfRI>c|I@mW#jG1hKqf^r!96$A3waIQ7`VjTyw<ddU8Ooy7KNYwBk9<Ea_0`
zPEpyTbfRqPPlq!XE&5H$qg`uL#crKB>0Kjv+gj>Y&YA@au4H$n8T!-vYi9Nnt#A2z
z>L+}T2_1s03YtEhIcELZws>);Seq{%^t@w8+CCj<^25ZkLtTkRy>?>m;>_HL)U02(
z9*Vh}*j976B2N4LvM6rX&%+<J51YqkhHePAQxlJcPp2JAKG#%nbD*kwbo|!$c|1Y$
zOnbnw`cYTW@?eqvKL_?YZhZe+$;qjVcdg&t9`^0eE<UdLKA<Nz{*=laWbo9bfuPaT
z-I}ga#|sg{>Tz??Ubk(T=TdVXPR3g_G}cE1TI5e~8?4LMQY)_=$PJB(K^HiUX>^+n
z70{eoruy8jKC$a4epc0XC8zG;{kI|AHNWRH{B);3^<aP4rhnK5P4QnbnKz|Ho5z|j
zB==4WZdPj8+)b~p&sj8G$e;1M@iZ=JN2KrPInT$9|E~YJ?ziIXh@<<w?meq0{Xpe=
z#p`vg(k-4U+WvXX$Mk1C!=^PibQiWI7Hm8({Xk$gMkkQi-O|7@B?RGoEl`V5TG4``
zeh4Dbr|9$K2F$Yj(a3SJAs=6njG-F<#wZw5ZXh<E++El3!MOlK*8<F@uvu~gvCYMf
zqO@zd7cg`^z%&YtDK`*1W=@Q6tC;q|(9HldC=3<3f!K+jls&Tk<`E3_0hmsqYsd}6
zVSWkSC7PMT80rqN3PnX#ZXk{#aoTAVNePB}0IW_?XUPr32Rnq9oelFEaU>96Az4V1
z8;Fmt5-It0%>-ho7+@-e%9I<3>uZHVtDxJH7`h2yE`>W+ZXj;(P}ydj;7!NSbpSId
zOorS*d~;XYbl@oKB?IvVSW`oDzT80EdZ}^Qwa;z#7`hQ)4u!*#8;G9_EC^~hC>X#{
zPk>nz7DH|ze$yQvm3WR}iJ@Kqt5Q@MaszQp$k(wJ;O`pVyAu6B@0&#b*FEDB!j}k!
zUyVxyh5{3q0crq*$OHfn0wJM9v_i;A!5SC>bpb+TqEHAiA*>ZbbOZ)41ZoAP|4Sp_
z6fa^E41tOPA0?9u%qv#OGMt^dU7&<Zgo}jB;nn;7-n6};Li8AHvsf%P>?_n2C{89S
zm;{_;LpZ^cKns=$R>0z<^G>@Qf7y}?GI6cA{?G|Z5M+|+Pa#tru@5?dOl|@%m}#f0
zL;E5h3wMx7bG7CMsFb9{+LXqD!K<=uFHL{Aoi^sEZPjsDi;PRCI#fL<T3kZYq3OXU
z;}T2<(}RPIOBgJM2J9wm1!$OT1Ym_Q5hz?HeYg!;U_clv6s{)V!NQA@Mht@HG*zLv
zgvnyE;RN6kRhB9nYVs%Yduh9Q>nI7Au-R-~s2p6P&Q{lj-NGdtHb)l@0Xz+qA{#SU
z3SGjd@V~5Z37^L|g0ApvAQjo@!4ObAMyNq2*h1hW8$bo8Y$c*l3A({D!11y{RB%gD
zA~L?XflKre1j1fGC4t<s@l*&>CwMS3><Ms^4XJ{Y4xta7#;d;dr1hBaz`YB|0#$gx
z$5&5|F8M2fPX*tUf)d}M&>yVxP$FM(eHJB371v`ZF|N2yA`!_Xh4`Z+B1dt(pG+i?
zA&&`(+$yUyHYO<6SoX3-X8!*Kz!wDZ|ME(bkmhfxTWi;Jg5)2t1qqS{M}`8pCd7?I
zR;EyCDj3dSG3Pj|I;=KOr8}!>veoBuEv=T@*sic!xrDRGTF=6e_^(|Jp0>`s`FwrB
i0^!2N#!F3>nSZ5gre$Pm&(RRMY(*WtS2-!UBL4%YXsdhx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/generic_memory_space_atomic.hpp.C9D4AE0A6DC5899F.idx b/.cache/clangd/index/generic_memory_space_atomic.hpp.C9D4AE0A6DC5899F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ce5a6d47eb509257754f8085191b4e46f4ab2f23
GIT binary patch
literal 2868
zcmYk74NOy46o7Bb6WjY<@li_Yw6qn`wy-iDrNIW}r!56^im6PiF8Bv16@>NCppGqZ
zj?Fpe=6^8GZEi8@H0m@e{yUA%AB<TRhZ-}D4jnqj)H#<e+wP+|@7<Sh%6HE_@0@$i
zdFO^IgTa?sg#5LYi<;_MXEO+)Wb!ArMjOVc5PG0SXwl|dla4M_3;eN7dwSGmnZ9u4
z;wE3Gt<15t&~;+$v2S0wUUd827k@rlnsKCh$M@IW`wI&vcW=9Ye@Q~sz6P=D^6VvV
z?}+txol3M!+;~3n+<)cHs^TeodeV|kymu(Ieb=@AuJ)O4x$KYFzgCwl7@WJvrm3lM
zHGHToPdxaR%F6WhJWX|~@}GB}JzaLNVA}hQe+OFfGgsvA{po{A=4k1@Q@%;uJjcDQ
zE3c}ry=QywOuU}2SD7D|uWv8&Pn>x4*QE4W+EuY-O>FCm4W6_IzD;~2!L_1g`GFS&
z-t;YRcX#c0+)@2!=eXY<9_i}|?OvC6dUf7QGcLDJ63kAIBbDpiKh)i^a>n@XEo<yw
z_bj-1=&NTnl^<-l^m*`1Z{Pem|C~8q_+;I=wHG?_^SxWQ^&0Q?J=2Z6l5|mA+%#t+
z$)R)Z{*%dbKVHtM2Wt?H7$TMdfixt~<SCkduC^fY=G3@4loCqEk^r^uMOkUN@QaI6
zClYI;!Dtx}psqapr9Q{DRE(=L^ciL>2~anDGx)#8e{$P7M63rMtkdbzkXEfq7=%<<
zwQ<3XapLU%w#DuF6h0kG0_->UYn_HOmI)jpvB&MJbX5cP%z*u;?2OKirS}HK?Y*g9
zE0zS<fBsq5-74q)?6}%ta9FS;K>bDewWH#3Lv{S@wfb5!mISCT%>2T2`qExvEoClj
z2wMgOsC#Y{h&!_m<i^!C$u+~UBtU(u$dkznb%)}<&B^9rSQ4N{(Ykt3Lr$gcu-*}V
zzx6Zbu!D7^$RT5g)nTW|=@2L?aa1Vev3N!)6k#JNa?x?b8nIL4j0hB!L@E?&v$Tx@
zN{GH$=snY0tl==v@_M;Un78mFDYEl96piHr3WeB^UT!WNvV=xbWDn&~G&U3f!n^vr
z$87qympC@E#uT~zu+eI?Q{*%X6qOh&6biH96uERbYz^Bfa)t$pO2QQ~jjVYxPCfrX
z0*5WE#UNJ?3mPF=Nv>PtmPwbpd{bn-a5Is^S*9$DvMJk?O;Mq#&?5IK9%C9qQKPBR
z0y*b!lc`B2SXJSH&voyX?2~mzRv*kB@zSdIZ{EvKWU$NT$^f4X&Ka6xR;Xxr(FpL#
z;NoG$=?XR5nls?-3~m|PVg@Qmx;pxuX>(SpI9$zE8=xleOG1%G1B3z@C(F`YfaPG-
zqaQy>+x#ZUil7&$zI5x|6#4YNQ4kY{1Ns0(qF$6ck;5^4OeUB&#HTc8z-PTS>CXC(
zGar$_>)Cn(eAf*NtH=LZ{dy;<X(pR#0AmjOG`?iYxh5bJOe~dlU{ZmD#=MrN%l^<N
zYdM_D=Nci=9M0qODDtFv(iL4PUusmm_<27?!L*=U5)McCs1b_5;a0wtqPDcQbkH>|
zexjSadwUx>=X7?u0ebh1(RgRabaYnBU?C(-h6ot!p5mUWkjLrCQpi{Bdkvyva3CC*
zrcf{|n5~c)5+_3h4315SO;xDP*_H*gyt}zHN7WG|>7K>TGRX50iyAQ*Hj#R2Vlu(F
zAoY~yKBRkd-=)-hv4ejN;;@(X8Xy!xHm{SSY;QhAe($Ra1ua245L_dqsM7TTUO-Kp
zOP2;LLyE|<bXmYMG!<EvE(usZTz~s=0AIYL<FF3taQv1+h|EpIQ)EO&ip<DNQ3lGO
z$blRbd6Ac*DpW;L2!$vLqcBA^sD`3iR7+6=MPxd+X2;#Z^LpE*uhlP9!r3|OE_Y8*
z$QSU9gLILvwIVQ4A<-{ZQq?VwO#p)aBq@__Pe|Ft@1`z@Jzlwobg`b*8=wcEU&wCh
z9);c4z1yE2ZC+v}yG^Xg08<ABG}$d(tFXIY%pU#G59@s7JTB}?RZft@<P1@$2p6R)
zCn#o$hbYvHn^WNg9ByG+h5$XAlD<scdcKVe-&||14PM9KJZm0BrMA-H@Hz(jZGMWP
q)~F3$$KY0Lt4z_l#@SN$h|#%Xn)HuW67dH=L~GZZPcSBx7X1%Q+J}Dt

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/get_id.hpp.D8C96396D4A9E3B4.idx b/.cache/clangd/index/get_id.hpp.D8C96396D4A9E3B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..710a2a623dd454a6d1d041c3ed55ee9bad3cc47b
GIT binary patch
literal 1060
zcmWIYbaRtoVPJ4h@vO*AElC7283cj2xTGj^Hjp-AWMHV6v$n^P?~sCkt7Tu_oE=Zu
z<~8{w>Ig4$Xo!0K^60T6vv!!wJi2H4pXGe+sb5`hF4($j(*!n#`D#q@7ZaaKKiLvs
zC=qz=lxWVZqE!|gGaXbXF}Q1QR^G7i{>kTm+ArLR*s<0i$oS2+7q4YiZYG>tw_7&6
z?m}B~;`fEmxEC5PUc#7aUiI7V?<MYOf1>kT`}*E3Td7{nQOL+pT$!8n9Oy=YV}CA8
z>VLw>!@$GLz{;t{X$&S9I2jokc$pYk8728`PC6_*aFmCE4X8j=SX2p2AQZTqcIY_b
zXkx>|zyVYsE-bD9CJ+iz!_Us_(Y~^Whk+faKvGyz5lkQy)D>-vx%g@GX&we{paNlG
zVRbNpP_Xr+i2Fs!`(Jq&IDrZTg$1?11VX{jtMw)qKW*H{!@vbpz%R_N2__H<7>ZKU
zirtwQ7<jj?*ssA_u$PgCm6L&!0~i8e01E~tF0l7E%+ubVx@+P#9#&yTVLzDrU_rsi
z1r7q3B3J+bg+acD3A_KU3z*t)cnuQ|tDu>n9o!GBqI#lcFrxzUbY<c^;v0a@5!Vqn
zh8YD5GN6&5@PZjx5bl+rTb#BDWTdm8H^N9sT}e}z9++!^dO<-6)7$OE-14^nzXdzc
z<8qv;Fc0&v@^kYG!NZSLkXuj)VXiQ@un^2V9#&CqQ6ZSaf$rrN7lOH$hgFhWQV8Y(
zn0}xOKyd|g!H%>CXI!{WJp{RclT#QL^gtId@N*$}f((L4JYfc5B%UaPC=yScK^%!E
z$sh^my-v03QJ!Mx#tbyvl-~}?djW!xFy{dy*I3vF=Di>9P2%hC^9uuGf`^+&7#2P-
qtw6tkk`>G^!1R=p2y%0AQF5_3hylv{P{7*qW<_n@^Ouc`EDQiE#XsNx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/get_shift.hpp.0A10E058B6B6AEDA.idx b/.cache/clangd/index/get_shift.hpp.0A10E058B6B6AEDA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8a962c7d1bc007d3a1973d9ee83840e6ce0eed91
GIT binary patch
literal 520
zcmWIYbaP{1VqkDi@vO*AElFfyU|<jg;^LB`%tjzR6^JY5*q+_Ub;v+~;lghbcJ^4C
zU2TuvJ<@M6**Q!3y2_WJT*3dbEH8g7WB>c{uH?DKu-km!&wnjxt9$N$@6G}B<b)K(
z=8la%ErGEGr~KDEb=_Qg^W?HMFHVPa?{Yk$Hp`f^=-Vb1r}+h$I(N)!0yaxk=1h;!
ze$(`yJHAT8#EPM~GB?Qq=n5}Ji>9`%8|rx&*jN}?IaD}w!2|;*6C(pVGb1bGqd!l1
z{Jq`2@i2iD@X7M2fe8i<pk^Q_N=+;N38eWYt~eQ7zL(9&!pbGjr2sP)Ml*4NjomO$
zdw=S#iQ9Nsxp=rlVUiKQXBa-)-E4xUx7&-k<!%3e3nm^`4jB$*m}(wYK2bgyn6u{0
z)qYfSsMH)y?~b$wXI!{WJp}0$<d6nC6-WT}a`B14c*ieYJbn7rmjh^eZ+w{(lrf{L
o9ZeW!BO@0$0AOALhCogt$i(8J<YI0RgLQ_|F5gtWSNj<l0IPqlr2qf`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_2d_multiple_reduction_multiblock.hpp.B962E3920F925D70.idx b/.cache/clangd/index/gridwise_2d_multiple_reduction_multiblock.hpp.B962E3920F925D70.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f3fe57c665b2b465322e3c04d9eb7cac87929eb6
GIT binary patch
literal 5630
zcmb_fc|4Wb8^6!#3irN)%8i$MU0aqg_9zsEvXv~Q6e5ZiU2FPjnxRxmq>(mDrIDhs
zWGNG^TBu~E;WJt=F_ls&S~T@L_s;Kl@8>gr{XTU+ACLF@e$P4YIp60w&v}lIo2%<E
zkHwm`+-?1u@VI4c7K=3w|Ha0|L?(}6u@n?oto4V^`3KJHG*cd)cm0w^ho4DmM&OpL
zCP&{Z)07I&Ej|^yvP>_bk6Sux<VDq=x^vsukJ<lmQ{VJgUu?r`8Ns=v4h_rv^QNUo
zU+C4RT=u@n{%o24!2a3Jp~<pyEw<@>{BdPWSTOfdnB%uQPZfGxs2iIUHu&4Yjhu)A
z!!<4)YnnbyUKGPg?=hTi+Inum$Fcm-7qWYfMAcn$%04@s8u778*`n)1en|*VTTpW%
z`Ch`vrS%t!v*gdjw3nFNs=L0V#3Ac#r^fBgbB=l###rd9HS8JRQ8KbAn{)P;h{k45
z;i`}gRauVNrX}-h&Ryv7T_c(vQqh?q%Cq3W!+zJtT_e`hyW6uru}*BU58ppq@k47@
za);jbwhLWq-+h|uDl-zLZWG=<RN`h`^|!(j`v);gic*a3Yk6F@&F@S+RF_JbdL(S<
z)OhNA@$ufRYqP3vn9YG)1+>lO!OMpxe+&Qgb~X^~Qu@g})~`ICvnR~o+KRfA{f%jb
zO`EYyOWeVcf@8WTk2c-PSlqdhzuPKUDS5Wfm=jA<o(|`Dyt!2nV(b)o{rIk!)#~21
z-AR)pbjuQYt|wczmc=Cq$8Nr)C$fx+FbsS%Q846o*Wl^=lBRCQ{4|%71F$Gc_F3DB
z9}+Gf=dH_9c^;}^-szX@dgs8ioky%M$j|QgnNS|Td(rEYy<s;`%0_y)zc`b)-qW=-
z-M&ooYT)hhne$H{UhAjuS(DD`pHbrS*ngn`vM3#Qbfr#`T<$$*I)8W3VfmwnO$2_a
zwbxvy?04H?bU5Y4soCHC-Kxo37tyKpjAgl1`MJ{`#g&yAw?yL}H4WboalPzb<T@KR
zhcD~fv4g*q-eq4kujRRk?byx~#}e-T@14v1MR{GZkDVHzpnpruYGVj)3_r6Xxc|3P
zE?4dk{vryS)-Wq>s)kCY*Qw~GYo80Qiuk6^V_(hOz1bi^<*yzMUA;p^MpyhiUG{Qa
zBKEp_Wbxbfi4GiKho0SOy>3#`q?Okz*5@CrlM^;idt26hMc5npgp=HO?RC3*@Y=aU
zJDPcuy@D*OCZzYwFS_TJT>G5AZl$a0`Be*!a1Xq><=9SrP?Xucyw1vZAocs|`&4-Q
zUt8SG7ChM(WIeFzW|fNi@P>D-bA7Jg@u-e<*#DyR<;_6_L4=d<@49luw!MYsL)oc+
z{CwptbzgpOVPl$^dunXbnotYGV!^AjxMyistGH;wL>3*%iByvqmK@%v_^0}6tIOit
zWtS+HkjJ8F7Of>QEO|C!X)6N_>WfA^wV!=&z(tC<|3c0}9f=|R^~>yA6AM(9aS?^-
zXil_>#1MLguHHqx6T6goh=b{PLA<fV5PD_3-N!-4XfG~Oz_cI7Pgi0HoiwM&y5snD
zKQ0=F>0nN<P+|zZxkf(9#9cp|i)1jJ$Vr?iF@)ZwgbX{2_pITfv6v3!gwhg2=-svP
zn@q}-6~`i3yz*JHvn(ZsB|iqRE_mg&!%%rtH~SbDjmQ0$a+YdI4C#OAQ+4Ir+hs~T
zq={*7HE%PCA#_{V)v491UOnU?IZTVxM0|-MbbIaE^ws4V?D6OuOed))X-EvAJ3M=a
zPu&fsc!-PXI6<7T#1OjI*70Dm*>uE1T9{s}w%9~sMrog}d2$N|!gxp?(=lo>e2F3S
zXQkJBJ)X+$<{~9b`*3`;C5F&3;SsSvD>9LTIFg7VdQo&ADD{CRSpj!NbKPT;omD{5
zq5ROPMD=gXwjGGdG4G@(I%q-=O~k~=G{~AkA&w!g#LW~f5{f1f;V3#r7&EE&e@mh0
zY@o7XJozd`9|7u!G%f+EL>f1ON+SrE_KjfK2vZr{4W8ZL!{8p6*#mP3R#g3MrrEC#
zR<YT13)r=QJt5fiO>nviGZ~zNJaUld;Y8kTr_0g4Y`PN7u0(SPR&B44`&e9SAtN4W
zehc$OHhl}MZ^0C1puNaxFPcd@vgtzPQHVSlz6v>1q3MJtI~O>T9b2An)EIttcMYv*
zf8Ft?^H;DtD!`xuY)F5#C1uBdoct(-P0>Pap+4zO(OTMCQfQNKlP>8((LaFj2N=a9
z7mVRZasU$%f%bv^Bqk|3d|J3K2?OC1{Tlg%t{Hebsw`EGB<EnGJ>&G5LEBc&S67AP
zBRoZMGGeA!f^15VGZBxXPb1aSXcEB&8*K|R8_X6WioO6k7ho16K{`@RN4BJ)$?}hO
z0<`g>ESq){x@nQ|*|fWk`y>W==y*sWPaRJw<fY>!h2DeydoU!QW78=}Ck0I+*!tV<
zdjXW$J}h6LT%eF>Y@=el{-sgcR{VN1GTDqANPD~JZ7YsF*(Q>u=<Q&)9Xtp@(V1YB
z33d$50nHpRV{k6;a$)qt6rBgUd0@rx#h_a(jnj};8Zsxk*ef_(4tCW|8bi@bO_q8~
zZSZj+R?G%(0OJOjM+S9DZ=N>rFg6c&+79~LVHzRCvsKB2$qdc`DhGrFJNu{?B@fFv
z%TRO;Sl56HAt=1bs21QexD`xVVN@TA&O!VfWK5<_;oYiQi3|vy{p%sSgNt2av3@~v
zLF1*CVI4AqSw<sJjWCh<NC)V4z_$$Uf*D=lR?uQvz1*&;jiL`C<AZ2)r&Dw-GOR_G
zq`jM~NNdqPREfQr0~R@8O$aZW?C6OmJF_qmJuTXW#P)(U6*is<M;7AO$B^<dG=<DQ
zU}T?pUwx|87bmUI=9-yndrimST#*a<xnM+QBo2!ydD4z5hAG9;I1O2*AsaHI<tu(Z
zzhLO6M|fZPkMrk~^PZvu#s%<6Iw3^@-m~Hgf$SoRt_P!f>9JeEsTI7*T*zSfd=gg<
zOt&U*Zh|1vZ1tW{edliXM66aJ;uj(VLJ)W2&x%V6rvEDQRj#CLqUZrI8URPqDdl{X
ze79Gm2#+=hQwG76tdtZZI30^C5i;5i_c5EFXMVQD$Dlx4V8~=$4Q&l66vvC>lQW;9
z<9YFXA__%s6mHZbpQq^Wgx^V_a!@D-9>Z3ELWMN01cgdzTm=eM(zqTJ>ZNfT$hCnY
z8E?m`ZuL8J|H#9O?E~#T7?m#Laiz%DsXy~ynv%4SeIhQnhzb-v1WH59{w43RSmMHq
zoLz_VjCztU#@l>x{Fbh-m+@Y6#FazQK9hV*NWp`H(AwXD#0^F+{$30Uunuq_A&CRj
zCBTb8=fL=!bn@GUrtU&c#P>K|p{beD_yn>!fo3rLNn~@9!PWg6s#VelPki;$kK`*X
zzqno{GE;PvdDI*d(sdDCEzX=9-Pn0a>?C|e|L0e4uQl$`w&4tUiarT)Cqb&zDg~-k
zI^A^w)hUhdfzmxtBT+-q4?y_=XfXIC$i0*v?+s9Ir2D@E>YX&sN6PtVwEr*H?k*~`
zd)@oh)oP6Sb70BG4c@?}3y@U-lb}gyk28a~u4iQQ7x-IB1&koTl7z_BwmZ$uPv4yX
z;?TWqub1pIkH7b;ueNnzl<K>LJRx@bR*>5Y<B8+%{2ss1NV91Z&e5+x{}os>VLJ%=
zgA68>`B;wRD#OT92ljO^o$+@En0A1<G`8$uPBrmyXmb!rmEyI?s+LjzuIaPRD*uo@
zxGKm1gAArZrsxBpb3l6FBG4@Y3-TfIA7>Tug^J9N<W)9Da)%<X#ifE#i}dl5gY0sU
zJG1&qWK)So6MV;;Lq03_rEmRW4|m(%p7pj=6~4NMNp*;KzxYbXEVKrw8YcOYD-~`g
zz8W&k&Vc<H#*O`R-p$(=xqldk#dh%84!%TQadGC5$&|>?KWIcfX<MKpDHh~%L78co
z2P%1>!{B02DVD}*NI4B@62)J<%%9O+qx1?#Xrfl44LMN<51nZ=E!Wch>Us@6?N0B#
z(p~(2m!Iz}MVIEBYI6Kn;&>lh?Dgqx$6LIDNI|3-$+++DZnC;FSkZuAmxDn$qw&Wj
zHHRFmCVs-i5NHj7K9P!?SU3ZT@8^t<0yP6|iJvKY`LyLe<j7^kgs)yE-qW!$VX?vD
v2NkjT=f@`xEJUw3J@RAltHdl>zJ@?WO<k3xK1#AQm`4HgsLDL5F^~TR_;xe_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_2d_multiple_reduction_threadwise.hpp.39A048A5A487AC02.idx b/.cache/clangd/index/gridwise_2d_multiple_reduction_threadwise.hpp.39A048A5A487AC02.idx
new file mode 100755
index 0000000000000000000000000000000000000000..efe8f75f84d24b0a81f780a07a82e6474e9482f7
GIT binary patch
literal 3626
zcmaJ^3pkYP7XH^DWBi1FFi0*lBA3v()r=%&s7$6wD%%$2QjIPy=|(r2q#OxdcDl+@
zNhMK}b}E!zZILd@-lv>Kv5QJK?WUaHob&b1^LU=~KhLb^d%yL4>sxDm?^^#qlBq(W
ztv-eY#7<qkG%htp6T>h=@|TjDyqM6yu$?%DtuFpGtYgM=H-q2H6q~f?&bH`tHaw|G
zt8mm5Fni0pCbm61xiBGPY==Pk^@;rc)DC7&^!$4_65^X<UyRZ=<+*CRcj+koCtmjM
zVRK}Q90#QKf2^6X<0D_=AG7G|l;%=p?EDl(6FZ~CsoP(C+JDl%9mX*?C%xXaQ@EvI
zZSSF7V;&u*UkbP|VeOOS^D{GgxmLSY>3zdFQ6IjSL=DRP8IgXox=PYZ-wXHbUpVs9
z^U-ygVNLR{#exPs?-SCMdSMF7h5(zn1J(lB&AIiy<2KbdD$ne-Zi8<>jkk2!SX;JD
zvlwUDJ$v_DYkjSW-@IqOlXfah4O|qSBHa(EmVXu8{Y5wYi8lW1v(&q(s46?jBAKY=
zhRotQohj)%-DYSQ*;6XmT3Aw<5wwjL+<HDN+^{yj_lVo2{3FsS<*bNDFT1a12FF$_
z3&w9~`yLPyylzX|fMgoK?s(7TwiRLC4bJ<cJ?iglbhL|H`EzO3(;Om;<KQy3HAa^B
ze2tMg*S+CX;phFS0WEqp?=Lo<+<jeC(>!y1d(?``zU1yKY2TJ7F}k5HM_bcw3tF5n
zI?K#2#SRLuJzVOw|FJ<$#h<*+oL2JJ_^tJv-1l6fG|xlt*H4#&+hZ5J>b&r+<Ijve
zD|wx}PgfXCif(uO7!cK)_0!|~7v#ACPyHp_gtFgVE3W*Zn;2JMzI(%pmP`D?8Tls5
z6!M*sAuSyvN1v61M_k`w|IK-|2j3wpEl$W@vH9Es{}We2Chr!<xR)PZP?bB&Nj}ZA
zOU`yYyRo5aW5{!#4T){XMP@TcBryd<{uGayKG$-ZkB0Cw1NScUniD1nJho4~x(O5c
zCr2B0W$n|n9iQv5!tU75>rdn^D&DuKo6q`9>vPjvW!{tt|8A-aJ@>l2d;09~y5ofj
zX8lEXnzO#J%lIMUh(7odGoY0oXRmepiKX(<2)>zPePL~4{_ed_ueO9Pxwh!m@q>@9
zJd%|is(JmOe|_=AOAeEy#pSV$u>$u~wlP!ZiR(6U=&LFVgPKgjKV8?2-#q4;U*f&X
zE?CN%r3>Z(hB3ofZ@#;kb?ef!3@{>!5oMwh7@e#i@?pNn{Cnnv(h4J*4ylf|h#svH
zNNw$_{~>H^*#cdf9;psvgqf=ZQjg2kf3be=;pxUSb5cFce43j|AT{5bb)@;6Pm3{4
zn^Z@dM~+enq@I_`&-IP-x}!_ONp&bA)J7$cI&$FF6}RO)S6v#NR7Wr(OjQD@W6rYP
zoZ2|WSeM2i)gg=!j!Gc)x|r&}-<_XZGJ-Z@Bu4mZ`*KwR)6pQ4f&l^TJ(Hc!oJiP<
zV`k(6GtA6JB`_UL8fKwSF8CKLlbqE0?xIW6B@dX%m}#XF=m0D^E<VLV2gCH-vVU1p
z<lcefgqRSUp{PjYD)yw1WU@qvOyWeORiqtC4^E`wsYd7~aAFhSn?Ma`1D>sp^T974
z1l0N>Ff9U0b<8b-2^2mLA?HCt;cA#t4Z#Q_f&fEB43sydCZ~s3fu>oHD(T9y3*$kP
z=z*Xf%8Yqr#qmN%#b!;Mu(Y+b|B+q7%GOE^#ahL3kZGKVw~GIPta|Kgbkd&r0Zy<m
z7DL5OQXmG>7%Ju=KX5`026C`Qwru^*%lmy=7{mN>$nZIocg1ZE4JY;hdk+Ml9XN3x
znD@a$ylCIZ#M?#(Tw;|xOIpTtElnc8D!>LEq)7y_1MMgzU<=fckS$b0LF}L(2$cru
z_fTCR12`eF6mb!Ey-l50zjcyrC1Y*|gJvom-tz2((w^+35jY{`NvBgjCpjhYD3oTL
zW=ElPPPzw$WWbSu8g2p07H~n)^7VZi<_kG%G;ku+B~+|NQ$vXdMbkoXE`(5I6cJX7
zp{iplrgCsC2Y0j#C(40a4xSWN09T=o2f=L+#v{Fd&h3ePPg4$%S0uHUPDIl<kz|_W
zh)|%!;#m48dLYTKBQW|1j6({XC;^ue@TKrmU_Au~6e&)81?E@aARN?I%-bI!NYN$>
zlrA<>i`-M<M@3!)coCRTM$3Uyu6}40_*KEwSJq4JN=6EKIH3Rs1r-ub41(PtxT5_c
z?_JAHTz2dvFFImkgaEa+@THaBA^Jz>kkfr&&<Ep@lgKNQJwttw%Jgja*}<qJkQEJv
zsxeZF`hnXIlaVc)Py$m4HV7l~k+z4rCbg{?0*fIExxC<Z&6H(bt{P;*)9lln&@@h@
zd!?)0kqo?KpoUw(wFS87tk{GzSHnO4Q$YqOG8BzMM<*66un+8<wu+qY0ag$EXadL?
zvg-`B?6U`*>f$ct*KEF8rPC=|oJ6i~1g}OABKzB?X=L^04f>GO$pig7Fs9rr1$rrr
zM0mTj)1r08lLMp$IoQg<0V#63&n3<Bf0Z}Pl~b#0wZu!03Q3-bJVn09brK0gA_^6N
zbAfu9wS#9n_#*#s;xTwWR>$vw`yP0d{sVA7P`LUxW>Bly(~qP#dElH!b*t)EYpYFj
z`cDq)Zx`dA-3ese>LhnWYethF$^U(7`8I}%>`%FLoX7!sj{56kJK)>ZnSC1Y)9UyP
z7@Prflpma^1ZE|Srf@yb>(%Yu1pKD@{+D3zlG=Z{>A}1T-utd$j@KnW?~7Ef6_bi?
zaO<XO=PrLtx+HYe?qNGxAFiC^WPM`;dD#tM-vFLe78}97k-`sLUw2f^jxHdZ+j58H
zUX<%=*lX0fR3>mTsbkR_i=^Jr#{i-p>Ih_!4SfP2g5pHDd$<r)B2HWbr)w}7oz&@b
zH{kOr77PpFRq#OQ-o^Vq4YfLvbQW4qwmRRWA8)4Kzeg|WxfjkJsyuoclUkH+n$D(>
z3`}L9hFid(1xBHqym@=jud~MB9m$I*n<#JemhUf?*SVgx;SBSv|F6!RuEx%?|J#HI
zxJmWFr*BLic4uEw%0%BEw0|Z~3$qM!LwO;;krHEu2$_<+FlGMGVCZk#{~i-Gz%XNK
i$fAa()X<0;GJgzdgyZ^cW0{|WZM98UriM&oBkVuKpJv?v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_2d_reduction_multiblock.hpp.6BD3600B2B4E50BE.idx b/.cache/clangd/index/gridwise_2d_reduction_multiblock.hpp.6BD3600B2B4E50BE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f580fe9d69cef278bbafc070f0f957ea1157c43
GIT binary patch
literal 9224
zcmds+dt6P+9>Dj^n%H)$?9xt_9->sYBJJL#L`5a?IHbdIy&@tNdBjnXl7r$#C&^J!
zNJk<SDI7vadZ35&JRwDNLJ{sNO*8-9fA6tBnYF%tGqYy><~P5Y-*3%!c5)JE2?SG?
zJ8xdQV#5+75C|0be*qi(y^|P$z*1fy*c_iX&&~IVg}8ZUsb+ZKlC+m5-d83o{PtAo
zic_3hDY`N>^we0hC3}>*8qDNA)*VUKjlS^LV{lNt(cy2e-M&r-9@zUHc>O%=+M3`)
z+=YBC{d^;t<0nk!O0VZXxRCUzX{nw=%%TxbRr2b!x4mmPvbM|jf{ez(^|Cwcf&w?}
znC*T&rdYYg)Ml{bTkUIiR0{P!{o*=2_3iCVmU1~+LDP$Om3pT?*!F1CXLozKh1RF$
z%S~&H&3!Fy@-VsfVx#WXsV&PTCpMgEbGu$M&0}SeOUb_G(bIQIJ`Ga;k`$hkv9sRH
zIa6d*e&W*1_<Y^AyU!gCjl9zI%iQ~M7YnW)KQ(p7nlUP_?w-bHG6y|5ot11HJU_<c
z;F)t1U#~ITb3Eqq^zx1)%na9(kTT~bi*F2ll2c)PD!{WRKH{vZ(U$$Y%k~Bd<65J4
z%`I6pR6Ahv3$0g2+(j;~1#@HV7qx^a$p7o5{SMQ_;84$l#dC~h6P>rTyJa3+@WwSY
z>}o(~(3_nxaY<%iV`RsNw}+JE7AVMl71!Q9Ij1!!rMz{<t$9MV^;<FzPr7HPpOp6V
zvfIwb{W5f#He4>;cuUuz=h3=wwVsU!#&=yx{t6#WPRM+J??K)46B!n@Q*;zh4(nm_
zp8BrI^r<MFynL0{$_Mft84Ch~!!|mH=gvQ}Birx(h3yHF=+A+sMNUo<J-a^=mp)?d
zd*3OjDl-U5)x0ln5m*tKpyRS9srZ5(<B~X}O3TXPXYovJ6Cd?xt!quvo%7~|su&+F
z-kkdTF4t%)m;cOe*PdV!7#*?i^Rl@W);CjE!x6c+mT&Vo-zr{t=k?u(r#g~G+s*J*
z40cKgSrfEkS6+LTgLFq&Nu!MCpQSSIZ<)Wh&AI);|9SVYS@()H|6INO(QglSUpf9Y
z_NSnX-va;m@a(;EjqYKI?nZ6bs8#1PH_cYO7CPjwF>hji(oabfT%1z5<;O*GiGg;<
z<fqt-II8@kgSlgaoqSb@Ti&Ie%3BhCca-b)j6SHFb${er8LhGLb2mkvj*rY$h%`9&
z+-H)1s{d{86)tyL##R)>L_NM4`ljQDW&U-xM?P0w-u=mLZtLRxmF=ykww!6&nG=}r
zx4tT><3BoAeC5729IzSZS7TrEX^vvaPm)7tEAsyRGYHwNBJX`~1bF;@v~JLk>U%;n
z>UK@9)Bd}<^J8i4+wAQNby^HwWIVRCO?#TQ>c}JIi=m+p?`jF3>o<H}TFuFFwjXz_
zD4BO&@$*nLX=bV*V9VO2eo!Eou*(m6mE2q*3zdcoxUHhCTwfs&^6Q!ZO@DNH-`M>}
zJY?Y@UK}6_7|~Z?vDcJMZ)=pF`^rKEUc5}SY;<3N#euDQi;Jf=oREczym-B6y+&Vw
z#lc^F7d}kgvttk(tSsO>2Yc%G6@np%AD94eAS8H%#zV#GHVG8+?Lz_or`mtPee8n9
zF89pp+gZUVz7I~!>Gl<ZAq)_Vk??)&2|kfKcE<JFvQUQaBLavbfO7u@_YrTbxHdEU
z#e4}I!dKY|_&?SD3$AjkGUiKD@VqWrDCVoI60I83S734O^x=WqZ@W3k!lAr)g=mFV
zUxCF1o-%&Zqi&Z;pe8R~!YwiCE3o+Hx%swAn{0oOg~NF9I?+1Kz5<IMO1|!Tu(pen
zKoKwA4b*l6gZ>My(zwB_D<>{ti6WHat3-m4kwCrwf~&l7zj-4nI?!Da%JWsu!;$Bq
zPX7g0c^5Ve&GmZqhXju1tE}PH81@xd{BfnHM`^b8A_<h`#bH1#j339h3$F5I%;CB`
z4bSzmkmaic15q#--haVW1pX_$0%oZR1hOlOvR@efk_lKYKsCS;Tbi6vMBM6_2~QMR
z&Ucva2&@~P#`~HQ@|)!6gw2cPZ9*dyXk&9>`N`AR3AFLNvRp10o(t5lmU>y53oHp&
z1CwfCMfeFYc|vnH9NrDpunJkO8%n#OCE+~AHIJD^IG=INr@4{&S0iIY<j)wJXUt^6
z&l#KNG<Ptr9d!Fn#<i31P?d+1jmm4+$Re%~SQi2tJaEL_0rq#mWWwQaYB-!mI0;Tk
zf)0ek8K-b&Djo>p5_z}GxDsw+bekAGnrAgJvk51o8OdlS;R57RfT;Fm$fb<1N=1gS
zC*`uSeDA<#6@uAhAa{Z3U0_By0#1y8wuGahM>JeOI2l?d!+{w?Tm)ko!B~-j>|spz
zFlL04n2AY@E#YLwGMTR5#Mm}5c0_&>O+1OH`sv6c9W5a8>&WstqMlcdEXxUx7;1BO
z(y4-5kmb^VZW=Jh182E*sNW6;8fvuYWDN<dx8oN9C*VYVK9?^>+g;4*uqY9I_lnZp
zZgKp;HKX}5EO!-*xeBHbMfCus9$-Q-WP6|io&lCS!ibO1YwA~|^eZyKa>M04)@pG3
z4TQKhFs2P?<Bky52{bx^CgBJ)CIV^WhKP$s8qr7-Gu|MfsBdG$dlt*p0k#gP;%-dG
zHPm>pMzQ=v`UrhAuq!nmZm{`{{_b$z5Z(j5_q5MBgv5u?DEt^}--wkbTO#}vSS}7&
z#{pN|G0P<YvjpHoI1!9W1V)6Dfh3u{AIqfxNeazfK&=ZHlKN%Px(p7iQI>lQtsm1|
z3C${@6CN1LRYS9Cnrq>xT1YvwdN`_{=0+%Kgai6yxn?M7rnwbLT4`>Bk~W%SkXj5f
z#Lu^ppZLdOC#@}r<<=RlbEkFy1wRwA1LOdM959o-WEV8(f&-4pKC;N_bzQ)*A-&qw
ziKAV60vMeDOi9B;z$O9>!evmWjI7sQZRtFwxe|`9gr<13z1kXGO>-?|YoP{_*F(0R
zu*2+;$=luw9axsj0HzthkvzB%u!Y3FSgr^d7J-RG9u8H)p)TPhsFVb?2zNl!4(Lc8
z9L|WtnE`uaxhTdkikV2{iHuSrqeVCwsU#y^!UagF08#QXq*O+DYI(f%QBTJJ-hO?B
zzKYZ|nflq0Y03p`E>I?eEC*WUU>xBpFtZ9Q(3hF~^6>UTcb3}+b@xGIB2R_dsc<~u
zZpd~+WirS##vqNjeO{@IK?cpaj9o6{JlFQ{xkAH@b6Bp5F{q;Db&OpdEpKP++G#$6
zbk87T^1KS9SAnF2Ymi<IVP_{l%>}XW0>g5Nz&MdMvs!3eOPfhOG_EI%4@11u={<83
zr_%zs7NCdU-_87>ud?AUQM}U$23Enq5ud-@W||Mb;}|=OcRN?njH`$`niQiM#f0ar
z%`l%XpE#c%;bAB~OshH*i8JX}y@kZL2zz{ro$&0+PLqG!8Dwcu*5q}~;~62#B?Fyg
zppTW=>o|2%Xzl{zyMUC)8=+1k)F+eO40W1mZiPCnG`B&WHkxD5_!wdw%UAwhFz?-|
zzxX-wQt(p6Aq2~LD|oBocmoGmjHoZLB8K2I%Etn4GAy^zYU6BdGu{<GYyC@ZxAOho
zV8-2G2DU_&`@oF*K-lNdQXPi|mrZ=dXs8zrDJPhY^wQCQQ^U&_iTc7fGKg&xwmISb
zk>#EN{byhjQCAq!4?~l%x^OEX>I?GDC+O#|$c@gzXLhluKd{5~SS}OTWP*VsAj_45
z@#SCw>AxIE%jqSx4@&pZFOdq3QfY(kgf^XU;7G@x%oy`DhC0NjG1h6cL(E{zGiZ4R
zW1T_Es~GbtT3*FiSJCn_NP31Sk>x6oNd+1>lCoS4GN~bq16f`_y-_UI4-Q}X|Na)c
zV*TuOVRpSb@YRBMxaxZ9Mz}f4Y3ONC6b2+=Koy5>EFU*Yb^}!$QnK8Bps^o}B}zR2
zG!D>w7Ra9k5+Y9n@@X`m1M=r+P6zVoG-m<%ESm2F;e9Za^mh^JUWDVZk+9rrDE}Ht
z2)9G|cA5_|!h_6EA}?jwQf3(Dy=xjqmd>a?#yiA1pk4<Cw!?idF8Ix{Z$0ooZ5_vq
z{DS+I2*CPgxr@N&B5jZFpv^mCPw~g}8@JuqX#S6Lh}hNetGhR^LEhKL#^(p(VgjuK
z=U^AZa@V2sI-HKxp6KUpmhvZmf?>Hl$mYR;4Ki)>6&d%niB5blzf60XA@*jx$20Ys
zNIagGqn8^YZ&UAycu>4Ibo6#3lno5B>7XDOng&DaP!s}9Luft<rAMJPF}HXqji<Q)
zN(-Pheh=P5LW2*G@|eev=`lndk`j<<0?pY-nvJOTIY^pA^F1WJM{^O977@PGy7kiV
z$c{7LeVCW+RsW}h=w1WA{2~F01pEhrZ$Fk#L80?-fUy<u#@L%hBgS|gI9vxmk%<gp
z3`0obg5~0w3GpOxam7Yc;jf7I`{Y^fG!UK!ilpIVz&@r=pp`&e2{<Bu0faB;`mI3N
zN^=wxM?tFoUPic=Q6v_0kr7{HIKtVCQZ}PbxP}qe(DJ8@(o<Ugicxw+b2t))BdY&%
zNPG?r^ndlvqXp-z|8D;7Ehyi=>Gi^%t?rPI&*9i;68Io49gR&VfeOyG@(#Q=(Ml}H
zOU+A)&5L&{YTi=pR&cPwpSF9GucS%PgrI5od{bCfT2%D1v)~`+-yhhF-8Ors>f!I+
zL`vqbR_O{nroqQDu|OCL6p53$4~6$3#l=uqOt{qLU5U+}xkLHw>lPe$i|lLom?R_W
z&$koP$O8IVbQ+-u*cSnJ;zptv`zXepEc;-j5sb!RuI!ATy(%^`{2ymb4Hgfkd`MFP
z-@P&6ndHyi-#(|{sc)UuH8a-x*Y|^ZQ%P=dU=FU?o2Z+X0A>&_gKlMT4xaPg)REg`
znk(V7N;reatKqb2!mmE6uWl)uH>NL86($2E(s2q9rqJ94#9ef%s1XVqp%Ur1848<e
zZiT{Dn%khTjpi66jzOxpefygPyM`?BTYh;&0L=(sj$;yBf!EQu?@3rk9q~E&<Iihq
zefL(<@v<SZ`Bp>9|6N`YpWFFeu=hJ2o|xV&Bbz~|H#;EPLDG?3Qoozi(r-`tpSx^#
zK#JR!N0lFVo7k$h)d+8r-H&#dly+oX<Ij0#f$mv4T6zujUqgFrGB_X|BI^HwNJdx)
z^b2Xv8V=3FA(j44g3=^7@QI7%!Wpx0hRSm$GFFL<Gky@C<w3T|$c=CTvMNB7ybM{D
z5&rb)RQ~vy!=3yaTt(8WXduR6xnd+OCfu_yD}JJh@)uqMEkM78BqneUZm_8T%O`oY
zI4F&SREFdvR5}R<zC^IxWhA{!GA4Zf1}WVjUm{rUE|T7*<qwh4Ln7zDtokhJ-SYzc
zmjx{Dy;12czWw+QuK1^lZv+Aja;rXY3%LjW`iV6P7GSWNhPui~)e(Xb146+_a{FHs
C5V5WR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_2d_reduction_threadwise.hpp.324930ADA73F028A.idx b/.cache/clangd/index/gridwise_2d_reduction_threadwise.hpp.324930ADA73F028A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..197056666708c5a14f9f1e9cd82732be7ec35d17
GIT binary patch
literal 5210
zcmcJTdpwod8proqYs<F1v0u5Sh{2ZImP^XkpmOOVGDDXUM~!B>QIx42Gs3}av>Vl+
zsKlTq>4ubUR3l1iZ*x?XtyGkmj!H^3=e6g&t@YRO*Lgo5%lq|Px93@FJ?nV|&++vQ
z*5z<!eLZLMx~PqjTn>jrGe2<~V^=vK4yTCXa5fj53;8PWo}+F{;5DS}7^2y8#;5$Y
zAR{DW8)u#D=Vt}gq2t6S3Z2r1{yaHR7=PDM%H4Y4;@g&<m$eDm3##{}9g^5>Q^Iad
z@002_O8pa3>!v#|C*7CaohK}K^6yQ*G|kOkGUh_XmJ3rKXkLG`E3@7_R~qpsAvx|@
zY|z<+-{u{An%7I+Ebgx;YuL5kD#T-D>0=}R<fE-o^NsfO{<Ux=rM9%)&aY+BxEC!`
zD!+WXW`^wEktfr2FYelMeUk><`&-`guKtYNsZTRC-}(B-=vtMCp07EKT8`DfIkR<2
zOY`*VL1py3gFc>Z`l(!hOY4OjessyVY_auneo(1AmfhAVu`d{3wqPhcUC!NAd}@E+
z{SVHCy^q}sQ(kQ{w>r`AwXn}&_n#^EO7%7TYqM7ubxn-knVbMDHBHj~aXK(Mu|C-(
z-dLU--}|AXGn$(}T-{JHvvJtR?VQnycr`TxS=#rS`%WHqR7ei2w0&;W$qVl`3Vnf2
zuDrkLNl3z;xD%&~e>h;Y8HIJ;RO@=~P{#H6zO_5|+)!ku>`vh%&u!KYE#D?aH57lY
zP&chPW_0XJ@VRtBxc*-6y2H*b1Id@Co%+?GxIa5JN)U5ka<;rQ_V!TBQHzw#0kacs
z)ikEdB*p7D8W~<Qyd0yKd1*}4qNwG<cJ5ya?+&5XHrGkUi^V<S!SxzvuFcSCU*vLN
z=i!)=Wh+0J#GKnw^SiQ2ySqyKrnF7-O=I7>ttDIA&J9Li?6Z^|ta|Bl1x`LHn|J5%
zn9QqZZ;kyq_+?ky+V7Y7*62BVUMSylugJ&sihQ-Z+AN>%=#n~pBc6KOXSc3Ai`;Z%
z&syh>jFRRIyW5KI*UnsUcEKvGRXG-M;a*O;(;9LdcIJ2{dF3hJFAUBXgzpOLo}=|G
zNcP?r<U3%Kl4_@;Hb1sVbZ4w6e|JcCbluzRj^pCon=eniFK_a#d!BAzo7=0oc6^uV
z$Dg=;Hb08Q6?q?KeT;zT4P&Ekhk@Y1U%8jX3(IwH9Foj;O1#+Ue@Xo0b)w>rD`)##
zBoa@n50)W%AxHoDq(Ry??%``eUy@{@nPR(T*^*ZVla5!-i`%kpxhddqT(pKlzv3J*
z(}bGZ96FpIZmKFcJmzFE&zy#)d*_0B2Q?wZ$YK02GgZOLL0XOO8XNyiq&S9J9NJi5
zY^^Fd8VGQV#sV5F@w>XqJF~h<2Wl|w`+;#k5dQZU+eZxKt9fTlwAO@trpglj5(`zq
z%JJL!1O9zy{3%VS&d8DcNCQ>D%3EsHuX*_P#p^&GBS#rWnWzd@R@AQBp0s;a(io`5
zOfXnI*j80=G`Ntlj+x+t*E#{q6w?;#!7)tx72{V-P!()@PHfc5xV73Gj^>k+6|Una
zFNPGYgtkgJg|&>LlaXyQ`ov=$Gxn_BJ>?!v(aXjyGh!XckZ74JrU<VHU)J{&y~1>b
zB|8&}P6eK+z=ySrqBDR&2AD{)a|W1>xg5BZ19!}afXgA04?xcY&<9Vr1R9jUi6lFh
z!0DJfphpLsiMbPcbdo#_ZHA#E);~ax52XGPdVIvJZK#;(D8Ii3a_M0>XBf_98*=F#
zC};;-7~h(j-dODCY51?>y!T}_PE17w2vLAg)*>!li$ZEqDBI9T+vA#7>BTF|T1bJF
z6u7b#DY_I|l|omxqG_+NS|h&86Hv4<$C$6$(M)0Pd8&3`S}T9lfvgj!?U$W*T%<<P
z*?`Ihc6gxkKzJVbVLk+@L(q;Ln4;^TppNuWF%lLdKUQZ~UX8ChZR{?k==HAa=i*M!
z0NXRb3-d7W8wO!)CltLO+3rVPEV~)lZGK^!md?1P4p4Pq0&ZA@s3J6httgJk_~w`1
zj15{89WRcbhpS1PBr`E3>LprYN&<RGfZ!%z)C4ADJq3zW;5>E+icW{(bdsfTsucQS
zJqu3FB3TCYWY7}pIZ!W$<UFXCM{+*Y%O|-K8C9amY<mx_8J8n`Ew*qeI?O&Sh&Xq2
zgbV9^W|F`z9Ry+{ilJRG3}id;mNvWp@gS~n4C5zZJaNDzCXy+*Q3@2MKzFv$h~>g`
zlBLjGiuWW%XF>BU%(H`yOA-covo$F?7ntV)C){y9u+IlGFgF6TMnDYQ1E?Nyl}|$R
zlh6rIqzc+s!5NsZLaVEAD&`%?a0jwxj~_)>AhQY_$SArQ37XL)T)zVeIxzdm3&eYu
z`osyC(8gP)L%7&A!V|kV15g=29~)5wtcyq^%RpcmSW@hg`$E}g<weo0z@im6;QGD5
zw3pmxC6FqC`gl?|pxq79C^@u~lYAF?-i32+_jJFidG|Swq6eVe0I9!)o^MIL3RzSk
z2i$)bGVQ|qilQGP(}$Q@s}WyyFXDq7;~Fy`8}SbKy#wKFrvYv`G5YpdJD9K(=^p9B
z-f-+?3izr^^w>Vz3sGl&st#Ru+-<(r!9|Fo+kr(p@WtJpN7Q+8t~U{N6Z7KXyeW^*
zrnrn6c-eU|1OI(Pvt}%BuAUj)BSMVZ5`~Gjc;F-;OacTq0ox`Z!nzC!Wzd%Gm!fl^
zFo)zkD9j@{9}4qHu0*z#NW`}P`kTs&p|49HG5%UgFP*?93~P5T7tU7B{f>!u3Sg-q
zPev`Wti?N-ZNTNLG7Hu^W?H~LhYaFJq5V<#DM(ZFB{=00j$!ONsq<AC2=1l>_;r9U
zvGtDu^%xjou1BJJ<iT1)(btjSIwJ00ITFc9y%PyKvCj6xtj<V!!w#D{7p3;g{+h<z
zO;W&<0&TV)MehZ?y(E`HUMbYZ<JUl54auh=?=<GjHA>@K3+@&&-mU;96~Kb+IQ!{|
z4PSUala1={f<oKwg!)Zxqi(z-o6o8Rtvl$;TtrKqmb$a+#L&#8b1|g>yENd!PJ^QB
zfqy*+!rTD-8%XX1ew}1gzYqNGlY9``9fU5}A7#kD3<Y5>NB-p`Hy}|1a>M#%B)Uwp
z0*MqPw;@p*=Bj6sDg)`j;Zd`V(<{5K&*nEN8UIFcqqNv;g<=jRw;z1svmR#jqm04q
zu}tO(c*#HqJE8<oC1k8E1%gsA9_yz7?-W^I4k$TUe=p?ig*w>K-=W}lI39C5)NO|*
zm|sJ|Yf^s$b>EPBHWFkbqJ2K%<s+j1(}+5aKE?97w5NrMy{XO2#v1}QL*P@gP}kVD
zq(c0%XH*Yfh<*4+*n3Hks-n39Z{3j$(!3T~uv7o-_O0gTzE>AV-Jv%<tP+9)C;U9>
z(Co2fPU*;31{=KFUljy@n<q^lb)b6%das_ZKEz{;-;RXa(G2!lW<LwG_^J;Rb{vW>
z0?tKb>RSZdipZ~mRv>C6V`49`=_NDR8_?|rnf2d*;u~bPAct;pQkO%qoYV)P+W@H#
zK=A;nS0Paqj=>b&g>1SIks&`sHV-i?r`??O=7=E})zQ!VhXV6f<qwV|WdZSD;LR<W
zEg@3p6c~^KzhKvUBq<9>Cs_*prEme(v!H(#=Kgmkt9x&Uel~LP@e+0MAW49i1PE>d
zR1+}7x(xDUP#1Tb19>?l=RsZ`$@!3%PjV%qDv=@EexRWCg6No;#i&jh=u=92t|-g?
zubJwQ^M(fhqgTCDv-w}A>up?Fz{j@sH_Tp+HjDlgxZk$@;Bsr=_<1J2NrA5vgkiUp
zLf=yO>BA;AYHj35^oxsK5f?G?LLXl>xz*>@FJ|hxdItJB;{+VNkw5?ajALIo90UBS
Tk6)Sp8M2Rnec^C)@aum7Bk$77

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp.E109D2D6AEE43DDD.idx b/.cache/clangd/index/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp.E109D2D6AEE43DDD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..025c0c8040e36d1fd5f6eb95d3cd6e9d479ae66d
GIT binary patch
literal 15548
zcmd^md039w|9|J)=f07<#iOTvCn4&IQn`~PDU=c_`!3tq#tdTyElAN~OG;8HV=bv{
zS&E3z5Lu%VGYqmbsNZ?2^PDrk?;oG*dwo9F_5EDexMt3IUf%C>?z6nl`@EN9gpZfk
z5mz;}fzy1J&YS-IWTd91)&c*&`1?h((s?yCD<d_vr72f}p3kp!*1op0c<r|+P3U75
znlG!EptWSz%@OZkCQp}f9hQ~-aQwx=4=Ydmq^#X>&SC$hKiuNal+QO!HMl00J{{iw
zO1qUYK2hBDFE`>8JFl*IzkKJ4*x8RI84A2mwzST0{;6u~dk2R1${ICu#>f-te)E5k
z#mBGwsJ|?#-TC`5#p3q?VKc^5?9MD(cqGThC-K)o<Ae4b)19Vi9kakEJnGbgNwGtp
z%-e>FKJPG_m-Db#_wb0N**{h;p5JzH#k&<zTN@mBuju#t&y}m?h0lyNxcv0Du*}=n
zhL+0?psLiRyM|8a@an)u^D9rkb+Q^;yyM&Ux*Zo&mj#5+Sn=5SR&`v;!jN8D<05Aa
zJ9~6*<=*wrH!eJ+@G9Au@YCeIQ#u9K-<meqO0W3i`fr{`oeYCk*nGb)Jz2eE^wqa<
z&)dIQy=eQ!9y?4!&wMsI{Wv=FZB<BBe#EU$b-P9y-!a(PRrAi)itK(?w|9JLR~Na$
zswVbT$bs3VGpuX6@y}vi{S*ByoUWH0_?2^tR&)v5m(wTb+PRtzMXzDlk#`sKBYqiH
z^Vq}l(L9aDHLnd%^?h@GTE{^v=6Cn&8uH0YzvgY%^n-d+{IbS6e=hVKclP!4E7yY(
zADmhgofzKp$LNon(yEWXUSeJ2CELHt@1Rw?(MviR20OO9<sCP7Xt?98M`Jho8{awX
zDVt$!b@uR_t<jyMs;8A5c<Acb{?>G-KV}#nJQ6-{+0m^(mA&@7UDBzdVA$;}zenEP
z{|rb~f15e(cHQxwjs`z0*%DJ`>)3B;SzUF3(W5P<J$4^+w^@5&_p&7W0G9<7-B)(&
zefHGcqiJhS>Ym#E?&jk;I_Kxi9De_t^^u^)zuuq!H16Sszlv_<@A0$!8d_oH@!YiY
z>#;%RA*X)%lHhp6)=ieS<mn$DSIoSZU^3^W%i+K)^@A3E?mG6)?7xaKz6?8W>y#8d
za_i>IOrPWPH&$-g?KS_b-q`5JzMH1>@&4?W{!_MgrcJk9ccb1v80dGi-PplS_Qx*F
zwyJ#6ogaE{RNSQ5Rb$uAeQ5Jn(TzUQ$3HAejGyx<^7BpKVI3={CP$1MZxa4jbwS#|
z0<CP<Jv*EtF^6%t4K+>~cNdyf#Dv!EeF$R1-X*-~^)aU6hty9R-$pEud_Q;DzvR?q
z-@iI_S$<<~VN&(0Sy{_YfBgOYxsM)ihINftXVug=#XUJ9ZkD^9X_nTftt-NQbXmMd
zFFUJx$%AcEANu_4T|9AH&s@Fgm@^U9r9Tf(;okjp?fs|Zg#!zmUoR^)y|UBQ8^msq
ziU%!DSRcN8chRLq&lG8I%ZD^1A-iuwo?A!kj2vCOa)iG7(xMBSk4_oC%A&~P-DGKb
z-8t{6+rKmEU4G+nW~c7o%)9m+8Qfz)H*4Rmx%!6fZv?MjfBtrGaA|QLUA@xU>%q00
zOF!RVM=lpe_|5*_dA#6kb?K+c#@j<(JU+_j-ikbsCQk~rDk=Z&Zh76^M}9`rgKPRO
zy}$JC@?r0|g^Klk?gTveKA?E5h5DmeTc_zxU7HqqV)dGUPQ8LRYL9$s+xcLw{q%lD
z6&;Ij?EgJF-Eh>+%z(<hgAbL?%2{>JGxwq{eBE!=gONYCe;YJFky0{kW545-H9r21
zj^@W#5Azcu&bp~zoOCTGCh23ZsEk*Umc~<>HY}U_$a8?P{mHP!#%(p$^>6UFFnnCE
z1)7|nhv4|=Nxj;EIoj@5e(L%vK4j?n&{ZGKt?k&D8&~%>_&`E>)Ah;_O+9(j1mT3U
z!L;2Lsk2glv%T_i-}Axt(&>=Y2NREE1jsUsFZ`_aEahn@t)k${uPdDwEu1ho>Qm5|
zlL1>d$QJosmX6k0YDNvdSTl*|o!w59jjtYC;pLzguzV1(%nn^-H*>>@XG3Q6yZU42
zbLIwD*G^n^^HkTOuUkG>j9F>+*aO~KZPkP41KNJ$PaJ!xcG~jc^eHd(o8Y0}KDjPl
zHgAfWpr+PdQ4p0p?Z!Sil-sLGBMnCyHy3Jcu{Otl#Da?bc6tY9%Ap}9hZu$!Hy4C#
zsAKW{k(kq$WzYnZw?niYnzUAgrSGrKlSU`Kx!WFg#w;_`XBadWgzPTgq24fABU%ov
zd#Ooz&@>M^w^k~a-b-FyTjsD=4!dENlhE`e?A=-s7N6^1w60fOJSc;j_#-=k?M^VR
zwIVG3HwSDQG<d{BIqZUaISx&a!(OeGilugeme(u2D>7(`dzpvkb#E>Rd6d<+WZxyr
zcgvwI?%^ynJqs1B6=4aQd9YF#|8BS(_Qx!_&@>m?v{r;A`0001{KN3Ja%hEF_CnLW
z(4n;=ER&}tZmPZ7=d=u(;c@(ke(c^{5c2e&c&~}SnEA_KdrZCv&_&RpwIVEY@2)*N
zX0e)@92#JjV8dYJ=7NwH{I=c3xo<(b3|e4v6hKj+duv5l7H51Y=Y4}G%b+@DX#}>7
zz`L~~EKBXGm(5H|oGXW2@k|av(}U2lwIVFbYMf5_=4V&Rp&e#94NXtOk*yVBS-H*o
zm(Y}&VRG0Lv*bh5d@KMh6=8|kcfGewXxk-n*d4PJLeoO%(pnLg=;0}TUpub|l0i#6
z+)`*%3Qb!p!V<S<PS;ZjhKJh2c9<m%s;5DzwIVF>UEb_5GI%sZ4%=du$%d0nnhO>A
zs^0A6q3Uh|lqu9Czczk0&4pS!2=Gw>64=r%!g$&#-xN7)gWFFuoM_TqkoJjpyghu1
z)7!}*hshHRCzv!BguHzUuj4e%B}5MOF*(RE$hf&6<dkEs#xV7EqZ|sDJk4;LX>&oy
zna9f~uM>7IkV7L(9%DGhxVa$YtV7ACw901pltUd%4m1ojZY~J<WT2Ztp;hfHIYgK|
z$#9ZMb3w>OTEXe3Gz0zRkjLbyhEq+O3qmd{9g>o=USp;VcEIF9U|$G^wN`}X`qkN<
zon!Xo39u7p$%gH-p+Rd!SZ+oP&srRC$y^4_F-sUgVOT0#DizCh3*C&wQ>HR#j9Jz|
zv<6E;OGQ}jjVus5==}Ol4s|ihD8o_4%>^MpoM8X_(fT%$3~FI=BCt&aV_Pc~%Zjr(
zx0@!PZ4WzQmQeLj!{&mJf8VEfb5l3p{c>o7rExzr-H#7&OGQ|oeMBSr?96D_9*Ve!
zOsJj-%~~tM^2)P%R8zmQlQO7{S@r<52e`LZgyr)&pXI+zv2KvV9{58ipy>&$v05s^
z@~z}<pEi+pwQ{J3dl_vw+PJw;k!xPxd$#VyDjC$k<Tzj(hbP-o5tc>MXDs$nsHq7j
z9L-$^9W3HRDHaZlg%fVJF10AdAp00JjGQo0YT`Yb`0)gj3&cRZ`QnIgJtLyD8Tf1l
zL%+2SL6lN}PYM`H)?bt|K|m%5B$`{44gsS>z=W)eDCGj9TozvkPS=48QGTLy132Aa
z@lDY0Cg@M;RUoJegiyQ=+OLDX>9k(MzOSJxnT9BR<@~;K{uHMmuN34>aT*$u#x6r6
z>d}a7DZPOYY~V*xyiM@gCJd#xP;f02+$cUOxE^Kkal!RCi%$x!Cs|x7xR(kZwEYFa
z{Q`^Y1@C%ch>G5NaU=ES3!83B81f54%rv*6z^#ZG_Ip0?J)MSzB5M2WE%x`!L@CfL
z(28ggQ5tDB(u!=oD2*~3Wkn92D2+B7ZABzgl!DBHtcYq8rC_sQD{?!DQixfI6)`bH
zX^hzzE273kX@dELE<|IB(nRx#U5I)WrAg+Kx)7BmN|VhecOj}<l%`rtwI=#Xl%`os
zvnHBTlr{n5O~8`M<Yu6^8JJL<42+Y3C6&<>pqBznD9!*T8K4uD*-W692}~%?0Y*8%
zg39n=V04(p@o+#q986^}0S-uDaT3x?LMEy`l#-BJ67pp6=p;0b;vC*3hwo2u8E;(1
zTT&b;=tl~s6vqhqF%)a$?>XI-Ur?kWN^{X%vDx5L+I-Rv=m!R^0%#Rr@LGV@vN#T)
zI2I=Yl*r<p0PSRPHbB`dE(EBM#TNj&z~V-L8d<y=qScTYPZC5)6l>{MdMIvG&*71D
z42(PmMv==2NkyP{5pbaR9@pa@*OROck{)tB9<ul*>TwgP==xQ;ZNFw0_R>aJwe{N!
z+=$p9tlIje05>8g2&=Y%nZT2X1(J?~QOCh(`r#rlr3g%=_zbW+0~8cj0;v+X64iyI
zm(cAcRBbboUPHInEN<ZJ8@OI{B1y<Q3Hh+tHwpPsoQArjA!~|1AiWPrpW+<eBZnVI
zaT)Jb#t$agK+CnvJ?qSQAWF5+vla#t0`Uf*daDmWwa#v6z8kubd(q_O4D%JEC*<HB
zt_gjv2_d8h)2fi7SEmn`;aeydI^;qpVo)owTP_?xaS10Y;pC*DW&b0`-Z)pN|Km5u
zdVz6?SL(iYP?XMdrsp{avO=Qt7<GP(`V!o;S4@7GtgSiTgBiv%Y>1oIbJ|zQ1pZRo
z21!*ws|s|WQdkB0R<Tlf9~j*SCX{|3*xv`eSv>eY@S*qtuzCQjDPGmaYE>JR6xvlg
zd(V4Wy$+A30-01GJ2FuFC#eUGCY_(Jj-)!!wGMQnJ;XrM7-&Y+DU#Mh)AcN_<%ZRA
zzN91Eo*P!j;s!3bfl~<|k|NQFNTd>Hgw=1)BznawajzuQoA%d;6phH9;9lO3qXmPI
zJ_t$of!2MXO$T}(SltI*Sd3Yi1O5Q$JOFY+cea|MHd$-p5{RUF&ZM4GDR3k`<xHMZ
z?Cg=~k@uH{)_?udxb$8#qPk??d%U@ENG}elbQKXhUcEUgKov$4R5Vc)CrYWDB9&9A
zKvBv=iaf;N3PHO<Fe6(b&5xS7`;VxF?L?^xI#faC3cUjz?y&eCbi4-#k%pr50NOr)
zjG1<qv%AZFxSF%8W^n~lR3LZSz7i=aS^R~UzVIqD&{ebFxrtt#m*Ist2*iWHj@SzQ
z`<j<1mfTgU#J`I_36Nl=+7cZho|K4vfj%doKXDV>Tf-zky)`&O7~R7=3cr6y?}ocB
z=DHShHe|2^qA&M-`Ec=$w#rp9uYxMc7NtAT{0@unL5q8k(IpR{;RC3;8b#?YXLOfi
zjNxj|sG7x<$heXoX%R1p^%q{Hi#$gdW&Zr3t!D@2jncaos;(YU+RABf<;=;5MJb8X
zPGWIBr=8E@qn!3p7T0q2wcG&OzK*l6V{ttv)^o;`{*)7+vN#g;ibRaneF<q_Ld<xs
zAnhwGP8W32h0e6U{esSZ7M~Mz&at>o=ujtEQ2u(MLp{ZVZ=^`cV+JkmqBLE_O+c5}
z97^{}+|1%+ASMG{VuL8%Eir|~89>Ycy2K_?x?f@@iw^<qL%@)TqbS8gF&^p?-Kxa4
z@z9y#1ZbPU;*C(;$ntN3;wBb9fm%<X0iEauByK>uEOy<1yeN)EuCXkRM`Aq7pMb;!
z7AGMwiN$G1Ok?q9Bz|Ub4sV;oJJavw@wRy^zR2reWKHREUca2hVS;v;U`X3X2sROd
zBgN5zO*D&R1?^atKTgn&V{wY0lR`~%ynlj23hVyu5Oj91^i)A7mBm?tPS)SCLl)aU
zThPg7>3am7Jrw(REwl*P1`i|UQ66xYV-#jJH=vrumB^zKF$(hw@B4*UIg~`eIQ7=}
zD^=--pm+%OB~e3B%0TuRXd+oLU#E@pb$hOl$I2*-^AF?35`rjg=M>vHiQ+WQKaHgy
z;}pkOdOq^cM@;*xNO2W0?T-k7M}!HqzjDE$Tu|L5zQdjLE)LojfR`ei8ye23>}FBg
z#SPuX;sP|ZfZElfbVl$yBLt9+{q)Pveb>IvC#-irBJGbzB`hQqjWu;k9GYlq|JDFB
z*7eHhGhMpNocuDKAqhm|_AOCp(th;3T&H20=|T95If8bMpt@#>*^HwBEkSA0LX?ic
zzDM9-a?6NP0qk1<2a~9*C>@3PKW`F(4eofse$qjGE4(-xf!0P~Llx3?ptT*?Pz96@
zw9<hMkq3N(BSk8WpO7#(o@z@BoE?#P1X_|21%KIQ|Kvo3!++eL8|RwbF`fVYyZ?IQ
zX~N+-4SttPYyQ_x(WKg_nZiT!YAg=EU45O1IFa-MUV3G2f&8^7m2>iPDrHl8`JCuD
z`P5PDN%-6PJ5sh6T*wP<5|Q#LmkT_;tGDCuM=Fp<1y$L^n#EGulA|C!%-JckdRepT
zSKRRiq_=@RnXyPOmg2d^+1-+NydV27?O+m6!b(S(>mVNzrB%pp6&vh&#|OTn!L4~Y
zTgzrX8e)Klo&vQ}pe=D=Nqz*cP)l-z_V^shpVJX8L~2RDD!x=G*T~Gj2W5M~4SB%@
zku@U86ucBInH0iElp+35Oa6qO0=(`=VJC9#NFoI@wq#K#V=)wqp$E|kKfic1tn#|I
zE*?<{H>8A)0h?CHv6h-L<b>lfwPagJkK~h>(2{B)1Tq^;Xi2b;S;SZk@Q=GYEf#;X
z7|4r(K8a-f{|0Bw5H9jXT$8iDD9xA6Hz4Xxl$Obs84$4)rD!0F1_opkL}@RO?F9zZ
zG}#Aa`&e8Ex)sv!iHPGqn#X{g4&o%xJjvqIK=U+<8-Y(F@F$yur3PxQf^v#qK+P9W
z6>1Qr2B_J<;$lu#%rU0TD^Bwj4NZtr4U*L$hW-s{enU)uO-R<n(pU3sSMyqAxkagt
z*RA83sNiE>_c4nT1zDoN^uJAzZDVnxpxY>@+Q&TeycIfQ(bs?5P2EtJGQ8)?km*YG
z;^He2X^4%H^-9f{Ux54nd&nC}e-C-%xY*x9-bng;$eSeWu*|n)>&O>y>>pa)fTPI@
zk=Sor_2zUPWlIM3$-seLNhzRz3h-dDX9^fhaV{L33x|-)Qk2p;pLCANe3WoTCDi-I
z8^ih4ajN^Cd>5Y&Wo|Gh)#mi{;|T=`*MDyaJF)!WztmzRyp6S(GIdD06QyLJmkbQ)
zFw;4gbWWAI*gB*5shfc%*G81GdF^ao6)nJbH?Mz`??RSIluq*cCs}-r*SW^pw0C*&
zF0V3EL@Bb3c4Ql-zt}d~u@sZwJRYwymr2LF3FvJChSV~M;9Mf;cu6h*+ak&wCvDcm
z`8BZ#n74f3TdJ%@>5|}nNl>|BDLZUxe=c6)^dGkiNVcGjdUJl2jxPf^WdMm9G?`#f
zCh%eL&`dC#;#}yP3j?S$MWT_BD46a>AsSvtqfO~6V&cj*od*3&F(DZeyzI&>FCAYv
z@(X7#yAOQe2fBV)tqEXL_2$GcZB_}KD%m_*73fg~Y{~5{N>R{0ip`rnfg_&4k#rg_
zVc(Z*9<82ptLK=2$WzYkDT^Dpo(-HTx`#snoZlCA+mn%3GMjhH;e&JdG4y*m{J0#R
zvEK7|mpq=a?(_J5d2CK3j~|dngV~}~#*Zsw+n4c^%J?a?|8l-lId8>c>vFyui*3qz
zCiL-&4}3+J89NO^_i&*H<zFXsU&rDt!k{gJDwr-xTLhmiY@Tq7;JZZ_PTOx00=5XM
zprj}z3O<P}f4yK|FQ`H{zXn}*Q!kv;6C3F9P%j?#WIcAh1QtgKIuSy5`k`n+Cz|4Y
zH-tNDD%vf^m-?&@vuq!>CdEbRIoIPkr%EpDUwVS~pO@~{4Lbx-ZzGMqDsh*M(3XCG
z6V%(p(l^0wo1i_VM{tS=j`7H2I7JML(>X;ti}!Mhy)2GLdhtjV0>tx2cJauG&MzK$
z##0AGloF6$0?VI(x+KtOuafSVfcnw)`;oXG=}>$Ci3eC*h!lm08UJadIL+c%!6;Vf
zMBB#--D9bXhV$-%ORT_zs$vE2SV84j<M{|W@d6X7iWe;7>3%8c{o)12F^?C9#<PO6
zLonMRsN(I|Ars6}S)47HWwUsXV77<F`9kk}fjMYJLhm9L-w?EK2>Nt>w*~FnEPfzp
zKVb1Q!SWe9{^x?_a~A(8SpLc4kAme#7Jm{fKe70;VELKigZEQP(j3mr!Sb6dXeSE_
zI<Vb>_HK&vpZ}Dv7yV}MKR$JIbjG7{PxVVRnm8S-F;|aC2W$MmpjAL)6=3jMps|+4
zaX=%E#fd;8k;OZK#!ePz1C4AJ7Xpn!7GD4w7g*c~G#XjF8fvVD%y^QZMiPrlAu5HM
zWHu*UEjm1o`mh0u#X(-@AZsm!w-LkJs3MZ&c>y*mmCvE*eYpwPZvqa~y4ei+Zw4ML
z_S_5xQ=AJua^WCq21IZ^5q~!-bxT<9p_cQjrAFn+i#J0KNN?-^@zP12bYQbb`TmUV
zM+VT#07lg8Dnu@YRBfNxc}NHiZ?hZkcL`@vLeu+7e^{)gS?^-&74}X?hV1^A_SgCD
zl~bRe%U_Rs3>S351;#X3C+MuB_(G3%yxru;Fnkp+)LPhwdb`WDmiM9FZUoSZ0DZ{g
zH~icRXvKg&#QVK){kO}PpS?TtA2)M}d#rTxn;-0v6}|L$WA(sKKK*dGsR@crkTKO#
zIWd(r-SUu_N3%MYKAo)m)c?;O*dP3Rp6RNnR;*rX7l{d3$RZ1=GP2iPyZk(2xbfD1
zym0?^Ye|redFe{L=AkB`Hq_YK1>5d|y2KL}rQNXYZm3J%QIN<CPIDdu-Hrh#vZ*3|
za?<TMi;I9p5ip|kBH&mA22y+mh-ZKy!RoC~Gl@)y(ml@h9`)fw{8q;HA&YM!+nY$`
zlM-phu6xS^PTJ8kB!5N&XvF%%j4jVNDdR3^z6*{bQ?IIz8l82c8^9}+2M6TAAQB%I
zarkOLJ`5u9ED=YoJPKhj31i)Umog%BTlCuh_yHpC-P)-)KQN^OuLR>O!9?mnJ%uBl
zvJO)N7udj!qUOjq&hH!R0PR3tJ6H$kH6Qqzbucap?iZ;aEJ~LJ_sbMlcU>P?KC;`a
z|8WXQ`~WM^mS?YY#0Q{w0J@NetvFo)6(8YvGD0Ga*kM;bhNaCO1Mx91rW0)dVgrlA
zp;kCFBXP7E=W7E$?$-zZ_`_bi;`pQv`i>*;1@cVLenzvsc;N-@j}(&)!H1*eeK9%j
z$}Ejo04<5k6McYndCOyCvPaLoKMmhDE9nEif~ufb6=Y&rcc9lD7T<$|??IJjR=$I9
ze*jhaSQ322!qf6rnS22!y`XpodXoUPD1GL&Kl7>&y-7KC+AhbU^FMyq-#n{{u0LOs
z{eS-0;;r@{e@3hu?(lENf8P~rbXjQlD*k_Hnh(E}uF(B*R_ovX!iP0JM>}6S{Qkc_
zMfdT`q5XQ#<|)036u~TohGVeBBba4TOdjjtliKo<pB(8gU$Rcy-QW5SpVV+p5l#av
z%B-_u7mEv!qJX~cRA!xfpAl3Cjy$}^+FJSV0q8kx<op`hYvv6f_=de!$^`c^dW{hO
z1t0j9e=tBh+6Tpb(1j#$Md=+>yrZ{;+M?-mCo8?J#fzpcj#K_{P!=tJ`jQ`NP%8hl
z#&64%gdXkPdOJAwb?xKi+|Na=w~9fnAN}b>e>&5jF7#&~`V;@FWb((2{7_SKq(A>3
Dx@#U&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp.FE93431287EC85FD.idx b/.cache/clangd/index/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp.FE93431287EC85FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88cb12e6354a254b3263594861727d3f5dc54c7f
GIT binary patch
literal 21792
zcmeHvcT`kI|2H$Y?7euAbty~75{QV>L_|SRR20ia1#9faL}S!g6E!ihf?@;<3WBac
zB4Ps+iGmfy5)c$nK&dLWh#C;Z{+nf)oyq&2^Z3X6Jm<XU{2otoXFq(uJ9qBfnQxyR
z85j^?G0MQ8-<N?u%$c!hDr;b1&<6d#aM6NUlXwGzr}hR0Kg3-f`^CIF-lq9MHB;>l
zIILsGds?kG*g0UR$Bp4vpT^EmF>Ov<5i6DqY^W+(zc6i=_u8ka<CE{)dFFGe_l~9?
zL*6#kD57Vih?|CoDHc`~c1i7Txh!Fb-!tBR#YXSq#fyP;_tKF;++}56{~ceg?sE3!
z50%fD&PvC)$!*{4^13$o#j=9XG{f)05}&yG<xDmiI{f~n5$Bf89J_rdx9DkD?kMBU
zLw0k;>6@!lLcMn%>R=t}zVzzq1LsC9YCY?$X7vxDt6IldHUyUVnN)por+BSbxm)SI
z=_yl_R(<Z@t;fR9n&95W8Lv+bPtSPGOg>@mTln>gdo}84qqv1lmZ9;dmwxs`MAuix
zuU?D)HemSCgKd`teev;`%Hl?!sk8m2jdZi<X+C0Lf1@@PVf&o+^f|ICY~PnYhpwME
z4XQ?b=Mu5J>OwncY-I4ml!JdT?}}Pk@V9DiM8{be4sSAkG2u;AaEH6=Oiyo&&92)V
z`l0{uZBN~o_Ec>=Ip90F>(JBH4<l`Zy{jIk-)}IQa&v3$!3S#&+dUt1XKJ>TH28d%
z`Hy~GU9@d_)xyP*SE|xtbH02Oy7A4meifM=`>a`4@NMaUqyy6SS&o*)^R1_3gx~#T
ze5%`tUZ-nJzYh0y+-)B(6}<9%H}C70zgTxzn&LVsbJ-Q)h-L5ln&*FnPrW|P`^7Qe
zZS!LmAFx=L?D?eQI`^7oXU<q$Ob@*^ZsYQ(z60GKAGaUte(<jiQy<q)8s%R4>sMYu
zGkW{K>CkQ9se6aV4m(|1eQR3m#ThpbO^e-In3`DXIOd7@xWVU82Qj~(4uaCI?k}z0
zKP_m{QrEOnlf$J^zioc@#hcOX^S=DeZGN%i^brBuqLo=8i?a--_A-tSJ}s{PX4><*
zl1F3YefQ4B{j~n>mw(x%ec#&6vbbu+io)KVUes3X2_C;7H*UY_m5IH_kKgh=n`v{X
zXGCH0OK<-_=H4?-7{2jjge-P^GkkdE@5XDJewfkmLQZ(kG{cjVOtOc^7(}h#_x>sW
zet*ij*{4c6uOA#V{(IBEVxKPFQ@y(NibJ)>ixziYc~g~A-gDFKNwss0V^aKo`Q>C#
z?8KRu3ixB|z0I0BcY4s*J}@(MTk+h+#odpO{QlUxk_|(CSa8_ljQW<-mfHtXnpT`_
zdt%-<og9uWKfC?PsR;uSN55lzmmLfCJ7jkC?Tf;-OZKm7ob}m<jN`venfIu@!?80{
z)U)Hae>}YE)S;h4?#Z<sjB+N&FI|&yLar+}GVS?bYrv&dldjjh_Bq;Qc3>=cXShpw
z3vy3;?CJ5r_II!PM5A2i#Rs}<D(!W>&7I{dUhbbC9NiSSqGPOA$5UxwkwN;vYnFYN
zOSW&m?sMs2Q$(8m_41kJ{SW6=4Gx|WoOO2jqSi;C^`Wy{?Wb5fjPH8gbxGd%4k-)&
z?Co>o!NHSX_4vNQaYkIwvX}>p4h;Cid)cY!m35r$gaMO<tMg}sjcYT`uJqROHm7gP
zCyw-LzpB^Xl6?lZrq(Z>cj~LLxm|bu)w|V<r3-7fdks6K8NS&0$CWQOuMDr>c5Ixy
z^o7a6LyPAar8f@xBI4Mw_(Q|y^tygC_-erCKhEhl-)q!B(;?4hcFvc-%dt8*V&5=@
z+tlo%onuB#E`>wp=Fi<;T-f8qri}cn8`>AtDQiz(D8Dseba$hx-xSUKFst5nWyZD+
zKV`{lRVSaA?)hk%-et!dlTC%urkcL5V`S#g+{w%4yoq_a=tZzs;Mkt^qYKQk)-5{j
zJhHa-i2bfb7KLN39ZV7zxNm&5yyxP3H?}SN<%P<(e`(Q@RY|surq>a*)$S!f?fxa(
z<lzgpcz3zi_LSeB#N7K{EPD|7q0@9tLj9rf*UF2&9ewiTGuX7_;<R4LM|;f)FZ<5h
z_hLIcyWo5A#Xs9Q{yjOmk7;~$x5lD)>(Ics1lwO5*95Lv9dCc)T-o`4uRQZM%yr5x
zSRVet)OOwLr>Svgb58ymC`5Sg8P`-7{n;O7qsDLen0sihYQlu~8TNKt8d7K4zuY;-
z&TeK*O}P3(;Ig-Kj^2q6Rj4#3W9tjgCK>k|{n)ALTjleQ-j`o&zjXG=WUm`Ooxd3o
zJoiMy%H9o~Te*eD$6lW}C2ULVyrmtVSskCg!|1P*sVApSdFwN0@WatN?lugI_4l1J
zKXT3fgtHNIPj#Jrq5e?msiW3EUvHD2e(S~O{<7iSuu8AX&Oe1d<oHQ{eiil7*z`z@
z^}?ldrmYu{x+#dYTJ%Mqy(;M3-$2e}*i6Q*r7_Tw(Ld&@Z}|C=FEY)b&j16tn6WEn
zy0tWzMfB``&?$b^FcmaMz1#!rJ<zwM(X%v6i+|-%Ft`;oLM)dV!^=#smIkvJp3HPh
zGMsi$1>F(L1;Abaomv{qqGYn}c3pSlk{RrPSf*P|w`sm$vdKm3%*hVwVP?=D4fO$I
z_kbDI(qI;K)}RTE57d$w?1NaU7`rN_M@xfQERQURDe0NmSp{9tO^yKe2pWg3!7SEB
zk;=XWbN2Jls-J-zq6q2Ld@&FqKr;qpVAK8a%=-HqXQ`kY>f<C}Poh3_4erAwvBGei
zjnx1Z{0y-qGHfE#y`{k{9^vQj4eu1RLIp*{k_lZh;h>fVvvf^g^!<bqRbMlxLM-#F
z<~cN9FuBJXPw`S?TDS^YAaWeAjsu-r8a>O7$HoP{=hdp9Ct_)2tQ*l1&^4IF=UQOV
zuWu&)qJp0zmbJiYE$T(rV3q;LTfO&-7-FY_T@cGY!0tm2tZOjKzzW+76?v0pnL%GP
z^>-P&yG+-X2D1$PDXsg2obvZ7=zw}zYPGa;^97UT-rs={tw%0U!A^+$5wahlQ%j>~
zITN-&WF!Bf6>N=Io-&3{8H<(%vy448`D0MG!H>;g8^ki#YOX``MNfXu4s5y~l&OLy
zh@1;ta>0a_2D40rE90uVofy&zwn8iy7{d!_5$PH|%eSxG0?RHwGK1|A%QUNLHq94I
zo~8*~Rw8z3RKd=O9LcbeOwX1EvrJ#K=T?54=xqjV5z9oYi8jp_O#a$paJSE1E<I)j
zIYj=}>RZR=3ntH06=p8`dgz>1&<l|l87{JHzF_j~H+OpdZLoW=8SIVT#ZAWUCfd}x
z2D8j*HKOh6&~>lOAd7mKX*JWX`GUy{ET&cXvp!j7&<2qwSWU2LzUawc&WJUfIP{ts
zR3P#Ks|AkD7ffDKG<DB%2g6fla3s2a8Dm$*IJY#IWqGG1KW!~H+o6JvsE5N0dzkTQ
zX)wzQWw&XwR|{e*XpLBw7%s7DzF_jI-9LZz_F$N+8SIO0@`$l}#JIOKm}PA%uS=6h
zTthO)5cRMIvTI<QmIkx@dj3AstK`B4GuR9DQo-0&pqH&{Fw2%@Pksq$TC&LudZ1p$
z*^cYje8J>shw$shJ%=n;K{X;r0_#Y$H*^hV*|YC$8GkkDh6=hOmiv&s4}Dr1%(Cy`
zSmPUtJhUF{5X&U1NjA+FOinJ0-lwdv2r`4+5qXByjP}hJOg=nj*r?ZQtx8qU6p@R7
zOA%Vtx(2hPbj-WLxeiJ<gY6K@RI8~r%@;j+kIC$n+rS4Nc0=R>rd0uB+0tN^wA<VE
zjO$-qWCnYpsk+VB-9}TTYcNZC<H~joO)-7UU|ZD76ssvV%@<5QafBWAeCyi>Drkww
z4Zytt3~gyJ%lW~f6F<%xk6xJ>Vz~`mZlfpBHJBy0^-J%GI}Pp4pgm%lY&F@Y`GU#$
zx6`wJIWy&q3U)-~^N>A{#E7oJECm&tvNiLZL(HHzdXhrMt`O}HU4vO}p1J7HeHc<_
z2K^ArJ;v@H5@)&wvy_12niUf#%`k&K5X&9L?haznHJGKmYfh@)ph=t=?25)R-gbP)
z<_jiQUGCN~R$hqIm4L|et>)V`UoiPm;Nt5&53222K_w!eWem?UzAX)Ad6xCy3wQNM
zuU1e(EOm@w9TJGT2D3aLp84jcO+4C+3}TsWHQTZIg30x-0()6^TxzO<PKaC#*<v(Z
zx(2hn+^c#Cf9)``6*NXHd5mEm)3&9-EFX6nWi1U%W6dCsSY}zxvTMFzvcZBGUoD(D
z0BO$_iv$|B*dbWgO7c~v?Nz2Hz8AJj5Tnj0NoiT2l3dA1l}rb`arvxnZG8tEKBbW4
zKcVrTuoHg&l6)H)--ey=14{BUX#5Oz!kbZ&*D=QHm`+&jNOCmOKbo10g|{Sc;ypI;
z!+Nx6P)SY{`XvgJFveyPW7uL0kw<yRNDrAnY_CZ2E<w3V(03%Z7)$8Ube0nN;yM;Z
z0mkAxV3P?k>g*-L_<%Qi!29SK%kBkdk6rEuB{>FkjsbnJOhRJ@u8E)@HWZB9!ghWW
z<$k-fBu}uP;DDvPBu})T=zvv{Bu}!R<bVZ|Bu}=V?11;2Bu}xQ;()igBu}-U>VPGS
zBu}%S<}k5EHA?by`{@qYq?6>C?Ph+Cm5wCOYB%e1yk;bMw$p4^yi6teTc>YbvHp|f
z`A+j)@jj8{1x^cG7qloJNnQ!0l|Y5f3Q7JINWTIVwr3={05}u?=dmp+Uy`o@^)+Bi
za0zfI0nS)>N^&VsmjYXYtAIllaK<7@lB<Eb8rTy21Xw=-?eId7<fp*;DUEkBeRnbg
z$%2Yy`o_{YnURti6*d(m`4}S|V^ml*O7c!t+R3W$TbAT_R*I)_B_~yK6g`ZW!gvbD
z@lxDBv1c6Llk^wI4~^qT5}c&4NK)9fx7pKiV#jfpB>9lS;t-9C6+Ma-euTeHVN<7Q
zPw+m$VxM3~aH?RDO5-<z=^Md{&{v73t3(Qii>BcO8y(wsrX%;=DHBOf0&EhHwEl|r
zpCqRNHjTzvfX$-udBC2haV}tUX<P)@A{yTY>}?v~1?*iKHvqPQ#;YN_8dBrQglr~_
zvmu*J<08lw(YO|}wKV<!*$*_1VAu$T>OX;D6KH&hVGq&x1jC*n*w~`j->s-}HqXk9
zU{oU*jrA@ozXn}j15biATvrX}iQfn-CvsgAX<Wj0E#WEpUD55X=#KgE?nWX|>#Y&-
zl*0w(a3e4qTF64kPGL00STG{Frt{*6Y=6v1kC}ehw8QQkFl=$+Fr!6Dn(!{^4-XZK
z_H@5%pg-QzEZW2UN<n|Thgq~|`&EPfc+ax(OECH+7(-^@HJJ7qOegpqaC-;b2~J~s
zr?L7q&&ug+?{pd;<vfmZUGbx_ava|~j;HwJ_<nJGe?m{>9TWM_39jeW^}Gea35s?J
zWKXejg2FLDLGdRjoD&qC34gJ|zgW?a;5tP>onj!ymd3tU`s_WM10?x(cEIm!h@Q~m
z2BUYNbAHLoi{#8AIlXSde@E{?>j|gsc0MyBb907GiwlgtuXB7cpCsSs&F&N3WLGlo
z)Uz{3uCNmF06fc>?$|ld;@)LUUxL?hs&$+h9wE}-yy`Zu$M*S~IzI1JcMvJLwW8Zv
zu@7O46fGk~Cyd+srJtzxzGnE}#UcOe#qs%fHs=F&9}v-uuIKF6bKUR^B2CBJSMz$@
z@h5Lh#A&xNXcax;)Q`CCn6ZmT__38Lp)DdTvRUMe-O?^!zVoq#KV=K7ycHU6g>A4S
zro}#6VK0KCVcTffp2j}W|D=zOhGPjm20FyR&k3&O{A#(rcs#6}BRb@W`t5>MD_W(x
z5DZzRIZWFeCK&tru9e<{<~*%jj~=9uw=LvdagXlzcOSN%k~7bkl{K)V20D>jUxqf9
zp)EEyS-A|_ltEi;ZnAPc0?-bdo2=XbeHx%I!D|?sHH<AbKUq17v5R8tv9-X;+Zns<
zG~UaN+{*>yfw6KDH!_LFN4fDwxi1L)5;yV^O^@T-$MGEqJ&yl8j_*kD2j2Yy?}4$$
zpnDs5%W+;DD@Q})Xs9F)9t}H1LuVQfh=u_K$3W8<Xhv`?XIabX)jSS+A%UWcdEs}0
zgI!3QXd_+t&ypMk`bU8vV$(Ll9*wX+o=9(pX$Dh`CohJqoXpuKb9(!Ol@D;X2MG4|
z-{F7sO*`ZNb+58V*L5Q+&mU++BKRs)Uxglc%_3h4sw<%fUb9FJFzWS;2VRqsyoFb9
zA!`D=5okN<f{o-I)o|@<IQ`<miwM1AU91r^`pgUa`u9g+^Cd~%$adSv4#1qq+hV(I
zqH!eKEt1AtSl2B?@k;V`)_FUtS5wHzVx4!<csJ*^o1>IO0_T=M<1)^)jQfn-zmRt;
z<SG7}yxUD0YXo17FckA+Cj(7|E>w!25P1R2XHOX~xg}KT+q&0<$pN1&MMCy5kRAh9
z;x6QJQZA>LG=4qXo_AY(EA{^pL48)4@DE=mIB69h*KsZ9jK?6!xA-o%csGpuZpiKV
zv})llL+#|Y+sNweHtpoL+eG6?wp}ERx3Jb*SQp%pc5>C**-ixSV%587yqmMxO@=PX
z37ky=jc@WcH;E9HWR1{KBj{gN|B=?e{rFPoZ?5%M`^JC~WN#;e{)r$2FEp*y;`uwP
zH=mHrz$(kwwq#)Ctg@WOO{}ts#&22WTN>}>Joa+_<oijSM-q*bIVqX5A@l>Bbb!W}
zIFCy-|9j5*J^e=V_^x@p-W)?NDBrb!#_0-+bcJ5%O7dmF^s+#&Lcz3<#*YQl$249i
zTCWqS6|i2kUoYy{kR<OHZT5?L^AN}F(JIwN>&YsO;LIY3l{c^`&L?*KfQ9|F(pHKB
z=H!h;fzDCDi{R}*+D_AV0BHw}H9*n;bMgi?ptA;e5xfUTduVz*km6}v0Hgw7jtzRP
z(=J`3aS4!0fH^kswNATKO5-XZRRMEs@N1oRshY-r0p(x7itNHXDCI$OA`<hUOCIzl
zI3G&+G`#>y1vIXJQUx?8BD4a!R6uWntDsax)2pFWP2<0z^fxre+DwxFhAw|YZ-N`3
zO9PExL+Lfm{{~8LXdK2!VT?Hu;bDwR7~@Uw21eRI)594loW?sDX(wZj%@E|fF)lk9
zZ-QePmslFdF;X1Ozl)J}(KwkkPG&91j>_fSaycpp5y?xDyg8vq^1hLL0KuDi-_11M
z%1c{m{%yRpjmELOYb@V`d_R$w5^4Ht-uX2#ZINhFxM~y>f3d=)Sm8~+U!rg+q45)i
z#S`M7N%C`r#d8|}E+~JegB(eMbCRGpr6oB<a899dx}Z#_`Hu<8V>HedOta}AO19vc
zE%YYiJ1>}?r|B02(+f1t6HN2|i9Pe^?_U*6uhR5<!8D)7>qO-`5|okTNU>9-NU4_n
zVw?S<H69O610v0@%K>7SJuqN?yK&L*2r@!;bN;(IO7kah{s}a`$@|~rDY`~L`a>U0
z7#wpZB7gN0Bcvxci^|QS-a*81KqTz6xkV!C>X^^!7%v<uAJ%=#Jo7FaBaxQ8zz)B_
z>MdqTu3+6NXcOWtJNzz9f5*DLqv?nF&WCxuGbG7Jc()_O^OfY|{P5#E)&Dg<_!_Sd
zAS3lH1m7Y7Wl63VJnIF0>|j_>clG4~(Ie1|<*-9?SQ*c;B;R9)+@tXce#i-)qTd(f
z`((ysi@XbSTX{918LQ_l>xmD5g#vn6x^yd^Gi<z|snjJ|F+q}FGxo2UE_f(GLrj00
zVg{o>>6DM2lh%D`>jbR}c~el{6!fOzSd&ZcQx01=pr_0S#`$!(_9igCNrz`EfpH}r
zj*aKt;(5JAh7-TYeb8oo$%vys_bA{`6j(j%TMvg~0Xz0>wEO+6)t>+HU>VkEXneY~
zGZ|qKBNZ_{aaKx_|6tw!U?<}_p5Pm6lE6Kj`@ha_!Ol5ag}U@RnX%QZ!)g-D!Z`=l
zusID+*e-BE7w9--3?CB1>*J8f`V>N9NL+H#k=f7X+>146DgM8Os-}4aW|>SqeH`tI
z(JrIAk@2K*A*sYg$2t|Q2wjZ;VLZyHk21qZCg2>?_8c=BCj_VEX86yob!AZhWt?Lf
zM;Ypcyni9lmN=!;!LYfyfb@S2jJO7Z@siXUQ6ox0FfppC!H8-QOpK_%fc0O%mKaTS
zz`Bmck+6Ft?15c=359vaZG~SDd=YwFgr3BxDuEs)(32QRkD$jR=!y5FB(GvTRxzGf
zKB4{?kFAU+o-heTVSA>tzIca7@)<7Z4E>^WIiFmv7oq2IK6x~L%Go~U><RrDXZwuC
zt9Z{<yceOb=Dk<bnG+2kpy3A*dLlnKk%%Hm{=j$nK>I)2gph4yp-J*-Veo0<Od`!A
z_*T$Lxl-`0q;a+2TTSCTg6|y~KN0#o5h%B&PUurd<Mra`^`btpsC5&9_lu**`0t7y
zcZpKR**CN>wWTLy*^~fv3FttUNiEx>miXv6{KpwKSEG=xWwUPCwCi5Oy4BD)gLlgy
zF>^_d5S0<4{_V`xs2rZw{F01TdnB)pq!sLDUcH%MT=s(Gg|65I+aAaXV5DlM1I{OW
zo1NqoyQ}ewPlk+e!qU*NxjKeCQ33EP0R71RxdsMY1A#Oiat#C#d>PtZhV9Av&x3w>
zuphx?jDHyuKtkac*)bPs%VQm9zK(>@BzZ3vvX|4p&pD~vuY6rS*b*t%lZ^36#s#O@
zzw>@JKe%X-8cp3#9zXRRS>M`PIpo!06z2gR>p|Z!j&rtNkv0|1XANSK-QHf)$}9WP
zl;xDl4M^qm2|}DM?rPXv3`AyQ6YRc;-u)4<djyR)!|t1DycKrY3SEhG$blX?v~;)$
zJ#Ny{;UV;RNK1#6jK@k^I&5J)w$RccnGH^6snwmv_DrLtMKb4?OoGM|iWB-Bpz%@8
z<tV2QtDxP=1)Zj)%X`k}JuOXE@}4VcX|a~~UQ0`h-F(1qT3TonJv9nSdTA8BH3};9
zq*3^56jabdqX^I_C~2cn4Av<0hL6^TLH`XWyIG?csZr=tJzDx`jY7X#w0OKmF@ffv
zq)|*ExL7f^STUZ)6N(j--CCmPQKHaCb0oP$(W`{c3zsPRmXMs3B-be>)Y0FsQ%tE-
zOe6h2S9Ex;aG>$$&lUQ7j@HF;ey-5loJbo8AzNu_x>xA5SI`GNB>Aw==`fA6gg#lc
zG|UqEX9<*xnI#O$5(bm|X9+{I1j?K~B@8}AOY8Nb-+ED>bkIsu-~A%xl06iMJrsj*
zf4E)&NmFg59N9KiKwSko5G9+*_sArE4;BW<sL@r>5ouP=_>?n)a6Oi`BF3kZ8H6ih
zw9?L}ni+(p9kwEnSk%?okr7;B+FoJC;DVGN>#Bzq7Y;H<_RkX-@B{`EeH_XKgpyo>
zB$xAp%lR?*3DF4nfJ$DUo4}QOXz1FqJ#xQ2P}&1+$i_#>yn{484UJDjdz>M_X<+1@
z=*s)>46JsYy~6%sMi(US5=3Q!sE-HYp3ygS)qeQap<f&K@_Rp+MO8vC1?5Y@gbe+I
zp!`7Nt)k^tk%|B&i*1rcQ^H>+nwE)bf^UhYw`lxOw0KDT;c$n(!y|5exrsqG0WkXj
zEO2J|Q^DNu!OWVvh>za3l043AoDD8Ql;rVd<85$V84A#Y)?0uXac<&3*ErCVMBDZQ
z)qY@!<ABI;0;+UiiHo2m`65tV1eUnc2qkZU>OQa}hTa39dO+g_;M4%z$U`)O){Vf7
z;3m+ziN<e1>$fz{gn^lGIDW%OU%}QVp&7xcZ0l53@6t>1QMUC_8n5M4YdI>KbcAbt
zgw7e?;#IeJivFH&{hs8ICHW(-`bg8)DulHPWAZjb1@lnBmf$eKJdDPt1l1{l>i@K$
zI!)s!(IJY&UnTjxXntPQ_ZR-S|IL|SEqM2DPpaX)j~2eJcolDzP2nn!-HG!ukbC=2
zo0paUY4fuGhYiZg|Fl6-CB;8%P*(n@4Z6AT;q#SQ%MX9jrot;3*;LwETD(KG)m$bk
zfgO(GNOCe`lFZoQ6t1=!%48>NM_x@OXHrSBdy*W+n}qQcJw{;}qo6|Tu?o9bTB{@}
z?2{CoNVQX%!Y+-_w||xWptogv21$%ug)&z`bzH8nC|5X=fmACjs%hMyFm0eCf*%yp
z2ZcWWDaltv<rR_YuSirD5sddT64$yaVlqS3T)S#wm&EOIzVl=D;_m<P$OTpdXhgc&
zWip~9MxDgS_#Gvz2#?5X-F?6(*FZv{V}9IIjRz9oDGJjRg+6?^uU&ovH|({D2)~J@
zzlnO!7MCN6hRt=+WCZ)!;Qh3%ca$4-l%q7$SuW@-ZM)TQJ~g!MwucYcL)&IiLP!*C
zTRau|Jf&@izlA=36P!Gzu(#nKv%83roCno;unRHiE1<dpb|EhA-%$NG?1JY~LgfJJ
z1{#Mk>M*7Yad~$#>YYp%;_}8a>R1{l38qPcetxuZWz!UbaltrJ!P;tZGG9?(L=+f@
z7a6hzL3a)CBMRdc?C}cr#lt-KLl!rDPC|e`N(ln>c3@AwxC5wn(0C6}?*aCtqj;c>
zr*S@1=R<okKl#uxAG#1+0M!LFy#P8DKzBl~g6b-0Pv)}<I#xj!f~%prnx<Dnr)ua<
z=&zyrHMGYMh60(;@ilZI_zhIQq3Lg+>l@ge&^IvZ4U9c`KN}dw4U7xH;fy+*riU|5
z;pjv{%b^xYj$_nuq?|*G9pe}m@_u(Q>RmK_7vr>xaVPYPtlLGF%1h+4ZuvB>W!-9N
zT*tcA(fBE6`jpOOHgTp+G~UXqxAJ<Y8yVTW>sC6ry_N63l@4)i<JH?}{%ySDHl9i$
zZ{xdd<Eao#3NNMblpU4IOQ|%@<lQoP>iM#Hw`>}x3)bm^zI+8)GD4?x5=PeI9_a#Q
zl&1@W(gn&X$`DL51j??<5ZY&u{UXU3Laz*g$~t5SLox)tNsBbOV1HiFR~Dh<fna}u
z##aUVt2E9R?DJ`STj+Y5q`)QluF&-^jiW?mlxRV;O0=kqrtxl3xtqosQK_MEqG*~(
z+pT*=<zAYeBDPN<@>G&j#rCN*P7~Xw(fFjOJV}55l&CyK<I|$@G{J{rttSN}8h?)j
z_a?E&CQ<J_;$RW-esl*?h$y<jNH>_CxYA#eGg<dcRv-30Tpo9RkLTIh$Q#)PytaYS
zMD*+iJ$8d}MAV!Ul;;G>U(Xekxdh|-bTk6(u@`dx^IXt*S~WxqA<;ziV+#t2_D0~;
z2)Yv^_a$(8N#obR=ryn=^w*%<YtWD2cR+dvtS~lgIafmlqT#w}II3tMk?WR7;}X7G
z2`M7LejifDy3;p!q-aesUU!)hI0uKz^pOtG9>pQ|%jSZz=?C8-gzO*>j{I$*&vT;V
zjwM|Ik7kZtgQ7!?LZ?Q78t8k$={?2x*Ro^k4fXr~TX^}2!pq}dRp<3L{IdqtzeEBy
z613LUwQDQe;{Y2+<0N2|1k{9{28_~ZoCS=sXnY<Rou_dwFv_KI5ilyE@oivqo5pv6
z(Onuh0HX#PuZBjeAvK;%Xp~9gY-p5C<05EOMB`d$R7>Lz(C7n=BN(FyhU!0oF-oBE
zA;#zsjZZK}CurQru#Jp<KJk`AZjtV^8Que#-z^nSs*mR(ah1wRshmD|l;zv5&AqUf
zzaq0gR-ufg&GrO^GJ(cv3S}CNPbicpXk4f;Eu`IsOQLj1qzWBsMAI5#_Me<*zTknC
zpDPk~zbRaQQ}__aWWT~`zoHw#uLR{QT9bVglphH`dHKe;Lp}|)|J{Y3*o9{`Dg4Y;
zir<^VUD6Y1L2SKHP738HU0lv9%X!LnsN|(eQna7_`3m>$>4Sg&q}!9-XXg12S&LJV
zB>mSV>^LTfRF>}SBVM;C*2+jFOi!HZK$!sEeJ`(9nmL_{r~h@~*hX}Ry@KgpfeNA?
z7EBKle5td@xlUcP5<UDuV0;jGkq}V^FwOv8Bt&!?7@r1SHZ2AJl6(mmUjkk@WOS+U
zdT#FH7ia(DUg)3qwp8|O3had<Y1yol%~C~gC}Ueg;|yNPAlbJ+n@$!t`90~Zo%MfN
zKZy^yA$_SmRZ8ak92azs)|cCbknOa-d?xgHM$}-wWRPum(RB@yEy;YlWO9NK`!Yxx
z>CU6#I~4dj{x~woX8V7)>?gJ?E|EZ{kM6iGF@4GzshsJ6ZKLZq7msmnd+}!;CB8cR
z<W48PI;;ooIPoRP=NRD}V~!K+sM3ZJau{=5d4OUyjBtf9#}RVvDOKSrV~$U$-f(?#
z`{<}CTak^p+<3VyiCu*P<4|CWi-b`bD`WgSV~gV?=l~mIoW|JV$UL%e+4kF5%7Tq$
z+sD%QCTDk(C@D#Pz;$@Q>Ca(Fa;T^bCAtJ3UPcQ|cVrn)6F#zx2)cvHdO~g3=)FZw
z0Bx?@jQ?$BES^Y|nr8efn6Ws(D4~;B{#DFaoLne<5jS#X^oG#?`1Z#kZxjI1o}<QF
zM|<eg<X2!vrurCTa*VMfQyb5k#M5EJH-hyWfr{R(60KK>6b=`y!)g4R==_^Vl>*d=
z?Q2ARGODuU#^AgVr&*sAF8B;93VUh~dXw>20lzBH4@-P)z3PB!5J=-8)gXxAHH_UF
zrag%(gfV_$Oh1Cln4mHy7-x@=xaLAGkyJ7AviLEXbWvIbudE<`c#U^KzgGt>z<)*2
zCq&Vm6`51oSachSUVMtEOcC1=&pk_Y&Jukwz8f%_$#4s?*RGI%d2S60Atc;%XZ^_<
z31c0?Sbb3QuR%>g(X(P+BH^}?9k`LzpDfqb`48Mg<4AU3B#pPQeYdds(}CJL|6bc!
z%AwoE_S!{o-H*L3gU8e^M4D`g?GjI%)~kDC_syfi(H+pAOF4BZrw_{D5=-<1y5Arm
zPw<Mde?`g$@nLUd-)Mi3fUp$+bphx=>LxC-JuZ?0llr(5XIxX;efHnQ`H96@|F~vD
zZB9+n|Mw?BaCjCeB;9XL;HA)@eDq}X`XJALd#3C~wvpp}t7nm)6xc6+y12sp?K$Ir
z`v_jH3q1bWpGO-1*VS`ks|meq-H(Xim;LH(%0czxIlm%jBnqgbfHMj8X@FV-oJpuJ
zTd>a-x?<UivMz#s9>K3~{OozXKJ7G;Oj|(vEx;Rh^!DxEGp^;^U!XhWux>eQAnD;A
z>voUECwR9LJVn1RbiGfCKk$!MAT_T2l@t5~$S(taC7>UUc4{js2b6+98V@N2K?IjU
zyE51wuYY7_LB9&vkKi(9P#H6nsPS5MOzl4_!fe)Y`oahF6A4^M5>fLDW_&wU8#G$D
zVEV!c?T3Ndi|*er{K3Fr5cwEDJ`N`zgUH9h<l_+XaVYs1sQ(D%{BI2A{yA}%VXF}V
U0|yNY3LHFS_)vopdIp340mOJmJ^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp.A946F68C0A5B7C29.idx b/.cache/clangd/index/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp.A946F68C0A5B7C29.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e69cb3de047db51be699ed89986eab098bdcb2d2
GIT binary patch
literal 25522
zcmeHvdst4{9)El7dv<!Iw>RC3ib@xv5~*k>mt1md<USaKai7Cr<|LyOlKU-%p$JJT
zDoRl*<eJ<sB~*iv+sN=+Z?*PX^E-e2p7WgN`JLzRu-E(X{l4p6d+oKp_qB%w$@0O;
zIy!y64O%p7%7TfEj*d<n_;3D#dDBBU9Uaa>M`zK_i=#tl-|<vc4F02!Wkkl50l||D
z*`?p+$PV_&r|vCl|J|xgwhjdC3yS&crMKb2bv?qo?-;f^zP@LJo|h(a{EXaR-rw8#
zYRihgo*hm-h%jz=UZ0S1Kj&rP_7CmF!UJ_vvxg4y7;ky?p0m}qs|Ovs9vB;>p)>Q%
zW<UJLev?rLpO5wRH~1sfr@g+odu!v|jp0WeFCS{VASApd;@nvK7b8P2)c4&+Pw{Fv
zJ8!>NsN7nd{LuzDZu-YilXo?i5sP9Xj53aoc-`}_e#>V2ZH?Lgp03Wnw$a%3LApU#
z`hxB*ucBd-Yr(DW``VuOb39-%BjEemRQsQ+<EEWmlHzWtH}^}wlC^W*AFWyBk~01N
z#>7xnPI1PU&ZF*b)_odxd0;2+SwlLfI9?bwH`Q_FKL14%D=x^{nNw$M`sVygjoEKk
zPwgK2=i!p#O?R&w-u$NLFH6D=AAEOZ<0RXk)`i!&eCnVv>Ga3V5mP%~&K~#sE~g^B
zb8YUl$@Utu{M6!gU4oz48~pxf_>J4&yZ0zt)ZZ`LUjKUg;kQG_{y3wyAj)cWr-u%Q
z_3AZ`UvG8(G{mO(-ua9edZ6{Z={A4dH2=U>bz7=_eeqk<V!zoA?T7#PVW9D&2G9Mw
zEGGu<_m28B-=X~3v&_tE_ObVR&5t(sdbQEZt#NLfH|?iyo7T=FdV*d~{WsAQM%eEO
zzY#Mb^xpYQo6$e=r}qg{&QA_qaVsF;=#d%T>VS3WCD}utJ~%hxM6dGfq1=+o(<_JZ
zcZN?bdG23X*!FT}|H{425pDM!J)CP;6OeKA_mAU#a;vZHsJZUj?x$D7{#?H7o=$4^
znOb#%<M9bIru^8&*1^T>c<K1t@e>|&54#*PZa~5hy(|X$oSS>>NBy${X#XWoTn==f
zveq)qKE10&|L}-ge^@{H#m-mgem$~ixm9oDmvsy358f<2Tl=`6VP#mGj@JjoWgc@5
zc+mfZ;WEw5VOzf27g}p#&eR=zZ}R;<?q-J(2geWE|J&ka!w!dZ>~%4_ro*g**KH>r
z>^tSDLyUXC`$2KRtG|D3wA62<;q2`O>DHwOVqe7D4GBKI{QI^Qoz{MHuGRUAi>KCh
ze0luxv=i$OU3#<(gzT^xd;Z6d^KRs<K5>1NOQ)FI_cdV~o^D&%;jN+H(CoU^zkQC{
zxWM|(%lw`PrnQaxSo`E~(#6%IroJ6``^v1V+iw1FaIWt4_*?Uz1;pJ>UllNUneEAG
zajWbOZIY848mFnEYBHY=jZ7KGPe0$){+a*hTe9KV%&;)`B!_p;>^r<_|N7S<T-KhC
zjS*CX!RPIq^fr7^{+18_{Kvx7p}pG#e{mIy%=jNne7(NpUpVvg0+#`TTe}8&ZtR>H
zA9j_Gm!5p@WE7dMv*Au($F^R5j+@v_TQTxgvA<(Ze(xfc^X{VM^>a*K&3yZ4X;mMW
zhi6rq;`}xXr@me?V(EOZzc&7ERo`W`gH@d1`}A;9{xscOX_WgSE>Af8wm9GxTjk``
z=El_}6+e5YgdOvl(7$KU%70FqDK4wLIL_0p?`I}y`$XrD14j2)*Qd95ev<jiMNtXA
zN@u3~{TjP3?s8P7+h%(5bd$2(X=d#Pl#RY!G;gL;svd7y;%Fjo9ecY~^q?#H+mbWv
zSJ(}>dR<IY)!y;E#l2V0tXL85_^`NtL}yNuv-8UKhFsUdOQ<PhrS2UjxpkZHt;v*<
ziNc4wm1axriVt6KlkYgy?RvQTgQu@`&x^w@7}OgWzWmWOc7{J^W_Tv<Uf|iTnftRZ
z8C9%$IDO>2S-ZCnAAex-?uv!kTjrWNsN8OysBOLbQ|pN{OH(`kq6US%g6(cqb(>b*
zdt2?kh<!f}c^j1Xp!Zqp=a$CrTpJQT27THzWJP<kT-SuR_dEwqTA8JS(j<ym)_Y#W
z_4do&S3V7${O#_Titz9U`oFgCn3s4g#%8~Dk99g-Ufr<{x<B>i_Ku%7jX4!GXYL;_
zOF#PrHHOxmPfGT_9aHPG#Bb-(@h3ZPIQ913nAbCZOYF7g8uP@{=);^hi!SxLydmJ?
zm&o>2SqoR3|FWm!{f;Sp#l>au?_H;UYoUsbE1Re>d|Z7o>+aJky;)C|Zus4zrfuuR
z0hXJKbxxk^`14QKhCRDKCCA<?FF>lT`uyh1-j^;pZof?NIi6WPc%JpI%jsW?5?$9$
zD4pmseoVE`tF5`_dMAbt8}%Xi!TyD5KPG%`b!qKi89MV9&ziK8*U_Q24E*EQ{>6(8
zKpSrzc?adZgBsJ)=m<~;!9QT^kOi53UCRwX2gp(kOp1Y1OM_Tw%PYZOmOdV?2BwfD
z%p}aL>4M1YeIJkI8yp>40TGhl1KszaXG?=v_#L{V3iBox8vt9#asV&~fO$)USi}KI
zL2{pTV*}6$vaF$)HIze3gIG*9O&|MY=8|p(zzDMB0@GYL)aC}Um_DoWavg6_tOhE`
zGSg(HZPNvj&AAaJ8|J=Vss`P9>c|naMFj2M(kLteM@9{~@KdZB(6EQ8CR5FuE{JUL
z{gyKyFaLI!1I9gc<l+2q@1~262q-w2luTJJS-5fbnj3@Ez#M*moXI$|rVIMMgG%)T
z#7?SJ12ae-YckfX>4L~kugwOAO&u4h1{@^MF_~l0bV1}U&vwqY>3!`T2RcLYXVB_1
zXxGvp79R^=tM?xB3e<pyEYnP;nKxY!x%Z*4!~=`p^-%+BNS<Ue$*k#u$i9YQZr#H!
z{h$V2-~|0jTl`9QYiSgg9}K>@_gK7A4J=^~6HF$UHC+%n;H^9F+cyP10SibTZ!+Gj
z>4M0zpZSE-u?15NfDI&X0?a1RqoqMCfiHHLC6;UMsR0XFW}D2mXu2SB(C$prH$7%|
zR|8i##5md_j_%UZAeR1R9~-tb$h8KbJ?!B*V4eeymIkp5ie22wWy3s`0nk8}Ldva>
z8rae(EDu+Ywf*z{b_38Bvh1at_rlH7+#r^r7qYK@DEZl_6=(-p7U(W8X}Tct=$qT-
zU)G7NHvo2!yqH<+&~!oM(4v!zvQEYhYz3Sld7kdPc1;&V9@m|+D0n}An;LY5H@bng
z*g$t}X%v=c8QTkQW=}Q%9bpgG0dpOAw={@l!r}N?quZ3*ssRI8zBBpGyy=3-Q}-^u
z)-UT}m>RgjABm<dqG`{T2C>X-^;cv--fjmq(1$(DFqvW7bV20#cYa8D653j=23C+f
z(PW}o(*==#EbISQQjx(91JDYRZ&J)ns!dCyuzVQy$8@U{Pc;|_$FP~U-c0vuX%I`~
z&U97R&oew)0X^8uYoPlY7`HTtW%ZcFd)LSMJ~RMMkYz7zx|g<SX%Nd=(+7rr`<KBV
z>JEQsJ#DcbE}iBEvBXNDlU$2yCl~-j*voN>IZpYuG>Bz`U&N7}H>&j2zyz|4F&Sgl
zbV1}zA8Wp~;Idw-L4WuoQM5%A?by;Fmd&cQ&cSaBL)E|qvV_wX;ZO`UH;83tpN^~7
zcdiRj124$3leXAN4{vD@OT6od;f}k*`>VkK$g+*L*amxPZV<~J@2yqq#7fx!sNqmk
zfoUpmXlW2j;mQfZ7++?A0nmjkYk>0_(66OIETtpupU2F9`jZ+sLY7eT(2h+PM7}@O
zAu4(1r{@O19+D#{CIU+C<_57;kC5W|ol^xh@PI>IO<SyntE{;}EKkdSUT<}H`e6g$
z0DIX;F&kko%?)CCez>>klJbAJ8UO>xl0rGBP+=_%VtE0UUg}!0eXAM>kY&2bbn~VQ
zA~!I3zijzX<I@VXhU7+|+XyUM8pQJEtj(+k&yKZK19v#owY0@r+N-5OEREJb-s<d^
zbfy)MU@r|ow*h#yG>GL(*3H-b0ycM015en?F4|%j?b^~HmU&aA&aWAcbu_KpqM4D3
zk>usHw4C-rt3{IU(yn*uv1p+q%?)&0^fmtCYtZ2}=#DlG(%*KvO?n%%B~n3&)Bt28
zNOCe2luQj6)FSsJ`6xB?C>4UHU6OMs(;UhiZ4pU6MVX!=@gvIP5#@z;yd*!SJRXy{
zn(9?e^~UrXDzt_ggYj0-Wh-z;DngQPf-X0K`?wY%C&{<y9=GV8=>8?SnhCCEhG6`P
z@qWeh!uSuy`wtSwvF35CC8n=u&DWE7AM3u4?T+dDS<n4!4~z>~xq$75aS_|Uh#iP=
zDi@f_1!0`cb<O6wVSJTyy2?3YoW{GS@x<?y^P|f7&_yM>z0QVhDU##^!q5Xk2<A^0
zhNlZ7F|HCqs)P|3$B02O;sA^fiaigCJ{YHnJyS@0NbGrt#7D%QM@W29^gSv1<L_sS
zzS$(M6#G?*{Sh76fp9%G8-aMCrO~!&^eDv0<#e;^8F1tjBgyfgPdpfhpZGnXe-Aj|
zC!WT-rI9oITJV1@C_4a|mGBpvZA$!$pMcdT;Dz?PkTWbKZVR|aN%D0*@H#&j4Fy@9
zu!m-o6Zf!=Q?KK^5u=`K)b{T;x!rM;<S~|GY>^I-<S@%HTeQI>d93AFTV&!%@;J+J
zw#Xio<nfl{ZIN{($rCIm*!s4ZQj$E;a-uC#q>?<za*{1F)FgST)l@t47OPZ}r&&$2
zLpGiye`ocb9n$cUJl$%#9g={OJllG<J+h@Fd5-lQdt_Zoasp+RK-u7(pGavEDRYdI
zD6=HW21|luN|Q{PW1LEvr&8^)WH?G`j#B0r=TfG*lr@$Vrzz9ZBu)T+2_O(Z_e9{A
zNaAuJl>-AL;UxJbkX`}<WY<Y@5iJ$b2I%2P@(Na3!4f!yl~PEY$7=FebIhN|`sA^F
zNjx%-4Z}E=H;&~k@J`&s8*d_U0dH8qn_&I|-nxLdC$Vz@?@nUx0^SefaKS5F=#BAq
z!EC!=gK?H%o+Y%u-MQjrKuvWINq#RFzZWbp{gYt)31hvC<h+jToKyOeyogyOHTiY$
z0u>0wET;&(nqpRycmu_3An^{0*+JsH6tkDaDHM}J;&h5hCvhRg6q2}tVk$_y1~6*?
zF`kWp*+}AKz$BCS8ep!G_!(fHVXSX_-QV@v{TUJ?pQDDKqedX3o{>wat|gQ^#%0v7
zGAab)LdL0(>5S%+k&79pViGs9PK_)<PvYE?xGtE#j1MZ~`=j;2$d`Gi%Y0{yukcP+
zNL<B_uHr-S_j7~^Il@Hrd{rwaC+B^<Z2`s8Dc<}PUO2`#(}Yh+SzX*1Ig#p>Nco^0
z%0OM+E1B{^JCcFAy3bLnFWP~OTttm1qDJC7E}<rsP?IsfK!skQzQMShaw(@=F|MM#
ztEiqxH#2et?NdQ3E0%$}*yk3BD;c*+hM1u|wqG8r=t4%$V+ZB2!FUFWSi2%t*)fc~
zlhf?vj4}Nr*Xtyw=q~tu&OHz7QZ4SC$N6IZJT4%QlSw=vk5hEBmVam-Hw@E@In!d!
zjKtQ(oU#Mqir}1zIbsDAb6twLt|Y%_G3P~MpJI-fpAyce<nOpo38zRLxI(#sCEOtV
z{St0y2|2zJZe$6ks7x*Ym=aErP>dYSTSxQ6c%yl{Xe@!WbjN7k3HM*c_pjm=+kla)
zdH-s@52l9;{^0_Fw+r%iArRBk1j{tRio~{Qf*pxn(*!pXcTW>MNZd0`P;4_sP80mo
z$luEmI^_rsXh$$|ju4h3jKw%t@Xr<cAly#BXNm9r{Bv-}l+t}mX~jrFF(RnYn<GP5
zp+~aTk*u;pQKSf(-P$k_X3GG5Gr(ZvIYA>9_&)}NkV%3fNU;Cr@DU2hNb(w5T0`3)
zR|jcq*uOT2g!@mRGzpXm+Q^bzM|;&_qt^VzRI8syjyno(tU&BuASyBgbwKGh1)PrV
za?G2O9hYfwGxXHUykjM_1bs~RDB1(LpIYn_Mf+i#!Wg75YV<^qUBc@&#ivmBl3c?C
z*N{ea8W)&G8nx-Xdpc>frU*k*NF()$5b}gHN>7Qtr?Amz)BE_z*PfSkze=>P678!*
zBP5#r9J4}sSBeNE`7mQ~m?4bEXs$yvr?{aVLzh3YD$V{5N{BMfx{Om~ZpUSw39+YL
zM?u1BR<oK_gfDVP^mLnI&3LO<QDdv9@kqx@aydP$oF0xOgd|roA(af_Up`@ipOB&<
zlJ$;c6`3x{g)IC__{T@Nz@wz-IL5mlBSptSVdz0pbW{r=)uiai5q)#8=#b=G(Ki=k
z<jp_{rVW(i+e)Q8QYmj_3TeX_0Y|AI5)U{^4aPW%9u!3n#e%Y)8D7tf#?x?*9ej^f
zG)(6%F&W_of)$+Rd1mt*Q4qKDx1YSt!LkC8k!vXZ8mbMNRxS3fp?YClN14`9<|KBj
zqq>qfu#O7CxSq1Dr|dEQC-3l2QniQk4&fwj5N#Vo#k6vfHVnEnBjqz(SUG}Qj^K`T
zD4b8hGFPxc*yX-wzgaKtuYx!Bjx~SBx}X_w4FhJabv+ZIj(NbDJm8eBCHWy|@{q*W
zdAIAlVzs%|>^fpP;oNL}Mt)6od`)%2_plvUY=^P8mTUkc?*JA%NSqIR^MN1I=8XIh
zbbJUpAuY^6t=#c3iA$M*r3_(Cl`#X$NL<N`u4EKt%E<N1uzE&$rdso~a~`XNX0*6V
z9!r>ehq#f4NOSBe=Xw=yJqAu?|6E}Jn#wNys-wAfq1_oquA}tpC>6f(I?A?=vLmr~
z9cgFRQ-<}F8qtw;p(jBXr_(vg0~zSh1*SR`k+_7?E1^s=y@cvkLiNG8jB+fayfD5%
zIbI-fIVF`-CWx-vl2*dIs*rImWQgEfG2>iJ;zriFktOI!T-PLyaCwtCDVbATW>4Em
zIurHBFXR~cB4d7$QG8oQE@#ZkNnFKyRPml@F={PWmulV>;~XI%N08CfYafMUuHb~Q
zr~elJ^pDp1|8r~5tLw^P_X|_r!hN5?YBI3pE6LfMCL4RR-m{~o?Ry?I7p}1rwEGEK
zaZ(^7?Ve5IT)JB>9e}1@l26mlr)g!oLz9(uETEN`8N`fB0i!rYaLqC<MI`>h7=B?o
z;QrpRuJ2gIsgvZ7tm{V-@8{h1lWVGiQ&(`d`1>ye?-zpN<n(Oa>+E>B!(tZR2qk4x
zE;t_bijiJ1ihzRy<J^ikLi3eyQVB=Uj|$z63W`cV@i%DOYZI(^k0eo=B)Er^0D3gz
z6^-{u?;ci#t_yE#ZTr6qSQP8ybeoc~xW8H^xRzA78C+lnsZg_c_bgJOrV2w-Nrn1c
z2zibbDoXA`-=#UTi=>Mr$I?<P-5%vWP+Ss@N1NZp<9WseKO@JJ&IP8E<H_LNGsy8A
z7KR=s$MaMOc}k3Dd0Xz;yFH!Yd>mvt9%LNRgY=6o>+#~j{CHjMinKaGD{@%7BCWDX
zoJ(8h(uA#anl?F2D{i}XMQRFY2h_24MVb~cik#Q3NYf$`e`L)*V$CPX`#JOdoFZ%_
z`GsKrLQrh6zQas^Tl!AutCr-G!0#j&irvPupzm2Q94RoZo7%OMR{U#8{!FVr(~3VW
z$>EGDoFQ;LqlzbSBBM$qaVg_g$`InUjBzU?@kK_u$SCm*XxlMTIf?6;F7*r%+fQUo
z6Ita6Nb)OI^@=6N)4-}4NW7Cb-pO~w6T3+?+$0iPIYBf`AaRyxm__0Wu}y_&jlW+h
zwyDGz#mS*<r!67CEAS~3{FGdQ$GE^_<O)2_yB{Z4;2~k?A#w%Q2q86i1wsiX`kuxn
zC9+504rwk;z(4!}SbYE^QH~?<+D`8s-vrE;wV`w=filAWd?F<!k~oQyk|-lA6Ot(@
znZ&7-lu8+4YZQuyqa@CuR5_Fh7F`KIN&rS^r11BFa{}<hI1xA}l6W_ec9Z;jfV79i
z89>SaMtI?70H+Ml9pg+OWs>wPAZ3v_A4vJY2ru+};FJ%#V|)fkXGnShkP1lr07wsj
z5wh%{NeN6J0$Yq91JlPOt^!gO$zKhmY7$?f^)J!w@WOh<xV~bP9V^M{td!0gVR|~-
zGo4iehmw4p?RlKUnXHsa@@KJ97K!s%DUZZOtW-qe1Wroeh~GQONhkk~olkO#c>|4f
zPRb$axtx?s;?taTn#6gWZ62wr@;ND=q@UrWGbAqNq+$}6aIPiTyOiWe-X)S(0-Mn8
z<PFd9ig~8R-OurajxOT67m>J_H!LRk&+~@oNgOUXhYQ60t`MA8ka(M5yp2qS#0$pp
zBu*1lX@X)$N^+JkFiRMT%sokdBdFex^ml^eJAp{THVTf7Bn}r<;UdxBa#6LM#Id4b
ztVrB{tmqy~CP~+chU-XroM;$F;%%bgw!dTdZRGd2i-y}t`VP@>2gWFY2TcuaB@~uN
zsgx#_GR0!!9_w`v`{F_JT<g%S;21Q03K+ivhOp_17{4MCe`NhXvV={ypX;+92hz0~
zZ_o=tNwTBrD;S1sF1<oF1j@}p(b-&Zg@^hESiJ!usA^)6$NJeuoma1e%KbS#<T*VC
zZ^h+|>vD!LAETHdQ4B%f&A9F+=~q~%D>%#{$v0Tn8!YktJM54<Eb;v}tm_++o+O4O
ziR19^=ZWrlqGGNN8tkDtAFy>OOjFm=18Qj{c_GQGm;tL80^el^+{MWaC}YInba5!^
z82MCi&YBC&(9|HM03U91T^S<4geEPmpuJGE89Lsy>jPZ+5o~<n><_KFe}I-#Jg17s
zZWSs>V|1Ge(Qtnal+-|(A-7VJzfjT_62}4kIADp2fkw_s^BCBAk2aLFS5)m4m8kjX
zw#9A}ju_j*IoU<&@1h*B-rG;<@24EmElKhbO8*Guh!^Nt*7Yn-<)hL#xLmY#ad@EV
zj5-}Z<I#;<-D+|o-2dxSd?-xB>NeH*;ro3`yFSHvG)Z2`xUM8!q+6`(Emm2RFvcdT
z3Pnq_w8DDsXuF4f_QU_2cZ-T6;ht))j6`y1+_4$2g3F2?{XZu}ki`L4WOE@Uen!6o
z&F>%>NjOw*A?P+$Qew6Q$}NF%$2%gC>YYgWleljp6^L;Z?H@%4;J9NQ9aKjvDN9LC
zVT@83!jdUvf=d}is!VtsHH~{<bsz505a*CyXc&{a1m(1yc=`l%`$syD?1ugu3P&dN
z1~UOIoJnPw{y#i%q2V-rVeGyzgp2f!^?yeiKoy*Q1*dohs9+RMQ**T_ns^CTQg&ZY
z4?#tiGZGAJUp)Hf0XUF!R+CQ3-Q%q0IL0#zk~-~(e>>(s9txUubaTnHhyB~Z8A}2B
zDZm-k5B*^2J*DHCy=q8!%-cWamC2an`EqW^wFMeDOJ$%-8Sq=%y7iwkCohhY<iD6M
ze=&Y2<1fiEY?l~T!E+g%Z3kQ%FKeHI<wz&;DfGX=C}~|(S_>6`wyqYlWdM&1;E#>)
zhrs_KPC}weA?O`x>uvGZo-@JEN$)6^4asGR0RC|<@Hpu$W%BNsJW+slL>PL6^p>6p
zA<wY4gtQ0TbIo<U=!bu7Dd2^K=IULvCYJvGaNzZ;{fwX)atsVO21esMIt>P#2BWc?
za0>+90>jZJ)Gq%1DR`n_Hk0cX#VP)3__iaRJ}o>F1fSSx-uyJLBtcNkBV3-%Rl;bD
zlAJ_aCgIRPjIH00*xTQ(g-_@TrM^NLBQ5=Z2P0zo(<^?D{o-jZ$)V~{GgJ&K$x95D
zv_oo4l7FQPex=%>T3AWmOc`vZ+M%jeNzSAUGO2dhjmx48vPfJ=bt=U57us-d>vNPE
z59&PC`aFrtsMcj9zC^XYMB+6dXbl*G6grIXfYz}<jd2=iod(qSdE|nwxu6Gr9-HZQ
zn`tZbWTE(=Ti>9S>@my@(XA^<e2_6XND8Z)OzWFC+YH?b&LEN_gxM9Y^%aij?;2-t
zjigs`LKUZv#FZph2u2kGQNeXnFuF<NXwe{AB>ImL4Pr?AhiLSNsC++W(VvUgfC%V|
z%2m9572ff%sFb&_#yF<7Z^d_0=Y9IOBl+k~;2vqNO-BoJP1Vl9oqq|PqLt&lL;TQx
zBQXm3n=91Oo?07Y&}G-&*_lvW{=J@+k^f%L%E*7O|77I9*MBnd-|Iga`S10gXstrw
z)LaLTh6asJT8gLpp&E5bt`}7Gf}*+Nu0DGep0ni0e_WuWDoi)srV4yKzADP2iYz>?
zp*q)4E=Yb!@)po#3t4Df0lHL>g~pfYVVB56Xa(J)0@r9rFlFI$ks*BGa>l2e#Fb3v
zN;14$#RON8aygOpPQ=cdB>%>`|3(%bC$pZ(Wa04{Rz5=}R*!IjM@Tsy$@hrliQG#h
z?-NP-<dM98Bv1H&kvw$si9~HA-#?O9o}V@b5FE)XKA#p3i{y#$cqBhEl2^jvTKYGU
zd>C5dT0Ab2S8NTKf#uzgl6vMMKk6b`=o~JL4i~<`-wzkU!Ue)HTOoL@5D3R^h0tpS
zSw6W!@LPe4i6mIL6t<oG{&r!)c3~3kKVE1bFA#Zyc)>ni=tR;T;{_s$oGc7YCiUVm
zVfZn!5c{zZ@)+w$Nq#FhycLw<E11U+96pjbN(_h+mEtO>+{B<LvT!>}926xE#^Z?+
zhenA?wo;O##h_@CzfyFo6qP7BvM8ai(R};_i4JI`&^==4KxD&dYnFPfp#xFPk~R<E
z9ZUB^c>uUkX!k9&AJPKoBmz{f&8Jn!j-?pd26eo7YQa;Vb~dyCYxa5R7Fj9#fmeMX
z!$GSA<5jqh430=JUPIzU!7z~w|6CNLi)47`y{LMR!zD0|uB!^yRl+q}Chy*Qy}zC|
zr(MWNg`5&_MzRauO7r0wd@BjGltBBV6C69^9Um_(Sm^Oz?}4J@LQtVJ9}+@$jBEfn
z70oAvupC{<TCZf4?8ly;V`49~_6Yco%k@21?Jn(nk<bM`wJp5i7Ty%?gaf0l`silO
z=qzf(-kJo^nGAbt5=s03XdZyh*q(d@G>=GplGB{T^_|+VrsX-F2oo0ZmPI6fCm6mH
z6vsf48wJBgj8m?0w^m*j=R<p9VVi}{C=4UX*BGa3j8X=BXi*M3WY!+Jmn827n%%hI
zL5uBn1800cdw^ySN#6rH?E!9>o(VLWz!E(>7zPEJEE1mqnlr!>+j9j#Q$XS>ps50u
zcz&vYT@`S~xEg4xNqRMKsRkaHUPrsu(L@-xfp%>maWvx^%@FkUjO%(5$FYWSERmnu
z!5Z!$aVD$DB+uq#vM!lqtS^)8n~B3o@cvm%7RjH*+GXK_4lTV~7O8TtveH%75G^eD
zpjhb|iSMzl_gG^5PgvI{B+h{+-Z>=>qYc~J<&bB4ayd;dNzdgv<>G=7NzUgq`6T@e
zr#VC7V&1YC*YiqpG4EVVs?hVi<$02Ro_9Y_ngHRVX}H)P-(R@s5RT(`P)>?o;UbYg
z3>W)_i-gLL5Dg<l;`u~~HW4DR{v$+>2plVw<One!LR1QHBzc`^xega0NphTM8Asyn
zqUClH?+`6_kT_ZFnk*73|ESpYD2a<iRgq|n@2^Z$m67<esJcwzd!o%fa{Tv2oBJew
zAlf`2ajj@mOX51wrjEq*qD?)<NAB({Omxqm!NdF!=zbiG#2YgYc;$gmyb(`ut|vHR
zqEB(%PH}{fewOcXmM3hXwW4aRsK{VR-YTlLVw_R;Q-&t`^}(<8v|sCKzt+<d^|U9Z
zKD^LZxAc(#JQJn2NJE~9(pyducr~TBn#3C@y$vMZLFw%v@m@-AFNsqqy%Z9sQ+nwn
zE~NAdNnAncRgicM&|3qD@oWTo8%dlD^pZ)O3Yb*T8sASYU~)-(4d`7X`JVy3XC&TB
zGka;}`zPx<FBxRkdkZwRPVpm7;WJ7(J+0eRN4;AEO|1xC6+xQt(Yz{}#M^k)HWKgV
zRl7-C%BxCA{Dn9CLb?DMf|MZ;8S+d)l}VcTZ$;^?sL*p~8_j!a(%S`^TB~`N)x0;3
z!LH{!t>?QToO}L8=wa`MCtpS5SJC)YG!mjQKOvR-ZaKfNDJpMs3^-$B<utH44V<yD
zatm18BF|%|FxDyfoTN4?FGb-~d--3yZ-05*X~DmE+h^7-pB*&3gCX3r4>*qp9C2pg
zA?NWB;{yAi-8_!>-}fJPF;O5Kdc)1{yg*KG;U~M!JKZt_cn_t#Ybh>fKj(0La^2aC
z)$ktP3WjgVaM4G>@FT{>P9p0vaYZ=HkuTF<)*Yt^A}Reysyj{*#8UdPRCja)9s0VI
z{uZh`N*NSiy;4?I`zHUtKJb7tCc3&!uMI&H4NDS1+j3w*9`bJc5?G)#E3D0=+ZMrB
zg|xiaMSIA*?FyE_DQw#m5`X5}eddS~#0cIZf+wPSvAkt0?|{^UByZy_wqbf{_4@mL
zJ_PlGxxS~g^pqy5e^)ZnN-~^#i<NHSBGJ+hxz|7Ru6Kef=I>>!mn5A6-6EG2kZ^&u
zzQ8KpQF%|hrNaiBZT_m6zG|kgnu*X%S1J~ca<nS?1-j7-%ojM~Fj^UCUk2Q8_WcbI
z-T)(1QYOhCfbanrq3T>o{se?izzDtmLz}H{{}<ziw|)%UK898B4VU^m=_4j=fC{k$
z=$8O<ER*VhejPB!GAW<7$)^c#ypgeJWQa)rDz^P9R>?-+d=fQs|FupOR0YwrUo<@m
z1<0Z8O8c#*N8!l(TH1dtJ(`TX8*HIFqR4yYo1MeH+ZrADU#|@A`V_VetopeY%EZ(B
z;M2U4)VSYqb;yO`ou>bb6h-<HhNqifDuoruaYlU{*NQ#zEbsIFkTLaDs(zKKU!^J`
zRUiMHJ)w<pw_(t-$>BTY@IA0ebx%~?!>2A_p{uB>#TZE^Xe2kktPQ1%kc|rWWb-T9
z@Scq0)N%N=NvQl2)(|wmuMIP<q@|T~4|FC8N?z8jgdLBDR=Y~~m-fHW@UVnj#`h}Y
z1CZ>l>-S;s*6G{cL0Nr*hJT5ybv6x;2N5`z4$P$$qf2{Q&G$4-RQeaty$bM=oZmlM
zeE<CF$o9~_|IBDUGeqpXf)gq@C0BsnumtP>v@e*$ca%iACBc`@D5t-Ysou$yKZ*M$
zQ-K&q(ScEPf0V6-DuM}$W{B~pFs3Q^AS9d!Cb*1I>i=HvJelWm)Vjl0araf+eHC|v
zxO-juDEi5{M@j$RpB_Z!DHN~GucAbvq(Sw(K4R5i_kX+S>`j55-CUEG>;A>4LUtzn
z>9tL-xI~iu-Lj&ejT+DB|Jy(PZe`Gk4yDK5{?ApWzi&)Eq^X^?1}dlwpveG^IJ%Y(
zH2J_0N7wQ=RUR3zh!rhkaTyN`P>7b>F#hw#FYZ@fr=5lmGY#0J0Z-I1a)Y2YZhjdl
zT1T+*nzl{Hg=~$D`}17xZhq5rqQ$kAR_YS8Ct6%r;W`EQs!`VUE`EKN_C!n9bWzbw
zpXU@^`e#4;fp935oVpU9KS5P9a9uUO^c1bDc~fRi)aESa&zn4dhxUVy_R{nR$c^!%
zM|!kd^<E0<<<mPL(AU3DzrKDty%2+rjz9j`2Y>X%AN}yhApEf({@53P48$J;@W<Zx
G<NpDq^y4!C

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp.EEF70E721362F483.idx b/.cache/clangd/index/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp.EEF70E721362F483.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f1f22ea873922e7b4c327dc16f52548e1edc0f14
GIT binary patch
literal 22284
zcmeHvc~BL{7B}6~Gjqj35V@kPf+8Y_pe&;3D1t2R;=Y1gOpM~1m}rcm5)}{v?jnl|
zYrGf`M9_yKAcCTzxDca);)|fTuMqwE0)4yly}!PC^{T$4q{`|0^PHLK?&-7iIcH{^
z@33KLl9txcxxT9w&07(Ow6wI^<A2LnEc;Rtw6sD^w6s>mT%0m|d6kdeb+>$vpKkA*
zs^@bvaa;-?b8(!@P36g3r$*{jX!U%Ry65LPL0^4pe6!%y?kgJx_dN6_z<0%uzhCXW
z()H%5IUlnV2Ho0gp7QjE=batD?IvBEHKJqCcbjM3v+4S^|HE@v_PHiLZ+J3kVZp2p
z`qh!U^F_ypS8wGHeEH-?V0cq|l%(4Ce3<i{HK#@&|83pe5qaeol1F*%U#sx(v9y=g
z_$Z1@!p;xAukBQiX4veVe!RmsNB+p@Kh{<G)txpOLbTVhZh<3a1v#FxdO37^$it>N
zdO=Gr7wpXNTa{$l>-feK3%++9JUnNr{=vfUC(fR~Zps<wb7A9u%pL6a_}^CZw@i$x
zAJJ6OwfE+IX-)?3Pp&>RcX&?TtmP-8Z)Fx+E$CR-5cuR%bYGQU@r$1?huL(#v197~
zrT%>yoQju>>9bL})86ISip3+pnY_$mP2AE*zn38gOBaVPp0r~4oPK-dA6KvM7U3M^
zmr*nD^17FXYs-~3A8$-?b@pD^bY-y3whQO1w|338diCY#Hg6sdDeaf=wp-QnfoT&i
zmh|^s<+nb!tM0X~<4bk&9$de+$#ttk;<~h>E=S|Gl+F3R{wMt}zwLU%zMTVq>G0;l
z0jBS!6)${OTKjx%hjZ)Qc9`|FUSk>d>^l?xOA!v%S9`7cXtN{z-XE#oRGd0D=;V^E
zyP}SKvo*+i$dC7jxmvezjSA4seK0jDU{a7Lmu7v&?3=E?MFm9kf7Z*l$Moz)dp6#h
z;(K&|m9Lln+@Org>p}bX@r`v(kLDYx2JpYVSaR@H)cq|>1n>AodzSoSzrJhAl92Cu
z{*t+@I;`%CN9UKMPCBvsvES7V`}#MSCV`nNH|%gSTyXK$)V;epci6#+M_%l6=@63^
z(&qb;h1F@kvqm@E?>8lHb<JUrF#o-8M$p_a^IxqG_BHb<(9b+OwDi-6fIIS7bM3Y<
zYl2Ru1^ExmT={cqblWoH6`L;Sy4f^cKkU0B=<crZI!Dr?;DX<yzJK`kv3F|V#fv`9
zsV5p1bT^;3aP^o=?zv#X<7IDCeCD<9>9_Y#UU%QSOTT*6am}wcdR%w1nYdwlkYnyj
zqaojIQI0qM^<PJlQ~J*t?KX16lhXbA-w&7=TzTuKh=8wj658JS*OpfoUVamq`(|0=
zn1%1cwtTyHviF#XCq{lN_v>8?dSW=E<jvBZ;_C<d52#L=1WcXtt~@z5XUK}s6~$Sr
z7xokCwmr7l=h40@`rC~mYprsupYS6m*AI(#d1U@;Tc?PDi`xFN$MmHf5wpAL{hZ$}
zJlzwq)&t6xacz|OX~}VSjW0{pm-@ud(Jka-3%T26=RN=Q?DT4+81B~Yw$r-tf3Cl!
z-*{#J2;G|tFT5=rc-r&Xz_@&)<8wo=uKVI~b<_EuJgW=!X8dr}z&GH<cN5%~-`Z<1
z=k()>ZKamgiujkgWA63kXI9y!_TCV;>3Y<+6ITshb?V&qRm*OzYoBlNDsaH++QJ@d
z+wTjsA7^`eR+)KGVb|afo<(P@CNA-`4=*}Bd2;dW1$M)V3tT#G`?ZsCsO_9Bx(}M{
zv{T-oHlud#e$=NhIJs`qp05&W^?sjlbV`Tp;I;D;c4%jgx;o$Vh7f+H$HBHww_gfR
zI(2ET|Igb3kDZ_Xx-ff_E@x0s+4H>J8T%p6ziB%oY1rJOUt09IGsDw;&YYC*UE-~S
z(<WST_Hs^Dr3@XL7Q5Ja@dmzI_fW5a7TYgbIEn8KYL5p8<otBLTVBlAZFjpSSZh0d
zZIK?mHa|aNXvN*Wj{_gdZ=>I}O;C9ETQzS`ecsaDg;LPzG%lfin-6Or2r1ql8WJw~
z7RFniJlQY)((9qg0p_(MUF$O(b*{vsdr#9FyPdd@;ykXYZ$$3q<x8J$O3oi%v36pQ
z7d<P&3w7W8QPj3(<DVDiy!_N@`bVEwXT4(wm+p->IS0*WI>qWPe)Y9~{inc^cNrmi
z?|;wDIcV3x>OqwC)A+oSeqGGE?mBDJqwsuD#O>M{I*aB`i>{3|oHeDfq*v~fmtObN
z_if)F`ykI`cBH}EpT0HzbIp|C_q{R`vfBL^*~VsW=bc_a`@UkT@tbA8TfSL;?tzw5
zK-3}C(Xr;cS7x~M`NZYjThn#<cZ+6k#x{Ykch5I*s#&5BJ2+{{nE+*i!L5y!0^1J!
z*VZ@eMCBjVeFWGYvt+@xS+H|!Ls+B{2Xcphy}&^UjWNqiqnRen2SV0q+VZKBk@*RI
z*cOw+pnVwhX>ACLu3goGiml!J^<fvxk_=HYG;M7NOMB%v^8MPcSGR>am?Z;hXF&7T
zM#EzBA!YIF(W~{L0cI%xjs;*uYeQIUqx2U!SDo=xLN8}6`69Hq2nV({gvI%ZUuWZy
zQx7Yl9cC$oW~H!iYeQK2KUw#~%lzx*O4u8-oQG!TVV~B9uz0xVN|#5NX10aexGyTG
zt%3tu8^SVdNp(A0<74}k&;+y0GMZ)5d>~|>E$#I8cN6+3p%Es}Fq&b~d>~~1*=uW_
zeJuK@goc<r)o7|o^MR13EST7IxO3oUCDg^_g+>dlnhzTC8u&gkEoZSl)WhVxz<w{7
z*4hx3g<qZSe{aoUg%WncEYpmpnKT~=dF8Pml-H8y6)RycJO@S4tO&ZbHiTvEk|9a!
z$9^25ggr1z4m8VwPOXiG<yc_(f`h5k1gPw*B~K71^l3h5wc#Mp(gHGAx1>B~ep#Zs
zKGes(H~{Pq;EuL5<QhMXemZ@Dr<uPJ+Tz=sgk~q9b8ACbw&X1IciA>8QVDG^O9?b9
z!4IsZAuORAKUDcw8sf)pk6Ch|SuXBuOG8+~55*m}&DwHR2|Y2(WoUL8v$QmXW!oE{
zlOL9?zM>D?<L;&c`&2N!wIM7&e`$0pv4_H?EwsigOSPAFZaxrl+=zhRr@rj*gFdvv
z<SJ-f1<hI;!V*9IyOrw?IYcO-8D<GI3N&dx5OP9&&cW=`AI>S^P~6RYXqFFcTN}c1
zC?aQRpN2}z(i5|sfo5l5zt)DZ9Bwpu)Z_K{*7{JwH@Obbb<n=GAuLB;Hx~EYIBby;
zI^e6Eg=S~*RazRtqEcGjy0SDFw1pk<RhDZnH)=i*a$eu7qwOcUR4So}$w5XzX3Yme
zF7CWwfkU5JBlV#rCI<r)40^UUgr!QW*Q4@)zynIy8M90`nr_m3Amr+7uLj9(&m|>9
zm^{yDo@w)ekZZ;Z-{|*z0QI3cCU1vmJ6^;s4Pklk?~856oEc=S4=pguDzvI+^MR0S
zHx+tK^Zy}43H#tVxC+g#!tt#QVR<P0u=RMyMM@=Xhp+IJ(N|W@2SR?_vp``f^*pN&
zyJ2!PK+$-REscgH=}W5}Cwjithg~sC7(ii|rKKS(4fA$XNsb%wJn)$13!^Vgn-7Hi
zv?6zw-Q3sqN~pl(`9||in-7Hi!acGqa;fb!B{aq40HXkt<^v%&UU3bDYmXjMLPtEv
zN@!LI9a<X=%Z-@i-oGZ!RKhm+3JZ)Dm^L2>`BOp==XTrjxDpDOyx3^5S@VIAm(82M
zJjR#WBTl>33axF`50ZQhO4pzdnQlowCF-3LM-XWd@|?T4d#O%wk{knv#=x26$x3oO
za*RhBMoB&^dYlz!kP{Vo<NSZ#1%C(lM8P{o7)=Po?&(N!A+RX~uI*c`izF8Tn<5sM
z0G$$GOz9=SxdaTQxD?oy0ym1wfNdFz%Yjr5j0j!3)w<H5KjZA5aqgsJlKhIZf5qZD
zWM79E`cA&@PJRgG-^EM2cq0vNwbr6`vz4Z|*^WkSM`JW>asuz2!243%#5*<dE)@SM
zjQvyar}((&e_WhEafD)Ugu;Vh9f!?(zuMvWTQ^CbYChG9m@i45W<Je|*cwTmZa&?L
z$h0KSFrQ&X?4TsiG@ofjd<{vSWj@P_m{3U$Fb}XIt5}i)%>%87mXzdqUFLNsia?U*
zcbVUvh=U}5(dCQoM0O;3L6-&H8(UYhBrmpDY)uxfBnMdpSrdIF$+5sB7FbScHKURo
z2MptYDaDDvBoSB=>m|uaz%U7zQmg`|D$tdf3rS7~hUvhR;yhrS2P}wxhL<}qKE>i#
zI3yN&(^Vb^hs3d1g$z~5luW)PtI!}7@?`Nu6`D@*R>5GaK-Y-8Rj}A9ShIBdt%4Ja
z2W%CXHC7?GRR}Vr*9#`~f+fY9Mbph<*QFhnUpTw{s+}a4h@DDAGfKZGcDhKhPF8Y$
z4}QrhEPUUhZ>45O6dxpcJwWRLgF^udWpNZhQ7qmI&|VfF0O$aVQvph4aRxvcEG_`3
zfW?&nRkAn?qA<wxCmNz?7AHZJL^1J%fp)9^ON39i(=`u=t9KSk2rHZMg<t|v21qUe
zeM^87#id|eDe$NG8E5y5>qRCKVdY@=ip6!vt`0HuoxI~tz7OS(7JQ=xzvc7KP8^q5
zcpu5T1-sosFN*gFc6(TSOqhI3m_pAF69dA;K=NRTBM-G(-FGrVNzQ~_GNBuB<Ml%3
zB<FvqGQ)azlVG}uE-k`<=cvWWCs!at#)KBvo)9|S_*K!Abi9u}!csUe4h$kvh_KWR
zOag<5)FCWoL(_pLkuroe`Gg`ckz83l*|ST)9E!`plrk`t;&RZt95_&10|wLpcj77`
zc>_0S1E-M!B!_T=LRh>BIc`GC<g1X63Tad<!csa?g+|fI%|hL?kVXoT9LgJp@|`F>
znID+U529GbJE?eQ7RxGrFpIrZ{4f>|SMeI#i0{vjRq^8}J)1Yq=1o{^k<DwQ8p+wb
zT{h1wu57+fHs6=!@0ZQHv3O86&&*E_KPcy~cxVo<(Najx;YZ}~BkB1${Ma0}e>wcb
z9DWi@pO(XGv>K9+34X_f;gp^wcw`AfDXtJaDp*`E467HsDLq&;4;H(y*eY1;&SHmP
z(UHY2!J;dR-Gjw}EcOT%nfr%{Heq5<B0NYA6Q_rXGbj!hJ;KGI1b5bTKR-C}bP<-B
zx6tz~)Yx9rhWTCAPMmoZ6QUI^(F%=tn%$Uk>S=z)dF&SDL#KS`O7;e7JSZOyq4*}J
zf0I*^u3E~+PQ2)MS^Kk0VVU?WQ~z9<vR-0xc35m$8z{+DoLLpexH#`oxA&+&k%%5s
zf~&if<}Sc8mdjh@@*1fpAth{=TH;A$=g_N9Xx4gtMHhU=Qj?{&WZ&OwZqoo$_(Oq+
z<XWIx3))ljQ49LlvgV`?7}o()O0NTsb)YYcz3YH4#r43d9#~UcsIV$jXv{}%DSBl2
z!nBWgs9~aGnCL|GiX?}N=Ha3x!M*SF^I6nzXAAClEi$b|y~!LnOot}zwEM+lbyCh7
zmGhcQOY$Y&=n{)#1jiUbW0@Qu#;S}1ih{6z^BnYe4s2*2_dv5f(4083NZt$0_Okdb
zH{vbFn1FZOh<7aBgeGr78ij}CJ!sq>q|uCOliEv#?8yk!xQ_}kCTbr)aUW|!&hQRr
z=)*;LYW%{*;bdz1_&nGube`gZkX#3J>wq2|NFA`M1Kn9XppJc*^}wJWC<&cJ8S$!V
z2{DpY)6Z(QR-o>zl>(AKaHbzPjgO4vCeE~p#m5BKV?sYVv@D@_mf%2fnCKNI4kKMx
z55+cIv?I8m$1ab|4;H$A`)A_TH)PzMvuUsKx;}&q4^jV0l2drY6zc4fuq~eRmY6M>
z^8rDdzc~D8^U}7G{1`eth8hn<k{h5?1B;(Q=V#E1%&{cDfc7t-W}Qp&Q?B<@>hVeP
zORo1z7S|$&TBPyFB>53?c*NozyyFgb?_R_!i+C$~zkG2(zNqnN+}jN-oHeZ5cSw>?
z0O<tiO*f2LykizWo}QS)OF6tolJQPTbY3q`CUm)fm$MEltJEIEUqya!a0lU!HwSDm
z;Re^^24_po7_zmjf5YA7``f7JtjlAl5h?YYb!lMnGidP)GK%#DG<pFwUZZ->jGuBE
z5mnEb@k<szLMD%>5|ZQ{yy*^Jv(B+ABU<N+8bwVa?|2K+5_Tv11WEokwE8!kNJii}
z&iK%pHww@8YImydTd48NBsq=KOXD<tAKpE4dMXx|a(bmKzR2laWbs?h@h!*fj^A;P
z?^yhSlRj`7zgohQEj6)t59+fAX+kHGd>856MNEHckX{XoABqMK#cpI^lKf0Gc*f$*
z3WLoGhMuHopQNy${3(j|DHM~ailE(`kf8Io1C83jMu-mZ-UrwS(Vx82pKOHaqcHX(
z8zIUO{c~u9$op!{fPGWFmb<E>DN-!xNcL^&Xo?ia;zS@Nf{tV#r;esbNi0?YNd-ER
zeVsa*BBir97wF{zBPtxRP>O{eiP6FLhxW0sAH{LdK90rlP>N^y6QGp9;zp?32s_jH
zJb<JFs3VKr4<OCHOOg*F_k%1xjHJUXe<G3+S*${mip5z-%3|?0UfRYp_e<ub<iBG3
zWL{$_BsrCrQdxQ$FQu{g2rnIBv5L1+vC1Qzm(p4KQC>RA;%r{ZW^oShki%;v6#E;3
zLAao?Y-;QhE-+ego8YpI#Swx*1j`>O7(}wTLa?u(mQa!_1^Y@C*9o2K*l1LP(5ZpN
z!J=NU$V9(3iz7CR6N#n75noZSh^3zsZO@6CeYPZD5N$88xI)ycVEM0#dRJL|Pc*nk
zqg|4GPjtE`GFGHoG^l3jHKIWci|a&#y1!zlI`;f}(V(8CKNbxhQ%oYNI0CFrD^mH6
zMx&xx`Httk<5~G`<eeIMCIZ|fjBR4&J6rV6rt&Sx+Z2Ph(Fm~bu%#ALcEb!_t?`FE
z<rpjbk{j}p#gCB3Bg9zQ9sJN8G;X2R_#^X0O{jXLYgACjUSW~ilDv}}vy+=f1uva*
zNaqGnoW+gFV(I0aLpe)7jqFY%%?2Ka{*gl=V$Lr`V@eTo{rkw_K1+Wm`o9xr()&d!
zoFWyPfXB$uu7)LEyT@X4y^9;Zi_=7UBsrTKp3UMiG`x&Pf+TsPV$?>(SaRV}oh}qE
zZtL<6TjelbFN_BFNm#wBc5^nEKH3`KS_9mO_ld1Mu&V|36z_zNJ7Hg{0UP1CM##7i
z8@T=(*qp-$ZqNsgk=Q0~P!o$cp<bI<n;nTpMY2olC>nMY`H=BQ@_ycXKfBc4@lNm9
zrM5$uv_qIe!jh6)Axy3irqc5(gy|ImGx3#zTcyA}xJqGQB^#fv6oypN$d4q~3)Ab_
z^Xr9xdSN!beuL1pL16r_2En>Puwm)84FcoSycfp4XGLR!IDUhuaXoM}RP@heMWtBm
zSuEPo>y?Q;%UFC%^tvT#B2$ulOZ2@({XjJyc}vtJuhe+#Em5-#lH}h+-``mN6oq4o
zLgTYdT=c8!h;|1i;q~-N)O#grLX?v`oO29Nb+W=Mc0bVF4{WK8J`8jZ16yj7j{w~x
zz?RzJOyrPB^KO$HcRN1J+u-!Kk8qR13cQI>S5*)RBlChYmo3#4gn&JBD4l?Aq!>by
zw{Q+yI86d_y8FHk2lzi0|LwP7Gcy)G8&y_Z{r?<m45-=sg>bja9lZF)_ZaU=CT?~g
z-@Fci|H{GJx=HqvoXLMnaxj__i~@+L%r49JSX|qi!)&$4qn7omig@cHUbA5#IYx8s
z<~kO7y*S_-2ZoaesCF)WlfZDgu+zcVbl^`Hb}lf^1*X)QJPC|XviKfwy~nEBYT#PU
z;`?CWeIQfSTMH)Df+-a5f{wdbRUE?g4`EgB&uG-otSbE#4f~Z<rCG=?iwYigZh7xm
zR<)+^eN$+XK$6e!c4v5I`}Ll8de5rfUBaYYtm^$BjQzl>wvFQWjjU=rCi)*^RoiXR
z_cp6~8b$X;)>&y1-J4kar|AACi$96(pIE$AF?cKWDkM2nF*uaPDGHwyg&$cT3*+>y
z8miZ&;Pr6;86IGb^+9BKkm9ckl5F<ve?9Gg9(Y-Fbm#dm?)i1b>pd0frb2s?G+S&w
zU|x^#gGx-u6s$9ao@5M5k=C9RhsMilQ81rmN0)9e@(V_q{0Aw?!m6pI9*e{`);zg4
z;#!4IM|}QG&hI9tIRo!}k@H)+(<Dj?@10x9wa67Dxd2K9(2^7mkuKm{sf)Ib#@go3
z%l$c<2o2VPEj3z%Pm=SXln32Nl@<0|IEOH*1=sIU*)*+G`Qo962&N%|Cb~+ByKs+M
zD!XWpilI~tyOKeZ5-;4NmclN=7-}_SOhnb(ZFumPf$}owMACQvcSIx9kK6bp;!{5(
zN&a5{`_82B2m4+?e<SEjiji>q0qE}lok`t}BxeErEYO)MglwRn&Ef)JQ$VwZ>UxrP
zg+NKWaSpUQ$Kq1Zu9U?WLA#4A4uig7a12Q%V>cVN+Xj^sUxb}6!Y(8(jTI)>E|{iv
zaKe>qw~56!IQ<(OvwPi$+HFLdRFEXULHchHLqE;8JIyon0$#s>XXq8YqJq~Yix(?Y
zv15|R)FLE{9g|u7o2dVr$ee#$)W6N*0!7CHh35RwRqwwGha2#E#OXcjV^m7;gAlB<
zC?;zSFP)Y;KC&Y46M#}7v?TRwp>>|u7tCMQ^j`;cNr50<gDqu&<c{I@V@CJ-amqZk
ziE|r2fJnu*Q2jo!2ZOTljsL0?ajf?*N)gF_Rf<Udt5PJ!3Cmqe%^|r1E*yc<DcGNg
zktF9J#~d_^JRFjg;I*5J5b0^#Ifw0R+%l7M$YgN|awx(1_SRxWN#3r|+paJt_~(j0
zo~_SYt@@wq(w*}Q{ynI(C8xl;RFG2H3YPPN-g$wEF;oeissv5$R+8@so$j#sxnS^|
zjS0kxQY^b?FHz`~D46S2DD)~Q-Zdv-_ccEqb<PBDa5H#Ka4lxPZOxj3m9Bq#$Co@?
zytrGcQpsW=g;qFRsIEmN|Bd&!Xw)v&Q8~nWA7UMmPrTD7*3noej$Oy@S#m}HT<S=W
zR5f0zEj6t~xYV|(e*yF+wn;7C{R^Qt5pPK@h69SB53v+Dc?F%y;Si!A53CK1xX{kk
z>wmsjNe8h)YALy;BRCIT&ciVzK|y>$Jk>3Qw`5LJCRGg5KEAM*s18{g#=>57!yX3>
z<5--`8z$4z40VXjJX~NxR@(&gZ7e<~8k`d~9)~)lWpIJww5!6ckjrhBW2L*I{R(?h
z)+@>Dkli|3!Sc(hJbuig1H<~_vP)<fPxE+c+&v!J)BYwv!vvO|0BsVWBc<==9CmX|
z$TFUDh-dL}&fz$V^Eron79U22hml4hN%CRT`!E~XIgC6HvtDu{GE8Lo6H)g>nm3f>
zMC6=^G^HHaYev!;WI%)huR$agu=oOUxPX}cUqcSpSe(imrt(bxQu*$wtQVff8>X@J
zG~OnS<|T26n>S2n=|_3Pqb!aP%p=$Wz6ilSLSSNdk%D<7OOF(sA_aFko(j>pLhQ=o
zo)t7MfCDh1TZPD^I4VS+3L28b^}3?LHIYeiToWy?(e<RJyIvC+19DCDx+ZG!ypmik
znpcZVg1JUCuVHb$XkO3a$D;XT7QYw!z89GQ(0WDR^$G^>Qt0hcbfV*nQRv06I6<M8
zz+#ocQl((zBVA#c&f=pA%cCsLQ&{G)_@u(}B#TceEKgBPbQPBPmI8XB8i=GJ?dHOH
zq6l!}5lAn93DFdi{1HeWSsV#<BcVBQ3RSmb&c-{PUWj9_NxVZ6uZf*X@-MveFZ={L
z5n)3AFgCJwSJb;JGTF9AqTVBlv+BRkGTizi^>d-n=RzU=pIm&GGymQNPwmp`cDN*2
z=UYRzBw1%YU~nkV31x8<(1~L4UZAs=#Rq`S0T!nMom3WQ0G$jL7XY0C7FPnDN*0Gf
zoiNDsCmQNRvp5OrB(b;(qAJ*qjORpsuhk<><Xza0-6TxdL@Rws={rsuso&F~ibw@q
zRbT+wfv9t@Ug^M>#lzFVXo~aU$b2}Kcykgi*crbEO{V(i0vdgR7VzY`TeN?$`AxVW
z$x(t{6zkw83VMkwzAPAAW*z<zQ3?^6r07qg-cPJUU#yUd6&m`<lFBJZ2E45OtjInq
zvVXK9I~|uUEC?14HwObPZqj!J)WJaMEz<Zzr#}s-`cz}H;y<>uXCi}xe8+b)z*i{e
zUCVjKE55|LUZS|bdaa}DF~7tA^PSb%rtZCB9J3U-t5JeOlrV^>lcJv2=F}Hv;deCT
zV$q;jWHNSTqCpwOB)W(9W-ad-QUzDdDXVFY_?%rEzIWip^*AH8Mt6-1jS_|c-4NhH
zu7jg;KzAE(dEENy9!^aI-JKY>zU_Ce;!0Ur-K*1o`>;UigQz=0-}w&2X`k(!w4Gy0
z%rZGClf@-SDnS~0>AREH-pP;c@H6{sWm&nTRiItmdm|>~A&Weu*%VxI@4jZ-Xp^0v
zRrF^S{f}1BSFW#|WZR|qM=VpzO_$q}Js|e2pduZ1Bo&V8cd8Vbup@b8>PGG6iHTQj
z09H+B;E*$L5(zlrmtNtJvv3mGW#WP+=ureGvk@`<a@d1J#BRQd8Mk2f)}Q|N%1QGk
zsdn>Q;Z&#BfEhJl7TrBJ!EsG&xyvTxzX>s$;wUsKiVbCEpkWzo&NYtrj$=ccA9$w^
zY-qDl7~9C^OtVD)EE?Lx*)YZ6FgAx7t{5Cn@$ErpJ9yt;TZc7bh%h>YUg%DbF#ikV
zZNB^u9a`P*($KfRbON7M<!4p-k5-lBy>6WOQh(inN_Z?Vj0Hxts$)0jwwqQ_kT^4T
z_gX&pKu=7EUDDwMqGM~;<_5Iy<U9_m;thh$27xJJRVnmTv}hVTD+;|liiu;0@1Xu@
z1ik(}PI-^IljMU2n62gG5hU;@$=9KD9eR*&FOaWD;0U4m6B6Wlbz8K5?0Sex#u^}6
z7#J;hkrKIjpLe5of4S!kHU^L3u*XpImZ!Qte^>*HpF!_uP~&B(-)R{90&1$2iN}h2
z*z#2ha%&v?=auVtO-z-zx|p%$+Z2@X6H<Pnl@?Dvn0@uAVq#a^=OSKF#A~AH<VAdi
zcJs$8==rh0F%~${hZzTOiMR)gJ>!5k#re=9A9~Rk!*0%ZH^<DtO|IijHsk%48}*jc
zR9BL(1K>5K{!#|*Mj|jw1V*IDL6VPiZpZ1GBA*|?5fb$mH0WtN(WsrQ@km4dX{>QM
z$a^1TjmJmc=_6}AJ_%z#QJs$q%0&MY)OfszIg#J{s71HWs{XU8|3|C(7j@NJYm2Ir
z{=a+U=%wDH#~ZeccKWaDM_v`^bYE)pH1a>(xi>!)ukF})M)$wI!<!J_oNlEVumASd
zrf(b5e=)3E6prn(iq})|%*##pMDu&1rnFg->qPT9iithNcC_U~M#SKf@4jH;qyBsm
zRo{ugF%dXXt)B#NiKqvQJ(GYp#re=XANo<_bDZ-%{#WB;e3QP$Dar4+QSYenX>3f)
z?|pmcD{QxSaSppU&DSB+?+7?#v$za7l+oO~`W=D38x<NW_i;&a`8pr#5qL43<CW*=
zI|AgBHrTwWKg2`_sse^8U`!+R7m(Wpn(|&Y@9RKymSFj^Im`E{f4Hg-Eq^ZLhnCg=
z`qQ2MbfZ7})1MxipHN=y=|4L(A?sVo!^O4Vpuuhf+=mS8uQf!&p*4{H987->qCfl5
GpZ^CO!(S}`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batchnorm_backward_blockwise_welford.hpp.8F24BDC37125B404.idx b/.cache/clangd/index/gridwise_batchnorm_backward_blockwise_welford.hpp.8F24BDC37125B404.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c5253527eafd9d6e5627777fcea85801df49a7b8
GIT binary patch
literal 10148
zcmeI2c~lg~*1)H#+BnX%EW*grEV7P9b{GMHu!Hbi5EV^~QE}ghAc`0jP+X9(1Y}Vd
z^IVBb2r4RyfS^b~A_#&;P!_?B7)66dMg1D2rs|#VkDQ#G@8mo0WR7<C&wIPNy6V=g
zdv6=)=i`$t6$m`%`>kEEAbc((5C~NHzpKK-miY7Fs3s7s%`OUR3TbjsX~=nCp!~RO
zM){rQkJZ7Wb!UVpepX*yyXITr?tSK^Cq^dkwB$NkGSp4JF#dGS4rArX#~f`hwlyc7
zPm(5GOnwrUT`GA|Qfqw6m`yV9wSK#;wn1}W=SOu9e9P{J`d^BVj%p2`xR}|K+ah;P
z^zgE|{Px@T{#`>B3Y%kQMu~^FOT3<K7?eH<z3=Ew3Taq;WA9g|ea~EZ5_!<c+p7F7
zV>7SacGSrg{_*j8@#A)R*&2jseG#-ZLUUn@!KvyX57?M|dVKE3sg{{h7bnL&T@*OK
zcl04;>vF5tM}Jh2UY|WjX>lAfmbfL3YCY)x@SbO8%5somRyOm(TmLkZ_ie-U_H32y
zlo(piF+DnHtzmA;%JBx1#vQWT)+UWLO};pIa)kbUGrwiyG^SiE83k01TTfKG^j2qi
zTG*_smATWMZjI`(b-G_Pd;!>bEz-frxMyC$<tg17mFHXLFZ*WQme~b84<cqd1suw#
zk)5h3s*8AUTRu8F)KTN**<ThcXQDg&-+DUy{6gy286K`Zv8rj({Fgm5GtCmVR?PHE
z$a%TP>E*fRDmfUx<R}>PW25_3Hdr=YCf5%0pFMP6_swDtS=eU>Cm&gQ{@7x9@YH$p
zLLMGVS+}&gwRu)R&A)7;4{c35{%+I{gLCv6!VXUCVZU9{I&s&AiDJ{OHx_3vOK5V@
z)|jAQZ|8oqHDiME&|d$$uHHRoT8@AHRm!nPK8+7{8V*|8^v%!isj~(<+dW-VwpLAV
zNr9R7Hk9g6`01@Z$DEIBI$NAulvLs$dE-gq+`^bB^=pr(G~InXa$VPyI~B`M{S<m2
zXxaO*-wy5FFm=pR<<y%Op8VTVUbV6PNP|*fdg|qi^JPz8UYlxkrb6}l$%N*UN-1?0
zT*vut&fU=L=~Yvkq5b50)$20ucil_##AkFpj%t}5o?_qR?KDojqN6FU!td~wf{RUT
zsr<r=$?_TR=kL<G_RCza6?d#=WLbZ<X1DmWAl>a(l+&AkKJ9a-l@YJ}`GR1DUf$O)
ztfE)vB;JmC{4gyy@E6(i^Y^XF3_PBOOf1UU8zJ^@bvOGsa!16CweLk*YLUj}hQ&2S
zUwYp=d@M2kmClu4;w&zIeY;*brDjyU)PL^j3%WnHILv&SS($rOu-fRGp$(H0r_SH|
z>oxz7&I`gv%l58(ymx_SdeWi=UN@`DPhA`#3_eljUOk8T^n_B_2~}2LoB8XEGUyfL
zICrIK(3@fPY>s=%-u&$=qtii!meJJZN0#0n9DiQuZxc`xmbBt}Oi4xLyK}zjjLCvI
z7QO+!bKJId@2;3Wv+$krk)BzH&)=&|d2+=!vSa-4Pt>+(luYd`%}RIQFebb)rmJzo
zW<mAbe+SMBGlv!ncgDGIY*P7RO!t=GtNeer*v`BHe%_Y9L=1{#8~1CxSo)dAg81*W
zM{dq7zP&xIA>@msU_<-m4HE^2C0%EK=p3AP#?-Ilt+LqG@Y1Tt6~U_^?_EZ|s;>>7
z-z|o!8UpSs(O3HY2Z50HI{Z&vDb7vl%E_Z*sLYF3iB=i)AF$ZEeQ@;PCZ7jlsLqQw
ziZ*in2P_T@kgj><xw}maKjOtPKokSC2ToXJ+Mw3~UE^<W{s;~pF5rR(2ix@@1Vb3U
za{|DD>35$mv(>s?Du#pjHnsxMR-iR-!fnim-K%e>nr$F~VqPT^=w<@5ffH7l{kWa=
ztA8$$KqFpdx%P6a{sR^Vb0!lW<-EEtfyTUejrJPz{)0k%)2DS`z&T4+V90k;PesqI
z{~#EGfWSn;cXG|0&9lPyiLZ#E5`X_n(Msd~1HL~oepuTr*D3j8sKSdwL?MR#2P~Ff
zThcUiaP?9N6!GF+KzA44{|_gu@?FB#=xIi(L&Z>yS4jY(1b!qxoUlr%*O2EUS6Ac>
zgF|_h{eb_YH*mr#wbNSWc^sPEDuyhtk_tqrKx5#9Rc>bp{fvU<j}*gUyh@}f(xCr<
z#ocvYwvQaw)JmX`7pDR3G@vtZ!Yc0rKVG!QV11PY4&hbOfp$8t^5KM41Yrvnt}+`T
z5QrCDDR^N1^+~|;Q>wESIO5r6H@jr@?g@(E1qDE}0O;Vi8)~(E_tG@$reQ3%PHmkQ
z?k-Q3>tux1yRG-Zrp@vl)K3E@xbH0A8~t=(f;-A`?Lea)=;BB8v28nWB-{a<I)FRj
z^T_Tz8cVnc*%i@Tg6vC>i;vp(>*ns-t-^9;$hZuBO!Ujpq%t&(Z~|kOz)<ZaGIohH
zCo}fR4Aov5W0=N_B==`90U1mXW@XLm9#-{tmI)cI8@P4@H!NVdXy_UZDNcc&DbS0^
z_e0P9H0MLle40C<Zzmi_^wZJ!bTol*7V^wOiuUjXAW^?h!AqUxqTHgUVb{TONx(V@
z_+Sgi9)}V2dm3_6gmc6+yiQpz7FfgrU+iF5E)I;115>d1Xx=z3?9Dl2qsDO0fb1C<
zL%R4JSU(3ggkzvA22%1((0UW3<f+gx6)NV2;dVmVPFnu}lpUb?6qKE!`5tt<N8g`~
z9J7h7Fq|C8<cNBH37S)a=3?7nc-wI;L)4UQV~n>kCQWzO|M2o^UJ=7>V_dc|pAb%F
z<|H%J{i%#qDq~IBOJ|JI8DsMP8H`&7qp)o)W#>w_Ll?^Vb$0`dyaDX7qW(ka=^<-c
zU3r0?Ku^T`=Ki0dYyR114z;3x2AcbG11)WN9FR9+?dSY>MhT;|@M;*j!QtDaVyo#4
z%iRM8_h_fn0}OkB2c8d>%Y<wuG$VWqs@#HBn9b++I_bg<Wq{!tkbVO)zykavDC*y;
z$@+Q$&0atkh1$J6x!j!cFCoLFBinRjhXoARgGTot#e-!QZOt=St4#hyB7_n8c!fHc
z1PkUW&tA<iTnX|kLH@)7lNrBchT;q+ID?r-+RtPxG8s$E4jvgECtn*V|6xa*zTCd(
z;KS{qYAja-Mic?XP_SGHkdy!u+#>IffuxM)av&+Exe`b!X|4v6YMQSC$yJ(bfTV`z
zS|F*VxgJRBX}%65*J&<<V+x@U>0dFlFNVH^OQCfsRIE{!y9CExg8oEa0c91?n{XAh
zt0Hcj<?5kJJyf`9mTQD=jW7V~J1e<dn(bp8$=h6{d1L^!iK?!*!#<ql+JIFXn1XL&
zxemZ~0L3)9@3=Dd*OpcCA$_CI76E;tSO^UZp^R`bG%6+@lI2RFZYeR`zPB^I1gYJ(
z0vc36h5fNy6&z6oZAp9ea8x~XCfo?k8lfv@uR!j=`WHek5zE~IvRi-}rf$G?lgNVQ
zcSklFDk2S*OMxmWa0G6M<@Q6B{WRx8m3*2DpiKd!R_Ix1bC%{#sNM+`u9xMqkV+O(
zwC`J=?UePI_bOFBmJzNSMooiky(^i9cEGlS;iOX?z^Vf%_9vF>g63UN;kH?>8=801
z9E(h25w$jQkzp=b%`BIXjPj8=Y5yE@J%{}8>Sei`$l@k)Ao4rN;ttJS$gqoUuNN8i
z(!7Z=+Qd-p$1|?+4As7zF_tqHxP3pLP=il1;R)Wm6akwepfIBe_9<TuH~%`B7o>u*
zslXe@wi71VYy9Y)IT^BCBJ@wB%{~wM=Ml!S1MdhwWG2KFCL@((GBNnyfv7(<$=fr!
zWWl*_P6w?xZa-dW-*o;#TOhjyD*DWFIgrgEjDr$B4Em71;09T)7O}O6!!OHnb%?Da
zjPn|yXds&*H+BQdZtw}YF&bJ%LyA+NZ3=CA`=RZAn)9J;KFyubzLWMqS;#gEDNK3B
zn@sC_g`4gE_K5Ox=Pb2-gC2zZ;bwsIJ$}sll07nTp}L{AI9p@6aNTfQ9B|^Di?@fq
z1d!Zv3$V9{KMTHc+9RaX67l||1V~H3Xfg?9Kw3s~IgpmqTnVI=G*<&@HO*In^eW9Y
zKw3j{Es)mITo0u6G+zhO>x6>`l2+bq`!ZH+w(}SLa3QGYPzP_ep{k)e)L5v7>)^3i
zVpOI+VyF^t$DAo+s)Czhxg$XT2zkWv)L^Msn}39m<z7RR*ZeLukb|<^8)))|=5(Z<
zPL>qQ9YjV4kz)5?xkJe45Y0V^?LnHPy&h!VgQ!g*gVD-h6kGc8vYb&_@}8N0IoiO(
z!Dk<RSv)a<F5uAxJ}3Qo4Lo0iS%l-EWjyr59G#^7-HM<M#=Lh*Muy2q(b0H+6#F>o
zy#v2ZN*Sq`3~VNlWC9($;#ofZm+S&MxH7<UM}huPV2a!4w{W1;2qc7W0Hqr=zk@38
zpazj2M(kmvM))OSUlNXA*2Z0%cuUUD_hqDa8Bv>51=6b^obcd8=oI&^q`!T}aY==r
zsJ?m%nW%^n5e|4Mv)pE2u^ISaSI=_W!PxB}5PRfr>K;6eDf}w;FPlxe-n^lg;zIB9
z^ZlRYiu8qL_dn4xM2O`Qka_|#Br{NhRBDhO;cJM!hO{wf%SX4bFI#Q@x4T?yg1jgF
zP?91;T>yOwAXUaV3w_VRAp8b<x=L$L>Qt@zUt_XwV|-GzlD+p|9(7zY;Jfyr6h^vs
z8L^iU70Oj0wu11nwS|KMR_ycPeRzmTh&i?;-XqFFU9bz_$vbocp>SEa3-*uv&fy(C
zflv#uY5`PG)C#OxX_iBCIUNFj4=ul^gR&NA-9kd%ynW+=Pik&o&=Ui6V}Lof-&0-d
zPHF6TcI2OI`%h-u1q)lNy+l=Q{D|%c%zmJR>lcO22C5y~UnKDU_!VSdK`OA0M{GQz
zcqd|a(tH3351?V#+F9-lV$aa>0;Ezv)-%ifgoHoQ`db)w3q!RZ$Ed_HiuHf~$A(Y8
zcY5CW&s6`(sm2q+hdm!^;MggYNjF;C=ldW0%WkKl?%I_rkAA!Gx4Y&VXVdk8lXTMm
zv5`M%BRE$aBpUeofo+7ZBSXtLx|om#RnqwSgd(mz3N4S)^3zb|G?CX%U#j{08o562
zXU+lPIWP?SnVY|52QJRsvF&fS8C+K2_sqVJJLJWqq*0&X+901g1M5UE4sWn6vd`ON
zPIf{*Jlp|%b^t#dK=I)%a7hMkSh3Ba#`EPdP4JIJ{z;4M8u$C8%q6>C@Nsn^99IY@
z5{oN_e#LMS;Zo>XN-~YUa@qJxFp$VApl=1Ka`ctUe5&AhBCm&&>glMh5e7EG8Cd@K
zwcgT)bwNh`>8-E?sAEB2)l^tUb2$)}(_9IJl{8laVKvQHf$%ELH9%NHb1e|o(p(RO
z^)z1x!s~?b*E}`R!1qUD%FV#JnJ$iWL5D8#38t?k=Fm-ZEV7P8R6#Ko$zl;zLd!+w
zxpZ!ti>z~L{d{DRPnXp4kxf1+nD>>;9B(2gGLUzW;~kp2ka-u~elIfb{ZF>;rJuKn
zvDm~=CA>|H%_fE_kjfcLIYZ^*a>iE9C@N>qvd@-S<{22P|Mz_x`xSZriWGJ9zHejy
zm*2{sJ!;$0S>BfO|E}cyVyj<)an*^QKdh3yT9q^5UBk_n{C1T9bQ6FT_T#S_qU^3c
zJ6_BS8o|g$vY6hyIaFfVoc<fX6C^^*MDi_|<?^6q9^u~4PiMQ>4S&Zg)*{_nWJJur
z4(Zkr#$Vz2vfGFMBtX{jZQy*HJUVQ_^0|H6>8h}Kt0MX?wE7PHKgiYDSj+cBz116{
lKf*v$OHW5vTcC@j0)YX!)F+pE<Wh%RYLm+m<WiGd{tx&r_&ERo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_batchnorm_forward_blockwise_welford.hpp.640DA951EF640017.idx b/.cache/clangd/index/gridwise_batchnorm_forward_blockwise_welford.hpp.640DA951EF640017.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7241dba09c1d81abe379d92e3dd11fc2e160878c
GIT binary patch
literal 8814
zcmd^^dpOnEyTJEaOU3rhEn8og(k_>fJ8hywhz?Pz={A*0qm+_MOlc^&l*rH|luKx+
zBoZUJO}UH;4Jx{+G$PgbMSD>5``Wdv?|IJioc=h^dCr+9Pp$9%Jny^KcdvK7>wVv~
z_EJZC`z_NsoCTX519y1%xr=c)99ibSuix&id<N7R2u!)Q!g*(hnM|uoGe080{rpoC
z)uXOr4oS|knUCcCo7yxFWP2v(J8{)KYL(wr=xZ&N)5<CM_VTqlg~!Q?yUqS(Uv>4&
z3HMpWirKpk{VJWJrab-a#01Z(zdiQ(*9yn?Ki7#nlAMs9Wnb~suO<+dufI0ud-EvS
zGdAImHQI#=l4ABx`1{8D4shi@l*5KM2j6ygISc$dtmhWB7#`3n=exEjq&IoxIlKF`
zZ8BVbQTL#)mqGQh{h<%iFSc-3E_h`0OTLC`{z!DZO6U7oXZt6;aFL90-+jVOO;P%f
zm2lk?&Q@4wFrgq#-uqEan1)cX?%0c-$_=@bB-NIjoi67)Q{v3HKaI-myA>7*GxcU(
zKRGSOF1*%5Pe-U<$XVveH@7yrSy{-Je)!;>lKd~G_FXr6)(*CJzfy|tKcM+{X5gWm
zmK*(vIg&L478_3TQ)}Ya6|Y%fapY`tmiuXS*Qc^;Ce5i>cTVkA-|WU~D%R+sb+Bys
zF4+#{*HPIYpXZFExaB_ZDm4t=`*Ti)oI~8h_N@D*uE}@yzkb>Nt4qxL0IMr{(GTBF
zvk<)Mt$g!7_xtE)N{NA$$x6F^+@V}_psc50?y694-u=OI`88X&pKIOwXKsCfYp|G$
z)AkPG-u~9KSE*Zs8owUD@o2w(N7nWp8_nA1R%>V5Nkm8AA2yksd&{e4>y1Dmcy|8f
zrWCm_uaUBs>7nbF35yLE#;@U5Y`8YCra)Wgp*o#EyW)IP`(>NuiYGkd`mS8lzNand
z(Vipj-~Y;EW{m1f?ZEf(ZaqD%^UyY#*T%_<+SYB#iG7gbpuVTVeObW|`=usUKb+wy
z6X2Ap?ef95)?)Z|l-r&6Nv7&SwF!!07B3p6jWlom)8+N81Z%eogC{j6C?7wtqxhbo
z_28aWw_|>tmEwBwRNJ|{V3{l59V&eBym8e)Zq=$i7ay+jesy=^%WtnCVP6N;Cvdk|
zs@e8<v;U0c1?%;T##Pj`FMV8a&bG{>HS}`W!-#EK7w!fdUcB33^g5+L>&flBt69U^
zo{M%rkq}CKka`_^aM=AVhx^1)D<MEBTIo${nvAw!R;vEg!udr@HW>NGH*Q&ZcENHr
zg{N_vi4)fKTr`jHi?KKsn|Qm+rh8FAh?bbYrJMSr+<C<pEUvyj4R7>vbf1+3s_fqH
zVdNGOeP-tBd|r$K)NnoKpK$5ukG>A>r94~P$O}dqPnKy3<$MqS9_gC!LL#Qg=*L8E
z&GDJ5x5mC$HI<jN^Yo$3-cFSjlUEI1wOM^g?}yOi+y3LPQ*i_DW=*xL%UNu>JZge*
ze9+s4zi-%*|3T}(>ZGRNiJ96C>A8cA-ko<(TYU3a&enT|=^B31>Y*ba^3v9Z3^;bS
z=FiddKQojQTnK8N1Xn-E!#cmcPbw5rS63y)sR>m#_N~t3_y+CRr~^41dKvGw{2lvo
zl2De%p?$f&d{M&@XQB!7rzTYar(LRnl2D2f@8Is3A!@ML@NlHyn1|zPiZex?Lo3NB
zS&15s1Ol9ClK>5D6<;m>?PjGk4@xlo!@-nrFzfRJk74`hgxSgscdzi^R7Pcs(iQ_z
zgT;>bX7#siT%0WjDMlO#lq12^&kwBPY=1k`F4#PV2PZHp2Y}K6pz`^FRZhO~$dOQN
zlaPet7?r);y-K16i&H)IcON>^pCAdP8L=nVbGoR(;`9{3jpYvgi^s$9%rw_dSZge5
zI1*xv1v$(#GplabTgFc|kc1PN{@b|Q)I<&LpZR94X@}Py#)^`RI0SG*z~s*ltn#xF
zZ=fM+XA2LCGb#sx(m`eppB`A{I&HQ6`)}X<Aqiy|6@RY3s;I%@=3yDN%1}OIM=m2i
z1g0EfO#SJBRl3V;-fSw^wNDbtF)D%FKxI*b#V=7ojPK%mo;;|^h<7RN(iJsW%-QX+
z#Wz}+!=Vn&O&&SDVh^C`Q-JxWg4bxGL1fal<Axm+Mf*+iQ^dZ-kiMTOp@8`T_INTB
zvrwuBfI417ikY<PL7<N9LD3~(atWA%UopzYCBU3;DX=I7)`V-IdJWXsm~Qg@zPp-t
zDY_o2)kAH<O_1LNHHm&RRBwhln5F3!8-w2;c~N3?EwHKu^RYmTz7H1M2eyQJz=9q&
z3xSo8tzQG3Yv5x1VllcI+BQQw(tijp9D)vnqtS+F<W4vSt&2hH38$fjX^5p?juw<7
z*7s>d3mb_Xn;UXP8ydSXMF(jH&BKmB-@PP0a<<q=T8yp(rgdO8d0Ra&tOrJf+ktsI
zm_xV&n0B!Bdx2>$oBslaf3f@5Lx+0k$hL-{8JdythN1Z|WQ`w<9HNmUksn3#kD{?<
z7NcX3eGFPi<gsX8EV3e;j?B{$YrH08-b7eY+OmFrcI6Gmg)e}a7k~hFRDG#=MIzvd
z6(dmMC~-wWkZGn{UH+xttcm;8844!=#1#d8;zmA}KFe{ypy>U2`xoLEtNFOkHGf**
zTE>Jy;z5eo*JcKrrN)*Ud?QBDWq@DC-m;GY^%%^;OF+?mK&_ALEQ)>uW!^voEZ5yU
zVlf2<S3@zn6VC2rJ5eE;U5Lh<3TG$46=kLT(>fI%rEymbU!2ZU-f&|l)4v89*FY2E
zl;y~z9GT)5nd)^uSV<WqG86F;_ff^OHB;ZnahG22&-AZ{M)lB`^p8R=QOFhdH?uim
zQ#izz{^E-)c1C`e^)k{&hN4peFBPcc5hywX@G{t(33!=o&IY_}Hs=6d4x4iUFPF_9
z!JLm^5g9)e&I^T$35UVCVbGay1T=|&V?l<ZpFzcEP>aaBpkf!ByP;J#bRqIyIIovI
zei52igjnw@LDnT`3D&ointOe{y;>0CBfh%6t~iEKbbxGt2}v2nz@Qko;Z77?3aC<`
zkJ<Wg!<^TTeNTx~^hcoak)10-p>`;oixnw43~GfzOTrOwS_CvE+zoZQ;XJ~<P`j7i
zzX)j;A=Y>$NVf!yr3_sBFni}y<wJJQFt8s6%gMkaVBrW@f!TJca^}84aa%4$=L6Gx
zFqSkZCMi+1B;is_Qj)n3rs0kh-2-HL*enE;kgY!qjD~>?(H{XuBWxam@<ULK$kUK~
z8e(xHl5Zr8Yd2<wqlFuppVt&GV_XzbwD%NmV;o2p{+?oyyxPu}@pV6OKRMPCn+8~s
zB`yI}2~Z&4qZAmFlJrB-NpN}+WQB}8I6aTemm&W$$?6nc1*ccB<+q^KE$H}n^}uY`
zupq|HZE!{#)Ftl=L0Tb*<q6?PE1b<)NIMIe5dCvV`y88FkX{RkL=@eQ^x6sIeoSQg
zRMBF~ISYQ=sh}Huj0rg*U``0I!=)pxk7T%?D_h)yq6>g!0bn_464XtCW4mG0RE9}x
zr_O`Ad2IP*IO{Uh!(VCCRJv7cc@;FQV)t)@dTk^ckDAV?4YEuTf=oitSS3Nx;m9PM
z%~{Aa3ymdQiav)-&#}1$nYSPd@_pKoc{^ce7s2E@yTm1kqFceDR=|o?DQIyDi6s=>
zfEG6pUa{knz)33O8>Zt;K-~na8OI};cr;e)P;@ntsV2N~IMMLM)q^Jg_>yp~W%8tq
zaoyfu%tsp=uGn~Y=_4kBrUK1WFjgIp<|EAvHfI9OOg3i&&1^R30L>gW=K{@KHa~+p
z&)Ck^1$DXzZ~pFL<%)sq=a8bkWW7{a5k=Np1xJ+a@f$U5I~VO?lEyK}KgQn0agZO!
z<|t$kg;?>v9LblXvFN$I`m9#+sev_LOza`CN<Z&d5v<1S(rh?28}hNU9$AFK`eOe!
zV+_X^<4fWKgksKNyhA_*AJLdS0eIm+1t$oKjsdDMKm!j$(dU8GdB7ul0Z3h7a{{Ci
z;3Oh1gH##equm8QZq~iA|6_~9-RXWCdUbE!7cKJB)J+qwNBR80SmZw+34hvB{l`4T
z>4*`Wfih>{RQ#QBwaA=LMvvnpoMl5Q8;%_*Q%>nT-B;}|_{VR))4f$qg(|=K{uc}3
zy9XwRj$S+vL-<VdF~c<{MF;Bz+hdzy9VUH@UPj=;He>&h*qc%UyDzR5-bc(UNAD`|
za7^e2HdSCPULuOV1N85Ju|pWnMdP@iFE2>P6Ug`kDSVhPf#fsL*tVeP8f09<mNz5$
zW+KOjea06)U3rj!3xQ=JNp9x@ua0xwkzvnVFnFqa>SEI|(ZSSfE}j)bc3#eeyiL95
z;#k4dB6i-+go=Pc5g0o#Q1lgGaD~m!f$noKb|9eWeqh(nPC4(O;XBB>j5vvmPm)|k
z(RIkUj&Od46R?lZD`0$J@07i|csYxDzb~46xIg!+J6x7*d2-X1`&%btk4PXE38e6G
z`>G{h_Ic#s(eV@=3&mrh9C68XD4x#d>yWw*S%qQ)lxcuUME?k)jv$u23{ho-%j#NJ
zU9k99*H>ryYR&zIg3F7tz8EWTNzUxe(JM!?GlS*_Eyrj3+B?k+4L$EF|M5|}(NbgI
zQWuq<zB+%vrHMHA^PMqS=vts(OG5DIk<<7-`^eY>jC<HD1o}d@{ty%lL3^?z(~uwy
zvA7Wl8VTQfm$K9|@o>aH-jzDd8W+4jPlK<{m2gmCBF5-X48)P5;ml}~EOBYWsEz*A
zK-BU9l@C<0zw|l?T@tsR92DK=IDG_5v4EmOp=&5yK{yP$ghAG&N(8izfKEi-4VQGo
zZwU87*Isu2BIH_xSmTu-w-U4p>-P;!-S+a%iW#Cqk$5VQ$AZyAk$481Gl6&}o3nv<
zHk)&Rcn+I$fp{*PpF!#wq{;YQkm@4bpYltEem-AK=70avz8KmSldU^?-}(RP62HHv
z^KjRV&a)!d_`kg98w__WQTwf6fJvtR{f+Wptv{GF_g^Y!3ZC;|=6Mn&hljH(^dJ26
zPIRlG-wMWl=Aq~mWSoLnUfh6;8wl_A*zP`BHTmw|<m)#2L2<MZ{Xx3=Zmv_)rVaPT
eE2}6dDNg1vzun<KCH#TInM^*akdF%F<G%r<*{CD{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_elementwise_1d.hpp.AC0144CD925F6905.idx b/.cache/clangd/index/gridwise_elementwise_1d.hpp.AC0144CD925F6905.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cd6bb67c7134e3b49048add880b8082decc20359
GIT binary patch
literal 2948
zcmZ{m3s_S}7RT>|@V>$|2-zm`k_1Bt??(d2Vjx6bF$IdDl+sw_sgXwzXr-o5WGT`G
zEsIb<Eky(gqP#>DZLNw(%L=}_YPYg$sp6yEtzZ|m?j+w%?)^T$&6iB_oBzz5Gjr}a
zlZ0rgw9FF2A~r>rsxpc;(lHEM4nOLm!km`?!|Y8lth99~QC4@;n^=%flz)Asaz68x
zsweyW$HV`)Trt`+mg4I0<=JCzYmP=%zf2o%;?`g8pLxtJ$ozYY&~JyhWK5PeGGr3%
ze(^MWYTwAtJ{rIN$RFn(w(LEam>DuMui9FZ5dENkxAyZOc7=!IDZBcD=s9Oep)}cy
zp;&kPh4%P=>P6{4uaCZTQmw7~YfQxay@-4H?IT6X;y78Z`18&;{G*PWcKd0&N@Jci
zWPfsOoP}>ooxrWc?N1h2_I=u+kCXp3hh^8b4bSkVuWh**J}~v>caOmJfC$!}SQAb4
z<c>I6$;J=+n|{0f@b2l3iqEjp%B7i~XG_kN1<ik9<iECrc_KT1VU^Ls0Q3DU`N-A4
zW#yLk16irc#s?P?QjDGsXp%d2Kiqe>>4l?2(epy9`jgAGMXuxK;a`oe7W~R_Y@go#
z<ch+jd#uO0Jaxb8)td);o|pMtzqfa9mtaFzs$FDR(G81<*O;#>Yng{>r;awnPb!jo
zR(n^*@UJ(%m3lbFJ@ydoBBPdGenPM_>%&QIu~V<_bU@XE_t~PmTfV!kj%zNl?0rWb
z`{?PPf4ce4O`pDK;hf=Ie2afW6IUjg=zLqf*f{FR5{$K{RlNDt<S!e$2K`EV4{jef
zF4_FeS&caN%7xL8+Y^_wZ!nsd{wDZ7tow3ILDlcH5q57UrhZTL?sND&RD5xw|9C>O
z8IIZD7{MSITz!L?8v)GL5)feQ#Iqec$8()5Y39o?LTDjm>Kn`qPKK}X_1c)0#0zX&
znkkgs?A^Hf2Fa7&qH=jc*alnL3Ml)t{5|yzl4n8|N5h{^P1({+pzO=?U8!%7{C$0|
zY%24B*p_AnWxhS%Ro@_auIpv$8#ny6F=!T04q^v6=o=*ea>lNfUnE^?LNkV2TWz}9
zUf*EmbQ*LVZmlpQQ;oxYEc34ZEcY)yw=8hN*U^`YoWh9!VL&L_0Zv5WQ4C}fC*tvV
z3Z>vF6iUa_DWt(Q6lwr?1E6p>SkVnwF2x6acRqB7i4%QbMIXryUv(VLy~XsP;e>=I
z@gs-j^YRVkcz^Y<dGVbvfD;UiVWE!;Y<M4t_Sc>ER>1B;W}y(d?y>24h&}CtvnF&x
zz!un}Q=$_dP9CdB@^ticCdtdmi=rjPCdrU1bRsP-Z9O8NxqV02%JX@EPBa4hM&N*~
zio*AX4=g!bzaoU>Rc9T%UtK`Q2`NYFkH+JKoFk`@!d($UYD#ydf!0`so_j~ic@J(g
zz%0NW#Y<c>8T|MMbsIc61yjMN%w;MKBuQX*lr5S`;GhhCMld;OGk0?UBFW_1>T_cM
zHPApdH`^C=ue;J3(vlLXHiC8q0fERaPP73*8wf<haAFvE4g(QlnQ%{@okwjwTum4(
zj6~@rlM9&_Bb(x3cO6Ko1IcLo2GucPoK?FnidBGF1%i-}x~9dZ-T2<tS5EZZ$>H#=
zKYAI#O-Tc#lugM4<p#=7eA`^>e)SYwHjA6(hZaTU09feDz|no~XEH{24_ACSU_L9$
z(FIB3FOi^V<AjVS6OokfpHI;$0k0Am)Lk(~m5=IPicE3B&DL!tsv9`Lx8<)yiH;Lt
zR$;ct8R(i-7KO5{vMH2nl}n*UVAcq%(aiQ4*QwUW`+ec=jsf^)B0*bv()e{%yA*Du
zSX8_gb>l>dsDwhhK=3XQPO7yaxR&Ihi52UnU;AMJt{&zVW|%gTdFLC6#id!;9YhbZ
zLTPw%jq~!m)r+6N=cF0n&A^UawgccDR2~HQAjzll><WKkg?Yg^b^?AU2qX_$3%FX~
zgBU$SMi%<gjItIdYCv!eSV!9110wf;1jHx=W)}LAO->Uq1T2y~?K~Ywin57fqXlpx
z#U_P98XFCTngOdBtRlyD19mrXK|Z5&F}BcO2&92+;%!Ofq-qdR4HP7|0zoS<oH$Ml
z0nZ^aFw@t!DcqedXkc_BcoCHIisQvmD36y%p^pIXBl2uf@rQott|QW~oRFMQ(vO_z
zoLG|DfUu2BOkKtY4Ff}gy5{xp)<BZ-QxlvJyNG?rX)&A_Lp8yPM3+R0mdZ(`Xey41
zLgm1<92jaN${iR+-3>{4>|y8ONRpSG7lq<&;#ZQ6q}!xZr~%kC01Ed3b|2uNz~RIo
zU=0EnG#h=xilMuKk--O1zsZ|&niqBDH&hXUM&X1qRB50`OAfjBXEwO#Diu>|SQ6HL
zrrbbiMQA{Ghoh48@A=u9AJ5Mo(i!-_^?k7_dW3VSf8muH`RrNe8L!)oPhhUrf#5oj
zh#V};*s@WVx$44nb%^eV?oadoUW979dMWV_lc(ucjOEstA;U3@fren1HTm<uKT2#^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_elementwise_1d_scale.hpp.CF7CE1061BA1DF52.idx b/.cache/clangd/index/gridwise_elementwise_1d_scale.hpp.CF7CE1061BA1DF52.idx
new file mode 100755
index 0000000000000000000000000000000000000000..90272db2e00aa239f7ba6bc6449813950907ee4f
GIT binary patch
literal 3184
zcmZ{n3s6(p8G!E(Awal>BLuRW@KOjyP$4`L2&Blnl;=hf(OMP+RN9CV6s54NtgHAy
z&;<lVd?4WN3Wbe2zCp*Vf`A2FUtl!~TJ2)Hj^Gxt_9oMlduN<xk~8-^|M}1P|MNfY
zL>UtmwVEV|@XVM^s;vBU7C{g*sh>JOZ;h`BLA)_1h)wmU7p=%0cHxb+{Cd>7l~;a#
z`P=m5%{|hw^w?7y{~UR(lBH$;7=2#+vcwtHKfZTg`E7BB<gG*hrR(PF3Of3%G<&O_
z9>`9!bFGLQx@FfoSN28SH(oXSqfWHw#<Y{oVzayMLbY{YciGF=S9*44?{55hI4m&q
znrimbg2U|l+DX-D+{r+9%ReS`{B68C>fU4XySl~RYfn6oh+dag*XlfGJ#;#r&|dmP
zGG2UKv@+2x`JVR`-@<zlVeIMcWgogb9ynCqDUsGz%v7gk%sKgfXn)(n;a~Rb-BP3`
z@>lF?c8`A3Cy0-`6=-=u(KR)%%xPRQ-qGm#F6WbgWnBSBODelB#~&I|eij#|TlC=K
z^_62p=ZT8@iuaD*|K9NZTc_J-<dEHnZcO77bfe+ZyW7^SRoZUx4S(R>e`A%YVrF0S
zNWxh1lg#T)p_3K!+_nGtW%%b0E>q89d&cYbh1F(PZWy|g^B3XKjp~O}=*=i+%+>oc
z`;kiVY*5g-F0=DMM(d%r_FZEUrOq{9{jl__ykwqkprppLKQ+?ILZ`gVTj1}$RO4Uu
zWUt58z&7cDHA{Cc<CL#@wY~MklT)R;_g9N}HJ+=*cf-Dp^Kicu^w*zGX*PAvxqYyy
z%Bw|m-~+)Qy!y%2XXP0J*HQoK_r+v|OWhyX1u@}WX^S3it<|O^m3yz~7%Ud|PSjtn
znz+-xvu=jlE5R)1hwEN(3;g8wPH)ZL_?wEk3;mZ`MRldEik>ba)SFcq(^L*wXtMr=
zj>=hfCgw8;Byp683<F_7$*7MkEAgM--cM{L30smtJj4?j2Ex)52s<l4AaC0Cbb9FR
za~l&5r5`E`butWCKRn`W-%G*5X*&}OO6=q4<8BzRI4h1l?&<KHI-N&}g9X703<DPD
zS2>+_v6P>+GO?t@A)FAtVZh?*re)!~t6OfdP0T3cyv)5E4Fh4xqKbl`jN30I?oS9W
z)!Lb`DSf#>?qC?OeqL6#`WodZz5^?c!FRkZiTrH*1T;zQrA{;{ZIwdoJc%-GGa0l4
zY<B>I8$r+rv+30u5Ng16{`QmPo!~D5NYn+wE+%V1sAY052zzOs;azqhr&=;RgG6#?
zxd=N+B85z$z&A=FC8@Nbk6cbJXHX`Y$siS}V$cpCcL0MMfNWs026!3}(8iww-Z>_B
z0k4b6THtA!+zY&3nw_p3|1M{kKi`Bz3E~7_dfP&Ap^=<l%sFGZ@o@-{h)3`^hRmb}
zadE)mip$<9)N~QQNQRxApE(`mXi|KY%|d=cKSvxg7V>xTpG}k0Svrd*nTw1eON&j5
z$EL8*>bTVjm?X~{+Jv8LbAg49fa4Kx!l&X&0gj<U+8aDOv(E256A-pPOfxmh`lIz&
zk6N#<8f-~WZrQ4)7iH^NBua8i^24i0l<JnspmfjlAX=8=nPa42uHsr+l-p+NQWO>n
zPn-{-J8#NgJyIW{?m3;G?!nOI=NJiBs|CkUrPwAC`P=(D(<HN(F=)B%atC@lm92_F
z4Pe^<3_b_KbKr*Wl0*X_7ywtS9ibE+FPgN7plr<b$d%)15^WG|@W&K!b79%FP`_Zx
zzdTW%51x)5st-E2EK+Ssxjj!I#cq?R1?IJY6fYyu5G)vi0L<~SeYy7YcUMxog~&r9
zal?sEl<{vw)_z7!*Fka}EXC`Ws5)eE+++Hbs|0=}1Y*JOgAdvsGyD76kK&wh^Gw*<
zF)vffsU&F<!>Lqhs*$j{W*kHRqp!Lov?LiHL85&Sz7JApZiacyu#o0KSTG1-n6py)
z>UMkHzDU`+UbNm9Ph%fVIfiCRpH&6&Du~9*NK_Az^`OLD+NH77#jnvPW`bWr1RcW^
zaY_J9g?@z$*&Yz@0i$%+?@YIjS9}OCCy~TX;(*&OiG1vQ9B}TDD25wjhfk$U=dNc^
z9ygCc`P_U49RZ6YV2w4gUsQeR-we1Oakr#yTK_)4t4O2{QY)}7MKS7Fnp!~CLWf$P
zLxDqJthi%??)~+z4*O84cLAs`pH{pM=sJ`80rk^-a?=H~B-ODf>Po+afG?Sm+yVhx
zm|P9=YC5tcssVWo&8NG0(F318nxG;bGdISV5NER=_eOAfbkuYxE0l}d=7r!{Ge^rN
z%PD&sfoueO+TJ!G+n786<N(bVbM5n<3qriiNYoCZc976vKLesOOuhu-OW;S#ufp7`
z;7@Zmcyxml^T3@ADV{DrmQWWFFOD}(lW2)}34=C^H#4Xl#O3rBKY9H;tV#a-8#akz
zU1KG*ON-qW8=C})R=BQU$g<tC8L|Sm0tW2@_g!G@xc~U=uZgv5>L&H=Q;{A{Z=a*c
zG1B;yQ_hRQC9e8X=O-E48r8e`sYZJJQnzzVt83Bk^nV-i<hGczZaqIteAG?ey=(6D
z8mW0hg|rUjb+8Cql9%<TbbWx;c`MaH`Um|Vg#UevktYZnJVy}L^dtIU>C1H--6Usm
JZDv{%{{?Ri8?pcZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_elementwise_2d.hpp.DCACD2865A1A82CC.idx b/.cache/clangd/index/gridwise_elementwise_2d.hpp.DCACD2865A1A82CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ea4a848e2d603ca35715cb1393355b79af75e193
GIT binary patch
literal 4026
zcmbuCc~nzZ9>?D=288ffUJ^(O0b~h8SptC}kq|S;CJ2MbVW~(J3y2FK1QnEO6-pgk
zkSYpAMJY=M)G7!rP_PJ<#g!^OC=N0st`I%6;@aap(z$u>pBet}os-|mC*OPD{oUXF
z-TS*)92^+fV2ojYiNV>aNm&U@48sWWqsq!$Wy`>@D>E@H`%qJ4eEJ<X?#u9=8OA5_
zAO0ooSJ%4i73JCdqfK|}lm2B8D{Fpa?4AGm`NCb13}#^Uq%cFRxP1B51HK|_Kw1{J
zYVVmZAACJBRu=c#;VAZ`E~&R|bD2EJ>&N<Q-d8h)Cl=~9Y_B<P?6q}b$zMB!WhKII
zBptoB8Byzh3G;pyUhFZU-doiw?|S>b@>cDhc_oelCs)f-m&kn~i@w)4HS8b%+f$~K
z$g021f63=##kzi$F_LZZ#UjfW8;R?}y!K9EKd<wu;iBYV!sOLTlT6j_$y8;g$5^o%
zyF3uE(jxARq2TaZwTiu=C9mm?$?m$DZ|Az6PdNH`!9BiOsxA-8z4DKvHO%(~b0Ugf
zI4*BfTb|On<!8(bIJ5C(UDk$uF}jToOTX<YxcF7Z@lSa6rwf`3UM!9LOgiS%P*;8^
zIR3jq^-uHODR(s6wDe?jY-mdOZi83MyGJtht3?a<EVvVtSCq0Yub_>;U_8D_^yc>J
z&=W4zB^6hL%>S5Rj}Ed&cRCvsO^)|I+;7nB_OmJqI}+FX{piBuw>|jtg{Dq@F4v-t
zl~i*e5dybgT>7xX<Ki2Jd~bA_iQlF5No?26&7Qwk*IfDf{6Knb_0!>PEARZfwB>Es
z3e4_-qNLu5Z{^!i_F?~zwd%>b&X^lJ-JG&N+rpnS&i!X^*7I}s)4uvNCw@W6-BZua
zHoAG<vHE-H`z-Sh%Gvfy=3Kj>9P=HoQstzsumB8$6PsRJ@FrqS8G1$-;cDqB)EG>U
z%qH^DndI0pyNgP689Ig-;XKpXQe!ZECYfaz5HNGA*Y$h(UCC^Q0oh+^p|sK%)IZws
zK7Oh1f{Q7`l$5>Ay~P@X<V}m84NYcqyA2pDQdY1OJdHu}smk~^@5sCYQ^rhEmYK^0
z8iVAHsx@0(M^*Ps8M>q_HJ1uB2FZ78PD&~&F3rO+F6lelge}w<On(MoW;p5l_?Pjw
zvIRb+rVM?uzt~djpfRXFmYI~S`d~oCNS~#N4T=d)xU=2ODe^G$u%ak}6JddZgA<9I
zL>d(Wrx0jd3s$w@K(%fKyH;=-|1G*E?GOE@I57a`17M4EGzPFVYjNs`Q+v{d;$|e1
zKTj<KYGvrQ<J;9G?stuFLTW7)qL|@?%vvTy0mccrRIWgaMN*hNlA;Jaf{ose6Y+RF
zjS}%h8m+<C(5MjbLZER4;1zVf0P`+@KQ+x(5VV3V#T}sRfB=d+LC{I*{9jM&y{|nh
zVK9k72pEJw<RX(8hJay;t$wPiTYZ=3!oZ0TQHV!=PSlZZA*&xJaz(jXvVP*&tiR#D
z5^#czv00krCylN0%=rn~4c4@Sh`x|j6t%g<ko~vtwn&jLE{X5umW(etm`uXm!rc-r
z3X_n`lGsw@VeMf<k>@N=t>*sl$VrRkUFqZtj)C<tn1$ZVB&xu!3hXH!1g}BRjyx3w
zuVIR1!WCG8@v`-rnMek~gX%a8l3|Kv^AFB%d~0L;(Id*!cdV>_vLl0u6QTB@o@hLr
zh_#QUQG!!~oYGR9Qnch_r0850XrDvIRIVp?LJ8?x*ypVtRvjkSGl7>dht`*uq9y;u
z7FD^Edj2{%v2o5uZ>rG`p!@;CDZUB5Hz9)JLEsKjQN_t9a)&7n`sI+kd`W<coRi#H
z?uA^!iF$CW2ih=AFuw^xsod=YNgw#2VZx-}q?@_y+Dpz%;jIWj$6#2vl-C_l8bSVC
z4&ri<p!R4SU6$tfM5%%k+re@>)eI;0gXMlY*MVCd1fbX8L=y;_z@2id7bLw9g!Jgt
z)6!6*%BdtS23|4DMZ!{_a!#er<Ft>?i4}d#^V3GfGsv|I^$MlME7mJkOKUr7XO|y-
z88>aJqPWdqY03r_mIJ^&05-^boHz^%4?`^C?H3#MN5WQ3C1!|ch(C28M2n(j6y<v6
z(zGHF6@gaW88Rb(yy%0>5GQ!fJP&QDA^6UG8o8Rf3Q&c`39+eIfNBm-_zV4=kw-WY
zAPk_<I-_-_l$K$XK_iusibh#_S;kb$^?K`RbPV*40gbBwuL2HgUEKfp#kQ8sC#Q{a
z<(8dASB#_q(>LsQxOPZgAtBSJ3;11NjfOg&-J-K7wK9+lcph-`K!60C$Ompdoeu){
zAe~ErTS_siR}7ZsB0w#8nmkQ`UfR^d4Z8Ws@JrGSl~ASKb4|}SHSr23>NL~%fA!*A
z^8G9Rtlkl_zmlmmLZzU^$A);YV{*4HPSk=)EtsS6aH0WB8tB{z+(xjX^d2zjq3Z*H
z4}ck^4*@<zacjC+<|A{ZnA{_0L2?%6QQNg0T-sqS;@0llpPrEaHahK$IvlsSOV5N<
zPrHKXcA&@7+@z@Zrn;nxDSuPtsoH4cM4C$)O{)iaJ<!y8P}YMl)wd56eW0YXZy)#}
zzV73YyEt5MVA?6&99SRYH0$TBWJ(2#g1x9+8zqXO(Pq(R8tnnm9?<e_@z~wQsO9^}
z1r4$baz`%UM5KMB_T<HhWp>MGTC#mIP0O**q0ugI*ah0khRy`?J~nmtMAx(t*>%5H
z$N4br44*A{9bUSsY;vm1DH601!if|`ik8rELS9X$?yuB>Nz5eH6nUC?(kPx2PuqcO
zIBRHB0h|hIzu`nD*mZ(4$~c_p0jnNxq+&G$=0jkIc>Jx^>L>ltc2mbYFI4+8S@L<Y
zTADa?zFmCE#(vreOgtSa9PJvZor=K!EcBBF!RPFI8(w~NX?pjrrd>RA;CFJ{mjkyP
zT+pP*{{d?grcMD>=I1K;)ED-Y`ST(Fg#Y(p5PP@HJ~?}EZ8FoC!#2S<NWd@?>g#_1
Ddh6qv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_elementwise_3d.hpp.4B57404BC7344584.idx b/.cache/clangd/index/gridwise_elementwise_3d.hpp.4B57404BC7344584.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc954cd8f5dbd8f28c3ad03b95b40f523be08df4
GIT binary patch
literal 5488
zcmd6rdsNib7RP^k7zxIo1Du&*cq1|cBLnde7-WDEc?p6%G&Gf!pkSyd0vVbbpaqsh
zW)bpGM3if&7b8f~^lG7K6sDjzz^ho8C}bHZ1X;hCd(WKn&;66NnzcA<KJ)#Zv(ICn
zz4woV$nfx2B@7d?Dl#`CH7A+RFpL@Zmz|T9KAXodfkq6ITiLWQX+yV%cqH~pi2YLI
zZ>HQlUhJ~}-%g9BU0XWT<+Tfx?+uOZ5IQ^`xA|cX&*4zQ_y;9#bez5%6l7SkxMKaf
zn;&FW?tM^LGobAH(b{o#-;vS5x;t*MyW@H;jC385Iv1=J73acGT}aJdsnT+nB4X;|
zxqpZcpD2oszrCrZK1P%2bkk<<PM@Ea9C#KLvhJ66m-Q`*TVK@eDR<%TD@#svxNTee
zZNu4no<sAiUuq8A6Pg?{rKIg((azz~e?si$sOK|88!H?e7xC+?`t3H}DY$F*>$QTm
zW>xie59jUUmR}g$?d-j~Ej4tn{KBJ$m5mK&431qnx+3-T!2=5~k5~)#e^k8ji;d?u
z1aw)ObTw(5%aZ$<^O0tkRzJDCvgPxp!>Ve-yT8`tnw+@v;oySdOK~6Xod4T$vr`^d
z<v$dNKfZQ}`P^sovRP-2=5D*U(s_;X+QZ+ki>vy>K6I}^)dTOc!tPTCpDsGu<aF{}
z^Nr_>H{WEWa7E74@Fb6i`6s`LSiW@O)>CG!x5R6QI^Rlibw(4P+Q@zt$X>P?kN7S0
zR^`I*6r;X`ANeo(EI4^Ysc2Qk8)diOEMHSrv|LcUdUy41#R7}kfMj2}$CYY_<x~3)
z)&Ilw?2Z$4RUPByaR*8h&K)RK#2UX?`<6p}zx7+!4BH}Hq_=Kr&aN8ycKAj~=k43o
zzdYFjM=WD{OD-<&xbnun_0v*2k0n1`Z{||-=%=4EBPtBf$PezT>iBZ(Q03|s@>d1V
zVQZJUipk!bk>U>wgLZ_MfA~}#jd-jHgF>uB>~syoIhOV#Sy6@hT!w2TJOL+r*?KwY
z8Z0X-J&hA2h?9*u+0)iju4}Lyv38$fQ>ClgkT*rhAomyCZFLP}%;$WU0R%6%y`LRf
z?rAoaH<jxjEC`nB8r*;N8m~oPPmi>f@Lu3#wUydN*I+sKPI2|`NeL2`k#NH!QzX|l
zjIjYQ(rMiAw>JLg0b?9nA>kQv{r#={ZFCLp|Gme&f*sGVUXt+4IXTEG$V1m)x#tRh
zm0Nh*Cla0!CkI#s*ytK8XQifP=W&;DnnzLjnvXntOj+b_?=QzTS){gC%kc_W6ks19
z$0M>xBhpx6^Bm38xKpIjgb_+)6A?Zn7NxK$6s5E26lJoR6qN#73KZ7^TTk;<2)_!^
zN5`yRJoL(EaV%;FSv$B8z77%BA%<`#$U2FRX9NPByWoSy?jjKA+ywv3A|G=fdqOI6
z6-6nc6dODqi!w!-6xD;M9w_bvWhZ!(S#`sVZcvimgJ3rZ&V(nxY62WFTiiNnG*S1t
zmmwbwK=c5_V1bVwL-b>sN5O9tW)po3{Kja`<N4+BC_SI&mrvO4LhUDOdZcbV7R4%J
zX9ZsW)a|R%^a2(Yf}#-gY(L=KY`p2N23Uj`B+%VBu9>0v?DP}AYl2Cired4lfs%6`
zi#hwdK)MT5c)j>ubLXhN)a2y6k+&5gUvpnOLW!b8Ydipp(naYMm4c`gC~k%s&EP_w
zZU<#M%p&J{AMEaf6VZpkY8Yl<c3U-}w&iWR$mb((Yj0aT1wQhb?lXgsuf4AWA=Px1
zUh}-a|Ah65)(ysdbR6uD!*o0{A2onu1IP$Bf^#D%2oFHu00a?!41ten9)+2sWMzCb
z1~bQK&g0F@<LOt<NBO*&`Go!CDNM4-E3fH}e)a(P(}re0rg;?PqhyP;hU8;3=keru
zJpC4G4axHf`_27$ZsYF`CV#sfP;V$&eR!Z~BcDa_vUqR2eikLkk|;`3rm2aRq0G=z
zfH3f@SHfkRIj0Vq6_ki0Y4AH=`~Q3=yNbJ#Y0@<P0k9}TnxUuA1nb(ZPmMxfU{L|E
z1$5*&2fpVZoS3`^s(TPYcmOO0h@Y@rK(TmC^D_`XBYwrAKSBH_VeF<{Jk_SjhW*O(
zsV#e&BW_fxE0%_3bDN<t(jdGR7PWwP3-LM@wZi;Xc!}JfUhwUOIe2z4st-1pyX}6T
z+mc}a;4qwZV`5d(wy?5z?&msiuLCdqd}78)RkW~18yBiTT1Da>?r9*<<!L-?7S({b
z1{`pI79EB8M_~!(<pC9<nkf&~|LvSD<?H6T>-*geIH!pBjn`i`&e?pE^n`ObPoS$l
z$fC2=*@1Y8>n_&Ktf(4UHt81^7d@P(Xe%D<DFy#(H13N@*It`g@7Pe1ah55T`4K7v
z??Rx|3Q$ylUVk^rVCPWDpMD}1NnNG0uqhT<xLQzT>1s)lr^M3+=L;5jNxW=u6~Ll-
z@_DX!hgcLTkEAF{9!1e6;U)=r@^#_s6y*wYDcU03LeW;?R*JR>w^6iRxSgUmg>O=H
z9E^^G39;A!Yy*f0H-c#+NHOoZ{@|DVGjE)j^wX``{%%RTP!%@$hORq(Tlw<=FYbzW
z!SpUtGabu4GbJvgCY*Dm_dxs}9Rf>0Ttf395Fets7R0rLo34l>zO%eD!dciN-x7eg
zvgy&5Ch0p*KGCgkGHb9?X?NR01y}nx`yu=gAug`Y1UQ)eSoE~mm_^MXZU#F%F&13_
z@dcVMg2hErbg<}4u=$c?SQd4IxSM|dKA7GID>B|7m=4lB4D2wN6MX{M3BqSLm}mWD
zrE%wC(Pi+z4B;f2cYsd^1Y<ti*7fpn^>6)?zJnK?w|lBSi#j~%&e##)v^62trgYL*
z^xe(Z7AvP;dXsbR1&Rf}BrYyhET!meP`nKkmxH35aF_dFPvc7~_8PJ%+A-RboaYkR
z5`79~(JIGPl(ts3meO{DY$r*#EGh%1GSF8y_!V-hK=-;z5>A#|mOC+*rOwi)9~Nc1
zWm8&?TMk7nplN|%(z69ZS|E((h!%(>+zY|I5JGcUFJ0^OLS!$@$2zV`3<bK^XENPl
zuq}pJ_%Rk$gKag<r@{L)L=n9eWUb(j-JC@|;M)VySjR8VTpHHCixV?B>Ky$(^!as!
zJUeCqlkQ26=8eXdrSE&HEhhuhsJ+Sv%+S-&?{;ejt|vNb%L{3|{xy^<FQiF&dQy4%
zynBO#Y|^Lv<YC`C{jGg<+A06v3i#>Uk*8!=KOOnom1^VTN6x$VmU1zaD^|sI;D-Go
xD|KD6Hn?YJt;|+yzlgNWe?O-93&V)XSF|Uw{KlM>&xCwavB*sK57qyPe*t`via`JX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_elementwise_layernorm_welford_variance.hpp.61A857F748242277.idx b/.cache/clangd/index/gridwise_elementwise_layernorm_welford_variance.hpp.61A857F748242277.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9f9c5a8132539edce44bf30bda39149c7299162e
GIT binary patch
literal 9138
zcmeHMX;_rE+MXn%!Vm$00Ud_0D=x^s4<L(BTo6$<w_{yOaV=7pwrZ($L9B9=mSTZo
z0fl-*MG+B|MMZ_NDkv&=Dq`!3QtMW0)qa`zzD(lPAN})P{pDQ4bIr}2<jwo!NuDLK
zQ4tY4?In`1Sy9X8&;IaTsYD{_NB`13T$J(<N+eA(iDX&+(Kz*j8?I(Qd|Yd{VfZBd
z&9ERSrZhI-iy;?FW%ufv-J5z1GYQyN;9hyJF6$@hyzFhGic@^5mY+Lh)cMJcWw$?B
zSiXD8_7SOYW#O~EC&CslpS60__2a`fb{*~d<>-OFb1Mg?Y@3-rH{an3$imBdwKmpT
z&bu!;)w>#p_3^LR`CJ|bCk|hqzRt`yXZ858{s-VM4h8K8Yi~Ksi3~LU?$8I_=WBA0
zrJ4QsT^^jbSbyVnqb;xF&Kzq@Kaw-;LEjI2^3&>D&iR_|%wF5l?z&)!>!zxic!tTU
zhPs83SB_=*Jb%Cbc(R`3Pk~#^u1EEKf66$YGV)&At9gCiH7#24{-olFY3t7~X!~hk
zdP-OIpr3P=miG4C9r>gF%!=SdSYqV4`Zt4LUZTZ5FL&;B&MQt)FHpC;XU(0Llks5I
z;?_$073m2R=UU9ad!Q@f%Q<7>HY#@QJe%1Wu~fcCaiV2LLHqWECkG7{{aEF75TzC<
z3);3C)cmL1z$Ghj%5OvRqCe|QZjx4){*j;cz~{t-llJ2-PYyfM`=iHYv&#CpSWoJi
z{lhEsJ?RrpX1lj~+>-w1qEnJh)5UvFjgGEVcON`zS#rTJYS-1Ce8-ZGB?lYF&wi5f
zP1G-gi@%*(cyE=7v)-@u_Hm{5t4E)8D+uYk$l$|)x9u#Z>Z?M&T(hkvF{7no=lETY
zlR6&cY<y{5+m^9n%jNdw$l_;94lPxz+~2Nx?j8U0CvA?(?1Wca?`K`yd+`Is>tPFD
zede?C?(6E4=EG~I9(}2H{jz)FfhtRBL%yT&nO>Xr+PH)?=hXMx;X6&=Zq>gX2b(T4
zL6gcdjQgx!n`$0nJRIszblr62#RjAIQq0FZs*ul}ra#GC?s+@qMbhB!s+YOinmv{e
z$$F=>)3vTBea|`7g#DK{MC(TYr)i_72ZATDnb$qKH%}_Pxue@D#qQHtn<sV_$C$kC
zFnPU4>Jc@=E+M4kb+*%=8}C*t6JIxPvvZbhNIsdC<rKRo<2Q?{8L!u;eKdcDn%ceb
zpC2`()kY1JLw$vWj8lwLYK26P?hX1;7VIe6v@h_20W_m@h%CfbD;VvydyU^Vd3329
zB1$JJ5?!@|(E&HtZ!^67bb|pjp|rQm+fFMO9o9ef>Cb(}+?2z<l#aEE9i|nG4*%hz
zea@?(MmaQ~^aQI3!?c3Y<J0@EyRmcLd^xnIbdVy*St}SlwPTe}e&xqY4WKclePzB5
zTEXb#Egx=D$C=JCfO1Ml$RaGYg3)v5jsNOs>wF_Qw4=1A!qZhN7(L(Syh-0(3!_b-
z5v5ZtQ;AkEda-JM;ZldT<p!`Hr2}LE_FBQ{C0i<5Wslp`a@dQ~35o<atzh)B8(W<|
z*1HuThlZ4Pw{myX3Pyjv`Lklj@2aTI`%^kek>souj4tqX>+pQ?)vt1BLFsr!yt7s?
z`c$pIYht5GwjAnFI$9B})CxwQyZ*3k=J791_lA82NQhG(rvR;x=tE#GkpKc(qJP@^
zVW~%*3G7Y#PgYDGq!p}xtAixd^@FOj22e`rC|Q(3D|GbP<@ZY7zwaxDn9>uiCc0_`
zqp$wz=aAGe_NpApC>^7SQECOFe_i#-fTY&X%?+R_rNd-l)>^@6$)efsrOmLCNbu)@
zc`v_>UkWg(M9!7Sttab7++@SOYx}xpd^6`Ox2x8egqns9V8%_6OQ;t|;UmH$*t0N+
zwvKjWTEk?V^*BeiXE2Gej&Wp@#w6Y|-k+%-lLXHMf40(?Bzh+Lv)zJ8l4p`XTSrVb
z1H2iSFq^`p2$&TC1;-@-mxz2hz~v&Y0=P=VbpY3i_%y($MSKZZTmn{n{7b;&67b{r
zKEU_I_B}AH2Sx=tJg%~A^7O{!Ayhtub{s#2qn^STjt`?Dhmk8E?+9`^g1k7cM?>mG
zz5xwyK%SgGhTM-KUyiS!Ay>rq?Pz$r*uDd~bs!(s-l+e@P}d9B=Jt`2<ly8HLFzls
zQSXeNFC}+C<Q?!1YbGTbFf0RxbDRsqaz$JV!)irbiAGkU2xd!C(vHI0(MZ;x83mGQ
z4^Xy|Flhm*77)amVe$+rpFzU3)&FvZ-pkU1?)~U&<C5d~Z0~}=yS$r}+yj31fIk~X
zN`41!zXNxUUjVlkB3=*u)<dCvCiKdL-n{)*=(knmzky!gi2N?-wM)e1(5qa;b<nR)
z>|YPP>P0>qIcKBc{CU|ZEE@^hK7ytkK~p*3fWjNNHU}7~nnvuXZ=_pzEpS>3bgRpD
zDU@j)5m$5MpvWM0sFQ}0opY{QIMd!&fzwss!4}Ofxbn~=-}AD!+I~aZgE}9??6p}o
zh3YF(FLD4=qtgoC!mL`?1Suvz0*fE{B4E-3Y<oZmTRlu(0sIOK<9H*)8=($6&w3eP
z1y>vcQnCSz-T=n3^>DlO?QYBIjSHcad<|4zgFrTdl%CwSmB@}Y^cvpv-$d-1OJ^9Y
z7i`UZKRZ$Brqcu;+#2eDX&o5MM#JP3Qk>#Ph_}u25i=ePJVZxJu}c}me44N049UAj
z|5{9^c?69-f+ARd-_Wl^_rI_;`rBy+EX+vWem5f(W3mYh-2?*I2$*C8&upNxe@t?K
zTaKu~0x+xqsCfG_;7|sHL$LxlREW3+IMj&vIB+;F;$|Rk29A6@HIS=C+ysM~U^M5?
zz`!$bEW<2vql(ppZCtUzHo?A3`<RR#K01QAe3fC~iGM{ne1tFwaSn-L8x@oAe&L?n
zowfnjHZXxbCnQtt-_w~^&`YB*i%nn%BQy%L$p#)A=K!l5Ab6D~IHU>c3<i@kaPS$Z
zQy_~id&#shC4YfCz~>Iott_Tqz#{{O^2g*tk6aPgLXTRG!(%N=mOs-A*T>{6@Hz{0
z$3Ld-#q17n-~;^*@b5s$aRxNbfHoZGLgQQ!?}Em=L|hAvYen1$-5dEts7XV73hKrW
zSC7az5#p($CcO?gtphrD%c5whOdCCOMNIEMU1#~2Opl&EmF)w~*lwrb2sYQq3;6+g
zlSZads}9u*H4#?UD?FI5>?MFN0dv*~lWWlR8ut)XK``tN9LpRZCh2Hsx)>G|qanqj
zTQ5gL%0*m_!m5$bUyTN<kzgQgD6CE7yU?I6<i_R`6_IM2lneJ!1?7OS9Bu+*eYPw#
zcmCqvl*oswd>F}kj2-7=ae8F#c!)_E9970Q2PQReR1L>$KTw~fNkh1b&jNfFShMls
z=U4cQF)T`;;}t?&2zAa8lM09{I8OW{*X`Ef&prR)2xf{Z&$Ou5odth;%CXd{k4&3V
z@%Oy}d~bkQ_P&@r2BD9^6pmLx=T%(4m~4U0TSUAax@?D0y!}Dwa**R0>Z;HMH(gL4
zObS3y0T4EA8SpOS!4cIb@U9SX4e+iJ@p0gNT*S>_P%{|HAFKw0)Ev)B-cuj{bjQzh
z;r#mfDcS2|qUxtovRscjO=<+1Y?3SG9CSJdb<s6TEa}0j$trm>KPx{EVPq?n2OD`_
z_KZQ{U89!L$6tW97oanr=S66Hk>h#CN(bi^Je~X(|HV20nKn&j4`mOabJ66e?BCSI
zs)FN~QDZV@_?Sp`$g_locG2Xmyh{}@trAb~?~&qrzWG>6+gqkhWqG&Fz;rXPWb@Oc
zK9)tmp5yz#^uEY<K+*vRa{eK-e+Y-NGbKGsQ8qt*g#+EnS%`jZS;JNJ`}MoVrleBh
zD)77t!dYCwk}&!{noOEc>@+Yv4TJ}EB1I?CX@0{fwE98TpT06o+~w{{_IjAa%43zx
z^<XkVK0(ROUrZ*-Cn{V2F9F76iS-g;_%*<A4Un_xYJy(FwIW^z4A+VHGQ^jmDQ};L
za30d(jd%8?P6+AF`df?H+<NEXXNN!8{TGYLAv61)Sf6^ED(uY^I_JG(w}S@QAxeCK
zjK4sH%^|;F=#Aya7W@3ehZf7X^<~;J2VcP1irLQGp_KsB5^*`pfoZvj+o7VJZzLLg
zAeU?;959856moOM<N%^yoo~UU716H_f7{(~e!sFg?JwFZTDc+XjG@oSzv#QDb>P|j
zo<eup<2mgA9O|6Lu4RXN#mp~?=#7cDqqj4Q%jiku<rm1$3%`+mV>wcJsRG&Nrl-_M
z)mV-$0H+H)jM`f?21I1<)_BDzt0-r-%@1^ce83{}(ayJd(i?eF^?TRPgv*+{=)RgP
zoox8OB~n^%BB|0;9XhPC?e}xWvyJ^QnT=<ga~o`hdaY0>^Rz?U4uyEB0~&Wgq3m-L
z8s8N0EogQN>h>EZnMf}a3F8$a<3c3#{~qD*kud&#g!ha105U!xwy#2FReV2dawdFA
zY~O-#i-=p1ajV$>0y4WG_U}Zvlea&z(0tKtMX(FiT{m>^hC%GmreOkfehmE>K61J(
zsXXBE&9}DqhV302w9?I|a@6*}938ASX(ZEDa`_23EqYoEx4!d0b)MIgG|_p$H8Fj<
z27|9bA)`-6-sxfyE*<%&BVBrhNip&&=0!IR_A5q0zFUsG%0<2$`IU>j8hNUbt{#Pn
z8u_Y`u3m*n7jo|sqyH}C-Nol~s<lO}zW1d5ANDrqR0EgCj#K8mwaqtdli3`NFEv*N
zd%qo8H@jwY6*E2>5^F9BSSw7<L#Ojlm!Y;?TAVP{=H#c;2Xz3a4&cp#Gv<wGf6Yw-
z@1KQm7Sg4@ET*C{oaRP>H%kslp2!NSm^8wOMwrCHl{T0ASNF$F&(_0aJ+NO7biNyt
zt-yY(i1UC&o>)2A0hBw$P<Ico*#m@Rt{PZZ16`<yNdr(ei0!WcdjG4d++KN-ANzi8
z=DL5_lG&{V)ve~9g)42E?=*k5a9zIV!hYM;AG|g9H_W~Jo#*3nQ?@>%b<s^Aa1#h)
zMx`lQMP!3995=z3CK%808943?jAQu03o_?!YrMTSqqQ#J87Hkc0E;p)5vl+d6(X(y
z7BwP14lIs~xEbg*12gs@db|QXHOH)ILswsOjmB3$aztbt%VZzrAF6ZNYwI9?do>|m
zO-RXI^v+$RD^o~GH}dJ`u1QKBAnym<^+`z&8rj2($&c=JW!`M;D*b;ic|KbkRqfEc
z@9E#xZ+>t4)U)MLMJ+vW*8rzAJSzC(j~#U`|IK+qW3)2pQU-OGWtxhYOAW`1X3u+9
z6aA(wnwd61^TS#rv>!-L7&(@LX)BX@+bS(>tgR%rjFL!f_)mII%RWl>Ln5){KmP|O
COA6Bf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp.6AEBEC03A02FE520.idx b/.cache/clangd/index/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp.6AEBEC03A02FE520.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bed1f5dbedef093f563acc02a67fde52d3b8cc03
GIT binary patch
literal 15286
zcmd6Od0bD~|9|(~bMH;v@|JgtLJKWQw^F1cWl32IA%-!QVeG~_%x5sRB&ifpvKvc7
z+EMbZh&C#PN~KIqQPHGioAEnMozC(2{qcExACJ!;dc?U;=XK9L=XGAM*Lf|+G!Hko
zXeTA5G2eS^UbT3`0;r^<)C>JtzhT|7NrIA6gqf1k=E(cAR;?*>m%a7+YwxzDUiC5Q
zE<2;#U4C{r6yj!_Q*rIlLaX`J_E(zhZbW_l`|&(Q<FRzlF^_s0X6MD*rr5^Ezd5z_
zm-dQC6RTb-r-E7)H)SUA?}JYK^Xj9E?3ux{lck%Y^;D8Ge?GbTTK}BGVHfj$8!&Zw
z&GT5ZVe9`IE89NKbNShae~-KJoBv06&wk$Z9rL0Um)qjK`nOCC{$>55R^!n+CaZO7
z42mvJeX~T~`}jCp{~nJsR{E-C9`N~a6b$YaBTj#;(&|1oRH@&NpbNS4_k5Bm=H7-g
z%e-7)HmK@4?}>;$?icM@6)wNOI;LUn*%Cf}4tqJv^nPsixn=Vf{0Y~OcOJAPZQ!!0
z3uaaYE-$UoiHpefzZ@1fb&OeBMDmVZCtS}w`)%9I&)3>@R!v>_`bm%S!sYq_<^|=d
zO}ESbzB(nqBJ*Yc>bV}8p=E=wUcV^6cJq1MwBV|dqpg0EeBGLkg&vD+*xv7FOJDZE
zTI2qYml_R-*k0wmz;EBL7Z>`k30u>~+%~8*`V@0d*-_Q%K=0k7&dxv5xX7^W_S)Yw
zZ6}>`YUP}ii#CRy-?(CB==1Sw^L}mkaYD)X?@bMFZQAO8bERAKs(RN+o}Nbv)PJz6
zy86f1G@tQFc`r=Fkz3uovNpf-UwCM6hRkWTQg6GbA)(tRzZe<E@4S%j`RRp2dhpDP
znTMJ`ZP0j^vUv6tt%)8nZj&429x*TH7%X@{a8~G&b3=M)t8E^q5jEhuhXeDRj2C>b
zd?IIHf%A~4HFMsp4{7)Nv{3(Ep$m96bXRz5*wR-QM{o1Fp)tLQFV|FDUp*x5oMrQ0
zT02cIFNsr}TID@^MV`sKwvDbo=6XyLg6w<E+g+%dz1L{p&g1*a9{eF?upfpGJ@#C$
zrGDWD2k~rTs_CW?yW3tW^nG+k87#IM`!-=o(2jKz%zaIlxlHqN^ZnUwNM@nCL+zxp
zNe<epd?Ryb+6ES`zH8%jz<Sr}0eLZ*vlblhZ&8wcN!KDquQsvprAGLoNA-pG?~EOK
zVuP*W<P6iitX)1P{Ra1+`!Het^apddgnrVyYf;Z%OQ=0)@z)<Ow>qyc9oWxf_Nx#2
zO7rEWC5gR90$F30dd^=BI_qv5?DhWU9k{c^Kx@Uy1%7wu>eyEG5{7Xxi>wR&ntpRG
z+_ApbinY^bO|P`;2gYU^1msK&e17<d=QH<rx!~%}?a7wUvuk{+4%<(?HhI~&@M~v}
zOuJwztWLWel>6e~(1HzK6IB;PWSRIbme;2IUX;Uln(PtScQr+WFBBLZZQ1zPJ}*vf
zwpOrLeB_k<IvZ2xB;VI*v%m4s)b!kTF7xWNDV_;?;vd|1xpv^yE0Y=jq@>uUq@+}>
z+n%v_Hn=}8=TeHh#gknd{QD_68G@?N=vLQ{X1C%-e*3;@x0xjL^^kgq9t-_LqMG~c
zDH)-2IBtgZ&x=>gZ_SyR_VTdLcV%<74mF-Jrg!j<lYdIRp!idj%l<|0+?ho&rw{3|
zdS+oC;@6$^STM}&&ea}2e<-w4UcRyIk&9(iwCbX<S`|5GS}c_ACJ!96bkv%N-q!Lx
zuCW<Dg&#NN&hjk1zVnCBy!U!rYI+5vDMbV?2hWO9rv6m;^SH9okS3QWPc2s{zqIjx
zaA?Z@Qzy>UPIez^XtlE}!hio`3ypaefe+X8IX|#+rQT-M9Ltnd<C99aZa7;SaL-4h
zkyov_xY28k!K0CXR~`ti<Li4>4?i|D{PWeL7fu*IeLiQfm(sT<Tc3Td{pWq%`&B3N
zlX{<T*|+un^9S*({_w7t|LXz9ez(?rk27gLuf0N_o!A!f_HFBc-A8N-CT-j@i_LEG
z*eDe|eVDUs;T$6`?=p=o5qUGSwCs}Z+}fGeGSO+;6Z6rlJXUC4Oxlp>pRr}%^4dSk
z<%thohG!2t`Q+sJ-)21bls)XJdr7Hf(ei<@HTQ~>_fBu<-}lP;txHwwOD~N-Z+0nk
z+5VAnJ$Dbg9(e5L(kDw*y{)RfqAO1A_KXSeUe}uQU6X?<@0vfZHPY4RY0m1=zJHpW
z@vJl2|KvaJ8WUy@Sn!?RZpF$vs~!(aVi$Nu)aI>Sy7kvx)%ufK`UHE8Qpyh$8jr4K
zMueAVtQ=Sv=9^@fSnD!JJ@L2SEXS9vyKHG%rqX{(``qt;FJAD#?#hH&EA~xbn-;vj
zyJ6+oA^-Vsu5|AUyU&|f`DL|OmcB`N_WJc0{kC?S4f7kF+s-!YWhn6l|Jhe@ea+r&
z^1DBd(ds8t4}S6!`*dQz;%hzshMY}lqIch4H&ov8TKIhD!umYZKV%>02aQ-~`?-94
z;g*{!XPSMI9doNQS_U<|K9g|rn&pS^THQwXxcJ>6rcIw#hu&;;zp7TGEWci|-eZq3
z_lEJ(8n!*?KtO)V!G&RF+jou7)2xX+5q}}@E;De!-`kE?)JE4^CH&d4Pt(!U<|qHj
zwc4*u%(G1VJe(Wt6>T@Sf3x-PPgUWvl|=@h2k_GmH#KOlJhkckEp|=9o<q}h!cB$*
zd#R_GI;%dFKaJ|2aG_vh!e=IZ(_82DTUIR`i4IskU{B3}ods^1KrmL4QW)J7re9a1
zq>Rp9^dlHnp6c;&p^66Rk4PQ^_!#u-YA}hKtnrDZ&rS^u&=Zl|VbJfKt_G8+_xU{a
zw8^$Q4bTgbJYwJ@#-*#lBpRPyjVYYlv``cDMI_f4-D^zWt_G86XRJ7W!z}iesAM}(
zNgAv?*!JJQlzKp*WH*43m_hGdo@$zU$W=iPG)5IrsRARq8a#${c3?rXx5*6+ppQuI
z0(ci#b~Tv9VaDgtTetV+Re>Clq%$h%jBZzhNt`Q|ze-6@`A!39Ad(u!riPi+)nF2r
zgb5}=KQ-*t1e%B>hS7~d>*;JT$&}2IS>KN<?4t^d5Q&0OQ80bF8cgEz)0sCef$#oM
z1qO(u9;nm<tF8u<d~@+OzxarLrYcZHB;`P*95{6~nB-ff6`<G1rOG_$jTjfr2+_=_
zu12W`%b;yW%$whr_iR}?Y=<V`(Kt65-J3`lIvYIB_ht)f{$6Bcp$T*lNj#$)kJi-L
zV3Oq<e_Stw1v_g3EkqK_=*A+F&IXfg*y^xjr_#Y^8bF3f@)-0xr>nsv8>j1gHtLm3
z)&$CkB$d%kMRVzFFv+%$vBQ@J4KvXM+K42M(TzhN>TEDc!2G`&6hpS>Y63k(lEmmH
zA(GAplN>reVD6NFwUMeo9g#Eul?I^S)nJmt|J2C#ZGDrZ33L%jBBPs#M(J!Y$wlEC
zx!>o4d=1bKk;DTS4;;E0Oj6Q2$am+8>#7>S5Ro(k*bIcOrkmt%UbD42e2*$nK_oRm
zr3OiFXM;(~v|W1DsauK~Kn;=X1aK$l)zx5<D)scZP4n+u)&v|PNoI7D5$ie|O!CL_
z+nbFqKijPVL_`t=(64(}gGm~;?02qi+dM`S^g$#EjBWy2bZ3J}l-4a?vfj&zSP@pG
z%XVOK6Qu%1yMWQf8!k%4tV1z70`r1T8EP<MTy!cFr4%qG1<b>I6s25Xp9`j8Wfi4-
zI65DGi(eF_ByL<1Hy>kcq<~77X(BUd1O|=30dF%lRcNZ6c8U<zG1hg=w^*@7DT=j@
zV!OYnX6+ZT{IJbaLlNn)ehg!b)mucutsl!6W4RZR4C==*##lT>q#ye6j4@Un5phvJ
zficDm7Lh3FCo;xZ%Fxmn{UpX1%dUuQp#C++7@J2C?WI97V}iFCv5qlFVN9?o7m>;u
z++<9!Hj7AY3{n{rtWlzr#q`Z$j4&reDVx#GX8I9a!1OI(jL4o9GTMbqKY}Y5y$Z$<
zTXa!+#^^nxaVl_41>?z9-U6<-XnY^I-Us8!))qsZV%Q&Vr6@gxI!|d_!Rc3U#-zW3
z8(G1PA-I|wRn3hfxRx7J%XtudgqI!RwFr*n2SoCNXgn^GpGa^PKOl?l&*H~r(fv1s
zfj0zmOs|%BQDMgYnA$@`5(V!t;=iXR`b$J64d&6fn1RJKu3=yejduXJ15ocr0T@N&
zH~`~loC#nijUNH{h+w=yNYp#eTO#UhjM&Dw6H(8GvTQgBuc)U^@o4L!$`w7J^nn@r
zff<Ggp%erM2SH0r2&GuwF_w2CIG%Tmr*Q`Fm_g$^yyG1jm-3FKG=9oEKBe(L{J4L3
zH#|O80i@-f?g3fRUFdxm&cKA&PheF3^$%iBSSS`JA57AD607{_!nGkP5Xt<=Y{nVO
zJd_HVk%f#imhp~-dw*c2lZAg`Og=HD1Rn%W2f+wzjiGd!bw16G!rzC|IoA0cjU%8<
z1RO&8Gq`CPoHxO5xWR9@ZpQZMIU?KTd|D<WN{898huLnz{!eD7zE;F#v!c|%ST-=$
zct)ahfwjEATH_H#sg2cWV>L0>e_@)`V^h^AG@_Z3nF^LsqfrSrT80;>eDxeL+dBMY
z&@F2pwD300u8o_H7bZ&Qc$;(lRDu(En?!=KM~Gym(;Xx-vx8ON!MflVEr;w++@Yat
zfH<UJ`YIR;?3`JC|H-Kz*qVXVAr$ls1-;0whk~J@z@FeRpce-E(by&o*wT1>81NwY
zHuueKZZ4hyl(M*aS>#ZIQZ6?ym&R4RaTRZh>9GYxOY3xF$<kf`gBQR9`;JzX!`)ZC
zs@#W|SI3#waSnI{Q98{Vo#yQbJ|XlyAxL|{as7#}w0^Pq^F*}X{LD3I=@Gyv0t_Hi
zzYh9e2m08^L+J+Se}l%^pno><q`I7KD22dDA#gJO5Mm{q6iVX=I4c6qCjAB6=mKKo
zp;XL`E+%-0dyPNeXZA1-N@1X97?6>Pg#qI*U_xW3Fj`aZag*+G)I@T)DLGuX_zab*
z%V&w!N0AgXaE=XBN8_;4mD$ZxP=`4**Tf1meDsmg@gEF(esvsju=Sl*dHeS^fcync
zp3GTc)*^p^lcy4lgB^5CbVfQviXwp+3HoAIic&P_84dKY^XRn3?}y)B_^lO1$&a=3
zW5<#&__KEYG!9^;0M?oG2eMXy#5#&nFl!Y|<NeTTKkT-qq7(?N0%@EJopK>{V%Kqw
zb=(9p{%L;bX`b>3B7`v!f;X0GQ92=5oe)N0JVJfso$uTRZsC#Z!H7>8D|~W_QaL9p
z=WNNq)ts!F#<iTRmd3|;*)g7?zsJk&5sZyBI+i*kM$GC_js~}_H<rcxpB2<)XK%j#
z?U_Mnmp?F?9~d3{2~lcg&01LtJaz1?+NgAep!n0GbP(7d1P)}EL&4xsAQ7AhY!ZPj
z)^y}P0=q0Q3~RP1g|pMbS;`)tW*tux7hjYjpk)N4#7_ae738QErH9b`Asmb+DoUl$
zyp+bZa8xaHCH=K<Of94~w-%1Cg_OHj3n$hR%YsZE^r)kS<pyVdgQLE8lQX|b;|$I-
zgPTIWm&JXPMU1p46>?(>xo)3CltTGYp*%IeL%i1^a{i!jfj5ld2axe%`8lzCcU*zO
z1>P`@?oa0p(`lT+&&i<aZ}Wz?>HZSlu!P2?{G3vnzKl04qx+lrvCYJ%5~XI|y_u(!
zr=1_r&QtDoJ8#-fyTk3gMLXZ^s)$lMZ`ID*koEZr1AGPQeP6-USD@T&U%|pxpx*Zt
ztb7IP{SaY52>pJDU>ZWdA0k+U(C>!`Rw4BJnSw{AKpBTjVREM6Mdp_;IOGf6KC7#L
zuH&m`>yIP%-H$c!V=3kJXAS&m9Kh-Zu++o@S-n73A_E7rdcic_5B2s#sy`6w1=2Vd
zn&v`B@_vM16(NkqR%z@sy>q|3RUX^3!w0lI%u)eQ7AwnQ4akeRtSpztb*!w8#xGgf
zOB#njn-E%~LZMA4ja#6+1r8+Re}eK)G(N=14{=m~GbeARaVW1J%2OI2!^>iLYQE{b
zES<(BysU)a@kNnN$G#c2-l8K?6|)#E?4@@^s$w>c3mCD0(ZXIkQcXrIq;WYTD`#|x
zIZp**D$pXPEEPzpU^v0IfOLz-3Lq+I`ZOS>(fB?P?*lDj3hx8yJ{V5$10X)2`}2XA
zPvda5XE@u3FuD-R3ZX8+MbM%M4kfq@T9nba63QxR`e#t~jK+;n-bhE%jnJ+UQs;0J
zlsD1+ub})Djo(7~+b^-*Tl)QWC~v3x-$D618eivR*EwoE3Qne=aT+H}qj3c%ui&Um
zKs9Gq%~A8M<?L!{9LdWgd5S)ZmuJzqnRjUBDaC2$b=r9&!cSj8$5)_mh@cZf<7B}s
zS?G@1L@85n&J^6q`}u-FKFI`lxUDsueH6r@Q+z*k-47{;C=j{^(l{4R&V}7)%w+qJ
zHCmPd!OAFC#7>WAzawABW9{;2&*&jL{UP1oz}hv?{ku7z-P}U*ek3<MlA}0~&s*n{
z02?_tymdLjlfCS<v&J2rit=#@?8F4tlMGbEPAsBvFy|ReyknHu<1LE$ZexdIC*)Li
zX0?gC5zAV~(lah#tqTaAxhmFvlKPpBIPX_Z_A5s*-<Q|(<wuek<?!}7e0RivV;vL>
zbtcKl7x%F0dx#yM)%%{!{Ae9x<TBl3dfsC!h<V9pdge10*qqGzc+}?4lwEdTT^a=E
z`_WQ6ll^3=pMm;k(Cs+S9dWGp8LocyR~MY%2_r|ivlf9fH1pzCG<)XWtw!7bf6H&+
zY-bOZu8IZzM8_dwlf~GPUCw4kWiz8`JT{vdPjCZc)4<pfNBRP5bAhEo<V4mZk?l^F
zbtGD}+Gryl0zE@ucR)G+x!(``GlNPLH(p`Pt}ym^Vhguf_9#14(+hPBf`fE$K)EnC
zY4qw^EB4<|U{W3Jh@R%hpQiJWCj{3M!Z_?N;QTUF=_<(J4~bF+^v<9??r_c}oEuFv
z=@vKp7B>fDoX9p-=_=chfwCEoY-S>1RUzY1$V|ivgRBXoSI+b!=WQjUS4ra_U>`(>
z8o|Imn8u-CS}2%KLXRY1lSGFe=UB&cbZC(Vz0>H0)kBYZ+ST97jo3?{7k=CbKN|aU
zBm8L`z>Nr?aTw<kMl<RN=W>L`DctN7!Z=a7&CR^c&BBaCMGWqnTsm}k#EpGKhYq#e
zs9KH+9U^)CNaW3TRf<F@l6Q@y&#y~-pG&+R>A%F=Tq5BmIz9L~@wDro%=byA>63Yz
zWSagCKj#ijpT+BE(cjPFU9;%#7w~-w=<y18n*w^gr~I6!^mzHgq<o?K5Lp!ymi<H3
zgg!{(OF+*OAmOSF79ez5bQS}N#MLv_^>k6@3Tu6Z6lHKt18IF{ZIHZJz}OTpc6j6t
z*KkxJGn&R@3z_i*H!#B*7zbMGhh1Q)L|HNGQT(OW>$S0z>k$e)Ly6Y^(o{96_@TQN
zT520Rp^f#ydxon*f=XAZlFaxZ`1T<9jtGAQJ1v5at0SOy1TF08&^w*RM>&_Hv~V8d
zT#gaEGkKbYfBCg-Xh!B%=A*H>*_jsKqk8hPwdkMycpX393{P~wvFntevhM@`Yi42p
z1a@~*(C6VYQF_8?JYo6}^)F#GN@!fe46311)=|JC3QWQIBvHBy)Gq@~f={vPr$`E0
zl+Lp1XKCEZYP8apHyo;m!|s$fQU|CN5BuR2A)f|nCD6DIYScl>{jP@^^)$}swDLLX
z{UT1Qh{lz?RwduvzkfxwUZH2%am31W!8~2)&c*M4Irj0=CF?%@&y$t7M2H-L&Z0IE
zrcCIaNrdCTACX>`+h;FEQ@g@ATw$CrVc>p^At%nRNkaMmFKs21zO<E4`qEZH=}TJ)
zr7vw$2c~Tkl)kh@9hkO7Q2NrgbYR-H1eI<XuBMdG)FHlF=9XA|aak8R5}l=cyf9o%
zMXc$p)Z-3RrGwsg2y4QYWYjtLF=Elc4ZN&@L=O=`UT!{PGCfg8JKMjVsBq+Qi|4;&
zY_$LC{6ZYZc6b2)-asI0t^fuVU<P&^MX4GXRD&6~HX=&3z@Qe)DCxR?aLMPPv&yv<
zmK@3z1MO5`iANSuc&vSk#`l5teRO}OE6p!TMNnP@yKAGOR0ic`1pi*dm--g;SdWZK
z2Gc8pk#O=^l#W302<*-hUEZ9+O<8ru%?1@?fVP4Z7CNwr0!U;=X+S%T?oR`Q(tr)=
ze*m-}kU|8yNC33+X`IMfC$htdkw|5&Q)yhxS{Ku}oV6~eaV6BNgj9O35}H?%N}niI
zLWfE^xl#$;D@o`lN>xx^1^bfkS3#pH!m*D2kyUiMvkFeEg53o{QEGzvO|ZKrBucNK
z{wo@{L;ZFdzk~YkXuOBB-NSXK!$j!-XM2Fgr#Pci9JQmTIiu4wKEoNEq45>Y=n9Ra
zIiqMAU*(Lh(pbUiDL9IM3eH?XhkXjpK|%XS3eH_Yd&Oy-UK;&=8aF77qtc0K-0(E6
zyUHg@X`EXc*X@%aSA>(5a5{wljog?<j!GYX<YXV|{s3MlfF!FBzj&PkG`_^kF7ecO
zQM@dQ#>u=anZ_x+EQQ7eysUu6g}khg;OO$m+o$Z(Rv-;(fU*YIeaIk31IpS6PHgx&
zQQN;I>Z`BpUSCp`H&!L5TBYOugxVcO8{77d`x9z;G%jY;ifLTKsMXMT2T<DqsQ06Q
zS`>}rfLa`lGl5zrjUNHEM>Kv7;A^0c=aVwRuvd-W+XI3qJ?F*e^o^2_y!<0MIYh}{
zkoyaib9h9MAE9xapc6;unUVyZBpP26`d$-EuuR^}dNlir)0^6_ek&F0{&+R_)6Y{F
zC>;=N4+!0v%e0eM`K3FB-_Tjw#u~PfBzmTK@uG%1i2*3#^bpD)k^`Wl3?wh5@eNLX
zgX=yZ&@CrUev{yP4-0d1U$&-wb%%0JO-SXKcOD~=!p5><EK8-u3s|v$;GB0ii{FiE
z9Q<D$?)(uZzfAM$8}`+ULxIn7op+m#>!V$FgKph=RQPzKkA=bA?MRQ(nSSYv1#to@
zfN}-UqTLkbYM@2CDay4#3%e<z^a3cq09rWpCrXV#xe;h#r%;s63)bfaXPhB<WYzdI
ze#ZPzG@l*3(+<A7X85GmZ|3=;K@7^dG%=1%%oGv^HZzXRH1>mze&qDV8j9S{&I_H`
z<0@&5oVN6+|K6`o95L^ZU)*`I6wg?ccCqTa$aNSThoYr*-ZI66m1Y6ndD90i`>%3}
zFDRm@zVr4e9-$*8r=7*<k~~r(>ySwDI@QDT$FyJ8VZM3+f4(hwelMM&(-138ak5h!
zmASdX$*vHLLt4bl&dau#6%Fq1ypAple~Ydj__1z&bb`U3b@Qij0P7OKQf22r)-{kF
zk0r3<4uxwljU$8^5yBiS`}npg+KkRiym$sEX$8G+lXwW#>p7PQk~e6HyrHm)F&y~S
zZ$`_@YX92I>I+|8A@_~!MdPl0UUvWU@V_-f&19|a>)`*YiEp>w-KO<1z32aY25)^m
zt`E$KYy0X74DBDUUe<oO>HuO*Bh+t%R8{{i)PGCxJKMFl-J|ci{C_UX`}G$l|MRG%
z9UT@Ig}xUBE38THA8ofQY>B;vI+{7-W)doW{CHkrRUX#*Z}H4Jf$ctCP*luX7tuJF
zvkm5`5<)RAD<)Sg(7pPV3p%{c_3IX`KZaT$^@-c6|LJD<K5>YNshOqKAae^F>%mH<
t*fmvBG9i!V<k5mW4kC}%<Z%djv>}g!$s@YmkAJ$8E=o%1f8yew{|8$&r3e53

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_dpp.hpp.E915886DCA1D5AA3.idx b/.cache/clangd/index/gridwise_gemm_dpp.hpp.E915886DCA1D5AA3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b70aab05684de1e53e164555eeb40f83c875670d
GIT binary patch
literal 14384
zcmcgzcUToi*JoyS3BmP(2p1`GDS{NKBGm<vqSz25w%9wi7}02q6%-Lf0tzA^NR0|e
z5fK%^5)1Z101H;EV1ZbO82#pQ8E5j(m+yOCpO?q%?C+klvu9?`oO9-!O_0BzpSg>Y
zlJ^Y%<qKvmnaU|CDRsmDg)Uhf(t?4zmXgx))T5)n4|wRN{x0CzNss0sPqRbLTaP#D
z$N#u@VV9f6m$Dw2dF%DtWO8fj)G;65Pd{_WWjN#)bj$RgRoSxqbl<x@R()<g^HW&;
z#@>BaEF66Fg+p6c+d)Td*C=mUKElwf+ndubO}jOYSUoE_Bcr7vd2a9?vwo+F`^G#`
z4}6#ZyYshp#!z?isgRnK2^*LAo{%n`wPIR%yH{F>r9(@I=}%o-b`&G;l?(eU)*8qa
zw`ncf{Yy#v>Es6UTZ^v7#QznZG%6y*Xqmpu0IBIe(aoH*W3BeDc5OSKt=KSFPwl+-
zj=S)@#LvnvXIkJG{q0fXCjJ=o{c{KZ(zepFkrp0SrylND6>uV0mcDAzcbhC^Z}R?l
zq`Nrh^V#Y9_NvY+&fEJ(;?3thP3+ew{d92ZfRFx3i$U`@zt1bazr6NLX#f2Oo-MMy
zZol<fzuR`Qq%yswH>`K`&g<dgeC#_1z4~8Xb2GPlWykeyIGZzQ;E5TxZ~8n8aDTmh
zLi4@zk(X1oa;E-}j~w+sa*}BSE&dEoQt9D&_Sl--!ylEO9Qy9@qm7pf(>RmRhN@XR
zUL}0;E#8rFbbE-&-zUAcPaOJT<9BbJi{8Fb8`N;N?2THi`|E~hvb3i@hhGi({Mzy0
z-kdcVOU-=yjZSb6Y0R2iIeMkoJ8QM}VV%bF{5i7$htlUs+oH~E>OC4|82xbai6^}x
zXS?YS*}6TtaIX4t$H$p<b6>hlFB_rd`mSt!!Q{$}S213pHS>1Z=N9SOMl1=c_HYQk
z@b>L`yIu>u9YO|=|0eR#X}u4%8AfkqI@*Jdp5C?FCpvddt>&ZY84E{u*UZ`#U*2_-
z*XPqia$louRq1inPCKPb)9j};%}5$!=f0q3!~B^ES$Eo>RvfF_bi2>NqZTVYv$ia$
z^t<IZwK?w6!QF?u{59U~eeM%Dx%aI7?ss=J&Rtfx^~c`(bRw?Wg=D+^^3(ZyVEw^y
zswIg9p?ba4mF;h%hkM&&eyF}z@FwNV4a1kwh3%p5oE_8N{;|eOx3$5K=MDuu(;eck
zYW|Dfu$N)#k9&D_*H{s<`<iX>qmA=cn}41w4URvZX7$I%#5Rw%geie4PWKj=n)0g+
z)2{YYX|UaXOqBN88@$<WQFX#tzSq^4Ha8=?Y<G;B7&B?13Yr`%{dt<z6lb-U`;}%q
z4%!(wd~J1^EUqAC%&pZG@t4g{ua4gpA{)NIvFEO5>ca=kS+RN6oGnp<pUmho{lG9$
z(?51@(~zc@a~kq*HB4~7arIJ8(D+-+r#}2k$yWJ=dBuQM+q);NyEOiI*Wqw{c$aFk
zfkTa2>@KA*{FJ)u$naxJ4>ZNUP94;^pl9#)8zwrhyrfIp3{v_lhqM}6)fD%c=WSBt
zl-jobo7m9<*MO6~%<lJ9D}Qv=JR@~*_X=L>J%=yW&h;33r}o^F1L?Dezsu3e$xvBz
z=fb!nqont{u3wz{d4JsORqBQA#vgX<lW0u%nJ=mJpIZ=Y`fT#j%H;);v1yO*uZqf#
zZ%MxDc-(c%#?~Q;vVCt_T&BG<a?UJVJ6hVea3b(LKF?>y_n!BCijR54_IR_QeEEas
z)S+jM#6r_~Pp_X2n-(zG*2Vanf`SW=+-Bzm?0GDW^<N!+Y{KK)?M<nH?+e?_*Ub23
z&Vt;k(hb1|nH8M*NT2EcBeQo;im@KGW$s(+s<cH<L;JYT@0rDQ&kr)2Jo;4hZ<7u7
zeEdtMB)is4DfH(B)8dt}37iqu?ltAbb1kT9uO!aZoom{0QBuJq`QP>G2R$CTcN3Jn
z?36@(6@BfFi_*6oR2pOpM8Jp7vbZ@>@|OlQ!Oy27E*%+lHlz>ZunU(RgS|Ft!EZ3j
zY~9&CJ1&IWE6})l-s=|;63F*e5{rR)G3ege5SHFm-_6p|*c7G#O>qwgklq30(Af|c
zbKCH@i53TSwO~)oGDCNUe#eE7Ep@BjJP2EHUJG`?<T<)?dUjk0+2;EZsW;y&n56~v
zF?qV~bp4JCA-kP^pf&ur$O>g>fyo;{mkmIpvmq?v$@=#*wQ4R(pr3=1cn)+w2lP4{
z!s0ojR(Sp(Vx9)n#JyYrHdny+oeg2}8W1r4boz}tE$EF|=IhSy-Ekpg@91#5$9B1Q
zG@uld+kkx=7}nVk7XN{7tkq7hEYgAom}RE!O#O}vArCiy{%iaM=h@0o1(VA_mohxr
zislPT*!Ur-PEmWb;6Th0j&#FO@6Lv>jC3utyj3<|B84h=u;rjzIne2BzOW=E`ZnL4
zZK4HraW7MKr|NfH2zlaNu{f+WWR(_F#^ey)5YvteAx~-a>~*?OO-~v2z-l+?n@N@(
z7o~3zz&eT5ZlQIReYn<au?y73&tC+p7lCePL;6^hx~gD-&bmSg^ua=&1-fT}ac4tV
zmQ1=_=cAT@7fps$FAu2afo`1*VOc(Cp7S4TcDGBQ8D`lF)c0Z*MMGGAX&nDsX2qnJ
z8c-d}@(`LlgkonySXM57yl%ez7B>yp4YMrMS!U63A>_ym>j^5UDNR~X50j_qPSfwW
z5b~zi!P6FflQLKWzr`|?0QC~QXo`lg#LltU@!R{&ej2bV?qMC!SqDrz8^RJ_=hrsd
z>A0B`THszbf^Hi@kIsg$Y#n<c?4j<HU<owDEPH_Z9z0k@Ls)j6Ullw$RM|xf>fjem
z(Ve2-aUtZi8&;<Zm!FH1Kr7rsK2XmG+MNwy$+$B;Xu|LFkOW#|mVH2dA7)WBge7lB
z`HiDLzfjSDz3@w#A=eBoIvc`LJow$IKZ7oB>jDMLvPgANkB$o=@2`J+Wk{a>?;22o
z$zebz49|z6AuNZNhR61N>N{EkYGIZ}VAF`VqoN@!CyImTJaCR^)quvBB@%Iw*tjVg
z!g9fL*q#^HOZ+9UHy%w9P%pxxQ8a|*;-MMBg$T7>8c+@QvRG%aWygh(FHclob2>m1
ztB)lnS0lY@WYF0Vmg`;A4^^-F9oxU|m?Z+5L_oXFhOpFM8dNiW)(@WQ&>YK=45Z1}
z{wNy4a%aW(VOlT8KGc93xR;eWD=j)Mgnajq;lklNK4?l{AKXI$P%pqEQ8a|*USw$G
zqLgosOP~>EDFo_;m_^YLmX^xA+r!5=UF`yG@Jkk|F4XC`5c20Uj<<Sd{^>1&zIans
z1Kn!8DHRQ2Sv+%A=sqjF%{x64uiD{}h)EE{CScYC?8)*8Vhb>9VR0(wn#%c*X%)me
zZeSfZoM7V7fNF;)D@wDfnt3>`jQBD9v%-my_bLB0XUXB?qgFscd<%uQ(3L2sAXaih
zCFe>yRzA|tu)puVqnI%fsU;#yqCJ9GfYb_*C3%JF4}P`hlg~v0On`v}a6DPwu6B{h
za}(@uX$s;L*%VW<)dX>>Y^v$VPJ=3l(`3_3iNz4a>9Xmj#C-|k4A~4*;tB<ErfjAu
zaR`Dq+kCbS@x_8T$9#?r@rr^NVjf~cY_cHEHJ@ukrXSBAkktWG;=Sb9sSfm~xE{#r
zS$YGIHT)AhHL(0mK-L6I>H0MRrzX&!;$|RgX6Y?J*78s6)I#~Y9@t%F&M!JD5yZvZ
zVnH6W$o~X!8Mll<E4h^nS_imwfWcva3uEy{pt2FDD|(Xml?+spSzHcO%2|8`a97yp
z8v)nI;x@pwu{Z*95s-QRL&!a3aTDa4C|2uHGtln(op~H5hC{1x*oVXcoOqqHz0TPY
z739PS-Xemxq<9l=v58`-)znL8MjZKoHE$OT+y%!G&C|Axe>>dqV+khca5~y#;63V!
zRX+Z9(1GW~df;3ST*#txVl!}V1}<dL<bwH!!y#0#*T~~F@}z>aB9B&zb=3M-56C@s
z3cs%r`81+Ygwf!5_TIU-w5{<biiEb2(2Ynah*8isieiJhaYtWOl~f}^oNqV3FA1Ck
zu@!2xLM^&UD>#h`PK)#}d)B+`+n?@v;umWvX{nNd8he%=YIV7y`qc+Y_M56|N6br%
z3KrUNzB3iN4h*dWBgpIuc(z^Z*)3BKJnDfLr8j_~4J>~o^ofLn=v=izX)DyBlkf=z
zenP=y5(Ke=lU8s#1d|9H&t(FTB!F%tR1m}*V3GsOh;vcxjLS(pOyY8EG!;QP86Po0
zNVOvhCqE0~2hRTkH<Y|s5dY?a{^q_2M(h!uEqRDeADizwzdxO=CNQ)Kj3BxvU|n==
zrUpTdJ(__RrMG~gEi8W&^ojcCZ0WT^13Fu&Twp2{N<Le96`TPP+S(!dz&Z^TL##pz
z^%q(Zhi5(GqkB)d>ICG(9N0StT9DP?#3I<ch+><78G{`1mPKP-D~J8dp$BEGg8i#l
zd<OPE!{SS@|0NdJK%*LHLY^mH0IGHzUg+%o#rgk5C*A&W+CIH0r+&bK-h`Gnp%r<a
z6K`{Yx49rPP=~>dQOY{woW7BdNG}KKQ^q2wSH$9Ss8bGgDZL8nRI&IB)H%cAOHk($
zi))}(4V04S`!ze=m|y*RB%WbyP8+A@6n`!Qw-4O^!AR|21$XTm5p?Hd9(LE_4y3&U
zIZ#hrh_nll1M$Re1}*32DNk+D;>1)SO$9xOqI2R-Al>;-Y_b!WQhFMYrm^%iV3Nkt
z(}6UdrKbavbe6sgNO!UHUBGe|=tKL@0MZPWo&hW~Sb8RqX0r55V42C%vw$><rDp-l
zES8=Pq}ePz8(3zu^c*0~Vd*)bcMh<i^5p_)E=$h^y>ls@7ziA@DNf5&85IW)!XV(o
zXU{Uwi=0aZ@i>qi2U_HWDc~^IIfClE`wx+`bDl;1$O~c(a;!l<L|Fx~7CF|kxE|Tm
zBM(Zy%-LS%zIY@-yvEsHqnMoEvCjMsB)<bGkwCTc6i+1dkWRa1xNXf_Y%Gm>8+np&
zQV>mgn=r`K$kdh4h~4H?JCFNxHrjz*JD5n^rXaqAb}!*XYKq<>ySHc}8N`6F?#fT!
zWm;e%TcCLhw4#hhIP)W%6=5W{QdzYle?Z^Y00JApBqESJb`ER-lc=fu4BS40K~$b~
z&}|)L@I}t;A~)#EYogA(y}TQ`vtu<3YfxW8kgtYeEsN`sZauQ5^vj&yWzL4;Yn<LS
ziirT&u_&?}WCT7zdf6-AsrabM1Bbqi$eZ>M$EnA0h7@n()VHy?kW(*YaS5kh!s1%a
zzLxW%&)0GGbu5nOHRJhSl%B$Crm#4V*UY1M(DhXJl<~fyz2xz-W*sn~b6yWL>sj0Y
zG#gml1T>p~0bP@3pxMmg7NFU};uz!@gFL8>i$Q)dXe7mP$T5zk$05HsmY#%!BqXK$
zNysJ%xlo*pgk+YUf`k+nXCudK<UwqzAZ8=KY&4SM-N<n_OW%$BcC++CBorbk<u5`u
zMaYHXVk8u^^b#bLu=p@?Jd8X@EFy@9k>6o7lHy9_Sjo~WkzXZCe~Q$eqA$UaAa3T>
zH}kr5p*Qn(n|U{iV|eu#mcE5o-@@VqUOnNT*e-#6K9N^XWa-;^_3bP!<qb-CbMiiN
ze#4=ZJex;0f*`hY{_WJO3*sj(=o5`v{QZ738kYo1u&KC=wU|8*U*o*4QB3R_jv5qs
zK>7wnVvjHkcaHx-YQ6baP7pt$;E!lBkq}3XoMQqtrTDnS1t)V1eLv^8pVEg1IBTEv
zO$x-K{f&nHjYiOpQn;Zh+!s8&GZTsJrXmeVMN2@Y39MhwN2dAIuM6S{q;>-76CEJ7
z9GjDWj)@$qHB>vYleGVO;9U=V>7H!{-p#<5?#UQ5J_b#odn678#-ULZXQRMuG>Yo*
zZZv*3nn3YkH2yG}K;*-z1Qb|_Mp1kZ1>ZxX>BKxk1D~<ZtrfYoqW+Zrp7VdtD#Iq;
zYZK2H$1L6_i}$1aS^VHEeh9?{yk`N=7{7hI=RS%@FDP~%CdnLw1F|$;J&kuD5034A
z#(v7a9;PbT-~;Vi(3j5GbD;ek^krvEI|lWoGZcrk<5--Hw6jrPIzzjW_HGs*M%sr_
zUphmTNV}5avF%CrkB@{q{>Mk;3H?*LXYx-M{%epqF}eB0AoaJISc`ww`q`Xhcm;8_
z)@pN-YMXe_b^ZjiUM6_r%D|v9Fp6v@l54>ZTAuu-idP5B>VQ3+#s*;4z~WEH^%JWI
zwcNm3n%oj_2tVpJH;Jszq>_2BN1Q)>|NnVPp7J1aw(?!WJ6P<AmJ?lxbfm@GNS=}>
zsk3RTt-rmwt+5;CmU7Y_BwU|0x@XB9^;>HA=R6pgN0SY+)wYNJ`o`28&)jXGeH*yb
z(Y^-SuYo&Jnvkf*<jSc^%Q0gsG-`!?Xt-0s8C7t7$S@ZaC05VAH$(@&YLVU|H<BAB
zMFMO;6cqw8071M3g<JFhBZyC++7l=vnS4RK!>QcidXWA}DFF*|0!U7<!J(qAK$evh
z6mUQk3nZ~Xjh>_vfGhzRQk)26i7ZYBM(Mzq1lX#bWe5@{|1{L+g4I~3k$9u+g7{wm
zIV%Zb2E_juoC(ED7H2^*i^X{muLr}Q55;^I7eJhrWa#^#xQ}8|55XF>1xU8ABTNHE
zY3v9TRT44+GPhWH6eSX}7-Z*Q2P7|yke>x{E0AslW@N^~4hLC9-9ES)ud%kR_CTU#
zBoT_&SY9`wPq%UYZLCEs<UI?iMH9p#-m{2eQewfwN(7QbHU*LkOmf*2h@$X9h81n<
zH8S?*j4fC-{`ZhPdI(zoBKEUOU!143TI@iI0vI_h7Aab0xy+u-h#;=CTxm~iiGb@l
zX6t}8DfI}ro?#XStVyXuz{MP$jjT#V0-Z<}Zvr}-SR4g(qF5XabfQ_j8R%?g@gAVE
z2dI%Ek$}r!Dtmz%(FnZFfl3}wBYQx=bsv>{phikUINJ?W3V<4^GYGioqp}aEk$RAT
zD+wxvK#dd`@Dc%)BA`YpECNnGsT2b>lC2VO2~4E~sF9@;#3MlW2rwef9_MX<ZUu{v
z0^Or5t^~T3EUp5&RV+RRbdRyP4Op}RJ90J=upWCnguX<v<qDwD!e-spBXK?QAyJHg
zf!}&GoW%j_(P)Yzk#Qulpe`~JSw|vA7P~|uPZs+|qG1%LBja>rL0xD%vQ9^iEOtpp
zo-FoFN5d#SfchRlZq&sdKpqE>FN^&Tuz~X{Bzc9tq*n3xk)(yiM>vfmG|MRWO_JL*
z6)T7vc#RD_6L?4R8j%!7KliDfGi&k3|9bjqbL*kwue@cCzIuW#=3#1p^_p=ru?d^#
zK2apLYD=_+Lt54kWjN&cXP73(Y?wxNKDLyK3NPtFp10LV00zWX#@_hZrK?iu3%RY7
zYFiUq8NcA@=b2Y0iumXC$ZS1wAgV{6!?yXK=^3)uv001(k{H(7?*v9WS!=H-6BDzT
zuz$&#7x!z@@T=A$gSE(;xbD=QRu5L5SnB*A2Wq4`)=jme08OXj9_n`wITP0{h)>bL
zr)<RZ8o9k@L+6{E>rKv`^1tK!-?5G(oc9c;jzSQ_dGB!Em-0vOUJ-1loyU9TvHbbG
zXFiKdd5cp1OQuE;OL^;3-j?=P%G;OnOeUq2?^nt*<0<7`OL<1#QXY5lC9g%SEncx-
zf#g@9Mb;Jv4#0R9P0^`#mgUKcWy~MS4GD~R05~xh*yI8`5{hJ$ZWNw|T;Gj7TqKl6
zLK_--L_uj3#l(=~SH}ZMJp1ZQV4TUmT2byNuikyB*0aEO+u@ElZ&)iFOO!>PdLLH7
zjU|(_XUsJZ)r0fo)uAM$orJ8Z?JGjsMaY`iGSUIwDT=QLsCryO!Zl=1dd$17d?Vt@
zx1m^1FM)2CfDMTm^OqmtM=r?p>xNIWNIMzH=&O>ESu(PrI0b2^u=EsUk%H_gy%=d1
zBN-75yD?-|jBF?_LE0rOy#$$+u=G#J?h|q$661_0XBWXSIF++YWpOrVm(AiBUN43>
zq0evO^|r7$k=ILP@pfKsJBzdUzFAb^_T5fBxy$j`JiG?cymvG|op!v9_ufWv;r_*E
z+fvO^@zgHSTw*{%91@)4e1`n14Z1K(WJ|0GlOWzh!aejQ2~#|w-XU|dng;GG6?97l
zT0}WX@Pijx@pT97FA-=YQj=J^^lZxVVaK*&vDbkC>p(Curt&aqKo|%nJsoITaX>q&
zb?;X%1(wgccfm*X+<hFvUjbdOfG<^6oXP}U8(G{2y0)>n33hFQ4F5*NZA21c*72bn
zacL|*fVcxJu0~un#pHN}LuSS2E_BX50=tiFI{Fo~d&Q=s-y^&CG#y=W_K$G|?tjW1
z<<~6_S2_mG;C(Zj@G+_kI+W3>uOOa+4yRZgjhv#{<Yp#v%4Cy`<;ba=CjY92Y%+dv
zanT9924&E%3=XBX>=g7n#o}l*BpL<KT^x&oV$m3iGtrPt6hL=zE(*#;V<;|1L&{MA
z4W24cPz4%8aMgmFd5Z?rXJduVgn^m#8<1nCxA13w`|A=mqW{<O0+9)OZ^gGsWIfMZ
zyL$fo^Eb!-V_$pbS@PPQPo1u+%IjhVJAonDew>&F4ANMf4h+&+ybH*70VB$v0c060
z&IGbd7H0uj7K^iiEStqSz%Yk~9dch~m`m}w#`rtlZT=$O<sXpn0Woehi4&4=1*r35
zt8;A?HEo{FeK6zyerzMB54;}RfMgr%)UttbHtW<BUwsj$MobxA75OJ$WKAxa{1O@F
zF-o2l$puciz>Q4J<^E<Xg97vuzxrk(L&kzAJ}x7oTp2UBN83+3Wmps$C;MVfuGTIc
zV`X@H4gUEe=y?%X5tm5z3hqep2^#6B?os6E-0Kzq&(SB~_6dxnN2ImTZ7pQ*1<vgP
zH<E}{zoD*&@>1Q}uO4ZWNyVZnzJVjFQvW73Xiiep#{YOr+%exg>O!zZ$iFHOIlf~L
zCI94(%$Z!Bq}ynEgPe--9L51j9IIy;z&L}|GsQP~WJKh+j33+zBwN`Bvw(3H`=H`W
zy)O?wTst7HTM^EL;rviw7W|&9FFu&Vz<m_oce(8SKEDU}>YDlZ;+~^9D~AVQ2{K?{
z1~q~>bpQkNDQ;59kA-eUzv3~*BKKJ2M@$~xM94iC`BAf5f!r&QA2G8{gWCd<LUz2v
z57r>R8Z?Y(p1eTmSIgphG@u@R$&kNjpU^fuCBnJGQD~L1slroG>lBNlkxn!+q!Ps<
zgIHuraVFBqM21wtTx5`oOero$I_1caNQ^^XWKe-j2_~PH;BZO)F(p->56JEV{h$g5
z#GG9ct+BMG9;&h}GV1l!*R1vV{isK$?*E2ea}D~MCEq{QpEZYX3UK_dANSv!=xX+(
z?yFz_btk`Db$Ye5{kYnHUBY(}{^h;SmAw7en#B9|!hG%L3u5q4HuHL$*;8c#ua`h^
zdwXt`?d`-jnDK99_c!vV`yz$2OX0p?^1Uu*JP9Nx+03`%t6h@$UOaRDRAU^G66>Nw
z2&4q^<%iJ4(?e6_v1>mPz$z&@(p&t0>2>}=@!WyBN1=zJ{uBl{_H%OW@8Rs?=IO39
efG{X2xzbxVdh1MYJ?X7Gz4f5C{pqa}z5O3r>5KOO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp.8ED0C5830B90D078.idx b/.cache/clangd/index/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp.8ED0C5830B90D078.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd3c3c47224fe48fbd9c633cd4c0f8dded9fb31f
GIT binary patch
literal 17608
zcmd6udt6T0_xL+|?{iLaQl9QNUDW8Jhm;a^D%Z-rj4-a_HW)Kx82Svu2q|S;q9J1#
z(vX-ajaw<XOgAbDQA#KkD!0<~+fTLId;a-+Uh~iIJH2MkdEVAK=R9X$)?RDv^#u5h
z9GN~`O>Njbztv0Ue=`TEsj2nE|F8Hac+mo0O>L-wn%e5vOF>J&-tVjZbjA0)dGq6t
zq#w36PCoIi!q8#iAFF;HTytR3fm_DcgiUUB%0&go+gb(vOy%aQ)puGy&-BR8sjbhE
z=Pr3U^Lyv(ADuRaUj5~>9Lw0!iWl!T{MGzm(Z&+{0gHe1GhXPv#jJME`=4|R8+D6U
z++3;mWdB&CvF>(x)V`)9pE!+WHuLw7)E<_R7Anc}yt`=?9<ja@?OHf`z?OoV*hOCX
zrujSht9EPqP7c^uXubVS^~&9eennjOvclyt?_ZS%ul?zIvG(;w6AhE{nWrn}Ii$>s
zTD;Tj=%a`BAE$8(2cAylru_g0eVuVUeS~Jtr{zgP!XF>D4qp>pas1}Vk}~_>+Ftj*
zZ8OYf`BGOC*ZbNZ_Ll~&wK?EzvU6~@{oat#ei^RbUqg+n2Zy^?dY27Ls^V6h{BUbV
z`7yiL{2sw6w$skN+R*I1scg-`vha|SwDg$aUq!9h-qOwE=YT%j_{OAze>C_Gx2>O^
zs{YrfT4i%CW__LVan;3fg}&#m_Bbk;Z{Q9dv3rvly6m)TYUB1lzPNJxbaeHl{5Y-H
zotxI%_?<r3?}x|{Q!0*4x&6G!;@KsuZ6nvNGaF$z@Z9X}C-moh^C-X8KI8e}EoBo!
zL$dcCR)3ovzU<DQXB{#E(sP$Ss6O(|_E#qcXk0R@-?6uKUSLo6`Fl?mDE)7pjJ`Hw
z@~rS5V+X!7s{Hh~qRg)jOwaz_W!=HJdw+IM&+q+9NbB3y@pUt<e`EaIY@^?Rl!ZHy
zU#0J=FUDGR3qG$HGvvzyzmA$&?YuhwqS14&0gHs@UCwbEUQ`#v)GQwD^LA|G+QP5G
zeQ&?(d*{i7VGhD?**1H={CKF_@|gX7M_=l7+`_NrW%S`bgHE2feZBT)4z|v(DT|u<
z;&SP+eiJ@smWK3rV;8-5+Vl0p){eW^67z81OtYFRiyvW))zIU&@|LsX>Lwqp`2M$)
z-LE%=c;5QY?vP)~LXC#c0SDI)m}K#xb-d=O$4Vsc-<OrR=y}<ZIX}F(`(?K*@I}fG
z<9we~_3ZV2R;Z(IOiq^F+yzS&>%ZNiq0?;nebKa}Q||sTFRDLxT6eVa?v5kdS1x&P
z^kCgzFTKx=d2PMn?W@Ls^(l5gTi>$t+^atLX?^vRC&*|~9zWl{_KSVbzkhyp?3uCz
zx8|44=T|3{{jxTERsNQWwLfgwJmU0#oc%%ibJs{&(`?R+Q<xtBwcQSgznnST)2boN
zE_rBqkkM+(pRX1NwB5d%rT6XnH7+@!vx~0x9XM)Yk9iN4rcBx$_g7Y{ZHdkwORSGA
z{Or`SnK$~KSU=%amU`yYWTAP^T(96;;rI1+oh~W0+P*qw#O!-fE#;#UuMZE|G|45_
z)nu`A;Qj-P&+Kp=Qn=vR{o0}XXDsd>EKa_%$KXZYrURbcYK`MhmpDF6`uu*Kw!Ouc
zWwVDZ8&a25w9(%(IdfIg{fOgLQ!Vwxb%QTyPTb@hXHjSE*mkLKP0F>0U)2YICptg<
zHQ@)(e%rQd&U-jCBVd}x{oG#zJr7S`a?a;QXiCKFM-Nv|UuRmd-1lL5&nV4I^#a3f
z>w6wN{p!(;f_~mJcK-N$$EKWD6Cwge-tC$Ct+!6tw*IdVmx`GeoLB#0SbR!<`@1_~
z5#N1flYLa<n9C{{_R8R)-;eXJ-@HHIXxe}~6J7V#uTwL5(AVwmh_zV@CkA~RvB~lH
zPYY__FIsc&^ZGq;Yo1Q_(BIT+(wi*PcYB+fV!FS6+^f;3D&?mE=Iea!#w;_IQ;lA?
zR!%m$nilu)+-G~v$C)}e>hr@UJ*fKa#qykrYt}<T{#<G^>F}XS|Krn>-s+uf&U$g<
zxcPU=O#jthGsYBstXTEwtSaNdr%J|88ya;qYgP<zH%@1gGU-Kio7%2~4OT1GEdAV7
zP*YRCbbDRGqDe6dVB?}DKjhFu&aU&MrjD&L{?po0Vt4NM`jrZxz$)vJay=T{c_J!0
zm*Tr`4S4)c0J_d<vcKr>(0)+s1~^SMHBRRAFI<b9Xz=YtJ)p$*7-Kf3Py2yfZ|G_B
z?DG?=uPJ~779Zh^k8lp1C!%7}Qrz&wxT#tJx?z<KP`Ls6cAkie)qck>&llBrC_rD8
z3P2lxd*_L$45*lJQvI;CuO1Mw$_(QfX6=VAaof6zKEqp0x&t1I!=OeOH0V4Lm4SX`
zZWqSdE>!>xtda<niC{?QiKw`FxZSNA`Q3g6u*53K5G6yC&J$4?JnujC!O@>@RRD9W
z5(iNnH0?Zfsbs&uzTo1ty9(fhRWd}A46#?|iKq;5H-Gy5CGbsm&>gER)>v%Rejs9x
zwjDPN{2PlDz#5BlIh4!UcAkjJh$SnNZtfZvuK*TUC6Y6a#0}qZ>QbpU-njQv_(KJ7
z#44o_mEyPQI1!a;+>Ffq%TpHW0TZkeWE^DLe&`afhgYUOd%e6n7=XodH0Bt!ABgy~
zmyeu(=8C@U4tikm1dR!%?FS;B>0#h^rmv2X9)MUp(|D#$`+<lTP6O`c6GxTl0YfaF
zWIV~V{XoP^1{lpge|=+|0%&9LO6AJW+7CqhAJ2k4X<vHYRsbz5UaDOBS^I&Azk3}q
zbC`Y5j|$Kmi(@$B7~DJ^C!!J-SNB>yY04ozV2V|y7*8>6KM?UYw@)GxV$UAb1IAc9
z*?6*P`+<m~PON&^T(#y$Jz$2#Q;nyZwjYT2a6`iC>Al^~DS#~&zvIw5JUcs1L?u4l
z^<?O_)Ma`=AFBiy2bi`Wh&ZLvXk0<m(3#zVgvH@dBOH(DjuTNiXE?FM^W%6k1+d2|
zjSw~BvDk4UDwik4es<;P$1fD1CstX9)_J!dh&XN9i_4O`or4}Q!s1VjKQ(PX5OMl$
z{ats@Xx-NX0v69Qo@LX1AmXbP_tRUJohea(URWFjQ4}6v9Vepl=iP2cZrxY<bq9J_
zWv<3t<Msm)XBVhXj9PUfRsrm=cn?5(z~IgkQOQkQVdy<&a9=%OfK>vG15MixM4VSR
zWu9$;CG41NVaOy5^^!YIM5So?aErASyDoMI>i7;@p~hC+k{u_a^6+Eb!?hQl&(H(?
zu*z7ov3=SPL|ne{6k4U#-&GIv$Kr8j<NCB8i1_*KnEt_88n_)k!Qy8edWNTK$BC#^
zi~6ssr4L^!Kp(7f7@)(ztn)-v>Q{aEWV)c1qyRct<rL?1iW}2;A}X(6w3$?%K94)x
z3ajKol#4sO<3v;%-KW<Ft(jq{2M|`7Zam$r{XoRQ^B1g`Fi1^JxG>PdV?^R5B*{gb
zRuN}UOpYYib6WMBJ$WZd-U^0o1v80>#&UjWJwJoAJ6VQ+M(5%~RuoBoBldkGP8{56
z4K@3R?OPn}^q`L<2bc%gkkK#6f#!iWq?09ilKCVXVi_d)Q}a)4$QYI6$>x)7$h4H?
zAoCy_V!tGLiun{9GNUDVs`*qKGD9VKy6to&X%|VJVLL-fRuf5{X**L%43#9$vYn+Q
zgI<!CTQ9dKc1Dte4T8-mt+8HXPfU~~hZuyIQQFLzZ00P<)RW{e&LoWGXikde6vTo{
z@-Lk93#T9^OOg{fDS=avrB#w2bJAl@L0VdpTR5qOQxFp)$q~RO0*s<Pv<>)dV>uQ`
zu|PrQnk1iw(rKt5E14u0L#Y@l$l#LXRw%VX1+gxYd;m!Ykb?BMBp*T25tb8>l)!R6
zGRa4lRKFgX)U({k8#eO2sJxjsY-TxA(9IN#sXSZI&1Shs&@G~@nH-;H%P+g6E6GdI
zQmMoBQB#xTm1reHYtdSUPI2fI$8ZjZa#$|qP$|n#IrNm}w;X!Q@_K;Q1Ll6w07bKW
z2%tkOCjykn@&<@DK<4^Lh$2}&3ei!P&q8#T<s67|SbhM}1C|>gYG8Ri(pryjyx#Gr
zWmq({f8nDg5Xo8Gz%0&{*din+@-B(|K++eOMP;Jck1+8tIgL)glFWW`#sZa7dHQFE
zwE5?Eq!4dXQ={FBBwCUj&FMsQW^_tC=5!u&W^~Bk0Qxt81##gd>~XopK_5DVk3qL%
zEXPB)c$QB?ozu{a>K{Nl2ap+g4N1;NR{5we<to&>3OQ2F72R{iVZ@e6a=EBqE?N*K
zQxzM*4i}idSUBt*4qeD%Zdm-g`p0AE252BT*fUrr<^;(ho*^<ZAgWFX%;F}|PPoo}
za-Hi#IU5Ym1|vvoAh{3>FJ$=xaQ*;Xs5}kHX~>Jb4U#id|1q4*yC(B)L~f+zann04
zJ&Tj%Gidl36i5WUil<(BnU-`DN^%P4n8NiZ-$;_PkYg6=PZTA27jL?YHz#bK7iIg(
zvn~Nha-N{f6J#Qg<a>hAJ;9DJi52j$=?D}^Lt1`z>FD?$7LUb`E<zSXXc)Odk{gjl
zBg@5tQ?W3K$gRB2{?Rb#md3xj#fj;L__f=EAM)Cgybl=c17oP|ZWX0g(T5B?#{m(^
z8x-mmxU1%v&T*L5>90G^`#8i5tj*9wayrmT2R-RfNCyMcfg8)y)4@#2Ilv+ZSW(Ue
zmbsue<sx8F#LDjg%R8+69<aE_%I^b<`z%)ilS*Jl?^g+&DuFZ0zLmhA@^;Z?yV!^H
zrqk1(6HR7keXWk<0?wp>GbL^dl507WTF#WXElA!BdTj>gq-Bx31@zj&ay95(4Xo*d
ze&@&h&W|MzLUIy6CW+-_{_|vhE|CxLeG)DhO?AZ<vl3`k0&RMuN?=n7K4ICT5_l3O
zo=10$_V|@{^cwdy9>fA*hr%mlC_}?TMo~)PTvIrAI`?Whvs%uAytyR5<Yv6&W>d~W
zu35;Pd~@tRAjg)!&sCdUye*X{@#B*CiG&AqM%H+2c7)aRD~BQTFxF^hh~^n$Uotky
z8{@g$5p0tSNtBK6*b!nAfh6C726v#6dLq@LRki3&?l@@Z-l1ojEw%p7d*9u8bHJm_
z6K}9RyUV%W<tCD0D9Nul*H_#`()eV-!3#r2s(`*~k!W2cdXQHmXRO$fFrbQUqI#Ry
zohXpZK{t)|i~&`<##vwEoav}2Mv7wA>3WM4Z&}{W_ukD9AiXTfF}xJRn^69VmwsaT
z1aEkP@0y4bb_@&?DG&KNY~j)Bu;n7QSjb@?8bWV$2ss>LIUYI0vz&;OiHKPdj`RJF
z^IeORB%kE_on*O&A5z1QCinN~;dyn|Nb5D&0)}!@DA$j;-K0x!i*_VXXj7GQmgU?C
z;#_$Dpd&ngd)N+lW)`QH#Tn8z+{G()@p|M@KH=Ge>&sRgXn#k8eaM~&B>50BIK*;1
zGKgn65g8}4?!<B4<T$US*PrA~PO@CX_o?9r6Z!A}lOx}~RUfXYT9BR3u!~(DQs%LZ
zSB?5qqpp=)lIxJNj+KY;eZuHUFUi02PQUX$<Uta)dQM3!?+|o%2-Z}-U(nsp@)<$*
z49od~NxtAf^$P^^0-?)Yl;kGCv`OglvxvWk`=}#zMf)fW_6~zC#Q2QpOl2XB&IA@!
zY=H(XP)UQ8m!j26YBD6bL9}V0^K4XM?2!G_Myznc(VethpEzJRx|4z#S|hE|rv^7f
z3Snpyr@e_YCc_NdEY9H$H;@cB6_35c1yL^M9Ew?a3FlD4@=MO)B{z^*43+sF`;rTy
z{F-xk&B_}%hX$4-fD{4r$wH@!29*(D5an$^xsByWAVsqJQ9z1fc`tO?3uThq!G;04
z#IPI-U1C{209_8Sd>Oi2hB7holAH!z(pb)fE}1M}g)UcFega*dK$-NoBv(O~DweCE
zOEt?i(4~gub4cqPGNMDU8ktt3Pbi1+reQq8uXycOJQHm;@D2?$;*sPJyu$~UTY2qP
zR{t+v`!ALw1l<UMiEbkVmk8G5-6rU6W95;8ZY0amf^PKR*(I92evhELhn4RYboa8H
zFKFk}=vd`p+usvh==(PbJ)49sL+>|ox#iRtkc1uU<GjytemGS;$@`pSxrQHG!%rY`
z|DV#f=WVRj#1|GLsTlPn0!e<0q_-^J7qssSU1oHw+s?1`9k>06?d`8Z(67Sh^ui0m
zpbJ8mYd?0po576{F%$6<9-t8qkS|qiMkAV8J|&DhC3Lwg<45ZrnXLz6{%5!`dFk(N
zV|pA8!i&d2LHnT4wH!{-x$ZPO(a=U+l9y>MbD(Bowbp6}Y8duHgT2t1PL)__5X<sq
zXmA-i)2WaN4Ki7N0u7!(XFAlYp+Pm}DJ?NhPcLkA`NvT?$%<jKs!HioTUp5mR`Q>c
zW+c%9Hma)BK2^Jeq&ui<gkYnMq$ZY21nm-`E4U-O0NBq{Z2^$mO!wHYbC`er)&IG(
zMzVX@-Bj&Ckne{V0E6XLv}a3^UMaF5Z8I}z$*XbM*B;}s_s>Gz*%e_6g))mrcu4u#
z`ni#(k(8#O(Z0z+{#lY!Q9vr{idH2#nV*r&&mvy~CmQ&`a@u8*d`|E_Ck&<X^Mdzz
z%8Pnu=vrv@IHvMtfFcX@B2Ku9ZL?@jU<s%iV_W?DbNn?IparHR5k!VIZY$M}3;mjB
zxc+Cj8Pqe%M*Xu<R}y&X={+|WmirswyBz^qM}U%~yO#}mxjeA&8-wnWyn}Pv!Fk{K
zF|sJQ`+r_baxCW*%XyE~mWQm^5cWcn<2k2z&RhGl6(_$uV-zIGH-Y0#;4w>U&*!uH
zj?t6k+raTQ%L%Yw0+cUIi*`NlWphB13t_)PDF0{kFx{ao8=hga4f}n7a;Ej;xevjl
zo;ZmrIy@E!*Qe$MFWtLyfh0c@9iFjVCpy$oUNw4nj>8n!3HXgBcup8k#`x;Y$4^4g
z?OaKffwGN*j>N&2<YUnG7|Zd{HlF2NXqyWiX=sxNZSz>phqn1FH$vM+=t%vVCTQEl
z@*8OThUHw*K3D8V!UUXX5$*F>&KK?TDU)%JJ6g4OMmsu*>z~AS^bOSi2JPtY>dMCy
z75M7oX4%b;+Rb+jT#~%TCZ}VUjoxu9?7bDbkWHA4{g$jVf1cuq?@%jf*9u)TF7(&$
zL&nBuUB^Q=8X84ID{5Bv!Cw1dAHrJ}YI%CK`XfB1?jgN<$dFEk=g9Oq>M~4{{2uAO
zXXRUX(=9YBg<TUtJ5A7`*IyU3ud`e#7?ld9RNg4`Y!q}U|0?SKDjHBeBI+Jt`HX0I
zM(i@AWP1X)k!rV*w$Kwk@Clvc5jH*(w%?zJ-Ln4?k3{&x(7NqygIpzfjNTYil3SJJ
zv3g@o$&QvJkJB4xN;a@0c|E6C&lyqgbpxl^!15{1?i8EY+W`GGzzHNTC&?ROkBv}|
zauhU;g1t%KNO=3tJ_<UqJSYl!QQiT4c0ga2{dd5LEC=p@(<sM5WgHyHPBY~(mgAu^
zp5<g{nhbl>B{vz`CqpNe2PH!<%Bj#N75cL5p9&|k9GD8HQO<?RTsV-O#>zaF^Pw`I
z<x*%`3VYKfz7*P*LMN67l|nDdWzeS#`m*d_1}Cx{SO%w2ZiLE4IFL0H$|jcIK;;{j
zFQXopQI{_!$yZR1D=a@imJbk<3~EM}%`ETa6+3yxyFSbJI7>rhNiN{^3wS26UdZbg
zvb<SPY!;a7w+M<YEI$%#9?<|%k|Ra^NRheUpQ6p5G{+>#*`iH0%Q>P=4$HY>|6I|H
zHH-c8Sk4#w=d*lEw7EsEkEj`b_lpI=AOCThk8D$7`>HA(pmV)~53HbbU6TJ0y#Ju<
zcVuPkc*pOj&d1i#&)shbxhyI|G2rK4zfM6&juSlMXf%N2V}i#qmVXmGeq;HB;BkU7
z*%QSkNL8{x?vH~kKCp^@t({9Rwavdh!xMWyJD~jz=t}2fDzs09u5><>LHjc3O2*@!
z1sTs|BP)n)*Dg_em&k-c2Sn`yl=p2PKQd@o);Rp4d-z^^_#k3LV-Gk~e3!A(?H@b6
zWIGhUT1TA)eYtmH-*<E)I^Hs`kste^8-BSeBvm0R@@Skv;-yfY;d(*0Ug%2mkS*(8
z8tsJ|^#1vrQ$FWRBmQd6shV@95&TBrv=KPdi2OXTI}aSFYjz&Eod@oeF95p>to#CS
zyTHmX0=tW>{1ULc#Bw_5n+}xpemTG=2MnWJ1bm8Ez6X5nv3w2%oMU~`G&DF3dC+D`
z=B<<22s)V`kjyg?W->oMnI@1W`4J!Zhz>4Ee$Csz=DU(Wc$wzy-?JPlc!df~QYus!
z7AlOO&kq$wg$jNwj}8^OQumU)S@7E|bY%#!O%l8c*!wjJ9!+#kNb*_H@T}OC0r)Mb
zK-M_7#8FVCa$N6lV`#^|<Xm5JV`#@l0D}nNNIP~LFxbX&1Fze_ca0xaDna)H<r9U%
zz0J3}t-zCIUe9?-QYs_K=XvQo-?b-}u=)Z&Vd>$K{cxboxfXLW{XWIqh+=Lm<r2=d
zgq4?Yz9rlQDu2znzUE~5{a<q<UUOq9H*l^Eth|BqZQv$Qc_c81q<bjXID=0jfs(#|
z6flTl<x#*c3OG^u10+2_jQ>=Mq*9j4kyOs|QzSiQxfw~#EVm%3h2=IRwXwX3mp1Xt
z^ICb6R=zhKIIX;WE1Lsv<=tD^>}M<Q+e$MYlHA7Yw(*QV*v4D6(f(A)J==I@ptkX&
z+xV{SQ%T+?m~Ru99D1Z+9?9|^!F&(Pdj<2oEFTe^j|fa^Jwb3zVEK`t{YYT^;c`K{
zoaHJ(yNczfg7#CEp9$K}SgsRv>sbHym7x8Km46VdK2ZB8$<2aQGs`W4RSV0TMXSv?
zEa@!SlH@I-)fSe+M5{2$iKVfbzqq6?!Sf?r&<+=v?d9Eq_HN3_$8K`fUrY(bikAi3
z%j{IYCiJ-`c#|h4zX(Y-2yZz4uigwMwF206=qUH23r#pO4`(;ru8QVYsk1_yK)eEW
z6amr;OGN7ungmZ>P@Xef<Il(XxMD_gX;&2pu9(qW$<SKOwG5r&G*591=Wv=iESGYc
zr7S<?G@r8kmeYL8@_L}T9x(Td2Aa_<9|D?(SWX0*i7an`nj0W<eI(S3Wceu6Jj(J}
zsCky<9H^PY@&l;(faM0L*}(F8gw~@TBvFEkx)2IuIYUGlVwe1ahh@(vVQ(XG0`Ue=
zzX23<^}PkuZ?Rkg)Js_2Axb+$eKKTmHJGTIKpiJl>4|!|sGuJ0Q&IgX%eA6<E#-?}
z>jytG_Ueb{M!KM#&d%l=g7yu{7jN8|dU8l()xUafTy_bV$B#Jt0{d>!a8NY#rjF@F
z=z5XeMyP<U6?9vG1iW}ssfu9f%t+^?bk3J1wx09a&-pQ=_0#;fTGiz)%fMxl7oq<}
zIG#GbS75*u7(}@O`d7g5#1Y2%B^dAw1`#GjjQIJgDqH&e$9&*p);CHOyi;jBKxz^3
zWvc31df6I-HC7~9r`pliYoWQz^n<5_FMbe@wkK)ycT<Np!~2WOzb>pijmN@&FT)|N
ziXYoi&PyMh2n`aUlE&keqE#i`Zp*SSn)~8v@;3a0*Szj)>cps=Jl*${uNU0T&aP=n
z|Hp1V@sjXYQ&khwFZKdRFMuhXW}BeaCTLC)6@R|JSoGejwlDr-|9z8}RN7!8+EI>7
zpU^By&Eilp@kq(4g+_ZlGF1!q3m#3<Yq*fZFJv_Bs7;*7CeDluM!W!WCR<p(#W~$#
zBZ(T$sfLXtLV;5#8%bOQZWr0e;1Y1V#Bw?aNC%T>qvV32xnMZuJ7DM?mhS_f`)ni-
z1ASwlKN$|V5)^tJfIgJdpjR67p_AY$PVGQ{%2m*}3i^{tAjvh*s|NZ|&On1R*vO)c
z4=iJ6!xq7B3+u`MJ^IB-*S|%-lKl7RSCao8{o?Gp;GIlo7tZYq-YF~>2$KthDP;Dm
zwyOi1gsu$}T=XFJPZYa$iHIH^D;*W*bP)W#Xd~a_lkWn?o9_8n+t%B6R|Yv)T>D;2
zlINPubs!T>mASXt!S!b|_g1l7e>QV(70>l&Gxxo2g1&6#zSnJ*6QD%`RI-_Si$bVm
zGxruBpps_pi$tkNG$!*7m)wZbU6xBksf6XnqUmGNp3U@|K4ZB~G_9jd64ThAbo?fO
zcH&lKzLj;iGDY)D>TccXSNq`1r?YqA_S(*AZD*@j_i<YLI5UzH$9n~wRvc$WszfEZ
zjMFM(t6M*CS|2zwlEJ`R?m#OBm=RNicV?m1akhH(Hq^Qe&1e$x9n^XU&4~YpuSb?K
zh^e|bh%67X9EU99Sk6OI9_k8!RaGNW9m}@_?OSw<xwtB9a$KPuhZo^gZg46$fv%nx
zxWN}#KE@9|Mgyplrr3ZlV#2rlW1~-!*|=<2_1g^EL%(o3zpxG6!s)cI4S5Ub-(t&#
zbD>)<^q~!x2i@{m&WCRKEEhwaVz!}=Ae|#@*>F9ws%J+?DBnAjXUc~2ME5+lZ1}0D
z|CE*uKeF8xm@~<4(f?eONcvt|qy2X$^!*b!odkAZ#R8pJcHoo%{Sx;58=+ex^q~W#
z3A#10{06$cVYwCRw6X&$0qG>L@1H4{WC}LK4&rRAV3N&pkzi89a=qYCFASvm38Gbk
z$Yg@@MfZHRy=q1MTH0QZ2jvfYmte^KtIcS|y7RMp8V(G=o%%u0{va^FgV-!;Zx*|_
za^z%~(tnaUR{ZbY7|FljWv1gtMYR8$Im>2l1Th?B*A?%?cl^SLD85in*bWAt{t?gE
zeb93syZ>?p4!HtH(iQs|9P$i~BrEg_-}mEV7X5?^I&fK^=RUUb>=5!i#Bx0HjAuC!
zc_*Tgq^s*%roSJ%f3sV=BW-jMTF@t6fxWIk8_E^Xr~+D$Z=iCddp(0Tgo&-e@7VDZ
zDe{hFQsY+Y_(c>Es1^+CyEOW3i2Wn2-)T$hfAuJ8sCgV-mG$`8|MyzMM(yg_ja$d7
zicbDnVKe!G6)tR6{pyUi{&M%_9@HNW_6i=#(D2|<lu|gS6wZb8FwWa@W(AxDO<&&O
zX58UsQ_ez8S%}Fymm<f~zZW)J>|)b>zw_gMr+Mk749!oL8^2W5A^rD|;O*LL>-1Z$
zX#L|RdAr#!)%wq*H~;4)Fz;GUB^cB!#g%_@gkT;aIFo6K!)U=gnsQ4^Tw1@<Jx$n9
zJV5pj*!kRy?3-CWB{-iFx&r0k`Craa<wjNn&t0)!^}$tj==d8*KB%eTUmlV_ck)3E
z$aRlQKMB{mF-r0o=<4R_HOSrDW5{4NpDqoxA@rj+{WzF@^q?QT=tocb(T#rmf7+{1
A%m4rY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.E96B152CEA3EEC31.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp.E96B152CEA3EEC31.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e349c8662e4e3c8b4f9a5ef1a37fdf9d4fd498a3
GIT binary patch
literal 15352
zcmdU02~<y8+i&l)&-o{(`b(1r4Me2je@Zh&(Ig>P=J6ITagCRIuX}MtWGH!w26GCf
zG884!O_8z8B1OhrWT*(=ZrK05ecxK&yS}yFcYXJ@?%n7AThDX$*~7D+{XD<D8|F7~
z;A=M(6|e987KMk-pM+FYR65}Q=gprx-JMra@zGaNS(I=w<nFB9zVe#jCc6zcs@pHr
zn&Gf<cKlKI;|Ie->(v+Rm>Ion;Js+Sg$`fy)z#k>$?qP1IBkMz`n>%IPVCRsQ1t!%
z<9%a`*$QvV6*pdif#(+0?Y{T%-N}7hXPb$|>tm-ZQIAhQIX_q3RAvn>Jrymc75?_P
z#B*N!lN&q3#x09?9kTL-)>Hp+FGrg`T&!Z|ea>n787~o8^j$x<Y@KPKf#1FL69cEd
z4NDrn-($LcTK6XvgJSpRyiqmX@+xog(7}fXMm^Q&Kj?ne%aEt;{)IQ198X`iOj&mP
zmFJ~Lsv*w{8#XumC8kWNydALKe@Wc%r%P7*pNtF0uQ&R~whS#xkI=*zljE;{=$`h;
z;lia`bup)PwTj1dpH}R4@5P8ceoGTS-}$;v<!7BA>s7nTT(0Pq?qB*H2++|VetOaE
zCH?!2UZKC{)|$`$?LKE4{`mH&R*qDu>Udc3$oAbCz2^ZsSInN}t$RD%2@ROve#Q1J
z)1%8!Nl4_v)jluZWEK8!DeK9lld%cKD@OhF+H<UF#i7ba+1)g9|KXYHty|;%pzpHS
ziPBrIOGXa%H#OAcliZ%zBzVkxaCzFUH$|Hd4ZN%`OH)+t)Skp^Z|rgQ&-Yz6onBPB
zFyFti;MIf~G5IIvf7vGgGicg`&Su5&c9Q~grUoU&Tz5IwxTD>O>~n{b+ilL_Wn+)F
z8*ysusGF$thX*$ucHXI5>NIKYh>iz$>ZZ<wCT5#+!uQ_rt&jXX6g~a8^w-Kk`7u+!
z`Uj4F5#D>s%mpdppE_Z61^aT2{gO6v+o`zOqhap_BV_k>IDK3*CvTy<Zf}*c{l9%~
znEdOFh?DpBU!Q+(?1;du_7*qK&AT_`SvETOykf?8W6sRr$DA?TGkWs1s3RM1cIbc7
zE_;kemvjCV!6Vm*r4cdw$SkWeE1!-Z7QPsFT<1rNBWH#ux=)YsDzCdTTy6i6kIVMT
zhc1|2WIArqY74{P{d)z?i|syt_bS);y`P;1*Zkys=j80kH%cevE=ZC;KM<m2x&4Iw
zpRXLUSMJL1zNoB&>&}#7@ATE9Y-bM`8l>yE<h_3xdYv9;T=qlgpskTp+sA1?8uUyn
zX=>z%-0P+xFK_HN@p|C%8m-83vHbdFa>k3%HU;-{&w78nIC{#Up%3-|t1l*z_su*;
z*R45xeO;>TQc|E{<@;EtDZ-dd!fO4r>!y}wig5=Xj_r4Q*U84SQ;oec6}vmV@9?R^
zfy|+S6RL|FlNRsYyKTl^r&(rIHv{W7I`%vs_Du1?dul+a{Kex=_6ryPFn?dH)|g)+
zmzKU??N?+Uz3Sy7YtZFxX!7Q*iz+g=HOx^tTo)Sp%B?EDG^xWcD`P%C7mj%??))Jj
z<jp06nY~=k?wZl-=d-(J_v$r8VW;agWlm_t&OT31YFYGj?H1F`!0+AQZmFM3qu=Ik
zUYxdmQO%xpbKOUFm}(eld1*pO(Tl4QBhH)8+^=tVd0*3+IV%sFtg9>j{kb@1O6hq?
z3M?-)i|ahVX4jXfP7h~$&eY!dVBN=S^MBNzvL<P3d1qg=?3vHt=o1rNf3ZL36q;M~
z+bwIy!e05~2CezkbE4Lr&Zzd6Cr-BOZ99HvU$sG85OcAr?(LO|qSVHO-bo&ZSKf8I
z@#f%8=j1J#@sFlnzTY9??81WG%U*e2ouK=$$o$?{AM?V}Zm{Xps=7moKkeQ%R;pV+
zZsB;DTfwo|49kU))lsV|bi7t>Kb{e1bIfY|YH4&|=bbqX(=vL@SpBxtHs4^oiD&Hh
z69Tthy`MbvOaE|<(8WtUq4~(qy&DGf*i##w8I-bCJL}@ly}~j-oGSR(=r(V0_~ZiY
zP*v|a4ZeG8^G7Xc(nCdB2GKI;()LhM#l9B*6PLKS#0vu}wV*0ixd?170=KpYQPI5G
z=gZ4L^C9h_1y(r$)lNXYwue%s&U1;6<Jp+@uq#$M4Al<9o^21JqTAJI?jO7KXX-#6
ztBf`nZP8K?v3~!Wz~YF_9kiek7N-G}1}xhiM8#-g%E{~<t9-R!XRNXZqCL>K?Lkz!
z{;=Bq%^xX!1*m1IBKgXEy;};Eb`apc0tr~`o#zqgk-n!rRL6CsL$!41*7hKE^yn8p
z=+1_`1MOjZta1#h9fKxq529i<AZSUz3%Ritlwp-dVABYE+a5&4E~nqteqXy^(t;XT
zWs=RLF)alVE1vqCSmiglN(*+w;sS^YpsejdR9xzvy7zZ|^ph6U#46KmrjKbUh&W)w
zoYL7T>1yqv8WtafY6oH0wg*ue^VBcAQz~lGf*r6*G*Co?L2VDB68=+$)7BAQBeb9#
ztIS1nJzEMQo@@EA={+<<TF?}WH$k)sTDCoi%EFJuKbpUvGerj?tP)}nV%$;?@#3_k
z`~K^f9oB*RSUl8Vs8LHn#7jTj89u`3YpfPD!{R@H!5_e??LkylX}IofT6M}s2Xa_t
zl))&AmV$^QZ(g)$D7}!O19h=D$RNn5r6A(9<F;hZ3>>+sJruF{I8-~1jkondRJJY6
zI{fZ|^)el(ja32-0*zV<BHn-V*jzIojX)i!hs8q-h8VRJM10`VSI4z+T`uWB0gFc(
zj5Ka3h&b&@<L6!D<-@h0F&3W$=p?r5)(26^;3N0dRD{0Qfd*J5*dW-br6A%9?Fx4f
zX^L3c9(Kgy@oM7@S_&e*64xb_f99*(9=5~c{ZMT`wu9COQMqNrg*~0-{6GilV3olJ
zgN<4WBCgoGNmS9RSg8X$Vev47VMZ+l5m#F{K0Z15mq;CGh{eMVh8wjMMEt6Akb%m;
z#Ps&CI~I>s8{4U+AmWCwq8b~opXTboE?6939MHX`AmUHsHqKt+ST{-snq#rQv48iL
zf{0b-hK9}StWZ%Aa(kP)4@f^Lh|&?Dc?9$%{z8=g1e$+>o@AUwX*Kj(4aXDLCrVo(
z{x6VNvM8mY{%L3&S%9Lni65|upGYuy5302}O5##P=@2wM1ivF5gg8#1*5)pWqPkt=
zwjchXxN9LwfhK_##JWUju*qNxG8Li}WD;aSY*&<qm<+KXh9*iwO@>;Kg(FJCOomyI
zITEE{lVA&CF`_iwWVi*HAW;gj46z~WQItknj<g}xDN3U(N7<02Bub+#N81pOEJ~qW
zLwgd3B}!p>Va61NcMb1JT#+cv(wk*W(a(lIn-LEpO1~QZ%Aj>XTnDts`WB@q*e?qD
zkl_`jwXok>7RSSW@z94(Qw<brpca`HQF;Z%S1is#`dO$ml^;j?$61`qo8<BqRG!QC
z$>qH$&gVVy`2iFc@LmPHAI0^&yq?#kxQXx5#P?wFfF?eG;s`+=Au#$0f;@r6HG*l4
z(9PC=vdi(iM%zRwR%RM2>qg#JKbmsNlK=5!dr_K%CW$RubMx*jO-IuiGzZOLP&7c%
zfWZd=I>_QYfbv*;5ul4KZUm^2#Y-Vt3Yq$1A&O=3ZiseMtf_ykzg_wL83K}$LGNVX
zNVXd!z2thm<g7{GkQ67l#R={dZxh_Mu{cd|O%vRx{E*;!h+-0d;3e4_e2|qUN<~n$
z2<ng*HQS;Ptk@cYP{nd+Q4W=}O)Q^NYl%UqTDV@gIW_Mr&LxZUCbbifgKKXMPv~n0
zpy>hVO+2yw%}mv%ec4hwB!%?}>r0jrl4kXp)tA)XJZQnmU^pGL6woaNbf@?$?E4jZ
zkgp?YIoEeN$KVD|(ZESmzJnjJgAbwjGT-YmubhNVn(pU3lTQ}`QF;r#-@;&`NTMp-
zt=5o=b}Jj2W<wVe4|Tpd>g20aS?4%W+5)V%06WrDQHtQKBRD%!qbQYgTIHN_TFgC;
z9Itb}qV~5M57JBQzMI3uj3ld5p(kRF5$}#Iqcu1rdZKg*T3mu7XeX~ijjMEiESETM
z-;i$?0&(wohjM;6Rg4l0qo_5L;lv+qjV;L+NKlCdtq~;=SW8hymuab*;;9>JG}el2
zxYpk{xf#PHXGA331THs$8`+VNbPKrLV)1R@a+}4KK(i8bp!$`dcO`J5xEdH%17j8|
zs)0R=eX4;U#XI@&JLwjRO`D&b%zsZt07<F*<Wv?P5iE`f-HF^L=zAaOz)w*CNi{&d
z256FsZSOn#hW~YcHMY1y-n5Wc8oMYh7YvsRO5+x#PlDwq!IgMdvgGhAHb>rMAhFX0
zcCla_HQY$fE|OEm0*d4B(t}3qG{tjx32Iz|T}cb^oI;IqiWQHO(hVmT&2EpREMS-g
zjEHYQ(ivcQ1{jfO1WC_;`7_Xk^ax2SVds_5glvOIS_M0=V(}K(c?&cl+aQu&bAw)U
zfutcwdczHR!{P>RWCJ&f%G3G&>2#aJ?c@7rQB3r(jkV?+WN0McM{9&mqa1l8RRhgx
zAg3*^1{T$z8;jklfjh-}`9XVmriCf|kQ9C>`BHD2s)~`KT^u&vV&1iw3Y`0I>3{g6
zndX1o7`oW69Cp9(z<VT0$AIH8FqBNbD4hk4XTeZ1`JxoZYsc}nWGF-_p4X12m}~{u
zN}IDgG8H%?1$LWQD~;mpqNtU+&5oS9_gUl|94g#~4!5ClgA%0*=upApDri*&mHVnF
zJ%Kh)U|-svT4+_v%Aawz&*&B-O0}GAEsIl;TPk8=o<iQfkQ!I>s<c@yF!g^D+&&2b
zWL3Ir^!anbz^;pVQCcJ}vZaCJ30{7Jx2Mg@=jHh<F5u+_EdIpHKT+)AYFcQw;7-Q>
z*-LmX?<hR|u&W1N?#VzW8R%2LP{vu7aaN>7{nnJa{#7|IQMGx*O>aY{sTI()g2h!(
zuL|}gFE)?3!4t^%samL4%gUc|hR-+~s$a_)*0MMibx%dEMD9JzaL=zFRJ}EtH*x#7
zkO^m#Ie9W?N)$yYm6NBkxQLS%vABejm$3LXr+Cfvqwl}r6mM931!-SFj6=MGwC}LE
z4r$l1_yVtefoJsJ@+NPo>5I~4q4Q>e>G(lGeo#=h-={pmEq?TXc?z7Hh@rZ~pX1y_
z3}aBZ7_LiJ5zbA-Sqxe%Uu-~4>?p803VPF(dK3g61tAn?1FLLSo&&6MSX>0Gia>9=
zvWq}q5eT8U7+4jv@)BTG!r~|>MnPTT_M5XIn<(f^@mgrJmc{F!xQ^9d55@H?j)!7A
z)Fs;l_AJmQ9y(K;0L27Wo(RQ67S}_~df16h_aP)dgbXOoL{^!oH^n)~Du=~+NS?>)
zpG5MLEIyC4&$HRsdE{^&F`l6mX_vC{3rPC{i?1Q=Yu{prYwY{ik@j^~egkRWVDV;N
zzL{tGlgP^xS)9bnlUSU~Yv=MzHkr>m<nv5>3wVbD7B}(QO{{*2(7QxX2IrzwCg_(5
z=5+jO1w}250kMmhb=@Rm%s)})nkZvDd5X*^g=U9-1Lv5Hii26$#XaNtJ)=2Evy1Lm
z%i>fNn2Lsw4-=9Q@As_?A=zXCoi@(awO$*IU0Mt`JcgS<-VmkzoWp+B>1S}mGg$dK
z&fy#@uSOx&Xfl2OBXa(T7|*m@u-nacJwvd|pg1tdNv~u;+)%umZsY<sa)W6_nOs07
zi~mA{|3b>ZLX=Vjs}$-&MucZN4bs>Zg2(7SlHW&6u=toae9ZTu6(<RHNy0FKM|Lb$
zOib6ez}vw0n%`SdUo%s4rWNrnBR|C{9_Rky@NZ9c$sUAftd+1OGd9*ezT+<b>8$^_
zcximrjMsxp&sO1y`M>3j6RRVq3YDh!ap>q{>ElEOVsh!R{<CUqxt9HYKq1T7EEhbN
z3(P9|Bm{mEhLZS$>;^dAY9_i#eWJ7zh&zEV8F^9KhvfUvAYvck8Jo{dec;~-HzOHp
zCPN#NG5_fNYEE$Zd_8P9t3Z!cz&R{BW64q@`;DTsM(Djp@Sc7zf6t>S+out?j8p7Z
z3Y~uX2sAwcy$NwaxV!dhEgU*k3v~%~r-m8k5!RnU-eEozZ2^icz=8H93m9hsQ@YU|
z1>=r_i4;e0iU`hu^c81F+|XE#NnXo2-ExlEC|+}eUvtVuNJvUNq|MYmZSYPMyc2y#
zt8GA@4X8iCzrK7h=vtYtE*{cy&cB>fvKE9c#PzkF1<)Fg!!F061EKRpOa*m00UZc8
zwEXnpc2$AX=inNj^7>DCOY$i~dEy$Ij}B;kp?aa_bl74!mssjXS54LI<MG)a;hC)#
zjH(3_VnU+ySg?I8NCXoCS5Rv?bfC32z+nxL8O8=KxPfEFV>=4o&W_Iq<oSV)k0>=F
z&qj))Ecy-EaOZm*1pJdoHOik`@p8jgXDw0k*YP(Z?7Aog=mZ!MDqfW4Yt8ROR)Z)l
z&|1KtXwV~?<#rl@Un3Yo$O)Wj1C38WhvF|#;|rzWL}?kPv5ay*cqinv%2)%c=QQeB
zc?{BxLB?dd@GgLKH?p`0X%!*H0M8<=vn)>Ib<=p}{Y+jrlf_wrZkC{wN6n}*ygK+!
zBDRyQvTj>ty~u~7o_pVz7B;u(U!DLHi^Yq-_1K9TNg-4%q(is1?s&l~FO!=8L4&M&
zm=I+B`>0UdW*-+HH}d}aC@m>z_v~Y;m^MM~H8eUx(qrWQn8i<#`%@M_NAAyA94(kd
z3th-iASqgKiKbgLlJ*Fl_6UYlzDH2(5tLcb=CCtQ`gStsuyy7L@*IH)qs|EOGZd4f
z5F53Y12)nTe6j)*hk!kGN~eJ06tJh!^<AL2OZg|9=K@6qiz`8|N?=1M2@%ssmNlR!
z#eYD>AJCpg+nb@oX6Q)VvMANVVfE~4{R&;bLM2}<N)4QK1IOg?i6}UcUA-BoX9iN5
zk|<@Pp4lueL>`5RNv;Z!S0Pfe(U{#rK81)mdniN!g-A)MV455G6|s5u7QW{ep84K3
zzUMX;@8E~+U>V3{eta_J=kR#*zK3{!GM_k^<UQW<O#9yPLGLJ2gZFL0e2c&&b8*7h
zIAJ_}f2&}=m6ab9%nnj}!bfhxm^5J=)lV19(ph<l;8`LlcLz}_5xh%;fmFXt=u##y
zXE|j;_cDPwizySV%2+nAOt38z6!iUC!M&E{C29rFTK0Rjf_E+Zy;{MyR$#^_Q8pk^
z=0odCl=&sf7#<`==9nULCh~0?f(C|o6%EGg^*Y!2I^Eq860NF#J+r{+-yTHnD(iMB
z;jiryZov~l|A}BinwA<;)>kcihP6zTjsnM{U=Uq%MZmEL45Di;3hG5cYr1CELcO&t
zj)!{j(3-BHOr)KO80I4fY3EQ(&Nh)+%i%d)`a4kY4t6QkBhPxe@rhCbKPrJAL)uS{
zGjV{_e9}&<33my1C$Tza0z78*XOQ<SABwgB$1T8x21sXs@fl!B`&|Ua6@iHqM{tf2
zoC|4yC}nX&v%U?G49hvj<-Oqszo7xrfpX#Y$}8>WVc$5R!vq_0mM=>0Iq^NGG~t7b
za`_?Qy9PR8-vJ!6L0>vA*<e662&6a%IOeeO9N?P+hERDia4ZIWY5R-8fMO6xaS3oN
zVdW*jw*(BK@^w&e9VMhiX&vmg4%$#VSP%8qv-0(@$9kxs@&u@tKnZM7N`!ieERN;u
zVmakn!OR6`7ti7>&Mu3^CpfzkEY3rQc}TfeiBcZwna3u3dB`P?^|g7(H;*R7qLh!c
z^N|sq&wON_Pwli>-Y1{+&G{%GA1P1EM5z>+lp^KXg(zJ>CKp(I9hqEb@eO2hgT)V#
z{X@iXvM-SR3l@Jt=3m(Md`0G8S-g}tU&=G<EQ&XeV)0tud@YOD@#gDToX8s{^33=p
z@;wvTla)l?C6V0=5_#W5UP&m3QW9^N#J-=z_ekOyUz@}`C-I(ie3STrNj#I)ALHf6
zczr4_;eAT@Ac}AE^4qMume;T4mBgSZz2x;@vN&4MkEX;SrZWWn7#8ml<a-3>dwT`>
zUW$p`<LIpQQVw+~Y0xwcdJ_gEy&|D-hr{U^cp;xd@^gqi?L_i&ipkv!Tx0W99$MdS
z6ug^yx}*EffoIc3uD~w3N^q{CS?bXjKOfawQ@8(bby}Je_TYlI+W9IGUr$h<q{m!O
zP@m4AIqGv56b;m)0fP?$^@A+V1L}D!z6jJWvbYhbH?nvsR9_02`eLDaEQ@zT_1!E!
z4bf?+LE4(@Zr0&Z<cD>5>pLch$JnbG7X<kQfguYk1$iZlKML}XEM70uUN2)@ezZ(G
zn#D0P{TP{&{L7oIJNJo!hb^Azhl0^V!IC78c_r6JrMbPW`CHBO>FtMvY0G7Macj#t
zvvPWiqp0V#DKGv!x)u|48Av+=Dfh1Ca92B<#anpoE$mg9ZM^n2ii^vxmX<z$clzHR
z1)YDi@xE7sUmv_=H*n$xHh@p(#B_?!H{@Mw@Tl#D7612xAhLa9GNJh%9$ninfcOF!
z(S+kO(EJQcNRoWnz1y$Df{Zr*ttOh>R>4GL>uo{Oo2xe#gjkuL{k@$ijW-@|Mf%)K
zav7`wmXzc=096k_T~ZH@@}X)P)FlqH`M#X$A*f4kRN#wyQ1uAZC0rLa8>pI%FKxD6
zw{0fPx~&l$2x*2-6J`3#Wy%{<*KKR>9{z4(EWQC40W>3kG2x0tsS=u1LMA<X4h^3}
z#y%hLZVy<RaH(LnRABTY1w|w!?{C&bjz})=0WiO~6}WB%Lujm+09+GT{Frln%qjOq
za<|V+t>t<p8Bcs@<rfw}sis+evjS*1y4Y~BF{NUDG5m!=89<T2rp2d$;xwBU-vf$!
z-=@LCD#37Sg)5+91)KIR=UkU_Oj?_Sf|KZugpaS0-&r=z{FYY0+}Ag>0%qF2p%p~w
zTUtSs-XYI-Y%KpCdA?`yR({l08rx$YmLI*7A4@}lRDOIao3<Y2eGju~YXR?3z=kSK
zyh#&pM&EDZ`!%u0)8En@;(wz#L}{Bab{m`4eoK3Zm<C~K4{YGVm_uw@n<Dd1ktuoO
zTM;+)g#`vT{s*6P>s`XIX>n^-{o5TNxipGHk>*>y)F5}E;9Yd$?^~`7zA(JU^ndF9
zs+=!-)gROce=9#mZolGAHs4LAo!o<h_s~vO|8`<x2mRi|a3{YY`4_~PNfa-SqGt2L
zw;?EQ`qmHlmcnf~@HSM=XY=)nffX#Sf_<vkGx#Ua^9f{Vi(1&HmdeS!ExdU&-;$;c
z-N4yxU=P&NIlFX<>k^Kivdu8-`nMYNy607EYKp4%{lDLccq@Na`^V}ahkrS3`|gZ-
zw>bu{qW|fl_^{;cZ@QlfH2>u-_^{G1r|bEw_y2JRe)XRY9n^asj;}CC=aI>Iq$G+(
z=^8S*Mlm^%;MH3GCV;MpYtZ5vyP>~=9bZ8uhmZMMUay|6SW)^Ybp9wPk9->E?F;;J
z{Z>5=VIySv5i%<ha5rB6-Qj9oW<C~N;4Loj$}F6mlgQLs{`!D6<TA9l%r>M3cC29=
z5+UeC2#hO95cCpQd_m}Zfi?^i(?a(eK|$K{=~MD4+ltNaTF$=hHgbM+>&oQpGFki=
zvj2-7Nj9H-%Tp)=IyZFYq~<(l-rOnk;+ub1H5aY_dXgV1DmKcWP|{Cz?tSW!tF`FZ
v)2f%k-rB~tw}YLEtx`k9n*Ma4KOO1M-t?y({b^5s;y>FU|9X)hDk}d4bo_V}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp.75601854417950D1.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp.75601854417950D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..363b7b0660b33425828e3e7724b27792c9b5552d
GIT binary patch
literal 18098
zcmd^Gd0bCf|L>f;PHy#;w5LS7Hg#K+HcR$>RI(dm*I?|+Sd%3gOhQ6a*+NRB5Frvp
zDUzbd8k+2c-}zSOdyap8&ujj8e$VfD&Agg>zi*$<x%b?2Kj*XGK|Q;7KhRxGt;<h6
z=S>^`>lju|O|2#Rf7Y)vC*9=K)B+6D)aJ!r3o4&p?x84|mAzc+V!ZwO4|hDAGr2_r
zN1RVLa`n&j8z`>t8XC}kpxOR}viNUjx(w9o6|L>HZ{6afhqk%IR^8n&_0+<TX<ph|
zEq|EdcYR*R^KjI;2S;=ld`#OrX=+x&Gra>(6OF>$d~BWytCr;U-hN_L?(mkzBgYLb
z+BLaf-(75bo%e05cD2h%shb$HG@P4QBz7tu8@nws<@F8o;lD3U6$}CoyPWs%dvtfN
z`|TqGt&%D>=+?&0FIfLe`LK}Qi6tGq^3tF8DK8$JU~uG*VfiU150=#LupD@)&!Wlq
zY<EY$w2AC;L(R)%{+=#Ng5jYDqx&r%J^$S0aIyTf&b))R#U_t-U#Kto?EYhH&Ds)f
z%$=i;FCMk%c4V_%^!RS>LF>0xOsTyz$?l!{J&XC-4?me+3aa+co8hn|bhkgh{m2K0
z&OP$|ew1%obq!x|@t4c)D@UAMKYGHmOT!iK)~pPznmOfEt>4VH3&t-xzV)w|Wxw|J
zaarv&es1MYbMvOgbkCU&Ug|x)dv0W_&;f3*;i_7<c8=3($MNn#mg@JWoGr}QcmK8B
z`_F$G^;+yY)cvOSqPDRSHi=jN3g~co9^?K^`(Vg3E!#LB&$|t6^+(#ImtNkycb(HR
z^W-X*$S1!#KG<pgW_8~=v%6U*jTS;}a^}0F9u`;hW6mUrWq$}Geq5R7-%bCGZCtAP
zzH}GsB>VDDcZ|;k*$%mXPIF#`RC;Bhf7P&FLAFtM-9PfiI<F5@O-XY%UL5b5d2B@W
z1)nk2zkPaJIWOq^V*4S13%4J;GH?3<Gq1kEHJ%mzCias)4VzQkWxv&N-36L^pU!wX
zHR;lyuY;dfb~^fGZrDcmqtQ!F3_CM<PvEu#10Eay7%_8Bk!D$I`?=i=^d{|n)nm|C
z8`BZm6Vu}N{92VZ?`L0^#V6%qTW)Qi^DTGPJ29!_`o5PJSWen=-Xmebf`T>+Ubi`G
z^m&`y{;l)j>%%5SCHP#_LJU9m*qrHZHKc!G+3o?}SD37nVf^)u?{B_z>^HUYll{$<
zSD{aG)<wKp^>DkBW{dv6EeSj8H-58Lc2;iJh2akk2G^D!)q46d;-mV)szEhuKldbu
zqhVuA=WVcRv%aW$?8x?M*}cvVJ+rK6+Rs-n^Ow$8znkV$y6r&l$@#M$+0O4;c&gs)
zn@7jB77ZDL?M5xwpU1ro=#e_6%~W6apo2GBus^ODyR7@f1`U4t#}S+RoH=B-DI{=t
z>Ald2rPr41jEz-Wo*Urb$KqBuQ|*C5%6o?~>=N^+c4o(nteq=%8yRQ?-rifKd+7HE
zz~+kK^9fgX{g$iqB{HSkpFt09J<-VgJ=^k#|McZ|=R_=u%URW_eo)@*g!(QmuKEAg
zzd>Vtsom?h>zp6Y@cSv()jDSnBb;>7>5;+C@chAN$<mC($H^}SsKu`N8Zb9zzQ?h=
z>9OKnv*MUmd!ly_?RDs6`hqKN%$tO@rNgqV=M>FfTJZD9$;SduDD*?FM*KO_r?h9a
zZ<wXla{bh|553kWylMUBxZb%Ir9pq3)N1pfy4^eXQmsoTrXFtC+UH?=qjejOJ~3Pu
zcl3$Tx@AY_P3h1t^5XU2s?b3u$IKm{W=z`MlN%z%Y1(RyiuYWk)*@}HnE0^Hf6LIP
zUL|dx@2xj{#~P};Wi222=T)^iPKSp_9kf4m^WBSyTNW9Q?Co^d&phzOobp64Da+U^
zX5riy5tXfeu2Ou~y8B?-4!&j0hg_>t`!Q|*DjzxV^U*GM-c1W_zbB;Mf|mLbbIeYM
z+DvG_C@^I7$qTFJR<?3qyWv5)dHiohB?BD-)3U`E({xU3miPON|KpJTmXM*FyEpXD
z?0f6kUnABgTsma`=Zv2cJzg$!>lv!P=&nZa`)@5T*TsxcydCB8`S_+z4>yimP?^(f
zU`Ps^J{62iRtpTAyX}qJFU~P%ANBaDTcPlDdRlC<!J)|EVbP%#VV9!Pbb_~>xi!q#
zcKo+@+>m!iGCO|hHFn>$YraM~uSRG(zcD&h)_Z_om*Atfdpf=`9p1v@@T~dM#y&xc
z9%}n`o>6z>xHfEOrzWohhU<V`^GQuaHURwxAG>}S95gPbHRKS>IKy!kjR#ED)L&+3
z{&G-`HZ(`%r+|G5IyIj#i`MC~Eywom_+1;CA(r!iJrAs#PnboKyKZaC9^<yOh74jE
zYdF@T@qo$Num2kH(^zpu3n(dX0@a^F)8-RqF?=<}=yB^Xe{E=rSkeHShHlby!YpmB
z{H{0Wfyu<y&;YRn8wMLU9x%CmML^26qrV)~hV2o#7&sJzam^>pVrkT?RB!I;i7ns{
zh@~2;S3|k^q++qw8j|#Ef1x&Pi&$1eb~Q9<K4BK?sCgR}|5~oq8X6&%VTQwu8xNRl
zBc9Hhqjj-s3#g9Bx1suN*uMFMS!_K9Oj^6s%Af^ofmrT9^*e~A>7-)G+}Z!S>7(El
zP!qA#LiJjx*L+g3eBJKeF1z1RZD@^HUPJaZ6q-+%MfSJK`WklUoi?;UEN=n(7QNl3
z6K3(;Fr&ly^b*h3&;+pzHym!<c)(=esp?Y~Wm;-!!&ZpAkX`84c);ZG!#)R(=x2SR
z1=K*~8mL|ay_-*%WtvN5b#y^i>lRQKu}n~(VAy!T<moq!`f4oF5L-Y?M7{^r??LzG
z6K0uVGL+Any%kN-7_o#Jh8i~>FnQ+6iscpC=D%(Y+aU4?!x6@f2TWe2U10Zl=bBgA
zP#ckFI?N1hJYaIdQ}>iMli0)BP(<V?=nw_HnopP|G1$stY1pN+ts#$C#u$z<X*^(Z
z(tMv5$;XD~w}vbtk2V}_(s;n+4Bed{cRw`L<zZ`EHF=OQ$g}aF_5%b8H8miE!tQG-
z253L}p*7S)<A)j!HEulM@kN&w8)+I`vuO<t5jn&##JKT*$&Yi_=zXeQrLPTJBJyO1
z$)Sx0On&;NL&>x|r%GExeMAm23^Hy!U~;X~mv6Vq`joYRT8MlXs^3LV-*i&3EL>zh
zyXT~>+E73&H=x4}*rWM`S?XqPd2-G|=WPpUgILC^k2h#MsK{rwY}$V!r&b&4AaXQt
zjs|_2PnhMSTeW|+#;CWgp%r2oVlu?4@qo!Q$NxM_!;9!UsNSsQ@Zw5nZ?vn0uC~qF
z25UV=y;0i{wmEup%&|mBat-5I!*sQ6-bl>Mg<hWLANg{U{1J*Dp%x}cavk5bj_*xy
zoG>g-7(?(G(fEwmF54yW@+ONsH%ZPGjk5{nvt%>xZii3vlAH;&GJ!4Dc9MJrXk7ue
zSj9^6UC`w&7>RY6B!33(pTQu!k0rT^4X9#AkQ?W5-SarrL^PZtwom@Ux!gjMhnfzx
zz>+1&L8d_#SQI2V*fiJzYg0)MF%7Z6LMF+>Oov%u4JXN=rlA&i`$_U})8Q6aUr6!@
z(-9VU?@97#>(LH)CrI)b>oE>^+eq?Q>#+{lL`d>D>v0bF{Ydik_S0>#3`=sDUYH3%
zbK1|bB`<5P-dq!cRsf?Fpe<HUlDrBSt)lT}AZ-R{&o>!JNv;I`l^~ET>^<OrkH&RC
zssq~C97^&R;Qs{#l7)|Eq-aK)Ec_!@{}J1U;1{g^3mU)Tj9zhV3B8Ups-y81-gFCZ
zLFm!EX*7*@@rqr%E}>`gicA`p@J1y(b^TJ_sFcQKyipmA%LU_d!F>HEyK!}$-E}2-
zqiD5Jw8s*wk-jh6n)~InB*~N6$x@T+pr$6tGufFGTEH%#P&8np0fkcmn@Zyxz~<1n
z7_h}OegfDhG=2lvH#821Y&fK@9|hSc8sC8I4TABjL%N~KjUk)<5j1-Qhhkf$rGGQP
zskm~A$jZq8{khdO>nB#;!#eI^o$#Ww@^K#hxe=Vf%NaB-;5!xYUW9&~?{uBUD}>+`
zLI|O6745f*4tUY=u17O&@~X&;%b{60?27#aymyc`Q~FkfZMxocbFy7>8P8mz&9Ku2
z(0Vj_TzHuNt)pVY-V+_QSUIe7m=7M$%5yu<^}!oVxi*87!4R@Ghk@l`U`6mH=yD0V
zV)Mz$m!ZpL8s|ZmJQ}AmKB<f^S&MYWC!NO0taCEkk<d%HE+w2wH?#6L&gmQHOz6u5
zd6_`Hn-zld3V|BGRkYbE+Tyh~(DKgnOFDZINOBU>HHjIFFJw6U(7r9U50|r&{F;|u
z^DQw!lHc&s8ydgmrMEPG$4l=B#yTCn%_hGUueT(ZL9;U06+4Dl^9t&X{wrZS$P7Bj
z3?cVVW(Oy;KN4Kb4lbtg0nYCL7eMHTIKM+Q{>cCMkq^T3Hn&1fzsgR{V<dSeu-^%s
z@T-*M48}f#QHi}IhqK1vtSKI7?tAQHy~lO+f0VpnLxHwjQz$`R(@UWD67<39Mv^O_
zSq1zNdr~;YKw{7oV32{~jAl6FiI>%~InqFkO<@LMd<M;)K~KDat?s=TkYYA(I1=4*
z-m;u`$5*gk?6E8Aj8hPl<dwW)B~QJ<MBXToC?agn(ZeXi6>_5zc5n$TQOTTNGLaxG
zut<Y6g)w*_)=P-c6t$=b*!@y(j#V&0lHWq{E%JbyT`hb@S7-`cFr$6Pb?M8rMKkoU
z<BiALVb9Y3rv@(*xHw0_%5hLD4z?tECk~>2awj+mnk7MVg7-qRy);gNMk&yQ&{Lpu
z3Ur}yU<&L>a4FZjl<SKp%*y3l|8j01!IfP9N*Xutb`88ep&t+|4hU8RM~d!|qDl%_
zIa4&q6paWyQ*_M~DZQ2@dS!{7G2OA!qsO$Dm1_h^uIJ3^IhDeb<XGM)mfRS5#JoW*
zZ-)meeOQA!8pjBhF@gu7?+`3@2rA){<T$}GPVmQc=SOjejYeM#lUO+e7-axstnZNK
z2S%5HF}7x`Tn5^d0aLt#SoscUbBD$sL7R`j6uVZ)%0lzwupL?P1h!8CORab!+b5C6
z$!u^kOIg(iT)+da3!W$|Kjs1+)A*BM_eto0EkVZ~57+VrA#OY?r$DU~s312^ffgyy
zlEz*s(3{{pT%S9%C{%C*D!za9<_&y1JQ3NU>Ta;)6oW+h1E=`Fsbl~@Iogg*u{aif
zNxl!w?!%!tCYIy}(Ch&migk*I#aOj5TBBw&tlYq(KT1lL2{y|FTYPI4nMU7bf*-~n
z0b2voKDE{QuYY--F5!bJ&mVk?LbOxB?GzY@rBITu0=KJRAeKT&{tEQIf{u8tCHWes
zbB%MvD<a9)Ii2eS<FE#4V`UD6ypdb%;9EpRdWA(yNPHSGqopLTWn9-1Z%dLR8P`Y}
zZ({5=F)H6sl4BT$7{-U(e+T2dgHdfSwA~n|I7aoxBzX_hX%Ewj@b6?Dcd{z~Ns@Q7
zj=O1G!FpA&RES#7xzv+gpi~bIu{?GETE0Uq?@g{>$9Jfsag5*@BlINn9fHRWp$EZn
zLO`4_2xD)}&KE{?Z$BS}xXXaF3^<a=>?POfB}av<n|Q@0-USbo<eR+bO@27Rw|V#5
z{4j!}MBgZJEZ)Apoy^WV&AyZJf35)W1EF23OoI^#*$bq-pa=1$%2`D@+Xv6zZ+R>J
z$-8(PWTcZp>tvu$7{giZa8_j!{nzGqdU<cwPIaYdnXP506^dlcB5Ayd(c8qRHoH=^
z3}YCT$&ut8jL8n#pNnIR;}}Zx_b`@wh$DjhLDpy|>wssh6fL9OG_GK+D%eh#-gS^s
z!lHL#S52iW?UKZ(G%s>=8ATamh8a;dgHhb2@drlnfySR1#b+8Pu+9lAwI>o;=R_K>
z<aAbY?aB2wa5@`koW$uQ(YS%rX`t~jUjG<xO~zm271wxbzJG|`e~5jss09|sdhHn2
zZB|E=rI4oU;y?*yDWot8&5`El;;>1Q=Sp)av>Ygw14Ar2lAHzXvVcsKRu<@+1%e1Z
z4eU<S^fSQj42|=FT|SVBlFJ8u^Fa{7*MQwMnqC0x3TRvjq)MQRO|vp9;!p`Z2)+j#
z?$P)@knYp`RY0nu@fRR{0lGxVegO_&fCs_ffb@-~FNM-lNbw(Fv<@%^<b^z76%SZL
zf@@g28dfH_mbI&;aUH9uqxt`06@StAE35OB&dPjcUB9xF68y&Me52`0Ih~~(mDP#h
zbRxdTt`QtHehsIyhNiFObk@>1pHt*>)bkW^iXs{pbBbabzv2|HXuO%%+00Y(Dd80*
zJaxS?UQtHlY{4^Ipf+Tlpr0q0ll7|>oT~-OC~Xw=H;U$1;E{?F^^--4KTXt6qj8~V
zTqs)N>-FqDqiskGOjU-Itp83{6|E~(i2rUHSFn96*a3K8U-wO6y7ucgqIA;*X2=C*
z6egf(igB$VX#wN`F+(2E^ajSYfu=`uLDAe;GX5~<ahNuH=Xj@cBodP38@$sE8t)Q4
zc8Oic3hWbo_K7_S&Jn$H#6W`c#V+|`KaBhKch|ezEoLAxdl#AB7ny!!A`h9~4{03F
z^@}IYq9o_>c6mIdbQ46E1knv&SduS^?w7<)7!RFx)V+^pd=N@gY~mD~Xutmy?{tb+
z8DdG^F8Xd4$KZjId|LE7P3&iIt4q$K59?bXdo)dJnjKMUv$ba1VKo-~ImY=>)(Y4E
zc#3Se_l{Qa+>`15^^WVv)G2lS@^h>HcJ^g-b;Jbzo*7cy@<8i>?%3@Zn}0lDdW|E~
z=mvGG=O{-lmiLS0sjXPc`_%GOny8NVsiSd>&^t!xPXu>|&})Z4338k;Bu)s%beu&+
zj&f7E56?)Fmol15i3dBWU5<{KhUQ)*XcwXOMc4+%QA%umk)#HvZXRpZwW`-##27EQ
z#S7F_{}kN*BzS7hUYi{|--iC}@CTP5k?W?cKav?|Kq&(nlZll;trBQj+g$OQmJ*#e
z;eIazwDfDB)*9%5Q#Lq0Lu=4fuq0Pp$!M-5UjHv1uVxG`{#6e>S`=`O0)9UY56W0H
zVSJq=C-CU6E8eBJ#)7V>tZ$O*m9m3NN%(=wEr?B7=Ok>m*}=C7+oBiu``j$*p^GNI
znhjXZs$TBWD8qfzhAwWed?L|ujc9}4D&F{Lg_IRia^da3X*(D}eAY9J(-}tP%&(rH
z)!Fw;FBU!DZ9)IGU_|N*cLlq<f&<3uEc^#V-T7$)V*KB|;B~#2l`o>cd9;$`A+3iP
z<GP?EFV<dcK>VmB+Dj-D4Q!(6eC7?<^9CG%U!x=!L(O8?n&1dVj9|3zvrF<(M)N46
za>OP17^8WN#;aKERV?MD9Aq^Q62~3MGpl=%H6ho_Wp#6D{E^lENRR)-YJZ|}HK$w6
zQR5$Rx{ql5l{frK*K+m=y88qb|GH;gODFz3v*F*K(1?5x+u#1Tknu`Y<T=m_3rXaY
z)ER^QpQF;uzu3PU)@gT)`!nn~vMSp+?`_17V&(0e_jVe`a^A5tj^n)JXncTgdw^tG
zP_cq<n@r;z-XMpkl4=)tg9`-zUi74HS=PeCe_N46|D5yK$H1IHTKtHhI3iG<*-1fh
zlE$k=oz=v1+8))vdr+5){gfMzZL@||W#wac+Ep*gneF}`x3uv~)<s>_bno`J*Cf8k
zvLN>5u8!!@Dg>PhLAB~wt)d)nQ(FZV7+e@`q2Aa>f&W>O=XlQX#+nF4QoeHoDAaXs
zAVE8U+fLv~)Y4^Oav7MBI4B>C$Ooed&S2a!=!mz58CXM?TEbb2aFz-T6WM-=tSZTi
z>)L3-%JvR2;pv{!y@>}E<{K73p{`+p1nmUQJAo^ia0W2R0A?ic%K{^^z-WRq80QSe
zl~O6pz})XAY!=R{QjN%*u>BIqg!cy(`KX_mVvoFoEZ~*}`j89f1Gjw8hpbE`(5nRY
zbOfk(kH%ks-WOm`BDxw@r-r3+54EgLEx`wi`LgiqKg>e9U}DRO4%qNY@_ts@&#KH#
z^1Li=z_j@8-pD%zZl{3{ncr#9?KJ31@EPEChNhnZJ<fmugnkXUT?0Pkd9HzO*Fax_
z3xHbzO)mgF3cvtDzYp~8ljJnYzyizrz=5ns70|1q=~cj{3OEz`H=y^8B(o*?8?gKa
z9LRbtg?dX-w_9@~7IKoH%~I%0a4zGN%TQVP0>-I;#x;yn4UJzgPA_O&#~Rhqf>6iW
z){!zQ%KEdOb#w})j_pxL0x?uYV0G$QV<Kqvta&|opGtb?dbS%)?_JNTs;B6AS<`Q<
zDz%NCk276LQr$@1ai(i%yp}UvOXIDa%T|txNOy5AyJ(!snWu7;yrpsGX*52@nIEI^
zS<d_{jn8rB=V+Y6ndi{Bh%+kUsEt&_*%r~!RT1Y|#8D}!BCbafrwU||Z^0QA)8mUd
zn_^O?R?<C+IV!DN%ylp3DDV0?r+CikllA?;_4&Y2Mc_5O^BP`NE|TN}y#4_?Wp$X>
zKTP8sUXjC7*URM<xiqfk-Kyy{@NQ9mw`hj_P?UBS_4m;Dyl8x0Y)|MHMB@ttAHExV
zez)t{DQLfcVHICU?*#HZImK#DRgz4J?*qDTK8*T9<a@luJzka5N#C0Xo=yy2i3qm@
z*IPng>@uc5Uy!c1w*J6Bs_cwS3$mHYPQjEoyX^6qzaP)|`To_e>Q|~XP<xKXWId`q
zM`I?17HBM>P&Ckp1{6*O8mTnS0U9|pE(RLKG=2g!p3wLW(0D`RaHtUusq05UjVK!5
zfEqVwTo2iLsEOyE<=wXB!-#j0q9i9mF%fEGcR`Y)1u0t4C0k;Tpx7f&Nv}hKbcm*(
z7Zm4dda0l&rSURRS|(E03l|mPqUw65!gOaoHuQBw0$3*KmkI5$0G_&B5^~7vb<ICY
z0MG6^!cSZ-B%;+=&FZdZ&GBo<v0CEneymsG-}VnL+TI-Z{6hMAl(~Ay>O3S)fFwU>
zb)M7sDyMUmQ#lEeT*&DZ5`3xXMt=UYH)sFjuJV<Kn<~3}=-C-5y&Oi$p(`sWzf(o<
zl@F(Ge(<gBfEfSxsJGDDa?zmv#?k*MQN!J-Zt9Kw&RF$HayoEM2QI|Ly9k^w0vF=4
zl>+Bd;6hxm_rUc%a3e0&Ht4*KbabGED0JRV<2cwM4muEhlLY;fU>AZ9FoO=zzIH0p
zDV3orrOB*)GE0@yZm@%IkmQR}%xxdC9q`6cin;A`8oy_K-_tJHd$!Aa+QoX$2EJ#h
zgx-6$_j}U+fSQxpo*(H-?^VwBDo5R~kh3kMaS7*JLc1{KT<>zy#(?I_p+BneBb_%+
zr;859`C-TD0__Rj_ykSA#<#sjf^SJK;71kkW61Lt@@)%gdbZFtTj);c*+P$OfvR-p
z3GMO(sx>xGu*wrCwU8&+<q1@|Gf!~L6I7jnl3Xo#R|`~;rdsf;CRs!!y=ygn|7xK}
zHGThR(JETBAui8WanM$A2*Jr>yJV4?U$SVGOj2peB8pwINX;)zY?ns!r-@c+G=G|C
zmqznvicXm#)!36Ix@VD&A*2>Xt3rBwp=eh~k1rG*3q{p>+$fzLWM`K98>%Z$FqvRS
zLb51e9tE68MzRcwWl$G4eW7MED3(KA-0P+6a}z6|F79(v_Ns{Yp)SsaDtlGL2T&LH
zsz~x9C_aL^IA1Ku&!G4W>f)|8Wd_tTfp^8riaXF*^~TQ|h&H&%4!%k6u7jN4LDHj!
z3OZcK4K5T*Q;B13dD@SoH=_oTNT3x7Oh^ZKEzqh3CZum~Bhz*x(bY&z@NKvA)V7Z2
zo#S~`!^N$dh@nZvHUOo5c7aa2zyK1<$AeDsG~Uj3+D<~O+izkAO^jK)`af>ADy`QK
zzB<Ha(!aDuRUU<1-Y4sWf0Q+<UXnSwrM^4}txhVZNTpl#&vJ^h1mm0zN|h_WwjqML
z4LEJ1Q{`EVQx-{;<F7Al)f+#+A;ZFYhV{nRL`konb9<9C*K%OA9GH-__ex;2lE%k@
z^Ksfz<^tzj+EU&E&bMew`2o0ope^MN=)B{5%Qz?w4k4CsFZAC_TgG&zQ#x%4i`c<M
z^xpi$_WVR!rtj0dXp4Q9=0!b8-=%p`57c*QUP<1^`R$|k`hL!DKaC5xkOH#zCApLf
zETv0fm0a&i+R{DY<R`SH+svD8rW5Ob@TPyz`1>?7((K=*nNj03KP-c`jNhl7QE~gb
zv@_~C<40YmEn@@UrhyLLR|rE^2t%<hljKi=)hB_nM4yF_&q63xnMlQnE|H=tlvaM!
z(Qcz?LFgOB{u{-C1aB5YHjC8w7||+*Si0vuKJ<^7wCx?r8m?u!uVqv{Ov>iA?vXU!
z#B|<7H~z#helZMX(swXjb}*C+6UPL^F;w%|9;VM8Mpd)MHU*Wnn!Y_FyJI<{xtx^y
zaR~>7$4wuiVMg31j5ek6^E3R<N(bF3hm1$xPccqa1Y;eJuGjQg8Xj05dotTGrEU9v
zlnShWR<*X~V%6UN_Z?`j6;Eqdtm*IiZ?Aa0$<eT!Vfbp}zciWe7UwR}{d`XA-#(*v
z;XO0jUrBxYU*8b%{_~M!y=T)R(b9fpO~0~KyJG}r8o{YBE&>0b-uOu<-l3Ab8MN9=
z`*(GqRUPf$B|_aqXhw`#G}9`YQB_Tm8^`PIqLqCnua`;V^Fo{R0#y?#7uuBz_BcEA
zvE(<`8}&!ephQUnXVE}X)}KEoWjo%DMjsEMmLR8#v=)5GI6b6sJm(V6QCc>SSLE?j
zppzi#Cy35OeP0rdFNre7YBR@A9ivRA&6+uG)?Vcgd*z|&KL`9nO|1j@=}dlhBtIQh
zKcW0;y07b~!OLU~&kptuu5OOb9o=1=Jk&f@ENbrLrw94zLVmiDpRVMm6Z!c+47`Io

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_wmma_cshuffle.hpp.7EFDFD3F43E2C561.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_wmma_cshuffle.hpp.7EFDFD3F43E2C561.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4a42614665462651fabe0500b629ae7b33998fba
GIT binary patch
literal 26262
zcmeI4c~ligx9B<D)!nDtMvi!xmBS#2ARsaw;Lv~!O5!XkF(xsS#u#JNOlB~OfPis;
zD2~jd1E2^vD=OdsDhY!FqM}9(ASe=45b}x-iYnh->%PD4TQ_T&#b=%0w|7@hReM+M
z+SSv8gMv29*V6L;B>4NqpML!z(9+Tx0{^%C>tzdn;<U78I%#Qrzw2>m-IsN0MNQVD
z6rIOsLeD)Pq%Zr)|LRwd!O~5VZb3lZxUbg)WK3Gu(>2{<RBYU5Z!b)kwEASp?-fSN
zN{kYX)^}Vv)|%tAvLklT@boDU6lZ^IYfno#)phH#az^m)Lq?1K)4y@9`Qpc4JAO0?
z?tDKpX>D9)(n9^XkIb4ooc33E2<ro!l&^kMS}q-$D_gWbxFE;w!>rU<nhg=_huy#H
zI4kPQpOn5|tZ)d7UXT{VbFMobs}Hw)|5tA8J?6^Oomm$nOgu^g>+1ZD1x#A}<HzN$
zzt3Lv_uuwHX3+x8^gy4PUwl$(Hgn%6rBmlL9=zvks|vQuXzQ3+G_^@~W60Y(XBAZp
zZyQn)a`W=6A*Bf+yS~paJbB>L@Q@uN_H=0-Z7^k?pVRVNK7YJ<(CLx)O`kY{y>_~1
zX0F@OVY_`YW#OH?vp7Ij*tEB~Xtm$i;SniXmp2?(^|LSA5q9LysrRR@ziC_<Df5{5
z`M9T+%u3U{!xB#|efatDXPS4DDh-rwhyBgZIA0kwu{^?XQpf$$R*M~83Hk@;9R6xS
z+LZYdN*7yCb{qLe>Fi;}t~LB9?;PcpNsS*Lid$%In~-=hcZ6-i=%zuT9<~Xqjo;DV
za(U`%pCMa@c%NUGll}EOqqCNU9334xW0dXYf}_{#@9rzlHGLY*+w7mA_2tFY??oTo
z_SO%rbF;5T>!yzYZq4KGZu>H5cFd>G7Hq%%!J=h_F-LaBEj&4QQplg5@3Ecr>`Cpx
zxfS`@S$hJzZA(9#n&Bqy-4Z!1_$mEcc(e8B@{%9oYyHMd%DuJu<2j#NIqiS-B7AS)
z<k2g}UEep~F48Q%Xu)&+3)3Dw>)bj#^5KgiL5I|bueV;XI&yuAx!Z>0wN|t5d|;TH
z@@MX&Nl!HYocp^g{f<$h$=%VNogZico08R^&a3AX!=mbrv(uYHR@d{_8q}8tr#kuj
zxf(kv2PKPvjU$%t8(rZEqGDD*d`4Z|WmnoNyl^{fryJ}(F=YQ=V@y_>eD`z|eRkd4
zigL&0N4I_J=^Z?;EpnZ5?t`UH|FpknT5&JLXUI3-zPRvF@pY>?YmblYzVfNh+~Ove
z`rM#D8nZ9Et(v{!TXuR_lkyLiwWjXu-#pSav2>s2_BU%TuHH!Bu`+I><7~QLJSmEr
z+IWJq`egIS;|60Zv%^<4PP;Z!uT%d>Tfmarxj}ax?rv3P{QddVIj7hR?dr{L`^GKI
znU**_X4h9+9)?>CEqH$`e`AGp_VIV^rA>U1?O)$5id)&rq~FeaYuYb{51xOr?YH_F
z&s~OgKe)6dKXze4)}D?Pdm4{B%>3qq-}e0Ax_b{SSjXf&4O803PhXR9$8*H8_hK&@
z{}6i~Y#-C1On(`jm3%QWA?&MZMPqdiKk|#7X!=FWI@|jhr}f-D(q@CalvAI4eZ6v!
zfBD3MQ8BXuhp47q{5C&)y%l5W*{t}YE$nOgDAV`;eEPAg<<gRjf=z4R-M*&r!+ZS1
zFSAp}#c#+r{^7*$7U#0=**R=F7yVJ=Ow$pwoTu+<-QAF~NlWkY^P9y}BQ)Qf+_E-&
zSd!1}oAbsz^YhRhJx)05yIpH#so~UX1rM~Q9xjZ1HpJs{qnq!K0fxV<Ygm2Q_&d9I
z??qLdx2_N9tgc*Queu(;&UC7xj2Zv^N&Ra+P0LE(edoZ4^u-k(VW*TjNyf@}z2CkF
zm@#M89=BnCW@JTfnVVedZtvJVx_;2ljemF94*%Rg?~k2s?QyvUDM>BkU;gp1F1tGV
zptf@StVIdxqv-~d6TG}HzwPbieg5A5b>E%+DXm<)^$&OH*`!@TsVn_E8)`f<za9G0
z`A|1gv~}cZ&EDhRT(;>_+s`dKw3KmtHmdX6YtM!aeRbq=-Y*GTP87HP(6D#&BcHiX
zPAxw@XivA1v7gcUisZ!sZZ8Y$(s#RgS=@UnyAi*zVZq`J9b@tj98aC)bE6{MDrc5g
z%#HDFezq$QMwiuDc7H#yxoYjCjwG{M^^&rrrHk4=k6)7J5tZd0XWKKiDt`8j{0#fL
z8~JJW$8O|j*>~MomYg%^MyK;n<5G$uH+~kke1hW3#pS-vSAMzm(ft`KzSw$pa=HFC
zTm5bEqo<l?1kCW6zTIf!tkJtw2eiG-+e$U3#(z0s?um7C7ylS88+-M<xvE~jwq@h2
z3FmaGZ8n^JYt5<$Up(@s4~)LT`!|_TBi`0=7+d(B(aNkclY1jCX_-&T8e6yQ)8moz
zH5<-dP<>-DQ|tUWU$JAvXZPkGT+u$PS`%3rp7M8f`w2%It;&dw<)1}ddeQRZhnk;c
z8*RRdtuL5+cEiMK7t@_;kNe>d?-~x7ztHn(?y4ax$KHQJal>LK9e&uF*80KxtGmws
zcqzxBZJB9S<Ac*1cLWDjo?80#>DA!|KbcQE@hVe&WK??MidY>(hhL4t0=G6lt=68i
z<ZSF|tNM9=$NdvFLU+{tIInY;fH$-1#6H(kFXFi}YnR)@%0?DXoUUW|Yu@|UwUgF6
z`%R!eeATwsp>@h{8@fkGo9EwhTxM<&Yc+hTxJ<M4iQsbmhVpIC1zW#ZSkf};?CAre
z=db*A`kK3S;T7LEjQl&s?XR7`jk{l*XSQ-e%#M!x?__qU()NtjUcO@SM@w1gj~&;w
zE&sTvRF6`+Xh|0-<BQa={z6L*eKq_~KgZ(9m)5teMXhnpTGASse2vVxztH04Kx_OM
znItm}s|y{Ta?E!SWee-MMhVv_qy7TbW2fu7gg#|csYmHTmGhMGd3dV60;!C-r|(ue
z=;p_wRscMewcOg}KmTa)EY)`!m%+`ktX<qwU5bZw#IX)>?1cUT)#1`Jx$TqOMLTsV
zGpJHcJ5<ww{RL8SvoESB$}JnEM>)B|i)Br(v7Y?}Qc>sIec0M;u|kh>fGWSUroXde
z`wOJvBRh8?q4>O;E@cHzw2KA1Sdab!srX&n>6fOd4A7%S!&)j?(@I!NUx8Et2i@BI
zTt2v2ml9wt*MP${;MHFsl}X>H|EvlSSL;#su$FVI={Z*2Um%tD7ujza9XnIdrBtw%
zSlS_$9@k$Wl@Gk#cbuoM1&dn#@Y&v$$#2U>_ZM0`NB2G71?y-3IQ%1%!-FUh)^mpy
z?y#Ev0@brDz20d}q1#J6N*}6JP{tMTHTD%qCEUp4m%vYIcIZ)K;RRQ-rq%F*`-%aT
z_k!mcM2>l;N7=zzYANGdO4VN=m54pLZAG<vQ(0;hR5{HKI?WpO7f2<hVR2E^$=$!}
zQN~c^56butcrASeQrT4f0l1ejyk-zJ2&z0_g$Jyvzd$P63}^kY;hV#1JxUL%lvBp#
zu$I08scd)e$h$vJ@vSap4priSLmcqyFObTP#!HKIDh;O$q6R~id#rGe_3AH>O5*yb
zMzv2`T?bJ^pvnzOxIr2B7f2=fGw`W(w9iL+R1kc0rL1WwYu8^Ol{A6>a;W8OOFe2R
ztmOh_d;wl=Ux8H8t!>21zL__4DIKVC41i-`Vt;{Daz8op;+>wK4D_gxP^F9-UPg`V
zF9uYu#(n+IMk+v;vV<!8S+E~=lfD9}<gM_ZV!z^-E?vqLsx;CLjkI5Xfm9Ay+V;4m
zrAPNQ@I1yLkMWZF3ZznCbGoSd@}cE}C~a8FBUX6C8uS-PrPw1hJ-npqv@T@>Rhn3n
zCf2mS7*LsQek;vv>McFW8miP##x;~#e}Poah=1;q9?cn}M>)c$Th5x6!-m;cAe9n+
zkmYwRKdW`A;jors+MyVlkG=w_lx-+^ykgPDKlG?TSj!F8^ag9)Um%t8`nb0FYZO=Y
zC?i<Q70UPud`Eo+QmMRNcV*<@&wtdVETGC!*5N2SzP~^!*IQ0IzDTiv&D#*FT&9dK
z!&>?Zq;k7XyDazkNx2?n0#&Lh<7#-*eFajvbEz=2-R?YmN5i1XUCQAuHLkxvD)$yH
z$h7?a%wb(>1XL*p#^u1Szd$PYPZwX?nQUOmQVM83N+_;`GVd?6ct!^O1oQ&`{H(dR
zw&;bs9%TUQxkMRXf*0IZpn9~HeL8>n)!ABF>@im}O<?W`AWH4DK|4JbwIxy72@H0E
zv8d09QX1=@#=ej2ttf5b7jNOCWT;JOkBQpx`4*?T@gj65s`Hc;@-!ms7OD%B73wh}
z^lYk&lofJ%BJAa=a>@!B01?_TRRv{*Oq>X>U3H1FLboNthp)O!S)unL!dI%gLRq1=
zEkfI)`h&7UW=WJ<D9aXV6nX=_?9oE`U~Z)>TS<94W!e5Zd$g1K-IS`EvckvrP^uo1
zSJA3fG$9YCRpGC*b2#mRkB^{L5u`klRz;GWKwBozqmXwIr3Bg|f%d_iNLwb7@-*5q
z?REA@BlY*us(rK-K0b$5<&d05tMW)bNUILM&dvvE4}5$9ttuepg|w=W<PO@egC37t
z)(hJ51wD#nw->aA<Owh6$(Uc#UN31s%n=|c0!+jl34$U?jsihZB*%cD7?Kl!Uji79
zTXqt#Oah}wc1r>pk|!jA$(WOYS2FO!oCAV#z(mZsASjpQ{UB&R$@w5CpX9T^<SZ~p
z?n0DGfk`RJRY0r)y2w0;QU>dn!H&m0F_ZPnB)Nzci&)*<e#aq7QJfgX>7tG*N_RN;
z7qEW5<BttVKctIN0|);IzQ;-TNPfUc4@llGm-fpE{R4980LcgC(m|4o<WdpIJ9xt#
zyeYn(WZp2D<a@mG9&e20O}w&+<aS=!j#;}f`;-my)d_78+EZ|s68|$^hX06g>;+<J
zf{STTOma01s!497K_kg=0K@^}_#*%uA^92r*GSG|KpsQr$FU%eC3q(bc9NXOf;^Is
zvfwCY9mDEzW2$Z~8U&;%Zd2URupCI!Jf?YKn(i^3pb(D`g7(SW_Q~AQ+yO|rGLKxD
zC+0kvM;^%sWF7}dzR3pOWP^~a1yT>^*2B4@mII{og2Q>i5xps7(G^1f<OR7aQEH=o
z+vst_JUzzI-NWyS(stmx9gKrks^K;*&lIIcAmkC4g@)mx)C@wJNq!DOo|D`KLb^yU
zW<rXYS!k3aN+nE43CUNOkSiotF(FkX$8po+xDa$B(8O}V3EbP5qxcQ7D1Os`w-cpW
z-nW(?hc&MAzSl9MM=TTiCu#c=?M{7qG`7g_>H+zytV3hHfYeHPwNe`727nY!JB8C@
zNgf|gPrw{OJ4etim?LTDNRkt2uS8md%rcM;(oP5Iu_TW_NKe3AKsy)EE|?2x=R%Sr
zz>EkGg3K(Ca=?rn5Q5AcfbUH62zaA+2H-o>G{4TF&EOp@ZvmPXQr-eWTS$2;@NETt
zSpPZDJSX+LfTrtp4(%e3?*^J~Qr-<hyGeNu@a-Y%Ph%X@m@)W#>5OAK$r+4u2IGR|
z#f+wy@kVV7z&55Sd7VQ`n0K(ejM0>l@-im0jFg{ae9tj{SpN#6xkBn!F`BB^IkbvA
zzJ}4%kn$QPw1$+|GQPEB{ZXuA6gvi=FPe3XCOL+6j$vJ}ypDCMW5?pof0K2(Npd~w
zTF<&+c|7-VJogFe5kT6?eY}<A1kNde8;j-1oKy1aJSmx*jO8huZwlv!d6nS2N^l|B
zYn7nE940u2k@7IXD~yz92!R<w5Z2ESCS?f3`Lcw-EK;5&Ov=J?BOT2J-^}7tT9h;%
znhD7M!f~?9W2bBa>hPkpft$O5TZB%7LNY+;55!=bn9r&6x$(%Dic*+d9VRFE1+RF)
zs|G4MoxbPuIs<!3l<I+FJ@7;gQ<MrB$3n&vy*Sn74K`1_AMb;TExbbuFQFPl=_POU
zk{^Xw_1DNE{@Xh~@Qy6BEClQ}f6@AdpgNgWoos+Dd<u@VsVjvy_DgzWzuwrde~<mL
z`B|N`@#L7<&@9~Kls7pSY<Tv`mHXrajOOi@LVsKsTdb$F?^ALVw+i^L0)co`w+;Ai
zBS&>vz&{HFlA}EJBhqlS0Cfw=&w=_m$=yKRO>!7x7sl9Q!?%X9TSIarGdhxS!17{7
zU5pK*D3vklGLo+_>MJDIFzOnTZ!mT@7<*(`MX8Rlt0TFA@oQlGvAl^H-NZOxPGp@D
zS!X;^i{+-rlB3Lc?t^&lLu8M8N1gTw+`w>Jl#)67WRfq)oi50ov3`Zzse<HfyzMrA
zG?pjwwuvNX^Y3T#^RPUd|0tUu@YtfXhY#Mv6QkZX-lvTp@W7(9NziN(2!C=}FuW`b
zOw{bht}9%vD>svg(A5pQM%f|XXaC77ui@0sXA}Sim#u3loACa)Gms(~ok(T~dSkup
z8p*h0Ud!mLCFScGll9DSEMLz!t!JD`9>1Om#@xV|H82+FjR5H`V|JJ1CPt@;8G_|a
zjB69)j=7o9X(r`uj7b|a9Lw7nr#8l!<ne7xFy_62)m~vFdg6|^)DsrB+=_;M@G56{
zm2<^?FiUQkB_Cikx`5M%-t`wWd2o7=m}@exbcqohd{Z8eD9`V5x|UCGK4}i5mz2j#
zk~=BaPRb3-J1LJ&Qofb;*h+h5&-4jhlGyMbkg{oyY?61=uDfYBEZ<Fg>?Y-<v|B0d
zt`+)Wz>o8DW&o+2_9&-4F<+<MuG8+eFDugZ)+R0i(rwz~HtmUdGh?xtvBWRxC1dfD
zvBVE9g$+z$gHYE2QYssmO7cF=bsy)3o*$5EIFB066LT%+Qj4E2Y<QeoEje?2#ksyB
zkI#^MWXOr*GvzLs_<aB=Q|^{2CuY8n<r5yugVBo}JK=r|YxIsA3#9do&U!|HZ)`nd
zwVtsi*=s$c!Q95^v@r^N<86#p8)HqfR~w^2jF#SD!{}RgL&g~`xKTp?G8+<zQls3d
zQ9c1(s%x{`^)D_wnF>dJBftnfJnCC51FK~O*ULaXX0%!dPu#auhZcs^<F<}F*lDiw
zrcFx!yV@jLJcVzpZ{-xdF%Rd+>9@}A{To`Lm$d37?TM!N(BS}8GVnxGd{4!hgu?CL
z^)R9o&4|&AE^4)+w2=`vlKhYnA2PbAmBN_^Besy-&WP=py_Q5R$oMO2DKAPrtV<6Y
zfEuSLt>Ro(aRi5RcHx{CmPd0A(H!A|aya)KauRiwbGym~Av+*S-JE+jDbJF-WyuM7
zEbkSIU7lvJcgY7qBUZr95djP$0P$#YfI$v^G@^7Gh^K)g3fdw=MGO6l&$#;)um%O}
zL>!I1&zjw5XQG8lA5XKhW4^hu|NrR{e8Yyak2+FELGN0}>ejJ(XyWY`Q|{SvdwBxa
z`!LOVSYmvziZffq5gg7Lgmbp!!!(ZO>@nwXmO0oc_C7qzZjxj9k+Hld(nss+uz~fh
zrlSTHFx@2mrw)0*5K%fu8=Rv@qUAieltUX_BDsz>sG~<B;{`*}v_S*O8@RVOaPOis
z!lfimae*_#ueX9zRFM3bQ$FTKVEJ~ra=V=1U2^3vl26E$CrCcZD~|GJ_<SdL#R-zD
zd1W;}0?S)@lUCjia~rQ}<B1X6CSlknVPHfSQ0^O5;^(4)FWYORmkaXpqO?Oca)-<X
z#Sl<8>$SE2Nku@7X?)c#ubp!OmwWWCVHvhi7AP1f!eF9dE6MGYVLQp)l(L&L#rom2
zGMpwjfi_H_El@B}loDyfM3U2J!!(lj(aL?aDb_zoD-V*~K`T3GQ*2o}Xtxg97xQ!4
z?K#Pvw6c@b@1m7mBu4<R2oQ*mPXNjUV2Ugi96AHH1Q=89&k%su58M(-P65gkQa=?a
zQ%TMNUO6BTAAbfY&H!W7y+!FEFn9>e(VOWFP#ZLp+yM+aNbUj#T_m4k3{ElT*g%#r
z1|=k4W(+QqT*Vkvk(|Mb8LU1sGrem-4jHT(b0+JMNpcn|W|8{Yte8!5ET<F88R3R|
zg)_LqnPdK&Gx(e2B)LJ7oRF8w4N6Jg$}6_wH5nM-=k2!huDHG=-Y$vcG+vQL>ZkLH
zbdvLUMc(T?CJ&n)QQFTd_LK4hyy5`Kg}kEhbskel9$&;Oib(k(UU7)zR$kFc@=M<K
zB|i{+?41e?*(8{t`wI?QYW~hPDj)id94<JAn}{t?Hy7MZax6b7mY<3ad~5rujhELn
zYeTEv$tyZ}JtUaq9=Al_{?|>cC>3yX3%CWi#!H;bCC&$PxO{H7oRBBT)d_NcEHC9}
zm+~K>ilO`9)pfi-;z?874bBH{n+7kvkegA+&B6yp$Y(^z39jTPSMt*a4(v~ef-bKw
z8w#0!5so)pe`lwl0SF8@u=+RIG00&}d0Ux2PmkILt>Awjpq#b%fcwP3sk7n3In66h
z^8>Nt(4ps?=H(h%!K{VHl*wbt9ow~6l))>?8S^gMWEbs@?bKe{U@z^AIJ9e<(}QDc
zT>dpnY;!d`ho&-*zI-#<{$DoQMlA~VhQ7o<<g;Nq3+q<FzJ;=$L@ATA&g2G0Q79u2
zI{UtKKs4E(cQ<MQd)w?5Z2f1c5oal9RO3f0?fDy<?+<}J))H76;u+j38CWGF_@)fp
z#Qaft(YP=Fbfo_?(wuOWv`Ja=30d+1o3%h^>+((QRyI)a8D;W}%zR0uO;T|t3CdH1
zbI!gbMRbL5y_dCWV`n2%vC#62(oB1BmO_*yLDI+L9v`KT875!Bmyr1>{Rr7OVH_dz
zSNapO1euT^(?_Ex==){LM41`pB$<#z%9CZvWKy0Y6H-Wds!W;snw-&1VN9{_Pho6Q
z@WP-dMY6h)tSR>Gk*rN5UPxQKKmNjkJ8v1m_jrQUIl(%hoU^agtxIQAeQf{*SAf$M
z;QLnX*|jS+wk<`OPrT5d(utpI81N4xBd)uF|86qE+6pwSK*E>W12jD(pJO!V7ztl$
zEu*O=Ifc`tkkhMZxb7WItb2zYoo*L%^UzAz_C`347>C>5YMJ9|8Nu~3$9l|b92c!r
zJu0$?w;ClVqJ)7FHcC?M4O{#tLlrxbI#$09I}H>v6ovi-Rs2j|QeH19Uli(sj+Aoj
zq+Bp((_Y!MFA8Oe(r(&uH|>JCoc1cGeNpI2grO?0+q5qVRf#aP<W<A@Vz*e!Il|}?
z<_x)4hMbUR${k_!2+J$wzLoMo)ECeV!aMHEro}hX!>N1l+;!~&9dDoRw*Z{i-R2dy
z`Jt!=I2h#>4@rK-8@=L(VfjX3$VNel`IMkMB^Y2XhC-4r35J*OU<M`5*$Dkvy!Z-h
zDa%^QYT#YKD5TdT${RoGzbMDQC@0J<)Z`XwD)z>$ly@s>Pg*IzR>~jC+o{Rzq<$ym
z)k)UBhE}bi&9Qzs9UM--jX8pLji8CqKqT!NN%C4+wU#_SidIFD98IgDN#06(ZKb_&
zQ`ts4Zlj$rC(x4<=&8u|_s*KU6R~~lWxqt)AIsC|$!VniZrW=%S${sQ%BRh-{y{qU
zApJJx0@}5J9L*Hcu7xBQ(W)Zy_(Qbn5XpyW)nSrLX>}=%+KSS3T74ZyX?s_i1L{E#
zHYbHlKp_)^%||KgP|A)$>Cz(f$?p3&58U*NxX>bQ4(3;!`zul&DG!a56Y@&Arcyo-
z-C1YlH+AEUPc}fS@p{N(-CK0{<Bk8Q^+m}+*TD#T5KqB#2qul7A?R(rw})X_luS2D
zW`upjcA2nUHU#rdnXr@OOqpJ$Y#5f8QF>+6u-yLqZD`to?g?Oo=ZhzS?n#p4*x)#J
zDhl6<(q?w>W>ydLG1lQ2>xx{1C>>+{j<FLk$8&?@xq;b%2$L%YZzcJ$eDGsAA>YI6
z?%|2~-CkaIFUdW;RS&r|v|G^MEeuB{VqIfE?Pv3sz4{-IXRp7vYl{6Z?|ceh@mb3C
zEai@D&4zWlV-wSs6bW$6hJ$|t_7zC?IrvBLpPck3$@y|AAKPOf70BTqp<gJM3Q0c4
zYtG>;37FKzYtEB=k=I-#xt!OOW8Qdug<4yy;Bl`<)E{9_dPM!zGJ@-6`t_L6jtJ-m
z`nE%$8;9|F*0+(JiC#tg{4)<FBa7Z)Tb`iE6HJjnl#U3BBbZVCG#oYbZJWR~t_J?A
z$x%Z#@Xsbk4J{y~h0Lbw1|i)fmoXt_WVT!l6H-HRG8ddoX3O2+eQ)4gI22TYx7xR7
z17FxR`N(VXfuS&pi9*G`of}xOowaIbXQ5l)m0)-Gr!(KU|7)ggXwL~F^zSsmcNWix
z@ysx6V!Ifziy4MZY&@$I&#JJYXpzfX<c8=;Akzz{9evwW(C`9=N?ErHtUube1VgW^
z+a=Z?O+!)&){2i7R%OFxRS)#*0kLXc$mkd1Rcw^D?=JLjZ^4(|N(Hu3<FWVshf@7R
z5#D(<ty)cb(=b{UMsf_TiXl0X4osxSV{eyBt8z(iwVzh)C;0%aIzaLfT6KivXSB^T
z+7_oLtOY@9!2}#bj0Qo`B<}z=JAf@tSI7fFd0+x+zM^yh1RWsx3@|wZhGUyw0!&Is
zehAbL$@GP0pl&9)1E@Pl?gHvAlEWFh@YmBDrbjX%xT&mVtk*I&m`^e4Q)K!=38O9{
z`7)!vOmY>Yt|Iv+V|SBGZ@A0&-F-d1VR{o2f}3VDW8KWyU`}G4lE`$2t=tD&$@GTT
z4Y~M#8FCR?ang{B(2A3W90tN(GvqK{^_n4v`%2_a7s+&pOLC`6B(H^&nYF}ZMwIsQ
z!F%!Q4_vNz%@x3%EU&o&c!t+p0o=Lrnk#@)1>Uy?8$Or^!TVk#xs9LI#)qQL+Z#ok
zu}K&(_3(VcnDfFwm<zQm*tz?5kKwLY!-_R*2%0^?^*CO!5m%gTe)$<Q=_L=nYK)+a
z5eAlgcF+Dp5)Li0my1#dt?Hm%ac6i=tDcjb08|OU6*vAwph_e;gEh!t?a2VNK_<!D
zdF6I~;KZVo#4D39@2z65M^^I7dxv(ESsmq#EablLk1<mhrv{CM^Eg`7Nsq)=*hxEd
z(rzSschba?Ko_m*BK5myr!Lw9>vz#XUF5cs6rf50Bk>)k0EZOdMzVJbxg?MZRH>wX
zDsV~#9#}sW1f`POM6y_eELMe|P!?;Q#X4ZlW(~4Qc{V#Ln{~qSB2Ha|$2FpKl2e}~
z`3~oLhx5jR>xZ1{Ly{xq>PS4!fr+Vd^#+pn%U$=&iQ^B+T@R6*#+#(^mbkH|@wRC=
z1`9WY@g8Ypc2F8WAq_{iMJb&(Nhgm_=SQXUPPo2w-Y=aWNTKM}pOVfG#Eg1F{3iR!
z;}7sA2S_gBO^Qf9#G4!<xr}!%<B6r?OT6<Xl6!bX4?o~ZMd=@2@ej%2f+AcX<Pm}*
zg5+pH87&aA?q3DPucSO)u!zT94-VM{i>)MY6D+oooGMtPlAI=3q>-F1Sfpdl*=h2A
zP@2wnum?To)zA5N(1S&D5<%$SNs6sT6&P9thU1MycNqOUj2SlIMeNWbb~s+Dj^c(!
zaRZYAQM$*Q+~cj#iA1T1H)$fdoi}MGIYzLH5nQo;nP5?dS3}Wx;Bc>RKP$e_2G*>B
zorU6yxi@#6O?D|>1jj15yqHT)TMBrw05j_Au*SY!u(-yn@{w2ZNE6*M47&Dhf<=mG
z6D+KuZwD+Az?BA8^_U%4$}P;gAp7gH&~SK7w|JXdWIuSok9@%UqPs3^`k~Murftug
zgy}a4({DC>{EIex9G-utB0#v<Ft~SLq4r(M08i~>Y3*2=;9^?4nB;0&yPD)iTDy_t
zIG`N|h~tj{?IR>#1KQU}&SSLm7(zde)sABc-pOk3Bsq`O&LjCKt9_K@CKfcYgVFsQ
z)0hvrAJws*5v9ps@?abz+baWmWrK0d>kcEogOj-6Qnw)H2>Qs-!SrfDS%%{g$Cv0Y
zduZ(A2rc9tUU7$vPe0-nk1!uUU-M3m*Rwz0SjaaP@{NW37g<Qu^Wi9}?^_^fuvFat
z5fC_W?fU`<62NI7@J<5$c+zxI?sHNeh$b;8It7PceIEqD;jDVrxSn-EIV5L%KRerV
z?3>=S$^Slw1f?Fr8vDK<f-mJSxzk@bFS?{-{>;2(#mR6yeTEuwhVns*rMA@{H<c9r
z3Qx6-S8l`m=iu@VuS~>@t_?npzR!`Mi$C|r@8#u<?Zy9^aOK7Oaku<m1bf3<&gYc*
zIB&jplcF*V=fPimalHD4PxI*BY5Z$Rspw-I@LBYIa0NZhD;n#crcW6b|HjzAG4^ka
z{lCcASI++2@Wr<YD!9%y9*iFtJi@Ss3~ZGV{6q$xU`8vL(Ej#)T?rWin8m=FJ!C!5
zLfG%M-^{i%EBghGLhC4b9i@*Vg1y->7FWO+lpO<;#94U*tB;}sFpkX1@3Z=7^{BUE
z`9oG8Rb1oPTz7Ehyx(B5;3c4QiHv{bvpV@?yd#V=4<q;6J>eBk@P|pR{uwnZvuczK
zK7sSB>OAX;*H$W6RR!ycyq73d%B?Ep1EHsD?YpLbwk>AuzsCNfPb<Nt#J&$*;g_}s
z_^%<C688ZAJ>*hiD+p}`bMTr(4+!lc`5Y5^j+ukvFYq%OOlU31DO_j@HwTR@L}?w2
z*{>sF_P1<)oly}oYT<wJIOrQ(aLm#7<t=>8FKF)<v_A?C!4U)J)4~mmIneqQJh1Px
zT=>BAtlN3^E&Q4@IO`0Yx!0h+;{SY~q3n$-d*jO9xUzqdE4#b$=)56@uG8W0HeOJ~
z3&ggDR6&u7`Cia0x@b&@>i;>*Kq2+sdCPyk+lOt)Bi8Z}>w@CS=>Lb~;=Zr<p={r#
z2`{E>Te!UgdZ`{ZsE3`1P2egnXcb3rIOiSC4dm~`y?LB(G&k^FleS$)PC4#3AMwU2
zy|GGftkS>8Dz!B>#Qa&>koEt4ALld0U(IWxr?~u&C;PRZ(Y9V{+`8$%jNLkZE?cSJ
zbz0|t{0cfEgAb3mnE&^i?KA(&w$D88%G+nqxOhGEw0XQq9=V0KkT)sB{Icd3m*3kC
zoP_)0S~;s$&I^SOMd>ML{gl&SZsV-lNckzb)hRile_C#RT2APnkz1X?@~*DTQ;s*|
z+u>}lkW&|O({RNRa&?59;7Z=Pk{@u;J;_axT0;JO*vhNqVwHSg{TzM3mKFNH#)}36
zaBUfA9+TTNUIEQ3l1mv)DY;Ffn$c8~90O-bF~lqhe&6?t4|~^lmM{Bw`NrNK?!Eu?
z{pXE-XlZ%ipKkc4@4!z=x<BqSv#MDJ?+)^Fb60!&czAmGYP38D6tukXPYwR*gMa$r
JpKAQ`{{Z?~IFA4T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle.hpp.7B2A45106BD10E35.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle.hpp.7B2A45106BD10E35.idx
new file mode 100755
index 0000000000000000000000000000000000000000..66230c649cc7a789d0cf72214500d9a9301d087b
GIT binary patch
literal 15078
zcmdU$d0bD~|M)xS+;dBAdV6cprj$1+rPOWHB5P%dLiT+(4Kijh_9Y=ng_>3jk)<qU
z7nv+Y5>Xf<Dr*sC&CKsqPUp<`kI&~Z|NK6W@i6z^r`PL!@44rm^V-gt0N=iSAG)cj
z^_t?lFnH<@69qLjwbtan`9Flr?1R<R-s!2SElju`6g=mikK+B)KQ)&>Tlca>uHuT(
zi(!uky7o33vOc8Ku;N*hwM%l|k2?L*@z{&yuS;gX%DmA2>egrH)RLaYC6*?ZP7klC
z`MP;{&>;(rnQGJK_^8h2rY38>`?~oFe71c>R;B#}Z=L=nk2<J+a9kC<eTl#347;D6
z<mRkXE*tfz*N~nCnp1|tg3K8frck58+5h|kA;E9++Ro2=-XCDIulV*6qlg!a4pzG)
z^gRF3?P>@0eiMyjclA&HJ#e}DfhC(yk9&6BVoPP0R<`XmIyxKcnFar3`#d0cUvYu&
z(&W<pCqH-~H&?r}AZeAROXqDiDAm=jGn(#N7=NsKuj%ym1sh9LU;1>e&5lSv_3KE@
zg!s|eC^)RXm+6snMT1f_pC8=1XL0^aOTF`R#<sd`@bPfu`E;`}enF8BfA!t-u{59{
zak}dVmr>j2UN?D^_xjg@IR(M5#w&VS=wCjWdlrrE=2;7K-S<x~2!0V5(LKm^!}4(#
zO*KcB>5o45?4d(tyhmm)o1EziGxhL!ms{>z^AP^@r|WQ=lJAF`k9D0?Yw4P3H6n1$
zpMRRy9@cQEA8L@fIl1D@ur)uPE|}A=^}Sr1lE>RN7JBZRUTxL$%dybxT^gUZ-5A>X
zb64w}MXI7b8V*?hc<`?1dmcII*Y?(Co$QlT96egK#k%T9z`dE8QDbIa>GH*@g>UMK
z8z;u;r7bDh8omGH@zX)!edgcCo!+mE|8(ny|E0@SWw-5)KaHAp-_X}>_yCLRqfB}m
z4!ve}-Q}L{g1Z@o4!^Iy8y+9}=@;)?cFD%}I6-@RPK^HO(Npd_mO2H8hGl);rZO_F
z*xg?|cvar^{gUDp3zOmx`DM17>QVH5{QlnlbFG|rMV$OuaJ;sC{qhBMW8T+!E?xP^
z>&C&#!O6-+2iGRe+xunB*@%;)i;nhMniU%qbYtonwIrQ4!S@C|x9y+1Tp4WhxMjtu
ziFY&>Ts%`_{^qQrdPU2jW?x5kUEC+$%V1ch)0@2SbIM+CyLmV1u=}MIpMQKc8T`3t
zlHb|6I|rFK*o(F$)<cbN)Rf<AyP<!ztMj#%HFLxBZ~ky^>Ga2U?0+ehe;%LlveS6E
z|6lnBCK))jDS9>gqH&f}c0@!(<b_H5EG?$?yLG-k|JOBrme+d5F8}4j)$~8rk8Lje
zzI&v*_S_xQj*K{UZQV~E?H{W}9dJDA+e<xh%z<;A$J+G?`QCHvWFLzyZx;z!>wd0J
zvhw)l*HhPeDV#o-tF%6CnHd~u-0{elVOc+(@3T8PJ~Bgn@$wmieva=lFX@H${Ouug
zZ)fD}`$4;Y@y+r5?V_wy4qZzhIKJPlU)G`H<_Q1qcC6ZN={wor$sYUf4W3M%{J`mw
z$>b%I!u>bZ#M$q*ex~})V^}xes;?WCZYy1vk<vEn#kS3@5{Hhp*q)>8T>7@=*`13<
z{uXP)W(E9wW6p7-xS8ko9B-5VY#7KoGk0m1f;W0SFPvEUcG`neNrOl2eei17@%Ux0
z(v6h*3pK{(9~q;!qC)e<;4+U9t-pu~g^{0Um$k_6+hW41hhKUOKi}((P1j>zOF;g(
zk=+Kabr~5|@@{x!Z;QCF-M{26>@ns|&DW<_HtoFF`(SLz=uo#Sl_ix6ZjS1EctP^e
zprsv$srq-B9WYRD{l#@#XRh9&nm(_0x1={7Mn^tR3SZZ1&Q~v`YxgOJt=uO#T6-)V
z67|_R%y974v;CAqF1vnrbmiy$M?ZC(dATB{SKo0vXYCu+=4zOYy=IFUKfH;|A7E2_
zR<qmL4ugE{{&3l7`Ex%D_;SKuKlZV{9OP{M#czJeqO;2aHSP8E#=|xN=X=KZpL+gm
ze5?7hRciAW1y6cJOe?f$sjS&@_BSOowNjI>0O1O-{B}~)AXb<BX&GA9`B&c`SLs4b
zRK^&NF=;$du~zTbGu{4rzg`#W6Y(&^VaAOIDmE&OSr;~9?P?{|BI01%;L(i-Dt5Lz
zlJGLRW{NHph<KdgI920;irv0zIb+mptuZa3M8tzN2OBpYsMvF*@3kR8uQZgf6A>Rl
zMn_P)ZzrnaW$RfrvBF@i5}Fg0L?k34<8LRb(&PNRZchh9XDFdDQMm(!JJ9IciK=)n
z()I6hdi$xCP?M+>LCqql|LsIog1);Qz2;`S7+t7GRE8Q3HEujm@#t%Jc1(<2?4}D1
zh&aeF$hh%9#gpdmKe}dB<o1@Z4G~Y)oNU;5pyFvQMtcs(OuwWHjfi-p;Yj1g0~OCM
z?|eM#^7uGi*pY|_nGEXGc%b4ryDmjP6xKh|g(4A;Hyp2OJW%nx1m%ahQ(`A8p#>2~
z0K*7i{q00mLLOF4{wpfAjS?z|N*c0FLvG(rRAt$;Pl|vxtyZ^$h^WlcoMqT}pyHKP
zpM0+O&^)1pt%!Jm?Sj#b2PzI5>8tMgL)~2^)F$F(!m=KX2P%$z?lhqE>6{O`uq_b}
zHym!<c%b6MscS-)1lC<u!VW}S3WQSN`t8)LGV5w^cF-zQC2UVrs)0}q+`gTt%Fgf3
z`9;;}hA3eRq7n*iLt*!CC#tg7X8i9NnHHOLp*m3+YdF@V@z5-uHO##IhIau<XwvWy
zC{&Y?)O4aM2mEH`CwuC;HVxUwqS0fqo!oSyDo399o#=LHmxB_Tk-K~b!e?Oe?L<}1
zjBnvzekt&%2$h}H<N@jfIyWBFT0qcJO%2E(+qBZ4#VNmWN~mgh7bJutqi-jAk6ZH+
zs|;p{M!Ha!s011Y8aE!OxcF}3GWSEb@{~}6i1VRsKJ@u^qAC@}s|SrsFN;ya)<h)^
z*vEkZ-%eDey8Y6$Nsrz`U1&&DMi`DTZah$NO~jkg8?4@5XbCNec%tS+gT@0Df7p8Y
z!l(^v*XTkWA`UPNFm60h@yCqz$J?$wDJx++B0dg;<HUG1ov2F4)M@kYc4c-BYJRhB
zv<D@5qq_A*b$|L@#M82wd!IvR1xY>!w9f%6`VdLJ0km%bD{3(%xeW9w1LNq}ljJS9
z=N3Gc+9YZ=TWU6%OZL2ZuJhcfRg~lquaKS`^$zLBC<WN307u%(l6(f3oB?L6LD^tz
zHkiQpDYAcx9BBh2xeg7fLqi#73VNA>0pr`kpxeR_%35||zs%Zdf8Rus155){wC^Q3
z&@@m*2a6;RH65y=woa0VnGRD?4<O0IO^2)KRF&i)(;yW!dy+iDbcBkI21y=iI#NZA
znk0|09AiuSOOnT0j<ux|O_Il1j<cm+LXyW@j<==eOp@nym}kYhD?~rUgwdi7i>#>8
zk>thti%l4X0;5pSp4uEq4g*GEJZ}TiHlU<-QIekk?<b%i`_kpWyPW4*Ak_k8;x`i`
z$*D+6MM~;=B)I@d1xQIdT9UWnw%c$!#)-IXBF~jrrxF`7c{SFl=J^BG`M{X=KhkXU
zylkYc!GUY=Fm{{!xW|3mld@Lk-fT-e_pCyagN0zJ$=R!^N%8_=0Y^)Pr5wcpAr5ey
z1%xb~F9YE+&m}-8;kgP3RXl$J!Y7_Xp%4nW`{zR;pXXvI6!TmKg({xcB4I7!exHDZ
z1fJ87b{f*B@2cJQc2B!o4`(AmP6z(!U@)C)f_$^#uM_PnK`sz&3&gIJ+ul8_QJ;Fe
zTT4L>@d)vxqfU?)do1>(<E-JPj_E*VH_ZSp8NikCVbJj~=*0K}bhrQ=sUH&L9O#h4
z^F`=zk>@hxS%$o5qXfAed6x71K(K!xbfa<uZTG7^_MFcJl3a{>7o$Kbpv@-(s3AaO
z11JI5C4jNCeUkhF*}Xu`7Ovghk!N3LAGwMoIR)6H06V7m6xlpQcJ%&|oGB<X1zpOf
z1re5SJgN>rNv^@RHQ191B)J|N)Z@;S&ApDDeCtxE`LF6`n{FF3S1L8kNH<S3o@h;7
zoXwQ_?j~?)P76Vfh1#*OHT%r5ko=P~&tqcYIL0Z^ECre~-T}>a@SFyX(x3^Gr$PHP
z=)kj28uVkFE~?VSPPCKlE0YcxO~{?2ArJ>+bOsnxha$+=fzfqfOg)|;KLqU_0#j-S
z1^Exq?hl?bphX7kz#ep47;sw{NFOA~#lnDMo*xJ!9tb0;yqiz?2GL-o6Pb-^P&*AO
z*p1SlDh+nz**y(<P^RvIjNgU~f{ovK`t!``;BIh-u63w4eVR+p?LCivZm<16e{ENX
z(13>*_t%lxa2`0F2SaH8OY%+NbQ27r{ZEr2pxKxN(bw!gCv5ue=V9~IB{>H<=CC)E
z<Xq&K%X1#G&O_a44@hz$vMog1?7J>>y)HD{6CxM77V^9aci)8j)8D(d^7w6h-wun&
zKwkr-HJ~eXs8oRjTTM9`ZK@>4t82uox1_-q^^iodDKle=^<sy5(VZ&xSk*>+_I_t)
z(nsk)HyyNPrskMnc}%dTkEFRAxnolz$L{!1RDEQ3q!}F<Gfn9nQ#^)DkD)H~j*?t1
znwE=QXj8m5<hxb>F+Z`fyUlWt15?aJX1P4)A^kjL#pH#^un^fWzAhMD7i<|93Py!I
zZ^E56VK*x89bok9lJ^?DwHnN~Low<_1tfkE6k&oH<Bfu1BhRUVB9-SfL6OGuZNdJw
z;LUztEZ7(Gd<^Rx!yTCX9M(C<^Bt^nhv&E0=q+ykCS>Me(>mPjJS90sG>#FQd$`}N
zgzoX9`poZ20y=4)9-FWsQV2(jq(yqvwUU5NTFlW}pjZnG>G+f6Y+#)YWHynq!N6<~
z#P|ZRzQD_KfOQVfg}}NH$aLb7i2w!`f*{64z`BT+-vZXRczyz;CqR!H%7y^a_6cxd
zTn=o@d9DCb1+QNTq)MLufR2AaPdd0I`4Mz{#B(`xEa$lbI#%%f0Xlwwo_oKgr;_{;
zI)3E29y-?Z{1rNW<v9X5Mj%g`R7&z@<hYsVXyh2pa}08f;kguPmm)6nh{lR&7QjmK
zR&2c$%WQxpVCw{)Q?Me1*WZB^J9s{Xbq?_X^danchz~2$u}(TK&%invJfFZiC;rZk
zC-~o=#5yN=`6;Y(is!dj@s@8-J+`XHTqYJHwvG{v===Hho!5S3EIdL2r|W|Eb)nfO
zHCSBlLY_C_ft&DP`a{3C?9BzMUuhAG`xPs`;^v5IpmWq5J)7`IQIg-Gq3_Um`U^=8
z7aYTR8y+PLjS@Kde!+1+FTaF?F5yY+_YbklL)>g@2KqbeU+EJ&gfzVt^{+*Pm}0ch
zKU(0p2oEa4&Do`Yf1O_^=)&0l8uJYgKI}Z8)y^RDjyYJ7gPX%wnskbqjfFL4ukWFN
zdwiPvKnQ#waB0tK9JrcKgYMxT_gESv$-m<szcU`u=Cb{SLv2+Wk~~*?t~Ikp3$+(o
zGyC%g)c*rIuo+nn^~-tw0QEmW2R7sCp?*Ej5lBA*Ij|WOjr5}#kN6U6UwL}9<3CPJ
zXueJSUqg+Nby6q}3}u}##y!5xPW<fq|9S8;?#S#ngYqwx{o@>B!sD>%;vZ%Y$uRM=
z^mC>!O6#b^tv3`==|4;I5hxvj#x!Ci#T=-81Detz7cE7QZfhv6vY&<Fz%bTrGc9s-
z%(Pmil0lpemD#WzyI(f6%x0<CY*;bC^6uehq;=DTX~s0)q;=ZXnvL~Y_Lyy;%{E}d
z3J$fPO)W5C1%+g&mkiC=(-M($BJ!r;lO!i0=Omt!k#jQ7si;jVGGY1!s7(PfVRN?<
z8&zTz<7#YF&GQFr^nvFH(K<rp?DaL#{2FuUlKfh9eJ%E)A(JGpP}f_bZpJuTT`!t(
zaK^T)(;xX6klv~X+V#MeRzPW`hkVV3sx|AcMf!`(*<6efTw|Egoa^#>Uf``C^vRHl
z0`^g$N95P>`HEiGXG!ud)O8o~-2Z27iQ%K~?@01Wb?cStZq-toPZ8D&gJ}1YLC{nL
zXU{8w7DdpJ#*DNYP6F|!QaDxov48Jt)+3w-lNTB6G1#95><cfJl`j@<UMDrhL}ZzW
zY}kiLLY7H9CnL*bo?nYruf?vk<CauC8gRSFM~~b;OBkFbG^gUU_)ctNL*<>ery*MQ
zNCyL1wEPL#eL_9hht3iVvV`V*Tar%;(rLk%$qR(G1%f4I+U4YXG*s}}V-kT~B3~$|
zLUvWGP_Vk|>}96UvTR7R!$n27*c_>@*|mJ}z`ePb$$a`98vG8;nLT|B+dYPzD6gBY
z?cw#+Par|}W2|_L+t3z~ngmw7;Q23X@E7LZAX;o4E$T4+i=xg&QJ?W|qRwwT-xAy2
zVy>Eol;n*Y7A4plrvWJq_^^)Mh!q?00D9YvD(}IY?@fvLuldSGKeVQ5^ItAqBzchT
zAY*DE$czEX6~KT#o(xr>T*-4B=p4saT=JoBJ{(N#83`<4tLspg@f~P<2U^hnlt^}D
zbqCt>>~aTsF)oGPrO=0Gzfw4a=fF}pig7hGu7(!O-&I4aYG}{1OEvUjTm!vppbyV}
zHE;;efi-Xx<4|NAiY%DF3`JI<$ew4HP~^pUJ@Q_Ue0cU-kB0CZxE_sSoQP}_k<3p_
z+a#Wok!>>1k5H>es5$;35k6}5gy&;|@)%!hzb~}9&njRfB@*=N1kM7#7xdoqybmk)
zVJ>1$!^$+CSBk2Yyv02z>Kzoh`@IlVFIdtl$uC9KOP*hes#iQ$iK;4|t3_2c&#y(h
z*P=5uM3VeQRJ~#HjW2rNnK3P-{vVfaHa$u3w^=@NDlrXyu72I={EFD9?3T1^PL?3Z
zSF!t57K#gU9(K>;`5Jb=#&ZF7FJK%MeD<%YH^+JqXHW{QN}&@Qu{F@D20F13wjNon
zM^1D^Y@3$z)YHHmkr{kcR2&t#YW-PJah5S%?;sYUVO^5Gfh4cRfooalAjuDKj|VKw
z{AII$-=JQ(gUE}Y6xy8>f~e_8NVI;uBxiy1KMwt9b}49nTl!>8c?<O4f&na?*s8AB
zs@{iv%s6!^PTiX_)gykeX`PhaC_=qMg!(9|NaGRW2b)$)sp8)D1@G{{Pc6vHMTj~P
zVsj0IF6)p*v4+J{sxQflTo<`h??RG$uf;t%>b<xhqZHtj!fNQm&;gU{z>KBlg<xzU
zn85fca(aqfY44KSgD@mU_}ikGQKrCs-C|)-F<TU)!=s@u8kciflRt^7PpqD_FQ~{<
z^Z0BVl7VFdr))5Q6&nhHQz00@TJ!|yKLIxMbCLQI&@bnCE7sYHo4ZVs6R=JK<Nde9
z($K;d^U1)R(t3(54S31;7Nk<<%IP;E4R2aqW)1&)4er3g(|B<3&c0TXd;vIJ0G_Pj
z7eJp2U?Afh;FQD5bAV3{7|i5Fz^Mp$vKK7^eTu+9#<zgeEna>L_}l`6nY;q%SFm!j
zBv*ir6~LBtR3*@_<mHv1b0x56@;YQ!hg{gF{)+6r@*E@B#R!}{L9k2UIRzV~U<>y9
z6l|5k@=vnLfn8HDS0+orJ}E3Ym*kyTXD1(e?8N3f*@tV8d+fx0*!_3n{yTB=DhvtM
zuxUDOuFFeu1~$#$`6M<y$@3{}dWz@E*x@qf%DXqP!wsHSii(wDTlW0bqGC1AVWJ|8
z=k=muJ<pp&olPPaT1ASANM0T%n#ZxZF3G=$=D+Y9FPg{myhk+O!*i-=p33uH(R?rC
zgYib=`tH*Hi5Qj7*#0vPpf7UhLBhop$Mdtvj5vrD2QjzqbQ~*=Gp1<;u~$vID(IZZ
zOuY)8&ls_W+@TCxmhs>CIqvivHwVm_FBfO(Z+M&buUe2gG40W{-kN!3O0v+a72Le|
zNfvsw7I3swYbi%@Kr0S#oCUP9c)kp@F7sRhv`Toc0$No(e*#*ccn*bHp^&?OKGe$R
zxfp5{^IQeBs(4<DwAQkPTw)}VRszpyNJvAiXyQ%se<U2?`LQTG7Mta#-P^Y=4|~4}
zOL8*QNQO%4+9dg~C><8{=vb5FYog9I=KUK|Ta6X!N@{K;Ia*yKnlT+=WXLt_(#c4&
z&YpT|OV*h=dLd5@y}FVC_XI1R@J8b$R=i|<=1TF%gWca%{Hq4sImdu0xX;d)B<T7b
zy8I4%u%N3BI@R%o<#6N_&I-je_#z=p)0QRLeUiKz>F-9iEIwbYZoXRGm34clx++w?
zdA%sxZ@qcd^|?7D<F12#b<m&sabi1Qz-Ji5I2`$fBY*m2vWo`=Y(YVk&nKsgGghhZ
zCPpVqkg|kk>zUJI#>GERFUTNY@IS9%=UUyK{PMTVaH1H4bz<1!HdzeAI`NDz7v0Ry
ze^Gn>9|wMU<xvlNef0HcXbO^Q5Vz10DM*n*Gv|FgbNi#$t1d+GKd++QaPPPzz~6Y=
zziQOoyt5$4+U(MD5`j%NnQYAhqbOh=1?>1Lkl8L|%c?;2qEs&$vZ1t6-FT(C6|D-9
z5LR7^P;ZvgolxyG8@J1`5gOtf(w|j}7yB;m&w|#qz-TQnVcGmTV6=|sQ^5WdaA1*C
zKCsUR4lFWy0PG(C2NvnXz-}?D+DO(<V7IM2$HIVEIFuQ&9ngCR?8W#I^mzpR*tD#G
zUKP-r@ki+O5qh&p@)i1gg?@}TBcIL4k4_RYvXNH|@@8C)+{%$VGn%V#;3_^BO2fWs
zY#~|tKS`eSf08_6bpMv*kquRUOY$VS1otRmxfO|pagS1->+tY8Jc6|+MhuJ*xg~-E
z(XK#jUTG-q`s)7CVH2W>H;4q<k$hqH4bXnW7iM=L?Hzn!^$Kc#g`a@$u;LwCD5B*U
z;tiTMRMOFNw<2u#o?D#(v6wr7+fKF-i5NcMwwvcX!7Wc{UJ)v(O$e9~yJ6iwHX08t
z!vn7k?L70p+JZ8d8@)apXbb*Tw&3xyQxjUZl>^Az#fgeIu|50rdql+^#!vbVM_G15
zO&e_If82^hD}ChCHEndIpYEl`fz8k@Ba(c&$I#<3Uo-m*yMKm#*{s=ux^F>!=`4Ba
z^U*(cX54$?qjFH+9MoJqY}mKaH<#x;<dMg(_ZOlbg{Uumr>ZYwJ`Rizb#8Qa26fPk
zJ@GSa_Zg}fha-b<WJX)i;OyFMK`P30dnU;l8+Pi_H<aXcIB*@CzH~<u@zM=Dc9~iY
zva4a+&&cEv>>~N4#V=U#1)I`lzfCxq-8G|qhkw<;di$bmLq%>`>i_rMAMX^;U#(v6
z->@S1&o)QZ=4f@CXZSkuzfSb`OD`?c`*K11A3x!JsPBmmc}MF0=ZlGS%d(zk<2G?N
zFi!&ele~M)1NM2md%X|r@AK|83U-U)-Dx!J7R_@k>=FxYnLAB^-YL9$EkyyP{EBoL
zax3HA;RC_u0Sn7WaUTb+Wb>;bP_&A{&HG~-0!6EMo>Q?`DxajL;$EqIVIdXwOU1r)
zqLKZ`xPK~c-iIg2`>^jm-u>2LuR8vIF`|2n$n6WeEw;TaHt$FN@Fo3#{)^yEWJDdp
zriU=M1L_1eJ;9j9GsK=W?aF6gWWBoedba8Q%a=XbT_0?#B{sAc+0`Ph>Cu8+w7_u@
zb|_-&ENUTBznj>Q<joJ6JU_MJqD#Y}>5nzNP*WrS2LSzJM=#XWY@08kyz4GL9XK+#
sNXx~#i;bg`ZC5*Ihi>+2F3l=x-PolwyR>JQ4(!s2T{^N$8+Q5s04i=eyZ`_I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.8A2E3D6921803C3C.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp.8A2E3D6921803C3C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f7d353415ad2a2412d871ac605af58db6c87d073
GIT binary patch
literal 19686
zcmds9cT`o!*5}MAgV)35(t8m_5CTdS0j1agMUi9gCAO$BYMwF8L_tMCK}1AEMNljt
z0*X>Z1OW>c5G*JbiXEdq#ESY2xp$b!KQHU8^{w^Ripyl*UuK_~IWv29os-bOAw%vC
zS5q7Kb>NCcv%jCgsi~=Hp?{WrzjVGDR8xD=O-*e@(xvf%i*I}AKFlv(z0N$bG-8{_
zj;dbw2cI+CbSJrsC#%c*4IBQ|coW-R!4qE>U)f?Cv^TzN#Iq5-bK2YooN7B-U3lZ(
z$06;8&bxmSqaB~jiYfE)eAo8YFVNh+IW^XMqwa^Sy*p&@$2$$!=Tx-nMAVX$gd2wY
zH&pa^S=wVx>`3QTE`v|!xZY`vb<3Y>bK-<!&AU&#o)%pWEuH?<)LOfH&2015rXwTD
z282h|-0bp8w|(54ddGd?^PSypKiIT!d-3$Cr!Ayuva}~Q=I*`grw_g(9Iska<?JK#
z=+fg(pM}d`jKA2&G2zjabr+W&c=k)>it(}G|3298&i!Xgr>Q-sUPzp_WO?Px1;3ir
z6nb0DT~l>aBjZwTtHazRgQ#-7CL>8#I^vq{@`0a~=_2s6TwnZ_@~zb-L({Eu7rfrH
z{B`kh`*UY+mGJNP#tFvXectmnzne7}6F>2%?dg4!K2$q2^<8lJK>sCMZ0EGczW?Es
zr&a2Y(O=!KGq~_*(S;lJtNrJUdcEp)RHUPQ)0i!v$}&9eOJ{XrKe-1+7Q}^p^q;D(
zX;->!`PT!=k{r27hZ`Qf=)3l#Lqo4WesEnVeOoYXdwk-;2P<tB#Rk9iI;>v#(Sa|_
znAFcauI0?o5xS+1+!|)5zs%a-mius?^*51Gfg3(sR<;kc|1x9y*sN_9ZMC1$Kd?>C
z>ZJ=7TxuzsZECRC=Wh33A5Z)4><F8XTP@elAGy+!dVG8T(Py8Ix&Lm-u18JB_xzM|
zX?XuTbN239)OXaiKZY&#p1pa@laPOTMRheQx^T_SY2#kS52ddB?R{hK*gUNK>cIMd
zgKTTXs)=cPCOfBFe)sC0>%@?=yCX;Z-u`99m!6Xx#@-*gWc#G{c*W(E1h27C1O3mN
zd03BIA7ggnT&8ZFlzOiJcP#-kb{frzWYPk54m+C|8_{dVYH>mEA93~hL*9=GKDE2@
zL|m8S3&++N1?+z67B)CJVrS@+w4vTB-aLrCFeQDP*3@3#*oU+>)NS(`e&G4k2_DfW
z`aC=AoC(M2+-q+isNMXn?%C&)*Cae&Iet{o#FClayez*Gz^mkN&8k8>%jfs)^vBq0
zueIxObV#oaQ{1lkru&W9IxZ!6%|t#iajwRoqfctnej45X*R}UTtIgWA6<wB30~TKI
z0`HH{>3VK_?1msIsy49IGthrv`A-{T&-mRrHYCLNcdLEc)3}cluYK>kw?0izmV4o`
zWwPDa(_Xpp%hi@2{$^3%d2rcb=S6)FsO81J>hXPgq}yuuq90buGJDzlp<e5m$EW=3
z<yIZ%8?WkYDpHS~kNd&RX6uZ$soRt64*gtW`qkL?U)=I{`7b#1>BH)4-)29XBOUpu
zQ}7}7+5I`*Z$4jg56B2RQ8L}E{`XU!@8;KA2m3_l$!vC&PYhk^6*p!@mcuDK(>nbL
zfoqCp+5K3xt9rE7?{mMaWe&|Ws#Scw=D6iD?)xWt7xLO216|e_mmaQ)Pxc%;+9mU$
zMPkT}pvlHZCmJN5Pn=*f{??}FWzh#N*0gS}2|w@r+ojuT+r12(-MkDQ&YW?wv`<~P
zl{ZhPUdi%%G)?!yyVe~4?QcT&JTRD7cB*~(iWTbGzYkgy(d)5q;{D>}lVL9}Jyt7T
zQ?O=n6Pvulccck-*Jk>GZN`-sO-~vR*4k^;7SwIl)7#0%);?I<<aTD1=P&hQdrDXR
zuDZ#!DY4V6jpwKN{WN*&*wBe<?S}Oa7+E{z#IXGR7rM4R_$j#GQC^#G$(z}J=J`9P
zIRSTOM3!ybshIA*>ems?4cyC;+J>z0z1eRT$evsnbksbtt$N|_bB4}yA9Qq<tC!zV
z$@NP>vGy^=$buQE-(2zc+y8pWno}2_NXv(n>^U^CZtRATYeVXM55?SdFq->l)wL9X
zpW^)Wvb}LTZN?nczd3Er=J4j=#X0xnk2jogdOWK!bw{pFm$+WT4@_E<Tw!!%OaGj1
z^Y+bmlmE1{W$wgldjeu+*LVMwXBN#etebUX_0QRfN$Jylr#av6du!2iMcE-?`7Imn
zj{kR2-<~LxdtaY2v^H(?vU?g~N1}9|N^hN(reyMkJt7{<{>Yh8Klj@kuN?YjPq=aG
z`&Nsn<x_RD9u`hDhOq;i#4fPD`+|q1`RktbslReiFZ88X^sUMV8}8k|Q`qOpoGv<7
zj@><Hrfd3R)*Z{JH@_`Ra&w9s8x!7RTjb=kGk;v!AbQy^Y4qRtMeB3D=8s>leAoQ_
z?r$T0n_Do!ZP2Y^m*R%e(WQ$wK<&`mcNf2Uq%&z{Yz8aa{Ohiy5z^)b+ORCP^6M8v
z?>_i>@az0-i}h#jcNm#rJI%qdCGuYF3XQ32-uvkESbcJQ@h^=}A4QI+^l`Q@zWv-F
zKgKtu!60zkPs@#)r!CYAKY4oHKEK)b)a!P}ZfxGoap!%&)y3L=mhT7dPX2g0yZX22
zpXYzS?8ikjhj&p^<5qu;8gckjnl`ZWSCgUu7X?gIL`@ypOZ2BpcH^45Wx`!;V2@bN
z1MWPqRT0c0rUVvT=2t)O3Um-l8IY9$Llwa+rs;1kths2^+!gSM<q#t~#2BjxX0gy3
z%wD>^eVH8C3{;bvprQ%tsR(B2nVqw;ba-Qk985qg70{>xTB!(Tv6)|$*L%|3aBX0W
z20F!Xrx;%q!7P1TqCHKXlsRaFuMkTe;OfvDDG6qA{<$itE~ksP9N40P8la*9%|%Ht
zi|gbgM$z+I=W2uAsF!@e<pWa{!7Oew>KnJ0`;6`i%n?f}BTGdutt6c+rK5CZi(?wQ
z0)50%4rJxPKt(W%XZE&{ug;mtwShZgX=P1XS#uS^EZ&K!Wd|nxQqdK3MJ&;*ESfb_
z5zI0mU~Wmz1UCz9V2W7k7^gaBxQbvFzmuh3S9~2cTpJi6mR!a$mzks@m}Q`B;jz!p
zPIi}r$!J~9L!I-`R7EgLz^nY~#wEl2^?)_%WjkZMoiSAr%rb1eVUoXkufF;~hI&~G
z)z?CWieQ$JKNuE#GT(Aj05a4|IP4k@byNhijP3UG$MUQ>wE{5nLodz1Xr`}<sC8i(
z12r{9VkU$X?%cnu6>a2DG}LwIdL7PG5!}!8d)awLy#)(B&;t#Xz!)c>2`dR^nKiRD
zaF2QNVmX+EdN~L6&p}HS!7Q^IS2x}Wxw1hH`XQF9(DW*FRuRlH=lc2@!>$RF^?)TB
zD2_3XLjx%ZW|_w(1=bl#lez){v20{z8_`B831(T+RlTPEJIg9jZ9IByJ9Rs+KmVoH
zg#&6+e9$W|skSrxHu}dQaxf9i<SaBgiw=pB;69d?Up(<awtZ<=0MRg|Kvs&RPDwD!
z4{^&~QX<|2$-yAh%VlVA8QQ7{W?515*5J!D$6<P)3+g3_F-~I4R0Om9Y<XaT<DEJ;
zJ)n+Qb}`1g7;6>5EGxfizm!=YoFWGvXrOv%UJvb61hf3sWcOp`^-^^`pn-bX%^2@S
zy(kH0SzTB)D&WiU(R!dKVo79-6Vci#31$gTI39as+u*9MKo7CRv9dVUO+`9c&ixo-
zZ(;IQ8}vaeml@;BjJ=9rmI(DHZ=VMqE7b$8h-EWlwwbX}5zMl|%6aU=?9qd}0$s%N
zl$AYY4O9fPM49`gw}g)^*8|-V%QnV%8=9_?V3sXu)=nlm<@dXSzKCTHBiq9msR(A-
zKCI1p<HFqe+MpL=X=AuH#zaLhOOjRJ>M0eOujIfNt?pfDc^CFp5zMmlTW<@kZSUvn
z0V~u?JYyV>=AtB+CHeOY^IXSl8z~0^P%m}Rpbqv@5zLYbv%+89S>mAw<fxaOjPXt+
z3`&Ao4k_9uZ<?c3+!bgdmIPLoz#6IuX35-O@A#ox^($?lg;)|<$3)gkMKDXY-S2N)
zgZr<O18+2!8faRB4w{m5vShT>>6o3JsQ_+><s{TN3AI%Ovm6VR&1zgQv#2W&(LiT_
z><l^qN`hI6K3gA(T5@fv91KCdT!H#mU{4jnET!iLj@|vfYFt;Kje6OPj5+JBBABJj
zdHTI)$wy4&zzg+q2b$kO;;kf|EN^Q+75!{#rwtrXFY8!t9Xj+%f?3M*2I*YbZj9#Q
zf_iC$ibiCQlmxR}?pb%ict@JG9P~#lm!SS7G+ia>WT}OpW0+QC_?*!|sensG)?P_E
zS-yBY>z5C%%E4gN%QYxogI+3vS(-Pj85*l;kBqP<Vz~(QFQS2z1hYJyGb6(Hn|EL7
z0Sh$H7RGoB5-BCYEYGK`wIBS$veVkY1hKTRjxDUGieQ%4-{c&)SNnd2HZVjihZ)Di
z$Q~&PW@!r?G|F~w=of9^gjk*c?g{9pBAqN(kNQ>mkNT($91+VWz<mOqDuP))vDzhB
z25ZChKsUs)l`-CmrmG~F<;&S_?zaz)9Ip)&h^3izY-R&g1hXuiJ!jcEFXB_NGL<KV
z4V8!ltj8JPg=J5awnK~U&>fpYq0r6JdvInECrVX7vkKT@*GrV{1I_!u4m&QQ^a>1o
z1*YP+7p0TD|4Dud_7w1QXa-7ujm%&ZYq^OXgw2d7B|+;X7>H+p9XVd6a^J{<pMc*d
zFtSaxn4+|W6}Pa)_yC}eSpOt81fO3KnF{sGP#;?njXn{H-)wcdZH69XEVeM#SVBZ(
z04?GeYit`uwC)yL8Ed>HBGN<_@r*TAQX<ll7TXwWtUyI%sVx#1Yiz{PnlcvK8Efo?
zh{y<7Br?|6KA?#*wn>a5wp}99QMNl7N36a?WSwkxF^<@9iO9~_?q(dZ2@<7jrbjko
zhh<Zgau|ag#*E;6rbj+whkYqgDqsu>7&C$^7?TR78+KDg=^SHnj>gv+@j9a&uChd;
zR1CyoppCsvQECN#tzanG`q#kkHI2hqF^ttF8x_Ziahx_54N)rL4NG`)g3s}W=V%-w
z=*9@fq&-2<O`!1}L3a<q8o39GZTatt6e2pA++jvk1`TRzBJv!$TpAZKToH}y7_N@S
zuNm$&jUxdU38?;30GC4JJiz7A_#EKQ(fB6dZqoQM;2zVs9dPY5UdwW8S!%pQmP@2@
zGs`v8xP{ehVUd@q{G(}D>))s5?gGF`FPMohm`T{-=A^fR<6EIOUQxpv+3M~4PdLjt
zBs%?bm;qRHI3z0l3zz{|R607cdckxj9lc<DUNF7{zhrv8WNc_W@Fg>d;8%>%E5?}M
zTdd11))k*oPP)yy+@^6o>rzkS{oH{4oDZHjCmrGj9HQ|D&hG=~Pukb>PV0GRjEywC
z&-<pGE@wn(2RCR3C*cMh3kot-1W2^Z3nt(NGak=al=3<Ed~OJSP?YL9qk66v!3~_)
zz?l-<#yPfe128tfF`?*1an5;Gl<qMO_n1DU;|$IrgX@DCMd>!DeVfz67>A-jl+vI)
z4eDS8DN_ZcI6Seqyvg=u|K}NqafjfzL+BhDJt!EZ37u25cw%$B%kuj^Xfb-K^^{?$
zvGO^V|8I{f**{(waO92_Dh|rypaEG?Wf+fF6e~t#yOhB@5soK}_=E`{dw85z9OsAQ
zJ#%o5%w46eZi!ZS4`aHAvBwUy!`JOzX5iP-E}V3f)jY~-kppm)bvw%TrSasW>{Nma
zS<6D!ir|y1<w+WsvnJ)N8EG$Poyu7k8V8iKK?L6vtZxc7c*fYAqgPdCAIPi5u$D3G
zARN{sBL%H@5;N`ea95_u^zy~(oOBtOTn47t;B(SlU~(6j;+{F_4`BWWu)x*=sT5%T
zg2t6>_e!=0nc-@F_-Z~F&ybVC_~BtRUeAwT&riVZ&H>FE1fvP=EGLz-n&qs5bX3k-
zm$N-->|M_G#~9~VkT~px^1bu{Q)XGPIAFU6WUAZ>S)(%GQU-=$mj&lq&|)exEQIkQ
zD_&$x@zA35h}C?=T40aPO_hOBm#OkD<iRJv_X!BbizgzVyvI7$sZ8ari_&^_@On0s
z;CNPyXN_>2iChylIDwsr!ydGYtp84SB$>z+sD1_N;W)#s#l3Ff`IjL`fNeQjJ**-;
z{I~fZ{m@PGf4<rMT-Ju(J(J#s_UtNfzY0cSsS~9};NA#EVX5;pe|2GjFyj@+NvELx
zDQK9jipMyq80r_(xCGjkKs(Z23iV4#I~FrWrpkBWt?^nMF?ZjSh$XxzMR2YW+#t*;
zN|BsvB#mP@`xwrPw8wLf@x(%iQXcP>#}C8p-tzus(}wi;5$!=4BbG6}u{y#>5y_!4
zy@sbEqS#Zt2Abfo6sspRXJwKNGy1O95gxxu>4lc>4@U0~#t?Ig(k;&R7H5y=;MeoY
zfXyZA0uf^yXVS(w;094@=S<oO_S;bD`NzFwJO5k-%LvYeFh+8gku;9s3}QGt(jLzl
z$CHjaR>3BZ_r&ek13_m=nFYk}AWC<bK6jWYWa>FwpB%0;@WAOoH1%yzzKz!M%3L8n
z)`LP#_WtrleUL_nZ{V_n8;FM%rPG|^G-pY0DW@o<aTTYiqVZKuah1lad8gI9A9+5E
zcM7BNU0(Mt--EP2=5-&__;*3~cY$h85=@eW&dovbfMBtkcyppuB$yTnU*Z0S)+BlD
zm^gS@?~Xu0%x3hlch?aph&eRQXT*F)AFp*spdc2|xSUaxGsai|Md>r+`k5Jkb3UT<
zg>n5t<1pYF2B`LM;2KWjt-y6F7=SbWD6j&q@ia~Zu8B0>0bF;`I32jAg8`RSIRsJ4
z0InG{&H}DkG|mRD*)*;O_SL|RD2-|mTn)w%Tm$TD==NG*UrXZ#VBY}Tu#yp_1`yl;
z#u3~E?3?KJW?<h;<5nQH0)3**TY+OM@F4g#aC}W;r0Cjc{x?8;L*w0?=5EdiZ@eg-
z;uWWODsVc@+n?s$Nc&md{w$3vc||47f00*Qr14E&_a+^f-Q-<w@|1&F$LrS7?YDT{
zTQqLubsPVRT^s4=n|R$Oy1kj#ZKiROphzNtUWY?#mnOK9_bC#ziUbpkalnlfq%!AE
zbVd#@<?v%LW8jb_-6kXebQ+M>3C`m~a6{1c;^@m^EgGm{zo6JJbb901RYWRPSu4O=
zB)D(vVtt3G7?dzM!1^2@o|Y&bVh0{#M-Wr{k{|bypNWS;W<&6Z5IVI^@W{Rf=LRQ?
zLPNLnBii{<m~rGV-M!QGKmt-||JlFCQ$nw=ELM=$|LdIbe@8E)ELmjjvM-)pu=)7d
zMcI9a%TvaoIkfPK77|B|*E#DnJ=4%yU6isI%`C>AsFXZLGmo(+D&#X`@R@NTDk2OR
zgaHa~1qNGz3sC`yz#x&v>A)ZzxR4{B1q`wX9{(}H>0x1*>)(dBxafl7uH8_+8+zj%
z6r~eTaRLs&Sf(n);MJS#y+bF3fBfzL+}D|MqTj;FJ$rOVn-LEChQon)hEsDEyc|(k
z(ePh~lsG|!1VmX|!h(Pd1&{uAo{#B`5p#ummUq$68yVY;Okd16bEQL<TXD@=sKJJ_
zF~t5Mu1%pC#Y1^KG{lRG4gpjgg2oujR23`yDzkNSSD63nLmHCp{jAo0)|i}@gRI^`
z)*NG8+Cq#;P@V*PkRFwFF5Dwd;+e}-Wi34Q4&xV?!we&mTfhVsFvE!CmNO>hj2SU@
zXBm^TH2%T_d|`r!WQPNvaNtL9Jn)GJenhf&fPft!h~NwmkO6{-WM>1PY~V-mMc{N1
zxDaDh51i_O3o$M&z^Mhe5MyzKbv{B;CFo?b&RH}*%7z|gN8=Sl30l_gBs-AcL!9R!
z&Ks*4lsx0n-%c42rCL6)mbj%T>&JVy(#rca@BNy_Z+P!FG=9f>zoYSL!FRRLnW_<`
zFu^yB#=C{FyM^&&K1Bj@Y^eFX75cmtI#WA1e~g4#S)3#reHE%-g(g^J=65gEwbYRB
zN1}6+)xJq&ssr2JBrCQ6JQ!|!Bl|HD?lgu=V@$DlFUpEPKlgr+5o%b^YOZG;ak%xJ
z$BQLW&G!bV>#K~@RmS(p!6(DqPbS|NrGwD@ARK(`uuP#hSs058rby#<R7<hG#br|T
z{8lL6O8ctHVkwqtTr5RNab<~=yzF_<`#kU_QnQI2yon9PtAldStbZaM@Lh!J7oi?5
zpy1@VfEb}X5nAALB}(Z~kq)~NoCVFYpd-P#a9A#!OmGSGDuKgs93@i~YVjMaIMdMl
z6IXo+1q%7VHXk_Ps2Vw)z_x(Kg}}Cu#t~3I0-BJ4qM?2?!8nWFRi>)@lBLOJ`eifT
z<iO`JJ~@m(jR)m0LkYgeINf7hNnmjqm|X^zWNE9xlxi@Y;0(?wgQL8W<J`F8+(aV4
zw>kaWw0E|eAF-O)b)4Qu5stDbO;6k)o3SDjJI?tW=Ynv@zwExIV-mmk!=Ku@-$up@
zKlsb@`@?Im2I!-PdM>1%#6-BhjpS?(l<y&nCsS3q@h;&K1DcbvwvD?+yT%ygGgd@f
zR&YKQWKOFKLhU!ypIC{8w(V`}k53LRd7wk8EQFK(j`Gq`qHV%@FIZ{u__zalK37oW
z3Z20h&San-l;v{L!*!^B9lBscAD;T#Pr(Pu&!Ts1XN=kzE25ZJ0rOSB24h^-Ks_p}
z>!io0Z1<<EE3Tcdo2%L1=Ti^|MJY;9L<u^015o}@P{h&rpkQ=RpmM+`1+9~UF5zzx
zbXx=if}abz&uQE)7`78X7T5Q+WvZ$^nc;Ot=Q?ACU!mhH_00!<^jYd#K;uHtw~)rg
zK&Kd(5&ke%CyX^Cm6`%pzkszQHX@GGiQ~-3#A5`L7{Qw01i>VM#(M;lJv2Tc*q;#G
z2!ESk)kY#!<O9QA5zvJ`+t$&rH^B#>@d0A&QDhE{Ps1Kq$>5zr;;OvjKo+WvwP<5I
zLt<3og6ij>9(GhBt^GoxZ+*Q93Cn*@ghU3h;m@PLc%WE%Go!beF~xg>row2SW{ilt
zP{L@J(72B2RY#}$+ZfX}raP8GlvrWx+88Gqd$ch=1ixea-Z23*4tmFoqOtUjnMiOH
zFpUD;iMtU6?4p1ZjXk1(55dvEFB$~UI4By7qOlYWCK8+qOjAL3;;y6uyHwyrV~<qe
zLvR}KO9KHk4oU-~Xe_0Hi3H~Z$9&*MM6&=m7SOm5I2O{lnGI}aLvRXJlpeD3hpZmK
zX`DQblv9vJ=Hv%ye2de*MQi5>ULL{g5m8Oy^;39C^iz5LR2tXt+BG~Cd)M;XwKPr_
ztkY>xzb@!s7pVR=K&=h5ryL2jB5522wW4Ue5o&Fu@g}IXiN?{;G#Yj%M<)hq#n9~y
zP`v?a;;c7Hm_qeNsEKcNh^VTh-UKyq*(vhrpzHJIEN%a<)9Sb?j?Sd=G7&kGS-g}*
zRK6&!6@1qc2XIrQw)3vk#m7;~_OG5MC;ipa<fOlPnw<1kPcypiM-L6PW6wKmk-n`1
zwnedvioVRQp3)x$DeoA&cZ@qZI?=!`8n_ctPXl&oz#WSwF5+-9)!irZM!B4OE;pFS
zOgZOXPUC%oVxM4v+vDeyJ{n+T#Uia!FDU8-ssjE{P&_0U&kU`h@<tUIu@v+u1zy-6
zMEQMakwC6*Bu0-MGHzh`2=warIP-hlI6O4IgN24ph4NHrhUG-2x`~B7fTW%FjlY!s
ztMA_i*!VUCnxpcf7MbG_UOGbW^BTc-4cX^i{`b7iM-9_J0rn|qeG1xOZQg<Voq}Ej
z7enh}y1f|oE2i5^phpSxB>W{Xqy&Z%TnepA>E}yfzf$r%zEOq*A{ok)p(S1u6e>cq
z9Abu08vW1fW_XS8MGfR~DsPvOSI=kq<ul&IODbS|3K)MH4=P}W5`2$wzsK|=F6Uif
zb{ANZn6Cj$X#mp+&fwfLxPEw7MX7=tSMgVu)8sZs$yFFXB8<442fAHu<%8dMLE>>j
z&^;k^X5$ZzyE;I2Vu1q+1FC_0H5g8u+y>y@0EUwpv;u=x;6S3#*TCR4jZgEsr}@r~
zMd>WBdzN5ah(c#cdHIek-7eOD7kztY6=%JQt{>Dv^*X4JH!ZzJs9RgrWf>Cay^Pjg
z#t|3akXqrzwR~sQJ7Yy5AF?QAh%Ks;0QVX&fV^4_7+eE_39be1wRC$e2&e@iq`e8a
zH-Q0UDVo6GCJ;<;GjMOF+nYf^GYBE=ZNQ+76h4s$0D86oN3xV}fWaHO{SD~#1~`%S
z<DC0(j>^7PaPAc}uH;QB`R+tGDtWs~lEV|FO1@tuU52RS11d@GQ<N_8x|ev$i@n5K
zT_Ri8(cb?OPiebL{IE-W=amXks^cx{_|DQYswwdnw`knNTQt$QnYU=B@hjft6;IX4
z-t#W+X`C)7(gj2Eei?!ygT~o{BAdoJf+B~;#|7Qvw3nVQDDvs{Qo*W}oC$PQPOv&d
z<1)djjK)_4t1C3F60E9dTrF5t6MT4w$<!gKnm?e!w^wl5D}>?~!QOU9ko4zmQ*tIU
zp=KsD#d;~TKIu%d>*)nZ=5O$d8$5M)q=8p75R40fh;b*B?<5YiOm#~YYqIRgH=R?W
zG__Io`V`bU1@*A*M0dWRRxyoBVD}QDOGT*^YL$|9{2WSfDDS%B5%5Jrl(|q|Z6z;P
z2YS~5Z*n|h*ugPuD6Vp%Y&Yw_gU(!Bg6fx`9=>CmyZ=1%WZw92^p25&ZKObz+qMWc
zTLfRcyt&VQ$~D;V??eCi`z?8M?q3`vyU?K3aSu-8aL3mN%0C_V;52e+T*PP;(YTJ$
zsH5>~M&mV&BY{RFp!!Pz8Ywi+0~&cWJ_j_;(fB6NxJl#3K;tou+kr+qjn}doYstN~
zjwF^wB8{6_jb<9Ruv`l(C-2+Ja;*dx$knSV#y1E5&pnPp?`~Sn5pOo4Bv=8fUchSO
zqmOi@Al?%6vCJadAn3LchY%H)pm+f4V@=Qz2#cqoK0bCQ7wa#5WbD%$9fpm9Vk52O
zwhD@^1fM*2ZNg!%)(8Li!+>wJAdhATl<$D%Wa*NjS27$%a5gl{h7JVh!Js@ik>FBj
zTM9k!Qp;5L@3DeA<r?}mKRD$X@*>+Ak9NivJMZY4F>v2PSFaBO_k*MwT^zLD>iMPb
zN|El}0)n=HktF^|0HFzB9Ki=c&_OT~$0MSY2|_c$IE+v4J|fIp4fi3#QqPI?T&IOB
z^__R-bK$?TkW~EV$|1f&jVzGz0|KI?YoU5AbiJ&)l8?V6Ks}~Fc?z9KRDMr@R|RLt
zQFNgEih%T(z`7)`!?7dS;d^&WhMsu2%k64sJuAzNLL(IOy2ZqgL%uGrTR|`mToL0w
zDBnk4-%@^*fTwcyYISAh)7Pi}Hln%k?quydpO&vskN-I-RJruvsE<vHmi^E5$>r`p
zN30$oY5r-x{Z7HbSV%j}w_h=hOW^ntIGMDULcdbdj*Ti(OxvM+JAGPq!l0e>X*dGS
zj?ky27}^%or$hO<2VQ(EqsaU#Kk^`Qa0+xg1w2T;ryTSs2VVFpT*tAs*g+!Oj(UVm
zIw{*?iA2Y47nJXU{jmUv(lMww27NG=sXiUSGrHp4^Ow+(rt$xH4Olhq8^d?YcUmB|
z;Kg}$8prDE{~Bj+S@ByJlnG$WQW<*^I9+9|t}=Z{4XX^8l!0FOcB3e5V%0aX`lPfJ
z!>Y%y`ov+2Wz}O@eXLASO^a3E%If0`e#c!zlLWRGIr#^n%|YmfYh>4Yx7<EDX8IN+
zgH?=X6`kcS0-8m1*7^(3`~s+~^(L;{Ca&{#J92sj#ZF>8Z#;+?n^x0{L88*Ycs4L0
zB;0OhJez6!lJk7Yb>0TK`8p|dUc!cTf2%p}+D1tiNB5flx0<8D<I2Fd8HUV1UicbU
z7EISNbPGi*QYt7)g>J;Nxgsd85Zp9mEPK4qXbZ&npKTPbSR(OAfbs-dMJ2(YBw9sf
zLbFU-MV*4Sr)U+W{7?)l1pFD0zD)Ib7*SD`z^Rh1$lnG|x9N)fGvM@$uE-x|oez_G
zGP2XGb0&?CvR@r#9f{Ld$odu175Uv<=x%N_S*rb<=YG0Ezn*tk&r@=q!Alu*qOF*>
zE9R-9M+I+JLE~!Pr<zUxRPzI?>43hPA6m@^lJTqgVb$d3HA<lHfi?7vfZc-6ZaVNP
z61<D(?+-o+hM&mwt!L^P(ZDk+2A#uIp#LgB)ix4<R{|J99G*<zl?jGm<MAxu-N=Ob
z+utBF7r_mQ;5zTycU-j{5=rA2u73=DlQW+4jVB*(;1B*#l&bvN4{!DJkCWd8?^xUS
zPhZSv3$P?pNdV>vz?$HLz~~^b#8d0=#mqB-HOBu=$}jGn)vd=r{(%4cX~TvG<qiAM
zag5;pKQI2aDxS22tsm+7Ul$&5xg5Ei$`3U0<|4P4=~GYN7(T-FIYO%AuS+$0E-`))
z^S9a?XW~%uQu&!D8DSM`u!<c{>Zn^;%dPAfV*2u+ZXWq~0`sA%DnAA#9mTSivFuo4
zLvx{SE)2rX%$r{;R_cE&(fqF$=gr!{ydD>F+K|iozx`a}?Z+b-22U4lM2mfsx46ku
z7seZTi$;P!eoQOwT_2A=$3yoaUE0Y<Hu$ql^y!xJ`&hDg$y{hMtv}cE(t27yX7W-d
zjl%`sa8j5+A6E#z>j?h5@Q-67hb+oRQ_bbXT#hPZmvdq{!MNs#)Qa+}Ts$MSrLz~#
z==gtr;k$fl*yyM&TRLl5aYw=;5N9pa)B?!Q5uHDQRQ|l)>XKvBcFnQl0s@0VM-Lkw
zJYvkKkdbP^olI)O$j@=)=UDPHg#1MRPcr;x4E~{}Hj?}tMShMZKSRmSAoBD70G=O7
AjQ{`u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp.79ABECDBF9253649.idx b/.cache/clangd/index/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp.79ABECDBF9253649.idx
new file mode 100755
index 0000000000000000000000000000000000000000..48c200a13e87eeb4f81c5b9a99fef3fc89ca0308
GIT binary patch
literal 18190
zcmeI3cT^Tf*T9+ChPcS%LwV>R3L+>)(3Or21r-ob)MzyJ5@QrhF^L70B7%sD*dtO@
zn!O<^phQJf2!e`TK&l$Risc*1Fq1Fm<UKj>IqzTb=<f6L-ksf@a_7#yvwof)9{dO;
zrJ*xDe^@vx_$yXPNvS*f7Zki?o+GEE^h`@h>4%+#feRPy7%8g?TEQ8;f4t;K$irCs
z@tLs?Hy_o{8f);%`{2P_d_J!;;aFCpXX>Xn-TH=dL!V5pt37vncGb46ySDwp>n)nP
zqE4)s!S+zSpzpmu@#Nd)>{;hNmG;xv7n=}hp5HfodY97&GaZgL#NOIA`}mTlukMD$
zemZ;8XP9Tu*tyB3m-PDGEeTT^^6rxQOIPTf^joe*@BxE;-Am>zxHBfX@YMAST@8LP
zE}op9qsiuc2+h+?Y{=+!fB!MZ`}+2c+p?$pIMr^rbzt9bR;G9*RV%(JSa)=?t+nO#
zpiAFAcsKInpstI@UbMRMaPyGQb#J|Xo~UwG{rhrGch>%QVGri`?;ZE+!D6%7@yW{y
z=9!mXSTv<uiB8Mmm<yRb?u`tLc^5wY&gE|(n05V7Z*G*m`e)I>OM1E6>B5#N=dXNo
zX`%jb*$_rC?tb2K_KW_mf55fQi##&EuJw=Re9cr=6s_E&JTXTxV!7R^8P}w++BK8b
zs2RIQ>pYEC_;tw}u=~T%G_%7$O^gU`2(MXP5xn>L{((35t^Ub2_shP2_?wp2=~v9i
zRAhg%Y04PuWtVo{&Wro0s7~rxx2(u%!G*GB6Sv0GwU?@gXID=NFlktJIH2~)^M-hz
ztZ{F*bkTgfMD6U??3;pP;~Zxt<wXW~E_u+Spgwh&<;N%A?K|H$;PmY+SI5<7tzM9y
zu%#hHz2!pThng0rrwyMH=51?n^4K)5ZuqA8JBI%@K5kvI#p?*Oou!xOpA6sq)b4%g
zt)l*WChM+soT^xBz9_zG=2ZKKiiXmVs$YC!(?gEb#k%VSs3vt?5!$GBw6sfdQNrkc
zlXP@FE4z%Ib$Lh3tHcNWr+Pda_rhUg-Rd_>RNSte*?YpL=)^!@q2#s0!Ec6?ZVDb>
zzPjdgtxNB!A%FBTeUUf+%Bj)7ZFt_KiwgO-1z-RAAnwPf!yo<B<7d_?W<&F{@A(J$
zySje3Z|iZ_F7BST|HJ5C`qn-S=(}HWeVl30PkXmN^!*K_$od`(sq}snl096vQQyZ^
zSAFp%muI0r_9(r4YP*%z>OiycN)sRI$LIO{AXM!caHab9BPQF9so!i2D%ctP-Fa^x
zpYA!+4;cHKjwrhQp&)V1@D-xZHMhd7in409(tmS#w_Z191ebX4O!~(AOVzdE%*Z3f
z7tK@+?3XN79$gdCC-!<vbN8XTUCw8k9CC4fTDJMq*j~48?Ou84w4zVK<<zt>L;XHG
z1`pm@=&|qM#?s5vy<)!zh_rgU@6OybwZz#i^#Q%}yMFwt(y!Zf2=k2=_jv$ToHumw
zvWfk1WZWFT{NeSZtaG<`8|v=~zx<{m?)H#M-yvqq!fSW)(`H;dlb<%{+WY+aYw^4E
zj&_|f_@vF~o5#6vJIuML+QAM<$-j0<>r=I5ccN4Fs~Cs5iS`N8KmR<jw(zo_f4$WR
zlWm_;lztjtZM4Ak^jcq=Z;Z$2p3La2kz!%@laqF$k#LgvV?o-b3rG5H<))nuaSZt3
z&G?JwlzQ7Ie%kkP&8(-E2I|e%g(I9{$f6(jHvHCY!t2u5c^((9EV-05gY)!_T|C`u
z)q3F~XEZg#`@XrOcgec&kJo)iIOvyWM^4&de&TtoS?Z0%(=%7!AG7_c$>I-(`tPzk
z^0;sG$WZljbB2~qUBCF@vt6G@m?(Z5oww?z^|O<1DU!DQp_jl1EP1hN*p*4HwTD){
zA91Y3tma~6%nXkQ#y=kTQvFJJKU<Y9bAy{>cC5F*m95hMZZB_xMW3o>Ce#;hQ2sjX
zV?ji^{?sXl>MaZRzK;8tpfSDP`kcSnuPfJHP6(E6yD9zgXx6Zx9~VxyKo%5scOBY)
zUvQWlcIm4m<$zu}z`o<8q>L;r`lB+zacNL>%LX}=Bb7x~i>9|7uvlax_17e}sC0!Q
z67OJCb})*L6IPK;8xRqy1B~Rb2U1xB*)`C-<Aha=3Ll#rpB{Ky6Ea9;vhHNlwgVPR
z%IB^P3rUKX!=6Zd4YJpuW5)@rxJ@gmdRq{7PZMe*6<=Ln{k8)ZdrZ1E@c5A~WpZeW
z#5;iQ4q(}F!YZS@&z8Tq|Im?#T9!(Zm*8d7c2Md9fvS=ckiZO&=K-@G6wi=DbyP+q
zU?YKh#|f7)>!R8iKb@;f<WLi-Bm>)IFuLP}Rc1dJryoAOW~U}JM=GNYM)z(zVDWc~
znvJSWGwL)Uhs0BKr<k@Ku=tm;iN?#nvy78NH6&hUwd{+w0~Uv<mM%ZrpjoI1^^rJ0
zH$cDbfW@Kb-&prO7^~V9Dv<aqm9KQ$4p<yEttoAI;7U79$RqI=x?dQy9k6(3N^=iS
zwb{cop&k<Z>-y`r9k6)!Fhxp&<J`k?XokcmA$t-!beyoto`m1?Lk$g@G@&k1@zeFw
zZ#!Ud!kZj@_uVbmG@%X>kJTNk-*&*_6yv1_@_kd+$e}V47l2*`sI#`8u*#7ML+jpp
zXpWP^u1KW<vK6p<#|f*P)gDnibDcksLm5(;X*F|t+X0KSTb9PG(O&teE3`o3sVY--
z+74KpyW~25DP_%FIn+SnJjN=IaqKu@m20DYyPhx9eI$o{kV>JTS11^EoUqCbR&h;X
zJ^rE`_ChKr81@9C-*Lh!MMG?TkDi^oO%6?v%5BKrh6Wv{PL;WHx9op+Y_TR}k;)|9
zNd|2PEWVTd{!Qeitw%MX0TPeX9jD)Rz~Xyax&BWV>gUU$F%lO8wiuXqoUlrjvwF8b
z8aJeBLL;OSs2ix?cEI9i3M-#ql4qgL-W`cUfK>?iqT_^Bs#|^y${u_j4cy*HWe>yd
zVe~pqSfzHh?ze$a;AmIa4XJpmc<Z+vu()15@m=@(jglM+NW7V`+RTjXIAN8>weLK0
za>qSyukag;-VMf3YCmC><{pN_){cu!(1eDll<~Ub_1g|u{BgsF`3FvYe_sw&kT@Uo
z%11q`{e)GP%$gmP<3ubbROzsj*ldW>L!kQ*Sm1hP+1#l5naVs6rADaS2)p6F$!85P
zb{?LQtt?7UfcX<J6kj1qFM;|?U`6;HP=81BW@x?{4#oCVloFtN0<<EW3e{6-J`abU
zhhO4>iOT1OyyT|f`ofkSy-KTDCodAtis7t#r`jTq(t_<*!_kZEhCaLD1bjc0w3mg>
z+dox7Q7U1KN|@gG3Q;N+jEaTccqHICf>G(18OW`Q7^5PlFS+oBV01(1i!W5Q3Ew_{
zi*327DES)snqnI-N@I=2nqq4qN`6Lurr5NIlE0C^DK=H26krr!imjq3jWZf&ipPa0
zjW-%^ihG7A1sVmK;vO$blguYs6?F_SQToFC3oC3LMQO76WGg&liP9AFDOT8Oi_%>4
zxmMW3h|)aMd6tCcE9M&z3NjC}!c8Db!KT5MgqA9n8W39EYq<qB{Gzl%vBH2*DCiXm
zEbveir7)lf0|taQ0liJ22cEh_X$#QXLi0W#?gMf>J&IBe9GnA3kS>!82j|jU48>w7
z#}l$BU0}ouj2s&@QEFhs21ZUAw3+SM%=V$#wV548_!X!9iZdkqfz$p#^Ce#660b|-
zg}g=~&856XDPh&)`_G$m-)5_e(oA-y*zUn7DT&e|b`eF(*ku%j05$|rybiGIXif%f
zGR--F&7nCTu=z9>0JebUBES~Wyc)8rA$7kKkUc?j0b~nkE`w|t%}*iwl;+J0yP2WN
z+s&}MX+F-d$7#-E)bbbwt{XM&k|8#~KUl!9(pfOzEEtG=7*^WP+3e?Raow|0hG3N;
zSYyU!0FCN)pN@>`mr#5OZLq-;rPqS^TCl;RTFdgQ>%Vzty+;P67WmYHKy2`|%ML4l
zNIBVG&Pwxo&g+GT3ZhZ-MiZLvGT)U@aL?dgc&M;a@TlO?gqFH2btSaIWrZsqI4l|{
zE@8kG50uuI_dg2)$jhGtX6Ha}!qvdM8uY=B%1Yaz`F7X`R{|@oW8Bv<gYi4FQY7OZ
zN%MNfZarg9<nI~R_lz5^Oji2HxPGMhC>MB?n}FpyYR*^OlP_FCwzioW*39^00UGhj
z!@`vPapfA8P0W6A{>T+Zl+r=JbYO$uK$KRp{Z_Iz<ib)`zmzq?j13y6(&l=T#+<7+
z*9=!TvS^BVCWMwNmYd)z6{Q%~Iffn4S>W|(^Or6SY3Pw}@>Vx_S1b^vSG>+EzSAg}
zxE(v)IOwLzzs>%fmHyo9{rX29UnN&IMnh$){!~lsoc5df!N~yrk_%a>nl-Fujd9ho
z(o5d-CEqy{+g9z)(3^H?k$@C|VKp$K`>)}4Xhiz&S!j3`8ezlFO4-mbo90x`CzbOh
zH%jAt(r7-$O*qC)#Bw}AAm5^WxkmcaJE7M*VG{1IgG{F@eWf;aDa%R~U_u3$NS?3>
zI5z<o!f|j=9CRR0oCKYdpbO?fL$(e%`o4$S|NK=Q>_Yt><fXo1Mf3)W3eX=vu_%22
zdLO_*!g0_w4o<{REJ_(rF9Qz5?6fF+PEt+yV&oV;X6zp`)GS%S*jLb8!&ufZPPno~
z=>=o;f}sZ3YgXqqJ?kcN)`?uFu@R*d&N_wWdd{hy8->er?&flFsz<LM(a7Bh#Ern3
zxDQRNtckV5ir68OtF$e)NgJPrwx^*T?o6WehVgyF_~FY$)N6ICS$)DWT%Q=ulIDIf
zT<1)KEDL8F!*$-al?TLd648Ie`9I>u5pLovnmEd>e9u|Dr#X~&3*`qB{ZM{rC_kL$
z5uv;%&7(s3F@(c;&v3r8zC`JmpnXiByp#xK-w0(t+{fKR)cNYy34Ks~o&}m`fi|(p
zrL4S^)x-r4URU7w{9({8p|uAXKW3;mu3(HSXs%%tHH-zmu(b#1zF_(hPUQ3wIV-{`
zoL&me^<3|I&JoLp`ROJ8{91XKTI=*;*USvX0#Vw>%JvZhBTC7vESct9R+dZiC02Hc
z<_JzJf-@uKrE<2Z998vcoNXG-n|O^)d@rKEi`Upi^Ko9|IL$@8UJ*}qof6)tgr~Yr
zw4fg?P+cKR*)B|ZC~g|;_#!*i9uy*dcn8yS2V;-dE+c;5>9k|g@Sq{c#TVylktuMV
zIFF(rF-Qx09LT8`gDF}rF4rQH<qB~HMWH|p1zOk}X?5epFq&h5EEedJaaacI%YZAf
zEM>r>4EPc*2lnN(yc~Fx)A9-+Rsb!cUkSuYnyY|VMROyNH3D6t-w66Q0%yX_K-Nsl
zUjf-Gnp=RZ<<H!|g}(kHkbR`(pMdNW%@NQ(0=i;LDM}I0BLeyoj)L}4v^)xWMA7m%
zD8@l8qQ4c2TWOAmVm!@>P?iXFiGCvNp9q}^Cqr2>El+{66q?haEd9^iKb^il6Us7a
z`5`DfL~{-lbD$P6pgGVg2M!{f3$1c#&Vynet$z`U7imsm>{A$59F7sC{fzy7n$sBj
zG@8>H`*fO{S&e3vTBgNthB2Hu8KsS!tdXM(U^8dg%ncy&7S6JT=Fgn$Gp)apm#rkr
zI8ln?HKKUxeo?%A6zxGq^BU2#d;_nsf#$8e#@0Wx{Z{(=cwQr(mT%)Vw$WU~%Zljd
zdBt12;_b=%L<`-c1wG839*cWSjE6^%)1Al-PUNVWJB1sZLUTPgrk=P>Ufa%ZxW2kx
z75R6sdD&~eb0)xM2&l9Lp%3l0JU{Eq#uH<t*hO0X*2i=arD!%Fnw^TD08QzveF{rW
z@#$<pIxWv<?el5*D$aft?E*w`j!_&n1t#%<N&IwDel|ZSo2TZ5gUa>?mEEu%9pfV@
zE)S0%i=3bh?5GW_H@Pr_9hE_IC}$Z;&qRlL@5B69tnV{QBXOE0MC0*4tQy8IJS>gw
zmJo;(Kd?PMu+*dN<z;*M{<y>m>c81e%g{De5v4h5bEt646162%$mJKcUo44rUJKOL
z0!w0#%YdQ`*b#eI4ix1yM?ggcv?KN*3M!&#PGJ-&j2#(&X^bL`@Pw9l+p4V9_WyVd
zKG}JPdIDGd&Hvo#{qo3ym)-^WkN@#9bJ~;eIs8N82dKw;ntMvPpHDA1G31*{Yo>J|
zu$Hgc&Iue(c__-zo^iofh|)<Qo&+O_FICIQYPr$4+429Q5!9ZC!57XRZ*b&+tXvJ<
zYA5vB2?KGf;mnMv(w33I`p5+W>id8Jb`?-`6RH=ZXnaRfN|b6~-x@d&FYQGX$wz+&
zlbF~^MmkCJDMmU)^94rz0%Jh*%LRvWq0>b{<qKLhf-xRyqV!47`b2Y-vRjn0Czj9a
znX6%}+AReQjBrLC&KQyh3uny3Netra+>|~$cD?%Izg~JRJQR0j&Lb}!RHJvH+FfXc
zlNa9(da>C5_h1F;C3}GF9^h_#EPr0_K__1zN0_lb%ec<h-{hyJSnrD>L4swi;Ao!T
z%eE<Kkt&X-qF_X83XC*GkY11(<zFd+O$dboMJO;KV>Xj@&LrW8??w%~W;tP?FEX9p
zF5W&k?a&&7a;gPm$k46>PIWZb1E+eL8-P;-%{!sfPB;d4B2n4}op#Z@8#?W#c@K2j
zL->ch#}&)iqC&2<8yVC<6Vh5I81oZMKhiBuGUg{~KE;@yqPbkKC>N}8H~6{X(dd%m
zky@xx%GjZ0Z0FJ)8#y$hTJwjbZVv#P17HdsH==ZfwYkD}W(#l-9bML%S0tBJaQ+qa
z+AE89&*F#RVYu3Q!E&SO6aCO*gbK1yp)*v0bCIaL)<h&JZxti1V)RKh#4~E~j1kTQ
z;V8UJr7br}E(>xAawmNucucS_p`~t1hY<Qs@trQAP~Z>>oJp?<18!kp2+hO7zzD+W
zU|>3MAhom|T5N~?Nb@H`uSC+kC<+I?(rI4F4qVAP;OY^jOg12s4J2I3nwGMZXO+f{
zP2&QvJZz4di`z#p7OfC-cv%jwjs>W>c-a-2-|{+dd1{RvC3KGxG>HB_L1UkwAe<;@
zB+`6T&^}6BoJirq-tgQdD^WZCsyyUZ<q^1_MVb!w-EeniG%`1TzqpO`V%9Wl_&mr>
zltycg*6*~oQWTI!0Ud1Uk!J(s(KIK6KFM?(GY1UF0nT{qf`}4z1M-10&f$qDB|M-2
zIOA+_E4vqf5rm6DpJHG`27EEFF9uFDyA=Zunth8&kXl61?mlav4QcT;(0&bcqS<W?
z^q|>y4Vh(7ln829KqJz@E1`BJ%~eplisolf`x(vEP`jGu8rZW2T9EQ;p>{1TuY=lk
zG;d}+H#5F?K#5Wu(=CqCB)o^w-^27IzRDiPVh>|W^PoM98{q_IZ~`-uX0HThEY1E2
z%p}4m7^@S^0NSououv5`V|9w=Ta4)~I!E}1>Gp={^pr*E9n<X{%^#R<A85YC%5Skc
zq@&le@>-hfSa}`I^{l*}<_1>YK=TV${(|P0to$X-N4RcBIO_RIIIR-SfZYEUr*(_w
za9$qHQ_mN{%Ohxh&TBpAsq5eKrtf(wqLC%Yvgm<ZDwvj%94a#Zf@vAew*}MNG~W?S
z@6dc#FuhB2xnNT+NO-)9(mlcS9xcBwnBJ#3Lb+RnvL@D#db-o6--?N|km2xj@N~ik
zJ33O{fA_vcCy?{8pLgC*)@H19fOkGXb1LtgN^?5zoKARSarMhpSwCg`*IrEQ!tBqp
zil(@r2Tfos5}1Lw$HvXhedelT!XOK`Nsw(4sAcIcLAHzV_6<HBfkQ8OqZ010hIfcn
z-?_{3$<Mh<rGM<-<2*L2(zewGR~`yDa{kBYN%tw||CHu4y!#n43GaHBx%Rdq(x5Fw
zjiS^(#JDv9y(W4g69*M>(3%*pB&e4}7BV;hfu5&*gAaM4n^18RjwVj=N@dwf<>9y&
zAgie?u2CM08K>)&RoZp{kzzhFjvpCkGFT$Hfsx!G!cRE=C-jS+=G{+|7e#Xr?|zo%
z5`KIM-<d{3lZ)UVEsVqEB`lBJaIKp({J-uzcr${rO52tt@;-AFb9G3t@N31_6a^`Q
zbciGJz2bX{R(P%$g*~;_aG+-x7)6|rSfCdR49JwZ8R%`MxeQDy1CvQBmxHn8z@Kmg
zoD=~klU9y`W22xy;WN<o3?0AU&y3v9c;NwooEpY0of%B{J>&SEZnb9CsG047YZ(QV
zS;uCUYSjpCU<BQ&M>+qabgMq*{2$XC!Fxu~@%BHP4DtV&3{m>C$w1HhCzFAej=cLx
zx|f{d-A~b6!n>8w`GIJ`Ia;9dD;a`KhR_*dz^(;yW!tx-k>-B{bRUsDL8253bz`9g
zuIT+et~YUG-gQ9+BV5o37dqql2Lg*-RZcGGhf+a}K(7(llBR0`dMz|3LcK(2i%V?v
zH}%qK&Vh;?*bfg}QObpiT$-CXjb^U%#-h~1X|xbd{hhxPdb3LqvVGILPq)HOyC@yv
z#6w(X3ODVCEY5dff`_vxRRTpN$uqXHStY%PsR}5nXn7SdtD@!2K(86}B=6e{ESl+@
z+AE;<ik80uHm|@SqW=--eWc|dfyGDq`cFXb6D|J)Y(CM~Z-t7jBrhsTTcO!jdQVk6
zRK(NrcxV<+%afsAGVDp(BN<vG!~TR*pk4|sPk}Zma1fDaLcL5{o(U~7>FW<cy+gG8
z5VSc&U!MmRc_eQuN_o&M4_cA-zX%l<Y57Ii=OVNv@)*`8hNTQ{JZlqAb0%w(N%Kk8
z<|NG_oGgUXCPRNUCtFSPXHM@k*OSOUa~7Yu&Y>hqpE-xmbUx@aH}W$Hn2ORWUSk!#
ztXajItm3KuyNY*NMW^ys@uODp{^WkqyiqhskfU%VZ?u8tc-|<U=54&uHky-pyJVh9
zy=L%s88qMJP43dqQ_h=|(|nINxkvLe-sBn0)x1eH%{9D94bAU)*?XR<-;ccPBh4X#
zEJUE>s|DF=n!^N*Fo9YYt`lVIh&+9V-j^Qx)V@a+sfo93;yZJ;8TWSP{c3+<0Wt*Z
zdD(iN+D8({%i;**bqI>PweP>h#_)K`6;LyG!WvY<1>XDuZC)?)y)W~f8xD`xEjzAQ
z*Ld*X{UCqY4{~buqie%du0Ga8``c7!cJARq``c6(QM62T8ATyLH3U$+4ydl9IT@%X
z)0_iTb7;;7s`)e*0M!DTi-2ko&8wm6YDnGh1XMjia{*K>pt%gHmeKqasy?N8Go!kh
z>{mm(P#M+TG#_VFkJFsTuz5^3yuVFETY=ddG#3hNq0lMMa_-T+BK&m}@{KnN;zmJ>
zn2Iz(BaL|Xt#Mf88bOX-*w#3#@+ZRB%R{N;)?MYe+au4H71xmMv}YG-EqSKvW{rkH
zE-%ZaP1#jmc9rni%eN+`J2h4QI}H9Z4DdcEv<zw8p--xMJ!`X`PI9HPHmNi>^Ri~X
zbIaQ~d%v07@Ps-Pp1uVJ-lFs6VQ^p=olj0;1}2ew=y|UQlZL`?bNSW_y>8L#sXM^$
z4jn8CgI;0KhfMF0&@Yk>1|>0GNsJHiqz^KF2bn<3=NH~h{nq_<A~MJIz^5Kez<$n!
zJ(>L65FrT(%2=&3)&%$3+}_J=rDI-6|MiL{*L`l@r>v!iP(S_q?R)r>2xRcvzlXrh
zamk`&M%~5Z>rn|!oJJGz7Deejr}3WfZ^cCg1+~o={_*noYDMgWp>I81&>hyZ;(C%b
zLMs4POeK8v?b(vIZuNbU;@<};@!&?TeEXLh_+?N^T3L)zR>KOdVNmf`!Z@6TMt}SF
z9OS}Ep;x8gj4!<5Z1%ICkAB?0!}>478mEm=Pip_*1=k_oh>wE)t)IEz|3+JJT;{rC
zyNmXwj+y66XujM0A%qqv7U&XM?y%gMtgctMtr$YluoWW+r32e^U{AvR)xe?}^uxoj
zb$RC1NN0mufL9C6D_Pr>EVXeUh7E`zfgw>UWsOT&${kMS#-@^Re^KeuKuhEN70A$h
zW$=|HS$^#SCVPMlolY}8%UF?gTBRUX3c9$@wWinfYXu9OUPCccL3~H#x2)^SkB*zR
z2_>O+0kvIpuJbxnyH4ji&oF9d=v?J9rpGg;Gh2&tqr9w`Y@{fw3?HBTdmn(FKNC1+
z0$);*M}XrIn)85W9<U@){mVe}GO#33{X(Ew2rThZTa@;4j(f?v;CAy)zq#@2!v3)d
zx#O_EDQnR(BV-x2ak6b36^G3dEV6{oEqM5|5j4PCzojC5u#)qyB(VzIGmrsk|E7v4
zK8NCSdRKgtAU2U*@ej<``(F#_Gw(l|s7Fs%x^L`$s2nLa0G|f3V-_V@q0b(&zx45-
z8$;iuX@h?)&0j3dljUcob=Mx?hwAVyFT2ZAp{i%R>>1&w9^;u4HUX{O<L_rd{N)hp
z9<3kBkyd}k`9CAwqpI*gcGAydd{B9Hz^9H(=&gBUpIzkRDje%W{iyxRI#S|6U~`bp
z(_Uk3u8}+~b{)`IZic4K&>1htR60JqBL$sf6z3Q#JaZuzLolfq9Ej0)A(*}(6Bo{y
zBkSDyl^!mpPC0D@bUYM;oTxj%=?<MZi-b;*bRz5^<8+Ymzz<kA@~uz&ylt<M<NcWN
zc+7O}+&~#J#-oDf8pfrD-jn)*aeu*hV3((Emh1YSn{7SN6Kv(ww$j0dE1cRDnokRA
zr^#X$r@YV`wSQQM->Kof;ryE7@ja1ug&yxU+qs->Xqo(W%#Kj0&GFN@1&r}<Z+&^g
zJHV817}N=a#<b&S7zs@YCowunj4`eZQ98&N9%M{0<99>N-~J^dezgDOhcxayect*|
zk6!-{`o9eN#@fg0Dla`w`JeBu#@`qs+tB*yDQU_o&cBM-hUQ#Vv&FhEVvue4dnfgE
zsOPC(SC72<pYQs5(~_B{s9hL^o-m3xisCybM>O(yqpgJB-CAK^)Oh$DQhdRhykH06
z{*1msWldfZZfQwAZ+$PWxozK|%?6g*bDzQ5WY8SS$wJARL6i>jc8B>+qk$a-RLAWf
z?vh&G2z57-n7h)FSzmwE8XO2(G9xIZ^@mmKq3sWprase-nN$DviK=5iE9-%SY;5~G
t46w6T>W4Lylx)e*{^X|(`RPc0+L50Q<flFPIe`2eM1Bq=Kds5n{{u!Lu@L|O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_pipeline_selector.hpp.4F4E408D810A020A.idx b/.cache/clangd/index/gridwise_gemm_pipeline_selector.hpp.4F4E408D810A020A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..76837aa0677535f58f27f980d70e06662701e92f
GIT binary patch
literal 1134
zcmWIYbaP8%VPJ4h@vO*AElC7283cj2xTGla50HMw#K2H7=jeokUPl~w+MWm6XuL?|
z)vea*Zr}SRRA3L^W0P$QFQ1&?6D{`f?%ddzE4CH7Ee@a5f7&ONZI=)I?8C|ITH_k{
zsN(bLe>J=64R+VF`TcTS5$o}>;c>;pH{5JrW*&{vl{QZ`tIL>Gyer82`&=e7h6|rN
zF09U;dx@u0T;%m4Zg*AI^T|o2`HXKHgbW#6<(G*a<ey-u)%~>PQIREA?5qXa6P}kS
zZH;x@`ts1K`)m4Zy|RCtznXRKYu35w+mohjzkOicU+06G&0j85?mv0P(J+6NREy=|
zC7&Kwg}?W9xY%LJz3bwcsWY>U^UpS2`FgX6$Ne*_N|I=Aab<4OK42I~|F+YvDtax(
z&%npdz#+gQzy>B5IDm3MU{kt=Wp8KS0)7SoptLZHFh7`pNhdt&NN2mRaF3ONg^huO
zmxosxOfYbQH8XNB*37>uEIlp%7e50(P`?n15F40)>F<3!$6vQTGL@f!7bwlo!p{aK
zVA88Ev)ydFz*)@0zzj5AnqOK6OfYaVF*2|+0F5t7O)Itl#v;rYOk7~s!<@y#p<$+B
z0W(FXW7+=ue;%)5<l!*mH4}shTgiPqmCcs;k&%VNOvp?GCJggA&`6NC5k?xA892fe
z!9-cP!2aGaPkVpru8G@tIAj%M^<j#Du29u5LGW}`b&z-lss?agl9A_b;lzDAn0PoO
zj3ulPTI3Do;oivKm!{S6cMCtz8(REYYA{8GEDSSO8vMS%4AiS4VF(lE;gDC7M{<g?
zx-#54SmXe`2#Oh)XLvYt%yi(^!J-JL2oygEMY4vn#xPqZ9GrJqbMxmWpqupg_26!r
zk*P8HH3QQLps*3Y5!_7+bG%u+0^1&eLfqTT7iP$+V_H2kcZl`_RjUfC!p(+BGjf3=
z9~Q>vl&d?M?1DkTU>0l^4pa5E#!fOg{An>zSW#F}3nu(;@9jq+W^coQ!pdgK_Aue1
x)SN_6QY|h@F17|SAPE=5V&DRkAcBcQ$|LH+eEFPtEIgcC+zdQ$83t|y699*yfKdPd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_pipeline_v1.hpp.8B594F4BBFC9415D.idx b/.cache/clangd/index/gridwise_gemm_pipeline_v1.hpp.8B594F4BBFC9415D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a2b420caa1f6cffe696a1f3868b021b46a8cf277
GIT binary patch
literal 5874
zcmYjV3tW^{7XQ9`@8=-ikr6~<pcy}O@Y!hON}!^{L?kRBYQDCorDmC-Xi2NPixOeV
zvSuRrTJW(6;c5g%zVOjv(nxCR&$KM@frO+oni(2<X1?$EU4B3A;XmihJ@-8Bx#uu;
z^pj8i+(VMwuZ&*s#<Y1aF-elT!Ee^QxihO}Nt&xj(t>r><I~bF4z}D*znMdy3jNRJ
z9;GSQoQqRNbxUe|HT<t*|1y8p*Y-z38*4A-ewY}$<+-IJ&iwO*4NZHGEML;l_4(i5
zjUgLme))OXF7KGcOKsKf1wG&Cf7R8q4?6NUAGlcfL(7dD0j-b!{>rwxrm(ybGi!R3
zy3&6<{=)0?YDWC|Z>O$b_@G&8ZBv#eRFArU>)o4OD+03l&>b$@!yk;;<M=W9hXxuk
zdFTH41G7uj-;WMSUz1cb^+I01ajB)L+rLayYI4hWjq6$!GdQN(y~Rg!8rs+AzL^;>
zn}<)SU3DYiwf+~+Ty<F2mt@!r4)%_lv54jTveFfFapXnPCDXRxt!0z<um5yHO7Xe4
z-~O?QJsW*!Q)S-F(AD`{UhVep(4@y&Q~tENHDbE^%+-Ru&1pf`rvLqy@<rdZ=i2Vk
zyW<x9(WBq0yv3HNH<TwjMNb-&)p(5zpLTcCpe2F1=jy*36WeQjPHf52m}mRQx!*j+
zLMlE!oik&8=~pKmF-Iq?Df`v^@pqQ_uO}Zq@kPg$$StPmvs*IWoaq{{A$G$@V-|O)
zUHX;p8ua?uY35avy@y|JS@_A!N%zy&{PgDW6}J+ktn4?YwnOBkh01%5x>4VRk`SAu
z23vz8{fiU;Ecodfka}|K^q51zq(_9LHqwYjYSRgdSk_l=Jm6XMb};D*A`R5pKofO>
zB0Y}&dLmY8UlB^W+a-05bxx#zLH0x2o;$hMGU1a>Bq%&ky{?$9D}8i=B0aa??0dcU
z@ElMAJ7hbv6aC8r_J!iKsFBMCgB5UwBh-3?M(PAbqPl-Sq;mDRiNPc^Qd0AXGmpgS
z<bg=>Tbp;Ubx#Hn*u_Jg9vY_;6d4=)Q1d&Vzq7Fuv4nS0FA>uvV$;b3k%`BjuHF}Y
z4bIY2Qnxefb{44<6nVz}*yZKV|5^iHwMuG*5>cVpbb=yN%}+Nh8};9CF!*6fZ6VGU
zlBg3DnHSjl;=;^`S)d4fl4;G1^e@PspY}!1ZxbJ8pagd4AQ2sKRE?lW_NGgd9&N9G
z5e$Z-Rx)QLi_-~;EM9V~dv)E}WS_xIJxoo9Aptb<K;+oL3sGz0!Lwaqmn!D0Vu?CI
zk)ornYa^XeO&|i!&nM1&utp;&QarQo$`*-y)rka$Yq_VEGgCS1Cw1loK9E(p__LUK
z;mfNUeP=PL7L&#1UvP(#y$jxVO<H{b3<iG`$PopyO(zdTIy#ltn~n_x5je*>60r_^
zq7f9?{Pop$hW}CA3b_gwBE%XJ>0gk&wST}X*1q50^BvQqZeylxte;L$q`a>G@{ejl
z3w`OVTe3nitx)<%oy}CUWLIW(=RILH9xinYm(<%z&~0UqPNdFV<gkZ)=I%LtdE?dN
zgX?_ec2;MbX50J=s@Zq+eBHr0k6Z&=!0q)cqMq4wf+DqNW-Y#$(G&+60x58XSg(*s
zouJ6U<z<z8hwY9}iK6BZQx14UBPddrGU?K|$x~V(OTe#L&a629g6yy2)*p#?Ot}S0
zz^6-ybqOfZ2#S2Y>#d2l1<n|sM*`Fj71M`s!8L*+o`i)H$<7;_KnYxlBE?yx#OVY@
zj<(0&E16}Afylu1tf$s`xa1l^kyEM5e}3#<ZXXZ<gPVx83C^Ms6gj;)C(vw~`Y9-a
z_=Y+|<NV75cI^lA#?L5GAyeTT#fr07Nz@67ob7vi_zQdT$Ac0uIG0#+L5W6Cr1`|#
zS1;Xse+Gy^Y<4i~4yXYdL6M*Pee>?lzef2K!8yX5VTt|)*{ur`!pBX#oaKv)o<!Wl
zZN_D|RnqZU_xd%ZU<a7mOq|V7&oqL2+@F(H8~4fNxsV~CVnD<In}0#JG<VwctbJW!
zcl`~J1O2rhubFvp{+icTJ$jCFHP{~PKq;<<*h3t6O}LtA&valZ;_4jx90!(r4ryRt
zLL8VO9K2-DB@T2Thg`H@Ar4Ga4xzF)5eG(*LkYJZp^lz<b>|Rmdp&hvB5)|H_U+6u
zR<F1mTx{RL9O!*>_14<HO6K0KTn%NRTyvu&aW#yE2`GzY2`HOo3n-7UJR;zH!tzDj
zOjxssTL^0rv4=7b6~5m<S%Zix8LMOhzlyOc5f>?}ND=tO3M)1+zRnOh?P<npjc+nA
zv<Dd(7@E$3{udZa3|F_x30vh+SVFm4CMT4M_?Vn@OiniV$K|BsB0ec6o|K<3_^0H=
zQzE`4yKc#DgMV9g-8L|G00f`^9UBJJPTRF&C&xZO4<7abV3@B7!1H5<1O@7^KHAIG
zHX7ST<Iqz+Jgkk55b?-1>JqWLjV6gWxs8q%acUb)6LESQ%`mW+#(G74UOLPx>hsc(
zUQwTyy1k-4FHQD}`n)vNE9&#obg!WgyBhF^*3w`-kJ5V?U}%jDdL;SVb;H(9Ny=he
ztsny{$j}n~P2_4J9au<*V&C9%)xb77R5Wd%*TC2cLCmy1$cUMS@xI~pgrjKP5Wb<<
z1_49sf$*Kc<_8#B=YzVY#*E$_^3uK<nX7k{NA4&?@FAVUq>su5w;v-uhV70r02y3u
zS9-N8gArq;f(zj5m@t93S|TTuhzVaRCzOhKx16+FO!z%=(jF1l%89jN!ta+8_lvkm
zb~PCZ&(*84>#Bj#>0q(eqM^muq5(te&(L`oPjJ81jv4OP+cVw&i}m&lcF=k=++kS?
z&2260csvryV|=I!A4VvT?ywYmp12xkkBi2agsZ7MHN-$!@mVfx54f7`&Q8G#%GI^9
zeXSgA-1thFSIQv<&Li=8#DxjWp&fMRlN9U}xmr%_<s{n3&RW82Nr-`)NqjSL;WNk~
zQ`{{i1yhHs1=L<ZqtUTk-AMUH8e(7%jrUL&-V+Xm(%nE)uyS(s2DRUy(MI;?F`mal
z3|z_LE13)9!@;@kDwcxY=IU8yKg*)AOM^nj_;nUy;36fyNO57`!=cG=7b_{J_5Opa
zyA=B_B^vJtS3L^%C?SZ8rXOrd2#=x>T+Og#*o=@&v`iGxG(OF0@D9nAL$b{X!4cVV
zM8s!g%NY@$mo4W-+#>T9*=l_Mrp#}O_&1sVCgM6`sUtQcWFBJii1;kwXNlF|pC|mh
zh|8#@jM|KlZKsy)BHl~+UTQV?2Pi)v;zDLAWHuvYE16}bh)WnRVOE2`neojczRfJR
znav1UJF~Qlc&WmdDprGkLb04sY#1VFZxnt-u_DIQhKs0;K3J=9^pQL3BM)9>?8M-X
z`=%ZIm}_H?e!npsu&p-pU`<260z(^dP#^Yg!0^pCs0Hf+Xw&8zL!XARpyPl;I;+}X
zV+h_?g6=DWu-Mj2|H<=2VEu(qs4(WxF#JAL7;~6_2CxCXF4X_S)hu(Cz{@sg3#dTu
zQXq#KvU$XuM+AO8G3SeT9bxN;uy->tH;eohVr~&}2VoteyoZ`SRM@|Pnj1viNLeEl
z_OE2-N+!rxF>{rO>lv$Of_#x;E>Z;fV#Qo6;tHiph4R2&bSuP78yV5>7;zftpAhk5
zuFZ=NJbPUka$QNqUJKI}cGHGOj6b&H0fG9gh;NavLx?S;amIx!q{9m72oaAgq%IM=
z3u%&wlMCr+5vLZ?G!ds4(hS6yuaJ=1L~0nOPpG<S_ymC|zVQ^@fpZ5KqRpeYp|ACc
ztsrrDLwtBx1sNgYkrl)xVs`~e5^-_`87<<}3X&$`^a_%J82cJ9M4NgIL-Z+FHw0e;
zFhm=H@tZjE_=4pheDV3mSqT{0oQuK2=NuTmAr~(>zAx~xHr%2H9Lj*9&9dk$oTh-G
zO|Y0J*dhSKH@>0<oLr$g_$F4v)VoU1T_qZm7$0u<wl=lm!ST5QhHqrWZU5MNu5tX(
z)(O!1wJS00$}rq$?zCAi`Tie`QI{~mrL4IxXAJ@lQ8VWDpYhDaAOA&r^^TIFbY}zU
KZT#8G`1AiR89@>N

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_pipeline_v2.hpp.1C7FE339CB47DFD5.idx b/.cache/clangd/index/gridwise_gemm_pipeline_v2.hpp.1C7FE339CB47DFD5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ff45438c04e43d918221a85635e4f803bccab9f8
GIT binary patch
literal 744
zcmWIYbaQ*a#K7R3;#rZKT9U}Zz`!5~#Kk2=nOlLhIU@r@#hkTg4f75g@URBV)nGBI
zdOJZZqhWz<!vU`z-x)%jfnF1CU;j91W%Ts@fBg4-Xq(iQu)O?>anaw9tp!<g-oGhX
z6r`7<5~F#q@8RM7?K8LeS=RJQU9$46{!_SNYq;1fhJ=ew8!XaF_BFno8+D}XQ$d+W
zl#gtE%OkI~lNhe6C<`wUwF<p1c6CK!`dWkA4VBE_D^?ZWyOE*)eZ!2u%{Ert-;WoE
zY&q6y5PZA-Ka<eJk~Iv)mAOfEK-cH*OVjH3yM>>PiHVhgos*qY4oom`0_7MO7}*)y
zK58f_8CXXOF|n{Qu&YX|%7F<64zL1-#Tqe-{`H<)A;iQ8lor$zbOIAF>2>GXxBt4x
z<0r(#43y^9<F*77FzKSywBl%B`0;LCv0sC=U@sF7J2wkA55#B&1|D`n7C|@<=2S*5
zuoE}T)83!DYvMMT1A)RI*TGbEdoj1X?f-AV0yNTr+a2Z*ppl+}0Weqdup7%8tHDg*
zVfW?n6^8Mqr5$>1C#l5D%EKNi5Gn@~=3$Q#h(h8O@DxZP<Vtx;k$BZS)o>onN}w}A
zp#*a#%w0fXQ0TyfcceWy<HB|7Av4hRyxihQ0m3VY#8YKa<%7vRe)mhp>nry$MjrM+
nmOu@dFhfylP9lg?T$Ef4(#phMIWh0->k`HLj2y6d0LcRY6i?o7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_pipeline_v3.hpp.D2EF7B80D9AE44CD.idx b/.cache/clangd/index/gridwise_gemm_pipeline_v3.hpp.D2EF7B80D9AE44CD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..43e9e4afa3b3c08fea439baa37a984285f2849e3
GIT binary patch
literal 566
zcmWIYbaT^VVqkDi@vO*AElFfyU|<jg;^LB`%*8-jiIIV!V$RyLwtR;Scvv6I;o#5N
z>u|K~0YjjNPqOL8wSTs}dYif1VB^L;Pi=o2CokL(bX2f--{d4Ex7{&k*V^q~HKpyK
zjg_=L|6xgczHi@rIoTOHCJ82$9t@3gNIjY)yL(IMst2m8L%w@ur0eu>cGSLNJb3=X
zgw-P2T}MrY+B~g&tSgQ!yp@^0w(hQX$lvBW(yCdv4No01Tf`l4?!NrTH#yH3iYs%I
z;(=}zmsQH?NuK&%h>4k%flXdoUJgt!Z~*0iAU;Met2Op+HyaZp3j-S`JEt6&VBiGH
zGO{t=Ov#oNd-y9{h=~cPpIeXH5=_AKGZdw!72gCZ=H0qtzXof;UPc}^ZWeAHuy!B;
zv!0O)Y%R<>7!4E#*|}k!_Wsme6SsAHF}J+!|8KzrG{lPA5oR|Jo36aB8qBC^X@{QM
zNh&cj^RPMbI7z{TdDz@}+>v-eJV9{Yj<g47T)0j>1ewdrEe>-W51TxTJRgh)vkmAb
nkT+pY1o|*15u~QLD7hG<m5D9;l!W9lwPaUDb_RBsZUzPbLBgp#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_pipeline_v4_direct_load.hpp.7C3BB633588D635D.idx b/.cache/clangd/index/gridwise_gemm_pipeline_v4_direct_load.hpp.7C3BB633588D635D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e4070436c52f91ac8e0418dd37e6f3723800194
GIT binary patch
literal 1944
zcmYjS3s6*57(VB6VIQ1@;{q`*>@K@5#1gy2pw42CiH9H?O{kR2fJ1;_z9wX)2!VBU
z6O%%;$rvCjZPpbStqcZHA=H#dSZad-qm?t5HsVB<k4o+Ag`azOcK>_7`R@78|9|KE
z?=H(oOKXrJl(H(LVngw!0v;hWhy5Iz%GN|-gidh?RkU>MNqR9Qo}Bx4>9Z9-)RmvU
zRCMcFK|uEQnk{=eMm2jc@BO;kvE%0}x8t)?r}o!=?t0Yp_>-vLuPqxspZ}!Rw>Hk*
z^4v&GWLIC|!|}Y~D2v<Db8KH*?rXVol~<E$kn&!9R0l2FJlOS3PyW(m_m+~br9<mF
zT0F%yhs}%4H-22M8hmrxSH};BT?%};`Q7|2H|P9Biw<_&{W5(r!PY<jM)198Kkg61
z@{xsi%_TVrO<Ma4Bf+sNYR7JG>T5XA*qHTc7O8A*y}EEXUi;2)r@8qr!-jChSfOp}
z@juNEKdQUt+xKz(g#pD-W@iXzcq8jkf3oFFc0jskSKod2gWpabv(ycY9B8Xrd*14N
z^!<_Y=BfVX1v}_YjWO}>fZ<ACo#X8dg*{9y^}(F=A6#s!Qgg~cM4eV=q7aarDat;f
z?kn?3TJ`%wI60%MG*#h309rpZicXi9Tl_eg%7^M?I-L-Jjy^N;nXJHfCrciHsL^Un
z6avr*D;&2@1S`u4dLjf-f9kIl0#aa%LJ6jLesH<|#)a35LO2E6y;@TpE(BP=;4EJK
zK%RPE&H1zRwlJGc2tcQQyT7tx_~trqaXhV*RqBKQbhi1evE7;VN7S4jOYXu}7fznJ
zz$33LKAL}dv{cTzXB$LlBEp3L^sBkg_pP2fc8FzQw8?5p76Q->efut%kKWnp-G5q5
zs*!1QNa4#IidcISF<h{7^V#=U{j7pU9M*_+GZ)D5_@(ybbKCO+IWBlQ1)>V2VlGll
zqq5=>M=Bct)zdqg)_xd2?n7vVJ|Yt8CbUXl6$xiZXtlmN5;RqHczUCKOHcBI8hIm`
z$&Ls$@g@m!@=ghIVcvyjF?a_n1<!!qp^19~-`}|4qKwcynim77L}&pmkf2gh8YOZ)
znD*cpP=wF{%)SylhRGO?68Q;CCPWM^`X~i^LM4PA!j?lgPF&W8Eo~wO`*KRbwqNXf
zs<*g%cPE)P@r)#3V%DYC!LY`%>c3~ST=glQr%Pg&SfLY!>`O95<iW8XY=r`N>cMso
z&IAk-AXf^`htno>rD5f4N<s?_g%Y$;yHO%zFE;dw8xT5(4TDlVj<w@bA+KZYbrFLb
zm~h_l5`}BUmR3<HLff&WUBu8I)5JS=pb0EzdV0rA)YFT=VK8@W?ex&WDsU~U#XD+{
zl+cXW3`zU;So<u331qpv;|EK&ng2O)M?QWX6RP4>X4tvT?&F%=_1Gq7mS1E{f`et2
zPchjc6`^VRG>I{DNKO#6OQ>^^GYM=@Sp6w3Y=?#jbz{9-G$o<km~`VHIBa+yV1@7=
z0Cl=be)%p{Iru9Z7k}ihgko8ZDr2aKOe&KE*?3!k$T?L`3FpEp7na~g%s1lMr=c1q
w+4LI+<S#2;U*JvYC@XTrd2hUbsxzM;xIrAPUo$a&aNcCBk2+`$QqL0Tf68V@X8-^I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp.76D93CEF02B94AEF.idx b/.cache/clangd/index/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp.76D93CEF02B94AEF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a9d1d6782a6a12de725d0654727f134ce8870bfc
GIT binary patch
literal 13164
zcmd6NcT^Wg+xBl}XIH=lejvR_QHtPF6qTYVMFa~LRItVtqmQvf>@}idLB$eL5er37
zIw(XG5eqG#VpkLaMWTQOM8uHjI|PQAe1Cj7?>WzTj~rz7a<1K-x#ym{&5rZ&@|x3E
zCi9r%vufe@%Vr{(Os0na1uk1UKSq$rE||(>t0HdrPg?xMQT|}`Z4aZet!bHaoVJZc
zsmB&Q@IEFi-&;N)C~V%8RGTs5!+*T}HK)heE<=ZBM&u7Qit5?7!XQF->D+Un?{f5K
zH>>ZHwR(On?rMK@<i@6^l@+6Qtn*m6an=*v#g+|oix;L(U)yu`<UW0)_w=mwf8Miy
zFOB09rrtdNgF{ZkRn6u;UsmUxSl76HG;F@HyXesNF(ps!3WsQG&MQ;8kT~S*#J%==
z|LXPGzKc$u8QmS`7vGNE6&(0(lWM_-8E${>&6zqoc*NT{<A#sAp~JbO#go&|aH~vg
z59rrAKkS)#yL!dXTJQTgdKKP0cYP!X>{4OSW6r#Zv$S6A3AEKn8LrS}>YnmklUv(g
zuiSAWtnAGovjtb{-|Vkaw(T7_CMb6Lymw)5H#%Lj)D2nd-q-wx!w1WpX9x6IR3|3P
z$!#!7J5+7%b?}YeUx{=7_-@<7gw@mgPTg;7<xu+b2h|O~X)3Di4e*H6vpABVs?nt(
zymXsn_UOoJpAv<?Mg}QA$sgAk>}64<$kzt1i&Tl6dh^EvQ^x84cIbMVEcxo08Ady9
zxUYHhhr#@e9*6oj+ig7%sJ6!_qcUZ-c~wG>6TAIqzIxqlgZq`M{_5TOclD?}59ZEE
z4?eJQRLdxx<0HacT^izk@Ul-SjcmDK;OqX^3ena@(b^|fSh!T#CpCV`$r*J%sU<7s
z_^Bt(?3!bun72;~v6#7Lyn5okIcL1##$k!InVQnD^TXaJJlJ+q^ZhiJ0Pn?{k_{*P
z_Re&6RAZLg%Fiif3uaBTvbpy{;!|t}sizNqacl1EBt_qQFB5`J@9-FTqu}WFn^#})
zjkouD`h-4C{QakHV-Ng#XvKt&T~<37Y*e{6Z~kJp&|#;ifBE!cc)~5y`P;mY&Mp)6
zy6j))J9e(xYK8d*mR{3@oo8*hyw!&`xxe1nn!mO9MeJ6$Z7PeFy{eq+(IsxtFJ@_e
z8T!{UZ7QO?js*0HcUrroYWj=QI%CpTRxdWNP1S$e=nvv+3huZ5=#XY2Pdu&gm+BVF
z9lFaG#tu`6R$etmv~1mL4OZLNZ0+gxBUtJbmY^ha=#i%>3R;WKrd?mQOTi-A&Z<Ud
z&*H@X>rYkQFsKMvbs?#8o70BH%C&zQ6e^?7=MV4k<QFcVG*6*hY`l)9{n!N)e*f(5
zJ3Q|qXslYG5je5CUdFnQ2aPjfnk>i5e%nhm=gg?(Q^wS$^odoqsNTqbP6-+_a^<XL
zDgnKN?kzfXxa)__58npgOfyf{mG4#k_2uT~u^Wunu6dziaUkSbOzHRTM^<by`3`Nm
zAJ7ahr<L#<W>2z?O}O9}AZaF7mMLY$U7oAvcW<Rz|EXQ<$NlGi$WQ9pHuu*Q)xTbQ
zC~@bsWPi;wiBsHm48JhJCr{68p1<Yi+VTF=dMM03{<Hdyl*#)p?AQ^wI>B?jsrJnZ
z=a6?-y#2=Or>|D5eO-_=IK6pkgId@3+YiPa4Hz-zQ<9Ur;&O%B$osL%kF)PpyS>T2
zSy)y)G<H_~w3*`rUNoBKz{P2qr{`XjdPf-*Bv}RJ`P=l8^!7{{y=G1EtUGt#m2Y3Q
zXh?U(-FLHww|pJ`D(m@G(*eF2nzuH@Y5nnGedEH`fQKeY84oio<%7=b_WS&1V%V;P
zPum`b9N(}i?bYkG6L%gTC%e;UY=C37mqy{WuAZ9)W;<&X`uA2|e!IuRt&R7lX>V<O
zKh1wjA0cj<O6}N|*BK`V)@<5ZxmjokGV<7Lmm3;sxZZbR@F16llS#b>cpP20<I-lQ
zo6Dy*M@>7>&v)R&Pd*n-c7<vLja?Fd&VQ0`Icgg3HJV@cdSdu(jcE(U4m;L$%b_V^
zWQ~W@u4~rrmlNY`cg=h5a@_S}KX}WeO!?1aphA|j;pgLqs*mkg3WIdZ)}6djwctM=
zBK!Y(?z8vN6Z1a=F5I-Tx}|5#xEl?loQ}m0uiHKxzL{B<pS@{m_@E7U>dRL}tly);
z7babK{L|%%ieT5*i)YtYK3ZQptZIO}no8;?^@~4MsQhKQt3heYiY>zRx{I6Mn5n#e
za?!PioA}l_>-YC_o7O7?hcxD#P_{7ncsp#>k`Z5?dflF%I6UUPf4Fl|)z#-}Zx=_!
zoET*C=TURrg~rA6-K*?fFHH)&Z;-uqrh24-_F9|qI>p;YH_aLQ;?t7q`g41#r=_;+
zPXF<jy?&E@pQwa#vu55O;k#(*fx*4AYogX&^|JTO-SA=UiqZ|wHlO;n{ukf(QFT8A
z`<YI*aBCj6%h7uITW!ZNx(jUn8nD!4)779f6aMp(N7AE`<$?7%4)aDw*3J9MO|ZYK
z7`S@jtWTUwCTv=KaZ)bWqzSwAkVyxD-a%l{(a03APs2ajW%Gt89F|pTLS0Nb3iOU*
zN_#^n#?Gr%D@Lv>)r35zoB(<!fMG{NC_NI<puzfoNb1lCQ%WEzfn7QpLNPaf{?Ii%
z%~ycxrZVXO=yCu!b~G|Y4&b2x30U2kF!bm5&&@QU27XQ$&<n$(X>Z7L`c5)BYpS>{
zNF6F+N+`4mh29+vp}4pocpO@}KS2}fVM+whi@;B5Z#pTf?gaTHxu$AD1x$$odQo^z
z?G2&0Iv6&8x8&#=O{jw@;Xp4OKc&4P6mN~faOcGeC)J@GQ;L9n5g6Ig5Xw|V`^0|b
ziX~m4F{Z=-r5K>w(GbcEop|pf=<*>^W{)Lfp<rS4_aB)eg0em)SSmA&{g=JG_d~Wi
z)W^>_1koYr)X|V<tdKSN{ZZ2Gj5;*HlnW4DfQB6np{!J07S73Qu4}?Bm~s;6oy1da
zZwO^~>iZFQ{hqE?hsv082H2bdQ#u+#*`pG>I{ap8WLKz$DX~B)7U*>}gmTDejZ$S|
zlw()e9aFADrR&hPqal=t5q+BPua11C2?b1v1bUHpEbR@U9J^ZhxN&DgSy!ltDY;N7
z7rJybgmQA(pe4&==NGEOu9)&WK)(aEj)qVYgZq{5T6;UWD-<!M04fzgqmG79em&W6
zN@-ZvC+biYQ!W5>0SxJA2&LFL^Sdwg8UgCi5L4nIiic81LnzP8JfkO84VN^bHl`c_
zdPlIfv^RuO?aVhsZFtmM9jahT8n8*j(rs@D<>hxZXZvcqz0-tRm~t5C9mY>-ZwRHn
zi|O12THIfnP!m%Q0lh<*(%uk?Z0YxN1Dh=9{x?+W*!5%+ic&Ke&<vc21VyO@^lkw*
z6d!^1M_^x~d7_jE9TH)GA_!5ch2v`Bc(U(BsS!Fi!hsamaZ(*;OYvUhw--&II2u_+
zquvz9Bdd58XQIKG$c^@AB9BZol*Pj`kq?VUWTH_NpGQ9D(FlrzdHZ0#FU2vuRSeI(
zKaRJGV{sDio5YWy{ImSzv;0(wvw6>K-izWd{NOJ<Gv2TK=&$@(+8-r&MhOf*O7M;n
zhSUDLg3n!n8Sh<T)LmgT?JpOc$^{pKg^T@+2M$fXghc5C^f&>h(}hZb15)5Px=<&P
z`$;s7+$c(s{Ln~#2Eod9+r$5J(Ef>`h&9CMATT2?L&Rp+C=8epi!5RtGdcv!i28|G
zEsYKXGorB~R(7K!z>Mqz5t~Y*a9~EPsED14(NSPV{EdhWvQY#uBeoh(1Q?$H=ET{E
z*fbfR1m?tQ;ei9=NMKGZhls7UaTG9j=`iG?lmrZtfGIIyqLd7@lYt?{8NeU|n37E>
zN|``A6Btr_73g0DCd3+u(siJJoyC#REfNl+f{uc2Q7le_ZfS5B75pira|(4Qg2p=t
z=_InakT)vi&1ioiKd6xRptyt|T*5ONU&?!w@{Gm@3GyI8i}DMEo&|yx!O9t@FI)0U
zE-8xGfTJ@&Y;*Q)KO#0cD2>Hg0A;bb2%sVs*8o(*;!ub}A@jU2h{9NW2BI@8&Vne5
zVpW}c?sj*dEr3X>0R1XJf8s=uRLS+O<Rmf!qA5_RBOal~D-8LCQDdU>FiqjlxL^Ay
zA-v{;l7TB(bA;D;P$qCCYuxrmzX~v(zOf3JR{;x(n}JI+7)Y!j!dA_-5e_CJLefX*
z`jN$ToP8bFhxRA&<C1tkil6blpYfeD(p4Q;G$b}R4~S9?^sIrMM8mgpjW<l3o{Br(
zqyF#FB=Usr4<}uEcR9U?6Qxpsf9=RnvB8G;*NzZHDW6l%=QIg6dTo)ZxT10}eqs+<
z52f~{m&wTP;*>f9I4aX5pq&Kti0DKqm~#r|It{w%;B#4xj-^W9yo`wefKo^NNoGh|
z+|eF<rXIJx{QZS}nU?DvPlV>!`D0O4@(z`Jr&A-QPgkiU1SOAXvn@W!U<4IbCh*Aw
zBZ!`f(iNb81sGCWat-KTV{sE0(8PM*k6^$@7Po+LEnqy+Lu^Z-eK_llKSJk^the2V
z{PxieAxc@uCyVudkI=wJ$d$-Olpdpjk6HW#4Sd4lax}1<#lNE=zaw{g|0^`)6^ny;
z?_i$sX0iOFSbhrSC-W1N`2dPD`N^5AH_hceb6Ia%$`3AOy=l2HvRv>b_e*_3GB&9z
z7~>U-0R}O^ngrR>oIhO*;rgqJ2*)$3EkKP(unqTX0Rt#*1^TVPkj3_`pf8JuwE`cC
zlljTXbi_zH%TGT`-6*y+{Pau~2McDwf(03|?K4O3g>RqjfTDDVx46SQkq+z@1cOAu
z=GMsf(_PjSuCT8;o~%D3Z!w<1S77=T^q?b%fZZdY5%E7rItIHRV{sDfo&=4E|3T6V
zZsZHDGa^D#H8--F#dTaj9oK1TkaUrEzexQLk}mV^mnrV!UA<G#oz$O4QY%nx1#&vD
zR$$f&%vtQx${Nvheq=h&3?zdelR-lxlF#9N*`CaiUDnU0>S=&zcNnko1K#-oy`x`!
z|HnTTy&r=+%#fKTv2Bj-huqKpVWRrYiO#8S@VIB!5<cR^KgDaD;+d%UG_P@*ViF5s
zkJp~95>epb3;7+Q8{B2__8(5Y-u@GI?&Z*-9C}i9s(=m^EUtn!Rgl^FPdVGCT&G79
zr3%irg2fTYC4#jacX;PJ?9M(dc$^meh`P3^uWh0*m|zl(DJpd&({u(Iz&-;wP!pL6
z24@0y7JFubVHB5w{-wZ)^>_VSAhT~yaXzQGk#yquoPIvXgj>~|Z#CB$fDcp|bp1Q8
zo~v=doD0NUU`uUt0WU9LGlLRdUc%y1US3KuiD&T|wbugZ3=bl|gLI7sI~!lKTV9s>
zuUmddP&eUKb5t)(tN@x7K!?sbkF(6<tjQeRb{0Fotq45UrES%X%OR7^R6yej7FR*N
zD%cr!h|*I||0!of9@Mt#`V}mWKo$|mnf9L+Y)=dBWK})K>Bp>VR`675i(C5EKqg2%
z!pV<t#`MMrP9DMH^PK!Vi!XBWi!6S@*}veJw5gi2uV(Qzq;U=Pq|d*BG;Xl?CDM4w
z;w`+|7T$#NgL$=J7C+%Np76}`zVJq0=ms2iH^SxU<e`D)ZP|pF1hlA!PX=N#i!*?j
z0ko*PW&$yj#aDp*3eclk774{js70(d&Ss!ZBy^-W3fe@mI2wx4EI$T{F)U7lVj9#U
z)*c7-&?XH!Qk)LObhiI26wk7_7OK|bl3z#N4rfzH9)<KMjzQKjs2{}%$U1?=DM+5e
z@>7vKmBp8k#w9k}xr7`pu@OZM(#T=^FC&f1EG|GA1^>hj1?>HWNTZPLFG3ncERN>o
z(L6JscwQdQ;sjovz~Vw)qmXCJM+xsx!ZUWWly@j)ag^W`B`_PIP_Qf%Ix_)Lx+}Qe
z6}+kZ$_3+cnw^m?j-6h6)sR>bAFrQGCLM<9IN*QExjp4r%g(t~us8yZia?zeictgh
zFV>QF?G;363paiX_Z@xpG0x!_YvQB1@zHF5Hs_Gd_TNYT_t7kR{~P4^26aw#)aU`)
zg+mXI#UtFxjo8X{MrAk{<3_}=xC;4Jp-vM#abenkkt$LC*v{QX^4o|JRs*l!zz-sC
z4Cq#1KO<Sk3=<Q9Y9g?vyYL)PJqN7Gt_x^AY+rM6v%|lpk|aJvN*$$FB5qL%;_QOh
z$wzW_kt}XN@&;r?hBkHJ(QZ-vt3}^jj4?fZ!F%80{Hp(TqDYF*SQFbS&m<`#h6%g5
z_Ifjw*cE7e1x_I4*%{Bb{~%NtKf}S;eqgyD3?Rf=8zjZDU#it`M^DsKheT<!iZk68
zy|m^2c8eqLuqC!c!7Wi3N}SHZ)Ucu-Du?Uh`?H~HHnbrnZqnkRwb!}H8%3!Q8W%z*
zQu8IHBb@BEm$@l1188ReJ)%cAF5{d|(PVeso63>*?s#kASLbuX^ErRA{6QDSS?_%M
z>jvCmW^3k7<`k5BR<UZ-{GafDf8cdK@Ro!hV&*m`v~13ve|7C4!?`VQLcco7@nnH|
zHqfPx;}TH6#Nr~*tB9=wghHQCIEF;gqO=dH?1P#VH$asJTq^D~^i5EuiN$%GdLG9(
zf?7_cmevulH9=ZikRiQ)E7ID^;`2!TJYw$8Lh4y8zQAi;;F<e#c&!{3ZxOV%2%Y^Q
z3##;Q`<5NUqP!(o-4Z$r93i!y59ZBX`saT={~*Q%Yf1Zo2i1}iU{?aB5nC-vt<bI&
zP9vh&{UTztv~kk+c&Ph<(|+JWZX`z}_=)Z3Bcwx=HbHR{G$5&xC>?;R2cQvItUV#>
zeU3*j&O}K1XTpyDpM)Jr|4i7C^v{Gn^zmv(W!bqmZF;I@CJ{Xin|t+lS6$P#y|+`4
z?-Uq!_lqF^g<^7ogSS}w@eQ3p8B{5wk!E=4XfJ<{JYW2xLauuu-6^C@ixYwNV;%a0
zBw(Ke9EjN4YAJ)0fjf&mlff{GOM!hUaG<Nu0_|I%16hT(LX}VOKNA71d{&`faK0~S
zA`o%R`uV!6%Ln|c&maa4%b@-6iO%2=RCz?JfMhkX1#LfqBE&QPcU+ZzT_Evo(VliB
zl!#Hq(H7K>VsRSOPJ<GS-eQnO4C<`qic$j7NT4|3uJAaxR51{*V;oS612#nCMClbL
zzT$cj;UuoQ$d6eV<z*vE(NH^@Rtws&c{H@4BZ`6AF>HSf>=gs;X@5G@PN&rcQ928?
z&$2j}vkT@N$%{p44`;WB#iuyCQ!Gy6?2=fVg7i~R4|;zJvPxk?gcRhI!k!1EAnz0!
z{owc)X`~?odVd-+O`{TP>mQWHo*$*55oxHi04qv4$S4PO)@wxRGBUc%;zDFp$l@Yo
zRK(&3sP6;Rc|Idb&rsiIEN(*bCZt2h_Zi7Qv-k^=e_?S5ZyLfg8nBBu-NoYFyy<Qh
zALdODv-k*adW6N{ylFU#<9Ypfo{@JvZxzooAz3`{6wg-R;(70Q)>$X;`U&j)34E^v
z-k#1cfp<*chp^ZyfoEdHYrOm#&&cB;@9~gl&cAASc@5kDiP!nWGl^ykuhT*?art-;
zwx98mJ(&D7;@U}v+y!{6WFvVt>hverwIg{x#Tj*LGPHL##((n#ZW0+_@6vu&O{OYJ
zWk4(g-qcctBY8OLERmd_TUp|%^jnou+li3!8K6zvLfeUuavF=XfN~a#i-2+wi)(;#
z4U0peawue;7Y3EXSbPR5pJ8zpRL){?F+{~sg^cIoKohm<?ajNf-P|sS+u1X&Q-a1R
z>ZWilN6^U?OsSS#7j&+(I8QLh6U<3}cH#X=NiH8=e)DUbdm=@c7u4kxcB}cENj^Oj
z%d@&SyY70%E?mZnMjFwm)6$DlJkp3~aSX2!!!!MHyha?w1$S;27uPoB{;M1Ot$P16
zj~1UncmoA-Vi0?zAIXW46#v$eeXnJ3eQ!+s|Ch0dm&1;={Ywg}z<JO(5BiYk<JLg)
zb>l`GgnjeNRqFpAot72HjPP<fBj-*%zWrdizqN7xMy$uNz%UkAQ#XADDqMkDB*YM<
zJgATdwaCQV5*dX;s70)rD4i4R&Izt0HM(zG|K#k183*wD>Uo!XdMNzx<@SlOcY6Vx
z1>OP9cfc5G3+@Bw`z(IUIX~t)s|SyoBF4=-ymQ;XsslOR$JVF);}*JjzaYO~X!%o=
z>d=rndPelja+hDp_+In>RXwU4Z+U!5)B)f8sz2XwVTPJczj1g*A-p_<tuh|w<%cOI
zX(pCg``12HW;cM{4K~C00`0!g41;`Ify)-{-vH4Y4<Nq-tgvg*kXkBiY~lGyXXpW3
zop;OV!}E_cW3y2Xy~<&yPr!Q>dR4Hv3J$4)%y%lMg$bvHDMa(g*B@PQS>W$4kLZY+
z!L(-Z9SMf<^22d;><2=1oL?Qsd;qi``R!+C@EQ$y&CZ}64XLNNF(T`-ZK_GnZ(b`J
zYpZs?%&Us~KQGdMkiV|qykoS(e?38Nx~gpclis_%|GIOV*XM80YQ3WRUmrnpu+RCP
zzomct=BjR=T2m6WYZvauqPc{OE+OVbrT`fgQ2aSCZq%3ik3M0GohcY(3btfJeZIfZ
z;dWzM4(@oun?2z>Gl<sK*vqz0!<zmsv)XOtd`J+}R%W$}VR04eTSddXrQiQJvn`Se
zTsk}OXxoKZ+n@HIp5#I%v!s{3=q3JUko?$?3z^K0UfR-2tIkU({al}sZ)&`xNVMx|
ZW^Ucv!m^jGjg_o-CqrgUFU{%Y{{ZCgcI^NF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp.F83027A680F1BF3E.idx b/.cache/clangd/index/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp.F83027A680F1BF3E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..586b61c139e24ad183851fdef6d71c30ab3e2450
GIT binary patch
literal 19936
zcmeI4c~BKc*YN3{_2RII7cR1RaREUDLC^sNL~uh?P*B`=jYiS9BuYY*O+iq=sJNl1
zL`aN?;1+@gL{tzFL<ALyfv~uN$YR`Lz5(gm?f1u*N<ClIo9C@qm6`j?>6z)CIeq$^
z)7?z=96kE(1T{5}MV=c17O$I6tEs7Vm;PJ3Zp|`xMon$CLQQQ$LXrQ>RrklrU#{J>
zS^4JF>?5CP$(~HjowvAq$VldM|KrDR<+;x5R=Kcf+wGF$e|~+f*TUHC6Aovcc%z-u
z@b+ecUBi>5&mV4FrS;l6T<z`Q*ln`<4W_*|*|(j4I81%PbmqGQ9={#yky_D9-aIG9
zdm4D$5IJ4#r0()$&Xs=mL*cKz>8MPv&~W>F&(T(oN3UMFpuBWt#O-;Su1B)&#tL>}
z`4vUOVx8xw!HIsAR#)x~dvd?WJiYx%`$q1LGwAu#-6Z27L!(;UbH^<$@;D=|qYUS0
zS15LD;(i|f)wieEm5yF9;^1!@H;x|cydYpt(|P%iH-lgIEunYim<&J1KUkD=f`70i
zXKDHS@QdcD`=+kUGFhb<|7XBglOI^<-^{B@uv!FPcN?C<S#&q|`1<Uwh5r6>ZiU9W
zI-i`Eccb;WFSRah876FBV%6Qh`l8>@uAYG-H?O0#%Gd$z?B`#4eY*7=xGG=M>n7FI
z|CGa#9u<3TRDM}sn%gV6rhf9Mp^JXlGJO7pyyBA&0<;qO7i~K(99dA^*XGxS?%y@C
zdoM4wTKV4go3jracMYtnyb)9_FE%|<miFRH9brfK=2zogo*eu(C@y?u_QbS4vZ@Q+
zG=IF(Zyncu<9fdbWzCk~#La#^)$EdkrR?I=f>Kx6n(=KdD=T9^IToAeKVVnwO%OQA
zyGd`$^6L0wWjA|hFW$*om8RvIJg)m=;7zXyx?!`UW_e~0ztF$mtQnejcW{2ihc*V4
z{NjJ&+?F|K@5q-09<lKGxln#&`sZuzluv5DUU%$`&&BfwZ;DS0yff(Y>EFNqZu#-g
z=6}w7zsLG|to1RYo8^<NuV43Yto5-UCU2~WS@R~O-|Qh*lDffTR`U1(@4nx?X8(St
zYo88@>M=fZQ{ywIT#NR}#;3nrX5ZFYKI(Pbib>yXUVk;pVM40ech|~8C%gBj+s{t%
zxpra>T;}85>(9E|r!OT^W5&Giox0|(;*(q12CY$7?6@G0w4RMh`&-MtihosAF4RpQ
zImK{cag=lS^TTZqEEM+^%-c$xNvF!E3hEVKyq<U<$6!paVs2si<*&>Ww@)`+;uEy=
z+N{hzcOA-FmU&j${9N!uy62Q$TMZ-QKCQp^W_^8-Yfw_Oc0~7zh&PW;KRS6&F=|23
zoL&8PP8~lZXxXEkFW*<(xwCs-T%=c&MV-5S`JW}WeX3U_t$*|Ujqkg{4gG7A<7N)c
zD_*$T$I^A6V|~?u?hAKlrTa%syB1=#%V^SKr}b{1tX>wlq5t?#*XPuQ`<y)4Z(8Q2
zg!oU-t)6w+e7QNZGs3PQU`VCIN!zL2-d+&kT&KOER^s`)hrQjD<yT|E^!9kXgWAu_
z#)t#e<Bv47+?%!Qh+))u-L?f^<u54J2)JLG91w8-R`TqC`?r(74)`*40(akN)sa~(
zSM-ywq-k&0vDFxTUwuIOT}|1Vjb62%zC9Od+HihLWbGmb^!4kp*6#iCE6RX?x*x~K
z%=A2S?(+PW)t|0ozP4&|STXc~=8GebZB{I-cNINabj@?Y%P(#$o)G=2_xx7{)EUMg
zxMCeu7`<ZYj7l@_Ssq_}Y5L+&dE3I}<_F^oKI>ODMsq{Wgr@7V%Z1adw&wr2>`9+J
zecvh`)|+kH@4Y&v?vsR>XSx?`Kke!MMc`>?3%fH%`l(O4zVOiU5&P%P|1EP&W%ic7
z71>dJ!@Pb1Eq@;WLmU^?Cuvq`%jj*2o=l!B<Hu8>K}K^P)S4MD&uH^VALA^lwFTW;
z;XiH3i?O3-)~0yq-rM~Bm!7K@-8nXQ>zLigephB{Z!L>jU%0n+r`rvO6)B(mNnN=g
z8@e_yU;!stQgF#2+~LOWSvs)qKs7NM($UbV>!ij@7FYTQhyS|sUFoLdy3kB2GTwN6
z|IPyvQ=hbs8+!TBZXL)=;-ip03cGimP?2u@w>KS2w|?IZ>PbcBs?Rm(JRq^=t-)Vb
zlnJG}&_ELV82A`<9+0>v|4z3@_j7BzK|vC~gX-^~bJs~#Bnuu<zdh)VE@Y%4GYw{%
zb{>$};M2{k4!)Y0qzhR|Jl$ZrapwVv`^>9s6@51k)P=^9*w?_<sPllt!`2?qI#1_S
zb%Pp`IE+#cqZC~yRK&%i;rRa9nufZtmsG^hz|W}jpc1c--PvSue7+7em&DnWVK$}g
zI-w$Ng_eV76*=$JfhJOs96;xQfn6t5WR(5DKc9^pZleQDr6TRXx*aU+I-w#y>DQb}
zn|ki)1_ww*W~<Nc*?B<XIe~dImt{=Q(1Cp<@g7L;fyP}YRAlatBOh5_JYl5+`%6W(
zQifZpUR@_tWPaGAbL)24Kh%N!q#}EP!Cqk5by5}C`o?q>yYOu{s4f-x9jgBh`*odA
zkx%+fIeLVu`a%cxkctFZ2hHs~An~Ub&RN^9GRZ7dSgMKRxN(Cz4{9<9v?Q$+LGU}*
zXMavS+fx_P()H5}rWto0(DmW7%QNCbVqA2fToOkDIuf{bolud8zT-xn+t=!(1G`H_
zmRT>G+j&rl*M+B9?>MZc3n@uF!(fJK=K+aNu1sxi$Y?0gf!dOIzV-Z>od+a76YpQ$
zI(-VG1A9y2`+&YLea{akRgsI@YH77|ophm*RAjQjWTVal5@$bxJtiekV|AgvB%Ww6
z(Wvu)#Fx#7cK_mveX0)RBykh8Zh~XGPN>NBHM#~y^!ewyP$?CeVlc(1^MJ&6;*`<v
z{Z?)424#}C9jdoWAL+vh6}kVW?&aAHky6Jol!{C;m}Jy>K;mBmPoM3X<ldkIbtLgR
z>vc0b4@lf%zUdP-mHw_9)Re@pq55k$tm~vIlK8Foi*0(gF6=24@iy=_>O3Iv?|O6m
zmqZ%$(Sa6{xRo_*WtCkgRS~lZUpLL%IjtM)CKU;$)PtqI_2GnyygDY|bSQh-PdZRn
zDv|_jlfd|{6Dp#%X7Q4>neNygLiH{yh!jx}|9~ccz^TX@A*-ma-f0bSvBRLpVPK3p
zi6DLhdwc_pk)0L9T&R}|O_0?T#7e41C1vc=Wg-Oe6{CE`SYQrjmBB2*TUq5+l8>+!
zM_4;t{vm7jkhMXKo&(goY%;1Sn>ED5dDO{^oFEng^FrW(ERZ0U0Id=*0COEMuLB-P
zqXqFP(0U36V15rg-h(;FJPBeF<DSIKMl}$`8@&4sem44+XfUAEI|l(&A3?kWg*(s)
zRag*PpjHc1B0blziQNC`Zre&zL7b?ZXo`A=AbKmkO<TIuNDzILKBj0q5yVN#Nv24L
z1#z-+vgw(wPA7=IN?%iC`32EW>1T?pxFAkZPBBHjM-Zo(PqRkmMi8f)Pq#(`iXhG~
zpJ5%*Whn%4ruj_ko?T;+ATI8=cmOgJg1AJn#28b6d4M%CK!Ui^bfqPxRf<){m^Sp=
zFaUKlK@3y`8e<9u=E1-knNmRv0j43q5_2d}gaTvCalkMR^hUN((oJCa4atXra2V)d
z-3^8#g5enaWkcYI5R$)v!Z%O{>vjTVkU*KBF{DGcg-S{XYr=Po{&!3-%xR2%8p*F1
z{Z}Lhv--g-QGP3{zm?=$y#6h|7t$<2yvys~#jJ7myKHmjvx^#nIG>&`eDG$})C6%E
zy^NrB^g4oq=pce30387cP6Bii$=QI;Cixzq?~&XN=ysB$ARPsX`^7^#p5!BtK0<O6
zq?<4!#nw^pbdYe@IRS^AfG)V}q{3mTB-c?cb<}YD-W#aN4U`||R?4ZBa>l%y@!ic#
z!TgMId&an9e$Kc(Cpm(h5y8&F@)XuTg`JK$hI5JGhU4!P$4!po{4ghQP6?bd=AU`r
zpZO`6cgx&%%iJ;VmAUODIYTx+LpBk~we-u}ZEpXvoTkN0FeDQUMFRjW{zeb_jaC^)
zTC8RVS7YNzi`)1?+jtvvBmIhW{$0xX!98hliOUjK)O%@hW#5(kP!E;JYvp)Mt6WyO
zVhVH#bVdD@milMk5YP|x#}4)i0pl@;0+&$Wia8VbW&%H4{VZUX1^Q#Y2HdY<%T0^<
zz&syVAj>T^54cxj%T0?lz`O=nAj?gQY0y3m4ncZDi|NoI9XevJg)X(w6>}YQsUx`_
z+SNmQEN`Y<o3WnJVk_m^N^%oDq=_Di^nw=majyHYUeIC!=bAwB9)8Fkekl4JXy~QY
zyWD@=0d`WRJ1Ku`cnT=90%{t1kRZNg&EB%pR5#LcE^tfExk3rzern`?$`@Z)LyfGV
zeDQ_R+{kFoS9M{RZD-Xx;|u)!4T=qY&}SCJU|<*w%+P1-Fw2S%&<Ar0?UF)|#Sbar
zTuQjH=pkruk*jw`Da1n@H#p;m1P%{$Cun5g7)-&yB^bEj2Zn$NA;1T7C>S0J+%bm%
zMHn!^d>jlt4jhr536dU7ss?^YZv;t?2Gjs6%#C17BN&JIB{lRV<%k=YLQhGdiO-Qw
zo95HzSl&QSYM=>wl*kQD<Q&lpkvtM^N(ra>e1iCbGkw9CV|g4uDUMg!6XT{OCL5>D
zIRykULFSYoQ@I|!DyCn2o}E#^2;y;IbsX5>7QUpcUQ#N<Ac*<2Q9iB2HB98J5;>J&
z=%|6w3r>k@pu8Dl-t5wN0t#XnJ2;FbOl1sf9K%|n8}%7}CbPw%MEzfj(wQ@$#s680
zg8V<p7wri9aaY^R`0gbKz)QUEB|IQm4URq=qQjd=3O3tlwk3+Xtroo-W(>d1Rj0))
zsFek~W5bvQhh#w!^JQpq8TP@P2Tk%wE`o+d&=|{$pluPfBY8{_^u(M%^-rKIkQt-J
zT-G9&wM1G+i$yYvBAF#pS8MM@V+Q}QUdr}J7GEPBYT+B5g?4B0h+tcP=%nGiE33NE
z;x*uX4fv2li+45fL0z5}--2FmffBijwD=D6dPnkY*ylFviyu@^d)L!GxP}e1cLT{y
z^t>i|K9<Mw-m$z7Y5^_A@!oMH@8Rd|;pZdy;IXw)Y|rUKDXHmDs|af22NyxpB4|eP
zup;P;*uivx+I+3KpG&W%02mbjW8C83fYEQj7`52JeUJO;H@&s~>t2nc-ImF}T>hb*
z7Q`Z;C<22~3kC5uP~0ZD0Vo>4VANtkYyyfV%)?eiesb(_<Z8*iE~e~@sgdYo31TT_
zUrKTXWm!QD!}2Q1x{7ke{E!~>kS6@o8hTI-$%)LcM8*r359aKHxsm8&OGcNo4<mUC
zXT61U#qvmgP$W<IpV9oFXp-~!VfnllDvuI=l2!PS?Zb8<no>qn15rN{#7mUPB}!%8
zoi$zZ=8o<cDA|P=AjE({DAGcarF6%R<RHG^KE`(+X=ieI-&|~GP=lpm<q#ztqO{S2
z)w{BX=tgd#J=jMtk6K6$O(xLI1p0Vj%cphnX<c-uj)6<(1*eOeioU&+?)Z>%M0XU#
zm(b)T9E*}fBcckOnjWq_*!i(diYX%DP)eDUl3YP4DyRYYiB*(A6@Ev8_>eYyNLyoh
z4Q*IMaw5||k#WNEV9q3%vqLrR_}C_4ByZsiws2Nh9?2U<^46H6dBbRu^ZEYyyc3d-
zoNV}Q&`W-#L5DABx1aJr?^yD3D7zZU16Sk#t$lzt!F-6;K1A|QwDwOVU!=7!l3Y*Q
z*3%>K^$oOb1Ibqz`BkPLmX|Q{5|SS=@<$|xv&wLG5H26V_Km>7j3B14CMh@*B#6<R
zT{P!`8ZL-=oOT{(g83S!eU0SbIPKp^uIIGtNsi@hWBC#I`Z(S;j^yXO{5jtb%U|>I
z*Cg+d$#=*I`6-$1DVZrQpCQxDz>J0#sSka~ePdOPgNisf9;YGRK$ACc3Q9kYxt%cV
zz>HCAU8IbRuvia8aZ*M`SVB;M5TJ+RHYp<`tR!fIut5)d*MUMHLBT)>271_c4go?4
z$+18?78oE)&|!3iBA|ylo*>=?!cCHkfly3xDG*9OX6sVmfUhqDLK!K)1B5#yHvpjl
z=wT!N2ndfzehh@iBsT(~@ng1b1P=K6=RkN)%A0}EOmZ+3f}tMzT2e3st%IQh<`8Hd
zLUJe+Ldo)BPzWQrj?${5dSWfFr?l!xKE`MtV+^o7g|SRwhG0I)Se_&~gVD|)%b#Yn
zPm`R($aBaXR1RaG!w_TDB}RUUl;<+?T#~Od^6MY7{dMyCLPlOl%5O088zhIb@^CW8
z6~kJ`unxFADXeY^OGG9yoK_6i6Ze`}PAit=o4oc--T=!>c*_!g2<8gjvV!DFUb~Vk
z|A5ziK=NZ={ut-j1o1I%|ClGn+9$mH2`PWd%b$|m%FA0nX8Tt1`ZiwPM#|fHc{^s$
z(W`q;-wQL?j&WqfLz-|pYUmL)BquTx5*d|;DTtB$h)AAr8lw3T(In^d6Y{Y~Ic{He
z?9H%e8q!!E%gSR}l?F}_zgeYcwJlCl5Vz2NTj;s?(<IXNiKN?djP^T5%73Qqe<tO3
z8UMS?0(^Zl<Iv0y15O&}m&VP-pDLfT&&TNz$#igjH%WN|XWu}|LuCFTvIY41eKLoA
zGGYKYBy%`K@@bjlX_@K`P4E^K`J?tul3sNL?G-^!#1&1Zy^=|;WhT}#ld$|0=XHux
z<>Ms(kn<`gdAn@lcG)CUF>f#VxAS!2-oK2VQJ}@CcTQ{K0U(^V2`67{5^a-2@^)6f
zo$ZV3m&)0sk}v88XLEz(gEIL+SzlDYsR8NYc+I2!QvJ#p?J{OCzR@4N_8<IURB&34
zYqs-F>YGZg^&+iBmUyIDp|yg@@2}TdZ;3~fO<J2&8T)B(_uAHHh1vhdnaf$u2YMW3
z9<BVZi%w?mf7UogyZxy2uCIX}*MJ!w=&C`FYG8&2A~csRd9EKOvavRVP$nUi<L_M)
zgL5*LH%~0QTK)e$>4O|aX=?4mJUOZXnm>`M_F;A$2_)Z@>KjJcqX1)m*EG3whYxe)
zs37ttBwwdvW*q%*X)vXQVA=)i-}jvGJ?Db-Z&}}5xrv5miuCeopiT|!g`(OHHm|{%
zT69C{!5vfUrw&<WFaBxAxfQu776%Tt&-}~a_cMp*t9@^;Q*>n5fNcVBI~HpbXLI1O
zOHkvl<I76N>Pa;ypxp~-Z(P+D&b@{6MpZ@gX;PB2W9}YRNOI8}7dzvQvt;-ZcY;PP
z8G|XnCBThv*e0y>!L(}lDtAnuDLym6bR5_o2lhBZx&|g)1AbV~^1*<7V1>C9j41`<
zFu$a1UsCqCVUhHd$d8koCiygBZ0hMr^&clUZ4)_r^dbZ?otu*WadOk-1*bAT((uDi
zip9yzutCc=Dj%J*l3slxD^Fxq@dENNrR?~J<qr6ro&mxcFc!D?AftVd8IN0VmDj$?
zk4H5^Z%R_WjztmpHg756Ek!8bx18`TR=$W&v|NV&F^-ldl-n428>7JYf5*t*k-UR7
z-oX-qR5GiZ%<ALv<*a@=Ylyjm)vqAAo;9k+E)bf2m6Gxut2A(Z7b_O`A=Je%zzj7_
z^6wQZ`(TQsT_VYZd^+cnjuZ0HrXze}?=On^+tGFOI4Zm`_D=_;Adc4^Z-kZ%2x1h_
zi2^;*7n340pc74U60k@jv(s0B=2f7JRWKiD=9By@(EJtXBGcHBpVzD*xe0nUK_4`g
zE{M%gvl;4QZlp9DDI!VmoYH(wato!|Lh?0Q=Nhf@hy<~b)@&q`3=xc81VgN8h-CC4
zNzP()vKZp}3yjVMl6SHeJITm3hc(Dydt-%&=QQJS5<?L8a+-TdKEP=nAi0^-X(oN*
zcwRG}S0zCN@f5FjiZ@0bK<bjbUOLIOyiP4o_}F#4P94dGGJ`@{Z*+b1lab|1m#lgB
zx08?CS~4#^_fXdTWnMNWT4(T~q*dnxT8yL}BWYD?jTTdA#}tw)8RtqIe$e6r#`yus
zzc9|fko+s-{42@noMSpS3^k3GreYjRNDh}dhs#v%KP_&SId3L8LgpMn@>ZGiR?Is}
zA2o+$eRc9LN1XU2xecy8`%sc=b(NK0W&2=dC}HI#B;S(BZ^`<g8cT~1Wb#VP`(wRF
z`+HoOD2-5sbgx3%AK$2r>($2jqYD!bTGj;RelGsUzSzOnX`AmVqK%~|76GRsFd6sb
z2H?~HCZm3PbYpbv4NY;>U%rg*df#kiCcKtOW58k7^f0UX{NMYRx~iXFZpBL}SfD5+
zN3lkrXe6iXf}tW9T9IM7B821=MxMeDS-g{s{3OXGyu5@TfL@j$R`Bu)%qKQvF+KrD
zM^EU;Zzyg8C2sRgV0IH&V=e}YVp3iV%!)~Q8Bmmw@-kpnM#}F1#T`<92MoLe?D72{
z0mUPrMDJJ-9|5yRz#8*opm<Em9|N<;r2IKhJSXMPf!T9X-V7AYq`VmnYzFrD{-ID2
z3RM#$f*1<TLZLN&@nKLAM#{sWMHsZj@<`eyk|yH+INBzT<P_Q_h2%8aCXM6_#xR2+
z+>s1sKn5AYWH62y43U-0V8&*UVe=VAeugo^z2pqj=M1R`XBd|=%qUXsb%r5k&@VB{
zOH6+(&t;UkBo{KuLXvMV${Qr#W9;rRDh-$Bq8Yntl6SK5oh(tmcvc=y@&Q(UfaC;L
zo<MRUYnq5BXr<L_tUQU7pJ2^Sutd5#l{HHx`6O$0lH^R*ER*D)ShJr<&SK57NKWT$
z(m4m*kuo`(Op;4Dn-Y@mayEBKuH+3XdBR<+<OfvZkW?~%ykjMqm96B*R^l{=AXf46
zD&7b`zl!fug*$SG+@*>iMasRZc-4vsL43k1pYZ*${3)+|N^%>oY$LgySGJQJDzgif
zsRkIyH<H;!k$hMtKP)5acT6TfM)Gl){5Z+yWoGAPMEd_nnc0sdXUfbnNzRj*<&k_v
zW_E?-t1`2zm{SiJ&KaGgwO%s4+t?x7*r`adk)tWO-yMsAu@{;QY?8rj?0mnbY+h5u
z>d2#<%~4LZMB(Ivgv*ERbCyfvTsEVf%@9ix3K;DIk{|KfkMLS0X}u(`-GUhzASnp_
zu<RJu=oK`11;?V`6Rn73)H|0S<6?Ul-#z5qR1@RYL@xTu;eB)PT$Ld1k-6=`i?Yt9
z6oAJ|r-e(SUMn-Om9!9V8OyhfDsgxA>6d2}Q7y;+T@vw+SQ7E$l3#C(RKH%WEiIbV
zm@n<K`|wX%G^w$SpmiGS2nx~&A}9iAL;!-5fJPF@*+3(k<a<El9?9)Mqn+d^s1XH;
z`^7_zc#@AmjUyyCL5(Jow^H<0N)xvvo1(KxZe{6KRwe)Gvq0Ge#g8{6#eGT(Pw9cE
zhYR8^Uf9JCM5Zpwxp(*4$d}utu9^$^T&RPyEqSbV9!tzgJY@ABk}hhoOuJa7%E@0?
zrMITRVE7>EiS3NMogpIQKN$HRm@nktnVveVweH`p@{i~$XbrltdgsmzY&kCj`^#V`
zc0sFveH9prEV8u3AK1PEcG$%`3kRKrwpe}6!Wn1bEX-MOU>3B-d>M|o3_UQ{Q<Lk-
z<-Cp5;6^e-wUhDPNm}w`W<WAS=yfVHAeH1yW_Tv)5@s?UnWO>FWX5DNp15CRGG3X?
z1k72CXBN3!qk%PQAg8oqxyiBQ^51xFa6Fk2yTJQiAV-#Ben2tqcY;{X4=5+Oh96!-
zx;QnwM-Az+)bL|!cu&-CrSu5zRl`rfT+4gb;^iMXiD~T8Q0|!29Sdl^0-6XlTR6QI
zP8C|__TOkLo*8%SFXOdbx22cgXMKK3y5s+Qc_7NjNO{>0J9>~#T^Ufe@M+%JZIZ1?
zX5`7_{*Y8go=S2tFE8dTaO28(c{%25rMC(TpS0!t$5AcX9wB83J9ZYKib|^}7~lP*
zvR&qVFXNASz4n`5JYIRaNDT?6g>af!zK}!<NhGIoLMk^1H{=E<+`xSO)rGQG!=DZO
z+ex`%XS1Nm-bV5N8qt4*5xrIZxxb~!)lJeYV*==t0Bq1yLdOau{w35yE083!3Heu0
z5BaTv_y+QCpdMQ0(Xl~`{{!lwRpQdJYl;t{^w2!eok7nkPfwY*Q_{TsKx;p_MJons
z#gJRH?ohq&P=sb~X62jl@Qo@Yb<z*J<&f@I)J0BBzHI@dC!7FICx8!j`qO|@8p#i6
zrw4eoqabePoi^j3?C$0YefD#M^1sXj-fc^mymW6=#DCnI(IzRGdglf|)FMIJAuuD2
z9E#H6oOC!34?Ruv%qE(MX5ty&c(V65GH#7z(6EP}xd-nBlY$)HHyc|AL5!EV#p592
z7xQhtH~cJ?{q0VUrdOqCtYcduZchrZNg<=L_ms_hG8#+aY*KJER_##i@%ssV@UN}-
zN7#y*jX%xnu0LdQM+al&Cs<SLa%8gdOp?#b<md6)F=?%vOrD4Nx6xCnb2fg;zZ{T%
z+v)e&i%oBzN#XuJsI?CcLDSAp$G-C3yKLV}X|}7F8eL2gvuUN&=u(m^D3=Oy8DSOW
zRz;1*RS)Jy2Xm^~w~qCVqr*tv!ntkXMx$}CG03&2uYcWIsYY{+=h~y)l*kH73jblF
zEmry6l<972Dw?g9(%zI=Aw{IU-?3)zSXJ7)B_T6=(5c@2{#E&Zgvz6xt=8(D8>8`8
zUE;aK3+0>yF+dT}6HjNZR;(r{&@<2r*~5;#Ns3_56D=^1yfB~$Cpi`v#sXtJHMs*A
z?jSi2*yoY4%`d?I7c#ba4QyYNu}vDZNrQv&fRPUE)8SCeS#WX|^uwG7N92*QO&w)X
zN2zu)NK==TQ!}|$@#Fa~Y0>LP^Iy`Mb=tOxCRQu&VtjXz1Jnh^^8y*Gd^}$!{9p5B
zg81=#nIL{VUnZqP8Mi0ofcKPfdrEQy>k+{c>GBlTErr}KRL$B~v#Kp}(n501B8F4#
z+>p!==d_R9D*5sJT*p=^a(+(A%JH^)cw+VRMc(%!IpEguo^?3(k@VrC`8>%p`Di{*
zS_<~je4ZeFJfA0sJ7sP=@c=I6Ok{4mNY0QA$&iW2nzcNsj;gy-o$|kKS$V7W<7Z)8
zz3u<@9B^B%hS_R^=W&1A<$3w_)s1>@FKPYlC%)X``D4H98SUNv+cx<iC6Y$*laz3h
z+#d1aPemYWi9!I$WB>4%Ayk65X(tp<0y<K7IgB!gA?(C;MtPm&$Gq|}*^gRzWh-Xn
z=}S-e@b@Lq6W+c}&K~q2zO8c+rA-7)WKEK3n`Dw}8M|6?Rn95S<`hR{Es8mtVv@JZ
z?6%8PeuMNEELP0#$Z@V+vvBPJ=?ujm%;TRfSVpG_j#k!#h77f_9c<@duV#%Bp=xSF
j@lPxK(-!}<!#@Y(pEmfXBmU`tf7;`pL-5Z*_~-ut(bW|Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_waveletmodel.hpp.C9F027882C9D4F2B.idx b/.cache/clangd/index/gridwise_gemm_waveletmodel.hpp.C9F027882C9D4F2B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a7c0eeae3a78a2c09c8ff02829658dc371c90618
GIT binary patch
literal 1354
zcmWIYbaQiJWngel@vO*AElFfyU|<jg;^LB`OiM-vhEGfk3>9<M273E584BE+zlpss
z;o#$A6(1Yi3|u^H%pRBBFxxVB8UF-<f4gso`mh{1QZe(e`Dfd-WjkKJbql>?wIyLg
zj7G08)5ZyMjy<-j+uu*$d$)Ayngix?r<bQ3%Q*A+%wwTs-RFW*ckYWVw+im~#BY_Z
z^yT!$^v52b7Z+;Q?%$J9amMcA|JGXeU4BwZ_nc=mZ4y=r@7XIPEovZKG)Fw*vAMzS
zk_XN+)SsVl7dcwIsr}*1q7_q*_Q<B~-%u4A8f%$;IoG)0!R=P1Uk3{e&X`*rws`UH
zOvJYDD<qO9@fb-uI0R%r%V?h0%(|CXYs0Lb#SH7Eesb+~Jv(J>snCg6qTLzHm41A-
zF0Ru|^)@c!y*PDFi+QfdzRco`o;H?O+Ar_eM=W2lrR%m$%88<zbt1LP|MGnZ6Mn={
zT$!6>1q?yv`PcSu{&&%jjfsO77_dMfz{Sob&A`bDlw@EOU}TAS_3gqF{wG3ATzm`y
z0(t_LV1j`IteC-=Ph!z@dj}IXCQe=!flT&HX)pnkc9&tBI$P27I8Zk~gFud6jwP6Y
zNyk@*&n*l3YcIsa4K!RsT0;*^z@!^D-?icwQZp7};sNTekgm`J6ENvn1wk<uT^R~O
zOzc4EJiR<8FaeW(E-AcqGE3@tkkcTJ6-Z-G1A2`Om})>?d#C?9=}lQD6Ugt(0(|Uz
z(qICn|F`JBMvf21P6LexI!;Va%n3}uq>ED1itB(T!9tLU6C83I=4tOw-8FF=55z2p
z#SE}eWaI*e9?WLmtt<9xuomoP;SmsE5#WI;X5bMJV-Z8*<*?)+@$y*m;5=CH1I-16
zJ<M`gumgobVGdIT3vQq=D6C<wfCV#97!<}Z)!kmqEpPk(TW|m!XC>eW^8(Ou&SGA$
zz~K=v(Kj)J8O$T#!s8;0;N@85IKte*Bar8u=LL&19)TMD8Z(&9JOV9@JS~jEF#fc(
zL(lCbm6+Lh1Uv;i<q!r23j`za;(6kw5ON)i0v(KUFe7*bdKrNTi9dyrX9}YfOdjSC
zpr=782j=M=X%Ei0aGiPxv{`^xKpZJXc*T%-8r&M92)P{I93;6s-aI5;1$PBpB`ie(
zlL{y$!eWaDl4YRg!D0pIFi?_&DS`z!P#BaHVXBH!a}q%si;I$r!M1{mM+jgNsGZ2<
M-MZ5>j8TvQ07Fxb2><{9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_wmma.hpp.4F8B2CDD1348D704.idx b/.cache/clangd/index/gridwise_gemm_wmma.hpp.4F8B2CDD1348D704.idx
new file mode 100755
index 0000000000000000000000000000000000000000..32516562155502353b796bc8e12b486e162fcc61
GIT binary patch
literal 20164
zcmeI4XH->3)UcUy&NX-tM4F&niikq!awr!OL=cQ(!*0|VYwQ{mjlBUD#Mq4j0*aBQ
zA}CleLBuF3V8Mo{fPknN0hQQz2Vt0Dy+6K{AK!Xc*3yM{c-(u=nKO6Ko|!%SI3&Q&
z?@XYYn$M(wMYAUV_#IYLQ)^HDF8Fc&^jJYnEyzesZPC7~!^h0I?=BS%e0ARV#F|HC
z_YP<Tg8`%K@TioD-B0xO{&mlt-)D8)T9jt==3)5Ef6N|^{&`HhYu^mrru5x#;4jUy
zD{8-Y{TUe<A4}CUj(w|otNQSK)STs;%4@o>U0FZj+i~p!s&9^nG0@-O8{pvZbvK*C
z*VJ##ujmu9Mq%;sX8tan<NsLQS(#Tq^y^PL(DM4^KNo8KQ&^Mp{gM*r-#gq`V7||&
z`rZhWX))>cwu6R93O|EA;+(&S&j`s`b=$aJ>#sS<D^qSq|9a9_b=kiprL*&;8OJk(
z$eVqwug$-3B5`;5hJ~XXtUJHH^>+Qvs9k|pW%?Jq&)a2hydGzub8%^~R_M)=nDr;(
z?mUm~vQrfZ=cz5Pm}~Op`o^{RQ~TAiKjLwkLCLp$X1P~S`sLD&x;9bi8e{JdxcB_0
zeeucIRGRy1j^UfOgD%zYcs8RfwR1*nw)t3#fqSQYNEB9BOcJ6V)-TPRdgZw7Ye7R1
z<-KyOW6mrW6Kz|~s*))42WGJwZ&taREYbA5*ZW+7w^iUi>!+_{%O=Gx^i$fo51YLA
zf~)aD$E0DCJ!T)D@%!cU%6En*#_o7EN|kchs&snW%(GUddjiW!h7Id@C~a&|{8qgg
zgX1?u+HNbEJh_A8vhS?og7&XEZ#eh9)ah2p?5Bl0de70exMrzQ9M|ihw`0fKMZ1QM
zzx4fgAA!aFuHoH{kG{Hc8a!<H7<PQ-eU~S9d(GPW@py1S-UIv0-Y2v7{d~8be#Fl6
z`)|y*kDAx%!nDN|@o}c#oe0@F?tMt8#=JFYv)yvCqLjs5--o1l{bg2AINo4^*4EQO
zqfNWk|Gst0EYq-WQzyI?-^|RM={*=ZUp+H*&xzDtGb<7n?R2|p`rfL{W8Ilk<MeDc
zXMOS<{LU)l-8qwwm+)QB{ywv<Le-V22FLBkY@hxitbUnBuPe4ULJv8A=%zXs^vgFF
zJ+p$Wz?)%CXMA4Hz(=Ou8vmQ=m>z#gepCGuCS42CO&K3OH||+t=9Ey46z$1QFZK@k
zA@cR=%wt`8Ciq=dFF0X4`nGS!_SW}zSako#aZ_EQ%S^{V?FJ>y_A7Uvlh)0y#_Y^b
z6JJYlDShs337c_x>e<bHlk1lbo4VR@q2v35PSdNV+}!%GB6P^Fzy7qLPKt9r*Lm9D
z-y*NA?RI`d?HE&UUn!~XP`ktJqk;?S#%w=Sonmy?A#!TjMLcy`?qjvB?y7*zR#z5v
zc=<c}sSMa<{!*A^w7Nq4TSyPJJohT8O_2HGck8BS&iw7~Bb|?bd;7)HzVq6)@25)g
zTjhD?MepUm4A;MR<9MW{$KCE>L6$!k^qKGYWc||t#*a2;cMX5z+izRQC$l}5@|Ld3
zIpcFA5RKLj_lq?2O%uJoQB<D2mzezGvgMD0LVIyaCv<A?jVnv*;$B9?S`IAnY`CM>
z&~f6V%v<$tyRd8V+iA{T38tgN@_*PgyhEzHPeS+6SMz@fmB(K~ZBk9&&K<S>yjHcx
zuB@%=2CBY_-g)sS-RCh+SNi6HjlpQ#__cT9-Oi0!=8$HVf6cVh%nR_=goPHT7L1L)
zz0uP=u)wf=;_2RN*TN|wYeu{1Mh_~|ONci{M+-8)IydLyjek(g$jS{acV+8RC#|zl
zdE0uWT|beVyK8v*ntoT#JbbCT5|U{iI$-Vle9gNqk9OniCECfAL(4<@XD8@9M4Ep*
zI}Ki&8NI8^v@_L;xB6+c$?H4sGW(s`cw^S~M4Oyt&sQFbEPWVg5ScT>H~Z5!BP=sI
z<;IO@F!1|$;M2P~V}q7%N_aP5*z3B$$u+h&S56%qI&@#?!`Y+1eNy^eR!Bp*SKJ|u
zF5jBpwF~^=SD(PAZpBC3HAkjQ{3fF4i9=`eQ>9tkjV4ArdNtHrFDYm^tQvQ6<hyO@
zvz{OCrPKRB{O@<Xq?%Ft<Jug%aoqVtx>L=}=!(pN#`8Y>`dq8-kFiD{+XiarFIQWz
zc-F+8#8-prReDR_-8?I)xml^n=55RkK7Xivg`t{<C6Iykj-|y_dSfEBp~6~C4o9}(
zNYz}ZwINQQ{OLIN93DBmkDV^mA}T8kS9JRPLB*yW++E8&^aK&MCAV0tSZwh5L+vXB
zNM&*hbEA}p;qyD&>q0Y9e?2f<50uRXy@oYZ4lGZujM9ciMCA^&y#u|Q3##HYZ(5eq
zxZ8uZVMn5p4Q#VP@8*K4xa4dbeJOdUhc4{)wVHew8z08*%>`9)U;W<1rsLpvU1&j6
zVz6-xc5E)FO1~dL;$u$}7i}n#8&x6OD&*f>P?i2}UnR)>ev8qC5>Z)VxTMqP4=Nu2
z^5jq((_cDjLjxj?fVL6jX_^YEGI5a2s-h8Vur72b?MT4J3Ak%>K~<*BK78AM-mvbv
zunSSyjg5C>*XDw%%$xFVLgJMJp}MdgxzRGiWt~2MQ1Qb2f}nzF0T#N@n21AxVJI+Z
zE~v^P{~dF-2gUByh26<b4r1ej<R(o8RatT;(qD73(MN6AnOtQZwq1w&HWyT7%{KdU
zD@yO#XhTDy5&~^Qpl@?QRl+^p2cajceryZXiAoqy4+DnHMT<)Q^+9{jrtHv#T}fN^
zVdH(IElmYg*<-XT^}(dyX6r%~QQ3)&cVg@2f~rJEX%D)S=KiNP)FU_A0c>{=J83GY
zN{n6cikpuPTIs?LL}fKFTn#!k7gXiou<@tAyVCwdTPP5fO;~*sRy7w?C2fM{z@=Wt
zj<<zvh)OtC568;pqD5uR-)~p+_leYn=A^mN*f^S4TT?+*&R1!=e7~S{xh~WpS6K%P
z*O9g~6;vg=bm+HU=Mt`K!>@?SDuh>|{>=qd$@|)2ap=$GU3H-bQCVuZwA1GgD$dXS
zu8;j%i|e{jmxx1vVF)pVrh=*zEM9x<>s^ETw}qxeWg}4E2s$<wRHdlH>v0{Xjp?Qh
zI}w!{glmvnb3s*#`xz~i=NeXPLnTp3!A2?AxVfMzC81+h{Npv<Ko_<rDl38EO7g%>
z1y!ju?K*SjE?Z(w`b4D`*wzBy=AuPqgQ>-lb9-0oLT#e53K*^;S7|D!%Co^)gSK>O
z*G3m=5|!nK%R7DkpyJx9sMDwC9vvk@6C&P<+it}jn+vL<Hh=Pz1=>E$ivsn}-t%ar
z^|{HXemx<cb~f?wO+Jey`4H4U1Rd$Xm*mG#|1or=ZcUQkA)j|>9PL3#z9G)KA%+6l
zUmEtIzx}Yo;l8&dFHtVBp#DXYmnxT9&@We#mnoN7&@nH`%azM5=tD~K3grq5>Zv8t
zp~{uOg7&pUx<<JQSX4ImnnbK!xf)o|fgll!Q?3IR)EFf>8kk3e?sP0QvRgFh#W)6-
z$MEt)!2D2ac00uD9|g*zz=G9329(EmJ`R+}c|Hl0CtI`ANz#Pou`S8zK$*_VGk`LK
z=X_wE54zJiLz44>TR!N;xB!?J@bW@nUf7!53VHp9K=}|@u=<aH@)6G^Kv}|bIZ&3j
zW~XxC#_Cr9Wd$#<1j<UD4@2+6urITx1ZbWByYnn3Ko6e%6JQ|YM5s!H-i&jhUoPy=
z_zLv9!t*ugca7(K=$FrPDfBLdeVIK!f#y$Ocb?@Z(1U0HCoqt4IaHNHZ^pIIuNL-a
zTnGK?czy%@-tb%x{pxuRMn=KNlzK)<4n;<xJbysmA5dR5-aaDlk32^Ro)JQy4F$co
zyW!B?k{l^`M)JI0@Z8UHtl$~T^9#ZA1<%!jXEo1%3!Z=TTq}6i@_a`$xFZ_V;V;R>
zqCqiZjkMp-bra^EZ7Y$P3-18Z=XDtQkjRP$XY*VOa4pXv5QjicKLX+ip6@_>hvx=}
z8+Z;!I2>{Hs}Qc@`4z&icwUF`I^3eKX>hZ*{f*)o2+K*>HwpXEbq<zK3*^Uz_BfXF
zMYnv>o%R4-h7u3kv<9WCN`u?S+I%>4($N~rF+ddqJg5)A@=4%u5;*eQ_ayLVoDQ7Q
zK@Y|mz$t_00-!1Y9&}7%xg0o@14o|wmIHsr6~L(i^k7^GoGN+Fg+p^;5FKM!u7yKu
zVGy-(EJq=aDAbcaDVAfAM{H{z8H+|S`9b7yke44sBM<WOIOG+FyqSJ7@<`_OQ;<hW
zYaW@x*H1+rsk}TDjZEd`X~-*$zkUg_D?#?`ex=B+l;=e3kcb`GH=cwYl6X$W&dFG2
z@*DX38+a1604(3c-{0i<9(K5g9hv+Ac6h*ZF*Yd19htls_bA5BjH4A!(TW~CtD+Si
zjAIo}vAjH1p^D|@7Ztu26@E<rqGG^BMIhr#3g1h-{E}k8B_{8v>2b*`?vD#Vl5_Fc
zTs(tTq?5El{qr`88oDH>L7z0}YgcW4Vx07Bza+neJ};r~u+4XM^`C{gNOBzVh(n&V
z|A{{GNaMK}+ZVI`k>rcwW^hs5(z3|tTs_5P`S5W^NsIO=$j>O+BI7>~#Q$(#--Zab
zBfIU$o!&$F)bhkvKRxP21WGlfI`u-d$z-rLZujVN$gEdtMvZH}`j??9e;KO(rJ+g#
z6q0&#Sb`e3Bqst@BJfJ`n7KIBdFfI~P6BpGpa<joKy@E@t*Y@{-`6o^v?Lb+yCTqo
z@h71A1iY-()0bMiU+g8xTcK(z^g20ze@XVRb6p4vs;z<-ofRc{n_#z1=)t%|P?hlV
zQo*j2mxqa7VWKY`$laE?MulhEkARZAN6^{BD?Sx;o-(GB5>#)Fval!KAzJQW^RkWO
z#<W%1ZA^#<*a$jp1lDw)Xfx@9sxw@c)drJ3>$)Cvqn(K5vq<wSYEOMgBRii(E{rpg
zW+pGsMMk-(Gn3~chg{^ubKhJP!1$`d;;O=uKC@l1tN*O(;`QX2w+rUm1!vZIWrBGb
zV~6s{WTOce<``gkFL2un-08Q7$+zs5$a4~KP69HMCjqx4UVa(4T?X#-TgLKz;C7$q
zBH&yEWF{{HZbiJj9?12;g<1ucKLNK-z@70XC~tx;)Jib18n>;`o$+;Kb{(0s9(jw5
z-y$Vl?P9qe8Q1gt6gxk~GVK_0J;80O;Lezs+&0!xm{=kC;nr#yf^&wz)h`j;O8EN3
zBun}FrGi|_*N+$d<HZ2_jXC<4Z4f(-lF1@57is1qE%qR}$RZbY<yn=BJSfwp7gBEy
z{4uMIz;Xoc7v3B}lH^j{wG_9kE}fstcV}Mudl0c=Q*5eB=P?>DB=cBPu#m<CTzl{8
zef*uN<`+}>KQ<LNr{E#Q=M!F&!JP<|iO`*{t|hq~D$Aieb!6^ZSrKV_e*Q#4F@GTG
z52Q_PhAe}Sbdl$)NV>{%A(9Fi(*QBCiKgf<wF%Xn(5VNXhR$mv$yvBZ7VbmMne54M
zk8GZEux$?ZWb!KPQic1{;nwJ<$d4c6Q-aGWfs>aBav3kr6IFTQH}rZQ?RsX9_v^A4
zk%cAH&*l517V6isK9S_LNLq{R=q|aJyUBU`h4+sBKRu<_>JDP%o0#roxFlljM665K
z?A{x4-K!rih!7k5z$6Ph(Ka>qfk`&cIoL1<+c0?*Hm_nXNRm$p=BJoTk>otlGLKC=
zeTEnvSo)?-A1x9&1N!fPB@GFZ$QIOJ2|3;j^*2LH8i6LA3-!12oQuE7#ba1Qf(5N$
z!Gv}GdO>SF&+&pzywHWo3k97*p6?1ecX@s%=)B{(Ueu}=P1yb3i(2n_-m1{qs_4Sx
zM-}alDoh!tDB7nm?t5dOYVT;@1&)m&ID=?lMz^yP*$x}T@O%gu9OC&X&^Zc>nf^(j
zbCTzLV2}^Ys85yT0$@<Ub0IJ&<oO}cc?gV|emT%7=lL+yISh^206h%l!_bRyJe1>k
zPJ%i~ynZs&N#;2hs&b((t6vIrO36V+^GOkL@sRv@F)o908P63^r-Ik7ggTWx*Fse-
z^kwzeA+2@DkPa3S5<>cs$dvh!Sfn4za}v@|;yDHBr||p%Ngt3Nos=c{1G4>qTp52v
zwjX)^grrZregl#kc)o@;ud(f}BySS*HwnDq2>K~J7YX`BJg*k@SBsqfEm7;1Xvhrj
zwrG7@Y}r0b@&nQO0nd*`t;f86iKtb=^9xbyMQgTy!A7(sSBqNJy!@r8^^)gWQLDB!
z+t>2->qM<OUjACtdd)b%Z=UI>-7tlC#VQ<7h5NG+d`bv7C2*W44#*P+(Fdo;eZ=86
zo%PX01r5S?QvZCoNB>!p4?wR2us7SMABR52`F(pa^eTqE*}nY+^m)PW<D-yk6x(A<
zWV7gckmqFNn#^-5a!utq16gMv8)lfNko75^&!Vr-B3mZ^6IuR=x-&kHEYI_N0ohzY
zc1&J}9O{r0Gt_H%$Tb{9`-_}i;PE%`cZ~00n|s)f@dIr0fahS*Az0+5iuIzydY%hK
zt3vT>R=-HJD&o0Z99J$*VDfTtV!7B7H<IK^F`$x7kCL3Ca7a<K+;4!();W4Mo3@ZS
z>lGgR3QuK*v4L!egkFsQ6vqB3aPoVC>pj7T$=4`GtWiv44cVh`-J|fKJYb-U{$IYk
z2NMgd!9#2Ca8@x(7@8$;yjc;rSuvPa96sxqOaFE;Bgk~yAZj&;Eg`Rw9WFXdNH(w_
zi<<qwXg_da&in|_KLVT>=K-TU;KH19G0-mtPLxN!-|bL-dQFdi9hM#IvA08vP&xa{
zLiqo9Axx)1vSDvZBw+4nyJ)$c`N;_-p;N_&CdI^<7FjKFqiS>oN4#89vH@MzPF!yD
z)xGUy?TNk#HZh=EqlvjEde44hhd!?r{kQY(vrO<WV?0%J*Mco|7Tt(q959OG&-`uy
zqg(6@jixh@&0AA018o6W$zbOgJdm!uNstBWC1HEo<mu*FIwl(J4v{-NLfVgzG4oW9
zP`5{@2jy8ucV3$Ms9#6&D4|$06x-6wg_a{z_0Q=Mv`>iVg<c1FPyP^kJ>)%i6bgz$
z!|C-U60r}8<vAGzCG(tuf>L<S!$b0T&v_3A+`}ykc}dP85!xIsLQ7V!3a{;o-qbTl
z@-BtfF2*#W1FJWuaj-{u0((AzK5R7;iabM6AGVr#FZ6mZ_)?cd(>{nfH6?qnFC!7_
zC9=+^T}Tw0GC$ZIw_uYk*o|gEtUfiwdPC94<)rgX?M!=9JEH4yQokuJgw;PJIvf)H
z>7#D6@E#O?Z_;K`zxAGDV?VU^MR<d&I+@r}KouH9Cu~W6j@ms(x{QOdZ7_DGO_t<f
z>>Z5#8Q;L|ZeU${M~NIbw7bc3yx1;Y<m8p2cBR;Iz#z#_MeV0N=PLAa6`kqzHdgg1
zoHk|thkqT`2rJz;&}PM`$>h-&sTTF5J+yhFwqs=MoD=Hg<Vf(0Vo^XWM+=_OJjV#0
zF+9H%JYVwsO7MKebB*9x!}B_Y$2vvJh84@f3Xfo(LlqvOJg--HtY=J9dx%vw<@L}f
zAU%wvo2U~TK-EaI8Y$T~mLRrCASa*AStB%jLyV;<?T1-qGB!-cJ*Z0vzq{B~LoM}1
zqp|6k+E8QLIVG#Yv!hv~(<)GC6)3p#)%yyq`;2KC5%EP$IYjJ6zhU_|>_eRu*#%(o
z<3n4vFT%QHY1TrQe?5`XnVQ_ODHn<L(H<=C!Ts1BOK{f`7P^W~4UsCR-}s$8>UN~J
zovrI=?gsJbO&Lt|iZpeJ^kP%q5*3iF32c#$htuvpJmR{i`pFqKB%*N`Di1?vI>|_6
zqo$1KxfCi(p)>O?Wl&kh^9Q8=0okx6<42_bk>}f@&TY|(-nemZq4R)o!VU3maQ;^d
z$TMyRCYwP|YEg-cP78x(#rWBgIRPq@pe1Wz60}W%GS5Ae_}yYMR3`KK$<QGgx-tD^
z=$FhVOjJN+1+-*OSOIM-pv-g63Vyd(36+(+ekF9Mgl<f~68cs02@s!<{wEfamE=#T
z>nCK(daMEIH}LWX)V%>YFnKL@t;O6y!F%lbp63X`IYQ`3{X5yD2+jw1J}<bQXG5NZ
zjs(|So?i&gFL?br!MTp-$D+|=(VTsQk43A;Y^6#Po<z6DqBk%1f6O*nl3XGhmGJdT
z#O@`c1Jf@Ny-Qe#xKV##i8zAEt3{(~zWz(m=q1l}qEQ{suSKKRJg-$atyQSl^)@P;
zHuAh*p|xMp;)%%8K%sSz=XixyJkRM0vvdV#(HRP}44zLZ%uex~tuV{x`GUgi0?#=L
zvmD0BMf=Y0@A1bB;>n(i(sO<{R3l0?jOp}C276NmA{*?-u-7p>j6QK%b1EWn7flI>
z)a9i;`z1|(LrwgbQ`;}6w*RRFi}WdvuJlpA{8(yClG4}#^yxC7F-b}zo99}fQOk1(
z)Chr`egxEr;Q0>JxWjV;)M(&29BG6ju6`BLsN(q*(s;%5I;^n{bNVS5r{H$9y{A1)
z+n0sD2_w^Ao<hn~=+QBI=C50$l2o-%z6{MTL-Rj1G-)2OCD}E8&S+<|(t2RN9$3(1
z0f`*NtM-DP^r)R=CIP#Bz=3fz2#f}U=<>Po)T3t%A7qFD-Z8+3$q#|RL%e<xP$lu#
zKLL~{fGN{I2?9=nZy2Wo=X5^MmI0hIcs>P`r}+A(f$}uZXMpky&zFJfGU!QD0m!-w
z*yRET#`z#H9}J>_y2ewOo&_vu(#YNgz=z2TL0}=TUj$S|{Pmvz<r84a^vgj&IrxTg
z1#qt51BI2qxsvCnK>3ufUj>v^JU;`<XFS&f*Ls$^Ku#r~>n4_}(70*ovmN=d9pFxE
zyA#{f4H(%|W0$Acjjda2@yJ>{nsJ8Uk|A*N^TNpU0w)g>J;KD6tt*XpkZo4edlJm>
z;;<BlgJ|GGlCO$dSH+Rk)n>n$GVI9wKlT$J{Ik{1zO>o}tDE0H%TC)w)?tOBPNCSH
z>XS?#QKyLU#p?yRxm9(4{OgV3<+82CKJ@`T$sJx|otL;J>q3&x2|DLk7RBZIGdJsd
zJ^z}tuyufhz8+w&{`s8`+VsYu;u{0{GSf~0$`sx-Q-Ly-=QN;9<M}L5p5?g!_!fY^
z%#<Dj<zwCyN`bPJ=Q5xy<M}yIKIb_Oc8h~n>>TAP^t%fE*%mDi`sMNb2zGk}t=PHB
zYv}hH`m-(9Tj=+e=XJ<v9qLR?LXty}Q3%hG$TgBbSBXWgu{<Xs*Cd`(kZTIhr;+vP
z*5@um&Y~b@>Y1o(ChEpG3t49I=PudEGMncdWRt_6!@NcgulaMBn|S<9{@kT?REQi9
zw~7jp@4r=4h}c)Fs1P|0X%!VBrxvZELZrV$hY<dpW`pRkf#(Y(dU>H$^zx|~@RUD^
zX&qN00l8LjB@(7>6<3nv)^R104X5zh#-g8+yi?(|ljjtLe~O|X9g}$;U6&3S*lFjN
z<<ghs(tm2Xlt1DJgZiH$l*F~$WBZnU0ex$a#8sQ#)nl$b0rpIQKFn1HBhO$y+Mgiw
zN?_4`nzKnR)bwH>yU<JQ`jSsgIw!cE<5Q6$6j~7qB^_Y}g$qYmo8+t@N1U5Mo6SIv
z2I?d^47UlxdNlSZ$y;!nEm)6+CvHCp9Ugb1J0KPJVC5d{%)*@!SQ&wxX@HenNVEtO
zTLLwA>h=wpwtGX!zn)iV<ctJHn%-1o&m4wL!tgNK)Z%WNhF%HkKK*~qtMs8s#H8s(
zMtX(Eu6aIh6Ai!@Kl8=U{HOd($?{VZ+8a0zA#2Tag;qMtG9wWag;q9W`rasU7)@{H
zQqw1KW#n;yf9d2n@;J`(Uu0467q=*&iF;%mHoeHpE)^#_#IXZF`fe(DEvD%`UZ!{g
z`kdesZ2pEmfAa}7acE>5`F|>!|0jdS7h^`I@m!4i7vq)}cO>}|Y2ziXjnDk+2kxG}
z=MC{rS=cX&Xg427HKsNCW%Hbay>f8NJ3;jAThivn_lsGJQ=v~P-{NZMQ_Z*dAPPFj
z9~7sepj4h8;D87GLGf;d*KW4(so9rt&hDsbmoE<Jiv#*kIiQ-V#~Yqpczo#pd;V~(
z*3;)})(>n<c=~s*7|>G@QN8*90$^kMIo5fO%~)jOwBUN0&(c04=$sKadA_L1=TEc>
zM6Cj`<w&|NOQY*N!<SqB*RtcyvYh35@Bh^N*G+#D9FX4Sa?0y3xmy3nxms`EA4$}&
zniWQ-*%zYG3z18WtQC!F8UJ%@MUR4-W0_<)BnlRZf{Gbevd}eI@L+sIusFiYKL{2d
zc>W}G{lx1x2o?=Y{{DU3IlH2rb)QocU2E`QcHu0+HA~=lv%+b!qGba7wEtPKntI9<
zVx=U3a*NoK>@t7y%<me{FBZ)IenEKSFSd;zO@DgOU)0p>*{>~G8ft1T>|Y&PeueUu
p>WH~}uv*`?t~R!IJ)G?w99^8$99tCBoY=3<>{nOzs~!9Ge*oYDG!6g&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdl_cshuffle_v1.hpp.E796A6FD7BCB2B56.idx b/.cache/clangd/index/gridwise_gemm_xdl_cshuffle_v1.hpp.E796A6FD7BCB2B56.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a8f867a42e40c9bd75f9d72dfdba3be025910415
GIT binary patch
literal 20202
zcmdUXXH-?k_ck+g$`y~73(~Jr1(70%B1%zgc&sQGyGcxpiBXeSqlpzXBA_T3(Fmxh
zXq09_1f{D82o}JKh`j;!j`$8<=Fa%#pY?voTKTz>Gxzb?GiT0}J$v@-F>sK-|IOh#
zI{l^$S}^mwxsy2^9UXoA|D3tArzP+@Is?sgbQZ*xjhXcGjvj_D18z*VNZj|s^UJP>
z_HuI%t9KQTivDr_b!U@Uj(2mnuPER1adDgOUrQaPRV9u7VtefGDXEK77e}^kdh=rX
z4+mmQ2Ra=8V_xK!a><v!pWOY`4&B2>3nu2Tg#nIl&MdQ<dgp%TA6I(*wtf4rJtUWl
zWiw;`DjWCH`6|b}%%_7r@7_;$|L6*ovyLx$)2hk*+ptXm39;SUIg~z_+9~swtZ_ab
zJq9aQ?#vrjQRsdBMB-+bsb=3^9Ju^r*9<ehLu}FAZhzfccQ$6i@ew_r#W=KSbFZ^i
zfWr@6T3=uNV4%g4KTeO`Ik|gX-SV48QyZohpB>*}+JY~m>Xse3y6a<*dFR1}ji931
zL5itLn^*Y7v`tlQF7C?zt4;0>D<f<F5Xbp#EHCu^c#DS?RWRnsk#FWiOlb1=?zt`s
zT-!OLyVKT(wgW~UD0nfeUzcaGR>MpaU-uY)a2QO!tW*AWjzg^L+S!L423~yQ^wP~}
zNz9}rLmoc!T@&5kdd!0vzvO-!qyO8ZRly6dl}*fY@Y|3&@knUDp7(37+Yh}JGsHr#
ztYcqTXy7!;W>I0QsY>h0=eEbw(r-F_+!*H4!(!U{hV0v=U3dJ@`Sjr5)AUU8?+zII
zWOwYj+Ya~q^-tcLw8UkA>6qQy@{>&_FS=Yd!=>O0F?W~ZPE_gu{W?#h$?j)uj0TKV
z%$av3?A70~woQS*Kb%r&<(gIe<Lj6|t^}sPE1zTQe6`KQQ)80147Fd=sq%TUUBL5R
zo{w9tj%g5npSJApkxQ~`mSr8gHu8Gum4z32?mBtr^2gS8b8Zjpx_(GQ({Il+6X)*T
zF>U<ZxTz&y&(NQg8{b}e-hR!k(bFowt6cA%_5EKai9-@r6z^Yi-D+o#$(6ru*6+t%
zT;vrv&L}?lX}n|2pPg)`OBvrx{MGL6@~DoT|MGXP8TaFheFM&%{n<$OW;fM-hxTdn
z{A#S`9=SBE%|tLhReI>GcWFXutLmos)_u<`I8(Ll`?_qyQFottjp%cyc$6@$??L-f
zr*hnXL+b;-vdj`3W)7PCckR&*A9{FxG0@BRp>9G_Zjkq@Uk>?SI5ESdKK^O<%nJiu
zSB}3{{qwr}x2nCOvH~{#=`y!}px&TaCI=_2_%1MV<5!Dc1TJtY{IYoF`SB-4*SlI;
zA1q5NYBl4`{dE`OpVk)jcw*bwzM{jU$t4+HzZ;zFR2bQK^~Y89>H!@_^<2MeT4+t#
zy>F+yO`21g@nd*}@A{B}t6h~p{&eP&ZAQt&SB~{&wb`!k@8~{R(dpy#v5j`o@J-M1
zC3Z7nRxIjid+*#mZ{rhreU@(#F0Z}U_^fbr?*p5sI0jm;sT`Gf89no<`(oLpeI*&g
z$NsQlK%Yxr|5|Ev^7W!uu)(eE%zERFonL-?u7h}|)t+*AX28kA^A0DcmA2XXaz(@>
zZ~L%{{OFB6T|Gm;32;b1_3YxG9}3(?Oy3k2W3H1C@{9NFGoif;6}^53bK7=lb3`5d
zKxgRNIb7nmZgI;V^xX92vIlmXmfIA~v5KB>d+yr9mrA~=jPB6iJ>$w>>2<F%tIj2@
z8gu<2Z?f#8qVLEu@!|7foie)A#u+)kz3@8C?)u?H-nFxCbxiOL*Pl2?tWKVGH`24i
z6JLuH4;DtIPE`Lo>FL4mUcLX&rq!&a(N=nPO(~YqmYez-Zm+d_8l6)*!Df)^*X0w2
z+O~V~;Z&PIz1P=fepO@dyLnX2&+jLjyIGbTj(<I?Z@(k?AlJAlXZ5X${#91bzWT$f
z_-0A#-<G!D{bBcvRja=r{r!=0-7=>kH*dd~JT9kr;q{KwBs<quZ$3%~a(qty)OXd#
z$kh{L=i7{J4Eo~O<S*Am%`Q2$xh~Ad&wpF#0z=b;H@B@X9>@swH7UJnICOzft=Ic|
z@z(TfbsNr|_$fMdk&ry0AtZ6#v*2g%`}yk*IzB=7+3icgSF;w4sJ`bnYdSj7rcIq<
zX=7{enFGhoSvd2XBchH@>ytC{SH{npqJ)mUbmVhLbq*P~BsvPL(eWS2vao%4@0tt)
zsP@s3mveT@xydaF5m8pfWc+kz)>8v$ghfJuSt#h#k`NJ<MN@|Uz>j`P=zw2Tj#TBy
zq$MFDhR>Eq+&B?pZvYK&DI0j(4ZLqlLPX3)?=$PUdG>b(&;*OD0cLCPi!_N=<l<Uo
z<%AK7ji414nPWbu-KPhk+sB={);+d;v;lO*^nJ+Phw7Guh*(ygaOu=PAJ@Yii@Zg4
zZ&6@Nq7}K&!FHw7lx!vJfxo5{sY>zJXc8jQA#$vj?M=nD)=-H{*~{zh<xN`>BGS2k
zaM*f_t+!OLD=y^_GCqXVEs0iSh*5Zwx;k13hx_Wt50UC2GHpqSNVjheT={N8;co`8
z9e&X|&S4$buO%TOJ=e7k@7(s~>ejG57HL4b4amGDAtGLd#p87%hkkDaEwRXK^V#h_
zJqYbRX{qOuyUFEB*cw;$EK;3CZCer|;=9v+uaA1rGXrRfOGx5ulQ_4QgoyNi-SPG7
z+1o6PpbZw8XFjjprw5@2O#e~U{)z4HN;nLE>;q){09mvoL}b+Mktt(MGt!k%#Pv9X
zRA+EKGzk$IbL(6|AFK@wpbLJ<pHTfL?9q}Ck+Fx?D4+c5H&cMRSmXj~eF2xENr=do
zZ(Xxe95)?m4SQmd<v@2iFl$MO$h72)5!RzWCMjVjT$htbbrR{fBwCR}!3pow{8=UJ
zf<?|F)p<O<G>KNE(XzpQJLn|o4Dr^HO%<lLpB_4`c&HQL1!ORN%BXge2R-#OfX=uk
zg+N^hdbA{@j9HWRRPCFyM^Zvh{F)M^D#621lMs>l*<D^OvAp|@5tMK#i_I6?etHmk
zL7VEi&cEDwtArhK38#?i6#f=XLPVCez0fMut)M^&yJL}Jq$<WQ(IiA9cx>UyY{j(!
zMz9rr$rAG=wx1q^4zcWCbojOZJR@j_=>_Hs+I@NuI<(5@o5g1>Q<cyizn~1M%J2&`
z2@whF{Nt2(zd>dO&={AH1I%)OLrbC+nLfnGW!>9U1E`NhLV#HaaBNA4$c9Nby@h+5
zp%GMIkwxZ<Y(G5+9cP(1fBBItPJrDo9gAAWB7>HMi0mHV9(`rf_YaMrJr-GLzOdb=
z2ch=_jR`t7B%+@Zy5TP=LaHL%nwo@&WIky(;Mh|O10&c0m+-6kukAiP2%R-6bkDBE
z7m|!&H(WwI(vQdUUXu`!V_&M=W?hKFojw%Lmix%~KAtU_L@V-O*rWa#{jnx9#4pJN
zW|_FTGzk&Ozdvc<8lwht1L%Q8l6iG9Z_$!yMIM93WBtd<ni1Mr(b-mRe(1CkkY<FQ
zOa~Lrc;3P@PKC?Z1I+dSyOxBMQRFs%kKT^B;|8!Z7K!EUV)^kciB@E=)4~ALMe#<k
z4Hj8yzSQ>9LrYH`ow`&%Bf<c-#dH)fi^8g3lMs=z@dpmqOi>**g0@&>zWMxipB{w1
zGTO>I#BzeA04?xL`$hc4{nJCI6$kJ#hiBTug}YMa$`9R)pf%1v*L-ffPY;s6e*V%q
zw_?WrW(;+4gI_}WmylUYLPVasEjkyuso}f<)Wapj0J9j}5}JgFyemEV{mSpWn+%{C
z7I}*7p5oO-lMs;){%WU$_Ukg0&<S_`X{0)hCxRv+BD24nIw#&wM@Km7VdXO*C0k#T
zF9Fph;6${AB-a4f8ZemRN5J(Fi{F4wZ$N*_Hv!ismJfrjVUWqc7pnF`C!$>?ISsm|
z!NC+~L)UB;m%vUXus`L`Lf5k_UkzQWS^OTheGfa6^;wcvqP8m$gCkMfNEUD9`)%dF
zBm+y5L&d(K;v_Pjb$|B1Q2JL%s3^(tz#|^?B3&cNNx&ls^df^+l3xO^m!KbQh+WWQ
z7wknV9)Y|fP(NC+TyA78H=Znal3dCKlyYMUCKeTv<W!_gMTTT5OL76~Sb$u}s;}E(
zXAyNn>@1|)VqKB(D#_b9X*(yAg<g__MOCmklzeI%x6r?S{IkndV@aNEIopn`<dQtc
za*mx@i#a68b1moEkr6J*^DO7tk>yyD=UdLVBeO@6f3^J8j_4;zUSPSvj%aR4UTC?{
zj%YtgUSz+>nP^i<UTnYEnXEICyu^NqGg%NMd8z$U=c1Mwj+-3WyhILUDoFAxWbz8N
zr}zW1`G6dVAtuR9$fOCir+77|Ud>sP7L(*~P94tT?R=l@{6IQ368Js|EM6m;t`RLr
z14{A+(R2gFZPNGTI`F?_D<!;ea3MfyHiMi0;6VnsP!_KN+!_`~0WONgF@TF<@gBhK
zVR0tlGFhAhxEvPWgWNsHygwc(;*m;Io96l)LW)ByzJwH)Se(RhNgR`Z1J7;XwV2rL
zP`AZer$cTpGTw{q$=KC1t?cVoQ8OKKaw0c0ksC(FA}5~`x||Z+NmCLtACLYdq)b9a
zqy$OMMK-yp6CM4UJpdW~#P|ogEhayyim}t=@}l8K-{BJP0k3<YFP*5%k=Jt6mrl$B
z+{gpmI5H7UFYZ?~?#^&;#o^)Q`wI0U!;8bi$+rphBEzZqn>-S^?(~~>a~``{yoa;j
z!*!(mYQE=c-iy45lP~kVF7v%f8*p+p->aI%6+)K^+D2x2KE-~Cc_)A*m+}3}_#l#z
z>~3&(X!bOu4J3IVGTukKgG_28-A{WPDpm}}7lVnkh04IlG7v=Z7C3SX45F%SD;&QS
zPNcXVj<1IkX^TFEBcH+`ieI5GUZF9hA)Dvr7fomkX-GUhICUu3p3ciKP94VLYTm1w
z_aWaa$!~?Ax55_`Zx;t`7l%-sC;H@xy(rEXeezjcDEbt#_=M<lg2iV<zcXTAn*Xfm
zcb3HsqTC=3COr90-2R%4A#K<U-Y$byynVw|t0kkqOv62RN$h?}97js9xHvBRaqfX)
zJXm8mml)2Cj2TJ3&bwUa-H4zhCkqD2f)T-_39<6mY&hwwH*jtnxJk62JG|Q+Uc1g&
zw(Lal8~;Yizu5_Q4A~rGC!A(~N+z7;gN}z=&9ChZCD~BNP?xF{M;%8kvgv&&^I4Cx
zy8pR3M0VHsvsVJkmB1ye#q8H@*<;i8E8#nq@E)Y0UEIUdmm4UoaMvvd?Un;4ve9#y
z(&*hDE-nx`IfLt%!F8h3yMXIhz~VBlM;XUV|8lNJIg6_}`zp?X=C9)Xt2mkB)e5`S
zijJg4L?dF&r*Q$09_rk(_s2bS8!2xiV`6~nws-)@HnwxXl!5O1ehn8m`5NbQjq62U
zb({0K&Eh)Fr;f$XIiKe&e$Nel&kd#dgZaV1ycT!48=s>7`oz!pMcY8fZNPyTRh+yZ
zbleXdh*8DKtN8(|d4F0^sW7xu7)HLClTQmnPqVl}7+WD|=aIX7J4`Se*IjDv40Q(A
zjuuqFsS8+K#<`Snohe_=xs<cGic?py@>QI773V|cuW@GAICF|`b7r?$T*sN!vG_S>
z_MFA<IkWdH4(82*d2M-Q+TzaDcxK4B^r-88c}DS*5x8@WIb$O_{hT>xE$Y!XrtiUb
z)_VVXd-v@6M_^6izL&a^d;_R&fbMk3cn8$)SiBsnmqUhM1J!F-d=RP+!tQiQIStjP
zSzH0t6)e6E)z?|P3iVxuCJ|LE$?K7NJ?c&{@nrG3E&eOgp4bNDrEp$5P$YR1ueXV}
zBzA=)pAovB5&X#rl;pF5^I1W=7I|4bEtxJ%eyZl=!^rqBG9}Xzk3eLc@j155Kz5Wr
zf{c%_{1Ie(gyl1laVE=WBHK)s&qBsoET4rOvyd|_|0ps(%JN5%<58ABhK!G~{4wNs
zjODYDaW>0mBgbr(&q2mHET4lMb67qX8RxQmE^^Fe`8;Ht$MSioV;)_1I5{5~=d*l1
z>X^^+1<1I7<qJ^f0^~-&zYrN0vV0-xTu6Dc-D#uy$ze&y!#O_i96yTo_hsJeGF`$Y
zxk7NM5L_w$MhJSt+9KhiW4P!{`6$sbipB9_?|88f<>SSE@!|j$4~!QFv3N+lsMUay
zyiFXmjkR?e#4ZhDS0Yawu6Th<LCO@QPYvmOWSfudN&D!wxM9hv;r(;y_q!g1&JrX!
zhwqld_a`Mvavt9;kHrOi=K`Lwu8s??#|5n!E6GKIYY~fAh~6t`jeV59$G`ILun@1L
zt2k*D=Ss{zqD!&P(|CbNpZEp074E-E=!A7%B4?DynG)@YHw%J8lHf$D+9#~ktKs^b
zxKH0;mBVXWrTH7I@>pEJn-}mdB%>rB7u3gv&J-63>LM1e5Ie3Ay$IhwP`z`}D@A{)
z*%a?u#`hzHBrg+G%LFTmLj+X_i?<4@tt?IuR0%9D6}prP%<woZbUDr9eWKw$u>&nX
zT{KK*ahYgX#^O3rzfQEK@(rSX1I1)3kL|~1w>oKYNq#Q`y=O=DdC~7YwSbAo9rwsB
zq`bwN7#jCGF);?}zx2QKeAHCj*=Gdx89{4L4XlXu-ZFl`oN<y|47wGAUUcOs2L8oh
z1jS{bTN%rj0sk_VzYLr&vn$P2;Cz+E)xf!$#kD|H3(RTxwZOd=_)vTksBW_SEugx^
z;(DN}{~Wv5v-uwb)nk@_0#r{}{0^wzvG_euy=QSFP&Kl68B{HUOnsL@_hpc&Z!lB^
zvwR3tg|IjRsv<te?h$PMwNSN|<s+dglEqtKw=Jw+YYX(>0!Pr5dMoU<mF2fW|E(;)
z4?6E-SL$TwoXp}B=$yjh15kAUn$z+RK=%XChvGv}b%^EDp(>rlxlon+Id;!w^XEfV
zKFb$CRRN1nL)B>(mqAq-i_4*^oW<v%>ip-}{XF}=N~o%2`6{TYVsQ;r)qIZKYuNm?
zP*uzFH=ybUi|b*xdf1CtI+9!u{p;Zfil4%6Pg(vc^nc3ojnK6b_N4Mn(6x!hAEE0<
z7Oz68Rmhymhay!di&rDnY8J0Ys`V_6LaHbhM<Z1<i(ew?B{HVYf>%g-#o`Z0`hbkd
z&bQf_AT_Z#lv9Or=44gDUQk}z&Kpyw-gdt8cD@J234G@S7VqGt9jyFLUfRjxbG+U;
z-i&J0Fi{mIno}GhIz@<DMJLJ8qEj@BV?<R9E5AupZDR2*(Qp@aDoXM$vD+@zIguzD
zCbIl)(Qr45Q$@qn&vCa@Hh-FEn8xz^MZ^6R5AvU7J#G^`fDMS_LZ9OTV+RxoeTrDT
zLL9b296>&Om}k__#x4=-@Z$T3AN7d;iV%{#Lg==FH3-6lQDFkZ?-07}VEJruOt$z9
z&0i(<s1mhn?63e&lT!mWjl?yq<A>Do!>M4nFeF@HaIrYNSR6@satMqwCL?7sYa8Vu
z+dS4b(i{yF+i3L6{hmXW31jdomMN+-MXlSCC?)Lq(Ht1l1vH&^O6O(r>alH$x_q5t
zYNu#k?o5h7SGwGk0h2NoZ-FLTU{_MIByWW#TUlHWP3mD+I{ltPlcy|xf%IM=Cpvu^
zkzONmBGZK|%(%vygKJu2&B+uI82f%xm%B$py8WZ;cB0Rgwh8?GpZ@jD=A^&s>IbS`
zCgAQ114d!Mfw-Y1IT9E}0tey;#eR8Uv;jC2wH*CT%IG;`qJ0Nj+<h02=LOV{eB+k~
zrau{8dh!M?@&C{5Mw~w!K6CkW#7|rC0q^#J`f#ypQ*aAq$M$7Wby>6|<KpX^q2CMF
zt!i-9W;)FDB*lEQ#HH1h4Y&2NePzTMnbKXp6X&EwQ%AQyP@}r4hijh+y)t26I%dL=
zS2*k8&gMpDQx7)@K)@MyBIQn`NB0OB$T|br&^>}CI)TiCX)R$1IAa=8rm<&Kg{Wg8
zdq$-RR?uc_30A<SeRHHjTD(*4cUkKXlSzOFqb5Fqe7__g<ed)E=`%BJWAXR3gUxU=
z?d8>bc`xF}BM}U^MthKQ59`J}ifoS3?V@f=Tmz}m{2~2MJB{r=0v~{G2ELm?0C7p-
zHvr$gAb_}}B>5chJqH3vT_t%3^xXjiXk)*GzAs?_osb7O-ve9#+0Ecj;d~FW_z>rN
zh{frgZ#s*!Ip1s+=WxC`EY9V8b6K3n`R1|s6z6@48%Ez>%y}2HxP<dAVR0$vT}m-I
zh{HRA=4cPvBOj2-2V_O`h$Ke{JtG9IkDLU8;EqT{%0zZdWg_cLc1&r)Lr6y~s;?bd
zd0sZgU7I4{|GpqogT#&C5_cozZnkT)kWCibwVG%WQX)A#Y^D1tiiCc_OXT?y`BBaI
z3iW=4WEKy8g@P!K;kv|dUUaj0o$qp;*X{wap~VN(^J7VUn_ct9$%1xj;{#q{c&RXk
z@WEMuPGQv<OPW1*uEu@ooOX6~)*^CXif2S~oC~dclMvKI+h)a%rN0f^bD{{BRRqk6
zfDIj;rNE*TbR?L>#^6TWhm`x+M$ADrIcy_pB4$V<l86~x5ly%Ztx6g)N<$rp&L9qS
zJhe4JGej`luFr_bD^u3uI{x3g_VB@IMSbMQ9{rk+O`Sr3J1MdmcMk!+6o&%GP|$_y
z{WZXG4d_DkeiU$w0$qsS$66iO#{d^Pwf6w~J-~%d?M%=q6Sy5~33tJs56~$GxRGzi
z77$Pz2YRG2n{nIY?AGi!us#kPSl;<KyOtjZUdKTnmLGB)1X8>aDmFqrs!BG(wi{W~
zej~Kr2pw47c_V8JY=mALVIP(svJnPSd>^*E4{fPxx(_?vXJaMqLy!B=hvoh5!+|Uw
zejk28aXd1MM^>ckv2B9Pwz7B|GTX-D?Z|99ixZGp0*iMbvmGoxi4-T19$BC8;6{of
zq({P{Bpg|$IED0xuEDkxQWPUSqA4V-;}s=Hk1Ps!1|mf%(j#YG5)Ot@oJM*iP6j(~
zkfIFfkx3%qz*0py(j$k_*k^$hXOJGT8?o;fDb6B2;$y>015%trdL(p1!XbK!^SJ&k
z%cLY<MdnwLH93Eg<Z5JI&Ejjw{2GgEka-P@uOsv8EUrc7wJZ+d3_>_Fy46|18LVLO
zI?iexJ%Gk`5#MVAKY;qSH}cAjyb;B5yki{Sg>H@Fc&|9V55<>w<t5&TXf*tNyz(lG
zlLUh#f!P|J7nJAeVI{WL1>-uQJ<VS)7}v9Sn`p32Wb$tp4Ysp*mBM(HLYqJQVgCz1
zOr72MKY!X2@hmg{aX-sD*o6Lf58~F`iVbjCI_|q>bF=46Z=$N#h8wuY@A&yJ$H||c
zoMR{aXD8>J{Q1c_Cx3o&9(j3Tk2X4o>YLX%V`CR)jWc%EA`<w6EsW+EN>YjD=pK`o
z$cnDN^?c8IUTa}&oLX?NmzfR1Wv@}F)+m@ToEU{FhT^{>1N_JIJ24!8WSU@+Mpxk2
zIH#M73g&tK<Ccajf`V>K{3flVCYF*2G;b(|UO+))LxGRMkn#pHqBA~;^Gu=#rINg#
zx7*KaPY2@OrmnnX65byBkuL+!%dC}o1$bUzaRX2{uuIE3;Qfvb(bxw)_pw%K67)=B
zaUN9Xu}jBk=zW?E>1c#rjWCd^`VY|S1B+K7^(w?{JJ%!c^++4yv3<r6Zt%!IZUQ_)
zkVhf%BwGj)^M&n*=5SZqkROGhkJO+cDjwS@yO44hYd#)9)<;<LQ4<eKdf@Ld=X>d9
zOn1R8SPRs(pbK4G>w&tS#mk_28D#hfsE%Oq0jNFzyU@im7pikvd>*RLv$zJTYgoLU
zH`&g+P-U6Gn<TI}LNtsJ9jWq*77e2*-d7=9{^M+`Ie64QM*5FYXL4kjwBRT|VrGK>
z1gvg>=T*>~e&JQ%e-#X;xEgp?vwSu1uV(q1Kz$R~Qu&*}@g{Jm_!dy#V)<LZ<reUu
z{9~Yg%<_+c<6}1e6QF*=@=t)v6E^>QpnlKt?}6icHh&{fH?n*qaBO7xV5km;O#OqQ
zV=!du9|F}OEFS`0LLgKBwNSm5<<~;TwQT-KsE%a$NazyD=1+#6$*?yaA<57`84jm7
z1$w5id<yhWVfjN)eF)l8`9sk05Ok+F9jen=J{`KGLl4U5Lv=pO=R?PQHh%$B7qEN*
zbSYr-mqB$I%a=jNGB$rXRF|`SIdm*%`AVp+Wcf<ySjqBLP+i6HRnVo1eSa-f*Rp&q
zbgX6b-+<~HEPn&K++g!JL9Zsr1nxD#K}~QZ)ma~**GHEB2nT&+`B0<|MT|KdiX20c
zJI%iusaLc7YUH?@<)e@~ishq_V-(9rBXu;(M<d5*mfyjf?BFe_hS<T|@8COAypuQC
z$?`k-PCHq1te$tP=X+2M@{)IZ$>IpXEka=U4T9SS7RQL{7}1vIj}aYXs5299m&Bei
z?9p_Li2q9K;KiO2(Qvcaj^^Ji+H9s8teN-S%pPTL7KdyWwPA{QA1GQTirN4syssB6
zce6N6v`l01e$jG2i*v-TIifdx|4FgyNfuX&s%p`c@;5})4Hn-ORd-qZT(o)4*8hcQ
z^Mb`MMVpr_UZ${FreHMqa)r%u76&VAf+<d^jxF5QEpIwDSayo4ogx#&mm;cCC{8^%
z`_!9Q`&g{{R|3P8z=CXe_fNd!o)DpDfQiFM|1dHltLmY7r?xB@niq|WJ_?MEf=+Y`
zn+J^YKqs<=O@H)Tx=Glxz5lsDydzU<Pxsd?xzS3(u0O62D5<UZC*=Nw488~3+=C{R
ze~P%Lh~eKN?k$U#bKG){;n#88Iu^(BTrAJ<$vl@#F$rMD221np8fvh75Q08XgN5vY
zvD@QwlkljIwf^nUS2(&`4p&Iw3S%-H$-NugXE%}ZCVOVCxqU;7$YZCf#wB|{zw@8#
zGcw3<(`YUO(Z)$dMyb>%l+!J63O^v>5k3kk0zHa=A2lI1!|t2e!{SG<`y=*n@+#+k
zmBvI$@<Y!3A&cK~?r$m19lYA6zU-F*d|P8P9K4yecjI7S92`T}&PQ<YBi2j%7zRFu
zW2o-G%K2aAhET8RL(cypi{EnoZz;~5d1c=(ewTLQ7jFR`Ti9zNslX$Zy(UrtJSy1m
zg(B!x1pCp71tYIu_L|5cZsZ~Mn#d_G;1s<klD9Qg_#s%_g@;#)Af*UeZAXR~9*~;b
zSfq3ETdrYYqWSt2@x7mLto-)T@$?8h9K%J!aIq69K$4?G!zhZ0w+#2X=86_6xH-be
zAO%^FsX*2-EU3AtMFo%Y(otR;)I#oaU<YmUbuB9P0eOBve$>U>gnBn2nZ<*fP!Pp2
zoKFllfOXLNT<4ikst6$<LKsUwELpHl78w0~S{QzsZV!pu5bydl_rqw#wt=?WfGzbS
z?gwr616$HE&36`zlW3Sd_A7DvWt=q~Oy!(@ImJci&y|)we3AE$`_Q5Xe=Uo<-~C*h
z=3qjL3}i+0KPMkS7Drf|i7YZ%oP{j2kTsP*iY$+^_!zQ0#^P*bna$!HWSPU_Tx6Nc
z;yh%PM|Wk-!FpEtEG|GD3h04q^RBQ%A;l%PqiXuS8RUxx-veHHz%$`=p@I}j!|O`k
z9IJfOyUrO4{@+eyatjGN6`C*b(XMM0f*PsX#^HOSUj+?hD!2VTG`Lrg9-dBx$hQ#n
zAuV>s$9_>@K)a3qIVF5{%#WsT=Eqs$7HP+|W0o-w&O?jNU6?n<$?D`%{8<gCeFJhL
zo5_oJLPsZ7bOLyvy_xgc%#EOvCzkVyWpTdXl`m+Gze_7GnhNvGFZ}B+s!K0o1Ao{Q
zw(1`z2kBz0*P8ES61`UA5D|2GRHteGIAUsVE%RHipLP`=$;SeZSkRkJ?_}VS40_Y)
zT?srY*?n~)^hjj)d2gV{8`zt!CF$JAbZ!Ej{(0QUJQf#o0mbw<_C}90{az=Tg8$r5
zySXImYkgCXK-`qiMb&c}HHf1e6{=+lZB*givcI!;EgBZUOY(i-bssR7BD0`Z7Q5QL
zh90lk_5L6?@*um~<!~c&C?@v~u}W^fR!h1ZJKFic^K6iAg%DK1&hmFc&^s2N6aCK7
zSue?zqF*J&j}%Ff&?{{no)x9Qw-m_KUWtR=anPTx){mk0W9U!1`jPx5VAHhCI6?^r
z&HCr?+Pjs_0jK_XEH2>v3V3Z8IdKr;!TWjGMg9AUZ-#C8!}F6xW40Mu(O1Voi#TXU
z@gr#V2wIV<V7C#pcns|*zRDS0r8gLwO(~;?EPl%wy`{Lp)AID}ueTk?eHaHk;(#xm
z4k^GR1^CkG@B(<fVDHnNgkC4v=z|h2poAMkhX1qJ%v{$r>kj|9;r;C4jj%f>ZtVWo
zqlKTV9@K@b4(Rqje@;P$1YTq{m(%HlOhyLD)D)EPJs#`TybBovFACb&&oY0!?jM!%
ztHmdm7{AZg`=4*BSAPsT(xK$Q%YQv8>h=57B$J0T*WjkzC0g#HHv#ZHa?vuCVsfs6
zJ5Te^2GGun=G>y$tAaOqx102;;QRN9xvte4UtlFy$Gg?>%-w`=!7W^1aIx66m|id^
z=d3v6UZmX11|MW2n`}1tK=ZE@5U2j^?`BN4!5e-OxTQn3#za8-7jtHRJ14T)HQ?9$
z(ERZyKXjnnZ`wC&MZ3nJ)`NTd_3<Crw_pE319W<81#||`pZ)01{`BV{`coUEf&Za8
N`nQAW&p!0${{ykqDxUxV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdl_cshuffle_v2.hpp.E0B3C0FB7BAA0FB3.idx b/.cache/clangd/index/gridwise_gemm_xdl_cshuffle_v2.hpp.E0B3C0FB7BAA0FB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..beae9ba703c1b999b88b35097ac05c8a26bd4153
GIT binary patch
literal 14774
zcmdUVcUY9i_CHfz7I9sc)m`0Pq=*872%<=r*ic_YY>C&FXe@~uW7LGNEf&OrBov8C
ztO$yQgb*8c1Zz-iG$~525v8jM>Ths&n7P0I?vv;Kar2n>?1#^pcji4abLPyMnLN(V
zFUiBeVC4L9tClYKb}lwBFfb+mul#n!qIki;z(H+buxfwiwAm5++<UYJmL4{#w=Y<C
zzvqk}@y{>5`e~f!V!x-~pY|^4Yqv3K(4u2G?XS-M4(3_@F*S3sg<<-I>c0wJ6}(z?
z&En;H>C(2HCA~&JSz-RL>mN^+JXrH;-qf@t!w82}C)1*h-wp1&-NR$e+}N+fKh@Ou
zg!j4KyrOW|+<i&s^4Ih(D_XTwv$Dl+-}~G0G42aL+uUP==sbFW<?;y{*_q03PoJ4J
z{?c@BSN92Bwj8;(wD9)d*RG#)RGk)+d&WAyK3Mn*+<AZAT9wl`qiXK<{`|Pvn|b?3
z=e7*%`@rt2LH@;|y>ASAXZZ8kdwm8yZ;vW-UHB~M=0iW5MXM(5EQ(xp>Enjk>Vf{(
z%S^M1?7B_MN&fEocZ*{3y15-q``IKl);ZeBZ?k<+H;dd6@5{F-%yZzLhV!4V{9)!h
zzqPKByFu2Gr9*8G)pZ}e{Nk;aMgD`Dj?LZeGT!2w%X15Rcs7Gs4b>ApKRa0Qd}8?5
z4Yqaj?B6C`Z2u&?e68l#3BRPr`6cVq%&l7nZtvTiu<t_C#O-#UUHCR?+lq&2ss2Bw
zIqmT_FL)3VvfFyv@zJk;m~t;6WTNkO)9d&5{?@+ICZ%N6{84KfvLcFmjqbWAw_x+y
z%SHp8(s%mi{9t42;kLv1YE9FU{^{pEg>GM8F)B~|df^h&B(K+wqwV&rvkoh{dcr7r
zlj#Y&^vQb@vpzW!SZ~(2=h+n7c*E8C0lAk_U9xvC&i=nB?cDsYOD11g{B*v{qVb0h
zF0)>}_0^WQAB{cLbneN65#xVt`m<G)U0Hw7@amV<aVzbAynFh6PqC-n_G*_sMQ>Z%
zz7MGVc1&SVjZOEAIrZ|Vm)E<S9|_Rh{`K*f88uV?sM)mW(*pC*DG9gEt@pPK)((^O
zJGS?c@>X@ms-<^(!8ZHX2}_<v{N%kd;Edvp%E4`!%;)aqAz$7(v+0h<zKu&@UGSXr
zn+q3}XdbSye)z@e8_(8sEonRS=>5Ug!5PJ;WsBD@7?^W4IAim%ZjVE%0*+7oZfT2Q
z+)tN>ecSEf#qe8?(+*93o>lbi&owQ_yM~;5H(`<cwLyPP?Rk23^xE9Q>?dYdvgdxk
zGIjW-z?FZFt+UF>Irp&3IG=<wv+o7{`0+TO>9aE*`Op4UDp8&tw<KrsmDF})<<hM_
z0lsTL*|wtXm)n+UbLHhuF$r<o78U>XeB4j-rM`{Jj`a>&<{I^|a@4Mb!;4a022Ofv
zav=Wd$1@Y2c3+g2b1-M%37@-Dr#NO0w|Z-^YDxDEm(zUbWo+mc_SxdhjTTcHkIhNx
zJAP8lg2PYJ_D@`Y{;zFDj#Ceh(D?s4UfC8B94UD;KH|=+-#-H@W-czanY&27TbLVm
zrFYNhYbH;+p1(alDA{ZOnjvH7`<nf<?@B?Nw^MNa-mj+3+p~72*@m>F14fy(HHQ`V
z6CLB8?1>vN_}!xVtIpDvHR;RN8(o>Yb<^LbG2O0xWwfyUcCULjbArIaJ#&JDs@OZF
z@OXvNV`gmcpc}EReaBvj&K(_cFj)Sle@UCwtRG6ezc{~jLw>>0v$d%e<PWNMQ?9>X
zwROgo5|`Vf7iz516vqQ5J>N1$(@_7b43{|e`lXkJ+^eqs<f|UuOE*t#TkaA3#mh}8
zr|a97A2+w1v?Dr5F+E_<%Di=}gHnr&cN)D)>EAR-^Zf5FZ+7$_HOTn0$CmYLFZbTr
zV|mZ>*N5Di>%0AmbMx}z!q%LC$J0E@uAH-2VAVBtaqVgo+abX-&6*raUl#WD2%L)g
zykBtq^nj$x`(L?@%iW;b=QnKh+)}@(LnocI5y$UrX>Xi<b<+jU$N`&n?Ah9GA^G@A
z^Lf69ufM-NYTyGe(*f^K+dp%5@*n5%#eA##E+&ztHG|9rKRb`LcYdt1zg6YG+VJvc
z>r29KxfK+6#NU0I@aY@rjFcj$y(4aQ%~+-i?hzI9N>!2gd%u+L-o&0S-c&Y1vSY2;
zlkV?c_pCYN{fWWK)k{B@85tOW7&n>e?Cb>!*mIbH7!B0Xpub)i5W3?(@GkY;*K^0V
zD&PR3@($v6u%BK~l`hKX(|Y_h+(QOEKQa)Tkh}@?)Qe7)P{~)mrlv!b(1^Sx7-)il
zwO&va<CdUsgO{HsC}AI>vQD$k?8Af7($h04W?KB!O$J>^B`=Wt1v1qOs?syy`l0t~
zb-WTv$s0BU&1O=GPEeIzpU=yl`BpMT0cAwx08k$Q{q%yWST{U%pMD~FuM%1jm9?6+
zW*;7%bjyxe2L>OsP(THt(*>J!;Um4EDz=}>9TuqiUsFIOQ3(g?aAH-RpejzaxjRbj
z&pcN_V_yUDJu-cd)OtZxh9J-Dze2OmWzd#<a4V9x607P2RdM_B;ogJ8{%}>np+w~-
zGQEj<=>=7B|Ks=F3abPM1#}{9$rSo!3N!SAs*DM5RP}oE##{+4h{|fsYO@ayN{hlz
z)t`PHStT_H@G%h8SZ(p)VbDc@29vyj2>b&!tX}7EZJ+|GNckH;eFNC(1${?=+~?@F
znbW^lKnJ4o1ZbZCcfFu0lZ_(>NpJqWO9?HB${NiYvkwnSPo1zNdTH#%b28{oKD7bK
z8;CV@f~w4L-muH+`9c>N>`zo)Bl&Bj(hI8c>C@?t>?g;aP{4ttl5GfYL;duEs(g|0
zt7S+{(GEFuC2uK2szRjI3#zhcVoip_7tik~pecFFCBgcV;G-8*WvTD-XLrU*cXWkb
z9TfrNZ9uITROOr3P4^<FE5melbun_W7CRmWT?~=Vu0eBNX^zeh%~C))c}FNvhXQ-O
zpzm1q*|0h5{*{}cfW3%H7Pij9&U(RAGUv^h{6(}3+LP~TM)GE2IGvy>Yo_^TtGc~v
zQ9wuXmXlC>61wXJRax)4V`-H-Gfe?ah)NPrCy{xh6I5l(H0_*t|1P02*p*cD3dvug
z?s`F0Lc2V=U>RTg#1IZ8Z@G@TUPlVOpeoVPp0c8>!BsLikxYyzERVvy^n$AF@^PMB
z`f2toLugFi@)mV{ODfU{s<L}uTHTed-&|HeXQC2;a0KZ@ouDc|ZyPFhd9W>620ch~
z8<D(`*hMF(%7Ja6i(gF-1~MoiDs4#KMk>+?s&ed;pzRB*-+rZlYEn@Q?iYh6>jhOg
ze#)ky`;RDC3A+%L5TFSmUP&jY%1OV9Y|W^mXeBf!DyuZB%sxCi>EC@5vzAG!6|j3p
zB|sfR{E1Fbm9vcrY2SsFb&<irq^WO^{0$lVIzd%3d)HPDFta26B#>A(600Jyxn6Xt
ztnIz^%&YQn1sp^w$^Zj1z!<%tDp`|n+>EF>@?HVWh)N2!OTi=cf~w@$+mwAdYTO7J
zv?Fb)NAh~&sda*?<Xs80xxd^L%is{A(t_kI#P~WvRq{9Yv&~!l{XPZMl8TPwe#h}#
zy`U;ZbLV>AZ+Y&afNn&kSkM*==6XR@N<6A=<QDccmO&d*(Mu$MN!p?lohreHilZ)h
z-&Q~kc}pzr7fZTOCpuNqmEY`s(6^5gb|WfVfMyHnLY<&0)dx10j=$<6Q9^5?^1bGJ
zvkwnSzr30q)|%*a$Po4>PBcgoWc}e`&;^5zHC6--dG9`sZ+JdI3Hy=q8#EitK0Ijo
zmRlQ#cq~{&9B2<hrvh~<@v%BVRsQ}iCe(Upx`z_>B`WJR>&-qqDBZs1%d)yb!`H~*
z1k(KxSQSCKUni)_iUkW-nmVwk02=FK0or=0_!y}kBOAKz(AWSN>mvfHXn4uh(tC8$
zm9A2;9?0u~16=^5;zsDS5sD1|2>bsCM=^W|IvwKqR47k{4%9`F7C@)RP-M6m_AiE`
z7;c14jXWQRy2qh`)FDX4Ur_g7czh0ZKgTc)Wsp>Sh-43ul76aGEJuCHQGdET80!NZ
z>e9OC0~}%sU35cDVZ6*Q_wL?|?GmYY1h^gn!)OR16;A=zQ(zc%tx|CZ^xOeQGE-cH
zt{34j_Gzb)=V>&OeM%LcRE1~IH9#sh;lL*RFN$eAMOxGmOwsnxIEsAXJtVt_OxPC|
zBdcQ6hkc<g$f94^?d<VXJK>uv3MpBHa43*|2oK3aN@5!v&f{o+qj?+ya14(R0DOSQ
zNdPDDI2GVj9^U}?29GlU&fxJ|h~Gl4JzEjpiZ~pOa5#@|BBPr~&e~syj0$=D9vQvo
zaTdl|n0wzP0bde0oG##WhE)!8@86zq=`Uj8cHr3#Mz9_|jy#W}5v&LE@uYk_o%R@w
zUrCSYqE^~xbjToMeXL3eshEuv*~pyrVX|PAEI86P_R9O+sQp-q3&vt3aEk=)G&IHH
zF5tEcxYG>)7WV?Ty}<pXeyPD?vfz;{c+wYH)Y_iw5>(;~rD6|*9>&ac0}TdtqCUQ7
zE;PA6H2%l6(GZ*%q9gvMUu&%oy~&?-aW^ICR}qJ$3$rO96(1qhBj&IiUBb?8QW#kR
zDGA2RwgCqkr#a4V_v#J5PdCKkGi>_|_op9@#ae7z%i|X8-hw*~gT+?t-pb>4Y}1bI
zSowDB*N(?hJTP$n7?;HV3nLYU2;L#WFiK!COz;llakSta&Evg-_g)^y31XZuo|Qi#
zh$k4PI}Bnh-S&cpO@pcqCTZu~T!yfi0BjS09UWa*OaZnjz>YekjxjJgS@5GZxQwmX
zVW^%yRN7%9?K5mfi4GfSYkAy)9b51~#<ya}Rvx!wZ9A{uj=kEkH`5Oh)FFa~;V?lR
z#^Y!~9nIstf_g8H;{<gak535d6AaUx6`7z}NS4Jr`6|?>ig)t5{T6kBuFr?&fAvSh
zL^2ChSfymMz!^JtB03j|;dMJQTH1ZYo)M@2vNHLP4Nv!QWO(Z~aBO(L5Tq}J;WQBP
z7`%1zv+IeihEg#CY9ruK=6T|vHjc-oP+JOzvbCiJYFl`mghnKx+00X&N80mfD8*jO
zwtsQ7cKdQtQKK-VQ5Z#AAQhX1A<aB)6$Z8n!x$ef84@nx^mj>y?Ba2r#4AtYPxXiI
zxa;|%c;$f)6<Ib4gDKHbk!3TFTLn$4;K=xJiDkHi)88er+{NQOiEW<5lj@I}tUa=>
z&1e*sikA)Lmkli`Ar(^%<*7U_Fq9YY_^F}%DZ_L(O}f8hD^0s<OxAv{eKSU{w2_K0
zz<?Kk>zW2Ipn=DYU_c{}gP}YaYFJkWLziIa&2R{mhw%IsDBr^42q=%>aU_&S@;D00
zqj-E6$`AiLb~((~7Z2s}JbwhrkMQ^!9B_>vO6hPwI*+fz0oQq)4dvNT!-irubjgO^
z4Cg|5F3;ydc^;2Tp}dsGWl&zm<8ml3=W#QXH~%|!Y3A!|h4NOOZ-eqS9>*i6c;rE+
zf>b<;oR0E10XZe`I0?y<kcR31hUCBTI2p;4d3+wp&-3^Kl3(EQMI^t-<NXq|{Sq5G
zUdQ<@x0)UUPZ8%9F7XSOjH859+$Hhb#p66lK%QhG<!PRW4BU<+Px3*B)CCu#&e7}+
zaXz}#4js`_F%K#8P+#ir=zt`Ox&#kXtP&imgt4@u0JojXRE|GH5-)nqaL6@7UrI>D
zXNE(bF-%iJq(vS1Ao@O<3L;;4AIa|X?z<HADdpX_E<r>cFijAVg1JbRi+a$eOT}7b
zTZ<f7=&Q>f(SozQ_jQjGYQFxDJ@@Rr&+Es?TjNNLkw6&<>}aGyIs+(o0XrIhNlE;r
z+zaez{53np<EvRVeXU8L@)o(hMI-6zO*2@;Cc5+$waT2b?OzxcTRtNtR@$v}qksN<
zy<?YxpDIj=v6Wcai!E7B*r^jO{G|7(XYxXm4vRw1`)~xCe5a7-DL!JYz>_Li#7YfD
zTBFO9QH#+O86kA}G1?v)aCNl*LspCuG#y5Wjx-n@enE~gzsv2$+Gw%(6sew~UbIRQ
zgdo+kf5+C(P(Q{OBULfa7bELpo-aYF5}q$X)+Ib&id3aMUyAHY(E#@TGNdZw`7&f*
z#`EP!RnGI}$iAHCE0C&!=PQtX1<zL^RVB|?BKu07uR^LSp07goRXksfRMk9Rjclvg
z<iKJLQq}N$4YIA_`C6o^<@s7PuogM9@2^9uI-ajX1M3(c<Ui`3!@q`3G?j{nfX^Wi
zNW)C2m<)W9K_Cq?rD7HEsRDu2e@Mmi(C0i1WMgnY^4X69nFGzoKKVG1Mu$?d0Q(g1
zxDfjk^0)~56!EwM`&96_68lv0xC;AJ@wgiMRP*>H_Iil}Sp5yytAWRj*sGDpP1vi6
zVVc$@16`NXr89uOkPx~|E+cd)Trxc$A=x9|;Z>kM6|B#U^@&|N3#kc-OV(v|=^LeD
zjo?tjTr%woqNq#iGQ}LE$YC4IkoXO20*>FjP25QlP#1wdtk+6_c?qzkm}Y}Xn{$vX
z2bnToTZ61?kPRJ2#`+vFZ8MdI#`?@KYs$Ze(;c31{981IdPb?Z9m}?3CEfUt;2yhW
z;m!?~RICvCRR{y<Y7-`+kot)C?qWwi!yy#7(2nWALqmZN!{NX_91LQ<I~v$WgF(!9
z#{l~nFo^nYDcPjk8~~1Nz$XElB;ZJmOd?9qKNUFBoSu{{to?5QXS&jo#0*f}1L{nD
zHdZQT0!1c|^MO%5FrlWFius^>KHwt6d|;Ii?09}aJ{Zhn&wMbP$Nu?XGQ)|`C=r^_
z(IXWTVfREn0!oBdiO`Pc2PDG5JoZe4!+GqV2q!bV5h*qzHO(Fpdm}|CQq!f4Y+jIA
z1hQrmJ_6ZBupC_n?;3%;dH$maG=|3$BGAVS-$d%0$dZk~TS$G2$G4ICHjnQh^&KAn
zfz*HSI0LCOcwCQ+>X8Y}dy@zs8NEOzG}9&}31_31$b{yirR3bor~#SKIEL(@kWnKt
zp(hd~k%x?$kO|F2lC2IhYDOljKGV4YDan)@wICCkM3<6NL8De=LbtV2az<g)hD_-3
zg_O)NqgTj;PDv>_6f=5_OlUYHB?kpYZ;%N+Ly(HWSQCt`*hJclHJf=Hf;AyL-hwq-
zc)S&Bw(@uz)@<YPNvt@D)ilFT#urv3@;C)sreIro)F34x#IQ@kXc|*U#VdmBilAgT
zT`*4<`qA`|lw^m7rwd~lz9GnN2ug<U3-<SgL3CS9f*!&1zA&8OU52t<hDz!vq+*Pr
zEQZI`hKg!KHRBT`vIGgIpDIzMN_sQ?nnZPt$ITK&vxF<(B2l#PxWGtNVANSYtZr29
zmkU?4|Ig=LerURo_Hv{}<$pc*^<(+|K*tT!7j(qJ9xJ`*3KEgO+P#~>nfi|OjY{Rn
zVqs^Mb0?w)kVGjA%fhgVP6K28F(h4Ys3dEq?(mUTBmMspZqyGEMcol3Q>+vE)d}8o
zpil=u*31V;_JCjK%TS*(ex27HRMNGQ&U8{R8_BZy#kLw*SM!Ul?)b8^;G5sJJn0$M
zTiTJ8aC^ZAVK0H(OCE2A+Rcz#Ln5G81f0UwjBC*C8XvT!LANv>=Rj=^zebcouTnUL
ztqJkSGagN*wjzfm$nyw~laMwE4QBj#<aHiRp_rb-k_@@-5SH2aB6hxrXESGV$k6$a
z;cRMidOAyjF5Q_dEqG$u<6*`rOB_i;C>Uyk;UM;@hoSZ`kF%jR8xCTh(hRlD4AaA1
zQbWhdE`2|Z7D$|<JIbYml%!?5Cje^}KBR!|DZrYA4_Qc+#d6wYIl!hZ*ouaaWI4d5
ztqlLN>XI;VX`J5xvhD!42H?v+qyhLffC&sY0=Gt<Zv=jgJRbtJA<&xXhd}!f=)&+8
zsNKTzTcG0>=*sv=sEy?LNN69)myd$lD4vgk_E9_^54G_;9}n&0dHx909^v^T(D4Xg
ze>!wahrX<%(xG2EoWSsP=ysjwuS37<Jf91-xzL*F=R*5j=)!Ow)aLPg9(2ruu8c2(
z+A^LmgZ5>7`EsZ&=lODIU(WNbP}|D$t<b)e=i8vRjpy5-V;f)pQRI1)-^?FH<Bp<9
zbPkb$hddK_J^_tO;Q8N>_BTE*`wiLuhFn<rWTZ{z`DA3D%<~tJ_5#meK=v1S{vy&|
z<oS!p{vzXlpY_Bg?gtYE@#;^J=~JYnTZl8?-PyNl{IzJ}veSSv4fJP`^(~;h1^Uy-
z`s|-;&i35Vbn+h`P|;9{INUrW%i}Xib;!1k&m`$CVNi#AeqqVIQN|gi#NlRuZW*8_
zTZEnf`~=7tJ_+$j$l<rJ+gr%#Z$;g<A`XY6Zs9zR!8itU<zq39<?(ThkMlSM;}jle
z3OG~X$`=c`nBmKJa;E?4)m-_HO?h?5<oUwrxIamTzX-S&fsdHGNP<I?`19|bXy{Hf
zj19M7!6ld-QA))q!6k~vv4Tr1!>M9upZd&i(#b)55+wh)1j8vf`4pVS=IKr(?&K2=
zyU9X?rtuZ~1q=Sn;YJC5Q9O<n{9+kSU0RU%%|}m;kctlj*Teh`fV05$EPn%_2DsLM
zVax(W(6b1RWDPrkJWud901EM>LjDH8OC0!;-2k{<yLaP(nxT)nb>zy-pCL;+VX#<?
z%!_$kg3L>JT#778krmS~Ll$K`E=LySJgz_%6+Es)7L`1%LKam#u11#C?5aUWF3Pfo
z$F-<$Eel()Scm%7F-)&AkqN&8%XaWFn(kf`jm_v5nuG<q+e~bc4Mrxx$d2tn!;ncB
zvZH&@`_^l>2M?QKLVQgf@~K0^sVN?K+pL=$Xtw(wkNbzy7WaA`bijhd7Z%upn`f5T
zl0zQYgF~L!vlG#XhQt-RYgqK#=suJ<r@KgYm%rih3|T#6W;WK}%A(GRZYW8?$4K^=
zk3TC>pGrRd)ZNvh1)prm>t*<zCilM{*wZZ|=}+B_F#63>@eY>W!DFdWNY+Ur&ykF$
zVIGylSh{;+v>@G-lW)#MvP|AsCCI9TH<s?&82#oF{k1XD6y0Sp`bM%{#m<-bh%?U6
zIgUk~CHl)`q$#>9WVB#O`Nof;Op^<Zq+%R!jRU?c0z3^|PXk{T0oDT7T0T&_1YIxj
zfy*J}dI<S4hgF0p72%o8VO8Tv)jV#%fekFHU+Vr~<f~tLfqy*4t1?eqIl3+&P$0QF
z;Q1DCcbgtS&j-9`JBVBl^1i(QPb%O&TP2=U$uPZbM{L+})s9;5PorNVpl5O<>C6(~
zQv${^PkIV^oq~SMlkP@dyOAHY(4S*p2gWS=xs4QT6#N>6&T-Rm%<R|9<5uCLR>6;s
zr~1EVy$;y7)$PNgqE3RAtcFw2{1oiR@J^)Oi7aW2QgJsj-;Me)94sh<h0Z$_9jlo#
zipQ~nGM3>NZWi}e%!$20OdAhe<ADzwEN6i08Q{YPODOaVh1?~iLg-n@AMrHez(zcc
z4i0KdV%m;tiu4?Y-cKZd?zo&twVL){NOelK>ids}N>g3wj>?;*$NuXF{mt^)=bJ+V
zhy2fHQ)rM&(h9nZptMgqPB)dg>|nO-`&;W(f2W!J&wp{-)^X?iW}IsMkB?#L)mSok
zJMP6&L(wa-B#a2fvQR#Z&|QqBVT8eo1z*kWLqc<UENeg+x*Bx;VdaW>D<e9B_z^Uc
zH!v8>evV>4hj;#j;?AQpelB(smUJCIe8kAH<3^44^B-d{x>LcxkNxy#KgY12<Jiyf
K?B_`K^Zx-+cX-YK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp.FC9F72A612CD78AC.idx b/.cache/clangd/index/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp.FC9F72A612CD78AC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c94fd94705a1266f30c8778ab89bd347e15cf837
GIT binary patch
literal 16418
zcmds;d0bD~-@v=)o_lX{t8Zx!Ewo8>(>IkYNm6!Yq#3e|WsGfzF^r{!Qc^O;79l%P
z(kd-VX=5rXsfiXcDkUO>=k%-7J@fqW%<FkQzuzAZuNmjQU*~hq_ug|p=kxiT&pF&&
zot?EDWMn2Ras6fW(hZBCjEsyD`rmuQ`W0@RjLZjp8JS;>-SwEVw#rsgo|54gIC%g3
z(}}BW6i>@9v|mx~a!hSn@t+Y6Q{oNgDl_{h*QtGKwO8&q=@nkRchftG!IAd?Rt-fz
zjr~!l=5w*pVs(kmuW;LX6Se!5e+=I8@%y>T3kLiVadcU<_U{{B|7t%kX1=oNoux~M
zZ(DFsVeu<h*RJ2^o^o{?cRuCkEl-uhI_vUl8@wGf)D$|tFZNV8YVCLL_ur>pS30M-
zzWS!-*%@k|_y2OB#r#_4k@f<`G182m|B_3!{&B_SV@+?4+LasBp0iBT)R@xwLizNu
z*5TPjGTz){czWRL(b>cI+F$%8^w+U@4}#j4)OD(D9tm&sPcx3OS)k!LV9}#@VC@pM
zj{CgS`(jGkrgWp=>r=0FInIqf5d3;;rQc%PJQbUXyCO{%Ha(8jo~)j|=w+&|j_SNG
zOH<rNo7>*I-RdpPb9imEeX5$<^pnwUE6o);d>5UF>Gx}_=B@H-`4rvegPq!oUc1!=
zEI%G{o<F1Ae#CR9YwBym)L)l5rYYa^ZZjTtPwuU@CA0KUs8RS9tM;1P=F4KeSLnaW
zX1fg34YnpeE%_lX>cK?!lnFW69$7($-3x}_x#KsrBKCF6sb;^D<c7b>>pcGOt35m0
z^r`*I8IISioQ;l$n7V2lA6Z#upx5drmv~_3>t##kE9d&NTc;_PnuTpVva{05RYP$~
zWa-enm5YyU`ff{x={A@lcj$96NMFfbPI8YjUX^g*@zQ8%TvKlHl<lf(etYF=u(NLS
zA%&PFSC;HBKR<Kz6_fEQEwBBIkKSL%ZNA{@mH&C``Y9`ZTKi~WvvNS(m+f9Z<d_Va
z-&is;c+Ja4vyM+~D>!gf=MVq(t!oA+7|nNXS&=mB*Yc&Mk()FQ`YU9+#b_QZ>K7&5
zU~0Z`RU>Wf8CByrs4jei?T(*L?)qupz1HNVUoy7_4f^6dQn#V2F=y83jA<{t!=DZK
zv`i!9r&|NvPiAK3vidH8@2BMb?l<~)-p)h1=YXBNdW2#@p6|8wf=659>R(oQ+6~<8
z(0{A{m>HD=B|8&;**)`BHeB#smaMi@u)NlposM-YFKb>I*17J+_HB!XtU9V1A^kn!
zo6t+0(o@z<Sc8|m>HfNw%(f?&&)+ntoSeWFPqzHd_lV--kZVIrTB_9YW~+6~xmN1G
zIzOYU!DE=qis^OHSNvzr?qB&LVb9*u#luFb?|*n|k(@+haD3IZx2Lv_dponG{@n2C
zjV=?tl=y=E&3dPef4uBgQZeV$j|&!Tl#5Et-SXVS%k<LY9}O<e^5{Q(ugkWG3+tbL
zQ#!#SebV}kZ<X^WdioCAGkc=VH?uy}&72eCvfsmb-CuqVi=5;pt9~wB85MD4@*-`s
z;D&E1cMMs*jdHIOrsaKj=cw57&doR`GEM10T~+<Y;=9qcGZ$=8GB$WnJG07*nwGca
z^NE2!E`PSbYfODtO5Mc71-r`Hw<g?Q;9u+U=g;+Te;oMn+LoO9$gSytT8FnPZd(WH
zMvZ;%tPyrDqy5XS5oc!yEEqhzthS?F`r(g&drhHXSL4t8d|MjJTL*(6m!p@m?plOC
z^4zSL@cPx@ExtkOYQKd0C+WMt{Zp=J^QP2y2iBd>`}yi|{VLC&6PFonKOR+eWMkqQ
zkLkyERsTM1`L%%uCoNy|S=0YsVnx>RFl$?*&Jyc~W4oSzd-?m-tpkdmJ+z*n9x(fC
zZG6hKBD2^Lch=vYp6aAiVpgN{%>CIQ)jhR81aF^U^~U~N%V4ts-s@DX3tn07KB*Ww
zK4N9vmT~5Or>=%2Oo_`qWg6TVHGhDf$?iBa+rz6Sy|0Nn4c<hDM4CKbRX@vm;`1z*
zs-}j_&YIkix%)HeZC*-mH;<|67j<#MP<!e3_HBuywr=a(lKyH{N<jYEU+ZUCXiUs@
z{f>{{b=553(c&{^xsU475^YP=&OcIK>-wwN?F(s5dU2O#K#4}ntpI9s!aT>fajW(2
z^j~6-SF_1BYd3d$!^sWl1NyI~Ce3@Z#Qxmg*rrmW&~m*^UGnofvYgk%EAJTcJU~`^
z+1uRBL&y4!JMHZIXph?7<NKFWUbZ(TY}k8nlTEnk;DwJlqs#LfRLeK?+Y+@Zb&-^>
zIW^;euV$!zbK0#H#<4FSRqotp|E9qKZf@M}^r3CIyxUBtdeglppMUnAugM0^U)K<=
z7piDEI5ekre^lNCsZrR`l;czD{)l-14qvT(X!L3O<8y=msv7v+Vf(O8K4X=ZE!0W1
zf7zDqn&z|qCK!^EkX|3L=7;d~l$Pv8>+({5Kcf?Rt#$mjgNEpNSnk-f_*zi^s;$ep
zTZ?WUbb9l6!^VAgOEwgD*o=P|wsy_?z3r=3>6XlLT#>2RdT?Tw?S9Rs;;=7IoYe|y
z|J;-OVsCO)2)r}@&QRCQzF~JP?|;$pJvMOUz9oatW$aDX7?GzpX3~-gk^Xu=+|yNf
zx&F6#tK%B{8^$mF$(DLuR&}#%$A-WgmCWVpbAcff%IB}vo?Le7qtH)x)0_pLVpJSu
z6PQ&3*c|Jrz148#l;|R-$Xz4uY(5iu_LKaY*?;(XZ(hCl9W5in9@%W{TcI~z6{s7?
z2q!76lhnZ8LPi#OVDv{zd&d|(yM3koK|dt&7|1;aro9Cg8F>D|stxAX)s=xR5@`gm
z5pcZ)7BS3v9ipPpSk)gWAdxyCR|j-@3oK$f;qBoUj2e0u4J2}k(mI9SrKi9m!jRPo
z%MOfFQwIH!$VFg&5xDdgSj19ehiBqXLwr<$DiVpLv?5V2Jp~rA@q8aMNLlh+6{sPR
z6O`5oB+^r0kx>;x6yF%#+oB94NaO}(euMh1x4<H!_gVd7apLw4WuS;e;wbYtB+^r0
z5wGvSLJ-vNs48HP$OTI40yU_&=o1NP4X<i7*sTg=k;r9A>oOAQDX_@;$d_CE6>Ux`
zgF#3n2f!R)(pz8=pR)UZzF2G5zdw*eB2_@H3fT7+Sj2DaUN7&3PKK(0Ln0R`t&8Y~
zdI~JE%eM5@{8rr}Wxykm_Y{0jDfJduBxruku^Uz)Z&)y3sElxo;*L@Fy@gCan$nYz
zp#*AQ+>NO*%%(JDFc|f74Zv%_uD8JboV++lDMX>%mzR;EevD;}&HnY5Og{+Zt%smK
zPAT7xkN9YBp$fE6AEzm;(`b%*3f#x3)y4PTOLruw0&OI6hSEBNM0yG=awT)~4&OaG
zOO=5V63M1uHs#P;V38!BFPZbTr?dTmJ`%Y?$z7o|dkZXb%QScA6<fy)WuSvZ$^a|_
zgL(@rl69-4^mu=(y#7E1iNsKHF_czs(I?WG&}ObwI9VAeBau?dvXq+ITVRoU($^24
zB`6<I1p|@DSxW0H`l6l!i!?Tm%DTO~EkYHrNaPZwbqP&lPk}{VP5Gc7^Wj67D$qnC
zQIu8`>ZPZ^A~NfjF7s})64yT`*SqTR1oOg9Fn=dlfOUcw!fCf~dNy9_h@pEp{6^$t
z=$=e+7~>GeIEwZA8HfEMv#Hj)qbFWXV|XD2Oh^G<SZR2n09X}(=~(M|As0I4!ueuH
zag0kG^S#J7nTa=<g_!#<FHC-J7IK~9g$!<Z2IqPM?ry5DzH2Kl>^}PA@0L}Ad1MZB
zA}IrGDR{K@bxu$Q*evj9D0EIz2G}<8$O!42q71No;?ZF0L{SEKoOz^|I;SZEtg<{(
zLY*^+d-an?Q>b&6GQhLTBNL!|filF_fk(?w_abG8S16C>Q}+^Oh^LxI_CWVClJ7MY
zypTu@Or-R&0pf)uN<E1hByu`6FrCuJHVh2}rJg|z61kAlE~NDE;^2k5l=fYcQ^ELD
zFiD(|n_&D+l8eCjA}~pup+cxp2nXXC;Du7CQA%<nqtnP3i1m%km_}xT$W6?UCdP#1
ziA~HDkz1K@t;|G`A9A`6Im4eR<-AkgY&I`cbGp?c%O_mOFlK&A>&GK=4C5%i$B)U#
z@W_J08zg5@FpK0;3YL=GNx@E%F9LWG5dFmh7)$bP0B@800Kf+#D{4G&6z<ooWTDVT
zS+`LmvwBSe6xu20c50Hy0dRN#G{avE1z&cqFY76CC^t5g8z*u-H>RE&i$_GG;+kyN
z`J@r@5UIwPBx)>H4Tw}>Oa?U;s{$0BQ}drwUg8h8QSNQjY?0e3^LENY<ek8NCm4;b
z0u&C@V-M5Bi$u_4BS;R1mf>)O_@a4?TOQ*s@<(RTM`p3e2if5VS>nU%InR3TTl`^~
zilg%#&SmDH#T`OBh0uK+;UERb^@b**9eW29-T@2z3%I1@dIJ<JIOydv-R*O{Vo}!z
znUM#XzQG<`;hEN&5tmQ%LM3HZNeOr`kR74TB5474$qUW2ax<-pS?4dK#C|_N8Hait
zDl=3LYoY$Q_>^|rV!5wwa&U-;hPEfZ!$WJl&GtlCy5NCE@GvVq%o0u{p0$i;XW<d!
zg&fv8hwXDJ*y2LD-Z&8Vh@1v=Pa(a{KE`36=v{c>Hsg7l`3`q%HeyeLud=Kznu;sb
zz$=s~_WEW^y6gvmZCU-G5C9YdfRd=$0l+!{*oYhiw1dDPk}ZRP70Hu=fUC&)%(wZ>
zJp2YwxW{<i6WtUPN*J#al3Q4V7S;&MTh!RPtbSP&1bLyIF=}V*a0M@PFnS$~mB<0C
zUI0sQFsm0VG7g#1tn|docnq=o6XbfsV_ebg`fYXsJNzj;8k*q&z+B|(!16k<68*y+
zV0{PJi2h(7?Y55$6%Ny*4%4H>Hx7qp;jnK>B7X+m6UoWUhr{yWa8bS(4l5?P8jh=m
z<Hh=FIH4L6{-GL9s)nv){gi4rP2?KrS_29H63tjflOr0#xWq7%MfnV7SO!D9PZl#Q
zi{w0JdLA=Vl)uM(drzEtUZ`N4Di~+6K7$>Z!P<(P$vS7UE+QARBa6xU64tqdtoPv@
zd^jSs^5L9(IHLbRPA8Dl6Z;F~h6Hki{|)4f0y)Cp26Co>T%Xb9g+rXqA+r4;Zpb0B
z{UOfi5ZV3^XL^WiU&_rX<>unRix=9sA#G&)HqNMxY~RM2wvp}IIEyyUQhfhU+>lRX
z`%j$FC$jw~&h!)6{u5{MiEJMv>l7qQj8BlPOOR~e{8~PXysEu0XRQJhUQjA8C{=9W
zp>PD~9|1$K;zJ=33{C_(Sn-i}1%s1F&I5z<fDT>=PzZukgJ7THK_M7U4JJ7p&JBls
zo(2l#jAOYN7(w9?<M@c=PmIPVMiUPV6drRPkGZ+nJ3`?J=kbK(M$V;?>stsTTpEH|
zP0x`GdNH6F1SH~z1p$K~U`Vok5b5UcGE?s|#D^3!GmDu%yKQY=TRj&UcVu2e*-@dQ
zLAP!mS-C2|Z6>NPgodiv1>#71s9f*1TQogMlxq?-1;;Dhp{r{KH3i2jXmdbm7f^%5
z(6xxtE+W|v*!YpA%OBYIlN<ot0>Etip}Y_eEaOR27D11SAWc;gbWb8p(_`rRm^2-q
z;poqBES?cw_yR|NA=#H1?aL7L{>*59lJ_zWdr7|)$~c6Q9L=~ylcp%1v5Y58RRUv~
zK=Lid^OorMcp;yelh4c*->-!EwuCfIRg6;=X_}fCohF7b9Zk&mCQ$>C-(@vZSZ%St
z6xK3DbT+(jiye83G+kM&a~5g33RukovV8$-SwOZgWk;5h?eDYB_sRD6Ig9(8B{nU#
z28(4DDK6Lupiswo)RC%R&&{akX5yt`>v+iV`UgG5uYQf&Sp~S&WMA$;aYG`boyd$7
zrz(ljP9iyzQORToFPzP&WQ&ZK94*(I_Y&7=IVF@+^Rbmg+g%#{*SA5pU%Pwx*}rzL
zQ{;sYK>7hViEq#aq+KL`2Iijuu{eBasSi!a`_oqbG_g<IMO*Kp2|pi5O9M%HKWO0x
z`!*xIunSu3A~_4%XOWAyof*|mF0Kwn+QAUN@59>puzlte#X+nzfF<gKS!uAyqZP*7
zUf?`*GaCAQiqEGkL@V`-kvt<;=?g~kg5*|4(ke2xerV`>@`>Wmmr+6)IrOJ!;gmS^
z<3{Ob3mYHB{-1+J2fqRA^S4nZ0Qqi8wT;pcC%cI@ZlX=`oQ@AJ9QE?C_woMSldby!
z5MgK+(Cs4mGtl}Bh#<p<*7czYd4F2RpYGc!qRFH6cF|J&L%a}3>jskYeo)&Fnv3<j
zp!P13v!GEH93|HKum(P&WpQ%TKC|_$tP|4G5ZWq))Upg(l0oZ=4YO!T7RfcVq=w{L
zT2f1L5VQ;;qtjq$8BFqDP~|T;RP6sXRC!JE9!6ykL)2ellrE7bIhs+5Ciyp3<u{g)
z&tTOuSVOV@VpgqK<Vp9B*++ak(c7#$jl?HXYFOL4(@1<0$>|iIPN`u-(Cv%)43Y~d
zNdcuLYI!Q)Q-K<ej(8yzn5P0;k#7R?n<S?JK8=)52YfooMSw2?YB-wWg(6^H1Z+jV
z3;4Ta{XM|nBY7XKxR36OmUtl#O7ftVI4}j!v;bO*Tmnr?NG^ktGE%-AO3F#Dhbr}C
znyVg4>q+0=096{u`lnFkDakERrRDD|Z6Vvggeot|`c|mYN^(3SiD!uKNoFL;B&RTv
z6q3^!Njk|nj3kHTTt<>hawDVC$Pj70CPt-+<W@%7%8=%pv-S~_QfS3;DuEo){t%~f
zh~!FcXeDQkT{JKJ$qoIJ<Tg&DjnfzB=M$&#iR2(zyCB&<_u%ThPS5i&h(pHD503YP
zgq_|6$L}IJ3r@>|#D2|(o#Mkz$3HYVa>J^q^H*9>&>GE2qq(s-+L&guf32EX$Ua$=
zp`~YEq8H#6ypT;xv&FO*asc$~e6qfdme!H=AEEn4=q0xQjgkJw5T5HID_msl#rimQ
zWE|VK<DE9cMm=xh;aTWYFVj;l(|xaw;$wPB0m*^P^guCC=Y^}R=~cGRYR*}G&1R}X
zln1i&@1W$Jn5aWeot2(p=ioP(JK&Dx!mAnvNHBp?OrT6ftC2=2rctKYO3eLu*z#Gb
zzw|$z<D54-Vn7u0e9c!M(ak^F+%ieh5!Jo90!pD^DBir_^8*wn^c*dUVL~FMo=9nl
za}-V6MT@%{e2~yjuJ@!-Y?w}2rc+WpncXITTn6PxvQq{%N#sguWF=)MMhB7f$Vi&V
zRTa{%h4fUcMco-v?Pi)V3BhoBFznk$dd015n_ie(`@c@LF09+VoPDfYgT^R=GLE2Z
z@E0xKZr1P7fd(a1VE_#@@CI~oVWQ)jrxx_T0;Q?#WbUSeaqb{v!T{DefbGi_@Io-_
z94zvxtn(%j$2-3J#|_Bp*n|1YpG?+7FH!*%D}XsZRK)HKWv+XUcJZt8!V{qT1Wd%K
z?X4|OramZjQA0g6)03NN4?L>)@E7&abMh<p&;b-WK%ZH`2TVv|dJcX?!E{PJozfEL
zwUD+e6cxtrhMQ?{b<%e9(}os?j(8NYBBIn{&tWm{50|K-=ZH8q(lle~jqt=h2IFUj
zJz5g_k8$Om$@I>aurIdCC?!lO=TMqrl#xp*=aO7XnUs=A`InUPOG;Bz$&0}CBAAKO
zpvYnZh08!y<U_Q=Au(ys3rA^%qa-)c%1vYxu@5TjgMCSU<O-nLB{&FA2I?QGT_(8(
zD%U_F3V95bACp|hsFg89`$|TwlH?{<t%)V-OE|3(P7nK}Ju7RqOQt_Mjz%<vGfv_9
zOgk^!<cx2Myyt~e`47w1cm3OwKYQzs%`n^HxfCf!1Z5XN+2h#{-J?7r^7Pt7St$HH
zO9KB}mINJd{VhuZg}-M>;D5_HbTgTCfWqIi3f)X*6~Zbv+seztcRO@LHA6G3zBn*N
z>!#<90)BCHpa^uU!8dqeg_k~W^G*Hr>OU^>gUhlS$7<@+EHAubC9hb*(0yPfA4tyQ
zH1aqiNH63x3P}!-RSA$KVxN7ol6|s${T&UP;p{OXXFB?#m-OJ5qAiFRw>DF@HULH9
zLOi1#&k%=G35<4v$j6SG)@{w&X!9TU)A*(bvIX7OJg_CehvjJSdv0}zgP#iBQ^l0q
zsoqm`B-nGeLlnFT-EWG5QQP-~6)D(E{Kr|k=utXWE@`D1x_Xcb)Kh^ORs^(Ufcj06
zi-39&FvBJqMORRz01gw^dI?l15&80cwlbi&pEnx(SV}3DGRL<<kmrVcDD2C)NB@$_
z%v>GiJc<|6fO?vkLGEV5G+-{iQ94jhC+pLJNjk6;>+b^fyJE(f7w!S|dn8BGLNrYr
zHN?|GJjsQ$P)PE9TDVVg8PqO=L&PsGgTu<mFtH5Um64m<GU!r9hSn8Or2-BV+gCvS
z3US1cGlXL*Ad%9kfKw`<yC~lPbsAvbsTeOjg*s13ehGD6lH3Y)T1oDLR$Y(?q<1n_
zI~jtHGLoZ=hWL9Y7|98e&oGiRBquZ4$qX?b$&5)dLmX8kGd9U&AfL=QCo@ERp2BFS
zknK~LVJQre^+;jtQkcF#9~pJVC4~%6av1F#vV9ISEQf5L!`S7J?Q<BH9I|~bqn%5(
z&t*(<$@kA?Y;ze0aXfMv=UnprYZyrlIUZr`gfNykN{eJAkt{Kums!bWl2ceo3duKE
z$qkYVSV;lNg{-8I<O){5f+beqLstJG$(5{rCCLq}egnx*S^cLZKV$Wuk(|wGWOKUM
zyW*$=MNB=H$3){(Px1AXix_5PLP;j<i*&D6AIm-|&0L90K^>ITLBb9<LrJs92~ArP
z)PvhEeRa11@CiFQyYIPFCl1{OTDU-NhVRhA9g+_*l0%FR*1_auPl}x63ToxLuk*;q
zQR>)#cVFj`zd>>qC7(rdDJ5S@awjF<N%BP?e-RM<#RB<Ql5Yd~+ax~#@()OU2H-QG
zfZr!|w4PGK?zel<i?*_SE31aT0BssLNiaubYL0S}qa<JFRIZbmkpxa9f#f@!${mu+
zIF&Mz%Q=m5GBfg;Q+Z9+`^c*Jh|$B%yfV)#_OG9Qbvl=^zW(LBt}QcB2yl$EI3^we
zL%~nh%1^d0*Oz%BhW){hJB#9{W?HYA?n_JKT^WitdhU6OT3Ziv>wz7<<dS8y-O?t0
z^4WikHgF7pOi0hQP_f4dz(;_Aco&7PrU1n&=vH*^9cOG>P~g;aSrj+S8TR1&rrQZ2
zX!Dv6Rr1BKf)|RRN-@b9j7kQ1u`Y{I$rAZaX-Q$>i#M77@ffh6VSmkpPS-KW8eX9J
z3p8=)bcg2eh+NQl^Fimh=HW>2e?E68Jm{TfHgv8pn!u;r=%-v?1c5JEq6zG|IEzOO
zg;vl#O$-<R7;U)KZN|WZU)>@o_V`hw^XKC_XlTa3F@4slq<o`?sc!BLq%|?rpcu*&
zD`xj@N46TM;oVMm-a+;WP{Vl#UZ@AM^*{}4e#O(>bI#p2p-}i-MUASWW@1I<g&Jy9
z4atp^awBCd)_<W#eW8iG>jy^dgLseT;hSS_KO7DY{Ew?+jd6&3(QK0y|EOh;t>@MG
zUVjRZ;5kluj<dmW0lHtt>E7msV6L_OW5T;=4eF~?rn>Dn7Ak32yP*-NU?mkSaWkNS
zl{AQqFDs**V$W@Kaor_SmWh;9OhzVA<B}*xlAV&MNg`KLmX(xL)S5`zGLk$9jHX?q
z|DID+YbFykL2!DIm{V+W>6~$R#gVsM_X(x*2hbOy@j@4H?jrd!aQF;}YeGJ>vk&=0
z{`5qDnuv9G(UW%3eV1aoPbi%OMR|NZ92x$e%lG1l+<@*kL`#IO(lZX>;-!!FV<{OH
zv3f(lI%R5qQ5*a;r}q5+yTjtwlE0e$gJwwo?YXlzS@MSKv|9K5qa}K~EqA-x$6Jd3
z_BVJN;F>(NAg<%9KSp`?F(z95#p=CisOzClJtS@(v_PE}kvpx{WxHI>b@<01^1b)@
zX`jm;zC%5p<}^-orq~;xf6&2cTon0z*$!z*`!#f&F8szC{3f33ef)SX!=n1&n}40-
zp`$$Fs!0Kjj`9c&WUK-i!ggI{C0E(Lgi_bM#A9QmYM;@o<Z~wZ91&m@aVA9~uV1=q
zQTKkud;Ry`=er+_y8r6=V~8JQWK6}!;o_r-_-H9UqJKq%{|v(qGBP8?$G(KC3=lNV
jB)4X=<L%_F42_J<Elf;@k1(?wCSxH=n~RUe;^Th-CQYa1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp.647408B24FB9D296.idx b/.cache/clangd/index/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp.647408B24FB9D296.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ca594b41dd1e8d72f3d65416c021f020e5feabf
GIT binary patch
literal 12588
zcmdU$d0dUz|G+!vIX6u`Dc8L%_qJ-87A?1@5)!4A7P2IR;cFjbqKxfF5e6lct<)fe
zEYX7YGPs4xk|kzr$r>6<5=Fn~UY*XF@9z(<`Q!HoFLR#f?Q_n#=Q-!|`JB)BJkNv%
z`1+d5R8)pe3s^jN`hqD)MMb3-|6kOC$k_*YXcVicEKaKp8yS(|rrjEH{*1?1Kdae)
z`X1iqmHIs0`PomSUfz*LP7#h^eNWCVC@JoYR_*P0_W0~Y`Ng`SkJ~G`UAYzMPaD@7
zw|cwJ-24Ekj(RWASmy9!=OF!MZS(&$i4n%hHO<dEs|~A}opbBrZ1JP#E+fm$m`QNv
z_(J`aDD<@`s<3*+lwMbln<q$<{u(}@-m4|a-f~j$iq&~f`ov38_bqWMEH-Tcy_pHc
zlYUCMo8wlhYSg9?Gby%p#PD3-9Ro+(HU++&Xw$Mlc))boTlwUa?|sHR*{y%X6I~g0
zKjF}pm)A1elTX>n8<w9{S=(>J<6AbLCtdNi`xC~W^B;J4cKoGN6TDVs!^$B0x%W#|
zAJ4tNS9O0>yRGFC@q@gRCB>f>+*JP*;=lFQFM^}BL34HLEE6kUPdzbWRN{19$+NYE
z{_l1*Y(2K};D070rfEzr*!+w4*-y`wS*4ohhh3c6qkl-s@n7z5Tytvcy~0Zoa=0X9
zMO4D=tx=oW#nAKCuW43~eFjH~C&#ATE{Yu4QEl|uMb|L8Z2C;MC408A`+r(Hezn6X
z&j}+hoOYAL{d0=mIusRaZ@ja9f%MI-et*33`eWy$!ejmZ*cLSI!pv=nQJddR>YX-n
z_wMxXf}=EA3tw8gHAe40_}#@NTK6umURtet=)2~>cFw!C{`j2Oy6q#Q{yJ(i=tgs5
z$L#gT%lsy_ZaLA~d^0<H;l~my&x-BRnD9@>uT6XSV&CPXiw-PKc>jIN701G0)ykB8
zDJ6{;&}`Sn&2{q{m(|rfH~yV6Z0XpKL+bLc_$BUI<FPte)nVgq)n<M7xv^6oI<GEn
zA78Wl>au&srFIGOJNcs<Gs_lK?=f7Suupo=#kFVD)f0#6GqOYf{QN)fo(Y1V<h96Z
zS2av~bu#ht#4;n*kMozE(Q~WGK9J=QP_b}zPphUmk@u>cHI|RLw#uTsWT{=V=BvF&
z{=T{)_UM@A9sVDpZ=Uh)=$*9ar}uq2PbMBNtNv)(xOC_3J+^_dY3GDG>S60brhd?P
zE7*9o=#croBm>W!Z8pNSxr=@4b1M7yvRF1?)cEb`<88vz#OKS4mn>ksCW6}Gf9-fY
z{>6yAJN7^P^=tkw^P|-l9!3#Cfi@j?#$P|Yc%o0TXV#4P)kZ7sWZ#(y1pCMBdUzg+
z8V~7duIVUX*W4~(KQB!FdG3e6M_j<C2-DZQAI>&25lY}|=I9&0yV^@!-vlQ<F^O1R
z$~{a-T(r@>8Kyr|wR@#=Xn)|qiCHJYzArKi&9<-$Eh!VcY1OGdyq>+Pt*>gLesW8|
zhs9yrqA%8OZoK4h^VXgT^X#&FSLgNFxk5bO;6Fx#$LlYLj~hi>W*SA=?g{YiA8m3!
zV&%ReN8B|1jbg<=+X3qx+q4c_o;xPBO{|UkG|c}(YH;`izabCbUosxK-SGRLs=Q-n
zW@Y@mvDrg6J81h4TJqOpF70jqynj~3i+;6jug}E2x!jie;98q_(-yZChbo*Z8`s_H
z+~i$Z6@;!;fRZ08Uib#}pYbrE^uwW^<$Iz{riljsaqG&^Uu@(fX9m~rxMxrtc6VdS
z+5wT*LsH^j-~O~$6nSgRhk#L=g*%5${A0?A)|klik|8QlOXmKkhWA%JCja>%XoOa<
z5Ncbia8vqD8K5jw)Oer9|6%8^3pY4f6{-r&zgOW#XpAsY7FevA;j;PNixvIFP=^<L
zN<0me1s02JeSWRG;eS*JjjU9-(bl77$^whUop<(xnC^Zogj#&-RRFC5L%J@wic$2a
z3-8>IJP|_)U&Tk_W1uXs*vx&>`5q!mUw-ub4}|s&9iS|**t#yscgL&?JH=3+Z|yDd
zHc%E=>@=?RRn6>)k2RnMFTM=bFGJ(53$Eh4K>vp7k>oi33Ha%mYCTn^EV{*cEr*25
zTQ!8xl<%PlpeA70b-`6!*7wrA{JUVK5Nh&OBJCr8P!?G1`rV$Ns|uy}g|HVdP6Dzd
z5ZHCWRr)QFu3vX>-bgVt<g55g{0)=^77qv(iJjTzS*p;27yqd7qmi<};-R&zM_0~W
zazqH#d2tRvIiOe91y>22mgdk<wtbWsO8F|j5?=#lfyE&mpV$6^L^eWb%!^YXN`Y2g
z7hGjh?lN)7FyCQfsLEFflmr?o3oM?tMDW};)!$zXb$PLu#LGZgVDXH19?47EcV>#A
z9xom$8ET*`uy|HpK6lM=UVsn^`8Al;XO^9^z~TkJJN?M6Z8{``-|?**fPDk-@4DbB
zF%$l$V24ZS8omc!O~U(x+bN4~aj_uLKUscI2z&B9EVW<ygR;QlwOiZbbJF)-7eX~&
zoB&Y*6n9;4l`SdZPme^LdLV`ZzDlqp*g{!gamLfAeWvBFM+;#eUR(t2i{R9*3$BuV
zcgUXSQ>DBmA->9R$#6qufyFsyYN~!$E%U^X;l)9cAPZ%I#d&?doA_XK*bWWI^5Pn(
zUIPbqU2v7tfo~QK{VTIV1B&=6^VH`_lm!;oEGeCT;cbzQDim9)a2`E8+?9pOcM$My
z0C1otJ8e|-5LF`$*n@9BM}3ZsvcT=@X52KcIXL8#5SsDgdMK@jMqL+N<=SJ<<4^Bs
zGD2v=S2+mLL0*r(TyT|!d2OpEw7y)d0ekaRBGe-!$^wg9t}h;v@M=Mv7#i?mKZ&1#
zvcTfV=`*69SgNQnyKhwFf4&k7ST4daLWVn4Epah%_c>aISZ)v0+5>H{VzXQh)XIT2
zcz`T-8V)@TC*fsdxoqT{ji%sd#6A<KD}5z4K<q08>PjDpXOtg=Q-lv51j{8b8VQUM
z?u+G;8Rulik8lCwTtHaUaox_Jw#oj|V>wS_PYdjDS<cJY%L3aK%MCRiYJts&<-Co(
zEwJxnIUi#m3+zW(&ez!20vkNb`5F6J;8|rke`9|O>}^<XxYclbtd=YnXccIW&79?e
ztb**ZPO@CERj@ra0hU{6zR(7HJ(gRfyU37Gw0X2mOqaQ_T#RmvA)$C6jR&T9Q{X2G
zNE2xO8yNfz`Z#yFT$XzU3|`TE7J8nAeq_<)&{Ixx9x}*7eaJ)?A%h~qcxFZF$}Nnn
zUJ;NL0S>!SmMaFbVw%fA-*RA2<S&5i1>ndqH$mAZ$YCeKa;Z?33OS;ChoS5+<cRK-
zLRl%~$YXRuyH4mphP{p%u#Ryj{5uN!9fcFlLH0Swk#GUBFQ8eDhRTr-k;{>v90kxk
zT#kZh4w0jggsV_c6$&A|Rp7E!&=0fbiL44M!MrjJmWxD@tTKpD#1Y(5w3MP)fMNl~
zNdP6$d<>vtG&ca$K=U(zp3%Gkq79JhFBzg_nzJCvqPYm7BEnjF*Ss9B-kk#xmjzt1
zfGggW5!VJC+F<u=;(&&?d_^>aEgtJEquv$15benTYzFYh{juBuBszdb;Mvlv&sFO<
zP~fD2xJ82&d0;z1T+E;t4{RBVL5<1+qj4vQI|M8bf$s^|1Fjx8V--Z)O~C)9_!i)9
z(VPUGl88zot`o{Sp%Wey;?^-9>zKiW;~9^5!une7=ZEY+CFkAiYQ}ptGYVHUY@T5f
zJ$6zKFZhfGd`1(o0DB*PeiR`bnVobLmQL&L2ju$!`FA&~4=?+>BJVuIaw)(r1vug-
zX1Qi)*G!g$<(e3iCdLdm>@)ay(HpnR>fh9~;#ii~Ekz)Ubqo7(e(n?jESbAKC~OZs
zcWubGjm({$Q~ZgQLNyb9E#}zGamTy2-L#H@hH&}m9*8RkTIHY@QJQk#QV#kNt^isU
zv|J9Pa$rd0av+lvKZrOv@RNf8B0nauJ|?imOCo!`>!@_HJc31BHRw|f%t*yIK=KAi
zvBN~%TOfH$a}u;jf|l5UA}$H`OM-5McR}ecXn-9m;*LV;QD}f2D&i`jQ3W){4i#~y
zpwTIs+hJfk>~?sF>wtkBG<U*rov_=%I{810Q`HZ5WfV`WB?sE%LFB+f4lHRNC<pGC
z@o9r!-!DfEyp;Hi!3z`z3>;4234!WeCkkAj<*q=JE3iNA5PJ<?LlnmgQf(V>+y*8R
zqj?G)pTcgV#Q)C!?3ZH)ZipiQ{)ZKZ4gAk6mx;nMiN?8k?eNO~XsY$iuV{bnx6r$1
z4!?sew->tZg(I;lv0NT>&4VMcD&dvkCt7jLA`>&uX`T;`uUYOfqy3opqJa_XX6}2o
zZoVqZwLpg!*zHzX?mcvPPje#UmdJE_KbG6VxNV`iopEkwys<^G+(9(>AlZrCdk#AL
zqp$fA-s0B-c0F*wyV8)@-l|Vq_geGMmj%RGKo76H!fom_G5WYi&$#LV&G({yR}NL$
z0`0LtF;wY$niCoGM5cSwR}9sB3(f6}Njt-l_6L#eLF9oMn*^_VUk;>rD!oIcnaf($
zyjhlOhq89)iCeI|RmwVOUd`~k&rl*i#)ytF=7jSYQ69|~8PP?WuP~x3G}j@$I%GxK
z-$r`3X?}|&Z;=g=Ckez!0t>=h1>&uQ{jR1BOdanNWyOcX>_Q!EqI@{aE}|%!jn={2
zJRc6TF%-oC?KmL8i^y^xf%Zor!CpbZP9K3g;dbEEPICv)?x5>`0@|NwUJ12VLaM)&
z&~YUkK>CY?+Of2J71Ulub2_w5hc3ic(qT|K3?rNYZ8K<jCbZ3@ITzaILKkdXESC#|
za$y+Zd}y0b%a23b<20Xz>{+OT-8##ih4yEm8(}%Lm(yGc*-E;86=bVuK88fc5Oprg
zMGm>BKUv^H<WNX+DH4^^^~;c`jOH4oS3`~$ELVfvYLFM{uNLXm((;Q)?;_1Nklu}d
zvfB;1eLd2vr{y=1-c7;*zVl7PlVKhoJGDU17D%~`_t5h_%?HuQgJ?7!7(N^F-d~Xk
z!j{SVMijP}c>kb&NfA1B8#eR4Vka|hCo>hdXSsaFEuXf`B4%6>Ew5wT>S*~x<n|Cz
zmfVg!+R1?=D5Rh6IiKW_{DY)2A*oC#`JrMaq?qPc$l(?04r8$<@=oSU&I&)u*txm=
z0(<Ta;~nKaB)f;Y<JWP$t7Map>RIrP@>i{2ZHdt?*II6i&2(IQvg~omD(A1B6(+c+
z_TDRa`pY-JKjKrrhPpDTho_O{qWed?<7wf&*Weg0io9d|2&Dkm6wsgSJ4d16QD{PJ
zJr_>Qg_8+4L)T_V*<J?>?ts)8w1gQ`!i*!D(8O3bF?}&l%A4~lxVp08U!NOrK1E+$
znV%$2x+q}LaH0+|0WrhLv&8{v955t)dIOMdpt%IdN=Tf{@^N=)IT($%K$fcjeJkkv
z!Al@}N#_sV0mFB|nE2&ZVAx9YW+>ZC4s$G*24!h<e&Gm|9ij6JWl&Z|=NICc0r7M`
zVHXP9MeLmA&Y_@lB+6j9C&=&#GRAYwa!--rQ<|HQVH3?Skl_oO-yz?3<eb8Ct;n~P
z=B<Ljt%6|OKTc}#Ni{`cm@G*r9MuWO;A$+l8-?wrhy5P;zNfqWh<ra1#yMr)IVcj!
zII^2<c3Nbj+4BJJ0j>k#bznr6={m5wPR`Ic>CCtPl5xfZTlVsP;I%q`9p1S-XMCTN
z!!gc5^DPt!XnefGdc?<+io`SShUK;c$L)0R{v0|!C&4>D{qVhgNl4>{@fMyVHvKtm
z<5z>d_+iZImzz4>L@XC-9%_qEpDY(H2{*-qWw|NlQ*3d*ljUYfW|`utW4SdzxCZDG
z&HoJuf1^1GSSQhk_688p07l@cW4T7qvk{01Z-CYtpd)@XmOtqAOoC#<pJC6>P)wG5
zEz@%?Lvb9_GmhqGjLtL0kkn5^l0;-mI0;FTXs$v!Rfwv84(XhuIa?sf7Ie3duZo$n
z>R{eGex+YBeO@x%x)|T&ePh;)$c}$~LfjZHblQ~>aT1~bGBWgEMuxb58X4mLX=FGT
z%PW=QT?Ez*-ec=V_rV&+awW{55^}!SKI8Ng5B)w2pB`^y*hWS|Ch8%>KBW1)Ky+R}
zWm4+|qB_Dmrlu!f8>Y!0Lb%@``x`XHzANo_+Xu@||Jv`XTPF_U_!#8N%M$Vgwa}y%
zj=}Fc4u;ivs29w!Q&WV)(vLufe~Ldp0BJkT>CiPD2I8PV5pKHX!a%a+oQ1k)p&kBZ
zmXkwWInB99FBefKqC%usNchO&62XYMdwuQrg9?y#05dX>4q(#(oCtpc(oeMf6R`UP
zT!=gtN@F3_eif9iqB#S)W<U?}AQ{jn0|pV!gsz#iJQMn7!VyHC4_))22YK*(=#vkF
z2p@;8$7%U-=zkoJAo5D6TS?L<ELRCFE1^A^xGJbyMa!$8brqBmc^c!KM(;P7jB_T<
zC5&?k&1V?rGc=bX<5KiJQO;6iUy5959#l%_w@XoADG6^`t_&HM(e2BTaXHNu$hd;$
zQ^@!f&9%s^7Io*fSneV+yGV0AGOMTgCNjH8^CKjBgbc`dUn0><nqvi`SOFzpD-f-v
zd4oW-f#wu}Wr~1O%$)+uoiwKkEK_OTFR<KC^8tb70h+S}mRW@H3m`uDQ@oTUhPbf*
zLU$4pEgHPYiz4queuPqhECo1|;BOZ++yzaDLZ-us>2NaPW+-ch&SZkxVQ@Pg{G~Bt
z(*7CznKv=r3UTyy+L=Ahr{?f(>;qswfWE}CRzP+Iq_`T1suA@=*OBNtVf=cFccO~-
zt<;n2F6={|E{z$KMw}>iOuXm%@)njfyALhy!?F0WPCQ?FLO1Tsp>HZhT{z=@jko%R
z1|c6asD%p2%bn3`qbUkk3#Vv`+7ybWs!gS6mf9?eA~hp*NuNtKmr@i9G-Cn9NkB7+
z=3_wf7|jhpvw`MkK=T>R8=&R}NcEQtHIr%1f|^-07eUP;n(HB|hduFVOWaL+JzUqC
zz;Ag~0_`e6_ug50?s|Ckz_*XSsbxO3`<Uvi*dF_oF|6b{V~LgaROi4yJ0Dmt_&+;4
zPJSYF<wpfX>1zR73k*mU{Rn710>&idTY=aW#QDi>u1$S@_5}YXkIO-NImizGFv}Gn
zy#m5jbyupZn_5nNHIlgSFzN2lwtzu=kI9U7GM$nvV6+PeUuY}6);9P>->=5a_}Tc#
zTk(}d&Mw=E>BWgB@e3B-@D<-i;94x#iA0^msa$fmToxK)u>G6ThbuRJ4YM_=jN#{|
z7OK@k9qi>4Zz9yLKpp%h0xuUoH($O0!P9`F2VU^ygAXjIf3$Aw{;SqNk=+3UQo#sn
zn*alL(fk??cn!PXoZf6n3!Rl5xAv<VaMx<XsG8B%v%k8q4Q`i)zCWS|zNw|bpOx@x
zr}$b1Pb<q^W7um<cNFm;x_ELgJ(o~^5>f=B6q2WAx%~pse!@@md!N{XreESMe4gPv
z8Mf=^{%s-2vr}66WlCg*B{JR1pve3W+d^|Y<J-;zVoi9HR#f4TV`~0QP3oJbhPX%a
zh6Dfa??K*bKYOuieTehFJ}I`G*0h{2`FrzMl^TEB!|R9Qb19-987M4+q)S?tS4Qi!
z|DpA-2h#dmK%x1Cym#MxlJkB0u_L-oa}#(CUkRlvp#upUYmiwDqTY4gKxQ`xf4Cmw
zeC17UIUjG16^*sQUx={W550e|CV7<Uy{B8_7rlHQrT?rDoJVPI-(TTydwYw*1v1X5
zO!u2;{!=5yxtQixNc4(qf+~^If1aX<K%ye2Mr~5u*eZ%IfB64J0snErH!3PN<kq?S
t7IIBJge|FRLVpbx8{57zCp&uwXGazL?g}an<kpeg+L2oqa@&{O{x6k(RpI~u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_bwd_weight.hpp.65AB95ECB27BA1C2.idx b/.cache/clangd/index/gridwise_gemm_xdlops_bwd_weight.hpp.65AB95ECB27BA1C2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..04d190babf1755748594e01ebd94225205992fec
GIT binary patch
literal 22188
zcmeI4cT^Nf*YKIDnjXYP5D6nu8Hb#MWXT9E3L=7-b6Q<jU0q$ns+cjX5m3=ZOdzO$
zt|BT11PN<G6ca&$Vn8t?hPMbsmG3#<^ZfC=-yhF8a(27s=e=FsQ(bkd?ycJs78DS$
zFG4}VcYM&i=@Wh(gA^1Ly5j$5|2k_*cZq_+EdvFGdAt4@GJS@&i|YHwUzHB04EV>?
zOjR)`RHIX%^(wW%Tl-!t$}B#z)MM@)^P@}aQoerJdE&$6=DdEx27K86>!jV?H%`j4
z4~Z#1*L3@7)B_dyxyL&f)LpAJ%+1MK=sWyN;}>-Hp^~%rOPTS88JS1@9xW|OtKITO
zrFrq%iiK{^3SyTu-5U2%C3eu=$eVM$T|;jU==xKXuz2L3D%+0V*J&D8H?wh;uY8eW
z++R1VzDoY6t*S~ZI(REB)Vw67c+9s5=l9B;E$w>z*5A(PH%oC=agS*C@teJl*j@R>
zJ;+NaovLlT@R#`fg;_IwwyPwS+I30p+uc<6cGA5mqP&g4hYTnC8xIbCz4lG`%sa(D
zopu@Z$HNCdHTqv2WOlCj^YK1|e%s%<;&kDzh2up@VIO?$XOCUGKK$+AUe}g=s#>?^
z@ru#6^iLlSwd;Cv)RC}BlNQ9k7Y9Cm?p+o<IeTux!!5`DdN)L>U1<>`e?Dnqkk!w+
zA+l@FHic-;NnY|`O1JedZ&%Lni8_DK>D-h4F2BtjHR`%^rL1$j?kuZm`3E<@oAgWL
zM)QX~C&i|3bFD4!;<<3!vJEF1QdH5Q{IubNj_=g?b795tVcSL?Qod2_c=1r_=%aQA
zkET5=FFGDxH}~MxBYTTqq|TF1J6HVTXOE1O;ukU63m3)TpLv0kY|C1tF|4-G{RaFy
zN)g=JvM}B)O)2nPaYCkG9T4w!*G9XKWkhCt(WCAP8^$i><_hP79Bs^tZT2+yG#HI~
zvLn7axx_JV($(YFr;Xd+tNv}TEgy}~em*hs*W~QUHGQ{y+&;fN#5#M-%+k{72O1jw
z8UOs`7^S<5axRu19yL62!HcxIkqZtFwg`Fs=XC8}8zm8cJ+1eg;(TDk9|yBHys2z{
zH|)eG2hSTW`8nMThnTDw7o@8y)cv`1yoUC9Z;`X(lE)q+wt!r_9uLbR{o`%KlZuQQ
zuVn1ESe>929aEON_t$wT|9n=;EGm1xWY$>o**<5I!+!sB!51gbuR*Bb>IUn<gRM_#
zmW&1Q8)N4$+$nrMzC7g6#0j2W=lwrNEj#*zU-Hat$k5=)4{VKBOxkhd_7idJ<fo=q
z?)UbN4)qxlG0?&NvggGBWhaeA<0qTnyL?^yZOPjq+s2nX?y_jYrU5h8Oq{qkC~E7#
zqBHV!kI#%4pxQ0ylHUyRrCwb>D@=cI?aFwy6Pok8Y+4Z+qIPn$Qg&*Wr~>JL#xYxM
zE<7$6-FQV;IW;UI{ABQ#;@$UmzI`z%e9*q>e%3L^cR!kb;g;qUm7!Y3i)Q%5R=x>8
zGH?8b+`*>iYr~Cxjz6+JV_Q#o#ze382d30rX);|F@vtP#dhDQtz_cLa!orx<4<06s
zwl_+<yk+0)Q`eHZ+1zQoJIUzc<e=^q6DuEdj@ayR>9?Gr?n@pHD$rG_@VqqYMDO>;
z-SS774~rT-JV)(Y_Tua<O$m<@cD+=eF=oM<FXismwq-l3<XqqCx%C=zGfq3aPgrQB
z{JZg%ew!q}oa$DcHva7k?ZHi6k{V^@<e$=pOVfkQ9ZN<Q{`!kyx1`Aye^o7u>Q`yF
z;l&lLefLvmeA7#q@$8!Sr};(u^gRYzTwF6&29>rfAK&ZB>K8`_Pe~WMmDMlGC>q;&
z?vUJut42dE+*%lyy!^tw@rxY{2Q?n-VX}IUX1L9;Zex}N$IpvX9@#7AS0%kj?RaRn
z<CKdi-fE}*nI0b-tMcCzMBcCU4m_Z<b6TAJs6dCQdk1!2fA#vI^#OVgu6>iNv`754
zN$bhgTk8uaeS8<U#%00!{r6ml&l}oJF@N^H%!xy+4&@&Hb4A(Q+Z%?0TPM@=U;H+&
zS9-Tczi1%$@x8ZyF5T)sL?_yK+Gpde_uyoHzdbX~M=zgLCEn2aYfs<8ar>0_h6jvE
zNEmCS73>$A_h)yF_!;SuI{9;KcV$d`Bo0YlA5b--Cc}Ewy(;^Uu*$>2)NJ>bVN3nq
zr3;x;8_s1XTE9~A8tplIzk_QIItqP0j1N(Ne`X^;AwYGYAQ>27dLzy%sx~QNE!6+p
z%rNS=)z{Yeub5}ruYBj0UNV1^{az+X11Ez;o#(wZ(H&)2Irho?s^vRQP02mjW6JH2
zr3rhqqi(8>2-=WZt{qi0GUUt5(Z=}^`|mGV$X#x5)Lt+CFl1cUi_iIlIlgmL6Z^~=
z<zqa}Z$zH2>zwXUc2)K#t3KLHpMLw;cI8eX-A10)^Zb4Ci<Kv5?e`rUw$(u;%0}zR
z+n>9le3uhF9(}7F*f&czWbT-8`l;Ge3+g?SZf=<$zF@@DY6J7ZC16bD;U&4JuFtX1
zyrXq>$ZvknzI6@!aB6Ap@m@{x6m8R*IODzx`#)Y9IPap`fS&rw8y45N>pPu(yzJvB
z>t4(GAvvEDUbyEtly8`SuHodzzL!Jq{+eU5Tw|+x;8^tfl#1_)d+9UQrv)k41%iS_
z%a8Tz^m=qxwUGYbrs?in(4{<V>bbfN>%WTiEmJKNX3w8Kb~v6PRky}EU75!J%`5cp
zqaf@ODee-<+Y5zG5~L93AreFyqvi}gRJbatGjR4)5LN-DRiJx&p`eJTTl}YQKWmFz
z+tgnJ=wOv*k#@64x4j@L#zFprO6#`%-UWnsDG2gz@?Nb6BDPq2YUr%<2bXIAeO$v=
zk@i<yLt8;qY^K+E=*f?cR0X|p4fFKo8MYpX*si8$rswp(b321hxQ1s)=^4^*FNlgg
zSifBbs-5Z#x?`0!KxqxoZZC+6o4L`G{>_^<b_S;SBKk`D($)hJ_qki0w{Y|dx6VKb
z*HDj?>XBo6K~w^);H6Wkc7g^l$0`XxI|1}+FFI68JuN=kMyGZLF8Crnm3nq-JrHr!
z(BD#5$4q~!29$9P>k(Ryy0#ZYWvr=VW@LbAg&NSrH8|Qj_H8{7@wi>?OM|wl$EX4+
z7CTuvd9)sgc>MIn+1nfL$94v0SgfO@quzQT;$OZFQyOzuj-Q1IzWxfJy#naA7ewXP
z`<jn$&spKB0eWJUr9gWrR%t7U%A%uA((Wbm#;Jj>SY<!3+7F`I3!<{*=7(E1rVdI`
z19E&H<E+McwjPLhY4gTy@50g^X#fjc!)~Cx8_3!VqO#im;Qfuhr8;WB7}qcx&9-Pg
z5b>tmg%|3&U(!?qTDXQrXxj*b+Y6$yxywky$3vfY?g9ql8jQLab#FZo@%H(TU0i?Z
zKcF*k!(v?}UG>&Ohxl;xh8=!JWop0)U!Vc$G~iCxRuGlgE5H2Hdq>%?8o&^%ECbrh
za8tDvL?v#@%bR|ilyWtI0ap1Y(*A}!e_KIR;xC_BJ?H3$S!%!t-^T>23By|t9pVoP
zi8GxlcnzS7YxpeE{)}sAD~QUDn>z+(pIp6@1KN13H;E)oqTcO=LMIO5QH;lWVzhU}
zZ}D#)YXCc3&t9Ot7kA#ag4B~-;u^DVYtnR8z~eh{HFkAvJrMEX?UT5-I=|v3?}lrL
zfmSgvs=Xj8>B`SVgF|~JXaGyBvI}VM!WU^Ph)Sk$>4J>gYSkLR9;+k)?Iir@+X|v`
zLhVey#voUHO<;@f!^F_UzV*-{o;I&>`pVRwG=Lti;fqN73vQ9Nf~cI9U+ErUEYVW~
zs#v82Se1ZL?FCUet7-r6A1+~s8c@e74-tBZT-pnwa?wL$ruKzJ!5Y93-$!@7?!8(M
zL|m|W{ci7~Z5z9Q0a)BirI$|Yfrzgib8k>;s<7+~y5I}M0;O0WZ7+z*b=RF={yup|
zSq=2UHB7ad>eG54;u{qwR^IWpII0R<u{cyO)Ux$J#5Y$SyWwJ4KV1Vj;0q)I?L^$z
zZ3R)eb7)q?jzcf{X@DMBWeLz;f_rFNK~x?F23QZz(H+JsgyPP#N+eq)@@+2^IsuV_
zdrRl3_@^6pY?8To+Y>P}G_x044y2CiV8zKL=j|4#0c(7jT!eCweS1Mvo)l}{_s?3L
zr3TcnN*1)rf}`3CqVnwQaMN?0mlk&hDp(~BD8=ChYb%IK!_ITle$vbQs0OUCN-44`
zMFH&vQF*1fQQDNAcUBE_Zn+S$%0fZy1yOl-W9v@$a+RBEKm)g{y{&!U)&mj0@1!UC
zvarVx4%pxp(UNFMTMr7I6veXEp)_m4#Mu-3DkyN-jy>G{kDTRr;UwsD5=8A>TEBa<
zop}K-JOy4)!LWVt4Hq5egudg21lT75Ms>RLuqN31{&QZq4!y3!VTbbmDL=W(&59R_
zkxDT#BkjlwXE~p<+=%V%Hb~iK^^U3AZ0}Fv1%q&dp=2iG1;Y@-!2#`44lh`ST1Jp3
z#S2!0tfDA+>3PY>_~wN`)j&;3;a1@eq_6S9Q2n93C=Jse#!$1!pjjj*YXTnIA%o8%
zd3?Kx;_)an_#%>@Xdj?Fp5_g{isWPkz+=m2@J%EqMv=!|(O?OXlU~Z>dB$KVkdrBh
z$HvrP8ITh@%j2oaU<HtqS&YY~#dsI6BvuDEGcZm7mSmph@swh`8(5OrhR37BI1yNq
zfx~0_XS^3!64S_IJ7k;$EQyKYg>;~s4h-pIO$T=Az?tR1bTEK&29RX{Q<mK_fH&n#
zAj@RsS->+3_)~c{u*+ua&jx|nZ2gyk>@qMV^Di%41}>L@C*?dK%VXumz@Qk&sk|6C
z76UiR<v>;rOldc*0I~{}D}k(%<tiYnV!0Z~s=sH~YT!lds{tl8z>0D$kkzvKwZN?w
zcvE>DkkzsB=fLwh@Tc+@!1D#mFM-`l;7sK&LEuX;fN~R%H33sH8}LFCuxSFlSq^9d
z!IVD%*(X;26L9?myr{ex$eLMsG&G2Yaw?C87SYg-WuIsmNI4ej#zI4C24bOIEOcf$
zFcuD=90z4_(3EAjIOt6|9?Ifb`5x%G2l`WaBD70n>raG%iERCcq3kd;B@;U@9EL83
zp(o{3C`)DK8PFgD%Beg9I%YsO%DGUM3r(q6y9{NQS<Zv9JeKpJET829C@c7$T??QW
zt*;Q86hbS?MNn46>K8$`BIr%!#ZXqv%5Or?o6w)iOQ2^7%can+6gpFRDGV%y11MKO
zSp_sDa~UsGK${BKo8^EC7)-em$|_m?O6Xb%y{Nnj%Bom-Jv69?aw@Nf7WL4MWuJN&
zNck%AzKZ(NY2zC5zQ%G5(yBpvWU0dobx5m@<&zTIlM<&-4QAy^8SC#VpE#6bBL3y9
z7B83}6Ta2|!v|h4LuL%wBYTD%kt0J+$cZ6W<jPPe3T0><8pqHCG=ZV1XevXq&@6`L
zp?M6g5-F?_wY#b<ce)>-{eY2|093+q3`8-Ik!L}a#d0G=jV!N2XdPnp5z1ou5<-_)
zE=8!6<p&5opsXUj<zrJ){u7Uc3@|tYL=el5giPR(3Hng30X{Xrmsm0+)PljaAcAro
z@TdcQD96FUaWI06d?ds}k9gRJav}67gubLFBcTWmE`kx1i=js`>_d443Rr>q5-$r0
zt2p~roC8^qkm!O)sXeA}Y`1yNgwvb;CJ_as)E-A5z6mcJ5ZfFOdy*P>Ay;gZ%knyj
z>N<%5kxTESD}FwB(ynJqQ#oe<7uqx#AT$FEq?`$yGg)~aaLWVkWRgQdHE^y5F0`qt
zK}a<SrCb9%YS^Z%1<tjsek};C1p}#m9dNE=<)47_C*VSiJQ6;EuuotR<!0d44BRQl
zLFYK=LLW^W42^>WDaS+ScvhYY-BO`DF|tS~fX)Tbg+8hR7*YU3DHlSILiSM>LFXb?
zzX*mF!GTo27&;fT@=EAj30;VhK|&=AtAv9nS3$Qb=uUYBvRr|zXjfT@ELXC;3K^|J
z##Fvl+-EC&@JNUi_lafsG^cl((<dDP38y*3)0~WOcNO;nuLI{Vi+EwB82=YV1UgPn
zZXb_XIR)Tlv96&7T^3vQw6Y}&VZF*Dm+rat*#&uFk>aB6B=E!we<=RJ&{C1&Qc(x%
zHB9WWIAY{6e90JzRg7dHxg=Q#Ybdpc#At8dh1_?cJ|y0Qdq1)`fUGE|Ag2`MLOBoF
z<{^7($ybZ*R*O5v8!xOAhpiKblMD00Aui+)*Ks4faGo<e&sk7;9%q)v@(s@O2Io!X
zH#pxLoIlHbZ*V~@_q)OMr(D7Xm2mwi=SVzqB+O{ek$C4wd|A1Fj-)TkK{=A|&hMWi
z31Q^}b0mWZ>tEQHGWEWeC2sM}$a*t!B@JazZ~S+sxs{$+5GPWI6Dg5N%Fy$8)@$eM
zN<Zv{@WWoP_Ch{i%gR&B^9nW#23-wosqbvr#nOtH1roxBO0A)55&_@^3ndF3qJ)iv
z_8;cNWDuuTtkfRNrepsQvOk0zNtfk?WMrSr@?kFIFgtcvNW4}^d`RE7w2L{hR86r5
z?&7O}?kZqT=2Xk^pWXDpqKg_xurakUC%1(J54ndi@#Zmkd-zg{FpV%L-aHZ_f+G4+
ziZqR6WRd=neW~mKQaOOSlJ9`9{c=2joG2eest1u8<wHpI5X;F(HJRlUq@IE_seT&L
zNke)p+omCVmV?qzKgy>#y;EGr>HrB@oL&~or#a)(oGEETBxG}X*{uASr0+3F5LxM3
zhwGNj{(H6>ZsRhfQidE!V2M0kys&B6h9Dimwr=+koza(Ph>_r@=ciAbcMp>7LA~g0
z79jlsWI)U~5(<%iA<HqG5JSzVU0}^xu6tB(goHGtl7>`iMQKQ$hKyKtPeUG*k4fam
zBu3=>kR1}B)V?o5J7W`)Hz60&Fl3(uug}^xNvL8X;uDcBxdJboLn`Nx0STw^!g@||
zJtrj<IX>@wcUr-#P~3t$L}-UdgBW6G`B;T9Dx>BAB%F{KpOBc4du(CX6B0MV&OWg|
znV)*9{2#Zei~Y*5@=J%`De*!UkY&*|H7}e2vNJ3{2eRjM;l~Rvfb0d!uYl|o%X^?~
z4_#B^=n<6dW%(wQ-J}aIUMPXG5|(d6*=@?MsuvSZ#Q*jcM-P8j{Jjh5dAzVdaREaM
z6&EtJSaC5!OGJuGL<}zzDK4W-OdU3AExTQ$FY|&N$<>I(<Av!a(|eOZwA+l;llIrI
zo{8_N896i~Kk~S7Xb(AjVfh=f`iA;Yc{FDo%`pptm7MiTmXCAx$GMILf_oQ_+))9&
z=Hu@32&p_mcEm>|+i<uix9!2v4*C|!-=a`*MXzPuxF_%Tnqs@Y0%@#3QnJA0g?h1C
zy;y@Ndao^VZK#-?(E6QvG$VT=X!%Y(zOeiaX@5hOR36RgL~~Y@S8_TlSw7Ag9Ovwa
z+%HUL-@^Bbekv_1Q<s%uW;Jz2ta?V=gDCPsj#xE^<$Ge)dn`W`t3G6T31_>6UA-;k
zY?rcJ&Z(Dky=eVUIQ1tiFOjG(kudT-lCFCsJ*j@8q-!F}`z5OTCCvREl&BtLIa#8b
zOqs-66_wg|L+LY12QKNrm-uQt_T|hnfCJ@h;F8VCD}h-ha3H=Gk9{}qYS52zEikKP
z<u8HDOX?<*H9zk0EnBa&#~UFdWwIT}3qPCuY)g84-;!N!JBRwucE?^rS3@0A0}h{S
zTJ<1PEH4}b(qn8OI0HyCSk3~{ES9r@aW=4}-}5Yxo@M3dfb<;8mx1&$FeE-JFXRDf
z9?Qi*y%^|Ic{z}l14G&_)j(R!@^c`4&hkrO{1Vtw{YD^dWaY1c^fk*(K-vThiO<So
zpH}*b<!GoL4fUyfJCtr`1DkPB8prY;DBZ(yA~a5fw#4Vffp;iPV&(gwbRWxyq4Y2`
zBt9^Y*SOMDmNTGw2GpnWTqw<jhQ!C^vCk|mVEHDL-ekEH8ka&_s(%Mc@38VRC@o{T
z0!k~OA@PBE>?2DnS+0lb^-!P6uOj{`(j@BwUbu>^t|Dj3*O1jUmailJI;(#J@i$nm
zMtn8WBt{earDByBad$eP*NIiuv7E!H=5X3net|Q;z&TRR=gjk2F62}TS^XkTwTR_1
zPQ8r!=y(;yIh1kCY;c!Tzst(+aq9P2e#EIi`koygvFBHF>eZ~ghEuO$IY%PRkr<Nu
z4+@ysGb#>bV6Q%!^N!}2X=^3ty^`hQod0nyket}xY4Z$C%XNR^x#5sF{E&DQ5nzW}
z>~Nk&CGdz8hv%{Ka<M}>D}TjBzT(Ex^H)opS5t?$e~6RT75}(__zsiB{gTB4sA9gj
zUp~v9xdET4cMu*Yf9-2(QiQPk5?~NY=5%bWybb$NYh~$U8BDE}mA_R8p@`|}PQhLF
zM&g<FIj8!ZPFF*^<=c)vB9$xh!coxYC>TnuRW8uU1y0o3JOo}3!EkClJ_4POz=@9j
zZLrTaIFyd&6sVH|o#<%11idc7;dFGBL!ENyL`KJu<~ZBB>}ZGo`lu*bJ7edeEwzWt
zgTvi-cH1j{^7H@kQg`&D)swi29_9Gaz89Ij7rBt@ksKm?BP|I;WF8#5*s{~Ttu<Y-
ze!K|9i=<@DKK6swsDQlmlz-_XlguMrX-nD>sgz_K;RRbux)J$jUa--$>6kz#sl?9y
z^R3(j_u?X;Rm5&VE`wUjpd0N==};@3-BNl2wVpsX(pP>0wZSHL(x2e3unC||K$m=l
zpZ2K98>*Hh<F9ZU$WDX4v^W0+Wq-lGq<2n}>T4SnP0_)MI$AnC$?fxku9j|3k`pxT
z;z84$3GaseTPKEWMLR0Br%{ndzzYX~@j-SkHXV#i2czk2Wq_~@5KcK0xMl)(%Ey81
zIFM6513b^Ld&!r<fXg714$?f}k@x*xuy-EtrSf7R6oX*OrNFq9-OGFgWRHNH4)Pk{
zS;Owl)`GBF5KiTFz_kvzQ+^C&kJ<BI0nb<L`AuM06Bt1s^d}Jh2}DqC1|H48lk#UE
z`wYI@TaAo{QB=Mj8n0*fCS&2qSU8$IW*i@cVR0~=ay)d6hwhYjK-mr`r@R+>?q&B@
z55ob6VJJ0Xsn8?!`@K=`ROn0P8BoZ8!IZP1aW=cRcLmC>Kshynh0wE*-K#8uVMQ>U
z%8Q|EF?6T=50w4Go_`y9-e%9QfWs=_2r^JwmL%boFoJRw^r(WKlpjIaqwn_`BkN%l
zmA`?;Z=fmlsx~9P%_xA_I6Pq@<wHn?axzj*W_gF$c!#(nBEt(Q;z23m!F1lcA|7;w
z<$AGZy|^R3KypB^S!qi?qyy^+7jlG7kPBZ_2H(0FsEIAYQ*p#o@i=k?($jGbZ7GVh
zh6J%y0^K+uxr*4ZwInJM3r4a=IHmT?M0$P(h|FMzawhQ11c7v@)&Rd6(4P*CS`b+a
zMpLc>o^>FQ@Uqik=4&6ET#UcL4w3Z^kq@zHBn1-RSWEsRxv?0zcj(4@<2U?U&o@SZ
z%U^8#=8X4qyK8jUCFyLuU~g(~Lz19)!O_&whRhhe;HBuLLgHJzFhX&J3JD7GIKR@g
z5!%o$k%i-MXdn#{ok3mBAPvg5k;QFfPhxAla2vVYMqVrj+(!K=S0J?t)Sa$sE0J0y
z%d5p*R#V5G7q*DIY+-r3xXX5ypNQ3-u&dx0u1gHp5g6cwJ)GtqPLFgLUigdC{EOuh
zPOXGvR@JvSwOcG_Nc1x#auU#r$^T<abaKW!{2Uv^d;?t^#ys`AJ$d4+&;N0lZ+-Qy
z5X(QJCg3Z)7rDI`d5|u-Ax6z^chZc%@fF^1?r-Qy8VPSX_qQy+=iJ}3{E>72NO|Mk
z`OeA;$6vJeMa>-|OVSs~L&trwE!mb_dehA(ua;&nIP!0&O5%Ux&|^z3F6qtK(^j(5
zAgfhga0+o6N|qWpAxxybR3s;{vF#HtKJL@q5VlP667|#60l=A167|cJ$vlqlwk5Hb
zT#Gzu{LgKfyyQ!e3^Qyi+VXda0I!X?kfYSvkkmCCVQ9<ZrHZLsNGestF(8T8N*dKC
zS-$wvEh)a_j!5d9j#7J0FRkwckey)X|2*KI2l^7L#tTn?><K&HegU#C?0mTs%678z
zTPpNVg?;J#bQQ|3vhztblvT6y!A|79^ZOmlumdQZ+WUjZ{vf-nk&kTi*&WPkq*cuh
z-ddzp%kna@^D?n3c_b|^tZ9tcobpz&>sGNl<zy}-nQh-&T+l5VO6UJq$LIf7$Hyb)
zyN=Hb-*<eRYbfzrMxPh<Bqd(aEGJ5A6D1wNL-L&P^K47FCK(XB63ibhyg0|{zuuc6
zS&=xZ*On$uzA`VwLYG+ROP5Y@&@2u*P)>v{iL5*in&m+Ux|At^-UYB9<sxWS#L7#d
zODRoONj$!ae=?=y5OyA>sZ28_kvkl@R+-Mw&niDNv>9n`ru*%C7sqTY?&9SCZ;e!P
z<jp=xCx5cU*Sd<dt|Ch|M5A?$<qMqp1<s82&wNfjpE8+3asOyZq$lH@3_omW+7jq#
zqb7>266vDka0z#J<@HXpu_0ayx~>IQWalPzUbZ-N`rZIXoFoinXMs6gcAN!vXMror
zerG{2<#RxGj@3U0?9Q?J=YaP)5JdGGfvgdjlVu8?{()U1aAn!M5d^V3sF6*2ehp-=
z+4ElmyVvabuYvb#_Wai%^feel>q~;NBxp{%RT8vIg03w4CBb0I`=D$etG^H0?PK-#
zLGOJqi0a>gvOCb6cEdZ+?hbTi+4~L*VtLRVHfg&I%F5XD%b;Budwv=8E@RIxgP~<`
z2(9lr(z;GJ2XRt3GP;heXpg&rv~IBS8_4tqvZeAAu}zBDnfAa8u}uccSHw0~SiUZ{
zxz2JSr&CB1Yj~lMGb^O=p%!*2WOo?~xxhlMV*>^!t#j(doG!h7F=tpzM?j0*qnOPx
zDdzeWa~(Mzyl|H@xXT&S^Y3v6_gJpx460eK;S6e6ZshD6IX8NK6KCJV@<xg3MoEW9
z!3$d@s#{r(m#D^5KJsAKrN11`|AZG8cR1BMoDHqGic_tkOw!GjaPnL08!mJVr-P7m
z_BDVE;FH0Qz)WD4`Th1sNH(iq1<b10?Ts2>T|-mNc$@|3QwxIW2&n^Rb*%nN5b_d4
z5jn|Y#~w`E3pVsd<6vYQ>nz7Z&v+O}hejdvD}?=NS1p2(MQ}9bV(3{60|}ECc0{1{
zMIB=7$a?^|q1s;0q1TT`?(yv6Vh3{H!E%h)BZe-Z@#eEQY(1O)6f1U(Wfuo2Tu2JL
zI9M+6TFy@1t0Z2lSWc7#CQ1bIgik#C{e;%q*C{`ehkhgveV1%_YGP%vpHg9!6lX*$
zn`jYd2xmkqn=xdsY|oIRvLi!I%1#WqD!VcisvOGDIOTB+O;Db|&{XBA49!xW#n3$E
zc?_)*by_9rM(=b#P~Hz1c?nQ1VL1jW$3RA&1(mZ{ZiLE>EU!b#>kzAtl(SgAgp@C_
zT#A%SS$=?&AF$kjPy^~h?w5r3uz%I|)*dw(@1f0mIEMJ4*m)P*>=BP4Lpa;LXV;q5
z@7LoE7GFhQHM0D|uQW*b!xBxR#tTOzsz)RpJ7H(8lt!hwy{Y{%@_&r{??yfeLbYru
zB0d^<la8m<{*ox|m>HaE23_Wy-*c3kyiBqmFWjGsHJ^$*HYqO}FSd0$-tvZm@UIi2
zNSqE&n=S90lGzmRR~YJ;()|h>3mbbHHnAUMKa_^_JWM>S=@9VM^wpyjs1?|ez{U#^
z77_L|z!zx|$<Q$QFbg8PY<6qhv)mKw@apJ1r+%KM+Top1PCbuu{>>XjMNi+J|F8Fd
z$?Lt?6t%qDOCN)|lDQ^r_(j@_j7WT{ux4|)@5dky+(-w+`~k5wH9NUtK9_Rg$1}G+
zdOo+nivK!!>6*LI!mtqC?LX|(5Bv0;eIjp{<F$R;tL4-ltw6n3AU|TS@Qf$6IU;6q
zGoFfNPsO&hYrYhFy<`&*mUGhOoDq?ea0?D(w7j5BhQam7snU;g6AZAKOA(<IQO68^
z<My0L^B#GBV8I%ow+5J#zy+}+xT2O<*-1s@Wn66S+TLBKUnCZW#Ii3<$3dSs*58kZ
zX7S&Lok9{>{d{Pa&xVx>p>-jB(FJcsL!TlTOq>|(uR^n8R=*U6l)@+?CvPd^&`sML
z^z=s4xR5mJecY*C9dV$<RD`FF&(Q8O45W|Z8?^hz@>j9iSFs6c>aw@H!Y0S9UG-mY
zI>~PeVAI+5g90Qtig!kUtbr!%@xm=AyG28N<tFO}6^EO)yu<Ua?IN+i*yC*ZAp+8w
zs+_O+en^#yewfQ2=JGppNrGm0@V5Ot1s%LkQP-!)i{u@X{cqfL+kT{iDmFozCe{H-
z65AwE2jr=y{{f@w30k=8{=eTdc^3FFByLLldu$q-Q9v^aCPTj^7dhYy%ioaKHxxkR
z<hL=fxo`VL40>BDQSX(kxi1#m6n}5-{hx}NsM{;C*DKcCui~VusJSPbvAChyeqMuK
z=ovD5hJ47Y24t5PoAb6G*&vFqcV*>RAM4rc#~A!E2EQAFub)<}t-V}z@c(-v{u|Z$
z=h16I9RB0|_O}<6jb>`U{PW*VN5u2OgQ4Z;WXN<xO1$OW@-s4Y7Hs)VBO=7XBxVM~
zt0qzBH6mRyg}z^uw^+0Jg35pVUEZ$@I@PN%<K6#p4&sOAqp4a?r?1CNRK^*UaUC0^
z_{9#+;1T8K<^ws_54OL>rY~7+lgz#ykuSE%XZbT{|C#GBjk6|98`BaVnmudW?2RoS
zI`U%_6arhg^$!S^8x_3MPQNbb9PI7m8yMv0AJDI_f;W*XDBvILA%FhlgMvZ;{n(Fw
Q^rIhx=*M9C(U*SwAIN02q5uE@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_skip_b_lds_v1.hpp.8CF2D2B713EFE65F.idx b/.cache/clangd/index/gridwise_gemm_xdlops_skip_b_lds_v1.hpp.8CF2D2B713EFE65F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ce7e75651f0290b8b7cc31a4972ba5a7938ffdaa
GIT binary patch
literal 12088
zcmds-XINB6_rQ1N&RUjBk${Wr!UBqlaOu?*2?)j-6in;}TM~^GG_ep6iHMXapny^g
z*svmEqbNi%7;IP&P_O|OjN(g^M3HwE7-oF_AM)Y<_~qS4XZFWAckbMoGiT16yFOz_
zkG8fFiAF6LyJqRaRdbO@B<hF%2dwg6WFQua%%vjHn&fj|S1!BntW$U3d79>{LcOKG
z-q!OH&Ah&V|Kj>@eTV-%@cS*N#e&Y5%8C0g7igL-{>J1$_`4@{ja8cx5}s)V{{6;x
z_qv$K-G1p;xhHL!4}#kk?hR~PG$HPp-WNlQO1=T^e=bSgb!~s;N}qRU=brI-d1=*v
zO6Nh%YeTeGDdtCLZ}oJ#X*}h~{S8Ydm1ZU9ep#|1duDx(+Ew}4TfeDKT_-tf0h)&S
zrEnt?M2AHcD^e`ye0AX=axR+Ocya#OIibhC@^9YLJT5(G`(*ojA9~+DS?T|~MgO}k
zn^tBWzg2H8JvHs2^ru@R4tfo{ny^sbE|X|9ZfLo&a9hK&LW2)p#Y<is;_pne9VYDD
zr|asYlO^omx;`T22XXIszjd4U%s%nmDbB7rZq|3(Xa3ZdnwvPzZO@H{kl}$Jtmo^l
zsJzy{RraH{hjh-+n3|y4&`{|4_xXL1Mc*_(l3FZ!6lvhSU*-W94tAP0_KmJx^0Z$j
zl}Aq6_}$F&|H<s_V)n<vr#rsQ38=j0>k%WAmyAl<J*9a^ltZwuc>mqYwr)!{O}(!E
z?dm5pW7fLuIxsK(;=?VOyQ3Dhrrkdpl+(*>i*8GJlI608g2u%&T06`3W-Ys$qc<;Q
ztft?-i#tHM{NP20j=2|hRMuumCN6@DU&5ubQpYscr^_bJd3oU2;kd-{$}Ks=swW=!
z{kY*d(fy&<w+IGV1^u#Pq(|*tPXAJ5Rj%+|7cizMHvjJug_h4%?Ru9j+N01<fj3+p
zo(rp*vQ6r{_FMIE-;fC>{Ep6_vS8eENmAzTTK}xO9(QZuo2u;-U4O|O^>9N>uYm@m
z-*4{wv)!|41N`4C&^@4Sd&l$mB<X)VxUj)|@rV`IuT}n$s}Z$aW4zs<m0wm=9vl9j
zr2cbmjDKCE6!K$=Go!;!^b1Y{DZj}4E^n^;tnJpi#ZlXz)-9PXKd|V;=#tD4=ELUB
z7&UqN53^sE%y?xmIl8>;@Zzf>E!XF4-F`PT&Hm{9{f#k`-qtp3$o=NpHrKwxEcZQ`
zT6O4;Z`O#CpHA8FVb0dv=Lx2*SK7@J!*(?rD@}v;=1;5&2%LTQ%JxwkQ~%y=QFra~
zzVN9-mVbW!iczN8PgcjOhaam*S?qdMd{;T^vAci6`PbD)eh!j6pZs7+|9iXEx38SG
z^ox??*MHqURmV|tMeOOPkA9k5HliZ$`J$|7cT=I&IwQ?1B(#2dk-YSpx8LyqzvBjl
z9w7mJoRxj{KH6~WmZ6E?tE`%U*>7f9t?i|j@3q<Jpj5JKrOCkwwfp8I$Idf$KVE7v
zC#!JeU;Qo}+g{gVT;;c}AjNISv?G0|PtW!K`$pODcHgvKBQ?_WdJ9i>&qcedHWcx;
z4?k<tKg%1gK7OY#Hdx<p(Z)JG3#&qZ7@uGyjEmKNbgIHc%sDo$yjhX{{&lP5NLgX4
z!wkPX+3EN#E_R{C=ZY72ne__a(L5>nML*{ct9AwDpW>>PUJ01^)mPPVT8Gz%%__>@
zqaUY8^w3Frv~%r*)noKN3v&!SA2?y4JY!BR>McF|Yvy3*d4cPfo%S7hYVqZ_PkxVj
z=sRPG{V(g~Z*o2aD2ku`dCxlL;m&%Cq=7E#8F_8L?SDHfD8epGx$^Q+U)Nyq*x_Nx
z7JnLt4fs`@8k)W5smrhP5ADw07WG-v^2F9X8<Op^xAm#)owzO5|3<2RLiLih^;y!I
z!bt0a3k5C#(#Yg&<g%o4>WtCPiq84=?H!als{joh)UV+7<!kfCFA3Hxj9FuO`D-M3
zuX%g*+&!XzwM*v#>@xe9fVbr@)8FVqok1eOU*BKeS%~^zPl`VS%A)IApw3Y>*b9rF
zLCG^{-dzzD!-56HO=Fh-qzx5Vr2(P_*t@$TD#i)YNxOfXZ=ehHvC4Xb^#eN#A|Cka
zj-T}uH&b1xi^T!@0rJj*h;1dg`XX+kOc&~5aj-$~(9VL0M-~^Fhn*2^)`g#8@k;%b
z^3H;YKTn$#?EEgaQ5PCx@j8Qbrkw>5yJxE<o*jR%P!|rs;x+nf<edc(FYk5keDOWG
zt{N0$aXpmO<LB?Hh{`I_&E64nPX=qlK3Js&Sk{2C-4#&@jEQwlT7Mu}8>(ZK_dxa@
zn08l0B}8Vv+<HY^1TV54C=yKim<;VKM12ty*_Z$UZ1j1#e@5J?TwSP*+gPE$Lf%=B
z>o?u<ZYwg~cu^M`VsW5B;K0s;h{L9yFpKZ|euggWkHxF>SIIjIB96cQy0pZ)_Np$l
z#o{3SAk)r*h<6T7{CS+A+G1_k7mE)9bP)9Gu82w+9MngDb9$~e)W9m?P!<k{bXP=W
zpSR76kuz4W)rM+VB@|eOg6Z8=kIJE6(%Tei4cd^$D!YKqE-<0HA}Se<_Hj2nOtQ71
z0am#L(IsfmT@jURwS^8Ze4?;TOR>so{nhf$f{0HJT;i8{e%?u4sD;JL^q0vy3nI>2
zG=V>H-)ONKG{WK#APE5jx+|hmtoDb`_{kcJw4ofU#6c7Xt-7lol?PW2#@~DFstfyJ
zmF4=&<(&l)--!wz>0O@wy*BKP#T5Wm0Nw72s61IS&S&QXU2AP9!zwQUdI_w$s~(lK
zZOM)I3_5y2O>8tvdM|P6EJS@dfY&ZInubkk^EWyKJE%bkZsP@%ynu@CinJl}U-)gn
z6jPB1p$pM}E$Gt=^1^Yh&v8zJENosl&uN|K`V;OY8#0|6GgpZfv!G!X^d-Z=3vHZX
z8|T}jnCD<L!t2;Mzzc7{m^WZPc_iX=I7z3odD8tDyzm5hKS7hp^}MiC?6y<vPFPY|
z>w96^dN<tOCTOt<eooevx^?)j#XD^7oASak#WGVe!Mw0svD~z{dnR~cg<^#%Sw_6D
zQnAvM_zqt1SNNNfZt+5ZBEXb*C0<yiSY=8U5HGA&tTrVU#tUl{YfOm`;e{xmhytd>
z&{aGp3Vca98YrS!c`Q)Ge#~QHS^W|qF99ZG`SHRPAiu(LGE^i(Q#$_1&><NrDW^b1
z3M)^Ciu8}!A)VDf1eJ%NJH7r0R32eD6Dl)VE`jnAXhN1G?jMw2VfhrNJjM0QA1|EY
zlxJAZ<&?QB7jb+Mr%mP?+XLtLgd0J*mUFCS`5EW<jOAoxmyBGf{t4uD0*$Afi<G&D
z=}#$gFGXJvx%#P$f}!FS=lby2Eum1rcW!~W;IR)vyI9TuD1+s4fXZ2}0jP%MdVuO#
zj)f=|GVK>ZR76=*?~aG{?FT<VB)mdquh1~kLnORGW^X9#m1p&Nf8dO5FC;_*r)c0z
zObrQpfyG`R?CW-@NT>$3)xeJU9VDbcrxfT+rWy$m+{g&djov7l8yQV`faZulKi^+?
z2`{}c&NGazaS|XvNq1O47jp~LXn|IwMN$JvcW^)y4S(NzXz_iifSb+)R+(TpF?Yp_
zA*cKP_(-+f4Mm0$G8rV8z;{$d5~LYk2!T!^(3SWE6StE&jn1W#Pd+rnPvHKjLKoVf
z8suF=o!QWzop*0NZ#@-1O%=yiaTa7i1l!0{A=-V6upk7;Lx4HifP@9_l?L$V3u;Kn
z1)8~_9~l@GJLCc<%6UL54`{Qj%mc2J^MO`AD=z@D0$@Pp1;C~N*s(me0F0+x1Z<0d
zJ>^o+zZB@PJfswu5w>}ndQ9ebX}LNQf*pgMNc@I`BG9jh?Z!2re+|gUWk|RV^slpA
z4NR-SU~03~z^NKIQ*HugO<)+=RglmE3|fF8<!ey?8k7<DgxiPu*I6!w@=|C(<u{@H
zCd&tr<3Z#^CW<_VSkf7P(ur;YlbXOxa$8<F1U(MHiDYT<!YR(<6gQDHqw;F5Pq;CJ
zZO7C`^cpbT5gTd&&@2F2bUX@xX#p6_@`wU<T8n_92pCa$DbOlq&7>3<m(uOR;V;L^
zC4at|gbmFI8R?Ss<m{2)k^PUc=Kt7ix!8sIJh*tc1wU;okhTI_GC#`Y;ot3h5x!E)
z3-6%)JLpL!f*0DMeLKq?(6R%Lr1Ao8PyuI8IRaTlpdOz&Lc_J_o6&>T;<zXS@)^*I
zT0l0(XLCKSkVG9g{81fOs3AQ@-jAur;Dto7TOtjA$Vmn}8r3m|UKRqs2!Y<j@wrD_
z8UFfVKnhYlpV2#LNBi3jjoMl6fchO!pz;VLi$IoyNmwJ5bRLQ5{emV3`H~X@4#mKb
zO~8z7JgS&yb`<!Aax`#>21=G)qk$XcSTH-5)!z+<?8fbNZ(OQP!7~$#CPvN+dB8c3
zZW6qZ57hI4CY84WyH?;pW}6pwz#%)J8J&w{I4&7Zpqv6-QrH+M1-hm{H!4qu<I-9E
zOz4pbJ*oUW<j+G5%88s?A{*1>avr%HV`R@b*JpHN#p8$gO{gc>_VkhM4Q%b>siul?
z-8aF>WLkM)EA-q7r;vk|Y9qD_<3^F$!fuDtJj5v|ALTTUvizFUdd(S8`BublMRLm9
zknJ|)LHQq~^A8$CIaI6@DrPuNtP@B1i`&V{q#0fTW-1HNi2{anhNFQ_G|RC-Czj<#
zz&C>abjF&1tO=M>ZU&akz?t$}z`ten+W_CjatP!@V1GJrA<!ZOI#S*U`HieR6!M`g
zM?yXl_NN0K2`wU_BjqT_N3rr~$VaoB40V#BA+?AUsFT8SI@C#LIRjc}z~NNC3~H4@
zeX=d_LK(CzgCi*4g|>HDE{9s>tbPU5s$jVaYE^yA)>UvgZLb<?RkQL3Q0oE9*_?GY
zH=G(m5*nO@tSRRp%N*oPxf$s+BSV($n-SwiTaZo*D{n<Qtt_`At#+hOuWv`z?TESm
z2c-3Zm3JVm4wh5JhN<GA<odCrR~k=GgeUMu5P{qyP|s4F=oGuWzYw(z`|EhlI-XOK
zD|q1;XLpS2>G4EwC+TIc#L2jY1a3$I*AtF^sdu$#nVKWEKOx9G$btC8ua{;yP0&bJ
zg?eFPtuV1IX?WUa#WsG&^i1&`BY`v$*wHnT1f@yPj@ZDocZoJn^Fr<a^#CwyMAB#J
z;^)i$$2RYIK|<eq+iUyb5mX?B9tl<ET{-Qstj>zVe=gu03#d(gr<oA2t=V)a9@rQl
zjbXQ&G$>7@?w#cG@E+Qg#iP4gGt_9N#}$&*!^f4blpaweZUnc`mB=FkUigDE{)00o
zlT9*uxCK=@?|7<t!NS~_(<A}WThf`oqgUsFS$SX%-N*8QO+Jmud7+qdEasS1_6T`D
zVuOVQv0DNS7Vwc(?6!;Y##|rsh^jL|c<79*j6BGQlC&X`bfybwZ(4w53$Uj?w*`FJ
z0;W(7fo>ttiwp+Ij^II2<rZl(waB}cb~17t_sgrT9nLzuuugZKoCGwyuwHk)oSf@;
zAy79^PCCX5!Mee65_+o+;%2*mHH{9A0PQ1S03EfXK>H}m<-n{QSd$|IkB`9`)j*eW
z4H#DgCX?Cbg%3dE1JEVU#p4vFWjJ)8zW);RyadNn-+2j6z67UH-oj~Y;dF@&;JuL3
zh+;VdX=NZOl^;S{hgkj#Y5#==P<aE=ZlD}l?|Jw8Z~fnY^0UIPk=nMYdzPQ>i-bp<
z<0EQ!NO;USK4!UwbF5*xj&rPI`6=i4l;udVXQX&Ec|s&aipNBX$FV#qQtZR>v`Fy`
z!dolXI;)FLzEat7e<Oj|as0Ozo{bz}!f`lWL^_$sgy>^WfpiL3zJj!_(A|dkY`g$e
zX>V$|!Op>>sM$3ErzYS*%|8-4MM4i^776py67P&t--0{!81j#yF|h|?m)Hk%Wz9+Z
z$tmVf0xzs~`d0@}Zh==;S5lp{pmGSt#avJHo_2CG|9sJHoRz{6bN@_WN$ySd7<}63
z%DK}PQh;R&m`)edHqLSzH=PU}`8)1>S0<iZNS+S&sVo0Z1U$|yo4<gbB<*}~>aCHI
zGd~FUDAEX|jX<CS*bJo2ESEu98MGlS^TJ&yyG!}VnmqBOrRk#w@WNXleG3${p|@c0
zTVTnu!&~4=xeZ9$Sp7C&(Z)KqHsIL?CQ$v2P`VK+r~?Rv(omM8pfn09$h7c6G?YfO
zTn=UBa3Fn<ayYCU+ET87vI<sS0fh?aK;_j?R?W&EK-mM96FEMS(;;)j3%_yvZ!ACJ
z@QQVy^5>lJoaGjzXhB1$ycH>0S^j`zAJ9N5??AE+mLtV>k>Z|E=~z|r#oz1;f9S;v
z8=&I`_%$842<Q?4r_+JD1Nl49f<EYd$lqtV328O4QRRCy>OGoIIzlq;*ydC(bg0GE
zA@4f6!;sYfFiGb-9jbN~IGzR0WKHlmgJzLS*NEy=X_Lq9w)cR|J(@1%g;&7w6?+CL
zhBn2}jy}ys&T%94ayZoFhDWfO=>~3K17}K}h8L2MO%h^`=1IsQiKd!Va+f63Gt4S>
zO`@p=6}u-P=J1z<ypm`pTO}Wx#Ad#e(8MI<P4AzBTyog<bC7Ee+kOsm&tcooL0&m*
z`#ETA4%>bXnwax(`;&7}&zm4#$U$G_psDo!50Li*I&HjAkKF3n%zQl>Rgb)=em(l4
z9*t#rd_7`vYOj#{E4IIJVz)RN{qVwevD<c*Q^jVf;+|NI7gELeUr&Zc#Wty8W`0t|
z4yodvRG>=kl1i6>id|F1J*glSyQhko`AHRfrHXsrG@N?5?vym5@z5tvPM`hu(N#}L
z$zz;O2nFh)0RJYc>!XUK)OWF*0n{^CE(hx6EY|?_8kXyUdOgeU0eTNKX#3$1g+qp8
zp?WN2?q3Aei&%aQ(QB%oH^R7IZFp-m-p~^epMZMaKH`^-NGFSK;Jk1HY2Bd5KaxGi
zj-u<`E%gE0Iet4QBQ}(OxorATWmDZJpM$~<<-Lw4W~=^BbU;pfxT{_7kZD(+!+y`<
z2=a2J_}0xUSL&M!|20}5?>YxbI^Q<a_M(A%G#Euf0bV!_98a_E@C<M~!*VY0$purW
zySxu<?z8UjHE?{*y2BLco&raa2T~<Kl^M{L@<FIP$ns&RJk0V@s65K@F{nJo@?~gq
znRU0(-0*1Dy=HUDY|fQTEiatplqXp}%_&c_e3nz5WjTja=CJ&6(2K*mkAhxa_&Dh0
zg^z<?UidiZ<%I@rcmvzt25wvfH-YxIft%UD&8GYic|T;`;~V7uhIMZr2hlj4@KF%0
zVm63YF&jkVq^Q_!2kQ>w#cuJGOKM{ujA|R}iY+#dGm4{0#v3CB2l{x+<3IVDDV@4l
zukFVa1@3`0vhMLBH@}-XTk>?=JN)O9Nbe*vAR|T&bU0S+`Y{DD{W7b+?q`4H7l*IN
z=j8b`cT-*${`LObW&jJ)fp$6=O2XZW=47Am6C*bN>tcKCd}GwxBYNPI4+aTNv7hMr
z83=heUU&)ny@Vr3R)hQ<kAmvQASC7Ra!i|d;-a0c*q6S8qu)Ws?%Lt#c9uKf=N+);
zaQ5mS!^JOdeLWD5Q2{3{;CjMB65!(=b^WM>+!+T-(ESzsn#@aMa!!HOapOUse2g0F
zA4k+(dVJvj{nn#N>&0K8TfFUYk3a3}xYnQl2<rdtcg;_Kqp-1O2bG*?`5(&>zI~T@
zL|VTz8o$R2fzl9YPLCduP#Ou%>CvMK%Br9h4c42HqL~f8+mWmt^#p%^mj&D3Y|J{3
zt^NZr`2d_r!|&ehFR-eLZ^nWIj!&TBe0$~nc!N8?4#I*$Xjllxl8w)Q;gY$kh$z5+
zen6<|%V1T}^)V-3M4}<|EB+xC`3xmrL?R3N)r@|%q+f^8ulOhN-5)3%Ic!jw*6=<?
RVr4pbh`Gg3Gs|J3{{l>5H3k3x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp.76232B181B348ACC.idx b/.cache/clangd/index/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp.76232B181B348ACC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..da1c489a1445410b45618f47b50d55b64e906509
GIT binary patch
literal 20972
zcmeI4c~BL{_W0?Zo&j;VT#!vrE+VKPF6<znj0;3uuAmXurzUDNYK)Q?6OB<p1Vs=~
zgNT5*M3Gh8m*8$hjJTo3J%S1@2yTc=e$CSKvA%!2di8tXs&A^2dT-`_IH!ANdO3ai
z%o#)a2LzM_>*(~E*njSfN#9H$IyyQ=^#5Vs%%1v;!WB~;ow+GzhmW1P?h}*sX-E6)
z`qF+{{`0oO`ne4DnC3laV~?Vz*QfZ7x&E6FE`5D(`1`81Nl1$Myt7qq-Dk!79xV9!
z!7bt8^Q9WUQ-*$Xk3^Td3=SRom(RP$BQy(AyiRvmypoK6f5NMOo%aua_}^^fH*L1V
z+3M#~&)?23*MHJ$rGYlF&97PGuPn@UoIC#Tm4KB#5{fwO*(<V}kzPWrZ*l)aIYwcg
z+rQX!^=yK{qVD(cbNpQ!!_uDOseMark^=mHJo>I`-&seS^7*6Ye4ik#o%*8FcvXwB
zgInzw(za^->P^+{quzO~x_huD$Es`4v8|)`j4A!>{k*s7cean(Wc+K+pD$8d40x&c
z&32n{?Qq7ZpF&m?ACFiOUGZSGZwup-VQTFFblfV>ZLq(u|K3m5I9q>PZ4viF(XKDG
zH^Uqsx|T#NAKhz6aly@E2m2VkHzSJrRi9hm`oX|N&B5BI_C^!4rk}LguWRb-&@$(b
zM^7DUzw|8Fb!XbAd80ee%~+G$BkssIYkz+rzu5Pih2@arx6Ec;80@<(>Sbl+=DVkR
zO*FXwTZ(1Sknd*foRMNR%lrJ?jBe!zZ>5ikEc*KO(bqfPpDFD-Aox~7VSnEa3%o+s
z-}Sw)%uPO}{yKl3eG7;0UUeNPO#Av=ZA_lo^K<Ls+^@NW^%y+P>B2xU%haf4_QX|7
z^GknAJd)ULwAS`f+_e@bQ1s3mhr!#eN0RWx`CZ@b(7P6t=3Wt+d@FN^nR`WA$%8I~
zszw+@e}$i#g<t64)?KH|P3h0C+J$%7w99jHjNhx$Uw(eIr&HLeA3ja&9Oiu>@|4Fg
zhp%rA{c7ONTxrT7_c50;EW3)&q;In`9+zGY^)fJhIAX@AFH>(FaU5lTGdnH!>pxmf
zn=-k)aA=o34x@Wa9ikfkN!a+m5<TW*4ScZWn->2n@*nnH+ehc)7QSic8y;1#&Ex6W
z7rs#!NNU-kRo8~^yLk3`$d%ZaW)lv*nH#lbPF867&hAyA7hFpIWz;3<?aSw_9KY-r
z{<!s5?Y7Kb*I|mk#h}*xZTh$dXRpkdcXP*&s|V=|)xOW<O)X;99p0p~#y>Qq-TZS}
zkB4UW?I!g<H0`0avHPTv8;-{9n0mu!<l4~T0dv-Ojn>h&-LmRv>!@yF-zDm}XuGe=
zwOlyP_vA*4Q%27dO(*{=#&6tjlMdH>dUwMFbAx$P<}Qx>Wy;*h$OlvAE{gnlN<`6u
zm>;%2JZ`b#fZwp469x`1={71iZAzK(w&Epu<tE>CoD?!7M!0c3-m>uFM&DQR%&K=}
z(8S9pM*8WGi!Hw5*7suHfJMLly1%Z6zji>c;h9@qLcUn>uxM6I+Oc_WrKO)`Y<@Jx
ztm^9SLuKgdnWVXPudd8FUbyJ6o!#_SZ}SIj{_0s$vBe$JD0R}4rK>huU;Sp%qdnQb
zT(^6DN;NTTe%5LibDysGAnD-w_}nCuwjHMlfByAF=0@K!Z_wx6GA^Y~>XWgcAnfGw
z%JjH2kJArKx)E*rtnZx@Zx5@RKS}lZ(FMzYDE%Vi^*oc=vu^ahbSk^8dRymiUPWaO
zwuW!Lz2b!FgNq|VV%L7(&VBI9wjJAC>^u2%PjzhS><zm<eKKZKVfNQfKK{4IexW*7
z*t2rPu89j??wVaZ>e8Bu0sqoI$~-i*#kMw1Zx&5TJbcEwX8p#Xki&i69JM)-nmP3B
zYqRWUPH(3xMSm^ay-}}KO54PDuAM&&dHKcJ@^jsKT{epi`+mk}Yb6~WL3?EPjG*5-
zs&F?q9c3X>FGS`|CmlU%AL)O>l+az-ITI30v5BjWGR|^bn}!1u>wcnhrPaET^GtC+
zDxP6E!@l9rEDlMI+VF0Gg&7v<1(qVcrAXa$Vk)f`-`HImyM3oAHm5awX8Bp0h658@
zEVi8b{SU@wrr4T_r&>;J({Ny7%io{gnCid9#1val@#mJGw`n*qaogn)E6jYS9#G@X
z^!i(n!4_oPbYd!YOTVyvHNAzSF>X(5C`ERqD4^-YR66;jI3yf;f6f?ppeiNUt_1ru
zotR4Z$x%+@&JFcc;jZ*TKdFB*Z#XcqYf*A}tMP+Q7-KzJ!waG73!#70iK%#6JzB0F
zm-MA6ZcSAtT25@!aA+1!*WVZIba0gl+tLdxLh41ds~S&CMcG|^&-lk*Um0T~TEiXe
zd<PF~Ix&?2&XWyGd^>1WxC^aifqH>?!-0th?u#$=iaFWE6r0f+CRk2r({NzoPrv^$
zb7z{@7-MWl#fjK15eGD#n9A_sCf)OcQYuuq4ZXmR>K~074oo~^d&>%P``mIB?nK4&
z)$@%S4oo~L)iAHq&A22L?oY+vn|*KHaA4vYS>JWH`c}`u6g$xCPqUoXw&B3Ub8uN$
zyR<Af3Ab|5QO4?yb#6H5v=9&-$Vfr+kMF$oYUtJ>rr4C0A8$FnO~Zkek6gH`kIqH?
z7RI;*6`w|Sr;&TpiK#4`v1|6>mfZtdVu`Ax5&bk`-gIIrskd>9m_ENQR^hhv86uE6
z0$DVjm`WP?;@Q3<=SQ@}#<Z3&{V>ahL$i2K%Pwb5&hRwGEvdK=*%czUrV~@y5`FOG
zmAOmr8)JQ{QiAMCs7m9Bsbt!;7<E0<QD22M^ghCoI-Cxe#uHP?z8&qib$`%p6}F)&
z^VIW<8V*dH<3D|Gu1jo<F?OOC*-Wf96YHi^v&z88ZdZ!zDvYsARgNRO<McinPfTU!
z@)`55+jpVu;YcsC7^xT2`)E8dmHioK^=o3sUNyz-Xf0n_e%ZF+z{J0z;G^?By~Y}2
zLn>ZD$O_`!bYd#U!rwalwc+>+W2~ksj|IEOLg%IvQ#pC}xb=|KV=qmyB~_VZIjK#<
zfr)=Jj63wi^>cGmY(>SBEho2WI56>9Q^VGk9W!PaV|yxoCy;l7YtxCT{2sbCa(m3K
z!?I2w9e%cYw$2R)ofZV^3~;8y?^0CP%c@z{t`Zh$9dlaFY1MFG<uCW%`es1&wsK=^
zK*jMw*LcCF>BLm7R7IG$ZLPeZ!U41`mmu{L+LnzcrgD4ua)<Wn-X$%uiq?`&^wWu^
z>BLm-?wis!=-PBT34>@YKd67OXgD;B|L~jCVZ`kvEpc~R!z}$-ts4$ZTxn?V?O&nB
z(Z*PpiuWM9J#>~do|wub$GT;%w;vQ3<Bn7%nRHDiflVi-^5V?eU(+UDIH<y((0iC`
zHrKl0z{EPUCru9P)sL@Gf_~Gw#2Qgn)+5z=WUAY=!pKSvvduvb%;K|k%UHi*y<)aO
zRz3*M9|Tv{pR!U!LW;;RHr8e30tvoAhH{RSw2@LU=e3gKTFHrXn&g-UIaBJDDJfi@
zDfP*e`aurNl=?#+kSPt~oF(<ok_K>2()CNy?FCs$((MDeUy^Ph<o-#zAC(`Jq^pJU
z;3VDV=aUY&+xYg&IixQuRmiCd1+wK(R$if2uTWRci*Tz&81OROY8m7#?39HA*%~1$
z1-MlK?#lUB-0D}zr*NxNkRym;1aagYvqtK@MjFd*oH-g;zrn%a{dp4`-o%z{iDttA
z={LC<Y`o~t3@AO5Tofh9%2ur1ibwDpOBaIEg;8w!$;x&icspl<?or8Ku6DleCM)AK
z<LueEl9lnA@%C&*k(CLW3HEIHl$FmkpV>3VMpi!8d~VMi6<L|6nP|@zds&&JnPksQ
zldMeEOtxp1NLHrWOzpt@A6counhI>s7WsO1ub_TN)1ogcl~_{=<yBZy^)b6wLH$~+
zsfF@d>|P7yZ?NVKlz+gQ4<EDp2dE!Olt>cD`)?^xmO@@d%vX`N%$Uo{8e+Z%@;#!t
zN9<XD%gR0CevkNYE+d*UD6b%zijUd70_s;2O*NEP6Q^q8!t1Lhy{pM5oNI`t2I|)k
zj~e33<<E)cIh5BCP2I=rR0r#?BfaZj{o$e}Tm;V-F1m+{;Q1m%O$3xji<;<<**zNS
z$BUYHD32GN;zjU$@nY|I5j<als7ZkO38F`W2%axd)FeWAvZzV^n4OYg{mEkQWLW=7
zUFVg$o+&Z&tnGvcOAK~xE$k$IeaKLziv!t%<c8e@J;>CRA%&2SBXS(_X+%y#E<vOO
zaw#IEkP|UU#Gw8XOiCc%!Q>9)c!9(VK>x8o9z%X1kQb2O3gj*17(!wQD4$G7GG{}J
zOTD{ayfqCI<t#Eji&`<ak|;4kzZfBa%~)n1h<?L{hg-D!;?Vozr+nZ=OWJ6-0^O4Q
zZ%MtG+^qA2-%kxbP)7}01=3caP;Rl`qhasS7|w}8aH24h89HWJy689TsQC3~+Rn6N
z=2ljUkz+A(Vh<rJ7m?jX)PXsmvQmq*wJ4ZhBphqQG2j@ije)!iYj@#bZha49?P18L
zvGz3NKXJ!Du>)(gdb1Z)LH9Ml?0Hv0GgOKFDse9^ufjo9P`?%j)k1!Qz2Cs{Z?NAR
zSpEYJ`T+HF1=n1`oi!p|HU)<&p)+$pW#yUR{7i7=e1mknK^$0b)tj}Tdt@NzGU8nZ
z%~l!lD<fdNuOLAcP`{dlR+Evud<_{=Lqa$|C*IFteJ_aS1!>3SbtJS7mJb($!bNaD
z5u$elG&2#RUxWzmCt3`OhWhbhXuJsSCqW#NAcFfz6ulE+eXB&xDiPdIvKX2S%Wn`z
zZV*TF`^^_e=0h$L<swnV<qt&52Vxt>7FY81-u<%2#fm61?Pl7uVNB>S?o^B%*f6f2
zo9)jcJI)u8`9);GhBcwX-1Z`B&$$wNR$?zUtcdakd%l4jBXo-qx^sDqFgivU!#P&y
z5G!=#oGaMn3ifOm5hYh}&J|oZR|!s4LKilS2pwkLRYDKQeX4}MoSzBZo(bJKml4l0
z;>DVUC^f{bhPXo>Ttk}OLZUn;p3k9vgy<O|g7OKXTY?D6Cy0X+L{L6a^h|{MH$?9n
zVh?`3`=axGzIqVlfoSnSY|Z4Y4Si4hZ#jGd$;wI6_ayhfncGDD|N7t#H&V+a?J{T-
zZ%h8Sxlv@h0O|%d27Q<t*!JmUn+3!Fl}9gBCUq_2HE1e2?rt&XPA@93*0I)StCel9
zo%?FMO7#C3Pk*cN#FitvTWE~v@qx3Gblypt_gnV8U59KsQaFTm`&YJKIWc>|wi9&N
zHHQ0m`83iajRdnmkgV(__ItUT*u`b}u0_UrHnhH3ZDw`8+q9SNGVz^HTm0iuLX;%I
zE=lOXC;CRgZX@J8!8%W{;qqwRz-Zn6%&qFGofzn{{k!E7QT7Vndj%h^SSWZGLOvyU
zpMw0m;Qc$~QlU?&(2tkDDfGF?nFTay3)Dw5+5C}}6r@Ul+wN>+o6Wb~&i7Mutj3?1
zDb`!qwr7zwx3KqwuJ?p){CbOs^&+BS`vsybCf17~-xo*R=L>_&CuPf|)+0Ui>s!zw
zNodV0+9+6Tgq$Z><U!smnC%tJxqhKwRtWi&V0H@f?}FLykV^%#Qph(2vzw5kb?u{d
zJFp&Np<Zfw8sofd=(s=jxHkRtv-+WCLu}|9$HxEIR*(+W`k*h<mX+m_b~(3Co=x#z
z+8>QkU*03BNR<kEBp2D{@*eTB|4e6s;ka*w`W~@3i(2y@iPh~GtJ{g)I8kDCd&KJc
zF!t*GbMJz>c834;uJm?|47qh|`%5A#)yS$Ed9c|kD=V<o3Jf?0_s+p%xwBY=t%|S*
z;~v)4Mbo4S)#gOmip;hm3ug9-vJIJS`<UBrL-t&riOe#gJQKCgg!1jkY&(>1NA0&m
zc@{Fug7Pfnn1!5p{n^MY8_KhhV>XoMAhR4O&q0nkP@ap-a-lpIIp#un9x}^=@;v02
z2j%(5EFa4Akz+oT??7fdpnL~%-ho{C{TCp!0w^y)js?8D&&;SVHdjQ=qCI$sxE&&W
z**Z=)cf_p_@)6SY2<ge?36g7q)EvI@Rrwts7hv~2of8SfFoC$R@PPl~R?@>)84gqs
zBbder7R+-}I-Khne`INYEGt(;t1Ds`UiB@}>K12Z*_j?sZ-=dKxJR2q#FZ-+5}QKE
zM~LMS0xC+7Gzn7kPMTerPUyyWDr}1O4YArZ|CL@}y2U+7Tu+kbr=lSwQN3NX;aAKS
z)!C2_i0T874~gnSkS~bN7es}ZzbHCiguF*G*(2F;`4P$F2;^&$$u-Eel2NS$obEcw
zs1EW6N&P_r_q$M6y-*kMVqNuO&Vd(GeAbQX7Zxllcahs&)RRx(yC~o;8p638xs^kC
zISMF;@@k~6MwVQ^2B~WxKS%24kZX~;_G5ObMZUaz9a7gp`CFuZ3ppIC!!fvCICcrg
zp!^c7UIOKjSRDyD2CHK}W|tUPJ{GHEp*#+&;~=MEw^ZDd88BH%#Q~{!2<Hs!mI37%
zI3NScbFexGTXOwetj>j;ht+wIcVYFekJ)7x_T}aGVD%m--;33IAs@!-!ymKDVOahs
zRv(4(V_1C*@@cF-{V}_ohUL#-^%*EH#_D3om$BPr+>^OuvT_**T*gB<U&C(Kp!^yR
zxCZ4F*rWnmvngC3BsHmoT!l@lAlG7(TF7s($s5QYu*nC=kwlIpX3U9^X#})uB=O?B
zlyqGRIf}?pP(PZ;(U22~oJh==6|8?dVOT_(w@LNU1Cx8inos;PVp0aVf|yi5t|lhc
zkZXuZ4dgmvQU^I)Gzk|${Sl%`1mtMZBpPzOXc7-OK{QE#oGh9oLr#*^Ns=YM|759i
zvedk-l$CU;b2{XWl6oW5-z2FwLC%*<@+EM;`I1|{1eTi}lF1GzFOW<MARmxS4t&gR
z2VnVwlF30RKO~tPf}Eslk)&(Oo~M7ntac-l@Q(WBOzM{)HAi3Bdw1%+)xUJ-zS}ya
zS_gf%EM%L-eK+<(LDFw{56>1JSvf;O&cK&e=SlE+zG;?~C`lW|eMecjDf!>zyJK0o
zBl+LqJjgR)rdgMmm2^!#Erp(zK4Z^A?Yrc4moL|Yw4Ua_^-Bt-&v#iGe3@T?85HWo
zHl_veTis1M@8<h0mLEVHqCPo*H$*a0B_l)j`i8EcsNHsC%N&y7Gx9wLsxm_9g-=N8
z6MS(V(dv}*_#6v+xqc}!zl&V?Qc#Y}%OR&?^Hl80?o?JXuz3dL%h>!fc7>+J{2Jtk
z*!&@O<)-8{Hh&HIPh$Qj^eUba^JkD3i{^_(;DxLd%~x_}nGv*M>eC{4!>mQBwXk6_
zQM*juFeBb3Ip04J?)Gn^&0~DmwaO45e)V7X=Eo*KeiEpDnL*>v`;h%U<i%bP%gTOa
zzaMgyU|c2Gvl!LbJ>JvDbg*kr`*e=aoIv)(<Nu8MLb`2pi~f0Zr_r8lAbqx=ON&2N
zl^M}LYly~z?Vec12pxF!=_0HW`gRGa)*v%Bo+b}(yW^Jnx*@&EEm*q+hw@>t9}nA)
z$8b&+f>U{r{0qaM!&bhq?<C2}Q*8be28!Qf^Y@S+67z>J@cM?BzkwVjnn%Gv=vvWy
zE$8V+e{op1{^jWZc$>zOVyJsspAo~_i=FA8>oZ~UQacw;+8@7r&Vatfokw^Kv8aDr
z;8TGH@%d7Td@3P7LOzcmKSn-}A-_ld?@<UZ{{i`bfV>6!Y{7$g!~TMOeu2Cd`)q~0
z4f||^ydV4T$05A@0qlPOa-!guC}`QXURG8Keybp_7W`I2UL*Left)J%r9w^<{L&z&
z3x4U4Hwb<kAa4`=wn5Gm{4ybL7yPzE&Jz5xIL}YJ+{$Xr%)0tv?mgOvbr*Y~Y^mRn
zn#BHGZ{0@Sh3|68QE)jL%3D7bk4lBZI0FY~;Gw+rF5^*`@fc{`g0JDBoGUT)jrwrw
z_6E}sH{ibn=f7auL?SUyB(1rlvYME$hFn5|OW=xBMwBwRLwZRxFX3y62vLa;!47J*
zs96nP6WkZO+!wp?g=wjzUCP%VS;?0AXT!9KV#&7{F63t=-?Na<NxtVGpO<{kL%uHg
zU*{{NtlW_NZ*X3`Kcw@r8+#Ve-m~du)0<f%7Fwa!s4;zv_umC%aRE87_X$gDk5>!}
zejP!}I~h8evD6_JRiu-)F}IABUv94qiTh(>EY<&bJG?wdh<FnBmsdYoaa1{)vWjRl
z)AOsIESgEl|EoYi{l6N*X`bFZLzW39EA#c{8?v=XR({m`(U2|Kva&#Lfg$r`WaTHl
zpA1=)P^Q^AdJB;uODL6TkV<b6GGtLJnPz<HEk=ec8BC^`XnNtukR_?fG(%1=0vWR8
zIhkfP=`BHqEa6F3mLa2M$b#8(8XZGM%OS@gqZr65kkJarvB)SEavU;>gS-+Mt%Mwp
zjN&0DAfp7xiO47s@+xGs3UVQGFGRj+O{sh`O^S0rj(l^P^5v+bg*;9pKQ>_Mnf~t+
z&Lzmb1o^V@BGcqP{}L3)xfBJKqQQLrUqOLaAYVm+S0P_Rf!84afdc=4d>sW|hkOGC
z-hiBl0~7IJZckR>z*Ue}<G|IB*Wkc4kblO3KSNH!fk}|p;=r|#OR#SV?!)a<2@Wj5
zgCU2M;E|l~V23-n`Td(rlmDC^!`HBn1<%JqPv(2j)m%_L6-+rl6I9P2zYv062*E6v
zOS7Q_)hof2a~v^_gZ6GEF<uF|gs4i0DYsYWiRwJ$3q*AR@@1mB4EZWiU4>jOs>(%B
z-+R&Qz1Wu5|3NhS069xC&XT~!F<UauhP+(YdbzGWi}fzA>{~iz^6Yp2=cDT}HT#ZK
z_0g37ufP9T@w(;3CDUU@P>Y*F3{!{`OKOZ=ZtRk}aprFND(B-hhQ`-FT4RXv@ft&v
zkJp&EYu|er=<Iw_zlNCEbYW{q{Dp_l79E(E^B+5cYbGCk(6hCzKzDJ8x+aNyS4yKy
zx+bZdljF1jp?yvarWY<0tqa9aHi}Z#cm8w!(QiHf?Pg;Ao7^Q=&6l^OUU4<@sfK>Y
z8)Wqc?z$qdRRjjRvK(xc19x4!vDI$KN3hiq&TKuPyQ0QSYVP1~K&lPg!IzaCsM8MA
zooyIdiZxBHsn4%w>nMHeE68O+8}2o}77Skt8a}TpM5hX#6~*i~U632|uvtYLrVk)N
z!I6I=q7)$K0v<ZgSg<^<SmoL8Kkhze+Ucbl6S>)?wuN5kslR7h7n-73i>zvqGdKF-
z*eV<Yj=@$jkauCLUD%l$?ZepWFyzzN>NMm?VjfAl!046vQpm}YNwVa~hD&|4#w49H
zTg+)^HRhS~&PqqBbQtEyN4EJq%(4BVbS<)^MHn5B-xz(<mGxcL+yml}85sedvQmSr
zYEXOr;5Epx2Dxy4j;x+T`E%s>9LnpERUMSqA;&r>e~YZ%Lit<N{Vnq1^)JCzOE9?p
z670AHgX>3Pt4Ju1#N8t?C?AWhVxc@1JI2EDao8#j%HweNI9NUxTjk>Ryd!h5V=i{#
zoQJLQpga#d=0W)$Y_$i<_h838P`(#i?S=Baxcgr0#p^$ct&T$ZQS5jWmOqBAjzRe`
z-2E6Ve+FBff$}rh@eC|qjID~Hycl;chUKG(c@%#)OygUmLlo)C2Wm7ikB0JS;t)-o
zxqO4@u|Wi3!F<spAM!@YYNOPi5A2PS<3=82l9i2;_ePl1y;1sPqcnu;Z<b6pOU+40
zva(sS-OMLOJ^O8zKoa<7X~1TwISHE@FG;gQ>d5OWkTeC54@#PYkPk_kLy(Ik*CGkz
zl3tKpFF<}MsUJ!fy#6PW`U&J3NnOL4?FQ&XuKy^6Pvaz{N`kSdZK&Ng9*bhtQw!Gk
zu?VwZ%<H9s#&1KK1vgTTrMn%TMwhHSNu4KkXI0SHk)+<wIk#Z;Z?!2MQmBR4hfMY%
zYZhe6A9K|uBgW8}3brDnt;m!u0Bom0pP}(%6aExwNR<YSQ693*<Hl&$qaSveFRRJ?
zC*$XDmGQH8@}1(o`bBru_4&pIdyqNvr|R>K4GJM2M+V0spGF3!A(tS763C^<pcHZ<
zHb}&v{t|3Z0{IR$xC1#}Fo+j`{$s)5G2|D5!3)T51%tPcV~9Zv0p*j4K{Djcglr}%
z_Pp$4DmvNfzocSQ?SOAPqq3-1E9g{<(3K-}&G=MXs;geA+x$-E;BS{lX8SxZ|7ZOF
zt;Ro##L)=%GC^J@C@hqx-}I>*8wj$pQc|zvi_x!cArr-Z8IPzhe;aw+hA-<=uty4f
z8GR9ZT!b%McMC4Ng&y3OJ|VcAfP6`Cxx~3JDBAYP*{_ebsJ~FqZKUPy)jbq)4<@ds
z;Gh(&<?cs14oQbG`-?c}BG&TF{}mi^1&4CpEd=Zq2EZE&I05;R5O9fe;fz1Ff9-#D
z6K%RjNc#wl;5W8f2wu%U1(KCaAvlxs(T)q8J$D9e{*Qfq7IvU>wDDU-);Y&tP5v}@
z_ThB8oQ>)<st^0~3CByHKRUiEhECqKlF3>Lq?4yfCTX16mqauGS^tG2D^H__k~W5i
zOHW-mf96c(i^KnRqj#zzdExqp9_L%uXH;2lLpE%AB}yi;&V;-jS#O7&g)~{H9oNrB
znrz59NRtCO7in@K=OIlV<b0&bhr9#X??4@S{RK!<z~$^CD%ueB-*WPXSRrXw@P;U=
zJ$R|M*JCHz-XF(Rnd?Z)H-7BNYdApU1N^<t@9pPBE$%tgkk(LuybF*YyCW8grF~rg
zEhu{dbI!NJ{19!MHUF5`zm<6{3H{QdcFuYY?Pe$9)V$c9Fa0*Ov(3q$sObo68`hca
zTxA-*w@oA6Sv;Tal!<LR>CQgRWxlAW-|z)6A6t$8LIM-lyE49eP+#Wk%Jg&K-tj^Z
zH*A}U{bruI#{N!+T;u1%tl}%>QNy=fbU>o496_c>P-hm`pqXpP^d#!cJap<D2*wdY
zC$@OV%2L62DdZB-xJ2y4oCg-4r(SmBXUP24Gmt6+df5fYwt#!tSC?F|kiNAn{a=sc
zU44-fG9_u*(tm5}Zgq+oT0G2Q>i;kWtkY<#HhxacTJ_Goh5qqIdDm&HK1SNdXe3`Z
zwqflyJd!V9YlPr6!f4h%va&%4-T*mE2+rbs*XvxL*I5?m9~<$vvJvdNdb(!Re^k$>
z!p}(cGY?g=Zvtp&vi{?GHkzo75IxdG5G2SGJ@O!zOX_k-!=~Q-v)c}Bo<B%S{kBS^
zt%P372Q=&h8p}uHFIf8v{NVWj9(Dj`Rj(3)S8=a~8ZRL@jWe@8bo@7d=Fjf<kzQ6D
z?vb~W8oztU=N<~+S4_t~={SIo|0~$%3Jzcn8VjY<hHCuB0a%UocCtr08Oo;IqfctJ
zNmJLpqE|RX0uI5?;0j4VA><>({|E_Sjq~K~nA$<>B0U>?&DOV(4ZpQ}$od|#=bVCD
zr(hd?tLfM}9ouuhh+AL8Htg1E)-kreg6%o)7EE`;fbt2!^aSKfg6SpBPdzoov&W~?
z?_HH`Lhv@8Xi}51udrKQJG+18{NHNM*HqqJR(|5{FaO_J=Fim?kHe$2ZvW@sz`ZzX
z&|#Kk)yn^|PO^W}gU;&uf6s%Pztu>!nw!5@Kb~A*_V$S3|M>@f71@8UUD1x0|MiTy
z*Kc#P%qwTqhhp+2O}+$j0uM-<1DxN!-BQ@?#+nzjsV_^em!;;fbU#dgx>Fl4V;`*{
zU6j*BkU*O!%6XjWzoRmJf-MaVv1B_P)-CL;^QYOf<vHKXcoH`I^RUADAd&}*B<bk*
z^PlegXAl0<hyV2AKl?WS#LAA)@x#rNvoj3)^mX^}?AgoT%e#k9Z$Dq1KFvxx-u!2C
XDvgegFaN6_|JjTG?8$$6@}K_?lM5{?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_streamk.hpp.3A8F1B585329CAFC.idx b/.cache/clangd/index/gridwise_gemm_xdlops_streamk.hpp.3A8F1B585329CAFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5c9ba8bb9a0959aed780fa361c6e4d168c7efc84
GIT binary patch
literal 23270
zcmeI4d0dUz|G+!vInQ$Qw8*VnscuxL6r~c`OFT$uFk=|Y;5%c+K8*Q}B@!trNfC+c
zAti}2)+{OeHq%t5EEBSnEQS0|Zs(rkHGltRzJH9@%RTqq=kuKB+_Qep=X{=Lm=YLR
zHp;-jf1)yP>ZBjWBLf42=J@~YA7*_U$>Y?@z+m3SizA0lukbZ}6<j`EJ2$wpk2E#s
zYMR-$pSuTa8NU0~Q#a2QK&#kvF?P#)w5X@wT(h>V+CN!)bN3Yg{~WXO>vZ>7r#}jp
zYj>=2(00$;+sQB@XZ_u0mD_)>d%ksg)|i#8En7dV^o#NB6=IpU{&$zsMF<@z?_Vs=
z$@^0|_r=BX{x5#|)~@W7t@6d>V^>DpE=n1gwRL@mY{#*jIyItuIVwF9(Px<wzqn{q
zLYD2MDFY8~t?zqt$Lom6e4O!~8AWN)z023trftd^kbZps^zC&=<92r(8~;<$^p5=}
z<rH0>6ep@3+BN(7s|3Ys_E^_59e*!y2p@VNpmt_Z^V($lAj{O8cB3-qR1E&_(;La_
zJa2S!`Jw-jIpI%t#q3)7IOoj8-RG7grxk+xo?lNa^tW@UwO(%4<(qz&x?A?T@LgMf
z$A-Q9qT~$gWApCL{<(MmD^?dYH!SA$QcW;!b>m#~iQgQ)*P$Wyg6iSe8y2}Q*X+N%
z^3?7hzFX6+YZ{j|p=V;haG$Q@UKUrp`r*X2Z^~zgdl&l68Y{Nn!`~?b-`V?h{O@pa
zZ&q=ipTx=6#%${FpJRg)2M(TA6Y>1z+TO5(^@-mWj{8YSp5ioRzxUf_(^fCs<+gA8
z%`3$=gF5Z}az?aYNX(L<>AlWA?3>#%a>2>5o!iM{(|Z_x^}2t2#-*>8PIdKs_Rb>j
z%9xpn6W)FuW-;l?xDW2?lIGv+4@mj#%OwlO&u{(1n*1yN1MCeagj^WXF)4285|gzS
z_dZ;<y1Z&xm%5P`w{Cp0XL8$}uM_r-+@6tjK23Yd_w3b2_glQFnEL&OE$zz37<(=_
z@MO3}Wvso?R_Axiuk7J6#G3f8tq+tuIp}=gLBW~B<14z3&gqidclZ8d{ZChTwg374
zh|AS2K76&}$z1D-QAv}7LY#Nq?7#46+D!AXxL~({O;g+)J0I=cKBn0<CD%K3z~%d|
zinmuA>^R$Z>iJ(LmOME+SbAL0Om#!N_f55*`>yS&%hCRl$$2+yPaGUkEBzE=-hI@s
zj|!S8uD1Io+*;~2aZs3z+qnAi4Z+vWI8X2J@T&#$lgG{R+kZXnY42@qLP`hwZ}F<R
z@{Rx9bv0XuZOtA3S!dsUZ@X;ko^m2Keb~uun+>_>{<Bt@#azE8`*wVOEi^bj!*Hmx
z+oDqw4QF~iEGTI0=)Pc$-Hco3f1hzSzsF?1&1urCsjIa%FNfSk(@a|&uc?jdKEx)f
z&MIN$KBN2@6AWv2McaK_IbqacZ-2iz2YvoCrDgrLZ~eXtblKU>{H8p7WcM{a9-5pF
z?!INhso5sZr#t?^r@6J+_`-YJ`IOYyE2Zs?G!bQP{rnQP&%ftj|De?<us|v~vqLj&
zP<6x`hm|G&>1jW+qRq+{{SSk?c9t8)m55(ATPcp4FP!~z+<ed`H1WGQjY&yO$vDFu
z-CGPQ7mQw54P86LI-xEw=uZ3B1!fP59-pqd=-IE!4?X=}4WD0T*dy)T^6Aeu-3|qV
z7xc2Ve`&LOCueutclA5x7sD^{&#T-_bM{qT{%%0sj@Wmm$DBX+x_@!;xzcXIvBwT)
z)|3pq)aB54%kRuuf1TV*+h*Q(`S(;qHGv&}b@KJ-W_H$bakGVceIf$eb)0dJ-yI`Q
z|9Y3xd~d^dmR94fJ65^G6&(z}a%Ifc(6O3DW6xE8J7VAZ^ZQ3Xa_rEr=J)K($PfoV
z$C~|?Xk_g$?}pmK162oS1v!0N<uclAoXPHn8t+5gl!l%@vjzth_q;zM_MH&lvvjYP
zzZdTsHE`A5rG*iNeWRAoD(hAk=i0aDtVa`Uwi@(1U%KhwU2){0h1C~^M|QtoF)RP3
z!|x}`x-8o1wr&2hH!p{T_A#4j7uJyX=Z-qQaLkUOuUgIA^5^#Pb3+#Vc_u$-?6uQy
zSme#Z{_l>BSn}>m>HP7pLf<{+Oz+s8I36|jt6Q^wnmQp@G%zrnd1<QO<flEGLDQ}V
z$}(WM445`01`?jR@js(m&91d;5#neD&AJ;X;i{J5D!ZnHsF?jYY|QbSGnyI0PJRYT
zEHI4)nx=%PwC*}>%g^qC69w3|D^lJ|&E88UO$kwHn>~Eh@aTZ2W>CguISCw3f(}gy
zQE^*6DQcj{lI@MTa7vh}S(wVCDIqEzL4Ddki-}elLvLJ`IA9uw%c4t&ins5*0{QvG
zbYtj+RaOJj)mTNB5EUP|vOL0IcTVGt43yVWv)8z|x`e27UNC5y@s}?1WrM+843w5q
z%hn%13{(O%80HHU;BTIKS6gy?j5&0|dHihsv(3j3a(%G<lv=0J8~g<*VsQj$76Do`
zC3+RFRRJkWU-dVKc35S$^=zAuAA0e~=I<|r29Gy`?pWLaPy=Y$ln|AWRuA2lxP_cD
zgUzr?8FVj$N>f5qh97%*UkvzZml<@$D&+u`<K<VE5S5XYP~$o_zQ7DxW0eS%CPL-d
zln|9K%-SrxWz)0R7zX1yjRcxVoQp0YDql^xG4pbPYX$B|-3*jEsb!sH*OU;Iuf1+m
zpNm?x53BSrP(Da4KVTJILR6;pU%RW;IMK}*I^r9}0Mi)IyeZMEEZbbhWuGZAgNC@c
zxxg_OcsC_PWnRgbr6w)EpJxtRW0fDRf3*4dp%>rT<za1dJ-r!hj>SuW;S$_KbO}*e
zG=6kD?p_KsgO*rj2j{+n8_<*xm8E8VdR=P!+lyw<7^{Q>!*E<wT|!hM*QytM`)7x6
zb7+HAW?Ij*`S?M^@!sXqZ(g-~%%B>Jlc0MN{IV(0tDOGrgL~A)|2BgvtdheS=5U&(
zgs3Eh#lCg&%=yU}4#8Cx1vF8(ZFC7y*^n}Et@HlX-o|h!R#^#5R|3nXgs3E+4*lk}
zxw_UIcEmSYV7<Wh;|CFMIyb9A+M9989BQ%n2kRefK7J5!O45H998-U`(+s*{@t+X=
z2|G0<L?zYG$!F0qn@7g5GcLm#pjiVPn-Ze3rx_>=64$LThxRy^Io5M*K7J7K-nBVy
zi~IiAU<Nf<oW!{&as8SSqLOjPWa6k*pG7o-5>{CX442{`s!NE<f!srruCB7GG>5iW
zWtR0Un~xtvoHgQc^4v2bG=}YPF;@Z8Rk&4k2~jy{b*Z)6F3`&in&TW6b7(Qwy(!VF
zY<zltlKaf%rqBcDk`9d0fmu^RRI-B;XZ>zbJHQxrz$($eG#cMXmk^a-I!vj%^8C_-
zX3(VZM!;|}&PA8#RU&-L`@ghZWCmTZN<KjOxbbxfQMr)z%R|RkCS_*O0;@!GD4Gju
zO7tovsr$JL7w`<w9;?(r_c|EZln|Ao7q_>Z1~0Cf!&X>juJv4-j~_&QV@@)f;5t||
zgPvG?gll<(vujF-N=aRLn+qKWTbaQYSS1d+$Kl?tOQ_19V|K1j2s49LSS1yDq{2Z>
z2~jB>;nHbH(w0<n*aoXCv|i}&@q>u}n73`#4;z+*nL}qRo^L(h_Tz_Myt{sE-l&q>
z0&IzA_37gDP9HxER6M|IA)eJA_i$Jr_2sf1&7cpyJ`@;+0;{Hk<nc7e$Sv1z%X|wc
z;)Yubq_uca)g?rwI>r5(SGTbn&7cv^<uG(V4E>rCqEa)>c-Wbcr?%$M0jtcjo@ev%
zgNSEMnml{sAOiy-&)dFR-}GNZS-B3{T?hWZG&Q)a{02;Z11^-GgLcn>Kbh}k<rOe_
z1zadcL6a!x65O<^$jUQl@(j9=kt-_?gun;Fc+!x>bPx=i42D0Ns!qlehqMwHlQ~6J
z5<&Aspdr;FD|wtYk8>g!;esIbD`ZXiq^LeA>e+Db^NZb*8Xj24$|lf!6R@HgrGVxs
zEN=$QH?w?}lg@IcG~;uebdKfooOGV$0!}Jm`2r_hV7ZW!3Ry1Vq#~9ta?(YXFLBZ(
zmZOjqg_!chAt{dKcqGNMydFvGSx!Q>Nr<VqJIM47vZTMS9GRB0T*gafJafN$ynK&W
zQ~d|L{D87i$B3jUYdwC~%F0abOeeB7$;vG4EGIH6%F1l*Y^UHR%S2Xw(Ei{==6+e3
zqn+bKCTUrjtDWmarUhB~QTwA4S$Ab+o_3xS>Eg2TbDN*r6MIQk5}+ml+LGy@k=+tt
zC(4OXlgP?9LCq#;OBPGqO`zK**okrq)TFTTYS_CP_M;uU2KKIDIfBzfaJFO?mz79P
z6UlNsr-|om$+RgeYdOtYmJ>Kl0?Ub<CXwYd<erASXfy0Z>b-3LPe+#N$d1Z0kwYeO
zrksUbvycbn3_c)(A3!;iAC$=tqr6!1UM%@7eXpK2T|3;$D0|oOcKnQAOl52op<Ez;
zH1qL?jIC{SlI7C?oo4wDfc{|lJwWeSj)N!;GWjJzl*Do>M5!zvhUhTMWe}CI`~ac{
zEZ0F)$8so#LOJGst2nfZ<s=RzvAlyrJ1CQ#4KQrl&yekfiRFzx9j}*tCm|&ZdS^kO
zk~-b8gOpTGlge2VLjx&=oVJj&CsmD<2F|~M3!t2Yf|Af+$`xox1sY0u3m>$FAH1pc
z%3kx5ehWZKh0wb~=tJaWW5pRZZKuelK~{DG<DI~qPK5_S>x01gV3S#)ITf0aRkjYd
zkFs7RuU8op6W;Q6ru1%Sj;97G3q2NklEo4!iLiSj>_NsDQc_^|6xf4|p~i{`$%3EJ
zipYj-vSC}wg<Q8nt~)6>QuZN_eaMsYVdQlfc~j0q?eb`!MoJzE%3~`k4^7BJUsL^j
z<dM(jSAg0Tu-6x$paS;#0yLq3y}laxRikb+zi{3!oM$+dcS+@4X+vc59kTh3RR0b?
z<POg?+#P=O9exazSMfco_+FHs@jahWw!VI5rop3blm5~kHd!{~>H?M~a!83*IY+Bp
zNJAkdM&%sCa;(ZZma_I?+e50kfAj>hqA}1Ik}*v7*|<;ZHrZrVZr$_Xks9AChW|GF
z-`eoRsewnZ&Z9xvLsrg$mS;g5YM$K%E$@Oh#2jn4z&9o4SjUmLeeWUrdr054+Pxk;
z_t@+=pJPEJ-!hVKP5PXhXH@oLGszwcLV-;va3!Xy+r)Q%ws65I0V&C-b292eu0Tq?
zIIvy}A}a$@HcJCHOF>kgEe*+*^eYKcN~E48Qg14MBKbX$^rN24pSbhsd^6<7x(83}
z=eg&{C_JGag<Xz9KVttP<s|HKlI3%-%Q=>dV3#76uR;H7urIy-HuS&E@-g1?7_VPj
zk#d6fI>Gyp5)<zbZVsJKh)$@xf$?s3LOldpAEFbg$I2=9zW?e}B$SmyqURxTB*_t*
z-D2nK;%LGi4>qN@9Cv0quI!Uw&`A(PdH_<Ypj{Qz60-v-PoUismN#?Oo4J;BZAs-k
zQaMk`g<Qu%PQUDX_P-Y?SbgCmHrA=dQK+GHauQmcWceJlILC4kv?yZv8nn2^@@;5w
zo8@}Zxn6W3y~OLX#evQqGhXmA?oM5kk-mG#$`;9a3+*1bswDSAQXr`sSt*rlOC|jx
zAS+L#9#5nZMDBHY|FPodXZn4d3uWxWW9LFShUHk5981~Ty}aaex#K$A^S2_mtw_KA
z5>Fp)Zk>ORHuo-Iyo+t_Y|uKJHn(?`&u`xsJRgE{v_e+qWOgNPL7bn?8Av*~OnjM5
znlkZ9HZgG);*rtlDa_d9dVEsu>O;Ox6AZ?ijGd!uG%VVua&By6yB2coN!2v2k7M%C
zSCk9Tm;#oIQO9Cx6E^nyz9)En)g$E&KkyDeh+h9ebp0T<CvA$94`PoGVo$=peG>a*
z)weSFfA7UzIxQPkc6!fCJg#m6%}wA*=ffJ%qXv9I=EJUb)fc`K##iI@D+5?$083)(
zASDx6WPZvHnZSw44*-h;to#6QIKaxYfJGK7&jJovto$IbILOKm0+)lJJ<UHGSY)&E
zY~Yg3$`1jHL#%w2N?t`fFH%;k<kc+4spPm%**;Fynhu6FDtQenU!&@>M&(E4@v4sT
zDo@I5RbAJrx=~J0StY1iQchG^C9=FuWxY;iL*#zbBmT4fVZ=;BSt;hcim7ocD<zy)
z3Cm@idl}b*^f5dJP=_$olX5uf5YF-?WP1s@QTcUbd!6MJ-YbO<AT}AkpU}BP=tuR-
zg?8nFFX3*+-Sfu=I{t)YWf^C(jPs-$xHX)~8qSk$*fLPp45T;SWaR+zJ3t3huO%&n
z%2%n*c<^UI^DJmdtlZv_g<YTCnVl>(R)u{r*NNP^u`29KST5tN%Q)tzg`t*V$erqk
zqn6<;r||YEd?)s&I+O_dWlL7d1$DX5mRwJ^%FPWw9x2dSXkn*?-o&PF+|CU?35Jjb
z2p1ZRE(TxG5%wC4eht1NBdl?s-6sJCkUaqQ+rXZQus6&8iLfu_O|Z`<7(n(2GTt-v
zOo6>w_D_L*DQ|;qw?Su8o3c{DwXNWoeeFxG*Go>n`H_`+{<C^szdXpwI^nZ*0>kT7
zChJwrNk<JB)^gkYS5kl#wttXEDAJoLva%R?EM|E-^4^a6lG>D&9MmNTDU@@OIu|)o
zK8DoCSU!Qeo<Pio`z%tQW#vUkUBvQpWbqueq4_Q0Etc>MuN2fP1?GM+f;xufG(nxl
z@;*VmkL5_wGE!_u^N$xT<5}J#TJI5EsQiLxenDjXNJXM~5z7xn^M@=y7R?{C{7h6o
z6PfZ>i|T5YUx?}#l>1%Z=(k~1-`Ra+r4V=(f*!O77J|S+Fog1D;B}dmUj~7fS@}&M
z-vky^{}zyMv3wiIw^=R*uEoHc>KB8cVla|&32-f8<)y&2l;uj`RS9~~5mE^PE5Q)T
zRluu?l~;klDpp<t<QiZ>^<Mz_1<SQSu4VZ(aD5HDss3va^csw$TnAk1Sb05gt!Ft5
zdWFFr)P)-c1H<4D%FChGa#p?^1}<miE1~O3=uP#bp=&hDG0-)J<pd}vKnpq}Btkin
z<xNoD1TE+ckOJitmaCy$4K0XCEGyN}y&C#bu7U0~EWd#A3s%1t%C#&n=j7#_1)ccg
zIXRxQpcDTd&T9|XgPh&T%3jWEFUuL6R|d<OoLA;2jMSN|XD$n=vsm86JMZF|Wnn+x
zen0O+tSVVK!n+)yyAxTd6x5Xh<LP@OI6M-VWusbfsAjoVP}j2hF9r2WmY0f_OGW*T
zP*#?TzDq@BMhzD&!&&(<(Q+Bf(V}Jar|cWeULPY`#<22O(K43h6_U*giJ92eNH%L&
z-XS^ekeFpKO>#`5-0$teA!|cV@5KvmF>)%VrhpPSv(*>t-~q1Drsy3;4SZP%N4>*Y
zE)lSqpto~_I>$}7a9bHG$jTf3^EbTSp(HD@f^V$gPv^u0;qwH6k*5p3>8yOC=(|y5
z^s~esS)zW;2nz13Ip25PP<*2X)S-cTs$`{(53Xaqv57)(qQGpLc8MK!iF!|&tgM&D
zu9qgzg6)@r_De$ulN}5)Yzh{ji@)xOi6}IBce~|+fN4#q;W9p7>T?(lAlpRZ!p9jk
zMi$U)E(;--h2i9Wva(+c+%NX0d_W95K$*-HcuS;<K%g1raGp7w65n({Ei2bJ=WCqa
zj}fBk`^(fxrqTG|s!meZN!BDjK*lHj>JO4N2?oGztWt-otckl;#-RtM5h@F^?#tM7
zrjAxwQ;t!oV_1$=sbeXVm<YTV(*;D35|e2b8aBm05P_^*<(#ilOLxThwO{Pr@_Sd@
zPdBPsY@|gQF*UREAmh}LcsEmvTGb+^yCjMpiDC$;MN(E=23<G>xz*<_&UuVWw{*hR
zb|KIdf=+a;x(qazS*`?{O3;ZET2`unri$e-s0o9eNc+gja;RC(@>*!J7P`{8I~kfJ
zLsvRi@8L9ixK4Dg%-}Q`ls|vF&f`H|nAbn<BPWq3xYcy=CUgL70mfT^1@V!O?zW*t
zDqlJ6|9ww8CaLO4Khu|~cq4KYI2{GPbO`<eoPJ?B4OyqrX--!5Ag?`0zrz`m)8+fo
zZ5$nNKZ^i<5uhLWX=JG2?xTyNA&!Z$qwS`Ik1379SFAwoSI`bPVWFF<IR0L9EU-kD
zmc*ekq3}?jY4<yDxWisT9WEhe>AH?OT&GM1Bfg_99)~!Zzpk<BI%%`%1f20E&TSLd
zgSarh>FANX(Zp~lL`o@YU5XfAMup^3A?bZ|-=Es)ydnAJm;ZB45E5#HXGmR~5vhJz
z*#V4q029))WF-x>P6G~<vp}mXd?3?w8bjJ2SDr2!iHvq~9E$sHV+<1cGxjM8A=l{~
zJg+GnNi=K<LZXV<utPTNMz>U@f>Wu$E+(RSz1WIOJF;>@RG(nER8*I;Tqdf^D3igD
z%cqNQVs4DOWF!45u_SkxYh+2zLfi|iI#~4~-7757*R}{OT!jCeCk5qES04$DGQ&r2
zALF8E#*x4{5}1=wE#p8X<0xQGVwmt@E6}V3cC@M!fF^<EgV6C{<FuwbvBM@8=X8WR
z((tOtIpwlk#r3M<^g|B&B)MKsSbon*?>Q4%gYP-#_niJ<2ru`X)WFIcIOhgd{+{=G
z&->Bq8+fk<mNy7q8w5taQSjPGdHI|Cho237y%g7=tBI=x@gqj|%5K=2^L82@x0`vd
z&HMoBIVczUlnbAe`;kL@Tf>imWwe?fLXU^A6ScH=pn*G(-t&)pG#Z^oT_Un_4D~oh
z4<Tfw2=y+a(~GPu;fF2ZKO?u5m0i5oF6wu0WZzwUS1LckdmrIDv)ttfTP;WU?nijW
zcXWjBeS~M$(<<Jtir1e!$jU__<RTq0jh3}{rNHF>Oz?ij@_WJmy}-Bt-V6QS3wm=7
z`*VeX?**nl-V4F+1^xLSwo!zj4?-|4e~~z}h_3H=`jvXGkeKqWkovBWm~OU0!YTvl
z^(&;H6%v!*3TfyHNpCMS<~MwWq(60ta_T)Krg&mB?iQcAsG<gNOP|Di@a--uS_iEQ
z=`XTk?_keRD~DDLwRULDkb{E*LyiuP3^^M+o6|g30cjO5ApvpN4g*p&Fd=bkGM+W1
z7+^xCQEYMpDHfQJLv0!R<fYZXggEJCd?X{q0TXgYgZ(|A`8r@phx&TZd_Buapm`F@
z8$k07EN=wOH?o`znkTb-5_CKXI#d5xKCsURZK=hT4_xzEk90ohk`H>Z@_>9WkmX_d
z;0wxUf$>>jPPrV|mxH#{Vk-x(<-n8WF6E#X%K_zJAj`wbsYhH^?gHbxz?{k(fK3B%
zpcY>PXxjkXS@v!K-B|9`fQ<@$z*hqp+yI7CPKC~?up^yH%iw@AIF#7W*gpqd>Y>Ml
zrr5wn?p_ZASRPUjM^X;s%)&S;x;tOWnJr~GmNSdxtmtliHD|V(<zG1MFC4SGJjQ8{
zv3#7<9%uOkr#->)Nltr`<x`yY6w8l!v&Xy@9g|Oav!^Uq^Jdj7X9#8)0;8WVSmz7O
zVtYogKEv_{!Qz8pOY;vEEkZ?x--)f?iO$4Z5Or?R_^{mvUg9zSn3ErK`pGcrQNXQl
zCeM2Jx1ap2tT}S>nZNe_KYolG9cAXZY3uYu2*;_2wke{P)G0o}7Hv1PoGRL;vb;^S
z-9|a)x1W5C3=TfQb!=r}<woq5)z>PYFV6cp{eR9RCA%}c+tx*c(*=DObl(NPq=U7R
zbFbvSBo#w;;n)P##fXy&*Ot4S=yhz{BU}TKD$7WfBiWiJOg{Bz53AN3_PH#TESIug
zoEXV6hUFSbu95Wdn#5GYMx!psoSMz4z&Mp|$Vh-FR@6nDQ^l>ocq>)hnCx0H|J0n$
z|M*^$#K7ZPYz%-Wb(mbYKm1oslr47t+ytF(f?zu3{|<b92ZJfs08I_(K&R|Cz~>DZ
zOnD{rSqTTzDR>q1Sw(qMpF7>`h7L6POBfM}z{LHcG4!4kOjfQ4Ay-(lC_@a)pw1g&
z{^1tTh2+x<Z*uZYPH&cO`F<c8Fmzc3zVRX8a)_R8lAwK@gDzs9wqY7DP6I8;>YchU
zDyGP|bKn2DA+>wtwH}5!-?`z&D+WHrU=VGr*TCmB7(^SV8V5x{H&SV`QUf(LEI$;q
z4+SQ?;klrFPMO5T;x=l0a)DO$RUzam+eVpUU?y#&J?BqU-h39}ga_!S8|=OF^7tWB
zQv<za<rdJ~0$N(|TcFJ?;7<8A(A;L_w?UiRth@yHl(5GxC7^E!2%^`Q0-sV=UJCk`
zvd1znfaV3z(u#Tk+Pnbnlxu;emX+6nHnpt04*1lu$0&86ZygAt*VhA|dRAT!`q#6^
zEYZ*>nmyKthJB-95FJG^&?knK$H2ZZto#Mkyr7Xdvho79c>&#NL)Su0Ei12u&b82k
z%C~YRTR9E!e8|c}q<)BO$lNO{kC6Hi%k@ZI&+=QOe#>$_?^Vz1gA?)QO7L1DFuYFi
zTF3Hc!D}<iwSu;mo)%+gq~Ko5o+j4{-D}y+PpvShR$w-7F9q#O_WD<X_7%&o1?_8=
z-w4__EQgB@;i4=3{mVp$Wh}>t4lyjpiVm?XCyVN2k*UA!qIx^axuQCk<>R9IILrB>
zI-lj+qWU(=C8D~7<#JJ7PMNeLp3`*C5YeHUiaMtvy)Qn!WaH`0UMIf8yYMhk9VY5s
zTC%cARIj4EKWo<cHyhh*#AEm<P#*<$#3Mu=N0JPi9zvqGS_7mtz=F)LvT~Z!o@OHo
zl972ba-em*S#sViF&o}&$vs<Q`b>#rTOu)S@<g(FBDE#gXO8~WGj*kj87^T4Xr2Mg
ziSO#bZdGl;x8Z+$^fV0G%m=fchEe5$K8k`oAmCv5@o_2oLkhvDLNJ<6CYQm`%OHeu
zB^Xr+M$^Gw1%_6E5Xxb2R2Uph2lH|`bU6&6oCy0RLWTOVQ*bChREVX72Oso!#CkfG
zbDHJsu44tKS;6vNZopn{5OICUN+#Dclj}`+CmOhu^_1*GZu?LNQju66X)_VC3&=v+
ztWVh?i+zYA3$@Lnk<9p*26a5lo*(5Q|2)KaD)NzAK6}3c<X^zbm-54wvYwb{yx%j{
zQ*ucNxx_9vp<+;|$ao$;wWzU!=@W}uRz9_;W#v<gT2}Unf&1w4B`fJ-U^>f}#G#j{
zr-XP9u#;8y5FIrjwgKa9pE_A>(ts^7C~!rCR+-evO172QUd@DZW@DU+&IM@L^ynS!
zk+%i)ZR*=0Jslf$y65mn35X*D7ftur9TCXNc3`~y)1qn9fIThR0nq9IEn4>D|7L3<
zYtsIWz5ZKcuMhbgSoibQs{Yt$ja0RWq{noJ1J}+=@Ac=eIMB3Ca;cNrlk*fD1}V9G
zkle^=2f{~>F5xOSa`bq3@*hP3h8N14Hy*AU<pK@aFgC_o8J%SLG%z~N@*lwH50>8p
zqxUSwL8Cax<d+1El2}fKMyV_xhDL{3E`vs8EI)uo4_K~)Ms+NQaz>#XbH7!b(JGdc
zIHM$%cW_2KSU$p`Bb+h${kg`{m9wAU3;N5(B(Gbm=Jz6At-@zvNxV9VXJQ@p@HTtc
zAgGt3=}Y?5A3m@b)o(<-{?YR{zu4#Z;{Ly5=x-fE<OyRVJd1vOyqM0SOMuN1;7HD<
zWaTN>>nX>?32Z>0Z$RVdx_BAcUS`+58_4zs%lDD`K5GO&K<Wo9zeMVnEGO_o5_tVV
z2M#{xy|?h)sd-W%I9CYzaCz)a7S-#hkw@$gY~<;lhNeI43g>)<9`+ye8a9#doB9|B
zeHH+p0(Kkp0Qfv$w<U|A&ti6)kqbR@VOJ7xAuDH~=NXo-K+h{I@8*1Vv(ZNHCFl2&
z2N`3>2drrQ<l>A|%^C|bpn&x<Tm-`|vSDoxz<>uJn3_<Jz_3Sb7};VtU@;7)CR8{a
z77j;J&V_-wa3EXoz%wjgfq_?8-pviz%>`2vDV-aZ&W$8|s_jCL&Ibo<|DW6X=a299
zTbsL+-(jQokktQ>q<>EDjLVI$ALnPU#71wWpsf@b!~U6|eMb4*wX21Nk7`f+<DRK=
z50gTZD}AmSHHO36WdeI*)ga{nusgtV7O=}=`5@391g)rkHqd6Xd<bX{u^gvr5vMXI
zZbGE2QJJq%X(-35<ajnbeXUAf%W{HBPGC7vB`31HP9?9SOilyvIMuzGfR58*L0wGu
zALN(;n{m4L6A;A<_u|U@-zeR2W4_`|U$NnN@q%f*ppSXD@aE`^H$AJ`<L>b3p<3ar
z-9z8rojUt3t6Pz2!0optRbBroc%ge40xosqn-EA*a14|Xa*Z0_mmGeMSkfcJq_G16
z&qKh6w8G_XZRQUPwn_MRwfkFFyDKB7SiYH?ti|o&id^;9&DC3TM!MRcZ46_`;jDAG
zcEp=b7B}4IbZ@jE4SoB5#E2c&odMpA?15%`U^}uI!P_Bdmd^59&g?CxkKO(?{I;bq
z$NJVke#G)??Z#o>tczU!kI!smV~^dYx_5z)I>ygHL-(ERk;i?`{XRYNAXSbVNB5!-
zav^y@0`Kb@Uk^h5j3e&2R)riBCSS;T7E&L0qr68U4Gk%4w=$&YGv{yrGF{y99$m3G
z>psMS1b$!wuU}TkyGL+|8ecp@ZcI}A=f)R~kcY#{eJ}gJ-fIc|odf>XIpFS^&t_CF
zdRv8C?;SLG2ffL{a^L972&1Qg2n+5?4tMDR*uyu2LeoP#M&Uh9GWsl;eXwYY<iADI
zM<X{r0NJ@z(ucq{J^=ao6KOn|MjlIhV_?^eSo~z=Mc{W41QH_^JF$S@BM?X}x^U<h
z4g;yhmd^R5bAiO-!v08=bC}ADn2zL0VBWClRW+o`H7<!>uh_Vnc)=^44SUWJ)j1;L
zX1pP)Z%}^R|4s0^Z`Z%VB`)Ryi#g_}mT-Y3ESGV;%D6!CQ}M$>d|(Pch??u=Lf3LZ
zUo}ta?WR1uHo^ff+Lw^&CD!i0j!dspe)4woo1hKLI)Ai^tqOoWt+R{3?jnx2)V~ks
z0kC=i>}ef70(Osp6XnIwYB98@bs7%s!l4u8TxgyPZCI<={0z%ip!pS+cXL*|IeXeB
z>6~3U=R}w|i}0xbG;sUb^G!X=51CefoIcbEynaF=cg9!f-WWt$k-SF*TN{n<2_k=%
zl^a6H4Qg%FY&>$jLq;pde<zf`bwa6mR33T%O!>}#{HehgrVp#aRs?(f?QtWur;OUn
zw0;(gd-SJ<^{WME7h1eMY4W$1=+!diu;YaTFaO_T!C$}KzgP2U>MA^sEfpP>ip&8`
zwCE5`x$fp7ud6kg$8a5OM}gbXXfjj2eY@j$hmwR^{G3DsQa7;C;&r@N9nZulCJJ7O
z)US<WE=Bb&k$IYEy%exs(jVC1Cp#q1{cH?kp5&QF`Q7X>_qP^@x#FUiBGXd#2%|!>
ztdR5%ZZ^y~dnvSk+d;VK8+iK-y#B$$S(CmW-}wG#631>p9OwqKXMH_8t8sPe)%c<N
z*N=QLFz89Y_NQO*+s?>8fAYn^Kp$*uU=Tq6I*@)HK)(j*ze45bwMKIbeyNW&9@w*2
b@4$Y2`~wE`?H@GAps!xXAdr6TO~3vhhFjfB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_v2r3.hpp.66B1A28D309FB5D4.idx b/.cache/clangd/index/gridwise_gemm_xdlops_v2r3.hpp.66B1A28D309FB5D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..000b5cded4987bad2013c57491803a89a543bc4a
GIT binary patch
literal 18192
zcmd6Od038FAAjfE%Xz9>JzCY19we!hN=2J;OURO%83wap2JZ~UHjH`43`L2gED=(c
zWGPD`vPJe(T2v|$vZNSU8nXRP@AI6_@%!ifUDy1sYu>r8ah^}#@44@@e7AGH-#ct*
zP>^D<j?TdEhc1{g`R9qej*d=8{Quma=lozS>gdFE)6rRw^khuUPc{9N_jf(oX80s;
zx6_sR)4%I@X7ur~=T@t(4>&rLw{gv`I3B+JX6m31zuSkJKb|ry=+q4%0?z5<J8+J{
zvZ+yZfBoWrf8+g4uPwrUnE2bpKfZf^=73+rt&qt<-!5$GhK4P=VZHsupy{KpPB|C-
z<lLICWA^Q+DlN~vw|avruw<&*IMC_Z`}(L=skf?YZ4OO~`mS<z|2yy2#&&Kh{d4r<
zn62qnhyTTom=j;2*fhoc(O-$@KR-YJ*Z2+@F`U=z^Z~O2?F(K#d9Z48dEW4izb0O@
zSu|+=<%SJ|-*mq{e6!cPNP{cI{Ze-Oc2~N^{qiy(UgbOS`9Ygb8#i?pP1gD}{nX&R
zI2~NS6Z~NR+R^*x=4Fh3w=pqsR_+gLZ%1|N<uglJKf!itR>b=i9UiY%UA;3T$0Ely
zK2{O4`NXV~rHgZQ|D~G}l(9Q&@U^U+UWTK-tkXAeDff*_Tp!aj*+274?o!{-brbKJ
z9+)##Oi8vM;H_5azPYjD!@TT`vmGn--sgX@_S??y*S<THvUO4Q%8VV~c6P7Z+%RVH
z$olQZ@As}*VEX9W7p?=64(De~x0~Gm`n#v)cVm)v@96Qf$K}dt&-_a&d|uv-JiI+E
zV(j&hv%WKIiih_f-_yOU)Xb>x=Y&apCs@?)Y^f?ZkFK7{s&bm26dTt;zp5f`!SLV_
zlV3UMzB@RyZg4F0-*5gp{omm!*Y)2W4y_Xsdg-6+mg4km#hm0bjz5}*tFz5shYk)`
z|AM))_&??Z*6o_Af6}{ONZl7XOU|SRj}ISHz6$y`IX~Nh$qNJP4jk8em%8>$$O-+0
z*(u{EZCpKN(B8Gt3#~4t^ekR|fB%Ah$q#eC$&E{`eRgwk@R6rkRY}v&4dHkCb*rf~
z+_mAyFG3E_@z^-H&#k&MJr$8R&POlql@R%S+hm>2``31x)4ge^(fmWh8u}`94SHru
z4sVoaoi;n!J6~EH=M`_-$$Gr*B!$vzV}{ZG^JA7y`M$<*EBDu&aX%g#7OB&a`RMKx
z=Z4j0Jy%QMX+U4^tHB|~Yx6IS$?OzVmHH}m?&DmS5i`ooE>3Ga;_}V-J->zREO?MM
zuYBExitTpk$%Aquy4t5k4Ex4T=v%TNsA~Gd?hf;J3=U0y;pAz4#U^o$iAUFI2|0H2
z9GBal{!+To>xZw!8ZVFHUo&3b=iSXu*(3&ek2g8KG`Ygo`GR@O<^_93t~;{umrMHN
zZ1!75&6+6N?X^$;Eh5FmW4>xdcGKg)=ciof-;ibuf1dmA(o^@ptC`b0W_ZNzduAU_
ze=uyYT68%2%IBH?GQApky*b0YVDs+N%OAJ><6vA~JvcxP*!l4EAGbdwG;f>z^t0{z
zmYW~GIVq-U^wXUejEy!w`0S3=Z+<UR%M6z#|6G41eq{0E%=n1$e#T+XN>Cn(Pu7dR
zH~KKTdeQrNc+_;8!(*!lKMnONbTUnQzOB5=$J{^Sh8u6rnYhJr{*4=bnvV{TIlXx4
z@tilmq;~Fc?7}+7FAh(g^rqRi&_OB)Hq!UWb6z!}(za>JGKJ@_Z>IcJ-*N4=kQnn7
z;fdiTLeEZ}wiSId<=YLvcQ)_8Ir|}cwpDq=$>8o(z4hF|b!WrCFw<VkCV0WCu3vr=
zZaN_&v%}*qlZOxfEy*^}Y;=-Y^NsD^*&{C<E=?@Dn3yhAx?MNjojyvPcO)z>a@v^>
zwkM8F{XX;LF}u<4URjS_1`<DncNjL&_}bdrgPx~{FWX`|Yuk#U=wg%3%J<XF7L9W&
zH{Fspb4+gc<ez>jadnHcn?GpE`#%;v^g5QSAHN(n7XNkX&%}WGZf|W5Mx4Imb=tg3
zw&Lb@cPE{gnG@A{)3KRNZ-<Vb(&KucTTgtpjc`iby&&U-_3S!>nv8jGkG*(*HDjFH
zuOA<Tnd>??{`_{lf0KXEua|R6n)7D-eEMBL@7!Lo2VmLw-Y+Lc=ufX$X!r1HML%iz
z!Yi@QrsS%ZbvD`be%SsMZ$9h(KGOA-Mg1Db!AD}V4ELZ8`J=-Pp7%ao?^$A)JmHU0
z`RroX3Gbc{n>Xk80h!J@ZrA>-xBKRl=aoezX7c>RjIW~uRy}lk5-`QB@1@Z)`XNp%
z4U$qvnD<-I^un<7kR80v++Sx*I)`+03~KHqM@^WdYYMx0=*Vk;<r-k6ZFCgahv5G@
z4(rpL;x=xSC3Nbi`HP$_s<Opygu!^|_npJ7%k&MPO<x_k0w^m$M{Pq`x*C_vFb|yP
zYy_S0YjQbNE@!Q6+F3pi*H1tE=!606jh|8llvTi7+YlD}*ArG$&)#$22rB#O$d{1K
zCFG)Q2#e#j+tRdI$`&JNhhG%M@ljlFZ9`ak^_Vna)<l&tehOx3fHn=#McWV-e?uST
z)umoP8o{3UMQMmnLn>`USOUlHF3fE{_0bes<EN|wmaFi>X&S;Z=wgA3{WPCwCG3S~
zbQzdk1}besSb{z5doB6>@*)G+6|>X=Wi2q$HiTvPh=Bu>j$0)dL4Ew9^}u31aL_h{
zCG<dAQLmtbUm3v;m?aZfWP<M6hOmtBd&Adc`ugfaM?9lgQ7=|hY8%2bu0Ap7`%VjO
zB%ML-I`SaJAoYL#(a}Ym&d@$U2IF_=7H*P9szhjx8SV?{zTl>9baX|CWdk!!2-df;
zwe}4%f`)i38Nea~*lQc|qVEs3JQy)EAIpmio^&-(R^y%4G=$~HzT)qyoWVH;&=s@P
z0A&qk(KLi*rgiIu_cP{JYSv{tu-J~*Mbi+L*|#eed_FNeM6=TuA%78elr;@uSv0dl
zmsw)AwE-OBsUx=lWeez}Z3xR!lOJ@2b>kP9L0!yJ4?5NZ3vJWRl0NytcfyU=M$i_o
z?kSN!C91UzVOh3ygBfoVz0eeP#_JLbEMu{}Xd1#2Q)T6;6L)ly2z4>bU7&v#7;78C
zva;#wzQU*zB?iy|ugfi<yoJ|A(-4-`?@v1DIVD*b!5(-l%_857$D(Nn%Nmu=n_~e-
zM`(7stD>t~Zu>_^SHWv`daZEBebVE_*Nvbt9!U~(PJ&~#4SB|f)jw_d>zQMfDO6#W
z7+@KL=c8!|OHx4XhjHEK$@<U~v%CR%Z-Aw?AuPL0J1!jA-D!#uG{7vofW<B>Et-a~
zWYng9@$%%j;YQE|vmAo1hv0B+Ls<3=n>S?Vi-ikJVJFNI4J@PaSTs#L%Qx=d9G~u=
zZv-XG@*di}hkn|oo#pg5dXMkE3)6>gm}MQ%TZi{q(-4*eOKj~r+!J|A*o~ON<D$`V
z(S$G@v>0;7BDLEBBWQ%55e1#2@WN>t!g46UqIX`$vymcn!Yqe@{$Z^0nx>s4_UnYn
zuicfV&;qk82bRn6d^8PVIk7kW;X~)KlPsYr8Oae*bwqR|45x3po$j?)y}%M$5Qbx-
z(J?yI+_(3So*1;%O@wAxo1*}V0?yh-M^^wEZ7%#dG|f6^{;viw2+RIUVEz)QwGDYu
zdBoXeYu+E^^r0ReY!0WF!*$X&?JP?+ZGN?BB+#g860k_Zs;X%SOI3(%&y0sVe=&tt
z_$e!a<w`tdO+#32hSvsV|K8!D8Fa$3pA9-@V^!5OgypvDt(_Aa2V63RYW$QqU>S#>
zqG{S${=IP5(fUod4PbZtl$$_#6KjyBAuJD*N?mSk5C`bPj`%4{pxzQ_p=}6DL%8kF
zYc34@XaEP}b$JEMU*UDpG=$};6g~D_S>0>{*ag4n22kF>UQyE!mRFfI(^>;EKQn;#
zn57aZD=~|vAuMnC6J_u1-z_$UmUt{HfaMB2BTYkCbmmN+GPe)jWbuSY7vDkqPw7eW
zHL=$<aqv>DS(4;((YRc6rnpu#u4VB<(fA>Ym!Mut&|qRpBzZY9UXGk8UW<&^vN#DD
zC$Tsi8D}GB;zlJo7a8YLtT!{L;_9aSXei0IQ0Of*jyQHneuP3FQQW~Ldee^yt~H${
ztj_9aU`sR|j|ixj1KUGdO_A`Ls8;}6vRo22Md}z}ORSWHWl6mf*b*_5u<2FD0$XC0
zCG5J@tAH)>>JpY3bsVrI0wv*?NWBKwl5h#rp*jxQ5`AyO9&ylz;&`ZzXXy!0o$x94
zNMQMQLiJ8)OAN0h?}Q#Zp%2ArP@Tro)1f;3Q|yt>^2c%NIL?-YLwJ0gM;zxvaXhDv
zXXy!?I^k37k-+jNbLwQ4p3He9v-A{Box;*nId$r%*dvwlcQ~>;x2G`cR7VMiGW>cV
zX#z_f9SPe7K8eNK0l%Hay8yq7#TkIlU~wkkGg+Js6xl#YpMM!BF0;5EDC$`p1^Fn*
zjtBBdEItJJLoB`k`3o#=fP4ds-$VXAiz7Kcl4G8K3GtUuJ0=3A*3(Ae#7tuifOgT)
zo=R98432}tDUOF;@vKC}L%(?FPw5FTIDzF)hIYx&o=VtG7`zh>r#KCIrLhu~2K~~Y
zKc%O`;B=Nh2ioO8d*ZGoIUfe*!ypoaNpc13R{^~#u7n*bp&`YMFt8B@k?35KKf+-j
z;fQ@&$BpOD*+p~qRD$EU;5cqL#qpe1Je4p>j_3U1Ie$t|;DQra{$wsVnZ+rbR|?xd
zh4V{c`=@fjsVsjM7m&rt^!ZJkT@z<dPCAm@%8hE}!YJOzkJ`vH_&hJ2=M4zmu+w$f
zrL1-u&&y|_<5}2?EF>?N@a`qN2Vvyp2*Ex=aG*F+u#cpe9M^!Z_LxS(c5+PPB>5yT
zI0+1iLzU#KpzBr8jht8Xw1+nmshBxVEV?xEL@SWwVia18#u9UF-h0if5$^A>I>^?%
zwK>%>nU~v9r^-ylyIn`!6ucY{`^Lk5L?U@P4faig{fI=hO?>EC7)&Sr5f1$bgNX#-
z{#@U9t{+)(UQXfqrm*+{*Z%?MPZy?v>)$}Jg`w}2fV49g@HDEqz-sO@!f1Uqb?=Wg
z=1zEF@?r0MRz@3P??&iNW%3H@eFb@w`Kuf3vUTUz`Q!Op>R9U0`S;i9-;T&zfS&fi
zOEkKx_P~o@n+ptbfhpaOBG9b}*wg*coQBDMbn`!)^SocN-ao#gWF7F-G>35_YGjL`
zp7!ib<}AsVIq5Rjmn;fdS^NaeL7Xr;xy2j_H&S%Lo3##Dtpkqxw9&cK_wRf<!{77t
zdAS@Kmcx$3fV5$ca@d#R3aG4rMlANJfPNHLLS-dOuZF5>*qPF+p=&kl!{VXUa2Unq
z$gUjOlSSa=I%HSJ;znfOh#U#sIpq7nZt1_oU<H4Hd|#mcguu(K$hVcn5enZ31w&u1
z@Lf)^>x1O|stFfnLSC+gHnp%j9Z?lOs)`RKi|aO|euZc;)>Ds{tD#{vR8q!jXj=_?
zu-LbnU7B*FDo0jygmp+&$Kpn0)rdM%`U_<70$EbriY!`L9HFp?P%!l63XA0wdo+1g
zO~3MN6xJgv-pZ8fQ2-y%j{3<P<gC`NhW~ZiUVWm%YR{#=5+%7rRF#OHH0*sWsvfgA
z63LP1YZ}(3AXN(Tr1&gSokgBBj6I8d&Z5x-ds{c>PZK9L8}o8LFv|y>h{E#NpqgF%
z6x&<{wv=7~%nDd~0kA1x>4m_okfj#_n?jae1k8$9dJ%9a0=?+-uK}}bEd3gAxW>|p
zfmtz2F9r_9EWHGnm9X>@;84QSOMzJ_OD_cur7XP+n3b{gGGJQ<?8x|hW=2ol*$_Qj
zzy_Cd&*uV(#7c4j=U%|#BF?#p>rd&0eD6ZO-Na*u#P=?yxWCz?(k|;3y}~ASxMFyh
z_8ToZL=nQE(5|6gghn<KTOiFnAF*KrR$TRddVB7c{|vHAKG%m1qJZmCz~Um#vWRn{
z^g>=$$U9S9%&Uqi4h&PJE^JW*;$vDh*Qc5rNEu^!Wh~!?;&@&e&*BVTnZe=%yz&5x
zuk)_g`2n<lIqzD|;yA%1PUueQiGoQYi+2kqyD27zA#BAp7n{U5OY$2&<P9}6WMA<b
zYHm4+ry5+A<g<C)pt(V9=OdFiXia=Sj?SP-Jc|>cNdk*^LX(})n#2YYjulMOSey<`
z(penGnZ$9{#PPI6e<txPPT))uSe(q6B(pe$Gf81_Drb_);#$tImb0Ksw_Gq;E?BeJ
zeYwC0>I%VR1xt?+Ok!BPMo_L1ENTBWg3B7go8q;CaxF_=Cn(pkI7_h35_(epJkdQ*
z>`zTyo*0xTj-vRY=zfu<UlfBbvh)&BSt43e{!&p{%HlFnS;pc@QCayZcB>S9Y5yuw
zS;f+;MP)UM>qKSUr`WBI?SEHP-eu|cMCCmeH;BrHPqAAA+rLp%HnQ|4QQ5@e7SX*$
z>`zU7ix|`*j-t3#bZ=$ptzu9sOOHg#NMuR*qmVL+#mkU#8H;0(GUik47K40g|5&7q
zW$CMsautgckTT&@?3Tdx-++`GSb8E-CbBpgDU(0NZpm!_tw_0*rEf#ZZ7fbl?&+vM
zv6hmYj)KzBD2fjt_X8~b017(5(vPCvN0Arh&q2L&SbQAyKF;E^NO=}nQvN)o%wzF6
zq&!FQ(4g5}$0owIlSm;yppa)gN-;m6m|~J}!5-o~FgTB||FpM?l3WJ5m4WWWmg#BJ
zFvOIRGz_+Jnj8!<h9ml}nQ7*<GM<;@*WBl?xo@dXuIAlWQ|Bbf8~D#R@C^MR?|zWd
zM}+h>zdR^06wl!eH|z~Jl0GPrAC}0sW0LN{c3G3_Awy}KQ^)hbip**BjLg2n4YrTL
zLG)%pxmj={PbAqOK~I|lB2G_|kBG)cL<@?uMdNH19~1SCiAqX8F6tep_)FtUt`qim
zvc<k_6*ON3`%r_Q3e8htA8OE7aptSIKGa}u;mo(NI8QXs6Z=qua#1wD$l?~!yhZFo
z6{uA-Z)I^hGEYZ+r~n;6<_9SL@@=B)gA+^K|Mqd4#1U97+EQi*lO1y>A77nK76|`Y
zlPDvEuY5NfZx<f@^k2i+uQxS48LWJ@T_MSPfaxC4ljKSyc|S1S4|<YJn}nsnKNExz
zv5@2!u*VDNMb1`jHz{@>p*J~4;~-14EfD+BrCp3{7o&c3DdQDA;uT&bn*aK!*YvM?
zbhp7$bQcW13qB)xDlz~phnhSemBVwu>>TJp<Z!~>=&9oEF11*)BY{IC@S=?Mf<wLF
zMOJIlA5OZ}>*_mVMisB>L=OV~ynj0)=|il-XMn*O_S}38bi2l$n>9&AqQX-~cRp0B
ztTDtBsX?JN^d4iX;fA^KFKv5b6<!C;*Rhvqd!hMW_7ZFzXTFZTB-+85@1U5ZE%DIK
z0)w;c(27B~Vs>bnWF{HfG<bh_&&te4cxa0uzZhDPsn0mL{>s$4U<*vh;Z!-CH%YgU
z^d}ZDP12LBDLK?(f+p2T2=j*p7C3(CIm!q}-J(yvID!hmO)=o67)J32(dUCWf(pcH
z6tEhFQJjT*vd{>sTZfU)VHOu7zhV?Z1PGtQkY5RlOOanGi_4H-8H;O>Uk!_IBEOp~
zzJ>g5vA7ob)w1{j@_WGIhsf_CiytAsM=WkYehn0poG_NcXTabYFe4dYJ#A)~C_u#V
zFvk@)kN$xft$JI@WX~iypm#t!BAxMmXmZD-H(72>&}58>7LcR_)+<ednCu|-Fnq{s
z8h6J8b4#DZeeBn#?H1i}J#ZsVsttRt2YwW92ael;8xa}_%Z%rC;79Q;;IIq0QWKK_
z95R3_(eO4CW19(jQ4^C5EV4ltqTrHz6j&T(@iAa=jKw*?B8SDtfyHqap8ysoSX==V
z6~K^OSxPu7rl<slBw;4uM4I9TFeEuP2^Wees(>N6Qo|=Jpr{6h#8hB~2Z|bCNc2;}
z89&8MV0c@b48(aTptuDLNqSwv&R<ar3`w*g$q#|$L(rA%CC*a;%SSA30G16behe%h
zv$zphHnO-0ST?cv5VSu8UC7m^gj3V~-@`%FWPgMPAE7D5_c??6oGJ0=cs!iJBNkWj
z230&m4;PHW1?Kp&R4`h~;!|RmQ=%OaNeN4&<0;XN#a^ex02T+H5{I#P)G2Wc#m!=u
zX3>sXjAqfXS#)EuXS3+X;=pEc7>h?Wi(@EGLsn_XhFY97WS2%0iuiblJkpRaOAkmx
zgIPQ>4Shjy%%w#Wm+U$83R~4ZoV17D@)O;}s;8;HpsM$&^2R(4toYZIIq&}G@79P&
zW0!Oe7@T8WQYq+G%DN;?9SLzsE1&0_{d1uD{@=c@jJ=l>;<RY&<o|eopGe@xtf}Br
zFZdC=w<^ZSEqTYxZ2ZQpLf}>*h=_L^9<o&kX7Q-4LVGTyjsE3UVH}~yRsGtpgU;b6
zZFbqr%89YdW&!Pp$QCxgn(7!b3{o9~32jv|Byb~<E2pQegrU<o3Z0L_u~d1VbI#AX
zv1BwP2aP*w>SJif7CxkfnsZ6sC;0CZ+V5{R#f1co8F*o2oA2kXbNDgjiKMUwhwqxI
z8v4XKDAhqL5(7x`ODMgBRx}h0=M2L+HMJu(yrPEh)c)Wtj&~R4&FlNOqv9=Z4n*EC
zkLiqq*AmgEMEs0eh8v>l1{=rU5mk3s9Entsh>7c0BGpP3uScr&EIx|7kD@T5*^+z=
zc^{*A+w@_4aA?$Bz~l2>uQK3F&Uf1mFOweqP_`TEMLsmkht71(8lhPubf#-|1({tz
z&Lo;4J`$S<O`#H9{5)WmN2A_7V@mt$9i8Tc@1ZM2Ri)@k*Pu>R)v>riR5h?T2B~6@
zE0Iu1PC%*z7AGTBGR5g-VpUYJ?p!P<t)OEoa3=DRvEYO-YR2{;*S2J&ODorp?6V|C
z@GcQNgE#Un8(F-KciG0`6@q$&U`PAM2<jLXuN72l1sh6VC#cr3c&pH7tKdWFrJ|}-
zw4p0fDms+1{-#V+m9g|P(V>i`SBa`BmR=<~RI&7GQB}>-t3{`3(SyG4uBf`p((j56
zciH~;MAbc(eou6|$M$a&RgEmYQFLfz`!|WICYIhLIyJHVqmU{J+0Y%1LJm=g@sG=p
zY8gvkh8&i$^jM^dW$CfVA(o}DLaJ3PeHC(Ag*@o{HXzjomc9WwY+(B*B2^+wPee|M
zZ2zrDwUwoBMGjlp{@ajh8%y7YoVKz3bC7oqlBw3_prJV^l;Y#a`#4KKj)op*>3K+%
z$HoJB$RQ88(f;R<>KscyhaApPIx*GQ#A^x)i4K$cqE33+`awd#*F(^=4ECc|q8#=r
zXRSsBbgiJ-LR{{^c~sCuCypq&A+=n)3zOsroc9CHmukcfe)tV~$Ab^?e8>ydH+<y%
zKeBiEANheF`9ZY*M}F`}ekhBFedI?_93c#d5SU!UZoz*y^))yU68!hDI7_h45*%s&
zEWs&D=*?o+EP+XiWC<QwLVIFYlCuP_ELP*Q1ivhS$=qZK1G3ohWC??^gu&$b#6#g7
z(iCfwjUaho>`gVbnS>z8Eu2dWP4{D~$-5-b>$Uwilg@2*KQj#n1<`^sT5ut+lH}FG
zz}3Qc1d}39s4HuVJ4s+5$yZR|6%<4~!GWwfmtQCKNWv~85tt+bYkJ&D1}4eCx?EfL
zD9Ne7Bo$bb;}*FT!*kHohSE74fzC(RP~uO{`A-^3WPWwSZTm_?BW&sOLC1VxO1v2f
z+HgdmDJ7*()Ko|khPFGL&UZm~66i|u6JYQJn9`F?1MJ%X`xEa+3ha#av<0a2!SQfF
zJRC^o(w3nakOl{mFbO+;=vl%lOeyp%WpOzSD~F#`g^K3}#B&4bg<uLdAce&h+~*bC
zXtGO^e486`n^mX=&bxtCsF!@mOICp*g#nSQz<nBHO8-}k*@oE|vkkK`rX=qb{P(g7
zlP>tDQ+%ZH-$%?>Jm33|7ik=uQdbbDmw!iB!ot=e3bq&K;8Hu-A&JF@V2492z5qL1
zU~weZA(CVI=W={5XF$X2lOlgoWbiYQf5zgSh~J6Y>7;}i7b<B=mx-R8yj(tZpU;c?
z|M=3Kb`Ser7__|+r-09ho@dzf*-O#$C7T}FgFN@J>5pdQ)=Ue9@U5Q0ZLy-gaEKIV
zBVAcjluf2C$$8?qJn<_Us9Y37FS6lti#V=D{E9mARxz|y{EXssG%g)|MIGS*6nX%C
zMsRNM@@`KGXXW9z{EQfUMhu}ZIxmKuXHW1i#o(7>2z~HtG3+&aBHx38_n;8+V0?B#
zVf)Y+ikneTGa5#0W!nk_EmpMS+!@vBvjT3UVryKALQ84MFTQF)p${o0_NL8N{-;Kr
zoX4;S6~qUE6J5GhqGFZUo8&gmTwEnw-u0#kTj&4RT7dH{Q@%Jb=gd|-sHMPRDHudv
zF3I-<hkL>x@~%szC0DOLetG6^pOY>%Y+AhKk!Oi+Tgiy^RnUbPbY3n1)&(ps1lENt
zE&}Qz(3SFE1L|umE(Ypi7MB2Z35!dCx|GFbKwU<0e*K!-fv<=9;iYKdq!x~G{0Y32
zz_(-ag#c!J3JjjIm!+D|1&F69viU7KqW@=Yx9?@Zy$tjtL%-(RV_{f`)%t&2-HXTk
z*y;8BEowX!2i~FGPD%1merP)?shBs$vCG9pcp7g&=QqHS9Jb27-*&o5z19sAj&RmT
zxS_<8ksEt#MG}EJ5je47f;t&Ev0;Kb6*!SFK~MV`1(m`IPO9MAiv~zP93yEyrXY-!
z_oGLrmDvL<`X`~$N!XLbySUO28s)P1DsOa^Z!aUb5n0(uoM%~q<@)c{sNHy(6!x#g
z70dp%f{=oJtp1v>LC72>`7b`?FV?^B7yS2A|4Xi>@#1_M0g;k^e16w_8A6}f%7?Vl
zCz8uyyo!0iAdkH{D*@d~*qbxW_ah|6s=K#1V0FiXHTdA4FNWlcW2uq-AclMp$5Nwu
z7=;{0W2w<BL!o78JT;oND72Qv4JfpM;yeAW4g4#k6Zpq9=Rx6~Q#%)q2*DzLQ~dm<
z_%)sJYV`SPw!|(Ag=C>G>5NNIXbH9UxIYTLNiq2pM^{h#0S?`)y<EUvE`%r<E=b@4
z(pmh0bNRrY2Db1nTj*&(l27o`30_6R$&<WvlEp=Qry`!Y@hIUtl~7DR0K&H7PhjvT
zYxXps0THv;sK|(e-Us8c6rC4+&Wk~GX<myyuf-rL4EvDJJ`_Zjrg6yYki;K0V(W{G
z<bv|K;dH14Tu=dvi@1OyuKl+3$=k19kJud5_dh<~;*8jZzVW<heO|Ps_@!v^QtU!U
z@LIHfE!tAN2U+YvUC4;qe7yBOWJ_@~GHs@1p>00ibg`nnY?NeY@C;4^gVSumc_HXp
z$QGPyzKbFX^t7+KV9tcC=dme$fI=V8MD6pWoZQ|AyLSJ_1^)TtJ1g#AxU=j3dvV|k
zWkb`F<st6>^LYi5io=_G3K*PXH@5(EEnqiS^F<rk+?L-j{$cj^tl|Is3|pdx9_yZe
z=+%F%)cEV|feiD<Gh(qVST3lSvzH=k1l1aPX+pkE6ZN#8@KI}40l%t%-_XNwC3LN1
zZ<Ai}A+M-wlH?4*KZ80uNj@O>AE5Z{+qB%?H`n9Ke_TrB@`husg%Wv}M0!nxza5-D
zu@64Mks=QrVz<b4wf}*PRfhz@I&<fIH#e^B^yf{EfI2z>^k;wi(~JHbNPqg#pZ@fx
pZ~IRu@7Oh0ugg!tJM{y-ynXry4Dj_E=&v)Nok1s%{`8?g{||?(Em;5n

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_v2r4r2.hpp.FFFD55E6D49B5037.idx b/.cache/clangd/index/gridwise_gemm_xdlops_v2r4r2.hpp.FFFD55E6D49B5037.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c73ef86a0a06cab4f0dcee22d1339b6ef00ebe2
GIT binary patch
literal 23224
zcmeI4cT^R}_rRIisxFTQ(&VAf0xF0K2qGfNqJo+j1WnLH6HPQ~)YuYJObj+e5kagG
zkwn3+Ac%;lphmHv2o}IXvw@w5V!_C7DC1nu_phAu{hgCPzBy*+eYkgKc4uer+?lz%
zL4*DLl0G#u>OX$)kJBc6H;y+lGBPFq&;D-K<oSY;QNERt(T^L?2ak!2>1F;rpzP02
zQ-k*3o^o{9#}4go7Ij|r-R0$9zc@duw4;iaEAc({ye=s?B`Wg=e|=!oqYnnH$bFs~
znfRn##jOmz_qi5hemuU^rDjLUXFr#eP0arB#g^&zxu%QkjBc64MUP+o=R%v_7cP#A
zebB4byucHWlp{`gd98oiIyw4SuzjDar4P&wxWDv^xt(Hy7M=0v9@^d1Ir>MpoIywS
z_MGUoYg@vki8Ix<%N<U*Wp_Qd_{hDG2{(PUP90|SDW6-C;?iNs_@{pi`)Q^4<)k^e
zV~v-ljOjSOgT)i$7Wr@5jGSXTt90}7y_(UMT*|lK-`?q3P_oLa*%!fIJC?@G+Em=3
zYWT<a=B(&MEn1oX;#~Boilk3mwoJ7*S-z;srtrjCG;><Nh?Zl2(0&^|ZTZA8eX>(F
zDXlk}_x{rMW|q}mylUo&0|)JA|7hxKJLPuJN$zonUd<-7>>ibKZqD8Oc74Z<_G;Ds
zRdClggEwA&d2&VK9*cydF44NnX1^@9Hj4?m`*XX&WB1?H)onOy^RQ(5cBgZ(y5Hsv
zvHtyB%#MQBgALscCzYQ&EpBtVUq#wU_tX74CuX5E$GWkF7p6x4HY4x&^mV_@a9C`c
zIppoJHOqFM&W!r%&iUkZ*S|Ybquy+_X*7!dakkHBH?t*o9;NJZy<D^ZW%xri^Fse4
ze}3#cWuSAeI5ThBQ(y08pBtJ@unhe0hjV>`pZVn-Up;hu^qaxuo>j+x9}${ADE*|{
z#{ND<)r;)nE4Ia*aC3h6S=`>R66ZJ7mui<KRCsos@>kCt6&FvfNQ#)%YDwrmL*miD
zOJ`ON9{=XReXm|^R~l^3On%mU;n*nWPg-Q2-8MBo@&2<Z1E+syaw4x%e=(cOd}Qrj
zyPLbkh2FZ{?lW%no^MXSG}>np<u~n_=ZfVXk2ghTo^<y*6+Um2#wvB!KwIAguTMI>
zir8`LW1~BZKMgy){nJ63!^d8I)FMCT+bIK5Z>~Kz<wd$xuhg3SQ&XqzZGG$ey=$v?
z&D&aV<IG2MJ{=amv3q*3gHOivBUzKzRP%{JhK&z{JTvbHOgQ_=#NP3?KSuj`7DYU}
z`!w;F?O(|Oqk4ZfW4M0gt6g8+m}~m-b2&2W%RF^|$<IIA448~f-n`h}>gwEwH^+~f
z=r}3IcwgT7)z759g)YTk&W|gOE1%c;*1%immPF3|+$AqC<J|Arv5&d41<R)HThi<=
z&6Bw;#<wgP{Ojkb*rF!5efrNgeUHvso;IWWQ0X^=TCXpdKTtFESW(sIkJCRlEbo}q
zaYx*UkNc;+@tqJgvU7F&z;pf&f(9&|5O#0Y2i6A-&gql)g#0=xGOnfetIgl7ZB~8U
zI6rsz<!^+nU1K~7pa0e2c4XVX0`>-Mv3^oJ_4MmW*9Kl{v%%1vb8Ef1QrPhM=1PkV
zfBJ43Wpd|uPTt|`{nH(1Coc?~XZFp)6qmBK8@z)2TBQY9dJeq5`iBwg<W07nO<r%G
zWifBBM|YF?9&s0O*Ozk)J&xS_rem0gx80)e$Hi~@>iPb>6Ia70+*{qM^_0ksu@fff
z6NmlX)%2%F)rIa~9G$S??pkr|w_lY<7haDs{aaL4MZL`3XBDHFKESPG7yVPsNcm{x
z;uUd+#`g9aend?l(s|ycoRD+3#$8#NvUSL)khIe3jF+`Nmt9zuD(2O>UK!Oh`0(lt
zF`eCmH-C70TdzrnO*?fpsd+Pb)$X9G&^))}%dBeOc#n$hp1V(<TwM26yLI=*guPif
zR$NqayOZtG7QbCx7`%5)znODh3wJjD?s0g>pqn#&_pN&Bbo@~6rdeAy{Jy(t(2*Wx
z>${Y`G0gqd?YD%VHuXr`J?;5vgNNnI?w|Ueja-#ELk{xkUz|1Q^X#DKyZ7d45*M9l
z->Rgoyy8e(!;#eNG2M^zEgrpG_wn14E5E2&{o5O^E+_k;(d?h5eOV_O88z>uj*gnS
z<bfGh+>O*6WS4{7ni3<KSZwm2ZRL_5J~vx@$qbtkl~B$ll>4wLp(^crwcYMCcE~t0
zY)n+rxNd2jS5rb&>?<wp$Im&pUV~frG*TBM?P6rwlu#9i3@Q5Xcg0IuU{_MgPtAX7
z)o`F<r-w#zwWs&bX4s5~&&yWlWxJ+?s<;h{34d!mbg761qB5_=yp|0ID(+{q|Kj%M
zVFMK-KQ|*)FYBEe4o1y*Y&4_?Qc=GjYdigEHuk<5){^{JdAqB8m!^c4F=+C&?vKU=
z^w(e)avfnv8-}c!5~>oAJz-e0Z)~P(a66*15NQ{ZOKnW3$}l5`Nr$taCs^S2q@r0i
zvsyPCsCdMLb92%b1pTbRPNaloNV|+w(wI<{(f5~(=R4mIF~brmArskUBKM|*s*H0Q
zlxsd>?g%s7oTwBcyFzkr8xyKBAv$i)kUcXD8r+st6ppmvq@u=zs!ZK?uk_fIs+ShH
z6)9zk%@o^)0~OEtH2TB87Js?K47VWSY-E>BDrrop%Fi}->s=n*eQJhHh)RgShX|fc
z300XJ_PAoFU$AV3I}w$;Jb#z((Uee?-(0f3p538el^M1pD#ggI82L3N?^O0*zVpns
zSuYK4L#}2y(k@5VO$k*A>3!tH!1(K3EwC*q<y)I?Z5s|$96o<1_etc5rxw_Xh$q`j
zwrx02@rsOC+q6|??i$>kR1k`^p`^JQ6RHw9a>k4^L50mU*pa9#LfS>-<~Jr(C2Gvw
zCl;l>7n<R2q>}Z5ZoSZ^DWNLSJ<eUZo^tDnh|Nh)oghtcZ8#V;;}CfwNl%Tg*gyTy
z@?(=Nunj3=s?Ai}h6BwX6Tjoa&E0CT1-2&QZ*0D?Z8%VI!iZm9oVQ3#GsE^oyc65)
z#C}Z)RoQ;*tV>%y_kafblj{gUmLaH3Q$kgC{rbbJWy{B?EpSUx%G~C2Z5j?#y!X%&
zhfFl=o(2yjB}5=?1Zk|sgsLR|Jnc;C%wKP{z#=IjT4)|EST!Y7WnZqn|HfxAd1km7
zQOQGgd8D5=CRF9%!K#HfYkIb^z!pShqRm9xh65F6?RJ_yaQ4!2Gb|HvG2gA2AKa8s
zmGdj#rcLjcu}y;qlT%%Qv<pbbX-uffCHuHTopwoS8r*@XEJfO-<Z2ods#0*qwBW6y
zKn!I^QqfDk+e_ZBDWNJi?Ho?nF8}G28FnBl5kj{J!M`b?D#dXhZy#)#Jk$aYAS&P6
zd~e@ypyHDFOC27bJ7{Zx^+Y_=W~Ob!fr`r}e8;C&Rt&VjIwGECGtIW)K*hJ?pL~_r
z>%(YEERd^+L~<n3HYHT$p3C&+6YI8bwZN97gh@7&Y#R<#Ts7J!@<8W1lgw~uBCf%9
zHQ1{up(@q8FPt~)d$LZ0ok=B2kah`as>Xz>Jo_y(BVolKK^i=WsLV%}^GWY&OsLB9
zC>inNU)<N?u0&-oGTw{Kni8t=YRB>yE1%tM*#c|GxvUbJuM+f4303)9I$1oWZT3tH
z+=Zyjwwc|!;XuVkvnG5!+heGak(lA$w$H%5M>LB10y(`v{?yzk>RZ(EE$YU2DQ>wG
z1CGW{(HQ8j$1T@G`JcGupO8;rrxVzpK7xvR7PmZ$yD`3kTV8=&jGc<1{%zdyHk3cY
zEgwOC%$q#s9jQfC)Gd<l7HJGMI&{EC&6|dKYHFHK_q%#-!`U!SQ8RHsCLYPoD2EHo
z;Xb8<yrSOZ0&g-lb`IO{%^H{MZi+flKhd5(jfy%+Kgpi<CPkgBpKMRftD=6R|Hhsg
zDMg*4pJGqP8AY9{pK4F-o1#wBPqU}JT2a5%e``-|xT4N%JF_EwAQUx{(?@dl)Lp4(
z_ejo*aTKSIg7P(-e$D&rz6R=V;`E!Kd=uxs3CiO*eH@g>bNcxA**zZWKjqb@yg$1!
zRlHgSd5K`PL}*P%OGRBOSS^K|C+PD8dwL@kHBWHQ6TBGb3;KL0FA($v@3VUW)GrbA
zB~V@>IF$&lth^Gzw?r7kxKz-WLj6*qXQ|MK$;$+N8I)HD`il42sREW?A^29n@~Z@W
z71Xa1+^gXE8U(!o%3lcj7w@zC3#eZw=<A@oPH?J&_16i$b+G;rQ6D0L_6-qxhKS($
z=8O9IP#z}g!`^47FcGw0nCKfOg7%A*U1DXA;h!z&W#j&JsPRM_kIB1ivk<Rdl9fxc
zhWfY0soOI<3p0*r$%BaaOr$h;;pCtYzlzU>oP+or$a#p*gItLCLdeC4FNVAm^E)vp
zzXtO)kV82>lmqfKj!%PpmFKVWKwixA#gOmt{2j<IdHyBjI-akC93k)#!n^WJtS|dI
zUn-u$@#+<1b_KPhz5uT#bN!MzANpwX>OHaFJ<*4DMawSZE@Xd{@tT;jDilzKf|>DN
zj7KcSqZy}jf$7{RYTB#|_Q`LyrMcSh>JN55*wdkyS1+T^myrXj{|f4G1=%saj;yXD
zYwB?G>UGrqI_khUit~u#da?73<2>RZCv#58TvsMf;ha-A7smU!j{CVzjLSH?GR~f!
zHm{a(E@hl6<09Uvi0?|LGrU^FdlvD%A@?uh2Qa?IcfZCvGtL)0@`YZsJ$bcMa4Qwu
zA%9dV1TZcWJj$THLGUm@4iVi#M36s3{3t{O`R9us^F>f!wdhjK9wlCVCR#reThaPj
zne;j1vo-4^Qq=vzfc?U7DzHtl9s6p)qIIOlT$UX!%Z~J!B0WfUxB~gA>~NJaZCNrZ
zHcoKqsMz}P*KOy0@_7=es!HlsB?Z!|^p%|sG@Db}j|g-|y5{s5rMG|-)Hn^M1-0+D
zKi$ylV)OrXt$%gbviI#VEdE&LFmhYxw4dWdEdX7;;F>qBS<saRMcs@vn~@21tQ9o{
zwM#+msrO3PH2CHXYZ<gRDC!E)d4<@I_9jK$EjsUpY>>1DNl)dpACqp?xbQ&>>gpPv
zI^RrgOV0Nv+n>77>bj19<JB7fdYtFgqnzDQt|RT2yqe3|<w7p!+LUu`nS8D6zg8Yh
zN2zWB<NaNC{SYn^--Yvh%6T!xx18r&$Pv6}1P|mfyk`vLoqYeD{6LmJo$sH{*yZk~
zy}Gd{rwjGQzV#K<h8fpvzFRimon?H$w|T(psUO9w5BW9^Ay<ndtHn{Y>0AfhUM990
z)w6khQ&=D6TCsw1IqO`=<(zdn<fokFQ_hO%zvV37LXO}qBX}T>;Vokz@8m6aLQdx`
z(;=^w?bpg3X;afKB{XkXw4!&I_^}PsG%Bw53i@QC0dxUKA%?~`jT=HwQz1(U_UnbN
z^!nX$6OUeddeXnaCMuVipHXiUl`D|1%F0#7?hktwO+E9p{!wVnx3-`bu9y9nM&nGr
z_`YFOw!VT|u^Sa5cZ!ia)9c~Y7`b<h>_xek?>gU<*X>OH>v!0*M`%#-iCxcSMSX>I
zuTW1mysyT()fjLL*2O@M$GUjPC$aBIJciADO0jP#<a=0m5BFs9r&#xtaql)&XQznc
zsygs$60%G}*3=5{>ON$-?|tsD57{$$GO|pD@?_K@8Ol?TWeSw1pbjZez8_ichw}Z%
zaX)fm<)<RcR47kHj;T<709hV@@&m~60F<X8%QPrYLyl=seh^t6gz|&P@gS6^Bg=Fs
zPe+dFP<{wm9)j{i$mI~~!OlMeS!O_a26D`R`SWEZpLN}Om)iZ3>_N>YuNKHk0hAZY
zO5yw5woq=*^owMr2+E6O&m!52$*;-I*JM}5*X7>V<vxsW$gOV3I>tBURyQFR%QnTb
zEtPvs51W`!88%Z?)LnwxE@1$DcolVz;I;>HqR=f-=*#2}#qJN`xbax*{upwc)FV#n
zPxbp~y#M;b&+aGkOfMEp77MO)wn>-9$pBNoE}pchOXh?_D-TeQko>uEHJn-{p9L+&
ziki(1q@ZNZBAK(M9+bN5T<^yvvtt`BrR^@E2d%pPQrhl;oG92NGCQTH4@KQWu^ZEW
zEb1OZj+69p(!2T9fFND`FE#Q2lX@qx$A016H7aVnsErrfvVyjW+HH`Ji`wImPm0=;
zkSj%(N>OF`?~5+?A^#<r|0UTm`31@R0_3}r`CZ71Wz)rSJEp%xHeCWaT-Jum;Cv%w
zZ3N^<SsTgN|I$XUm`?}J{z6eJky|C|%gkLR@~cF{7(YU8kD&Y!@_Pj3uaNc?vSIqK
zk@hv@TBNOo9D=nW80d#!*AU!?<zIlc3!r=<)-Hq`j<w<MvuilaAAz+IP`(0dS3q8k
zwW}eo!P+&D*JABj$Q!VB!~5*I0oJz>Yd1o9EY`+C-h#DT-e=b>Fn<EpCP4XCtlbJZ
z6}zS4zHGQi#eS)H7~^#8mJa3V*e@N*Ph#y!Y{T?VVeKi%*;tzmIR|TV-e=bw+=u1A
zfVCH(JQr(oAs1k6!Tao50P`1NZ4s1T!`f?*OR=^Tav9c^K`zJIa>#eF_U`-adKcDL
zfwdJ-UWv7pke^`fllR&63Cv%GwN+4Vz*+<3TI^Pf`_d(CMXkktwRjlgI_y>l<#pJv
z4$31r^GL1@HGPU2#hFJzUc;HMfxL+`-vl|1GmnED&zZ+Ve#$FPc}wc)DrEVv+f%+5
z<0`&e6=Vai7@&SNuT(=06_ilHlG=PlJt&wQ6y8m%>sJ)b^Mp3cP~{8e`H%|)^8&~v
zf_Vw#Qo+0wa)n@C0l7*ruYzn4%nguV2<9&!*9qozkV8cC5D}bTm}njbIYH7UNH*;J
zw@F>LN$-{&$Ot5L*#kLA(k4OueUf$`<fD@LQ3=#{RB}5ifx$OZGS7taEXh0z@@dKZ
z^!w~~8s<MEnV*63vy%B)$O*D_g4~{758ZkulhDR}XEr5BK$-+(!6u;xQ2PVOfx1tF
z{bsfswH_yv^{|H``3KyK$D;Zea-1|UPI|XyIn*P1x@FhpQDiWW5=TUdU(i!g)a|0%
zcIap96-Vra@>8PQDJZ`x1z(lEWceRUy&f~aWN3hg)#-uj1IZb#6hBxge#8_L#Sao8
zmrEa&vt<~%*G}4`afh9?$sb7b2b`GiL+$pli8<YYC!S6H_B-`!hBq5{WZDGn3SJ=-
zZS8s4lDcRLSy3{-EZZ=?BAZ`<d{s8T%9w89lM`#)xMwG}4QaN)n{^PiKge#@C*#(P
z+WqJC-o&#ims^&zGx%iMK93JIn}dlUU()6?PiSPzY?rZnt?iY1hrp^5^<V??BV_dm
zaw@h;#XZ=7kdCd=A=hH7THFH~K&v{)D>$naTn}dOVmYf=$QK2xi_pa@7OaXPKNYN=
zLVhb)y=6SIcD>8pjD>CuQ_Fv{;5)icOv}we#kc?UsU@9clU~voVnO@Km<<mf`D>qV
zCLQ(|vOk7;QB6fXj_i*^F5=CKczfzek4f{KI=Z7>2XYtZdCl{uy<%*6*hI0UEm@sc
z$95j;L93xF$fW+pzzr(+a$eVF*H+&)CHl6!tu>vT(|8Wj>l&jtX!+!UMVjrf*QKNO
z>8#g%{YmTOV(oPkay>_Iz!4nG`bZugk%vb!KFbB3WlQE0P1enhdS>65n26P!)oQrk
zwS}|V0{N<7brtT<+!L(sL9P+3Y9KEWt(J)I_P6N%Ai1>qAQajVin>k;SjP+kF>#X5
zL&=vKkg3PFImE<1|Lh+xSIqffKu(vDJ%`!UZz6bAp`q-lG9WJl<Z9$q4fz@JdIoth
z_F0UBSpHD#6AC#B`$R!Lg1wI5p{&J^Vy~l+GqG1D<Sgu!1vwA<<l!KeKOg(#L%xB1
zZa_}wywkY=wt#eq^F9PQgY(XSe3<h-4EZeQeHQXL&ifqX9L_rj@_EktJml+~_jSlO
zIPV*fZ*tx@As2Jr#f)hL5_v`%<Bn+8pneYN(@C6?#HrM7HE)VOqKb+dD`{h;fwa$1
z2MafEia}zEI|OA1>!ZI6@6qo{lBYX)3NOhfm)L#+z1O6K#z-Yr!e*h%W??9;>bItl
zCDN-K<CWM0y8~%<AWJr;A3~iEA!j<K(@k(vLwytzD}RF&uz@`e3l0T!S#~3B9y!C-
z-CO%oJF(>DUzH;QUo0Yjb}(_Uq_h8Udv#dkwec&+pWmO=hY#TvJ&LUBHAqoAX*yX@
zeX>aIF{dwGBq!uE#~;w(Il~w&MDjvpLdR=`#6-!9kO>W!BI^-IUW`oWLc2o3s^m~)
zLL;se5<ns^K_)b|L?JGa9EMD&PegXhk-QX{&@df^jJonNWI~5wh1f$m9GOr9sgUND
zmm?F}{tAgqk|U4_4O&vvXk;3VtXcD~LZ+)AuSTY;A+JHEYap*hrfVUuL#FE>uScfq
zA^(m{e}}vQnQnj_gG^%}Z$zdWA!j4^Y}AK73<`<vaL++~Xt)RQ?NQG><V_cA>zVu;
z#JCW-7ot9Nz){F}<5P(I85bk}V)PO1*os<${7WF0BL7mzWyrq_ayjxZhkOh9--3J_
z`QL`T6Z`MPA2B1g3;XYayc_%PhP(&+?}5A*`|pLEi2V~GCt?32$Thf64erlepBn66
zgFk{CRD(w`4&@v|IcI886tX_%bd~QxhxvN;xXSlsT*zw*c?-tZc+EA)#r%+BK9B~h
zDQYRNDdjB~zv3NV@h&vRj4XBY-mmxpkcYkEgBdRn%oYf(s0~!qg@V~a$VUauQNe=R
zA+oC=Xfh#Z37Ra(Cj`w2$fpF&DaZy<V-P`k%OuNXQfpRzxMUd)IbAYKm%!ZfkYsiU
z@_M<=dfA?C9?-}kv3Wy4CNp>GLQuLef>o0#1ZJ`gVMSdd1*~D_P*LwoKKGe-ps0@|
zpGS<t?+>^->FZf<{`H$h%MIzrp7ht>`G4P+i@3EhpzCj=CXgXvtmjxS+VBxCT3lK@
zb@@p0=SaaMQgEV?jVr><TsLi*et_&e7D;`I*ld?quStEbLB1jNxdFLY>Ql^^T7EJz
zHb&U8k#P&sY=N%qe$;+Hb7kp}LNYFD2)|_+|3sQU;gm2LwM%AG!pNeZdKnw-e?+dS
zwPowBbgZO(i})9fvAVPkR_0ecou4sx?>}}eX#u3f`p{mw*iK{K$dFzi&`Y(*hMp9#
zp7r51WsX@sN8|64)6A8vb7ea^C0+aV@q2w+wdb&+{vn(H!4?k4u7Yg7i}8lY0Kee=
zCqE)3Tol_}6ocs{(uh2=TU{SB%+8=%99}JcOw|;1trW19osptGkbEAnr)Xnrm-1hZ
zfA8^+=W(&G5*HU*g|}|l$@6-JhO^;jKGx00U}n1<>y|@ajlEW5pnnqUPQsb(d8|7R
zxe)6LA(vuZDdSDPC4Jfieqc<3vr1&8gzW_VIrRg6NML9=5o92j47NqJd0u#Aj>co)
zKkf{4gd=yoF%X-zdc4pjp1F!NBAX~ShGVmS9g8%v%vIbGoZq*3+LW$jJs<?@La+;K
z#c-?(hr9vnHbCBjbz30kU|kM&p{J;*1z1-A`7YMoh5Q8To<M%eTRr8wQh%Dvq<E_;
z$lE0IZIUDFS9>J$J&b99Iq8n|(dDdRPV>rX-j+>R%6OA9UeCtL8qukSMegmoBo>8U
zY&M&8_g_qZ=|)T4{bPnOY}#hO!HW7C>0YA_teIXT$JfY}aV^r-LU}E6tcCIgShoO!
z{0p$-0u1sm#JYu0z7RVv#JyPg5m*-i<q_C10_I<Vbt|BJ1$JHm^RL0WHBi0=JFbEG
z*J9mTC|`>m*FyP5tlJ3X8?oa?D38UuSSXLh&atrm1guMd@&xRd0P}Cfx~)*Y6+3T*
z`A=cpDcpf|vs2jd6n15tjdj^jo{b%|q5J~YU4Zfn*zp3C=VDzhl;>jST-=M5Uxal<
zP+o)`i(vk1Sa%J|uVLqFFn<}=l|gwKb}WPW%dxH;%FD51Ih0pmT?Ld^V8;q5uf)1a
zD6hoMm9YLQtgC|ZD(qMV^Bb_v0ObbkY=HR<yp@4P&?u^b?`Yt=vB$QWx2lHnYTlum
zcVY5q(K(uh?<wjA(Rl;pB+)qu@<Gx0Amk)Tmn3yyW*|v&Ok(SsWcn_7CP`pJHc1+k
zBn@Nw$&z`pWDB`{GBYD&GAnr}v*~?3e~>J_iy9=pucXhEI<fp&l0FOa8A*Qz@>xlL
z7IJ~qqd<DMQbp!QQjZeI&n4}1$(oh_TGGCTTqkMk7}Gf@abOx>mtlr}C(`VMo=*m9
zpTRtzy*D<V_`@x03UL6ANZKQkGpjLM(q=QJ@7)jwr17O2YFZOhW}SYyvExQ!Dvu%a
zW2g;X*4#I`&~@{26Eh-6LZ(T`g04f*2_zYn8sEyHMJehvWPJ@e&>3XvgP&8amKk>b
zF97mi9RPWtXY%^Fc@GB>2Pao<naj2x4t`ylGobm|GSm71G~-M}Ug2;2uRbi@I2&>f
zGR}dVhm7+e7b4?A$i>LG81hbRyc2`+Yp`(*<WSBylmqfK&NvP7Ro?h259GzXaWUjO
zyzw2#FL~pakn4EkI>-@%afATMTQBhI1r5D^`ofD@^M=>XSpTo!T`Kq<%t{^<f)2u|
zYL*a~1*fX3q<~e>sop9%Zk3$aGrU`J+zmNR>X!zsVw%)H4O+Q0$v;gR%<`v6AEZg|
zRvr~KT^gJ&eZb@ivUh?ELQ)fCH9-bpH3{;-1lb?z4^EKZ%Re+heiziOkj#Mz@;fVg
zNF!f7HS+e*|8S`@`m{5>9aa-Ttkz0dSt(nx8dl5N)$+Uj+{34@jN0Q>edm7`<X>$;
zj=BYn7Y1&AKz4-gp`Q0({80+-nF8ZoUg4gvV0^`6&h;_ZoB8$OylXfQcsuX9opI)n
zrR^V`pK-ideMLj=p#bJkKR`hb;A<Z#ct{EkU=Cjz4oZW2T(9tuS2%zz?){B}{>H(K
zA9H?>xewrp`Gxa<xAT768D~zrwrhq@VLaJ*s73+RXe2wa3@$K(y*j6;H@LtXjOh%9
zjLVJhsZwKdywf}vkNrau{_*VcubQtvPMnqX2N}1&==?>0THZ;=%U?bCD|I<}=C(@a
zTO|kf%<Y!UcQek;&pUVS{<ExqJeAL`+z=95(KFAiKK8ZEKGc>v4CLh<)Fv5n3Tl%A
zc|X$cN9~w?D$=JyK7jNGAg3XH8svjWe-Lsy(x*c{gzOKYPOSV4q|bozLb+w3Y(ZC5
z$U2s6Q6yV2z9uWz*opvo*+*8cL%ty^Hz41Xm79=@Wu=($+1t^@{a+6DCbwdRsH_mX
zgIgggyBVK-dHC|neh;0<t$crT=G?3ufwi}$&HfL|N?&UxzFFhT%*>IDLz+0YVsN3u
z+^_|GhntY1GmvKn@}`ZG+o$6%K>@aF|7UXl)h731@Hf^k=fvvC{d3?Q-gzc@SHBPL
z()QGE#174EJ+}*8_#k^*Li=dJna%<f^3HVoRf02poto~Uk@-jC3+}AjY(<)_aQ=~o
z+NZJk2i<%i=Bx4TcWQ15?u31^^^yY;6!i?UID@*dZK?~%;sWZzwyAb<W;?mgY+Gdy
zXSN6OZPDzu*qJ^fg^LTU#qVvd{=YAS7e3n<G->^^CI8qC(sgn23^u;*Purxv)qAfJ
zhSOFr?!0_>&IpIe|M7U!EmTrM{nr3k{)L>$Le7Q`QuIA&=jII`1Yl}EwEw}5mO-MJ
zkmF_KL}zUZc~jBuI_gL#Y4t0H0U<b$4OrngARO`r9IyfM796kzat;p2!GUa~Ex-W<
zkniGvyO5vYfG3c{xK3f51H1o`+>l7_L&i~@XA}%7iQ>GYxPDB&h8wa5>TlwPY=Rue
zdB(x~ah!J?%pcDUiHG`ST#qu&osIlOyh9P+1#aWHT!Y(wXN68@1u&fE2}AOP4{3iT
zp<9AyJ_{hIXYYIgINb%pkOHV*A_SKR?_Opi+k(QdQXz<yS0;Ft!Sc!leYpT)%`1f9
z3Yfo27*YkxHwc~vnBO3H8{ql85Qe;f`gKBZ9V{<I92O#i^O-Mt&Syb8in>75FA%}`
zgo(jn?By(Cip1g3aAzS&9G(REx~N<iHSBtyi8jwfu+&o`o0Q0=bZ{=c72>nfH0e4q
zG|y1LGc<~I<SZPJg-5ZDd6)}4%zegg?0GKmJmg|7u$XaKuiXAGc3Y$W4TS&dfslSE
zg4oLXZ=WzD_`4MFJDXeIS~X%u)o-;AiLI^WO=@{}I&#r)4>E$i|JL2z^E-|t{4z9v
zOlk}$zyMwPP&^_Ok6|6{C=NIZzn+ngN94nI$PRIVhnUN+sOPxAbByT+Gi;hSe0GIh
zK_&XM5`j0@AECfUaDyxrf0_!z^wM!)Iv&pYdoBL77LTT*g`(Esz&bpfaTG_w*ZVU=
z8po+|kSlqYO1PR4Dp-ZWc!Ds&Dh%>bA@C?%-N+Y+tplqGw*>tyXs8T=YJl>EqJAMX
z9MxjiYUT`+XhSJrHM^sVdPo|42(D(lzxzy9Bj4M7CNr=1cApjX{oQ9pt&n^w*p?;<
zt(AN#89$JBN8;W|QN*=+fV>_cKX(7quvZ%PV-5c|_WB$9(T097=w-nA$-mc-jJpKC
zUC^}d5&ZT*P856+1wU%~9@UP1IW#8JqhYq&>K<y#YJ7m&JV5r0Q*f&k+?Jg|8g7$@
z?HRwqtzO}_^o$hsZ`|f@Y|r>HXYrT=+iT&xML67U+s<2TXZ+YhpEGOhAAgZ6iozqJ
z;MWyy;1M@qpS;cmUT2}Bv_j%%HhxElb@phXceD`59Mr9X{Z{s=5yQsx%<f6;?EYtr
z|JBCWaKCKXos(tTh|Sz3{C~a*SgoymxNvEJ+rK_)hej`w9^d#qEH<->N1Avzvr0wn
zQ`yXlelLrRYmHyYVi`9f%_cUJe|Frs<4l_;QU5WEpfh?hvNe87i<VIH>#2E`wZ}~U
z^^I2(I{2X7+2rT{^&tNjwTZi}?oX><fIceek4oT`)zgyxG-EndBW>9Du`z0}YHPP<
zcE7Rq88Lw?Mdy_wSaeSmof9FKOFhb^ck_Y2`JGC#*@4*Rib|H)vN&V9sX*d0IX(~a
zZH~VU*}(Az$RRu*!h`&gJRb=;mgi$3@8<d4kTZBb1M*p(Kg*bYM~z&1{YTX3g_G~6
zO&!;soHXrCM(fhkY{Ro#mecnf-brT9`f7Ge{pIzd*Jxzq!G3zNpMBWR-t1@pcR#Ue
wvb@_+i-5aZl>Q#QJp1|d_Uhx?*W1YZorcl7KxZSPzU;3)>}Nmr)06%DKcXlg(EtDd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_v3r1.hpp.F4E339DE2B052B4F.idx b/.cache/clangd/index/gridwise_gemm_xdlops_v3r1.hpp.F4E339DE2B052B4F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d4c706130d039d40a000375415139a7a0f5ee630
GIT binary patch
literal 11174
zcmc(Fd0fxS_y2q5HSbhj)mtj*UEY-z>D3;Ul0wLGEg@UiweMWH?$1&w+!k3vXfLHg
zt7t)Kp~Z4<ZmATKBqUdsvLwEfnqK3dpFcj2`?x)L&Fk4YGq0I5=bSm`OtYLPOo%cR
z2*xdU-nL=IZ%dFsAQ*uEZ~4uAO@vS&Fj5x?wnh9ocf+PeYtdiRulseS?;P7tD>!TQ
z9GQx|AuVH*HdI~SbbK>AUT1TY^@Y5uFYm*zZQoo!`O@wxuYv&<<=a|AU*C%J{L=Oy
z#QMs##X+}M84tbO@9CINaf!P=*hl!+cz?PZ_i3^5)up4RTw6J3>LhK+oTlolBNwE(
zU96cuCv;Nj(Tn*pLmsEq%xk#p?arw;?|MCSdiKUeSw$Q7j4~T-GBTq6*U~EqT8@?X
zz=DFcIzfTZ{HB(LH@D<i&`r4XNsw@X-sEL5A3ikBI`pZrR%XpdP2ZC_YbF{LUZ|V7
zGHPG(=qJvjPAv^KJielMerk@9VaY|<_<HYyNqR-rCK+x=76&i-y6*BUkNNHq0e*X+
zocZm~d)}|9QhIVB`E0P})T43s?yY|7Zb-5tmfmxZn=#|du}|W2a)!0)VUhb2)lZ3c
zly0~Y|KRkW|M(0%?WX1{n<KqSu5*0AFQ4mP>$u-P=%IPHSbd}MG@UQG!d)8sWQVM<
zx)PGCx~2V0ZcJl?Snf=^>f@}(%N!He<$TDVRDHY6JaclcN9I)%liiR1@VGSDsa#$^
z=JVXCaYr{Bf#&o-Q?<a#*KqA!cX(^2ZH578p5On<XwW>**S#YZg*-ho#{gu+hZGy$
ztc!c&)ezS)^=*dK&sn=CPF|Z*(KX!JO=t3{9H+o-(-x0jFgdSL#^~`J&o6O}2Tpj-
zVdiJ#RvcM!RAR9)XR|`?$*A{HKFj8mPP-?twzU(K$T8lwjvgJ$P)qJjbn%(FgR$+O
z3hkbyiQ)x^<bze^GiDx}-Tcsc(J!U>8fCI8xDT7#%|d7H{dlr$ciLAUQR2x4`2jn;
zP3s)IGi=&b+~YMT4h%gnyLVX7`QOg_TDDXeWGpkNmka8bbXd-MO?A)$&-jp(f?Ix7
zWzEr+w(gg<#^$`6D^tDmoLgD>mi_IGReN@6j=OTFb6@=-*TpU8eIk704IYS8hZp+`
zYft_3_;3Hkb6%SkRXSKUI?HWZ-g!cH$_dxO+S(4@cEd8K*F9A`+4{UH(e0>cmhuX>
zG`&~P$6OU3+n_di+TSB*2ge!>aN1g)V14%TklR+xClhO=muM}hcFNv5bHZ2I>&be~
zfw{R`onvzQ3*%*X?W@~r;bP;I<P<t>*!xXS25kCe_?~3B+A@tDpE_4B^fehcGSj`@
zvFcK2CVFSIBO!RYpZAd1<r~VfVO^;2iVx<%aM2a#yq5F{74i1>2UZw8v5X8j{>FZ&
ziJoir`-s#tn)mwoSpE8Dr+!T8<4XM-S959%zxu?Dcys=0*S%4{?;QQtB74t08<#9v
zGobonbyewAE-6vc>^izjZsoBJw?kc4NJZApUfCvod)iLA-{9F^oqO&Z<aan$2fy0x
zZQ@y>Iik(N(%<?s`?5@{ifcVm_J>rc^0J=;!!qNS_gh|YPUH1ix5^O*7tORk_xrd%
zBBtFq+)y^S(!s_eAjxn-`PpZoyZ>^YvS@)>-Pzbd?Jsg)TuOg>Rb|WCDe}_Oes%iz
zU`NOB(+M3Ha$X)otF798mrAxzjye>xcb+IV@4DOu``gXw@|hRU{5i4i^VIM|nIA7^
z71-`*^(fvsG7<WQcjRmxYa=r-euCc41Fd>Tmp|X8;gGiErgQ$ohvtVmvbC}{KWps-
zI}_BGNG_+H9M~cA7OFHV&u%&B9(~EPRDa|0O8eS&Vf(n8l*pgeTk>BO9@Ypx7`=S9
z**_<Z4n#kVzNcw@pk|+c*JRiH7WXCIGENOZs$J@^|EuSlJu{L6+N~V#OurPo=&}3Q
z@SLdSAhJGX(phn{AP@#Wc(}FUi+22&)ngh<UTqF~y++~4Uq@R5kF0MBa@ldPG<`p#
zn5^om|NkyJFAPb(sQDye*k0+NxcrxUz8p)~KDj)jqGEbu-nsh+f7>g4MCPr%)b?9`
zD!wm&+Z2%E=627bwE9cG+a`;JTefdl8p#L*Y_awq_vbANQie)e0?t#-Q~ldRAdUSv
z{!us)@Z{&!YBLm}93}?=;~=oGw;?PlsZbDHetv>7RK_e`YF_H!9)xT#%4+)#^B^~6
zI0%!yftoi^?`;Un2pMmaLH$xXlwm*2;s@0HfJSdaSnRGmjGP);@>vQhY6`d={dO3C
zdkAD0fM*Ffu<(@c%_XPTK2e0~c#JG)oCPQLHe`&|fes1VXSU;+sA3i$pyq>T($n;@
zv=<34$OXsrheI$+Adn3NgL@movUB2&@z-Pm@P}kEOE5GJhLd_5!s2!N@?WOo25wP=
z8knU7q7vAzw;?RPH>bZkdzWcbhJ!K7ZlJas&!wj!EPfvZreD@?ouLe+FpED>^T$%@
zX$Z@~g)2t5{WC{S8459rFHrNvEIkcjIU3<rl;)is+8@ebmRu;C3+MMXgeBZ-jppb<
z8+R!~70j|rZI}AD2O*~@2VZ;YVewuW4#Z>+H4pV~4?;eFAys4MSJC1Akj3OeC|d|6
zy$xY0bos~a{yg_iMc5y+v;x!$2J|+BrTD9I6C8Hzb$>Vjvm5}j2Y^~{Ls&`<{Wit&
z<HFw+p*&_u0LBSe6?+=Oa<fJ9t9gE4tRhsvEO`Ls0oUG!u-tB27+S0P>j_q1f)#0~
z^iavShd>5FfvE&5(jBR7Q&)Le?NfvzJVrS%D+g118#2Znk?a-wpgALzp&Dk{1Jw55
z&Fg6hOT+q+7uD`Q4pWBGm?Z$H1>l|OX$Z@6vwnr*>1KZ_LNR8EfG7gG_BMn?;J#wz
zmP1+sfz;&@ns!d9d64DKK=usOC!JX?1dR<rZbapXIS-WmX2!&9SS}2V3j<EXoLMd%
zj7tYjM5ZiP3?#)siwKbA%7CPd#vg&zM_@yClI1=Dt4}oU0<*fnY@Qwi&0?ShKg~8c
zstwxpO_TFSF8*i^k8_Z74w}N_8e~_4?0I|_+1;ga9kQ#V@qJ`>pT^C|p&5<k`#(hv
zPigEYoZ=^(M(A>;{$Xp6ncdT1IS+}4E>U@w^OSh%5}C4`m&8k#xI32HCE2A*Y=^}z
zP2vr7iMM01W0Cj(UE*_CY}}IFK$qA(i}gja2j~)2VzJq3`T{*-(Jc0Vntni!xJDLR
zzNSCWBTj+E%A^?p^oX9ZTqw{91;hCb4g-=fpv&WEpc75g(|}GI7|ts|I*_CTT^?rw
zolKfu0wg6smsgNdAStEsbs)J;;~PM7gT~2FKN%YF3Xu%$li_$Cr$YTynw|>nQ)zl6
zQjJ7I$cC|8EK-f7akS7ZT4*)5bNyE6Z?Q`*^K||&;jekpEIzGJ5MaNZX86Hkmw^&!
zoC{DcjmrTlr|}DbUeGugqF_jk7Y<Q4jk6%i;<1A2ow25s_3I$wn!t!AFp~He#MLqQ
zpAlIR*(pZ0_k87Tw*r_}05{T+<vO5g2kbL>qL5Ou-_BjWTPJYq1dDl-Y-C&-8Or1W
zkV^n<g3pk{GulL6Acq$`R;@WF{W&q)TpD3tVig9g$%GO1C06Ocny60qLS34`Y%*WO
zH3Pk7pwHt@VA~1o$ixu$6579nV~8LS*9Ps|Xxzw{H8SReK3Ku-x<f)iF<`lRIKCdb
z5Q4@(D>c35{+x~puaNO8G?Ng9)XdL&m7j4Pd(XGf=q)rQ1ePmg6bl(;f+f%Ovt)j6
z7=uTU2qdyZXSBy$xcJ8Uw(NhtbmCpGj`Rdl{ALFtmq2<mTaZHwzj;RHftfyv(wcZ-
zfj~VF7_ROObc~jNw$*^1MN)|C0t#JV0KY3;U_=+N;;}ardqWKxn|VVE8c+0w&O8ng
zP6-iCBL->K9F;2mrFawm&|av$7Y^kaW8sikC?TEzadB`+9E~3^Qy(yWt{QRm%+z`s
zH!|}YnfYW7%_lYOlNvmKBto1wRPcr({zKkS*Bk25*w!0TdKe<q4H4>*857rwWP4*#
zUf5|sk_NPi!n0f^kYv(0hOvoZ`VPAhts`sKUvHa%g|CIQl!;bZ>n;^6QCPecdxl7%
zVWf~FAMVEXk-{+qTaP_4_WUO;g@0XiHWt3K>aQlXV>^jJN)gC}$du(yBBhf&w%z2v
zD*n0uW+|4dgqD?XJkMAKEvsl;1C47Sb&gjvCe=)zt+QM$V^T}wqsaCs9|zdUkG{5e
zg3fj<x@$mm4Oo!#hwKHA{gwx$M;Yd?xO6znVfR3(XC1Kycm+*Bxe2Her^9juj6ngz
zk%>C)yE*Eghg%}Qtw6I9S`Y%uRYA=v8rMLz8rbKiS+1H9S2M;ue=Q@frSVaue-w=(
z^zpOAr#w2O$18MuD~ozKju2Qbm=Ogtnmj(ih)&S>JR>?!<19v$MdJsI*#pLr?_bZD
z)zi2LDHS0do_-4{-J)?5Qfi{{AHo5D2(@^AKj8pB8V3qRfkJA&2ZW*nG(IF09pdrC
z$_U%=1x{PU-Jyw6C>TP{i|$xdDU8O^Kq;EWX+S9r4B^cy9Vn&KI1?yk(zqBX6@wu}
zHCQeQijtrjaVXu`JPF$II0c%g&^Qf>(rEs4C`zYsBUEUFgL%P(AyF7o<8d@njYir$
zK7~|I(KsH7;%WXQBub)jE>g;+Lz!G;nM=pAc}OXbre8ryS7=;<luEwGmL+umQlwN$
z)60-j8I2Q#s)@qB`8!Y8tTjIbW-#5V?pV!G#$C%e*3$SWnsyY;APb*nb!3x@(f%N8
zQ(xiiuW&IT;3S2y456(&ikTh7Q1ooZGMlE~L9TbuQoesHvTj9vrZa82)u1w`kePUf
z-pmwl<|m#pnwb(!;|BCo10M*?-Ehuos(h>~b|<%y=r-z$Ddr6<HTxx1RTuk>5HKhN
zSn$f43I?SD3!;qkIz!BwFYmGZVK_(X5m2_b3PC&;%N0S*B4|VM=0$el17n5HHva40
z{pXBzZ9m;CdGy05=$A(SRZ<T%>#>x!>uk3m6JP3OBvTjMGywlv6KSfFV{Dbei7i3z
zbcf+#qrjjjV9Sf{0vL1w*b>o^R0vC?r|3ckX1V>yWj`Ohkk|wtygd~dp0Nv>cJYxR
zKBgGc2wHF2kf;qwh&rq<N;C+MY+v}p6Lv%T@$0J^CJ)9-y9gC7LSwRfq}YRxnVw1y
znTp5DhN*WdCaK^ah0Nqa#+7`Nw75r4;fM6teQB2AzWdo;xQCXB)>tAwk`Q94^pt-H
zKg%6JE(dt2kj$^YY;O^WXFLM*kI<=ID6k0ycKkks0h_Szv3(eD<mu7CCYq+#0M{C@
zfIm9lLbJEflB^7;wv4k6GnL1Mj7lLx`Ns#$PY;;B=sQr?amJy#<p+OwCI?Pt_P#vy
z)mo9|@_}ML7|eV00-#tx<8ok74ovwO21DmyID_mK%N>XE$DuNh<Dp_a9L%2p2~aVC
z#@Wy~8;&5m&T`q%F&j?eaVwN>g?(uP&M%<+D;gItiUqU;9x?Kd__TuMIvJHtMuV5w
zXGY~SjkA$rHlie$gA{XU93~tRCe$SXVPNU5C40_fv}0Ae#<16T0|;yxU%h&z`{(~T
z_C4@8V!F|;`77`jwvX6uMO2McC$MGqR2d1O`;3=_0&QBSB+)eXVQhT(koaJrqIp!@
zrYt<s?+ZPM`@YbFv*hmzJ&60h&~vDEyS1F)!pm-}QqeLZR&}`IS)0%0ovA;xd{Q38
z=GRjj<yUkI%HQI%?Bgqop4t!AX0R5=3YB8{<C5i4gi0wqCKVwpqwZ2FKVCD`ZiZAi
zwhy`N<IS7po+F3ne8v_LY1rsdwAJc|2TiP3;GtXcR!%=$>KWGxd)e%DMyNX=5+^}3
z-W*b(IEBVhNGS>p=hZb1DaG+PsZy%e_qNOyEc1h4z(HV4inqzzE(>RDh@F6|XlYQK
z28Z%rNQ1-EpgE7zp*WqUr$f_pXwB1o7*ij{nx8=cV;Vr?7{)Y)#-|z6(=?7p;&?Qa
z@1KCg2{cYb;zSysM&i>nK7+((XnYol&(b&viIZrYjKs+_PC?=n8mA(0Dvi^SIE}{X
zNSsdNJS53O`uzN_AjuUPmm*0ijmwavjK)>Sq6*pa{U0EU2Q+?*L~oHQPyd8OpJ@CQ
ziN5lf)I(Kdzm@)Zo$dryo!?i1aT&pwWY{bh4b7tIDsU6jZlZ1AB^>pVt_n9YMvV-W
zcKRU~KiWpZ(eQB8m&CDL3>qFo<4iOr6HzHqCK{KCocLvCqKTQvnZ{ExQD12R`+ek`
z&1avu2rD%66H>WhqHuJgkjh>Yg^r2BzPOa-5`|8Q!ioI&i9+W@Ayw5*6i!PNPN(TJ
z6NP;RtknAvS7R&-)?puW8Hp|<s!mmmM8!NNu_F#LdM;eZj%2>rkvVAJo3sCUr7CA-
z!>#eMHy+7!*I(s=z#!t8yX&uV2{g_Ha=A1v2Xf^!egWiO&^Q>%1w(4Qa3~i}<18qb
zMdM<KilIFD{$)F@0ZslL2e6C}3)#a$6@D{<g`!|#Us{)2Ry9Ax_U)7ZT$BaTX;Q0q
z_lw8wwvf>(<SWL+%&->p+^G>QDAM;bv$`-j{)b0G@$fs#UR=xEkA)GAl){k#FN_$Z
z6vJcU;qjr?b2mpmh3_pO>kHM%TyXvb6+)nd551l+(odK^PeYOcT*2(Qy5k4Bg9AFc
z%KZ_Ve&o;J(u%)t-fVeS@Wb=)M$?h{aUYyV<27|bwhL0;B!Xchc<*xKL++gqV_G@P
z`2VjJkuVg`x#voepEJ&p>cL_@4n2%ShtV=3#d146k6F{zkN@W?blY{U>WAMWC0K>b
zk$Imjkh@T<K;0Liq=@!+^;TCy&5~U>y*>am4gf>4`z)6SrSqT)37Wc7S?MCELfjFN
z10JvE-joQ6+;ZV4tLK`O5O4&6Tn_RczUGPl+=NO4fInRW6l=gRG9=5@0mV8RS2Ck2
znZDY?-FFeQR)_5K`=NT&58Ll@Yqr6fAMQH2S;P5O&%G@_=hHy{G<`Le25i!R9Z~J>
zyH%U?@3DP4aOCNkz$TNX*8$hM@AIvZZ=ns*B7Dj*&N1KTTjD~7x(cgjeyZp5tw+}7
z<K8E$g8y9WNHm9cz2|nDm*5qsd<Cv1dh}w_hv^|}LObxTSHcOEkV-A9;Djm~*Fc9F
zIDv@fO~l1~lXNYe|9q)$S|06tQv4|KUoR)W6+Lg=6ENNKf6nmU70KysR(lonLp$j3
zEb&t5EL8ZPzu4jHoTGCiqy1kmbiMCPOCHp+q5Fu>MUq@Z#cw4@Qo`fT&V+oE`^VpX
zt0S6vGgM$0&6q~hxB*!-pgz6Cf0wa-Nq0oC#eLb9@a~tP-H)ChZSo=z=<~0`_*cEY
zSI8X;iMg;UU}7<AtgSm#XSkuB{xBm0fpH&)pf9Br2z2@18t||9?{CS^aPlG$=<u)q
E2j8(Ny#N3J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_v3r2.hpp.A91ECFDCE099EED9.idx b/.cache/clangd/index/gridwise_gemm_xdlops_v3r2.hpp.A91ECFDCE099EED9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..68e81bed95e0433fd1fb4fcb5128e6a64c8ec45b
GIT binary patch
literal 10860
zcmchdc|2Fk|HnTwXHH26ANpAGDIaS(L?lrpNnEO{(sH$F)84mRTJP;vC`k(~TKLFK
z<W?$KOYw2FS}JK*-&Wm7^`(&Z-?)sKe&0X(<9;7`Jec!*%<G&p^Pcy-=RI?b^zrhV
zZY7Znn(niFe$e76NFtGF<KN)Li)L@*B@$pHkt|nU8Z&>vV`u4QzoPY@(th>O_fXHA
zHM&FjB&{X;Zj5Yr6&Iw#Y5R1!qd9+B)|XF8=bo44Nr!)azN_rA>yyq`HqA78T9cCS
z{%qa+diCA?f{$kf_8m}mZ14N(XOj!>!^%Ng2UnjoNU!dpCl5Lvbnf-WZTB`MpDDT*
zXsoq8M}G0)pfzgKH5(UvD)YbOqJM1P`2AakdA|#}=JNQqQ^dH3`C57pCq;oLUF^1l
z64}ptxKbfaQu5OZ3-N>Qk^8$&Kghi}$nDpK-L1m^s2Qntsk^v+hC}hi=V{A5&5y1d
zK4E%f&<Oid+1pQvGe$Yr1m4wsyzo)d(3N`<_7523Bx|^J?DZh$9rBNyme1V!H(|wA
zqQy1mL`(Qv?!d~~-qOrrPD#Sn1s5NrceS}5Y&R>YpT&gS!CGg%Ja2a%w$yrRU8Cm?
z--oVox-T5k4Z1mPv5Y=5qp~>E*8JoBLeGcZ8cp(I-=-JmpA;ER8yg?l(N{CyEOPe-
z`{P{|y?Xz<cH?TT!6|}mr^B~PoZ{~1_k#K!mRSv{qu*!Tnp!`8&zSNBRl}A#Ry(zO
zm)Upksxz~^ZYDoCF{{8S<$*z@=jHn`tETPvNw>V``S7hOyWp&jPI7&p1?#_<dw~4Z
zHM=fNUNK4c)mW=)(=o8f<|oew)1_;Y%ZBVN==Hw1(<9T>4+aN5_0;QeEM|G!^eJA0
z9Akr*U3_ytV?^m|w}Xl=XZ)Y69_8?PLnqIL%eE#h{5Pzw{aDgRUUldOe4aDT{k6-M
zLiuh#&Hn@%B#rO!uMaAdiS8Fv`u~nvSX=Wx_;g;67ydi-#(d%=q20~4oVzY@PkY+I
z_tqWz%5ff=mWc{q>5$M(m1nF9^4E5;yy)w2(A0UdclvLN<&%%4{5oRysEQ#I^@n{d
zt+cYQKJjSOkVRETrlr1}bL#Y(p4vAz<{N#9Dp}LBK0oR0{$oAstK0d(mnW9oFt6?!
zekdj8@Z5=$`kSAXv==x3P&{U1;Mm&PH8<Tf_P*=oHvi#cvk~V;w=0-BLE`#(UVisL
z>+88GJ7-QFJT)L_>mC2q>%E3t7?nGze00V5a^2+_J^tg=ea=R67ngm<Y&`sq41Y2q
zGc|HizX_R@p8_(i)pTB4@2Tu>yQiq|!SA^*H%;GpW7deuu`41joo!S*kXaEIoi*=J
zN`BP&15KtOF{hFom&~7eJ^J#r#g2aN4kP-9&K^3m<D%H46i1t3b%Q-(Pr3RIKAbjA
z$7;Y;hjKqL_U2jtU*h)ZZ@u)vJG4(kz$x9Z@ka0VcQU?Jch$kPw_$bp7iGrH*Eg<B
zt9Y>}cWK72bDsr2^G=vu==AfIq#WziH}Buf&cEaD=WuBfw<Oi5tiaZPsP&{hq2uj~
zWgkOMPS<<vJ!5OYx*uO(NOiT`Gi_<zcBQ?`8#Ml7;^KJE<R|Ax46PcsuSDPPsiW0j
zngPH5n9$|!lm*2a6$NEM=d7)lMQqsIbKMRH$LJ9*pSzg8c%im5pfEP*MdB|p%O=Jo
zoU9laI&a|NcT)q*UZ?HJvroGCtLZok9;`lnf6}wSgw_2OwV%%bzI~XmV@2SenquXb
zDc#?6400Rwmsyb8n5M3JKS$--j$J(CHw)*P!}l)L)J0}vtZr8R{?msd^Us;5wY?WK
zZS6Gd+S1#DDz<pYM?U+VZ{oEjzg6XrSU*Rv5e|J6yX?J@w&G;$lkxwWWp=4zuJCTh
zgPT9vUVaeuS6y`1=y?sh!Q{swT_g9tO*m!j9Cd4(+Lwf*mlyR_YMRfNuN~*Q@T`-S
z&26tN-E*$0@p{V)6RMpB_t5=Er|nD1zjxtLXqCb&TD!B-A$qz~lFIm~SG&F`&t@e@
zx=O;|+z4K}>+W=A#NLiN?=`w#S##r^i%WH&;pAuaJ^I<a+mLrFd%U;F+S9vNHh39p
zeORe<cj-HRvh~=pnzxU~e357l)SVN&V*b>*oJ7*e+vl&%;Z0fXpq7zDTm#y#0k*A$
zqyq<V3<n~R9aG!6sp^%B4wNe-;x&k_!S=0%L=8*uk72`zA46(fY&t?sEQto{(ZH~^
zAd+sumsk0By1GRNcEpl{!1f>rXf24u<yt@exjr8j3X&eUHD+pNiogFzIv^;q>55z8
zWv!d4RyXmyE|lVDgaX4*(51B?&+viMV5h6&{dAxqmgGU(Jm}q85J{k1db!=3kn0^`
zJ1jX1)z89!)`Cc;!hlL0uK#EqsD>qV0M!BQ)`Cb@{ahi4b%%0wp$?X;1%_*JTUv@X
z$u5tG{62o$Izk>xilBNC9M@V9$p*(>+v?ubztM$ySP}sYBk)sN3L=S_;qqQFx?w>_
z*a1r}L-os0(OMA6?vYnLP38}Y(t+w&vK`oN$F^-Lh-81P!*QeYAANM83`@d+VK{zK
zOF<-mfMVs!+`?(PP#a6afMFPxv=l^=YQ8z&WA1q`U8s*G8-U>kENLm)BvJdF4t|&(
zqXXMxNi4LDg+H_wL~?9x@9X`l^Q?5B29}fq`*JX}wIGsH{x+Ag6MAUtKmkh%04e}u
zTMHt|OSN?y{L*o{F6@LQ>ww`p+?JMtNUkhcn6RNk*lb;>izVxU;d(4-DTw5G#=8M`
zT)OG#LIW(>1PnJ}NlQT_ugZd?NqrXo(h+vSk}W`e3+UWh5Xrm6ar6Gvs`1u=ov~y;
zMEl{`)`CbRi-KkZ#~Vu|?M^uwyAD2@2L&+=3Te=i+$@NPkjEi3k<2#@yNK91G4^+j
z1nlFAP+&paU%;cH2m=;mzy$1finYLkxTJv1pjZbih=l|^(28(iL3}{KHc+ew7Q{RP
z_BcfZuprJT;2~9P02XAl1#A$-CSXBqC5TF(Py!3$cFpXl1TK{0fg+x*PXvm@uh}t?
z)u#hRI<O!iLJ-q|V>)o5oB<RWZ2bwKIPo<*o?!JwKv4uNh!Y87F;En<TnZGWEMEeO
zODtzV%M55k2POl$Wk65LS<o_zt<QpPS#13=WN-`_k%1J%Ok|MBat?2w!#k~21=yUN
zILAjL>kn0T{^vY30Z&@A9SDENRQyN40TLZ#Sp|@a<#K?^S*``BmgQK8Vj=UqM2Heu
z&VvX$B6**d!F3P2Yxn0uBvu2*YS4$wR3yIQdc5LnNRP<60n}Spi?qGgs-WDh*+hUV
zS{4eTG`MwK?PE$#ul7ic2Tt+8nK%m)Gk{YDa3+J@{A&MdFp9pq26U?dmXwvyUI}}p
zw#FGGhH-9T+yL@sB!+Wt;gmaTxfZz}C^!cMu^xKXLw_P9YZg*(U8zV@vAaWI6YN43
z0YTggwe~^<d7~h{<<#DC2IN6Te<mH8b64L6zo8NIYy|ztjf!f^><&vG4!{)(i9(%N
z*>phW$=A+h>fhZI#KfF>>%vQWd<)p#0s+Lvf*8fwMsWe8hc>-7W`*de8RL#@0bRBL
zYZ4M|rhjzl3V$o?fW#1}6#}(M*PGcf1UgX;h0;){!?H^#>`yriO2gRta3~LlU1@zd
zv=4_4Ec=ARA(YejA!+<DGNtVQ+Iv(!`P>2?iILDO5}H$?5}GNYIT=hO9)v~*p@Ph9
zBp!lBhgfdld>gnS#1y@LsElmaIiL^1eg?I|p_D#299o3KZY;Zovpr1XEz<aIq=k+z
z`rMgUR688o)C8I666bKXm@1i~6}SXX!7Sc7i|;}7&Fq%N4<PL9vCAXly@}TUyoY@q
z!bjdebGQMoPe+mLDC$YNE{GXOmO<HNLB!1D>WGEy1@Q^&{RDba;ZxZADa+5G?K9{`
z>uX@o8rU{v1@Sr8^EsWeg7}i_`I6;i<dBR8k^4!Wh=WXXl1O5Yt7iYRfnMe-uthHb
z=>_0G;=g-nr*>71i6XA32D;V2fcW{q$dZ1q9t6k#-2>w%(1E-`5T8Qhrz}5%hR<MI
z{1e0)D6fIGRR5flKj#>J$;n@`oQxF7s29<Dj+7^@Y*h1<HU}k#df3)AoC9#uL!2=Y
z3gQt?dW7X7PFlorDJLyuxsJ21;|9|EUvu`aS-y#6H<3B5e~4rcS#CnICYIxP?Ks|q
z>f?Fsc$O1*?F5$h^V0o1)Ba>$n#^(vFHNCLQW^vG);ykuq%WY)7vM~$lYj$Cj}T}}
zITYH5(tsd{2cZ1{=s@PNAlAY@wa}S1{5>>%51EdKas9$LS6aUZ`R_raDIY;rM^F#S
z>BuUb<svko2r-5(LW7FXV5%=d-bKiV<sn6A808Y=Q-a#!iXg`G_VGO9feE~Q0?R7C
zzlvw(jEWzq;u*(L@q<;oH+{Z}_fhf8^HuyX74OH^k5KV#tD*NbrAyqn!NKz8WKO08
zMr3`)4Fj@xmJ@+2k>zwCO9w_|alugn$TC<y0c0mwz6fL&ff1RUIN5{JbZAHiE*<tt
zhpv<}VXsV<v!OJb)#pHI4$DnYs|j|d9Zf~jRAfl`I5IenOetq0gKU;_kTi$Y=OJkx
z%VkJb#wK%RsCOA+g4<OjyUNy=BUw4i6-ZX`HTSMy@4t&=ciH-TNOq596>p&8+uBby
zyimP$n@eZV6=-|~_9gpAAFqWb0lQ&3*BovKKIfQl{gNB_lI3JHI2rvwEHun%+X6kC
zP0{THF@_rz!v&HGyg+chkFs$(lN*)E)?ea!Ut;U4(AX+8mEOOGcV0t%c$lA){>8z&
zN8lFj<c93zex$-IZb%l(_2|cXG=kKRnSb1AsAl3=9N;UFw35czpE_NzpM2E70<UFp
zKtB#RP_I4?^p67v;<Z10*lqvUsj%MP%zP)f#&t^MtN!&rUv|l<1RlqhBMCZ=6;QJR
zwrvY1KaQB$?t$@r+}YLUs~w1p1V+3lwwzbARjjh<P_d&@8|#gcu>namW@_yUj;^=p
zj!m-<=<j29V>v)Shwi546{d)B@eSj@nUTy-+gUX8p>JpWq5`N@0BwnDuDB_g*Tc~I
z4Xy~~ZA1CCaU&bncIvGsC#20-2q%QXNu*PF#|!Pl*k~Vz{NvbY{Smo;L>|N;1n~=U
z|3Z0H>q!hYZp$$YebGIrc@It^H-_epw2r)&vkEuWq^F4oF&5bw<NI3<XXyP2$UlKL
zWTVBv;ZYUSx8VD~PBk|Aa%*2keR0+n#6qA`2s+dFcLwO3VYwXO!w@@?qX_u$#3vRG
zCkZQ#&rnkdbt$JpomAMF#>*p6=LpOB&^8}BlEDzfd^j*4{y;gD(+uTwi3RZC0jIf+
z<pxf(f#uCeZ!_vj^;?kM7M4qpP6=Y-Zz<9#WjUQUO6M&|9=q|v>M3Cd(;KkGZgawI
z8Z9<H_q;i4#-fk^IMd$zMCoU<Iv@zYVYTCGCo*(fHtO`+yMIA8etr#dt)USZi7$}r
z3zlCY*H<jpBG+2VQFm52Ye+I*Hv5d8i4F0Yt>sm<A*X&l`ptPm0yK86meW1bLP5L=
z-LJy7oMh*W!e?%sO*uSeb9q@VT>}L194|XZnM66f(Q7#cq))s7Y;Uj|y-l3$Cc4p6
z##=vLS-8aMn_+NY=6a!O_O(B7pBsT*BWTOZ5?5`ED%Erv{LNmI!zkRwmSafT#(3l(
zPp6b1en#$}>0BnKUNZI8!%5~wFfbkrBE4_UL<VMnK_tct;zgji2u$eAD*=iUmOp_$
zpV*Z1GwAb~<q$YB1dgIXDFxc6uqkIa*Dst+DPxg;EZdP%<WtI~d|&&W@c;Zy5Wn_2
zoYH>fcY^pHxxc4|5yTJ3{R7MK{E&E_NjVevAqkWZjxBdn&zWn3no}@&I<%)DBooRr
zSx!T;G-O4md=`>rQ9gXF-JS5V4#C(NwgT;~z?K}XkOM^Q*)4~}bf|tu{=ZYxq%J?j
z51*gt)m0F)p*$O!(WYcWt8Cbdat@T|u=P36E(bc(`dyseE{;k26F9pBma{p#Y?kvm
zyL^^&kUR&O(XO6A@)IndMDmj?=OTG7%cqe16w9ZP{4~pXNS?=XK9c9NtU_`X%LPbY
z!1A9+{wK?YNM6YDRiwCzENT18k)oXCyGU`D<$FkRkL7CQP>o#Z{dLHpj^%J(8qT+!
zeF@@bUb>m(9lUf0WwJ`)tfb{r0P%pM_mpRT?_Dq#$FeI(dIj0hH{L<gJCw;fh@a5%
zrGPvktLB$1{m9pczWXuSi5U+|J=KdJb!a}W)z}X7NyKkHuGKimvI=OZSS|+|<t*0%
zjarstp++oZo|gzU5?RiJ8hI?2LsSkm$@@>anrK%>G;YQ=J;)0Oc|B^=R9>3Ox9x#Q
z1jMno<<knWDL%;o!V?fkj4p`#k#s+rMjmwf;*EeKE^nTGcbk(>8Thca<r5EW@keO-
zk=@rTk-w5g7D4=k+&|HvSZKM*-YL^J`J3z0Ijif_YR+ciqhRp}k{v-Ju>p>jNS03d
zLiyE_lIQgW-&`|`E4SSr)a272KXE%JZ0DGRHM}WHr(E20`g+rV7d>!1{W^)f;@WNH
zNWU&Szq|R#mm(a<TRxvr^Y2Ced#Sn0#?CQlS{kpwO?O0&ZJv4c=9012#^+b#r6wA5
zjRw|ag}|8$R67Ut$nFP!s)1_7P>&?+I66SJD^QQ*NVlFwj5%;k1bB;h59r(j-KnQO
z06Gs?uHtm6xb9@sZr3YE&e|Qh{+pw8zxyWt(oy2<|1$m~=PQHWrW%0nZhZ3D4JSM;
zA10~sV}NB05NX1r1infzlyW@i8_(|S;z9p-Fo4!40^dYde;8OE1|m(E(t&R}7)m(<
z^v$3<!{!95e+C#p>rVjR6RiFYu)G6AI!2Aavk`dF1S%9dhr+h+CC$4$jWDQ1>y^+!
z2_4BIg1<|0mK!)`Chp>f@8U+1F~-RQH$0o=$DI3Pu5DLX^Fx#0?%8`9u@^ppUQggq
zdgD{*^_1mj(ES<oqV+X!U=3_LTqB=9a5QT9985=KJ7>3@?qP7O;Ox>_4&$X^yn;OV
zweqB@XPSxmclW8+&mToTJ@+W(yPtf&kygJ5+u+yxAMfhw3pKheH2hEWf6ddz-_Ebn
z`|zjMKYl}FxX%gm;`E09`BBf?564pVpU-bzr^}F{3^9kk6-ZG*nPjzioo@M`0_tby
zq3(G&i-ye-s9OSOk+AvU!vR&#dpqm@K1s9N$uY;!S)5%K%k{{i9<|K`{9g+5rZndh
z!HcE^M>SuVH2-P&u_PA~i7~yz|M*FM%*lmB(v4nP&`YznODL|H<r-$vcf!h!*2X5L
Wmfg+FEv&jpy0-}=*7VYpUj82{<7`F%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_gemm_xdlops_v3r3.hpp.E0F0A6CCA772EDE7.idx b/.cache/clangd/index/gridwise_gemm_xdlops_v3r3.hpp.E0F0A6CCA772EDE7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..acd23df4807572896afb04d823abdbe7257abadd
GIT binary patch
literal 11116
zcmc&)dt6Rg_kZ?2`&4qOr*wblsg(4jlu98KiiD6E!;o={4C5u&VKRtPsaz_TT$)sf
zq$H+0x{4_wrlu$vl8_Q=^zNp1d;7hA_~U&)^Xv1`K40s5_HC`b*4pbd)78agiJ3?=
zeu?X*RZG_`L?V$$760>Dx7PaxFA_b`7l}4S-J1RF>RNmC+wMiz9ZJrX?iMTZdu8F5
zDxC)`x1IIJPbtBR^$qt7EQo#?{51PQzdviM^at$vYQeZIw>LJ{)%kolxAm_Uj}yNi
z-?adEEK_!B5C4nnY&dhmS2?cnR)OBSiytG$>!%;jTNE?IWpIRw+gj@g&nbUa!*su_
zKOE(xTc;SL?AiEYSK^B55lfVg>8@WnvREr4E9Ub0og8Xv{|+>@YFpk0);dX3K;(#=
z5>elz;P8H~4=*fl%{p~IVa5ElP4{yx-aQB!@Tj)WI#a(CuY>9{Qgm!f620F9cyIk`
z+?su+5zcvoJ*wJ*j;vXDew+EkfT4ZOme#svg?ti@S-OV$W-0iu-MQ_N!Hjy-)FBC{
z!sZp|jh^w0mrOaaL4Ewqb=Q7g_)%%gx0m}WR*bfsd)U*%;*VPS6$@pJyQgk23*7Xs
zRzECfvyIt<Wx?9fc9+e<V!WOxI?2!U{b8Zu>pHoPsX3G1H{OUWw>1d)(Oh}(s@DVL
z4;}WKe)FCF^$Lma>sQf*2Xl=s25$V|#iv~ezP4+yr|J1+=dYd6I1rsS?7Ux%w#~zP
zel?cOUh5n>@{@GtX!+Lt0&4pEE;fQAw#<INcg>=&6tb5~Ts#*(U7_{EaOH><{+&kS
z!0}l*H(g#&Z<@X9ZFu53RkQ7v%)gmXb4vB+*iyGq+n2V4kGf`g;Z?0?y`$!k3kNqn
zSn`9*RNDrhjStq>SpTb}`O29S%3ASp*H3$moL(8)(PTKO?$-}@9)FBn&z-&AYI1(i
zyzF~^<=;-VzdB>WC#4dTqgo0r<yJZs_i8M^ajZ<&s&@|Pe9_?Jpr`r<MtFsMlozSX
z2mc;$X3Wpu-5hxE_RYnGseb*ZI$sOdpR^;mK2MgD<(J_Qc-Q!ek<-`Bb2SP?UmI1n
zZSVW;$B{w1T#d``tuMQAb4tsNQ*~9d_g#uk54qTsQ2#o?cTds}E$csQn<wWG-DZ@j
z_0#BRcb)OI%L~3bHK?LII>Du3MFD?i(1)_<1n<nd_lLXN47u_1d()wlhn>^RPJA@c
zd+>@kr+2vbJ@d_nv-ZUz{d_*BKJ!sfb855a^E35DGxq7v(mXu6bz}9QJHx_Z_`2}P
z2iNT?2pICp&a_3&`Q^(Z&#2Oh>S^xhGi=tDCCiEzXgRwbzv7&}Lw~c|q(PgT-Mo@t
z+j%O#3|?~Q_vJHAC0q@<^~!JHAvKQ`<L>V-KmIXmwo~BHEcquxy#G>LzU;2}_Tt8|
zDvs8##_sSImi0|M;TLav)BTOJ?}^o$4?di|Fg424qRcvZbJ5ZJ4xVRkmF~UL{@1`f
zNlEDsd5<Etd|KZA;Ns|xU~h5jkI%P#LZ@Rp<J;37F0Fj&HMnAUQ&VKqp~+F(Hra%|
zKebZU@>j!K@7ih)caL|TTuP!vsbhKQ;uaXEVe6iIe$r;ePdV!xeBKP~_ok#MbV_cD
z%}Cd*l>Q@K(^GbyFsgcZ#N_llt=Jbwzy38#Berhsj497fmOd`ex)ZPN<6m~%=j+Pk
zhey}$w6k3Ijn(er0bV=2TrKKbw3lo>Vz*sk)TkZvBa*yy(ogHv&!1TTed30&MQ#pj
z-`gx${4wcprq#K^>m!cT@L=_vCeJ??gl?O!^4f2ZuSi+W_Tc6PhhG$LJz4UhX8cm$
z=#q)WThG0=9`Z<U<4^O;j<nk4?m5zYDTT{zSQNV>PM&`mU;e(lc|!lp$8OVXr&g+J
zeV;CxYnh~S#NE@9FWq_j+wbIc4HMnrm8Y>O1$VE9cx@@{V}33=Ge1e|b@|+)+|5oa
zEH*xr>k}U6trczY;PL4N9arD}^|o=Y+P+Sogp~dvrv^-`{1~m7X8C4sqUOq&bdAW(
zpH|I@^4>ZrtW>^_#lKei9cZuL(tr1?PVJE3(`%+p+`sU^rycKy4=Wt8F~sXq0Qa5I
ziKWYG)*ijoZ=a8E^HTsqMGYCZr!@_=P4a0PGR*wLo{bl#riIK?_^`COOtO7m?VAno
z{E}7M&-b4{LR>lX6W6-K^pMZyRf|G6kw|?)!HWx>!U;9lM_(lS2+&8M+S7;>u+PUo
zx-}Uis-6c%t3hQ<sRUM)U`kIzD0T%4?T)XwIHE7)F{KPDmch9_4WWEvT=c?og~}%l
zsER2;KsN~V?`a5SNy^r?H5=MYG@v%7><7C0F{Qg96z|iwX8(G!FhUI~VM-A|MPPPM
zLnz-TE3JfOC$98`DwuK@C>{p7Jq@A!SoSh-&6S+}YES`FVxUC~oYd10%J#8KL$3T7
zd0Gwj$CO79J%aK*4WanYSf?o;p*^uL9Dpg2Krs^N^fZJL@NO2$yQ-zC0W~ls6zGQH
z6?Hd+vS*omL2;%*wj2~oMY26Weh;wdX+(WEfcGEBK!lU4i_`P@M>SwS+|OR1yBGJ<
z-H?7F4RZN;IeLdRpeCk-0o^c6>27){sSh@7s>|<GgAz>1hZgy8VoyUTC+kg*)%wg`
z-51JX%6+JKA1kuEA(WU(`|yV62V?p|c}yvRiY0JPPeUlNk>d-q+J^4YfLfTc59sd0
zbLnmfCCN;$BdgXaQv<4FN(j&m!IbW%mok2-Zuqu*9Sx|1Dd9jj98<a*Lb)_wS5D1y
z1~$#Um=XmnqQIP<hEUR)vhyBk=^a#q0;VJbi)2jcZU`kmeyqvp;#YbaPz_Upfo?D!
zsJkJQd)J;0*4=yjFEyx(Dd!+M2WR&*gz|@8-PP6?CbtEVIaZ{Rf|2y^U!pz;iY)Z7
z5}&UAHu|FS#{GSvBJQISDpo>iPeb}xyVT2PvVlk>mu{=?;GCEV1=&?7T!kiNd_k6h
z#$}-S#5|QO1CFc+x2n_=@L7@u0b_E^1iS!gFfb--6R^2R_X1;LNdn$%X$UYTdJwQ#
zN<)D$u~-4yjx-Dy6SESq{-pbWF;T6652kcKFeYbKz?zYU<NkV-CdgueG!_^W4RvAL
zSTKs>I3SH<<q1HV@Hw_kVEO4lnhuPKIt5ueuuTV}D9!-V3|5{Cq`9AC+gz4^2T1P#
zW8zQ*SvioFv-lp6-ed87AidAx3}})8%^AhOF&S_I#aYlKi<M`=F<GoU2T5~~F;S!-
z%SY0D7XQXu{l?qru3KepYMXDPlzKXQF#oSi1p(VPiUPvlt{MLlurZ@#78e3k$l^+X
zDp~vipbspLfhYzt{Utz@z~X#}@+nr)zB|scy!v~HWX-^~84M@k0m*(wLw-i)WJM&h
z0L7koLOh*z#YKgV__Sg2NEQd|;($G|2_(w^b{W8)=(KADk7h87j?e-IwEz=}W1&?n
z97fzZl0|Z3BDt|-21s^<8*_wW9TkTn$GDt4Ajkr^2?3l3A(9U9RJ+3)nW|pysk198
zwaoDwonUAu7)cf-eQt8O&xS{1u|Qv>uSk?(IQG)jX8Tgbe|%}?!-7(`t10N?`9y;L
zk-&@uC-Ws8qXxjAuJu8(K&TQ3Rmlpvux%i;qc{kPgP<CVM+L#r6bD0bFe?v*l2AB+
z%0r=5D70p=YbbQ1xR7@%<fjsYwt9LzQSx2hY97fBz<~##5hcdLfw9ntYBd?^B||AW
zSxA-w^-@^uk0$#gH=@j8lb!_1>C739kSr9cghDaxI20O(!a*!{2xV7T$Qu{(gUAeR
z8;3t!S@e23-cv(ls6lj0aviAHli?6!5o8TexdD2TiP#%25-n6&upXb3TfEsVehBHf
z3y-<Qk0scCT;#Zvj|M9L^Q{<Ry?bW$^@LU&MzWD48x1E@6J&Wvl1K5V)d9<5p9ieL
z-sLT{c?%~{Vk@+1WpNv{XoF*@yd4g0hrLG*pA$H=lf@m}&<=V8v78&)$>KC*orcDf
z{v4D?-&o*cv>EStDG-+eYvP|Bcl48c_U;!`ENBKA%|M%+S*O68Bb#b`j{QA#{kPDX
zbS21Ip?)ii+n{b6?DbEAtQ|_)p#|l4LP;l!J2*)P$I3ZLCyUdNGz|?S?@yR1Ikokj
z!UWYW?_wRm^{z>fWpd(7PM;73Sq>-8VR0oVu4HjFC$47kPGq$cInnq1kd+^cpCk3>
z$cW0{BK5Z{4&v2=ct(DaSG~v^P<{%pn!@5#UNx1)X}oG0i_>{=I?t>plNV>QIGYz|
zQ#`ReYShuW&OVZ^96&u5=n;P}VE?Zk$KnK_p1|UCpq>u&NIbyF4^YowaV}8LWpNo$
zF9UkSW^j%G#pzI&c#bYSEFC&foC$|zvN#)xvsr!)6z8xwm{SSnbm+!qA#oPcrT99M
zUPq=B7a?g8i;I!CnB|uuaVd-Ika```qwm)tn>xfCwkJsa2`jHh>h&ysiPT?yj%{AD
z?>8a!CRY9mslQ@z1+QJf8<O!|UDg=PIRukALDs=Jb#RPV=;WL_S)7KP)6gXH;#9l+
zt2NDo!{h|n8E)1YZUGSpvRuw4mp%Rk+^hmtUc=ecu<~{^yB#f}?;qgp5AeN5YpT1Q
zR-yBu8F+?gIk&UiG)gSw+zMIj&rkE`XAt?<t1jA2QT}B%b|-B}+(u9FH~k8%zDv|L
z#xC;|&^iUIsj>bFw0;HF#287Ut)tj;i=|;801gj;_9V;*IM5CWgccMBL8~AdC<WOC
zXmtTvQ>*mjhWl~$WU4q*=L|zQ=J-Z(BO^HnDo;Qj3Fs?|vyfF5V*KC@WOakZRcLG#
z>prT`_$uT~->*UwtB@;;-Kx-3imQ=pHI27|EQPm9;hFJLd8<?wr}0*4EUw^3SMW^8
zsNkI{s7?i01@BzJPo(2j@U9hnZ+!2PPp#nHS^10#zBifsru~rB)AXG-UrbQvIUMcx
z3;*ng|9RN`t~OwqT#bG;Hm(3*6acJ=X)N+J@ALb?C#qOrfDE)rMnUdSIHTyUb`U<!
z?U6?^RqZmBNT0A4<H6WC&jPKp?CmHAXywq`4e@}SVoz~|_Sgu7MzDZHPeGQ7#Hna8
z*-TPAL5e*k6Z&E}6vAPDYQ^zTB_2vimL|ydAcZ|hn{>4DTD<AeW3AtQG5J}Q6j8M7
z(PSOGCq+=D2wFt;R4GW$I4SS0(2%K+A+VskE<*%6ew}F+Se53ByD}VVIF4*DNrmy6
zx{EY~FUaCK%Xrq?<Z+gHERNvC5xkTPNbbgXSG#LCw4)2i;{w$azAW>O`}lF>MZ&`~
z?k?~o9XHEf`s=>SLB>wgLw{cqf@l1BRueRt^LrWEX^)eMn?UU*&><QXWF<hYgvFJ>
zv=UfSg~mYF7&x8m9*((CITmVAoC?)ap$_#nm!aBa78gK^0%%LLAjk@!QvsYraRjFv
z!D*1C39`eS@?jSHBV~WY$WI{66KDYCpG2A`SzL|Osu2@YYmizEi;H=^V&0hK%0UI&
z7VbQg+=_RsffE|2vVs~W++Xgsw&TA}`}fvGxtnjBvlNdI0Bi$*9noG$klL{0F{>}*
z5k4V@Pc*I~StoMnWbt<1VLRU&w~=fo@352N(1)Atl|+|bcG;<>p*gYBeRrO{-k!cC
z@ry@>)Nte!duo7WhPVg>=LJDedPswzRWKWB&LfZWteuAPj-k{}aZ!hN+)r_Y*R?;#
z=ooU?-`wHV@6gOZkp01{|3NXS5@8S4U2>!|N<tn<bVOoCnAlx^q{J$yUj;{yB3aZi
zv)ZlK*4us2i<62a?y<X0NqcOB%8m3o7JbQIc;;Px8t-~1(Ch@g0r?l-pwJR!JLfO%
zJE?r)p}NbSbf{-g`5El3%8?=&CU#dnDe)rmxJb9-%<Ma36w|&pM_s9@Bpq7OW0?sh
znJm7BB-fA?Jw}B{Qb=(^x!l9urG0$x;*J2-Bfx?<@^hQg`RS{EaT$QGfKZYR2htI;
zA^xvn6z4!m4lB=rmO0R#$`d%t1U4;6;Ve^FT*O%xvAB%0EMsvok`$wXbOkq&<R*(t
zkfemgw~*u(i*F;zZ5IEAB)_q^6iG^1T!tiNEWU#zcUWAGB;_o=izIhhd=E+PvG@s+
zK0zjQ{`E*&&*CN|ZDR2&Bz?u=4rJYdM$z~EcxykN!H0SAVZJwj3bK>D_#}(Z@ZvKR
zCssvWKWUTmJwBVYNL-67X~&I7+(<F0@Z#{-U0<ih7{(cfv593O@<^n{iA%t|V;J=Y
zq_nH9*i(h49UTWI$AKwLm1Du=STKd+I4~lP%~s;T=r}Nz$`iok1eSjZm|Oy;B-QS!
z^-fL)Qz*^=BQj{V({(Q#odL#Dc`lfo%kpc1Ni8s?F|rd(=maj*O$I^xAlO^n!?km$
z6bx0UJQiBVLR%U`k8vi)IOd+0z)er!W>O6lanp-f+`>7waJ}gOSvSrjx<6UajjRS1
z)nE>t=t<7vBsYh2ociL&RIR|~v;TN`k?bs<ardVaI^##c;t@OJ7|tSw&iIPgqml`V
z#gF@Rm93ScfEEdMU1e*fWEK|!r9u{00;Nh8e*j7!SR4bDVj$CB0#r(1aXwVaXK@Wg
zHBgz1pYC9w`Xu1pUVMCVc_EkABzr8#Zt&t8eD9?`qwwCG^P}F@|Kodp&BWKr?$YGg
zFNUQ&+1<r2Zlng|rDP#V7CrfRDM)gIV)89WL$T-U7S&=LaEb%tiD7mfD5nfCo~T`r
zl>uoPFrZ23T_C;7;@xogZuZLS4~P4+I1tVZgtO>@N`h8N?3MQjH}VL3r9Fo{&as}T
z2D#R-SI*B9FdWf8OTYx#=Lwkbe-beKF^qTIOFbLDV)Kq6EKcFwQh4Udo65VTvN(-*
zOQW~|H6|yxU!I1eZ#ZuePRk4hcgk+wY-q{(Vir^UWPkPe_pYO{4&yl?o*PQdF^?1S
zC@y}Vargb$#vyp9&u<LBISkr5)4hMhKW?X`vsY-p-*8NdXEXwh=sjR%_t(!h)6d_=
z;~N4Pa2S}8j6{&#f(o~wCdsC7ZV46cKuuCt$M<`vPysbbVkgL|ph6YYB$**eNN_ys
z{tip6_%_tI4VTl7%b`X&TuwWF05u-K<)q_^`hc(F%1r?-9@PW2dN7!}-Dg1U8H+!1
zY9G15L?;hgqGm2X6zKm&t&&f|_}<s`J(%43ss{&ql*}^q{tq2Kw*PH>+c|CUkL!@^
zEuLuicWJVxFDAWrKjeM*9oFJo=<*g$A$!nOU3Y0^aT|1OgDzCw4xQQ|Q^o0oPMs7t
zM_tVxnq*+~k1w^k;c;Mn-sAZH_wNO7#m^ge?s2#IuP>}E*OUgW(R~&6ALjJ#r+i<{
z_FR?!dW3hoU9T7wC%693D`jo%7tU!ltm=xwbx2yr7AapM=}U^++vBo_Rz<Y@UFfwu
z%Q4l-Le8>~#s0jtKi_MoYnQHE*cJ1A)-Lu5?Rqfm`qlk2m^_F?_-`)A5B|ef@?%UM
wL?RRVXha_e(MQwXM=1Mt`_zq`VTMiLAqIv62b&rh4>B_m4e2F_2GYlW1FIgm_W%F@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_multiblock_batchnorm_forward.hpp.4F73C2A5286FBB35.idx b/.cache/clangd/index/gridwise_multiblock_batchnorm_forward.hpp.4F73C2A5286FBB35.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d8f7cde403e5a263f0b08dc085de8134ebdc1cea
GIT binary patch
literal 12750
zcmeHNdsL0t`#<~gn$zi3F*W4vqNXO5<CLVkC?P}`5x(PYh7qGCw-HfjLb@oW)TE2d
z92JJhEhSw?jA=xwQAtjs!5}H$x66C>p7mSn=Z{&xS?jl!7Vkcv^*npO`|>=`{_JPJ
zGu<alxMnF5xh`|x=(GIW#aJW~4dDO#ee1h2n+I1lMWT%f7rkaozh|p)eR>H}Z(ZrX
z*j3f%bAxeq8h5Kx%tFkk$E+*QJnijp^6*ZnVw<zW`hgDKku{PhJB<_DKWm69dF9cz
z{OP8IV*fDH1@&z^qFzLZ9i=1PpIGhs<f!M7cIUma3hiydj?Cq-x>lzml_j4Sw-p_p
zkg#}rp61ki6N_2<hgMch$$o7ud6lR>`h2F%*{#``v)D({#)aPF(onAWV5!I}9Hn}E
z5$vsff9bYh_k}pBAbH<6`xfn*uA<rBQS(iC;en78y{i{DKNgK?cej7yzuu?#OVx9E
z&m5+y2A9PL6uez=-EwM$YN%1eeRfLTl=BxChS`OuP0?O)AxYy<y~$A5c^O>L$-=P%
z{iTMMyCW@DITn?uxP0lE&slQ@*O?NtSv7C(zTIsnQao_+-LWh-$fRLQc-7Z0?Val%
zZ$EbbSH|M3y2Jg?e-WpOxyKXzr)<*rY5vQ_r4sGD#j>%w?$<{)DE!ac<oM2GGZbeN
z!@k&?I0zlH-}-9Ku7k0YJ7<6KKym8IFII~-J<@q<vf<+5+t+*&?pej`bW3#k{^)AA
z=Euv@a>rP$ymsQY-&a!~dhU-|SCt;8(09|TQ=L5f^~%bMjqzJI-<Y41_UFW?Tu=8$
zwg<m`P_3S$w#~<N@V%SUY+s#zijwSGvaaAY?1seqKUmrpo_R4R{=%b^<z8Hk>B)rI
z)$?*r;L65&kErD-O%gq~`SmMzKgsqj9PSzx6*~9r{5N~OEvGuntu8QK>TKe*dgh<C
z$5veU`q`~y$<o*%Rk!6Dp(@<|g1bwf?I@bzaLs>$#+t`TcL&USS#i3v$a><2FI{Sv
z9Dbwi?Uj_%vS_dMf`H2C)7wY?e)ddEWo@AISF2Lyow$;0n&`9Z*`Iaa>TJ2Qe5dW!
zoL^pANOHV&%(Ck4YWK_QY;j>~SIuhnTO(6BF7Nj;sI=qNoS+%D7c}?g_iq_nlicOD
z=(iPVSF{${j5qFjqc^{F>*b%cwq$)-FmvI#wZ_k)p9fyppAQ!6@07Xc)>%$3&oNy)
z@!E!dHNX38a?j0C_x}Fn8u7!~Pjz?wSmdv{-tDfPpW<P{=39RxocGxfHorYL>&3jM
zgSX_jTvL2=FS{%?03_|m1*7l2F-=d7n71@OB{O&Qnas@f-<)2u-jf@ytM!-1zBSel
zFZqss_@7_CU3154ib=~a_M3DLMBnWH?soRgvwj0VDu6>4Jo)Hg!y-7jZJ*&PU8mSH
ziI>M2d>%2^GJfHM?QhJ`tdqCu?}L`y3{d2<;7`rxc`a8|9)@3U>sMA6d%bZ%bHvGh
zRs$4?D;8Xs^&gQ{V8fv6KMniB!+zqrz{uHwpSaKL(#{O`-Qu}?(ve|@P8o+a!F12?
z*#}NdO#3c8f3-9uL}Px_)V%ME{Z-3NuN%6|@Jh9H*)wU0+2A?DS6f>gdf@i?rR#5{
zc6Uollb$<n`)STR_l&LEEBm!Pi&^{o*Au?Y`g+Q2pDQObH>@;${?+xd4X+pFw#6Ka
zt1Rp~5I5H*Sp`h8&U5i-e;1$MqUM_HWnp+bB67WWl*(n_d&UR+vO^czDL!r36KAM<
zTI;y!%onaD6)QaZXKq*p!{R>EJAW81i#2J}FKe}#YZ=+^9hmHtUFx^VXGsPWiOfrH
z%v-Z}z%nhUGFZeN0Yi=eoxVaO=HnaxH#)QBlP17cs6v)kNdl52FsQE(RK~s9R$<oe
z?aPV`28lR54Lz%#he!nhk<llB0}~F<m{_*#k40MWQ@)KIV8{+2?JI;frsBc#<A1px
ztqKS5D!YMXH_-1Z1Ql=nCl+}Pmv5><O<rXmknH0-+gk`K3yN)DY|b4yTnmbMm1Hm^
znOEs81eN8czE0~G2R-WtKjJ&P=A$*HJr9uz=I2AicXq2&?$YIBijEG1{rNVI0{)*~
zUm>&+k#RI3rrlJo3e|X(Es`xpJr6;AKv7tz>Epjd3##(sRG^y*hW8bM%F!JQCih$S
z(+pMEk5`EUk~qNi6@tpK*V^Cc8AVj9LM>h;97w|XvGf*#N~XV>t%aU}suq;+DrrDB
zjek*ZA*h@`<Q<{5tn`g4)ZkS%OEw$!JOpuGxw)Mz;X%GCROiLpB-=jic?jYgv%WI;
z!_=u$3mWp`jYBqC^gIM{?Ss3iuXk^Is0tZg90ep%{3v=0LFK7^rQ<L)JttK-kXH!>
zl2BmKR|qPdYXc6oT`ZlZ3P0vmb^^&xUZuAXRD74O@cZ3VBtpL5IoK66IjgeVa6DX9
zX?UCzODXg-(Ek~HCM+b&od)`+Y1{~;jbId!D}Yo%;{qrvfTM~0DwJKNaS)OQA*y~b
zk_OZG5E36k>dN{-PVsLOy*2=r%Ljw=fsP>PZxM26RkUTb2FvXNAMXO%!c25y*)CvB
za0nP10wxGf!g71S$9sUbFuC1Wwg*@foDB4nfwABYEbr6$M}V>5?ksl)47vk!g^qM%
zt2@As;A-Gd4aN~%0|wR5`Zd6+hSpEP#wpl5#)S*q8$NYD%cWzZbZkcOacpoLn~?T0
zv2i9g7qI%^-&`zi)qbssxpv^t4jcsm=H7sDZ-5KI7oqb-IF{fW(D??9yP$Iyjl+<0
z7^3uR(8wBOMe5fg+ghX?IObZBb1SXifgCyz)t_kW5REAugU7{Ss(&fiBLz<r29CK*
zJU$b<k@wxkuD9`cf*;{UkMLrGTksbxcp<@^*tL_^k7r!t8LGaVag`H1`1T3$+oL%n
z2X#;6jCL@SG}HmCJAl%rF?R{tU4r&ReiPc=r13jw_m0Ma$T|>F_4gyY{fMd`j;zCJ
zc@3IWgWO5`wP;!`nojT&Wc`HJe~T>NqLD-%jb}t-s{aSE^+8O%KOIj>$L{n*+hk%}
zQojn@Rbi@skMN{NnCg!Lk5yo$g=4M-PiVmtN&Uxo%wz0Ka2K}gqT4&p*d1q-?Q5$$
zmW)l$E98B=2n;I%HbO=HM<!=f*8kzm3-m;KlAbM<e;7LdeD0a;4atXv;{M%*GO26z
z%pSUN0lyw$KpY0N1y(UEz&2_BC5t)zSndfhctQui=a794hYKspazTg<LKLn?Y(0(J
zkWm{_`Y6klV~uibA=DRwFp~5!ErF4++*ROl6)1g#<$eW@zk)GBMd`A)4!STPA7QQ%
z>Q_PoLBKmR_PB(V6)_jic!V=bY}u5U$t@~e3o)01ty1tP;x+NisCdR&XvliR{d;p+
zi&%bgf`B*(=nKS6Ab~29KI%vM{4<pO43z<q<?^6L9;9}bA}A}O<+q^5En5BpTfM-_
z0L^mmu;x3g92(1|FjgrHCC_9uGa04a)_93%vHC)PerV}<WIDDYLu<m*o3N)aG+UR0
zE~nlYsK0j(><1m&aBBAoBYqBAA=?VIgccmuhJ2a!C}iCrma79=b--8<uv|URs;6-S
z&}yJ@BhYH3aTCyLqH#0OYNoLQXens?4%qX&oxCp<4o!thgJQW1I5dOCayV2@<07P0
zgsA?OAgvM_mm#e(8h4>FU1%D4-*)V@9eWTQj7J5NO^IJ~teuBVi98=`=hL_lI~Nl3
zWVvEIrWjN0M>1m~845=;<D!`vLVJ#CPFELBFxtd>=1w5p2|Q_9G#~_wBD5$JNJD`q
zp+Qx^q6*9s8e+L>z*Ykp!S{g0Jt7|$T4w*G&hIeKF5iLA-qHLf6|$*NpH$3%YzB?x
zkd@Q83z>BxC!)U{n{Fqxf#rhn&|s{jAuLyj%?q*8)4ONLtQb@EBN?+uhQiT|G@5vY
zz=MIL=irr{XnuKlZp4}KGr80)FU8~4S*`+DR{$jqVR?4Iwi6=Ca-CpACy6vHcM)n_
zghNUHZ$OP3H12{LT{I3u8exdik45&eh_cvwNaG$YZ$%odH10rb2ht($i^gm;rtod7
zc^exMc_-HFq_LdQlru`X5VoMC@03ZxRRw>P1<Hk-cq_pnex94(;BLk(-USBIrez%h
zoQY}e0pdMiATg~wfV~4gAw8-F7S%w>2UxBKur;(i5*kL*n@ANLQ3aKJh2?7Dh*}yq
zLi0wbjLa<e1CsuLl(d3p*hrdAV>vRD(`S%SY#55EO)4B4gcHWia*5b5k(U36oqxpc
zhaF6>o!*r7gyl}*PfuYf8D9-Ht)VTp0h>0^cn4#)gQ3nuA&glFjZZKmPB7GXbC?l1
z1Pk-bZ;HKVf3jj>U~CweEo{>)7XjQNz&zn}>|qnR_7myea6Vio0sAE2CTvw6(`^P7
zyT#66SneF0d=66Ks~k=)r*RdUT!kokJD%2#mEngEugtUz5{|rlPS{LUJ2HoFI2*Fr
za0Hp70>~E7xB{^ih?2KqjW(?0xO|TpjU+~i=XS(eHk}U|_2Cr6lktCY3KDFEcmLk>
z5HY4;XcSCN@(Uc}KR&`VuX!&i5DwBhlD=#bY4Ij-yh%sSNGOej)D{y7Eh8Zn>8qe!
z6?7mrQw5!>pz@s29S7`cX?ZPls-@+PaCGC}<Ai-9&98nywm*<_S$EuU`~gurOFFVm
zr{(F$F`bsnk+uBqam7xKsQ!jxX(*<~7mkOAW6H)8u{4pECt}M)I!>IzBTf-A+#P2|
zp2AezsKKLZ$T^SY8t|wF8t-7NcQ8~O3}LK8XncaPJHb%>&0*|v2wqZl#%0X|8DhIh
zoJk!}rXadWoJl>68-PgzjT?bUBaNGYNfV8mfk`us6~IJ6<052IgsAtGAd?armm!lf
z8s}m2JbEJYv3Wkh%f8IY^ZF~j5wcvsfB+pKBgt}sU_c<y5%PM1o$}0}JJl&LgVj+>
zOvc}zyn&}w-$R4%Ar<FCp+P8(qmW?~9m5VF!vi#K!WK<5vx#Rk;~6CrSe<{=H2&~k
z^WV$Q1&81}+M99}dLFocNQRGBfDu2?M5vz#4Fyd;n}+oBkrIcFJ>8X()3FAzTq?9r
zg-Yt*&5!LfXe@{Ja)S3vM!~K7-`VmuVS-JlG-4{yN(DMXHkaiyt6FJ5M@YJ|To%yJ
zqSK03f!bA|JTI`^HK2Bl#zjD_h{nZ0t(e9oK&^zvkAd1_8t;SbKKL;ie;j7xuoCZU
zKDlnzxR(9@=W8n18oyU}r@n=SV!7wQ<vCbD`d5i$l_Uie@oRJ2_VeFmzSk6l3GXlI
zOPZ56v;*^Ynq*ysvWt)kZa1Lp293L*tc%8BNESv1yBcIzgDgppTam1lmUkfY4mwyy
zWAkW|1Y)_{*y=XP{_!*#TXoV{&REGAWp*$|Z)yLl->-YfbCtg*f@4<XKOCcNh<F>1
zhZ^xvkMt-HY2+a#PB?5{za`(_<^x6ehZG^P<;smyI;DR9V=0oHlD}5Hm7T>46i}=n
z?sjDBp8e%&Hg4}VU*T$pXCJ-yI>dbQ5u1;cxe<ZF@`}CJIz%xGu~}r*<r~j4o<2t=
z3}Cs<#+#+Wl@1^Itpl6{Z{(3%fCoW=z&a2(3Hw0zzTg%JJP7^`4EqgGIfP0utdhpR
zgJHkZ_zaYtfz-ZR0%avMehsCs>GMhml7=Aa{PGY<AJX_Q<n|X$x^H63n^>8MXStWy
z>LsT3@ifLNjqKw>5awf0?|m1UyT`!%F&!g1flDV)via;IQ^AC&%#*yUMFG7iFhaP)
zV)@n5ivuHs<vV#y<weQLoqzOfCo@ywY~qs{aApRL<#48);L|OePY()x_Tz`t?!TPc
zJ-gzM^6`?=`@HKO1^6gX6O!%c=8O(^D5xI73nG9x0_X{zciz!pKx4?uePWgif#MK2
zkgRbW6vxpx5{V-bCBKBkmuQ@f*<?(m12V8i1|CA%YsPFdEstZ^I7X>2xHOjZ+|CkX
z4gy1ifUS_vV7X`@iv|;k&Bg;+JeVlh?1kb=@8b^7AAB&qf7$eeu*$Dr_Z>RDevUz4
zHd((IAd3N$$oeG!Spt|OtY5y#R!f_oC#Su4kI7FR_r2aF*Acu$XQ824=u^Sa3Xty8
z1Ia@oc|kc~%YmjKDEjqgS()NR-UsIJ51B)S<7ZoDPS=U~;A8)?j|tzR@LPWOhbY9y
z0(1kcXx^h62&@E$ta_0!^Q+j<-5<>UADX=ovUsxE`+*Oc|3t_pk{iAIwiT|g<b%Nn
z|M-{vqh-?T>9H&0UiRExO?^kc{pikzj7+6-h8b{V294!#q@2cG$g>N130nutZO7BM
zV~PL=;|aldDv=lB8HIQb!Nu6K7*q8l8P7<D!qLpEXvSO6Z+)Y;>d~)WpZ4VB#C2rP
z>dwiD>uKBo#0@lV1mZ>-Hvw@IjhlhEnZ^nrR?xTziHneOk7T(LBrc(G84{P#I1jUV
zc(BlM{uqzhe1e6`A>K!NKXfBL5&?!s09S!`K1(=vS(a^J^ub5|WgmH_xF7nU@cz;N
zcQ?E*ce$T6Dn0oZ&kX<fPdHvb9kp;<o%+W;1kj)yC~xK4E4Pih`Rv3cUT_TzyGE`o
zIy%xX%kIUz=(#A9or9Bzd6z?3IgP83tO`-`c5Ky7$AJvSDuYqR1>r80cZc4uIf*-j
z0h2H=M!2W(UA}s8_gT%)cd6gP?jJ+DAH9D%!Vi&1kNnhC{)Ajx|Bpfqrw;n+BRzd>
TU7aCAMMIQAkv{pUO@96tLrcx;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp.C7E55A256926B340.idx b/.cache/clangd/index/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp.C7E55A256926B340.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2dde4d948d3e65d0c5cd0953b7e4a366a61400ee
GIT binary patch
literal 9462
zcmeI1dt6P~9>C99>oiWMt%o|<z4cNa)#-7jCnbs!5gK)yYc#?jMWVbK6zOq`Olllu
zA{h+5NhQ(|Jw&PLfrJbSS1QCOcbE2A>zcpj^SOV_{d}Cyr?t<w-*2tG_gcU8d#vN-
z=IZ*wfWw)-%q=W1V8c=_hr>}~{z5ja|G|=hxG@|~n7GutWo?^{N@La&U4`eDXvvLw
zZ)14-p-+reQigm?@7^PZt>aIRHt1|t?rq<5XolDR0Na$B*_yA@Lc+t(egE=d%WIou
z?_q$@WiXvv>6wxdWNfN$HFJKNB;xWi+o;LD9}g#O;2*73Hh86DvrG1E`0K~r;bYdT
z1Pne@xSXb|zQaDa>*b}G2$`I>X*+A{_C@q>y)`=ijaG9tYBnwVM&16XlE=+8J9I<c
zdNeD_Z%!*~>b2+Qy-cJ}ruQgkM^%h>tFoVyDp$GSkEu((ZCY0Ha;eqB{gDc~H~jNa
z%@Tt{g6q|A-Uo&Kg|EU>9<}fn^7Og_S}H}hlNMyR?Haw<AUq#XubYL_xfbFTe1~<H
z=H?Ep9Q*uq-MC!}C8-~l9FXPgG}gEA-wrk;R+X(^_s)4NJmKJ1RL~U=yJ})npX_9D
zpo^Tb|DrdQ|E!w3<JA7T{?aF5XYcED%RdcBq2}J`vfu7GFUDg0)i)OvEu(6{H8~lb
zJ$c_o2q*SmH_-G|K9!af+cDXv{9Iab#IyW~o34K6bz<z^oNmqP%$%(;dX*ajJm$vq
z$qTjHJa1YqnXbO==%_m}O%rx&uXydgB6AVC8ycZuy;f$au+gfxZo{TIq6>``o|3S-
zuF`cr8ZWe)xR>9aJ9cbq_vi;FUGHc+`B`tPQLd33D=b^Rd1Z6^@BK6X*?)JFN{+ZD
z*msUw;NhfMYO76k_Eg!mZAz*2eH1v|r(uoiFExI9P?AmY?AEl?4ksrTU0hp|VAgo&
zkoELO8y_inwNB4Q_vZfodyK7z@1)*wi=HEEr1tKxuU(JAuP1bSpJmBu91&}!YMx0o
zGnD)8sN&4f9+eXBYZp=n3Uanf%olCoZ}n4%Z`XIX=}B01e`WB$PF0DIZ3+7A)!95z
zkNxc%Q}2Xo`l_!uY^Rp@ZT_zO)A3&RuWTk1On&j;$;D@!jDj&gD+I4-jq^3}d3?ge
z`)X72uc~KNkEouy@*rk^=!?@~A@>yZs_(len&(R@^|TVTy5?x@40(9{m|E(qxeK&j
zzAkzb-hCH7eZ1qWRYzm->a8F76FS4Is~ZO{{`=xjsad%;CI@F5<VxBE^EZ3?&v3|^
zoX*?P!Vg`1@yHrU(?Y9+uTLbt>ss`^P1c9T3yoI??raWGSzEeNd3R6U53=`<8Ab%q
z+{<>J^Rtf&cdG>L(u;8BI?1;z@YrSS87nTAX@93N!{PwfCU(Dt>$H4Xx@zUzw_nx!
zE1r(ANSvs2ZIQ{1OSv_V92Y*_!+Z66==_qVzRe=<bf=fIMu*oPigFkwx9@q}3XNj{
z<=O66-49;1JCz?-Ww%3QrXsu(Whz)4<6FA?rbqvQ!9?>_v>+A=?@11v8~CVEuI!#{
zyZ7VK*|P^@W%9NuhG=p8^=FsyH@T}h>|Aet!^wrOC#p+(zi80ub)oIQN+ZCe_pRBS
zknli%Gsxk{XoWBQ)!;>d0FDxH=s2Jr2Q)`693JCin2+4u{6Ch-w8+XqMMh;af3wbT
z!-|zOE}pz;a==gy$}!>?z>fiAM=q?2!OVXoF8noml>o{zDv3ZV5lkMruqspD<t{(<
z{$$x0NHHo0fyP0gJ#t}HrWy`bt7I?ySpc;el~DCili`LHdt~rx_6oO;5<ng!jtA=T
z%$t0=uqxiGruiv4)dtH!B}OHTAErCpu;Q&$_qff1!H05Ch7oV$Z&Mp?Sn-~Ru9hG6
z-e?g(J|j*5S_#Zse7dkI@eS`T6ba&1%fT^>N+jS%0`-v#t8!WAx0s&_qBG>+XhtQ7
zA7n7xu;Q!R9F~NIhn|#!3XC|AAE-Cnu;Ob2n|<#2`fU(EEk?XreYJ47Va3&Lo^EBT
zoR+WPS1KHO>8Pcm;f5p2WtN%)XwYhSwQEyXl~4}KGyT`|*Xs{AZ2#V#(QksxXO~eN
zO{Tw^qMGw?!;wXRqod69=d2G{9wMua?M8_|VlC{gDY_0g)Pbq&f>N{?I*8$av6C1&
zV|fO2$ROnz&?$qI7eeboXp8krpmhn!r;y1hWQOHANSK2xFrP!hb0k+FlL}J50-0AJ
zORQgoOskNPWQ!_fg}D=%bs}?;Ejy8jWw{mA=bxCw<pGM`rn${lYEvVd2=a$F51y^v
zz<f7wWesPvF#6|JKbf~FP;|I*xEgD)44H+CFl}+#;>vDIidg`iIAF-`WQs9*op@l#
z?gxr)0%Mzi7W;rhY}o{e32g@U&0rdqw}P>)q<$-~Y$f#@kyRtI!}?9gs)^)gWZjIW
zIE)#cX&U6nr|7##?=CXN`t4|5JMv~(fxhi5Y`C|Y&!xM8V>fVO1zdU%I1Z9r5AExr
z1C}>H#|B8~cR>3NQvLusJ|N}2(4rSgSDH)r!LR$EH6AY#IYuI4yd>n9gqd}8GW-$u
zh8M99nDqhagHiO4Q1M4d%=j~C_6$E5MfXF+ek`Yd&*6Q@I%lE8rJn=a=Ri8OT>1qN
zy#Qb1C;SLRA4#r<^Xs7(mfwb=+mO)jfb%;@`9mmrNXmQRtX}Ab`}aYQJ~#(+6cR-t
zV!Q)Lbb#bKWKf3;Z}>dxGxRpS!=>xclsYsOb2Iv`87*Z^flId`VH+~Xd=Kg0Lk9SM
z?Z~MeNlj5r!Qr~o@k<hBFI@x1*MM}Xbe<X&%5J>x$OzOpYW(3a#WZ@3`9(c8H~C*C
zlK$_BWCJkbkAz)px)L}+0&Un8py&(8<pNsBvhg;X^n{DTZ)7OC15h1cGAp3yN~l{2
zrAx(l6{y!ly6tz6dIt&pXhcOLDVsk|JEQTPBnWWnQZ%U)3q+<JcNSB^6lQ5R@;2(Q
zLEpyEpR-h9Nhspd&B(18xnmn@NB-^Td-lP&^nGM{ANzM3=Y!691G)--x{-ZQ^oq=<
z(dz^hodd?@0K((v0YM%xWJjdv0w5?L`8*JuC%Fg+ibyU7f?|?OfuNM+G9V};xf}?}
zNxlLES4frsfrR9p(0M1`)f62Cr$xc}_<5tDC>pw8j)hjSa0cdN^mQ_F#hi+)Q_)P!
zVq`AHyO5&OkzG1+$MOthn}MWWgQ7E$LniWI<qq;wDwen!gfli1W)e1!h+0ZptZ>w7
z1j0tJfbB%l&46kKL>O-Y!WJxN^#Fe)D`YL0H4rAS4yA<bzbU#FP_;mZ?Kdr^+U`Yr
z$YHsm)uwW=fTDLo)19OtM?sS)95pC98tO(vBL2t1aj`gxP;@fVOUCh^qEnH6D)x~~
z{)4o{WImCO#-}49{AM7-3`Fd}Ok|vi?AZsO;iYkGYd>!W6Gv--s1^{*+6@%D!8lft
zV#1~3Ajt`kN`RVJUJt2yk~<*PLGo*;_!<)9MIyyWL~s&POd|OLGQWV$WM+fd{*bRr
zh$54$@WRIs%P!jDh?l(yP)$G;KU6aiHsgp#(XD`LCFNphE{4+Cpy&)}o<VXUoLq<_
z8bz1D$t5J8LVBmLcVSi^8RlT`K+)%r;W?5kkX{9;UxoCmkO_X?2ISa)+*t2J(VfVk
z6PdDdH`jH#J}EGl@m@K=A_qv*LN*&?%=A-Df(`X1aJY%{<vG^-gH%o85}6TVfL#o5
zVbhK|p4MZpxund86g>dk2Y@t6Q}j#d{t`!RcIKH|4V5_fMp3{#3d~|tP4B=QYY+L1
zZ<z630qPZyz5_)+hSXyy<;Cw)%pVu;vijR6Xnqcdlrv<Wto@U#_pdH=UfXGgm_3vS
zjPk%FcDaUJyHNqj=Yi3Al8b;*5y{2CsF>tZU{p$S889j%xf~dklY9jjT_ITlj3k(s
ztvGwhyYKiTW=(>Xf;HJ%k)lJCLNwXh?FapGm2soxvlt~kuAZW_W~0-APC6cIO_IOS
zj1Korj2qq$_4h-<wI)IRB$7)JRf=dlDW%Ayl=M*TNUa@7<Kdd}tO<t>_x<B9hhnxC
zVzQ8-;u;%fHl(tl2EKnTq;fH{bvKVcQf=do-N39HZ=E3zXf{Z4J)BezrRg-2!@x-$
zB=<tSUPzR3k;p6(5v6JpGE2f7>y9FyCw#PFiZflVt{k3teXc%123!Mz5`Z89XtM7>
zG5MDu5oodr8AYc7ois2WzxxrObA;qyfX*)@R|5G;Ai(<9fc!O*uLJq(Bv%3XDw2~S
zl?;_}e+iVAKw`WyM3o__Ja%=5#;v)nhZ$STgxZ-<nyORuaj1P9GwW5Dwf%JLffp$g
zQki6!IS#4gnB$-1u3IpzE8%}`ru*uj{Ch|7mNS3ZOp}+F|1nif4>9k21}dI`+IX^Z
zAe95vSr#9j^k8dwsP*68F=lL!O{kH#cKORwaXPG6W?o>1D@D<PhJhyR;S}SYM8Q+A
zcM5h1o`Wew6f%YNRgC9y37Lbb9ti6Jk?S-7VFSsxf$%oTkAcZ!97HMl4KRCy%Rpws
zLz6)$4YJJ92QrDrxmZEh=7O;?Jtx1^gP+xd=a=8FnZd7ON;)Q+<{kv{Yz1BHpsVyK
zx<8&v(fc589~{HlIz^v`ywfDtBB~Y<iQFxucneAQM#;5CpWp0XcYc}r&!(Q;CtUvU
zxi5B;HPE>R`r-L*gVWpKx0t)3SvNkaDL2|~Zk_L*`IomeyN4M^J#@N^@4s=}MjN)6
zqUZ>_2ro8XSJvOEu71{g>2L4n*wb<5-G+|F@w)|U1zWNc&$uM55KGo2u{um$Idoi}
zoi?c~;KIJbI*0LFy*xFOznJf5=6my-RrK49hjkdwaR%_t0OqIXPoL&qvDeC!m_2`c
zBVkvbF_NL*KUf1`j+-1dF2Wn&M_~6OSio+9c9GwM9eJIQ$q9A<*B!u(6;SjZFl7&L
zV)LI4o4V<5PSfCv;e2K|U9;YLru>lF&(vT$;liEJ7kjoSxF8DpV2*|}qj5P)(Xr4q
z7D|toh7Me2C8N1mKNY#9B2Ua>G+j&{Po*P|bR<0_XO0li><mO4mt~^4nfS=_`GEGS
zr}f_Y!^I*m52)aQ3jnWx<nw@cp5!9HD<ZiV@QO(;1-w#{%K)#8<Z{3(C;1BCT_ITl
zcoNJ#;-WIMY+ZvdruLbs^*rl{>6CP2{okGSu@xl~jD~)DWrI=wZnyIWRk?kCI@fp?
zk~L?rvH1<-j3a?oBoMM*=v`xk)vcbKi;SQW7+2zw`2G9iWoCDhdl^4A0L%u!ENt8_
zq1j6)<qv*^;we_DgN$M})XIkXSTPrB<zi-j5ni)&C<KJ8|2|~q&==LA=F>-w{lejB
eNWVfRi`V&Bu`Tq9jE1VZ+E~t5sfeSFzy1eZ(C0n?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_multiblock_welford_first_half.hpp.25F962D583FA62A6.idx b/.cache/clangd/index/gridwise_multiblock_welford_first_half.hpp.25F962D583FA62A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24f967f82813d3864b05b87c418a93d913c10392
GIT binary patch
literal 5070
zcmbtXX;f3!7S0YRAzULUkt+s}LBL{67$SpV4j>`S3gQE85uB)qK?<lqOUqESMdTHz
z1+^+DJ}Mxhs91zp3W{2%ic=MeRJ2eYRcggP-$}ZT=e}R>2U#n7-S2*9pJAUp-H1Sc
z|IV2#mTy8}#>T`g?`gAGECcjU-J)Foip^s2^jNHn%5%#ehTaqFkMG;9WBTsVS$9W_
z41Cz{(D8hZ{?C^kt~599=`9&Hr2BLB4Z8$>wN!ZTT~Xt&6C>NxA7=6HKHNGI@M^wJ
zhyy2N-VU>h+|4B$KG?@~o6sGSb-nx~u)3{uOYSP($y0lE4EV7_kFL1w^a_`D{?cx_
zcaQM%VGmVP`O1bD(&t_F`q@DPjkYZ-Q{1LGtUjM=Vc%R+&51PXD=kjpzT)=h=I+?6
zJ7!(t)nA$$DNGPg`K-EWVu|aLI+F`|e+ukPRhB=3NTKXqRv-xi$7#FQS+{6q*KU|&
z5S*-7n>+n6Ec*8sWjmkF{l1ttAsI_+ZgkDIol~$ei#M^+@y8P%8HBX={h{dC>LS)n
zPdRqJ+~kB}WmQ9}nNz;6yj-E`yeaQ?Zq+^ARljNVy;Humzr{N)P5Sirthz4O<0oR9
zRz2DlRaG2x_~#R0`QI2fL~C7%)_QezSpW9?gFp0640D7N`%gZM3wkOSKNI%e4_IL5
zt&+?&ddeDp^Jy3F<f^ahSB_4<ezfw<%4O%?ZQ6hIAB)!2-rlRa<*;>1UBs^7<f@Dh
zX1zR8`fSR-teR5pp2==ME?av0pz1O^!Cfo3JSovBATHeL<S|pX>=%d9Tcegog}iZV
zO7rwun6!8}yYt%MP{7)=PwI-VH2!|hUMtb}>&;0yf&X;4Qxv}^cHqg;M@M@8I#QuH
zZ@tz<yLw>)HA5)*IZWPhxS?ZA>*Z3Li@Tppn<ufX4jL>A39jWo{?GQtFyqSxRemKC
z?;pF%vZ*;?OQ~cD`qo7kWQ~~>J9G}Zb_BjURKcz7QhwcauH+77Rp69RS>s2USLGSE
z%$Tv`{xebKpx={Yc5RFD<Q%h7X->VJc^=hrEb7;hZ_IzoY|0ww*mhujM_x}#)F1a^
zVofi{{AI0Jpyy)m)VenL@!27x@}h$IcFuK|m%ih8=8bdSsjB>r);9{zyv9t6#;pFI
zT$25_^!t?UdJ$`(+p*#Pn+CDa@AI-#BjtsIx!*l;t{k#tr)F-9pNS%XYZMq)A20t+
zk3EggqP@6YRvN+LpeR8<@t&!t&#`W-*JIB>bO<+Ojz(bGHTRR^ip<1|dTbt|eYw7t
z8iDCe*&lgk>IV1dv8N;2gX>|T5t!cg<3+)Xn=`lQu?-L%$PF~t2uv4^rVLy=n4QmO
za}d44c!fYCFuiMo{1?x`)eU?$7tygMu?sW;)4Nwp|MR|0CzsFGMRc4=+yafjbk#wZ
zk7N_nnHg*f(eZQQO*I13hrby0+0)*<hR-%dwA5HC)(A|0d+oh`r{&Tdie+NRq7C&8
zoiu`_s|BpNGl7PSZIY=qN%|Fhwhq#dF^&;v1lI38-Jx?}-EEtx?5Rj!q9d8F5iDJ8
zbR{gLe{WOVA-fTkK##45^rN}aMjC<j`$sQL2r@Z$_-r#o`y2ZUH3HMWZdjCaURDyJ
z$JR%55I4w7BQQO3$MQPO=Wo|zPeF7zH{4hwFwIgXu1U4yu~=HlZHrx1gPv0;TA(e^
zdwXblz1ezeY(R=u3YAU-)q|)W77^S4q6UJoyRo?%H^X+I=o0WLfmp0IO<cHV!x#4V
zbSYXsQ$8Cz3PK{egdl~N!XKZNqGQZr=3%>2^a`^T^RNd|bUqm6g9)BqjUDsBmEb}!
zDrEA7;8;lb>GVw>@r|F7w6*DSSX>TLtf);_fOiG>5L^qMwctf?D|ogNOy6$gygq!&
zK{HW!IYf|wUxMUIVkd3-D7YR4H^QF-*K-8Vo-Xb7syY81P;`cEhBF>!He_>B_k%~C
z=(QoskgEw{BrH5noE~%efjjACZ5|GXHO$S3Me~PpLTBUbV4LN#ulQSWG<v-g%y)t}
z-j6Wh^<ootRtIR)!(cWH=9s{m$Pd&>h4{f$u(%42c$j2i-~Hv3xB^WazzHzJZjaNB
z7FUyQh<1Q9U>P1n(MuOD4Zyy>cvIop>Y>8T9EwijC(XkPP;@dsnSoOHDGa3KD;Y=y
zZYq!y%}(&_giwOJ!J`{O2<`!|9tgu&%5%T4%HJweo1z1Sfnf|MHc&VbC(Z+L9z<d#
ziY@?g0m0rym)u4krdFXzRA8+l9!k-jVAlyAM6ny*=>~U#dtgBicw#Jzm~&|RI7fy;
zJ4=$~hhJ03zo~K%5L^rTwJ?|9R?u%{;&yOoC;lIBqjK?AEBsQCmE|1y48|cP3Quy#
ze4z4S7MZjVs6r-w02U8m0pUl%VUz?NMUR8SI1|4H(Q9IhVCUVNW(jxhLQWIy7VU>q
zX>h2sQI}suIGduYL0%2PWd0VAw-6k)vC&z<I}nWqRsvNC3(3HGpy~-;K2c#mc=ltb
zNvDPQODBDjAB(imwI<ml*^%{<ZIT%%#U_P;ls3w@C}C~=`KT8)18B?M2HpakWhq)_
zAQRwddA~H?M%E`!M}uy{{F@*l>)e9*w+O~%8-<s;W;4QTZh#93ugHaw{y4nwO`~{L
z*Kwi}Y7^>#gZ;x|<0BiRvaHa>41wto$p$!2qXpGvo@l29r!fNTLy&$*TzrRI>yzQ)
zH!gY<O>5JP>J-b572xK8qT~2+0^AEwbfQ@zqjIJ3()h#(>AZ9X%HU-%P$n;vfwuCt
zGEfolihzN8f$D{sc;3&R9N!%2{jB)^S{-j$o2%(#h!M)cuAF3kimnAwEinT{w}PmZ
zV0>0&OLfyjrYQouBIY#h;NK3*@HWfVbo}Bm+d>Q7%^>Ixf+;2{tL*x>cceK_+R5>R
zz}0kGcw@NJLi<&>S@<<_WD{>kh9jHE9I+$UWt^Q7w()>J>JB{Sc?hw)A{B|ZJ5llW
z^A09RCXu=02~b@4$$|;W1>0N@;|)^u$6)(06CVK40eD+vDY_X%%>*|bP=J5wvE#^e
z;U?ii91_jXwl*6TjUAnI(QR4ts7J=V*oWqi*N#Vh+26mjQgefW!tu|LW8DDG4VXdp
zKLDHof?K<8#eB2)_s5go2yOt;4XRt-hP`%eiDybqet-SirAtqrpP#gc;O33X)ile*
z6>_9GjLP-^gbzSOPWu4tACQ)TqDR4flvFx3_8tXaGJYJ~$C>;%c#kvr*WmX0f2)hf
zYtq79m0E9)2o;n~+M4)y$eQX80Bp_P=amubDvAmwZB1PFrf@Z%4tP0=-VdVv%=6WO
zejT&-Z30mflWzn4Hp2J0T=o5<j)qB}9~Tl-GSnY1WPe#ymKnaSXG>m&R;>SWT+{F<
zRHR{6H$35;5RC^?bT@=`LlnV15YYq6F~-dYT21{uL{<y+3SEW^)bK0Q4MW$XEY)?G
zlVpf>QIp6?W}p;K3Ii!Q%C`u|K3YZnSw&VE0&0k~Rb!Q>+e8iKR+F~j*wenE$KUlG
zo^(I>=dB#4Yh#Vyq8yWMoBlS~CMIgyMEA>{qa0HWqH0(|Owt0P7J|_?<NNQa^HQpE
xb!w#gVYXWQ{S)8^7HbZ9G$xNmZy(vTpH<(A&oA`#PMsr|Wjx!6HR~<M`Y&}CS&aYy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp.BEA9938289B2ECCF.idx b/.cache/clangd/index/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp.BEA9938289B2ECCF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7ffd94845e2cedca2f2294fd5bdd7301dd534e1f
GIT binary patch
literal 10526
zcmeI1dpK3u`@qjyYbzZ)At#&MhZG$Va*gB?Ceg*EX_{`T=|WAWq9|b+*U-gf#6(3+
z6d`peItpdF2&0<GC=^Q5q*7w~Zfc*kzR&M@emy<Ezkbi;slCtVdEd46+WTGWec$!2
zd9KdR<0kQV4t}md{wu#<M)7z&dG>$c_nX$NW8rTF9xv$Nd2ja(ZPVnMJnFT6yeB^P
z+nm!HV(l5H*WA_IJU=hO)K=cq>Gw$8d76{-0t()pKjPwgr2V#cLD%Io7QN0@!R^~+
zFQ`u+v*O$p;~$h|cRtby)`9OL8Z6tcP+vIjvs-Fa5vCw~x7?;|`GEZM_88C9aE;Y}
z?a7=~=%z3=Bz5fz-Eq=!7u{b|<GST_`Yq83jfiKRZ-TaWullk(+Ffmh%U@dWs`>iU
zJ#*mF(Uvo;ijRtq=Ns7@T~?xEMyT`8KV0s1sYWd-ZQB!`?HhNiuFYHhYagZFr0jt5
z^;=EhMcO072V(1Pr&u}1(l-8;{9oto_MV~8Buq+maAkynKWaP(j5N%O+!|9Z_bB{`
z*SDReUV4MelR>Mm%bY|yaK^+SQ$KXCXp+&k#h_#foHR9E+1qcde!{G^9@{H*<3}v~
z?p;CeWseDW4hQQU(zzF=pJR5#+W4m}==j;pf07^2^N*jp;(zkGVL&FRbASFQNL2Ks
z`bcQ>>bRrJlJq<u7%v(tvRGn}tdw&t>4vdsmO+$G+k|DzU%4)W6X(4-e`vu04IkA6
z=g_P^!?MFG8bqG%T~^KC_tbdE8rl`=(oc((u7~!#t!zHs{~}fFzO-9vPkNZi`V;n>
zYpoZ1f75+vdrV$pW8rTBdFxHj2jyPTcPx_f(93r&{&MS)!doLNf|uOT&-zwr-Riit
zuA-Z6rpJ%C-FvEUbYamZdf$dMTKi|3DeG;nz9yrb+oQ5?zvds8?9LbzKlr)(byeiy
zuA_;4zJJDlx!0@trn}L_5iPa;DF1NxQ?L3>A&bjJAurcAt}oRwttx8pviv9gdibeR
zFI$SX{&@STrp=}APUQKd`AgZ>uHLF!Dm>HFq`p-xa#4KA!j`V%5gTLE&!}{^9X=K5
zE&J8Qw?V7jUj{V4U~d1F;?-|IXQA$2f7jZdtx+3!|H})Vz3r9{TEmuYEiW+dys)ZM
z_nmdQUq{HG-1QSJ*3aiQ*h-hpd6Z|_8|Zz$p*63p*t#x%$335Ijn-hvE>Ve9cEi)_
zKCuD7{5RwLRma!)xaX}~p!T)Dy;@_kOHJ~C=L+hYXoJ(WZOxGf7oX{=>DpHI_d%%+
zKh?5P{dYVz4r-lMA7OqdFEF=wLCWar{6`g8a%-*fd?#<{ov}Ep^4IR{)EkBBQC9K3
zQHwH{ss~lsJyN|T)uQlbY{L&<^Ftliw_4BPM>dTZJ=pR~%Q;I=hq9*uDwiMg@*kWX
z=rjjXKb`Qeo~tNaxcl(Yuu+PBY8_AH19@Ioo#o`Uk~?Nsrfo_WJXFk;&zdMyP|fnv
zF&4xbxE4=S6h7D1AG{Kh9KOcupXQaB6-@^DwmyGW@=M%<>&(}UOjxz8a=~cE&#gU5
zxl(-VeqPgVJ-2|$c}Ym+*&jF2!wwCdX=5i}R;XCnciLv|s}t*g%NZpcG2<7L$&L2!
zb}o8-?$J@to%^HyaJ(9$b@QjU`yQ|IG`zTNfw$ShP+wV#RbOd`-LaeD`~`ehI1tO^
zDZ7Th<9RHlOn8C8{>x1uk2mIdR{qq@%7!vfPQYVA`Jo!a7oIfhm)JkeMD@F-zBQTx
zD9wsvfl4e;{cz({3~tRnI&)=!vJ4!<s%+$M)E>TYV&N%Ybjm6Ei42rs#ZiDC1(ZJA
zI2FetQEqW)*~}4egd&ge8|h~@eBq6v*a7kY1J(#{zv@?)ij#p-Y#$MT9|2T8+_*mc
zm0SkB()Dl1KzUYWJAb>@@P!kHJ`=yrl+_d|K$;aNg7Jw!{lkq@iCKC0x0_XAKhQii
zb`&adDl>;Kyio}7G{>@|*f)9HN^S4n7cx+m?PD{4^B2PxuKh8!J#)+Vt@}m*HCS<w
zN|12)!ilSL$E5Ut`!WJJiWTnxYJ0%s4>wNbQrI|!ppN2g0?21o;(=N`tMdNFsnm>o
zS+Y$g{)hmovnl~90Y<|YPTVAKc4O}Ri2X89ffa`Vei#`4;l`<Szv5q<Uu^6ufFoIz
z7@!ivTJ-&mQ|YOzsrGGoI$H*gW>q%vH|Y#tI5BV2%2k2gT09=wwB6dW>Y?38nwh|#
zATRlVOghj_2a~yPPBR5SSO82hmjYoa$(=yhNpcquc99$lrDNe3N&ATD=>wVGTLI1d
z0@#15oM5z3)K6;<82?VwOvu;}Wo}BaG$q6g(~cQCoVjUFv#+iZ19Z8sLbH=wBNpg#
z^PXnvfl@tC;~p`@CiTD)a|5t$0QQ)hfKn5w-vmsWNc~)-m5cPS{xPI`3>k1dhH0E>
z)X=g{j$+z?T^pFe2`Hu)*!7ZJ4Xvx84VKqHyBbL7-+@!_Kocy#3#Z+M=9s&nT^Fg}
z2d(?a{xZ>*nP_%N>yaMgpLZ^&m}AKC7;?hx>ybk}a>U$*mbRf~m^;un9cT&WKIG6x
z>K~vS4p2n<9LgaFGiM;kANEI_i=dfZz-brwiks+6bD{J=Zt2vq6w?kY+kq8+UI#Gi
z0MjtP0G2OE`AcB-l9X4&xz%tU)~|(TwUB6k2fEyWu2_B-&b|xXF?U0=Zc_g>G=2@I
zVtE#F&O)=u?`eJvSz!4sWO)k_<8MPQZHRbYJ923!xdS<Opjo*6L*)FB<k!gZHQ8S-
zWtmG!`d1!fb7e+O<z;peRe_0BKr%5jy7bPB+R|#r3RHM1{NX4-1OnC`@1p>=NFA`L
z1Gd~iwC#$|J-4{Y|7;QF(=0ME)Z*ua3r0R8Xy!JcZ<GGB6VRO`mqD#ENcfx<NVh=3
z-#&plPoQLJ(M&axt42ngg}8u?_``u38!}ik*n*2<H1h*6`~f&~p;zcPV66s2#gJmo
zBKNb%gX@iAVkq|*O2S;cm*Rhj^0@z}$mA)Kcqp2AgB0H&Ev_S)$)HR!C_;XWQanaU
z<i-y(PmW)Dc>_CwOf)GI>EoqzfHFNmnQ<MN8Q%ZHmo`deXK=7|um<-=7P`xM%f>9(
z%(l-$6SL4H+`bigv?5Qgy~WIwnMM6tV?O#I>kUz>k9I|Ur$jS(K#&J?xgKcdBoLe=
zxc~?XNG<|`B9cphpoHX7ASfkS1Oy_I%YmSr<O(3DAh{9<DoO4GPJO@&kFOs%_JakO
z2Z8e-@W$K^t@@!G=0RvX2<Ko9L({{M#QD)o1Tu}l&YfmNNLhp?Vf}KXTuyQYva7&z
zh-SpdR*Z=L<0#uWir@swK7pFc^=BhH{n8R=onR@N*)H7fNrZO!9j095pqZP%=q8xY
zHKLgYKsSIXm>YpnBbM98)LA|KGcZGjX8M3}A0XCGKQQVC4!GeUm^ug~p^#?!;e>uj
z<OG9Ie-KW``eA5N7?OnQp|7PGfh3`sW-5?D1(JBsp|7DYMnwBKN<WSwIDrx-;DzEi
zZ~T#vSJIAb+EWY6YRQ4M0l7Ayi3idP<a$YthIBNP<WH=lfOIv<x1ro^NVM;Qa$O|%
zLAsCB-;0d*BEm7%BgJ~8gP-4r6#Gcdp%il{iJbEkfIpo77i1b#tP~}knbmO6I{hoB
zz_Dy1Cmo_dSleue9kzBopzFan{Hz9G)PUnX%`^eJiIgWnVG_=VXeI>;Q%KH$`WcW&
zTBDJ6G$NdA0@6yri5tzNA?-9$eiYdqMXpy@DMLM7hZ{7LkG{x9LOi}^WYCNxPK##l
zA%lA)4<P*kL_9y7(hsKyK0z6tpd|SN%@k6Gg_zyV_ih*`j7wylU=*;50#2M?cK0w>
zs&z`8i)f}1x;5eqhGu%voL(eZWHgga&B?}#&fCAxe70=158JQ~(sj@fze+cf>qe4P
zhy4_kTpA@|-?yp856dD<|MfZT*EZ>6v!yx%J{nk-*PWRe&@lzED>@J8<pIg|U?{NY
zog}#c=oOG$1oVnXE&+NaB$oocQj$eLPegJ#&?_go0_atcTnY3lNfsdk5h8qQIWj26
z?6>+<rT5F6c1SaU@`0+{#*${X$Zt{Q_S0O7!KRk)xAvSLpqU~_7h(Uhe&2FE$NO$u
zDb|BR{Ro_F(@Z4Pk0iMj8MPw9;~t<C4^R?EwqBe&DLv!mqK~EmoUdVrHdIo;1KR~A
z?*b0oE`(bNtmArLTHqAJ5>$@)dxz?<tFQ>F7ePrf!|j#X7DJT=&eg@bp~#+ucNVPS
zwJ0qaPK#!=DQ$w{fgm2Ja^(V=Eh7l_09CHSKr_ieBN<HKM#Z`|ppind7|4o&0L!a@
zY!%6ufb1obF9X@jB%gz_=OEF3Kce@ev0Q)gou|K>Z~r*q|7@nDTb<8#m2E%yQ8RHf
ziOpz-$}-&7XY)UsU@tEJqnX{na5tF6O}AwA6{9ajfA@&B&3~5>l2?fz*(z%zR!|7#
z3ZXjpVh1w}9)yTDoB!(_S?1%5H;YuS2L6vvm%Sr4;a6F6r;l2UGYY$~hAL&)Vzq#-
z1sdEJ&kHJ(^7PMk9z!!5bvFvRoeS$t%mSxlZxZMf=#FWN*_P?tj9@*D(-wD3*MQMA
zKm^U}!00;3H-OO%lJ5cGJwOEM9w6)?xeW^2AR+%7I{i(C)H-BbhltGZ88Uf>thv{t
znV%_>pD76!WX}f9`%a!_ziR|givU7yu9iIBR-zR1{Fl$0>QCHMpIz1Zt0VvFefEnc
z0xA*6a`lXJKGPDdtA1Bu1-pRsE}+i&@-iDO`Sz$+@$5D{97=~n1#GNSQ2G?f2N8V`
z5gF%UBzG98VEqS(en86iQS?4aqF-Lyv@qYgx8t)-KhdUQy&cBp#coGG>SV-6?4#8$
zI2f`vJ_I!mL2b@7T%=^<#rr}$_CM#~q;rsP5POl+UL@Hlb0L?{|4^654cow!He3Z`
zlW8!em*i?_Pz@#JH=4NvweLV<ZbUTG1*ddDVh`8{4f;s^dSp_MW@7z5WYR}+4rP)<
zN#xu)0sDqShXi<-fog#!gsW8BV!}Dcnp^dCb=@y2|Mk)BhRuYKc^<0CpBLsQE=<ma
zvO)d*(G#9_hai0jN={L3zc@JW+tipJ|MgStzQqlPzB~qeo@$>s)wuIIMgE7gOFY%C
z0-LL(<4uAVNu(o9f;LH{qfLPpDWp6F+N6;33}}||-*q9&3@Az2XeJs7qw$WHW)jfk
z1Vj`^(vUEXl&2x%G_v@Yj|}s1aek;yG&P^B`!u6z&G?*Xs7^HP9?1j9Yye4)Z`iDb
zG7F~&K0#TYz$xZXoyf8f^JACS9;s{6UJX}0efq#+9AWx_cR%<V^C0jY1m9rphqL?f
zff>yVLeD`+)R4oFM;P+L`Vq)E0?o!;f#y~qAIxIpDMm#5ILb4QA~=DXpFl0-^tm9$
zM!2C9YHa^XII|KG!>)%@>hZZ<Prv%wu3O$;43}1<^ME4P-%x2q`XtE(K)QhBA|PEv
zatV+wA-NPtmy#?3(jt<}fpj^^6+pUz<Vql2NwNshBE;bFl_R<w^Yeq}Dy9@@>3nt_
zpXfTCci)fcxO_kN|6QH`)o$0bI#*A>Wb@wt{Q&Gu(@xX+=Y<#8yjKh+im@|%`!=Uy
z${)!uhHJl58lelGT)oJomt36Llu0%vS*)8@u3t8^a|+zFB5?lD59OiD@E>Hvi5ZiX
g-r7cu7_T}`MOleAo)htSD)^@o{;7(8j>A9y17X|;I{*Lx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp.03EA80728C2C5A38.idx b/.cache/clangd/index/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp.03EA80728C2C5A38.idx
new file mode 100755
index 0000000000000000000000000000000000000000..63250ba5410719e066ac86e894d9a59ea3f4fc9a
GIT binary patch
literal 9856
zcmds+dt6Of`@qjyYvVZW((O#ju5_D<F6T(P3th<6NGM}+9hXRvTq0p&3S|yW%A^eC
zGSh?#O{H;EG$l<Fku+*Tq(pinm!^4lX`Qv+kAHuE%ugTIKHtyt?6vn=>sin9tY`aq
zd3r{CDiF9W^@<2twtg`q5D3)x|KaP`tqkDdYYl-QBK24QhR_Ck)mx`p4Mud7%)OOt
zzt5KmYQGn{Upe>3ouRMx95uiDsl_bl<?v$Y-#slJ>Rw(Ov(oEh1MOEVx93!<XLdF}
z-8?cb>Ra<a?{_L0RTwqDUXyBNlXP8rEvRdlck;q8ztO!;ivxc&c|B2BePrr{&c#EA
z6eW*Uj<=LFM3!GM)K}{|8m?5O<{hTI;8%ypBW1bjRwc39XYaAOu9iQ<u32Mlg~_zq
zElLv{t37@R&VAk)weW4@l4<hE%TdYc^E-5xhVEA>=ysd9Vu-lWbAz#;?{6ndH!a(%
zUH3RpwLFboSmNF=IbC4AZQ1=;>9OJ4pkd(g!{W-l+}f_QPwuasz4QE?=u<9tc6g*5
zQ&x3<WDIT0rg`rB&z<nl+1o7Esyo$b-B11QeDv=RJd#c{>pbfW>@&`sS-0Laa#`n3
zO_v=jtjkzIX`m$9-Ap#eL~=|k%I#iMQ;+adP3IKTqlcHM2lmZ{^}f)2T%Aqkvm?Gz
zy^2e6)vPro?kYe3RR2%Xean+mTxu@{Do@B8y1%6?O7jQ575=+lc};!Ln6vAd&akxE
z)ty1#nm*XN;D?SMH|5$ryLo4`<5Z(Vk1Q80NF08tRX2;1%aylgc5AO$zyC#9toE99
zyI<(6zccgf#_#7Pmqi_PZ>@b;<y{f=%0h2W()haTA@j$461?Snes=oLXTHmiuUmR%
z#@<bjY)^cVRQT&c8}_^EmBl9dU&y>V!on{<DEj`ZBkH>>Cx$E!`b*2UPtvfxblTF@
zo_XSPwRcht(B1M%owSaIY~x*plA9h)kH@_JV`18`#gX2+htqt2Yh_P;_3}|yTa~$?
zvv9_ZrfVJJpM*s8%(^!wU*4!v^7)Aik1}-HFQy$ovoqT*U|6$(N67h>@s=IBwcF3e
zZ0lPTWMVcaLcc2Od9TR@E_L>tS*Gu%7KpxZJ@DXeP<3bY7?ncf`^#LO>YwDA9OJuf
z>aY6LSXE9QmDZE@%BtbSWLf;(3X8jC^OoM6|I=sYuUhK-mivULotyjRrFmf&Bp1>h
zRRglPpHE3|H>#SgXjfTTDC(}C{qw%QI~&Y>i+_q12g;Ss#1<?6q5aInC{sDQq3o@7
zd4cmv|E$}aHg{e)`n5yPsU-zh+IFqGCV9NcqS&HS=gk4>w19l{M&stex#1g5n=D#&
zIBc=i&2!g=CI&b=FFKoRACeQ3X>HaxBuL}8w{<`HBs}}9evkd$^bL7eYf?&?b-8g}
zPt|M$=H5N2ox)`e&JWfaEL*ly`|3(}v*~fn)F6%F-kZyeI}UEqINfD<W6IKP!G0Fz
z?R6WD$e$&p_$Mk`C}sZBeAMcJ?Jw3QN8c(<w=|o-AbiEk&1#lMCk3ZS-^}zq9(`ze
z)-{K<t1m>Xb<tdOdg3!K|9Hy@9W~`k%Uh<%!rn~PnpYU+;WVjStE|PQcl&3r3o_=F
z%cO>{W*VLtV)kv_=6wqv&p&T9dXfDsW8}S0pq(_l_e9?%cB8ge^gn{wYcAe`@XaAh
z_CtX{V?t(9^Oydcl%U2)0T(U`HyCIHLf!}Qe?|o<_j2S%GnJq^FWw;9U@*{NvCWn9
zNbT!CHi@B-7assR2LLyCQK+;89qawUd5IDn!K;LbLW~9)EOuZoTy^+z!DDr(%!^aO
zs8pagc)=<jPdsbCnXXZ*1c&e{(LfXpv<EL(#do}^|Nb5=hoNwomVjF@Y=QMaBN)!`
zLlgiG_-<XTUU&N7?_xNd?;`=|B=9f#{(}3M-)gz}p@vhc7#i>@%XOAZ1{y409XdlL
zTCe-85@dO?42WbvYw&_qLiOR{XE)Tkl%OiF5+n*T9B8mOye1^>%}&Q^F;wEki9k0I
zSPWjU%8q}E%^ItRH7G$fUS*SLlm0-1#oy-ow4A=DEno$s`C;m)>NpQHg5d}Vj6UIq
znK<T*r4-oti=l|W{~*vk$Uox!1^00<B_Kk#*fmxO4&_y%L{T~e4Ho~t{df<0VQr-Z
zHF@zKAld_T1}|8p#$&Fhbd|ZY5)|_)TSQw%4K!H%^s5UOgG$7{Vra;V*Xpdb9B8ok
z#ihFX^aa;%DnS)qyh^mnc%Z@Jb<38ABYlAYt&4PWD1GEQl;zAt=4y(ObD6*-6Ifug
zVYwn;Sp>!tE(MmQH0MBR4zwroTqw;YthAysztuAS9AJ5i;{Q2dV^fywPFZzO+MvdA
z>owME<88u|^?F-E8#Oj+<2}N18?86m5{j4<;fWWN<#&bt0bqhR49gpzegZJTcFA(B
zK&ut#lJ{%{Hm$&ca2s%H1Ct3q1zJyO{ind@DXsq)jem?BiGC{@-%4{EvTH*wO_Ap9
zYK7C^uv`Z+?m%OReg~S-f#zTyF|u}w<Yr^Ah~Z+OTMTr^0*1?iZdo)JLAN5Bd*JLI
zsOX5{;?UGMG>zP!jhwQPGvOTMnu8`0mLn%QEx(LhFVk`v<0@km`gm!eXuywQ$G~!X
zfpjmJhRu6qU5>E#)Gy<d87>w&#6rbHG2ChBavDw~&p!hl&d~B-q06td{1SAyM9brm
zOFW|P&qgz|kss+V7un{b354Znrks`+BimwH-i8*np~ZNi7=EE#JLvK6XB_r3iuczZ
z;aoB4<mIcpLEi;tcYz&#kp2_Xyx|-Fa^(d&0v*wScjp^J*YgEk_J4>zx{1L1kT-<>
zutf|Z(jw9gyD+mT`yXVNBmqNM?j2Bl2PAm5Sne29Jq9VxfvPz)pFjpD5VarABg6A(
z3clZRX|Izmj4FW)_ZB+7g_M1#F#0Ks0alb~=hXVo|GS!BmL6!^0~I@7((#wE`p+hB
zKM+VC9XT^<(bYA8;qS2NfePP;eT$}OFa(n^Hv{Kpps-1ni--2{#AbPC1nm=Ou7=Ll
z#Ku{!8cwQ)iovnmE$DoUmfwPtZqf1v=+FS2@EBRH0lG9mSHeliB?(cSid<4@u1A*j
z$eQTaBWXRdA>4$lnh>=(O=w&bQcMKPr7;d^4E6jp#wCrRo|nNmWYF>q#wCNuZ3a^o
zMl_JBkZBBa4x59+8_VqgRy)8{98X#98{qg2n1us~tyS}%b6LqDe*MCPVfy$sdy^%C
z#UmDO;8(8=d9|VGcw^g7NuF}<t-**7mKBd0hz9c_{60==cCR?`WOtZ0%jJR5d9?B7
z193hu!FRG;ArKeRTm-~LG|PclPIEC37t>q{#HBQs0dX14<v?6cb0rX0(p&|^RW#QC
zaShG2KwL}nZ6Ll)b2o7520mmwJz!D~@FkoKJ(5X`VYw9Oo&skRc^dRggZ_kN$VrAi
zBb<OHCZNv=m!b)!$dhmxvMWPV2`4hniHtYlWX36(QS|4m;&Nr7r}1V!cL}!)_r)QJ
z<u+Ju@Wp<K<s#G~#*_TuF_1h4vq+b1fNcZR7H9{Ob|Rm=x6<)hQ+OKNzxHe%&?h%`
z1M_a6@Hf0$0Tw-gvi)QzO(yZVf374c(2d-m2FImA55h8JDx(vJ1T-cAxe$3NGAu>5
zgv*do8Hrykm&lkU(t9nLF-<1%%H40&@vX0f?jn|J0=7+ndT<QXjDg0a<1DC|MRO6<
zETXv>susi1L_ZFVj6-IGvk{w(IKpzo%4x1Z(hAz6%NSJ|qv#LsYa|*xd*E%za=nP{
zMWaZ^eTeNNJoRR()8Tm@;p%+uC|s{jtYM>YqdH{`wh^wx8d?F{3O*rSwgE{SPy~6F
zdkWa6v^*YK#zRG{?|0~y2{czjX*KPTZ$arTnj4^11Idb6E(wiGB1sp^rJ`}EG}j};
zdPHU6O~|MTS(5i{L9Q*x3;VWyhisL`P+52eW0gVJ%QMX2i$gG%_mz2ITpmzZlaF1(
znolg__wxc7aFhWLd<yfKVW(B@acCxFxme^EOA=L<i(~j>xWZT9RpPy8|FJOp$XNNZ
z!UK6Y@8+|E_oqOt$8z_9^L-L4{X=r>yj9Wz_y_et%|56|-1q|n!$vb!c+7L(9g;q}
zxO4o!K0^+4KCF>Gq&4(|K!Xp!yoL9le#wjk>jvB4V;jqb>W13j2)U#(Z%Sy3HBx4|
zd|;Xn#^TxN4>_iVG#3HWBAVsER8Dg-FfFFJ6quIMTn0?bXf6k)<uq3U(@L7FfN2%Y
zHNdon=2~D{OY?1DdYkal6=yH|zdZSnU-n?NV2&ESS}2D{kBxDNXs{4KrYsuSMI*%_
ziRJd8k^2x8kYgC@7>07v35;q2qd5Jo{%(n>d-L>7d|G-1T3w+tv#ZeRD$PAmvxjzi
z*U^~kwD+w-W2y+RE;(g>IPK+^A7mf6bit=Q{dEg6kg(BV_BeN8x$RclJxRi|3pnlq
ze%LYM@`g|}Skxe$#6X)EIz7ySHd!<mL7O6)<B(k(9e%UX*laok$<bIj%`(PX#wgO;
z9WtFCL;RzRd8hFl8a{`LX~*uBH}L+_3VFXskH~rWNWeP~AdLf__@o#&9qsCn{bN6Z
z<wi53l}I7NfH9!RkTIkv5r`83hi8Q4OEBVtfWy@p{%`^GkAY9gW=RM7=`?2m{S2Dx
zfl56P6aBkD<u1()K&64^KY_}hG*>{?3aCZy&p;{}i0ZEvv9(A8%j1HZN7a0OH;p&`
z7f9~~qGE6t((59;uQfMp*5r2C|GAywS?0ZVf0YY+?E^PjJHNgddFBK2#WTyV?E7jL
zS=$U`o`I;yQ-M4xkRmn2#>~fo{!$pZ?KxC^PWDyO@{-3BwG9#C>m*RK0~+F~V!5|S
z^)1pP{5N9%Mmm^N)5hN4TC&0JUmq3Hx5de>tJrybFtu3`7l-(Uq<iu%KFB1<5<3ds
zFWLsX5Wg7g5$r=K)Hc)wdr3ZGc!c^8dI%&B0kt44K+;0<BOrN1^A0H8L5{^NcM?iZ
z(wqmSc{J}rcDv}a<Zi}#H={_6veUhR=XYoNQ;jIyC`-Kl7uvU6(AwLT`O%d@@|XX0
z-CcC8nICT|U{Zk!F8mY(OxW*K+MvS=B88E9*!^8}Hc)%G`_;Z-ELRHIQb^^X?T~G!
z`6Xmu(p-szl}H_1KA%V-_CBKSe}GgUke$JDuaWRI(Jv~$^~GtY|FnE`D^Oy(O=@?3
z`iT#AR>|?nd-N{3S@A2-2#p${A|Cv9qo%U5t^4x7-f7?%$-CY6UkZr-e1X^(^x>)t
zv0a3#oj={`H-k(1=ob5rfK?-3Z%otY*Yq3^o&)N5yW#!C2b%sbCS;--p=u*klmqW{
zr}_mQ+8g_?U;2AP3IaCGs6NaO^ndG?O}2sezdqLj`Q#x6c*X!PoFDR4N8l0<CSjv#
zwy$>U$r=ehIvZ%8_Ik#lmEXToV!3WGqZ`a2F0lvr^?<pAlVLzIoKH9f`lrA^!f9}B
z8eBkFhJ0j*DiS3i-vs1O<fUj@DP6}ZLtbTQ29YN+pC>Yk<cH;w8J}cE(I3w9c^mEj
zY)5Rg39wD1#MJRtZ`G3<{ze0bK4Cu4B>fixVIj>$Kv+bx90=t!7Xx82&80wCN^=<y
zmeE`egyl3>0%0Z1RX|upa}5yI&|C|IwKU%b!rO##{=&Pn{_m1x{F!J<CZd+50J#+)
z#j<s!7Rar$4U9iJL+)y8-rI7u`PBdWBIBzVuYBWcxi9$;{J)<}-h`j>>AQ8OhxcN8
zfbJe3!OrQ;t?lD$x^gb@f_h+9PwGMMtiOpnB=J7`f!}-g?xfuM&!leN@5Lg!SVUE)
z;~48WM)Ae7H!v^NW&9_7ykalX?L|gp`}85*KEi@^%T_P$k00Udz6zh!|3jz0dH<)2
qe+UHp*K+))hkpnJqZB_O_iFAGzuI^H*+ccUb@X&c3A7b5!G8f~bpH(i

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_bwd_data.hpp.43B5DE0C3D25BB20.idx b/.cache/clangd/index/gridwise_normalization_bwd_data.hpp.43B5DE0C3D25BB20.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4e368f0358218880dd43c25332e447da7552493f
GIT binary patch
literal 12016
zcmeHNdsL0t-#>frCvtk^py<ibA?c#a>Ecw9kuHYr7ez#p%PT}O#_gS<Qpq7oO(`Zc
zWsI1ZD4Cjx$~gSoZz&^1(oC5YCS`ulIlr~f-t&)ly~|qfJAZW6YIi>8vv1FS_Wthg
z{(N`f%<0pgTL}bii)U_FwPalcBM=CN@_(`G)-E?h0zsirAlQ&w671vGWG%Xy9o~or
zULN5(ufiT0<<$jOuMfQu73`ZC>G$yHj^PiqS551E^I2|MM~FDnVE^Jr=N%%i#V(rv
zd*x^1lHFU=1^$1P`RQ<NwQua}zHEDTDWB<D``Ba<41|8MYcvyvyj9F74y*}17$X@F
za6FIsdZhk@dq0}q)-`C0&)<+NVb`~A$o*MpzOE%{olWP0pntuw_(Xh7L$to5HgMae
zVNNEYH5W4FzQK{rn|0QnpBH@e@sDiB)KsQVYu5I|9U5kv>JpwcpJor<E^8UQ(D%nn
z{&odUeQUj3tpe^>tjv-fX#eHJaqr_f9Y+j~&(hr+J$!%q=G~(bKhd`Yhd=dg?5jB!
zwI+Jh9iMs08NG9p?-k7zm|Izn>8sJ#SbZkoYUakH?U%&gZ7!}+goQ-P#w=-()hH5X
z#=c+dwdrca!@SFZ`twVMpFI9ubAHVRPt)?28rxTMHVvB|^OCVBSlDgS`nsX|mYLDo
zv-{%qrN(XhzUSKHpCn(Du6wQLJ#>hN%a)D|Q)%=4@#h8%Wlw6~KesDn)kH6Qw_h7w
zPFgH~qv2IuHzR9=b?8Rh&3n%Mk$F0)D|P0&T^rU}G`W}`x-je0@RG$<WmbQT+>~)e
ztGRp5#p|9L^~T%$(uTKh3vsOP+O>O~n`K|=a>;0e!|@BW9#sEpaC+3Ufb8cZHhc!>
zEuU_@Q@ElpQNFt(Fx||hI&MYW6821_(=Xe+jr+<SbIQM&^ulq@Lfsc%u2{2Zu+!bV
z6;l>A?J91`v0FU&yKC7dpWp6v%fG!w@7lA}z1=#o-(~e`z4gAjq%-k*&w+g*npMm8
zMD5#Sxa2Fp)~c8*zZQlR#TIUwzkAigSJr~fR)faf++2KQ%^*13x3{D#->2K$df&xv
zAaz;ua#d>3)TIk=Z7<rY@PBL#BU0|4Ha*j@_}WUl1+h8>HVzKM$|WWp8w@<cgAKN5
zZv8IR-1o&(v+&T31#cRxejMO^X`=te$rjwwP5R*vcG>zq2=gA2GckM1xbJVSPp@<T
z>)Y*fHdYkgU4Lpq?iSIm=(*zZniE;$zplx%NggSQ*uVeL#iB3sn;%?>-MDH|G!zKf
z<6A=7eYfq>f@p+*ixftRRf#~8UrqeS$gHncntHxg3l8P!P+{m;Rf1`={9Wa>=5Kt5
zK@m^KjEXT(C75=4^+@!EZ5pEmhw*f<FnF9Q!L*mj)1PnN^DGzuHFX7?bfDBil?Vnw
zFhC#x90+=w_h4^(lA{*X;LCdoJ&jcfF2Btv<+pdaKjdjahNlCCfySx?)7k4<UW%?c
zJLp0oPp{BlAyy@rKC~(>(PNssl@=Vr(-Fdmk*WmKg$DDRrfIBtpbNEldZqqKu`0pz
zv5{fjpI7(w=)%!FJx_n0wJO1Mi5>TLn0AF!3l8V$8NwNcssz*Jvu>^*HZEPP3t66C
zJ$ki;D#3K+IrqAo>lSEgK^>m<7J83YC74zmKJfU$pt59LIEbgC^rOV81k(@XUeS-U
zj<{*TfjsRi^fgl@nEv&zg{bdndZI2A^YjA!1y-sA)6L@7S>o7^xw_DRr$h8ZEL91n
ze?R)m6t8IpeOgePr~QTgR;mQko%b%J|5?Y})`f$4I(l@pg(|`H+9gY4n@t4*4ck8o
z*JLld2pKL8jfg{fcp@>JoY9vvVvMzxDN3JP%1;B9i<Cq<<JpS;YY0`d6;C9VYlWg#
zXoyR*TpJX%(YP9ksu9K4B2g{DgRPVIM`hSN)MmM`;b9zp$2^jT$p{K}2%mySk>#Q!
zQRDF&VYwBO731;5W4V=*mE-Y5V!69e=PuMY*LWNpRhu}F<?5hb9W+S4d}rIMnUCZw
z_Y&&7r1>tW*G2QiNUs<fkn$x+T!IWq`5vU#L-X&E_&wo==-qa;zV&E@kl}8DskeX&
z9wml*0;WEpaVK!=1nz{7gKlw<!fDVgjmG)VEuY4TC@2xlBIT2iPZIJacn>ps4>O10
zy-dJfCJ<x2d&Qc)g=cLI8SXZa-3IoUV7NQL<_@qWxCzLbXucWPG}HVmFzFR=BKf<3
ztc%Vc4=2S#icf&D1W2`)4kxA4{5~k#NAusnN#D?X0hATcd_A054?Riy4bZy*`Via*
zZ5yE-!9CEW2TsFpl;QfIV;`JE_&ylg2j>#J6`5~E79@WhvWY{~_#`0D1ms2XZ$|;!
zk$RvRE(zHtA*%he$mcBmqDql}DGDIv&m;Tu$bsMl#yx@YAnhkH&PmJ^!Y48Flb8hr
zr!dkK#+u}pGcq|tjrVTGcQ@ll@~1MvsZ0pLdl~z^jJkbpi1RNl1wU2u#-S3JRsuC)
zIQ56JPRHxQ{u$IigFZNeuzNF5sB&w_K5_-PT><Ka#&QpV>q8KVCnCSE0LwOFAzAJn
zuzW}3ozQY8v4|{}3#GYmIxcVe-;Q6WR#?y0V7W`c@Dgyv1k2q6WAA}!1h)XT1y~UL
z2C#2voCMh<sOE7#AXNEb_<3xe9Xv5=VfAW;;od{X_q0WcLxFK92$y5HWX3s}UcquE
zM9$2{yv(YxehzD$#jm*8nzJR?S8yD^NDwh3{4>7(9=M<fhLaV(6`5{DW&|fNixQYf
zf>Ria6vh%`N7t`hPxp)+^3ff|o+lEjJWf0ie7Buv2jgzD+|nsaL$EI~d3Ex#Z(5UM
zhOk^b7!wby@k6s*BAA>A)Mk|Bb^xaxKyB7pP7WO9U<S#b3Z|q2wV7wRG~kj3{0W}{
zWEsGN;A}7<n`~1ocMzBz1ZuO-az}vW5ui4!EO!)G90ks#d=ZcqfvE(a0uxTr^_2py
z6qpgd9B}0XJ8Mt57(U%-BV@UGcJo3gKSwj%9Q!%bQWQ?@zs4Tm-IjR3#sdkdEfJU`
z(yi?P#yjZV%Yl)c*k+bX1!k#q@6&)q8r}O0Fg62B#Pz!ej>_NiTGM?H%T)nc6+LXX
z0DB9lT}766W7sD&?gXMvFa}q|a&b@;2PvEeMQJq7hoXEM??d+ckb2RwoC4V^2>$d|
z?&O@g9<dzn9%_aSqlV2Q+#U}b%ay@#Wzdpzq7oWbk_F3h)zG*aQgclK$19+e<i7z;
zZ$R~yu5=<@?m|!E1lGanb<iI#Af*%O@{;De;Pft<mm?!NvLf~8Ad?(qL+~MFd<aqf
zDMS+rk@{`0TrqMlM!ux|669HeW|8(AkVykwUlTHJqU&o#6I$u|I*?fhaY0zF8(DNC
zwM)QqJ;=R>Ztp$vd{4KR$(Uv`>hYO5J!WiZ7CglpsCZx-4`lekum#{Pqw=8n{Xz4}
zCFk~@@o`97J;GAHhtD736ybr-gg$-_BQJSm1@J%TLC-wmvf&dcZ&#GZQCus_1seyu
zVxz-ySApAA5RTgpSyk-dt$lblU)xuZ{R*l_f6lusn`h<m6aVopd6;v~@WUF<qd(d%
zFH)RwjeaUcnk<(MoU_THLuvFTAEeE|5is=#aKmG$G<u&LrFY6AFs+ET2B(1YDZ0K=
zU|tHSg;WmA%L!h*?Bq`&-38737VI_DON`$u%lQoT5##eW%lQrU6XWwV%LNP#5aR<d
z%LNS$5@QKsW%?o$_eRgp5zBoGjlZQAY!Q?cL38pPC!pj6!7Hl@&2kQO&->`D6ZZrP
zRU!zPnFpZs090>YJ3Udt@2P(|@^(C0H(HF(>nyigceNPH5-j(*<a1-v<qDu(0dxtj
z1lpA}&P1Y2G?MW5k?20byPltl37XcH`oA8Vy*Hn~+*ZE%#7B=!=F+N`PdG!wuv`n&
zX`uyy7f|#98j!AEKspzQAi#1pNK}K=JU-d;gRIn($RJk(=W6f?t|s?@`I9YGpE>;F
z*EP0xyg5?}VB~pnpfm?k9!4P)6~fW@`3`@vGxdtLgU3g|7_5!(<Ed2GNc+d2<uS4g
z;=NU0s1n&oHU;ojfJL}lyrKfU>Ocg+FM$0E;7xE3u<xO93bapw6rT+pv!Pn3V7XE_
zu9SGoELQ;yE1)B(?;<q5NDtFxIQ}x6O!({2^g2{e3#H;Cse>}IS$0857mar!qn${t
zN3h&}WU?Poj$a-!&O_8|DL@knkXp!Lxe_$41Wh9K*CUg9y8IJl{Ddz598Gvmmv2XA
z?TB(DyO2c}qE098(YW_?{TYmD2BUty;~UB~{8t^G&a&K$@iQ#32FhDlTQ4VKVZB_v
zJPGo#@tH*It<N-{nFRUC{H74=>*wL;Lr{QIfGe@Z0iFSY1O?d!IbnOq+iBMzPl7^?
zLM`yI5BHFt{z{3K%&XaBW;>Eo!jdUVg7GP#q-|5l$dt}w{nA3ewD3V`p>*lvEAGO}
zjl9zr!^CL+C!8V>lsk_d+MN7)mln$%hnmOXFk&ojL-sZtLGXPjx)1dUehS&AG~W){
zb{cm=Q74`MBoduORQc0L=QP>MSng-U{*0*h-XhUkI)5+H=_UEGZ@{k=WxRr{m3_!_
zAEKlY1@clLwRCd+(w)$v$*-RFub2My^1<~|xwff2tap<??^D!6i+XZMVz|f9;xUcS
zBI~oX%3Y4E%L&HY2R~7jkux$;>w#-Mm`~QzZ@}_5qS{oMd|~78gXcFN?bFoWyrEFM
z==$j&@BCd}?(%0MGxaU}pMr2Wx;0j)iZhZCy$GBy60eu#jzPC$kkYlvpj8>IZInUV
zGDv9&mC&-1czL{v4{a(*I7F$9*i`==%c`MT<>v(%Xsw`+S_(K(LBcOAcLUnpfNBXq
zi5+gx?wlM+<+N5IM|N^V<<CL3IkZ-hgY0t<<^CN)Hiu~K;t-M@BH<C1D@3xwzvD@T
zh;k1bkZl89e*>~_pzCi!HcfQ>O-R;6*Vl?<t$)XpTIu>b(8LZ}OX)z49khGdjU2iW
zr89P;$=$TOn90~>GSqlvG7g!HI{!6i<1K-H;>>>2-f!AJXxgvOj?#O({(yvcg|-6C
ztzZ~-hVFOe2L4x8O42{xRN`X?uO2GzQ(z1Dt1Rrvw$o1`|6t2Jmd(It1T6mYw-c1t
zD{yY5c01}WloIXqB^>pV#>GfdjO+;CgCso!H#%N(dvjb5^sl%6_4dK_hEJ}%s-nEV
zL*`!<V5>;vrftSszpUliuLmdtTt10F9R*YdxV(1&^*sS)fXh=(2e49sUn&SDu2dQb
zOatl*9W0jtKFuHjFnloQJu>BmA<~~-VA%`Q*9-V&prstD@$WqbQLQ&a#;LU8(b*(m
zrPPYW2Z1SR{|Fd!gpP_H1(KtH(p-yxVG&Tv?ksl-j6OxzR|?otphxN}2W&aPow;R|
z((lI_jqAUV@=w2zas|!2g4B^&hWr1$pn}zP{(6D(Djb>kw}A8(?KnRH(kC?T1a_Tt
ztRoKE#X(AkOM`Z4G|q>1`2=^qXiRxp-B|d49;<zwIJ4B~@~LkA4E1sM4BpzsT$pj9
z#^s|AgMZ%na^j87;tGC;t^lSLWQ^avE2xy#XLj*K9+c)mb@aYB>`d;Ii6h?gv4U1;
z)CwmMtI!6G+Gt#jjH(gE*CL}@g4Zrt8KHD^V%J8-PF8*ysZ9R!;h51obLTv{Tgi;l
S*BdF&$G-~%qsUi1^7Y@yO_!?x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_bwd_gamma_beta.hpp.8028572B7299488D.idx b/.cache/clangd/index/gridwise_normalization_bwd_gamma_beta.hpp.8028572B7299488D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c1995295652892588439078b1bbdcaafbe38fc30
GIT binary patch
literal 6944
zcmdT|dsvKV8=w0r$@Hom#@k^Ooo$WjG&;>}sircO5=CuoD`FA0lWZ#+2@!2YDrxAG
zEQ_3`HigS4BqXVvR-122&WBJI`QDlDdV1dL^2h$=yDoEG-OX=)&-=X3;l6+O{X8L4
ze0=l_IUKLKQ<g7|O`5IF;c!MUe~C#63$`H+=ba{pv%KW^jHsa7wxeJ7u9O-rPtUr%
zW$9OIy?<R3{GiaP_+oz7t*Qg9XS3oe%6ER~GoPBZLua2Pd)pqD`*(8g{1lh;dvDu=
z;$)QPqx(w!?)ue=$FG}KR`xFIDU8o_PB^Q(w>G~qx06?Yc>0~Q)A9`VUbX!8@;3(3
zV`oZJ<8%VHiWC1jFtS#sICs4Mtf#e`j$V?^^wSoeRj;$1w=s9Q;7GEzET-RarpMdB
z!i8!VC%Bf!JYRLMvUJX9uEu8fee?IeIB(wZv(bb%hJ{(7(cVLTx^}MS&z3wRr#Vd#
zest>7zb97RxD<alFx0{=DdSGujO{gN1CPvFB#qCU_|1xHV@E;H?7+Ss=U!fJH2rl+
zQ2(;tA37Esc)s%L^KBW$R}vyyrf;ggCfKxh%$|ok!)gs|m$c`&|5D`^YaP9)@_}|v
zT>m}p%~c22tes{cK|%UE2NGX=s7rFqtQ`=|+4(x--p=5YTOtP>!bS|Qw-%Wc)bXm4
zd*7Qy=#I)CQsd^4GS+dN;!M8g6RYORfp1qf7nGga+onIZ;^<qi)kf2^R=kj0C@NZH
zALrY^361>W!04?GR`xUX+~?Mok5%+6d(z~tW7^UiuC3vYVk-iY44%A9elEI^Htgh(
zX_MC0J(gZDZ(jQ>tux=I_TXKwmbfKR*)PPKIZdLj85_-~-s4$$nh$S}`2Nz7f9WrB
z9k$BGV~XbCUu*(gzEhmm*u9yTH7ZyUCH4;yH>9jye1LD-GUJwealqXVFVafQ<e^5M
z>phD5EE^uVG|UexjWs-bEn1!#r^YJ`2rIe4%Uj_fznOpd%58C+rD4$(p;?UmyN8#y
zeEc=@`ugBPQTNTdfR`uQy9FEavyQuWz7_C~uf93C<MeON-eY`!hGS(x37o_q7suE#
z{>imX&PbCCwDZu=Q5-tLAcC(l9L8;xp9yBx+j>TCv(iH94DG^ou}~RITW61&d*|E6
zC0dBf(C%FKi7JC>fyFKB6};A`BN4^W(T34RDuZcz`5IHrkLhwPq|4AwTqhHi!StlB
zhBaQNDkMC_XJ~H&Z-L5S+AHhm?(9P+Zt~DDhK?K?X`?cj4hpU|zwIaCYoQSgE#`{N
zRR+^B-P8USI<-`3AsvSH<a(N_45k;|yWSP7$nWMMLxz?a$ZS;x(`)_GEpr<keYDVU
zhW6w78LJGY3uHsL-V%Q-;~`Ckjv5<fqcWH-_K-z(FZ%uuEu_WJ!Q5b4Wib6~X@l=p
zpYN)*kQzgWa6|YigXuQJ-yONYCyj?R7&_b_oUbyN?g;8MdiA0BffmwcXkV`Hc$LBQ
zGwsAS^Ol|*9x`C)K!d=EDue0QwC<4Y>tZfzpdlkUwCPY&Tb1DqLok}d0U8ow=Oy0e
z5x+rPmA~NUD0(l5_rlkB=e6cHR`&?9j{-%97>9`Q)P~!xFJ8FS?)FfM_8;Z1k5`*P
zLVpoKQ=O(t@PH{g+&J7EPoAP9j3dmk^HQ_|jaDFo+sh}$o?dZ(DMi;Hy&7axShIN9
z{7E6(b@Z-#+P2(Xq@hkLz+C|zczEh`J-FAiaVvPWf;Zv2z^jXm`@pM@jkAzf78-1?
z*M3;@eMOZ$rA}9Ys0t>Lj;g_~8te%^2cmOq{yf;7XY*~~(gv=ieLINS+3mZ*rJK#)
z2hn{t-w&dGHr{|l8xU(eb;!OBIgtMA(DXVq17m$1af8R6)Ah{0m4Q_mIAOxX{n+m>
zov(7g_!(NIIm>fS+Rjv`=>1@_A3X648RvoVH87dri$Glj8-i~Eb&Fsuk!oC(G?Ga0
zwD9!AGDFdo;8h9HxDza#%tK1mBo8%#pb4hoW)$54f)0Z5@4!_FDSk{a_r2R#ly(Nx
z=_|<i3L1~wVTEO0qm)<No}#7Wq|R7hDSC!vMko$V(ZtR>vnbm_X6L1vQe)f*|7YS|
zsl=qGa-hn=7{AcjJcct{CpwuK$w_2?k~o;#()IJpAFPibPSNq>;%#wzicWG%@+T<8
zHAO~Hsw6d#pfr!PV1m-YB^{;^oB_@m;79N}_-Y-A4~ou)3EAL6_*}5b1u?+|AS?iP
zf=fWZ1SSwH2YorgVr|LUXdkm55k<=!WMM4H(wu6IC2R7AM%Skui8~mTi>KlT*NB<3
zWzt||Q*??&3Ts@c6H-~DO0!90jVT?>)4`dHDFe(hNND1OIFzeOgk*~pAXE@ZrWnl=
z*0XUdIJC0;zY83?*ticI`q(%NIb@;1c<~M~+`ZzBrne?VmxHJrq-1;wpcKF-n30ou
zHf{yoRv1V4F3{~_<37;sW8*BOn}t~I^N@WW8HjI7iQD!FX(E?NBbxprS$kqLbrRka
ziY`K?MMywiDn}-A#ER~6WL}Phgx`;>_M<`JR;C&c1(LB-jZX~<#;Jy)n~_yByL~&d
zY)5vuzbQWP<HHNl5hlFi?c+uG=Rl{trDFuyzc4=sy9P^fg%B9zL|>2=PGj7x7i7J_
ziiS)i%S3}1S3692DoY)*_ou+~6hz_Ige^Yo6rf!;lj-jvP!C}uCMfznQ11zz^{&va
z`$UH0XU~{1jS9~tT|zYtie3lebtDxlBi1dO9g(^4buQZt3t(~q+uKUOx&&AoCkJae
z!B~kIDO6T_WQC++X((1ViVl^8hG8k3JO9w>u-AM3Qln_s5w3hJ^b{=_A>rc{r)ZB6
z9(=6l*y*@jRlP{wau5OzLJV2%I`FT9*#uvPNtYpj;6Gr}A8cHRObZc9CQ6Y>DPl>%
zZe+e2xsm7hA*+38unbccM8-8pguOH_KgQC#Q<lQ`!DeK-neFa5$TWxGxHA<Kw(ofT
z%_n+}D_~|hmE|v)=ra&LBLxyguR_9Aq(IszQ(N;S=c6;T-I2UVK9($sj^ah}aoJ4K
zvBt3$WG=fvdl&EsE(7f{Hr|4$Eoc<xH}xEepFa6f&OgpA_gc^MRVRKr_{rSxIGA8k
z*6w6(dB`UZh2j?#%sX@cYkd<n=0)d`?s;T{35vdmsEfz|V;s#xxT-pz^!F5mPuV%>
zE$F_5vG_5HPDjFY#L5O6k?uw`$d~M}zPI8`veRde(Xy2rb1rE+Nk3UA94L$vD32**
zVhxCDV6R6bq8bVQZTX2I!Hdg$v?$ui+)03o6eh~-U0q3(xq7<F2$I-ITu97GCQGIe
z<l*Y!i6fASN0~<mLB3|bHuy|L(SBxrHrRD3`Z)-nll)21tC4Uu8Z32>KT17LZ+N-y
zi!XJn=lyxfn|ra7NiNaq(c1qz*>gB2#N$W&wf@Z}rUKrCG&iA<#PN<H>KGbBa5d7c
zCjQJ+5J<D0w6DF;8h*g7x9f{vs8b35c^dgXH5jSWwaB)XWN~%+G!mXBS^SLUN;{`x
zvhq(N;{3Jd#>R)QPJi~8Y!X|n2npgBeDQ03b>>37{$)Fi8RuN0xnv}E&cTB|S9QBV
zq@@DH73|`*2u&<Ptn)+>vM)l8#C7FJASd~ZqUFd=PL3r^&5CTx$wn)&V>vk|GlxUu
zupbT9X-e$0pPh)CkwY_^Z%6j+<o!2Zm4tj>xMAIA_lFKq<h@l#yAZRCVbU-^tjt*X
znAD@Z3L?vR2)qtK1X-q&;CT|l3BCe?E9^9W7X)_+?sRVQ>ibO(zIg9@<kufmxS;3_
zQ;#GCCQT>vPYFl~At*H{bujHybXrK-EW)LOe>yp#Q*;K%GGOqa`s8n;g%7TUnX2mB
zvFpeQQd!^fv%!jVlnchWU`=oVm=utToT5vBDgiyh%Yl*;j1M>(T-Ajh+3CmN{Fq%y
z4uJCj8!M2Z0tv}^<|?wf$}T$EkYyXGEMJ!Veo9z5-t3Ft`IiqnH+0o^R{Wn1=XgJu
zn;Ye=YGtYQso0uaMZZJ*zx@uL^Z3V%^S1S8$L5!ZPcren3#@h#1$_5z&ne-Zf>%rx
z+6%&7_7zAb5@w>oE0On+iV}(Am;t6?0~*(WSOs<?8rMiLCm}X&wo-=^6XqnkDZeoH
gz4)Vtzi>Eoa(72fd!L0=U0-hukE2hRG34ui0YU!5T>t<8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_naive_variance.hpp.9BD4CEF549374198.idx b/.cache/clangd/index/gridwise_normalization_naive_variance.hpp.9BD4CEF549374198.idx
new file mode 100755
index 0000000000000000000000000000000000000000..825ae9834bc7cc5ac4f4cf9c7a12cad70311f14b
GIT binary patch
literal 12224
zcmeHNdsvOx{@?3@?A~Ehk>2hXblbbz?xsZdqKOV=P7Jz8W6T^f2Q#LlLye*+qZ);#
zhDM}9DW;<ksYxhP#<&y`O)?=Q=J)P>p0(GSe}3l=&v@n!J&*P5kI#3#>wVYy-qv@m
zkN{ubN|{LH6&0{%dGxBqSR@jS5&p%kS{eHj6p3=gBGH=M>bVP+U9-`=68zg~eMhyp
z?eSk7y670W(W>eD8L_)JM%*+tu81;qSr9&WGo~>u`f9_Xix2CPFH}8yF!+nfgazw9
z?;U+5PQ1nRmV4SJ<7Y1(6yN*ub3N_O5dmW((!WajVR&?$Us6u^nKC1IIMH0<xa8V2
zpU~Fi<l`LyJIzbGTYR7P*_zh0YV5F?F<s*0(lR|J^~SJNUqdgqI*q1XC!Cv}j86L6
z=Mit&=i1dVF;8}P_pA7znSE{0f0af|NP^9n4XzP~Cv}!CFxeEVmgMnK*wWw5$$!`0
z^ayy?)oq$+wzktS=s|4LgMfdwG!NkHvG;s;uDh{Iya72a9IvfC@}9~5zh6ZoVnb3_
zT3=h`wKYloLZoaB^pCOHcx=Y1OHWrf$AzCra9VJE{k>5?_TKfK)_AsQrF4YTiDlVs
zBcJ_{7WrQ5)1r>!(%w5)UiU<~Xg6+`wvXz2{ttub!t&~4UzKgQI$5=7%8*RIXKlp^
zX?kdN_J`ZdR!gFd@*<pvt!(u1j2O(1`FuwAIQ>m|ropGr54YUyap}1z_2p8xX&MJD
zueZ$W`KVayJDr@0i1;z)kJmo>xP15CZfWHAxfedGnD6=9p|HhjAs4*ysr+~tzQ6p;
z-<(n=jf>AQ+kAZ5-7HHZweaajR@JRK@b%$4u9<1&7uQ8h`68^c`0g(+yFPG@kN@@l
z>}P>F>!x=sac<rEuWjzl^SX>jeR*@?V?0!DIwpNpPvOu=JaNmJ=8Qm(ADgYSqQ;;b
z>laC0ex0~tT!O(T>xRu-negIa>)-*0$K&hPubb_#aM7A?Y+w0x)?Jvquq}3Q_BF4y
zIlsoQDBCmgMg6X@s==wt0!xgpuOAi}x&?oF;G1<(!GU~(%X@Rb==E?|<al~+w$;Pn
zgZ=iVkI}&U)=?>E=eWO|{ra2Qy28#~Jvw@g$L>VmhHF-+&&lX4{bBg5^7-AC*ssng
zIM?uo;gq?zyGK9V&{rR~c6mgjpxq<6WHfK%ubNtLgtmx})``|tHX=2_0))TO<=d{$
z_4vk49gY#`>Eh`o%7)My!;c<Zkmhk_IK<;c{Nxdn?UapZBn0C`BESQkft5>IB4fj~
zprO#;Rmasv*^vHxSX2M!-K|U1;V6NgA)aBNYzVEpdavt07RxVd!C?X&sS~NIYzS@Y
z@Z`ws8&iH&hnfN%EDknOHiQnkJMH@qzwmC*f@%W2RBx%3vLSSsd}r<KpLR^tf+hm(
zt>bO2YzVy|shBspvp7#3z9-OA#Zyg{4WU0=7292LN5@MG8Vj_)j=znvA@q`;>_3j!
zo>{LB)dhO7c=32;L+JRYQeRg+b-6kmEzo{qKO<#B=uD^cTRr36|4|E$6zEvJSQ}+S
z=x@6|32$nfkgg6j1bVi3wt=!Cbi>l4HNTBm7N`Y{1bU{<OsTRVbR+tC<BR@>5_O0K
zdX9LGuCgKY<pbeCddGH&)S;F@hl|7YlntS;rtdas%7yk?P%O|fdNEeYhR_}1M@R2}
zZgP4o<OF&-zg%D05c=gU`_Z2Tv_02?!v#7@CrVe@5c<`v5%IbgBL!uY2=p?YWmd|D
z(7#tcUtQQHKd27J3bdQp%~;tGTC_5HNnDCK)oCbJ>M~JMjz0pejzB4SNA;M-%5Li|
zM*zpqF__~<=7rnfo-?#NbT#DoN6_vObR_Z|--{f3kt@OP*=*YR(GJ^dCLAB25io&h
zgn(oLa*BeSg51fHaQrNdSrdrf3P?6fPEp9T5MQ#k93N>AX+{iy<D(3s%*gC<e6&Hd
z8L?cBUnW~7CsW7q+X1&7j3qD0@wtG@WpN?k3R!#s$}T`Tm2ZYt&CrhGi_q#Ki?<<*
zZODf5=}4B2nEpGE#SWIwMAA&;K;^TLT^4etxCB|0u=1ryTFT0oA=@(KMCC6d`^(6M
z;ud7nf|x~WMK-N0-jAjGF*AM<wlBg=|3g@Mh~-PLT?uC1uM|6$VmCV8Su8!v%Adn_
z=UDlAEU(8?i2P{nR!^Ij8_UF)uLmCWU>YGX-wwRmfj7n7z^j|ZgTQN$#i`IM6*BUB
zpw}K2KY{^|U?6>7De^8wJ{0$&pkDMo#pkg9IUGPR@$^Wn^zhVZeB6BI5?{&jVeVm}
z#8QY15G%cY>2AxD(Z|y(wA;$mh6<0_X=467ke>&twZVJ?ux$V;-NO7;;BpnH3<~o<
z1NqOa{C(hZpXDC_`2&`J30z*X{1A{2v3weor%?mOd>)kNvA7v}HbXBm1(@%J-o22S
zvUD^*9W9`ICi2c?r?dt6wje*s_aN^cmQTdCiJ1A`MC_Z08Tlf-s0c5n{Y$Wa30;y2
zqutI=+gDR3IH+<kxg4mBfJ9YctrQ`ZtOLh;=y}+az{T;NdY<-VSxKZ7x+uad?UD#(
ziByX?J`2jSC?*bQf>;@OX}9CR?Kqf9-dy2YJ&(|C9A67<YoR^mYoWXrsy=|@&q4XQ
zKVp}2P-Qh7{{p(aU~vL+OF+zcdB`D;wemdVoJU<V$5$e|O4f^1BF9SB>og#F1M5{9
zkV^w%#(#=jp0fBAa(l&&pMm8WnE9Rz?2<vPp5t?{OU@s$dk*%X^HGJJtFTJ*Ilc<J
zRbl4)nz3Uu>s6YuYcusi#vVtj`kgO|-<cTyW+q00uaGJz@<*!Q^Bv~TAgR#gFP-;n
z{f|u;;`nM{R1K!nWo!am6R@KAI^eETY##8z5&NTGnF*;x1)N#|RnC{=kHM+OD7KCo
znxY54JOwfT1e!d7X5@*Oe}qgPAr+PeNBP+uSiN}!=G&k_8#JPV190L1oJ{d+X#N^n
zQe1!x3J@b-jSQ+;+>a*qBc}f#G9N^${<dAa5AloYJ{^wvJJ9(KR0UAX--o{UA>-8g
z;LJW4L<jCf&Yd)1WBwlUy~jGuUS!cr2a-!WewoMFWD8EuN6p88L>_0e2+`uv3qMnS
zxV(*Z;=NGX3uSb?bmWtcd?_wK@&f8AFkg-2)hzBo(jLV0FTw#uSQST|J+nNI3>uC8
z(?5L5r<-E--ruxh6vrn3!vw&jSQ~)D2AT<Qd=hX=0yD`Ma(puINCv?arvcM6z@$7`
zz%+}+`M@-v#Rq`a0ie<`jxPpU#VjrXS|uzl16pM)E`+XyH1pv2A~>~(D!1@F=w3`U
z!fllMnT5WlYca?B+xv$ynX+1tCC!wBt%GL~^W^xDF(D2lH=dr-==$(x+^!Ldnc@<F
z0sZm~z;FX#!cG#HoCFxRo(wFL*|1dzCl*p&R=l}I5p<`|D~9I9P?h|7hw$I68&LBe
zDa02buLP=0ffJHDt{&*pj_p9w4h$*o29j<T4+6;`i&LQ_6*BUBpkxn=3!q&A^rX+*
z3%R{eHJ)Ef?v$Lr`^1eGHXmvMW0@(n3UZ`VI!kTVSY}G)As%E(6#;u<Eo5}=9JDwG
z869eZUQIBl;8x4L*r=`DiU2Kr0hu&80of+70W%eOr6NW?51Hp7W@Qc{uY)XKi6&Jd
zTl#?w$fAJ_;Fpl!B{Z9k{}f4|vH|@SvVFylpMfnh*fc!{OLN$Ox)1y9!_4!luyqw1
z&`)E((=6YNEt|1wn-k!>!svr+cv$dK3BWD^sE8n^j8Df|Z`mdgi(MD{kUBt6u+#W@
zpX}K}2pIr@13={-1>=Q*$<#Q6%mD`GP)v##LV!|KFld0<3vKs8)ovnmd70D9QTd7x
z)eE>@pxWee{Cdc(r<-1m&wyM8#l*h~&sUT{=<_Q9S4k~v-pg#;t`i#^{_@+H`EEHH
z`NO(D`P1DsNwbHEm31LHai@UiDKMWZ@?~Is8K}Y|$9DqjP8L4~vd2KR4I`$jDOT2#
z=woZZj2fWYDscQo;CPXZ0k?qTEf)6y=RUwB95rxa4fSLke->JtrTWA1O;EQ<@G&a;
zPDQ$@h{1c1`5we1um_RuL6$#;CLKeJ=RA)r&Lfq6a{MJ^cnLYu=ci+fbj<KOv2-U^
z?QS`KA2!^F8Tm?VU5Qm&HI6@x4NtTD1#Ec%t1`2Q#tP3*?#hr5BGQ0k8r>rb8yw)6
z#o~P6n9t$^VA27g%6k;Sa8fagOTeTO7MFoZWfVunl+=Vi-FHVw`8>vW=n^dvHe+Kv
zb;)Locm+Xs6s1AB$Qz+-BfCgjp=2x6qq>v{WtlAhEtGsq`H#0pO!DpsR8+1G!^wwP
z@7)I{_fh=usY3IdT~EW`dNM`op2CMKik|etYk_Ai_<(-sC18Dt&3HP1bq9<4fvg{>
za=gufXk+iz-<{PsK1M4>m((aZeyP?{UE)DGK2|GMmy}jHKF%O+A|3E37<CkAQCtp2
zm9zK^lAJ-~DZd3vwqO<BysU%2^#0Gggp{Tgnzhojo8zxQvnwoaN9OG`+vNDG$owkB
zTl%YN`@9V9zV-Y*wfpdjzfJGi`ag9i{c`uCgcDyJdh5EAsUP{?ndBAsgg5;E>#}5t
zg^;Ewol_+%fW8G#rDT~)PTiY2!59f^)&Vs;pfO$CE-2}O`V`+pnl}-{-$9Z)NX6&w
zvg}-U>NBUmd~Dp^Q+x4<ZbRJvoI#Qm3NxrUU7$0#z+piE*%jx<Z{GUjC?}t{t_TO$
zoESE1dA_fZg}a%#S(8%%!Qsk1+^NI$@bL(u$W!j=PTa5HjD0+VDDt!Sn?gJj$ImdH
zVL_^8!d45)lGr*`Hk4&kT;B6}`S_InqIY$eH>|@{F1huix45B0NSl}A<)i+4_$U&c
za5EZnXVX9`FH}IGS~+BnHL4+34aZR%ZGe&ns6%lh)NK5N4MXlGE1!&nlTW693X-HC
zrvGN7xtZFyaEyYuTvq;j#C^}={YbK(?O%j6i`f1Zh^t`b&mzfLR{jFgyuivgBkg9^
z=UhkJb+&&RmZV|keY3D6i^cg^lF#A-tXY6nK8xeO!`k0r#;28INjWQj5=&09xCU$1
zu>EVXb}ie#2}_!&{15e4J}92@`0hKyctgWDH7VZK>1g2Ix9(-nUB1-V_~1#+Up}gl
z^2P|U@|c?jtA9i1e?z89_BC|=n#Iqd^f_BnS&wYj)59B%e}be>5L0p)LbgMcCx>7{
zDyuj{r^DWclG|)D(G4ZtEbc&(4%V={kfe*^X19s!LV|TO-x<yu8qUSAkF=lvbC-b-
zQR3C&$C8LbvKL|56n7iwvb91<D>R|;;|i2qVR1XsY)6wQe-&w7rTB{M*K0@SF4`{C
zU$y|<Ex>{b?f`~6*m_JJ(9NTKdt#gR@XvI9dg}>)`;**|rP(Q8{pIbO?B)dfR$Mos
zD^?F=^=vJ-9oV-6#$LODeK(5-f&CzhQ=xq-YyW$o{T_;M%I9|`9JvdH?ZG^sdH%%Q
zIsQYp4?Tz$Db96lD%f+*W6<Up#U0MgUe69{gLl^ahSp5XNw{jDxKu=xpvV7taCYoB
z1H(Cf1MuCz*1wX#`$=F9>8Cg&nVk&6C@zFE3t<SwMQ~OT45hdj1{cG*1osZ=e{}zH
zsEM+kG(HV5jv@<;&th>t7@yDL13>Kn&?FC1)RWYTSzH3tN?2S5)XFIC&#gKsD>X9J
zes`JT%}#oS15@nZMlV+gmrwANtMo`6^VhKJHF{)^`RmyEI=v{=|DYq~Ze2&=TR(If
zND8PlZ8-e&Pp@ddh${?weWm@G5HZ6I!|jQWdvWDo4ww4N{v-VNQ84)^^|<6ZnouQB
z+~gxR$?+$E+X+VO<-U`$UoxL4k3M7rz@M5zGL$7l2IoLo4#h+BD{|c(#=RCi@NH;%
z8!|Dq8=7{rxC5DXAcpTkrd<@1yKzEIU2&(7zUX$Kv7J4~%mo^`EG`5Zg)FXyI@ORl
z?Qes|ZM3S&@hM0@g}us_i*#}kBVT}w3)nNQ{YZa5D_?<hDp>hyWL(Y4pGEp-S@}e4
zpNN@5$TX~<hM6<GEUcf!;sR`3z@CYfWBqcLuffJOEZ>Cnn<!5TEy5Kd#obFfjXs7x
z&g2}4<3rs;=aOi-GWz4iip(o+Wn|pn6h9^?n*aXkksl(Ff$AsZm%Lj2lkQP8VYt4*
T1YJEHk-mx(>C&G%^ymKp`i~Kr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_selector.hpp.88C73ECDE0BC86D2.idx b/.cache/clangd/index/gridwise_normalization_selector.hpp.88C73ECDE0BC86D2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ca921cfb64905e498542e641e487ca896c0ce97e
GIT binary patch
literal 1186
zcmWIYbaR`<!oc91;#rZKT9OE4G6({3aY<38ATtAlJTC)7#hkMtvHrIWM9%u3Wac+;
z<(^XJVDgb+?lMk8l|yE~Qg)tfjVgVkQ5MDZX#e@2WkSoh_}@-`XZiftzVAOG;`#6S
z+3f6We8}<UJtyaNmjkvfnF^fg@29u3=PT<!@Hrtfw~6C|?yimT7dHQLN&F_e@agVk
z`8{)Ib)2{8+R$iYHu1TF@14%h-G&1DzjSOnpzR=-u<Ua2r<)8jH;4)uz4>_F_K;JB
z<E6BfzZ7Q$%jry13Gq}>v5=2&+w);vrO0xjWf70qO=F~!95XuS$xQ!WyXK+4E8{|z
z79S6Vw-PHEgYAxQxiBXr{p_-xq2cGEFNgi!9d!7vmhQ6kn*Mulh8<jXRIYlT)cv(m
zU-`FvO5G8<YOQOw;mzHydv(@&KW(~c&LwCtGxC?7p81j*W?#vVKmMM`+TT>AU%7qa
zH=8dHr}wRQtlhQrV0O+mCN<uKeYu%WXI(7}&b+s5_sdyJlcW82S6MCf<c&zZaOy?)
z%%ioRxTbPzKlqk7k%4c0rq-0N(#KY$-Qh9t+ihFp#gy@(aJw`6{3|~e&u7`MIKgJ-
zq!qVn<a3s!o6gGLq;D^;V4|xx>*9*q?^Zp1XDjX|T<Iw=6Wz2}PV#`;EuVGCv+YXX
zoV8s3O`oy8!S17qPE$?Ll+F*=89vH#D;WPT_!BQzH|I~^-sl(7m4Ek`Uow*~JfmrU
z#F%AIa{ZjQe%}?tHpcx;-|n#9e{1-Z@Kv==cQ%%H=cg;QmvO!Od-Uyzb$Z*PBVL!S
zJ^uE_vyN-8e%;ve=}dreyt{Hq%!2L4cOC8Lre!7X+?Y6bkMGL)J9)49-=4Sb+{JmP
z8n$e@8j^Rpy3&eeYwBGdS%+JT!t!RF%KbHMP01TSuQ<7~&mT|xIyspm_|vTSpV)-h
zu7uA!Qx#z`^$ypbsiHcz=c|G;AMRYY`+?vN?nY;w6Q_EAN(r?{E|Yd_k=FQJ(D+;{
zVp^r~^bHNm6n7lFH1z<B|CF=duXmg;e3hp8)ab8BrDnarTBD>HE1qvX#Ch=5i}~V(
zf8L*0ITiU=etD#Lb>8_EKfGsWuC?G?yI=47^>~Kj%G@L;U=ERZf7H)jLgEGw0}lrS
zmk5K12$*2tWM*XGWoP7KRDXAH-tC`n3V9e<feL(CeHFk211Ad;0~-TSK~ZX2@hhN7
z!fadC#P-=AVdCKm;0Q2=>*LDg$b|D?*0FMdt=TY7dw=S#iQ9O%wlT_XV>Co)I>0D<
z0GWT5QT8k{|2m`Wbp&7jrTR9`vIq8zJY2#I!dyVtfC0=nMlP_MVYUTLQ9Ike+%5s=
z(k4cZCPq`3Iv%dsj2yEO{BAGimbd-?Er7;KiAu@99AA{0lL#`uxG1?;8N}cO5fH!)
VV=yprZ5QFYyn1)kGbT=M1^|ej3Pk__

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_splitk_1st.hpp.6CD10C30DDA5197D.idx b/.cache/clangd/index/gridwise_normalization_splitk_1st.hpp.6CD10C30DDA5197D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7daeb203c75097d835fdaa995ed9fe406662e2bb
GIT binary patch
literal 5352
zcmbtXd011&7SFjv3CR@&j7lc#djkXrAuLg37m!6zs?-f}X;oYR3tDJvc~o$z77C?;
z6>-6~ZYVB@MX<`NibV=a7oQf3f+*A?<tb|I+}jS5|K9gL^1;dd<u}WjGiSST;h~|w
zj28%k7Kdl1EJ;ri3IqZp{GXPd`j(0j2=?m>1es-D&y0%s)ot7#F%LE-pS1taZ(=Jb
ziY;=nOsu-)-2RzYg52-;uHFgNud-7d_JzlL&v4Cual5JX^3|;^KVDArJoPhUdoAg)
z_nu<cTMzE6S<|om$>{ahlo1lMp(h`G4n<>UW^^lEeq3>_&|p&NQd`5$_W7&)I^!IR
zbI+_!b~Sk#`N(o%+MhG?-&_6oN9!9?3l%eDlCuU+TZ9)k>Al;hXd1Qb`t_Cvb@k>J
z>KDJxJ)yra;ekQ3A~pSe_KU9WL#qeZJl@&tvb6c`gx0mEcekkC_U>MPHfQMVzB6@&
zU!QB-H~%IaNWNwD!=~MPuiyI+Jsq=XfBFYU8X7ao74tgJH=oc%bj^J1u)L<=N#Gh3
z)>!etX=mu0zJcxrffCn1wChB!^}-Whm-sE;s@QvF-G(>HpT|{v*}bGDapP}8AGo{S
zPjD#rYg6u6Q7Y?NKOi>GcwFkSz{~h#!^Fh+nh#H9UK^OzXXRu%ee2;r)AiTAUH9UZ
zY<217<G*`rZx6PNI$Wkm-n#b7)cP;>7^q*2b!{8%xu;<B;f~QNP5qVY`vnCXi!I$g
zo&Jfz!r~;!_l{9B!S~>=>t@{DpK)%(Q2*k1^`Mi>=HwzF_)L4=Z@*`iNov!T70Wz|
zQ#R?1mvz@)%<NL#J><XC->WI;rFdRjQsYrY)hcEiYTlmGk?l}keaoljQhu#Z{(1SG
zIhsnZF$J<u7R-vvHb}a0q+;j{`%bj+`TUw%so@3F9(#THOZ|-2pFR+ztxs9hgMA0u
z)IE#eZ|spW`jZ66MrI@Dg+PS;3xC9l6+RPQh}w)9A*N%6u|~WYrZd>+SBu7!Nf<*+
zJBl4`c|mA7+~{4^)qYCKNH8647VgXoLfanNoiP+$I$OdRV_G3r*zkhTc5l5*Ol2ND
zk}^h^j*>+=^J19pk&Zog!$2%ySWJhB!zS{A(5{^pJrRy=ucV9^(~)M8ZoDA0x^3Cn
zH;l8t7|o2DC_plOnH?_#1`Ob10|ISSR@Lp|x!n?GJf80-_OsvxnIE>H{Z#*abl#X5
zgXvV1YR(HnM=$bOJ6hFrU&4&Tw7b~diWh{QZ8p>6_hiL#DI>zPt;|-=3qrqP2G=4l
z>?oHqmYBAbS-SCp(DSx`RPvFY=9z?<faxjXDTo(@PO>eRAK&iXC1Lb19Uu;n@q*CF
zA%pd%Ztc${%vemT#cC^F5c=a)eqkB+UMi)G6{fvpUT(Y~ba`vcmWJ9RpGz1MOxug?
zt$9J{>hy=(LcbL{OBriSd&oT8ctPk>2PcmTuI_e}GB%iYmASg`g3v7ozu66uAd@nq
zF>Nigmh*zpKWxsE4^CRuEMcUW_7VG7@q*C545WL8ese-4WehMKVHV-c3qq$ZS(@fG
zmU{dsG4JWb5m=;&*UTcPXVDgzu?6DEHM6J?w1qH*;zo#VggAorQs43O`R2Yquqe|a
z)1LfH!iN#_VMF(71jj~D4-fk%jPNinhUl;|2rPrSWY`$Dyq(L7+<%s^$lAi%j#!ID
zHWoH^B#>ETYhi0g9E?RO6O}2s6^xWBHANbYW->jouZgcIF)@pLmA-0<v{SXABqCT8
zrHoRO=w#7GU^jv>QF830wm`Sm+be`ZloXT{{BHcso;i_Uc?!{9Q11mVGE9i}f%iV}
zp|}#<E5U={N$pjl7nP?yxMk)fCF(8{xfiCBVcas&`yhtmN(ifjX#|^&)n1-@p#BoR
zmwa%{2Mr-eYSI()DT$hVFWBt`Pcn=}m0(o~ZUkHVSJ(7=Ul)&1Et2{0?p)fVyDPQI
zStLkU6zm!tMxvS6TPWtUAf2+ne1RMBGww1>Hi8{dmqo20YXxV5?f%J`cwa7I(nXGf
z@+fG@FczJL@Y9e$u*03+*khKNbFmfuMgFEF|2SrQ@7`YL_B#IgQ?U9J0?2rm#V`EK
znCuIBLevHL&x{a4^bjl_f)&NjV9GO?N-#;<SU)afQ~i=cl4cVhU{MW(*1#MxN$uKk
zYc}gvg4IxplorH%WHN~P%zo^1#rY*?KR)GU`E=@{Ui%t6I@vxMTmSC@=eIg<+5QK4
zYe%yv%Qj2r{&PT@qifJykml;*TJWuf2%-y%>OfNmVH7ujR|A9+tTmi`aY3l<dLfHE
zR2~UBp(xTQXej!)_{8c2pOMy+1Rwdvse-FX<dFDcg8k=^iwW3#)wl0X+Acj7)q;I3
zjWSL>n>reWENTGj2AYV1<IujWK~XSPZwIJ%KnT&BMHRqSfSlq3@Y(_JrMLz>Y9Nqc
z@{Ntp)kdObSZkSIX<|s4KF&lUm9ft8oi$$c_`V`vW1Zuxwf^MTENTb)b{dZ?>I08H
zUC$i=j{%CKGzF_CxNO~lZ#&F4EQI)fRE%cgw;?4n@Xyr{Q4LW<UsAWRdAXuZ^kz{h
z1eQ{lOGv5GOgF6HJpU_TzXDe>pX6OUpR4M0{w0`piMq+`=Oym<YjZqDm?TM`fWw7r
z66kJDW=<;cKShN_&SuUkl5vQOU<J9xflj#yju)w;%-i{>H*DM=6%30K925LVl(1-}
z^-48CBsJl+xGsaP)d_4T^_4|ePX(^Jr_{sVnB|qF<1#tmkOK;0Iu_-ELoUUnPr<8m
zO$uH87$}d?h+4etXnn#T2RgB)PDW00ol_e*%gL!%7A|rK?uf|1gRX(qHE^Mueg{_H
zQT+aP*};^AY+HP>onYQcGu4&|y?^y?f8~v9Z@4sEPRa+1BBT*=(s{5bQW_~INsLA5
z7U{#S0E_a#Fb|}3=X@~C*Tq}FaEmV93Wi&C@is8rri+Wfut*oT0NVm4Wc_X3$JfRO
zJly_2k4(Z27Z$GDqjo+F-iINYtjD5W@aUzLl|@g$;|ayZOD}XyF}2pimg|6V9bir-
zmhDp9&APBoGeTb{jVJbXuIr>~DlL`XB<Hgz$TcXG_~oI@+EKA76`>*)+1uH>kQK0_
zsGTNLM{x>qilWF_?L3*d4Gx_U=O~KYRqlQyL$F9~t9B+%%AzVzR?+t2SjBV*Ej)4@
zTO+_Mz=dc;5*|JT*BjGA?1adjkVH(zqMKlUlV-7#4>z2gxVit!5q2HKN;smp)|>8*
z)8h`XAiJMg`b$%=_(})%oETw@A^#qT1&dOJsl%vNYh~1#H@FQ~iUs2q$Vf_HQ4@%o
zXgA0;AZ#m4AZxSe4v6kh{@k}W<{b8Wa&Lqu`}g-PSFiRA)Q`}|nzi;>aWV4Z5o$vk
za|1E|`a<no0NMiGcZf?c{}L>uwrL0Vc2MbJbvyMbuGUTN1Aof*K|mh_Q9J;X2Xy%X
z2pFLJ4dtVsPt00ah<$Q{e8X#G8Y#5cE4WuGs;$4mKaiA*9|p?e-YnU?X@mwQi4wni
zaqnkzui2ingQyQBMkmgrC`FYLNKp|eifG@_;oThc-$RpNgg5qZ+S8blw>}%>8>Z>C
z^u3O2n8NGm`vuo9O>3Z8v3JOP`GfBhY`EGsG0TS7ltnop%7O6|=YlAg;{LLl^U7mZ
zwj)%#zo%pKy-OXH|97AF?8bYZE&Ww>*cI}?F^~G>^XCW7D{mDKV4@n7)ik}QE?Jqx
zWrDQS<g|G1LC%T4KNjRcATXzoX7o`u{K%lfdC}`kA{BL`%`HsL<T8QjFems68JU_*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_splitk_2nd.hpp.FE37C91CA5D8EB00.idx b/.cache/clangd/index/gridwise_normalization_splitk_2nd.hpp.FE37C91CA5D8EB00.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a16d1b9a62e5c8ad0f6803ceacf97e2b3ded6dc9
GIT binary patch
literal 9936
zcmds7dsvNG+i$HEYWGMEn=HGcBh_wKYIm@ccC{rDNhOTrFnBSq;>BR7gmxM!8Y&`^
zB$Xl|ogA7F!i)|i6gkX{OwMDzXYcD<tu>cF-s}7CyV}>)-Tw7kYdz1p*S+rhUiT_s
zf<jSauA$*Mf5Mt&3s%n6)X>oA&;LZOj96R;H8fmwG&I(vmrbAK|JZiWHUEyNd8ZDx
zos;{F3)WhtlPeByN^UdHPq=&W$EAG+-~94!Rk~hNT($Pwq>-cF?0@xhnyGWkx?z)^
zOw=9pt?eFkWXB%=!i)U`pT>w@mL>EpTe%H+Hb1nk)?K@(#wyWf_53rtJw?O9qcRRO
zUl(r(t+_XKpZ;CPtDW*_&y3Lf<=Kf~_YJR#4Z5O<RYwQ!TUEEddV;K1|AQ{6v$L!l
z7vEph^3CxBC&w>bduG;Rr`l;vk9?he8*^}HP*sffLA%c$5eI^2B>wt*<)P{=d)!S$
z|6aduupBxm4=L<kOcj-^vkn*#)#~495c+v>wPc9w`^`S~W6frG?_XPUWaasmg2t@L
zAJ*J4u@zswjN4X}R7Yhv_1zwuo0FLL+@#-P--O6Ku<G!T%C&P2Z;4r+YJE*RDYj|o
z$%ni9ys@}4pjbYy<W1Y9N1rkuIlPO#IQEpgVObyB%+><~wb82Turtd!*Hzaytv)8(
zKKjqG_ja%DUA80QYGd{N>NDPT<35E2f2q`)quJONG3W=+vvXy-wZT)zw%F}&-F5!?
z>yRg-b{5$s#5Dc&@IQmDC7xI}Jm){x*QO-?Yw3gf&llw*OJ)>K`t8lbV2_yEPr2bm
zQJtngP1Te)ztKPA`IpAo-Ere0-IWTr_tz?aS1!Hpa3cFjQlwu=m}JK2`!=WCYcDLB
zGCQO!=lhzL)$`WjSLOpZe7HButZ`(C`CNT->loYq&X=QQu`cU+|7d%uBYwE?_Q2D7
zD_#t`sp}l;VCvu-uV1xji0^^u6&D{Q8!UfgJnN_LGS8-5J@j<Ryp|-{-^a`jeyly;
zvh$mgo$K3bXD*2FS+>>ju+zf|pIvpoC!YLb_0FVAcE_nKKGeK6UpHw<?^t~E?{l#?
z_HB|6XnQhIKewf?W%=ujW~XZVTVj_2`{PGlwUx(;dv`q!wJ^A}V8Cd}(MKJztD~Bx
zKCX1HT{<!^a_zG4=e$+-k657$eC>I00Mr?xftL(hVyG4xTD;}+pFzzD6}^*k)(PQo
zp7zK7Hflj=<8s%^-CkQ=g;0;D7vV+LYC&j|M!lt7pJI3QhMI#l@W?(R&D26e8v=f8
zfPr1Wo24DWi=XO1A>Th#5IRgPNPiDen3raZvR(&rJnbRyFjWgedufNiJ7-}PE`&xr
z?TTG()q>FDO*Yi+{cCox5cc9}8J3CEg3#Z7sq_q}+|56`KTmrJyiC-B(3?BT|4^)a
zxlIU-d3qwAXsZ^4PWiz-tzuW`MjhCXr+o#!Bh-S>dn49{9UHi7m=N~i=>-N0tkr_h
z=|{ptulFlg2%!m2kH_Q1YC-61<7i2AX{VPEYV-7bgZb8KLFhBT{DNDzCHv{X0X*$4
za5q*9LSN30+W7Ct(-k_fFHcVvOg2&rLf2(r?c;{e-F4s~o(>WO8LLG%U0{CWsKDf+
z4ixZoxFCFpS`hlK(IH>I>&CNn;6R=Z76cDh3qm*j)Gpi<T~ez9b$L2K5HLb52>oKL
z@62z0{UXqTh^J=@W@EJ=^h?LI7aNCU_7cJoJUtChvsDX1cbQBr-t>FKav{{>>4gRh
zZPkL%?|fUV%U6CW6vEy-?S|b%YC-6T1q&lL57*E@5$nddRJXhL;_w=?HP&PUa5xW$
z^MC`zML=A{;s-$dfW<99+``}(x3tgugH{6$E73S5@*)HEwcVDwWVc<@01nHHWM;%F
zIqYWSW=7tb!+rz(h7z0Ok;KoDqA5;O+{jMkaG+D58_^VpL&k+D$f9s~kz|o0*_0f<
z0JsZa0C^D(Ujf_|7FQw5Dr8IfY9y&fjuc-(mRDH52HDggMngX%`=60BZQp_{TiEsw
zk<CN4eJiqSMNXuBKm9v$+uINRrK^b>fT976Cxj;M0E!M4zX$U7z?1SH!JLmEjN&+`
zh=WZ3csMT}&Zm4TRHU+eI-Hlz@_A5^$MX4bUOvlLLPaIZ--e3YELNgmB?_U>dx*Rq
zB5#7pZy-?nR^k=OiuMwyeJc4ghfhQ2)6kFf;&2Ugu3>Q!vP?o_DW8hOsmOz1{a*{U
zK4+KM^J;AbE{(vI5SsWoaCy$+55VyQkW&5=aQwt#C3I0jrvG;6vYo|g&?Sw<`=HA{
z7N3AFCs<qnT?$xS1s$s(^ZYx|@eYft(X47To4#KS@~WX~9okR!^SC1wm-rnQ1FT}`
z#EjjGOFNx!3VOT-$WGT5sKYL`pbKE&1z<)Lr`q@fudp~9;cUdXz5;{`SbQECo@Y05
z9l~`i-+&AoSiT+McFL3Y;O$)%^9%M{l`lBE^Yq9K&KWx{*aq{Cw+?W1fO-A{z&&8`
zD=^{}=(cDMe*xSVmfr!n9k9E<c>d=x25>_a(8NC?*B_CLOx(tQ{&@S`)$vdhCnDoS
zWJ(B4ya$=>K^6p)JOm2V`3IQ*hr1xx1r55JweLSMEbP*E{DYLpK}lVuqfKjb2xq&8
zH$@*UA0rapoz22E=JuPjif><yysMFd4CE~TNnZ4EM8B^F8uQ(@Z?ihKt?19;1mKnc
zCXpT-P6V<<z<7mZphyNmlurS!DZrQFy<qHKFqz_X;FivgmjN6yKzEela3+vtvg2m~
z_blK~`{w{*4r@cXK$y$oJRr<taXt{{v$zll3t3zQgheba1*WA;IK@@Kw2ESxuG_^q
z3e&Zk9G*COVh|G>wWe4RPs+zetw1X#Huev3AhD6;lDxN7C6aWBJAu3t%%by*f#MkI
z+c~@oig&U22$URw6G(dx<yF_;S|ayps}^ij0_e6q)q;&o1Y<};4krWqWY8THIGh5^
zQ-BNQ_X5kkz>VT`Fe;rLF9QtE0HZ0N2}WkJ<7WY@EYKYmNOFt>>ZF*?S8gC5O_af5
zZv*eqBop=u7<PF5TP-gg4sQjHTiJJQ0D28zIDK{p(Cc9Fd%(R1BFe`>y*S9UPlbA^
zEY5>^c`UAkdX+4`4fSrbxB+qvkQx6FvOk36<as2g8Yoa__;jvuKo&<;!r=yJ&;X4{
zHx9Q$w|3~xVvly%oo8_PC3Jhq@-LysOP241uAPus@=iFm6ON<fB_qdVn(%Nq4LPTw
z?tseSTx6Gvm?h0cj=606V&qu-PwY~R7?Y|;4)v%zfO5DVIoGq}KSTD<*aYkul0Kse
z+xXk*W70ysBm43Brk39TW=~m9kurPAI#9~&DaR0ZvZquEv~On>@Dkd;WN{~ybVA0)
zlaXaI`$B0*oW@Qt7n$d>fu$H(7PIZkk!LxYG^1Dkl#NA2eyRzNs%I70%$HGE{u#1(
zhV00KOma&4Zm@WJBJZZdUBkWW{`aUe$=_*6jrX3Zh{HM1HwQ9Kw+#B0vG@k`y+Lu%
zvO*_6-OOOV;{nJWfHveo94>}jF^exl?lQ$8pZ3_bpNnzm@u}}DS=zwNUeA_)73jjR
zo|Z3flOQc#=fENdSW(sG0*hQ0=K+g67Uu(td=?i1i$WF`0gECQmjY=iO%6Feut=*Y
zp8surMNrp~Cy>MP{&ErH8T)&Sh-W0fd1tCB6i^4_BlDR~qCJPhY{JHqBkt1Va0{<i
z-_^WrG(*c~n$~joKD4~gVkOd3qHgE5v?_a4+P<!tUnLEjcHpDSf;O1)*1H86Zb2i6
z%E-pm5vWTP^zHrt6aN5nNiz;dL21-K4iZqB#NtfooC%qf;TSSMMzbjn7a_|c#JIO|
zBq~StbT{5Yqi!L_oi-y;Gs{0k7Ee)kgp2n@n_unt+nM9=eBpc%DSB{tfpCF{6i_(4
zP`FS;PW~K@Fp4mz6N)s7WY9{Zl?;joy3s&L+r|Lh7#42^x|>-X3v^>y+zPo?IFR<w
zLR=Q=#@jnfD}VPiYWuILWx}(QE2eq8*zte1gPrxy{)jra?!;GikYYHm@5Q(kPkB4|
zXQd%!Vb#;ILyZwnY({#Uks*0`Qmo>0A62D_8pvxPe+_1kW*m-&;%La^gvn5x%;GF4
z$%5TkELkIdSXG6K4*L<<e`M45cqomBOy-jgrRglrhthnC)Aw0DT3@xwsmD7<;yCzI
zhpGff^|uvpTS0f4vj0Wp%_324<X7vR`9u7UYr0O}Us(v*O8mJ{RXn5vHvnk^o7Qyz
zX$Onn1N-;v!o@*p99<X=r$T8ei}Ro~kHwWxTFK(uP<or<<7>|KnzAfY!EbnXGj}m5
zWAcvIQSL?^uiRTci6U=DZ#Uw8`E=FWdlE&yqkYE^+vf0O<H?q!_Q~M@(*SEyjpXoc
zAifQl%GP5be$3*tP;wS_SEe|84N9(2T#)I<AGS`M<X2#}!E7;c@uyyVe`<)b^H`6s
zGn`rY@S2yPww0g%63r#Ls*3^Dhb$l;6;(xOYGyveeVmDLb9j!;90gHXX|ObLOm!3H
zN1CZM(~z9I&&fvgf3odu0&f{jP^$?tHOpwE6^$62jks(yn7XV2#1*jk9OBNg_&m}(
zk1%ath6a^U&&Zb|5Ld^xzlXScEN(!04Q&5LG^mm7-;TI;+Wvga-5Ggf-n8|2lc=)z
z!Op4JH|s0kQ+3!Q&hUb#4Zp%&(5MR<6V;Qg$y<x+PJ&uXDU_75LFOuyT&4I*{jIB4
zpZ{6W<D&t&m*Jy<>QaX4XQu5;AF@%%k%bSrs;e2==nRycVGmx_P_LR6t~gu^CABPn
z7wX-m{0*7;`T&1XN{=^Na&qOBth&3R`kU!HGmI$tuH<D?(e$un{zy;)L?ytIm@tPc
z!0-yteayQT`>THMRrr_xngrhaGd*C@9_7{^Z^R_x@|#?B`A79z55)Csweb-UKccwR
z`G)7a<NBb-+aHO8y#1+eKdFAUAZ`mi1HSNi@4sj9-nZ(iqJILIPAxDIOil!W6eoj_
zWbh5eDZnoU1W>#eOxX*VGI2WaPiM!=0E!H@?3)QDXR_mGfvH&_n2bm2Kz#)2n_2q)
zK{i2N#Dj787LeXz2W<t?Ru;EG`!?u92IcSzX#ax6uc72M%^P2RG+feAA7rAg6=>yv
zL9}Bo(8^_T9?;5TaX!$>XK^9WDr9jH&?;haDHu`;Oz8MkU`Q3ko$00Jl7bPYJ>E<^
zpSLR8F12R=zpHAKLIA&iRQLbX{th9<A=G`6eY<7C8PnR6U3_-(&x;yz-_H9d)jtT(
z{(;j1rxWAm&(%Ptq;=yDcQ-oR>MZ1M({VJgilzsoPoIvIOPW*uR2L2<IjjYjK}i{l
zZ$QZniX#>*ovX^ABO~TTPE&murV{E8#InGtA~0h@Qf~u8ebKO?8U}>Y&=AqDhV<)D
J`c<EP{U06;IQ;+s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_normalization_welford_variance.hpp.FE77CD9DBE7C3CF0.idx b/.cache/clangd/index/gridwise_normalization_welford_variance.hpp.FE77CD9DBE7C3CF0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b290bf26d1b8a128b4f8b3e147f9c39d088a4a22
GIT binary patch
literal 10872
zcmeHNc~Dfxx}WYoDh>x7kr5o5RW=bAwn0`!_C*#kxCIPy5shMuA#RBnmn({zi;4n<
zEh35_A{rM6pu!bbj4R&6AO;sOqKJZ^pz)oVS3TYSm3mcqw<=Xq@YVeIeW%Yk-CuuO
zPh^n4|L0aBkxzWk@`ZE0oXv?uqG7^c(wAR+dI5?=4q}mLc}`7K%%a<NIxoX7{M*2J
zu+92ad&(j|-Y2#Xjw&hhm~pmpwO?GoakSvTAH8N_8}FM=`tE3lnPpqf*U7HEdROa<
zPJH&ssWYd8xaO|a^Il|~tXWnazQ{AAX}hJ#kQMVTEVydBEWUD=5B@kV*x7Di-G-Hp
zjcfL}9WEI)tUh(%hVci#d}rXzMK9X5Hi+Lo&%9z-Zd5|1wchH2<y!($+!x*FZ+Jd(
zG&n6Ut!a<A`f^82>)fSPX7;P1gN$lJvn;i)TYa@R?r6Z`>Cf!7Mh>`S*yw)q)&1??
z<OTO}A8ESAP7cdEy7hSX(fG^e*$?M_%^k0=JZ64p*&d%~-{;Clb(gOHVa@epk7sFq
z8gek{*w0RT1Kw6jc8|LQ&kPT1d8-}c)OF|i?L9HI2a3y1|1hG|^OAI;&DqAK3kR2l
zO&h#uY}0dD>ahsN0~rhU9i2R__0Z(TI+tCE>ta@n-q3G(qBO6@bIZpM9vswK`tY0c
z)6XT{SdzM>{g<9q-W3uDuM^(~^cp4QM;zb3rDnK;iSgHpew)H@Ux)3^wbx_|wvE5m
zDblK%_p{IF#J=%Iu2wv$Ua(Qix$M)9)X3k9O=gst+kX}4a&c}&*a*$@CtpkVet*+<
zpppLLL7jVNjq&&_;Q9I5XWQq^a<;slWM<Y9vj4bkRzl_=4Q+ApffcLto$lPVeKKIy
zlBKQ5UD$BKgk=r4-B-k)H2Sfl_wxNiiQx-I6sIQN?VV-<zu6n;rSt2n!#}S*^Uw57
zKl_I_x<_dGwHp7X2pqEeii=jdXyuPRiqFkVM^w2#k>5X)KUec&{x43;v%@==NY5EA
zF@12#qAEP`CmS=v6)zlKHnm0lTPuFx=fCA@EuWHjZ*<XhbN~7cyY_A^eqzD5J5^YP
zzI><Oa<|p!nst0%{m|>DbaUSNWu4UaU8=u$mG!>Y$-edNO-*xf{0W=+1D?w&_YMp=
z1)lFd>nBQDu`uqjV9~?-1GGIobmvH+$w(3Qm-*YNgh)fMdf`VmDEMr9*6KT&5D9dY
zI7+4xgf{QhigFscD^>~z33QAs2CD?2Ei_MgKS{|b)`Y_a+D+_cq7sDm>|Js>W9I!c
z+K?0I1X+TGN)X!bhSk}b6N^r2LLGq)7l#|F1fc_TOIJLtoFkS(u|UV^$Jwa_p(A&g
zUrs9;A3G2Z&=X<vLFP^>AsP&Uu1EwhnD%Vuw8lq?zi2}(;rirJ$p$Jxu8&U054+Ov
zXqyyj2=px3EUXfQPH1=h#4+&2Oer)HXcw7_ok|dT>8l;D{Y|@THK9bH{ltF8DnaN~
zJFA<mV>g_T!hr%EEsMq~LFkn4;?hQ-$UIHR3-nC!OkI^A^yUMrR!$3l6(EI!1$wsr
zY-^Puboz{tU)Eo`9VvyH0v)R#YpoK5E=^r-yuj1-rY6)D=n3Kpqg8^?$8VoF_}i>;
z5=}Tlpd-bRqf~;>7tUID{Qmn4cPTU$=wMl}tx6EO@%F+|gNrtMOW`nqj+ez-s|2B&
zT&sA)rkn$saEL&Ah&@bIg3zz_2i}eOzM)7HN(DMp9BQl*gzk@=3_hPf<ZCG$EzsUF
zZyS{$^cxVkfcvu3P7@9l=pb>Bp-K??i#Z8N?xrFUzs@uJP2SXHkjJ;6!!77c41vcr
z$hii&5v*mm@w@pu?SI8Q4j3LVlI#=#$paJ=g^UaFB(uWfu;F1N$!iEm9;Tou!Yjg`
z7&VWh4WrG79q>5DFvg7RCLYf+oMlD~kjL@zcm<hQ9`6AB4xmjQlE=A#&t-8BwCaI&
zl<$S|UZ|kBA6oUZ{2OTd20BsxEp&JbT`1m(tac)1ybNTUff)TvWS@z~QT^A*_BG4*
zA;&)CO8GyK?H?@v4mrF-%>7n#&Z{{Os-MBxW^nBNIEPG*(a+)(S)4o3AF6lL+wMm5
z0!<DlO-S;(Fu(t&)tR7`!_{CyHSi*89Bu$U4ZxS;Cg9V=;tt@`!QwU0XANZZ)1gl~
zi*Lc8TQHc8mw|jUkRQc0D6|HJQJl#IW^zFU6LZoQtK2VfUOYYuLXLvjM2*L1fb$tJ
ziQ=2U`6i3+1DE?ieH$KELGvoe%w{dLs)g!l;BgleejT#+PTw$3w(J3q*P}7((F8jF
zc7(ShhR;P7xyY09hmh4F#EjR1tXf$77}-8X>S^O~4>IaO&h&m8Ih&1~4;?R^GfL+e
zK93ul$4w+W2}ek*3PW_7ZbHdTXhhU_T#Y2vNR9RWb6Dd|(FuoPf>A^zO{GRr0~9qt
z?EpA@64;*v4rCY(Uk0w1f!ZN(_$Q$FiPgUYT<@@aGf*_M{4?PCjO9ClqLby<Ld9C@
zfH=GZDt55A7e@EO82UVKpzj;V7*!*(Z$!-VHX{E<6hQU+P;4JkTMvidp}==YZFD1t
zdR*|@ce)<%xENR#1NA00nQ*l7h09g(hxvMMzKDN@VpWzyt?v-<I0VL%X(w5aV5G{F
zhpPD)`M8i6#^Y(W)BK4&TiksbQEs$6T3CrG8dFA+7;KsBlAm_cZW`k8VPJe1sP_Sn
zYXM&itjR5Td=2o|D7J}z<E{^r>mY}3!_l{)8Byf$X*BvYQe)fj_yC6k%Qg?<@HuFB
z4jNO%CTQLSEh&Bg$2@>zDNaR(sff|vgADhuxDJ`uA?EsvXv{^VzTUnyuWVFo{UYHB
z8{woz7)l4OK~^<XQDNKCG@ZB05~jvS!^e=!pR>^`uehh?mBPQ5m@Kg%siccpoM`sY
zPnHZ&zK`PE`+Lc+7y8rNq#{KsV)p(Xq}apaJ`~i4)DGRnJJb7zKI5T(`n>LoHqPDq
z`^Lp$9;X4*G{8izEMS_&;sPKo0P3aSaUqZvvN#*MWz&$t<2*P%4>G=eH}u?1Cr-E?
zIu+88#N#69Q3RP-Q35BFz;H4H9z#6OeBy7qg5$BjqyJPAfp{FE5i*u0G@&-35ljj*
zEW(MTFl5e!?V!xvshyWW%QBjAOx$$It>bP|?m!-AL-TCNqyl-+Di1QjWH%hMn+;Zl
z(5MhPk_T44kYy26hk72Dz_BGzz3+S@M-{H@(eNG2<5Hj~1!_Mjgag1=155{M0Fnlv
z4i7wT0+J>ccK}HTi`PKO8aBM7LrFS|GoeE!^rqwQg8VM1z8{$|VWN~(44tSczEft9
zy-vy*!D}dM8LE^tG3g@`v3Q&b{WIYVa;JbBIqumr{gR|YQmPT4&8(qyh%;ToFpV&6
zW(^e)6UZY8Qv=Pb*>tcDTGg@X;9Y2YmrV!TpnV&=23^pn3x?8wo`S}tAZE?BBcJVv
z;q#D19<nDdz~gdcRnDe^^~koKO$RR{`^$(~^LxnZ9;@GuY};A=9u&}nCe!Ch<!n+p
zb!6jlIv0@6F?>EZHlI_c$D!l4FB)O9Wvk#;X1dMvBZXLK_&DA3ep!<RNBkHBKL%=V
z#pAai_$`amU~n4Ulw=u!SXF(N=Z!rz=f`cwr!G7^ZqkqfrDJaad<(Fp^U)3XZWgCN
zJ_V|Iq9=@}tXRqL)BCgRTdUSN{mru8XpijS1p^*^_NPvj)I!1zQdZ*RwRt?>WxgBn
zLvfdmdw=#&js%C62AtDq-o@iA;GD(c0$@=9)UinEc`XVlj-Pwz^wgL89tahp=P*x9
zidG(b4fDdJd?pP5VZSP?fBHzC`ks!=CHh{D<dVenI1AsF;AP5i4vh~(8?u@_J_3!8
zu(%OgHqu>~SXVSAJNM<se|c0RZHfV6ReOTI#4+G~45-}<kL!U=J*|d#+z4zMS=<Wb
ztw0@INm)BmtZHk}Vdsa<k08s%<1*k;M$;yxE3&U<-F-DwR6}O#)<H$x``ERPbwzif
z>s=PNL60^X9(bIBoKk3nS7MhGTJ8(+6*=Uwt|bpS=drG(94X4#O1B)jmb0#@9=X=D
z_%iaijMU+Y$M=xp9((;g<a&?QZ%3}}?_<w)wiZj}Tv9o8HNxXm&Lfp$#?R-R^V!Nc
zpL5HnmGtT~+5Ux7lT8Jyz6nik(wtCPEtyo)oRBy@!6=liC%SsCfczC;YRPp_vJUD~
zBi;n%o7nKZ6H0bceseJT=hL)zE*d<JmBwPyrQ-2y>1<3|T0D-E#$gf*dAtr7t^?-u
z7W=`F{Xj}_F&I+J;u0h&LAsQ0LXswm$z3?Hs$oXoEKx5}okH?BF)T5LTInumzYD4>
z+$}FEYoGZTKK!pI%-=pbvUtkG_O$=A^VoL%QRkYfugm^&=RsC*kXY4Uq_bHK<kjrH
zZUBxAfXR-UfMXMjJAh*ci`PKMHEcGO4jt1e&M^3g)|u4BZGYOo%Sfne8$Z$n35wUC
z&NXO4&7u<NR3bIb$sK!dW!;i-f3xVB*Is+|2YfN<PbZ>a)#kK@A>;i1atkP3UNs<M
zVS&Gpal4th*^t6l2%8E|PZ|R~{X9b{@=|ztl87jTR6nm!iu@e?+(}I0aga%n6=~}7
zxDUwtXp4f!TcCUkjRs^4Asj1P+SC|*Jba=^(a7WJw$lSh(Oq6J3HYa%9T9T9DE%lK
z^6z8qUmerk^rGZL^W(kd2dPtp6ij(2KqvShv^+?Qx{8F~8hphUTLhy_;1Y)X?<tbd
zbjQhDvKm!+7!P}}vFBSo9xvt>kD$J`9BPz9rng-I`3g9K?%OMnzrx~YkblPFPAKVw
zGOGU^>O6-A6t6}6TEvX^FU0?g#c4>AhM4QOBAu;>xjqN+Ijnvul9aOgClG&v<(rYD
zndRG%q>aTdkmLo6yO2&7d%td^*G>JD(4FNZ>pAB6(l|*Pi??z*TRG-=w{d#gIJKV>
z&I&k5F4eC--#Fu-d)LDcmj0fl*IBHxA9pl(?_cgN&t1QE>C)rw(|@x#L;46poL8Rr
z(VgD|Tzc5EhmFu>BinCkhPKT#Zu7Vm+PBi~wQ!b!Y|kJjFK9sa4fM$5iii2i$Z(wT
z!8+fw&PMr@Uw?>-O&1*Ow?O!{qTVe77-i7b0XZ?$5vz`1>9Q@BEw-jEcLx~0gKawI
zg5kL=u7I)%$TXVIL6dW|GUah6H0WgeM{AL6En@Ujkx43BRHY$<H1;?#2g!0+{XNKJ
z5365_3`$x3M&#H?jfBU|$e@|!+mJyUi@T6X7dzg1&R{*q9A|FjOt!Mei@BUZF6Dn-
z{~x`9OJukH>*@Z_-8qr-vNmn_oB0Zo<_Wc<@^qgX-ErW397I!FJqv8k0`*Cia1;-0
zZn3xx$lHMWq^iZ`icfE;9{Au3-tz_Rfv>`|KHb%$YOPPohLLoB^I$|CoJ#R-7`_`u
zQCtZ93t3-X1jCBpN0cvtlS|+<f}i#o%>VuR)X}Qenr<2xO$cRcO*f0h1wf+!=uo~8
zXcSWXBB%0{yxiDS`r#y&J&EOT5f@lQzYf6RV$Qvo_HH<QfO9)QzvaN;gPhAjPJNW}
z;&IEShxIK*|L<okJt;vIrWcRA{L_vtIc^Zj1?Bg6s3Vv%c8WKtC0;jv<8<xC;gdoI
zwjWsTr#%BwunAWyzve@)E(LrkFeKL1-@osayeXqw)s~k(27!c7w&mq-S)2ytX;96R
zs#2&HmERMh<G%*-*KA*NGn8+pM{1%k<|NKmmNZFU{39tq`9WVPR6h`3JKVo1r0)FF
Y0fsUI{ZS%A!iYqp=tl$kQAR)h4_kh(P5=M^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_permute.hpp.E425D88E25D439BC.idx b/.cache/clangd/index/gridwise_permute.hpp.E425D88E25D439BC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..92605ba22f1302645fcf2722543d4eb07f829ef3
GIT binary patch
literal 7052
zcmd5=c~n$K*6&+1+O&_w)?QeiEDA^?4Rk<O*=d9i!T>rV;GiNV3TOmu30n-K21RgU
z0xF9(G?6Hbh!Y@+j6_A;COR^~%wQN)ViXk=*944Tbwlx9&YY8Ta`KOy!>#x0d+WZc
zTet4=qC*vmEv6h!;HuE9wB*bcTn>jb9sg%!rmvCca5!7|98Okc^P-eUE9aRHBlj7u
zt1-R7uOE_a2o*VpSshzin4cQ<!)cey|AhAXFTMTq^M@zjF*4s?ZlwBQ|LB8z!M0v6
z9v)d1VzGwTcl^%J4sMT3GbQm^cOMSj*mNR1G^xEhyg|q5^6lO8ZhUa7?M&dM-s7*!
zGtM<Rgi2dX8_vkDZI1KY7}|bidT4>3tJ>;&vyIDkp779*`S-3Tl6O$RU*_8>%|1L>
zW`ECLS0MasY`w=R$C>fUg9#mlioSUs9~(3cg$I}{cu;-E;_&M3xm6dVQ=Heg-&r(s
ze{1}vwR=mX-7Sidw~s%)Su)TZ`7-gmm-LbSrieohou~G^4Ayxab<0ivWZoZ4w#m0T
z=4B|~ac#fwpm;;=mO}@xUG|UE(f^mPu&GJq=r2oME<f<~{oMRDAIX+C<Xzo&`}(oH
z+gD%jXt<^`Yjw++YxA9yXOxRxR5(dLS1wy!yK_Uri@sG)uU?3`HT=-0ezfdG!N9cB
zGuEzuKa$`6D0l0*X@gHj`sz0>ZSEdv*?wZ%A4}a1#a~VPtmR|~I--hh6J|b^oOqu~
zTiYBpTqMo2ZaZ>Cx^dY)Vchjqe{QR8-qyA4uOFF+3*SDpyh$bc>0+GFU`7`|{D^d)
zji>%*hsHzwKCLM(n>~uRrYSt9R=AZX$n3h_zPRR-^z(Y}bF<-)(eCa!DwDD4Tf!O}
zZ@o7q+k3dtK~>yZux;wQ-zv@+nEcZ^)#2>%HmLS(3hLfmk$PikT+cGgmIuo2#^Upl
z)y4fKpPqCtxHEL4Q!-V`HIod+oLS-?-Kx?nUDTlOw)c=z^tP{W^u`vHSvc(pKkKu+
z<*cYS$}_K{JY@S$=Swb8E5km}tGW9s{m`wi`h#cQe)@1_MM;;#3}MKeAB?{;HmbN^
z+F-D8boR6h;)uupH&XmbjE_9ySU`89OY*17ehjg>e6Z?uZD?!BSXEfzcX7&L;pn>J
z-K%KVTZhlDr2BP~M*nEHqGPDlE7m@FLu#h*hnTuypKt$tr)gEF{i2hWyw(eTpXLYh
zj6YI&1TO6g`l<7q>eT;)y$-VP%zN5+-1ujmbrrqc8Skg948!+BKk;qLjkZ}_0h%G^
z(3imU63m(yI6Qoh@SoYb`5iS;!7fvfzKBD6PW2RN2}h5MFc)aB;>Nzp8`!){hzzif
zAd?_7Eg|#mvdYrpeP63YNLRw4qj*t5Eg@{#-!DnrB0GY{*>Th3X3-i#*nY)jRu`|B
zbH&I8&tA-1tfwV}9WX8IvG?C?c1=Sw@w*rD7aD5`VF&qGYzit-{lf@N#j_J;C%9+{
zVK1qVUOt$be#;0=!R!?K6kjbN?D)?6RVD-19|;hJmDhoD9fVH|L}VpDu+eahGMD1a
z#>SBtNi4O5qo)HLYauqyyG;=vlzV2>2$2B4BHS|ER!hkI4`ZDQ^EuM7@!d!BqJ^}E
z5cZz9&=N&pgkqc>!;2Bp8ba7rA6*&l7}4h#p*dLj67dpqEg@{R>|gPf;W}pH>{xNE
zIjtdtT`RpjruO<ry$~7V?FE|zn`y};yWe*UCx6t;02yQD1z=o&U1NM8B8|RJ8y>$%
zD42qDv4*f|VG=EwWWTpH-oL|h`!uA7*;|3X6>AtDh)CBLKhL%LE+tWjjIfBpL}8{S
zg#FFvVXF@=6m$y^AF~TVRtRwu0}<&B`0=7yv?xG;reTp%kd=bh#6U!bT6^SITl^LY
z&~z+v6l6z1F)<L4Cm+ZCiR&z_7a$%MX$Dy{)-pa2k*BUf^XL4x9GlAk+aq#Tq=l9c
zc6##ajQp7#&XhJ+3-90~?TDhw!Ll5@$pR^QKeF48g2<KAcPcDjTk6~^rsyE^AZv1q
zC_30Y*qWRTMJvn|*2GE_9XUJFjx2<t6Fm}riHlS83Xc^ml;n}bLg}9Ae#F};I@2?g
zg_NF37AgX95zxf&6kP=Kia^fB{zVYN;GH1e$(An$aWNZLfuah+Sw?_>{jiY12O!`8
z8>=8h1rZGYEZCj}2Xa0X-3fC$LB`;7V0(_u?}E8q44)Vq_!Ci$L^J*7&ri+OeXo5g
zMJMnQjI@ys6N*mZrLa&Ta0`KAxTU}?W#c;F*0J$X;2ve;X5cn6SbX}0SjS=59w=G`
zVinkvZAz{;@y6;}<$NxkY?mx0XNr+5*^xn6^RgVsDRSv7=PWscvK_OX7?k6b<IJE2
zh-!dn63)1EANcfvFR=}mz6m}z*?0ha1{iE?=zYQOU`HojbuR?=LNu93Y>ADlNef5!
z+n-84@o|g!J=+#NWxp?$mvAJw02}`!4zA1BrUx>a!mIJ@18<@fn+8RxPz;05Bh&MU
zX7JaD{u-GRY~}mK@e%i{{5R5cB==zpX%Ziamk#ox<_`rt3v8i<W((~}5_4EJ=0&5t
z(_Aj?E%$aI7UI%rc4_v^X`Y1zXTg`;04{wF7Mx?SEOJ$dyna&=aOrlCw}T52B#9QU
zMU!C3TAc6fKVrJPb3NkH*>>4d<~2EPIqn3>7Y^*0V!U*|4)!{mEL-w3m(G#quvV69
zmpiGz-FhWwg<;}GycwUjd{~?DbMbR!wl)IBBVa-flS_|)_XsQ?*garRz~6qdH2l@Y
zdAzkPy6;T=GhG~ELA)R2<W^Di0T3TxV-*sskeq0z=xQXcX5$(pu3@m(`oecUyH~hD
zhoZZ|xf}eMi9I0ifdB>%g3}=Q5lo5}K-yx3nA(5GCC{P%W>jhQVU<oqo}znTRu4Ec
zco576!I5C%AUb?)bwjL2(XW7d1*SxdqCZE}=V%g>z=2O(Q~Ho)1Vx+H=04{IVSDvr
z<FL1}8wqwJD@I=#5|puVI})_BaR(A~FgT>U(yJmaI75%3^JnDK#7i_egen3NJ7-Wk
z*|-v@N)R!}rUFXE#%iF{Y^*|*3W=Cysu87T<61=3GB{MR!E)(dbPV6NK?oRxU@|en
zZTEVS!_L28#xsyVW85wx(oK9mcyAQW;hT}|W+WvuD7qfm)-#xF5nEVOJrWDY&ggPp
zeneu874HDU9biwkPC_u2A1^1#Q6_j-%&5`5yY5#VeOdW~hapDKYNp`KFG{o{$BQ$+
zXax%;iIP|-U6gJ|mWeaJD3gVhBIP6^l`LMGri>-QoYb#8zP4^9JEZ6c(+EojCHf^Q
zSnKgy!9q!XNi3A^pB}<mkAEf$DgBj`DD`xmT}9Qi_}>*jNO_Gnr75e)dW*tzwmjVR
z+TDPn!;Ql&$R2SiV47gXY$I8aY|Nl+DtiXA@Ej_Kg>tD}7TN=XJz&hZa~;^#ffG4<
zivAJ=UxG1-%M|?`7<>mp#>e`>pr4IzgTZY!ehh-gV9dxr0l^an7YzpWyt_Jm>^H>f
zog>G;8V)r7{(o}D$Jg#Y`minM$nQD?(!}6H&@?aP5J>ZagQTW&AwSpT7wx)i;1>M5
zUP(G1tXtFZ5Z$Dz=JK_j4!L|3T>|0~FebN@!Ud7%O?F9AjiRcVlB9N`Gr|Eh&Lhr;
zBvy(}m!>;1E|MwDWFe(gIf?iaeH2+&LspyaAJS?R`ZyzjxEjnE->Cs{4I8ULto|jo
zQG<+;SA(k>7BF}S?1$hj##e^HewdBbNUUaxDT-Dj8#R(K@@nL&Mw9uNqHB@3mMve4
zY--u^waB%WktZ9}=W827#^CH~i2B9d#K<h#j$B;o7cPU5C}uYCCyHLIw^&48LD5Ti
zON7L!FtT0Z!k}1QtdKYbMaNpk${571hA7;hT5bhd<3#I9M#P_JT$wE#1H)r5oAj2=
z_p_VDJ4Oz_k!9$YtNSkp@h{%Q%}&bn6#Th#{Exz|w0<Fo8C(kbrEFXW`gLr46!ed>
zaWm*Qv+*V1UIGK6ug%+X`at3HVw?s~0`DY_QWIVF$@AYV{j1lL@83vcowm()cKagi
zv)>hfNXL#VHqA#3#%l(^bATOLYEV=S6Im_}?CuME8M*)$9%C?Tj9qcoA&WW`NG|3T
zZ=2lcNVBpx(t=l)td+b}R+-~-H|Ls9y2Ce)4|@N$c=;#z=SrAY37#ZMk+_SEHU5c*
zG1p<B4ud<BI`0B?7u<;zzTJ9VGG(Jl&;R5tq8V!(|02X_{1N8<2(Cl}MOPrZ3N)YK
zo9<TvUo=SIjkN3ig)bxbuG#k-dq+0}c0(vxwWfav>VYtVNl?NWcl>J?+5Sl7@l%dR
zE$6(E+Z!3Yx#RoJn{~grJ$hnv@4=t9N4ormiiboZY@zY5e8fWJ;~Eb5nooR8c3E$`
zezI16o_8)^H2Rg{Z+PMJZJ}Szxp?f^uNL?(MjIQ%gK3(8Sqb7w=KIj-=)qHtH_D!B
z8xm?WvL+|3X-KGg2B#;duF!<vjP#_86g(oj?v<nRQBuQHo`sdAiMi=4Gma&R&m7KV
ZD&TNTm|xAA$63sy8S`ktJX$f2{|{kmGzb6y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_put_element_1d.hpp.59E84F53787BFD31.idx b/.cache/clangd/index/gridwise_put_element_1d.hpp.59E84F53787BFD31.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2d2d044e85bf042e9a466980ec79b19634146392
GIT binary patch
literal 2964
zcmbuBdsNd^9Ke5f3>f?U+#nh6hu~l_#EqO|3<eu7+0dvBg;aX*NI*A}nF8^MAlyh9
zWC|jPdIV2tJ|GN)6A+r_ECfv@HNXc+fQMs77MfkCyY1NN@CWyt-LucV-|zjt?)`3Z
zX;jp99K*s>q*=<;&2N!046}hB_2$g<et==w<`|Y$)2z^|Zg8#d9-bHsms3(d2_aFm
zu`g5IsrJg=7#nL|YnhdT*|kk^ZK9)6p7LQjBVyZdSYcb!%{{e@(=}sy_SmAdRc3$P
zJ=9X^_SKnpKHqQQTD`_cb#S<=rTbAzW9adg-jKmo)usBX!k-f|aw=6{*v1{2I?%cE
zX`7oMxiBwQG~q0BeLTpdU+-&(-s7e4dy+@DpBU|@dIUah9}92R3^}%FJ+|eK2~+y)
z>|F4^G(soI7#p$`)h<7FDsL>f$8~D!>89lk^@@JhlAmfn4q1BU#7^$Dx&wUGwNEBn
z*SoH-xsBgAxR~9~Yxs^^Aii@m;16e-Vsqftk&$op-u78XJ9OtyUbV~Ge&pQNWez#j
z={=E*tj=#(UE4G5dhVz8dU%@WcQumetxF@Ht&Y$;R@WUk{_4Wll{N!$E8inL6)G=A
zTyc5xr-esF^DpZCP7b@dq*i}gTe<HY#hN?)a&YVY%lkD~)<%NjjnUZnj+kFZ6LtnU
zOPk2!z29H;4jR0^vORIl)atzMAHn=vhsSSqwabo#7bX9Ayz?oG>L2aZU8cF~q;ydY
z<aIvva;o38yrS+FZQ%BilyvHT%g;7D#%trJE?;s|?y$X27kXK0F?F^wsSZ}!jNYKE
z-({g@kgTZ~?(O8gcxJ*Va1RcAQQQ)XkIvm>kSyRf(~0RaGokI&p$joh4I#M%Mzg|j
zYinz-nF+Hp1DFj3;6T$gDU@EK{rm#S66*Un`?$?aNWVSd<#m#N|E(4zbExlW>FF{v
zVOC@k9mAkLmYMptx*0ly{4}S|_=gk;T*{ZqkQD-7$zN$gN$yEYjC)F^Qh*4>MW)Wg
zV?0h0A&Ejl1fI-JUSuQ{N9B)b9#>PHvCIDkiNKleOfPgEfwKZwBIGRH;Yt|y@=xo@
zWIUUd{UXvO<N2ce5F-@<Q4ug<EnsPZLE7QkNy?MMU%o_whd5+A8r8JGK?@A*KEgU>
zwPi3A5V$o4Yx(c+4E>_@q2LmA7VQ`F1a-QvxbZNYC?d)vW;r1@5O*!fg+>+ZhbH(!
z{xal=|M0$J&TpRG2*`M%YoZ%k7KBxXHT~&9e{MNwqh?+Gh-MR+z_|fj!yyFDTfpNY
zT>=ke1bP`s$PgMRm?k>&Mies}dcmTwd{M3?rF!9m0d*~mfV+^(H|p{To;#ulx$~ji
zNlFA13Sox`(K3ODv%?J(p%-h;NO-}}fS4@~Ktjtd*97fd9jP{hb{Dbu$S%Aj?#Vt#
zh`?1$)nY^|Vgu~YM^?QG&25R?5{;G#ybQ?7zzT79|C`Gx*6S0ds{DuBig<InCToND
z&Oc~=U-ewNlk}&;RQEV$(1k!&ps&#dAxmhW)cEiAZ1=s=4kHd_hZ-IUdfk9J^Quqm
z^X*7IU2!VC(el}bC)`{po$rwEiH_W`Gl>;3B-;uLo3kLtaA^p<fL9<!LIhq6c*Vek
zOTmItV35uV%D7PAanTjJTLNSyV1-e!0>~<iIV-9ZxLySL1yC?oma8y|>3~cJ6ezW5
z9fmO7e9$}pNM=?%u6rghCvXu>#6VIod|IdpiD_aJ3ZsRYP$DIfj`ViiZ``DijOykT
zdQGkG?cA^54Vcp{sF|?xvF&zS0c%qP9EyOUI!8wRWSh=b+QWPm0$Cw2$yhm%l^Zj5
zG^@@073IMw7>mqLwiJa(;4ywNCbY_Ll?lcB#T%&kTia#5jsw3z7es841Rc;kn$v7w
z`tY+^8d1yJgBK&H7lxp=57~!?+E`n#hs|Bh<PPY58j)s)4kd6kp*9ezE-d!+PYJrO
z?dN}Z?em~t@6BSlBlFIKwNG+hfe%OvBw7%TjC7TMuu&Svw1qD56Zo&A<#T<#di^ht
zI_Ffco?!1>d92IsS$w<xs79VvRl03XwuS^f;lCes0JGQ{jeZ<kmA+>ZrkX3|#-eax
zgMNu;Xbt0kx@C;?ty&0$0x5yEMj;_ZXrPBRtsR^b9?V$`KKx~<^xoN_!_x-;w{uUH
zNL!c}8h?APy**yCsXO}EH{$;kZEC8agL8BDBMNN2>@9{`!fwIIHEG0!fKv!exCC%Y
zOt}nj%0M*gCIqen9Gxkb15UXqR{&0hk>S6W4R1|fVRdG*dg1gJVtV>#;3q~O3@%tV
O#Uafjuz6?|!~O<$bbM?8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_set_buffer_value.hpp.1407258F85A006CD.idx b/.cache/clangd/index/gridwise_set_buffer_value.hpp.1407258F85A006CD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3d89a493cc1da4ffae1a5dd3fa0eac2cd8c23490
GIT binary patch
literal 962
zcmWIYbaUIq%)sEB;#rZKT9U}Zz`!5~#Kk2=nbM353`>|87%Jv$ov<<Qu!2Bar3k;=
zBrUF@+nlbFi#%!@%G@Q+nHjMiNnF12Y598Rq@_XA%T65XynOtxynETt2!mYFnol!i
zdnUi#{jVu=|Ec5YYIBdy*vq<>tMtWcu0w(Wb#rV}zTSD#`Y1K4*!25y-)a5&Z!{Kp
zoVfhe@;Rdcn{M*6%G?-DAEo&BPZWIkFVA{Y*ZC&=fANi!`i}pMOP`od%u(2s+7oh%
zVeu5Rb<DRG1TcMd*lT_EbF1jH9;uZjbIUcnnD?|M91O@^`cgFD<|PiX*azpF1p=+T
zv+oBRTxjg|eyZME;J&T*d%xGdHyW~<LW+8~z3l5#igK7ZO0_lrwk^C^w2vjRP|TC-
z`o!*AM~z-iJecXakFk>NuPZ}wWp0upFsxop-`~!4$it9_frEvCjaz_Q9ZWEAGBGl6
zGBdI<G8Cnz6)ONklGmVPR@PL*+e|!c!V<!AaE)woL2?l=o+Q7`Nrz<zjxzGFiK&Pg
z!-VBus&C^gdteU~=4RmL0@@1(*7+UjOJ8;r0*#aslv0GL=3!INQm}&YZ1`-C#Jj0f
z0`<x%%eldX9lK6jy}MsL7bvVCrl1KEb_tSRQuB?&1t@H&XlMu%cE3EwdUlj+G0-MC
zcey~AFb|uFw23N==W9MGUre`W0#L7@m7ptJ*xy|IgzJnzps-$?ULs65EO(>158tA2
zps=cfsuN5&>GU?Iyr=riKw&9mDRr1|w-<BE+y4I+%sgzOT%v+7VPG(Ei6QYMxFq1b
zH5I#A19BI+0o|<StmTc+qV22==fOl7xxn$ZVV?H>)Lj#|z4rLwKO=MIH=rvNY!u**
zd;3c+>*b{|Wsu`oqy%7^KfQX`_4Me&<sj8aVfpiFy~)K-8}|W)MO8%M5%T-@vi<sZ
yrhNel>pAPizziu$%}E5sYH?9=u`-ZmVml<W-%sb%w8>1Y><sMS2nQ3a2qplD(PXy(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_set_multiple_buffer_value.hpp.77B0D514858D3743.idx b/.cache/clangd/index/gridwise_set_multiple_buffer_value.hpp.77B0D514858D3743.idx
new file mode 100755
index 0000000000000000000000000000000000000000..294da0981fa08dd288c8cd1757b5345fd892d725
GIT binary patch
literal 1274
zcmWIYbaVT}!oc91;#rZKT9OE4G6({3aY<386C(qI7c&Dx#hk4ZytA1MMOr@xr#*CX
zlzNjbc%+r%?o_9{?K5K152Z|+a6H2L$9-vID>0p=B6n1t+?iaxzxZO!m9ri<qA#pz
ze00>)Q26wYyG_xJ&tGSV$}Bs&Vz+t__m2e&D|&tNPy9N(K!57Pue<-W98R)%CC6!g
zt9`z;SZdHqp9{X%R(h~XwMF`;9A%1DUtz9X%5c4Y)0U95!&l6oe}3>rS2tMD?<?o5
z|MRk=VqAB5IL|uzBi?b@QF)^npYR{TM?AYLd^v7;v9ct`%w#KgIx`_|+l^VVQYTD;
ztMzT4E!%i(b1zd=^W_B_uEe-9&bV&T@a<S&^1+P?y~mdJN{E;&IQ(;&^~;AzJ6EsS
zd-ciVOP?mXS8;6V)H*Tc*Pi2hI--?YPB98E*G`_36X!Wkt}k=GgOStyPhB5R%(V>6
zUg8iR`N#Hrab<3jA}~a`qM|3doTzT$Vc=k4U=v^vPyiDQoJ@=ioXm`Dj0{DoX~l-X
z@P?^m<N~YMFi(4b>aK~~cnvycWlc4_&BDW`D4{3^Q^muk&7sYQ;F)olA@MANEFxfX
z@-Nl5ah5%>XXIhyXW-`ox(N()Q~p@XGSwbpVqp`N5rvxz<avsE!}MC`ccd?U*-^;E
z!=|aAX#f-EVKY`YHiz+S_-v2FyQx$HEwE6waDxducAd6*cfWWpP}o@5SPdrZ5+uE(
z<{O6#P}p1F+de+$@RMk}8-Lv|&#|5z<ys81$-><t5T=@k%}>@(9mexDpOi1A+cN>E
z*T&Aq6Q<YK-(38J>x@94u!on2FHATrccZxv-=c7!usx5x9!xmt^fsrwr~1r5VND}V
zGnm;;2R3s)^vx^;3QMU-nZQ(cdoj1X?f-896jtU^7KEu@Q?Z*hAa{`)(77)9F6J=R
zJZ!G|u5jMglOpaHCGUR)=~Y#>f$7~<&7aAqcWN;s3!5H?o)AoUw~eCUla6n_z=)P$
zkwA(}T^3y=9<ml5HY*k@xZHlPPNlbJ3l0ITleLp|hPmnDwbF2Hg=6(Vb0t+IjbOq&
zY|<*yNIZ2}b!~*4wVbsMjCWQ4&F#&h36p_-Rh3rNgt?c8%}~Zr3&wlx@xy;c=FD$E
z%Z)9K?P0=if5~ONycDJk6xI~hl!pudj``s`yUiTrR54jum`$HvJ?wgV^x<-lAxKH)
z_wQx<_3upk0u=Vp_wazJ2IiccL{M5TE=n#|1~J${1O%|c7z|8o<~P__?C%sxGqJKW
F000d<-4g%+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_softmax.hpp.5A728FA36C708E60.idx b/.cache/clangd/index/gridwise_softmax.hpp.5A728FA36C708E60.idx
new file mode 100755
index 0000000000000000000000000000000000000000..35ae5d9bb851dc7c31784276f2d52802ebfddaee
GIT binary patch
literal 6866
zcmd5>dsvNG7e9NS6URAQH=Vb8ryCumsZ@#%Dcz}*2FX2YXd*TB6%Abwk?Tk(xfRkd
zq$ndMw<wotOwGt8ibC^MAEKLYpEJ)s`}Nm6{_%L9r}ezQ^V@svz1O?;T5InDy*)f`
z>2Nrc7J0`<F8*NwkHg`p;lG$4qL<u)98Q>k!->x>{x-z_fwjgv|C?MDYj@$gjLB=w
zf7_Oye$Y>&{eE;#-SwplJ<Bz;o?f~e_jlyDWVd;>&l6wt<lI#*8yByCttXG;7tLSo
zlPR@PQrV_eu;<D0XIE0|cV-Drp7mVr5|Fs{zwmlzZTYjS+R?wR6sD%u*BSXQc$4(%
zTvjJ<|7qptdk$E<GMbVs>J=IcJi6T2pgMcpl1vb$W7>Ip?2V%Lduu}u-8`&sY#6*<
zeYx|gS$jGL^##td&OC5-QP5zelF&l$Fpib2joBWRi>-Otso&Vny*};m$9>hle^te;
z?ddDI{eH8VWv%xdn?EP!gqc5I*;|nC=s^2*wJ|5!q!F*Xw;t#1=&Rg(Mmk3MMaGqf
zzg)Q%wm$f}#IB<yEKzKr@9=)@UDJMUkT9XDOI?5H!N?A8uG8M19$oc+R;n6oy$VY5
zvnw8G&1wJR&HE1hXyw#y$ygM(@w|K7D+g3^BxHAg*3#aM+4HYFIXr9p%c*}%S4lNm
z7gv~5p&zlkICG=Nj-Qu=WLTD1Uj6yJ$D)t`uGhTy)OV{d=QmUwuqv`CD7GIa%-I>9
z{Xit_b=w(gr+v`rwdd-yA^+)YUHz=zA>dYtMfK<$ruDa|pfPJg8DW&~tT>nB$EQp@
zWT0Cgs9wGnG;4O``RxiVhrf3G<`=l^NO<wi`c<xWb4|+9Lq9gy?^&l=((jmlD5zOA
zFWuDbbZpQDvu#;XJ1$D{jm^7UC8N)+oR{3Z&3$uSwW(^&w?~pKj801mQdKR5!#)gC
zzuEKC({)V)Tg~roiFh6{OQY9ZMZ5WOZ2zV?7ynqCyX#@bQ*rqb^C6GJUD}nby<&&7
z?|W5{T%vWd@h7L^nIVxDt2h&mD}^qflRxj!)ko7B<^<$spKGZpdy$*JtNl&;UVh#P
z2kn~PeR+*plHF4W{r;Dm`q~Te=Ed`8J(3>#_|kdDca}3#20+oWrd2Tskqhqw4rka>
zkH8_7^<#w)4dXBq1rtr>g2TsNj~}sjTe`=8+to$T0Mk+(skK}X+USmSgjh}Wx;hkM
zdb;j(1Gyly=UBD)ahcU+La2i2$%4sZxghkc=lzYRX7HMYP#Mz`1QU$qg3z<Gw$0Gg
zc8L~30jB)~ewuPY=(!o~N-r0T`9TE5nD)@|u#^izFKL2{=KpE?qX-VhbbwBPo?H+*
zez$4am8C<s3gJ*pPZCTrkqbg6+di>&tw{YKf<rJpO@G=bxgd1*po0@q!dz<yK_zt#
zV?WqlEEk+1JRBrAfB|*<enD(xgbU*6;|3yCk%3%rhHwGLPz^VDw(nGqk8Z*W5#-|m
zg7kw%$psmprKIfi#X`ScAymb*x4_#-E(qOvS!w7(as6Qt9E9nqI#c!Jg3z4k#bGh)
zMI4SWyy|?b<>n%Ym<C|d049<ThM4uxWIa^iq1GGrF5PAGKp8PU!+f;K@EBS7*i$rR
z;uH_^Q4up$Jassk31R}o0mF&W5witoZUOpqxGlhT3m8lBQD|}$j->b)v_1yMP+SAe
zYM>>>wNP3MT`B$u%|5dEB(5}x%i>eGb}8I(q`ivP@2=K$k0J*1m;&Hb0HlQAF(-h_
z383i6V@iQjDVr|?E@f=K8jP+6wzU6k;CP!ZjK?H%ZIiingeTjJCy;L|;x~vn4BQTb
zxuh9lihx@Y8(#%hSAjR-wdzmv`;ML)Jw!JDfC3Om2p&@i><fXyaI*Oal(P8?!2SZ8
zuLfS#R56da4g78cf2yw*8rMP-ij%m;Nn93A;TolIjR~)<BE2~=|ME3}m?R)c0uF>A
zfrBfM2MoG!7kw97Vim-=>ATsIttf6YD;W~^#2M=naYQ)N{__xxKp>ATM3b@e+0rh@
zdxEddB$6K;bGMdT&%k?bKQP%3CXodo<|HsVN%2VU;3C_zX~Xe)6UQWakY-kBr|+GJ
z69Zht9D@eOpb;SuQwz;&KXrQZB(7qU*gW5VQg{Bf<&cLPhStJGq$7_>;)W)171;ip
zrpB4bItvf%$@dhKK<c15D!#*FeH#Ax5)@v7x}?A3@P(WOD)VBwJUk=UV^Fao*+71^
zkmWr|T)!l)0z10yb}i~LQu)^zIW6B1p4YZv1s^ftW5N}tKum;Tgf*E4Vj@M6!zqdq
zMX^wfD29a+L<ua^2AtZ!WZJU@YPZnfBUK)Mb)JW50yem}rT0`;v{Rd6N26T^7?nXc
zic5OcxW6CA<ls)>#^Lr<cLa(MQ`82`+JGVqxdrMTNqon58-$p3z<wQYrwt1LDgb&E
zmjYDE##KPI3K&wp2A~=?Uk6Yf#bnla_sOy@*)xb~28PXmHQOD~a0gUi&$?`<eY4zS
zhT^!-_fcnA!hVV~(J71VXbWKIT)Kd17x;$aTqwzfqbbgZ{`oN2!m_<lko)iyVk+UN
zN_MQj2c7RhZ^}2okquB0ViD5`M|HB>cSA`xTmJxb9$@!R<Bm+@D*AhStS}1Bf@iQ>
zM~se8tku_H$8yGU^EUjm-#9;aQf&D8J7`wAX9eJ&ZvdYgz?Uo|sBTV?R^-C%Jj6T(
z8jpbmArSKdNL~PEiZh@_24wNOpkx<R@Y5qtJ4{hNGy{)N4p2EY+|1rAo97o&?Y?-D
zB6S!}QnDh9dW)x_rz43Dh<O0e17J#*xe=m`P=RMl_YFIw)Vl0n_d2mO_Dxw$M@Opw
zRuwe!BIXWIy+ca~G7l(_mwB|?I)K&z#g<!GRpPp=#fqzhqlH-{>mx*&MX^wfSquv$
zm?f}K3$$o~R%D%v!cSeE{x-iEhjSM-7d=+$QghQIsf*M-cyh8zlTK~}kZu4Ih_3G|
zj*i@y^ETwO1L=zU<>$7%Ju?TdD$Xp<oajT$L!j{xXj1$b*gd967%?qC((<XRdcB~L
zFgkdJ_UkC=3hat;_Hl}6L{^3UK~@;k)p$xhza`%RF^vE<0#=kwhbSE?FsY!i5?T37
zl@tJr0vZw#E?6u|*|-`QR|6Zeq)k5Dwe4F!IwHhqiZw?l>S{(yti?jpMAP)hh=>Ui
z1?iCj8!<D*Ge;10hzS-4v(OslH6qF-0p%pX!pT56nT=C`ata%#1Lbr!ehv7qfeO*L
zDWXZY#_z8j>^X0N(OY0f2*eaZqe7^_+iP1pk}s?}`Ppuf8CLOPg0?XiC!qm=27nG7
zp$PJeAS*_fKvV)1Joy^9rji|Ae%WtvX>r;1lI$Fdr->D@4qrS19a_C<Yqhe2`)6CA
zF#dwl)W|~~IHOJ&KEaYiGQ3yqUB=VB>f-L=OOczs+j!#RxCC`~^QCAr7_}KF%J<?|
zD~mPLy9&O#BJ+IMlWI4?ttPy4r}L&OlUn(L)JUy)!@CRtVq*ER>eL-p@mH}>AK>=^
z7Cr*e5w^4BLDf9!EQmP^`DY=keKl0AX7hKT#vN!tN9%#AJ#4-YYV=Y5*UJ0BKRf-e
z<*PGlh1qHwhqFF;pY4S=YHO;hUcR~f#rMr!>4?NYf4$7F&Xo6NEY*4+mm|jhxtPCL
zo%kmS-#G5b?nFr3X?T+z_+3_d+84LqWBb`plZ#p)&Z~)%L{<&332JPDth|~5?J}U^
zaEq9oP_pyW63=TFoJ@8=ljGe<-Gy4<tNSb@DdMCoyK*9F05R#nBAuOtYk+kPyQHlH
zHg)v&h#U-X2`9UnqH3N2!)NUBV>>k5PR&N{ic|&i+bY^E4M@{ym4uisVAVxS02#LJ
zVxL-ap=~ZbwIC)JI^;su<x(YdsQeT=Rl@P)>)^XEXxG3#P&7ct2B<it$qpP2o$UUd
z(5aK%zZ=?iv&Zj-4&ChbX<WNBuHsGxF=<@KG_FGbE3fzdSxff3!*xNpSGXUYRvYkb
z1Jej@?=e`~Ry*B9R>|@sHOSA1iQ-4GPz*nYg%bD)EYt!~3)CV#yRu6!S^Z*U`qg=5
z*UP5#mTOH%KkF?0ZoT(;(_3fW{_A<;L(Ea%f%^^J*xXM6dP<K9h}j0wHmEo{e7L{H
zuBPiWK9n(QBx@8!dv9<4C9B5FH~8lp!0HC@BKGW?U6MWCjv2s;`+;FUV3n1{(6AUP
zaP;Eu7sw)DOmt{WgzQB_CjNQwKby$I_;+cSxBl2z&)3$`)78+_5^=N$$KmMFS6%w5
PO<(bKFL^Y{3y1T6A!Rik

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_sparse_embeddings_forward_layernorm.hpp.9F9F7C52CFA18397.idx b/.cache/clangd/index/gridwise_sparse_embeddings_forward_layernorm.hpp.9F9F7C52CFA18397.idx
new file mode 100755
index 0000000000000000000000000000000000000000..88cb9c2df87a4ecb22fa98b013eb2293e5b16f3f
GIT binary patch
literal 8282
zcmeHMc~F$cwx83EFpMpX3<G=vvMU&N0?e?ljsikPBnX1yhCu~X)C3o}KpY`}vKeqe
z6rzBMco8>HqehL<i#KklR|Vq&1_L~C#q0CN_szU&`g@fs^WUpVl~ctje)H?~^x3*6
zI!vL6w&3wX=7(jbC9jyn=ka)B*uUi~G8P_2Jl+W%9xt={`-+)MZo3$)|If6HD@BKc
zBWlg>C{-OB_XU4@V^7nK_VYH{&6E4je?I%*gOit%EIs*Oh4rN)|0iqum#;PZrth1f
z^rU^^J6m&l_ZgpTeAgGX;>F#?r>(dC@MOc%yMm!syFpRJ@iETQr-t(zXL{a_+^Ti;
zQ)|E24Ds)#>z;UBSZ@(+ou%{biS}CFM^&Mx=INM5oNnBoHPG-pdq=O9KGXHcVS7M@
z?AV$i2T}g6*Ok+Q(Gt_HQ%}!U`fR-W$}Avv+q}JNPkbL_mz4V1$&2!PdvE=o;q%UD
zSNW=6lSZyL|FkxJL`(7IJ#l(U`K+wmH-$A37e}TR_x-Tn?8vd#y$7enjq7Thrcf2X
z7$d8Gc>R#IjrYL1&qmj|<<505<=<Iok~c#f`O?DF##0s;ant=eXcrxNowM+#xTdk5
zH}1KL_f(eenPC$jQo3@LgF-1Rm-QNF>CGxzB+FGyj%)J|x_M{G9`S&`wX@;D$3On(
z?(yd7-x;l}7-hZk*I6q*vEAPIa>Kg&gDcm~c<uE0vqjxsl;2j}lk|7Hzs-*pd@alO
zyR5e{<nHzKzkT&4{anzdn+peWZZ0ypbLg;9-|fr6D;&q_KG$(OaOJz{=WUBWJOA^u
zHmS)(=e~Dnpl!j^?N7%&x;0`JG!z@9tXQv@>Hb^o6|WYY>KW6%;7miIReSEdKIo14
zIq{-%dy$LyitL)n;h%O`d>0+~_e_vlZF*%i`IF0Qb<=`x+1li1oBs51Y@n}LekRZ1
zLAhhD?dAt-#}yPzi`_Kk-ZI?_iOIoc;;oN9)mr9VGEVtJ>(Pa>FEeMvK8=$tIk`YL
z;?Bc{`1tIOGO77<S8d_7qgtIWYo??=51aMHyRsiAfBf;#JYj}sSw)PaynZfv`>t|u
zNxm|yT-<iGKJUu>`lw~2<DXbs2aZy>z0m$;Q~B8CtJ3C*S?d+fwrOuK=N}n`L`FO&
zRv0^0BX|PV!r6~_P2Z8WI;9fx#v1Y%#(?qH2%Zif@ysN^fVuOLnAhdYv?WNF?LSF6
zX}m^|{;n?Ttaw`9%@U-;(q2+8Q;i_BJiBgj$pw@&3W-@-DO8Tr2ttSOR0+O!yL=>Q
z6ifR^eM~ii(BZfH#$M_fQtBf;mQFB9Fw+P^$0r+n*O?}EksvLW4wr`8X#}BXYLypj
zJ+Rdug+{V;oG?zR5ri&@^-oE8d8tW)v{_mqRgBjNLf6eYoAAx{%vcE$vb49<+f*Y6
z{Y_WyrMxkI>m&%Xw6E0HR3ixeLy8YF&HLS1f(%&NZ>--qjUe>7;1{Vwr|#&ELi#Kn
zDU7t#2tv1e2wuota|K#Rz#fWBTjs10Je?819vOQmzkO?dbMUgNRf0yd{Uf9ib{awY
z4<!yYT3S`H!;fL<X~JnH8bN4YM)HE?ViRf`NT{(1GA(A}UE@QE`bN7HY+dxV>m75<
zc-eV5kpst!x1F~W(Gq5S?0lTaabm{T&ew^k2s2TFC|$B;7P&-uQZ!XOl_6GwnQ92A
zhKc9I>kVdqHRlCp8o{Fxe25)j<_H8H0R`pTz_<;}C_V-u#~_U2<6wN8%eRAnJA~5l
zE`r}h2%-3&;P+21?gal%2(4Y$T9>ya?^Db?1h<DEr{i~ne>Zo$9`Nts;z!{02;AxT
zz2MyoffNryz#vSdxDr`bB3p{9P;eEROz~!9u^El0cnb>Lf)o_jBZqqAMsXu@X+)kB
z??rBVk(}Zt<km#7o^eO8%eA|UF`vl@$QWKGpQ(bdDhMZP;4@pmcMJGYyaOC}fD=&*
zpXq>*4hXe*)!J2{8~>2cbV7V5%%JUiAfX2m$ujs%KLqqcAm!hJ{4ID<UWI~GC|F#k
z?egmWw14xNd=!|Ef+%i8zKzI_v?nnQgqm2^Tw{MS`Doc;<D2^hZyS!d>9f0w%!neb
z)w`2dfft>93xsX?zq?J?0*RF03Laa*lj0rVu>(B0yPw(t(RBCSFryn1=zKk3-vbU5
z_rsKah@|87!>oQtqIdx01KjngP?QRNLU*r1vsGvg#p_YvdaiczQDi=fqI^C|%15&)
zE<pYTD1cx_&-X%L{iz=TGs$+zGO|o*VpHR$kGk`;F=L`{VnlR|8B={z4oaMm=tR0;
zX6A&M9Fzw*55`bd1z5$!6(Fhr3DFx#Ji0<n!lA(_Oc54MQG{2-L=q0!*dmJxp|NG9
z-Aaz{W!h&B?*=og-B%AAIA*dPvxm(MGdbXs1C!`NbCF>#vZA;FNgKFcxEFctMI0|~
zLS9W=+=`^Fw7vOb+rv7a-wR+B%<vgqO&Vr}RlvcvXY$L!59aw^^R)eCK-p}OnaFA{
zMi9f0ROfK~dD_3GSJd+{Qvrq*FoBp6W^RD!23Qk!!b~2*dB~iYJ7&H__)CgOej6#&
z6f)G%()`mn2D1~Qc7l?c#0C_-0mV>Uij<`&nqb#QHTz6v|F{J4nWeEyr<1SMfzhxN
zjL6zNobGqVVVBM9-coRiHEHJ2?|Ccr!l07<+K6YQzM5!0@^N!{bM%r{0-xyyhi-5r
z?d8E$!3W-0>ix-VB<r%-UtO5doyG*lBoMv&EGhVG_v3<PtlQ^-c^){ENid@Va}^gC
zfTRG%QN9o)g<M<=l433{0Z9oLmx82}Vqaaqvq=h@RU<JI=@_Y`POWrR#*n3BX4QmM
zQ|OdgIE%Bu9KbnrfuuNLvx2(NB-;&*HrciIl^~Q=?qZL{!%jTuTD8x%855U}W~ZMT
znLwwXDwwKIr(ca%8*-+1$?~Qs2V8Q1gX_V#9;`^3z|27~KFG!A!1x>&p9k;r5K7zM
z2G83xNn+*~F#d(x{yzBL=hjAFBl*`TSU8Y*X)vHMjjW5^zPgYm+mGoGshp0VS<U)Q
z4cOH{44tn7L>=Hu{iqOG71A_<nIdFWL~-1IZ+3lj>`V8*c<n&$6s%LIFBC*gm@N{#
zt6(zm)I{HH;<}N&>3`y4q=sfkQ{O-6Xa|t!0J0!wGB?{%=k})VF{}+5@(qoNgU@Y0
z9Gu?cJVN7pu6f`=+N*ufRmH^xU|Rt8lrIF^LM|=_+hQ&*0oxKTE(P0Cic>4*+J@c_
z|AL)(4_NG>wF+h)gYjdqp%Yaik4h9o@n+<{8FAud4RWeM94DwnmbGX&QembJS=LdU
zdb(l4w(Tzy{t(h~N~Dcxu~{~(h7KUB186wR<Q<K+FYG*$t&WZMM*c)~1@ouRlicf9
zjKEAS7}s)lun~ziBF+xVkf@B}qVN&x21?$^wJ_r+@iQg&ILyq~oiCx=Ox8{2pneeb
z(<siK4#FxVrWqz5;e3iW^tLRU7Svbr|E^tf;}r@uH(hFtVU}Tz6oq?*Pb6-@R#bXX
zCY<OeojRUIzf7EIOm(u_X!UT)!ps``HJoZI%O{IdPwhsAyJ<aDa%+`*H1FVm`Uv%n
zCXgd6yV^ULdu+|VKiER$g44hHGgc$m_ASVug$`7`-TuMa(<^0v@eoAr)NIgDKReKa
zN%v15_N=<>qLPccvWb7tPTke<KUe?P4mVai--E&TU_uUvT*ui2pnk3(!|u&IHd2{Z
zr(hF6tVOIHc`jn#02J|2B!02=h)v4yjfF+ZlsGbObKMkBRP1YECyW=w8<Q@F`c@w@
zDtz|shjRA4<!tMMdly56|G3X;FM*#hd>J_AYdNO7;ME2L%v1ud1Wu8$4Mf{OO0<ue
zT_D=U#SLK40A`fm4|qSfe+!6Oxb4pXKEvhDg6J%lzYh31m+t~m7ngqx_%)Y*1EM!v
zJ{Jjckv?5d1;Q1Gv!Crqu${~AM1q}Myc-F2b8!R04cz|C2sd+aE5fZ5A75ga@vAhz
z8DZuS*c}3QTHhRn2}i+$;N$17$2a*r@A+V_@3q$x33ZLR&I{P9JM#oYPiVbPGUW)N
z=B1N_7t91%2ia4!G;Zl^Vhg8jvt4Be!*~B7hLL>3MnLu3DD53%8{<!ICO=)h(%$~t
zi&KB`XiqlBhGq3bFL|uG<U4+C^e3j3AI$N+=J>A{)zK-N3-kZtM$k1$cFFIf!n2xN
zm_ZGUB~b}8*8pFmk7Af9M7R)1Xl+@9a1q7#<(EQ+_8P+nvw5%CkOxE7>C``-5FZ?P
zW47q=)mUrIg-bt=n?+Tiuj1kY5EOs`iDc}tfS{0zi$PG##U&sp;o?#dlv4bxy6vQM
zv!%_4V|(why;>jEYV(hlmw$5WCK-ZFAL`$}$mwEch24tbSKfbL|C{@jXZugEiEO^z
ze1CG1>~CYKbJSOm<r(vq`>6kss>T0)*q_YEhq-a|_n+VX-6ofBDm5|YnUe2$yy5FR
GkM}>x6}jL5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_tensor_rearrange.hpp.73B250CAB9F4E6DB.idx b/.cache/clangd/index/gridwise_tensor_rearrange.hpp.73B250CAB9F4E6DB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1b6a64470b07ae0b4e9ac33d5bafadf427465529
GIT binary patch
literal 3378
zcmbtXX;f2J9)CB4Ej$vy@E`#}1VjtO7y>A$DEm?hq)0(j1la}ISCO?DBZ)+<r6Aj>
zl%-YFsvuykwumSwt0-D<z$wFk5gqVgEo$c__9mGx_<?)QyYK$;U+#Z@Cmat~*H6?D
z#3|e(Gd?0Ul!zdRCjO+PCdUjC5F``DpUO`EV4vTZ)Lx%E3KZrJHL6mr<Dv1>--k8w
z0)E<gF6D7xT3k;$Y2{2_&A&|UDXPR7i|S%}-}FAWpW<h<%%?J$`&IS3#)Yk)-le#7
zc=d-5&1Y2qsXk?DlNoE-n-yr+btL8=2G-Y_M-mjeoid$oIYbp&n7#TU(&5SDgag|*
z)m*BHd;ZXR@%6=SHYdMBtu$I;#_T2c$mNd81N+Q~R%hc153X`uJGEuRQ1FNPO~<Ao
za+em>#{Z>x_09KnHocolX)TVW&%VvGu0szb&YtT}GYYe}8l_uT_gR<yHz&+ui$?W6
zr_=ZP7lrq)+)I2rt9h|GZXiaP;@?4S#b(GCIA!a{=~U$w+HC#R)rI>%uNiL8w-n`3
zv9X$fP?xFqI$u(wgKKm{PP2JOXpx<yOULYVes=%V_{#3RRmWZT==gbz_N~xSW_y{r
zIh%}z=XdD)Mh(`~oo^Hh4!I>I8Mtj~tqQCOKVWjy@mC;HXdA59@<KFOG5uzJjaTQM
zsHyT=LFSr!gO_(i#I!TKn|4|l(^;Z7FEsa>9;^}-*H(>8L?ymH?;VieR7{`!*|fg;
zFJ-$=Zp4#j-n_W~*>RELF!h}6E6$Dgd)6I%;zA7164KPa4?1}M*!0nbYXb!t-y|l`
z{i?>A_V*eWeeG6SRIr$3I+{JzGPzxYc${a%W0QZZKeMe^*;ggy?t@b^tK$8mH(wgE
z9Q%{dt?K2da+}3uG}H$Al~?C#nBDa~IUi*{-R8TYI6o*U=jeB$`q&PWt(W}GI*vRI
z*0KoKYBlc39olhP@#cnXwES56;>BCLOSQjpH|b6~<wtmv8-H}z<C}P-NVj!e^=-ck
z+|-~=@AAVfRvBxoV?W8d<8tm%+`tLXy(bk0OTpYk*I0nV;l`(D1|lX!y>o8O+3!ZS
zHgaz=hH`g*?rYIH%p}i`(iR^5=ZR;3|Dt+W#rZQSS@B^Lc!g61kICI?UwCN|RMip8
zThH4-vLPxs4PV+PLPaOuYes4lRB*Bf-9ulpL9&G)g`G$AOjjZ(p$O)p>|!9<5LE?$
z3j++Wb8P)`G^56}3Cg&<JKbGhvO#%{nyOLNLvvLvf+9}l0b?Fu{{4b1u@hfK>vK-_
zsS!vxix<g@CfOi)qBpg0Ho#sFMQFGMD^*q+N;X85NI>XnxCKaZL}W@k?v6T>cPjSl
z4Zjgk%-zu46k3L29)=#K&`<T|#_MYp6R!ajH-pFnc;A0m6y5ud%%Lay2}CR(=;i|&
z3>*<F1hhgxhZw#ANfJmHA$WlzNfJQ#Gm7o9+qD_Wpja#;mMKF5uv!32V31Hu2#kfm
z6k^?nKdQWKxWd9s3fdav0L6&duC2SIh(_6bXIuS^PTcDNrvMIQLa_|<3_FMnr_EYa
z)5jffuW1O41btrZaJg-6V?Sxh3!o>szvBE~s5ZYI(SE$elErT>&{A22tMt<MGKX1Y
z_=LN$&Sdb3M2rtue87gz5HSH@3FL7Juqpx8GP)3031yh&{(G?!EyzZqXbm5j$ruG-
zjX)lk0J9RXRz?>BGocK_Iv|oHl?mNJu^_u3xy10nYQB6bN`P^REEU!wn@4eXUi#wR
z>Jjy{VF2wDcu{rJyd*sMdA50OFx)6s09XaU5#sH|J$AGIOsNvH8Ryx+pT%qj1z-cj
z&K$k(au!sa74THD0;JL?mY|nl4w+CaNiRu?oO`zMZf<99#_dWoPumEY+=eRcD|Wi1
z;1?!nBuGmam(@s;BG2;c!rmg0I?lA0vDXU9qF5*+REoT9Pb6TLf@3&S_-ElkGNxo^
zvOPpF^mvJUXmJ?iz{Ro+|Grwp{%BK$<yxvfKKZ*Q?gyPn*M;$gRYf34W&%`;Vj+x>
zE%5fjc8+=R=#^Y?Zr1BLLfOMnlr0YLJc&vu7G@hJZ3PtDXS+`haRJH&6sQZuxWJGL
zR?A~6E|5zZ7r1eOC#3T|6~25f{A;5siW#jjvXl-h%+J)%0y3dkAURN5hJwk#awx(n
zLaGkMQpu@uv~+U19LgqV%b^@{jui1{oF7C-Ccj;Fvk2zeetF?!IJxZhE#Ygis%sNk
z0P+7up72rDMs?)!b8)UPJxiFY$}01xIlrdaezZ=`W?ohFjX&**CpCf*A(xg&Mx-2y
zW<<-O7)Fd7ietpdp?F5T6g5@o?QyN8WZ=eypdpqpVa?C7n{|rkepqr!wnmQMbtZLB
z;IiICZ!&c3r`y9pP4@GXOUiZ)4)pZQyt?sG@B22d$>I3umoBMog!@oeDUJSM*(n@(
zRmq7iD?a>DvnDq7IlVrs3zjqswgx^U#2<0^8F}XZhfi5d`GR;fdgS@Qn2SfFbAhbS
z<|^B+ufJeqvg8$WGZV#=of8dWT}zDFd5vjN(Gef^vYTb$(5+l-@fXUn6(vW+g^DXC
zC3#;;k@!FnZyy#F9uNXnqseuwKj7Fv(j`+gsM?xZIx7%qsUe7tloipE9aqSX;(t7U
SFX<-?L8Q|bK{Oy4LH-9AYdpXJ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/gridwise_welford_second_half_layernorm2d.hpp.52F6FAA1A972E598.idx b/.cache/clangd/index/gridwise_welford_second_half_layernorm2d.hpp.52F6FAA1A972E598.idx
new file mode 100755
index 0000000000000000000000000000000000000000..61f1acebc0b1a3e0b0f5a325417f930d26e96af1
GIT binary patch
literal 8388
zcmdU!c~n!^_Q!KJXhXOPs8N|NvymZ$F(C{h0udt0Ad^FFwWU=ME1*k!SZjn9j3O2(
ziVRwu_;ooJaKaz73eQ&XIiRmkaR3yn1;kQCOL+Qj^47^YUH<X=@3$6NaJZj*?{xOr
zXP><faghqed0QSY^yA3v^prJA4S77C8TXsDCUeElz~ePBJYIIim4w!L4?QMbIlRVm
zj*nf>`*x1|3hb}j&o!&9+m_Xl{L<lj>wUS8A6&|~(Y!u>PmAG|X|ey14}Jdnk^h;I
zA@jLy(lILewr7twwHPPPU-`Ut%ojG>8b9Bjw(*~~+f%#QtnHSMGUndjv?5dFyUum}
z9`ohb4lkKin-l)fdHLqY%e($nyGLC6(n%IoT-aWz3GrGwoG|66#o%Js6f+P17Z$58
zZOZi+h`%^rAh2`^|5iOMzWKST?Wg|615N7G>WuKVy%WdyHKvqF_ocSP9!&0wwL1MO
z*-N?aarB*ly4UUMrbRA;&>ZpEt6FvC-1-%@zY2T(Z`Qu4cE5M(xbqo-`>Q*>AKqMB
zn(rIc)77{)@JxzR(Ta0FISiTCoJ`xh?#=>f`omvteY!h0_L7V5P2b2XYx}P+NV4AW
zX@W_+PiUbmJ?7#22H6Meir%=azSZ;1Tf<YuF5?F$uYWMdHm=#bEwL!uSQ*iBDdYL-
zsexfXTAjUhwd8C?Zg{U|_O#8_$I4gMgYUtSJ*L-&BU|=a+;k|YkE{zizuctpNz1zB
zx1XgQJdvE+^u^C>WG=>a>253A^U}+evi=E!<zdpDTXII`o<6cBWs#TBMUStoE`&Xt
znl?Q7(4h6C3s8IXkiX;5uRAQ7yk=T9Oiqp%vvdJtcRO!vUa6o(DhhAE`>oy8-QU&(
zE?RW@l=<osUM>5k>cCTv*hM==F303f60K|6aAscZbKiBVe{Z`gJgUliVdh=-+i^y-
z<n@ZzJN~}j^5{2haq+*cnenA^&h4q6U2+)?o*=jM8aI8F>YnA<tQyBq#gY!YxHZUp
zOTo!l(@Re~?J7^-UKJy6liA!@wddRB?X?S~?9Gz!B@JPVH#^T&#iA$aF;gbLas9yk
zZ^w<Aq+fm!FxIQrE%nsdf;mg`owv7czPxwf*1e;vI_+0U8kZ!$_)!?4?mad6Evnhq
z`V}v0efqLhoJaGM)${p>H!L;=gGoF#N)TnCYj}Ll>$%Sa-%6P6Cir~<7;$ovWfH4v
zusj8=TyJvtRDdyz=VTv&kAtqkvXx)knm2djJB-1YlY<07cDe@3qQ{*-s(St|7J?He
z&$5{1u4}NYDE4+Z=C#LG2o9V)!(xVquEBE5j3<+i{@!aU1V>KxxA6DSHCSHJ`^U&R
zp6{?Rm~yg0ps>|7Sk4TbF>JG{caIR><K%ejcsE^x<;|C3T9fZBIARRrI5|WRVxw!Y
ztj@lp*2GrJjlqPIy#!wNx(3TTw-q=U8pUP`VGJk7TEtrE8Z1||zYX10W|=n@3@7r~
z>F-T<)ipdLK%B`Di;g6B+uch3Xsr;&a@WUM#98SYe0{6kuFl2Vt=|{}!^v`i+)meE
z`5$RU9uf6YOCe0-WT}PJQ`cbm7sb*A#kHybHiikD94Uyj)HPVXZ*FvH`Kx~p7=wV5
zy#?Nmx(3S+Yu+wf@l}kg5cr&&V4dKmYp~q&V%tsWr{+G!V8+Sef^d6XgXPSWPqGeJ
z@OTE9IRWyEkAufB>_Vr7ZrIBhwg`!fkPqQfBrc`-8WLZl`8pC`CybpL1Ue_ic4pY?
zz+49_{cUrD%15da*CB?z56<^NJY{Ciz3%i>lW`seds4sg>=@6m0TTnvv4?Ra3GgKp
zEDH|8&dIRB^Me-<3YCQhVZUbB*)wM=up=^TtWB&Fb`6G&vx#%U1I@5qVBG}{`Zr{|
z!M+=&6CMET0a_n`fB^_0`UqH$(E11jj1b*q>djdm?f27+4cP{yXh5^^g@$Y^Qnb?C
zffOAy_aS*7@+ar_BY8i~8l=$B=c_@X2I~G5Ag=)G`bv;j(tIBx?n5LV0Yi2Klp`<)
zKM(&!0$tp}BZ}jWfj}2;@Yf703zQ{b|7X}}_vqQ!C#L>#l0S0zv=?WB^GJT4?m`=q
zx6ynT$?wwKhobvX41N_uwjV|J(>#KFN05~0TfuiLP+S4g6+m5I3DK1_cSCeH%>$qu
zASN(33H*8Hp&wc}Z%vz?Cc}@xvDHwZi>br}bx2%?!tiMf+l<7`G~Yzxn}qFy>&_1P
zT@#Etv*1ZKMxa|XNdNyp(m#-r^nWu*Hv`2_!0`$A5&b1Nyac^jr+@CZx2Vx0X)ME@
zK+Fk5#ikz+^8?M-5Oa;@PQ-N5{2VdQX?}~Cw}d?(R~@tbu;nvk$OR$m`@rHU;OpLX
zH<|I+&5gH=FSEhk?B}$Mx76f=wcPpL;MomcxCef-cFn3CwljI>MggA{e0uO{L53N_
zuJvC_Ss*(+I{{zKusLBl@q{)iHYO5Mp&2SPk8mCe%0qJr=Odqd;!6y>6$!T@C!!Z3
zVIj>$NLWO3F%lNjT!MrpG?yY_Da~a_SVnUd5?0Z?9|`x<T!Vx)gagM1HGiORTyMy*
z{@(tHR6yqkyHX*XAL@#|iwogqvwd&~4^g-JKYNf>`5tG{X=~-ApV^+-)QrpV$O$Fd
zMz4+GgjC2yg@Op@A?G|4NH`za<|A*+VR4oRHVpB@j2L#ce03!Ln&G0jX#=t)+=@(F
zY3@L#9W?hM(_Wf2$W%kK8cfwd-LD)>%W19#rkX6E*j9XjE+ZpzFWfym45t7LTZ^Q%
zD2lw~MI^mQb32l@(|iovj{#-JIuO?ZWyy0OIR|=M&Tg*=C|?+!Wx_4b{D2A6Xu1VU
z@f$L1s9UI%8cpBX!FV(ocIEVy`f0?l$HDzL9T@6CTu%Z6XLFD=05#}$!2S+U@u3r<
zIw5gOirLrkH90K9J_X08;7+>#45FUV`Vcq_ftctTh|<uy%D_Qopx5U{%clGkzB`r+
z3%fwM3-q%PcbppxZ4yid;~|nhq|Lhn{B{5}nQw#rZMwgm;M55L<o%z4(=(bi;H05>
zlYzq~1HC;G(@)Bx##bkC@3$S8?WE($n%$m{8*+U{UGA|suzX_mm?x{=39QRnPt97@
zBQX$Y1D<FrawTI}h(v`n7a>s*&BaJmOmhhmmC#&@M5Q#BAyFC4RY+7t^L`}SPjd|t
z)ey#>&gH|}0*JVVQmOn9*8wYcE)#`yDc1qR9t4+zBynNbL*R0V@X8B^oy#j<E_!DL
z%J36>_vJ!fvynEtlr$g@Y!7XADQTs-14%k)?nCx{NT2p<Q!D#^nl(tGq0d)?L=Dva
zE5N>jq`ii0CD>OI#zo#(fv&hC{ck|h2BeQ94ELJSR+>AIw1eh8<lKj-oS_1oE6CK@
zRAf<;p7^;V=la)y`?16Q$5kYkrfN$_^86P_`~uCwr!j0Uh;u>DxVq#n)0URxGMTN+
z59b36JI{Td0%x~dl?HkJMQ{DMl__2rZ-qO~unEEhD_pZN?8i1APa{uFu}PsQb$qIj
zXldiqDEbbV?_eTU3!i+SF+ZfY=>PPXuWvkgu4>3R_KwHkDrk~GR~VBXW&~!;z>m?c
zavlwciqM=ZwbIPFQU}ePEA^4(o?$uHr~oP*sRrw6up#$94)Ws=Ky%P>2qs()e)XhQ
zVc2>ItOxr2z`p^g#k&D!G{8)9{T&eB0SV2XcR;_mYgY)#Q}87E5V#K!?`Bw)fxF5;
z@9kRlQW@y0<5Kpcam`yYe&yWmuVrlMCl`L5VQz2030J}NDzNwo6_u{t8!oJsjk-{b
z9fT_bwc8~ps$82=&V7BVs`tlhwN{t2-myUS#;rwd<7MIRI36yzxrm_MR>AjW*v!C8
zYGAn-)K5DS7lSmkqj51vO)_wX3p);##L<pqH;jo%uU7CG)=T6i!PA*xy+z&<yuZS*
zJ|Z6pE@T;2?kx|%8#@de<QU|JD?Ek`u@7;<v4mk?BJoS)i^r5<^FW*jdOlGdg%mrF
zeb3n>$~sDn14MoA-|DBR2frJ2b>5jzI@`hom%F(BBpD`+*X>Gj&DkXBj)03D7TniL
zzLLd@Swo<m*`tD?LeBIDfPVlc5Hno?{uP?<0skJ&4}f_9lgasg!0)5=eqj1({ZHWk
zN$bPF3={qQPq!8v4fvyH)Y<!jXrrgBPI>U1sdM2#$QHIS)Sa_<BiJ;8JvP;)8`oP~
ze|z!6sEaCix^e5Qb`zBZgGWgGh%Re)f@CKsaDVX%%w<*D%~f*QYG$=QmAM)?Z9`nN
zRq3c3e>{`8-f8!SN$(ngX#{I*gnzxLh+Do}oj>XZ9mij81<-D8lm68saXs0{WY`uY
zZlU=W65pb^2Z?)VE&)ji=u7l2zpJ6Ij!Z?PZv1hw$nC3XclU{X`jK-#odau-R72;;
zYLKc)b*^0#ohwO>jCTb%x6=MsfZSH;p^U&j*BkLVdiRR5A?^N3)Y>@gYYdy6l>I)T
zoWz{Pgf=E`TtY~NVpNFQ!OBC?c_@MC`ACsZisJrZHudR^L<e0Z$KQ%3krN9MzmVo4
z#4n<`81ai~E<yYfnoAMCl;$$TFQd5%@vCUwkNEp(u0i}7!h;oOn<OXf97mly4gS`p
z?rG^d{QtIPHuP2G8OO`tzvR}hzwWF4dFu<G>w_m7x%De`da6Fz(`XJgN$!@u;J)4k
zl3fsiT_7`M<x*|7k(K#T)?w|zWNq``j}1QH@r3$AU`1w6|7XQK^uOcmg_EaP+f22v
qw6!v~<4w_@!sGt;g+G?~fXACk4(-XIEjhF!hgRg!njD&w!~Xygd=KgX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp.2CC919AB5B044CF3.idx b/.cache/clangd/index/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp.2CC919AB5B044CF3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..01510a6fe4552490ffbe9274eae4dce4619f9ace
GIT binary patch
literal 2688
zcmZuz2~-o;8lFsGfK13nSfUd)ABz&%5ky1~h^!(}f`|bG@(3#66N(skR6s?s2xxuC
zVpSk&tyQbCD1xF7K@<?HQbD1(1p$i%TUk_klXosBJ?G6i`Oe(=zJLC2{{O%CMBvBg
zr&AEbE6y)HQJlI-4M7kl_@$(7PpBs&2nx9Libi2yQlAUs@hm&`PWd@!kum*XY;r+I
zi)8Jq;uCGfK1WS!u1{Wy8aQdzT2Ss&jP&UrNO@hl)Zq}m3%wb7<X<bZ-(PClAnYrQ
zuiRCBcOX-`qN(ZF_{h_3sg?DaCS7_Kw9741$rkzDE&E@H^QS(?3ZD%GUbykaC+|Yk
zsm$M}Hr;7GF!i{IUGa69&E$);Pc9jgnXRoms&{t0Pq2SWO%XO=;oRLU&kNxrONOpq
zyfvCb3%qvfE6;DSDN3Wy?e^+|m|p+G?4|`@J{R*#A}62Q-M+n(6nx8I+H4F<tk&5e
zv$C!9*10s)I5BmVf8#(+?LFSsDvR;&7Rl!q=U#ctM9cC!pLqHms%5s(6CZS@ZJrvu
z-5St#^G{EUd6E#V`8A_Koez%;s4ruMc>XeExn<4z{c+Qdl!!usx4qT3{B-6*>ks^r
zGg=R%J@u0nw7T7P!`MPo@8fZ)?nK}^Zs^ZLYdI^oPFqdzT8%uc6VXk_o;z8q4~-ix
z=11)2r@wf4=97M7Zq54*mW%H$@?&R}+H;qm;OBW1<oQ>H6lMIlr<<J~bDC9kpg}LA
zQPzIOjKj!0otBtH9+G|a=)vk9itM1Aae~!Luno!6PU4689tk;v^u0%~%UpFGz3HC!
zzM<tEb9&UNGi!To*O2jr;mClGUmn!;_y(nkpWnH~PcQ#rB-&`J=Q>UC*+n0Ua@N0S
zE<#BiFL~#NU9*PX;D!5wwn$=JUt=wlCw9pJNrI@b-#C7xSwD$cLDA1MynkysUdc)x
z$UW`#LHgDwLe9P`u$z$=+tidSr-aMPEq(fLou-A5948;gxXT6t1+%r)lh?&-=2?Wr
zx&FFW(5BmA)@*;R`}g5<NSMd`E$zW~_c@ffg&+AdjHiCQ=j+sF$+FQy+BR4`Sz5|U
z2`l|+1?3mLv5domLWXIayd?9jsM73BsVFju9cD`DXH=EB)k`O|liNBXcUabFKCAM}
z{DFF*cTi&7R~zjZcz@Z`n`5~X%^a&5aZQqpcSq2_=GWL$U&?Z#H0xsI0-pJfm^}`?
zh1)zPjowFIJvNlC^C;WOJZf`UN^aMhq-{T%t+tdVd9fYKTAsN_$kR64cFlMN$PLH6
z^1EoM?$V+M4Q?m1(oRlv@7diit1O$18=bATOFn-&x6F51y{<7ql0`B+J{r90`!44b
zU6UCmf6O9BH~#b~JMFs$y@o>bh}!+%f8J#m^znoI_{xDg|Gi0qDQSItk?^c!cWsL^
z&%Nwv%Fe`?ZW4m9MQ=lez87<FG#o)h8kh!C31EjSuf;QxOV+WiF*v#!&=icqQUaux
zd~As`OE%nsqw4{!jcL=B0BP6y9q$HTHrwK8FreueouLFsucBlZy$<<rA&!OsS_{+C
zQv#$V8#8*U^b20$Xb_+^F-@itAib|^VU4flk!C8g3D72LCYnlsbY=c$_Xu~c6i0=C
z#xRVj1W3yo4>FrxIR)bA20-gzIxHnX`m(T2{h_}BhNBUHreQR?5+L2}aTYW0vp9pJ
zYXOa8I711L?r{ip?vF`bkD~%W>tp%`N`Ull=Uo?!^~!`e8U|<<#$qV}(i785lEr=g
z@i-a^Xa>fhDFM=7UI(SQEi-Gz(RF}kV{E1pAdPGn|1(7w9MiC~DCKV*rHY6UVk2~D
zH4@1Xtn$brQAOGg+Agr4u<amL(R2825SvFL4wGOslmnTf_bM_2!UT%o5CD;40#roS
zgbNh;u-`FH>X4<t2g8Bk0!siV1jH)FLnU?#kpW}<OK%i*=8kHkR1saDPQZM?Y&MD5
zJ7+dJB8<JcAzi3W6=58V3+IC>(%@)tVX>$pEsho!rlpD~910f}gDS!~I2X1aPA9OY
z7)=#xkwgNtmav_Fdp7S%FZe9>Qnm+dU7Bs8QGB9}5qL~rPWOP0_Xw(a=L&-rMIgFJ
z#NeEDIO-(I-$wlKH~qE(Ba34|ufJ4(GfWD<3^-Gay^1q$A{dI``~#6<_*Fy~NyPQJ
z%(A|*7<+?3bC81OKy!iH;-cCgiT9n_#z4tgx-4Tj&{UC*u8uK$H57vsiVdWaVgbQ}
zK43Y4NU@DnME@ZXC+2)WIjDeQJ3-lCNEcE)6Yiy`7*6Ox)v&eJ4&!md3h-EyqsfJ>
zf;K@e#ioNV0kM@N;{05$@mC%Drw03CLA1#tlTcWEIFle+u_slcU5S-2^W;c>vGtp1
zQw{Lac9gh>LISM&P~w^DdITj#RoC@o;wo81f0s<OtFE7tiC)$9O*Nuf4eH@&0e6Zc
zLB(A&iMT)aZQyj+3XV(g|80A0ijCqXQUU+FHo_YOiJH45i#&z*4^>#bdqR%d7>UFt
z#%+xUq`x?jtfr1>Xlmh9Z3dm@%k*Ap&ZPLV^mW(<9G9ib+?Knq@UYgka$T^*(&+y^
r)tgUnjZIAdVP>&#5zoff&cVsq&~e^kdoNu*zF#!zv&M53DFFEkVjP|f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_bwd_data_xdl_fp16.cpp.618FF0F87C438165.idx b/.cache/clangd/index/grouped_conv_bwd_data_xdl_fp16.cpp.618FF0F87C438165.idx
new file mode 100755
index 0000000000000000000000000000000000000000..74b14dceb66377f3bc3b7157dd6f2d2b8cd7cc98
GIT binary patch
literal 2630
zcmZuz2~<;88h&|+50V!kfdHZ}K-d+<G;AvSzGFez2_TC?B07ekQY>oKVJlDpWuOQ`
zM9`|miXhr5SY#BD%^6#T;e=MOf<gsGKnl!F=3FK{XWlva@@~HGU+(+A`%Z#=JUtz0
z2;!FLb0jrMyqkg`2orqL#G;gM$q3R-MUW$vKZHaa81!U}tx!FM|GYlV>N|N=!W2Zg
z{?b;|Av$bk(KM7a*Vi9g9dV{O{-KKo`6<3_?r7Q>9cL`=NpO_K3A-Sbu+hmS&E7cP
zvR$=f;pAlf!oRLM-DV8klkR<jpL{X>u?26Jj=BEM>-gjS2CA;<v%9^=nC};!S;b~W
zcTe|=9N(7c_d0c#Elg+4T16c_`p8SoC?nwgZu|Y3X`wIHMU1w-ib)MmX!%XF|Mqa^
zXnA9Lg2NW8F`*c<Hm@Bh-e4#$oyS&MG6f-9CqKMa&Yz%HwLYIUnphX0Pb=7EoE&rE
zLvDJ{o!Q|FgW102A@PUW`+u+Un-El0mIjgR$BUB_YV-rkk;Lpyb7AC-BXg5Wn~ma>
z>T#--Su537yisU28rE|zyj0q<F^6V!&VJnD%Z|<(>4|@+pB^hznw~j7;3+_@dY;=<
zJ{T$FVj+eZ^U_-7=L?d&D-wYq{qae)^fKN%(LDC5%U#-8@p-Bw+h#ICa@M7xRgzJ8
zdZqSe<4EAa%aoKgH-@gouk%uB2p!(}Yt60(<BKuo*EP4r{C2q`?#FxWf%_YH0w0tT
zo10ql5)mYAvTizS9o=R5&Y?8=jijx{tEI-<SKZ8mI~2sMwcJ&Is-n}aZ@_(VDRjbo
zB6*p6`$Qb){J4v6#{D;gLbrlFsc)aRowF;sBHXQ)UR9>-vB^)&aEKiVVmgolCb|pu
z*=4;Q!}E`)OHvXdN3<5!*>L!Ng5a>Qah-#@E4j&(YhQ(AoAIelO_LX#j++{6y(Rko
za&S&^!{2^fwBM2;Ws8F9{8<M4k+gg>a^xSPZPf#oDwh`{i;vymnKfw&gV_h&+8<VY
zD2_`B`G)&_SKFSxQ+dDa<W`gy6gCC7a`K1kX7q-*cT{i2FMT~6Qsj8as&#k8&w4s<
zziavY@n8N<Wqvm?cyx5%MZ=zF;+c5!w8!LLVMT*Me)O{VuhltEV{Cu+IbRlVtWP7L
z<daCw*wu!Sx#9GwJk72~Uu)MyLyk$}aw+DRB^6foe^^Xdo-%TJcu8x;`D4}7&gzXX
z!wc%8HHOSiButmD*W0=|!sh+nYxiGgu6MJsHGMJUZ+$5>#>)NhM-RFwxwcoULu}I9
z5NP6F9gy?J-}lPGj?U#?PTk$6OJV)oESF}z@Z8wr4?Jer5)0p=tVajd3@kk<55GD4
z%`J--mFD|S%>sjx>@^8_)8g4<sH2tl{GpvQ0V~|i+e@yFbtT3kBTr6Ak+h7|_<j<C
zu=sC6LcJTmz|jZ<;VWTEOgVs)Ex&B$9wr(e&=ulnIG|N96*W0P+KYDLtI?p2A{^ZT
zXevh4k^`iZzW9_ooA-DMM}q;ahN-E_0n*9QnLkOi3tr)90HBpIWu_b;ozuJFmbcm2
zb~>^f&>RX!Sq_k{&0BJea%6voqoIJtFpMb&NK0CN<-DGmt;f+IK+`Z9Lk^Ix`=O9|
z@0GPbjz$7n1Jh8K1Eg<;wyYWOU5nvp6rfcxRW&(4y5IQ<W-@5nfTKGBjbk`N4v_xI
zGQ@T`UL202p8{GN(_SkFNKanhZaDR_E)++@0IiOxtIGk>GaqdZBn|rR#nBKzGcbm#
z93Z_s8gR(o&bS>%cLAD(v6ymzG$KmcllBU{k>O-f+Mgy$=MzChH0;oPB8Cy{<(x&L
z^Hq2%ws73=8^p?H4s(OpwFC#oda&LxYc$B8fmjBQVGG-ao&d2jLny=wiA)$97Zt`5
z+&vOXq4O~vj16Z==PT(bv0>+QK2?XxhWY4xTnA^vCl9R!s>pn$P=!lug;k)zfJo*p
z1yKx<0!62TFK)V8v_rNHqJu;{CqV-Gk3{>E@A8$}YH}Wkj-zxwjZHIvKBM#1wA9wa
zwo1Mn<_7DTo(D;_w`mJtO1MwZkj$_O4Y?CLpa^CJBAIO!M8}BS)$X8yflB6H1yvo1
zxbZ*P;8Z+s?T86zM3tv%3w_;4`FCo!^`j{=ov*H`t`8ka=WA$c=)=@>K1-9O5B&y1
zK|$GhP$+0fY=XT%weNQ?%Wvoc<v1HTe(({(sexG8VNr<9ClO~?n}u1x6PFztC>sgt
zMJ_HzxXC7j6IxI;Zb3b)KdoB{epc2|X2W5_y?|P>qXm<I*fJ6!S*-=_3S_b)rXbr+
zxWW?i6T@dL#^MZ=K=gf-XhR_Z)_o|^ueg??#E9bh8kx9GR?y!g6ZaL@J!GO+aea$I
zG*h4+K6KEX?1U?H*GwYXS2rO_+6Uf^;9r<1A#IE7CRzsny<*`FLgssR8M6+C1@XwS
z(J@HODuE=VChpq{mcB{;WXc*$Ntue%RTyfj-b{B_6DG}DU0Z{-R%h!rJ9`I5Cua*y
z^X-N<W_thcSFQB~Tc2ZKWNf-&<EAa#Pb{r%b*<KI=DBHVdHTem9)T`iBtPUoZG(t;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_bwd_weight_xdl_bf16.cpp.3B512EE28C055424.idx b/.cache/clangd/index/grouped_conv_bwd_weight_xdl_bf16.cpp.3B512EE28C055424.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b842af19e713123087fa9f936ee1483d9f2ab88c
GIT binary patch
literal 2742
zcmZux2~-p379K(jlVl(Sval0E0)kL51eAn*3A?O;VV5O}3Kjz>pkhm^R4H}A4HQJ9
zB32P=gKe!+p2aFE5>TrmC{LgYR1^XAAlf14`~ly2IVU&y?)T4KzW?5d3KEGLNf6|l
z8B~}f$xp*W5JbhkdHK>+Rhke~LWH2g!wPYaw8uqv@Ka{&uJpJtfm|zo|LU-e#7e=l
z03Fya=sT~k+WZfcY-X}0Z995T*1jfh@0InsZ`9bOkqI^1DF*d^)}K6XyWvQgn{4Ev
z`TGxRizg+usrI60PHpuo$I}a%)6W0;uJEKmYOPJ}CA80-)7CrrVYp}Pr*~6-yQOZv
zoI22bPs(Ha*FLd}dp~@0B5mELNtYvidxp4ASGbIP#r+{ic4jI`dd}sHA~D}}Nz&Nm
z@z4h+6}3#GH_ifCH9d0GjlEPs@Fv+zkG;ayWW#i2*>pK#_RIZ=6TRBiUW@oq`|66m
zeq^n59qHO4TWh{ra+)(*Dj#y+f8W{7&_<3cyvsdXv9**V!F77@Q{uZD+R7+LJiXcL
z9R=N9<$Dg*`BCD1En1wKImhrVV(+`G^8=A@*K(S57EtgZv}&Q*^^Ui76*2g%nw;!H
zX_@f91~IbCepxvi{d=dbB;X-`IEi3h?U7kM_)7CeyWi7m)tfWqWgTNgWjl-iIMQ=j
zdy6vuL2mueF{KgS><!}3f1VSa+c(;}?IkI`iCNR8U|)4Q9EG@yCXP;S<H`4Md|uT&
zYHmR4Y%?0_Y8nTlYVY^?-q`11Yf_wVJbs+BJIe2>>DRh7N8PymyJpAs3j{_5r_P;~
zf4urmuXJmE3?*U5v~O?9>ec);85ar7kE`}32TXrkl)|Q+w^Mpg{PrJ9U3Y5Ovs)FT
zwk2nVE|_W@itYX*zukV^Ion5I-w?PtrC|l9=)<C8RK8EF>2DJ@ot?jGmaxLhpT-Op
zCR#I<5xlZb4XNWS;lQzFyRZ8Tw(c~QBU!I*-jgnSa`*VhX3lhedtGPj8KPi(TWPuT
zZLYHGbYvRu*6%Aiw^e_rbQ-Rz_&%cH@i+csU9Zhkc>>du>GGScxn2is3fArsOsyeY
z8qHZ^_;TrDpQ&{R<2powKZ=IeresU5_>1rLZxMd{^1vv4^OZ`0L%Xt&^~fcI&$|_`
z{M|uT-nxt4-jty5$&9vWEP6_5XuXhl<=TR<OP+)h`{5fx`egk=Bzt{@F*jFqsIhtc
zb(3V(rl<UfhZU3klFU`L@9b|UWmG?zcDb~(KPY@e-)krwIveg4w0-8<1L04sn#F4f
z1=F?OC8zI<G5h@A3_DH?X%;Vg(q6Y^dlh<NA|l~H5989G<>H*oaYtcj<l&N;b8?$a
zVN7;pu=`Q(J4TJ6Q<UV9ECk9c%1L+8fFP2F<H1ck)2qn12na$&1d+MQU^}e((2`$>
zV*{H?$v81a6A47B$^bemY2#mVX2k#*7lzSf0-35ZfL>i;cM(42lugEkW3(PY&q8GY
z{p}qaMc~4Htq7Ea(GmCvSY-fxY{!g8qK9cQ85fPwBm#-5GJrmD`FrY>0q0OME*7J8
z2s(Ny18Ak!X@XUc^(itg5~H;UTDmF&=$DNG%dtl%V#&A|jHVDMdMX3xkAq?B-Gtm$
zGA;_EwF%l(l>xL=vLbK60tmuay(xaxZ?#+lK@%7WR-hqxv^C=DT~asZ7Q<o&a09J1
zNS|p;6?bLj9!Hw6(<~7y6o>|-$CEyj96{MITOS-kP+x|x#T*J02QCL`%;*mrE4jNa
zYv9pnyJ$C1#iEfE-KYSs5-gp$$UFo*%P)xQ_>{Bq2<A2t83hBkYr$6dg>}Pi%&pJX
z4+3t#%8BQZQvE*6Z6GpG*X&RGhk8_Bvlw$TMND<<;Hs9L9kNkvO$24I7+eqp%fvF8
zhs;=JY9uarIi#F;zZZ)Xg^$t)#iKteI^7Qjk6~_xh@q}B;idB6i@!SGVs4R#C<Z8y
z+_kjG(*6+xhoI5)XfqJxb3MQe<4B*)E`oAkj#_6>GIn*dV+`it%R%nUm35^xSY9)c
znYx_ZrDN-&yYrb?zjTCd3?4>ML&R_%GDS?)XkDSl9h3I1ng6*gP%@Txc0WMg?{f7T
zmni%mV2x?%YcavBBPfYRG5~wCC#wG3)booO*dZc~$OQ9*pkx}E3A(N}p4#{4M%y2;
z4i>@-O+cw2h9>E=okLJd*itPDf(D6$)Z+&B5qo5I89}KCY6V--KwJb35r?FK`#I4w
zmd-Qix5APei;UIdQtyb9^Y$TLn0t0`L8K=9WX>(;>tmRz+H`FeSWyJkq3f_fb_Au+
zDJ-zv=d-qj?$7hA!5&GXB{={!5j2^WJP)PNQq-v9rE#pGEPVD#eYw8&AUTj2>u&a4
z0^A5{0bA(Jbr&EGNCn+>hY+O)-O>lJNG2i^^^Snrut>A7ow-QXu=PBhLdBtKo!*V~
z?fznDIa>=sc`%PSm(B@xnuqu>e;#s$UFRVIESQIcuy7vofqm2nsESoL`_`MQjtg_=
zU3{=OSbg!3RFa)GdoWL$k>`N@YNC=08dr@H&$ejBCB#DsAP5A-&;5;?`<s!InY|K&
z!IDr-`~rd&Oe7=Px)hy2s-KS)l@zGQ)TbF3E_D^UxqEneeL-I&uyk2y{{OrTV-r)B
k1)IaQwz0Kya9r%<yu^%eWY6=ZGekj&IR9nd0U9CDzp?SZk^lez

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16.cpp.35241A5239C4258D.idx b/.cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16.cpp.35241A5239C4258D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4f89411caf39ad8f89c1fb36e236ce70359e2823
GIT binary patch
literal 2742
zcmZux3piA1A3w%4&deN_VaB)(W6Th-X$&zgAv5k_G8~u4WlEy0L7~#F*|O!ciWG{}
zbkkR=ZE1JgYHh16l?-Z&C>yr0DWQ6#zBBVZ@8EmBd7k+{=l}bi_x=5^M-(U&Hjp4_
zd1_#8rX)KF4?z$G`)6fK*HvjkPyrEwat}9(d!#)sdV`-+qxL38hw|lGF$dB^Q{u|`
z&;4~^`@nyBf7K#5Sg@VWlC<pUJz4vPyt`M{>$z29uSP1=WZyWb|D(a=al0)?iacZ^
zk1alY+>k#hsZDecK6h$4wRSu?rzz>;&+l_j8YR}+)?PvTJlQS1lOKnB#(sD|^|wdj
z_N$2loe!j3mZ0{jee{RnI}=HpK2N$F=_?!JJX_;3@)hUbnX<D}@zV1yXB*?P-Im6W
zT^$d3bh5FQZt~WdFRP|Sth>FR!VlUeo9S_Av@_c>U0yU@jF|uQaN<O-cD45+o~WWO
z@9S<Gh1*ES4%r5abjca^XrX+_^T0!A4`W+7F1L$wu4HE+TY_u%;w8j%*0&Vt9$DtY
z+U&^h^e!$tROhD~v)uB!Qxp3b{<_$wi+OP%;@t*zlg<KNd@!}z-TYSDySkD{d|FLr
zMy|BT{V$_PS!%zmn1%jbsweS(%o~m;SX6tZRu8_`yxr>e>}K`$6nRnG7*Wy6<UNU4
zwxadABIeP?Q$I!)hWoI#h(rE)L3D22YUi<^)c8&0`j$r44X48*#AP&YbaEG0UdHx)
zUDMrEkJQ<v)YsKC42o(W_AS3%;bmu*pKUsRoV`!vcZ2b@Ud>St4zJ7n*a1G@B<J+`
zbMj9&-s>0c%#PHJ-7~$sHz7Tpw?5@Eq3KE0{#E|d-{vK-s2A-OJ`=zEW~Jvz34MOI
zWYn(U?9e5K#-XUr-?LjC#+@^K`405~+Y{>7u=74HI!597MlpVwux)SuS+jr{R{Sh-
zFgMPIt_bHAeXdU&XSxp@Te0t!fWNbpAxF|)-+3Ti@wDssrzZAvc57XG)L9~bd{<$y
z^F5BD<4i;n_wKK2+ILldEO#2ND)}zF{>e9jv5q$uiCjM8WU~BD^G5H3wmBQh_*3f%
zS4K0J8ozQ~;ybmeGP+I3|2=PbLqdk+nn3)Ze~0_0FAt8=wqGmfFYZy~GP_+;c-*@&
zieDGYiktV+S}(*l`lhb5Y{+}2Ti<*s?%K@-p;wj>3LJ)SyVE95EkrUlmzZ)k3J*0j
zjla1dS-0&OFZ^-IWWOYJUF~~^d+{mNPp4h3l=cUPjTm?jg+b@SyaRX7+<fHz1G8qy
zdP2^0txv(3`(yMz!P{ZSi6PDW6;E61cI>V~FHMBU9_%r^@<*{aGj-h2JtX3A!OVHN
z?Y2-lD<a7AsLy?qhL9=UzqHf*psc*iWETwxB562QZYxc$BICj#2o(?n7Ak}7u<A!$
z^->%aaG{Wl6Js<?fGH{i=(PB)f6D1417ut%Mw196ipl^wy~O@9eA+34j0?kP3V~v&
zGJyW}zHMW`!ir`DipS_Md>E`UfIhZo#w*T?kx$00#AqUcNKqL;pSb!R<=TLA2pJcJ
z(b@!UeU$;U!ut%ty2s`;85e=kS_CaUl>zjt2ENr;_lYPnE)t`42s-*I1L#kKp_@J2
zIn8982%`}KLQxq&OC@Wv7A$}ueAV0h*ZtP3H4rq$Fvc1r1dld{-{>NBVsIoJX$W9Y
zYYo!88dJm_X&a9t7qGW2gcc#7XrT0X(!7!*s3mM^0A4~+UqfHZS<heWzZ%Gx(;hdJ
zbM{`<z@s92kq3xk(MW=x$ltpFlQS2Z2ZOV`oanaCnQM<=u(8lM2*7SRJK>i$_4hDX
zpQRrNV88N-7ZFl}J`AP{>FSsT)PE>P4K(vH*idMwzFSZ-{wBTQBdUp@T2w7MXbpnG
zRG1F(L{K7?NC!<1s=8j<CL7hpj?kDi4$#GdV;ax#7)*xR6X(1dQcQf<kBLX%BMd;k
z==Y87o|Qpk7)%q=)QQEuQdGYDv;7?g3%mr8plGW)T=T3Px(#s%Dx!(Z!6oxc2*w^q
znm6tUY6V-VTLVJIx~-giFgjlja%Qe=Dy+eD8A66SoQ<wyn^tya)3JQ12-OrEMo=Tf
zXpU!wn5n%@xnB3pT03U`r@tU%Ojo&5K(|bE>ipD;%PE*HiAth_6sk?9D*o7N_dN#N
zz&2(;7qmf>G;ik+lm)ZY86s$aI6ysaU>~s~%4Gz)2+DyuRB$hX28n}`K(SBsj3sl8
z`mHf>6QPNETuwRS<lG9x8-taD3vRlApUk}Ld}|CV5kf<lU_}vBo2JbK>IkYs(_w<`
zzL>Tv<UrQ48th0MHEuB|4T8o~<L7v*sH@ao+bh#3V_BH8KfWAa2Oth+7RyfgEdels
zGGV6vY<5DiFcD<e8BCN`dL$2EH<=1e)jI-W!){W3c4lv~hOOt+DOen=uG9OGzI|T|
zt!8N<=n{AdaaN8CbLV(Yu+tpR6?UEDxxsF8JU+~y<N3foY7bNui%$8~n~jbQv*$EE
zP#makJS3H5Bq?9alBQ%WRzAck&3}I}-~oc7)sLF!R*ysE6yrT3n$fW_P)cTM##)RF
zl7wjD7Z9{yA{o)v)71%}`1x8>NCEnE1FDg+tDC!rr`Iy?FKCPSRxS%I{vT>+YQ|t%
kve+COTRZ#3j!T@JmzwiT9JtGAhQh!&oM45IzeX_hAO9J@$N&HU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp.BD38E20DB8B3998F.idx b/.cache/clangd/index/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp.BD38E20DB8B3998F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..27771b24a7378dca1ab2a22a5f484cc595887991
GIT binary patch
literal 2932
zcmZux2{@GN9{&uD&&<cnFf$n27{-#2(85Hv$zZIF^p)&O;$*!MIgjc%r^S&Z=Cp_?
z=_YiX>J-&cH*F?KB~H;IA}Uhd7WbXG&pW33+<Bh)Ki~iVd;jn6zrMi&K0ld^ARe)T
z?3B37C=~=jXyBibnZE9XGJ@PvLy+tPEuy|3`W91rX6P}Ggue;3JGQ=lSkxp;&J1}T
zbhx9$vx3ud;bU{elPc?L+xL%gl-@;Wbar8q{u+#}+!%M<y8P0?x>NP(lPwbavxYNe
z0po>7CQ9=h4AvY#$|YkXPoEczJRO@<OHIos@F;fY&j(!k`+e-xq@F?kx24y$U5C2!
z&Q^r0qaBm`dpQw#Mvntty8bXZXl-muTlDPhp^BzQ3p<h|+z&}t@$>xhjQ02Utt8u0
z@X?$qyNGK0x>dBYeSW%CN3LxAc=1WdiNDg$Eag{J#=W`Jax`MopKe;i+xYdG4z_`x
zW1SMh{~b})SBQ(}>h2$?%c0a}sR#Jqs6!011PL)!Y+o@lrO!I`&))W@-N#Oq@80(M
zfIl@cC|zssu*Le}8|Iz+{Pqq!nEMOMbRX}P&8{o=s|E@Zx4UOtQ~vbn*WXKaASX{3
zsTU8Xda$-8?dIjB2C0mGDX1Q^*v~)sXuzX8TpYjTi&^sm+S#V~zK6q6Mjtv7<EX6h
zI-})pA~Knw{gyUm?9koU!x<sYPHsi-4v{Fe?$*XVc-qO1x7&`<x2pFso%;jhSL`uG
z4_*307>9^-edqigV_p6(d)m;aVM-}=xa*|Fo;{{MCD=x`duHLgy>|k0z42GWk&OwP
zQw-i4*>oC*d7t>xW90FZSJS~EM^>@UFKis@tQnhbrW|<5b*J6bH1L#I@u*#;cN{wV
zJ2qdX^yE!@UOL+2V3(U(l&5xL_a9uw@tamXyXF<ocsgw1chraPwVA`h&pg}sbcA}K
zUVp<Vox}2p|M25jd4I3Qxw>v4os%-yzBY*H@bN|T!m20tYQ|Fp-`3qcd?$3?cBkRO
zEjzilx!3R32MyL&7Ti9p?s)ZB<UM;mgQnnfR?Jr4(I3+~6P^985?{9F7q2~f@7#%d
zO|A<b#!lpx4i=VjVk<xAyhyI$c7|-6jJBS1tgqMDpz&qKxYsJXt;ET;@3PmS>LITj
z!Tq>^-&6jTm|t<ay1=&e^x8_sjYho1J=VjxA!p#j@xEuP+pbs^wYk=13;(iO8FcQ6
zO+&(IJ2$&XO}o6LUCry)SQf<WaKvx^99|=SThQ$pzv%wY<zjKpt#$EctZu@szGTC8
z6aSYcZsWx&2a7!gS3j0EQHG=N;PsobK9`hV*c|BMzjNnM&$?Irxl?KRr?!lr@4T|7
zR??7lE+uWoGFGG~?pN0g$qd-|P14J3JNm#%;G<qGyv}Oyn<ld=f3ik0a#Er*l@Nra
zG`}LRI9g051S1G8APS6R0&KPHM-Lwng?cw`Arn>t7$u@KnSgNOn%oCRw2PjT2|@sq
zh$Na!Ksc#r-WjycKAB7i0<acQ%UC8L{ENi0#oMCnGKQ=Lut-IO$^?WDNoqe6pSJSI
zga81OiDXTgfN=GWuP&=y3^tJoA^@uq)o3yS;o8<;X>HFP{m6tc08@w*O__jj*Rpz|
zS)X|wnGgbCRiY|YCLnyb&2r#bgZLzw5C~w5h-u0MgvT44OeP-HhLQ=P0HzYDnlb_5
zZ!d)Z;5l<IlL;#TtU=VE$pnPc<Ki<i=OBoRcyiO5L9<9D43A(&n89+OW0wQl?~%HJ
z8Aov@G{a6Qk-qB?O>{RgwFYYhr;Yi>ez4@Qamu9c8iwH}sEHOlgyFs--$;0|A)~jU
zl)JlCNd*s^7si9UR-<dF!ONHB0|GXm?F)S!vQ~D^q$E@VvmRg12bz~;{e+H~AHNOE
zbPinr&7P%G!{F(A49vQGU3rvV^w+d1E#*zXti#ul-|bcO?rl=TmoQ}vS7oT`z+PcE
z%0P8sW*Dx<P}6}^;3K|R+_`m39k^(-wYiWN1jp7>(2Uqda!r&qKGZe!X)hpOsj^ZF
z<{Q@Ba%X9U&jc`Q^R?xP;p1HuBM<I;0A_)UU<E8%_}#@hCf1Lb1Pl*jgc-t1zAqtM
z69VbG^}uj0%9S?*h79^E-4r-18(wg~wxw-32zZV7M)Gh{7f+;xbZ2UVe3=+i4?1Hw
z3u7s0hM1vTOUrhV=-;^e^?&*cLk7Ik6@<Kdf{&e=9zGKTcu5SB4ou-Uz3H-RxpSL=
z*%GzXhrDoC!MI5m6<&hjUZ@ux9>VZIQD79T9PAu$kuLA7iz#X<FA9eHiTvaf3U?Z~
zNEdq6#T+$PVDT6E%UNpsCZZSU44Q#>_4)ep2|G5QaAZLlwhWl1vkq@+RC%9r%duku
zR8@na!G?Q`;S>gi4cRfA%Am606VQ}c=(jJ!^&oJGqDR@nYGQaaJz7DFp~uKIIL{yr
z>B9rbF`R>PG-r$DBl3xW#e(k<aFISnvo3s;uaJ3=C`g`px35}y1uyzJNYQ|AAYZwG
zT^A~i$vf8rGseKQ;l(cu(p`K*H@E>anL*Zp{e)QpUg;}0%WHvJDDVb|0_40OM<18i
z3`KHOG29-tSDVep0d-K&98pIF%?Wi<(40|c1&xRD6f_UiL$1O8fas);>}+&qsF}hD
z2}MHr2&Km*M@i3Rq{n2~N<Y>}#eaXR;Rm9Ox1QZ`Q5|V_S6;PdHL^+$ATcSi$q7L9
ziStudnL|`X)ySASm7?KI^K>_(k-Rmvwdgv!i<da_mb$nuvtd{{nJlz0`hQlYp1uLw
lm}AN{x12lAcD|jx<03-`mh}P;29qxcCwTe0Em!hI{sZK(4B!9&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_conv_fwd_xdl_bf16.cpp.2A5915DD526E33D4.idx b/.cache/clangd/index/grouped_conv_conv_fwd_xdl_bf16.cpp.2A5915DD526E33D4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0784ff7125a1e255559fd6b4a679ec57e61b0646
GIT binary patch
literal 3254
zcmZuy3piBi8y};l!#TN*VGtRHVH&wim?=U=G)qi!4ac3#xC~kR*)CLCaxGD4*O161
z-L@MQOOZ(JR;xl$tyQ}IY1O~}XJ(&oM$fb7dFJ=L@B4e-^L_94{oXS?cURX|1czG}
z>7Fi#N)20t!{Jo1UrMSlZX4#XlySK9svcj#rU@IB*`repIz3vnj%h?c-F?Q&e!2Xk
zH^ZVE@j>bVLMqE(Zu%4LTr!k8(v>@CbS_OlZf<y{e<ZIx&$8@A|C@qjdjDO@_Rqqr
zhG`jZ*RwkvR)2Ny95U(aD}4IiI>YK=rA6hkWivF-!taSg9^bXS&Urd7%=$7xC#;?6
zoq9PNU0QSNSc3vFZeGCV#*7WknArZ}H}!`GvrTTy+&mD~K6I@1eCnDBPxFb9<{QVh
zq?}E=)6?a1HaV-67F_&$-O~)i%_J_vr)*f~-j(e7(v+L8GEXL@5Zjlu5^E}gOJ=o7
z1|tWFN6UmxXYM_C6H*%diT4);DjKC^C|_vTP|mL_uVL|6y(1@CMveB}YwmbzKW^%2
z*cj??+L3=cc76BO{`0fE)Z*c!3uhlV))ibh_@OqE{Oq!Kb!EX;?ymn_B`fqW=KNo<
zZ!;nMiI&EK?N;lH#{Z1s_&>QU$hE!qUzN1Ik9~xG(Sv@~H@$9tS={GrTM#e0e=trr
z{*_W*>0$pK#yd;e_R>}SzLv>KqT^zhgnZ*Vk)e<K3zhV8E{_bimv3=j@#tDRC-L*e
zW1em8*11M1;XzwB?X|E7h|Sq??&9H?KP=J;pKWOh3F|8f?kXK{+^}oUE4eZM{n^pY
z&Y6ykNTz$>ZwGrRUtdmEWOR3VKaPDe$9WXO-73-?X=Vl`pXi<m9I=0%?#UmWWCh%_
z3-MBBD!Hs<rBt6)eJ%Q%zo((e_jrZx1xhE%)%WB=ykFW>tlbK~@ZI18#U(oQ$0V;a
zX*r2EBF0-*tZ2K5)($@ycy2#Qe|V$(JT1qh0%~~ka}?2%_=Cfd!ncU~U!CRt`kv$?
zbVJgP%u)Ve4}P0M;0E72NA+9=T$*R8FT1sb(b_PxY?hf&we<SeNxR;u+ng%hA|=<4
z6_k_)gNQ4fmzJ3i4;ARKw7j#b>UOVMs+6pK>Y>BC&9g^ueJQ*4akF5)qA|6`#Kh41
zzIQ?Rz7|!DPuo(N?7SDe>-EjM^9FP6&QI2i?S16zR+F1Q+3I3%s?GGva#F97zh9PF
zJv^>+?Th9ISL2HoW*6(~iK&C$YYt@BHz|H?Osf$+s9nA#?L}DAl0mH^joq1@jiMvH
z%5Ew(PJK29klQT*obQL_`@#e&dlZUG1+nph!lci7*T1gk%{?2jnNNH-!dUjCH|xaQ
z%&Gl7_$%Ij&kFct#Y0W)jg5^Re?F~Ec*@`8(Z1+OLe=2II~(8wyid?uVk_gl^76Z3
zKel;H%uN-I%)cC)+Q8O6cKP~y&$~pe4`$r+SzSHFo#RGV<92+?98oXo<x>C0e%@r^
zJAZ2AwEk}g{Oe+oxmQcVteX_%Jh&5C`q4~&SHJespuke+fwjd;TC+k*#x4!bEb{Xq
zUm88#*7D$oDwC|DR6E9M2iv(cvZnJ-G0gswm_CC9;e@UIK-sHMrJ=X^I~6`yym_w#
zt2N(dZhmd?^Dr7{IvQs1;jrmEYu_>b>J2{0jiM{K2akSq!=+>h!aeXfobpeN;U&eE
z*)X1m!=Wn)D-9)peV-(cvEnY)R6F__#(QG4HbGlW0)RH35Q&5i3ED8;7o+tF`f3sY
z^veE9X!?@haTxD`(OLv8H3<OPDo5q^wj(V;Fg_TgsRXLJ1OUx^_c1VIW97c3xJ?+%
zATbOi0BEN(EsE)Ieh!TH!)OzNiMj*;y*?e6pPYEn6UK*NG?_qFlK`N%40k7Vwz@us
z@!l9sC(son0O+(&PNAjmUf+W80T@jq(9|UW=)57*vu?D5*AQF^M)Mc(AqfC_AE&-W
zKuGF<@%|XCOVCx50HDi}J4L)x9lyeOAB-juNU9P5bj{9Z#t};0r(k>_Mw=1L6eIxX
zlMylda|{NiVSFe?8xxGxB>?CS#fg1ImG8V@{6>t{A?RpI0MK2(HmV;qjURyVn=#sm
zV5BAiK;Qp(uliBY1h%y;7_CXrRFeRpKTeO9m_J7;Fg^&QDFn(A2>===jEYW)!~WBN
zs^k>^RZUn}EG19|RD^}a5&~hM9=zf=szFc|#3F(ffQGOnv4B7=g06wqXn-VU2TSz(
zVS2V4g3{PDS8%yNU|6kKU9wt7$Wa~C+U?YObFyAEj@4RFEL=bss1K_ZOHWp757|rC
z1scSX!~&Eh*+RC`4m@m26G@DM;zbB*#5Ur93qnvTo5}(0A*eCim;<gJK~2~u9Iyo-
zWqHMKB#fX|kQEH}i=ft!wbTnuuS!HaE>A2*P?8>r3IvIuntGa4paBFW>yfFT2L#pC
z)1?AsBd9)Ee>u3gG_K14-QUIwJ0pfO!v~xzIA^SV@dU}*Ukk03p3%<i^SRf?pIi{v
z`Z9t+EtoJ&6XNj#wP1bhmEy4iVFa~<?4&0O+QE{<!zD{v1+9|aE0{9u)WjnuOR|J4
zrRRj8Oo%D<n$s^W_7c>(Vq2vsQAl7Iz^!1*iKkKaCI`qt+9nuFEJ-|*vLrUdCWA}r
zWbz^xb84rt1<a^s46sEY39ME;u(I0K&}wPZ;MTAt@z~0eY#<wH($qyochZF4A#4G@
z8J`Odc^Yai!n02nV_`a#ZUF=Z8iW<UnQ{d6g1i*Ko)Od=@?PNaAl?Gc5As{!`9uB-
zyZ|U*ffoXWEbv01&;=d}p$ohiC`RgGUmSjzxYrb6M9O`!ArQkq&sJdIaFJ3|9&JxP
zuEv}4)lyAJOiB)qjEIVfjpqxt#zlwW!lh}r2$>lxGh<|Cyv$6Jnb9&cS!PPM_>UVU
zi^j=JzRXONnF%s8LJ-LpFTNq_JNZS62}>Xa7+I>KtmLNZ;%uy{=%%KrPShgXu~x6K
zcUbFau0dxT+R*g>ElSeSUAEkSVnj7DHCwUL!qR$`tsYaGVYN<!<mxWObG)3^%W-l4
E2dIZahyVZp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_conv_fwd_xdl_fp16.cpp.F0393DD4FF0F531D.idx b/.cache/clangd/index/grouped_conv_conv_fwd_xdl_fp16.cpp.F0393DD4FF0F531D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d48c20886a0efc558de70e081d3660d5edeb72db
GIT binary patch
literal 3252
zcmZuy3pkW%8=eN!$M@wp#+fw+lWF8IVWvnjj+>YqbNX_o8HX7wlr8C;9DgD@{DfFS
zrA;kcNlS^esFY5lWEE=rWwX`)&FuBg=(={UYwqiL-ur&v_j%v<er7oCE-sA-4!1ta
zJ)IYwx<wX;!>M3jK`K8!6-$^(I9z&Vw?A)_&_;Rg<Yb*rw<fh?3Nc7`pSH69PUi9K
zV0F*dV6`B=G1Ktl)B^Ql3Y0o}eNR8-Vwyqx$HD1aqX$|KSQby*n$1n2-TIxp^KbrD
zqqK~7uB=Xn)n6UF226W<^IrUEoniH;+@f4hZ<^|rw~aWk;rr&9oEM+?JH7~MgtgN*
zC*RD)6rH$JTPIJ9|HNZ?#y%aGHnsgZVDh0lqj||wbMwup_R-p^OQ~ywUgpBl#-7>`
zL0j6r?(4p7DLc+mLkosazsNB1BYDz&iwAW^e$A>W65M{9*^(?Ewk|zOJW&=}IHy_I
zAJtDhS<GK|VdTMVSW)N#=MOTJ|Ad^O)ZVJDbog}X2_}bmbF_s?skiT1bI(hAtf9MZ
zW4Ob4NACGJ*RF(Hm*zOB1%t`$Z4Vqz=eC!8sEX2h`HRo-^4tW^-P^Bf$#>H~22QZ<
zG9dg<P4&4utz7e;y^dxFKL3Tc$9Cj@%4z$@eE9(}{Q<{sd*A*d*Xv}PyS4g$NxbgX
zw~7afjs<qp-&<057Omp;HjS4P9p#)851X8>Hu80!P);xP<Y>XIr6KMs9$#x^C;ff7
z)~mVIdJjc8A~<2wehZ7BxSU-VFCUA2XpxrpGNd7FOK)N5^`bsUkKOy;E_oVwBrB%T
zDbtZ2#c<y|RC1I2_04!$M%Q(pvAE|S*^k3K6RI^v8yUeVKXpxS9<_g!?!|pF&I}r{
z3-eZDC_1la3XZp_%vAry-B;J(f2Pd8o!p6f8hCj^oS&POt=;mzaNXel6cp;v#z@{5
z(sGh|BA+#_SkZhNtr~pL_sV{p_Nb@y5;e!P462*`Z#2=8_@hHb-aEwokIvFS123%#
znh|MN<`ZszH~t6t%^v>uPU<`NIXBK3zv<HC$7sWh;yFfQ<+9(tj@#XwyvwfC%~y1p
zFCz=;3?qMKzp>1GbTn6=sp+$$^7P(S%M??z&pmQ@?>Bey&X?kA^M1TfW%Z2a;^M+H
zQ&gAM8zTn}eP*$=^==62g`5#5=U6A*5dOf-`b0y@#_Mc;K-CH2vSYN)Q}X@UQ|=x8
zakEFzuW+W3TBxRa=sk`CYl5nOtCQom_p9&f0Uxf9viX%iErw-cbALPJWdC$%aD#9u
z1sXUNrQ<<(J<?cp<#qOWiw)Ez>(*;?q0PqC?#F!LSEKp`K2PVfZN}TuyyVEv;qBWV
z-i*k8dS=Ih%|&Fd=kSA}p`ozU6d&7^i}kgEyOV9bCV25Ravt?rd(su|m;SZ1VA7)d
ze(k%353andxtUaBy=9x82s7&96_Csl;}D1E5pLnhs?Qi#v+Um<bQRVV{irAK)1+ku
zy?v;Z;IV?k*tam3)1o!JXIE`aN~e;ww#}C4Bm3xdcf(f(Ju(3f%{?<?vD<dj8gfHt
z9^UuouI*RrbY)eR>Yc3EekB(+ylbufvgiD`vTggj3jJ5Tb1U~x%y(av&D5OF*e@sV
zA9t)_TiM+>izebsi*n{2;^#`|-}AkU8!inEh99>&v`dh|i`an2;gn9*M-&!VX2Ez4
z4u`HJtTYk>_Uk16m?-RKPPU>mFy0HJwF%m)VgR(cu)3P>kf;sg{W02rV4x}nK(D-2
z4ozJNI0NH1V6-MdQ&kLrw#rey`$I)jFpLkyXk&u0niv4R?n0A7I(#?>#s^@uDZx}t
z41jh`#~n^dy6gqx!!TNlprtAXK!*%=C3c>58H4dY7)>M4<i!B!w1suyMek?s!1y4H
zrV^-XVgU4kf#q#()RJomPJq!|SuP|7Kp$k+H1P<@9WXu+qjd?os$u|iDRM8LbFSkC
zjQ7Q85`m;520&L7O^tluclRyBZNg|eiEbzcK%dBdX%eZ}bq>aF#%MEwnY<VP-4Ypl
zB*(CC3dV<Hv<bmPO$>nUP!Jx>FMsb1<2Pco4napl41m6Vqh76K`PM!d?}yP80!38}
zfWALJa{O_y5UV@{qcsQ`s$u~2{M3^|^H(St#s_0GnLu7D20-KZ(J_K}>>mxdYKg+X
zx(OQ=SqV%7F2aUI76QY-J$Qw0T!WxYh)D!H01sgzk%7Qm1YHBIQ3oPM2eW$k7%giF
zf>K#j7x4CW*_yL0HPz3sVGFW_Gf)HHVM~fkCtcDWvX?9g+=q!o){}~CAzMjt8*Gb`
zNc7wSSp=o9C~R=|2x`nSW`oKR)P!Zi2Cs^srYuu7C;^yQM&TRtBB&K)1%rAK)Ecsu
zWNDtxeY8LuZ|qQXC%P~28aOjnugL7ude=g0C5Ot+y0quov*+!ix&HJ}Fn9kfTbDVI
zt1XA1Bz+PExEnz=^ff5JF$k)quSEgZf}pzkx)jhR2x_2ZU;v&O^cMEqqQ(NH05Mjf
zsG-0xg4#iLk{bnkV<J&oNkyxmRchcuK;4)~)Lc@LBjhL%fiA<YR@7orktJj)xl9CQ
zKnzLNn0`ghn^)z6T@P82OajdSo(|hi)Q!@W93ThD)q=LfM53OQidYa!3l!eT;6%x>
ztER9G%#6+GphVysY_6zbrE^z9t0gxB-T@Pd+EyyEfovqA$;%4vBw@f&Yy++t*Au+N
z6x5iHXSEby!!%=>1#lfWAgu6Bmm{b*<Sh?sMo=HfXHmw1IE%6XC}2?*2n8<6f}o&9
zSr`<yC<}+e7iB1fF3Mt|SV@Nc;&8W!&MM$X3hs*no(TT=B!B>ii;{*B(lAmQiWm7O
zjgyYXO2e(vFj*SLNW&CKD1*lAuRZULW|^oYCMBmtL`6o&#%<;D65?aF;G!imT)Z^o
zO2Z^+m?#Y+c~M+ZH6GF5%gD+RmO}C{vP@Y?(M`qK$wWoLO;tmUsHtVgT)oELVXdRN
zI*nyyLpAufQId|X-gky%im~Z(vlS~XEUj1B>NB+IR_oPCF7A9h+k2hs5>MR!0hl;I
A*#H0l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_conv_fwd_xdl_fp32.cpp.329BE83F917399A6.idx b/.cache/clangd/index/grouped_conv_conv_fwd_xdl_fp32.cpp.329BE83F917399A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..132e34fee53030c83a44cc4e5d75e79936f481fa
GIT binary patch
literal 3174
zcmZvd2UL?+6vso9_`w&MAxs%z1!9m9ghZA|7)BAo-eIqVkst*sMFm?e8w!GgR>Dxr
z5Cs)0Qb4FEf@q5h;sO<lfVGOKt@cZLUV`T|=kP!GzTbT>_kG|02Bw>{b0`*t+7{)O
z!HG@_UyDMa<dHvK8aIA7l2Bz)sEnEcUyj%GX1V3#b4}_4D#Y%2tX78GqPea1ng_4O
z_}3DH6a%>WRGoM8ABpEvp|pv<yb-<g>00sc#ul$n9Be;mR{rAptAbS0^*e-|zq!BY
zrf2@?Lc3zO<+GjpsNv91(bK;yGR-HeO{+I-SR}d^eT5s{xub0<_vr`jo=?*xjQ!%^
z+{@*dvST+Jo1}2@A2>8d?4!{|L#qpZbN7`fZ4!@6OafVIha2lIrrAupn@mqMUuz8J
zb*A4Q==1JO-P1}8DfzwrX{N3Zo<a65A5$Oi$!;j)-B`#vlfuKbuWQ8}s|qPyRw*5c
z8o?be=Wgv7zxOJ%EaW5e4+2#DkdP_c)vhF)Utf8Q%A^iXoT2KSvhBCI?XEU;dZ1}n
znB6%C_PIEh{)FonmzilLV<}yo_Z;dAx+>n(MX5f!>{(k~kigjg^)ISY1LSx9FKD+Y
z5c;Q<Qw2HZF2#>$qv`%nE_3p%#^1`Ne>dgL^@|zttG(fI<CEl&qg6p7|87ORM&g3Z
z!LlR%1LW6c#GEoq_E5`AHO@iODJkEep0Dfe_ChYBlEGAk+be_JHa_TWrzih?snNZy
z-6Bs<E+Qzw>pN4^fVkYf=Pw<Jy>FUc^ep&vX!uZRNMG5o!}k5(E=WA`FU*c<cFb}h
zM^W4Yf2$ZIe117omD%6tITiQh9sNNlBZ03x(M$<S{i%O3aKiR^hCBP=3^icfI@Cj!
zBIC4;%B$^^U*f-Kf7^80_hglC7vTzvq2<nim>1gQEL@8|v0dRGN=ns9Q+SV#^xWiY
zk&jz8Zfv{3svEmE{M>eiG<mJ^A~Dyn3Tk?FA{u9g`^N5Q(VtkiKdw~zYq_f)CF$b#
zW<6w&44`*Q1#b7feO%LN*r|D0|7E`lH%1MnlrK|~YS#bydB%Eh?iRg9qgcjyrHa67
z(uwS$zckC5JY1kjRq@<YQ-8p6y-cdw*-5+CKFh~%ek$)>@!@=^Iz?$KmT{?T<*N)R
z1nC+t&(FrC$&EG?F#=%uTB+2OA=@6=t^*}DQp!Uihq9QIxYm{Z2G;9Z4d2rEy;l4G
z*)&_YBQ4xJ&5dC+Tf@&Fe%$V(^vP?%#-*LU_2SEses9?tGfk58+dW<`R)tz6+BUgm
z9HQ!vLFc|-tXo&sKcS6P)YhN8H9n#7WW~1Vb8-FRjZV&+M>Cy4w7$akzDrZl>iLr2
z@Lw-X-ZRIU_(Xbnd3`-LK%csLzwt<fRNs_!?S+O4Jh5R}{;5%L&~q(^#(rGb(ZQuV
zTAgq2_uQ#SqwL5meo@<KWIF3pnesj_UnjfQZSNh_ySE*NnuUkAf7m_j6W;tLx--`P
zq*1%;n9Aa$Y+sDGBKk<>;Cs8_xO;gOmTKoa^bS<zXv+q_zf9=)>iCpY>q+zU?%1%P
z_J6E*t!U1r1g7svT3j$U$#WQu`$2cKY}<pG!Q`we!^s|vW7EoO6lcR9zu17+^FI-P
z`eC>uzqf;Vx!-Zt((4URVW&N<;^b4NOO8bF+3zz@yi88SPBaQ7`{SvI(h{?57|leX
zSer1LbcKTaJ_$b>5nkotnJ0|aqr4DI#*=k~0<eMfe(GF1YY9fXBUlZirXUo6O{V#L
zu3eHEjP^ya7Dh`!C;)G|UJcFv>~|7I??kW)Mnyp=0GsE^-P(P$B?v}`AXp!xuP79N
zw|2BhXTbTnFxn5nh8ROdp#bcXfyz%!zT^(0LlLZsQB@EMz`<kvNmp8(r(m=vf=L*X
zlu!Up|F|`*?Df)37#)CMB8I3a6o3zo8g;r7D|)dg9)j6x*^p2GK16S5;b2m_VYEMj
zH82_qLIJoEd%Ku<w)-lK_C_!sgO?Wyz{mDIGl-PwKMSJ+5p0YxmJ$lUXCh+@b9IL2
zVRRUR4KN0ZLIJp2diqdt^=l6py$iwW7<FZ#0Ni)=lwyTZ;xLT%L9iZ1PeCXE-(4B6
zeGoK_v=)qDWsI_dPyk+;e^_eroJD}qK?o*b2<wCbFp3);!&5{4rvX<<NdK>!(9&PH
z2+MMS92CJG;3;IUz);|9;3Z^PU?i{%JcJx6Fc4VAvZxRh2aW_zLW%@75*OJ(HcFs~
z(oN<6euR`Qfn^bCL}%at>wPM(&NT2JBg>`)QzuXj+=m<~u%7rxTgX;)B=8_oBru@3
z$O^I&HMi5MEE!KOC|QeT>CyD);00k>`ZRqy=p4&3pc&A?tH-hoX@+#r0<iKL=`Spd
zWtl_fFz6S{vVbf^SvtdMnB>3N1GyC0k?ako9NZb{S73Q@zxI&5=u-Q#FXr_=e$pk_
z>q`y+d%;K`??*5XU@sU2<Ve9BfMqPp8nPBW0x%Frkzg9cMV64I$QNKNkZTHNLR@49
znTaljWl<oCC~MC6S<-`3=Zv&UkRjlKd%<HwzD_Vc;*;zkJJB`493e%5ArcqSAet(8
zs#hq?C`o$VJn{v``o?6?B6x?$UcrQk_ill<h%OCA4Ji`LnYd^(v{_U%cS+g}KkauI
z`2w~vn*kp3Jk(r_rkyE4mPz^~Q{Vt_L0I~WvBa`GAP*_fGnVBEd9KQs5OY=L2l=ha
z{2~8USpXEUDhq`|S7l*P*s6>Lu~ucVP^>89Mzg~OSLJadd3OZ?M<D-w5<q}LMTx@*
zaTqBMg@^o?#)(&B#bKg2Oc94M;xJVdu3>%t)&DQg9EHpBNy#ax5mAxRv2lrPPC|T4
zI4W9HhKd)5Y;l+@4wJ-TBqxe3=w^HLjx}o~G3y{H7`t9hR>oD{$<aVw+EqbW5vQVR
zP2FN+YiI9ZqC}$UZYFB|-zr{RV}rI1K~LY%$av!>Q!@)oD@}?T*?gN4-r0?drh9C4
Ikzk<y1sDn%jsO4v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_conv_fwd_xdl_int8.cpp.E847303767D0E187.idx b/.cache/clangd/index/grouped_conv_conv_fwd_xdl_int8.cpp.E847303767D0E187.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ff226374186700ca4347ce48698c4837f789b2cb
GIT binary patch
literal 3302
zcmZuy3pkW%8=eN!XTC3IgP|}O!!$w@&E#0dp^1=?<L7)Drx{tarPAMt&1nfK{~D2U
zY&)op7EvUnHkDFJWu;L6zjV5^-^^a`jIL|vy5_!~=e_Uu@VxK+%y_#wJ3mEmxDAnR
z8HrJ8VX`<JP8s_O()e*{Si(}o;W8?Fd=s~fTPe*|P1b4mXwurJ5Zw&7X-iu@nJ2SD
z!W#*Js@wQREdBYZFSHA(P})e>zCpta>AG?AL(}~u2U`zX6i@We=B3j6?@_Y9@_#o-
z&wTH~?yy_;&CcVtabMq&7ynzGY5BO^ynOlcX`088AJlHUZ)%>&dGVS5<H9(dU_aeE
z`F1Y4=)|qsI(fCY&xveq%(L6m#?}}8CLgIYo0mK@GYde;M{BFw(rm^(%*IC=Zqx<|
z&Zpn)>GC<B`ePF<xZr-xi%f&9Bre0Jcu0Hr>duozf}8KMT2cgRtxKEKPLu@~&S@47
zMh>b~74sd>4L_U>DGL7L{e}YNKc!?UUTRfW%&jRs!SZJHj<m20>utMj?s||%PxsVq
z4z>H$f%j{yOLu&K+nje=!BEPj^A8<r@-CJ9zdBOu)fKPf<$3Yky*qx_lJ8;6`%kd%
zFd_Vz#`?T$OPBnyKchJQ&#xrzvmX9SDgD5x58p3((C_$7&zlQ!ee13B5`+&*;&c+;
zDI6@S@b6)Ku%KlZt>N`GzARUBkaJ4PrPc@yeB36KGD^ALT5xM=klV^9ovobYua|2*
znp;=zGgJx>jNfv=+<aSX&h87BD`Fm*ryqG0bUGxguQ0f)Xux6P-u>^EJoEo)XLQ5*
zEC)s;(=FhSl3vQUw=c^wySuzbW1r7+o`i7Yg&HFb%)rz$-O~Xhwy!fhcu!xlwhgZh
z@l<3gIBj4Fj-OYa5q{$BuRHB~s?7Hir32;adT=4{i_J=_U5_mAT;ZP!3bpB@B+ql{
zImtI7#u`_yY`%$B4?P@sZTph`_(o|PEyuVFs+;{KO3gxTm))@=?-92*9i{%d9$Lrf
z2Bh6tPkDnq`0erm8-4Fqt#BG}YM3*6+pWouCd16)Ic8Gjvg_YouI-(?!>QECS8)DV
zMiJEMM_lE+wa9vWG;ak<)9c5|n*Xd>rjSZL``GTo*14)%3&ouuw<dlrt5=)+WTUcu
zc@M4hA=#@n)Wt=ZlVP?!s6XJ~pPLTmaHl`*yusM_(yh&PdgC9()sBK54Q3}R#kT%|
z*0K)ex7m)#k=3xUnj5guaYCnO!y~`br)R6;Cs>CW#IC3xR@7#V<-Pa!IyG^OeL3#y
zZrSHrKl3*9;YKUArfJv~=nNCW%5Wx+_Fou&4PBo}{hDpKJ+vuN_>I~;dRIAV;9}N{
z*<=5ev)U!e&w5LLNvS6vtEd>sH8Xod9kjZy##O?);|inkoP){ve5K9LM=ZbLESJ}Q
z9AWONS~%5sd*OIfjWDc#hs_MreUD@F*?l*SgOEws17AsKYp%0>AA7~;ZgWHvG2-N^
zGqztG-4AgN_oh7CrDP?nP`Y!DIXEBY+)Zt07-~6bWOD7ROy*esg!-rUFK<0kI3=x9
zwqHgLzq_8<HSc=h-q~@1Ao*yjzHFAWP$#og=LOYz*S&6=M+>iJb*^fbIVj55&$Q@u
z65Y4_rJzb)R$KC7Ica@!d39BB?B}6+C-tN3&v0sLZkZr6G29)G!zuo^KD@BNVkeCE
z#^KOagjEJ&z<!;?A1ZM#YqAxcf$<&~O(u|4!~kfsaiNfJmqdo~z8I}b&{YuwpjY*m
zLsP%`or3Z17_CXrR1pK9EpwFaY(Lf*2;+k>+K6DJDh5Eavn$=_3%R?O;kIBjgT&Am
z1E3wxH4-!6+#DG1htb9aV^uK#+9d;*o0@#t1ICA7v=%{2MGSxr8tP8!XmTEf@m?5B
zC(z}^0O<5Dj-f>#W^TdwZ5T}>&{V|$=!3US&b!h|IuV=zqj|DCNDP2J#5viRNJwdi
z@%|XCL(owX1E5QhyZPQ{+pocRAB-juNXlXW^oc#Ms1XX?XJLE*Mw=2$<;4K#mWY_2
za`XqLV0<V>Qwda6F#x)qIDRO<{DUWq-;B}P1Z@p50J`g1y=sX`!T^lliqVDyLlrRq
z`oYKH<4*#|v7-fHv<5*#MGS!cIQ6v9>@`Y(@qrjkAyAfz0nj*pRJ5QH`$v=E?^ws`
zer+#<pqgY&ec%HGC6meez*7jSP1e>2=LJSwLi|?)V#6Z$ft`Tyuwjwwz%bYdPyEhk
z2+D$3YG4OoJ4__<9GHusHjs@v5HZ_X!uu8Uol6ju#-=#~KdjxOIoonlIED?IQ_P)!
z8dw!uQsh(Vk`9oADi8s~Vj_`Sr6OC%R<bX!EhZ9qRw}ZFtR)q=TNfph7<mP<2x`bS
z<bXFtP$RYx2b71PR5q0Z+=8ISY-0{M0&u?!@jKTes3l|xgYzQjYG}12i*k2t-_u$f
zE{C8bGD#137(q428hXGs2&zTa(gQUhs18|24>SXU>Z<E%gNIA!It|eMtvs=dWUOcS
zfC~iojFm5HI%)a#kiF!N_Uvri*E#n5l4!0kBN)sDO^JCz)RAB=*dJR`)QMmiLDxcS
zB{vGn!9=2dl#13sYa|Z{8WX!UQCCVu7LbMHo)DA?F(p|;#&2?-iPg^7Q7H-(66hcB
zD%f_S29~a52iZx=1f7A2L~Se;u_3k=c%%-dccdJrdJ5aX)X0<pjtDG)%@y^ybnZH6
zoup{+YM4mW=~9suWF-+zUM9Mc#{G_B8}LkdTyV)#P(wbR-BN%J(~aonz)+w-nE2ff
z1A=-&p7P+#2<ioSEy}zh??ss(<hLmEhx`|1+n{ZWvJfa_Q5Fh?F3L~{U6jQ@F_MfQ
z#S0UaD&R*59*BZO5&ZLs2LTQjDGkG=VT3dkFY-?sD;<rIh6&OzMH)s+!&GS)B@N>w
zp$tmz?dxjXyL+W_QgTXacw|IWOl$%#F+MIj49AnmaLLjzNg75ZM)E|b*%-A+Mplln
z6q1LLWlD+)uF6j9smerG6%AE2O|7-8bvCwk_6}z1bhd#NP50kMN!mKg_4Fx*M#d(l
eD_5CYtX^Zif=OmrZcryVyYcZHPe+#}T-@KFMpByq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp.E2C08E4AF953E7F0.idx b/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp.E2C08E4AF953E7F0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4e188e8fabab391d25c3576ff784b61b8bdea330
GIT binary patch
literal 2116
zcmZvb3p7-D9LKLq*Ua3jF=ob?#Hd+jun3Ju>ybw?L`7o9mCy*gY<cvQWQ%RrbE_4N
zI!L8Zs+np%V%cSBa$+TqLZ-AHg-)&X+B<X3e>;0lbIyEc?(g&e|9=1f@B6y}LN6~T
z5(If{6Yhx@Cr4;N5JV%sNy&+^KdM7e3>kv<NTfl73Hpv$$w$*?wf2=a!i=a#qrS_r
zt1U2TA4})%bjlEPS6Mbr^%c$;*GyZ=$oZ^L1+CJ<u=_4#1BC_+lxZZFH5Lb5>I%Cv
zvTH`F_b7ZgRbSV>e9Ony*}vNj{nlG%pJN@AI4l(Ma>p2BZ`%|%ZyP*vdN0p9|Dq*w
zVFjDn*E1WGXRu~g>|}9nxQ4U#SXSNRg<R*`a3h9<Zk6=OO6y<q?_FB+Xu$XutTiDG
zc45h4tJ?wst$EwM$HkjOPdi8k#tq9pOc#hwEPl9<ACXwlY#S#(lUlcoKJV#uk+Sla
zvj=QSu2~+iYOnUFzVjvf(nv`Cc*n0uV7`r6IxlRtKZM^OU;+(u$4&@BZ_Lk(zT|ng
z+PU!zV}xa7pDCkn=KnmCVR7n<W~(4=f%nA~6Jbk))!ajeB82YcEloSqGzZ%K((m{G
zky)60Ju=r#&`NW;$<IhhqwY+>{mT90XG;5>i=rJR1(G6COx4%+cKaSYbv6>QGZ2Hx
zojX4G-u(?Oaj(=U&+tpWc2e<x7I*xgp*<HEvE<4=iu7?4UE+(`RfP_`(F*><LU?$D
zcOyPff3-aONoK+U3!l9X`>9N`hTRTx3SEg%Xi{-9b5MQRs+JU=m=MErzc8j>$u3;3
zj(T5uklgGp>dh$}cAgx(?3ojtFxB0pQzGn*k9}dC$8=*~`f;gep6!78J==RL@5U%(
zV{a1-ACmsMUd)T0#C7gQ(7I)3O{OP&oTEB>Tnn;CXjkKoiGyKk_27v9Tq?<5zEqat
zL(^hid*kxV*j~nEr0hEyc-A8IkxfUs=I!F#tanA81MfwNsRK5ADKjbab$zL)_4uuA
zB9XuhPqDqizFzwDog`au?Vh#SiMo#R-LdcA7CL^HRrI8x)Bi?`>$$3-pn$0^Y;*r$
z+Mh8;?))(KQ-XDp-JA|f_^?dWQ5Ip%{QgL4%G|K;sWTB3?%@vIwIfF@HwUo9y_35=
z2csR=Z+Ko6GZHmk)+6^7&xxe&+xpyxBAeRI{-ed4|8;cc@sgoYiB~-}ZT+OA=U{#3
z!-HCFxpHLqo6~WAR+sbS8v=`mg`MN{xTu${b#Eg5BOUPQF@@d!+#1*XV4E_RTd&tm
z<R8gNn&<oeUh>AgYgjmXz2kmzobJp4&OSN6!HUk3f7E(l8JP94)Asaem7Dk~XK~Ax
z^rXG<kyq6qX!*MEg$=J7)}pWof^cm_o2e4womG9z)g^ybn&*g7IEcU)g0WNr@LI~D
z{3jb*^H6vLfyoG2S0w-ku%<33mSj^=_-g`dBAP6f0KC;Sgn#R%>@*4o6BtF%r78h9
z2H)RT&dO1supfa*2#KZ=fPd&Rm3s3}G-FUWf!P{tl1c#nIi)CkXt`M(3I`IHiclFU
z0l1<em)58dtV7`t0y7YXj!FPN-^H2sYf3zh!kY<9LugEu0Nl1~x;WDDbt(!65Ewxa
zU6lab`!;hlZ7f0!h5ZRkK`44E0r+{%D(;Jcip?mziNH*RsiP8rKRof<<@A+tGYW4c
zFdd=OR043KI6BD|{^Xn*<+F~#l>a$};WRy(DYzE!jo7O+4W#zos>{Bg{@Y<S3}>6O
zgFuIYjYLYNhhQ~^YwKy7f&l_%5-F8Vf<?eYBCK=}gh6T|taJ{9vv=s!*h@VhEW&Vv
zgK$BcF<g_Q$py#2a0-XQ1wF=al!J0XbqvQi7#B<}m;s`SasWUT;5`vm8V|ytIYd}#
zI0#?Scoly~(Eb9(a5|H21l+-J29sd~QeAP~=q#PAttC3gbYyw}hrtDi&MAikq{MKl
z9+eMH2W~>tRgMZ+G+{~JZ8T>nA?lJiBra$qhLbsDE-(-jBK9ih2<!!iB*M!10bwXn
z{B4BtAt^B`iA(&d<Kx;D{TIvWXW-B<DD=N|P}rxxQSsZp-A+9Eh}WrWEJ8F%WE9iV
zrY`oTdAgg>DBe0OIzx}`;P{o3vy1C$OQ!iMt{tEAxu~xGQUjx9%fB$@nVNlRwZhs~
ZU~jl`iG_^^Q`bwl1zxk>ZLOLw^e-9l+&};T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp.9CA386C0ABEF35F6.idx b/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp.9CA386C0ABEF35F6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a502ab69f2c259b0021a0949fa7ee1f6937150fa
GIT binary patch
literal 2116
zcmZvb3sh2f9LFzLcLnZUhysdbn21S<md`Xz%S3#YrkUnV6Qmp}D>SFsddRkF*}H6(
z+9@?N%SR2<3Y%G(wlvGtD$5#{mNQ*wD_`4%bN=h>oH&Os+~4QN@BjFISr#Z1dXXW>
zFEKDPRhph)2tg2?c&DW+zWUJsf|4i@lv$u!{zPFBh!y@5xO1DT$u{<e!OiLuCR1B$
zZtsma8QYZJao_uinZwvfc-majc~wotI4mg48x;lQU{(gz(sDtN@60pLuHakb66B(?
z>c&vR7Hw*DSFYz%=jJlgY!~m#bB@MMtbf%Pcidx1MD8Hxpk^{7Q#<U=DQckYdfk*5
zIb5cR+ivOqm()ves(W;AaZA&xj;oUPs;!R7#wq5#htKh=yn=>h`R*D1mA=Z2n#1ki
zpPSmh{#1GNKIal`OytT}tz=7{ddB3y{*`y8UAi;3sb`jQZRb&+j~Su1LpyULc}WHP
z-?yZ<T^d%~T%;U!(k7l&O8zPe?w5ruY&{JXrMIk=B;3q%B7K<7t~&du{>FCNai4|O
zwO@K_ef>;A!^E_hrlpgNlt(s{&80b~eV6SFAG|M;O`~8~u!)5wvPYYJS15;F+8(%F
zDJgbM!_L_58EW*WpVb=esL;q{8{|sb$YhJl?H4&cU5kAscTUxIG&u~?J6er9`=a+2
zYvUEHygo({&zzZZ@>sJ?b5Yp2E4?@U%0<B<b<}C5QPB%rT%B<L?z6A!oaBMlhCcPF
z2OmQ+xxIUOZ*o|x=25m?_-kd;{+SnhLV}$5H(a_%<PLUa#G*Uz=BP@N$}aeQE4A#W
zJxwOF?wygV$KU#QZ|2EdP{|4;sxRb>JYZ~VmOW)njeX0);dx~S#@?o(nq1*q&&_?W
zT6X(xD9#dRv@u`6c{5UjW%n8?JDqF2CQMt%(`QuF#=PeZ-Mu7e;I(gm-r1E9QOf98
zH}}b4ahxJ~{-W^GBlG+pU+c9j#xjR)zs|gMJg{V<iq~8wQ0CW)8}Dv&|251$(;+`E
zN<~#?AZ|+A$b#bsYu31V3QHAguZ>3@T+J@jX8e@(tX{5)9hbDWjvNqgSySEi?0v8>
zIZ=2hUbU<K!u}(5qBpFUK`hr@<>vN16$k%`dpr&w;}3t&wM2{h#DWdsZiVt-m(V<|
z;mfukEf&;Pt{m8ODkQF4xZ|wjbpf&|WMlWboT}I<j+Kp;4oTO?p3{DI@R(@6+%@rV
zUCN`BA@v{S4~E~3{*Jkpek^p^w4m6{ZRwR`j<0Nf&`90d<QIB-Mh&uf+&34J+e`Mu
z{9gCn(({M9@~^%dFJvsKNC~YFle+&t@i6z^nA7@>@$Rhy&Z-Yry)^7LS)}b4DTg0d
zmzI?(K2L%m>yY@-un~0t3a^A9+!!%t>V^2v>R+=B3OdVY@0OzQasp!rW~LW_1E|~f
zy$rv;7lp$JOhG6ty#Tz-?9bCZ7P&MOjv=rSVq~TlfLHUQ?dt1li%@t4fl&mVrWb&d
z@a%`hX1jY(IF!I-giO~9z&qOcDv{mcMhuE4Fx!w#)(gM~Hsr^5S<kwR!Vv_fAvA_w
z04`PUp*QrnhM;gXff)$HL@xlBwQ~kSe^DGl;V1&r5js;Z05@k092I-K--N=;2#g>I
zOD_QbHul4d&6)%f3NIlr6``8z1>m=pp4`F5rBNsxNnj?zG|>ydlP^Ovyyo*7Q8=8y
zridwBF90i~Yto$IPtK93pLGnT{?924r<>FH;8?&n;<q}}z-ZBG7W={0Yu}MDoIRVp
z9NaLlkr=7-5Oia>vAHoH1PGW(jMO;^npVfIV8||KMj;rEa1buII1D%97;(X145x4?
zT(Bt&r*f!VaEBO<a!@Xqj^P*w<APLzgb-`$;sMqK`zFe|Sb#D}1X0$-0hB=!h_WsQ
zpbV~xDC_J8<!Zx`)L&h14Z;|1$~3hCUST+c$*=;Ws(r$}RsA(J#62=Sn0~--aCqV#
zb&&xhF`Q;jvje*W=O)(G1qU>}aiDx-^=^6rF_O$7bHT~MXrfz}BhU@DMU-{<0cA)b
zO-ax_q$wn6T;kIJUnH1`ZQp0x567;8V*l%dR(<*^Nli>yOFRZiLktYRK#a%~6#LSc
zHbq2VyvUYL6`7csGR)cYJm!0OFYxhoV9xgBy4i6)S7l9|W@$CU+J-liKg-^6&Rl0#
Y_vtPc0w+HvOBfgj`!8J>KnjNb1=1SdcmMzZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp.77C435750269A2BB.idx b/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp.77C435750269A2BB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b2fcfbf15b5aa6c3cd26f309a12326c446d66e11
GIT binary patch
literal 2106
zcmZvb2T&7f7=||+Zb`DM5JG^U7%&EifQk^1c2-Id3suo22qBaRo+g69^FT$#0-}N<
zN>Kp?6(M>=MZt^ZfU`tA!~zxs6bqsPN>RB@X8v(*#?0)S&HMcM_W!>3-vEJ^*K`U5
z&5aVs6U1vHbRY<#6PIkQG;Svuf?}x<B+su7Zk3t}u!2#+j!n<y*%dEHqKAujjOlvV
zRnKdxh^Xm)UD$)>ej5&!*^|yzD^7id?FtSI`*?50%*j>aay#EyRwtic##7@&$oaJ2
zZhWXq?fYYCTZZd{nB8it&#~*1A{={C++}!pfy=_cj9zwj$M+<8U%xTCFa_K7`dU;_
z|A`J^x|!$4M0dO6&qlI}8?S{mT@k4(H&&*p#+tm^U%{(%_wASDIVX9ZnU%J_W50T9
z#kjX?j+Zap<51EU8no=?KNK@=t;P4Ay~}Qoueoi1t>m!EL2|oj!zD|`;5@IB299@p
zMxtC}!iI(WTDSP-B~NcrDdg(XME-GeaS*hqrt_R)nrMC=v>_{FNz1#o7uV9K7Fp0+
zk%HtS7CS7h73rtjtx{>q_<cH80%ix9mtQaY5;uvzTL;<3rDl#L*AIEba-2z~F)POj
z?=>BAL4q0HgU^pqT87Bi@2}xkyPj92uOE#%)U|rHbk)Qois-=B?ugtk+hfHP$=cKP
zO+$Jv==`;`QiBYdQ`2qHKGCAGkf*Yaqd_Iy5}{4=PO^IS6#Iu&ak0J5IP3f*`sSiG
zr62*~J0BQV)5o{&`uNTxr>tMq-jKaqYRS#Zi8Or=ZRy4y@Qy|&A6ypj=f97R)}7rw
zkp9+&Vs!81s@kvbJ)fm;C61_Q2@=xncXE~Y$i`$_YP7<(ad!WV6J&i4#)6IvulKGg
z-7gz6XC)RVM<(52w!;T366Q<p)tzZ}P`ZCJYotuDIHe5z%>8h;#;%U5PJiCq77<u#
z*tFWdwYOL(UFkN*zchD-=i{rd%!)DjhnuhE4Ml>IZ<XBo6Lx8N%1G7SP0qI#+sdcp
z9SEtW)g~cMX_i6xMcImICs(ghX|4PE+?LVd$*+2*^<9?Tmz*E0&)NT$d*ge<ih|mn
z+>Dln4C`Lyu8ZgP=3el5W7Oqq#NSnJV%vEtdsO)ND}0#Ozm;Q#7IsJ4CHgx($+By5
z486n1akG&OHos9m9`eh(9FVCW8oqrwHh8eVu3hJua=Z{vbHnd$|Ds~YhDnFB)?FN_
zY=3z6(R1$`ffcX21{5rbZ=YM-;0CXR_(HjJ_$>iee=|P+Nwm6gK=+fcS98pRLV?-2
zsx{Q2DBeYf)R#kUy!?<yheB?^b-i;dS*2gZFOCgeA+^RxKBgxB?T~-J%<uEmHTA+O
zS#m<;SrP<I^a~$eJW%V6!pk5C*GKf3S|Q$9?Z<|k-&}5!DMsO70%HhftQCO0X`A+R
z`CrdM;l%`|B2*)-032ZauB_8ELl1>R39O6g8fyjM6}+X^7cVFaQFsY~Q3M^Y6@X*$
z-yakkXLh3S0s>PI3SBDzZ&&lGeXRGZFeseBEFBg_D**3H%nNUuIQbF^2NGBh(KFNv
zz@@dj>2;laKNMa{U_->vKq~;BP_ug${3Xpn;Sd7T5js;V0M{q=9Ex=Lya9y+2#g?z
zkyZe{KeDYor6Yob!V3vZLue*i0r>qHS5EKa(hw95A}|wS8fXRJ?_CR$+}*e;6!s@D
z17Xm$0<csZEt?GgNRCANX<{(#e@S6D-Gt5qZGqf~|7t=5rF~Wyv064>-9o}}mJKTy
zOc=ySl+<JhR%5uni9Qb;AP^=|Qj;WD1VSWiO@hD%r3qV;9I!!Xgsn*o*dQ#z)+7aN
z&<$a05(0LW&OpK~enT&e;S46j9E=mg4Vi}Kpj6dNe~;?73I)+B(}g(~L;`{(TGgZq
zN@BR4iJmps9f+5Bt|nWs=*<-BI`dBx@^vvBVIv$+3d42Tx*RY=45zRu9559Or?ROW
zkOqd+*fb8Pj^QX9<$$qbIL5{};B^GsC3c`WK41r+Rl?RB8(>St@e!JPnN%d>5Z`3n
z+QVwN6Do^>mxn>i!7>O6`*ABuh>DLPe)@|2$U46ux)dskjnUT|>qDP6$C6I-F)(Hr
zny_ZLxVd}$I&;<(rj0Ad$(sE$*J#{$GjofHlekvA$+lCc**ox^CpengO`pp&@)8JP
M&qcGnN%Nuq0F@TtaR2}S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp.8BC00DE6D83E6AA0.idx b/.cache/clangd/index/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp.8BC00DE6D83E6AA0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6bc6bead4721f0e7bb607f9c0be52d0bd1e82e90
GIT binary patch
literal 2116
zcmZvb3p7+)7{`ypJ!9saX=a!iL`~z7cgZkb6P4$n2O{N^3ZdvR-leN6t5QaJ<tmX%
z?+0~Fx`;_KO3{rTveJ~2T&9;&k$YzD+Uu^nR%fmGuQUJOo^S8{eg6p*dU**nA;>*a
zm>M6IxJCnlAO`VCNEF5HRfnJ$G6bcT)&xHl89HIZlO<Zz!*Tp8!2&w$(&evZQLuD+
zU6<rgvRtxlyZmyvn{Ba8Q{3>(;50SA*8Kg6#xlqj3JdNzrjdATq~mP&z^rwfUB*}B
zjF24qyjt$orX;<nb-mpBpa6Y2*s{qq_uL)r)-8M6CW1u0eYqRUCSNpv{1Px{d%CX9
z&cPFX5#9mmCI0&cuSy?`$=_ep?X4bI)>JItjQW4*;TJxdZnCvT7FK7x?CAW4&$kC|
z@Gp#puD5vIj}6x6_};<wDvqB>x!LFzc<!%?FpjMNk9?Q8;-hKbMwy0dUfx+)Y!X}U
zKXBf!4;2k9vE=#PvT$T8DlcD@GEd}Co(q?y^Vz<VLUPj=S!&eXht{s!zok>NAJ2Sq
zQ{7r<H$Qt?B-QO}P*qfc@mMA07$RhOHTW5wNiw#vu@c54>N?(2%+tR{?sR_o_}!u0
zkoGyJcRAi5>3CM%Q*bZlo{Q}|Lv^*QGksvSo0reiRukI~-SVP6ttWX{x_2Ipm^<ZO
zhHQE0(IB&;)k~|gH?j*1To#Bwl<Cuo46V7ND=I<^QZo~>rMr^+TpH!KYfZB0<NcEO
z_3Ecfs(U*xT=F!hMJp;6RDPM_UDlh)GjBV=e6V@j43FsAMP%xo=C1r>4bRx4HA&~)
z@RbXy4umAVe3jDi?XzXT@e=(yyDE?9z@)IAZS4`BOKzPtB;9|R^LgGUe>*X6>b2pq
z>xaD^VEYTNv!6H@XC;JY{b7AZSoddO$Kwy5QjCpKv(EQKJ}GK)AL5Vg5^yquc)R0|
z1bQ^L#9gOtZ4(b`dLCKzf*Fg3OLA<VOD~-)U}Vh9OKv=9P%goI<6BBA=58_zXiu9}
zaxg?pnb?T%ck;VR%L}U4@twUYMH7Op#SbUn+71lcT&YiaGMGHpT39q>+VVrTx~M@>
zoD(9;F>8Fo=0zLL-jdg4eYVnSDm!RSsLa81G9&gQ@BGp?sJAm)aKFX(ZKi$Q$R5dE
zQQ6Vn<@cKuNA@jpm|fJip_`IN&8ltLu~?|kU-T}iS>A15-|Qq#<BXR*bUKxhR$Mp9
z&YgCc+W1J&eP~s9`R{or?HXy;<%>@*Z09LnpZiL0mfPx^UVV~~IFVPM&lqj1(Fxpb
z6<6Bqz1(7?y*%y5e9Ea{$;qHAJ2>Av!@LF~N@`<%Y-ma@aUKsZdyhp^Eea+e>8tS0
zgtYjzm(?K1&_Dd!lF<fl6kY*AI31z0R6_h`Rga~5Y1bJ`aTE#%6Bt7<9hCs=P5JG>
ztALyPQFsY~$q1RP5`Y7B{yyD5HHU`6%Lz<ENIEJ3c(r+`*|l0p8452WFp8iCDgihK
z-}dOZj<_F%7ZO+#(PXFu;5{AYH9lrVQVa?wu%3pVrb+<LPbvxTHMFQh;UEIj5E@e@
z09Q8TGOqXA`=f9uftd(XOC<oG>fj77Y!Vfsa0r1J2!o{(fLoJ?53P0joQ}eQ1V#{q
ztrCE{zweU&`f`mL3NIos1)=Dw1mKa2&PIx!$`BM@N?;bk(ozY)KVB_N7Py#5Q8<9W
z+K4toB>;<})+O-Z|0btK`RT-9%Ks*X;S61dIk*;>8}VD^&_HRQ)oi_oJ8opDVYr^9
zUNGn|7$Z?qIYY1-!|A$oa}c0Vys~d>!}~Y{!x0X`1s%e05{JYE6)>F4A#=g*7*63(
zxS&}Wj&e{g$OMLC9E=O%x3Xn#jJDLhoH&>!N7E1_11u)4p-c}r7{G+BObxKXbqHIT
z7GQ&*5VkTUzy<*!Y-Kut4LV5J%2WWmN@Fy>-Cm}EF<hIaJp+so!<j7R3{a}dEnsfV
zP<1tN3zieh9mD{%ifFGgNuVT#({yQO;B=r%#J<WrfkmJOVJmY4Y_LCJEAs<vNE8*f
zM)@T{6p>)0d<j#^zt2kW0;%JLa(c7lGpQU7TM0$PN5(}H%s0wkUE>#oq)A3GDxEgP
zhvBhcCWGRmrK8Q%)pK-m5zKXUn>UMP>1@O|<NVBJ>l;j)!8J5CF*Ucanr$<OXYVlG
WZmPAdJB#flTm^eBp6{*Z2mJ#xH{H+x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_conv_fwd_xdl_fp16.cpp.F004628471E9028F.idx b/.cache/clangd/index/grouped_conv_fwd_xdl_fp16.cpp.F004628471E9028F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eb93cb35275871df63e45b3f98d948bda45f1657
GIT binary patch
literal 1946
zcmZvb3p7-D9LKMrJ7ea0%rG+=86u=X9z$NKR33#4QtpU6Mj4MZ-n(U~vs*Mhj`Uzz
zl~HOb^rY7!lEi8iLS%)=BZH8~-kEd$XLrx(o-^Nj@Avcn|Nj5~?|aA7#nCYX4?&wl
zToNOBv4L_B1d*{{OspX6pezK<5FjWq=kCsqD29u2?owKM>fLHR)|V912#+9j-Fx#^
zYbwGo($;s~KK!99;OQF^HCfJ`(mwj&4X>|W%a(71_rCPxW~R7%AKh9$+gXD|IR%{f
zvM}D?@!&z$<lMgW9r_a?@u4PJdln8}9Sq1baNq1Yrg^q^G5(Nva@~gCl7%PQzaDcj
z&2hu8_ff66@TutHVRd_V=96w}we8z^WX6m!k{agUo;^{YzvY%sQgCY@kzIJ-f7Y+9
zqLew1ovv%W*(T4afOk{)C+D^LeD~<}H=UO|m^yZTym28a*5|f{;GWhw?>tg#9hSFc
zplEwHOVHtAaND^`&q{Bs{LanVK9P{v<8oo%lGbl%ppr6Dm&kkNZ)AUHW?tSi(Ds_S
z0((`noGx0&aG$MjFLS2Mqsus{B1(Q#wKIQ(Jc4YeFf`z4vC)1VB^k!PaCQHx=3l~*
z8+y~@m0QV0V*K%ke@F4-jE^eC{c1KMc(7U!sC}t4l-n^FsQLByejb4z+whz-yVjb-
ziX|3MghX1iQP7oqqe7qe;-1UiB3+R$t0i5dDLTrutt7JWjryq2DT8=q7r#Sfnu4mm
zl<gX3WKPUXGq-uqU5f5|QEZiw;g{KK)+s}1UOwvO@M<zTEUnr9!0EYqn|{r?R%Jtx
zS67(zSWRcLBcHuJxD|do=sFY?FfAkAO=$@B%{V%Ea1&nZ2V-`zIioypCNNu{mOA7p
zuKj#2yh^voq2fGSbnh11{Jh<Q$_s-3%HQKT?`Vl(U5ee*rqewwRe@dw%FWT{565%@
z1p2lPoC4Ps+Iv2TpIku_zr1XZZOBU#O}Az&{B^F<k6Rit=&`PWW0<tI+_0EWg}aal
z*sajY9v|Eo5i8vG5l)&k&i!f0W?<Ol#G{0+Pw_L)GqcoJ)h;&dJ^Og%g3z*#%1M4>
zL~C@k@6|5lW`_H=xLyq^IrCIwe(C2|YfDbHY;B*M(`aV<Hrk%GW$|a)%7<db<z`6*
zjFn}VmG7&Ko0{5HRO`AWb}g(N=G9*)v1<&_A8Cu2%<rvfnal{}*Ji&a+Oy{uD~89Z
zGtUO5wC$fovqL{LWLxO>H;GR58|Gx)?tk{E&vL0=oDigwn`@Dq?zHqIE!pkcxBmEM
zzuw{l$CkL}?@Njtd1XsbL7dqX6q6Jg6e|Nk+O7d}9#fB<NVwe)get;{RH<P9P3gsw
z&1t#LI>{s9yfI9KiFBy|>_kjG^O5uHGzqr@!}72^O)3C;(mxmWF@#Da+%62`VLV+b
z0Ppwvu{EE5vX6vw!!Q9Rkfj3f@g{@2&a2N=BG4WT)8*&{sQ{c2mle>dz2+ea=ZRqj
zSV2`P02e$yMSjv};Yz~!V3-0^C{h9VT9f98+aH2V5^g7km0%^RQ~-V+Kaw3}J$;CT
z<6szuVVYC`9=&O!Gxn~)hlKOOunMd~kqW?zAKl{ZY;`M1I1dbyVKP}N01J4bF+Kn3
zy$tcYJ;xFM*J}h-Qd2Sj`v#rI=1STQN`sDLb0zHtHi9auDH?#R_i}fs@*gJp$RQ}q
zfSKT|2#ROmnV<rK$}{AdU>O7@GKfs@HGtr-#U-hMRZ{mQnEJxOw`35M&Z0Yk{J<R>
zEeR2f2KiyOBs^e)+%Q`b8nD5+F<TNAuuJ5oA{#8~#$*wcOeJf9#v>?&O3?zP5LAV#
zq6IPo$HiJC$rV&WPz5!G)nHRY#_|bT6U-d!C<F$92@VWuV7rin4(5VoF<TNgum!w`
zK*@8AAUH+`yUU_pC0@fgPTCO8mkarVVGzXq_c1s!Bw`=-x{c>5E4K{B6XZ#Vg0hmL
zGkJ@H9+~J&p{uB>Ew{3^wOen$VWTmXWus%cTJw7@O?`!i)=KSFy7~rd4A+{Nnp>=6
XvKdBZo2fKM7eCxq_f1YR+o68|(E_jP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_convolution_backward_data.hpp.04038859D7920D2D.idx b/.cache/clangd/index/grouped_convolution_backward_data.hpp.04038859D7920D2D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..912b03ae24777309c0ee71cf5316fed733f09a2e
GIT binary patch
literal 13936
zcmY+K2Ut``*T;9tbzQy5T6Q%S+|9<WF~$n0(LIWw-ZhHdXquYW#Fp4%OtI4xq*#K0
zXv9Jh0R?+BcEl1jHi#Mp!~!Z{K`DNhZ@oK5pC^IeGlz3$=Fa@*%(=&RK>z-eyXka&
zCk~i5ZPM(oxK5|5tNu4@_RMeRGWGY?I^DeJ%WJ31i0Wwad)ek?M@Qm6A>mWUFa5)3
z%Fb^_?4EnE?VOy3S)^4eJ;CZ9*;4r;^0Qe>YBbh2ZDac{D{EHBttC_Q&n6yP_vqtc
zR|`&k8IrU7@vn;)l-C?`_j!%UA@6jV?G|e*J^g-Z*5RgoY8-mi;OIZF<n!`fT@RF3
z%*%Q`eL>Y8&uPD=wW<BrlJh@DbhtV9r|sh>UQMrxyWiyL0sl4&GAE9$`1|pT&vtxV
z{9f}KB~>x4=lpdtAm`r6J~zMJx^_;<Hm~V3mC+4T@?U5Fo|Kwd$Ng@Kb?LT2)6YNl
zTYI^*#i<t0DyGle+v`@~ce6sT6zw>0#LcHN{-dmqHkS@Mnl-s9?a9UR-glZd@T&Xc
z>7v)Z6+f5v4%(H~YW7Ocp#Ha8Y`^3C#6RiIrk?Tt9R2XxldksVUFY}u+rQJEny*?8
zZt?8<LH8bKpSL~hGw9*tyMwLi8?)|~_?>f0{~@dJncv_h=^HH%HvG_&bUFUv7pdLm
zr@0TEe?qz1$hJP0`r8_IDy%cU`-Iu!Z@C?bnPE(+j428H+jH8D^1|A(TWo1hTl9_O
z;+vsKk;af?%^>K`?JWObLsrNy<0ih}q;3sb_`|*QYsTKUPp8J8A9C!)4@LLWZjB1(
zH!p3C8eZIT{NgF^jcGO9FJO{mJ$2M?-cHwYz=K79&AQd?uh#$d>e<MeGOcFIiH*-3
zp4gznin5!!2}N{r?X#Eec^qE+wBNn|l83bKFlxNt`RIwAE>!+`{q)uP8+VS%n3>-;
zv%S);Rl%my7xEtlgx}wlm^HQF?xdlH(XVJ^r=fE)Y9CWBmbS`VIp*cCA)ot(4mU4f
zJ2ENR<cR6n=J==O-y{$54gKfy$acO-ajQ%d&(&MiZi}xkU7fII__)ZQewsgK-q#gF
zlA5>gsg#{)(Xw+{Ge7Sfw;hd-{#|goH0t=&q~x{D$|{n#9ea2Fg!0U?jE5_rVRpd#
zFA6J^tZ%2ZiaIdw$LkLZ<a>F$4rYD(^Pxfa;Pg8W8{GAe?)KZRyoc^58eAMPBF3|u
z@788rHu@K|okR1e&j$1SYr{U+@7}EGn3`uo2Ze6_`s>ku1`eNaV!@29t7_HBmWRd+
z`NGk3^WD+EeLJs5*V-MYSO=6Rl`Wlh>c!@ViOI>OGf#aoBjd@bWz+Wz4DC8CyyvgY
z(#+Q$k=(d7Q(6Uu@A|Ih$&y_1!Q?aaUdy`y@ps?tw)*vm8AIyM`f1wNL)5J(Eh@7N
z&^22v>-u`Bn^OI!Y-iGT_Jz)1c~|G?M-)=avEr*m*WG`uM_cySDFv)f0c-l^hpwjc
zZ)%N-=p8t><w9B4PV=&wNUO<kouOe3rx&@!$1B`DeE*hpJvEb1CWW%WI)kmQ)8yTa
z)0&U|U*#QH_lah5n@G2bYSN^Z)5Pp)38^)`X`R}%x!RYjtkzXl@68XbFV<~~XOBJ-
zK1<epsCl_Sqzh!2&d{i)(@X2DA71zMX|hDtb<j+X66q)zBusjIHsaFrrWw~{T{q3-
zF_9jVPeor~{FN`KW!3YNbsuRa4#}bD4E2cfF+&p@KN>mm$|zaaSu^-a`bm9=HYdIX
zQ|4QPf6kC~tu%v=q>t1|t52OChRu8T^X0wf?v-_(n!#l%U8ZU;tVXBFH$8vue<$#}
z5?R+pGucd}&DCRhPwzCDvF68GYsQ@Pkag`f6MrK46Sb>aPrkpk?0g=-K3&#*teJcw
zeWLm>F{h6opEqbR?zoa9>w0Jg>zTBkjTEnM-h&=T8-AP+BkMY8CcB8VtNJkwB&V03
zW>qD$9Nw_KtZS>8jFZNxk1=zn!Q!swZ>Po%=^*QRYX&Q*w1TQjplNNVNl?dIexn1G
zSXtLvGkITnUww!+X(1b=#H0NtXUjU!3`(d}QtjjYI!+(a^6Ipw_UwOZ&=y`FC?}Zy
z1Z(l;hqjCkH<W)EA9FcH*4Z^LC#iI@+KV>z7yI5iT&Mh4OIhckndDL_myQx9H;P|L
zSza6RWSzTal0c<|YA<fGvoANtw}01v^zJFLuB&EJNu)}mdeK(>ZC!Ye-e=N1WnCZ5
zWD%1VF?AU<d&lXeFm!iuyV`4-$hz*D$sQ{0q3Zk59`yX9(5P**2ZqSHUYbc1m7=N_
zQ)7eEOJ&W0`6K6M{VMA^Y9{ri`s(}9=Js{QceRZR;=Yr0Z8U@Lr0>+NrM<wJlfIc{
z>mYQ68s6wh@R9NjNw<4gWhGcL1xwR4RWW>01Q{=b{;4S|W-xoZdZ2s&A9%aowem51
z?CshvN+0Os?OL+Rr|_w_YcndJ!{^?vb*4;(sot(mE00)*N365<XzeGd^+wLLWueKL
z+IgppiO$DyQyib5{bj9J8*VNc)OFY>S#bk5hv<AWXokoF7DVl!okI+)J+w#kA$*9)
z13VmJ(4OFls0Z{w<O4nqF(CtB0HT2~5YZ?Y<q(rK8b%`;17i?<1K&8rME(!{hiEcP
zMl>5{JH#Z<fjNlg!dyg42rMBEmoO?z2`t6DjKDI?VFbd6!zEx!IDv4?5d<PI#}bGo
z4(<KP$_4@(FmEKV5%UfLJBY)jp-KvY6wIjvQZfHY;7{UkDV=hNz#+_s2^_|JmcUu!
zaGf{GIRfV}pC@o0a}I$V;&2^d%D)8u#e9Rn4a^S-JR}a+!KoAyD8yVupa}Cz0xyZf
zb+{?72)x2vPM{pKKLvm4a7mOBKp_BgAca88YbdOt4ws56YbmV7{5yr;F>j`@nL1p`
zq->$E1@l%4TQTpVu!}les;Z<@NXMK(Ap`Rf3P-5JC63Bb3P&*?qi_uK1qv6a!=?So
zMG6-&U!rgc^DPRusKa$kEB{gW5A$sbw=ow{D54J6!J#~&@Cfr`3Xd^YP^h2|*NLW7
zQmDjSMWG7wat6zp!*yOLD;TW69LyjX^Ew9Wn8OvDDNzifFh?_p#=MQeHs){z14=xD
zc+3e55-{&(u$wttL8h{Y!5++e8SF*;CPq|s*BKiM-Hn$_@zQurceMqt8ag9HVR(U?
zE^wD&-5r6A=OjNk`iZPGf<|6Kpc_MDL`|V7qE^t#O9*spXpN{Xv_;evx_SwL?grfu
zxq~~R1+c(N$lgK%3kl-I1QufsCJ;=#gzSY72*DgmAQW>Hfhgi7WG|XPH0JdL)?+Rp
zP(Zx2S(KG~1nyyeK;Qx9X9S)RFCiPx2|UMKMxYGy5(-PGmynI66qaK4qu_`6EQPbw
zOKiJy6wYD(o5J6ib139cFR|_ZrSLE2n-p$hzE9yk^%C3e0fh&c3n>(0-pODm^AgK;
z7lU1xGZ<uGE@qvJSvPHU%1SZwEoLJGm#{7+tUKo6C2W-7Qr5MUxeJcrontui7sGvH
zIP$licUjMozx90hdXD_X@~*KQ`8&%ypT+*pa^JJq-#Omp9QJpP4?l<fo#$Q8V}BLA
za|QNS!F?;Rze?Vv68o#<!z;1BD&DmU`wNuZ1En5f{DIP_KxwStAju<0>Z!4I38b~&
zT>Z4TO;+x+=J#29&9GU1cJ#O|g`*9!a)!bg>LXTIHic~JBUaEo3iqgwSRpScyr4cp
zc>NgoF&`nUs~N0jK0-J*G1$a>gfONtNMk-?1(oqmWn#qI0k7(p^KL5i>o7MR=3i?g
zXeni^YC6SjxEdYz()ThuZ`(SZti%z096`LC>X%c*#Z+I6xq|5{7;4|f_1ie&>s)^w
zbC{$LlU(gu0HgNTc~KVqJ;Y6ixa(DE5}_J8ugk*l5;tAqpKAltLQvJP`qC`S0=X%W
zf2Nr^x`)qb@ZpNk8nSYMwZFi+iy64c+F!(ciM79k`8FGTn~l-dimcpWgYRI@WrK4C
zYuc>#r}~aA`m>RnHj43SVy~K2-_(U!88ekJ*S$bnSE`xwzAg+GOQyw=YZ`_RuGi{b
zSWl^$te6ZYqd1>z2AffwK^+YpjoPDSrMIECQ3U&g4TFu^Y|6@b!+4_z+NT(%7)4+{
z-!R{(jZ;<vh#`O&#Tm1P7}gM@I77A&!xmx`p>aAfq!XhEeUA~tF=EtKhOAs9hKt0g
zt!#C^iQx`0iZkX3F+3qgZGp&26){v1qc|gisUeseMd%n!4bjvnLc>IANTfy)`t7BL
zz0@c|yHnI~iW<cZzd{XHs8M?pvXVy)dDJL&cquiMQlr@6i<x0DGm0G^#tdQ1D0X-(
zGsH5Z*x|{{kj#u?hu>%K+-LQ*g*&0-qGL||FFI3MNn$ONSZA?t1EiJ#l8Yxf><1U6
zZuz>stb7e$`-pcp0VW`t1d|Yb58wNU_xc0;fM^!XLX=4$llX|Ybeh0v%x4Ik!F-j#
zRpKMw;xz)-FlQ6U#+*VSh5CrMl1d>J^G*soG5<;7PwFGy+93*uF#kp2FU)Hgtid|F
zmcd%gkqja+$1#Y*I=q>|X3Sd|Y{eYH-9z{%+FO&A5I#19e<^q+_gKk$3l8O;p}dda
zFzy+Kc^`M*hyCs2WA|Zy`?<${?C${gJb?fHAoo0oxsbaTVt<8vY$5hn#660zzen8j
z5&rwf-19N!ILSRuLgS5-#>Po#yqhJD%@P{#7RhsqgvPs7^4uy|QwLS&oJmKabKALT
zyU@)UgEB^?l#g4ao@1ZsKXaWcvT}szkKl7Gp6cWAxwVq%SK@OknCpY_IklJT_hNp?
z^$#&`lJuLzIi{(d+Fxe|RP^@`Zu&!<JDR+zhShmcVYY~y7Kvfae&rUH)NR&mby&Z_
zZ{9*&R}ff1ytT*4$~pq;h_?{e1Of@fTbn*v*-KzA@fPBGiohx2Ew=3y0#}H)*p`vJ
zQ=|y6G*MD}>rBdu{w7MML}`+?UbP3PhR)QiFihvBbTNKShE&7q1g$WO;HC&Meod`a
zGiRDs7~bTjo8lg}<jcnW6F=@eS55y7gJJ!}c{Ln{BN_=K5$z+ekM!62EGzp7?8khN
zz(LF#DQu+u#fk9;g+DOIQHaCr&%mGc7m6RiAOLd^gCNYyxcf5hib7>&86UljkJIK>
zR{Xe!AMb^Ete=Q@WW}F*`s2Nv%H313zf?Xt75hu$9%<NL8Xud6{q5wQJF&l9?w*VN
z<?_+F*k2y^$ix2f_}Dz`FQ0qnV}DVSdz6T}WhF`)9VMahMoS*i5*lx`G&Wj<HnI{U
zdBzB>S-NS#e8;xqC%=-FBG#yg*~CN@vF=3-iOeI`=n-x&W^To-rLZq%J&IXB!6nSC
z1h<#4o+Zpnw3jlsQruq2Y?Z8?Xs={_D%k+RRm@g}+pCzpiVYO)ubJ&NZePZ2%Q*74
zjN6xSWbemqez@I_gC8Fx{P}a6KW-1<wve~EZwOa}{Yq|IiQ7ZDEfjMYw}oMj<jo>^
zEAjV{yk{gw^RtdOTZh|YxLXWI^Bu!`#Bem<>$%%{+`gXoT#x<5a<^FA9?xy@xIgi{
zPdrESoxp7gxIKZ}6Y%^ca$6#9Pvy4Mx7nVG?bEm|4Y#LpNW=5JliPOU_I=#8?``h8
z56}00ZrhLB4{+N7%m=ycAm&WoD3hc0oypxZxn0cvY2N5GZa>T2&f@tx%X^%~^L>uH
zox|<tc+YcqzRz>F^SC{m+p=+gvU#6uJl{FomV?`KxIG8Y-@n}UFK*A}w%oVbo{R1C
zxGfL2=W)ox^PSIa`MAB1+X~<2zJ++ci@2=_w?E>xN0=XT+hfdSyjdAuzh%5v8SgLV
z`z3Go61P`yw+cL86}(3Up6^QTR*BmydCy8b-&NeL3bzMJwm=E(|3IlvpoG?UkYo$O
z?Lm?~NJ9FsT(T|4?ctIw{B7<Nj{QYQwg}uFA=x9azg3cL6>g7`Y*BBsJqp`LOSWj-
z9xXw%g!Cy!vc=%`ILQ|GHusH_kUnjeY@2cW7Rj~+^H#~WRd7I-+vDl)OpLB0D-ECl
z`U2P-nxij(ouCu?0%!+2`T{rvhM+HiU%(gW3*a=EhQ0v)0>7XzEr>u64%2=o@H-CE
zwh`Ed!?fK5cH=PZ1c4JcO#6qxKR8S)V(M8pR4l|uJ~Wb#)L6UUtI<>S7i<wbJ!hup
z%#{PtuBfV+^E<XMTqBv*h^%<nS6@8c**W&O8jb!8Kcis=5(vb@TuWdr9_Cg8Tk$Y6
z2xQ=49w%@d5Aza%OL&;;c;|H@xfNB=&26ymQKs4rMX#V(e<#uJ#IgQ5s$Yj={e?`w
z5XbsUxqc~*<x{vm1@j%Qzk_+5q+cgueJv?SYQ6cguT5Lbj@|FQ_|x=jT{R~_AdmQn
z97PF%65=Cr6bmUVq&^}?v68|{>Z3)yva+7Sdg>!0&g~So<J<0j3j6VG_B4glI171=
z!Zn<Qyi4IO&O$z;@C;`mmoiw2vyc%CBJk~ZBZH0jcALT=1!o~Mxg(Pg)7GD?lyToO
zK3d~Aw=aL`KQA&*eR1!>duSuJgqCO{c7ZNvBld;9Xe0W9FWQJ-!B=P_&VU(cBNnj_
zi&!V^(L2V@v$g4TU<HwteGK-A+<Hp5VRwxtJ0jJF{S5ZAep<q1S1ZGa(6bxws}0K~
z&*dV{JkWYTo6Xl#&RNuyn$=`SDK)Fju5CG4>1gd}M=3RHZ@YHqla;~N!FF*)8E+kL
z*X~2IGQ~Q@E>bV^t@G{T))zpm0mLqDeQSty4Y7+`-xgxsLhRz!mrktd#4av4$B6Y9
zv5Q;ZMPj{3?BdpUhgk0ryI8JIi1i7vi{)BHtX0G=mTNGz22;COuF=#QP3>a2CQ@r6
zwTtA<UTWP-?IJmIids)mJ5H%tuTVQqsaf-=9jDZ+rPQwJm8>jg*2T;&E?8mA8piB6
zrDly~cJ1~mE6L27%<S45mX(vtc9QiKnX_xG<~7zvOX+I;QgzW;nWJrRSy?EV77ATF
zk{f(Hx82I6wPdB9z8(^!dHQ*X{v-PTusofh`V&~5l9-&t8jIdMVfrUnuA;d<8q3ic
zu0Mmhoa@Um2S{>&<nni{@y@gF^m|oT-F&$Oa<M8E6DY>2^qRnHf>bGlLI_r+7z#01
zm69kVVO82kVc%P-#5<hkU9_!u{L=^9^Appi{(t7OcM?;?hy$eF0g{Wg_pZ*rv+POC
z|5C}cRQg8qc>SHRBl_O&8>xPV$)u2p1@J0`t5^W@Ddb}Td`jUd7QiJ8mS6!4XAq7B
za07!4SO9l0*ntJ`CGY%_duScgR=zp{XH}LMK`}QKi}Z!IrBp*_ZI&>+%uScMYpB{<
zR}GyNTEZ|^GR2D9aehbXsovOdd3EC7Y_cGgGTSUjrOX{INTtlZEl8!zgDps<%;POc
zrOZ<-NTtm4Em)<@0mOn;%DjeHuu7S?5DQi*b2_nLl`<bA7OYa{i^PIe%6x}duu7Sq
z5DQi*a}}{@Q>xa3QFAc0V3jgQQwvrpb0W29?@G;;QS)AE!761wMJ<{d$jTLJzCtZn
zrObKMqODe0DW&F8YQZXHUd$}oo0F9=W)5Q(tWxG!X2B|DPG%ONQukS-`|Ja4P37lL
znw&Z^@U?m_1xaIq#2YUtzA-L#_51_s2ER+-F5ciz2|UFcd@+T^IP3_c5QaDS1_~SS
z2H!zp2j1WZDICNbJd=Nz$z5e{+RUkI)L9}Z)@Ug=m2y|M@M%hM+GmR{v{8#E8bU)<
z!PW$tAZh_EP|aE!XoILTbVgNb-Jv@o06=wV{h>dip)eFxqK$wNh`xj`QSI48n22aP
zOh?sbGhrs8U*T6&dA5kaB3x;<oWOEiNw$)}O3aZ2B5~zd41pNT@dV;=rPy`?+cEDU
zum@L$9UyQ3^GO0HaV1z5fh^3I30%gNU)Kp-$DB_fA6I(aCvYEgDS=X4+4X|J3(Sis
zEW(vs%P1_v97-V+S8lDMunKc5g;-pvwTZ$e%*hmzab;E-g*40uC>+3*ScfSb#+*eV
z3s+v9r*Izgbqd#UrPU1zH!v4aD8Q9f4=Fsv{G7sbTuJqc!Yj<n7%an;Q-KTuF|T5<
z3Rg<~&fs^<e=zt1S4M4Numy7}gH%)rrKB@R$DGA#X0iI>EI!X_p2wWS+UBsXT5KdM
zw^-X-f;BZ&_q?;_SM2!$ZYmI2;8z)I|LV{qYo{t-&a!CjiP{OJdZ;c57JvVYnVvCM
z0YGJ)k!d$(9J{L~`ujmYFL7FVffpig@J5tKAdz^96DNs466RzA$(X|_gi|kZVnk4g
zz`UBmYRqL6%BYu6{FfA7VlJmpj=7S#S29$NQprYDiXp0>u9-&_LmvXF*qADY663F#
z=WB)%<MG@*o+E$pd{n&nh$1Tq+#>-e?i2W!1e~}}<erH*ai7iIv$4NyJ}Mjg%i$h5
z*k2AGlY{;J%RT?a{=y~qa9n;9E{zJu<u?(MM+E*55Fw3;z#jruNuI03hX5_^P*+NI
zZMaw|tGH>EC~cWJY3f(ii)+@*3A4Pbe;H8yKj%NiTHbxnwEw<Ll)MM{3>>8U|8{57
PK+Q&{^AW!e62JZ*%j;0{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_convolution_backward_weight.hpp.CF4258D0B5FB5557.idx b/.cache/clangd/index/grouped_convolution_backward_weight.hpp.CF4258D0B5FB5557.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ca22b0b5b07f00bc4a8df92fb4c2dede5ba4ad10
GIT binary patch
literal 11580
zcmY+K2Urxx7r<}Eo!OjOb0)b{QP(@sXrf|7K;T9X33|Ggn5alJCidQv5H)HnD1wEq
zi5)N|_5gxni4-X+8nBT9SRz;v3u;6Z^}qZtd-L@B@^Qbsw{K_PyfQPpgB82o(@m!f
z7_WRcW5Ty%1f5Rzw)$W2x8F?px`9qNzKu@z-R8n|FQPYgX?CJ^NZpvv=d4T$pXnVC
zQBb^j?%~sy;=cH%Y|)esLnjNTCdUS!epzE`JnG}GI`r-<+1&1~8<eo_LE_QHkNW)4
z%-eHD)a;L6d^uh0Tx%MA>+3A<oZz1-_6=%aXc16ZByO#aSUIISzI1woIX=96e$B#W
zWpfLU|1%<EYPTeIx43RqvjwSni!PipKYsq=I4kcOHBwH?YP#y%Bw^a@U2g_n+wt%9
zx}kShTx)miS?pJrirYraoBTtcu1)vSW5%`>UY~jH9@xd>TxH|mr>E>*R<iQp)cpg}
z5>9t-@y9PdHD?1`onJWJE2_iS^O`=q^s1kw&Gu?KzI)QZ%IoL0#LYkFe`IXH!J?Q+
z<M#M3ytYqCiaz3%5f{5^(uV1YIlE?CKiTj_()YVN1O$CK>+==6_NNcC1?HdYxNAd?
zExWGS(v2w{#(g%mW<<ju^O^>?&g{D2&%xPk)At;>%EVy@CVUn4^6^KvZ+i@^_<KTe
zcyX?+JoL!MiFxBUriVXh{gpkh_Qp#2<jig}FD=-g81hf!njw$e|JgO~PE_MQgWfcp
z==f)YeG_|}+<vUvoRrgJ`mP^cY8_GABCNO?N43fP@bhi{S;=i*&U-xU_~7a}x35k!
zl0{o1+z0Jgzf@|_?yazi<->Qr`PXi`Ih0D{YI2(9hHY<Ad3JNgtyr&RYj({iJGRUj
z8<R6XZm4i}abT%Shhd%4R?Oa85j}Us&)v6Pc<X~F8+Uy&zQ514BozbRDk*uVp)Bf~
zx~$Av)c!4x!GG=fu)~my{)c}$pVv3YmUy;%?~bL@llGT?`nBQ4s($ZwymhR5VVeVS
zs|w#Z9(!O;;_jipR&?y#`BJA7CfkgenYYTfv>*E0SMvu?AAfyJ=AS<{ojoI|YSz^A
z2R6=mkn<wBW?|&TJ>6EWKECyOYqRhAaaprV27BBnbqU@QaX8u~IHBu}q^TvJ-5tHC
zX#KkhX4e5j#*S;G8~v%+V0G&T?p?)JjrJL0S67iXQm-lfTsHpQtSIIF*cOe04mN1e
zU}3vO=1ynUSROxHk^LCQw3SZwy&5Ul8e~6=muoIgu+4HAmf~{x-#FJn=>>UxUXER3
z*)o4%YKLdV)xXXC$8*=&w0C=dT>5*1sbcj~{rO3)A9@5Yo!|0llP<F}@;^A>a>MOF
zqwo9qzOg=Q(aW+YJ?oYn+Ohh>e#d_L{R8*4i*M&F+^P$nH)CvqntAWspE<3K5RxY8
z{610>QvIhqMC2j*M5lLctn2TO6l7^ZZwFQMlPPZh(R<2qK|C&Ke{^p+u4$9HUb;D_
zS+b=2P_r_DObF8HT||cskIyGnzgLrZRnqxr3R#$B;m>sXw(mGpyoc@Wa(icGfTVNN
zRQi+t19keQ#G%mrjiMhFYwP~+;P(S`%2lE9RiWj7f3$J*aGx<Or_;D8l8$LMHX^dI
z{%Pbl9aj4O)uhR*7a@gj;&!T)VxeKN@XmjKG%EuqO_<iGpjVir>#kYJLL{r+N-N1>
zWe{n9xJAC$P}12nm7g*B8IRQITN@lI!)$Z^T{L!Vyrk=-so023^+^-Kp)h<>Xm0MZ
zQRR}Zr>3wDlXY0VO`EqJDiMp;U)<qa{zB4q)Kt2VF8y?RQv-*>_G6LT_g1fcU($8c
z6jouf3J>Sgw|7dv$;15@BT46>scc7NJNlISWbeWLk=;8y>LKa6YAUISq$1Ud=C&hy
z!cW{S!Yd@5yQZ?3ELLZyh2Ale%==1@v<J4KlFq6rL?RN2M(Fe{8#+{u>2J&%RoUmB
zr0b%od_%rbZ8U$wp>Vv=J!jn7eG?>I2Th?3X`?=_5jzwLe%t-a{x$`9lI{adp$Ta+
zRHuIzITTKHU+{C<lf9Cp>#Qk!Mm|$-(TqA2uGZLGPrdnlk)-q06w)wBt9QB<h&R1o
zRi0Tew@T7^Ybq}hd0Fo;EhQ?w@^rh$wB0J{yfl^H5&6A7Ld=aEH+txny?XlkrRvo6
z&{VD?a=qRkmbV-#wL51&Tc)@dOFB<Y<!?m(Mru-MR&?J?m>B%Sdpe!ocK4(7+EvAZ
zq<q9a>gfzFWe^+G)0sNTNH(&k(+5f#>6}J9w7hBfEb;QP{$4}+N=hr%%I@@~VrFJQ
z7G?p|k#)2?r$y<+IstNHZh(B4kKGxsiZAm8WMx)Bc4oKp2k6840P4&70vf`G*!jZ^
zWkUf4vOqwi*=Rd|@G)!*ps{Q$pc!n2-C3`cnQSJYS!@=d`E0)38MDd)wgAxg?0Z1r
zh=n7&Ge?vah^>IU60wz#e?aU9WOt^F@*`qDLSBd1I><jE_7k!@tDCY7v2BolM(k(E
zNr)vOyEF5ZWW<sory!OB`3Pc1RMyIX@*85mK|YGuQOG9{JAv%ZI;osQ>?Gs@#0nr6
zBUX&;&a6@XK<p35e<JoL<bM$R2icwZp<G4mD&%s+$|2uH>@Ko9Q%<>u*geSi5xWoh
z8Dh_n-P!w;=ZHOrT!&a4<S@*_u-#c0lyJ<#A+Nw}1>`u);;`LWoRu}0t%1B2v$c>D
zFiXI8XG2gDF-wHJ6|=37_hPme+nr5F`4zKYAtzy$1o=12e#3TW?Ng3ob`)|3W*LxA
zV0HrAolRRgiP=fW1(+3Zemi+(^Vvz2xveDSefGYMmx~YB2Y@=T4uC#kpV)Y6f66`u
zG?)zrG?&e_@ihF7eFtbBn+GTXu>@q}X_AOoBIK=zZH2rCu|3Gf(_}AVdm;ad*sqWe
zB6d(^XPO*B>=5L`h#iJpgIEo+@i2Re*i*>0h}A+~hS@S~<6#zxSt#V?m@S7Kjaf9d
z@i2?QECzBcX0ecWVYUm~c-ZX5Y&YaRnC;=LjT`I#TM|a*JWt$vm02XEoEppND6O%y
zx?Z|@q-WJ<dP%v7*iGcm6Y()(kCDG-Nm7<zwgmh0Cyc->0{inP*nrsv?9UTnJ7(Ll
zKOgY{%no3G9>~WqJBIyvpcYWK0_yB|+R7C*{BPN+J%LvL)kmver)vLOQkGNWayn6a
zuy!PUZOdu$(1w!oq4=T6vDVe;Bqc};0<=h61Sk`UnF#QYSo{$Kt`x*d0q_AL9w30P
z67eeJ7%Ik4Apf0;zjN;Fnz11KNuEtz7c^(h{diMv&)WO(O!Q=)e!Mx~5<1@!x@u3U
zc~5nQV@=N8u|za35hrPiTDq!=^(%9(cA6MZlaZR5_E+_~W09^+gO-@;^^R3Jx3HZW
zxATC~*66t5e=Bp%f}~UkZ7YOM+ARip$IWWu77;@wrAT<cNbulObVhjp4CJ%I`)47S
z2}8?-QQDm(<)Se3BIHZL&`X@PR#)$%|8E8O&^J)y20nBxR8>XChJY)s6pbszY1+_-
ze*8}RTe01Vi=;HwH#P8yYpZW-;1lPrcQ^2f^V9no_{wR3et<!n6G<7aA8z0)po#j4
z1}(}ZC0HM9;3aek(l0><zFLVw`Y2@JtC3AezX=(%*^-nUNWTLac$rK?`ZSfDWiT7*
zvynk_k)#wNeIYXNZeN1*CCI?L`3<DMfegGzJVN?M$iR!lE2Mvg47^CJ#`@LRz>CCs
ztY41}yh!|l^}k?)=6OlkkM;YpfftEvtk1><EzIj<s8NORt~L>xKh+tl-=FdMDj~)a
z;;aSQ?Nl|#E{!YxL5+X#2pIL<=2@dNj#d~YC4_i~5Sw-*Nm)vKmJ&bC%ZTqX(vx#2
z@ePH1mUy3q*PkOk=iv3{iSK!+e}VX3fLuqs>)`b-h|ddn{Y&Ee66(JqzONuBQ}1L7
z#-BocQYaYzKI*%Vg7K$P-&DwVsP`RseHHbog4f@rzIUPiJ?eXp^SFoS2hM(Xln0iS
zZMfYw43axkY!@mzdBQ;djF5fX$2O6abBLWo{yg-~BX%D01;j2uzJ%B%<j+I%GGdn@
zUqS2&<U^Ppg8k<(W``jk!R!d+Jk0W7f6B)!AM$a`jzi9%J{izs8FWMj9nB*ollo>t
z4`xwo7WbH>WK(N4<XY-e3+>g?5w+0XGwS;c+Ivo|&*AlT)LO@RR{!*%l;<Nu)Mj%U
ziKk(gOT}U;>~cq`c$9Z3NqI`er<}iiCdcmW7W}PhXcC*$lNYLQ*|$A;p$b7P1oh;F
z>N0V=Oq_jd+34o>J?<WJ)b8AZwE(S8W-_1v7T~AlxTN%Ey#Y;SQ~mhOr?csR;t-2N
ze!MEKMQknPO2jIWAFpCnh*fb8-L`oCz@7OgrmJfr!R4ml%IEN=(CwxGvgnrJatq4u
z2=X1F9oN4j^tdC~IadjC6_i&Awkn|ym){lSyHNf_ke|HHwkJ@(Mv!ZuyhiY=5&Cj_
zPX+lYl!p*Ggn;pdkbn>}fa@<M@=_>YM&xCXLx~&;Ihwdc6X2g{(k&WRsTkrC1Ldj2
zCG~aoPK8w~jku&ic{-8P;r-G{k8}e3d638lq5L4R9VEb?hlqR#%5#aF`#Rflp?)5b
z^PoJB_~jAc&wL{1L-|=EpM9MJ&O(2lBl0;YKTqWIkS`GV0_4lY<uZK!%f$OK^v@OI
zas|rEiChk^FDE_9p?@lfTmj`3#8v^H_ZpF}LHT_m-+!HL_o4m+B0qrg2gL6I^iMUB
ztD(G($aSxCKpphY3nIUO@|Q$@3HcR~UqOzfE|C<(TO{>~r2ag9qo_+1l*dvzmV$VX
zr9EOPh@Utr$3b}<wZ%~oKWnJG2Fl~99RE7o;-P*5l@p*mf%+v-5I>1jPK5GgDks0r
z0m&5PV+xg1pnM;d_d!mjaw_Bu>Y712a(`sdfDC9ale%U?c@}lef}BlVvmqBymjdX|
z0_sx${e6nMoPzQ~Di^}>6w)4r(BDN=E`stRYAb^NJ459&P+m&q(%0Em3iZpVTn6Q3
z)UOQs`y!PuLirsk-+7$_?m+)kQMn4r?^5|L<a<=U2f3EI)<S>P(tuiM?-_M{2IbGG
z>vPC;)U^)sa#3C`f_Pmn+LntT9>YXA49deqzc3NRSGXvLb9wlK-YtF4L^N+HDNh8C
zCxR~zj3+|i6XA2tHG*f2V1*o1BYeU6so?cgV4QP_M=k;Oa!Ft=0rv8UXC49e@<>n~
z0rv8VS3Uvu?h}vu(B6F#cput(Ks+Bndk;v^18A?BcvVAt@zf)pg7L-Ez<3JAmq0xe
zC>UP?4N9P3e2LU6k%IA+Qjbz-uapLsLVIP@vkclRqd{fR-bLzl5!zcWdMp>g_?C--
z%SAB0FwrwibdGPuY5D%lM&mcXAu0aMAG|^Mn0*Z1Abid~2X7E2v&mp}Hk-``tFxtu
zErqMIR5Boye9EWyG8u50v$ly<_fPe&u=wujp<sL{4A(A;9W&~2nnyyux-y)?rht3T
zVROK}mm#(c-unvae1*U1*mSd-{ITv{fx0p@i6(G%(nIV4XuLQcPy`Yq;EMJd60bpC
zLc}HTT;>cB&p>`o#OIKcsF=i8ypmE$#Y)awk`Fek$uur}TT)^Wi$VT;p}7UIEy$m5
zzjq<F3;FZy_W{HXsO;Q+=OC7Y{Q36#3}R>C9;XbkGPsAih1e~4M)xmb|H3o65X?g0
z8C@i1k?@ReBW4@n8Ql)dcEB^bbj;G>9w-O19Jt4cqyr=A5N*OGrGO4B;GD3o`8V&H
zXKy8v;=)|OI=&ri2M)A6m<Kq}x+(O(DGcH!qe=g0;^f4bJ=es;F4vo@M=h47R&dl}
zX=?>XEf#mH7NwHnXYsT0ZNmV|04q3Zu?)9@qZZ3VE8j!}TY{~82eSlOmLMzN!9*cT
z6teQU-h?chkd@E%4rJMZtbDH1kR?rJ=UitaOE$9dxh_POLS*Ip-4bLeK~}z>y@4z@
zkd<#z9wEykWQ9j9mRHCMk6J9Nu@xS*Sk_~!cJ?7DzhKKR*b0wYEc>w)9<^Aqv6UaS
z92a^W7rx}Da>atFSolyob^BQ!{k{FW4cF9U@5y?Ck~EMF1SM$%8v#mEDsfBYn?P--
zYFTl-_~wP>1~uNGleM7Ge5xusUVn4Noz%FKZ;CZ%tBQ`7;9PMxHSXpXlSh2#`a##D
z5lB*!F-ztL=P7aegS18R233%PSqko<`D=fBePB%C*4wHeT=WX(wfjhbkDKSPrFT?E
zG-J&`(7(st13}-Fb$u=9-EIo*+Vg2KrVhLQ#XKK$B{5d=!&xo#R5i!zd9HYf8V}Ja
znmsLsRmJ+3^jxim7;Ct_oQlZ&ijL8tlBBqbuArRu6Z--BTKpQ&1|)8PrL`J~)sVMg
zX&Y|Nqvl^h{1+aX=MXUmmfCwnya#y;6}RwGD=8&ZEP)&<N};0DeqQsm!bZKGy{$IJ
zSj1vsDUL@h9+u+Wi0y`@I32NcSc-EI%Y~)*EMjM2DZYr<MOccXXy+*EskuCVQ01@J
z_oPn$|ACj^HaviPAXM}V6`icrEOl<`zmMQ^^FlDb;Jvf}pX{_izdc+Hw`k%PO`Nf!
zRY3Ln`p+l$^{a?+6&a_EM=NovTKz{9TrHCrGkFM<G#U{Ya64dw`ugWKVz*&yd4kvz
z*jkoiwiLFOD9oZ@YuSX^CfHhbVzv{umV=lbgsmkPvs~C(PSMV%xIeYBq>kV5jR${%
zm(=)@I$vO2??N7nhNi3Q&*Du@&6;~8rLC#0S#!3exSQO~ph1}Y%%DM-2ADyEFby|r
z-jkGxrio_IAWXq#ExFVeV#u@vnYHAVlqh71LS`*XC1n#bZ9-<)AWS=uSu19el7>ua
zDuV`L%0_0`AWVhG3>$>01esxjFx^0A*dR=gkQp`z(<@|#4Z^e<o3*MdDeJLmJvPGz
zVfqD|wSrfF7H`^*&Du<;2b9>9jm^A4R0u8=LK`jOu3w)pdEc;=uhfVQ6NAD;=Q|~B
z0P0jazQy5Fd5s#c@#=lE`qIdRALbublQ|l(Xqe2K5!(!tc_(5!VKVPWY(GrqV~8Dt
z$y|h35lrS%#7bc@M^d*){_at8xjJIUr$KzgIn<cLgHQ9ls_6JYh$}v(#>c!OYm=rb
zIzAKPih0zSM<;1sto>`plnd>zTv5FkLpsOsX8A0A-LcN>bebw~e8#60m=}$QrCyqq
zcSC(=(UP^a@dWT<UN#=`-B>q3-pm_NFV@S()8PyDg^ee`m+VVGqu3}w)7Uf{PmWN;
zLXnN9!*axyLk>eM4DuSp)*u^CmpsJskd23bK4SThk0W**@*TwPAe%M?lCl}I&Dh4n
zc?)J+Aje}C&sqICYx<aa9|nIjE_g`&4<$fa{i4C`8DG(Uz3jdF^wrtx|F7d;FYSj;
R*N6Y?&42dgKl|~Y{{x?S#kBwc

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_convolution_forward.hpp.3DF8CF239601D4CC.idx b/.cache/clangd/index/grouped_convolution_forward.hpp.3DF8CF239601D4CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5126d61e64db6c868c4547f0e31294096e16c488
GIT binary patch
literal 21098
zcmY+M30zgx_Q&tp=j^-B=^eYb;ab<UdmGHD)c#`Skas<in|ia7Oe?Ll)ao>w^%`C!
zh!cncDmb7Bhzg=8iijg9&Nzc6j-WUK3M!g_{txfA?_Tx!yr<vKUT3)bti9H656jSg
zUhl`D7EAERem{LRYU~KbVzE4A{_p#-V?OU|!p$xg%TI|Va~r=YedY0;H<n!AcIxvP
zdp71DUG{Xu*cUo1ulw=vJEeO&#m{?uW#95)<#2q>{ome5e6;r?-{wENBI59H&%F!h
z+I%$hx96um_D4g|^2y!ebMjwGEpu<0{#)%IQ%B@)El@j6j^6oHv|KrC;2+;@8ry3}
z$7jN#?&q~^R<wHmog)bye*f+MtdGw8EDbH#F#Go6R@V-6PdIkr=igr#8@2e4fsdZr
zx9;B9!HuI{Y5C*5r++Vu`R3fEW4=_WdS-RIoC^mVKm4&>?=~-#_I@Pk16X@+sJi=!
zp()o|$GlAL57q8RZ*HE@``E#XF(>ajKZqRqeb)E8zT2+Eb)3Ixv+IL5vwYuvarsu`
z$?m^yO3K_daP9kB5?(kKyJP5v=QmB9mf$^h;AZxvx%Iu0TMpUxR8{fu+o3mKt6#A5
z!=9fsG(VIxY;EYJCBxPpSYFtxTlU9MNehN7&w6wJukE|Iw9}#8PANGbPWw7_^4H|O
zk3PyB*!|^yFQ1?FM5o#F;Aru&s#&|6g<Y+vj*mPyap|pFKlxgJ7kswO(uqm``ngB(
z`sMov?>&>SBY(lr=G%Vn6WmvB`LuWGH}Adiz}N59KX<)dowu;Q)7fn?zaGuI<yq*6
z?pE~J>$N9H<;B?*Rk`=n7suv4@NJ9D_p)Cde(LQ9qduq{kyOxn$}3|=Tv+1zHe{1~
z^ebggJlHHY>p;P|0ndFmZ&X8z*1g_yeo|Dsco$5QL$;O;8rh<>XUSr@jgR$urDNjV
z)oIO^+$b48|MY>f#=@SH7RC3;h<Tvx<9}3M|NYwRvh<4E)2>aa@f`?m?R<Uo@LsbE
zz6j|Z`*b_{aPfDQpZu_E!rV%=m1Fl)uWcE#;w?|t|DF#@=vQs~@r|^%8qd%4wTpQx
zcEPs9x6@Le?)Jm{w=>@A(6wUTkMk1-CA(LD^L5*v^QWBLvE}XFVc|tVcPHdjP9D@?
z&->P>KAZjin6GDqzMcB?#r2y)YrhB$?LRPO<lx0q`!)u(4qh_7|K;5At#W+Jeotv%
zQoOWn|B?~YN}oz=JMf2A@3ao>FyQ0W9ak+r`Spf3L!$deznUA`r`mO`*{s7ikM8fE
zHs`Ake=C?mp6%MF>)XSsH+>xW)fXcwR~)+W?%ucuPku1J`qj06X54jV{kB1RAa;`T
z!3|j(M)w=~X7SO_-|;*iuiW&No}Rg8)2TLNS5$N>c%*2{q~Loyo-5qm+%nwzsy_Om
zF>m#0?q29?+57x6|8c~s&vsn$(E9(pn0W5s&W`TQRXZNE)+PS&;l6MPJ6Jb1KWK8;
z^#|M5e0SyJPT$i$x8IN-e=lj?0{8sdpN~f=^E!=fHKu6B2V-0*-{$m-dHJguJ*&yj
zca!S!r@WgT{YdM6rDNqE&IJwcy{+{FQ6Dap<TCyBuQu#$`@=Iu>qpxnX8xc3wX(Z;
zf7EsSi(T(!`*nA_<7K40=h0nikMf}O83j{ZpN%_z+v)x*#QRj{5OqVpbDPt`ULG3T
z*!z6nh~FPumHI(J`L363%|G3^BmCi=cbYv`Jh%S)pTGKe5HkbItGRbtMP1&lTYR6G
zWd3K&mg#J{zeRiILCbd^ONL}gTc7-Y-!?v|TVCTP#|b@7ycW&%fZv4vcSXkAJr{na
zTRL+S2X*wZXw9Xs{&x?@bITh07Y+)yk{3SdU^G}+gSGYFe^^xibDMuD-Tyl3{&n5*
zCcnpZnO>LsS+v%6|6^LMduq?4|GLy(x4h3yB4rxcWWrze$(FNzivKjIK(}<`Cbbf+
zmEN#utyKSAo=iM?cI5KTVY+1iH@PR%d-9v&K^+I~sej)$yq#_d<tA4ZdQ}OvXl)<y
z-=)jmZl#W?*+X<oA8xW(qKhT7SDvu?O<wsq=gi21KUe6M54p)+O83%%7Ol<0ev?-V
z_pH4zyv5tP<t1*iL!vvR*Th{Q?_&9&8##S-%MfmIm(sh`{L$?l{=0aOzV_R_&dXZq
zmhRl-s6>x8y$k=*eZOD6@oJXxxo#QGO-@pJvgt1E9`)a)-<`Ma{gt<=zi#nzlUPDy
ziTN&1YJQXc$@ktL`?fPlxAf&EmkGVx<O=S8{cW;g)0L0Ir@gOR26B@V3O%8`CAx0F
z!8+sJ=<(sY<t=V9SD|wi^BeIm{oX`Ra#n7ERkw8ICd(*YM$K=;d-sFwXP*wv|7oCZ
z>B3DyXo%@^PR0KwAO4s+d++Axak}M2ZqS;x4z_5I5x>ERu8{g2`(>DJd6^q*pmYO$
zT|DH|XEuhGU;6N<ZuviM(v@~K@A0_ozsKjnZJw0f^|89e!wvQlx|f*u;Dh6vC!Z_s
z^x(8w-7<)q)GM@JF@FnprZE!^&4-Z*H+4%-Zjwc478xM!GM36cN)HWLpj(D=lfyDS
z+~oVu+WhY_{*yUbkMA4aQnvs%8AV5#Z}OD>w?XWGnms;yF4Hal<_15|AIxs&1{0zm
z-2Hl+A9m@Mce%kLD_vwYO}MYcy!q!>A0PZ~x^8)eo1B;EdC7bX_q7@Q;giJgor82s
zFK&`dXmXPy@jjX9s;L|Huj+ZaWdt{=q_mQnVTV6x){L}8ozLugSGV-wCgTVlM||Qz
z@!u}2SXt-Ts9QR5lX!{7H;tYa^y~ixp7`a}XPOaazRP>u<SwCiiTR+Xn|<>?lR1+j
zN4FelJfU0O;3iWQI#n@;DR;+(u2nZ<+Q;A4EpKy^y$ao{n0MiyaN(lj_uH-i_JVHt
zh?{Is=my37KHMbrp?BjST|W3H-4eo0)=G43(=ctx{9kC%?zU~d8gk|t-SRp&*(%em
zO|gkbxHLI4{+Y{{_vn`1+~ghlj(HFM5m$DZ`QNO|Jw3W*FgJ*|(s--+dwAcjx-sZf
z@IA*Dx+RF4%%gN3HT{hbuWa&Nc&qlqFYA{6++>YR*T`lF<ZqJwLi=$eTlE{ITi)R&
zldW{J)qGk0ply%+KH_TV&K<gC7&j@GX?at6;eESja@pKH^MY3EmSAo&L822R^E>lj
zbmnxPW8kj;8~iWw{78!t#U6-a5C8oK&mLv3#jh@r@2=M^AMm>@Q|Pj$FZA3)-~R7A
zm%E!qp6b0fS+~5(O*Rp_iI}g;-SSq~U%rSPzw+wCq@($uxy(M7wfy@J{-FC>m)ymp
zIv&t1{rFuL%XG19#y8$CV@8es{spg@)&9<{MEg6ZMsKZOm|51ttsC1^`!@9*{y3gt
zO+$ZX6^1ia`%HBpH{^NNG;GSR!fXw*uVHU;Gp$!u@3ntL7B|z4c@oT%LIMNDSSG<T
zDI}1*j13ZOkU|2UY3!9?uM`qUOU4NaPDmkv1Y=y5;Ib4F2m;1!32sXvfvjsxATWW1
z1jeorLm`Hu`z@rfkcI^O&sa%eB@GFLdn2DhJ`D*(W}|@hD`0Q%ANIiAg{Q~+y1tz#
z>4pnjKHeX?;RZLNr{F0>ouHF1kPM8@&>7K-@FF4)czl77ZvX(If5X2Ky$Nsn0vX77
z3*JIB5C$Up2tM)!#;h?Mh9epQBM^N9-}nMC-1rv0Mf4qfhv*OZ!#AD>b7PzY<0Qmk
z5`<x%F2Quk7Z_(oj07>5V<m{iyikILl27<dvINPPQzS^iyi$Udl27>5DhXC$&Xyn>
zbG`)ml27<tfdmDZw@R=T^Fav?N<QI}ha@<J`LG0sF`t#-tmG3uTP8sn=5h(jF;`1a
zE%}5`Uz6Y(=Iat%$J`)6gX9xF-zY&N=6e#{!yHK<lK8}!h$0Y$c`AXanCB6gM|=S}
zW+W0w#5|wCe9X%TEF-=^)-^H-WMW=UU^(VI0(rz2&>Y5k0_!nvAg}@RP69iLFAzSB
zT?BSv-c4XP<`M!W#1~LT#!&)CF&`sv4D&ex=ghRury0h10_QPT5U9X>gTM{q3usZ}
zCV`unZxOhKIYNdA*%wd(#snEAV4f($M9ec~m?`@LlMQ2*46`uD%Mgz_MTQjF$D_S&
zq{@(rd65i@Ft3zhrR)nN5@VGNt1xHFkd1k>44Y+NAR!xDWY~f^Uxs|lMKTo0zJQc4
z_Q|jh^L`ojV?HUvN!b?=n8qm?PGLSR!)eTwGE~aGKx#LtWT?ViEkiZtyE5FBeS!I%
zQ7=P1<^~xWFi%!svf>L&(~Kz!Ou-zfKqTgP1>zN7K*kvf3M62jt-x%|ixgO-_yV&K
zBTa!c%!?IRj5$Yv9K{z<7{*!!)?&_8AQ$s?1-2``fFLw>D6j+bP6c*iE>WPwWUeud
zqY50wd`y93m@5>hP<;LQM9a9Kzy-_~6}X7`mIAjFUtmsa)G1Jh`L+VLF;Ac{f%*c|
z7GolXiI^u*n1p#2g;^BE=XeV7m=h=@U{0fuMtwp9ET*s+^AZY6Fy~Onp+2Dz)>2rD
zIhR5%=Is==Q=iZfJ1FeHypzIC%q0{`s4p;EH;z&`iuo9YW0)^cxIle@8MSed!bQxN
zC|tr^N1=}T0#hF2Hig@m?@+jdIl>ANR$pM=Z%nYl1k4kyFcEW{72>SEfEqVuSYZa{
znO2yId7%{+T77|Of{|>6WXvg6NWr|q3M;I>z}(QtvO*T-l~!0O_@S&REq+*WWm|}D
zbbt=QLU=t7&m-yx9T8<pkSPTV39(#)<(RW1$iiGCL6H<JB*8uj_F+CC!2!%m2rMDN
zB1JAGuoQC!feg$!1ae5QNRev^ti`;Jz&gy^2y7$4B1IMwD8#&jzz)oZ2pl58B1Il1
za2Rtjfnv;O2%I6oB1N7ha29hpfpW|-GQ`NiB8A1u5Q}++3^On<kYRxwEK*pK3`v-i
zWk|-HDMO|lEK=BV8J1(tk|7K8dKuQs!6JoikYNMnO)_l4yi0~%a<E8YyJgsod9Mt6
zF&8LMpahGMyH$a$m<tss#C%ABLna48?qLNEV=h*p7;~8dWlFFJx#bF!V?M9IdCWBm
z)F{Cs<kl)si}|JkH!+7%2&2Iw<c3oS#~eW+0&^^dSc?2VjzS#fnG|MXPNtAdgGI<q
zp^$=k5rsvVS5a6+gGI>ArjU(!4TUwB3n&!OU=eb+QrL>QkU}BmLlh3tU=eZ;Q#g#d
zm_o5&zIrS@@b{WApZIc3Va|vBYsI`lH~v)Zf2xD{#F38#)6l<SEDX1*_N~H@t9sMW
zziuoH>zKWcy~pp!<%wzNUp*Fv$5s1r_5FZhMQP%&pvre2)Q#l|ELXfjNNiGIlj0R#
zyHA0ACI`Iolme#|ukgAm1*#OU@Tx`y8Wpebntc@ZQLpIL(-clquSitY6soCLB%%fi
z4b&?VO{5hftzMBxW?Nyl)hiOkVk<1RdPO2AQ$1yB2>+qa(X}&Me$i@>`9p(Pki);K
zZ?Wh`j>K{##Fq%WL=Yd6*&!M63Cd1T#H+3PYHJH|zw=i8JmzG^k{Pn!#@IH@dzrqM
zq3g$~Y@CYhC#(8o6}9K8EEl(*SJ`>N9iQIx>!iBEUQ{<)K#M+N38Fo;?<2BZcj(?n
zEG6`Z{(Z!9K@#hm#9rt9#=ocON&Za+;Ykguy+Qqe8}jsM8v3^ygyAt}KgI&fMBD>R
zL;r?@FubnXudBoO9eDzkHN(H{zzsDciqa_h8fSAnG5en^lwZhw#(aW*GeQ_n)9ll<
zz`*0PaMRGgB_RxNGW$*O&fH5)L;uEvFubJNFNqP#=f$R>e|th0rm6NcHSmsnf@m80
zHz|Z+gldlvt7?3JnTGzY3SoFxwck~T@?SmR<+yK}zc4k%sv8qo+lkEFi}(Mh8<SYu
zNth?Iwv#c>Vz19)L-<0bZp5?K<1r_&*AoQuxrzBk{+$u=MkiJKNi`6K2EFlcn@h1h
zX;<B7t~Iv_XM9e3&L*7kCG91fa7M4@wFze&tPQq_^{LU?Xq#Ar`ceDQCf1oIO4>xp
zCf1P>BrQR*iFKkSlD0&$aWB=4TuIB7Y+{{fr=;zaY~1H`<Cvr!lWg2WbfZGjDkPg&
zhq)nXHzXU6$hr|mv@l{5>nL$Viz7Df54w>;v=m|!B5x(pRuY>Kahr*@nb?F#D<WDE
zu?Z1YO0-g96C&#h(XJ31cO%`XCt5wRiFJf1S&NcwJhkY?99f$q+k}W(B5O-zn-EDk
zvX&#;gb3OuYujX-5IIL=?TBm>^YC-Bc22g5d3KGg)yOuU^K|3BtlgJwJQ3+ew4y~T
zHZc#Lr)cvO8}}vM$W*jU#U|3vMn&7G*hKm%QnVtI1L^0aqMcN1e01nWrJ_|THXg2Y
zqe0Oc6q`stQ>iwUqVJnXwM1$Yq9c=Pnbamk#YU=aq&6WMil|mZZ9I7C#wn_uqBb5F
zbfb!DRn#WZS3TA0sf~vs-I!w4rdVwveI;161glM?uQaQcX0?g*mB^YUvPbyf=I>>W
zIsaCp7;~k}UdjRqoev$;uxZ0lm~CM84Pp_Vk3Q3^Y1>hlWiorFSknIJr^IiDZ`ye7
zQQcU{x~yc~`TOX`Di*Yg^$?uRg0la~L$cY2qWu!<atU95nFU?O_E%WYm4EV(EBN|k
zwM(*!o}Z!yrKsrnscKN_KY2*1ik@Gjb}7Qw?^A>JVf+1R(Efk&kp1}jaIH(YhMqrO
z3mUJX=SOHk5&z^N5nAB++>gxu_V40~9zVhCC)g0)QJ>Tv9`M5>A9W$Rv4^rf^idJT
zD^*sh26)tfe{Tv~wfN%!y5R$#mw$KNcpY9x^ai|vXebQzif}p%h9UY8K1B2deBl+L
z_DlE@(bw=bqF>=xuL#5ch5sV@9ezg?B|((r6(M@61XD3jlVBR=lM<YiydngjlHe5P
zGZLJ^d_{sQl2?T2s}fwrTqQvj<~tJHk-Q>=|0Tg+nCm5|$2^h1MB)`8eiDI6n5Ph!
zg87mRmt?O9k(Xt-jQOezS25SgP$zpu2)!-CZOnhk@E7I?1tJu$2(c3sn1Feb0+TSu
zDG;Z4MF^guzzoc@6qtp1p#lpPuL#k}3M6ArRUj2}ErnXF;BQd4f%z7NTbRdLVVo5y
z{4gtoVIFUV@tCJsVH%zROt->x%&}I8#hhq`L_9;7Z-x1oldO=0Il~GWcm}b|3d=Ar
zx59GF>sY6CtScY6y0MN8Tqj1dZsf7fdF&<8p2tG-*t>#jSf?6%eGMB}gRie;oon&+
zwJfw2U!Sf9rQ?v6t`1HY@>w@B)UFvgq-Cf>GK65)jbgP^v5MZmSRGib1|pGel&GCc
zRP_EOYG{dy-ak_76se)>Bej8%8oEA8>l~$_>!Y;LC@pY3mwaYOY&x|f0%97or->BA
z6_8|37f93D6>i8=yLr9;#0qcFjeDy7o|xl)<6ApqW&N-)bHlk0>w~6NgV-QMBiTqq
zW7$|l5fY2Qlf)8<m0&KDSs9)rUZw0Ro+RcnmWwBS*BQHxd8x{lib<Sq98uX3%u_Tr
zMGL$S7ma3@_z&XnU(VH;*(LttH@rbNHmLRuVwy7cZmZbLm%bmX>&6fm5+YK<r|@Zr
z2>btm|AdIpK1qT}QiurOGYHHeAwuD%5=bQ>LgB6=u!@B6m(`6e1h$Y6p>X#R*hfP6
z_tlLP1Wu3;p>VGfxJp8V!fjxl2G)lUpYg+6c$dA{<wu)tJONLjnQlvHiRejq5>cWA
ziBcaiDVQ(8e9TD_Bw<b<kU;u~pE#SqY|L{B%*8xThH-KqA;QCC2*W&HhVhtJDX>cE
zBSdeu0@;|?D6mE_|1jp~Z8|z7KJIeWzFZ7k{>4nQrUO*MY=vrHA!cV0sek-3VCnXq
zpXkN{)?on);ytSyNz9SNp2XZGiS-nm!kj7WY0TYHSZ~3p%$bVYQ(5;^7A)G+nKK=?
zr?c+qxIKe8GjMwb>yaV8e3KS+uEg!Dm~$28Z05}VCqp*tC!W8KIoIL#b*%e3{QNxT
z%){+@tVbSxem?7vkNc~DISO!p6|kTJ_NsVZA#)bu`xUZog}A@AGv{{PzMXa7j^C%4
zIg4?7G3#E8+e?_U1h<#49wqpFFEQsO+<uukFJr#KoLBzIa0T~Q4RhAu_8Qi`20y=+
zIcsryE$dN>pWnbbG~oVfWR6DMUyUrN5%*V^>I_qne}$>t!c^p6;i@wnw}-3U!&T&8
zv8po`x5uj8V{v<&>WstfacYk^H4s^JBUyDO<MtHQnSwc0b*BE4Ayq~Gm99RSuC@{1
zKV9{tt9=D$s843#_6!v=@bzod4r^329&=Pjj*7-(j@mUx?IrHFPIa!s@3T(rwhs4K
zp6blQ?RjeVJlr2!ROc4lzD4c61-Iv`&V1aSulC5t{Z*toi*Wls)wvJze$~1EpA7r)
z?=4oH#kjp#?Ou$ZU!pooaC?c`qXa*{OzlvH`>R}al;i#?SG$(u{;E)&75IJ?YPSm9
zUl&y81>Al??S28jPnGJd!tGUR_bS|8tvai5d$ro58o%!?)p-lI-!fGkzP?U%*5UR#
zwMQMkK3sE#YiN9o*PP=qM`+H7e=<a9$Uh@BXC!Wq)VfD%X#7NJ&M4d-rS*u?0^@_P
zQJDF~fBsfvm^jTIrv=vkCS4C{+3mp87LVyhI_s2<rEWSKn2x1x2J4)GrEUfb&A?K(
zm~|?~*B7&a#rXOX*0}^<U&2C5@b$53r&ug?W7UDNB4g@CoZ2}KOWim%G!9GMEo!GN
z`1&pCz%BUte6@2vzCK?K&BxbQshz6u^;PP?Dtvvl+PNBEe@pFj3)|l^7lg5Wo!Yq$
z+t;a~b>exG_qMwBb+eI)PThDH-bM26Gx!Y2yPx1EB=06mFd56cITFmlhcnY9NXLgW
z^CZZ_hckCeup1xFJTAd;d^q!>1Q+q)%v%!N!iO`*6Bv&VXU-%r6U)m*1Qua=nN1)Y
z%gcNM`B+}=C$JyO%aa67VtH9fpc2c=X%wd6!=MW&EWn3BS5R1i4})%|u=$^dLE9Iw
zP9i8bvi6OF`Kp9DX#6MA#h|&x?6-t3$By{u=ISm<+f8ZyKD>{<odE{=cE7+c=-Wj~
z5Q)ECJb`%JcZ&%u#(lSjz#80lTM2B%eRq(+LELwz37p1#S52TA_gy3F*C?iJd<tOB
zPy9#g`ETW`qh=@g57hGpb9q>^M{9vI$9#>&H1r><7lyZ){WcrSyO~RQ)6jpoUKmy~
zd!^VD;+otv^dGSoh6h#qL2=xfFOQgp{)6_yuvE2|iuv`ti!Zf$!&1Hb0o{0nJrbCl
z>BdZn&BT+#7{X%kj^Z7{?qE(=Sh|9)&tfbKPYy3Kb`kSJl`X`R!@Vloi#beVVPdj4
zKP&g~+nXQ#?FrpzB+y8_yeG^)mmymAiY<b9GR%{`Vv8VMhIH90wg_@%$d$cfi(tD9
z+hwoVA}EoeME3GYsJUe#Lj_*nx*@|2ynYp?Kp0-1ny$cfy#6#_f%$lSX@vqS@DBB6
z1vcXy>ir7r$2-)g6*!G|sH+vI#yiya6u5_XsP|LYk9VlgP&k8ksIO7DhS#qeDKz38
z>L@Ej;q|9ER+xj=mzG#z30^<SQQPOJow+OPM!DL)TyRoCi!qP5en>XAavz7s(ImDl
zv_)IFufQv4E4MH7{l`{r`y|$pKY;I@n4ayg4-uaImumk@3<19BVjB9ZM1<iT)qY1T
zBk<ib)6iclA`Evj`%bZcx~SFg|9XF#bKYEnnJvL=oK}`fuoS12brP(@X=RrLyKq|h
zQ-VKnTDc&>1)NrHN^ldWm2d*#IIToeh{j=lK85)>%rB>~9EbT$6gJ^7pU*t`Vs68o
z-|PW@DU9fWU8;STScKwBET&;oL5wiVRPC8!x0mk+nP&d77-1Nr*<-Z8=J?7uE!X-)
z)*Q3e;7<yF(%yV$itlKdckoxuh&vo-_Txen@o{V#`fF!|;SSZlLo5{XJuB1DUqK@b
zr)&1<T43*KO&jf<n0?8Y&3CF+2Um+9zdrbt7lH<dUo~C%F?bBQavNxaT=`{q8M*Rn
z@EUUE1<bQROgXoni{5^&!}JJKSw732MX@tYVre+`ZkO10%nJ!yh-2>*nN7h9ck30l
z9>?N63fqIZRAHr<uPE#a=DP~Ji#d|ANW2g@m$JE-Gbqczyq>c4m<z3XAwEq~Y}Jb~
zFJ){g&J9NxJA(NP)6d}3B{3?C!MS3#s?Wx!OE#-)Gj6}BvYUeWR)*;b{z56?3CC3X
zF%dxuTddyKtk2zt%<+Doz<r!_rpquLC!P5+%*RP*nGDNt(#ew{4=0_SGVH`j=a>w~
zaMHOb!$q8QZpm;9C!O&MjK@hQR)JWYbdnTE!bvAb?VqE*#bYR6{WN>PUk4_7AWOAp
zsR3bfyxIHj1YZuGW1cr!slZB{w(=Fo$7$=J0ta#0I;+50oVKnja2=<uaTLZ;bg=3m
zg@ZV4l~E|eX{&}p4NhD4Dcr|tE7}UtIBm_d!aSU|(yfq=(^k3Kzg+YPKQ%%g_`B$h
zpNG?pMD}bVd!Bb4U*$Br&R_f{x^ALopD4s1Um-T-qrU`>8&-6tH<|6RsHXEIj^-|O
zp2YE-i|=Ub#!HTuTzqpuH@pt7OKgq~b_{m$P^=rH9iv@*QB^m7bo}VzYp3QJ8p$zH
za`Er28wrvlL2_}usT)fq#}dhf&yzTEB^S?#y0KGo?37&iJc;9&<Pw{i6_TSua`9<_
zZrqR@Hzb$Xj0_`=Fyi7u*qosfM;vj9rRWsmNFgpRwM{Q0j+MkER@*ld$7bT<eW@Eo
z#8E_CVzs@LI7*32thQevjw{3^R@>`|qn@~gSd5Y#QL;;j#W}KLj_eX*af$3$BD;iG
z%#j^AvWq7fv#3yZY?EC)Uz%GSvg3&C5@PY3>^LX8c!bf78re}JyToSbec5qec8Sf-
zXvGn&xVTz3YY7#{JjEq8D>D^GrsCpGryCm;$414)rH$DiilfM6be_a<QgPw)B#uhO
zh0l{X8Wb1bbkdEf)G?K!@0&;+iPXijuX&z@Ix?w?pW85>PaPYni=P89pHCe{)Wx%Q
z(|HodDeA)KNgP$w#b-?B`5x-1r!IV+#4*L{;<?%MPpc!r>f-uAH`1(*G^>lR)ie!8
zXCixs55|i%r-miX`emaTpc@D@-~c^UhN(C}&y`^=4$w<wSc(JmS{c^j09`0UAr8>R
zG8E$geO`w1I6&9RP>TcfI0eSx06k5CX*fV9Dv*c+^cvN(Mx67!xw0mEXxO24U+RVv
zoT$#S6|_Rs4%(s0$`{}TMBShps-=7tUPaUo`l0H{0Wbj3d+;8rm;3-eK=e6$jw&R-
zg0B$$3_qj#$lu^ML{lV~f~y~+C5Xm6SAw~?+Hrvd3ovI$kb$ckS4gk|^Lh!^<7&pu
z5^To2M}j@LdU3x5`!SbFP>QP+PfKtb^Cbx`;p)Un2`VwyNl=HY5${TH7jp!G2wZ(Q
znZRVsvk1(>)rNBj%)y*SAPrX+rV~iVyqdsjTuqouAQy81fdX7TxShau%m)Y@z}13B
z2pqwDiohvc9au)7409EMDqIa%L!bupJp%U#styd7Askl&PLs`lY6n#V#>o(eIZ=j0
zT-~=&hJ~0jWXQnPd@E#Ffq9(_>u~knMj19@-XX&dT&=f9hCP^%%5W4{=atG(iur;J
z7jQM+6&bEzzA3{^Tzz*(hC7(U6$r=Gb`uqth&e`q7+hU9Q-PV77bvg*SJR~^kb*f&
zfh=4-w_1VKn71gf1y{>$Q(znB0}33#)p17@ID+|%0%veF+&Kl#VZNroHC+96LxCHZ
z?<;U$LDg{MDU8R}Z__DE$JK2!D9pf|L?H=Rv!zl<#hgVU3s<kLrm!0G77AN%wc0ic
z+b|!XZ~#}Q9ieap^H~aKaW&d`3g<Cjr*IutpWUQz6Z1U^_b94H3%5czu0ES;g{io@
zEXE2knCDtyF0Lk9V1)&kms(+|s2(%3x)m}pk7LR>W)o-hBA60^c{*!5ops?FSvO{|
zwlf6tY;2Z}`2RTvuAT3$O(^KtWAAEnChh;{i1-ug`>cOQJ~~T&$&|qE<{l)0-OqWD
z1a`mVK@!;Q^&koC9_&F9*ge{VB(VEO50=30iIN9PV0VJ#!4lZLMDkz>?9P=uSOU9u
zN**kM-Nz&kmcZ@`$%7@Z`-bGf64)I^JXiv|<A?`KV0Q}fU<vGANjz8ryEhXLmcZ^J
z;=vNwT}nJy0=use50=30dg8$n*c~N%umpC`kv&)fyO+owEP>rQvIk3G_cqysC9wO5
z?7<S)eNOgZ3GA+sJy-&}@5>%6f!)!H2TNf0JjH`0usc)nU<vHrsCcjhb{Cn9B(VFW
z;=vNwU8#7m1a>zl9xQ>~Qz`nsiPW7)Jy-&}GpPqlVE0Dq!4lYAL_JsnyH8OMmcZ^R
z>cJA&T~9q&0=uVJJy-&}6RaLAf!%3VkC4C();<l^Pxw$WB|KB7vNn8hHrfYgpZcco
zf_b2*2lNOQ!>1?oMAQp<A=)UxMk!dh{3Z!DVcsIa7R=WOTqD84#jX>$j=7dVE#^OE
z_)`uRE^=Ik<Css#Z~}8Vg>o7!y6_x@bC@eAR0!s-YEH@h|4a;@Y3oLzYA+PEC6+Ox
zz8uliao>;m<ol1B{_;iB|M~wjGHM3ulLqyi&%XR()aXw?``q#=Z?afMiNB5(fBj7S
M^>gvpFU4Q~ANDfmF8}}l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm.hpp.76EFEC97F28430A5.idx b/.cache/clangd/index/grouped_gemm.hpp.76EFEC97F28430A5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fb85a305bc7c2ccefbffe85eb14c2b1308253e4d
GIT binary patch
literal 4884
zcmY+H2~<;O7RTRB4G?^qutW@7;uArgAmmDlfN?{KC~hU-LR}(Dix#k=s8EWyjJPXS
zZK>12WvHFHFfJfeJB5l4XhZG618SjA#%dRBr$=qAUFN0u@m)EG^WHD_pKrO}z4yC$
zAwDin^b&;FwD{Fo=_~Ag1VQkS|8rL?&x}?GLYi6-Rv-9eUslfB;hM{jhn{b%NRo2*
zWF@6dn*7PlxIMvHWj}T-?@|nSy)3KrVCENveI}fW$lO<IoS;Z^-1*>(fZw0%xHbFV
zzqHj3?2ei^{jFJM$B-wxpS$MRc7KsWdUN33r6nPTi+A^YS-f!m{f&9bx4tTi?JSzT
z^`oM@UmP2elDqe|-`TvfeZ{2@9={zf{uY@q&zj!n(NB9@2UOO7J#^8e&u;fm4*J{o
zo#Q{UuX_B~JNFm9^zhb>d+7Il!z(_vteC11Ki!k&S;B74P0zo*{g9aI`q}W<H(X)^
z5<R=mt*c+{bN^A_P}_Y)(dS~-<okZ#-V0c{V&2`)p6?K=3nL#5zHy=K&if10$rt}V
zTWk3=eP`bs#R=2gotI4S^_Al1tKyqp_)qhjd*=R;QJAqVt<)`k&fEXMrPa2y<EQsc
zU08a3iQ~n*DL=0F*j5~Lr{#e6w-54@J*JoI-g&pO>s8B^A3l#U)?f1JJ{+LFlG+?|
zc`tsKx#XzeTCuAAgZ{q;q>TMxzp(c;Yggc=y1-jcELG}1+#mcauPszN%5B#U<C~#Z
zKF?45^?Rj#&X~2j?zJ8h-t%1h!qn6yH~cft3b99zoUgokXtL_KvGBj+Fx+$0V*GNA
zIJBv$d;UKk>J*kq*3FJt7qp+gb74*KvG=1E6e^}SUcPboV8zcdIjQ+&MaPd+eATYQ
z_p5(v*u3xfnnxZ(@9m18<vGdnu6K23{%?R&qTWb39seLde1zKOP-(8@q&yKG>PsSL
z8z22H#D2HACU;HN;<F$KZjEkQ_15n9L}7rx99Z(7l%>p?DX8502t)ip0*~NdZJCyH
zb46hw6D(3Lk_D>gdIZ7YPn4-iC2=kawSPaU3|z}#$kT`5)-&X=@>=ohQRhQM!HbRI
zsq`GnCy8jVf0sM8YP2YLGl5>Imj!H+)PBnOdE<^Wi-Hdm#3|$AxF93h`_!^qc8w@_
zGC`CwN*=)#dX~9t+LM|Fm&y!L@L&R?(kKhu6g`4fp~~B1ZY7)+1z#rE1j<b?Lr{q>
zJ(7H*E^>cCPPr&(nIut}D6fJ!>x5^Ki*NQfKlfGi4-ljZaIXN*rw`_f#vgyz^_g_z
zl&3-?OHP626!<@VFv)H2vvb^z?fF;~bZnY@Wxl+QzRI3;2+Pwma(DI@1Q-3e9kB-^
z4Z))1g<j@fKTF=|O=tiPAY?#;x!2*+U>r<n1dcHGdP)k%a6%Ten0wtUS<!0lJr!vR
zP9bDNo4MClQW7Q+nvJsw+0kz9jW8(<(+Fi_wz)Svr5wy5^eVner~nJhy-t%}!`BGC
zj;|9c0W1M??@>#e0XI|L3b>VW4H#=6jGdV%)qu4Irf{wWV=cAUfvFB6xV;`s^^_YS
zp%G?qdlQ(NsJ$6X&6Hcg*h<H51#2svzYUCS)ZPik&S%-$N%dV|?BaIjU4`n;q~Biu
zr2JS{j6sw>1kHz#!j|d&OL@oospiBWQSw7S3y(5?^d}U60fdI)Pz#T;VK|J?a2!s^
zjAjdux##hDLXj9rXaY{K@Cb~>SV9wVBB2>L!@{F;Ce9=@3uh5}5nr_MSe=ja2`#_{
zgfcPH!eey_E+MoOmlDduJPY%!D6PU(gjVBfLhAw7gN4WL2EYxJHv(>?ybEv_Sa|&I
z2HZ`#6tI-?AxJ(1_MR#%2%>Ztk`Gfp0?9`xpMa1PfUJy((n$z8Nx2e2DmgQ6$noR6
z6L|bw0?j2JKOQGH`>s6k&By+t<byt>5)H(GgnZGLP!I-@x-<lb5DLK%LMAkkDusxI
zA~1r`7#u??*jOA(XgrQ5G!>_kdX|6*gr?(kLUV8qsb_O>E};}mA(W2kq`GBb2BDYn
zWkSnwIjL^Bm`i9St|U~1MWni|16&7$w*zjc)odr=PR{J&2KT4eH9J3+;PSCLuW9zN
zD4hq*d0tIecFQv7WzA(5L35GIY|}kM?(QC~>?_|<Dlat~T$BtdgPM&YN};MyHLn8E
zs%SOObgRm$=Gkmh+0;B^?JB#P#i1x=tFqNR^9oc2YMxsqpeg}1&z&>S?+kdcBFDy;
zcjN3V_--0N(*U!Wj4eu*Ir|GPy9SzTeA8?vvdr0KaM@MRT;(#h7+KcSYj9yVXu4q@
zo0jcY7CJi)F1rAl3w(7O4j1K5JKAvBPn4=aTLppa`%X4j!5Gffpsl9%8qn2%#O*Z@
zQ3DpvwV<n|_BsfxgAi`7gQz-~z_}g*>#4m_p6pqUYNYxm2yCKu2Lw8v<tPW$H$z}E
zwYPw83k>G-w?K3Y#B#n4zSpU}6?Cn1KdlhaO846ax;AQW2j6!3{dO4D4&(Xw9pKwR
z?VX_Oq~Gs^h)z0w7wEdUebWz<{G*z;`MNp}YIqfd^MR{iN)^oHTn)pksl5flTENWh
z*CFgW<#w3X4vE~}0bw27e)L2?<saw1JRr*1Y49?U>@*lmBs&eECKl786m5t$u|yCh
ztHEkw*&|9egU!UERg~-oyNP6{A=^Z<(@<cd*=Z;N6H5qDE=d#1Ac@~x6`-wvV0JUH
zr^5Ydd(DcOD4l_zGhk%<tt?elyZD`aOCd^Y0oOtVpK%-DHi%#|o(WMU?fT^4*Rrq$
zX0%X`wL@4tjASF$cW-a#{>#n{AWEXEsAVf;X1SX7H1%vnY(hD8on1Un-J77f$us<R
zxblIk+Gne?OlrOK<RG*LJvj(%sGb~zHd;>(LTlBNgV5Ua<RG+mJvj(%ww@e>wm{ET
zDM}@vEdf1qi_@RFGd!@ru^VbbBkqlK7o|Zsh}^=V7)oy1XpAPeX$-~?nv9dl{ka$y
zlbf#&66$Dw+z6&d+7COx<am}799;jP@_(mNimweDC`ww?l4ewgIzkdkr2P!WU_xOS
zMjFnMIFit498FrxaX5}pJjRozavDw}l#I!wjhu(`2&G~wX&hg|mk2GxWu#SHfh!0V
zVj*c8ivf#io45sV3vB~;0PdiC0Fn>TwyzwL%PAj)kfXfK6QyGia*XpYRr~8k#9TQo
zXF$)-eU=0A@6uQPY&LV3r{*R)KONj_Qb@ymDz$1`WSUFR;6Q_ZkYMPw34#1o&tC`e
I*C77-e^n^3i~s-t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_bias_e_permute_xdl_fp16.cpp.D04ED4238574C098.idx b/.cache/clangd/index/grouped_gemm_bias_e_permute_xdl_fp16.cpp.D04ED4238574C098.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94f4af8d8838345c12a9426f1e9e42018cfe4271
GIT binary patch
literal 12730
zcmai)2UrwG*!XYWJ&=RF1JR=%9Z{sIC?`m>fQS_=v7oUMjlHckDuM=VfE}X*k*Fvl
zV!;N6Ac7H5R1hH+Itn5f`xo(>J2N-q^Z!3ze3H5S@y^c9PI=pG=#ap`%$_3Az?nl<
z&Y!hxnph+fb;SQKUAAP-bquchBGJl&`Qs*xwDUA*@jE-x?wBOQIr93z5f0OJI`+7A
zq^`X5@wl~bDq8&CJzRJxR@}=j+{kF$mg(R9RCoE-%c!_0gI{L-*{lEln*3FXKI?p^
zN=##2pU?C)S$F7e)w%VV1$}=C9BI4;?&p_OS=Q99jBN|cpSRA~^zXKrA6|m;>4L7y
z7BBVrEi%9PYAhdot;lFu+ww)q?3_!<kxsS$=zQp9*5`L;m$c32T8H&;dpKu~)$bwU
z`zuyHAH87t9G^+ulbjZ5kGY(dVwl=HY%EM!-2S%fo-nU^kABjJgC&=K2(I!Ap4$Vh
zm>N|SJ$>jX-<`!1j34&7IJ{fRO}jP|h0!bDvm?%~3^Fr~?y#ro$@t7c!3X?D$+qn)
z>pidi`Td5`h88FFUoH-kY)r6Tr8gnc;aA0lppDk;{k5kUAk&j$P6qlv-*9$AhA4gi
z40PwC(K)$8#O~rB%`d!o*6}AkluK;Q9qT>5p`Xr?yxHgeIu@gUxsT}@&$SIR7j*w)
zTOafJP9x2-`wbex4RZhaZF$cx+CSVp;CVBuHhPJD!mX3zca+(<WEWiDHEg4yef`8<
zUD}>m?y%7|_vmtNZDjnk;N*d|lFGe3rGx5kOdnuA<z`g*%j~YEyH3pL+xo2R_z>M?
z!O7<7+<-d%hGX=N*GY5FB!6BP`RNa@G(S&IhuocgPxLW<xo%&UmHd1!rx{lKs;R}-
z8_lx%6`kb^Gym+hDCM>D&$8p*US^dErX@44-l}(x5nUcr-c$NFD&LYe@ZjG4DIV$j
z-BY~(85>bmK44tg#0rCqIg__PEUt0ro6v8ZQAoh79{X26E*<*d9Cu^uY{}T`j#n3Y
z2kUsOyRd(FtDjM7woBXEgS|4x>^*Jp;va{hH@++1*SL4*#La(4MwLhHKJ5SbI6pVs
z>F48dE#9ZDZmn6r<Y4u({oXIF-pf|6OqgC*JR`q7sBp&X%fIlJrmMxzs~m41I<+Wz
zj%?!n-)1I)OYU>s`~NK5^W{Lr8?9}=x&dc1*Was|m5@^In|d~E*V8qFvm=)|4xd=K
z^Y58GuDYOoJ@uMh$F1!VV(>q8mxpN|d{sW#T$+_uHDXEcL~on)a^0hS{d2a~{AMle
zn0{)VXL923z$`7<hK;BEY922L?a*iQ`PxO3i?`|XlRms&;yd!jIwy}uE!71pwz#j3
z2yZ%KY+$=+>+tO>+KbZWU%e3CexduVz7MDW5}WX)Q{|*>8Jj~(m9JpoR4J;PVS8a?
zd}C*yM{d>QZRSpSe&)fxm#a(D|DGA~zGQO5oSR!~wHH0B>9Hd%=CReOl9P5`ekuDt
z?<+Zb_SdRINBs89^NHQKXUKw1frlM^&vyw3UUT_%?+Ly$+*0S1hyCs@O7@;UWV6fC
zy+^;iPrG_=`@6~UujTEl^VcVL%dtw!bMF85@R5IK1_wMYon!HH!IV=AADXp`P8yDW
z+>-UEe&m5cH64cCh%sIsdTC+!8_s*~nb=!Fm3fbYtx|qIR5;1`mCx8C9V|}jhThB=
zvI?yYXj}EESf~h{d$(-<lSkvG{ggEPgkQ1YlMW7hn;*|NyBc<=_D<S^@9S)Ddlo2@
z$6CU3tLhaMn?K#oEPYp5Ww73Hy5gzx@sCooMDL{3oxWEC<VQO`o%BiCy&}#2gUOKI
zyBt1ta2vDs`=zVqPdCvLiNt0vJbOm;8X<#4^F+cF%PCIk0&6?<Kcm5Q`}9mNWg0;U
z77x`KYM?HN*z|EwS-*>-Ry~*?5(&Mudzq;VBDRTnH>JBtXn;0MohuS%>diD%7es99
zz5T<Prif=UScV&)YCY9WT@Z1%i(`7#)r^hPh7f%B5WOLt)CCbcEnohs{>@NxJs5+Z
z-Ceu8p}HVqxA&{A<h%<?(1Vd!+)cZifw~~#9(8F>Kb-duFoYSn@g!uBgcSe&kSg9M
z?j-(vc~`0)gyI&=waxX_1rY~qf8F+D`ak3KAPkF*wT&g}f`|v}M#R;HT#wa*aae4j
zZDFJ?h<MZ_&uLaU(X~?1Y%I33u(MYeL_Ds~`_{kmlx|Yd6fAZ&a<)_#L_BrX@3P;;
zV?1TB1drVJmft(63nHE!w&vjx<L$M2FbdzT&{lL*7eqXF@UdUVcNwW=0Mqch|AHjH
z;CKJ}L#ixr4Ji$?Jao?hzQ<M0AcHf=;@=-qWf3PY($SuBL=Q&ehd63G8mJ2*UdeqK
z+dJ83iBvQki(QRe73zYBqj%(IJiD>}PZ=!5Z!pPblB2rNi0?R_59+ojmlthD`1Twm
z&Oti={*XHH{R2_hgZAlqFalRG(>Bvr7eu@>^w{Zljaf7FU@{h)Xq!mX1rhId3u<+D
zoaG@EjmKg)BR7S*AmSrgH8u@@zd5D{6S3Gz+sZ&)5b^2bt~)F5M7c^uKVY$~g{{51
zAmR-1+xbO(3Sa5L94vOwcIc=sh&VIIF!WjNz!Q4#9Tr<^Tk5L|B0lfuUZNFn)Y%Yb
z;=w<G3{D{Pe}71o?1<1iV)v{NsVE$`U~ge>uP%r<2YG*Yvg>4jJqVdE62fKSM(Toy
z^Jf*D3TxOIErZ2)@F&_#bW|5ad@n*=aeZ$bX8;TE-8+zM2kQ9m52;d__GZY!z*Xg6
ztIU?o?kK1VB7QRD_7%_KybF3T33tz0+ge{;5OJ+vOGC!<%`*&O7Jk7rG&l`S`1glY
zc`<Rl;gjKM8?@m^+(KZ7KznsT#P37mFHAo?$wUUr@f(DhhuNtMBL2|yB;K`ts;wR@
z#COZJ<sH-o5w~BO*l;{~gP|cz$Ist`%=Vx@|Nf9FOJ;>Hl{kn*=+&?h{<(vmgIHK)
zv&xpdw^)cr#_`C63>7Ty{!A-cyj9E#m%;Edm`9TZEL;J@D=c3H!>cS`1Ji3@OX_ny
z^-_N=nvQLvkOv-l(2ods;W~I+XE`5y@?ij#7eN042&7yHJqn=@%aaOWD&=Z$sfJ#Z
zYr&<K<$7?bXSor2H-aave~IgRiSwa+h4a3`^{1T8`Db$hl&^7vu5m*s=W;>0+;GCg
z5CJcYk&Q7YhDTRzjx?qq<Nf&c`g7C%V>YyluQfGy&(P&~;SRXm0blYEUbqV`cUdk0
zmlDc)C7Cxr+`qg7@j@ixBas7{Z@%xd`}KR@JVIij04xf?inJ*G!Av>lq~dohi05SS
zoB<ISjJReqzd6zl_u&z^JfhFwg$i(~V7U@pDk&RWvYhX}C+Lnu)p^r9;7nRnb>8$Y
z%Ozl1!tw(!eZX=ln3l3!2Bu{!KLXQ7ELVVO1<RFST1lBq3DQ;j0Ax;#|FF-f8@R#`
z_qPHWRUix6-|ZmV4syynK(>SBSdhiC90&3^P|*7OK)#RV{UG1ZastQ`SUv#q1C&i}
z_gsDRc>f1@GLtx)B+iL6OuQD*{l;72g%8l_19T<T6hj^sH8$w}h-WYlyz?NK2vl3q
z`#Q_{FfboNh}?2c;k|U-aXEM?mm=3v#O#rF)V&>P_K=k<!ZSvvs0#OR6W3`IXH8ll
zj+0nd?JyB#$YU$e^#pPy27(vTk##!iO7y9U9n(<{!Zw$e%x>P`y%{ggO`L2K*O9bO
zJTKhCak6n1WM*~M{uh~8+pXKDwmRLui0{thT=O_LayO}fd#CcsXzx`1ArbOIc;9e;
zVu*O*7}7t6jEURjnD=C`$A@Kwcx*GlD---k+q`fd6z9Q)@<mWwq)fa%PWKycj~A55
zNSThv)bheRFnb3Mq#37w_L_%`JhKUp^cK!y3ms|VCgOKd`G`dQ<Mv{Zau?Iy8!&qV
z_M|;tNaX@kxlx44d*LPZ)t97$lEFD<a2hq-+vmD&`KpPXwRqtYXK;x#A}#R36|VCY
z&Wv(4XOhiXQohD1u5lVe!3(*ZbuOoQL(lA&S%Ym1F5ovj3|@ypqiDSh{y20#VMRM0
z>LR35gk-eE+eqg&%f(2in6g)A?@``i+k<e6N5ShTUGH9&ZGX<yp4KK23;1ytkc@bX
zD&=Tc0G-JbRBT&7M@=jgf?**mF9eT5=u7LzaWd?8u<Q`Wbz`|-9JM)OA&wg#$4w-1
z;;~C~zwy+0AqI?NK(o7e;S}hf0uyqZYCkKkLl??-!S5~vQ7#3)QkE;guY%=T=u}G;
z4<0Wr@F+Ku-0yvV%>CEJ+d??iMjDXE4I~1waGe`)o#lM4Uq0te<bE$pThfmw3V7;X
zfW-^8hfQG7#BwuOG*j*`skx<)+*R8#P8vy<`+$YVT`c`B{g(k9khI>mZC~POJoPci
zA_ny&c5t91^IZH7y(QQH;{|aPhLWy|g<}wWjPfAQMf;w<@bSPcd;!-l)VvM4+%oJU
z-#b!UETn;18r|7qJS%3WSWXACbe2zp*=f*dA+d0fvpmT4q&-OFEE6e{_)J&#o4AY@
z%uLM`BzE8hb5nB#SvS03VQQfuksB{qnp!G|ug(kBrq&AL*kb)KRym$(8)caCdxjP&
zmoT&uDK}{}N#NB@E%Tbgf9=2v_mJcs(kK1rg};#WFVu<hL695-MxF@LL}28(AkSqz
z$)_NF%F6d~9rtm@q&@OlSVgI#QL5>Vae>FUaim&s!nN4q$Yv>?I4cV)TM`EiSMFSB
z<gzUe|Jk9h!$1<?48L(Vq|~Lg5Zj0^Ao~KEtxDostXEaREq$7@zcPr9&3SM;4<6*%
zyl{jIJVJfK(UO4d`LhgE%lHhEoI(0T(To{k*tId<TdSkW!b(cnK`cX(GL|1A$wQXQ
zk))jEMkH-y2mLvcK4-ZJNt;-1M$%@M-y!Kcmfs`kdzL>U=|`47A?YWUqd^kQPI3%L
zVp!e+k}WK61<6*HcY|a%<uQLv2`)P@%<n%OWRhOMYeAJ;U}on6Qs!vZ>bU+l-bOx|
z=Z?oef|EpW8rN*xv^^8j4p(|%!DZCxGBP3)Gdat1UYLWu%Ih7{Z^&Tk4D3K2JJ0|U
zF7Y_s(Ek@SoMa*<*F=p!AHHy>+8r>+WL4=ous8>fbcJMrMHc1nQ&TUT4E6j0-*^LA
z-9UD<Vga%$V7U-k6;dV{6zl<wm5pV>4OKFQbbHngtu7l3mkz+M88|F(JhAEFe~cKX
z9F(cjZ7JZDLQO5nyWn9}C0*#S9_Ip&a~l6}&XT3V$PdT-wRvGDIPPR)tSV4efj@nT
z8c^1-TnEZJmK#9X!15Ome*r1EpA0Yd*uSQC$Q#TH{jI{fO`Q85%VgfJ>0Jj^1+Tz^
zwA^O73o!z`u)=1AhRD?8hkVTwk%ugtIw>zpG{y%HlTt_NK{h?f8)40$N*d8YQtm~{
zB+c66g|}e#mii9Z_vHdlP>*ZzO5Zz*Kkt3k5etpTrV;fZFD(|HBb(<eHzAuQmYb1H
zGt2Lg%{!LgBb)aue?&GPS^k7<K2ctL{?YJvmXU}4V=R}RxOa7D^wgj5SSCWxL^d#g
z3O%1v7KRQgavJY33J;sJpK}P=swAI=9bZ*4jTyE=q%3BJO<9hVPc*zTuc~&nxa2l&
z?IDssL>m8cb<3$koh2o;IAc)-rd8B;;Ds76tzo$iOzT)~0MiD_>t;Ue_n;waiF)sj
z$%H9nB2=!t|9R>*^1?;%zeqXar;Aa!&#Eq|UUancMDn5=PYs(t_E1Ry_K;h^sRg`f
zd#}Lh70a#Q(h3@Hi5J?ysg0FKa04T_P;&pz)3g>0tiFc>d+e)tVRuC%Xbn(!-q~h`
z1;z%(M#OO-DM~zJp|VgT8YHQ*l|)65Ht_2InxrHXvo)BDdKvqrFE&xOa$5yS81jOh
z+)hEPBQMy??G+@M$K#c3(ug#<1s<<tljkfqA(JMSn~_N~%kPlMJC@%gllLrtL?$0u
z{)9|Ev3v&jGpq-Z2{xJ3rt`u%;Lox0bKrW8u3TQo0zQkCXMt-Lm2XSg`*`8n=ao`1
zw%P+q(P%PwVxbIKlp&3Su43CVq;bf_c<nkrq+3oblq2JEx|YR4G}uMcOobRb{B|)c
zZvne4EN=z7t>8|dzZ)!fgDcBDc2m777Iwpc-Jn@XVqp*H?4bd~_QuS`qx(0-{-1i1
zxSTjRQRS-1I^l8OjicG%Bn|Y_XxpmfgCdV^H&s^1?+zP$JOIB3EI$IjM=V!Er)t)s
z{>BCV#%Y3+`1Z5u#V40e!hJshUMHwA<Anp9?*ZzJk$gH<K&oUq?Z3iQp&*tItNK~W
z*+kWo96Pp$ssuZ|McIgyEgBuQr>gt!{y8Zx@cuiCOwJ;EVp8_JpE`M?zM>SXmRMvG
zi|ncY5R1HHQ4r-VoYNLgNhWab&7foVH?MQXEvBQ+>Bx#?M3PSY@W+GP<t13q-Ltzl
z$tE10H&h%Pvbh>ZXjzEQVk6p{V0e>_SmQZ|cp8x=zmhxFzOeg*wdf0Qe?dc8UT6aM
zCYGDQy_w}C&OeE!M~R<}he~zWzzo$g<vKE+DFZ7l2K@2qm;dLrQliHv)Gg0Z;DLGo
zE)Qt5h5ci2DP_3~T*@dPZw)#;B{H=Fw-|%uG02?Q__UQ*v_{QO4#WbvOm0LH>Zi1V
z9<~L|@x-ge)yFl27$shC^KlCy(|yX_YgGQBsrzuHnO-vkNHUAZDSods6hP+Uv}66)
zNo^s{c+lfGvpDKqs<>wy$866yE;x=GMrN1jQf=LD4qfot>#Ou5W2H(g4a}rUUS*9O
z&VvhSL}iT}F0!1=1t!zwP9hvF-EWR=c%hrTn>EQA@q(k=(VBQPyl|2WJW2gL{7BB@
zJWYvbJoR0XU@~$xUM$t%Ud`s}Yr(yi<$7?hXSor)8bOmQKND8oD|wr~As*lYkQRU@
zugBw&lf-cv2ap7>I=bH+FyYY7&dQlAIh+y$<1MTPN(cRPRs)rQUkR&G%D}IT<x23Y
zWVs$X)l-c_vXyvge?7}04kF&1T(@*O&lfy0_pd2w4#L6YS=8w)jVdk%?pc}c-%yTM
z$_-?H18Jg6oM1xs1uPdL`$EcBw(BQuFo?@hXWM))aZF0*3OC>iT~fS|%>`yt?~)g;
zalzNPF?6-$awBrN@q~%V#!js2@QvB7!O9T21TTWyMK-)m;Q~|W5+lh`tWi}*babE0
zHkoZrHX@Htnw;X1J5hDHWxu(LigV`thr5|KsAVSqHg}Vj80Q&-FL6ww<_b6B3f2E&
z{N|zA9Ak^GapSIW-;p;J<2>WUTy7HKyyK%5NCS1E@Q{A1T2+3eQx*HRHQ-dkaveC;
zvD^So4V1~DBwo}ZvJl4IQk_o{ck55{mDQ4!GA(1V(2jiCkwz!r`~&j&!tw_2*#L|@
z0(>G^-Uz-M>6Q`;k>DH2auoPPvGq5B&nA|Wpi2_%y%=X*yBucu2y{8Zatd@wVfiR@
zIm+@e=yHtZROph*@(Jj2g5}Sk^BMHWw1|bzApHzX>hv=heg;i?StU35OtVEQw)o7t
zJPDj*0;h4d#n}2g9-v%w`mOt}(~q}dd+-mK{zLtdqJ~qs&BgbJsVu^1*9nY880|KJ
zScGD`HPIV=N9$wzl!ZEHAroTZdEq9Q-2^7f9nZPObDC4hlAzdITQhH&|Ho=CssCf)
zWxodvE@A=CNfKxr0oAm49A+8M$q|<EoTRXf=j15McutP7jOQekWjrS*Sk8l9d8}u7
z9eQ17IUhXpX^f)Ekoy!+d#uWkI~P)`q~h-iL8EuX*!z>;VW;&j$nUaT0`d};AAtM;
z%cUSMWw{LGWh_4e`6HGqKwiOeCCDpTt_H_y=uVemEjZS)Tn~=*EH^@rMqrY*FTmsl
zn9}=SfZGdTg83#eX=3F~;MT;-o57@+l{bT1Gb?WahZcIKCl+3T!z-3sp<64>Cy8+k
z>Cnc?@t#QFG^g;YZQ*r*<s?p##A(hv@%rPulh_lA49+iu(`0nS!Wqu*49odk-+WGU
zA}AJ$IL{)w?TFUL;pNxwFo==3ZzR=CDw=PL_u~=pIzrv{GRrGkkDRQFu%L*uDx$7x
zh0(S4)qd8i)vl`N3t&zqo4~V)<!10~W;uxqOyV@*a^<GFhdDJ`nfM8Jkj@>%lgT3{
z3mbaX#~`Guc+YbtS)AsR2$DnZp!>~7A-u2^jJL9G$Z60&&AK7^pvb3+2S>}$sgb%F
zJU$QTbdY+HIM2d$N~HWW;A%?B=aLNE?IiF@q9ax{W!ba71LF1mL&p%e6;H+2FII>f
zQ`bZ5O$UA4b(MqLF%6lMnyU1k<xUzu*GIoqggnxojVq>tS1RoRFC=ij3DhxY2)(#2
z=k}-0xU2D;Q9P%~YLZ7{k3scG40SugO~XBiJHZRdT(@KzS~TrSe|N~~&aa(#VW{y?
z&0#Mu3^N|a(D07KjY*D!7beOkb|w~?7beLjbtdT$UI^0(laUW_@L(dlB;-gAK=J7#
zvOCOjGO|cUc2pjXjTBWNs<SzR7?!ty!4{Ucg27gn(?BN;m_y7{pmU1lbkIp>`84R9
zX89n;ALI<lnpLe5K9TavrB^PTT-mh>yEmVa*=L$CR%N637WQj}SHQnwxfP_XY)9LG
zZ=>>exj}xv#+Vf2IsFJGA923o-!Fv7Na9s7S~i;TdsJUF5x?g{aK>ZDsGL=}TlbLV
zJ;WRXKS4%MXexvko+6{CELS6=YL@GeaUIfZMiSWJ2toCk6kSoutDwvV>dGo_gR+?A
z$Dn+|@^Ma?$}#mXamvdq-{h2kvV1|T%o3COpJHC`XmPC=iL<#cz#JcTlF<?4INBT^
zc2aHzb9~rIxdj|?o{;h@aKw2+%B|pp^MsV!z!B#ODJOAO`126L<kK*`%fEgXMz)-2
z$*g(P@W*zdL1}u`onNj04iVivbd6O`#O?lz5lOd`Ll6;}J#vm9B8`Qc%sLPr@cw5+
z{2&rCi-^)(x@5-EY}IV6CsQI4t)`bN=;c~^`4hd2!k<l&e=Es_h!dvQ?rcwr>rpL>
zL{VrXia;w?uUxfe?N960Z$KjapXT32=6a3hucB4-avi;lWGY86*E8nNoQJ<}6$Q^4
z#)-AGJLpLCc&UNDEU4q40S+B?gA9#3bv7~W>)FrS$Jfu_Rqo_v=h4mT|MyoTbA^Sq
mjjer`u8!TE-FozN@1wNrZRXNzpxh{M$R-2<WBU(AL;n{}G{>(1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_fastgelu.hpp.81337D2196A77665.idx b/.cache/clangd/index/grouped_gemm_fastgelu.hpp.81337D2196A77665.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b2bfe44041f69c9fc50d575128790c69a7e7f5a
GIT binary patch
literal 2640
zcmYk72~ZPf6vvl90)!=mK-eV4O^_pi5y%HY6Nqvs5R8Ziks?(<pdd&CAsj)(18+sE
z998LvXDy?E6|G{i1;@5ZtBACySS>QON)d6CQL&?4V0QP*Om_CS|CfC4|K7{XVr^t(
zG#f!8(zL~E(hE|o5Cm}~-o}FbEIS5*^buNd-EY->m52PfTS}aL?>AIOw`u#@n&XZy
zC~}H#=!vf`?X+`h9{J{i>%90!ZJdhZ+Ae%w&<?9#{=<KIw5wcN{C=>PySz3vDeQ68
z_CtBz_xA9%6?$ITyu-skprL2)&CqRc&plprH)*({P<f#CV1)C!Wu0DioTlE|uP#Mi
z`udNX#=~BnZ*ouev<gS^L-4Jif7w!b!)oY7QTy2f`%aH;r2y-@%o@<OUDn;2YF)Uc
zrcE%E6dTMLl26^apRvli+b5!m6)|-QZ=fZt<1P33lj&#TpB*)H+ZXx|t4E*IsMSs9
zdp-O5bBFZhoZ_yIwRaA+i{5m$)_H$={yR>_U+U|uzy;dn_Sf%&=qLY5?GDd;RdW4S
z)&6TEW~0kS=Y2NXyINIRUMgo%-OYaEkkS1j8-Mi5*Ja={>)BUaHLRj_z19asK0j^{
zCfPS09SNNh+0U-en;$Z4D;Uj`m|ONN^cd@5iEG$mtLUhjU&lbvu6HZ(yRCl~mAB7I
zu83>2s(;X3m8f?;Q<H2{UDaFjNtM@;>Jtv)Xp2Zjl+N>eorTBdIS>Cy{J3fE`v+ck
zJXV*jH;YSdG}LJXs|SBr_TqTij@@U1w_Fu|%nigtf^Rp3wfQ~cZ13hA5T|Q_Atui*
z*I(><tSaJ!vX8%VkLQQbb+PLc$_6SIc73SW7;`r8t!0$DzY{T*tV!)BX2UY6DQ;VA
zYZ{7hcmxtJY!-7?G-AuNKtzrfn1$mh(dpku-Q4XMY$pUuU?=b<E{M%!w6<2%`E=A1
z;W1D`PUEJDCoZJitdH5$=n@{qV%YI)F`c!}W8y-}T`a-goS7$kQN)2f-I3`?oNi+^
zdHT@F0>_}<J|l{7$pD;*69LwilL5objcz!8e=UmGlL6D2(}@5UV=@5APtP#kW6@3-
zwi6C_cf7AZ-5)GPu`+Ji64C}sa<MYv$V@30#dv_Hr1Ha@fD=S~z=uczBuXk#Y#Nvb
z(F`y{N#%(70Y8XBK!}p6Kb9@e_9N3kvFxbq7@BhAIgn;h7y@Z6Pmu?aNoQI}YlVtJ
zNGny8Qp6ssf4FJBatVrJA|TSxMyG%&5Xpc{LtFC#UJwO>Kn-n836v0pfiMkiEgXbH
zv{b(|Wpa2B1jXX@@en1;k^wn!D3&5ip@@9SU{A~+Ianx`;hrHU9XmE{7nvFduM)0Y
zfD6<j3}6rmfe<1$;086!9k@f}2|OY42mVm|6hHw{00@9c12j-;XM@=gg@RCu$iyw!
z6F(c$)zVm=?CBRFOeQggwmb*AI}%%o1Kkg~t=xg`fXY_oK-WCOKEsLp_KIWW<^`v&
z{p^5Znfy$gd`yrsb2d#`{47Yz7G_VQwd}xb4NW=197r?Z1{tjxREBVx^6)%JGvOwf
zR3<305HEx@Gj4`SWriY4@ls0LIx<hJy0uGWNtp4;^rsKbjLD3nDa$8|(oQwnGe2H@
zT*o4cE9EJm;z}h7sJK$O0xGUlrJ#$eo2!ebM>v6>;7N`r*+HVA<J*qhGUOq|@)=)n
z<YN@e<>pe$=Wl=JYikE==XlroJPG+36qE2J5-2ggTmmJ=S4rr^GK3j!r0c;HY3q!@
z=K(e-CIwPxjO9QMjj9T$sBuIwHBduj&>5g@Vp5o>{efa;g&8KDnTjM+CCrX5Yuc=l
zYoyGQ4v3)fg-r()bBn2<*Uc}#UB2PAH}QLR2F}o45&!{2uD})AOCG=jB46MO?Ii#J
zM8O~!+Dmi59Ef7|vCv*xtX~XKqAZcxODMKNwt}K}ElQ`C@(wdGTVfzq(=GG{-fFTg
zC>8{Q)MNorECNKR=@#ns`W5tSrOHzM$k`gZP}}Ys+<DFdo%|Ceqw>>pCXzGeuQEmv
k4;fgUE|n8itBhd^gw8H}L_kI&h%<f5r*B>8TOob>KNDts9RL6T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_fixed_nk.hpp.92D3931E2C242798.idx b/.cache/clangd/index/grouped_gemm_fixed_nk.hpp.92D3931E2C242798.idx
new file mode 100755
index 0000000000000000000000000000000000000000..706f325710612f7d3b90ce864e30b40754cc75f4
GIT binary patch
literal 4728
zcmY+H30PBC7RTR-vIL(JkqBW8@K}vOYgld*L@^LWQ4tY~xUkj@u?7{jwJxpJsnms0
ztKve%#wu<j2rAUtve+7mPJzNWV6C-QrKqj6GumnAMZCEOJ|OS+o&P)c+<VWt_hj0H
zxVWWG95*g&LP2hJetLh7;|7a=#{6XqUYBs3qi8EAZrPoi_ss}d!>@_INMz%K=M0#%
zD>*r}JvBi&=Gdd*5f{Q|cr9LRn)S|qx+5oi^UsFDfZ(z5MGxD_Utj-utD^OxOZ26g
z3motGSM9qtc1hsK=R3H;se_iJf0_2-_=EP317xw8zu9;0nX>WGA6L#s|8o88Bj1vw
zR}%F09~`)3xYJ#cdF%I;^$x#B9B*!p0GB_Xgnn`R`r5jTqYJt=JnN}nXsfaNDsNWi
z@Vo!+Xj+lc?fLxj7v{XbHtM#mJ3skgAer@cch9wRJqON7>-QzMdt}TG>~KxbO)b+G
zl&IVLXYMa8*wNYMnqD}~qYCQXGV1(qON5rK4X$g5d%Z!CH~O2u2mWh(ndbbu?)XQI
z!&~zA{cX}#cjKDw`oGN1Ws-*{X1Wi1{5&}$bKlPe3lqldQD-MM`Nhu3N)+@%1N?sZ
z>D{b@M`OQVXa2Ec(dW<VMmda^6$J&9rONI+d#-H1boXA`_5-f>j89KwJ{^{H=7Dop
z(7W&NX}zBsbT8wNt}Uj!tE&6WPa5bs*1vGf#;&yGr##j)eY?DCtKaF5t~~zioldxQ
zso$r+PqBG6by8cCYs`Df)*sdf7yUfrO8EaiUpoKm^si$+v-de<SodaZ)bTba>7U%r
z>mP3VB*Wo?&MEE2YxV-C828@Hu1#C)t`0Q!w7Pt>Y0Rac=KDBJ50ZH9%n80WKy}$|
zz^}X0vTL?C3cKzfdbs;kN6)vrk^?SB$NU(pNGlGyxzE(JVL)N&wxOqP-YDMekZ-y<
zZl|%($v<M)r(K2Scg-{VC;oQCxOUz5+ct!`d>p&?h|`wv6}Q&8FUmSwQSpQM+}eJ7
zJ@XfG##Ol)Zve-sQ#%*uB(F~3IX5pc4aFZpYoj%A(gFRr5Kj<b(u1^Na|^z`&2t`9
zFvDhsDB%0|3T7W1y>{-2p*L+ML%aqHo4{@p_`m$%?0a3lq1KL$85HF=K;q=h2}R&g
z1WqqMsHAv*a`Z~uz4v%dL5ES>sKsF%?Sfv!aag({IyWxsI?uUMfzn1fo|8IAdgCzp
zKS(=tv1vEYxl_Ren+c)^nSJmJ51-84kZSkz{yd(OQ$e;(_5@Bk(5BbJ8N=Wy!($!{
z;<=$zu*_zeD3C(0puXqelVG0@PVk%y6%4T%BE~StHt@v|PhR!9&{%!rf|JBqoaiBt
z9RiP+A9SMJvg{nA!yt~c^{)D0TuFp7h!>oRv(D;UkP|r~7vh4*hxq8MD<KF(Kom+s
zb=GARMvxJRULmg_(h;4`y759ZiAEGp;&s-g6eg01h!RKwq7;&%vu1~oN>UN!kQ|+L
zy@l7w>xdSR1&EBqsIxA-u!5{WWCE25Lg}LNf(Z;Jh-Z8ZRL8Kr3{+*9PlD<s=1Ndi
z_GLpQ_E!a}Dr~O?RW;@sP}N{=098X@HZ)*=S3z|Z+s&XdV{QahBjz?xwe@8~8}`=@
zs&;Je096O(PEd6+c6fHUbK}IYBrjg@ARc;_V4lPikr(ko6ikBkEWtuZ2%=$R7@`Og
zp=aq9Ng@#uLJ;YRUeD4kmc$~8BXNkPk!gCCd`Tn;(R4B$(OfcD&ysH*nTIHyq$65F
zmgre3=8-%^OUY71YseZsOT{<In~2^bZz0+S+HDZmd$@5NFKh?xcFa3Ky94unQ11sq
zk3n8226Zv!1E4;@c<|{hZp%+S_+k(*xDr=X{R*N$<WAfX`4fM176gz0L_s77(I_$s
zodw|}9MKpu2Au`t$T&ok$z)Xhi6jxxEHVpK{~R*MN_68T(iay?;pfi}m)+JXdEqk1
zF2gih%ychBndN$6vL=u<G1>IVPJut|(Ah}ELsRPP$a308>f^}rHdGqw$Z}IB)j85I
zcp+XI@5pj2MVjJB$L56`X^tby6{FPX$nv2C2A04O8ZW)|z~SXxM<bqFP@FffruvSA
z7g`{s1xC_0c%c<SS{Y|2jICLJ_TvmuSPSvBFqH{gU}6g-QqG%rKIP2wwDqDeD=;gJ
zWnD7_HiMOEc1avwUad4lUMK<aS4I78oaYvIV&te5{dr*_Ss2CUn@e&Ltt2a>*z8x6
z)mGZr^WWwJ6`kiD#7$5-N1{zo`b45lP=-dLO;GA0(IzP4Bhe-(QzFqOD03oNP)4OO
zl7+Dud^f|ZYzc}$UIfGG)ZaPs_UeRVbqxY9l!Ckzyy%oI999aW86O4tQEWGX!UVp|
zZUT)7q8J|o#W8Fz14UV1)|6rWNl=``_HvMy<MGNNtQ^Nv0rCoLuLMOUe!mhlmDqn3
zD5|i%8Wh!iSyPSmHK3@$_F9nF;_+%BtQN;p2l6^>ZvaIDe!l@U4cPxxP+Y}!Gbqe`
zS!2ffMo=_jdke^0@OUi{-U88V{aZoaitTNnXv6Qffu;@nZwEy?ws(M{qc3YZu)Y%%
zoy@-ZajXYv+UjO!Nt}>U7{LN5h4@ljt4AT^D7KeFNIB*T2&uqa3n8_b>ma0#@zzb-
z+?yYVhlsT~g-p@15=kRzT2>;9$RaH(kpfbnr6sceUbyFEuKP>NVZmp)u}WzMz8S1c
z>ry;!Ep0YEym;Xv$S$(umKL)pvozaGb{b@-A&IUE4O5g^T5cx$3S?g~Z!}_2W@)^c
z>?X)=!gT75PDPYi+HWTN4rJdU*-mMtt^4V{;!HI}qop(Ag|TF;mYy`_fzssdjV1Si
z7s?^N946C2=%FPJVrk*oAjd&=oXza+2%D#NjxO(8?i;zYH+_p2eB?gfbai+kR37Th
z0?^5I-ZTL5zLCd!(*@#%6nTm_+s7PvjyGKuUNFjy-gNSoLt9Y-el*?(>B@`4H9wEC
z=LKKli*EBM5{2&QnPjH*=H-Qal8?ISGBA|kZn7E-)tJp-Fk|ijLkD9Th<M@key^d|
z%hNMY{xNI)jbLY97)pksR?>~QA@U=BsCf(|frv(uk*Hk^BVmX}lhLR_j3F_ICXq>~
zC7ep8B6^LyhMK_HWHzGtWIk&97L&z@mXqbE(OX4UA=(1kEx4^K1Z^SaU7+3tsLk68
z>b;Dgo!@gst-W?m%*Ebs$+S3^WiL**JR6r~8j~y!{+55eKfd&V<CN^t&-y44Y}J0g
XPj9_-**3sm>FeXi1z2sI4}1JS9b4@%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.E01D0ADA6F0061BF.idx b/.cache/clangd/index/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp.E01D0ADA6F0061BF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b084ae928db958ab7a4491d5624c4e63a229e14
GIT binary patch
literal 4814
zcmZu#2|SePAAjG$c+I@yDpMH8xR1$|+vLcdFo<c8TL>>9N4nTb7CE8=DN=TeilW+V
z_swowNvm9&-L1A1-L};KnW<-9{XgsT@txoEet*yN{GR9cd%wRS(A&e~iXwuz#(C$a
z#%D!KA_$@YelxQ&k`91@mpp>x*8jo}OM75FyMNC>kV}tWN2^+e^n>CDdwpXp4gWRx
zt*_E;=bW-;#r(7DG<2sFTco<;u2bjgO{u-Ce=(S7AI*vGWp3)=CJYw;)5;0a*vyT7
zbm!*4<u|!6|I(`K^=Oek_rde&z{jYIDeKl`A@vO`q&zr0V#Iy)-;Xgb$F=oF2-A+K
zdvyZqhRB98ByOOkC(rt`#hJRHH9z0D!gpS$^WJiIc45!)j@sKdUe)L4?dgA85PpTh
zPc~g^)Oxt>5#4>Ds-9hAqVnU*B`S@z^$jaOP3hN!+dl7iVAx!h`S5Ga1)J<-3v|B4
zDbCJ^^gu_G-N#I}zWcMkvFbo+NpQN?@Vg84zl^by;-VUs6r5^LDX6h5h}KXxZF>?m
z`FyD89Y--)PEE7*Kz*I2yJ-hi@11NI$1vfO3pJp<o;LhLgs)$x_IZg*b~2aM<ZK`1
zH>S!<hBzEb{r%uRC0pu`GWu)M%dq~bGHm#~>O8r7z0yN<rJbyX3Xbo^lkLmfQpW@>
zIr;&&<?jr#H5?}VGqw+9>gSnAHB)?^-_Sd-PrrHpJB(|r^JD$xr9OvsWYw?z5py(p
zSx~uZ?&b=&g(-4ft8j^J``fX^qPyIiIhij6=l9=7wsKG9Csx^cdH<HK>XKMru*dF>
zm)sAROSgLL?&J5fKR<iDQK~2AP?y0o$KnU&Rrh!9Z<pJ=AU&ZeM{T^U_f_HVO@_3L
z*2(VXvb~d<Dm}aEx=Y^c)s|;V-tJs@Icrp6Qunl6dss}x>a#JuM_<17boJP~u{EXh
zwaMPZWyWdOV-@Q;YDW(J&}5!|KrLtL#)D36lY@yeMlo~Ny_l<uZ;I?SHJwn~);F@<
zR^xQ(ylN$%E!N!wg(bTGC?2owv#4jxT0SeJi1u$IZg;HgJIfu<IjP+x<Lg&fH90Ka
zwbhS$5j6>ld(K(@?8L^lsz*a{#~((yJ(&#X8r>rCm%uwTAtF8HO&3;sB=oU{>iZ7|
z@7x|ZFnlfe(A<FyhgObl8yEQc*%WjHAMR*U-!1F5OyAImvnnL<?bFd*v&D06H`lao
z^1AYfJ8^m`V~p<JzMc3dj&mtp>(t0~gsWICd-=gs+)10E-MdP&ClaOoPdv_&>wf4{
z+vq*5=QrT<-G;y^8;af0>#w^OCL7&XbZW^re5}1STt&n7C&OZ=ck>5n3f*nqI{ZP?
zTaU*-{&ee6;QD7*Z%*glEx$LNyRCO6e<*k~Jmfn6LcF}H)e06d&Moxx|Kq{0d)h{=
zPhBkEzN+$SZknU=b5oD}lB7UQxw9=#K6jss+_OKqcg*OcT_@VWz8T+lHY%|FQUY(m
zNaq6gM&E?xE=SfbXpSFh&I{{GdPK7r8XjwV&uCrb_VXSM$u|kRm}int+i2eF88$KK
z8(LFmo?p?B9TpKYHM)$YoT7Q+lun0!y`=5<r4MD3k(p0=#%dBRU5u3b4F{Ea@bz*L
z1AZe&?eSGl`CY*Rjej4X*{$1GT3pOMI^gJW@c3-Xk%U8BDQo)+uP$wVYjxu8bk8lx
zJ+mze?y;VAJ6{}*ELrTW5z$1GeW-gda_@!FyU!<b<kOD2WhPafMY?+?ax(K$W8{Ha
zF{`|}k~<v>=@dSK5JprZIS~Lo6aDB%xh44>6r?kd6@YeRI<iFoX+sq(rjeyIL8k-(
zT92wHF9JxLzRNp(=B?K*I>isrmQ+i55kT6?^i%EY^?%pUDa!#}gi?!8*4Gc{vLI|I
zW-3duoK6V_F8Wk`c@aR`z3k=Wio+k5(kbD9)}(66hyc=_w1SGi0xp5sp@24|8Zt!y
zY2StUu@d=aJUS&2&=n}P0#*O|A$D=C_|SXw_$oTZAGjD(jhP~VbPygYx7Er*ih)D`
z+Ma2zE&@o0I^gS1bxb!ikXS&QshR1B0Md!!`S<qCDf^R72?VqO)j&Z6kX})E=H!|k
z^Nkot6rfGiOmsv5>E!*&c4bqosthCs(57mpIwF8{%5j~or59SR&?!p*ZAG<G5CNpO
z+%)X;GN?I^BO6e_H=vRYDE9RO`c%4jpy7R=V(1iK;G#>_l@kG^s{&h&ynTEkmQGm)
zXf~BCBLYZo%UCq;b9b001BnB)g_?zq2q4|ewqP3Q(D&0R0f6REIZP2i`l#fa)LyG=
zFX)tbKx<RA6+{5(V_SG5)!{LA3?v%R=4$3TB7k&fDW+{drT812k^tyxlv)ic>&pjp
z`T3-wp<V2acXUb^aB-!&&J_WquWZsk(f?|p3Ij<7v=h@wO9YU<)_Q*aPt(_3>6E2_
zwx!z2ivZF+&iX9_#q-ZGkVrr~Fdf(;fOK!e=78w~vT1Y*575?B>scaz^!=+|AN;N#
z31J{9fOcj&Yl#5TgWh}Hb~JJt>68#a+fZ#3L;&eA&l4tye8D0Hk_u=SrVCpHkbeJk
zxYFeBC@nfA3DD|Pb!ibmIwO90W*r7bFRaxo&-b*WIKj2!`oaS-b7@!CzFp@ne5}Q_
z5kLN_A`-N3)B`7cF<%9E(>Spm<!wg=FgfbCZa<7ZH#-Nc<)-Bp2;HFMP-x#s4Nfe@
zmd=7sP<X(GLe+uM<Ae|9qX?aFVhhUKf`&q!fpR2C`$mRvA`lClF{uRQm5`HgVkgSm
ziH5?c;bJK5|D?)hpCh(BV5%G($J*Z73Fa6o7>HCTUu2|<vBfhstwDJ;<R*@)+@|#p
z<C7^k*u<d-^@qFSpo$%;QGXZ>6hJB3H^PS#y2`pN`2OOAp0XYbUKvj4E9<jhSU6#%
zY{Y_H0176!d!bsA!3JOfbK$l6+Ek@6=ap7~gKFDqyTUgWswudbFH#C;`+n6>u&7W#
z;bWZe!~DdHLL~)GLK!8U_!wUv?i{4#Z+q->$q3lF64k6k&ETF;Xu)iu%967^F;6)-
z8_KOT?HifJ2_D817Zy$wqr74?jJ)vgQQr46$CapgCCY_63nTHaM3<1`9Vl-H8U_;t
zMI1yXRB$pfcg$TJ8I*G1B-C-z$piBcJ3&DQg(OsS(kTcFnh|~}$}1(q&GvEYHC}4X
z12<;xWFG);xoE<jWp}ew1w7W2*JQ#ff=vOW@JmC0o0LeOYus`nV*{v1V>M$fm{FYI
zsByHQl9w!!*f?JP*;WcCOq5Jm@Z>mQs$|N7N5ly;B{LSx7fzTfnX{mn<AjBh1q=3#
zGMQ_!k!gK5K#UEjQ3E;;MhD9p)V#1K!5DF35w>WC*c73>B2+-;u@ZHxL_NtoRinIW
zQjB0T1Gg*eWaOd-$_6Z0IXGdcY{-Hy2@DiuP1xPYDc+d3IBW27fRnJtkxu@Yzqt5Y
zb5Be0Qky)$QPpMDnNXtJq`dA;dL`LO;)J8CBOeMQPB^(b@nM6)31?SlK5Q8{;o|DT
zhb;iUn&75{9hBVMl4;3?YKRk7Oe;35E}XDo+OT1Z!3kTYEgLpFm?v-+VaFwx^TNEu
z2Zn_YYUsZkE+B;67S4v}0E-H{EPRX;0!%O)zPeWi2hXh7!9Oo7;~@J8SjI3gkThWn
zCg+A=Av3hK7Uk8F+6sFX*ihJ*$w?bgY$Gaulc0WswS^6ubX$TgQG#*dL>bB}LzlrS
zgJ}S33)?n1DHID89|m47SXtP-NvB{eSiC(>>_U0FNF9MG1uF|%Iyor}3lmR*g+r%(
z)6;Px8B3NSV@|<RW{gs?)ET1;EMvwf6U&@2%EGc{jPkI&8Kc$M>KUVaEPuwR5XA~n
z@%U5p^kNnbT`jC?fn(H+sunm!!D;|iEpU#4*9odx;1UI|0Lh3?0pI$-e_S$RGkb)_
zDj|INSq%*YNh8fn(o84K9Ma4s&2^-iO`0o7Gec}j5CZjdgWsRjS;?<mw<bS*WoAZJ
zPHx`nRoQ9C8nG9WMVfh}xt28ZNi#MzE(LsNMHa;`mXMU9N@Fr~oFOkK>!skfP)~v8
zr8sBS?77Oeb`IPHj!w=dO2+ou)&?5?^^~c~QqyE>>F5go6gM?9pKoQOZmBZQ!c~ds
P;k};XE^zTgeUSeF+#Dz$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.9FDD703AE6D2D15E.idx b/.cache/clangd/index/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp.9FDD703AE6D2D15E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21374b5298fd764b31d5e069472950923741dbc8
GIT binary patch
literal 4820
zcmZu#2Uru?7S03+Lo%V7C=yCA(mT?mgOo@M(Uqo4M^qS)BB3uRMFEw#C}KgeE22-~
z*+9ic5d~q_g1EAFY@n;K=mPI1dYSmX&G+%ox%}tca_*V`1o&J}&n{&W$vuWENQ_O3
zlqQi#O5i6oEoH?XFi2G-kpw$$1wF(c+A99u{kE@g>#ms7@5q@acb%p8b3C`yZ6f<s
zaIQ4Ra}EA8_GXMHG+iCK$8YmnaxwQdd-?5w`1z4WkzLH9Qx3}q*1xVdS~6VX5c#q%
z;KB9Jf{)|cJGwj%$rfg8x_f%OwfMn5f4_1`&Pve|)Ia(7#e~|>ZK@OQ9P#n<sP>Nh
zbK7H|jz52Z78-pO*mb|CZOb^l!!+o@-&El+qIsxK{;aRzG5zBX-RHKAtw_{tj~z&E
z8tpxF<#m77Mc%K*9jg4Hx=Mq914>6bhE!kH=JPv>p5iTi$*DU}CN0&^i9VyFVm4>X
z%|iW)TF;;M+P!%q2*A}1R&<^XmMM9bYgpgspds|6yx=mb)owC_xhk_pD?jws_#5R-
zO-$G`*Ef~ObLtQ7+@a;65F)iDyTo|wua;9<YYzwe%Oov6p>wFeau7-OcU#>ib$eW+
zkzROt<?<$aSb8J5P0Q-fbqWnit@AGx45~NZ`Ivkwr>CVO_~xH-9li6))mGi@x89we
zl~F1zu6`jueqU~iz5Ib=hVUhy#0X9avNl(DWE;154Ti9}&y8+`rp#SELMkb1El&4q
z8Mth@j-wov9TnDQW)?EL?3XpCYF>`x?;Q=#E><g;6YyT6CPtpGJa2%4R#=W$E}uPw
zG*#cZbJ#sMR{dkQe?ag_qyMfnEPAB7`s>epwbS!PTTjZq_x;qcucg(&wA8q!POEKp
zw@*K*LiSGZ?SW^@uG!qlX`HjUfnjh^#Yr*z)K|p|b16s03~ykD0r@sD(hJ)=m-q~m
zH{5L!_-Jeyj2RLv^I%jISiU=ZGw6>7_0@qMK0KuhBVlcNk$V3`<z*HOSMQ(v@3ry!
zu?c(YgUS6PaWs>yhkbhZ{F}+vzSVxg>ISb$ES_~)%g38M3h5Wjd^q)q<B?Hqp}q?%
zmRfqSeITUkeyv>C-0lk<@z+~Rd!O!0XjPK!JS}Xz-8^{zb)wR|@TR(O>yf{X_VqqE
zIBCB7_h)G;dSTBE3}4?Fa``P?Sht_QyeKXo+bDZAjqMlO5n39Pb6e_C*^p+q-iGU`
zi*%#e#VQ>W@@of`9Okb(^H)c5(s_GJ)pytS?Gl`?@*i6ll*6@5JMdUq=WFbVK&+ek
z`(>TJhtx+0<9*)ke1Bk%f^Eu*oWc0boXNuOo>zf_ur1yVXT$3<9*(E)>pn2qv#IvW
zhohP&zJAcz{`AYaV=kxag{JE!H@3{Ql0UOXW4%|v^)tJi!v?L-T^-(>UfCu{a#DS7
z>X}`(g0Iz^e_&|pcwBhx)>U2I%)6w3^)D@6hqI3aKe0T#d$@mhUC#^C(|hh+B=w%c
zdz+8<HKxiONgRH{%#%8%960<^)^#HAVA1fGxAgYU_aDj>oxho28K=<Ohu`!u(riwg
zv$mo=5ifcxr;GU83oe;Yt}@BZY*wz&@t%8ie$j=f?SHPH>Rj`M(zE|~*rij)JHPaJ
zjlDD48|Ul$?AXI=GvDe>wflZ8Khn!b9U6~+oN@mZs-ULUU-Ed#qa-ht9DW2@!TE4r
zL2*kyd%Mu7c>&`|sn@Z*0_s}Xi&ve!n*UI~b$~l5lUO%}re-EaDFU5BmiTKFta9(8
zlcGr^++5w9BL<*k;tx~Eb&#%hDx{Hv0BuMyR1gEC^=G3|ds&PT8hH_*bt$@vVwk31
zy4jRWZ5yDI5&`YTbYqJF(xzjX7cP!^Z>N#{0c}OGQWOKEtxP9tKIOltqLJeOU5Zdj
z5!Uwyba4!O6g8fvyopBs3ApG{^c2MaX^)DJA67Q}6HFs71vHz&mJ<V{y{HA7p9fw8
z(V>7gpcpX40BJwB>}aWMGaii`59msSQi*7Me@wf$Z~oeKvN4@T4gfAj6eFe>ARV{t
zmY0-CA%#v_320}gGg}Ohj$fL6Z{N&{XEZV&(E1d8B{4ucp^>w({K}ylH1c9VTT`r+
z!~p4x;-|BQ9<xm7q@Mw8&$QPN1Efp4^_#r)s#-Bp0Rs41M7kD1zdxW)dHKScz<GbA
z&`GO+iwo05TMUq{boE4*{x%s!Bl`hbhoYk(21pC}hfa*XIulJJF9WnTMO#h`klxR>
zVCr#b2WaF#KpRtxnPPx+gA_A?%gtR#Cq)6;Ox=tl21uWj{+!rl)jmuk#{!x|;V6j#
z(oN;)ocZI*Jv8!iKvyA@Do`8W9?+%btiHZobmth290pt#P!`M*1EkyQTP@E|bhy*V
z!GN}-*eQwu(p|fY11Ap3C(+0}K-*AkRKx)30q&j!+x8mmrID8a+LmIgBnC*o)37$v
zl0I{fPKp4usk$jg43HkVaLHNo*MTNFDICxaOb50YApPM{JLApT=uSE*0nn~YS8Xvs
zdbAI9QLyMXpp#+%ZK-a_5d)<Ee)C7A$%_bW8hHhvH7T01Vu1AINqOvZPp&PU6bWbx
zbqkIdAe|B$m%0Nbbb(C$K@Tw88}(LzsbvbNZMma)Tm-`%wH<BXv7l&3QGXB%4EIO<
zm0&D}Z$o(7kPvv28h3i1L|&ee0p{}1_TWP|s4!qIk-WfdVz@8rs|=kmd_BTjkAy<S
z(EqMGVDlo;A%d2oOI2VHR3B;T52}OVi_k?9GnXU0a$;r--;VIMBcX8Da7Bdrf68RD
z&k2nS9Nz?vV>92z8I}_iEN~o=S`j<*L;WNc-HPzG5{o#gIn?DpiCsm;z#{ggNB~?F
z10`=?g#^IeK)sWp{-ARhuA{2Mf^Rd1>#FLqU|um?PgRcvw}s(`s)j6he^4Jm8b#tr
zY?hDmXTdc3*$9)E#^syAK{>V@clh2y@dSzarkpU^@4IAzNkw`Ie~sY*Xu$NOix>g6
zr>DvugGq&mmJl(6(?LZA(IPn|qC?P->6?YRDog!APcb|Y4V-qu@C^uW0}@8$vJBys
zNsfhxxe#%HD~oo*6(Wm?aW%rLM#5n3p<08R6-hO*v&HCQiA5_AUInqp8lMGShQZc6
za87$?`#>n^@Ci;S(sg34U^G|}Ccb3EgJsuUJ{!cc71>O9_fYpi;=gG=j4qSCytnj9
z$}gabjMR;^;XPouvAVG~6zj5uQtRGrnzEI_a8rgU3+fn#n=#B-P_i-HoMFy_-yAU9
zf?>gevV!533`-W2DA-`YK}6jJZUw_b(a;(2AQf`$(cwwIb%KM`A_ldHF+2jSSWwQQ
z=0imLpgs~hQi||Okr3h@g@}_7@gnZ23gJ}|Is>~FxF=D|A|};W)n~!V!Eggr0~V}*
zm>h5lQ5z#dcqnhWWa|YNWO#{np5PRk@|sMj!pCL2?|txIVJD5@&hE}ZuzoSz#oZ+c
zrU%1a-CcvA>|?l_yIT;{Fqm|(xTr}Ii+iD7(?Sg2I*=(*t0bJPm{x2kbQo^Uv}VIP
z#c*4uEgSYR47X$2v0+byH4SR$-+L+`L=6?Lb7NrO;>vA7t)ensWdAcPUpO5|wWz;B
zZwwDYgCxXm2g2Jyh#gc)5G3llM9@#@PYhTx7`_qVZA6yAONYk*L84wv1T8_ANW6SC
z2(N~CV_|WF%!)cR5#)(_PG=VOG~gua(}YtIf)*jubHlp>ZlZooxP_r%(}fP-3g9H_
z+Jw_Wbm8=ZuyAP9AKEsCC!h&(#Q77^M2QiO;u51&G*x1hhNel3($RE@Q5KpdG0H}>
zB}O@Dj>Kp!g04lT#}l@OpDt$3l>!&;Xy@nwlQbDQ;me{TS)yVOagLBE_7Il{iDD0N
zjerRQ#UA1o0gD}~2*{A=n+cI2F3O$G5Gf@#0ep7>|6@ssPQ52G(nav?lLZYDDTy$%
z2{V;2lL<3}Fa?B}Pnc^6b2VY65N0l6rcIktc)2fCm3)-is+5<Xo0GgcH6<-WkeQX8
zz9xy3IqgQuBg`Daj82S60DpRr7RLHYNy|`VQ8^k$S5%PqR$AbuqeS&qo~bfpma46t
zy@R8Zvx^DCaQ+-?J<WgnEmMu9uEo~o%+=L5FfuhWx3sd+m^a(l!kxkN<o<$?JwjZ)
H5TE}6KaV#_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_bf16.cpp.60909DDCF9BB055C.idx b/.cache/clangd/index/grouped_gemm_xdl_bf16.cpp.60909DDCF9BB055C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e05731409201f13b1b493701de9234c9e10d91ad
GIT binary patch
literal 3386
zcmZuz3pkY98Xi9eGs74eX51RaZH8e?;g57dF3BZCF^CyT3>qq8S1Ki!B%5w-Q8`K?
zmy%m9)h1iJ-Mw>iIjHug)K2%*UNdJeGoI(n^UV8v>wn*G{p(xnd)JJwyNe4)8G~`+
zyC(}GlfvaN7z_z~6O)9|-e6#>g25ygw+DoHkD=<nmOZJ;kHHb_8`Owr7cP3x?IT0i
zqd#hKud&<yaC?wl=dQeajTzlb&CT7`Mhx?YQ`43UOV~y1sCV4pe5+A@O6WK-EYG>y
zrKj`W+q|wD6KMPKnL8?SVWk5m4g2H^p7gy<ADNn2@y7qg%&idhZ<JiiUAuZlUc5p}
zo^ic@?~2_ybdcuR>AmTXDPe+bOGnFUg}@_ia%l@yY_I=(tl1!RXTH*<qPAu;VKv+F
zQu|-rr{|kP2dvNHg$EXw*{<HirB<KH39bL}-oELi{piC~<wqt;t(rk&xev%E_K;4j
znOKv(Bzj@h;0Dc358<Q_)33MBDM4}Qr??QK{~Aa7Z!AqZTwfo1_yi|CT$g&O=1Jy6
zm3LTYom<VbBwDz2WVf7={V!U_(;A;dtfsy*BlG(>{i`bNt*H+VpHbYi=x7j1s=KqJ
zTreURU?2RX&d-%NLh)bqcywP3ziu#IzgVTJ`u=F~XP<4xJd1;y4Dt8a;WZlqo8QJ(
z?#SC5nYpf8UoUmkQdq>w@H6mJ`&Y03)rv!-FBRAIr=q13SxGl5E*1WU(<Ub9SUs+*
ztuD|mCRiNW|CnBXu~PduGbh8z_Hq#a$*w7f+Zn$Hr_;UTlsuWT40)|M{QjHN;zc}i
zx9k?HirMej_h+|dF1NRj?7F=E{6+QcWuIT2Ew_mZpX#qGOfW8ub4txl$zO2FHpT0(
zb*OryycV&-swAV&MsIlMyS}Uy>n`S7X^&?0?qxaE6umt2ntONC@suxB`t*FqfTo~p
zbXs&s_LXrfMK7|~{Yd}26rFRkug40T5?vjO?yC9sGTVmPt0`*B%RFx--=b~X?J<?N
z?&kBS{qp6kcpFV&dYc^UB9lCrrxvv!;Go)<(*&!AT-lS19fR7_5ueLSg#HH^$v+x*
z-0!)go^PK=`no8b*f;r|KSR5onfxO$2>o|?=Doe=|6CsPHCOL@^3i2{yOI~P4W%`c
z`|0$n4VNRzQcc?F%K{G^J72`1<~XbFu-dt={v74&q#c!;)g*A_91$vnO}T!}5YVr^
z3?RCWIdIz4@F&aJ--_bCa2(LtkG-ij)S)7`5KsGeP8SMDtim(#>z7$CQ{;0wf3=e7
z!&X|;&SPGjm4X|B%G$Ij(~(xS=h>W=`U~SWeO;d|2Zxs%|7g&Ebi7LMwZ@0vdO~9^
z#O&r;yqW2LRIuxeY4egv-{5=a-oL4cOMZB>;7Z$1o|k8p8s8|aUzgx#yQ`>j;l_@h
zq*LY7(RnsET4v9gO+<H&vx<nd9{F7X<B9^4=KV8N6D6-ybmzAcbn(h)@<VQsp(dj?
zdOEi1_Sr1v<iqLM3ysT?YG+Xo#iB!N({`Ln?mGOzd!VtsW<2BPtL7VCG3vfsT|1>L
zcsEr2b?g&DR4~5VcKVC5mfkZ%zl&+f+00KQ&AdOAnFNbyl`Fd&-<(nFH~fKI`@<j*
zlb9liIFDd3X4|$+KcDf?#v|)67@h^zLQM?d<`I7^Kc=*Hd~iE~N4x-Sjx|>m1EiOp
zzf*efT0u1)Sqo@=tUgf;kaoqTADLL&n}<gN0Zqr!HN*gE?%kfaj#`&zcw{}Gb+9^0
zVt{nyULW>MR^K!p2?jI+%TN&mq&IU@A607Rj^hy@KpS8UNMe9=f-HGJ5gS^GM|=Uz
z!m?Dw0O`y-j3zhJ!Y%@48=xKJ92CU>>6`?$TH7Y)$9TjK&=f2MCk9C8uXK^&-v1tf
zM?3*d!_tUifb>z{s!JbUUg6=9P(V|$R3$M$x-@;ln5WWn8IObmS{tjaBnC)VUK!JW
z@ouIHj|2hQ5NoI^21wUN2_9lz9<x-&gaMkZ!PXZ8r0e9~3kIxz0lT&l(0W)sk{BS}
zewD#k!t4KtM>YVOgXK`f0O<~2r7f1JfjB(k4QLariLw|V{dB@zIn8uuBOVC>G!x5I
z76YV*+<$FdGT76MN4S7C!Wt39Fh`G=@CJ-ejNlP}K$~JsmBj$*Pm@FW7SA^t;E@PG
zYhX1L!~khwWK?1g_$7oLi{R#UDuIU<%AluZV1cl6flSo4a26m-Ll4ZUU;_h{sC%VU
zcBm7aYQ?M4%8|zZ?2#ex9LyY?VC%r{2Dzfvmdf=-BVewl?a_D*R?ZQStB2}2!P9}Q
z4{}9)FUd6}nW_){vrl2k#qk^4+r~gHg+!sio`k*txgt|Ya#0=B5WW|>1BgV{kP@vz
zeV_<>1c*cikrEw}K?~+KX}gE>fZ2K*AFvL#BikP~VERv2_jiqqwul}FutVTuxIHqs
zd5+3fxPCgbTet@y@C<bg4PcWKcnmT_3zh^+l%H=m0*^+e=|a;Hc=}X*U3hW?9-T_p
zg=azF8BvXNVIRT_P@l-7QuSG)DBJ{SRv;2NRU$H=8tA}ngqD@X&9f?jXF@Vjho&L$
zm?Wk;^a!jF%q#M+WL}hy#=)1@B)7`@2r68_9%`s)Xg~|Vy#%=;gG=R}lR@j}4hD9U
z9B!V+2|RPMISm?sz+;hFG<ehmo(0*020a9q0nSQv6~Om&EcWHgud0{?tI{^oX2IfN
zonTd>n*pkUrEn+-P6f*cBGLUIB{D`?Py`PMh(x!9lqdmBnG^kE10WLJ6;h%Fs1_7W
z1Y4+Y#RXEpQaMXFUhwgxVoe@mS0}Qaj*$)<+70fnB5vN*0xI|+gYH3KHW(j4(0u7}
zIf9;+9#10Zl=S!;g3d~hb7aw6S*iKH$f958#u#B_EV%l?zk4As@wO<~B!d5a_%Oi8
z@*F%%bcd~D1**cttqGBvqc%my#0ui#x9~A>b1F={B-|_sBPC&sB#f1WiIOl%5^j-%
zTP0z%Bos(Op(IR@gggO1R#eXF$ThNZ@>m5$B|Je{m8jxIa&|N(;oQ_T7pPOH=o0&-
z4$GFWuppb+(QQl@{=cUh+B&*4eFH-y69&`Vl5Mrf+E#C|7K`IV)^Kr8LRJN=bd~YI
F{0*$>Uv~fi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp.3B6907AB640AF24B.idx b/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp.3B6907AB640AF24B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..23a6a75d0277e39776f53226bd3865b349fd577d
GIT binary patch
literal 11882
zcmaJ{30#fY`|mlft6Oh-xAoq3Y0;*M!WETDmJ*eH&Cc6Ggv{8LZb}J7C_<K$83vI_
zB}-FUY+*2C3u7=wO_uqe_x^t8o-_Z?|MU0xIKAJ#=Y8Mvp7WgN`99Ana#(Qi40jEU
zz}R8Qi)JmEF4EA@Fu?y8FG-yDBL?Am8XC#FFT^a0vvk&f)_DJrR^0A`XHQ2C>O7Uv
z<`A88#M(f-)HH8{V%p4#i(jp7Ef}zL@u4Y+X|Z?fx%z>Bif0Bm7q&h<n=mQz&Cune
zpB&Zewfn}q7uvff>1~|aYoYJ&d#$~GvA)rAw`pVh>)+cAe*647HR15D@4MW+O6zhS
zcZ82ixmZwk_s3n%mi6u)UDG^2{+)Pv_S7EBT6V?7ER65oFmr9={-)sgyxh{r@#bfq
z40$-`)I^uYGf!vBmt0wYbiQfleW%kAzqsW5Y`G}*;Bn84vEjb|e8>rIRCK($uyNOq
z>%Ts_{K&Gxb(VefwXlLQ)z7c34_fg@O4o&$Z{@=mr3W1=3VHwSlZKdq^5*Cei)~lS
zPj6`$GSAZaqqF4?-&>we{_|P&%KX)*Hf{0!Av*is;i0_}O&qOdOKoxwOfT8*R5I2t
zZp5MBBW<o#_BD}`YDK26Y5&3n`l1OdW6qu#b13>Jv!vp9yP$`uFH-Joek4C%^1*9l
z;IaNWCEI`Vu+hF3!52)9DU~eQzs$p{&1FJDX3Fb4|J-FC_N@H5*d$9U>%Qs8u#g^y
zn{930eslX;X#R}jXPbPJ9X6#m)g~S9U%6q9<}egwT{gMk(%_`&a&+$T%}Lk3^cUqX
zUUZr}vG#q(jRQRg&->}^#32jMZ`hk19jdplSo!gj<vgc?kgx53n`dk}7`^Ch*{9+6
zt%lE^X?n}eZE|8<&N9n86URR~TlA-6WA34mk2bVv<!kIK*?l&5S8Ck;<z@Cqj`Tge
zQ)h8W=h|QDTHTyO3#K~H-nX{v=;-Ku>v!{~mbmYmyUI1LYNo*j_kky_A1Er1KltUr
z^Wmp|h?%tQs^QoJ?w`wt1UYH>@0~pAZ9#VHgYCY1&nX)+jI>ivpK0Hn-u<OvnX+`r
zd0*RBg9&wpU4@+A;*3pW=YO}=r|Qs0fB)+@<fpH6&AE_c<vu3%joqx*;Q=vE=Jm^)
zkYE!z{vWyP`2{y>jzvWMwfEGQb$edhRz)|LO*%2St^2M8R^z60G_EWv9k#LS%6TK{
zs@p5?4Y>NOZFX_#2Ag$WnV;RCCVG0gL|qy=QattVHlMBB`a9{x4%Kp>`*BuFcZK|R
zb&ub<U$uRP?pW`(^y7eWTQeX2Rr37Gi_<|FmZt)J5;rQtmp<t{u&!@($ArV-5y|rc
zpE$+z(fzcj@Lu1$<&F~@!qOa*Ze6)M$+Yf6U8ci@>tlY1iyZW(|A*l=F>5CE=u<r_
zA?DKFA_wn3vTnRIDgGL-BYk*3`@Uhj**U9@O>_U4lWm`D{r&4xu20iuSIqG?(oR|N
z^q(RpMP`F{%+a57a-~T*SNrBZXk3$iV%jO4DTj0Co;Vi$Su5Jduxn>^k9jHGWd-#=
z9DkSnvGtk7Rg3gAWs>p4m-{|i?w$G1+*`waPxn6YcJ{#z<(IW1H11CMe1G#NQ|`<(
z<(5|?wuSXJn_X{ux+OQd*yZ_ckA+8<*e)$xr5GN%edPF*gk7C$<wm;oKZM8SNdu-e
zFLpc}`sq-za`M7nnFZ(kM{Vki&613;iul6m=v|&Rdr4~J=ncLz<{#NpcxlF^51p@?
zr%V}fM!e4Mpk8)F(d!xQ4pjqcTK0`gSd<p``tf;@rrC+RHVb2_AIvun?K;#vEA6w+
z)&=PX8jF`Nno*1W7qnS~Jg`1Cl+#!b8XA05_b92V!7fhqPrtcl@Z`W{t98*vn6JTy
z$U_`e4dG3u+AM!%V0TCxjmEdzh;1aQhS0J?JFTJG;ual^SWNqv_}Ho%Ld)x6^0P0`
zQgqP<;pxHhU`JI$c<YTOGZyV{+$cu#@a=<p3^r6Xgtm{>cWn%+PtrycFx^|+TUXT(
z+BGTZk7pl-TWX^yOnZyH4OI=HJvx?GRDKN2(?+8(?J4%uS2ctlc%*U9rgJ$Z+9(p!
z_F{WIRYT~Z8SmfCFa11D8_mG9x!7EyY6u;owI=K7=$cGz6ocuW;+`g|hR{(-Z=N-M
z_<NfeO~dra9wSXu4WTDBRr8gl!4I_2I84jMa*3)T^sKcJo*y<}{X-i~#<ZK*O;^<r
zdhVo^xA&W7JkdrGn06357^oUT$9I#~Yl^4r*G6M7?IrfoS2cuAde=DmqUQ|*U9=Xa
zYw(epk$S3z@M|N_rIyEb8EB*N_;yFJqrR#kbZVpXsZgifDo$gaU{L3g=seQ=?+?;v
z>yTiW)cSRXHX4b$*op1*R1KlCBM+VU_~OTz+Gr-GEyNZQRYT}KYo0pJ)U7_LjTT^9
zCYDK54WaiAZL{vPVEF@GGz*(lkUYp*)ew6BkH6VI|MLSj*h%>AUgBQ*s)o=9#`iP1
zcGY&UE?OgeeS~I&k*Xp5k&mBZ#!MO(qKm%6PY=@%vs5*NE))H=sNUxaUYWW0>DFRv
z164!lvlDV8J8YuTbkQRG@Nh}EnW`c5g)`31elxFiXrqbv?tbEaW~zqJm64M99xG$C
zv{5*wUBoUrs)o?@vsyj2)k+5GqS<(80rCKARYT~ur^9qooL2mxji%ze-No)Ys)o?_
zwreJL@cP$vQL6Caqa~w_RSn^P8TNC9Uqkh2Z8Qbn?j&~7Q#ES(QLmZxj(Lx?(O68o
zid}V74WS>|u6oruv2dC$+9YiBP<g1Msv&&mt6#RdKbvN!jTYkDO~s}?R1Km3hH8%w
zR-u`?Xb!$RP#$QlY6z{7IBWLeczf*4#LE{3kE~Px$FD(>HOP>-H}Vr{{mZ*?{6oY&
zL^5KuD8c1X_f%1OH;%uCtgj&_LU8<bWPP2*^~k!O;%;8rC41{<D78602ZrQ8FnJ-z
zZv)S5;KSmmZ7_l2T<Dhz11a7KeRskDiu0gv9*gtAHy;#~-vx?Y5I}Jac-Me0#h1bR
zGK&k)-~u$1@ZyGxuU@;9X(5g;1h+yML<o*A0=FWHb>>*~n|D<96Nq>vvQ{En(oz4c
z)gtfQ(8eCZ6U+;tHz9;4m>03Q7|e@Vd=ShJvbY4yOIUma%#X16IG7v<XYza!n!0PL
zA`*!e9Iu2SN(dv}IKBZ~8{iv??}5cVu&4Msl3qtvWT~YGde>;j9tUhg`;qg0q+U7`
z*Xa#)qbj;E;S=pMg@iT|0kLTb;g8rP$9qe?%}59`ozW2Xm(HM-_})&Cbb^e$-Sm<}
z+0(#f{@8FP=}j^rzvlQQx=W-KCFv$HXsO;(rsp!FWfrt&vNT!!aE|{@`W=H-gLE~0
zHQpm&vIgYz-nAfI%ktlY)%Vbo@UmgI>t8(An%|w{cf!D(><s6@z&sY`!=QW^L2l&u
zT@bhnMpImXf(y`Cg2_zc859%(ok7JxP?Q7libO1kv{V{_uJ>jbvYCB*8(6fl>w67J
zuhI3j)?MS5rdj_OFU5N>dJi^a803N7wEiVBI6g`i)r)lE_(~+IM0&(pY_u}tH<%f2
zM<RZ9z-;xl67lo<&sT3!+jEI?U#}jx5$~P%p!*&S$kS|pNzCtI)pZ4r<p?r6f-DI^
zlntH|LDtYIakF%@C*j}8Vvqj*n_dgLi}*B<rGXu}Tg0b>ES<#}AkUyvBH}Z_HWSnz
zCE~Y&{Z_iOh}Ghm5L6R=yCN49h3W<2_+P=|S8yb+C4veskf5fR+fo!8nY$i=#UpSa
zcX9l76tNwtJz~FXx0@@be4~XYs|?wdA@xS>uYalFRNk_$B96a<H18l?vIZP~7ir#Q
zaU0TXWAQztd5_}$R`0GY5Kn(65n*R-m;=USC<3<2p+gg4=WV!+<+nlKZPb~H_*^i|
zW%+bun2wm|r6ZSgr1pq-U66k|8baTvM21S*pGZzEt$#^Pj#u<o1d?sQ@o}bcYNO@w
zQroAqPBIT<d34hXPT0Q)Ldh#R{vh}tWN`%;Rj`}>GK{+n(`o-46p@3*6FjhGY|Hxw
z<>+oAz8pE1BNsvlIG`K_vN)z3jU(9q_nU7@5ARuuO}!4R>ZsKW7N3&lTD@C^iBDks
ziTb0Un9e6V7gt=uL_fcNG35J+@W#d|sBUUycII|6qAWSy-rQbB<}>*GhBpge&)zHM
z`04u78P_@8cDi~KV6M+}CT^|=$$HQu!{_)NAl(7#4a)KP$RHn?5}xRNykP~gPe-MQ
zRLmel51rzf(XA-_B>tifaO{A-<nFM%i<u4UUhCi}+)Liej%ZYlcayqVknKLQUv`{{
zn=%XEZS89vNTlY-%Qr^fbbEURd!P=`?V$cl#CL&i7xe;TM;Z2dv2|c~LyixS1n3dD
z$nk-aKs}<cI6g=cq(?L~#|KM-^@!u;_z+2m9tpb~A1VpeBf^g3CrBnRv&Bnz2E|F@
z7!)swXHbG9fkBCqL<XfuQW$g+Bq!Mwc@C22EPer!7c6cENjr;Qg5)K|QQi8=LL<tS
z;OSJNUP`1E2qZVd1|}qCbSa}0QA|v`h(s5W5%EDJV*{;!WolDPN=_e*^qYfyRugh;
zLN25~$2TL#W)`;~#}*d1BF9#WV>Q!@S9|SA!=v4eY<DA9Vmib+u;B@bARVnjrf?vJ
zM>0bENkTfvbQ-FNAe}fq7e(ZviR8)i5*IHG?>saZ3%eE0EBcUMbNm^QpP~Dg<Ih3A
zb1<0ll_0NV`6`fCvA6?79iT(|cY&yj;sudY-J9P}{R4k#tZ}S*!;!2N2Nxk_Wj-Md
z6gf<Ec?=ehsSC&Z7Deo!(Is9tyJqT_@z3#2>jI}Ps&6^|3)p`FKQgR@<Ux(HYt3J9
z7`g#gH^7<@i;py)&rY2t_^W*6o{!WMLlS2^$wJah%t%N@9M;eU<M>vj*NRMtB`lTS
zG=KkcKm{IAHOQ-3*If(pT8ftr52|+^*Ea&c*eAeeEKyC%{uy#*_0kdr&RI&ptpq~I
zK(VF(w<9b*4lc(bi15kPkDsg%HT{flya|RkLG8;p{yrGqXK_17+d(ZdSG+l1U@U2R
zf{pq|F#nN`P%s7ab1c3978ls1{|U^0qWr3}HFlH77=+?4Dnw3&$eRppb<2&jsmb5j
zVd5#6J*CQp6d%M||0+6Qw`t$Yjk?Wv(|C1$!toj4kOA)G*?8GNmd_5V5<DvD!mm9(
zVo_8<Qyn(@7O-mpFM4Au*tN3wCJed>BPo9i>~67qCk*Ihh3O~o{=}wgtI@#KXgIll
z)7zmkzec46{%CW3a}yHhh@D{1@}FV~G5z$gZtH)~{Kr=l`*Y&}OU9oY2U;@z+&IXR
z@#n_DmW)3)4zXnXxpAl^_2*%lVS1$6Ay|a$K3xXo8o&Ib2_NU<p~gnW#*i;i9@zdM
zE~WjRo`|2|J3*b3ig@0hXGBn(eH?@09pcq%BjOY66Vwwf;uGx?)ixpGQ>;>$o++Lw
zYM+C-fRr(07$Sa^<|;FaPJ-)6YW5=jIe0&(`Jjk@0p2fI+z#IDEPe^zFDX{uo9Dak
z%gP)0+sh!R3?`EoZ5_1llVPa;GQ5%3A#NRFBG)ElzX_>zG7%|Q<_WbgvYuGxc_@6T
zH|9-wM&_k~eHttCc7kjt-AO|3?O({sx?=DzX7L&DI75{y$JfBP8dm0Iqlj!O^Rl&X
zbI~(ihvSd@3RYi1y$Yl@hm&rhFh|GkVD2CzVTR*-n)j3uNyPDv=8iJ5b2;A0+(|~N
zN!tpCWKa8)C?sPO+$X5@P;SSx$!qmwH&s(<v5rn1$8SPjn~*xQBDFv~9H9_MhvQ=4
zVod5~9Peu2YW&N8D^wic+n~2GNw7HH-N4<L)W*m@!7C#a4oODH@teVIGkB1d7D^1}
zN>&W+fWRG~&i;fDIra=h)BY6@P(e+Q<Lkk-9^5E@7cB0wBK!)HUO}d0>kHMWzSod1
z!F!)|_K7Qnay)CRkz_Ser}+Dh%(>KBmDGd@8(*6Nq%^Vbp`_Qo!0{7sEUQMw)ksPn
zwC}Y`M73#$AhEB2{0a@5g$sv^!bWfW6^lxo^5ud*8LCkI>8C{MFj-XdxifSDyu(j;
z3p#J9`YifG>ip!TeHRWA)!<#tnqDn<*Rr?{eCnV-d4nKu2Jc4d$s}GK3ml<ZOP5#?
ztKi9QA!S<}dxerMjlBh6QUFXWXhwR?$bx+Tfz)w%Pm?NT*v^W<t(a8<2f^(i#l#U}
z^&!-JsbR!8#4u)2g$$}_HId_MkwY!=Aa6UCT+uyZQBg3SD^p!l6H@a$;ruLW;=9qk
zaY~hrtkRKs(+k)q9rYtKBj8c#D1u;;0bt`7D$R6eJdHf9NHWgxUPfM4L_Ois4T{)F
zH6F(wMtu*{jI`|Gpk;Yx;dAkX)PP?NTf?~wewQgOpLC~Bky6hPcgz8u99q)A2`SP^
zN9q#C$$J*kgqq9a@JD|F(I+q>qdyazn_N2h`5lo6OJmn2I!z)hja{2r+=5(NSlo(S
zTPZ%P+upb_e!d2N!zpBa3OSKCkZL<N3Zc|ajUo%_W+A5T)r9n#Xn}~>6gD`a1W%{+
z2^c-0F1PZV$k^^fi(cUwkoA<g5l@68Vn}Q_83J)ZcnCtho<2*Vr7$2b65LDpX6i1g
zgWkk)e^qU;6Y;5FpGtS72zRvHO#Os_JvM_Ud7FU4HiKGG;LQq_O18feJd|vITqjks
z{kOn3TWGOb#BYJnEijBeKMO3g*ym+|M;80MEC|bDpO+1m*=+x8@W^KSXG2&v+kXdm
z?w|@vgp=-}J7_Uoz!P?WIztg*KO21p)cQljSAa(a`+V$lE2v)<@s%*5k}5V4Uj@-s
z?ETll?K+i;BJ8K#>RF8I$PMiL+y>cgTCouEcR+TB#n`FeWihTxwz2Dp>(KXD`~d79
zu=p5qKZeu}M#Pt*KBZJ9S09exb%Hh5V;_4449?Jr7KFI#Ia*)9RZB3aWceyEsG_*8
zJ%4q<i@r@bK;(cthbBAMER%ngB<ps^#2c`A1L|@QQJXkY3MWm}A*O;YHd<maf@zy$
zqosHY^ub0(aTb_kqoOz)%(2l>yaU{^QBZsvEbydLd<^x*lSFYfGN`6$d4v7R)U|`g
z=;2M83qy0MDGN!zzmk@#nx0*XJLli}+(noHoKLFHL4+D$?`<?z3wZQ4h$L?m0;G2?
z{UQ-Aj`!I~1(S%+gFbmI&WCUESsU5~e!FP+6JaGMEd+IuN`wn|(jpcYgS42%2SIv}
z#U&suVet`=9%1ouFg#A(FqQ;hRt@I#d#b^sn$EZg2Lg{87T1DVE!)2qJZf3~GI(5O
zaUGb~(d1Et%YR;Vpw5~_Sb3OTVV~Cm@)oe7<7owXD~oS}-%WPDZh`z3%kM_A-PF5_
zu$&mUn<_>TUw{S{uy{Z6-;bD7;sEkLz~W*QR7~ZFQ0H_kN9wYnfc?wa{^e+FIf|m=
zKZX2GA+@{_;UG2Q6jEo6BCPFvuOUAo5=8i<(f2xw>yd9g#cd`Rzpn7NS%FVss==?C
zJqfG@zgia8!GJod+&KO-_<m+j8h4}M-AH{-Ml>*16T;yz6RI2)ZZuSFfru^CV{&{J
zglEyv#qrq?o=x$+4Nq@Z{?=Wnc-}{v_Yp@vlN>i=K_whD)7ZQL>25%3YZUbUs7*94
z!8tk7YoP*;<Pq2n2}jb*pyL(G>7Wx#6V$4T;}cC28I%c5nKa~Z_$b6>E8VWb*_CV;
zjrtr886H)vdfkR1wy^`;gZk~E=0LIzEC7T<ab|!<3JW?wC4?*KtdS%fFNJVYPMvEO
z48akWn3gcvLvRG8_yJhp2tsi^lGamU^Q-yS%B0#`)~fTO;51qp!om`Q(^<R)9JbJp
z7S4-?XHxMeoEMGU3Zux2o_cow&_mBO9k22Vs}*Yf^-A+tChA`%6s+Yv<({NO`f7e@
zl%G*RE%wQ?O=qhMnH)dgbiNwB$|?O=;M%xXi{p8HUhNI>;fH=4gW~n$8I+)(z@S9^
zL<ar)giJU#VNb|}A`yE+#{I7oGQlj^6SChISDZeY-0QJ~!$)t1X>7ns2g7t0XMj-#
z8?Z9LG!q#6+Y07eX^|eMdSG;o?cWS`&7htae0qY+>X0qjM>ugqW>+YFck;V2ns>Y`
z@tkL)fNZ2zqJQQ@AKf(SQ5+_I2lL<Q;*vB(Q|n)Ua)GN6pF#13DXc0okYWp>Hmks1
zbPtPPAVoWi|3-=~7N>PnZ0^R~-_=d=NsURtN>5AmrybPhmEeX|D={?=OVVJhN-53)
zH>^S_&IUKEGAZ5x<FIO^xB}d;N~E|F!mui&xC+K%RY>u5aK);S;yYl0RUXBq$P=qF
zitCWHj!oUKqrTXq)Ag!HeX&O;xGU}bmN)Kg;rJ}F8Z2=Em^?_t*McQ308?BCGJJAI
z@fEPdNjt?YV1o;m6t{v6E?82069(dfCB?VE1{W+T-i@sB(E`C=erj;cADF!R|J|jD
zvldPlT-xHqnTxLr7t4j_KmS&c3ry3?M0y!VFPG5ErSx(cz5G^v*^T#`Wzz54?>{fm
zU-fOu%DBY|iSbL8E=yj%B1t1b-Af~hUdGeQ6ndFVFIUmamGpAvqS%G_kGM3#W{v13
z>Mrh~DbeP1^!0Q@4T1t34YWcHO^uAr%zga^3{(vAAM7r5?Qh@L$@>5Lt%;>fZewfb
j(6g7bOK%TPFK?fIR^M2-^$C=k1P@ycXlV415E%YH5-9I?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp16.cpp.C49639C51E0DF485.idx b/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp16.cpp.C49639C51E0DF485.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c29908045b41b5c9a0cbc786546032f43a1e9d37
GIT binary patch
literal 10730
zcmaJ{2Ut``yXG6#m1PgRv;~$bMWhIbD6#~lDS{PH!QQbQAOa#{i6YVkL~KzMTQn*l
z5Cp-94Py;NjnP0f8pRSZ5kt@@B>$Yf|LjcebDzueIG&ex&YU?@-g&?85E&K{@{6;U
zR`9~GRdF#Z=83elv<&dSq!o!vzressuBElA<ZN_Xe4CGcdrQZpK-UR(VrTS|H`WI3
zX<8E56|w(V(a2qPryLY*y&r~ei}_^bS9J@G&R43up5FZ2Lba&Gt2u5T?>gAMWS~dx
z+CLwU95ljuUD$wo>5uP)`IdCI6q@}mfAO+o$)y$D-CfxUpRH{2nVgWi&Qy8t$+b5h
z2MlcddBW`<>-H@5-8E}=*M-nKEjzzj(Ve_}*xA2UZkJV@PaXFrbx5m@d-3h4ZQEuq
zym;c{;o7l@#|8zRJo`4MW=h=IUHll+r`*eh51sb*zaIKvVb`WVCi)+{`;(;ZA%8mc
z#(VjJp^a5%?2IbyijS4>o&#Rnzbr}ae7El>-)+~8i>CI;dfI!n{KQDV!<%(~DsHaw
z-q8~1Y~<5j)Omfk-i*f7KOPS+Jl`5T+H&?J$8q7+M$@#5<+2%Dwk4dK9Js^p(99tX
zlZqo=hvsS}ghWaQpyt0OEU^%!-D{5W?)d(eu37qVodIi?=W{>&Rj%{!!k#f!cYew`
z?xo~aJ*aw+p6NHclVdDehOK^QSk85h@p-zSr6;WU?V88iu5Q~jd82W~lKH;A0o&hN
zr#-K?+uLcz_piBl!Cp0!^Dg6meJ|c;>tLY2{(NmRryZS9rSl>)I&Wax$+OSv!&l!}
z{`A^|;Nt6AV<59lH|Po1b8}Tjm)7)~OU|7u`f%I#X=!O`=iJoTyE~+3&NaQRI_$k@
zUdQb>Uo3ZAU6*!tduowU>9EuZ1;b_=*C>Wq2P{xYi{tDX^PgQP2$)hH;c4+~YtzP`
zv+lcjhVGr?xoE#%@2xFGPPQkeyX}=aO}}@2#+#|4qq+lr9ejKVSF&{dyT7haZ!Ovq
zQzKGTN!Q2fdlp=5S-jI|?G=aQy8g{q+@B=HwWEV>@gqzV(;7qr%Uu-i6U^#1-#f7B
zGpFF@*eCtI-F7C_{P+1EOH2E|kWF&v{pw+ue5do^6a%mBPoB7T*X-FGg|eMqo*OG&
zY%wHZdia#y#c}Ho3`h+<d}7?u`sryy@`jxJ_<Z=ZdjkWPIy7I;J@@MD>%9(1F<T?=
zMpj>#DoQZD<Fji??x^LHzPb=yRXcvJwsXoak3T#0vrmV2!8gUjlUDRM88RyAz1!0+
zn|^lM|6I)57jvl1uuu8YteAyEE$*D%+9i4Jo-aqqSHJnJZb8Q<n~SaZ>h`iw@2Z%P
zAD$lfIpY7>!p;kO4|*JWn{aRS9KSPfJ6gX|#;2Vztjyc`ZON!3$7g*1=F-r~Yrh)1
z^2Xpj2lpm;UhV!RMs@KI&#>85qJ2xM*6145pFbM?w@2<T#fxVx<6HfL!y<--{jjL&
zqw39Xin7{-KZaiGHeYu6<lHH?I}257w_J!mf9u+cHigI1LhqEO?qRcBCVq5lsaiXv
zew1j&4*Rb+bS95kf59PbsmEvW<-L01pIoMAB-@SMnrCeEMB4T;_xQo<4S!18T*~FM
zmdt4}+!{0AWX?~HjdMqT)b@JKPj}4NwyUkI+j!MK#=oD?du>@Y>0R4P{Zi5392GiV
zdf@Se@IA}gFC|U3E!dTIeDRlqc9~lJon3P;$|WwnO0L}}&naT%fg0tE!$}5L8pk+Y
znz=xB)F$A3Wr?h3<<OtXL-t+%$7<G&`=z{LDpsr1xcQ5)T0yJ%=m$1sqd2WJ(9+@~
z`a~$y1r{{*kN)kZfLXyQ8)PUQ7HjdrR>6+yg78*1Vb-(WXKQ6>qreAQ1v#n<!rNq;
z%#S<Nk|{=uaPxkB{S4Iwp<R=ce}C3B)<TLVW7<RPVW=($?eQ-4Ok;QG=TbBQ({5rn
zeRV--zvC_YvKoqyNl^r*?ZkF+bwTJ+`ENUyRrE}iqB)pWij@*|LFm!CX<OPOnhK>T
z3eyf^2NQKc=qbs+J-gNQ*ETVlh3T-qVJ7N=&@*pcif^n4c_2m6n6?sINz?_QV>X5l
z>dL+LyA(~ww2Rn9rY;D*WagTChfMRIN>Mnb?Zx&6>VnYAwG}sY#Ip}cQ6#3_#qRp*
zg3!sGEfMDieP<v;>t}26<8{W%)diumTbxgYI&HteX>AbZ^bisqLOTEbkJKp&41t;7
z|2<!d#^WlsVq3YoAoSMAqbIvx)Gd&rxtKN=n@iM%hTdBGy<=5Vqn_40OnaJm+Nuje
z?@McUTp+ttEk#Q(Z7H^ts0%`uj%u|T%%?t(q4{{p1Xu;ws0%_Ls(WJl{Ex1qQZx<U
z-Cx{aUtJLT@U&qj*RR<I$k1o_<;UraGg23XKHmK?deY3W(K56aHy)xNVxcYwT`BrA
z?&i>|cwS;~<2GU&19d^@`su}zo&Bcd$j~y}_*ltUGj&1ev!|V%RST}alcK5kZZENy
znYti!W29u6$C_weDVm6BC$W>Bx*+t;nD0HdU6zcHp#`{WBdkW)s0%{3wukAhby{5`
zMKkf;u3}d`bwTL19XhMt#p~aYp$x!%n;@BBtS$)uFznkis^&|zQZxhK?kIMYs|yYN
zXuyJ-j-Nl0qDh!`7CY;y3qn7#UH9t!j01CJXp^vyM_P?^R2PJQ|LS3p`?I-rQnVD`
zZYnnIt1bxr7hLk_vJNejp@sNve=C0*bwTLFm_<qRY_QshQ)5HMHx37me+!bgU`XVL
z{Eu}1r9K>g8F819C0WrZ!R1k(Y*C&z$5$eoO5{Wcj;}&CRV=PXHq{hsdrFU$-kiT#
z!tuEfm<u7~fgGO)gYsY~i)ZA)T#5_9yAVF1cnf%MVQ~=*FM=S-p8~H_;6rg8c-67E
z00k7FQG^#apL^BeQmKnLekZu?gb{?`_+8+(i(<XS=EIhrusjbUz87qI!Io6iueXl#
z+8){>7Ftm5gn@(*T2Ss{@g7j_VQ~p4OITbA%2F2Z2jzZ>$x+ZpSA7ByW#{+|2+V*n
z^s(o`^*jur_$K(?gkXxRkfI7%lc`f29@r#ZSPpni6eH(iq?roS`ORbg)Elt|6Hh?$
z1T0AlrWYM5+k;d5@nlYuPctFE=J@5Z<qC=tWC;vPG)gq5T#_P5(;UaADpDD=TCtiz
zYZPl3bQG+QLVwbZW!Sx&FP`fz6LEYY_!Y8aum$|Ku($|<iy(sB$ngazqySAOn4F>_
zUG?Efhc+Do)7fTjfbR`<C@Ybok`ASfEKQZ8bMrACu?{fmfPSQX(xA5PzfNC{kFbmw
zKq_(k7fAF4l9P$)r(3u@!_06867dUy7HC$ih+pixcsRMqwjpsz#|FPlyevCF)&T|@
z;$h<MzSbYE;=b%gX1kFEA)L(j=^wi5$%{n%MzGunc7zb|o4|4ti!;GC6Eyu3@!4RX
z4bG%K@eX*<1mi%ohT~P+L3IF#R&)FvFuMZ|<Tj4aLE$-QCc(sG;ARAa!8Eg0mC3Yn
z3(RiO&T)Je3eQ5*Nh`kk7x$j}Jmqh^6dI9EBa)F(<oI()=NyaABc1aszJPQtQ0!~n
zd7T%}>(mipr83L~WAX$6+vU>t5Mkvq%wzdH@Xn)3BjS6&u!rS)z@>+7XMy+cfk4`R
zFBtaH`ovu6>i)}A;o<dF1=GQeHI3CQGY-$7{U-Vdj^6^7Tj+if6v=-Vgpx;cd<poM
zu=pexounI`<Lh8*9n7QkH=*!NXd1x*;!}$4)}8C{8fypRcCaGE$msV^cO{*<hzT#1
zS2XD@F@$&<2yT#0wXM?Dk~kucw^Q0#lD7}}I^(z4jzy(dU1#ghX0&Fu?QG4;#N6Q7
zjO4GAtdo<j;H?gd0?;gfeBdC1JY-6E;zIGsCs<J0Csm|sKIwSq>_Pe32PanJ_jwGC
zkHMSVJ$~5MPfgr5Z^49(k4-RfH{-wgF5<4+o2z*1J_6YzuqO3I_<q@ADh!h+7!G(*
z<k!c5<3~tF$caPYcz=n%oE+*LA0P>k6W7M^L6RUjQ7n!RmITYm(ZKN|B_rj;OK^OY
zB#Ie`Xh}4KcnQy-SV=5{;v{hlikHMQC`FROpd%nT0t~(nlKU+F5hOpdxD_O=6i?Ci
zwhRrgT!DwL7Y6jwLyMSyyix?`PiG}!c*Hm|m>houiH;y6veijQfTvQ32k1@C2F}J0
z|Lecz96MHfB2u*&3q>7rtV1s3ZjL{V98a^j9y!*txB)pfP`pql?~4td`*LvK^O0>n
zawXj&p$G1U5O>gSs4P_uWJQqx1izpVeK3_qslrJmj?YBlnP>)Sd1+$O%8Bof2H?$*
z?3_H9{F>v-z^V)!DXxHF6%asiHCR=%{38%O0zJw<2GL`R`N%o$x8KhB9lzN;<9V8$
zLxLhK)IvnWbU6o9#f(zl2D96&QfH&^Y^v1DWm}u(^iF$@9oHjpcm%$*8_&W1Ij9Iu
zI)3}>t=V%0b;(2Sc}VlVB>ci-C&XT4tq7jh;Vo$9f#VyHyaAbz30`S+SNZnWPtM>6
z*ML<GD~q*YRZH>8u_JG~PW2AQ5B3c5oV@kF!B)yYfmb)IJT@HPUkYxeFq(9f<M)Hx
zeu`IJdi->?=+?LR{#H=5g2wl({;hnkvE<fMY@bhn@&tV%$5()|g2h#!tfF{beUsg+
zNd}>~#UkWXguKY3H{AKIK6}-AJ52llW<P*EAvTtei<`3d)|I|^6kuQjYnYoLsEM^g
zmm%mfi?2Y?6&62%PoA(Is2x1osW0aEPVniZFQ4^hl%=X=vpMdOQeSC8B%e$I=>D%j
zIIr*d1KAf7H{en2Z`I$91fLx5qHr-MTPbgh_NLc`|M=jy8DTuaf{{sMe+xz?jRPzg
znKTZvU}Vxb*n*Ks<B=9rCPQ>W<Rot0oTl1+BB6U3?*F{VdC_D%Hy_^d;M27)+hig>
z$|p(_Zix73(`d~O5%Ih|ukrsPKGr^#L2>qRo}@VuA8#M8nMDzwVx8hnxs*XEn$3&3
zpp;0;9Rb%PFp}bnV0{s6D83I~_o<GG_#eURM;5n&S1ZMv+m`yI_pbR4zwKcdc^GDq
zhKfe)|7aNMpMs~Q3%D*S*&P2K?B7!{A;&hhDMA*5Q3Vf`4;3=rlxJ+pMzG(=+LS`D
zETpSl2&DXXvbJOo`0rtH8F-XYr^WH7VCpH>rlh0rbZS$!O7C$|^E)QumwW-%FF-Ty
zB+G&oO-QlOzS}G9Er~*Nyo1uglGp@}cT_rBl31Te1D1Lr2}2uD9RXF9MiRFj2;4gN
zW1<ijM7c+4%)|C~b7yUoTi#WVv3VCAQ(=rf--E{FlY9;qJRzAwALC@;WK5DC9Peu2
zYD|(CWE)6z)wvxKfN*>|xTS*!nPwq)RA#X5r~rZsKoe{TN6X|gh@$mRLeNRN{;=By
z*G6!o{B`iX4t^9@AVmc-CEH)f8+cbDABwAxcNN8@&)yIIv;wN|NOgjwlUmXJ9WLRQ
zOy3EP@-$eTrZ)UQ>{!v5h|FKG)!Gb(o2h>m;`Cv?^f)=#)bl=+hfdtWub}rU^<W2o
zS2#cY)&2uk<r?s+Vb@hHc-68v9|hzi&EOEN!m}@A$>_AG7OLXO(kJOMJa<B_j2>3S
zU{Xx82!i{NpGM|nU=C+b{k%Q7(GstXJ>a&7)!PzqE1{U^7GA?bvW-skB!@}NG=GT<
zzNC2^jz5kZjw289w6axa`h>?F48db#Dl;`9$)=Oe&!)`ij2Ngb#46T3bjSoev<HTf
zX%_H=9tbCx_+Z>AA;m{W#!cbYj|6-i@2+s~M_dKRdni2mk(jdb=MgENn@wDTU;h-S
zP66Zj>OfUTan;QGgAZ<&8{&$&pqESIUk-OkuZQC5HgiQn(>GJ`E4G8E9gIj{Pls$@
zRT1#~K8}a$kZT<s6cL^h*V8PnN3QiOZa}UL6xYjMwq!0}riGtSf^15V6L|tjYT_jz
zWHjj#NJFwTqzRe?Ln^PM*#xqJ@bVHeopccIg3(>7mW@Lq7xo!-@D&~b%l?*bL@V&|
zHG1Ae(gh+CxC=sBl(wbPRT+>63F0s@oo@C^BY#`S{dpn7O2nsueHz_!B3#iToo-G6
zd!)l4@-zXDNr$l%V~NRN>t}#R23sHJuQJ&BSzwXH*3SZuEVh0YjLBl_=YT~HTR#Up
za@hJgFeZnsUjTy&fH{q^<Q34QqzK21(+fc3yG2+9qsl;I21NWx@Hok~k0tRWRZS6J
z4N=u>{j=b97MK$rOJ*aBalY<6O@fN}tKfH)>c5D;27cFAj3xFui*Zuw232JdkMn&u
zS=<c4%@kkyGCW={L}wFLyfQE-qXQ@SP}d3?-g5Xw45}%A<>l@TdtZ3p!r7)=u*#*r
z<GRJFUnHw!eK7G5`aJ{}GWEo&>FBC+?Npx9z!onmQcYNTO1!8j&H^P~1Qh3h5>GP4
z1>lY+k>ac1iwBM3GNdS@=T5Wzn(U1uCdu(8*$n2JX+SOnp8mZwrFZMu#ZMdjzkhBg
zyy5U%s<T46XJ8&Z=>!~|2UE!tMK}dHxR81a5zf5~-ooM{P!-W=M}$p|Vkc-a8X}yV
zQS4&z9#HIIaS13&SX>H<QWozA#eNpofLRSFY5!`#qlS){2n(ynDHhj)SuI<?7CdTM
zz79O<SbQ4HPP6qJU_b+CL{G#w!SE(_{4T@r%PhVE!>_P7A6e#8H5B2f+AkkzJP_WK
z$ghCK#mK)HF+O=4^54ed?P%n7dRU9FD62}4CfXG6xDqs;jt{mal}JU7L=pa?;Zw!p
zYUEQ*ajVI>zgPSBTaEJwHK3}Yn~~#dK~>A*XW;XUU7GnQBp+$=^~AB^vr_ox#7MKF
z%8g2M7EH{d(u|X+FfoVXwv6_BjZgXr395@o=OW_BYml#2xOL%+6+L^pLDo%YPjH+Q
z-h(Ey$MN+@UQg`^F}&F23tzUFPA*ra(oV*i#xX8G-ZY*;nc$R3C6~j=WS4BZJ#b(P
zmPPbb6O_lJn)O~8C_ICDdYth`!wTtiKfHAL!`;TtD0NmlWFw77g{*eSCKhMH*i335
zgsk@XY?wgS0m&2LlT!Fd$GoYL%ACGw229ML!%sp<{Aot`az|Az3j*;`McNi-Ef60?
z6gPt}mS&2pk)oP9fc8Osy86mZ^YE-DTPJI_*0XPqo&93%cp;wgjPjgAf{<4_WrfJM
zQpiPG^|u;CvSF{5RZLMC1>p~Q`~uSjOm<|k>0%9fRb0`%*R`cohvRvDUL#5zAFChB
zpg8?F2F2^gGw9zR6NFSW`!NBhvi|ikL70E`W5VmCGqopH4S1}};ja>g8(F2_1csYf
zoC&6x)QSoXDYMyhQv>KVP#xs>ufgbRdW3OUm+YF@EKwOUD?_$q!*F~#GAn0s1v0Ck
zxU+iwB%S-77I>7?QBXS4Sgb#aqfTT^c=RbI9)R)zUCP8w<0DY`Zp*Bw9#Hjaf<&Cy
z&qt~PMAw*VFH-Ge@e8DS$>P6|>H~{&v{kv<%>5s<RUb8&q)Bi@D10BLYcvDgupOgo
zFbmwU-J&=L+_0UZxB#YNdqnX`aKm<p;%b<T?GMFg!4=ydim!n$wlfr$BSkqqsqw}~
z-dJkMe0<1xTkxBE>qPuLrUoo<;*Q)X!a8e#6L%D!1`8Y;QrrM8IQK<y69nMg7sZz$
z0O!6az5)R__eF6&vcZ{Mf_u+5JMQ*dRr3GtBdx@kSo|j$`0slX7bM*fE>;S~|NNwo
z3oWhX^zu`B8A~sd=;aD}xrScGYc92Uzv`2BCiz5K>#tsux+->g{HJjVNh^|9rX*@5
zYHDf4(M$aIN8~4gTxe+})63QLauvN?5VtUPF@|Ge#%YWCi2LeDq@13<To!6DGRV<D
zH`LJ7$k<Hjqx!^e_z3?1cZI93y|<Ii|MgoF3rnkhwssEv2ROS7^cdvnHFTKu5OcS|
S!3vX*uyttEl)%w2R_j0B5@W>x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp8.cpp.33128DEF03F0DD92.idx b/.cache/clangd/index/grouped_gemm_xdl_fixed_nk_fp8.cpp.33128DEF03F0DD92.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7ba6a001911e1919997b94d8dc8688796e0fa394
GIT binary patch
literal 10828
zcmaKS2UJwawl=$HTN)0T2AZ4%1)&iW^q^z}90gGX)4(7Y51^=s!WfVYiXb|oqGB8a
zf*=TDM!g1<(J>7?<5kdW8p9X|=GEzStE=W;|9Y%taX!8}b*gq%eYN*4BSV9O1D(~>
z0un;glM)yHC{j~X)5rf7EKHrf7Xu4jHMR6DC!#x(J3aNDw!a+j=Q8#;eu|Cm$wt4;
z=VnEI7`J_2Y2Ze?1_x<p&({&<iQlEwPvh$(ad)@KQoGt#TMWJ?>mJtLAeEVn-Cz=Q
z^6NZ1Q|pqGyVkup^T!JZ%dy|OM-_kS`(a_)kslQq8B5Q3`g|8}vA^l~$)e&HZxcTE
z7#crFaeDr;^Xd@GHaGd#xLwb?uS`;WQ#5V9pV<=DH~OKs<7aQPB+<{Ivt7c%<Q|W1
z{S?{RuX%!L)tw){1a|Ci{Gn~=ZmrbZg&ry8*9;POC3`GA`(^L8n{!@gPkA%+e~u^q
zmG`UJgS*p8B{8L!CR}USxb9==!fVGqXM`tPh9wq;9Lx%Mx9rL7C_Agj;6ATIyJO6+
zbZ;-0SbV#fo_ICa^}+Eke{3BU_o~WXr)GemWB=a^4z9e^H_tWXa#QSno6%dxxtZPD
zcj=cjhoBXn_WN=Sp*z#*h<4WLxiekN)BZE7PqANAw&BR8VDXE;CW_uX9J71P?AHOe
zDvBE{mcgWrxmwe&9;;j3ZGIx7U}A21KZ|#VXZAYG{BmMmSVEY;v!7=4kU^G~+ryTW
zoPIyG^jdmxgTdrS4@)iF#OAk$F3MDI*DKS=_lfIN<gPq$Hb%6t^+3>uS9k8epW6RK
zqUXCOud6~6)po13CiSVBFJ3m|xc^c$*BM71JP7$|w&lC__V)Bz@5VV-a_+@G{umQ!
zUhkeZ`*WnH;fc=E_b(s6SkS)N+v#8Ca?K7+Gu`cnx8z(JrvD)LecY(+`)~iUQF}pX
z{FyfaFE^O(x>jR4IAUw>oQomlYon?+T<c?0J@=_=`oOI5Ifjd5QA5Pnj-2|Indtf|
z<f6yqtiroFKEv`&M?LDUzalv~z3f=0LHG2Z-!?a<!gb@}+q9Z}r=blK*K3+M6;HYF
zBEsLN+;(Q;{IxSKPjgtk_;62qyTQV|0cE|D-t5Y|VpP$RFZ;Ictu&)!Ot1?CmVbD#
zJEvZIlc(!1`+Msfm;buX+u=skp5LZc`)&%_bmvV*-JwZYmqWI!moC4P(fM6y*JX#@
z`*$a-I;`>7ZN_}}>qc`Q?wPTwCO&ENX>sK1bFDpdPR+lYWU<;Qqw87Nz*&X&|Cn;_
z`5S$^-oIZ5wAWovDwuQoNPtJ(rq$1X2>qHp{pSmPSK1{^Kh-(tqvhF4Yli-wI?-YC
zUAg_-)F_MJtbOV}cZ|JWo}RF?cmAEFvED83UUt;Wl2^7EoGvQ6x+Q4O{wX*AXdQYa
zvwn2i6_3q3D(4Tn`0;U~*W>qtLVu_cZJkw{Au((|y*Ij7Uii3Z=CnC;JC6H@jvF3&
zyX){buX(Sf2OEEUKlIW^<2mPR;-YNVmw08aJso}e`lW@PCi2-O!xo?WIDDFO_&3+~
z!<j>xf<#j`*n6*dJ3nms8Hbg#<;#+*d&KImrPXH@rTs1i6|Q`n^J-yG)VRcZUvrYy
z-_CP8QZrL>$zxEFJlQa|c4yDZijPBG*ZGGPCr!-IJiff})2-aJ;=B94{d-}4II1{j
z3b)#S{UdJAn1*+cy!PvjD)zlKc!Fthk<Q-%9dljQH=enfw=yfJynjnXN3rXT=KS%o
z{mOqlG$^lC*CuY_*uKkJ%wA8lZ-4w?joav#DMND7&#En0k`$kd^$I%7M%}YM7{sY%
zftng0*(Xw}EU=_0fAnsg^PLv3c!dtihFNNSpk<(=vLL)=8%*o!>B`hWs{}s4GQd$;
z5Z*f1I6i4-d#)JG#LbWBJHkL&5ZYzYqBmV1Mw@G)37B>lyBjDALd!obX*u~Zq(Tdg
z#k8B)O;1@6+IxTd*1Tiu_GzJUn6?+&>naOE2d#ehcFv){Vzf{!rp?4=5@kW?QJO2)
zJ{@<iL<>#CbU$%FV`V|;s70^4u7CJaE=JQZ9o9F@SXmHy>h;#-lZS%uX`yILTZyeC
z%7V~|SrP6Z3NO9ULX$D=B6iVH7KENPHRH}slhseOPz0tO#18t(g3$BSrEMDGA9iY?
zNKDJcay?~1=tXba$DMNjRbL0KfFIQOu^MA_l?CDR+Xpm+^k095Q(Gxa>MkVOg*5*A
zi_|Ig3x=sTd*ijx7+l3pY^SR%2wfJrx8~!sqtmre9H!01<`QK==uI~tYTVcvSD~#I
zk7*BM4_jqH=&dWCI!@PVt<yrYF>NWflqd^AZwu<M_L#fmo(`Ih*NmT~pS7|e^v<LI
zw(WlZVXqdNgx_`)JL)M5LhqV1-1zb(TVEZt91nhs#u!6oLFoM-zeSIq8ahe`{e&AI
zr8ml4SrEEb^k-7r(2ICtX5hwc#5VfMg3!&A*GYc1i7L=Rb8+L}OTIT%7KA?0G+=<&
z^vj>LPz-*1h<J#pvLN)yNXZ;|Mzp3D3di&S@c?aQLFl%`oAUDWl94(n0e8*E(#Kj^
z5W3@OsCH)mrAM^TRQ&cp@jz{5LFmp68tI>s^{(ikTw&J3CE-TOg7Ei4ueNwyYi-m*
zQ}F9fVkcc?p`sr;O>c9oc&LTOW7=8ltgS2v{m}NO7oVq8#p$3NVIv1w201DV6<>UO
zo2^#STrKn?eqAP(X($Uqe|~Yl)U7MdP7BS!w29cHud*QY7ig7#unH;BL5a9=UrS$W
zWkKlF#2E{!tg#x2m+--3PL2SMe-D!PU_j)F{6?DpP$!PRfVd0Df-HD6-}zymd{L1)
z#~()4hf#k*aC{xIu48dMvaY9CeUR3^ZEf+z5{@qdzaj`GALRIAa4&|TES^#faTKqG
zVQb+#ic4WwDT~WsL>c%~z7YmDf+xjIFt~}uCCIk~1rc6+?bM5x&b6A@A{)VVBa9@3
z_Dr+kvkzFD1`*#2*1cd$2)$;jq`~V$+Qq^fvW+m15FB3tvI-V&0ofK7SAwjP#oIx)
zoyAoktD=}34Sh6~CjwD=j?V?ZTnM9|Jp(RhU<k!mL2(rVD6T`&I%GwrO1f*{Ijw|h
zz#}e41Im%g^d|Aw!v54AnSqHXAbA27qy>|+4z*7M7W?4Io2)z8nEaXJ=j+UuQk1Hb
z%Af^?3(P3DP`Xgn9LFz}E@e=LG=o8z(o6>Z0#?63Khlmx=$*D_-I{Ym9KRO4*RnP!
z1@BT8mq9=oj3X~{d<hCJK@$ijr>aO(d6-huW`kcg+e{lM+E`N_M$*I7l-4>cy$Uqi
z9$|~U0>f8eL)s?|s%!q^H0Jn7i%2I@iR1Sm(H^8rCdNjyWL}P`!3HGa6Z{iYD^$eK
zQp_4bUa~!wI_u>M?_9hrUxCgm&{q-nQ#bXs`g#%fWiv9}jLZqq-)yVi&Rc`#ibVV>
zuvi6lgb?vLV3EV(Jh06JRsTeM0oWJ708*d$64+>hBSEx<;}z>cQ3XW0IsPV?-UJ8o
z8pju+h(a`#VB&M&W(0?WX(m&V%d~O>OmEQ6aeO|C$VZb&D+;}{l?@e(d+|~@g)~kf
z9b!d}KaDg_v-k|sIK$$zNaHNU3ahu5=Zb%Pt0BV5WKaY~<O>3}E285NVWl!CX8B?m
zR!o&e#P@(f56ky}a}V9l0`JoUezg5wFzBWAiL<4t`41-xn^&O-pvFxyNm4B{4$q){
z4*dkjmx4tp-7kV7`BXp%`6S0zf=?xjYr(LVZgh@sf|w@wk=D;a5jkiQ!M@@K>3XZT
zKjAgj1x8(9Nr=Gc&rdckXgQ0CAznkGNpFdRh_`{@9nz_`li67i2Z`gIWKI_3a1B15
z^O}D-V;ff2SiM+AYhrC<RVx#79<hwXFOw|OC0)T=9i%0oTK@QeLHfnWgz&^G#V4KM
zlF~jYA{FtZ;~_t|uU6j?UWdo?7#ttNF!J`8;bketuElFHVdH5NK%AX17k(Xg%k__o
zc<Vj_okw6r>WlFEI*+L^Oc-n6^sLmok3PryNPKjO!@}{t5?@_%h;zK3#7~zvaU35Y
z3D6~q#qoiXKwWY)aD0#?NSC-493L%-W(*M{iD3{g;Te=9Nn%j4B$+`ek`xA|OVSy%
z7bJUu!M}s#cNTYmq=UtsAnBwyN`06`NJQ;IY`$J_>ZOMk2?p>=5h4LPE0H52zbD4z
z_`OK97a5YRPC^7cl|p<#Uvk!WHoE^`cPw__zQzNQUNf;!G$F?(<V@b?_-5qT%;ICn
z@feF+kYfwQ2^vMaR}9)(fcw4%*{(q@q+28u!QBwz5ZVoeg~EZXC=!t10SeIwQz=Lh
zK`L>49*W3AQ%K9RQx~L#f8OhhH$&QhG!OD;j;{vGYH*~u28P#wFU9p>S<mv1K=cT-
zDgPKmk13uT8S8fAUF;h?vUsC-)y^S75*BJ9LSnjH0E%^tQvU{~zp+YPfFcU0QqR*V
zI~UtCsT*;4Z#X;x1?|Q&uzv<#1TWZs<9JzqoS-hn$gLQu#!JF6Y&#(oBWp$QVIAIs
zY92Vg1?jdRV=}>MmbYZ@{_|Z6{%`|WHn6gI6fBQYoHjbJ%_U}71pcsxzsCd;doBLk
z@8XKIeIxMu+rf1^j3OQ7_$qL%qBy<v(UYa3>sRsnognQ5mEW`Ub#<kY<oXlrE*}Kh
zLHb3GuK`&Ni|asENAXY1=j^79*AKxhmZAP-XfXNoir;>1&QD)%hlx92dI#(YkyZVD
zQdH&jfAz&yfPpQngM1$R&$F)31@OPX;)~#ak;PBoyC<v<)CKY`>KAkTJMer*2cP#x
zkcC%!u^H}>nVy+3k$f@<p!xsA#YKHjSLL5gU4gCYXz6H2Vo{EFk-C_XtyC1Ip8cxi
zZ%;fLA0r=gMkbAX%^8_A@-t^-(kQ^3kx8RKb4DhOg3PH*j?x&VOTyUVm0p_;%>OtC
z_dh-|KAPCGc-Mw|DVfhZbwqr$XS6E35b-f4F{&LR;(2>s<^PNLB>N-=CEF(tBF%~T
z6#Ep_EQ<Jat8_QYrMsuAHZSJ<(<3Rj7hLv2AjPd<)e6=W{|<wHr#dR)J791Ji#uU(
zC&k5`vprY!Wc-RpyBh*`!!*)R>B#Nh3_^Ss<7xR5xId|6bNm;u|3bxt9NXAU5t15=
zD##U{RLFP}o^exFf&D7hO<4;TYw2ni0x6%3tXr}Le73N-8syc~)8hC>h-qZqlx!4{
zP2H3-tvlSr_?O{$kk7&DIjHPTk}g<Cgp3RAyMxTZf+#e{_mlOrAZ|i=m0wxhw^YH!
zk9Lbz`2p)c#ZAl7wYa4;m)W1xJc79feF2rfPQ(Ns4nh`(ep0aq6o*wJy6N5yr@N>2
zea4@u0M`n--AJYfcUegD(Dw)E4=^I>5RM<HKhTJzH^`3A(o|-BNG!tf+2EQDax&vW
z94X6X9i<WoC;?U6A)GQ3s$n9nUkm=VbSYxT8(dC-E9I|%;tF_ET!W-F$b=|?kTMu{
z7<p1$hlbTryshiAN6H~Mj7{(kB=4wKy8WedM61as!MSV(%Vz44SMj4oVdHWiW3Ov9
z7_6qQv=Fin@1-Zpj&px~4w(yYaSJa%`vrB!cD#`ec=8|nuUMcPU~mJw#E!z?qby#7
zeAghAafoc;X%+HkbXpV%ie$12cB%Vqx@Yowr!w*iUPU!ih`fSxSeyr=^FZb32$9#A
z0vJn7NiuSH5eVrxdH|M#aXHOQ2p)}YGcqF<-jyFy@pREi3q0&C;JSq_Dx!V(WEb*x
zbcIZCn82)$eMo;F&HZrv0pxH1$w})6(_8vPB<%>st()kW7?T85&48|`DR0LOR2SkM
zs~&1=0T1ng;bc$(9@_&E1QTBxcS^__Qj@w#-E2tE$nkQi+=h529Pci5w;^$9?Sqkv
zD@?;@;XT<1UX9e8LRjO~MDgLNcRh9#>l)ySMW9_oV`vU{NxO&Qx=u6c{Bsv#@DRH|
z)CGp5uSc9DQERVOJ;#JhPiFM+zqz}n;PvT;e7o=B32#C!P1IH*JTorMEIx)@j<L7}
zxwKGBvZQ#zg%l~B@GPX0g;Zge;L_?g(R>71cz9I`sZ(mg+hBN`>g~xPkqLc*cD%sj
zx9n%>N+bp!fTQBW$q<RQ;4TPxR@#<AQ=v~jBxp)_Hr0sMz}E@fpJ#F`MSK?6XVI-H
z!WGT4>7Ev_JR97}*907v4WlW>%9P92&joodTOVh@a@qR%V4ly`&j)!vTR$Jd^4a=@
zU|z`9F9dlZTfY#(3fcN4;9dgEd5)E`gr+e?IO?2S0xHj4gvE1WHK^PP5nl`PTDE<x
zq_tFRMSML>tY_<=1lN<m_y|~OPq7$h{Lav{sffP>-k0dHBjPWE_hlAiWxm2<oYrci
zVlLuw#_%eO+aaKx;?`dxlC^_1vawKBgMKwNoZ!j2)X)H#!voQ;r~JR3Z(33LY}j?2
zuPOq|BI>zZHcx*nN!RIvi3ecw0G!Fx6K79DQ<=P{;*<rpcu|pR!qSuBMMZHw$nYYd
zxDaG`k|{0$H#~_HUjhX-8pYK}T1}6uYxWuWStG~m;!Uy|%vRHoUI<-%dTCPdde_;M
zV?H;#?Sv7ID55GWB#Z_Y(*sh#(Zvu$z9_;;OpmqHI}zdRjYlbq%fPFQ#!DjXCP_Df
zDkmbsu7b3J#alqSg~gR1tz_|bkZxyj6-cXC+yJHxAfx?j0C@wom<Y?NyphF6!SpCw
z|0u|hvV0TBn^@corp;{q7I12T{`7n2VZ?dXeivZG1r}d~5f@p!23f43YAC{Sx%V2R
za=q}LMBXJVE=NA)i1D=7BcJsw-hcu(&@*1dS0b-Uq>5Js{Cy=FL;H)pnZwA79Ec)(
z2jW@B;(Fv+Pw{Q5mOj6`*tFq8qY_zHBGte<j8FA0^|4usQxpy0)j)SC#~%f+qb%+Q
z&u(_ju0g?Tkm@Rcc!2l_6>hf}&306{Qq9hX@O-M-I6n*Fg%o$@JiT-B-+hGKS1ZzJ
zMI0F)xgW!=3wLAmQ2GQqpXjs-9`M*Nph_BY{4u0^jCv&`(}x2H;SP=I<UGX^+R0>-
zWF~+}F-c)i9`w(nTFhZ@$+>{;8yqu(MHxNr1WA(Dv%Xv|ipZtm0=`T@!`ISvKr%=8
z<P&c3m<}2$%xDL5Av~8_jRcqYvPQVtqpFn;e)#AiYeq1(A3j<rZU+UHT8itDw4R>m
z#IwLBx^N{(`(iJ%Hzyh6r|x||^wl*f!ZVg;m8M!&UCRR>pST&AAgniVt)v^L>*~G}
zyEDfK@sY<wk603-ywEsUf)uqvZqurtl{?9<y_j<-%FEFIJQl(PlLRK)GRtI^3cXl&
z=wqcz`!)@Z=k<7%L~(qQUJ`?n^^zHsqL;#;e_l=qscrUh0;k6QaXBGOI(s?sYC%im
zfpn)wnjF66Fj&RvehwJqus9D)@~B%XG$bov(^oB^-9lB7<4=I$33{M${CTiD&t{#f
zk!dxuCA);<4<XY-EUrPOH59+CTRvXn?jUn)<!t1ijZ_ZT`*jl!<c)opf{FVeyHD3I
z@vQNIDBSzfHK6zl6g{dC5?>UoMv66vu22OQudOV8h7`|P`~@k#vbaE9QK-(m|5aV_
zO@&F?1&5Ww{W4wQx!{VOGrEHF!4*4S6c>UkcCIKcff(#8QCthI*g2xO9wuOCh~ksr
zf}J6XFM|R*HxwU2(nIvz#=9L2!@^7E<7>gYHLu+|!two11DNAPAbC-Q#nl`q0x50=
za~xn&+yc%xmqzh<@Wr__iZ6gK&ZSX&5qxnjjp8-P8fS_L?m2zUag%rYmj81fsih|J
z`2HXNf5+763$6$ci-kh@qt3658WK{Pp|@EtBXjAJr1>fHl2aEhN?V-1Ky5y$q^35H
zJ|@w}h4gU|eaxhf>GUy$J}#h-$@Fn4eN3f~Y4kCJJ}#k;(~}Z-{J)gc!V<q%7xfYM
z)sSd$+IqS=A^L&-j{2G*1}26^rZP{j@4QEh^zn6*x+v_2^|${2{ngmq!qUdp&Y_>v
h0Ox^n_d$b)4!0U&=IRk3H4YA4hJvE}M!{&c{{?t^kCy-d

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_fp16.cpp.164525B4D87B1687.idx b/.cache/clangd/index/grouped_gemm_xdl_fp16.cpp.164525B4D87B1687.idx
new file mode 100755
index 0000000000000000000000000000000000000000..795190ee066bd4c6001cf00a312483ba39fe5abd
GIT binary patch
literal 3386
zcmZuz3pkY98Xo`fWBx&-8D`9+alenD;zzb3x8#y*<P!gkOD<_7q^)$LBH`$ykl5wu
zD3o*|g>IspqKkI9>{6YhcDqvTE;wuE>}96sIrBX8KHpmJ`>*wV|N7oF<L~L_X01XZ
zxd=QnlZ6|i<w+zG4Sdr!h!T9kz($os%B;8)9O3&6r{6v@R9Boxp*oyVS35R;!Tl~j
zIX!K?hZj5>>^DBx7-oMtr)Z#MXV+qLbB~oVBg9DdW+zc4{{TPsL-^`qt1&@V<a4#C
zqGerf-IoX66?I-4$1fe6>{peKs_HX3vt6NdsOMeo=<CTPe*|8eyb(eF&MdUd$>|<_
z`3A3i8t!|qGilRs9@qP_Z`{9Mi&AYbbX-`j7_z@ze)D`)+be(jd#i5brefvB1MRJ5
zqI$k_<E8iEk59Bl_ANSwiFPeKV!J#poK=6gAhLPpqr=%l4r33pRUR5Cw`qnwE4<Gr
z-%2ZAF}`B2UBdjjJE5AFy+jj!rU6%bTvC;We@=-o_{~rlxTb2uzUJnnedX4<(eqf1
z4MX|kb-q!Tn>-qxZs0~Q5_ZWOINa1axVdF0W;yGF8AH%xeap4hVG-;8zF(BKE+`4Z
zX-)kZ)ybps!49j3ngZO_Mwx-GkH)qq3YzY$)2UFcs~;Sz_~MstNbvIF^sxbcbVF!J
z>${}djG}m9eo&W=cJ`R1=zzt}0Nnufe_RbbU$cAcl~T~HY`kiG&xY$Yjb;C$u+>sI
zR*#xa)t9m>sJz`f9_clotYsfGE!gQ|dpb-ol=IrLcjvv;xq7}S%HF1A69uglY{zw0
z#R7uSMP6%FGc}EVJeHpC<lrFeJiY3~N&3bkU)~(6wuy~?eXFS~)vzkXC3|mH@!Z3<
zSw8y~MbcXowA5;>DtGqSXpe0A(6eVr(8*#e_Sl}Q+bmof4!rv1_wc)M2eZD`>F5<Z
z2cHeQsK-r+*n94|m68v`XHXb;m&rLk_4~83vuW<m2kxo|UNvnW;V);ZJ00=9k$Hog
zzQyZxQPA}lk8dedTdcFu6y>(dTbwjy+$mCzog18|{`Dx;>P#W|kV(cJ_S={*M=C{u
zc_x`NEg6H|{q$moBHFhF(P}*t(}GFvmHf<^v@rbF>imIiC;sD<_^nWTI<w>_f&J}?
z=}_MX>0QONC!wLu{g+CIX-CYuy@I^TscyRYZOluZs{#t9(krRFYmaSoSRNO)TYjpY
z>e`z!mb!2l%L|EFI(Kl#h_!cAEV^Z;Dy`qJ`&aeQe<s9b#iv`>94nbr>8-o2m@Dvd
z%Q+YG+`;WxW&9J*nEs0Dmfs$%3~^jA_%=eB;#&UXQIP2H^U*2Imv85n%na>M><h5W
zz_R0`8_(WPcj9SgZW^EPbn>USRixLSOc=<@Txq*Fk+`*bxWjO0njH2gxxw^&J=>{w
zK#e~VebM{g=N9+F$kUvGviB`f?=^ZA&mFtD#!R&1)LaMWuLY&@b&edzX&=s^>J~~O
z&#PoF{ai~%>%h3bYgzh4Mr%;meEWwv6$bLXSB*Q`yxZ5fe=~}G931@KdHIvFhEMcW
zGtAq!ceQ4`9h_RfIX917`s%6Q6sl6hE2$6usoj<FJwMz1Bg#NC7UuOj*=PDI3rn`h
z6>95xlhU%1V@@C>l39BC+ZU5wYzzq`kq92jQ<ng^c_bgp&sl99pFGMj#0SvksQDZT
zKzi|s{;Ir-rS%xH63{xRj+z7@?OOQ$-P%*%S7As5pqVIBSptxDr{wM*UwO3%LxKUV
zhw5oa0Ma2ZbiK?>w=Yp4MFE<x!Pk)hq{Huar*xchdx{|;faahaWeGr9xXq70xu@qX
zhJ*py1T|5W0Hou?vme%K7Cy%iKS1lEx-<zuI+e`mQ$izaF~lFx7O2G>2|zl(-{h=^
zaakvoln!V|c}FD)K)QITn_T$dbPR@g1DcC+)g%Dv692l!Pp{4q7!nC+7RpkV0HmvO
z#|;V9?$a0&4QMvXR+a#yYtKE?dHG?o4nsl#t&i%@kpQHd<Uc0&ExHNnUIS=tRGTIN
zNMAZ{Vq!<!`ivo~0Bwz0GbI4&4u9pfmf0Z`4DkiD5o)9&0Z2a{_f*+zyr~64!U1iH
znyN?u(!-v&+wAUiw_?a@KpUV2Y7&^GM~#R+!}3uK2>`S)YOEpwNdGx8T+Dm1Mi)b3
z0Ih*)C`tg*B4KP=H~1xl9g9$Y=u|3!m&xHr<zRuZbAe3Uwr~|7+l=p;Rlx=ZDslJ9
zsO)hUxYUw2RkfonPqxZY2}d(W7kK!vyA>!ubTyUWvbY@B0#rhWrNe>SPzgPj9tZY4
zl`vo#aNtXWvmh0*6=YKJ$75hB-nJ#{G%O1CgH*I}Z5Mdh&??|ru}@^4JuZhg&sGa9
z17u>qNM*(}WBTx4rvr@xL}J%SMK}l7hj~Ez0Fl@`G9p*p4~n3PfJp2h8PRSzymU6J
zjb6+8%vRa>ftvWv{6JVs?$76UbdHW)5Kjm5Bj9x2AFdp`{sJv9&78y3fQJANMUL`=
z6Za}!Ihx-k+KNyK{dxMj&^uJZgkhouD}`qT@)UbjCQnNohs8kC0+HCUQjso8mjho1
zv@V(QgJr3N5zUAWy+tKVX{K~&C72ReSL|i!y0`#OftxpEwkh}}*SLYAHB>b;pi|)S
z%Ts>vIF&GGm~){4sDuT>f(y@rO7IvwF7yywAG|g3RRA~dSm-ZMaIKjD+p!JU7O*N<
z07zAQGe9-46a}ZjrC?n^B)%VHM25Hp6v3VcBJnLDBTB`yW<`HBBoK-33K`K{TnmcE
zSM%uWDIrXds<oZ951gJ*YA8bNo5Vih7;yNo@!^Y8qWo~RfC^^h@T~}Z8jSZNc(Lrb
z8o`gsjwcZOwe0vif=|hg3&?mOS!TVjWc=Igm?RP=fx8|2yB85@z2YEF4FCHGU_c_J
zNW*o~FkTu8rD38pOp}I5vmu%AT6~=E|Ehgvjwo$?sxUq_E+H`~Ic42i0V#G?MOrHj
z*Gt0$X_zbxMba=;8WPEZByl;*g)7MN3aFxzGDcOIqo(RXTjp#?qj;!m&ZRS1xShjd
z$Df>*@EB(HdN#)M|KF?zn=_BAqpNRVWMXP=$+udt$X0uymW8zoL&ME;1L7LA)LqVt
F^j~C>Tc!X2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_fp32.cpp.F97F080B232E6F17.idx b/.cache/clangd/index/grouped_gemm_xdl_fp32.cpp.F97F080B232E6F17.idx
new file mode 100755
index 0000000000000000000000000000000000000000..18ce02b5d578091f99071e6681dbefb48695a0f6
GIT binary patch
literal 3384
zcmZve3piBiAIE174(5zI(=Zd`GRECd=^$GOkz1%-Q;x}9aw%4H)3%9~q_jyFMM?cr
zq*852Q7$WO7sawAqJ%VUh4!8K?`x*#*?FG%KEL;TzUMvf`#ZniJLBW(?97!TksQNa
zQ{#Ba>m*4ek`nkPB_~FCf`OGhiIiG=b5)4fOI)?>*t7Ed7z)+)va-T-u|4J}m6Mti
zJg8qa{50VCiR8>_&GIO0SDOD*9j84Rc^U7=r<+q6{2PkelH#21W)%#i<uA^(Q<i1q
z&3JMaqm2(-ZHoCFe96#~w(`q2rMjNAjF<GiP4*fKZfeS_Gx*!0gt|4ev2XC5-qD_5
zude#Hf89{9|26ZsqSKcj)b~t&d2KW|*nUD|EMtD4#GD(|YU2<0x~G1;9aE=wh@v}r
z-?-!S`RI|RVBhPkh-DwrN|Wyv=3j3alKcASE#Gt3RN`y85;FE*Oh2)p*FPj*mY(0@
z*Ir#zz~4AroHG7(NxYgHFB?yz%GcL*oIkhp#i<-hYeBn()xkCkUwv++?#sdjF;!EO
zZ${?O_S78V>s6My7ab3m4v;ADvI=rHBr?zLFbQzJ7MWbGU)r6oN=|V0!8i#2vwXE&
z@$KjcW5$)?AtmO{3K`oi3AxnApXIW;`sWvRADgFD(H`D9Ft#InorZqSf^+w+O3r^A
zRXk6va>MTsoGGXFOVa%Znh!`CS*Trz48v_43MVvwiMXb7v}!1ZUEOfkj_E?Rs*E0f
zg>rT-*U{!LjU5?6%P2(u?=_XQfGdpjj>Hp_&V9`V>D%-Rw;w>{YRL{SsqKfWDE;Ib
zdS~X=+%FGp?i@RM^wQb@lRs8b>ddG3f!>VE+)YtWeQ*`8+OnJVH=6gf9a+S=r(sU%
zCgn<Z7PU8o$K-9e@j2G%c*W!9qM8KPuzF30%7a6vMuR&eiqgK6YwPAatf~pTuFHxF
z*<1U{LdH|g^AXRllg=!g9{sDZCdtL&V5hR*9itne^W5mli;so4kNsjlJah1=*14Rv
znnG*|qc_&<6_O?!7`W1?Sc~1NUn$R?bW{&p&yb$)J6gQ3w`pRr?Z6p%-?5pf_h)t=
zGt}9vsmqUO&@8B(ZW&s!O=)~)usY$2y;8Z<%0IR~<}>W3-soktCja_`bI&f|iL8rr
zLIkhrhfkhAx@@^XHD77Uv!w53jfPF4DjpX6Cn=U!RgIr#g|h}|nxoEjyBg#BBNsiB
zby*!|@B1k0Pp(I31iJNGNm7S?>2u|Pb5SvA;VG77((E^kgQ0XS$^EA4U*Ba{y{CIz
zIqAyTW0#cuFgxg8$j4Ui-!`)iRC$TSp_aJ)J3oaMYyF`2?XvchqQNzzH2&`mz0v1J
zGK0;=Chyl4WS%tGWIf^&)OG#C*va_R$9E6Z-pKH{G%Z~{Chfm6!FOS1e!14#TaC#_
zPVKIC89{osbe*244w}$>kY0Pbd97|@)&&(?hc7z{B+KVA?J`;N9^Qew)DE#*oU((u
zx4%oi=C^{^Gkp7s+lb>}ci!c!wTJ0VJD=adHdSw3Ulzw>o1eN9(h}{p{bOF|i`dBY
zPpf*g?l~@E7A<}M)9~iV)Wm%Enu&h9iawVYwiTPtaIcnH2hViY7q}S|Okuvab)1sY
z;=(Er5=niAoUoE#KO#p8C6Nd&jjJsbaP9~{#wjW9Ur#P)V2B^UW~iC6P#`w{lva0(
zzw9uEcmiyKnkWhd;zbo5CA+T|oW+n80BfV#3POR{DR*K#`ofeyhJ*l2N9nRcf!Kwz
zec!-}J9!wg3SeDSmnIa5gF74JZ(VTi!I0GeGf}3jP$1^*@a9ftx4g%YK!6QVLwTV<
z92uPcq;yX1D-7`lSO?Wn5(>l#5^Ak7=$cXt@d4NrHB}S}#6Nc!)+{q9tf!Jv0Jf8~
zlMxEU`S#A_;72oI7~%mi3uP$?1>*fa<^SUk)e;!824HnmT~;U%muw%<C*&J1VaPgw
z87M<mC=i#{zSJK4Fj<Zv0RZcvdWu4U_@d;;xYh;DVBc#2)<iXxgaYx+tA>Ww#Qjeg
z;t#MTYDpIg#J7B8qs`M-Q!vB}U=GTW6AHxL1Fmvg41TP}kYIp~P$M~^K-}xvR%hMb
zcm+d(0A{0X1)<F1J`T~Uf4C1rd;vB<4djFZ@!uo8`DU-z>R?D1z%-O5Efk0odFzuJ
z!S5ifScLLjrBVsJjEq;2;Ra#l0-K<0;Vi&bNXAdh+F%6(o1lEfY+LZ3W?it5flE-u
zVlErp5k7IryOPqr>KEI{RKm{K&JorKta38tKdT!|DJX5RDShxTIHkwJ{TpbeIs3qr
znz*JTe7CUl!M%dI7u{>1WT4vnZ<b*L0F$8lMJAky>%j-WHUK6;4~UtZaBpaW%>Yb-
zP7pKgA>#$JN!ql0X{)inl{Z){*MaK?-_!Ptt6BAZ{nrGSS8+q&<=taWOzVyu4RD#R
zNT<Op!K{I73EB!iYp?X}6F)a3ZbPVq-dsH$*aK9;P|Z*S?i98iSf`-p;Jx6Lo8vg#
z3~WA>@?FzW36?sG31^@Z+UnX&_|~a}j=ByLrix1Fs_Qc0;Y20a>TD(~eAv1Yl<!)W
zN^q1os<5}HgprbwDr_aV6|iVQFN-ZY9FK>Ku1u|y@{T*{43b5Yr_o@iz<Q9ReAnYt
z!bHu41uL6Mn5vnw;5(!eW@=_E*c@<vkZr+H03Uv9zE8N6)5#I=I0l<x3U>uJ0G28^
z8K4{Fmx3$7so=hVNpL=hne=f}XoC3yCc!BoW=g=*W=;PVGB63w3Ne!kt^rL0L1wBO
zDXZyVsg~B3p78RBOl2Nob5YPZOg57X8wMs$hVtFf0(S6?jBi8W-C(>A!Slt(#}WL5
z_;>`t--?f?5PVvEoFjqfN{G$(MFRgiJ0>OaV!_!C{@W)KNe=`;gh2lP35NlR6fX)l
zh{8xw$eRr%h?j<;XAj6tb}A+&ZA{=ru8)X{iH(cj5FJj6nRSt3MPZUCTrUcvMd3zK
z7$pkhL}8*ROb~@cTzITtJ8nF82}vnbT1FP5$|)+yFH>6Tpsz$(raVVQm9CCk+b**E
zVet|(HDeoHD+8_nx=Ld(=d!eQ^w=CjBNKD3#k>UzHRo%XS~{xHoL!R<r<L|D<mIG)
E0BXW?GXMYp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_int8.cpp.01E8A20D5ECCA353.idx b/.cache/clangd/index/grouped_gemm_xdl_int8.cpp.01E8A20D5ECCA353.idx
new file mode 100755
index 0000000000000000000000000000000000000000..86b63ef023adb0d545bd491243607d8fc1898106
GIT binary patch
literal 3264
zcmZuy3pkW%8y?><G2cv?%4CQ+G={;L!so1$oF#`GLfT;_IW|tkYD4vxkaehja#&1i
zbta<UrM6Tev2}IWW+h8(vXauWTj_si{(a4KUH@E{`+DB*zTf9PeE0qFb9ZrRCE;*R
zLie<IQHnqnhr?08Uvf%Pv>O<hEyUr{D%u0P5{7LwANTdtuQTy}ST>EbO`EXd<Q1JP
zlC1SNIK0w1#;w~k^Rq*DW2Mbl2YVQ|Z^M%nXZ2%!S4Egc^a#t$9Yp=PL&kl9FTB%k
z9o&Dt<U#z!Oh=Euhu_@J9(;FukAq|I%BreLeRZ?mMXLM&uLS!3!Cf}5zkDjV(Raw#
zy#2baVf)z2Gw8bFPo-WZW2uvgdv<4*OeZkX(b-knaarzRrNu)xr;gfR%`v&mE*IZ_
zEOlRo`Ul>bDytQG2bSbUHE)hKZT3m^nEd6-T%|{^Y0G%~!M4jej)`?|@4PN<>dZ+o
zwm3DLe-!nPJwkq=_N<wIvbRGx!rM1<lotB#l6%Zp*@|O=bL4UXr#ULvRq8IWj+h#`
z*WyXd!O9o1&($7{<vXJ^a!H?+t`}F7z->1Q_Mlx1ILoQ)E*{hud#q2q=lJJUswGzO
zlX0xa@b1?;)C$6nAJFUlH6$(aIrZk4_=KN+Z`pfMtoh5vOlR|_*s%0fYIIDe)weS_
z#V3NhctZ`#=2)@6><jJk=VlSTuHGNlEI97Nt=vDnh2vglIvbK?&@hGTq_j_s=Z0Lp
z(=}MUHpg*Exp8ZDSD3C%%8IjTC*E}SHspIYYXv6>m;KE4FD$MWv}9d8f5+B8kmXQK
zWl1Lze1f#|W-sR@kCM?ksheQt*oy&T%=zD?rdVjw*^V)*rk2uSTEmt~k#NgIns;dR
zE{4v-vhkGRsq@y`&b{3nNTgk+m>uqWcJok-hIdS#)}Di9lFXQg2v;bS)Hx!>Pps0M
z^zl1I3F!f<TFrL^Cni^iM_aVqa;BKBbN=z>X|g1TIdyqVcwo1>-^Z3)b3UTDQ~9NM
z>AUR~W}|mKfA(uUI`^6#!HGD=ucmivTKI=n*Ea<vPQ)2~$$PK;xaNIp#(?eEU{vni
zt^dmTOlt_3zP5&T-bUNE|MQ8LPaYW!u}?RC>XYu3817x|9$4?8x&L0lk7tZrTy8zd
zDgJJvBGr4S>`@}qgjqXKm@%ha`&5@pY`A{5b^zN}E_HQe?h<FNk@Ej&K^Js92)Akd
zEm=px=|3L*&dOr;<4lD;N5^65-4JK{dw5ALlXtRuGI;fa-eh*DWlEct(^GSui4Bs#
z#aBqbo++zMprfU|d9MacC5MdgFRl;G&eqKSKz<f-eLA7-(*7qO!b9Vq=G|Vt>2%tf
zv8dejyDh9%X~o-lR|KV2ST*G94qRY8c(fOvxi3J&x_LP0ufH@#L%p!BbgzsiZqJ1E
z77y2t2K0tKVoxIJ3d{4ARm~FiMH>yVwByFG>yS@I{mE^ewTxAp#P;@$l~(`uFg~)C
z{wz>Xv2*Lk^y8YQdbMdO<!LtxlK4HAymL9pOtGtZcTv=He@kq*ysF)QrZbwpiR4us
z<@8~Jr9<0Vj)&bBO2FF{Z*)f%Y~Im--_-2$tA5|Ex`H{i3se?4IXyli4#D9}Qd1{K
zr#*F1WHSyYFd>*MmIL_W$bbBq^tNk%x|N}bH=wx$uBIFy?Qo{A>d=2m>QTfO(E0@Z
zMRI_&D>1up#P`<sC=v*027y791Ej+qbR=B6;PM<rf&fh;&@|)#Y0*BPmD9OhlPD4b
zXbyq1P!5pZ5ti{(vh2u96!8Hxoj|9^0n%c;R*yO%T!JEgfHo!=YsvxAIepw_H=|=W
zNVsG`+o{^A%K_3w>s(aA9?eCd$VNcx5%d<x0n%03Bb>;E9gQfm4ba*IZ4Eg<TGBGi
z{Qdp(X%yK4XeNQFDF;YjQvDF$gWUxO3kS3=L6;&2NVm6gxwetrGbj=aXg-0jEeA+n
z^V5joX9N*ZWD}s-1U6X?kd}_PllK^X+k_%vfHou;lH~yD0r&fDw!Ix!P$U%4ECOqh
z9Omgk!^j>^*&vGe1KNOKK$ZifKQP>AKWx`2CgZjOnn&d^<pAl~@qr?<(QjBN@(rM=
z1nL4gKsrgZJ-HD455p&A*xPoDu2CTg=rlS5zDQ_Ag!q*ONdoMc3Ra^6uLW8X$Yg$m
z&4FwWc5q$=oe5MjTPmsSFelh*?L?Jiu&F-_Y{bsQ&Ix(~+7`@}c~)tz9~J@UZnQ4m
zN#*4ig1Ndjx=!$&LQjLaGD9oo8c+<D4*YYwqK)g?vl~uDsFDPFG(9>r8%dx~)2G9o
zlLQPJgAVsX60m42I(&OD8ekP!E)=U^%P}_GA-q>0lBGjQ<c#@15sV2C$ugoOI;?_~
z%wOj&&ka2$!B#%tw7m7a&Cmzg-?bjNG5Fh6+2cUoR`?iRqzdt?P~|Q7sV3)UQWioI
zF!h)$m=cnJqs7sIM}n6DHYAIk(uVk$4cs-%AD;MC?nnYQg}oF8gd{Me7%qi*fUAMN
zvOp^J6=Dgn@42)#HJ|ue7jPcxLMj#d8NOUq;#WZ>3AkEZ2J|CIV60`#fcHrfm}r?W
zprv7da3Qih!^N*z`U%yXYsbNI+6LM@cq(`RuvJ;Sp&B@xh*4lGcrGB4g<MI*!FW&v
z{Q^X?s4I!YSo*x^A6EjAEbvOAC72Erjf9#lO(X_sgRSyy`QGsHxca&85xYyW6wnN4
zE1~<L5kO1XM?$IPDHW_?-c|5^)rntyC%{3!s9;$LTm*~@5v)jgT#aBg%Hwea`$KvB
z3Bf)qkMr@^5xi2rzwy||`EimcP9WQ2a#CdSJy{`=!T){{u)yKs6lJ`k+^#4i6=k%d
zj8T+{ic&aV;sy1kdp~~m9N(=ePTZLgxm_rVijIkmi{BA}+cB@g?NpReiZWJFiWOyo
zqKu3e#>w;^A}_qE8exID21+7pE?VeD*|45NA-XMIwq&U`&BoT=!EyDPwPso-b}Oum
t=>P92Rd=}_gUMnW7;=qG`72jhVAlGUIy`eHEvk!q65<@N&Q-+|_g__eMY{k1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/grouped_gemm_xdl_splitk_fp16.cpp.5CCCCCFEB9008060.idx b/.cache/clangd/index/grouped_gemm_xdl_splitk_fp16.cpp.5CCCCCFEB9008060.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24606faa17ef91853c3b83444f64de772564007a
GIT binary patch
literal 4056
zcmZu!2~?BU5>9?XSQ3^f2?WC`3Hu@_{(w|u$qGsm6p%&bCm^ze2nuaQMM2q=MQNY<
z1l*`np%SSmE>sak5lgFLT@VdgYeB&U1=^dOcawO|;~eJ9+;8s8+&g!^A4EJh`?diN
z7ZA_OOO>o&L%`v16!4q9J}W5_7;*YIT;6}qhd)SrfExbR_m_By<I+1R9e(BosbwLi
zb9uWtg-6)4*6bYEQ|&ta$+tS*Cvz|+rB!D&_Zc}+)*BMQDGe&Ea}sYqS{fycmfe?)
z>r~v!n@Dt@^fZYL*)w$d+swd^6Y-1pTbS*;)wi@l+jxYt?WjY-<G(eA9^|=C&1mi#
z&8>?6Cf_53IDTJZvCU@E`+M8*A1+N#$liO?qqn~GxO%t$n56gV<!U42@}qzBERU{R
zaDQ8Jvvld%h<&Z@<*r6dLF;#6=Z+obW5<h{#G4AIZZK=2Bht=j=`Bp8<UipLmHf&~
zs-V6MO9+0WX>Tn3-E__)$JVG--2FRlUX8LDwepZYTf$xUvTkn8nv;4_t8Lp7QO+Q*
zA*%J`*wDBwZSw+uj?oI?^|rnB7+tTbVw;|{+WkyztdY~<(z{LKDv^YjVX%SkpLxUg
zS#<aO1zRpUB=d(q{JB=6G`5zYGm`3jt|hH|OYYc_>h645!TmaN+g}F@dvk8Kdv1#Q
zZm?y=XkkMmYem+s0!H!TYo^ktE8d3bW@t+PkmYbWTF0tyS&cbWHHv6)&Oh%lT=cXn
zD1Ay=U*0@K_n(Dp;lzuJ+}!Gtw{Di6-g)%pn%9gQ?O{G!n4AroA?})XIwl*s$J%}y
z^N=v_XsoqrnEvKQPEpIX#T{tkqf@t=^r#ySUU{9!JNZuZeOp3uTFzK=(fMJo1C*$N
zNOtXS<s+xD0e!y!uk8I&%HyE%9-U>9!=dw>GeZh`R(!Qype-5+FQ#uOpK0Qesw;Pw
z)!x!HI+^-WFVE>*&Ck!5Q^E#JZ_OVK=)9AVUb4$NV$06kMF!6G#fQdznoQY}$Q6BP
z+g9+xT^h=Ex->YHKys-)Cb%eSt)6*n6UU0Has98w^-2$MRE9Xob$Iap?HS{dd1nJg
zGXFW3d#9`0sIG%d|Lhz0{R_z|_rVOy`u2b3^_)E3{|j%aWb$C!wEwcHys8vYO*@}=
zA-J{a+y2?hsaNbu@0ki)E6)G)h<C2i+a;bEINv0@Z!}<**6g4|gtVr$8CT96dEf6>
z?3<q#VWzL&v9;mNGe1VmI5VW*G))Vg<ZLdUxv=TBNB@gvKYPc)VG}yFEQzW4b>Tp;
zPoJ=`|K96mr7=6wTWv$U`6i9xi6}#}T;~NG@nd7ZJ;wbfBRRQ3qU%iYaPEy?|4Hfl
z{ZrD5j;+3>XZ8;M$-<m!Kfc?!ck4HecENtwY_ZU%=IObn;nTd-Kb~x$mdHNeKK1xg
zQ{cX{?_76<w@WP_<Ck`eLtb|t_FdPJ_hISpclw8}B|BL>Jb`R6>zTz0+h^Y65sd89
z|Esca;pNc$lEySlSQ)`;??2zNzgfz6D0|r1a`H0zxaj?~lU1a5{F|GpPw_t2edqn!
z+hKjH=;Hb7cANw6md(SLrOT#bAEqA-n2H-C$QFfN@b796B-S^M|0pa%pO%X{{fn-C
zed}oja<w2V>v+QjJj-|fbzJtw)VL-DhjYxy8I?^4=8%vu91e3Ky3A4tz(I;1*AE*z
zI^T2aNXSwEI}x1>6#~MGn)(jxI9qj`ga`p_MYJ+d2nYw2y?wK;c`Aa0!~ocYXriMK
z5a#F>Rz4U0Qb9si0C+BOE=?gI9NpiQ*4fN{NJ7E^%pfv!6avDracm#cZsVJJxK#jl
zqq$ir1cW79Lfs~I_KcE{l>oLU+UqL>gx5yr-)%51dqzS+0c=gQrYHo2Gc>5Z+Qirf
z5+VYyGtt>lAs}4bXWz!PtGPtR<p9{1;H#|=5Z=9njgKCfi6bE)0JbDr8Yl#W_lS=E
z`2NKijD*Ajm`<eYC<KHL6h3ER`dzIgWDS7l5a;M91cV#T3|oziPaGv7kpQ+K+88PX
zgijIPrS>kk2F|@2z~)4Aib6p6e7n897k2#v35fu(JJH=lAt2l-(pl%4AFfM6mI2t7
zXsf3X5FUKa(<`vsbee=j1K5G+pr;TJe#E=k;dQG^N<yLl%p@`m6ha9<vBi2>bx%l$
z7{GQ!J3WPf@W)q=cDu+{Ta%DD0Mm#xEroz^mLwrNLJNn(51MiwTAT>Nld%|b3<I_T
z$#i(4SxnfaLB{wNd^@;I#sn4uH7nR6SZNvF`Nc0LAa^r}B%lh+;TE%Ae2i>bRsfdy
zRK9ZAVmmD^wV`i2SYAw6YyiFC03o_x3=SDXYw+kHJiKo>NWe`#MetC-tpMGw^zt^m
zyuYn3cLp3HTsvGJ<{)Dc+7W72q;{mzf+Gk5%jZumus0e2v%^8u)cs-r$(SYGk^x^U
z8MC5WF<?A0HkUq^0Y{XKG3iVO>@QpeIptHUmQ#er!JHur_oUODODn;BnWN?b@Lev?
z`CHaMF<A|k{Rn=u;Su4mgOKt`SB12r*cm<g-@}0;51i!lu5v<;;L#T4Ev`C#T>68^
zD-qnHkR^10Dae?JB~r7*EV0spYyl$6sX{F>1GRxCN-SmN+n)Hl7L@Mp<sJo>A(eo=
zavo9J8-#{JCrBv3NzNu}P9=D>N?G~_L13?Agl8y7;pXoa1`oVrEQsONS858DO$<$F
zu$}N^c-=4PXP4H+L&cY~HX~%r#>~bV5(XKwr`nsslftfpg5(6HR*)--!m1%t0Vg?4
zshq6o)(m)V$X6P=U+@$ev!&P?K|&#84ipC?$R7uT1})r^H$DU%;!?TB@El;^Jo%`~
z5_k%f7CZ$gMb3GuQqXub4c6NjxsZ0f$I1`flgL113L`bW{~Ruk=5fG<_*49qltIP<
zZ311P(}}ze&Ct|(HrPwkr_mrA!FNm0{enlym=o2>5>7Q4bEZ05!XA?`7pjXTB!+f&
zO<uF#$UO}LCUg`k(-AO{qe#sXJBpRo&nv5R_F6740i_5Hgh~d2rv&Ym6F0nq&iSHv
z&7k^Mz<UmJjx%frym@e8avq1?fN^wD3JldN8E~z>qVpBl%cioGjp&zMPCvVNU<B06
zNBLgxiU^nh6{uOkXt2_Ptp(}i-vzaFEYumMdr<WJZs)LOD@eCednFmheNd~N_aV*j
z23Qswh?Q-FbOS=lza?rRGtiANB;*F*B>$qQIn72*q0<wquU~zPI<y)j2&M)rFZ9Xq
zX20E(QxCwh+X}bUaN1sJpQu2*Psy2!!DP5WYI>dBA^joGb_hILwh7w`7Ei`FCLA@3
zYr<7p@DXY2e(@;-J($9y)5@s_b{T?}Bd}ax??=!B>h=o+eW`9wBIqY|yF>%ssiC(2
zlLq=(X;0$2Tqn%A72y4Is2pXN!DV;`D+eq~k^=tT0RMYtVc9q2O1xYsel)PrdBRmb
z+36F8S=pHxlC=qmNy#axY3b|YacR&Ahf7qcYgMX5r6#M?6qTB-Qqxsxf=XSdQZrR*
zl1fcgsaYyDL#1M=@hS4^_>y1^f+kT*TZcr}Gc?fWQUd*16kYBt<Jm?gbkxgdk?&WF
z{amPy-g7<eEdJjrZ4SfC(#qO~X>0G`<mxuhW5Gi6`KHe90aO~Bw;l;vzJ!Ap;Qj@e
CLU>02

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_bwd_data.hpp.3714A9EE70464716.idx b/.cache/clangd/index/groupnorm_bwd_data.hpp.3714A9EE70464716.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e329fd2d4983420b281ffc3d9b9e937f3cb55caf
GIT binary patch
literal 1234
zcmWIYbaOk!!oc91;#rZKT9OE4G6({3aY<3;N=62T237`!iaC2jVzX}>h+Osm=xiO;
zvHVzEyvqq^PR-gIVG;7X)C-mGswiz*;<>5W{MY^GM+zsIDfVpB=6L<<MbWvP;qp%&
zz44#lsv)^Lz~wdj>m4Sq!aMJ{+DjxZWvSb|k6~?{bI7$jI!`&2c#c(G^X{u%SM6yh
z!#COAfBKV0`zOXvYK$-tyU6l-XX%UUEIlVpwUgCvxZ8(cc7D5Iy3M;o(HdMg{?@q~
zlrQ@w_>@oj?b7XG9;-j!&28?TY%Q!L(U=?|8bA4C?YxQ6J+g8f9w(!>*F4L8>GOE1
z|1{nk86Ps|6>dE%x<+aF>$gRczn>qu+~@z+KgVi<)rWutm-%r^!}mU6crAZ$lX65I
z%kN8Hw`-=aGBNUcc}L9SIim_oXkhyE7nTwiN&_0XqAjEo7ERQ@Abll7a-nx;h|a8!
zmGACY3+#QPHFxKvZM}DW?!5V)y?d_wg8h+~18y7bbM;@{$sYgyb=d{)zi&C;uCkAL
zc=?+}?T*Ev*?jM>8cSbuef7LV%|&YMRmS97=T~euyJYlZ?w!iA%RIN%h(@f@`}~^o
z*7gY@VV<AFQ{4WDGZa_mCe;CBj!8@Y^j4<7jY3T9+zbLb0y;8af`J1l2L!AV1{EDL
zmS)^cEI?@$E)^j#0h4~J@prz{+WucW3>;h#4FbZf!UhbSOpI*8oQwjDA9!|42)9i-
z&c?*b#VjDrF0BeCVET(v(~3oz5Uyb40^1JL1+$%r3v4S)m`6ZQMNbFjay8a%Q)IGb
zLzs931la{eA<kf6;1MujH-Pg({DjnBOtfJIDiYxo;fK3hK#x~X7-mu`)1l-2w+n=s
zfr=DFtPqN1lw@oW=BkLQ=)+vrvbes=cg4};tULmSqK0xXlYl&9LmLFoNYn^P&dAKj
z9wBEeY7CcyS;@o+4#(a@zZa}|Xsy5_V9H}^4%70~fb|PIzj!yuFQOvyFg?Ifk<?R$
zDT0L+FhoJ24m0W3w&@4;tWR%d<blKk)K`B6!ZzK_Kk|$VXr#4>9l}Tv2N7o^o~wul
z63<J-7m4RD5`^FxY8&dooB|9C4?}MR&q&Y65Q!IJ6pr8-Ya7G$F%+fdB!V1RT$EfK
r3}S$a8Ytj{F@Q0RU<&kX-+AvSA7dh$kRT5qFFyk>LXIE76hbfo_3)xz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_bwd_fp32.cpp.94FC313C790AEF66.idx b/.cache/clangd/index/groupnorm_bwd_fp32.cpp.94FC313C790AEF66.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1308e7be6001b9ad1456fa17dc0167e06b51c26b
GIT binary patch
literal 6438
zcmZvg2Ut|c7J&CG*A;dzz3)<`3AliuG-X*Z0g-0Ia>WoqG%6%C0cp}ZE@gS|p%ldu
z1tl1XC00Zc8yY}F)I?E0EbuT*jhc6N^Ic~AzPI1UeDiZ=&fI(E%sFT7Qa>LbBYhTY
z!Fs>s=*Wb11dGK|r~l#;;x;+cuveMIO0K=UVs)^Yo7%YNv0$?+gkItX3!{^cny*;u
z9fkL-u1Bv-S`&N!tVrYg^BLFEZ<{-)^0KQNcQjx7_`JEp*Jq(mp^zJW%KSmm^OLj2
zJtrnhYa2dXwQC6qI={BH_ug+``A%-~uNNj--q4x;*SzAr9fp^kFTH>6Y~9jrAGAHK
zy<IY)8x<Sx5n)yGc2Y~bG`-G#VCeqUS1;d<RMv$yj<|K)J$9_y;oIsXSu5)O$L|z7
zCo1oV?fAXZHaPEy;KjC<EMdF;?KzS~{!OK~mt5rapHJJr;?$F`oo0j@uYbB@OIcX$
z+8iZUwpobRoLgRhnY(XX`z|c&e(s_gN2Nd8*G-OdUu>?|vEIv`+|4r(_{T}5O2lZ^
z<h7kaLQ&eA_eJL6uNn=NH;LmkrOyKUcVG7{5f+;ig*BVl%-Hj2$)64y{OQ@#=lHAb
zoM#kMH&NysuCee*RCI@e!2`1^yEk9Z7M+}#y*0yOU4mZVk;KTM1rJXLoHeK~pA~5S
zVAZih(ZB8MFcebyCUcWkNtV1-b@gBIz}l`_+$^p*?RX&Wx7^Lc9+vf5+%|LTyTx1b
zcu6ny!dut%iobKw(?4LYK10%X+;5Xf(;OYm!7;wwu<IZ7TA^cJk1yn`F|DjzxX4u5
zMbA5d^F+F@^R>wj%YPbcJNE9H`(aNvck}kjwvaW#cj=WUMvM73dj@_m-7+3f+_KkW
zV7`6k)1xylR6N_Jlw+D=zH_ifa4#~gYox!FwRP{&2=`(Bo65u9gF<V+^ib=KGq=5&
zm#seFz2b+N>52nwEkB%i8mm@oWB5%Ip{~s7{V4gp^2p1~5}R$_g>AeEcOU)sPJ5kS
zdG8O*`2Aapmf+K+PX~)Fc5;3;9}O=`FKTOR3UU4Udg0B5nl_hw9S+x&sT^5)DuRF3
z|Nh_4%7?}sEb5Gc{;*8`Px$bTWtTH<b*2CDeHA(G=E#0GSzg`U$=TJk;Vt!X<H<1i
zNm!(Pv;E^|a{_&injSpXXlZOUy~<;!?E>37iWPNhUt8R0-Tr}ft-k%jrT6TV@MBK<
zbQNlUz3|O4-P#84M|*$Ei`ZT^bSw7F-Hz6}Ub~!X4F?xd%*)28)K?ZUPA<(8UZR67
z;pqyFT-TLb6FY7?3MRLYXKL#W8%xh^Z3%mJ;QXe8t9laDS6zuxc=$Nry5-V)oCVr(
zV^?Oq4cfK8uCKCOZ=U}$%R7JH*cVV8;dVyB_;&e@!;c$|R2Rfv+cJOR3|k||YnT<>
zo6vLAs+Nf3-(0iywa@zS#Z;cMwsxUK$Hp$fUt`T*EQo$MQkEXlKmA|hGoudgOI1x;
z>-WnoP22W;<J;mJ-gtQqesgNmM1JwVydT}~*sPir@aN?5);(u=#*Z@YdtNsFeAk83
zYJ0-6{`+iQXWQ$?a~J=|_so?SE5&+gzq+Ocs47antyWWuSsIxfwqfo{%fED=Qxk>p
z+oHp(fW=a~Exu9WVtJa!Ud>`rwj5h+c|*q+`OomCV=A|R7sg{Rqs76TU_*Jsh;3UB
z#Z(Pk$>6brXt4puKwI80Vh7zmy(i;wt4Y=tTD*bSpeSz`v2SI1ltII*IvzWO7VB~J
z^yCdAUcNul?VMfnB#*tA7VC5Lwd4&W-dwoKz;E%WM47dM78|P=^W_aAP9?faKDg!G
zQ)aEC#YQSd8uEq_r`MO3HDA8)md9R7i?uo0YVw8=AJHt!*q&Z;k;h&_iv=8kn!I7e
z*Rys$5T=%0;jvfIVjYf-s=Q&ugBgqbJP)<o^Vq?(n9t#B${R-f#G}AlKj=d#kG-50
z>vD88<qacd#YJw2e?f<Mh1tAA4L8C|RY)omgqh$tWlt&#Y_h-}aW*(-!#u<};Fbd-
z#JS*^3!fp*17RMnFCX0U@$m)VT!8D7fQ<xSPXc}t2txYGVP-i9arP+(f5cL-k>c~E
z;3q}%)lL~jJ5?{dHI1ZFrl;`Hv{X*2KB6>knjxZerE~*C8DxerLyS+exu5tlN&OoP
z{|&R4VLIQ`4|cwXmq1d%qTqFmtLxpIopSY{SDyk&4TI(|U`vjG#t2}Gj)LkaOg$i@
zfL>O1P?+r`sWz}|0}p0d>jRM<HQSG^;Sp4fZp_qdf{N9N)nx_|R2)BU>V*g@-Z<V2
z$r6kcED<FN5={{$@sqH5$-2qd3O!)l12$-EFBtX0OvHU4=z|#yTX+2I>-;96i#CJ=
zJS5=F2uP|N=9PmOu@pR{NX|qBg8NA{ASnmJL6g}Wl3GSA(`0N$QV&7&5CR#McEX$!
zdPOWhdT$<q$s@FB%mCni5=tDO){B1`TA)Rbe*#)hV9Jp0Mw8dJaMn$72&xG*nm~(D
zAY-p4Sb%e269gmf0*x+wd>43i;p4j?unUcMpTByz`Ge>f{REw0)(OsxBF|TM##;{7
zC(_Pc1iT{9WCSEt4BTSSKwJVUC7^@26qHLr8*v$^m4P0^pB?56X-#Zj$tI~*&}_vg
zc7s7TOt}FQXSv)@;wwqn6ZSf2Kb8_pb<lpq^WrhjPv9kBly0By#XJQ`WjJPFktGL=
za=-yebHOwhoDkmw(LGqqaA4={C3l4X?NTJ@5T}xfn>!0svXHR|DjQU?(Y_K?4yfdy
zcP6M@P{~Dyk)ZNGB@fq^4=VY{cm!1dDg{X2@*oYXXBG3Oab(V}Bmq_C31!Z%RE~26
zC{^HG2}+eXSAkL$&Qeg4B4)B0hx<u3BdO)Yat&nuL|)?5!6m7DFv$ly^hiCR+JgtC
zK@bf>03-kMix8{5qwRC$2RBp#VT?fbF00ES6tNUSrAQv}pm!R7cCQib!~HO`AH}Rq
zaq)@4Z@v{P$}$m=3Cc)A77$rDX9JOqa}E$WIOhV9i*p_jc{t|-k&kl$5Cw>tq_4pJ
zB-xXc6XB$WTzMt2QjM`?Y_jN<!L^<bbiAzqVFfZHK~;jV66Y!qR^eO&;u@Ipbb{Iq
z;@vnO2bbgE%GiRSPJ_#7oX>#E8Jy38%UPVyfy+6ZAA|p6w2jHlnOB?n(F62u90#-G
zD55Z>4Q&KjU4s&hBjKosjIe@Op~x74DR-uGKPhuaY8Ej|1x;B@ELLHrkW>sW1`9qZ
z!W0+gRb_5qkO@;sNaio1eptxJGg2-uHczq{ryu<$c-%x$GW$qCbV%J#S4fiD1HOA;
z1v7)B8o{R#LJ;qTMSEd2;(g$^4^|=G4*~mO4dMgfdjQvW9z^F+LLsRO;Bf&Nm!#T3
z+>RcMq%OjOi?9^w>j3W#Sc3QxcwIt)t7MTv?kmaPt~`?3thgDA7F!gzU=*bog;BI(
zG)7-5evOd?v?L%v%awyxInEWJRe^IQXjS4|1zJ@&OF>JDxcd9qTT+I*?$aA`5cmgC
zsaE~xmus_C4erR3h_Dsh&=YinXE&<h_C)y;i<aaMu?eafG@9|qdKEOTB5s_}wto8D
z>>V8|ia@srua}BJyBM(Yt^^EAP`yD?r64E;6Z8aSU|NPd*)UiPqxyxUM!;+Y)h{G9
z3WlRd-@d|CHN#ua8OW+iLCRFwM^dT$R9m!2no*iPqIAP_EY4<_WXxt{2cmvidF$kZ
zRN90!FslYU83A)wpd-5MqJWO$Y+^Qcwvu_t*lEfGgFO7z`arc0WiOI?0HO!zJhlEJ
zN~+Nd{z@Kyoif3PS(~J90uEWIDk14KIX@eB>m2aR!QDC+KF`IEQ64zu;rjByGanyc
z0P_owzN4W7j*X?N8nnAKfpQb5qX+2%<u1f0eX5gN#G?ZW1T_ZkW3Y@FNKntg^*K67
z1oZ+OUw|(oKczBuD?es4i%wo+AQ%I4MnKXN1<!GQ0fHBZFOD9eIuHFw(JM;8P=bt0
z&{@y094r_`1SJJSDUx3q+nHJS%=stUvzuUf6N(`O)dkDD5O?WV9X@JvDO_G7I7q;q
z(I~4Kf+cvcDhIoA6xn5YH=rCTBIze^kmB>D5G+OWuL+Wel#*4ZaS3WCh<Ab?vk*b;
z0`V^77X(!c;#w5d2&xXmb?DF%R6U66Q4}Po1`s#k`c8n$2~^1w)JbqTiCmkYPJzoQ
zRM`{MkKpnnx=1qjU^=bHF2Tt4g@jNUrIjVb5@nQDVtKJsRXBYZXQj?%UQg!Df-JP9
zBy|@=cad*&o23?Ji<YU<-q#HLW-vk%uLA!n;-3Ec-jdM%Z2GKz2e#j#LS|6s{Kphe
zlN5PyGAx3ri<7Jv7ZjtKnWRd<tOTsk3Z-CLikyg~%D}1&usAXdw!`=pcLc0QP@I(o
zC$mwckNHGE2MyWx0c5!lB19e8Dv6ggRqvB@$!c7Hu0kZ$3#z?%e7z5%`)Ko-&kXbu
zvTqD%2|L109c@k=FAiHG%^_{-rYZ|2#@YCS10Whe3q0+gdur7oj~v>{w$8T88I>=p
zTK=fBz12Xc#uy@oi!2yR#9|aj#9<Uq#AB2|Bw&<CBx00ABw>_HBxBSAianr$R_q1E
zUYz?tu@CXPqSrgd?FNJCgWn43^c{+sNKoCtr|(b<e+X=Nme5ai60+>+aS)vs+0Dzc
z6|9+ovb$e!Cmue#!L=KGNcniE_V2<vQ6p`!Aut|7xs3ki6SYpZL-BDD@#kfo2wC&b
zZ#~mtvE*a^!7%jSO{n`N>VAc~p{TnCb-zH}4XC?zstc6E&(kkvR<7o%Zv1LP<fgDM
z*M&!{U%Mvs3s(5lAl63IjYQpu==D)D$)d;r1!5X!x}p+~R8~_}S*ZS*x1~CFp@y!e
zmY%+=+r0T6B2TfMwynFF^GxIa_g@`>fuV`%408)BYnxg24zr!+x){yj3mq3|>-hL>
NW`7>&<qQ6-{{dF@4fOy3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_bwd_gamma_beta.hpp.A4BF65E232956D79.idx b/.cache/clangd/index/groupnorm_bwd_gamma_beta.hpp.A4BF65E232956D79.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bb8653dea919f2a3b8eaa561e75c90998f73248b
GIT binary patch
literal 1244
zcmWIYbaT7H!oc91;#rZKT9OE4G6({3aY<3;W<~~vRjdpQ6?686#O7Z%5c%qVlG%Co
zl?SGc;;UadhnC(t%)lx4;N-5Y`O&NU-rn?E)$rrJitpuV>vB2{s%Y=DyT0%Fj~~r3
z`+nJ_D0&^!61n*6qM2UrF8ztSU$_UanR4L!HSq;k%{xxbI_Y-LA@Bf;a!*>Nna8=>
zd%TlsY~<9}*T+me{r_}<t6S392p8cs&#x`KAC~Bo=WYB=<mi{zB1$#uWceoPw}0R}
zcxTqD;3@1myXKZ&$>)~YHCOGDU|qrQ1ofBKS~%Pu@I+~68U3D-{)8>LpPeb_lb>09
zRkYRRj%S~HpPZgH{k*i<z7xjlz8SNuKJobU`{PF6jx)<&|NCU#(PKSV8rzwae|@pq
zE2qN0;i06Cb@S8o4cqr#Jv&RxOL_5>bvcbCq7F_0qNQcc^O_lB*+o}uXx+^COzNl9
zUPs@lg1e@89Nm?fA9n6<v(xfxmzHX0K7aS__}Pt_>ej!Rw8|`bx9UDF{;_3N^@p#9
z`TTP~z30z<fABC*e2Mz&qoP}{{du?OuhOwy9)G*^T0ItTwMy`5m;8FoEsCvPVebhw
z-^IS`4lLUHxITVCwSi~qR26$hy|X-i48@hXNp-+D6fNz1v}{FC1ve85H-mr*mx>UW
zVBi4C0fCosWZ$+rrU)S>cA&J5fQ}59fJtY6a1`!{-lD<7z`+I4ARx>tY{tOJ#K<Ph
z$tb}1<Ca55hG6JEHYQdsW&vq-X;m-*(_fUDRxHYda0L?=*mjsM9sxZSJsp@E)L6Go
zk;#?~Vd4=GWET{L*vG)YBVfR80O!FJGjf3)3NtChPe}d6L>pG1UJ*_aez>~@^mz4z
zVR~SqOq^g}^d9=XV9i5o1s(xY9#dPGqL#(=RlX~Z9%tndFcdYEg9!t9#)dWso{^{#
zlAMvXkqbi3SkxFUx6k<b2Nu0ZZ)Tvw6h!P{PT&!cQIc^)cu_@E#RL{rPYqbVu=9&|
zgZv^YB9G9bC#k0lvjygHU^sw69p;x`+om7Xvp&6<kp~hFP#gXVgl)Q;f8-e#&`1Xn
zXM~X=t|A^tJTDPnB%Z%W5E3s$Bpksr)Hc+EIRzLr9){iso{_PU84@qlC<4JV)-yJQ
z@feCya}q%gEG|ke4hAuJK?DSV${#4;gE4?<0m0-)Fa@T6|G3XHSZ5-ekRT5qFF%72
GLJk0T>z)$;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_fwd_sigmoid_mul_fp16.cpp.E1BDC41DAA5F65CC.idx b/.cache/clangd/index/groupnorm_fwd_sigmoid_mul_fp16.cpp.E1BDC41DAA5F65CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d6f38e12744e4f566b0748450bc21cdca099094a
GIT binary patch
literal 2446
zcmZux3pkW%8~$dD!F=N|&c@lq7-mRPPQy4G$3hNMv1o)?nNf~Qn|~?Cznr$R`ehXn
z)+W{|vzv3#wz~=~Q%Pa1({8H8|9>;r^|I^w=ep+ReedUa-sf;X<HK}z&XRy2hZD@C
zSaw1v27(|0^5rIQqH;ta=$<$PC6(Vf5)?P-EHgVAMaYZnZMdfVI8&NJAWoiN>{*mE
znNZJx7kR%|lJ0kZE6W~t%ek2qGH~U?*=~z6Dujaqo=#L^{<xv?aQ>Hm5|ux{S{R+g
zO<!RL9?$!DODs#;_>UiF_&t%Fjjf*3DV$G|m?quH16kMVUi3corlxgi`WJOf9zJ8E
zBK<L@=Jx0-rr8J1*53B6j)sxO@l_Ik?e%L#<+6NJzXK>;Y>WQU);%GWPF+$4#ou2U
z6D^OvJv&32<6JE+t!}m_GaHvPx^DUV7N&QdY@-|T2`#ha>jzGd+)q^BwtH2aFmQW)
z<jwl|4a);VBhAe@71`^7R&jf+D0MoRCuYNAykavpY1yLL&Z5ICH*v`({apF8zPTPW
zjd}O*MJZ0HgAMm={pEi(`5oS`RonNq6_I&|{Q0F&rhQfaKzgOf@t1mxg|FgQE*AKg
zhx=tcBRO7oOtAS)tD&IIWd2aLjHlARCC-M?;1$nvW8*FHl%CSW!br9RKBc06>V1A^
zrfA9+J`&ENb#G5n3ai`)YGm404jDd?<uocp_ZEnmmnk1V*81}}EufD!D8V!R_7qo@
z@u>b))~B&Fx9`dv&JjzdRIEi;;Rv70MLy+O(<qbV;WKV~tu?%8-gl>`Z&b!typg^|
zmQkOnYyXDxeCgx6s!Bbp)=#5CFHX(Q--}vmI%me)|BG3&E9&H35g)$enq^hVWWiGV
zhnJ3feu|#@@=iR{JlpAEtM0+tI>Yt67DfiS`Fy}Yt)zPZ+W5{i|JqE@h2Tfl_dIX9
zmgbIkuCHBI{Z%eU&m}A;4c2T`QRvAH@$Pd-pVdAS;{I;56@PQ$qJ=|8XJ}5?Na|oz
z6x&ZtO@WV#4F6%nqtBE(XlhWRSxkAKw(0Hd)~$Zl|DAV=3w`U$)fsGZ*nOoYgWz@D
zDg}<SUqxqQ#8h=@*^u9#K3j3^86<;{NQHyb=DLPO2}fEtf2xR_yfO1yze^6AzM#Hd
zFjU*ju3~BDz&S}@5Hwe*#*Oh_q~PD0a{V{OEE1Q<FV~O75B~YJt9NBXWF;m)_QF7&
zZ<>AZ0C`#Oiq&Q2<C1;7>J!QFg>bvx-SCdK!qqwcb?+x3sZlkQ#<keLJglt@qZ=0L
zsnhj~AEYcMPqlT$d2Y(deYWs#oAJz`M3>YJD{4{zE^|n+L3b<J{J+DEy|&hh%A(xl
z*suf?1fedG5rua;!RSK}#L~m+$qND5S>Z=zd|n}S;WZ6L`yn(Qi&ql@pbeX9PZvGw
zPlwUo2rY?~R1gB7jl!1`r60Gr!e~!~#$j=?LI8ARD1P`qvyKm7v_C?_SXf#JfR5uu
zOwGN0zJf&uAT&XYKoA0;(^F(Hv8u(vl8`S#i%N;g3IWg;Mky_>`XxL(bR3}>7>2kI
z06hq=rNup>smq{)10j~Yy1cFs06qNq)#tR-{C*gH1o0Nfic1Rt(68v(_9|YRc`*7g
zLQ7yJ<b(j|wfUjIvnCZ*l8_HVW2LY}Apn}gj^K78zlNaqDBM4Kj#~2YV(=X2P#Bhu
zjt&WAifvjC%#c1+zw27ErRKiA=;<yLo+Xcw$AdS(;Lx~#3=Ger(y6}SVK6?#NH9C#
zi)R_o4LpDm7$RaMnBa~Pg-&4tBOC3e&&p9%JxCU09kMmJwvVf!Y1n?ZJ7>S=!ER$$
zJ#2TCa%AMeo5Hp|y+{xSiD3;U2xNjJAkfH80)})$XYf*tPxXY>3Wm2UWK2;Q+3|&2
zA`((ZMn?_w5=aXPDbUqUNE)5y0z$^OjC;P*$oy(oNT55!S0K6_Un9Da9q`Rwc2ct#
z%_Cy*EOC-J1;`1{k|0S?fO_yO90^AO{l&9j5=;Tb;#o>GB~y?uo~24t-9CZFAi5Fg
z8f1|FwmII{ZlQ2jt!)wfhmN?R7s$1|q`VTa=8aXGLL=Q#yQ%{(qR_&90F43#AVmw7
z0R%y&O($>9;78ZO;oRn7)m<T<7C$cg;Zz0-NkEILwOz%F-~_(8{UvypHE0g<l3=fP
zUec%QyMgvD{&noV=e4YZNP88DirOG)<YcL6w{ud`P}1Ew5sAdVSrRRYf3sB8P~El!
z+lyo%SlyirjOoVPg6FVfLIn($6V4?ee??f-3EKx+Z#=J}gN{N$zzKqm?))1bdm`o}
za&%*Rh+ssqV&W1oUQ$L{%9Y?~uSdYS$|}jpE324WSXt9;Y#BxhhL)rQ`kMdmR#8<=
nT}xYMk1m;FKsDZLVrq6!V?U9$&p|=anRyKD<YniAa)<s0m#<yP

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_fwd_splitk_fp16.cpp.5CDEA132226AE994.idx b/.cache/clangd/index/groupnorm_fwd_splitk_fp16.cpp.5CDEA132226AE994.idx
new file mode 100755
index 0000000000000000000000000000000000000000..28672cfe3bf4d51777b3eaa3687550471fa4bafb
GIT binary patch
literal 1954
zcmZvb3oulB9LJAsJC<{H*S@fe73;Oa6+Kj|LaWUqwVo?^<SmtKrCJxIBCng2N~uH-
z-Gsc_!cBP$5*m@Dl5mr1yIoSIanJ6|KXYf=nf>nW`TV~B|9`*VbA;<19rKhS$RT9?
zzU`sAgYXapQLwjocWmU|c@UIFfS`TnOFTt82W-@OXSvO3TBQ;wK@eZry!2Rhk>Klk
z%^Xg;CI1F5exmeMf%!nD1k!rTfAgs2+i0#;5_&)PO0|hn{&hG#_RA2?T9Y_y(Ol!*
z7tq`^Sv0WPy*AOvEi7EasG_|lJT2jq%jff356qmQ*A#x9>~o8s{%&az_BAf&P1I7I
z-J(tOi`Oc)PK}K3ij7Nj_Zuu5^}atg6hn6T($;&@n{9ZYs!i*jk@qFLy2-Y&2fdk-
zHhmsut;{Fic&6s!(8F0L`~zoWol8bZgHw$$ac3K7Q5|meTXIz5^*Np@tg)issreNa
zb>45Q3xr|Whh1Y{)LwgXmz9%kWtwEX<)40^GZjTHQ2n~*h1;g0K96>!d_9gEVCF|&
z(mQ5k(rw>c)~=gjo^H@k>Z!&Mm@XR6brcQ@0t0<cbp35vmcP?ZOG@an**p+aR7h;m
z^l*(HN>11J86&HOxAgan`@y>7y`gy{GWX*MO(}E9z;j2k<R$I$DZT66yH;;#OKcVS
z6u6t53iK#T-9d3*`8+9V#8!jt#P_woQDtB9OoooplZgXIed~rE-7(-Un6}$~&I8_v
z1WPYTwj>KLJbvkL%t@D?#gj6uxvh3w(nFmjwfrHTVb2RT$$U@cz+hqh?W`H{4>lf&
z^o;YJnsFaYc-j+@+kA_7v-|AH!BrFUM#~M(wA?mO?jAlwTJx`V1S@=TY`|xSB^L(?
z9?SBK!%F>V7D0?B)kXSM`>*7czkMe$;Oskm^LcqCSJrh`ryjXLZYVzT@TwENJI3nX
z*W!R!<CXTC4BcF^t7Vb8n_%AbNc4))@F?L-Rme<bsAZp*8FNJ=QA~Hf;7oG<)AvXN
zA6}Yl<ngF8<b}1CPN7h0$F6gJ|LW67txi~WoFqS;l;|QAC-;=cF{5vv{=GFKS(`Fa
zSJU*M|Cwy#Q@%8J_T0y~wco6ZWtm<<qLZ#QuPClLBRmah#pR=atvH>Z9)HWlzt}N#
zA`h2k`8qgiSO%Y7Jal$Sd0o?|{nD_+spsYW$7!htvq+AW!b4Y^mEVe5SLwOErqpIm
z+P0cR6pc@?JFF!48G~wh#Ufk${@Upe`@WCuW^Hr&BIye;G?p#u3@E9Lbi4V^qvO5%
ztaMxHhYr5Dw{t2gWp*#K)4{9Isy|UkRUQ}b-@esT34)ZejST_{YWWCmBLs=4FqN(l
z?D<r@G}t_IG)zB(2njI^lVFxY0A}`&GbHav%n_UihEW*RRtUfg%Zg)ip0y+(I5!Lv
zVIo5z09yph;>cZ90tDxZVO3a_st|ykcN8p8|Djt>gxoPapERGU5P&0s(DqBCYa0=q
z7lsiSAu9yn)J|@dz~oc|3i)B!3~xqI2*B+J^In)sGFuRwCx!_yfvgaK2lt?K=Hmz(
z1m}%m5=>H42*AU(DfY|_KQa*9CJd{<Dl~-v48?|qiIs43=_;xIGF6=F|I$Q7T6irr
z&<MyD8<GbL)<#8|cuf`<GInHt^p^~GDxo5h14#gm4T8tK^3><N>Ueb;@PhcUA$j(5
zLkv9Q=P4|@^9v~b>@;ix>Jq9a@PckIue_K!FAL8ifFaMvjbne@F=)cPW*jpfXap38
zdF8Fmc{T7FMBq)4@wJzCX0Y(6h`=FmL0zawg`>g+@uMPDjw%=Y4^)J35H469?1n`z
zFCXjz7Q!}{mkutYA{w6db8`?QHYBfpZitCzlE4YUA+RC&4KO!E$I~fb2)h%agXBmY
z8!YBvpYufONfYH|R{nULk1yl{ToB|t_ceHXNOU;%(<#(t9)3QoOi)2kq8gbbpsckw
zqNoa}8Z^2l(`NZfp6#mDYb+QGS8$e^=>F=ex|X)i0zG{LLu0O~*>9GMtd}ffFJ_rr
TIWW{6*Kfx0H`uLHa)$l`OKqGG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/groupnorm_fwd_swish_fp16.cpp.FDEC9D120BB249C9.idx b/.cache/clangd/index/groupnorm_fwd_swish_fp16.cpp.FDEC9D120BB249C9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0fe62992ff090339e8933dcd4d546cc9d1577bdf
GIT binary patch
literal 1956
zcmZvb3pA8z7{@<zGna1$!%Su{5|)W1<T_Ig%`l-0ktwH++k{4GC^UmD8%dLp&{3<j
zPE%^6bX!~6E!Mh4B}6KfODSb*Y9ss2?0IwcoW67Bf4=kn{?GHg=l^^YBoqkBbs@+n
zT9}#;lN^D8AjkmuB_&JaQqT~TjfbGr@^c}gZI9gbyJy&q`%P3bD8o;FsnIpJ;FO<Q
zXk28O=g2?9k-t_Ao>=fmA%jew^Pjageg4$OW266sB%=5{>fPM<8e2DW2gj_=iMNq*
z*^cR$@bFdFr=6R=Z2lO%^KyjKuX5ST5p}z12fOM@e)D!O=BCW)+j%Ozp~#(?ZtU~o
zd%O}e{<}0!JMpITESnUy+nonwSq4_lqNp2ORFc577i&(gJ6nHYixkVezd<E;kPL+7
zCk5UqEVnq;QQ76RPB3CSihZ{*|9<?Ny9uk{loAH#MEBJ5fyHGFc1}UX728+#+1&GZ
z)g|L^HJKknZ|=b~waMn6JG%(Wu0*AFS*sL#JvRJ>d6FJoUn_n~-}YFZsweDm7->b5
zg)Y2YdTn<8ltoXikUF&<yJq((YUS2AtHlQ~xNK{Jf`V@POutl+2PrPrk$pY;Y8orz
zpW)5Jjc$2n{^;)2UKKj*`f2qQ-zpJhX>NEaKDN5I+jUpZ<Z=CTFi&yt%XaBk|MuUB
z$Gnx8$~}R}d8K`w#52X}@Kjq`96{Mltb5krTpkR&N=)LXvz><ut#}hzk6e{-1-((q
zQ{TY-!a4U_4z+lRyJ?!#P8?LoXCHY=S+=h|@i;YptDe8V+4xTNX;t;mSD%}@x4f^P
z@ml!vm*9bn2lr!-G@d6{ca|L<*!WRrGK|_9B`VAs+0N6QyNiCbk)h~cH?_KLvlr@t
z;$?2H*TbLu!}ca`2rB5S4{Gaq|2MbLfa4uH9@b}+8aqH4%pK+@#y_o5Qu!|Bt#8uD
zO8R>bu3kH8RpobOB2HzCW=%<5%_f}}E2q~><=0~zABNg9JV&xU_FH#Yw{)EejYurm
z-Bv-MO*%8wt%9c0YquWqecScnty+4+Ok9v6JC;Yu47iq*b^k&NLsDCk-W;1%z3=_w
zN0Cn@lGM{<C&OE20@Yq8r^-SS>(WCG2VNO82rPQRp<larGW(`y$+0~7`GAc|!M@i=
zQ3oBLM2$V0z?HZ>E%=y|w12Qmdgp7ZxJm32YK}G=^}Q6dr>ar+Ij?z%MgE{cz2XD6
z**5mn*lSjsldOR@pnp^;;$qVJKaF4gXZR9x%Zfi_52LMHW=e-up%0UV)uU_MUIouw
zh*71N=Dn;g8}Fvfj5@Khs|7K&GtfUZI02NDmJk`D13@|kHddQX)bmNG)et12;;2Si
zLH?oIhtWGmlf8PefP`9$U>FCRY6W0cK!H$zkGf5Uf)I?O;Fwwg*seyISk%*$K|-xU
zFd0XtX$4@1sEHK4yLEmfR3L(hI3iUm0LMnaEfp`9-zK3#5lq67^t1x-z7BSspKb9i
z7z#(QJ;old6@XiIAMIZtQ#6rKAqd9f@OoMSxbyNMUqiF@A~F<=U@Qeo)e67^J76P5
zd#pPNwGP1)97SI%01t6<ycsK}50Fr65KO=k47CC<B#jX#>7ZuQ)ggX2RTS}mX~H67
zu5mEf2*?&OX@Uj4VUY>fL<mea0@j<2vY`$gSVZxm_<?JK;E`HQ>a(?UE<FI$g7^`W
zCi__vjZ6DBV2MRddbv(z12VvHo*@U+hP2-v&cAAP2RXL4wC8|bfR!OFHT#-vIfpw(
z0BUn4_@*8m2bdUGgtx@A!J=Rh!IHoR$-*L{C6NuT2#ZLTBsS;{rXi_n)()nCM#yl@
zy1_AWajxOF;b8BGNwf4>6O+s2fg6Gt5tC;1vnB?Y;R{UQrie-N0nD0=xJE003HcHv
z5gL>vjY_iAyo71wzh^yofzTok`xIx)4a00y*m@``AzHE-VJl(+&=@RE7f&F;WPLq~
zpTTnP9}I|o)H#Mm#te54FAjIf(q#@bJ5NhjTl4Ro(oIaw=330NvbJH*w_oVE$a%3F
W%Z0hX$%jT42sfbkD|x;;{?NZ#1EcEz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/hip_check_error.hpp.B6401497C46CBC5E.idx b/.cache/clangd/index/hip_check_error.hpp.B6401497C46CBC5E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9a8aea2450cacff7de55e7a9af26fe41d524e499
GIT binary patch
literal 744
zcmWIYbaQ*a#K7R3;#rZKT9U}Zz`!5~#Kk2=nZJN^1|tJQ#hh4wU;o1fJZJqsGA{95
zWGemWlCqG{9gBi~qvNW9mv?#H>Erx(|D;6^|H<jkpMOu>yz|$MQpI!qmfqXA-p`NT
z{Ihrav(N_zCR9$D`j=P6c}c)R@0TobI}WIDoziS%Yg2h*a*E}W^ZTSEe*g3O7xaba
zUe{v`UAi#Ov^Le+g~!OK?a?0X!lRLWbJnh&zFdfVQqN|Qa*rJLbHeI9!b^`DUkaG&
z`gW0_t8HR#uBHL+wDV7I8=RlcF;leofcYXf*Pz7Qvg$KRUEU|{r(OTuZWXaJXv;r~
zf29IavwweEw|B2b){F9jIJI8o?Fav?P7Le3SHrmLXMII+Wp0upFa$Q;mJ%&Iax8*}
zfrEj8m6L~48cZ;7GBGl9vM{nTG8Cnz6@$V_pl#(%p4%$h8F^SaWjNJf8s*wHF1v4W
zQIe5`l~<Zq7A7pu*J#<vapoaVwV<$|0!-LS?&GO!w#1L@EUbLoe7rDWAWwx4ZX}Rr
z!)J$(<LBl_QYpYKfXtH@P(-K{FcL6<@jR}qmdgEcIU49NK|Vo2n6UT$=l1Pq=LG;A
z#;49_0TZ6L=X+UHLCqA9dsPJ0VZy6TC;g6|{o@rQ7po|Ts5mgdzyKx<6b1zuO!)ZA
zSq*bvm|X-~DIhE$3R89I%B4RGSZ8Pe&F1Fe7J~^tKenXcpo;k^pfE2luM|wUC^aV$
kWMFYoa<MXq!CINO_RfW)*IqKQu`{rPL?D0-#$aFo0LRn~@&Et;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/host_common_util.hpp.DD6CA1FC5D70D95B.idx b/.cache/clangd/index/host_common_util.hpp.DD6CA1FC5D70D95B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2e65e35f388af695b55398206647c1d2deca8b99
GIT binary patch
literal 2672
zcmaJ@2~-nD7@m+s+^~T#A&@MXs2DDhfPlsr3YJ3z69j^y#cB|cqER_KAc#h)Xz)JK
zhgwg3cyK8~u_)q!M+K``6ssb5;ZacP?Xw$jl012jeJ``~?f2!s=AYS(7#9?j$N<1U
zZrmbOe43I50I(!~sc9*{+@}DbpaZZd|A1oZtY&ws#<1)DlQupK3Z3|I`ZB>~r!Jw$
zL(M5nQ{_Ib4KA%Z(snwmw9G-1W@c7*;a&T#eW|{2UX|&)pWGz`?sM~M7j97$#}W&8
z)#=mPgw<@<;+l_J)gR0f&s-{N?+OJk7e=O?Yss!leQek4cEGY@(+a$(I&|uTp9XnH
zO>b|AS=(7Nj*?aMXm?BY&boTx=9a~7ydO4QFnwBzyJ7cp>pS9qFG$>;1j~ozIoWKh
zicP7@xVn6-{Elbin8KF(4|8s8G>L3&rj(ewDa||2x3SyH!e?RyhXSNaR)wv5?@*sJ
zX~yE;?k1^ZT`T1$!=^@QhQ3UlO%M4}hYecO=ttDk-jq5XT}&OjU77U!fTQKTDfK^d
z8kl0LHn=v*acPGvrtF~V>$c!alP=zUo7b`Kuvx^VV&Pa`L)++4*UB{&D{kiBH=VU5
zXP)mg)=bsFoeNFxh;j=mg74by$f8$k<4q41O<(^uNn7&YBv#j%?D+zkv&*e7otk;)
z$IBjDlogd-q!zPRSC6_?{y?fMESIkc+DDv~v`^@;U-O`>EW?M9Ya4HOr{-nYm$XxJ
z!&|4{F4XQmuJvd&m$Z*0zA_WJ@4{b~y2`JHiBdEfL+KxxVxQ&)keaTFJx8uLIJ__Y
zM8Wzj6O#evItTz{F<2ac^#H56huj<MUB=_U&H_L+NA1!*0IVMs@No*H%!v-fvd@MT
z;=q;^$bp>MJphbu0tNz7P;{WVX}!7T7!FvZfX!yxcMkxgQNSQ3DX4VmsL<{yyo>`L
zDX?eRJ9Q5LWAq1RItyP`o|#H5Em?pAf)t1bh`hQ7fMKRUpcACTC#1TX0zmx~`OKxY
z<{*Uztxc`Xk!K6;%^%%w+?!=WgK-XVjwoUpOjaaEA%qg8FnSXFRu=BxCH+4aoE%P?
z)KF|S6g#Bjg(;VMZ08=M;E?0YaYeoc`yQ`+EpA}+(!0N!9PA@zh3oWg951Bz@;S$7
zN}pmy#Ub0DJq|s_q1s$+g?!;Ki<ZSeXoU6}aJ!YBY=Xl$W}GET0Efw(WFA5ie&^*`
z{gj<_8YCcLh3t}O7$GtsKI9v5d{@4k5hvzL5DGfA{MGCi3A;(3UYM7Kenp@+=B-B|
zG0s=d)@)ru<}S1m5@=mz<+1ejyRJEsne$jY4l)^u1^%ImXRW2<(B8{lg48(lz&!LO
zaVRFl0}bvZYzfbR0<ZwRQXz+qPd7;E{ctE^i;P5`n5SNZLm6A9uMrLvOa&WN42Mdl
z(g-Cm6Z9x*n$wFC=cp<&Gw1%!3>0hhr#z3xRoPyBHrC8gpX0QkVjfw#k=&7XhUmxG
zjnQX@f-qt9?4EDkhr>j>L}x@Us%>e@pk4oyL8c*)4MC4_DCf!f2BhFAj8L>?w6%d#
z@{~qgEHf4wxCCB;5vSs)2m_}cp&pI;4~Ll}GWBRx^WHgyzI$YSc4KY`_1Bl{zKYGr
z$b~XjgI&2m?qEO)fx-wy+e911qZBBOxL9_qQFIA{1S2j9Ptq?O4pp|Q!3Ga%soLKt
z4{4?zt^4QmF6$`1FPVo#B=I&ZQ-CmFC_)=@*X*63-c0Q?+5DYDA7-ANKeErR*wyb~
z<=Q!O^1th-J+nh;A>k_?lK&+}5u?#wp+SBKe>_6yP99ad;}o6XB{5CI_pY{pWN-wN
zutX{x@|k=+I`KT?gpKxfVXyP4B52E^V}4JXNWz2Vq2E6h4?H(4$lPB+ZeAfKv__qe
z?o>w4S*`C^DNE{vs)fTq+dxl*s#fM;7tSl6eV6g6y2i;zN56U0E1=UUt*>G}XH4iO
zEHo5Lv~Cywx_;$Kd)Jvh>wC>!y>i(PkrujQj3EMziV?w|h)})MAu}S(Zfh^!O<nmB
ziFw+bK3&|Q2vk!<t3T*Ybc~~~aErL!|J@NTrI*ovFN+y1-cvS`dhuYf1FFmG{L0g!
z;z2`t>AS|x@4WKJeKkpNAcu#7e(#xF8+`17U$@@4F32F+7P@yNYKyP?hpm0Wx&4-O
z%Vlx($8SpfY^ClMpPDiwb((Hsts8nSE}D=Z1Nn|Sv{Ju(fvF9fV`a^?1Z)GT`(FHZ
OS)mC4Hijw3F#R82dD6uI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/host_tensor.cpp.BD60013DE8F91330.idx b/.cache/clangd/index/host_tensor.cpp.BD60013DE8F91330.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a83a8bdeeeba5e55bb40fa7bffe6d4e6d3fc10f2
GIT binary patch
literal 2800
zcmZux2~<;O7JdndA<2V05<*A_NeE#HF+>7{AP7_h1T0HH9-t`Tj$m0-qKH^6YPC!e
zL90>(fg+WNsK?5vAWBt4QE{dSmBoE2aA0Z`ThaNW=O>vtedpx8eD~h}{@-0Sp~1n=
zX$TUyG*p|hY;6h&K@bb@%UPSfYO4u?a7+<ITV5NR5&1}sbscVwJhsX?^S*1}<}KZ>
z{~p)K8z!v1yCh97?=?5|cTfJ|6~Ut=t?h2olx?iI`nz8RNkllYtX}6FP~6}&|Lwac
zb0wLclRk&~lNJu#INAAjVi=it7hIuWxZc?Bbd_?k_XT0>m-dIFZp2*tRHtxoV(d}M
zK6&flty8<7q$-z-I*)TY(>LUdt6M*8K9qD{`5!^ntnXj9i**^>R*lTDD^IM`^^2BW
zGQF_wdTqB#w@%#iovs*t>A?u&)GIYRi2juyvOd>kGf(#~IO`ue6uR=};+Fo(@lCU1
zjxtYQi$3SfOxX2db&amEtp3Q@@Yjdl#3rWAiPZ#e$UGYEXG8Z6zpfq1TUmbchaFGq
zA{YAwq75-K11f)>ZQc0zA1AjbwP?;vPbI9WKHpGvQA4`m9+MpIT)5(&0rh0+t&M9!
z`Yr}-*_Xa#a<})Ur`8wJiapEgL>C;qqRMSo+&vn|E&pP@>!ay^E#BHN4@qh0RjJg^
zM@G3ibhiD@znUlo;)Hm*=15dYLdTIxm-h9;#I(|!g<5C!OJb_7?Dj4;LmP2tL2E|V
z4x6rT!kH#p@e>(KxB>I8WPg^|zz%Bi_o38AZA46MyAR6iw{Fw$r<r}lSgO87<gH=s
zu=1HHc++r9p7r6uu;ux}iCCY~$@iL{-W=rJxPGtb%OVl$)NhvsYNxQTd`lgeBqtaC
z&LQ<}b=M`!{GWHEwKP9yTICrzkFh4IA$jD{>aT<Pp0HP6DY;xg<aO9Doxft)ej-D@
zZWG5>=v%*Q&VmqDTc)!AL(AeICl)OsBr;vxa;eVm2uVBE`6RV3f5p2ZUEj#OGt5EW
z<G$<cn7~)X^yY^z9uJ*PD?QA2NM^0fzq#n1B)4l?>X|5y&~4RcNnH^!o9%xHyINOW
zZP6GzIw*NP(nzY<`|Zf7l%LcktH#1)t&@ZM2AbL&%WpkoHhGj}9IEjpOA~r89<JVd
zC;Fgl?=L@=DGv<nVs~v=*<b1Vu!dpwO}dhuxIO+nWt8Lp{;|!U2d=2?824*>QacN6
zPuO^rHJ<x5x2jpusLg$Ku4qr&zdZBuhRc(x_SPg0xIB9n+L7{o-HR_=hhvW~s(kY@
zCMSRM-Ma;4n`iW|tBjCI-zF>=xqW%!yS&tVuVh8r`>ahJt)<=_baIE&{iE$?uKn8E
z-P737nw#2Nkh38pwHHhzKeXq@^Jla)j2JZ^!P#Q=jJRAZSAc*4)jye4V!w>xZMi6t
zh#<J!TyFL8^+|~!UsffI==vYmV8j@3nk8nr$F0ZK+riV@>B6d>&TV%vVkA&gY$&og
z4OPR1>Jx*0xd*HxS7O9ypyrGDo^e`CYY)}uiWn7Ju;Lbsr~zuJN*WNiQL@n$s;^l&
zH8i=jr((odpq8m*0daYfyeT!3y=-~Telmif>Y=+&_r2;RkZ^^Cf(0XjffHySqn^Cj
z<#^CnTP_)$`}WovOFIIJ(=j?5M&MZMb26GcSOA=0zSt}%mI}`nMpz|J5K$Zp#-gAY
z#kt<xKzM`n@<51cV(m5Hl55UoL9xq&+V$h@A?85LA#uR6{PfHej9*_pEoFQfpi5Cu
zAzgsv3WWi=QQVAl?i6<;<Vo=~LS7WFDHMFZa5Q~*`C%I18;314@MV!%bXZIjXOr1<
zNE48jaur4>lpAS4d@`R7%ZTCvs(@iYLNg)DfTSF$EvzVt%Q!MySirgaB*n7bk}LY$
zI$(~FR5+`I+td4JqCU6gm?abg<c@hm6dS>GdcMb(42-07sfE!6beRRbU}2R9b70}=
zXka8K%Tbs_fCO>}h>{o0cdCBY=BcleEms6XglYoue;PmgI43^7P4zplw#KY^&^qVG
z?(-!(lk-hboW`RGAWIZyg|i|J`4u_{r7$y4N~C?9AJ~T|uHY(M;O)>6Xjc6cL8G0C
zyl<(|{zHHT&5Xu^R{O=JHD!eRZlKCEOBxr7LFP;`Gb3b;*%}Zlh!qC8fE1W=&0tUf
zakzW~5(<S9$R5S*=ysR^NthCj0ZBDd9G0X?_&VgBV|+K@YHnt30WXHeATs>|fRUXI
zF4G<lcy^CS!j%%GGt2`CS4mV;sQJRNr<QvL9e_(rfQg|?cq7QDerZ6LD6X_pilC><
zb4Hh<f3y{vkZ?Q9&H~aP;S!F-+6cKwWDvo20)F+|#^6_FrJC}4GqW)kTU;ujpg5I8
zMIk3p2@=f+(J6F>L1#`fH_~A)n9`t=h$SwN0hk+&ED`c}$*a9)l9wa|;Zx11OnBrK
zFFi6PdTkcyb|pz^)a@#gY6`vkt>R#-q&OIyWl$MY!v$jk>?vf5QRlEwmdMZ@Y?M7U
zASlj9`62^DsYvPpD**V4WFC+&m^U|h5G(-L&*>@pm6wyfB*z|n3Ha5U!rXzeGf#=}
zaY)>!ZK!|88~!fISejl8V0aPXa}yGI8pV`KN6j#Xd8kEDprZvXbT*U2vb5%^y?uOV
z`S}O9vSl;vRnEe{lUmvEZAErsiG!2WWx8U9o4co%z=J1O&S6{qpSj-0-x-I5B@lxn
JKATIJ_g@sfFvb7?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/host_tensor.hpp.A2A2489B4F3A4E06.idx b/.cache/clangd/index/host_tensor.hpp.A2A2489B4F3A4E06.idx
new file mode 100755
index 0000000000000000000000000000000000000000..62d34570431cd0893ae85e131b3be636844c9241
GIT binary patch
literal 16564
zcmdsec~n$K)Aw}upv-V#W*CNH-(dtCPyq!2uPCCpMGQpau5lqz<N7K=qlf}ZL?R0A
z_~H`9C9a4Iq9RKal;Ez2dm>R$5lM`w`TBaNxYy@7XU=)g`Tk0dnVKJ6w{LfKbyanB
z&EV*$s6E{!lAhzE6Q)j>Jyt4_NL2LyS+i$;9YG|Lsd9-VVPnA;Q>Upzlz)Eq<uc2H
z!M|J@+<ewYYcg!6Hh-JW(slOe_SZvi1~pfXn>QSPv#=%OFHir5nO#l$th^TR@Nr|&
zcMdZXUN!`;i>z!}F#BbL$Cr^+6Q6%rI{WG3mp(JP4Y17n{7nzX`F;_zYzEei>0lFX
zvS4=2nlZzMcU3%y3Uz)zVdLneIpw309)4kU=f=(oyU3H1+RW7r>6DRZr;Yka-}5*1
z1APZg^YY2vU#z?6@}zmnn{Q8KhNfKHVzq71A`<?*cfsc;2D5J$9BsJR`4^XI2i{GJ
zeZMo*r8IoScISpAH3^3u&!}uuJ>JBj_1;Twwcn?4u6tj9F1O}>X7(KIsf3cF@iw88
z8+UEB-m$5`=Gzg^dJk7EKce4y_Mgvl&Bt`SrCQ)qv9GRZ%=(Q1=lY!stEyR_?Ed(t
zp(p2<jn`?b0uJT}tg~s!X)Jm_I8Pq^L27Q@_&-DDk2xs4@^_)7RXnR5J-0M@#+%%4
zcir6+mYDQYzhBo?@BXd*rYEs4hq1m_ufFQxQYG8E`1fy)&Az9*wsY?G*W>I4YhO<d
z9%){i-`{fP*9U#3J?*tB?Pl3)$G<mx>E-#0T^H{lMgL91wmrF`(A}StGA=9R&BD`m
zgKIY^9y}>inpO3k_Dj(C9_n0&<&(bJb3dV5@x+1CZ(T`z{p{V%^3quo^9%bYq@~J(
zoadHT*K|3)ZF~8ZyTAVWo#U?KMORNA@%$_Euenbr53;jsJG*aAj~<p4GrqMdJ@%D%
z;zE1v?-uPRtUs5n_WM0bQ+{Bw>~)(xZEVdhOjy^WYTAQKX5R07ELI1;e&^z!KKhrZ
zFD>7VJAco!?A`6;=9WSIW=vf-a;kI9u3KBzRxQ_b8*=LJjSeT4zTK6vpeg$5>6xKv
zYfCCRX9ipxo$!^%sK2xox&uMw726NlAAGEv_ENhwBj9m|zvp#W`g>|aL|#9WFJ6Y*
z%wOTRZJshPx^D4(>s=kMTP_(n^uVFZYx}j2oj<>8LP&6}q;hplXk|LyQyZxJd;F50
z-~R8i`jyLPeHJ$@LmrU6Qx@|^61-(%g?CJY$B<`j;{USh+$$k@d3e|`|9@Qk!U9hC
zyPx@Er(<W&U-z!~>5s%tm-^pM{^`Og?W_H^6^hKS-e{AecSK8OeK&O+Hltf=-Kn5S
z1-?&|SmQ5Y&k@-<qA(DNDQ%|z5!K9IPP4{lE3qSG*^QOEv6F$cvUD`d-&{3rxdrY=
zS#A@>ZDM60z+$#-c8K3^s$ELlm9i8Oha%!^AgwGl2hY8^e0&8X86=Uge;|`Tkh6hE
zWHKV@C_yYzF6jNVta@NyC3g3cuqjBHf^-G~{V0z)O%2HIQ?0}f^i>bA>;ZN#5MWWC
zKXhl-oh8Gxct|vz6O!&mCI$j5ni<5++jiJ7M$(y%Ccq@X;p2xyW{UZd*qTO6v>v}_
zv=Zy-YYt%90jx9-=)?B0WJ4BtUtxg<`byYiNO=rd83?dAL|?C}e=ProO4o~*H#=83
z*QvF6!z*;(+dEf@W9Ul?h<O3g7znU9g;uN0T<TXVu^nZp#j;vD4xRvu%lpzRtJ^;P
zPKiAz%MPsGfgKG5SiGvXm;Lo%%w{G2oU)V?$8zFoAgwI5q4zfHe+#n4?dfbzAn6HY
zYaqbV&MsHE^=ik-*0=-hB^^uCvD`p_#rvlT;p-M28qTvcBWW{|vpfM7{n@8W2K?sU
z-vSSyFDgWCg>;R10xbUF4_=)8!D|z*x0i|RGSyq2082+aLy?KMgemc0y2klPlaCw?
z1Xu!hELD$4x%GV;+=s4l9%`G1tPMmW`?wlZ<GMs|++4G<VgxVWQY<T_^5qHi6Y4AJ
zX;;`*&PdwPempEZ96o+XWL8Q^Yyj;e>~ieoR~MJZ@XNTJwB1gJ%M<7$;@hA<ExtZo
zro>KEbbGOKFLpN&VCfmvbZhm|#u{F9Nm!ahrOOjwi9G(i|4}AzG0&2Mr74t!C&1Ec
zYvYT_2@iuANiZF3gh_<!#}A3j(t=m)sCui1<C;vjYw;I!Bpa}F12!=b=%de#n!U$k
zg6!L1AKJ$u)b<c^Fc67MWllv$Cp}Q*DxEnh_A)<~JZzqa)hth-pHVB~R$A}A@Jfm8
z>8lQ7*<q|P5MUWIKIu9uAMId`J?Vy@A)R6K@dKJC&DnJ>=~DGiO6*2^$j8cj>}nvu
zGVRa>^98?6Ol2f-bg)@SlZAR3h(u;)BJm8QgPqlXO=9DmKV~YimiCi@RT<dcK%k#l
zQ?d#R&etdKre}e2fsW-qTAOd>>Qzns_gUc2sMa1uZby-ifdI?=h2Qq|N$=E6iCySR
zwqx~nx(|5*EQ|8KY}i}(Xd9gm*n>IBIj+#W<jI23Z_K*Cp-N4A$i+-9wlWZ4Ni}oM
zyI#4tMv1MVhrzPJw$Qxl*}-{^W8)%uRoqY7?x(8A6JS~6wt8DOy4{iPKPvJJ<d}in
z3<Ox#wVyZg`Ggjw5_h8GI8EG66P<wo%clIG<yRN7L#%Nhy8pHz={BS`5Mar}mU!p+
z2bmVQKNVR4YEytr4Fp)WywG?*dgA`p0{5jasw1*GVqqY_lJETXvhK#99A3}fBI#Rd
zQFzkI^2N1Yx)U#d;FV~QbdU|peY7?&91!(%<<=@*4J)y%lByw3fMw4qRmWZ0n1A>l
z;-qmlEcXG;``C*)&x-3^d8zIu=DVq#<_WOuzuRr@I-Aep=>~^&Nl+%ZL-UCYMQ)Pv
z$6~4*(46Rz=m*Ue@hgVsAAZq|m*!3+-ASd%6JYtpTDRwbPw6;n<)Mdp9`jmzxL|!G
z&s^Qrn-|j>EL}rQ15cob3)A<-Do3w&qAE-meg$$|fjkWaSgze(u(8==>S9I`L^o52
zNr?92heXyE@z&$ULg&Tt8*8^Hu`6BEGUQf<{0s#8xYd{at*hCv!@Lr2!ptVB#5@6(
z+D4DJ|ID#3F~c@=*<T^~D`a6Hz;gTRdutBw*wvd?!Q;f?IQ3XO0hT*n7tX#MF(#bu
zSlY{GqTEcilqbM)_fXiLpr#Mq>B7<BUP7`<NM#_v^5Etved=1zKQ-8!vK&FKN2pYI
z0xXYL`4?8r^*_XG$4Kc&8<zWk=K4o#MlK87ca0BzN{Cwt@nm@dEPq@wZwP$&=U#qw
z)?(>ex;i`omZy@TapSH!A7&(7sOWo}^mh38A(6?Yyy%~P^K)E?x<?s5iiy&RHZ1qi
z`uwV&|J>wTy}1qcqtCxUZC@aB1Chwo3f^cxFKO5tVjV!8Jm<mJVd*+L4xT_iuRbI=
z*Pg#=tHjQfWgk}V!#)NAEU)dRd^R>__6$a%qp#9fXskbeNMuYK-XZ?k>we7RaW4{;
zIE405N?c2+M0f&yyvf)-`9XA_wmeHQu`i|(BTs<k{ppKd>%$T@w!xnCH3QoYwEy@a
z`0jr$PA^QJ=F!m%+tSxGB6%YXz<2_E%$zWBR&qx=c|#;jZ3iC)KQ)L8jlOa3&ZZyd
zA%^`KS^kWyA$UYD`VZ`WCh`w+DO-Z<OOS(?AwZI{RmiytxxiMCQt3qP!@WUpQuZje
zIEoeE{iJL)wy4I6e1jR1vX8OHW9$imtdxC*{hnce0XO0<jW}4qZ?NhOwiNJUqFPKu
zIFqO{#kiXEuO<Uv%2M_w>3@?9fS?K@ekuw++y_w*L!94{p%GjAX4DaaIS$i*kpgZ(
z@)o23)6B3f$g2hU3OEV7C1D*zGz<-6f|77o0jFYlDpo)^!_XkoGadUte97V9bQ~_=
zOl+HpoxqDT)T{b$$ARE*85-^S@4$f&2s1Q92*}655Irz-q8)bQAecErBd?&nxGMx@
z40{Btk6;^EF@`O|PDR*Vz~$Jb9D72n$gmf%`U19r6=m4V*y%EM7jPYRsl%QSyEE(~
ztaya2U<Dbr9{bb_E6%VD*r(xB9MOQI1pX$hZo)QV?9(LXZ^98xV*VDaYr*XW{&!gS
zPK@7U-Fq?qfOQ|lIF;y9iRk$>qDvFwr9`(>jF%DJGBHjkx^ywlBDyRwUPW}P#CSE)
ztrp{KqRSTJ9HPq+<26LLMvU`_B9B;sL!kOVG&_hL#N~9oiDoac12@jF2Z`n&v4dcV
zVGD_-kk~mGj$?Ae=@vFdoBe+cU8tD6=U-<E+P<)M>b)^eDP`EW{&6u71u<;ACf*MG
z4Mkz`5dum^UdhN8!eu!5Q)|K>`$1@ChVy<~Ih^u89=RkEPWzd9>?LSt17;e;_!VYe
ziSb*^ycOdV!lV$<^D7CnQo!bu?7Dot+u<~pvW^Nz6*wy?+g{OL1!p-(G!S}z`j}+^
zXT#}+{Xx=6tQ+=)q!Z`{jGE4u&pp6=EobQhl4qy_FvA8}25DLo3~W!!o*FonwK^AU
z$Vb2#71*ge=8_QxQBhfin}xeURA!yLDC?`F=fjDVtwc_h$XRef)nrID87emULvE6x
zVDhykNjZn^tbI>K;$-GzE{H^LrZ)$XK*od4j8AvK%plJpl^e1fa6V+HXEZHA40uL{
zy@GUCkS|1C3|ou()S~`??Jk$ikUR;SV00LeGQqNssS;sbcH_>waU|GqhnU#SFJCpT
zGb(OaG=vRV6_}FaB-O_2)$uB$PlNfU0^t)b@U*kVqL5{#RrhJku%0-qC*GhG;DG28
zd>RIxa6LXV{#jC3Dw5JCwCjl-JmEISdE|*jKld{_5O+Th-Q5elpBrU@6c5#CKEDI+
zX4q=dtC~cEaxgR==yj7sgL1&?Q6ceJA6Q;p`F`v9%8|}SUkg5(KFueD;I-}6bo~9-
zLT0Ver@ftzrVO+m^M*?D401Yy+QDET)S%t+*(2!I`|O?Go!`v<g?5{adM2YjV2Btt
z4fRYzePCz%3|W`4u;5o>^!+|t-t}(d^})tC)BWvp8cTay&_O04(<Ed8yVK`U#)6>(
za*Z#Z@9A$UXV@<9Ol@qePi&B2V~0Bp_lA=N!;W_v?=9HaZ=JsNhSW&=OFR3Sj@#G%
z|GeuV(Pknyq}O0!8M@G&PM|IzY=$jHw#CRvz^9PyDKRcVjwQ$q+Wl6@O@fQ>{nds}
zDKxg9Fr^XN5jxS7{KvDxlqUF35TR+>X*!r7!%o*uZ$+S%bPw?vKUjIl_+f0w^MUP@
zvJa8XLu3nPCNR8Ie&f4gR&s_Wo4gNWe<6-LjJq7h;Q}tgOcAyea5=Ur$96)TSAqR1
zaG-!Iv415F65_l{9A1fg3ivW+E@Mjp*I}zVYzKyfVejC;JGd)gn1D<UX~|X>$tP5R
zfidI)#Mo*la)-154S10>AE`k^9Cpq}egdW$@ZHEtc>Wr)x`y0AVhoj_-9zLJIPm45
zvTLq)BdNxxp@1~h88%hWmMtDnPmGY$hGL>FCU$U)>%21~hw;1ch&GfHbvbbWa|E)Z
z%gJX|VL7|pe(`<u*vyl3>jn4+go6WQ*a-iKaPXJG+4KL$^v9C1M)zm%1DmD?OaE@9
z@4iSJFOqg(>A<tmcXL@=*tQTBQrp5OX<<lk9v}>xioH^C2pmx4hEy#uGW5vTXCEF8
zN4_4@cIutjWB<MY5l(DWRD32HBmk134SfC?8hUkhx&Gi-jd6<#%S)pjE<+6q!?69#
z`spAM5>=G)e9E(ln}ZlOR2^yyK@h`6tD|j!iCUo8*kK|I6f<Io$O82lKTu?W`b-d^
zX))7=h%8XdbP<{(o1=hKEyK>0%~e2D!mtUl1O+7f7&g&9(G_xg3_H($o-5?(7<Ped
zfkN<3i;(*w<SoYC7ojKtC!?5TG(^BDNS1;Wu<;lgriP>-R={Z}CJhY{a0b$3AbUve
zGBmNLS%K^!=f|+wNRy51TeIbC4$|a^@mgfQ7O92yBgpOua(Q6Lj?p6$vMNM&;8Pfy
z`?flY>}DC#;tUN9Wyg>LVnxoy^gV$Ffy?0hyWbfUDBv<=Rfg;!f@NqDF|Z7E6>tTz
zsX$HwK8IB2kXFE#knJVp3>J$EJ<YEmwSaFR)eWQ-@GYdeCC0T#RV&7KknJ7h%pAV7
zaNm;5kz5KvmX8(U6oPCwRzUd4r4VF$u>wp6XAxxkumUUsLr({~2Urhr9zzq8vdu&x
z98vcZ_x;3MIA#<O?*ifv5i7%15^W`M1XISaRYY4Q##f2<DscpF$*|Xm_L_j<mWS?7
z{)z|oC&U0WOyn<igw6M9^vAGIk=aw^0_V7XZo!KmoqZZ=)D**|5PdCw6$G!1xzfDN
zWaz8^HTv}6D_hBv6W<LrdOrAE*6ZZd3Cb0;Cr!AfC!A#%)}yCKG+6xDftI@Zm0>0d
zx<eI-8qp3_%+m-vGzIBXP=I(3$uf{ayoY2fkV3qNWZ6g|?2#NK%Ms&i$o!gE2VWtF
zSI7;r-CQ{0b{Mx4ha2YQSPd?ct~qX3iT%XkgLxfRL%6`uiyZBK;waV`#{%LGIum<)
zRNrgs`?LSLG)3Gb*@~lUI>j1Qo-Zf5b!_&j{*p!k2e9h_><Q=D5qW`@cSanJq0y?w
zM8gOQqjAu<3&_L7g8}uS%H-zZEueIyOh?-3d2ZdNK3hDOrYDhiI_dy;ME#2WH`8~G
zGdjFcvu|{&??1MPo&ZlH-_xi)SVGVS8l><yq_AbcIZ-Fc-HQqb;ap^vi=06V!9vmJ
z`TJ3LKK^8{Hd|CZ`WhWpd;@!QaQG-w%5<8jPZL+T`lDxiqAnq>aJli#kkm6pUMD3+
zdz&(N`z@D7>6w3>22NQuP8WH<?_8=}S-3|Q9tf%Esex1D<2KxxY*Z|;RCI~>`)RPb
zGYW1F_|s|erhlEhX8A<+eRRH$KNZVyqCQUS;X03|)`_}^*hAV7<VcmAzm*5=WvBz`
zlZOX`M}feI%8<L3hutu1*QKJYrDMMTc<u-(C*g4JfWV}U+;B?|j*nqaBGZ#d3Hy&K
zKC(THTm@W$6eUP2;7a6DiP{N~M-_6dLf(L9XKxyE{>A<y|L5a*!rA+e5~Vd4Xkwrc
z$qJE$IJS}<MHVmxE;f>tAq#jT!&V^k3S<fRyTf~*Y+7j+N!?~Sb|}Zr;83J&9d@Y0
z&fq{2J0$!*KGi+XsIn%$e!W6<#`J<wruka({W|jDPZNgT`3G!6U7;L+VNYSFQ`i%%
z8XYUPtH7=>BpRn+r)$_#@Nsvr-5u--L!2M~g@1lg%|p7LwWwXKkV}NS9=iT~Nd+iM
zl4rcVdzkm1^odfUE+sB-`3M2LT*h)W7N9X)B<!1w!eCP}G!pjB5#zPUbuH>7wCCfH
zeB2l8EO%LCdj$IkGc3nm<+zJ5!wcB<0`?JRSckppa2J>%cNwI4B&4$!J^m42eE!`J
zM&-Svbl`;a(5wPF$zo()jBG$Vp<;vvTzt6*hz06Ih>YdxLxkB6u^8ecoR>yhjCO+a
zQtIKx@sG~-4f`(_m6n(ow0PY1Wk!d%{8aPu?yGc<>C>l>*D0ifA!gJcobh?Y(@Z+{
zPi-f}F?8&Fg$j&)wU_O2&B|mhby!W*)kG^q!8eKeCecFF3y~UC!?=jJC}CP7JB$p1
zY0>)&dAve62F{evRDf@#R_9~=3y252AJsCx>IH_hAzY)KKDa$C$@mZV4T?{n_%vse
z+v9GSQ%pZ|qrohc$k1q<ua6O|)CuH!0_j1B40{jh?g<&1y!XfVEIf8@iqR#3>KZHV
z0ym@?b%U%bcVX+&f;`0+w*E;t5RUs?{=+2|d%`K7yO8aaj=Mmzms?3D3#-6-b6A^&
z9Rz#;dmq4l5Q1?Ra*7?q3NkO;g`8q9u@Ya%DGm}V&<O5APEkm#AhPAWi2gX~B=#b{
zMWh3G5iT3$TTD7YHi`>oLQa!zu$){l)1`!jK`_IaiY}!j3`_-=v<fIE!QedPh8iST
z{hcrS<c(UKe~U^x4>{!_4<V7Z4e7QaJ!k<4l&Umeq69tU8%mU@nsB8_pl@I#=%4w5
zB;W^CrVU&n67Yk-X#afq5%5DsqK%B@%8r1aTlikhXsCGM+qWPV#>w5xhbG~H!osKH
zZs|BiSm?btbT1wV$J_&Jyko*k&i_F7@?NCci)`US>(H%*dfz{unHkl`!|X+*%lB|K
z%2W_CcWeJ~>zV(uBS*_-cm8r?yJq_CG~|_reBs=7+`{zN@}Z9g7~N@b1hkW}T;Y`<
zp!LXpJ@OQ+{dOF*9kbwWxKaR*BJ3wv`^(tlviOWvk2}=kuEIIL0e5Kl6!&hxF+zJ2
zwr#@hV%(ug%-@82H;MUMu)YO%5%}L>{W~#!kM-}x_yg8|5aU#$PbH$~(}+GzjF%Gq
zQZZge^vlFJo#@lWIE(1B#CR3auM*?cM88^$vxz=ijB|)SM~v4H{TczEv~*4C^`rfD
zqbhQ0<dW}$e+h1Dl<Ca1?E#XD#b5vHV(YV!&&M;(XHv9Ms+%2(a7WNhDJ?_ka2a<5
zqavjj?P2ve9CSxY%S^%=KE<Cm;8>x(341mPk%^RT!eLEf{wDl+lbF8+hqMT~D`nr|
zkauGI9*4Xa;}1CGgBYihkW|vTccg3@2}u*<r6gpj7%w9s%fvXHgrtjc7757`<5eVN
zl^CxkA*;nWn}lSGaSjQ|5#u!^WQ~AJZv^#J9=(F8+J(A@>O~1@_h>yNq|2geN`I{#
zH17X<WL!3Xgz4RUGMc{qBJsOOx(h+k4dQr%_(M<xmqOHa@D-%stRU}2Lqe`{R4{=X
zko^X+3EY8$c3>7Rv(H@_ZZkBb0vX+H=j{_7$Pz3}s0}_y0uGYS;L4@cl?N1(&fv-`
zXXeMWJe)eqsIjh0TvV_!xxn4%E32@5+oNBczxQ8`qDmQgy+!e?tI=0Ly)5;le3dMC
zQponv1&sHN4-mzTzB2<LZsabXyC$O^a4O|oQr9%p16-2aP*)2wuIZCm`NP3V`+uD=
zs`<TS+Hv;EV-s4~fYj@e6*y8Uy8)>;i19<Deu%6fAunZ%v7#6&1><!UE3VQ4e?yfd
z>@J!Z;L4C8W5UJ3R`-w>bOUE#%ElKX3jy>!<a-Zw0Y<1wrt8huBEx#uqJp;}ZJO`W
zOTQ6;BO-)T!?eI@5yE;;51ihLAfZd|ZTZ?|=!=U}&5sk6I8HS$A}Vp5YF<oK;yBg(
zG*OA;RPz#|6340Lr9>r;Q_ahX3gT3T#;MXeVhVviyq9i$zEm0pcKha}o!gSg7NZ-g
zE_2E~{N;FBokg|VDjB(owc9EUxq^1XB?f&BUv3SrxpU?3uPW;=*0!bRF^{-F2x1v_
zqQ}HQVJA)Um?T1rkoh9yD7>o~DVvdt5V5{R%D2cxh*;ku|F=Q}3+qeQl`roG1-L(I
z_PyX9E6x6OTMn+|Xl9EqQ->LIc_fEI<RZ=^IUGeU;yjW=nNYaQ<&j(}kO%A`hCPS8
z&mleF-|t4QNICmkD$Q(VBgbsiPT1Qy$T3HZuOZKCs6DI&>>;WjT%9}^LC|XYXTCZf
z)CQ{K=^FAi@z4NHl|I20#Dft4O`?D1D&V1icuFIOx_83^+$o6rAsR3+|8Y3uPC?wS
zt%UYZ@3gt&FZWwFp*;gjGq7AZ?q*`sOl%?GPhAT0AJgKnYAx2nyct6YzR>?VY_<-o
zpjw}ycX%cnu$j=l2}?I&IoLFYnly(^*iFE>SelFFkQ`!YZpb<pI|%sG%Q-IBYL|yy
zAlJ&_PI)*O@UxMfdxqY%DyAo|mDp+}c7%yGD9#3r4fd@zYF1ztbanY(8o)GOxd*@Z
zS?P6yt3DkTH&U<`@$T_nu*TF24~QQmpkx$~j5>q0fa_MO3H&b=pl`U>p$**c6F_so
zkkUW%KSuz5u*kH5{}DoKL-9{3^*i1yG&;q1PjlCw^7t`|3h2}H9H^$5od4-Ut3ryC
z0Cf7IBR=!Vo}s2wq~Xke+yteguyiyC?xAQF21&96m&@V4S!fh&TdD@1+#E6Vw-VY$
z<O@He`0%7>RLP9&vvjKA9m6BR1JkOE&i$fcD*`jkSn<EK06}q^M3s*1h19`qBE3y4
z;8z7RCrla3{mLJ5DG~@vC9`IJHERPWncfERP|{ua+D-V{Q~26T_!=gBjS{{_3ST3H
muc5+MHnnWv<t}b{u}trX9z7#_MTLfick3?cA#j8XU;iKWc<X!s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/host_tensor_generator.hpp.44F3F167E8AE0CAB.idx b/.cache/clangd/index/host_tensor_generator.hpp.44F3F167E8AE0CAB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b15a31f9af9223ca21e61ffa217113ccf64357c2
GIT binary patch
literal 6738
zcmZWt30#!b7N2t&XNJ$uVHE}hM{tx2$nJugD5!v>g0E0YO=Yo1XevlTSgDA)mGTG`
z+(A?{S1@^|NNA;`W=KBETv8Fu(&AZenD2b^&EXrL{C?iU@tp78bI*4FhkuNSj*fFz
zsfMPG$eErtdx~16QaRwitl63Ab)ZrW>7Y{OtgHBBO2)-t`$bP*X`Pbqaz1^;SG=eC
ze0G;7XG4>;PkN?4@-P<8-Z1;jwU3@YXvl7yoix5d2t9w~!B0B}rmhRv@U-;iyX#wS
z#a)QoFrn<}kyCe5@4C-ysK2pw+`HSipXqV$!GX~4-@Cu*CBIuaDm#~;p724GY4`Xi
zwL6A{y*+u8-PzD5I{o3~)_zx6?xVLpz40h^?Z?H=2Xr-=38QE4Y74cke*43Pr7vw;
z^Fz{zF&%T~x1Ids<fHV)<|O~BnCeiYzwa-x{{Q`YovBa%zfw*2_iP<F+BD&f6}Us}
zT+O|=&fZz@Nzk*xxHS{YUN3y*=AI?PPP;z5{FPVmnC#du5A8g7<;I8Kp6l28_|Er%
zYreg^z;Woay<eB5nN}zMvzu#FQ)*;WWU=4X=-&5j4!^o@S>R;*(rH5@J8N#ZAGYa!
zHvQbHZubZ5UY&ieZBOAB_5N1|_5SJQeUbS;4gb3CUmu2yd$R9Yc#*qHt&9JzZ`Yhq
zU#+@5GRNCx`tRC%Emg~EhMh{^S@V9>k&J+xrudn^Hx#E<?O5_3^~FUK4iEa_*<#Jf
z9XTal+as5zy!oK~&9T#)qf1;um%V+kd}{25byM?ab_{8|_Gwn`^yGtDl}hc|uVCWP
zD@XJ!Ot0c5!DMG%R62|`{%E#`+)Z9Iyimh17x+EU-UEY$sBCo%HGqTW%+Wi)`t21o
z%np?XXa^X~i=tpZ`l#3RU-P?bSVvTF(YhGSi=yB-Dca$hdYhj;3w2cS@g2uInHRFc
zDa5@s@H$_CVo}ArXu9al3yC}5xOvOVy<|pb*2_)BSAp9q@V5|BF>L9R`th*OKhm?F
zT~xdkOsx=XA*AB&Qn@I-%hoe`7Tj6I3qU9UCkr7JkHo8stKc<PRB=%8sk*5S=7q%F
z9j~ZP?CdK_eo*rmZH&RZkhq5dzU#2(;2SQ?18uku>if{yLfTbAqAOqTsZYcMoK!qc
z7w2GJNbEBr_#cHCnR`&|qT;7PIE^OCq+O+a@U}61-|4PrA-KZ`-3SNsLSnz{UxmK}
z_b)Weh>=*!v}>8sLR7X|OFYBY-<~|K|E-tN-JMiCtYesyd7(Xe|M={;-*yMzKyg<U
z&u2nDbF>gr8NRlg+muOHMO4A7J_OA}&|3(pjHrBk)7g31VpKsxN7;`um=_X{%DnJ*
zV%jU8q1ao+i<nS^hscCfUiCd1@c!VGS$Y<TVJ`u<67aB)c9p(<XRaHHHcE-7;epyf
zgLxtG#2z_$Ualq18fHLm6oR%8xP>UWIp<mAiFxl-4$?3$R9eZjE18#tsBAl6E0jtg
zw|4IIWBI+t>sfcSWgQ6Xz{Nu7u=jsjeDH^s(xVs|4BjaaPGRt5LMlriRQR9uJlR3d
zdf+al(7P1+SqP~Vj&@ATO>B<FUF=o-CEZI7=7q#X`wo<;-}Bs!Zo@;)LZ`EsV=^I?
z)n$fxD;LeEK^0dOuLb*BENPjvtL(jZ{H?f3BQ^z0fZ&e7_U47erI*TlE*!0VDCM~g
z&(dZY%nOOvoxZr{_VxKY@BlP%h-Qf1ypXu8Uu@o}q_r)bSpZ&gHMp(DYLyA8Y+0OC
z(wtrEp<&MG@&%w>fF{aBiAzO6OI7dB);8cVxX(t=Y{VRs3GK6W<A456xYKVO-UIuF
zL1XA>UP!!aWyG78`)F2TIiaD0bb}ns3yBZ3T|W6$<tYrj9VSnrHql^SNc``&rzc$M
zzBOILY|%(jD;ms;Vqk4u#*jI}lVKPtY-;0l;~dNjeP5TtSFReCy8-PNRNPKur#CMo
zu1_|c>t_-+qwU>Pd_B{wXB{nsRE|e@&wQ}z+{aROtGTz<+hATueDdY|D4X9G-oyj2
zp)Cf@Vswy9NadSv{S)%*TI$h6Y-nk^GzarS;zs>kekrV?5*>u8@;L~fW6H{eRDLMf
zFnm(s;$wOigq8jWIQ#+b7D6glKb`r#xWYvW6h1LN_C5ylLgJgz?>2>xU5Z8MiGG~N
zwDYjOWTHaP%Jp^ggJIW>G!09YGQx&m0_`Qt2$|46zdV|dnrzb+A*rZ&4QOjnMJA+j
z@8^^An`d-t(lZmL%vuoEVwK8-RQ~$yQQ)$GksI;K7_=l^l7o36@k5XBVXqyHU50l-
zhoyH)?`&R3{HXD@J#H(Fb?8rY`9csDqQhiDDo?7DGJiGr8*vAW>n5hz#Pk+IDs9Kk
zcUv%YCq6SF==MV(9Ku~>LMp1vw5eHd*emZh6U^^6rG&_R1AIft^Yn_NlKh05-HMB7
zd4EWOUUCrR7u=#}TYPxwp&oAeYLPF6prz1H`MwT<>a6)j2x_$En<1!K;Ry)f{t^nH
zZS0>;9=jzoVKEbVIU88c29uLSUcrJYSQzny6M&iV0aYmB1N?6Z9MCC*3-B}<6i~l+
zSup=!Df8}*2Q{yT;A#k=UP3UzMzDYba*W9Tgf4%A553_ce+<r#p&LDP1Rl`CGIF36
zCi3MFvK$6bF%pDO3ufp*tyJV0ri=i}E`h(Jd&fbM<(cx@WgatHw=KHzV-fnQ1RP4h
zK+m9vWoP>b*i%Ik8bOmJC_=s#pX2C0=s6jX(1m+gkO+P)!4TR^<ngBX0D8L#kKkow
z6hdVx@+Roj1byf^68KdRu7Z;yZ-HJdirmM7TF~zjWU)FeIKDrfhG$2_#}G$I27@cZ
z7`kAu#_ccblK1p_A-I7b6fPK0`Dg(O@Les1aC&Y7cQiW2e_rQ}f)WTRflzWO0U`8`
zgo0kRpo!jtjtduZUw1aAtFD~sDLtih6)Tt{@r0D{dl@ECuNz=rzy3hUoWDWjRiLW^
zJtbh2P@d5>((W*tvW@lF#(Ze2sK<Th-YB{lsS(k5zfuS#<3+w1beqARb`bd%&~35i
zD?nFa&2I(WR%?Dg^xY2;^gU%X+AQNX%822|ems#A63`i&z%-|6M#@!@&vlw>#hnMI
z^I)XFi~Ke?-B$Q$+l>()b*OH@`#oP!6jnS|!h)0m1OxHEW$;I6j6#I@DI-6sZOI=V
zo54u;p`f9ng#Rc@#OJsPBft2jrf$_FjK^Nk?FBpfns&u3li?wKKPhM3jLAn1U$+&x
zx2Ly>{#)ceIv+dgyV!+IzWxdq;~8V6IZ8Lmj;<~81XF^)!qtFh4Or>cg02?q2)c><
zIGB!uKk+H8jia)iCWT=_P7$ZLDaM=t@dUK<X`>>K%-^>)xdUF_Ef5C)`A6iz>R@5D
z1@nnKOdTdr=8HVhEzwGIyn4JqO?X;Svd^%e<L02#+Q6v|jFdW=yIqr#zUcd36f9!S
zi&$5xgW1LF$2HvB`x!?6`Myh0MK{X;n{3ND_VbN*)!&`A5qSaYRKOf*E0GtoPQ}cT
z5=7)>tW%j%dIZR^HOXk4A}R9yFl#@&M-i4P(6b7B6}}omtCe~n*sc*QU|s29P2ko9
z-b$)BL3k5H5-;#8AY6IA=ef05>$gC7i=v;uDKT#Tv+N4=S}WMLf(JcDROy((@<^au
zWiuEyLl-iAS)9$n-&Q;e*GkTEEl~U}@i7H#wBo!aEMy4_SNLM)TFkmC&MaoJ#VlUo
z%b05!>q^Z?a&~MPizj~h(La|DDe`H-xR$f1ayFC%AKltN>y^kmD_%&Em5aX9U+DYE
z;TKYb<{Nwoq~V4DiO8b{L=C0M0=d|zF~k*pd}!dkkT<k1^h}BbW0nz!LdW=Z&|e1^
z;)@H0yWJm?(I2fT0q+tBqV7TkhV4h5kO*W_wO~x7@ras?$ots!ai-#-2OM)SMVn&P
zdaOA?o}JojQ5vTx`{{;wyfj9UGu9up&+=~beU(Y?l}8D9lDSe-75)S#dev#vLM!C~
zOUayl5U>vhQ|1s_!TPg|Sr!vW0WFY6E82lZE8JR|s|a$5+|$m}neZ45Sg0gTS7bki
zFEFYfU9Q^4HgV<Y&A0JoOYuy>8C(7ldAcrLR9=~BrfFQ^8o;vwtaQ(U?ktEDaFL$}
z(|O>;ALu`)d}MN|_(FaXfJ6hO`A#|8nxjDvog+<oq(tK#W~wyX(N-d#=s&SH^+@FU
zPmCgtMn3eWH1E;IH0q%h@|;K5M&y;urIK}_aw2q(iqf=4mf#a)#eEt7%dX~~`={Y!
zFiV`(-uy*=5yXqiV?cluZINd}%7Jep1|{4d8Iyy~LV)`iyr`|yu!x;dniMI!5xNFv
z^jAWY5s?u?oWQen*;bl)!Fg7ic@cT-nx|M7ZOzz*t1pD;)XfQ=mEU)~f&y%H__k9@
z(r}8-kVjIgO_A3_uX^Z1#Y=z~10~O=glHsCZ(S}DsP}pdestkWm#a5l^R1JR<nv(c
z@~SVrT*5akMFCb{33Q_Zq^BO8D$ThRujb~of9@N*<Qb-TmS>i?61QAkZYS~;0x8p6
zUxm8_o|nK%_d4jVLnlgU30RqKgD>%fEHOzfv$N$C>T%eF8bI3sR$22rFVH+~`B3~n
zp4^N6woKLSiqG>jWuqzbY)w6bW^3F*8mz4<#yF8vjgSMdX-nfZWf|kS`dk;{9>(px
z<h!so7tvRgDT-4o!3n`VzN><wCU&0aM*Cu#x@UBy-9?_~o@d1^1>;iaPBkI&IxyBj
zcgi4<H-fPdx|6d--VDZOg{QF}GfW=#sg|Vauj_K~Q8r!<gO)=y@sDF7X5|ida>s|K
z6dX#ym)?4+cuXYuPXyEw=x+mfl>8R~3X7EWZuQ`;*jx|c^$<z?-@&^EX9kV)K_unt
z<ZGl%Md0a_U{qXo9Gs4Wkv_^yn>Iy?a#m(a)<C3;pTAZZI^mOLKdOEG{C!lumJ~qZ
Nqx|Zx{OYIt`ac&)9xVU>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/ignore.hpp.67D37926F4126F70.idx b/.cache/clangd/index/ignore.hpp.67D37926F4126F70.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15ce55e538ce41b270023778543157ce92c6fafe
GIT binary patch
literal 400
zcmWIYbaU%qWMFVk@vO*AElFfyU|<jg;^LB`%swD}7Kkh6#QJ*jF)0c>(Uh~fT`nXX
z_uwe|glU;eo%wwm3XS%cD^|_$Ub-!0W<vkFM^06TU1THgT>Yq=t$N?r<62#!i}lCV
zwK-?k%71Izxp@L>!wO@^9E-%q>LRf#tIlXdKW1}$zKA#9$u97$$^pw0MuIIY-sunb
z9h_lvH0aENXZ4F(waV6}e`Z?6=Q*jkGB+s^=pq)DL%nZOIM{d@*qIpEgq4Ifzyt#)
zP>z+6k&V&q7>~EzuLIg_OdQM%Y+USI{9ppCfPn$1;Q3L%|Che3oGHM>$-=<KE6ghm
zCKxz?CV)UuYFcqHkcQdB$OX1x!#wT%sk<g_gPG671vYneM)u14`d{05*n}m86=8Z{
kb}%x5&4%d#nw<mk1VeFAaxpiE!5059u${@$%!Ppg0Nt{EDF6Tf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/image_to_column_f32.cpp.E393FC8829D80520.idx b/.cache/clangd/index/image_to_column_f32.cpp.E393FC8829D80520.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e0e30170239e51c45661cc71262b267fd7e335ed
GIT binary patch
literal 4960
zcmZvf2Ut_d7J%=C5EHlr0x1ZgLud&YdJ734NCF5!xQdD>Qk7x=k+$f%JaiS=bwLqD
z1VKgDf`w;UT}0g{R>by1d}2ifEPyEXH@G`D*L{!gWA68#b7tnunKNf5QL^CR8Ttqk
zm?+E7NXnf@K@fxq{&RA((`Kn7$nXRN$*<iL-ILklse9uKYyKZoH)e2O1wObnYkuXe
zj_n<dllW6O^X42nK6W6!w;{=6O{H)8l?k$1uiDct++Bz*GCOYJ3cV|<laK5^9N*cp
zY*Sb7jj_-D@AGEIMZB-Jsc>={Johp2bYh90r*ih3TF*IAVJ}S&KKz_q{OOa^oZ>Gj
zvoDz@SN1RPGk(*pi0R!vvY@1B)U!r>CuaJR!2+!@cf-Xe52e>O9H85ZniqE;(|)HZ
z?2n6VU|l}scrRdDt3{TqdUeww!8*3L?7j5kS<BX$`LBOVDeEpkBbwdX3J#ka6$IhC
z3|$>VR-_lcEdIRr+{=W<L1&|$r$-AST1TikKOe9UFpC?x?LDJqTRP?L6nEPcpTlFX
zpAB$3PQCHBF-eteT_fk2rM&BiUv9j~$~--qr5&<M^h|wC)6L409W7Q3UGpCuS%kzr
zu#gUHRu0)1XT;8rF7vdQyk&B(AJ6=9#X(X3^)lT{_d9$q46W<f7QXCe_hzM?^z4W7
z)Fe7q_~Pur5y>#)=2N!Z+JK$DHGcb@iXY4x1MZy)+!GSo(TX>-C)C}mO~_UL`)YfZ
zmy4r}Luoq_x9I}1AkCP2Xr$Hl@rXC2S@Ep#^fa@!WWIH5Rl>y2v@?%{Pwd(k70sCW
zBDJlzp+Q;qL8qFlYy9ZfV{eC#?(2VpPpgcJJ-olgKPxA1^PnZCg+HU|7+=Tk-Y*qW
zO@ZrAE;GAhKF+wr^|7eIN6zG4*1oTKQ`EbDc2nS-qsi$S?>*MCeFLUGk7d}sy7FH9
zJGXQ7(;HXk_av@L^s4#v_Go-|R#8CLqsk2yy{C3eRiM6Oqs88ROERzC&pJ2sZhd=z
z&;Ep`%|;ws+_zc~>d^S$)ox_E)X1nW#rZ`|Xq{r`$Cj}6Cj(0h6vNM3E*x0e<ufqU
zS>@w#(fC8-vW%hql8G5>*@|}sk#;V+Q>TB#s^0bHzGSXDzo_Wfrw2Bd@97`$k2sW}
z7@k~pm#)60=*fwL*$wI4&G^cM+!YQRUm+_j7mjl1)qUdOI)}{EJ(TcE!(D4?>u&3=
zT~WVb<E%SJSq1HDi^?A+XNX!o42DFV7X^Q88Ca1RKUn<OV2S4HjiaAq*SGCSf5-6b
zS$HI>ZPWV`6^ll-emKvOAkW{P4?}VaG7=CJLAc=qkv8)3Lt1K65d_Dm81HKXPT|)-
z`{(`KotzT}T51tMO{3BnUmK`aL=G3deGrh0A<;lxg;J_eqrX1T<og{Jv-+L=?r5n+
z0uwrw&i>j!b=i}#g0y>cCt%1lpk`^X*k2o{UfF5a8e+Sy9YbaS^$wJ>1Eu}-fhJp5
zee#>-XI!YI76nYGR4Vgp1Jzy9T~vOLO|zDo0;n~p8oFN_sLoDG&hcR&h`LB?=a!2J
z<rEB;m`E%Mw{WSARG=c6tISJ9azA+>BudRl;j8=nxhRI$q55^GC5(-ROX6?NS5JHa
zish);a?}O(pk8hp7_@KE1d1TbAXg}+4Yo&~zOH!<h-*+<4a$K7VXceVX%>qdfFemZ
ziEvT(s7Z#)x{yvaFpN%+1sn#$gLQ)mBe>1#n%`_MhAW&EA*wFXtY~{k2A4#I=bHVm
zgXzjqZaHcW6(;QNy_MEKpFk;iH7cw|{h))^1#9Xeuf0tHJ^oa{Aln+&={DXFEmWuA
zK3t!Pa3BiqYv4;bOu@wlVl$|-qX=9V(6%taScQ`ceF^btPy@$uJHTRjt~}CKn7W!0
z1e2Dc`AxE7I8(@UgOR{+i!h65I4*|gP0I6xBv`q<t@q=@UFsN~qMgEoVhk@unWZRO
z#r3FmJ<5dZ7d#uivjM-B3Vi08a*233?w+2{+FaR6r{E%*hz@&GaA%q`iCkzdByy#>
zlE{tbMk05bJBd7K9whRld6LMB=0%YEs>_Ju%>pIx&dJ%y9lGIdGIV|+b>0w-f=d_@
zT_~pDQiha7DS9a^l}^b}l5|SG(w3x?DE(DbigHR(0~Ob!di5v^MojYj#*4kpb$P(2
zFugECXcp-E;CguXz(F7uu|*D09K?FgY|vLP1Y(*eO$@g=`1n6wqz@)<0pcurmJYO@
z!W<30H4>c+#0suL1jV6hjt+ds(heY|yVC<<c?$p66~ztVs}3~^E@Q}ypfv@TGvqu-
z&}}u%zZE$Q)Fk^Is6Q-Mif&=Z*&UD7!6XKz21F*rEPm0~_4-{E$X*tYMXU?M*Q1p6
zs6HG%?y6^jrSLWfq#rAc6$57<KUM5%)C8LS{rJT_ES=Gsob?MB-ALDna2mr+bWI3~
z|EYB3=1FQW@C*}U(eN~+9#8+6Jo`v15U)a6t57bSH1$PGfa`~6qhL@^wx=}|WB5|p
z(pX4o**SR;V>_h4NJf!qLotTSDRP38`F@?e^H)FpPua*neg9D*rRx$M!>#Go3{`JF
zolhbgx($hJ>9!=YquY_lo^DSf2f71^1atw3gmfW6Mf2_^?bA3M0v53e)!Bp^!Rm$K
z4X93oii;9gdGw}vRRcpFk4Kbq40mC>5LZR<>J9R{jdR;TK=YmQz2Txt694UU{Yh!|
zc;6`AY$(QX(<oDdzNa7FH|wHo1oX9-Xb}pHV5*{;-`t?f_k?A{Z0<S>hQvMaU}%lu
z(><q8f%HREwcp6&2|2*XgY7}eyb@VSJT!t|kly1Lv1<KIDAt7XVBWzb2Y&o{Aw$C-
zd$wlok~D+^zsVJ5wOqaQ`A=UwwK3d~VMwZ;M&?FD-N0~ThA~NJVs1jxS?gO9#R0?l
z`g{`E=-ZIUR^OIHcKUWCve&mKk%PVii3Iur5()K%1pWLdU~PElkN*U#is8lBz#y$%
zy=H4<QPubnCogv0+qrSeb$B;`IMX@P4_a3{^o>T|*WL!iI1eXQRzGMTd7At76c9(~
zM-Wpsq=!=Ein0e(P?N1Bd^iZG$wH}<iqiDb$TeB63{&aKP)^ymb=RjJ6~irrZ>9;F
z-@KK;4H?Ls3Tpy}`)c`e;E)(D))JFQZY3vN0fpBqYqH8RFE~#DY3ob2-J^cX_uyd_
zW)em`NX8!(!gACZnqhba>QsReub^?3RG`G5O|c%F^XJT^AW{K(0miUf)5;k&j~D;W
z1L9z-U^gh<-NqE{;wr}9T$YBGj!?YMt}kZRkU{{8ywXrf)I>lcC2@{hdi|Ev7(}Fk
zqayQ`g~Ps}m`#>XhqQn8E$6M}I(nc)OXw1!#$mXWE+vtYt|aIv<5_3fq6G->sRA{u
zK#5Qsomok*HjDWYC~_D%#L|A-8GCSG?9&%8g(=UJa1X<s+0I0n!SG^P@m%QD8NDmp
z_WbAg9pJhb+l$B>3}1%ImZ3ADSqH~v_kQ~`3Bb&i=}O#n81AO$Mg--&QT_vZK0`x;
zf)|U6LttMD{+;+c5|xNcNVHVEltjx=@iLU)%a%nem-|iAp@NVJB$MEPAY^V*;^~Ir
zX@+SgDxES)iL0m#H7P@h2P1~pqk;A46qwXktS^Y#f7g_QIC4xm#H9;=0MInQdH91|
z_h9;~a@|MEhg5!o+=3*kd4t`9r7DtH%7iM)ipV1FLGX(7&LdLx+Odak`j%1=5N3C_
zyPc}bce3x|Al+)NE!7B}*#(@=w8|u^%WY2E=R%2DA?O=s6=n;q$4`w*IqCtqXFV<d
zZKm#p@!k=c#0KA2bRtb3Vgngqk!DN>)gpOf9<d)Ft;D7zvgTP6(}Nnz<CDmSXG0=e
zo-K*&dG;i7;5m>;z!Q*2$P<z%&oYlVn16rDjQ@tRsY9LN*7iHnmzWKk)Pitmurv(d
zmSQ-SMKyrMY@l-Qo6QcLTfl+wq4;RR7=bcI@g<R%A|_}g{Nmy=+HDNPFAvKjGHYz$
z?pnvbxfVdoOkyU(_<cHgnZMCDzZQteRRzgTN}o4=2j*nY&*6fvI(|*-=6Y?5*Cl8S
z5(95X1eu?an4SVqo)o4|(V$MC(X}vbT^&XUb8-NmsTrcj)n^$@^z`!a75hn~j%){S
wOLtq-|0m@b@r=#PEv&3<?Cb?jlSD3V9wx4aLgzp>Cs;OJEhyZ73L1+17aC5vmjD0&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/inner_product.hpp.D15AD8CC8FA359B4.idx b/.cache/clangd/index/inner_product.hpp.D15AD8CC8FA359B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2dc6ce05b1bc4eb3cdb855d56ce650aecafee144
GIT binary patch
literal 7484
zcmajke^iuJ8VB(AtFKp`eL;9PX&~oKXcTb<6b64VI3NgA;xE+1Tsv#7nk2aDlz|Pk
zp{7tn4h5$oJ4w2bpqo@yTDfh>5sq!L8tj2%I!82EAwi*EYp1>A?B#vvkDPOSUp{m1
z%)IZt&vWm6uZ!=`$cQ&d(wsH-=jW|`>|t4wq+tHf`Pi16clr5gNK*c5hxVixwoeU8
zcfMpzyJP>H_J_9S{q17X=W9wneR+p@McPMC9y+k<`;E!hPJ|{!)&F>U|2whGKV>+N
zERC<6a-w6wjdhua>z+EZGG;>5)=lY8tW2M=-(2QS+xonxdip8pjmI6SckNWOUbuMv
zcf;>o_HxvJM*VBx>-1)0hlxGM-Ys9ZXjJ|7@SHc#%5%-1p6%#6x9#aSvj1>=t83Gv
zd%9MA|II_6d}~|%!@*xlJ<U1kUl$(Qz0p~axB4`{?h%vMsCl^~s+3`)WQk=3Wtnar
zq*467_$ze$S?{yiZ!Gc`M@SI?%y$sQVINk`jHoY?{KeTp*(T;Yh~nstuf8v@8TVg*
zaTTO01Tg<eRGBvB(S|*H#%1`cR3fMnCgwkhDv4JYT-$!-<FEWxnjtmwZ}Fc*m2F?m
zTCi=?$Z`HE&LF3W`3|DkH6!`(?$hTl_=}4WRK$P8e-c%8>})S!<6}PYSE+|o&ws>!
z5><|bEg!b&?-RfASIG>@G%?>n6xV<Hm+CVoW+(cKy^y^88~i6x<*0e&&4PtL9Z)3R
z*tcHH5M(e(qlZBnJsb=dhPm?3Z<u+={~ZBA0Vd`<h|ho5^vm^g=7}DEahwzvz<dW$
zyk+e==khQ~8n*LpZ*1TB%d)~E0wY2O`;bK#BgPX=iI_o@8ZlRB_?9OV6F$3=9H{WF
zh^fT1!CvBiuoA0u??uR7n5bj-Vy5mYL{uSaFc5g}XH{Xj?s`PjBWkeqd6Q?=W4Z1|
zgf${ucN4;zsGAYdjHtm7;c=1GjODt$FneJc>>J)84|uUsSl)2#`GOy>oaQYR7Z?{Z
z*uE^z7&o41O56;h)VR4qB7I=El{Rm`ZaJ~`nCIx|qsI?lzP7kjQCKmo#fTe>Kvn{4
z3AGzmH+3njrPLl+J=A5emQk0(T25U7YXx-+tSyMs8>AK1R_Zob+o;=NZ6|)F>ydr0
z{QD?xpk@R%BSinAW*D0>o;tP}3DjxLm?s=Cb@!`}mdCc=p)j-2tcgfd*aYJQqIhGx
zCfZhE(~Q%IrW>ad%`na&dffQ9HaG~dt;VfHJ7C-aP49u7Fz%#&7RG0(cfq)edOwW&
zq3I!e0LBB<XJI@GP28x$&cS$&`Vx$npox%ESO<(9)R$qrOnn8$E4l-|*mg%&>X^g4
z2b`L7f_{_ENM{U@!<k03(78-VU4K{C%GWRK9woC*sGTs0E0I|j)Gq37sNK{(P<yC*
zq4rYuLG7dNhuTj)0ChmOvGvHPo7L~f4U<`Ja_(#~cgU;}DUC=KIW9tjr<8B15h*3A
zk2KtxV#M1D%a`+m#l0%bCA)}<Ar~W9?|L_6H?;?{hq@ecx$cPMcR##7^Nis(7i_sF
zv~*W{N4!C1k!qw#T))gFsgsB-s)fj^T8Y+&ua6QpD6^dK9HMHd)i8<hlUWVa8tPi8
zwbXS`>vW6H@z(TDyrSXa&)nD1SsooWOkwlY`P$&*#^z6+KV3+S<HHUApOeL>6;=(o
z8o}c06;=zmR=3Co{<Hp>T6}Qfb5W;nUaInrQhpVU%CDkPVMC))VMC))VMC))VMC))
zVMC))VMC*FY3k|FIg>tkh2Q@edCcAZ=~$8!HkOSI7g~O8t$g(B6<2RvZFDDWB6oaO
z8{I|Sjp%Ob9z^$0_aeHNx)0HP)cuI=ryfA`fbJD%k4Hr;-5AM7fhq)6Aw+NZDj2IU
zo;tP)3Djv-m?!+l{MY`x>f?_t@h{5MGAD?R=8DXWA#!A<5iQJICbVkNg;D!zqZS1z
zEL+Q-AflElGCPLIk)1}gFngJ<d|SR<&)5Q60g=n*A}X{M5*5Q%47;AqC9suHyJ2%v
zm%>&`?SaihT?ShjbvbP1)D^H*=w3TN<>P0L{B`v>-=avXgjpnq&l4+QrFK+e4)uac
zEY)2Ntr}+0CB9s;Rnso3fmTDyYhbIP<+adiX?ZPdwY0nrS{*H~gRPF1pG3$>j1vQ{
z!Wt0LK-~yiBNFuTCfJ&&TcEYTEOK07t<YMj+n}{kw?k{E?u6C}GtDKfi@F<HH+2uR
z9_n6bz0`fs`l$P%^-~W(8_=B_x9P_AfziP{sUt0sv0}Vd*d)s&B8$aBWVKj{s$s20
zoZf*ou+~u5!dgpR2Wy?~+!t3zrg`T+!7pHmx<nIeNUq6Crt8WL$PLw#CNCh5XfH<Z
z#aKO}_QJduQ*}2Ypb??82LqaPZ=Pv5xO0W{`6NDL*&`E03-B4sK8eU;w-8zFR-y!Z
zLZVox$;@uI6FKY-q9l6~QHni<DAk@ylx|Na%CKh;<va3|#F!zo0!IOn%i$s_bQBU5
z!%>VRF=oho`Ryp7cEjPOE`_6%+5?A&x(tpo>T)>BsVm^9psvQOYB)rs_-5}}HPp43
zRZCrmS#{JcNNPc{NHv+KT2d=@8<N_n+mX~x-HGH*%oeHUTk<7$QFkM`o4N<dJ=DEO
z?xpTSavya+lKZI#kUXGU^cY_T_@2kbB0zKpf7|~QE()I7yI|+?baw}zor<6pVS=dX
zb7T=>s2xQ}qh45qWx^uz`D)Ss5HEf~VJ(nb5G-0$VQrAxs5>EdB3QJTZy6$YQ};sd
zrS6B^uX|T-`nZHIpAX~DIaykkS<mLIiCI>n#H<ved0C5ftq)m0PS5+CkQ}1i*xVU<
zSzc@&QNEUM77ebj0<D0^rMZX-wL+p|XvHv#mQ#3&XeHEcXm09KXr<I1Xddb^Xl2yp
z(8{SRpjGHDzVuGg48xcg`8uo-)<(pMu2WbOtWCPzo%NfSq+ET0$Hs4sjK4kdRgE~d
zsVUU*1z*x{2;C4Sno?$`5PAw>;<w-W#&CRpV)gu6&+lq6Otf{>5^6V8H+3o0Qfd!W
z4|N&TGU{@u<<u2WE2vwbw!ox^Yb(@N>Ncot)a_8)bsxO<v9gz9Hnj5_Nw6fuiV<63
zc8i_JVQ~;8S(1oSEGa~(mQ<p2OFB`8C4(s6njc5P(ON*{vbu;0t%ZX+^y4yR&*jbc
z@Lu*J(2G0uMD!xm%Xf>!HXARuc`=LnUN062i}flWV|=@GBL77(f7`z|Cko!Xy6wF?
zORgWh&BLD9!o%oWTX*8qAFR80I?d4J<vV5u$zTxA4+;yA0*Jz;aH3c#mMBxoB+8Pq
zh_a<@BB$geDuPr5;z~%B)cYXqqdo%Z2=y6AXQ(eix~TiulVfrMW18OdC5*fwkS2_L
z3V|YFj?W4Y?%4gEn`betrtrL?)jm4))q{r;Z(P1jBcE7&@4$$+$2J7YEI~~$iKyl4
z4AoBLP#r``Y7$Y3nnIMSrV^#A=|mZ722sA6ZxT_<*BNR7kxO+E6{>|qClP*9-^Y~s
zk~+LW_lYf+zFm29+alk68*&xB6&qn_gd$q8q3o?oBOkfq;;ooFId`h4rm(!pd4u}r
zq_Uf)pZ@VPe*mjTU_C;_r+qfoV?1?iJzclgBds3u^zs(iT98P`7F#QI8*FXV?Xa~|
zcOt%%W?&cMyQsSn-%Z_v_#WzB#P?G7A-<2gAMyRv1Bf5c{pGBQuEh(-Roo_+#Gd+S
z!>zq^k+FR7FGmK8rLV#)vV|x?P6!rZr!a@?AWD%_h|=YBqWSXtU@`LW^}M_!Sj_7e
zZvKAayuRJj_y|E`nvcXJruk2im{u5#X@$|4Rv3+Gh0&N+7>#MZLq=knFAPXb^Zhas
z(+Z<8t+10YoWxi?TpD0#(CwZ1<(%*9jJQqndw-6-R(2z(U6J`#IGsv;YjHZ2%53OV
zDzl+esmz8>r7{~jmC9`BR4TKfQ>n~Xesn7Jt)=NyD)W^fol0dkbSmw2T4uV>y_9O;
zkwdom_+um4=Hqchw)qrBxA_#d-mreGp42&p9HODye9Eu3`F#6pbem6ML$~>^zxLsY
zDQ}0HZ_^nfr1|XU`;Q}$3qRO4e|F#DhS|3np0=#beb_g%Ik)`Dd5`ap5x#@}#lD-I
TR_S?QN8zv$Bc+kzT}k>Mw99Cz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/inner_product_dpp8.hpp.5615826ED9AC4030.idx b/.cache/clangd/index/inner_product_dpp8.hpp.5615826ED9AC4030.idx
new file mode 100755
index 0000000000000000000000000000000000000000..109529eadaa85dfa12810052d8d53d8730d805c5
GIT binary patch
literal 2696
zcmYjT4OA0X7Jdn#CSia?4UmBd5h9ElAOwXSP=1n#u>q7Guz-jdApSwhkAR?7P!MXN
z6$@3g+f}NZV)-cxc12{hsMRjNM_ereils_bgt8PA6m}-*%jBJtJM-Q5?!9mB_q}%#
zrd+q~5E}qlyfP~(Aw7-(0N@h8wDi=SxCsCyOaQXZ)P#?PK6K}FJe%V==W^CZ{_}fD
z^^U-|GkJ|d&+e(7y-(iXJ#^#LXk&We`~i(~Zg*a{)7r*Qx~6i*24h@%n4NXC%N$m`
zK;GTyrw`sc>0=hsh6u(#$q;mQ``;N&Lei4uwfWyAzAMNNw5?AMw(>aIzQZ*&d4Rd=
zmEGuPK^*RhjE%P9ivC>JYvNV3-r~cif9woAHL(-9yN6x;=g3{tMSa21De21RKmQok
zeEB{iQo1-*{IO$lkNFntYU7b1Y5g}=0h?`Xp1r<Zb~bux#6Pm~%Uuo~4LZpm=>2AT
z`poL8`*SlaW&O6En=(x9ZyoJjXUDGmkCi$*X<}^hBjKy&dd;Pl8$E{@8P0;7;OYe_
z3x7S<;5E2-)A9U(^<Af%`rfhvs@<AoOMmUW!ufEZ-@)3Jf9ZKn<eL7YRegKUUu?PB
zHG$dowRfAKTRZB5$1A;N$M50Owf~NK@KbZhwbhYWQ8butD0AxQk6I~h%GTsfdG1ZQ
zKP!G4nqGW1Vz*@_+E^}VwRmsdvvj}Zz*x<G2D&h(UHfqQ>#)9jXGi<(4_Y6EN^5wH
zrU$+I23qIOJnj9hWUEK`OX&h1U%R&rOF9EzJBYuyuzhUUGRLhn>-M`Zg<dBcN`4Q%
z?fNQWRYdE)<dEk0zb~y4-27npJK4wA^ADR_w`TMuhh#S$Juuv8jVraCcv+W3ZF5a9
zTI+1zmOUt$#I)P;<L}Gni@e9Nf?P%0%~SSb?K1Amkzeu_Nf?W!&K+z0^I-R<YwKPp
z(|`GbzY4o*Kl(ZA17X!vQqZ}_C7JK+XIjQmd`c$LXJ1b|z*F`gXfGbRSXpzdCU#e#
zt@UKI{7q6tfnR`pPQolLyXNz>nHLj!_oIii*0P-5_v7M#WX1fGO~Y-QguDyN()PVY
zcluMW*KBcW+IlXo<;kvu^TDYx|LvPl39p>Ld}R6Y@LDuidGNGn$nxJKSZ=1L<03tN
z6m-3q5uCjm*S`7XS1AeMJzpiC**9<^bJ=(WpO=t(%XG&~b<UD?mBnvD@f^FEO+Lwy
zHeb)~+xWk*PxIecC9SV2{AncWIB)LOOvRmsh@L1%xtUloX-U5|6Q8y(X}cpG07i79
zYO*)Xmrl1{L>w$)pk{0{3&7F?tNGmUnexKp2oNj=D9*thbq285@0mC)q@lDW6|yZ>
z*B`Ay=u3zNLXHsA89*FNa=3r&@0WJMFJC{+=kPI<G;n%#LM8UY2H!mWG{(U&lr(S}
z-??q&=Fp5c2(TxT7tlBbG~r(lz%pmibvy8JI~}QZIJjM(gU8`vC~060pZ;}&q1}=K
z{j`81z);e_X_-nE@bqs{!REk%D4~+50(FM`ewr1B_`r;Z?k;OO`0@I92}gpVq=E4#
z*5+8A<K5^n2aZI%#X<|A&XDnzJT4JmRvr<tvfNCL0HQ&<fpR%q46u*|K)3(%E&-ay
z!RVa6fS4o3P}0BxbrmC}Pm7vk5nx9Yh_;A!(HX!pW12dU1-gGLc~(7pONM}@!~!)(
z?W8kQf&Qp_;~T#H+*Myd#1UaAX<&i52Zx{ikf1$+03i{t<!J47hNPIG+aO3yNKCs-
zT)j{Xbhb_eP&x=&L}L`utRY)a5;V3>3UC>e0^uel0Nqdmgqx%ux?wWHt>+z1K)Q(u
zl4gVZFPaHANiTFmdWkKQ+(I{k#+t?2!PZ7Ts++wsBG^r1ps{wb0_fhh@e%XvMQkIA
zpj>yZmw`x>8^~2sh$rRYa8n4%Q}9$2;!F8B{B8v0EBGo338VttAWl@E5U405lnQa!
zMg$cqgenS&q$1ovDJoKkR20Ib7!DhXpqK(vQAjKm;|8svVue^mA&FFi8-$KZ6cQDM
z@a4D!GLN9RFD|2ydWG7<pg&Z-MlGjMB0G^wAq`h!Lm}-lt=OPDRO_wvg9z3ocqES*
zBrL0KXgHhE(?LT}m8EJir1jValOv;_&3hAW3$_Ip9!B_NuyuzKCP7eNZ{Hwl_bLaK
z8_a~DG2Ssjkb20j2}^fasRaH^e<77X&XgMvY$maB^7z6Kn2``5d3K>2K@$WCUa%qc
zQ!Nvs3+w&p2&!RfIM7WXPmM1`|4Mw^E@%HdWKPHtaEU)f`awtz@Qe^2@QsiF2#k;r
z2#t^kh>Q>hFe4-eVk0C05+j5I+z6?G+6ZZZ)(92QKmpBw;fI*8br%%XiZo`L9poQD
zGnttN1eHxRmAtZG$n)7tD)-EwYJ&Dc8l#Zr1e*b~5bMc{46Y|`usw0)nzU4Pnj7(?
cqaI-mf#aI84^3?Nw!j9)1Hh+VZK>D)0}QR4hX4Qo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/integral_constant.hpp.A156C83DCD53E7CE.idx b/.cache/clangd/index/integral_constant.hpp.A156C83DCD53E7CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a78929738a4fc01357e804f33754aa61a5ceadc4
GIT binary patch
literal 1574
zcmYk54Qvx-7{}kX$M#<DgYKwsp_SvV9J#L7v_<9^YG+)f&V;N1#fHqt)XkCcv8+Sp
zK+!R9m8c;EOkgYw5{E|P2X&4m+X5zvbD$9tjmDvKiGgejflxn|Y`%B;^sY&t%kQ51
zzwh(8p?clAWkm>8Jzw3ny}q@MLI{<Rf2Or%>o|Fyav;>!z3a^jJN8FCi9a{)eN1ia
zJ(2J7Jaa*==>70qc=4D0+o}#Ib3@ggdwb7+Fn*=6Vfh2Km8JE{eB1XYb1M_8KG{)o
zU`<(*%>VH1{8UE1a`|{wUf<mH!h8Okv+rK>b~JvT-89zZy1R1mM85ptjk)Eozu(^d
z^!4@gn~sMPQ`47zEFMgy?pr(i&2XTqoi06<J9F-};nllcd%jqD^l<xNbkxyzHunZI
z^V9V`e{AXN*fka09GRFHXr0_Td1?5;V_!`Utr)%C`B8RUIGXj$yc|8!)Bp2Bot4e0
zBTD9_?a$Vc3>S*m9*JHZZ(<Q&gs>10LZ(5q(}5Ns2X-ixx%0QzeCK8nM?R1OQphyc
z2Z^tb-F)Jo!2Lx|ntXtZv1l4ZQ!Y|Pa>84OgjGFfcJ_--cQL}8!j+l^Vd5sg)2|$Q
zdyFrCltl}PUJQsK(^wabJf2miFE?Mj$0-oKdx_g`8qnYI+?d**ty#??hUoQx9x{!k
z&oBMGv*XUIa||jWdUhGBm<IIugVh5oMnApFA`j840X1YAOaFK3#nWqRvb79aMDz^L
zh^7Jk+>M>H7x%q$)#GG|n-BA4rU5r<sc*>aCZ*s$J3FE+XdWtNF-LLif5{`nVxHn{
zBp~7c!Wcdgi8hjuWFr-+Hqwz^AZV(~X*ToYp3q4B;NgA>do^zmoP?s@7!b6XNQ@p^
z5?Paj;H0n|m198A_!76-aFDQ=W!Zb7=qwh(LL6=(loZkyc^R+YA~7t+3)6tbq?oqI
z$N2mfNnt4tVX#<AN@<H^M)q5zhShkX!&ptKX^S*Q^IN2c^>|^Hu%6V@7Dbqd9|*?T
z;WWn>ssO=}0va<6)u6CX^98|4DC&y=4K(~ZmT-^$W=sUfEwN1m$J<Dtg+-Q6q(vJ^
zv}7ZdR&AuydVyevTyArBAZT)NihB|=BBQv0e+uRq@+AowOJyY#(IP=`QaBQg#DJjb
z4!5~sa0}F!oMfz*b&{rOL2y!7i)t|-*f|ofv2bt_i8sWR+2U1r9pZ&2>5H+1AR+s=
vtxh5JE4-*Ud8;;3*Y``|#q@0LU+YgCEYGw&mtl-b<F1NZGs3<t-3a{)M*Y|2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/io.hpp.27CD569BA11007AA.idx b/.cache/clangd/index/io.hpp.27CD569BA11007AA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..110000ff4319187ef9467d3d8ce21d7ab8298f8e
GIT binary patch
literal 1856
zcmWIYbaS&{XJBwn@vO*AElFfyU|<jg;^LB`OdUoBhE!$-hKf0BCpl&_849$14i?)I
z<XC<oIU*&q`diR@?=#EpCfwmPUhI10|GfobtEXsbJTcL~KYQl!70=(5Es4xpJF#_6
zWSZ@hy9%%Kj^9c;J?UoaHRJ0>mu4q>+i?XMKDIfp9=YCcX>5FK`OWUO6FFTs6=tiI
z)LL&ZE4%-teU}Q`&ECcT4*xTJu=W1_<FB+=?|!<)$LVL>Io|4M*`1!xw-!&zNnRmj
z6lcNt>M;9q#svqLHn;iCJz?tdrzm9RwHDrLw(ip%InNG;a=WkPeh@7kwv~Cehl=-B
zA^ANy2a@j{`N0r=^LWJ7<1)6}rmHU5945iKqR~IQy-DK6$pRBLCy@xlxohSh(|_rB
zDThnt$4^<le=|PViak29B=+3~i#d&lDhjl?O6yEtNk>>~1r}H4CM5&IPW({Aw9WTs
z{N-WbV_@JA;S&)96AYZJObq<&j2w(omyZ@Qys}A!DB$Je72^OA44iDt5CvxhT(1A-
ze`5tvAk8N&#sMN2I9V7W3K)t~(~8wu85npCI%Z`}HN4Hp!=WsztPImA`lVap`m~qq
zOgtR?!u%3&3pvDu#Ux<v6Pv;i)!TIB1yGTqouWHTFVG}@WeFG$rkI5Z;t8PtK!Alq
z(M-_-rU=OMR`f;il-ZRzVLX`KKm)*@gy<*10FComvqe|5T*T`a(@D2nuZ2%^$E(P!
z>l}mLwaP+}B2Io@n9F!L#5l!xVIH;EtSB3s(*7HyNQqw+Ns*G6Dol~3oRxOlBF9B|
z6<NuBJeAFs_>qZ)L!Mn89w00ns_d$8o|}6X-^DM*s=x@8HkEdTg<8;+#mujES4J=s
zAJ8FwLh3Ij+OPsm;x*&7fjK->$yvT@yTV<(P7Pmdyfs-=Efk~(XL^ZQXd-kkW=kL&
z&;nLo9+(9{o*XYyIEk=|@W8@}heJe81RiQYFS1JWz`O|L$w{lic-h=hU#{4>Hsf_>
z-r;h#>Fqa~K#KJF?GY{z(-*Uc*;#S;;gc$si)Vr1AYvwB4HNFXy(n*${X$=mUYwEN
z?Zw>kw*S8cP_L|ztR&2F6She|4Dnr8jMs4!yL%S}R8L{XYr$-5CH|R~ZrkvhG|%b2
zja;BNFA0&qi;v^Nv4|(WBxu=fqbT^K<6AG#uZkRsLNLD`==98Gly5o+av{#7aJpz^
z;P1z|#Uxk<GapzEfXgzH%Ylp6O2f4kj@5(WMpf1rX5K5u#cyAIoqP|kdx7O;P9n(0
s;-chYcMyXUL_h$?#j;z?DXp{gS-H4*c{mw(U=qOM6~W|2Fu4#+0Oat7ssI20

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/is_detected.hpp.993129C9FD0FB446.idx b/.cache/clangd/index/is_detected.hpp.993129C9FD0FB446.idx
new file mode 100755
index 0000000000000000000000000000000000000000..32c81bd745d8c8f4ebe709df5fb7afecbeab8dda
GIT binary patch
literal 1390
zcmYjPZA?>F7(TuI^4@!T=m(Uv9duT@)i%&rRx*MmE^eKTZp&0~5Tk;^U>mZu_`x!g
zO;LiAAP~`IYQSO2Brb)_5y>1uqOz$dAcHYC4O5sIPPQ;YiOijQ%scm^PkP^v^Stl#
zK5GMhf1whgmD>WlOA5;J6bK<5{e{Yc#pyDHnmL4aAFHWdR@(clc3*TUHdP0H#)t3a
zp1$c@Jv`hszPje1#xq(McdGmRit*OKkW+mB;V%{6EQ#hnKEx}37~F8{Les5}2QFs%
zirw$18!ARR4*a!sBH|UEnF!R!&J>Lu9o=|6bs_AFO{#y~9`2lK=pnf+r7xu`^KW)J
zulTJe%c7U==Gmr&iJXJ|_V3)YTV_5``b#gTXZZTo{gHaS!T9nEImTKcVbzYTIm^DG
z1#2)VSu>M35!w7@qsxE!XxfMKZJ+I5|9o~h<Ey=`73shBWjn9FIC*w^-H6c{3EuTT
zjJTWS-kjodH_i+gGh-VyvkA_WjOP8qxuUr{E%n>S2OD2?W(P`e^v$YH*+9#wDSxu9
z@7nl$<6PO|{+{Zz(Y>_`Yc8DJrtnt&GxFoK=6c~|CJL38<bO*Q6wh?Za<J!nk<%_h
z#HO;D86aLQr#j1sOj~uk_v*0Ww2Cv(f_Qa&3<Jc|G%K*{ucY+KBTXWwqXl-A-OS+W
z3(G2&_ODpG@+Xmt6A-}yCJfLb_5$*S_Pp^5KXgByi^qtV6ecADNN4G5f?Mugq|X~^
zw<w4@29U0tZN>TWo@^s$pxtJL88d)%Q|_TK39L^NxfnW}Rb|yNK)i}NM_RALJ=QaI
zc2A9p)9?zSRcm7yK>yhC@?8gT>P{(*2(RX27$9CrwUheCHoP_%_vdfCmW!qQ7T%&`
z0R0bM7b^!B|2{8rddlxmIm`^6iZdO0wBxx1xr(kQBBB@cItI``x5@R<C$)D7)sChO
zf`KrAbg-Z><fNOx6=eduUgU(UyP&J?>;G;iu$w^`6O+=Uf#egx{`aLZ$APMtSPT{m
z?1+gq#hM0^>bpCRl}48bWteCawMkEoi4fR<jv{#hIKc!oEI9ylh2zUOc668mj?M|}
z*icOeSYurW1MGVAp#AM5_DMQ`-lVsIk}x}ElI#Fa9`-uqpSL7b(rVEpE&+y5s}tX9
zzuz__$HeBcc|b@^>@K?pcEQBqa(LjDz)uMzL&1jNBpQ}%)K>KF7q{JiSHbLZfjI}l
zR5V%yqS3|d)?|1X^roUnhJp=%pAtw$0$$jgh9&#J7~mWlmTUvzV8MHN5^E^9H6%)Z
ONLG3H{7$b+gV4W+GOj%U

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/is_known_at_compile_time.hpp.C66FF603A6D2FF7E.idx b/.cache/clangd/index/is_known_at_compile_time.hpp.C66FF603A6D2FF7E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b939712fe758f01cde5f0c2d2b6e0a79044194f
GIT binary patch
literal 1570
zcmZ8h4NOy46uz$%3iVY>Z)vfRzEY4CWwpxMu^P#S&`D>u!GP&P78OgODIzTnm@c}(
zaB2i=7MTVlGn(o$wuS*)bQ{7%qJYHkqb{36#8^PxVni63bN9Y-Mq-nGdG~zZx%b?2
zzW1^+EEZJ=LhH6=R219Fwg?eI;pFcwa}}u>gw6&bRMF6!)#>=fz)pL|efhZjx>S(w
zZMoyR(O<Fu=oiT?%HD~t8yy2z&0V?HU7?Xxr@S*KsyCL-R5lGdnY$=sI;Ul9>8-V!
zCEcC2O}RB2?l#w)oRdtY3`Cgzit)>fs+u>xBM$pD(~%if7UwzOsQ>ZeNKS%K-RKBE
z9yj{N2Rq&>tet(euE1wqcRFOVXZPQq<Bt*-WqiDo2^D`f{9x-t)tvo^U)$zlEuojn
znjV%oYE2~<{)u{;=d9mwy?f01?I-^_4SWAcU3Rabya3&L<>q>C{PzBEOHAhE&4ixV
zs?tZs!;cNx!f$*dKQr&o+_p8gMM*2O-n-T}plC5Z`}B8b;+;3`?W<nw-j$cUvcEI_
z>{?&R`1hOA-Ic{#TS=F~OgV>ARafdm1PXB|R)`cF4=9v$lKj}0wzoA7&-6wJlq|w#
zm6_uKXuKzO!hiVj6wMZ5sYuH40Ce_~^!xSc6=q5cuui1ocmO&_*OS!YXrH5W2-b?U
z91lPjp0e9#)nXSpTZ(X`Dw5*?=-o#I6V+qx^MqcEaGWZR;{j;T_@G5_V4$3C9)$HG
zJ;wvkU$*XJ{Y8UkC>^*emWU)A3KNm0AO<r%$A+v8f6fk5K`@RMMRPnre(z-bSJ_X~
z7?~i9G{&Ger1AjtP^&%n^pi)Kq&Nw}GL?+u0qEcAm_t+dmc|fz3BqcXn&ScJ>G2CG
zN?XG|LN7vixoSDb1JEv8fqO`>Fgt;mpPBE(iVoX3Pdgi!VBl8&PmmGwlY_QM+Rk+7
zf_MDHu-G6pu=8RB10`G1X)bI@S0IXd*+0^Hu%?LQu3oa5gI4GjeWjLj-@V2Nu`%12
z1L3sGM<=uAVogM?)o542Ce~vebyutIO9U*|YxK(?%;F?ADI9XaK%{u;9w-NlL&DTG
z5Z>n5U-MzY-pkJw06rpWY9eraQEZo7BU$Vvbwz8U6QKaG6;V@9!BGdTLDdiHo+T5t
zR3m)>)X+|%rWS)rz+J&%ysIEgIzY;WTsH7B%r8kB$Oa3XQE6NWZTPz4eDK@FCoSY~
zokphv(GLZqeQDw=w@KI_F(@Gn&LuOXwp}=!u0RA^6A?8p+=Intt=R-`{k4&u@2=2X
z+CvU78jLBB1Lh`|=AAwN5dc$@Fm*JPF+Q5?%)1!W^xRay<3vpj4wH7-oLgv(Zdbm0
i4SkSP2>c@8A@HG}U&$7t_;uE~W>j9AC=8d$BhY_iVbJ6N

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/iterator.hpp.DDB0074CE39B229F.idx b/.cache/clangd/index/iterator.hpp.DDB0074CE39B229F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a1c2cef69738048e8a49971db95c444033c9882b
GIT binary patch
literal 736
zcmWIYbaT7G#K7R3;#rZKT9U}Zz`!5~#Kk2=nTLV&6h;PyiaATqZOl6Cz{B>SdJ5m$
z8C@wRiq7*F?hZUxq^mQZZ|<7}r~9^sD?KM~WYztz|6|F9dfR;M`Npa;FP?t>E%5Do
z3CG6ut1fH`U}c!g#}a#xjoW|1SC_D4r<vU19abs26(XyuH@Ps)S=7Vp_$IML?3<+6
zuf3;#u3z%u;)Q4TzeKH7zj5-yRk?E+WgGbRCEeMvO61eoGWKP!48(R9=kM-+eV}x^
z(SL<HJ1*`}^E=y8rYK#l%kG)i#hDs&h&5Gk;#7v>%G{)6pkGQO`D~^*J1cTCaWONn
z^0M&?fe8jqCZH!77+D!>OJ2Tiducp_n~4LcK$uNf984e-bTtOPPu;vDo12Lfs6c>C
zKnzSE6fhK}rWLybjgz>(k>k(u?M93&tO7~`$}o*e_FHy-vbey7CJb{KBNy038|G>6
zPu(?f8_YdGVUS~B!fmVKCMC-}U&+M7%B#$41k=pJDy$)F4$}j(5vUjBPMF@AopKQy
zU#`B)$iph2A)pUawe@w?R`IS+tI*uF=fu^>wbq}Xq6we3QL72^mzjnpeDPXoxVFNv
zdXT$}c^#45Wg+YabJs(Q^UmMbNHC)5{WNFie*F!KDrmyM@XSdBMh-)9QF5_9hyhBV
cP{0ObFfgzqn5?sZ?Rnw5Re2jT8#@aF09O6}5C8xG

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/kernel_launch.hpp.4FE4793F82580CE2.idx b/.cache/clangd/index/kernel_launch.hpp.4FE4793F82580CE2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1892ce568ba300b323cc966066c9227b76783126
GIT binary patch
literal 1756
zcmYk73rtg27{~8{t{&<ctw7;!l*wZjT404%q@x{@BIU6THf5n~Ac7z&6tHayB{T~U
zi;jf=6`W+?><04%I3q+SaT^L_kbxtzfeJ42(9C5laf1;yw~OTVo|E+C{?7OR{`b4z
zJ@<5XqFQ~2O%SoUiDkMxV>XK*h%MC5VALPd0D|zO7Ro;9ZJJ4P4(A<Rc}MSOza%{T
zLCyW+x1UsTc@w#PE<Y5hgazk1mWu9oO%8hU#Wf8E)nbM1^Q&J?%^q3s*37>i*YKCE
zSbY7}lT)HIQP+aMA6#x9ALisZciRHGb#M5O9_$^yFy_57J7DgO`q;fb(NMq4-{qhD
znYPf)`2L<}<jI8G_MndIPadYopTeh)`MzU!)r~yAKlS60-}r3(woP!et$oPyw{<Px
zaAI^&l9ZSljWs3K54Ow~JX~^0**khF-J|Ejk*Hrj3ZE<3$L?r3n5F%r;FxO1gzmQg
zo*S;OEDTgWo;#&GQ4kT<J5t@s8+&`WOk;oRr|B!%L+mqX-sG#^`d@(kfoJmKn>+G3
zWdHo9o!sLKX%~xy#$CnSeEsRno-cJxU6rkN{oYRd7bEB1>t9<t>)xlmmwW5cPSGq;
z<6p)$m~}ZSs<*Dz^UsR3ORb1_ksx3yJ2m9_M7Xe=h|Mnm2p)x8>yr6gup<Id0>nmw
z=M&-L3^oM%yaR?GsW(GA0nUa4WxhVPU)kN9B)Gw6At)gwLWdm;Bf~H<lV&09fM7XU
z?jUvl(ccSyJ8;?wL7qP^h%rc)rpxIR(-w*DPJVYX8$qE!7|e*|0y#$UWPBhKD-o0g
z(*_7u13@*w*ak=&2&ChtTy?4M=v*{ag-9%tGA0JGK}w4(ck|pVDck>{L;<3J2u7q6
z=^|)Rljv8C>47|j5-CX~R$MJ0YncL?SN3IAaRrl<3<V$w<CQ8)-AQ|0zV|Z!$gLr*
z2ld!U=*@^EP~yc<7!1Qm3Z)pypbR59l4C9ksK8=MsKh7+#$Xf=<1s2hCEj$22%4eU
zi=lE@j!^}yz{mnE7*zvQ4KN=ofL4ITYXPjqxemZOoNWNw81Ak7ZgXJlcs<om559*V
z)0|9xrXOA2_<-(a5AWSoO61ORciY&(1c6`%E5kw4FEviIUE&UN5ZtEPreU09N|~CW
zNL3^jQz#U1OiUfEPGG25Rji>aK+vMJs2QpPsw#j*tw3P~aZJ1kL^pv1n&-)mkcLI>
zYt+Z)>f?$%B-e+FQCetPIGuu^HdO1Ng-&~ZpSI;^ssU0`8p70AK~{v&DfBA!rLUg{
zbeERBuFf3l_sO7gW5^h6W7h$)j@j6&C!{me;v5ZCql}bcM}UG<U}PaJjw1k3FP^jl
z2d_0>oZSDlx{<n5C=7L@>p`#-mb%eZ{>wU#F1z)$QTOp=c~Ym_#z0Wa5@VFb%EHLV
zGCGLR=M`qJCmHnn4T|-PP3zOffZb;erWn84RThWsvWei(>jXjlfziW-z95MI0V&0H
AK>z>%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/kernel_utils.hpp.A2D72F83F06BC2A4.idx b/.cache/clangd/index/kernel_utils.hpp.A2D72F83F06BC2A4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..901aa0396e9ad4310872f8124cc5132262b85dad
GIT binary patch
literal 210
zcmWIYbaOk!z`)>~;#rZKT9U{DWD5duaY<2T5|CaE#1(Uv_BwJk81S&vPOScJta{C(
zc_VM{)*scb^A+BtO3aWA3!KNWWRvlv%T9XgYf9EN8ohk_s3UdzBNOfR-t4s2Co6Ij
zkG#D(<4oKCg>ff)v~MZC`cYHy&~^u7(QJn$#g(}rix`Sh(~8A`^tx)V1&70OQW?3J
qcsY1gKvG~(l$w(WViXr8gVixGLd1b26VpQ3`-j~Af6ZlNWB>rba!1Gj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_add_instance.cpp.5CD53C62DE24ED15.idx b/.cache/clangd/index/km_kn_mn_add_instance.cpp.5CD53C62DE24ED15.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fabaa1ec80a2221b1ace8b56eb8315ce4e2c2382
GIT binary patch
literal 1924
zcmZvbdo)yA9LLWv<1ur_j2SbG$4nl1B}MX<S4K0&AdQKmL`+oDosq{9w@9_Fyo>U<
zb(<b;B1ObS;p#Q_S{^B>$Vy&O6h(U6;kLPd+_To%d!66+`#Zni{_MRsvzW|-@(|=2
z%8H8&iw;IZ5Tq=9MbW~DY*`3u$3jqCQ7bPp>X|kC@jFdz{hs`1GJ~##E0O(sG{!Bb
z#pgCvt=^b-FsCSge#X5u=7!)8tvr0sF<Tv@4SI~6-Evu|TlQ95L*n+lT%G=gS`w|k
z_}hbBgqm1};LE-Zn#D%j9dm~zmADVDXC&)dM}|04njuq@Mn?3$wznH@i(7_o3l7dr
z{nQqfpY|Y2rRa8$$=v&viOX}Z?j<B8cojs^`n)ce1vjzz+N)B$bw==|FQ^Wxey0=L
z53~5)rpHe?H!`C-DtO}bkY09|Vf~%&LN76Y`|MZ6fhpS*gJ3vuG$z%uQ?abmM(w%_
zIcKEd!)R<u62l`hO>g}C<spaHZG0V@h*Dj&g-vl;3We3nJeLrj5>w5SsIKX&KYFsK
zX_O<o-Z)sCSNy;qJC|Cit4ujkmT?1}S{u%h7gu~VbknPJi8!K~{l#ThG(H=;8G9<t
z-TJvth7W%eUHdbWIqVvBeypT`eO5vY=}zTTYTjnwsK6_{8E+JN9-@8<HS5*c((LQV
zefV>`U5~=G)Xto%IU5Jk^lyqiN;<WPQ(1c~u%sseYMvRpVvNn|Ze6?Hg<Lsd$Wm-l
z-OV>pcT#<<{99Q$m0jGX?POXtN}<bs#%o8ou_D+lpN{s#c^oh<DRDR%FgCKkC%|>o
zeulxcB5GaWIfNU%<<}>c{+_-KskvD}>AN)6Ta|ziZJMt<5KXtEeRf2J%d9N8>%X<z
zysvU@ocZS!OAE6yVOj{j@O1Uc>_^?FtaXiE_TvLsx`WfGi>nSFeasj?HaJRbFs(0<
zpgXo24^1YF<=ft;r<(ZcC2vAz9|=yP@JWq<96~=W(YLhDD%ZQSyy>BJt-kMBhpR^d
zyP`J>nS5Jrx%qzDqw5z8(X3)ox<P4YQggw}F5R_t!Bx7|J6Z~xx)1%jWvpLsW+D@g
zc^dZ4|9drhjel?aaBb=J7p~=z$CDL8-Ck74l#U%s`J!g;d#i&KZ~LLx+@6u~*|EBO
zw)g!*VZp5Ujnm@;cf*@c$KFWhr@U#RAQ2MwHjgu&?^@M|tB3=%>5jsl29u#XVQE}@
zx=Y@^MEMNHKwaX%&2qDJ+kQB6(h>9FJ->-~G2dz1AgLp&Rqp3J{DcD)|CKz)n@%E3
z>YVpF^qpT7n0aq_-ke_JeD~G2zPO^-7SDIyO?tKclEJg5&aAZA`P{j8?#c8c<3Zfh
zuP;NQ_{b1T6a>k|e^hWur0vGZc=<yJRt~GQI7sjA;-kUd<^APvk^qKy{{Kc0C5)2i
z4;5N<^aO>hkmUzMd?2e}RCqvM_sTrzJ}sOKLtCUjr(&p9ivy5%*G^2xqDrk`$OmNO
zF?bIk|H8W?JJ|cA8HRj;OvDhqfZS-CTeJG*{Y)7012PFiTB_Z0g-(qY?5)DdcuQ;J
z<Zy(=0c1-=%8sVWaqcj*6=ah!Wd4$z5b8H^vbqO`HUpV}AuP2f3=@c^|3eRjU8I)Y
zJb&)$fZYl}O-6B%76`)#4Nc<!D1e>Rw7_iH#6UB)027#{1s3oD6CfweT_6U+FrtQ5
zV}l&PPik7=wrtWsYjA-HP?VY$=q;PH&{|7f0G5i_MTW~JJO|GMy}*bPSBVdn4kH8(
zfhPeN5u!KIA2b9bBtKFRs0<_It>oaP{XkUwRDu8-7@=BIok6d_Bkgrzua?`Oqv?K2
zrXMAxriC3_HmRf4m(a8waC@Sn_|yvpBRCRH1FUb>Jv2Y^4(*8yj9^JvO>j4)CQY0^
zFbTu<1}_{Fk#MI-RXWQe7N#P*zueB;M+HCu-~d6;&dAWc;nGhJfu}4QBPXw*h=cJ&
zf)Y#F%~f9+%W@&9l2z1b);4zb4jUYuOew2ubuA4U{}-idXlkv{*3nyOU}S8vX06$}
Y^;UEX_0{GZDO7iMpbV4iya~nm2YH*>4FCWD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_default_pipeline_v1_instance.cpp.2F582983A57414AF.idx b/.cache/clangd/index/km_kn_mn_default_pipeline_v1_instance.cpp.2F582983A57414AF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..64874d79d7aac87aa6aaea98bcfb6481178caf5b
GIT binary patch
literal 4048
zcmZveX;c&E8h|r1$c=z3&V0)ZLO_wFy$CJ{2r3ao356h_Z50tj(5Msy6u50&uUg=M
z;!?DA<7zKOjk~v6H5LVNMFADcajC?mqTte4K`PwfIgs~9Ifvvtd7pXm&6mu{M2Cfj
zUbK<Pg6D+gq$X!2n#p7`JL7+5R)!{-lgSKnnJi~pRcvaS<ZpYgJ7$1&$)*Lx?dzlF
z)%x^rIevN4mByo|@X}w8o(XDe$iCXPX3ey+(PPRLp9TJrld>Z>lZleG9~l+zZmYa~
zbZ_A4Gq(Lsf8Sk2vKxFA<gXO}9osza1#T>PcGU99n@5Vd`Il~OZ9Z9W>+X~NC%*jY
zM91qL*0Z|eU%n~p9C|kXU`kSSdROA{$IsR^>@0b^^NWr*o|>g*Uv0S_kWe*0t@zdy
zce^)5K1<iFuynhcbh9`t&0bM;!Rh|&^m&c{y%jpVb&m$=e(8GjZD`s``8qe7w<Y}o
z<w?JX|K#~|h+CGbB%#tg;_7VqxQE>iYpP=xe46}VJ?rzL)KHPXaL@Kg+uBW8OB_eL
zXiAp1l@>Ieiu9ay=3kxWBM$Gb-D)nyo*U}<<;4Jtj>5xkcBt}9<Pr1LKg^4G5t_(H
z<%b-apb4?fdp_Id-M9I&h?lp^+_H0=io#~N6l|IB@kX0P%iRMmUuek`XCL|cV&Ijh
z0UI_iQ2M=a)TRx|TB`o0*194;=)Ipuchzr+ZY!}We8SMY!Lg}dOj2KLq2({5=8v+-
z&>r{x&gy!)YIujCG2vkLCULV1zcTHn?ozUBroXyhYgS=YOIDGcs`T0}IUWA|;>Kwq
zD!$NLRz9HUb!<ZU*bvX`tiZM#pLGVkUwa9;*yK)KH$A^|=G~<?Z{AMbem|hG;J1t=
z4y{bVe?l6MCFe0s?jH?WK6+qnV9mMm1I@ucJMT`OXCKVhd%Si3-=Xl_85zazPG?-3
zoZqf4h^&+6er>2L3ktR`YA+6Sh`uq^He&nCV?`VMyLP_&p(rEB`PrbnxQCu!{Z3KP
zs#ym<{X?IvynknJ*EJWf+=hgt({~1rI<&TM{fzG-2hTlrOS5Fka<5%sJ`<fwZu;Lh
zpRn0_<n0UV6)P^G-42V8{j9=ap@u!t-6<#2mf>{w_&;|HNDH^x@M!zn6`kg%#y)JS
za`|!j{`u3Vl^#Dk=wZX1k*fndi=T`f?|%Nut2yn5wU1_0YxVCBy$bc+^5owyE~pQL
z9?@8Y^3^L#aQwPqlgPr|4*M1lNgjR{&4{a8Svj|K$NuuBZkx5mUenG{%jr|P6!9JE
z+``p&BFmomY>N!2u78@IHpWw>O26Eg_Izr5(u+_ZAN|{n9UT>=&K^4UYiIAjWSP0C
zv&R{)y7?ztTRCQ*uazuniY&&lpJVuk2jk?!-}$|cI4fo0F-ul$99x=O+I)EE9cLt`
zWF}e4Waf$!MZu-MPPSGtypj*;J<3QbKEgzTQUEhSC4d=HA}~WL1ZG%?!3?V~m=Prb
zGom73MwJxIsEUG_s1(6WREc0FDJ3wIR8rr};@z%$tHOO^sPWX){B+Z)@p1fIKpH;F
zM1ooXrVu9pQ_u*&L}~;~6o-I`Gzgejje&{dFfg$O1Cyu;Fi9K%CeaXJQZ)r8jibP%
z8VXENEdo=F6M-peL|{s437Arx1WZXI^_g}D^Q6TGTF&=uEamM@o5w5o!GK(O4-*Mg
z0H&Z2fGM~Nz(f=Q6Dbfdkt+fwrWlx5fq{u#F)#@wz$6L+OyWv_Nht*;RZw73R|-s#
ziog^VA}~c)5ttH{fGH^?U`no1pE<tsRqKzBa}RnOH@1S|&oQR$<Lel{9&$6o>mh4d
z{s0Sn9m}7ET*LC$0Sgt30Q*86Bf!4U%m}bAXjuXF1syBEzEHyou#YMj1pBCtL9maS
z83g-C%OcoEIu^k`s$miA;|d1DKCWXh?Biwz!#>us81}J_#juZSSPc85f+4U^>KFq1
zq?sYGPqZw7eWGIt?2{Unz&@>DDD2ZZhQdB=W+?1aElXja>R1Z<w1%ayFIF%j?2C1b
z2>W6)Bf`F@WkuK*b*u>cVht<8zEr_TurJjy66{OOj0F3VmX%;%(y<ckOEs)y>MPei
zmvtDjni#Vf=V8viy+3>j&EL#5zwdNUZoEG99RHd#ecuH=CqPDe4nf9x4nrn-jzFe*
zjzSjooCsOcbCQW?j|e-(C+bf0ETo>}o8Us~IRP@N=MZFE&tb@<o+FTHJx3vn^_&P<
zs^=sV_wEd1@!?jW@LcDv{SN}17MaPmmV0;J=viMS$JfI3RdNDkRLLR8xRS$=NhL=h
z(@Ks)7ArXsvQ)`QCN9)(Q+qD^Jh^8f29Ey|F2ukIkdc8ykg<WokcojKkg0*AkVOM0
zLY54iWa95X>7TLYe0*llLgYLGUmUpr2+0v3EXRO|oB*P73P_ZTfF!xpNBf?w%KiMo
zwj#zV#;U)Sb?+OWiyilLmZG8B__DK^^Wfgkv3JZ`e7JY({XC5lxD09*PizqS|Ee|~
zt1wdx=smV2(dlQ;QLQ6;PSKAK=uKJ02z~+}<c9!ZKMaWY5kS<B0uud1K$4%-N3Yt3
zx-T2OtgPpabNE%@S4qebAfg-u5$9ltB!@txITWIpBSMsNBonqznp2Yci`dY!A=V1o
zyCLswryaSQR{6MRU0(bMupwRoAmoJrVJ{4bco9I<ivkk8L_m_4)JL+6<b{bn_uZKp
zNtwj>%*A-N1y|>tdhNkP%c6Vl$Yrvm)Hw_188JL1f-~!9{ttP7ORE94Hr8Qw6N6pt
zEW#!XB)mXHKmUNhvE#-E4MQJ~88X^k@prFc`#}y)&VyZsx_Nx`PcQG`BYk`wM@gO|
RLJ*#$ie*D1m6MrI{ttg4<jep7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_default_pipeline_v2_instance.cpp.FDC41D8D362237D7.idx b/.cache/clangd/index/km_kn_mn_default_pipeline_v2_instance.cpp.FDC41D8D362237D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6afb5fd2b2ee078a1195c978618aa5b5ba58d486
GIT binary patch
literal 4070
zcmZvee^gD`9>DK8C#GLg<D9+s*?a1y=m#%XgsUib=r=OeGozOx6%oy=aOLMrM$A&Y
zGz>44G(A63Lxh^>2l+AUqC1OtCQrqqFi&L|<@IjXy6*Rn*IM1XKKJ`spR><i_ndRX
zg98FuhAI@(<Ab%!6Vex&DHIAj!+%<OYEpurQ1qB96x#etk;_x)WZRbhh#@w!uAL5A
z7vxs)WS`lqs7FN!&+g8*y7267jN|Lhl%HR0*)q3eqW2Ny7gM^mi3irCadQ-JPI|?-
z*=lZ|Iy$Ag%63ro!Tzgw^_@uvQD5TZ1NrVPzPt8xp0c|B>QU$TOYI#qU*5me@gj8Y
z{hwm*uX{Hwt8Amko44KXa%0OrKjrDJ?mrvV{i>u>b-njSW#8M~UVgl<?%($1ai=S<
zbhzbh&1j?-8ZUk`?6GIZmBW?UJK`4YOz~8ECF_?KeZQ)y!ZLP2e`;vaB+cf=U0J=?
zd56uOJ;nb-^A8R4E)E?Vm+6w|X{PMzx2a!r^z^ZlUj}`(@AaIw3;ZWG&34_oB*fci
zD9fz~o%Z^s|MU4JyUSvJT5Gkp`PP$}!jMr9b9YZ|xBsT^a`UdOmaQ?a^D^tgD}GpA
zRBqws@AO5Xm($i6tA6}weW?E?7j370m(LX2#={=D{xb`|yPBn*mS>-^=+!xE&%3_y
z2Ojm<U%htPF@N=r+m&|S7hD^bWG(P7*pQ#szu-UR!!%aKBMWn9*d3i%S>zhOa(krK
zZ2}#1?sgQJwPmro`1^Gk$>n!^u5Ay=wWyu6^g;W$O(*h8zFu`cz-^^>qj~Yo#5rc$
z$*T6?(N4wN0?4Lc1_!PgRsM^^q1`^-4y%_OUtjRw`gY6Qtm>nI9+%#oSmC#HY3CW;
zZ}!jpznol^^ZO?M&zzF)oorAn7bTl%Y}*ot#SbGte_6Qi#)=UY*@fGl=B*X~(p+5a
zaV}_m%(=u7{(V*cu1TA#eKVd$6ci2#><c<z<NjpS)*p0Zq5`je(jWFNJaBTs&XqYc
z#&jXy^zD`fwW>SE{&-nckh0>>{fW7Is+W~}#;mZQ&Gi9SJZ8E6ZT!6suPV2T_~zk#
zhjxZ`<fjOQAzAXWMOSuQZ2$IF&e?CX%L1@-M);A9r|wt!S*k_V+8bLcM-IN)rR%-8
zKJ!<H)aP*@M^2kN!Q<G?d$BfM%U<lQ+rRrt_g-b*!$;b2ea*MKO51L~+|aBoVn3(V
ztu8wlogJ?ix4YyFQh8L%^K-^D2l&kG47e7zVf2}^hX-xD@#)lg5vTi=HqNofGdxkn
zp+3Lru7h3F>fHX>H<qdQugWRWKWs^l%FFv#-j5HOk5yl8$zHnj@^7x)-I>#;hxb{1
zecsb>wXInm=WDp?>W@00^u*)}D?aQqMTFI0tD)~-43qbN*Vo3+eV%uw*;+BjFm@c|
z82bLjI8IGSOv|uRC<ZIf=1eb~<Y;RhA!@`K#;Acp#91biG!mFeArhD&4FWS1g1`)G
zFqq*G3}!?_z>I_tFq1Vhn8_hBm{AP{Ga5p{jA<B{u@LrQX8E?DB{OtF1Tmaim>6X`
zH8EOT3@AxVH<1)30aJ>WfGH(Oz(ipPm?#<n6D1*F;xG(M9F2jAlQ1wz7y%}UCcq>~
z1ekJ|3`{v%2Bw@O1CxePVA5y`OqxW2$-)>gSu_JCOJW~PyRvO`<%y2^fsG|1Gi{!z
z6rBLMi0&qmhy+YYDFIV*k${N^0wz)-U?LX;OiVB^u@VClyI^1vLV!t>1enBy08=J1
zFlD6-OxZ;SCM6V@R7rtJT_`XaVZdZc22AF{KA5c6cXUQ-Co(r|Y>y!RE`aUp5hTc{
zM?jEqkANYQ9)UoXdjuIW?GY$s)*~<z`}Mwj{*T9N%ElTNQp<^ToM|Cq6DR5+w{l_|
z<WgQd!2`dB7cW58@!}1@QY|OJzSP7?urIZ866{N*yafAF4KKmIq~j&nN3|S+ebmGu
z*hj4#f_+rVBiKhZJc50s;}PuRS`NcLZsIWP<5mvCJ}%`k?Bg09!#>vW81_jmM_`{c
zaRl~BD@S0Tl=1}jNexe6pXhi3`*JNO!@k_a$*?cCax(18rMwLLat$xTzO3VA*r&A|
zg?-w@QP`)g9EE*a%2U{<H9UoVs^cl_vs#Y9K5OC_?6X#m!9FYH8SJwfp20rT@yyiM
zWIa>-(UX3UGmCH@<2=duh@Bq(cI0=Ly8Ain2pcmScjF^=LF?AVgC97y3|zl$g4hMG
z-!?&ljM@YQ8Mg@-GHDYCWVua{A=5U2LS}6OGjZIw;HqH@Yt9ZVL@$W<;6n6*1R3cC
z1R3iE44LQ!0$J7zGGwY3C}gG=n2G;trr3LAZ12s1g)|7_Ex3>dL4u4L1Oyp32pBSH
z5C~+sL69NS27y9m4FWUq-)CB+ZmEw+8(4_BD4Fh0lDPy3nIk~h90MZe1dwbl1ES^>
z5Hn{V=vZgwnt2cNa|~Y<YYS@|<ALiUeV@iE?=%`db~bZ%GJdg)W8U)pPrK~8-h$%~
zeH;xE`1ToP&hq^aJfoyr>3>JGxEp3j4mTS<#AtOabUg3j^?dxmnX5$~;}NVQB(($x
zsSzNo#(;>L0Fu=*AgZQ-n3{c{moG-TZJfBVWZ=MBF%vu^B&`GyX%R%M#Sn>>K$NvI
zM5?6_nU<OGO<??<<rn2U0~=y9oESIco$cI{zos-i9$1%390xW;B>_Sz1PH4zAfh6G
zWR(nvswg0)Vjn0qA$j4zts*UTQ5rV93laJe^+k#4ufw=-MY!=MqEIYa9-q9#fT4-A
z1hYYdKQgzlvL0eP)F#-@f4YmEW$>pT<6$C^)suauO!f7fHbxrl?L5&<`Tsse>=S#(
l;Z7q)y1I}0lWMHz_z9C7yr{>x84?N%iR1%jYl66${{z5O{hR;*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_default_pipeline_v2_opt_instance.cpp.9D0F889F526341B0.idx b/.cache/clangd/index/km_kn_mn_default_pipeline_v2_opt_instance.cpp.9D0F889F526341B0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a4604b8d50abae449e9a299ac2e62a7299f56299
GIT binary patch
literal 1768
zcmZvbdpJ~i7{||C(2Q{z{TR29TO>lF#=XgCLhf5cE*UennQF(#NSocYrQB+5sa#r7
zX<U+)M>~XZiLRu}6f)FAh$LNxjXmRej>jK+p6C3Y^Z9<?_dW0LJ--t`qf!UuA&3$~
z<FV=7C@cg)MEH*9a^hk!5QN1+5U-#;h<%{XUiJBc`J>GI2m6%1QC|0!SDV@hG^E?Q
zB{ahaq^!#<MpL7wSGrD}W^SaL>#kwvj^xQgIXBec{T|*WH*82(+4R)>7BCexzb2wT
zTzn*SD*r8yBj1p0x?aQ}<&NL4V>C;~2)rHDg1MYu>s2!Ht5Uj$#v>FP=Z`c=MB~p`
z-q48t&V7yZ-0$?H1!-=CSZ#Vkdxx|kuwvJQ<xMi)q<EnE&Ig-=L%%0wl5I!XZ%STh
zN_G>JmHKLW&8`;ORkrMW(KOKW&PR3VW81-$kmgkD<Bi<DVlVENQZA0l>TX+ie2_9P
zj^cUO9tze_Huw9aw_U%YliE9qCHl}9BT<x`L%4T(f@`<OBc;O+=Mqn29JLEqcNeKy
z#l0+VFghpJN>=lC?;khb=eH~U`HUSt|Ca#c0E;tJ>>r(0U$%<_h()-8*_o|r3h6d+
z88O5_CVhIf^4fdZH>+pEj-O|Ju<hc^uI{{7YfzAQ>UkZJJn-ioW?H02(b0nD6l>m#
zlSRjRa%yt663guCN>n<>H2G70Kic6rm9^G+gsz!0(bJ>VLcPQ@T4#+Fuwr53T{;xL
z*;j7j2GuNq+_B<HcPxaa5*(v@(tqm#R87EF%AO92<U51V$#=Z|xV;mY&-$i2PdiD1
zor@)3(*)lu_?3O+ro<a0;oJ9u8w%mTq~)rm7onT#KUBmtlFRD|+4L;<$hG0Wb#Uyx
z%xLn#Q_+g`XX;Ey{4NVo=H(M9b?=O`PK!KqUF5WT4?fjTEv4WE_+tJBk8!gLZx;9w
zV`d8A=A(-z3=a%ZNd73hZ*LE8bUjL%d006n_Kv<=FR%0FZEg@{<_3}D^@JwWVkhK<
zgF!`RfX|*2IpmQ(x6muCnYA_&&qz)?(=)97&d!nyqBH+#Q^;iB2RYqz(QQU+pIT~6
z(#X$)jl%sZ{~WTxtu-jm%ME5uWE3`JUNrBIJL&Y7)Lm;&_L|mHY)dhD|EdpZSq=iB
z)0!h+bGJQN`D&ZJk8nkl&`Yr?%gU3)Dp^&YU_`fRQr{iclhekSyja11y%3>m6j5>5
zMx&<e--z?ll-TTJ{X0rWJ{Rn8oF3)gawya=F(K36lfGD1)g+eW<xP*LCcRXB+ziB!
zXqQ80d=fj_1^(;OgPSpuUQr7H^7DrRwdAyX7Z=!ByjsijG`O*>8b!1(Y%9tusw^&u
z&7sG}C!--qUjJ4$<&uK|nh*%w!B!*-rtQE7A^dF7K&scqKnyJVf>1<;DGV@{Q5@hR
zM65cgop61uRhFU&bPyRPePFnZOo2H<I&&@-hdJXKC^%>m1Ew&D;X3zm;@WUhl{{<+
zQJ@8d7%=ysz?0?t8U@}g=SnoVj$Rhuhz3o|`4Jj)EwOvrdq_IN?L}b&b**$AkO@-)
zW=+nve_DVWQ^qQrBO{D_eS#q`HW-lg`w9czFZCO3ORHZQ)p$#0Ji&n8CF3V7bb)U@
z*DW*JFyQf$fxHQYU;KtyEptT`LyiXd-Z;+j);;mE_!U4M*+w~hAVP);vSFEkBPb(s
zOOVTGgubW)mUXeZhDc4sM&lO`08$$cTDHYU2_7PVd+7`3YGbvPkle8-1A(R8#0FUd
zn}8!y>f@w|S&Ds1sZ8d9)TMcU^NcyiZq^dXswit~B2~Vt{&4fvfwo~;_<CTARDn2j
zR+Ma*@tnQ!B-ulNY+1Z8fzt1(j~B?$9aI3cmmR}m!m?Ma4+bkIudo8ANKjH$QKk_+
zD5gX_ZHpRdCD7E`wAs<g*~Qh(O5K8NWN)Vbf2xL#?ka=ThQ=mq%-614Z*8+-ql2EE
Vwx#V>bq&w$L1^j@_iZTe{{W_L+;ac`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_interwave_pipeline_v1_instance.cpp.536F114F5BA9349C.idx b/.cache/clangd/index/km_kn_mn_interwave_pipeline_v1_instance.cpp.536F114F5BA9349C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..da0a6e851430b25aa4be1c542d58c40988209947
GIT binary patch
literal 4072
zcmZvec~lff9)PF22OI$fO?Opw)tceZ@rW{-AYeQJ<rWlNHX!P%AfONwhBzE*7B9pj
zteOz7sKEdc2iD+QZwN3(4F(l_YrIfL!Sz5<69h%w!S}}b{jt2q@V@zd{dITI1rrhE
z?|;itq40|f(x=4h<|-8mML+X*x-Km#QBWwlRSJcE=cySfsn<N6{_}Qd!-m3hi~4l>
znXZ-{)QmcYx&klU4RSq)H||SLs9qirlFts4?RzFo{H*mx-SO<l8*L_Wx9S`9dnYNQ
z&+mWM{@7P+U;YNqxjyS=PHg**7JXW={#EGg^7HP)FBewt8mjqQLjFDXsHM9rO^;uj
zTI+Y~I-kABt(bU2+x2JGXtpKyxntSrIV(=QIdS5xuD&8Q==u0tbve!4M^V}94{Y*X
zSGx4T(`<)W@3|iqZ3@l0);Yw}whOKFtMfE?<|hrFTGlRgR2=rm`e3qtT;`GL;kSH<
z-<)G97YBX1H}aU$%`<Tx@@sc_`()&};+W&Z?AMB=nT|g=M#Z_c>2mXm_f<asmrKxb
zrOt8IrRI!<U9G<zKU}bGz{UoT3w!#zCtdMto_?lc!}xWj&W|%Tj~+67%AzbUyB%Lf
z49t~+-EsrtaQ216Gn9$BhbN3bS|=<Cz3Lim6l1pK<O!3nO!ti1UGKD|&95~&>}J7%
zSo_+{7m3T*-sjoA8@^CQDIc5;>eAXej_r47T1@|8LE_$p>0`5eqfQ5|UK>#~y<m}H
z^~E^-n#laKC8uq(PZpgk9k|UY=ZB7|?!te6%^a3kvS?LAN&1Fqb}6<~e$Ja4@ZBNT
zi0z51xWJUfTTA$mHkXA*i|#C0`^(i{4<?3|J7=CL%eoYCJ@rAy$>M*1{;<2bX121h
z@?oOmXX2pR;t?SZ7yPH+Rvh|gUh5JRmLaM4cRlj=y+*cw8+J6{;ObewUfaC>(!H-2
z`9~O6#OOBrpLtkRIKhTqU2Od7?)`$?^4!D+kG{{?dhY4lOWKUBf?f2zz`L%gxBX6+
zDrk4w$?fWD`<Jazf2L|O&mOmBcUr#L^N;b#tFAZd`j06xgjKHRbGWJdhuR(8*U;tZ
zv|sn_LC-bas?c)MmRI(B=R)16?+1VV(yM;itD-Oco*sLVte-irVEB_0j@P|@Z*#fq
zuJN&n`)kurc7|n33JQ-V*e#85J)f|A+M2>UXWQq`tUtWzTJkp|AD!x(S;Wq5sa<+!
zaDv+v51XQG^J_xvtG{v&z{y1e$}-&IM^#HRV;l302{qpx__2B7`cr3c+3XEAhM{lO
z)!n^b|Kj_3Nz?P-u##O5@9U0u?LvbFjfsDRzZkBqTw1i9K6`lfL|jhdjQqC>jfcmH
z^vslb6U_UL1t(1F{;mH42gS7Mib!W0=c(^r%#(Ltb^AX=x&_IH?D=r>*xuIO@!gAM
zoEE<zJ<DF9uvH(~<hR$$*~uYN^bsdpjxtlQIMqs$j|65?umonv2Z0$1Mqq}0Fqq+B
z3}(cKfEfuUU?%&>U?vC4U`Bl?n9*PgX3U3y84G6bXLc|5G_MR99ZAf$7B0qEZ%vF9
z696TNIx9)x5-_D$37Arn1WXi;fQe!eFi{c$CJx8I#IYEdI0*xjgcD$rSOQFvM1U!W
z%fOUlWnjulGB9a41tyK9z@$kOm@J$Dlf^P%vLyE2>{qdlW|Tju>seSLGV9`rYS9&t
zMjUD-iAcbd)DkcyjRZ_Y5HOJ%0TXEuFfqZv#A*yotiiw}gaDJM2{4I<08=J1FlDt2
zOj#oXlM)I{s;0oC8VXEC7%-Wd0h4Lid(*PPIlFfodt>dZJqK=&b)NX@X<KP)X2nSJ
zde?AbEoWU-v5^y-Ae%Vx4rBu_mh-@`;>Bvn4ZL_BuvEiIurD=o66{MRPJ(^Oz)P?%
zRq+z+OAWjP`>2LPu#Xx!1pCOuA=pO-9>G4U;t}kl1|GpauHi83<3<j{J~nX}_OXG-
zu#c;F4Ewl&$FNUoI0E~mkt48AOdNrIV&DnvlPaFTK55_y?8`Nr4Eu5;C&RvM;$+yD
z4ZIBdauqMbzTCjeuup3^3j4H?qp(j+9EE*q;3@3WDxShVZQv>Fvl@=UK5OI{>@yR`
zV4oRy2K%gvXRyy2cxLVUtbL^D`a^e>Q%1VEyLnl*%im3T3AQ(#n^orh(ISX{2-fFa
zY7r#Js6{}Kaf^T<lNNzMmRkfFGHnqkWY!`uD=$$hw*5Hr)rFo5F$&^Ecp*kXf{cs;
zf{cv<hD?kCfh-#Z88S5r6f!dk%*q?@><k~VB09e3LOKMo6JAJ%AVEeQ0)mV?1PqyU
z2n4d+A;^$vhd?2-4uM&DM`-V~HFdModoDyJO5hixk^mtU0)$l<5K$38vPuR-RTL0Y
zvG-K=aAjun?VX!Ahe(Is4t*^rE>B-vnQ9k5+5C22W#1u|&qlQ~CfuK+zI@vJwwuz;
z)$(9jj^gd!eUh^LJvK|=>1)WCa#`$sCER?hi_)c!<=6*vosW(1X#1q+&b`I4mXt%J
z<ShY0-Utx(#(;=70VI3NfT%YG#Jt&idir?KkQJUQN_tML7gvIR29jQai1Y{|)?<i7
zPaw*A86wqFh)mC{*clL4nDVoHxo1K8x)95PympHE=|-xtwdcCD;>Tb?v=ShsMS!pt
z10q@iNY=`LsFng^TK1maoV;e5;n^0|bK)jJybd=*lORDxO#*_9n*<D*GzkQ<+$6}5
zX_G)9vnGLAIW0bUZqE~!o;E)nn|}(C#Vaq&n&kZ?l8aD8Se`|NVtz_oa-tbS7EBeC
zHnzP~z3m<PI63wW>Nmwt)6Xu*w?F<sB(nFIvE#;1m^jH@`sh<P&mrpn_by_88Q|>V
kI&je7p~Hr2M~?brwAY6obi~J#B@_@mgZB^f3FJcm2jYV6JOBUy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_irregular_default_pipeline_v1_instance.cpp.5937911B0257B674.idx b/.cache/clangd/index/km_kn_mn_irregular_default_pipeline_v1_instance.cpp.5937911B0257B674.idx
new file mode 100755
index 0000000000000000000000000000000000000000..001ed6cbef886fdedb48a1da89ba71f68cd9a46c
GIT binary patch
literal 1750
zcmZvbdo)yA9LLW*2ot8$Fpb2Z^2n=%jJL@M<uQYjsPP<+p^VD7u3A?fafz!^k`z*i
zyvd|4iKZA*4a%U081z^YLhc!B9m^m0tabKTpYQkk`<;FE*_&!lp>&A@0NLL@ni&v9
zM*{$`0lHaHk-?!T08o?!fau&dH)e!_30AO7la|T87@9ek;lk)KloZx-?Vk5lH_7|_
z!PB$)!ax1uOJbsj`Zm&C<wIuE(Sdnv7QzK+sM_tMf%SP)U1G+0gcXauvfPG0%{RV*
zUko(K%heQ^XJt=RW1i1^QVvRa_WJ7k#?;q?{IZ6H`wjCyGi8Q;`+S?pUfA}?ry|gg
zdi-~A?dSxLe>8jcX!!h$PH>!9Y!1)#SX&66SQ_u5I<_TrP{|zcaiu*Ul~o?>jolm{
zn<}_PzgC#$JK&IoYi{{HbD0vs`F25Fb~angOu}!>v3_$+rW50m?!U>>IT^NJuk)5o
ze_W`1Q^##li3t&q$79v1GV4e-R<xJ>q`r(oCi+2=5hKX7Hg<Y+$+Y81IweSf;VQYg
zc2Xrs^P}w3^u2D_pCqmV?y~AtL0Ze30$k>?gqIR}s{IK{QRh>Dd#BS<oxgmuQY;A1
z(6;$uPx;&mxSqOrpnw-3Q8`W?^-Q<SEDJ-@um2f#M5=q<T!rFwjuI_wx1ZOGS!X&@
z>v9;>vd1+$G;T4!NlCIwNTL?FW`y2L=<$zEq^0z<wafqUu;|5|tyi&&`KOq=s4*@{
zi%}e!KrLoTxJWU@6WY(l+2z%!Qgaw{h~#HpT$(j`N+qPR;KPZe!~wR@a(_9I)KMDq
ziu&$`@Mr1m?M=1e5p+FHXaC8Xb=t`wJ&OyP7ZSA;&*G99N<nn@gWqrF@WS_o9VTz=
z)tuF+yPwMTip>1QiR|1bHRo)zBlujdMa(!Y|K^O{;=VkYo(XPhU6<}ZMgHl_#c_8F
zCS5y7cB*(bMwvtr>mJUYc)d91-ECZ7a>n49qrk9llX^W`dXld8@$&7g4#MEro+gdQ
za@|jJZe=(Ma~<<sikAK^8lG0uZ}hEEtMmO(+S8Vt=O!G|o}J<$W!feL)gw$vpDNCD
z7mM#(pLFCcR!lC#o#`uQ)R`L3d0|d#=mpsp)A74Fiq473Z;fI@+2lZh-=(q2;ibOs
z9`#q$6Td!ncCxEtVkp(^oGfY17VmW3K8lyega_9*Zbw~1p}QpSFAGoF;BthK+QG7z
zow%C1Ma#u${4d#rcTUBh<?@`Gu)X8QBS<<94l&%mh)I8+$RImI!?D>ccITlU3ArIe
zTc33qz+y9fEugd4Ov<X)8Sc2S9&m95X!2t6?kfZ2tX!=nO7c;GVKo?}Gh}1LF|sQI
zY)1wJviy+%Ag<hyPQGbOM6RcSHlQtB1(74@1e4USd_laUAq@q|E}$1IL*xxI)=&s|
z5~d9m=M_$t3tL4c1W*x{AtHimYe*f`f@#wy>XPbAaS8$ood(Y#VBK=z`_!FLc7-^k
zw;;gQRRiK)1o&z#e?WjA*K#=$tU$tzp!_2w_;@YzkYM*JTPIBd^FN{n5NH}fmtYJp
z@Ui#FLE=M1foNWzP~Z;~+|PHHebYvIbA!m}MS%mW#tAfV^{)Oxx5#*c0((}Cv=6z?
zI%hlsM8;1PIJatC+b<cJ*y_U)84{oz{D#trc(&(AZaO4m@R-$AglC43TKVN^EerW&
zBc{X_H;eJ3%KUaHO+XV=;S}5qDYdc#;4~Ck^O0RAZTUFKerKo_kH$;GwL&`afzIwJ
z1JS%pKr^`2oSlC*^D%)hO3UCk!RaNeN7Z1&lTlGxA2fi|kpW?J(Y|9v`mylP6Gbbk
z41fCYt5OMq3Q*x42>^af|1buGj)4bIXfg415|Wtp(pXs;`wh0_?Hi=*Z9x1++$O>v
z6H_yD3rj0qxg8`mBMs&MQ{^`+ZXv3uZr!G?sim!_zth0bSZTL{&aS<3@^%hxNQ#s7
IKE(e203lr33jhEB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_irregular_default_pipeline_v2_instance.cpp.0318E39A03EE883F.idx b/.cache/clangd/index/km_kn_mn_irregular_default_pipeline_v2_instance.cpp.0318E39A03EE883F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..54f5f1af2c7da54642fcc71df24bd32aaea930ed
GIT binary patch
literal 1774
zcmZvbc{G%J7{}jPY%|tg%P_W-#?n|qW6&UG>=IJx+L<(2%D&#*OO_6D5#`FYgo|#*
zmi0uIgWQYKu_dl9iXqolvNZP{=e&+T?t9MrzUTA(exKiYp69%;y#<Z7zzqN<-WC!5
zJ|P}x002awJ2)ih#!VCeNb><eM1GH>f1nguc$)1XAZYWd!92!Xss2+Qddqpf#AkWJ
z1>3$n;VS=?75L}p#6%~8qF#mE6@!fkdO>6`!VcK3)pk`9HXg03G>B;u=8DN@_e+G2
zk=(@h=;VTY<!QsLTvi=+U~B%JR?pm$_1aX=(&uYVQ?Wlx#js72R__qEcUb$GZq=9T
zG?jJPZO$88xvao}-Rj1@ogD3UBxUHM4AZlrv2RH!J2|RXs=c@4F1Vn%)c2$@?Y^g1
zMxdsyw%-V&<YCyGdW@R~JIbnrWSr2Om9#s6lugjwEPW^LQVi;qT4{L3pQ3A`<<{AP
zU0#(9RkgRwf6Du6x8qJd`^J`#>}#o~FNDvmw=()VOkHs)&k3B4i4b};{`!+ON<d{c
zGw0l#%-y}0<5|g=53Yo(@tyYdW#J`EJ|(Kcl_G70WYe%3o>(ht+^GoGNS3~VaPJdh
zCe^y=L4T62QMQba*H$yX<^;vNV15(qf7KwLAAWzdQADqu(CwS#PA&W~Kb!6TFB3Gz
zmMRuynu=7OZ7d;pho(A4DCtOXHIF*uY!a{QmOhM$@?(zaze=^t<a;T^n3+2h_bk8s
z=dc!<Qm9@pZ|N}I4xNe%o3l_=C{3Z^;@)zbMXE60$`<G7>&b@uK8-E>-Zh8GOlqq%
zBlhf91hg|4tmlKPGRxFk<ks{rany?nev`G`fWN4ipb<XBSN8g><K?xYyrF=T^=U;Z
zi`mhl*TzfRh|T7)uFdq5)V)?J;YLCmC2G;3ut>mc&%8iT`IAa=*`S)US-&*<!oIy3
zxiBL%-BfK|oD!0XDSW9uR`qSIwJ<Q?5}Tfx+jf(w=^DT%HQq(*Biayt(VAS+ZdK}#
z7=DshoME+;ABZZlOp?6m)px&R?%wP4rh94CG>H>Y_7!*PrWy?~x}w$5Ly3)w-2Ll=
zyB)Fd?`4BlJdZgVIq49qh9@O>A|FMU-6~gHk6GaHi69s78$NG{VKph%2{RdsHGl7q
zuhB>k35l!I6tYZznKJG#S*)vb1gS;Zg_#!ibTloyXV1y+wnf#k-?(CH%UakoezX@?
zwmeV!`hUB#{ZV~W4%__Lc1YFWaoF*)k=f}G=j`mw?3$VJ{<gX4GzMdMm9UXlQ!JwL
zoonC^V<r$B>F-H_{yo*iR56e@!-*fTy9hYoxNw&a9*}czbydoWhtU1%vB(RMjpfD)
z9Xwz=$cG*rjRXL0xu$fJB9c6k-vK-iQsF9ytUz0sFgu0Q@m4wxC`h&gufZ}zE}-ub
zVt}DAiBzXZg+5#A;*>~&(y$DXJb2;=5kM7~aP#*23R~s(5Kt%_OhCXoewU&e6~q$f
z=wDFaHVQW24InTFlc#m=7CxQTmp9y9Tlsw<vYH#pcOt;9!-F6mL4a?M@;?Z0?kGP)
zg4Ia)d???J1UruM5E2|dWaFe|;M-=%Bm(V#SHqLwc1-Md6d&|#u%Uhh&;rWv#@#<8
zU+12YpXQ7)j{-Lj$M8I3(F%IJXyO<nC~)%7xQzxLRcP!EbBt~j`1;WJ)tWCTvCB1>
zWAK7vaKBZo_{giX`RS01mB1eEywj1g=q5MT3(0siUIA_su0moCJ`8w!Q0T}kgQ3(h
zQND=_P^~0dQUI=9@{n&KYOiQ<8X$uP@KM`l-nsrAl4G2-pd=1X?+ZKCz7OnP;G{J`
zEjSGX`S^Kowi6uW6)XY0Q4Tq$v7|3k<LU@|z#iTk0C@R(`}sm>MYlnrxwv_F`LO%~
z!a{-;B2<%8A{dMF$0R^eoTM&U-{2g@&`3>8Rqw=ECAt4o#qrWI^2ZfUDiV}cPOEEZ
ZYUz+<wWWw>OvS{_EFF<FTVr#C^?%Iy%C7(b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.0A7ACF40DAB7236A.idx b/.cache/clangd/index/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.0A7ACF40DAB7236A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cf87f69f3a46c9bab0da6339671f843f2b33ebd8
GIT binary patch
literal 1776
zcmWIYbaQ*b#=zj5;#rZKT9U}Zz`!5~#Kk2=nR(0%3@V%q3>9;ZhUDhow&3~pUsA&S
zy54ry6YXZ++a6pqyOI0C-Em?@bcfDQ4NK4W*ZaHPrk!nLj*{8HU+7||^z(zSqyyW7
z?n6~89Ve|yOD(LZS2X@%@ZVeK-MV}`{Uc|KY%GKKRIBM<PZyu|`n3Hyk;$u#j~0us
z-o4(l@Oxt6=eNr8`~L0~c&-24yKryZ*<))29$n7Ty*7XSe!chm?{vM6zV?Ufhspf%
z16u2AdMibjOgs4I)3NsxTS_k3u?ZJFeSM65=SGcf9(=j`GJCJ?EIhSMDn9Xlm3Ow8
z%FB;W9{w|GX3x~E|5?_n_la|^8EeE&N9jo(Pc;9&pCFf<8Z)!SStakD^TU<XPM)w3
zR$QR_<Bd&DgpnT8eP{8W^0`raD?k5}RY`JRY|Lx2VrjSA*9qk^pV*VnD`jz13x~2w
zao@akulHD6z{wLE+)`Cfo+w+$UVGZLp{?fa<H^=G2FEk9Qo03JhF%PoP&n7OdF9!i
zJc`$9Pnewks6J_FwxiX1x$OIFFaFFE$d+-^?fhLB^soC#qWVdXT#?0HuAZs0Z(S5#
z<;xLMp}`ZyyYiY_)Zw;ti{%ovH8~ruTc7Lkp2TsGKkNPyqXx~n&zs72xn<Q_wKKGu
zlm!T|Es?D@n&m0E%#m3jx@gYwwpq6iUkP$Q{+N$<s`@*7pLGgV9;>Ua6y2Hs>f7pN
z+wxxjxpy#blK0KhE$licm=Z6Gu?lT*oMyq0o3JQz&eF@VN87XPb_;II)A9FNerNHW
zLwd_{JzZaY_A*V@d|<C@byWGu*^A!unfP^OKPp$+xfrRQT;`=0b^N5vfB(Oi+-C-?
z(sE4_&*YZ3a+Cb<K>YR9KR;qWm~UNGZ+z4Hmg$X%_1k!O-y|wuowj`P$;jZ@7Q0Uz
zNWXBTYw-r}cXRSP19LN`ZEw7!Y_RiGrF+cG{`GF+%5_0I>o;H7Zf|UnvcCF!+ji9n
zr!JqX7TR|(kH580(MvBO&MY@BDR#fbiPLwUOO?AV^hhtf5GL~Y<cHJSx8AS{+$<?L
zP1j!Rje^qKr+lu-Qa;AYWslXjD_=E<Ja>8e<Z#nHvax-7pLz=`Q>B(}`nmSbqe;_M
zZMc7@oSpjr|IeP9ZLf|_`1q>pe}_VxjABZb_1t;aE3Mu(`7A%Y_oa!p#mCHf|Buxz
z3;muKQaHKJe&%&vasHrhvcG1#<`h@vCfNaV-{DnqVkK!gTezA00~vz3*|_<^1W*PH
zP8SQ>>j_r4^DqPer6pJ;!odVgx+pcRxSEN9flYo#pYuX<1ty+gVJBf%n4v(Px3C|K
z$MtK$y=HIIU}m6jfN(TS7|4qi&OqX2375ing6lh_O4q$Q4ALSgECUk;@)U%Xk$CFD
zS}>l()1bdfUzc|=0u_}Dw=lxQ?SFrKyW&~pdN!c=Ax7cDa0wv)GNbSnWd37D;U~!a
zbxgwRnP3J1)$eB#K7h<W&m?>S&Ub8o!|?M<$t^~fU}-&Rb65x_JnBeiyRUE$<gTa8
z!e5wS<`uFq%v@>k`vOSf2($1-xJ2*cIsUr!k*OewZ_L6!;1aWaIE&g2#}$JlID|!E
z2CaVGTp9UrP9IR1TbvsnSnG-lHs^Dtx&VcFS$H*IQFkcB?q$6E_2WQcX%=ZEm@{F9
zF>!&@LGPj83)VceRsbqO@(WBcBNsR|z!aS;b-gaadD9<gq$G<ZFHF^|M1@_tCU1>_
z!orrqHZU80xTj3dJs^1<q?%7s0;c+}K-i|c`A43CR2v8z!&Lv<d;3v{+1oIX@MUJ<
zJMh3@C`!#q1bM2sD7jc1NHYg7l|C6Uz1*FbF^C}u=3NGc<lK~;3?SD#&6k;lm5rT)
zlbeTEfS=D((A8N@kjv9aNL*M%Lej$0+Q!z--a$`P+e+C?Lmr!`n6!+nf})a&s=B6@
dj=q7Rv8lP7iIlF9i>R2pR|u1vpQ8t(4*=`D!^;2w

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_add_instance.cpp.09C78F4C6D71CC59.idx b/.cache/clangd/index/km_nk_mn_add_instance.cpp.09C78F4C6D71CC59.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ea2909a4d6f79ba2e871af0267911ff3db6848bd
GIT binary patch
literal 1924
zcmZvbc{r3?9LHbGSZ3ZaW0+wqGg->oLbA)2Va6aaCSDRDLb^ki$8syFo<|glvfR2<
zL}bf$Nr<jIc`PYWmYY3_P^5B)dk^=I`##Tm&hwt{_j}Ioe9n0fG3fMYAqcYYW<&(}
zgnFVN2omGHA)#FVWB~~3#6VC)Rtq~Qct8*SvPO~G7mz-{KWdS=9XS;sGo!PZd)vBV
z=Yh27l&tjCC99UOGT%Ss(r|qj4HQ)OE74L;2wvr@Hqy0IAsk6dRT!$TAyVqHf3_XR
zSB2Ajzn|PMo2`1pBz2;_9J?^K)V{Z6a-20U>oq^8s!AQ~__)6zw|Sz$H+p6MmyY1{
zgtn^^Sq&Z<E1#QZ^H$zJi;9i1&j_Xr+UFH{HZnQ#>Tw4ZCUJ%D$i|Y+H)1-^GB~}v
zE?qZ!Ob_lVVdo}#^)q`^>Yl7|?Q=OtmVbzh%p1fhd%`hOVevZMB1PT$Ql;kNDU<aJ
zQ{i#3G@GCVrJ0<(apSQLj)J~_p(09KKf5T7#OSBrj`E8Ot7NxJ?jEeWaILR#iY2g5
z%`-PGyUhi&5?`(;Mmk@VRECPL@nZ?)mV8mMR4O<3KQEd5-TZhcE*W|hem%iT@2x|U
z1IL;w|BX(cun5kX&d*@pYA1O0#<R+08<=G!IN=X7kGZyI$X~rR`xTs;98K(=|JrHT
zC;TwJJLN&jfsuqAk8*ADyX6V<R|B*$#MiD;wn@js)HG|WAC~qY1(#JAB8`$KILgwd
zlCQ*mFDfQ8vpeKXcU4T0r~=<`^8S_#e`fR73w;qbr`7WFjjy>*PoC;?wU{zmqS18;
za(CFqeySfibuoqa6Wx)jM<t}ed(-_DQ3yf5sh}-nu@mKRG}w=Sdq%6v;a;u5@|79-
z(*hlB%_44s7cTQg<@V&4z1Q^=RfmUgt_;P|MZUZ0XD_^>&0HLvBGm7y%Wp?@9abBk
zi<(Y1_>&s1;iz=Z8d-kndyNkl``C?zAELxK7S`&f9_%h|e6Cls!||5!gY#}Zp@+D1
zj)7gV)+x%%(mN_BMs`S|a$$FDQ^s(Q;+|U13dPE!&6$n8XMS^<9#UGGy#j~5@%iMk
zR*Bl}(jPfdQ&{@WqB!W%Ibm<hcP0FV(--5uOBp#<cM&5E7P7UBXi48pDvPV#7oKx7
zmUBOtzB=9N*K{Mi>>MZVLn8_CZ)dvO+_e4FB0W(-7@;h7W%ku;j6d;7urs2Xr=5%u
zN}`R_#*92F)=V@Qg0IY(pcg)K8VPsPP2ER{UBNAazoy}4jmfwl;w$XMSp1wqj{TW6
zPKDc*XA`Sh)GD*q_df?Cvc|OE9&3$#f8?I>z#B70!t!eB$|tLH)bleQb~R&pP)KBu
zmku8U2}XVqHjkm4!1CL>KnO+<Bf8P>{@sm7MCNAM({7eC3^}{}GlCG&ME1WbBopc#
z{bD8aEDRk6vN&3t1LUYAPPkF9WeyBE@g^su$+{a2kdp$ZXGwL+dtitI(h_Ki&B=?5
zU2a~U?GS{aLqNu(@irh`lVdg@^x}>o3^@XsfF{@jxxR@&E?W&C!H@%xacJCTy19T#
z4)qPJ!15pDm4+3>;x`(Q9xidHsd~m3g&{Ubi=stsfz0*s4O#pTJ$}pvwZx{?+Yd$z
zx4{Tj7OMo701Wv>HrQ>M<WO=>01+U`Gp!TcGEq@fXJ7(7A7+CiFaeId>^eU%3?noY
zZF2!YlV@6|xMh+?Ni)I36@SIv7ik!_<bx3m5hDvO2SClst|Q*c%AjQIKo$V!nbtvX
znWRusn+t3*!EE5(GT~S_HkbiMMD0X5n@51)S$H-mA4UiV2`-=@7$G_nJpf=B5kD;M
z2}r_-gtLSPU;`s$J+c|7wd1tqYYCAz_Pl*!iC7s>6<CtDh4s_ks*Zx9ut1jU6X>~q
z4??)dLL_;;0P@gE!)x(vj5?ny<O*&G1RV?V4)o)F+W6WEpwNOs!Xj80N5G3R#4Ig#
zh+!D!L`iW8DT<!Hp^@=^6VqKJbpu5m725xel4WG&w#h3fZC6%R)7ZU7bMHP~s<!k_
Stpg;o71NEMZf9oA$NC#v6y)Oo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_default_pipeline_v1_instance.cpp.7011152E4CED8319.idx b/.cache/clangd/index/km_nk_mn_default_pipeline_v1_instance.cpp.7011152E4CED8319.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a8e6d099799df6ec2af6a25f6c7be03c95c98270
GIT binary patch
literal 4048
zcmZvec~ld39>6m*$c94>XMV>FLO_wD-3T5i2r33d356h_Z55G2(5MIk3T!K0t5$fR
zcoc2DSnX1*c%W`=mAD{yivlW?=Se&&-ZoZ|3LAV6^8I7;UXu68_mj{3-Y=6k6BQZ~
za?VyJ3z{98lbW2BU?!8v?2XUNtPE{3CzBcEGFi^{x|q~7$<OX~cXWT7lFjpr-))GT
z+u&o-e(b`ei!Db^;Jn|CR0Te7&c5_~&6;UtqsEjgz6kguCuL`DCKD-pS2;4y!*0Tj
zBl`kQR@wDC`F(dC$!;F4Ab+L!?cDBpJ781ElOt9a-#k#v$v=O6TiY)M*Ka-AfBdVT
zk9WM@X)~)U?&X`Z&LOAb4yGhUrFSJ9d-!C1^RAM&yB2o5QEKzdzTSGpKfZ2WTJiNM
z9`<jFeDc<>uyVhYbgeiv%|TIj#`#WS`rMX(Uk@4fY_AsS{@eB7+mN)C^7ZbvZ%g_G
z$di5#`&s#GuzQxeB)-->{8FNP-2HCHHD_byf1Z4A1MBmm)KHzjVDF9yyN1nKOPofz
zYD<<sFD+<25uuz_^^Z>T;fMD$Y%`Z)YKADkI_GcMQFz$h9@SPwRGP2;VQ%<~kOV$5
zKlqSJ8*I1kX`=1>Z}Vm0FK?8&XXiK<h0bs-*sA(;lkKAA9{v~3v}cNml}pbBTzu@m
zam)M(zAv0A(gtVcX})Q&sm>4l;Op64cQ?U(C02*2hO8SDle%z{=3F~1e;GM%q-93M
zG4JoJucWJobr@RW4`y!`x47~v)2`{xC(CB~Y5G0ODvWH;DzaCXUfwOI!=9eoG%Z-o
z7kbOe`xm{Ai7y`;tjx{|cz*SZ&cF|A&m&jc+^OrQ=XcJ$m3Qshjno}?{96j{W-M`h
z#uWT3xaDZ_I;Pd*lYz@e4QL3cuPHy!7UZ+**5tVkL41?vTaW)73d^05QT+a7#^uTR
z?=%Gwjq=>3hQ_kMAcvxN#Q}~{SEt&A@0fYCXro`(uJ=C_WdyoB8MrR?zVhqeDGFRQ
z>%ixK=(8u>xw)_Fva45ab9~atn*&B3T3fhb#&;2e<{Z7QT{304*X~fCi7qAA{O*{m
zw%Clgab|;J#d);HaS?KuRX8-nus5nZ<(IT&INc-e&z=3#!mKww*ztBnr}>Go_gm{+
ze_XzQ-t=ju$4(8r-+Xh#YJX+%qY>jhPG5XA``zJ+2MaG&=sz5K6*7A3qkq0Qqd5>#
zskIE@&#o-NaqEXpA`A97{<L^-@~~5AMr`BC+Bu~=_m{VJ+paD4npQ=vrcdco#C2$L
z3s>KaD0}3yJ;MKN)8q8CF-o;M{X$FH)2VSuFG75L^lvwHbX1qRc<R`tO7Fj9nYpQn
z<Bb2hsv0|7**PCs%Oa=9qMiCVg?)T6PCh<P?{mUgDGQETu^Qvp%G}EK<4f;2BRM5A
z$x0?OR~#=2Djn@Cw~pr3e0cA-tfK|B00^lOAgsoKh?)SRY6?hHi-07x)JK-@ci&zW
z<`Yeg2iEY@O$X*<`8j~Je3pp>jQ~s`Rsg1;6@ZB}2$(1q0TXEvFtG*$6USm;Vl4(H
z(GXygSOQF<CBUQ_3QQVHfl0L#n4(4mrWh*%Q`CyUlr$1BrC14=l2+<7?GLV#79VIo
z-LtWjcQ9=puiys(a^pQsBv1jEf<geM;3fbQQ3On+K)^(92$+~+U}6OZCU(QXB$NP?
zC<ri#8v!Pz6qr;&fl1vcFhwc?Q&fn+6x~E%N>l=-q>z9qxk-KI_|8|)etejF(A&7N
z)eK+5n6{5^WcVh?Z49r6T*2}OSm5he{uJbTmcIg6sAdG%7aAD>_JuY^fPJBY6<}Y`
zu>$N1^{fE<sG32rj~W>S`>2gUu#YNO1p7$GBG^auEP{Pp&0yHajSPl;+{R$o#}zDw
zeXL_K?BjYC!#=5I2<($ahQK~)V+ib%3YNe=(Xj;fNj*zopH?#z_Gu$SVV|}!6!vKa
zOJSetSPJ{Jo~5uaRx={(i;av3`(hg-!oFC+im)&0SP}NcdRByeshW{sUut9|*q7QE
z3HGH5R)T#=$4am-)w7bRKVj`tS%)F3l`)HU8R{~+_d*ok_RSpgJI=RzE+u-7f6bZB
zcR|kykddB4kg=Y_kcplnkg1-dkVQQwLYDNLWa7l(p(pqR-SM7<G;w?@Tu2iqKt@d*
zf{dFu44E`>1Tt;nC}gpT6Cq1YoMhrfX0mPN-kn!_7E;Uc4R9f~oB$csatJc6<uGJY
z%Mr-5mZOlxT26#4)pC-F3-#MI%4J_B_bkM~@qfaF7&rkkGH?hoHgFg+F>nMjHE<NN
zXy8Q1l7W*<{QYMZ8Ea0*W%evY&LeP!$OS-1jsRgf21Mip5S3FvqFe+d$)!H}>B*|x
zFYj$HVyvUBEv#*N4}7#OuiWC)-HFDJ6*damdn)(7GnOCksS0a4ckjRF?_3iz?(r-|
z^I4<qV&*ca*XoV2mLG4idTpZw{vGv;CpHTfziE1oWTxof``z{g=U<g0pN+6HZip}M
z-<vX$il1mA!B+q?K_!40@<m{VR0zzlF9tKL!eB;x378QT0W<1L!HlXXn2Ek3n29P8
z%p_k4%p{f6H+%Jbh{v)~%gTCw`5b-~xHt+q0z{O9AmSVhk>n7FG>1YIb3}+zj%30Q
zcGD_vrPV&{S%4Qm9PF)^00?;@K-dcdB3=X#^`d}8FA<RBCH0XkBY8nW&wY1hMp7m*
zzH`ydR}box#noj@lq{<Ej$9^7N}au6t`Wmh!a1{k=6{o0SXuYCv$YAepBUt3ZyBl@
zKzM<QzJC4zW5<mT9Ev_2GkBDT;{Sb$9R@l&y9{z2;_ms$-@Uwtjqn-mG*VIy4@P*B
NI))92m@t|7?0<Wo<&gjY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_default_pipeline_v2_instance.cpp.4ADB4E587F42B7E9.idx b/.cache/clangd/index/km_nk_mn_default_pipeline_v2_instance.cpp.4ADB4E587F42B7E9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2bc4f6e6d80cf6a6de054a2716afadf942f4495d
GIT binary patch
literal 4070
zcmZvec~BHr9)PF2feeQTvHQK(?*#@yIdpMEQAbcfbT|eT1FVTGqJpA$;R5Q2@mMjU
zDMc5Fi7qOLg!Mp(V$eWP6z|liOiIx;7(iVGHOpa41eYDG8t3=NP8CCa{rl$YS3P}9
zOk`MC(?Eq{^885svIUv*tP~1`v*|A*Gd*>Ipip!x6$*X9rMP8jbi7kjU+e(KY1dCj
zd=}wd`E;MvtN6#o3!dMbZGZ0hy+pUy?P)*0+_Gh6>DYiHs!u0$=~E7@$>3%ven07-
z=<PJ=&Z(mlYO0;AYYz5Z!>jL(JBWHy#vdr~X$so?Rr@LX8?PR>kG}M%ZEDAZOKmTs
zW<K~~&V$e1hUAoQ@csQw*W0`~<)57L^U?I3iSK$<+OEFQbF-@V%^v??KFIKImx`p*
zRae`*^S5SQq35og|9sFBzqYG~t9I;6O5T;`r}1BETwMJ1s)kCtIdl8cql(8(+I(eq
zPR|YAb+cc0$-mI-LxTcJqDCfVd#3nVsXF@{FD4&7eeC3lh`;T7J>$(>?bwFtURxLG
z0s;rJyvnGM*SEAUW|!_MPyAu6{ocmgPp1k4hBfEynfS=%^WMviySLi4Bzn!tz7SLS
z&FbO`TW_uVr$zqmTc@o0_Jeg%+Krz2cH{2A2~JlI`{rq<7JYdwM-!6ovLN}@SqHy+
zLGuqh?smC${j^)b>YaD0oCD5z)i27Kt1Vn#kl#1=KNW)}*_RA0%A4YRbZk|z*Zh^+
z<MiHRsP)-9@nqVTg_@GDKg(KLaX0Y#c3qxr-MGaMAC200qM&rss_(<RR|Z^BmfT92
zVYQvCdK5X_y<}S$+4z%9_?lrAKe-;-6ByvSdeQN9h5x<y$SyCZ=4iO@rMD-R2QOaS
z{+;1hm*?6Q<EwLj+o=86ed+y^^@?Sp<QT%XEplDh9QR2_(Y~9@hg9w;+V(7et@xM5
zk{aK$5$h7qrVP>cR%^XdH`fGZJ&P?Y8W7$aalp~%>9DQe7)Hd0U;C&p`fW`3_`+Q)
zbEk~xL_wL`?F#GEcaQziQC*m}{Lg(Ud0*8msqjl&Zc7_4hF$fY=Jof{_uKrdz0c!Y
zhxZ-Y71dUdCKTy%<R!^hcb<Rr#qHcPU+gFk!yZ{NM>d>#P!(*a5!GvNZmAk-bFI_R
zbADa+&#vh&l0J+JnK{Py*sc3>96Oi1+<RgFo~vDZRr$@2^`m+l@1$2fzSFV3QD4k{
z%($?+{9wY4`9^WOXRfu{w?>|wJEAcxaAteh^`!N~zdLi-dfUy=iL+u)_o*B`<{Z!R
zLzRbmgKIhuc2fJZ`={SrqS?PHx766&lo_9&|4aV24;zowTyEO2c=6?5y}G)xCr^&)
zwcm8v&vdn|S(X%Jy6Va%&f2^Eg~7%_G2>%Jtedr4)Vnv+<lXDyQCF0ivb55kk2a0%
zZR`iWdoz#I7o=ol*(($_sx!Hhi^jPr9b!eDIL-VihgeA`0YW+i2<tE)q9cH0oeYTT
zC?KX|@5$~>VN-V0m{?-kakLn3*|C@)E(DY+W?D#!mVhZGNWhd*C19dx1Wc5GfQeEO
zFmW^nCQiV>#HkpVB$@z|BoJVdR02#nS_Y<^AOllQm4Qj4DKKdQ1tv|Uz+}-3m@I(-
zlclovrgQl=y7ENZ#r^|JL}od7qDpiJ<SF`CNFow2C6xqB$x{L*A_$mBg@B1X5il{q
zz{DyHOzeq)NeBTZQ4wGgPXbJt$iS3UGB9ON8JLt%U{VzYCiSGiWP|~esTeSsCwp(Q
zo}l4b>FuchQt1}N-vn^@x&;X`>J|`W+$~_pq+1}6<!(WSOuGdNnRN@y!ofWqFaG&t
zP5DUELh3m20%uu>*uaTK$Ss`M3b~9IPw>F6<;8Q54ZL_0uvEuMurD=m66{MYoCNz)
z885-URLe`SFBy0V_E8;&U>`Mb2=-A6hhQI-@d)-&EstOy8F&QyxQ@fHj~h4)`?!U}
zu#d}l4Ewm2$FPqLJcfNz#}U{k4IF`e(!vqgCuKZ=eNxL4*e3>_z`k6^$*?asa5C)6
zEu0Mdav3kfzFf=8urC{U8TM%%M`52fa1{1w3rAs}mhlw!X)RA-pBi`y`>c**u+JJe
z2K%gqW3bQ4cn15dmS?cf3_P>+C*?d>{L!6xpR<bf7~wI_{6w4_^JeImm%92mhge4|
zM<4SeZ*I%hg*Fe}n)<KbRzd89*KeyJK}M|tf{a@Q44JeF1hU*J$dGBPKq0eMfmt|d
zRAlv_d9`Qy7h)8|`*0yfL4u5o0)mW<0)|YC0)Z?W1sO6m3KTLk3e3WPwNmUoGP37Z
z|3c~o@itsYy&yqG^#X#7>jey%)C&Z%TrbFwX}v%pvwDG9_#abk)3;nq%;;Z;Qj{$B
zCrK#*LP`V(D={FVB!Fb442UWzAf{yR=~#RAnpw>Sxu#Ra!Pddiykqm3#M!<3O||1f
zPvP<AK)1@4^+#;a{h~G5gB=GGvvnaZ=-C|A-7BWe9#$Ui=FR39Z};w(mpNjRz`1Hz
zsl6-O-iz)}S`8jx{%qSkx9@%ZUyOD#9T1HeXik|ZRMc8X(nw$?g-T$CGziR4C;~IA
z!C;0%F_;kz0W%Uxz)aT2U?zvkU`90*%xEYDGp1o+#zNWqS;xzv-W$elDDD5|^<p-7
zR7rXXBGMy>SdSqRJ%K3eWr$QyAu>I);9IAeCx1?>f6~7IwKxhKTeSoTsSzNo#(;>L
z0Fu=*AgZQ-n3}z(^aV@j_1~H@(vvf=>Aw)WLPsB5(rn7*ViYmvo0dY6ylno`MJ9|&
znI>3S+kBw3wRafcG|(~9Sv%R&*)B5lLp(?%vSxhXgo#1HAtR*W0Ul$$RsU}*VjsD<
l4R#+g)XQhspVTA$Mvoci>Q8+~O_5NzE{+eIJ}H8m`ad_O_Kg4l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_default_pipeline_v2_opt_instance.cpp.AA43A9C5EED3E40A.idx b/.cache/clangd/index/km_nk_mn_default_pipeline_v2_opt_instance.cpp.AA43A9C5EED3E40A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..266bfcb20bbaaff7e5580eb073425679bf3c7c59
GIT binary patch
literal 1768
zcmZvbc{o&S9LLWrmRT&LH)G4b7a>t&HyOshUMXb>!;Nys$f$1ay>7@>*OgnA5|zdh
zx_T648cVbil|sfYCX!SaH|`nFb3FdI@AI7ZoX_|B`@Qe)yuWk2=``AqC<IZ%=<Jwq
zRtOe?AQF7Vv6xX27zo1RAc$Sq<{NXg-%R@Ts^0U=f~SWi{-KWaRn%%5H^?qJb&b*0
zpOUk#Ml-lgjxOnn^>TT^wl;ewtlCe^7KzxRhMsh@uh?P(I?LzhR_wuC$cj$Tzybcr
zeRBolY^LbF6WZH*8RR^{lX}L3v1uaPy|!UF=hsH5jDqT<u3<qCp=sqLcdS?ND%usA
zRPjIDw8Huxp18W07sQvD-_h2-c+aPD?~K7M3SO&Zu=e3c<JjTf6Ei6$Q!%5ZZ{)`I
z6D1}3>-u!B7nxQy`@i81cE5L%9{$uCn-uUM*(jxn)nDSmvM*!dXwhA*+fs(8EBp|)
z>+NHHvXXk9pOu|eDm!R>Q&^H4oiQ0g%{hj9uiS9sz94vU;@NV-IgEuuky=-=jA7K<
zihCM=@#Rm*c-ajIG!J|3O@F;$iZA%ZThn`ME)Dxfhv8RezBj2DH@LL0Csi!nI4UEI
z<P#Y_uO_MgIQv%ZQeeu3=#M6y%q6vs$G24r6VAS_Cs79fd>EM;>`;8F@IjIh`^}l+
z)7?3BIeSRuX7!~~9n*50InN(W@yx2*M)}0fh1}5jsal?LLazLIO))H=*YqeHI<VVA
zWTu9u8&7E$zt$B2p=m^m(C+l#x&ciNa3r$leS<kxVBhR}_CVC38O#?IZU1wYV}4d8
zW8YF6zE^Up`YGIm8ZvMELqm<&0ie}<-N1#|Me`aij?@szZx7h@%J1ZjiN6(bF^3{U
zDY0im35~h+T4YY=*51skX-W0(HM7q3I_6o6DD=g?R7ozQ;v4WKoE;7V-HW5EoS<o4
zF>veor8M=U!&I^t%JiS{<2!9mkr$p-P4ittA2o_9j*jQ~QWt7SOqU)yPd+9-{{R>w
zM0&d&Ov|B6_S^2e){=SKc+4@F*%s*-*!IxBG=pTtdC3i!?f)pEl-_%v(b6xI9F{ow
z^H3A-h}7L<#yEY|iu^pk$eE0ydzqK?2BOYbUS715KbWo4ayG^wVgHdWA5*i;8+evF
zC%@%6^=x|QWah>b58=5GxLJmd<ml2Z74aJ3#@x;O1G{rt8MBuvIU}n<N*Y0x$Bkv{
z%Kr_zu$UB)eR{yXZ1PK?yT$wz>yCMmtd<rf`~&%`L3Le1X@35^AUW}^^os|85A}wb
zcf=*egj&OY-Q5zqez9j)Er^h(7vv)^BJZ&_!Oq%fDOZ+dMMT#T&>pZ&5G6>hO^D45
zkBB>gh9FUuJK5AL=Bg;74{!!<NEUvncre*zrw;~}J-|LhhA9v*Hc%8`Aw;S^qY!^{
zx<!~G2NV$*CKaH*fwX}hLW;}QTaPmZvck=0F<=gZ7&gnFW^PO*R*S+04+WZ0hyimq
z3iNE`5fm8R$W>@?6TKn62@SX#`8gVNuCrb02WWAD)rZ3RC>bi5BWFtWUeY?>_IVX9
zOcE=pha6h)NE$<wuR18K_Z<d&Sg$wKgx<I*r0I^(=)r)#b>k;2bdh7U+$A(xG2q3z
zQQ(SWrnDT46B^=RGg5B_-0Ohh!faS3;)s%nTsLxxgX0c4z_Nm*f*evNa!29UTA$Ut
zTwFP9q1x<gJr$(V+YU!7VU^U8C}c4-e(g&@qT!$oZ~i>qp%>uZdI)#IDo7x?(;=!2
z2BtGR&~UlCfCaJ?O!mM5p-;0~x=Ch9WpFNFgXHz5e3$;=*1Myv6GDAEFhRm0W_Wan
z@Mz<hhvLY>2XEoDHZVT>?_Rb{)F&jMyrDxeVbPJW>=NOI!HS5AiQ@=F327-wI>~{m
zO~TXdWyqU=ocylc7M51lHnxVFw^B6BbXEROl~q*QqN=8@simW*ziqpb@s6G5%BBhi
TCVMu^Iy(EJY3_DTDA)f0zUI`g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_interwave_pipeline_v1_instance.cpp.37A32BC5170A6461.idx b/.cache/clangd/index/km_nk_mn_interwave_pipeline_v1_instance.cpp.37A32BC5170A6461.idx
new file mode 100755
index 0000000000000000000000000000000000000000..883704a19213b3151fd9b634cfd8cc413ff35d70
GIT binary patch
literal 4072
zcmZvedsGuw9)L450V5!wow@hUy{qA&QLGYe6|g?gASei;z_JBVR|Nruq7Z6$s9k(N
z%c|A-s#>g|7^u~=iXto!ix!Itp0&PEqu~0$`e;#nvB7gfet#_IBst&wzI=1fW#&vG
zf`fu?yC@U^@xg}F1pO?fLZRqsd1dI+lamC6;;l-dFzh-Lm6~?l$MruQUiIs@ouA+1
zRe<?g@e$3?lc;^-#e2aX=kbR9DT$R!CQMq(j*}gG$NGQXa<k@C&XWyxW4YUP4TgPV
zl`$6%J%9Q{Cq7;B9?!fn{Z?+=)1A$R^aF-Bp)*P@cn`X=t#Y@Q=5L8>?|Vlt+H=(W
z<ej;tZjb)e^Ot$0{x`Mle`XJ7oAX|{6c3-dwETT}d56BPG%fhW=-V~9P29)PIqME@
z)U7F6bog10Q=H$dJ_Q>?v#-Az;A7v8mIc)K7=6|z_YW(6Dz%j!AC~>mc*Cfy6P1H*
zk0SvyPo`e#_u0P4lS)sYq}eFose8xoLq;8lJvGpAl~|PJ@`Fosyl1OEZ}oxwM_>Ht
z9(+ovcbR^<DRW+X%de-7=dbC#p?=uKy}i7XuLd+tJzKhd^qL~KCz+du4;U0UKik(~
z=T{MZ^P~{ZyovER=i>1wWm4YpF{8_BgoUBkJYq~@?DpK%!uYFGeWLf&xo&O^Xo;D8
zEB|nuV|CWcq{VFCiyYnhFICaXhqb}&T6>p~J&#R^?KLo1+&3>{WVSB4cH)Xv5d~B8
z=NngCiZ`r`Tzjst);{NS!TF-T+g)>iXbbZe{`*VTz@)<Y%OeUi)=zOrwGaIF>RA)M
zJLVCwBWXD|F?GR~LVi-K`@FJ(y9-zSe68C<|IiY*th2@0mm_YZJ#0ID;9p-ndfQYr
zL%Hqfqa>Hl#eUTX22XOj7&P^c;@Ce|w=6`HGbQz*_QyfG>tx5blglO?Suy>W>zmeH
zzW>eqpa|2_SpBA;vyTe4jj^NG7MQ-i_aHy7BroaV<L@)KoPXAFS(~{<aEQ4-@t#N8
zoq*aR1$~=-dWX8w@pVh|pJ|$`bEoXt-R7-(|1mmc`HcpBuMq{t$w$}mxm?&GFNd=I
z_3b{chxFeb@mcv+6<R`CR~P^GYMy@R_x-<l?OV6_O~F?I&rZHfF-)7CKj`;zmm9vn
zwYp#N){L`@|7+t<4#vd`^S6~HIxLFyxRAJH%F1nb&pn+pt?uU5>nU4@JU-)`Rlv?~
zu3mJvf1>BrVRi-E=TuE{to+)00!}IDU7YEeFtk#d7T2)WlvwrM;e$>7>&~3T#WU91
z8NE8xm2bPf`&suzVdIOC$%VTgJ<y-<-HrP78<FrBe>q5dbWy<$`ux$k^7!1OsI?sm
z&Dde(8Cj{b$5`&WnnQh%es0P!M?QS2BGS>`(dEONW%A+g_K#D`j@Jg?b5cyPjNR<q
z!alrN$LR@kGqN2O3VZd5jRE_7-BeDIVu%=KjXFh2Arc@If&k$V42XmfKyru-h=x!=
zEQEED!|T0G%O(wvB$gA0i?Oy7i*aHipkz^RBPm=0rW7XuQ%aVAiNX;uQ5*s$N=CrM
z;TV`W4g(V>V_=eS0!$J|fJu@GFy(L=m~xy9OgUKwCJm>+q;V9OG?@aEg)?BXI0j6X
z%(~2;rE6$r$-|n?jU^(pZJwwWJpgG$FB?fj0;Z&vfGKGtU?PHmiPQ*~NP~ch2?i!s
zV_;$p1|}f{m_$v0Ni+nQGLeBPt7Txy8X1_BP+(Fu1t!%{U^2ph$<z#(OvAcN|2NNC
zi(0ZuhgdeYiW93j+xCeKoY)B2%!zj)8+oyW2Yv-FRzj}l#S4I?Do%oZsezMVUovwN
z>`O*of_<rimtbG2=Ox%jRUCqS)W9LwM`jMeJ~HwM_E80oU?0`<2=;LmhhZN#a2WQn
znZvM;jXZ{ZT)|`5$MrmheNx2{*e4Affqi1;2<#IhPhg)^@C5crJx^d?uHt0amm4@4
z_GL3C!@g|fW!RT1cp3KPdR~ToTE$V=rwts1eQM??>{BC8VV_p;6!vL7Php=`aSZlZ
z1IJ*WnK=gg%*Zp?XB9kyeOAviTYucD$BOnp^w&6Lq^Gy1uk}guck}8*`&({JosW`c
zLHt9oUGGw}AVEgW0)mX21q_)q3k0&<EXa^)vp^xUW`Wswp;EE^;E*>LI~QUS#7l4?
zCP9LXOag+8O#+5YOag%{n*<p$H3<|lGYQPb8}9B3AG|asp>rW^g7^w9q)m_@qc#CS
z#%%(IOxgqjS#A?#$h1wMkXf6+Y`inHd-}?n85x}mQHc`x$EYMgNQD4l6$V691dyzf
z0Z|nN#8j+{iXSb@in+6EBj*(9)ZNM1dg79d1xM2y62@B&J1d<BSbvlpGcP&Vz4CUT
z<*>W6JF#9^)lp*}%v4{gwH)?TdU{w7TVuS#haWy0s^9EAd}pMaMFJOPeP-YlvHR8V
z&ZN@4hxNPGPtM`PuOjPrAI@?+Ie1v>r>>TL@)Ji|&tM^)7-%EOPXaSZCxIFALtut<
z2+Xh_1~aU~U`G51m=PTTGuclDGg&8t8TF%JMs*a-m>&Z(rej^RXHWVKSn9L1u=9d2
zh|9pEN-{_gkpV%(1`LrH2t?T+L!<@@kr|i`-?>KrbTiG=(zyVw_zBostpo^Z5g@F^
zfQXg=lC?4*s-=LKmUYql)7Q;WKCRK6CvFtP8}OcN6eP%~Q9zJ!qktikMu9+<8wD9M
zZ4@YE)+jI=rzfP$>b&qX(&uDg%cl_87@+p^`Za7b7omu--VzGMoYeS~BnwWO8zw02
z?7OMDJ394nb#V^v85p4H=@6{zg+CIB>^EZMsL^Bm$9hX2f9B~kK>h!YBKDWwZtfm^
j`}Ox4I7mBW=%>Sd`wXLlKN&Bf2_aE@(ByFwxzPUs=%V{%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_irregular_default_pipeline_v1_instance.cpp.322AFE9D596702FC.idx b/.cache/clangd/index/km_nk_mn_irregular_default_pipeline_v1_instance.cpp.322AFE9D596702FC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6e03792a3c9b07e6ad8affbcd82d376013d8c2b0
GIT binary patch
literal 1750
zcmZvbdpHz(7{_OKk+oq>ShrZKa>=C>YTa#CF1hTYq|~l^EM-+zj(Uz3afqXcBtjBV
z?y@Hx5^b?A6)KBJu3a7pA!pX}jK?2mo@ajZe7@iJ{mwfxzoFQY$sHm9K=QGTVEBg9
zga81Lf=*_5SYQwe0F=-G5Ru*D$_Q21kL9mY#U(PX1*I;eI0tkYpeL)jHZS`s8)UsM
z@boOc^qzn9B{tSweLJyO<%r2_gkKJeiEsw$D)u^RV7=Z}-Z$YiVuhn_thV6J_Zr>C
zE&J`y$=2kXre#c3VqVUDQt?lC@h0<Qed3#e-qO0IhjoiTQzZs}dwrYBSlZs`Rqp3a
zIsH4ZdTff<dm>~0M9AWtPGGcfR2I+pWXq{ud`XNmVSH=QfU+sh{YGmpDy=N=I94$x
zDv^JOcC#Rf{?0B<`bpF8xpZ<U=i60v$@vUn6H)JR`&z}SRL6kpy8p(DXQfzwy~SHK
z{&Dru+Zt|@YUC+CX(CFkBDIESWkK!kC-$WjFoYh*83y<pS4Yi`tr)jGOD6j(1h}9T
ztEW}{HAf|%Cm(Rd{v>i0q|*tR{G_J0`O>K;+1;XggnqVi_>~0U!TF>_r!U_ul=4GT
zw5@*Fl0P>CF6XWt%IEotKAIqnc_f>smIfnfxBdt|Cf2!Vs!H~}OpcgrwO!PU++;jl
z?R*r}vXy>fSHIOXj~r*gj-%weqy#-+clkubQWLsbT4gW%Q~2uMwoL4D?m31oYMe{d
z3MdL<Q;L|P&SDG^cI%~Ro17{FB`aV75&y!IOSL4;sGh3NA2|~j`;Ikf?ofs&wv|M_
zrhK?P`LpEij)v-xP?{d6z5i_0Chd5Tmc|853-DS>m!#tZl>KRLhrj1#@j?y+A0=(+
z(VW+)d6>xZ3`_mR32Q$nw%}y7Gw^b@S>yyYH*d~n`CyJj*AzFgrbGAdLZ9T-qUhrM
zX_q#l4FSi(s1V7*okJN@Z<ZI1cN*2+zqsdxJ>Q^DPQ6x0e43^<ntnH}O@3f}UxUU|
zna<}~cT(&pv+Z-73s?Rs9GX?zRZp){tD%pSbhX6ixK0jg&(H9X5-n5w%3%iDtAaDz
z!R)2i#vOZ)6_!bHW6)&`+Y@6ruS}_RJs`_$Hb$SL<P@v&-ta>Ri{!`mzCQkFXr=GF
zd+iPN*ssr>9BnEX7;<GRCrzBQ^?0&wAKBA=%AHG(-dR&u;D+Y?Y34y4SPeB)J6syM
zTe_-d*?f5x_e+xf!LjJGOpap%wrApWC{f4GE|S|9I_={X=5J$QFg~BgYCqB?Dl>>^
z8MQ10m@Ed}4Ep!bikTJ9x*gm%1I{i0RaRKmZEb;^wWImIvTV3ta1{pW1lbr7jO5w^
z+hM+bOdliwh^W*hlk$x4h|N^c4m<>R!Qlw8wa}!drQFhCW4zd}Tz-tb0Tl&RJAp^x
zYKV@3{u?L|42Mao{EEWavdLyaiyWv3%Mhu6gblO<)Pzay6J>=kSCk-_DFS38VBLJ_
z`^>#DR=EhIHzC00bpzra1o(O*k08L&ja-HV%aL#pRNsgMpKfFx66{=O%eZM^@kjVO
zgb-C;SKbJI87Du+`-=?X`GUDdQQ#~J&ZFzwHf*6isS_BlQQ-S^<BSlHS-fkhQ(&~9
z!0vS;X(Zc8=c0$Nz?egU^XtY<2Xt6$vlmlfh=MZk9ZIHRSRTXK$&ie}VU%FG>SJ^<
zx^c)Fl5r9^Ie5?}SF?!|D!r|cEH5NafMvJ~DYmv1;ExIwZFtCzv(~-R=zb?bq!3OV
z?wzFJAM5O$*@J}U+6NlLQ4m>y)-bJL-75w{pN&DipdP3X%VEC3G{L@OhIun_&=p11
zi>M%2d8aENC;$cCkpSS$@Cgop(BAJ5N=R5_lPDUqSsW`VVJl@#+94%oYX#!ANXyCZ
z+iz@QYG!VsE3=cRW~ia^f2ypa(pJ1GVcT|fO)YJ`UAy-f7%A^n(9u62BWq*liX=N)
I9z;0&2WmIbmH+?%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_irregular_default_pipeline_v2_instance.cpp.6202EDA536A25A77.idx b/.cache/clangd/index/km_nk_mn_irregular_default_pipeline_v2_instance.cpp.6202EDA536A25A77.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b8b3cf3c45c627988afa12bb4fbd9cb025d32d1
GIT binary patch
literal 1774
zcmZvbdpJ~S9LLXGjLf)pT!wLLiWzFiE!PGy<DNn)<<2AxQ?5glU4?q0b_iKnld$N?
zxFwe-A}iTN<%y(`c2W(pZd<uDd&ct|k3aT2&w0=D`F`Ku<^7#=yj<w?DGUHOgt@Sz
z!sCL`001aJcWhkD=`$z*P?rJ#c5#nS6idxgo-gv2k@a}i<ecJ6Zv0e)UiSM?7C!UZ
zAJ;MSIzapDH0#eV85zDL9jhv>V>XLyMoCgE;t=qy-ZX$LZ~v;H+9u_pyhKW|s81zv
z#5_=WjbT|*tjD*_Etqb=^)G*TW6~oKy3ddI2)~^09Zx+uo+5J0oV!T*wlclW39LQZ
zV5DauYWG`QE|_NZuhurNt>l|_AgO~NHFzOS%}<5oysYG2wT|A-%ixre@af&=?CT++
zH&{j%rp%X-WrYbZ8nJ=FqGY!+bNfrZxtXi|NX<(|OXY8r{qKNQ<!*aJ(sda|Xsrd2
zY<FmNQ%(I@=bwtc9$E>a?R#;MlojD>Wi3bGG`iV+9iq+pSLU<$DQvl0qt8FNqh$0a
zIr;kq8kg4|kLG4!-v^M6r+0ZZRwS13q{uYlF|H{w%Q2x&GS!WirpumwnQLt$-+PzB
zp}BK!_GMbw<!OY6F1JV<y{3khd{_eeo;7I~CtiQmtYFnadJ>TtL@T{eoF@wUhX>l@
z%5}ILM}_Kr&1Iyp_-j6FvYDDh%PT*;N5(0O^1_s4CU3<0*)>;=)ML5G34w9i{o=}>
z6I$uyc&pxR<wJ}^=xg`{fy-`U`4u`oZ5ZQ}q|Y1HypwNjrI{FUFSYbXw*bq@Y_E2r
z^sHA!cSJ@`KNy(Pn4z7uY|HsKjdqyG9IJl<L@6rC+7YfqXr7w%IXcfR8jRl6n9aTN
zIq$sUiP7?QN{e%9Knr6RZLN()I(?~~n*7<jlq=)3=3FAH_enpiV!*)9sZU+xx$fm;
zS$ZQr$I)OxnHqNuTl&~=q~`m4TPZ90h={={Xg|X<3W%0c8||h)rFf8jF&Pt@wvjtk
zhVB;Kx#1=(W}&#QnW|?(pI+}2Tz#JN@M?A~U1evoSJlOa@n&1Bg<|db!Hi}dOy9!5
zYG-QtTg{l+knKKpzGjq~p)nQ7q+91J&Q|Iyq)bVMvn@-cZ67p^N44lQ$nzpU*S%XG
zou``@k}~G_R7;J%3O)Q^nz#Tzwt1_%KQBAs-mC1qo;6?QIhUI4ze%LlSG0;InDj@-
zmP7VLM18;b?W5t67Ov&?!NJ`&w-a_$yqx66`Q_y;<<(7$_O%Q6*^!Y$bEL(hx;qN`
zKP38F?0G<JQd9^P`uB2}!dazu(q3uc&|$zEFM&U@u|Uqo)m^Qr9LHcb;*g$@jl<yN
zHWt{931`HfM*;vw>tT)q*IXMR?G3tu9&i+PrIerUX6B88WKYl!mLWO`p4vjE!8n){
zYOkop-xqd^L+~I0mLbvtiCah))Po8C!E2qkTzOSooDE(^z&c@-x;r^$T3M_wqrk5y
z*sxulpZ$F*sTKnnT?nvyGXmn52yl2S|BV0zTlqc`tVP0!p#2?4uyZR9BEg|ew$Gda
zzAwd%A<*6g1A;j`YX`3totr(2BB(JLw2U4+%%JyKr!W)Rd~v@56u7wAZ-}u=8yNiP
zq1YHkfp0dAb7<gJ)!x-1vC)G9Uu+t1B?NZkZDx0^mfoAR*8bI9Dkh^lAXdD88>j>y
zqh>li>G))E4kY7LaO$w!`2IqbRL46SB&*7*;^CZKKGe^FS|h!XOh6Neunb2bu^V3q
zToT%}<>UkLv=K$A*Ph~(XjK_FRv4_^N->=^K|+05fYxv-hz@`>nAYWe>)*1TOhK`r
zA!rQCKukC@SUkemn9x`i=!x>i_!t&?^ZzJ7cmZDUZ~-7RDvTKcp&P>kg_gibZj-`E
z%gD>gx+u^bbQQ2J2ezw#ig;BEOKY3`R9ia(rQKFL_mQ>!PgN$UYiREv?$ROY>F+Vz
bYh+?(u4$@9F?Li^c5?MW(hu4@Bi#Q3mWS3o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.667185EB457E88CA.idx b/.cache/clangd/index/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.667185EB457E88CA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..401eea1ca2dc72bcb581ad1d854ba70123f26f76
GIT binary patch
literal 1776
zcmZvbc{Egg7{~9-FqScvE@s9$457^MS|_|DW*Eb(&_YpS$TBoTGS*g#v~bc<O1*{9
zYZURSB!rPeWt)=hZI~==Bzf;R=Q{p)?>YB(&*%Gnp65Q#@7|vmgHHd51^}mEMm(Dp
z8z=z)fD&}a#B#%;PynDO0|4>)%|7htH&nUr%LKf&kWwQlNVj5E5BFKUVt;gTyl`Om
zDdPphO^j*bxeQTAM;gjcqe6k>Y_Gd8<3ZSr9?on`C~?d#4PkNOQ+3SeO_!O*pUb1|
zO`dje?KXKoN?dOu48`S6wmWnx=aiW@$Kc8y@)?{Tft)WQlJ?`JI>kwoA52bNKu60R
z#nJ9?a$f<zo-$oOkTJQt@0aYnMZw`G8~90CJe6Cy4O4S1(-%&}cG;sr?z_nrsoJ}S
z)vogF$6;B$wVY?wYI_5h6POWJdY@*;M}C_fmkJ}NF2rR~=47r~N$#q3RL^!DH(HuL
zuNB1FlXt>NFKW<fq$IaZKm~QKljo=GB6pclkV8(mGl#GG)$zXk*3{eYa?>1Zal0^+
z{_Xr>jX9~H=ept2iQslgHO%Y&!K{|Ft!;t|I!nJz5O+f=xt(?-ElD(%V{2#HdLTSF
zQ?X?GOK&`}^K4~FN3EPrU$VfWebykmFv8JhS}S4*{qff|#R!cp<kO2Buiu&Dfd*}^
zY~`C7GzRNR|4XpcLwZlVp`0JKq>t{`lom=Y#wRC5RtRl7&oHy48x+EaZkZi1y87W*
z+yi=evQ0X0(jsoFBKnqQqS<8zp~w-XxSN|_oObz5Q;(NR>zF+DqQPee_i|!_Ygs}M
zcc5V6ds$I+)a0+hhJbA5>qAvi#sXwux2oj2D#u(Zz~1+FSbkx*e{*`c{X?a@QO2I`
z#RE478Yx9=25sVtn`MyEumjnqnKa(<l39RMAZyN&c=pa_`fWvS6u;Isjb+cJE|<Km
zr5kAbabXy38#-ZT7&qDbYu<mx`hIDu`D<ps<*Qx%YB}swAgMREIH%3m`wI1;;7RC<
zrx`aZn4j~bPj6!%$gMfrMKY~@#&g+|caBfTky5;BQ!9IF9L%Y~{KV&JHTvhDb$2GN
zd)yGEaFV0rM%fo&#SYl-Ur!aZ4}4HNOuyk8%6YLvd8}=wz2^SG-P<Y&O1WeQ)hVK`
z=$$+*NX^}x6gOs2L+Z8g?d-mkv(xgCrvF*WTo#ANQY*Z-@Xx?#_9cBg%wllI#pUIN
zGfCAG&F5z)GL}yf12lAk!)>o#6Y^|C$J~pX>ONU)q|Sz2TW(1y+WsTThm(`ykSE0A
z6uiD`{w$_N#_-tt?V;1B9vht0v2f+69I$m8;Ek8W?_60RXXR=t)l!WOiLAmPJs}%|
z#we^Tu+3$K#3UjC0ImJ+S*IJ;M1-6-=mvVgS+YN`4<2V)dZQrO6ZC^+i1vV?YbXqi
zg-MBjTJ2EzM3XoL530j5MB1S48d?t;!GwS3wWRy4I79sMc<>|w*5T0*vMVEHH7;Ji
zGh$10SezBze$V5hq3}io*t8lCaW?|&S<7PxaC|M7BSAhAt`Ei6Bf%$Y`8g7NvC59=
zQ^3O4*nWhBw>m}L8g6Ib=&4Zj5OGl4?t2vY4F%WZNC0^yri(AcMl%ZRS~ar9@;%9^
zzAUlv0|hRu8duz9xM@uRF=9g+RDlbXO&;g@j^v+(WDE|Y0n6nv2P>mxS<aA*mB1Rp
z->_;lr9`HFm<GuNc>*48v(d-?(_U?1D<rE+sO!QqTns6@@*}{T3I(khvGWj3s4Vl^
zQ(RJlfQ54>0*MdE79w*b)GigYg-b!S1*E~Wq`2eW+vun7p<K`eG=t^eb#F#}tVBD+
zat{i8ySfJemlYW(o+XAG6oZ4FDDNYkUt6`cgwqHwzzd!d0D{=Tkq01ThIpVPB+*jR
zG8j3mqJlg_iRQFkNtUr?9S&5+6R4YP?f$fP*i2E`XhX6x)c!wJRb4}qsH3Z=Z(y{+
d*u?Y?b4zP23pKKtvx=&Vn-7xym!m7f{Xbui(gy$l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm2d_bwd_fp32.cpp.7066935C4C40FCB0.idx b/.cache/clangd/index/layernorm2d_bwd_fp32.cpp.7066935C4C40FCB0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fce3b22bd6e8272c9743ad31385a5193b61295a9
GIT binary patch
literal 5730
zcmZvf3p|urAIIlEp2W;!+;1}&BBj!W7{t(4a!KP}J?Mr?h>V1o3%Qkak=m5YyGtpw
z%GOq^cCjvL(Pb%g-Slqn7E8Nrt=9HF<NZ8yo<5(~=i@kEzjJ&3=l}noa{|3RJqZH_
zV_B4UYJ7C^CXB&gXp_I<<fIrE5-w;k7^$b)Lf0h>IcfD*44lc@a;j<|MeaJB6dPA;
z6tge6LT#Uk_g<U8TdhU!?*Hb$@@UOStF@QG!l)knuWrHggN26*wZilUlf19!_U&Bv
zQ2Ow>v9tS!;)w9FQ<AG2&+Px}&hqhR1x;)HtmlPJ|DfA;=h^acz0GO6PU@}s#qi7%
z;j^bh_t}@PU0vT$-1jE6h}AA!uqN*9i{X@=nc|zP=Lh_5-}*~qV^2d>NxeL@+;_Au
z*D-}%w6pbX&s<SXJ?~Z8d3k-a@$JP2eSFUq-VSK7?{}A0hhBZYLog$pA2nPQdtk$<
zh-`Hyre(;|#l1^ESh;MCn0O)^%JDg2r~a;a)A%Us)%J1&n=hH;m24B9Z_?4D>ew&x
z@vhR<wvVrWcrUlwJn}_B?qc(+>OOy1$KHPNWt+i@Lq8;hXz3sJAFJ78q<%o9S*3KQ
zTDPg@m7aH=yhZ9&J1<{;2At8uk^T-(S^RLlW*7I*7rGRl4h=f--v8bFpapq+uVdCa
zJL(4d9MkVUD>A6Z50x&d>!|sA`l*?`QN80nl_8a#5ynBz>7QMCW2P%&O(TvRjcvQs
zlB2sM*C_6R!|hiZp1C&Odpj*J#Rqcp=NxF;n{qAJ*SYBWtZ8e~{hh|1I_6hd9SQVm
z)yUN<Yj&D>_JXUIZAh~4x4Ih_Ij=-T9Pb9J_`cBn5^3j9?>=t#;DdULu+dff&VT9l
zz;%Ia_`>Y(55G)P&$dXnDjht*8;H*6{CBZ2{OTFmk=$iF<}k4L>#)Eh6ITLz{NIf2
z#w;6h>b3_=Xf?z{#3tN)ZtfJ~@;F$11{Xg)$$vF7Yyb>H-t5(iF?_tM&@Su1hPq>h
z@BLQ%D)J4lNHM;7Nk4Y6Mw53>8~1}wOiV~^V$*T$lZUN?^AB!ZYtJ(!s*^evrysxk
z@s35?qfI_8?x#k&#}>AAnV#7<S$_D#>$F+M55-Z~ACqt|BhMh}+Xd@xlzcQEyYHVo
zqwx#p2NeaUZ_d6vZCB_|!;`N!US27V-d4@ExY+1>BX?TL1-|S|jVa6e<?%85lVuj&
zMo&*omz<n<_O~9BUoA3O60i5&!Pf80az{?37{?uzRgcRf-J?bK*E0pajel{6vwxk_
zFEIa7CJxyucv#n{{qX}<_r2H0%0FzA{#mtV^7^v2($lfcE?amni?$Pvhr1lj&m9uE
zZ@IZ=&}F3R*xn!BuPe2y{b#Fn^f|Lz&oti4a+75mALD9W3xn4Goc6}r3s)^saSgk7
zwL|jM>d?Jyl9$CtkFA-Zn|*iA>zW#mI%A>F`T@~~x7_WoQoed!dr{bs9JwiS=A*DH
zuY4v_%Gat}@9M4n@Pp>Woqq~W{#jJB%?7)%yzWn}&J2En25<1>IE;QX8T5C+*Xm8i
z+7r{%4wlEPn#*6$%t`wp?79T@HX9ylJ@ZyJmP)7tYz@hMw<S~@eNhSn7QgewY4Mwj
zfx*za?eXJ&f%Ug+<~jz0a8Py7pDM^4GxcNIZMTP&$KJqZ29fe=)zzj`1trgIIGb4f
zphLoDt|ny@RTKTGf|3^+{$%ugG-(~qh$ZD{ELv@<pyZWDGUH4tM#|XC5K=Z$H8Pqi
zC^@(~+WE%(+Hp41pOlSNjrFDqO3o}VJW$*A{a<WmASvss>T69Elw7ZSKqAfD-@;}F
zkTOq|r!`ej@-2DkL)$$EI@rv$q->ySpgC1ga{nJWD=q}As?=btA!Rd;na)%}$%7If
zZ}+p!3)oB%DRWi1x>E%uKk^Oi|Mq#5iv}Z<l=&P!cdDS|=WcnQ8Lyr!WHW<F*-+I`
zcdDRdMpE<^v85)1p|Xg5w&KUlg=#pFVVhy6v?h@`CvyQJiKAo*BB`@fh)4$RGFXmS
z4z_Zjjg$q>SwL%_4UXBM98K%0S-gFTSMM~ONSU6(MZ@e?-EE90m6d9WC`~=h1W`Jk
z&Zmg-H74s5AH#`4FfD|6R5OEw^1=4M;WgmIYT@cllnWYlFG{~zv-BqwoEQMz0iaDe
z2s(p6n{)^?hd{YNY5;Otg#%M&#)%rRt^qe{TAS0+ZYQLTVQh?uGmKOA#)x=>ctffY
zM(p73P<{|bB=8d~k(S6$v__P~OR_*D=89?kk`0q-GqeN09p<369bnc0vk`X!uM=ia
zOa&dr`XtzJ!Uc2DrEH87fmom}_31cq2ZVRPj~asd3#?B9(gpk7u<EoZlC=DN(7O-H
z&vr2z?}}C3G_Hyf6`)fAdei_4URnXm=<HVkBE(Ig(?oCI1WTLf?VG@_3AJ}|UH77P
zQrJkY;2Kz714n8g_mRHQ^EKruq$}nCI|p<r0Vi^Sl?ytE^MI2F28i=PBOmk;7l2j)
z7*V{uj{URVZ_+hPoTvufYI?^OFlhnhrcv37#rh;m;e<2hY=F!jfCU&JvnQ|<XnvZ=
zPNY%Vg0!X7DsUp*E}a&qQZwm7q{+ZS2KI=%LD&uc6#KQ`4(PKT>r}&t3{D1ZYce^R
zN;6@Egd<VzCq_s)Ql*J7LIxZe3K)!#14mBZv@GCcA(LT5HgK|$cTkC1mGw!2#)%*-
zNC(*>iJheMd7Q`s^DLN8eFILkgJwHDrh7ox1FI-KCP|zkdVj`4kqWU44eH-Gk%?u}
zNP<adB*mmOk^v?I8p{Eb(>V*UEIMZcmW_C4s<79j>+YluMihZZ5h!;KBMyScK{_7-
zk3)1m3?7H+d;~m>(76~qis^h5JdV=21UyRUd=3QXz=<*eM%03!md<q`sH1Z|2<qv4
z9t7v<{1phkqH_ZX8t8lh1Q+Oh5d;_M+y~)(5J9bH*9-Pn`n&Y{$O!HOtv-~NQ)^{6
zYq{|cNZ}k<o<pHc9puwlpB&&g;fOhMP+*BL5eGR_96OE{(kZqn0%`%IhnQq2(=bkS
zfv^j_C{>c)=5Muo&M3KxW^ijpnJ24$ReVU<oemr)j={=f5K8sHiR0jT96}I(4nCj5
zI>aZy`vj~-Tnej7Aq??JSb35@UL%kR&JXEdgWK20)p6nycw9mY!->nV>@ozR_TRu~
z-#`H3E3otmN(cLWRQ8S>{OCj`v>3G*P1J0=+IA(CTwD~J{-Se;G(#D1%fOauR`PCR
zM3$yW-&8uZtp;aGz=;-cZvkJ5{}tzp`2^(t%*2RF(5a+ni|e3s9q|{x>DvtdYB@oM
zZVnjcAcs(7Zv9-KT`TgyG!LC~IFS#$d@!eG#EAm1C;-~l4#3O-bZ+9rAXpBfa}y_q
zz;p<WSCzl^#EV@wOcb{nUW)R%ffKvAyXT@wQq59nm#Z|>G+L5MH&0(gwLBg7WKHjt
zhkHm97QwtCaHj<7=1HbQ#kCXNH<n;aX#SkcPNsRV98Bc&PwNEDPLz>x;x-6xBQI`v
zBHVq#NVH9n4(&5MshP=ixG+<B7UP7(RiZqJaYE`Yr8$udR><g1D+hZy&_>Dv_bj^8
zWrJ%ra>@(g5A41u)YKu}qyjW5KpQQg2{f7zU-2wSJ@4_$0~L&T1TK#th-!!tkHP6N
za$Age0(MVeC8b~G{HHfJaXW*&?mYtDBe0?bob1SZOy?)SdxE&*Kc%v=myUNx$E<+h
z3KV!4(FDOwh&wy~HFD2YZyLFQ{lM=>jzql;$b_VL7@!RlU;+)4ko+-!4U~}L+40Jw
zmvpIlyB*MiGkKY`<=+C~E%co|mV5HEgh865B`bkj31+C<b>Ln{e7C>+-v03ZEONP3
zFt-X_Tz?v9b#~0qEhL35nB9d0>Jg*D`sDF~6Rw!68v2e9EJTg^jzNR({z-Q?PoD}_
z(;QG<U=_j2%SE>#oX7*qJeY-6n-3QGFcWbB%qjp{Tn)h70iZ?JAlM9|yrT$K%ONx#
z^}-^9Nb$Tvdk~BTYg4Pm2{Bu&yvgC@wS=FI-Vr#_0h%53eA5fUUX;P8R~l&^#S;xp
z;fA?sA@hV{p<2|0)XR-rmEzfkx;SA@+NjG8_72*rQWvHw9~op!@+Gv?^CJj<L@OHZ
zceuLttXnp@lMap!!PJ1SiqHR9Hn+EejT3QL91EE^9*d{Z4r~XF60igsC1Qy*O2U$8
zB*w%vO2(3D)DCLxz(F&1fLaHgJ3*}z@kGI!qS5(-A~N)<L7TizQynq#@ZgfyX^JQP
zDqbe{6YZGd_;D)(4@yv6!mX>p2JtlzU8Cor7I11oFI4ia9Jfia4dSFo@pp<!q@wuW
z&w8rDU_>daK(s!+JDgFw@vY|8ZCj#aHf-FqIWj6DEPOp<YM}o$+@NgFh(XnjsJabR
z!%;O1RkxsO1gdUE)vc%+jjECHQE>{zCwi3%HcfT9nmQZT(9-01X)piGTASsiW2mcV
zWbEX;#MMpc?lE6~u8XDPZ2tfIuK~}*)ZAi*)y!EobLK5rxX9i?V78cRYqw0_z|(s>
MbA{j1mEg<xFGYS8yZ`_I

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm2d_fwd_fp16.cpp.6079EB647B2C4778.idx b/.cache/clangd/index/layernorm2d_fwd_fp16.cpp.6079EB647B2C4778.idx
new file mode 100755
index 0000000000000000000000000000000000000000..91828e7461c3e88ecf399c9fb5d99aa4ad98c06c
GIT binary patch
literal 1970
zcmZvb3pA8z7{|XEjNu!N+l;%)keDcRQTd{A9ZVQZ1}DUzqU174tSw?)$}Jr$B+50?
zmex9irkXCU>EgPwH6*E7U1(+PH*?P0oX*)fXZ~~f{h#N3-~aP|OuB<Zoh$^|hR}~i
za$<v|AqY}J-!ZY#2jZn5C<zZi$4VN!N25nA3C~ty1MktFN5<bLW^8N-yKuMQa>2Nv
zf-8;8SFIJ?JG=B5_Wf<VrH2<fucu@w9Tinz{xhNT{FPhlF{L$WkLSOB)Ck?AyX@_H
z_*!EJBXjmZa#Y3QCt6vro2l!x+O%P2mmuS>D_O&H!OXF<hA$>Mgp*6W%S96tjI8>v
z5kq129*LzrSzmvz=g*8UGmggbiYG7uj+NvhL1BZ@H7n;&2TQkcZ5wS;5@rkgPB&=H
z9=%lSlV@@HqrI2!)FZr(L3`J~bLV|cv|Hx}c;ogsx0)^1U%8BR!i;P$gK+=cFZ9^n
z)=&EiM$?>s<tUe>yqZD^pFb<gKIeh4n#u3Gf9I3|2P;I(_pt}-?xhniTJP4oWwB$x
z##Wi(UXgNjGS^zek4!%`QM1N4*K7aNQZ4!!%E^;KbXp~^lY3h#^0l7LtB6QmmkwgV
zFF&{+UhS{iwWA~5a%XQ9tJ-|gJ$~)rhV%d@Q-<{B=Mjgrkeff{{K85rci=Bu8Yk!8
zXmY4_H}0wlR2nz0;>M;P6d-E4&i1~vx@KC#kWhMBHAP`GZAbfLcZ+_8;<wAU7{*dH
zSFZc_vtE>+<KI|tuXu9eG{x<3PuaN;X6AtwM6tr6q&l^Dl=qbBxU@|ETo^Z&tnN>1
z%4Nk>3tBp!&c!p%gmT{9@5GhThnsIM6!G%|l^*A7+vT{7NDWvFZ0-#Y@%uYzOTZZ3
zYg@5lmG@!%-lkOt{Fr&IZOk5@`A<jmDaTT;jvSw1zPuE-Gtse(*j0OlSLJM4F&V=B
z5Zv*44>z^KId8~*<WBW@51iGaAaUzt*g@t(bI8(7PE$YYs`|*sG?`3NpTXdhI#$rJ
z^Cx@C<w#3mVPh6rI+e^nf^^$mX02PgkMHa@{}Aw{H{qL^pjW|@zbjN&(mv<n(DxvF
zz4eMu_?Y>fh(3<T;nyqN#n$?7cBYT2tlwl*J&e}P8e@C#R<~#U$-eEr?wS|Hj?2LC
z)rDqpM`Kf{#q&-}N#_O?8(bQn*I!TAs%18<{^lju^^ie$Xrhg7?mKvXNT+kuXQCkc
z)cv~NsWz{i!W&1~oP}`KuvFtrT7P%<KWiVP3{0dCTBmzHY+LF67{4KGgm@@{q$XT0
z?HH+FAPOqBzs-EVKPNd#qtAv_X+E6nHoDRX#T<{^?~Q>VOrD{Bz@;{O80!u}Y-OCX
zs#wq`RQ%EyeoHzrD=>kv-YAyC$!UuPU}{rsRKXDcCm8F7VtJfANh|=H1`Fef{mo7=
zmVshfoUF1~0N$kbRQ=WaXf^@zM=@2JiWdvOwA<k(ei!+23Xm6yiK~dpVgWcji14_4
z%I+?V-HT#84o?&dz^Ut97AI#K6BHmOiWOHWYKjHmCh5<S-CG`ffU%w^hH<cxSO9)}
zvUtqo#2G$}WuaIGCqonqz^{;Wn#P{xY#8f-VjK>qA{Kz5XijJh5i3a-EB`~P1U7XO
zbr(23$QU(=0tQV8Y;{E46U+xBjhaL;OH5>h><UaEZ`33TTVf(1BswsGWij$Uq`hlx
zQ&I`0vJ*pKtLUj%0XG-|+EO$DNlQ&cb2l(Cw&foon`GyrwXiL`2V@H-gw~42BdJwG
z)Eq!<;)O$#N&5`mp|w<1>PE0TFg~<aG&@PH7NYh2DCt6b?d^Tpn$iTe3`NEW42{6X
zQE*0JMg+DjMb-#JLtx`6cq6bguv2su(G`%aqJd~QfgOUjs7Z7YBqkL^<@+(fYN$zc
z9V8}IMAaE|LhnUHkO;*@2ggv*Pbv2Gz}N5J!xtD>A797^R6&rh<ZEzbNJJR=)0yKc
zB`t%K#mm721x4a2CnY=DIwg50Wi=I5bq&kyJFSpi);6Xj>JG}*^{fByU$U0A&YHD)
k`UZwZ#v3+nGPBsSO?R`ViMcI_>_GR!+V9!zh;f1b0|ctVjQ{`u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm2d_fwd_splitk_fp16.cpp.7A8CCC9ADABCC2A0.idx b/.cache/clangd/index/layernorm2d_fwd_splitk_fp16.cpp.7A8CCC9ADABCC2A0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5c34b78667a16ea09bc1e3e495d9d52b64b7a257
GIT binary patch
literal 1974
zcmZvb2T)T<9L68v5dsN>gd~tqR6>AYXwnj;cp{OYlmnv#K}DoW(I6lvD%d!x296?N
zK?DVnK~O~D4pFLwBAh2DO;O@m0C7N2(R<0wnsMgd%=>2N{l4A*?)Ts4#p80T6(PuJ
zCvRV5(B1$Tf*=C&#_Sb_CCWq4DGUVdE55>i8ugT|_UAWJU>omoq{g(<^Yy!~@iO$C
z>pkuiXbX-qy<WEEA3Ts!A7LAnzT-%fOK}YQBb6!N2}KCP(}}&Qm5yeAP4$=`c{b*4
zaqIV6U+s*mJBCM;rYBF?ojK#D>0#FGku>sBX?m#j#hiuQjk|rb`S`^78FS<EkMZNr
z!&m+jPc=$AkrQUIm^jP+qp_J;9i_RT^L~%{6GBc?t@X>?b*WoTKD1YNl0!YCn`Um-
z#0J`TTamHC<9i*_UGMu*OG+<jylVFIzI@8*2FJIdv2#r6`h?+Vbr<X)5>>*PmX91#
zy9N0p>xD(0OiPbDE7?&a-eLZkJRX&NJI8b(g7v`1%K#dskLOyi@@Y!=@Kxu+?Rr;3
zjFo;L%<#QOi(AO44HD_k=PI8^@fh6ULeT)5s9BJwZ?zIvCEoIgYl^b!er%)tpsz3x
zS84Egy2Y7L^<jTn-FQTJtnm?i?3)!2h4YNgo2g0dg>I3hDW`|&rMb&4_3d%Y8jBZF
z@?Mc#83kn4`J#@2ESqz_2|dSByhM2?!>@QQ-(>fU)3*I?c?rd~n3}Jz06VtS^MW92
z+dS;8yR_eC_RZ<KZu|w$o_UcM!Q-jA?$M0$IQj}FW|mCM?GCLd9iGgZ&5a#8^vZut
z&B;x%>9wQVUe7g_bL-6Gg0>&)zVJ3KV2X2s(Zb4OaDLSlUo$)}Ufs_w6*Z<!U0!(0
zJ$I|m&SKSW=Z2&pm#5ij0~5L7?b#QjMHyp++VEXvZpkT+BNrm46EW@&b!_-62r7nk
z2bO$h+^D6K_%p2rk0=iB^zJ1oHW#~dzie;|Z_cYaTaOF=J59m+b3?%K9W60k0b%^h
zy^i;92lpOLdsk;Mc`7qYP~qozVx(+#%73CsC-_Xf*{TqouLVN!r;5O6OP>(4&)I!6
zj^4eVQ3}oW!R~9~TBFS~hqS)bReV`%npn#BAqDlni(T8lljE^&vg*>av)#^O)>&rG
z=1PKIV|kJjZoZ(9)6k-jnNYhpGFWjjB=2V2p$|h>2b0uOlU3c1U)y-tpk47rMVG_U
zp6AI;nX~p?MxmEpN8Br#^WkWT`Ee(LMfPLaQJ9CVtR;u-j27eB2keVoyPuER9crUp
zEW29ny@#`HbbgdJ|4G;1YF$;RS=_uh|NZuz*2X!DkYfTiHzpx+2VV|?<V1%0+e(_8
zu_#Xn5)jcul2nktRQf1?+lA}T(kWAcyb!FWsz#Fvz|_Il<o$16nqX0U1S_GHw50;@
z@|ud>IZrwcVo@FlRz@q6r2?>N;B-8`_qq!f<%VEIv?5U|0Gn$JXpT<`1vtnL!7P}C
zkqW?}0l53;Cf2rMQJWErL1XYz0XXLkksDrC6|DkoL@-_zPm~J4HSlcY1MAKyEXo_f
zSTvR(6@c$27Y>>1&+NdWHX&F6t$>#bz~lBQYpCvv8CcZM2u7pP>QVt15(Wjw;8C)4
zQOZB0iW9KRSq|WQAY;TN2^cJb6KKI&YM?(LX~ZOnS!U9NHEF;E@<vROuw^C+Oi>3W
zkUC<LL@qOtVe<EKrhTVN2!pz5ayS9en#c!p!ItHee`tl9ZFV%pB;zFF#yVl$LAIbn
zNF_;oWR)~v4Hd9tV$rT=2evZaB6C?JmKms!GVQF*>B*qMIDrCPfeHG76QJp6CWs#=
zP^2p|!GUlB3?0J+MM3|O;*#deiql{k1{4Q7MNE<_ATv>6Djt|XZNwzG2r`p8tp5EN
zU`vQeavfwQ5=<h1m5_T85g<V^!oV21WZ)x_{~lj3faC=Y@<}HP8k##$KE6<3<j#l?
zgt-Q7kcSn}iWns<PDKr`>Oxq%#(<#gLex+vX;RsoRd)6cj;l?{%U9BEjF$Z0SBjRl
m&eCOi`V2#+F_N@}rM0cD70qOY6Pd#0Z9zG^|FlkSJ@hX+yRU%&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm4d_fwd_fp16.cpp.FF678597E26B24C3.idx b/.cache/clangd/index/layernorm4d_fwd_fp16.cpp.FF678597E26B24C3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a9c9e329b71c016dae335ec284875c4d6df1ec88
GIT binary patch
literal 1968
zcmZvb3s90-7>0jP0R=@-yx^@uD?>>^%`7xS4D&*oH|#blv}lqE=IW+)(OF9^OIbF|
zEVG=<G)tqKwN_i+%Ct2{H_a{Aq#H3=VOg|4%$%5JmKk2)f8KM>ci!(KIy5BYG#Y}K
zywDsWKP$-)f*=d{EzT09W*b0IAqIkSs+(fD;>Uitp>HobN|tQQxJEBv=e5-c37m>Z
zIfrz($o+s@bnop|JHPIcmfuZvs_HHc{)fB^|GNaigXHcPS;mTBp9IMfj?Lklf*Fq2
zjq6u6{?z6LV;b4B33@JV*)>nu22EzYLfvQeU_*F@bLF#a?;DzrGW@vH-8-)}Tcg+C
zuP>xhN}ly*#ofEHkrAm>W|TDA4}CG2p*hGqyw9aa<W5GNh;M8^zeQw38eDfSkD5M~
zP%e(RQ&Bzl=!3e4wD=Ik!WX6=GcFCLDf@*SCnc4hv$E+^_oXeg-5lP@ML`qQZJZP0
zp;PLTQQw#z_lg$d14q;`iSa44=nx~-$*=%>tXL{B?ofB-@FN<nodfbdWA+XADZ&j*
zL%oCcF?goMU%H1cunAuhjiPp+B-A~*c4*L#U~4ON@%G21PpOaZUvDU=ZW8)ktuqzl
z&disqZn7-SPNVm>PNb)8rxxNr1n{aRReS!u6}%$lzFa=Qe5EcD@G)swqve(F&`YtD
zER$1&G85|okL0w@nl(-BQ@5B6>@fDlxYI=Yny1ytElH32H)i<JsiE@>0~>^uJy2Sb
zyG`38pM{ezUNa=8-v>SIs$6$|fvZfCOj1Hc3fz=i?emYG7Y7HnUSj3*9o5dC85rw_
zXOwLYYSvi6qd|f6ZnNpx0hwe&*zB%ccjM>h-MfR{Up{VN&ajM_C<~Dj17oH~_OIAp
zmz$h8S+&i$@u1w5n0hS!j9plbrj%z`elT{~;#sR4!I;_jp(*CJw5Kf%6{P|{+1Tvh
z;wzPowWp(9vJc5s4SWd#U$FGZflsokwr`0S^8cnC-a?V$9CL0v+^LwVNKK7x^t{~F
zvAMsf<ZCNkDcgS_s$q0a`IA#0-l|1U%v!syUGHNR2iA{X33@ra$u9lErTlI|;k~`@
zH2m7kUQFwo?i1^K$FpayHhu}G_S8RqySKhu<WxG!jk<m8Xm1}YAb$S27;<oS1Y(t~
z%4>UOPGMM;iapDeOq(}_H4inO_xH7@P1!0G*})W7L0aAlr-QHQ?}i<3K28{~EZY-u
zrf+k5O!;rEnNfVTKssX3toXG*{Qk?iU3CM0bpEBDXy85BIrV9ol>VmJ=V%FbM$@<H
z&7;|3(LlRmLi0=UPO%+lWn3L?<eO*=i5BMylVcGOggCUoWm9cC3yWL>L0kfgV5JlI
z|EYV~jl3l8cs1#PMaII|7-j6B6M!i#4cjV*I}5N#4vbAuCPbY8OifY!jPJk5#v&tN
zj7FgeIsw?zX3%y*DdOUw^)RLwQZPCJSdfJ4J^noSPb@MX#uyX^uM>dFoAcY|am#t8
zP&ACq&CE$U0l3BRgYceD*IO)dEsU`!tc6Yh?%iEA>ak;gCl(n8V<VIiUMB!gFp5{&
zMSd;ABBNl8LZK{m0`TNk`e;q27t<7qfid0;PtXa#kchuoj7RFzMVkDODvnF>qy&KL
zgN$L5Hej#`j%&-XT?_gHl7>y%nDr)WhIKeFfxKapHf+6#$RLIS6F3%O@<ZAIb6fJO
z5p^919M^K5<#OPTSW=N;?NN3RevD;eBSEU5H}FZ?#^_J7Vc3L#wZ$ry!_o(3Btsn6
zh-~BrBFAx2WRx377{^7E(Qcq}92Y~zxPje4i{VqW{nejh&#?b~AL%;ZX+MLiQn*5j
z6=gBF9;g&vt8Ko1tsTRT4JrgB!6xkp=uMUk%kS3!mBJ?NAm~k246D^(ComDPNjnaD
z6Ny0z0w#Da(v!4EEJ_iR;ZFmuO^?gqg&n-h$b>{F0jz?cME%zkAus(W_~&YVxPhS&
z3XL(w;!Mr)W^9Y#mGdo3*aRC(D_c9irOTEx0s>c1i4=db??UJQduna(;5cXQJQvpm
gZj0O(dwS7*mN<EnJZMa!bx7zsBr7s#6(S7!4|qYiG5`Po

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm4d_fwd_splitk_fp16.cpp.BF340D61393FB17F.idx b/.cache/clangd/index/layernorm4d_fwd_splitk_fp16.cpp.BF340D61393FB17F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15bd7b2915c59a266c3622ae92074f977729b02c
GIT binary patch
literal 1976
zcmZvb4K!3~7{~8i*O=THW5zHHk<XejK2k!$CNoh@LqdZhVj@<dupzcGt&hq~gd~;E
zI{Ff6%a$b+NlR)>Hmyq1gxcE9DQoY{Id5}1XYV=pf9}1%|9zk5{-5`5@ZoYVX+Y4b
zZ9b{-QOR2o2!bfsnUtIulc5Gdrw9;~THX*K`%&hs{b-IBdB^ag%yv}NaeU~B)nYMC
znnCS+(QUS_kY)GV-;=Iaqi*c_b<wi!ypGcLQkNu&Ll*INadq<|+{rq8UgC!tc!@q~
z&bISf(BrVqJ0GiL9{%a6R{Uro-Kw$2qmMTGXX#(p^Cskx8=fBSd>rO6^L(<srDWt8
zFZaAOxj%~ImtN7AD?NU`>+Z-bZ`bbjvM}5!QM;S{ed`eaEosMIKVl`kEep-Ye;l$t
zg@!l{jfMJOwh(%cf2h0KTSQsD=BI<RlQvCduE{cq=td8*!gojJdwJQqQ5}Q&HT}*D
z<&xLsyC+)Bu91oy@+{(h9((0hH1wz>|Ck?c#f11l=kMi~QSh{z!xlkfP+rQEduEnb
zoG`g2JIA7|(hC)|o9=UHW{Z(qfh#vIsPCu?aS}6D!)oo~16)zIJRKeG=45qBhjY%l
z4%>HC#zvqfp|rN;`?8!}`)UHm9X|VBGaX=y=oWM<n6aa|slfPCCjU@iP9?W%)|q)r
z(s;?U)|b^*7fu;5tJ!n)K<rC5y73xLu;;lZ&xR+{#Bthwk}M~<^?7fL#V^`ltD1Q%
zwYr0sL;G7VtYq?qQ>VSE$(pMxp60qS?^e+Uc4{o}5<2teIbIXhS0CPUWA9+l(5|`D
z&-RC#HoRHivny;+9>gn%j(IzLn{m=dDvvR-KVsRoqS(^qkTE?OIwL$2Q~t0!;|}>z
z`W|!Nk?j>Ft%<yr0|DMLaqi%&qS#JxZ9-xGDCJ7*cr`CuG$fy%pXQF#>S>)|T|bl~
zk+$+@#ja9HE{hXEk3mm;%6AxTvl9+1_VAYsU$m6f+U&pH9C5cl+o{)&on~8F9N46}
zKaUuP=hZ#&a2>0VzSb|gM|%HybMRcI-MHL7@1|@rH2HJny_1Gjhg%<nOIqKp<r;3|
z9{a}*EePS3#PjuD2LGENnv=6UB<Qp4y2Tu6@|XD!mkRm@3SF&ZKQ|^xuXm5%dr=g7
zp!))Il%VEb6Rnl7(?_QwZXaQf@kFWc!pM|aO5Cy+UJ;d%M=1@*D(h!(X|Y%M-HoY8
zLSzQtL6cFx{KQO+TVP)&r)cD?Ae!8A#?W(#<9X47BiuP7{+u!2QaBS-acgVAz#p3i
z>RZYaroN_-(!U(CEu-I$znJowUNV!q&5<RUh0dKNJ%f_c;v)iZ5QHmSXc2b$CI^Lm
zAxJ>QQ)x=U{!`^aAVitN{Am*sv;o7~TG|Yy0Bj(Aqq}$Vr40%PV0a#Wo{>@jX3pkE
ziM2B}lc0?lCTo$YN&(pVQvFWJlden@=3`hBuc@mPfNdkEQ^@^I-YCq&um)a(suY0j
z>4W-X?-B(>C=|nNgiTNiz}vSHdn#Xfb)aw%h6#8ASt$TtLO#a#I(AQ?us?=TJW5dt
zz&+VzPi^)VbfIt{hSl-vWTgN+<|gtqSU;PO!hRUW<MBF50T@b*icTWKs&rw^?@}cS
z*!FC9@BtuW%%lhyEJ73*B8J*vJ|Jn#q=;E%(ns_ezy$KfOp35oCI-SF024?ZGbtih
zne-4n9bf|4V<tuLDw8gv`>iM2hbK={q{g{8qJZj14FK<g@4{(*HxQnaI6==Q|0w20
zSE1{{0KlBEN{S(=D$x--2`nj^<`}s=$Y&r#fx5Xm6HJFFz?<WlU>HP!hPeh4bWRiy
z%n3}edU43I56)_3d025ajco^71WjPY6<a`6+yF5kgW{kXW>Rbfl}QKD`PLz50y8PL
zgUUohXcVv#wlCtgDo|2lWRkh!Vv|CYhnj#o72eR;mM{Tt3WkEgV+e|j-xeptZr4Pu
zQ$y788iaW$k)%!5@}_usT2VB;sdOEhzJarg>k2n_j~{Gxt(TiGTV(QoPxTCq<}WZc
jv#?yqWMLUBUgGFvyp&<%uu4~t>k|TV)~{TRTMPXMgS@H{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm_bwd_data.hpp.D81DD3681C23DD68.idx b/.cache/clangd/index/layernorm_bwd_data.hpp.D81DD3681C23DD68.idx
new file mode 100755
index 0000000000000000000000000000000000000000..570dd29f0b939a2533eef1fb94b72f1b091a5cb5
GIT binary patch
literal 1638
zcmWIYbaRViV_<Mj@vO*AElFfyU|<jg;^LB`%(ILP41GYkiaC2jeETmO2z>QF$ta+$
zS}}Q%+#0R+&?{HvA}(xVw}@P4dpD}re#6`l#*g<4XV-0%o+{cpRj1m>xGMiyZ}OLK
zSNrGuP~gaHWPhwCz5C*)Xyz+y<|`LWmC5T3UNh}}<OH|)wJt}QX8U|DoW+ytI_pVl
zq*{Jm+=9&<Zi1N^zk46-53OhWCeOX!L3p2vO2fBffyoDR88`%`L!A5;gvVb$UYoAy
zyHDm!_A(LcnmUGM`6a(CJDS&9dF@ZGnyI$EcNg=;f)z`&ysvDo+ftnzeZBWvV<unl
zwrPu;FZ=(r$Xm=5>wJUNHp%J2KE|lcd~^DqiZ`7LyC~cF(ndb`=btx$g;UOYcKdm6
zJFWWUX6N$THl~jRT+iqf`#3U8k4>AN`RT}X{vTb8_cIQBud8s`xO3;4#)KYmH^#UM
zwHecOx23W68J&G~&bxnS;I#^~)s1?Kw*(}wth~!{phoig?$)jU8D4BW_T!-0j_H-v
z0kv!H)$Pzfu)p`{>eN{UeDS*JeLLUHD1TmcX8->6yWh=sd$DJKjDFn3s9AsO&1^W5
znajBU9H<oQS!vs%qrbse?ZL6#%y(Zu($bSj=DPCB_D6NYze3lw6Pf-pUGxY!!BAY8
zn=~I7yC!q9n!cXMFXCom<7W`j=F%1d6AT<cIUq2X6Yu$9(eBK~#LmYmqQtJG3MOFE
zjvagS<Zu7*5n|#3>NXZImH`tm>B_aMwL28%yy9Wt;e(hbBEc$Q#K6hI#3|0pD8hKa
zcI%esTpyx%7`R~ygjt1w3YZx=zzP_OQqzi4n1LyAeWz6Ex>tv}ctk|3MeN`jMMNA#
zoRN60A|6OQFA-lPp1(*Cf@iL6t_O2IkBGU4xi^Amp=V);#0#+qNAN7QE#dmqShr1)
z$(9Xa;t>&J7ZZhek%56n#F5<*&Vz+86BjsSVIj&RVya@Q12Y#U%E$!{-`+#N7p!?`
ztpJk-3WEX{rYgiwNd3h`8&+muU~r1?BRnO^Daj8HbP-ctQ(>41sZ58C_unoMW&vtZ
z5V1niq9AF7(4wrQY=iKIwy3r~%=s`iOq}32gt?GM#6!Tt7G_J!;`%Dz6-SQ)W6509
zTn;7-<QbdWAb1v{7D#dyW)}7cIZIJXBsnKhCuE+n6Ot*;qRvQi&SuU?rnrc@z~x|m
z1-cxRxDYP46|jZ7{LtgxqmQ-wTYyn1$Sw$vJRT7<b~8BdsR8R3c7E}0P`HbV$iocg
z5s?&?l!pZnFk&T5m0=$GwQc%AJ?qn(8F?TX2O7A41;RGn%|G%C7*LYdl6DBak`9v2
zNIX|b4<w$Kq%RWBUor^6bJBK#Crcg?Cl4p2<m{~HjFg<6L!6P4vx~M1TwhUYP9n&C
z#YM@*!5{|Q;|vT!AU*^Lz!(e+B4rOIA1a(S;|-gbsDO~5Fu1~iN(&>Di6NK(Iwvz<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layernorm_bwd_gamma_beta.hpp.8CC35A2FBD2934F9.idx b/.cache/clangd/index/layernorm_bwd_gamma_beta.hpp.8CC35A2FBD2934F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a8dc6369b10353d5a15c953445cc64ffc7ef8097
GIT binary patch
literal 1628
zcmWIYbaM+~V_<Mj@vO*AElFfyU|<jg;^LB`%=3&43=u%NiaC2jeDiM`@O<_Em?#}~
zU|v@G2}h0^jvf))HPXA}woN|Z;iGh9nu%xG1K!{7H&1?+ymI1}(`F6}G=o&+<yG19
z<NtcB_}p|!YUlN`E!VVF_lKV7*}hmN`{YNz+d^DFW%pdk+9PBhFwMY?f3c@g&mp^s
zZi$mpBY!-59$HX2)p7PZ|36_*?|&=IY}~9H!>YV*$(6-Lw~TTxtlJ!*b#0}lZMf3^
zS<UV3#}Y*9ZmyOWxUzPYef7rePhYQSuAQE+e=qyio<qA=?Qst*Jd-*z_29BU?wf<<
z81HF4cPyOa;O-;rW2bsJjOWq=w!hb2G6^*0S|zfYWQ5*!*Sb-ndSegs;bpokX5MW+
z`%Ka%$Q>z&Vqac)XWr5;H`)ES*OdL(muhO3u+XL7Y+v8{#UDC#Uv6A#yUnrxZSDUv
zxA@XGIe2kIz7RTjsF6WEJLwCTE;Fm^<0Z;5)7`9Rh@F_Gdpbg3Wkn#TO3lr6<=n~t
zoE9Z7I(a7~<?z?X?JJ$R`k&Qsb~*X!-gqW(_F73#R{7LzGv-y@$T9x@zi)+d&UXR7
zi;H{J%N(E0IUe`=(8B)7QzF*duAFmMymoHiZP%#y&eZP}|7{LF+b?k?s_TJxmxlcf
zhT_WHr1`-3RS~jzdTs8Vg=|dhe5}HX?24*jf`J1l2Lz3}aRw8-uAkv%V&i8J*5c9<
z0uwOl!`t&a7gwdU^DuDpK{N;pu?kr*a56J;i19KCGk&&_@H_Q&r9BS=4@`kLt2j^r
z3lk?;!G9m`MJlF$ehV>i0WB~RFp>ciFbf!pQqzi4n1N|=eWz6Ex>tv}c!Y%=gq`6Q
z3k$mndm!<=gnf~C{=z{>yb$4V1kX&{Ob=!`kFc4CnKy!GZftIb#0xc#K=3T|EDT{h
zSl}^ofdj4g(C-Cn9$G7?v2L3plPw#<#3L-qE-DHM7zPF&VFz{xI4{IcNd3h`8&+nZ
zhd71#5e}2!l;DR4o3IJ5i7?DMSY$A9fg=HCJCCr5iir+PQOn}`D&G}Hj{^hBOw>#c
zCJf{mo7o_E=A!0Ea^}|NE(kdbQ41tFM^Q&)p0Oj6DNdqJNODfrPDrLWi#o&Q_8C9_
zz@iuF%>wkSg0MZ3XB8yu5uR02QgTFumX@fN3Czof9`_!7tli%N3{nAh0eB$t2%EB-
z!g(-bnK;3z5EdOg!Zrdn&M@yiHDLY1&M)2#3U^Uqd6*s^VF^(Qd4y9;Bu$iIT416;
zb3ut0VXnJ?JKWq~+om7Xvp&6<kq452p#JzP5Vq-V{*h<E=#X%baE2KL<he?CAo08;
ze35wm5<y735Q%UE&r#bEDcw4HI3lH6Cu1k1bn6uAgp_Wb^_=197UnTvx&`GpSVAgF
z%}E3~ySOO1*d4?GRRB;R0%I^R2q2h32&QoJ{|mQ{+P+%CDk>r%D8vAd6$S<X=M_oI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layout.hpp.08D624D3F8E5BFE6.idx b/.cache/clangd/index/layout.hpp.08D624D3F8E5BFE6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aaa21cca052ea0f464ef05704c87ab13c9fbae2c
GIT binary patch
literal 7000
zcmdT|d017)6@TZ<#pQV)Ajl#Q-US}3vda$Qg1CT2qcsKzZc$@gVj@~B(P*L(;tG`{
zMzG?-5+jNE5SOyXCN>BXm0CqogGp_{CANst05u(W?_}=hAMx+<={?`?o!^~1XXY$(
zX6BTM6KCi(n)t<2)~{N!cA=!vXom9tX=~TKnhzSyHm1?6&o7->|5{6s^|xu&?_@g9
zbRBfWU}0P~W8k~+=ghgIhor3WA6hi|ouhZojC)dL^Kp01z>N0y>YG1uS^h#$#fE?j
zXVMp+FrIqt^5y!liVaUTNShp2mwQDOyH>8b+`W8R)wf^Vdnx_m6SuQl-Yj+6`}O;&
zUwi-O&S2x{->-h|Y*4Z8akS;m8IPW8^4a!sP4@5^uxa@s{puL6`%Tquqw?KP9CF?B
zd~(_s)q_@eq#Qq%`QVeD+h3jUcFCyo7`*Ev_Fc$<c`XK~qzdDhIe*i9x#+EqOSj_Q
z`(W|@>xQ+PZmpYeaKP^CTZ(p1eJN(x$royD1Mlvh?f#FkFAO^za`9}ccD3KliSmw=
z7l)nrvg-a1Ya5C#ZaA6Ua3MbJzj3yPH}blEGdVIYuyWCC!!$?tA16A0dCD){aKJeK
zgMXgB7go>_R1mOOr!V~Y@A;=s8E+2Ab^J@<kLP<=gqC!Tvj|xE&y_beR(Y<ToVH}^
zkDhOQetWOQwYpnH>wX+#dH=|jg-)r(Ng-`BUCW2temuQ;&~g3G%H7Z6#yO7q^w7cY
zM<4kwT*v9}r2THyB5%H7Qd^g9w{GQ38<bo$$_0=vfRkBh2JyY*e*+3@{^YQJz<nDW
z#3^2qmvjGv)WhdxWLeJd{lX51akaCPvvdE0)W*1<N|U-f!-t>~R~MqT5RGO*lrT81
z?>sj4)gd^5Q@TOh&F5eeM44QcF?dzVn(Ovx&nczoSc(R-Aj<TO3#z{LN^rJDEvMA7
z=vp?#EQs>L+LD53o9!BF)Nx7)dY53dSr8@l-Saj1=dF8&pd+WejoP=-%PfAOe12yB
z+y5SR$TXL7bT3CA#UzOG+r776TQTpI&DJRMQ3{!NA&WK(qO3oXx@OUB55B6yIi&ys
z3LwfXh_dmcGgmG2WtWDa1E*|8?Pjzy3!-eAx6Z$DRL2w>WSo*7kUqNqLF%oWfBtUg
z2W2O$(VDAIGU+5Um<3VxB;30B>G<C7EYXcqnwg=Qg_{LYvKwz@ja(S8$qsEfMen3{
z?thT_{a1Xtcdc7nZH?}IBcN+DUk{TY%7;OtPNXfpWMPAYIVE0-AJ+dM^}g?qopgHe
zuR?1a%GJ3nFqcJ`1yS}lJ^8A2&V(^T(1KHXLE8%kvmi>|m5XN`!tig_ID}L3n0Fo<
zZ5Bkyw`(tK*;LzUjW(Rpz{WPP*=9kM!+!fVzO^jwhBaDpN+)`EVvt!7<&%_(EqA~E
zHpFCb9ge8O5XB^jQkm6t@ulKUD@*j?qg-YBt1Qqgh|*{i9#wJIy~!G_I3<gDXR%<j
zAWCcet9ftC|7@it8ad@Av%bmf&4MWRYacH8#x>*W0JL+~DE0&Gb^Q;`V1$7h4Jh!R
zkC%^gJM3`S2K9W5T`29sp=Lp2^mZEu+xck&Ez!U!*O=WkW;6?;Xx1!Qnigv(@&#(m
zS%M}YD~Hhb5c<;hC~Ginf3^^`UWn?2=V?D>rI=Y2GjEER0Tp4cu?e46F<Chdw#OlW
z;zU+Xg6&B)?gFbW7(oe4j=V8_h3}lf{KX9W3?l`GthAtg3mVBuWu+DMt>{Bo7nHSs
z`TM~wLu5s-=sn5RWrb%kPfBgF5+;Y)kVUeR9FshbTv}GL!8#l4#RngN`~YlznU~T5
z<2qoHzz;$H5ZnZQ1c8qrOyI}p@fb&tAIi!BHt_(PO5Uw2FRAgdc)iM2Ry-w7*_5I+
z8d(`HjaQ*`DP4sMKq>$gt^=t~jd!B76IJ>WluFdN6s1x%E=Q?cjq6aVQ{!fon$@@y
zrA{@@Vp0}UeJ__uxoVupq&zh)WKy9TmoceKjcb`ytHupXY7p4U@h{_pnr^N@Ntqin
zH<orzQWnfu@H`>wZ7-M9bl)&=|C#18EtGzil?5&fRHz92iXfhx((&5ogL;ckhwzbm
zFrWtmY1brW2b;TtrBINQC-Jc9tB@iZaxf$ZLn&z7$DBOd5qU}b%<-tI^5`;^R54;1
z^C@HFDD%ramrKpeh0?tL;@%h++<4AHQW9Jf+{sKy$-%%J45o=lN<N0?W2C?>IJE_z
zqp*>97(L&Lvj|fO=kqZYZZV%)=2OcesXEI%u}Ih0Fv>X*Tjq4XcV=CSEGvo8iAgjc
zSxJpfRiV9L+zX*1#&W@!tHwtm;3$ltZQ&n*=xdNPw=vwmZr79OzT6ab2ouQ~S!qPy
zMvSIJE-OD^><>7NtdW&=3~k2*fq%jgKVb~*uB>cl9@|+M?QW#!BF#dp`RhPZlBXp<
zM@B@A+BWm%pFVj6vT_gf_aK~NJoXQ-7Jjwx_c16dr$B!S;>5&nfpQDR6LG?|&BmiY
z2gm(iug;{V{FuDC327FxvJ3b>U$KqFU@QiIVSNc0OVqd&>`TE_=ud(5DR3f8heAN}
zQIMQ2@uQkOja%;N++@Wo)GJCP?l9XhM}gw~;)7|gWMxv)q!|Jw1to<Elsqf>MS)V1
zQdG=&LGx71^a<%xgnjELtXH87j|`RVnJ_jJ5(GCJ9J0Y#-~t$1010HP%*(CY1@I!z
zkd=A}sfP%GuR!b-7$<NOOlpFu0=L10Hkcwpq8;MfVY0vv!1e(+3jRZIe+b?JKZ5W_
z5F_v|blin5BBXYs<8C$1LAxAu68d~}%11?wqw+CM;ARYN#&8jGE$GpLzG@uNg0TX(
zqHik(sBvs7jw4L@2DIiLfP8~`0Io1K0Hk=aXJ7C?Dz=3FYL%e&fWK|(4e0Gub5p>5
z(bNi%V&%DF-@F5lQl82UuSY(3V^-5z2Q7EJz)5k!xsrw@83jrSOi}6P4V$OZWrBYu
z#0rPG0=`!uQfyEYgf>CEz-<uR24lsBw?j}nj1jmCM|NQ}<+wFvGn!N996rc%)KPFd
zs*cKQ;C~IG=tVlY@g*@eI%GMyuB0{hHKdT09&qh}kyJEfB^PaT(I9-G5?w0Mht`z(
zApQkY6GQ`%?Y!#rwLbK-tn3Bbz2Hu2SviQ_2Qi$+r`pKJGqptY;*LbN_5PlpBYCFs
z*ZaGROp)l6=q5HL)hShl>d?IoeMKp0M#pAU$PcORO0rUgaus@zQGGSX_6HnEL|zJ*
z)hXsplhc};Fq#{kOf9ti%@_SFD<S$27ctYh5p$!&OuJF(MmO40nIB5@JD3ZVt(+@u
z|9ZRP%}<|lO_~vRWm6YN-}=z?-TnKI6ovd>S0h>+?lXOTDA~pj^&^s_Nh$L~dr$|A
zrQo3km3xe-JEhmCjp3u2dQwuzN)Zk%LQguV$x1bPRbw=5KON`!h^9`JUZfV4D@^_A
zFNy;rzu$0fUBs{Er-HfJ#g<ck<1vV*mEn^-ggq(4Q&eahxNQSJ5g(Vq<udq?wX)I#
zh9>Y8xDE7e;6)M7k4NCt4xR#cp<NfclAn>}CYbtXGU4!BJN7hKMkGAta1=Y-;Y?jL
z@sF>su$Zyx@I)Rwvm9nA!kVQHOI4@@0!kp9(v7Ti!h}vy`A#%3*H_xv9lj)CZegAx
zRI*`2HUx-8cmO^Rz)u+c5F#F`3*IA0dIZyjJ_mhr(2tUe%umdL`8Z0Ar{?1f!sT}}
z${n_M9DC-`{>;+Wh6&mWxB419ou`AU!PAXbp>$ok3Kf8^093dRbaiUH6LmXLr7uBU
zi5i!pu2hZ7QCF_Ub*QUT<7U(~t8pjlI@LIf>9UyWd$~-PtHyaum#4;sOjoGJWlUG5
z#<fgWtHupX*C6oOyZ&!Za~|_yU+<PU?U(bntSod~=q$XZ2<8{T%j73?MKe%Vj-k~t
zG}6gHR!*b#G#V&WWu+2>D=|UfYK*JKnF6<ALK`Mi-sIf{vub2Uie#<%;)lGt^6RU~
zr7;h0KhqOx65n&bfARHmyaacndpCMhVD%+(!wzPoM1Icut>B20iG@$a)t|3iuCKq-
zS^KLQgT~;&YPy4>X?_#ywPEIT=MSDa&C7pXH)o{l`As~)Mi@r;i2yqW_Q$}3)>~Gt
zgZp&|5&_f$gL=SToC12mtrr3*rOLds^(w?rkwO26_J33t(0a75M-R##R|D_0l+9kS
zcYv(;Sox?*;#P3m3V|Zd8^N^^0>m^rK<)r1fjhyqQ>{P1tPU_2O4!#s^JgyGw>|S$
z7q!)EUtn{EV_N^t*><Bry>nJOm=mo!-S7?WbGoq87$cmj5`8N%P~d7bRI9CT!^k#_
z6HfOSogRxE+}EOpZ)dUeA%4bUQ*zmKs$_llZa&2<fKG2(^SvVN{;kMw;=7JJ!ZUZD
zUf>n5@Ah$R`qIQLX_=YwQ!SbN)UmiD|7=a*2{(^tj^EYu*>Q(Et9J3FECshx2owHv
z1C$#uiZ+fCEk7WcZaHaYeU21Yh)Kecs?oX{UBuE}MBj@TPL3oiZ8)M0BWP*qs)ARc
zN|Y<nNxXOyZEvC*5oP5ZvpUCIsQ1vCFHvbCYnH5B*q6@J*1Vk7pBt5BqobCszT`g}
MTs5wJ8q?qZ0CJ;OE&u=k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/layout_utils.hpp.9A8926719F7B5577.idx b/.cache/clangd/index/layout_utils.hpp.9A8926719F7B5577.idx
new file mode 100755
index 0000000000000000000000000000000000000000..54325797bb4c68189546b8231295e0b3aae04a7e
GIT binary patch
literal 7498
zcmc&&dpuRy+uv)gqikm*g~JpM$E93zDY{H@t*x6>ql+#=ku(!Ex<`~Qr%8=OWg1dq
zFmg$yQ|PLj(nu1eMy1K9qm-uM-N&}<-8-Lm{PX@kKc7$MS>N`v*Lv1-S<iQEAKKH?
zRgWOt0_injL96@)1VLyp{~}g}hs;(Yh)@whtT`;3J#X4GC(YXto3|Qxp2&U>_@Le_
z_EY|Ar4Z5XpU>sYO#VG#&A|tA>QeJ0oqr8k7#^D9<6R;#jsN&@!{alb22Q!=tiG>o
z-)C^}v&TaBoE_Igl->&B%*{h+l@GOhmbvJc3HKh;43cgcxXNnOnXQ{cj~Rx}j1T_V
zImYXSyXB7C(J#-Mi%n-Zzqj|Fl;$(nd+kG)d4_hy(*#!<3$0J|%P^09L{WQFqXwQn
zwLw<ZoYI)|D5S#9yJB^F_SBra4_yjgC;PQgKdF9acquFQsqKRLH>P^(=R77>t2g|*
zK48UG!~Ki=?)qwNci+6~TA8llu!Vbv{d!~DFGkW5+jsrncCAQx*k<jOp62DE-Kt@U
zG@d@c5SiT>D9)T5u`S}z#P1ty4a5HgTH*1H(>As3ZP4A{Vyr6;*ksV(F+mjtzStWu
zzO}t*aaUPC>y8xqn60Q^^TgvH%{C6Cb`P`o+jT=drTe(F;K*8P&xkOUQ9p*5czl!b
zaMgb4{`0X{RgT}9SR*@J9er@+<MrE14NgWRpRa+74IMn>cj3rS%SEp3hWcB#>IGD_
z1{s(YMVblYI}YXkd~(^Gwwd;26$cG3>^yGqY+{XHoweN*kLd?KU){G!E4=jXnX<1B
zxUO(4T&Of5Ecj?@etqSi52OB^OjgI#hJ@KDnOFQYeCga179~4%_vLDQ|4?K1quOuL
z{^z2s&tubC?>juPUF}QVTc~>>Yx?CEo^RjXD{kBMn^T3^gJVT<>7$5YJ7K<sceC`c
z(_`6Bd#63#B$r2g6LrJFW2(b>OTX;uKbki!Gx?=3zU_d@=%~_}k^%Y`GH>*cKTs2F
zwQPJz+(qjJW+T^(@%%KhBS<gqk-^1~He$O~^rXO_pE`MJpBnCJq+Rf%FfTvl+thjS
zvWwxK4co8jUx#H6lB)N(jVaXH@_l;ivnN-#y+2+#ud&11|IWR^M-9!2jW%}u<ayxA
z+goM@pU>SsF8kfoY>DdH`oy-5{1I(0ikG&7_MB6mEvngdcfw~E7cB|g)ExMF^fIN-
zqt8~vY>c`ZH!Id+Lq`6|bq@oo#yWUNoH9&uXZ>36GGDaYzDm7u`RMHp&Er?Kclrz=
zeZO5};p28wPx_!lM;5s_Pa}4w^tWfHUv6;A6W+3vmmC_hQziZ26VD*KkOSvB=fpRb
ztsFSF`Dw*{m8G|LT_*x_ydCxg=XfM6@SD*5>S_77`}@984sEcUFL9Y_R8k*VSrPbp
z?A&*17s_UYOo+UCrRs>K=hPv02d?k%*k^oZs9}ZHtrLr<(I)rBqJ(=-J#6=<r)9^c
zsXu7<jcON>&qAgzjZe4uyrV5^bd=HVA2h@Fc1(6TVj46*_|zN!t0<;?K7Ho9)Uo=n
zu9&^u?5cO!Df+9-%j>gW8ON9<y88I|X&w=2-O7(?tUf(r(I*psl_^z=+OPD#f3qd`
zNvW5Xw$dBTAJW_NY{Lpm{>YNN9BVoMF|dmD@;f^Bfk+tD;v|2s=YM}ToWA4g>C5V2
zOR1DslUvt6OTYiaU-K8Wtry#GN{X<2G%0r0;?ClP^LLitInr{h91S?QadnH<!doM2
z4Sck!^jfM<IIigT*Sye@<`tScE_3rHjEicaJ}$m~`q>I=-`4?o=RRgO3P%<ntq+~D
z($}$ZaLTn;J8rJ59#_6A=R!lX!HkvzZ@1pRy?xn&EgNE=M~2qR)|;tV<-0hT>$a|N
z`0K;OmTI?>H32VmHXAPp75$KM@>SURv$8EKO|IueBtB0ZYw9>DGvoKW_p^siJ{9<y
zv=2XV@=W@x`jeUG7FY{QYP?TIhmU%cS9B`N?>CPn3E{)bUiz(lynV!(Mhm-7YC5W4
z%Oci>Ev`lcfws6fm0r1hmLkTECTKa7$l=f)L6AcRe(gvA8Ys>9Y2{s!zbQq`WR#kK
zxCt2d2!b>e%^Imt8k9K2MHm<Gn4KvJnI<6%niT|T{FVA7Cq~IgdvW?OA<CtQ>57@4
zqzRh!2!gaRQSM&Oc*cjuRa+v?QG^$xlnNxNz_dpY<nW<_ZY7I$i+t_JVO<n4l~D?o
z1Y39i5TsapO1HxF_#Aal;g9vK(HT&Z!7RlJg0wOvyOriNxWDih&ZuPVDS|{16ws_t
zY~QC_X<@_uk<UCQ1+p{AfkY0>XjTxU?J!}tQc%8;LwLa0t?c%@prng&Br6EgNC$Q+
zElvF8v}Dc46d|0<7%lE6R_p#DNC!)%=P~{XOG|chO!AblOG%+b${5WGg*__Wv#g3-
z;uRAS7YE_&QOs-1AFZvTtwE41*m%3HA-ev1f^}aFkkl~lV})Yl{k!LoE{_lkepTJX
zZoCef)<JWc6^dvL=#JLjImQ~%LQ88_sS=u2GD@r<NPSIKDLM2`rfj^IGi!1ukaRNM
zVFf`-bk(|-OD>pn%TKFDjUv8bHeQP)waC0j5Tv!4Z?{tV2c^z~r#e=#%aub(IkS0I
zC|s!2t(5y<=WWzU8?$kbltfz7iXVbB&{FMIC^$65#Cb)%Aw|q!mdZuuxyYtR5TxE9
zFh=3$8xDo*wwKs^O%c-=B^i>)kZF$~$iW)k-7{4c&(Mw7EMCSsG7gC2fP`j+BJD?X
zr+w9<#PiXI?Sk02ry)ri8cMT5k<tCSXKFke7Nja&`-z<?0Z0;<kg$Rv?aWcPQnSst
zhT7*N(^y9)LrF5TeO3^py#>{+ByTztDDZ0|+2y3jRLWEtD-<DNtBeQ_S`smN5Lew$
zrKhHG>QHn#5~ia8*m_*yNTsK)ImJUzcnGTZ8b@bBVJ1|^)zz<Lgm|n+t_)K24WM}g
z*z^`J9VJkX5=e05RflgsxNN^|ohC)wi|qAwz3_IKQ{cabq8%lU!#FfXd5ju|5`ktS
z_^Q`?bOEp`0QTGxXMyrrpvGYtaF+ou4qpSx*Z9v5LES^pfZIzYbghIFd()h8q?;7c
zy*{QdBAbh7L~pV%*!d#z;BW=9uRx>t*u4UIariY-evQ<y3n;o3xwj%O4*!jm|K>m6
zB5>a#;60BKh++hKSYNg1#D$?s-^m73v;|?I%vL)=P_!rE$wR?JFb_o%kvtR!h&aH*
z34lo8<5WPT@^K9yYWTPb5KVk62ZWrDI|0$j$5Kd0A#c58NF?)d1|%~0xDXPBd|VER
zaz3tvL?s{BL86Y2n;_A|$8t!>`M3)bU3@G>gcR}YOF~2vAEzNAjgNB?k;}(2M9BEK
z77?`^R@eE>b@;u8P^PEk0p~n09%rV2E(Ok|9LBwqsZ+LNa^?4Bor;b=f2&*Nb`wza
zdn9^~jBsv<%0GyZRjUQ_;387Kh=$-K)TueC{5hl8j<LaUzT*O1TLL;7PK$=#`1BS(
zbNx2bck2eGj^9J!d#Ke*4CrVij7D0xzu^wfn98<q&eRuB_yVe91taIeGjgXpDt#=B
zUD98rr^R#grGvrgKp!hmbS)6p0=3@u#$cUVFqp%ca6l#;gij`legOx(fP=6-iX%@W
z9ckht4-*KW(oG!jvQ~kg$LPV0<pKe1YiKLs{592bs<ljW<GMRBO3ng$mgOuSdWHr+
zL%ntC7~}hC$LZB|$`oxawzkDnP&6&3ZSfQoJ=1iiBZp1^s}o=xx7kc!mC46hz$%N6
zvw=7p*mC+&P#gtqaX6h!78CyJ3s(aHeHjeB42EIFQTE$>8ZQ3wo{4iY5ETPE{D7jH
zk*FEj;e>LpiISvu*^c`!Zo(`1!k*hPuT}d~w56G)CFjQkV3h#uxVF#<tUG}{-aExK
zP3vT6&mHsS(77DC_tFdc5)@v71326TZMvWX&TyPSOop>`5vK<Azk1*7L04@T<?5I)
zzVuf2w%}svtM1#2X8)CB`@AH^q0ch_Q&DDFmVk20epy(UjxxQPC83<+zxQgqL1qP(
zm~t<EdCFj#%9MtJqhht?<r%{crM#c_kEFt^m9dEBs`w3x{&$PMb81~V<K8L}qG&m?
zk)v@q`zZPyvV4ceh;LqsuNyvkEk$=A^A0o$V_fw8RC<^#x8ncrDG_^g^SUV~E?;HN
zRWA*%uW*ca^)F4`ULGFB817~6WrP2lqSww`JExbT((8TK`(d=p3&pnV>2hMSxf1GE
zLKBX5)It3^Xwpl`86NIh2WbvVk-ij}aI_=|=_er*+%zbrXSgLHFAjI0uR4$+_7$Fo
z2|G*hIbVI5SlA#&ktO;6r()I1H_;{QbN`VD37?N>>}b_XWW1FHoU;IpPnQGYtp^Vt
zJ#P13JWcTNz$7VqJK=Ws?@5}xLx1+k=3nkHn-Uv{O>uGHRAq{Uz0ctJ4yRZ=-B7;T
zvCoSCd2W1$8a>mPIX{kpfyaOz_Lia=bmhnxFOj+;VmkD#KFnBf1sPr87-`z3U3+e-
zIePpf5cq6kLdjmBuyKm6$nVmbzE$H?ibJiKshg0f30dJav)7a;d55KWDp9lySjvC{
zXIvQ&l>q||Uk0MfeB6pGTag2vh@w9s(I;epF|G&3Z;Fc>HZ1QR5-)8=uNXrv!NHfH
z0p0?>&oNtJFLU?}impcn^_-WA{#aKe+9}WKgK(T)^5mww%JoJCizY3YOA`CXuxQeg
zhl0sq9*QI*c_<E$ae#*t0GYtYsenx7;~GHL@Np9$oA_7`NI4&O0<x2jrI3_D-g?Q9
zOy=VZNM`VHAtVd=xEzw@d|V01N<OZGWE~$jL9&UD<&c!~aTg@J_*jZaDdO3egvca5
zPD5lGALk-6myczLl<{#bB5OH(?ppPnoH1=peZ=*G*AAnWTi=y43FK(w=z!Tmx!ERL
z$3m}T|0t2m=D!8}aj8hk=obJ{0kFggeYNWDjT^7tmG)V`m|`=T%zi<@nOQm3bc4^d
zA&GrhjK421)R+CLfZu1R>VK5_VhYTZ6#J!ux9}IA1*&ac*A>iK^dm{pHbNU+>{E))
z0pc8BjXkC~B9vu7o5Sb9kn_Nt!`G1hHDrk$M$wIENFy@GM}p#FtlWyUIs68R-ymxa
zZxJYO5%fNP@b0iraMI4WJ|k31_zG&?D*xT*t>%eol1*)iP6VUqC~`F7HX03u(NG)r
zjz*_Sw-4DmppU>}X37MX{m8@f?|(F~nDR0`j{R`NE&T62uBEgss8YPHT9u^r`s=B3
zuZ0RjHSkS|qSp!6Y2fplV(x2Jsr;`oH-Ydb(BOhD2i9`nfaxm5a6@Y;biibhVrcG=
zB4~tpFVpm(jU0~PE}k-EB}4XH(%eA88%Tr0waB&>IdVy|O(5JR(7;%J_{;^HJVWC?
zVue{1Qw!{8G;Z%-lH%~7<^GCR5fL5`u}3lRQ2ele<MDtXT)Ck$H}vF&?%Z$!H>B@h
kC)&@tj`~|T&V9l}4^Ox8F3zsRI8MQh8&2egF5K{c031wHqW}N^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/literals.hpp.AC03AD34F59B947A.idx b/.cache/clangd/index/literals.hpp.AC03AD34F59B947A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1446589e3aa9c21e1301752b7c5915a4a786ff6b
GIT binary patch
literal 514
zcmWIYbaVT~$iU#7;#rZKT9U}Zz`!5~#Kk2=nNNVU4N$IPPHDe?_F)5_qqX;z8>F_k
zI9)ZH{xI9WS(eMTPcX*a(<S8gkMp}y-t5-TesLppy865&--U&aK7Q_(D!p>^4Ue*I
z(JGf(;+JatlbrhK>yEXn-p;OB*y6a^=IOSj{5N+gbgKs4GkdMleYk4x`cv%<YbVuu
z*=aRoZ@-^A;|;?<c9D54f#(nLc-?Ig&zY0!-BZj^Rqps=abC%i@(=I27G4YcETP#{
zDE@wto$)nhwr(Du=L-&Vr0_mxJIJ$t?VOBf{RQ@$UhWkySDrIJyC`w%s?T4p`$itQ
zqOo<4>ORJoH*@C}SLP-q0{yDM)nCvmAvBGLfs>7aol}-m6-+R2GBGl9voNwV7P9O<
ze^$J&g@=I)sDMwFPZdmn6fl4lOuxD}{l;3oP$ni0R-l1Ez|O|SCd9xA;V=}XrWNM{
z*)W?JnZP!{w7KnReV6klZ9WSRJEs_@3``XdJGTtC5`xDk#wP=J3cG-efD()cqk-mv
gTn953Xm1Y4gABz*$;I3t2K$DSF+tA8!PX260FeZ?o&W#<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/loop_scheduler.hpp.CEBBF3BFF5C82A80.idx b/.cache/clangd/index/loop_scheduler.hpp.CEBBF3BFF5C82A80.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d08b4b335e29949b24824a5c3209022f90e6f98c
GIT binary patch
literal 574
zcmWIYbaOLfVqkDi@vO*AElFfyU|<jg;^LB`%-ukGG9v>+#hkSh4D${f@VJ)F;CU<b
zZl28&=f*|e{0at__*FW@SUt-o+&*sj;NaZN1r_rvSoZ(?VIqE~Vp`t5V)wnDnLaO$
z`WMqMpY`gFhbI5}^m^kT?^obGY{plSx?Mw}L*oStSFi2%9RU}uS3N!RDP!Zr%PTV;
zuRfr5g=gW6xo6~_2$eorcD2eRJo#nd=f1GWmr>^zPBFhTvwF%6`SnXZH}fX+m+wx^
zJ-%)4JYT!QL%;4^+im;o8Rv&T({3^pSLP-)0R50;<hffoao-Mp1}1g}E*=&hZZN^X
z0h9xRLKcRZD-C{MU}a!oW8mWA;gSRs44hzDMlQx9-_k<Y7adr@!@vzxARr-N3?>*j
z85tQ^Sb+-u?Y;dd#O!StKLaz+0$vtgUN8Z(peQx1crMVegbKUVgG;V0XXN1$Vi7U{
z8xJI4&SL^v3v?UIIXqmVQlg45yJ4ctTwup-n5Vrzb=SmgJX``?0^%@5KrP&&5-=W&
zX5<1p7p4W~SfDV-r7(jTic)hDK~5?zN-kCg(o9^<Yt&!O=Q@?e#K{5k3P=GYTtF-a
E03>9<SO5S3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/magic_division.hpp.7DE3E22991AC58AE.idx b/.cache/clangd/index/magic_division.hpp.7DE3E22991AC58AE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..db07d13b8dd2579d1fc92b860f6cd56e7164bcde
GIT binary patch
literal 5530
zcmZvg3sjWV7Jz4uKQNBx7tRdo49@}Op^TtB92Eos6(kUEiq8unX%eLfBobm|khXjw
zmJ+DhMX#h_ND?Gcnk1%(lEICL677~|+RYwntybML{Iff!wVJg&XFm3x^Y4G3{W!-*
zXV03oYlKP_yL@)Rsuj;HlT<1dW&iS?$$j!7u)o`>R0aEv&F@(K<plE6$;aMz%*+Tq
z`Bh)~#%asf_kJONTD{6U*ymQcsrYo`zQ#E_)$1puf2i|pyWX|n-kXUpZ`)EZ*F9|W
za_=dzx3;c7e;~B`R$Ix>*~zQS_a}MYf2Ml+o~<Q!et2u_ckZ1z%^i>AU%uabH)6@t
zOV<2U(D+=D<APD+js3e<eEJ!5MEW07?@w5H<yOGCi~m%=_*O&7iRd2t+R=rx=IJcH
zPn!20jQIA8lXlNPIwtm>_u-fJFFc%O-u%Yoz->|6j?VveT*ujgeK%VsWF}r7xccfS
zOLV?rt9fW*|MkO{I!^6M|7A^@>Wlw;-%%JJy!Y<mEw9H6|M<f~hvX9*?Y*X%;=*2-
zzWoFDL(jWho@A6SFWv9r{C#_2S&~#1{oJV6M`n~g_Hg>L1?wWttNd2zcDxpH?lnVG
zc5T&-o;xGfy)$y4^MhZ9e&e~g`NYGgj!*c`XZNaCe0SPCs156`ulzQ){n$B|=taja
zJLhT&Z~h~9miIdM&Lf)Nl66f#4vqicqm?JBpL%%bqk|9c{k*qrjQv69Z42Jf1TRgS
zn6F-^+ZY*j-}95}zrUAK=-u(|xR>kxpa)*{J^bZM>8GA`n>(@S+IFvN>n}WcQnR)A
z&b*sWdC#uO_-Keq1-ql|A}0TMP^VT8Q_*-$y!+sw(y(1;A5GiR+r9UF7wgo(#Mzo`
z)8K$&{pIZouSmH+DPl{oC4pK86nhkZI5n$K{<o9bs8!JhaBqOowh2|nt-bm1l-GgV
zPHMtbTEMFX#@MDmRodj^`t#>!>(mmnp$w!l(Ay?dnd#>L?3b6f+Eg}38}!sVpg6(k
zl(zY<KG;b;e5i^JfHVMF+w`Z(ds*eh!8@y&9ZaPUy!x0O%7iLOFBIPH$s6W0Ts?}Z
zTmtzLjIvFrGN*jq&z(2(=IPW!n5mJ$kuwJe6fgSf=~01mYIcoO+p(9e2SYu$*(Ow3
z{Y!gCXlxfVbp%ss2d{Q!hccncrnOg&zTtAEb)?#!+0g}tE^xO^sIn#XR%*)I-QNsX
z8<<Kb$ep0GO{lW-^H-Cf)2`_su9n$@szI&>XWN7-Z;m{+eaqGtdz9Tx(WIEDbwF`*
z%At}~rn_gA(8<(fx>M_b;`bNz=^Gc_(=!{`0}g?Fh}ob_sM3~p=HY{|H$S&m+i6tv
z05}|gVYW#%bTIya!lk~trL)tIZnjt3Go=gQZ~=&I;!0OuaW1@}``2d0@%fs3S85&b
zL7&bXKZg1`8x?VyCe4*v2Nd`3EVIAquyld4FEwDOVf&&?sPff=-`;xsXnHCOA@(+B
zKsv)hNSRROrfWq*<B{S7d$j{I)Zk!n9vrMjj%0~O;kNmsL&p6z4vNEE16^aOb->@>
zf3R%hm_JTsDErzAhF-R>%7iKpDz!xyKM71@A;_#a2=YM|g35#{s@xS%<mEf^%vNg$
zGa9WUG|Mq-4DQ{K-1Xs8+P;o*AT)DArUm~@Xm&{UR7`$C^8@q4(8YxA0Mice8)r*P
z!gA5M3|u+y0gpZ4!?^;SE4Un2Wp-KX3AZc9F~<*u+Pd#MBdUJ?I)+f!8Lo-wIYPaO
zw=<_8k04)8As!(j1%!lfnPseHB33IxEst2{af%!pIT16L(8yVlb2!CMh@Xr#lF)>Z
zgm6xYx<nUFNxCF~lFiAXTsGG{SD=)TlyJ-~LKo^5x(F1#a5ksKUW><aN()Szgpo{W
zx-Q*CAWM1_rwp%*v7E9jSy6mvbEf9RpqmLT2rHPzcecn><d3zK&<)NT+ywI55F}8<
zh8diSHN{SxO2DH8d^wkbb14Wbm4aU>2y2x>L@5ZXZ34|E5Y{V$ab*z9<Eady${>z&
zIcUnkN#wERFp={v(CiZB6)>(sT(1J6D#Z0FK~pK}SHjpzalI<gREhFxkg9=TK_x69
zrGr2)R|zY6(jg#N4+*O*?ls_pMTfACIHm#2*enTae})z?VP~|mPYaCad<KS{0X<eE
z!pe~68SoQ1^bACBZU;j<n6R@D*0ZEbK(L(=))=HtAlNeq>m?pt;EQdBur_Gu15<*n
zLt5FV55{vIfMElm$0kcyXZ0KaKaqn6V5-P*1CWd?Pk5;+Yxl&@w1kH1!;L&GWArgb
zo`{L&#9*FKN#-PhYQb0wp4_2zV5}3l35-qP$sODb#%7Vb!PpI+{DphK*uz<V`;F87
zc5B`p!D^A@N(TE98(0;QECNMJkpjg@aRSYk<_nZ6r3$oCS}D+KX|+IWrL_W;fK&nk
z?*VC#$Tc9<h+Gd+y~qt9HHh2-Qj5r^K{_pRJ4o#!cY)L;axX}|BKLvRC-OCru5reV
z9ikmPBl5bx(0}1{UgX3sfIG%ot#Ww80|Vv)>mkaK5e0;vg%M}LfWePNfGH{mM|`Z7
zR)ACidgQ=!0WOEX{%)5<XaETiT8x>P1=<9oH$ea|ZB<~df>3<Qj0<~i&;NDwB$?27
zkN6-SuE{RR0i5Q#%oQlbD<zn}{$j_)CQfN?Y2$ev&v48zamw__4B}Ni%O`6Rr);n6
zU|!>M{Bx#o%6H2j&ntex=z>rrd@V^kczQ<HXlW@(rQB77wt`2iSdjX`qn{UqxR3XR
z?_Dx8ZwR45`XG1SgG2No0wp^p8__3N3;sNo<Dn0W4BJ2DFpdI3Z@`EfV8CBuHDk6b
zCv&dY4VrGy@$Bve?_MzDX^YT)n9>ii_#~`2EKIHEcXR^lx1-aA?pxqk;EjI6V?DE8
zIn!e_V##F!<p__0oXOgwFD_Aipp|I2E?kePA<-CJj2<nJXktL38C@jNq<|!WYQezD
z9@adG)`5YQJ<d&FU}X<8NTSVPU}cYUHyBviL$64*2Mnz2A>+&}S1uUnPV5rQ&&ss|
z1%$1j=>ZvkN!ZbKVlPY*{XDZD641}sF736pTL=2a+8={!*q+!9<%Buas~-H(%Y>c)
z%?WVCS0S_w0^1-AQ=TwiPilv0=zPp37DCqB3*S(&(j33v`h(Bm_$W1@<zOlYe@tmY
zE5TIB8FP%SX1(n2)x=ZvJ5LN;nxVaT-P$o^*HJ>pkSzj5%8>%a$#DYBm*)$VDyIsx
zQeG+0YI(IlYvr{96-h<I`O|lRv;zdL0l5YQ?Rt>wMQ#AOLF5*YTSPt$@@bLVL2ehh
z3*;`5dqM6Mxew$%k*|S#jWeE??6kI9BM*=|(A0q}21rCbMDqY?22C@_JV07uW-BCM
z)4+9DhAUUj<F5srUeQwCWQ{!Q)ejY{k!QW4BV(wuA1c>S{=>f}i8c1Ew>SI&777+L
z)>|nL8nKFCWoJs(yXoI6;XAXu{Hv`P{Y;?TO3`Rshvk{|O3F<>0{SCh!UYL!1briz
zM5j5nf*U%G(5s-o3MQW3{ovRSZkWzk<u%&DJ1sty&?)*U{#f4#ovxqmkM)gRr`!{S
zQ=FJVgdPF!BVZBD^KJwSH}5KVUj+-AgH?vT{$E{>r*IZU$~_tVLTD{$YC*@JP!Hql
z!NQ}c2{cWh<4<gbDXkET4*BE1sk);6r~>w%5hh<#5N28Kil>)ZzvtyXo)>5RLWhvO
l%cXng8|)mNbk2H}&iY5?qsK2Q6}zwF#|givRL=bC{{bs9#nb=*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/magic_number_division.cpp.853FDB10989ECFC7.idx b/.cache/clangd/index/magic_number_division.cpp.853FDB10989ECFC7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..47987dc360adefd27e40fe1bc27e27fa3ae68617
GIT binary patch
literal 3322
zcmZvd2~<<p7KU%&f(cwm3_}PIAcjZ?Cd^S{FcAzw7y==39T-bUL>VN(nL2<7`iNGZ
zQ$c;L)`G1AARua$`cNtO#9~FCI3fsEMZpQdsTbb5L)Us)i?jB3{(a8bXYYLvs)(?#
zR}cV%Y9dx;XXK}W002xz{tWs0%o<Ywm~ROHtM)Y{+{^6_r;SJr%C3js^yvp9B<`Mx
zoq?364ef^-o^KnR{GsPgQfO22XRVP!;Zn`ybweFx1JfUz)Bwy4UXg>FOnNece`>E%
z>ZTN3JGGwC`{7yTiQY$AzhAo33ku{)?ON9;En{G9+0R#kFE1_q^V7)1rZ1B7qr(i>
zbsznO+6OLEFTKqYJNg~sosZI8dj9O~NL$OPvyaBM1(kp)Il0pdGm~c)t!8baZ`{4Q
zH{pq8T5<9ui_`f(=FR?3puOcA`?n{S_PyR+-`=LU$CreBSrr{^@cuXkZn}K7JYbA<
zu|;maJ;X4@Wz<k4dPO;QkY2y{koA@~_eb9(A53Yj$l6s}_2hKAOXPqtT%qv{ktMyl
zE#B)d7n!Ut5q)#gsqy1b&mVKc4=h)#bzU4xaf*01MR&W;GNaX<PxjhcXI^}G@p9^+
znqi(?QyX%mn6cw+h<Cm~N19R4`^evUR_*p3P+eH*`CXl(HFkOV&c^S*b_@wP*A^ff
z_VDn!Q&`h++_=;<YuQSAwEYG0HM85kUqWA|dP*qr^26m3JwJ<f9a{NYEP3B59;Kmh
z=Tdsb{5)j`JB)G5>RrK=9eIlK(4oS?ik;@)wTok(Y`#T4ykda%o$Xz(;Ffe*v&LQZ
z+_u(fomc6xQ_fW;EsxV~spKUs#(*0ghEF|D=?9DIUB$vXpFDF*wsv<HFZZ(a>dD@k
z)B7&Le$E!3+$;3p`q}CxyZVi+=)q}6yEO%sSK*X70mlAqw|!uqeN9<rqN;qXN!78}
z;(hKv67-fqPh-buO_^VNyxcbs`bNFF^9JQ$gLBl#{jSL6^>f+JG+Wr$ZD-|3+BY~g
z?~88zSh@AaiN!aS@#coXzC!t;<}_hn;fL~j*EPIt9e&H00kzr!^WKWnH&3c}xIWGv
z$}CDu>X)z@+e`Kooh0QhJARip@A-D<Kwa<dw7#iVjMd@xje&-QxcVQC*Ed_d7u`sE
zArnRKHl+03k1%bf3e9z+`?8YbIT<U~df3nC5m~eJc@9OdUyH_m&8W+yF4bmMe*Sh~
z?tRkf#3BFcpiO^l><+JZbcWg&RA}?-$csvi=1RBYKLgw47Z?|3-id2+-%>og_v6_9
z?5=H2+@yss&K+8lcBK0EDGwG@-Igp_)7W^1{nKlO;(k}mq%t4LK&`&^4(nRp#}(^Z
zlpTI;XN%Isq@GLigvA!aw!E0BFM|yxpl(v|m(l9``$pVr3wJlA4<BqRbF2JB-92h+
zV_2R2RRQw-`uB`*^Y69nfq+B+fT=X9=fr?06F+_B=hAe$uD1{ni^u|Nfycyv%B9yl
z8^U?JE<ivEB8x#Wd1652>jK}e67P*2hJYkQW&>>c#DL2B46R|$Bmgkgubmmx(l51v
z;K@Fdr(+?(bUxh|C$<gS5!(@F^Vz-xlCr~a@)mjrVv)l>JRfhIq%<i53mcZvWCV)w
zh!NtNlCSi|DV85gAQeqTSQk%=Cy<(^CQyn?3K!oc)gjdtr)-yOf|lcugJ~8=hptL0
zKKGkJFqKdB!;?Vc(+DKs34HLY33);SX~-HHdX!Z|bu-Uw**Ovf3jhHbEr8WvRB9u~
zA{7MTBvuj|^+t%3G!>^M?n`(C;x7@P<Z#t@O<VQs8zvB($IA0Sry*F+(qrT#dbzTa
zyZu|lDc>a@n}*;*60?vr1zqI&r$-}Y#X~9LMg^!AXp9i-2Do9wZ?USeX>E-_oSa>q
zg?L^>t|B3t3E?F##~=G&lv_iv4c~^4E<4Z;!j~{W27#DlCJoovD)vIWpEw`~!ifuV
z3FJZcpy8S?&zDdFf3iOf4GMxKpaeoM3&Fu)Fo8nJp;)mXID#C35gL<;#os<m2wp;7
zf>liHGTgq@JZ;zngpDM+kz|XW1H$DbdO6P0!LAp_s}HO|8X>_x!3A{-`=$6r)^qK@
zkm-e13vJNpxgXZW?%L3D6v;Njjln}hfncs1mxo?D`tt8_U4o(COhGutItJ@t5LQ|%
zvEqVotaU6#u{i?g<nWXY$i8+$yFh$jj*t_G?yHWWdpy}8F+)B9DMd;}V}f8AMMfYM
zMMa=^N<4wo6g5WbyUB{16;V<oMRwEd1n8L{%$dd!pl4?44J(xIt7N7S90SH!;b$sA
zC4pkWSd5mzid2C?TL&U{VcU74`yv?=Fo-u}b3D;gko*a_*geq(K?*$nzT$mfKrX-v
zTzD)hd*1%eR^*fvaY_UpSnBlDFr2c**#s?TdJd+Q=%~TcPYMvJuolu{SwWYQEdKUz
z?`VE_T%-wcCV{ZlTkC^*eV-KxD&jUjnrO2bMv@p6#@lR0InK3-|MfdqMx`SKHj=E2
zq{*lwgg9E2<J`enTUt0Po;u+uGm_>I9cAS>clQ4~_l)$$0P<AH?d0~T8wf|*M`Eun
z2q!a>9dS(;q6<X{O+J$5@qa{knv=*$*phDh=Fe-SZflU9OAsdbp=*)XO`JCiC%srt
zpmo0M0&#79;QAnx`X9{MR^0M<vl$3WnNq@wDr3s9cMybCOcftp0>bgkcmk=JYA;k9
zaN1K_e`LJ^=}u>wGYfAu5lzHG+X#YnG#v|F@^o6~56t~llaR9OTy+8H^s|gvWt^S9
zVdEdRm=TKa#1(Tx(M$pQjI6ZrJQ?)qhN5v}(fIJcp9(bqK#iLTxH%6u<8X5UZpLFK
z34Y9)^Sl;ttF>6TXu<qARlIs$0<aL%f%&+Zo~_BsL@-JhV+xv?Pa;z+AuGD|WExdr
z6E=ruW1+BR+SxlgiDw4SlFDX>_%H+#?hHQr|IIqETwJGc+^2eYdQB7h`uPV=4{{SZ
Udk4rF4&jl>CUcdc5u|9~FUB<dPXGV_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/masking_specialization.hpp.2AE80C7347AB00C4.idx b/.cache/clangd/index/masking_specialization.hpp.2AE80C7347AB00C4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a7dd984d152df06aacaae21d661711e74eb0d2b
GIT binary patch
literal 2250
zcmY*Z3s4hR6n$BiPr`-`BrKn`B&lGGi6R(GfMAWmgrEfh2aAkV#3(BGAq+(+v|3R|
z5XwMhwBj&hOHonKQ7aX-*!sn9)ml3WwmSGh#1F+zN2M=YFWF`$XXo7e-n;kRd)`g8
zSu7{L5Hc^Y73J9rvPguGFaA3U@)x%Y5PIx}P|^0yY15XoN=eV{xoc{h4t%TodDeoA
zJxAgO=6(0LanZW0_VO+6WV3wli|*K?N$<Dy>e`y;^%PUxIp^;GbGbs(pW3L|8hmee
z<mZvUZ0n9>`yW0S(1q0ujci=d)p<V~sWJms&~x7Un%J-j#<6E#U+h_{?N5)iIv&QJ
zJlIjGc^T1L{CH`@iHw{MdwYxTv({s;>aP_Q>zMPW*D3T&Le-lksh|D2sBYHO7rD<h
z<i@J23A1i>o^qTYEjjR3{c7##(mBoLwNa&IGdBE`n_hd$GHPf1$un(BRvv0!-BDY%
z+b)Yvthu(zJiGVmMXOeNW5te+FnQUTY7g_Js3#S~rS8R5^GoVhM(m>ZcTs)jCn-&5
zfB$nzbawOL*wz@^=9H%&JuBWgC1!EexzNT>+1KM+auhMG7Yz1T@${%CbCc7{D(iDI
ziu3&+T^Ky<vu5U(;g@FmWQMAu<*GElk3-a6Pb*(u+hH4&1?&mlwzQ#xt&`rlX}NJ-
zkh7^ze1&ZLpd@11wC63eyGmvkPcM%d+O+A)7wNi$y(x81zv;W{aW1h?xcaY=ci5%X
zHJ-NZqOHd}zrP<I!sK0dtje1oMBrxjPYkK<=-f{u${#TlNd<C%aNC_frbu+AWU$oc
zj(mI(Bcy~%4v>)Z3K&5cIfuFRwQxv61WFOZs@QN2V1MFm!-QBm?-Pdb@I#C~P_O0y
z+%rn{Mh9%zv0g$5@ouG;GJpecr&NfueXgt?q)-5kPD|+s4v>%#pkN7(KL5S-lgrVv
zyAr|^A8%5bCUOA#3wv*RCui+4c@W-se~@R;2o7A?ugxq=QtIwCdl2q;L6m1yAO}cD
zy0Hl102`_s4mbYzx7>#a!U2-~ljR&hfGw)<!Qk+XZaxH!-APoEF9+b>Q@H+3L!Kg0
zLU>`fQDxL|0PYs)azT8?(E<-bjH9bORiPZX((6jq^%(Y5wBq8*5TjP9!#RNc4=U|`
zkLphs;OgPyt)x}T0k{V)Kab56)zx4(-mN9IQVziV#-JHEzADU(Mk3sKZ_-=B0k~1V
zeW9bt`Q+azgb@8Fk9Ufo7@0z*gAR!gHJupkzT%h{#jqqxaq7T8F$z+_V`@^(V;WM!
zV|r50V@A@*V<ytXV;0gf%s`-|h!dtQw6MJT2XiUK&<6TL$c|#9LTQML$%L|D1_Blk
zoNz%hAY5#AsuqL-f)$83p~7l78k?Oeg)Gv9Q;+4@C6xl44n_OHDo`z6?UXE>4ToW~
zQ>t(XxN-fR0tGYmzq~-FB*hp43`)qJV$8HT2;xzUjkX1WYDAop;q>xEw{?A0Lk54`
zz(_V4s)Ct`O>i+~pi&Mspcth}IgZCBD&u%8QJLam>PU4osF`Baaq4&%)5^3gXq95L
za_uM{8>5Zju^6q%#q?wJ5w43Gry8fjd!rbW-ZaT|ZPOH!&Bd%TD+@)S7^~bmipR!S
zBX}&vYI3n8GRYerJbc}7y+sDvmBu95q##H`#0>+eFj#RPos$5}xJ4t);~_g330MQp
zc>v8mV)?3Q-{c3sxsi-CN@@lnk&G-(mgZvqGXF?mM;6|_5HC9ahx0{>*#O>)6f1a4
zEmrfGMy%m6y;#p<MzN8{Okxv{S;UrM28|^MZdk#MNM;;8E&#d*RmJyo&MhbgxUtze
uuV62XDqK2lP7RMBsY0rS3+LN&vz)hd<Yzm)@t0t3ZudOcR6k-TLjMA>z?g&p

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/math.hpp.234B356D49034450.idx b/.cache/clangd/index/math.hpp.234B356D49034450.idx
new file mode 100755
index 0000000000000000000000000000000000000000..65b9a6a97467b6c43cd2cef7f50a6284fd57bd98
GIT binary patch
literal 5434
zcmb7I4^)&@7Jqjfz>oQ6h8Y||gb5jD5F9=M1qr5*KT4p8iB_04i#4_hEXi0gZH2l5
z6{Zzoj?(6zvbN1XnILjRsg|&ivd{<%Gz&vRas3y2`nT_!kIVPY@ti%&oWp(dd+*JC
z_uY5zeZR+otl6{U41}aVlvP&jSXH1Tgy`s>b5-dh*MX3Q-h`CZ?S5rW$<7$RwjXy6
z&0n;X^Lp&(%BaiD1J9+Grfrz8){t^v(c-V8&OBXQ_3)<N7sJ|r?=Vff`Rkfj-`d;t
z!xK|J|N3x9-j_Ykv;_5pUA=z8X?y+4!QjTIeeKZ~GOLQ_oj>@*-lB%vBFav$tGs^x
z(Zu!Rch$ZZZu)%ZN6K}-|Lgp__b*bm@9qBbs~dlwuz&rc$IsN?=-;=ZaAj_LUdFEJ
zg*LC`x{APqv-iYlH$3y##YI(>J+8~~VH1mmv)^u4ZCzmxIM7nPws%vHvE)Mg>LUBj
z9r~<<JKTrf{G&VAy*7QI=$d-Vs;mXaRyW);W=`*#r8VK!mrt*#Z%v6dO!=_3Rh8iS
zAf;^0?dih4eAn3N%e}%cHU3b(Qhn(3x=ZOrQ>=Tv+T!=L{_XXGh2K7{tc_pSw0%{*
zd1dgui<d2#uz3DQDd$cfT|CcVX#UQ!eDPPOPrTpq>+>(9>71<@_y6hmp36_Xeeubv
zjNN@}zq#}AuECJ9L(bkSR<*&kIsKi1&;IjU^kCfruhfCMleuMgxeC0N1idl-HS5y1
zPd-27j&FKyZ1aKz|DIF7ulAMvo|bK)Uk6@`uIw;HS9~>X++8npR85-gH#cqO_eVb)
z9}#rr)qzc#*{!SnmFu(DY}0HufN%2FGbLHC9{TO%ww{AkO?TG)+MID#P!0LKb20j=
z(^b4|t`{L=cJ2tc=DL~2li1sc;0B8uB8MBojq{D076Jl{J6Cz`;t6-7LXm8waYH`>
z@5^flCxO6eZ$w`%N@?^{q)#G3VtAs}Bgj~3*w~fdJa}~fCp<}{OSFTj9W28Q;dDME
zn9wDr+^EV{EP1h>>If%7TzH(-BTz>SACrIZ^_a<YIFtyT5YY)!hZ~%d5IQt-*(J*%
zo<!1RY9O=*%)<@g^kd0HaT(kB7iJZ#{$w>z?xd5-E#;9OK{)jo;@RrVwf^4yH+Mwx
z#6~ALERINzAe=uZO_;T5QmHZbk05GE7!kT5qMKSGHE2opsS{7G$!z<{TM<unB={r*
zdj#Qv0;or*6?d08_~%x>{FJvMgHAB`7*rl1PguCwd^J6Ca=W)8l}-rt3DtSTu*0c_
z`yMx(nOX2sv9u3osM92ff^Zt8w9%zo1}<g#-PpmC=~P3eDKo(%2<PW5IsAeBg;(aa
zfBc<vDy?8@rKch_;-uR>r&91>&Gm!VW>nD4nuuTxw^}^{H(F42V%J}iuRJYT;R<z`
z1W^!9>m^z7;MEC9j%|L+c#=k~=mAp?OdoCt7if?yDQy16tpCZ};H{WRCr$L3$a{qB
zjbdKECE#q#&r*yY0n-s`s?^8<^pyA1u=10YhrJa^R7)55bb)HPku_DDQ|E6y{m=&K
zbPG*|7C{uUmRA&7zi;3D{9bRxEUHE8qg8tZ;U<iij&Ad+u8)&my7-c`<!UfhQ(L4)
z*0QzkT-u^1JNv2ICKEvyuCsasx~(od@`Ly1H8#^>3lU5arl}r*!*6bjeQWQB`9<D}
z6lz0^PmI<h<X!BlpSyMGd<8GrPyv<-nzK?P$LT+hwe+7FeAX<vz7I@&)b&zBI2A~)
z-=i<A&d$iVCV932Obv85QbRaiAx%2)*s-l;b6aY7l0fZAHKoq*2sxBK9Xfn+C}qGZ
zdDU*QM+%~lld9{OWnWQk?<u-rGZ8|<L#!Tw8}90R_MTmxk-(D~bb%aG&g~u{FVH)A
zecthfk&j3V<cH>)1X0Lw)>}Ha&GO@<S>B2}sD>z?D6L1xE<2;kn|5!va}Tv3f(QZO
z0alMd3ogFzEP8mI>%6xjivIo!@c9Dx;YQYR{X?~7WzgV@k`;NOc_u*=a?X`HmOD$;
za!uuYhbt-8PhNNpw6B4Fr05FULA#yByFt5~#qWXkJr*}YKqCZ=9GlPrsus|UBnZX&
z7BI-T4Ge7%h`Fwgd3MW++SqSYykHBo8L@8jLe{vfppl9p*rV(*BWEf&;vLB&RZ%F9
zDvuc{qBQ8@D<F9!9%<-AG(fa$UIWB6K%$JBK-B~~8TUbaA0&^2s&D~R7eI$tUH|sc
z$zG2(8+k#e)bY}7LI^JeC<7Q2q6}e>Noit`RY{|A_`?fv$~XokDiax$N>YdKD{+NP
zlF6VPlEa`plE<KYlFuMJu`|d)91JQXg$!~MCxcwX#h`Lh&Y%h)6~N$XAk{3c0aC+a
zHxM_A8-O&h_y~|AEN%tT%Hnn)?JVvDWhc$e5ifRwvYW*{Kzi8ueL(tTtPMPw7Sq{V
zI7TT1WCzT_L{thkz0DsZP$^_-va}e1*b52|kHG5`dBhmsVfG-WRR^d#K!X!_p#$_C
zU_cy}a^TR_#1p<srJ$Ron~l|3DVTL;Jr-l7kgLwsU@1}xZJ=obEoKhV6KaykO|V2G
zEuq~KC&5C3WCVPNDG1t%xuf*;5D)C;g^i%v2%3@Wg<}0iFvz$TRJEW%y_k?xjmS}C
zH6^MN&B!e$l;}$gGO7h_E$C52n(xj+W2*KrD5>o^osf&0L54#&CGi@(TJa$0rbJSM
ziz1t$-6EC2ZsahuTO=^pjm(8wE;1JE9<2i~RusM-UL&i?)#irbDdOqTh>@6$Jner+
z4Duf80+ICK^@4nd+Ai`OS=;Ci4_OiQj6{i05i%h<AkrZ0#&S-Vl~@lB@T08%c#u@T
zM37Js#t2<O<U`oa3;%yORcDV`cWL7vi5gx=nv^sdbC?(GM!Q+QP3-Y@n~WSrhZ#ee
z7aUQJSQ(WY%gvZAyigurZj(_37%RXm-#Zl$RROUwc7xFkX8D?RLzEk0Ssd>Mn~cwb
z<}4WSTzH`$H2q+}d#GkZQ}ST!13tH83bIA&PmwC3+c0UVHzj6>1EURctW-1%MG}e2
zVacN1B6-9_`+M3MoAK^BYHDUb1owjx?cecC^~U3x8MC~ULP&PV9Asm(1dq1`;6c$6
ztk3etv=UXA`XL-uh&Z_)Y>2S{(w&P06*q_{KzA<jQ5+cU&T$>8U*fE&2)P4YLE@=6
zFxp%|R!dh9DJ`mLTz>X=hVQ4}iJU{N6W?o|59J){00xDqLl|UIn;2wOTNxCmj$=@w
zI*~!CT&lmkLME5Vpd2oTL3vyrgYvn22H80~gB+ZLL4{l)gPfd`K`ze4pmMI9K^4GN
z0E4T6t7dTxa5XG;1LtOO18@y2J_6hk7PkV|%Hnq5+F9HQ>P}!h)eY)y7WV+x!_Mym
zu205Inz1K7zVC;eTQUJF4~-u2g@T7ZnmX5B`j5MR`M%hBi*8ISnw8=U2WvUrMKsRD
zmklo2(|ch1xyTmlEj0z-I@Cn*jf3+?V<KKxbV>0kges7FQ3J&X5MJ9zS?Fx(;e+!p
z{jy{zXKUpQ1+D1glR~giP&73u0TDimsJY@}2!A^ob8l8=HHLq>{|2RJHnYuwnw3Jj
zI$eWlO3TxvtT4GewSj*djKd4Q)KOF*CX%!CL1(7;#VQKvGZnwk|5~L7t2$3bei7z1
TC3<R%l}wR`qUEnK^4I?YosEJS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/math_v2.hpp.562B4AF9F5A86873.idx b/.cache/clangd/index/math_v2.hpp.562B4AF9F5A86873.idx
new file mode 100755
index 0000000000000000000000000000000000000000..badf43a0d894004a74d31eeb6c63517431933a0d
GIT binary patch
literal 5860
zcmeI0c~}%j7Qp+Jle7#?502v;%^V!V?UA^C;xHZ|8ju(;Cho`v@xbV60&X&5BpO*!
zgBNU2^P#McBtf$xn-FwaOh6ZpWTJ^;^fQF0Q3&9L`(-_d5_GqEnwqN3-}~pz_ff^K
z>(#5*b-eegUr0?(zNqK8i7%(-=g!ZYtKv9r5dK?|=U$W!9H$$=arryzHm6NJ_JmOM
z@I|*#yD;RNlk;y3O24?Q=HjT_4>~s;w^T<xa>i!OxFWn@K9s)GTRZFE?5sCB3q2F8
zSE3Sfmd5VCapt9+Bc#hM$;03N=wFkk>^t+f^Yt^G%i{_x-B-sZl=lQRym8|C#n7+f
zoz<otoh4(gKeTIBEnWTniUA+Ib!ytxg-PbG&W2a*U%Mgjw)u-soD1tVM_!%zX8fnm
zemCcAw6(9Hcu3-?iDx$i{$b73?Xz7eCeNsu&bhtG?$<lUEvj5}B}nTUrFyXU>H8D^
zdH>Fh>3vV$vyD#-k8it_n%rC1Q<y)!ad7sO)*bdm4e|MRcjZ^Kul_84*PZqU>BSR|
zZM*Ztr{WCnTz5#G`&!d|@rxI}?MW}~8D9HsO7gQKI|t_)r@T3|Xxhl!HJe?DTY{?j
zs58b}i9P#g9hh_qts0xSyMFRh*K}{8VLQdGOWw@QzKv&8STk-%-!9iOfr}c#NiM4^
z+*fb`!eg9G2NDPszg@TNkGpdOWWWNI$W)2U{S`M5FAx4{{&aNv^yO2d1uhtiSZtPX
zU%>?gJ<f${v5xqC1=lwmauo|)I2Lf29bvwL3s4W>1`WmnL!UamH8(f*B}IeXY7dj-
zf?6>`*qV0w)#^tACtKmxx`+4*E?^*V8V+li++(<WcuaYSfTFR$0{sHXR}d4Y<=r?k
z?4IScz(wHZwV+)KL;5SKBlGj|LuUo#suUeLhMW*dE{KkqDZdO{X@9a&nTY~Jfh5TV
zX`a<<zkJ*=?>PZQ;vp)*Pzh3hMMM_QU%h<w(tU3WoNVMA?Hs+Ypcc4Kj?16je$S%J
zgWjqSljMR65P2?4p2?-}TK12Q|M`7oCR5BQVUk?<CdxOHD>735I^yeXTNDjCt1e8E
z3*S`h<@NmIGo`ydFE4shKt^ms9T@7s++Pth*G+l;!KrIaD;1F%(APkSq*O$tSo^VU
zPu2#rfaLky2>Oj+>aU2%JF5yaTQ#vG6_HXfl;ZhNDk4&5ofmX`!|PX-&6jD(43p%7
z3kXzkR@vCH^uNEG`tA9p%IhYXlfopq@a-($>v{&ny1CCc)+&QEAX5Y0ok~Rp*>&^C
znFHpNOO!#ftXbibT+kWW+g*45F{gQkvab31e3K*>r2p}P@Am}!c+Mtp@`P1^wh9dW
z6<xSH)3oD1zI(kyG1O^xhDmbaTR5Mg_kJ^EhCcWUk0Ov@P6(6a!uP5^frl5y-I{)2
zS&*`DT8lPRk_%bD?wgB8dp|#2oGl0fC-1K^(3asXtW;b;08);@zipoMwA`K*$|m%H
z-h<agsmS}d=Py_ijUDR%_X?-uSl5Jsf@CpRto}nRSt2ZPlx!+nAf=@6q*(u=NM0~`
zA(XayA<_#&DW3q<31E76gS{K#{r!Iznpk9>(}o180)i_*=QmzzMz&@Y;eX?PABEID
z-=u3VYQJ%&{GeJYM4^SqPMCNqQ1|;M(hyX)_a8Bjy>`efNLo>gLyYnhB%P>ZNH6Lc
zvWOOj?4q5a1Tldjhv;C)DLNTS5|bE85mOj)i7tjR#Y~2>#4Lv9@N+cO>IM7)hI05E
zhTOcHp?p4{p#r{up|!xT1%@{Qzmerq;7eI91HO!95AYtAD}k?MxeE9ymTQ2oVYv?Y
zI?CcY^LKrbi)Vr$b%3J-#*>giHWOZZIk1o=RpY%E<^})S8$v&7`p}v9D(=4rg7?5+
z;*BFRhSz;&Szm!076&iZ5f@L?W3_TPqG~(9+5r)SNnFBWay%lB)lv(Dwoow=ov@f3
zny8ou={(3l`03u!Z*^?Fkfau*C{>g|lw-0*MN^7V#V~EDs#L-6^`%r>YBZ%*P_+Uh
z+X}W;U^HC-)djY_4^(}?XzK%8ANci=5Q^VIj-SM8wX_)-HzPCQ)S8F$!lo{nk7u+I
z5*i_qc&@bg;?Z}HLJMvvM?vLi5SeTee6eCB@)A{QsSfGt=mce`f;xA8x&0dU2n&N3
z4j~2!QV9wxK~W@pzEo8fx_n3R4|pI)qGJMCOEMfDMh@F_kZs7c4Os}2Fg`%-3*Y3i
zTB<>&8v2@oREq*@kx0a6@qA6}cP~G}w*AUGN#{}Ed88rUi|jnCUr7S!2(8G}%8qaW
z1ztcR=`b(C9WdLKbsxW0JB(?EaU^EUPfI#fSo>i%_SAZV-r`RZBz?F(hLTZbR8w+>
zJ7fG&NooR96WC~L6WE&|j`C4Z9R;R$7udTXj`YvjnjJl<?b&5|LCV%-hmdu^G(0<z
zQl2I+gd7Dy${UvFpi~5!A_yTnL6C}ISP?iVZwJkG2=RM-sT}m>5YF<*au`p!8Z^}q
z;*ZTzHH@eRC*=~PDM5^-B`CfGB~V){kfs7L)>fdn3N)J9TaPsLh_SaG#nq$H)ZS*K
zX-15_%_ycBjiUB7qe;!^dCF}_(}oy(+fZB^8cpmacHylnXN7dfY(lzC$Vm9LgDa2g
ze~8(QcgXVK<vNnMkY&WZ<!q5itMRIbHc>A{{00`2b4DuG3)WsLMlwh&CTEUR>@b8L
zreZ5&-x%h++Vp2^i!M<&j`S*c&-3or-m;(Y<6;Od1|7*1N#2Q%g_3kq#W#`ZCbAM1
zq+2NP778XENH&V?QF2fssFpgBv6CL`bw530b~G1Dc$eA@b_;c=_HcU)r8rfbnxW9R
zC`u{eDKTW41j%J^S^W1S$ra&>rIcaFVA@*1)B-lDrv>aS5J&kEs4f9B>Lm!h1kAX-
zVDE)EI&vX06(SoQtq?^PqIk+{kYNpC#;-)$N@OGnkbv_aOC^e6IkFP>Cs#R~A!{3u
z(VskHwlts!mLnTbJmC#3qmsK;Ti4(<-2%Z|U@%GT$=c$jQLX@VX^tS{5#;y5CD-=9
z{$kS2cmD4{$UPJXk^iLwAuEDUmvY6TE8#^ZFB^1DpuiJIBpt{V3@@R4^`hO1kgW*$
z6CskbV=?&xM#Z|p+D)Cpj#oM=|FZN)4Gxn39kK;!D+=0*v?QdHt%Y-S<+esNlG_>{
zNWP_!$7*Rk3SCcg{r3~5PY&Mo%09d$>mj%vh7fN-d?M~IU*c&0DrBsp8}gH$#@mq}
zPM+{MPQLU}R(ew73~|gkPSvMc>8Wf5eJe1sm%wm|J@tK{?*qU5!39^oo}@n6W|i|K
zu@Q_kgFg)7VV0XfY+~69qL<|s5L;Mo2eF;y4iGz7?gFuk<!%tWS?&d~m*qkv79z&J
zA|w{EydH_`S>A;BO-TMW=lcmbLF6~HybbZ&SoR>^!*V6!D_O2Wd=<+zh_7L}4)Jv?
zHz2-&<s*nc!m=0fUY1WG{uImS5Py#4R>Zfmd>QeVS?)x9C(Abxf0OdD_s)JF^0;>?
z-t=wA)`lWUR2QVHDDWy8L?Re@j`ybW8G-K3%Sd+_`FHD=w|{@YTlCGRSYs^&*MgpC
z`1J@U$Ifp3`J=Q9d<+hQ<1maRVYt~DJ!<643&prWJ4WjyDf+d}S)UBc+?#?GXV^1h
zXh6wyWR4@Q;Kuf_yQad@Ufg#T1h0a@Bn2kvAhuurxIwM&0&5p_0wit4V)7>rD%K9x
zb}IHD@R0o-Vai}UOoAg}Jn8kzh_^Rwf8x0s{5Z>*HG%j8@&;I`{8@x5Z3J5*`27S)
zxUiV=X@m-%Lb_A*szI`KEU0`Qp@Qd-@f`IOoO^!GTzQ`_anD=gl>afyh4QnIe>l!W
X|CLsz-I?Gj{WwEqF@>195Zdw|YCkPj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/matrix_padder.hpp.0539595BAF18EE86.idx b/.cache/clangd/index/matrix_padder.hpp.0539595BAF18EE86.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0c57d494039b8318c6ad1af044e1ee042d585f46
GIT binary patch
literal 9614
zcmcgyiC<Jz`=94=1f`1%F`x~LiUKoS6+{rju$f_-L0JcIK}`$m2T2+vn)#NM0)l{M
zl#=v<LGDW?rmq2NW-b`GFPJTeS)v7LmYV*~Fg%^}7wG5XJ%0N>bCzd6_m*sqj?VW{
zC?Yb<3$kauHdCokD4f~f+}Cn46{t|O*eetZ)_$Ej?d8@$_1T1z5xU{4Gc(VJFQ{$V
z8u-pj>vuG`f89_Ll`-Cwb8gG+#)Pi<Q@8&zC?mjlckQY-wv~B2RayB=^MgqBr626I
zyAtPDxh$%<bm!}@{F?g@KRV{Y;^$`tl_Zo5e}4MJ@?TFiy4(#u8}Re>+5cNpaQAlS
z-0yz!n6__aL&V3A-d_=T?BblP)!MmFgygQ-y5yZ}i@$m!aof#zdsk$JjnL%&cj1Tu
zxf8>`7?2bC=%t#&c{=4cqu1;hnR#PF=H+t-_FUXNc-xf$KmGB2!GgBE;m`asdhsWj
zV^-%Ed^YgF#VN1-vT)B6yT7-y-?_JadDbwm&7IlbtIzbGt&BJLn>N7J!jB_YwJykg
z``W;hi}J7K<u032Hel&cZ+Pm{mma<1lhFBOpq#v*$kELqW>$l8-hV7NzTbZ9KO2%d
zHa<0b(w>v+;&)%q`=EZ_hl`&-)yK8q)gP+1e-3UT7hb;J`fb1v`_u*3KiD*NPi<pX
z*6?+;&f3ZM8#b<<WJ)~S+IC+zdlEJ@7F-E0nSb=r<g6jBo4ulD-zgiN_FvU|%REe<
zX2yoJs&sEohT+=pvpzL@=jei`U2>W9uK!|}w4q;4czfu}r}oswO+PnB>U(GKsxDL7
zt9F;i{Wa5bpY~|%2Q~4lw#>f$*_FNVr9V_mj2UqJSIOz>#7j<oqtgA3Ma^`n$^Pxx
zC38=!a$nDWzOFZOh+gxuuDlbsxF4#074ms-J`b9nLD7fN)<3Z+L9Ih#<6Uu}k3wDs
znq}bCGmr>uscJ6-`EwPzvssRU@+i3V3?za(g4^y7KG4e*J=rYFL9?9AVjD<AsUA_X
ztzy&Ie%PPQasiw#u)Wv@5|O4YELb$FxpzNwX0vPp=S@t+Hjs!@<ov@qFAYBEicWqC
z`Df7l3?4lLi9ET?eWPWAw1jPUs6w8poaxs6Alk+7(V+Q%dJj_JQ*0jxKzV?j#x{_M
z+bgY)>_;VjVu#WYg<KEz^`Pz<6n*>jc2Oumh9PO@ij%4@Ut{yI&Atc9_t-qPf#&fs
zj!LXK(3oh4Pq2BGfc+A1?imz)?UlB9hPA6tEna^27gy}ZPIL`4*Vu_{1I-igpi@&k
z$L!;ZDkgFrG}oDkZ6J~GH3`2DOSv1&cEdzEK-s}|V;e|Bzi`yp;O}l%GXt;<Zv^L!
zOuM#$L?-`U`RZHxFPAY9Hp^e={1<9^1`^4zv(KAQ>gl3F52n78pghSIvJE8i(g3g8
zXJ78Onu)N5)`4;z6R`~>l2!hH;rWVTkL=KmX{HVA+n9)L;2M5;dEfQhUm51a+<`51
z5u7iw4ci8qXO0K{oj5_!TZI~C>nvqf|LzCTa~&uDQjqXt5%Yg`p1bILmz~EpkjOm8
zdz&7uP*u93BNJ%{O*_-AZ6J}d3)j|WA8e~)4#h-1Lgh#3)-#aEd)f7mW}TcCZilYy
zJUhXDC%E(sR!jHo_L$XI21Ko&Gk-tR3)^h2GS{v9L332Tvog9UWb!!{iHwfYM0s^T
zh^~GmHwh=6T`MUDvn8HXJn7Q?Q1nGmcv~HE-I80w|7Zd~rY^S4ztQ<`c39g$b8Ikv
zI;roQ2|HZTi&-WgH2Ew{Z3BsHAM>#Er`!EXnF!m3UZeNweh^*vAZmWyf`6WM#X*ct
z&?I<uKZxFUFXs2iZ#`v}2N)fziS_D!5Zz!3+ug_Qa7{lP!01KJi!|L2qQ5z@Ix4pL
z@L}e}jGm;N<ktNl`b_SU3%;(NUodO3-ERQr4a^^G1BsmLJLP;_(*8qCgw61}^Xr=K
z2hpwR1D(?M{^`mrz*KY$mB*MRZ3BsX_x=l)mT9KWwnGP|=5YIP*Y1a+uU)SvyRG<J
zm@&A2*YUSljIbpZIxp08Kj`;=W|pmb`oo{bF)J|o4d*vB-4CMgc^|m&?!k&;=1c4h
z6`-tOj&B=CM3FP=#oSm=mYRFAcP~jk1iC{IM5#-X&w%a>1W_zH^yGBPTa?p#IrL<5
zN<@-e3Fb<Oqg_aHEhN-}g(8+RJvj8_cbbv%JDbs#)%lF&kWda5%7)ez5-Pz$D^S*F
zXJ`Ul69mx`)-wd3fpC5XTP>g!?e>KXdM0XHy@MnNKN0NvulLLMz~vr{pnNCE@1o1Q
zDB$JjvRuR)(Pbl!pct0qo#?Vt#QV@?pNKD@%LN=ko+8OVqRWpW{uNz*<=A0P^ts~|
z#|o8_9HWgHNii(R6SNaXQZz_%G3bgRi0mTCrJyT?Akr$;J*K2`2rh?kQJcY)5Kh{3
z=&6RJeaFDUimdXHt#*<u_mTT?B?ZI=By!Xa1KYueGM&V#%fM^kLxo3T^>g6Q;6tTV
zVx?u^b?~98<5;($QQLcNeLqQdRyb4JppfK=%862U3&9?eJVQA{K$*%+0p%!j1hf*A
zD?z|(K)FW5)u5~v@qSS57x58L9uaX1C|g8)1C%#JT!G386xKV6%A*|9v6;`?+6k&n
z&iAT<{{DIHE@qX8)QIOOF-o#2)$|-Ubh0)%h_WKJ9&9da-$A*A;+FkqTcbfv&N`EC
zzlq;>4<*^4H3ae{3P4u?BPm^zIQzY=88HvdD_;4(h&KnE|E$l0^<M<pE9F5^gQAaA
z-OtQgkT+B*cY?7KA}C0dau*o8L|lc&DvY3@P|DS4tQK)I8k;eKf<P&^ps|Hx?`L=I
z?HbeK@R*mSs9<Wa^;r}KG@?UKcSZVP(GH$F!H>!_xxCV$r^6y!OY(0J@*9Mb#w7V;
z4EY#CNkfu+0z*z<C~0JP#)Gk*_@9HQl-)FL9u#T}J>xcx5Cst1udP|5)yDWdubAod
z+&rc~??K*!x!G=m$8GSWO(^9%;BiO9pP<Jl=t-MW$~EXwBjQu&aSA<YlS=tCdYtC?
z=~4N~SH9YO$6k{C!u=-EvXZ<M4NGw<Ez7bKx|iZ;j(4G97fz*RCAl8m>v1&4ZD?r2
zskE#lcc6O*jwVd=Fg4g3Qc?poq%6QHL01Vul;*5uDYzEGsVq74w5eq0h?@2LM~+gD
z0!dEPCI(XU#hhL{_M<6LxjvGd?v<{kZbFhXyfU;Dnv$G7Fk44`qa^2f<!Pxgv!-9K
za?p~aS@-2t30iSqUbUd5eOc?W*CEhSd|PXrR}*N(8s~Kev{d6H`91{Pho?y+lKc+@
z{3GIr5b#jMOHs8HHC#R)Rrw+=KvjW=OHoxS;xbf~iFgI7R)}~L25iEoxh1z?z!nj2
z#el6M-i4}NsNwQ;sHzk3ZdC0SaXqT)Mcja@1`#)+s!_xjG2kLTO`apktr*ZM;>#Ft
zS;TFqYC{c|zly4>B5p@jyNEkb)gj`WsJbcQTd2ClF<pmxOL7&4R-v9Byb2?$FplFE
z3~fO@-$)BawqP7#hn`CmdH%GDj+-&6JL^~^4jMctm`7qKcy>Y{w|5tKc8RzOJ*zN~
zFI$bC)go?2&t?pyWhIukJzF>?KV*?;yRlJWc`?m%=M`0}6RShf+UP)D+=@U~1S6?$
zA-`w8w_WEbAn9C;uwCLP0n-(Zh4>)|J_O-ns024bIE9Ks&&7^*H2dIs-A9#o(jRk*
zq`bfmV!cdK0jGN;bA(O^?F2oyP$xuoLLA4<7}_j4LNi7-i;iHseUiecCu0T4cIPBQ
zl5auCEwLa~Vo0S}keV>0i5H~2@sUmPq|u4PSSO)2`0`G|pf@CQWK<gk&S)@NI2WOg
z5V#0Kguq3rBLyzf7@5L<7p;!=<sK8Qk51+&Mjaz?F@_im=S*soz?lptfitVk0%taw
zQ~2-3tH=BD&U(Cld@@H9)Dr}5f?<M%b8+f8fr~T530%B7Uf|-5@hSXwiRwgODlx3!
z=o6DUN>V2YT#_Nl!ntI1vcM%9lT$cnQCkGgVzdZc0k{@`AN78cTmXgwuy9-mu7x6B
z2*yH4;d~Le7J(ndtRxqKp$IG-7lUiD$QOgL7*aT20<I<C_wU_AE&)RcSU4^P*HV!$
z1!F0s5TE~!@A37w?cQRkBtjc8lDkP_O5!w9KyhrZB{$0+jbQ~`-B0aG4HM-x7Jat2
zEEEkZPhJ20?R{@-d(7E@+D5h?+bbPPYpX`z8&T+U@dVqy?@-@=^Y@3}Jve{&!Qt)f
z>?W;AxC|y~lSb0nso7wYT3-QCidj45#G2%X7hh$+jnl>nSI4B(r0Miq3IcWwx4s{u
z-?9$JoDnB<IEmRQl*uap-oErg|6)%^w$*UoR9<TCL)d*t=85Y8ggp@PQgmO6qc~rR
z<4Q4&<4qX036sfWlDrwiHj8){y6?hKoUg}m^_a%-MGU)$$-Gou!mvvsZbSDr9L4z#
z9M^$qgvn)|aOi#uMuw5(DNa+odFS_n(+l2QFSDIydy`%yd7jfeZ!(4?zYk9DgE#N|
z)`HVo@aCOg4LH?+H&sAM{tBGF0&iZ<j)T*2@TPJm$)~~TG<Z|RUH5{1>i&$kAF(=U
z)|!LJj*=X&jSu3f#-g?SdzW1IW^q|lztE`1?DO;%k(IhG2<;L}T^B@liKVU@L#xG}
zs2U@yc~9if^X`#im=YB8UE3Q+Y96dzx49hRxbADgTnlle>n#iR^-j!Q7abtU0RaJ_
zWO>%)>(fo#)EW8=6E}6XA=}JNou|(;aZ{Irz8p+^xk}Jif{B~D7WB1XqUEifi~bOp
zxT%{!-vlOZ>NB7}112)HwW$yL2aG&Y{sHqpki_vr2zw~<55fFU<nz%zABS-HeDu%9
zAdU;ry+Gs((7!<B%h0_{<jc^%OypOf`wEd?f&ME*ehY?eK_idUEok0?NgQv*u&pA$
z70p{kz7E~%a0r*LL;pGq;&?Z@?-u#p=)YU!8_>N$<QvexLF60Jy;0;F(Z5mTTQRH^
zjg%>@mk4t!CUJZj!!C>bWi($F`K#!D6^C&7tLT3fgE($S_jZwQNB?$_zlrWQMgAuG
z-xT><=zdG&Z=wG!&Qrx?abkV@&U1XaIX#ZMd!9Kjj$FF##=^S(MO~X%s(75Qfv766
zomxKx@SSGIX2%P`myl<no$h(@ho(q}V?SEoCH8e@AHZyX*t<l>iTyd6;W$G;nU0wP
z%5lsQ&`NMz2?AaNj%!3*4UW|!-VcuZMSKJtkBGPh99u+u0~~LNxB?w3P+0FMIv(Yi
zGC9+T_1g#6NP0rLh1)J8A;Utc>dPZ%CvP45NBd*89es&m1u5G<dz4_U;Jk3MmP7Z~
z89GH{_NC2p^%rW`t8Sy#D0Ft?Q^rrDlS^_T=n5f-vc1IG$J7#-PVM1Azc)vQ?uxB_
z%vJ61<;x9OA5J{L_8yrQIg?_HRpT`COraX5#n0p?v7}jMl8=+|Sl4F#E=4+Ne#UoE
za)QSPkJ)<kC5&0G9CYOn#EV%i=xQN|ikCyr_c5xqX9GG~KTn;piglS94~;*sr2%~c
z+z8P}G&Xy>E<K2j&fb;j@`AW;Sii239w<37_1HeAQPn2_Mp(b4QJ?lh+hYAn=N;A$
z7sgbh@NPXOHAZ;1J|T6&bKD+<kX8sYxMvkZS~1Mv7nf2<D}@>4X+Mr>j<~yJ06b>X
zTsyI`zIIVu0`oLOieVb}wCI%RY24EaAhiId^RpB|Y7tDQvve#!+gUU3`Xy$8NNr>g
zpCm3dPEbg4TJj8X&|9Gcn{sAU9b%RpuN@!AwN?VU5*SHZp|6*0<_uj%5M?2j1wyjN
zQ5JCM{-Q~y>sq_FF<{5wp^sUS^dZ&ig~93J+zT^;Gs0=#4&5JDX{Xe}F=w-VQYE*~
znKgT+_5Y6|te~5nA~)xS+_BceqowMjK>3q}i;wCohy0bpv^sx<|G$i482_h^|MUL=
DW8@;0

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/max_pool_bwd.hpp.4F15616E2BF6E596.idx b/.cache/clangd/index/max_pool_bwd.hpp.4F15616E2BF6E596.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00b1494d99571ea9b93f6f8a255dc1fe901d67bc
GIT binary patch
literal 1614
zcmWIYbaQiKV_<Mj@vO*AElFfyU|<jg;^LB`%q&I*hEG6&iaC2H9?Uyrz|;0w!`A2W
zgoK3(g&nKp0*)R@OlrToc#EO(BHv9ax*?bE)#}DFUhF(Lp=`74^UeS7zj*lWu==$^
zCoYdUMhErIEv-1;n3dkd!8P%K@$G#YQ~k_sW}N%ud4G1ld+mC@!v<g8*;l`)y3klG
zy-dwy#>b5QospNe_N<$=np0NMRzxM+NycL}PpA}6v6H9Mk9YctvTj>W#2%5U)-??=
zIGI@Ov3O_BWCNG8_gn9=-~Y06OW#SBv^8<NUHl*1n&jG3e2Ztf<n32kTD|6Tj(OKR
z9nFuOZrbba?#ropp_2PnoReOjjsA{miEv4?3uiX)$jmrcVDP4T?yUdLYj-7Wy%w~@
z!QhzoRK~c-V>3lruVpc}8J)aziK|<9YG!KW9H!Xji4&3(C!0KPm~3h%+IqjFPdfHT
zf?}?1O-$zB>nm4w@ESk;_+?wkuA~2a-tyfQ-<4zY_j_+vv(+>auYxbEdV7!Y7FXsb
z9RP;B(ht9u#WR99c^G*3fm|RE=40j4WZ-0FVi4hD6lPqrEi}pZMV*8Y6PEyku)Ki0
z447cxfT-T1Z=$j42U{Qy12;^+5UY?D11Ae3ME~Q@vl&#w(kJmS@WK>`v5EmLU}J_T
zc(LsFUbjV=UEE9@KnrBKWQD*4%z}598@|QfTV2V<#L3StEX6LR3MOFEMX71U2bh7W
zXnm(t>AF{kS$Kr`4f##r-V)}w=C?!ely#L2U_RmzR`yZ$hw&7&e7#TCd8D!e4N%}W
zKxh$G5H^7OQdnF;+yL$+VQD34BS<(fFz^V=ipr|Od;`<Z#0d`1-b23^ta)gyz$2{6
zqY8I{8tb+xGTE{rOgzG(?4qIwlXTd1;Jgq&A@vs%ZCF7*<K*Xu8O$Rr%qh%|<S<Tg
zeuTpmc@>3WK5I^pa$9X^Tn02i*ihIMp+(q6*a5-QFw!uCY2gvp2+|0H@mdzwSNX0u
zdYls&Vxr1&Fkv9iRM`%}QxR1`lGBjVP)5jUSZlZ-c$!k0NOHPTy2v~$U1x+!Jt;l7
z-0bIOvVIKjP69(qkX;ZSK0Lze?CNkHOe-T7IGMsy&(4J_YOSW-S7!qTq8Yy>QXrZO
zTOq<g+)UgO5r`^!Duyugc!V{LHIc$c&sY!6gXssxH7M;P;#!MG3vR%n$Gt}%YxlPR
zUC+zT3wJM%uoAlxoCmWTXc8!sAWYKd(TAs+rv|KF*!jh~K?y>XUmnRXqQdepx9|vy
zi;By`%;OPOlvIQVBrJ_Gae;FhER2APR1|e!9)^hm0}_<!U{3k9ZTdky>(iSVc_7&k
znr8nBgl)Q;f8-g^P2z^)CJ1}Qt;OvSJY6GQQ<w!j!n%RFAuwK1YEB}^tm2~NVr39R
b`1|es*~y7=c}#*r;5q^%2n0e93Pb_`hSm%T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/max_pool_bwd_instance_common.hpp.F8135D216BA19663.idx b/.cache/clangd/index/max_pool_bwd_instance_common.hpp.F8135D216BA19663.idx
new file mode 100755
index 0000000000000000000000000000000000000000..08da602c1a76b688e85e89e1ffe566d03a218f94
GIT binary patch
literal 828
zcmWIYbaOLdW?*nm@vO*AElFfyU|<jg;^LB`ObJE?21B4+#hkT)r?U<l2wYkIV?y$a
zNi0<YitR#7c3&^;j)=Y;6``+U*~$6$`>jf=%oj`9bUP$7%hIRms<%t}uT}4E+w`@1
z^0gTs7BlZ#e)EaJ{=%J8x9nFIZ|>i@H%NP1jh4mPdD*-l8qaj?%}rbXCZgN&^lq)}
zbHX=YG!@KaQhTcNeA2d6D(9y5a^}`2t9wY?seZmiqbw-q#tFNrYJ3b+jafonIs<bL
z?v`b36f2buZMh`*UP)}0i|@fnY^mONJ(@PXJLmoMFNafq+Jn>P1yXWTK1$`e9Xadi
z-g9NYr`PH$O>d0E{|fzEygF3SV2kr#?k_Js@8WgPJbLp<epydulE(jx`_FB8#NSP}
zvlpJbr|_G_IesGzqkqMfxk*bwVYR+fs&w6}!`w_P><qkuY=WF%f`Jn#$H2hI%h;SC
z<+j?;xQv^L8K^*rO^6FjAQbFexT4l-+I@9yCN`h~0X6|nFo968caB8<;|~jdb2D)O
z6^OBkDS-(FP9|XdfUNi{5Vq-V{*h<gOsqf!!fe8vU;<$RLs4p4F-SFYQMI+@;b-=Y
zJiG$D0wOT0VV+>*0z1C<(C-Cn9$G8JSIo&z*t_R3P_>YtkOa&pO%MHlJ)RfF#>~Sj
z$ss8Na~=<`6o(WNPnts-&Vy+MS^x?Gm<8=0TxBfJ_ecRP5EK*?hdE>Q>*mVHhjaRX
z!V;Vkd@u`Ot_2zb3MH5!FlnGLC~RQDus{I{gF*x*T$GxV2#T5FqU2(05CalXAQl5R
cm;@7C2qq7L$-B+KTr=F`l>-YW7dH<B0GizxZ2$lO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/maxpool2d_bwd_bf16.cpp.E133A720E87AC796.idx b/.cache/clangd/index/maxpool2d_bwd_bf16.cpp.E133A720E87AC796.idx
new file mode 100755
index 0000000000000000000000000000000000000000..52057a21a24bb755d92533cb1219ab77ada9077a
GIT binary patch
literal 2018
zcmZvb3p7-D9LKL2Gu_O^(1e&7#(2aq^pbZZ#;f&eknt$;*z%eg^2jQs)D*4epfdKf
zoDMcCg`_RA&eCgD9$}&fDl$wSC9&+C*>jmar+dzP=l{K*@Bg0pf9H37-CSIRYXHFM
zfLm%@WYT_l0030r4>yVPqW}#6j%xuRRZ!1-8UNA}-}6-Dw<$`HTKhLD%~g<kD3-4=
zNJt~NjJWE0<_CpljP;726<M;oaukCcezxRV=5Vhn97Emjsb^OJ_hQY8)05f9F$3q8
z=?UWN)jM7n_VVA(=MN3j(l*-&%`cUdkc~PR3|=pxYoWdBc5~ysrEltHO-o!>N0Y7R
zqig-MoCmy{i(5+yYM%z2=jWKEyz3zydWEvjZD;K>4(iR1K^sHql=KF@uu|t*C2Mf(
z&MQn6RKT4sJXdkS!aatz+jmxJVUDF4k~TYzZf)Hw=M@7rr=>6~rg>GJ+cc^kWSgXm
zGnVKNu18#p<DX>C2H!0g&MVdC-?HAR<2NI;t_bY3K#8eZb+xSe#ZPaCs0Z3DTo{_8
zZolRRP3lC?KTvPWbN0m-T0c@F^jv(B=i(!F3qKH;zjbue&5*t<I$QP1Ts3#<Ql7sk
zqMLNX`LzAR(TqMPB8ky87O`Qc=rH4bXx?k9{yzDbl8_WSrf-^zFEe$_`DL8#ok<n%
z%L*;)yx{Xt_nW~xj)?d%2^6GOY}<gb>ebARYNZOC&e5xn2**qrZAK53T5$E_Tuq1V
zFW5U9?vM+ttDKH$*7wtFG1IuHPF_i;=+tZUrnJXw<I&r;qM+*0#M{nJ<F>up9Sl5a
z#Ah&Oen33^#ovct!fL3i*_FUeVl~VK-oJdsJ$X0T_x`Cb4ho`}_~^QW#Zy07giW08
z{(Q<N>RPaB$9|QWx)HsJp+{B`M$ekhRE?-{(pzuhvWj0Me9E1<dpdMGdBMU@?{Q0P
zQ%-tiu4rv!u~#_H_M{i<yuzEd>(O1T%n!N~@gBFH`e6&$oQ`L&5-Pkq8qNAm%WOzy
zUJ8kr*&C6W2E&&dM>a5pVLU<xd3h%P%~(yYMjqA>-Q?NGDQXNqN~ke@mBN3n&`(t@
zJA1R0DHd2UBkT7p7f&*F4HfM1|NJRKXQ2r{fdA8`zeRkaw&hGk-TL1rubQ-_SeCDy
zOX%YlCL|_?XFAv={&jKCBR}4JC}Q$>aiv|Awr_@7=xBNHbWiOpr|ZcdWUXqUov<{j
zeTgs>Q2wEp%wO}Q-*I8-#Nne9%7?{8z0$=7c2kmF+%ix)#`ms!cHzs#mU@pV!;+#>
z3+rl3SXsvNnJoSO?vnwcVtSl2RsZ$z4tbur?YpLA_LhRDvCSo+GgOVPyrabIke0#V
zq>8vHZ9y_4i<=e~{uO>#)xC~{8Dvjx#G$+ZfT@gCR*?!!q!+EGGLhfco{2;K0AnSr
z(t4>N*j&9^V|bFY7XpG|OpqslQbDjy5sf_06E@>e-Y~{v@v2fma4GK}dzL*Z9f$IR
zF@%Lwq=MiU`J#c=NB9;+U=NH{)l_w*g5c*4zhMnt8dl&?J}|~%ad@d9`0Gc%laxwx
zF%IPmV@0guI;kL-6B)(zM*#qubHv7~_5BVxh^eTdsEzDGYC$RduMUW5N-;G-PK}fU
zha}A)kA|2g6cckKgnY{>e772SucVjNy>trQ6xkD&DoEIO@9wZX#KaMCROG}E6Cy%X
zq&<kKOjM>KKNw=-iFhhfJo{j_Tsn5V2|k?>-be>|?pb$q_Qd&ROC^Y@32G`Mw;`q$
zs3qfwAW_DVK$46jgJc<}4Qk6c9Z*NcQ9z1}(*<>992KO>I6Y8L#_5ClGR^=rka32f
z;fh0gfSZ(z&}x&$6yue@3YjH1BpIaDkQv2nrEa8OI3yXT)sPXzh=!DeOb8s34Ap9g
zMxoJ>5GOKrzl3l(5nP63u~#DhU4h5~00LH*{wqs4=H(+VMc#dl&6uEoz`g!J&<X+~
z;ts?{!`4006D^Oyt^wELAVs{glB>$DodzljuBz(m2pStKt!!-V>~}cO*VC-278|v{
scUMh|NYbI`>ggLA8Jn1yZ`!hLJ9(?7>1L<(YA$Xpl=E&!hMWiRFQ%vEp8x;=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/maxpool2d_bwd_common.hpp.4213849B15309F3C.idx b/.cache/clangd/index/maxpool2d_bwd_common.hpp.4213849B15309F3C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b0173b789316938c2fd41b6dca8a2f13a37c6bb
GIT binary patch
literal 4964
zcmd6rc~lcu7r^Ji3=kM5ECL#lEwYIm*%3_%gpd$G5kcU%pkNXr7zhbf+$!QqiyuNi
z6%-Xwh<s=ww%V$=LtQ|yt%@QQ6^jcNrL}1FL%*Tz%)Ea8c@BU0&PmSw<=*$^&dq)I
zzL}X4k!Y?pf<#Cq`AS({8j2u@CHc|hsTU~>5afU<g5*~=CrrzE;A3$$?*7^6L&A=j
zDo5cgw%Nx1nv)y1cJ7|ogEyUT;NEd|YFa)2QB2W=i)r8FxsLLCzWLGB;`xmi-(`<3
zR~i&HU3=>+Z`#rw)Wu#tw5!{PpU^X7-{4-0AR;9DQeE+v*9#)LgI*rom)}u+qU_bV
zxzTsu)<^~0n$DV8zSY*h>)+cHQPfd&t1v#jd{lF(L^8(KFm;X9S+o4t-c|KZ3zqfV
zdYKxndR*|M)yr`&PN8DrbyDF%<CSvP<$YCpo8^zHDpr+_xhZO}P$kxO?U7u*^6!lc
zJ)L`+|NH5#%eVI*DSO_(merb;rV^b0&*ywz#f(`mPO?7t=4zWB=ZTJ;8FVNqS6!Rh
zd&c%@ihljlxgF_paeaH`uf{<uPE6@NJfT6R3vX}OI(GU`FMcxb?s`#id?3rcKc&~c
zVw~&Pot4_D+HqNxwM5*}XQ~Y?x>GB)4(7dg3v*2@1f|6wa>qL1SAo$|BU9_}*Gc}H
z51dleU9Rj%E?5xpV4Wc+P8Azebf|Y@{_gdciVlA=I%Xg+y6lWidf~4Ds4#i(Rrq?`
z#M@>Y>p=RDqGG?ty1;z}<>Ts?x4U2X!>8FJWZ?3f8!mQ9^Mqr(9kwq9lbf7khf4fw
z_e%b7`uVy~TdS?7Ny6-MM<J(gbzR%-xt2$j_1Cu)VjdPtDvYfYrshT@pT4Spv-+Ev
zFB`s1ar_;-X7bo{?M%Cyb$qu?HJzy`US*w$eM3CgMzpm1VRPUw7oMtK%_%JkpCE{d
z&M6N3myyQ*Wog+DRf+YrU3TwJ4mr2@-Z^1lwZgmoPW;yc?~r-2-&SdsDbt<Fyvxm8
z2bLT4oyHL|DKUG@-saCQgk=B_Llj^DQOo3-0XAtm|4ewBzbD%e$L<1`yTFT{igg2*
zZkX=@mwOC@=dshn&IzAjD5jjG^r5DqShjDrA495O)nta${^|gRG`<?z($LqeZcl4^
zF^*$WKWQ*!gk!Q0StvvDaJh(~jG&CE3@JhskfkzI8Nujs{Boc*bEo9;8J$|FmN1m(
zmj_wy0CqcoJ)_+T#_a?y6dQ(R)>xezSZ#n~0f_<0)XzB92xc^b1d5Fgze-un&N?_k
zp|Uvcp!DAw)lObD(=u~Bbrl@Tvd(g&$nJc>wy*A%51FEv)I~}gku+I4g)%}h8CT{C
z5nmS0klbDF36WS%M+wC;#%Iu%B8itlM;^r#Tm`hIg0G<0<owv2i#+AYJ}Rfi20dT8
z)3v1HgLi-<u_mO*r1d`eKWqlpn}HJ*EF7x_*3}F<nU<*5kG9$dP%KUqM@P~5gj(^e
zfVaT_CFc{Hfh9E`CFc{>43B?!{Gw?<YT^fbI4N%5Q`=mxgC9IG;FBiPTgzsT0ytI+
z&{|+hxi-aN;7XQJ+5n48BtmLI{Y)kjA%jRx$RWxgG9XeA3W#)o(1CxjzYYX4>*@ia
zhxK~kuV?gvoCDqt{Np3FvqX*S`xifx??G`a)*{xH@`5C%SUTNF@`(+Cbo|&zilQTa
z|F--_WiuJKW5DAWpsUlg;IbTB@8T__!kgzUrtGJEcP;*g_lrwpW@rJSWvXl9bSwAA
zr92G7u}F5L4P}pGBDM%3F<T5#6gvtc30p!_VpnSP%^k7)5vo9o)S4x6ueD)DIHp9E
z=8SjQXf{MDR0WY5RYRmfH4te5ss%JJo>|lBG=QEPA#W-7fazaNkZgLxBR4;^-TulC
zL*!K=xsf516^e<tB438YT(LhxQQRoHWucgaE1`9{M{iFXbXdD<gq$us6sw%I>-&qw
zIJR=qO1iq@*ec&uVU#~O_F3>}P(R!b{I>%Eqb&!)<p9!G0N)A#&8`FdIxwBGFT5Z3
zre}AO5t*Z*CZTlG`y$$K-7h8YC)?qe1INLcvc@q-jw3`)94CmJInEHda9kkba<~wU
z=ZuHQmE#JL8^;YIcaA$m9vlydJUN~aP2fy`$cy6zQ5X?MZwnj?C&D2T5CVuIhzN)x
zIg!vRB8~_mF-Hth6ekKI2}eTH)>jMW6|(JbkR788@XJ696U>8ve~{tbf^MEee2n==
z)hPEajzOd2H#{Kwl@9Q9fKEyj(*vHK;p*o$o)4dL-;l*?iN_Le<|da;TuSdO9P0$Z
zoq)$^?*RTCAY`};1a~p|{hLxMe*W~d{qU*BH5018S?*c%2~+RcJ8RBB{KP+9<mv!)
z4$%Xyp5epJD~@nq+pZ;Z-N(&`J`6U@z2#S3%(fzTos=b|3lez>Wyv6tv*ZwEureT0
zuoMvK087UdHk=F;OV98z+55|0_XQ70FRlWQt4zB%PCRMfpvpu(xN)B#Os;*G@^IQm
zl{n{a*7zfBrx%S-X>3u~#lP)QYCfuaY_$+}yk$R|@=-a}#$B<gFf@TZoX%lVXlu#D
zGN|^*<K*<S3XWw6GoqMHs1Pe4OC1pE0HoCep`PLIqe@!7IMOn9g#PE74kxkOUQVqa
zyb127=G~htC^Z`)Kb~Y=Uz1p#$|EaJfG}Vt6Prd5*2uKazD(V$5NpBk*)PsI&Yd#r
z`%*W!%d35l363S83G_i4$CA(_h?3D{h-Ra+A)15EfhZMCg(wY8gD4$Mr)lVHo7)av
ze&zr9Lq7R0P8&Y|Xw(Ze+F`+B`0KAR{Lbp{VMGZCf-|BqBbqZJ)=vzkch<ConOhl|
oni#Xp3ATkLZiWzyi5VlBFe3R{j2bLTK#<|r5`PaXN<fhR0k<m7XaE2J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/maxpool2d_bwd_fp16.cpp.AA5B517C73C2030C.idx b/.cache/clangd/index/maxpool2d_bwd_fp16.cpp.AA5B517C73C2030C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5a476c72ff119b91b9290a778b342701d5203ac9
GIT binary patch
literal 2026
zcmZvb3p7-D9LMhr#x!#=<I#A`7&FW;##<yKkMan$v%+}Oh$63cCgf4sCGsepwMOTp
zt!Rm|N>eh4tSH&YQ3@#;uX6NI30du(*>jmar+d!*&i{Kq-~T=Le(!H~db+!hDFVRO
zD9?il(FeHl007{jlYf90dqEBWT-E`=!Sd_=!--;B+>_o<v6g$vZ@#X|BC8s{)gCeE
zv@LYN?B*HZtG1OABIajUpJ*w~gLze9&L{2o_uH)rI~AM}qqPt0{IbGr+eN`)XSj&Q
zN~i8yc{8VbQzstYJd!rHu<$^a&88Ps6Q&GLbhqV;Z92wZF<82Z?-E;9+Oq49jFk49
z7mMf8?X~6yeFSIe{k)~crP$?#iI(P;y}coKs{3NF+g-)Msym#Ix|q983C6vJF#|4&
zl(6BZEbFZk<;^txJ)Sx=VJV%FesAeZ{f6|(zIqj~n4sWLX}LT(XgYXbzribD*M4_;
zRDZ_Q$5fZYB%Oqr_{$YjW19nd8cz{oFX_u$X%yFGYlRHCkETUsr=-jX$c~|ha!lf{
zP2ZIxF@!$dq;Zp^G~u(Wo|wOOdJJY&U`hu2jNUkS7J<xPQJdYv2`T5i%jKt=<H9n}
zS8Bu+iD=>Y3bvI!c^>C+uTQx2I2SiOB4{cTr&8li19jJC2L1p-UdTlXX)%5L&B{GV
z<0O+uael#8#)GaA)uAWOoa5x($<OmD@j1z_>x&6L_zVA7Uq`YBt99bO9W4k)Yziyu
z2ShhcsEIx&Uwq6ZBK<BteDR8LGQy&SW8!rUY*mb$6NvpD@zMf=|BPvEb=xX-cuF=?
z#F54;MIYu`UQW!`RnS?DiLu=L7n9RB8omVHDcB{bEpWJyWi$J;dSFcAUg5se9}d|@
zJUsR|C6~2lCRDYDi~ktg<vG+*%J1ZyYgG&CG9CYv8%;oNta~}C^TFX*_*30mbYFw+
zw)mF3^qS+d>!M3iBMUMyXB!JJqn-6tC9K0!)Y(Mu@Uv{Kzl%!vx$`1!F0bcMS3^-^
zGVynfQOtB?`||cq8-AB{t&m)xp6R$!Cf8QoK2YH_edhCm)x-lOe#c?vF8}6t^h%8e
zs%rM7h7SL!@=XB~@ii+pBgMz=GA~^gtwgg^!YDT>IjIdtBQglLtu6aG8$=}ynj_-!
z&gZRMc5q8uTmB*<xmQD9P(~ezyj%TTp;0YE)2V79lDKFXI`^b7*&+5#Se{M5EF!_@
z>XnyA7d&5gop~2u|4RO!=f}g}yzfa!IYOaKEiV&pd$=wpi8Ei@53L|vM`-6`&eLWs
zv*XIfz6m1=hNf?>&pbZo)~h3k^3@5ynj{k9um6^bwO2FR@~!;qK*z(HOzzi-<y>;x
zsgz~0UOe{gWQAKo;SpeWlSdOjEg^CN`p**B84>iHw?<eb8vr;eC>6X^AR;}OS8Dc`
zJiB)oi~Iq?$|z+GsleER@K{s)mKOv9p%7M=S4T?)#;h`??qb1ED;D`9gmEaGs#IWH
zbW1aBIqE`<65t16)%B`Wsld3R;2*~jM^ZW#=?h^H1>&Uw<38tF6n)t6A{OZbVJr%Z
zlM0L%-uND+)L6X3B6mVq38kba6&UlP_wYGL0FdKlur{?%Iw3%glBN<F-UZi!#Qd)g
zkYh?QHG$t6E(Z!pnt^W(a!e>D7H|muMPR;L4YXI%%i3NuikT_A=cj!+h;-CU3-oXd
z977MjocG?1xT&{DTOQ<KiC8LJ1jqr2AQf&6<fsr;sPKD$92^lxg}+Ckmt4+l@i#kV
zkfV*(R)Oyea&*u-GLDEQ$~Y35B;$0^x-yQ8Cd)WIw4RKkpeZtril)jqeYC!eGe8^2
zI2xKJ<LGF*j59<Vt~z*%pe7|FwAQ3K#eB6%xD_ZQ8Kkw4F~xZGd%{IRA;~zcg%}hD
z6D|p!5GW)Ws<jZ4!Zd?Jyy$qYgz$O0`5ux*kVO7F0^tP!1gtI*96?$)+iqZ=vO;h`
zV30rHzlwn1wdL-FsQA5*^@`puC$E4~M6bhwN;nneZFtu$bUbF8DnU(MQ_F4>%VD#l
zle3uy(_Y`!i2S{)>vf1EJqp!;X2>u$F}JX^w%Mq=LEF@7tHyeF&k&^B4i^uEH}Efm
CKIHQN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/maxpool2d_bwd_fp32.cpp.3CE2CB1F186065DE.idx b/.cache/clangd/index/maxpool2d_bwd_fp32.cpp.3CE2CB1F186065DE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b4ac47326dca4fd23d61a2cdfef940061789f103
GIT binary patch
literal 1962
zcmWIYbaPw8&cNWD;#rZKT9U}Zz`!5~#Kk2=nM;@%7#ukn7%Jv$4ZWTB*i7K+`7fD<
zX>)eGT-G2jr2bCouF5g9ZX2%TA{D`99#h5RzrD8pqFT1+X8W>OhY1$NCk`8n2eCRZ
zK3H|iBBUYC?C`Fr*tv;}+ZWrP-rfFs_WAOsXLr<BKKk+Fl$^i6TF-Lv=jwf@4|YB_
zJ^JLKrnOw&`@DVs4mPi^zdQA^e*fc7RaTPUZU#=9uXefU*Uw*>e||jOw`X7a`RD`7
zFJ^H2xIB&$@^xr+)OUG4?TPo6jBAeUim|tMHW)iUowG+(In7gAb<0#W&8p+SzOOf_
zdUkO=KTp3ntIcA=KhJLJ-aMha$tygez)daXa`p4i<&Jf-(z&m5)+~Dd$Ts}k_8H=t
ztCd-dMEln^OGe*vdr+CuTvqman!H_1ZHrFM<`+kpWi_`1o|b*0lV7>z#zs%BLqYC0
z8y0a*zIjpOm5pZ)kM<-sTempzvIX9ASYGVOj;&d^STwt5muj5gB7Y+*`ELU5$1iUA
zb)`+<_T6bar#~)N%AU=zcJu3NOBtf?F(++N&A3=>%XcpSiLB0r?2srUt(z|K%VYZH
zE(mKs(An-aIdDSp+KU-+6-~t*7mt>@Tkn5*!a_AtKxSv`^vevpw)F|``dYf;N}>!?
z$cnS~9%@dAH<%oz<F%P*KYP-rX^%rL6jeq<t;^WI-_`lC&2@P_b^+NZi+6qewD11Y
zw`&)vS#QtFOzgb(?8UZC-y;rmB}`k@WwWHg^lhU^L`GivmaN$yYt7=%c3dfIvrc;*
zBXllN@N?!#&s#^Q79R^+uwOXxr0$b1ZK>i+=4<aikp5`X5qDi~w_32u>4Q1@+N)Of
zzU54vT%Ocb!!>_<7uSPh>z7Tou6wTZHqSe5zQ5#|p2@{+-*zRo6`iX+xv3|wRA!#&
z1FjcIhyVB-Gc7)8ylRVlmq?A>zv;{emmj{i$o|FLuRn~Qp5iS&TBn^5y63&xV$n@X
zLd~l;9Swax$1?nB&dPtLcl$ezXs=$gtNZi3H@ln)oH^d7aPR-7SCyW$Y;8t~daB{B
z<CeWU>krS{Qq$Zl{F(XHJZ(1}eb-$_JN$QY=c#`)sPp-F)Z&x>jwh?dUsqhZvYRh^
z-7V%@i;@!ebAOzCX|nbGn{kIZf9;gsd*$KW+i(6kIi36YQ@^Ue<J0|PKQEPT`ODm~
z%iHOqllQG;w(ELF|7Y$O*?O^1_lMNe?s?`qCzqM+NxEN>shp+xQ>;WSpyi+C_qE-#
zgMS|ScKRo;eUstaAIBYUh}}8s+I>sxKSMyf`^MtR+@v4CJS*;B9k15%PJ^4tpMfEa
zpN(G-OaNuTKzqZ=g2^|I*KsrX0j2ra_(Z`3Oxi&Fip1l0MUgxVF+gb%77-3G0h6|#
zt}XYo>(+j5rU0O{0Gogin1D$)Mc*xKbz?ok%Mb#T77`Ou0uwOlMO}aGqU~g>xS4{1
z(mZTDf?xtBebHeRo7!#l72HgLKxu9^ZUHa>lm784xK(kb!Fz6|AfPla8?P{!fJqmn
zrWK1aF)%O}Ra;vgerC_e6UHmSD-Y8J3l=6Wl;GeA(^b^ff%zI1I6y_9kbxP^6Q-l6
zV*pbGqZzrd1`SXzC|nSF^%V7VVR{O)S{SR?Uhe}sPE$Zr0VWI!H=tfnpds|?E9%4b
zc6%|myzT#Q!N(IO#UaHH)5{Yk%^{7%li`p-;>mKzBJt!n<dAss9P&s!1r7xyo+5`L
z5>JUk35lo7p^U^+;ZQ;1sdA_y@zglfka+4G>Tn(`DFA&8N(TrZYbk2MqXU*8fQmq=
z0ij4!QBxb{e^>+q6@k(OLXozjwjN9oLs4o@B1o#ZD7n}j#E1kD5D)=lFffF}nT%n7
zS<_a}Ubn@8H7YzJGL#_{F2xXqU?%6L<fH?+Ua3CJEUawo9Gu)dyaN1u9)d1TYJyxI
zLgK<A5|S2{);6|w_6~ZY+E&VD8uHjg#iV6q6%>_JRMj=Lbo31jjZMwvOr&&;oJGal
OJfoRheI4Byy%_*cfzR&%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_add_instance.cpp.F2B6A1DC94E17C44.idx b/.cache/clangd/index/mk_kn_mn_add_instance.cpp.F2B6A1DC94E17C44.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3e05c05848c6c9e2681b7a6159b0b9e622bda9a6
GIT binary patch
literal 1924
zcmZvbc{o&S9LEo4EHmerF&M@&lYLE*b+XGaV{9=7r;Q?t)R5(h%AlUdB}-A3HWe2m
zTPj(?r7O?liBeLLCnZu5ik9mR_j$R0-19u=J<Ip|{k^~UbKbL^$zU9lfFP%R%=pNV
zm>>)UK{CQKI))dXBL+c_a1az<+{KBEdSMBFU8X3S9V~buI_^}Yg&sbrJZ(N-dfByM
z<F@=`xy1!P7F@bw>q76T<|9L==<0g5nzY;l;@Kj1tSvlriM#Ui)JI$Hk|@n3fA{Ps
z+>WD#{yb!>QlhuZE^nf@9zXYXp?6Ey`*-YFmEhSKJw573-zVF)($0yt&|^!pNBW`)
zGJ3M*iraP@E`90zShn=$L1J>EXJHg&#IsBg)Xw6n8SuT--y@Z;$TkXo=aU|tWO4_M
zPM>pVWkmJYa7r_Shgpwxo9{34JWIK|7Jo^N&C>ZgL2%M!Y?}FilwiP0vDQ&G_kGLU
zWE?-4<`$WuIeoS4oz2@muDVrtr3S{-szkt-XAUziCx-E38#%oSn?{<`&knUuvc=5w
zgG%#DdIE4uY4sX1@~MKXI!xN#Ft$W#%~xG#&3ec1RE3<Mj{9Sf9OzcuxeOP}m)=?4
zTvw{vcLrm^DeCG}c_HgkFEMy9ja{$O#;U79BtJ~I^4w35zwI*~R`>1jvGaI%{L$7S
z$(v~dxmCH_#xk^TmAaJ=s1awg51Qgg&jJ<Qv-Zd88#mp#S^F5R_+6JN)vj=WtD|JE
z@KolCpqk7o=~J^eYM7L#ihW1a!kwAntj=%gL-B4$^~=j`&IV4sKRguZG-<s+V^|PX
zuW@X`^geN$lPdqn+=<@4RU<!gV`{h|5hYr6RP;p8Kf-wLi3$_dD(nvMA2b=MUz%q8
zSz&H!EZ}7XBSq&MwQ`;ep0m`@dp(K-GBw8MMXnp1On*w7J~ci`Y%yvs@5S`{>%W^x
zoGPIIMNKpG(M)kg7oUWl6+x0)cd!Yglq8?ZCW}0;f$H{$mUp#%F4<J2?sy!toyXwP
zJ*rI(Q=Zgb)5S1LqBC_W2a-DqUq9B^+!WNH(YU9xsD1FnMc=7W&4rI=;MnIOp97W~
zF`EL06DICf*1mG8jy#<rxzG7kjcDc6DgIAIYo9y)qy+k0iHS8W>${zwU}^ZvLtf!x
z=?D9#N4vv1&d1fIaQPqF<<an7)=swz?w`AqCK`xil==Rmp%%k;_d_x~tf`Lqhms_+
zXk$%DW4Ee}GwGx7nHf9m+!t;;@p^&%&T&$IRG0YieB`4I8Tlo<#F<Yf%&1@WJh9Ae
z*m35;#19kdZHMkRe~-i$zcqciw>$aGt{XZpo;xrz7JuX|eRfHqrcUqnsD4`pMJGfC
zn~OjY=2Cx*Ip<{+4EY2Ie^}vx60t<je{Ug3JmIUPV-n>6Ueq%HLUH0a>D2+qsCP~T
z{ew0YhWtR9ge7r+9Gf17d~YQQi*E-q0ZVWLa>6rnYJ_rjEDUjgERB_R2l8(!f<|2V
z5k(l{0$C0#w^sZbzt71m&+sG+`3hSjW62h)14!Q+`?Mvq%IyLS`Gd49R+bCoHoD;U
z!_kkaFyswn1dFV-)>%O%$Am^S;6=TJ#qr{J!s-ChypYi7`Tx)pm0G2i+419Y)!5bz
zFsh7E_5ck6hN8GtcIzg*3SJXTgd|;ruQc%uL|_z0!l{4^04$tf1?+lO6{G44CIFa)
zrWL&FCMAp#3zz_Gp=ky3x`~3Juz?967n)X3ubXHX+S+u$EP`7F4om>JFuMX8yoOOK
zhUy100K3q%f_vShh*4Z?6)=&)twLWnA#8*LD!{0;hcp+g3PuTR0tbMAQKA<yU~Oxo
zBtOz_KoUk}{bhqdvoI>>C$}3c8%D{NWCze|-%;mha#GJdg`WgZ!YhNWcp(u%EBhdt
zw>Mfrcojp9%A^O~8r?NS0-->#Lm+5x<i3b7;iFrqyBG#5E+Hv}hY=z{nknP#q%DJE
zI+7G*<rFEFR$HxYZ0+og<PGQ==DM{1drMYUQQe@XuBoM?r*F7vv+)))3#zHoMw4yw
PWEa*BQHF<us|foaMTp>u

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_default_pipeline_v1_instance.cpp.B364344C594966A7.idx b/.cache/clangd/index/mk_kn_mn_default_pipeline_v1_instance.cpp.B364344C594966A7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d328b71dd41c5e7d84837faa2b7688faf126bec9
GIT binary patch
literal 7000
zcmZwLdvufK0mt#Y?*m<ca`U}C@AY{vkQQ7Gh=pLQLcl6v5CmkKM6MQy!iuFzx=lb(
z5jYg_GIT?f>BdlkiaOnpuwW3cQ4$d3sLaI61OYEmKqjo^oYLPP>p5-DxB29EPM`BO
zX)=CPUEMg{=Noa;sKmVbrs<u0K40(F&jn45@p|I(mHa+mV*T!k^BT0Fy*@rSp?i;Y
z58kr&)P3XT?5*x{<i$6ydF$Zw+vP<cJilYu@dFFrK7P+VlQs>$d^7u1?RSY;8=4nz
z<9w&KT{fk%*YNkBf2#JC9lbig@~3mVm4yd}Fy+6qhHhBj@8jAB)*XGm+gsm%#%8a4
zWB((Cf2`X7(HC1@{@qJ2pZ<13kEv&;eEt2VGkteXd1ltk@$=74fARC9YYsfN?&M>?
zIsJWAd{L+0Km2aZwB0v1tlfWIW$*7>su!)fvs=a6Gxx0>)u6N8ui1xZ%%5}culwr;
z9(yt_rT%sHvp>`|EcLIc=y`Ho=UV^F|BQaA>YpPknj-6_Wy;39J;Q(Hr{~Oj_DsCx
zm-V0A$5)@&SW2&)`{bizd+mL&>2_;ypZK~Z$2YDzw0&&V)E&P)Q+DZdPwai9Oq;l?
zZ`JQ!ujzVv^>Y=yrOb}8+sf|#%bYPM>ZS|hR*rl&93R<h#aA<W{`iNLzA<0Fzo}wj
z!fqKgxzDPH!xul$^VTJmHE+IlWPv(k+a0ghzV&6z{SVzTJaocJHh4{oqWA9YkzP6M
zr%=ChyZ=4CVyPS%9qzlrojC6|*F;}GLN|Xs?&iz7HYQ&j@W=9Z=SK#fE*+fq%)$rN
zhx!Oh8}_B%sQ2ACG}`%C)9P_Ynp%2CHoo(?pAP)$^#>-6j0md-_%?TM`F7&8%~y=9
zTG&*3{JmeD8TQliH>5s2o3CGU!^$%^eza)czW3)ndbsA`s(&}$ZXV-S{dwfU7wT7V
zhbsFUO9o%Gw>G<L^VY(M>c>91c8)$m$oD&0`S)i>H&1R{`{OH(?_9g`RCLwYoWJ>w
zQf||*5qit1wYBE>_pa|X=FuBpXt{sr*~fnTOH1Q0=cutF_G#7c|3jr=%cgGq<#)w}
z!w-M>)Y*6X1e*^`oB7Iz7hU%3^40fE{^MA8_6z&tw_mp;`1q*mtDSZGh8`{pKh$H;
z`>)-{?tDXf!n{?|r>?%FuJq*ibF==@uvng7Ipv29-5W-i-~ZX8C+|E{w*87v5AE*r
zw<TL{zG2eF7k3(;9{6z3-8EHfzZi5?<*RRfbJMAT+drGUCt3XI*>CEGJp9F7Ctizg
zt=kswS|{vTx=x<5=8|iaxlfo+-{#d1+$l|t<(6h<Z``nD^PzJ+m#+;@+CjVBaNSup
z<#e=p_1zzi-SkEE`mr^8@?XwxxV$P7ng8a&hOe%lGV?@Tb#?LN1E)`?H#+@N{2l!U
z{MWajdESgGTmR~IUY>mE&BMp;D))`M&Nrc=bH(WMUs`WD|MBWmmfSRJ?hD=cXzR`0
z%DVMD|E2xr#`;+cj><k?8GE^9#Ks}^pz;YqNT_MQS8IiZtDz#on2r)dB1kbTf)pbn
zNGT+Nl)@57DI$TCLo!G?EQ6FIGDsz)fK<W?NF}0xq#+8DhABuIp&-?e3Q`TLAk~Np
zQVVGywXg<Ki)bMAkPcE0>mc=r4$=r2AdRpA(uf!!&5#Mw44WX$hzZgPSs<;j1=5OG
zAX$ikWMKx9MHomsWP`NBHb^^SgLFa;NGI%obRrH&H{^nJ!!AfS;)3)-9!M|jf%GEY
z1=;n-$3I>+x_Uy}B^?!R=(uc!m@peEE;MzN7!^T^F%hH~7ePu<38WO0KuU25q#Tt&
z$}t(F9G5{VQ3a$DQ$Q+l1tg79kTga?(l`aFMpck%Oa-aNRghX#1F6L{kXl>=sYi8?
zdQ1nY$90fK)BtJ343I|L0BJ@|kY>yTX~s>ER@4G%#Vn9k+ycp>3?z#&kSxwX+EE*%
z9kW5&aT}x)bwD~X2c#2sK)O*Eq#JWVx^Wkz7xh4TF%P5{_b$lpcWfx%_|Ih*bLA5%
zdsNc)hyLErtk7=TdgRr%E1e2@$JHz_!G#J4{W?mdB1n;mAjN<PQlb(_iAf-(fCN&e
zGDw-pAmxAzQlSb+g()DFfC7?I3X(Dkk_IS9m8u|Brh-%hDoBlLAT_3e)B+kvo$4TU
zri0W2I!J>WAPr`KGy(=llbRq+W`Z;WCP<4~AT4Hrv;r1LMj1%P7)TaiAZ==cw3!Xk
z4%i?a>VR~Z1JVgNAYJN$beRj%4Y(jZ>Vfo_2ht087vxoEzB%@{&zql_)p|*%Ibj#q
zap4L%PRL^yIH8E0<b|y~_!KYf#Ln`<yU=2q6LDY6aU$-E1y00$G0BU#FQ#}A_r)wP
z;=YvTB;1#BoP_&Qfs=4wO7ar!ODSH$eJRUJxG$$U8TaKJC*!_c;AGsFle~=ka*CI6
zU(WI}?kj0d!F?siDY&l`I0g5WB(LDUlHwKISF*f<`!vl_+^0E?;yx{K6!&S8r?^j3
zJjH#Q<tgr~X->s`HOHyAuNF8J_thk?;=Y>VRoqvzyo&o;n$vJ!%W)d+YXwfjeJ#mr
zxUZ#n4fnMyui?I)=5*ZGbDWO*dV$k%Ur+Kn?&~RD$9+A^>$q>EIRp2N9B1IZQQ!>R
zH<G-8`$md4aNo%C2JV|_&cuB)$C<cq7C004%_MK)zM0}p+&8nliThTXvvA+aaTe}d
z1<t~KE6H29Z>4w(_pK~%;XX@q4EI@%W4O->9K(H<<QeX>6wh#<WqF4CcAB$s-_CJ1
z?%M^<#(g`<+qiG1cpLZaEN|n!lja=UcXFJA`%Zy#aNkMt4(>ZC-obq*%R9L5ra2e)
z-5lrQzFXj2+;@|_i~DYhcX8j%@-FUsY0kraFUNVf?-e)?_q`<V;l7vRJ>2)Qyw|Zm
zeEC<t)1{_ET&D^BFX=y|{a>nSg?ne09kxI2-TM4fB*M3(<M~7^5)oS}5(!%_5*b@5
z5(S$U3B^{6M8(#MM8npLM8`IY#K1O-#Kg9W#KLAp!m#Zkv9X;Zaj@MYak0H3@j80O
zrK7eB(^D_E9Za4GhwxzXM8uZzM8cNyM8;O~M8T$cLb266QL(i=(XjPA(Xov@F|f@%
zF|n;Yv9MX5Fl;+dY-}e_9BemFTx>5-ypF!Llkbtu1J1nHb}$(t?8Sr05D{C-5D8n(
z5E)y^5CxlN2*p-2M8(!JM8noIM8`HV#K1N)#Kg8T#KLA7!m#ZOv9X;Daj@MCak0G&
z@j7~SaecIE@vrOK4yHteAMjvGM8uX#M8cL!M8;N1M8T#dLb25nQL(iW(XjOr(XovZ
zF|f@NF|n-@v9MW*Fl@U-Y;31Q9Bj8lTx_pIypI0UFS<0|^XilZZ3p8QBzT1Ji%^nZ
zf|C6*l;T&QsGmZqeicgdYf!phhcf&Il<7C2EWZWC{0z$W+fa_*fpYyWl;`&@)YC_o
zHUIjP^)0P0V#~Xf_h|pM+aGzkuIJysT)pee<6Azdv3E_p^2@30z@Ap!zf*s=UA5nf
z?|T02%y!+1;6Xfl+tmleE+0hOUM_d4?B0IgZ;tn^Ts(O3rnYlW2+QE5gP0I8rG$hj
zCuB?|p<vPk#Z(h2rk2n!^@NUTBn(V5VPaYd3zH=nrk$`corHtwCR|J};dSIsy(Vq@
zs3G%t+c^Y<OWTjMe1aGhp`@S$B?o0FC8$8rAcazcDwGz~p!A>)Wdsc<GiX9tK?{ln
z8I&Egp`4%t<py0SFX&w;Ut|5;>22?SFKC>(KxzF>Ce-HL5z9Z=QpJt;jc<Qv+~=D)
z@20tPT66TQF{D%HvY+|8bSv-Pt7nf<y{{e-=-qWx_##CRsTvwuQ+vggR}H&Fy7+Q$
va3%YHKb3XEv>mrk-->?ye;yn#a8UIS>oTqC(vgyUO=KcpH+J~7+%NtI3_AL>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_default_pipeline_v2_instance.cpp.A590A56B0084D88A.idx b/.cache/clangd/index/mk_kn_mn_default_pipeline_v2_instance.cpp.A590A56B0084D88A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a721ab32f09214433e04954e653f563f7070709
GIT binary patch
literal 7022
zcmZvgeUww>oyT+UGte+Ij3Bx9KKJ#xZ-n7B4k83lUV<|~28?A`TG<H5%SaHU%WFVU
zTP!a<fG$$17LgGt6^u}Y0%aNv1Oe>{m~;hM#U(*k8MO<<+VZlqGn_Nb_m9mv$vL0g
z&;6W}-+As!?u4;5HT9~;Giv78_`F$*9x3s7JY8Gf3l}Yj%_1I8(d+TVw;Y=|Z@x0P
z%lYQI&K2W7*gxjUF_rtidb#BL$zN=r_01;_m%aPVCsWKHn&$uIyH%?u?HoAd4ffM}
zug2@QE?dYw==sOn1Ey4V8Ts)$yY6i~+@(X~tIa2+rDq1cD&D9cymd>@^TS^FP18GN
zAAbKu)7{4|UK)4#(_@#utDW@epQnEM<joPQ_dMPAAJ?wl+&Fd5Pu{8O8Eif{`ReyO
zn*twRKXu^7wao*D^TRU#t?!+_|G<e$m7CTqKd!uc{Jmdx`?BiNiPsOTe{TAW=jT@i
z2h7dQ+5X~^?7q^e)0!95ZXY!A+2b#)zWyO^JX=-V@t@+uuXP);qxP=pEB*CVCG1Lb
z#nBnN_V0fCXJdZ;@(&MQn^rwA`#_I1v%^D%-l1;XS3BZ|)79TTymRxODSv*V?4`3G
zeKn4B?sZ|~=KC({zr6AO*%#K7=BM;{XyuU!`+m1{``%8K)%H&t2H0!vU-E|^t*EVj
z#vgCWy)g9NF2`T*yRmv)!>>=S4vyHQ&zkZ5q4KIvhRxjiMNvQb!G3eg(&s)t&~?bW
zJ<_vRPpjVc)Rs-n)Ba~~w~=K#x;JdRzw54n2e$W^xp>{gc;!7xheIDvrsG#V8r<>X
zlgsDsJu~!!b>WSjQiJAvezE^EZ*AH6izR=msa!ndxOd0t`Ugwa(Ipqh_O^Gdt)b8S
zwd3ezz4rdqcy060A;!|#Z?4$(zeg{YZd~2CYjodZH{W`E_?$USf6V+%|EBt9gAZ@`
z`!m%)w&$LEJMEe06DmfiYiApeUYPik%MCA|dfeT&zG3axo1XCf>)9QReGiRUG38Lb
zTYckjb&uGyjl-6IUAL{F^XMC6wpR4~s@Iy|W%^AXeNt<V+?+6a@V4g{Z@9nT6>-?2
zb*0-<fit`RdHL|R`H%lgbN$BOG(NVsYRcoCl(R=`PV^n$<KOQ-cWJ=k%J-zxufP1-
z^R<_@%qI=u)$(IAPCWPC#b152;oz^<@2QcT<rCg``khY?3@;7(0#BS;b)b94lUFj=
z-&?WrGh@NG)9;)(V$wZ*cb`5twc^TS-@SC?mCYxvzQi_N_#)o_#*3d_NfbW5{M6a_
zcJ(g{k1XBuYIOa~oNt|fLx(`$M)~0l{m#}5ozztG!Sttk|MB4K9oC)--S<%4{$^I;
zOnq~CmALP<8^ar~yn00`JM_u}rydKwvSh=~+=cUtCU4sGUz`5$`Pto#@1I{kXU_Y7
z>v8q!%2A^x+$j6Sk*b!XZP~o(!&;8IC97t4zLvc6&2rC!Klar1?9j9J)<?^hTd$+L
z3~^EY+<j$yq-Aqi$Fe(aeY9;}FsptcFL^v2*})B?8U~qnm)H4%zM*Y<wNS`c4HWi`
zZzmxrfGLCoFom!HrWh2#6hk7IVps%I3QAx~Aqh+=EP+Xb6igbTVA3!JQx3{t${`s{
zIV^*z1QjrqkOHO>R=`w)Dwt|W1yc>HU}`}POf95=sf9H#^`H)>9@4?o!#bEo&;Zj2
z8DJV=156fVV6qSclZ6?WX3zxF44Gh>VG~R%Xn|>kEHJIG1*RRe!L&m*n0D9((+N6Y
zIw1#4C+vXf23;`SkPD_8c5j=d*S4KsS$j`i>ynQ6CbutJU)1*~P|UZeorH)0rVtgt
z6k-CHVnhT}jEZ22F%e8DB7rGIB`~F!1SXA8Flm&6Nn;dDIU<88M`bYOm<*;8QNUE9
z3Ybbv0aJ~rV5(6SOf{y0sYNs}wWtQB7Sq7gBRZIRR0mU!>0lZW156`ofN8`GFj<6w
z$)XHQ7Gq$V5fe-^YJzFTOfap81*R3Xz_el(n0CYl(~jC;+A$kUC*pwVL>(}lm;<I8
zalv$>E|_l2y=``0v9-M6Pb+(K<#oL(deOF1f7d;0mBnveI@-F@sZVWR%|7O{f&9Ln
z?IchEOo0ht3Vs1hk&0l7OaxQ(i(pDr0#jlVn37)tlTr#MWfV;6r(nud22*A-n6h66
zQ=tl&3RA#T{0f*VRl!u53a08;!PKY*rp7cdHNOU?PIWMKrh}>bbubNTfN3xTOv7)0
z$tVMpF$N~{GcZkRf@v}nOw(_IX;BMIi&<b=ehW;S+F;tu2GjQ2U^>(R(_s#nj^6>(
zr7oB*bHQ}|?rl@OKCJih1x;etmL**zzP}S#xJ4o$i$x+LOGP3f(;}hBa*@c$N|7kY
zYLTeOT9IhTdXebJMv)lEtVkHLStKU1RU{U&T_iTLQzQ<uTO@8f55Io-+uwh=Y)^g5
zU{aj#2-iLsUzYRbkn^0cfSllcZ}H$KdEdLp8QymaSV(aK_Ju4bU|+~{0``RjFJNCt
z@&fjS3@>0`OmQOi#VjXcU(9nN_QeD*VqZ-1BKE}$FJfOxaT4~WEGJ=K%5xI-r35cw
zUrO>4_N5FjVV|ZrihY{pDE4Wdqu8ejo?@RSd5V3S;VJgz6enX}&T=yL<vb^2Urz8c
z_T?lmV_(kjGWL}er(j>natij9Jf~n^N$?8xl_alVU&-(a_SF=pVqeX2D)!Ypr($1C
z@GADzB(Gv$&G0JrwG^jeU(0eD_O(2xVP8w|8uqm$uVG)y@EZ2@6sKcf&vH8U^*pCz
zUr+El_VpyMV_(nkI`)keXJFsRat8K|JZE6vNbm;sjU;bi-^lO=_F0N!*k@UeVV~tW
zhJBXc8TMI{XV_;Mo?+iiaVGZ7EN5ch%yTC8%>-{^-%Ro*_RS1$V&6(}7WS<yXJOyU
za~Afk1aD#AO7a%=tqgBr-%fEh_U$ZZW8cnmHumiVZ)4w1@;3JE3~yuKNpTMLoh;{I
z-^p_h_MHUpVBbmd4)&c4?_l3eaW3}VEazh1&2ujH-30Gq-%au^_T3Ecw)aP_{>Jmq
z;-Yh0NnM|QeFnANYmS<5t^2Q!U2Sf;1+6Hl=-KuhVOoC8qa8mt&$nK$3&eK?uh#`4
zAd3YeB1;7#A=3h($Z~<m$V!1I$ZCP8$XbDD$a;b3$VPz}$gDsZvRNP|vQ;1!vRxoH
zvQr=qvRfc-J5TRF_HegHk_TG{lOw)!IG7v}ki{Gkk)<4wkZF!kWI0D<WF<!wWHm=r
zWGzQDWIac8WFtomWR@cg*~}3W*~$?M+0GFg*~t+H+07BRo&T-G^U@o4T|eDAm^ATy
zgo8;F0a;8F5m`zT37Mt|MV8Y<Mpn{9K~~d5Mb^?pL)OzoM>f*LKxS#ekj*qPk*zeb
zknJ?Fk)1Sgkli$K+qvyX<4SKm$XkEp`nESO3^ji{uG4~5N2e@ootM`qwBN=FUI9q-
zia?TA0-|0DBzt8b#j60RUKL35YCyVI2Qs_{5c4u1(`y1*UJJ<f+CYxi0dl?WZQ9+m
za@j){wrprQ-;{SMuW0+(M<x#YdMZ0}yyfdYC4KC+nYC@^OK&~pY@1z+1gEgf;_5TL
zPM<_t|3#Pd>fE;P<?p&zK0Wa1ovj0o`&Pn@vk(^$#khzl#U(@<r-*V~MpWVoq8e8b
zwYY|;$8|&_ZXmKaLp0+iq7}Cg?YNET#2rL8?zZF2E|cE=Y<~L7);R=x{o6)bUMB<u
zATb~UNdXCn1}Kmmkb#td0;C31AT6K)=>Z+c2pB*tz<|tv31kH<AUj|KIROXA4Y;>y
z!K}HDv|h~?E|{^f<$<_|)ZKUds|)|}+z^wS;F-{No%48R%$qrPb_>?lk0&J^I{wJp
zsjR$nmpdxPcC8-e?^-%Gbf?tKN9Ew)q4(Z5Z1{+NLhm8Yz)JT2eJiS(Zd$h6y+_Yp
feFAq?-F?p>V}R1P|NVkEIy{lDd0^xiZruL?x;6qD

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_default_pipeline_v2_opt_instance.cpp.7FA7EE3992571786.idx b/.cache/clangd/index/mk_kn_mn_default_pipeline_v2_opt_instance.cpp.7FA7EE3992571786.idx
new file mode 100755
index 0000000000000000000000000000000000000000..73c8cc8d98265d79be672749dd881b474e416c36
GIT binary patch
literal 1758
zcmZvbdpJ~i7{||C&|D0oUxp#$HX;!cHSQ)Ox!+cbT#8|%+A%Up*<BlQtF>jzEv3?g
zqSd2CG%nGVbeTei$ekozu(4-6&$08z{+{Rjp7Z&B-}gQ5?|pv9%ZbH$B@IE$eNMdS
zFm4DIf*>;d#Be#0;TQ<Q;vk5Z-{KQ}u-i=j)so)RjJ(JDWxp}sb(L1?7}qGyZSNR#
z(tJ$Gyd1@DsCRTuQ>j+U3AVM_Ib_v(bgDqg7S;EtlXuY;>)(ETW_r;cOouG43+f3J
zCG48cd(Y!Y*B#Z_C}dM|#vfI)?+uO-d2W?83)#Qa%BAPsOzh|%4<ghrCNvBR$6rRd
zK*Kb@{q>97?_u#vb2&jGg_%t)t#ftWWjiMgZZYumOTCr%KNuhC|1Ca)VKN#$Q2bhH
za5s@8+g;VAd$qu{ywUG<LvQDZt9<{*=0l18_mYf~>$%;RoVoTTTpTN^qj^JeA9GO@
z!gINO*jJIH=kZBx2d%7))isJGyE?IlLzvlzaU*Iq*Y1u7&ka3Uh&zq3pcZI!6e$=+
zzA3HK`ctHQRKe4(XIy)~$Ii4@v!?jGpS`rb^v|%czqc8F*&*^G7vXy6XSb!uq!~x1
z?<0Fhgw1G>)<4X;RXHD!d@kyPNjqm=qwV2s_58R~ud2z6-aqa~qy#$@9nZg)XvBMc
zvgkx-c2)K^@^!Q7V!5_4CH}O>kEVD|`E8?IV*Biw(3#Onfm+-d<+Iu{SdpOqK^PRc
z)m>_`f~6bFXkB%sBOF4rh!&xpX}@*?mJ;C0X3h8n^R2+HsS#dJ<laflXPS=RY0E)h
zt4o7lQ)<4K@yojz4RI9|!TX7t3YkDaZ@g;YOzdEJ_7_EHN#(ZsZ+YpPaBb)>6<qY*
zh)~9%Q=x>~Gu3nozg=ILaXB@ydPF<(w9qlfT8i3r=mjmQgo&@gU*d0a7}q^Nu*44<
z)0F|Yo))HR9_(jQJW;0K-XGa)bDT2!q<l=|68fN4T4msUjt_ITg3NJ#?j%r-j?E1O
zeS`=v*FCA(jNxwEU00eiZW|9e26I{>90OYJ`xU2?t@tk*{HMA<NU5d?@3Nb^6_WPF
z4<GBR7aWlL`>-)?y?SYGj&H<ddO=-Ap<YkqNy`g!cFKFQ)-|1aV<2K2So<L*)4WDt
zxh~;r&i3c4-)=W^6|4#oI1?H&4IL>_#cNArwZe=W*6a@G%x-2+6_)YeEd{A+1(h8!
zR;;@IPtdu!#PF;WJ#Hn#pYz=;W=6Sp%nKChbVk?&<%_}1s<`6Z+?nyD_&4&;?g0@b
z3^s3ziH{DohW|S3g=btd-<~oe<lzZ<D@!T6FK@83yqd17DRRT3st9Oz*d|C5<d!$Y
z=7fdEaM1r%LE)FPGaDC*%K9zTrD=Dvm>12}F+^{$1y~@(V6q1e60`%jB7~nW>P>Rq
z>@7}l2fGj%rU1Z}ph&<)NcLDE@nTJqK8ol~B~g`-6OnJ8q{iMDYZ6yd0xC!}Of;Y=
zK{`MWA(aJd{Ue-lMYs}}!hmTEV%RKvoV+#^e^VMZ1SrsmLJXKYQQ)~Gze9lmNiIi&
z8)%96dNgQ|<fmxRzQT4X6VTi&w+n^!Ry9;LN6y0Znx~&_`LqN#M#7Txknsn<OJz%o
z)O*GKzQurvm3~7_oN8Bx)ZY;s&oQ8D#rO#eo#z`Zbcl^+40yI;<hkHD$xVA=#KtPH
z2I;pHzBT^+`B|_`#1Tn|+~{LH8blMe!m=t>RTF6wvC;VDTY_8zI7s3wPh%Z~0Qbfn
z&ZT0hvPka7DY|Z=Sz?K}0Uk(Nnc#Q%A8x%p*gPcGHv$vn6P&Q95b+pdID2C#;ukOR
zw!ARj$2<<FspE4WpuC{H(fgtzVA(m`6@!(MmRW@(5M|}%NKRx2rVbhJWUoM34V09(
zY_+hovbM1`T%*s>GSj8~pQ@;$x>j97Q=7g{Z~caiM#h^qo2!{p4NSJJQFPqlgJ!we
IZAZEM3kZVGH2?qr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_interwave_pipeline_v1_instance.cpp.E84C4EA0690C6426.idx b/.cache/clangd/index/mk_kn_mn_interwave_pipeline_v1_instance.cpp.E84C4EA0690C6426.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e799a3bf3f954eab573332ecfdf3f4e8b973491
GIT binary patch
literal 7024
zcmZwLdvud!g2(Z^?^9|lm4cJ^eV_OGyf-3XB{~Bo4g#9WrQs53+3vtZT@|E}G6*e_
z0?G_{K?G#g8AiOp)By$vtb=EfTi8GxbYLSUo@E3?4W_Q#24p6Jplr)IrN2M6=d|a1
zlTUu<^gU0~G_^G&Ml`4%Pxa)Q`1C2w50!X4o?h+WP0h1n4aDQQ;`MmqufIQG`i!Gh
zJ^$xw-^|jL2dCY1u{wWb<2(L4c8Hfo9r~)qJ}51FyK(BCr$&xmqHd$BR^L7B7j2*I
z+d1#-vTk>CUmVEB*W6t)@uLmjo;w@zoqOs>>A_?7e?C8Y?zKXE*4p^xF%NA1sDI_*
zm3!Xk>;KoOOTO-3Kj*Ej`LjRd+YY?deDT}w7H%2#S>Vz?=MGd~UU;Em<G=?O?D}!n
zuB*)lw#=xxFyxDU^N(}4*3Wxk)AG>s>*s7bKd(F*e8^~BK4$LGi+!p}FNuq)_f;jT
zmc(uuyYZY*+_G)J+@IVVzpLfFJ(XYFL#rR$G5zD-KVMU~qr|Cdcv!4HXI=gK?RTx6
zw6kCL=X~p1D*jkeKiN6cym0Z_x3^yS536QpNpr>hpB$h4*rm4r-nnhX^ZK&PfJ3Wq
z?jQTK`uO+{wk#d;{Ce~3?3V`isT@9S?x3>QeqHM>6vEEJQIn;4hqg^9X;`>z=#cIE
z$jmWE?1^dLq*v!JCii|izN-GM13h0pQ{6Uk+~+GcMZ529`L5wfb<Kr&p`}lF>q}1V
zuelT`tr*<v-TNlp+^@#B=CP*1b3^s}M?LdgZR_|I(-O~oJURYs-I5R2?Ju3TyY=9D
z_tl>B|5zN`pZxD%TlzJuoAz|=x~8S~l}#@l{$Gn98u{kCcI~Q$r@2wnAAe;XKl+UI
z*!I>FGoSm*ksD498?)JL`C#MRPil|NI9c4i_CJ4h>dNud10^fBo@%K0g|GMCwYQBf
zKQv<em!5b3eR11NaolWyZMbxLMCd48_4{$#N51pS{eM0B(hHw_{o83HYSRlQHNQ0C
zgHx?5hjvqrJf42>t8Z2;+`O>i<mo@me&yi#tDgjBze36;em&|dd&Za5``3GvE3<a5
zVtcxO-&X(68UB_Jcb2Lr3cp+Zn<0%)AImo1d}nK7+}0QP`P|qIeap7Low-!ib3^m*
z->G``ig(OrdS>y)zg>K+`Hnx_^4sqR9eDC`>#wWN@A$4U{>z6~RQ`Qe#j!zuJ7XR0
z@4u(p<bTQixh(PI%oQuQPc54>$^K~SQ};c)^2CSd9(m%xw&h0~e|P)o_j|Ooss~@*
zJLkkLQ=Lx-bZdR}k<{q!dtU56Qfh3~H_mpZ+_6XaWi-1aJvH^_rav7Y_QLxgNE;tm
z+AY!dD%*4Ah9CYC`qjGJh48p_Z=Cw3`Mp7Jh`oE?IpwtUL}g&>oYqy!x2HbbHF<u+
zge6x!{<{Y3YHFGO@X+?>UCZGrqfuS)M7ig_@t(TA-TIEb_M`otYhU{|7-I9Ijqi5n
zBkgy0FYR7&?MLU`v!*=S#7iDeDSK~u^_oHE&&un3LEqrcN3~bTHykSL8{1VvPyi`}
z1du{l04WAVkYY##DTYOmQcwaZg(Q$tSOQ6d6eJB%kTgs|%0U^V9FjrGVHu<nR6r^r
z1*8&IK&n9%q#9B|s$mtR7Sup$Aq}J!)<Eh(9i$%8LF!>0q!Bbg8X*Iu5jH@wAOp!l
z3?vIPkY>;XX@*RYX4nL21uc+P$O37FEs%E525E<EkapMx=>#2+PRIf2gdLD>&;{v+
zT##<qy)MhXUwwSh=z(<|M>^u0)OBoqQQuUkn6J63goprAhzcNum;h3Yh#<wN2vUrR
zAf<=|Qi@6-rI-YgMkq)cr66gHf|MgNNI5Ellw&eTC8B^-q6$bQrhrr<Do8b|f>dKF
zNG+m))S?<lEvA9gBRWVus)N*HI!Gg8fHa~8NF!!|WDy3EMHxsIV<63l3DS(3AkCNw
z(u!Cht*8akidi7-hz-(?+92(i4bq7?Af2cK(up}B-G~d)jk+M+n0sB`^yHi6Yj!NU
zl`F67*P|cpeChAC<#}cH=9Bw6PCE6eU8mW{d^VKd*SD(#Du5K208;P^AVn&I6qyK8
z^ot-RDuI-k1XA)#AStCFDWf2%pMsRB3{qw?NZBugRHy<{VG2mauYgpk3Q}b%NY$@`
z)TjnhV;V@!uYuI54pL`2NZqf4G^hd6U<OFTZ-8W!fn<z<WPS$Hq$Ws{nIKKS3DTk#
zNQ+q@Ex!fQrZz~M*&uDd4bq_wNQXHf9lrz8r7lR9xgcG?dtDB@eE!V(ww5iAwjb#f
z=iAG59b8|Q^X0JfobLp7g7<Ca!6$j&9_$S7`v_V{aRTlOSx&%xA<qf8FC=&Y_k|=c
z;J%RI1>6@?oQV5kmJ@Mb%yS~{iwR!DeKE<4xG!dS5%;AOC*i)7<s{sf@|=YGQi7Lo
zUrO>4?n@b7!hM?JDDKlNM{%F#Ig0x<!BgC)NuJ_9&F~cW<rF95zMSP`+?Vs5jQete
zmvLWC@-ptr8D7SHCB-SYuVgs|_mw=S;J%XJ72H>nyn_2mhF5T3O>rvjt65IPeKpUi
zxUVL775CL7uj0O%;Z@w%Qk;hST9(srU(0hE?rRBN!+kBuYq+mvcn$aU6sP09p5=7h
z*Yli?`+9=cabHjJI_~QkUdMeS#TmG7WH|%(jXY=IzLDS!+&7ZEf%`^=H*lY&IEMQy
z%Q4(%d5+;eOYjW$S(0bC&oVs2eKW<GxNl}T6Zg$LXX3t@;7!~&le~%hW`;L$-%4>7
z?ps;T!hI{xS-5W{cnkNfByZuqmEkSiw^N*r`*xPIao^5!HtyRA-o|}9$=kSZXLuX;
zofPNbzLVt~+;{SvgZoZ`cW~cH@(%7h8Q#HtH^sTQ?`Am{_uV|_;=Y^UUEFt*yo>v8
zhIhO6?|JUD=h8o#k8mY*mHjIRb-tzgxBTL%rJtL}d$s@lDG=X3NY~#Jp+E#|u|Py@
zsX!!bS|Ai#E)W@8DG&u)Ef5u3D-aD^FAyEuC=dgi6$ryN3&g~>3dF*;3&h5D3dF&7
z3&ic}nI)cA|8)E1Lmd|;O?)5Yg-H_uTTBxXTS^lNo2CiHmeWMWR?<YlR?|eq*3v}7
z*3(4CHqykvW@*B(%``Ewtu(Q)?KH8moiuT<-86B#y7T04W!FA`S+@7`>P=q`HkX}v
zJ#yQEiBmeRSCRNG;`J&L0b48*5nCz}37ZxP#g>ai##V|%!B&ez#ny^M!`6#L$2N+@
zz-C3lu+1Vdv8^Jpu<atTv7I7uu-zhYyZW^;H_m!?-vdn@7sl%o;Q8ehphT|-C3z(%
z>ZMS!SB6r&3Y6+qp){`srF(TK!)rh>FM~3@CY0s1plq)V<#-(^*Xv%dji(m1O#Jfo
z<?a8%mETz2qx09^-1PX?8D&%MZNIrkNsm6AAJ6TbF!Y-T+2Q@|H&>Qa+MQ>t^HF@+
zwGZ<;&ssZzm$1z2;fH-Uej4fcXR)N;O`VTDf3|m@1yu{ybzJwjZxOt47UBY?7#A_6
zxP(dL6jP4Nm`Ys1RO2e97S}NKxQ=PW4NMkim}cC>wBi<~9k(%^xP$4&-LCwhXZ@c)
zn~`qoxDEl|PdhJZd7Tgtpu~U(B?TlX8lX^eK!#ER3X~d9p|pSor3Z8<BVa(W0E03E
zCX^MhpzMGR<pdlkH{f2cA9o+kPpCRm-|=hbi0>G_9_5IDE#`=bE#-)WO>=}|%Q+%r
zD><TIt2v@#YdNA}>p7xh8#!WNvm9aAW{#NHR*qQMc8=KCPL4R(ZjQKJJ!?wiLmf}M
zrdf|PwSOS)A$4sDZt3pvFFIVUr?&G6>G3==eR5+%dyal|EGg+$dV}}I?&UZ2tmskG
zYk0N4S6NNyX6YwBDhKZzeAken!|v`c-1>8;st^1BZWUEcH!a)keM{edm4Vyu_}Rcg
X#sKBEpWZ8oBf}H;5##O|#f|wt6{Gq8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp.1D02066DCD48924F.idx b/.cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp.1D02066DCD48924F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a59a645844991c13616c159b30285d60ed3908fa
GIT binary patch
literal 1742
zcmZvbdo)yA9LLYhAdJU2VZ040??(@0yiJWzj2V<fjpuj_Wr&fh)|E$G;;KZN2&stj
zmbuj>kES%FLS+!itJjiO$US4NWBKErwa(t_^ZoAMZ~yi=zfE=^65E9UfZ*>COAm;m
z3IG5g1-*=@$ly>E04Rw7Kx{#q8$Ck748vQah)d*N3e8^3BGGz`MW*UF_Rsrknq_^?
za1Cv~_QkyTnw;#RxgCE;^@zo6Y#@utK#+jO>OIa{7@xN_cP-eDF@o_|*V?c#eWo|C
zD}iRL0&SjUPVRIK`uW@^)u6OzZ}L7irN0^Mt7u%l-?%iNEiv@l=lfjl^7h9*Re^rw
zlfQ%O#;3V`$8#5shcC_P2PX)|=X1?Zw1xEH$`eWI6Uw23DwbG}s~v@?oXTKtjACMZ
zI`1a+T1kfQJI5U9M=ihSt`H;G-!E!PF60VY2>VSqH7M3*JJT*3{F^GCpSAznb?%z^
z&x;S<)^l3a;zD?Y$#{+G?0USN4dvAUzCWvkF7PnLgcf987e70`YTo%YlNh8xa}`mn
zn^6nW9+Pa(+~<aw$8!{<uc+toGFsjiOJ|=*dL?YAK9Hmmbs-J77n70h^7XroQgL{e
zp50Fe;^$Vt_4K7f#oPel2a|+x&rIv=iZCSg`X6D(M7x(P)rel_iLp~14oik{Tg*r5
zNJl{(2kA$SP0E%<#1xyP6mqd^R_MK?9{<>6N?K1_hwPbuN?+XGmWNp>JWV%1O>po!
zw6f47av4LIBuW=b>NuBR&#F}?=hGGusn5JP6kEa<wUDOb(NihO@0e582P<*-&hog|
z<PSHd=F4yGXs!#7pc=Bf2BK@X=%s?x91dt%g40ntC!I=D38K0m{!x_A4c`}bl(4l|
zdqJ!Iemc`DGW!=hvg?58qKlnw@c9C(xJgQ3(VYFt0hUD1G$+0OrNQ5&{+VlK33rNT
zTs!gh>R2XP6;BlG9?qS9v$E*jZQ5}6?CxhyJmY>j%?1JS8LGzEm0LNT@`Dq5o3)<E
zbhqc<%yOD4aAJ{4SN|#<p4BjF@~zdV_Z=<oX-j3fO%3TSeBmM`+NOCmBXkj;YWD0)
zMxSp(%CYwtL76Ofy045$S9&7*g(ao27i3z^ChlS@xg@KG*2jl334uJn%M%ZVSNnf>
zG+fn8{?_i|Y+p@B6KguyIpS<(?@WV!qL;_C2gf%-x4yB&U4;9ml_zCzEy6_Oa7Els
z>Du}g>y=sTFUh12&SjrvSkBFu-pP{@czs95I8J}WjK5E0kiD_-#6k|U>qw8V%n+h&
z*tP;-FzLQl(B1QL%?`s)?@$o~NUi`yR#4V`eSn<xr}eIiY*b)aEgI<p*=QlO<oW>H
zkpY1WFC+j6sWxU3icE1RF$!n~TEI~d5kTT5asr)U68%-kOLQ`(@Iy$T7c4{M4bnDI
z2pA2M_#Z5>qKApP2r&v)0xJiXE>(3wA-Zy^m7hr-RDxGSgab7;ktV1EliVlrs`^}6
z8k7l)0nZ~~-Fo@Qm)qmaDj`U3L4d6r2E@Gx@bzXMMSx?Qxe^IhA>l;O`o~D{$!6vv
z!R`&VO_>3fen!1R2vFn=<W1qDz%xR0{V!y+EEiUcnBz`-Obrqm!twa!j-kL=6ugb^
zE{Eo=)JKhc<24F=zhRsb0P^k_EqC*cHWc`3!^jvdaM3^O8NfH@P~gIbaqXZ;WOAzy
zgKr3fGVl)N(}_&ak%CM}Mq|+%AGgKLYVx>hUk4=13&^X(b;D6e(e*Zg>xUL?I^@gf
z{e99R11?Z3Rsbsw$F5>Lss|&Uj`IsJ0(ZlyfXIL_D!<<tk$wy;G@~dQqw{6o8Y`6$
zWPl8JA^`Z&{ljPwIt3m=2?z>p5f(v<iDM)s9HjOWc1Vdj*n!xs(sJ^9&CD$<t*mVf
zWOVTwCR(chr^+fSDdW`Cw{6$d*3mOG+PT}<RArBX{;qv8vi6Q{NTRdt0mQ-o0NgOj
AUH||9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp.DFAE93983C514BC6.idx b/.cache/clangd/index/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp.DFAE93983C514BC6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c97f6357a2e99cddaf715aef834050dddd805165
GIT binary patch
literal 1764
zcmZvbdo+}J7{}kaATzGLF2lH`G%gV#*9Oas%ODx4wC+q2rre3mF1d72%cQchGlbA)
z+;Tk;Imm|MxFuFQ6r-%$Dwk&8an9@bW8ZV$-+Mma@AEwG^Lx(k=W0(TPYMEnb%1?r
zXker-8UO%E=#7XBzkCG+01CnY5L?jY7D|^lk(lC9Ma7(6)!QZ5sn>nVN3VO#76&el
zdt%!b$GucGxb#22q^7zP)J)5j&zrBt(hB1u5H7&i8UrtN39FIX3iE_U34w$HUaxG-
zsFAnKHqE54Kx4`x_a3(v+qXXZPQQz@;J7^5weaPl`()yelL<VVw51!`UpKisIo?(0
zYV|aXd95C+>-V_yzOAZ;?ae%cHY9Q2qaxd{zM*G9osk~jE#KDNeiNM2Tj+VzkagS7
z{|;Tx*dSymxcGkb+d7Q5FE5@_Y-E+votw7Rhg3?@TPt}d<5>inmQZy4l8@5#(8?=3
zu}=Sr`pTNCc0cEDxNQ26PQG;}qz5^enu+6c>L`{QgQU-%WqI_egjn$hV{bk=qC_=k
za`H}b6mM?77|Tt^eDES%NbYd0D~%~;3#*e<&$A3v(`}-wg%T;Gq@%Iip<FX_iS9?*
zIV4Bc-QF}~ONL^g|9X>%-Z(L!aCQyseO0eg5OaH^LDIC1@G>aPhs6A;fWh<mhYea`
zOVn67Hj))58;S`5QJHSB>W1<HO(PySr_@WvCHE8JL)fEcuQDBSgkOjUPjgNrJuN8v
zCAyic9%b5npk$Edg3iQ6bL_QLOESo~q+vnZI8FAjQc<3nsZvbPlSJnCP7Wp~t+m2d
zyKARBtSva0`>cOSaglV@q$T^`B+?nxkcpa?K&X_om?b_VNa@my+qq>{{y^A~x-3@4
zJmb35#j%oB?IycKuO`|N(sm1pa5<%w7(Y*CvP5mS?Fz*-K53?x_Um}q_A2mB@3`8U
zFz-ZV+vu#w5F;}&%ooQ;E59wbFzI1ud9<8+tykE3USY!WW1Zw4Z70I7`V$KVE$Z#E
zgOBoy?obvA=qQ#$n%outp4;u5TW_)(Z)H`HWe>-@mfxtIY_Px>OI2MTNNrFP>|N>K
zYEMjluN3~-@1UEdyP<aF;DoGD+=J_-SIg8^66S;gV@(Q0ENBf0+(xxp33l*&^<O(<
z%VeYbgw&-eqKRT}sRr(~64uKj)~H#*lbsd)WF(8xwe22q&A#&B?*fT6rOmwQ5c2c$
z>wd?ALciVk`tkUhGPdcV^MIE5!{|e$Lo-v69t_4Bqk4L*x0N%M6&yUcL|DzQE|S#z
zF3?wR#RekcLj8!)-HTWF;~D1hHJk|Gat5H{1aN0}7s%OtIxCc9B55IYSmbHQ#tLG^
zcNf?W52QteBLP5AxiQ<CWu$@<p@Jr$Ih+NNHAvn^6wnza%qDXxnPN!gr?`L@VHqM%
zFlZkIgHbSvCKigYnv=B<B2+msIUL-&WL1WI)YFAdekD0j0gi@91w6cu2%siRxLMa7
z)%CJlP$e)1OhLdpev7CTAI_EG>+2|R0|grvTg#sx&&5><LPiGy?A*(McnARw@8`b}
zAZI^6MS@jGxDXWIh6LO9^8gYY++(Y>Mc~_7<OBju#p~dW;5}NqZmQkwTIE5FiK0a{
z;Q9M}NWUaFtun>$mxBUV_xklaVc!D!K5yh3!zl3Go^cHgJSf-Q8sr;YDDdr`0sjes
z**ySR9b3#NGiEBkISPlTc6vqd;}3w+@HQ&B$#EBE3bG*?D~sKmd50Tu-dlOJ8<O#8
zyeeE1&O%~#PY+xh3fi~D{wUI@l<@dzC|3?GCkp58NVwO$r@x%zH=qmZ!==IlLwxyD
zhzR$OkcGY|YSU8E&g+}1)(BU?6`l_O_=g6B1VKolIib)3f<gy`u_B@p;$rraB<rJ+
z82cX%%7RijIb#zu^HW3%OC4z~)59m#mH$tb!7C`L98x`^M$phac3f9a-_S_OKwkTV
RjkJudgBz0UY-NXV{13W9%)S5s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.905D27CAAE2E6FF3.idx b/.cache/clangd/index/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp.905D27CAAE2E6FF3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..74a955f0bd06d69aff8a8cbab432d16c182fc148
GIT binary patch
literal 1766
zcmWIYbaT7M#=zj5;#rZKT9U}Zz`!5~#Kk2=nR(0%3@V%q3>9;ZhUDhow&3~pUsA&S
zy54ry6YXZ++a6pqyOI0C-Em?@bcfDQ4NK4W*ZaHPrk!nLj*{8HU+7||^z(zSqyyW7
z?n6~89Ve|yOD(LZS2X@%@ZVeK-MV}`{Uc|KY%GKKRIBM<PZyu|`n3Hyk;$u#j~0us
z-o4(l@Oxt6=eNr8`~L0~c&-24yKryZ*<))29$n7Ty*7XSe!chm?{vM6zV?Ufhspf%
z16u2AdMibjOgs4I)3NsxTS_k3u?ZJFeSM65=SGcf9(=j`GJCJ?EIhSMDn9Xlm3Ow8
z%FB;W9{w|GX3x~E|5?_n_la|^8EeE&N9jo(Pc;9&pCFf<8Z)!SStakD^TU<XPM)w3
zR$QR_<Bd&DgpnT8eP{8W^0`raD?k5}RY`JRY|Lx2VrjSA*9qk^pV*VnD`jz13x~2w
zao@akulHD6z{wLE+)`Cfo+w+$UVGZLp{?fa<H^=G2FEk9Qo03JhF%PoP&n7OdF9!i
zJc`$9Pnewks6J_FwxiX1x$OIFFaFFE$d+-^?fhLB^soC#qWVdXT#?0HuAZs0Z(S5#
z<;xLMp}`ZyyYiY_)Zw;ti{%ovH8~ruTc7Lkp2TsGKkNPyqXx~n&zs72xn<Q_wKKGu
zlm!T|Es?D@n&m0E%#m3jx@gYwwpq6iUkP$Q{+N$<s`@*7pLGgV9;>Ua6y2Hs>f7pN
z+wxxjxpy#blK0KhE$licm=Z6Gu?lT*oMyq0o3JQz&eF@VN87XPb_;II)A9FNerNHW
zLwd_{JzZaY_A*V@d|<C@byWGu*^A!unfP^OKPp$+xfrRQT;`=0b^N5vfB(Oi+-C-?
z(sE4_&*YZ3a+Cb<K>YR9KR;qWm~UNGZ+z4Hmg$X%_1k!O-y|wuowj`P$;jZ@7Q0Uz
zNWXBTYw-r}cXRSP19LN`ZEw7!Y_RiGrF+cG{`GF+%5_0I>o;H7Zf|UnvcCF!+ji9n
zr!JqX7TR|(kH580(MvBO&MY@BDR#fbiPLwUOO?AV^hhtf5GL~Y<cHI**WR!S+$<?L
zP1j!Rje^qKr+lu-Qa;AYWslXjD_=E<Ja>8e<Z#nHvax-7pLz=`Q>B(}`nmSbqe;_M
zZMc7@oSpjr|IeP9ZLf|_`1q>pe}_VxjABZb_1t;aE3Mu(`7A%Y_oa!p#mCHf|Buxz
z3;muKQaHKJe&%&vasHrhvcDF)<`h@vCfNaV-|mO+jC<F3&*WhU2xJJBV3i056F?a-
zIJ`<utRyXG3pbNLP@0>Kn;%TTq>ED1ii?;S7}(@@^f@myS77D|7Pb_&f$0MBoQ2(x
zc;3Q(Fdo;h1^1f0O@l#N0)(Ss!a!cEa0U`DOSlxq<E@>?vv60lHX~24B%h=N%;|#b
zJEcn3y*dmsNm5t_rW(jo5LQOwsS9htcoI*8{wjT4-UT#CxLmk}5hiZ``{Uac&nnlm
z0mTn73Ll0`0Qr|0g|8s<A2SL+LFTVx5?;>)GYF`DKa=nQWd3<3;R|rSWBVJ1pI=ID
zF|q_p>q(o#ypr&!Bc1KO!ab0?o-qr5Wrmqo$igslrNQqDAc>>Q!k6F@y^rVk>()o6
zf+W5(3;%>m%=Y0dYC9ZP43gjw7KItK`gL<<<ik0AKw)liZg^m=D=yfa&z0%|6y{~&
z)quq>%w3FJ;8f9j==Xv(53Lmrh1k7}m%n}-s9KsuS_!5KW(N}&IFZ0q0~H~8@m#6v
zbrH^+{y@EwERwu1y?+J5Hr>rX@(d^}Y#?k5GwR>o+mAxb-iCpMuP_VWg?q0kH760|
z%HpErVsRkN931d^>5bioPPj4#F$BTf&%ltJo05|O<a($1GPAI<v2$>8^Y9Ar^LYxo
zI;#nCc{&M+3yVldT3A}!*xK1U=!t4uDVu4?V-ppVmXTFZR8mn@*VNL{H!w6dHJ3Ay
U(lv4s6?69rVRG|x^kDP>0F<85R{#J2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_add_instance.cpp.0D22D3DFBDE4264A.idx b/.cache/clangd/index/mk_nk_mn_add_instance.cpp.0D22D3DFBDE4264A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0fe49d8d3b95ed074c19764337924d437e0038a4
GIT binary patch
literal 1924
zcmZvbdo)yg6vq!{JZ8?AF~%?+Gcn2|J;*!q$}nRLVoV&5yrxk}4S6h4k+RlY3ZYA?
zOT{IUS1Bc2b!9CVQBskmh*X3kl{?(~<Nk5aTIYAx`F+28pS?f(cXlxujH5CT<Pge?
ziwKJjLO~FuAikrbh2fb}5Y&Z%ptuWd+=$2*=J1y#ii*kpycd#h917MWhxV&Yn$8wq
zcCOvDGxuoLg}lW%r?!}C{#}h+T>oiHExjGuw5)y7XC-c1nQ?Rp0lC>)L-!hpl*YoJ
z?R)XJVrl&E2Y09!>IK+jk9E{wXWq_rY-@Wr%AQsao}SXvqYid{+|g9rI@ZKLx-fmX
zGcr%qewK8hDbQfybL)qag^@=INeOQGk(5EVl8T^a7Eg0yio4c3T=^@qwX*m5#I6%e
zUf<@l9J~9B$nGj`aeD9o>xpjTgC(I`F)v{LhurYAWr|J^oH!nnYT6@L(PN=<(_S&_
z-MyLd*pwuiON2;!vZ!R#`fVpq%Obp-jxx3=tVmH}4lph!2vTC|xgE+|1{+VE>u(-s
zOPS~g73UVV`(PGQYv>9}Co3|lQK=0AwoGx=7hOl~8vF2*%9-Eo_eSF~p}N={k(2pL
zj|>l<Ggb2&gE8h1Su|0a&$`q>2<}T|*QhtKs;h9a?<emIU5}H$h8hlNd9`@ja2_A)
z+TJgFBef^%de+Wi(T2KWm(m_h!t~kw#u(x=UlrGky)pWRcW&Rf`2;EZMVBepth|q>
zqiU=CRN=RZN;0dkQ`2^H?YI(E>Kjfo+>sg1YW;euKhEWderc)oIp2wQhx&aT#;xXP
z3^Rhp6|S{F?<21<vHbURf8<tOmD1p~iGkV#gkaH9)*dz6h4R=FDUe*B|IlZ5pV45=
z!X)F5GE-y23ZW<%S8%?5eP(Z8jyYZL^$^aNNq;jdadqQ~Q%`A=r{9bd?rm-??Lc+!
z)*qcpn8>sIlbUMash#YM%=hxoN#K(1`?2vultj<+J7(GLJ(bOm%^Nm&Ub4P^((g(1
zE+K<w$*D9tMCrYGMHj^^j7rxj?@4OOfBl5M^-fSNy?#$?L37{ni(V5$+H)Vyz%kFm
zKKU%wqqg`A#E&(U-+bjz8IhJO8|wI~O0s<7bjo)XE6>~A#CXe@LL)0$#y1=9iU$9g
z$HM&i;`g>sk31B#oR6(e=B2!ERzku%SpF`*x_)X?9jhe_Q)ash`tKQxJ_r+Wtf=<6
z2NPv7Xv23BhwCa0(=CVKGgCI`na{jt!qq%m|2M?$$TsO?xwsG3WZVzM1@3GTeoCvz
z?f4R}*6+-tu|*^5ExU&!KL_J3yfuCq@-S&6;F`{h=XOld{9^XPC#PiU$;m*@UvEpG
zsQ8FrQwa!?j{hQSpGet<m2~rg5R5cNeq|B=yDOI!DM+iA?3e~ayM6xIK@`!7ynj`Q
zzb8!~pr`r3&@Le3(RddipKuYFkVc2)V8|251T?`7$f7#i-lF!yAu!|tWE>i|8l7%K
zpf^|B8^90`$Rsq03!-a>L&)uA>{uA$0$Cm{?+WDBGAcQmf1nmC=`L;zD~-jkEI@7-
zSzOHV$diU4ZxBsH6IYWz9cj2WdgikX40(y8$!N0K$^xQ=Vf?7s|Im}ftWZmDS-gCG
zc>6jS!K!1mK?M}`_qiENmpTa;!4NU(fH*)aDYt@jEtZO+dV}u+z+%%f*fo<XN|gmn
z0JGS%jCakXhEn4I6M!u?Ekj;2si0I=GXdmc(=zHc6AeZ40+|4Kv1u9jnu&s<uz?9Q
zFNaydzGlL)aa>RfjL383c_1H*;MsUC=om%_?gSsu8jKLViGiy_Llk!_1_6LDLh>dB
zg3@7xY)-ZVWq=5At;^$GD?<aNvD#4Q5yxjFx#w=;0bz+)HBdUpA%R($J%|;C9SB-J
z2T{V1C}r_e3URxe*eHbx%$M+me8CBUppb~r0|N1@3*S`=g_f3)mBYd~0$!e};OMYH
z0mHN>Dl3vyDCQR1t*m$0*lt$ZXh}EKrTyP7SxsGIou-!ddL2D|gDqPPx0#qxja4@p
R?NlN=vHT<%96M(T_TR9h=<Waj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_default_pipeline_v1_instance.cpp.FC2B955D88E3500C.idx b/.cache/clangd/index/mk_nk_mn_default_pipeline_v1_instance.cpp.FC2B955D88E3500C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..094a38e903d334e64c95a2e9cb1e90e150713b82
GIT binary patch
literal 5692
zcmZvgd2~}{8i#Z5g-)O>?cKk-HBdS!I)xw<jIB@t6x&D}DB^-COD#(T3vCfHDj+Hd
zoE8wZF03wtC<bI1QI=>RETRTTSY%k#1W{H|GEgcoDd)8P{V_eK>3Q-#c}~B(<tD=l
z^7DI%CewhY3T8|$uO8FNWHO~Deygf0CzdlN(>1fnG-FwC_~Z)3o^tw5N&60UuRp!y
z#$wlmL)mT4Z~12Mf#aLDi8FuLv?K5GvFYDlo;!EMn%)nrO?xc&&l%%Zc&jLv>Bh!h
zrI{)Hk8b)ncl(Z%*4y8`6O^VO>ysw^HO{_bS=ZCKOY1IdN<I)jn>K#I{v&Vy{Kdi}
zr_Qb0`rPMR>;G8MVbtx?-{Wf<GW@0M$CVA6a(m2{UoI><wzBT}%4h21Srccrdj73Y
z&gh`0V#$%AnW^#G?3s&ROtyYo7FkkIp{51*Sk8={GU52Yj^y{a^ua`7=U2DSzL{S!
z$Gphe@p@hBTyxp4&d;;H9B8dBtQ);MX~?%@&3$8c^tpS7KmBO=PmAg7E32=4y<pM@
z%ZpMDy<YvS(YwpUy4jakFFdiWC~MS?e>NoD_u2c0-cC}62Qspr+n3X}zUDJ)s<3-U
z(Z-}#KAbS*O8ywmwP4@|$Hajt^DmC=c=OE#rXjx{U1Oa-!%|x?vdhA^9QQ8mICXYr
z&Nq9`SCO$BU)Yy>;J2LD-gvtI{a1`nE7GfH7QcF^!`BP)Zr$JYPVmPu);VIK(~&X1
z^YF>f3@+YxUS9jV%hRiE<)>S0@3ad|DeO^y?fB^R(_bfVbm8VyM0V~kH@WS_tuIyA
zxXxGCrWUULZk1W?adF?$5d#ak8k=ct``SN-k6!!Wz^v)jxt9+<){u8=-hQD=NADwx
zo><V}J~cBEIXZdynVjPbf2@2~zeFv3Z{YFG<@2c%nfGY3d$WgfzYnbS{X8Ih<*A1z
zr~|lg*Xx=8-r)3(tXy()d*yczEx1v<u;{Sa`@*%uYw`xDwKtaJ>cb8`k}_nudvon;
z_S-9Oepp+X*Xe>b|H)X^^S{bM-b<r=kNz2*-v7+Wk8gk1rMvgo=(6o6*<Ksw)hr(Q
zPEqIan~zL<cIfQxs|vCQb*hWl&m=kC=+N`%p2cY|?ib$IrwZz*nr`{mJ{Wdq+!qzI
z#3`Aj|6S3(!rAV%v&*l)*pRgC!Ptplmyc$z^E@$P^%lPtJ9e_?D>+$9&h_k<x$D5y
zr*8DvcJ`SApGI$OxSHSRt#dD5*;DMx-#D>tKDT#Homjf4+hA$Z`}!x-(#w1Jg^^Dl
zp0j)W>J{tOp19L--jePkcF4(3485IJT3_s~dF5o$nseF9igNabf16VAKvrSlly8n#
zTzsUo>`H!icJ%tv`ueX|cj~&6ej(fTm#NA-d2HXrS+{UYUgW*~Q_I?!TtiJIom+Qy
z-hD_+?*8oh*buA7P1>AH7bnKaNy#1WJ~WLh%g0q6Oq{$)X<KUttnOn`+Lf?&Hn-_j
z339SUEyUS5m~l=HX1tw;8Sms_CfEg-2~GiKqFscU=oDcl*(I1sP6=kRU51(Llwn5f
z1ZKoZV5Zm=m?=&LW~yC<nd($wrr9-^X-*Ahx?P8v?$lvs*bSH&PNR9&_U5Y7FFCVI
ztclI+Vn?=YX4cJ40P?WYT8MLTU~+B_OwPlB$-8(kc{dLx@8Q7|TmqPaTL4q=2w;jX
z5lqo7f+>1LFeR4+rsS5ulspocvP%Y2cFSPO9vMvHB483X0h4$Ln2JjQQ*kR`Djo$)
z)un=|x>Yb$j|!&d(!kW*8km|#15<bDVCrriOx>e{X}An94YvWN;W3)c)b;a~X}<Hj
z8V?o1YAuHYYhlxYtn584#1Rfm&ccDoSvfFy!h^|McrbY@52io_Fa?VMreGDo6p09?
zXc55_ts<Bbk-(HJ5}1-z0#hb3n6gC%Q?|-r5<<Wv76K-*5-=5_fT>s%Fcqr;rb<*W
zRf`IyYE{A1hz6!+(ZJNK8kjoK!PG4}n7UO5(;x<zhQ$EWuo}%K(Js7MTj1P6wJS;Q
zkS;g9KlE$3dg-HIyz7e+M^J!bcTp|J2^*%^2y&ESW5_<5T~CAWr`fNNLo|C3m<v!G
z_PH>{VV{dq9QHXM&0(MO(;W7>5Y1tq4^TYz`7p&}pN~>J_IV%8W1siaJofnz&0}8(
zPy+UaFePALh*ARf1s^S7U+~ic_Jt5FU|$SSBKE~FC1PKUQX=+6A1z{E^wT2t#SkrG
zUkXqX_N6c-VPA?;680q@En#2s(-QWj5G`R}4p1`o<uD~<Uyf2T_GKR}V_){uGWO*V
zEn}YqD1v<wrU>>)lp@$CKAK>k_-TTD5~2zAl>nt+UkOtR_LV55U|;dk3icI0tzcgX
z(F*p}0HtDI4O1%i)hMN6U-i){_EkTvVqXo>D)zMirD0zSQyTWQD5YUv^U)giH9xIk
zUklM1_VoazV_y$bI`;J_rDI?B(K_~ZKdob5579dIjR0j}-w0C%_KhfIVBhf32KEg<
zZD8LB(MC(Z|GbN)`fJrEs8%Hz-7@+#UH?Y^{Ob6mGnUis6W6~O!~V&%T+g@|!y)rA
zhDR1+jDRf07!g^DF%q&IV`O9!V+gVmV-#dH#;C|zjM0$w7^5Q_F~(@&vG)~h<HqdV
z+PJU?!~O>s7GXGKKEm+GLWB{J#RwxJOA$swmLrUeOd<?HRw9gotVS3WS&J|lvL0b{
zWFx{DEj+cA>Fu?)hJ%d@3o`6sTv(9dkoh3PBMU)BKo*0Hh%5ye30V#@GBOD=1X&3(
z3bGnxRAep4Xvlhy(UFZHW3+J7AxX8p`%$eX(6DOVPdS#F=(6IhS&x@Dt~buG*KxgZ
zhC}A#438|t839?0Ga|ASXC!1f&dA6l&JbiJ&M3%goKcarIHMu!aYjcr;*8P4?>^k7
za_+9us>X$xSpiNwGY7<*c_6_o0EuQ1NHR-6vRMWqW&)&`6(H5D0%>LqNH^<1hS_MQ
zPcFRVef+0owTT<5WGT7%9xO1t?{A~hj_pl6ozW_zbJI?3S{>c??g#Crotu#0V)XsA
zLC5$uKNL6KqqRzJ-}Kt^V=P~0^}5uvaYH%SK~1T4C7go;;vGDY;1GaBhX^D&Bp}%#
z0}%%SQXC49>QI3+hX$lObRffFG}G0~8JV+s&sx*??|a#o;SQ4Xa)`W_M-;pQqUaS7
zC9i}idu2r8C5VbwK~%jeqUO~Qb+3+Sc#Rgkl`>-EsfyjdG_J+Q_Jmzw<A8V@4<y(G
zAkiiQNj3>cw#h)mMt~HX0;JkhAkC%$={6n6uo=x%Sw3k@<99q&m1R{@;xAKD{oJq%
zlXfHy*I}k%P2c{QOl6awnlvGS&T&JSR;`o%W^R+*u6;_!4h5-$2K+s>ZGnT8I9?|A
z+jDXs?AtG|n{e*~>Af@4{@=5rs_B+aox5aMyWaDU?zSF1v-=pml&t#(3gY0x;dFjc
J|A(lD{|E8ADLnuH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_default_pipeline_v2_instance.cpp.AB0F8B7DC8CEF96B.idx b/.cache/clangd/index/mk_nk_mn_default_pipeline_v2_instance.cpp.AB0F8B7DC8CEF96B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c08f7ebfd7a7e9b7d8f1e6b1b812a3f350342d4c
GIT binary patch
literal 5716
zcmZvfd2o|;9*6V3Z^6*g3cly>J+nYbfi0w*4Pb#l%MrlXvM9LWrq*z7IfEjsMJNtb
zw^S6iLdzj5I4N>ikh;NO<j^5#GF&1qHh8pv3nhy}*>1{A>F1B_Ow)Pteez6y@15My
zIXPnmlWF*@(aYz}UOYX)WHNP(|1Mfw_|j~~WV&lMnU<FaC(K(Q+dEus%1cfecjoY@
zS4O2B`gv<YZ1VM;vu|8-C;4w&nqs+Ezu>D|Yu8TP(|_Ox#`A+4mgjHtE}~qf-wyPh
zlGb6^`Gb20*BtB6re=H7X>r+w0o#SfeEYWYZdbE5y;FZM>C4#l`aZ#E-Pqe#f_1l?
z6R&(a^~x&`hpwwymHylPhKHq7tDZZU(Jiy-$mE9Dp88&2J_uDe-rwAJ2%UBICvE?X
z!_}wi(#ncgoRs}1kH60Tm{E6XSM|m>XUu$SK}KfZxna-Fcb9&1C~@kvrb6e=0mIgu
z+_dh&m$ben<L>Tn1@{hi;BKdFMp4)Nj0EFOQ_6{%dk^nBuyE90x88HzpO)SKoAK7-
zf-wUJJ*ku)at^(BF8ii?&*rKrpT3;5<>J>rk7bg(Un|}GbX0r2@r#R_iW7gBVx3eJ
z$UXG_vYq?erDb<|zOrwp;t@+fdSazBdv(|4_2Er}26s5QE4?&(Y~|ah*JTbZ(`L_%
z9c`a+DQniY>vy%&XAWD+m%Vwuy5m5<_49&t)3PgGEiY@D_V4}du%z9om8By(?(JW_
z(>iO(h6&5l`pIpMo}Vm@Tl-?>?ss2VF?aukL1#9MDQ)-J0MGZ)-m87(dtO`mNlx05
zfhWzo&*i%kHb_gOqkDAPy*@`;9ceq#+kJmT-?4enKz&)khbt@odm@@xx~^vL$n@Yt
z-~1sSPyNSdf7NbeFSH*k`OoU?XFJWkeBg7_JdRHps;n>2U%WQqx!aXnL-V^F+E}^%
zm$H|+7cTCuNk2Mj<&>lOU9uaGWm{ibQ<Jsgm%NI~<dKb|wxx9YxqI>ZXM0W_d0K6H
z=3(wgd&OHzN=Ed&BV;Y!kXZ3quM7KrzkRG?!Tdip<(IxwGiQIsl=<!CizjkUrH`}z
zt<UAUzQ@vzi|2N2-SL*Qu6zMgIc6Q1GxOA&$D?n2U2^1&ja4~f=M}jhtU7q5dPrg>
z*X!lb+UnG{r|+D7aC~LacY5K?8Ba|ZI<a5+zH^tSrreoxYfIq0&8Hf+7-iS4FYn#>
z?stXN*U#U6_2TlK%2$g5%c{2jdE=}wx1noEn_lTP#9h+!V$Pt6^*Lu|yxQaABfHwH
z4>_Kmly|tvNa;NF!xb6Ap&gAwYVK^mBPShwZ+vJ@=6g#^_JprpT|Bw0>|bRceSdLZ
z%@<cUdOTnJYHesJ8a_O?G3m9l8S%5tJ8wo-{H)8LR$Nffxp-20lj~ViUY9mqoR41O
zlSe-%_Ugjo{JDpc=x5^Nq_#;<K6+^x7tYRK^i%xgZEGAU8D2TSBDc?D?d;%|PsPc}
zj&CK-&cTdxaxmlVJj{3}4>Q3oz)WxoFca+}%tWUMGs!N&Oma#vBX$Ba;v_JW?J~?{
zrwlX2uE0!jDlk*+D$G=;3Ny{F!Ax^%Fw^Zi%yg#?GsAAc%y1fyXNmVKt`<4_<yqsK
z*~Pxlx|vxwTL9!?m$VY+;=ttG9GIMk1Cw|0VDfGrOy0wTDYyhM1-AgE;1R$ST_TvG
zTLe?|h+s-C2~5c?fhl<;Fo}zRN!$cX;vrzlE*VVOErTh0WH1$%0;b|tz*Ia6n5s(!
zQ+2CgsvZ?g&82~<xiv5~j|Qgh(!tc-I+(gg2h(sFU>a@%Ov7V5Hak|WmzVhJPBb4X
zgjHJ)2iC%N2C}l<TZtnan4E<Jle2PQ@`MMIxA0){Rvt`&2w(~p0ZhRvfGH9YOwl5O
zDOyD^B_e?-StKwes{|$?1WaNfU=k|<QzkN)vPA|{w#r~CL;+K=C}1j91x%HwV5$}s
zOx3D_sSyoK&7y&+Sv4?qqJybhbTD<R4yHj2Fb#_VreQT6n?$YOTr%2um};MwnvyED
z+#i$&Sv^)1)(gq;BPhnO_ZT=%Vho4O#~2=2h%o}P7-K|aDaJ_1B*qYAImXDyN{msE
z)fl59YcWPc)?<v0Y{VF&m4`gIee)kbdaK6A7Z#w{6IAQM*f7OLkZUP6itMA=DjIx0
z%^pV%(d;>3E<kbE=fV_+eXf?`u+RBu4*Q&+=CIF&Xb$^)fa0;whbbQWd@aRepZC!`
z_IW?eW1kPvJobeEC176&Qv&vdT1vpa;G+fX3w~O_z7V1X?27?P#J(7&MC^;Tl!$%N
zM~m1O{j`XEF+_{lmjaZ8eJM;y*q3T43Hy?dmas4RX$kvMh?cNV0u;eM2~!07q?RJs
zCq9~BpZIBleG;Mx_T>O2V_yzaGWO+KO2)qIqh;*Nep<%99HM3HD*;Nuz7nPs>?^gD
zf_=qDE7({3w1RylL@U@=1C)w=HB714S8FL1`>Kytv9J1R75i$4R<W-IC=L5sn9{JX
z)lwSvH6N{EU-Q!%_O%eLVP6kWI`;K2rDI>OrF87;K3d1V?x%I^>mgdlz7e1d>>FXq
zz`jvS8Q3>`w1IuYPaD`bLbTD^AGYp>>G!*fFH;G5)}Gb@Ef@CTx%X4w4mLE!ulLCb
z$=zC>?5F)u{9@bhEmxbb_fdvz!0UaK;gI<#!y^k(MnD#$jEF2n83~z08G<ZF85voL
zG77R9WmIG>%4o=Xl+lrmC}XtpjNYS<vD5!~q<LWxhW!>77GXGKKEm+GLWB{J#RwxJ
zOA$swCJ}}p%MnIKRw9gotVS3WS&J|lvL0b{WFx{Dt=zIhsl-R$;5G&xRD5`CkmWB4
zrY#@X9-M1lZ;)ZX!Sx0i4w(-!JhBjE1Y|MDh{#fqk&sD{A;@x&k&%@kqadq6Mn%?w
zjE1ZS86DXOGDa)^eQdkJwI`-5YF?O`<y&vVoS6gS%{-7`7Jx*v2qc*$AYvv!vRMXF
z%nFcdR)I9L2Be#HAj51trhWBA-bvTWOXBBUl9=@P(JU|_>zAp<g_H5OtqImnEjzVk
zb#&sR@82yuH!i`ce0E9p1+LwtXPO_~5>k^}K3g~4@=1E%n|+!$l!MJ~Nwv@892^ku
z;DH2(03<p@Aju&C5eEU19Ws#OP=Hj23ZywSAl;z@84lwy-M*EYwyOWCJ<b2Vmt6zT
zU!0di<h?wi;1v)>uZSplB}C#Sh_Y8kRJ;nJ>QxanuZF05bwtB!wBo}K6Ayg1;PW4w
z*J5M)z^<@yK)j6y5^MsHXcK`Xn*>B`1W2~YK#ENPQf(@bX48Oln+{~yjK^duoIQ7X
z^BrPQ;mk#1{4bN2ytZhe|Eq-)s9aNS%e}*7nmKRQ+=4hd^T#m>ZQ4FzZkN<Pxx<qw
zqdR5~@7ghOwBsp}<s_16A2j&stRX{t@;wH2?w@A-zjp;i)hvJL)Fsv0t@|H)*)sa{
Y8=&`<(|eELg^^<>&^hCWjiSc>52s@<H2?qr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_default_pipeline_v2_opt_instance.cpp.38A404E476FFFA7A.idx b/.cache/clangd/index/mk_nk_mn_default_pipeline_v2_opt_instance.cpp.38A404E476FFFA7A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a1b8712efc14f79217c0ba8dd11c6b750626ab0e
GIT binary patch
literal 1768
zcmWIYbaQ*a#=zj5;#rZKT9U}Zz`!5~#Kk2=nQ6=n3@n@s3>9;ZhUDhoHWRq^U-Mkg
z>@(TCKb#-kTD?)taF^Imj|)#cmClIwF3CyXzt7#fOM1J+v;+tHh-bFPtKZFKbzppO
z`clP02bRe5t3G}B=OpqW@t=D9o!I*|Q6FYMttevKQ>~_dJzad->(krQ4?KCrQ{lI9
z*Y8P<yZO6kudl!G@O3=*zJE3QpIm=^HOGhHxqNu`zJDcuQ!D@foECpy=#%l$<3INV
zuZei8yUB`6b<w?z$KD$jJZ!1#u`+s@`*`sUi6_xKe7v`}+|pP!*JSOX@EiN@U3%dw
z@bKf2g8Ik<%?54zN^URmE^%63!s(WC;fPM#1LuGD6DxeSmW7J(X$E|j^O9e4!tK^e
z7C~Rn^yi7rlgc<>$n9FO_jSV0XJ>zxPhhr|nyYwWfskJ2&DDF9r{9;X77BE{^I9c4
zAfoHq7Za}84M8eF+Ed(ECZ5py=5;?vZ~^DNUte6>*}DugyHf;%Gg3b(@@bvw+qCgl
zOxui{_eST7ekq<fwN+tGdH=QTf>!q?AIoS@a9vP0=Rmzd#f{Ddt(PWkndBn4%53{$
z{u8exR(=S`GT|y(yVYP8&-pJ?l0LoMcu=l<isV!kc9#1G_nk~-h_&=*eY?R;qtxm+
z$I=UF3`}l3)=8JTnl6dBNr<fE?fVpxu+ml}>fMWqJDJIEnZL@bg-@_~5^B5X$@lhM
zf7h&Bd)sP%`3CWWPv7j?z#c22dT5!hH_rvPz=sPmlv!sTkF>ZNTC?KWW@*mc<cuV%
zg8n4#-BY%!imyDceZ6N%NB#B}D!mh~yHB%YmAX}MMZS5tGuJMzMJtV6Uu(>K{C{Qq
zD-Cv$P3PuyDCIwN77t`J`SG;M)V@yq%h|QB?)xO2+|4Ha`02C|=Pw%si@Yy;9+b>2
zofRwcfIB0|H?Cun)$`j9QA-c^Y&Cr1o=|i&!#(EcvGB#+g0?HK?2mkR`#q~n*Y&;W
zhi?ltr&K<#eX#Ff9{<iVLryJ))w8CBX1wj5yQgQK=AFz=oB2N-CFA<k5BJ~HEsxHV
zec#?|zU!ckdd>G~9+!phdYJhh<VZZ|&Aq=@&s{ucvFz$n<y6D{qR}yzCLKwCH*d|#
zNB`qxl;hV_8H#OL^(TJD&-%2!);oU7pMRa>XZ`7A$qw_mVydcEsc*!;>8{^WzIfKG
zPp_LRZwg#IAabAKdXD*t;>z44J7BJB^v*C9dG?T#han)4Ay|S{!XHckWx(L@Dmk%|
zw45#6O#VP=ZZ>XyFaeV;N=+-SW@2Dqli$(jywF^Mi6>atQrHHj3&`^p_Ji@bel57y
z>}?v%3=|F!j)n;Xd9lJ7NW3iJQW#HgeWz6Ex>tumS|o*KV8TG2g0M0YPhD6G#*>tJ
zJ+05vdn+SPuq2<P1k8WZzwNZEie8HWg@xZU3x8yWski_A@$HIdmFwAn5{DRt4>Q6f
zfc(ph!dH;_j~RuZAoJHT39n~DQoo-`_y98hJd^MRIN!1T4a3hbCASz^f~EDO&0!96
z4*I1!_2}pSK!@?N@M^+BHsMi6I@^7Pdm#7TV-|i3cW;uB=WgM|eLFxBSD1xw!zBt?
z7-p_C_<aE+ag<s35?lgi6C)QmYI_g;Ua;n&wZd#4&Z4%%am66@9Kxb7Z>$EUlgNj2
z`hddR;@t2cKNMp3GG6}raiFj?i?kBVOqebvE^xYlnGIBg<i&HPuGd94Z~6oEO0r1u
z!t}mMRM@3!^41vS04HHrm{A|!scO`l)h`DLR|>bl!~3s5*rvPrN1lPi4TOzh${C7M
za}q(JR$P=^EDmA>fe0{Q4qldSF63g}|Bo?<AvrfCCj%(#o#xBT!pg?Z!O6|TE5Og^
zDd_5~CdlRKBqS~@A|YvEX>DU`XYZgVs%@ogrXi0_R7_e%RzXooMO9r>OGn?p(Ad;m
W&O}Pr$VF7l-7AF2&Ck(;(FXuG1<p+X

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_interwave_pipeline_v1_instance.cpp.1B80E7B3E985B173.idx b/.cache/clangd/index/mk_nk_mn_interwave_pipeline_v1_instance.cpp.1B80E7B3E985B173.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b9ffd374274f7c787c24863ae5420d254924e7cb
GIT binary patch
literal 5718
zcmZvfd2~}%9>??EEj6}afxPX$H7q57P{Ihs$eyy*lp0#5BS>J7A~h5d+DZc`T9F-O
zTtILGtAG^3D9j++iKawA7=k2ckX0cVMxY2X7%C`}a!%XtAJcQ1o=<+Cd`|A`yDu*V
zxx<F#3no)eac<?b$>kGMOeRyC<ZoH|%x6o8$#mCjGF8488a-|L>Fn13zSrgGveieX
zwz{2j;Z)6MoqFyT>PH;AnA`q{xcuYNDS?-UyOt<BrM2q@4SX_wwsy~)Ys*^<VtzP$
zymI}ZlriDWzg)lO<gUN;hdAMk=li+iuD|<Z<;)G0ck;&j!&%)<tPZ^2rPE)gEV-0b
zH0#6d7q0z&A%6J7^4q^8s<sV0+oS&0?B2?oRW~zgdQX_Q_m91M@0B0kHa+)d{~v1S
zo?{*^nzMAv3g_ZYv$our({kLS6ZPs9d9zR7?wp-kFT9#lo4qxA$+NbRHP`uHx9#jT
z+cKoGpYMx6w;u*eITLnI`!@5j^@Y1rI%JnTE#zEpckgF=zYSyebWLB#ZSrM&l2KIL
zp{~4Y(T0z=-~4;K+&wAf8J_653U7V<Uwd}0T&yiW+UwZ5wpq`f%sDse;I?J`7jLp&
zt9YY#=Wau%&hC@;?#l%otN6z|RE;PW=N#KPI;Es)=YamZYRQbeQ|-s>=f<v?yNC=q
zIV!v8!^5rLtjmdy$^U-kmT~DpU!vqiW&O=L&Sfu{i&8Eh$gS^@n$fq-4)@r$U30nh
z-m<>4oka&mytc5Qdeq9PTVMOOxN<?^l0zF0q|Vt_ePmO|HLd4<^6SVf@}F;fT}w7j
zonNrAY?(W4TI$d*7fl@g!H)I?YfI)cBc?t7)<)J<*Ur1EIyPhB*QXx1JTT91^&PC4
z9W6LB{qnE-HvIFsD|gR*H9lqa_A4bBPjZ>T4c%QWj}04j-n8Ryi{dkc{0iRKTz_?#
z^R%?~o%~(HKYPvd&FMFmMlZcRby&gvd1K4p7<TYV_38mF<WtY@fBoW5E35ofC6}*$
zRPol48~36;D&8V#V=j%j*nax?oCBLo^4*#H)*6BI-{M8Lrg!ok+LNlpeq6QgmHwsk
z&m1pr>!{wEzkMk?ml?UaOWLlFkJe|m-dz69XW0wxn)Ccp-J+WBZhOmneq?+5w?2no
zyi@&h&W+uP(#ofvUfJ#Ey%}fvd{@`*L{_K4EsFmf`8;jwi!)ZP-Zdp{*4XypDKEJf
ztd1SJKIw(SXIGsrU1h%-ddOF;9C<T1D`uP0;bgBC)oUhw<w_5{o;6%7t=4KPI!x{v
z;GY_Ie98VPUwyFU({lrth7O7~<CnGA+U1@Rxck8GUpt@M7`gd){>JyO{8au$pZA5#
zOvmJ_;tSn+Y@b!VR{rJ6p}oa(OGYobXX-Sf*WNPUw5JCo_uY>3lY~Q?^JcX)xks4_
zJGSUJ^8Q0|a{s4xvo4fBQ@SIaeIhwdPfgFb|Ij#|Ir*8gOUdn<YJ9OGXMG>5+_KQ(
zu=HzuRgzqm{AS`D9LzWu2Q%Km!;E+FFcTaC%mkMJGtnW!OmvAblN=JvB$os;>Yy;A
zE($Z*A;U~|$uLtK3d|Ij0yEX2!c29kFw-0w%rut<Gu@%XOn2!pGaLrY442U~OZ#oz
zxmR7i3vJ26?6y4Be3&gBO9_zIQqfGDn*)>caA0y?4ou$7gUNe%FnKQzrr;L96g&c$
zf>!`jbc<k$9uZ8@D}pJxB`_tA1g7Mbz@%;pCiPG-sh5H&yJawCj|`^lmBCcp3YdyV
z0aNiRV5)8vOx2@;sd`l~HMa()=Fz~^yc(FgTL)A3=wRwz9ZbV*fN6LPFb%KKWVYG1
zSg!D2u5CC~s6}l)9V}K$2OyiJYcp|_1Cz6IU~)DNOrG*!@>U*9-o}F|PytNADu5~2
z1TaM^f+<=>Fh!dPrbHz$C94FcWRt+8l!8gE6ijNPV9HbmQ?|-r$~GBHg(_ewRs~GO
zrhutZ6-?Etf~ne6Fg2=ysaZ8JHJb*ePIWMKs}81a)4?>T0j6O!z%*<|lS$Ppt5)Q?
z{=u{?%zP+QYW#c{c;`mlrnqlge)0?oGL~AV`8=^i7)um6##k;Q`&r927W@Ef`4TzI
zT8;s8L59OV7hyQ;b1{a)KIdmS>~jH@!#)>gIqdU6hQ~f1VR-EGF^0!J?`L`J^8uE}
zJ|AXz><d9gz`hV+1ndhjM!>${X9esF0an1i5M~AJi$O-jz8GOd?29o*#J=cfMeK_K
zR>Zy-W<~5vK}N#96k#OnOEE^mzT{^m>`MVw!oCz{CG686L$OaI48=Z;F%<jM&r<Bu
z086n?!z{(V9Asqd%MnJ#z8qs@?8|;u#=ab2W$epgR>r;(WEAWx5k|qj5@QtXD}GkN
zz7k*+>?>hb!M++~RP3t}M#a7wV^r*`epbc48emoIt6^5fz7}LO>}wH5!@d?{H0*1B
zR>Qs)U^VP(VOGPw9%OXv>k&rBz8+(A?CXA3$G#q5b?ob5R>!^(WDM*Z5yrs25n~MO
z8-CWnz7b#z>>FX$XzmYQc-2(@-||yTN}(;w)~E6M_ty)HrlfvvJ=ZFE{fiUJEz*2F
z<Kl!v=HrA%7UD!e7UM)jmf}Q0rg1`%<v5X%l{it5)i_a+wK&m`^*GUyjW{uyc}9w9
z&8PM|#~KzEB9`O0un^&p`4Hieg%A;t#SjsZr4W&jX^2o{IYeY+B}5ctHAGZoEkrbA
zJw$Y5BSeg5ULJeziSF~pOm0|Mf>`QtVF|(^^9jNu3kf11iwPnkO9>(&(*&W&a)QXn
zN`ffJYJ#Z9T7qcEdV=W4MuHg4{BGWZGZ)m3FKbws*}}t#Xy$-;GY=$~1t8HZ0!d~G
zh?*&oY?gr(vjU`=RUpl*0qJHP$S@mCRCDE3-<b37tw`Qrq>JfI_ay$Z=eJK!n>-}>
zbgPtBog2SlHts(r?f&=D<-t4aww&*44UQi0(*)zhfuwCqv9)is8jl5=cK>T?qn(ss
zD<7>GdV+iK<P!~dI4PN}8ee;PqIGxoUUiQ&99E}gXk)5nA?M_Pcqb1eI0YclDFR7O
z35Yr=knEI!6sH2DI#nResR8Lu9msGRO?2a0X6Jd?^ENiTeV=7M+<$RC4w3iqh=NZ*
z6n!G1<dYDok0Q!G8By^mh^kLT)O;GE?$Z$spV5rJw=Vkp?DYNdhPBu&kHEje&H?du
z9!RhYK%!j)lI#)?wNoJ3E(0lc1xU54K$=|x((O8sVK<uSkA0^vjLxnrYIyrmVmXUf
zo+#mv`6%I$g(wk_#V8Sxr6`e*X_Qc8IZ9+?B}x=zHA+-uElM<GJxX+BBT9^BHqD$|
zI<aAAmd%`0CMN%q!V7nl9^PVlizzS_H11iGY0|Xf(vl>)o*79}TBJT;elWddtJWD0
z<+d4`)2U5buCuLZ;UxN~qi?_d0|pMt;vasjLw0B5|2-=xs%CB1zGJ4XOV@5a>^&dp
Z-AC^wcmLB6UKsxPXm(is;1NvT{{RteHDmw)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp.8AF96AFDCF49C20F.idx b/.cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp.8AF96AFDCF49C20F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b5a67557dc6355bff92928c4df07f18462a8ad7
GIT binary patch
literal 2408
zcmZvcc~n!!9><ehz=W{GI|Ctr5TF)CS*ju>tbztZEQADUEd>p0SOP6zA(Ur(9<o#(
z2#+dCDaE3t3dPE9$&)^4M5!7O6tohxP|Bv&gAoysmvBzV{PA*5ZqDcazQ4~*=FS~H
zhs{1jArSmyI9cM@451EzKv=K(re~xjBw+-CrK*-yB?}d&n!D2!D*{8Kim#H&UX}7h
z1Fn?m*1mw>Mw&ZKBR`Wn`#v4cyzw+YKg@n3>%8>`o^x4o)sl1~kI>PyBWM#X@^145
zPiYrTH|NxfjGj5{_APxW&b_*7v%;&qVy2n;oAOWV_+!`aoc`l-(VfxZhK|LbIu>4*
z89jUx`Ak`{xUnnpVq7%;<2MPd6EpJRgB9}!Qx=qt2}g8uD&-!BWQoJf`dpsv<c6eC
zD=&K3sool_yfGnyW|5myr1(zw_1WU6`+?=~O7|P(Np`CA*>QX0`3hZ6a`Ys(-J+!|
zNOZ#KzlDaCrF;JVM!w?l;`pVzZGGJ~*@+7OsT{kevNo2Vui*AQ)=24DvChv0J4Nvx
ztvPcO%O2N$En&x-i$W+Ct+O`qo5zi>mh29ty=3*7!;`kB6~*0mYhl@;yxU}F+k1Id
z8DAbF{Fqr>6#Vp=uVrmY=~llN9QI#5gpg0adaqU<OTIMaKXIVMr>r5FB>d(d$sg(W
zFL>Fo!;iAFrh7RH&e`ic9=Gy7#2q;BO5o)UUT4?^zIg@w+K|$uAM*xcvhoGT24uaa
zpZ(|DjUV1RO<St@RP2OJ_OTp9bxC>rx^yy6U#yqc`^Axf>K0pmrRWv0@OpTkV3+?B
zo5ahtW1ke{-<M4L>}zDQuGMD`@gICU{j&c1w>w)?QiaaaoA(a4tlL_M3(Na(ud_@C
z%P(M|$SPjA|NZA@D&;A=lRxwagPZ3!wf$5i2~R6~Els<*SN~P8-<E`<Ro>ZCf|@hR
zfTg|FMguc_MQyj7{(UZ{WTo!N`P$i#YpeiUx`b-YV(a!ls+hU6^eUp?t^LC1+plvK
zt|Mml?K+0DLc8&k-<Mxwj84AWx#<^^{;QSWm2#)6xYfLK%l|s}XwJ^%a#V|5Thv(n
zfUK}Obo$}e`6qIck!(iM{8&thY?991N*|7DFZk$pnyyLdesPq^&YMNK(i>iajzL`F
zJ(s&pY8jkwoz(UqMdBZ)h(0lS>Cy7Y^RV_)_W6Hb4Gs!u5>wgDz0z_+>4u0BrxA8|
z*i2Yo)R8T19cTAb<p1<OAQ)Xq-D&rJL-sqcrESS)X^#HdIPXDF-CrivL7lY0sgF}x
zj)8&MeIu!}F_CHU0j{o-^W~D89}JL99uj4<yBY}T5^<Eb>fP(PU}c&Sm)t@n1&0s>
zR6VNk>P1yq{ai0{uRak{p=Utgsmf-$X8Tt!nsQofT>4AZYpiG8QR08bjj3lKz}<0A
zjZ{TkJV;Bx9UwKp1xO9LL!<_|aIL2Qx<--9brn!lIvyUb)!`Ai2%$v$u$BN1AW0Ab
zkR(V1ND}fOl7tZuNy0?9mRx`FOY_5x(>>~GFmOxlP&gB}L&zR?&=Oz(Bng-RNdkL-
zBq0MLNyvmq657MHq}iYRWm{$4F-$d$O#CPj87mW@2s5D;8J06AOE}zCqJe;6#4yu5
zR-eV^Pkxw?T*OtfZX(`;*ziFjK7{fZ5g$jnk%V6)X-CJqNcb-(%Sm`Y!l0W7P(K(X
z0@M%2hyeA2MiM~%po;`hKai6E>WAG#i2C6m5u$!LMueyzHj*Ifhg~Fy`k|Zzwf<cN
zvxJ2g8TW}g0t+V#H_iGT`lB#j?;%s6ei-8zK8I=71B_z;WjKyOEk|wRbOPa(4t3)Y
zhX0N>4q*UgID|nhf1*P;ecolUU)?Cf@Y`sk3<D@b83wgnJXRI#`1yfYb)yo)=g~$b
z22h4d3~Kr7eU!BPp2&1{BN;bA{vjD41j%rX>SuB#2Od|IsCFBbPF>s9-J#x72d#&D
zRktm4ENnH4s4<fCSC5|J?!}sh#~w^o2w7=On5e&c0BJ(E4(}O;lzYMIeXYYV)TE|`
z%%2^1JlU?Evmd@kGlGEt_yL5VA6%nlT3FL)>aP>(78iUw(&7RTf-Z24(qfZ^>XR@%
zEjpd9`o#qFj@ILg4HxJ{K7p?}BMF3PaZIvEh1|IJFdbdJbz}<Fz>sEa#96<`|Lyhq
z96y{6pc&&`cMnf5Zy#SLlPxT}otv!x-(6D+%MDB$+qX8_Z+6(~?DEcbS2wF2=8oHT
Ro0tX!hLYGpyY>?I{SOj$wm|>@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp.AD25F4CBEC60DDF2.idx b/.cache/clangd/index/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp.AD25F4CBEC60DDF2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b11e71dac6f6ea798ecd4b59ea87862865f23b53
GIT binary patch
literal 2432
zcmZvcdr(tX9>;HT146>ndk*hFieh+(f}jwf;UNiNP${8kaX|rzG=f;30b#Yp0y=^&
zQdE|*Ac8A65P2*jOnI3_bXTY_f=Y3z(JIkegjD6xY{E>)`J<W1&3x|f`~Ch-&hMTZ
z9>V2{bTMrE!4Oe$LV7F(!!TpzH!VH&%fmQ^St)CxN=Za=iUrGPO2OAR*m<`-I5*h7
z?PUc;9`(F7;q{|vTF>i8F^u=Jl>hu%P!P$q^SWlU*Jn;7sLD<wg<&6B++*yG{2#V9
z_~hO+(#ow=44Gz)d+aw^60oW&9j1KGUy!xZ?#Z7&ag#`ALT4u>Gr#VOoXq>nWUeBx
z@a=Ku4-2y8(*4bQT3sDH72Q#D@(Z$*dyCB-OAF=hJ!JOqFLZHSd&iv_`?8{(L5rTj
z-fxK)t}}P8bQGVBi$9m*>gk>|Dy+Sj`Jj!uKUR^$tM%|dIe5Ns@gCXwr0aa$6O-u6
zgjXHUB`$x1z?EY2PGK+*-_YLJay0nsiuYj)2RPdv>|z!r?(p(9v@C7o`Mn?EyotVA
zo-&myGQ2eL<I7N7-*LLM{8K6Y+oisV^F`F3W0+s$_lLLDXVr>z?K!r4Yus&%0yA%{
z$>VWOY!Jyt&wKkA4PJ3B<%HIJH&p29S4K~Wmv`#9K4KrNdOlAK-EC)7W}SW5VeHkz
zyq#EhfK&a~$}+`)-^7GJt<J8dG|;$VTSqPPV8)pUk^NQ+t<Hy0W;+WGdDdOb%}ElE
zd*3~?qg1!gQ2123<wR5E)o(IyaqTm_26gI21YwjjW|`6uC)>JHT(c8n+CkZl;xX&X
z<=$S_S&7&4s{a^}QcDZF8-koA%h!(d2!*nL-G57e%{j{ID*62c=QG=+$1S(9WB?8P
zEKem`ADWKXGh0(Jd}MuFam}e$Wyiq2iMnp*&fvV5PQiN4QWuB$<;iY#&MSU(jegKl
zaFv0>OUI)6`<tSIhO893m&1ct)#oxw0yn)gVW*#=R`+clZ~QphRh@F=bA_PvLib^@
zYs?W{i-`g59p|0Qzqvi0aqqJ4H66K9arqo?rZNSu*->bIIR4JrUg_x{OKzSnZswY<
z%L%`Bymhj}m+A?cj|~@e*l7>FyT90*m;aM>>YKQ=5q^<dof}6Uo36>ebgcg9Rl9e&
zFV-Z8SXFwyf{w}LPP<kkvGCQ6|1D3<ay>3G3*Jt#S@fZL2ebcL(_*4T9=EKb#l@M|
z9~PHMmLikB3Ta&X53RhG`df;pN!%ay%HuXBCVxEs;g`+xHnh&~cMUuFe4qJA{pj>m
zdQ@52eA$hs6GPq7sbZmU<Slcq;>KlT$3L|0_4$jjwCv<Kw({&1`x>q5YYEt>hlPEH
z@h!A0K3};gD=VLY25XaaK~fuyyjxkOY10f>F6we>f*?&_3&XT+ZkBAX@nC4{@d*~;
zqjo8YN9@uNU;(5CcmSzE7DQ^02Uly<h3ct%-d4V@GA@kRr->t?iA02iM23by7(kRD
z8X!s_1c(xbK|~3oA)<srxT-X6K4p>7G&7)DhB;xSX+<!IbqFyDM-2gUfG7b2AWFal
zh!UDZL<t!XQ9>qMRhm5yU$&KBJ&h}ukwu&&A-%EyiZBam(A-K^UJ}A<C+YDm4J^&n
zQ@31XJLROxOjKcVoOq8TX>uH(4COe~u<zpRn}5E@ZeFW2_LGPKWEsRLi5Nrq8Htdh
z+(agt$(lw)51Hsid6-O$APo9R0L=%ZB!K3FXC#2;gC;US^Fa?8p!r~!4A6YoPl9MZ
z93??CA3h^NG#@sRA({_+$Pmqk!(^z*_b+^neVk8!Orr4Vo9G_ujo%)=VE3J5PNAHK
zK1JU_eVz{dTy#kLDPu}C4=GN}q4SX90A(n}p@!qOgme+HKipIq$8h2aY8=A>%5V&a
z8vcrcUApG7IHEF2aN+@Kl;8kmD8ZqI)&GP<U3p;?_b*mopY~?_JybWfU?3(<m8U~M
z&6t4>KnUu<Rce&wXMZtWS)$wvnkjAd5cEf|U+uRUA5`9^Q|Pwpe(FRrb>&T@?yD3b
z+j~EQGY)hg?N&YR6ia=zcP5r`%h~;nn`&_Z1V<gE#|HrbAs7HxY1t^U<)@U}FH{y6
z!VR&w0ED0mTqP_uAt_ch(X`a~G*jg-&L8`Clv7dH9vxCR7OtK(hQ%izOiENDPp}iG
zXlbv}(WU9>8yOme7<0C7Fs6nCtTiQonYky++viiZuiqx<<h5>_z0IH9HL<j!Gd{6h
iZ^v|S+_>4r)orVXwY!D$mOy9{v?GGd-Q^!l3jIHxny_jB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.EB9B1319190E914D.idx b/.cache/clangd/index/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp.EB9B1319190E914D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b2d7fddb1b59ec4b013aa01241194a5c4d9ca2c1
GIT binary patch
literal 2434
zcmZvcdsGwG7RDzT-Vh$nIfMivfj~h^BA55;LID$q!GZ*>f>05{BNAu;K|rOqT2vHj
zVU;cvd=!NOEwR-?sYMh7K^CC$s!CA=Y2_iJSCJxe6V?hjf81FsGwXNux4*q-pOXpY
z^LW|X80H_#-zO2JN2_5NW+eZmrKct);}~WokL@dK3zMWgbklpYOfzvE;xwpt9B~oU
zKOS`XL-Kck@7D<tN9`7D-TAXaU**rlb?4#XW|s`48@(+TUTmg)r1AK8OLnzyQB9mk
zx-Zwl@vZ%`!0vfniuZ<_-Kk#g!ryZ~*f11NEt%@_=>f&Hj%{hw+N(0YbUs@8YKCQg
zZK=s{YQwxh+7#8@{)6F@J~4antgM+c+dP~<6*2fu=dIJ(%<JoA`wOz*g_0Z7FWYCo
zKAe8fTZ2e_I@PY(_>*mYpuXf<V!=S8^maY{)9B@F!5$Z@XLIAD?;Q?lCbDxEG72~^
zwaZ=9w>SEl76p#4Tbez&Hb%6w^sv8G@`(RvbxCK38)0#d{dU?bX}bf5{K%hrBC|Za
zDeKjHhSlzX3P+05_ZN@zUZ2b~d#M@IYbn;sA-1a1b?-kIDQM3N>Fl_~6Ipk5WSrML
z(8c{cZ~x3#v8R{)tpsuGal`7ZcZDX*p2Ex3-Hmz{g9kdCy5?+(F7EO5m|eT)k;dQe
z&KT}7^J5=dl!m@PJ|1n;87Ki2`CPu}^n<%Z&1S8g`)u{XDb<6#@YcL|w<?nZ`;#sW
zdG?$T6lvWs5I?%$@cFv(zrM)$jVC_faRmF)DI>&C;{qec;S`@%>5Cgiq?T3XoqE{X
z9~y9LOrP?N&2yiiI%alYZFYa^@Y%^XwUzbBQ}0G@L=_3{@4cdF*Flc%Gge=7#ka%_
zlkEB~vFu`BWZMz3_f?~xlI?<ns)j3uTR4>xK6mofCg+%SkA2u4ZLIO`yMnW112$uh
zmF2zB!Md|@6DRyur`dAw(!GGvkec<}-PA-~a}OTv#bfH!z`M7RFI<1A$#uLhc;NiU
zc3Hh1WjdNQP*PRgxkGr`?P|yM_&YcAD=rD1m!%xrB1tG|_|H9-edFz{fSsjZ$#_)O
z!O+Is%l!>Lj&8BCoZh?!>yx+pdUDoWyOC`0rPN{*XIGSqBx-kLvs*{k@L%*y-ub|I
z>78w0tn)?J)Q_J<Y`IJ`Dq;H=PctoNp6YXB=s}LGj4_)A)_~KFp1yC3KXLw@5n0H2
zSs=|4(J%hI@Z<24qHnFebQfd0zgb>hII+KevhC#DWd8C|W|Wyltk|>s%utrc%ojmb
ztxeCIK6IN)JhR+>uyX5sa+tLEpik)#g=!G`hVhS#o0OI%+3hX=_j!S6%YI|-Q9UeV
z3nrwi(?40c$s;RQYt35Y^thxey5wMaSXV>WVCAL^r;6gzDC!udVg75O|9MxYhMtgc
zC%lwdauN_9s|dIQBnAWki9vUW#GnAK#&rHUH*!eeEYz0Ug%A-cJ0g-uKuApNRS^gQ
zh!R8sL<z(IQNj?2C}AW-lu!&;l}55-^u2YHt%^2igqf-o!6aA+*%0eh1ZV(J0wzF|
zzy=^nNP~zHG9jXbHgHvG@-%eG@^w`{E^i}~I7~unWdamoCR8EKY^tP?FZhL|Cp6VJ
zHBs)m@=4NJ+4X{!#dAN5dNO?$=Vrd2$dav**JvRTtw<4~k3{sNJVqkMQLZBsGP0^E
zqM1xwN4b|w+(8($kO1lr`bYrv2V*3F`hz+$K>a~88KC~4mkdyU*g}G+KkOqx)E|zK
zAnFh6$Po31&18uB!(K8}`THK3#uomY{(z(=WN;X+%5m;`ax`A!5pzVbM-w>l8duE&
zOyB@zIDtbIOVzN_YWu}Iio`aYxQ8aT;Q(dWhC>w>jFknmb9aaoiSszIfF{o40A)Cj
zLlvJ6(oW55jY?A_Y7r1=R0|*kwcsk%P94hHF<MqAKTEn)-PJ=|m-gA^6dln<`7%X~
zVypZ_w1jy-`_z2smVB9^#;{b5NLh%iv+_Yvo;<k-nPbmh?hw$vAFMdFYNixr?qoFc
zH@4G^qoQX&f~&OB69PYg5cGqqv}CyL=RZ?!PAF3B2}dNw9v}qm;VQjvdiW&FWoDZq
zr5`8$L`T|>1C(Ju4poe$ijtxgub-A0lSY+4apCB|@J#us<O)(K7OET!hQ&x?lM>`8
zh}(>-scUFzY3u4y3=Q=8MqK|7jCA;ZYp4V;p}Dzxdi~qm=OYgM(1Yb-YySUMjZMuM
mObbgZYnyfJ?Kar|%hB0&trMN?un`&uYziauKK2bH1^o}p%BZOT

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/multi_index.hpp.46D763C8845699EA.idx b/.cache/clangd/index/multi_index.hpp.46D763C8845699EA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..599eeeb37f4ee839c85a375b7b68db9bbd85052f
GIT binary patch
literal 244
zcmWIYbaQ*dz`)>~;#rZKT9U{DWD5duaY<2T1(4QaWMHV6v-Yev-(dwF*Nba3=A|1?
z<JLda^^?Onx-33)Zt0w#S9jiC+1A_<baufLp`%|t*Y3Lz*>$6+l*@Gct4{_WTH_Ly
zsNU#VP;gNu%j%l>sjUS~E?;c7Zo8ZQJkvSS`^fLKo9#TjN!#jdl)T?>`ceOrQTISm
zOL1jx63}{vqSUlvaUlJF{`S6_``%29JWT8??B*aTFepmRNdz&9i;|0#K@4UP0Rcd#
T!vNFq1tK=OYKk+M7?~LWd?ix7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/multi_index_transform.hpp.97E44055DBD0AE25.idx b/.cache/clangd/index/multi_index_transform.hpp.97E44055DBD0AE25.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f23b7750e27c8d61f45965b198c3a9621572bf8e
GIT binary patch
literal 47022
zcmeFac~lig^Eb}S^u2eso{NYeASx)r6}$oWh)Y~>O`?fQl$a-*$;9lM<cZNN?i-4t
zQBiT<5f|KW#}#o$R76o+a3$`FyYa2dOx@}H^*i75JMVeF=Q&T#$*uFTik_aH>h9|5
z>gsnpv~9aM%g3j6pAG}Q`RvD!iI0y@1^EAmAN%)x&F<r~(%;8tK-!ARAADETu*=cE
zE44Ys9lkJO?~Q+{OWv6I@7Akp)tS0*P@#Ru)FTZmjPY-O;O)eR*>{rX*G!yx+<hY{
zqWE#(owt+iH|LB9Slws(nW4!E1NXh%tC#=ZFC(WeUw5cuy{fDKa(>t1P+`Oy2R8Nz
ze{J2F3-R0gOdPiEtQ;}DLS63jQ(gMiUfiX}2)~g#Kj895)45x}`1{4ddxwf^48OLc
zc=@eC6Qj2@{P4rjR=<U=x%S~=j-UFdMrYx2hol!>GYYszDXp?+-pVXpSvO@zNKfhX
zPdh#?-v5y6)vR*fwYnp=G&_+nC3oesnU@AH4Swf`m_H8`jvsR4V542LJ1Nn5decq^
z4t@4`<_mi1&7J*SenTSG&M6J4y-f)CZRnobzeP0P-TLjD?fT|E`ysy1ktNIDe<y8b
z=cW~JoDQB;^vYH82G{i6C(a7_Q@aI@K6&EcS*3pH(p4G0rAf8BbX^j@@XL-npZ~CN
ztoD6b!P`DL=M(IiJ&%MeK0QvpvbREfvvU{IFHUi-`h4Z+f1hkvsq)aGOMm^e;SqPH
zal)PZwuL9}H2wbbQ={HH_FeFWdQbK>9rxhQN3R`xu{(cf@z}J)C6j1jug!w?b)#cX
zzx}jFr$;yWhw|*W<(ukvTJ%8f`bCP8^j1(&!>o+@AN8!=aOlEMmUmt@dBW*Ui|pTi
z-!N*}H|NuL@}nDm7BwcPU0V0C(XQUYmMLE*`NeeV_OH?El?9n^hkvZxAA4wqHhtHS
zd0D%v`>$K8F3J5Udd;|bQ|a|V8!p^9-uuAQjx}ra`lR{8rT323o|^vUGig)s++17m
z?)!Dueo`~adE|nB-Gd!I9$Y!?-lTPh?c+ixoNjQqY5vaIq2H+SHTrW)m&xaPOzAbY
z;8KHK%ce)I$UKy4AC{QfGrvm0#hOvJZqwhDJBHobUw^|9_s6v8ZU)))wJp(IGye6e
zb%vLHuO5<f@6T|ydu?X(KJF`P3odqFaJ50BH)|&qw7yVbVe3XWoBj4t#$$SX)P`}r
zzfBQ;oVPPyKGyH@mhe$i=ZzfMxyy=q4g5RaFYR;a_tr%pbXXS=oAD@U&Y`hOGA9-O
zEBW+?Zy$2^TJd>h(rak+Z}-OzJ(qJPVfyfwXI{%_I6cuZC^JGHwS4Q{!Uex~kLq6g
z^+~HQT|WAx@2NL-%OO*4`!|<^e*CKH{$9J*Z?AT-@46kAy1hR&&*zL@yeanfnl{(+
z2XFX$XX8$BZKf~(Ri>ZsXygv;(EUuu-pb>&2eU^0`t9Fy7vDcKyL+80!OJQ<`B(c3
z&o)15e*buW=H}0zH7UA3=!cyZCjKjWl-i`M>eH}REuQ~%t!e7^S?fRg<MXT^%7%XV
zvd+nk`GNk8?zgE{ePd$MBKrM%(zID6jjsGOCV2Y#4_E$r>2SZB_d~bTe9T|F`0oqS
z;>;~u`~CKnbLXI6%39uT_GYzL9fpnX*Ribi&ri?3H|CE+GhZ&6b$MFjcT#pX`)=&$
zH7~zkyd}Em_p2@Xf4cHi;*{k-J!$SNST$k5gs;0Lefh6e>(3XI^w?V3sE>MT=f9P$
zDej(S_0D#wdgf^Gnn#DbUw$j)_O~rQc;nYYgFZ_7t+4)=cRud6;)9Jpt^e-9_m#eS
zF!k(_&u{Om)BEpjiRmxql}7%3Ve<9Pzb)FHR2uN9W6OIPTb`U7n$d5i{pGy%5lelY
z%j(^k;2!n&IsNFI8I_;@kn<qnK*`yl%w2KGA71j!nl<=lhtk6*<1YmNk|O1G_Wfk(
zPTxOP|CMvhuPFcPy3;edYm?VsbT#^<UcW&f9f~-5dwgugGnJ<Fd;X%%7t-lAxizD^
zrq;}=n6#kfjmEagU%b;L;qA0-nbj7T`BYw&bjh)FcbU85quHbTKkQGN+FlO(%<(9G
zfB3efetv7Z2H5Z1`EcadO{;b~^=I3Cuhz^7sxUPnFz$<$i)LQDxX-Trkk~v(JhP$i
z7sqR@>lT#0d(+ovs}9<;veQ|&_}8aR`V_5iU$AaTePx^b%IK-ToITsO#!P$fY2S|c
z>(l!ma$zMMcBCf7zhByA?Sl)0w+{U(dcf3kqpmg{X&-h!K5@a44@S4k+_K#Fv%+sL
z4PEkg4OjB@VPoo6UjO0S7q^^lMz+}wX0N{<R(r{SX*)i$J$z$(=Yh^;?_GRVW88w$
z?B<I;tdcNeU_tV}k!N~ro%?#XYTt7c7OeT=H?27O{YEPn?6nU{tn$*aGvoIUeSY}m
zn@@g-gbCypcQF3rH*fp$Z}s-k@5!O}<T@tt5rGZ=ac8&W&P!Rq`SNXg_~?h#(8FrD
zNl=8RX-8-N_Pea{gFf)l3pA}j3o?mUq{vP9cCSy=_<w`QNliVe`I`hqgtzkV-_9I8
zAb{`ssgJ&phA*ViCP5K0YQ(lKeTHiq|1H#Is_LAo2ATv#>~{aw&i6aVLUrLmsm{<;
zXShjF#33L5s!oG%9{KUryZh*qg}BK=JCk@t;(APfwdv$0DCI*Ry<2Fv@bVuodwZ4Y
zkL;r6$Gdv^=+8L!Gp>b6P(&QrL|@tN>|X(VYbYTgJfKeb53-d(Jy*8~`Lzp_@Da2J
z?>@=5FbRsNAs@uPyDg@p#(&?-N53gJZwetMK@l1;qsPW;4F@ZHCwST8ntWWVXc83B
zh8)~e|Fw42;R{2nX4BAYTE`?_k%xsFeivpo_2b|AmybT4t2v&lZxR&o9oMhf!a*Ny
zg!X_JPUb_C`EZk<h@bZTp4)5B{HgHupt>c3Tq5|J#4ECF(2CH14Vet3B>Lzhq|gyk
zxJgjN|54(Y+*jEH6#ku0eDw8-yk7A)35rx2dwomR>!JNM{)hK{^o>%*jZ#gMph)13
zDRr{8=Y8VGN4@8xCsB72ZDkS^35xqf&aQvq3qQUZG~IbE?!4B{Bq&nlMfk?h4c7)j
zbwN+^gwQ-8+$1Ox+Bo&+t970vgDQROqmR_$MnWkDL6NG>yEHueXSHwr_}cKLPx0<k
zd~cI@MV^fPe#7o+Eua+8mC(4*cI7|Ft`_z0I>MbTP}N{4VFT~pz_&DsS7cN8rId}&
zZ$J+Ok$Jo}j}I~lic}9v)xJ7<tb&~n>E@%q;vKJe)g(Srpz<a3be#*^+y9qCZPcBh
zme49AobD0M7AC=R!nO624jx>bQi*R4&A(6zT?ox@5U)s|*FW7Al@kl90HexQS>7u9
zn*>E_L>#*N!?YIl0{FI2;U#?d627)cydrNkX+Ca2&%fa#f=CJPF5z351Vw5F#-^;E
znl%ur3k}!O-Lggb53(cr1_@e&Lw|!Z!3P{J$iqP;4dNBKUY2+C<j9Xe4M5*=l(-zF
zok>t6vK{GtdiPg{1NfHEJ-!M5CZhbu%l_=rJ8SzNkm2LO%NEFS1#&x+ctsX&?YKW=
zQ+tj737TxPRCBY`!6YaW)#-5UtHqyJfSwE;Y#%4@<0_g2MO?|>&3b=#(9h5wATpYF
zkB0Uz2#VC6QSzeo7Y&95@R6YR#cFu58f6j`i9Q<NJbhKhF@=8<UN}pUvy?!Sph!$)
zk7Jb%9hwBq2HkXmC{GamO@bosUw2OXYf-<OzI-s$rHG3x;+mMmD>CEh-xEhR__h+?
z0vdma5IRJtVG<OH`*MCpQA|X!F-H1_``6JuKge#-ZqS+^tGfJr`OYw+EU|?yvBjDM
zMH)GTils~Me&EZ8LE{(7?n1elNxULq&3`#Q_nX$xi=j;>($I-C+$3HRy~{DTqa+DL
zpcik~Lbq$-CP9(cV}H+mb7uPG0KN*mO_C6vBt)9TE7EjM)ee<A=*IjtpN7we`O6?)
zku%%=Dt*4Rk)2O~J~)GS%;23S@sWIOp4nFCUO#*5u{|!NCNw{c*fXg-llq$k%W2YX
z_TncuWgi%^p=<2nYwqFen*>Fg{@OF~vj@ZZ0KO_rl{3`v8K4RVL6K&AY8Q4V`7oOf
zg6}*>)aHmmCP9%FJ#UbW_Xj=F_#Z)YH%YEdQdg6p$QymT_bwdz<?jJ}9Z-csT=*f*
zWfBx=xg+dVpG#|s0{D3NO8e#T{c<CdctxHc9{1JyAG_K4rqD7+HOEnnn#4z{Byr^}
z(<*rDplP+oRE8b^AK`%DJ|MI*36|6P$JN;{ck$z4=mC*of;<emyg^W;P05X~+MJ=Q
z0{D*51G1IyY{g{~6iJA^wxasJ^h)pv;GOnz@?Os0BwmqKC+=5Wf9ny5z|f<~V9t1c
zklk*mywNA8Mh6&IpdKkgTngA220@XItJZF9zH|9}7&t(^UkY(ALBt>^(rL(7a}V}Z
zM?ont^bCj_(60Ok*>B35m+Q~2oq|CF^!5RjAJB>>L6NuA_GhLYKm8hrfTqmh<vDys
zlc31kU&&)1XYZ*B4}p3-7vi2nJq&^(okw<VI=*e~Z6E><8m`3+hiT3rUXfF0R*Y=+
zAkLTX0;R0ChpxB3W)c+X64Ump*6r>zg>D9FFi#1cr-YjXMc$3O`f=A|)B8bxhEiS$
z?pH!HlXyj1e-pWP=kIG_=z;!Rsz#QoO-zC!UB~yl{CV?PBG^2jwYR0v+fp5qph!2r
zQ6taQTvZ6F1f8syx{JX=Fo;)V{^<p8{W<O+h`>u{Y3?kol}S*fhtKxSIop0e2z7x{
zGHGNcZDJCyNZX9iT^rh`fe5_OP!c(mG%*Q^^ql#xvYXpxUIGgN9+aWEGqm0&L6Keo
zU0pv-%0Fs!vLjsh5$I$FL6O7?H-9^&Ol$+y1zlOmxmR+nOyU)(S+CXeXKxol>w@MV
zP}~QUW+p+A_s@Si$an4g<Do8K0E}|FM>%_&1VwuHe%d7G;ev*sxA35WaRb|x|9IIS
zU7LAhY*8mD0UAG9b|=fdP2v^ldu(caC*PVN0^REc=YGNUHi=hcZD4%e2Q@dr2nS#J
zkxCw^{w6_@4`y?_Ydp9*9o`0-ZM5bdtu-?Viu|kPh3;cYyHXep;QP#=@(dVC4T2&c
ze!q9$>5J+OP)cZ%QEKQYSl$=}MLzC&zJ?lo>OcVB4BF(37Jf#HG6{-&5_94EJ;%4c
z=Eq0FS1xdn0!NHVydpm}U#>bISJC)^P~FFrKBj>t@rpc|G1mF^Fi`C<sP0PYUP+so
z1VuiLIo5a3@p(L0JTR<pR^6M`7AEnE?EbdG2d#dp3-&Q|vmCi%j$G3uDALE7aW{75
zjK?tY!w1;UY5Tb#lc31wa~n6Vey-O00emBv92SV-3&ht=f+G4iV@3{=zKelU;7f0j
zv@KGQNl@gA78@F6udH%K;|GAQd{yzQn&m&p?)%tYHsj-@TMF-nhx{qZe~PtDf+Am4
z`SVp|<87b9&;wOH!Nr~6+L^>F5?Q~Q|E7XJK_Q`(HDcTvv7Jd!<m>fOC4p6QK7(Np
z+GLk3?~*H;#4A$p#moL(9xPS(y6`q@C3&snG6{-&Tla&nC&qGt@O|Jx>qPfDv6)Fw
z<hunU+g<pj$7ZMte8LPGmqFW^#4ED4=98_d@u2<Dpiaq7lI)BziC5(C*{^4YE}ssQ
z3pCqj@@M|#KgjO)P1w{Xe;-{2Zv$*y)+?6(c-aM2GAG*G^n<?!O}j_sd!Y3OL6QDi
z=i0~jcRLItB-G)d7Wz=DV-ghkp{!S2M!zaG6h0Q(W0@o`lPa48MSkpaWZSf9t&2bz
zVUpV_xVH)|OyU*kf2zuh-F<zbs?d)MRk={DXcDi;ituXhFYFTvb%9pR6GQXFIwnDp
zpT5g`8apl{H-LW~>XI#mXG`%W@rn#s{CvaglL64G&@pzaWVae)60gXb#hFbz9)<}w
z8mhZjAbW)vlc31IQ|f02-aT;#Mr(M`Nh$KA)Wjqx^7AEe*sPVmeP!n(z)~qt9R(^i
ziH}rStyONR{QN95;O$oGSx^Ib?JH{B6}6p7u$+O)^v(me&kcZq2ZoFyM`)2F+$3I+
znZMNdG5;4g=swsOA#ow?%72hODCF&wCi@p%h7lhc?vdzzB(^e%SLE#M7CSyW5D6a<
z9&%NXuL_k-f+D}`i))rd26fW-Uts{;qWEorS=k_7kqhpGZ>9$wgN_a&n-y)d5@Zq-
z8LT}@>bmaH7#OcX=ca2lr)%|1f+9oOo~auVd+emf{{~%SE_KeOAtv#POz^QgB91Nh
z<$vzuqu+8w-g3NV5)>J_X;Z6y-CphluLyjhEXkQAg_y)EvUh+xIBv`^_=xbGQ-#P>
zp@~UQWLTg&wqC@gW-tSR$acZKUFdBRugL5HD^u=o=%?|YgUBeM!YHAtNl;|?vK~>9
z1G<8*>F3X2NOwd&hUyvwMMhK${c`ZePZojx!w0w_$Tx(FCh>~2T>V|q)VV957lE$4
zRNXIONH++IjE?X0P1ED|KDG0qpmTq!jz2-?4B{jCdmMGJZW!G?{qwe~x~_rOhEI^B
zMJ8!YOoHW%{r+J7qtT1shsJ|paIfs#D~Ff_MUpE0?A+0*md7vxk;#gCveL>VD3aV~
z@xjP0XZt{hg_cPdLeqshCP9%22d^~iQF&WSUp^Yl&SIVv^D!ntk%>KCB)yOq%!b|w
z+B;YxgQXagct!f$?$bJ=V1Y657HZ*zF!371D{{VXYX68r=xEiT@k-RV66k0KL6MX@
zo1UI{GzZ4px1blr){L!R{)6l(bK308+I#F%=!c+|B{Z}I)`tc`k!kgBz58wQ)16>O
zfX*$J<i(P|NxUK*eyQ9?Uv<xL5_}i_U4-uW@v?8$lXevk0uN6$DB*|{cSLGu5)?`8
z+BB~FSDl-{kO7@6SBc97PlQ3dB41~Gn{;?`70_DvHpz-RSpn0koOngfhjb1ZM6ZLs
z!OQx}z7@-VkUjI?y^jz0rdwab?XrjuUj%L!gP=&-*PWh>`nq{Ad^(tLhw)?>d^&@m
z$gFq#mis(8JP9leQ0iMk#alv6lc31#5!1vT7rut6AsFfr?hbER{)6nfufG%X`one5
zbKZgyvK*mVjyfhmk$Fk$ibqC$EWpdcmwlzWU#Y!Kf+7p1HLd9%A9B~ucYtoOT5+sa
zG?Vy90kY987F-+C@6NL9Kqv>kZ$5SB)7~b*au&`zGyLMVyFH=lL0dEVxJ<sCNl;`_
zr&~XL({289XjjnI6j4qQE1CpF(t9*`?ylML1w07eDc9l7b@Vm~iY#xN?-zN$C=ogr
zv`S}p=T_xEUUt_ORqFgU><4IinAewZWC<5z60gX`u<)Sz6>mdT;p-G|aRp!w83aW#
z<|n+~Z&=TG=veT=rzts2V@%=|$qIVB_qR<VC^Woq8ZW2u{wDE?92+*V<?$Nev+4&g
zJWGt2C3Z9kimX^RwtD@eUyK7g1B{4#P0oi!ut88{Wvl90>)sj{0(AkspCyx7a*Rp5
zA}h}11U{PB9GVUEzD#kKfl?a;MON+l&>2#1(%WEn!Alo%?m`Z{JmmyMR^M5aa5Us2
z0__5Ixv0fmgcYhmP$V;8?vTshHrNd!U?I#_<7UH9VGtBq^V-1l8eLmJH;snx(~)$H
zDgW`Z+YETxzF&M43>DC9rF>{9c)<<g6`6Nw&o9aL2{2W{C+w^B4J!ZfviF|qzGz76
zPVgpB!dl+FmTzVf6j}S(u{XA#I?@6(4L;v2D$jyNy+OPpOW(b)L;VH(DIRlZn-snc
z%prrI$hrd&XBt)w$~0Uq2j%dC;Bqktifp(N+GEFGie~2<Lhl(XI>w4sP2wX}<ZF~q
zDw}_}^I5&W|NJ_j+@p9_4L=KRB!gf%n>WrL_R*UA)1X!0mDh8j>$z}~pvblumH*>|
zIf<Yt@Y=I@c^2rKLA)Xbm+y9oPEG)o1T|Qv$m_uIZx9sO{!{LyQ=fE)5g`h`@nola
zvJ<AJa^e+9dvjx@(G5o$zMXAK_%`tE7{n{GxU<@(RK5#!fo^(PkS~LW$RH?^UD#>U
z`NzjR3v$r?d!qZE*xMv1vLh<z)}yzU6vB6g(dmq;ol%2K;uYz6E4t^x-@x%045eg=
zky&CBlXyk?hz~N7?nnCaZ$n*DNoXpGHHlYb<IbORzuj028UwBStNYhx<v+;Y*{jX)
zlZTh|Q25v2A*rIADn^*ZEAmP7gdsh%gTcOlmVKbOA1EzM;uX13bnELe8_t7IA4(Y~
zhK>_!n#3z|JND;0O<#Dt@leVT)jdRQWfBzGoiKc2-WN0rCOIf&wHUV=)?o(miuByr
zrPk$~*6@`<Bw3V`!L@7<ugE7GN_Ne;dQ{=#K)F{*@+!$~60gXU)h)WdU-3(5S?I|N
z)W`+k0x}4S?CDh@DDj&yee8UFQ10`B<2-Dh8N{R9HOrNI&#J>dqk9hh4B8e9nJI#M
z3V53gg5~UO`O)ph>gs`p8#Pr8PX+xqh*xB<+GBdNN1*><P~m*Zoi8;r35w*rMHfFi
zTov2{VNl_@s(Y^5%p_is(Gw~yntiu6C=|@?`($mO9Apv{IneKmp2@FHErrnuKEVrx
zyij6Hf+7b$+}rr4E5rI3tM++f_&iv(8^kN}?-QHei*MT>z7jm>gdBGQ_A3nH6)8%T
zXDJtCPzO-Ui$dr{nBNV8B8T3bJ|=v`fCW$r^ypQ5+$xwd4dNA9n>M=AG;W8T{|Y`q
zsl!p~NHB@#6Bx7ep|Zg0%B7m%iTDPF_5+;X0T?_Cg5?~k)uQqPp<x?4{|1!viQ^Nw
z{KrRf3Pu@6ewn@W`J&AgVE}>;FkcRxFNd4NTgI->%Q{tVJPdj?^rE?(doFmK4B{0z
z->Kx=m>TcH8^O4nPUF&H<!%rZ$@_5Dl+T}h1j^7JEbQkr{5g#@35p!+J$m)c>?#td
zCOqhm$UmBt{~-JL(24h3%^F%8>>l`p*Ok!gN*$A+NWm}P4}Lyn2h2e(c$;yudz{?D
zBwms9tSJv#)qHB_+d+%mrjFZGHHnW@0m8vN%Yjq!^~g8H8C{IZKBeJLp|S?Sa!yMh
zzghjr&39lZghwrr<RxI&8pJEo?#9%&{roS&`@oocRg1g|iwc8yMS8Exaetc7Q{#Vu
zR(V97k6>MA5U<F0L$~)npxuGyhk3QNyLIpKA1`}q!S1(S|7{QGJ~aIts?DLW5m8RO
zBFATRy18Th4$u>5*Dw9P3@`ujvi}PEc}?D(D6sLNgJr5UGu8SgL6I{lEgcntlCojw
zffD-2ef-ORyzEmQGA`emmJLfI7@5*UXPOvd5)>)=ENw99CLV&B8Ja6Yk~5@=CP9(2
zJ^X(u?9rs9#(xLhVm`0UhkaOspvbwFp-=Uf+mFF;0c}yuU9DO953-9t&4}&&V9q>f
zaL|y$O5|asiAhl8e9f)Vfv<vp0u6x%PuJY(S__kSMXn5y8aKNOL$M25G>^LTXbY2|
z$OUEQf_{neWq1&b14VLVkqldS<-{w}VMNHC#Ex%5lR@X)!^wMK8^j<ea?y9?f~{c%
zZK2md!>6fnX<$?t#4FNm@9wnh>+Rs{LUqfiRt8_!AYPHF^U3&A_1=b&4QA7kl#B$E
z)gWGx)WG#0PdGma>H>;*Op}j+v1t$#xmGq|K=U6{z{5HShOOIzc3TKC35s0*`uESr
z)On{5v<rOYJEHrJ*uo@Uk)bK6GmkEK1ujeYgbVnJ3;3EQ@rtA!|6^R}#y9MI6X<^%
zWyeOjib*^pTX=c@yV3WZ+<(`t5C@$L+IO2QZ-Y&6gJ3zO_P*cK<qf`qaTY{2%b}aW
zQ)>_u`K#ypg8lDh+=Xcu-s!Nzeb~{;BwmpZmIl`Oy1^n)C}_NEYUDMwp-E8W=AqxO
zeSYBTb2}dc72d@;c5#|Xd?d{iObP?Wt;L^w{`swDub_oNF{unrL(dPEaVw`@>=%Ou
z9klabLm3Ypj)#snOkxe^ce?&DXTX^w{h+$gMJCJc$ugK|<pj&Q`$0j?uwA=r!dHUF
z_-TGY<v(6_|1If%MZMSoRyvfBC&lH#O4%S@k&?jgYrXX#7D|9=d!(R^gzsz+6uCFG
zN&8&^f#!DY7$JNNY}XpZE3%}|_tT&La065U8h@PP9tUQTK~Us=%Vu=xdxMq+@C~5x
z?+fAgg@z_Ukq3j)IxbAB?q}!Ug;q(Wj#TPv5+5mu8ZG-Ev%$~4pG<aYe1Fio*<$tC
z;u|Kxavo+asBRzl#un%SP@6P`q`};75ELn^cVP7Nlo`<N--6zFgR6N1!gLJc6=~p;
zKdRF(&)fkO&T+VN94$=Z75R&-xmsfdI6<SJ!{(~)T(IK|f+COQaUJS^v{10~jp3c<
zOOE+cb(46$w6QRFyqi{-@NSi2*xrY6YK;=QMyX*EEaz#{f?wWAxjfBizPm#BU1&ao
zpvW`dxvBK!pzmSD3uD6?Epm<4*(6?(2UFL_d@&e2>Mj_!XLIh^5Nl-+6nR-?%};lm
zuj>mv0LF$A&0PXJK?d=P%-uiq;?y<ULAl{Ucd2|A)Y~8^@=AW=(6?XdL3X|sytd*{
z{K|iPBtM7o8DGttYx`pO<z$$}py{@%?ycaFFo?H|XFJvem5kj7{#N*m*|MB1S2hWX
z`1Jqm^B)$Cfg;V=J5Dwu?yj4|3P#)k!w9!S_aDaD5!sBcYiaJ9t9p?ZTBJqcPL&F8
z-?>Cf#D-DzS<c{D&Ik;S<IQM1D98xKV+)dtG(W4Ln_7dLTC>xp8KCMzoUMj9JL3KU
z#_Vyt8L5Z;0VDO`2}YP6zJm%u2yJJmZ)1FdsxMQkEmI?Rn>M|w=c%>wR5vz)t?s=g
z-%r=~h(sSxe8&^N9_EHB(Kiv_O~h}9xrQS83sU_B3Ab>r%{SNPhfyR%A8W5Z)*f!*
z6L#MdcE8T%#x>DPB;OLr54Qk_K3w)4F8kqTGPWIj6(gz;+YG}AfnmfFVx*PB3?c-E
zCxQ_5%V7o&!ap04gE%^4$RISPCteWc#oh0h7Cx%gOHlPYj;ePY4TqRsa#dd})?F-i
z#H~72-y_xCBXz{x5{xl~2N?l|ubbf~7+EOGW;`LTmQ?+LAU_ZSaNtw*Z1Q?GNx(K$
z^_`^gPSVE0M{KVju_a)e!pE~U&a<_#@Ev>YJNAay#;X3*-r}jflZ9tF-k9Zh)4~@V
z!5190aBl?T8wJ^nYQ#2Ip#zWtjbe{b^(oG%Db6@_5zyGtGrpV>S2R%7qkN+R&>IP;
zR#c3I>Nlt#kM3brZ&Ia6xP?Y=z9YB*wC7cQ1XpVW7h~ZRu0aYHubHuvsy>aYGK~wj
za5`5xovXUsbm^)3a<207f8u(}xz{cJMb7sk7l46Ps(z8Hb&-p)@GY*vEiN7>0r*l}
zl?Pn7g-7twBltM<-m3am-ghe>fU$!fTx%;IW8qxBK`tMUkwmI~gs*ah54Z3IKKcS5
zhmL$zpDp;#76OKvdn>9wPw<~71Y3Be;9e=z$GrhnUnf*rCxloyONh)8>R9-i(CwO#
zILO?fQuR`ybE(iB_e)g0Oz2W3^sqGXk<j&#(96Q3N$6-&14jc@pGbTsk^l>*k+?L{
z$kNpHBz!%IvhY?CvXz8cxReByl4>lDFC{H4Jj50>#K!Q4*y4xS7=D^9D2?T(+2Ye!
z{&HK~a$6&-XY8?s@3BQ$_@FK1pe@Y8Pi#R?Y}HsC|HRhP!V~O46YLCsf<1nMo#Ch3
zgVI@kx;;Lf<!`dbZen|8o;^Iz9%c2!Q}&Qk_Am<%b5tMZsAb{Bj_Qk9yv5<(;;4^~
zeO2G?cyqhsT?^lG)V<|sY>i!y9d#c&8siq4s!tL-PZGPMpIz0bisDpJ!w3PGHbilT
zsG+}0)t89c64Bop`_e@%oyE(=uFJ(<7Jr)<x=pOH!*t52damf3D+XBjgy=pY*8kDm
zj8^qCV%Qlm(!!-;rBX43#dS-?#umOQ`rZ@+ESxTNOP3OH7hTm?NS#+m-Ej@6>Y0+5
zDQUY+Z<ngCm&Emw_O%&>pz6D%F1w^2IF_sWUP;?4`CB+w(sEh6U+TJF>SghZq|hR%
z2F_-xepT|lDg{{hzU00y)yLUQ)t^XVPozi-kCH2ml0#VBbd=o6!eeCLF>-)~cgWp#
z$ce8^14z|(%bj=2-7)Y2#uvHE5xEEYrB(f?-1Vs3%fc7s(2H^nT$QT&4cYgG9AM!x
z*<B{rN25vApUYLB%QY?hLazFP#T%8b8<k!d?h5Uxbl#<O$KXFz&sY5Om0(<MdvNW1
zCDy_xm0Bm27z>|P{7<v;r<K~LS^2X{<+DmvOa7cv`y9)^skm<{_0a+3F?)QMsR5SR
zvtDhmUX4F(Ix<u}ORbcphFJKK+Ub${PH!_V3i#AFiw0l>ng=(?qVX2qP2Ia`eGLCs
z_5HNUej0A!6wNn93&2fdRZr1srD!o0o~1RIrN!gEzN#<Ksw~jLEqqh+y{QFQhT%=E
z)=e$O!cR2!6RkcPimLv%R^@Lk95G$LG&{=n{kBb<s{0dv70luC|5QDI1TZLsgfJ+S
zgfgfasm7pi63(Dl63d{Lq$Pt|lhzFCKsqp}GwIBrZloK7z9HW*=sWTqg9eZR3>ruV
zGU!+GD}(+Ze=uk~N5*pup3RZjES}4exh!7Hk;N=t$&r;TUeA&BEI!1MLo7bRks~ZV
z!I2XzF5pN3iwikY$l@Z76tVaQM{cnA8AqP6_ytE^uy`m>hVsn&j^@c|7ANy0nZ=8E
zvWUe?c(R1Wt9Y`C#hE<GWbs;_tYz^Ao@`+89-i!B@kySXWbrAUoMQ1go}6QG2~SE`
zT*{MD7LO6g7=d};B!MKcI7J{SES@Zo$t<2CkSQ!q6-X+J(*=^w;;jPN%Hr(;+0NoU
z0@=gj0|Gg~;yi)mvG}e)?y~s4K<=~nxj>$?_@zKzviOxiUa@#6Awvn%9;t++vUrJ&
zEU_{C^>(t}&fqKu$#O6_$3b#foa-RDEI#ZYhgp2mK~A!`$U%x&e9J*@vG}orJZAAc
zk<1gB=PwY+0v4|p$!Zp_5y=`BuM^2S7H5ehi^X{&$z$;yk=$YNJ(1jF@gtEuV(}}H
zykhYPiHwk#`fiiRHWp_~B%8%YByxnsc@oKEalS<IS$t9=Cs}-3BDYyQPA21I=K0Ao
zNoMh6nM`JJnoQDIJYOdBS-f8+`&oQYCI?x3LMA6zTp*JI78lB-ki|tZDPr+GncQRX
zIE9Q;*!n9ZnZ=V8GMU9Q6*7~>^As|V#oH9Jjm6mt$!2kmLULG~tB_n4A5h2v79Uo~
zVHRIk$aNM!P{;!omno!-#m^P;oW(;_GDKzGca%y-u{c#FsVttMk{K*cQ%M?&XRBm3
zi|49jE{hkaWC4paRg%f#%_`Z<;;ky#%Hmxr*~Q{qmE^Mcuu2ZI_^e9Kvbb0!#VjsS
zNePRusN@QZuc_o3i%V5f%HlGWl(G1wN?x+~l}cW*cmyRQDAV3alq9iuA|(@9JfD*J
zEM7>-LKde}lFs4`N-|iyl9H7y&ZH!h#o3f(vpA2EJQn9ulF#BoN(xzAOi3||ODHK}
z@l#5kviLbA&sjWNBf~YOJw|F|B#TFDWHgJDG?K*P=^B~N;&hFqvp7Q|87y9-ku@yN
z(nuDIw`*iOiwiYU$l@~^Im6=f8adD6iyFDe;u4LNu=uJ*uCn-{Mjo<wgp-VLGVL+S
zNk*}FypxP)@nk2N%;HohNwqLq|KJfcP6DI57Ts3BFwXztrn|H9iB_@4i@ypcy53N4
zgpucf+Up%G)IX+wooQwWB+>gf@87~g`?!YtxJG!uk?46`OdeMUPqYv?jnr@s-v|#X
z5q%jSvy87(!#p5Lz~$I}C+UEW7@`-GxMEUw(Z8Hg-|m03fas;9eJRTyXNwzWWB6&d
z_Gvce@AGW!^K2c^kwWyFwz!+Nx?@`^?W-Tz^9IqM*xEl~`7`WsGwcjM-QGUk&iwsf
z_QrqNUq{CX(Qnz~ZrSTjUGVejhsWY5(MLEMj&L-x_>&xQlN=0xmZSMBM+=L8PHB8j
zc^wy$1QurXQndQGFeBiBuQSBy#)TA~%!V&#9Lq*;GM>klWb+iZ<>Y-L1UwO{TduU#
zL|kn-(!n)B#7q{iA>tYqXWPVV%dxKNJ8a?(7Vor)J6U|#E*`cV?SAGdaHz3y1RMj=
zDjpgC-ci#3;s4mtjfcjeMB~Ib28QBMahq%&5XY@8aAZsJPASB4m+X-`?U8z0c%RgE
zpY)!Em&w7)<Y){v!2{;-MB{Who(iaZDpP#1B&)7Qz(MMsC%lFZQle)GZL@@S=m7O>
zXH*?es-eTuvz_7EL~2`lo@;aE+G<;Rdcy8HVXtlJ$sV!h9x(zPSVX@p*1Rl6paaUY
zol&(!s)i09&vr)D;c_)}*bx1s^2SM}rKM-ZN}XcGZQ(2$lSS*GLxJcAY1BdL!XYHc
zJU<UDU>u%D*L+B$>jmqQM)nmwF5N1nLX73oeI$fE65=eqn<Q45B*s`SzvW`saxu<w
z*{zVOtdL?Xm)U+PY`+v|eUsgCmECfT<zhN2haHvU(4~aO_o0Ty*?ruX!9)ANFplaM
znDI2B7gp{*)1-M?h^qGzdj;YCjjC_v#LZj~20E(xWL})i2VvKPeHmW7!w0P~!$Vd5
zq#&LYg0KRreo7EevG{^Ryx<7JT@F>h<Pa~h__9O1%;K?<I93Y6j-%@1Byk*zlO-{k
z#S<lQB8!*E;u1Lsl^1+Zvbc=J%Vlvni&x0v3Kp-D#Z@d`tcZ)1AUw~h>Pr=IDT~t;
zF`dQZRB@adgh%aEJy{i#Sv;AFlW9;R^Q5+_U#H@A8ibMg;7ioR-C9s1^O&}0H&dMD
z48k2b%&-98(@3v?`&?Dcyb8eZBvrr-HQWP%e>RdSpnNqmivny~8YvWz0WHSM%lJw-
zDX9>%+j%XYxW)|U^=Lrw9X<jz05uKvt&Btqt7k`Nf3`VggK7pLpnz(gAxDh)7ociJ
z@&)Wwn029oZ01_9MzY0H_+rV0XB0e+WoM=oz&e(l>!kq8vAkERv{$NP^`cy<QZ9=R
zN-+nedKUkP)b)szIL$o4pz4KERH5`5jw`BuP6{|Dg<ANk6nIstYT<`c%tNUjjytfi
zBgK@l_?Z;&ObWI5qvR^1<Qgn)JxcCq;W2XH7`ZCq@D}Sf-EW*H|Hluw;bu||UN%!{
zV7P#)ACm))$<=Y(LnVc6WFxNzYH<xSp#~^dNqII6WMCQ%V0dz9U_Zh18Su|W(hStn
zh+hKTEdpEQ!VrH=sBle)vv%>y1g%VPTSI)hR3Tl8vvxstN!l*SZ7J6dxxx-P&e}CN
zB5OxvH@=FhZ&WI5RN`1KzP3woqZdEI%-#VtG*Wlq#E&^Upbb0;JJ6bqG;?);ypgB_
z+c471&;cFRNY7z)uN+CvkzTWEyh09MA-BM>6wO-rLnCPinqS~>=g5AJ;!#^wAInFL
z<>N4m2U4htk;4Q35Oa7y)r|xmXj;_XO;=akx32>M<3+5tQ*<|}lzM;C8Z}(ii^Xci
zqPt)4nAxW<$9}Eq=cQ`rCAZZD|CXx#ExEDFU~591jr<?jYAT%jBU_XTcqCBOhpSP;
z)doAw*l$b>0(CaBf}rJAr_YP{Q2w`Fy@?(X5fO<S?117Unpvo!t6?;5uoJyu>xKyo
zN^IEMLXDyt)wa-3E@CJbiCgbPAIilK<(gS|GUuAiMdKQV=##nDleq*ICr;*iTlh~d
z>QAn=g^Rg}VlEQ5+F^yk#TRqUEPRJ^-Ql8ftDWd~xYl>L1QsXW;d)#6E*Eu|t8L*;
ze8eU`5~m%o^!WHqd@~Cl;9Up!Xxw@y`T@T60X~7ni3j-J7Cy*F9pr0Uc%~3BQ;4)I
z&6z^{Ore>DR|~GyLbPRht`=Ia77{F+DY!COex}eOQ)p@Np9roeLNs=2SkaJ(5hT*G
zTt|?WBS>ot&nB+fB-*lMXOj-INhb@>A+9+re-25QL)uyV6*kujTeM~QuCR4jVe4e!
zl{VK(mcP=Lu+rAf;%~9Lw%DUBOL>dE!xk&(gy>uCuB|M8t36?>y`3e$LUgSVqtOl}
z`erd=vlwYv+MC6eo5j`^E)ZP>Vzgy>7l<7S#7-7ICAv<r{8M7WDY2c!KPb5lO3^sP
z5SU>j3Z+Qwz8>7NP-<=Ahmz}|6ph`P=nth152a2PE|XkkEWb=jD3jV*{A04~m>i9R
zEa<x&Q7lK|R;vfMES6ha__6GIEJx#3E72dz9UjY_Ec`@vJz@Dz<b)@3JBz<qjaY0&
zcoBWE8oyX=hFhPo0#_q$s*&%Si9I|x{-)Z@!cSD!6EzyQK4FEfwtk`}usHFF+S|fU
z)u^XxZ40lX5$kBAwUSv!<JZw<7T!x;ducSTY>2*>w%$t<Se&?*_O@^ijmn|5Eu5@H
zBx{k@3O!kiPu7}Qc#h_pqebJ^DR?}z)^oH37AMZpdRus|7ByF^ZQ&bQ#0@PHw^oUM
zLyNy*#pDtFndW+?MWaW+gIhn-5?Gx0OzUmo=UUWrtu|tJz@dI^H?M6e5dEA|{hShJ
zEdh&_>c#)W@x@A0<l_Jd(}9rz3a2o)nfghT&4f>AIKYG=gbx>@(6WG)l+bmu&<jK8
zRDGt<e5TL_hb~p0D+J6Hs#tin&~~-(CJs$t><X<ih4vQSAOvg>svyP$Q!tl$5=)_B
z6L)p?%8-xterG58I1)OJRJDe#jU;L#aalvrK3mj2o68znj@zS-+g;XBaYYWjB3HG#
z%1|X@s8S16zi#`>W9>n`8uFmiTD_ZEed`+zalSFc*%H42J^-51$a96x8cc8n3?sW0
zS}eF&0)~;(3e5yuB>}_8WQD8S#u1<ReC+)2r>Zc3bM;1W^{okXIrqkLt|i8z5r{pj
zcaf`)-<LoXV25nd(Q1_vQmuqkw{RJWD`Q=9kI8F~Nv!3f8)9oegmuA9u)Q|H#_%WE
zs!g&n<sGqgIAZIFjY9N$wrclm)h#^C-f)<`k>#?RU~fNxb;-@Lzc$MrYw_pWtIf4D
zfB)3p?5VxE^%W;N8c%dEF1)3VhD%wO-5xP?k66_j@Gpy@m&K~q0Det$T@!0txKxZP
z6<rn{D!YctwXFeolw5a|>|GKOeV7t6OsRufL-a99^cW?^!VhSJ2ecukdm{SZH2QBE
zgU6A;Aac5fI=vmbajR*q);<e!{$E#bOo|5G%gBg^{rq(^6&f&%1Ze1I#uW#co<{aF
zYhi4hLPr6J?<&s;h!~TifmRzC(XhlOzPo0vpZ@dTom9Pt*h8~?Ht&h=F{r25lR=4M
zB7?FxF^khI@6JX}+{ofhoVba_TR3qGi??#(Ru*sL#BD5|$cqzs&GHtd@L~#!C-LGW
z7XQhMf3kQAFHT|cO<ugoYv^lG^;^7ni^aEj@ivR^^5R_<-{Zx5EWXc+_gS1Ti1~tM
zd9RKO;&B$A5X2KKE)c{578eR)A&bvB#B&bK@^%$F#9|hocZla%e9<9ZWO0c@EMf5#
zhj@j>qa<;Zq*>mv(ULft#bYFK42zQ_F^R?FC2>59CrIK17N<yJ3X2!Y;zC%anx_s`
zeUU6KV)0^GT+HI7vbdDR>9Uy4;tW~LVDU;>T*>0qvbdVX3lwpIqFLUzg^IY4#fubi
z5sQ~7;u01wQ^aK~UapACSv*D+$Ecd+tsAR~V_BS}ib*UUuZrVYJV6yFu=uDd9#u6A
z==TKpi+L=bM8!!|!wF5*uTb#{)hzGdRVrR(@ii)5WAP0t-eB<#P28brI1j>VSrd1%
zc$X&bV(}hL+{5C%nz)z6IhvTm;u%hHhEqfDBiJlXaVCq?oMIY_XFJ8&7RF^8jATaQ
zMjYbsL<9_)MxI7IGJ$y-!7XVdX~bnW1eJ^Oa<LMwR3PYD?7U7)#5`3H6fFf$lp?Te
zU^Yh>eT`I(mX~9<B=45$;}Qf4mV=keZBWNC$0JnDNbrc39%gm~hLP40%@g!n!lR62
zj@CeLByd3i2(&FvsD>DGJVNsr2_A7y!OV`pF!DJf1D!!o2P2syx-$~0MEz27B<cUZ
z!vxbo!lR8`kofuuW)etXXpsC7?aOv%@<;e*BjqDjwcYWw_?d|bKU7wA-)g?DR5MwN
zss{!HR>$KXs;*blUAVWZ>iq-ySI6B{6}(K<_HizBJ3(jS0`j=(xSOiNZfdnTybE_z
zJ&x;uWqft?33~QqDwGN#?4C@(BX%!vl2~Dq7{cz!1S}V;TN<`Ps<1)|VfSPL_Dj_*
z#n~-afF%0to=m_|%Qp$$W3^hL>iW@43kI$#HQ>Bj9i3j_+NafKQ5Pn`fFmw6ZZVC=
zb4GCNfmVA!U1$nHSSbzon^s3tNQHAE)sJhn@!*K(V3ey!tB2>AFy|(G86(ps+S!<C
z6BwQ}o9Hma=YSd;={3=q1~)b(Wt3ht54WoNCM{}{7T3*88i^kO{?JIeiM<>SPVi(N
zPjRmShC4p$3Ll5gS<J=>6)|#fVh!I;`MK=ywI8~|?&woT)u)a)D|CLDSa+G2V11)~
zQr&%00)7)r-U%fdDLe5U(EbI6k*^aUg;NVKj4Yklt1&$%FpS)s_!qcV2ZoW66IUZ>
z^8&-j!-+!<Cg22yk#-Xex^Afv_glN74*GhW%Wba-T`cEvsnD`i=w#_`nb4|CcoY3n
zs{TkwcqDYT@FcP2B(W2#%L&WH&Q?36OKsDoE|zn8h17C|)XCEAT~e!E(wo?h;L(*5
z_Dh{DyhCogL+)ZZuXoEWcgvkD-9I9?IwHS`eil#2d%{t<vxPS*2^*Ep=#KS-sJGmu
zbi(lrRS8<iNIr_+4ILN2Fj9_U)8R4}7)GK|{5+Up6c|Q&Q7o$G^VGPjTZYyDZ+M3{
z{NsODcn1#bFu57&UUBNdbg%G9ja;udWF?wOUV&j`c*W@uQ@aAg$m@zzZttcuE6%rF
z|Mvg-dJ!|nLJd4=W3k4)&5W_|&qlgf93y(0d18TKB#FhBN7o)OjMT6=++a>vU>FHu
zQPD9IEHI4xuh?hM%ms##^%a{7rwnNMe_AM*>lOaONc4)m>=QG?D~$O@dRKh7st?xW
z!J5A{OYPMn_iAxyqJHu?@WZw_nd$$}_tQQx6VSq|8QEvCl`!orFpQkDXq$d&CYyzK
zHZslPRH5pJW%;liVvTl7mEfgHG|nx!F9t@fk$)Dyt;eL1^SDZ=`Cu||oeQ}{OakVK
zz6hSoN8oz8Ps*5)16_mn|Bv05I1j)FHqzUoS&eyZp=w4_ThvF-3?T29Dp@l?zSKEi
zO5AH^08;f;a_}lS0?lU3ehUS8Qs1J!tNM@fk0F*xx14LWoa>6-H#n)uxwmpH&>w*(
z;Gl(#1i4sEu$4qIORSk;W@-bwK#Y1Sw#T%fm^l}|nUOXZTlgz8UoJ3=WVu-4*OjXV
z*E-lYBv{pHAPvKW5UQ>PYGG*Ws(PQmK4F+mMb$qG{4DG{GhZ#7K;{DXaA6nB#368M
zjSJk%g`okhLXx?_G(IfD%s&UOzz5Fa!?v0E!BjBq1BVG=XzP3Y#eu_xFw0+DA_SHQ
zVHn${LgKE#%R<<8Gp83smWYALV%Tmo&6EoJlz|h(Fx<EVb5aUiEQMLY5=*4OB~lnB
z!hyDz1GmayxYw>i!j8agav0|FfCZKkxK0V%W+p3xQ+rC_dL;}4(BPCJ4V*#4aI}H7
zAq|{K!_a%BLdt=_yEN>jnT<e&w7h}$XjnfpV=rvwYk|kKu#0B06F8No1?FpEm}w79
z3(y54hq2Y$_Q_<QOmoaEQg9quiJGRwVK&dNcP(+voK^Pze|^7j0{uVP_lsdZpdm&I
zYisP<ERoGp4UBXE4Uwa!$?Y&>3;GOTP%si(<Dh^Ut$|^rv&K2$J2Q_pFpMPDXiwm1
z01Qv+Y8>`(MGIB)q^!n&R`tH5FI{8iO;q*OT-0hV4$T|v8Bmsy!5ZgO+&KY;k-i$A
zj^;cJx<=A!s|8j|?N&?Ot;nc#Qmb`R7i=;JHj&6i$q$1@F}pQ9!ANb5t9V?zLL`on
z&>F1|@MWmvj7l*%AZ`Fa2}W9LY*U;-VCpfFSz{f4GP792Ul=K@ai0Cj%v}u(BXKoW
z26u&lVWg|ZqOelX+(wdW%L8NNZ$tqIA1S{m1~KNYh7yg$)yO~(E-;LA)oAz(H1kvg
z!;_>M-5@v-!9N?xsqyu3MF&+gQc`0<;L8@sEx`{H%z`c8h{|&`#4Jb{m;&*<MuuuM
z*1*Z9$UBu_TsuJsrqb%9@-A*w!aFL>ij}u4oU8^XtF_Vk``yfN4N7CAx3&u2p~yRw
zN>;%Kl+Fi~M6|LY7E=u#r$%5wxM>UTY$U-(eZ}n8@B}0EHFhJ9i$h*a11uLu7Oj>=
zA&a?jw}c9Tw;=~<ByKn333|{SBSH2dGtVAo#)eik(qf|~K!CO+?~npe>tJt14qhVH
zLNf(-SE0^EvTW2X%#sZZBSp5CVdln$moXA!BSY243pK_IvA8n?$C8EiSweT*koAPZ
z2A>iladQEE81OPi-fYY=g^PT686#=7r6RWla$BI7><gw$N7OM#9ID6&Gn+Ql*~p=d
z+KW92N;EQOTm8Yv=ZFHH3^ktHauDM_K9p!A)y7YX*|ecVBbBx_8y}M8Lvny+`z}_3
z7b~^U_Dyo^m^XS?|G{wfd9K=FuKG6Shs5a`9%W?S#uAe{3d3(rc-6S0s)x8jVzI-i
zdZ;Tj7Q<XseSm8~EPAq3eV}V#EH1eqw2O<_!o{Ff>9I0vZRP4%_yp%V!NsD{2nkj=
zR{<Bh+{`zu>VNXJ{^aYR)eWXD?^?yjVjf7?iRWFJd@QC?21|zzyURyeyHxl1TKD)m
z79Jv08X{D|Pysl9F1S*JSah9wA_H8Lg;*;x;G$6JqEO|anVJ&L{|m0?LM*z=RPg7y
zUJ9}3;8XR{L>x_==n?@-jfm@s6ZfpaF-X)RQqjW2#J89P;-0mt-zVyQmS0AE%UJ#x
zn>xnE{QWqa?>HOt_j_#O9-9;QtRbI)O}$~OXyKbS-<!5T+@ppiqFo(qXZXYHzQgPc
zf0|vL#`0&_eP^)zJiC}@ccM=ZMpC=)EqfpyLxKDn4&O-*29FafjT5WvG&8@#yeqm^
zi?Nu)RfVO2YmFF-YXz``MRA+x#655rQ$%%#Ska0>JR^!{L?=#p5Ed+|=fsLuOyVU`
zy~N^$Ql*7b6^v_!{1B4sh!ksu8Rto^JSq0BnGO!Vmn0TRPMq#M`6|RGk`w)lpbwJz
z9P%;Nx*6>%`nd<h?8}ul%T+MNo(d^WT_@z&gJ$+q&~MpQAje`vDCFCb#f!2NBRRma
zEQ^23PTYfs#8-;n5GCjWZ*QR-I{#ADM=9DUrIN*8qf}a>RKc`pDg>mua+Fw%P=$OW
ziYr%%#Vqfz5>j1hYAohYg#0?HYqlDTNvb`&WL5Lj2y54@M0J&@u@lXF_uw{GU02jt
zoUBzyRq9Hou}91-rYaa~t_&KB)*Gy(G}mw~7M;N=9DQ(&)MD{01gzY&@I9IfT@tWz
z)2i*&YFYTa<~py%;;|2CFU@sPi^aU#Nu_NHzWBWVi~r9^7n})Th%;^uz^MZ>4FSWr
zF91E}cu@c_j9UWGjF{NRxE=r|kALndVI#p`7*_${9FUTK{qB%0gI4}`8T22%48qu9
z&;ZYc4``9%p0Y<V$}f6ABMcn^VA31cJ6N{UK`G#%wZVm68_+q={SLT@!C?yuHZFI-
zzsLR#4C5{bG%QoiyBvUF+~9yyJ#GquRpGh50jsX+e#DPr@=R5KC+eLzyiNdbb%09!
zGl~YYfx$zJ8ywIDlr}5+(yzOIYV=>f|Nm0IKdyk_^rmsYhGqXg6y%3Oh_%bLk<{Ku
z;w?jF6N%nLnp$|Dt@b`!yk&Ug+M;u9O)Y%fUi-K`-ZC6c*rQL_n_BpiBlwad8W(>U
z%?~Q&xzz*J6Z3q*KO47sV14l#4`3K~c;Hw(Z`~;Mdf($q;rPNk&UYeFmpu*u=V~qz
z9ROfha?VT^ui>0)SUi?@j^!iqJO!Mn;+;t>zQQ}N@R7K`3W>0I=T#P8<DJ)7d`NH}
z5+X4U0sPp4^Dv8#I-Exxk+|OQEK!|#EI#IN9%Jzn(fLG-#Ko}3OY3~f;%B1s8H=Bb
z&gU$iB0HzZk?5y^#Avc}DvPJd&S@-8m7S?9o}xIXD3Q3dfWKFqQ&~JsaZY3LFx5Fs
zjl?q_U^uAG;VeF&IuEFkWo8~(Sn*Nkcp8Z(VpaVDbzY#6xWDV!7<X>cB5^kp3^UES
znZ;W)=N1-EaXP0sBQcZX{NIx%9MT#k{CCR!51+DeVuX3mxXH&_pBoo=pa3K#Ex)rB
zF<$TkBaCso4^Ha176pcJrw>|`7!d%EGH&ugixR?TL^4OD7;y!w3Nh-r7>At~ulRuy
zjr)CYl*7>r7@k{wa8$%S9r$PCHXpPXa9Iru<MJMynDO!+U>G;|;LG5a2{4TNdhlg%
z`v+dexTVK>nMI<!NPG?Vxje^)gU3lNFd7Uu6`(}p(jN4nt_)Z)er(c+lmFv$_;}k7
zlxSS_gI3T=^Nt^27&rUix5G_6U>NuLU_0Qt0vN_EK3F+iAOXX;y9eDgcy|vlJQw%i
zCqRb-{IhXW54Ib6oWYmqxuyrd48&W>{VSoFr|Q3U{xxxjc?S!G#c`3#xw`mSJqN>s
zH**o_M8V_0@GRqMAAA;uGkLyI`Gr3CRYBn>nM3Jm^9BUZ=oxi^#^LC>*1YKlEIZ?V
zA8b){Edj&0)d$VT^?waqA2{|t<URQhPe6cb4sB}OU1W_H#)T&+0I3VhZ!bl>@$(CZ
z+Q<hkmZu)1{((&DK^nlI5E{auP#VghYP1@I!f7~zVreXcTGEybYE4@+r~~c5pw6^2
zgSydf4ElzC!=UfzcMKXp2QX+L9mt?x>8}j>gZ{yw@f;n`F?cpdXR~-NN9VG5F-I4(
zcqK<yvUoj5*R%K#M-Q?12uF{w_yk8!u(*Ju1uQP)Xd#P>I9kNw8yvmC;%6Lv#^M(o
zeZk_PJRQn2?>m~OqgkBH(_|Je;^`t5FX8DD7O&#zDi&w*G?T?^dAgRx8+f{b#d~<V
zhs7s(dXmMbczTM(=XiRK#U(r~VR0!>OIbWdpkoB)eUk*5#Nrfzrm%RjKqs?!ia@8Z
zI8~sjEKV0_I*YdobSsOu3v@e+_Xu<kiw_9&0E_bkn#bb10=>)P`vSes;^zW=&f=E>
zeaYfi0)55ep@a@4Onamfn#$rOHoC;d@YmbvdOL%&95l<p;2a0dVR5d5=Cb&(gC1t_
zNe4a2;vxqvV(~2py~X0k4*HnI^F%sNWS+l3qzhQQTBNI4yhfyJSiDZ8>sXv6(kvF|
zi8PPJcSL%J#rH&dkHwEf`iRA^MEZ)wBP2RPV(Pn1qT5)UEzxWiACc%07UxMckHz^C
z&1dmRiJoNfZHeAy@i>`|lbPoy%QTtAlVv)Y#c484WAS{M&S&v{neJ!tL75(8@d=rp
zU~z#=3s_tz(?S*($+U>Y_hfpH#p4t@PGRe>&}0@*R_J6F&s6A47SB`YJQi<L=r$H-
zD>R$MISS2TajrshS$sgD2UvVqp@&&~U4d&ynfg9Z=mQp)DYT5m&lUQd#Y0p&L}mVd
zluAdjI8~*oES{m#87xjyX&Q@Xt8_Mt=c;rrix;SL0gE$Ln#tnLD&5TDtt#Eh;$14;
z#o}C*=Cb&(N)NO6tV++axLBpdEG|)L35&0&^a_iwsq`9)OI2FR;xd(%vG}D*U$Xd>
zN?);f1f?S=)80vxCb4)Tr4v~^pVIj(UP$Re7N=91&f*M8Gg!Qm(v>XEq%@Po*_39p
zIFHgi7Uxr%&*DN#3t3!DX)%jSC@o>}Q%aw*_&KG|Sv*{$!!@QoMrw2<i$`m8G>elo
zn#AJi8lBGKbd9F7I76cuEMBA0H7w53XcmjNYjit{3pHBE;xigO!{YNAJ<sBc8okKk
z5{;Iy_^L*)viPA!AF_Cala6pQ?J>$pN3nRkla6QcWG9`>;#4P1wJ=^J3JaEhhM(Z#
z2^LMp-J!S=+tM_w!sXGA#=~Kmxq>)X@WVI-Ro@_p8w5YxN&=T35tFQ&VIb&$h*@mH
zy6r^V&f;@KJx2m87wJQyK4fu{O--^f{2ZH@^N)mfHt{fv|FWrn*#az=>QK8n)Xw1P
zc6B<7^X+23l@JdO<=E9ydw}I~o#0R>I2fEQirJ!{wfA~f6wiu&%wDS`p0jSQQT1n%
z_)PNNTmw;(vUpkcv-UOzE8<|q&)T~@p{OU604q4ANKuO{yfyb^uNCnRuKdR@DKQKc
zRMEJc7T50>?+Fa!N?P<^pqCOD#znN~rr2&?MC)0&m)}C`{R=Zx75>7wbQWDm+5Jmy
z_kJ<3OAyhU`!@H(9W0{H<a}pxepa@c$DHqD&d<s&lh6C+^L}`&(6j6hn<v!3&f>{C
zSU*c>h@Azp&j`Motj<I9E2R1r_S&SI#Q&z%8Q^fb&3~BH8HhgBR(-0Cxn^m$&3`t_
ze_*Trz{Xs&^vve}%<3GF&(iLnVs!?hFSJ)*$X?5oVfW8q`GXxHgB_vRIfy>WQF)Xj
z*uo1OAq!Z3x}$PB%Ri~qJE=54<tF-F#rLk_hZivr{V?@CO#Lvtis(x;-zAzK26EwD
z!0^3{i-6IMfI)r`17%$QiyIrBGi~yDwUTwF?V8&Dn%ea*^PWBkZK1)FX#^g4#cmBx
zFfIx9KEb@%3K+)Kz^E$dj{=70LSP&)@op6Gh!__EV?hvHAd=Z)HN1xo1_&|gnb;cL
znmf%Kf}uR)dSFymj1Hl)c^j}*VzL;XEV`^f>#3r1su+O5>YnRboHN7#tdi%tmP$*+
zDwZ!YU96PO;uT`d3bCHWUoCcBEhbvwD4WHo&Ejj=R^S5`D`tzq7A_ED3dDNWka$Xr
zImP0$V#Tv!F!FJC1K!xU${62xSLE+;jn;Ks^xy4<|M2Yw)GW{d<BDojZB%Ip1u!nA
zMh0BrBL=Ln&cb3GF?2WMmTL6P!j_jrHc3%<+}g8c7d2IW6Swd%@ew@QxU?Ew*x0|J
z#Pj7hR$DGj<Jv?NfatXH%P0}!Mb=QFaeXxg)!_BjFb(}P$PtgMz+V_QR%7Kc6cHH4
zJ=It_JX!$bopC!gdYnDy8s(#MC3H!{fgQPXp`3`RYald92~JlcF#2&{{;`7xOCRp|
zZ|uSU(%6FoecyOGlC*g8|NP2Dyzm@!(YWm#M`^s~9Mt-szJ3hvgTFAYG)E=HxY1D9
zPb$C99J>x&l_`Xm2rlbL(Gwx&iBQjquXrwWeJ&(gfu<u!)ClsL6<C@?0+L9mg(s50
ziB?J}n0ragY*No!Y|J4sb6A{30<uV`#otQ8w^}Kv;372=u$|?fBZ22gRZIRM34F-n
zBwJvTjp47b#jLQ^vlb^SZ80lZoMQ{fv4vXvgSPO4R>~`gf3O7{X8C{F0{^mAwd9A|
z1BcofJl!5RoyA-1F<b2Q&|ZVUV|&b27U$ao^6g&x4kE<u;is$=Tb{*HV5z;T_4gAT
zffF2Fi~-2-#f|Id(Pqc6co<BLyXURJbc!fW5!+g}%4$hmEw#n*9IvE@jX~ogdfb!r
zoJ?$)$Gu~nOuWc7xX2}7o-DW{l8+z3cfzPEPt3IYPrf<EOrxC!HS*kIZ`FrQQL6qU
zt_(x;;q8oj?QxhsVqRzu4C5MmEF1$sfni);k6VpLo7VdE-Og2}{>QK4j{fOC@RHxW
zQUBva33xmgMwRC@{5g%pKz~e>25tIB<R4A2sc{Jb|9o8uy{^>3;pyaSH;dMG|7bFt
z=lxS`_@~&$QngeuE>&z{nZ=95*hOM<%j{h$MlTg#x9}>d$ttNmswU)}mKv;;T3bh!
zc1sO*ORcSw9XV2Lj?^4+!7twrem-SK3#j2Z**#8ff$D>A2C8_QI&M?dQpTq={HgWo
z5CQRzrZ^>Ek9<>{(FNM*pMr%C)873eb?m7B&g%b5XLUT|3c9dFl9x#Ss0*hP>eQ@I
z%RlIU++Q%6A(VJki@d5eLFdQm-s^JQpC<GKQ}GdXKB6J0ZmRx@%CBeu_B79}gAG!&
zmUzU*6NO%TkJb>Q&QE_gbbIdu+8ua8Yj^A37}J5#5>Wmes?DK6IL3L-QbjDLvDR6t
zm9+CpnuzDxJ(=Z$uhR&O13rCxMyH!Q*6)C!{!70v!?9zl`VaCC!Bz~&2(G~hE&)dY
zuxGibbgr=#1+s;!vxRGhj!e9=5h|Uj*34AvqXms;rh(B%?&CcddFE7yjLWyCWkX*~
z6P;;dh_&EYD9Q`P0BbZDD+Q00YGD8mCMtnH%#h>^sUm(+IP4>lEfU3p?5dtBN2SVf
zxBxlZ&Gv5Pl*diri?1WLb)=IO0<qg>+ihd;QM>J^-HXrm@c*T-N0XM&@bh_XJ|BdW
z?f+@->SE$5qA<G`z4y-Dd)XlDmMt4_e+t<376P^e#RW{$s?kl;rxIQatw6#9E$M@Q
zwOWHEX-S%<6e6K5G*xJ0l+Y9_l|Ts+4J|aNX^23>8oI_5ke@;=tmnpj%<fa;gV8kZ
z)80KhXJKZ}&zza>8|o7?YmX$zpET!9nz!RnzX1ozm^J6%G|tb3&Xuy&Z$_vqQuLI3
z3@nN?^zo3AHl*aiqfU6bH`~LGgCuCz^mfgU88z|C2Xb!y>=hCWtNO6&$5cSElLTWd
zZH(n169hJ$1pOw~Z%ROT2NZ(@O_r=C%MOg&o!px*-M_TK^w>vEkQveNE~?t1nuK09
zXqMQTn#k?;s1`V?6@zZWxS_#uUGu_j@6@-%9{(x(b?V9*HMU0ep*99!I|)WKHliiK
zWdtX4D@*!?Gk=@KVS3}~FFV`be}x&hqsPsN$-qY%G)ZwQVlp6x<QpXS1u2tTQ=8;&
zi+t{tti6(xBbvq~>$v3P2&R77-7jY%wh|SW?EW_Lc|x{ML|%ljCR^9|#RdQ@Qs>?i
zw0lAlUx6$MY)MGuD~_-b2n)p^TcYy4=m?5loWQ<YdFajBXLs$TKAe&C8OaZSareU$
z4WHa>4U<1;H9ysA_HgUGXl5798#xQuoLQeU`(ceR4QklM#4eW1vw>(<H^QPa*Y|82
z8h{sp{*-HsU1JFl0CuGio|ZYUG7kg>!-A`uWpIXt5#}3V*}Ql#%65;kVl3ncxXaSV
z*%p|yv6o>BGADlyF?)zJ@LGtR+IOF2^SzB_=39<@e#pEJ**do>;qDaMVa18>|7zQQ
zVbjz;`@c@bg7f9nwL_vlBo<-1po~m{X-nF)Wh-n8evC%3&m{JlG(L*aq<w_7Gr4Y<
zGSC1#AEZ%b>N<-Zptz)OK@=Cn0<N!7k{YE#p2c>`Qm0&qtQ28}^qpk-Nmhio1oR4-
zK)e>81g}r@cg&v1>F-hkDa8m7bQ6zbj~A@2Km(HIVL5A9-m$KE%3ihUwfRiaTq9^T
zLJ|sK1|XRo7ZN!j`IHbiB^0A#X57`|_(AXjkUUqPx3kY?4^n<xE4XTfG$@+;bZ=GT
zN_B{S`k100Q~a3GFk5O7U6lL1>8qYCUp(!3?8C;qHT2t&+Ca`Dv44ooG@}|+u|c)+
z85K<vY$-dhy3VVc(EuA4)Ua7fZI%k)^62UziG3zL!<~Aa?5LCTkRmNyX-$3d?Sijq
z)S8soCdG$Q!}b{oLMBhhlmlZ}d@knJ$FCkSmP`YpJ|OyGcK|$)V9=ruT8fY!EWI~y
z$o5BX+#~<}rJ9$!Jl}OJ{^LAbunOv~7FleOQ+Yo}pCYDh6xr1yKaU1LuZqMBq&Pbi
zH}Wro=*D$fN#{g2gUaqfWe?zO#`Wo0x0G#2;0SK3I7W#*zVuRN$^4mfma^MQ*6m37
znBo}Y^5yLP*LQwCQ1Jie=YMMRbLE8}-`aH|c$$1$^ajKO>)8f%^_?m*h7K*QLkl2N
zd@x^s*;jt}DhZlRdb7!oHPvcz$HCbbJ1a;qrRh^zJ~BUWA(3EKWwUAmR3SXjELKD%
zp3mV(Du22!-ep@S_79&s`$YLi!8+2_ZPB|eC9qL&G?K0!#(LNW_yNLcv)BO^4<0iE
y9+h6NYz15@=D@znH;m;`^?}lAxACTI|El*!>EOGg=JJY|a^sV~UKMx~6Z0pb`-*P>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/multi_index_transform_helper.hpp.7CC4F9B71E8A7CAA.idx b/.cache/clangd/index/multi_index_transform_helper.hpp.7CC4F9B71E8A7CAA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e346249d06d4c5cdfb0b1a45976e7ddc93089604
GIT binary patch
literal 2414
zcmYk63s4hB7{?ck$mI?sfdE1R!3$5(D5Xgh<t-?bYC}Y0M2!(r#3&G`2q@MEg#jO+
zP#pmkm6wXz%A+y}qo76bU2TV2$3RsC)LMMjDRl4Bg}a&i?*8BXW_NG*|9$d>VPWwY
z03mA^rt9KT)hqx22j2SBlmwY601M!^^wJJRY~t@TZEh@j9fCRTJXD#<RyjMov+7^%
zec^0yVcm=TxXuMBJ-$P&iyEd$V}?#l*KBFN_^f~HbEhc(*dejq7YjAVRd17UpL(S)
zZ{XY9t0TjkQ%ZdDXS~zj+SWc&@z8Kn==#JW<88snRo_H)tjc2K5=+okwOM1oDzB=g
zxx7BmOuDb4pt7=|v(o=f&jwk5=smGdlhCbp(ubP16!D!@+`q5+8>)JXt5@V!e4df^
zsQ6N0Ek8@OXyznh*x|am<m7fyP>%2nd45N<ea%ot^07V3f0`6uyt2<a^qx**-Di3B
zd4KzrvM}wPpV^tKS7&ND5qM4Ic3-~Ni^kpZHMVE;-b<gBv;XWOmTeVmaem^auy1;F
zq{hEmtY`~A;8R<%*CReOZ(fu5a$nPye+9kA-puh<l--h^EB~V`$mL^xpTnF@+3WXg
zGJCpe<L$b2{5M7H7;N2Fozke$Wfug#6~FHP@o4c9>k_f;e%3_qUfE5BNpieLm04r5
za&p<?g&jAK<{k2ii0cmnok5XSvm)HQEN69T@ViZ=eB$1+2W1`eUA?)dyIf<U6Syby
z#Y*MKwN%jrL;SnFKg8Hh?1_!?I*@waGkZ_$?T6D-&zp=CHK(~bJjQ~>5fgfHR*vr9
zIbxccVKX(U=!L6VF(^>i*M`gV^4m633{CU<DkZ_q1Lo~{1BVo+B6lqz6e_tU_r`_V
zyWZ^=M?G0RVwa?tlyk%1w4RsVZ~bn@*B$!Jx;3v%05B~yopsvK6HEXn2f!DTVksQ}
z#$$siz=UtYF8W~Iaj{=U05=F^ky$c20L;eR2#}AX%NnCj2npZ?0UfE6(g9$ciAF%w
z-Gi6r3(pWh1c3!$fg2qF#x*knUisj%&GDuO1aN_XBk4GY4gi}x!3fO$EJyz6%Bi;m
z@PGhL;{J31n1zWEm|LE3vO`_uM*tE6K_G~v!}t^{;^z4uU88OzfHwrv$TUAX0L-2+
zj*xhS>-y5Lt%d-u5QqlRB07wZa5!LNo4FyUnE*lvXi2S%4gj;mjU$wOJiF^RyL<rw
zJRu+=MSgSum=((i)GWR0&u&bdPJn3;P?E~obO4wwW&}FCe$foJ*B21L83JljEu#a#
ztS1?P`yaacID>T(0=Pqf3%Fc50L;?V2)zGv_O#uAZZ`p@LO?=Fq;wen3y=~QuQ$LZ
zg3jH9O`rF<@@cJc7YiHAUVeGCsv)k{gy73qGH=*u=nI(u(v$-bZBbv{>M=2?Wir9X
z0d9_Hg3kdQhH?as3?%|0h7toYLrH*yp#ni5LxqEIhLQt0Ln(ohq0~UlP+Fj6C>_u-
zR2oQQs4S2*ri3mBEfa6u;lPdX!F{aIb|JIFji8*4HUe23rYV1;G%_|!Q?5p7WM`PB
zygZifX&D^&R4s*5dg7ke$WO@Ikfw}{=%7zG3KHv66_C!xvpML5LOV+a6#tMCbAr#a
z=847}!{-I?7LO|@mJ<)Hj^I1ZbP5_*?ksm6nw;Rf&vXwOS1Oi@H|Ar$YK|%rDO6!_
znN)e9Wum%*X{xADdQ-3Yj<UKfiE#A^cmn6}|HYN-7pa?j@3jV|HMoX@3IctGgHVNs
z20_b$X{y{%`uO*AMK*qpf5Aaw@mMZeAbLWYsy#$EJQ2jS4dRxNj>IFUqLXZjf7~6)
zzHXqbjp5b|YhxUSa>N`NN`#3RN{opaN`grkDi8}~sBkQtq2!pHp_G`Cq12d~p|qHm
zp>&vzq0+E4hRVXS##FnKGiA9hZ!P@e5In>g`3r3}+#jm-p#8b&7GOU-K@tjSJKW9^
zSqMELP4yu}Kd5ai9u5|Lgmf~VY=z3Dzw};qfuJc6rkCJLgvb;_-CLi<Uw`@uriHkW
zi!3;tSsHPoRLF(tbUfV#ISVy6I3?BD(3GR>!0ru++9{B>z%6W1aluZvUQHdFK1HJ^
bsTV%y=*#=B4%WRj6~K-}WsH{qy#VkZtZ(O^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/normalization_bwd_data_instance_common.hpp.3AFBFA5B5298FD17.idx b/.cache/clangd/index/normalization_bwd_data_instance_common.hpp.3AFBFA5B5298FD17.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b23ae11376072545e8aa4332d34a43ac9289b124
GIT binary patch
literal 1978
zcmYk64{X$Q6vunVR(f~c=wQxk1DVz2Zfm>UwK>YJzwOrUvdzulA}DdnjL7CBJ4Eh=
zvH(s&G)vqFabXb_m~+D+$Z(R$bi_H9sgST3gK_f5U|<F&8!>T=pklA)XJ3<h_xrql
z^ZmWwd%wF?6=h`u3kf1xTd{RxUE{hGf*>-?PogotL2?kp#RUYhweyp`#>=BYrZkb6
zB|TC8#@J|y6xIXJr_JeQn`7r*?aKXiW@Q)OuAW@AclgBk#HOi>lkx4H?!o<6YPr+h
zd5^#M{d~dnV?AQ&M^6n7@AiGz|J%`6QBCdiU1{}hrX|*BXjP6SNPO#Xk+`^Zx$oqy
z{4d+D<)1qH?1Pm<;g^ocA6z<LA1~kM6enxuQ)f^7w3%#hzjmon?pro;zh_7A7ICHF
z^Vp+=K2y1~uzUR5ZMSE1qq}r<d0W@o8#QCCe04}uM%KJvxumbVIxuvnin-j({k}c5
zW9dlmibLBb$Bi?y^?ef;9<JKoQyD9`@XFlbpMPi`etU%@x?_5_@=i<teP`c+Cp%94
zQgujPN#8W?&0TMPed^6W{`@%h%)avn%l{(2YH!jLO&edJ%${XDH!$BY^xaO6)168%
zK{6P!0^xF+F9`?Z$mnn3wj95C#^YRI3K)`cS%GkwuPwpilDj!)bJvf0oRlfZCbL6U
zupOc|+#Sar`R6l_(`5=W$;><}*n*1LuRrabDBt68lBS@TEDl-0-eSw~qO}`lQ&o@C
zV+wqvFV70LLt9Ct=iKRUmU)~DO@T_Pc~-C;{$|z;T;4P`<#DE%f($alWd&OhuX{0J
zmiK=@9B!)*t)ut4j$K6$;VwMjwoZ@JG~@Mq1$;o0ByRwa><s}@y<tGQH)<2>XMqbV
zvA~6uS>VE|EO23U7PxSJ4!Cd<2V6Lr11_A(0T)i^fD7;E1w2!l;Uzu*NajO;R6Yzy
z=cAbLx7o$yg~08?2R75mYk<>u{$@Hke4NHpH`B?H<McbXOP7Wx-sK!LV?J4UGWZIZ
z>g4>f`snnors}JmZ6wWPxw6u61vHagm~G%PX(q>&laBYJnP6eiz&{L{DR32}V|SV<
zE-W^1!PkCVKJ@gFqFrut@|4$So3j)Ph@&__JjMI$jRi{Z*&B<L2uPt6pIwGVX+FCQ
zL?J*1W%zJW_^Ri+tt*ZjVU~!z5l|vx0da@}#3MceO7shYU7{cf0YI`40;CFIKuCxH
z(uJr^qF)4_q9g)OQ5J!xsEWW-L?ZAMbrE<fenqg`rASHukgS9NsY)0SQX+tKC2EuA
z*95!inxq8)$yx}Is)Yd|Edof_qBbEv0-i`hz!S*`cp?=6PlOQgL^=YVhTjnI1)v#2
zG6H~PBLqk_!hnzw0i+vIOhmjcwl4V>pNOwdpyYz7zuw$ClK;hE3Z0g|h@g|X|DJTb
MAc!=3xyWAr3;)cfp8x;=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/normalization_bwd_gamma_beta_instance_common.hpp.84C16E4EE625B892.idx b/.cache/clangd/index/normalization_bwd_gamma_beta_instance_common.hpp.84C16E4EE625B892.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e3664d2a8ad52c7bd1114eca7c2a07ef7f0308b9
GIT binary patch
literal 1982
zcmYk74@?tR9LKL6mTP<95s>t5n|MYD_KNMb&Bfj$?Ui~^n9;FC&}Dy=QK!s8EHZH#
zC2oWv3L`U`O|mfmj4T~E2yV;FmJAT$lFjK16Ee177Pf!}GaO5XEPL(pcE3yd>-+rj
zz3=b6-}^OvrnI>D(^QHIRhKqKBL{Y+P!yGcKlKM<dz}Wn#)-zR-twnvhg|Hp<=a2m
zbcJedZr&bidU2<>zh&gav4JP|ocR23<{I?}xn|osXR&AAQ`gnoGv=$Txb^tO>*D0B
z_uS@=w!>}J1-}noAMfa%_;umPJJH|b`>sA;eg1A?=cx@%3t6u9ulUXmRvvGR^O=)<
z-)tzzKOX<G;cNBi+fR2(v@iF~shDBK4ojK3b!pW{sbi~q`#*S4@cLJO->f?QFtx2S
z>%zsW*S74vd$Ru)R~6aRT7LiCgQ=FL?1s#3)kEC;+L`ydYqU~Nq`5CwD!(#+>ES=-
zqT!B6^W}-$v}Z@>UQK%;|Mu)M!+$u}v~{v)OIiH3tF^bbrf|8WGG4fI&+0#y#$A8i
zKWHm6|M>k&JvZ*;nWU;;rjFgab7A-#Gq^Z9vG7ICPv_f5sQN?EDmu}(%dLDv-|Vo}
z_}B`HWlgL(87MREEkzkv!`jV1S6FjXPFsy8EC`qaN;2qzKwxZP|4mP&)o8|oEK`;z
z87Q+6+b0~}I`ds*djB6)R-*+ATqc*247#9w@#v*9ubn+&HKt;LY?3|6pqJ>(T=cK&
zdhVLlNMk{cDMv{L-Qlz7qu@YGpJp{$u^`Ws=Sc=#u*_BrOx2CdSdA%IkYUO&Cxc!h
z7TH~o_y2!CR#}pLI5o7?{qZzf!>~4o%}I`mNwRqX$u=J#)fNJz*}^(;4i30*5(iv3
znFB7I$^jQn<A4kA;DHM-@xX<bdEmmUJaFMP9=He&!AUyHu!1Cb0m*_7kSc@#X+oG#
zcHR4vqmK;tEVw1KO#}y}IY2zk0}^OKBnAv?r|lxSKZX@)5s;g9i)4HlR-qM<%nHLI
z8UYH@L6M9q`8HURqe||OY!0UrQ%ce;540zIH%{^fhQ&`d`9YAfSe+PIq9*lmIx(mu
zO^$=piSZ=q_}m+Zc1(A*n;16RoSjZ~z_2+*IVI#BhRrwUr|T_n6}d`Cb%xz&-k45`
zGHjqIP(qGJ-oP~z6HjWAQD1LKjyOr<3qm%(z5@<%fOy0M5|H56cV~Cloun4NF3Ii%
zB-?#}klhbRwTA#{_OMQ(Lj<0pBmz%S7J;XTMBpi^BJdP75qP>CZl~Tix8(K$lHEQ)
z$n6KDx<i08cUULIp*Z#KE0W>`Br85ZNbv(wl@K6J3G0L$2zVk10Z$|&;E50do=8Q&
z6KM!|1|2~snG1#uN<l9mIp_m~f_^}1Fa$^oh6z!z$i7{P|A6|~%k?O+U=rcKhpipo
ZJ?oeeqbW>U`bsLD<SA;UzD(1X{{j>qumJ!7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/normalization_fwd.hpp.C0D3E8108CC38C5D.idx b/.cache/clangd/index/normalization_fwd.hpp.C0D3E8108CC38C5D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e6dac2397a662086464b7533b5ae4abb98ffe910
GIT binary patch
literal 2588
zcmYjR3s4hR6n%+?kgy>U@*~2Ml?V}*;Dp2&NJB9gLm*5g$oRtq3$`K%5<69@Q)`jx
z_(27=rC6;xl@16+kx|icR6vE1an$ic#ZN&*s1?-OYFg|D`tmk2*}Ri?-o5AEclM?y
zo6SNl0H&PeH3hk=vnK)oCQ)D4>f+@gOaS)#0kEd}+@7q$?&-p^SN7sYTdlA^oS`CH
zXK$)0X{^utqA_bhy-O}0xtt%d+2(%X_pph#S-3H0L09AYAA0W;2D}(LdY))jT{s2$
zE1yNTpYA*Fl6Qai0C2k=EeT4D9sMS*zOSh2by>D<xb_S?@3|{4!<aT9f&0%d4_~+)
zqgiXO?e6|6)KPXz&z`2*^YdU=!-Kg;PSkuZe72|Ln!7&vin%GA*dIvzx$U;1FXQy!
zy`hewMozo`k$GiG-m+bs-ke={?+4!=m@qi-D%JGpVLCDJX2j=y*Ug6K!&}6|AvI?!
z+A30-NAiq0@(<(GPnzp~n>SLPk;Z&)l$7u6!+aLMc0V}Y6<QKIKle<<?YUQ-mZy<j
zXCIsHMz_p-x$?rXh+}gqO3CunzCRXR{j^G8XC^1jQgd5M@AL%DIr*k3dGYrr?%FGE
zz6D2L?YLC*IAq54lIZq1F@|-U+|@6vx|14h#)b(u&#cU1kG9O8KWS^^(p9tT)`vIj
zwQgJEBPA*$8OoEIANNWuwbP`tre0gNcTsEQ^^P0+uqzX{m13)|9yc&r!*ktx^VXLo
z6_$V99yg~mKGA)@i19j4bvSlUOZeK_PE*UiQ(47aeQ!)z)xDLhq3gd|XB6IA`0}sV
zoTjGtZBbQQSKnn=<8#tA%j%v~c95qsrNrS%(`oIZ?xzDk<=0dP%KJlm%OmEG3eU|N
zX?62AEaAQZuC)bA!x#VrTsibEU+*6nzzmoS2#dfHJoW&-l#{4Wir;qs%Ne70F(8zU
z|3ge+O_>Qe+=-JEK@5ViiZSF~QE^X$fj~$QvV;_Y%@B+c*0pv1-BvUE7zV_U5Y7sx
z2!1RuM%cLN#=4!q`&DB=0tvyaV2Z$*FizO9XfgYCoGJ<fQb>qoMN$OJXPi)Xr6@Xb
zxX0$h<jZ)3maUyS_JB3?Wbb0fx@~b75W(|OmXtd0%N#%7x6yZGLDeUR{Fwe!4-&Wu
z0b>t%9u(&;bH(`pz<qb{!PhBTyAMW)En>U3bV6p4*}Yi^+=AP^MG*>%!tU)Np|mLN
z-i{EYg|x%!CN*r4eO5nsNk|2c5vx&fu|}+;A-&i@Lq>5T4VlDdgk&)?H4-J$$aFNM
zml<ftC`+Uvlgx||9)qh<F}McT(U2ZD(2x;Nq#+Y-MkqBhH5xh%BT`dR(-5*n+oscy
z)n-E|O_vr25&UIv#~cbTJ)G5C8N4DSgNYGxwp=uBQUEYw344hb!3$t=#_YTOqWfUo
z<9Inn<O*_4FtAOxyHFgxZ6_5eUKt;YA|-v4v<P}6BP}3eV3&c##t0Ed#7FH8;so(g
z+d?=Yd{llICyWmzU_=Bbf{#29#fjoW4`76qXAOjUU~EdX=aoXD(*@~X(XQQ-i7nOb
z2Yo5qNX0661|wurnF{H_rML>|QAiakq(>=Ls*oO1N~(|^ZIX5&YD9u4AqqAEo}kou
z-e;t4xnQ|h9ZX4)Jnu9jEfg&Dl1dMMSz@bcyUeAWAxgzz`4m#8#v^1C*=Tz<t<8w`
z(nM*rJ%`9aM`{Nxic{pI?K!nhT9j;2ws#M<fKu)mHb{AapunpfdXTcDXX-%G#oq00
zy;0?*l#v410bUbgL<&2_i{RZ*2|UvXp2dhvL8dnWY&MfSHj!XH$|njd3Cn*no%Nc{
zmu#ghDdveKGz;)#5|{=fa2_s!_kj@#o<c&i08c4_7m5)ikCedrsZ>GMFqrCH^S%qJ
z_|6jyY<$B0Z;gKina6tyTl=ZmrHE0eQCx*ap`#(a!azetMIsHE6lR2!F-kSkqtqyM
zG^AG=XvnBcq#={ij1U<^s!<)JhSbrJo;1*qkxZl^6KO`s5$T}KE=P)kHoKhBPTK5p
zTAj4nm95M6&aUFz{A^FJT*V)|9G(U7w)%Q?N3Qw+J8ViY9ug`CQ#^l<e?s5_030oY
O(J~Y*<!Bj(mj3}j1#u(*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/normalization_fwd_instance_common.hpp.38A39908E71BFB61.idx b/.cache/clangd/index/normalization_fwd_instance_common.hpp.38A39908E71BFB61.idx
new file mode 100755
index 0000000000000000000000000000000000000000..821cc2330efb068f6a7bd8cd28f78b0f8deeadc0
GIT binary patch
literal 13126
zcmYk@32<EXeFyOUi6H*Agt4(GwQlW`R`1Dv&-Qy-y{Fyxw9=C}jC#=MzHdlfbgu3*
zv_>I7b>DYr3A*n)6r+S@CJhafbZANkro)6}LYo#c9SRwmPFI4R{d&eLefRsmA4#jV
zcK6x-`mIMFO;RkD)XH0})m4pUU$a;&mps4SjSYYPM`W>FzSm-Do%#EWAJ=^U#>KZ@
z{qT)T?^quXz4qo>;nh#?yH;_1^M#V_uy4HZ`Dcy~ej4<|t*pO&qVL0(o_^}-oaThE
z-#i-s&lduJK5#elw|UR{Uq1N5QrBDeeeW+{9{l8o<^T5&fBdfbi$7(RF{Q2_e^<TM
z+m`eG@BZSsooAA!D*wBF@Y_Q_W47G4SMPb7%THapbN~5;rNZami~rtlEH^9u?pNpk
zt-xB)XL}}5x=`}38*hB`v6DYJ?q0s^eDAxTsjojhcKd(g)8BGFbt1g<%i1qKT5Ol^
zM(4fk_XNu;xW6*&@IzAHd{=D!;r%&HZzSJ&=O_NJKl#QJpZEu*aG5dBJ}<Amy61l2
z#_?-f<uBtJetl=5`Gro`zpPdM<41SDrS*OGY~J$oKiakpJ@%8S@!vgM@?5~%UGIPD
zN*OoiH}3o3>)RDS`1z~-k6(W6^Y3&jU#z5*a8FbEbG!EH%?JMJ!pBAKw(9Z@&uvP-
z-+1rO`YZo8!0-IMmY_T2e}}Gr^~Q3}&vSr<f`}!5eDRT2fALX(U!dopSIGYra(?y3
zchIGcZwEK6|N9RD{O<7_bjknHxw8w)IgfFsJd|=P`1bPXodCc4JO{i;{x4GeS8seu
zRO_{u@A64pfZqkrLAU*HUq8F>9bhh(hJW_?Z{H5^yWexbb^q(v&n|oi)ct3(`{RIt
z9|idNdk(lu{&zi}?3ZtR2i3om=KjX74~+o7bDo19@_&dwyYL-wckPqH*UtrC3-Du}
z18(`>x_)-yJK$YQb;eu8SN|)(FTiualm1WgXBWN$et$80`~BKCe-+^OHO~PL`agK?
z?80|@SQ@IHa({Z=_dTHVU)}P5{rxW%RzGk?dT)16@#T%~Exfn&^M2>gZe;KEfk9We
zD>m;J=SlpPtB}MiHz2839FX)YE?)}ff<X!9gFy)vgFy*agFy+_gFy-5LO==OLqG`;
zLqG{pLqG}9LqG}TLO}`TLqQ1@LqQ2uLqQ4ELqQ4S!axb*!$1iW!$1jB!$1ks!$1k=
z!a)h=!$Aob!$ApG!$Aqx!$FDQB0!1YBS47|BS48zBS4AJBS4AdB0-7dBSDE2BSDE&
zBSDGOBSDGcqCkn_qd<ugqd<vLqd<w$qd>88R#2?G6%?yz1;wgbL9yyqP&keQh2uF;
zIFSQ|Q#nvLodYGBiv}f{j|L@Lj0PoIjRqxJj|L@%ivcBuj{zk{i~%J^jR7S_j{(KT
z*+8-JHc)J$4HTPd1I4D>K#ApIL5byKL5US(L5WpkL5bC4LE$+b6rSfn;YA)4UgbgI
zbsiLf6F?Do0Te+LKoL{{6hRk2iR0oxiR0ryi4)^MiBsc1iPPgiQ3U0h_u(KgNRgDA
zkQ5~WlA$C+GL=VsDPD*NC0>dLC0>aKCEkb!CEkn&#V*)Eu}gMP?1~)}yI}{#ZrVXf
z5E4L1kP<*iP!d2%FcLsXFcUyg1r-!kQbAD_6%^G_K~YT=6iv`T(IgENP0>Kn3=I^`
z)IdoT5<y9n5<y8+5<y8c5<y8c6G3qZ4p1DD0~Cki0L5WAKyjE3P@IAj6sP0_#i=+!
zaT-ogoTd{LUC=?%B^?x9(LvD-9TeTvK`{ga6hksVF%$z7!!STGOaqi8AqkWuDG8J$
zB?*)yBMFowGYJ%z-~z=Zxj=C#E>K*C3lx{>0wq~U1|?Za1|?ZZ1|``@1|``{1|>yE
z0VPFB0VPFA0VTyq0VTyu0mT$dP)x}L#Z*jCOv41lG)+)ag;Y>drBqN-l~hntjZ{!l
z%~Vj*gfvjnq%=^{lr&J%j5JWv%rsEaxpbTNIcZ=}I-h<Ol9+x2lA7*-q^G-lsgJlm
zvU#6~2L|<#+D8sw?kBDv`+idUu^%9A0Q&(_2e2O`ZV>xHQU|dgB5nx#AyS91A0}=X
z`(aXtu^%CB1p5(EN3b6yZWQ}bQb(~LBW?`)F;d5{A17`c`*Bjov7aDr0{aP4C$OI+
zZW8-RQYW#WB5n%%DN?7fpC)b^`)N|Av7aGs2KyOOXRx0oZWjAlQfINBBW@1+Ia24a
zpC@h}`*~95v0or=0s93~7qDL>VUezRU$z7WEt0WF$-cZq!V>mNWGrF7Ou{nu%VaEL
zze2(a_A6wpV82SjD)y^ntYW`L!W#B#WUOJoPQp6&>tw8Bzd^zV_8VkuV82PiCia_T
zY+}Df!WQ;hWNcx-O~N+z+hlBGzeB<f_B&+kV82VkF7~@*>|(!1!XEZ}Wb9$TPr^R-
z`(*56e?Y<k_6KAfV1G!$A@+x49AbY&!V&gIWE^3COu{kt$7CF1e?r^|*}Sjt1A|UT
zJt2oTdw-63UM_!mcES6y*n3jJQ=Fbr@LnzbboM)!-VJ|2^E`QE3J*^SSv({=vqMT?
z36NB#LeiK9DUl^Yaxe!ZCv!s5nGVTd2BajG1j)r*kUW2~Af>PrNG3BOrLt5=X)Fy=
zHLK>mw@_eE4Xc4v%W5Iju{ucgtR7MWYk=ftZb*%+5mFOtg4E2KA+@j;NUf|DQX6Z7
z)Xv%=b+8Udovag57wdx5&AK7=upUT@#1@IiA6IM%^D?nz%qzrJFs~9@#k@vr4f8s&
zb<7*YHZX4z+r+#@Yzy-?v2Dyd#C9<665GYRM{E!CKCykw2gD9A9}+vnd_?RB^D(hw
zUrvhJ?3$s!dBf>B&%~}+;p`C$h7`g=Ace9}NMS4tQaB5T6u}}OMY2dpQ7j6Qm02Ni
zjDr--q9Mhw7)Ul|gA~hRA@PidBrpL|9E*dL&eE;EXIL35!|FXX7#NhvG9hKLEJ)cb
z8&VF-ft1T~A?2|=Nck)uQUNP~RLBY;6|o{n#jF@o2`hnA%1R-Xu`)>ItQ=AWtAJF=
zDj`*|DoDM=ddccNzkK>!ll76+`!hN)sGnFr<^f^@m<NdsVjdzkgn5|QFy;|rBbY~t
zjba`nHimhe*f{11ViTAriA`djA~uD2n%Fev8DcY-XNk>Xo+CDgd7juj<^^I4m`{kE
zkk$Lx=6!JTJcON|^Y%W3d27$rU!LOj7T!ONp5pZEw)g69eey_X;@d9;Q((}gi<kbu
zdmee^KmMvM?}M4XbAdq*pL_U%cN`dW^W4n~-VYrZbW6RZ`|j#b-gwgPyQ|+;Z|lBC
zkvpW`A>H@Qf06VTDaH5Ae}z(Cp~rncY41|%U3%R6)p$RK=aWA@^X~oRy&ueT_38O`
z@73ND?W^~kKRd(jz25U<NxowDeJ#OqFr*MU1X8FR3MotugA^`@LyC|iAVtcNkfP)$
zNLJYjiIX`<(Q-7T7&!)#O}0Ubm180CG7m|R1xRsn93(|koWAc|ycQ41uGt|aXbF&1
zO@*Xs8l*%m5t2i5KyqqMNV=v&GBg8Hl9mL?rMVy_YsrvOv=m6DW<pBUQX!>jX^_(8
zbi4OH6d07jW!SuT9Z%vju0j$sZa|VV;2xc!W;meJGhDuu$z_6>$!CI@DQ1G1DQAM2
zsb+$isb_+j#btq+#b<$;C1!z{C1-(|rDlPdrDuVe&1Hj`&1Zv|EoOt6EoXz7t!9In
zt!IOo!{vaP!{>mRBj$jaBj<pbqvn8_qvwE`%jJTZ%jbfbE9QckE9ZiltLB24tLK85
z$K`>U$LE2WC+2~fC+C5gr{;l~r{{s0&*g)e&*y`gFXn@pFXw}qujYf9ujhkVz!iX5
zz!!j7AQpgGAQymHpca5xpcjBy$Q6QF$QOcHC>DZQC>MfRs1|}*s274+#1(;A#20~C
zBo={LBo~2Mq!xi$q!)o%%oT%K%ol@MEEa=VEEj`WtQLb=tQUh>!j*to!k2(qB9?$z
zBA0+!qLzSJqL+YK%9Vmy%9ny!Dwcv-Dwl#;s+NLTs+WRU#+89t#+QLvCYFI&CYOO(
zrj~(Ork8<P&Xt2%&X<E(E|!B?E|-H@u9kyYu9t&Z!Bv1+!B>D;Ay$A{Ay<G|p;mxd
zp;v%e$yI_`$yb6|DOQ46DOZA7saAqnsaJwo#Z`e>#aDq@C02o1C0Bu2rB;DirB{Jj
zEmU9gJ*%mfs&7J4suLh-)lNu8buuKg`Vn8M5o*A!k!rxKQEI@f(Q3e~F>1i9F>AoA
z6>7n(m1@DPRcgVk)oQ`4HEO}EHEY4F6Y9XMlj^{%Q|iF1)9S#iGwQ&sGwZ;t7wW;R
zm+HZ+SL(s6*XqHnH|oKxH|xP{5E{U2kQ%^jP#VB&&>FyOFdD#YFdM*h3vMvok{e97
z;s(>LxxsWBZZO@Z8_Y(b5zI!Z5zI!V5zI!d5zI!T5zI!j5zHo`3Ct#`3Ct#?3Ct#~
z3Ct#=3Ct$53Cw1p8O&y>8O&y-8O&y_8O&y*8O&z08O#=;1<V$y1<V$u1<V$$1<V$s
z1<V$+1<Y2V70gzt70gzp70gzx70gzn70gz%70foF4a_#F4a_#B4a_#J4a_#94a_#P
z4a|0-9n5yA9n5y69n5yE9n5y49n5yK9n21)1I!Mo1I!Mk1I!Ms1I!Mi1I!My1I$jL
z6U<Jj6U<Jf6U<Jn6U<Jd6U<Jt6U;853(PL53(PL13(PL93(PK~3(PLF3(Rhz8_aI0
z8_aH{8_aI48_aH_8_aIA8_XV|2h1L+2h1L&2h1L=2h1L$2h1L`2h3jLddcQ{4ctq7
zFI~kf_R<Z^axd92tG(pFtoM@3m-~qCqpQBx)O{rO(GARUAK5YMedO}xe&YLa-%nyc
z?)ypZ$9+HP{kR_>egO9aBo5$yfaC$(50E~9`$6IdaX(1nAnpfA9>o11>4Uf*B7O+>
zLnIF2eu(5D+z*jHg!^IQhjBkl;xO)qNgl@iFzLg%A0d7O_ah{Z;C_VU5!{cEK7#vE
z;zw~mO5!N)M@b&V{V3_9xE~{a4EJLsj^Tcc<T2cjkv@j|apK2uKThH}?#D?U$Nf0z
z<G7z7eggLsBu?Odg5(L@Pmn%=`$^&_aX(4oB<?3kp2Ynm>65shB7O?@QzTB|ev0HN
z+)t4{h5KpZr*S_`;xz83NuI|2H0jg0pCNt*_cJ8U;C_bW8Qjm1K7;#N;%9L`OX4i<
zXGxyL{VeITxSu0_4)=2;&f$KJ<T>2Wkv@m}dE)1BKTqO3?&nFK$NfC%^SECiegXFj
zBrf27f#e0;FOa@~`$du#>89@-;36rDlz>@VBqwHbksk5oC6bnKzeLIs?w3eg!u=AN
zOSoSqX&Lv+q%7lpnY3lxFO#{9`xTN_aKA#z3hq}(TfzMbnJc(oC21A+tE8;rewDOU
z+^>?kiu*N^)^NW@${Oz1NL$1G8kuXjUnglD_v@sr<9?mAb=<F$xsLk{k~VO^LCOa1
zH%QyS{RWvExZfmc6Ze~>Y~p^Cv`yS^lDUcdEt0lyzeUOx?zc$W!u=MRTe#mQX&d+3
zq-^7Uo3w4*Z<D!=`yG;YaKA&!4(@kI+rj+~nLD`OC21G;yQJ*mewVae-0zaPi~Bv2
z_He&P${z0bNZZ5x9+`W%-zRAw_xq&m<9?sCecbPpxsUq;k`8cxK*|B`4@f(}{Q;Q=
zxIZN65ch|q9OC|vv_sq<l6i>xBa)7Ae?-a=?vF@2!u=7MN4P&G=@|FNq#Wb^n6zWu
zACq~E`xD|%=&JYcEKdJ1f_OqVFv}-o$E=@_%bPv_`u~}-)BhRaZm4iScKXi6(~GlT
senF2vkn-85uND66-UlvTdhoL40dHfmJm`D(^#8_u_3N_t9gF4v0reiy`2YX_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/number.hpp.25A61C6490A403AE.idx b/.cache/clangd/index/number.hpp.25A61C6490A403AE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fa72a7b0af8f98e6e655fe174f101475be930322
GIT binary patch
literal 450
zcmWIYbaUIq$iU#7;#rZKT9U}Zz`!5~#Kk2=nf*Ze6%bd<S=;Bxb;v-#wOB-ao$Q^q
zOTr%O1(v<M!86y#Y0@g!J-dUB2=Dl{XM07>W!d@Zr>!m?-g@}ME8p+U_hKfUJ1pg{
z^~}0%<F%N`6HE>s6BR<{Z1}`mv@5u9g+}yKF1w?1wECR#7A{gb@G5mK#~-Hb<cHg8
zLl*B`U@?D-*_ARb%l{kK|9)-zi!Goo%!Hx1GB?Qy=q6r+j#*h#4R3QZv9L0*bF*>t
zfC&apCZJ~+7}*)`&v>6WX{D|kHxn~Z0UsNm0GL21C`wH$o`GZ%BNy0&4fC}3r|z1#
z&E@2@S0TTY&6t1&^K&b}?BikQ6XsKa86GOhs<uqF=nqgY3kQoR&@?c}R6Zo4z{|Is
z5vW>~&k&}%+l#s7ZU27@pfI;0w+75em^9E3kTYQpC`!#q1esP`lw2$hVz9rv@nMF=
P9M_YKYz%B50SEv9D(aNp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/numeric.hpp.E1739325ACADC629.idx b/.cache/clangd/index/numeric.hpp.E1739325ACADC629.idx
new file mode 100755
index 0000000000000000000000000000000000000000..91e8596204629f302e6ff059de29ff9e42a3af91
GIT binary patch
literal 530
zcmWIYbaUfkVqkDi@vO*AElFfyU|<jg;^LB`Og2UahHF4vF=y!nLw{yNp4QL7Mj|i$
zb|q~!NPQE~*~$1#>+H@sS7T%>IiIfhb>H&VV%Y@ePmjJod2bsTy7}>;HIaN_x0d%L
z8F$~CxTGfQ%bOi0`d>P46>^F{Jh&%|t0YqHY**9Nhn98nLhNThhOB<&AGbJvg<HHv
z??2uZQ%l!<IeNQi_JTY2B(5&{T-gz^^o~^V=hdAx)0pd?$UmGDWFCBiW9`*G8!3ht
zJKSGP%FSI=-cjRl^G3qo2^x)$?TR)<_s_QVa^Ro0&B%;}>)dp)gKyUdoqVFnFW}Yh
z^lo4G$61FAn@c`CyJKVgyQOkpM3X_fhG*WE-A}V*By4Bx`^tYL#HE&TUh-1q;>z44
zMPR7JK5mKFnwCF_hk==efsKomOAt&ja56D6ure?L1FR@Dt+)-S(sVn+^eztm9ZWoI
z+&bI_aE)xd>b%-89*kz>0voeop7#FKT@$wztrL%kIB516q*s(%9;S+ijaPtI3dWnz
yc9HLf*_1ta6&0oCBmy17P+XK;tPEnXg9r#<3$>Cy;C_lPhKZenfdeMWzyJU)*vZNO

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/permute_1xHxW_fp16.cpp.5780304278E72678.idx b/.cache/clangd/index/permute_1xHxW_fp16.cpp.5780304278E72678.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ad3b77bf3951ab9d0694ba32cbc7e977d7dd025d
GIT binary patch
literal 1446
zcmWIYbaR`>%D~{9;#rZKT9U}Zz`!5~#Kk2=ncmC{41d`f7%Jwh4ZU4-+k)rn|B{5G
zv#;cq9pm45qj$OQqnFDjyYsp@i5(YN`RS4O{(TEnkM~Zt%Td!8YIK%iVqY+YF@<5<
zsx7XZFI;TnLS}y!KJ(z$t8BxayLbFC)6`pj_~OLJpN*@n>~9&&o&T@+iNsu4%Xz_P
z-pNgR{I06#=VNaDx<BTc&+flpe)MRKncSwu8hL4-lm5MZlvwru@3tK~QqN}}ID0je
z{miFZW>c=qsa9=06}?+6e1YTMucw|LGF$sKB1>!96P}lQdc01@NG@Ew#Qey9+o08#
z7T%VVc;|0i;j;GU$=%XMjo~RvRyibl-crA8`TA0|Yo2sgs@TaJEXU7QuUV6Az^b{J
zsp698`rg(_x~WVb?2O_fFYi3s&i{PAqUttot4|KzLcRfu+Sfc({?D;3XA;Z)Fvlwj
zD-W%xmhxPy+}N1t>9qLpfx><4xhaARO1Eu%*x6?`Cpmz(u!Uc*D`D<^=NQ@W%PW$l
zYchT*U)E_C_tf)YF}ao_Ep1z}huuZn`ND$7bux4IpNNcD7_2chA!I{mo?x)`uNJAN
zw%q(vm2?#|*9K_w*&Nz=B;oM3rT6l87)?`AWvSnhJ3ED8tHCtZw>7$6l@ZnaqFQ^)
zEj~R7JF`SZeW?ZSc6O(dw%1cG6jWPmy>jv42Ityb+3)gh=>j@SR`2@!X{UZk_1cve
zb$Vl^BNiL4+@<%{_Sr@yjpWdY%`XZas}-vJOnQx_Zffoe|9rF0!FP3KlGXEB1)QrV
z<arCNT(6yZY})eq6MG&AZpu1-Y5D=@*JVHSD&ts|i}BsI3UoPq^v%8Yx|4I?a=6a*
zOX@1&mfziD^x)w7XPY_ggD3x5uX=oWG~3c`NhdGZ<!c1$RO(sku9Xr!rY?Dd=T?BR
z{p7;)IVC+_Pmfi8(Tn|mG4kB@!#bZH-h7a9KtA~09gEU8!ECeN<i;D_Qb>3C`Lgoj
zq--VcU-g}FK~MhfmdML9>vz4%=gxd_TjV+C3D(+DKP)U{=R8?zRJk@vantJPm`hXA
zuf2cS6>bsHvp;}8r_5JdHRIkv<HyWJmp9Mcd8~myQ{-K8MO1jrmHn0f_?9k<+?p0p
zT>0Mpm-yd%?-|6;PCl?i^X~dZ_st&2`y7isDHWpjo0m7|!W@R;%G{(HVBS)Tk><_*
zDeA$^<j24e#L31f4JLpxU|=AAMdI<hqDUTwP@uFhi!cY6fJrav`fC?$CtJnM6abWF
zV`CEp6ENwE4y)MIZmX~0X7UG0v$C-YfC-p%QEFN-CnLfQj9ggl1`2~LhUs$c{q!+w
z-Jb|1o*-5cR&kgeJV9(CY~nBnWx4+IW0Iaz#>f-IX~XFY^Ce80i3`P@JVAW&d|EI?
z8(6;Po-sY~9;la3htC`)46_KR7vvL|bAXD3`D9^=ek`88<f`%3jX=G^qQVBS;QR6V
z-bB4;2}(dW^JVa5!c+r!t&Dtaj4(b!QEE;i$l&6l<YHwY%^bv)@^SNvWItb~z+i?z
zn4Js^!3bt@Zc0u%kn5G|!_30U#?HaX&BH6e&*vfN;-n_X<sl?4EFvLkVQFn+YiIAE
zC#r3wY^EWPO;k);Mpi*lNkvs%Q%gtRz|h##T+T#F*T`8^%*``|$<^1<oza^C05>)t
ABLDyZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/permute_HxWx4_fp16.cpp.FCA2351A6D6E3584.idx b/.cache/clangd/index/permute_HxWx4_fp16.cpp.FCA2351A6D6E3584.idx
new file mode 100755
index 0000000000000000000000000000000000000000..989ea697de84ba54c2efc7265caf0abf7804c5db
GIT binary patch
literal 1482
zcmWIYbaOkz%D~{9;#rZKT9U}Zz`!5~#Kk2=ncmC{3_sZ!7%Jwh4ZU4-+k)rn|B{5G
z((7-R9OK`4L;Ac~=W<{5eyNTYoj%Q=)7|=ie>F{uG57nmVdfqW7QR9bnT8;?35+*Z
zT@q7WApA8YGFShcLG$d__1!Q3KD|0Uv(Qw!&tGQ#yqlYko@~4M=+AoL;}s>3D*O!h
z*ZcJSp8fLb_xg&De{XMzsHvSHI=%a6%89;3JLc+l9i4vq_|3!DcR%{nXS<$p`}OOx
z2IsSH2gFrRd2!V$|7}}LBiq+=+Vi<@Z@HGda%IFs<0YSsv~-J8Tlyy56ZrqlW35+v
zUU^D|#rX%ETc4@F&pW{qH*w-Bre4*|`99Crc)vH@8M|U?q*}7HjQRVOo7No^S<=Jx
zz%zW0nV8qMDVzs>pQuRneJ&<%7gN(T<%ZphBh0#<CXStQo7(#S=-$}mDfGvUHL!UJ
z*W!1XDqH(l*xFP%d-<4;|CHM@(Y4{)&0}rxC0m}#G-q`v*AxV@y#B-Z=+%x{ZgX!v
zaK0=*GtF|3M^UEIf{kg{Z@gIe*j6b@vWZRh`{NgX9M2t+n(MTZWpR^l$>|xlE~;PA
zcpq7@LL`&ZGrKja@B!yJRW|7|-LG|rr*2fyRQz!^_qGc2ilrNMGM;A#o_c(%dPTs?
z=ED=!x2LJ8T52B6+`%lc$T((sOWW<kE0^~1wadQBe!b(rft=%p3F}_nJN<m$g|}Om
zxotKoOHOD$x8&8XE&MZ%cP;1%b>&-hTvfK=h2=$)a~Ubie>k2^w`MY3`>gZj9O*-%
zYn*rKdM??wZi!^{yno#(Ed3KgtEcT_k&D`X{QJqzN)}s7Zd~hBlG022Z}-jnU9P*v
zs}iB$CDA{wq)uVvuiHAq|L4qESL$^w?d}?BY!%gR<EwP;OgdV8GBfk)3Iog22huLI
z&H6dBr#fwMa)_4s$;IU_fAzat-sjtVrhW0jn+^8M?(Fn@b>leCwwt&2rEGCM*YW9d
z_afEnF4MmJ7uvhz#P9OL-L<#+x>m{du`kNqJf~fuJj~?7_Xjlv)25$#vNfyq#JY$A
z-~N@!*UNV-lI&k^Zy>(@!bZJ%i`0b8(mXq!?^?K3Xv>RlXM6ciG*1Zi`Dc0l{=#!6
zE3WZmt$X{|<G;ea-?9%D{Z!4|DZcMTIq#3>lM3B8u}wW$$MjtHye30&Wo}XpFmI{F
zNb_d@6!qX{3SeLe;%4KP1`|LTFfb6mBJucLQ6vvTC{S9IMU(?fz@!_><TQ?)pE-}4
z$sZ`q%ErnECScOHjxBuor@Ot4o5>d_&Bn$i048A4MX71UdW;A+FmhqF8z>C27^cg$
z_tVF$b$=old4gDlS;b&Zh~wM!|MU*QNkCyXVKy<C)mg6p{FtQYlmUggZMc15-i7I6
z;zDsSPY{O|hY`%C4J==C&zPQg57aB5BVZ2G3p1L93&l4;y`miQFug!6A_8(SEkK@v
zfEJARdi~ngJAP*rfaZ#dit59{_{Zye6ZM`YC@}$5X9#4%R0Dafi~?<pFn&>LP9n&b
z;-chYWe_6}M1TQv(AJM%o_~wzSj-d{%n%F{W?)FpP02|Ia=lW0m|0la*f}`4d3Xi*
z`8)(&oYVxlJcPuBMI<CGEUj&9?d%=&M76Dy%{1h(iHb?f$SNo*si>-JYU$`37#f?J
Y%b7^&8aa!Kxp{^#x%xS}GkP-s0Kr5^fB*mh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/permute_NxHxW_fp16.cpp.6FE6AEAF71058087.idx b/.cache/clangd/index/permute_NxHxW_fp16.cpp.6FE6AEAF71058087.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00d66a9b695063e1fc21b838bbb983137863080e
GIT binary patch
literal 1446
zcmWIYbaR`>%D~{9;#rZKT9U}Zz`!5~#Kk2=ncmC{41d`f7%Jwh4ZU4-+k)rn|B{5G
zv#;cq9pm45qj$OQqnFDjyYsp@i5(YN`RS4O{(TEnkM~Zt%Td!8YIK%iVqY+YF@<5<
zsx7XZFI;TnLS}y!KJ(z$t8BxayLbFC)6`pj_~OLJpN*@n>~9&&o&T@+iNsu4%Xz_P
z-pNgR{I06#=VNaDx<BTc&+flpe)MRKncSwu8hL4-lm5MZlvwru@3tK~QqN}}ID0je
z{miFZW>c=qsa9=06}?+6e1YTMucw|LGF$sKB1>!96P}lQdc01@NG@Ew#Qey9+o08#
z7T%VVc;|0i;j;GU$=%XMjo~RvRyibl-crA8`TA0|Yo2sgs@TaJEXU7QuUV6Az^b{J
zsp698`rg(_x~WVb?2O_fFYi3s&i{PAqUttot4|KzLcRfu+Sfc({?D;3XA;Z)Fvlwj
zD-W%xmhxPy+}N1t>9qLpfx><4xhaARO1Eu%*x6?`Cpmz(u!Uc*D`D<^=NQ@W%PW$l
zYchT*U)E_C_tf)YF}ao_Ep1z}huuZn`ND$7bux4IpNNcD7_2chA!I{mo?x)`uNJAN
zw%q(vm2?#|*9K_w*&Nz=B;oM3rT6l87)?`AWvSnhJ3ED8tHCtZw>7$6l@ZnaqFQ^)
zEj~R7JF`SZeW?ZSc6O(dw%1cG6jWPmy>jv42Ityb+3)gh=>j@SR`2@!X{UZk_1cve
zb$Vl^BNiL4+@<%{_Sr@yjpWdY%`XZas}-vJOnQx_Zffoe|9rF0!FP3KlGXEB1)QrV
z<arCNT(6yZY})eq6MG&AZpu1-Y5D=@*JVHSD&ts|i}BsI3UoPq^v%8Yx|4I?a=6a*
zOX@1&mfziD^x)w7XPY_ggD3x5uX=oWG~3c`NhdGZ<!c1$RO(sku9Xr!rY?Dd=T?BR
z{p7;)IVC+_Pmfi8(Tn|mG4kB@!#bZH-h7a9KtA~09gEU8!ECeN<i;D_Qb>3C`Lgoj
zq--VcU-g}FK~Mhf-f{a*j$HRD**>;Kxtr&-E0l+sd?+livN-9T`YCFwQ%d;u9ZNLN
zuepC&bbY}FiTZ_dH{Q$&^SZF7IsF*>lV$0~#l4L3mo)Ai|FC7<j}`Tw|HyjzZq7Qp
z;Mu2p{a^Hd?Y+;aH%q<IGk90L>;BvW_A`2RPcmKM^G$lO;6(w3;>z5l8era1i;?Ed
z{weCg&E&_x5X8yGDGeroGGJgJensN(yP`-QhESliFpDq;n1D$y>iTOJZ6{mB%@hEX
zW@BR$1QRgniw>*U)NZS<;AZj%O0%-D3V;cibWv(rF()I!4UAk^?FI^iEQaZF?fvvI
zYu%p+CY~Tx5ms@S9Xvs7B5dL?2W7eb^J9{pQ^v>>#A(Cn3-cvRnu!a=ojgH&@_bq_
zMH^VY=AJP<@gAs`PlwMOCJeI(s2AiDm~()Ng!yD)iheAfzT~R$){Q{D!lJ?ku;Bag
z`rbsnX9-F`H}hrkWx-Sfd993mZHzEJLs4o@BFNz4qU2&_Ak7>U`Xe?wpL>20Q(!Pd
zAk0n%hF}CUIX5LI9mw@c^<idVWn<^y<mTZO;OFxYba7G><nj;_7Z#C_w6L_cv9+^z
z&=b|RQa00&$0jN!EhDR-sHCE*uBoM?Z(wL_YA$CYrEBCYD(2=H!sP1f=+5ZP000j)
BBYXe=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/permute_scale.hpp.CFE8A4ABEDAA7993.idx b/.cache/clangd/index/permute_scale.hpp.CFE8A4ABEDAA7993.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8d2b410014e692d75cb7fcec5df1e8595592834e
GIT binary patch
literal 5034
zcmYjU4O~-IAHIKT%;AMiVCYS7^J54K$i2ed*6wzP83L08`~pn{!~{hfAN?}Ts4y&z
zmNG3N-u!?{A>a3BiSi3xN&}>V3R8$TM~Wt-iIpF(mcH+u^Kj07!0xl>;hgjQKhDlc
zON@`-A3#x0<Rorbx2ibnQHr9*NWVqJg?Wiy6!ogaZFsM9SF)ojT2*!@Jh<=Bu7qam
zi7&Pp77Q)gls9eN#`xtk6^p1%_jW8?aI@tdYwwF|UJHCaZkCs0>+=)5TN?hnb5f(3
zUY*?phZ^5I`|Srg-+KG+yHxBwKK}eBKbiCHYvJ#9$F_~0T==f*uhgrJmj_nNKKNv4
zW5%j~jPBgtaV`1Hl6PjW*Szn$<L5?)^WvFK$3XAB8@IQg7L%fbopwXm!1=v%%S!!!
zj8l}(cW(U2QJTLuz3^?{K%3u@SF$_e_>>6aDeteoh<n6Pa`%H#-f?66Ud<LpZR^m6
z?XL>IS{=Q4{!iD&Y`UD(A9!rjm#(~S{gYRAELr@-!Q*p}Q;9JHy0*M!ngQEeTMOf3
zGV*0VoH0MY{zk^du+6Tdjq8&OVqB$v$F#nwF1Q*~`^$x!S?}LTnU<X7b33Q!;KzSH
zQanD*FY$CuR`Q|0Pvj>2@Ra`d(>X(P6jydL-pjbsuFAO8o&Wmencj?|pC0^lXB2g3
z%>1>Usa~7bRNz|PxGt)#I-q~~hQ-3cx>(VF_D-*uQ{SH(>?)d(eScTFT^F?cjdJDv
zkN&-~qww=l``#S;GWTTK{$6jzeJ^gJuH)FY@Yyd+JU(gO!k)8Q+o>_r^VX+Xi<d21
zF>_+XsOn<@A8(0Eo^$)yqkiRU?q4uDUEf~H`MUi6`!{cCw!WqBd*Z9eT}8L1)Rq77
z?q1*Oil)2!*4?=@_eg%w@Mi(PRTf`3AK(=gvy)kp`gr$ymwj%WQoeoh>B$E!?(1&<
zdFgZe*1WZ7YR<wlKiYRc{y$1g|1E$jdSPAm??6$(-J5%VIN8!iQ`%`#p-C@xo@`ze
zr3f5F%?t+?w2N)CUhu!Rnx>{Afglqkfxov$U@$Fftje*PXiA3!F|rs*Fh=GPSi&4f
z?A~Qgnwp3NQ)E*lf$|ZLV4M1x7ypy^%PpD;LxN{z&q{)Ev`4W0`?`~-efGIVdp$Ot
zV#~m%4E!IyQKLMoKA<)=E}ZxHW}2FW##k&{ERCU7c*dwZ_=+aMcj;3!r9pyVS+FGV
z^YREj>AcXoX^QJPZ?Eyw81^OkOS%Vt&;%{26Adb7#43NUU}+8Ipel#phi^!7cC=j?
z6t;a5O@*K-0%d{H6u#h@qVHBq&sQaLF3{9uB#4kjNCMTP9>LAjy+_(tMf^xpQ;=YV
zY=tBU^6?06M|{2~&S%RKnwo$Fb7XTQL4e#NxEJUhdF0NxEi@I11gmAMCBfLy9syOj
z>Y1WGKZ^3XyVq4}i=9u?>{xEBd1UddhSQiwqQy?-rkY1G$j;zqm`7Z&dQNX1Nk1!c
zqIslR*d#8=JhECgjY~6+lp4E)TVftb7@N&yn@5bZYq&M$5ff}7S7=7%G3K9dJoen-
zUr&ysS+$@R@q&YdAOeL5AtGM>1YrV!bb?OAW<rHf0!0duBDNMKL=h-fh!wF>Uf>C2
z7R(~HZ4oR4ssu|V#G%rm*(wmLU>3%8V5uYddJyYLz6mT%B;O2TGs$;?rIX~lK<pxU
z7g$^*-w$Fx$q$2NnB=d6cpdX7VP1*{dqVk%zxLLjH)(11Lr{GPMq~q-mt^jp0n7dX
z)gM?EDr;~CuMJrsr`dRUyxP;Y6h*T;LB10Rd;sJJNc<_tKP7Q1$XiK#7UXA1d>Q1I
zF-B>VR_ET9@aoQk>O5YZVZ-a|oi9(@FLlE>>o^-OF0EB-!{roa4YT0_ind1EaLE{~
z1{<1<X02AM4Mm1#ZC0BNcf)+^d>aZb%`UYrwc&19X<cc<-LTfW)`q*G$XaAW-GEjh
z`Eu{3*jEdvTClIAJuQx|(yPAFQq%Bi6ZRFv2N5WQ4>4hv6Zi=P((yVInvG^d`A`By
z@{uMSfhazTK(TzR3CDxyc><Yvvk6DX!dnQGz$ch+tP=S|0;TXNCLFm`K9xY}e7Xt8
zF@w(_P$r*g!cooQvj~*S=bCVg^Y}ah*?GGOMVw{}_yPh|fT04+s0(P;2}UPaFs=kc
zCCOKTv5Mqtz)(Z-wP36z`8qJvk$gQE>q-777><&B0~i}fz6lIXB;O3iW|D6MLmSDr
zgR!0DJHgON@?Bu;BKaOL^pJcn7<)<H1qK(%_k*#Y<OjhpNb*Bq93uH)FbtFYbueDX
zye;>?SFg3sI_)PFwpPvJ!Vb}fu(*_US{;jvH&PqP;&P4E#<I9T%~~^yOEN*5z~UlI
z(WbDt?9#RAEH1Q6Z6=FLD_5J#;$pIE?JO>zl1U{q@wQh6>M{sLN1Fvt`A;3(A;^3s
zI|{Xuu%if35_S|iC1FPqsU++uVwHp)g;`11Q6wk{JBkz~VMmd!B<v_Mm4qEdu9C2$
zuq&}02lzT50M&bL+ph65OBP;{Hc*CR^w_{eZlWF=2<O7}*Z|A1dThYR8THse92cj@
z2Ig{e_1M4yZh;;fc$#}!j}7E-IeKg$pUc-{1I1ji9vRqjpk!n6N3HGBaiRj43YdhP
zxI39~lDHC>N)lHAQ$^w$U}{KQ3rsDE>wu{vaXm2gBt8nvQ4%)*(?H@TV46tW3`{eL
z+kj~!aXT>WB<=*Jlf+%Xbdk6Rm>v@M0@F)k7cedo_XE>U;z3{rNjwD15Q&F@87A>{
zV6J0~5+Z$*xsUz$o9qgxuHbJn)G1y*50C!nz)rIcnBss4ls{Ao(s=GG03Po+sE*_D
zP)$lQ_eB894uk42mZ4d|=fQOV$<Sod$DX?Y(w!J6ZylKK*diTs0^|Y2v1Sj*_uyj*
z%^m{zArgNE^3O>84amPC@i~y6Bk?tmU&9#no8-oQ=RnJ(*=kT#V>hVTB$@l>fo0!;
z>N_mE6fL_USB~B8-i_314H}$gL)0M}6gZmIsdXCM%#rFy4H}<jW7V-5+{|XRS%aH7
zL7kw%&77i6(corISEp-G25B}^ovFdioU6{&;AXa~?HW`&?iyemFd3x<{g33+eSyM0
zJ3!TeeWFE3GWRtK%XWimH<sOPc;|~K)7PH>?YR;@+{Evx79BbF?3cetJ4k|%Afmx&
zHc?0<P>PTuVt=VZDuL34bP>DF5HbjqDP)Ss2hC;)Sp>=zazz}8JRy%jcEK*<xD*Hl
z1gZc_1$p3ig6PBtewwWTOAX1_f>=xPN5OKG<QqV2Ao(`1w2^!}i0vfb1C}0=?**}!
z<OjhrNb*A<4q+ZSmeS?E)Z%pYfvOLut8i6*mODj7g)58ZyC2vy`G-Hh8PxTWHy05V
cIU|~ia)0W1MWP3aiolOE@MAQ7<nZJF0b;wZ_5c6?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool2d_fwd_common.hpp.3724B9CBDE79A373.idx b/.cache/clangd/index/pool2d_fwd_common.hpp.3724B9CBDE79A373.idx
new file mode 100755
index 0000000000000000000000000000000000000000..338bc7330ac800976e94dfc043b82bf1af9b8300
GIT binary patch
literal 3740
zcmbW4dsq`!7QpAi3@?WSm;!5r2qZ*a5dta*lxGN<$Q$G(*aUfLB9Qn1?JA;BtSRUs
zCL)SLDO+0v?P`Su0wQz;#FbUVYJF7JRsn(96}HrEhixY_`_KB}{k}}j@0>g5+<WKT
zb7!L03Is=K5EQ&=ZN4mhn;3>5$Q1u6w#l>WbRY;`1VQ=L?J+44_uSsO9&xRkvoAIy
zx`TDKdqj5)whl9!+SD>AOS`@Cf^|^w?4;t!rh}qq-GU>vS7)ksPoB>yt9a?^n)t$k
z!8Sdd+Q0Pb5uc?UWkas6PfoQoZs1%0c&py5Rc><ggYHv^Z|LaF!#}UJI4d{aV|C5A
zw4!?O<<Fk)m>2sTUVl9KWXaU}jh119*?v(UHfWSXn*>((P3}I*1T)63IValNQpab5
zC-vJRXARL4?n0}WGwgmjC#8?ANGNoUi~L*fnX5dIPZQ|SX9K@@_tN-BDZ8-FZi6=T
zQ55W_H|3?XR43nTU%bod2Gd;+Ycoxg=8Dgr-(F%~*gwUZ@%;Fft$ng8+U>zW{g}rC
z^{pKre%Hl*>b5<hqIOI^yEW+E<E)ik&54UIdhPMA+w)D>;GKkBN6#r<LMgl-3J>~P
zJ?^T{8>8LJ`a0^T-X`vf=I1?w-W6p-=DRNy>!#^VYmO<q`@P;*c7Nd2CdqqWW7aex
z!xEe9cOQ~{5`UJJ+#b>1$VkZ6ji2Q7ALp#7VaN2^h8k1~At$b5X?nz(&lo7@PgBL|
z7Ew9O^lJxpYM?F;(<<FQTJx*MUgXp;@?15dGO*WFCjIgmTyZoWvGJa{p}$i8QT+L#
z_o}+;j-QJJ5lS7&e*z|U<#Js5t5!8dj1(n4INfWq?^Quj*7Z^8iEldt3%7k^d1oM~
z=7+$(%Czd1o=2rN&f+hJ|Cn?wL(Dt#7qd!D)|9`bIt=9<I+LZpSX~63$zEo(-75dG
zL3J)^*Jjb&_ay<M(+ad#n%L0ptaHHq5O+&uU+a#xnuY_JHr*@lU-)Eifi?4?UcI%u
zeN|NJcWrqtm)oUEWi%~r_{WKkGR>TRLD|0>m)O{4E1!0@j?!7zF)pMikV$OuoIQ}<
zc_lis=q3tb5QH!hChfl$$WRABx-dWhBu~#!yfDJMZqx)0dN|T`P~<YOy9}JQw;~$A
z)PQ9KgUNN_0bTrOx-cSjmAVmX7?JU0?j+^-<oFV#%Ukqm&45ISporLA>`ACmM6z1q
zN0QV}8cfnw_pOwkxk0(1giWt`CSfmq%jXLif!?5IY}~2Y2%tz7H%lASBwS-F8{f~u
z)tTH(Qr*Ioc~q{sp^K{*XSxt^@p~6%Q7D^{%_VFmb4^RieVyuXPcg!na6)Z)rSRAn
z57e(=7!liv*@PM=FR?dC5|+fCLjICKf*Akk$%CHzroXj=C95WPp0F=J{>JCJl(uj6
zL4$$E_zX$`Rtjv0A=ns|=NA9L?3fOWM8-zO6KWWV6UK!RWZNpw8ZY!M0Wf|ytrVCN
z2KH#nS4JyRW8WCxHKOjhMV|BLZ|vyJobKDAC!W^BSCJ+*Ap*gOgeIYo3eZ&Gf0&~J
z4rFf?@KJ$xNmc`zn(QBx^Jgy8zhxmgCTJ^uAsx@C(MJ)HiO7QRjFXLsLm@ZODhm0F
zLP#nDhGl?GWven^NAh7{c$l;cga36|eY?5`e_}Pjp$7O69b?1ImVJN37s2;iFg@6u
zP@@Q+&ZkfaJ%mC6x`0BV^iVCu-cJ<{SFQ73z<$e?D|ST8HU!{1ND51h$q~q48HIA-
z94$$s>%X>{hPxIJB`@FahE?CtYXnAuS;4D`C@{ij@p&YLutGdY60ii6G8M`S)!Ond
z-VJ-j+<$TbBPeQICyP6I^DBIwJ6(5b3kZtj^YR0Tm?%=<Qy?UW_y<P6m14Z~&bLEC
z-2)Tlzj#@qh%IKzAbO#Q9cD)%7RI6w8)H+*9<!$q2jfs^8Mcf<4wwUl95F`<Iblu|
za>kq~<bt_Sh>LM4v>aPbAwQa*Hh)mWpXN`Y09pWr0%?I13T6b`l5^)X_!J6ZgiuJp
z5Kt(T5vrxinfFqP=vKG!eW3*2N+2K{o(g*C%Zw5kzjZM^BNf5IsFH8-DyIfkYQQ8s
z*G`%{jy`Aogy(_WLCz)AC~^sSUIN;K5JlR6s0~CC_K!;wj{kFehZR23?B&`1gu2P`
zX<YJjlnb7SD!@<yE5b0(JQdJp>^uuqU>#v^w%vV!HEXdS_s8S#wEoU-7;&#v(#`PE
zi4Dct9fzO4h7t;?fT0R#xrO%ZMoI2w=#BNbm1gkw-h|Pxw_a6^@Sxqq_1r6Xr0~@6
zG-5LNNg5Tio}jMbdJkx@BYOdxvtRyc)Z7!9;;jcZV8XvMM<4foYX^g@-K9l-G4%O`
z4-oT-x!?;QDiM*S25dEXG~HZN-hb#q|B?mFbeQ>9<gT;Lz#E^{<J4`Byn>D_T7d7d
z&FA7)V?jN*??}r?%4=bqWt^7AD$cz+>3pLGKXDS_L?d!)NpKQ{lHp_urNAi^N`+Ik
zG%>efHta-+=X^$3hihM(P$b$e+L3Vnsz@cA^=f#ECwq8&cru}W-8smq^3AXQzvGps
zOU3h5S0PVRG|USo^OxTm{&`2V`Om)Jn;s!R5K4;1q^M7d_{SsB=HIM;Zw7<_L58GA
gCq*+-MC5y`F>~R?F=I15qeTY#hQ>6yi75*G3KY?-ZvX%Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool2d_fwd_fp16.cpp.09424A28FF97D19C.idx b/.cache/clangd/index/pool2d_fwd_fp16.cpp.09424A28FF97D19C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d0aff7c6a125cf0cf48dc38b3dff85d669c3f0a
GIT binary patch
literal 2890
zcmZvc3p7+~8^^bCX=gTLWXQO+G21Od8JBSxoSAVQ46&yXxl=@gPEq-CN*B_~Ii1sm
z&N)#XCHX?SNOY2-h+2_QCzY>GLP~did*)l)yl1U%)|zL%zy1He@80k8KJO#)=ktB2
z5agTazab+jCr$-|AUg5S&XJ~_RE8ku`4F_Bym7fWbCj)h|JN6H<8Qh6-yNoGyz#8I
z*N}gRXT06PwG0lg>in}~_LF9ktFU=v+C^2BG|ldgOiht0SLyM|PM3&(#jwY68)v24
zRHI|RibsZGqen)5`sK%|2V4J|HrVXQ!R-%L7@92X7SfPR^!WRb2U!K1vp<g0bX&gW
zhIbpAV9(mq^E{J%#ak-6J0Fa<-+lVmRBiM5gs%M3Gf`2Jj)-NotX`Ffg>OYUYk4i-
z6(=vq(duigpt?9m*xisg#vf@2^O2;L9i?3>aPRxi1;LdEX71%WlO@+D4@Wg6ngu5A
zuWfx?9(3d6`|h!cvM(5W@V;0q=((L=Tb}7KFYT0>iX)@zep|UyhS{XsS#QoIr>>A+
zTQ(I1)zIAPxB~Tzd(te|E=h~}`kJ!By1p>>Z+~1(R9ncau-LV4cz2ce-hHE2ALr~U
zhAO^4UA!c#A-OcH=^7pVWL`NPH}>R}%i&0qms;_h^n)p`v)r@i=HGF<l(tg6u6;Pa
z<eZ9a-s_H$4;9Tnu=4f<1kMCR_V!yfB`mu4QtSL=XquzC^t2Q6_SSh@gbo+azeuf<
zbe9B{&2O`*|68ih?Z3ayyXl<f{=qJkyM6E#a)k1cWtP@nwe0RBM-Vh3J)*v1EiJsu
z$Y^iV$Q_rSuRH3)cWS-~8$aDVLCeiQQ+?B}Mk2<4Z*sk4J6^G570*6oInP+sdT4#A
z&Vf|pl!n!ksg+KTYIfdRUu2QICYo-yN`E@-CXe~1(<#CB&ZAv5PwCQ&0)5fr-``!n
zV$tLC_TtQdlfJ1~c>LSz&&TqH7j$A1*R($BcxLJ|T7QVWecbs=cQA8pU}?_@<oWfR
zs+ZO`$CXOkvTrt&tj%@zvTkMEX)l`6FUfyimU5<VHoEHZ^pqEevB@pn_WFs+N_y<^
zp$pSP|9+&7>zQ1-nY!oJ&TAj4UX?HHI9U_<GB)9IvFS{_ICr#u->Zv0lZNm_BK1nC
zkS~a@J)N3kk+d#%2lde8K#@`OM!2<o11;z2ZkyvjnXXK@c4BCzdNXU}$@OD53Xi@k
z{)o2szo{!?U4Id!yL7^>=7RWrYOz4%-{<SRoY!txp!eiNzh%hf`pTkHZwoS}`if|4
z?Z4!2i=Ld0J<vkw{I+l`u)lX^L@)Q3ZAXv&+%SEx#nmhAicVwiI<rl$Q%CB|4-ZbN
zT{h|C-xIwV9t-*n9!)To=zXsLP&m^y#quiLJUBH`SZBN8K(TJ4?W|Ir(C1e6#*BEj
z5(L@QRcHM;(6NO|5ke5Ip{Bu*2jWi4f9&_|iL6ZHHB?GCA#1B?)8v7a{iFdN+ON|2
z8c-x5Gjth?<bjmMhFsVf=dRX(mJ_mpu7RmMkn+wRt42S|{cQ*oL&!!dMzB1P^3mO2
zJtdw-c~pvskP$V6E)S&K_V-ZH9#wXQ1{6WaOkE}_52W17J*kF0wD^Nc2_a;v8dXak
zNO}4bJ2UBFKnj%-O2~9Ix~4pk@`sn<+tF(K2`VLwkZEeP1@b`3(xl|<b*d1g%w2s<
zyKSmK3Bei83^y_z%F=v}>z~iPh{rTtni08}*5@+@d)=)oDF{w?q_fFSKyYnGZ8n){
z@<1r+|2PN)U&UNyMP`gN6DHZ2ke`L%3CsjD(&VOfwW9Ih_p>Sp9tlTkl8+HQ3XTE@
zhjD<!uy~Hh?GSH~9j)RmiOfWEayjyh36t!A6{aL+5=NS0tE<nQ61k@6AUMZ_GxvoE
z?r!1^kcWeZ6FCLJy-mCU&Bw$CAYTVxz{x{-EX6!N$_Ge*3IOs${Q&Yu{Q(L<0{{v{
z0|5#`g8&Lfg8>qvLV!Zh5P(9_P=LbFFn~m;XpXiewO$p47WNVq6QdQQPi|(%B+t;b
zr-VsV4AqcoO4c-jBZi16c|S<AlKMZcPjP_qw(-5+Tr?5f95$zsE(B-6EPyZ=1IPll
z0LT)y1jq`u0>~P+2FM1s0cbJ27$95N79cy=4xlCQ5`gSsdw?8ZhdCnmN>rJw5h$uG
zmKi&j#Pj_98)~^@13D@=$DCtH?pOtPH+KifgYDs_(7esP0nNwU2OwXzFW}^1JZr@~
zKE?+~fC&Kd!~6j9$NT{bzybgY!~y{d!h!$_#)1J7VnTpIun>Sku~2}*urPo`m}rj3
zmPuq$)-M%VOlBs}Wsxj0!X)dO3R4U-X71KM-2SY;WmNSnQLhqMf{<x>vF6^^hyyY1
zvVLg6a3UKB+3AQSWt~>BWIQw8l$=X;Nx~%Sw+d4nGtPoELDHo3IGL0!O~@9=22rx`
zzn_F0K+p=sP&_wO!n3Fb<L`dnKcE^dULlEsqUR(iAtNz8g}{KMAY~O*wR!M-Dx#r9
z)AXZ%=Y!GJ{j~KK=q%KCapkysczSW|7&dNZY)g~>_f*%w&}b2AYR<B-vR=Hz-eIYe
Yv+**ft)nkPm+!xl!VC5mC<Q|Q1(`D*+yDRo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool2d_fwd_fp32.cpp.1AF35A327BA2145F.idx b/.cache/clangd/index/pool2d_fwd_fp32.cpp.1AF35A327BA2145F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21940e2cbbaaf71a787e1aff3fe35357d81b7fc4
GIT binary patch
literal 2852
zcmZvc2~-o;8pnr_;7AgZAOVR24ulDdtV$##%GMAdtN~$SkbN-<Vgp!(inUd%f_S9$
zs4da9pjC=sl}$kf%OX%g(Bm$ms3;W$@hNWjCV6Mb?KyAG$(Q@P|Np&r?sw;YvY>!~
z`BVt<PYv3Tm6n@Ch9HQA|8jCw%1V6*S~44gHk8#ZmuC<2nb!|LyPEvBXVBGqq>Yyz
zSNG2k_(^QJ%}r1WM^$wH)j9p%C`}O7yis|ULRK2}c4iyNC_Z{amEE4v-^BAr^6I8l
zTPbk~pX7sg6XFI3f877W#H}q4Cg*K(7h<jl)$^_9_J%QN*~rni;kVXp-jwrhjA45I
zQ(jcBr4{<PJu`n<y1%?o-P?U@to`bv2NTuJjVV0^B{i|Jiq7c8)wccQ=((?DxogGe
zzbHzdlgk{aQ&T;cL_1$nxF;W~jr3C}OMhduZ59nYJtMhr%SKdYGG5#<{&Q?Ys*N;t
zUv<k+S;(czx4k1TOFyFg+t=lCN#Esy>auLNS;{|b$nNZ(>#b!TSvKQdr+kG?9zEfQ
z&wsr=q>ABn(nn&LbxpNS(4#nSxW20N+mi#4{9le;Of{G*R`Yi5y|=5vchBD8i$l3P
ziy-w^CyHFw)uxw3Hng#j_uTTyq>+cuJ%5g|8f7L6GY@76rhQJ;&%WZ-q+Druvi)8`
zaXr~7|3&BEl)Cvl+x*?Z(l@~|{Wt6zQWjhrWj2mSDBTTJCp<Wpx6CRGb35DkY}H9c
zZ?Uv=cB|v*e^ll^H?Du_+fZ+`?{*L3v+eeC+9A?ATN`D2#p0{uLP^M=>X6}zwT!49
z3yVDsgI7HJK7D^WYKPIQ$gva6FBy3SHOIT0s}yqV*9Jk8)0o;NQS2JNTx=<8`DuNL
z$$?dt8MUhw6DvLLSM9jAemgIHO&rTP(R@<bCFZ>9_DFHMa(`#lBbMr{#9TJ?>wnH&
z;Pv^vKKtgThq<*p?C9qgACBbTo70WHY-7GNS(a_iZuu_Z^3f$9dqX*Er6qm8)1Gv6
zQJU5_CzYsLbGmAa*XD`59a`8|+P6=b7Z*G!&8Qiej;k1&obVR1zxK*>>iE69oRx6&
z?wQHEf8IC5%&eNaR_*?KN841z^Rh*ql~pmL2`T4_tluQd^LU*$xAL|P37CUjriB*j
z;*z#FMgP;V?uVmBW{V&ENf$Q0Vt!LJVfpkDJ2u2#e70jp_?fY&55djtyQO<6;(yF0
z{dd(oPNGYTuhuL~AI!sYPQU0k-BVFo@%`G$UAqL+&TZi}k%#g#B=p#4pKm`rLbtk`
zIQ14jGc<C+bra>s&u+c@tINmExr{o`7E7I!Ba1fW6b=;HuPz@Bn7Z5FoD{^bdqB_k
zuGoF-$fKsdgFdW|PEoz{n~Vv{*#2Q@A^Ah?@O?^>V%*itX|>{`zMARRma{P{nXd;y
zdN(}Po7&S$=uiv<VQf?O0&T$WwDxg4d3@auH#-Zdq%d5j8_?O>K*;>J?WwlemTRb_
zC|ou+FlJ~2As6Jn7ECQi?C8*PT%Kn-&srM@c}JgpU7+2*RvHwK%NAq{SQ`lWw_Tr>
zDVAB}Q%N#hrWw#!+Ca!1+zjR611pl~P&6)cOgV@)5OTjyr2%?}cZ^C3$7QMkm8lJc
zJo%oVopvWUgG!3PWtIWUNE--wYBXvaa@_SLl@y7~3<Jg-Z6IV-T6#_hMVCF2;lHv*
z!%{h^Tw;6rKC6!yw@z%<qhahNY%e0?`l<q<;IAj%csIk8VL`0O{7^G*kI12%M8jC_
zEI#p`G|brDm`~)JI2L?e&6yCdMZ*#~iT1?0gc&z!PDW=+;iT9QCNJhib=|?QrpYuc
z297Zzx@lM}919QzV*trv`3xl-KVJWbOpswh!-Q7CnR#iL$VvpzQn#fZy2*X5d;!hR
z$`2raH-Eq>M#Q$daREpGKoUd(P#_WrP!JLXP%sh<kQ9*u6oP~R6pDlb6o!NW6pn-g
z6oEtl6p2IvBtv8~L|g*=7&JFPcMNHqG?dud*0h$3vWRc{@sA;%8E;N>6aRXK|GF7P
z!TMXr_I&m<qG4Q^%OG4d%oer<2!&CAcrXtjJJ=2&d)OWz2iO51N7xadh44avoM0z_
zoMC5xTwoV~Twzy$++epEBCa`}AI%-t<tKrYFq6c_fPEXPeMWAYkTD@wXh-arjET4+
zfR^%?dg(M@t}meZas2@D=lcUrF)DV@jSD~n0Fs~*fCAA#fP&B<fP&FrfTXAtpb#_!
zpindvpfEHHpl~!Cpa?Vqphz?lAQ>u~A)*@MS=5w5T^7?h=`&d*Y9emZ6hxgVo)bS)
zuI_AmeB=Bu<rMxYDqsbTNXxS|*S16-h!<(97?15iR57AH;ZtggsBX$+PO>#ImZ)vG
zNmEL7rX)@hk1(myGLtk)jw&TbqG>2J;eXExq5(njnWi4r@=4aQ+M0Wa5+_%zh=t;2
zBq&zbOvy^k%)l`?EkvJ8F_;C<rqbw4hEX8v3qO=)7-(!Z$7HU#r$8uLy3E_hneFIh
w!?&~gf4!#j=36X4thu&4dxwQCu5ODwmRK(4IJx_?O#^~flEk6D5<MyOUt*vV>Hq)$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool3d_fwd.hpp.32DAF50D80F54A30.idx b/.cache/clangd/index/pool3d_fwd.hpp.32DAF50D80F54A30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e753637edaf38734d1e1a08b4e81a09b81fe1ed2
GIT binary patch
literal 3978
zcmY*b2~ZPR8tz6y4jZG84p&POAaXPzAuln|gedfYnBiCykzrH>1cqy=;!tsQ9S%=Q
zXLJRf5*Y=r8O5<g-NoG*7deC-cMBa?c2vAjaCIHj8L=G2*-osN*QP43`m6r?z3>0N
z|NZYRm>(ZsE5$H&#r)N!d6n7D7>3Q@{uPzwg{zz}tZy2It#0mZOf9>vn|^UJ_4#1Z
zG3Sb!cBwq}_~ZZhl^*&!<VAS-leP@!8yz!)zMdE!i*P+HQ?>3$lx(l~?&kQLmk;JD
zl3SP_Cb{2f&%(os(Jy4@LBeS;`C*30WypG{O+9|Gyt&gSq@ev_a>9wG-Zwc{BefU*
zGE^G3c=w#2zB^HtC70$7Z=Gd0qw^$MU#?G1^}mz#%d_*f`v$gOa{c>g)vbp1ln3ld
z<0s7>7Y2{_y*_CyT8m9I_HA3$;IgOee(CnkYZIRRMNbZ_%+Z`mFOtl|(;o@XxIdn8
z#_ACfw{Et7PtECn1X$}1BnQnGNrPgi#r8el$*lS7mBFol>mF4#QNCH0w+HFQ&O3{e
z%RVh2`k2wK<G|yVH}SiO>ATifWpBQ|1K;iCwdE#GepUTh`Sa$rq3vxuyEc#dofv$0
zUD&_r3D)*_nTfbD^>$ZdOlan~m*Cl8{>bL?7Fn$NI(>CyV8wv@z>yD|Zl*9}TQcVT
zIF{MfF%&%rN)B4xdZ#u{W~qDn2c8!7=XaNN-%BfFmsVf)+%j2KGvV>E*Ot%gX1@G#
z@|wrbo#czP&t7Mi6c;_Ds67)Oy)N0*-qC$+IQs@U{zFUazy5r9?as?*=XM1hWZic-
zb$O=zn|gY6L2`S>^7f|XZXt!|PF=`3Jev8@nlEGZQ~XuEv8N0ppZ<6$OK~Jqp&FdP
z?+EwZy+=H4IM;uy6~Fg$ruVP$`^*X1)2=kWToD{vvgGZC=B<BV|M@7-<xpYNpI#I!
z3YlD~{zA6s*6R4qguNS|O1?SWoxfG#SsXQ<-E}5vY00kgh>rZa>Y;C+dQ22w`tAGb
zfXtw-%UQQ>Igeu%)ulNSuKv6lmx$7?-XsakFN7;g?gzXD-eH({n!w8#i~}6pKYv^J
zLX=Ek{xHBSFmnOZg_u3SSRXp^>Q!7lfvI4CMxfDRVhR2(AZpE<mcG;Jmv}Qo2qL%{
zW{7<3GsNvXupW=d9(Hl^3~>jE!bG)wz(nr|)n{DBedVeLJc_;oU+#NX=XVR#7d?M)
zb|5#@6;p&@KrWa2*au9^eIFXg<xt-;bR@*BlX7#KNn>F7^yQR&fD3HwnYVxcCvy^c
z83YT0xeTT{k#>Tc1BWT<tJNb@ojkbj(<SL{_5pr>=E~W~xxX*Xbj5tR_1)#}3j1JR
zUlYWw|3hl7BK%2D1A)zg62%B&44BwWOxYK>*F5Cf^JJ`!zye@^UZ6K(Vq%Ki9x%50
z=*DkdmBFr<l3PM1mnrQ7TH?|DWvcQ^Yt9jv8qVM^@aLA0x_H}Xc$*oU{>S}wm9ChK
zo54fw;bR|=NacC?6_pYUn=-fPkXzTY^*90KdbtsbJL8)z-c=_ab&Ci<X(*keZUCjH
zj0h<;N}a<npwugk2&pw{or5P(>(xet^0oPTXcGbQlk-y{f^$2G?UlLh!t4FV??x#J
zP$nyjhC#CgXLnJ>Gt-<1Fk3jAd{;OaMt~YzSc5xo@X@6$6HO=asaz5wiiA@WKq*p^
zJfb2siKx*cnsk&1(2BIAqvU}>WFQ^p0Y=1-j%o)Mk%e^Z9>9uN(oqK>@q@%vL^7RB
z_k%SCzu?lf?K6~a11@X89SK7UT$pXQp|Ewbbrz_~?A!fG{jVmwWL&ytDh8!Xu~Y&=
zO0zNsC9h<a2?(jp>KG)2npGztL|bSU2}8%xR)n+`EekmkKpUsEB4n@_SR|Su&R|7|
zF*7kpLWX4$5VDvpF;E@?SXfH}LadpMLGrRJn}ASmaBdisgaEk-xk(Vgxj7YWEeWak
zzSyuq;Szgef?;XNZh;SXm<197_=tU^NKl1XAw?1fhy$cZv=DKK6v-4O4wE8bfEY-T
z6p`XcDI`dMXmPX@r5h`bm7)Z5TyxxEN_ES<<Bx@dmE1n1HFP*yiUza^A+1ImjuzGe
ztqCE6#t@EZ7=XcqP^F<V0xm^>%CyQ1h+x8;mbUf`MS6dq!fqmw*jqFTqO4R35>Rbb
zry$AbL^>5QHzXQT5g*3Nq#%)4E1QC3DP#)gKz0OJ#jN6?4LGv_cVHOB8EiWppx^av
z>(?Y5?Y+pgi(E3-7e;Xi<eCx6_saL<AyfWbgbKV0c<(BGD+AHH%IL~?gsOb2c<*X)
z-&*`VY_7%QP;?#cTZi65bvW&9Cl`{H-wj}gagx|iFPJ0&asqNf;mmu-X3tE<ej4B!
zX`nFB(KrcEA*^tqw*Otb+xi0E%pDOPdJiK!THrZINbKhyVnN9%1u|DDS1Mq!5I`+g
zD|lARX$4G%Ya4Q{f>)MugMw$goKYa-Epm$j?G9GXDo`0DE=c^C*K7ULL!lC#2ac{|
znr;UYTz@oC43DBo$?zze)C`YeByHwVjMSQW6eA609z_$w@F<!r43DCTWe~*zeSr}=
zK!7S!6@yS6ZmPo__`=Ax-Vyq<dd?qDp{K(U0^|wv90*R#m4oe`fsELM%QoRGoEtLb
z!fdw<6jm*({s0C&&<TDKN<A8EO^)>8*{C2DJR1YZ0G^E@WC+j3Ffxp1BOn3K#z-=f
zXJa%O&9gC<jOE#w<CWtNNx|z7S82Ap5MoZLD4J)ZN=fr<RH<p6jVhYv*{IUeJR4O8
z8j3>zM#a!P8&wvXXQPUxq1psUT9C8^RoY_z#Tr;?|1G{+leTA>nX^sqC)dF?$n76L
z4;|b;kR0cS?e2!MDwY*HO#Aih?yj(??>o4?XS*rG<tC@6pBgUf8RgnOqhWN2FQaFS
z2w5~19b#|MTZ{;?8disLVfCyLp<3Kfi<_YQ1gOOq*5c_9!vvfq_B$mauCyvskOA66
zZ7LGkVzs2e#mn<bvTd1El;>6~v_0r-!@HjrK44fFdJIR8dh`hHw(kAyRiRMoVhGb|
K!}XW}z574+OMf%~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool3d_fwd_common.hpp.EE0EEEB8E82E75A0.idx b/.cache/clangd/index/pool3d_fwd_common.hpp.EE0EEEB8E82E75A0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..634cf91401295cfe6ea5e77768d7829582df1824
GIT binary patch
literal 4464
zcmbW5eLPgz8^G_)rI~TvJK;8rX~M{uByU4Vq%C=ux2VKYqd_CDYknd^VN-s+uEo@@
zYGt=nHkHbj3ft0@UFt`cq-}d4y{fI>zFhUY_n!ay>3)3f=lP!V+;g7iInO!I8NuG3
zp7JRW^vP;(r9ze-gFp~OB9xY%8eggbL46nmDXW^oV-jz<=)VlKJzStu-{{?2=Wwv_
z<`a3$fg#0pMReiJfy>n=SDQ6PIldXF@%n6krTK<qRvJ5B?@imEmXTBDR~<gteY50j
zV90E{*Pec^qOyY5x7%17zYq#U$q$c-3RgN08^oN#=6bttydK<C<GZZ#baHgrQBioF
z<Fbl`zouQBwBSi?9(S0-I$oE@eY<bclgE>7`<=GcIGwvVIAYvi9W&#}Us|^P{P!)h
zOP+@rU9F$w;agnu<bq-Lzw2M0x>C2vG1NwW>&>oot8col2Rg1@|MqfUmsZf?#D~Fq
z!W2+v+hD~>y<%6kcl3!%1161mP8pY;PhahzalX59=GdTrhN`jZ`xolRR)2l9dJ(Vd
zLMf*R$LGa_p9nh>E3{s!=lDWtenrE=?e|jt>{|n+u4|O+yTV<07l)L+YlN4FHazff
z9gUu+%)ekQ%r3fcP3YQk@{VxcrPgQn?{imtx9IpLqxB-A>aBMSxdX>L4~b{Ik__$L
z)mkG^_+C*C_V##Bdu84bD{k^WB1q!#4&*hO5LaDle^IfIbo8hZYg@NIza`n*i}*u>
zk5Owo;fPv3%ueecZTC3WvI90aZ99KXY4KFHYjN_Z-MRFlmla7|_vb6LH3QBn8?LTW
zMBSd99jA80A_W^-26=)F{Z3VXv_3dDWgzLYk=p~&bGP~)pSqQuf$O5PMi4&{KK#wI
zU%Q@b7L<$*?zJ}b`{U7lpM)Xt8cu~I%CNS&dw*F(md%&`!@g^ur96W=?(nSZq7U0H
z*Z*1ldBMtOZC7rzL}c-<dM8O@1&iw)oJVc`_LEMT)m|xAYkTc9yCP$I*Ts7}<a*8Q
z$ukYT9$VWOP~Nbk>qXstbLqx)Ck}JGwgq_0PT8n`Ja=WsORMAVl7q_1@b05N$<3d;
z6cqk?>k)3%b#ngFulk;tZTU99F7uvQU(%r_d~9Lii0<=!xnHvyT!Q){*7x4qed62G
z5yK$@j)h10<4S4dC70iD$<58YzF7rU`G<U)PIdp@kUXbiN$bP<lWm&+-4@Y2svp+A
z<V1mfhsVa8knL+mWG9;vvmC;|tL+V$wo)$w8{V52y#7Y~ufYo>-Kp1>Hy<B+wd==8
zg{KDhI*b%`OO6?>mg(*mYq$0r%yr^-R1R7M1y>xnk<irF-swK(GpkMUfFn0-m|PAk
zOka)ucs^WR!0TG{La%dI{8wrJQpBzz_vVrK{TgG9ULq6{KoGWAb20bb3u0?&LM&|<
zgOB1FIlkp~U!#x?sX!PK*6&^rTLXrsu}Os!CW!X(O1Cdi2qhK7gxKoc3t}UapvfFk
z0ZNs{rIFnb6n)-9Z7;zmp;!}a+5}tDlVg`)^GmQb#qg+K!2ITg4>S-g+dA8ZYDTae
z+Z;QFa_8o{Qlw$4UAwO%wp0_vGMzH#QOzi(bXP89DBC03lc5a`8(bL5gBRt&J``&m
zdliw3Cmx=F4ctbd!hW~cH5kRR^s~g&NW6=Ad%o{4{}Vy5%o&+<Gf8fl^C=|+Q<^I+
z09vG6OcDQ=mK5lrtL&o(FmfEttejEy<p%>l9oy(!!s-~*KrvI?R6wsi=8QWF=!M40
zVZCzLm@?0pQ<l8RKz9#m-gL+5&eTT;W@2Ju#*kf<T?{n@g3WQ9<4lpsnN<03p6g~Z
z)TSI$9p)2fj<XJ>gpjk6e^aPg&F(g~aime9AN0vOaR0-pDVB#nsOTbm`ctA-%u~{g
zOe~X2O+>B?rUXa@V=9>D1?vu33Edkfk|v(#Ha%WD9Jh~5Rk6L8jsl8V*joT3v6lce
z(|#sEmiCr3`F!&7ba{7G3z>BXV2cCrT*~m!0A2C@?XH?yC`M=!+EiFkOoqz{N(sf{
z@Hl|vxE!DaJOQ9YJdvi*z9^p?6@Km$*dFnzX`cnYu9;6JS2mKZ!&s1m<N%b5<kGY@
z_+YQ`Q>1wUv!p29t_gnn`A-&^jrepf<u#d&xCuZaZ4t`!$wV?|${qxZ6UEswBp1o;
z8A=c(04GqQD3R_<Y3T7E6Bei@U~B#nKSh|TvyF^M5o}!qyHWF?STSr{4AZ9rij}}~
zOPIrHi;qU(NZH>`Gsvao5<JQglBC33hNeoV0>qQ@05Xso0AwgN1jw847BHsx5Iz9;
z621WW5q<#q6aD}N5CH%M5`h3MA(j9XL<9j8OaucILWBSmN`wLwMuY*hlvoN-I1vs|
z1Q7wyGGZA(%ZcRxtsqta6iGw^6op068vw<kv1ouk#Xbcn28#hG7K;UF6}Ac>8DD0?
z#5#^22audE2PlD`08k=7k*3|RK3$oQ3%bc8pbB=Wg6R^scfo+GFvwTuqbtYPYkd))
zpv`w+ggo@jRWMzlsz(XQ-G`z#Wa{R_miaJUq*3fF>~a=%XMAdg1DfFwiVqY=eED0(
z4}$R`X^}<mI;tDyyuW{tw}tVC%4LM}{?_rn6+*GCFl#F;q`pG2Z7^#a$Yn6AjNzk_
zdtob{2F-fs244l!YkEBLKX%onxGw2NCOeZpcgSl4TM3W~W~*SDsgk6{d4JQPjZ%me
zfZyv_>*yMZVkxW?fYMlL0A;W;XlleBT-%(S0DZ7roxxbwTO;oJfArd(F;m~Xea6|C
zoJklc45dPiVm5&`p_Gs3eM(G6t|nhZNTqOLB%@>;X&l3lO}I@Y^$CheBc*hyZ@zKR
z9#a1hKY?ShJ+(II?T{kvqdn{3F1+~`A9nfZ0o0=P#rbX__;^})X3}>;GA;a-K}^!A
z;MppeE(e`nW$jz)+l?mR=?&po%^pkRjt?&OP3;y*NXd@nq@r=C@iHdx(opF#D)56*
z>9?I0Y}cNE0gn>u!kqOMw2+S4;dUZwBouSR9RYIDc9~A~J=$3}rn0;?etl6O3Jw#E
zBqRx-WF#4&R3sIkbR?aok++er{VO-m8BZ`m2wXXh`Nk%UpEiUIZ~>L_r6!a;e_l|V
zSGX#x$N|KVSf6VC{~aQKpRA1;-@|FCtI}%51;g>z`$wM=Ac)V1I3r^En0!wU(b}tH
sgzM=~;p$H185r^jh+tHvG9r%=^%!ydpCtc#xRd}vx{PSRh(?U~4-5=Y*Z=?k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool3d_fwd_fp16.cpp.176B982A3BF061DE.idx b/.cache/clangd/index/pool3d_fwd_fp16.cpp.176B982A3BF061DE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..605fb0dd7fe4b90f6a731d502846ab4a93112788
GIT binary patch
literal 2502
zcmZux2~<<p7JUiRGkF+-BnTOpK?H;#2?>NSs0k#1BS;t(f&m!|f+)XL1oaoCs5n4B
z7QUdM6|_`ltF1^a(8@wl%diSm1S}{}_<SgX|0RE2@?5J~D?4}Jea^clcc1fi`3Qv%
z)<Dn}nNLPS^r8Lg5CoCYZ|b3x*k&~d@<X+Z;@W_|#Md017oD@QtOLd6Zxgfl!z0yQ
zp=X`W@Ci{V=HwgRm-Hj%#$7MI;ZBcSe@k^3F5(Rj6%nuFOQGa|X!=GE`+v<e1s~%*
z!dKr}u}yk!zbPrHy`yb@!gy*S&%KCYTllm87bLvALMy9Y@A+Z{z2V`5#-%we+Q3Sa
zwLG6x@Mk2)%_z0h_MyN0&BWaAv42iw><#b#WY}$R{KX4uVeXH$=3gEsWQXTIt@`Oo
zOjrDgwO&niskqTo?HWfU!k-H`;usshU9(xsMgy_4{bd6a226iKhA?Wr?dy=5^L91j
zz59~=EoHAVM`u1<ShD2wzU+~TI)BNmE>2=4Jqr}UvW<0B*}TkN;b599J3XCtxkT{k
z0wN81_|_QDu)epm1DA3+{~xtI-$)3L#G;<#*9Zj$F?wlVnF~)*e6pEllSH*KU45L(
zWUzk3Sj;bmWr?33&KWB>`}8#9anC25D#5JD<8-cLXUdYNEMZ+~TKe-Qy?q_cqO*1C
zc1Pbo5zJq`pKWn8Pb?8yG>?s6>|?<P)8NpIN+$!2#+Qb9BY9ffiuGarhnybU$$zwH
zJ24pWYg!{x8aLih+81<M{O+AjsaHZeZhO|!g`#_rJHL#cXXq5v7gwCP+}H9fMfBvg
zouT1q<A>yQLYL=*Tc2L+T>f|BB%9rHGR=l)Gs}_JU;3^EAE-XFCt%n_<94i8`S)Hk
z+_Hea+{s%D2ey^hZRWH326K~f9fe7fP0!Bltv1NaTJe`z<ez)))^QF`s+yb*=hM2z
zd!t^-a%H@tm16_;FBZSr!C5fB@l*ZOk*_Nu=F&mVcSG{TXH#{<%L|u(=;t=>i>aWE
zY|!UlvF#WOd@@sp55<MW7nZa&$TuAkUYZ*<)hN01q37mc-~Kq3fWgR{k8Z!}+R0d~
zI-WNDjk7`3dCGb2a%lR~qICg*HvhTP_**nN^V{1^(>>nlVI~=Ca}L*>j1XCsnXz64
zIyDv4G&Hqn_+`Ymd{NCeAB?;&a!F~WNW|x>S-<a2Tx8qTmK)r;db_*()Ia&S*6s(k
zg}ovD4SGF0?tQk#y?nlx`SNb-#BzJ&NZ4t1{S-4{+IyqV(Xku1DuWg#y_RO%-!u<S
zR1WnnM~Ln_jb>{h5#}cPON;j62=W{I{F&?HF^g=Sv@1D(G&i;Q4iuLh3eW8wYaj~_
zI~fIC@Jq;O`7J-u?4b2R-k5JG{5pzxK!1_;*Q2?PX^Ri)QN7cmoV#`zkIztR?N)HH
zTnQUWJ(3U^fP*0VzFjZ<!iFeB{B{VEGBg=9C7@HT{Mg*RnOyMelh25Fe^l1i)TSr_
z$PW75>tDZ5ks?q8D%+^rz)Aoz=fABoizH?o5x)nObv1QKN&xbSNqJfOY}b%B6o|?+
zD$P_0KrTQ|Ih1HOHfcivsBB0zG*JSOujMWaLIjMXMEq`4Ml=z!5`cUgSCAhccePs^
z+Kb8*D#cg{K<@ImrD@q`RY}C}KxLvPQAY_tp81QD6x}D56Y)DynXE}(s{|m=|FQd|
z$xVkzB7PStlQc=|lmKKXCHg?B;ti?haqznC57WnRh?L??aRD`{rDVGDT1MQ_Z4#Bl
z08ay6;0YgjfJlXOAq!kJa72ypkrRm2L}20vo`6Vw=sq?e1^nU&f9no<oWdGVJR<d_
z`&xh;;1L?6Fh~{TPxs#df=*9uqw_k?8sZTtkxn!PcZ*07I${cZIwMxgd6)lztA|Kg
zFpC6o5ve2Wh+&)I%^2ncJ7E|bW@8u!=3tmJ?2KVtn2TXNn1^95unUH{!mb#`hxr)h
z2D@RH02W}FJM4~O9<T?7dBUC;<^_9U*cNyThHZtnt}@W?sI!V@S6#K4z|2SG>~{Kg
zD{y(>ZqXFQ0H{)o1V)$?Z@Twt%4!c1KAHzaYAP`GR-NKU_p=12RK=F5%{6+Pq3tTv
z5z@eRO=jCH@0tY_-Da9H-9RpIAFW9-R;rqq0w(4F5nZ%;z$;>__3qCv1f%^!b|gE4
z;X|alj=Ij^CGs||J8Ms2IA1YKRune)A0QZQP{o+4j+4+O=AcwCVQ5W?VO0eg3ydY8
za6~Gmi>*NlI>Yh%6@jTKQK=%uB2)qY|Ac@Agp(RSJr+F`mPFME3l0elf`V5iC`h#o
zQ!S$sWbty8ilcqi)HO7<;59@<TZgpPo4nP_l1%W{)nBK#-hj(<<+};oJvLEnT+BEd
tjQ(zyN~1H3O-#)#tgJWM+B>i|vz?hvhIWoyC{&?OIDVU-rwAv3{s(C~d_MpH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/pool_fwd_instance_common.hpp.4DC931E3BBF2D1BA.idx b/.cache/clangd/index/pool_fwd_instance_common.hpp.4DC931E3BBF2D1BA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d9800f6acd29a0731f8a751eeda946751dfe576e
GIT binary patch
literal 990
zcmWIYbaT7L%)sEB;#rZKT9U}Zz`!5~#Kk2=nemJa4B<e5iaAGv4(2f_3Oo(|_Mk9$
z(Kfk`j{&~H`lgvaRVp0G(@!u(M!Y@t<9%x2>vz|_1x;4tNqlzj@!pdQ3iI!)@7uH>
zlJ8{9`5Eb_C&*n|vRyuAg6qrZ3mblV9B}8}v%evD^6g8`+;s};6}<SoFPszMo-K2^
zb#8Y@|A|u)q4kpU+`b)V>0RI9GdcU0)YWNQ-1aWk@~sn@q86(w{N=|L`GpdP8E3Uu
z%(UI_D|lbvkutNX7}v9<7h5=Pc=5G(ytTRfl~dOH3frTF3trD%5S3!>cRpMu%A1uX
zdCp9kf@d=l{PH%MT5D)r_6hZ~I@8Sg_=XqbU5AB^i7R(%u_o-}e66?n8=Fh4`;WWR
ze!O5feMjp~fw9HTm1nQqw)e;|HF;+)6j-%#$@&c(#?G%De{rv$w7N+&bn^e0SF`@k
z;yTl^Kz;fBi_t3zS^q!%z3b(4zT?R`Uw<<39@%*@@PxWG%lx$r#g(~9$3S7ezEi4n
z-K)dgOe`D>{6cI(oM3{16DY^Pz{t;NcK?L$iJyNDax<|56^OEla)Swkg60G%x7CKm
zW!y~6Kn22V!dzeip`dlH$A!+^TT{51IDrae*<{th1Oq1%FrGnH>|D5_)@s^)b#5j$
zpaMZQK~6A%FyXI2*rvPrN1kyru>uu{u!(Sj350^8)U;v)plQrS)z+4WpV>3=@C)(@
zioy(rd6JO}?77}UzZa}|XsrPA8c-PIFPLzA#hm<vy?Y)54G|U+mW25aMgvuYJPA_`
z^Bqta<TaRZ`v+GU%kw=_KtqItgd|`tSpB-WGV<Y^KA^BXr#v4_^QP*PUo01?iZS!>
zD{v?%!-RSG6*&}<cuE{fa30KWKzl(U2DA4%_fgH@&Ph=~3q<uqjbUDdg&<HhDEMHi
zi&AqELGfH%lw9l%V(`~6s%by7F<QjR#m&RZ$G`&;ga9rWgMon;!Q@6T`4CJ1eh*d;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_avg_pool3d_bwd.cpp.4403F8C1766CA1E8.idx b/.cache/clangd/index/profile_avg_pool3d_bwd.cpp.4403F8C1766CA1E8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..694bf2785331f601c35b171cbed230816646d2d3
GIT binary patch
literal 4198
zcmZve30PCd7RT>|utY)<0!H?LtPunO*&!fX*h6lR0tukVk^m`tS;U=!U}aT7swjvp
zT11N$trfLaYXw^sDYdw?h-G^UQfjMJ^xdF0cl!G7_hmA_Idf;uIrE>BpuvHG2@C}B
zOAaoQrR1+ABM732ee&{irRS6o<OKymiuN5_Ez7#=%osYLd!o&?Kcc%+qm~}5cl_&S
z|B1LqrmN~KD^~=h^ZM*wKJysP-R~M$VGgLG!kX!k`(9V`RBFNVqumMVwe8JwXX7>+
z^i#R<^PH|3TiXlb9sQG!@^8$JRC+D{GogufmAAh7c7xYdVb#Z9WnIhpYx3{kU$S34
zoj;x-DD6E^wd3Jf=_Q&8_ZRK9=ee`8k~hzy#rL|$8BbqKr7PWgv@^>-YwMsxQ}Kfr
z%JsDo!>d=Aym$UdY3<<vd%s-CT(?W~__v(98KtXU^_w2!1P@FtrW(HJi)lB!GM}SD
zABcNB_LC^T?!Zl{aLaD<;o7jJ{K>-VQ&l^23LN&-#k{HL6Iu&vb1Zy)M6)O7BD+(>
z<c;+B6%S<hrsiMPu#HoCXv1?p6>(=JgGljJ9(U27nQi4KRtTI9eqP^eC|UVISx8{H
z(=zMez-o1+iS^T@>=(ysyeG1+xdq-{$}0=%t!lZ%dN_5adFb(7Mv{lY-Y&n=J;|~S
z99iy6&QO=?<T=q7=jkUv;dUol>+Z`TPxierE^Z2SK=rS5Bni>`mitwPhRQBh`J~)T
z4_c}t)li)t5wHIv=f{y>3&jq9Ei)(A+&dpS4}=~!_CI*;e&pG-8@a=be(P6bYEWLP
zLvw?}J{8%PI}M}mPNzEcKi$*b66P$Cc+Rf5I7#uROEo%sOQ%nk8=+qnI*u=)vFgP~
z-*5PnrQYpp_426iR_=Ftr+wJ2w^oi1Oom@QeZTbgz{@=+Jwx=zNxxCY?C&2MY;Wxv
zbnGF$zJEyQDSFTuTh1(MmhL>>FkTV#iN()Om2HLM>wg!1+L&;+40W0``y}Y6xwQ=f
zl~U8A567A>Gz5KPG4?!v{@|&E?{$X9G!92O8pTBPHxFH&%TeKOvvt0HBe`?$Or=0J
z;W?vp_{ZFm`nfN13b<~%J9+0iV!xbgzcZt2bFxu(js4Q1_U~V`Yqp&(ubqp_e{gr*
z^1cs^j&$|svb-&i$32(jd+LoJ&M98!ez^NHk=r+%3j6s^orys4AN+|xZs>QhEf+rc
za@rs+kTY;mEq{YYy6dEi_h@#&rKkSqu5Ec^a!G%n;i=*KL#K^9HM07CY+w$RxD+4L
zo}pza1vVMXlAoVx`AwEFXvb@Obxqf?B{;I@_`!){lc_&uOygcOM~yYTW;z(%y#7=#
zrm87+CcSH~ONn)(^6hM)U5BmNY`WJY50CzqFC-b{2X#rFzqV}L&{)8+m`yXBt!8Qr
zzt6w)_(b>57ov5lZS#s{Nw2WGj8tUq?iID}rIBJq2ugRRdoMI{i~Lbe(VL#%dxoY$
zibD`oLsdg>p?PaW5Z7(oGmm9Ij8Kt8%<QP_XtdDaX78Yo<a>r!7z|PZwm6HzVlFhe
zIlNLUL33v~mqCih%oY?2&4mUx$M<erp3pefNk!IRiwBSc^cEW2T)po=w8N&dF%3kF
znY#ee1@zy2A@WLvJ&&)r@e0D%Y}O;iyd6ML^+oFP61MQ<rljV5x(MsO%K>H4LhN|V
zRFutN>*6s`QL80Z#<<H=G?|@jrSK3<VW&Wp%1(tSjhzOOge`$c%9bh!-<P!LA33I?
zNoq;@_*Tk!w>FI0elJiVqsdmu9DJCJrdXvwlxmd<QJPg6L=r0rL{clMf{Y$ttW%qQ
z6bq=Rjg5^x9xR)-#jR59`Up1M0kk@RF}|zG*<8uv&7S3ej2UPhKod7uDEI6#$S{l|
zQPEu0+{N$mjZo1>pxOu&jLQd$SgsvnTC`fU3BH!Q(QIFa%Gy~f8FivLG4OZDs58wO
zA{UwqM6NVfh}>vy5V_ObA@ZPkK;%jDgvg8L1re9Vg~*%c4UrGcM?tt^u&Csc$Bz}u
zsg{iao>TmrL|V{1K8cL3W3JP}LyvUS(2x}HTCp_p9C<GIFcmGaF0sQ&ND1hbr7)_M
zs3_flz7!WT6=gZIJaJ#Bs18Gig;P`n({iHDQ&|;zg+sR^UJ0j!)A4nwXapq!B0hx=
zQ6wc2A^}AJ5lTT76z#PC+ob)L2`X3!YCLUXah^HPo*>aOkuyOhh9wG3Qqf9aSP39*
z0)|Zl$M_8E-1y-9Sxkd(YXDvs;`d$d9eJWv^>4$wx`LgLyr>YW>jK;^5QV>=ShO!|
zP4Br;O!e7|vkh?X5|13|`3!YS!vbR4vK{c0Qc;nYXcaz*ii+7{Q-b2z@en1j6Cg@t
zCn_kFS@!cY)ti<ic`ycDzy^0l&bD2^6X$GF(Q6C!b^#Wgj-ieb?nU;UqxV_{8`801
zwk}%_AI>S_-Z1DJn8%7nxOun*?iU%2Fpq$UZ_bA((mWC(fw=%8)ErgNdR|b*FAWzC
zV)`p(l|pB_1F&`gh?@bcnc%Y8T|M-#Z|}#_`}W-MbEtS4VBmJ+BR$L@)0XMs&h7|N
zDtl7%#uXceYxqa)d+ukIPTs{rqesyBxRl8#Bb*UIkS1R<k|3sli4vq0rWH<*c9^z;
z@N5Ip|M>XvxU*pdGJOrZm|3b>ih4#xw*%GfM15<`3?+vM${%9K;Y;&X6o^ky$}Sdf
zs#;#dnaPUVj@58x3Ph>QREW};X%I=65{RTsse(Qi-*IUBh|a|F<V11Oz{};yJwC>a
zfq~wC-G)<}>_(0Ca~<VMGSxB1C37ZlcTtzm_#G`Qum1EfB0mg&Mkqm={+fzxkx`~U
z6FLDcf2|N=rnbK}L^|p^G<+*etB!0Jf<(TeK>XCnsMt)b_=O~+@n-Q5C72~ZlxUU+
zQ8Fi4p*<|KoRk&BnyH*r2ZGW#X)6hma3ppFNjXvleH~*oWj59N3d?L0ux<j11uu>6
zyfSKCQ--M_Q#BJ(L$PYHg8KA}9#V>E%1ko48SpoQ7-EwxfWHMq5xf=fw}NPbw*me(
zczzk+mjNMR{}Av$gt-QwH9$nzYXMpd-XpjUpmiXY;Cg`8!}A*e+5qpr3!uATZUOuj
z*q;M{e*pIXAmAT_=eGiWD?Gmq@Y`U24*~umf`2ewzkMrLNXOFb%ktI31$@iaJyg{r
zZx+)j-G{yk7uL`PUddp~Yrg#c`_cXJVeJ0<F(A_P(}alW$An1BPYWV#KWzp5q<)1X
ztgMg2Tq$rUQ2hSh`Pg#n$l0*P*l;DVt^|&FDn~0aj#LVcDPaaWh0emA!T*K=%0Eii
zlf18bJ}a`6|E*%P*;e=k_-ofy)nc;|?S-2P{~#f`DVb~KRWdJkU7k?hiIg{Qp9Nm%
z0(29z&$+n$Ki{ZCMX!z&2oV832|=QXZWPg7O?21Ek~5_khNVR)k(E_c)hOyz4NV4}
z7Qzhlw_+{|(bCagtgGkl;l=g#S>?NewcOLpb(!)18(m^xXk@a~)ZCI|ZDVWi;OOMy
Z#&*`XTN%JwVi6Sj9*M{I3kKoHe*rj7F0lXr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_avg_pool3d_bwd_impl.hpp.C8240F2385534214.idx b/.cache/clangd/index/profile_avg_pool3d_bwd_impl.hpp.C8240F2385534214.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d47a808cd0706310fbbd2b5cd21b3600b8845dc9
GIT binary patch
literal 5756
zcmdUzdsGug7r^&Il-EK+*d%}$o>CEk5Z;CoK%j!W9|EF+Aj-p_JX9=&w!YA+<!dPv
z6pF3&Q!1#`whCGUTJ>u|s0u<XmefjoK%iP$_)0nL#?#%M9{%(%_nhR+Z|<Ghxik0P
znc3LTkdPuihAmweS}e;fTFb&P%$WQY78PXg(893OdKgw*cXh?;+#f{5Z&9x9JBh*U
z(7UBI+{=f``5}{M0xfsj3=P<PZfEC{yl`f`GV_&|u=^b$o}zcRjI~a#x|dNKbfu$n
z;jxqMnKnBc2Ch|1dF<11`5C*I+a8K_zq)mP+NAJ;-kA**rQr?!rj7FVnwvIu&Gd|9
zy5m<Xy?YJ&f8D&@(z&5>YVWnoqLlHQc|6_NK;OUCD12HTd<*Io>=qs4UH0YQ*Dp@U
zwAM{6Pnf6Js&(&>ENw~C_4Zu{-S+h#QM@0MUQ755S5}BR7CU$F#lJp!SY*B6w8{Cd
zSDzgVSnLR={SA~tVd=7iH*OUf_J)=I(7o=7e!7v(x!^12^Eb!_D+kk#u9EILXZGJv
zv&rL0H^f<um1}aYogd0ue#x)z%;B>mY1u6{$(>(6{yC>w;H0Q6?R{kDr^?qU&Divz
ze4FRuTWPm{GL}|O9jpo+%x_s;goocR8;Sg;;_&bD3NjL<*_(R5y{5JQ^<<qw=cBI=
z8As@xy<Xc{GgIT?aC?9Ge(ezN5AzNPw>}7tx6FStZl$xCj=FmJ#+lF|wT{6d-r*BQ
z$8|prtbR*>sW^SW+@Tc??z4I$w<e%B*~u{ZmwQvjCLbi4S+y1PPX{(ioxV{7ntZM1
z4+fZ59|?#a*y7B~l4b@a^yyV@IW<`KOL}E}W7RpsH2VV!BIgaCGC5u`GS^=5X70yT
zsUJmL`uu>~h2Sr+w#!F;tKA|_a@!!&JJ$PyTQ2X`!0+>?w}pA@#H4KB8nLZ#$>5&{
zyT&A+_bas3+D1Lt`eXOsJ`&G=tTp_>n&j8q(oTnDe0uDQ6XNo=iF;EgFZ^e}YcjVZ
zzW3;v{f|#f`>6lys11>04eY8*h1+H6Ye|nZ&G==sQK>blr6tfMk$ecQK9`G;4yt)P
z{GSbvwWZ#~F?~IZ;F<AUp8sKbT7a3(1p-Q5Tb4UiDL;Z^d{V(-j)T?nKTMCMjm^;^
z6|jQL^@X>&WM8TBfNnPjHsC}T*ml8Us;RcCL1RN#`T-VB$X=6$Q`I=J$z#(JhFU?`
z3bqWlfn6IoP&~I=J~F!Kj6T_Wwm5q!Rc*Mz+$;N}MLV#Fao~;vK?&HVy@8_JKjxBx
zRtRr}2ui?-lv~S2ciNI2Wr?%sjySDCnOoJe9&NH(nB_`olW!AdBa~yABc^16%WqYm
z{jp}ofJKy;msn8MB#BGh5eh5`VyFqsn*h;m0<R|UVdU+=ZwFJ9-P*x};X6Rw0Y1Zb
zfw+tE5SS04{SSfH5YxYDe~a#)$IppbEJEfcLk5}Wnn&BEnK6%6@MK=n29h@LrP%!T
z)qL!UZ{|y6_dL(>ZHE_B)XwrvOQW}a_UU&0L<+1;tgWbUu+p#CyxWvpt;HfNtu5`S
zY8H{YB=uEl9<196vPR$X-9x&VwUM<krDP-hqV<{g_xdk1$94}|uj{+7iJvuKm#Cvp
z`1y`avjo)Zwe_DzFctSWQ3rS(*ieJUiF&~68D1iM`d#kawNE)LLc*2M;X^vHQ3;1?
z#3Gu(wiy;MEC;?EbyGP=<q&|%ljN|H;XV-dffd6F5Gqhsf~^wmUkOo4h+*_qAXK6H
zYM7^n`HXxFxMNK4vxqV97y~a<9ybOF3{L=eg3*^aVoS}>>F8_WL@#i9K}ad$!~k#x
z7!J9(du-$A`i8kUk-|x#!;mDa6nmyoiYSGS9#SSv2}Wd*De(-Q2gCDVLUGiiA5S_y
zlwTzmcNZ+&1rkaS+j~^|$%XAnFHy+F=>>l+%jCRE+ES`l%BC!dkL14c=sn*#x;Vj`
z!!w|~7AKtSo#-6{C!Fn_5puD2K}cjTLdezL6(O;`7$K>ilzws2bd9U(!cO;D8^Y=j
zMk?~UQdQ(Y%e>3zDbFq_+!{4?+@JIb8B1nJjUOlSSa~$%7yI>Ce0OVx?9`HN$)T(%
z{Os1n_xGlil5XN<?8Rf+dP%+L9S0}8`QGM?%v<V>$b1Ao^B9?rzYo2KP_>}{d?%&?
zeMfV+ETR79b#lBl;8g=5lpaphf=4aT?e~Xi?^0L%<;^BnfWs3|V<yR+Ct-;1&PPb#
zE<niC-4r1+cQb^XoSf)=g<MA`XM|jwTo4jDi4byiaz#k&Bu2>H$Q_wP4<iqRq#UV$
zGPO!j&JQRf%E;v?<&@GnuWB-F^$xa0AGyL`f&W(!PU+XxJKlfe@)oaIKJiiDkgqUi
zx#3F`Zw-5W*-vx+C?t)P!y-92Q$`*cT@-yqFnO62oCU#ICYgP_aXBk2X3u>cGW9!4
zM3lZJ*+(`*G{Z_rR6;VtDp;t3MGULKR}Fp)kHf-oR6YTo6W~qBKN;saJa}UJ2N@js
zF8RnRcY<FhAR*8N!CeqWHN%M;u=EBjW4I4i^g#;43W!xeBE_HXO{yK<a>;@;JI5`D
z_Wsijs)UrMF^kFFSq^S;^k^lAFgfEJnmxpB4Ct(`VgE4*r221JH6Z@v0Gn%w6B0v-
zF;$HdC3+=ve!z();5C66BW(s@Gqd|@*2-88h&%UzSs$~p;e-Or6ete?ZwT#A2|^{>
zp9;)WsJ<HbYCy8g7#NNLGTI3+oM8CNk=MTzyz}g{7dpF~52-3X<^OO%TjNQ~n$am$
zLa35Cin9ncc&UND1GE}F?%A_B2P2p4-#3UhBi-+>B|V6EiNf*X|N4@fCBF{8&`iq1
z_SvX!Y}SOXFj2UQO3v3$$m1%*Rz4-Gi*$?V%fj_rH#<9@Ola;x|Gx5c7Bt38|12wf
ziKefM6VyM=9$oSh4Y}L4WbY2YIJN<q5IiNcQ<DiHtRxam04tiAlQB-TgQOj#jQkE*
z++j|-IC;CXxQnt9Qk9Uw=&L}a0wjy7AxMp8$#D>kqw)#xn_#k7kFWh3u~8P)FXW~-
zRhKulNqZj}vxop*0DUTE5rMowgo2EM=+hXB2sR2vD9j}c*?^XdmeY4`77^|lj-2Ep
zoFiN)>sdskC=$_(=0?+JX%-Q^DB6+HjN`}A=L8lJ&yPnafuDfT3c(6GrdULxAQ7P?
zK@v@Ogcs&09Id)$rK^5}s}6Dd-e0K++bCf?6}AHtb+K7B@08Dq&!4<|mQFUA!V4{e
zYAER4-B*S+w^6B?)Rr=QB<ECuk4d2BO0mXd%`0?u<*DE69eWzc&mHNkbbYECCo)(W
z2(4qSqe(rzYAWLJ9xshmJfQ_;)w8-*xyu=AyB8Y#g!#s(wJ}Bc<ZzmS)6Co~aY6|k
zCCX~hRReu9om$Tt#w2fcl2*&XMh<qACI46JNsZ%M_K>43vo51O=8vyF>~NNJo+Yb`
zMa6Vj&a`*iAM!0Gugd?LCGy+WM$Jbwtgs-XaHpoA`N;dvqxt0Z&x2D13^QbkCQOmX
k6b+ann<)})-LVfsypI;^^9+m(*&Ji8362>tat>4c8<bq<1^@s6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm.cpp.8A238D0EF3B6DB52.idx b/.cache/clangd/index/profile_batched_gemm.cpp.8A238D0EF3B6DB52.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eb1b44c88a963e02905d9f160f423d3e2649f11a
GIT binary patch
literal 4950
zcmb8yc~}$I769<M8Z<&A5MW4T5l|Ro2m%TO1Pla_T?n!mU{qEqAQh~lKD0iu){2$I
zph2SoQsfB=A}W=r(WVqBBBCr>RNN6;MXT87(l?16=JxCR2fmM+`Q^;qdo%aUnHh0_
zpI@i}f{2m=ax;@xFQ*|0qEGzitj<omHxWTDX(LGP{<^4JD{l$(Zgf72@ZyD@=yw*H
zXXm5~))cJ##+_y8R7{UJTza@T_xwgC{|^=xduSDieYAhjv}30TE3hu;ueciZAXu;@
zez&T9&y1q_eZy1E%j5BZftN?SUSu%ez7I6-7<DdqgtczC&_87UqAbCE!`By{J$m!B
zHQILL`@HtYcb+eNpY3^$J>l;ATe(*ruie9}eD^$JqpQ}-1NUyd{-T;)Ev@6$wfwX-
zdQn>9!M0V)^YQf$BhrR{Twj!SPbd4U)n!eAk82#or(N?de7s*ONe(^p5HNN&ZoT&4
zl;J2gWrLyQ{`Oy;WM3M|;*5fK6r9{uk<+8PgJrjeaU=W&tp(TrY^XnmcKLIR_G0nN
zGHu%}xCZPuMI)EXbW0C<=Kb+=MJlh-Bj0Jaj=}C{kv(^%^M6FG9{#-Gs0FVy!csEO
z{ouNw*h(lbxqfjwGXD-QbBC>c{?c3Z<(0;7*^5K!UnrX#V;=X;_Sv@Lv!*$7r$7BD
z9TDWagqGe+@m`p_t8pIR;K#D$%-Bx1hf$SM%P5^~ftkupZA;$F+|Oz*z8d)TMQLVj
zM5l!JJX}9&)-~51zG2=x_HRS_8|s($hwU=SFNpKq)~B1XZ}0T)0uzgdbkltfWcA)|
z^ggtG{e>cIlx8En#MN^7MWQcB)l0qd+4ZT%RAS~fc~5jtxkFz+SEp3vQ13l*+ht1L
z@tbu^*zYp>8b7!3>upx_tP9s`zgc0?Dwz`~;YpYa6A$W?tmQ0CeRbno;TM%%fo5-m
z6O;$M-4{9p<i$Cp7P^1vE-re)KN@A5M}LxX`r7gRtyY3ax30xS6;AY&*JtaZ&lQ_Z
zS)y{;Y-QfrdZjXd(8uuA2>(EX|A(1h<+*RP?T)fLBzf22XlcA6=1f&&-9=US#f*z(
zVVkEX9g$_4oovoFP8*Il<)>D2zun309vr&6;!Cfcm*ao<=7Y3;&x6E>?73MjCFg@1
z(p%76o5Ik`>jJ*;C=4k%JS*>MZEgR-EhWnEGHyxKe~KKpkLurXFCMOmxk@+9?M=9y
z!z|qV=R@<pnUc{vNvpGeewX#qD0Zcf9<xDb&DK}1qynQOqwVsTo6aqu9=o0RuUGyy
z5uUP<AzQnYk~4cl*VA_BZnry7J+yhxq=Y-dRray1nWBOE2ih}&4#~Ptq;*I4A1NGS
zy47!boUwWI<rO2z+tV$-RmZ<FX<KXX<j7+7v-W6EJMCm(5ZB>p?Hj|s7i@m|rKe9a
zwO_T47<qD6vlh+?ExE)Fdo+=~W+m&G{&ma6sH<i8bm4OQ!)71RyGdBqH3jN^Z1Tq4
zEz!R}SWvLB?QG7v%!FKkAS|)A?eyx7W+oE77{N_wCM-=u)B??)$xlmZ!Yx};m`H3C
zg6q@t8JcD+JX|Xu-pRkuL>3W8Poqs^X&N#d&K{ZU<#5uPiAac}^=NtwO+$vG;{RYx
zcZ=v`B1?#{0nLD+X~=Nwiq|V-=si6q5<`SpG!{eCkl~eQzdw0HG|zl0l7b-k06+$S
z$zLCcx}Fh#TUshQ|I3@0tA#`myPyn|t!X|55oCX{h1Okk;|B(kN`x(GmTXN!hELop
zD*1iEvnDMhk_hWEbWJr48UE$uw!J2Mqbrz*nE35%8k?zU$nbCXcn(WqciAzKa3aj1
zahRHh3}+{&<P0zn1UwH86)Js(CerXEaZ*HCc;t>y%giVmo-9s=lp;<U)1>;*m`%*|
z+Ho90uCkW4W_8K)!oXDkQvs*36M?q@tPOStz&c=e0j!HMnH@y8>eL`l7z399u?$G4
zZfk(J26i)uYzFaEybVOQQ8xK2tB7vZDMih;31Cgs=nT9WV9l@<08_wL0!&HS{<5<H
zKEcolF!0ake;!PpYB#-$@+R5Zk2(z8hHj%vO$wu9kY>?mLE_T6kY>|oL*mhSkZkF;
zkoa^yBs;ntq&f6CknHL9V{$Vextu;}`3Ruly@1sV49Q7)XC2|2FK8Z5DulQcXQBLp
z!ifw#LzEFpo&^S;DawSjQnYeR<Yi2(hx!_(mRSlyq#%-<3<IwM;wpHVn?Pg}wXEuF
zi&Rlw{0kn~Qx?2n9C=Ju3}PngazV|c9Ei)|ksE-x0k#rEDyfkfco&H5q8uZ<$?jZp
z?t9|=eySs6Z6Qk#F`l~qP~*vgSPu7J2gG%-6(CYU_0PaNKx7By`29a927fj9<{#rN
z&E221Own?mNYP-UU{mtUF0H9KbyVz{O4Ls$p_32!H3N6{arUR=V&gKKlE6j)$<@Xc
zlADbiBzI$X6Y4k*V-HB4#-5ObW<pD9PcJhsNZw}NkbJCsY^Xhbt$ZQ*S@}WoxA31y
z?wK6orP$o?T>|m@L-a$8$*)sb`TegaU&SV<*T-}K%pliCZR-JGOFBbbb<<*mrW{8~
zk%xi_<jv!+61}RaKje-17Wp{(A}bY3`)~IdV%7LguH=tl;8I{H1s0TLfFlFel;waU
zhg}Xh<*=&&rwVor;MBmb1Drb84S>@Cy9scbU^fF!Gi(LmC}1lAM+sX6I4am}fYS!M
z18_QEcL7cp<^00kr<lraIkCo{E;jO(m|wYReaI12>m&vq$BUasRi=1eJS3clLrQW;
za-;T4c1VVl;*c_?Z7+OHT{{b=5u*kf1&yUNd8i%@-viJ+U@YZm_&$K{gMAR72VqwN
zv=VkTK&xRN1?W-O#{qgAb|XL=VV?%*Y1n50dKUJ1fS!ka380r?w*s^kb~`}ZVRr(w
z6LvR1yD5{|H$nSv&p_%mvKs*2<W*9h5kw0N|LUIRPUPJS#%fVRT%Th_0h2|X-(E1a
z7wD56k~NdKCe)8m<YE)Gi|_0~?xemZ=1V~kWf>64Kp15?5Xxbf17SJrDj=+aT?2$Q
zu<L-Z4t4_&Ho$HIq9zbQjn@oB&9D_fq=2miA|-4U5UF6d0Z|+54j}4)-33Hllp9Nh
z7JApNDgJRoIO8YHJtlnIN93oIuTuayKN{{F;2c89h37Jdl7J_G<jQk}<i>M@<ZkLd
zR)=V~hp7i7Pg74wLM!1|&8FdAR$h?2t-K-mU_Lx*F20yABtOg#lE1Y-m)!H*JaLki
zZ^bZ?Ka04F9LTSgJymm&E9cV2QH0xlCdtZAxl4(j*4wOKKwjZjW^DNS-<}anO@apY
zf>3gZF8^KaJ5IEljAMvfTOanGyRUVE=*yAr$RK+n-w24O1@-Fyc{flex0Mv|$=%eC
z$CwN_QkDZu4!ay+<&+2U)prHnMLOe{;4r;**95ovLSlkp+^~7%NaQ;SF+ueU2{pkg
zfK^cgGVmIJ)xfR;SRL#JfHhDa-Q7BQ-ONElP4>do0?AI!SgyX}bFve2($yOpsW%X0
z>8B5M8&7R6qc#($%{Xc^X>1eVC7<26WK-FDgT5>=DPej1(l{h>EQ%y%CS{}%3{DN5
zNYk1$nLb5_F;!oWsT;`h^O?oc4m4mJ8gZt%&G+zJAoTLKNBQnE1+y*w-?zr5X6BYw
lGpsQ#&(_Ytajuh#>vU(6IrDr`V;ldV<OvJIL;)ZK`5(<A8n*xd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_add_relu_gemm_add.cpp.CEB2ADB0FF2D0A07.idx b/.cache/clangd/index/profile_batched_gemm_add_relu_gemm_add.cpp.CEB2ADB0FF2D0A07.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c6b382c5df4137591dc6c748884528566982280d
GIT binary patch
literal 3930
zcmZwJ2UHW;8VB&XG>0S*z$~zl>?j6hX(Ct%QU?hwQUVDO3@vmBti%EqYzP7ZVFg{D
zpe`OOz3Nj$5ufF`0`|7*+V=UL@+`ZGSU2Df`8>OG4tM64|8UEhd+&_UkIxq}G0ZE$
zFGrrZI8F=0FhkX!VsUorW=#wm)WxuznnsaSanp(QYR8Se{r^%Xv=8+^R|a2qxWyH8
z)w(QC7`b_MazM0S!C=fWpIDN7JD6V^cYN6F?yh&c1IS8)8*P`xEN)H1Y0=1=R7<0`
zE+g}{-HDd2u1&hue&Lbf^A{Ujqs)%lN7d}$psgl8#{-H+h7RIC=WUN({fW-J`>8sq
z<lT|TrJX;N`(SNP&#kPm?TKEs?e^`v3E^`+^7t#O1xru5G|Kx@hrFkBOLjL(+t1}2
zE`LxJv|BED;@!lHYIAJ3)E#=*x2oc>^&8H|XTz7RmYt9X2B*~8RwXS{9-0^Q*fXUx
ztNTuiZccpYzVAEEM7~RLPL{rxw%*E@R9msjdw=}I+`Xi9W8lB{d2C-&lq~zUaJOJ`
z_vusC=}TMQKfia$r?^61TV{~<PlIWeH$t!eawfUoc$&$>YgT9eL9_GTyP|qs)M=sC
zpmu<54gUN$*VK7oYn6GjzRhOt=BIYD3ksP{n|qy)?NxHxQZKhIaGUj^;)Qd5xTpB$
zucXf#-*yL%uI&7t*4n+C+}I8WLFspoFQfOp{TKb6ENB+J{y&@$NuP!`-?V&o!O-ff
zC+p%TRu0n2d|a}bb223VU$nk5q~6rPb-jGj6R!gf##MLX8ZC?a+K0nvj1Z>N>!;#i
zq4qi<-ZxY<$XoA1PaB+@Zye>y9*T&GD(u~Pj!xThpMAZ~X;|dB`sz@>O{sK}gq|F~
z^hnEajfq9*ohplY|LfkD6SVl7wWoL{d#;OC#+*0F(#n*ShTL46ZX6)lkW(IC@wPku
zd~EK=gR-eJqr`I2TbBL#%yU`mPkViLAiJ({*ZgGVPp^juo8!~x8Ml>9_cV_lD3s3E
zy~nRC8DZUR_4_cbV0Cw1WxDm&ov#;#+C4PVENSUT3o~U-Y}4)hvB2%*pyMrG%ZBU+
z9lzN%Z8=xMx^^n;zmMbX&0<4#DIQ#JbvIfrxO32A$LrE%FV7nd)DH8v&?`5WG~C~F
z)_!ZX?6zy9=8Z1D;yB;hD!mVGP6Xr9!s-rjW98Y0e4AXIugwlF{QadS>z>|A&255)
z!27FjGh{2WuKE4`-f=;Y^wfl3uGO7o9d_MU@nS#fZax0#!6LhM+h^k0`UhBD_kZ>F
zzZHEo)xzR+{*~>TrTv4_NnMH~Giu)tJ!S7;)tvk}z9uL*r)$1rsWPyl?aW4R#k&07
zyzDnVC-Ldc#hr6q=E&v<@*0cIuJ)~4{ti>*%Hw-fYt+_BaY0{Fr^mpg7$#zAv8=`l
zRWf$W)OHtUOWLmLVGC996@*<ulm0wmqsv@%c$dC>k=FD+g02ywx`nPo*BvW9hZwf-
zLSJT26aO|{BUlwPb(p$i1r>iU+}r&6@j)3~L#T=wIt<;hf{M4VeKbv`-`7Uh2vx=U
zI{La}1r=u}CMh&%7)I!?<rU#S-9rq~0Gc>}CR4+jayUEX=^B>;q@@b<2aq8ZP-hL=
z7vp8924{B#g;C?C&PanA^`A{U8=R=^tX}q!#t<oRg+5gsL$nxQ473Da0+fU20OjJj
zK+EyvKr8SSKr8W;K&$XoKzVrHxTr^=NBiOtFhrXWxe0lzJ6VdzQs^=yC__Q&yd06`
zFjpdh66O_%tblndB3q%`5ZMOZjs)%S|2q)b0rO5I=!AI}BD>VNfAz8Q{`F7waE53j
z;%`I&>W&s8elc_z;+H`y5nl=2i1>}r&4}L&-H!O}(4C0isdj|NP1Av;J%?4Z>GM*c
zrUy02wP;^VZic83k%j7kFhmWAY=CY;WRu#`nv+NG4NWZmOZ_qX_jlC`dC8UxksHs=
zRXu)pS9edfJP41eYIzc#KqNr|c@bVfK0F^#r?0CosMC+|19$op{y+gl0FZzX00puG
zzft!kh#dqJ%nk+$VTS+-*+L*OPYik%?ivny7C}US_9BT$APFG>iXx(bq--f@A(|Zx
z6vK`Iie<+F#j)dnWTrAR_4pD@6Mzy;6M>RUlYo*<lYvr9Q-IPqX;aj7rgPGPGB_DP
za*iA*lamRQ#mO31l2QJRR}%)OYmClVvnyySHDgC@aRu2>&R14;F?%o7s&>%NQ=8gi
z3!i(N@dh@rRRP_AZcJ@4CH|WEL*<{G3>hLT+={90wKZ-HWP{rP+2Xc9cDNnT415NV
zJ#G&)6Q2p>fI9#=;*LPG@L52!@!3Fg@Hs$p@wq@wxDyZ$=K(q6&f}t9SBUnvuPGYs
zi`NrFRDj3=_4*#&zXTiMMpuN$BA6E=K{3or5Lp8AdPLSkHzGkJyuTR<nxR_|*`mI`
zeT&CrgNqlB{$+nT#IMR}^7zknOD&O`n;S_zO)Zf->CRWngX>|bmM7N}h~$z$UR*CA
zA2%ORr!VOX>h$CKfjj-V{y+iT03ZQZ02D|Bj<0DgQ4kRX6ifsIg%BY?LP7{6b`yi1
zg_Gf+XA#^8&|V}r5=g?807Y@5fTV;Jv=B{11H}+AK(RzDP#h5lB;&}&SFx5Tfs+7~
z$VmiB;v@kjbCQ8lI4MAB=4sQ^&oA9P9Vo**14wQz2g)?h1j;hc8duLOVS=`I-4oUB
zpl_jXOMR~0+VJRF&jal?)z51Xz6LR<R-fPv<r>Z{N>#wLVA@mf&0)OuMDMQHLmHzi
z#<&1^Qui}N^~kth?cvJP6LPutSYtC7eyC+9X2gwdS&D3#LiMSLVPT`i=R-_gU|7WX
zB@%_oKaF@Y3SQ|%N+QC;#F%6}!y?s};p$77JRu`hML}|)rk1wO1YDQS&@(h(>iZk<
z=UW=l{3n{SzWSQ&JkQn5-NTc#H@0=*I$2HryI)PrINwY$pK3AvTWcG;nGTM#=kka-
XlV;5FHa3~&8xW=86XfNG7GQq>2;d-A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_add_relu_gemm_add_impl.hpp.CC900C93F3C48179.idx b/.cache/clangd/index/profile_batched_gemm_add_relu_gemm_add_impl.hpp.CC900C93F3C48179.idx
new file mode 100755
index 0000000000000000000000000000000000000000..67b088cdcfaf1603f98e5480324176fd1cd857cc
GIT binary patch
literal 7496
zcmeI%c~nzZ8UXOz%abLBzyk`Dtq6olK*R+b79k3<i3?j57m&?XD~N0&PEoN)tyb17
ziiki3QT9c}u|cHhw5SLOsWoF03+)-Db(;|n*5kZ5^YXqqo-_aWm*<!ha^5fB&3pIV
zd%54eNqBHTfY6%7@=FZPPEN{*BUmgJN&Tf~q-|P`S*&(F7AyPcw~_0&mb;mZ*FC%u
z+*o#X#*2XUCpLZ`A%7aT^Q=|6af7+#EywpxP5u0vl2uz293ctTO&?e&Ngqq_IOnzF
zKzzinwn5se-Z34gq_{hQs_$=i^nK!!d*|+x{ZB2-?g+<J66(cZUFX;NzNx=^9WnOH
zz14P4@$8sq#qC3<mpdPMcKLkc$zK|8mrp%CGWw!Ex#G%FrybIu3)A{OK4{eYO=^PK
z9An4Cy!Vgp2~N7J?BDoBf9KvNtHjmX*;*m2?awm*Rq){|MPy&QbIR$vgX8n+zmiP0
zE1M+BnG^ZHEn1f1Gwa@PwNLV3UFZ9mkvhWJhYvkAbhoHJ<DsbJ^4!?I6E9}|Gwbu%
z?B~BfTamIOtG{M#jAE|+zS^IY;(WXFi+gh|`i!!rs}mJ_Hh%u8&&98%8hkn)R{hYo
zyP;r!t9;`3r`F5L4=zsTUeEixjrRN)?F{29BhIVR`UErA{xNDkqwELH%2UCM=ijbg
zC(FC!Zy2+=mJ~(1x<^0u)c%ebUxL3(nRw!U?a8i(_F9XV<|lDad0Kv2VK#HmteTaR
z*_ma9^*UB_4Qpc_IS=0m_o=yhE1Zm);%*M=VVmf(&;7Q4&5GomT?fpJF6?vhDv=a&
z-Hd-SbdE_b+{!9-=wW9MMh3P;nqNNbR2)@n$qBKJzEa__b4F6QC~!37#DeEHUsjgr
zhfdsGc)s@YXzM0l`-s?&zh162;rp=CfVVOFx|eohN>Rw*k)hVs?KRr-4%G$)Z(3Qk
z)_N+*v)!O=;dYa}Zo`Hy{xki~n%w&I9m(-RYIK<=Je;Uk;z^86rctCGqzT)^h(!%H
z^<z@E;!;n|hI1U2u{Mjen`Y<s>V>6;5lf3eBuY!#n11hls$^{UqU-aV3=oIBhMlhA
zx%4Ns=1i}zyc&Oy!y!u%zZA`;101pp@ynn^h%bT`Bfc0~g7^|>DOw;!Ud;I#kgx$Q
zVcdv>jnGX<*aY2-gw4<`NZ111hCJGk4}CuE>)rP!_$};d9P$|A9YYJ~0EetZyh`Zf
zh<6;i3h}C-s}Zl7vHm9(A8k5qbrBO}H@5G_Jo>~0slu)*>_*$L%O-iAC}@DKt^xIQ
zpabd}P*33Fh&&Emg~%%CYD88;pFrdZ=vqY9Lf0X(4*Db_PeRutvYxTorQPE8JJJ^>
z1l8v@rO1}9sM>C&NHaJHY7E$v!FU;RD}!+nvJt_!2)T)1T#RhQFfK-JVi=bo8wrd{
zkedX?rN~AK<5J`;W%`^T<;X&gteE=Ck&7I;LVL?$`C)86j4hb>C^jF3R$x;FHfQ1r
zJV$}&K}RX@D%$4qR^31A&d%i!WK3wxGWu<Tj17$i+7P+{C@wS(C_XeEC?PZfr~$b*
zFyoLQ8<Be>bQ5xKf^J6c&Co5#y#=}zxwkU581CD~y62trZWNGjHVP~kJlI~&D?I$#
z+-l1%OwBhp5>qorR&>@nW_aX;<wVl5^X=$*EV#jbE60$-Sl1Qy{FunsZr#=N>b8LM
z?TZjW<r}MO*qXlG*(+-9nQn7#c<mE&4CV<69%Oz%1$vNO4+^F~F@NEA>>qcmF~J<N
z1KW0BM>?RkpaW|%fTzu)A3oELdqk?IF>Wa`qzkI2F>V>O2ysQwV#F0gOAuEAEk#@@
zv>b8e&@yZ)gVTl#yULivK#(%*C&QZ5r%oiM-LU)+wj08l{6LT+*lq+`iMdJ)?njB;
zluQmMNF@$Y;!x&(Rd|L9+tc=PX8Bp1)nQXDIEYOKu^k=YkT<c(O~yW!-qpMACVfT~
zR3KjkieL(kBj0hx0T=TB*fO!PQinrsaoM8D7aTI#B^fBiB?V}!%T}ONmsFq*Om(Ii
zbGc4Tb~0Y}QOQ=*SqF=#vEeQ7_M?kM=q>p+Imz%4Ro#57myZqT0EcYG)Khc5#Nz$?
za-4R5|JrEWlortYvY_K9>cm1Hp{9d2rKM-BcvdIW=8z4DXh8bRi5n5o2;GE;Cg^5F
zG()!_q6NAY5v`0r3y*ajdK~+Jx_*XXhNiyB;?3rsw{3o<>f4KLd+`kV`l*5k+aiWD
z%m@zY@96JNzs(^790Py?9Rq=a9D_8<?p2ND5QEpKJ1fG})08<y2_{RR%Q0CFU4h98
z=qgNBLDyiihH+kFXmV7=;1%jZhp@*X9K>9x5PK9t7h$g=9Kytlu}3kCmte0F7%#;h
zrA+*bnq4`|PWD})7O*QFR?cEh6Xp;G6z&iX6yXp76zLEN6y*>F6vL0vq<;<>%Z~-x
zz~2BA$BzSw=f?vj@DnsT5Tf<P?}aY~R3DaO%W~$wk3+SuWd(E<wyc7#!Im|Q5AN|e
za(dg-&FaD9mZ9llky(N1qFj?)YVhP^!+fk+9f>A4tj{yB8llR!VfQw?h;FN6i+}dE
z30tcFZYA)BD}m#uKDl(WCv%W$;l)0SHSJrqWhoI7w(st%9fWH)ObjaQqryVEBdQ;~
z=QMWDeoB{0ZHV6iH&dO6-wAyN@y|e?Mf|hS=MeuKbUzO0$4lu8amXPYHH6nN9>LKg
zIEMB&U#>npo+-1UHd7}M?*y}G<WTz#UM+MT;?+T)M7)#G^@vx`xY_B~sI`y6T(s3W
z(M|?)oD6%&ux7icj)%*zW|^eUaUUp|nNV#nC0<C^SiKqJs+b(vy6(rh)dvlXUhQl3
zN|`BFy|2|PgBBq@5wsZTiJ>J(PXaAPdQxaP(vw5WFje>W=!&VE+IonoIEPB^)+5jg
zOe$b`CDv17efsmA_j5YUzIuG>ty~{=asM%}FLg7E`iJ=6uN}^Fb#Gi=J+|!Kl=FsD
z&V|2?L^pap8GkE>hira$;<P2z_o?CG!Sm2$-Tt)tu;;%dr@tE<Z#X!v8v4th8%%h<
zYsfcj$m{kyzs&QAFnE<s1f|HEo}Ja%L{J7TLIM%A7zxDC5+sm7OOd}6fz`1b`OBfZ
z@q%vb$y{$3+YU2nlSAE@?I^SYuUFtWro0LVsF>CBfVbmc!dF;{-^%G?a5Qt3v+d<w
z>fMi-zvg>bYWHIn02F8z2oz)%q|vC&U#H!2w!QjVI#LM+{3aXB(OcOpV@ko6b>C@K
zzP7u)uiH?FJzLb<W>21{=6mY<o13CMj2HJ&+vY6etPh!?l4G2s(f!h<rxnfvmDHEn
zyNF$ErqCWl>_NbL5wRCK9})S`1&Ans-j9g=&<7B40J;znh0sNaD1t6VL^0#XN4vUb
zx6E*OD?5&ERK$Mus}22I&(s-7kE>UIJJoGB(ZZW(EnV)##p{koyt9w~-;PU5+7hS!
zX(&A{A-zz2z*WEe=V3%2uviAnp&@gq%N$N;4t1Er|NiczeuKljtIr&gLZbg=>E7<`
U(~Wd=^`>+5xdw)87Mm&XZ{5=>J^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp.F57E6277DAC86E29.idx b/.cache/clangd/index/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp.F57E6277DAC86E29.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f862c54a7625316bf76eb6ddb95100f717ae4a95
GIT binary patch
literal 8382
zcmeI2dpK2D8^F(6XN%2tT=vFEj;MqhcOjQXDN`yQ$_zzAF;S`9&Y_uU=m-@`C6`>A
z%INZS8xteRG<8hJjEXK_LPU3oq|`LNeVT9Yz2@tAzW(T+{%Aj(XZ`lOZs%R^e&2Vk
zef;Kldj4d@<9US5Nr(!M3&nUmo)-Nd8y6E93wgZ7sytpo;W__S@7vazzn6C>=eSR*
z638u<X6tTCaNnss)D}||x$H`AgE+~utK(Ybf`YwQRw(Vv*)#7>;jyM)mlsUD>tf|}
z_dwF}uFJj_N2*UHiuSr#1!~l)-3!$aDGh$DQLBE>SEE+rUi9NTI%boMUnkBNz3w^T
zUU(>79-1~a%BJdU;o-c4g5)%d(dSch3l|-2)4pk5zuW8h&E(@ZG=r1NvR3BYRdS6P
zmAL-1-i(9yJ(MyQ)PJY!994EiuV($0v)>%5oUJzBu>J|Yyv%ZSos(m3%sEl^EwA92
zFYZNcy4b8D`m7-u4;$r~oc?0Yb>nqX9}jJd`sK!(yu;3Jo5{B|*uFco@pxY6wc74e
zj>^&>9pSejEy}B3oq25(e6nWa;lkC@1b#v#SS<NzZr7B7*Q=L2T0Icme_Euh94W7v
z7+J0Ruy3jG+}bf`za<>cPVAcQtX1&y&dmq*-rpKGcI>>N?dP&^xeVXklUU{N=~KnK
z-8`#NYsS7^NnUf+Rt%3nFh~k#=;*}KDqr;ZjmyUZ95z-h-DAl+mR~2haNuQ1=F87A
z_xopdMLH&DR4g;0!c_gV51M*kzMJ=A)^YI@rNj1xrLn;^Q4f-N_yvpciCwy%n~e3<
z{btn<`6pchW{jE`;y!$0$-_;<#h=HuZXRiyYlkmh@x7II)$y!#ZMXcEPmulfu>XGS
zaPHL=G9^iDw@Zi}JijWie0a3iwi}Hn5BaZnsDHc8`?uZubsg3OH+P)MIyyExaQ*p?
zxb~zw$$z%4)wY~`DRSl)f_&}yqgy3^b)4PMDjRLP$Yo(`j^V3a!7tSkEwg402s(T8
zcl-%DB3l;oViTj5y3;nO7NxFJz3gfS!P6PVBTbY{oZfu!a0qyO43I#$?8&2wTktuA
zx=M6?`XcpI^|X2T-w+f2RqQwsUH-Ce`X6@Y8h{{spm`6pWjf|9T^H^_etnS^90X$y
zg0ZZi&%rReUVe-gJOxvpf;o(UZ=+UG(zrAOC&)O<I6LMxLB?ChPvJ;xEk;P<Ecuj^
z6#?fWfN(KzD+V(-c?obULAeyTm7=@{*zW;TnfjytsC)Qm(ot0<g4_s9Hv%U{K#-}x
zES0NBkZHgy4drxTmX2~ZFwN%LtJL@{$j*grv<e1+;~+O<W&oi2W;K{KP+zX^7<tg(
z48+I=IKBazF!eApg>RX{w`N#F*fPVqxqT&uk*&a}6&N!DjBE$S?O+_ooxr#g<!)fy
z&9SC;opID7>3BXt?gVB#(XWyVOmn%fLXZW(tbmhiRvSjy<;-r<ROpQUo5+l)&>58q
zOgT;iR2s_ZfJ#Ss3m~_k?K1$Cfyy%hm5Fi|pt4ZT24ptcemkJHqw*bq+JW*;K<z|1
z7m&GV`#eD9q4In{<)d5xr~;046I5m24y~U<ud&+%H`b=~eo5MOqCM|5iIH(eamaea
z*~d9By1~eJ<M{CiImENJhmm4qF(MN?h!L5@NP_e!ahI^hkCAeqF9!x(uX6B-99W|4
zDhJazehh_=p+3h?pzsOG15h}C@-rxWhVn2J4x>B*bw{9(>u&_wj6n7=$H)=rKLWpC
z<icymRQ@bKX^UfI5wIviCt3`wiUDi$7+C_WN>DBZR;4JH0gE!UeJdE#iYlTV7_<Ws
zGcZPW0)tMJyMaMB$NJ6ZqInNo!#}Pdy<-1w1;;q_Np_oT&8MG$e};x}pZ-`q;B-;-
zgIn4O-PJNx`y~tCnz8x*aJT&LPAmo(*#Zn&IIl1%{*mgCs(wA~C$|H$?ZBO}B7)oj
z%yyuh56tp8w$yEm_Wb$8bh^hLXwU;$^Rx2#{()}v<U?=G!8!;U2vnU)X~7^c9^?$#
z-tAY_>xqjfI!vWP?Nlf{-Z4~EuDth0IxIo$G?X_%?M*1xLX%oJnVWMhbgP9P4BLPC
z(l=#a{ZV?vN-(Yx%wh!2hW!^-jtcG9rk$7|PLsJ!yDdRHLSlg!A&EeOkQ@l)0AU%_
zm7&f-1}$aKn(J8xJ!FubfFcg+HlX!;p?xnz9B3c3??bsC+V`V82*(Y=30(g}(0&M&
zOZfsRADN$&Z!P7s$qhzI`Cd}~Jg$8T-#CTOx@DKiJP-Yo%8JaP7gD{D^~nU;2dO@k
z`ythj@>57XMR^EPLmW>V<GSa&uJDzRAl+TueVFkHa=y!amOM{we7x$>S0!|%TBubE
z^|)%Mq1I`R=R0m+qid0|jb30oXFFEAel6egAJilUeDrYd@Nf(8S+!B&npw1>w?cd?
z)Zi3d1+}U;MK93taXK-wrGr+qmys81)(gG0P48v9DiH)3fCnJ01MxtFg76@Og7IL4
zLhulTLh(?9!tgMJ!tro~BJc==BJoIs%Ar9yXL1Bt0SziR4tSMkbEigX|K2&~mkEVy
zLTXRlrXP}4XwwSWG=?B=Lfe~=&6x<Y4cfM$+zxHp(fS?Gwu6(0Wp7GdXP)%iTfQ}%
zs!TA9f4Tq7h{9ILrpJmAg*Q=dgTgkH+o7-><qjz9;5afSR_ybl+@0>Y2w+7(oil`D
zfEA-$0<aR4O957jav8wNIF8!2Lg?ITc!&NfjnJr(8*!~uU-W|JI6c~;T=iYenen0%
zrZpK{xcpibBLhSMR?KaT3={<-6eJ2_jSwS)MZpM#h(Ztw6@?-cCJIA(2^WPU6d{U0
zYDS78S(1!QZ%7hXxY2XX17>-E%}Eo^c6P7BF8xkN&qAnG2z8nH5@b2lDo42jYE^K&
z;qbgD|DsDWTHkrlCJ!QEHy_&NLv{pudjjnWP~Ht~cBA!+pkWcTU<Tfze5Ne@mzX4a
zs&&At4*0MFFMdkz))6Pgqe=ygVQv#-AQgyE5EX<_FcpkY2o-`*C>4rO7!`(4I2Ddi
z1Qmf$Bo&EJIW#SYY!0t@R81>5&a~b3m5YYk4cdB+fH6ma88hu1y)@nF>#6m$U@O$z
z3fWyj&hwDq4H^bl=<i<xr_{jdOpC&z$z9+467TrlPkr%>4p&3px#}N-%R7Qg*{Zpi
zmv3s<n@zuE1DXt_@TaBlS)1R}r(@pzVEjupg4_#C_JT>wni1qaV7d={!ts7!x*z3o
zU|NoH6PVh>Z2<|g1x{~)o?QE0=-10_(G<yye;-`L$@`&yKgxqJa1bux<j<kUb2yvf
zz3Bn@J@KavFoNFbn{Gi56*7QX26w1HkeR?N6Xh&mmc{WA^E>{F`+Ywdg^@BiPKM@n
zGB{NRU6>ISaz7bd$gz|^R?27NfFiGRk@DR*c?w^U!sUU-g09={$<WYNr;n)cIK<1x
z<M9ZIaWO&?T!N4s;BwBt6ZC|%WKf40P?0ZM^>Vvng^I2EQ0|9T{U{GY(I6KV2yzHo
z4WaT8s5OH2FXiJ>K0BYvhwCc!HV+*5;I?Abg8NOWYgh1SNA%Cl<Ie(>i{}Hn=YRAc
z@9-a|8m<H$c6r+K!7Z)q@+G<a!Qh$q4kZ`HBy`~k0wvl*{s1+9fb8}QBM(5$102g^
zDtuqvj*9(gGw-mOM$MVmUTHjX{-~&TDC&>#U#D-Fwm?G>2wdaXP)xt8m==#DpW<m^
zge0>i?1=$EAIHv=12z^C^ofQ;19avB!DDFjn2Tft`2-q0L3scg4WRrC8a+dK7#a<u
zJOYD9U?|sr3g0t@@5OMF>%?#Td<=GdaNgq5h-g;CMtvPR5G3M>0%pEf4QrJCJ=v)7
zt@v=Y*E;u=wCo%$uy?j+wSRrE&@UoybJ}}5old7FkLQPk;Db|C_bsPYLBI0p{pJ6h
zC_dcruc4yJO(nFgMqyE^j0OpEHMSa|HP{-2)?#ZBipHW5ios$Kx&%}%0d`#o@-k4l
zjB+zjX-2sPsI;Jb1*lv>`6^Jkit;s}a*gAG!s<G!qs9|HIQ75$DYxFzMSg;|y9VIa
z02VWPjtO5Cs(5D_8?!97Tv1S0eEfYWehu*7OPeY1m{+Wd8_L{W0aw)EiUe1rxS|SM
ogyipn=IZ<$O>cEd8COwN7ief|X{!-D`nLznrO6cN*TP)!KZmJh-T(jq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_gemm.cpp.B214B5448CC6B6C2.idx b/.cache/clangd/index/profile_batched_gemm_gemm.cpp.B214B5448CC6B6C2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4fdbd0130a479eeeebd342669993bbd3018e7c1e
GIT binary patch
literal 3482
zcmZwJ2~-nj9suwe&Pj*~mmI-35CU?E;TVpDa3_HXsRV{VI0Gc85V-`jiil7txPle5
zsNmVED9BchRs<@dm1-+pbiJ_qR>9TxfGQ~3#T7SThkU*^@8!$<=0D#x^ZmbxBwQfa
zy99zlGQu~h6dUC_5Cj>izXcof*VSr6P%R09HZ`{5Q8^=ChIfwKkIviQlo8LpM61eb
z`{Ce~LD^W;(bILI4NPor_2|3Vl8!M>-h(B(^bzCdHE*VV>l<~n5q+K=zuuW_EWUH=
z%+^`98L6c4-Qk~e+Rg<GwU1vNn+}_r`>ZF9p^9s{IaFi%rMa;BOT!x@tD^GW=52d7
z<L%Nn=Qb4krg+!H`p!o2(ZTbd+`zkLx6UZ?ic9N~ehqxd8#r~p#oOi91FyzA$)=aj
zUhC+zDmuUp&gq?Q?ug9l9mF!Tigp$rEEbtJyGx37zwduPd06&D{M{r8X>I>{c!FwJ
z{FKB)TBm+^y)*Jsb>uP8Ff()V#Iwf2skeh0tkx^eIQZ%6&!M7mx=~YoRGHtnaaOux
z<yR5b6X~P-uwGf;mN$=hR|kv4Li+j(hV~A=4K4j&3^lF{bsfx!*@G=jtQOPQy5H7(
zjj+vKI&+zqhG>G3W8{igR#tueQzmJYGxANp|C?jk@^XvG`4F$&oc%vmDt+9OD_=Mb
zibfKjo%X-n5+S@C*RINb+U&jDMP_+Y;KMFlGk5jo@QGZJspztk4Ov?F#VNQm>R;tX
zPD{skFiE1dy4P&E`82QWgypaF98bi$S$}%6OVvVaz8)9TeYf+t12yY;v+3I7L67dW
zZ_(N7WLO_;u^|U(@)&GCs@G%F5jK-CoKT`-(92y{ImA_`r=Gfcp^ZGt>*BV16{WAW
zZLGRKBY&<+u8-I(-9K^q<?+Y8$6g|yuP?CqDN`j_$^`PTgWh`E+1usLD~rdxYCe6s
z8Y>_B?@-O|(46?r5KBp*_r<0Xb3e2G=Id2mHHJ+=?iZhzwWY1#nw_s%0W*__D*caF
z<l0{UXMRmFYhA;Y7qxe5^HopiCrcjI8(a_!FkM>dOHYngWCyEGm(kd{O0Is^v4@_M
zJ-bdknPy(?bi+9pEOrYmZl`XbTfVhBc4|x&Jw4JhC7#Rt{OQdtHaB#y|HzNcyETz)
zd%r83JQtm*2nndYDIE2b+PC7)3ipdfS;?17R<ua9A0wOJdpsVr_c&YkyD;IX@AFCf
zszarbx1xo=1x*b=?z(g#HqDv&skH6k`%OD;WZ!sq@8@MwvdM<G%w3G|LRokfH>dMy
zV4cS=%7JtGZ*=ztG$oCO%j=#6b^M2f4C$qvG#YkXV`llyTJd&v14b?Nex;l~6St*5
z{kmdwtg^?hRQvX<+y^SytV-)s=O@Zk)OmO8JG2B!gdn^dhPq*^KYpP399Dn$yQint
zpZi~&fmW3I5(7O0(n9ee2tkSe7}(HvPVlRN)*5w~rbi<!6hwHB?4hGFUan1nlGI^S
z9aF|aL4@~8PPM&$_HDX>mP8$<=ut=u1ra`6_QXD&I&j`VD_R|<>QPAx1rg3yC=2+A
zdKT0&-zwy%M=%BNfK57JYob?TdaxETsXwaPbmy;~*zUIz)=}4>cECnN4UXaErp^P!
zb7Tshg=A5Q(-b@#$p)lCRDg1kTtIn99-w?AA5Z~O0B9q!5l|sg2&f1tS`_gJNaT-h
zJq6#2ZS~O%u>>mtv<=$^s1z#&)Bs}*FyI3)b^!E27&{315R4ttI4<yk#ZXaSxBBLN
zSTV$HCf0%u`J**J!OLK*OjDnNH^EpF=)*8}SmUI|E4_~<O{)LYeoAw5*I9{gCX0e|
zd^tSL%(*;nkVbx}pRGnbln2Nk^#>Gy1^^224FTE=<%I%m@=-o;r2rKG5~4ytVQ3g2
z5nZ%gGnUo#)qo=C5r87;k$}W>F(8Sr1Q=N~FB%wG3>pLUhod;4STq)p6qN#sr^f?5
z$mlXa3G@U&iS$H3N%SN@DHbWqH1kWfNChOfkONAyNCT9?$gt7WQ!o^Olnf=HOh)FS
zl*SeJXY_wrp*6qqmUY9n#LAnuT{rAPxMELz7p?zRlX^$^utbS1qw)oRneWiw=<0xh
zp@A8ZS-V8q6_d5^eU?ygC&Y=S8JRQU49Ep>0kjfX3CI<31+)rT1&EEX0l6V=fZP#x
zKpuz(AWy^-5QboYyb!NN5#!Y&f3&;KPu{f~77}-b*ll6*U+%Jc<d1e&3cel2wrkRo
zg71K_9iYo$tQ>SBj5UHj0%J!sZr>khZ8$vK`=_-2POxv&xxlCQSUNbzpA$?>MhE8x
za|IgtvHTn~;<0#u{8|2h0$2fnLi|I3Hba9$fj0RpK5(UgB>*I32?2$%!T^a-(PAp;
z;H%NqfFjTcK#^!9ATcTiB=MI3Ba04>21XXciUIn=Sva6rRxBVXOA06+jR$&=p)x=T
zXab-_G!ak|ngl3?k+PU{I(RB06_A`E2b9K01C+tcu-7~z1ycb?$y5T$WM(d^&r_12
z%RfG@PD-jBbtUoEe^9VabZKVWMlJP=&TOkE(JlpVfXy1<MUQ;;<Zj=n?s;|XVx*Wt
zTsDrh)M|15zDIqUW=CTa@1{5P*Yr2D9+@tzLNGa?d_}f=ew!%BPcKma-GHFj`C{P%
z<85E{%qDzVKA{&ci<8F0LNP=Lg5orXGR<MU=1{6ROjl)OuTwK3Q>?9{tEZ2U3@A&C
z3~5x6u^`mJm@G1}FkNa%=Wu!c0f9lmY_pYqEFUN9zdCEajIrE?X=}H_(b>h-&D{g@
a_C>v{R(bNx%<aQgCuj*j2?>WIq5lJ;TCAS{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_gemm_impl.hpp.95E513E815A63274.idx b/.cache/clangd/index/profile_batched_gemm_gemm_impl.hpp.95E513E815A63274.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c96faf96d1779f8026e86991d5d9865a9a27c50c
GIT binary patch
literal 6562
zcmeI1YgALm7J#$IgFrk$a=@5Ufha^JJj+vF@)E8hh#<7pg7^Xj#VX)iM6J!^Y8y~U
z3o2L=B0<#JSc9c4Xj2tS#kR}G)$60Jk0PkWw%FU&Yfo%XX6|yWyOzKFW7Z;TKK7nD
zduGqsd(R|ua&mIr3p`%p{K>0|^ULN?JRVQT{*{)M6nY>Y?}i<Zw`%7n)oHJP7buYb
zG`(bY)61Je-}7jHak<a%!E?oHf4JiJJ#9YbQFUipQuCwox7<1lrv>=cj84B<JO4^o
z-^s2oXK%l8*bv7+?`4m<U9s<c;`q(cf|7BmCw84m%Ib=&`64m@oOe!=tYWgZsBd6-
zt8l>MV1l_K`QdBp)`ng9<J7H`fkV?)In?=yr5f+vt0#{~Y!A!dHp915(_%ONQpd5g
zLAxi%Y3w_F&M>R4jrjDl#7X|;_r=?M8vd=gesOI2fm>Z^SHGIG#@sC~_{X(=yG57#
ze(uw@DrUAX>3KUMH1+DLkW|l=7rpLFqaO*5&9i$jD*DYpsYVl=-G1q6q3D9~Z0m+7
z+hn&Ow+9*L5{=4V?><~1y!-Rk)hl)??%wJeoZ?*9<NVR7(vGZ|eq!4@Ul(@^6PlO}
zDT(~K7hJP%+k}kS_T%xuH=3E6p!ug)7K|F>ePCEz*^T}8^0UhmJcoTf?yYQ_i}xfm
zUk;Vc{&L(Q=rQHl+RWM{SpOf?d-(Xnjuj~}*CcTlbIk$Wo!?fe$CZYipBO$sMd{{%
ztL~@6;m*s0-A*0+Wo+GzJJh?A=c$|@46eI<r6V}0_=N?B<EL)jAGol5?ykY}B1!d<
zsjb<vd6(Y$Z=}>TqoiFs_I%lvZ@#_h8Pu*_cED(RC+!h^EH&$6twFY>Or@!QyEST^
z!-b5n_?g~Yib@xzb)@O<ChQljh+UmG@0D|16^9aIYk8%si|59%`$atO$9}^#^Z=pZ
zHY~CqM#L8}Jo~@kVrgMY=P&EaXr9YZ9wV`lcs~B(*(2Z$p@4ys{DRT~5v!cn5q~<^
z+X-l<9(mWJAg<?-mxdV{y60}SqZv7jlY<{8pcw_o6cB(}2{I+gDhN?Q1eR+dMhgj;
zbr7e6B+N|^*F>@&rs!cBCm(vcqpx4~KA)l)HF&8ZloQZQB}gkF409Dot4OW|uUf3{
zu)sCEaPP=YL@|}fvl2<Ujug{|{M%3<$AYOHql)|A@U*N6c-7+(1V`M9gKH-B&=yk!
zYVhQGSWFRCf*0m05LS^~4T5S?zXpUgq+A0+4av12s3rC5Kv+l0cYtsQ$@L(pC-oaZ
z*g(p6gK#(I5yv)Yj91=$=uENBb(4b!XSjv^<zQV3igmWT0wh?jfItOUcL~KP!Cgtp
zl@O>T<tlJfk>jf%QU%d?JTthOVI;}EX7DFD+6)sgYmr!sTroExaTCdVka!QtElAu#
z@;)TqN3s!#j7W^fGoqJ_D467IBYK5n@%e*xf9x>@*-=c+l$=!V-xQOZnwy4cR_ZK*
z@}}ewnpz0fLI~E?L8uOIT8e3c&?b`g5Uj`guHSsVjCVUa|GDRwr}Ny4VfR+<kg98+
z2y_y@dg+^DM#F}Mr8Ac1V8@hYlx1-=I`+_!`{U2>pPApit{*wU51Ezqg!!`N{ro|`
z>(L=JQwP#Ih~Z48nH?bALGo^p?#Apl|Bom)RDKL7<^+gNfb|{l7j-X7{>T{53hL3w
zdgR897?7^LEn4Q&`h*Q3F=qaS)iYfXn=6o1j(oTt7LJjlM2;iHTjF}7UE0|ZK7fD^
zV4_vvK5%N$&^ZH)B>*|fLBI(t0mxB7vJxDXB&)ztMY0(j%_OUlNKL-rYUHoRv7Tbo
zC{c~9*~pR`M3tm|BMNCm)+9qQIuxQKS&u^WBpc9Z1F~iUifKioTS+z{M-w9EXF`D{
zoOUV3gi=h1h!<^WR2#BpnkXM$qN~Z4Z$mSU$fXgD<`#-(4Rg_vtVb?-lADoBGs%0A
z%U;YAUW{(6_%44D>%-Wn*mOMXq^L<&O8)rGyNmk^8f<B%z_-Af1!!i0?*f7r`Yt4>
z(6^AFBHto{<VYw-Bk;o%NT|S^7QXFuk>A!zcDsZ{hFN_y-9G7);(Wo|tf`m5{xUe>
z{u_~SBkn)HN+n<8y`k%gom^N_x+4A8J#lPJp#@3{4tN9|P&$&EfNCOH50oDBqRc%1
zukYvGV=Y`FSYn+>v8G5IdCC1I``~IMsYau?@mD^IKe=IrA%@j&0BHlna)M3!Qj4>9
zeRYDh=q7~Ugk(-YGq)h(7EHqY14R5laxaAUV*Sn9iZyA?pPyjswoI=~qH3G&HJzX=
zuPlPHy|M|)k>ywuAkE~;atWFxn?+EbEYC_?H%8R$T{c)`S&G0V)+()LghCX*-eo<z
z6qiA88K?T12eWfGIgLEa>TgD&n^8PB!}^4)(e4>j9G<7Rr%ZA0EuKQ9WXQjJT;Tdw
zl0mKwO=v@L+zL0|8}5Dmw&!p5G}8>@n#tPU05Stv%eR26h2))R;!ZS~8^{ubQX5eQ
zW*th`A#2*C89hqZlWaiK4d_*@--=#oMb@ONQ@&bzWBI2e*<@A?Ue&nrrkNU$*5I;;
zW;7txU_Rh|GyAoH4Bw%acs^Q9##=RtP@^dBi57XL8e3$Ew;m?^f?7DrgsfTAayD>m
z!}0oH&V``Htxk-C<yd4d$C1EtEV5UStOR={$tti{k!%KgGs$WsRHNaXp_Zi$XvCGM
zCB6sfNY<kOJ;?^-X227n*~H`4O0p3NjpX<yWN$(a-19!{Tl3+Ff8K9-W);p6$KD?K
z(jp!k5B~Qg`jMmEnd7fsPkSB@o-!VMeDds!eNjK%cxEpBG$eDrO;TeY>(Y>bkZ^9R
zeNke_`2G9h(&r9+N{4m}CZ7M@sXy}h!ae1}o%UR-+%#dj)8nWeCWlCFD_WxVcm=+X
zpxOJ3@k)|akf6fV70sAoqM59|Dp6P^BF57qi590HnrT9kCX)9c$sUqhkfep=eMqv8
zWFyKoqB+>!Hk90kQaJ939<wepeWd1@z2fKG+im{kBk6u3YI7C>|MATucVGGK_@2ZE
ztzvj)24BB6esPX@NW&9nf<C)eHHv)Sa@YtB2(@~&uV{aEgmc2@>|tZ6^U|^0v}mTx
zxy(v^+xHLd^66<9N;9je)qJe9mRd_tIaN+j1yw=NMxZtV!JC2FjQRe~BOL+nkMeqE
z%l6yhcaHi;ePLu*=P1M+h1a+_KI}X_rY?Hb&i}hUCHafzST3eZOXijCurz4P*Iyq2
z*ZAv)kiY5gdu$8YVg(1>bi_^eCll`D%r)5i@HJ=FxK-H(e?~jo+S~DI2S+D?a2St|
LHJosB7;gRvjJ+^!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_impl.hpp.65F140203E93B240.idx b/.cache/clangd/index/profile_batched_gemm_impl.hpp.65F140203E93B240.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c68960bb70195d33b61a87470fad555aefd3cae4
GIT binary patch
literal 5272
zcmd^@X;c%*7J&OU4U!m=04*>Gh!{}WL}ZaoSrkP<lOP9BKygG|U{qk9fXXBq8J$7o
z;e!#C#r>$@##0By4RJ?s88^nD1LAW4HHm^7=<7Jssd~qA-Z}oZPm)t#-COro-LATQ
z)6%fe&;ptuLgs{JrY2=1QUpO5GvD-#wE5vALCob6MCO)jQ4x!_y9wKOFMc|y>R<Z@
z7Upc-D-C)5Tk5iVCuiiBP);8LFYE}8`uN`0;wR4igiRUI!QoE=nrcTjDIPs}`n}a=
z$9`dyV(IJpf`?wq2CklPKr*2ryMD)zLYtA@)B8k4juA%X%hvh7sax1~_LqbWgRH2D
zFSp!nD7w45DQ;QSyM{NF+n*<18y3^rbI+u0W0>hK?t!3c{q6N`_VE$ZBE#KR>2D0*
z$uFx<-}z>jV&BeHyZG+mM1s8djCD@J(#(v@<4gT(9v_}!6PFe>bkyv5Jy+)y*Oi_-
za7<m2{y1geRIBGbt;g2b2^Q=fYV#XC<wWjU$DLVzobPfkCer+=D}S{2J9y!T(`ho1
z-i}cZMm*4WIC19Aqu<-C@8s9HL{8Y6oV+13^mzArSN|H-qm_&95hX?i^@C5YY-w&e
z+?cmv)~Lw9wyePDC2<eas?73M9sSecg5lP1PGD@1RC>_!%I*I1!bN{Xo#l=zx)c@k
zq*iG*yHMk<KQ6lGU0M5E8_RDBR^N<iujou(SDdx>X7f6SM{noOC}_1eR4<x#Xjo0+
z=svPLx2^l_m>o919;;aQX7zyUE^qVPlhY$I+NWL*wlu2<YQE^Vacy&nZ}!ZxAKb3?
zP#Z7T37rzgmVI0xTe0P|z+kvhTf?+V9XAFTR0%3?)h3dPo44gh+@m%=Ik=*<&@<w@
z@PJ=;os6|kSRT5o&$YI*R`1&S4>Jdw`n<Tj=YehMR{Nd1FXUu!-FCMb&fm#%{CU4q
zpV_sp)z1IP6MSesm6W0Cm+;~4?DVYE*@4X7viMfjwdWD}nk3Ej8DuWnf@@(yaQ=d;
z;|xl+9`!Hc5oQJiEz%d+eEK0cBoKNO(2$liH{C+JNoSr8Y8*9O06e-2hIfHGJ5q0C
z|2^dmv-1piw0M}<h3)3il3@~zq#jZqB+K;9^hcECorO{M`0Pon%%JgfXP4taF3F=~
zL@}OhH;;}L#p=i?+0uLdKC3e%MYq6^7O-d4C|XXs%1Jktg*7&*qlzXrQH(WV%n-Jl
zu_lbgC{7rMQM@o-N9J|1;_8--eL{*BTZ;#=1{iV`>%363#9D&MB(4%nCbgDgGO4dr
zw{nVBfTaSgkW~dZE5H?Je+7&|tb#}tL?Kp#R1MP*zXXeyV2QW~EP8P61+!kTK=NJ~
z*$bm^j_QSImM!icGVI-b%w3P7GhH&VX=S-&VU+EXts~1@=a&#2{z+eDyasx`Si0GM
zeaUAh#+JK&@oEc`VrIXNjU5}u#+RZ)Bq6%HXE*lWHJZ`axnB@%(c_wnT#tllpE-0j
zYgiI`ygCRd=ImN^fi>%L#MJE_W=otZKO5n8#mh8-q&*oyA$S+UM0SL)jrz($gG9CV
zj`GDqwwu}Wd<jNUz7!(`@D+fu63mqNzEgs$65Nn2C4?wpJmMBGYr*wZ;GqI9Bv*rn
z8fOi7XmIWZyKWF6eHqD@k=T4>q??S?B{Yhbk&|R(IMSDsgXN?y3HUh?A(qGbahWWl
z0_J1KD)MMGFdvLHKx=S*1@tS#V+Q(f`=K>yArs`}Vad8Y5L&zXc}myZazh^N>FlYS
z7LWFF_QJ^9*&CzL&Z9B%arVLJ78u_Gb2Q6aAi4#Ph;M`OZ6uHIE?i{hl2^dw^L-%R
z2Q!djC5S6=t^#ot&UGNJ!}%PD&*5AT;(El9obe}9lZ5$9ZqEfyE(nmLMlfzfGm1av
zlM?438p6aWAKdd%2AuQHwYk~DWtTtO*YnfT|1tH=?m#^rok?XHA;Vcz7Dm}rwvHA`
zXS&|h%zVK(DFDj?7|aezEnZ;Z`@78(#z`yKwj#mGs_@jP(i<0<K$U@{3}&$-cyu{P
z%5km$Nd?Y_L2?-9YLHapd;}y%5a-E#Htk!|vw)e}CWzhyx{Og|{-aq?4xPhHEf<8j
z=+RO1dRF{Ofz@qh#0~Jg0YU84%7P#I+ejuGeU%BhqRUZy&ir*96O{(|rU7i(sBBG{
zOihyHwtgxgTwCx|Bqztn$v{?d+lv9iA9oCS#~jV&cFTvd-8_1Q-3p9y+;Viq36Ji8
zF&z+sbX71x1-k5_O&XzUn2h8ah}NKc%%i&@tQ)4V@*QhpO4`3aWyPGDI8hwtCSDY;
zqXWaAM9ol3oDC>i33f_6_bI_g34W|Ktz4?aYX%uPNJi>5zcv^8$;betFDLnOlsgZ_
z-E-fTCot*H6n`AC?l3XMA4h_blq1DR0UQNjtOR2v3}9Vov+4*HDnqmZ8llEn10ysz
zcY}2|I+i@T7mRyxdoq$EBX$0cbY>qh|4Fm=3m&XB{6Cv=7cU@~jq|^??{U82^$RhN
zBfiSxRD0=mY?a^3_AdzJDG%u!{qfs6nLq~k1_ZOW{Pd)?cF(RY)*j4&@d3KxpA`f5
z(|;27pUl;?J(Blrsp(&3&>Mu~?!Fgv`F@p)Zyd60&B`&60&TYO68p0IgUL2t5{#t5
z(oiH*z&Hh}UwO=5M6VXWv{f)e1v8Pp8m6lu0r5+)eTfn%kM0569-Mn2z84aazMKq|
zle#_J<nORtI@PN93l7O0n-lu?oopLE%Qk<EZVGu_X#rng>+z$}DRK3B<)2kZ?QZ(7
z`U%W`W}M`MQ9dfRIv4DZ@-YoQ&pcnonvTUDY9Xc}I_li8zo*oxshlZ+Cr}f(Xz^iG
z7)Fz+$rwdY5juMIKJHD#=3H;>iSHj^!Wz(QsX0C3;9$`gtnl5@U(UMvU9Dy;$stIN
z?qpv?`O}O$(=RX!xeOkc!JBoWt%+Q3z(|(+&NMn~^3UAze^>lbzq9gwTHrG#kogdp
zkDlLYNhyiiS1IXf-=<e<8z$PHzb;d@K@iNFJ@)FyHVA@`ngY}mqNX8g(hVLRubR)>
s74rx5;~H}M8}S7~V-r&zF#zfGN6o+A?`S)WP_Ony&fgc0Z4ktN0cQQBlmGw#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_reduce.cpp.EEA3A1DAF90E2787.idx b/.cache/clangd/index/profile_batched_gemm_reduce.cpp.EEA3A1DAF90E2787.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0e0d9391a389a418677006400b314d189931c35a
GIT binary patch
literal 2518
zcmZwH30PA{9suz7Vj>R=6c7>-kSl~EArJxy$s?SGONAud0xckyKtvKQk$MsC6|f6b
zTcD_`h#Vr~@ll{Qh>G+RR$L*XB2~63Abgd2v9%j;V`jHI-}l~|-~8vzd^7JqPw4OC
zb6XvPyc7KMM2T5(SO|i26qh7Ryi>Xcg31)Jy!!S~QN|=w``&l*=AWtPP*=_YcT!8=
z#x&iNX-BFsyatj?ujNRM@xu7Ram(?t;O-dLtJQgLo;B{BVVeb%rbnhyG;{;+jh!th
zU=h^&et5_D_F+uS{`$o6S3@&Lex7IS>34k}+wba*3Zp+avS|#azI5{)e7G$C?Q8el
zZ+{)9M(ykQE;)R7Fbl)0`!a%hVdh5b$gAa`%kqENIZa{Q9z0mOYThmy*g4NL>)qbm
z{;+dcs*^uk8+1yveU{h3jylh18|w}3@vA-3LwT$J`rPHawz)D<P;5$rU2Rfs(-rri
zXPzmCt=uvuu=NvqXGwPAdv4{oBZ*UKIg@-!FmIafvh>IC?hc*5`x)rpyg<uj)`|UY
zVP(1Ly<Mlnhglhq`c7?--+t!jZAX-0E!A#ryiyRKZYK8Jck)nM%VS|yev{3WY$^|m
z?z}52gHZPam%R0n=PYKysVBbIHiQgxF8-M981c&+?JjTTVNTspNz!hEbKZG638zG~
ztPqpTu+tWCZF1(NS9c55Hd6-fA|-*wHGgzo=q3b~dHFZ#p4c5qTz2YY-rAcb4H)cc
zn!s9iADY&wd{C%4S*O<LdFxClzO?a-$Z*#CtRtcJWp=xD`G9<{Ev`6)@A7)?UaSyz
zNQex~9~@;Dx~@wdWl0H9++#~;&b1H3j?Un4TQW80x)+yEG)=w8Ne;Wd3onu~=-<1|
zasSC1s>w|ps^r+)K3db@+mv~`$mV_dLhHM^imTgN=>dJ>fq34m_JV+Vr#igNfHbl{
zH=UF;eAall8gp7czZkRVf^D|7y3Y}2tsR<sRI*PWm7lN9Uc#2=aZTQr>dFpVe_dhF
zc;0`2T>NG4zM3>j_<`4%!L+9uYbrXsQ$vncuJ2O6czmt<%~8fAyQxS#*8PgsQ8!$n
zH-0_j`O|o(O>u&glCg>NQ@Z(EKfi4G!%M@3r~6OmrS=lLqMPcY6ZBcziq+>yRSA=1
z8?tG37;|a%S%FiaB=OGni$XH%a{QoVLUtpSy}PpdU6qS3vutr;Xa0+CxYqqo9D`FH
zo$;YH=l^~z|Mx1l=%lEZ<94ng8sCbRsZ}v)FS@_d&3%|Ua*&*yQm6U7e3Iz@L7a2=
z3yrPuuN}U!a&V259`o^Q=0_%6JHb9Ei@o9Uuhuy8zY+)UuiG?fXJ|F`Jo;CA+=sZo
zN%n~1y%cA1#|{tKPuT<C;Gswe5@=zyOjd7-$ZC0vF@%dvD`lEcgd*GsLyfTEe;&}v
zu~rmuVv<Ay2SFxU2iaxFPt!17a1>sD6gE{3ZPp`BnLFTG3}Y1mt)sAxGC+D;5z>(n
z!^YwTo&-;(vKcQ3CxinM5k!CrVL~DN$sBE2!^LB{AJ*XoriiJQY6vsL43Ifu4#)zr
z07ODafGiP9KxBjr$O^FnWQ|w@vO#PB*&?=|6yb4y!$?klk7D^B_fNT}G4<E_rsvm}
z|0(cAOUtEJAv;+UFQEC*0#u8#-)!%vk^|YnRwX)_4#<)02#7&u0CFZe8>-%85}ANl
zL>3^nG22YllVi*Qgc_rOxTai^vZwNi(fF@a!Yh_fYlQuk%TLmk%70TEwZg3Ii%}zN
zp$x3pvJs{#Ev>G(^x2PN4T{A2xPc$@4uVP+c}DD!zih-S@(1<wRL+YRc%fc^yispJ
z;b=IZ2s8pvBpL}wgo*&|LU#emKr;Xp!e}83SPG+3&_ys>r1FJ24^!<scl!SH-LLr6
z=5=`dGNFkT(6-Qml<UC??EUSxspLR$uvdvrp#yTHI09l&7=WCOoIjmOtbl361jI68
z0b-l8$*LJR<{Ut%ISPnt$+c1T>~|6-;CLry5xjuM=S3)+@d6*d4<KK@FQ6!X6rdP>
z3?MOI3`oM404js|W$-6Y7R`+JKf!e=QiOs~@XD!`83begCv#?zwkiy#vs&@ziRMQ~
zDf{CEnfy#ZB{07P23!jBOI3bTb3<(pX<BdfJA#=lPE3zm$x(@Tha`BVB3!wx7Fh83
zI?J&t#e}{#G%O^TFN8vrAqe8Ds$r^XsH(a;*hlUVC8Y0ERJSGvuEFBgsv+v@@R~Z>
zpJ@f?`fRq=)d*O>K~LXcBM0TWxqEneIS}kzNiMdg|F4y3WW34D+``hzhC-#$9U0Cn
Zwh7bF-ib#blKr-XV|;_W{o!rUzX4CJ<wO7g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_reduce_impl.hpp.5DA2586197AB1178.idx b/.cache/clangd/index/profile_batched_gemm_reduce_impl.hpp.5DA2586197AB1178.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2b1fc5ba39c3cbe963c366d2b0e2b99a35d08cf
GIT binary patch
literal 9058
zcmeI2c~lg~*1)GO12RJcgY3gF4A5-K;({!K1VNEaP!J8MxCIxOK@=4eHLfW7FeYk5
zP!qz#FV8$7Mq?6{=K{(mI}ErShBz#4EJ|Y3rzT$wUwW$k`p)qj|8x%a>0jUKs#~{~
zd+WAw5n*9TjyzuIM-dwpDeIQ-c|4v8{+GFK?WzLc@dkxF-p1qqNmvqlcbw>1Y|SX|
z?Q=ZRZq{1Zvq^d(%+bor^V2sgPWx?G9*|<LzS8)EsXETtxml)MGt}9udfDqa_0=Zh
zifJ|l?AY-AbE;OHS?abRc=l_4(#Z_R)i>?JPE50t8aMs<*K}!N_vDhqJXttuerP?u
zC-1y;RI{q$UzyRqlJ17_cQYg#jAGB<dm|BTaY;+G9u*4(ydYt?_{H^oPn?#NPn)AQ
zcl~Lx+xTc<R>i#gOLkZ<*jB@?fAA!JLf*lStc(l4H#LRsHs6*x3x27Z;aPFsx3b}v
zr<LDKEG>I_+d3<3yzpCT0l%p&bO!SN?ep)qL`E0Xf8><AWmeVQJH|zd+dA?;;3u~Y
zzph+o6rXOt|Dww)Fa3a=uU1XYfk`%wP0dw{UJX>Pp7&|m@5247R@Rj@pUT|!XY(%)
zHv3Jz-)(ZSSjJ3D>*>r>+w3V1`T1me;oVyK5}%!ye_b2nck$Ii6WyFy`%ZP2|5hM%
zO-_nCy&&+<Q2&_qb3H61PnP;UY~R?^KEbPh^u|Q}x?MigiVL$2CG=csNb1?rmb9Yt
zyYquahn+Jo&2_CmbbVo@lV$GC)M%ZMsO%I$?fC@<UxXyD|2fd-OhexFHOFf}9c>!v
zJ^#kZvQgQ4Wt&U#KZ{R19Cyfn^|49I9eT~P{ptsXmd!m95dM3aYGdNuAG|Ky4oT_S
z{^Qd%zCBxpx7+==q3U^bsO4G5zS?o=E<qcHl^@&aol=Edy}iKXywTpI@M~ap%co-g
zo|u7e0<!#krzl&bs^MaPJ!MD!*BgsJj7av+k1Fb~Di7{DUY~0=EByRm{@6ozybnBk
zb5y=FCpqA!-XAS*Ry>=(@=n$>(?-*#*B?L4yteOPK>Hu&`;-5Xu49%I5!Sq@^Niuq
z%#RgIA0r-5|H2oG3l05_Fg!y`JaO@lb<=fA;t2rpTul*+OupYMUSImC!T|Ivd91ab
zwP57J6O0tb1<aCrUj!U~c9!88b2T{XI#M-QEql^`>Q{IU;9@rN5|5oEoHY9H8hGrB
zf6OUez0jZGIh*s?Al)FRkqb}2*WsCr#tqD>{W8Qo{+Npa2(boIJ*oA`g=ny}zHO!1
zvvd0lzyKEz>xnH!E~LO32oeuQBqtkyE-oO`li7}3NP+$3uL@7!3y=XgS^TY{h0sE)
z;-RX4s`V!-pBsPxzi+E&Yc+Bq?-v|?IrMDG2h9x6h_ivat~+G|W%9wjx9=W}Gyr{E
zgN>ez(a43=plshUWt8uuDF(pc0unt5GjbsX-iYR&?Oc7m&j3c@0#<rff{_a;@Mm<Y
z>*+<an;4#m)4^HSnbLu`Hf4F{U$(e^i>^#<nCND#!?2fO>}BxQ4gq@w#$KWMDvZ5K
zSZBP!sjusn<_Z}$)iu?Fyq4ygMv>fG?oVWj2?}3A=@Zf^x`uqOp+F)nN517WSD@(?
zD3Zu;AfFp3fN&M^sUob`aOH9Tm@`6P*bVH4Nu*d7o24ag!h>LBqH`n9u#b@F5lZ1+
zGF&0{UG;CtRp7Iw;8+SyoPf`EgJ(C4<5+Y-s+jm?cq1P-D%#-4J!aS}Q5Hp;M4Kqu
zEZVFkZh;};$im_j;+7TGQ?s17#|(Q1iS7_REz7qZIj1`MW;CBowN16-n#bgss@0Lt
zrrD-ZvNX>$N+!3JQ!@EvxpqMKtU{uYk~$TBiphl1CFztb-7lSzRYQC=B#^Q-5LW|p
z2;T&&n;<6K09Fk&H-S|X&8=Y7O0yEIlr-N0t6PMvI)2c9{q4^abouNx<Z%sol7`BW
zM>)+E$fJVh>&Wjq(#~E!TZ#NC35#2Atl>QhOnGl@%Gx$QLs`e14&=Y}c4B9=-G8rs
z-95yxFOlg>B;$rjCOnXlZ8iD`UVI6O2@AN#d^T}*qIMG5UtGJge_P;g#K&(Ny+kHl
zF^3O+yfEz7Bz)(i&-PrX5_%`kdFzgHn&Bm`@o$_%^osrGBa1a-Prr42ej9Vr4{sY&
zgf}`#uq2A}9Soa|!m`n9uDQw5r?qK%OP*Tb-M}bS#68Blfl(Soaw9oK3L^zY=|<@k
zm4HPFSdn^4z_WzxgnYIHLQ7yek(Yu+DP6uAysN>7$ZNp6hUQxEuBEvi9O{83<r~4f
zk(RfBLklf$1MfCk-T@9Bv|I%iDq0^k7^#8kpBl!gVFGDS4UuYyCfp77-JspL1IF?~
z#XpY{>N9LLm{o%VSCC<Az^sPmS}?1nxem<gXl?|vMw*+!teNIEFl(c^9n9JZ2iXJ`
zZB?hF;dPXUT=R$n4Jq8}@HFqMcr3^mlM%$}H|?~6+H7d+Jp8y2YzslV?!wBl2j%_C
zzs5a^c8}H?f?;R6&!i~EJ%*xK_gIQD1R2I$I}BS4CdFV$vs*EE5k3hfCy6{V|H^@e
zo&AP*J&0vuCoZgGSX-H`6BmOqtW+j-;$p(6O?H>J*H2x8yOIy)`QXacJ=6E#Y739t
z&vhBrCfFv5i@z9lYuHxp0%6!~C~O;9FbumLg>9#K7Yf@&m}?cUV@(*!`Q&)P^#3SQ
zM4#d1_!<dbBgz0pU{XXnI45Sb%Y!chb+J;cfHme^Jj1#OT#U8x8#_iY#+VCJ7}i7J
zVXTcW*(M}tA`NgR0^vx+%PGXL_Wt$}T-2MgZf>r*=i#l`M!Udy7ihiPa;ubr<<p<<
z!Ga?2ECOv{wQ6nVhS(S124g+1A^tUF$XS(P%Mrhv<_g5Gp!quDU#Gbe@hb_Z#Vzt|
z8D2DmTZ|XQYbS>M&{CIa_ad|Kz@I~sb4WXi40{Ppyo5d=?U$j6WrP(+QmrQ6mG)sD
za28}|!JaF~^<7Um66q7eup-ulvb)H|g|fS7j0<IVk%tRycfTeS%T=HGT}#=)oW)yk
zKb|9#=g6EBY^=D~^9jGX9!G=wkYFD&;sgx)1rmHg^FbsyNOKJm)DYfL60Jx$(OiXx
zpa%KZpm0)AE%L9Wc^Ays1#^i!8)jtFoC7m*2xs5<>G-$XN>5?c?H%kVafaISWV*z^
zF;|Qgx0YEub3rM?wjlgBhSc7IB3g(c7`7GRKdn4>yMMvI*Zh&H8Q_@sc<!~_;v=1_
zH$S|K$EgX;ZbC_%VE==&YvQLqKJeBf&us^IDQeDR+yH-RxyNGar)HnxhF&AlYorZO
zzkI%U;SOVo5-U~+lL}!frvW#_IL_7_*|>@rb{L6<Nu2#vNLQeAc9h}!$d8Kr3j4>*
zs<3=sjyJyJV0RpxIioV{39vgs_#1f?KXUeNwT|XAHLjHQu__2sK`{5C=5!U@4IvyC
z4Vk<3JaYW6A;Z=ohdSiK2^h8>In>j95rQs4D3M=+&`Urqn=%M3qq!PpR1@c}@!4TD
z5Jldvg?Y80jXiOD5K#{?MBWJV8fke8M6}TIHkj8&%R3;VgUE|>KFsf3e_4V}w~5`P
zJ)JV_W_Gid{`DmMo8+CR?%+wA;gF%-6~A}uOPK$3w!5At;&LdVBb*ZOFCqK5Mjlr}
zpS#pxrzX23pH)ME8m4fa(pcZ9n?yusT6JVA9_l>O>=jY=vD{wq!-88AigJxdi*Oo1
zkV-j+G(j3ga)F#8g+M`3x*(mR5-=$N?H;NLr@X3RBBzyRXuN7@t_81Jn(M)~p7e@g
z8^NoQmbZXy3oUO0uQpoV0k$2qTm>d7S|2qC)S%V({Nqo~TkaeF?wy}6E*L#*+?|oi
zd+YPSMI-&D$`5;HzTcdv&8e(fnOG9=Tkks`rpw-OAL&jndMv;`E;OFnF_b7kiKrc@
z9R;-0+<^i*XzoP8owOtGLcv`$_o3iEn(w3F`!o-s;32|SBKJBB)Mu0%Ykamx3!;MZ
z78KWl=8^T(iac6rxe|pb5!I1)6xL322MX(;xf4ZoqFJOq6^c<It+Qv?E)>;8%X?8w
zFD>swQGK-hK8m_e^8kt&pvw=Um_eF{P}C49e|68|lPM3_d+%pIHT!R_JvIAPw<7bs
zQ7EI-iKcbDGVY)J-YAq&8XPy}><EfC)_8NDRBaBIVbgrlD3XWB!^nzP1S_TyN)Jw_
zs1*E4L7PL?c=P$yu!xkdfutHpCVUfQH%Sb`uniz<pt%WTO*FTHtd(XZ$dokS0@*E^
zRS>2E>LWrogm;s`uQAYhb6l+C$UAov&41ZWUzsg7#}VFo!}X&{e?Kw&L`z&=f-fBT
z9z;$bw$8Sm+!+U_Je;MTT=D_aIH?~;T<#BN05l(ss0zl}`fwGrOE2l(9xJPR>%gLG
z(gOANbG?B3umgGSK;ua-b|TN6gzxz^hCV-Sh2FUpdr!Rnz%|nFe{cwz2TpmQ{dn|P
z|MUTveCBg}U{#@D744pSQE)HK11NZa=0OxZNceGTX@b9L$PL`b1k(iSfRSjLsHMk;
zO8+=9wo#KATgYE1BxXqFCsVYTznG$>{G}8v<1eEqg`Yyv3jPX;R`OSBY2ej@7c-CT
z_8sx+CM`&tebsn*pBAKz-8Cn1lU7=;MCM8)A*!|`^LCm$ka-8qok-k?v>6nQS;Z=v
zyO6kx=3XT3rMVA@`)Iz8#P?|)K;i+K2a$M?<{=~=B0PNj!WFMG_Ut>KQUBdL{Mn%A
zomgD)F3@DeI1BOd40FVjjBq}f;;bU!La@OJQ^G~yjuU4bb6;q6gd<;O$x;+Mi`}?P
z9Zqe!i@i8{GxO^Q>)Mkl@!ZP?pL~*<(_|?;3qiZy-dt^PE(qLs{QsS#TqX;rGBjU)
zi2`aQQj=5${thf(qWQ|5xprCRG0lyG=JJ0({GVC4pP0Mh@vO)#+hFgt$XarHC1WKK
mTiZzOWcCh@wmc~*AtSeT<W@p%ZOE-PxwR*^ns3y9|Na+Fv%WI`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_softmax_gemm_impl.hpp.03AF7B0A8D358951.idx b/.cache/clangd/index/profile_batched_gemm_softmax_gemm_impl.hpp.03AF7B0A8D358951.idx
new file mode 100755
index 0000000000000000000000000000000000000000..103a6d4bc5fe8137efcf14901291f7858940b8db
GIT binary patch
literal 7122
zcmeI1YgALm7JxJ3K}bA6!r@f}8Zc=qARwYZ01F}%MMOxYs91y&s3=vy2U_X_MAV`o
z)W;P?Fn|=33M_dDSe}N5r9>q@F1Ehvt3pd_*Hx*u-gCV7Waci{x@-B%Kc2OcHDC6g
zIs5FH*?Z4P)SQrzvlBU-pu{<usj@Zk1c$>B(4UMo>B}P^hw}rM!^u3X*>`=_A$NYy
zm^AZ#dGy=|1%D4M_^7JyTC=C=@Dar)t{uB-6{#1FIz4}CJJC#fVR_UC?>g*xdbnwH
zY`~%8m%cFb;p3b=8AhXR@w-d%8|}8e>-JO7GLMvC*VCc8U9t9#gV8t6O2_s%N2YvP
z{e`jo`V*f0^N~A`e>;%rKuk!E`Q~25-;2zJB_d6hX4&WKQyzJzPl!^i47eZ`$4EPV
zX^AwKo(MWyH9!5_!B9K<(YW}8JLhtPitZMu1Jjr7i5nuP-HTmx@4V0DOB+@X>|T5K
z>d6}oS04sumS3Hjzo-3YL3Cr|zUnWPCPg!De;<9V*nY(=Vb;Y~NzRJcUoxj;3?~ox
zN&dLe|HDD<CDEZ>7yLFEPwLt9=^W`PuVq&pzlyX9w)6bgj_<F9-@6$-zvP3S4PSlH
zM_h793f?}&C+>(r%a4YlkBLOPo=4H4wdG`a+WHwGMlwgky^{|o1bcnW*9xirsVeh@
zJ9eDr^W0(v17`cg%XmA~->*Glp?N>Fe5<nq72lm&xhpz&XVjGiuzrqZi(OEWTBxPO
z`z1YX>OX*3H#JHAY01WRq2Y;%hy0`ldE(K?q%{t;>SsI~Q9Re|)X!;^?<Ay1w*4h|
zHoEu$bx!(LwT3#ouRG-O{Hg<GPx4achQD>u!b6!@yk)Qdu&(WtZMuW%NWI6=HU7hL
zsYlnXmcB<PmVXwKSM|8pGj8PD0ehY4s?9c3k(Ih)3U7Df_ULV7{n)AnrFrMA=Qwye
zT@1|5vWe)=o!4KSePgt)Napdxm?$dV%*oi0nh-#*FY8Z!8>zYhA46zqKqGxoRy->+
z4!zu%%cf0Jo>QRiAUPJ}Ih3P;qw|X|4i^H>I08^0U6z!w$ePxSYYQBhHrW)AR5P61
z3_aMMoP>?CAnNmTw4e=4Yy)qzf^kzPRF(86<QtKc9K0t7Q&<5>6#%CK;D%TMoD?`0
z0?$I=gXDF<zYfeoTn_^3K``P55YT{gBM55*3t0L1fwtihry3qXQl&st3Z}CHk~#{+
zN5KrlN+4F^d>n|5BYmTmD5>HA{E>dt`@riyc%SWPoNVK@TxHu12}%tosiBw^5R?wO
z>YzKz{P}GTsUta)=n<-cs2UB;Z?aFFwl7pmGB)x{!6dc^W24|G5Fu6qffDBmz^}md
zj{(6kTwVzTl{g;<{NuR33J6rVTnz+joT~x98rQD@f*M?23k0=@tva{tKY8sRe^CTI
z>vnRq92j<$1HCH<dKK6g03rKe47(Qq{kkG31+Z7(as_Z#;POIXSBQ^a2qcBT4~;hj
zY=(d>&Mrg16=%O8FcWbbv~Ghoh&!Nl2hLs4x(nxSXx)u-AGGenSqm+-&>D@Wg;TX~
z8qN!~@FSM3FV`A9t8DZzBB;c$#Chz`1SOj%3rCbRFA1aMuw+cL4os^9o=CSIOs_|)
znV=fL^ah+8!L&xC&-Men7a@QR95#cRG|mG**_#QaUTs1r&KO;PNGw12Qp6PU<g0V7
z&GysLB6Cq@EbDiI%8JbTkR`|W>zW>VpX0r@<siD<)Wh@q_`{dXcUn8bd&INp7)Men
zAXWi?)?|`W1F;(CS|F}P?6UMpJjYVVIHU)-^?)$8qpM|qT1fNBKtRyB#kLvRu>!Y<
zvImxF&gz$JaF2zyCOq#_T5umY+(*0G)Bg<b%7(?}kffAQtc1?&Pz?JkVGzq+WiySE
zBFE(P2oi}zAIT;5zn@((KK^$zCR~}w0iV@p!jnk>&I(|nz_}2Z6yiJtOonh)K}!`L
z4ph)pg)##{sbG)_>XRCiNi5a4ek=5Bh5B4YQ0>sO9p_Hy*@?3TI%=RkZxB>3bnL}>
z5Sk1^?D+<v`yfi|1T_ff4ni!t=%9lR>hqGXGbhNV(U50AQmxRu6*{txB5A|S+i~uM
z=AAg7f#zp$?t$h#h-XgpQ*OQ`TS3oYkZ(`~8g_y20zHMC&i!@eNK%O*Nqy$>nLcNb
z)GC)%7_D|$jZvCQ8b;|Z=@==YKnbnT&B~yl3~{(u;VR21`33Z1@sarG=V-qFuZ0Qy
z{9Jk?cY%pr;62p;CJ4O=^np4`S2pfJ%Z8XYG4vH<sFWN0ZK{mFhh7Y0fI9|Ek%x<+
zpcu`N)FQ(B<nu4D(=!w<4A&PKiHe=uH%{Kt_flk8o}RHT;z?xyT@%&;q7E3Nf$ITL
zk8=Yc8gOm|L?hxAQE{#p9>$H*H(tkIr(Z#-`%|nXSMBf8H?D!g8tA}|zy5h(@0N8{
z{<MA#5Z9m@GW(l(sSAtG_s|w)L$7QY!uBAk9O#n+LlJL-KHG54g<iQxe_Pq+jp1Kw
zdg!`)v1suW)NP4q2}ZG^Sd5m6mSU9Xl&G&$NJ{1;!zjrq38Q4EWIg3?^(j)NJxXC*
z;l55^W$&}fvkcs1zKM3l7~qegifrHGWlOV7Z7<ULrEq#F3}j6`GV8jZePo#No9*(7
z+vVuWFk)`x_FFH`A+BmXyXfFd9SmSStQ@tNeEY$qKe;5;1m0`H&t}cQsTn_uwE(9U
zoKL~or{ElRASU|HYlV@B+hIgI)R!?N)d?dyan?Yo2F4)$UieWj)E7bJJ7NyrS=(w$
z7Z(*kRDljiBy|jkkD=<7q$+{95^?q9y9<~69_cckNu-V{JkhD3j|%#-cVy&IDr6Cp
z=-wVg8<b(+L8z~4m?NBt4kgOk*ef2&d{Z-H=2**>qe#FUYq<qDD*#u4b0Ods;yeVn
zLpZCTKm{#WLmAh)wW0$9lUUu_aqfg}oj7Ztod!J!NoRb!UYxa1pvA`@gxo=B%-*kl
zcw@cQu7}58TgBOA^0RjP>J$zg5B~QRL$is|x$c<T;cw!>E5?J<y%!gK<NL>**EXfC
zo>7Sg!OCHJO1<2?yx67oU3yjI)1Oi^-c0B#PG~=W_T{IhBa%0cdqu|$O!|CB_Drei
zi>N(A4kYYSWTLir0eWpj(yyJo6*w1yS%m<r#)iP`A%In5YB)m;vGLlVunnajlInoM
z4xGE7unXsIDD1|$4+{Hm*21M)7?14L!4Mss%d*z*ovf$`+x@R?i=TgP7yN0ijRn2M
z_X6%-V8U8<$-c$lpU%Smm%{l?Z*9<HwKaHcCf~l{y)yR9xRRGn1;a*aIh52e=R9vW
zZ~dGOr<_>mL(S6Ar;to4^Ii4<Nh*uV($nx)Cmt0$YndbIRzPe8JfxNbh#Z`A0g;Py
z9w72?-U*1EIOhW*ALreG*p2w%;ifjX8VAv9bLXE82RdARzdK2L?K1Gc43@A@{#X0J
zJ4Jq(hySnNIO(#L@yuTl8R?&7R5A@?=IcKf^AFj7FEh5mVgI{eelSG8nxLjBYLcjF
qj+#dLCZweNXBDZQVO(Q#L#`3e*u<1CFq=SfczP8MA2lbS=Dz`<n_Dsf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batched_gemm_softmax_gemm_permute_impl.hpp.14377DE4299D9A14.idx b/.cache/clangd/index/profile_batched_gemm_softmax_gemm_permute_impl.hpp.14377DE4299D9A14.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ec8d4039ebf6e9384a25813a7689f9999d765b8d
GIT binary patch
literal 7574
zcmeI1c~leE8o+ZegC=1J%S6b-1VKeMWsww-&4p?~PznktASe_?F$6&nZ7axD#06!m
zQUw9EwG9;>)k=+(#SO7wgb1S0)+it@DEhQ6EpM=Ia&M34JdgkQmvc^X?r*+t?)H86
zyWgGh_nJF*hls`U_{1w&618qQ!D6xW=wHIR_*D|fV)<}ctmM6C0+#w-v*Z8b`%oi)
zrqywt$~wYRW1_L{y~ZcTwJnlJ<+xq<SfU%JT8}O)KA=3bz0k`qc5wQ%r`MlfKh$)q
zN2}?$)o;_+M=xp)u^9{-9X)7vuS^HH-7l+1-E_a~DR8@8rb;bY5$m_PT>nx>$Uy)4
zQq3D{=7;LIT->nLRMZ>s%k_2ascV|-9Ob?3Y4?hDF5p}Z%v-@LP)_O?BN#XyxFoP{
z!pfadI_)jB#h-tdy4LEC+?f>y&o>?&3=yuRDmAUv{l(7movxKw{pZki;f<UVp=s8M
zdcWuY_)$j5oPP1_9p2~rmhrQ*OEY<jvGK-v(VSYg*ShhdtcgZD^LAd>jx0L6O>R8#
z%ad+`{XN{roBhTeAJkBRt{+ZhR_=>U|3oXfM)TUbW8Qh%LxbJ&u3vjqLobqq!u>_-
z=U<MWa%IM8yPV(Mv(}R98z<#~V(xrVdAXJ4^E6(Q;<3q+B%j5OUY)MZ4V6<jUeFYj
z+C*mA+Z-iAgM&YA9=G;jY2D35e$`diBMwC!oU$)l`SEV+NdwUCs~2~^k$szT>6?`H
z<UplUNfI^0Nt9)bFWp>Ft6XS)JHh9r$G?*vfC};6{aQ;;EFIX);#{bV`0Ml(nRidc
z<UdY-TodhEl}7s1rw*+=X-_?JsJwk_%<!j0E2t##+;G>v3(f<H>E13&o5FnpTx^bq
zEVFItoi3<}+}CSpF>7^AXLaD=${5#SCo^eru(Ld2e##@~hMzxL@t;vCZ5uR-bxx%O
z^o9nOJEVUvbTlimPwIb}lBK!c(|^qmNym@HSU#~l`NZ%^==H3dt>U0tlb&&Azi3&`
zO4uNYm`Qt|Ao9W8!@t3&5E^LGNMDp8*HE9u`5U$!`F!hb*&ZE|HD(lxGSf73c=gBP
zK)})<00rWsRwNAni&hRlHLJzJQU{P!6SQoCQ<*1O5gVgCs827^f)X&c1dLM)1|5xe
z)i+nuf^P6ZH}GNv8unU;OD{%jW0RD`M&iKSCaGAv*bfj%?W7pRImgXFvOM6N2QbbD
zZuwv~k{1BC0-TG0xCl&V<fB?n-MMR3#U)5;lldkm<~B*CSfr?xNGcUrq@ssNDjS$*
z1GP5$g#W@_U&3yBcs*cR59*EfN~3A3z7y6#f~tfQDxn!uKSA|Fn|^4=u#TX8wq09K
zj0Qn91CwT8$_NOm1(>#g@rV_`RDoD`!6{S8-Si1`)7ydhcGM@`T4RYrp=T>e4~*J0
ziFuwL7(c}vQ7Yi40*uoEl@2iN48YI8<(Yt=iE}of=K|B;0{B~Sc@E&`;Jh8s{lxTl
z0R9eKo(uT7hz*jt=RSY3$%`JH%S0EoJq#OCi)wmy|49*)#6*J4jYKSQWX48Nv8J&T
zFmjAl8;qc&rcz8Mb(CVVIFmT6&bXOzYHuPaIWUp~A*xjlCd+{h&aQIcj`%VZT!u!7
z+o7Nx=POWf1?Q_!a24lnDCovn1r1bCfa+60dlgi#8-h~702K^k<bsZB_RFt-oI0AI
z^1zflyrcQRHXo?{kDv;GZ2``OU`ipb-weh!;{$I2!WJN68X_nK5GoKGwVheZ8gPwz
zZz0+u|Ib2W9S7HcWtO{#_V9nk((yBXs2y^;#C_+cn|QPK99JB+=(QeM9*T0?lTs4`
z2<i+Fo<XbFD)BD&3D>AzgQT_qi!ER#;}s;811xfI&IJ~^h;0lmuATc+%`CdcCMaxz
zYIm~rtr##^Yb|@N4|btY$m2Tgrv*L0v<Eq;*zGj8b3-s+lcXL1{Rcqsg_qy?hgApT
zNa`Wbe~9xVp#KQx3TRdVtx=yVpj!p>U|1af+%Mx`eHGo}Az*t5_%niO74Ga6e$R4g
zf!JBB-hiBqA6|$Vwfvzz9h-O(T}GRZM!Z;zq&z7`alAN;<bWp!7|Wo63}5pyXd^>g
z2SLf8hYZd^EmcB;N?cz7#R`bU#WpBz!&wQ%N}N04_)a(x)z<~ZUAVjt^7<gwzdmT!
zhmrt->Vxz8pbvV!ADZ?<b+mP{W_cL>IGRgaN&)!_s1CIx)du-(I4dDviE{_!ci`Lw
z`CW+J$GVoM+=z;SB<1en?#r}KQr<4!YMLw0yuaq|ic(ss0_s&jL!@>9>K#Ble_Ec@
zV9K_gw0Rwz9n`bC!0mos#6{gqdPgq@W6OaBdcFgAc7XW|`?ptU44l}o_<aO?Lj+vN
znR8MSrMrW+$TPrs26UJiq?1UfSAs0EaEz~0&G6P=X^VK8c-k>)ix%jc_wRDmq*KU1
zP9XMl5GM$uU`{YbA)FA5!Z=|Vg>%9&ir_?G6v>HH)8ZGq?R!t8i{CnD3s2rFU3OBg
zqI1qVXkQ1_sRv2b!>RR9o#&8L1Dx7`b0f5GMEalX+LE==GPUWo_!-4Nnrxi-{Lrfw
z1a(lIPL8x7sK>bh3L0>3gn~xItKt)qeE;}*CS6e;Ao74VGwURk4~Tr63jk4ob0HuK
z5lf0z3!JYR_tIUhh9=dhg|*4<t-=eJ|6~(XpeWFmsh^;NL_rt@i-Oe_C#Von2u5L|
zFpR=Q;TT1TBCuK_MUiS+KRoMvYSLjh+MRX)iyc6n<fT>mNCNgL&e3B|gZ60<+s@LV
zLpoF|l2itC$iO)h+GisDt)uJ8vK!)4Y0o|af{uVtW^g1`1%j#&Z=I)+aW8k+X~c#3
zfn$+c5I+c`V16)0A^Z@G!uVkrh4aHPir`0J6v>ZN6VnplzB=kL2gE|dFAQdF;%}mj
z^$h4fLup^((`8H3bc9NJ&dT5iWpEZ#!QOq=H$H1foc7jhp=9>2uEu`e+V8{U8^UDS
z8gF8rU)GIRM@gGXeE;u<?)}K?$_K|--W-_loJ&%>f!S_g#n_0X_5kxeU^3!jU|x)K
z9hhE+4lg8C3uo2Bxkz6D{T1kNF|uC=w83B`SHb`#&Yf^cCk#XKYtZ8w^kn!z_ToLa
z5`PxbX=6GtPe(^rdR+mF4E(sA2`n-ZAF=EW2zlr?dDKYiGhT+LDl#}-23?pIM&$l7
zxCn6{9M^|p&q%7`(g)p;ydU!VQEF2i+Bvm+n~uIV{dR&Q!D3%5Cl(_qM~YD#Ck`Vy
z;K)%FB<T+6$>12K!I2cmR)NlMBU5VIhO-jdDsk?FqE0jblInuCUASBY^;CHMeUQ@!
z)%`hoXX8=BoS|>u`Jtkgce^z!ZZ(Sz9RED#{KR8lI=lGh{P*ks4eMWiUb&>w<<YHo
zZn|@q70t~9kL%t#v^U2e_Isg{B)m868%%pq_rs1CI(MDlPt+SGs%65b*=g<zbw=W`
zs|3r-V%=l0tW4@I#VF1*PJI?7>GyXZ%7J<dBI(z%j+M|E#pBD+<TBb&NU9y0wBvjQ
znq0y8Dm1x@b2l{U##sfIs^D@|e?Oet59cx5>N@E&e_vtIJ73g_VbPrEOrtS$JdWqY
z^O*izF+QPr)!O9ZYhk+c@m~Mv-8o;pbrAU7snz4DOaR{b2yg!8v?lblM(Jy3qdWQE
zKiOy9TB=1-D~Xj{X6PgpO+;g~idcowYGO4;F+>bT4S?MM)Kf@Ojey;V^J&07jq@46
zK7;dFz&?xfIlw-LcxZ3!DcdU3iSK*@pC9|O-o~Z5hW3S8;8qJln3-b5N3B^t^2sP6
zJ~H9!k%IQf-{05BSI+-k`b>evd|?{7F&f?Fp`s2dlBmc>MYg&KsT8M|F|{(sQQG{`
b95z>*r=zQ<uSK%xZ>r3t%M@5FEmZt3^(@^y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_backward_impl.hpp.30164435038F0711.idx b/.cache/clangd/index/profile_batchnorm_backward_impl.hpp.30164435038F0711.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6bb0850861d7726b9044b7b2dea2230e1836bb39
GIT binary patch
literal 7454
zcmd6sdpK2D8^F(6HaXgCw~)f1q})%(UEz>RiYQUjMWmsSOY=ojeN+^ckRq2xbdPcw
zp{CME^<hvVjTv+uxu?r9<oUjRn%R4;>5uQ9=6UKoF8jCMb$!>n-uGQ+d$~F}DXDQd
z4(nZ`Lj$+0BRCw69Q_-$B{J9>ayTB698Pr3wPmX}br>tY^|<FS$>3y5+RjiNV)+QK
z-74r#dv@ycY~S{@16u}e=}inBww3QVw{Ew-ct>et)fHin`lHVG{V}}($Kw3u?_Qee
z@UG_PH9F7wW7_2E2V$#wO`gYA4Q4jwP-nA~w>{C12QeM>L7U7J8)O87S5h`tKk(Qf
z@y2=WS;@;F&G~ei>evr^)gxn?QtnW{+;!d;DdVQOR_XUJp~m|qed&%?y4CxRbd&MR
z-;TX}Fw;&ez1j}mw%XYJXs2s>{i5I7t9_&o2m~pyo+)Od<xS<KF8WT_*K(XgUo3y0
zz)?073a8)LwK5<iyC*v}=#0O8gZtKL(H-ggR{)Q}<nyPud&c*<)pY%l+E}9~xqkV{
zFgNAS_vI}LJpV`yedp@q<@2_5Z?vu5<nUd8T!zB^ijwSuR+SbSuP?6ThldJe-fxUM
zzEojkv--hWwaOM9jaPw!gydDx-Y4%&G(SxodU&Xz=ZMSA%3#y6^!vVUc_o2zW%+r>
zI%Bz+{=vcVr&e86Tk)dwvh@kUEB>L_kz1>Wp6ZWNiL&XfI?YovuheeIeXi)0k{50+
zF>T$6$(E-h=Ll*RJ@P5uO%_(iK50DWu=v;D-kxGBgSuQL;qAlHbMGZ8s`X0lvHq&T
z^=$T=4ZKLjbJb}hD$(+>CbKRo*lVuJp6L==r|>YTytZwq;(}tKz4uXdpD5wHq5Xq#
z?*gCYNeDK)%sbndl9hSU@`j_ZzBo3cWvb~)?Vm$4vTtPnc6Gm%d(eN<Ep&E|*WNtp
zkkGj{eK7jz1b9|FWilr!Ce+`KwzA1Zl9{g0_lc7nc?k}sCayN)!ygV80*)8~D4-HB
z+<&ec3L#Vwr|Z)nN?uA{p2PhUA_F%>-I_{2uf1X4ICnNhkW?E~YlAv$$5bcecfxsy
zyCA;{=WfXFMhqk7bc^pJ_5+f#o?&gkG!rwKeD3=O{{yllwQT0HIZQE0dFy!_Bgz1C
zGJp}{Okj`+c!;xrRTi*CoCEA~zyid%z%duNBF+N_d3gJLFh3tSAbA0>DFF6}`{2Aj
zC_vl~?fN0s{{XZbz<CfZ8ibyx{t#R=g!3@;9fqqJxx`tMLH~pWskD;))cvq!2IvHs
zGnGgxhzwFgv_pLdTT`O(MqmG&Vkt;c2Z8oMz(>`Rfo?LeWSIF6q&}DfV+7LQsG0{~
zm|hJCY81+kLPe$?rSd>v+|s#;7DU)a*fGVj?l&|;`*&;8<0S&^M0~s?pqqroldVz@
zoqNQkiKJI1<%i8aKt5n5(;G>Bqx8*eL_4H*U^U|b6%VwKGyy0kfLVwW0X+vyo&*$=
zaQQ(%>jaZ01I1)qo(iZ`T%HCL(~w;Gr`V%qotYDo1l0mmTflT?U=j0L(8@(DXaSB$
zUI<kSAy&T#suum7or<6fs$UP4>Y)nGbL*ig!zvF;B;TK@;7JiwI2kU@6w~`+7QxyG
zL1h4)3^aR!$^^QZfYmKQWdY+XV1iob0DcZ&2O+3jV3rH8{_=os9^O75nC1gB)V=`l
z3s8I2mf8qTuT9`**VZ(%p{?074;}ktx1}KyXPGG3@R+uAT?O5*n1z#+se&m>npYx&
z`eSVpAVKNM=*lsb2+CB(lqD_OAF4(ynx($z4V>|;>W&nD&9YC{+iZ=>cy}egQQF<}
z7V?%cDw$o=>1;N-<*qngJd=mHd%ikfo2f)nmfDuOh|0mVa=<E^q^^Uh*MT}CH}qKV
z%)794KkcHzri3Xo#YP?ly-E>!g`d0<<M1CG&3IxR^hzvcE!qA}UL3zT&98Q|$TLof
zP-2GV&$}hn9<x$$5=m*AYtCbeNlH&vkF7*fhJ>LMk{J<381cM$Un5x=lq`cXh_6A(
zYl!*7OKRIRyBz7sRRWz#bW+Wd8aak-F%h()%K^6>unvQyE(7jmoGSpg0_Q7$dj;p~
zfO{RWg&;#vHl|5^v~DZ)S9O~t*1h72{Gdz=kY#2}P_a-c7WZp)P^s?kY*+``SWi%O
z(76t>v6`TSP+o{4l88-(5OYyNxKs$e(RdS3egf6EeO%diVPwRC*3hp&_gBDfiM>no
z)aCm6L9`&mB81Hy1bP3~dX;f}DP8O{$7ep%t<#V3gCYGJz84j9eJG}Qv3c5NC4&Pg
zba6G9UyV|X#kZS1TMdTq(rPM)^5sy2Y2o#tK&<y_jPGZ6kxRJ?u7w85Ceu@I1>9D^
zu4$RO{H&f-8*wq(gpD;>Ur$n6##$KZarM|HzT=rDFRt$9i;MzG6IR(|E98s!gneWy
ztJC`qg*B{wL66)9EZTrIqZ_|9-m?mRziBG+4%R$Y_aw#V@!1$fA479XE2fFayiLkc
z7#4X4gX@4br{Kt_e|e0avE!0-Xe2U_9g}ouBr<W%0z?+hIe^H)ITsMQIOhQ(59fS9
z<l|fbhyt9;p=3E^!$HVl|5=WY7jC5kOSn=vmY+muMHn)xi@d1p>wDAeHLWS0G>`Q?
zQRnafn6}Sv2i;ajQHM<#Xfn}Z^%Pz6_}Mn%-p{l$F9Gf)fccPOz%9nP1aM1mE(P3D
zoXY^W4Dqg_MWNog_wLXtJqG2EK}}}1B$WjfvLLpWY^ahAv0N<&8s|XvsF74I<mEyv
z&CY`gd3gI1Q27LY<;aIh`4Gz+E1*jS^kK%|d%>{rr{t}nw6<_ho!F5dcHZ{H;(y!~
z?q4XjXW-C(E$9$)3D7^7e=i19G0r7`D#5uFP^E|uoILv5=It*#=yRf{qQ_?S2S%!H
z4moH%r;G8(cCe>3VyS)Kq5o?+srwJ+FZnKA>>pqB5#_x~_B$1K(vxZhvaKj{J|gs6
zy^EUON^AcroPQNAWVR(ISMSO8pT0F0c>&Bh92a!isC&9ee|ATBBE*eY=$Y_DjG*G6
zQXC$dMxpK~WU~~4nt)~#(1KZ5{?L@!&w91T)JSUMoQ>?Zk<=#LP3#4iq{4K<^q3|j
zbqY*51ym8A2Fj<=t%9V^0L?Q1yJcShvo8SFA&^udFet=tuCIa2YoLJodjqI9fW5zw
z)LWqV7EDF*cR>Fgx&V-r5Eu%<TqGX@3S$6sDdRwS9G~BNp!6P0L-ij)%}4kV?S%H7
z&<V*qp+_fV^FNa6f(yEkZHU;X3$8--yP;z@ivA?k4ZXYJN+j=t27M5_`S(Mke#l4i
z0cbRU^B^=Igf>V%1kHzV9)`BVkiE5?+~<4j`M1}n(`PFcXs6<FB@O7Np^NRQ9Bziu
zp4xM?wby{THR!vBv%@YY*6cp_3oY=`@?o>sOS4~ifBni+pAh9YT6J&+Gb0gO)<HG~
zit?KoLZnd<+X<n7sekEIM|+A)+(o+o9;n>|*{v&C)y~U5K*>vxbOy$4!6Q#Bl#7K^
zm>y^wf>R11W)pQ#t`3(Ap|lW1Dv~yP=?TQ;ecQ^FlSa;e(HEJ@+AU@)a&+I)YN`M_
z6@Wc~RWdJ|62dofKKWiy6&AYWH|^8Gw5h}c?ReZ&5`b<3GL<VblFhe$pDq6Ej_$|1
ztt*PGUqAn%H?GEUa6S3<W;}hDwzjgyj)jfB4eL~Ew^g`qyLB{!-d#S)muD>d&i#A!
z^}TOIUt{5sratnV=lIjmsL#Fu4YH08$Ekj+&pwex?eF&On!l9#;42ok0Cz@%Bpm|P
z<M0%$2)Y+RFC?#r*7bM_HwukMaeq4j=T9JS)nucy-OFS8(J$KM+SIR#KhpcE;gelL
zd;CKciEmZgK1$fy-{g1&WhU+Y$SfaQG&u+hRlpamo4&&r-{tBxeNTUZYfso?-%<z&
z0Y(ml14avo1sFLIP8himE*QBHZWt{h7GdO0xMSo&cwpp7cw*#5c(F8+Q+30rSaZe~
zt<`r4u9Y+DO5chs0{iml4;#n5_Iui&eB{>t7~tB<Eh|Z&BTa~A2$pBI0nIk_R%GmQ
z#!W-(`ZBt8GcayOhdMGaWS!`(PE=$-)Nav(wCK;D9~tI>!;wXglh7mmmXi5MF%R_X
oCG_~$3nNhhiAt%safy3*_j;P8WaTDFP9~*gq_`hmfBseFe*{maUjP6A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_bwd.cpp.8CF98960A7C44AE9.idx b/.cache/clangd/index/profile_batchnorm_bwd.cpp.8CF98960A7C44AE9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c8d56c32530eedcec7d580d94239fb0158122e34
GIT binary patch
literal 5480
zcmZvf30PCd7RNI&U<fzKC4>OEi3zJB2!R9@P!=&z0!0PEi!745f>1O8LGdYV5do{j
z1+i9|x&UewHv|=eD@tu2R-tvN#VW1)Qq+ol>hs-1hdcfH@_h{VH~*PCGw04Zb0#4s
zDk{@~!HAp{Q<OV9U&~@J7<Sa3K3|u`ryzsFU=(fHo0L1rZ4CaxxU1*>Z$3ct#PBk_
z<oUK?_%E@4&0Bo=P-}AAi=SS2?vVZG=T~D}UGO8C73<{>zUUwKEOgEmWmkHNc(_AN
z`_2oypZ@psC$nBZy~STxpbz~k-@~)J|Jz}GZ|s)@-)L3T%A}9|<JSy0O>&ttW`=#=
z?WfsK9~Da4|8-RG>_SW8!cb%F>XZ1bz8v4%#;)V>d0{hFJ)6HEL$I{%$X%;LdlqqL
z7Z^0wrthb>xccp=o+Hig<Xmp@?YwRv4K1?{wsl9%z9y<qe9n1udA_)|_}=Bb@OdWh
z!!E}^J2L*G*YV=dW~AkHMs4XFnXYrw2e_>54KcRIK7ByWPxZJJx}@iAV}ASEn3z=Q
zVtL}N9O;FKEF$RfTJE*0U!}Jd58NC->8sogFUGccc;xnU$aNx5nOBHmM-h3C>`9yA
z6{0*bB73|;z=S<J6_>tan-uKyl|BnfZyb5ah#vUE|9;umcH5p*-d%d75|?hvdLH`k
z$Q<kSS(01I;og0^EI~2<&_QX<Pge{2j-4pJ8UIm}NACw!nn#<KuP<EcW!k85G>GJN
z`;JX*0EaKSeC^uSGfOy0{llesBOZTiSDe4_mrEyF*v>~ay#DB%%k{4pq^GQm8)!=&
zc{pU{iP#EHvC~(~zdQ_0pC*bwez{k^sMj^VgXrNl)+JP(GpTyIw)?;Q*1u{^L*V|K
zryd_FZJ+w6%Xd#|O8dIr`nyY7S`Nx%Z!Mm-@wdf}v(%{(j=#>kug^*{_UDc^UN7%y
zvfkUHoaXMh)VW(y=gwVNx8)Pg^J&q|;dw!(2<HjmV_vOZ+Bc4O@xl4;n&a_Wwc^Vs
zjTzaI#p~3yKEQSLo%APmeVfX4p{=L#Mqb)qvDJ9C>dsjQ+w$gR&(cJ;Dxd2C$1e6i
zELnK2h?xKD5G+t%yK-vQ@{I*28-CR3wrF#5s#jM<4|Vg&p8D7)>EW2OPp&>_{_S+T
zP5a9mcvW>vX{X1fir$=w?Frkr<))Nfo7?NaeOr*P3hvb99-7{wKI!>$>-y^8fgjVB
zG~NI1dUxWYg`*r*E4VKYZsIpbcm6PtTO~;h&>XmD^?T`kJkqh-PWpMuh|#g-Q;4RM
zW43OYbCnnfy<VJIFolp_4#%eKIJ+p>y?1BSoQ$lG->$vlpYZ`p%vJ6zjk@y&-~KlJ
zp{)Gziut>idz3b4X1eS%PTYU*o8WIMncnBW2wz*pznJCwan$Xt;xK9aA2kv6Q>^01
zXV&E&b&h@D*Z#egu75-Ye)Idf_FreDuWwSezD`_Hwr1|NWv;2}Jx;5w^nR|f57a8*
z{01lSe!-isu!kvkQ(ZGVDnE;O{QAQ2{Tb>U`PG4tlAA~Ga*p+Um`OhNzUbHcyp5NX
z`L=DG;VpCb=9WB$esS(cNlXTV->`wz$$RpQ%b3YvkX?Y$1&H_m80Hv4{p|nzam}JB
zoB2HE?8yw$ne8mJ6z@a^gXr&9bYJ+rj>nuz$ub+6lcj*N>*k>qCVa1UD031ufXG@z
zSc*aUjN*wgduKP!oJaNcw(+*N6j090Kd#ETSfS!E=TfrNK`OHpP|h8{udjNH@HUr`
zM7?~Oa2a7Kpu8aNOU-n3$q*hhk?JkCk=t7eC>PCkNj>n}?Jyp51|>_JBrcW$%1aUm
z2W@?2rHDC+l1H#cI9m!Rmn`<uvNq<5xQvgevH1jlxTS#dikaJwRqx8#hcjnUy@S|6
zxTS#dcEL5tbo-(*9#c!n9&8VLO9ADEM?dlz56`w@j5*ZHO8~0`4150%HtDi|ar0Vz
zvJa1$MZLk(##3l1pnQWdtih^1caI}8o07d)URIU@%J&Ww-~A=KTh3!9P_jGQ-Of@#
z`C&&;bis$2C1Pgg`zq3wV{1N8EWamv>y~7zqwnyT=~Qo58&`>?fHFfjdyf7Um%*^^
ztu)MiE>N&BGDa99hBm>-xN&h4XqqIPglGxD3E{MsBuqjyGL)Q1YpKFiM9T=u2!q<t
z0)0{D#8-*d7#YS6<H2f-%(Kn2qe(YPr=&?gN{>+fsC<M9!V1)|rR`q!rkLN$t*kKe
zQ{eO|7&G{_$aMf)2fS(C3b3sx?+4g^l#c`KIL*)*0Ta65dy7IB9Kx}<9`q@UjGP!b
z8BPl$wLaRhkf2*yb1d#icOh|71{-1IdO)lPBk2im1H?9z4*=o-$|nGEf@T@-lb{Or
z_ctj^Xk;26`n8i~lMzamB_ot3OG9XZ;{tcs&*hkI-ouh{WmYUwBi2Y^HH(}qo{Uhk
zI2oZdaT-Dk1PdHsKe%QV=Wihgv$!!7G1AM*OH5NKV3h)AXb{);;fXT5IX2PFW(sEx
zgX3Ui7M?Y@5q-+*kKcIRw;Nz&wnMfvY@`~0hI3`{bRiQXCAdTcV-MvDs<>d-D0o=^
z`T5q`eh!QD75EC_=oGp5y3yn(@Iy3(K!K3Iz#pLiK>$KZff6B=K!s4CAP}K!akd0b
zD&m78=VTwBdDMcI1uu($F*nghZOWaEuc4~7PTE0ZOpO-|zq@vvm3f^u5<b07sg0Bm
z!#Np#&FR9T_G5tcext*kpT6ac>E<K*%6|wminiosn%erP(N+P6RX_s25PT(vV>ub<
zFO+~M41N=gG=N|Om<T&z<R(CD0v<FQL9h|k8vxOO>Wv_{5!IUk(TwUQ5NtyA4nTC!
z`W!*&)jx;+Fq~;#ucQkMhwC-7R~Hx$85-A$^S6TxBR2zlGjOEmy9?mEP(A|iBPgE(
z_&J*4K}ZDy^C?IhDFe?yj8qMXYI@QbxdRY8P(B2RLnxmD#3`EfyDzn^E6@BF73>W_
zZUFA|Yz;td05LQhf!v7dMxZvLdLxiGqIx4xH=?=;$W5qj0%{Yj!#ziB3G<;&uQm>k
z8}vu>Bt-@wqbF4Zh#K^zMj$exdOaZOQN0m}8d3c)AP%Ft35ZOneg+U{XdPx3OwQlJ
z5sNGZ;!?U9%rd9|WKJ>Q2}?!(z{m*bp_B*8wBC?P4I7&IC%bDYS9zDu!A`INlW9ps
zpj}vG8L%q@0%(bqf2buv7d{tL{VRY&1?33ulaW=LA*GM1-j1aNG#S{M6U)*}DZ6Du
zg{cv~Batxm8Sq;HXjY^u(rNOq0RF2GH6Sh^o>rA9%2b+ERY0{CX&V?67(%PrLD?a2
zZ7{E*cDgyKf@{Rce0IJ)Y=n^;0I>mh!i8bvc0g=L`5+(;qI?n%Cu!dCTx4*FeN|6o
z7i17Qo_=kHDr4{f#z>u;ZaA&!-Si0MyX7NP5LgfdTf#|E8{=Kx2S1<rbB1++6c{TG
zq+JQv_Bmfb`I|4-*AChPBmKC32r0M<g#5Yw2nBEh5K?lL2&uR#gaWyNgLG(PxE#N9
z$@GsP-L%}VPfqBRl%NgQPJk;q5>;K)7=Gu9x!PVUg4IX4SJmwM2feB76loVZc-)-6
zxX)2q7#z*QNPZ}vx-`7|VI)2d|BxoTAUh3Bg0X@aNZ&e#m8{>rypsyHapE`v4#y%B
zh=jpW&J#tqY>IeRY!+Dq@ERb7W@k}H1HK96dVtrX+yL+fln(>^Fv`sUZ$|kHz|WxE
z0q_o*&!3zYy7Oq#G|CFvU~MRzEzItzeQZwc>3v+lE<pD2MnG(&_a#Qw0iq7&CO|Zy
zd>Rm^Y3_0>F8?$%k#9roAce0A*+Hv-U)7+$VWeWFVlJ)vuLb_qh#D|8fTUHWRyms{
zRSi&miL?!j2vpN*c0~4IR()Y*&8y`bI)0)I!58wyFz~^943!g`FEP*%@cv@UvD{l=
zHRW_{9g^S|+7=3-NinieTIh+8Um@bGg%O1sL`y78qiGEgtf9joM%DtyS|Eoz1|#<X
z$2};w0>@UG;av*Y{4H$4I~4VW%@-(o1$wrAaMNPs7l8Nz_|jwT1jJ61TL965@>xKf
zrTN*uM>Q9VPvuiawI6LC0bOKZ!?~f0-ERsl9*oA6zaH6y*-brAMzvq&w#*-<BX7!2
zqz^fGV9^ka43UM5rY#dEi9;?)3DN`=)TldPj@CS$UN=*pXnsgB7na`;vMTVC?-TNU
z8my<MrX^2HVI)I|!APSYQ|ZU)^y4)8ac1tUoGc3C=Ee_USy>Nd+t^}UJDktc2%;vA
z5O6g1PC^G~ad61^(6I1`2@27uaqeSB%Ku-jqeME)#nsJYxR<w&uYZ6ta7>VFG~uUG
TiyS>ah|OR|Pl=2HA2R+2eZe)4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_forward_impl.hpp.07F6CF6EEC2BDC7E.idx b/.cache/clangd/index/profile_batchnorm_forward_impl.hpp.07F6CF6EEC2BDC7E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc7f958dd09199263efd5a903f025f644e304da8
GIT binary patch
literal 6872
zcmc(kcUTnH7Qpx38I;9gXDI?AqJX$m=|u#jNew8)Pq3l}jUtMG3ag?Z(sT(3QfwGd
z5v2q{5Q5Jb6^tOFi9$3M>{LOc7|^^on`GwRJpax2u^%k+yXV}QbI+V}ZXv+S-Mw9%
z$8-JCD>fo5W(CRP@f4Wf=$NRrk&wsRCd=c+9=;G15q{rP_<r6{ieQmviSCXg{Dd5H
zwX!KOtJW7LX)j%=&@0p3XFcK5>+DscS|y8dOD8nd9(@^ev-@au@z5=C{^RS3tL2^)
z)=|F<e=F9!;?e8x`msm2M&9gsqw5iypKaRAUuCMw$LSQjj?9zaI{u6QhE}*P*A18u
zTckF}^un)B9kX9KF0`TI#2XKbd0Elt5A?)EZLj%5ZKbJ7vdZIUOQMo*zq&Edl_IZu
z&h0cWP{$xPzjfb_*NwFo-hCE)saVdvbl1#FYp)r6H0pBSw$CAby8f=>O{+|H5?8&h
zeX?pcIi{LFVldJuvI{axdr;eBH(0d4r94!2>&Z;9nY>5n-qVvWx1<l3?K-<S(Acq8
z`?oL5Z}euKam(MH>)d);QW&hia8t)(L0MkNt7~4Cjv9&1+h6UR5FYZ#Wp+i~U~QLi
z?9E>1Uu)|_-r3pD8ZfK-^{1-M!q<})5ypn&hb|U$s@APpex$Ma{K4+J^}^wTbFSfs
z{EE!yUdax*lIL}{K4i@cb3!%uxrJ8Jad+V5pw!xu82nZ?!~DkC19?Xyl@2}*xwdHP
zQO_IIOL$!^kr(Ebf1=ZbH8vUQll=l4TvV(@nF&yGU$i4f^I&G4-kPWv`HRmPcDe{&
zJf2pQule&8S6_Jt!<RlOzZXS6)htk|3~G;`6qKB0uqO1U;=0q?9tB=Y6pq5>0g(d-
zHj~QcmLUn|8{S=CR5}#=-KQ<yF`yylx3GivE8|2BL{Hvjvl9WolohWvQq0>g^Rh(k
z4^_><7+dW&wdwU61Kw^eJ`|bty;pI5uE~$W__V<*msb8Z>{3^=_YYpx9o>xRxQLZb
zO#G>>8tf~542L09l3|ef(28=3iah>b@NoF6MSZuEDk+}gI3BGbqoMQX7mp7Ck01dJ
zqQX{3f2qnSFTd#AW}#05ioOk}+=esQo^X@FW0_QPFC(y@Zg0#s6J`@i3tLz2Q=sVi
zy7LX#W{UPV^fyJ60t`}s3F1^>kcx8}uuB7TkUSmOqyu}z8Nf0F%tD+6T(W=%;v8U?
zgZIw^E_t|o1UigB7o<N59Y%5f2)#Z+KO|2imL(Fa5bq||?It!NP9ioW5z&bE5aK<=
zPL{`=G8<T#JU5G==p~v<uwk#DR!m`)D0(%unxk=T7km2*zLjITglmUeAf+98JFs!b
zi{fornfw}co3&>&z6T_oK`3Sr%B(=>cTp>%Z?_Jl=s3MNthZPs=6Vb3H6ttzc;1v}
zY?NEV*+J2vnxSUwohUj?JIo4ExFB2$QG_4@qs_{jMM$<qX$u!Rj2X&tm`toB#$-E`
zcVM!3rFcx10F)B|CQSrNiGY)<{;(^z`eFVjS(0u8>TN)SeY}*-+khp`E^WXK@kywD
z5@Px#P`%`z>|O#rk$y8&X@+VzTQoz?F=`!EvL8$8%o#&cz>E}RF-fO_8L2p@0n;?V
z`In^A!OV1&2qY~5vn0R{Jtz}6Wddi!*}y&<I3UgiPPs^~{?pY>ydL|oG55%1>(^Tk
z>3&`CSs9rsNSKu-+uFs%roNK?G<F~qZLMU@kyb<0SADzelOaj#3-lFOC6cxlSaYO3
zr>6dmZM)o<JAL}I`er%HWkH{fRe$RFJ$@>o9~gm!w?z;;4E@UwJ*+>(T$5p%buF-x
zX{BkU!zxj<osOLzqAD=C3UCgh=ms#U0cf&vlX=BGDw_<;J~P2Ivc}EYoVd>j?1Xlz
ztiGkg6}is1MattTdYbh#tj-z<jZ|1AiZ&rl<Pe#XrWjc)uvm;_<xsX93J}*p**e6Q
zgY&Q6)q3Q{SWyFJ)Bt@}(K@}2H}h`XCdL(2fL{gVSph|#1^lx(R|9@E&b5GFi*p0u
zHz2mt9J(GpZp9E^S_%YhK*-*hq(^|t2*6Ug392;Vsoex;HbHLcBIzdR(FDCve+i)|
zLA932Si(v|*x>R&39$fiHldh}^j(ha)E>$^8pyaJ&>+x}Rda9LGqApIb)mGGA4s#!
zzBc<esu=IfWSSd+V<X_q@NH`MzhgXjjj8>WP_YtDWqSntTuk(Q7Z*I{6(Ep5_d-ON
zFoQ9@9q`)$H>MyzMbWD)dl_b?n`SbNtDzLFWuk?V5#NaGvNWgn>kX~TUoiK&1#EAD
z+3aYS#~$7odZF<tBj|wE9gr)Gq01JCieFx`lIGZK3(n0HZDnqSJ;ct&j;poO5Hzbo
zlZG_Mj2i$K31Pm@7ZR#Yu4JCIZss~}Zdn@@y=C5q5+^>z)P6Drg&@V$eliv3G(e`|
zoDRrzoF#yi;G7A_Oq{a;nT>NUAail9gtC>8tB+9?s=>jRtlu*BTF6^)sULm%cH{ok
z&>c(x7AcFkJYdL7#Caz6a@R94dFuw_zB7P-24I<82KZ$-mjiw|&J}=PfpaC`S0dil
z{()E}R_~I&Fre3gX0pWcFCrEkxz)_LaX(bt53z+sHdM=o*dibYs^*||DMjZ&#XMX;
zAKK<a4>m+7x&Y28z}I$%pzR@Cei+&xhCWEY3VK$-g)Ao(?-cu<y4uY6Yu%)E+HA8l
z;HGpSGm>F9Sm4wnVFq>)SX=~iS;dqyGuwX1*c!nI@CuxEzi7khIp6sH>y>Y>H<8r;
z^?z+i>8oRYuyS4o=rWwk0bP!B1)wVs?>nCRVB_}g`%H;3QZwQfu~~a1Mf11i5dXU2
z2UJs3oa2?_ndh|wVLRYlePB3rX(Fw0hw<P!=y(o#v#ver^2lD(-$#DTYuz{NeMoQr
z<R=74j{=oZe0Iqo^fCx;jwk7C!aAGa7U4y&C+a`z(HYTYq7+Ol0P5_}DEc7KI*69K
z6nzLxJp`sBc_DBp1lW$|7?^tuaIT@~J|O4=N=UySDEEU&hzEeu05X)Khk)V`;Hm^g
zzX7&yP<m4ITQKJ>T0v9vJ7D_`m%j)0@6ke(qVK?|cOYhOCp74UToh4sC$#B=*vjJ`
zG`NS$??Ib;xcnh>eTY`%6#WSLJc9Glc%DG7CwP`Wfvcaua3p^UeV@VroHsm$*bc56
zI&|YZ!XD_;gWta&y7%K%q92C#!!UHe0q8q`-+usZ7{Kp80*yzYInp16#-liYgjOFR
zHW?)n&WVIO(%(&Z?Iy4tY7*g{M9f3-Jw(7Bf~#7`Q-cehZ~l)4Gs#A1M_}8@1fZ9I
zZ=@4}C=qb+cxuq|n`OI8?=h(ss2#{HKP&ZLELc3`Z$wI0L)uN~wMuzb6XfP&>1t@Y
zgy42)Qg)IMT$om#xZjy6NT^`$KMZw-A!k_Cl1}rYeYD~@idnDm+wfU!1SpJviD;lF
z;lz^=3*;uK&;&8LgpijI+$=~j87H5O_*}2}oND@;(__ACIDa+9I^?k4JH{2&U`91C
zWJg#p=)Rd8xsJyO{&@mZzae7&1D%s=nVD#_-ezoDzeTi#dxH%E+4iR3XTD=z6C1DH
z2`Qib_W77Q`m)==m(;sUdzj6ny`4Q4st!gD+%eWwah;d=N^TzGbSuhAk$o2}QKkN)
zPa~qo+?%aJw~kS|&u+{OY17%3n&=ov|Cy-XZ2+rCqd;>M&Ds=Q0{u$xir5TiH$&`<
zFM}}6Ah>n3G*KP1QK`6TKVy5qJdNBjkDt4m-^u)9IIZ=wd}M<J$o+Yi%AUY>+1_*=
zE>i<zZg_8Y#rq?MHw&2)J{Qsj+q%1wt{AzIZWwux9vFF&o)~$P-Wd6iJ{b9uz8LwD
zei$7G<Z*!U2|%8}`4k{e;am#HQp9f#*IzXGR%`m0`}22uw-o23&L{s@BYyAa&rt}f
zN@kjG0j*o;FmI$b@6ycKE!B*m1(>#=REY{(ze0K}A04$SdbjjKUi$Ud=kN1+DG2^<
zV*b!3mZjTNWFExI2^Gf6PN3uka(tc;$;YEh=KoLZX99b{<H?~*3SCP7arO7-e-m{D
AQvd(}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_fwd.cpp.493E9D388D70238E.idx b/.cache/clangd/index/profile_batchnorm_fwd.cpp.493E9D388D70238E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f79468237d5ae4b13d0ccddbdbb3d5c4719f2ed4
GIT binary patch
literal 5850
zcmZvf30xD$9>+5wV1P{s0Rm(}E(^*b34}z#EudUY<xp5I(Wsz;lslqTv8ad#DhLn7
ztBSWA79&tV^oghy5NxedON&^sRaC02qCR<>=&;kim(M4&``zEn&VOeA^Pm4j1kak)
z$D`5w7X>TQ7H2PD&}cM$@}HHRxpWE%1v)gEqUy}twDcdvoQDTY>kb_>vLkM5U7a-C
zajM={*ql2Sbn||&b=XdR@ZAibC+D9Xn$o$slpPb%UZ==w86GIvnsuegL1*a5{l4Mv
z3F6+`89TPlPkd)pv%+@xK+WZ`?A2#G72}2byI)NY8=k#s)z;D>uPc#7ce^SsH#W6M
z$KQCc=a08=U-X=h6jt_hUrAz)W{8bnd{*jw>&KoS7G|=#L)VuW8@DEv&#gInS|{S#
z!DwRW_Ja)#eW~l&#M9reXf_F^ls|B)OzmBE*d;CNmVN$)_mADSB`SWCr2D3{I5b#Y
z__EP!;X4`s%eC>f*Jf2+wNLom#(2?8>()H)FQOkv>id*om%n&<d^}-rwNvRXrSuQq
zqwRO4ZM5pN2aU7EDaosH-#`6n(<{%Fx2KI{DEveEkH7GH;27o|k!u+|T6bGm+xXQ?
z<&t&b-zlyyE{q9Zail^&e*veNNBqLPkUI*tKOK~R=jLY<@bkmou{t9wX+Y%TFZOJT
z(X+iZ<8rdB{EE|&;3r-;`jqjp$&z!+HJ)wo*xXyLaPqQ0CtqXp>~%=K)9!AoF29iZ
z$`dk;WoP_-0~Lp^P2FF7u)257y;7UM9^F5Bs32Y6C3Cmg{gOL1!DEBX4eZLNE4M{0
z%VQT9b3YXcJT?Dtuw1HnOmy#goL_2dU{&!C$MN{k>d7UmmL9K~Eq`7rpKSiw@`FQ{
zo_<}}7BzBv`kA=6w(^0akMhr-KWnLbe~WVJ)feKcyOit2?Y-X(_nq_}`Z3kJN&mzO
zEcQflP4M(F`=jn5!TkFn?ixmqo|UI=AINO>H=XG#c`F-mFY$lb{POGxZoqQ=uvb5Q
zADYu(R~dbcYuTq5_;a#w`{t<+zjk<c%$l{=<jvA|uP-&O*6zKX-nL_m%hEl+wI;{Y
zBk7)S;IhtZntxaPW2?FS2NtJ2UwW>}EZGE)TQW4%dy%<?zuP0!=<T-ARev=*jn$>E
z?-`d|mjCKx#pC`Kkbe8Cm{Z{cB1iub*Jpq18J9}==8JUGz7CPB?Cw6FcWk-a=&$kl
z4ZnYT?^a}PhEs)T1MBtKDm~@wt2bYzm6(;L%oW{<I_j4kGHY6M_Z-=Ap8h$Zd4_iB
z&OC?B;RX#CB@6a1xr@&SGv2l@J1my2SZDH8P}<YeK`r4MJX`zjy)Uhox9s2z3wMO1
zGz{w1kNv!Ryx>t{<Hk33{tA0H<%@^GjD*s;MCI_E=x#>w>eWTctopqNHZ6RYt=K+a
zt8$jUV8x%f<-G55zUTcidS;jP^I1mzr(Dx~b~@SQZn?6g$#`dVtM#(t-H}-9yy7)}
zT7Nm-aMG`f>$qr{_kMVOOs%yv=&zp^x_TO<$ts%Jj!FEQ;Qq~XUB7wb&>9!B4f|$|
z(;Kg<#2*i_Pp1i{XOGN_`K5Ht);m7VJx?yUYj<d6txQ|EmrkQ;Zt4?7(gL#C^r$&B
z!cyDPKwXg0ME%oj?!8*Mu<Ib3o<hnt+BO#I0?Ov$Z(<YVU60xH#iVS?G&NKgP<9s<
zTpK>w=F6tfC1ne33lnt#Wq*&;2Y8H^o7nW_q-?EiZKf`uJZIfN&*;qH{VZApjYg~o
zhU<aV#|x|ylfUlQOWgt%n?9d>gPDn$mAZg(M(tq(v)a)PHa(J*KVg2NuP&gRYd6xo
zBxFhohrX1Q9km?|)CH8+?<?CW6TUWP)0dI5owl8!x`1*K{ki<}ip)K1`T|n6W!f63
z3n-u4Rr&B`(Lp|kzKE1Pm>wK;0p$zd|4`e@59#C4KO^O7jA^Fo0?M83asBSOgFmro
za&qkbfYuN2j~ALM+>^hqAt7<=NX!)$Z6R4j57)C(7av3#?H*5VU6-=@Ih&qLc3^H~
zZlNxqJlMVZE=}BZY7+f3vNcbOhpP)HKeX8r6lLLa1EVb=<wC$H1m+(vu*#DT_t`n0
zCavSslRiEobaix9`;huKlHR_iIBL&0i#CsZLjf=>AV=`w0;{}#Jg+0@a5>kQo=Tc%
z&#>227f_~UE?$ylphx+rmX6vtq3IZr#7W}AdB%uJfLDS~CM+jT0{kS(-vIm@ly3n1
z2F1|R$xc*mPIZ#Y%$-bivLE34fspE?7T~ogw*$N#<z9gIQmi#lyd(J~SHQ%GP{UBf
zcA2MaCe=x}VK|~iNuzuzE!r>|(PF)011K%QFagn$+>?BuHYq1dkrePYauP=PG5y%^
zF-BzQX6RFtEyxyAlq1MNC|8h+P@W(UA;nBZ5Nroyf`*Pd7Eluz7!WuIR>BCmqkIM=
z3wD9~X6B7SO^gV$2y=vwF(Seu0-+d-7=#in5)sNY&YS|<Sv6;-JXz<tQIkQ0@x#pF
zV+Ikyk3cAfAA?XLKM|o!Zl)1z=U8;~!q4x9P6CWbHA*#wk40XWb?&T;HKb#N8OMwV
zLlH*EL^3I?1S25nRjTkuO-3vy)&f?-hzfvL0DH=-jsyHS%1r=oLiswtuTw0YJbHb(
z)`C%82I0(gHiTh~Br9hdiUeE%q6xV|gj~2T2)S}y5fX7l2)S|H5E6662uZjSgi`sb
zW^jmpK?>7(j*cnh92NlI0x%nf2e=226QtSzs0qqp=HRe3M(hLlJ}{k{pc;VJpnL(~
z7f`+l@S7CFokUaTANvPJ90Z($z?^z<J>b-%d<k$ap?nK)Zc!Ym9oU+-n6sT6mfS=>
z5hkK!T)Rg*JvCLXp>3u|xrVv6xgHG6F~0pKJ-KbofblV~!<Y}N>*%OAGipj?9NC2T
zVMH0=l>rON^1T4x3#L$94S3b4ej4DXQC$gmN>pzJcq^)R0$wMocLThe(i5sKw+@aN
z6#tL)^A%O;Nv$nI<dl>EqY_{SEn31Yy!+>*n|5^7c+L8O9UQNUr}YC*$Z*%z)lu)(
z)cCWR*+{I)*UdMi93<a7-xeX~{E6U#5&3@kA&3^4A5YOPz}*FKXd^}(2F8bhCB+TE
zxB=x3VBA4**2%t>T}4S>kny|>Se8-KqvF6a5KO(O8dz4NdNl~FMs+2yRHC{P1S(Oz
z6IgbldM60%q;$B-$)KysW2iY>Y`A#BEiqy_XZgei0n;0D<*9NTs)>TBFrX%A55V^T
zTWTDq0DcPP7J#>)+y(G1iVJs?DfKSi+DCfBha?7W?wg(mnz{FE<&(lTV6=_;1Td{4
z%~Iu5lv#PqJjASj0sLQ7HyCjO;3rUS26!{d-vazwinoMl6poaP&men8bCyG?p^T95
zjrVoP&^guF9!@HQ5PS&)8vw<^IAH=sE}MbN7DROobCpr5C{2`3kz2Z321VkXKwOHn
zlt?7*l$z?E>H$B*k(az3Mxk$yk{fued#o3HOy)qMB*fLTZL{quEyp$op<LTsgz{|j
z5K_!gc*1t|@poLl+N#IZ#t3JYvp%#3BLpk~LPC}hAs3blLar=VghVV6LT)TKgv2Z{
zLK2o_g1#>IwdC~mwf;|pX;|w#WQk|Gl7ie<K9M#-^XYVTq75B~K@<c1V!(wS+&F8u
zqQ-alj;ghRoCjMqvF-*ke?RL$dPJap;KaTK({#FyIz6XeS_Jq-)N=fHZ>_O;o^&<?
zBlNuW$c*~K4<k69oKGpzch?W0h&zKD3<<_QK>x=&xn}CKuJq*Et7O!a@nsV`B7=y+
zqb97pjw|kKE7+P$2C)}#_5wZ}EQ6>5oGO%01I}rb&j8LDlv@F(73DU-X+yaiaJo_M
z0h}I+uV0MzKHfApnw&+sRPGHY7amo}=%hNQz(JE!?H1?<E0Iks<PMO!ZB`Yn_m0%l
zA)Qd*EJB=c3lNk{EE$Xt%7w|4>ar8K>_$}ADAyQD73GLlQskEBwu&NgB@pjJT1tE*
z{*;>Pml_C7>)BNCc5PY5ZE}!$hI)KBIJmEq+qUYo1ziOrqBv0#L07deNXtL~{6HA_
zf$%Eo6oL`eKw6FJX8?W%)s;Z1MD;d+x1o9`kanVa55RjUec)9}?32F2S~An#2uwBt
zF=a?Oz{;td5+jZQ>=??80Bc0~8o;hm42R3m`P(@X``fa?2zyO?KCFrnMSxKROrZ(z
zO0KmJ2TsaQ`hj;ph=z*<o;-o}-=a56DakDF@8?X6H~=^YfHCX<BkBOB4&{r0a}nh}
z!0Ds-MeRsM@2X4L<Oeqp8~8!9UX^uDS~+FVP#vI9`^HAJee|R1H<I-Wz_bD&ftLNJ
zs69dBY<;c<MtE6xA?GKVnamZs7wItOQ6fn2j7$aw)d3mK9eFlRU7#`~D^s2osk(_(
z6(4>v@P<Z<p19Q@oL)p%Xl%&Ewc_FvqGMud@e>*?j=D{tZslo<QkRk#mMqg?XlhMj
zYU^SweU2VGggYx>Dpx1Oz{JqVlrQ!0^7irdn<?Nqc}|&OZ~6b88k?C<wz9VQ#BQ2{
g<8+~mtDD5#LX10$0(r)^L7|EC+2Q`d;8WUv0sBCBUjP6A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_infer.cpp.69EB815355931680.idx b/.cache/clangd/index/profile_batchnorm_infer.cpp.69EB815355931680.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4952169b11c7078db07d0aeed7ca59437e736f1f
GIT binary patch
literal 5364
zcmZvg2~-nT7suy?u*FH3ggrxKkwq5EmY`wZ1;mOdFaolOs89qN(ilH#ajA-kA__vQ
zNI{kg3W5toK!vJ+Z2_^L3$|*-qqfyrj(U*R;x|cs%<Jhl=WsLs_j_;VzU|&OM2Cll
zJ~hCwpw;2Z_383dHilt*`Y%r|%Mj5bpMzn_!zW_a=iKoi9#xy3tog#wZp96)0qw_^
zo7-$%&nc$DM!pZXiYyg|-=6S&+VQ-`x_{U8EGqiq$pNpsUw7sIwzoL%N~a@deDzR$
z+<#ZB_@Q9IzT)JhH?}R=wvVe@E=|clZnsmuUF}!(TvRz!!kc8<_2+MGBu*W24$sZo
z_|f#MIH&W|pV&OA&S_lbR$aBnl6Wwl_1>du+vJ9h>ekB-h5K{teUlPj7OjZw{Jh;^
zAW^I5s<6Y>xvqSTzI>ppnKXXbw~yN2vAR8^dGV2q=Ey_W9iJxWWW4NKbYY`L;P}cK
z?v&`#GyeF7QIj|0%Ecb5LOgD2&X)z))}5FfYJ6Vw#x{2BU`C#~&+7}0hg;L#Mq?Mw
z9x$L1`i*a8kkX-2?%3_5<9B=h{QZdCu;)=9-JH1<9+{Fa4__`XDHfeAdvWr{(irc?
zz_V*M4<4JSTm7l0U3y&wUy&Mh&S=GboePQyu=ls0(!Q7US_R*0dh+qDCPGwu?A7#<
zos?c&gQ9Twu+yG|umhL!E9=*Ho*v51ys359af9&ayovb#xSY|S4!Ju~{xbeid_eQb
z^7C2xOEfZ11YvW!M^?*@gpG`@Gx_aje!l#}>mwE&I;NcwyorUYZq~iYT^YYKa<V=2
z(NwmG<!oKFj!<$Q#`%v*50B=4onzSc%IL#K-C8-h^OpL|o*avkmh|1X>-Y9Na>d{7
za?h~fk9lVkx=v@WpNjv_q1s=6y!*>mr$XnkT8rU37K_Fl#;}FUsumrZG`+Xx#o(^6
zkJ6$ddU6-b2IE{`EsYo@;;x!1woGRYI(&cdy67swYP!AQkD-lDwKwZ47MEMZ?@l@Q
zd~a*_o&5bx5%Yo7na3-i^T(eQuJY~bE3nF_ymWHOvwq_%$844-T$l=Mvf&tIo38v-
zskr!6TK6u)-+|Up;;%K|o$>o-?Edw2Qt8l$$JuYM^dH^E*OF9vEh_%2@3l|<k7XZM
z&hKr0%bmET?BdVbC9n1Ppl03^(Xob9(_6#c%Or0%%Eg`oMq_Krv&&XJzH|MQT>0Re
z&X;p<Ry=HMeEJ~%y^Nwm_pix(l`ertroY=(n$}qyEtiyfm39=?@5$H|5xV5dY!lba
ztsbv!4_|l6AAhrRn_jI)=FxGjLd&A%_lJtC$(Vs*$JRRqPKvF@(U#(?&%1tjom^b7
z@MGu1puf^Ay~~Ap8(&!OX!)w?_;SvHtDpIomgp_beos;L!vBii&d0CQM>6(p4sE`N
zwbj1N^~St?W1`u2zg#u<!-i*e?dg}6U#qmAv}zYG4PM+PlLnnAuQlJZ#b;Jtmu_X#
zgOm}kx8qBW&*zJ*^S5a4cw7@BKlbazt%=tp;a+W1-?mJLJvR)s`nouI_b1a0af*~J
zXSknlZ~2k4Er08D*dG^?fBC)ut2OC+gXL{pzbco7VVtv1B%A{=wLI49D2y`IF*Q&d
zI*zC>v*V%h9;a?L^H?b~w$-uKR~v|Jv<q(EXdF62u-4JoN!LkVZ6LPWd#N;<eA3Fr
zR$&-52(UpwzWc#c;flWO>y&|Q8#bqNu~fQ>9;s)iHh)753l7lfITyYwn8(VbUtpzc
zWu`U|hgd!?coloHk&DI9tv4BLBGm@sSdI7Xv(hu0c&s&aYmuSITx}puK0@56pJY$+
zSZirKM|Y07+CZEpo@3|Z5fo~~T0&!cwmqpf5Py(O^wtdq0d3Y2`rSraMk2L=xNxBH
zGi<x^9>GeZTT6AMgxWw{(R7^WD?Iju$68Kfb6s;2wSoAM*_!@8vTpv&V@1>0M#qM)
zHV~iC-}~KVc}qKwwT#9Vx)%Cs1M#=7`hV-F+hmVp>9lJF>;h74ARbW=UFFrEOc}F2
zpj$hz9W>Pj;va0nUf%tXW6EPC(b&wy%u;P2eh~a{ZQ$B7B|KIJjqP;p4AchVr<bJ*
zKU$o!S;$I$S47R?%u*#L^(9MRlb6E9Fs-{q`_{e`xai<ixIwrOjuuWudPfE_YKg%T
zghhKs`!g)YAO>NSFBQzN1cL;GB}tR~Ak+RRPni<@=Q2&4^3(C-!D5`snU%w5l*~rv
z$f!J<JVeRo%Y7h&`$vPL-ZyYj;8aj>P!z0$Q>jj=3!s8~N{gf3OH7kIojiS538xMK
z@&K@%nM$e#kTs}o1!OC#djQ$PXmj2c=`Njbf1<k?VIJYcJT}Ta3Q@7<v4~1EPefF%
zajrFNXL(Mx?&)Ul9hz(^LKq=}#cV1{7=@@<VJxB&g^7sD733Pib}(+xqoxWP%n+m*
zq|Jtva4Lhyn3+WwO8^T-kasHzMiBa0s_^ly6r*QA5){BZE7d%8W(u7O8!p^?Gu{Mn
zD$_926jl=ZT;kl?7;nJBDKo;%2<`)VO^PEuVI|Mm6IVabN}bSUQ_ccs1K1N<Svt>Q
zl#9RxVXgvKM7asv5alj#N0e9~MwCP#K~$zN(+s*15T-PZb#huqcVUy~rT`cSVY~z!
zb)19)3$Z5WAMp{VDgjXmjG0B&0Eh-up9aKfRQCa*kI~C?LweSy6MJbVQcY54CT{ur
z0=p-r-kPd$)HTy%#!;jz(t|M>{&G0#m<&C|sc0g4X3}x01bCK!V8+1;Kvn=7MppsP
zDwNj)vL5Blz_S_UrvP~h<(<H@6Xm^t>}5FRKV~wlIDZbX-z_@03RfP!)brCb{i1(?
zWaZ(Ul%DQ!dPs_ZVG%HcuEJzP&x|V5!2ZG{!{Vru3oK?++kn|NW;?GDe0uwp_J|!z
z<%roJu!oMQwBsP~hkE7#oPT5^{2x!pDF;mlGguX;-^JbrNa$%^(`fhp!jw&PXf6R3
zCBPa2l@2KZ;f$^V7F8&(0wGl>Zw404C~pQK%_#2#7M&>X1R<Ra|8!s3S-sxtRrFZ?
zeRhC*d*|;VX42teAq@(FVIku>+#M{II@!V=vZ?KWza0pm<GUj?K7Cg5b^+ZU%}p>T
z(}7u)kYQ__q6o?yJ{qU?19CrbfX~9IF97)ks#^frg6eKSb~74gc6!~b(mL}JnT`y}
zKZ;oj1Ll<qk-`>H&Wf32f>VkBMFhk40Kpz+JaFm|Fg^q<7-x?G<724q0LC4RuE|=&
zj##?mfeszc<~lnd;cOdl*$x=(n&7&UQEt0{+irxqN4meyP;n^`A3&&NzC_B<OlhVU
zbieMU(LTdPe>Tubi;_}4uo$P3#7RhxWuy$5jyy6CQSy0mSJ)1olIZzV9g&!!P9l<!
z5K{)oGG@!;R4pKDQGF7SCsBPKkmniQmUzppwpdS~jZ@BCXFlv0PPuSh5ar5sMU)%Y
z4N>k~cSMP~Vnj)}k{NaCpuYt%I@<HEpmVywdEA1OxzixkKXvB#^i^oNvc>=LEmg68
zsu3KPPVVgxonP7=>D3>^51Pp+UteoA7H#lc$i^u>Up@Lv`}c)Ygg3F6QG6*sf>DA6
zf^et?P5m|>Xet<>V^pLtl7z2gQ_*Df%v-LK$_Khio+iD*4g%sJ5W=O&rYZnYf$BOy
z)S<c_5cQ~T2ShunPXXc-sxJcKBC2}<(aY$oy>Y%@cE-ffqmb&E>I=Pv?$GgAb#8`z
zQ$?V_L9hab*`Yc4g<E}>>CrLL*~M9m%=&KNQZlpHamqE#buB~PN`YHBLfx0UQw$Y<
z2E<hemH0`57@8T783Mf=-dXWyLD}USwDWofdO~;vf(Lv0jHEi%!vzl$FdcGKr+HXR
zpKFN7nZu7NxEqxKb2tDvJwN0@G&olQqbiiw0kRI|&A_M`<?VoMM|meO>O}cPKwe}x
z{H&litLmGAaZ;`$M^44%fGlUyB2FCz<WW?&0kRF%7XW#I(eUF0=>8+{zi53{aV@`3
zP7kNPhkgK@^vSY*?Tywy8L0OgI{U9<FrUcSK0wE*Ex=?85Ho{T3h+`em(kS#uSWF=
zfS*8h7r?t1J@ri2kH5Ta2kpqFIh$PJvF)$Iiz_wS+Y}l&<z?>W0apP%fo75JjEWXT
zi=mRGXVD+j^nU|oX?e?3H!D=;?<)@8U|1}3o5<WIFt_nDw;EKFVCsdb@;R;*pO6?E
zw*pIBzd9>}=8<dP*I;XEY3u0D!nu4xj~5{b4W29DMCh9s7@7(_y?lKA`~w1AjOKY;
yFL1E<f2GD|qS=;Kb8PJF9i85Db#s@@mzsN!&f*XwW81JrD_IMd28Dyg*#7`7Y9`?T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_batchnorm_infer_impl.hpp.B4DE6B04DAFD57EF.idx b/.cache/clangd/index/profile_batchnorm_infer_impl.hpp.B4DE6B04DAFD57EF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fe9e7cf1961d7f35c162b867f82cf5b85e1ef4e3
GIT binary patch
literal 5968
zcmc(jcUV-%7QlCgOIf@yuu2IM6oJ)6L12-(7OE5hK}bLiVnHJaC>W*qP>?1?ii&T6
zh=PKEsC;fHMgv$-69blrD2doLLI~(1V0mBOtmfW3dHx$e??<`6In&O$=bR}4p6>1*
zIy|1sD$n?+@QuqU9*?Jqf8#dBt~G}|-UVeIFTU(_U}R*k#e@;RtC{MH*H@cFjanEU
zzmmU1`4`>ZsRu6FHe`i*++2I&%AcXP62`7tD21qYG<LR}e6s%Nz^aXn2CJH5!<v<z
z_ZJ)esrk8|$bJc$cz<(db3|5`^Txyd8L!NnQ`Hl$7mUB;l9|-EepyOs`Oll5BB`~#
z#L9T|=NS=C9~hjRK9W{BZScs=(Sp4&`cSf7d3pWZ^LoZD;od{BLdE%OeNHD09j!W8
zc+om+`3p74#j5Vw2=CB{HQqnBomI;XP*}3fcE8P7knjlK13nIEq*A)`Q{O1i7aB&J
z1g+boD>ZFy+*Gcu%&{DHyxDuEbVk#oKmc?OW*5&ay846Y(8EXD76(V#n90P&+IRMO
zj2l!lGWpHkx^BCXZuumS*4x)F%{DoGu4&%%{@rhue4qI&@nzM)(gC-C9VSysg)0if
zM}ACj+@-o@>1SQJ0gHpxWGa#;PUbZy(A`gLm)57!fjv(hhkCy1-|J-*C{@tPi9c;t
z_8R4BEI7U|MEN`QqRuZool=yxe>*j)0F)H1+`XqPaq#mfWygzWX^F1pi{Fz3R~NbX
z)fp|nUl?}Hb-;DgD`7?OJcE(`w9pGKI|4kH`ulGSwiv4JzO-DlvQ1AmQa3(Q%jWK;
zYsae3Jr1i)FsN>MT{K#e6!!Ys+=>Me)ehO+h9in!kK3?g*iVx;Q`zNNz1!l9s)l%R
z*|{8d-^D-XNvd4lY;CXPWnSuAH|S9=Q4sn)D2z*pS}_lMkM7E$fttH;6vFX}81aJy
zssaIz{|?52c(WILb_UQqff5htE9x7*{lnu!z*C?A0<qz%;|wO?%F9}vI>jOcXmkjE
zdI$=cmT-gVUBxUa2hzwv&%uZ(rjetuqZvaAmg8#5FR#cQM<Z_&Z&Rk2Mhip>%sGlT
zji1R;f@uOli8hIjoGeM4WXI7~^Q~fzQh`k>AnG>3*-g-ylQ+ZJ%_O%#&lc#*$d$yG
z@;~3Q(*scG6>xY3Tv)+j%fS^HZiNao3epRj&J@!qL>OYr(K3}~pD?7<+1fu~dR!IT
zCR!LxREjZ)!4dGz^lEGLY}eN5x8VWXp`aa3U<4Yzt%gwp+l_GR2w?=#I?^PPZLQH_
z5GCH@(S<D=q!h%CjAtbXF=2Y5QK)vPh@;hl)r4$~<{D#879ofrWRaSYge+PRO~_(2
zV+dKSAeNBDX~wZKoip2tPu%(Ue^n?X19~!`&kQDKD;cmR*;xi$IZlImX^^O&4)xOi
z&hF{ZgR5TxwM(Eb$>I_?i(%cH4JvP{o2-;56fcaQ$`n&5L6|^LqIn|gcN9vRk;J+&
zg_7-(?YVX-K#~GxGpu+0{081LN%%+CbE3`fn@dgdrSHY+B=r>g9V(aaVUH4y7fxm-
zP9y8_)+`yGjQw(8n<N8LNT@E<WGYd}THTtZDRX5l&+V?LejrmCKWw?6Y8@K*o(7^x
z&FQ|{VPjZe>17$nOipz6&U~A(jTaShv56&-U&IDtA|=@w*-c}b(5L`REC8(9G?Ic(
zq`-iYo8CLLY53fHbv)93pt2vRze|`r-b}yBXWAR4A9-(X7VBlMHdYF~SYW4Nr^9q)
zJ^P$$Pr~8}N;H~mJ=uXNrjbZPq|K2TWv0rJIb}|eWsv0(PL>Z<@}WAz)<X-=cN_M(
z;!$@2;V!_2xlMK_&%8Tf1GY&4;1>WjrUi`(0l$#sBET;qc@N<4Az2FeQjQ%AhQE(g
zT0YE|$FI5!XfPd9=nl}n1Bm#OLTxD-k5XtZg%(_WDV#5bo*cJ9K^vC<DbxmMwLv!X
z$$3#54C3l{KtTsr-+BMmDZ_jBEyAYoHT9jtRC7O-KDcgRb-BEl?~9mXFKKIj=k@_L
z91OL9t_7@b)2I&6bsYQH7R6|r<rd)fr-1V*;KTF~@KdG2vu_iEKYG9~;=7%W3fI_)
zXV?h%jeylL&{r_^X`w`s!oFuVnOGwY%?t?=@kMMC8kzD<S^8pke}44kWsh;sO~9^+
zTUM9Fm&JsgKJ@?#YN1UnWV2vcXwcNkC+)N3q3CGIegqnct;B35#t~v`M}$HFuq@!#
z0eL8zNdfC8Yh&X!`HfZ2!x0ou#S?y<KqU~ANF@@KL?scFOeGVP0;m+upCeR{JaP!W
zn~qg!1=Cu=T;}`69@PmBZnt@X1(s@-Y(U4=UOQEk74|ReIK~r<S#QT=Y0PREZ+J}S
z%g$BqxNQv(*Kmtk{MmaC6RE2|Vr8-cKO3-tOrspY&mlP%@N-Gd1N=OacLM%Sl1l-<
zl;hO;1yMm?U%h|_`VVmW4_L^kMWYemGXj=!{0H#=1B7ur3fx8kn_Ou$hNlk}aq^eI
z=_S`+TIJS6pTp-bymf_7Yq_mqW}SKGnVp-Xuz*a6J9;akYv&&D`NtDej+a8(Z~1@P
z;f}Kb{von!4#3tWI2WK?lJfw{Be@iyQjT*E7XK2n<=0-kB#LxJGZ{lr$W+U83PXiC
zc@+zjb}RhjYVFa<)OJdqkc>BojX<Ljup{mn4f`S$>EFQi&4Y9DpcgZhy=M$>nU|#I
zV?h;gr~<5DpL3tYz~5W#qi3uG>%1v1|BU+z6zT)oedL<a4yUz4Q$`gEbwHaA$Zqy4
ze%BH`d}j1&JdM(TRvOS_D$*z&e45S~hejD-Vg|XsX9BZKV8zwX0<*FJk%X>;S=UMa
z3E2Mxhy->6*xlfEdo=0+PCdY#>+cqDyTzHBMz?|eZL<AgIO#BCcg1*>fu?nkbtXC6
z)<Fr^Up+LfC*}3fww{zXL9-@E=+g|%nn`Yfvsxgl-@&Y4>7%c{)5lpWRv62!sWgf+
ziDOR_hle~4gl@07jUC5#iZ8oYIwpD?v}D+S21Ow$oFe6vkh8rMvWG4C3T@oRWdJ$P
zYvT@E#}4)O6sV^h#r?~mkqokHP{Yz5tBPDCz{yGm_%bdz(|Avzc?W25W=VruX^=SR
zNTH?_67n{v)&_~iyaTFraQsbw;x{_k&uc%>#<ltUMw^hbX(L#{A|NaR>@l%L{nxIH
z^$|QQ`1|>}B|2*1FGf{s@poA*TupqJH70A=?{ZRI<w{5JL!XZxx2G=N2&s2`@#q7Y
z@2u<KDte?n9Vc2x2S>uKBr_ytj0>DkZ1PMzSG*USQcQ~lOz{t~HU6(3M8$n{8(r3L
zz4;mqIDGVpbQ&dPrOsJ|-UgMI46t_`G)}q(ePpgohra1BfH5VFN}yv2WH&-IzL~IS
zhwMR}#z}cj2N%{|65*BrKmFnl<Q3fp-zwgkKKb%{v2)L^%VqS4F8Dy>_U}6S$Jf)l
z%9Uy4L^%=XF=xt|AQ#GoAXmzjAUDd5AP>reAWzDZrRQZWtuv1s8h;>7m=6n7-(K%A
zmx3=Pu9;40J@@}Fl3xDCpQ{<zkb#SvfMFA0*Qi&Gd)v(&e>i~!4Zxy-bEw$xb<5@F
z<G9$BaogoDYVtqtp1*Dy<$xYJG<tVG89AV;p{cAgj#g7w<@4BsHIJw9R}=gXUv@K(
W3iHC_jpJTbxmTKdm0y_u_52UB<7l)1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_contraction_bilinear.cpp.B7E7F8EC02159FA8.idx b/.cache/clangd/index/profile_contraction_bilinear.cpp.B7E7F8EC02159FA8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a260245b162a8ff0f78a90a2ff3041791bd32fff
GIT binary patch
literal 4384
zcmbW)dpuNW9{}+4Fv`?14UOv<iHse$WXLeg(3o)<w>eD7C5%B9Y3ZVn6}li~v70EH
zqNrHgS|6d(P5XH_+IC&CTU#nt+D*0H_w^p_o;lCH`~LNwKgRPtzvp+(bDrBdFkgwp
zKpR0sO5e=X_)W1C1VL~j8Jp6RYSj@$vH(FcYkTCWX=5wtLx)G3?w#aaYv1A_SeDd1
zFWIOed4HLjxNcb=;{=Oi`D^PPQN-!RmLpOOwNJV2-aWBy+x0yTX3@h#Ek%`+FV39!
z?GEdB$?fQO>mql!@@F&G)($jUW)*#1c7HVcx4x7en>w3#F1JJ68av-Sx8MkR&biIL
zVbY>!J8d2wPD{8v`fh5|0<kQ+vqt|}Sz>g8^WoOwFO%0FQWr!<rhMGe{X1(SHMnX)
z!hx6nE<F|;I4ZYaI^6pxC-~R{#f5t#4<E#&R56<#2hS{=j^2<I!0}t%8Z&u$_)7cq
z2OCNqy5vLmn8W-n+Ll|YPF!oQNcB0g=hNqBm1{2#b+V~1WrMS0x1>!K`#)l1F%jd_
zZdK(a#RXZ@J7+dD45i)f&XRK4zq)Z~*Y?wu#gek}ts7sT9}(Rz>Z{w^b*w1aX=H!n
zy>($Nnw^T&s^gkx?RW0*(AE59t6)ddc-AF@fU9kqvU#kY0^5S=@_@t_@e96SMm0{*
z4^u|NqOA^opH;h$>ErU->bfOrspnaWeQ{Hj^&$KR&U!&o3yEB!{sX<rQ8ab4|JDKe
zpmwv?SP{0kq<!A5EM9t<;PiHr-)WqN)g069^)GUo7ah2=Ep&UsBf0$X^`*h5?7bRR
z@!D)ID6J^*^j#@=ygi~rx|=7NtE8t61v;-^-xFg$dH=H`O=so}g}=(d1}D1QI}4v}
zE&a8!v+^hQw5?v~x1qGn->wn`Yo+BaH63)E_(p!PD6Qr7q~!kTuMLjvds_O#&gvCC
z4WUO@w1#qh@@9=klKxfgoLs#^!~e^ZGn=3AyjJWx-W$Nk7dbq-6LqfnE1Q#>4!Y%;
zxets5cp2UmfB&pSn%%_SchKO)FRx`uJ?oHLR|1ZGw-B}IYj5PvJU2BzEL#%sBp-Eu
zmXDk&seLFaKNB8Un&HMAzJGK4+679nQ5`>k_12={?R0#^)3?3UBeUGgqVZF~F+5Ch
zVoaQbEv)k2bSKe8VyfTDf3q#!(WPHH{&eVhc%uIBhZnd87q2M<6@le%gcX6=)lZ_I
zz5A+oOH=0cCeJapZN;nTZmk<5Peecd(Q@UThb8vvkaKYleFM5>a!O|Q$Z-8LNpgyA
zn!0z5iNY#7GU#=O+nc6~8QJMEMj@qlb2TpJS4D&sJU6l`D+$}&TEB55Vdw0?r(7GC
zKN90J`J<nFm{$0)IMP_z6EmS}8}Xxc*Pt@j86Ay2GJO10#<o;NEWTE~B46Fy(G;)_
zMIsOcqoZ`=_dj^#{gGwW^kCxC(qa^e#KRm4$MpRV5uT!k)ZbR-Ftm_xJUjr90l@t0
z4>GqK%pK|R2^ka=FHrW_Jyp(L04R1I*qsNS#I>~+#_s9lJNc@iSc)*k`^UH9-t33B
zCQxjn??xFRTAHr3Eih*}Udff?YD?sz7@xyei4kt)Wt66|nSx@m^jJ1oNkLaYilfIt
zQqq+wnd~!;EYfJ&g}1R9cvb^1;u2>3vmdTCxa2Y)#XKpV2E=I;TSZv~NkkDrl4?qg
z$(sI}{*VGR10V%z2CC$0GV@LHyx18A1=FGF(1`(}V2f#sAu(tSNV+s#NcuE=NQN{+
zNK6_N5{t%yWI{86w1l<<k}1s;(o)(|NM<xMNai$iNXux;AX(5XAk_oCdZ6ks3f2bn
z+F*ABy>8h3K(C)P(KBib{x)S%Y`Nxg)kXadoGZJYKdpt2XtF5ThnP;GfjX9x``t9~
z@pPxUFCs@tKovj|QiYJjR52t8RRYPI>J7<<>Z4Ma$1Q__ExlcU^4H2B7ER>#$u9Os
z1_gN2&H%eJz@2DC<c*6PdQVG!!l!XQFv|xPL;%Hh1LkgELb@0*i~nx3VzPcz?WLZ(
zj~5sH<JQrcwP`Uu-Q)PJF<neG^`cLl=xD)Q67gQ)`f&rv+c?S`<)oYfoiI`>IV)8Q
z5XGE1&X9N<9wZlzi%RipW{uFHaV6e@1a^Xe7?%W{g3;gS4_d0tb(Zk}Fef@|&K3i}
zhH&Eco;n%B;3l4b6xbalJCA4>J}c(7tGOL$?m+Q;y8+-w<Y&huA)>XDPc%@>$==DE
z$U?Ck-5eWIdB7+StRP(oj0$1z1x9;GXFr@No4RiQIo_xm;8_F2L}EdCWyhimH|y}g
zpPS^LjY5#!PsGM910_ayK&C5W67`Gy)C(S$&GM*thtiyNiPQLGWw@&<p_qUnfFxuH
zA$ia|pgb`{48<f22_$cZHzXg1k4hy+tA9$%{$*@F1#@S)tEL?V6R-r3ge)N>F-xqP
zuoO(fl0foic|-DH`KVMDb<ZU64sX6DiaFXl`jJmu0Zc1^E9q)rQ4NHoYk)-!>_%YP
z2-j~07R_+H1DJNe@jhVMNBV&9-1@VwA9&*9Q3$w&KtSB6=56^Ga+$*|JWvie<$y;7
z4&^;y$I&jf;|<Sc=BoMumszgr>2Ps!p=p&5nECLs%?F%(0L6C$v)$w}o?G^u-Eh1Z
zaEeKPo+73A$#&h-pkQno`)~WJydJC}S~|DC%G+Rf19>;>ejx8B{YB)6!;z9jI`|UN
zq3UQ6+bxRK1MPYMO|CYe-3GfGXm`Wz2ipClzpVB!rGNKb&p$S$PDxefX^)8!Dh1;P
zaYM*vJFa#NC*>641U=G9yOj=P%-PNv63>nY$;HkElB=C7B)%PArK^KMt>driw&IV?
zwdab6i_eWaqa0Y0J_-zvl7ld}i5s4QeI6K|C*7axFQm$jUB)L+BXDaZr{s0>tUWsg
zA&c<)vGv$Sa5qQQgIMx@Z6K-*b~lLXhTRXM`bpn-*q7EN{C<K?!D8KGRr45^C|31Z
zgUdHowb*b4#Rd~Mq+kkn1(c@{DWE)suL8<b1Sz1rIANRzd6hVEoCK0I&L5I24kM)$
zDxo~3SPA7RrAjDIDO0NQZcOpV)U-kW|NYT7qVYq>S(IIZU!VfGRgfctV$Hy_nfwSu
zu@2zY0mu7*TOaAs^n|!@5N(bRG5qN|dUd(s)laMi`02fXyO-Qk#tI({_l|3{;{kh3
zdz9#ipLf|Vtfn-d!*|_EU|mU0Aqw^xu>K5oEwHYIeH>UHhkX)QpM>2BtUF;}0@jyE
zPt<;^<y+xXjkhNcaP!E95-&}l_Fu2B8D#qQ%dMX@;u$%1IWEK#ysW%Bf7`OVjQ2|%
ze(R>kr_BA+fn=n|WhBoP!sdSdJVJ;9g2>2Xu&StzUG6(v_CjCzv*zkG^3bqgSqLIm
zMUgdRaWz>CC5v&X%9JFW{)vI=6peZFHK_|xEp7TDnxBru>jRwyev1tlx`sxse0PD+
z!*i9ro~_$5o|WnUI;(HYGFfV7Zehu>wy|5`z;$wVS+bJ3+)=Ej&-Rh7Rr3xK`GNrC
EKV=FV1^@s6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_contraction_impl.hpp.E5FF1E81EBC1E7D5.idx b/.cache/clangd/index/profile_contraction_impl.hpp.E5FF1E81EBC1E7D5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b96f441f009839f6caf144ff272d1d2cee041288
GIT binary patch
literal 9162
zcmeI2d011&7Qi#duwN30kjqv!DY7YB*h2tiv4K<+sDPq?-~vfs1#Qt+tAfQXwAwzq
zxRfVoEx1-|t*mMVo&sXQQtE~%P#y}nAXqK>Zs@yrruAR{G~Y+Q`Q0;T?#bMlv&?1o
z8<CM`C-HdGGv3HkWiC(S^LRXC<}Y`7&eAQw<GE8jUf!;{w^Z3zJ&Z<X-rw-{@~v-1
zKM9P}Z*D!?*S>m5twBW1xv1fEhkz8Pv7t{pv-Xwd+A4JFj}2O1&y@vF4hro2r<LQ|
z+Y8b~w;FAf`wT`U8%M39cCWPhLKyV$<oSe~{{2=HdpoL&<410%uKR3dS8DTWkz1vE
z&QB}dM@Fq@POO<Big*6w>#1-0q~xD5&*_alkTv$J1(*i9oXHG4RJGvKvi6dBPQz8l
zsM^bp8$^AehFMoPY+7T3ME`oay+Nv_|47|m6{!{--oE176-z&HTNk9dIN`yT$G2N7
z@;VHv4y}u-KE2Z-@0vkXu6Eqf&wZX{#QpI@eev6yYs>F1-g#^C%;J)VpFgoU=f2N0
z@v8*Wj%3C9qbsI-c+<x`<?i$8VFiy(?^Cxz{hA$On`}Qwag{A9mOrsOIkt7@2Qv<A
zD%y6tXxkjiNt269JoV#G$1Gp%zxnXYk4l0r^24RSTW^mppZoSOrDYZ}rD46%CrkAl
zj=N=wKX>>=nCcW4bI>jM?AWvH#;;4RPt4!E#a`<y{p`uEp5HC`7rtBZ-8-^1C+)xK
zbSj^adE$ODiaWox>0$ofqv3+)(ACR!r&SHlPSV{lzg4nY=UM&aa`Cm-7qoc}%&@sM
zP3PCbq??yLj_Z3bQD`qp4gDRSZ=X0CuwwAv&w}JVnP9Ry$g?FSVe+--OWS16gN^5Y
zKk(79B|FuNPtzt%hi`n}v-C+!{^dm{79Q?5DQo=WgP7%l@@qfs&V91;H{*9LcZ>|L
z>0Rk75oYVyUzm6B+y7oO{&ZWwZ#oy(lu!AgplhSqq1T+1^uN}a6^%ty2~T9`pWKqV
z8#$iTwTmq|K9}$Gd}VC&u-LAD=6mZeL>$>(vEqw9VY=x`5ifV8D*YU@Hq92^yIXb>
zMgeSe8Du`RmA;jTr}rns9^d(1TZYbl1DK%4qph{A1>;};7b6;F&UtP@wcpO5c-Bk@
zSH7#&_!mzP5bre|rh{r|v3|teuWAi|&vdZUw$mH`!Xqpj^>Un|a6e)I+Dr!<Z5xB}
zFWjMPQ=Uggb^S&I&|*5+YuoFMf8h=}nTv80Y#4(V)P`O3b`~OvJ_gRmz>76*f~P@Q
zaZ|dQqG-8;+?l;h(a{dk1jRVS5Tx)@_~D+ZlGFf9X_7R8(x;_IVc9~-LPE9&dF?^O
zIHkz96b0el`%vIM6oPp_lI=&+S=PQ-d*`mpQT+)N9cLfs&0ePHc>8#Q66_N=(#`9l
zhc4L+GLujY@x?HgRWe*`;kWdF)oI|<Mc`QkoCcr12ZDQG#C9|~Wv%kwKBGmOxv3z|
zj=jv>R1i;4f*^sQL_s1!NrEJTk_E{enJ)J`B)QbgXUtP0CpC&-ho{f3{p|RSFUCZC
zT5c(SjlIl}i`)%Uv}H6Qi*|`7WHFX8ge=A-hL9;N6@*OTuL#6r79fiPWJR)D0h&s(
zv;fI4SHP?ah{Ie7v6V0zb2XS%g9YYVFsmiG4#ag}hUInOSw}vOI!LI4Ijr35mv43d
z*niSXi%+MzrA}cl^XW9VG=kFI(g|AVwveD=@F)gPR@0)bZW*s%nrZfGrvEzAe0l)P
z2f&h@C=ELgfGf!X1HkQGjXZXMeDZ@}I|%l;{~*W)A(UlH-=X(*zW(u-FZd_>knc`C
z@V})0(jKux_o)Vb$^7^i?&SRAjVRy8%m2eTlwi+bUuWC(;kUOIU8iW7yG+8`Nzoxb
zA%U!Hs<>%cWX16?#Am#!)iL0<k^9Wje(^HbgD>1f5`0d-V`Z`lpQ865Y7Y|OflHB5
zDH3DehfMY%ZmUrAeq^>ES+jD<z1X^TyUuXt4l+BL1AE!8RHnOV*6$xPmu-A(0@(l(
zXnp@ewszWm8)jV#<eWR>)6s%xf?@<Q1Stdxf(nqJ01>PPv6|eeY6ww77;6WgRztiR
z-oji24n^RIxdJ2=;ETBuB$XsrfuxG$8j#eG+z6sZFv0pQAZj7`GKemd+ySBvlCOj4
zI>}ui>LR%pM7<>UgQ%b60T2w}z@S<0)&pRR?HPcn1K>fjYyhu!K0OG6K~jGZ+y;U0
ztb-sQgcz(p1cD*5{}4D0feR^@58?hnuDt0MCw26hbyfi;6?o%ObS0Qnl3WEQRV3Gd
zNe$-6st*U1cNZ0FF}uP%*oIyC%swy=At)b=^MPiqq3C>Y$j8=E^m;H}kLB`{s~#@U
znm5L*(%i6IE(|ODKKX}uih3(hbV^W4Bpz*k(0q>Av}=O?c)G<VX%t<C0?W`etki&9
z8W5M~%)VSYp}%rv;;R@+|JAd{=|xnlGL1fBCcOsf)gU3B!7*e!hPa0s|E6ei-!`e1
z2}OH2d2lP0qCK5F3G#CCBFNjxn;?mkgdiU$AA)?Hd<hEl3iQW@1bGD!B=wRK6l@qg
zkyw?6Ap~U!vrO?Q*}`msltLvzDxr#?R%F|XIA2B4ZOFEb<Q^pLA(vV&lJ;VLE7`Pn
zUaD&~v$_jWVj-HxnvilbKsnFRVG<*#M_%<fZ(5MI>mRArRkxXEPznL1*iBIMUI^U_
zTne2!KVfp|Z)bcoF)A!L9RF@vh>TeJA(0_(u(H&QwZDYuFR)^68YT>z#M+XvZByYY
z=e6fw2wiLkq(5G*u={IiPVT!iAMX!k97-xb)sXE)(P{iNg3|fv1TEw*<S2Xgf~l<!
z77UHg%H%?vM5;c^G7G$F-OE_A4Q#gomk+b$5$KPXMXV*7S)E;o5?B+K<;hyCF0_vE
z`E-_N7PsN~bhc+UK}t_0K`KubM{JN}Jc_2o!uL6XLXMzt_C6GS1-V~ATq=<FV^{Yo
z{!h(JEzpEao5)bzicDKcZbPPRBzGduPQ=ANita(4JtX%c&tA;y)~p%Tl6^9FeCEW3
zh?}7NLoup^-9I(xQ1k^PxqxdUif%-bMv|M5e-q-C5k)s6Ni!+Gj$*H)1^B1lK(RMS
zzKLRQVqRamW|iWbx&~%3l!9F;>FD=@-CmN*!Lc0N@xTYb@c`xx_adwUTGSTI!Y>C|
zIS`Y503r@RG^@BpK4Ja9$ItyuDcaS})rq}K(QbBb1bKLSaK#Jbn7lm+^78f~$lKeS
zAc?nxARliZf_%Mw2?`7jB#aCS4kAbzEF~yQoW<pX6rC;3CP*n(5~LEV2+9%VnB(8q
zid<W96w|mO*EW)SP*e|k6Zh{$QN5Ve>jHKhST>S1zH}lB5jUUP=?&tr4@^E_HeC%e
zs=+1R_R-Xoe4*9PjK*5<S&K8eucmcMt!G6Wz6zj!3!udtrNaAJPjZ<VQ-g~d=WzeN
z>k|7TJ$jZ|+>4DDb44vhKSkzG5tl0NQAYFSvkGsGA3jKnfT*VLf%QF{Q<e;wI(PR^
zdZJ6wnYNj)<7d9eb`e2KY?l!9B}l)-4^fj1MpS?t>sCT|CA@*T3g%P+7qOU}0Fr88
zJ}ck5G4acu6*X2&{7JM=<c>)wI>|nXpk(`Gj*d9@Cd|D*%S~HT5_qe@haF49ay3p_
zHD@p4BC@0ygt$TA{5+o?0`noV1RVmuA>a;_zMa?MRiYM(nEXVpC+Ff0bCRVOO;C(p
z3_%J#1wjQ!uK*FO24gi4c|irZSK!kKO)BkPNpcmqSCL!;?lmOWfpHx<ULzPbf-yTG
z6x{-bEhJwC!^<RhfMEy8*TL{Q$z5RBMRG3~_LAHWhW#WDfZhObzxTwQRVV1p56WL!
zsr^Gz$wSk3x6#bbkQvLkMDmp2zb%FDWbv5Ge?5z=pk!@Hj$iF$Hq6)H@ijSfhrr_y
z<_dxCrTWBf#VcR*H(ylspP!Qs1wHI}X^*f*GCM;nqT~*<;akD46^P^5tKfH)<ZcM)
z#-65DzR7-|`JJ=OxpEnbC_{wHIf9~&Ag;2d=mzA|fQS%#1^Hb;QoJk9gX?)*Ni(?%
zcsApc0F7bOT5*a@(d{5_hnZOaDondd$~!^cNy@t+yqk1^eGt<JvsnGJza2`hFRMsq
z+;gy9urr%h^XU+~5I2^73?Iy(9-aA6%;ccnSzL<9P(T*<%NPpD;!eC6ipb)sS|(fd
z&L;G-1F{Ld?2v3iFFPWe&{KLV2|Z<ilF(CzC<#4fgp$xxd8-ILRe*}nQ-!DqJynE?
z&>KeP!#J1a)5FMi7^kFs`VlgJM9Lo_+ef5)1euSJ@)2Y^g5?)1^KR+o3AC8mZbkmB
z_zaxSWIq0Fh%>7(r)<{K+p64GF5qu2pvfrwm#4zJfv+a!ZxeHI(myxm%i@HZH(g$G
zqWX&ViY6$=D~2FNsDeo83s86gn!$Poib+_)i-6z?NUDGoEU$#bN-~+M2HR?I!17wK
zttGh*QtBWL%kP2XJzPw+NFCpwJ=5y5m*(7mKJm>YpUu0wm_TgpV;#(%thHNLYW?bJ
z+w?+w?(%P$K6=m$y)*;Z*OyG?t$BBeEAK)ST1b|sg-B6|V%dr5zUi-=cWy%Q3uhNS
z9$LdXLV62hc|9_$$7fe}bPs+Efk(G8VXPgb?SG91(oT~5K-Pz&!JSVJkL-4BDQ1fN
zaDKQxYXL<^@FNI{<VO+|#g8Ir27d-YJCN=UBw)uGsCobS`yKBUFa{TcUomj6{05%R
zd#u=5=r<ldUoS)4>pF@)f?SRu?v0fue3~}k`b~36VQ~dJREj<i#^-TXtqGsvX8fi?
z6F!YwNqIYn+e!IVFuqF4J3-t@%Dch1o0Ru~xDU%8>^fEJe$>w4rLCFAi}*dfpf!!m
z!&ig78gHK`-+z8)YEa{GW-iyFu(kNKVoXuox4b>Mo=GboBgtd(mGfGdx)#6v8#{f`
zafdW-7c&BRRQ6o!&cR=ELI;>T)}zpRoVxIGGL>nXe+J0SS(tlU(=gS1{rM5I4Ia-B
zH)-m4pSa9s&W9p1b4v>=eM3rMWGtF!Vk#CI@XT;OOWf4QO%ZNd;HDLBGXF)xekQUF
d9?t+bg}7;io0|W``Ts}2Hh4S}+%(6{{{eCRG*<us

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_contraction_scale.cpp.C10C8FEB4E77E97A.idx b/.cache/clangd/index/profile_contraction_scale.cpp.C10C8FEB4E77E97A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9b77d58a8790c2f219b82d0eaac4e6ec9087d52a
GIT binary patch
literal 4342
zcmbW4d0bP+7RNIfgkT5(!cN!&1~g#UNy5^QAd4&)5D+Lygj$T0in10dP*h4$L<FA-
zQbkm-1q2alKnfOvS^=?A5S1b-xZ(1Ak6K&l#rEcopT57|{d{sW-}%kCGxy9nClkb9
zD4e4b2>e9<tdyj!Vj_V+&_q95wx%T?T0$Urs1XQR)!mUPsS_^L(c1A7cfWDF)S8j+
zWtH5uRH{`cEiY4AQ)AUfYoIf2=9(w@u}#`G($mNEMPpr`Wes(|tkz8OJ>U3E6oYx~
z>anJ|N&3msJGBcjadEYWw{<NaJU(5)>!0v@x%F(O*n{J7E4<qH8Pg;@?!(W6OCHbb
zZAf_g{nV8Nz5XZHuby$Mi4VRQO6%WwC1%x|tlzS8{{67gt-J7n`!5$>zh2t;qtRsA
z9b#mos49ErS+v^yeeufLX-_8xRh&EGRzw?5<lZrgeQ+^(V*l?-Z}!KobJSy;igMU`
z*KU1G2F=ya;I^&w#jeLIWyRXEk%+__)hVlw7Vm#0ON{Bi*y%x<4;x;XxEy$DZ~10x
zL=-#qgnQN8K#w#(raXOH`0Zp@_IwqRD6QkEr^ngq?J_ZY_RGtH(>n(k+d10ld`47b
z$C5RdYFgH6etPMqZJ1_;n6RzPmO7HU?RTMkzIfHaDzWC#R~|W}Sk-a!L!(cw&?KDB
zpdoWlQtR2grcKDR&7(~=_q$@-&ug3knZ?e#8zKvXpYBVD;jXOIk2F+|;71_y4IxXU
z(ht1-PO>)z2CQ+oY^XmsHNg9=^HJ6?=UC~w`uy2~o0$!si5q9a9`0%iPx_9t)4Q^d
zs;)NmeJinoskzTbzg1acPsy(O>1e9z%Jv-Yov)9)DD3^ZVtvWs5yNQd)V#k>?c<DR
zx`UJM=fdwI8HZXrqC<7AcQxDUKKi0Dd)X^caNIqr=~hH;>{UNki;kN9<th6zztKHY
z`jB34p`KlT&`0g}%%R}*XS-tB$V}0}q+QKl$jwU{1nFyz&eHeBEX@5D*_0%%KXi{o
zdzgLa6N}5L_D2uKfA|Z>IeF}dmXM*c2cwUbzfQU36qQp@Ac!74V;xLkzcffc4ZH@1
z9VeF!j|^Bn`{LM*w6yB!@lFp@(=v)M`ls1jR_RYVO6S`4r(8;XV_p1qh^Dct<AZ{!
z)nC?sqUU3G#qxG~%<i`{chduVcJCIdyuQ`hwW?-x{B`BJ=d`NlE&oz4wY>Mowu|LM
zTEBjFxF@53jyB}(^zqn9pQQ_5b$A*Y<(0@xJMP)!?CHzw8lTzzX-bPnR9S0uX3r(*
zx2J=;)QyvKi?g3^=G~gOIh#NkdVZ-One6(}TOG27ONv-?M0%3TmfSdB*eCvy^wuR|
zu;t#XsLH7ui6M>MI!gad-MlBgFfz(%)8^)n`89t$wn|(eWarhj6SichNW^HaIxno?
zA8u?@BNJi?1cXYa8o#@t%Dd}R>oD4d4zIQ9gmtJq2nd6~=<hGWV!ESOcMd_&H~d(S
zOo&5=EGI6fzq|b<qW83<kG9-ghBk=Bte1%ky?~7L09Ft1!3-&>I)N?0+w4)no$0<3
z>!z)<^)*aYPA8I)1Zsjc)=fr4R1rjCsu&^(RiYr1UB+=G$|rW9&#eSLmB1H!gat>o
zF-HFfcNH??MfB3gy2*$)(HkNkq7Ov=s{Y3K$N<#<hyqmuAqr9rQV`}|CAIh5N=8_!
zEXBh;OkP};E)~DfCL&rCEh=V1M6@Z|5YZ?!h;%4A5b06$ATppBKx9NQf{0F`Lu5iR
zfyk6%3XvJb3?g%iIYbr|3y3T!mJnG{tRSidy0t*@szjt2=r+SF1G+Mp<v>@CvnP9&
z;Vo90S0N%5z^($Uz_x&h><4!HVLk-x4#9i^*qwm+6tFu5a|f{NfcXNjyMS{rZ*H&T
zN9uMocv6L_f$>s<0X8TLjR@AVnJwokqoKngaWt^ikP$A43lWdRgNRS!L*z^Hg-AdW
zKqMpy6%_4#Q-3Jq{5e4UJK?ZgjLVrlXT4d!2whqmV6_1rwu-oG7d|=PB)pAgOd+r+
z1Pn|dBRhc64q$?F2{0;opDjx8@$uC^bWc6iF8RmQ30c*tqVBF~bm~x@P(_p{G&Gz&
zg}84(y}}M)2jkN?hB-#!<P_|L;N;A7R^&PvabdbZ<jQn~$c^cyproh;L-Oc!BD#Vk
z>m*NXUp~?*8vl>Vu#M89vy2CUCFZO}W(<NA7;nhwt_h>*Uq|ih0ILpn9yS0dVfF3r
z3YIHMCh%ispDU9QCkH1#`~`A#auwl7M)HASK3IjTMZmBK=3-!2jC1aT7iEtI9gd*O
zH~@SO00CxLv}fO0jo#}usPI>{#p>wJ=K^E*p;ALs2+;{K!p484^8JzGzO;O_dn142
zYRn&B1uQ9gTDIUuLOn}y(ZNQeQA^_}Y|*HtaUtT-co2C~JRv(ijSp$QG+&4WGyz0H
znovQd$Et3p=02Q2)0ac%DB_QZaOqr#cyt~_d^%qdxkSX5?hBECE`Uf#7b>VM{;tV}
zNjDW$GUDjq7=XWVIWQ{+9ynJ5MkVmV`2b)XfVm!+)x+Z(0HXoc+ksg-toH)5UYrkj
z^?WSzSnh-FM-gx=0-o4J2jA>D9T~>Hi3(+aSq9uN;c&ry>-Cfit?0t@jPewIK*=Ic
z@pdRN^Pp%e1V)8${uTmeAzr)|^W9<xp3jTS+yUz)fLVg`kxhZbfUq5Rm5B(0!g#NE
zN7jNUY^96EJF*#O8Hkj@EC-QtoKMG%v5%H&XrY;-Mbc8oe%HuIEzqn5P;fN^&1RTo
zKvM>@9B9gMZmaS(qmGPp|D!r}3M;dkyl2K#i3mG{9gZ*CajoMzoSed)pjUEcIkR!i
zh2;W~E6WukH<lYj?ksl&VfN^+{NnisOWDN~W(@*AthC67Lz<_5t;t0F!FFKtF{{OZ
zpp^j|oa=x=9lkq@W!|6-<{n_sgY&@WL0+V=`hGOB>H)VNkJUlT&x#8@!!^+P7`hBY
z*#GghK#b4V4C0$%mVtN~%yJMf$NAcWzSMJGKh02yNCGcG5#T8C6BK7Gl>8GES%^|d
zLMS#N5fSl3ke!Gxg6u^8BFIh@B7*G1USe<jDPn<G2vMLo2%<1Cf|JBc0@+Cf639*x
zD1q!GVG@PiwMU!@B~38!f1f~qRyllp!&&psmB@(HOX`ai=l72kW6C$eWn}A8be3|!
zEyuTyj5GkB2K?AcM%n?l9oBmRw-@JexxInB!p;+ogJR%VjF+B??e~8^Kdsz~3ihh@
zWXvry)mPR1_pYu&yYqo#K0f{h;mMu(%!=cvkju(-$Hvd^ldEJ~P0`-v9URwbNt+h`
zN)WcBt=}SDY(y{K{<^}k27wTUH$xT8C5RX6I}cMo*%{TfQIQeRp<&^KNQFv>!kcUH
zW(3|`pOUyK8O5Lt!ApqBOI1`!%gE}QR1Hdime6;(mRf+eK268K(1XMC^z!!kz(IE<
y*UHV-?0@y@8PiS7Ei4%}OuH4VRcuFRS9envBYP)-uAcSkz!)XJ5WYWHL--S$Hh%>G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_contraction_utils.hpp.E0F0B9314CF06E36.idx b/.cache/clangd/index/profile_contraction_utils.hpp.E0F0B9314CF06E36.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ac4ed498cac427f215ce4618589b17d26772334c
GIT binary patch
literal 2384
zcmYjR3s4hh5dIS{K;Xgw;Z1ntfFwMkK!_58QXYcPDn-PKVg*HlRnUSDpg4?Eryxkd
zT5WLvQN$M(tu2hg2nag7)QZYLajK$K5M&hEB50NVceqV%=DwNT|KHtj_ut*+5n*9V
z_y~n0MP#OK+@44vgiP_Du|0iLKL;T_9zvOAbumke2fR%Fx!l~Em)g7d(*=oj`ZXha
z7RSrJ8lFsBYM*z>ucBjTg6ZStKCvgE?2e4}$a>M46EL*!{+~~7&K;yuCPT7+`S>zr
zGql-tPQ$$P>zC_Gvn>Zb-aY#IZs|l`!}#cq(z);69~LDvKY5nh(!~|olP&VTt*xIu
zs>})cA!T=EIroU#lj>}&IPp_=a<}YONmIn5$cVpQno7@(M{V=}`}|<tjl5rvZeJHL
zamDpuTzc-0*Pez-22yHT4xA46wG(wz_AE+>>kV2eZYfT!u07%RQu5aJjfa&<b=G^m
zHRswkkN11WW(g#>t4e&s!|l_;T>lMjf1Y->d#`ay{GFXP)XY?V#>L#eJeQt()mKnq
zUEayx<iU4%+55&Ks@6KMZ|~SrHecc&mi&5G(Y|jFOb6d^ynfuKXvQz_P2sf4;ayC?
z=iU=8S7qD>qtzCrUB~7JiAqg6`&@jy$6NC!`|ocKPER%0?6+9uzB03;WmH+St*Q(e
zC6Cr=d@7{bQ=7hD<)%55I1m}5Dd>z99W6NfW=&;R;@#)V@=vw&m`>!pI&M5LRXBFR
z@P@?}x2G~E+w-lmh|&=9+@xm3^rG@>SJ&6>_Dw4CcW&lfd{DO{BP(@%6Bj=;kw14L
zY33@042%&iB1Bd!AOn1aKY>wV{w1|C!ibA|geLV!V-^sW@W-J!=BwR9ZMX7vkjMdF
zP$88i3y4o}kc%U};7}xrFW{Z}hC*f}qMZq6eHI`9t#14x&ia~)L{8XVN=ltrKzzM9
zv-a`X!wr%zt@lz$hzZ&ec0v}wKE9_dt*JKb4<oK2wl~x><g=h9sL8%k9&K)YmqJ$9
z-j=WxvH<ou37b$z-0&+NqOd(hQqC-B?Q@=u7mwWXIz}OV%-@A@F<=4gi`PDO&^+|h
z=OH6(??^gISkT&+1~#WO`pbeSWNm_IAt9t#0Q)0vOMZ@x_8Fj%1*T_1*a%sG<BB~`
zB^!*}E*fz~*gap5&u2mFzWsL8`6pB5yLd=|FIbUQjx1;|bmV>#$d9hRP9aOopC)L1
z79c>+i26R)l+!W$ehQ?pq!~bb9@mt8zdrqnlbU*Qt`XNB2N389j9AbH=;!suG?J42
z6f(yF#DrL%1qkqa`#hWWgC2`0WP|fJC(J1p;CQI)rXb?GC5I_whmXw&Gl~T`9!D=8
zyxw*2B!z78u_<9nu>i;E8<R6)nU(dSi#djD_d-#id$F2nU8ZH34|*1>nTDNJA_Sd^
z)l8eh%R+}@HPfDOdqLGAgXHMxVh%|QlmdS!(~4zg(%}OEJRV8QBr;DhCTWFOA^pGD
zv?5dysg+b>l@v&kv?^2;2@-Ta4$ibbBnDm2HJt7C52UcImC6=u*PS_2U$b1L!3^aN
zavG?>Kwy{5G=NJO1gvJJ0Mdl^$7-hcp(ursbM}rE&-fYOG`(eh;G3k~JluS=lDmhy
zj^s&t3bkhnPlZBT0!5T!rB+gTsuYk0PAW>Z5+opi`7px<G-1lHni(-j2Br(EnbCq=
zVXAQXn2`b}_>rs>w!otDAzxY<61SR5&`MG%05w6YNwomzTnamsS^LpYKZl^@&T=Va
z*W|lAiMROpC@wwUl`n^Q5FgWFCKcksh+;J}pO9kDv(~K*f%k?;OhBkK*Ah_LD;;zs
zsZypRc`3bhB!8t^M+#AfY9+N$Z4PPRq#V^wI+B~(T}M)=eRZS&^+FvfOdSpq_{Mp#
zJLE%ty>+LZl?CO4xEQ38TnIx66N7IvTLp>%>jbNrT>`~`HG<X54uJ=T^?}vQ?f~_u
z!MS3t|Ai79($Lk=3&v$Eqp`jwQ_3)rQ*vu~Em$(x3A16q2}<9%C6S?(k-i~=9SFLs
XZMWA*VqZGfS&YQM7@_~C1)=`{bg8$J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_bwd_data.cpp.E30D0E66DFF68BCC.idx b/.cache/clangd/index/profile_conv_bwd_data.cpp.E30D0E66DFF68BCC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..52dc185d2bd7aaeff0e75db49d76ffbf17a11847
GIT binary patch
literal 3784
zcmZve3p`Y58^F)YDPtUF$YmNvxil?flqnI$C1Md8cMgkaOB*F6w)R_>3CX3~u#8rg
zB)Nr%(x;7fyD3_!wnSDaY3r-ir(Jz#rhR8lzx~edH_x2^`<(ZAp7*@Z^SsB`-PN^A
z4MCh^+>_%r?u<ea1W_k`i92`1_A4R?N=1<5(z<|fpHXYgw`KQF9@3?n*=BSrYIfA6
z`rhO{yK!Xfp7(}<GtI5elXJ8Dc)w7pZU1f2gYKEI$D{AZbk~@a|Dv@oA<TDr?Jj}#
zt%!W<^1%!|X+`^mj`LALm*?*5|NdW*UFmcj!`ml+Cbjtvf7r4o`So|HSKr<oXmdF8
z?vFuL@xQ)dG+m#3_X92UOkeV~s@+8em$MpXre5p+yeFqGvN15X`PzW#X@w%YiIcSn
zDq{=V9-3vS6>r})s&FfBL*iH+FJx$o(Tcczu_tCz|N4oW*UO)|uE=1%q{bb6>}yv_
zw;8=My#K}7CUwrn6Xw+;?L341L$+l@KG&8-pNK0Kj!y_agpBOHdE`*5P5<$AX=it>
z%&Ls5ymxZz?cVyRSCe;&Zmmpd<vnoTTw&ZMIN&{ImjBQqi7u*)DzB)n_Mf-S%grcY
zR5Yx<<&<#bd$)W0yeC6r(hUE6(4%AX7F$$NuVuF8O<i_xRvxv_pd|F=^_ATnw!HAH
zcUP|VpDPukFB1w|Pdv0aUmLDF5!yb$h|{WTcRSHK>~PI?w!8DGugd0mF4Nan4V;Ru
z+eN9Sa9enXs|+{Cuf%5FWb-@DZS~K>g}Gr%vL@7~N{%k7@qB)8((>+hleXvd&6oOI
zSi=G9=akol{%E3V^{1;kzD@mPrRB5vgTANU|8;hie{Y4w#C?;z{3?saH3`#V;l-dA
zcYEvb;Ev&;s)@Vt+jgeu<UEZNeCG3Om*-hQ3&^js9}lcwaq;TY(=5S2o)f3;(!M8&
zvzmAK9`4Vcc&yIQ{N-_7lS1d%o17CW;L%jjtmr?%E)AZy#HNB4?Mt;ozkYpHG^M2z
zV56)#`E~DIef1n&!MdNUlB*1N9bx=2IotQp+9GSqsMXE(?<%533j5CH>$nMR)`)}C
z7R6K-f5mU{eSe@<)jqGV`3^RCr7+|}%kEz;A9o!y_wAWm{HA#CR*#D}#KS$y{HB=O
z(87|8F_o*v1s9f<WX4o~^YmEt<jc+@LxRGu!~q@Z>6J8%Jt+q-@ykvde7^nO-Uo+f
z4yJ!#zg*2jl2l(=4D1QAc8J)&d++3fbIs`Y`BQT-$vbXzIJieF`{VWEBbPQ`iC1#C
z*nIWHW><E_{nSTu@9ZOkA07U!FQt%W^Q5ioynvTiI2rKsX!-bG7d$=~o!?Y1c>b#Q
zf~h5Y@^t`uI<vaTmvTdUQ+iIzx4GI02?Ni?-Pw=x8-Blje8I%7urx(R()?vBwolSc
zv$8cCXvZ7Bi`9=1JXCzm4$(D;ENL?8PsYpxO)|czRa=tMmi=2q{FHDmF(qz82tW{C
zedF#&GpGg)@$p4)9aKj{Du~*bUaJ~?gqr0coOy@<LGT6G0)|vTwf)HZzMXad_Mssm
z{s>M-=^9c2)$Xmu)81HuH4O<MjBBIXG^v2<q~%AlihR5r=OKQCaaBxJTPmPBeZ2n5
zl3?m{Rb&I9?g2;-(Ea-lku)gcs^%TM&FA=!(vU#HfflMolZuZ@1ZgQg;Gs~q>1!Gi
zM5q}kLqjT{`l5I|MPpc74?|)I2RxLwP%5DM>GN6hLYJX<42dSx`k20+R6zB!;DtvU
z+B!a?A;HAjX(&xYDxezKvGL195d}e1rVkVb&cyTp3=iT3Swg3(%d<<ih%A0p#PA5c
z2tzoG;gNcgB#P3Dl9B2rrd8}|y)J;_Tnd*8<D&<@i=LYfs~{ZgG~8(?S}LTwd5ve$
z-#6$M*qKU({dxYYpaY%i{ItC_VY$RGSC4B4jleuWsnY$2`;Or`fRh7O!`T>K1~_G8
zZULMYIm3Bs)KBsR!)>{?BqQm7kq(UIMl%5`6PU<32e5L;TmV=FWG(`XBJzAOV2R24
zGQcV$b0uI^lDQTzYRU5(0jrU$Zvm_pGPeU(yPU1!IYPh0+sec%4B>@X!LV)gUtRlZ
ze$*=r#SJJ1vh6`}Ly94Zj3`DVGNu@lXbEKriP#i2iI!59lE{Q&LLyU&DTz1~4vEYt
zW+dtYYF&UX3)#Kq+tktqoLGs6iHB^1U|AE%AgO6srx>0JEK`9qoQ>fkU?T#b%ee;F
z)W{i@KVe2v`!F{!JQHv-fh;32Tnsp3GB*NFqnra{gC`=sPy3Mw66fKv0t$9~z_^uk
zv4g0FkG%o*i}2F!&XE@ja{i%+NG{zL*?D1<FyyV}tp~%7ICZM^wD0OoN<@}A+Pc7D
z47anlTO%hvn{O&7d$v7^9M}#da%4M_$cgPFqm5opooQ#PHV_f4S6{CK&3&mMy#H3|
z_EH7OQuKO&5nPIdm-GNL$lE+ygBst5rJ)$^qvWFndku!?1EqXG;tHTtLFRKn>71P5
zhR&mY(r}3tbM0j5EZJ?7RImaDDcNmv5wMnX4KS~fGmH@^z$X27c*286TWOc>93vL}
zaid_^!j`0r*o|Hb2}L+BT;@3I_qBSfZ{;$HVJ|H&UFh$ChobOB!3S$%c#HNH7Ho6y
zpQt(^U<!fkc@)nE%xp5}0%k6m4*}*OG8Y48F`176<}ost17<mys{pf#%=Lg-Pv$1T
zY$9_jV78L^JYb%ebNN<J)I*T@gh&Fe1J?;QW)$xN_Fcdcz6N;Y5k)0A@8m_54%q3y
z0zQV}MSxvI=32n6mGhafd%PcVX&MR`zF2v&D)fdC-jut5Dm2qv=(w2P-{1L<T7pL_
z5r*Vkm4{IPI0f?a0K+Q*r;^O=fYUDLbFQM~ddKm5N+@o-*47=mMR7Z4J2yG;Is9dE
zvgg>7$bsWPB1euRiJUl2GJ^9JsGpqjlB6{60nRXeyFpW`cyRv=k!F58KUufwHQJM%
zZtq7U-WS)6>q&lJu3f+b8Xs_rR?aP;XA-l6dBL(4H>lEM;<x{BC@~zs3y@_xw4zA;
z<jVrXtteIuXcEI+DXubl=BvW=a)~V`EQRsHWDelR2%t(oLSTW<9Q{ct#dt(hx_#tT
z5??1<qa<~hxFb4|`1(SSP>J|>g~1y{5r5WJWUz2BQg?nt<QL(gVaOLU9TF+Ojga4l
z%WtFOVz$N-xNeiTB06ur5=B`BQ&p$a)I2m?T|U#GdMsdQYAs~i*z)Zi9G%uI*XG(8
yTbnNaf1^5jEPVq*qa{mCIA-P+ycL#Ks~4@(UAA(qwhsGq&nN}A_0I0V3;7@8kCKuA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_bwd_data_impl.hpp.0CBAC32BD7F22884.idx b/.cache/clangd/index/profile_conv_bwd_data_impl.hpp.0CBAC32BD7F22884.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f9ee2014df034106d54c3e7353bdb663be4b5543
GIT binary patch
literal 5652
zcmdUzc~}$I7QpAgfFukFAPh->5ex>|(NIAotcr+5NK8@}u%Lv31vjL1t3_%Rt+r~4
z%h!b>1*BqKkXi)M*7&H3f)6cHq(}i@H57=xDx%mo*iPp9`IqlM=KC_azd2{lz30ri
z=iV8`tmtTh#IUH;S!=R1t5a|c!))kZ_Ucs`<4iED!2-k9e11mVzO>C>@MrPw2S3Gs
zJh}R>@!e@>ggdg1Mf!c{{r<KYsk35E$mS09Y3-&RZ(HsEi(TgI(E7rw*S5Lsx#HWm
zx7+;F=67<ggw)(6Dz24yF5DQ_m{xXR?RVQX8)oJwm_03`hQB9oEI(hHe1GF6-@b42
z`5j(r7%Hu+oYNB8+5hpNPri-eS<KUl;ru~=D#L%z{!g^WCD+R?rjJ?bT>50{numv)
zPJ-8zyfKXnN)GyMvx!}^V3S|Y?xvoFd*1e0^5OM1*~wc`i<IH-CfoPApQ#<6==C4@
z>!lAPgKal$*>$w@@kjI3MVBoS#y)YqePgW9x2bqAy2-WL<{gVHMR3E01shD9OHO|I
zg+=-%Dz!4aSUI#$PFempY*s_%yuy`xw$@x+EMF%$92l8rb@ccIL6_o<+}ZWqJJUbd
zx$Fl;>Hgl&SM1kcO0Qov_2#+AmGd2PANwXQk6p4T*`;ub*$wl|OXkbtc3P(6TkDE4
zmabjcG=B3C_k_&7nHy9_#T*KMTvT#o#fqohQ!gpH_qHzUP0sbsJeS9<4!y7YpiDfx
z`1oJv8=@?gn{%$_t-ItN+;n2j^s9s4Y93ZB`|VfTy7RSoU2*as)j81*ys93Y_fzC{
zsRiRsPxxuq6|bR~nHGbUZ$DgLlNa08*RY^C^Ue5e%OWnQhC{A&HihO>xpmU^Ql&Cl
za;EI=rV|Iea%$?k#_fJ~?y|^dd#~T>uml~Jy*6uc3Y{<3>cJ(?`Vxx?%z=ZEL1sbY
zo}VxdZi<<i0||m91HI+9O&*z;2<Vg=IY@i1y)DN18?xdaSCy}>TT5U<`hJ0#K=k~C
zaZCUcjiK+yR%z0*t88iQH_t{~m3r`jAe+IX874Dbn@+MUFKk%6gG-Q(lN|$%7D+0I
zlCk6(;u^t{Tc}$UOG<y`Ym7O9R7q4`EU6`Ggi>du#<DU^uqKS9wBR%&nO{At@AW*!
zH6h4Eabf_glq60vlGR(Hz>LFA=Kv=Mz-a&|MveIg?(=9eF|@9xvZMtujW}(TpE#b;
zBFH3u5<<!RWQ0=qDMoVmF?Vm}ult7yoK!k0of!i(O;HB1q;gUjhk}zTZ<Q}AQ#+{<
znc7>8$h08Rf)lH&1#d09AP><(C@a?ksRsedlk^}(S*3?J7#6e~w|H81YO*O#ruwD&
zGna8vGf87i2AoWrlx8GR^M#ezy%5c)^w)7p|GS?}_~gJ#2|qsUWcB@^ChjW)!mYcy
zo*S@8`I0S3nP)&@(1KxF5G#zG%>>o^?w_~ztDDp5%f%oq2IK7TB^y>opRSCcQ??nL
zn!%Vue(}Zk$dz6dFZB>i87{qTo?;N;q=+Z7W+oj@I@>zivE)L!2v`z}#7-<pVkEIF
zO%zQ;WS*oaa-X-YH)7M5=Zo~G<SDHg|LCmdsSr~0)Cg&Trv-#{K<dyu)q$@L{8?K%
zh|<9fmM;K#0o88-ss&tFc^gn|D0cwWfpQm6T`2bhuOA)H0PqH0WbXlVo*95=K;;JT
zF@PV_fB3VHe%?KCcMkoeHiL9COlJg9A-`RZYt>iMf>Mx{!YoD*UHf5QW^Y;{UCiph
zrVbn!0YRPxo3kv(1@2i+`s~<cN|2$l&=^)NOcsVvxGdaA@>ZXi`m!EVy3-Pnmav(o
z$obyAKt6d7Er=7wz0TNDTrW1gSF?8BsGQPsIWx{=X{<k{16cr^0^l=V5#%YbImHGt
z>EA(X#Xi4~wDTc>A>oW#>h5h@)_LS!d};Qj_Z^E2cyjko+VBS1hB!u_AP<A|Fhnt&
zu`2tW_`$Mp`uU{dsaDKog4Ey|gwpUdBV}qAyHD?oeV1Ny9336qnQ2Inlq==KlIsN5
zNi4a!yGa>h{>`}0S9WGCBgmUzcN09=G28;XTPWWKyW1#tgIzbvS$mfXg4?J&bSN%@
z%SCWy46RSy9@vpFbsrtNLm)i_#^rhw=hV@K4YMB5f|<gZ#!{De<n=7o7uV|Pn0y8C
zU%?{A5J47!ya<w6J_z!IEPtTb7y6)QOg0_!^|JMIS+#FK`VBjIcK&{1<=hDUC;xW<
z3mmqR5gY9`(w-K8bpgA;7CcH$%HuoTpdG4(*J>ex8N!#59U+u_w$-SdBVYA;CY*P8
z^ds$66^N?X2V0am8;_N5`OO@stE=k;5Hs%LWGhguXu)g;svYG{pgK|R0jdXO1I#f%
z0%PxBzr9EIJ@F8o5$WReiOgk!%n)ZFbPZ(JAd-=Ooj>nDPfnc^oejz2WaAc(AXCIC
zMk?F6ajyLzu8Fj7UxV~(w%8qxcuNrY<-G)2kQI<+ERIJ#?x+?T<lbX&QbI{wn5l%*
z%h7lpxG{Z?@Kha4Ls_8%CCdX4IsnEpf|CZI3@-*?ngL{t{_(_{lMn9T+X@I$#1UCB
z-4dj;m9q^?E;cUqEQ#&KLY5>#i8D(R?I$8KPa99<K5r{;#HKID7wJ#QQ5p*oy)Wme
z5K?p02x)<%1%!2AqXT=!A%biHVGFxiA?S<|wxZky!ZwuKLD-IR2M9Y*?gU{c%3UDr
zLb(ToJt*tJMvsoCA2|K&c&qg5svNf8|K=5Tbk)4OjaybM#psNA@pb4l&!XvH^Sa|k
z<>ap^pI@mT9R7>;stlxMZ1JxB>Bjsc(;oD^LWIt3O)k;=?$$Xf1EI5RDml_#dy6h7
z@5$bq!zOYCNGsSa+r`bZ$IQ8IyM~Te9r)EjAftGBRk{30S623@biToQddm}jZ}6xb
z{8jP#{JhY)d`p5<hANHg16{rXRK|B~f>Z~n5z@kREf~u_K`(F~7r=*km?Py%(h8m|
zw}Ye|<xY@vqTB<L9+dTvq=yvtegmituz=yl5XlBbywl!Sm|V@LyB^+o-yEa!bDeCR
zoUufZ<sdDGnGF9-)tI$<xiq{~r8<KeqaIZV;1#;;z8#dAs532mY4fYse*5qXuf{?%
zf)qFkT-jwO9*^fTBO%BHJOLp&E=Nd#D~xpi^XjvH$6VcBVQ53&?fb!Z+U1cqr(F=S
z3*KaUrN3Tfj(jG=vR5t1{%oYdA36Pf*fR|b8~MKV&p|K^4CAs*OSWmvHc7pg^Ld?9
lU^KxcEUoy~W9{s1ZFtyNR>O*I^4R9cm)?I4JEnnQ{{x%12c7@`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd.cpp.7612DDA826C4C111.idx b/.cache/clangd/index/profile_conv_fwd.cpp.7612DDA826C4C111.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a530ec2b19c849b70f03dc19408645c320209170
GIT binary patch
literal 3708
zcmZve2|QG58^F)Y#LyXoY)vNFCLAUvdl_R)LLvKdFxe84F0zzclC4ttE<-|*Xp-bg
z+}xx^^|jp6N}K9xQC*el%Xg#hcTD%4IsNW;e!qF<{NLw&o_Bkm=RJNNZf@x+2;v&+
zk(C^m9z{VAM2+~RrKcubl}C`*N(howb2!k;f7o92_TIZU11?ZPG78O9RR7Hf`ZkvX
zR(P8&sz?ruy@*}RzjEZ!fj)hG%OZVTH27-6s{>#Cs@1Fvf;G->smZUr9U8iDql0P*
zTd`|0F@3`indjyeeko&~dlCG8&55H?_FTsYzBRgUO(Ojtl*}%RPS0LDHkx97@=Mye
z3qrlZ$()?DJ4JaPURPV`J*st9(K|Wx;luisL-MVPn>E*z1-1r-Y6{qS8!j>Ig7-|v
zt{9AzdIa0Cix2wdG@S@!Rd?QdTGe@@<M`*?FBdGf?^*Xmd%dOgkmfy;-uf*2Xlu3e
zkGG5-YgfbK(jrRL{NwoQJ(=8|z2(9_&-)yuYu6jTmYw-C%*Zf#?f$x@n~truF4&)J
z^oPIg%J|^KKVROgyjH#tXN;~C`yOxey{wX|wRrW+1;u6uEC+mp$`A27ni?(!4vRF@
z*0k<dixdnNsH9u8^QgTUV+Gz$XD^78rfpOG5o`DmMbmc__`i;*;<x;2cyn!WN{#vU
zYUQ$93u+pZJwoEOk0r?;K1GY}2n?9M(=)l;S(N$F=!Dks3ZAjx$z|Ky2V)MerHZMV
zVO2GLY8{r0!UguxJK1c;?UxpOx7MFrI~p^*F414!aEx8om?YSfaz}ree!rmh%%Qr;
z8N1R|4p={Umo%qy$lz*C;w@Thqu=-hW*u;~!D{5`u6+&G3Kd~B6Ri5TfxjM`|N42w
z^E3A~9$p%?(Gt|Jyr)(0{+5DovrWH({qE1<gE^wHwYI|1;^x(8SVChBX4P<D*Bf6?
zKkeX4GxyhhO%{npEDltfM^Hyr^!8O=jZy1i_pEFdTo0oh%}IGBniqE@zvYHUo%=vq
z`}<<G)dH)N$Ir~yJXqb=<}<Lf>HNi+b;llAYo-<!7T|6Jj(eDvsU^EZHtEHRtCF1h
z!zYfm(74+x%7)Q_^A(|g{gT;#w$AO@Rljpzqd#8Z+8PU=d%b(#W!kxYK+fm4)7!F#
z%Dm2U%AG!V-}xB-R-9Stb?i5@hwc||bQSk59DE#k<(BFv`*_<oS?i0xMhxwVkuR<^
z*m@!0S@M&Q>83+l3X+EJsDDlm8;?KR;5zgr^z(|bO2t{U*13}z?}iuXJ&g1j99hSC
z+~s_dZXeWT?vS=SZbIeL_Gmj^>vh8(<-7OC#lnhb8;w3aut~JJbUe86{H^d4sT=Ns
zz#z{p4!<v5_OH)HBTCwpKlXSRO%A`4>qU17gbfeJmk%=9ADn9ZQLJm=w{K|kndlAK
zIab&I`-fNi9sa4OPZe9nbo@^EDNJNBJv(vdPkNfx>uu7qlcPfcf^b^fG9Qjp)ai(i
zAA)OBwACeo$a~3S+vbz6Q6Gv;LwpegpMlQMkqD^fKAE_lez@0%j)VpvxGF_eT_T{`
zqqFLbH~ND;9SJ0yYf-f55&_jgT_1Dm113{wNHpQR7a+Z0*57}~RP`aA6}R<jkHlKs
zLXlXai#5esLn6K@5hTO1tf<n*i#HANC)6~QrX>+jE#{Q=J7EFCbR>uvUz4Iqmk6jn
z8g@~ACC+a_kr={yHac5ZBA~jXYMZCr-uQKNB$%+)q3Ebf1XRBYnOT0MtJ{Q*gb-_|
zQ|Rgv0o6!q+=?`BDuPUZv#la%JoX$w@nBA{4O|h;vUqo*$m+g4iVJmx2Cx~$m+LMk
zQIu|!lxXn`yM#l!rvL?KQ`yRJ`l=I0X6;;P)<6spzzML0p0vd~Hsv&jmk`ZtUG_rQ
z8y+}CS#r?ufKhxaz_tPhXpQ2v0IMZ)8^GFR4DD2uzY7Qycd~FIImrV$dB8~Kv=A^0
zfvJqQ0_IjS?*PmlWEKG(5!t>9FssP=TEMI&b3I_zlUWRO#AN#xz-%Gw+W@nT%pHK)
zA!EDMSiXN+zY?*9p`1`VIBmz-W0#T?qQ+1PZa_7Vo&yCpq#BaQh-yS4W2!NUSX35?
z=2GX9$b@P_A~UKPi7+ZgB6F%ai7coVB<cn#-9S}3Wsl~Kxiv>{VkJvWmr4%^<}VQp
zQ^^a{3B_}PO)hYS)+k;E9Ls>ajGKUClZ;`W6K<xG9L5HU7XqvhNFx%(s{mF-<`#gp
z$T%n=WJI`U(-~rtI0u(z3M>;q4i<;Mb1Z~+Pv=_3$!;Paz7-GHGoq%i=gH`dt^ZI6
zvuaY7_na6eMp&k~Oc$<4c<^B7AwP$BMIwB8P6F7B;#@A5FC%A`vzd%|EFOtmSS}>u
zv-l)(Ww}Zz&da4|)6Y%O#0-4Ze6^vw73%qS|D)JHS8i%4y1l><E@g_1dx1G*XdtKj
zT_Hq?ZtPybh22t+7rd8owb?|7<^x7Pm<8=9crjoUleq*iO2}La7^P&c1dK{DR|7^h
znfC(5UNScTMgy6f0i&7Bt$@)==61knCvzuYbjlcp<TT~)Dm!sB7F=mUPQ|NfE|?D&
zITf$$GGH&`CctizaeireC;jxzTH+AC1r99Q`5hUX6XLvP5{d{;gw%i0yG6PV*Gd>f
zvyY~abT7Ag%IA;n{L7wbPSi?d!ac8FwS?m7TlhP1mTYIX3oN`8yc=-40S`_JUQR^P
zOx;Yfq{#!UJYWU8p|}XJL}V5NmRQC=hYxc0mC)7YP<)Qk92y)AB}P-~1~ljfw#zGj
zcgLZ)1JywXc0zG?s=JhqR`BMi-nh~8k1T>0BQf)-TTwRi9RS-QyCP7$9$@uk?f_Va
zjDK+xWwr8N3@TD^r$tU4a1s=p>&jg$BWKKco{V@Hk3=q*3yJs`pG2;htCadcdv5i>
zmT}^w{5k&8%5v3keQ_SwpHA!@dog<n+@q<zbL$44aMat2W0XpEsumK~A)F9tIUAUM
z&eVU)PhmuJASX~7EU)~gGrR-|^+a<xCtTVael7rI$!7wL>+$jnirGd_G$fl$&frN~
z<I%65Gd~e-(hbt>U~r|zt%{n8*0j`^G~$~FLBgkm<dMf8oc3fH^tmf32$x4jgoPtv
zP=z3ove)IZSE1}RLiQSy9J?xkz@_ot@|0-`id3cPC{0b3uHvcgCNNP~_MD-kp*fS`
zxWI|)%yZ#eYFR8avNxOa|5~+mnX~l`42|cSV&-fs&U_m?2R++a^Q;$XX|vpyM9D2)
K=IQ~wkpBUej%8K=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd_bias_relu.cpp.2AED2F5E417190C4.idx b/.cache/clangd/index/profile_conv_fwd_bias_relu.cpp.2AED2F5E417190C4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2a3dc843dddf5669958a70712143b0b02aff3727
GIT binary patch
literal 3942
zcmZve30PCd7Jx4j<O0_l!eS885`+X<!V(2T2oN9(VGE!jDTrKF0|~UGD)G}kOY@*A
z6btUM)cS12mWo;^(ub9*ZABjnDu@e;wrEgX-m~t@yNOORuiv}hm-Ed(=bSl{nK@@B
zO(v1#P!S}yKvt<P{E(m^2tp%&nhz_K=PVIKXp10~P3`Gfn(LuX!>#(W1AQK{vsX_2
z?%{DJ$V)d`RCC3<-v4CRbVUU#O!w;9N7D}49|%{+79SZtw3@$S?rcfH`k1U1#}!}p
z_^#g~_ih@T%5A{E9-066=)<COKk;%`trFK4S=;U!9o1Xhxc*|z1J^vk>UFjHsrwTr
z@9|bW&EGl@8THxF?a?QDsgmtub;bHSubwF%K6&cV`w?<eTJ>XJuBVIlruM=#ZHs1K
zOXDvAyAuwLul!?^T-WY)Tl-aE&xV1x=7RPi?U%J@)B#tC*B>RcbAP=uiRykCjZx~C
z+t(M@PGpw~cIG5FzTd!JWTkA8O?m2@mOXgsf5mrU|HjWNZSX%^>BHv*u0Fe_Y1{0Y
z>ULurxgUSwUAJghb>aHa8o9gA%C1oUwG>ZHS(QWcZ?4&gidE75O)t`{Qd0x$#wKz0
zmb%6&>rXX4^HvCs46KN`G}}&hztrvU(G&D1E(<xt_$Q^}j)8LVr}*Hx$qj7x$-DhL
ze7vr4ch{_u&sVQeY~S3yzbO6E@L5TB<I+N}^Otv~g-<LQQmmXvevoYwXBBt1KXkY)
zwtds%J*ad1WK$l^X~eeLy5iWos?3lBH{5i!+2Wd|_39m4-P3mO|LtM*$bG@p=H9QX
zcCF=CH|F#dGus%+KHcM8RcH90@HBf<=XcLcxN@)jABPX%b<y;jPX5P3Iwr1<&9`}(
z+L?7QXt{FfiRQI~>dRyOf<UiAb!wcn{pRk=w(n7;`K_A9a_-@OP59L@Z%cM<3O%*^
zce|DN2cP9N^mOM92(?d({py6}icX0>MmYb+&Q((aysvHO!B^|-+oOYSKmIP~>GtIf
z>wU*4pYmG%Qzxwrq8(QWj?JCfcIjA0{SE21t~`su;qw0O%rA)X%^97xi@a<7cd-{9
zv34mb%Cc5|cR77%C4JG!<JzkI!G%Ln_nIGe>`6WJP3HcqF1;O()hF&f|0VoEtH-0d
zA4IM@JUN@S+kNgX@LqPwG3ltoJmK|k_vlB)6!*MNY?S$)NBgEzS9wT-jytbcSdYeD
zdyR}9R3UvF>I<LED_gt&ncbLw_WSXceJVFxpL}EgjD6{kW;G9F?>%sFWVBdlz2yA&
z&kqicDlfhqEx14Bw`0|Fe(u@6D@{rFCO&zxwL?!SsJ+9LefCm#$0h1?#@;i_dTrBr
zkB0u1dh7KA+btI*Ws#0Y0-ZzuJ(=!uyJc@f8hU5WH@(=oDDjrjj+Sleb;km~p8tfk
z!R8bu(WAOvtCo`P53!A6QNxH-=yFF=voBb(wpTX1`Tm7|z0Gs&_=T(2Pqv~Otvc_3
z1%k|IjAA^Ra8OZ^%v6NHDLB<A$m}<MZXK-|*J^d!t&lVXA!r!Q%_xAHS&8)&dkku+
zNY)~RaHY7SMqyIdx$qL^>MN~~g{1d8*gFiP0BUw$Y0s_FB)JumL8|E(onaI}?Wd0a
z$@^RT%o)g1QmsX_ZbpF^$`koj9>X4cC(|!bk;MxT!inO98pRtWg6MoLRX^M+SxrT9
z(h-75VWLK1Qs4cFwKJsqXDad@IrCYRS*TF}^;SOnm5#6sp&|-$S!PgXP>lkp56$_p
zbHkj4zEorhIk-E;9W@G|?!S@iv1hqF!U{<tXKshtxflgdUy7`W|H{=p4?~osnoVK9
zYZO2|7QB!3Oi_50ie!_2%#q@V8U;{S6c%YNqUN2OW^3Fr5ECOZXJ>Li2>a*z-0B6s
zyDTtbNyd`*KsVS{OWXh1PK*fSgb9EVIZ|Y15+HZ{<FK{psc~{7jiZJMy1@w`XJOcW
zu%#HG@K*SljNYCDGMwQMfPaD!;hgXYU_`n^C`{ozdMcLQStF>npb#2dLjy)&d~zhi
zc`=W~=kVc?oE0qK0*;Zzv^U4+tq<S+=jP6BD_+@RgfHrg18a=%L;WD~NBtq<pd5$-
z&;W=6(LjiT&>)C{(O`(UC>Nr+=v)&mOm^_T^|`>x8Y7f6CEYx!5?TpFrL<CrR5TSt
zYML6NGFll#<+O4WWeP94oL+vohg@JoYJpW++;psWe9Cd%U%HpHwRL@y7FJBh2%eB9
zGEXclHcV<JzK9=Zrf^X>L;|+J*Q|?RM?e(Gj)W+R9R*P|Guq4S7sHH!D3%!uk%%dR
zNX!&NBw<P*k}{<b$vCoLbA)(KJVXhc1c(wji4Y}mk|0XvBts<U$RSGMq?oA4K{xuR
z^%?JJ<iazwo?zigVrl&Wy}#_WhmF->4Vi6OGqXa#>^`WBJk97IdTC7|l$eqVx+z2n
zRsvBeRtk{{Q$eK0)DV?nWe}BP<t75<l}v0yZ3PvBEDG|>1t2gevZRrr!ZP3HIho3G
zh1qDs#x82#*=F*;7)~0<S>4+TY+-Axydbk{?u!{=r)ym;F(R9vEd-G;VljU)L^=E%
z6RnT8)J@gD=DA`-G#>42c8S4bAd1CfArj#th{U)UA_*>mNQz4#lHoFl;_-Nh67U3w
z67fWclJF#mlJR7S<hUH76g&l@G&~KW1^5Dp((!bNGVlzDlnkY(c_m61B@mS|N+D7)
zR1m2dYKY1hWe}A!%1s1rNOJoO*CV)o4sI4Q?T3bX|587?C7UY06W+W)rw}|bPX;DN
zA;RLq63xVy@l7`Xg$S30LnQDM1ez@){30NV^oxWj$}b9{XjZh3*)N6_15qq179tT#
z1d*5}hDgGaKqO^JA(C-rq2>tj+<1r*xCsy?auXp+;wC|q%uR+!&Xq%y!c8#|czciw
zZ+LisoC6y`-i(F^2gowAUUPV1V0L8oI%S&f?(@=oo7#H&YBDi%GjdIj17J)hwc%}I
z9!VRZjROVh(%=6WIef$F0O=cohG4)4yj{r286Ga6YD}itxh9E}J4tIIgD^#Uyzy6S
zt=G7d;jlnsR}`uULsU&gzGi`;k!uidzC6%CEQs3b#!9;Y9({UYQ9*ug9#UY^Aw}k9
zp}Cn4HPr<w^7Vxz7AIR$tfpI|Hg?zynmrYdcaVtYIM~KJx;Qz{a^;6dL`DgtV}fRK
z1YSH}=Kt%ZGv0aEW40&D+sDs8AecKhWPX@?s9WGX@k}~9E}?juR30m{NJ9P#JlsuH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd_bias_relu_add.cpp.064478D11C1695BD.idx b/.cache/clangd/index/profile_conv_fwd_bias_relu_add.cpp.064478D11C1695BD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..870b378d50917218e4942a5854882be6d7fb618c
GIT binary patch
literal 3946
zcmZve30PCd7Jx4hN`T7^0W^w1BAS3e5P}N{5CoDy_6Uds5WKPm5~zSkYz6&9`&yww
z6|E>*>hc7wD2NKKK&uwFx&dw#R0?S8@&xpC(ca{Bl4-w}@8dB4oH=La&N*{tV#FfR
zQVN1Zri)8-8AUh=K@ckOS6Eb_>9s<TI6DL>+40T7r2I!gF5|oI{JUDLIr(6lWL5S|
zEtM^7aSM2z&i#B-LQrwf)|m0$7g=Phy}Pe#)wWK%UL=3~eA;FwiB6q!$-cX8e^cXs
z25;JWju!jlT_a;x{-}CT@hi>d+=WG>*~JY1+Pb=_2yMvbT(0xukyy)p$Gf)ADtq*D
zuxZ-%Ri*SJ&+fc!q^v%2@>Ru^=S{mr8+W<)jagN>2R#2^twY+5+sCTazvw$oue_Rg
za`WfQy2iSFcG=!L+VeVDdaUYdblu>B5sz5WaKgnE!)H@t58PjNoYxugOxtj@DA;*p
zR{7Poo|1@D)vW#*@_I?5-mg`;r?6-5o0g&jDxX;9BWBR8vXsQ$mwWb9*xiyYex&ui
z5<b(}>&c#1-L($B{Ey#`R<o}4NNmWuj1IZ#>ooORpEsM?(Q4J%xbRy?tC$<x17k1t
zh&!zM*16OMPyK2rcwgm3OX1r}hlf|br@fsNH@$vf?FNs8q2s=SPqJN?N5{I?KJT85
z4ODKbImYd^<sVjWi*7rTdAo0TWPRJ??2H+@6Ont1`@?Rif4yAyXq{8PUDf@C0k<1k
z?s~mgKuz>+u&@2>s}u36kpAi8OS&3s-0~bb1vT`qKRT?7E#2Q(A>21F;nIesE5>(j
zq04VsC{~W_#KjXnlyo>8&nO>8ANgjlZSQD}ds<r7HP*U38(q0fv39S<;oyS%zp?7W
z8jEM!r)^#D*A_Ulz4Bht&Y`_~^uE%O?9o$SB#A~poIS>!_Z{o;J?)t<$jzELi5e?E
zX6g?^RX=$#8#ee1qMNstMQKPYFEzH-_>qcQl3w>-eeAfSk)Ho7-NTjCQu^ZQ!h;&m
zw*5Ed{?+FT&%FAVN+#e=p}dj_pUw>!z4`Wfa?%kb_u0HFZ3$oD16$%N?G|~KCq~>;
zb@40eGw1R?v>!awJiIuOO&gkg{(PMJ{izi<y1o_-9*X(y9b<IeseEba!+Ae``d+&<
zRWCh!i+(V9qVJ;nfu76Lhp+C;PWqzWmXjrnEPGKmrf5#yb};@M$Ml!yXa%I@$t_<l
z&-{q|j&^z6ck*ieC95(C^P#8TtllSmM+5G^uBjZTdOl-nS_!+EQp722Xk3|^>Ao60
zbLil2o37uP`RrXf=k0in?93{+s*(OY#l97<!&k~K7FlwHE4F1^DD^5ncP;R|3-9ov
z-L4l>e%~N`@7}cR^09)~ch0%hbgoZ1ZI|&Nb$Pt)n!1eEt!u`W*O$#bn3cWV`_;V>
zhtkO<mVxEhd7Yg<Mr67Axutb`cznNox@^+Vl24if&wqAG+_U9n-;JNke?L|7B=tbW
zzQahNUbnQ_0zn)$h0(`<a?(<egartWkuZu;5Ea1qn{s$pe1AoGtu+#ZAUGAJPB99g
zW|W}qS<|lTDM->H1a~F5l8wTouACGk4IC-4M&bx-H`I-86hO`D$Za3al`E~0ctTA>
zX>_9i>XL8KH=YVw9|c()i{LIK7qU?Rb-A~d_V{qlDhiUa5WyKF2H7Y~>gQ*f^>a>i
zQjkPqX_HBl$wmRx>mCku+vZu6p@@c<n?+)I7zI#&$z#1M$F1g25H+za4kQPPQ2=!<
zv%k}+N%da`B%83-Bl;;uff%YM@oV!s*;?fl=S@LUiKV%d+{s4qKP7^6KVC9zf3Y&u
z8c`9}_Ne_NqX6n31Dcq>sx!JMNHX!=&Ln5DQ2=#8MrPq%vN>{=cE*SSrzoD_k-!F@
z=lEOGl#L<YyDd;WH8C{>bc3*3+5Ih?D9&YbL%<-y6F)%?$Q@h5{gixpfbdl4T*v_3
zAOnb17{U)iisEWdwU24gUwJ@GXGj9@B`6-u4h{u_2$N8?X}W3c1=&BZ4q0zO!V9rN
zDi{R(6P^tDV)n#i^I%UC)J$N)bL^~Z-L~SL9g2ICy)n>@;yz>_h<wSu5V6T@i2TTY
z5X~abg2<oj4^aR)03r^V15qG3&_r<xC(q&P5NjJ0*HASy^P+O7IS}Pib0N}FwGip3
zI*9V9c@X7O^G%d6|L&w~#b?@yoirpWNZrLdzCAtgi}Ppy=uT~DICdb0n?*zMp!q=p
z^TN21+$b~g1iWZ71q*^93SotKn{}bAP>A>}K15-xFo?n#;WNy35sV0kA{mho2^a#1
zgbX1>B8CW}C`J@SVzxNIJcEQSfk?`hLL_6$Ad<7?5GmLSh?Hz4L@Kt*M43+IcgAdf
z@U$d0p01w`Hl84+-WSla68<%M_l-YnNVo>oP{8U)cn+EaQ7)PbkrvfLq(gNO<)L{H
z<)ir~0`-(Qs)k|;Y5;M}l?kz6Vo+6yL57mb`0L~nhf}FG531UHjM8zlj;IBOlme%)
z{I8O0VP`D0;H(1IgiAx61)~XDKfMg>!zV9=bndkZS1S}x<|hk)4T>-3FNP?EpJJj7
z607oGHogyXMe%Sfe1h2|0*ins5{raLfC(THVnT>Sm<XaMED9nqCWc6YNg$G9Qix=j
z3?exuhe&}bAW~vVh*X#gq8Ka&qF5{zqJ`K(h~lw$h%|J~bn{N+&~qTlrRPGVrE4M5
z(RC2z(eoh6r{|k!Q}fl6H=no&h|niZOyYv6z<o)C-f&rh%Lqi*!t!s4(z0bf!|}$A
z)BmVFCqy+RhvpCU(MWiZFh~r1lW=Y{S7s)jm}j~sNO-U~7@`oLkXh!Dp+2Dy@qPFZ
zh53X*6wVBvX|{`CMnDwFjD$$Q6hI_o3Lz3PMG!?XqaYG<#B<FvNH`LRq#P+kGL8%)
zIY$nWf}?;)$x%Y2;;2jnULZs~49^d6j6qz8>(lV;FdxMrc91ESL_u&#FkSGsqta|E
zHae~+0=^`1iRlG#Z{1+;na9@6ME4x>929&Rc)Jku8Xhj7#ypL&ca;k?^~7jA5jQz`
zv-yIJe#U@{ArGKhP>`X;4NEI5NGpsb{#)Swyo=Qp#eZ~ukdc|5wqz-?6s#OUmYAFA
z=4Ph3nPG0GK}}t{mU#IgvMhxa$@&8uvaLPpKy{>G5+{+s%gIjSJjrFkWLI8rC_ij|
zc!d8%cF2q%Z^r-YrP1A{PV<<~^qlGA>leTYoHLK>K6lEj*}{o5R<tzBGD;aKwvZ!#
E0{3*pA^-pY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd_bias_relu_add_impl.hpp.660B8209627A1FBE.idx b/.cache/clangd/index/profile_conv_fwd_bias_relu_add_impl.hpp.660B8209627A1FBE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a6079734b51399bc900437f3d54b50e93a48aac
GIT binary patch
literal 6692
zcmds+c~leE8o+bMVM!PQNSFi!42zaU5RfepWRDU+Rx2V^M3R8Yq9C+ttAK1u6|~~o
z0zyR=tB7Jz1j|#T#T|EjJW-K~8kD-d^3*q>ow@fN&v}3Omvb=Z{^suAckj21kwFrP
zu@Q^qpBVIIiY#*>&SJ6jsDBxm%afZh7Ar)H#rkshiRdLuZaeT^hCMHIiqy>5(vrMH
z8o$J_WMj(GysvrdGfg&hZ!Sx$-TR_rPWi;9-ZsfLzx2k8;nu;Sp_snA50ZI{{r};h
zkx+B&MB#3|pSPI5KU7m;c&%ftqQdHJ@|7#^pRFF!=3bcoGPP$UMDTvpEBKq*8xChH
z3g7+O>+T%L!9KgY#^CjhyWUxk%m*8~7sQ{qc-rQL#k3O_Wk>tm;=0eeufE)HjUD%L
z@ztC<lQTyiZ<BtPm!N0SqOdj|=nC8(uRbY0A;`qTLrW1kviX;Rwp@upVhol9q&`O;
zU2?bFaPjl1XGKp!Y@W4gOpm@BsMFSPvZpU-#rVA~4bIbM&e4|`pSpAESVCw^vw4^H
zzH^5go;rKh?_YfVcJirPxXWgVBgbX%p#xWT$3$}4sN(4^^}_B-Ua!ZR!&4VDeHEwQ
zm~lPM^M#wBWn5O<EniR9rw04}bFH?cGj_o&jR{jkFRtHlf8ja3k#C~Gefz`fKYI$_
zjimnmepheSgS1ySxSjIsr^2PruIAj_U;RpX^yTHgvs*XJI$?6Xy3jiC%IS!km$&ap
z3c6%5P$V=_M4STWrl?u>_{v&}w&8Z`XGFH!g*)#QPCoog?%G!~x37}9Da&k<vo-7I
z8|lW{SVo1X9$y&49@cgnNJ*GGmB<yAv+<@b-jZ^Sy0}}82{~U??cf%LMAkhoaoel4
zT_~>BEzN)XbxH3FZ(i)M+xFVI;YW-!Cl5V*G`XmFVQaA6;2Z82R~(h?eT#-bA-*p8
z-KEXRF+a8i?r7f8S^s<1MIVlNczjCE{Febs!y2A7&007kZO6jhJIsZSrE1uZ8yllW
z?Q#O@?{&vEUD<2;?@{Y2QPD`|rk|20eqQ+aXt(q0`3=3;%buN+ZZ~Y5WS!l4?%P2V
zpQPpAZ3~v1;QJqTm$EXxN=c}qPOxCnvnO?TvDX-8s7WF9kd_?F@htXV;JatnqH|_(
z;RH*aI^wE_oS-%#ip9nN%SZ!|;M>m8%Eo?AA%^iee>dQ9cufB-ht41GeZUJKSOPYS
zv{SRQ8v9|f@o|6k?|v=@iCL_PT9}#^izL-av$3DQ=cguMEtknNv`A{~1s%TUr&;R&
zf_w`myakr@*y9|v>MG7AlyC?#(J9f5eoc@v2bl{bxsN;$l47#L7E+RHk|(4JFuel!
zL#_ngl^_7}9w6ESrbDg<p4Gqyat-jW0W)b<Ki~2AiOoR{PLQeAsUGxef=shY6G2M1
zNOy*`%yyZWCT)e0TXKW(Nem~4fyFSGL|5Qs8D?LGInb=zYML@_>&y$9)Jz7cR`hFX
zCSKZSkkWP1%^@x0FSCJ^X^@HZTftj_^vlx8Li%O%vzdOnqX#Bz65CW$stJn@i=l@j
zQ|oIw_G`LOuccAaIrM9O=i2h-zU^<hIGJdk$f%Jb_e4fKI4LueS<qc5a+SG5l3U0T
zDaa*qM%p;3FjkntmWn9~dq_!^Ns|#0Cox*Y$y`j3iy6V%b(mls%7vJq5asolU_E5P
z^(M{#)E;$G!^sLTwE{3o!O2SCSP7iqz*WGp3gv3xR1K!V`WoO=1KFte^fK0vSoTRt
zYl7==_O3~rc6@M`O@FDWUwoTyH;y2ifJGBvbVrbffW;w{n}J0$<cZuIQL*OPerm+y
zK>s)p(()8?Hl^na#+OsKCp0>gIfOVFC5;ND8<`zjp7dm`xB%c(Z|-rRM^~78A8s9S
zzsULMYPGmIcv~xQRu{G2d|;IiLg*PSAGDs~IM0v%Kr~ass)xvBwv8WDc$hx_ZHfBA
z=X%r);wI|S15tN|D?>=ml_R9!DiBKICLyE*q!I|=I7(ozL=UtQ_$vWpOI22h`a|01
zV9q(1E9{?#Ip?8VfSDIy$h|Md%!^Sj#ki#yvYt}Rp%g<dP$?!Y#TbF9te{yL#@Gr^
zTb94kQ4N*7&B1s%7-F;XFkT+Y`4}%B<syt%1ex}I7<<g;(VmVVTL94lV(I+_G-}=B
z4X>F)>7^f-^uv=UY0FnGeIl>WAV`U>grwV3B$O~#h$06GBP)u$C5&VV@;=bJ5BRX>
zePDSXFlLt^2Y}uHtPfr_u|4OU?=os7cYyvKz)TX7TK#T==BzVR#VufQ3ou?W)NOMb
zX;)H8$%-%K&!(L*K?<Y-gjRa3WDG1puJT&NP!v1xM2bvzBSw(zfZYys;9C2DULRar
z^sMop{az+kV-ut?+gOK|EkQclJ0pJ2#oh%WS9@24-0a;Da<_L!NMtWU$iv<Pp#y+@
z04_`}knNNQzAmI>(FdmV0mcO<FVDyddsFM9Mo<<UuK?O`E0qARM7av!RVY^jyc*>i
zfY(4yDPJt`xNZ87GA@4r;UB<^9yp`^a@*#j_^+v@v;v)0pifui?hi?s{r#mg)Kbm>
z(HR&^5M(#->;_W0Awm8MTz*Akliz^&HxLZ#dw{qH^1A9ZD?|64J~L)s4VnPsC{^aw
z;1J5qz@Qm&;j=(v(S;JDv5r1XKmunt1bh#H8IYTSZ!_c#(sApaZhG&oOOTFsj)+#A
z?3@sqW;YEXXOS}!PPvF&5ONi{BIG7=L&#m^j*v(sLdZkpfsmLhHh@d^;(8(E&Gkme
zhwFooAJ-2df3816B{k*$Ok4fvHl^hQ!0rHGw)1WG*8Du|xjv=kE@0UOm;jZYA8^KE
zVmcDv$}c2+zu|pD4b%>F+o2A&zK;JQSI77&HBdir=?A{_06+NsOKchuto=!a2vtji
zacM-sy|Jj*VHgdV$}pcY%#Ysvo?!#)2SXNrQkG>eu-Xgk=n8_Y1y;2v?*o?mfHkbI
z15@gNBjiya8HJZc<v&7mut->+hlS)}5s(Y8umUU^axoTBjD1dXO;OB_ffen>lyS|p
z&SWB6f?Q#>g0VUTnPrj1oE3u1w#{Y&{{z+!XU}^cF-0Ax{IR(bW;`kuDFKtRs5nvy
zn3dpUDK@bbV+@UoJxeh!dSxn?$t{B}_Mmr!*OA9fO_UWklp01uH?Z1kHFL9@;%+$Y
zDbX3PMWtEnM8rDE*fNCVY&k*-wgRCfb`nBLpr-`P{ZUDKY7WN4!33FyP0d5O05d6o
zdm_kU%%m9QQjA@SG2<V3yz&Tt{Yd>skK9_hk1iCXFJ@8K@PG5rV_eOPT`><rKdEcj
z)_ryEe$PJ!KDrn?q{Ze>&!u;&6J$_q&;t5sQmJCB6rteQV1&Zvg~h{`;q$@~ikKI{
z5PgC%&e+!k7`hqj8r#ybN6+%Qh|&8g8K0EC&guqSAJuu{@kwp!yv3@b-04v|D%0bW
zh<u5WO_wo#fFR|w<cP6WI4hVh1_YTDki-P71eLD14+F-|5>)Ol%!L^UL9WAu>rgJl
zgoP-t$As%qF2iP)!HndBc=GDVFyr!%erGN{>`Xi6ePxhx2x12@<G66r%fSmFZwGIL
zd>nib@^kb<GAVyYe}=BkSi~;bq+3AQweaZh82XxBGd-?$(^h!)19R8!e!>5BqX2yL
zT@E@;+*dPvP?MT5I4YRQ*oPvuj6wsFtEkrz(Gkr09xwiBwn*RaH08TO`Ju=+y(oSZ
zLyx!o^y+)t3l)^HnS;;a(0e1u7(51{x%ga!=Hc@YT7WM=C;?Ak=;>(Oo6ubaZYoP=
z7$k&)OAVz61set<6lNO6e8HjwZW@kIglPmrBfDEWrXDo2{OD4CmAg&)#=Z6-CC;Fz
zAV$i67LROH+TRMJ=Ia9HRKQAa2d7d;Gb($8d<W=JX(r@8U_)h8ko&;|Dj}wMlziKN
zbSvj1r6)elpF~?|f@}nujW8&Cd-D9`o#HQd|6g@d--MSgRDEsESiUGDPgUTle*SvM
zRrj=hj<wo9$|1PgI(p;vc?P;HE?vuF5wNHOi`uZL`j4HzA6>e@VvUDIeONSr#s2~&
C@6+G_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd_bias_relu_impl.hpp.0A6D3A36AE3BBBB6.idx b/.cache/clangd/index/profile_conv_fwd_bias_relu_impl.hpp.0A6D3A36AE3BBBB6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..257cfa256ef06354f91b0a41c26b5d9437dc0863
GIT binary patch
literal 6212
zcmdU!cUTnH7Qpw61F|qIAj>RGq^Yp<vUH@n2#82ASfjA0krJ8`m6zC9kVxcdl&Yd2
zp@<ry#)m;6`s{@u2@*?uKrB%aF+mVS-(BL&-1qr??;rniJ{azA&MoJjd*<A8hsa>5
z)JvDg3y2S1oxF5qEXm{Xw9s$H%Jda=1dnH@#^bHtbuxNU;w^jOKzM(Eo=tXRaC6wj
zGxL=zWd^sF6!>pD)Ds$iamIJ;vxo0ay_^|W{qo_%NR_Rtu6YlBylA48+qAX%=gvKM
zv)wSSDy45<pz>!Sh2;-kX-EF49oT>7QQA<~OLMm`mOczlT`*a?;!Mh5@#v-6>Eb=(
zyV}|<zTaYKe%1JuX<uzs^PTHX%dPHfEGh6<9C26eR;BHUm#41F#r5ayvVQgW_Z+*J
zfc!hghdys!xNU;-lnV=6L)KpWYWeJn$$Fj9X$6Atp1jD4NEOM!eX;8EuB^Y}8(QP)
zZFn|`{=wSND6cG7ea}EkuD(~_hR-Cc+dM00wREo9DJm>pdpdCauA4vqTz9#*j+85%
zoR@iQ(S7N$V>bsf^)if1l7kOy3#e>dvpc8qS#HDuY4ylACdLis)lGSmo^H9}7G|&^
z<7sK)_%7u$nJrsmJU==A%}Mgu-Qnk=&g(@zJ?e#q*@>STW>1{tvSsVI@v*I0qxS;V
zt#MU(b9IGhw`*2X?vf)vWU9CfJ=)n88q|BI?|NvKS&Xd*|NA>%!M$Cn;oIxVwWeO&
z^>vV~irEyK?SZ_Kf32?Az5c?=YZsFGFVr2^3(hGbs~Vym?P1I1y$NT0(huJ5@O2GU
zyKWH{Ex6g|wDoaC)*mh>)#{(0jSyCeH(5w_L~UBS<3MBCq@B^$RO#kg>zJHbfseme
znXIjpyZP$vK3VC5HzsXso7<f2Do)(1aSHPZ_3^%BT{!Kpi5{Q0b@}`n&@m45X54!H
zwY&aYc~jMQEqULpZR_1DcQ)Vq<GhVc5Az!LeDs&etR+E<<<mE0YBdNu^VfHbKh!>C
zy=mqjXE!8PUFuwRyDuhhrr$?Xv!0bc?Rup0XQ5;7=f~@hkflWu&*8_yBMXNrdTm5c
z_@AzsxI=Z#Gm+TFCO2bEa@;=Txw^7vPip@|3=)K%3PN<z=Bnlsc>KR1QdZ&-lCsi*
z;wdB7R$R1!l0i6+PXJGUJkXH3>Wv9ssV$e{>GFBBjgpP|*p0_0$NgPF9)0fPhgP=>
z)CeUt9!)FLMq{_P1E3zhIz6(tzRS~?AVjKfhY+d?Sq*sUOP6Ih36W%Rv)?r*OAVmt
zelX|<b7tsq4r;aKSK{(jC_2_L)|Gio(Q)>15{_iPvgsVfPl>nUs0gMNK>)|a;8P5N
z9G8N7DNN(I3_Q!gmto~=&5xgmzf~Pa(Fv9b9?WBkPP9mL=P1c6$(bSb<@&BG4j8r&
zB;5mMJusOOkhGkzl@s<1Yc?7sJMEm&s*L*9Nw8oZqrQcSQ#ne~NHXOpMJENTnJP@h
zYNlzVu{AYE4jUABi7Syw!=uCJG9A#L*6nXTs3Jj+L!v@rn8%{i>kE(kRy<0PbgXGC
zYkov-vFxNtI?g1{jA??%CC-f_nVAfmsZ1)H!O7wc<Bc$~j<@CL1?av2eU67fcL?VZ
z&>g|~HR!(PSocPQ%B$+*jwDGJfn5=>#wF=ua3}^xuHzDLD8ac59LvCok?XggP2)ZC
zTKd6Rz`&*Nv$DwrCGWgDMnCSaOSmPnL9Sa3X4SyjoT6*MtOn;=FstQwl3;^-p30Rj
z)MEo^H-Ir?BV*NqRJrbiLgd+D(P6CfkaSc?R2Wmq<V5->Pu6?o0Et=)8bFH?nED)T
zeD3y>>U%Gh`R|SARh>VMerLvZ;gyf*G^_e|1Rh5HmlBm@`!$h63t}}HjgSiq;xLj4
zWEjN@;xXz5x*K%4mfc|6jeA=+1at!%gcPqn`oZcCfb#&jaP0@dc@XDeFdgRN9!Zaa
z=_t<Ggdm&1^vNddvkA<*vx$&wVkXzWoG_6S?91wD#S73sK3<ii=m8K802?MKdJu$z
zIKKqpOPpVU@D<0NW${xt+TSZ8C|X2{Oqm^kqANh90;VwI2iB`~3;Wi^pq1!=i5;8^
zNt@(Plb$RqA5YOzO)1UPM`SEzTO+cUvbI9xBV{d1(Vd{x2_mj#CzyAFnB$wEb(52a
zW=%S^;i_L6(xeTv+c*`&T(_stHu?WVCJ_aRBDtC92I*q7(t9Nv$S8W1_bQf{{gdF2
z?VHSAN72at>fs|sp(y_JlgUdpw<3!j1^!XcU<4H14qEM;rqQz|Ontb`ONmd>hI~T}
zrhAHZ7CYN>jU-|TMlNC(j9kU87`chvFme~WW8@+Bz^DrNRh(kWbeEPcn?9I}Ox6z8
z?ZEDHE7GNB-sB#=iKe0#q&JH*S*CX;yQ8j1&&UiPs`f>8EdsI#)H&6Qfh@+k1jrJc
z%YZE7IJq!E*W;GaeH2I@fbjz`VLD>|g6dd4!lXGFb=RA=Z(i~_Qu{Dy90qNs-o}G7
zljrPiJC8Iv3GOF>T|tUI1)irMgsDi;%^+#UlY<uUYJpHjo?W?aZCK6O^J5WJryAHj
zO%YLbYH+Rvom!3=OA)GL2@>ZjhMI;J%tGfrn{Mddny)|ByKgl}nPC;Z`_(`Y$F<;B
z%kh?waoN2EZ`?E~+QHHRTUJL)M~s{-oiK8iID2s&N+c4DTqG_Sxk_9ya+A1W<Suc?
z$V1|Rk(a<rha16L;Ej=wzy~8=fiFh>0)LDG1OY7N?=QTQ@{d1mA&;m6aTPc)-R~6T
z(tckEzd(Ycpm~%t{LaB8i#BQ){xTMaBpu+#wAiEMN2HULH;|wS%$tBsE%*4}^)iYG
zRsZ1ffPG1!WJp9#_gLC#--Bm5a>7?m_%l<l=+m+M{gK%)%BgpQ#cr@+1QfjoEcW1B
z4(8=x$;tab+Q+>PioGOk03td0Aj}+u2#$v#d>EoR9)*ZeE;8=lJh$ZDs#Au@;ZrPA
z*ff)(Q!P^2m_yNNW@&70U03wwiS@!6%h0l|vslNz$cHWO&soqPVXaJ}P1>{@-#5F#
zy&F6k6%_KwZcZhV&L$>h6KrTvuxB>m&DEC^0y($uHZ1IPtjO1xpo-2T{8%is#qr}X
zlJR92#q;Aa>ISWDU|m{aL%RVk4OQ&Nc7r$%!^B}O_E9J?m^g}aHo?y(*#3__UVBXR
z<@37tPO=&Wf3)VLCh(A@|My6CLZI@~xw*fGeQ+;G(}iE=9rXP3-g}#-IV8`t_iE~C
zbZ!ftAH10H6?CGSAA(Wn{7{U-7lbe2YR+CT8>5H?5iDI!uZ<Y#NzVA-%zH&M@WzP7
z6ORv0@oQ#Tn>S92(onq9zOif`gT66Li(`W(MayQ%ut*p09M7J8DU|8E^#Hc@y#V7E
zTo9t@Aut}oc?66{aDENOuW^<WGvr({-|98xACcjPh41aSZTC;7obdVi0Sc2|_FnAn
zNz&f--Wd7V`(Wg2?~9SYgFklY2yh5s>6aiGKfgdT2L+7T(X;0=zTRQ<gVL{7##i2n
zBVBH-0fY6r@ZPg};5eyff8Vz%XoS$HQ1(>%C{j&7EO5m>^f)3qg4O47!l60t+Wu!z
z4jU#4!@d(miK19~{Pm$1yRBNwl_)xfj8SFgLeX={xfsnO=V7#fT!7JHaxqK2BMXPZ
z%5q#40YfjCJx-&5p%;Qts9q>W;YQ)?_Y`DOquCfm7)7x3d{<+$-M1#@?`_8y8;e4Q
z+^X**g9S$gvvz#>V7c@+m*^AdaVwa$f`qZgaNhH+Pi=38E0HLLGDTSwBOp;l6I7Su
zHqb(cFOJ(mj7}CDcYpypCNezQa&1bf*XmvWcZtxq!=zZnH^q!}S;i(sL0fVA`%-5L
zJf7m~;@eA;De!nYTv5OkDXyr&6&3#kcze+`d20sC!<=`hQW^qvtqIyf9ZlW@M#|&;
E7u0xqlK=n!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_fwd_impl.hpp.DB77593E3FC3FC2B.idx b/.cache/clangd/index/profile_conv_fwd_impl.hpp.DB77593E3FC3FC2B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..737b2c49296f45919647bbd1ff37f164c7e49cb6
GIT binary patch
literal 5714
zcmd^@d011&7QpX;047`#;6hlX2#B&Mg+xIJNf65_ghec}h+A+0UMslaf<~;`YEeKG
z>Vi@&j|zy%Lq$|T4U3?p72-x)H+&*)ND-e~-;G~y=JxY1|K@%lWPWpI=FB~F=FGVh
z6B-m0#>KF}1)-}lq|4?qFbp#x|FV{4E*__YVNOg8Tea_OTyom2&&`G;1A*QNA?0J=
zOg_R%OwpfjlU@^isP)>w*w<4F3QJUf2}8dfU4L2MF`_a4*D2?Qd)B14-MFkwO?WPi
zKInSKd+2}H{C0`(!PblKoEolou>P=)nm+aYtaWp%{<Zde%9DYV4Z3%S4!lyof`QcH
zzP#+0Ja=Mn#jdnzXKkeY6}Q^AU49d#=nRkea?bIoc1zT*dh0@P*FR@$Jhr?<)z?0+
zx^}C{RF4y?Wro?i#CztLr|&&*g;iZNn7!<$f$>QuvbgBAtl?)JUL~R(*rcU7>x+JS
ztGp2UzV#e*`FS_4z8Z8W`M>vW>ppBea;sfeKjMYK@6%t%-0GE$J3`-!B9o+^!h1@=
zHt0*Otd81X)b7LGv@;>xzIOiDtWyp_$C_e(Dqb|&_k1n?=JC3RVfOz%wJ6MftyQ%}
zt*2yXYi{Pvn2bi}l-Be7!<)=AuN(1e<Eq+u<DdHE&wlGNB;LU2t<SIirRjbC{9Sri
zCMcfNY-G!$WP0ZA>t+y~Nl*I9zkd7Ka>0yp%X8`6O@))9leRgV6rD&*mBa=21V2nX
zpLRQ(r9P0bP&KP+|FhYF%)CUtaC(|?QxQ*{pJ(`d+_j!3QfADpmlqRszv?WXW^wTD
z@Y;Q;D?^6M{A>39xu&p;KhNpN_v21D8(ulK&gb3eSN(n4<E#7^_Xr)=Pkv)L(bT7_
zz(p`IcaU2waQNb_wCv|)D{3yVYj1oey^&UoWo2i~6O(>pIsf_NirYHFIy#nmB$5wd
z#j-NRNH@|vZzqp^^8WS8mpEpjj}Z=f4%Q$3VN4ysbQwTErZhF{ktx|Sr6J%7&xHdx
z(F86{;6?S+^)##~?U=V+A18!PLKmtUCq&L7H$-AL@#lyly(0yPB)$?6qNsqVAd2*_
zH1t1qtzqflM36%ekE+IrV8>uSMMev)y%!&~ZPZ~93h-2bwi$!y2HS41qZ%65*=6vG
zLOL1bXzZXdR5dvoC)fc|7(0w63+IGuWftc*>{k3<{LYv`2yKL;sV*eB3f&QjY(-j~
z3_|2Cnuug#TQN;0_7u}*D+6m8*rK*FaF+qCuM7k-5F)t};*{_u;%1oH4DpCt!Lk*s
z5nlq!OLSHNR|S?xt^!XL-FhmBQ=#!KuN@!tw)~`*E`x~lh}147gOGSiwBEoVqCBIt
zWHl-vmAjxag>;$>vkV(bll4{QQtXjHIwH+zGx^!dy<>Nje6*1qME|LK;iF=Gf_<It
z*R`<tvuAuXmUFo)ztkt;9Vy^j@xv*@yY~(C6wkc6h^)>BUOr5s1QWR(OM^};0)Rp8
zZrdi%y0(Yp$4A_yZj~Qt@U<Jbl%_v-z=U*uw$PXwh;(<hh(=<zm_{;S%Yepm;L7O^
zC<l=o0?=4;m?4Lmh!x<Z0B6KX;45J=;%4AC)42usEp%=JejA;;fZau(PdBi;|H<y%
z^bM#6wwf+i!vr;WQ2PDcut4jR1}rkXD8Wn#+OUBW&0y9{=N2$)p>rFUwILS0j;s4-
zDZ79)dOSaV8Z~~P;BHsMjeZ4Ly$g7|Ae0gWo%!Zz`s376GD;o=lcQik32>qkOezsa
zco%%dov?kUE>4X1A0L96x%j)#$kpFfOVKwbgtV0R>XJ$efmeuJCT97maYfNynn`+F
zc~7Utis|}M_fdUz{D|D!vCJ>dW=M_m$c4xSW-f3jtKh_8Fgc7CBKcRJY{v;Vxuo%&
zy`251W()Ga-MrdmL)%B&aM9D6sr)y6L*&4z{;3gEdz>f%UI_$Jyf`y!h2&Mam|WjT
zMx+r{jS~`vgho+}C@p2|USjEc%kBZG@+8=w1ntQuW1c*g*bzENPSMlKQyW4yR!z@{
z+y6@osUi;~c`zF(*#XfzU=HG)5WN#|Zp<FxV6Q<I`3h2hX)J1147_6G%S9GjxB)q4
zIplnD!8jKkhegA4k~eb5;~Xa{U{VDN69=aLA+U=MHyV*1SGrZe8Cdc<i?oUyT;(WO
z9enH>b-`lzJSivxn=%w!4yA`P!lF0d{g5pQ6!d&T4UuYyqE@CXcpUd?lFuq~gx$c~
z4PjcrGjo^wj~rhc;6%)r7;QMliCD+j@l+$6XaJuE@JFnKP$fhnZicXCkRWb>nJtir
zxDDdlAerLBTjNW5m!GpG-6zZ;Oq(>}M7U$PHWMuWJ}1`v$(ST^$_2nHK*8#m%Y(SY
zfoOMq25Baq90@cWCI@Z!(xi}F1^rCw24OddscA3>HQ1@ap3Z)1^qR+$u6vbj=a?Fj
z*)UV6-DPAp%oNc`%oNi|222^ySPmv~FsD>#P7@C$%F8sTiAOV?Tfn1*&TZh)MrRe6
zsOaN$0kaF~tL<N1Yw_KSpFbh*sAKnaZvJ`+M%wwG9|Nb^qg2iD_ajE+UaYT}rRab4
z?l00RMZhaU8|F-F_pB<v!QM}>aO0LaN2Cv(?u<x7UNlu5srv0q54qPD`Y)t~j30pa
z18DQ1woTy%kzJ;m0Cf~Rj)FI}W7;z-qTk%h$QqHx>oA^n?G5L#?}&W)D~@ZnZWP9H
zG~w4vsQoe^!!KXtL%9b|h^L5Y2a^n@$v~U%;^YZ1O+jZRBr9Pq>fH>pn_(W}R<Lgc
zC&ZV){t}&4kgTHjSHny-Bv9=;1<q?@B(}RhAvmZQW;Gic*KZ*mKa-cKJ)5qczW4IM
zdVP%CHVgb0L{kcIVh`~4KrqE>yL!DFZuT7?1@AjPoq@xZmhcHPxqsUyJ?Wxu=|>4`
ze_{2@{cfG5<hRfi#uOH%8z+1jzBKY<_|Ztf5NPSezPbjFnlVnFFqv1Wj9$!deLFcB
zIXKAC61+aKr)i>JyXM0wCm<(<nhf;|tPgiy)J<8YG<~k-YAh>repZ8~pt=3}`{7Up
zjJiP6RG6V!^5&L$45$Kz8KI&vDjK5VC{!fAJ2$ZZ;MnYRj+r^vfH{hV8yTCJa@d9#
V3pHY+qULh=@5h`fU>J^y{|BS)6D<G$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_tensor_rearrange.cpp.50518BDD3D0D27F8.idx b/.cache/clangd/index/profile_conv_tensor_rearrange.cpp.50518BDD3D0D27F8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e1a3dbb47373cf25d8244efd7e15f6aef019eb33
GIT binary patch
literal 7144
zcmZwL3p`cX8UXP9Er(7zq6Z3<c5S<k*fjL!kVhVwZVzQj_VmzWdQ6j&%Jf98BqF64
zgCt2QR20ROjLIu2H9bTz6Eo>;Ou3u6+y3Tv@A>^K`+vT@)>-SUy}z~gndR#2{M3-5
z9G1EUtnuBjNJddqKjF`B!}^sE)F^74CPf9rXL_!g-7rq)ed4caNB8yq*OnWKVa7V8
zF}oGdL-r4Kie2Ne^yh^?+x_{nW7K(W_%GfIs^c!N*pypQc&2}HP^oV4cQjjcZj$NV
zdt5I%>E_RW)YU#}zuxNUZ#g8f#lrSVQreq@K6l!y>iae{2JC6mTkQW~(BMsn=IK8<
z{oTj+H$2B}znHkJ=}leTpztx5;%;ZRx3)LD{anAw|FdrVu-<D;qQA6oH=UP}^P1N8
zQKjwMIy1hv>8-L=iEi`8m+uXpSvTvfr^&*c2aW0Ht8?=H-0}H_^**1hr?o@nT9>98
zUN;QwND2BO$nxX$ck}CZM4!tuFAWVIYkS~UL_=rxGtUz_W+U1R2EE&^G`dGCS7xYf
z_SKBfOi>@TMy>vbb3T(Fo_(G+gSVaic6DaXPf5<ti#5!)MrB*%rMB3ejQ?wI-*d^{
zy2;iy(@z`Boh1wPWA=UB@5iRCNvopNr+&=s^Q6AUG*FR0xAN7_1%r(`?1P$hJTf<}
ze4X^WN6B~cKjO#iTOc1$H~Dbt8dtC7^Kw_&Y%Jb)ZAHcjy>PcRhqsURe>=u&ueZAk
z<5J{n__8MD?t`s!f;2Oye|c6ue~{MoArlvCo1ZeMePcG&`E;^F&UJr}TiU-?Pg<Ja
zU`<zyU;anB&Whw8ipos4X&fl8j|nlHBA<Fw@4FXsQr50Bi`_Ei_hW_Y!%KdQUZNTC
z%)rZU(T9aox6f(4Yo59F>|6SNN%N!&z2fwPlbsd}3`>8PJZJZh@77tp&&%3;sBQSs
zZ`aB^$7^KPSn@Gm)qbIV1EXth==y1tWvCYTGFxu`YV9_R<j386o%?2j>@KtSc$iM6
zlf}zUkAm0Mx7|vW^NFKTMOt}EkVU`if&RfyypC-f;Gl{aW)>dp9;$I=<?zP-T752r
zpWZlOvP&sb6Quc(a=7h%GxEieit4+5-e&BL(Iq9ZzkUqZU{Rdx{Jebe^=<CW*6sIW
z|0?`;%FRy2*~}?AU#*T9vv%N(!MhfP-1uhXQ4`lHwb7Hmx^pHarOH0Pa$M$y5x)J0
z9shmfkw9H{YxTY@W=%Gq<zIbV{#TT{NqJ80_ZzY&ov)FN|Mq%W?xL3gTaI{#84q3J
zr|9)6S*_~x+h&WO#^)7Y+I(bWo8?QN9P4K(FBKzod}bWjQ}pNe+23o--|*JT<HN$w
zh9xt0S6ChSX3p@RIv#x5wDGTDnd@E{)Y`}_W_{E0OYKpYyJoXfQ)BK=Ef3IL5g2N+
zevyr?LvkaxNB`roNZ0A=A->u{nYsysFXm{+I5@i1Y_;=Rtt#2N*VXkgv&wkN`?da&
zR^hsbJj<T?%<UN4zS1)K3K;4iz2did&EoSwQF_&$v5lM7Ue%@+%%*rfnO^ViLdXH#
zFQZq+1=Xu|)oW8<c~bmP*-)+SLJSX0EmFu^JJq$R8A5o7Y=~BOA%>N!J>v&EMZVOc
zmQfVnMyNJo_}_n2SE><SUwH4>u>YNtw>sr5oT8=B((f+*4pP+QSGrq=6(?NKro4pi
z>B@9_cNb#VGx|c+3roZAwW)c+=Nrll`*as#c;Wk{Wfid-Hz=s(LgOP9BaONXG3;*<
zz30#j4?A^gmJlAG7@*f(h~dDjQ2TY3x{tJ}`3ostTc+KwyAZ>P>t4UOy!NiTf?6Te
zo+O(zyt@#?si~Xh-hcfdSe=?9ga<1I59}_)@OA1@M|HnXi?yk_LeGt4Mt!;qG5l_p
zT7-7Q6f<pVrtsN1GM#?ig&1D%yUcHvE=6fsgdX|kpyfj~1;5Z}A>Gp@-p9xXi8tbV
zq_x~&{K^cYt03d8<W`#ESELuX&wPhr*#rw2zm48DLTpLK?*e)kAcp`Qf*l5Q7<LrU
zQP>B8K8PI$bR2d9&<WVbfIfzO0_YRir+_|%odt9jc0SPg*o8nBViyBlj9mhB3HEKE
zZ(~;hU4>l>bS-ur&~?~PfqshJ2y`QMGtkZ0EkL(mw*%dd{R!w#*gFZmlc4zsCUh`%
zIHAL_BMBXe9Yg3C?8AgUjGaj6MC=qor(mZMIt@F6&>7gdgwDmzBXk~i0ig@9iwIqW
zT}tRu>|2Dsg<VPLO6(d!*I+*+^h4|?gnokEK<EbSCPFt!7So}cW_KDCGlYWA0h1gs
z6>kd#e}<TxAwAYV5wSDyq<1(dcq=2TDdJ5fW=}26?yM=M1O*=oTqsPCT008dQS8&e
zoyNWf+%@cbz}>_C4Y=R1cN1<mLG8y9E*3kTaOv1r2zLd$nsC+F&k6TjvY7RSVRfZ^
zX;?wP1>tTS0`3s@N#IUmUj^<e_8s8vV7~zF1$HNJo!AEmcR=cvf=?w}D)uG9UBWIW
zTsd|<;p!zDhd8*>na|hC75r0RpMse*KaIdPVmAZZjNJlk3-&|8J|w9A6T&{hZXj#}
zb`xQnB#X-?5OJ~mZ{8OdOPS_B7D)x~YUYY&+h5<mC*3RfHZW=fRve4fhn(h)0bzH9
zD|MY>ow=CyCk<^YTcxq6O-sg)l8@?H6=eKq`Di4iT!~~VH$@sFAA@8jH$yU)n<H^@
z4#`4pfn+JSL>em}i}Z#33#3-kx0UFK<C^4pa{G?>GYf=q%^p2_qWHGBe-Q5HuFZpV
zKL-pCFcD8u@NZ!F8|blP^xNi{>jyj*E`(#i9>WuR0@xGSr+__$ods+bb|PUDrHNJW
zDTGbIP9tm@b_QWHBs&dk?Uyt_ZHv&6C}5*tjMU?Uz#hbo12zsj0oVlWaKeTYR6mli
zk=QYWjln)l*u#>YI<NgQWXP;)p(FXg=Ho<I2y7vCF|fthCBT+o=MpwoN}LKlkFa^z
z1%xfYE+TA^WOrh2YGNK#I!M89v)Se*eIL8PW)~obfGPy0OYtyJg<(g5Z4|gm@q=J{
z5IYX+;=n_SCxBf7_AyW$!|j~_)d}oVV1Ej}l<H@JeHL~;*yQ8>7J^M7b}^`maeF18
zD#5-DwzqMAtH8DjyB6$faewQ;t`7StsGj2X8bQ^F-3<24xW6r6-y&JucMB^(*QQ%q
z0RokQmeQd2gK|G0CxbE>`ywbWV&4SiP3*^@e2o18lpnCeh%$_z_7jLQ0Xv&0v$0Ev
zvIM)1DC;D9**zFk;a`|9?2P|@R1x#A*SXwEo5OB@2v`3$c3aP)uMqae>@Iw5hX5Ob
z9R_R|b~~``*q?y?guRonI|*t(n6SZ;7sg*JsBaw*_CNg-w;O7j|9HGn@G7}#kTe@J
z<THB22gHB-b;0EZAzjYXoo6JT#HXj9&zLoFxiC_PDGqkx+X~*%)^VyNCle>7Bxe(6
zq-iG8kX%e$kX%)+4pKcg6Stn2uZMlX&a=lC3;prx=cO;6w5)$n-Me0uEa{pOqc$=^
zoRTgxZDfpSF|liC{^Jo&=(w5Vc(FAFAD|mBOi~UQ<iJSDZ$S4Ap!gYLaEA26#m7OS
z+5H$OuEh%8f24oUH5e2bb6&gXPJ%Fwzpo+jiEYm-C!>iK;RZrruI^kz@k95zsRgy}
z@0=jKy-aV}5OFy^VwscfzuPiPn1|cI-o}Z%3fL;_T3~Cj>wvApE+uR!L5cqsVQ*nq
z61Eb%hOjk~#eNEd?s~!%N7FSr^EToxe*H<S-DPewL!JrqWTmpQ7cWQ|-%4y-NzZj8
z=83VI|9BP_E;cg<Yw@ELe4ugQC`mhj*@2hTAYg*9LxBm!-Vet60o9KNCK|^N0dolZ
zC@@E{lfgI{*FO%-aU4Gh%t`Fiz?{au2*ww2{mZ~y#__AbT*bZy%r)$rV0;tTuK=b3
z$L|1h2m2l{_pl#>@nc;788FXq`~@&Cuzv&QH|!5!`~lbR0Hy=SJAvuM-c6X@1nqUh
zh;bM}-+u&QB5?cwVGdx&5+)Wqffy&?`bmUI!tqqXq++KNCLKGQ7-!@97YK6!$1f4)
z6806sT){3O#wECZ8DYwByqqxQ*wuuo#;zmAb-4Z`!aTz9dcxFWKPSv{$!ERqk4+BO
z?oSlFNpBM^@lH|*cT(?Gq9xuI*|B!UI@Q$$|FafcbUqkx+OFxoT*f=tJ2;8w%Xmiz
z$7zzBOr1DM&Zf>t(@du!xtO{jxjMSKNcG%I-FjkI&OR!4h@ygpomo5BwBx<ZC$Ra1
zy_2YR60|o9CaPfUaAF%y(B3SP*hXT<5W5(H_GX8P-C^uRqDsW=r4UsLb{eryBWQ1y
zLF_ZIbBRqZaTABxwMDhb!!97I0^D8^Q59jA65CST-&@4?7Ir1EtHk}SA$B#`4~gm_
zZtn?EJ;81u_6@kdO~k%QviQrQus7@aSy9Zn3VsK0J7AnNtZ3k(v5y0H9Q!hGm$55=
ztH6E++%xPB;5x7)2p2)n=Oqy?3Ht)!E?}1tt_=GT;T}mYAF(xbm+fqA;l{9<X624=
z4Chv9cEle}g$1kYS5s+~ZiCTnQtv**+~~D=Wc|SI`v<RN>wQ-*>RRRe)-Um!-F5JF
zSJC}a;~VB@PQ3c>-7htkEM2_F=WEJG3{li#>2RrZxI{YaZuM_2S+jKYO5t$&@-Nk7
z>KeV|z56J%`srx*b?fi!II6#<+kim>bq5<vnrvgMva_FRp=UN_<OHSh|9e;8Xvomv
m^azF>Z92x>a_kq@<0lRqXUJJO>FJxexcic6|8j5z59&Y1g0+7D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_conv_tensor_rearrange_impl.hpp.ECD3F7A8AA337DB3.idx b/.cache/clangd/index/profile_conv_tensor_rearrange_impl.hpp.ECD3F7A8AA337DB3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b910d870981f257576f6fff90f9e48d81e957878
GIT binary patch
literal 5620
zcmai&dtA-g7QlaNcS@&zPU@UG)#=^QPonqZ&<jOEBIT9H6t_~`DP|hFRHqsfdO0bE
zP+{sKZZQ)x%#2q?Q(+!BYM5J1lrb10Tz8-5v(Ns`#~=FC@3X#pt>1pEz4qGcS5)}4
zX#=JV6OtI7nYLo>ask6I#`rH|ZF=$_h+!Iy7$$S~_w!e+QoBo@Mn3xZXllvqRnL4!
z+y3<RV)F+J4L4+Y6#h%RW#OqQWsUh|Z-T;4iZ4AHvWu9gpO#ge)!rO7I72vFw7RG!
z(Q@jn+CMnoyUj~kWP8L>8gj8g?z6Rb%hJq$UN3H4lDp69+YEcdC;4s1-EG!)<`(?W
zu}b#n&ZnlCiz<DGKFF`Cu8dQ5e0Bz63JZlxzPoVe#W&ykyL^}De9>!8YO9HhLH?%K
zUxii1JN3qU&idD%A>LD3I}Iny&gxFE@Yos>pgVfs;d{?xRQo!d*3Op{^h`|8StOa@
zv8=8-AYg6I!cz~oNZ-^=Ek7DNW@zxyi*-F0EmGzkHP4+b*{&Cp*`nApY4@$^qlZpj
z5a0Q7iMsAc`o+>`ve-GYm(7n;#m(lwT5P=JqZ{$+&!5|EWh(1~L3OtM%76TJv}XVM
z<rN7JlH%q$o>`iz(|7p#VB5TFAI|!~Ep_BNx75~*ryb@!OP`n6pM2W2=zMTfSA#5N
z!iV1978G6kb=QhCt6ioio%S21hjh$((-HVl_r!)=lb}V78|H-l{NqTp*qps>Ga}sE
zU`gEnilY@PCyW{VEZcH(g=AfJ^Q)&iLrpmwIETE%mYl?02csJ;N}tQ3XX$?b+pgp?
zyNb{5jcJH2bT93mY9QZvtNqHX@4}`hW_{%`W6qiO%b{lSe2bam28J3J@4Nriv6+LD
z;{LbA-*h0S>iwM!-rc&=<KKLd>h<x1MN`64O`{h-lkfjFeUo15xr)Zp>XfeTBOwMC
z_f5Lt;HQ$$zihBN?cBU8$!V91x|^onxE?A?dGXpeaA@uA>hqn;TU*1+CVOt1To?D`
z#;wDRXZl)gde`l__gGQv)pEpb!w+{GQr!K2JAAN8C(P#;j|Jz{uRU(YMtxD#Rn&MU
ztD^SGq^Ad74&4X{oNHO_=^0+;n9z_pvbXPxNJ+JQ{?Y@HRZoL&8XY$enXRjKRr=01
z`Nd&so5E}T7QyeFRGh#n_T>5`r5f$Jv-;uv(PN(*)?U7Cm+?{Bdp$V*%@h6@sJnv(
z5poq{#6O&+fu#u}{0qDXoBCdg3$L@xcnQO0>Sl7o4@M|NjIjxD5MF7tF3iaE6w8=m
zgLpRHYWTqj4Rx7O3^rI4jK;b}$*WmLiVb|EKJwuQBV_d$BaJ~z2D|tC4c)6OGY%U>
zuo2SX2P4$i`KyQCi>!B8dPJ!q8i@_8^{ht>KN#T%z>~y1^uY?n@*%%YmNCNy_Gb1D
z!w*JiAYeoq115dN%8cJfVe4gQC%1Vy8v)CeAm<X~N!F+1VOY1f^}VeEma7AK9XJyO
z%hiKRJ-AbDgo%w1Ot}d{n;@L9UfbFJfw9L7@CXgy)Bt`&G4dale#!r~`WgwiJY<%K
z%!xw4)gY@HWJ7pVvrU>;`Sf<A>A|D{MiT|g9Ru+(a3XBd_ufFD!;3OJ)K&1h3bTm9
z{O9BPuMU0giHB+g&qnYf3IW#y-c8^`*g}8uO0&en%Ww~^AZY~|QOJBBe)#F=;+=2#
z$ySsz;S>GVyVy_eVs|^O=ebaEsGWAJxma<mopxWjHR3gPBu?xKLc$%6^sL4`CwnF<
zNVEN^^rV68fC7Pla~HWAlYRwUs3_E!^ef<EMX|=DUjes9w8ofp;1Jx@{8-Up@Xj^M
zyPGc=crT59YiZ8sy25;)uWiA8&jGg_@Fvb+xqNWTr#$Z7)7zEvuP0%{P2x?a)T|$*
z{a``Z)vWdXX{Q<|g8;{xRSB{t3b)9b`)2RE9DZvC!~hJ24;eX{zJ+a!>rWrOkL`DW
z=ML~Cb`xzLeYZ+)`J<5n9JHc37)2BU9JHc(KC3~b=Ccx+DfuC&M8QflnZB<?3zcXw
zWfk&PA??5dt{i!n^SJ_fSMa$SiK-EAJ~c>G^LHLsgIsC<wMbM;_5LnQh~?)a40Kpd
zg(NEEKnz(tu%w*N6-ZJ+IV9lMA0lq{H)2;+f}|40k{3g_|JgS;UEGBgiQ^N!NYqS~
zJTTrb)>)666QU9pXq(@hzin+DbqzNsc_#Ui=4oH%|DHOqaxZSqM#kAlN_r;-$Y416
zk&R=yLXZ`L3o)CCTwR=8^IGt-<bLu{667qW@Kc16AZNKqzsOJ$<a2M1pML33j}DGC
zC5V;4i?#uf4$x?uC!G3K+KN#{*sv3XonS-;Ww}>i{0g*lUHG399f6fI9%`06KEZ=b
zoeZyIFr1$05)Od=0sgRuK>rZsI8}e~>egkwxVJ7)bb&t^Htw3@BggT#%&;OWIBUAL
z`M-G!Ppn*1hNqbgY&K}KAj{<en@2gZ^pk@1&e@mWIz1~SCo*l;4IN~$6V2qd)QQeA
zXE`N>nZlN=iRA)ifpSWbW|6kU!7LXoi<VPLFiWr{KEWRHPY9)yXqL#+CCQTHRF^a<
zDV$QWELp2d`u*4x@0a(U;wA3|$zG5XU$9&Y1hv3y!pZ3wA4EPsGzl9vf}jyb5(SIX
zwxEe}T4_$X+qF}>htIvN4z!s?lLyAu!8r1wCJ$KEgZ66Cutz=rK2U?D8my?j8pf%C
zcP`ZspoSo-SD_dcT0nURGT(tLDOV!%N<LR1vnqa`sX`uAh<6`Vp_nS#U)qio^NDJk
zUcBj>k=<rwPdd&#cei`J;Oh56mfH%#tso)_mfHrxZG0{QVG*B;L0HV^5)hW~xfFz@
zd@ciF8K1X<a66w>AXHJ_kXY#VOY(#Yyk|Ke&7rPixm_UHMbG}m#Gn0t_$@De*f1go
zv_90_E)n^ZbA+Ea#I}Y%!3SqESZs!wWU0AFW~IgKxq1or&<DPKFqJ4+?j9)aK^S3W
zR7KE}o)H<?unHtqGzqfYB?!6%(Zui*b@l4{;QbMJsJ~w<TW09wJt%&o7{zk&T)Zod
z>-d295FRavUrK4CXd^E>WrKA#Kh$$UluOfw=CoVq@%39lycM*8q}f*SHa-`DxQNfi
zATH)}35ZMhTnge+K9_;GjL+Layq(V~5UVH`38&7KW;iObXBMyv97uX7-LkEE&IXmv
zTMjFierz^*qhuov+)fzPNl#SyV0>Jzk=2j*jW^))2DDe~E{~?V50oCqup((e(h}mv
zy_*A!9;FOr;ASNntE5-tzJalG&PfNq#fq08dkNYraewMe!Hl^DzYf=x_Ib#W?5=>T
zMQ*jogK+gjQ|IpcjzhXEcM|MR(mImmijkri1rkHer5mn7+PtK>bmy0&SgNl;3oB3@
z;o8j$clNA1V}+-74DhdlC=O?y)}519Jq;_YMb>7djz4_X*rQ%L_YH1dX|&RjEWN?G
zHzxMc+;O-BR3a}W(pIS&o}onA@>WwY2i2fSWEjmtt!iljeSFCc&+4s4COBEtfv}F|
zO-(H?tmm^Dglax3k+Bl-qF9C8R5V>_9N|{Z=L+Oj!RIPuT*ZIC8VReBHjkg^Uw=Yc
z__+R^HEOfyUVFhmQyA=_zkjBju3OaD|NQkETtVc6G@qYtc7kLlO*bre0fH_-4Dr#I
zEk7<g;{T-QokO2J-q{vX%H-n_oCBQ0>4h2?5ExFc(CC2ZaB?cnn{*sF()s0WoNxAl
zWFP;7<uU|ahDb6A@&YcCG@p62#yAbmr)hoN=D1;VZqPiV;r;WhqxG(4|0dJ}FHqGG
zss?RwfwKy7&O_tLc$z?tQ=#RQcObhRh<5>3BD+dHSE0BnzI`q7t3}$HraizZD=N}z
z$2*<aG5U(fH)jlHU?&zrXdz6aGq?}Z`?PAqw+52s&`*Qwr~dn_cat82UAYG&dteNG
z@gf9Wgeby4*__tB>0;OVR)pMDw1*5HHivi4vHOl9b;)_1y>As*<e1{G5Y4wX*(DMQ
zc=I%0*rb`|YCyjRBxGX!DMw=_h^Jh@6I*4pN_)8bw;z4B$EAHQ-cx(VUXOIfa!&eA
zB1$g$E<EyKeDr9@P1K(#q7<wj%%f@$RD%JT1IyKdpq9@EL2!`r<K4|?-HzGI-|3?#
z5v}VA^l!0vF;Q%k_8NG;QqaTrZ@PkAr$i1)8bB`|2!oBIj~8J3F0k(cH!`E4lNH~%
z`d`s}m)wBKH((i2y#Bh)X;(nzZawT5XZt|v2?u8f9{ITV1k<(qy7=;lpF{ud^z;>}
z%QgST$w*Jgc&HJqG>^Z2WJF*XODZ~1(UFSQR2-uf5qHK=UHfn8{`KN9&Msrktt~CA
tWOnxQ(N2yIwl<85){wEM;#exmsc23`TPoU7(SnLLRCJ)?Xev5W@&ErGu>Jr5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_elementwise_layernorm_impl.hpp.228FBBB77B1C2212.idx b/.cache/clangd/index/profile_elementwise_layernorm_impl.hpp.228FBBB77B1C2212.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e624bdc5bf23a8cd94dad739fa830fe5e659e71c
GIT binary patch
literal 7016
zcmdU!d011&7Ql0^mnGo_;zf20JBU$0qJV@&*(xYIwxBJGvM3N#kS8wHx)4yPl!6<D
zO569<s#L430!l@wTPszO&DOZqS~LhG)k14a-;I6u&b;TJ{-b;$-~G**nLBgl%sFQ!
zap7TMsa8B*aC&%=JY!`V$>Z_HGyerE^OsH~c)U729<QkO>xAohR(7Hr(Rb{QSuc&x
zc6p$av~wUt9%$>nI%xOAwUy2MFS;VDH@KHAYnV7APV&Z&s+r@m?o%<3-II0i?Ni@@
zue=8hpA06?U8(kTR4%c(vazQArekcMOZS8)pKd+zLtH?@mWGXbKZQNGUY1p#R=(y`
zrfJ;3)Zb5Cm6J8gZXVth+WJBG)|(M`C&-_y>X%)0@{Rj?O4#Q*A04URLribZ+UP(1
zx?_pnq95%J8J#iubb{(V?;j@VzcuJpb+;w>?SPnywTrA}kBjvu#$@zc*6udAw8_I)
z_<N1{N1^qCAo+nU>m8(v(`B=2LW6Yodc}wY{dBEvbDN%{_q=|=`eT{lSIT0}Z!eE}
z*R^xMi$(7dhl=lOFWm0fLfz;K9BhqUvvvCcx7gQ*e{+!>3)wJ9b9Bi!22q*nWV3bQ
zbJiqg-4EJS*81p!)bgB9czqwv?<kvb{6610qvqRl1{KY*y+vMCviVb{?>Z>(NI9ok
zEPr$>%JFQ+w7z@BgSL7F%j;Lad3UKKW^tQSl}F>j@|_J^t1R|is=jc<&t!%7(a*}w
zrWw6gKdZ#$&c@^69<Cq8U)ysuNyjd9iO1oM?yf$oGV27r_B9Q&Ec$Ab4sQSD=JD%)
zM{jkT)mZ0$GIL$X#0P(eW;Pa%_^wTTI{VirO}8u(CP`!`&gl+hx7_}7@|ocF;3iAz
zh}8M^j=*`B1F3!L_DbWhX;X768!jd&M$SI{@ucXDhIc>PcI%ecm3MgstL2N*80*p9
z|8c7KnWq&L&xFsT-F4huo_+E7<T##=F3@1Ucz9s{ZDNEVO!b-i%pYwmFs6C@e_-~h
z3+w7m9o<RsCNS+ybxh5keew7N@XYj>_PqRz%mN#8rglo}>`PKtBcNy%SgF8~ef79$
z2K)DREGidKboS(Icea?Kb0+6tl<So%Lo&Hk?t>_AavoN*5hRV^fuxNP-Uv~Mn?TZp
z%bOs)375BkOAAav^;;mc1tJi4fqfUaAnpeHZk&6-xd$Xj-V08>Fd1<l*!Q9OY&i&?
zX`a=$?xj}Mz~nCkBt<8Sld;*Qh*L01wNI6@Jy3L-U79CL!YoVgr3bA~14)+<o+ZR|
zR$ye_9puq<GlwMUE{N<xGbib8i0sC>2j=&{>uf`k?uGbXNMKnM)oLdn*zCXzF;Sd|
z4Jt{TgwX=~1=tW5+AZXUC~6kVr90+arkD-{i4)m=DLP4zgwX=S1$IcbP`J<*QL-Qz
ztC=E5!6?-*6>FC!OvBox3(~Qg8G;OqG7U4ac3HwKuAN!?J3GI;x%)2zk}d_yQm|$R
zqGiug;B>6zpi&4y@*7}!1I%%j-hdg1hah4I<{=)2h+&-7kf6qwLIV*RTs{JpBlvhm
zz<UH{vi+I&9nt@DU!$jhq_dr~Iol=a9OoR2a@}&dg(GSC6!}!P1Cq{j&ckYU!OSl3
zMbd7V*^P4#`1e2nlJ|mNFZd(wgPDD(zGc_h6}*Rj8LymPt-OX-ec)8F=Q*clE2vmr
zY)<cFc3-$}I5%<oz^|G$K34=U={h*xYd$ZSo}BR9bljsqd1$&q((s(XSFHXnM>kDP
zYc1Ze6hvr=m0+sGEnW$MN|=qy6P2(4@ep_n0p}Y?dKf&0aaMze8fOi7XmB0{hf#1s
z{f$EKD1;&&1MwJeTbZPb39Dj)+xA-aDJC$JRuHs;n1I?V2&sa=cC3O(R1nzsO9-bD
z0^1_x1YM5W%Ot#D%SJt|KOX|o5Qy21DQ2RgVVu<<Qsb-vkp?jvaS6e*NXy0ximnFz
zYOrLx4QnYL%pJ(wt54Au?iMrHRunB3ij5K3jJGjE6y_GjxiE?j3k>7b$52cd7YQi3
z5-cjghJ6{uL=s^a7_%I)r?ukJ<^dsNE~}~4<Jsa!@AsFPPAz|rDfaU7dW|iP>;7ci
z!;`BMU%A~M&ky}to*~-K%x4?$w}BBmql75hrN7cohtDYA!NL_OSUOdT#lj$d5Z8pF
zgZaT2h44c#ssv#rQZGd}f=MG-v9F`(CNOEjxdq0zfEg=Ke7bwu{ga!gzi`BVt6Ei&
z*O79cnQ$5SlmVBJEzK|ZSM=z<SvpJ(M`jDz7c)5=nS)U-nTwH}lw*`f=5e&5DB!a7
zcUS*n1{@>_;!-h^4weLC6e0=1s1lqjQCd}S;A+eJrKxW-L-jHAq1h1?eRJpD8nW|y
zrnrz_Xvh|C`h1>zeqHBzX8eV2g`B%)=(R%3>Q95|Y2;!kx(!U)P$Emw?ZAAv+@UOD
z+?GcbfBOh1S~^|I>58Idq|AVQGe!H6ei-?a{ul*`g1A*=EJPHHQHUr6qeif8M6=gg
zgl!YfEnwXO4(wZZn3b5$UO!<y6P~t#XdBpbf|1m>HW^v>F#73)fKHV5*49bxu4^y!
zeqkH>FzA+_ICie^mD}xmOZ|-0*?$x;s!_sZCAhIVt3Qy{-rZK%$p~in%?M<RKgyj;
zM#L)aG8!9&xuX!t3P^ekCXJyJfi^*(Q$lbj43dc(p5-WE|9Hs6_1;5=KMW{(xyf?Q
z!&CGMlNA`{Pt50*k)jK%3b<HD(fwf4j~l8ATvYfGM+GxfC>_&=sDL2|W(TYdQ31m^
zt6`2BO^Bj35THTwf0rii`ML0vH52R7#p&*-n+$OVMw#}Rm@~|>%i@CnVb}Zf7ph~Y
zFsf66REg)wN{CS6d9Q+aO+kdSuh6QJ`EXJDaDGYl9{+^L%+gq!T2Epf@sT&Ld461O
zL^B7EQs9>Y=XkVBJ6;JUY=2r!NQM9#^e{+<aaMywjk5+M8k|SKWCScwf1@BCMca*H
zOu=FdxGaQXvJlf^0$WuDAyg1t1f>|$5SAl8rdo5%Y@6o5OPWtMh3zOhH0zgZgmw$~
zkK);1F(EA`yjZ<78$P(KSf0&e*6DvQ2TmC3Upbd}FY=Z17A+Tgl0KLH`twV=GVLe$
z`yZXUC14a@BrS?%*WsIx!F1|LTQRd>Vn?wPyI)A!W|9q;#+<2g`nK=RW=p2+E(q8K
zF>LGeS356WeE9hE3&)7_4{F}2>-Xwp!jpn<QlPB*V*dWvr$5OHUO98<5QX$TH5%}J
z<pzGqVQuNAfOsP&i7ADMQgk~+(KkSL1GsRfb(L{L=!!xcG~$MFR>PZWNM(&h8#LlH
zxO^1iN8t@59|MmubfrQuem$myNI>#(!nd5@?ieonIj@V0w%+-YZn)0-?aNLCcK^cc
zESZ%IyEyW*^5dvid{XBJiEYPoZt&XU+PEdVx0Sr`&bs4sIe4VO99~i_sj?5vpJ^O)
zmQg^sQ8<^bs;UnCRwuc<mx<T&$aw-}0+D1SMp0xGM$u$6MloazMzLfpMsZ{uM)71k
zMpd9+1>8A7d&;(|Mk&6Q-KudTt%ljvz#VNhwau;WU)VXmr0RZbnH!xJyK)6VG4WY6
zif(JPd5C^7!QJMJW|BYiWE(Cr4p|BIN|Z|e{)itqKH-RhS+Fv&D?^#jAIEpKxydem
z$q0(UqZs*(zfNCruJJ3X{l6t<ekaJ){vg3C$X{HrO8Y=*zyA3Xu@C?J<n>=NMx_Gu
jXn-Dt=#iedD#f<J&~^>qSU?F4jYQ)Oc!o%9gdYC`S+&3}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm.cpp.1660D1BAA0C70B61.idx b/.cache/clangd/index/profile_gemm.cpp.1660D1BAA0C70B61.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ed68824fd5eb8ef4eaf78fa56158b51218670294
GIT binary patch
literal 5226
zcmZwK3piA19{}+4=Sb<8k;^bH2{E~sirhyNqD4kSk)xY#v@P9q*?hZQR9M7X?M@}W
zrb3jgrdB@niP%!BNb*TYidHIZ+WmUIGgGhg_Ut#$<2AqYpZ9%d-pl{IM;zeiw?da7
zM5_XlHpFgTP7(w$iTT^KIdR=hRf2dmfgqBkXC)8CJ#(~LZV%qgS)pR*F4E(6IEj4b
zuFtieEIHIsDao=do2q)G?2*MoGqacFtBSRwzK;C*p3Betr)irabn6<V3(ATvlqZdD
zox=au<tn}V7hb3S5|=Q8e|(ULd2}Otv}u3CU;SUH&vEU9PI1T0$=cI@xm^8j#3VH)
zIVN@V_MZpm9DgKDhw;Nbs~!5bJ>_Qqc|0t^OYK!wcm0#tGI@hdL|C}mkzezVm)Ch_
z>o*5p&~RJT*pfamaru0^hA5{81ua&$D<ra8&vf1%zc_2h<C8HjuHH$CSS#8sN=UAK
zS2$~2NPJm?U&)QW%UymBH}+25F|=!kcUH{xz=uu~9DghPkg@Oe&tLepKfj@RL?v$E
zoyy_Ltd*t@@{hh)KJw((i)9|$Tlp`hRGnDbe!3+;k>~#<-~B+&>z5^kMM=T?;%^DF
z!+HW5#%~F3obPBccrAakeR<8K-7mUrO|I0`pN{+1_dvDex6yOqW^VJs2BMFoEav*l
z2DBczE-}2AX?4|IdUC_1il)-__s4bGx)>JP4@d4;v|jat)zG`_H<5jjKE+c4V=B)C
ztHzxb34N1JH1vAt`79`0E-&;S{O!6#r26F~r^N4#yQ9v<iFeB#&WdC^OR~0CFG;DH
zc+kp_2>IqOoxIJQQ?oi=#`(ry@s+xkg}TUd13f4I(Z0|%qpnL(@8Nc=PMFf&v2DVC
zPq&|qj{GBj^uilw-}}E@X;oXI6;t8sbL^r|Zbf(nd?)kz!FgzAbyMF@CRxwx^1t8l
z@<{Jt@9Ud0#E19G#Z6Px-k-88z7rL^%}zAFfAynh?oZDjpnss9B?W8kzkSyE(~c)c
z2Lp<1j*^Eqdz$XdFsM(T*5`6RNoKB@ZSeNl`{8{F>nF6dZ>-K-e{lc4D8Kt|$4sXG
zZm@BM+V0E~t4l}j^tB(4-SBc<h17VB+_~~*U47x6N#kbQY|FKItKZ`|xGC(_mQJ$l
z!_4cI!;xB+EhF=1YyP_M)lGNFYj1!5!aeaPHiSR(v(Bv+b&p(iIF%Tn)7G8#WLKB5
z*{f|*i~7SYFV|!jie~)vptrbx_n`Y*x?!5Bc=Y1FyLW@7fx7RX1eY!~dIW9RUdx#J
zU!LmTY9+f+k@p|j^*39E=lZjHd)DQ?a364~+1fF;FxF2``|{8X)9`1>Q%A1Nz8A9P
z@tJ!D`Nb~ZZPN^mJt;P@ZXyru+I4`Ja6WE$i{U-f6#aLxcdRSsxFvJ<pWSXtEp1O}
zG&4UuGFg{bc(bSH;Le&(%Q$<I3Qedn%-C*e<J2a2zOej7`}t(hxV2Ml)AkK3DuE!>
zTZ#u4^JnU560rn9j{q?OMxXu>iv7>Lw8a{Brl)RQ;u4D`1Z_YXXe*14L4q*45km_1
z>{-nvmPQaXhvZPoVk|sRa&#cwzK2VMFtznbeQjld!l4Gk>OM0nEV)D!Q(KeNq?82;
zhewZUo4SSF<`R)iSc}x6lm!Y$tsPl=gx9UbC6+K@F3F{o1q$z}D!BdLRy9eDSj7A|
z9ZJVQS)g!n*d>F8hnN1rCBm4m#wYodvOwYT?yT%L-h)4@5#dbjiPS`6Wr4y~|Mk0-
ze5&X>ide&hok%AwWr4yS6)AZ}c@YI%Ld-O-N9u8v1q%P(Z98L0RE7hW2xY=NlE+mR
zC_E}mPd_<a977SSnZ`{iQ)6X;!ilk~H+5?gglghWch`USc>_g9Xh-nTS3;8tDrhP{
zRYX%cPD43`XnZI-4MNf&8r7%hMi4Z@Y}QR6Xu{nLf@a*UAZTT++aEZ6T5O1wE=lJC
zFBkMtLnQqp@P5R740y+I7XYsS_i5mr#$5uu65Qv2cMf+2@G5X$2Hs`d)xfL9eGPcm
zaMuB^4);yq-Nbzdcz1B$1KvH{ZNO{8-3h!-+}*(I#@z?JKHSfN_nbA_Tg-eZHWpfu
z6rBg+JcvZYO3_6iF2Y>}k}7;Yn?TZpy9+|PAOh7l`m{%xZWJ37Ul!(RFdco2qU9iv
zv(rq`RUoLsT?2v|+_fO6Wo@NanPVJhvWQu&*`S^cdT2SjO&)HH8@GH|i=@psW?E?R
zle9U<9Fqmd0@GB^R7{o}OH5WAD@+280Mj(iG)&eUYfLsA8%(wwTTFHwJ52T*drS@-
z2TYluoe8>V)JS>=w1-%ujK>VNBHgibngik-d=6zGmf@CzM2^p4BS;!?cR)x7JEtgf
zGTkWBCfiLui1YDg%RyX@yA~w1c(bh_X~o?SA^mJOi-qk5x3^Yb2J%y?LAeK|W~P5d
zUS|83f>?_8R|?`%+%+Jn!TW0lNi*(V2<c_}kCxV6dHh1>(BJxxNs?|{cBQI^S!`i?
zVa8}SV$PhYEEc=2VeV3HLbrM7GDW-3bDz)3!^*>km8X>_CNC>5Ox{-Bn0)kojM#c}
z_2*&|>5DM=n)sTrpY${F!{l$`k15C?Xzc4#bg)72m{z`O4%(s<GgF14m#8mcPVOIH
z6#Wp?AHtYpgM6-hb6&QB8H#Z2a0ArqYVExbM%3>JR1{NZJOUPI>J+vd0XyW43o0Yd
z{kilTrW=X6WNg<`bOWe2u)7$g4I_=u?y}54+q=1<PANJSbW>ppZm(1bWSs`OX?Q#h
ze9|C@ji-ZdIv&pfehyf&@jT$?;g$kliaQ_p`M71km*Fk~ei80c;Fsbq2YxwjIq>DU
ztAJmHy9W3*xNCu5i@Oo{jkueD--Npv_|3Rmf!~U|1Na@dyMW(?yBGMqxch<M&w6iW
zP9^v1-J{If_;~n3i)e3Q@{XX>*Xva&`YXX#)6p%Rq7wuOm=XnvW7^wyc<YCxWz(3p
zwis@)MO9K>`Wn05PBUaKhw6qJqM7>Xsk2Ei@p}Was>JGIY*k$Z^^5GPIuXBs3=G}h
ztxD349LLYHtam2(pw<*wcK#5%j?meTNk58%9a@_doet*dfbG~05OjbeYKEe_K+uJ|
z7X-by`$5pp`ux&1r-C%DHZw|Q6U;Qxor|I~K{FF3vCR&F<`8R?4S5=$9VwYCWbf|D
zR>}ck4lHDy2f{quQV>dU=YudGw+w_b+(jTP!d(i&QrzVrEXOSeksQL<{;NP#g}Vks
zHMna*RExV2M2)zcK-7f08AQ#vTS3%{y8}Es@U7nko?W<m!Lt{4KX~@DzK|=NqSf4d
z<!`r!3c&<P=B*RYm<1WB6FGJ`mHU04R4RPh#vy4pU$+4Cvq{=Lz<n(LkhF)b$JlZq
zX-`{EOkTELn7nPhG5Hw#jAcENo@+c8lgL<v$=BT1itXLc+z*q#xj&{Llc2F11xW{+
z1dj>TSJC+FsYJ2M9Y(+nrS|Gs;#F$C1p~}bL|aGCK((&*q+jo`&DhIaJ_+h4+1=P=
zvGc&UUSV8jy*u(84e|Bvp9u?4eTvzQ3x{A4^22>k?pODyl{4+RjCY~XW%P{1Y*NJo
z61z!L!8{f0(Z?t{2Lw6b#5xZIdAOw@kmAk<K|bq&@g*55Zs#(Y-f3%k7HS4Pf-$`-
zp1<(kzer_I7!)l7fsAcUao83V;VuP1DeiI*l(YULr(S*g)W^EYWr&YG8a!r-6c2bP
z@lv$&c;~U3bYg7$az$iQ;>t~7iZ?NeLitjqRpO7drIhUa=eQLsmoJNsA)--;Ae5Cp
zawU6p1^adx`*!7qRq^W>ez7K4l~fz2K7PVPN^_DHmlLS%H*cD@Mxc&?t{&ge&E3<>
zTj(=)I?vu?>KvOX|L>~4vB_jJa|=s>wXL1Q45yj1U0h9_jT~qB^7O6dFIc7Gzfcqa
HLB#(6net??

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add.cpp.B4A511D84B31E476.idx b/.cache/clangd/index/profile_gemm_add.cpp.B4A511D84B31E476.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f6acbea7247d8dde54a22628700d4ce97f589184
GIT binary patch
literal 2530
zcmZvc3p`Zm8^_Nv6N4FpFfx<WxD4aY3<krHDR()!q-II38}(oJtTOB%BNDT1yNR$B
zyU~R#<(4ItNTywvS%uh3{z<h||E%`V9&`Ha|NVT<d*1K+d!Oe#@AE$I@ptEP$K(OP
zHNrhDJ~C-9761Sh_(@Dkh|R_Tz#dpjE2|6GaZqB5U#w`kkoDQnV!8x86mw%*<Cihn
z8-&Bw5&2Cyw@Z$flty(d(<Hk>et3O3%DMQ|(mh{YC!sr_3IwU$l3$k{!7wZ6B|;}U
zdrx;X^o;+s@S9soTN$s2^1NcSZ{hcxng1=tO*gk{+Lc2W=i0fw<Gno#KXkmAlsFZ{
z{WWvsYhGQ@s}1&7|119VOr2BcVcw(6mj3z!qp!5)|M>O1Twv_Xr(Se3(R})8=7Bts
zJpJCI{S_WTHUkB}gf(%l#tgOo#636WAvVfzU6N1>2iymjvAMZ>DhWd;rj{MD<o4n-
zU(GVcI(=ilH0?Y}DLk%!qWaAZx3N4AiKW)@tl;-~T{}#|i<?;dZP0u}a&g1Cc;E=J
zZg^nvz}va;yw}QBv)Nqj7(aP(BSk0r?QMgbMcz9nQdzeZ)HJFXn8L1HKVj!(mso6m
z=VwO%ULOxI?MRuF+sIBR;tbzbVf)d)pl0%$S@q|C-bc=;$)i3=#+j8E7xv}k&#6Sm
z?pFSJ^zgF~qGQ_DV@<egI)P7rUn-goqi<I{r&nF#&50x@YX$Fr=17;x;a2Mi&*nU!
z*mkunPIR^3ym_!CVlKc7SKiE9{|<A#vB-Y>omS<;7o(rku*3%wI)8589W^8@9V(OE
z$R8}3VW~g7NWP$PuDn0}VNg+;`NT_^>@S*SS!NaQswywlj*k^4+bi00n1RC=4tKWF
zlwUT{iq9@obz=t8eXIvsE`Rs@P37tPk8r^@x!Y<}=ltTr$Wx9&Z>^_7$9;WII!r?D
zG|VLRT8cWwUK=0j-6v)xxmVS0aogW=YVLK-S5M<yHh-U{3U$xE9gG$MkB^48h0Ld_
z#g24SHeK2@ed`XvHLrCuDd9&<U+f0nta*k0glddS&yv@zZ(2@j27L2t?5tt^$-V;7
zr9u<&Xio%d(yjk)e|31K?)<6Nc+ce#oAPO2$brDt+?Ahv=7ob#gjm<Gue_v`vBb8k
z@DlHlG!7Mt++T+#CDNJA)3m6^7v-AdhnI{b-Fo94p4eqCF%&fWz=CZozBu&l)RM2`
z(@gR!#eBm=!eXB_)tK?T?yrVvd%|~(qLPDGt+X-s27=W|*YB_zU6)f*`YS{hcS?IO
zsWLy7GQQg2^M{&IuX<&T@7c?(oTKj>(*7(V^|Lg~Q-he^w-WT;gM;ecyAlt@hx6eZ
zNop;qF1(cbEgskn0FVk+MQyc%mDQ#>Rd-=A?{qx@2!UlG2na#)KQCbAQp4@oe^l#j
zu5BGrl<|d^5M_x9tKA0?0L0Z%k}DCB)rvAZU|C64Nny1^<b=qm#8KFjY83QXK`{4H
z1`B0qWRQ@sun-^A;Dc+7%*!Zz6eb>y$y5lkk#LbWD1!w1&-Yl$U1zR+ow}@uhjeke
zN{Bl=q=(Z(kv>izMFuzn6wz=r6dB?SQDlTOLJ=KDN0Bkk7)2&H6BL=^OxK8|{o?L^
z*}X47EF=I`1)v%dWU%v?=KYlR>#*?QAdp@1%DUSopRE`EO9)LXI~dm1`UJ#7L=~bY
za#EqSwN2mq+eH)bkd=!S7x9mWtevg5N{OM*Fp?5epNS$HeH#?n>f54-MPiYqcJ?HD
z6giL_P{h_|Z;`Ha(sn|Tv$iveINF>wiuC5RWnDQR4)02dptK$-uqai*=z`p!0bD^s
zkRk-BNP1RSPY4<#oHQ>E2_n>t!1m!=!flYs6BLV@h&RP&;OP(PMM_jqotf@k)6<K^
zLp&PK0$GoTKB0YrqEK2Wio$4NYm_(dMzOnpObhnxs^+SN>=-!??8g@d50^ZN52Dh<
z2Tl2)C1SRsGx;D3;mSBqtjCVrsnvRM5rAtacxBIsWz{jFD*RaiNDzQ3h(*ImXC39C
zp|*e3OEdS&w0h^?MhIBQYMYfOvK|XryIH$Si9u(WNr_2kqR57BgCbkHEs9tamX6fU
zo??$82Z{rV*i^QGbfpv32}RCSXB2U$oHc5<_K#TSRy7N6m}|(jKpYHEKR=qyhKQ?q
z@DPALh=5cM=Hp)y9?H<pSgW>`bZ_E=RK!S<Gc(-&WZhNRn>o%Lk4(Q7^~xR6om5-R
zDf(9gBt*vTT`8r+gnfz7%D{i6`_OhG0{{qI8)6{ggq%i83i+;VP;g*?|4txyO#_0Y
z!vN`UUwlMdER0^!J{aseSvj1%BA%e4OjPnv<!+^^DtN5dR8!xe#j<x~J2`V)%r#6L
ybZw1v{@*N#vQc~U7OI|sA>G*2!qSRiW4DP(HnVoqAn9-S+$FQk&($6D2L1+;Fzn?3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_add_fastgelu.cpp.374CFB86EAD905D5.idx b/.cache/clangd/index/profile_gemm_add_add_fastgelu.cpp.374CFB86EAD905D5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8b5d2ff4ac2b068b07904776f8499f68a0cbedde
GIT binary patch
literal 2400
zcmZwH3piV69{})kA~c9g(@08)OArxtOC%Dv#3imZ$EBmIwn6G^61vf8DbpIhDxX#<
z?JA76GnBQc`51j@wMCy&T9<0;i!r{gr7ddRx=ZMeoITrlp1kk-JOBUtPu}-EC!ziV
zL8LkWc*Xf2P7oawq5%NdtoS4!OxmBL0szH|*x~Yqh=c=^EZxz{iJEB}m%gTyqaNn9
zody!z1xe9Kl#k532VYxMVzfB6SY<Jm7uFK(YH~Go^<~A8S&nJIacPv8u7L|49jQxC
zXA{&rufJuSe-a&?SuPq|>YFW^e!)oXa&5QYQ&DL|v$ZS457G#SDlbeg{QCUo0b}R%
zx02Fr9#00F>t|!Q$IJ%YZAvFzfB$gq?XNYH&z}*#H~<<aCZ_Ez8f}%gys6gR-7qqm
zI4<hJbguN*46_<m%*M(a#`Q<{Kcz)q?Du=3_fF%Nv5&-OT;#1(`<8p90nIKUk1MIm
zCM~z9jch3@SgH|mn!u45`^;U=pOU0ZhS0-&<POfuzgM+1ZocoQuXm@}_7hfVl3zc%
z$Bnf+YNUSlJbUqJXn{2T;Kiy^O<~q)l{l?k?hdWG|24S|@khfyrg7bd@Q=BrW`J#^
zSnhpPX3U$c(#T#Wk-COoQDU%l!pNWg$JM{QbObYDdR|~OFyqB&D;PBW4%0NuB690a
zI@3pr>_?@f$<$}vD%EF%;%u)P?f%U1sGXWM-c{xy>QGx<euhCLY1WPFc_}73M4E-$
z+5HmZyU?ARF!1?P6VU@}HSZ4|wEkzsG(tZ=V6}Y*zE}R-?F`C`X=QjMMPgrz;?uI_
zZ+P?N{0dd4d6H*-Lq$;9gRxPe{ru~!tV?ZbDLa44fQAly+QFd5vU6EI=gvkOzacMG
z$MttK#revJdW*;IQHGr7dhi*&s}IvtC?AiA3qrqnd?oTe_vRsdK66G-qSm)GWkkxG
zXcFIcym)kfic31tRlB>kKd-eAFXKCQEP7o(A$uNqhnt4-Nw3h!@K%@mlnnG@E3f+B
zu*}U!|Exqp|M2LWPr__xG*t4hw1~sc7H(-$zgDH@ac77z$@wxTX{2Szwy|^|A3u02
z{QomCOtKKVkUTPebqbdjwEJ6&>kFjCnar;ai@zf_oj7lK!*$`eK)F|MJE^aAd+1%>
zv&+)?vHNfTRsKW&#cSeAL5YdIC+4o?mOV_`Yh=1=+*~Nd*}bdys;2SL=hGd*j2uFt
zcqX}Ss<Qsk@AJK)*4fWq>?@?_XnmVe+Skbcbf<qF`@pq1UNNKPZtfH8f~G|#r{mTh
zji4AchcBr1u2IL&3H)k8qT+5=bB^`w9dmxyF0?Hs_CC@ie`HIdPQ8eJZI|jODgu&I
z6Jk6SZzV`Sd*i+}E(;4p0RW_p);3*l6p{5yXRax5Id5)73y4&NWgs8}Nq=_0+Ka8Y
zl0^HGPb<c7C6vZrDqt1qwlXjft_1fW3a<0M48YaxkzG4?=r>I)M8QzB;c{5W9Al1<
z1;zp)ON=E#R16g%D~uIFGz<+PYm7BQbPOFK8;lJ?+c4WU#3jG#m#7I}foP}<#Fc?~
z_(ujTVDC2h)z^bqDBdjI`-j?7-z*D$){Nvm+B%rh)vm~Yzjg59qswo0%3o|b@wdR9
z+S=CoP)@uy7IO4-6u|Yd5QE3?R*GrKv{uT=(g`7!B?}?8B^x0Qkwa3>;}W?Dxe#3t
zawEH$DKp*4?g)91JrLrNc?j{z{0;5HWlyYZ8ls}smTxEnDe&^wY#{?}U}v8yY1Qc-
zktw$C{XYHqNC4RvGSjwy-FsUV3kj`+4)DHWp%|+egkr5?5sI^lLr7#L+R(pW`WSKs
z3k($B<AwJ!fY(<g4MYb(xzp>%>RASEZ1mM#<Brt@Iygms8K_kTZiZ(xm-0+>hle}=
zc7APwGlv^_bK}}*$kEp^5Y9(K3?GKSQcOD2PAMn46GALH3n4b0jS$C>W1^hLHRK}X
zV(5a98^w*P%yg%?BjiExK!``-A;hQfH`K)pjZ^ih{!y`f0Zrflm%B$fay-jDOlN&p
z0?NQ(I53$rJJ$8RYLg;hhq1%Lx8YMmqoHh&kPQ;y02;~z33*5tf`me(i$OxM(m!G@
zpG0x5oKn>HrTH@9GOH!GH>H}(@#}|y{HsY4Nrh`qQgTvka@g8ny;&OoK<N8!tqoBQ
z!<EAw8$%VSEj@K>;MrYNwTK<zVIiSF#6}E=O^B23S1>3(SOu-Bwh5!IiPhSytD_x&
z6L?$UGy=A4#p~%Ca9!NoJ$QUid%`wXb7yOlf4WOFG$NatQ7ou5x{a-aBg2W!F=mnM
Um_7uerC;EODBlmf{K4(OALQYPI{*Lx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_add_fastgelu_impl.hpp.7165E9E08D31883A.idx b/.cache/clangd/index/profile_gemm_add_add_fastgelu_impl.hpp.7165E9E08D31883A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..db1d7d541f80a063ad9b781245ca2f9b6966dd9e
GIT binary patch
literal 5046
zcmdUzYgAKL7J#$i@(A7}aMggoAP|ItLFFMJq=Z*qSO$WiFd`KcMie9h8nskLr&WqN
zSSybJ3L=O|<>3plDr$X}hnkF|F6~$qVojhe+KQEN?2XHtduF-T{Nj&ytz@0=?0p`0
zpR@Nq_r%M@;?{`_M&wf2+MJAnG>pMu*pR>cg1oFC69yxI$zZIlIh_!*;?b;0zs8R2
zo#5Tj67(vd@%oOMlfxOKD<}OvjeEm(XH9za>oKpLi+W-fa~9b;3j4M-th-_9G3t?<
zywbtUBwsq!_Vz;|`*f`OtIp(^y%y(1M>p-*Sa4_V)+b?ICA+fgQbxbabG{c9+O6QN
z&_B6*pso1K)<e9|>P!6{)~}vr^ayK$H<Zsy)5LQ_t{22Gk0+G051gNy5Vug#UdA{5
zRMpn7G_7#(a&x0&iT5(_E*h=~wNN}dH<!`rBrsDvK9}2)#VmE%k`>iwS|<8uy8Vgd
z-i(dk6rUIzOU^5*{LSsdZF3&hf5tn`33NZg8+dwq&x4~&9-Hp(QMA8?o7(4!^==)q
zm!2)Mld<ln9u=Oy@Hj3tVx3iEpNsi3JN|ytT^E<FD<9fT>}(?r_vOEi|4XZ<>F^I{
zw3d6L^TNu`O$#X4va8T+(c+ECOAdrb<xOLycC;o(nWe2}7H5C|>`1&$n7uUIu6aUp
z*x0^Ehs1rvnq6TFOIV83U#gN1-s27i+I8ALW0pR#z=JIB-h5bQ|6^0(Mb(Y6ODDew
z`SAnRitEnzt~&i1l@fdUl<>uu>t76|#HI1q?Ei04;lmUBSGn7_f6r)<=o*gilHYZX
zhdbQ5->!rXH)_u|t`t8%*wB%4<^JP=*6p{Of=@PmGG}V$nRLyklZ1t##cSdmPj59F
zEMXNk9+oax7I~#>b#1EYb3wI>ccbXolTHs|X8x#ill*E&o!9T3@<5SC5x9(`pU7XA
zlO9IyBfBNP3^o5?GHPNn*^ET;A-Gmtf<bO8!xQ}z-g{Wu(t<N4nKKBU8PDU*AA@NE
z3{wmU$jivgPq!vDKkg3i_w(ccPCNjw2jEAwH1(O#T-%$z%?c+(ZX!>r8YjeVVuU2_
zlIgTe>Mlh{<|_-NWfd^H0)lC-gxQrSt3aTFa9Umkf-01&A+j2zl-#_(TR-I0ZiV4Q
ztZS@s);JO88i$aaC-<kC;6ywx-bk#!+6QK}IG#1d2o3mX=oiNbEpWA9M>WF;J<QO<
zEQ+nWoO1liq}MFTk=d~>R5dv=I}RZ^NA60?;yLjMC9)F{&H3#42rb|&Ks1v$Nk+0g
z|F`lZ|E_$+!U&Or=ucGVBzcMWG>IL=h)nDyMr0C4i4(0U;Y()FBz2S`GAUn*$Yc&O
zq)%Cp%(#vip|n>z(!D78N~E1qpbVvCcDFh#|I^gzYlaaO;9UWTS|xZ_qO1Zx6&P0v
zBdWl!3gv3>uBP?vH5XSgbb^fc)GiJIKP#)Jm8#zHKxeB`#piZJxd}!DL<EQ^J7Pp|
zaPWr|xfFl$S`}UHW`Ywb!ju@Q8YfoqR~Z+{Rd}@PY0!16aW}o&K-Xm}|I~zc<~}oa
zpN_j~`q%FWf}KXbTVb9y!XdXUTVzem7bC=MF+viy1R*I~ija&gLrBS1BBTbc8g*SY
z2-S2D!3Z@Zs9`=mss`LN;7+p^T(rQWSqE$#ApPrLmJWPTd9)5bqFE0vdb+*%)aO5C
z4`tSx<Ajg350_FS$;HP5Azz;$ny!M)Rj@_b?J9WD+ygc}v^-kXUE2TmAr`s*S$LKW
zt(Jr57>TL{t2bGI${jI*J9D|-+jf}Z#6sc1Xi5twa{0MP;8ySzMw-VIozBUy{s%c+
zJuvHmLpSOMRyRzfI4duIP3()NFf*L=cB}$elmI6x0joq=1(*uuD!{5xt_G}{=A827
zcENX@w15$}!1)#!L+9FW9}KL=ZuF5;YXh4$K+N3<Hk~vVNnfV!oHa9%+?$zxnPO^U
zn+|=HleqiFWpc)cV9Fu*fD+(DD@<ucxeeyD(WwY0+F?#R${jGJgVx^~ZTi`8>FXeJ
zyiRcF1Y<xiv)#fC-!N$d`6~5bT@S`ES~j|P(I$@LO|r#hn0=Yvz<VO@3!LP0S?_76
zt1T9AMwY+KC*MmAUTQjo*A02aeQP^<o)lbwDHp(*vi;udxmdJ(^TRh;TA%@AHZf#r
zkrpJBq9IF%>LHBc`kyCz4(NC<$<^Eq&bwh6CBTV2;JgRrT5zug<k(UN?sX{N2f=-M
z2^xx%RD-L}IFbjCMAt;*s4?GlK0*t43ycX2Cz5zcMr!laCN3P2^UV#>$WINkshJr#
zT8%~%4RAGRG}FOU9lhZ(LI)u_Fs1^73D|l%T6Lt{_O0K>A(-R|jwwPMP|Oq~Bw<Pr
zk}{<T$(S;PluRW;YOqnmWNPv_xgIuJly$(=(QbQ8zy6qQ$<u@5o`SmAPp)lNEN75g
z=YJ2dC)t+QFC`7cyr;f$>dMXJLm|%w$K7CO{NtCJiR$&F!Tdx0jp465`U}_LK83*r
z{bl}trq;D5uUS6!C@23twPLSz)U7ejQ1E*?{eIpj#hXIsaSSf(D+-_%h5QKf6(c0^
zmmmiSslOB<S(q$}Zl{b=8b6_M^1~-egXS$Q#ArcIYwJO*2jjWmn!vp<KGv~(+;`-^
zoW3!;Gu^rO9T$EObS-l9YdaYC5%tSk6X#5fxJaHA0&D}2*dJ^gY$W|xt<QH)zgBC3
z6QNkB6>SzFCPXL<3qvRz3r8pdi$Ew6i$ti<vXD*nIXsr~BBpwCpuzb#{>GEb$l;=S
zmbZNz$m$Id8)7N#r!`&OGuvI<#@#G0GO<BMNKp@YAW?&>8r-N6UmmYGHzTCy2sx30
z@Pd!2>Q`s`-Kzv^YyPhp^D?s23|A8Qc}w!E4F$uk!`sJjBlh-rdvReP^4g4gET{s5
v!J><7x@c)Enh*)_rDERw15A9P1=G?BXR$ffHWL^Vj2a9px`@+74qf~M-`uH4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_fastgelu.cpp.CA0713C21D4F7C04.idx b/.cache/clangd/index/profile_gemm_add_fastgelu.cpp.CA0713C21D4F7C04.idx
new file mode 100755
index 0000000000000000000000000000000000000000..05b280636d214da61ce8ff8d91f1f1add73fefdb
GIT binary patch
literal 2630
zcmZvc4Lp?D9><@D!INf;8DmC+p+;kv$#@xz8JUS0LQ?aj21yuHD2qf}H{GU^gcz+`
zUKT~Qk(bae+M9CMDlgmEQP<Md>sETvC%5<Q&geeo>FzzB&vVZCp5Oof|D5wb|MN%#
z{r!E60l-fln3WQr5rY8$fS~!(GtxF6)(3z?^8g^Lydf-Q+n9%`y6RT_9k$@-mdqTX
z-I;D`60tVvXrW$6g<TKn%u%KF^U>$kcB6+>6gJIvd6!_~O6S)F^$O;~7D;tOUm!0$
z<l3F-ahct#KOPJvXYSbk@vCD|L-yYU<aG?)d;4<Jm0j6?$vk<nizoFyDEn=9L0;$U
z)+d}j8>jD|4{!gFAslKvcK8t{kNv($mT;|bUDd<aRa;xXi`;Z6{!MDS%96Y}K4~S3
zlQ#J)BW3DX6eE9FaDJC#%D9ZP_IRti;0iygE$HDHS80dWk3FAeDkB=oQ#;K}sfqJX
z-LkfQHe|NbG~(`&r|u_m%ulR0KiQktR#=m+?tOk>+-Xg}a&;Eg<b#l;{&f-Gp;j+j
zZJ_?g4dY_hy-kHj)~R1@dHZ0Xvx7u`QeMCvNp)JWD$O~3AH9kfnOk0R`>ZXi|9jI0
z+Qzi4R>6dggW+2y#W__r`Rx*O=+UUhuKjldu3D@ac=6(Lo2%^MYli$_@|NxF68uc6
zyvt-#Qgo`FRbBfIYUM7O7dOz@-cUc0)wC(TY|G@|N0z<C3xX{tm3^g(Fkez*GCfp(
zOGB3teGOCXB_S$@jSB)Y@6nQPx_keQllnKrie8CRPRMMer<zY^oKF&0+NdYsg~Q@~
zJ<W+7U~xY|>1(ld8`R*|=T=YcT6kgkjQmFA?s#kJJ|_-SmKXP}>dtQzgOB8~hH%SH
z_Y3BOSEt@Jz4qW3*M&urrcTe)-0QBMfw=$v-Z?vZdbfMkQ)s;>wPA$S!B$_GH%;qZ
zgv;(dzU<|P9oPJ8I?g>_b$~QH(Ac`;avY)Dv0~F%=ohM;YBzq@{fa4BpJOE^H*2ZQ
zyIf+=5$3VO3nq@9j3HhTZV*OnW{o;Oy<1mWl*DEC+&H^7hQG3pF`_Vj1L!BVZ*O|g
zAEHus@*<o#()E^bUCH5JvJ}g359*FK-zM%{vB7ikkAs%a@9jB|^|*`DQv8JPCrsG?
zqFO#)AXnKrya}x+b^plc+Z<7qx^6y2btrjOpOMZ|dmj`Ogma8;5{f^$O!YCwKQW%Y
zSsD^*-EP#oBXocSS%vleKH|1#L+ht$%s00Vyx)7m%gH=p>{&S3dTCfh(4n!dZ4X5y
zZl(!-qr2bh{Tw~mIDN7yFvI5fON_$4)>#ETA~6r=%SP-)7AqDU{6zi1HGB2B?ti)5
zZu#nE)I{+fF!y})dmueKB{oO%HuF=7jh=1doAdQT0RYAu;*DoJjWXNngVV^?o*Oc7
zK!irF1c6G>^1lP1ecqa;s!bScmdDr5*ITVIF)=hTp6x!10Pq5Kz`2hekz;{KjZDB0
ztY<q!PK!@S=NbS2-?XQq?Qc8&n2&{-5YuGN3wD4UP~-?XqKE~tP{fAVC|Uw7L6H;W
zgdz^aL6I}$j3O7v1x2oq>m2cE)BQ<?G1DLhE(3{WAPL!|O!QBSq0CD^XoSxhLb4WB
zezkk@saf8igy^jDZ5zA0?t)ktPrzFsmk@pCO#63IuS6Uc_Vo4iN8)2)o`@&bkr%_u
zQAd0RA4LL&07c#mZxs1Zd@OZ-LW&SYB8mt_zBJ!OI!iyAABw~@F^VKK$sEOpEbHFe
zTo<c3l?l$o3@MBRqT<$!(XhRqHa%9AARS4MmKRroY=jj>%69x^RfWd?a~_c*QPl3)
z8FIS!Ckz&jVaB+e&@Yy5e9w4^g=3kqD2ijop-9e@&(Xmb5-YEve5xjL04acq91Y1k
z=F6hT!b*^$1X1}=f=)`%6>-z*9!ijpaFHW${rcmJ+H)r?N)RF2Lq<akoCA_`Knh}y
zffXQGf#zJ0oQvjskerX^B9L5!W+h10Y!Qi9og9n_T7BToY+?GAfpevTBo@^BpM*BA
zzv)_&k0)w&Tn6IGKmrn~^_a-k^u~?uKUM5S|I(~R(ZicK4D1=;8I0ItV4j2*s3R|y
z7e_~Y79T|dmH<WGEN>L~SozrM{Df9Q6p5@vDDtKI+UqR+=zb^?)5R!~&?R$pi7S;G
zNNS#F4jkkV<bnhkq<@|JS05NZn}?t>a3vxjwX6T-&kPF}fwUrUuHv<oV50=-h}+oy
z`=gic8nkHQaUm`iS$+bYE7apR6={}(9fG;YGBWBJfALGyR=-mv7=YZ~YvoTy*>;np
z**ifCY4OQ1+9#5p7MBidCu_9bY}1Do-wv0TUgx<Ou9K||ldb{QA`$>dbyJycx=uF@
z(@o=2<jI>gxFT_-KE}Y%2%0w^izArgO@fI2VkXfz$jpLdzJSUXEcFqJmiao9mw4Op
v9Bu#KD#gm0wr~-BvAqL}?d0O>=I-fb<6+6+O2`z(^5Ag2fK`5hU<mLZ6tulZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_fastgelu_impl.hpp.23CA5590877F722E.idx b/.cache/clangd/index/profile_gemm_add_fastgelu_impl.hpp.23CA5590877F722E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d8715452f96170f66744f7dc2145296e047bb5dc
GIT binary patch
literal 4718
zcmd6rYgAKL7J%=DE0TE0<DvlsfdEBC1Vca&0+<J=Aq;Ov`w&F|gTzLn1<_HKwT`Wh
z1hPO;>l>*StX8d}RkYQTs64Dy)S@N|+TqcPGo$#P8|%$Ivs`Qb@Q1rrvd(w*KCgYx
z-us-JIxRjvhr}@XylF+6?8S3%48!=;Pq(;0Ewsb1T1N~k+IlKIG3nj}UQ?0@{PgUF
z7YkSGzq;u>Y~!tyosn05+TSoQDvp1*;hP1c&sEP4&e!{tJ~&l$yW<}fJ+FS#q#lp|
z)X(lw<D^9$r7gFoJiYwuk<mHJ&czk1`Z>Dm+nW3LqVMm|6)ZcqB(>K?o!)D=-0;=t
zhnfJrxp3N3!HS4JukfHvi<j<uC?Gx%DLv9cwr*0{o%hqtDET_w&7wY({zc=<xa$u#
zu+F$oAm4ux;_<f)ri`hxvI<6(*KBHE-EmA@ySmZwmVaxYtdJGzHKn_e^<7bH)9>8e
z&<or~!)(($@`#txF6eZmXQC>OqhE-r_pOy|U2b_c=5@!DF?Z_S+kUy%e^*BYizMe;
zi>v$(hgOwd$zQa&H!t_djGV=3!{=O%uX7!G@5)DkJXr-<5|E(Gx*<^ihwb1zt!Qpl
zTso^RaOR@}(dYa+m$23@e4(4Ae(A-Yq@V6q8@g$>;Qj8A6$i72-z;hK#sy)8s^?@U
z*YTJKS3D|}8{w!wu&%#i=5Lu#H-zLS$W?|<TBQ!tXHPE9<reQcU9;`BTWm6Sdxgkw
zjUPNRpXEF{`}4n^I}!G<ct=oPV}#CW*N9dphq(^d)WJuvohtFc>D9FhhQ81#UtKWX
zQVebT_qsOA<EW5Z+&=jO`}yJhX3s4D&^uSsZwk|PhE?Y_zb+Vm{IF@bTeZi2P3C6L
zq}*51ZF7}8)8fY;mu;AMBXj>jaS5hds>zy6?HJd&4}12R?Vi}#jTk~9^^gL#fW%mT
zf<6GdbgAoj7J&_Sz)0Z`q3@eNjAaLyJq{!kWasFDxs-BdgS1&ZmJ0;g1_5m#rd!&J
zhV9#aDQmqWK}!9kW9e#wl=;aJlKacYF)~GnBAg*Tgy|uY;W7v-L%AFz<sfC`29O(I
zD$Nef4c$Ef2if)nsq|6W=0%VRJ_!h^d{v=z6Q|sf;p(qF&e-Fm2}CAlF*w-{f_8AF
z6>-u6K^B-mGq2uLBd(a*;y{hXQ3~m5YAj9yLMpDxhmj?5k`P%kCmEp>Zi-FQ<%g2W
z!?!oR9EOupcj+j)1BwEqfegvqWr$1`AVXww54k6!DG!tfF{E%;Ang<p3Y*1oQae)X
z!Dwm&wMaXyL>oiPT-y#gzNk4Cg5#tf#_Iu5D}(W6D3^n{9BdZHNdt^GF!GUC&M(5c
zB-w*_t?m>5DBSX1Y1KPB-kV*j{mhlDw8KfUOdLns1}DQK!=q^urU@-mx2)J<N01qk
zj45<ALFNVK*|xq<?2-CM5v}Y2o71nOzL6a|EB&3h1}E+A5-b{5`;H*ev+u$JhdF&5
zYI|{{JbHdODdWfxl5^w;DL4v*v>Yu$Mi3ZL$1y^zk#Sm_G(x%&rZeMEIm-n847Y=@
z9ef!!1IG+V|7MtAh7eS)G=qv^3kWSt`}pra`#rxWXFKJnLA)RVtwxbBXf#40K@ki!
z0KWlTQ1)wp0EW*3|12X{RyCA1f7;_jz2Ph(i_fU#5IHtV3@`s!Fn+zB3Qp;gbYvUM
z51en?1_Y@NRNE-^m)-VVCzs9`#IBLXiaVvr=53&sUInZw;4+Ip3QkACnPzo??vtec
zn#njp#S2^yPPBj^%K$G!xg79vlnsCz7}iwgxkmo#c^`1H3A~%Y7D9`Dv~(2Xmz$`m
z>;?W_Kwfnn@aq^}IrZ7B@(IBilqY10Gvn!rtg2IKGQParNX>8$xb1;4w16P1!L1tQ
zy)bDnlXeKQ2Ha{G`8uWj>fX}V5mf(m;9du|uugXf>^G^EPRJE5e-}ts3|~R5s0w&h
zV2i&MPiB9#itBNe>gW`Nonm}wo9vFnGbO=k5O0j_YbD&iyk|OU?nVeOGVy*#PeAgI
zE>C`-1Si1l1bEX{*_oe!D^u3oeUpkMCa|R%YbuUw2b;@SQ*n$1Cesaf{V`%}N0;w2
z>cv%ncLlsh3kb3jyem=O2>u%ZIelCO$yLUzd)CgV>RfW#gL2U%pCshOknEF;P>OGg
zEt`Ea_I}0(eJOzs*03fvQpcCK$0#c+jc8ahfxv`@MKgGrnca(%W{5I_Eo)d8$+0kD
z@6d<WLw2p_lGei&ONw}pj3q-z&XOafU?~vNva|>pfp3Hn^mGZz#{71a&A>7<UU#&+
z_^8VlkM<8ZKh<;o*|J8Phf%Npe@~AmIF7A9&FDxR#C!UC<JFmUQUB^3un|s&rp_A@
zx2u~nf+$90i$D#^&wYBEv{u&;O%wfvem55j_N6?(r_l{!%Ut3q+MaWJA_wsln|+t9
zT@{nYwK`CUbRunB>UI($Lr5MfN6rC?$qEIdsa0rgcMXENFDOiC5@?6Sc1U5gEf8-3
zCCx1o|K+Jk9+d;$JJ&o<4!JST`_elOWQk~zKdE(v0gup~cOYYuv+O*TNX0H<BwB{M
zgxjdwaNzNm<65l8swg~)O)rxmCAb8kXgnIB7(51{SUeV?C5}rtbT7TnKkQH3vL@W>
z2wagZk@tPOGR^V2$epS#lock?+K;x@Hv}IP`VDyD{W*9i7Paysm57btW28Q%ybb2h
zzT0><DC**2YG932lSx;<Jk#vISyHt1{~8J_$j+Z*y<+PM=IXXs3q0%JpO5uR&fDkh
zHQY+hbRC8bWr}3j&$8X?O5%WU9?EiL6HXj1k3S4^Vl+5Rk@_k`KLlOCuwhJ*%@nP7
I`nS)20o)-68~^|S

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_impl.hpp.A44323833CE8C5AD.idx b/.cache/clangd/index/profile_gemm_add_impl.hpp.A44323833CE8C5AD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e1c62613961c97e813d2b0244d8ac68f58a86d47
GIT binary patch
literal 4680
zcmd6rdsGu=7QpAiNC^xH11ls1DTILK5hMfwAp&`T8p85!tuHDH3W|;518tR7)YY!5
z5EMnQ_&`u9^-*+rD2S*eRY0p2Yv~$RbdP|m&}C)OYA2qZ%(usL_8<OX&Lrpi&As32
z&Uf$qzDb-F9UW<hVT#4Ga<kJ{FT^nnBdK4`>Q$LeW*C-liD9{A7n6U^dgyOETs3mo
zT2%3k@M3A=(wSdxSpU;T$FHP2qE;)kN$Mxf9~BPF$-LaOILY6oecF#@gS^87uX{Hp
z*G1<G&4${%R`#Ci47d1?{$<@{&$(Bk&gTCh>N{Ii*(!Qmy(DAql{JY^O9HFcLes1}
z_9H<PXFhMN_}n_8Y_y#3pl-O3oNKd9#P!(a#ap{G`pY!!G~Y`p)jOZp?f1F*a8z(I
zrwunu<g!GYc;}n0^m+?ViR;<=B=rjUP#f>V%Zs+GuwQC-xDh|VK04frmu@Uef6*v$
z>TP^D>-6FP)_sl#EO~i;hJ8}uqzZoI9B%=q>a%(8_da;#(#v1>Fl;zv(}-Q`G9pe~
zyP#4pjIC@9EIpWBdY>nH%$q-PYfx58$o?a`pfiz{DJrbm*K0@mvWmNWu3Fghmvn~&
zkJ__~I~qG|i}rMRxHb%RK2!C-iPwEu`~4@!TJZ9liCxFHOg(zj;sWk=`gr}K`};n~
zi>!N;%l(IV_=|k(NyGLB`?tO>QtX*l%1=8`o6in7jhX)t)_B`~c>3_au8p`4KcCN&
zWKJ49(o*cdGSMZlFK>I-u9c6Sls<=ih-=OtS50p-+Y-=Tc+NY0eppv|($M6n6&CAP
z$G#Z~?RF8|{v|&(G&QpE(l3Lp>ZMH{$7&Uo@)%X;!REQ5x?gW`dB>cqvr`LSJH@@q
zm$vP3)jnDN>C*xGAgOyz&id>$H)^G<etvt-M>far5t#i1jC7me=JfUtW0?VFjspqU
zs`QMUDI7{QwOQ67@!|kMc7b;nNa(8OV(Sy-H_~=m5~NHh^P(RUq+BRRNa3mQVPwhx
zWiUepFtq?e87_pWg(z!5ss$M%F9Jmo%%s_(qq%>;`?QrgL8=5Q)3^vSMi7IL+Fcz;
zSFu^*6rA~u^Cfef)Pq>hOa>>r!J!+tv?5OSfL{;z(`;Msk}cUgv(tj=i>-2_A5(p?
zV-Qkv)B;8p$Bsi}@$7hn5;zGaO~=b$6o2=({l8h`q>L|{L^nW@x6GFzIbV**<lb^b
zrf^ocFq#Tqg&#voz7naY3{jdaj*}WEjWeUE@zo&pG*V4CE#r2bv3yna*8m(R3qVu=
zh*}|t3Q^XAL<=U1<75$tiWs@m?W-%XK56<mo-04-*}77X-6ij=cvq|4nyp;BVl$kS
z$R$y<ZE!L;G<Z5qf{5Dsp^%$aW44sAx4z9Pa6$4r!<Z6xw9jFs&&hWLp)P~pFSA%U
z$fi~fTV_j-2q)!iIYJ7y0wE<^iI9e^K}ZJ<I@Bk05TRrI5hrz!tb-4ke)J&JgD1n?
z;MNW93>$!L0Hl2b_!}Sqm8%S(X1E93dYJmr7q<R$`9MZF<%52<eh#!6MQ(nR5eo1N
zVW=6%W^hDV*bLqbw*c9~$W<lHyF2m+Y|IEUjYuOIwG1M|M6tozPaQ-%3#ho0EKQ!l
zC@u9}iUhYz-%JxF-Z^UC*SLQ6I9~Nhtf-6G>9)J5sh0q&1USsR&VkK2u%|h5RnFSD
zk*Y|Xph5v&05-IMAPWI6L|F^C7Ud$qix|!>&f<pt?D80JvK?I8!4w_o2EpoWQsORp
zK5)Wd_9C~nr*{K>tDPEQB~UlA0&0M)VK{H*=mM?(loZNAQYER;baR_))Y&NqZ?#ba
zuK->Jd_W5b@)+=rp<D^mDw#w>kX68|V&s2TnQwWz`%MVdehu(zz!b>$I{wK)w&@R>
zs2P=jZ3&n{>)sci%-_s$zDw1(2vaXIKJ>Nxfz%}-#%3HZjPmVL&S2JP4mEZic<Y!@
zeyDoM<vT5FZc*<eMVb=9SToBvYYI(=2E5~29ba6gYBvC{0bFUp;pH*7Dq-8--=<=z
z9!#mmn2Mvi!Q@xQR2<#|k#xl)&n9~H_PLKzuW>K9?gbB8K#;}YT8#2O@Z1N;?&B^<
z?=rTp*gm`D(V8Y_%8%m&ambD#UJ#E^f_s7~o1OA{oHBPX!PmkV`y@J;O80GKl@5(v
zdT`LAvC{zW8<-`IlLiPgfGKMjC&}(%V&j>4_X3XW<k%TES1cLgVse%oAq7i;kdmcD
zNW;<~qyuV%6Y1d+l#QtyVFR!XjBB3j-*C?HtD)*KC$W0=lg@3LER1^n|GRr^U|Zh2
zHoG@=954I}y1O6Og#G$x%tmMmOk6x6>PSCj1aY|76pxx!I|NVLHAdGE&l3NUemB=w
zok)23Fgs@)TjqwXvg;*hAaooLvBUkd?VH1sI7SBwkOk4krA{XSa)cCt3S_gOj8rNa
zO^s4xI%^Qrc|oa1Q$jbyc0&TA-2>4*prW}`>iKzMoOAJ**UpbFjT3(IalP@51NDS-
zDqfuA!kByL&-ylHn!Wrg6~-lw5+qs%I|iGmzv$b4AN1)o?!ChBFe`eR1S!R(2u;VQ
zBNUE@BNTy0AhgDE4V!M|>C1T|v8CIBjgG((o6`E=*}NpndtyHIc%6J*9IZW6R^L43
zw3~3u6CcUIA7Np6*Qso+1Az{NbPuEF_qF(iUH^{iI9ryTN<aSXQitaOX>Qs7H57Gj
zUcQhzHDfue7Uh&03&u0_e=l3QfML`Z8v3F><j|J|UBEDcDcUeaa>jYht7WXZ3xu78
c<$Ek^!iLSUwZh(GG_05+iz!+&MK)9X9cD8nvj6}9

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_multiply.cpp.136C377C7D5B0E54.idx b/.cache/clangd/index/profile_gemm_add_multiply.cpp.136C377C7D5B0E54.idx
new file mode 100755
index 0000000000000000000000000000000000000000..23c1724f1e8ef30ff5d1e7ebcce9f0934a3eeacb
GIT binary patch
literal 2398
zcmZwG4?I-a9suxr8C;C<e`rioYcPY68uQ1@#0)ayUuh<!O)LgmQU1raZ<QvZBBro1
z)}LLaB!6PJ^<>kV+UlW2$|#9u_0lHvYSoq<de_|d_MOjX?ws%WopbIzzjKbrm(LH;
z1pqIxZ$?5)Y9tB(0AtlJB{g}cL=ynARI!YT3&GQSM;JzTxvhcnZRy1^zX4f3qW$zC
z8~#1(Ra*IMUx~elXwMmq=o0uhnx5e{WgLi{dhsv9^ox~!x<XI!ZP9k&iNu!55amQi
zv(MA_J>LvJfAQ??1f`=ga7`h8?>ED9%3t*7l=H(y6SLz#&s0Puy<xFmjP^Bl{c$kY
zIjg5ddFEYadRtfkDN#I}E2nP%>hEvYcP*rz-tn+|&S~sg&qhN^zk*RQ6mHgiK24!8
z=m@Z^{yuy@NKw}o{-nIsoO6xc!};!L(2Dww><Ht9w>}q=Ne^D5a)yRQJJWUY-_1z~
zlA+(PPZ^v(Kw=zr@I6}5R#=xZB!5KjNH}|{zGlV}|0;`oVfO~Kh_5x<;J@C=pGAq$
zc<jaq-!{;pJk6XQ75y{bBK1jiiGE~uv8LFd;7gPE$YhVNYO@>a$3&@LRN4QKM)3y1
zng+Y`4!Cj;w0l`VN+#pW?<NFKEQ78cd3iOB8anyL=$se*IJ^C3ZmcUcJolybO}~+_
zr%&C-8v+E~p{)t=ll3}=M?!MyjWK>JD;AnsFJU+3dHPnF)FlR6q~SuXlx}9#qUJT^
z{3=NZw7y%ozbSHLc8p10(V4INi|O|X(OEvh*J}nkFPeKNWxTHRJNO9>U*Oh3FAF;O
ztyzEK{?)0yRlV7`r#XW)gHqqZw+c2|dq?=@CbN4#^&U-$puO6@YhP^4^Xqr`uBPXw
z21cX9PX=UeEqQ)+u4dfrPNofRGOVm-`u8iR`$nM$mvFU%93~~?5@v=Vw=+nSS6-d{
zW8V#aRC|M>8EyMn`9sVj+rr$jAmxVbg8bg9vWL?i7o*}h*p>p)LF@3oT=!Z@BJq|o
zDrcJ9-;`&(>f$P;;+A>U-$qaBU+(_N+UWw$ylz4g@0oD+sv-6BSdPx6&3nqzN9A{u
zGmFL*We&2iW7mn4U1a9*{P!hyqGx|}nooO_L0WK;4~%sd1wWvz_;uaxwfg7ojugd@
z*c|>)-`Y|gA(*U9DtD9)ojXOneA(Qwl4wF|DAAmN(iX~+#8T)(W!;&^`-dhvbYk^d
zj_(>wdiFJUdT=2l`*!^8_vJnKt$J1o?;7%D<g`zH^Ei9jCV5hqpT*bq{QBCUxT|8v
zgj7RX#yU2xiw?woR#((-AA~h;{(Mu_+WStJh1fFr6Ytj7Il1n`uJet1Zg%Y@h1aBH
zCPaCv?qp!4wsg0#UI-f41^_|~3bS(Qp^7Y(;rUc)Vzu1V00>ouWgs8}t^V@|EM99>
zNsft4DOT0lM%}WuulIe>M+=D%5wkq8&<3(W$QH6i$PThYhy;-kB12?^>>+!E93TgT
z93e-9C=dl9C&+18Y{JZ~-P(~eAWB#Ynv{ZO@JD*45Q=N?hj1vAPy%8lpamR23G+Z~
z9?}IMwg71<h?S}hkJHfmuQZ}nIa?zG1#mh;c?!$E^eaEAKICIg;lPfHe{??n%{2cn
zfr$G0_VXfE90o0<dQ$oDd(c7`p34TcXm&IQwdi(qgcx=VgqU_rgjg0VEA=?G1sfrb
z1qUHFf}6Fvr#rzNAufT75Rbq^$b;aqtXLDt@V|BLkTe#TkC%Z&c=?OAm4Ou4k`pJ}
z4X+N!z{LbEp1qGp4ge&zSzl~CE5D(I7DkdI*TMUT7Dkby5Q-*8BP1q^5sD$lEGze=
zH=fmZ45xaJmzft1USD;*Kgv&-Grn|QJWIjlMAvsTKBMp%bXEU2>o^|#`sT3v)S6?4
zCaUIA(4Z7Fh9_z*;jK36@9+G}w*J6BmC?wXypKT%sRF7$+#e-$@pkc5i{?mkR*UXP
zM~LCbK#1wcM2LlFtyYg?<Jkyt@EnBPh;AfxPj{j_LR=yjAs&&3kO$FYSyyNxv6gqu
zlxq2Wd;U6jx?bY`BiZg-441OvR|;-|1Mn_`8vnDqBZrmuu1sjPs75$L&S<#*6x2|t
z!EQRCY8KcFXmInJlQ(oSZ62E~B^P<il4If{7Y|HIa&*d;#meTz$HyXqD*zC*T-6kw
zANDqUHF{`PJ2)h0%VrS}vK#|~)zziRJ~BE%9KTal4T#&MiPF;6fpqoJ2F6B)7(WyK
z2C|8spQ)vp`3fAH<L1ugd3dhHI<2>Hby)rXS}pJ^37=RKZAtcylr`(9E_5bq6~oGz
U=8d(m^YP!RAqe#H1vdhJ0ueo1KmY&$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_multiply_impl.hpp.FBBEFEDC4A78D751.idx b/.cache/clangd/index/profile_gemm_add_multiply_impl.hpp.FBBEFEDC4A78D751.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7f3d50595e5f6c4596b31cd6aba5c320aa9d9237
GIT binary patch
literal 4998
zcmdUzX;4#H7J%Q04<ztN;E61P2n4X%)IeB`tY!ld!akzXN`r-%QNixejt!Qg!>H9b
zI09i+Y)2Rc#0_a}(bjgu1`MO1_Kc#K853O4o<>o0UMyeUow90v@yAz{)cx)`cX{XD
zbMARJK_(U#PQkFKFJuLZtaZz948ts`U;eth9A6fOd9pFApyr39*xUzmtbUCf*=y$W
z-NB&ofc?EYYrcP+H9E`cm7Dc-o3a`*`rT{K@+Dng4#e~u?K_}q6ZA|Q9C|zC)Af&~
zn_XGzOGaylN+nNs{W_$$$$n9{#<F?Ov6)N0QZK)q*y7@*^XRWVGRzMAbI<G}sk;w8
zy3^QNa%OuoFSPn%cbmmGuZ|4PSU4}NOSLX%?=udO?icNl=<D6l=;p3E&syBHEj&Zx
zwZEV;<Hn1>^P#{@9C<#YKc;@t$)<BLSi66L>Ae=?&e~;0!mSQ}we;sES<AQQ+{@-y
z7(H>hU;cE(mGL<`S;&sy`whjs6J~+#M|nff2kHioW&R8LvIp)Ih5L8Rt8tXADl5FY
zvdpoJommrCFcSQiiuk(7oV=2ALgVc@+${d0I=9IOiZus|fB)jN<1u;e#d8G(jlq$P
zNv!xBS%mi;=jNs7K6hc?P0zMWj`WI2Ijykf%f3#=8q(jr@jm{RYahHb#qN-xa_rx|
z;?%~C)A#jSod)N_jVCj2ua<AvymsG6$RK}w<7Vv9?QI#VEl+m02IZgdy)gWMv*_yC
zX@s=s$mY7nr%k)Klfx3Y^|kz6H!W8$nh{eQ<gqQ_KjYFRN#Cy3{8myP-fp|Y?$&{h
zURJShW>I3(p<LG6(3wS9L(bQ)&l|t@&B^g?v9F`vJbUF-#diE9>RPd8UfllazHYbU
z*FSkPyMNQ%p6tt?J&eyf)%>XR)-Pd~a~H5(R14oK)K}eKnGwZHOY^^0EDxu4kZtC(
z;ih}6Q5I{e5rx!4TANyv7`37FPt&Out+zP9pTMk)F_LG*<GugG*et*%;Xp!OR(5`n
z1*MtM8PV<M!3BaG1kXY6qgzhmn>E#6TfW1TAVsbs54xHl#jaw6B<>O~MkaNaA|&&b
z1v0V<@UMUnhAY9p5@i(#RS?0*t3X(Vay3L%gOrvVcX#TBJrA4W1R3WNXP7lX#=FEL
zB<INmbQ6M1;3XKye8sfDoP!Q$C*fot@cWn-$4L!XYrvLnhLbv&rGq&%TXZ-o{B}!w
zOsJ7Lan5u#H8Lk2Avss>!pIW12?!-~k`c`mP6|S)+*Cv}jhkj5n;$opAN^<LxH(RW
z>_yY*&MEQ~c{3!o7b7yUrx=k*93+m6ro>w^iy^6l6p=~2rHD*sFGKp21<4HSh?B}`
zN(ZJFrMD7krxYqfX_;+*o5`C4?Y>4hSphy3fT&f1PbJDK@Kb?drEsze{Hjo{2A^t1
ze_G#9YcZ`b>l3w$eZW&?wOgs`BM)>kEnTz4mMmxCWI$wqh_)k6hJ=L7qsb-gSjY3A
zUegJ;s_Rh4B~!uDq>rXGJ8qxWdac)?j|3r(BUf{cmyK|#9mx?{&@;qIF-MG$gd;&n
z%8?=@<H!(Fa+C<E!CH-ar5eK2OnkseH6*Dag&DOET>HSCVGTHIfXA>FI9fpZ*TNhv
z_@eS?Ei7PI2hKXCz4-K&C##0D>x~JLZ^5^w)hKf2&qT<VAH>iVu)G2`D7#((Plmg|
zvWt;Nt2#@&HxHY$2r?%$CyG&1genXa8(6m5+God3YBkb@=@MEgkuCZ`k!A5M)wTiH
z4Zx)Z1lbAZoiK&woV@(6;$9vIHzFvn#4Es@77%15;FTz=09T=01$Y(8)qqzstSDb;
z8*;}{131|aPW@mAaRuiGht}iQf2O9^3YM*a*sdKc+ZisBjx8;lGdr1jwH59wg6WBE
zYF?m7uDyPVnsGDOHN$7LfFN7It_9^*2ySJL41zoi!G}?91G_dxe|z+#;>V@$f~fJ@
z!M+^~fp@pfAFU$_tqQ4EX#k4`Foe6^qn|I?#C5nqwYUWSmzdpH7kO9cD33A!L_<=2
ziI6+8ax9;EFEx0onNYiD*fajT&FGJm-~!lP04Lh^dsoHa(ekbL-=C(!J}^Xm{b?%F
zfMH|mPt#BxgwqWho=)``((=Zr)vN`lT5zKU1X%}8btu<^dp#iO#U60qgYsPv-es2H
zyOK23!!OS`Pzge^OEQuYq`0IYl*&sr95e)(#!E90JrOW{pE}VASF496IekR#ZLE(t
zeroWi_4FLAMk7WaSofh3OA9l!j33~n7UpWf5XtqH;pmu%)V6rQw_yjDWK#(kTZH(f
zm@P&~!j>Q;WlIr~v1JG;*-C`eV5x?w^yCR@nJhIZYk{q0y!p6p{c)RZ&;K^z$4m$3
zQP0*jD=}(E{cryHJ;$W?V%kvbC+b3{FW>lG^V}B?C){9X1PNali5hg2!33cKF};AD
z(SLGz{PP;UX9{G3g>-dSUX%Rw14aHPYQ<|7^ZQ?OheJNm>38$~P_ijBk*jxMUr_+P
zDAYHUuNWbTK!PL*Qh^j9S-5OI(@r^GY4`#nsPC5feduwffmjX58EqYibr4N+kI;QX
zLYzbSgzw0kj;D-nc{yGC$b|=kdZI=T*}{a6s9V{R96TlRCn|*qun9npgb<q$1L=0P
z{8sDLQ$Lv?L-9~k+6)94hKC^(j)x-@fkz+|iAN$7g-0Q@!DIu6?(^~M#V=#4w+8B+
zk1H@FG$Tif5={p9_EdFYWMLew{k*25bM|3p*9kYvXwtxZ$?;l0YP`k5#fBsIWj6j0
zn_JXHr7vo5QG+Wzk+BmM=Vs0AI!fuUi&&RISC60VcCQi^)cjxT=4Gu~roU*&&&$lO
z))(}52_GK)4cCX~!*zt7sB19#nM@Zj%$zB53`G`c)i-4<ca?20F=aB_#FQ}Sa4js&
TFcX6YX2ui=rfA9(xlHlD=GPGu

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_relu.cpp.7B9D4CAF29075E5F.idx b/.cache/clangd/index/profile_gemm_add_relu.cpp.7B9D4CAF29075E5F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3ee8f4d173f6faaad440d6c7070882da99f0cde8
GIT binary patch
literal 2534
zcmZvc3pkYN9>Bkk!C>5mHDY#3Gt4kpGG?Z68@GvZpK*`eNk~f=he9SfLpl|<&=y)&
zF0oE3x1MBIDioWDr;ba9vK>52vQp=x{mj?d^FGhK^Z&j7%lp3n|Ns4XP8`k)Spcw$
za1z8v9t}kS0H6Z>l8z?E<VpfSGNcofH+Y98jMJ5$ULLENGBEG^DJ9FARNJW;i>-=1
zdq%?b5~*9Q_H5DD3(poj&&~HZ$r}(tM%&WcdfU@`lWdEW$0{HD%#h0Kt0R}*hZ<o2
z5d0oZZ*On=Xz{Rd_WVO)?|SsRjI`zMjI`q^Gq*aJi54|}6UR;$wx4O4Psse&<alsP
z+}f+--v=4uhGp^o?_!!Cm0Z5wHL!4JBFgvXVuy4_+wIrc4gq7O!w>QU1vi|u+9SV{
zB`yz8FLU=ZM)HgQ9oIk5<mTHjly<G)*Z6^iyv+2gK2-9=e^9x<jM>Jd$cWZ|7ZQXc
z?|+#$E6XC&Z|ghee-*pab+$hLj#f&$haR6ZZNi-U<l*O=uW_|NZpI%cpqKbG70C)$
zM+K3t3ndFLjtSgNT8!rI+bPU;qjY5MYvR{Pqun#*AAF?zJqJ!6Tw$M2n)VX`%;?&a
zs^GWBf%tqQ<p<Z}-@Emmxo`4gsmf$dY-H2Gg0|PGLx+;%&MU1PiRe(AbU61zE489(
zJ+FmkC>!42*xFD-u-Fx_qdI^|Nj~!N-oOxtsgtOa&ArfF9RF)fNJH_ehfwZkRt@C@
zHb8q3z3=Va+@O0oFRw%_>+9OPPqhAhL)%+3-)*&vGqW~1madP=J$Quq?l*b}Z|^yt
z<Gr`pk6Agaq-zygynLR!uH_r*ys^g7G@HA;K*>wE*{*|gv4?I(#s4<)*oUcFH$VI$
z{D7KET`OhObKapZ{mP0XZQ!<X+RLS5jLV%Ww+_=dwW){w>tj@>EJd!GZ$y?kuhtey
zPWQz5KW)b7gp>xIO6{B0W_q)pc%7{@qg1b3&K0M&hHKXeu#~Blj5$*-no;Rz6!pGv
zw?di5($XJKxchHQ|8ua+$F|X=$<etk;N)wjcA1B&{$dWMbZi9dFD|QYNd5Lk;DM{R
z=+!P{T-6dfU-N-$RoES!<qIA0&c|OeUi5u(S)vINB6OZ5=cYZv&!>>ebe$@Wy{WH%
z{p`uw!Sco;C+btji$^mZ88^fFlY~=wD~VU|6^lFhg>ge0gZ9g-G|S{Z0{;5<7xePz
z*F5@aXE!?BKM~s>#@*XEA(dI%^e|;me~wf;#N$pJuCtUWEw>?QwZ)7xI3E*JPP@yw
zggrMFl8r6=k6q+&YdrkDdMd55yF==jAI^kg&*S=^YZ2BBw8+;Fm3)vdK>I#QN{bK6
zf?lP%yWN$ucW@^#z<vPWtDsbNY)z16YlzkgWeVM!Hxz&XNG$>Z5vcX&AF%nbp;2%I
z+sL?KKPoTb4w)!PE6Ht5pOpYGTopCm91&S1FX0KP6{QvBwkBAe7#Wo`21SC;AtCq0
z{upR8BN1o^pM#~`ed)%<)Q|EQJ_$`yguO6)9kdRDbkVv9BBRL&qM#`V(nISZNFS|_
zAS#-QAOo}kf(+4y2r@z&Z4-;I^dMF`bO}W9g&<Z4s=<e}vKu4x1ir#ap!gZ;88~<u
z#m@rOv%qckV?WmzOzCQa{6FU#zI#x4Pv_)YRncD*!GiLHkj{>25W`neQPO}55?ou`
zTF2WPt$^X1+M06U{V;qQn`SR2GhH)%G0}DD2x90m5X97FB8Y`!X^H*JapnlJz*!*3
zieR-{Y-vrfMi868Mvx7`W}6~iZ921?tHYpMDJm$c!hIBl6^^aQJSIa~6oGgVNQBd~
z$vPs?0A}GoidvQK4qt-&Kj#r{Ibl`i-&}wExfg}u2T_7d;Id%&2Pg*+6if+5PzWVt
zn@%k{;P(y|XhM<g)a*3jtHB$C692OyToBZkF9QP=9+w3o5|srSWr2I(c$<297RZ9R
zBF-7b^~{~xYAjnJxLw2Mz|lRG1xi@xJVH=G2&%vyEhX$-%0ok)f9aT8oMJ&Edv;U-
z#W&q&>I~bX_%sKalbFn?X5Wa3PNgG=L1iF_No694g=g&&`<dg-5oCe4K#&#DiY&IY
zCR!tiO=Kg;hG?@*T{K>Vltbk_bYYGj#{>>AKupg&VZ~S4$^%yjy1@##cO<|3_F(uK
zgpBQe+f4V)ERYC0jh}q=tZQ28Cn%aR+86^buNL*nq>(1owsMO6G_BL07OIXt8iee_
zc86KOZ3QGo#)WRSUQ%Ls5`R<h+MKqABtO9^a+o}FiX!dn=i|-u0(h_r0DQ!vpIG!2
zi{4@}JU${W211u;H%XL~v<zBS9;2Y5tfa`ra_lKsIj*XP+E+U?S>~2j)@&PFV|7Cd
y5>tQI|69f3zb0t!ChCy&s0KzRdrZw3dw0^czM(m&<8&RJ{U!E!*g1i&z<&Yyr>HOh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_relu_add_layernorm.cpp.E30A7F678D4C6FAC.idx b/.cache/clangd/index/profile_gemm_add_relu_add_layernorm.cpp.E30A7F678D4C6FAC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1fdd81d44c80bb46ef379a6c65a212c433035d37
GIT binary patch
literal 2402
zcmZwG4Omj=9suxj2onXt4+2w%CJHD{hll}zA_a+MbugQ;nioH#bWKcC+pbyGHT9b9
zs?Dz0)M~S-iJE!6Q!}q-nM#^pSvEJ#CqHuata-gzU9=rJ_t`zq!{PnC|Nr~_9=S{^
zRhR(4<_K9*Y~&Fo8UO$b{V)EAT9vT|0JQp8Qc-1a;NelO`M|}IAXb4UqIM4D<5b?n
zsh4GcmStyj!A6$uBHt!Yd=%Pp_N#sL9R$@6$CdF<<Ho*F7nwV6sb|!!<C)OESM3%2
z**YuTXKYE+{k*Emx3^|yapI-_i?@Y>P|E(VJ$id{i08>$GpC4I|GWmpEyN|hne(t~
ze#7o62z@>^GN5K@MCT$elDc9R7pG^I-gd<%CSbdJkMG>LQCRKI{3KqHd=7ifTNQp|
z)>D~2mOL4xn(0W<_7zo*tRH-KHJHCM|6F7DqS<eUZ#+0SnLetyspEda=zcDWOD)!&
z)Y&{~__Wr9tGyHKr>F>d&EQAxUwT|UdMcpM)AVd>$veWSgzFj7yWfe56Vht6btf)u
zcX$;3b3Ug@(R^%ST-e^pFQb}qim+CBb`FPre0fHDs*|@^X<AcSguOqVMw{x;62Z@V
z<XAdtb_sY$3U;^@I{M4W9D9rKU7WUC#zS}iq<ONr6F>4`I)r9F-5U7?eq54!xt>{4
z_MePeuMbVayQ}IezoiHq6zi`jxJRbHS?cNOlX9Ka&S?P|xAJ3W`)?|k^Y<-Wi8DoE
zR(U$p(_6V}g7>Tq<=3??WqVi-DW&y+6@>D>kll?#9LHWl^nAIwb*bO@aP8=E*(4@l
z1MY^gcn0_KUb*y%xgzFOL|<8R4gRFxc{0(hougBYFG&1S2j>PW{`%PFp=oqj;+Mn6
z?q!`&4UmjB9c*GG6Y3~w9Xv8^p4oiuX{~MAj5ItiWwN0Cg7=2zvXDy~u7$Aui{HhL
z_}?oC{-7%tb!8+Wi7ddkmiL{xl1Z%e^L#vC5}HikM5@Uw7Fk@J+H4PrlPucRthS>u
zebr}__809d>+%nYUN@+mCv4VaRo7h&{yN91#)MOX^6k{INBQ6UL)~BZ%Dt+fD~s6M
zk@svO%!{Ig%HsQnwZ4`q+sEpiQ)g}GCp>}#oH6X)I|&*kg?+nm4zHN~+&<F!&J~$_
zyLh^y;@QIe#RElEc{1jJf8nD?nVjBMFmz_}rciA3X_v4@_QP+{7m6Db37H=^^bjxb
zujY=HkELGDc~|Cfe&qD)1l-}vx1Np)yS_cNH@`q0Mh=<d=_YW2btQlA{__6B;pm}%
zU%KfOWko%PYWtg{P?dhfG+!cho3>)$5_8*HC^SwoRMGUK_DxOP%S=4&9cX+<of4lM
z8z$C&m)Psrgq%0)2F!pE0DuNTU=XzZ{SR3FcJ+^Ivbw>)<7h1o*sD*$qw(~W*ZUv<
zs3W7|zgkXxo>Zq@avMUUp=^+p4U*w@Xh;K+G)Py0q$;FaKvIjr@RUFlJdxjrqU~#t
zKL}CEK^zq07!v>w2#<p@{hxW^^sU_oK?v@SZ+=U2z3}K;EDmzSIO5?}IEad&BIJZ|
zLdY58j1Ub&Lx_%{BgDWk5OTq|Amoa1MTm)EBIJf~TNR)3;_jif$`>FS(t(ybkO)sQ
zMD)<QH?g@<A9#O6@S^q?{i|tw&MNZ{fza~u`tRkuXgm&LiCI#(BOK%@^4ww&o5prE
z$cyHM5QoM=$eZSk5SPrgHMHZAc?j{ze1rrPfukW)ND(6BL-9e#m*R_%h$325lx6zJ
zD`OoUwY+@0L68bBf7#B1pd0M;+`M{1YrjT+L*DNbd`Z*GQd0uIYrk!TgQ6HwEcoi+
zplC)kLI)TJ5K=Kz2pwb`T-9HGk=XHivu*U_h>2nwczq=?+t2~f>7Q24Ppku1H(I-I
z*t0a7V5!g7fo3|;0&Y=TAaXG8>udVM`87!Mld42h!+12r+REAn=c6G{iKomUHk0jP
zkQdVnAr6y+kT=sCA+8<Q!O)Ip$3uv3$45v&70?ZtLaGoUAF2;RzEod?L{!nLn!V%^
zMv~GQ{qiL)5*FND5A|5a37;Us%B}?Hz#VX4^wi`~^Up>#`hYve9S2v>U@Fd{_|>`k
zYN?CV3$9+syKS88G)`REL*y5)j*L+*Ux0XZczn=uLB9N2dDcL6EqUy*&bI!w!FzTG
z?UDn#;1B@pHWc?5iou5BN@wpG9vcy((igW!?^uI2T5F6k!Q#v;%nA4aOX(JdrD=eb
zHF4c~8~!GtkFUs2{1NFxffLu&;s4c=?LMGvaHKlZU6^j}EKfFPBhTL3*27CeBGdf0
OeT>?=bF&N#1pW&(rMR8|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_relu_add_layernorm_impl.hpp.5DE61D830E8EC4CC.idx b/.cache/clangd/index/profile_gemm_add_relu_add_layernorm_impl.hpp.5DE61D830E8EC4CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5af399f968c3d35c70644b9709463cc2a3ad5865
GIT binary patch
literal 5142
zcmchbX;c))7RUS6GzPF?SbJbVb_NlI2qL?KEE>QiYzCtQQ8B1MRPc#=#E2*cQ5Izp
z6jWph+klFqL`-yy%D4>3i+dvalyy)R1!H=AX}aI>ybpY_56r3G`B&B4s=7-}2l~#O
zsj7uxUdwzFVj|-sa16uLssFh6wJYXuFpR^+u!OuDAz|O$`CO}Oo{V$EIP%zw)-B#k
zZ~IwHx_FtB{MEH}-DKGsedpMKokKlNYkps35w;VbkbbspFi-Q5-lKgg72mwuyF_GT
z^sR>YMuzUdjjtO~-^=GG%>@Q)5+6RkUo>~&N|)D%rYtep=e#QWp6lW_>rc4<)zB`q
zv&^3L%@L!EL-!-E-EcGE7upV8Z(X@UKec;FqD=iS>)CmaE1U4M$qROb8!p(apK?FV
z`<8Z;%*e^!v%Nc0l%rAqd2pg{$RCeiB~%1I8{qZl>lkmh@pyjo@xlJG^)@k=em}E*
zX0TpC-jvcGD`%$QvHJ0eM|SU+t3GvtOZ47qam2TZx$ic*CcLN%e(-+7h)Q4Ji?Wy*
zi_#q}GpBx6y1T4l>M`{PM(UzMpP>2qQHkDhXQSrUm_IA|-L|S`*_r+*|K__sp}e}Y
z`zv3$T=X-v+EnEJ!aJE?q<lL^uYJYlxZLeO+}~`eSoC-xSUc6=N_tlQ;D1CzmtLRJ
z?!`AAc$LtqeA{5lcIDG0e+-vqyp7aK+p1tpXmdU9ExfY0meV<Zd|^Y`<%lDze0rn5
zijTz~AKWBfUX`$a;K=eriwwJTOyLEQ_I>c5P42RrtD9U;-#^!$zA)#>_dhC4JK-hS
zQ{^yzHD6KdxHx$3!*s>eE4IcZcW2F8x6Na~aqF(0{>{4{-RcVSX_UC`v2bb7yIS$^
z)P}YPEwG`mfKyYP&gt{!uJ4EnA1LGN_a?-?pU=6F{O8+%sD2~6o2})oPQ;~PU&T8I
z%*(aPs=gF@1Z9b$!}i!c^&?A{y^0R{?WlOeso6$P#}98hNd78x5=+;w5~{^N2z|1&
zn$QeYQoZ46IKyw!!EE)N?j}cd7v5W%ktI>{ycF4Q=5(x4a3?qOFE#Vp1C6r0-Sx{i
z{%G3f!8h>mitRbW(>y(;E<Q8y{e(+G-{E|Zv)v_C=OyLOU4-k@Ys`-FYa-7($GiEP
z+C{V)&)aT0F>1?V;hfXHX?~U5POFgPEBfAT&X3l99umv#zB9eNpnOEnF?YrJevv|s
z%7dW&!D5qF4P{L~xTqL93#Sb<dMQ0~;_v+0Qq$&ItI{_2oVM^N=c8uVe5RbVts1Cv
z-`ZPj9R0(_sz4Lnk#Q}>v29q~hL|N0RCMVq?d>dij0O?nD^f^3Bu|;AfhqkL1dGlD
zN<$+82~1ZBBNG%7CV%+AlsJGXi~|x<tZppqcrObkFdeFcKtZ7M;R92`5vKPU)d5=@
zxjb$Q6}wn?<BVo|OD+)PMKtLmvZv?fOc`HNaDB<%&j_*;96G^;76`Hn9J<)t4Q|~q
zgVB4yt%uFMAnXNiMwfw`jIH;9LmxAr%C~&S6=nKAb8xa2S=1s+dLT}U5nqh-8E!#j
z3(}!kt==HUKJyF7ID$M5s>i{Y76|eLsGeYRF>s5)gwam|_avK3K&=E!8NC#^rHsyN
z`aVna@X$NTLJ9CCz_LgJ))KIx2g-Ff2~4M1=Wey)NbzaA&v0r_CwH<tcERK>Hg|)4
zH%z4m;$#ol_prGa9C~3Iqszcv#@72_avw7v|IV*#u$OL;W42gtTF?4i^L^PLxrdSR
zzSTQ*G_nv*rh#V~_%NnqfJ+9r(`+KFt?zTXsr<>;Z1%MNit^NNf<H2ywhWX%L=7El
zw)2W!l54EWKk7f#{TJo74PR?f!cjyVMH=)3E(WsSR*j30Dap@SRRU_XfRiG`7qLB6
zguFzEwMVsxU(43T$XtvpnDNa>qZw&2EJdW0@pPO5Cn*{YNjcA#qENd3%!Zxa(Vfc+
zs7Ocxn<U0B<Sa~L4vQdDKrIDGdZ3(*Q()8(IoqXx6Qh?Q^)kex*)O|sU-OPm6-1C1
zfVcqav_OzeKr}Ht$ML{-<mA0+3Iyrk<}hnCKamySRspk^VVA-BGB`4P1+1^I`6^gn
zWq8izn*lA8d&Pj0g^2na4cS0x5$EN(4WVOp_#h>p^D&WXsnm*wL1`Gcj48QDB^PPZ
z9G20sFZO!)Gs-JQz-a_rX<^0MxPQ$XD)vy6Cr-Q*RA@n-IPoqvcLUzd<{rR%*xU<v
zFPmk6%h=opcpsY!5mAV=>F-+=7;baBFT97cFcl0^!JHOivR3N2+%tGa?b;b+bOwzk
z%EX^$$Alcdb%ioL3F#)GiS!79+=X;^u{jwTCZo|rN07<LI~fHr{dc1YyOA}+dyxJf
zWW{g_(oJE{SA_hE&;nYg{R1g~s9SW15~LIqrJzT5+q>O4r)<r@$`30HO@%O8$mH!J
zXKdBlO094h)P|W1ojJHTERn0<LX9Xv4kc)`YUO)Aax)0<R~fUz6ePQGWh>vtQ8S8=
zjR@J%HXP}k^yMYqAQcDX38Z@h8PUSgXn))<An6ZWUf}F%5i84!k*^r}(;acL8BJ?O
zqtS;`)lSgD>{?-uw&l~8rf;bd{Ths3GY3qN{b1D3<^d240BidPK`_YXArK6)`3;!8
zfzirKkd4Tuk+~ZjOAgI$Uw1+OL!~u24x^U_d8IWu!RBHxEoSGRlVEz1%_U$~0uJ<w
z39=MSOBubqQGH4&FW@~DjlQP7j<j14q@RJG9Zi*%&q5Z;0<5Y16d^|ua;9gK^IQ@7
zl3^*bkg_inQsgCN(sGqBz`d@cx|;g7N^q-$ne;${6oa*xDYw;&+UyGVay6)ICIKZ0
zI|M~YTZD931Ch3v&CN)q8L`r&6sb#@kge<3SjXGbSNh3|WxeVPNz&?-80D=0dks9R
zs(ABq=+il4j*A9y%YvU=U$uYorCgXAxJ<#Pu!CAkD$GfR5c)8FNjJZG_sI@EB?w)G
zb7=Ruwzec-_|KTQF|Y8tn)lsdZl}wb2R55-N=}>}#Qoqt-V*qNwveC_f=ewLy)ROJ
zhXTZCbOi*J5T-XHPo{qhvT0#bj>Ju{IdGnS)+gWdo7fuZ{_;~dsI5H*_UB+8ZDFgy
zIfZ{)8D0O#ubwzbyavm4;FFu$v9cn>S<~}Z4pq~DyNi9xPXk^WI|*h0FN5KZ^ooJQ
zR+0iuf)wZq3~BcvNDD&?bA~1wPGpg#p(TqZ8BSu6m7x`jtPQPM<cs?%Gjsakek}6G
z{aG{{pUt8}fFA;7diM!38}Mv4=K!9==3KyY*_;P>9-9vXewfYqfakNh0Pq5aWi{J^
z6x*igQK#)V&C!cqKwn;cqivP3*(X2iq2+ix=DO_~bqx?9frzP`Z@QF(>cQ1Xl%XSF
zJOZ}#j=eo~@S=_DH4!CThZ)x)oEF~w+$_j(OUR>Q_WxI+vAPuXD?D2Mb03RayELv?
z-k{#J(T_IWz%YVolA>GA0)1UhjS5eh7_XwHuA!-=t;)rCOc!OQIi6|CU)ufiXwVG|
WlfMD@=TV~@7^ccJRhXtG)BGEIJX$ya

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_relu_impl.hpp.9153E795377986F0.idx b/.cache/clangd/index/profile_gemm_add_relu_impl.hpp.9153E795377986F0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a6b73a4747c619657cc6e274ab1ef0b9043d4cc6
GIT binary patch
literal 4714
zcmd6rc~n!!8o=)tu9(2hcF}->$R-wuU<e9A1hN<;ga#0NYPGn5g5(&%1#zuDK@`i9
z0!6J0E|e;^?nOnko<75KL7xw)%JU@^tqLw!j~iOvjrHbEkLSF9_=h<s$^7Q~=FZG_
z=li~yB*m)LGf50n&W_F1r!Aa;V;Cl+J~<1sGu&Aiwu6mfxmz2O<8^mN3NGoIz9^6O
z?>M$RhF$pG*n@@}4WAv5>i55iJ7%{*(6eG%geNaemsR3b{98lr?e<OOoiFY$n)su7
ztq<#2_pk-+#m%>UUtjGwHC$x(t9Xq3!RqiQwRabunfm-<X7p6wmX)RImcTQAE6ii1
z*04A$jbF{ToSfCO;M$tl%1Vc*wKH7Rg|jsIPKPJ&_<A<a(NOO1EepD?>H0CsvAazI
zx&76OjoS-$6-(;bO%t?%%4Y)OzLK|#^UDoGcXr@YRQG+qsJJ%wUC7yq^M~JfWGV7}
zZOJ9YRoXQ#l(A=8=A7SESZjiRaXm&1EIoNgH+6ga(6`&#mo2%oInOP1d6C60zrbZm
z(6!m6z85k!*L`|u#iJroN73UAUgDeIa_>xEexkUa;@c!vLRxIBc$s=i!|Jmt^7f>3
zVTvLk?)s^Dd`axmX{_pqp11pt{J~%6Hhv#5aj<rzqkhpQO>X*to2!p|;G$7WwB6+0
zQ1)>*zI4b0ei*ytz{XeWCf`qey&-tcJf+sS_=?<q^z`VwIsClp(>2>}yGAAPhn7o>
z*M&iYvp7z}(+asy^LFzx&xhr3s|Q{gVn0iLJtOEawo@xTG;z=Vc?I1oG=H2hy$gwH
zeZH~P@**JkKh004{WfH7Xs)5Z#A!4C`B#fg$#df#)c)Q$#qsgN(J^Yjg3Mc$d!p_+
z_INZ)hOM6+-hDherm8dOfp=hK&Qkr%Xll2(%zD~cYi7M<u?F^|kSe6eK}2GlKcRbc
z%a=|!E;bX`0DFw|?B_Z3{TIez0cM8-3E64sIYE3%Iki!KS?b3Jf@}r<R*=#??Ie!1
z+b_;6ViTm?NA5>A6QsgNfsoQy8NkR?!KzS(N?=q8gfm<Uqe@XO16di!8MzUZMi@`C
z{pH4vPX8Yq><CihrLoP6AmhB^5Yl>UL+CEtIfFwp_PL$0!^tL)G%<_8$u<zRfeWpO
zlNJcHz(|?}_3nCU`S@mgYAl|{lWwNQ;>97P<!ikdnU1GJWC^?kgcA9QHcjWBR#zOp
z{q;LXoRquDhtLn8$X_mINa3nLWD0);B2&64-5E`#SQ*HW%2kE*Q-!H)7RO1$V1pZ@
zX%HKbeg>H#l9su&9%R3*IUbDTWC@HY0Yt47MwFsl2GTOHSsW*gFv7^l2VeVn0rpUq
z)`!>XI{L|yEyIedKHBjf4#kFbE@TA@C#4GMINCNi85$lsmL{*LBlXY1t~m7CQa<hV
z=N&>mPyT2cLAqTJMGFG<e<TQZ@A~IF`x#w4YUl9e0(wR`so*IPQu34tsdy@c3_Jru
zCJ>oWpD;m`iSb9AG(oZnCNkqRflm|oGTa89ZQ#wY8F*$uo^OVcW(Y>*8Z&4awt%OF
z>90Pu?r~OU`gY0(0|kL1T8$#lz~Kl52Zk}!2*O5iM%kwk{24wA!n2HAQ`J~}d2J_`
zMUXRznL<V_ok+J)d}!Hx(TJiFD()o9l4BU9x#GD<aLW*9*eL1WyX+pGTso-_@A?Ux
zaSi%3!I#w1tAJAld}duofqN93XwJx1pX(_&a+6X_k)<f<W`fKVXWGJUMt06(-K&~t
zoS@<bUIJW3trYN5l*<4wL)i$pkzsvBrc3xO_j`bom%!r^*g|OTIZJyUe)SSHjlCe;
z3&?(`17RJ*tH!_iyliAp3gs@T(o{7)ku`N%eabgi&r>tp4TE;WC$xYd_rRb%DDQ<a
zdzrLDkTo!<hLLa7*!}%caZeca{5o*016x=xcP<dgR}NT7O}z>PRbUIW<u9jCTf=v2
zp?WmHs0PNPwkiHAb5D%p_Tg=^eZ7p|mH8%zn!5@7O-x|l(dnOX&iUm}l;8vmIsqQE
zJ$Gis;hMz!U*G3kSrgcDiZ$nsYXh4DS#xfr1)}MW)lUZcwLkQJL%pGL@F<62w16Nh
zz@q}?O7N`&<e1R{vKGdyyVp;uy1VGK8|AP%FP&|V5@do`0z!%2iMF)$z2Ci*&$<%D
z_SPULHNhx)Y%6O_Xpm|GQ4<;@%`nu=Y*U;xLxdS@>A|{4o`ng52dDlLTwTN`t*0oC
z9B~r`M}d%%qeMu>Q6XgD7!Wdn&;$eN=@OKUg>5LCfn#Re>u5*bQRfZMzUy_is^>ju
z&NpOY)a(E2G4TYCedQmM+T;81o<3!2nOqm~``um};dDsS?0(~_J18SaA|<v^)2J!*
zdUVNPbqz_Z<Wu_HT*$6X?7pMV>BE+}C{VR_^E<=)@Dwk5FI&GRasuD#K*93Sv~j6x
zNU#DSWrz|v6{w<BDn`?wGT1H^1a&b`HKEbI4dUA%k<qq*+5#Gyn`OSslXPwsz1}<5
z-B0$r5#Vw0BL}jCH7j54cY$7y(2;o{WsH;JXDVY#oux>$40R5*QHSxsi*Ev&t%s}#
zJi>uqCPB(@8A4<6u?R)tkqAZMQ3x$!FXGWpdDK1iRs5FxP^%;GCAN&+b!gQD_Ae4w
zs(FcGiH_EOwza-7=m$@qUN8JrI(`?6SapF)#3t}Efe$^xn`4z{10yaRrUur_^{I3-
z{R<B7@5J;?`rVn!zMEyaTmP^9u<W#~8P+RyPWG&vE!K*_`t@hA{)_XWe7J^Ni7dHw
zI%!kZBqzeDKZor=aCv-z&=GT@RWYnTQ{^&M9#bWlD)r|OT^#8OhB+`*>z)2X`5%D^
Bw~hb+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_silu.cpp.67ABB3D17F5C92B4.idx b/.cache/clangd/index/profile_gemm_add_silu.cpp.67ABB3D17F5C92B4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1a064c87711d1b575e0eb9752155e5c21a716e63
GIT binary patch
literal 2536
zcmZvc2~-nT7sqEHK!C6XGzc_`AtVq$kdTD11(3)tBy2(135XC;PzVJTB!VWiR<^b*
zrGg(85tQ<Q%H|P8)M`YqN<KVV1PU50P{pNu#SYjZ<7wxdd2@g7zI)%id;c>WA2$0J
zbpY^8^brd9`{Gal0O&z?=Dv)SVr2l>5BY>OEunGgV>I2L>tv^&S~=bPF)QEAtf|*1
z72A+{;<!>kty#Z8(~0uWW*^Q5m&&jCXj<W8N4s*n2D@_yGd;?6W%Wa0lV&w%8~O9^
z;;gWL2wx@By1ToUoPTPaF8hfvxS0GVFXwfCUe3X+$%{RV4Cm7kj}CnCW%u!RdHSJW
z#>ZmX)80=V{4UBGH==IBdy~?3yRz<FpLF)}qonW)b3LkgU6-B}cty&pMs7)V2s?a?
zyZK+M6JARxb-vr_4<r@8rrnXA4-9V^&N)+ZUm#6CdMNi)7=`rcH&oGenMX>Nnt1V*
z2rqi@?z-HeI-f+lWa%wgle#`&`mE%#aaMPb1(!Wx$C&vL6cH{lJXNAG$-@b0m0{=0
z)kSYdh5Ufo%Gt*Ugxl=eZDwwGYEAc}h-x>COx~M~_D|OQKTL%eEIqPwfmN0{5g`T`
z$xTP}!Mz6ofy74l);Ynuz`^4;?0%@y8!t@dpO?;>h7|AGwLh&)XJJobkM_7%>3=#Y
zCmR-zwo|Ru6YjKjww%U0Z-`vi7|9^--?Ma8I?QGeGl&Jg-}F}q?x)1IRJ;umX$&w=
zlMiDfP3O>CUtTGSx?1?;`^49l=3BNs>OA(RX{eDT@NFM^^8L6h*Ai8<a}VRq3tBA4
zv6SO|^<}{jvyhp2=A<!4!r5l-@|+^KHQblueuc9`xfVCww{s?S*M%g(iw8qt4E<*L
z$m4`)1OMhu@@TN!>t^ov3*J=eCEJ`Q^9Sg4y?PgSQ`t?~yLo3*^q;zj1B{-FT?(h(
z&sKhMEsghc8-^HL6;+&lbHbDn%DfkHqTYerxacxdk=>bK+APG9pDyIh*!!aC^$|8n
z@4noqRc$yw|HnPw?Uz*ld$&5wqt))bw_kJQk!K9k>L7i~xk5~p>;ZVEqPnpqdvC{%
z=u;PIjs7HD!#rAIbSt1C{xb3PH$4KsgHPy>Z+`HfrwY>(i4XS|<=pNWe(6mMDD_yp
z&*P1pYj|~!5oJ*pcIL$9XG>r0-aSw?jULdGPq7?5#@!lCWTzJEj?K)beEz)Dg%Ryy
z$JG^toGj(uUCgOl+M4yXtY}fC+R<iBmh3m4O08LFr0=mYLB}+!%>HPiCta|3{YH$m
zb>64I$KSYU4pl^I-ZbL<%)nA2j<`<>GE7@j{K3vY{`Y!b?$v}k{={VHozg>KS6uiM
zkeMTh&xgLt_{06~LoXg1YXOk}z!igl7&QL-0W5#J&{eVr+e+`)GODSx4O*h3s-v+o
zd~^c9NJG+CTOz+fQz;m7YpZH&tPHSwRC}1GU)dak0k%VKJ(S+Ml>v5V@RKrS&_+z%
z{@RXjvzIh6Tr;$p_G%QDh$bS)9Bqyu5}Je{GMbDa3$z7-EYX$-qM#`VvO-%S$Qo^p
zARDyJDlzf%w^CK(=0Ox!1Y$*?0sN8?tp6F_&Fbrd93Kk^7d5KpTJQKveepjWF~XYk
z*xsHA5X05c(=miA9Mja)+01cF*1~Y@J?z<Vehin&qHa-;gSmsHf@tP61kugu2x6Es
z5X8hWjTLcDI41--<D3!Xig(?pSm}m$Ll6tkLXbP&eU;#2gG#W1{_ZkAz`eJituY=t
zuC|u8KCGA|tVp(?HbhcduA7M%B)|!lnJ5OWU>1E<-l=nCq!x<*SYTK^qJ~4C2mCNF
zh{ABA$WeA<J@%t#`V6fx+-Pz%f?~)q2#O`gu2S)wmx<%u5+f+Hr-7#td}g>Is86Mi
z;pT(5d=OEne9$Hz+zf{;`)T<g6Xug?ekkAIqNgj0@eqNlJ!~A<dpij&{}tLceq+2d
z{N8OlMmN`&=wP8~5vU~s_25M9m8=cA!^6G*5Yk0<g|L-1J*tJ`+HbY@gV&?DR4=NJ
zf*dFgn-oN&&=5qY&=JI-Fc8EvVQx^wIhi;i$l1ghL9PT>l47MB!3{wy0t-Rz1ou_y
zqjC~eyz1r9f!P*pJ2-)qkbCs7D_3Ww4!$BV5O%=IQU3HFgW=}k^HycJ+=cb|AOQ}7
zpGq2^{CSqOWFrPWVUdj;nR?haq4FcNsV&+T14qHVfkc|OHV+3+qo7UJLTZ5dZTX;D
zj@h`uN{5hXfee0H+_GLXGZHen%SOoZuyQGLzjK>7E5D%qyJ~nuSSTk12!mY!z)_eH
z3Nu_`hAPYiL1J18g#O8a$|x08HMF`WMoUjuN81<6-a^J|`05)PtXXHobaHWZW4U|S
zuC;bHV_0tZw^^LYXL!?%1R}|TVr65u+1`QfxSnRbiR!f$XYTFCQ`#Ej=>rA;e*uHM
B&9(pl

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_add_silu_impl.hpp.CC18A7C19DDCC841.idx b/.cache/clangd/index/profile_gemm_add_silu_impl.hpp.CC18A7C19DDCC841.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9db8161a1f7d82d9b01167b94ad2bae42aa7656e
GIT binary patch
literal 4714
zcmd6rc~leE8o=iZBPK9R!k_^G*~PLOf`Sl%EK&(*Sgl$WML<D#1`tKGu1^%DvZP>f
z!>WZ+#nxK2sEEt4hKiuIXtf+)QoyP}Exz|$K<k@WCv$r|=l#P!%sI)qzx#c6?!8~W
z<=(_aMnoi27$#4N%+E?)z7WSS%#!}(F3(AK;9%H(E{5grX-t@|>YO0BrfS|$5$@S@
za?ShPqA#Z$Q{QR)@TfTJ$UxLdlWl_6AI%GKCR0_Jr7k7EHRj*z`naO^>4UtP-$iV4
z<2)H0m)%v;cF+CAjh?gPg(kmxPnJGhAN;tkd-?gf|7pnxpX=Vfwk)FE_xuM%1)QW>
z4sWgIlV$odi(Y5n+!(oYr&-vhg?14|i<E^{$7k>VG=(%*S9rQg{B9}xzY8-T><~fz
z$Ozf?>MaLLL=D{L_-J4G6M^P%>B|*`6>7)Ndhj`l2ksjxZ!UQiaAD@M@pt<4#qKZk
z`IL86^v0+1$n))suN*9@)54DyPClc{&UC8gR(CnR+}E{wW#_H}`=m9+dXK^_Hgo)L
zrj)r~P2W{N<C~BAiiJJJzi)H#zWaqm=Yln-OGe1Ph~vbhMn-zCj+oQ9{(^$KH!ICD
zQRWqO>+Dj#D00<2&Vi8EFOPitCx5g3w8O+q+vo}AS$Q8T^V3G%U4P067fxImJxFy2
za!=Xw#be|7LEO@#+n;Zq{UGVZHvh#-<<XiI*QKVD7K9fp<`*0|SG(_?U0597u|lM|
zW$9;|$+H@tTIAFD-)Uoi+L3+VR^hcjg!sPr$j$j_AgfACsaMXvO+Fj^<YivjVET@u
z53Rqt)b~Nk>1zF%){yJ}_(Mk~y9YQv@aQY8yI%QY{BP?{2s<8T?kJ8U+fVvMt!uIV
zQ@U@FQqa8o>9P$w{O=?Ul>64?uF6Udr@h5y(I35aI?gi=XY>df>4y@U2`P;C7t~zN
zZb{?z+$XS6rWoZs!rAf77slfNW`Y9+IjL#6etcRvsY%)@_TU3S-3HIwAZA9Ih|KG%
zTat^p1SNHodNADtC3BM@BzKp4u`-3fB9Nt0m{<zIESJH=GL*|fQVvp9t^v6QrZH^V
z+SJqQ`L&q|K`C97#&r=?luHys(XP<}%n*ykwt?w~?a!OwR5OU0+0Eco2M9aBhEc>R
zJ^1Qj0>gp^hb(c$v^G<EE>h{tbklQ@Q3yrzqg_~;ic}%87%~Q-SbnTg)B582%H#Jw
zePxbQQakBb<^?o)O1)W<*~t)@%u|NQ<o0q0R#WaR_hm_8r$ELjf)qxJ<CNM~ZO>||
zz17G#wL~4t$ZT#O<G!pt<&Wc3DfpBEqE-ezWhj?}xEzcY$0-f?Xjr-J&CA)?BT4Eo
z-mBfD$18V_E2(<N<DJY()SGRnN)AqmW#XxfZEz|uIB*I>E@9s`JPEpPHe^e=bv0Zu
z3z(nq&NBQ|2OkNuy^g#i2zKcIVX5iDev<YaQYv6pgi|t7hLD_;Bcvb|2&qXmLRt`N
z(U{Oen3j!4oYF#q7G|>ZG=p0+xU<{=&K=;&vJOZcAn(_~1ReOJa-|NUS=NKIo*f@?
zcJuF<y=m2S4EhRug^U_a&c5Rj^7jp5sR=Bbz#3(@Ch%nW0$5&P<;tq2lGaVV794^~
zCXy{#wKO8lNYew$mkE7}OX<9mAW3+iRa)Y`1W9h`-swh)`{|&`qcf{!4P&p5=S^+Q
zN)_y&w_XLjD&VvGIsq0Zz>47nr;AHx@bA{riiwg$Inzy08QvMjw40ujyF&H6HXJAD
ze1VsO1*=vDcp1v&fS04J0bIj!R%M1w@GlPc0jI8k(={-r(ELmKt^)kVHF_B}U|9o*
zKh%R|J<IE+4a_f};Fm~;OOiM#f?3GM`sl30y*IAVD?9{a4#9hjfS^8yF`uJc1Cwjm
zvO`d{Fs7E3Z&#Z9v#;cJ5dHpou&W1STCcF)B9yKjwU%Cb6$q-pm}o1WEtt2FZ{JRj
zXoQK4Y((vo{Y&By8)Y$!-K2V}gx{YrkV~&!3!YjwvG4EojJag}>>@2V4P#D&6JyWM
zGNW*1Y~inOYObUij5WnjbEkHIF@Ov;H&hSd%)kSWM|*TVavh+5p$c%SfN_j~pen(s
z66KxXz7vpRMmtE_S+gG6I;*NX@0>jyuqqdo(MJg?#w7-!Sl3u%*<z*wvp2^g<}me@
z$Nj_)`(wRL4Y^OOg^7%wft6Y`cQu2s8O@zKaMZEhi&HuX(SfmQ81_!;+1z++?$7=Q
ziush`aK)1%(Iex@5R&uc2q}09gw#AWLRzrY!f0l>1Z{fD4wQAk)3H%^qNm`5^|mKp
z4!MXmkPq7m)fpJ=6@NcHo+i21zn|4LeHgp?S#A65`jCHj57`Lk0^(9eOg+#;8$lE*
zGG?GAWsytYHMJpPM3JHy%x`lwr!ID|Gb?u(Tc$;zxIM`44IaizY;|3|bz^8e-w;6l
z(n*YQ>D!6F3?X@d962l~!W9ZuQ>{=N?-~SsUr;ooMW6$wcR(zwt%nFbC>d^(xUY#*
z*;fwPcWybH8F9zUspXvj(g(H4pB=G*A<xj0aWrwVmFzNINX6D-B#8!E2O6nIbM)z6
zuQtQkD+CWQW41|95?q4N6nqLop?D}lVR#rqdE7jbc}w5m+~?DG7X}&vfiE%^^8Rnu
z#dCia+0orAWh+&T_LDsgO@3cHyA65c&(rX3EM(nPx;|^cMGJ1s3<D>3UhoaMdYqm(
zOPZC$biX>^>b^^ozvurN3d>2&Txhsr=jJTR-EC+H3}1gehJSJ1K5wt#2C`!MFl;2-
yv|yVQ`3vUPZNGOCVKtJ+H6tuYzQEEPBUuglA0g&3XBrqruuU_zX}HtBef|fV9*ar<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_bias_add_reduce.cpp.CF67B6B3D02F20A5.idx b/.cache/clangd/index/profile_gemm_bias_add_reduce.cpp.CF67B6B3D02F20A5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..07ee4db7a270e9d62a50eec4447a37d8ade1fbaa
GIT binary patch
literal 2636
zcmZwH30PA{9suz7Fp-FXa)(O<!W{^ZKw`q<CJ^wz7cRLfAcqI35)uv%pvsM0D_|{$
z@j_8RZe0r0l4`~5Cn&BE5m~IPb^&1*Dpjm)(1y%zcfK!g<~RS$yf^da|2Td=KHf?Y
z<Q3<a&W-2CU?2!mlm1fpynVUrAgFW$1f|zD2XhaO(^Maym0W&q<2=}&n(b!V*r#_u
z{lbBYG6hzh>40WqMdhaX(fM;`qlJN;QOK?G^yOFeN2VFZ{sl>4`$<abJ06c*J$jU`
zy`k^hx73=6sHm*k_|e6|>5Aub)Qo=QU35R<=FExw+`y_n(B`$P@9l{d$<L?V4*l|<
zQJaX&p0kOe;@f-$R`s!6HXYMHTpL<k3AiEoyRGA-g6q?#OSep#xuSh@EaTqr%gqz5
z;vBWZGgSeXxZyLb7DhxHwP~a`u!mh$(L?%0>&^8WkF3)MxB=07>ujqM(u6nN0$zFS
zEwOMtIEJYm)4XbG%X{ovbU%zRc_4LsCn=CM<*;Sxw{x8>Y7f}DT6a6h$+T)7`yOT>
zEvdJ?AyiC1_^j`ej^zF;wwEno0Ar?jSMTQ0*d$|~N9Kj&O^yHH@DB^EZVGHzP-N?)
zfkMdHExX%G3-=S9cJ$JZzTfKviCP!F6Hs=&Tvl!OqLnbK2lEpS>0b9rSBtyECG&$e
zCx<i;W11wi?!`wZ6un8JN4WeQn<{^6?da6rQRwL>R6l<xn6Tp5O1pQ2pW}bKM>vKt
z>pVWCR{HdW@_4mkpU1r`!T5suD_s2<ud5X8s@H;M%Oa8Fh&48E@6Ii6W*<j$u*W$#
z(cG8Aj1$Pl{loMe?Fg4MOINNpi=v08vDmH2%CnsdE9Zri|4K~^x$^~{n?rT@#&y=^
zpRB>kwD`eNrk(Y(b#=bN<omf+?~3NHy`3$-wf&lde;-d2%bHQ0M{ORKhc@Y&4rQe!
znI?#@ZW5O(G)U$aq87GbE?ZkXVRHEE2WOw<XKFc1+R6n>n4)x-&F>1-2TCkY7wguy
z`H9T)j`e0%9w3Egzex@xzff9N+}gQ6=&MqV_6=R<*1O#urj9d&xxA6iMRH5ExL9-a
zPS8&;VjT@*&<iOeV{Mb_hqrxx!|dDF`tvWc8q)Xo655Mu%*9m;x_J~)LF!+8+qtJS
ziRDa^{yPERRqyg&Df^nktxnZMv^{5>oQmr*VEeLv&A&*j@qWMT=puMvs5!ZK#%k$c
z5V0n3;zQ8j2VV<vhwU<nU}&8?*L|$|Zi}-Z+^=xui>{aMr>INfyovX7V$&?fH+Sr?
z<_=jpVhx>o8__o6#lR~zl-LKg?lG;21|KB$m4n5xbJ#NF(<P_oyU1?|v)OwNK`9yB
zSWjtI&Wf??&!&FefQP~$2vxzT7=64*BOlAZ4TCO*zK;|tL%XEm1{i99_5Y}#)x?$-
zUVK7ICKiG=t9;2Q#Qkjw#-l=5LkJtmrZ(>3?VZg=@?eZKpdy6TWC5n?Kwa0Fv=1Bc
zs1eRcMb?c+jd8|+OmHTEh&Uo3Q=BOvGn^S9bDTLK3!DWYOPnPjE1VS|Yn=6(cDX;%
z8A|Q%kv{vU#mhD%s`mT7sX2|JKLvI-Hg;d*FcOvVDA|YXFJF|Mw;fw9dvkkhxg5+L
z08z{-fT-qFKu!cFefc;Vfd+_9paWtUGEC(EWEwI7IU70yLJW~LC8*~}epVc|P*`2C
zK?6*bEqK*t4X};OIpvkzpM5t{Cw<+YcIM+@kY4f^SO$!tlX{wX)C2MGls|$e;t9wL
z@dD(HcmrY~EI?sM7@%+@98d%j0VooQ1Qdlt0ZK-a0r3zXpcEtp5Fg<K5+DLVg)mYG
z11^G*BGAP!QVjYejGP2r0wX1IUq42-_%&q64I|wuZ4bJ9tHQLy03~ZD11f~I3*j|)
zRJ$9gK782sr*nAIrz*Y0{pFZ41|@GL2gv5epmu(C+vTz++1ttGKym;?AyEKPNmM{i
z22N`UjzMV#G(dC%Iv@s-VIjW*lgI?*Omqf>5Ro<YJ96T%tn)KCJnD=3hRG(tqimE7
z$Pe`cv>V+GC<cuIl#ZqY%0M#!RlsNkyyo%T>CyflvF*~6O~R4zvRlvKn#vTIt>w~}
z6^%wm%DVAr8kz>E3`Wafz~wMnF1Kuyg3^C|W><5F+yMK_l1o;@3jdQ_%auPUW|&TC
zetg~F2wr?r%xVUv@b;txN^2Okh!B-N2`&(<4-N?m+{uABtKm;&r>p@%LGosZycsNS
zu1@wx*~5)X+9z#qOWd&zgI%wP+prO@tfu;ziod##x23w0zlM&cmaZPt*~Qh(-NVyf
z+ZHk1Vr}&QdI<)b42?~QW)@Z?8?pn1>O^O3rs>-`va|{2?5&{+z5!l-@OJ2T&~gst

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_bias_add_reduce_impl.hpp.CC4023F2075FFC72.idx b/.cache/clangd/index/profile_gemm_bias_add_reduce_impl.hpp.CC4023F2075FFC72.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7d54811a97007c9d5196f7d8b81d221dc0ea1dd4
GIT binary patch
literal 9850
zcmeI2d011&7Qi#dfCMfiAYcqz$YoO`s34124Jc7S7E#<+6kJLG!3r)<0T;lczKT+<
z;QFdvtP5>Xii*#r#ky53BA`Op1PlV!f`UHZg!gV{zR$n>$G_wY^W9(0xpQaE%$YfJ
zZX$z}$|NTqZ`|UbR8`E%g?t{5XNv!mS0*hx13ca?BOWg`|9q4xyn2|ZE4-z@XZohR
z3yO3D`#Y<oSJIYgN*aHizWl2D-nLCqS#QTwgnd(e_>drD*OlTYzBS1glC6I#`RmV+
zZC@{&<sZ9yOo8I+h_ifqWzL@D;0@sBaW*RDs_^^Uqvr4L{^Qw(&<Ep=g_xGOD9lj7
zUxL#OU;eni+s`ljjm7X;tiEgXIi>Y@kNi4G%Go-nM<Zj;Mh9eks+qj&aY|vt*ME1c
zaIpBd^W^0t*A)$2>>YD=)i;+cmU=B3bb3<I*drBxndjb!MypONP3vu$I`~5VrSfrU
z`lxUc@0Nc^EgIWl8lSPEd0X`Y=KKnq95&(M?3*t6{hNI<lxS4;!4=Kpyfd8URy5vg
z+uoK^w9jH;eZ}j;mVq7L`I(-5yeI3MwPu-*0$RiJ+b-PQ_~w@K?{2@zk3MC;OYJBX
zTvmxsT#lIT<)u75eEYVb6yclRUd4cEhdzHSn#=SwZ_W7WhCCv%cxR1|LH|b<N4Go+
zF4<h{(zNsY!RqG!qy%|N0|p<RXF1kq?&!}hr3wtCW>)PD^WR2n_waCE;$f0{-zvUl
z<Au`X^yKW#S0+h97V96iG2MH{P4+x|M)rWpIYD{TT&~~ybFASHu{j%0gc{v-*%pPW
z5-L26Th0Y}s|)H!%=^mW+8n{erY#ZCGj^M|ntD3z?AW}}>Dn%rUG`HR$GzBeE$2(~
zokLS9DppTkr#N_XU2f*#NddnH%~{fRrrkN~zIAv1@%z?@F5Ha0vg5|Z-y1jX3b}sP
z&%tc|g_wu4{9C#W61~dDyGI|=zdNj-ar{2@4}nYK@3^k1EYC|?YFTj8{`&CdqWe9E
zk1YN=wdMM3QQP`qmR0wUml~shmcZ+?l4If?-MT#`$!dJjFH@{SyAww}JU8K}?~(Z{
z=cV<;)Gsj&^SCko-{<3!y^eT4vJ!3^a@D2BXx`ch3@>@LDw@Ib*I?3<sn^H!o5S!-
zEb+9)f9w#wA<K9IKs;wN#3Iv^*^<vrv|cuX{+2v;kp3V+U%?afDPslkq7Kcswa^XV
zG@$Q09_wZ3W%hpsUc~Ra1BOo7=WGN<_<lQmyFq;gxj*{mpG(dE{&A}jFjzsNFR|<^
zh(b#3xx1TA-q>RV`dGnA-%8L|5QS}*dJi5~`$+*DEdJjDOG8Ut3vz!uQfJ^%-faYi
z_<mb`TXA1O?myN#Dt4R3^>c<N<ZR%k=SJD!<UhQf-%Ptbmf<;?^H_g9f2Y2JC*b${
zpOMO1r_L<TPsKHexEdVv9H<(~=k3g{tlkrB1OnU!Ykg~zzJj!&{C>tK{k>a1F#=<(
zV54ut^c6(msm;|@V{iDjFg#Q45zczf)FU)sU6beLUWqgUJzRsmzP&?VL27s-npXHI
z;YyPc7+?jlzF5##5QU_e*yKtnw&46y|63k|h5Z<|609o0LDzq_3aqMVt_G`W!u^ID
zeS6?WbPmI?aSm~V$!#hJ6-DvG;ztr$f=|LHgc5xcDXKstE6^w+eSk(jpt%x_t3(rt
zyb6u3LjHuSkxw<E>U)HIAE7ZsUW0sV2<zV}eX1F9%8<{nsj^fLqPAMLnxZxA8c!ly
z%dXWC=S(2zb0ThDGVCT0ZGx5DO~y+k-pjtXz6gBw8sLv3C*ZTaAn%1?91F#N1`NCR
z_*2AZi_n51w2%{sifvUMyC>c@z%4aNwdcNO*wrShDOw|3BO|i4!nG7_Fxfz9rXiCw
zL~%M2rXxx_0|_$-b87-J>{|=Ceh$2pmiOc1J#WnTY@Ah`4fkYBL*j-KQdy}enQDlN
zlEquaQ?hukcpq|Kf?a~Hg?u()OoD>Q673QxS>l*PN>&4tYhVh|t%ZnMh$LJG;yREJ
zZUAuu%}+r5gyvQdx6<4O;x?K)K-@uCTyfgq^|3P}`t#We<W_;?<UtRR+XI>_ky|Cr
zRmh_X=_VJStw!$ENH;n7>_g=95c!h&9wVQ}geCW`uHdzfj`=tm)M!|FwWsWJ+_Cq)
zmo__-bl+@)8*0R`xiBCXq@3BMhC32Bi_P+|Gfxkg9>{&oXJ?O^{Rv0*za%Zyq>bK!
z9XS_Fa>0}n$j1Iu{M_dc!w)_?92$y$GxVG{^}P+@9)6_Na)s;h_rB35H7kBs+K|J5
zcQ%Dh2)E%e>~^HwjwW*!8D-n~`=|N~J1z0DVG<|ezUH$k6BR}ACh-&{m?Th?Xp%@#
z8CaHqnAB4S@-ni-^4T&NR|eF$T?5N&w0;eE)_@nezZN`eX;y=$n&x`&tf#pNWKCp?
z<g+axYoS>KG7Zh`AZw?&3oN_n_Vj>B4^aK<fnhx`g0!y(f_q>R;a;%s1(st!SKc_u
znf``&=cobm8lcwrS}?DrSq<iDn(M*5p5{g{Z=|^y%$sR`3g%B~eg@{x2#-H*RA=74
zVLEoGVz4PD4yE+$%#|H6Ex_}{N?{ekMSl#NgOoXBo+!_4eio;RJz&VNq4H4ODv4=q
zsBSbFHc}L6$@LwReWWu*BO-O%3&SP}l7vKd9!$@}K$;!S!w{PN&%*@5rC?f0^n>@7
z?zpv4V~m|dB9%CDAuhw(N^Kpvc$8u7q;`&6d>pdY?n3%a{}tFCg<w$#E?nKA-a8X4
z-E#hk>#|l@hj5W3!)`{(&BPuIyA3I~5$5W}d$cy_=Jx0*g23~t7}0iY-8>NF0oCOa
zFfAcnm>Op0-2BC8Jw1kX6gUcX5iaW@a1rW)QPx%9D&)c*hLsECLR~b?wjx0*(mh~C
z@1EgJB^y2831&F!fMH+Y&$X0cWxlc?E~<@LIW5ORzH0+E&wg;)54vqHRvhzHY+!dL
z7L<Uz1pGL=ElWyX75?&=qCbN-UVa5K<^&A;0P!EtT#5LVG*=;h70uO%UrqBv#D7Th
zW5j<<I4(EZdAvF}1rJym5|<$=e7=iZ?;>x~kbB7W9${79a`C8Y+a_$Gn;^Xjx`-zE
zySwLhW-a&!?&B9=`UO~UHE_<+&#>=EL`-MrWH*?b0eGId+Q~Uey{v9n!@qMA2f4XO
zkc&*Xx)?Se3G!(!K!O6A)kvTwobkgXRaC*9->`kv$X88=;dRKjj^_O^VLwbG4>|x7
z4$yoMf)Bz>B0mH{hhPTb!!Y46(cfHElz%Mk#JAWQGMQ{7H$yTz0<C>-=SXnJ21y4w
zaWN{xYLJ5lQLd{&K^hcJYHvdhZA6~)ue?VI>ss)T<M}(ImBifJnhRG<@qfDGy@!?a
zckGGU!zVXLy9Z`k`mZ-%j~|o=qCBwWTC}@+!Q2d?bvf=|F^nt*e@>8J;99rt(r4c9
zc+mcUn$fnCLySL$dQ_-)P%|Xr^1$v(c%}Op?0yC(&WsFO1a?J)55<M>gC}pP>!&@A
z4ZB9~rCl(l3#g++FDQB;fUD?myT#y!R<=jTunkDsfOO|LhHXUBMw*+DLldHoz)z6u
z38J>5au{7s2AW~-gYrIva1X*O7$_@fu7SWB2qyAc7+(uPgw+tGhFOH`VP-wdBisZN
zn;?{M3ruPOT?~eO1}17~c{@yMC-Q$~&Dh)c*#&FtY8zx5sH1TjlBFSv(-E6aPB09c
zf$&GSS93cBJ}ztXEu3TEogPFv<hXota8uOm&dEddwXvG4jE;@Uz_*O7_*!{n8BFE6
z#b<lKu7@n6TK4M!T~w|;ZkzOySc<z3JE}Iv!zH#(-Rmb^($cS0yWWJz!nyV4^qdDH
z4(ABXL^#<Yh@*BFl|V&Nyda*U1VI8tiGoCm%D}XYEbZF(%)JIEW7dLuEzN3hSJPY%
z?)5Y`flU+XJc9#In--ciV56bA9c<cZ?gG;;x;;H0=mA}Oe}1~==Yd<FfA_&pF2yD-
zx3?rN=iznu{jyWPm>B%=>x_obkE<lLN}ek#pZSB|UyUDp7%zB6F4hY?{8YfOQDdT%
zobNGgJMwEs;|O;kzYdx^kzXgx-AK_*2OF=D;uXzlphyEMN0<(Zbegk4kxjTXI9t|s
zGw~Ak>ZqWoFwzVSa?>E)F-sd&L~77<B5y-(ZM3`{DccdXf_ETg2hE*G*-3LZ3h71>
zq`sFZ>?P7=tr+$d3VB7#-=eU$v^)($(tx@@9YWG+&V;Z`k_BVfEC|b@IU7Q<iGFEP
z{j)i**F@k@;~X+ShjeGCi<t|)j%jvKe{>?siGI6#=lb>57Z=|N%(zn%-r=<G1IP~d
z0*PLrmE34wPx^LpZ@nt{qfsfN-Vlwi=oM;4eKZPXlw0iDtc)>Jgxa8EM4T=w#jvUo
zDvILA#OtC|hD}f=P@!U?GLfQd;ByT|a}!Y;bj+-QxrA$BMlH-CTnExR5`p2s7o-g|
zKLP0znp;8IN^=`X+i30pX$Q?+pzNXp+g_O1OR`zFM?0;H47bkv;JekGw>E)GH%lz=
zVIsvi#f;cvz43Y-aj6-c$k4vwBi?T7WGm-Rfn44bD{0^IQA)#Xy*Q=&wk3L%uF@Ot
z9pgNBksrBq&(_R$R-&4bbL!H6Yy{katw_EV4JTdr9LYZ?toFG*uKTzceQ*=k|1xO!
z&`#%#xas@Qh<#*>Z!CB{)+N~ZD2~5Rf#WIA<xid(9Nz(>PW=-fLtY}qOL{kbixh8Z
z&ICm!eSXRUMHXQ$orRZa?K@R6Ca;m`HCd*gE<X|FJ0RdH9+fa*m@cKtu+s-lcO~|m
zJ#e;;p5~tDDsa86%~H<i&o?AW3-}8tisnaCw1~fmqQ(5h6vgmkD2nCBQnZx6l%i$)
zWfU#vFV_)wg3&kZdqYgR>nL)Rb6Xa}x`<rloMjl+RpctC5`7}MoIAcTI8ox$igfv%
zw%$1}L-%d*)`n2RC|%^B4Vg?eNSBG%vX_Q_JKKg#+h}<^vS>%v#AY4HqJ!p6WYI};
zH<EPIFEL*t$xE7FA;~M6-y+Fdn$tj%MjwFEL6S~$CP*@A&H_mm&DkKyCfr%LVT!?n
z;Z}Is9zp?!NXF@TesQV$DSL+x?qM%IR$QKV;#b8xL0wPw^X<>%RpEMkoLqIl`Bl#5
zeC%;rrz2bl18{npa4}fn^gZDc7>d)d9CJT8!1+n-?+wUIlQ>BRb9rB!BOfYp=jcu7
zf##L>XZ?y7szUH8B#B;aHcVa&x`pe_#am8ajZV#f4{zip#l$Vt{z4=<X;Jb%?S;Lz
z_}`z5yTCtz*?iLR%c8yUWz0aUL1KxGt)0|fX3ZN!GzOAO8**t!E^W!BHMz7RmlAR*
JCYMrj`Co}Gb3Fh6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_bilinear.cpp.2A04608A0019808D.idx b/.cache/clangd/index/profile_gemm_bilinear.cpp.2A04608A0019808D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fb5a07c889155626cd50132152c0eeaf0bec67bd
GIT binary patch
literal 2496
zcmZwH4LDTk9suxjG;z|*kTgCgjgJvC7#U-lFT;%S9fu(*-ODg)u^vN=Y^>NNxxSLz
z&1@}N85LPYg)Ne0@3OKjQti!0YTJ_S!P>rd=pJ+K?!M16XWrlYf6xEC&-?z*3Gx<+
z+SUPpAlh4=5R(##0RUiw>X)1%+yC(z0Ekn?<W=WGA`{0Mx_3^E)lHmZ_BE&Sd6xUz
zHRJSZ;!2A(JQS8)`t`(m%Lk1=)!2<3IYtQMTyHwCm{Od!STs|()#BF!2^~tWBOeBw
z{C%!FyZ`ZI&#9j2@5Yv9zMc4p!5vF#&yo&4Vmtijk!K#e=-4E&H@TqUWMf~_%eh})
zkDJMVTNsIPd?#a51|IFK3uYEf4Bu(5%;LRbU3(*osEX)*Y5d~JL=<k%^!Zt4VvbTw
zTTGud3)7x1kI>vGdv@)CR>&@U0<|b%FaB2Om884)O<$AVJ}dpPygK@dmcXHtQ4!;t
zlu<sUXeRt5?~oYG-Mp~PIXt8p*QY;MmN^?<anbpL_brM3<E)(-PsG~W287!2aU6Dl
z6?M}5Avg7gyfyV{+AOcVu6)dPU$20>v)`k}^BcD5lI3XkR6$qHk*jA4;+#iIzns_`
z)&R9dB^*DS-$px}$v1$e(|DP6lb%zC{=NAStV!(1#pV7S=pUP!pncJ#*l&c?>D7F(
zQ_Q%M#x!v$kv*_uyJB|b?7nH-`-4q((C!&m>gUgo8-^I+#NFcCSns;hOHL&I1|@F1
zpbQd)=UgY6Uh*_M!7a|rD;sJk*X1U?;?3vzO>$Tr3-e0l<v~*5d77|Bz-+pg(<k=o
z-+~(hl63P}l%G$V1o;L9)Wy+T_R<VfWPU@zyVg(~1`Z8<Kk~!odtFOfo;p5zlpiV<
zG}kMows}{jUy#mz>T+GTWZ!4W)5-(AF84oL5KUcZH9T|aW%ixJ&m~tq>%Aor^|4JF
z@3>9Fm3E1tRLQzZY52p2sX5&qdiC!N7>n;Ym)UwSgt9yA2}6M?Na4=@XcUuc0uMIY
z&nbk)E{+q!^EI~*b^tAXCw<Ns7>3ox_{UB>aTD=sc$fdHWki3O1Dr3cdL$@m*|no6
zd5_7+)bM2A4a`yFeg(NF^V`Rc8RHeNI};B-e&;twmrL|<X%X;ewTiMM&tFmV7Ou$e
z4>caE>2up_e4_Qs5!SV~(Dk2_w0}9e^bN!9J6Gw+->Bf%GfSNf2cF93o4WW)@1w65
zxwA9(m(0e_yx=yDd_cNq7~eGgm3F9mOL~p-^On|Z3x6Hbzls}-#_6JWGI>z|-cNFi
zb})m+373E3;qxjmU3^nHp3@b10(#p)%Njc>O(^`>F4Q!-@q>yt>vP>gMw8PMqTE&Q
zV^r^anQo~p0t<u#0F1}rjaDwI$V#g&c;n?7CGExmVXCka1eBo3|NemGS6bC%G1BBk
zRUe<Y$+v08AqNdCEFcO<tBGMDQMf9WZkOWHzti4gv9JYXfmctmge+0Af~-&?Lu8b!
zA#0SjLR(R?foxE+g=|rxKopdy5EUgm$POiY$bMC<&2#;6nvruL2CfD5YC(Nu$zlAv
zhJ&e{S5<-c+k$LWMAglXhtD<^{3Y;@`ufWaLEB?-SlG$kNrWWA!gM}esFpL?*;XwE
znSqiE*##vgnTZn1fMue-k8Qw4iDSS)iAUtE9tak8CAy--C-PBpBf72XpNF#j2i^wM
zLl`(0B;<ky$ipyjK1j$%yAUK4qFn+KO3*F`3FT;4f`m%6t3W~(+NVLnX|<8(YH0so
zsbW><V6OxP$T>*$a>wSizPHv`J|I&iXn_QlZKVXMh;vKIFYDYGR)EVHmJ9rST?0UF
zP1ZsGvpu~SEG)5>(2xzl!qL{zD8*REpd_`Ht}5?^hw1jgLSt2Ck-o?nS@;+6z8D|)
z(}|Uua<2tf4<A_(sPU(jT--Cdt-4T0PnD<^#MOcukW^n)@VDsRzTNScVrvr}mtWw|
zjOt)uCr>9|WIhI_d(gera;7*tsKuZ#P;#NTpv0swQDT|0wy5u8o3c^jm~v3!ne)i%
znXcxpDDlntD7l%tp(HRDpd>UGuBwwB6us7?`l)I&{A~Ojk>mpw8Tp5K!8$8@<Wmdo
zKmvEmwr6tsGpAL7edhZ($T5z8GCk5cxwcs~OoM1xWcE|&T(JhLsYEsGZ{tr#h8N3w
zwbCsg>aT1xy3ADHzxc@Iib<A9l7p8!fy<ZoEePoVYhZ`tt?j37+4D6+LdC&>L4bHQ
z2861+A?of*viFoEM91$}b^T&@tii0+)PmNn$Kp2V>fn9!M8d6l+CCc%^*0$Av)Ejo
zE8oqXMzG^pGHti~-Kc>naWl!>!pho)LbZ3?=H%?M-Hc)4K=&XRkiC3&YIp_+yg`59
E4;(eyfdBvi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_bilinear_impl.hpp.3D87C8E5D761FA48.idx b/.cache/clangd/index/profile_gemm_bilinear_impl.hpp.3D87C8E5D761FA48.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1a7f7f3867e4ebacbd0eff534607f1ae1628867a
GIT binary patch
literal 4594
zcmd6rX;f547J%P{2h#1Mmxne82BAR>%_57)VuL_0!Is4aL8DP5f<&YTZHS^~&=I2;
zaqs}m>V_I|LEIM<A?PGXH-b(yXmY^NEsjw{QKKB!d2M|CYL4g3AO2WB`qWqVR@JTh
z>fWmAxT$jaJwAp>Ql{ppQnQnB48thIva;2Se9SQHh82e8Y&}0CI@aLLyByp44^5b8
z@$5#IbyCgtS!epXJzFX!CB7cp0<!$Kjdk028<xdpY!MdTS~bwnTA;bze^<5rW@4VJ
z*}pezo%Oh-`uXQ8pPsT?*BcpF7&AS%$jj1g467{fh&^kj@>tTIgw20G{W`_*YJl=h
ze&gn2%b&hJ`tro_f~i%d?4~uzi{%?r6)$X_)R&c~2!<cgh|JGlVt=1=W5>(Cr&nMN
z7B1C>39qi!U|u0PoZq#1Em&R1pvv04$9Z}H|0HouCGD+#n8rWJz9ycNCCgBpsf*kx
z_#|P?*oWg=yWWgQ`{NDf;#E2KUw4L`xnBKqb;r>b$Kr=RUD^GG+_u!x+{#NHU4Krg
z9T;(8Q>?Ri!y~(h&IfiL3%0b(_n)@A?%Dj#j^dR6R9vZzygNzvYS-r3ew?40=eF@8
zG~}8Cl_h%?PKZ1>)4M$K!f%z=^F`GO(d_QB&E<O<+~De=Rk?|?-)@i!QVtP`_63>3
zU+-FK9$ok}8#2SJJna8gvCK`I>D@oEKP}|@F-v$(sxbrZ2a|Jp-P^a0Twgw^Zf6Zq
z+q@=wTi3yjXDcmNbPjGzAM7irsP$aBx6w0KZ()cs^RY>|GOO&?sr<yS({^pwLc3B0
zP9=-RKRYl(b*?9ONAWo?b-?A;kS_f<Ij=`ny-6GFc@m!OAU>N{&bs_ru3}D#=tT9?
zYiE;G*3YsV!{lz8Dpy8MTkTQWA0D$n!E4QaoS8Szw>2rvJt7jzTA}(fjM_nV3m*0D
z|K04lnb}AS3aLU0*a8w`{RO`k1?7shxxEB7!V)8$Eu6=^`@vXdfSKb!f;x3!)@}}^
zoYWNF?&rn@g1iZ$o8U+HH1`?4e@EMw1y%$pah15y%>*fRl_Df_l_4bel?O1IMc`iq
z!3-CJe=*7>5L5!8j9dc}4Mfsx+1_-&PjrNBPLNT;D3e_T87+)PC`K4#BI`8!fJFyK
z|6q=jI`Gjki^0im5OjkbJvL4nz{>#MH1p~mReqYt4ohk-PLwm<OwGlKMkt08gOHM=
zL@16Ehfq8x-bA(+)|4LaE`L27CnXM&Pw2rZ5=lH5k~&BcnN%c2WHJXCB9n<^h)nJv
zM`ZFqxyj==skT>-W@b=(sJ##pt4+SQ>pW`ps^+vWj*~^;Sp<k$F?bfETmn8N;L8l7
z0Z$DhxBvOarPw`j>PNg*he;2YSB)*K_+ZC7u?v@cYe$xv;iP|rzl>R8P;gKPO~NTB
z>K_MQVGr3;uD9!3*c0c?_`r+_vAgdHmb%w|AP9CG&@Q%29^g<rha=(97U86nBSlEY
zks&1K$PrR=)Cg%ophaUs3sbZZ&Wx*t8CpnWSO=~;7{_ooICp~!!+PN80U2Kp-g<Pa
z=s}@}C`R7^&IYExyy4sajJ}0Cs2KF-c?)PYik!X2BjoEH$WSxzo52=k*Jcngd=dB;
z8M&gOsjxk-&)SS2Q-~BkqxKc?m5HJQiZTVB1zV`RlPFI7j8RJWNJo;}Vvofpif75r
zt5SLYq{giRRuyoWPN%{8G}zF*NS(DTcCaQ4C#X2Xi@=%|5M(jn#VD5mUV^d)a1Fz%
z(saAv-yH7%PIiD(2bhvg&c&YFEAgwBDO2_Ve;*)zJPiE946lxSF|WvbLISnTlKqn9
zv>|H`MX3_DU2UN(+zX@j!l$%=AZuV$4a)l<a33K4*$<=kGxALe^K}mj-v(0S9|niR
zAfyM<*lrMn=Z?swOs@i76_^r~=K1`&Yq_JZQ$3o%zlqt2+aqp?9hK9pKjNgaBVWuN
zNPm$<*{%hVmPv>^`$W?(+CINP2^wHj131xcsLq&%E0pW?@6IZ*jyVU7XO*m*ajfyI
z3N=6&-Er5$k#4u|xxAphP#HLt!B|>AkmcZ1j&cRKRseE5cY?T+aqFJ^SrxrsH;$$@
zx>Be#9ijvoCyYZVUKnqp!)|vH<_su3EOBa?ep>LS=QgrJiw3t21UfXi=wXbW*`_$D
zhsk;{rB&l1IR++q9R1>k@2&zaX}o=~BuJP@SyF^#EEz&_mK-59OO22g_*xiAn@dnW
z=69p42bP`*uaoyzp0xe$@qr;{`FhSD9qX5*W7OyW-|N*`j@6Z)X5Ehdi1&2ErR$#`
zn*5}9$WCaS7?)xp+jXCEf?uegDb+M7HVPkHUSbRlzo~vP^taKe_bXo-R9PRfXWDqu
z&X?T2;E&kEc9&K8YeVC?#sKn_Oro7jUEh7B2+4e92+70bQbtoPRhus31a+a7>X^&9
zF@U1GLCI(vKyCm9%^l)#-^9g^E*<jTxqfn=i~o1qdZ_bkrg$b|U4}=-M0&Pg9M4(Y
zba!g|Aj)p47cQ6$NiNh`ljM|Sx-`@Mfc@@OM8Bna0y`wnYwcjjM&D1bO9--w_>nT&
z-`3xqo{k`cY=ca6zqsz%HunzWeJ2DDVKYjhcql?s@F@s|;b91c<KYM`w_47jM|tq_
zi^1rs^#R7%;`*5`0RzWY$6MX-aiE%4M68IVwI6S-Z<=t#*>%VZA6$s{Vv|?vD1U1~
zs0CMgh8L&Gn!P6Lj#Cq7hG!<x&98rGA6F^P+4_GSg{f0Bl8w*WEcJq{Dr1Fb{Q0XG
z|4n&c-ambfWJ9-M*f6GQZK|4)$MW4pzwA+I2%BLnD>h-x;qv&yG2?^yeWCuDq04Z(
Of?)(xWiwSSQ~fX3;vGf+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_fastgelu.cpp.2EC621FE126AC880.idx b/.cache/clangd/index/profile_gemm_fastgelu.cpp.2EC621FE126AC880.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9e8b5dfc8d626893561ad533a279c5634cf72e7c
GIT binary patch
literal 2304
zcmWIYbaVT`!NB00;#rZKT9U}Zz`!5~#Kk2=nT^a04B4Cv3>9<Mh9>6UHWRoy|K{>X
zI@TvQl{MI_tlTS=Ewng$Vn37XT$Mv2D<@8oeSiJ^#YwM^cyjB=CEVCmz3cey>f^;u
zQv_~qx)LU$GH1>5(4RjNb%cIZd`>spy?fU`o73x_Pd=@D{CE1dn#wPSYAWl>p6u9T
zQDn0`;!bUE-|qg+uk-64JiHyhIp^oox_Pm>;+HvPBEMyBJTq<n`hDl#@4u54zWu`<
zwwhf#UpBf%-EO;ltJ|q!t6%i)#Oa*MU#@A-_qx61T6awGwxslk<(11zCTb~{hJTQL
zbYE^|=)0EN^A23Fiq<;&&Y`e=-i781QuizNCJKKzvMfL~bK>T27fa6V`msB0JMUJb
zP!qZ4=70Q4OM5#UPpxp6SK67Ltl7TQM`6$LW7^qO+iSkRJ6zp0^KDa~&6OfWpMau<
zH3uj9H|hH5oyZTFU@98Z<;QQry1j2HM+5h?C<{S>=j<2CUN9~EbFA%s$)66H<}3+8
zyB&UPum8w7m9AJ8c--vg1s?ZfuLa^%A~R=m%ieXLGxep1ME17cw>DoNFkU=mV0y@L
zW%FnGvT4^2X$ER=h8z-E_iB+JbN1pr?Mfk9*H~Qzr(ED_bhpUjH_M5Z`@^QXINVUN
z>e05`H~cG4MU?FNzIBlX-}!sNng%Y-hc~8gXPc}wVTzMf9)HU@$t#fo%y%b5O{p?^
z85kekesIBQZmwyMq;6CezUyw=x2!BZ>{uR8Zj158U3PD6pShl#Ql5IU&q`nG9`juH
zMTO^DHGlQJ-~IDV?*ZS{hF*J%4=1_47I+_+b15MFjik=``tlr=w9~6(PX{n9yZQD*
z?57G2zfiv|rAuD$blv<b_eHC0Te9No3L%Y0Uu#|%dvTaAiqT8^(Wk^aTjJx#UxDk^
z#xUM`?z?k>HjmDwOREACt^7`?U1^j0qtKkTt=NJ4>9Wk{u`#;S#WmOUuDo>eXok(9
zMQufwSKP?wv{lQ$J?Z}U-ttfO!8W}2q$7{|PM&(Gy{puGQts>}7wk&I%)Y#SQDfp)
zoO&s0Yml1x_N1f5E{(NYWp22w_<1}>dA7@g?<TgTTjgZd%T3etFkLnA|H|io_gMc|
z-hDdf^xvqnGT~g(IltF5%@rt(^6a$E-TK^Nvav_<;k;wtdMECm6Z?2F)8rY3J@+_y
z{KevyO}wNTA||>q!*7AX4OPDXaWnQmaozl_?YP+KV<%V7QvJ{5dM3<~p|~<P$qASv
zr~4k{pIcMJ!ov{Bz!1pC!Y2tPfHGjPYVyLF_nF+hc^M*r(jANp9gNaYiUE|VfuJZg
zt=J66wh}lrck9gZf80EQ${fmkaDxL?I8=~$svN3FJT(qAB%V5lIucKVLj#GY$)SnF
z)8f!V;%ReeBk^=NbdY$u9J+9xmE6Zu*=&g)8Ce3G7=@Y`MPQDMaJVILzHHwvpajH^
zFn`C+IkNZRTj4GYiTIT(x2+DcOy}VVG;%U>gBipVXzXC@g5a5`nP?$+rfQ~0JTo;j
zB%ZmNITFuO+)^5$&q~}1iDxZtjl{E+wN*r@w3D?%;@QjE!+B{!EjPciUr=WPc~Pc=
zQ5oh%5MQ-}Q3u8^`~UUcoL|abEG&UVVnyOGvsnVm#LAF(Rbo|0ygIQuIIm^Glx_S6
zFVAJ<fu$5kLSblGTwmq8;^=W^p1^SRa08e{Jb@AF5lFm9^+-6c{e!EF<@p{dpjnP0
zj#4nkz-T5el++5cP_Kj08RnIRIo>Q@fo+e#@$J;a2#*+;0!Ah@CvJ?nVYH}+UkGSl
z6C-aEqae(t&C?td1TJ0Li;)m^xJ|8C=kWXn4@;ntyOB3cJxidmtFb47XQFMQkKmbV
zn<DYdw9Sxs=Gx{+JWCl%1%y5;87m~7wTv|q&sND+4WZIb$qtEUuVfGB?K25VVRc>f
z8t6qAO&3F$mQy>oc(av;P6Y}_D@EJDBIlfPbw`t3Fh4jzyqXyOVCrwTKE1y0G3zFv
zgg%Ep4@~$q$C`d7s|^!?!mgUGCNSaeQ;xG&syq|{2On}26{Y4Rf|5sZQF3uGhyf-*
z3<icE2n8Yo5X=w+GYrmT4xH(2E@AWT%LBI1u#n(@Acj!597A$$N=_z_>yz%s%)-jX
z&cVse!z(Dj&*vrN=At3Q<s~d3A}T3mWo>I`@8IZUAf{`hVxgsgO;lV)R!&h#Syf$A
hTSw2($k^1}Qr=8j-^5i+T+PEfoXOqa*^|+i0RVAWXkGvS

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_fastgelu_impl.hpp.090E8EEC37D7278D.idx b/.cache/clangd/index/profile_gemm_fastgelu_impl.hpp.090E8EEC37D7278D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..471324fd2b781d53436cff2c7a8dc90ee0ba0bd1
GIT binary patch
literal 4054
zcmc(idr(tX8o=)dmq+4F0v8CBw}?8tUr`_-380A*#mGxhs|W}-3J9Te5n64jC=eC4
zj%Wo6inFdoD<Y3#OM+0>2QS4jp>|iSqOO7u)E0`ZyEoRGdv~1K8UFIjK+f-c-}%n#
z`@VZlY?M^m&q0s{8=~?v(>_c_5d^U${&GK5DvXT~#LXN*@~b=Iqc=MQSp2KDX_qyg
z{q>FCHXX1iIWAWy6;C3ydpkb=rTA!G;?h@zX$q&FoB~;}EAP<6(XHM4mgf{Uy|+>w
zRCjR4vsZJUkI65tlit0Q-Zb}bW<N+iDtd1GP$}<Ymo$v^`e`3%7ftf7UA$FwVku|1
z=Bo_#%D;cSYV%oB)r^te`W54y<rnKB<+CTHI!mYR4fj^7nw>A?J%}+{`&q)3JNxoq
z1kAP#*ebh}w$wRRX|r$mW(sS9akjrsQIr+<!LP~{|8y=Cc$R1OHTGV`YC772Ligze
zkp&ZvmkkGG#jD3N;<KD&xsT&+OdjltV{$xRY#X`6lc#Sy-@5u-!`5>v6Kv+D{$gY@
zRFRb`F;_IBgCQ$|c5P015tMk@+Nx&3qHimug88=J?KJoFy<u;jR6O1ia8GRgtggB>
z-l6fTa}&?!c=8^{sqh1?v-h8T`axvrl+}*HWk0NOjOCz<mv6Iu5X2wa@>@w$XM0*}
zk6<W*%UvH~va!72_Tgi~j(OYaoNs!@#rL0awa+XKR<G~(i)c2^yHa5NZ$X#CEia~)
zed3xpoe}#}*7}oD!En;!`yR#jkF>Uhew%)<Xm7U5Uw`~$=PCZFa7~@rmWm_1+u?Et
zdB%du>e`o2uk)YYzr#4ax^VF9C#zni1+qRz*)1EG(c@b_b4pWbLf4`aBzJ3ON;u&Y
zcBxOtj^8qxFfy_+A&@9Im&wHu#-DJxy?KdgUc!D1u{K3;7ZVqk*Z&ZP5g^7W;Gj%P
z&uzsB<s?l+e~<?UFuV^u`yhzyX&f-)cx`XWZc_{wx(hwXW(*g(i)bVc6bDnXQkYi?
ze2U9pUKyQLus{WiNx5miW^~N6k!ghC5;uuqWDJ+ONe#r>XdkR-aO^ZjaSa4$s0pI@
zAaDo4j<km2I`GwjKgkyD&Y3~`7Z0F>b+%+C*-ThxOKB9vj-ru_Ei;g9*Qb?dZXW!D
zf#O03p%ZC<AWxw;MIr|gO(yab(PUx=F-;~85gYaf#TE7nN6L!ATj5KiFolRDJKbsX
z-;cEg5*ttoKBYiYD+8Z0I;+4(MX~+0^Et@Fu(WqPe~0-`{!%r&{L8mI%#~T5U2KO}
z8lkv<q`!n*FN*W{JRwPL;^y{eAy=5wUcLKB`voR%UHn_a`pN4ba&x>+yd~f}k6+$w
znmo=TPM9sUAZLK$BDRP|Vz!t@3bukqYT&Br9;pVg8X_o5YKT|E3W_!0u7Noe55mkr
zm_@M`*jk|3*Mh$m0_k#@7M4=1gPA(2zx4a!#~EYkwL}PeS$J_tHG*b(&81PGR|rKd
zVA%q;barn6Pm0^XvW=2Q2Y<1J>$AJU7{jB(q8CwW$==DdypZCZVxTyNq$4xUVlNT(
z6~L$f4%MjvSPfuBvO<~rk$m!4IAOaK&{ANL0t_z$w2aOwKvfiHR&KK6|KxlRP`n#l
zyTK6A`Q1N{6rfl8h@q=NSPhYs#d;9dQ@nlg)VeSH{SpWtNeoJql4?aKmu4o^T)jZd
z=nzN_K|CqI@G6j0(YXdBH59ANIuDn0DL*9!UW2W%q`D<xNd{8MjCVXL|4#@pXgw^d
zhh?PNe%oE#h;7!}i1k!}MFkkr#{P-*Yl}FJ*NGlyVBQ(Z)4qxv4s(uK!g|M{Ut1E!
z8Q(OOOUzFVo@y#J)s1;B@3x)jA_R@_UL&}YOFf*i1eL|?y!|@chiSl&1@+lpG6)6_
z(`S2;4#LTfN1oVtj69q*MI6l@aNPs5NdboM1=qcF?t!o#YO9WxBz}1}Pwhw?O_W=d
z;j@F`GB=rlPI=r*STi2uZHf{L2~xv6(z>2yYI+E2fUBXW04=y^DGx+(EeNz=NN@UO
zuys_3pMHNR@W^fsZbp1{8A6)Jix?sri5X%VDHsYGslieWHsk~`V&5zW>8u5YmWq{@
z(SjD+vS(jUJHNKG9}Miw-h>cG_`es2vuyJ#|41B(e#gG|{l)96P70pio%X`iyx0vU
zk|U$U!hA)(hS=B0O5GmyW$RtmH_F!#o8&$N^YsObyq5JSkH@^clbQRDH=);Jq3$JT
zjQ@_O(LbxOq(~IU(R+2Ea6Y+b#9cB_L?dyycp)WIEL0e7iWqVCTd1LKj(V?-9)uW5
zTL)4d$VeUtoAYt3+_7@nhp%=1&g4fg*WS0h`WCC4)#Wu1LZpdhVasUKO!ZFXk;DER
zGO%FcgdI$K*wIbjB!pT;o+qsP+xmNv_As1h%QMjEzHff5@fxT#!SE0?gh?rdqM<b6
zqkI|(Pyvmy%(K{Jzeg|MpNy{B8LW3`j<4Z%Fy6F1&U`4qfoRT-%$Aed&#K!sevLET
zr@gevbo4GF*xo~Yywu>P26xipRO^9rzJi`J#6am0=}BbsAD#Vk4u$1c|KF`pre!4S
zA3C|p)Z8lK*9gzCZ*Nqb`kjYanK8_n7>muZu$+Nd=#~B~`iJzJ^5%`CCo8fILCmPC
e{$2E@u*eF6u&FBX5GTtFvVtH?s!F`#$?{(d-@6+C

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_impl.hpp.DB85F2B55BD1CC04.idx b/.cache/clangd/index/profile_gemm_impl.hpp.DB85F2B55BD1CC04.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a1e8f5bd48d3b16f452495da9f434f0a920ecae8
GIT binary patch
literal 5994
zcmd^@dr(tX9>?=LTzMHDxsU*XA}>XtpgaVGtPmiHMpGUV5v3JSd{A5knpSDIQc4jV
zC<r3eA|Tx)YEkQI=xDJjZ4+Kf(-x((wYFCKLaFE~D$4GSHus#4Gdtto-kD_1=YG#Q
z_dI^*_xs(Owm>S)b>VOlS1s7IKJS$rio@aXm|yWLMeBkQhm&l>;cRL?sZ?yZJlpYX
za_`<C9evQ{zOLqg{GF2>C$6hsudNQR<4sl;SUK;V99LbVxaqIw>~{|Bin`PF<L1f+
z!_nKbk1D3r+s<#@^0jp6&dR!$gR<o%9>xpHOK4+M>h{)ezKXNI_xbwt)&7Mp*Ru=!
zRnyiL(8<578JXrLe_s%|qD=1=mY=%d)YUU<Qrgc>8BXK3Joo0HMn3kcRmQD&JBz&3
z_)j^(Z&|&0C8*1C<MlJkIZ>Wn)W(rBpA0QMUEDV6lkc%_qH;GC_@9tl{Gs+kX#C5<
zwc;&Vezg-;q4UQFy>e$*4e0M}y5ygA<+ZHCzqFi-E8JW+>L!}+`*N(~F5&r86*mOk
z3$6ET4s|tT)A1XWwaXep=$#Mc!#@2FWa=UA+_hH!evg+RpI*A7Lh#KG_38OjVzl))
z^=;mZs^o4lSO3vOP5(&+<-y+-E0+vynYl_VR2M7Xx^sT}W@Q7~+0!=lo=5pzqzf@T
zY->()h_<flti5MzSJ33~!9d}G{eQn~Un2jE8!>;)FXH5Z+s72sM()Iw3}>866)xPf
z@8hfG-8#v;o!%W(XODQqY2=l+gW_jiY+l;0{Mmo!%*G=Zhu_#!@1FGSx!&G6n<oza
z^E(gUuTL$Hig9nBxcuKGbGPM>cckZ4h4eNbSvWFES==d#_E{FNF#qU5{iikhQSXqg
z@{}rra`qAnzHG<Z!{Qy`x&Yqc52p*4XEj!RQdR$U>!t9RdsF9T7I<)qH?Pl)XROC{
z<&A6me?a31xtR=#xoDxSkmfMPWV*_Ku77oMQ}#?Q$8`#a_N9FJkN!9|2skblKtoa9
z>f#$tOwARi#OJ~T9DqxALqIo#nOj<fPTk)+kXvoZr6qon0P}4wE%lQUBny*8VJ&-3
zb_}MRIXML7Cg#pJYj$gdQR_bRIE^T}6GA&7+$>PE7AdtT!^}>WvI8{`A!dem0jLXL
zV-{SxOLrdq{=J8e6fJR=JZHYmP@v>_Oj36#p-BU!geG&B5t=Mo#txRERYH{q?nU*y
zDuf`BY91y%r0QWYW&@-cAPsXbxb}h&b3eHDlWYWm5nQougis^79!5|afn5*R3rDOU
zXg?0Kpy=%2Y+_J3!8ruw2IsOQ{Q8SR&hMhUXKbb1Bma2q{pq#4o-k}L+uDuqxY9cr
z+lx+!mf)2a#fbi9CZD)t$8Sb|Y5NrG?04n3!8YQh^d}w|A{Z!?o;WH7iaBB)-M|{%
ztmF@V7Ze6{JkcV?bFA-W%bc-kj5*s&9L<BGXsNxFAep_4AeFs}ARP#Fq*>`8P6uM#
zlMd2#ump1__;tbz%zB8>LnLMcL>Nf!fruWG`yir^<Vg@t0&#@OkgyC9T!DlYByU5)
zZ6wzsVJ*pxNZ3g7E+pJVatji+klc=h?If#_K#hoFrbaW=XeOSY8bzv6G|5smV*M3G
zYmh)gw%4FA4T`|^HRuHmO2Mo}0xelzi-NQ$nAGQMk=)F&L7YV4$5yr!m)3)`9(>FK
zmo|X2f#e==?jgAkocl0KyWSZsxVF00hD(P#h6`{v;ojl?1cimObLG+<!0P}PtagB3
z2Lxh11iV9}ehYZF{>py0$ojW|cN^=9sQL|p;OaUHF1<8%sT6mc8JkJave;!TEw+)I
zT%YIoF5=Q!;AwGRTzV9EqnKB0YTl52vg;@21eK%Ea=d{nZ+@jvIjsn1glfdCMsAOy
zBJ;OudsRhAKxyw2hcQnfKKa%WO)k@-5`|VGHma>FDt<Njt~Q<-e>RnEhsU2o<q(uh
z<+4<EaM60@9;1PYoeiL9fGpf-2Pk%s+z5(Bk}rbdBFW!_;(L-Wf#MQobH8BwD7>4K
zf~_|s9&-n)h0I(^QD`ZiOXGO<i)9WTL(IY+0KWq;+uQ+{9)plEU?X5l!X=Ss>O#9`
zENHY=iX6uB?-w%*r-MKputBu#T42g&F5}&dpawTJc$pV&Z^1%Jo?3DFQ6?km1R^uo
zhXnhu3uym?A8_S&-}_cvdX?9z>E`xaI?pSQpwl2ajs3LA9Z2-R`U#UeNHV|z+`0!c
zdaylk=|0HpgO|+u{?$d#&FdMy#O!yvQ#x^^l}<{Q4hQ_KTt1fiyd}j<I$Q^H%&ko<
z*TEvpognBWqlyM)Xiz5Bwa8nGd`XVcB9WOt%6QxTO^xy<Gnz$Ci>8@xb7_T>f~6x_
z!(r{!4zx3qciKqUBgN#MHd2CQHZp=#HY$R2z|#RAcdv(8dYn?3)^L`A<Q|ySLvkO?
z>Lb|*JR{lPB=9G}4fkJ$_+^OT3dFA<c^l$yBe@pwYe{ZI{6>;@A^t9sTM)m6<aWew
zCs~bnYQ)a>n7@s(?(E**7*nl5p*4uz&SNLGo>;qL%XDV>Q=C%R*v_R>ol;rSjg;tI
zw%t7N6p7Dq``-pDHs*7f2mbXH;)K2Rm-?kw6wjCw>FOEEJQ#cH>QgM?rx9tZrbyaH
zm>3f|FOrReO!z=?4M>Jftr^8NV?HIX^0|I?W4A5SY(eA;xY`4VcmQ)Tm!g<b6o<JS
z#g?OZ%#|pnlGJNZbPZyYVbd_=4JaAcZ$fEJ_$7@?HzRp7skfoDHmn<d+g-Zn*GMDd
zVARN4O*#lQ3RNTG?N5!wYLsYxkg0x#nshK4<gY>O>!&F#OVl7Xhwm@ipE~i=`r>EI
z;07F%E=)LFi+RSTea`pw%CdQj9i}poB22=1FeXxjNePn0%aY6n#idnADmJU<G9Mq3
zI)Pw4Wa{BXT-yLk4P<of1+QN4!+Jk>^^<Ib7mbjE^?k^5A7YdFLD7si(vm%PK80gA
z@A=7;Z-cxBp75>bn@;~4NXOPN!kX|k>{0zOdN6Uk!xf(55m$bQF39@aqV<W4c*MG+
z7R7v2$Ha>!l+^Us^PSX&c_jaka;{Ut7fi)ym*^loeUXcZrICh1clHDgwlc3c@l?Dm
zt|X?!1SL=j1SL|5EM1?-x~q7<BFg0K93lm{g9qUD04HRol-IKq5h+zU@+>F05_wjV
zT!TDoNNzxG4Tw$fOq2I)BK2nE)=cVc$g>UWWB=N+*m^kBo!OX@gc8<g-fTX8YSu^I
zeowJ6cUMzaIk8*%nT^qbj}F_}{m!N{A+h~OnMV|fi&mId;^FCYGj@wMHUHmjD9S6y
zF}({F7p*MrHx;Z*e~&LmbK&s^>OaXE-!*+Ne0(8VrrEme2=f8O(aD)Fu(Y<anaZ_u
raOBzAbNJ>;9F8R}a&gfP7wvITfQwUc(H0l~{IqGh;ehWt<Kq7Sm`EHj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_multiply_add.cpp.0030E28460A11C1F.idx b/.cache/clangd/index/profile_gemm_multiply_add.cpp.0030E28460A11C1F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..846158aba40b783d13f87dd2b6de595f20b90a66
GIT binary patch
literal 2432
zcmZvb3s_Uf5`fQvAOS;o2MjS_0whRy2SO5&cOZbuiGX;;U?5<%V0jC*6`>IjQH*Fl
zijnfTMX7+iTgyc%Hd?iS1&TZrk(i>zia@JHtK5J+$+=(ee&0De|IE(J&g?Fa%i(0}
z0D!-co1L7HAwU5DV5t1kGt%~nGyx!2Y0H*2M$UaS;bHKzciT4E&aBf3zX7?yyhH9I
z)|`7JORZuSS7gT{+p#BOdxAonj4yaMXCD#FK7UG_d%k`^2l5l%;q8J>rL<N?Ddb%(
z!B5`yeKGdz`LD0#w64l+j-`Z;z8G6n{Iqsau{0)@FHHY1FO5rm#biC79BAtKE3eQg
zx35)k;Z06fM@$$cML1R{bKQ0LKd(3SEN7hC{h)WzVd{F{7Jb^_xQBFvZ_;}yb9`K{
zE6lw5YyMKicwGnoaYdUc`#P(S{pIn9b@g5Ov4+d9gB#N*_g|t4M#gyivb2leEQ*Mt
zkw0(D>YY16@i@AXdtBO4T9-Z|d+6SkeDQ33&Ad6`_gre@hXEKKM{A)WbdyCWla`?I
z$jgJjbEr#kjxjgM`!UHZ<8ifQtswujrcm$XCq_wvG~dIu`3?0`yo`^l?7qpQ1p+b6
z!@b9jczPe{^fyB*7#^Q|DGz#V9&x?s#kEY==$TgrU-`SAV0GRqO!Rc+7rr3f3Ymy`
z@?`ULLs(F6bX#)LOue>#QB*;_AvR>abh){$9k=zEAGgY=E+x_|6CZ7%@G`09wK!6X
zt3*-gO})^eX2Hb56ob01yIAKZ<FAwBbAuyq)C_fBHVsV8epwlkw*gN$>DA?49+CHN
zlfjfjRvG<O{rUJO1;aJNW!%!&<1EaY-Td%olY8IwA5V{^|GsPg!Ni1TH|}ygjW5j(
zO~&)jgyrmzJiEA9GwpRZ#~MEqQ(iOo=hbrqlj!^H_}XDMgBI0}ohQm{^)h9Z*B1Ud
zc#{*?*)ZOMvDs7c0Q=CUv~Vgy5wI($xWB6W!JO~qxV>y#>q$x;i9b-dxmJ`yzO9HW
zn4=ChA2YPPY^fN(ZCdr|<e9ZsdcPw%G~!L`<f0_M<crtzT|Y=<Yqy6VsK}a>{hXE~
zo*pmXcs{1&2AQ^>$~aN{R&qCf;ai8L%!k>OWjEQ-RJS<tK7HqB%<#s^MDriTNfXvb
z-_^IZR>ua-)TUP0myLXN*7eF2Q~OG?5v4(*DMx26m!}HL(C;eiE;NmPChyWt)NMVn
ze>nBm&%Nh{m$UQlB;9#i(MQ;!Ymxk>q4+#C^IzO!>;oM$K4oq{2(KUb<)z-~9_em*
znMPSTvt(KavkkYWPCRHAfirFSU~850r~`I6v0U~zu<d0*A!p2Usp-J2o`aY8HR(CY
zaem4-={x3nzoaVgI0o1W01y_1UBCKMT2@DTOuqQpy^vNtAX;gbf`AmX_<I6Yp0#qM
zB_yV|D`l2J@A;0;55CjEK$d7rtST^Mg|<S_2J{95k<cUrk<nxXS);8HWP`RrkS*F4
zK@>DaMJ(d{?GM)o=0OxB0gWV}3A_<_*3kW*3NgxZ9Hn3oTxqn=xcNZV)o+#l@5_Y^
zyQSZEKl;tM_#cMY`ufgGJmy|32BP`VIB*^e<ly5Fpe9FK$Bk-ovUNg`v#m3NTx?ws
z<Ywk(q4uMj(Glcs=8hl+kwH?g^dx#Bh)HB3h(%-}h)raxD1KpZOQvzGn+68*#(1xL
zf5ZTQfr>y(5vbxsBhlC&+IJ}$D_tT;K{DK>6}FLrG??KT)cCudaKs!=|6ym^(rpw9
z$^&tEpc%Y#6jT7>3Xog^;!2P#2619Fi%y;C)W0@#9$Y!emHP>25Y^^>yyc?orWOX`
zQ~54%LJSm3jYZHN>K+6Mr~(AVQR7rp_#%+N94Nsn^LUwf;o<F8CxxOypn~bu`{pMB
zRVUsO<~yqUG*VN!UPu!9!0B6JMqO)4^o^8i38*Il4dFm-5+5sr!NKl-+^=?uPUNY+
z>ZeK%A7dX?pT6Oo&2I9U8O5R?S`aN14ugUm0v)(&a<q4JQj?Rt6M~%Woe|_>?}8vV
zf}54vk4~T?$erMhAO@L1QLpqQdm@NQW+I41W+8}8W~=BbohQ@^teI8T=x^um0*C7-
zA1cb<yj_2_Ss@Z|D{PoJs<?MmuGOkE*rDw(uzD8VP^!UdKBZI#*agsG^{X>CwR5Z=
znXI-J`PkDEk_0Q)B|R-Zefx?KzVdrNc(4Ef5h_s=+J0IfUYJpk)<i}{Y!BxFQ7Rh{
zsTNm*y~p@uVbVUO7`At-CQ56KHd<#bM$gbd9~)xC37{J3h8UZhn6AUK*j}5xeSH0#
zaSofTJvUnYKdl*IJ#hnxY(ufLr#ZT~y16r$mL3*P^gx`MZE)xgji7D*TyP8UU(6xI
Ar2qf`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_multiply_add_impl.hpp.A03C0417F818F4FD.idx b/.cache/clangd/index/profile_gemm_multiply_add_impl.hpp.A03C0417F818F4FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..98ef2a1f0c7645ea93240e849ae969e448cd5b06
GIT binary patch
literal 5000
zcmdUzYgALm7J#$ifFy*39Nxl3Ab=7<<RKuW39q~mj36Kud?DZhNmZZ*#f6WeRX}lR
zd7~h=)!K^!T8ntCAc{yu&~Po*YQ0#9nxIwND^}Fr6PG7v?sBdBi$9*Vk~QD#Jv004
znc1_?#L2|sEfzFdM4BvLk)F2#qtR#_;+LD3lj&zbqw(l8T7KR4@iFpWXIu8kd$*Q{
z^0bFX{lca3k9KW0Eoe)*f2vhec6nW~%f-#h0_R^g6{VaJ6~E&9wchVu>2@Y$!@Rgf
zHVxI=UaaVMn9yFIm?79u`$t!{Q}lI4{}h$T`)ZHjDDKf4+O8?@4m~XvY)(?}hOPyE
z?HzeHb@j#RE5}NW{unK(=n($rm8;dQ?&?FyEvarnxm^!h(r!t^O(j;RZ>McNC#}46
zd3Hy(y48a5z_@8gMMqYJz-aBAE%#}kOnqY8sW4nxXJyAPa=5g!#)ud8uT2i8>FRx*
zF`Tx1Vaem?Cmp*de7G#P(wl;mX!m7%OPZfEkL*4$UN!5&`8NgewLiapXtQhgJ^A9@
z_m=psxcFz0O?~(V)~L@H0aKV6wtYQ@Uo8JNsAT<oj=Yj{_sdK_%O9JHFzxX1!$o^<
z&v4q%>zi2Y<<U1Y=?DM)pL#@f?dC=IR-9N}VRALS9ER=xdevWYPnbOtT6rmM;TFZ7
zJxwLWX)RU`x59@SEECG}8lNPm376l_Iw5;eRp=D9b^VgBnj4GWlt<KkmGp7O>f2$G
z#YtO@9?91?Pk&sX8NG9U+R@Jk`r{v{eAe7@Fr9z)uxH4O-`AY^x^8dRjiURVftF1l
zosbtCvYN3{%Pkl^oTum%-XCc${l#r#o?MteZLwqO;A?Z?Cc)12;p2~;t-jf7C|?+q
zntMY1>h_<@KA}J6SDUXc<tF^J=IZ8{j5p3pEyO-v#SaSh&;RbakkhcVjF!7Wks3;@
zA={Nt#*Y5mVBEmK%7{SX!L1qAIE`3Rx~DZd>?Y%3f+3S;X-vbpMqKXOKN{TtXoeWz
zkdvN~8)QysF6#*E_w!%_6CVcNF!+%z4Sh_H*7v2BGMKo?RpdcdGjXx27$J$fWF{q(
zx=Rs~`N{$*StZP=gkXxRU``dv)xfWYFiKtn{2G*NA)*$fq};f_Lp#Q6VPH&L?jqM)
z&BPbDEI=rh8|zOtVd8PzI6axIni`n-4~I@e3?BraLF&aZTn*N0uqB&exE5w<VK&L;
z?T!k+a_N8xF)~Z;OjZ*kvlbu}%Z_!SWO3{`gc4W@h-M-y5uruwMTlk+J4sJAU0W(n
z{!;bY48ukCqQ8-y6T}mFQY5w)BQh~hjL0Mo5=Tl?;whO$k<>wo$fTZ9L?*MBA$`h%
zWcqo;aOG5`1J#StQ;D=w@|6NoW;@hs^6E&NuMvh<f>$LVYE|G>g>p6cRfB$}FuVr*
zYEZ5PuUblf>R|UenueeLf%?Th;5TKhTUqscJJ5+ywsyNMUSWXY0pS56(vKJ(931=+
zNiM=~+n)tpV@$eNT}Rq`82(G+-!scw{&$V_x|vPy34$G;{;=A3#Zxx1B3UAH(n1Uu
zv&0BVSQ3P!EGa@VmJA^!ONo#QtX1f)RDn=M#Rm*mLA(kQsZj^Pbr9SsR)ez|xD;!E
zr2(XW4b0YnFDj4HKs3c#aMn`o!>#WWdh_f)CcKcsPmz#n;<MWyXOCsn8#8eqb02H6
znjmMN=?MAy1X0uroL;a&*|isV6kh?(6-pje-BH&6*_fFD6VDW6Mo?-BfkIC)f&122
zdzJ1b%uD7c>&;t87k#ftH{V0FZ3KEFu&LG^VAcT^BxmO2u9v?!5^BUGHY-*MW~6|L
zR{>Uqay4MpDAxd1gK{lkwG=BVvTTEYc2om~4}sGV=p$eLh2fEn*!8P~Y0bcC2E=!5
zz-gm+v-IWCeY3q2h*w+XzG^OM*tTQQii88#dkBk<f!#5fK?<1oaj-j%ax=_rrcM$j
z-U4%5P;LdgR!YAl%CP8B*<V4#cx_<c2KsnhZu7ZySfOPh@hXjA-U#|ASU$dd$u_pb
z4WdO4%;}*P>A~>3e8<>%W*=yzs^7_HKh1iXOT3o~cq%FaH;nNXT(BAMA_Nz~?jksm
zzCV;b4~vT3aqsOR%O3>&VW2x?MQYG5Ox+<X&_XELu<<u5j}Z;`B{7=^!07<Ekpd=u
z5S$L8To3N`fMgvF;NF1pUEtrPrr^t+N!6pDo^v4bh6I-cBvVLqNknK7cai@5Vd6>L
zBt12Is1uT(#(Em-B95O5=8%@@I7)>^j6tv-L?f04rfH}JkKq~!(SSaZ>s-LnQW1&l
z7Z`6-H!|UCO}XjYSjxufL|#T0A^TEH7b7I0OAwOMr3lIBGK7?LB|<9TsKAOGn@LO)
zM~$)u=o)G_pVDqTWmEj@n@K-?+F1_=cC5{!5i9C{)6z36lWUifMq)nDEp+zsjm5`8
zejlB5hn@3}OEVHRY6*w=3;f081a?I2a(Q%ht!`)f%lzk))mL(k#!mdI$o)XQ*k>L&
zG{GJV{y@#|=N9bTCRoVUZDC(g068hdr<Jc5A&I{PNfM;~QiNoovPi0(GE%AkFk%v)
zGm(Snai)eCHN;ZdS`cd?isS*l`=&U#L&c=8$R8d5Y4p=fr@r^L@Nm#T#CVe}O!|o0
ztm6rDEyBBr6e7ST067wZZG!cr-GBV~ftdsKQ<%5_6EH|OFmWL!L?{#sMJNmlLns^z
zM<@b|KxmW6CKlP}qluIkF||7ab(@dvuTN;6p4hz5<fe~3QC%2bC?~a_)wOqcw>Y~_
zx?5gkV578<%~yy#MFlP@a3x24`CaAtSs_<W5{Bf3<t-ztUw8Jq*YNY}{;wNz(z93S
zE*f%kR_4~~3c9<5caQFl>)rG2IzmUpMHu-^AqzB`8C7KIiw5}1t}2)+i%e!(Ore`F
am}V@tImeV{qSv6AQbi_JWKcymRs1is6`d*o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_reduce.cpp.0A54AE6B66160321.idx b/.cache/clangd/index/profile_gemm_reduce.cpp.0A54AE6B66160321.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0ff1b9016ea90562783702588a5aad2fa51243d3
GIT binary patch
literal 2496
zcmZwH3se(V8UWxK2rytkz(6SB4GE8A31l!ph!T)spaz&w!oziIR#1VU7$5{t-k=ph
z5!Qk=)fRz85lbu_Bn4M$7X^!gU<A7=D-_Q`1?^U7tFS>f<nDIwIhp+5{B!R;|DAhB
z5EdHRWe7oA62oL^Ntq%X1VLu%Uq+@Rxj+|!MCzRE%%vF7{)bG{v9sgN|D*-nQb-E}
zU0QmqQ_UMvtB>n&<t}}Mmg-vDIn~_o$=MrW>u7d6A9QDT_bc=IGlFYP#}%VOg3FoK
z=9pg=MKtqYWk2m<DwW+!?7_B~nnC9$@6fzncCM(G9T=eFNv|rLAyg<Yn|1x>r(Akt
zVAuO@<KU@}U(Sq=NxbFIO+{oyQs>))KWF7HPcxO>daqAEz4twTF7A9sn%|`!x6;_z
zuRBQ6davO9J+E87=BM`D+_;A-Ezdfg9Zqav2(tCAIWE1ZPMD79dSQsabft0Rxd~xm
z#&8pI=~bnQ#DpWoMlmNX{ZsAvGp!ZZQl$^0sL|Y~-hR)PqY_j{tj-@Znu`Ba#ylm7
z8^HDDiN8@a9iCQZ+dqoFEhc7`Mh%vE1>L%yWL$O7Tr83V<u??!wmcDJ=ALu!%A#=#
zl{;I?Hoyn+(<e-JIMhCVFfXsNuih19?A&Ldy0ugOSQ%w|A$hc&zuD*8>Q}5%VwiB^
z$Kb7tN3Jkz4R?Lt*)DH>lrz6Ksm^-x+$j9^u)MJ&J)98UM->^xW?{`m9oVBSW*!!^
z#a|-fd&~aj9GJ46Dc)|EUUG8yVx8%xqBPQPTP`px<uBWXgzqND^E~v*_d+&#(p0Y&
zoiO+FlW(f-`j@cQrJN%PDg=>~9+m38$m`kb;^M|GeW}~fq2ar#(U-Re9#no&nNAU^
zntK*<Fvj$bctyid#nFYTA?IPgnLfgy{)K-Wd=vVlY0Rq8?uW19hFLfA2%q_kTIA^u
z{`A<E^vSp)rJqq-m@Exg_RL}9nmt@6Sb|Lb@1Ezc&$S46*nT35ZAf^PyZK)qkV{FJ
zClNikwYr>tY*F^kO(kr5GBLxydV-3ljf`BtM68ILl+kkQw+<z{J2U1j2l)4flwAp)
zbip4<(Tz0~!?IGY@pKc`{@6Ts+oMkMwK>w<R6&!>`Ws)x(a$K~CQ46kRixDr2AYQk
z%JXgI$N!$(wk_eJaA4|3VXLglD*f^)T@_-nd?Yoo23h{Rskv=TcYXk?FXlB^Up<>U
zap%|9ffD8I&L1vEpNVo8x4pRXHrsqOnM{7E82en8Kjd)KSUEo7%UzO4D<h50-VlqX
zB6i9Y&lr`p5nm|&dW4&yQ+D0tz@n;eicCM<;mhufWk&o7qy0iWl#Go)898Zx4pyH>
z{QSPU&T0D$JQNQ>m<i5=`2M2Kyl+)kL+*Ep4<9gw;?(JS7^;U!|M`Pfj<mW-l6Ge-
zs$<PaI!6CB0$%;NXFVLG8Oh$H$o~`M_9TqQYTykuu!E+bW4C1Q^Fnqxj8i8}YG5->
z!l_|oQ{AU^$v8a5vE+DaZsV~S%NRf@mMMS=EDP4eGVPOh)Mqc@@t6bRV4@ws5pe|M
zgg60mMw|h;ATEH&2pJFsp#XA4TmiWuZh+hocR(J92Oug!T~i!q!s?E+w_Cjwt4pp~
zyd7t5^i00kaP)VHk6T(gFADtks7s`W(!;eSdWCrLwDP8SduYX=FaV(x6p#<a2N09U
zBx(Ek5`6*r5&ZzMY*~)lCj)E)0I_Y^YuasII{vf4U00oz5^d^XXHAJKCfCC>jZ06|
zcCNpEPp;lCt81~k>mcaxJlDqm&M|9s$sCm9ubngq4FVL51_O#gV*m+JA)r_^7ElVB
z0w@(t1tdnrfC|t8K!q?`2m>yH(ITxiRlz#{Q_YoA90${pk)GeZ*{;yP$l5qF5AZmQ
z{vkb5Qzj1c3iIM?<xTabYsH{40HIVAkPp=d5YvXab^>vjuZ=GtKN~+lEGHI4J41j|
z03fy#driGQf<!&;m$L{S3*m)CX}aJsE{_W+lotvpo)-^jJ8wIn{k;8v(s}8C%3xj@
zyyk~x)2iO5dJ6Skp&>N9X6jj_^|%h}s}t%ENCGb*T5}tZ?c?nOR08u#V8Ep?uT<-W
z+J1u^mr26=7X*ILBuQe?%AU-S?8=B<38Gdmt4*K@An2pD5XQ3Zm)#jX&Kc5=5q=aM
zC4hu$IVeUOz8`FryV4TH$?9<1o(Nr>p1uKMxDIb@X1d-a+&nbI&D<z_gB8Kz18Y{m
zrp<wzpkOb{Kd@c=JRJVtU80SxouiX8+0~s&^Ymt*8-4xlnIyUo*OEx#eHg2=HF8TB
H%!mFBE@9)K

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_reduce_impl.hpp.FE75615CA769411C.idx b/.cache/clangd/index/profile_gemm_reduce_impl.hpp.FE75615CA769411C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..82c975a3f571ebd814bd949deac6070b27314774
GIT binary patch
literal 9070
zcmeI2c~nzJ7r-;ufCL^PAYcfPgveuGTtGn(K_yiXSyZZ4sY??`l`6q~g;KYEF0H7w
ziq>ks4{cp)U9pfrSVY-TEQz%hs+P^Q3;pWC_rmFWGk<;Oc#ePS(K++Wz4K=7eRF5-
z%rj+@QmG!o<BflB(#lNL(s_Iyk7v*P%U-J0Yylo`i#3n8vY=%8!o*uc9sfwYIxr-6
zTHrz()xSgzyK_ZJ|2gw$Z}*Q5{_EenFeugI<+IrlUj|uZIs~S;_a4g5-ns6%mFf1X
zp7g5wGjH<F$H@-Hl>5xi5smdY?K}BK<{#x<zudikU21xvuN|}WoX^5vohB;e;o(E}
zKG!x)4K}WF3)p(yY5pb8D*uX?yR1_DdtyA7Cfxt}f>ati+<vWi>$%5gMzk%g`YJ)Y
z(Y0|=mCu>Ullly!8irmCaGq<kXU(}j`hZ{M;KfUp$quj22^f?U-{`QibfC_VV6`b~
z-$!$Q9{9FL*yFN;EdiVJdD|5oA<>)dLzc&D&pp%z&m0wIs5(9On+wVb-5ImBwwiU3
zKdq}Yq&+p1q&fWersby#Jo>zr<~PpScxsuKaLO-pi#Pw&x~pSbXh-qbale+M%e67H
z)W7<CHT&8lw~z8Z@V`Fp;`uq{vz@ZbT6Qm4+WWniM07Q~V9l7dkN9W)T=iFC{jkv3
zF$c0r&%VqsZW{1&TyFf<jTMt6@tzjH9ZLD`$H~6QA1s=7>PAg+V5&>gqOMta8K0|O
zCTl_$xGi;^^@&n*!mn|C_j@yw`n+#<UaE*K2}rm;PB77CymtPK{M0TK;<f8>(LAqz
z$4$HTiSMJ4bFO*bTPE}ky!|LUdAh65@7m(db<+#yRd-%KDU+MDKgw=qcEyB6&e<v3
zm#ezAY4^gKOA(dFFD_2LI6dj_F;I{Y+VtCuV|gL9(RsJjD>IvKB&?nkoVqN2L5Ej@
z&0*TI<YZA)w~x^$ZOt10BezwH=iS=>z$lXMKD2n|H%V0&YW|xQ_d#Cu$%i+lTpv>)
z9QEka&W9c82h$I~IQukn?veJili%-ZdS|Blm&r$G?05-Pf8IZxy()7)#mq<h(_1CW
zPIZ;rz(6sNcCmC3^cOq<GiP}`L?iKn$34gIKreugQ~z%~dZhJ8hkq1!ilfbA_I<Wb
zL-D+wdGuI|vF`l^Prx6*a~Q<5Pr9~eT)@-=-Zo&vbRe^ox%3yfgZXtgG93Rn`>_pB
zjDo~c;@n?gg)E4YJe-u4W&;+Cf`_Gtdw+oyR@^CW{^)e=J{z!P6kIJ`1^oq9*lu`s
z_(WTT9KdVvKQ=g9J99SVUpd-tHKL{02CSL(QcJ10zrgK_XFPAbI9K+Z;@OG+X&-12
zn8*_V;`ur<MpjO(*3Xw7Poj7tR>9N4lTc_$+qKnrYhRoV2$(M1EZv0t1@5Bd=7(bj
zj_Mj?1GbEUyQMqTUtonF4pG6yi`Q9EJbShaZwqgt3sZyFs`Mgx55*I*3Vs%Tgu)9^
z>UXynSKPM&E2ayvrC88k;4ZW(b@oel#=v*=G4&&S?FLZvC2+k2p4>~Oi@>#r<jdfC
z8S{W)HpjoHoxjbRqE((MU))UXsU|3WXnH7?WrSsn!K4Y(5LAppi_vH-EkU6rB<s+4
z9h!*cr6{}<MPW9ea06z`dVP0~-?!F4(OHVD5m-&D&~jv(E*Yiy#;p|i^f4H63}{x3
zPn#gf1VdRC6-hHk?3viWXPzd^@?d|aXsu97&~o8&f>sDuaKwflvhEK#c9IS@<m^5B
z)4mrDd|KtIa%Z2*ke@05liF2H$kcvnLYD5DPRP<nribBn8L|v+1bjLpDkB=pG%^h#
z(?n?qnGTY5Fcs_SAxRG@n9D$11`^EWATB4l3dB_;Uj^}1l50U+OL7B<8!(H1IAis^
z@ZxX_K3$9gi%}5ns00O;kgP+2I+FD$Opn5`ei;fY!z{U8nZ@ghRK4*w;ceHEk5(L@
zcYglL3xG_R8@|=5whdAA?`Y8PNY1)lZoPBKMzO<QCaO}#rNpv7^Xb&lsa(8!T-Ijv
z<V0>le5SSVcVy41DaM{FdK%Vf{o3C?+W`}T=B<m}>pbrkU9k>HoA%1~2PYowa?bKU
z@yZRuR&M!VcaGl?W@&Pdc8_Ow1B%W;${du;j&rp1$+d-+^PV^}tAS7@Vt?k-YN48-
zbYVI{8Nv*LG(rtQh2UHWV*H#!2r9%&9G@<P@r6LVu4CYQjMUdbhz>^L_Ie1>lUxcR
zr6e05#6WT-C@S&(!KbT1QB86UC~8Qq2Sq)}E#TZjj>iZ>BM|l(VW<&?<MA0G&Ipq+
zo4~^aLs*XR=Z%+K9B6Gt(K>L{fr1rKv>qJwB$t9?Dai(KG>}{Yjuj+d0mmyOUjxT$
zB-eps9p(uqY}y?kZFrlBqy6B%AGkPDhCInv9JN2l%oEpW*LXJbq3B$s%*FFWS-kP#
zq8{}Z11UN|l;DibNbpGTCTMs9w=+=mLcu~iEIS1DhhQ+to`=8>^I@<*jOB5k>vz`Y
z_1H2KUm};#Y)(hfQn{37^DBy$$z?Q~5XP^O>2m90vKSll!6_en*k>n<+O^m@aN8#q
z6zvx67SE<m6unxxnu`RAUZY$?&|2kMf;K2OaK!e<#JD;8WVhL=g4mKwm1qa!$zCMr
zMTF@EU|)bun>NM4yW`VH3uZX9fVSgOJ=#a$W5;D>w7<aLj!VJlAVH8Fmo?HgNKk`2
zV4VZj{fRC+Ac|ImD<-kYu4-xOHm9K78yN5OA+J8XVX4Kc{p#4>&5WP`f(jslb(=<;
zy*%-+!f0lk#fV>wY*_(Cmmq!#$vVW>k*r62J;`N=Uxs;6{(SEVZE-7@cY6|vPa+~6
zyomfSqEWb`OUVBc=FGkCi$~v*-e=bH6CghUTyDV{X=&Y`$zu=DBHEj9u*k=oaInbV
zn{cow$eVO<SPe>K$C3SYbMdY%bJsC0dW!6yA}3bBhRXo!{sRK;SLP-2W#<VqOZ{a*
zEUheWyStj-RL5j_n~-1=60*;t=p9I~gXEn^u#@BpB&fjr;mOIF(+@UXVO~!K3a=ou
z)=Cs!Npc^0s}H5(j&k6w93VIk-pa$g(R87pFz3{9##)6!5z6{@^W#{z@P=&?X5cPz
z7cVy1r06>2S%-)~u0xaRu<t0k9(mSd`L?h2-d?=6^A__CYfy3xn#B%ld&hTKQ)9Y!
zzVfVKR|6&<%||WvCF~iz**Rvt<9eo}UL@*8QnsT#y>sV$Xy?|#*tZ`-_d^UTI5h6(
zNNIAs?VHq%NZk(i_$8GO?lmyW+Fp?D1uxcH6#W^<KEwRwqIiB>@}~9y=5x-_W8@~;
z0#PkMr2Zy|Ho-XdK}Q}r`QGgs^0zHTUqSLKh&v!rbTyJ!lY9;$&*Aq)(dR*V9^%;t
zQS=2+ULaWqu{wyuay?AY!z9e5FufGG8A#Cvm}!7CEU$!#m5_kB8YWi*w-Hcu4NR;d
z<@GSR9?QSlGUM}m%XDsx>y|5)a|c3-UZGgQ(P`iN)8BcL9AIhAjTD7sE>Z~Lg?KqO
z%To$r8ao<3Z3LMSuYqQcFanqSnva)46V6?}y*@ym-Z`Ml9M<uQL^iC?%x)ciWSgCX
zh&f^lRKyyu7N`kI7o-!EA;=&|BhV032=;}*t&rx#Cs+rZx#q+tSWj{(1ecO*fM5g3
zmEc~9tz_l|xL1>01MW2>*MoaK$t_^tLXO7>0wZwaJ>R|h{NRtDe*M}{0!6~khD}S}
z=e=^~epzVMST*Br!kblyScTY~83eTXqKsYRDf$5tJ-~ZNadpeglM%n&d+l#i7n1Ux
zMeLDo0YxW8B`Mh$r07Nz(TK)lZbA`FBsZgoW|CV`bSt?vn^3fg<Te!DMshofZYQ}5
zMR#G=$8A+Ss9R!Sf;VYW(iA+5IuuxkxMPZWk4~vWZ)15q3alsPjY!#uh=sHXDVs=c
zM#^TATTy%~O2W@GqA5l+6|)J&n@IT`H02H{Z$t5Iq`V!)x0Bp~rgV_{ooGrY$z3SE
z3+tC|p8IcAhiBWHkRd{bO?#P;F<%|ykZD1p7927)+T+R3e#y*!<2%B4wW0|>JhST=
z{l+PTQ|gDT+45o3R6BE~FkHpm(NMH{xSF8!sB|v1qv#A}29bYhlp2DLLD(^f#EXUw
zX6j%LW<AW%!)(lDATPuD5k;4Syqx4JkXMm>735b*t_685$qgWHAh`vUEo8Q6f{7-4
zx6=^mwKgTuZSQN}TblaZV>32NoR~CmiR}^xZ0|DLWgM}W7R(vn{ILf|v(!r(#GZiI
zTRuj~{K1D%8Y&&hD&3TRZ(;P8*S>OM!k(O&1;zZK2tWNc65YlN3VWnx)XblYu-ZK&
zx`)+%4xQbed-fhMR<8?MHw;_7K4?8hZD9@LdryebYxnstmtBShJn_zB4u{*(@a=eq
zzIX8XSf4oC|1eqdr@;K>?jpLaPV9uy-+saze~l>GNN$FAQ1l&=J5Y27$(<;=6La_b
zr>2Jw8du2}o-#Pa9~+)JIF+OB{8N7(^l$irxj3A~pJk1e-r>JP&>a38g3|bD1ij0D
zm!Nt4c?8Yp&nIXBe*s4ip3VL%;lQR*=8VEF$r*QChwSSRcd={EC`Q&H?yzA#<=fYj
z@<!y;h}>}dCgjvaax-#jCb<<!S`l}lV|KE{NU{k@OeEhyk~<`~AxRs_?MTv2atD%h
zklcwRog{Z5Nf+iP-))#`)jG_TnFjf2Tt2?ecv?`T5B}D}^R>^&zqdZ!@i54g$UL|Y
zc`{dYtn2v9EF8kjMa=myh`HdzydPYdOI^$b5Ww8zvCMv@W$viVpJ;JBO1vb#?A0N2
z%NigFX6Z%37adD)&Z=UjXg-Y02kv8+`Nkw@KXCE+qO{&?f8@%7|2vnpszvk6AGx!&
z3$hQGE6o39VgGot6&_EHtFE}}f~zuI^}tmruA09z|MSO<t?+mfTvgyIz2atMW2b%1
TOv=UGRpKU-dnkr@N_qbS*I+l$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_splitk.cpp.6BF00CBB81DE4660.idx b/.cache/clangd/index/profile_gemm_splitk.cpp.6BF00CBB81DE4660.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b4f951d643c468b70ed509c9f201e33e9bcdfda7
GIT binary patch
literal 4858
zcmZwK3s@7!769Nq5H#XS5O7I&sVGE<pnwVqU_yDxOGq%FD6-1K7kX7JqP4YGv=%A?
zA_|pgBL)gZDp3PcN)<Iwv_YdHinbT9)n05Xw)m>$>n3*D>Fv$;G1;H9J7=>qb25=d
zgoQn|!m!}9h@AE5TUM|z44X*(Wo_BKMrVp)Q_L|er{bi%cSEnA<*lyA@m>iPX{Xm?
z;s+}m!&bIGs_K{)6n`+J%=zr}(o2853Ox9rXOnvT98WT;!sp|Q=QfpDx`fQ@I{Q6u
z6TeehEZ!;XAD5hR|J2ay{-@bjJ^2?;2ma?*!8Ucz@1)vK_^qjK;q&)z+uV5hF6)Qe
z16L-eG`tvkUgpvIT};pAl_EP6@^SB$3q+&s^O7%q@yYfweR{ON_m;f6tV(q<EouAF
zO`kj6xH~_hy{T{ePo=SYPOcq%DT~q^&@E7S+N(}4;wg_rzjuCQ;82o><JLh_?%GvZ
zDpjxSo?`lM=g(@bk0t6^btk@FwD99us%O$Iy5<+}9o*m;nVvr*VC>#k;s@RG06Vuz
zlWjFwbp<My>FFkY{%cy)Rk@d|V^ikG|5i8mZt$mvc)me-%hM-l?s;h})5At0%%n2c
ztZyG3buBqtvTN+8K5lci2AtPDkL;Z?e)kiB{qjc-e0Jcq2Tnb?FxPqULtpohy{p~N
z&MM)^ze*3V=qo7Mqlv1Y*3!PD?1QF;jO07o`09IW&v+kj|2eM4<3+qEs_>rMQ8zgG
zen;~4OQPYr(x*<7l*=L%(wV#ZlP+wGE$qpiFqrhLDX61EzU9Mga%ubPBKLivRsMB9
zEYT#t6Dth29)7=Md2{&e8hoK!Q{TM&Yks+YKDBYaO+-|nLv~V1(q8Q+r{m2J^>bHC
zAOG&E{6+qSmmu{R9=K`M$}Ngsd&R?ZC(=tE>ZJqQruyb<FJA9?crPtk>w9O%aA=YI
z<n>6$7t>$nI|T=Xdqupv%xhIq%j@dmuumH6)-HKvdF`9}TaCKZi52b+pI=#Uoy)q>
z+Bn>qbSJ{Iv&ni!|Im&{0g-IJb4ql3j@Eg&g!^ph#dX=zhFxn1d_B%BR<7tflu&A3
zS-Cs&t+!`f;CSD>I?v*6^f~F>9NX|Kch=;mrsi+S5&7AXah01Z7dL!mwSD3Ic`hkG
z6m0E3?3P{NYa&RKd!?+Klwc~o9g*Kzt;^3Jaj3xe>s8)sYSvCYkmJ~JIDGTm0f7+A
zG)W%Y{|i=6Qf`_a2^kpq>Whq~+s%p9LSe=q;vYXDM$DVnIlXu{^I&q{>y}BY96W}0
zT)C|o_b+UR_2;5zPZw;+c$}fj__e9*yGxfJSB0N?RK(S8``Br_AoJf_WfyI>XT~`X
z@BR7N&h$SD?(M&ml-&qf>Nsx8VB$!l<D#lPmfX;UcX+FF_8nC#1MQBS_)xrm)=-qv
z+Rx6<YwYaDSJzba>8CCtCPplGl&qL-;^>*|Q1Ku}J#Sb0ll1Vf{U=3@nC(83*O;|!
z{mNT_VOBEpnUkx(Imf}05;2m?;#wISYTX-O_8loKNp5Zi2V1@bBQ01KxUm@x57sFL
z3p@rmSS)q5Ez8!**wEpCw7U&Eiv~~N*lG+Tr?RHn8XGz+<qnMv@;c+f!IG)YC$J{q
z#)b|jB)_zpEQ;^qV9ThmCCd^wHgtID>fzPpM872mTS|pFEDmmL=x}b!@vc__(;PD_
zfy!--TXT&K9X@2TqATy@Ae)25Q+IC1va>KYbohvNlg*;Lk8(JeoC@<;JPTt(ht>VX
zCBKOupEbi4Q=N~;$2%AsI@}euQM^C>x%UKY6%`%;Yyj;4dczE5m3mz(+9<LrnJndC
zGAe;R%ihY^ya{61y)%2u?8_35bFf8JcoJ(8$Jo%}pZf(~OO_UTaxf_sCRhZ=*wA5Y
zb9zQrEY6HBm>WlxekSy&jxjfmCjDt}vH)ZSkTg1Mqyl6L)b${(hb2tB6{M}Gbs*P4
zG99=7Yj#lC4O4^8F0yP>oGsrvLJgP;+r@%D8z)`au83x^XCQK8yCHICyCV{?1&C&{
zXCm@odm!>;dm@^}o`q;Odp063wihCAwl^Xlwhy8?>^Y+nISqBL8?#~vSmb@Mx(_z=
zhoj$3%>T9maB5a~4nPn+VWOixvjmeu8ygn+5fC4NEqxt}ECiwubqNq9s1E~i81)e#
zj-WmU#4*&BKvbfx2BI2u9T0V>PXTcX^%)?}pl$-933Uq)EvU}}aUS(0ATFWq0HOo+
zRUodSz6QiK)Hi{+iMj`f9>#QOp=P+Dt}rt^4`T8_PCr(hEC*RRI>T#0T8qx;W{@_c
z?gV)!GsEc;MWr#+CnilP$V$;}RUlKL)__!lcB=)c7IinsyO}hLC3m@9*{#jgw0cua
z==nm|Kq|kX05bWDK~{|BQ-Vy1S`AV)ny(F{ZK(Ai*E9K(D=vN6_mg%0Kk}#4)M!4F
z30F}|*-zpZ!aP=Ch;RWTk*mm^k-w`yq5#(bM1ih>h{U#Hd*(Wctprh!Z4jbh$Ka{V
znIVoLh(aAh5rsL1A&TZkJ22P9aAQVA=5FhG+dMfAaB`#bMiIR-Q=jT%-nLHhF~Lb0
zF5}S^7bj}~uYpl#Sna+uJoXyj#4utG12ByqF@s$OU?y$4C=%v>`+B0Dr-z@AxwAa5
z$%8433xHPuE{uzTSB$z8c%`Vzfme=N0Xzk2CGeD}RlrlBt_5B#>U!YSqgDe?jamad
z4eDm#HKT3?UMuQ0;I*OF0#A#&6L_7db->f1)&ozEx*K@ijCU6u)Nn4`P*7v}X8q8^
zw!3oY_LxsEx0~W*CO>mFJ;*q@p1&T^2L6Uo?R^sJAne*_NA)GOk=oHiK<_8a%#Hg9
zdRTF?8TifMMW2R~t-x<Z-3I(N)LP(c8Pod^s$0V*gz2^x__fSwI9U(;demy*t5Iu!
zuVJjrjABJfKj=4pPrY>?MptgbvL5f_`#4tAr@Idm?!!d-#_7^R%~Qj6hkk%KDfAEq
z(kJ0$9{A@$6ypMr6hIu~VvrQ0E(J*`>T-~jqgH@Kfm#U?C2AE&RH$n~Tnn*GzIqVX
zqgI1hjama@4eDkPH=}L^aVzRJ5VxV$f>?{X6M{P-p2@F+U>#~b1nW_ELvT0a(}yHe
zEcN;?|50bp@W;uEzES>24NJUr{OGQrDXcW7Uh=Sq!y^4c{UYeJEK(RDjA0}ah-NYJ
z7x*Iz5Ck9!6a*p?JBUY14vUmHNDu`%1R)A`4t8ZSggA#F3Uv-e6y_X;DB3Z4v>3C<
z7{{1Vwa$^HnS~x7q&{(iTf*r2(M@~j!(H>^ImY=Jc^?+iS?PKTCjYa<>5VqEf(*NC
zW(9QuzZ2%t*W;uP_&U^j;OkL$1HYSb@1E~(w+@)8sT-Ngo{Q7n(ED~On_<(=WXl8R
zJn*1TGwiha1>nuN82H7gOMzd?c#vIJXd*gQL}g2KOPoVrM*j?;vKf8@Fxkq1U(Q^O
zQ~P+n0<{wOO4KUgs~A5&*gkgK^gbKoG(@+j)bEPS6^5mswK+8_-f)m&Xx_Yxfj5V#
z^#3qZ(&4i`0debO9b?kcQ&*;}z|u!!SQ>M<k~vIWpO(3XV)UworYy5DW7*@z;}a%Y
zax5aP!a`<PnMYc4Z6@*TME-&E#gd@6W)mI()BW71{J;O&IygE_b)M$JcN5I?^z!zZ
b>nof*&wkdNP{P)AK~#oGcwBG<#9)5{HbH|*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_splitk_impl.hpp.1DC09EFBBC454E8F.idx b/.cache/clangd/index/profile_gemm_splitk_impl.hpp.1DC09EFBBC454E8F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8b7c3856117c489fb79a5f080bc7ef8f932adadd
GIT binary patch
literal 6752
zcmd^^dsI_L9>?<=gOG3&;3^OS6)B=T6`~L!t2`t@5HKicTLD3B#g+O<Rn%5gL;*n*
z1(in;4N$GMSbS838XiJb1jR=Utc#^WD`2hIEv{(0H*W8p*?;yN{`Q_j&V257p1+y-
z&2KIda}<h2Ry<z7vN=hK%Ri3e^LRWn=1=|cx`e5S#}f*8yrjaK$k5eYQ%!G#U)`Sj
zUixX{*X>y+gHvnHJSp9BAyaB<oP2eIuUT($Y*Sc<<Aa)cPJu_Ny?(F0TJYV~Ukp1`
zKUj%!jJ_M6Y^MAp={N6Y!;WWe?z)i06MZH>2#J0gal57=^y{Cz5*9o`pBAjJ|1CPA
z?D+VWksad=4!BJ*k4cF$R-`Wt-X)q9uxn0F1#jUAkBZ~h<4=!Xy>ZRC%sZ1K_6oGl
zQ|Y8fO(k{n`3pg=Tr*bdvZH_(Qy4c<uITkFaoArK@_B+L{%NIP?b(dE{qJ^(+>hP6
zHE@0JQgL(NrVYum{J5>M+?^pKb1H6UM=6%x-fTGc!$0bEhTobz-<W6eyrExg@7eUP
zaCdpr)IBAYQiB7d=U%#7Tm9l69rupkt9X9uZu1KF)}K#Kt_!t3x5OrMccP1_e=fbv
zDI_>vXO(acQqt|K&otYe+!>YEys6VOK4+kcKYQUHFJls33sP-@zM!J4=Ke~%uFncy
zBo#($Lrl^8H;xPmZ@D$qe-nPf(1suMy|OyaVnfJYRnMxA(h}Olo3CjR74)>%FZ|Bn
z*^mom#h>pUymoj_Uz(-p{3mT^|2!V+)ANN_T9ofXN#=|*VS{6gO|MOxbnoQB^$%8D
z4!fG{@^w-7)#;nIoHzR@d;jC?TCXCnV_Sk+t($lDY*`uXv&`wN=C`_kzjJTLB}vwP
z)wPoLUozTO|CoMxU#t7Xv}H|^6)ik7<-rwyE}iSP;P<Z%^^bBGGCp}wHK5*<81Kuh
zAIZ{R`%gSXg9up~FvvW#L@1$o%yQB{OHwx7UoVWFNAWC1@@P8)yRmP6cmf2x5qzLw
z-SQ9ALl#WV*voz`Qb&^!6kQFj)!@mMq`N`Z4StyWz_kzSBixNo6gS6bj-=>7`#?vW
zO<}JfNa>;U!dmq7Xm3n0(_;vV4Tue5wUI5C`}&<KgovU;?LwvOYl;rD3*$)ip_Ny{
z7dBeNr`ut6J1E#}e7YQ2m!q*Pn_e83D9sIO<ukR3LnpAWncBo*1gXR-f+EBb99h<F
z%d7n7;n${oI?y_htCu0CKo_<;KCQ4;5SqeCL1;>AC7~&0N^azQT5YAa!L_Je)NTaH
z)iW_^AyNwqG1tO^T8P5j05%OU4s#ROG?6?6RzpCvH3ZX#fSVsa{R*sJVO{e58KYO_
z7d%Gr>1fwzqEj)hF$BfB#&TqJ`^s8guYCDiW+`jWryC0<W*-^$F0vJ7uSu2AdCVG?
z`O5<F%*nmwe`CpR*4c{#vYWzT%!B=-i%r65F$;#>AvcTWwTh}rKc=l}u&oAe5AYd(
z;M@(wlI{cBKCDj(Ki+Gx*7?-1Tl60H?AqT)#yzuTwv#x}lpQp)ox}=)lwu`8YO$Ii
zEy%P$Y^Yk0)skEf4)rh@*W3sWjU?;9N{5#{pVmQu4v5KU2dj2cZwIe-n1SnShwye#
zVNOAG3L1@hH==iwya&;HNX|xdHpvGOeURkCh(1hm5u%GoE<tn&$r>cqAi_QxG){wT
zaep+(O@pSAJX3>wF_$87DOtW0Ih7(}=1Y;H6mia)PnRKa8JWKfO)f)DxV|#vUxsF5
zE=S^WoZrWp7hrY4P{_DvEl9NBz-FZAT9DL|TnCamlCOZ|3T8!h>MyJMKPWb&XsM~x
z0#_rQAe}^zhg60sAI$Q>67x+kzX{ePPreD9f1~I&FmJ<ps8`Nv3)jqj%&L#{iR4Zo
zijMM$B50w{LXINt93RnpY15*&%tg)@1lA-jH{HuvDgy)=V1iA29n7y|lTvgSjOl{$
ztiB+Op7is6xdBl00T3O4QLGS~RJeLcP4$nAkd7SF@q%VEfbdPw!-h&r2iii{kBXHt
zLB#{09>AD4krMMm(fvU6V_vpDeb;8kZC8iw1DA*T)oedg7SHr33pr+?DQr(xSkG^%
z3EKJ*BP^3H3t(R-tW$3Ye^Kts*d&@CEn;6&bPPX+pjdt^M~QhWCEgw5dVo(4fc*eW
zWDBiLn%Qb|t?e}<9Dqp&uzORVY^%;mTk<K>c0NqW2QGfv@#Wma$f8?KOzXv<DuyL&
zRTO;`R7XiJ0aXd+oyU_mhkkxVU$7I#cH#}FXP-{sb~;5rgs~5?zWdien;ET{R!mn?
zkz*>hPp;)|3%{+Sw=$j20MiWYYjX#eMDH-Mxy#Jid6;k>TyY;;U~&sgWBD`xPWd=h
zi0CcbRmEv?lV>Ymsu_#wz)c4ptUZo)3ML&5c=tWVqK80o2*z+iztg;W%fULPL?w)^
z1bbE}Srx($R_*FyLXi$|>;Pw0(1+8`<%l!=v8U#ak9yyE$++1#+c<6kQgpm+JV909
zTZIo%y>FkRg)p{Ky>Abyg;31(FsGi>8zH0-=CQgaZPAhL_21es3qfV4;?^)lN7zMh
z^p#^z<onN5E{6Ie+)GCu(K-mz;o9^!2FG@Ch@~j16fI&~(;sP;Wr#DHo~31II;pG5
z(0rE9EWYn?EYpM@#axC2f!t=LXoWyQkW!!|NG(tkqy?!KhvItkNNY*1hY9s~w^MW@
zOlTxo2j)6F%M?8X=0hZ>AW;f3V>?VS0f}fg$$OA!56Rg`luhzMBsxg)VI(?CauE_0
zkz9gAB_wN*K!dpPo}DC!+;^ko@f#Q7(TT5>dKWUU6Bgi(J_L`4SU=ad`J83Oz{z36
z!rv%zlTQ2obf4J@UEt7#ue2A%k6U-GS;=EMF}%BPmDuQJ<Dy5QZ<({H{_gI(U-~@n
z9)|6|offgoAn;fpV`9&lo?Hw>(K*N~2hGB>T!6d^FkcQ%w|jD9O`U)#Hpi228rzY}
zcI1wWr6P}1#9d(Zfq*m}5z$>1^2j3f9OR#al(>8z3eH2}m<y190jU?G;9{&dJ<i`&
z^t0zJ#=mrcRR=z%__PL%(I75t)U%@o5iV4NW@wNP+mK!l*C67`P>RNvBI2@DihN2D
zcLTh>?u2US$3*p8w&0pgXMaCr((nD24ZG!?EonPu&Nne;P7{wn;)|TZLqU+zS4l)#
z>e*`UMy5Y(qqPvnE;s#YvbYupa1`AD_6;~hr06EFZz8z^+&h3f?kMKNkliaB$5Bi~
zAT7uHs#WgrNkq6!-Y^`)t)9BoRkQB=#5m<ba4N({@x5`E40O)6&BJ!eKge1G22V*~
z7@c_ZP_}Au!-!(0$nZXP^ckJXM&1?sm}n#q$@Bi#W5|m!_pLk~Ib)Rn6(*z1QpWiM
zik4f-IqExb`u8H|)?(%Y?92BR;#u?K`w`^N_a`WTAHdO*p~Wvk3wC+w-KB{qaq-=b
z%(f#5t}qo@q~h<QdUt7&PI4Bq$RarhndOk>^N>XzsTUx#0#YwV7R6Y9_I2`nqx<gG
zOsJXUpTv3Dfx?TIr&LU^AI3zz_<-Nd^GUwWtQH+O=<rH?dHztXo6q%1rinFvYhu}v
zzSg!l<ja!^|F7A3>z1#I)87Hr>z1l3^$An`&wn5N|7iYt{`#b@r%~)X9?u*n4RMmf
wNfAzpancwkjc`(klO{MxZ#+BHJL0qUUh$~W=7s_zAte%<n3@^$M&msH0ihxFp8x;=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_streamk.cpp.46F4E02E0912E672.idx b/.cache/clangd/index/profile_gemm_streamk.cpp.46F4E02E0912E672.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3931f5448c186afc8cd28551cf4fa793664cb2e9
GIT binary patch
literal 3594
zcmZve2~-o=5{A1WNC=HVz(9b2EJ=U>lCa1UFruKKpauj4T0j;T7Bygu5z%2AjT%9t
zBI@&jiW@HAGA@WmeCPv}0D2Tq1cEazxPT7w)Ny>B=uP*WnLdZ={<`k1TdDeQUxZ<y
zp*rIsXnI<hBs)DXg#<y+1ne(2PqeH_7ki*UkYxX<1?UQuui3qW-A6l^?c%GRyCg64
z?@pqGoMVi3@Z>w=e8msNJ-q7$+ryj`;@C=_iDTpQ&o+IN`y6K&l;6wEUK{#kvncMF
zGU+qFYTuBv)xF7Qa%R3>&i9=P1zq29drOhgVDrlscx8bYYN)BW`_94RIvxG`KUwqk
z>oc=PzUJI@H{Uq`L8*N#!;6=$HQ4s8KUjB%UN&adclu01v!h!!^M2=tt?NcB9)1k-
zzh(3&^Gv==)`A4(p7{D*d7-cJx}6RjX(jijz1RI{`^$snFE1t7-k2A#kK@)WSH1PB
z(m(C}udYR5re}AsqWqRMDtcyi_zkJP%)heQX_s^Lu(kE8Z)FLwBTEl&r93;*R_C*O
zcbUSVczvfNJ;HFy+=GqzH7~kvy@0CXGoFx`{=2+=Mar&V1^p9y;*7Apch?>YsGRdK
zl4A4kw!`gDMxcN`=j<X^<EArNlB%7`VeYZ;rY8MIo01RooiJupAAe9aCn&nNe8j(a
zN+hkv{jz)*UXo&|x>dP4-D>VB(b109Stl+{KO*j*`_eT~bW~D9I<I5DCd72;L0wGi
zy69pNb9HDPS(d4HQ#Wui)u%SIdtjWMLu`j#Pf<4G!}Jp#w7ofXl}oyF-UR+?9q|B;
zl)kjNahT_Oq`&mBOYk6~x_d0Rw?E!>G`W5G`eh8uqBf;doGdlDNcCij!TK{j+ZNKl
zq`lZ2V0tLC_vtIXlIgaOzejGVG`zgP%SL`H=s7FXv+>?<CmGXXE`)L}^zVGeks>2D
z&5o;AHSeGH{{6vF%b%XsFTdGYvN1<&_4U?T*TSfh!PRtU%c5kTcj&eHZLb~8)A<Hw
zBhsMrHq2tH^2sa*eroxZFyxEm>(b+=;-*~~-M2fjw*BnN*(bwGQl{GNo5u(rE<(J|
zZ7cGL%Bk>q+T0;3*sbcRuxIQa{TzrakIu2+e2fiNw%o3HcBf$^@(*5f<48<$Nv%ui
zRfn|~UTwY-DO3clj*0$#$-DE9=I&`deB?yQEY;YiYy2HU^rRN9@5+bEJ64UA{q*hG
z+iu7BHQ&<mrA=2hyq}n`Lef80XZW#X%a(1vgU7}`xFtPiHNa=t1-HBIIP7if%4^H`
zaH(SM;9{RQ#;);>S0v>%x8;8@h8K0ssVb5(-%DSw`j&9MZ`Rw2Q|SS(_wFod;ts{a
zS)UKhF|_koR{N3H-cMqr-mLlX%d!a@_9qVP31+u{x@0`l<Mp<bnt@Jk^u>rL6U@e$
zwhio)WtKODJ&YNyI8}SFA-4Wm64kJ<C3kK1VhIdECPIp<?a_OesZeY*gj$d+Of&}D
z;+o6qzeyxPQBejJN{E5b@ucyH#%SF`wFic_xeZdGx!7!TlDUb-0C&i-hk2FV;x+_I
zhal93WMiQ*z&+n$M1LChJd+B=W5JC`Mu^4$cMST;#MVDbNre_-Zex-$qA|douyl0k
zUYg373dLb=Dv63{3~=Y>@jZ9lX@rcRG;Hlk7^;M=e%v7S=M1}w8mg6_Il2-(C<+^5
zikMnx%y%aQ)vL;P{yz2f1wCj1<~BqOtu+R?pPc`>!m1*+nhFWA-#(Evk*YDk{Z_@{
zE=<_&PKD-UZW@V3)fnIwrDx=}=wlz!+TnZr%0GY8LC~p`sV3Trp#nRBGa%3;n4)P8
z$Qv|Am!fG3$QMDyLUBBpi=d^juoTv^)sL28vDEznvHXpj4qVu~=1&Ac?aB7z@!1Y!
z2Ld^g9SOuFGYP~Zvk1f{vkBxxb|R27*_l8tWETQ)$Q%N>l3fYpMt0MZzs*Q%=D3s*
zn1m`|69sGrQU-S&%3s8WV9j^0goA-DfcUWDsg;KpPZS|q0Ae6$AuKG!ix*`OaxFXE
zSPmnoAV4q;3<FBVVyR_{$0`sOXbX>^a#$$GW5o%2ER^ELE7%YCu1=5{5H=P~t!g})
zR464xlfgn69u3|9pKAkeOl5zImL%-GpK`7>Nu9EXz#F8Tbo}_mlS02GSiiXf?hNo6
zL3uNHvvBfcdAi`_#quJMH_MwqJ}e&s`I`G$;q&~={Rrf5?oS{-gKvkA3}gfnD2NfH
zr3{ne?os`AwvM`7)|IdWD3_X<m9Q&daEEn%)D>(wAax4mFWVd|fE(8jPeEiRnuL>>
zE@t8+rArA^KrbLrAxtlX30w-(ONlIl=`teAVY(dW;xffW>W%vcur_^H1IW7g&<1J5
zvF6*l2pY$Z^JutBo}e%aX+hA1?1co1XUA*l=iv}*Kjl_SEY@_h>6Re<s;t?h@cE@G
z&5j9Hz}lwQC3f+uw^B{8@e0^T0Z#xk>Qon_t}630Fh`<&VgOhhC=x5A`l|>s#kAwb
z3j}i!R1_eZjgwd)o`;iEASF<NpnyPyu%Hkoa49S(C9(__$cQY51#+Cv?G{Wj?(Dqw
zw~cTqbeH6`;CYu33FU@x!$2-1lo!T}z{!*2IT<G}ju(NvIo<^F;rI~9*V<Pr8xrbg
z?MEPgYkvaq?fEQxg+Tj20tMLz5lG-8&<Yb2pG-pYXnE#%00~V(pnO_Bfi}XljWB_W
zVOlYfH^H<`MBWP1wi3Avrj_B`=p{_k3#lH$<PhQ<!UZX`Ijq}M7!*a-$R}I@M*_#=
zl3yP+4(ioo4tKIU0$ziIQ;(uKFu}IO1Y<991B0+bj4kHjBxOqpRKPCK(h#|JyN>_a
zGHjJl=TI*&1)RUwD(b@*U!@Rc7sA@v>f@1JN@N+#mJwMFv*kFCDQ@epbsR9$$cuQM
zK%(@l6!lif6{Y54|4%?rqS}1F7J&x{O4dHYXuI=a8fU!8ig786lamq`L5sCMNE7Ni
zCgE>Wv(vJcVK`?=q%KKsoIcsW5HXryOdTI?5<0`lgc5FQVK$L&=^qdnG*vJy*n{TA
ycl33c^iOZitr<3U_6|(8Gso4P%k%V};%Dn)HQ6hKX3m;5J5^`qyy;<Z1oVF?vXsOC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_gemm_streamk_impl.hpp.73FA01C66CC8D3E3.idx b/.cache/clangd/index/profile_gemm_streamk_impl.hpp.73FA01C66CC8D3E3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bd495584dffe698603fa3b1db574f18f0a75754d
GIT binary patch
literal 6022
zcmd^@c~leE8o=`nBM_Vf7-OJ7DIygK$`U|US&~RG0zsB4n}7lj7iti<ZrTJC?Xv=U
zDzz1<`kp?Owx~5>(NYnseVWikpFVnu7Ligw5qXvrQQyRN=H4F9={f%G9LTx9`+awr
z@80`;-<{;Sv9TK*cszOj-1Wr;tMf=6k7v*Pm#$v5)PvyhzPIM_*6%r;mawwh&(5Iw
zdY|`6I@_V|{A-s#NxkO$!=mPO7eg$LHO#KQ{mpwxj~_UNi&C23+!y$+^i=TP*LVJW
zAz$x#bi+@bPJt;^*B|F^fB0+PKWpmhEq`b)^7N0Fgj}DS{aVV++V-^Mk-wk3))<}R
z>BN6~=jXrlwS2$xie6S!QxjK{+a9m(CHs@^&dx6vZ<nYi@6z1P5v3@aKCKF=OpQuT
zoc{4dKd1a2t9@tX<XN8U)J7_oK9j~%8Mn}Prmm??IuQ4x$7R#487)8R8*7GIK2mQz
z#t+}HvLJB!ZT<Px)4FQExS8X-aY?qXDt;h(@q(<!c4wDRI~~I^ll_!u8f$`G1y+|A
z%<5k%-EqNT*5m4&@!e~#rt8F)*83Yw`=_@Zm=!XK-c~WN&`u^vjh0+g7M|O2*+gxY
z8uFKKdf6rIy<G{n3Jcu>YZCb7zT*#19vQGL>WR(_6fO5V6DR%5Ik_!a-AB}xC0`D+
zh<|(YvYMTFCn^tH(V`;#*8in1yyg?CNPqWm-oXUVYe|-cxzTTCjuv)rI9U9~?Xl{@
z^+)43>9)WAZSKWaibpqE7VW8OJ6XQqO9#=m(Bp}tQ@oRkI;SpAE6UdE>kjp_icfc4
zI<S1K;%fD;cDoHRrXwTD)lq8`T_mY9*Yz~4ZBNZN{?;^dsD8q_u9xh}`r?-8vwy2O
zbK!dX&!ptwxx5#<L##eu__&fk`|M!d%BHVB81(hL8TGHy*NYcNGkeBy$=$)jH;7S!
zaG1a#^U#hqjx>+|I~?oQrL=UsRY>u~mOOgOgek61et3KWcornku&SW2)JMc9=d{Kc
z1H5d3qK|{uageetE&L}QuDi5&n<YidJY-&MHATxk<QOTWieMzm2+9aUlo^zXQI<R_
zj+I#&TL%Uu4K@TtE8Ud=Y&AvCbDzhNP5oU%#D|Z1nW=fd<n75S3I663y!5Ecmjp>0
z!QTj+K1r7o!g9inRkUkzEe?1;u7hL-5h|y!)yyEmc^D-M6ERW=RUA3Ay|e2?*9Q;n
zNm}MCdy(ycA&G1TBDu31lgTA=Or~&FU@}FBf-?w7tDV#?s2BANwGW~aVF~vhBwdnR
zl8R(Hn6HBr#Ck~5Lo(tfaBKo6#4X_1f^#Q`JHZjjJHfvbw_+!xb;1Hx?s)Z>_4pyZ
z)PkflW@cbh%$%8tQP#{Xj-0+fSHkO$D0r^z<~(a?-A6B1?tbcEx!F{zH#yR~nEe|h
z3!1~)nWRI*LT9t&9`$+Cy^wEgp26xKJx%9r0vDw|HK339Qb}y{i5MW6Bkpn>xDDy6
zI?yjJ@veWWMVRaGzn5F)4ZAX1T`0pgqFgA)NFh{Uq!y|%(m|jOX0fkB(s~Hg<J<^=
zjW{<$U^C7J5F7AKYJeyM#Gw8RkZOQe5gWn72-6VX0Qv@q5cdGxgL5CyeK-#SJ&5x#
z(8D<22l_tFV?dAL{0QhrIBN)@2Du`V))1~5!VS$wL-=S2Kb#{qL^NV8A=KjSwS+`V
zV2(vg#A*r7pOSPLAuPl7%ZTY^gaq|hMkvdOc$UMxd2%Pcl?~$>bs*A#J1d}=g%as;
zZUj*y&dnfdMjU%;)14KAg>_aG?PupFM&0;L@#8!;Ba`}tAlY6J?1f2)+rhpa9B}q%
z2X5{Z{T0}Mh2#mrJ64KkZmY7O=;ZL^Skx>fJO!iF@Klac_;XGd7uaoQ!c7J6E5H^t
zY6bgN)c31vDo5(>M3T%=sRicGnN`eq_2|Z*)}_ZWM`dxqVg*~BvwqLY+|#FSGS$7{
z)eGE$=2vgs{D#*%Uq5x`3KC*YYmOaU%(UnSuYT}lyIQ)cbgjyCD4JQ43^GH&R#S8)
znTb&rnT1g;Sl5C*t6RKlnPb=wuD#5l-+|kAzy&hxC-KE;wcnm+`mTb6D#&76P;@mU
zRO7q@5_aIcA7b}I8mnJ%caBSF$2KQs_TAvsjb{J8LxnhIgJ=UYbOqQ|02kQaAI;6y
z+PYj}OkWTF^}u;Jif)6THgu#vxOb&D?9O4D>3ScC_aQH}S9U$Zb^biTa~)0UwnW$t
zFB>amtYZL)0k~NE@t`E>tixy<BRB#t90AV7>{~I9RL<Yr{Ul9_FhV4|hUPTMwH*1P
z{X>(zditl1SyS{Pw?!|rY80L0mV;5QPcD~fQFH@DG@!LJdyzRhP@=YaQ0QSU;zn4|
zh;OTANN<Kkto$Ea7Vf^a<_i~QlO(z)V#BE1RUCch)ti<%Jb#8I$(SI(06}bPGb;_i
zIX81&C^q83LrbW%#C)V*M!Zl)aA(`>JcG-KP*#2{>zZ``Hd}fUllk&x*lCmV<rpdW
z3XIfzHAXt{)uEKi>{NaAI5&cCBhJm>+l;dT><uuPoe)KLf_*2>H$ZR$oqLK2ZGs-0
z`#{i#^B@QYaUKT2FwXZua3AL}5RBpc2n3IC))0IR!M!g#Pq2A%$zZQ0CkNg*>G1Bo
zBhQe-Gzov{*sNa0V|MeOZzIh@>u<hV*pu*FJAqRduB3kwK62|B4EjZ2a{h!l`v;g{
z84wx31w@J-g5V*DVpoo$O%QBC+^XE_e%qjK<1@|X1|%Z2E|}2;{)oFl+6~wRuonV)
z0Sop0AnnKHLm(T1S5W&AP>z5Ku?b`*Ts{uUaU?(g%ZKmO_Rad1aRD0Q1r2_LHH4Rj
z@Mb$Q%R@CpIL;~!?gF&LOIm^pkR+`o!nFh!oGz|9JpaMX;?n1OyIrzVTzz0W81`Hn
z!8rBJE!xNw+lfptkjl6V4f7d9D#u6>t>BVEiZ010;bNOPHm2(!o85}$V6ae+$HFFX
zYXT1>ZvnR!oQ<F`0+;HWgF$vD<RSfXVp=&$^*bV_y_KwT+4T%g;y<n}6aMS%cIhdv
z`%_4Ve6-#Xo}sUv8zC#Q+AZpyN_+>b>nn+{hAJk;k3hu8pAUJ&IO2h2pQeRQlAU7|
zgB*gmBTdnv4xt<k?ELiCTJMfJCZnBA&bC3uh$JI1iXx*hiYB8uy8R%_l<?8!V6%s_
z4ZtogU0~mZatCvsEAB?gs5#FU_u||S;(nZmz<vmCKLX+rTy6q;6D}VI@i>wXAK#c_
zea+vQS&Fr?wVY$Tx2LJqw_%FMGc1Iuki5kU-*}N(2m`nq(1se*)tvDOzj%TfqBy2F
zhc(^9FOAbaj99<tFU`KHU`3w!17qo`C8a0K1v~T4?~nOQ_8-q5A0f>&iLK-DCZnPi
zDpIH@Kt&-cPDDlK`#Ag9umv8^78PmFiBZF$u0aCfB++CmzO@Y{5Zc<=Pvn`ubpG-D
EA7$+KkpKVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_bwd_data.cpp.FA67D9F6149A96BE.idx b/.cache/clangd/index/profile_grouped_conv_bwd_data.cpp.FA67D9F6149A96BE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7768d9861d76029eba15d1c2abc80dff7498a880
GIT binary patch
literal 4052
zcmZvf30M=?7RTp8fB*ww4V1=U2%uy!2H6ZGL|K$w5U30wqF5fTY=VfGR*)h^pt2TI
z3L-)bt5OP8-osX*3LglqwMvyLwA#8LRIRAB9panJ?dzNG<7R$y?mctvo&P!a66WXY
zJ4{0muLQrnJGZ9CQxF6(#{bgNQ<H`@5rnOWAbaJl5z!%RSF>lI{_@S`w=~Kh*~}vE
zBX{UKqsAmr%#y@`!ogZ*HuZYk<mjcmyp9{)C&KPL-#>WwfbNt$<<ZdizNlosQxe7H
zDna(q+26*F^z?X&&-NK5PTd}(Kb|)b*X|Hoh8RBk{Jn;+_&>H5-hVkr`Qg(zdCu7j
zKTXV7Y@T}>&kJ61ZHX7MIwm&c!t}EzDX$(s4VK>m^Tt!#oG{0>GdEmC%3NN;^3kaK
zi8XoY<0HAcAI}Bd&mC^1uMY~I;LQ~A9jD44>iq%cO9HU6vJ>-$R@hu)XU%sUrlF<m
z=y!K6{N{VM;8?~Xhu(_bZq`TMwc|5qPG+?nQJTz}=suNk#LkYoZzbB$ZdCTTkEX2~
z@F~u`p*3~WB=S>9&irs$qJBL;@AEH8mswpsb1wIiK|C$Y!eclv+*y$B;69;rNDsg+
zQ8Vf*Ds0))t4E(KI^A!4U}{P!`a{FIA<4;h^@(7gkob$M(+f@tC!A=VzXY93DVUMB
zw2U?eJ+2(NyeCF9^Yf7zjX1|^jP3HDkJyH9GdOT<dj5<pTfd<{VzFIB?5Xwbp9xuY
zHs_pf&^^r0uid`5=i-_NV}n`B>>pSO)0~wTTsKXge3iU@Wl5l*Gm?Js1ye&{SJSMs
zRwromN`ju-)*rQ7Ej!sCxar@j+m+0(*yk3L>h)0f$oUe(tS26=?`!FuoiTaQuq|`f
zudjw@Kenwp9{ok%9j&NswV7KLhcj12?z?zTgAtUoEVgs_Ip^5!jovrA=&fCpL@and
zEikS8=(mFvsoI=-{}DYj+c$LVP^E;DzJmW;|8C0R1kTZ&pM<9yp64C7_LM_+N>coq
zpvxKGE8&YbGfYE++XTrcZJr*poJ)GOcUND{!O!N7-3<#G>qs7|8Msg;9uQCMY^k@t
zyl%y$anGUpvKIaF$f2?L{LYjJ7Z1h*gO29V-m>t{NlRJ0^~BNkm-eMEMma8r-pOnJ
zI&*`Gjq~BJ6Avv-DA&B&&GVuqOh<erTf{)0o&KSx#iccj_lNNCwcI0*ro&6>QQxMr
zTVErJ=ZS31QQ!WNRg#l0cAE?i>d!@v<XE`2^_TtfdT2oB<?%mH54}`2ebjSq{x_Pu
zXL32Cifemw_lDGaZbKP2pKSlJX_a2Cq1Cr*@3v+989VlR4cB(RvMM^Rv<;;{cok&(
z;p^rz{thjS-MPj4drFFDQ-m5yZ$;d*(Z`qi*IR<X0YfX7XPq6%Nfr@I<D}~k>b{J>
zYWzAzNPiG}I?}uUsCjma-GddO+ds@au511>NqBv-G@y)=mbG)U2q4JPEmhhc>#B88
zWD9~|LW+>3YVfRAU5;%XnKyr@n9z`rFa)!pSeU39!tBx!l4n*Q#nnPW@zJKJsimqR
z%&WFvDZOJXyGlde2}dv*g=Vg52(w@J$B%;16gL_Yfjg&D=rmPBm@}4_eNYt=B-KLJ
z<IW9G1G=gq%q2gi|J*IuuRxK_xOow}h@ol-bF*XV6)`t_iiSku&dn+2G*v^Guh{FG
z)oD}i86ffa=w5*Ig17(sgDliN{JLi?nXj@-r_+!Pc=SdTBXd>r#)u%PTi;C^)I|`3
z<pnj{DufdnC>F{HwNWo97RCrulR=`DYm$=D4=5Oy%GD!cH;x$pSMmw8LlecS0lOMF
zkfY^*Er+=du<J;+XfA$#|EZWFfMQ$**Mso4IN!_9ruC;?xW6Xg+62U;KP7Ng!rTsA
z+es$+Y3Tj!^P~8ZEb9dU#A<QxZ!`*9BRzoo&tzvVB}SoGHai<48C#|%LP1)3D&-K0
zL9rY`4isJy5EPN&pja&s)WX~f1g#_!UJUjArZ*H5@kJ}isP-`SsN+DfELIkuwB@jJ
zAj$)bJOK4A0IUK4*^2<H2<A$_sD%6109FlbuLZ1Hm>U735$>-5ECp<D1*}$*Ma<dD
z+qL6o^(h#O%2KO^g4s}QAYxP55ZO{~A+n>|L1a(0hiD0P2}Db&OCfTgIzYssav<sg
zhF!pjSVEECDcO7S_Gny!ehz->RVQ*@Q%{xaL=vM|8Y|6_7>i=%fK?9E>|OTcUTCW6
zEma=4R+AqZ#pJ+M4s#uFts~jT?4fbJsph@~K{{uONdMW+*$~N`Woja-4lc-rl0!%d
z*M}fW-3i4CfS>?CTmu9(Fe`vSLGlLa4a?Dte}94J_nTTmWK7h+e3hP`zZ1SnTWz*_
zk2HTi@}=-`Nv!-r-|dH{h5zUuw^zP<Q{TBMynm2+5QCUdoI=s93|pC~gG*d0@+Kam
zn7hP%HAx=!9vqUy_F{-6_7aFZ?L8rq+Dp~+u1VhYr#e@bXe>;H(F@qbR2JB-7w`xs
zKCQOi-|8|R6W`sHj3UiE&6=cx!0I5dCAl1!mjlS&0IV8FdtPaIH|_j4pWv(UW?6~i
z`@v%$#>&wmE8Gxh9{4ul??ZphyoV)!y5Wzb>Cuab)h5a#uE~Xcf^dRjE>sswVl0aJ
zP<_-?BkWRS<O&xmb3Sm+C%-g`6$0l%n2UjPG0F8i0x15W`FHT;<9c#e5wE6TT|m+W
zq{Q1ACuTM~THdzT#)Zqv6cUf|O3KTINXC;vln;3M0OCTxD}=ci@QO($c0gR{3)>$l
z^gWzC&g2X804EPXTnRXpFgF5DBgvgH-#Z>JqM6{kowcsD0Wm!kpPp_PFd*EXsgy1@
z8XxcbhrWA!EB3ZY?|!dC!9+eHKcYVcbN6yzLy`yALqL+4D~3qIl|bal^@K>um8xlk
zon3HHvYv*A%3Z?^fJ61|0{(=1LL?{!%VcJ%rAWcDnb{D@m@<g+0W+VJEd|#Evk>ND
zz$}Kj8ZfJ2mIJ07<~qQvgSiPXn_yM~rV{3Mz-%X($Vi~^KWXw!fNpBF$L48VWz~RR
z4OS47UZ|^lIm~r{Uk7s&;5Wgn1bijT?SS7-@?@Xib%S3w9ag0aJ0ITa3IV$i=3>Au
zCi(XBU9B&r5mxxXCcthYPdF&11Z*YD?SS1*GI2V?b9&)uMie6yI|%F#0;v8p0Q&}#
ze=i@@$+Ep|rV=K6w877FJMg0*{@-rumNfjpi6Aix&6_KlxIvI8@-~*djZ@zOETuYl
zF44<#QhQ@uY)n)%vQceAw(LyUk%VJlVz4GfOIwGktB)EO8_^8?O?<tVn&|nPTAG=^
zWhHVKOFX5kRxhXXJ#5`Li~rwK3&tWQ%Z6>Y#DU9OCUjik?7Gt0<!ynJ58c9kZNMgt
LHS4_mKoIg5hT%8;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_bwd_data_impl.hpp.59E9CBB730B20A37.idx b/.cache/clangd/index/profile_grouped_conv_bwd_data_impl.hpp.59E9CBB730B20A37.idx
new file mode 100755
index 0000000000000000000000000000000000000000..399286bf2d2515dbe947f47db49ba8e2e6028fff
GIT binary patch
literal 5452
zcmds*Yg7~07Jz4m2n2^DFd+$wGztyigUG9(BtaA`C}<Lat0)y(u~kGXC{@~}h+d0=
zR}B}vwjg3lLa=HozOeYX+6u@^Q0oJ0ivkv<TB?>L>hezPJ(<(xFaJ6VSo=Hs?DN>O
z&pw;c#7><W<wg)w7R0W~Oka`4CJ4fX{>@&IwJ3%~5H%cvSXFv5v2%0hc>WW;DQ#}f
z$hjeXKaG8Q@xk<p!$ZeCRyTBb3)`I+l?FRMeOOWwc64yEpTX|Dwza+4V?>iCXXVi{
z;b(IM9ge2xfET%5Pd=$v1+ENjY&%-R3gwcQwyoVaFtqd9FlwOMYfXN?%fay5HGN&n
z>~}2{A9`zfY1Ogy^X%@WM0BpJRYr84+WJL+XvcvwO|{+Xnh?L<Der&1JtibIJ$kZU
zR@jkzCCc}^Ymt{mY|z-H<cAvPK3<XZaR#e*OHjx(x6Ssm-G0!UPNnbed$!t6b@lN>
z|3mwH<fYxb-|nY5+|8Yxabx`96w7#%%i-)k%QKhtQnO;2bah+F^F4R#y<doY{0~W$
zccOQNKwnQ?+4Lbw?w3a;8xoT23@mDFgWK{P=iPzlE=wCG9lz3QpH!X1+P-mLGSSs!
zs;eq}%G<<`Gwx1yP?{$T4@g-_+H<EWhAzu+$k+9hWJE>>@3@y-u_8bpwoa?icWe~b
zc=}#UHD>lD+yAQbm^G)dBj|MBn2$#CvVTwAdxaM~P-XbaVScK2z~I{|_hX8}E5#Ad
zPo^f8N&j6WyV|_9y4u|1z0~nwp3HAmL{8Z)%jn_8=YcHNU`JYi(w@EV`dr>VLwKe=
z<H6cboo`rL8jj@N+I_$^C2vvA;PQ~pN6%D`iT?1J9;KS_kgALkuJgAKxw})mYzwh!
zLYd35aoUM1`ph-g>rLAl2J%K7cvwKT{2)4*ZVo@7Zp^Fydh_|*R7-(C_v66g-s?Ya
z$<Uknc{`+g>a$m8zOSU?O8mj^{f1vygDjT&5E|)+5;=)n2|CQI&z$q-olddSf=NPb
zPf(H}5}%iU2o4JfJ2sJ(zA*cQH9Y5^j%*F`;{i!kfL{d!B0IZ)p@#D2_lxXFN-mMV
zjhabHAyHtY^iujVI#sAj&X77(jgcl?GYRSJTTkEbA8mAEkyMl<DgZT;RJ0`8M$R7&
z3thC&;|z;U6+=KV1fgzh$_yT6@I=g?ae8EC|9UTav^Yte&7P#<CGj@m*Lr3KmBh5s
z`-vi>-a@wYlw>qUF=Pxzv1BYpabz4u@npP>L}&AM)ZF^db7wXscb5-GgVQ9H+uXyZ
z6z&R4r;sWzol>kE#@H$Slw%mGgoH{+Vps>Wbdbn!6^N?<v#thlHO@v58bQqTH$s3B
zcc&2&jqneo7hnC>@tOWupdFi11*mKR!KTy!YK$~8&3I(u?ijgHxZs=lbQCS*FBGBX
zVV8bbPTZ5HzY%U3{?5ZszI=P@u2)Wq=O=q6)q{60t!Re{?Vv#09~E)9_K&cOPOsre
z@9x@qr{MPzUYSIyA01pCmi(Mn1iA-KM^pF<o0m^LQ4<Ml`uKWOzzC#}&HU<~aQT>T
zUKwbj=fH&w`?LW*9h_V_AN8k$ldHf;$yH*c1)&xXAT2~_A(FAxLV^|#&|>f|1}USj
zgy2eehhZHA>u}aXupVav1RHQR0oMfBdd$Ez|DDI0!Jir50$dBOx4<|H$dEpK3^8Tc
zF$X6$NmYWN5=J2fN$EhK!&wgkJ<bLY7#N;f_s<7Q`WKediKh@;3PDt#@a39u8%MQ}
zR1vroF&rPNT_zk?w9SsBqzY*aGoO#b2P0pFuZ=V-kNcFU1GV%El)1~|kjX4g^vTS0
z{=eucb2vFX)J#$vz-0q7A!-Mwmsua(6)7EItG&kE5YmHsM|j)jTCjc7#$3O=##f@~
zqX)*x!B6|1(6ePKGUL&xB-IImPVh#&C@VW>rb$1E-kyrBaz@Q0rDm%!(y%o)%G|M7
zJh9WWm!5JbNOr<VWVGsBS9dPEy~T;7&I9K>aFK$f>VZ>_a|3W1aBc)nBhD9qa{=cj
z;56af44h`1TY%Gob1QIKac%=n8^h~3#}3YIj(tS?GQ>T^b|Ull$7Lp#wb#?W=mK>Y
zq%!Nj4eHxCcf-tXcn|5<Yj!Fh^*ChHzFwkOG7B}6)C~~afYFEx9nMq~oXyIkyB&u)
z#~FK)s)IRo3>Qwb+wj}g7hx|QN&C`vKUf_}|6tqzR0Gb9KsDlg0jLW&Hv!dzb2CuQ
zIJW@Rf^#cStvI&<)y8m%=yPG@8o?Sm6ms}E9%!Q_gQ+R`JdZ21e>XwMCWu6bYxm@z
z<(}&4&Tqu6QeGhE4J@`~)4SBdXf2b(_VkZdof8e7rQh>x*V*pKD<qW<!}6KqQn6$@
zdzyOV?UzN8ycm?IBb&0o3=60b?|taz*L}~+;y_Xfk_4=ROOzyHR1QJq5QdDb2{NV<
zu!FCIcXeRPMkJ+&8G68i)c`63Q<{8TFnd?e%2OWnCSoKpwpxs&VkNOQ+86nuICS?t
z743p~!SifM>7ZY4;@koCSbH|@nIJ8Mpha1EnwD9FwLBCS<JXfJlxDCMn^sSFT0p|s
zTOi!RT+`pqyB4^&i04Wl6gfw3%TRQs%TZvY<R~%Hf{Pa1(2!O?$tszW%6gH=bU5ol
zrpMU;G6T*=a53WJnSf&g+k8jv=N=Jl{-gReDv4TdU)#oy7ZbEs|6bUi;5uGBKf62r
zjW|AazOjCl*Nv_8E{{Oa5lBF9q^{x0oc-aCdS1h0XFgA@O#jup>y2>C_pTQ!_y1IP
zlRkM%6-#Fzw~*9L5ZnaYfp6R}-C@RM*HyIt3PDx~p(x2VWf{~@?__4b5zc7lM_ql&
z>!0{WY^K%g(}H~EES@!8`O9tpkkRSNUxAS_KxxaEB%L9nioupzt?6oBC0>f?V6G14
zGf`3nk}9UYBdKbTRO4)f6eFZD{VkAWfn>yOa_MK9nI1b{<M3TNap%w8we|!ZuNjIA
zH5!Seu7lt@_#*z%^SdF}$4HuA$!c9;ZBqun5yNXt+5Jn{l6l|TmA_K;-Y-7zXW5vx
za{B*+P<E&jnwzA;*kKq=U{AnkB735Z{wS?IEi;bvevQSL7P5PY@U>0!Vze+y%lO)I
zbmxEl!<%a89iM~9bIeQ1N?(#@{m_%0^+9%%RiG|hWz|mY`67ca6uAi;I8Khvd>2<9
zcPL4?TJ?WDJR}h4OOds~foXMQL}y0iGopwQNk*i<e4@t*2?Q~e5xI<L{qFStM~DQ1
I5HRBZ06m*jTL1t6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_bwd_weight.cpp.34186A38546F70D1.idx b/.cache/clangd/index/profile_grouped_conv_bwd_weight.cpp.34186A38546F70D1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3dd933d953390016fe6a1828c94ecff674b02dcb
GIT binary patch
literal 5272
zcmZwK3tUWT8vyY0)ad3kQ=`jtRMU2unkJ=&B%$l1bWf#pWRS=uT~MpFgiR{tk|aqH
z(k7*{A(6`v(l2Z!twF5I+IHLJYq6j2%=k{{_3bympI85R&UxNB=e+N8-Z#uoA}Q3v
zFt5da=_}*Yq$~`>bg6%-X(`KIt72G;7KWuCsEz1f)$gKn|IRPr9-^RzK4*7Ru60n}
z!@N~mXM$OQp<L<qd+chLUweisEPHl@9~wvp(j~#=a*J4MC%yHoCE0vY*5ZuFL5Fg-
zce}msnUAj?e|+-m_PD%k>uDEdQGceNJuP*c?)ngdcC6djV=LeU>!$y3{_IFf{?+2P
zx{*Ksc)G7V*gC-rJMbvrXy(`J{yk~K$ILr-hBkf7OBNGr^88}X`d3vMmi+d9YVAOn
zu3_Z9qMjK=Cr^|arXC*g+!-M#oAO|Bb#wQmlCQHrez3|vxckt3Qz9{?#d*84t!V+*
zdE7O5?wg_p5XFnj#1_vxMrLuRZvT|vQxlRn&-;m$&hX5eCK2XZ%W72r=op+kPxky!
zqc;9w&)Wxm?7jRMneVSx#@BYS+Y^g4$2a~G-TCDD1gUyh{4Iwyg;qD`*b{eaZp;bv
z-;S+4dPLV!Wiazjns((y`G#lhoHeh;IIEGuStIWG_>PbJ#|+5>tJW=h?scDU)c@6a
z$A*ck>;tx}OEYMew!5A?5&JY+_HcDx^Y)0W!fNdia_6EuS&i2lSy5Sfb0#?kxBlqA
zz98>ZtY~-rKGuZ1@8oR$1$D#7rdP>hi*;+3>hx!kCQmP?NlnZLq)pcwZa-MRcdTA0
zr*L@t!3w{GiT;=I7_q09!G<N@d;YW_CgAPr-sEpyYHYm}v?F@8-HuHLS6h-)-k)AE
zWvz6eyP?W#eCz95^Zz)1UB7fn#>s7y=j$cjUh0PDud!e8qfe&E?Wk#4ClWV@y}DS%
z9sACBA^)n`+YMLuZ5n=6E{V>0kaG5Z_qmOUNB2G=3Vh7S!s9oOSJtmS`ghWBQPU*h
zS2xqSbFREQu*Sd&H`y6%9W`>z^6hTkaQ)xC4_#WY<twL@j%{(j)NXfJWYWpzKIcYe
zbKh6sQ&Y7!l`eCgJw7<O@qqEMrQ_!vx%~9fC2TWMGA-BWgGlOd<HOOxf{dqr|4z{F
zZ(p8eI7r}{rny`06dT3al6$0PuU-8@#|WoJ*Uwb<-Ml&N<jR;=F&&ok12Z$zdIv%e
zudfeTwDfXy?y_$?u=<g*a5Z})Xw~y@yUOhq1O?9<4*EWJT=U+G^Hk5STW>NnYsYe@
z2S2vmczgN9l52HIzK<>pzs~M2H7YFa)3|=7Ze-y%r2#XB-}$+db<xS+lzOWN-A#K`
zBUSxqZWDX>g__lqwxJtsr9XfA-T2AL$bpDUe)>ds#@mkK<|7@Ah1}E*qdJxGrj^%Y
zmws4cV)vKq@tA{pZaEh`roE|<F5i}_y8B(|Gx<*LeN$Vz=N6A85^JlZtR{h*Ks<9+
zV!O^@l~b=w6ld8sD`D8-)5XKrvDA!}3yT56oKFRB(W{6Sjln`OjMT&Rc*=&V3d)bi
z@~*-sy5HXCV9|3hQir9ZuWabBUvtT?L3pwY2a5>9NJEw(N7>Nf(6y_a1;63H<Jb}^
zxeLq1LfO#a==Rh5-pNny;b4)}*+wiQ17$;pr5kUw9X+V#gk$m4*=D$zv9h7VYbWl`
zE)EWKAA^NcVQpO7K-tjY4Q=ONjC=S&orA?t$q5#LD;qj|+^Mi@x@gW*4i-g)4Oj*o
zWkZK+4rK^;N|G`-STGglv3R=5h7O;8X!HJt%h+ey*kY>hmw>&5@&Ek~Q&d&z^MXI)
zuVSk-0}eKix^ucLU43QqIf!9@<*iJ)(px*k!RAt74vV9&Z0InS62Bx>n~h=W&R6Si
z7j10=oDAoOo6~nxd*ZhJiA9e6syIm+lcQhb<b31#h~kXnMx~uVOj%ZAd=Xfr6Wd9L
zzPq{+-Ah*=!B4SpvIWE~FqQ5CCtE?>in<-d?Wkp7BLjQ-c*Em6Hf5fO&7*#)lev>8
zo!{&`FF*6zXRF5GWGC=D!HPZ}C%b{)jk*W;J*fME-;Y`jdU7D>^F^=c>xB8bZ=h~L
z2{@L3GabOmGH@(IT?vkrj3=#$>io1ap`1D~2dr~I#GIH9*7>M+f%PuN^jy)?`eKgY
zq{jpg<hEr4FB>eGq`AP$MO^^A0@TI8D@I)kyi(NVz$-_640y*-*8r~ubsg~PP@e(b
z8Pw;1cMf$M@Y+yc0h22rWPblG;N3#wUBK%?eGho|Q1=0^5A^_;3^4gzO@HrLt|tAR
z!y?Vu=A*-qMOv^e5LvP<5m~XV5DC}<MAmF;L^f<2M7C^OM0RXDMD}cZL=J2RL=)K)
z5jnCQ5jpud`O<@%MGgXc5OnC9<?46h+l>e6=23GXL>MwUCg>{4(o$AWx^m&<R+zjM
zB=p%hSp<`dP&dKkCdS^o|9laeGVUkkROPpzQ==95t*G09-;P=adNO9}_~^aStr(~J
zwQ}n4OVEc^2K+MAmB6oLOxJrgEoGgjCj?IBf*=<hnF}ceK`H7Q5Y(VP2ZD2`yFk#z
zn9if3^~L)Fr%J5-O9-T^?_Bo>ygO?zoTVz-=Mp>`XlUx1zR!P1Im7umU_mFu$$a4F
zquvGlU5x2B64kNdb;Mj>9|-!ug-$~rxYGLg&A9FV=r}I@K+=NN^G~VgAEh5<OdlUt
zUEN$0HZ?(mDi!V?-t^Zv={dvGlaZIe%a)P1z#EZ|zz2~;AVK6S@I@5H4;#H+dYr0i
zeevqV$>|!?c}!OU8Udq<5A<l=RDX1#7Ea2*Nyc2X!frA^ey<E9G6-PO$RI=pk*H&2
zuz;}~oaAUeIk?Hud~%S;(R^|Uk)!$K5F=;u#c#D<W3Doo$7NAvoXCOkbVn@eArtwi
zcLA{rbqNq9sLOyTLtP0(CF*J*s!`VhQH#0(hz8UxK(wH41)>#oI}q)tuK{rlbte#=
zsJnsaM%@EM59)p(`cXdu;t}H|+%0`?)w`@!6w}1`C0NkYL}9C!U`IRYnbqGHlC#Q`
z)5akO#Pl6iOdE%M)Vsi87h}4hscU_)1f$0H1kWi<lC2=z%ABOAcETdmRUoWF-2}oW
z)Eyw~U`#LYsIjeB-7$0Z3W%;ShvDP^hz1yMDcsY{xpcRTnxdbJ2EF9U{>{gD>YY4-
z3PkEhj*g~Wf7LB-3O>20;AEIa7>Bu<M1w>VdJOJ&Y^+|p+3^&0=c<8U%@j<WtOb58
z>IUF9Fs@iJo8=$+-4p8Bh}mM~Ei(u+2Em=~t*U=$p%d@1KuxjMu_=KGbU?Azu_;4c
z2{x6etHGukbuHM`qHX}22FCQVm@4p!H8FEjvq6x}6lk0*20=0EV<0$&`V0uppuPoy
zTa4-DHPww`ZOwF30D=PMS>a?k2+C2{fuIg`8wlD^-vhxt#<lDBb|q!I7E_OOo%uQm
z{j}<1dz~utI9%$HnroVC(?t@eUgnyEpiSS~(|g^`bh^7+|54@X`2s5cGlUwgH6sdd
zSG=V9o2x(9h)(o_q$vHQ`_H`^EYicr!;k)&MS6O9`ZDqoc{woh7I`D`5&0mJh$M)7
zMZSo_EW>Q*ES(nXw`I);=TH|aHV_l&g-Qm&pFZsFY|WH?ZVLxB6<Yvd3yfw{YypI=
zsM|r<j#>t$GG;U@miWR>G~NxuZqz*>>_OcR!hY0pFqJd;9<}=QX}?;KqnuWb)yxuJ
zF|8bHQ8$2N1LMadt7<>EM-a+Qhqx2nv~+{G8+8wedr<d-xF59~Y~;))=h?IPrS*|H
zpQtAstsgzQ7oj(MR4r2M^yqqslT|Reidn$nWCu*{VEnH=GW870$9l>UhOTVr-63cx
zcZGC5YFD^Is;F0~DT`96T`z{kDw@xq`Sc$cwqW#6Aot%0iEeIrC{>GFAYB+cA6qyY
z!<4B$GmiOo(aOaumQf5y2vTK@QB&8@9E)q~>TuZpT#2_0SId7KPfy>7aGf@N#!Pn)
z&xr;T+$>#e&HmrFhQ=nQUzuB2SqnvW_Kr^C$y28CoyR*&@-Z+J%$mJO#W%#u4+63O
E0Ne>@)c^nh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_bwd_weight_impl.hpp.A8CF2351C95F34F5.idx b/.cache/clangd/index/profile_grouped_conv_bwd_weight_impl.hpp.A8CF2351C95F34F5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6856e1d7921f2646ee7d4ce26f69b61054a7df68
GIT binary patch
literal 6012
zcmeI0dt4ON7r^Jh0PZp_EQ7p6Qa}h0h=&Nmvf%C-2nq<2p!mq}0g}mTDp;njsF{d7
z7AX}`sSqj5CuoXKsMr<sr-)cCilQLkVhJhWZ>VPO^`rmzzw?3p-0wZ-o;!2qo^#Kg
zsi6}mE;A<x$w#3pQxccWB?*EMGXL^rsf(8C5QHJ0AXe`DIWHl(={?ho{^-^9Tg|R~
zAA0g*uVV8VOZ4ZtEG`c#s&RZe>e+<syBq3W1jcQqD>^PsRIWQz+!N9&N*mMIsQM*s
z>SI~d9_MDqzy8km&%I4Oskw0eU2$B5;Tg|8I~096&uU{UwNYnowTp?Is%wFdk7`3T
zlfQnLx$3EC*}RsttYqUf|C(clzwh3C@b-qUKbw_QNp6w16wNQYv%kk<9bc5ENluu!
zJT<EJh}rK{QR!Ci#t@s)b$ZkFzL7<ItJm=B->2-Zt~CAOfg&+#HPw?ozo2&Z3Z1bD
z>0Yg=RrQv_N9INg8dXakXLdV&L@zQJTmNZ*L3H}N=YP5UJ+G?i>Vg|vyWhVtah1;E
zrqIZ(-D}o*|5|ufu3R4KHLLa~H>)L=+Wh$CMNJ~Z2_a>pvu+;B_M2(4Z$^2plS8av
zM031V$BpnmmkVQNdXF|fJ*44;&wa6d_nw`{VuqIs%}T;0lcw8l>{x&4sJf|J&-%EB
zaK(9#{-7_&d(RzH>d$^&GfJ(qK-oAgKG;2O$9lbCS%D>+H-7!JIb`*9<q^GuZR>;j
zigs4ZBB_(R^<IR0|5K$$<fZr{3-5ekd3TGl(XQdzF26rr&LwLM`PXfl9OLGe?drNc
zg_k+QJJ37Xc<mL#>q4=sW?N^4=#I($W2JWKqZiLsZ;stl+5g#&1ur`ODCktAtpB!p
zv|puZu~2tGO2DD9CBEXG=69N_3M0O~kkk=+_o;s0wLc{%LN*mC!<*+??mS-AGRmnh
z;|@4gzT<PC;*a77mo|K`;xpNaIR|WWO0^ez)5J%eEK^*)njiR`lNP%^i=V${{i%08
zIy`&R-={wz<f~FX3}RNG#k|LlN;En>IyzQ*3^ErzoIl)z;Qb3_>B_n#8^`BUgrz<~
zJLow${qsZcbb!z$friw?`SPD=MtSy`V6~T%5m59kaJ&WYu_JXo3`+`YKm1Ccq9yhc
zC$^iSrJmAph-3~j7evwHq5}|3^PGlh#*U8_vod}4nf6D{<$N8A4zmsOV!J6i+%DXk
zrJ)O~#w{wfuF@fCHF&6jQzPkqu<i#NR?)b^CdDf+<TA-v6NHUoyBTXn;kJk(1QD1l
z(kPOXnV<VYdH9bV{U#(W87>*Y4q?byG6s><TFT8y(o$dPcqEfq%P^U2jLZ$m6kw$Q
zYn;a@;3av00s@g-1MV8|z<H7e!V$N^2dxl?xE(Co!3uE~Sajj6g%~ZwB6%;E^@0V?
z9=*7Cdtq8He89>r8V(KV`|h}>E=fnbM|10gq+>i{JXs}@j`fJ;$m;6Fr9@jm;#+?*
zAxP3M`*XNY*NSgOZ!35u;%vo7D!P5^_;287`<9AozW1z{S8O?a=j>hR8{f|e{4D*#
z*w^G9v9HZ+smsAv^nJcm6K>a6EoB4+;9me-q>i`I{+g^mS3CGB1rlS{5VQORQjBB*
z8Ab{aC;(#>n5pmurvkAGf{`s1#HjEEss?*CI3T$OY&0+uaVvPYf)C<0ux$f-#GPQ<
ziL(~GwYYvC2>S5*>jy#q%RHtZ56}S+4B+wsa2)`5c0PX>LSl8?fX}3#RxoV^F0D}X
zLoj`aa|f7q;H(8xE#ir%Hgqn2G{2C^iq#;j26J`>imm}+4dTh;vX_{-ewC$5(XJxb
z5TxcNa>K}7<j&F5<tH5TqCB&i7aU+2z$KGtp6ut8MB_|m$_(ITfDy7)1;Q#cWqfD-
zj)x6(J<Rw7--IcwTEZ3IRd%j7&6sZIVCP9}_rlKd3Es~i{>7M05+zM$yD54fSnUI6
zmKUYUKaJ@9E{J)KXfk>z+fC6iWDG{JWGqK1$|Q^NO*Z#|q_4opE5K#+mFJsV(#U!h
zpQ1B?mk9!7Z5!~m;hY7$ES$FkZ#&M}z{|!t2Y5L+=K?Pm=RDx$;j9Fn66bv2<s)9p
zJO1O0+R!JA%X}<-xX@fz7M?O~SA8|(ixP+^fm!JO4?xrbn1i?!qDm33pPC=|<eq_?
z@%19nqG+VH7p(T8{d{xN-j5H6izYMEz8v9OCh00Vw!QE*q-Ym6m$7JkCNQ4lwiSxr
z2FxGESwLsuydCK6IA;T$jdKprIXLG6or`lG(0Mp3fmY(24|G1_?OV3)jZDwgedT%b
z%r}|^e{T9YvqCb!I0J3Gc|CIy))`q}V@6zru@_MS+CAZRfK61`(6`&wYqqPxO#w#Q
zq-S!*YbtP7p|rRstF*<S!tx#?kbB9w7~k{AIr6-D&pD?1Gz>qDvRl#OFmh5<#+`r4
zxd1g@@C?9|0f=N*-rlZZPAzSA&kQJf`l#vLHbBvFc5&~rN)&w)#5X~TxD~=$kr9e+
zgNQbWLGn%r>%`?+2+;y}+xs_7FSxh-v^5h&5w;OrsY%h1c9C2~${GXyKh-o_ga7Ds
zf7%Rf)EIq|agmn_xELDbNh-W_Q-hfr-wFF6upfe0yCgjTHUs!hHvs+vAY$c*X5R4J
z`<0Oi#nj|H36~R@nw%%aNXC<4qyU}*Fjj$31;bdA6s-YE4RAH@;2!4QimJguKaOZa
zX_lfpVMHg+T5#9m`n@3R#qIY2uMf?Cv_0*p`Ip@X-k`{-5Inq`@ktWFtf-fdD3yXC
zb(f~MOnw`;uZi1JHP_<G{GZ%=14mbFoKu|mz`psdKVRd|3nq2NWxt=g#UzMiQ8HHm
zQS^SW+7H|r_!mW(!Ib}*tYj8zHMmy;cL4q^wIu5K-4yv-=f1|dwZ_th=SGjlzje-Q
zoTIi&-?T0;+GsHAdrG)t50mwSq}cu^^OpIdtgZm50=Vj*V!ohA)i`Tlss<go2DAR`
zR+xvzx5LPGbcCYlE*RN`vleD*Q4LBlXWWEdn2W{_K<og-vwS(gVfEAqYvmgxqpOv7
zpWa`iPcZSeP_!_LwM5ZHU{!>wkn1)l^%`A9*1obiH2Yqb^c=K+H<+?L=ZB}eTrL+f
z2f08pkdLMmlVXg5$RLb@$zYDUcUGKnFCS(926Oh#C+`vdYkp&@RTbE(kc*!k%RlSp
z|Jz~a!PY|XTA0n=PHN)fxq~11<f-%I4TBAf!JmIGGqynx6l(tWO>nTo6!kKnwb{#n
zZ4ksz)HFs-L)0`vO*-{GTbX!H!TUlB(_v=j2D~ABYN)Z$#7JOBm>{(wsL6aTXD<QU
N`1gzYe>4pB{vTsxH(CGy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_fwd.cpp.8D8BCE0560F621C4.idx b/.cache/clangd/index/profile_grouped_conv_fwd.cpp.8D8BCE0560F621C4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..706e95ee710a3a7e8892305fd115f9b300a908b8
GIT binary patch
literal 4522
zcmZvfdpuO>8^F)Yl%b<>Hytx&UB=}YVr(wCU&arMD7TJ#A(RT0q;lKMYLf0M@<Wv5
zvSL;?%C7CAen@u9rP|S2t+bW2sV(gp?U{M|>~H>fobUU*&-31%_jz6co_v0VGJ?3S
z^-SLon-)z&5QIb|H7zCKwk(1yQb3UOlDfd*<VVYt#(Jj{%!5i!|DK#>r&n0YytruC
zq~Af=(%m!9C_ul)`a#39a_+Ul3&Rzg0Wn%4V)g9P<#*+W`7;etR3d`nV_!^-IW28t
zjXW2<fAizhqDaA8gXLGot)|vBU5K`{u^j=vh1;^eReSW;HQGqzLbHzW=oe34e^!>c
zllkK1jhXX|td~z-<oPA{F7aSQKAU;NE?Qz#FPD2HrX;XFC{*3eG-p>2W)*yD0e(y8
zagk@Rm1%yBe`fWiK+Z|w(67gZ{mmEOX1%}0-FGTOq`70M<pcF0!^`FA)-lt1f=<=p
zq*&dc?%&$9b*sDc^fI1hW;o)P^N$E)RU#@p-j^wiCmj$p|8yY7V&2a`-&Q*)JkaPj
z_}|j-&Z#?R%*u?m3oVD7&-!0%^uOslW>Pp}u~jM1H@d#^ykYQ{>wG*8)n3JobOtUS
z<fkh~YuZ*N9hm1b)3UcAZDO7SKO7IqdrT}UWhALrxLUS6IAYY*w%j6o&rHva{??kM
z@()BN8@#`_{qAhI*1D>Zbpl~QqR78SaVC7bNALGZ2a@FL@(eQvt8KTOZ&?V`%U$q8
zD~)qB_&;p&*nh8b?_U0vR`=`I15@&aQFl+jD>AKftqu;+oyxqhFk2y;Gl4YZO!Y?f
z+m!Tg`k_Jnn^zi}e_K_Vyor0H?O|ks>zEyu>sx4K>$N88>(P;I^6&jyh~*cOlM_rY
z`P{A>iFlG@sOB?VvEtN_wpU9~GbpTf&|DT7n_2oltK}8vv~q)1XQ%4NO|$OUdwEWb
zdfDWv4Mf&8skDzxAFNsc?oS?jTlF%;t-;&ugt33KMyJWclNn>)gZO8)X)1>OuQFF}
z@1k2(6btsGSL<v!#QgL3x3>qZZE_QzTKRlgQXVZjdb_1CUQWQO<OKa9w<z(q)r<}G
z??Zo{U&g<AWAj%@Yi^$3J+l09^oq)q_?no@<{8mTP9AJ{$w_}xW+dEqSH`S8$y_fz
zQ6=Km%Tw#Iqvh|8Z)>WK+~LulyZ+pUpx^mkM>{;<z3QC5EvB0POXv0rp4T3_YRn$E
zJbn)ut;!!Y58O#u+lLo?8zUNOZK0j=eLA~7E$#Yuj^DLi-i1A{J?Z%2is}=sqaBH)
zldSMMtJAL(z3WTk20I?@l6_|qR)5>MUBTx5>+8F@Lm$$!#_RoSvwPc(KaYDd>Z5dX
z_H0vt*B4sxIRyibZr55{Cug^0{A#vGb0)UWGtJTT$Hgd7S*+FC%sAre<~6WmMaUO6
z`=1%t&)aeAe2dc3$TuL#j=!kkaHm7Wo`dht7fx<0?A(`+q;B636ABPy!LX%Bz(xHP
zkr)Iaa%GTQ8LfZ*AmS`Q{w(Snx7~kDQ(+)}0SKW<(^Q!gAC(BQtkEw|tvnQ;hxn5o
zR8dvtoS@XsNqt45%DaDJAi=8;f<a@b&k0KHDLnp%FS^;9fdrDyHE0@)IYFtnE<Lp8
zn4gc+JY*&5ToF~&m=l!x<bK<Y@76w!MUl0nnn&ZY<^-iaYf;o^kFR>fK!Ql;>NIu6
zoS@Xr$M<^4l*MN-kPuSMq%l?I1f{<1ZX$RWyyOOotRbB*Mi;Z@1f_nw#No{`{WJ{*
z5=t&iiKe7JCnz<N5*wGAzCgOeK!)1af4V3x%|DTN#C%v5CDyUl=}6W>B(M@7N@OKU
zNHHF>N~mRB0W^Y3=PFP$R1WN4x9K$62v8!N7j8p&(mY$3op~-yKsIw(Tsx|__E)Z+
zI(1{oWOEEJM$#L95}_2};k1iv4&#McQ%<PUms6NKb?UTH;xNDu1Bo?CoB{Y5nA-r}
z26GR<dtep;Tm<tpz^A2bxh1&g!-n`WayG%dU@K}i)bN<XCqaf1c3eAXE&{+506oeG
zC5iyM2pCEEFkl~sxfrmEVJ-vAGT6Qnuq$Ex8NfaRb1h)k!rTOyO|ZQXu!XR`4Y1o_
z?f~o#n6ConRoK28u)AS>4`BDe+z;6OFb@LeAZ&jhu<yfq5nzj89s}$#m?r^q61IO1
z*w118G+<A|JOkJ>Qnp?EqC07R^b3?m=+Jc}QA#6p>ADc<(e)tGr|Uz+p>rTIpc_DB
zNH>JYh;9UtG2Iv<oQ^|eLN|eE7A%|vN>oVLdY;>zRnkC^;pT1REr}#5dC?T+k`|SX
zP@({QUI6%%HA<8M+fv{y<wjuJC}p=p(~tdAR3DJu19^cG@9t_(mCIFSGsxyuyj7AL
z=l-E<gv}1<C!3>qQIgQ7G9`Ht#c7hd3F6?!i@-AJ*p&dUgjoo1A<W$X?}qt4!0*HS
z9N^ET9O5*@?A_YYMy8yP)q$!QAx$lP_QhV$$Z4+Rt&~hNyripL^i;LrU;QJ~OEyNe
zUwTCL_fz*{QS*(gsS(x&*u=|`ci73!jcP^-dk1?LDLHT)jHTqrafHZ;;{=g2#~C6Q
zj*En1eVp5~zp0KPXRtzfg(l@LP9<;fm0X{JjCh5tTp3+zg<{svl`)}s<Cy;Yj?Fox
zq&>kSBrB$(g0ApMZ6a5|wU=}fhm27H;8BB$L&mfeNDdt(8i8pe%w52=3+B7P^e)U3
zz;r^&c}0hXjLtjaxqUn)st)aW?#{C56=0;oSKU{O8h@{sY~IuTAFRpdRE<=2gU1>A
z;*Rp*6Xwk|TXW6Dl_Z)MEtwr<C8O|3eHD*onk&PnHt3Yv=`Bs^ReMS6wY;^GwU@8=
zrg`~)HA-$U*OBW?)gc-&3mj&F6Xm>W_*skvQ^X;yLwF&Qv8f_X#=W?LOXFSuI09f!
zS))W5;FQ7K1UOAFUj>}2Fb@LGAk341Gbv?iN0>CC5MYHsvU4=zAixg7d=y|uVLkz{
z6EK$ptQ_VlfK|a<1F#yH>i|{<b2Gr2VQvLjE6kSwb_wPy0J{S7b%0%m`4+%#!Q2P1
zK9~mpHURT5z=mOd2(X7RKLXezn8yJ&4)Y%X`vc}FfK9>tC&2!M`7OZSO8J|xdlnS}
zhKdYIXv=FWQqhBw(IY<#6shySaMVd#slUJdUwMam<&p6te&R{vrx@VH(gz1h)B?N~
z<_>^&z}ye;ewfDqJ_hp)z-Oe~a{Aj}BIakOkkPzTXQv|-PObbS>Ge+I_vC1Vox7bU
zHG3Li?`rQMB?sKWOiGTpBScQP6GYCqGej=9i-dY~cjV_dtYnZI$MxWP!yCt+1zwbU
zs$Ev$ljki;SktYUR40`1pnFK@j<<YDv0cos<k~`cp^|#`u-$V|aU$w#vN@I)E4jd;
zndG{^oC3ueD2^8==}o<#$=WJ@JX2W|CAxrZm$Y`G#9d%}7v>3IJ0a!Qhp)+P*AuDD
zojaV2QexLfi<4k#%9>R2p9ch4EfybtVbljAOT04cnH+aBcsM^YDk6M!7!o1TAyLxL
zNa<&|^mEOIwd)f|{3700mNsv`99@0^s;I2QSm>p~cQaH`@KR-}sV~B8m)koyIyt*6
z)!^FcSsQErzgA5ad+}#Fy7~r2xQVH`1@Ch!n<dM%%q-nCG&%qAj+XIQ;pz!|kiP-#
C&Ez)#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_conv_fwd_impl.hpp.90847DB78EF4B739.idx b/.cache/clangd/index/profile_grouped_conv_fwd_impl.hpp.90847DB78EF4B739.idx
new file mode 100755
index 0000000000000000000000000000000000000000..27e3e0c6479716d36fae28d1e53ca57b7e103532
GIT binary patch
literal 5934
zcmd^@c~lcu7r^I|5eyCqFkur)0Rd%E79(IFKtM&1rJy8HS(IHMer&c%RaBtm!&a?5
zQfpPT*5tHWrH4{X!d8spR<L4ish<`<kt*U2(Tdh@;^)k~9{=)h=A7ib-@EtS_vX%h
z_uZMK=y~(pZ7?h{D|)peyKFg*VVDK=D=jNo;ctXt{#*=Oy}LCzwcxU^U^L-DWWbX7
zDz~wKgZ!oG#+hD)4N?0~1?t_lAK5mL<vh}!ARW@J)c$xX#k{U;%RR9ys&02`{JV`2
zM{14Yxb3S?wWl5Ud*$5l{VP>}Wd-_lUfN_a`C-rC=HHS?%fE+eUf{zPN4M&l-A8Hy
zpU3{{QuB84uAR-5DV(7zPM0@oL|+VF?b;QYW|>}j-0h5f=I4SDXTLn(BD4KU|NV`h
zo!B(L$)fffvG8&v%(`zC#2K60Gg$3<J@Trk7hlwtw{cbC+S(rbYgOI#%~<oDC;AZW
zv&s*9pKV+xzS*C`58fFt@s2c<SdZUPm#xy*NP3^`4eV|C^URg6{_GOIu$kK37!dmD
z36}!Sh2x?nUmDq{6zRSv^DlJd`@~!;{>fX>QM|nB_wte0?LQ2hh*;(3D@>Z@=Buu~
zGgon^V!ohbb;)k;V*`^86&{a~AE?^0aE1NJi<}9QV{ewXN%&o|_(bWkqE5w_RoS*m
zw;*53CzZRh4&R!uQlzdt^U2V$quqY%b6rfQ7I;Q|T<vcq`6k-O`?R9U{?K0i?UN1J
z5y=C-7t6EWT3ne@T(q)q&)&*ef1Ovn&rK73@60QBr`Tq1!<Mqe;kK!z75lEosZJ}R
zE)A_2D~o9{>fKx4E7W>Dbt=d18n;w9Tn^rN;D>yt=m1|?W6YV&<}bjp^F(2{L*lZJ
z!*c3+2H(!i{3P;1Xa5SjwM9-NT#w)1p1R8|yCQW(mu5Kd^dH3yrE3%!;Zy|KX5JtE
zy5H!rk&*2L3aN{<GqtnCsDLtDO<y(lrx|Z7Aut<bjC7gc;_~tj;~D|R!GVO5?3~gT
zlG03Xmvqnc-~&N6fkzYg(<3>4lfJG$m+_G?K}ubv9&|TB%G_mMEXlp(zAWVg=7g}6
zE6NRHDNm9&k0#^p_S?gr4W>o}ndXu<lkO(S^eO2=n#^)12d+5e&|!p=D)3VQqlS~C
z;4lh~w4%A%Q89DJ{2rXL#!GXiyD4k@bQhK~co~Q+lb^}RY&tj8HD2BM+#DyR_R_cL
zAryH^y;+hu$e1~CQsyn2!OG+gazrNglOwBDL$VrVtgZ&68j!Qx0=6wMndLUHZ9`cH
z);h3d<vQ@wq2ACzvJSQ12X1{ZmDTTqz&@Cba#SBg(`@_ecN3qfn*BLAnd6hgJWQO-
z^~+_V0VnhP@))w4D9N$TI-E`gO*IIr!Jf7?x%WgdHWZRQE*)tfaPPe@+_rxH+E#LW
zZ(n#9xDHZ+IkV@8X=k~FeXo8Pbl&t09Pc`y?lKjoEqZMVca^8GK4{5vN+7foE}*BF
zVSTQ6o~BU(IJLVSnt%yypM(QL*2P|jUo#|fd~l|~c=-bhD)@O)bJ`FU?mQVna-JL^
zC0Hxb098Vm5+v+cB`i{clw}pTs=%FPH3-!Zz_JE}8kDsl)S}!9!d8^~fY%4edPaaZ
zg7PTvM*qp)qkwE)J@E9XTo2Rr;6tw`*b9rC+-zb>r57~_)WD<-g4BROgR&L`T9jKs
z(8_Yy(`0pZF>gEN^de!=9NPZ8mg>8O!#VX-GOP!Sdax;K>2-eU>>fan`@mu!%W;9*
z|7AV>qir05oF<+&pB_(;9%2uKJjI?2C9OI(bw{GVVah<ez*u@n(!~RuA?+IZxZI};
zxpP|;+2&g)f7b%H7WlMR2y!P_>}36&dLd}dl<B`)Q_gc2xku7!S)cs-gL01zr(fGt
zd3PHkgk!fxDZ}xCcvo7VAlHl6$I-N+r1af{C#rDjA?DyYW^^|}=Hj^s<>7e@DeCfV
zMVB21sR<8((;;9^vDL?~4V2?Q|71##v$?Z*tSu2&gitUy7@-hu2ts167@<&ZC_-V}
zFoeRn;Rs2%5`-eS5e#kQHXmJjF8UGWH9tE)CKYY^Hdc|m=jSf!zOTT%E0D~t_9{eO
zWzz#e_Cr)Z%auu=g+97rQc8I>P7vqG4qGQ)w~(gICLQ~0I!iWCN@qcK7E)*dLH2;G
zhvm&NoT@upUj!LkM$TpqU4j&mB7}m;V1z=*5QM~}7@<%y6rnIO454r`93crQK`4TZ
zU}%TUht`tyg7wq`t_Aa2DpS8I-*!A+mb!`W@C)V3JuqtzNa)nCH{y>F$HWC@<8lMn
z*M#sN<UcK?yrKk8C7a3j4ST+G+~#p7C8z`YIyUp}FI<4fB!2MUm!)Kg3PNcI<D?$q
z_3#eOU*5Cw7#N!J)Px}Oo%5L^iXaQ76ilO)2(l4^8i6^;4d+6%8sb^G24Xajz_J#k
zT3E_*E67_RmFBN%<e%SI)#gA2Z-z?-Q#cZ2=9EmP8b0bVn7s5sqPH<l-SkW)%%Vpd
zI7SIfZZ}lR)++Qk9)-|R2xs;6;HU>Dl!Nu`-F%nwoBx*|@hvA&g+EuyBq6Hs=gJV0
zbL9vr!9odEv`K?QeAH}-YB(r;G$?DqM~iYR__U&|0}CB$uMfC=fXp`n+!2=PdBODM
zTMqqh)O5SN$)@(<H*Zj2sd=}0J}An^s0jM!>yU;w@%%~oK-{?8a4n~PS^RDAqZ@DF
z$c_(}9n8MzdTm_#OJ{lh;I~&={zGl(1VMru9i9YPA+BJu#ObO9Ch-?7S5u)@4?gt}
zNayymC0{3wU00NjOV^z<pY!V&e^@jwZ@zoV`!$<F7xE3M*I&xK5K*aDC<|hvNIqMR
zoINU-qXMSjH>BQFH7sYfHISx(43=BKsRfYYv<;lvP}V`J4jtbI34Op6S_E}qO7*aW
zw$~HlzBVbrq3#V*L$Bt#ez3wAqjpEEAQm}Q*NWFB(snO7YA5u0Ih}hg5nl`HiF|z6
z7T#dOf&QSv6dkAjwSxWjr$_XAyn5=XH@ieU(UhK>Ad~PUgcjlp5n6;VV(8&+b-Pc4
zv+Ely=1C5I0}HM=OD#qTE=tz(Pk;FA#~HzA8>u_K1d>aTPA_L{jwYFNX$OZAoPwFB
zAc+=~WEU<seC#VN$t-O*G%O8&|G)Tj1H&e=O#$1qW}7B#(~NBrY}1@=lAD@#n|0hj
tV#F6%S({AcP9n_AEiCywQ_PfAn#49=eUCJB@Yr6%H}F@N72UwF{{a9}^#}j}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm.cpp.7EE7588374346C13.idx b/.cache/clangd/index/profile_grouped_gemm.cpp.7EE7588374346C13.idx
new file mode 100755
index 0000000000000000000000000000000000000000..65fb0e7c5f4b6fd177b749acfb2afbe3a88eadd0
GIT binary patch
literal 4578
zcmZvedpwlcAIG1=BaDY(Oyf>PGcn^nUC2G+nIhv(sCbAU)<#N~O`=*+bdhvVm#s#M
zuC}%s3Drj0*XCBK?EbnJ-E`f4&rEyfIs4k*{PCUVeZJpw|D5kRVdT`Qy#^RIg(CB_
zQs&RWF$|*-%blN--mi{fK`ac*-+m-AYsSDN{q|jrGlUbllhy@1(M;^UbNEEWlG^ua
zjq5hX<UPCH4Sv4xm!-gVWz&xNZ}v^7Uew~cbJ(VAY+0kn#1LwFg(xuKug`CTY@Yph
zaZ79IHCdnYg{?%#yXCJZex@21&TjB<yQ|(FGQ9jW|3K*{Ezx%^(cTvd-+y?x#j)Ua
z@cTb=+)dT)ef~86O4aGA;f-gla+kSjzFL3p)7w13mIM1`5pj#xYag}xt=H4`K@;ov
zQ14GSBTv3gO5x5DHeB5&o4QT(d(>8b_^umY&mHI#a!$Bi^b2<An%ddke1~kZEOyR+
zkiS0@+n;$LH?>(}Vih|kaOlgyG($;NX`6?p-xsT#z{(#by-zNlb8$<1+RfU2s__2y
zhQ!Hz@>kcYjy@DEzGe4(bZu1iG2d&6e~9Xq?tgaiaM=fe1$D)xvb5z*s!Qa?4P$(^
z+ot0sm}kFELXpU+cvj<JQi$iJ*4;u4U8{rZ<?H*`=v&krdgIlx^x#L~z@*}ZA-t9&
z#U+C~uQyGY9y51hWroMo;>jhQQS}miyOZh#wmHwF6};-r6_R5s*6?o~Ivtn(Qfja{
zBgKVRx_f13=<SWx@9s4ni1Pf-z$Is-MbgpP(XF4I5@J?QUl?QGkQun7Bd3u61cW7a
zT8kbNA?as)XZohKjcG57JZKhQy5CZ7@*ziyJ<i=dwmR!BkH1lKW;nsm`|yI~u3Mh5
z?`+yy9=V@B?c$hoIVyOkFlnvP%3hW!J1}{d&bB<0{?ymkOLiab-CgNENS2IWLd;v2
zwBh5Kr1p!8-bU}O_0M*@6Zd$E&%|(H_6|*hAS0V6E?d(?E7JSVE>5wZA!&2DiyO}K
zuyQZwtSV`%`ptP~eb5*Fs%7Thr(8nnM)!Wa9_3fpdF=aZk0h%*rA_foPePl#%J!tz
zjSg}R-g{z*8-1-OeNBCRfV+o<x$d!p(ne`*x0Uvi5U<l?XT4i1_^>PZT-05cbECRu
zlY@Sdzx&+E+fehy)-ATaUsF)DcmDpYxF>N}ZI+e4zh>RIv-ITFiz8mdIWFl_r|wy0
zvEw->KNSS-TKLBm=a*ke&69%@$KCyL)QjZobEysvaSI%7Xmp%T8rV9hcd`0?-hg8{
zHR0iLja`n5lbp+&lWtj$TK?<YFV*WFRlJxUi20WHpFN~o<)5^%In?pyn;SD)_1=&B
zA%3r+=GZ5>kMh38eg4`xX<oGK#a=l%()81!*Zk$DI<B6m;Fw)AIjUx6x%YK^njmc(
z??d0v$Yr(^$FNxLez8fG*RC7y-n?mwza&C)%dzfF#o4Yc!bMfJ_5<~MKS*CmuWYyg
zMr&1hW4Vj6lB8-Frgh~fs_0gT8-XQZ7{xc>TPh8m36wuh7)xaJ)AdF+78`|8=D0aW
zX`tKqd3Ei~odKn6Y-S8b8RCY7(m=Oa+brCts3?_<&8FRYxE`T2D)-ac9Z%QV53sQr
zbmL~Y8AoZL+sdG)z`DEjPhBj79@;R#hJo4Nf0!a&(LdV^heLM$lRmJq1iAw*&Lx!Q
zn-jyN#zUIE&L_sQu|(Rfi|Z0f1Km*xA2?R-;T>!&o_6cw`h?Oz_sq1nY1O<3`fMzY
zcC&Fdp)}B)ZPVUA?1GOWuuOV*6YvSfN(0?R=W04W3f1#9ut?f%L>L(>4Rmi-o6}Kp
z<f#@L3!~?5iCgL^4Rmj7na7RjeU#6}B51b-ZlR|%(B1H0{l+)m&rWJ!(RAZFgpRq=
zK=<#1Ib=<k_YO8ThaNcx=MYK*-M3Gy-eR^Tx`vH~(~X<pCTyjF?t#p@(6S#kTd=WN
zbZrCNfKVFfe(^x)96PhzfsIM&+NQWETWO#h%SlPiJxt#o>Yu_N+YQJM0Gx7QIp{*6
z!GBHGKZHh5(cI|Ksx_rzxG@ODa$^yiX*yHDM8=!OBa~p8pdu(<z*349?g$huU93<n
zTuUfcHI`DOu$Z6<!V2PGcV>S}maePND{RtAuh6YT>(~2$pk4yQmtYi(hZ+WIP`Q7v
z)99ab(w;5%SVgbSOQ81>7{D4((sU0BJ;PzRTMk{#7%^vvjZ>3}$$Bsvr#uJ`gggmP
zguDnZguDrFgnS4egnS8Kg!~9Ugr*Qv5b`Jd5egsz5DFv$5eg!L5F!Z@AqgQ-ktE*s
z*CzSq0(vRzjO<LHK;YrUS;|8T7lNSDgVQ77{0T}9q;e3;a2JqvF&wTP@Jm*T{yKVN
zMA{-YY&#-^BY3plLqm<A_!@j2SWHmkw8yD}V0xehp~wx)0AwIc22_%W@#!`0SaAMl
zdidXR6x?%|Kj`|3<iXUh1Ywm-{e<n!zubRrxaME=XB|3pN*3lemqk!cW=>YjAVYmZ
z{h0|%{iQ(+h4I7e7z*cyBNV}pKq!(QiBO^_(TS-?8B^vAr5LBED3w#(^H%eQKus}Q
z^I>2OXRF|`!$1TXR;OpL$c->P5OkQsFz|z6@C0eG{*fLDD%3vI4Yon=IX7K(rsxxC
zQZSbBRRWVrfUpdh$WWF86FJIVz@&@e;)+eD*z(_Y&<gpseQ>0!2LsIAI?7DxTOmW7
z;m-7zDb7SFOPqyJArKb=gjWFZ3Y3e0xCrGcAg)4r3lMKXxf+P8Q9c61M^LT<;yRR%
z0`XCln}N6)<;y^P8RZ@z?m@W^i2E2`Bbo2@F8ur}dTkuo4m_y9vVEKHWG{W(M=N&R
z$Z^K7xYB4f&vzMj8C@*W7ja=Ryy<8mD6ToU;RMwM1YN)xt}DFp=rF}~$Arm&K+c5i
z$()8uq^lm#olGO9sd7kte78$Y8Jk0o)LqwIl}zE8qt&IjB;nQ)l+<48345m1H9376
zLj^tsp$t_5pGttR4EV@UmIH4&n8Cz%0q-t`k8k!FrGNA0FaJ8+Cj{D&71#GXrZdnb
zAie}9F)j20aWBJXrdH<H`8@1q(b>_|PL&04Dl{OJgq3hgN=kzn3KNDoG88ThM<_xV
zfl#C{5}`!r#PLi$%7Rkex;T|$k)onA_gvR6Si0{Fz1()%b}C_<7l%<A0X0wQcV0G`
zErl%+)c54~2+bqsA+!>ZD*?i*0l6CG?SR~l@(w`mK>0WzkE7fQ$X1lw0NIA}6+m7=
z`3@lOpnMmQcTw&KWIw}A18Xl22+ND<nWt%{jf9G7w_ft&3a@ZBT|A9D%^1%7#sWM=
zt!P%(2s(%GGE@sqP=$b32oSCUyegEB0NxRlF9F^qlzRZLhhg~Mqu&gQ#~xID-^xF`
z&JSob(1Qxr3MSy7;Pa1;R=oY-NrTTmx>)h%gN;1Zsx4P@KURSg)G9!(0uitog4zbi
zZ7A0RvL59IKsGS^JooG=S-wE=MsgyYM#A&}pOYFa<uekFhCY;I((ReP3Iqi%3{?U_
zB|umP1Tuz)9_OhoOupetA39eZS3TI%@TOMHMK<@j%Cm^xK{+Xza}*nyo0FUyuJ}k+
zm~X!n_<^ZYAGQB-Vqf~sb3|JD+|=Y0Y_7_Or8A#t%x5a|nVdysDuSlZ4N=E6Mrdkj
z>kzsI`fNQ3XR7}=4ohNa%pGZB>OR@i%iG7-&xvR6VKd2Y)c-3rGPkfCJ;r*hz*Z=7
faCUJWKhe!<f|;W@fM>)Hn)V;Hz|biq2*&;g^YeFK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm_fastgelu.cpp.AC58B27586C8FEAD.idx b/.cache/clangd/index/profile_grouped_gemm_fastgelu.cpp.AC58B27586C8FEAD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5b3e8c346c25a29ca8178f3facec52ea389431cb
GIT binary patch
literal 3858
zcmZve30M=?7RP5GNJ0h@5(ujx3SkM*A{4_C2x@>Nh&oYrP=_4@RZv8r5kXWyD@33`
z(N>Bq#kPo|U=<6zqP9q^)(YB|vQ&#mU7l7e&-P7XZ)W=X=KFFozx%)YIrpAH#lgV_
z3<z40B+koB-nNMXL69;2%h{HlKA{Ui8w?;Qukn-hce3vAnKxArWI<ntca3=qZ5k_L
zmz+%8CY!GM&{?uLtm?Mi&(ezuhMP}#W7FIaef0XU#mJUlCFWA|oSj*Xm+KZqCvbj$
zDcfNG;O@(J$HVtUK0o`P3@?5l6wo&ra;il4qp_JQ@D@cbbl%(JXh87B{G5VkbDsO(
z-ZlPY?AAVG+mWI7bfLyw7OEEy6fgMsFMKBkX_dCLzPJr%i?_y{lN>*8R{to!{=2N>
z`{}&lq0JviB5#gV^sX&DcG>0dh*f6Sy4duFunQ~xeAN)F5?{MZrN7l1XC*3NGg@f{
ztf8OErfRz2>aO~cdgGkr!lW|mFQ22|aI0@~s-pFmzxp6Mq-MY5_Wdu;FZD^A9kp&e
zvwNbkE5Yy1z>~p!t8Js7$~(3db;cwQEO7IyI@R-}v`e>R>D@@*%gSUIvx76kNgs0^
zL`8hrbt((a?_-}^5WSIlF?(XHF=p`6hdez)+qSCys_`=B!lw48cB;a*bG#}3o@L>j
zp{u23k9J=j@D7VgT~@Q%e{*R;**6g<B}|uY{foS8xpQbmm^9QODoDhh{WkiO)6d(e
zUw1FHs1RKnNOQ8eY+SHqldQw&EY+g7+~}lPXGfx<&i%xSme>=V=rgSsQhSB$?U3_l
zX*8*CiOe$VS~%0PYrkir>HP(_uy*~Di2KrzG$dCsRGhfgP`>v2dG4{mk8=~vzWc;A
zme%EQ?z_1o!xeEScOJf19yKL8mRmeD_|53Oft1Xl9b<*t7q@-vopziP=X`SG*UWoP
z-1Z9IU8FJcPp_{ePit?)^5Ut#q}=}M>_`c<na<j7vC50nFCU-JPmVoWU738I#+}Ry
z=(ce{*j1{bvw1^{x4&=xa_fIazvuTBrQKWhKD}2sIN<gXZ=w4<Er-^zf;Ru=-iF@2
zi>`Xbw;XKki5XF$2M)Sj?~nd$e6T6_XZOS}|B#IO?i{Zz{BhaQ>0REB(Vdqb6~~o0
z^wZa`_cseL`N~TWeM?+!6ME%J`KvL%xo0<e&!o3+c0T-K)1>OqbpB#a1|_j)o5!sC
zjGgJlzUF(%w3d0tk8ZGke05mGFIZBXci^4J%HE#QF7|GtsPT$}n+L3*)XDb_1_mlW
zr(Jyi$(7SfS`LSXJ@lneDzC)vXmN;X7TjP7uZ<tdk`Gpf%9Z9nh#G9>$3}GnubSUV
zyOdCN+q&T4!|h-EnMdCm+3_2DKx}GxC;#6r%FI(W>2?mkz2}*r=gkkA@9jL&8s9W+
z>lpdsc%^ulJ@kM|f1}oK*_Ih|nUS&c+J&12u(w%#-`1U+Ue%R*;L_0DSJSQC(m(XB
z?rQFaa&~57Y#j(<p$4wDtyg;)P~<uYv!+<HGzPy5G>`dQ{fXG4q6!8S9RXo%3LDWF
zAX{DDKoORdq%fcaTsEQ@AsVBVXHOiNJ?J*YfL7z3TT!f78UtkK&F<|kEBU`Opjf=M
z3B?4_7$AFo+W>prEcgO}G9U=^qIlV943O8FKh_WOIPJiI;_%jn6hlN~fE*G3JImHr
zreZ)bxXh$55sd+IY}&K5dd?)10Y%|5gTg>G2FNAnn^Z4(x_Nq#3?JJJF*DZ~AeW~+
zY$!b@YBq$@ak&$QI$^87{vh>Tia%YGRfnGjKIqni*5g&^2;Ev^UW*W<N=d+W$<td9
zXd^D$8{6Ax43KX=$`&`Ffkzn727G!f3JcK~Ab)>)U!7H5WD^5I@r|{hSTHmO$W!u9
z*Bp4Kb|C|b$6Fgyj1i3i@~@LTkLcJ#?hI%x-rAC4$<P=eL)pnGIi^$y(w&#hxJ>o8
zz!c2Yz|{~;8H5aMpa~VQRUlOGVs)H=^@9-6G0=ntC`PahG$RJM*y^u9!P}~%1axU@
zwA-hQIy?a*SR2f4gBOANcIUFw9u)cS#b?$A8@0j4KmZ|ycchLdFdC5YNvPuptQ5h0
z=H3ZQfjfPf14p{+cl<#|Fg}%U1X>}O57mc6zEocl`BD8y<WKb{k$@^7Q2;f7M1j;m
z5(%k75(QC%NVJ@~TuYKT=i!0=+Fkg{oy?ppfKR|-z!#`K31Hm_7O^T~Jy?f~Cc2QB
z%zPK$P8XUB1GHTi&T{;v%3n_h!R+<y>7W?FxHPV|5;#uy0P3?u44{NjC9HL1l<>B>
zD!1<(KE2nc0&FcfZ+Lt4VI$frP=&U2eB-|^Ow6&%|IvCwd;6IV)F)LR!Q8FfZHXy{
z2t!ta<sn$;%Fs}P#9XlpK@zTnL{hGlL}A=8Ev2ybj6Ku8=BT4yfpsUm1gt>K4xO+o
zU=XAW4E}aCAXtc7hz~Jtgh_-pZ1JGC$Z#Mug@uAdO4vdPYk5y)^%+M0jU%|TUhgCb
z+I^2i);_8OmI#7H38GdIy+sS6NfaZ9A(31lCsC##lSEm9ED{yLf+Cp2yJ5j@G8e;w
zVlr35f@(6?z=9ew*TRBYg3Be_0)CVAJ;7Jy%5dcXM}j?tb^d3u2bybtc)M`s4jvI)
zQ?5CvUt_k96QnRv;KeRX7ZXtYK~^ijeEOY-6ntwqQG8-+<QzGPieOF=OyX*oQ%x`k
zP28{QTN`XU9$jvs0;0<ZzX&No8N!S;1Sy0H5-DM!64vsmTHzw*_3Ib@5#v3<HF=*3
z?~Ur?C)?6V8{-r#L=++heN(Vdap)?7#60mE1W9-j5=nVd5{2=?wAAyHXH{-ta}VwX
zCz_Mi3w>T_lAfq(mWp6;Xq@&`B3L{cPoe}gfka!-EhNfDvq@A6qou^%fM9!JbT643
zV6=f?;6;6dzokM13pWY31B*nk)h4S+w8mr&iBL<_k*F!Nl#yti<vJ~a^opkmb#4VN
zzf=12t-dL}E_`2BQdc6NCy;IN+Uisb4nD}Tc(FRgf=*_sCl2ZOc2?p+xE0+R4?01x
zJutclCUG5%){(g$M(YWl%lZ6FN1mfP_qikPreJeG?$$HV<ZLh({GwkR;YN&=;h3?M
zAcdoXL`v9E32XWB%yyj}iPt>wc=Vuq7=dw~S6|ZKX+L422~%>8#J>*Yo79IQCp$3*
z|2Tr6M77a8x>$M<t*HI}s`Cp{Qj?Q5B|@5-uf^oQD&c>r#A^!innb)NW+ut;FDoc4
zHC&gXw?Lmtqa%jKOoow!6}-}kWguajo0wWy`uYh10);`#-{iRY+w)x({l8K(>xDLU
ni<dY!I`drJJ(haDwamxX+v*K35yy-hDoxV~S+hb6uY&#y05Xi{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm_fastgelu_impl.hpp.DE8318589739E396.idx b/.cache/clangd/index/profile_grouped_gemm_fastgelu_impl.hpp.DE8318589739E396.idx
new file mode 100755
index 0000000000000000000000000000000000000000..10c9f63e4649da63f66d6023d4b1b228d357ebf2
GIT binary patch
literal 6480
zcmds+c~nzJ7r=9$K@xl<;emu*K$apXU_^mXmY@`ZF`)1vE=5IY1rZ6f^0~3J5+g)|
zwy2fXqO?k_c2VmmLRqYZx>m&^wRLIHx+_8}M5vwg>w7c(dd_za|C)2incus2=FQx>
zbMKunJuW6D)}F_UUKm%9y(oV^pU2}_p})L*&0<%I#|sehcm-7rNhg<H4v@6(zIDx3
zc<rr2J(*_5BJW)}fAjnm;XAY9-|VO2LL}t_D;IBDf4lFsWyz6}<J-=iS$`m5{GENn
zx9tszb$4BDjB@J#d3ny@!NU$qb>E0`k9vM6|7vd1`t#0r!gsFOG%I<3@hJC=f;Ht%
zRM?ky_I<f%YTbmC0%7TCd(RC%o%N3z=5Jjn8nNty6OQ?hg0}k{@5ot}yyX4n;hL$u
zmy0TzCiMmEQqI~kYs)L0%$^w;U0EA0oNc@ple;pho3UB`(#W$>xrIsHm+dQ$&G`1k
zs&lH62i@gMGq3L7d~QwCV1C49!FHFXos-k^t<`gK#6qhLpNMKY)Dip+^{%!SFN?i9
zTsGbgexfN^8Qy%Tx@yMMGh;sdEi|<-UU48*Ax-`6hl7cEUpM#NUvXNZ_<$)=#%mUI
zGK+t;7;cTsx@}65aG&3{?%#qXo8L+aajkpQv1R(*!i2WsgD2*F&^*DuT9NXC)^nG{
zc8*>0;<j8+448K`b!Tj->m4SdYO+!WPWUbH+!CBOQTA1PuKC!7^F+gIBED;QZ_mTU
z@o6i*_g>L=t9;vzjGt87({(?8owr~v)Zf{&`9$EEIr>+H`Nhp!H>_@%JVJ7!-FapI
ziTd=i>8aA2d(yoNdyci-_ndfN{OHGDoCV^(Yq#cm^*)FibKl^T7<F6U)0jHqK}Nvj
z7$5DJqhr+(mfpWB-b+_XzRA~I+4J_+E`P_1ukiBzoxLCu`HNlpqkFa2s3#OP(hMQ`
zFg6x83=g@F>9b?$?%?%Zs~^%lo8dgh+05Dd&o7>k0^Tq_FrZnKk!NFzG}9WR&d7Yk
zKr@eE^ds=&dIn#+aI><va~Pjdk5ap_j}Zl`Wte)vw+F_uZ25}Aq{VxknkYWg27zrr
zC{|G8DyRTfQ51CilbXuImx0gpz|<a?#tNh@9-rQbULIjiGuc+zcI;!CS!%VEpd70l
zg7O9VVyyX^;5C9?7rai;3h@eCR#VziRC(aa)<F?kQ@A>keT*nUJ)NMJ)$=eVFbOsU
zIV3m{<d-mxptyv1Oo>b)(J#><k?5BwPka$;_JdPDxL`IyiV;#V4}#MmDKDa&iYQLs
zuH!So!0yB2(Y6(a7RqtAVhAk;g5Lm_H((Sicerpomv<*@(Q{=B$B7S@Z};4|?H|59
zN>m)86P&tE(M-5?xGQS}nyCisYT($_V&k&4c1$Jm=0HW@6jqJT1ce518|GH0dG%gV
zSSeBp7lm74rD_mW1IO+W2ah}s{?=lM`{L1cq}5`=%Sr!Og1>s#9ot;rkN+VEb?H8l
zHGF=z4f10~P57MJL+yoUr=F+|#?%V7t-u{9KGO;jtq?`ZlUm_lnA^ak4T#Ztz@-P<
z&2**%dSE=(?}Okz2*qrKU?a%`;57g*VEHrfdPcIAvei<={PYx~r--%bDK9-m^fyo<
z10})Z8K?jQMa;)QDGk(AEU%#4DyY#c%YAv#4u{Px_%zc8c71rEG-CuiBgs#}?kUN?
zgWd0#W9rvFSaL68hXu{}N&L9LgUHR#n|&Li@qXN1&`gLhga}wuET>pw%~h7G2&xCG
zdXQpi1B`3{N6d}j+z8yNXr>+P+wqZCP1=-WJGR__4#5KXf*9N_U7pU-4B?c9>_w6d
zNT~*dH9*YlFjyVNZ%>-Wc;BiBGZWE_o6Ic)kLE6OCup>6G(jFR4}v^po&<Txya<XC
zMiH-82o(fH3!@3j5@r#UE6gQGBh(O7AS@uL7Dm=$<1N<Yy{_)t9ce~0nLe4E=g>@{
zb0O#VG*b-D#dy1ErUYaq5QZ23IS4+-F(A7#(=PO~%T1K@PJ!zw;F4<IN9XG|t(&)6
zfOg+4#1r?IJIc8bO*8UIaxQ=)N|n#Wnqh8XM8B};u$Qo`0me7LM9l3lu^pejf@9aa
zR`AcAMv<fz+-mW1O7>07PJZ`XD>?>c;9W+JP!8VZB<sOjPx1knb^y59m6<nHuWQj1
zAt%}dahq_mr<qEKtHfMhQ~6`gs@s=^d?w2!YYaPMM1EP^2|*N^#W^US$#u@<A~quT
zT+R-BMkCcY;86;K3lx}iFiQusG1r2m7P!;(R&~^-c~3RhkyQ%D6>!FTTdK8<T4S>Y
zZD9>aYH%KY`^mhyC1R(GNWTt(>TvL^{3dqyh0-;C!;tS-6>?eM<U3ZyKyV3om*ANB
zp5khlOJcm`bCr-;V}r$?)hJuT?pPNl?#?vkA=|aWm{#D5sQ2%UNjN5b(t-q9u-D=Y
zQ@tdfuS_ic$;?z|hPOc^J2Ic?1;1WA0-w=R;aZB2>#0~hrNsIbRAdFkxy8<hBYnE=
zj5eY-K7+Vt*fndSFFPw*jN6B}jeGW6nmzl^PgQ6<1IP?8nSF6>hQ>L(?b20rCT2@#
zb1SBqInp^Ced2R7d3JZA-*8h#mbF3<+uOv-Rv<E18`!py8QegPGEiKA;WGwGZlJiL
zh0jz_q6(ZLKTUbVam|OxkI^*KtkbxV^VytB<9C*e8I-@&mTIn+F>RAoD~x1kL!*3e
z-viE=`#{zQlQ0`WW+Zt4+y;O<rzTswJtJ95S!pR^d_85Ur-=6(D4~I}WXC^Pr>%Q6
zt<V$MGDVugl^HZMQ#zBQLw8pklCFRJ(GUx;Bcj`9OP6KxkY)aUKdKW6zWsV;m+HAX
zp!IDRQ}@YxZVb^#Urw05&}_=iyU0lX;r?7~ZdB^r@1I^~I==pK{?pj)IyU*iLi*R{
zwT^ryHYAqYAViU|M9mqiisicS8D)r)&{Re$2~DL+NocA<RD`A~QblO0R4PJK2d+At
zg!zmPyme&Nse^Gk@MouOlFM~CyPG&x2XUl4Sw~jMjo{meFVuXd5oC=p0qZw{ypbHQ
z31m&Ayb0t@f0w5;!A#u03DTNi9_D6{HIx0DLEcRE?*v&VDer{PPV!3F2`_a*EIZz@
zJqb%KV+3ywG2)%n)FwXtIXmyUIv1xTlRKV@?}a{Bi#juU<+_sa8DbRO6QdJ}da(^+
z+JH-uG)k!xdV#CIXr>=L`tdn2U5e6-kk0xI%?yIaASo}RJc{t`4CQhvMvGH0jjnPD
zdMXL`ub^Tp@a^nunAfW5YNyH}#-wvMTXO0m+J7<UGi$+SEzVYarW9;SN!EgmmgF*U
zFT?jHK2r|v<s|FDT~BfcxOU(;a?$90sbOEg2a0+<kko_A;O|_Hnsxf3y8nMu$u-N3
zmMy+#gOOmNd|@<ei0i>;qn~_i2SZ$V*Tp9;24VIURHF8ROD`^(ZUmV1nv11n=rx%x
zncQvpZsz{v5NpM8lsyBbfxhh9X(mV-#L?Z&`+s}a_v{Yzvs{!Uinz%tBnpC}CD8=M
z(y<I4PDv{XQqd}cE`iM@aA0-s=N<iGZ-Mu9bn(!EwT>*3(d5hUR?%o{=yH<vK<hEz
ze>$g6wY_u_TCu;_pPO0t=Y=x_mjWHp;~!wc54cu;Ty><;|5G=QAzHjIgMWi3FFc7X
z-U{xm;K44$sC&HO!MN|$fBK+wE!eXHlMhNuN!EfCmAF`52A=5Bg}EF&(WMKs9z4;d
z3v&mI?!Xn4@ry0Z<K!m~pcPg@R28JLii1sOyxs{bsQSN~ty#2WzUfEgJWYDuc2k2k
zef{@i`d`IgpTB-;Hjy>^jK?#_O-tMq<E98VM{rGwxtf3a#dyoPy|lGCZ7Gt7twsn1
P7CZr`!Lz_k32y!Y{&aC?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm_fixed_nk.cpp.12EAC200C0F64A96.idx b/.cache/clangd/index/profile_grouped_gemm_fixed_nk.cpp.12EAC200C0F64A96.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a492754b7733f2843f3b5bd684d763b00085ac5
GIT binary patch
literal 4264
zcmZve2UHYS7suxfOCMOuvNY+!ii;4M6bVh~2=3rY8^m3TsGtd`!3Zh>5etF=_A^J&
zs8NCl7D{a3K?Q7pU=+b@RKO1i#GnQwmgE~SZ|CLY;~d`3@BQy9_r7=U3>oC_|E&gw
zElv!|OiS7*VqqAjN&jbTOkZ<b4#Uh<Ff6mYMi?7D=*sCl(pTBNX5Lww+#7N~bS}R~
zb~p^!ryKFk{qm?UL9zdj$#)lntV_GcZR4tzcD4?`zQ{8RE}#T@ymT#_>euI^ws~qO
z$HhyNUGDF_bNSGCx|2trW_k0ofVv8~<9*%F_CFoU44B}GJ+cdSj!t$we`9Cb{MVL}
zj%I1m`MrJ9SxP>VY@@F8ebUzJBY!@gd-<UEE63K}>gq82Fn(lp{VQ|bm0l+bx2`Bx
z=HI=&?nrQ?OXuE;f<E7KLaT}&`iH6-d+IN~-SMG)LE+)Lrw@#k)GqrPw#F8`JDz8p
zZ=X4kS(A#LNUc?M5ZCT2vZ))K-bP+gDQiqhl;jBhYQOgRf@237-HRhnVrfPlEo%&w
zhD#^=`@(FzvfpL}ch@y^@6R(RQ#Gs~7j_QZj%-suefyTzm?{6Z&@ATu^`TQDOTQiN
zkv&RWu(>(%5;xj3FaGibUMZQdABkzXrno-ovvHUCOLnEKZMsv!-n2bw&aR2hTRj&p
zej>P%8{IF-+q7AD`C&xmPiyxz?pfRN{f{fA7oR)MYl#_(@#VLu*glS`G;VnK^KtX#
zDM{&j{O(;#(R(qXx%D&AgYQ?a*K|nNF*+70?!EE8*tTZz$wko?ud+h6hUa^0PvR?9
zzSJ(>s9x#x)6+up`c%!;Va~rO^opx?T@M@Mx2F*9ZYQ74tIcWPylD;@={hCt&)?J~
zIV}HJFK&3sf7tJ6%Fo<<@!F<&5B$5obTxQKJ-Tvd+Un4TFqY6oq2@>BBV|$dA__7%
zB|o<7WGHl2UyVt!{PNDIbFc}1>6_lS4X;h{v9J9i{0bhLADrA8dnT)+PV4+07ht}B
zIAK`ed$2&fu(`#NTlM6|*`mRKjKLxNRN)gR-SnKV@?-sb-OEkp4r#6RmCxF7G`Th@
zC8J|_&6r!Qq-2njkZ@;XriZKED#7ve;dA2cbL&Gr8!h9f3RnMe#B8PEhJ&XQd&NPC
zCr*u>z86xw*}iyb$#)|;r4^?~-?Vf;HLX~8XvL)i4!ccGre^hLe_LxMKI*^YTG`&{
z?XB4E>0!Qt?e4%2SC5Cs-~2paTh4<9r2(_4_I3XqPwfa}xr?^&R@AEf#&UJM*s;E{
zFVSH4XxiL+$Jpyruglj(S$4P9IZ3~Gy+QsevFQHD_dfZ*zu5ZQRsO(VnX0R61=2{<
z1mRknpB@|veE5MpaiTQJr!!M0C9^<1UG%Bep(VGTsN1~^Ia;Ep9D3VW>C99{^3xvA
zW$K#Ng5QAX+*>`9*nL5-Y@-d^5=N61Z`rN;&wI5A>g&mjEotItfMF`B&hoqFRkf;N
z2^dDz;#e(i`1cPsll}Cc&bwWY%nfus*jR)Rqqr=t`fT%2#4y8~@htD$++;Qu9f?uu
zEOlbGfwDmFrJ|30y(JrqqGdIf8Zp~IS@=BG{+m@nMFbW@uWzDhVm#YGS=4JQ%zba}
zLtu%tY{#-Qn{A++d*xWyTfSVTJhq&cH3?0<*#^oIT-3F@W=x5Vk@U`WSUT#n4V2Fh
z94PtSYrH`oTS1RjB~<6kHc+nT$LEmkJNwyKI6YdMrL8gBK>5W0-#$9F$d-)>Xqm&}
zuxA@6W9doB8QFBblzSI3I)AX`Bw$fI6`mS&7zQ5TFx-ESgyEJ`nGG{^k)WmkZwlDU
zB0?47yh2>Y@EGJ(X2TDCBdF|%>{w<{4Ipbk48u|&kb-E23vodqE@Kn*9Jk#{cb@@*
zs>HP`aZ_k$b}2n&e22$Zv{RM1dL^z21$X0@x7R6%UaGMuC&Eb`cC#pF!Wp53#6pBz
z2p5E030H*N2seb>33r4R5sMJ=AUqKABs>xFBD@gtCcF{yA$(*MOj+N()*{(X`(>eP
zp$`uWUSc}9GdGwCZgN=i3g{z2NdPH9gWCoKZD??(fMAN@2xY&fv?PuFbZ{+|E!l9b
z<zZU9kpmv`I6;}no2bHWf-+M!lNrHYMU{V~3qgf(!pxWrg>%9YTE<z15NSa27|RHQ
z2!xg!ESFJ~cfVfO=H^RuNIsSacyr^*Z!|soReR6Bg!s(z4Xc{24APFp=)}x{i^Nw|
zHCB^ut5pcf+R)krx<F6?-T_OX5kUnm2@GT?$SP<)L%~+T2!&XMAQWm9ijdGo_z5#k
zq$iSXgP_EEVi`djT;(5yg`j=4t;Iv2uL-YumTuCHccfiv1-w?^1k*fesZUeRxszgA
zh}DeMh0`T#<@HS}-s9mjL7P*HTfm8C*s>P4fxIE4G4kv?Du+JpBwms;oR**_0dEq>
z7*+ulm46gCI(7>J+~A@FmCQ+&r6`^G`q`F@CIR{qfN%}auR*yD=(nLP1^QByr-1$x
z%7wUoAugLQueh|4-EzNzj>E@W0zKI^;Wx*vYo7rvMCn8s!u{+C-sm|M(fo?;=COHP
zcp3J6Cy{Q*8huRPDKi~2J=k5QyNm0yjkAsJw$!lXz;0OI=~6jU++mOistxek&{8P{
zJSoG6h8G%zVJ~jcYbOy&25_fGQ$tw6g0BYXMSKZgSxGoYJ#sj<kF7;pJE}R#3LUHg
z^oh*8Ip8Gp{aq0FDfE&)5s&0hhO&LK1q?}mw*)L>xCVIFpxg$$+CVs?OM#aZ<tgAb
zg>oV8Rfx;%>m=T$8a+Kt|4Pq#p0cpG<LD^uP_$08KAa!E?&x67yz7`?CYvYQLL>T8
z@v<EuUGOi<to0s#q(WCq0}EM3v#0>SfFNkZq5^{gmogN@53*w@m>-N#2tNd&P<|*v
zLVKYzGfreAlGQR6B{mYv=)zOS1DmtHyFj~Vp==?$S<OylqP*X+F%ChclBviWFpW$@
zXg#?ep<F=bq6Ix4kohQ=1F{_D3P4t%d>D|2QLY1I9m)-WY(V)gAn&5w0mu%NI|12=
z@*p4wQGO1{=P17i<a?Aq0P+LMe**GPluK~31V{WY!^tv+uMHNq4)Tlg<Or$^5M@A?
zIWt9(XhnGv5R(jdT5K)Y=^4SMkHnYbs|Qc9D|Be{!i0ngC4x%frntgxg4)jAj?h8C
zJqQp!1Gr~Uz6rQDQGNorPf(r)+-a1HaBdMUo9}V{^WNq``SbMNmMASD;3meDP8Z?s
zXNy?`l?TW?5C;W<st05}%0qx0Lb((tOL3X~uNjvct222spHS9>wGJ#@@avJzwwX^y
zCfg){Cjk!3YBhjY!|=<|P52iH_Z;b@wo$cFhqJvYZCBhf_Zeq4*U|SD{S}%jnyF10
z=?NJT^sgLscEg+M4&O{eg;k>Xm9cSH94uWJ7Rx-wGmo?5KjJE;cO~<fkd~ObhQ`p<
zVR9^a1w|!gRYFZugRLH{<-cT}mP)X;9!E#tz<trDo?hNQzMpUxc+7R3Z~Fgwb>|ov
no0(f!@~rtbw)PH=&Ms~y3k~g@{J6SSfgxf%VA<jzuoU|rz{m(`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm_fixed_nk_impl.hpp.B583CC73B5188079.idx b/.cache/clangd/index/profile_grouped_gemm_fixed_nk_impl.hpp.B583CC73B5188079.idx
new file mode 100755
index 0000000000000000000000000000000000000000..737982474d2eb5496f8f4481a4c176dc0bc7df16
GIT binary patch
literal 6542
zcmbuEc~lff6Tr92L0BB_!3D$>P~;FnLFCx|IF(Zb6hz_?qv8o3fJ*R05ikk_vLb@u
z5szp90YwE*Gz0||7c?4ui9$TMJn-{`koPs_o9WI!@%h}h?3>?IO;2@ARdw|+$=}B(
z#ZV;jp6j0w6BYjzBNB;p@Lybf?7|QbiNe%HqJ&*%!)GkLXTrS>?(%l3nB<!GYVomh
zzo0k2RM_2<nY|nCgKS}!?UC;LsqT&69$YB>BBFew%aeD~2O|Fz-^|{tN7)4jP0qfr
zk@pNr3_km+b$O}j&(nT-pOPD`BMG&-e)5O<wC-zDrgj^h`$lf|&NX&Hv$9M1$J+@D
zc0cypF{@_6vd1dz-J6Cu4n6)NEXi)vQRQEeW5+9QNpzoR++1p6$K6nws!?98YUulP
zP5YX{sK5u(r=xBfANcb5^zWVT|9EPtzGSuj$>NsSqb->y%~JEHtXkpWv}DC?)p)(P
z@!9OEe$Nh^{?+HsLw%coF+TYTEr~L}LCvl2w!cbj+xj4JAs2aX%f<EsJ#+3h{C2+j
zw{z9S-Y)wMhJ;O+?$A`D_<H7k_oj7Cw_hbihUXdRNTYAL%lwbDS;+H;ugQvfomu9r
zryJcmCB{e1Dg;>>*YDH|oc#jXMCD!Cu9vYatpDKry4UKudn;|?OxEZWF5NtONXX?)
zD;HQ9m98k7yCfj{(xhbz2DzS`YIM2HSh3Hc%kE5jL+MhN$mi9zy&fCk>FtqgwKd0P
zzjY2AU2(OfMW@JgfJ5n?OWn`5ZAl#+)4J$anbD=d`PLWxXH_4cnWNRvo1UeZ>DL^Q
zEV&@zik{w8eXSc68GLutAluW+;$?55Grxn)D{qxW{9q$HVR_^I%IE6}C*Rptb+@Qs
zbzQsd{Km7j^Va_1-Xrh#w7w%^gnnIgYU<B=)4p|ia>(t^qs}i5^`f%0LdBBFh3gHs
zUGfHtkkKP2#NJsjvh2$GH)n3z#C;tz+XKg!!JHSJhabRO0Er4l{NW7L4Rl4~zoB8m
z!<A#_%z43z2KE<mrYfcuAAUq)K%#yO;!teVyg2bdto-G9&+86WTEKFJATNaR!j;@X
zkRK%6|N8lkPU}*2tT@|hwkuJaV>L%0jrm4S3x71ez%X1T*jB<QUX9@z!LJbl2seXu
zGuRMr2m5vy!?Wb#u+m4~tK8LCZl%sj1O7D2t<qUVQG!kaMQhd8Y7xzKYU?OkueP3|
zjYw-F8c1|gkTeCE^DMczA?MidxxIK6#b8nlX1svmj)KWin#;kYoaPEJsUU1{yG-rR
zLnrO~G2B+<x)sUDK<UUeo#sp=%|t_ZxlvQia#6cH>a#5vH1_$beDlnmAAP(-G?QGm
zs9kNwisSUgnec|N+$p4Y3JGkYp1EwZfo={koRgQ65G)KQbCvyvzsR&Ic0uO`c?wn<
zr#ViCKh5B=G*2Oc&D@VyzI48!{>k1M_OP;6eay_Tk50-i`1|&O%SRM_RMA!1RkyhR
zS6zAxR|N7Ra3eOCfV_m}GO#YA<MudMAE&tzq?I7}t1k~^l_0o=;i^DdMaz{SRsNHG
zmEceA-w4u1THXk<MtZzP2ycW4qTdWg&0tKkQ!}^_ZUfym;AkGx24iWS*ajhl+rhjY
zENC9r4n91~Mu@zPPO7RiEOwl65m@sAmMa0{5}HfFxRmB&V0?_QNB;`_?b5Ebh~>JF
zQ5PD@59D)Z^RwvAdAl`P&R*M|%2|6;d%-L$v$uEVFJU=%u{)LR6Er615!nol85E@<
zjWk4Y4brJW5~7_9I>|tBIvA#dkf$t{1A}uwFqY+h0>htZc?}G%q2;w;UJE1m@dM94
zaZkNg81iA(YF)^X7X&$FFC92?+qeB$Zn12!;BJ<SmPJz(Ba5NvF`D=oMe;+i+!HkM
z3C*o&Vk_av;_+u=qO>z`a-2ru(};3d57Ou%P6?Z+Ytfn|SJ7cPa|d%Flvr-K!*Gf$
z94siZbg-nz%E5{vYX@tJe8s-ht^LG)6#0w&DGCq=P!unYrzlaJNYQHXYKoGOIEkKK
zG7=}#T#fXr(IDctUZl}WrXR&sBuH0uAH>fnA58MWidS418@Dp}-60SBjN+`~?D^9y
zuJwbrqTz(o(cpBNGmt|Dk`s9+QtKpTk>#EtwP%E5auyl5-jhB?40jtWZ<G5kPjG86
zu50YYf?j0Ni!AvO<Bl|*$xfTKb^uNUQ+HumV>!=3o>b_1I(Z5U5|%}J&fo`NIWJQ$
z>N+oPFJB_7LAEuBnr<@KC4&Qzr^DEE5L$x16miWV6_n*_z_o^O!lmYiiH!0Zc5gK@
zttNijSUe#nyg*qi#v>q$93<2zmfL|WcF>%QEOKeiLvDE}h**@5-12EIKyC#zcOlO%
zB-A{XdxIvvL6kS%Bk%WU5<lLi*Y9r@RQ!GdTYLyP972J-Aoay~W0!_)Mjs-@B_G^*
z!B*7^d($q)ZosK_7=;{0;k<z5icv^0;cfeKep|Yx?VdWr#Y^LbVv5NoUMO#vT;qkb
zXSl?{iPU<NXqG5EIfh#;S#8AMjNy`ya}x3*ycwx&My7-n=nDm!Mz|VjSCd58c5k8c
zrry;z@br?AZ88!<B74AXcFOF|U@UlrhP^^Uax3<FX3y%0ZN`EmWSc}D(RNAdK+kpD
zIvl#Ek@jgM<?p%u?W`FawTy3JPi28=7L4QtEO!AcFOUW?=ZfE<rj&Iy{rb{PXDbp`
zJeEsGI_WfLAd3tVuZ3Rs<<ihVjn8hacBjd;x)ycE;mIk%UP;!_J@>^Hd%PpRz>`Y?
zgCsEI=dh>KI^>e%?L~Y=2AF4%u3>L<AQKRp(t>TO0>3J{O>TmbO<>25z;G>K)Iu-s
z9iZDmmh=7I_Xc}i?9TsWuk25oy7SSBbH>>3^Q`7k$vfX_zCcH<9*0ls3bj#X@Qh_j
zx(iXlBqf9qGaF%KBW-pY3~d7{F5AGp4Lpf{JLtBPX)s(ngtwEfrfm8h`+eKAbn$MW
z2%L(5DqAJsR6=tZn3s_f$#TcR{5Z{(U|0!*`5X5=IzuJR&7j{*o*2uufkqn$Eo((b
zVuj?Jmq$LiWv$d~Ye-qPNQC|S@7t^@O|=_8Pkk8l+2MKS@~y~X*?%8>vZ?3BOq#1Q
zeqRTUQb#vOq0XES*kbnd+Oodjb@X?f$b0ZY>Z}7%t;3o>+ulp#p3P-nS8q0DxL}uH
zA?7f71XF<$926|vo8dxTLMhEqk5Ec8G$@qP408#iG{Zc?D9x~-FiJDRC4$n7@Q9!^
zBZ49*O$9Ph{Ih&oD99?za0+CrKveE3kW4|={XV&$0#U8L0);Ek6yCE8SBUHi5w(Th
zjqG;Qya(Csp?NQI+)HXK!|g+k`)J;e9QV^)3l6oUqB2}9$ZA38OZwQm7RC{Ku7K<c
zEx!WZS7>=1$m(c$9eCH#@_LZf)AD-ot|#(KKZGpN@KH<t<l0r+mbamxb*vKC+a{1U
zk<C(l?BUR!KVsrOJ2tOskGtKY)#>`#16?1wI&GudWG!6Ciovp&tN<*2vtW6Y=5nws
zr?~<wD`>8QkSbDfSv+;eCbG|C@j_wN0#p;)0Y)7ltg{XB;cF%Z8|Qp-b>O1`segE(
z#J`48KEAlL%Khe(0SuRg@K>MLW4LUDzZB;nvmCO$V{kKPwu9zeWR^>F9&*V;R7U0_
zmwcKFkV^s0sbHN-c32EvYE3f0l-RQqhV3Lv(J#`|D!+~}x%$z<e0SjazY^U|5^?=|
ziQHb2HmcdV!TW8I0s3TDJ{qO+N>xj;6^|GvjT5@}jzuNmE_z-y*diy1ldx4}IhjNz
zP{(&AZwf{<?8brQtL;neiv6_xDDv0#rzk)>fTCbFSeN$-%Z0L`6os*26h*KR0zHj8
zed5Oii)IxTmoGg9-Dl+>HitauzAl~JL31u*b7{^)W_d`M(bJyk?}GAEoch)mEl0u5
zuJZMh)$WWQgiqHZpIS7TH~HnR%JX)mro%qj?eFF>k3_Qd^|)Lp!CFajs(Z`JjAyoY
zgK_hng(O*IGu{`&l5CoDkU8Fy@TT;|usPn75Y9#Bcuzt&56SVKgm6BR<2?!C0wl+K
z62ciU6nBV(cY?)E(#3S2*imgOt3QT)atwJMLtpZWz047|T=zr)7EEQPYW)2wSS0Gb
zaD8~5JYm<rMwBQvD*CIwE8=43#2xE9(Cz#A@Bz;HvTjRlzqh9}^;N}c>a2#QmX5BT
z_5hJ4uOi~#JolZ@Cudd3p&mIDOcRM%a#o8RYLmkO<PiTah5z*ZkLCX#HU2;(`XBrK
B`-1=g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_grouped_gemm_impl.hpp.0459F4415131A8AB.idx b/.cache/clangd/index/profile_grouped_gemm_impl.hpp.0459F4415131A8AB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d4c9fd9671d7cc78c13064b9be6d22c5ddf3ceda
GIT binary patch
literal 6162
zcmbuDcUTnH7Qp9T24QiQJ~Uwkl(q;Ke893BMOKg!U`0tFDi**NP>Mtp3n98xMOVaF
z2sS`z8hDCm6tY+nK+)LYMP&ga_7W_J`eu{N-247{etgdN?SA(+=gckVo_o%j2@CM^
z(=j6mX-q(BeC)EhBtZ}c%rALa(xNa(5LG$^k-F<l_^hSvV~o#-{wno2VpufvUAwJ(
z`i&XaPsKgioalO5`SM7dmhqnVuePr;o~9#_c{wRx@7j_UH~iej{sk>dRDt}Qz^vhG
zrcU|Le=Epg!i<;KPA7bKEXrsfUuExj=EjheVX}rr+3PFroM;@8{zM~lh+tQ-OM}DA
zde5e1y7ExZeH-_cN4=gpc4p$Xzy#5Jy8UF}*{pee^2SB{l&^nD<n>Z{aBA1db$e=D
zkGcjaL}M~bpM0&WIuv^Ji<wWROFVLR7Ui7tOFS{{**WtcewOY#A>KXA{rFpr<C|C4
zXZ3f0<*G+<8gYN>nvAsAn<ANe(q(^kLc;ryl|`?E6N)^uDtpPfDXQPcKDztEtAzZt
zW9|jgoX$ySe!1pWwBca%iUGIB*H&+JJ+OJbuEFER%Kk4+FD>pj2s&%9(=MQIA3x+6
z(RHH2-DX|4<G0;~^RGM?T{=c=3G?+`RBuO>_ywKYJ8P1XHu4>^?)<<Dm+So+IwF}#
zaE`_2wDb1+88Ku2pmx|7S&SMV(`&qcU}E*bs~IQSzE5h(UVOXwH$Bo{Rvmv?v+i)A
zW!I?JH)PY~BNp0d7>(IIaDT~$>^EbpdZYK<a+B9xaC$X+cwgz&npIb#I^5Egi4EHW
zl}U4Ijq^lD^vXp!dv{cno-M9;U*P%Uf*1UNO$(ci(%v+yPMe6nx*m74E%|wxtm@^x
zUq3r@zkRn-5Pdny_0WVv-Y*usko1HtoiA@U*s{ebdfSEDGfn;zTQ`N;J7WE;FJot~
zlzz}}ODve4X!WS__Bi`~t(MBZ_sJ{b=Xo<8W)}VQ;lVrb6@+G*3^EsOu48UU@cxF2
z-;lBJyV>y+VLF7MZ8dELfBg|W2nY=l&>$&xesa1oqddD|(gimsJxI~ppky2L<({OA
zp`;k`kP8hD9tscZFpAMm(OyU`#wms){sIe+MHQBfBuUpm*BUs6RU_#Z=-&bZ5w}BO
zI~;|$3y$l8<5@OUjy&8WT|QBZqL&*iH)C&8bgDrrMk@?fV6;YijULimtGyPZbfBFM
zFkT1r)&Ub%+f<oeSbOXHcg(y?p>-*=Wd$UC09qfwxdK{O;9L!@s}Y-B|51COtZtkJ
zNoNDEY#>G*<pQr<oHqc$1~8nJTU<VqNZgmie)0^4xj$LH(=LDeM>ll1?zf_B?bc34
z(bve=nw<zmmjk16z_GPXe!_Y)!$L^Xo<5$OdyurmOY#}}kj?R=g%8suaz@El*Vll(
zO)`DymIID$C;q$sx#vZlL0e~Jcl|k?@ln%1x+s^J0kIWatpp_f6o{T8dyfv?ci$w@
zx$>hucnO|0e?4UGGb55Nfszt90WGEsO3H99hr)8)g{z>j3g;RqsDYdz)ki{H138<K
zbS)Ir;__xFX#QWG(hLL8_$^S-g3DW=xCQU81%|i4FOhybv}lKxID5215n>fIR6%2$
z$E%<_&LJu&L)-=Jx}XEjle?fF%VK9jYEh@9LsE?4EK8t}6;N~;v@FB<AhbM)b0xH_
zMC|v|`p0n(=NIczw41&gcSaeqapOj3Xq=lT8p50Bjh*gbey|aeMe(CBk^{aRU|a_b
z>VPTI9smXdILo2A9CGKBqVu6uJ{*DcPeSvPxcn5fI)%#{p<N?%X7z(T6iZE9@-_{j
z=s0m4XFQ6I7sq3?M7#u}8zAHch-AB<=$jzqCeH03q#f}zp6|){SpAKRt5*SD6(03E
z0lyQieR_zY!@bQCO?`^C7us{NnxY+q4j4HK9WineI$<Of3NdmPI%DL|^T)O%fER#K
zATJQ3AYKqgDZCVn(s*eYt>Uf1C<E{^@a5G4x)uyWcIpHCKD69uk{^F--aKzW(OuxP
zE-;-{j5R)zDoE)sWdsG#x&S({f<;Nm--Ny{^VX#3Wlqb+v9~EEwp(Qbd&IfGDi`NG
z;Fbp@NPY`w-$G$9zHqUbSBKy(6QnLe$BSsZ#MB9Gmd(GtV+4J`p$|B+JtUX^_LCxO
z&g!8Q?QP?2$KIxBpJ6_jYx;Qja0!QzMf%J_no=7n_MB8Ioq}X_z_kuw%Nqcr2Y?%r
z%b~j*a!Fo&1ibQ5Af)J1(CZZ9)TYkvmE@)Kj5%w8O)auZM%Cx>;YF9uF%hBwI1~Ub
z)KK(R;II|vLf}w{^BoX!2XL#&eClf{YRj|uYr{kZa3VX^T&>1ES<0kzW`vy}U?&J+
z1r%Kj0*Z0&0g*i*3dwsxWG~{p{e{0TUDeycBk2@D3Kz;48kNFD9frJ8xMNDvX;x`i
zv`@25;|xyHt4vo}u!E6w2Jp-PK8O|ID+QQ=xEAQwBF8P*Gf6S$ZBi$*+zjBFf$Xx?
zG}C0#8sjyLzf=MJDj;Bc*!pVDtPDNNt4s*Ufi^kNg%wbAJ#?%`bynf8{$-ctYes3P
zecB)!aM^;QbAdrF&UwHg4>`#XKDQ-;sX_ctu2+h)Bzn&l_a`%pYlh>RQU2L|o9D1w
z8u=BoxHrJ;4KQcdu=}A<)@1rh$vlw>?J`lu-4ho?22Pc$m^sx#|607-Y=thZa5UQk
zNq0bt4m>q>L&I*A3inGptbCOHI|pr*{aF#)ds0qWGIn3!v;aFX3!N5nbfB8A`Qeyq
z7c>9uz->F2%uef&)7|hH&!&#jA{kSNoAK1%48xis7qHaTv`Y)V9u*v+f|$Fh;6xRi
z#CA;5UC^)#txwJ2U8q#}apsM2d-L=RwbiNHqXc3>qzrnL;am>w%2B{#G7Pk<!np>T
z*FY<FVD+AwH{;w6>2}C%B}J<sUj?}W@yLUfM@(~`mk+u?tk>;rlP4@DK3d)#*VVpw
zHljQDll}ImU#>=0iC^>#da9?!hs9|6?tQ>)zxxDtE_^fuX4^hGpP=??_W*Y;a5ZMm
zDT%!|vh$PY+cf!c4AuAZ`e7t36UjJdWXM~FIf^V;#*IzVVWKchGt4^-(+mp^!!*N1
z;h1K)cQ~dQ9vqHoMu;LX%?R%ZOfw=l0@G9gYXz`jeTk$Mz(Ij>qMBV506WwQAXcEP
ztCssK09HRMK)3>Yf%<ENZjC4ul5`^!H$pCJt6ACzxzkC~N+?$1awU{1ad{IIH{tRo
zC~d;!r=j>XE<X*Wr;)trfNTlhPkY^<gU-3$9qC2)+%GYKvK0zi(HlZb(!r^3yW*2S
z*>_s?CtrW7_t5K;JGw9;Eh}TfG(9HJltRZ+6uK$qJ=^gB&K1zH0_SSzSdDWnl+~h$
zNHI(QS1WqCp_mA7+X1l>ts7c&LoPwKN$gjJg<2L4`t-8i19blSpu&EJxpY^#^!UUp
ze`u3*J}}M)G^<C_n}G2qoC|<$0pQ*rNT&3$-HLM|ur0*-0}y@y-1`H`<UH$4h@T2L
zQh`GMHNgqZo6gp)9}R?^o^8@sm1bbj7U&tH`KOkiX)Y6}mI;<|CF+C4)!`x|pEJzB
zo~EAMd?{LDD&gqCmg<*9&TYla{`>3uW3P4r`T-aP>IY&Jq#uNljFK6$#-->mDh#7=
zDjcH-DuSa&$tR9hq&jpmg`ol%De$_c08j;Jf7GQYwH4<=Kouf>^mgX!;GJ?0b(q$3
z=jQdSX6-cX8)Jttx9fplJ($K$@%gU$hS7&@Mh@D{uji9Jgm`TWbE=!6uo=02fA;f@
zk6qhCnJOV4nC7FpNbPE-n{X}wc1)qoPDSl%c1)p-xDePeg*M_$ID)CD5G$aA0wwqU
zquYLV6}QwfYb*hiO2BMZ@f}%C)fgTvVg!+7B>(R(kOc9r@q+yhN$Rfu>k&z@adXuV
zBqv2DH>ev%>c77)J+?s*%y)G5GGQA8p@o{<s05*l?yA4V|8wcH4TAkfs=lHA0{YKI
qu?>RIM$Mt9X~;Dp{mq1xq5LbmUh0`>>1gwY@^$s}4ULQqDB`~>HOMUh

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_bwd_data.cpp.B5C8D8F20885019C.idx b/.cache/clangd/index/profile_groupnorm_bwd_data.cpp.B5C8D8F20885019C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eee5a691cb4694732bf484f6cdefd037f60bee1c
GIT binary patch
literal 2880
zcmZux2|Sc*7k|x+VVHeH#=g`H*~W+&dtn%3$U96-qGE`snCgyjZ?{C+ZkF7>79l0+
zwn!?nw9q12ZbY}`igKgAuA-Es@142dlj-|?&+j*9-v9rc=bZDLbDjt0^LQL>2=a;J
z%TuK5mSZ3Yq98xny0pY*Gz1xGL6H35Z{aB+{Ty=VVeV^jbs}@^6FaJjr1ZjHy2n3m
zX!UPD%g77A(YfnT)05N9sRv+YX(PJAdhAYN`m3%&yeax`X=39;Q+`)M_@&8AQ}#v9
zS5efnoozApt#U{AUQy;o%aCIk_Cd$|b2GLV_ItF(Zm&M@j-M(^<xY=inEW`o-s;%%
z)r=KRjjjeY3%A)X7O|#Wzj$7n+F;9jbTXYdp7vvt+Nt-CUG_a%kfeXdea#e#AHaBP
zy>HLF=e3?Uua@sK<JP9%*RPiK7rk|Ax;i~p+^u7oIEKwB&P=p<u=~YF<`%a^Td&V!
z*rfwB*Ob6BHQojV0RvW3ck{((olgZ`b)<g!dTR9H&A7OOdE|S=nH!9?ig)g4c~_?0
z*04*GH97v7R&leNmqA|{D$v_CWcIG-B|7%5d6K3!Eu26X^!AqUVg<+J;v$aJ6vV|f
z+}qOBZw*yGAD4&Z3U;&hV|&eghr&j6>c|N<dY$59a>xgsN$ZIt{5^Zew$}0+Z$w@q
zR(x39HE!2}ePY(l$>vdiNUEIFxOFY6ApbFwH)}`13A?eXgp?r7Cln?@La4&=B-e5>
zID^JnQRjPxjVkY)4^+E?KEFnk)e_bzJL*y|PHJ?^g^muG4b93{OY98d$rj~YE43xN
zr5&uy$(d*?J1QpS8aO>Hkvsf3FuOdr7TYcPI8jl0cYTA5J*Y?2Er`0kn)XpwyT+GU
zU*o&k?xxlSZ#~XEr@;rWg{>F5GDmsWkDYf9G<b=6iGRg<dL*pE^K=-y7&ZO0Oys`g
z*~!Q+I`aLAyH0Nz^syA>ClCAH=LW}5H4pz9-JT71-J*#qhSM+PG0{zwm!m@khpYHg
zbK@u0O_iRB?wI|!jj%P0Z5%;t-`{rq+!lN$Hs$2M4&S<%&}6|oUv}=Yc9FRKR%+*?
zw(!q?%hx*m>DO_{O<*+Pu|@i^-Zs5?4XWy^aO9qR-nQ+TkKQLcmq)BDlNO=7>LNSQ
zdkPm5CY?ES!-1`-=_!(<H6K=Sui*~V?C7nnG+g__tITlRrmQl3x_eH8$Z~Z)*1vRk
z|ALO6n;U-l^&B;)vVZBp>D-*lN3^L;tsSm<*?apsE^RDotN*k<+N!nT`X8GpIl1lx
z&r=B*OU%dG`}Y0BjcsG&ZoeYv7!98M{ZW9^d)u-*tSs}I3CG=fUpDXHlqXr|Ego*a
z`8u)9qC`J>u7JWGUfR0%#k;G8X(?M<CPap#7b|1D;!;)wiX#P=n|*iHWgR=(+y1n3
zT748MISI+qQ&x;4_n9GB{4SZaDH4lX3_&nKgJ7x@h(@_QaNS0X%O66?sHG7QY=$$V
zC<Rb^-Ck(!={CS2qLv_Pk~7ImDS$fY@o5XM@+=A-T821x#kd+P1yEmnI(Nh-;CMU%
zibB*<NNS)Ikm9XDRugkuxkR&I9T^pkIH%w!RHgVw2|+!HR{hm)2>o~{2H9<=ZfC9(
zK>enb>%U&OY!e+N`7Z@)YiTQ<Fy%^<#>+OVLy)@cepbI-N2nSG7P<>PK*qZAmefl-
z&m337!{PLBE})IywB&34dms{_88!?SsE!HoXKtS6ojw#Emgq?=fi)hEp~aX2WLBTH
z^35jq925qYP$jbfje%o`F%*D6J5XBRwE+(=#4I!dMy|#qw^w18k6<w{i^w8_LSSG=
zqN5686WJ=riRh$)<`d_uAZMcUHw4avLMb~962ibi#2^yL0Q4QBrR+B_!ozejT_5xf
z4_oM2m;!%zSjZN-0|($AIj3SQKv(c^f?<L)m<tikuQ5eR-i^k?Btw!p2#SX(&J_2V
zC^Rxn7of0UD&wzwcXgx_BrB5r%oZU@_^ssP`G?w89AZQxe*K7kbdU>zT>N}ykjPNP
z1SrD$@$9=Bn(C1;QOqgGaOuJ8jeT!v+y94-sj8~0hPeq?q#$;5N01y6AkZ%uxWvOe
z8=k`q^0)C<K>;=aDu{2x|AykJ*}Y>L_ZY|=Ne!hAAQcQ8Zy3)4Xe~-UJy*LzG1fFA
znlZ>=ZO@4(B@YXe5Sp$}rvqAc;=$#DZOgKdu?{7MQbA7Hdy48vSAIW)(27HYChZ7B
zXTSaBD+i%fZ`96;;-_meU;lylCIyp1K<Uep1(?9loEHe~LUj2yW@n>%9Lsl-sE895
zjzs`7e6HBrl>Fd9BVw&u=h~m_`%Mkf*a%Itpv?h!U*zqTpYk5;!66}iNkJeaVx&zY
z0wd(8Y`V(~@(b}3ffHh29*Z}B2Klr6RZsvcKn3wxd=(`05c+|Apy+C-@9G3XA~UXd
zyk;&5gT^of*WyZq@sQ<E*(6d2k+sMkxV&v3e}R(|)sZF(nL<ZEV_>n5m=92=MaH(x
zo+1*m*G1Qb4r+Mc&Q+k{CL2*S!ke@JRHeJwf8)cF=@13&L-GYQ*o9=Nctb&x@vx_+
z=Nu5JKX+ACUg&RVBoYZn(gl0L8>@!<ZW7+g8tSU$j6TGgZbD~*I(#a+p^<Lcr-LM*
zddt(K$;%Z{WofZ8k>X>KLj1Er!3P8>RZ574ryYXPT-x9d*f4SUqEHbeo-qrX`3#@=
zj7^D4PDEf(LNFSmj@8h_Y2yhLGKm;S<@qwHT7k6LI<xfYu5KQlUfv59I_TQETXL*T
u{~xPoXk=_=ZZU^pHP6P5<;b4z!ZmR=uy^v))teh2h(!5^`0&+)(0>5DR3X3s

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_bwd_data_impl.hpp.05BE2AB219AD2046.idx b/.cache/clangd/index/profile_groupnorm_bwd_data_impl.hpp.05BE2AB219AD2046.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3ebb81cb4cab068cef52ba82f0133656c85c7d44
GIT binary patch
literal 7216
zcmdU!dstIP7Ql1hf+V>N7pj0rLwUd7AR#1xfT+A85Rmc|6ju~>1xiJ0wQ6f=wY4v;
zRgks%Z5ONxZKXWo0<l%MTEQApEm^m+RwRjSW3^ga+r4S`-kJUU*MG?O!8gA-Cui=N
znRCv}P2QsP^Z<8(Aa(ho=DO;qDk*_L;K=?pKGjh3IS2%PwgN$O>xumA?c)*59i`dx
zg|hDpv+swdpqnXE2j2T}bpP`q%38-o-iPXY{7jpUO=Rny{@|n0H@0*szI)~R=kg~n
zH}vdlG{~hjo`VnSgBNYgy!!0xBTsMN{8OT5Ir_xE=1A*n$82wj7v8;GH<%au&t}E)
z4Qa-bEy4#Yw4;9==-+$uO2f^7V~cA&ZyYatBc^}j$iV&!Id66)+ZQVC?|-{x#o+ML
ze|i7l;=t61tE`7NuN#__=Y9H<^UTVkRZE7I^A78uJKWJXcz(BXkLudwuRFS4iuL|Q
zzsqr5)tQ#r>Y_45W$Nqg(#6*fUUe8=Y4ZB*^Pw*<Co1yQ=d{~?`1Pvw(c_z6(>&N0
zeah{lp}hgYv+wL)r{4O+p*R1wT^6;=Z#W>cQMLT5z>9@LTE8QcuiV{Ka^^<VnbB9o
zk^_R(c3;o5vvP^FyUncMy64u-mZR1^S1P;SRUZ!1o~se33@nW5_)vByYARVZwI*Df
zp83^E!=6>ib=31~D%(UoviREJsvBYNk1aiPGWWM6-`uU^9#exgl5p+88{TW~F_W8G
zMs@pM3Rqutpz?TL+*{w)QvL6Q7bW}XP95l3-*fLZ=gijh;I<zxd#GOBaAqS=30d)i
z#&va7$?WPnHxJeK?dbYDEnsW~$koQx=l6dCp%n<MDL~-5?0Zwk*H(9}tQ>7vefEPG
z5#uBf{s9fuD;k?-v+at(dFP`89e_q7;5`B%oMG!oaYy^PsvU?%QXgp`*UeI@bRnkV
z(BddeB|#<Om`dlCCSWRyFH6K!o>ZQOsUoz3u+j_Oy%2z<ec;sx{+Rp0uOEUj4}k9g
z1YuS~fEq$Ep8)R@*q-hAL8B?COJqx<43CTmtd=RsBpyD?D=QkyvOTkj2h8!wiN!Ld
zhmyBq|0lP&n#1lxKp_)2nP4WT=HRMLiMV)WwGD-OLD>t7IRS<GAiEE8F!w`2KNMmf
zfF}o_7_%Bm)lkMUlQZa9XWHt`x|8CO!aoR&QYEQ`J852Nggf&+=M(Ov`=k@@Ebv&s
zTVeVnb<w*PjySQ##VHbQOxCz4^>IvTD9sboeEa!NU`l7wy)i8iFCeUBh%*Sv6lD@N
zvrracGuu9!u$jZ;5H^)!C1E8`oJUZ;D4(#YLMp;$fqemCvydqyY%UQm;jM5_52D{g
zj0-Fm&2zgyyt3y3O`#^YCU;JYrKqMDf>N6JjZUEsaO(g<w*%rjARcR<1m~0BN^<l`
z;1`ENTFBCpdu0p?$Dj!7>mW-<%E!TNoE&c);>PiKu3vWB{qo+&QLGEa0mXcbQ>etR
zgpW50m4=n_JBdPNQDyv2qELBk`CL4FML-2%r59wqz-Iu<Mvw0U8P@IxML#5B9sv0O
zC@`xbNe#)EPk`(Mw&!;7(|UnHUOkQHf6m<R*6#D&`OZVh*DL9xzFX2!(X?C`pKE=m
zMKR^7hr9|`9yB=D2fzQ2AV%?t=+e64nXI4-?79GPYc7+RKU-~EX?9|x3N<lYH-$RE
zyAymdYrt7U#*GG&G%$~p7iq{Gs)YzG@F|c&V-PV$vJN72B<sOjPmZUDR6Wec_9s9x
z0em*0&=ky>0zUItSUN>+eFMxfka8n98F8_okP%{y5Qps>A>RlEn9bm1#`+1tf>gJU
zXNat6HbF32;8PdPI>wBVtOG`eIbHq2x3#7f?d-(&f_*PM#*L7DYBYB&^!qbx9qR_C
zZt&m)dD?^4hJkfO(>Ut#h4W9;RWq-$POlZNb>KWx<v97@cwKHIq>+1&J6{566etYj
zm1q<s3?e8_7)OxQUdqoPjpBvzgiI!s5hNGN2~r3ZJQe-CC-VBh)|f|rP$+xLxbx5F
zDz39`H%puOw7g7z(Z((EqJ-_<7_gDgzSUNm(>ur4JjQmfj9!_|O;Sxm<I~D}?<LQm
zQ886)&vnzNgeoDZlqw~tj4C6joGK@%f~p{>8|=Elk+WLgoH*kC*%uEe3I+NH@&`f+
z1^EXNBoC2?a$lj4B1FMcW9Jv@H(HiH%R1qOyqq|tEl0BI3feB5W}Vmz^Y-Gi7>)Kp
z@;*><iZp74v{uN++y<#_kcW9cr0>UZNuzd{*G}5|02Y1#OE~@Q8P?C=-1(c7X17?R
zyOisukuT*d#^gu&5#&$#6BNh<x?s&9CWxRoCXOJftCWvgb~7^ZgiOZB2$C~$f)tE`
zr(G^DI?vlUlf8SOwai*CZs1*amzHgDa35pmwFeS=AcGUMwgpdY&@{z8^3A(H<==8i
zZl?V-PCM-_atHIO`;F|RX&^`gVceu0sLiA@a<@&`TFTnOPCT0w(nF>mvN=TxO+df|
zzUffN07(Y2%$Xp-M9R%D*9`nVd-uCpf!7S)_r)}-bgAS|p){(RRprJFLZfAmFXLC5
zM$6|c=j$4csy(ayu;zL2JP!nq!egVj3e#v5yhh23n--F_xN2I8Q^pt>Z8}KSVfmq!
zqIbS&`iDDvv&r+wBNB4HBtINKgvv|BpGRp_;90<zbs81=6!M2>8ZGfy!dp2Oc)g&+
zoEyUC77ave$TPMEGBl9I4PdFoK1S@Ng_Dht!pXZ!FGan(!@-HYCpQV3h+NhIvpQfl
zr$)1DE9nGcVm08Xfmv9tg<vfZevLu!7|A*a){#68v&KpLdXVTz`x7vG0v|DG_6XuU
z1;jENz}Y~u5$ugP_tMA=_GZk-jq8uQynL&3TAoOI(|0u=O`g2OE_NrRbV52e>y`*p
z=*bA0g6JtAZZrmv8StH?4^541TeC{Q#_<2YE$K&gqo<0lWl!V$q&_qD<dK9QzL}OQ
zeEN$EXU`gL4?Xh2^Vyu{31Ks`v7%7CPdxv2j6yOW8Bb>$I&y#hPhI0QUj1`S+Lu2&
zm}Jw~h4bFev}{StcVNp~Q$`d2Ld}wS7k7e+i+LKIgoKljgjowJEq;k&DNE&JP=)0>
zQ0Yjy9`g10P)D;>Ibs6%S3Wdbl@}Nw6Mx?X5hhY@hD0+iAR}_W4S7oU-P1DeE?n3A
zr9VkKX-y$tXJ7u@NFhIGKZ5+7{dv+!25c?|KXLA%9D3<?Yu?H|FTeQ6WB96cB=zq5
zt}rcU*!bqrAKQXQ+SzMq2o)m2yDF3lB`AyvBPg5-Cn$o7ASjZGBq*ASCMbrAAt;uL
z<>^*y-(Yx`r_Z#U!`p7P%F5iQ>M2VuU^>Yosva0U$x~oA1^kO9ODx+NFyC82eIrQN
zbe6pjX~0VZKHPHrbW@n*sOsLvZt7;wX8sg+zvu1Yu!OT8vfT|44SY%xG*s7CT7HLV
zY*^N~+R{LG_GY(DKJ!epox=>FNKD(?3G6NXe|+q?hCtwen=^3pe?RJ2Iz+fvjGHuW
F{uh<dFmC_=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_bwd_gamma_beta.cpp.0D9708552C2EDA92.idx b/.cache/clangd/index/profile_groupnorm_bwd_gamma_beta.cpp.0D9708552C2EDA92.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5080ed9b10dbee571cb064f7cd66364147b78a09
GIT binary patch
literal 2888
zcmZux2{@GN9{-HbGK^V;8KcN(6k#lbv1FSWWGvs{BH1!XoMaGXY4Ir0;v}~&DoS0s
z%?W97IPE1JqD3xJOpX@naw<vbelyQ`Gu`LD&olpdfB)_KZ|@)R;qlxwA;>e{Cp%S?
z5eq{QL_&Y!jP&&#$`G_#1A?-v&W5CJ9Ap#wj#1Cm92YUe`&2vBoc0dJ2;wqUPM0~9
zF^jFZ$^1^sxerbQ=~WH}4@{Kc;?0D2Bkfb=*fQ+My$4%*Fy4>HLoR=|D)h^3`y`AS
zE@ehrjKy;}E`e<YjDSQn{Zq}-6Uw(<-Q7OQy7jp0;>k<1*^aaG#k$YFwU#&ka#84b
zbZzziFr%^Yh9gyZ)Pl<4{PqgzhQ2TApK>=fsz{~>><>RtvW1&F)?D%W`T14JPVD7r
z!5d|5oiX~|ny!-WuxkC&DW|7YK3ZNpf6DOq2Q$5h>AFVmt$PNgenZ<^zWgz%)!Jc}
zF;ITpV>_Es)*<|CaiFh}E%Cj+g8by0<ozQzYwK<=Qcp!&R+P3mRafPmr*7%yH7#3a
z-rm?+<~*opkmOZDi?ltnt4(UWL~pB3P{}#lpvNBTu+@|6Z7&pREWJTHq_U$Ut41yF
zPB@XCFO|9&Nc!(2(c(*p^}=3ZzdPN4|M0M9cwzs0{;4;_V-Dfr(c1*CmAgALF;$!2
z+}`Y-a?K@fuV*i`sqSNE!GX3gh}y$roe6`_v<9C(+QKqQ<9e;{NiN$HbCt$+zUuh#
zrL^@{)H{dx?vQkE=fFAk82@l*PWSuJCp-G)GSzG0pK7Dr^$zI#9&hU;>em`jYBavH
zDY-Z8X@5)b>)xc2q5+D1k%J4Vx5QSJ9pO6T<53Vi+1*tg&1r<wm$#lBbTW4<u6@}_
z_Q_CMF{HFMhA%v}^Wk9WyQbMB{>@V@L%wUq$gvhWtfwc+Pr6FVZ7MMHPiurjYlc(8
zJ`*ohuHU(_YMjdulFnz=zND))&%S+AXK!uvPfqns?N?idQ(ZdYXS>t0GVi_%KABH$
z&TGxDJr~tQ9hhCI5h(sCEbHDsZr}Vfc3JMSfiRm@2RC))@=KFF$8&T|_f|YSSoO=^
zhUwnBF%Hk&yAC-8ESfW>G)#zu)A_?N@~v{TX3GD-qFKq<a7ujRqb7_+ow?O@x0cij
zvHpt}A;$Sj=!Bx2*vP2L(E5o{+8+)gjBAqXn_@Nkd;p1%Tk)}KY<7d%+o&3&9ex5v
zZ)Vr=gqdewO2^|47R`+b_RYT{w~x3p@=69-mRPeJl+3TDhXwYJa_Y42>upr#u{fji
z|Jn5L&tEg@qxUWQ@}iV_pP}AmX3E(YdHPCu%5yu3xz$%r(_FlN#b=@KSfP#Msbtlg
z;u-3v33KVk`?&p;hf0rpkc?JM9h3TuW_divn(cCV`|8QFCN|5Y;a?_4)0%jL^NGdH
z<68{pbU*Aq*DcOTjr)YY>$U#kvB=p#9v-tM3_=KO!YVmICGvBE<Z)#1X-JkT=5G*$
zXsKzL$O#bF)ivzClZO2}4vI#_7Ah8|astGX_5i8|T=`rJibTaCNTe?(kZh@-r$3>t
z4;RwYgo9#GF$d-_<U}q*$;<g^O-3$JM8rg)p^>m8vYY_%#4SgDrXXe;4HNMf3(?fj
zl<g|{`6@VZ@m~0ELqyD4)Y=GZM3NIAP8TJJcc?&+inupVYS9&>1S0}xfeQ#}@x;|K
z^|vp=m2e2toXG;tmf*KL7OVe0g25pXx)BVZjYFa-(S`sSwWcRcY;!Kcz(@o+Vi91&
zNHjj01Q2KORObfO*eO*QVc}UsP)HcD##<{O8@!DIvc=mfAUnLB0%GIY-;of@ZBG>u
z_bKBL0-dlF^caVb*d%9ARvbbhQgi?c_9ru6mN~1SnGu!~EEh}!1i|+dLhY(=$5k_<
zP`_L}mj-e{kv-RQ0SW0sQ-H$U2K4S{{nm=sSJr<p5x_sJAs;s|W;ldKq)|akIK)`j
z*bvknhX`y0&cN^5s+;X2laxLG<2Sm#{z3!dn5c?+vY=Tnc;fnU{edkG;hFQS77*W@
zuYkPGy%ms;xzBf$Kra0E2elq1+E@`?#0I&*NCG{91yC9$dw!YbN!dVA3@D311Zn+e
zhK@cun2fS%R2mJi#b+M0{k%IS51q3hd=MGLRCu8L0^!=jYLqRj8W`FVU**Ed{ofo=
zw&JLlo~(GfG56IV>YLzC2mrOMNm&j11{IB>Y<s-@_u&Aw#%TQTlyQg$?6DLC0Ok=@
zlD%pW01k0gcQpe_r}bx&$h_DYG@KAx$SS}Fl~BU`5PV}~zjbXXfsAUiuq-X`h2{!3
zL*j!6?P!#W19HB=QfjE5ZiBKZ#uO7!uS>i`*%G($ek_`*7r_stiYjSR@IVPIu?@|B
z0dWJkLJ%B`@K`*%1;l6Z6_7W}TLJm7d=!wtMZg7i*TNRRX1uQXg61MzA+!YSPW|jr
zY&Kp69T0)3z#6b&B-At12cT}_t-E))3JIvKy^cK%)S=gcvs%qbJSA(Q8^Il9)^~-!
z<<Zgk02$kp;04&g8QLG&-dyNUB0bR=G(fs*Lw#}3S!MJ~1S~-Z*n+~Kwz8!P<^kPw
z6eb<KHn3flus@92r1;kh4cW{Q)Fo{R%@nwrDZitZ!x+&<qf`t4IKyD%yBB;3M*Q%8
z1du|yC?!@Fzc_uJScv{BBW9Q5Sne#96{^9ZA%Q_cC=jFvK|u>wneJaF6kH%Ec;Ond
za9x)gpRyi>eu@6du!^dhI#v^>MIsXLzGR-4DOtmpqNlxxO5-@WxUO<@_ps7g>CA9g
vVfg>Jx^#oZMoW!Nn9I$~Em+nzcJ_`-*!q^XTpitI-mAkg`~XiMB?0tbqMTXE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_bwd_gamma_beta_impl.hpp.F912A6BA3766E9B6.idx b/.cache/clangd/index/profile_groupnorm_bwd_gamma_beta_impl.hpp.F912A6BA3766E9B6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2ff61fc6bca588581c110744d25dcec5a4bb6a7a
GIT binary patch
literal 7364
zcmdU!c~lff8o=k12AQF!XowubILOT~9K$^@9D@kq6ghW5F`PzJbOqg@2#LfsMibpA
zmn*Ku>?Tn$#%B}=28eM(23;{hjnTXrCmLp>kf`A>h{-m|c31u9KmNnK2d{tiRe#lA
zReeWwXC@^ioRtU!aivMs6=my-D1ks=#r{>TuPh$~fk1605LCB*l9jf5P{I5sUHf)o
zSEk+I!+AMTK5}Q&dQ0GWVcom0=iDs$X0K1q>EGsg-dxlAV?FbJV(y-eYlp*r_PCUL
z>dE0l2TbJjoiP1c>h%)=XZIUVJ*h11PTSt>`e5$XC#7Bf*=d2x-kj@nxudCYLC>P#
z7*Wajt+jhU{95hW+1)atx9!hd=KNFdu{ZbB&$&@_vujI6u2q;?oNF@K?Pay0C$=PN
z{!Y6IMrN|vLRa72XBU6ys8___w$}Y?wP}p<c$N6fyF)()E{Q6-68Z9QMt#&7Pv`X3
zq?RKx*;|t*yQ|Ob*jV@7xyW={NhRI;c`7XJh2^!IJaz`|y?NMPdg*pT`Tc-ly=h5X
ztVx^juyyQ)iw56nzqZR=hi2_PnL;1F@_hWeZC5HxQ`g+yq*G3`&o0l|WO(U^Ezf`P
znOUSp*xGA0Dw+4_)A=3+`!ApFyOuY)t-1Z`z8&GdTjY+33peEU<$KuN|M*~}|LvB6
z?9`NG`Hg^YUir52{;wA!Y}R{EuI@dgKah9qQ}t_Gev5Hm_Ug#&$7}!Iw`y$uq(Bn=
z=)|wjeelNMbLH#LZ~knh?1Yz2FIazC6SzCEP5e&%<}+t^?TlaawpDKD@a~Ux<wZXh
zp8O>@L{PP{qFBwYv3<$?;Ul+ZP0X5QWy&J^LyU;A76|`<_Z%lySSJ)s&;rIxfb31}
zy?_5F5Y7UD2?Yo$%a&JB4s7QadSWjH$t-8l=sGxF2RHssbOW4jfEULm0TxFNUMzlN
z4voUy!ev}DOL5_em{hJR{uUakVpR!P7Udfih$-4Pnjp1Yt-!Kac`QMNzJ)~3BJU!9
z+_N~mI2uz)Ovyq_rQxN7tSq!lg=G!k)Ib{B2+~Gy#quU_ZUPU?&EVDygw-wJ+JYyC
zM*F~JA9!;8%`Wxa8}|NCWJaUKPKyasGNc*A!)H2Y5{6~DWrc7GXq4@m9gHc*B`1g@
z@rw>&<;NtwK%oh+o`5-=z|#K80>z+il_`Z9Af*9PIRS+lA*m6PF*iX@6XarU22C^M
zVQzt}7Rct9S=8fPG5o44Yfhq5B4JLFG>I@L**TdoC&ev=FelYDl`tpGC5<<S>6BIk
zH6{*Nv&t<JrQDb_N)jc_!;~yeo{uSoNg)<1)iTuu%hFJqGp5BBiwT_>q6|W3rZ|((
znZ;xgI<qaa37t78htR39&=5NFM0teHd~rUZa|yGA(7DudDWS6f74SOk&b`ufO4ss)
zrcjMTjf8uAmI7<|d1Glo4ZpuB)D8~qK=f^guyzQ?_x6KrKiHET)DQfcP)G-<I#`6|
zw_wgKu*GbEYy-JThQMKn9B&B1hVXdyU!OFedapf@HKEX}karad6?qi#jzOW~z+!&s
zDO3_t5{d^bRg?x`D)TGjH`nxz4sC!iENz64Mj-BO0%a3^7z#Cmsu?12{}za70X_gw
ztiy)yg9uLUF!=9vg1eDrGkE?TLw|Z{zx(cePmSa*F&}k4R9+U%$c1vd)7d8qUY7k-
zEB({oxof*l-uBdv!BOW$*Ei<R0UC9Kc{d<#Pee!`^sh22dSK1^9jaltW(swHYX`Vv
z)`G2;^fN8QXd#xAYqXG$SqBOo@R5>2dQj*|Hh{uFvJqU3<akDiGeSJpHv-ZT;DZi@
zM!|6u_>eTs%2B-4DKrL-W2Af>tjBQ{q0l&ljFYZ54q4-ngU6c!>nYq{B^Sgww9giq
z&}=|qbU;K#JurHb4Zs*MC!GE3kE@55A7m%K0fY@;$&HZq`Q?nOzW2Xi6I>5i_ka^8
z$h`87$=$Avni=e+S;F{FD$1B$tkJc?THfUBMb@7GeJj#bL?ekp!iO~)$%Hb#6OFuu
z-UKOyN`k`0;m>gsMWYB|1R+z2RfH@`7)8j`LN!5g!Z@BZk6QwUx?T-_=JiBD!@b>W
zF1|3tPC}isjt|Nu-wv8O1Pn^q=B?nh6{0wE$}6k>lK$vD^=ujyQiWo!nMOra5kbXN
zF+n9%2|=Y)DM4ja89_Z@-UC*g)^*j<1Csu)o|sTb<}35#nknS%>rGIcPn<6&t2+AC
z**)6}YuN{PK`z!@bN%tuikt&q_OXxsKE%Eci#Y*}j)3|Iq+>n`aYrE&^D#&`hFy+E
zZ4lpv>oFP~hlR%>3-|AZ#9sIl$2*fvwvO!nO(|w~Q<x;0Yo?Jq<!*t=gYqEAlky};
z#>nh&PjAMXASI(DDBM1rcTjfUF%g7JWv?P+QA`veQ!{FU;+Qy|8tq=Qjjfwg$NGOQ
zQ_ELOjT40h^_G&W>@1&$=<~RoZ9O3WZu7YfVb6S@{$qhYQksz>p26|v;Pyz%2P-G5
z*h$lZw-)#v{PyY;Dmi1@cV^T1EU^R6CWVZUVuUnq^(iz0UL)X*c?@F4$ow_|UK6Bz
z3PPtq$@M?{la1`oUDrt=jb5~SkuO(iw8CbE1J{X0<&Nci3Zv0V=at0a>ng~u;wPuk
zHSoL!KKNc8sC76Y(riM_)MIa;kpU77*r0!I*X;X#L$`!I&1jr7#Azl^n&*#|<~!%}
zIhsaG+?Mc_GmVzIF6FCo8Wp${@FgvGKZt%maC67>6SAS4B@Z%uD2iQ%AT3#iYav-n
z9@WOdZ5*FGr?ato92RhTPcFS4c=!!VYeb_OVGR+)+QFtB=5pO=c2T7r!28bh>RV~S
z2FrCI*8yR!9^`tG4Inp=JOnmFq<$kvjimk&m^*@t3YslCY)64u-!ZTqBY7Oe<2dNj
zXbQwrm`~r^blPsm!=p2@M(s#wto}f4yv{Cm2Q26S{uRiyYegAJ*BOPNQSuZs1`%U8
z?{%Ud2DbfqrQoSG?8nIb^5>d6X)`!bo&Dl!-f`8>-_OVu?+!BFr~L9#*eJWkV+~)L
z_kH!{g$s8_d!Kn?x=?u}<MHi^su{fA7ny}$Keil>n87Aoa;@23ADv}s!N%g`nngIa
zcEI8evRdm0RX@aFxel^*puv0#B)9MZZaQBrH9#Sj8zIX`=Cct{i~x~<#^CueNWuLl
zKruneryzO?r;ve2kIk9slBOByX<rVVTl;a$6(d`nyW6_+B_oAAY&{6_wDsianzYOG
z8~OZ;PmQA2M_V&jwCs58nY*h&IS@C|W)CxR#_#R@VEll5;2?W6^`U%3c$fK7z6AME
zegyeb{sbu~1wjE+06~FNAVI-YFhL<y2tlD#C{GVtJA3>;batDOGdk+9I=v`k{W@y8
zN@F_kt0S7N4;UlKqhLM?{HrILeRX3#hWXKQ>U)9e<qPaVNeeDo{5|8B5n+r~_Q`E*
zKB{xB<Ey;MGtHm+sV<yio2wPoe88$KTU|8$*{G_rq-xdl1wH-mkDr*kU_USg?0jh;
lernefbIaL6kp(R_7nozICBB@EFGcwB|6iu28!T|^{{Y4Wd)5E|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_fwd.cpp.B71B17B33AF365C7.idx b/.cache/clangd/index/profile_groupnorm_fwd.cpp.B71B17B33AF365C7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7156c186bd99a7093a7422be634c837f0bac2954
GIT binary patch
literal 2984
zcmZux3pkY98vc!O8D{3kNEyuJPKn&c48su3h?;Tv8I;Ow(ugg_F48XLn(SgLmC#MK
z<xr_qMADpkQlY8sl5|12Y__y_skU_fna(oPdCvblv!3<7>s#M?=Uw0Wf>yBE$2B3y
zJ9<TGLQF~&7J?vc^e0M57G70`pivwIrIyu%Bqa9Hw7QPzUpPU(&#n6x;U&zCyLzis
zJQ6Wv>Q!vHXX)}dPMh7A_iho%<@DYRb0z%hX=X$2^c|6Eq0+~b7uLoVR-Q`=kN3dG
zs8Y|*>OStf_xe!PwX$E@_h(G+AMc$R$?l!`dq|zma>G#k*Sy%_-!gG`^z}4v=IxjB
zRfl&r9o=(pW*bxBnb*-Rvj6JviPrefHd70i0w!mBbHiBPA2;4PzX>UfP`%V<aFeYg
zS<y(a7;5YO1x}6%v?(%ty8RKkS~Ped!MVBR>+6ax8(+zXc^ZUItzng!;(np(v?OBo
zP4}jhqGO%H;EaQsmyM73R7fT@ga(TT+oI!5lD<xA@~*7p-Z`SKlyIw~{r2u9gC=j&
zM|MxuX8zu|mHON}KYlnR%QG{gVI#szyUX#|{2~5i&IMXv$nlGXH4`yTv@AdN8Rtcd
zSFk^5x{t1|#cck3Wk2in=G*q{VAJ$0-<tf2){22~`q8I`eOiKk>XBOS?K;uIFUy`{
zslt+4mC+l0UCo5fBDFi@+yMUpu2BosmlFDG!4tN9b<Y8$zKYirm$2;RKF_GmINw!E
zCFCv40mk&mVErx1)-s3gMJiZ${fo-a8P?^)ep|y#uXr$Xq3j{!F{Rdw%L4OL46EBw
zofUaY812E{4}Tu)465dw+DABONx{=unIpzEyNud~J*-DW{Xh0gK3&PYVs@zRkJjfR
zFFp0W^;U()33X!kN2)}phu2dYsy{ZD?th*a=2F6`Qtq}obRmC~a@yp5+Ue4Z^WIMC
zl@Ij3+i+HLw&n3kw;l7DEFTB+duRM~Y6hdX+|1oETE5CE<RDkDC&hiTG0)+pLEwY)
zXJ^VwpCk%C@Va5Wu4?n(f38It|Ng65$?du`v?S((me@{l{`>HP^f#*op$6pPgyWYZ
z+B=e)X|)eHv6-(j<~wiwvCQ=YNA>7<%<6Z;&A7_ht77lrCS$*M3AZnW?-kUjofFoT
zr5`KiJI{E~jqrbEC#A)ZPuEo&To>!7U;Lc>yqZ8C-pM!kIR2P&G=OnP{P+t#C8@}K
z*Iwq%2@MY@f^97c!=)s+|3<&+!W!%qXTQ3bR{Ym&V{qbc+*X&KnpCea^7Au&9R2Z}
zlWqQ|Q*WjkENvahx+}ihe`8Yf)DFh=Azpjgl+EtC9oatjN5}q#&Er0m2814-r8!g#
zJa~JqC+~3hq+k1ytFy)8Y4*zrs~q>q(e4*@77ydzwUGxV%obgKc*3r?m>csIvVNWK
z35n7Y1Rv3NObv~4J$B!y0S~Q%AjDq9USBq#E3(h{c%gn@iG@T96B!60#yDeb*#MT4
zi>v=EjxEel#RQ<c32Fp=*#MTW3E|*UQ5Tf}twEO)p+p1O07;t&`nAkM{-LsCO|&p;
z(d61VZ5`S0Z4rW=F}H~~y46?Uq4nr)cdYw7*#MTu|IOh1v}%0@858ke3enWil=dds
z_enix)3br`d?IEwnjEIVEZG2-lVdiBrc@wEMfB+BLEDFcN?3$%&bI((QpQ_~)c+g`
zSHdHfHkS6l1}bqEV~_fSS12tc2^WHBU?N82JM-{}uez@x2m(1U7@38@IxG^Z6H1y3
zYDnHVp25sP<p_yFZ9v=VrWa<LP1|Dd$U41sHh{(>(R$Hy$dxkHwn;T=N)3xR5*@XG
zC@kVcbW%XhL}vv=BhnO*3(-XZ(TQ{g#2_-}P=F(Wu<<9(2~-cuiA9&|5iAzi0Thf|
z-joodwNDw3z=p61xDY&|P19zAyT>CWEs`EU2v^5)q=>14=7p_cJNXtLoDU50hB}wu
z5tLi5RY4C|XXyeOMHVavIpQwj(&Q*|L8O@+35*2R0EKxzUC^<mxdF9YTA`q3Bg@21
zuO>+O|3j~<s;aLJV#KQA5wb1W0SLz<Tz9S)NQp;yRGz6E1yBPNP#`r>0R>Tm=FkS6
ztlsx(ot79>ieaKDkb)xfL>quMV^U`=G%KYQOEMx+fP&3E=blwODUL&FvObv%=%iGS
z4#PId7nJrf@tHdbR3BQg(pL*!1s(}e4<Q5LS%nAc;aiW((Q21E1A=njE;nR!6p&Fm
zLN{VAdas{y*2Mm=bd*+nhXs`-{fXO0`%nQe7v_O0JrVDZ^$W}yKxtQ^>s(=-3Bn!F
z<BaV?X+F%?0!l8f>2Ww&0PCQ+9dV8XP}-MDJQr$7Bu!|v;?-T=X*amgAlVV6Nv0%o
zV8u1|!PHvM;T{~S#0Oppl%PSHBq9hx-E$<n$PsrLmkm^45uO>(N{#}|0u)f7S)c+6
zG7C~bd<(uU*az-W3G<z+;7IgFq@5645ZZr8mWH5r+zO+NnT?&8L4Am{h+GU(wi%=j
z;8KYyXx*%5tzQmkEE2<tnM3WS>3O?7f?;&8tDY+vc=*WH)nCm`G=-W4d%|A8m9Cqd
zZBHs@c~aUN_5n0VjjEJ3WKa@lf!HRpz&{(z-Bh(Futpg@3C6*CU@vIcO1SR|WJ1(f
zU!7_x?Wtrc*$VhDec-NIn%OH|>D?;i2Fb7*vUV=FiL6D|2Yc~|>3q|Lz%)F<H{_cG
z<xp}=e3VqTDES9buyhh69lm`bU;-&46CuXTp_-^d)zC10P;dw&i~Y79HWvnk%BOt!
z^oNA#cp(Z`#&VUhDynMgI88i3TMH)o>9BpQbTs@(3v}n{lU?21Jy@Pz%a-abVVcnw
vE&PA%d_yCOv5Bd<rS)PPTYCpbXBP&QW?<*ETyMUGuYWj(!}DIDv<mtU0jzcU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_groupnorm_fwd_impl.hpp.DD036B9EC45A3789.idx b/.cache/clangd/index/profile_groupnorm_fwd_impl.hpp.DD036B9EC45A3789.idx
new file mode 100755
index 0000000000000000000000000000000000000000..275bb734c77eb9a7598e10452cb506e7d11744e8
GIT binary patch
literal 7184
zcmdU!c~leE9>?<=1|(q^0ws~numr=B$QA?<gvcgfSpuSHwJwzwS)Ky6)mmv^V$mSB
zZul$`z}Dqh#cFqjJ}jU`(L_X_QqNN?NUaJH1UH_iZ_>V*d(U})dHhFnPIB&Ne!qL?
zE_1)X-<@P8Cnj3E@py4dlUEhxu3Sv=cszUNTe`Aj*)`ztQmuHrRW(0uJNNNb566$5
z`o6yBLgjY9)F;XYQEqU<thYNat?ZoZsaam=*yHuXgV0UGgD2k!`A(@GKVh%KSLYu0
zNZ$W&X>j+skYA=AX;D47ztLkz`rtv=%CB}k>^T^eSlM$`at$_hH`cCA?Df~!mtF5&
zZ~y4Lc8$9AUEzaW4JJSLO~=p8-tc5;TiNQ#-D}To-y1sPb?uyjiI#rf4Ai7gQqQ&z
zS5541n7XE`szMnm)`#Bl&a`@D)nxha-|s6jYI7$i#WrVztzTHP<ABv`8ONH_#KyRq
zeZSpzz14clYB)LHVt7SjWZk9EX|Hq*-+nc=Ys9;z^<uz|xVMDup-*QzUQSj<ZRz&k
zvP)6mZs?u2?B3C<P0CX_I;$M#_yElppSI4pH$P|h_tPz9y{0#W^0h0Mo=I77^F*KQ
zyUTk^|CpC_uxfl+q+-?fs6#(#t4{C>RmRT0FSl57A=_g9szWsi%LBgIoVDg~p@+SG
zuKndsC!2<{ck`a^T$m>k`8DKqEajK9{xoJY)|YY6=`d)^)-3+g!D7NYkN=1mA_~9$
zJn(+hwY@%FJ~!@lx*PjftjMjAJl-wspE)~eb==TCtNLBtHLK&7WZ&vIGv|Kem7Jx5
zO-Ig5G}>1GI-}d?177^^b2~~`7cEgStL^m0y}Ji`h=&AWZ^<C@p&bMc4m|!}5a76G
z^GDtf11O%*ibuOzy1M@P!{Zacvmk+nlH4~-6Py|6h0W6qN<TXbioOcctDxX&(mf#U
z!MP8leTXfBZ4T6(T~g^r(W(Bafv8$qSXv~abY=P!L>c}Wn5TM%`c-7hQe*`qDhGKv
z1mT=q4rz#WAlKn`9VF{;dj%*f0K4->P;SJz62dA0>*pp2+k|s9$gA=8t0B1>Z(k2`
zJ#N=SvL3hBg0dFBUmYmxaIS~2di;L7A#69!E#TDxK4`qHAZrCV;x?Gr2L6aog4aph
zza6CQ;EU|1L3SF|w=y)}x$A#Kz^r<lBo14Wc&~URs+8cF5W<q}TQ1?t4!WHHl3qtR
ztRoy*gPqgG*r2X!1tdwgKxzw`1d?ur>{gh~I+Ao7WVOL8#O<)C9Tu}JOl$Tmx~r8k
zor;!3W1WieiorUi@>Fr15+0KjDXWr<e5Q&ZdIH-LMaR%Fo`_VoDi1_)f;h}Go{q;n
z6KoSO&m=(-=9x?<W1cCtDV*o{mJfG)eQoDc0ZHe%<hil8Vkjgp6r<R@ctl1J8^H-@
zr4c409)Oeq;PwYe55lZLd@@atV#4hs;4%W(y+<H?1R~h>oGu@={?GpJLnx9?m8EhI
z7D=ZCq;XSC(&>uyNo*C8&X8qbp6W<-6tZPTXT~6Efru8EintY~wL&!FHi&A2X^2lk
z#7W%09b(!+h3uyx_B6z??9z3zn0GrW_XVC`*D%w&yC#0}=`-u7SIjrZhNqktyp#*$
z-FNI~LFntVo_W;l%7Wm}xWq=Y!1T{=i;Lxlp1HlyX>EcltLHfaMIQ(2<3O`Fm?{}M
zQ((1t$d1{Jbe@pCOwtAr8}RLCfEWWz$L%==n1^@(f(C#S9`kk#8pPQIK_;9>K{^T(
zQT<Vf8-)bKV;~s=PC-aoL%3=PPHjk9OE_qe$eMYwmIz1nwM3Sd$VRLq9CXM(QqGHW
z`F@;P;0HiB020=bVmdAy#MuNw6XL|Cj|L0xzFEgiWDD4~z$<LU)bOp##gi&GGX~En
z&l%k1Q|Hx}Cf#di!a^fBG=e*8$h`Qu#qGw`IWO>@&EhBgSd=ULm>E(&Ki`h+`K&Yt
zpIcj^ENv;;-P4^DK#KO~`*ThdE$7QIQt%ZRMerjqDhE+HxU$Vtv<^f%oGU<7fpaB@
zDsip`Q8mtb5b1HQ1yL=|^&qN8oHMpFxW7>w`rPq(fvbDkS?^uW%ybp|7IR|1VEdg<
z{&sd@KXbWYN&%NrEh{N~C;icW)i@?UBvWnC=+nqFjMB+;j55d!jMStWqf9arqeif9
z1bepC;#JW`w=)->GGp|Y`*S%gNz3JOj1+POMiD^~T*6tp)zwp1bE=*h#60D^WVYSX
zFD^7~-LPmab6HAD9oWm7hF6QS^_@T4Q1oUHZ${gkqPKv{7Vu*oDS9h7Z^iSlZ6Mx;
z`)>!yc9?|xtH84g&nx$Ucn|Ks7hLw@?e7EUeR%u(LA)RLZ-S&In8UWeF~wrtz$br9
z5i*aquX`YSnWAN+%m$GU>4T9k>5GxS(0@Gglndn;DTE4)B7_kbm4in)o}lW$Lx*z(
zcvRqA2_BU=SA$12&U*0B<6H|KwK&&<M?K=I@$1FY%N)xXnaUUDd$9wkdbntT#?I{$
zv%V)F`UIq~h8n&6*Y{iA4u9@VQyY6dN}`@=`vS{g-G(T;q5LPM%p@4V-v9yZB<wDn
zNv5dFf3+m(QJ6UjsjPvd$G~?C9Y4&`Yl?>8B*A<{@Lfl6#}ks)5n(!lJC1#368ZJr
zmOfz=Hd~bK!Zu6MvnR~v)`p_zxX$4MBt`!M?!SOH^1Tkq>-dT421ssz6xnY;=na(r
zQgkmw^rDoMqWfTKA8;uwMc;&&o51Dg=C~O>fKpI%+)NroQ$W!sh&Cbno((yl-hBI?
zZcM;Uk|bd}E7>a<JG`WLrf@mLmwx@(uMMdc%rj;Hr2)_V4Ul5Mb9*h}ttDja*v!d*
zN=wADe02Vmkk2aZ92hyu<L6-_X#jfzh*&rCkmLh^^=%O3gE*T&Zo+v4L?hsg>W_kC
z6tD#v1LrZ|au$kt3d9-$TTU%ut0g#*W&$5!t3&+lo%g;SUorf}OPTrpkXW_qu<Exf
zgn4gF8^tq64WZN!VeG6OqklJ+ugK>y55}{-U^z}(pFfq;m-+%LM^pQyxebx`Z@!c(
z*j&;(IR8KIWiTP39mMVEEb#M%&W?`TBPX6)`#VYwsK<JXN?+g=JS&X9JZ5)y>I>XO
zgLKUXO>~x>fKj)UJnR*q5h9JafDb_C0D3`U7VP<hu!!A<X2H%h;r3C;8bzrQ#l*Cr
zG2mX;nAJ?BA+Uh3jtE*uV0Il5twSPgjFP>dneMjZrQ8dh{VmIn$6Wl)f}~|)8JD$@
zw2#;aBVVyENB@#ET6W7l&OWoQT^Xs#%-dP9{<*i;jVZ>shliZtrJV6Q+YUd_%Z+u+
zo67()K!BdoKr#@cATkJ}U@{n^5HbWKC8@+{GC3KeFft6Ia59{u;hJO3fk!;OU&=Wg
zaVbn+tX^46P)rOKjv@s&?_FyR!M*Kyq%3YtI($m?e^USGP4Xr$@`Db>U;r-zN-rM|
z@MG*}9W7^a;sVbCE@gW1-9LT`i0t^9xttr6%V`m>B)4#}`Hz><k~c~V%!U8`4P!3a
snt%TK*s%p3ZyYLGqaxktw=G3>urtTnZX91=L)lvM1juTGiWDmT7qhN8DgXcg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_bwd_data.cpp.EDB022306AF7F904.idx b/.cache/clangd/index/profile_layernorm_bwd_data.cpp.EDB022306AF7F904.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5eeac590a67f4bc6e2f30b4751f3091b65a536a7
GIT binary patch
literal 2966
zcmZux3pkW%8~)6UVa$0()EH+)Vi+TgQwqai$Tv*R$}!bYrus+N+D&v&n?vdkMTymM
zlT?nKC^;k>rQ?dSQGeM@X&w5%ncdgS{@4Fq*Y{lS{T$xsexLVwADGYM9l%47dmLXX
zj^Dfn13?fK`6ZiElbX>Gq^kiz(vow*;(#FsO80T@ThW;$vyIPKH2tu$E5B*|^m%Km
zSI0%uJ;8Un_m?(5yU?<!7&a?wLRXkibQYw&=_?=@pg+bZH9a)o_az2jo6a!cTy^*+
zjChgT7H!okwQ(L0W^6MFs7$x=tMtlA-&-)`(i*$>Oz{W)Cdnr5?6{i#chg%ISH4U(
zjj?TV)Tvpq+iI1NJ>&SN>$RD!mb^!2(@2x4-!-e8|NV*mkw?qdE$DPknL+WrP2Zaz
zIk@Cyt?S*JCyp3$YttSqI3pP<d~e%)b9SPrU(+OM0-Ifwkz~<(;MHfdolZ%X%f3!v
zBZujZVxOOD+;sB2hZoP>&l6p=JMVMThW6##`A-k;#>JKFq4X4GY}L~!%H4PULpi>!
zVSiZW^yF81#oc~hI%931ulD{?!w-*Nqhs$Ity9O-gNY2^fq`SZSl^%G;zCZ=<j2J|
z^z3XNGKZ>OPD%rEd=IdXVh4;oMuR?S)=?7g4A>?_XH$w_#MhI?`3Db8?5gEA-3h%$
zs`w+hZ<2K#`^>Q4A&W=**Se}{wf0*P`FT&wc=Pt<pJq)|CyM>lpHa<-VZ>@2FYH!!
zy2FTGX2j*kW4cwHmwi-jpf9HgGp`4AOHSF>i`F$c<v^!~4M*o?sU&s#@gys=Z<Se6
zoKj1wva_e!%1?>NIXbovk4de69iD$8rxx2E_IawJ?EaPp31>u`q?I3WFPZ*X3t!`5
zR$t@s1M9BF6*p~%9@~-Lw}RFyeHov4w<|9@`{=wzy(YY2KR+2%;d&v6Q-qp*UM_SF
ze{nYSPfh93r2Q9mj<}l$^EQlmJmC5#%(RRhj_SyQ``YQkim|jSd(6<y)YqRz^N&~a
zXBH()ZJsInIjU>^lQ!b6AdX%LspDwd?Mpie8Cdbze;;qZn%Hd2yIg*$5nm|kXy4TR
zs4e*GN9jiEUp>1@oqTntTz1BPuD48UNkLVA6O8vrm#8mUm7CgoB4llOd?C88F0>nc
zuwWH&+RlN&9Nx7lO&oTr=8yH<Te#wyeFL>s%#E*>l`|(T%B#|5`xiC{O_KAlUS(rP
zmv?>N((vQqOSJ5&p~#Ziob1Mvcv^F7m!o#pp~0?e+X~z2zif$G+}d#a=?-dkjx*8q
zd}4aI(L~4Kk?q{rHV*FI8=|J}$k|^WdE5SOS>DOcG_uj_@BA1RviHdI!<!p>v?F?`
z!9spxLNw1mfHM}^dg#@Mn+2)jUDv0C%uiRVqL;;qYy3r_zKty&`|C0*PYrZD@19lF
zlpeE?q={oDk=I;&!44m>ozp`>MTSAJA<mF07l=jvRlqf?@a`ieL?{A+U@;`tkqbyR
z43Mv_o@`s3^g@e*ibA5PI4Vsp{xU*Pdx~+$?t&?O5-J=qlkLdM<N}y`S;m-#%0n{*
zC>q(ErOGmr3t)a)>f0E3&b%Fq3PsFBHKKuB0Q0+6uGbd9>KzPJ$lod$uYs3MDET*|
zuJLEIb-Wt^T8%_IVjT74;%_UI8lNC}q6$H1PTH>K<oHY!0S;k=aOa}nNJiuwQk6W&
z9Ab3^s$gJ&v%m#7>Akz2r~aSeP!$4fYGKL-rPL>V&)uWmIf!f-qa9-cya{k3J<$Ll
z!}`><?{+w6qcCs`EoL52W8g$mA{8J<z43eNv1`V$7?@3BQ$W2iunoyZ32{gqC1gvo
zRYFTiOO=ov$!-p<<PeE#w|JE!>_z4x7RV{Yc{YjTDF@I5n9L*_fh#1yR6D9OxNibX
zr_i+k3i79!{+8#giX51{m~5p85s(FQClf6#ZHp;2jlvLMPm(7Cq(hLsr@I1)m?AR;
z3O5Y51SrJq$^83Uo9mHFko6(x%*c}4O@r_0d;dorU0q#w2IeMW2{41jumMz&bUvQ`
zU`qnbv*1}Pke7v*67sh2RziFW{v1l6WerTI^_U`;CuWMR0Tm3K%uHqjv=JqpU4*ZY
zH4R;tt_K)we0=)Zv4;ig5H(`~g8|f%)4h%PyH{r-w*ix28sL<5u&|DN<CjuIEju*m
z-kgt~Oz!<$2BHqo445nCJ|A?}`yanKAZq2OkXupoJSF4pQzRwXpBw=02lOinC4chZ
zxnW>Gk{=lmDBs|V@d?a+g~ZyE?B|*X98v}K4_+4|9^0JACDRZBY#f^i8un6=n*pV_
zw+Y!q`RZ!7TMZfLq;e27-I%@*l;SGyko3IU$YUHr*MsZ_=ps&dIte%-)#EVi705Hd
zQwTU?U>=*dRDrzMUP{QD?X86PY`zi_xClH!+>H>uQIpZ~Y2+;7R^hfl4N{^s<f9H$
z4H*G+Qzj_s9A^#rz*9IeD2xSQ2Lf!YZEOHS2(W-7a0XfK=}2GUq$E`|GI7iVHb9Mm
zL)}C93KZoYHHW&5(|7;iDkLL(?X~O~p!N?~j=pM6l5yGPyOEazX26UbLpFFpV-a9i
zb=QR;L4S+awui@N17xx8WDgJvj)TO?rm`Y-1!jdVh#ksVU%e;r92%iR#*wu^ESTCV
zxPP?hdr3px8L8<Y;?2-!n1K>~DZHbWW-_RW5J+LB*nzGCIS>M}!LA^n?35b6VU5fp
zNsW~VWs8+E@z)m!77!#i$srm(9^ISXw(_kDHb@k_GEfMK6mCI^Ww2ryD~{Wcgh0PU
ze>6rFtEP^_6Npp_ndC#`d6>~OeCYEv=Ph73I=Q$mb6dW`TFcVe#KGL)|7o?Ex_X92
l#tTgsFR@^;Z8%Hqx%zfGR<@p6+Kaq>SE0ND-1#a3=)e80K$HLg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_bwd_data_impl.hpp.CF00FB7A122475B2.idx b/.cache/clangd/index/profile_layernorm_bwd_data_impl.hpp.CF00FB7A122475B2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..95428076a81660494669c6a28b2f0af6b69fa78b
GIT binary patch
literal 7288
zcmdU!d013c8^Gt4%QA2|GlIzKAd51B?Ax$!ih!~V2n-Z!LM~teR!C6hmZfQBv^<%}
zQhk;}ex{_zS~?^qrD6<I#u$qtGBajqSYO|r`tCjFd;HgbnCF4#{?2>mJ?A~=UCz0a
z8Xp_G$&Sa1SrcDbTC}-<;_-NMnZJt7<?9r{<HZ^9c$Ej+(-QY6y@bCdYaMpx^kyW^
z96!bn`r_rwF`josLkHd%jbG}xZik=t+KxZ&U-(_t_d$_~`w}DNuh%QD-+cY;Z)275
zQOES$=#f2jCQ&y$cD0p-4<)@YCf{25lX=}v@50EATi>)SoU~L{f3p4!?Wz6;JI<FH
z8Qg1Ys;{3atuN8<l#KqoBS$@!*?8xnku=xgnXb=1dTZOfp_;p|gf$*@+_-i9(dnBb
z=~BmuDZ%TX>G=gkHqjdYN|Twb$$KKxrG5+fe@=?S`uztj2HI|yOz$ZB<&PgvP_~1$
z!{bSdMw||q6~DQ*>DtAXLz_NYX>H|m$=q#URAG<grQ(M_WW9Unn}&-cXT;Nj@v<X+
z53HV3Z}ZM*gbYLJtA}mIG@?+;jGF=f>RR)orGa$ffMZ(5M56wx=3lC^LhiTj{KB<e
zQ<|-zw#c>_y1i^Zxb;w@UT5I$_}2Kf?n5%VT{T+sY<;n)wX4?Dxgq$8bGZf!G_Lz*
zqDxES)-88G;NR<SsP;UX7S-7Bl>t37-IMKa5WH{ZuO(WG!HrFSeZ1vNil9NzrQKdv
zCv4t(MNhi<>*{YaQm?0+Jpnh1eTqbH4lcfTa&*rv`IoYhc{T;Leg8RJexqjE_T%3(
zD|r=LN(&>Io!0Z+va@4jQ4P%#8t{;{zP0n;|9Jd4z|*4ufz?x*$s@ya?$4PsSD)$6
ze2~ysXvX9J1LZ}<75}zop7T1RF8MAI0FADK(^Xi?DxOp3I;_8&+QSH3&0X!-W<76{
zBXt)G_aYieoFz-xW`<%UahP)4b9^x6y5@Rf$_vN~!IU4AAC9RYqF@=OLia*KOFKBV
zgBVLYz_9~dFn7Y@PH@BA1&g}C6>~R;yTKiE4><K;eQY}z|Gkjx@)=y}ygljqxv>F_
z673Vc*j_YB5+xBfB|9ejVp)nq3Nde~v(z8UQteYYEv8S~1g`tct_vun1v4$oW4j5g
zu10%ZxxPW4LhX>;4k@gFLLHFU0ZEuUA-xkaFn2*#7i44Zh9|orhh<?>r$g!Fb|=Q1
z<@U?DiO?uk6ib-1!f^#*PMkv=VNSerJYi0PeFCRNcv@8ITN~GJ#wa&gE@H=IlpDt`
zz_bFbaKIF28n+NryfEGgQ-VnXp(W8Ik)R~wBtmC0N+xusn5Gaqr9vs8Gu0%O(2{PF
zPEdw%2B9+(WfD4NrZPfjmN1LZnQfBIX|e3t`p${#@BK+rsLG~_iyMY~s{B~f7>cgq
z9GybVVABjl-)0DGh9KO#7v}eZ70JH6!0iHs6tGG`I>iuV41o;ySHdbKDIW%#VRF1-
z2pq=aS^aR_=#NiYeHasR#5r7yQz&<FE*EbU%JazMT!upVKKY!>P^iGaAOH_vC@v(l
zv_o(^a2Y`7(Lo&$jC*%NSSN&I?t;)R2*cbB5#11pxd(!Kus)kB=gN2!p+&QJ{cQu(
z&wsFJ|ND>RWXFgvPT!S$XFMx6#(Ctl+}JO7^&?h!CO1yZFLOKkh`=xGYvaK!+4C5|
zNiaGIh~0BPP0G1z4nHqr1TN7o%h~2&k@ms{g95D?6JDrF$Tm}`1)N%75oS5emy_Wn
zhX^@Dk#d=w%%KYKQUI3%DKrFLLnJG~OG&Z{oK)m^Du_`5H%$tSgJ>MMY(k+au$=-f
z^XOPIMMi`gY}KS(17;drEGVP_e+>lU`D-9e1L>Hj!E74$4{_tg*t8lN>(NZ!5GsI6
zT{L5uaEN3j2$h&)&+eQmnJlhjmc1QJ+rgY2A@Tfo(gAl(FH^@(f!QgrX9cNO8}%mI
zw#a62FsAXBx0e<P_c2CS^Q#4{Wvi3STyDM@s?Vp<0@npx31l3S@4|JWk(e(gD3BjW
zki=BNtssqp_(6m$m>*0~C_j{-2z~@dvWM?^kF{<0d+Y~=oVRc9-*_=^j4``LQp08I
zqB%YC{mM<|OmnGkX(GEC>&h#hOMdW4q#=!Rs2o$anMS!(E<t%z9zpq3K0yUk0YQaS
zAwj3W=oHLlwU$+e_uE~#{3qjW5!Ml0VX64^^4Yg`=hrY(b3#sLZ0~KyR+XlI_`RH&
zS{+2y;iD9d4nyQ&NM<|Is2-NpLmK7=h-rXS%#9G+h{K9TO%T;Y>T89#R>)-g?@7?x
zacBQu5}k*5FOaa!jE4waEHR0NVuAvNfdomcBwQ#l9wH1PWWmB<f<lF%1VsoVINICv
z&QH%&-%>KGCAJoGhn8B)SLR2(IPXQKxmsB5$j+zsetyn2f!zSJh;0zw1_`X-;D>HE
zs(YRdeC&>PDEjA6k#wc$EKU)1yF&%q^)nUBqRGKk4j$~H)t9WK5~RCs80bn+Tnk=J
z3aMbF3KH3lx{@Fs$CnifsUbp5mLx5RwWNF+0;Yj;ss^>il936g850_<v0TF)BxzJ+
zQDnmoLZjjZ#oSKQ=xN)hxf(>HwGL|+<DOT*;R+Bu0OkX@V$x^;90$m|mjWUcxWefQ
zNx~2rT1tpiVtLbU+54l<cGxkukaYWWB6((rGCc7lG95FyBPESw4l=F;(<sY1i#z|)
zDBC`p)6%kJEd2?sbSXtKyWuM*kI8aKki#l=0A1a5)?h1j9I1h5R(|}+L7#@b0yE~S
zTg9&;(or*5G=nwUjb`>%)B?o9%3-b?EU;VwZVDjm8UnW=l9k}5BzYJthDrS@5UEK0
z<6u3Gj{`JwT$n!v#5Sv8zM5nWm}+qPq|r2(PGdfKyXvInYj;1Lm50hNV{0p$BgY4s
z&2E9{7Kmk6T^C_tDl&qmz;_CWONknS)%aTR<+;rPSqH_xF){4o;ljDyX>_Z9*QWJ6
zruqM0kDNh9-<_9@B+lZrb+&gP>sZLIqqB13d&ZAtpFeUYo2g$e)-GJ8l3)J5uW#aB
z*JCeJ7i=qU@GFtN9)*IOgSfY36bg0@CMeV;lneB}@*~oRzm!(YVi&n6T=v65!DR3(
zHsO-f3%j?4rwN$SSd~!4y*Sfl?UWW+&E^Xl^+HH5L||6HY6X6WqAO{+Lr{q2N?5HV
z<tj*1;nN?@RCKR#;NJMqOhsRzhL!mHTJX}6@@WX4#>J;Uba8cRvfVqgGC02<>v`r(
z#MN;oD|@<m65-+H;l+gqg~9}3#3em)UL;pDuPVL@3aB>U%cR}yAl{DC^fghNzQS$c
z#Ygh-;Jt&XYu|hAmB$|VDoKCL{i9YeD>vfyf#&HC-TLd8JIGRMsWEmwcgme056Xie
zPs)=ZFUpG`Z_1k>U&@ytKgy3Ff6AYuy9ZBqdVcQUJS*4mo=r(|fpl{jrAryY7V;FV
z0-=iJDKMG>&ZTuBZ=}ZjpqLuvg}l_qlq@+o%E6i4j^FR_Bj&C?zKe0{8iyJ#KxR(8
z+wBq3cY<kt-s^cTbCnmB6zG1#sVIM{VuS7hQ32oNym~8cgOR|HZ)`%F8u3hZ?f-m?
b*asd@fFBL<<NtmT(lr?4RulY4<H!F3^V+K!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_bwd_gamma_beta.cpp.58ACE0ADA4571F63.idx b/.cache/clangd/index/profile_layernorm_bwd_gamma_beta.cpp.58ACE0ADA4571F63.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c4990ca5cca4385d525cf150ad98d2216c93b3d3
GIT binary patch
literal 2976
zcmZux3pkYN9{=WsVeX^B7$Pz%(J;ncZexs$%Qt8yxke<D%pj@g;;BTtl&n2bQR&1c
zRwzeXE!`B+MV3;sZAB8No2||H#`c;y=Q-c=%scPz_y52D`}@K|U*9qv2=a~>X2}xL
zW3dnfQIVe_UB2!*27*>-Lr_-vnJ`)EAdk{_)S$ldSOQzxr*T8mW&dD|C@y{JRIy_*
zJAWBJS<q?o`Mt}4yxh^~?m`qce-n9P_~K*<z8F7z>uz=r+V8=OunT`K%L~}r_DLKy
zRLG9Deje}a>=xXXYaW!SX;|GjavXE@)y*BxI9DHZwVXIVo#izBHDB+^OlwKw*%q<W
zk;sbgC8p0`)c#n$%^>&iP|n5Dk@bCFv_5T3Yrxe_4LFoNL~*f=HzF?i9C&)UZWsQ-
zlxPg2%XrSZS=m+4EvYa(m2zql_ra#+T(!xu_e=CwO`UA;+p=$PB;e7GrZ0bt6I*XA
zNgpWr)oTaOy!eLrPwS$-240>2uZw99XX@VF_pr6?_M!Hawq-`?S{dl!CcO0uH$^mM
ze|*!vrPy^)-zdqafVrAmv!`vuY`*>$Mo2+DH{^lW8hpk0I&O2G_JVfiA>7UznU$K`
zu1hJboRJYXqq_d<NzC{H%4u=0xZjgyB)C`REzRqHC#W8y9CegRqqmDjG2NXR=<-ct
z*EV^kTycxr@7)Wfo&3<5Thu0j40?PyH4<!1Yv`#XO&rsWe4lka$;JC(E;D(qmz_Sm
z9BI88HQ^ZF9VYj44gSn~E-35V-2E>6;m*F#8Cq4?4OLN|`bCV}@m!aLe&PVC!R&fk
za_`2+{Y{~xy-5YT2j~vF9o?wC1zZi@D))~<uiVh_?yibx=LW2NQR|sO7b}nas+XNK
zVLEE@BUEIJKzwx9y}`nXpQn=q9o0>b{3D*zVyzjR$Hz-fxYw1~m7>2st`t9t7)p`+
zNjZOb-LBN~7kqOu^;|~POO|Hi^qaAh4z{NMbgt;oeU&{VbL+%D*(=Y?xH%DeB8S$v
ztu?2rKB~)LV0x)`uwsKG^VUDEb^JTlVyne~#BTY)w63j!!es9koAoUBm)<*Aes+KD
zRPW6g$ETiMhg^c@eKw=lzD^KN<qTornZwU2Cj*NZeTDL#pC!ps)DJL2Kla!F@74vP
ze5!_eiTs^rd)qSWv!dG2%rv5++!i&Td-Qz=hBR|(ucrHoE!|UJnlPKr_E0>lEe^bF
zT^ro^IW^~+qCwdH*OY}|COv=Xt;vC3u8x=ekauV}zcokqcD3PX8@w;EumyEo-zY8j
z)A?7^k*)V_1oRyFwZcSN-XHUHjRbY~qWyVqnJXD(IYtI8scW-V{}g}w4c$dL8noyP
zy?eB*#(;9?Y2wqy#@pE|)4D(1nKVyxF@2kNsNMMEyRlU1;hzqM_0fVHrN?D6jUB%!
z+=~9v!ELM$gcO@)ai5TTO=$W3)VoVR4ill(5CqF0nc?gLDepb<>E0R@;_9!Ylh6^X
zAehV}FP~lh7DAA0P3|J9so59>9fgRgcq(mn0phb>G76W{yo!R3M8u|eQ|jyj#LW}$
zn|d->8#K^T#5YlsxNvp>;@4N51R0{3?M$@f8wKlV>nI<i+3)j89<jZ?!-D|DAg7(N
z&gQerH!1W5uNyMYenLioq7kt*&f0Qz0b+T=TEzekf^dr7Z6nrQAt)>?auvDF8U0ok
z6#DOSd^-lpfz_TZ14w`XGbv01Pz(ZWre|gXawotdJCQ4}YyA3hvDTa4r6>Z-wqkRD
z(tOemr+ls3!)O8=rx#}qv<Yw`J<$Xp(^h%X>+P<)(O5W+7B>&jSU8cCNCn8*c(QZ7
zM(m^p7UqyR6i@{$Y)i6LLv|!PHN+)x)sQ{OUJda`yg3wXM<ga?2r7~2!xx8J1D|3%
z|Dr5`@*Rc%lUd{iphyIm%A>lfszRsG832U_(%3JHU2#al<V9o~)e#X{G?!VpeZ{r7
z3U(Bh0P{(FCUA!!2fnuog|osfRVczVVktlpj{*HVnZL9m%~eh|nDxl=j*G+N^nL$B
zN1s03TnjrTY9Q^jX4-;K5qJLlKwy#p`&#)fQz3zsKn?j>`KckHm2eKNrRDwhUbBad
zG*8Bo@jxh8cmrz#2O!`Qtv!2<RO7{_vsqwPHlnk>TIifmPAuJsZVb+B?5}xr<o>~A
zgk~BrnSfT*+-*CsH)b2sGMEI@K;U^rCC%h3_bL!tSs8FE_Pph@obD7d5jsdWXs+S~
z{+PV+?`IqlTKxtVo+y2sy7ko{;*uOl4gyWAOj&{T583?;p&dvLbHfATN26yS0gxOP
z4j=`PL0k>1`)#WV$uz_h2hSmb@og;iFrnPNdl50I-pF&gHX|m6ayx{ko6#47oPlYi
zjLJs^Bu9X~uwDy*r}Mstvg$lu^y87>eaHbIc*K$podhf)wX|b8s1QGhF9rs%urJ5g
zUWEi4fg1AT_^BZwN2rEGZX!N7cST}6YCc-|1xYx<CV~s-PQ$Ec_$(3*gFH)?B3nRX
zVX3!Ns6tWRQFEx<Y|GxA?qV`x$${a(1f}b>c3z?BqL@@R$%E_(^6a}L$i9E%Ymk!m
zCi?*Ti7{FCn2$yVaxW!J$#UR{07o(-=g`QW^{4Yg&R`HvWIUMx&Vu(Eh5xT98CU$P
zxi*VEjA%3GGc7?eKbBnA+-yFqiv*C$N_7QS0C+@tr+m^?y;D~Lxgko*hh3%QuI0W2
zSdb^A#405U`5J{7`NzQB4>4uwyR)BZgocL&hlrtIa1eq*R9mI)w=W!QAShI|4O4B`
z$l_DhAuu2@5QD{OXlmhg2t+D{O!BAs`dHGm{ptF;^9-2IE^hA2Jv_aZF_yZTJ1#c)
tf2<zM$k=p&*+TZBC05oPTRVFPr};cX8!n%rXW_SEC0Y>VEkuc+{{Z9<R?7eY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_bwd_gamma_beta_impl.hpp.672AE63374836A4E.idx b/.cache/clangd/index/profile_layernorm_bwd_gamma_beta_impl.hpp.672AE63374836A4E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4bffceb08e7a69ea571c901e34340947a6e1fb4f
GIT binary patch
literal 7360
zcmdU!d013c8^Gs<3o^r<GXo);I0i$A4zdf#jKd<3vC1l<K<pv#mEZy<Vwh$|`B}+s
z8ueA+9#(?5QDQVCYN%8`IWR7a+6N{v;6mS>`tCjFKmYL`=6T?`zx%%Tz4txmUCy~T
z?xnD>V|F~=qU@JSi*i<HP&^*bjQJ~BU7UX%c)Uzw9<Q|V#0L>OesZ_i@*w_8=Sx+6
zXO4D!((+D*<9C(99)qN6%{8Z_qIA{CZgJw1VZZXo-F0BUWM`A3@^;tlTdlk2IW;T`
zI$<cIZ}?t+m3sSEziZJCx^p)cpM1BfX12!5<Wcw9^k&<{4*$drftr-colpF`A3Il>
zNp(?!X@>9q_|29tr$1JISYfvIL~2^3&u~*n*_IB&LrLL+xL4Ml9bWg|m)noQ%EP}j
zw_1h1{l!~u-dA4F|GH+ejk;jOax`s^)f+8MLxJWG&)ToAD>yuV{;#US`&38%WbN{%
zg<;jTqE}uxZ=JC6)2h$Pw&#`Ir|Le=rgG|M`=QO$ciw*qD^-Qgz1gw3#i<s@8y*~s
zv)y2adMiDy?1d{-?Av?&Z;c>@Rnn01i__VU^XR6&b0MGY?<+EjJgE7zW7U6Ooqefg
zv{R>j^TqCO1&U;T<7wkjyP4xNpLa@a=ua74<vF^wX5Hnw`m%n5r9S`OrrrPEi7J2f
zpNB=IHdap}e?0o>(WZ~Ojc?R%*5#dV-W$KO6*iv?C`sD<1~Q%e>-|#U+T{KUl|r)b
z(S_FsMmj6LiaF*Py`$-(*#qIe|8z8!ZZy1bx$wNg;fPP8`Hj|zH66*L7Pc9?RFlqk
zlW(@}*%+U(npaX*lo`m(w{_OYy=Jw6&cMLTh(YFu!~(H7kN*!`mxwI(h8VA*d17N8
zvN5um`}aQ{-vD@q6d)+h$t@YRVU*KP1)cYrBQl`TFEHa5aOAE;*I?E)aAw)i!?bx%
zPiB=ljpUB<IqYeM7ReW5QqEI)VhWfSKv0k@$Q{dqWx)hx&dVe;v*u>GVa@CX*-A_~
z0XZR<au?(hvOJ$WIhIw!tZLGs8j#e06w7PDz80J?*MVal5T4b8v>p!tjT*qA0cNxM
z#^+Dny*Kx;z?epHv*HL>;wACK!zb7$5RN4}CeCLK&?rfo<c(>m!%{Dngn72U`TOin
z1BE8Qd;%<3fylZq*!`#93XCXJ4H4B4$qFb`125G;IObYNtc4`Zb&y;K%P`kNd_5$v
zERH^9UvzJ?lyN6~Ryg5Kgd~D+C(=HWa3{(!if|`d8cnzp;}FBSBW{xvdF@)<%LGrj
zX}E;FCygQm5i>DG3L~G#6eW%#CM#ML?SN%5D8?RBoM{|kGhPr+*h~;65H=IViG<B0
zQ4(QuDOyU{OfgL%Y%UipCv2_|t{`ltic<-jX`(d3W;#meY+7}0s&4(Q_6bd)a@%q{
z_VF3=Ea%3Jq2O{ZzA4lOwrxP@wt;UOEWp?HfMpL@lkC+4+?-HI4UuYy#_~b17z9hq
z8c5KPK{5okL*)I2z;_7W&-&^SlgFQb<H@*?>72=>3Wc(qvbe;cP_}0_H}w?CnV;i>
zZ<y<z>wziHHIEA}eMI|IgD;lW!2B8@uB`>XTKq5+ssni)D6oD#_}2qh04OG57c_uB
zE4TgW>?+<JMa~o+{~12R>l$9zQTNnp?GSLZ?GL{Tf+;ysjyKxQ3B1!1p6aDr%qMp&
zSIPE2RndFlcY^-16bqoyX)rkrh>Z#FktJuZ*k2D}1TLX2Tp#n7-2b7#IOD!Kla8od
z%xY7p9i;8>0%jFhs>t+HL4XQ^NO`gfQZTE*T@73TrO*Jl50I<@cMZu}kZQ^MX<?BT
zLb1JJkPHJ?izqY(Gsb|cRC@Ls!x=y!9n8>?@^LU9$IXR8<1l}m%)oJoABRMIzezBk
z#QJg>Z;|ac(*%YzQ)k3#APVLHhzCg4fLMb$tYgc63-9IbVTQdL_|+g{?-12>A@=gT
zk?)zlcLvPQz${h}*Y}Cxo#SQ6Q#g6!`JpF@a>Va3PFL_NIF}Qm&1VnoR2cDTG}C=1
zSM6vd<I6ZDCOr8r1o`v*36cxt&#?nVBL!bU$dr5~K|%Z=f<pKq93?-h^|*C>v-dM!
zSfqY<cgMea(r+<tZvwwfxIX6>m#mE$|2%LSjWVfBA$A~($|5M6$|fj>${{G1$|Wd|
z$|L9un4AGK*6ONKWv|^2m!23<NaiFXdR2%`2-n9-S}t|GzwPC6=24|cYR+C;c_6YV
zareb@%!9T-Pzx+!1vJ_Rf%_l^^L|*gAL1|{fUpBNe`s_NLJ#6K7mW@<$RUWw`kk=2
z6PB|4PPpNlzwh|VPahh-GnMRVCN#t{D@-n87lQo7{shUb<y;Oip&?cfGNo8aP>?u?
zpb&8gN4u=vvJBd2v5`s23ULLuy4j_BIki$`cbOTjE>L#irDx-A+2DrGHNMY0ls*c+
zu8_n=2&Zrn*t1O`y1(*a2{ULan5zO;b{an}jG)3}w+<TXo6O>NJem~JLWCBgSVamA
zgYz)V#jJw>9ocjyz<Gj{PlC@R__6v;!_(*7xFdbY1XaFOKDV@_(aPy7ZCNE670f8$
z`U;H-?F)&G(p8vq6+bzR`eAlIxZrEm5U9poie|c8+yIVh8fjp$20QfWw&c2-YfjoR
zJCBrEDa6iWxn#K;HoC%o1=l!fl<Jtut#WCUCQaj3&@@VSNavQHhvwW$Tz)^+g`$}8
z_EM2cV-<v}$o<neIF94(puQIdjzci}07q8zdp1>x%n^;s`Q=2NX@lu)V8g1>%<M_p
zflCtoG@GelI+m+JrUt^V0gw%ltO1#Z<RO?oMB3MaL`&Kq2Ag5r(rKpUTaE!Skvg!{
zkvtB<aa<Q^Gzr2<%wOGI|CQDDKU${b!tq$xuF`{nxBHpNZinD@2xCWGpF<0^WR8r1
z*BH6A(1E`W7P9)s&aU=Z*692vlcg>mE?iG;L)zZ0udU=gb-me+OfGaK--w#Rg{tHG
z%gYYPAKaXhfpF66_6X&=EP|Q8*PLGC*2|r@52SQ8x2G_T%*DoqYy0OeT~w*=jGlhx
z{pyDqrGxxZWW?N0AyvrO=u#Co$3D^&l}zF2=n;opeI&Z)KZRX5FMWMmr7~V*$`thQ
za_$w0zJM=jC)a2_An$<ycJlOFodh)`V;%&%LA(dkcefQ9cp1yJ5U(Y>-!Qlj1JSN@
z@S+YPu>J(NPmuCSP)_1L)T?mX5Eo-tJtgz<;;qhC+XMP;8&b$a=E1EPDCFtt$+biZ
zg<6Di4YTj(Zx{G%c&>^`<*nen71!u1lH*1{%bxFfYBBB~ZH#-RcKch;JiXZ|1Hs=f
z*V$L{1Ae`Wbo;GgO3u*T4-Ss)mi6ut(8!f?6|gSS$c=I%$enU0$b<4A$dmFU$cyqK
z$eZ#e$cOSF$d~dZ$dB^l=#R#>Q*MXt9jD}oj@lN+WW=ssMd>SxxE;SOqM1dBSWEI4
zn2Z7Uwuxrm)|lupkLOZ1dGZPs(@j+1pu%4^CVuA!m?a$9%B099u-}9ij}O1DIpZo<
zwK9VB?(4Y{SDaIrq5n)&Qk+#%pg*AX|Ni+2*#qW-Qdjqayu*!sKbnZ9@dc)|(1d4#
Wr6PPd4Ic{d;s3ux=}(yAv;PA=&V>R1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_fwd.cpp.FA9A3EA5A9ABAF4C.idx b/.cache/clangd/index/profile_layernorm_fwd.cpp.FA9A3EA5A9ABAF4C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..007049bc5011bd927a34d326fc7425d5eb781bcc
GIT binary patch
literal 3112
zcmZvd2Ut_r9>6alj6lc~kgzw%3IT#)RBRR_1c4ig$et;qAVdu+0*WXPq)3$_!}qXI
zL`6Tu;s!^b3bkMnsSA{$6j=hISl7Er-wE-3?|xs-cYptJ?th(=5I+vb1PejT1V2f7
zqBs@}K@bu7Ws9@E`B@o)CNK~rscjKF{&tj3xZk@%)Vk7+-zRp@H?BQ7=vH0QI><J{
zR2%pe+J+d0biAJQJlnW=q(P+`O-=iaQGMgWrdp*0sKe&u4-|{$6j-Mw$eOmh_m8}N
zU4u(6lgv*2bfd)l*(iCt$C_=i9&0?kQI0*|<u{SqE@fcOzxTTE*C)?4J3eQdy@;M|
ziu3yC{vxdUTq0d}p=R#f#K%82I(H<FS&sa=xOk-M`JvO!r+4+!ssuwzZWZPHN9k14
zu#Rtclxvp66_4G0Q8FN`({D;UF{l0M^xLyZYb)9_)`@QG-ZV3`nCYcS$_M7%JEMLq
z`B2EOL7j~FmsXd>{K1`JzK?}^VOH_v;`g7|NscFG(Rx_wBF&v!jC9iUbPJv`bxXHy
z7v?U#nlNv;y@k(JcRkiv!r6Rq>y9z=<X)>3^;EJzg`)ZJ5uX#!JeH6U{^Oa7goMt!
z2Rfdwf|{b=FEMfM2Jx<LXSn<NkN;`=dPVZBhy5=e6cHw8Zi@{^+Drl_b^ENAeA~nh
zzT<=KZ_@L<b7D?zQvcavreEoUe%B5EiUW^qlC*0oy6KC}p)G;xPlz_yaIIzxC;Ubk
z*K6D;_X+*Ma9QY!@W%XS#HkwCnA*4yG2uPn+k)>k^)b!a=aEk}X%TjbaNU_J`^l}A
zw{lXBUC1y9Pg%@7pUiw-tfuVPW4g^ZHD>Spg<g%nLe6bGC*C<7<IvSO^(OXhM7hfK
zqYfGtqTTkb){m|h{KWWyKWjG>w$P<doV+=2Gkwji+wx@GMQ*ciWK?rfn#zY0>?TS3
z<%i8;+gGml%ATt9;t%lr&NNrBCRZOGoUM=EmQ3~+G=48Mi|pN-%hG$b-{wUdCa=kA
zb4}I9dyBjGET(k9#_skF;qzCsgaiFnWqjHzJ=fN_T%FM)?<%YA7jbg*ZCWZz7HUuL
zw_6x~_blORNwufDzu|j_-GOf|I=qX2iVpd_aildnL*4G&ovgm2!pVY^YWwk_@iBo`
zC$2};`rY;OmudS$9xk>-j2yfo%w^2>tJW@bp3vH{<z9UJs__cR#prxTr#6es!{Zr|
z*VKG|wP9223Qo1CX^+2|Nxoclp6W%P%^J$v`qJ>e-;(NEyC_0s(Yc8?eyW%1kIjeJ
z4mk~}mQuZ*JI`%$pUsW!4{8eg&~@152z%W?M!AR)S@)@)<?xi$t#R)u=VoZFUX->z
z(Pc8Si?J|}S1HLr50#BHjCz0k5HQfNAuwtK^Ku<)Pdnqa<@SN*DPyaS;;T@0PI}xD
za*DkRXG|~k94%2rt%D#~OHE5pwjdhWv++2)lXy1G9t%YyYDX1EJ=p@(o7OvB?~Aar
zB%mULA=ngSN|Y@?eF*b>PgAl=uND-6c;`WR`mzO*ULnYH)_l9IHYJFFg<=pj1I^Hu
zEixUl{Gu@(Z{OHTB%q>@&50NyNwxs>)V~?r&4Drb6jb<sMX;uZru1f#Jqus$%I*!U
z3d5l`BHnm9evNDa>a4`1Y_Tc?sbv42JL>pbuo4>PdGWkKNJdjv52(+KZBW9(*7nxU
zAWvgl0i#&`&ag5Tj@6B|0M=MotSeTa9CD5c(3xJ$5>4fM6)B_PSW>Ju5Tju+NvuFQ
zxEvzTc^J)gr>e%zsiI*jj!FPMK*P>BX9c3+XbR+lb5S5yoT~!4;oRf|8jjNVqS07*
zEqbjEh+v(ImX>HUw;qLs1*-&(K#YY&t3+}N#jvlWClbn#V&Dz%W}x6$m`Epjf#zXh
zGJ&iE2<DTl|2*KOg6xa8#XBv#@bEl&G1j})-HfZVj#9<KEF6on{Bf2$i@8h!1A)yl
zg`0*uER)DUq@YVTO_%G0K95%RZt1v)^jkWQU_wESk*Y|C983uo_Ez^c0Q*MP{&Hn_
zn*5Ld@r`b7Za)#iNJiqNI8taJS0s9W79aRxVYWS+x=b8<jskJ*xeDZK?<=PyQsKiz
zwY%0RBwm9&I}k5I4tZY#O1?uPWkqWxqp@%pB~0#%g(D~t3Y0=gQJ@S;hMY1{l22Be
z4boFUHY6K?6fz$ie{!V%`xHb>(W6j+7@P(aN_N`8M4{mTTmT-Vn0>tOa>c%wT%^^(
zxL^_xW8nmHf-a!K@`LU8YxnAqV<zneDEr=5Yy(DbDFqSpSMcQ%P~xv#IQ`w{)reT}
zq?a~S4rc6_c#K4e=i`IGP=Vqop}yF=J}m(4A6ztoSkrNI`LSq?ctAT^iYFm^Q!!L6
zFb{2&J|={|zAK2e;<$kLk#Uh;TOe*E3s#;0eE4jh)2NAlmJ8xbHYZzx>Ye8tlAQ7x
ze}F;4W#M@sT*OF|i~~kU!Cffs%ft#|g@TWwVYV~dZJ9XE90lS!a}~(f*;j#h-aIz&
zyCyVxY4NiD1Ja*xr*Kyw#=_}N>2d-AAbCpfU(jjfQY2d&fHCOSmyBQ}I29G-3bo<U
zz)omb$Q1f5QzSD|fl`^N3Y5vrlv9uS_I<m)3dJLxpzF{npcB74dIYF>X3t5x?1Ntm
z$~Mr+-PV8PQ;<~5#It~SbZ=^Nad3+=vI!o8*8yVipP_{LVuGf#FSVbLSPvuC6l01F
z$QpQ|HDrGs%h^*16g^;sh0S%%O+YQMFwceO1(JSy@RnMR#qbL0^e7BQR$^MLbY!x#
z;<H1gFCo(9??(V$AZ57Jcx+Drk>jemUKkb<Du86Rf9r*EGid$tYuNH@e0oCKHwf~Q
z`O0V&RW)^tCRU3`z~lT$9F`49!=Jozg|;5W!_)h#H9l*<rs_C&S*%`V^8ci|28Kqa
mX6BaGws!W8&NLS{cZM-t-^rD&qif|GumQylV)`isLjM6*4=-*2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_layernorm_fwd_impl.hpp.BF1975AA887FB0A6.idx b/.cache/clangd/index/profile_layernorm_fwd_impl.hpp.BF1975AA887FB0A6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..71db8cc202906babc8ca4b2206c7ee07cdcec9ea
GIT binary patch
literal 7462
zcmdU!dsI_L8o=|77bWp#h*1#%QvnSqD3JicfCQpE%1gsbMY}@lV|@Vf=!(`_)Fo1z
zU0Z>mHHy`q(^{=UeY8GWAAnYID-x)%-3_)rs)((iSc?|+rtQ5m=j>k||03rk=l<sV
zX6`(``Mz%^$uCDoE5?XK(-*#6l)b1hixG)L!?<rjVgAy)AQI&c7Kw^1UzW|>VpYlN
z9~$y~nhrMCA8Ba+xFai4@BT%#du!P(t+j6R{hH3D=2<-_Zv@^~-c5Ipi27{nli!pN
z4=f3tn{e!~$#qD&OYB=~<>S7((3bOl(y@n%wa({Dhn>3r{g}pWU;fk<6PPnybuM?!
zGZo5xW2OJHCz5mPYI>&K+WX0>Hi@Aswe#ZBgf{ne`EjSk^X{8xm)+mwV|d@|(AoV*
zw{95m?>D<TTND%am%8nipHv*4do8?gPtBbVS9`w~aK`t6bKytri`RGi?T?iA)}MW2
z`?6|>ij(&ic7?SrT@!9yA8slbTQzR#Sf?dN9{!qWRbNfCW-V*juU>z2{N9i8KTf=h
z%Ob<9kDSiV_s(AJc{t5;zu$K{%Y~y&*KZtobBU=&F0BqIozJAHH*L9oyjrh`bWpzL
z;!+p)(wOa^CFka6ezy6>qPzcT4{lm^t0dPb-LiF4P*Jr!WWDsxRPBSO#oG#Zz_)we
z{#WCUp_6Vl@A7qvov|dn?)4+*_4^l$$<W;%l~HdjPj6Pu-EbhJ=&h_O_hln%gEAX6
zqaIJ3(wJ4*Up3tS%(2kk++7z=K3dp2&ON;MMw<d3G+DiLg>`E$N+#s`*qTL!>K*DW
zu`7o69_y}uHW@o_*LxkvzgV#G*sq4i7jt`j7P*NER%gE&!L7W@YY#g2T}QtlG|YiR
z?t^6#nOr3P1D0Ht%kvjVm#`w)U=en8aCQ6rM<hld8pHsG{6&ikX1a393r<IVqn_Y2
zh{c`Y)(PW;nz#$Ry1<Wc4|w$uMtKvjJCri-0E;IKpWx1`4Vp4^U)AMTO_;^OUco+m
zF-MWX(S!^W4gQ3ZgOfuErKnRv2&GO;r8LuG(h`Vlw%_b2gqpy!3A_ngL1_hF!p$(g
z872~L0nZj%zZJY%f$DJ^DBDQ=!QY&|*Ez9T!mv0>5k++=+A~^B8jbOY2_O`!h!r%Y
zOI$Te_q(3~23I0^B^tr2Ik`0JRaV<_hQUpc*aS(ufWcNsvqC!IW=Ls<RKl&0*$P=a
z%aTreWOr`#;<~6)=!6%uIKnf6>SClvB-KT|LN92_jw`a&6|q-Xt^tW|4Bsx-0FLk=
z6e*1yOGq!#Q<_ndD2k$SG*vG~8bj5KmBdn-@sfCo5^w@lFHxFE)k~5j37VtM6`K#+
zw*AI3IL9?dFhY(3asv4-a-`3hL8um7Yk|_O1x+nXCGvJ~ZU+~d)$I^OxDVp`K(Imv
z_d{wwy-39<t{74B669KfsP~m1O$id-=W?ya@!3b82XI3$C=G)7GB|m1vam1=PVq~b
zM7ofwOr<o_Les*CEImA3N2m#=Hi4F~6{cAsoNzOQHN!N*Eiko()^7z}D?|`^8|d3$
zI?t}w`aDrb*rEYE|8BuQui7zw%l7AXfS$S|$GbxsB?EHd#@#%AK@#|v)aM?x|IED|
z&UwBEp1We;v^q)a>WmRwyto98mjLrN6!>6P{kI<NGdO`#uZ-f0wTiAw%Liw5jo`K~
z&XMuO46XyOIv7vb2F^BmyV^i!gGgFF#|9aM`#{wPf&;W~UsXTN#Yk0*D7g`N84>k9
zBbshRLN^&~LJAWS{1b!AkXsoNJd~Y7$`G{!%aL0-EiXsfazyRl3glKn%gsn`Cc%Qi
zW)x&b8ZsU;N-?7}!WJaA5dBbJ(RA0(he){f>H}FHD0oGd8@Q~W=3*o(MuI&0RLM^(
zIu}=Q%hCkWCK%3FOw?@6b)IB;mlKQ)8yh1OpKmm@OnUG&7v35`-T*K0g5>7CgE~&G
zo-=@bM2a}-%j`w6e{w@!DPHNsk1RDw?tSOIFoz*5R(L1`hs@crSScv6*jwyPQII%@
zqF`yTFtIGwh&7Z<E7l4$=jpa7w@+>ieBm77mGXOAaxO2p%?%|lBu~h?_#uJh_dJo$
zE|=yP{B7owk0Q7c8W@9=*J5!plT1+xlR{A{lS)w<lSWZGlTJ|sI5xmAzSX><@GI_@
znttOfNUPEciCw`~H;=b1=XY~G&Qxc<%xf2X(sXKTY37?;u_yMF^TiuJO~_8$dG*3j
z7ViYjPM~tmT@bnp;&??CSAn()g!Wjx8#KFV{SQF*0ZGPKya)7qz(DFB1kFKO{}6;8
zqT8<qZ8hC~4QOg;{mT%08D{hCy%RU+@AtR#huDqDN8!&GbH*f7jw0kO^QI_B7DQ37
zOR%u(IAfA&D4ABK70C3V`M<eq?|#QkwD)LlFTU}LQJb72*Ns@m6|a=7^c0GJ$$Vvl
zlY0v{(Txz^NJePc>HEXlbFXV&*d}-B+rt!wc<BK45LKmNPF>4-3%E(L!9*KO=4bDN
z74b}*!T7^qd*F(#BeTa~BZ@boL|)MzNPJ94IG-@M0_iFc704=)PbDq4pkNCUPD*=z
z9y#G=hgYu?i|3A<>&iFF;*2pF!dkL;p4&VjR<XDPUh1Hae4XIkNsf5-BVkY{$p%^c
z0MrjacqNOwLD3C@0oggQo21e#egsn=k!+L2J)rFYAv<O9W6(VYAvL#$sPI0LUD`ub
zY(EH7#NuKUUW|n7^TX0P+i$)8l{<IdiC4r^TPDFXfjaLbdL#<roW)6sBti4ggxhKJ
zx(t4t1=~Pvqmz6a#M$U1-;BnYk&>TPdlZf^BR$VG^IHS<n4IL?<mZTUC=XW)BWqza
zuf}pVpr`|CIczY@1|x~Q4}AMT=%0N8eEVrGM!v;}l9!;7C5UR@h!jRdwQoYBO(dCM
zxn#t-3{l2Wj-1PBu0YNeG@Fssj06wOVhfU52p_$-=IE%+-JcA|6!?qiilT!NKejs9
zcc#9McxL;~oMxnVUKvuCp<sTM7+j8m%E`{N?^JCC5)M&c)E5TNwfOXKzCbxyDH!5$
zeD{iR<ucK8XU-?EW8?WbHxmc2dpz}Z%iK>xAKn^}hH^Up#!vH~t(nEe#a3``C1;Iu
zKYp5Ve&2}<PM{pE6r4j{+4zgszTfD|t^X>YRYK-}q3LQvL&u{tFB~o}xfT5&F2WAn
zYc;+a!M+;u_Zgnv$SxSbW^!3J<JwcFPVE35!Z%)Tlx_%5apLmNoVXm}D$nluXVsCb
zI~KP?XgeJW`yjcG+#=aM|NMT)<b94k^d%P~;VO{jd{l}N%_edaQkjr&=f!e9DW(F&
z6L}?4Rnl?`3b&8|b|p-?HhHGIc|dN^tGCbPp3pV_$a!-AK!56{PNh)^$qIu5rUnR`
zk-_RPwcxQC95OsaxR3o-anhm9ckJcocGRtpEXjY}wt3SF@9^(Ku1x>sfC~)B)wox7
z@bOOHD^+3^`!c=~Vu5~)A4QXxNfb?HCR60k_*0}}R1{5Nrce~X1W=@A)B<%|j-U3g
z_82!H=kSs1ikVr4!aN7N-;mXjkYnHXG9wahr|lV?V;K^z$5<{maV)2~0y$O?exlB+
zjXiii;{RgVlf}#}QRrI@oWKU2HgYKV^}bj)EVahSMYBAQJR$Y$Jyv;ka%jV0u6Vg>
zx!~tS`HNO$+5dtn$bYS1x&7hKKcwuBQv1&zA1D4nBpN~<9myl!U4Ccz-kh}y9i4`V
SB|}-Mqew!eL&+mc9{&gO_SWkF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_max_pool3d_bwd.cpp.7793240AA812E755.idx b/.cache/clangd/index/profile_max_pool3d_bwd.cpp.7793240AA812E755.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2c523db25726c832a5f9d52d4bfa2d83199e6c2f
GIT binary patch
literal 4238
zcmZvf30M=?7RP5oAVh~SSpdV9un7TyDm%z#*hC1Bkww{s5C|X?Tq5YR2qGYYtgj-V
zEUk*7_*~kmAXc;#TM@fJE!G8Oai^5}Xi?sPy_xCPH{Zw1{oQkO@45eb&Y3~H*2kw1
zg&<+#T6wx;OFR{VASV8oy(KGUPzi!kmOzla{z7PahK>W$<!%!vG>ENF+@QnfRocen
z+Z7L7==Jq9oq9=q)}v!G`^nXRRqrZen+|ofDj|RNj{mo4R5-8~-HS#HGHRG5TXKK=
z{+2qe4dq$S9}NF&T<gf0(Hn_7|1&P1?`aHp6Z3IEIpdA4=~uaB$2fzZ#hx4aedLE5
zpFNm;^z)JMtUsP?Htat(C{5VrH#?bQe(Xs9Rng0?;ofPT->*=2Ub-?24NCUwRi_90
zD7#qaMnP+z)I=)l_#cQa_Y7r>BE9(*Cleb4?Socax%b>NqlR|Bfv0MSYka!iowCO-
zkE3t;oZgKMy1cG9(ybG@^q6_3!C^1I+3k$;CG9hT_O%re^M7iez?^z>biASP>2tmz
zXC=EhbNc^{8^{>`<Njs!SKhbN^v;hnU6(y+`9Z0mv3w_|G;I$vKE7r1-n@+_4tJJK
z(9f43I(I}BCAwNFD0{nS=B(aLnsY=}*}%4Lrh)h_Q;=qGy6oiS^DfIZ&kK#Fw-)tR
zrs;<DN#5v7N2-GL-?m-n^r#FQ+|EH848vsObBuybr+dv4>{{vqy#xE4)lN02<!LZ4
zC`-LGubE4oGS$cL8Qlz3&wr5p_3+91gYwavZf|^tZW?XR8F^N-^71(9)h;hmp?N}e
zHdS`@)lY1bm-cuyKC^93xbJvj#f<!lr*MByzJaLB+pA)*``)c_x73MsGE?{I&tGq9
z>K%V-pOSI*s0<BlOXXMWk;N@-?Jgd33jZa|x$DJPm4o4y&XJINj#@{PHtA2h_62G`
z?Q>1A=#Hgd8T)C?<s`2E!<0|$x<rjU!=DP*KS(MW{o>=9S0d|k7wOfTr`HBloEd++
zyk*Vn1ILdIhR-#4Yum5<YW$h0`M#W+ckDAl1g;}_PtMoBNL%e<M%8RCx16@&rWO`2
zv#Y(`$1yuoTQyR7=0HZ@)YL>=&;Gg^ms=%o!-^zRmn>&HGANVt0aK3h)LxOLq<o^R
z^ps{&>z`d_8|Es5w?7^n7??PE>Xdcl&#BQ-zZ~ZVIF?RIzsuX)<2|o*Uv+<JeeDwq
zEf-$sSj|4;v-`f)Hb3lEqcRqDT{By@S5UUm>S4|J3+w5I>&K5L>Ic)^w$Ar+t+xkn
zZ0zP`J-k}<PHB=iX-wC?b$G{2Yr)LU3z8PQSLR#b=4HHJd*Y)WG|=wvcE0j3=<AQY
zm%~4F<{3^=>}Ma8o!q_7JiB+dr*dr3m%rZI&==wOo7&(v^Zel<!FQQh|A|B9k3-*8
zzWVLF{Kw{Yrrh*aKJ5F)?7y4)S01`LQM|o5?b=K(OVipSa!zfB?2A??dq?`FIXusy
zeQ7nFkxhpfl#NjkhT5a<i?;>K!lf+He);y$=L{7}Bm`j$8bf>W_TC6V?tcD>!+JMR
zloE^oT8pm5TD$>sU@<#}c_7dYrNrQ7Q@Sa0@dnIdt1oj?;#3YIP&_`l57kF|@dnJ8
zOgXg2&E*mT#o=aKWn2Bl8!#t`WS@>smYb+S(YUz-hB{!KzaG#+6~nLRH=KO82F87)
zP1*3i1HzCc$U=Erys{)o*~jqu1~i~AS=>9=GlKC^J{N2e!30_YKG+e0Nq7=7#Rjk>
zUJ{9<JSmBidC4S7;iZsB#*-;XDJ$2->H3&Ej9}Z)ZCYS*;0|TU-=s&d1l0r`#h~ng
zT_aZChNw_6iJ3$Ix~W)_SrUn)W>OL*n<bMd#VmzHGBcTi^q+TEsJ<NCfIDWn!g3|h
zfXB#lDpvjG5#HSfv)f=kSk>TCR`T<YRu{rl+`wpqnZU4I`C_$hs@{4Ef@RaPHU2sV
z2*D1)v_r6ho%N^sQdQ!o5GrQRutz~0P%#IF1Bo0NjwEtoIFZPi;Y=bIh6{;SGggzx
zmElSvH-;OD+!^j9@?dz7$dlozAP^(?5iR5ZI5Au<wxuqR3p^>5#UucOs8}K^kqs0I
zwoy|{-s;<gE97V6=K#79EY~8}8W4Em%Pk}gA!7&Z(E*2n#K7p@o#_&^ObNkIUDOan
zIf7}~Ypn(@BN!LuY61dM<1;Vh8!>aZYPbPdcD;yYI#ID&nTF3RK&^>+MRbum7=&O!
z^dJ%i(}PJALJuL4m@Xy}M#mHsX4i0YQ-f)Y3NDALABULS&%|#fK~d&W_5|hX<tjoK
z!AfAg5}3qwuwEU(5gzxJ^>6FCgoo<;3IfE7uJ7r5^nzXT4@0cHJ|nL4$_O6yA8LH4
z3zm&NapK}hv6B?{lyAjf1q4N~DAy=YFbKh-dC^7$#qeTC6w8YxQ5-K$LBJ`P0t(z;
zCzHe~_-TfA$CR?*zHSG+0*qW>s}9%=a3)3mZu!#I5PUmaJ+3}*FLS7Exbgcc8Q#t3
z^0h(thjO=m-M+hT@oFM65iJJ}Q?VeEAQA<e1d}MlB!om_6ETS}6HGw`2dgitU;C~B
z7x2C0Aj<Z=_SSXkFV)5kn>m}c!K(ZHl?q;z&pYDX<l<f0T6;fZ=j6}0kEjR@0#QiC
zn1M_YL98HFFhT4PwwNGJ04I<jjQ|Y=9UiyW6G1P(#?vs9mZ>OA2v!W!iitwil(vrQ
zA5!=jmruwLD$2p9VtTcv@6)n{V$6~#avLwktRxaiSyB=uvyw@a!b%~Lj3rah*{Gpa
z&86xrTn>A>y&A~abG06P^zPlxf5hwg-PR*~oh+M$lQ-g;fFu6gr$&C-V|<8>%jbm#
z5Hc{mnCl2)d9f6tP%*X_n<O;Hi{np>)bP?E5qBw<0hYpJ%ZBejP}G_zA8@!-EZR6)
z@dl(~F~%_@iZzZUQJirci6jDvq5|N`3X(n`=9CJgs|ZRKB-;{{B1o|&NG6ae=vsvS
zQ{$&ibGWi~utgoLnDEPy12;x2$_wylNT;Qfq9KQtqo6(=`D41Aq0FLUd9XMSjwBY@
z1B>^-;RNTy;(T}m!3D6mfINOLEZz%;5%xk@Tu9~$7^#3Y2zw=rR1z;)Dpmy}Rj?vp
zs8}_ORFlWoz(@^@62Gs7ky<hzg~dn7{xrejCbIv{u(+A*e+w*bA&+l`#jRw2KY_)c
z5PaJxw`h+WrjE-j)D&ui03NV%UPp7up2Fi4^*}vAVEuU2H|P7tcR>qjE=0XRH=gD~
zCW%-=7KzwGHi<Yw4v92`8VVXZY`W*sr2q}wl@i#Z1hxU1jO<H2Q5@2)gd0#gstHyF
z-;*%?Z>8u(_T`Hw<)#ZCLVO<I3`F|OIqwY1HAU_C`*)Y-E=8~+SShSo3X`}V)~qM^
zP4!o_9mbD1i<wM*pF>%aH2gyu|KB_-Av+d-!q}i~TGr`U>?$oZJS;>UydDYx#T0_X
zL^GIZt|yw|L^G6Vh7rw#^u)9j90Q~xC91LtO_jbBQDdU&41bo7ml<n`KYJNRgRAZA
z;_Bw^;ko7mP0Q8Bj^_OT*Q=$gr*B|rWMV3?SYfqtm5rT)6VG19+SXfBYq{UL2#Rlz
Ja4j4N{ST4lvJ3zK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_max_pool3d_bwd_impl.hpp.885D2D1066A38A2E.idx b/.cache/clangd/index/profile_max_pool3d_bwd_impl.hpp.885D2D1066A38A2E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2a63bf051e333d069d5e110a72a941322e03024f
GIT binary patch
literal 6066
zcmdUzc~}%j7Qp9~tz6A8J#vF1FgQ`pksFx+atI=aj9?V-W(X)KFqq&*MVNRH{eZy;
zL=L0TFELT0F;Ni>Mpq!BfKd*2GZ+s@6flagpu%Q*lI`x=^$-8@5Br1p>R0d8(e+-v
zde!o|bLM<$#^VJi&&|zH<g6fgJRU{=D|52b91xFZqsQarR$Y#ZD*lI5a4qJ?{D?YX
zqFI;b6m@rg)`LR>``>Tw3SN1qtL*Lk(9}OZs5Y`WZo4Pp(4GVKH6tnAyY}r)^`8|f
z{5i4BY5V9`OWl$KE)A}?F6Q5GA3T4-Ahjb}_&6K<Ykufie|LdFUFiqW3md)|-R@p?
zh`eKU<LjMu*YB*d^K7`-=Im2jR<+;b?3K|tTlu_$zROcWOvaNs_4KYBj87ii&)<JF
zf9-{|`Rg~9<{HoaW@1spHHEL<<<Kv}O@1>zxz}?bW#D4vt=M(J!I7R8-$#Ek{Q0N5
zk>$fxuJPgV*Zx_uZ>?v#0XgM}IzeLjNIA9hM~^j4BiGN}BX-?$d760L;CA8@h4-dg
zF5?ZqU0$-i@!sf_Zpn%FPA&dcHl=ZVZTE$`O=E_&?;YQlGx#WFmD#c3Z(25M!g_wD
zVir=j?dPsP*>dC%GArwGxBW0#qEos-Qc<3`@*3Y-vG?%#@|w)uiT7?rOgUW?el6a1
z%j2js-S2k%JI!|5J4Lmd>VBDdf4L#wL+7JU&u86Sl-bUw*6uW|I%#H=a`3C_(c4Gk
zHXYIFyO&TMF`%BfqI8?rMgO*Z?*sMu@{N*7RrP^~?H}p2Zdo_o$NPBEmZMHa=Jy*0
zbULo=YV6G&_nWr2z~!iQlKamEMT2=_<yb~_PIP|nR#E+yOx;V~fr_>5TR%M!o-{|^
zGF93p`{hz{MAgi2{k&t%x1;Jm+(|_4tx8M#Ote+pdth14n)FM`4H+u~Y0oiV^|=4=
zFK7fI<B2rV4>ehTvXDnRlJ*%s8M@9nXva~KH+d3|vYKdR^XCsw4*_oi0Tg5_Qk1#I
zwC1w<*)8rfjDV#2K->pTOv?!~4G!<Sx^kB;NoCE-3Sg>9DmyGYlA|2YoLL;zf~Xe6
z9G?aIvoM3>4iI(V@=lo62@agR3+%f%{Rx4q4+(#Gl1I-iZ+_kaMr~3{eP6%BaeYKm
zsZObHGu0%O=90$F)9_tO@3gP1F91Q6B4H^qX4H(#Zw5(kbgw}K)d7(m5XA@xsuSWm
zA)aADbiHjxze-FGlkAj?4Wn>Tu)_#yMH%jU!doZNO6JLqOk0vtSSp-3N;OZ#WNDUZ
zm@M5q9g}5PX0S4|i(AT0c2|s(1XT%^m0-<usAacG@W45!5<)m`1j|NTzY)9}!IzVF
zgK0OI<LurIo*WN@@i3Tj{0xkr;XDq;<2V-}%L0VWuK;-$Aa*?S&J((0pPzP{L{ORD
znLb>%+3wlcODJb5-8fkXcy@pn$DJVS1l9}$)dhZCAY<f~H!f%K9?BGN{90Lg{r-OC
zwBk=*+K@K-#hF{pg=MryFPycIT?iY8sCmts!xgmt#~}HbTQR$UZ)*pA+Vo%JqTAoE
zZPfQz8u!v%U1L6bXqx4C<RyW}+ke;Z+K^yO3tB+81^A5pW{I9$Su<(H6N;eRgl_C^
z)p|O=5(JDIK{bK638rzZ0aFd`$r=dIz-(N;NCOER4}oY1Y&aeT(J0PiAR5D2g~Td^
zji*8}Diq7LFGQk3g!NZ~tV+-nPOe5mH5YvdN{u|#$Qze0P$M?z5L7u5mUH@kj=W&Y
z(>nSSNa`sFo`Q&JNKzvp7{U1k2wreJ=iJsu>HR7D=)fcw$l0AtlZ{-$bt0F_*)2xP
z0_35XELtANQ9Bs6!(@gdf``vkv{zlG?YLq320x}c%DZ%(sq?N9y4p$R6vkFJw8geN
zKW?I3w-tn~z)oO+Ug-G@g`kMmuLiwpFk)K7MGN2hO)Z<KLsC;EQ`tj_q--QM7}-i}
zF|w1`VI-D_F|wD~W8|gh#SXRTMa9hCI+e#O4}>Kj_Z6?bx~!L;Yz_F-0Q=Ty*~<4~
zMn4ZENGg-aG-Td^J}QZ9jFf~DqgK#u1&YyCe)-e6(jCh;(dKfYTuhj1ngU#cIdV01
zwZtgE6_ZKLq*k0v8YpFtgMy=x8S&NE8|nF-0`F5Whf_NZey1UV<1^rShC4?{stsb=
zU@0Tt88P9*f#N@W^k}O%Ph}%FZ53x5jBK54F|u>E!$|Bb#>n2;9-|rj8Q7aU@EtJn
z5_p+0)7WFS-E{V5<IS`?RfC`!>{!9b@+AdE*6p-~Zb0A-h-3s+)sA;JUR>|}N*mV%
z-Ij@B=NZ27$H=}NGNUJ}pDXF5(13#mT$rWUpFWR>h~0KqhYqwZZc?UN8(5>8U_Qqx
z^o|NG=C}||FGOsvMo=Y4R)YLFc`2G+ip$H9S2<$O`UAh4%((wh{G3iFOTf1T*oGvf
z23Iwg%aGIq@P7c9WA{O5AMT_55Zn*p%)lfy1dE1X3CE+ba1;_b9s~Ipyu<Lp9g9Bc
zS^vE?oz|s0rL)H<NoBZXuxZ`bGkW6_pTs&51igCh8ay-7K!k>KENwiuS0gr{X*p1h
zf|$`xBz3tR*kvT-la!mG8@t^}`dHJe1S}jhf=MG-FwIEX8loo7v9(5||FHcq1QtWU
z9y2603KpX{kAcM)&IQP%0Ac;BkVu8tjYCp}$fA%twrC?E(-MTm8Z|OhBW(WV$grH_
z@B04so!Li&U%tj%<CNl$A06)m-KQP;AWT2VMc*@X_T(N5{IzR>)`i2HaNDTRTosC9
z7FFxA-X(~=uGI1$v~J5>&7*t#->c7AzHVE?qWe*Ad=NO-)V}zr-_xGgXqLvGt~WP7
z9J=tz$l9EJIQGT84CNcYs8<CcoiB{~XT9++(jxv~M}hwWBa)6yUT*BR(8i{SN{D8*
zjW!l7Y~oUAlIjM(ZeU|2Ne$CCO)h07sb^sQ4CisM9>-aQmZ{JxuKz+LEksxzP=Z2A
z5EkG|k+c+-mm~jjF2uIVBpc;1)@85J$*=cbTz59$=CAbS#K**kO{NLT*TfeiKfWKE
zRT7kpFT-fI>ul_-4Uz`UWZs{kg1v&V3qgoWh$|-xm4;%PVZtyr1t+L5hcNcb0YOEW
zMz9$YLCrIrhf$<yBt}tYQP|Z#+AJEQ7_%6bIz)97Z#mjreJMb84OPigD?Zx(%1wXA
zr!{!wh&jAQ+w|TJDcX9eXB(XidRcghnU#C6=4O6e#ZnnP_q@P7)|vZOAB*>$9HhOs
z#7~ZYi_z}eee7S=j;-3;(qdw<K2uFn2}A-$i9{ksNkkGxONb>HEhCn(G+0$z?|R(U
z{xvSbkY#$zirAbi`kL4Tf+p@hOj0T&P$AZ&+JjxU1hMyN=1oEW&+miGSH;m3qKD^~
z*Gw;}2JAFo&*(iH&<mjAPHdysuLf*ufc4bpr^<hD@oPRw3o1Zb!G(|U3oVkpvfQfw
zTk&i~`U>rbN@eycWwExPtNr=wF=7hzN6pZ6!F_X|dnFkNbqw@LL%xwfPnRd)6!f{G
ZkSpqN#s7W~)mD&PtpQgw<cfT*_#Y|u?~(uj

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_max_pool3d_fwd.cpp.11655872CD06FCBE.idx b/.cache/clangd/index/profile_max_pool3d_fwd.cpp.11655872CD06FCBE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3ac6f2a5777303c4319e2f1abae0b8b545f6f06f
GIT binary patch
literal 4654
zcmZvf2Uru?7Jzd@AVfkE!pMeDg3`N!lu&~~iXb5aHUt8!7!sm1eFkB56&_XuR9vNq
zVtMv;MNwH%tO&LhSy55<fd#OzD(F7M@&@!~?)&!5_i;1-Id{(7d)i#^@Q@H2E`=hV
z8?K(8x-gkWp->Enznq2HiZfai$~y*yqOLkLeSSu#C+AL$@sV2Zt5HV}>i?S^rTqC^
z#l_yZCze4QZC6j2q6}?y9e(E<mtE~0w9*<d=PWMi?|SIpoTF0;{!@D-QCYfcf7WbO
z04r6e=J03ic4y1epZ1-sD!x%-DlorSH*`y{Zm6%N?D5B!SKe&t{_uWI_JgGriA%o}
z2+n-*7(4LuV)uz#WhJ36Io$*A;;)^JJafJO%%fE;`gz;50%Qq`nieGv@dq5joC{_@
zEtXj_`gT3kG0cAY;xhfz5xdP{todisihn<q)mqP&alX6_OE<Do438-)%1?G`Ti)G2
zsUT0`ylmiwN5d6K?=kzPR_*(xk>e$Oi~o77xI8P>&9iLQm*Mj<&M~D~Hj_njUe^sA
zNxFJc_oI13(&-luo_`uH5m=<w>-oGD6~r~9#j<moLj!c*t7?mmcuJ>l|E2WME5#(w
z!l@xe9uAJ-A?x-0d*}31Gl!0s1omcL5QfB9h8Bh$TC?rEdH3^^757a$IVleXTMvu#
zH_TOhntcCBK(|wZUR&;oLqC{o*J7{oVC~p^ajIF>7mM6k;cmFerM)RJ_&uBIC3o%@
zHmwmQ|E3HZ&r|5r&Aa0B2D3Vo{#+dI_LqY-js4(E_$LtKbF;cXzqtLwYf~z=VddV%
z9GzWb-Hoq@8%Gtq7pYpZs`<)0#W4|hNr)NWLE^{DT&i<R9H-^$S~I+B$IX2Y8+H$*
z)6>IUrg!XaY+m8SF`6&62r`+wI&ZsvRa!+(-@=YJKX?fm4SreRRD$=!?~Fh1uqSU*
zMSDm2lFKui&dyo5Mt-%@EJ8+=xW9Ux@b)98Ws_B<gY?+?wf8=m54QPMM(=m1zKHcL
zd9>I#waKe?i%c@)yShCiWxb?>ZFJ{=wRre+^0?nO&^O;cy2mri?^=`mw~K3rwyyIS
zaGtndzj(kiI?lOAE|i(fe73XhTxQeSigFX%OG|GiZERFUe7@_oy#31_HSe@nu(Ma~
zn@8J}(u40r9@940{YRKP{X#~;?Q;`kyq{+7kvja^P!>^@#k=j=<Nxr-wVxReo8*jB
zR|Yy<Its*B?WXW~%MTx-kG1{uO#af*C)ScFxln2PUO+uP&Y$r|(uoG`WQ96nfuQ`z
z?S|Hf-Rcar$%OO0g}?um<<&e~zgHIJGPKcsp5Mmp2W3IG-nzB-)vP?QY?^~(<TAQ3
zTRu;%tqQ(&aItmAwazLNUwOs7ILY3si0<Pbs{ZUApVwB~XrK;UlU$Pb;a$R=U0W{J
zcw8I~k4>wzWYqPPU;nIaTDER$PVW4ak3=<_R?Hldn&f8cQ{pKUd=O9ufyuWY%1Ff%
zm(`=Z?EI|T>p0ZJITYN8VZ<FBzA7n{D9X*W=6J^yOiB`AA%qsfA0422?)hIg8&P(@
z)S<=^YJIvse{_KA1zjHuPA;D^kwZ-&jN3453`Pg2-ur5%b6&*U0Ty*Op=Nus14jp_
zKHpVvD>WhNGLs@BjJs>Qn~x4q{qmA9bkVe=ANkZ--$c0H7`>4uHhN{J&dXWAq)@a(
z%H7;^Z<YZjo<Awy3vNg&JI_~m;aMOdR&&+HP^_JE>xaiK*QGi%Ty7_~hhiF@W|xML
z!cKvZ(oTs`x?MU#Dm#^iU^oC32I1RU&0jW^>GnUFO)SRA$!Q{Vn}D@kxLWsG7a{Hk
zhW)?-E@xarcG}Ay{fYpMP_X)e0aU<fQpb!&5r&5_Yv^PO&Dv<!Z!%3ci={I0Mf62u
zzpa77#IFPTb)aD%^Ea*OI>~SJX}AZ=g9D>M!#!D^2zjx*5b|buBP3)A5%OXAAmq#P
zMaYlkhmb$ZAE5wN078MRK!ijr5kf(%APqs65Uv<WTXL6!Kr{$uz*t1P>+7d23f)0C
zBg{QalO9aG5I7bBS7;@MF{N>SDrdbG6K9*U$HUBJ;zpiEe$eYooX6oAL5hvy+V+<D
zY14^E?Afm5BNB#$4ec`VXht+bQic>EoPi@0!-zpBmJzF=86MS_Q>twfb%;&qhGC@f
zFzc|1BqcZ`c#>2I%nE@fH<|cOV73$G8^G)a$+JZFc{l%Yrh!PguYCkMFrn&VW7jLg
zHGfmc)KwWtji(<I*_$&q#}vL;wr^kk0bD32+#+xhxIu3*@dW>bAUKJMCkhfRNs<X<
z2qg)U5K0y#Ba|XY(GX03A_YcT6PW_3x~V46_&jdmy?6BAY^fs~GaUp@aAPCvG6(`7
z!}6h0N1GTlPQ&MM=NUq4nN;;>8@*jp!W%p@o;loI=H2}dwzY3m5@La|fDgr4>VTW3
zt!<x(ZX>al*gzvRJlZ-MA*r<#A>0~AD8@Plp;+r!4HcGds%M|OT}}A^YsBFGi<<XW
z{Hx~d9U@y+0nRF53*XDc%K@hx<voDE2Ot@A3h+;%+y*#p==?tb=MR!szMf(#yt$4~
ztRNH1<ippjovKy%s^p6|Ax0bax?P(eX5{yE63$>pu_drxXgFsYCyFEki9s|;TqzeP
z$uQDz8cA3rrlHCnPqQdWe+yCO+4O8pk7MGsfL=@XydA0t+Ei&#HxVZ>ORVW7dvL}k
zqtNH;N4^SNxhB}eSAm;`kb<i~NXb<ql+H~@NX1oY2rig0X7uX^^NJ`Bj=&gZ3KO^c
z&dvg+AQP7h<aW@XOgv4HhLA#_Ku9T2B9tyjM@S`5Y3P^uyKc4X*jyqi9t;nC*zW5#
zi!3;8ZH<2`?_<kcA6uAYyN_hLCC?fz^khi6`mm_yE=EK#m>mf>M?^806G4(euz@Bv
zG@KjEMdEE3Y&eyii3MW_@$`5s_$s27+y!1FB}`5Tfj*((iB^f4z0q))l?<UIt0aVy
zt&$N+u}VQmZZFrQKM`*Gv<c)J6!r=?l9cw!NhGD)r@NA*vR7&7+${6wR?l~QB>a2>
zINksngV5tdw;DO?z-ndUvvg;TgSO5;F23~Gv80g5;|zKRlE(#rUO?t?tBJasp=N1Y
z(ePq`6$2vozn<f0_<De?2b%K^4KD#$3D9&y8eR&pQgnV9z{&uRe0~GKHlSPqunKg3
zCBQ0C`&9s|LhV-rtQwua4Pe_)do=*7LHQ`~J_<A+1{z)myz78Jxt?Rd`xpo$`8e=C
zj?S+K-u1wb)Sm#}Cs4iwgqOf%Qhyl;FM|M*n}M(yh)BKygjdk{EkM|UuBR0UTT$)+
zSO>b_y8ye3`sW_N?xFs<53u{_{7!&%qW*XQum>nV0oW6C{!@TGMeRQW*fZ3A55RiR
z`Mm(^MeX$gtdHaymU*jI1jMk3Y!w@c`7lRsx%foTeRJLr#m^S8gJ8AZIUBm9ecNZr
zNDCLUgQ1uxelZ6j1F-=@Trn3RL$M)3m>APghu$UonAIENh^6GZ<!Vl9cPnjIbTved
zCB!>{<4)iXL-lxN`o7iDLt2D_&0rhB#q_1OFOS`l=tqe2t@1Sm4F8(|jQ{y4y~_Eu
z{(#zc<OnVh2<%|9`D;@%J=dxW8|{kdaWXqql{`|pIoaRm#EjfTj|^Y0QT}=>g)$1j
zv&I~_@W6NS3!T^*)1##^lo=WwC6>IMPTqb$f38YFV5B@si>9qZ*JbE2^$j>|)>Ljt
zupM{IR6`y%)|l_(>mLv(3Yt8@$l1@z+ri@hNxw5SGaonJ(%RPE(aB|^o4bdXP~d6e
VI%$g0cQ#=WGpV5xaX6Sp`7iyzkp}<(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_permute_scale.cpp.A23434A32C62DC3B.idx b/.cache/clangd/index/profile_permute_scale.cpp.A23434A32C62DC3B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..73d2b9eae79a65d1d6fda4ff42809cfb913f01ad
GIT binary patch
literal 3164
zcmZvd3p`Y5AIHxz2V<Nulcq6^8JEG#A{v^B+!lig(JM~5<rbBqS%x-3K9!J5R@7S(
zLsL7iRx8E3Hfn3TX;+~nbdyVVQ)yo+>Fpi1W1ju&dp@6e&htIL=lZ|R8gDPJAQXZ`
zvEDlp<5Hp#2!a+Uf6|oXco_kLs<j|!$Fc9QL`lCrI#u<<YQnRV<C9HlAMZ2|><{aY
zx-T;%d!9@*K08!9)b!xYMR9)Px%gMvLV|9X=~c0*;r@<1TTUd_Q&(E{eB?y^ahGvL
zp>@Y63u}r0?Y40Grh^lmhX$W%r`bO2;~z12Au%u3X{IeTuHW{ubZ}(w&5Qo3;EWF+
z&i*s2wUTzKyY8S3&2!@Yme5rfk3I6+b}>Tl)crdb;@)Lswb_Ob7fDt~c3WV4O4_Jl
z2x+)%J+d%)yuU+J*ckpx2&+G%M?Y-jT73WE8J$mWcZM!oXjZ_cI!u%|EYZ*F-&6Rq
z{QQoXowuxGuGR|{g$Ot3mwqv5)@m3`Z?LKNzZ^(=y(=;;9!s|E@E|2vo~}H#*z_4E
z+VIDy3t@`MTuPP8n*9l5=j5_t-{G3hJvSX$?|*q98Y;b9o8KZUwY=94pB(ZJt|Lls
z_E`ih$cpAq8(>B!9#6j`6jxC<2&}G{c#Ng~N4JL?o0|~YXcd?@{#9&?&IDF+esWpd
z!F27$?at<-4{IhftILx%G+eVe?{n#{*Zaz=-FtLveQKlcQOoaU4#vAC`Vph{T3l&5
z;jufQ=Tg$G;(^)&1;v`JqV{hwQr>Yz=HcH&C$}4vM~pdOhbmHuLFR(aYO2l>imR>J
zZd;<LXJ2oD=(ySWCuEt~zpg(k7YZWoFj7)hb_OyDzE&NX*q!Fymk$Tti^T_g!rQkk
zn(CtmT^0OZw?AjPdP{`<bhc|?mF2p6i?;fa0`@ok9W~`^+r?|8>3L%Yat=97e$<UL
zv--AwZS#eY8q~bHA}&5s<A5hEu;90&NT;J|V}|6N#o@ExclkMJQOD9Bo#C_-?P`uj
ztRFnGlF7B1>7J}TGMTQ^?Xc~rxwH5RXUW#kg{!A*-z-jTfL$(KI`T@rr|X9IBTetx
z;phD8=aW)hTv+2p{BM66@Q!(Lpz!$Fw|+Mb|KR5RH$HE)W2@|WaGy)^f^%n`y4D{a
ze&V8cCTu!0h2M8;)3xEEv2z_AXYL*2y;hsfaHuXS`D0gYQO0&zO^fePz25cfi`x=B
zr(67VrnF}ohcb?h{#g9DXz<r;62U=CP9B)N@afExr5=Yb6Lzgg*M8a;)ov9L`@L!1
z&8^=hy=o=1;xAC9Gs35yw(zH$KQ5a|?iEZ^$A2JPZj=QF?>(noCiXhr$2MY(6g|<-
zHTX4R(N-&DRZGm$-w}~~b8k#!=RSv>2Y-GMd8BZY10+pLjA>J5tP#CI{52`-D->D}
zK^PsSv*sG5GWVEPyvkM22wY8u!j-ZDh7_>==NB}a^vVZ2*xWf{qn}BE0s|n7f>5;Q
znm<JdlFPqd+xw<3M;-cyax57o)8-mnz6w43(6!)W3<X-FoJ~cj+H(yq-%m;UFw?i^
z6ADErXPYDDi{~0#PLA6sozj3HbvNqisDdaSieMX=8};#72)2c}1)wBm5<scURDd#=
z831L&%xoCoT$q^)askXN0J#)qmV$f;W*!2$5@uF{EQ6Uc6_fdSN4J#Q^lPA)H_MxY
zA0NehSUv#xvU~wr#aacBAIlFQf0q9|;T~2a{l&8=#wYR__*S@|VKv;lpH1vwFwwqM
zBYHv`!I(rQ311As^ojZav4|{y*hDr!21EmZ42gyS84-;DGA0@W#36D3G9j7(B!_k6
zFcm+2kiRbP`8Q7L1Qc`Ba-`s&QOrrp2_R=J=Xt`TR+BUrt$388SS%;jT(y9NBLOI#
zlMYZ0%*lZPE`d2EAXmVg3Ka*t_Azd6Z)<^(&lbi_zW&(Htxv}n?*E$*Y1cWB+thsn
zMzM9wbpkxbBNd9Kx-}~{5m8LwByd%oRmc@usAxHNIY4$?JAmxD_5isVyP2uxxf{C!
zBr+Dw(?;FB4<<G5@YH5+z*4{__zh;+OaU*!d2dNs6XmDwO69@-tQdH@=Dzl1t+-vt
zR0?96n2B#%;6vCue&CZmqI4h4aX#ym?xQ&Y<V<r0D1aWo#!Vti@I`zA_6;geW@2t)
zsR~j#>>`KfO@upHhxC^cq_m@zK!_hk83Bw~BNc6CZ=SaRib)s}K$EhifR@fk2ed49
z7NF(8j2swH^I>*A4Co~=qXg8;V0Ia(SHO%4P_Kg7RVuw9yw|p-kfN)GVhk+?8NWY@
z$zd%yOvW$Mc*xa=+SS$iw@CZ9SNV>5*U?@gf(hINp7?kK6Iu$Fs%W|8a)9hC?Etd3
zv<JwI>oy-?2<Fap2S~&f%@bbzm037j`f&%L*k;~lYkXrAlk%hhW%05A%7=ORFu-Ln
zuMFfWm{+CZ4wIdEyIle)N=up8m<T~jdC6fPe9O*}%aWEa?~W1>%#GqUUuTq8S>#T~
z2O*e<F4DsZe=E!E?B5g~^0O&mt$;;%Zj3zkD9*Vuyid8ZBiqpwpM_vfY^Qm%Aeb}T
z*%H^@mR;9OGkr>%a~rr;a$G|6>^{=u4bs5b&f3{#?jc}ZxswXD8@8u4LPCSr1+9gG
zaS4KgRNYWjH$>H4r|ND<j7^AF;;K#l1Vmj!lc=SQk{3`ZIzGBy9$)B^d=@fjixx8-
z9GzXhbaiv*)2*CL?JbP{-&Z{r+rXG(V#YJKTw-lwyG&@e!f?6%Qh_I3kGs-0OijF6
J<PG~l{{s_hKNSD~

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_permute_scale_impl.hpp.9398D48C7353608C.idx b/.cache/clangd/index/profile_permute_scale_impl.hpp.9398D48C7353608C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2baeed23f6ff4320b52006d4abcc21a532a4c415
GIT binary patch
literal 5136
zcmdUzcUV(d7Qo+y1W61@D32f@N|9Yc6GDk0fC)(N#ezma2NVd1pn~99U=>6r8Aiki
zwnRp(u>?`DptzO~7!=V_2OUKL5zw&#iaR^|!p<u$|MCy_`*_^nJ?GtW?m73Kyr@7w
zKWzgD^7$?>H$7=hA_;;ZUF0u&O;$>f1_UKgASie5<=D>Dm$R8eHGT5MdYKD4*PK$2
zj2ljeUu3+w7}1@=_hngr-=Xu<pI6QfJc+BZtG0Pm;kx2Pc5mMcaps3ihmBtX&OcW$
z^w;rr-M`}1JJ_9QRhv`x$ogf;zGH7nr>-9A*i81{#I6%opNe~@l8LXr=Dgc+p}qO4
zwYl3<t?9{yO9SWR&kJoE-P(O3-luK&E<@FF(;#mC^K$aKwI>6f9$xzDwy|swVrhuL
zrdem6oxAM!m22jrb<lEtZvE!XIY(f5j%Y}yfF2lk@E$Au({Wkes!B%bfpcZ6!gYmH
zv*_2hq+EYCuUc44UiqQx>~b@XW`<u<dH8#`kN%si8;)5|vq=CRUtS#;^?v6%w4tmu
zFD0gJX<oTXlz+3(Y(-`D^ysX*9{-1r&swbbv#V);wU_2_^>Bq(tFh}I!;UqM&YmpA
zKVq&ulW+N<A=H-p>ZkKpTaHlS)z@q{F$C<IR_A5izjikJySg4%aCzzd8W;B03@l$1
zyK45`yvIH(_w}ld^^)x(d>#hm{z@4!64d2)O+V5;KiRQSd^d;Lp6F-a;X7mJspFXs
zn(OYbZ8mS!su67MsVj)>tJmB9T07R^!P9`C$i*EdA)jrjslvkA?go<QZe~;2`i7zJ
zu8p4NJ1^L?ao$t!o<C&Wv?ZJ>|5qW>e=UDu?sEKN=Fckn8po^^I#-^2x?*7#<kc}+
zdT+bQmbNy3es1_BNmb%dQM;9hpV)UYEBCgxpuVY7+}H#goP~0jGvCxaJt6KobNfGF
zcw2#P_WJZ?iHK(zEq^z7^tr}}hQ>5)1d#%>scbz6aWr)a8p&8+9DJ{c1{rHZuqoNp
zbnF39Gyr5W1;8NQz5Pj1UB_)2WJG~rd$RqEu?IvUX+mTzL;=c5O3uz_B7Ma-3fed}
z3_ydo1Iz8e4t-VAY3k7ft;<TM&|rb7z#MI+!9r6Zfr6ca-EdimQwV`%z*z>|acL#s
zRRV9Es{pSGXRWrzfkF0ZDl!v)3x5u(rnAD>J>|IRRSgoX08R>k2}p1)(5nRos3ODY
zzHoNOZ)t!8SAzMKU;!#%)|;hsb_LwgKsu)T6XqVk2(UnvSWTNNRKLmoppjq!OE3*>
zM#w^7g_DpaB*=soLV`?WEHcG4MRp<w0`Wy&I7vXL1cc)(1tC%phO-hFDuEG^os`6d
zD?zLhU!T##Gur>IIqys+!NK;ygh4~>LkJ`T4l>|~cUlSDD)HSR!BxPm3TNYv8>^vy
ze$phghl$(oIaRZyd&ar6g?DZJ2ku?!gc{1KyZ#o{WogW~(Hz4M^&76XJuyz;)Vp74
zl5k@c;;1qZCc_=Ut)vZd_+|ZS<c$>P6fBV34|!=PM}Cna%{(3tiyy9Vw@@`|&KTDX
zH@7&Wd^1~t=+EKJnTPH;&+L73npWa_^^Vd7OthH<3+X}viRdB%Nq~z4U^5}XQs5#b
zvK$!7iJmG4-f|$o`<8<kIhc>L0$3@4HO?xar@}*+1gn6n3U@*^N2@?As_$V7`52#{
zLPa830t_Soi(VQsF#{>iewVhs6%QsK(4-+I)|CMk+Jl;{WPpuxCD5(J<w1KIrEP_S
zI>-o7fl*jU(%|U8XpGR~SA#ls?4I6l+nCf)R!4)AwUcL{JqNG+Joxs8umjnJC{C0w
z+8pZs<0?b@(jSm!E{}_CPMGK0AD6W<w<6;v`y>aT+B7(io9BsA)cr%6{TJ6SoP;qM
zO__H&J&7qn))!BSXP_&h!8?KOPTc%4VS0b-E8~-qeRG=SWQ(gg&vGUZXBG#eg<oY(
zy%!5z#vS`h_*MD}a$m-{_z;y3NHbdFdyy?j;3jyWgP{Kbb?mN0ui%ud>~-OvYvv*|
z4<-fcpjtFIgcO2NdPS-cuhXm-8S5~xJPfc~A@_P$cOI$zXEF_rphVE|w<9T$1WKi(
z5-5$5Mj$aoj8VbKko4Gn?YEH0M7Ty^PDO(wT_Z8t-l@nq<}D9J)<-->b_8ku^NsXR
zt`=fjJ16WKrNPd2&IIDvaR`*EpK6RxI!!-~Kw^C{MrAv`S48BOYyRJuyI4hrf{g|n
z5xc}Q<FRnvHIg7IV3<BYmQe_t3IXQfG`Iq|R^SK7-hH+&HZ-kuAGf*th0pkAk-<8X
zFcc0H^BM0`N3#)A$pKpq98min9ArmaXN@!=f&wtD0L(-Ma&a&zD0172u@s1}0EA4U
z7W};yKg|@s!#3UhmZQiPhna>EGMys8F2e7A8e9SlOMnU5znZN|fDO(PAdmn*oTXr{
z6kvx{b@9SIJ!`L+A`=a;3~)i!PTBOvE`A?rhg`36z>x!2RG{V{IS9d70SpzyYeNOB
zRNr0;b5%fy>YrKi$XQX!(9=dzQHlTyIwTdP2ni&jhzKMBjuJd)QeO~`QX<QNt{lMV
z5Hwf`bd^L_0UZ_4MdjxP^3JhJJ{+6S)S;gK`p&k@R0y%-x4Cx%UHk6Mh25c((7P|)
zez54I$H$%tHPo$p?JX_+pRSHO23oR?Mt*&no;?W-)ynjJ_?0opn}kzuv&=6p@QP-r
z<B=`E&Mzb$*+K$|d_{yrMgn3bU=bQm>UfNif`vFMfUg1&^p!xQ1c|u37R;^%*jaOj
zZ@nQZ+_Yjs-Ru{$i{!_)Gh4^S<D*Y|qgKjFipKe07x#|O$O$7bq0<=HsT|&Cdq<tT
zdXhY;xCy*SUIg+cd1LfpZ+)ZVX>+RyoyYLDzzbH*XNQngfBQS-b26z1^4QdZ2;{&*
zj%Pbx-cY=CW6o?t447b^fc@4St=oOg!J|cmG{*_z;?V^m-yp>5j~!5U*7EEcbwywO
z_+LRjx1mJ`tw4|=UZvvIsd!cWMdPmmcWEjYOOMvO=oqr7I#cO122)p0-(c*^-CrsU
SR2PD1c$JP<S$I_sul^TG+jQRm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_pool3d_fwd_impl.hpp.4E7051C23FCACC74.idx b/.cache/clangd/index/profile_pool3d_fwd_impl.hpp.4E7051C23FCACC74.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1870c6cd48dfa8e5614d61d72982306fa7589031
GIT binary patch
literal 5902
zcmds*X;f547RUQOA0YUk8f3Sp0}ZGwv71$P5Cjnr!X{`yS;Ylnq}50uMhs{}96N(?
zNd!^jm|)asfG9?7B8q7+;DUk4#%L7TgoqFw+8pO~a^9<&@q?fIVxI%&{p#LY-mR*C
z-For!!^3qg1cK1)`5W?c)@INGfxwvkldmmU5lsmMpG_7BHdLKUNL%%z%yc~VL8#uV
zhi+7k-6-|XD{|bn?!)s_!p@uw%bp+6r4H$Tw<go4yxKvWmz^#etdfs3R>oASFa0QM
zZqWHIGNNWIdCQQeOOIavr5~4f_$PNCXek<a-!Z@J>Vefi1n4{&`ruUl=(N}O8mxM9
zJgZAG?0){+FD-;-Ta&N;ayoZ!%iX$nOd5J^_gZ~&{gw8VUGAkL(Or+TE`~PCvIE*Q
zV{Um93DZ3fU%2Yv?X$b;fXB&9@1u^j4%bai$gZV(S5$|TyA_tU7>6vP(l*`LFWI=G
z&~(vP^!AnocUpGd_NYH(uy%V`;?LsbV;6QWid)?r5$$Ofxa+6!-TQY?*8T5%kr<IQ
za@shk%Gz7Bcu)HFPPc2zU(wXZ^oH4;kDNYlvuA`d=SZ3OV(q21m7U*YEj)AVRLbhL
zy8WwbSL|Pyet7r1gKiNh<a+0&4Ib|=l<BT`*FNUA%gw*(t=Tq5wcKLUj`!9bKcC#7
z8v8uEZlpA~E?Yi4-sbDR+aNXNjg9Uh5gCy-Wx1OR8|}Y8+7x&Aqs2x2`UqtEzU%zG
zPtlp0L`_~=dM_*EevemtLGPZ|PfdHnR!{!sexKKlW`+D+C%Hy=>#LR5yz)M1Pre;3
z7oAC{4frxWdQ{fZQtZ;ceCdHilde*WCJWOlzdGfne==JahphN%;>wb;2i-pnKb0T!
z>H98e-_P!6H!DY<X4d`j`Ec?2;D>V_!J#kfk7O-#F-iWQ>ftKJ4Q&Vz8Rs&Uts_OF
z=#!3Hwk10bXD|NY&fG_@4p%O9HuQNaFU-#jX4jTw)^9_#|E3;Ml!*?D><<|m7@G+6
z|AH--k9_=Tjp}<wV5}=ZA{~+F&wm1a3IvmAKv0mgOx`+`bzZ8P)9&twC<b+bT^G3U
zHPJ0_x&`hWPja17TXQ9In=XSg=46C(#SF^w%<{#Q6P6Q+Nd+<$cw%k>uO{%rd=X|{
z#9ipNs|JUhP8a}#(qw6KI5)#w>)9(l7c~KmDkzf*%9Jx8%gzwljbB!=1{I{LAdNE!
zV^p^JL+{zK4W-Ib`G$mz;(Yg=5gn7*Vo|D;^JaS!r4h8`Z%bsH&EmqQU3I_g`NNP#
zh1P{4u9zjaLJxvM3g=>~2kUwueCuI$J@{aIE10)}CCTosI7&3?2h)Bq$MzvG9U^%I
zOh-r_2g`A=#{LsvJwf(A0kbD?|CU`RCqMc0l$$P%GNc*&o6;!LHPa0@lI@r6k4Xjd
zR1kr=2^KU#4Cadvb`khRVSVG$8o_{n&Px=vX<j35?R8Lo^xS=7W1xJq%+h2RyC`FQ
zV*|JlNrLJdN4>8ayuhg2-)X#T;ISm}xkGgPEN$QKyG{VRJyyD}4CaO?js0Z6e2w$b
z=hpWXKWJU0oAJPy-3CU5LasiI8o;gr95AcFTum-?H3X@F-;Oj=Ly8&}WB)!7_kj)O
zK@bm;JPhJtl1IT~6uhv%2JAFs{~8F_kgK+Y5|>a!`wD8Bg8Cb-Urw2nV|k%bIptYS
z%_i*$<x~=8C1s+-{(jDaQ0r5A1~h~2f$$!PIY$N!fN+51`yjke@+g>&f`s!AKU+4O
zH?*vVmD(s_6fXoU*+fm}oLG`Y@$b%3P*fOUi-}6aqyj?~OyxK-bo}(5n^l+C=`Vr6
z5{Tmr3Ho8@@^ggSSwoS25#ovyV@ziB{li~Jk3lvPo0(iOgKQ<X1W6<kg6t%A1lddM
z338A)5G2)?@?E89_PF+b{hr4&ik|l2pz@6?OMBU&6?+%+C&-Ed`TE$$p9Zsvo<V0A
z;-O{InFM9i**wXQTt8c>NPmY7u9LBo8Lr|K<P?g@+1!~|R0cT*ITJP)3zuow(<Ru2
z*Z$4NqVtoguU%%xQw{Fb5Q5$If#*Kp9T`*uK{Z%GIX^J?^PJ{R_xB?wmEKoA_10%Y
z?`XqoH5+<ITY@Bx5`yd;?Fh1Wv?s{H(Se{DMl*;2%ru%wkW?tO;6}C6VypR_jiwvf
zX(<vG@#}Qwqx7`RNOY5(lyeYx4)}w=s@nN>@r8A>pP9yeAwB-$xG2My=wPT(_@f6a
zC**7p)ZnCsSzOQ#<VDeuaa(S)ir)ZH4X}VS&`1L@8h8zJ2{p5Xa>A^j{1g;_bZFC_
znH8j6NqH$LZ_fYVh?(P^0lNuyk>on(64QGV0&fDbZo43?3r`Y*x*@b1%Oit&Aif8Z
zxn>#E2N8XE*bEwk#6ehu?Zc2VjFT<~J%q@I5X;#QDN;VVyY9S*4a^ePCHzkK!m&5$
z^#^e-v{wGz)nwkIhDbF;bM>|IYhO<Bf~V!+a*A*0WNNqD!EMOIh(U$=h5Y_zP(7H{
zgB8~d!$x1+fW=iiO~z^<R(c;;^#OkhXkXK6kmO;o8YX!Zq@y@#WzYneO_2R*K&&DA
zE1|4P@F~L}1!b<F2yrQ=49h8FJYFSbsKneb_;!Ou#psb2m^hpc-?`y<@V)Ly+LapC
z0Q0%AXjf{m2CEj$j@?T^@u_U1QGdslH&+Utn=za*ntZh-<<5eaXy2dx_U58veh=@y
zz<E&>{5qWa<n7m38EpphW}Hb~zJ9H(ZD34$Bl_QqPg_B4oaWbj`AZb%E5f;58Z_kl
z61`HpU6EpQU;<)T<%%riZ&6yMN@ySjxE1_bA&A?J40|yY_2c}PK|>%KB6$QvBP5T5
z%{bWO`WjfM!TC0u{!y|LioeEcl_^v~&BOi`l&pfZE2%&w7T^y5>BaG}qFpa=2VCpD
zu=-3;=RNkO;cDi}XTmgcGjk)z!^neAA!+1k<VldP)R#XJY2+vKbLHNTM*d#@M0OuA
zD?o~EfwDltGuR}U&!1@&>=f*bJwwey`AmsMbIs=x6lNYqP`E`naqXXHF^`}KiwK^+
z7k{nuFK3%8&(+oLv8wpxdn&d*BLug+J3=2FwS*TqH@!V`x0f~F-NN=QwUXL#@fcir
zJjrKji1v07Gc{%gU%cb^@72y7H98E6r{fK{Vg@D92?Qn5i3BaA7ZQ|AC-XE~)u@u5
zu(f}I?>A<h7n>2cb`6`$H-NAKt1rWzib4&^3Tm=~5_0eK*nLfN#POEkXJq^FGWxE-
z@7*@mpazK=@67+)*AFsIJh_F9S+Q*~zX&J3+WpNezqUHIc)e^ruM>YXwNL-pe?!&(
zMYtd*FGKtBPhOBEe@pv-dT8mSe&)1|zEDr!U<zYsG+AJ%HT?B6;vNJ7A%4`ukL*`#
U?q|R~2)GZ7+JY&#l);bx1&Ql#A^-pY

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_reduce.cpp.7CFAAB4D3A545696.idx b/.cache/clangd/index/profile_reduce.cpp.7CFAAB4D3A545696.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c961d17fabb9bd3c39aa50eff0a7c5a3bd1ca2ef
GIT binary patch
literal 10466
zcmZvi2V4|K`^RsdJr!^{?~XglAx#iW00jgUkSa}q*s!vOCMdB*Kp_%iN5!BZpb)W(
zBA|#h7!3AUhzJOR9gK=%Lo~jYn0M~tFq8lL=ksycFTdT{-I;mjDeIWw0RcDqiA2Mu
z4o{gCn-nD$i9|B|FEJ@0J`O`7FA}8`92p-x=l&pR<JQOVKDJ-gIUAd=ti0v&)5@&4
zEK8qk%heA4#a6jzQ`_de2pqmP#>T_>4y3shSXAY=ROUHuaD144ddT*?mj~VK*K90H
z?Gv-nVA0Exn}<5vZGZ4IGMnYNc5q{Mb7R`Yr;&@+%uV|FT+{NrJDyhOpFL=J_x7&O
z_kVY$*F1k={c+|X50ho5-W_}VW6k~=Q|singFim)Ua(Glw}1Zyr<}`#(C?JX<vv~y
zGE%K;Rw(0)k}edM7#p0uC5^~DHucqaB^j5F<nA8#dUfYkulU!udRE;++dKAOP5Y>7
zeQqAJ?zG{@nAFfgOZ${0o%%5^TRZ;oiO$C#{C+U9opt$e%4*Y&MRo;c{%*VWt+mVi
z!aMx8hZEzCjsDEiz0thIzUupT3#!-0-rjC6nGvLIe5AME72gM&&PamC*bQy(XS=&d
zy3Z(5;j(4dDyi-IfE3x>{=Dz?R(-=;WmCRivD2W*a9-#T*^W7}NuH&CFC}-bElB;{
zx3ov@-rlY5mcjQQp0@f^KQ2Dpy7bWH8%LJBIep~p^x#r`)2zt@7iA_t9iC#JvmtYY
z-HHJ>6?@Q;Mho*krB|maf_<)@p5FV``ZI6aS`ThtGN;Aq`Ikm#?;Mzt{e0!)I=5R%
zH8Ylvnc*fLuVqzx@}=T}7aDGBSQwQSp&i|B|1qpI;8x^hoi{1Bf`>gADqm(|a<FaU
zH=K#Vg1b}N0&8bKJAHiG<t;0YUHs+F(_Fh%OPea~bzYU~?%0*PP1nKsLB)aMs*Z+P
zv!AXC8{B>GwSilQ+ck|1nU~W){-y10qwz~VeRFZd9~<gTF5QEJ&B?E>q)2Pq>(=ZZ
zA>FrP;O%6S0S)^+v?uiqT^#h_T~m0DW0;@grmtJPy2ZDh9kTV(y5BRhd#)Nk|4>C{
zPb<4>r{j$sNdsc~$Ap$H?&WZ{`iSY6!oTA0bl*Dpa&6SwgXOWFp1q%CTrHAS)GNyG
zI{mqR;r`%X|J-!%!J@R3)RZIlf-a4(3mO!<!LqF)`CwetT1i0aJK<(!T}Dpxxr~Wd
z#W&F}dAbgkLq4GD3$br+M<01ny|IOBym;~C_ulS_pY~sz>Yw0fKDV$vv#*(Hroq@Z
zBNDFlbOMX8wI%wVM;)^+e0weZ=}%!Bati)<XJe!Nef7<vjJ2054_QYg9GqnmpnU!~
zyDhxlI(}05_#0mu^cnd;ym0lgHOHUT9**lL>o)R1<Ga^?w>4Gw+&?<d%fGg2*YlCf
z-`sM!*2{f*KbMD>63)EdKccuYN4~i+-#aVg?dZY%=jZ;dx)bqdO~+-6?4s|NFX*ux
z#%C5KdtaSBZPfvH->}zx_O5o196EXdR6^zLabv`HiY{F};xw!5;s#l#&;2)jzqsWy
zOVz7#_e9rRqiMl+YAP}(MP+}S(l%ku;^7zH9{j<*N1qc@+lJaI%~yK5jA|dW;_%W}
zSLdDjz54g}HL*{UJXTv9?h1;L-T#>XcIP0EqO|3|JN)vW%9ABtvcD`#Ep3-fi|*qv
zaNVIQujwcM7TwEDnwj|BtmyrmNThpy;efDrjn^bFVYEmH(+)G#G<ZMK{7C;QNU=0>
zNRdDamVLB+^fV2TRkcz1TGkN<BrqS#6SXHQG!2ohlUyE`EFXG+7tIxkge4$e0`kwl
z$SE!XMd@KRCMpRmz&-eC`^q#8k%v@VLE>Rnr4mTQa<E>ov8Exi=c8SQsaIxW`8zCo
zYkSKy4Uv62dQTfPAn#iVBx8A)_Ar^IA+n!Qe8eS1fnWfsSPmD5chfXP4%+o}S+(cK
zkrMa@%VV_0NHh(Rga6(u*NvO_QUc?!JW_k4O4AVeD=z3^>{*X95}1wUP-Up8rXlhe
z52v?N?DzES0i*EHxOI0k)HFmEI_h?=z9@UGf-#@}Bq(*1hD46?<r{uz+*c@tdARc_
z+Ees24Uy+Axcj>DxZyYnOvCb6t+8@VL*#_M)8l5n3YY4N;_<x=lm-@>hR7MlIgalW
z&gV%W3|}3h3^CI*M9yBSYZo`aI7A;tVR^E6vc9Gva)Dc5&|l^DB@#%$a-cHMOw$m#
zL@W7NG&e3(S2P3PyPOl1bH<;4iPWAD{~VeS-Ej40;?Gj}7Wc3iX)Q*2pMU*xf=G1L
zXj5yg-xuFVU_3r$oc1_9O+)0m;JVbG4!87`z)UQAXn80!4Uw-|e|zV{qk}gjFbm58
zh5>yvO_$tJc%XVy?pp~=$MOWN33{4_$PF`B=bck+cqoB5Ec+Y!TWA_0H|JEH{p92}
zPy$I<9&Rw)MAH!Y_VdUr>x-pBdC^$>{{8rVCYpxGj|Qwe7o*bglE55%^(gI8Mw*7m
z39)gBi_AnK&ik7}`KgXYh!;B1h)xtr#*9lyb$31U&<hJ~(60>!5EJZvCp~l8Yq^7h
z7c!9|6RF4*ypWBCW+Pw98<1iHtLG!PeB@2_0;DKl^{vQfD+;9g4y4$@>V+tz5RIYw
zUZmK|>ZK^S6pg0(0i-y<>W7f!A@l{+k0QlUR<A%d7090Il}J&^>Q!h+74oF|Iixtp
z>KBmb1>{Hd%SdsV)vqAeE69uJJ?15*MEZRgE9QkXq?LvwMBs%-sOKXjqx=N*eS&N$
zx1+)B$b+yJ_%uj@jz@94PzPakFo{eRFVur=JvdN)2#yb75Mi=q>*#2<YO<5?!UZI~
zfXqn$ys#aV+rf%l#tUWOR|aD#p9I&F5JdR~c-(*x%5C7;2BDM}bH0nY(Udbew@hw0
z<t?217H$M#;yO4Tjmw~qc?~J9AyfL8zk=*nFeaA~mxA9;?N8`s?MSN~>Cww}fpQm^
zlgoJFAov}Gk(AGX>lyH)d>cG&vv1oDp6$Rq{W8vX8OOYBHs_Yj`P1vSbMD)@0K#UH
zC9bQrs$1l|U~gb=)n(R#i-8M6S;#C4+0lpIh|D&!ycwBoW_c$v+sSe<GAm}e9GR7~
zd<>Z#WBCL!JHhg4WOkb6^T_Nx%hkxNn&mcR(1xt&^JaoxCOgQLU{J{}sW)KohUF5@
zpoHsMau&xEraxUgG)*oR?B(`KGDu><Mef2-7E)y)6LM3ruo0;?vb-6oHnY4Fsdlnl
zj8w%emm^g<%g2!F7|SP+>IBQDk?J(d=aK3>%hgC#&2k%(w;?5c-b|2Wf{d&Tu}}%}
zN>EaM1M)X4mvHhDu1mLi9`vqew!#8$+0T&t88RRplMX=V|HMo@c9;(3oG0<6cB^(*
z-TipK828YMl&#2!^kBE@#k<}=JVaP{jci`iy~x?)tWMK+lNEUH@t6AR6PLyd0nz|{
z;vUtW!8sebk={tCh7DRBr%}snM2be#o1DfA)gY?|Bl7rzPQ`lsdF;2r+PttF43~o~
z5qM!6@Y}$QavAVtET06elXPd`g&V-%VD&cO+gM)AX)Wfub`tFAI6jkO`rpFwTPTy*
zvX_n~rljw)6-l=uL;60QsAng;^K_#AooEo%U!$I{S$!Sw>p($YYaKYRgQ1l7fN~F*
zP(A{FN7#M$Jh+|*AFAI2k9)xEM9;zVIdttr*bi~OX>{M^g$<nB2F{mWzl(F<#SJ6u
zX4-LSM)#-=nOKOBM<_@riOC|ul2W8RlF_2&(F{$MPh}`h9>-9;Jf5Lh{45FWbB=rt
zqb0}_7)p{SF*Hv;kD>YU`3xn?lNm~pr!bT%Pi3eTsaom6B*_9^zv@hZ&UYPB)X~#;
zVLkBc>HLu-0$--iAn0XHNYTVz_7-Gs>9Qxu1p-Y{K`tXn1$J-hjDq}`7Y-rmA*4?i
zRT0<}fjyZtUN{QlkAgt?HwgX>7-!r9BU@k`)nCA<7cidkN-k_AH<|KAE@UG&hH@bn
zTF7+;2PC=arlUz}=wt3i(%r~{KIVTA{~x+H@<I;Ca-a|G_6o?ZQ0_{KbTr8j?Y0z2
zOX<GF3wa>R0~PHi4_xzrS<?9+%V+fhkQK082eP_<vuhoDeLcwPS-lZtjV!+h+53O9
z$9vY~K7i~4tA7I7C(5G+>grCP7q}hAh`C^w%Z74!V3)^o0oWGMD4iF!fZZ0VlK~Rz
z{3|=;H5nHXc_F-8xB;a!5T}7Y8N0FC{wHR|O4oMdg<51-%LZQc$fO>})}PZYUT8oj
z4J=<rCf8YRMkdXa!@iQ+v}FzLhCA-(+RuwPckEHL{d+MUMLVDu^(cYbfxSqE$P2@@
zhxa1+(s*?5R^NCZbL^|uLBDm>O^kQhd_H=!&DTA6Aw(-gL9$a!)*()mLbXB_WO?yI
zsCB3lr9$Lf$YwC5XkaOg=y_ooNS4vro9KPRux9?R$C3DRVoLmM_%uBcf=_GlBX%O2
zPV^-iI+Cp;?SEzIQwq+WXnCfW{y!Z>q!i4FJW+WccwsSTzL>Kilci?o#hg1~vUTFU
zN1f!;IlhS$H`%FdLDCkQv&4q_oLEx6I~q^T4KTj})}%i*_qzf8S$4QVQww$0Xx_%2
z-v*9tFqoVlCtu$DrpHwqP!CP-a?Xkjjhd~?x&DMn1p;q#>hc49@CqcYpbyFmd%<Qe
zIFkN(;TViR29cC6K=1{2TWf`pt!%3F5=OlQW?Ngug{|Vgrq}0jA$c6L4esVbcXJbn
zPAVZibpBNV;e~YMlTNd0wdt7E((nmqXvnG_Igt5Rn~qfj%h!?Bb(WivRWs$p;=0Q9
zt0EWT`9FnJr`U;H1Ey=}ez`m|=V-4}^_%gE`=?Yw3IQu$`uBFN`B8;ucpF%4gD=S&
zso7y0JEPlR=r(riD+8-C_WUv!Tm~NGdSXy`;?$KNI&p<aTF6dZF357(*lY_(w@@46
zg?*sh$HrzgAgf`|Zvtr(d;UXEK4i~d%SqRAj0Y&-lm#?)BSk128>=f&<T1rU8W^O}
z+mNlMr;erwMNZ>|d?d|h-zEbrGU%4W3%fzNn|+$Y;CC3lB7?*W=fL$GbcIyBa2Gu8
zLLk+jf#)+e*<8W-uHcx|I+t_H<(TAjC+EJCCY#%528&0GPHUCopsZiSmt+?Z3sLeY
zCcKG}$1oHtk7a1Od^$rj<TDtW&Ch0nvbplPjFu=*WGHED(l~nedE@3ypfrE%d`3&2
zkUWWM3*-x!UKYw1GSrG(Tai04on5c>GnK&~cH`_|D;m0$-K7grU?B=7C-Oom3MfS*
zC?CRhgZfaeKz%BZHRTJ)^#U!o@WK`3d<6}m`V(aQge{k}Bl~vbNOZE3;rQ*}rdJv{
z-TH5FP6oEgV5eXgl0}pYO-?(0>?QBT3lX{zGBOst5UCr<P_%9|LsNC9G8Cs9$56a(
zJVP^eXEHQLcMd}dx(N&==_WBWPj?<e^L6Jll&qV~P>OB}L#eu{T||Zi8{0qCSZYVB
zfL{d$WZfUy;B79gt2^;O4)5@y0r$;~69#q@<DJyImR1|ZcqjF*r`1C--bwu%SiX+@
zue02Y{F^EN9FUu`-}_OMrenQY8bGT%Hmpa+<k{678#b_f9T{F{xfvNYQ?Af$_SbrH
z&=${fI+CT+IRDf{(<fF>wtv*Z0f=*mE7>n`$l)C7N<)r32+L#rR6f||)3{E(Jv#1V
zcj@;q@jXl-1IY^?!1e=s{#wp%E!X8I&(xL~m?XOe>hOYri$MVK3cO%A$S{zSk+YFM
zC6%*^A>)C@epKu2)O#2ulR+j-FQ!hWjAl8|lDW>x$*N1cXymmxyLizBPK*l&s%$!z
zV!YiN=hGn+3suOVie}iCPEGJEsu(vx6Qa7+vbEZJ<W-M`kx5khS+53`uOqMPEH@*s
zX3AGAQ&%nZ9NSBZZPFpanc8HOD$0~nj4FnqSXC@T(^b<MnxUG(&}_wQL)yz+)m%nP
zR3$Q$6rL18&zv`9UKFMI;qw_SIU+fRY70~gm|hmD7BbX|d|Ht|8Je1v8$T?{IoF8S
ztAWHoK>{1{p#-~g^;Z&VxL1($3hMxOf^sM8@XEojoOO7o!Syuj@NR*}E!N>Z0naC_
z!&}PvE@d6wdd_V<>+rU5?%Sxtt9^KS&T+46kFdil2IFGrO$<RSlz?#w%Rj@2pCOp)
z6)>WL<%{5T5xl8>3A`?`d>@A0r)6QW@COY0gXIqJ>!8IJvG5A~Ua`EI3tCN!-eMu0
z3rc4>pYzV=x{5bq;U~`fCzkhcK6_|!Ml2L@K1Gztmn$5Qsy|!N`ASEl(vd)BTOE+v
z)w0g49y!*t&aMGDHn4mhIbLVE896pnCJ7qS(R>Lbx4riHOBh}~>d#>0&%E#iX+2>p
zUOPa!gDn#u0KWrlE&dd^o}z=y3pc^zCc9c5gXd$m7Qck^UBcGlvpBabx>|6kz`1Xw
zwRn<j;wVr3IgUPNHj-wuRfRR6Si=_H3PD*2rev+)8VvXyVn2AC1=q7|=zIq}?yyyW
zr{MXNHG<`w?{cndZQ^1G=axeYZ@jRBbKgOYplOrM(w5^PN}L6}2fO#+OdsSE*nWb6
zgh`65ucP^tNZ+ptNvo*KQzsL8r`Y#>hI&3@KS;em{0k(fi|8fdU$R^bN^E}Qwmh~y
zB{o0GKZ74OKgt#0hs}@jMR3LDNBI)CV)LVXA3U)6QT_uwu=!E$08eawlwW}-Hb2U%
zIbUpkl+!t1Y<`sUIX7&6lz-yfu=!Em!?|Pgqg=$fWAh_SDi}D-QGb`E?l1$y8MJ;+
zKGACHXg<x-+1Q7q`;a*u-FJw8N1YxoYy{~>)<`yiY!g_JyJM3EX)UWafUJSl-+}ZU
ztN#VEzo`Cr#<eA3d0!934uJii0#9C;Ecwyo+7MidssXziHl(NnyE>K|!M2etMKyt4
z6V;#QMAtv?F5_|k>tNJ6nv*|E{N-p#icJe%#xIcG3-*m)BE6TCI}*RKG=IHQoy-<+
zjs;v-RiU%&*~Z$`(@EI>n>m=d(}6d4GWVcl;b6gNeH{8Q)YqYJ7ro8-t;cs(EecH(
z%6=#LYt_jq+`;nscr2D8v!!eaKO6PVMg!^G<s*}P<V?8=8CB8zU6c?zGfMqni3u@@
zo74witDAp*CXoY?XevF7poh`)a56odNDpJW4mqK+E&OTT2ETGXGUDr)=%}d^CxuU*
zBGTOapR>ZdP8Lm}hmrI!X4cf1`2Tr`LZ*GiiMw_0p{3oE*Of_oNk+&6{QAjtM(7zT
zltwBycMnf5Zy#R=efyzSL%uZs|Nd&w+r-qOk7Zw*FKq4lJ30+?8SH8{$asLWzrKNW
PP;e{+jvO`|g^K<Qo$!|#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_reduce_impl.hpp.17B63EE6A3A6D55C.idx b/.cache/clangd/index/profile_reduce_impl.hpp.17B63EE6A3A6D55C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..494b10d5b6efbc1255013a7639bf67d160543a9c
GIT binary patch
literal 8812
zcmb`McT^Qe7r@`lJP>eUr6WAbO4(4PNLjk{YC{DzYV6Ti;wRXGCIJ&Eu}4J^B1S0+
z1_40<K~anvMFA^_kr0*7f(^wQ^_z#8x8r$#dCtdq?AzbDvv=>z+<WKFywQ<iVKWCw
zB*8NymoJ#Lc$!oqk!a)pgvIe+86b&7T}vWap7VR$<b~Iq_0Ela(l9p0c4Aonigk8*
zk>Bohs~mILZ{o@70c%dIpP#xe_U)rJGy3(Y<nNmA#QDGTe6eH2${&Q6&xSC2|BBk}
zGiq%4FxO+bR|oDqFsFa*>pK_EhMukp8~LNt+0Oj-!o_+i_B%g6-9G)8-Y=19@fyn<
zVs9Uhc{vs~)o<OnVPJW~scwI}|DNg`mL`3>agb&IU7g9}kEH+py<OFRwg36f?}h5Z
z@O!)MBFyVdb2lta>DoPFtX2{9e)09X#Kzj2+no(f?r<Y>?Hmi+zOxPU$ul~Xml=0y
z{Zen|iprJgTV^h}&}$Otw&tcz^XinGyrhS(Oj}Mw-0Nv=T)Df#Wl8lfiH%EU%)PYu
zk1E;HA2VyY6PLBPlJ+3I^p5ckI?GjlJ~z<X<gC`wjAv_I*Ntf|D=lPGvj@lQ_;S!y
zgCnJ-U)ug<9if%}`<s0wt>*P7hp+v<-;w4r_c#A}_Hf9S_LH5y+98%d&vYqkJ`w2J
zeCy`CSzAmV#C?6;{(9cHL&e{<m9XO;bTxl5T5s{uMMW2u=8Vx_Hp6sW+xqU^8Qr~a
z7Su$)m}`DHv>@f!4u8|yi{7D&{Zbys<QmSMb56B)Y1@%Kr}EwJ9LTWVTYcuiHRChD
z-Kn?tNy$3nVy%|*;gx5`p)pTyhQ8fjo%?t6q446GBdx``$$m46|Ng4*Q0|<C(d+7d
zf-6#5c6T=JKI`mYo~4(Nk+e8?a?1Kp)p9*I#&W!i#<Z5AEjHI{O>0YaW6KUb7hqw?
z#a7p6(*{V(BDggD^t9>47P?0RrXP<ij~k{kR`2qdnHgha`x_NTngs@I3-rmDy6JYZ
z=Bc++C+0aV%J^0?rDVeKXWpN&gI&Rn!h4w)bv-tRl-zgFx0<G<qY=1>ANl#M>R<f*
zPx-}ER4<C{8a+SSG{<dK)RLbaew^t&Vq|0RjGfWU=pC7-yxdlgJ>8pLJ^fcV$E5|K
zK=ZlNoty-N`gKcA#Kgz_IXAclTwilonwg(9{A*BCdhe%*u&R?WE%|jj^1mHFBPj9f
zf*VJR?042aH*8rv>sj@yY0uult}Et&3EwQ3o(3cmqZyC8_x=T6KxnOtg9iU`vo&Vx
zNYp=ob#lj^nmD&^Mq;Qh;YO&8FzCBT)FF^)>i`aT9Gm6${Lm?Zkyz?UxU)#*EHe7{
zOQNo!`k_(C`Ip8g)_nerK4c6eoVl?%-*=IyeR#==f6gqOf1ziRKGeiwnW=10--U?x
zMD4yaadqJ)eaK?5nU$GS--U>O6X3~;>;1;-!~R%oYHDiNcOl}&sB0s12h8584+mhe
zg|UTY--U=TXpa2(-rq0V7>O0Oe+^Js15E$@lBlbze6atz)tR|R9CLo{4{h)$SgKkM
z>bv|qNO%;wr#7DJ=`9IiB!jS)d6IcLeHV$knpDC|aDxZ)OWq&Pt(c_>wXg=`e#S<9
z7m0d5oT20m;%9xCU|@>xryLDD?`q2ehTDp4x1u4$K;+VYZ*JrCwVDjq4ous@AR=J6
z>%i<fuoU?QaJT_nMD76g9l%-S*TC^La25GIaCr|rL{5TkNzhB=ROpckeMBxqA%!SX
z<UJ^S4;m(N5ehG&c|UU5k0|}+$h{nSllCZJZoba3C*P|xoFE9^WULv^Q}Cq7OYow|
zTkxhRL<pfMR0yRgObDYWTnMKq8w|?^BNc|kuki6|=o%_zxJuw#34)1$;aUOP3JgSU
z18f`3+o5bbRLp?t`J?x{?MgKu!#xE2L*OYs_!u}p27!e8=_T2=N>}USH#!QAzGBy%
z1Sg7|1!sy}1Q!MMH@a6`Q}IR}d&>qGvH^ZhG;7W_7QQN$)dMM)hK$pYj0gsdILa;P
z{=pL4b{Ft>Y1{4t=lh~<1FFpyxNePTQk6H<y$fuJ28O#2WcPuS$Pa<;Lt6eA$R5-3
zr$F~9E$;=gUQs^a?bNuu_)%%N56!@!85k4AItje%szutaSa1`0+(ceCZ{%#*v3A=T
z!#(qZw+~Ieg$0{n$R;?EJYje|DdWJ6%(qC&#T&&Nlc<z(xyUIODI!A7BXZFw@|e*%
zm&8Bz1YW}WvXE64a#RSemZ|+ycxVXL_Z~RB2ZM<~$|XU^B&ZkzDVGX|r@|3LZhZ0g
zMUs2Ivp&%f6OTvBvaL5{;mP?p_!(|H^4*R?iGEq+)y6yj)tHML?nlCYQNVEJ$fsNk
zPFc*@+{ZmnvXsWqbhdWR)uuVxN&>TGI|EJut1cCWtAKhH(1^SfM+MZYr1>D!J4o{(
zsCS6w`$+FTG9oL4xmL1Qx>{o=ws{{k+y|{h#dS!x4p|W9YU1Z~C;6s8DQ>8<59$+v
zl&eEpbx51=prIAjJ)X@PpFBN-mYPpKoLmyE)Oo90)u%Mvz8H%IA5Q>T2Z2W4`nT@*
zo1ObQeJuEsr3NbpOg2|i^w~C-uI58brjo#J*FWK^Ht&p-1P*DBUdg`e0yqu&2tGby
zF7Oq6De@EiDDoHlDT)vxD2fy!DT)%JD2f)MDarwS4zMSqI(X(kKBn;7<2ZFR1Jh<;
zP6Uo4XNC_cUbaR>%Eb<hH6cTS$v4(tq&P0llA@@%VIobineZ7Qr;)qw87#8jr!>|s
zT{}ulrvAPV8_)rQI>bEe=66D)^_$Q2`r&K`oHx)*RwK}B6f+0I-2iGgfEKZw!HU#6
zXnqa!UeoPUp;{`W+BYHlCgeg~mwe;d|4`wOs>w5GThM{m9IIri=hl@=)HWWUrZmmo
zW>3z{Q=RriUCLd9de@+0QNbw$8K;VgTFygKksIlQoTF3GaFKJ7S*{p0QmzUaR3SyU
z%Gt9@TxR7QSB1ul&o>~Q2Bb&WFVkYfq0R60u*J2|p;lb)15>*tubwTwhWq#@GW!$x
zlF<kbdwOk4)sq$+r>(%K74SrrlxqV<ZDK_dJhQpY`AYmktU3n@IZ&~%lDZI&tGp;A
zMN+7T#nr#7Ta;rYtakOo`Q2aU;uO*eG&_MVc_YImA?+k&NO<_k=9tTaAD_lHr$e1|
z$Pq!zrQST1@jW}0=4Fg}Xx)NY1J+}~E2#YnDpZXfp~K(V<g0@7;iy5Q6y7U$LE8=D
zD#mb)K(mpys0mp$Av-cycyU8k%_7InR$pJWZHPS6e7v|#oUJPX@ZEZx2V4bLs#tLo
z+$eGv+$r)9JSYkf0w@X;0x1d-f+z|Wf+@-dd^WHnLp8}aTX&brov)RAd(xftrENbI
zO;<8~%JQ6rQ|oGPVnIE?pU;S@S;GVClFJIFBfO>n{84xc!yN=Z2gRa;;cCFp8lYI`
zzMN+KE~{ttGORHR`eeZvqH#`LsOm25peQ99zKTy+I`U;<0Djebp!Oap77OemP%Q~k
zoDS8}MV>c$isOZzDUbRyTprNO1B$?4xO~9p14U85a0S4oK&%BAt_Wxr(d~<Ybun-t
zrpo)QSwYJy0bfb$I|ytJ()wzEeGO0){_?@ptf!x^2j2DKm@RNyFef(q#yq8=NGQ5c
z`*YgVZ*jCtaGT&s@-D-5A?q$=OFVOVL+7nTY0Ek6aC3F%8j+%j;pUmmvlMBG`VuXX
zHUis?bQktR&HYeE>~bm8ETy>&YL?Ml4mHbZzKb;PA{|oTB|dw$Ui+BZIi)VGc)9NI
z?WIN%EZ7WfH^ZUijbsanGl_ivNHWQ)--a!S{iWp;_J#*A`~i$3?HTS7jCcgciToIj
zdJLx!PMzYrbyrHkDE#0CsJ{U!qC#GB>TjYMS8G{ff=NwGbpK&`e!7zBsgFZtUQKI_
z@oeM){Gno?0QeM$^SLGp{rF(Rdrvinlj+N3q-P9gs&6VIn_Y&>1X`IuQ5xYo5bExS
zirSgsN}+Bk&1F!xjOKEvTTb&`WOx^ukr`N<?6tjk(et^C+*-RlI?v^UkbFSx{t7@y
z0nJ6gwup9(Vh~ge!pXqOqrt9%)>jF_DnYa;KL|n&()wyZL=8|Zd-7<oucvpj^<Z2*
zm?S>mfr2|w1o7^*c^Q8$TzR)0XR$^gYZSL3YkR7XJqk9z{eP-vw!WA;c@-<WjOYCw
zbbkjGd%PV%9lmBU(VCyQ75H_2v@|Lv<!`B!YlnL6be)ifY}1gUT$6Iu$e@~D_ZyH$
z15)fce|=lC@B8XM<|{emu1ET|x9(ZK)?&Cfz~~Jy6^HN*uzy1j@hWJt3M%SzhFcBs
zhuRP}!?0#ZP2mL$P_eyYxQlSiMX0De7;X#Vw;(D(W+9&}6d=C85V;nLJz%&!$bAn|
z92nq)i(HE61BDXgQbKbXaxW9x@9w_RxvLkX;T7Cgu%*^JJHd`3d%>O}2f=}&S`b<b
zB8Y)`Y2&kQFF7K^q5Kr^Pw72$FL3S^Lpg7=d*E?P&3kz4JCI=q;)qw0BQ2b+<mXu=
zT@~N|;hPDaIr1XJO5I9}>;d*oon!M}x@8^ifR$iHnPn|lQ)DC9P{a#7MTfzp!$6VU
zNEX4bm!FrB*YCIOik<i*#sT|WDzZsMiuGI0eyK>2ZRNEjUxjSRfXF$xO5BC+&u?$t
zsFhrT`;h@mGQ_<_*`(GXyVkNgxF>B;zYPu~8W{Wr{ZvG8E*hAND4(oC+Es|kcMV9h
zLEH*f3_TC{y;hu(msj3TtTg<or|6USh}Bwmn^OKWSH9GIycMj`Qfodr{?_nM4CtB_
z>{nWjXLifWDJ1PcTErI)`*Pa1(w?uwFZx(ed;0Q(-+Z3l{$$hZbhxC@jV4!>2IJ`D
zAHVfH?b%OBP_N~E;JS1EOf1NRnwf9_nU_B<O*9(sbqXjwaW;N$%$u7F5<YPh&L$>T
zZdz4osid!QK*-fMY`6C(>LaJS*puZ)yTp^vTdqi3<v)hSe%20U?R4%+LoR8^o%D~v
zM?rD9NO5k);B0DEji|%b2ISv>f=K%&Uz?SqN6Iojc}`lcEE#8Saq|1Vjt{(vc(T$G
z%~0dtV`s9r#3c(de1M#YpIk9JtkPz0(TFvC?2xUiuNtW?F~0JDTFI^%9DSEK)FKfd
zN?wE$F2c#gLiy=fV3ru=3|E4L5=0#gmm!}rF-ws2g?lDHGACyycuJ5%ySPGMZ=1a*
zFA42Za*0mQrr;O*jKC+4_1$@wlE-F#;X9b&%=FDHNwUS+lecmwRxzA4Z%vU6Z$lB!
z^Ay?gwiMa%b`;t3_7pkr4iq`^jubiZP82!w&J?-uE)==)t`xcPZWOum?i6|O9ux_@
zK#?c!Ns$-tMUgk}O_2}pLy<4<OOYS%N0C48Pf-9LKv5tcNKp_UL{TswOi>6QLQyCm
zN>LaeMo~B)PEiCOK~W?hNl_FZMNu>#O;L<tjJ3G^$p(hmz*_X`Y~Y&>qD0OChB>tS
zFc@|iP#K~YxYvq@QViFE2DTucgek-IAm1JoMfm>alIOqJH|63~bOxxMp%0<10JSSL
zUj=GcX>J8-tu(g*wKkE-W(<dy{0Ap7ZJY;|=Yb-_JzyK$r+L~pDTNowmDU=4KShb*
zX~kMuP*K8KQ)I*1P{gx5MYgOhMRu$mMfR*cMGmY3MUJc^MNX^}Mb4}<MJ}uhMXszX
zMQ*GcMeeLSMINjNMFJ~O<jHzc<i&bX<js0h<iq+<<jeX}<j49^<j?w36u<^h6vzfr
z6vPHm6wC%w6vBp36v~ED6vl>86wZcI6v0MN6v;+X6vakS6wO8}=t08agT>4FPV6Tw
zz_0}<%Ke_4>Z4AjoYg0%;pZulRaR%pdzB85ULGk}eW&o{hqaLYVdwS1ZO0sa&K|%v
z6hgN`IGO}y{H*!Y<i8as#Lq}rCchadzkK+!AUF8OZ1L7uyfs(cLhh%RQMp>YLyh`C
cBNLganX$Qr#6%&M$i!PS@mBt;?7x5i11hnjod5s;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_softmax.cpp.31A9AF77C4279825.idx b/.cache/clangd/index/profile_softmax.cpp.31A9AF77C4279825.idx
new file mode 100755
index 0000000000000000000000000000000000000000..34318f6439f93a9b6ea06f0e5e63e1b533284eca
GIT binary patch
literal 5040
zcmZwK3p^Fs9{}(<(-oKN-s_sK7_LVY$y+XULn)7oDc5wbRD>j!RUTzayPn(Xjp&`-
zZ8vL+N@WX`+R{r8v>rWddPqy8&2DY2{=;_6+0Xv(=VQ!we&@{mX6BsV%qaZ4y?@Zf
zFt4S4X$jFO5gZJ|beO;7l%&|QF&Gxk#jv!!<?|B~AG`4WDA7OKS?3yC(^ui?{@SJ{
zveo5O&o{=}p$8L8E1v`<h4r;=-LTJp&nn{<zJ@5DFwRf+od|rfMGFkptC<{H6I#?W
zH!9%aZ?BH@{?+&X(e>0-4tw@WEZyA;V}8%pI{tQ`G4s*uC!Pa_kGlFyC-+rU*q!}g
zwZ8w?isNrv-z`|>er^{vJ3qQ4v8(<4^|wFw4)r9g`rLYVE3d@PZtc&3p*3GN-Hpl&
zvukcC|H|paa+k1o^y!1Cl4P&;thNHb->E%jzbJQAd@X%*?m=_+-h64l=koTVyTX3c
zzOA}xv!k9YtdK6v-g`w{&B=DtH=eWX^1y1pTJFiJ`9w=*?uW}wE3+H@78=zRT0M<7
zdSVt!diLcH-fa&ncTa!wS|@Vp!qPF?C+M%WL$xQyJiR*UCN+C<-I0`eQwcgUX6(+a
z(22P-8`P|_H_d$ieo4wc8kqlGpIfrxMDh6SPkm+S4=J_FpT`vRn(y&8zV^_|Nv40<
zv{s|i(bh-z7xkS*FR=3Cd+JjqYaVw*p0HTHEw}8<EA?a97M1n0tdC9B$v<TK%K2-T
zyK5sj^1y=`zFFP}E)CS#o>td$vC&>*ZRKAc>G^W@i@PDlw~JC0&%;_HgL-duc;z+C
z&d$23Yp3R?C)!?YpBAHGwD9Wj4DYnU+h)HM`XvsYjEd#C&d7XdQjuA%>{u6cyS1&a
z_d<11?zqa$?Sm2h@e=>6T5JEvZkh6gQ%V@VdEJ((ZtJs$x{i~XyIXv%2d|Cm#EVUD
z22E+&raHXKcmGh|mC4U-N`JES(rmg>donS5hu4zD4ppf+-G|MO|NY?i`SH;c>C|*T
zlSgUpf9P14=my<Ptav)Tz&Sp0VC!S`(<|dl4=%pH?LesS&+-s&>9l~$+*6#p^R&)p
zaAPW>j)`}d#y)eaC~hvK>+6-zpDcBjT>Vzubuj;!hbUxH;=KZoKR3q=mYD9J)}PZ*
z<LG=!QdSkHh?o&wYVc-Vip-$&bX9Em!tV=`E4r>U?R?kVz98wdjas|88)~a|3c4qh
zo&ThJ!jANw@2z=ngzZ~YH=ACc@n3z!HO-}Y)~`PlH*~yfIyb-fi;GViPTsL=-&EFA
zdCTQi<gW$6r9}$I&YNkP_HLr;MQy9@SB(E{#{sRk-$gc4o*H{rd!|%YZFpmxKlRQ2
zdBud@YO@F1m+QBribap(E!?hmtjm8xy{s&$99XjYeNUnvxz=uth0XQPuf#bfB-I7H
z5;)#aG-Ao0B}9?{!%PkzEQ!)NQ_05{&ckSdi@;-K!OZEApSyC0S%inwNE=&(VRQ+O
zmEh#Zf7tNkXMW1&(_Z)Rd8G!p;-3u|twm@}__&T$4Q)AmPw3b%7N5)PL0es0II@7@
z^BcI{;T2wccvv_Swo|hcj4WWd>&K=OqJ`15JZuRQ_TYHvjx1of+qv$n@6gmcLOhJg
zUc?oNMiwyKd&$iwHE>CW0Ur8s6RoMKIb5(tenXFj=0E@Rn~OqxG1J?OYo;@@fMF~t
zdRcP6HioGsH?DhZdm~7lLn~y8889af9V`n*M9XMIN<*a)+fwPP^kb#6QQ0C2Q7wRy
z^w#BT^gUSsc(mmdO9z;tF+bBSPotqzokwQ?G6Pt`_B{FkAP=A}1$w2xoQ>ZH<b5=r
zhm(0Y(tk6qw;4y;7vQ=DxC!hJhli%-$gqGnlSh+6QXgJ@9&N5?ZUS@V(F!Mp4DNtO
z2Rj8LqMc|&N?oNP+fwPF^kt<oSJ@z%s|tZKtxj6rn<4v>xgQxooB{M<fWur3@d4DO
zKvW9GvGMyrd>@VH;o>|T>Ax8lZN`!I1vtL|*Ma@b`+VQt`jjxmd9-q((i(2&(JB|!
zbf{^Pfoq?PjTq44&<=bDAspZw+L7;w$cgWS$eHhqXexgyA{V|3A}L>r$d&Jk$c^uY
zXc~VSqUrqUqk^*<;BYQ~9P`24IcE~0g<Hn(Xn~=?4BmYnEpidb*xOAA2{9B+i6(XD
z%hWhb7b}4s+eINzj9y4)6bOP5(E@r@uzmpUzv@Pa=iTwQR-LT|oR2xeq!;hKc&77(
z?uP#a!pq944k_GX#xkSRmUM&(@@ToQJOJJu9_?@KZ^tUYIsj3ibs!?jni|zIQBM1y
z`t3>ha5{!1xCKmS*po_dd*~H-+R#L;?}C{02qQuQvslsg<Nf`23gei~q&{f?H?N#t
z7Ij;7WssTgXwgjmv8lG;t0n7%JUWygDq_<O6NeeFiX~%>S*4IEh;o292Ow>?0CEdJ
z;yVGk6ZIjWdkE;UdpiQeN6`M&fUHK_UjpPM)VBe78}&mVeu(z}4UoT~@z;QSjd}=>
zL#%VQZLboXZ{EjTmwz5Quqdy8ImgiL)+Pf!j}DiHkJej8^6+_V_HoiUBo-%+Lt?4Y
zR3w%vPeo!!f%GUqqDO)JC_v&hKw5*wYk<54jn@NdJsPhE@_IDh2&9c@yb;J7+4x5J
zn4A||Ke#eyjXtL~eOLS-G5z^HH4ZJ*6%ufYF`6JWN8}`QMl?(4g=n5I2<n@7e~vup
z%SX&ml<{SwMWI~x)v}t}IDe+NEwWl<&tA)NU{wxAeQZ&i<F}gxA|^)%u7fr#Z^w&0
zO@vKNXZ|xQPkdqb*n~`S8lDdm2n#sy;gbZ{{MYLMkM`mCXv1Tl@-9lN@O;*$%bY4#
zr~^-ck-pFrk(JN^k(<yHk*|<qMQ9Oxm;o~pok%Gwd7wN99+*S>+xt7S3a}4A6lfoa
zh_a^;DP}0<u>FE(1S6tn(1?^q%F)LzhgSM2M~4}QR@thMeX16y!r?yQ^=7Wy@XHdM
z7ChQZ!)vroTv!zdR6SEVh9`+6+!6NW(Q%SEM5&TgM7cnc3lMJwlC7wB0m&}ZM}g!h
z>KY)aL0u0d^{5+xq!D!|kaVK%0g@ingFrHf`U8-BV112Dd&EuStJyH`WR{B2w3&A@
z%V0#bC5>n)v2?T|a_DFx8qqRh8KOiY5zz`_1)`P2N<?djHHg*{YZ0Xr>4@?Gl?P~e
zGdT1HKy5(15l|aZ7Xqpf^>#pQM_mM{BGmf;wGZ`vK<!6e0;m$yCjoJit<W603J_JO
ze*(l$sOtbxhx$Aq&ZE8nhzqE%1L8XB20%2Rz5$3Es9ONhg1QwDt*F}o(T4gdpq`@V
z(*>w5)XxC*40S)C`cc0G)JxO@fEqyk7Eo_d{{^VOP`?AzJJi`Ym5n3!HwUM3Q0L-Q
zF6)~U)_#@c9xPz$s7x#~fb+W1cACHX^yC4iTnan|Ghyj!?7EOx?Q!QB^ZKC-QYzR%
zq*|hifNDACv*ynJp}E7)2@N3^_GL6)sE^1}XobjC=!VEg=nM6D{U>F4LFHqZ3<X?)
z7;c6~V=B<__ZVA&N^r*#d?tJkf`yf7HvC-%>pk;bqibmhzZKvI8uRXlKJZ>7*Hj%!
zo7Bl1nKULX;jUh+&pkjEr3siPR2Hz!0#jisJh}tecCh|)`(^dd%sYvZhY^|-JT8up
zeCffYICKy{NC5Z7eD}pi44+nVQdF`J^M4HeGIJ&W;@-k9)RZBr;Q82mc+waa%&vy8
zt4elN#jZvrERBz47!X5^;i!#O*WhaLv~`36zFg!z+ft+{*Bwuc(>ItleTKV-=ge95
zVjG#c>tvJvw^h&3NMdR>!D5n?^%PqNN2jS$H)9vl&UublZ=$dN0^BFa%MS!%{{!uI
B(0l*@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_softmax_impl.hpp.28DC8A49A2D3559F.idx b/.cache/clangd/index/profile_softmax_impl.hpp.28DC8A49A2D3559F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8e742d8fc3bdfd8d18e5451e8929681917cf1452
GIT binary patch
literal 7920
zcmeHMd013O5`P6x1$^As01d-2%&1&)MQ{*?J5f;?ydo+nhX<m8F&?P#hQug>iei={
zgRaI~Rs)U)2}d+>L}El3V8z?GiZPOGOmsEd54WdZzI@+C|Mz_#P5-*8x_hdstE;<}
z@5II3b77e1^mmqIXDpn-G7Mvh|5OWe=bScRm~LZ+SyJ}-gsw5ShFUc5jj8ggR@U!c
zcxv9d{k|Vhy!m%8c(nb&Xr}{pF~96S9`&DR!)v8J>vt#zSKkV}r76ofTc7<Xctg(C
z)tM#@16H&Yjd*q^Iilx`*_3Y<TwEN}e&}9Gp}cB__d<iL4L3Y4YptGEuivA|TPOMc
zJ-^njUr#?7-s{d;M^8%qYU+FW4fk>{w<z6@EA}%<Tesay2n&p}bK3pxwy%$vM5<T$
z*3T{SMAN=9+Zr+*;?ov9V9s{CmV9boI&sRSf+=Q)-u`mJ+_jB!Yc8ypMqSBk&zae~
zb6wFl-Wj%U%{%Jgy6l-tPwtEKyPxUuM&0=JpIPX$w<E&FZ#3)N7ah^F{nKghcZe&+
zu65(C-Z(8@bnmde^PXonzDFy)k90g&s2j5)dIC;Pvz@x=qt0E`sgHkH5LES*g^l>m
z@)4iDFm>6}GVHi$&6(`T;?ys1b*B%1y!y9dW!?1nOSk6)q>cG8VE!3V<@3<2g%#4U
zjO|q!23HTSdS^~p_1QdKa(Tby4(IL^y&Gt6)pP33bjA9q0cTE>Ivo(@XNF%hTsW?0
zqp{n*M1}I4mH6I<^t=7MGd}rdlXc;Hy8>>N?mg`5Z25^c3>tR4(2PFTGmvweJ<B`T
zLDw5r9J_rhJ8@0)h{Q>rOGXyBMy@`axyP>|@<4BFg~~>GJzr9%JRXvNPFFHBbLZ8C
z4~G@451ssNg=VoPw=&D?g~7}S<Fsp`ww-Q2eK6|m*axT7pX~i|ht0te)s;_o4;goM
z{+J$Z@^;yxV#k7i-v~XMba_|J;TpyBOJ#~7<7@u5^_wEyHt&!x6xULnug3<twuvN~
zEYscvYvsRN_*uvK`xG9#o*X>%{dFJhZx_$_^kiF2e-odSvP(tVe9B@PwFk6`r~Jo?
zR#bcc^n0ntXy1*trTed2oH3ofw)e$B51&qp*rXI?{mSy-okvwmvuExEhB1(@U9~1A
zx|uWZG-Hs5$V1v!Fh<zV@h3|Av1i|PgC_<C9;Po_kiCh$6@xvW|6Fg*PoA_?9>g(N
zy6;2oL*2s}W7fbzG7u1i)tKJ>zQSVD%kE_&nH2E_={{!KIV0oPL4_~LQjt`O_=1pU
zjo#7qbIi?PbEcm;gPd7s)4qa`vxk`36f7E4Va^y~vJ2~C+E);Aj(lI*x0Am9cR$7$
z|1<~A!KttK-`w%!4q5kRo;p2ulsVHM%Oe&I>???T!=eP3^uW8DW8AU4Uq3&GzJf6}
z1WOYR%U=&(s(yD>g3rs@dx|`zh%bozy>%@|4o<P2{gNyZNu-D`2>D^b`dv;XOU%ug
z0eEyL)@fj0LC8#QMy6_~16FFv7m*hO2b&vk=pQiT9}uXY7pj9Hbrcg*f=QnViRlb2
zb7NgpGfJ?0Iars28<npBp9+u@Z0J95U*(0F8!=;?I8H*EITY_1KZGE|+-366#=8-~
zA(>3(OYCbp%W>G81L#X&Q5T52=pg2z)~KPEyR$H%974-sIO)i~Vb%6y-KEd1S)_7Q
zA<~SIuWG0Og{u@4Z34SZAQIxhO%P0R71&h?=~WO|C8XEDm>L*I`A>l81UOQxfpHp8
zQd|$?>V>!wM2$lJCQvrPc*=hs9M6N`x15Jz=OL8RTfx7T8jMA)kk|@Ilzs)Qu7HgY
z2VSA`7ZvzF(RzNk4eS0YSX>1=!YFF5Hu}%r+5kgd$&zyLBm@>!fUE)r32}G@L=fzF
z`Kuh}eptpEBfPQ^oZfo0_-~$@cmCya6q~IM{mAIT4MPset>vyn<s4F2D+Fk<^<q6D
z;ndK$FN~5E%wazXYK17;#TBr<0v-grK5uH>HsHZhJf(Es^bpdFQG9wLL2f5=XZNfM
z%f~(<H<MeEW)>;T6auu^Y_T4>M;zDmhJ0)C+HNJetvP2BG=2PE>PYg#+rPEUw{gQB
z9V3hJCX@0TyX(GPj@QAzln>t08QV6gI3CM)gK0NdksgM-JU*Y*f5u}E98b(t7NnU)
zg&;1Z(TPQy!D2HNV^I-U6#WqoE&?wqUj(5=G-R@<7{tZG{)%B_F^s1CYOqraBa#{-
z)DS78PgcVeifbUC27)NAg@9Ti)<A$pi0dJ!9^Rt-4VWmzO%T*1#9Hvs3gg2Cu(?13
z35%M+rumQ9ry2Zc|IIL>nZ^%3f?2f+BUT$YwShC`Z-eMIh^4q4oZ5x-c8G2l(skgZ
z6Vi2{&{5B4Q3qIc2*>My!5u>Tbb?i<klqQhP8dwb?*v7su)i*_>Z1H|FDBZtc7REL
z4%L874Y(6Q4%LE9tq^O#MkB=aU{f!|4Y*e!ZUUPoipjo-_i8>LA}gKbL%5mGgGh7S
z>GeO%>&dLdZe9kKWnfRHl`w3}0z03LMVK%vY*svJR<`akynkxx<TqjcmBV_HaqOAw
z42!}69C9D*?nfdqhdeAj>`7+DA&I5LUZ2+?PfJgGeG-VImQs6SK%5;}%It}iIh1Fd
zXHKk>WIHzG*zX@-!$wP%B<t7u_%T+pd&Ob>`f(^w9H`&IITR!g5};slumGhQr|J8}
z@!mLHk0!WBi#OD|UhNBwgSzP&nP|2x;+K`77cpUWz-)agF!@=j|J_q-gZ`ALP22Wh
z^ZW}bcQOC6kY)O8F=a{Fg48poZ{cQ7nWryNLq<a6nbp;MX5!E-l}UZ5K90hmqu@()
zKr(ym=zLmFQuR5xszqae-K)UTDQt=<;pR{(o2o}y!}^bxJq((NXS%{}g?<;zQl|KR
z*PHSPPv|Jv90eyLO^yk8EdGo@#v=c5Gk-Q9YfZKPyVF}fOkIJe5b6-Bk7OJQ_XwAf
zE|;&IKXS;1fN;!|=9?zians_{^gHgFL!+}NlwUfBnerhrpXQw$DuAd0&@<*Iv)PR&
z4kh84uP|Prw_w4Ku_&}<qa$vX$Rs|*?LHo9xbETR-$Qsmkp}vZDPcd61_@BGG+2Pr
zEYkEr1N(_Zx*ly@9s0?EoL}blsm=c=-78A0Ty|Hctk{AHf6O*Z><jE7Kd|`#>p#yT
zPah&nJ~NxAT~*hNf9@7=+(Oe64z++|3&kZ*Qm3skM^|virJyLKyJcCq*UjY(?+^Qv
z3syv33v)>vW%?$by>dLO{4mV?@!Y2>tYkF|QPY*Do;QjenV5eQCw0Z3EEb+#njy9s
z^qUBOk_+yFVZ_0z9@!4Qec$7$35Sk>>=@{?H6Fipj3!0=-WgRxJL31w(Y4gwIi!IY
z4ZK4-=I}W-qz*<1aeN(h3l7yoOudl50md}YV-ts(Ag&2U5&3<Yxz6EV>#ku(R9Y+b
zCsz(7Stseyfx3*~Nd~vu`tvEDYZ2X<dF)>#Jft;)q#2~7cRoc5Zl-GCQxew>p?oKZ
zJ84QtE*Wr`;cpm7)HyQv?u6vWiQX)W?Gvb`hddsSRMTJ2qE7JZ1b;eyCyeTZ1cIxl
z>H?}Znp;_MNM)=tCCwb#1eTk?mf}KiECd&di@~v2h}B@J7V_7CPYoTGH=<9i5Np6k
zBgFOKS5MW(p$42b3UL$oH3{)~usshBbUs>e(NZ<?XCBj5(4R#()CP8K^z6-{cCc#~
zVjbA&gt!AtJE;8elLaSdr>ykEPLW`ppih`NG{$<29)12G?{oVNz58GLz;hz5WXWO0
z_qu^B@{oHdNWQ=#iCm%}fty91a?ePzQL;!XmqwEPkwr4OEQ;g{{4ph}5aNh(_@x?C
z3~y7cg*YvYq__>j+l0r8b_j15VjakJ5Kj0tX8&r*e{U{>mGS?tlunu%eS3ED?Swah
z_mu<h>94Q6dr1D%o!9muVjLVO_`61ukDUANl2&{FXZ~jGj~8g#+<l49Ze^?9*yfdO
zbHO5}{h4{s@HZ-ZWtCm@SoYzX@FeqrczcOc=_3w@Bhr*&m`<FQLscNJq89^v4luEX
zUTkpa1h}36!KJbWCTn0Sm9K}%^+MbTu8l(eCYaJB9IqJ?nuR$<7mVnl5w9gox?DLH
zZGY`m*yeJoUx%0Lg})4qx}VFGvr9J={Apd10FT!({~AEn@r{s{=qCq7@Y*vX+XwEC
ze?B1nJ=l8qmvXO`N_<rj!^W6UMaKBY=yNs>#rnqzP#ha4;EZSE1!y8WQIE*m1Rj%r
zm>@|@Z<*!`zoTxF*ZvK?FP4bA&UZLqt)_@l0_bE@9aHrO;GZ4mjh&IWFbDslLf{JN
ztqkvMT(J;q!9q)K)%e`Ns15YDe0*+T)Gow2Fw#-{YbJY#k+0NZ)2YE-O}*^5hsF_>
z<E!&=(!9)VnID<v)0!RizH;p`+?*4UGmSLk7n*r9_(v~QZkp-~zT(CgFF*9S;>i19
zw=V8AcXM_TyIP4H9G&d!t!-=>XTriTR<tUjRYzLopMU=M*^>%>b)!`=t=iG5Ev?$n
QsuQi^M{V+1lM2K97fU=G00000

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_transpose.cpp.3798E55CF7D973A7.idx b/.cache/clangd/index/profile_transpose.cpp.3798E55CF7D973A7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4e3cd18a8be8e9bb40ce5514b75c04cd86fae817
GIT binary patch
literal 3004
zcmZuy2|SeR8vl&3&KQH~GsZe5J7Wt&%Fy5#%UJvDV`*rz70yAF774k!B3GLhI~`pu
zN+qSTRip)N3Q1)-N2SxfessQ>&dZ$pyWj8UnfL!Z@AH43_gP-PZZ0ktaS*g&lUsTM
zFEtViL68Famy(*yzb+0zQ!)^ge*D*f7m4H6#HanF_fh5>k5_$2+-3jvUE)K}sQ3%r
zl}tZczL|5w2E+V?`E?g3ejeMTNm2f{ai3b>=u1X#o`p6yV6gE}>HBw8O%>~Zt1qPY
zeK)2jdG+52hht0T2Fk`B<1#E>JvTd|F`Z;kM7T^cHE2qlJv8=i?9=puW?<I$sq^JK
zubm*(K5i^ACppheYzg*guNd)6Y!6ee9eQ|OBFol+D~4o@hx8pg#a?xP!^nSJ%baX^
z%C~Yt7g|0$J}_2JfxE0v&UcozZhQV!r~UTQV%gw*R!@{vu2&Y%_+IYVSFOrSp2<>k
z8&A183OQPE_Gq)58)@#)h`jNEVKSmOY9RYDo14K8u(0T3$;KO<JyEN&dR#kF^IF99
zkox(qgC`w)KE?AdR-Z0%pExrh{#LyxD`uLS>^3^Udga}7_~Q)IFIn$w05<e|3)?>@
zjLdvZ!{z)QqM7}0PARLfh8LKl-&PdfL2*jWtFd}vv2U2*S~ErD+6IL<_*vIqX2cDZ
zB#)Q`zm^K#t$K&j^~BJz+PXJ8(0-l1dBYk?U2a{~tltUA!NHte1>U^jveJTxQY*8U
z9SWIWwhhdkN|RP;WfKN0Fa74bv-?RO?Xgi<ZEx}g!fwNTx)f5!?g=rEnjAmL#BNGn
z8{<xp?}4kgF8P0)OW&8<iKFZ-A{cCEM@Q7(Z9S*>-;I|tZPJzo1}YW*bta^>t#eO9
zdCZQ0W^YaZ2XUw5x2~CQO`b3upNhXRtk61R9L%rI+GJI`zkJq*{fLP4s#M*~{$qO0
z_It;4?h!FpgUY@=$}m1t9_&A~x?qG_K9T8nF4+KQ^s1_aS==<a*jrh+mDcd#sm_vL
zajqMFA$4(cOk0I`@Pemh_2lNjD?HD*v7T_3^_6t%H&QBF-ermfQ*3({wD=d*)ZbmP
zZtCmWU)iQ0X=Rvp$na~X3+LHF&D(<+O04g(1(Tg|r<wlr2-?BaA)f8WiUUI?i`I7J
znF-7BPo1g3v6Dkv`ji59Wvo{(cgg!Bp{BT>?(d&gUpIAkZd<NtaP$18lD&&jQ$4eT
zp|!r8!*5)!pT=<QzA^gK)L*FHZH$oJ3zMz7A2h`-+E0C_T_D{Ue_BV+iQ8L$C4F@9
z{7ujBxPRD=CcOD+d`IcaF5c@Vr8C*j9B!LR)eKzTJgH8ukQ?aes1B>M*N%ER%N#X3
ziRH4s{2bHPSZscO3E_v}#4OJbt8O)I-`Rcj$&MbwDXe9~=3$KSbMM%5ACE2K%2IY2
zq+}#SS&Kmsq3rh6e9YW!JQNK<2ouX(B^2nC3qNYoF}*t-``+TADAa6)w^9=dV7AF$
zviWMcrJIO}2!IfM8GQwz0A@Shl`oyN5+6B;i+U$PNvc8t2_6#ki(We@9qG8oND||Z
znn_Y56`}ZHgrJ1MA_=qbT6-cU5+x_gkQIdjnD??je{&x>Tr2^3quw|?jwBSoTv@#{
zJ1?ElClC4hKnMX#P!<Yc9{bG7Ch9cbBVfW&a@ZPn5DH-a)<QL&&i`e*JmiOZ%VFh|
zg#wt9c`+%sB_K#5WnjlR{eh1d7U4K@oIo<cFGlun$;2Z`^dtw66piVd#nQ8HgT(L%
z-I#6;QVBAS`U=twPQ)WDX%-m}gUuL>FyA1NSj3CqB@Z^)5j##?I+C^|g-2|#HnM;i
zk1((d5oC+C6+w1bI}x-NyH*6*W9>!II_$b-WT!FL8z&h#hvtE=$R{mF@K^oCFhA=4
z*Jb=<oZAw3#3FG#0;|E=pcHt7Y)^IqRmUR~A_WG>-%C;NeX+d+26ZujO;=nv@|;!R
zxIpG}a{n(kuA)&0dJQBx?D(Cw(HW(J|50=6>suOqZDS?z2$fE?04L!QH+wfGpu{6=
zUAEy0a@TbiK_0puBFIzMa~Xk(pj8y~^-2|8v92(ni&4yZHZS#14}-?5mP7;bqDUvn
z7$6WA$|M*8K#Vqx5?u}S!B$NAk}j@F(1#RtiU#o6I^6KG;z`LCR83W(Qh*vv4p#PW
z6NBE|#LWr>hex<1uG$KUR*9wpl=IKh7WrF`PoS+UNHn-zu6xh%(+%xfs5(eHh%qdd
z$)J}VUD*G-vzl#xAq7?Yl6)zE<N#YdXXfB{23k@Mg|p0!M-UiM2fl?>Wv`O5KD<D0
z0CQk(Q1{dE?pU@@?ii}(JMme763jfBE5X2nd*YF3TC_FT)Eqi&ai&mS5e<hg$5#OA
z3uR0#;-g1xD5YrgT+B8d*HTTkMAZ};MF*ttip$~jCg$X@3>p^;_5^XEn{X5Y*o0=?
zl4`Sp+#KDU!EsoGZNxTRLGDKGBFMwYLj-vmd5R#86UQC+^{S?i$)pn`(C9cO92=nS
zqh;ppS??z=xFVbH1=Is{J9nwIDS|15nXn@Wda&Jf`;&?#jzGNuW&t%Q0ot9BaHKLI
z!6H1Ew~Suwn;5z=E^!_`$ywSN4>kb~w8I2X>q^s7iBuJk5Ikb2Y^ViV80`WE#}N>W
z?~kj`_xV*E<%MNn81T)cTx)4e*BeFER86WrVEUu9ODbdaC<(oR=nD%TakL!-FZfD3
za5cE*;3O!S7au9;;FRR(6hFZtR3LtQ8^HpC0#}y7%S&;j&gEXtk*=;J$&k?CppAi0
z$g&LzT3JRXY>MZj&?DAM94jFyB`t%)%PA1$32a3dmcF7aTZv3sr9!b=yKcP$)6vNc
zHnCsLFx2{gT4gnL4XQRxN6)|r{jXtRX=P)pX{~Cy#u--D{mDHH<I36KCgutK3E3Zh
AJ^%m!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profile_transpose_impl.hpp.8CE368E3C650AFE1.idx b/.cache/clangd/index/profile_transpose_impl.hpp.8CE368E3C650AFE1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..556406af5861a91cda4f99dd7a5f83ff4dfd2986
GIT binary patch
literal 4964
zcmdUzc~}!k7Qp*Kg9Ji2n{Wi=)ByyAa0v*+MC1^|4Ia_W44|ld%HaX7t|se=phVF`
zMRXLL1TPeI6=rc9XIN1|yzrPq9lSw>ofX$(cRTKMSAG7$|MUm>-ml-QSN*D9y{hii
z_}JK)HUuHgiC<QbTOub3g0Q0hi%W{~?F<Ma-h?2Q?fy0+Y4hE2mRF?x@@50?8S~}u
z_nEk@PQIBYU4D9#(9-Ol&7G~uO%GeUFRneh=FLP=QDFbBi!Gt2gz;q-3p?*fxxbtG
z@2x%Frd?Fpn0as8gckF_HU5?n<doaLNx~?PyQ}1H_7`oudg6!XpWkf#rLM&L{qp*4
zE8?z6V=Kno-FE+gQHUm9b+qu-hPx5<qbue0GG41u<Lv61{ATC2=d;}7#!OC8cAFjc
z%r7Z;yW`gWno;uArk7{@K0jn~&=oV|9ex9S&Ta*wL+2vK`F)sRwr$Qd^NNQlZIOA#
zKdClacFRkSMCU%ambOG!_EWR#-#Fo2JCA#4e(s5rxBtiMNNrI6hW5w(I-8U2$KPa@
z&75l(HPvZXVTW1MxpLQHhmtFaGnM>n@h3Rpi`JaF;QAzQ@}-k$_3f)0G6c2d$3IwX
z_-2B?LE`Ja>92FA^$X@5-YYcVW-n^G1utWC1^NEt%H{r!-OjcpQ68b$U4H*QJFq0P
zY4J<(5qU`2%er3$zPCjqt4*9E)|E8AjI}O(HgU?WLxUA$`hkji^DLups)<(%FJG_m
zGPC>m_{W)(7My#yV$G{I&+YejUTGOWr)83h&5?a~7JoOP)8(nlg$E{kR=mo#9yFTm
z^DjwboAKU*PhfEUg@#3O+9>n-!q-o)&#k-Ooomt~y6c-7)aun|r*7J{<i?Fzc^4n)
zmL8MNyHPqY!fCZ_QgLa)mvTBw?7w>5tG#b9Xkg$(=ZXHJ>`m>h3C{mvZQ{k<`*f|Y
zT*8q<P~Jw~PM>}V4rxdj83P5E`Zk(-D@s*d!htq$GIHX7`XM+5KnO<A21HSAUh$v>
z?QV9fs6*&&0bFW@$BHqGXTtz9ZSCbRH<4UQED*afx4D!=Ai+o)APqsf<bY(Bj5}Ii
z^!gk#rTxac#|s%F^Er;8`3-_|1|+3{01dDTl4<~_25@E!Ezh_V2&>~d4e0i~c&y6=
ziv)MZh<|?7_7lITKXOS*%ojT_w`p=0dm@tXC74d)F2Qs%N0|Uw%Erk2F^ZH$AyPn+
z0;GtQFhvPd5!ZoT9oXX>P>07?2N`vc#pvz38%;hOJRL|6gw#)pb(-v#%#!2v9}0=*
zk-5V#KBtgZOLuu~)_j&%?$J$W2EwkI4jHy*&zUyUpjqjkwH+(n|J<(7_t0mz!(IB@
z=NrrWt?6m!i7lA{CMgL|f{~0T!$<)^3Sg6!q?8b(#90k?YJ4))5UmCg@~wt+HDn^z
zfTspVA=Uy<3s`?zaM6M*uJ_l%IGm%j5QFnnEhHn>0Z)hf)4>-y@WJ&Hbs$090K5id
zAK^=g9Z!!ir5BR|Y!$#RATAxatrBr;%bJ%9dh=>cxRkTGvkhBDC>JXiKBAG-NGC+D
z4z42+xjDOeAgTbX3g9!(#ic5N{&PaS1+2CpeS)TSbBD6moPN$Mmn;utmKBsWo}r{r
z)j~VJO<U+RND5Dy#N1B3vET4{b7}T4Tq|juN#7RaT2|6w=Wuc?n25O4I<Q)YMkjr$
z_2{Qnkw%7G%FoSjG%^e2gkp`tIAJX17Zop&zC9R2dzO+?bJQf6OlGNI`#k&bdoGWG
zq-wym2H0X-xGbtu(0=VBy;wuNLb2=)^9o~WncGqR+>V@BdOEU9ve-Ji?E42jWn|YC
zBQ6!qiRPi6V>mGwi8vyR(mClEWpFZBQXHOAkWtgsMhB7}oX#dXm&ypvV2SAjOh0Xh
z%qCi%VEEU*&HoKEqsQ$Fzp;!JmkJ;Q%n${Vffxx%Ax1%D5JtgdFh-%ap;$tP*@j^h
zZ4>Q?dWf-!!AN8yVyT+H&Q7$_b|pOpIhHx>hEzTHCz--RaFdRDX~5D6$Wu84m!opD
zXX3AsE-8uT!!Y%0*GF3P&wE!)XO9|u)Zov=xVP6Q`FsB0d0J88P{LNueG3xFgp{=p
zKb;>UHGu8U4dB}VeoXtXUs2vY&pqB5bE#B8Dx1<=YPrvHECp8Dt;D8lrRPd-#(}<6
ziWCrwSP3yoVC&6+_1T(7i~la5A27iqA&4<M>isxlMt{l}W0DR?sD@xhq2~lOOhK#x
zI}N_a=zy;Sb^(!;4um?qy$(`zkjAucoOwI&>rEEcbg33|#Oz|GPiPzoMly~JBL$3A
zpficS`j1uOtOhGJP)rwGst&B`aMl7x3#`ABFUn5xHxB5A+|1ALo^`HWG>@Ri2|r+-
z<(XXlF}o*e80KutKW<Jt9P!UbLx$b<Z0D+)ry*TNB&7uhE!y++?5zb~W{~wfUJDU8
z$7>-G=V@A)!SJObZOZ%K3W|r}z%j=H1Jh4$BTRkfQgRbHTM92*O1j@$^o9?^_&Pk6
zuUAB+S?HIfkC-iG^pf<EU?h{s;+WagSD6e2Ohc@MbR}dX)__C<n0+0{>OhY44IpfQ
zP=-4rNBucfD%d_`ZgxFBzwm7IjVHDw6~GU`UReV9ffxz-LX3j=K^O(|gE6XvQI)_R
zNJ#pK7f=a-%rogZvJ%+sOV5)lQA~PHtwfcLq^iKD3SS{r5L|_y&Z{7%3T=OS`@|}|
zvQ|N66?)YoDHZsraDOTYR^k3s5TnBVsUT5>`%^)t3i-R`(rk3c*Y)ye741P-r+Dy?
zJq$VYFSZ_eQ{&rNOP}G!ljBX9A>vZuWH?3<WCTkCyU(<aJ?7>)q~EuB<Rc>DFPG>8
zgc{t{z`o(WJGJ9G|A<Q`=!cyynm(K9kbeJPp#RW76c>F}d_phizXkkpQL%e`FGP6u
z-Q?Nv?M=<icw7rhD{C8Dg8rVxT(*or5avj<L!vzrO_9h&q9qdbAIAQ;c#J?0)=1<d
I(F}?I1L9j;uK)l5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profiler.cpp.A1998E185346D702.idx b/.cache/clangd/index/profiler.cpp.A1998E185346D702.idx
new file mode 100755
index 0000000000000000000000000000000000000000..735e09e6661c021045373137e31e53da6ffcbb52
GIT binary patch
literal 1590
zcmWIYbaT^VV_<Mj@vO*AElFfyU|<jg;^LB`%qV6C25k-ohKf0BLt^u9Tk(9IfAjKz
z9htlHc00)3(!1HTHh07Bh5c<h7q<!2R!o^=^84%bO>>u>{Z_P_u~GR`hr3Gu5grBR
z522}h0u(&VHt(vcs8=%jQSf)?jQeqMwKK)TyI&ri`1r1I^+u~(8zSb;m9x}4*DrHk
zbK3U%lOBIN{O8qg_qdw><#B8F-P7$leYMCe)jhB#YF*dW>8JC%cds`;`jof2j=g5r
zKDma}dllW4!Yx9M`Jb-kPvm;oa+6iI=w<HfW({7kY?XF3ee2wib#qPT9uhyY-{#Vb
zNi1*wH7rz(&ulJPu*3TFgE)!9@%QpBy2dRG-o>QsWuiFKW!|O#Y?<QHZ(pU@xFr33
zuk9bUMB!~oOT)59TR#VES{lo8uU#+WJ=^x0zrX%kd(8am=q{HL$SJ;6M)CcR6Y2-&
z`GyE=vpU5l61K=#x%betfUXl8oHAKDPZTDy)E<vz(7XTZhetdAoa8HFmP#(KQoe|M
z69_+kHKI1hxueX+$hQ55_oS^Kml~X`JrcQ_ujO-IdaZZ@1AoUwzJpvzOZm>t_mDGd
z+7b|Dq;*pyq}+4)#H+6aA_P~MNU?gd==MB2pgbqL-|k(M%-=rmXK_ng9(3N`zS4d{
z_Vb;Wo7ana$zD@*Q2im)%Xg?@OQPCZ3(d2c7vwr*E!PD%rrx#Ks`9nOYQe&fN5q-+
z&M8)TBro{>Y3F{EFXdU~d2jz*I~X>>`zmkd`;7`(Et^(rsk~@fQOIA?DD3+^#=^9u
z=Kk4qYo%M)K7>lmi5AJeapX&1(Bxygg3jOI`mxwq{byd~-OW9!A99z;y`TKtRV8+E
z#%!01mU_$n+rL;^H9O#h+T0@|FDviyOQs&sUlgVL{Ex3v@8lVeFMn>nRIxKHW4`$m
zr*%IM6-Bv4%ciSK{@}TF!@z#>#<YFPZc~><-edJz{PdB{^Yq8BYO3rkXMR|a_B73N
z<911#6#dzg`ai2p_;X*xzgfPwGv~HW`p)7{mk!N7dGm>i{?^mdcfDrD-tREgt56nR
z7yes9@3Z8=Bk7rc{pbDPJLjKT;=^c}XW@<!*>a-0JwHF`+4o0n{n97LANwBQxE!3<
z(Z6@b<AU=~S6|lt_q4}TY_jP;X12GMydq|g-O|G{rf)rBzAxiK!KT)u7wotEXLw|A
ze?oC(Zjv1^r*)jy-0Gn*^&k&JC<8+PFAFaRm;lOvLFcte*|%@J$>m`P14?u7a43KY
zm~>HUTJdg11_n`<lUsf6DXTHE1c(WU3BnB3=~%Y^{-4LI7<mGuc%|UNR&pOtWwRxI
zWM&EA;^Pv4*})PZB_t&R(+lM3OBup=5e~N`&X?`m#l#c9rN`wC)4~%VtSoE;QxrSr
z$lix<g}XqCJh?)U6xj)T!W5+mwcPy5enB1RRN(;O5SVb4cH(c3k8{(3P8IeM_J;{q
zC*Dh4&$`_MC@jq{Ee-QGOq!7kYiKcXp@b4ofFOsU3{20$9B&q{z_v$BK;MURMZ#PO
z^ski^+$(FVmS(yon-_orR9XzKcVo;AqeVUZLO|zmadPp(^qy0$?r5?L<_CE^o+}w9
z3=DNADL0tC=X-aUvcxGo1nLzL6Oo4rGZdxfB!c`^T$Eg_52P6b&dg!Y>v~~%k~t_i
zkRcG}X9k8~1TzT1OwLWo$pmtJ(*2lOSlQS)IJtRv1qJx|yoB6bG=#Xkge62oC8eya
zZSCwG9GwirbZt~Dv=p$3ip$8#DJm(es%vWN=ouOro0?n7n@Q`NxQdCZd3c8~x%)eN
IGWs$A06)El*Z=?k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/profiler_operation_registry.hpp.C3738D3851333A5F.idx b/.cache/clangd/index/profiler_operation_registry.hpp.C3738D3851333A5F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e2a512c4c458f44a3e3ea1be0c6108fc63af2cee
GIT binary patch
literal 2204
zcmY*b3s4hh5WZY4AtaYeE(v#mB!mz=K_nAuAcKNHg2+p)P&=X^5_|)#5(g|*9ICac
z)v4A;r!_kGpgyKp6-osWt97(UMTBWPXh)?!I-T)F?Px1>?4N@rcr$GN-Tn95-T&Y3
zA96F((|1ZR%(*CYP5I*0#RP_7H2QO|u2`NWz_4y1hOP0|?<g+4U{!8<mM5J&yIb_1
zcF}5G$-~b0W~0(oedC0UwHyuqy=JhdVZsId28YAa*78TiyAs{=syy0dA~t2z9V>g<
z`1o@Edb>KL(9;*)-!r$4RUA5y`a&GNqpQ48@a^@Qi<{1^y3}`$HdoHqJUm~&Ey+>#
zATMjqOyQKTPJU9?xmme>`E5}z9rN;E_rB^MZ$JA`n%>#GYtOB9Pa^)z-M-~e4w<=7
zM9jEk&*%u-PZAwp&ilZ<z$I!Uk7cyQy*MrS>`K%bIk$4_uFG`{b1z_<uyft}*`XOm
zyGl{DJK*r~>?fbbbx)r;u4sAE$Da5UZ_ZM$()^8Z)!n8B*`1dwDnA#`x^b$qwLj{n
zz%{R`HO$qk%?(JaZ`IbeH_AJRyKh&NFYK+&onA8Iw^O}&-<f*Wc6Gf98f+U9wPU5x
zbEmrBD|hWcCu)x9U7LM1#7g0$6r*&(x^cX~&?Wf3ypHhs*u)b82A2e4l$?+ocmeLY
zDL1s&D%)EFaI~M2iDgD!V1dKY1-FHJJlAtBF*t$18Jb~u0nvrti)Z^EsW+=|F#@-6
zmRMeZyX;z%s>YdelEhd9ZVEP;c!34t0!%Bzs-uesGY0l9RpKHPZ8O-ycmdI#A81CF
zp4q_S!N^S!l!zDL-reoF*KpLyP&k81)2sA)UV!^x-TfDT4;{{u;Q=U`meOiofV**b
zN|>^z=d2PJB6o6NGRX^Yw+QV^kLTCsP`DhqjhxZK3vmDS&Z8{4?@qrMCy-ks(<pfX
zZmeSQ61PEwVUmIS`7LiQs}hhDt)tDOH9fr=e89y{LuO;3ksM9dbn;$dZIb{eC{D*2
z;Pj9d&BcJGIfbo#wm|;`#l}a(N5UEEw@=aL8deq}vn|j@gW0g`$WO7`SDZm+huooo
zu3=_`B*ju}2;8!ipB!k8q&ORw0(m2sEeU*!RBpgYDv?bzKmwAQCZDDT+_=Z4p{sq~
zYP882;+z0xl1d9n8`UO<4qPb}7Y~RC%B(QckN~x*=7Pp(W^UXVEtPZnwMa$eXk60c
zBd52PTD$@&l_|;ey*2#j_`Wr~BKV_bR0^BGCP9m+l%%v&C~QVStux@J^S*$f5*3L`
zXbL!kOL?vUM+nLqWsQZI2dh{*%$Sk9<ZB@$g{C$Xwvv=7#gq?ikd#Gci2yi~vMQ{R
zpcU-5scscC%|_LmlqS|6Vu~_ZL4%%9=RRh_y#WR?`xptyi#(o*Os#7r8ifRHf)RQJ
zFabWy&5<_1jX-?F1dy60@49Pu-Otg1C8!e^*tK!n&Su5=-o0pdjwHuCf`I{p5w?%M
z00PYQO<xo{3EGhI{Dm&im>CnJAi<s~;fgO3(wkB7c8#3{v$yXb*%fI;i;#G2j?pNE
za^cV!(k$F0(13#?DgPcoO(ey{F;k%wlG01`G}MVGrLyTDCweFx$=NVL2Y+py#qIOd
zixGgH8Se+s$#q(w8KSw$XaJfcIL;P322Pc105}95m23kIe$lb|UMQ#)fKUs*mkG=y
z6%idV+9!G*bn#|?7Q5CS=GRgcsVty(z0v;lkj}l&_eO*`V!-_KoD)UkC#D-w@+?Uf
z19$o$W{~v|TW+EW7<+keZ9*yk68dWzJDKPljK&BkAH9-g#XeYfMTxt}w@~}U@D~J&
l;eQ`SG?m|~@h=trB{l0+$uHI=Z8HNDN?N7X1Yt_Q^M53VxEcTe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/put_element_fp16.cpp.4CB3E91BA1C3425F.idx b/.cache/clangd/index/put_element_fp16.cpp.4CB3E91BA1C3425F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..46d3175f8ab203cb671dd8c56b5232635725da2f
GIT binary patch
literal 3146
zcmbVMeLT}^8=sk#?Z>vp#!#$n4cjtd%Db6nVzwVbNhQ4;mKq7kiRUQFDWZB(ICxU}
zP&r9O?=N~$%A+N6PEUC`)H#vfBb^`5=l*SdKF?pzJ|DO1`hKt9eP8!=eXm;?;P3AX
zVKCxV0ja9E<P|s!217yql9JWyuIOVh;{*&Q_3*{8$i!Y3^5gQMj+OUZ?=&?Vj2B#g
z&G?Pg;<Bqjjt$z)x@J-{I?^8VtYWw8#?gN_M0ytt{3jzxU97hX)5vY;pK*?=Y2c-l
z%Tw{4i!b+|e!8pm<mqv>XPKp;r$_G}tyTJO`@4n;m#Z^bsmAr^F3Pi?zUu1#7?byD
zVvHA)*?i~^dh5;PMcMA(evKGtJxTxgwmDJs`KD38GC1RW(@PU#co!qG|9syvRYdIB
zLG}7`zi;V3cxqFuN6fl=orBO1jXPz5)%gz<FD(9iw!TJlcjH+v>z>lMkt=7aMWrX&
zag`6HVcQ1mA9>cdMIN|hkvtY+3v*V}EuDuePMsLIBQ-a@QYFZBp@rYDrEY7hcyqTT
znCF@CY1~k`yso_v$0+W>vD5cC_l@S&vuW&4b%aW6p241`9feO^P23BjDj)W|@lLt4
zxLYdO*3#9zeW7cQBHu?NrFov?W^c+2QKS?H)$L9jsdyb-9cer_n-_#zU)$B`l0LYB
z#HfRO?_0g^wH~)9%`w`%GmtH{&+R{vUt_h^`^;z=ft2IRx6wpxth+rTmD}PiJ{Wya
ziw`OSl-?tL35P=QAxG*%9|k|(U`uNF7_#BWv@G$S^<t?^8M5=*hJP^q)^O#9&S!S`
z23lF|ajQF7TQzaL&oVnFIc$mT=dFsfoh6L>-B0SW_?tO#ojD$Z?sLiE_1uR(?=IZ%
zp?DaSv>deP3d=Y)aGI+3%D>}A@-_abbAoRZzec)sS<UJjY2)^l6er*CGu*8Nr-l=n
zvsbB<!**7Q$8i)Rc1I@Z>t6nF!7jTrW^OWZR#UFyFIVU8-(`LgF8ZPh_%O8ps%%L;
zM?)6Xo{Q*Ms*H%z*e#Mfo-%i;U}j*e%DM)n_WKU~IvCwFQIzC=_IhzbhUWtFcE=-a
zLGZ22w&e#ds*0ZPhW&1Pmh}Ze^M6Q)|C)Gp*^U+R!3%3!iVD}3<hOGj)<}$IEnHJY
zZB%X#J5FIXH$9I(+|l<%&*8@`Z-X?`lvStWd$?J|Uj3t)YlttOzb=`US&o0cXW`2S
z`p&Ps8}If;7n$~brnme)@1v+`aI+^hKHXpQ4O^Vjyy?8^)YG%I3Ctpu;lawE(!WGb
z+dDq(-0-H7VHGqylsKp<h++J6z3|(M&x47jOfx?rx9(b59^vMKqh9aa#)ifsVpYFQ
zlx_`syyJrX`#o=q53E(T?#^h8uU>hC{7(OLckH$|w)~kE-8o-!zjh3<J@S)xI;{Kl
z*u1W)cHr)+Z<yt2?@uPBsa7^2^Q>1mhZR$CK|;g^V=%BO-qd7jL71tZ1O1U%?`w4=
zgjt5rRyeEaQwyNiu7Fy98SuM7#3~S)j3<+(7C>(rsm7jpA6QPrDiNBBr<zPHfNsqy
z>EUnr`4$lyg3u&9i8QqU`qN7|$EL<%n241lGzCwgPAz~|$E{8}ir3x%OZYBx2%cv-
z&kp3TFC^BK-&>il55YVk&jr8`%op-?Jb{nEA8`8WEpGF#4|wT8Foj030@>i%v&(B@
z{@RJaG$zdj)Q{BJIAOHn=N$;_Wb5PxHX)y%shYP>`T&8sc&<5s&F%}+?;P`aguqHu
z<tzZtGR~TxH~rpo1oonOSp(R5G+LM{DxHJC&bH3(0JdpTul;kI`*uAX>}Bjl0Wg<Q
zMy}Eb<8e5cMPtzcjDy(>wvE=~&^S7rj3Gmh<0qJH$m6jJk>fl}J-FZ#@L!Mcodg!l
zd~-J*zv2xR2TP14WDtdT90~fZdq-%;Xj{Gl$;pmnrxO%BCm0RF!SWe$s&)gpjhqQQ
zufH^}PaG>t!9uVD#R07&1Pkdxw0wTG#1|&RKQ2ZR7n_SM!M8!8y@@n_egOhg-KhRR
zltFS4{kG(v3It|^Ga^AFgIXRcH2jg<`Z$<p%wvE}a4_GPuj2`f1!%#*6yOQpsUL#P
zL}uP#UkK(9IVNB`M0iZTz1DLzakbKVGLp;;c<a<jDe_N8y(g86;~4=M*#O7a@dP*l
z>VX^T6TTZa2sRg)OSJj8O?RV#?Xv3@slrOzZy~3+@Lf=eL9j?9LMH-(UHPskw_`+N
zIsZXE18E6~L9zzJ48ddu*&38&-^b<A*(6KNWJ{dIKH#{Dgg~53vHb-iUCCUD5Wtng
zrnY@AXMaJmaIkbh#fo&9E98S9WMr7iIl%i@Zp6X<O_wZ@1Go$>TG86e6yb5Y;u{Wv
zS!OIAh=yP`ldUuF95ap%C$o_`gUFM%17S<vD0q_@G9;*$PCfJ7e9%n3cCs5Ho(x1I
z-4OA0Ji$Ui6yWNYK5?wfC(S@Ilo(3TCV_e(qM3a8z^jJ5Dk7SWXjnv+j%c_%E}9{d
z0h5Pb80njd3o4<u*;{h6Jz>)1nIZ}L8a<xbY|v$IeGh?Q20R;N*iE}Uk?v)kK3R6A
z0|&rJ*#$yJt*7KF(PF(SI`N^@{z3U<eQ_*;Hi(U5>v$X-N5_**ml=ad^f@Jb_ozcK
zT}1cOYLrc6qmux^0<J*k`XUpNjzof?VDwq|k~g~hwV+#$G*3(yqwv^R_2s!e``;ng
zXSgv!!6{#Z$_JAlN3|eA_E_NI0Ww4sD-fg7{CcI0Q{tX_`2Q9ZMje;1VzSFg>e!^s
zlSb6!^50(=Ffe+sC`~-wzdF9laA|0G*y4!Ds3n*sD2)l%nxR^AvDS>xnvq&_sn(2D
ztx8yfxU#rK`ZxppwCRRK$cRibmQwtDITV6)2AyhRX6`!Q&E3P(%UfV-zhJftch>(x
qX%?1N);3HQdycIgZ?1!Ip0jAC6NB$4Hl_IoL}MjEJ`42%F@FQFD~C$}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/quantization_operation.hpp.BD94143B871C7D8C.idx b/.cache/clangd/index/quantization_operation.hpp.BD94143B871C7D8C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f61c5b0a61b6194911ca2edc909d95500c5922e0
GIT binary patch
literal 5448
zcmY*d3s_Xu7M`^|Fw^l2W9pefXSjzMm6<_9;!%JkB!;QiQ1O8ks3;=%K*26**Serr
zNl;NRkrcc5$VXiz9%5HC(Ja#**KYVA6u(cjLQ3&7%k=B+^H}ba-{)Tbv)0*b?X}lh
zdo3r9PE3p!4Tgl-qn9qoSTd6{7z|I+zj;e?vl@Z^+hi~--S)}kaXGD_!o5q!3wtb|
zdpNN9ZxJ7o7Yb%R*SD>-&1cG-Y}bdoM(2|Ji*N4e@)<im<+r)3yJlBje?4iSGx(j<
zhfAlQYT0<AvOV#Iw3)N#y^)q#`{e!C8tOY%?2P;Dm38&@`(K`KoN&>9|B#FkPm9t2
zNNVo$eL>XdjV0^v+}a&|E_dD+{ISxhV*iBx$DW<*inB~udN_PwN<^ah^hfEF!5G@M
zVb;XD8}aX6>h&Nlt1LDmwZHGLt8d)pZakdyO32KZ?Gq>c)<(WIEd0TEtm4MmSuf1F
z8N21K^~xIy4z_<W>iaL+#@`%%;Lwr@hrGe}cfIq(+~$m)O-=hc<Ida|`hcvMb^7=2
zGY|6i@PEwMe<<^P<Ayg+{r9V!^zJJGuib2Tv}{pW^4aR5*@gGc$8@he?s+kJT+rvo
zHdRLjXDwOW#J}D@qCR~3y10H-uU5Rbcw0h^@z{)U|Lj%$VdPuEGu|qg+I%T;*;~G^
z*`rJu^D=u3&)zh)yX{7=`Q;18^v?FCtu1z6C|Y;yVq0iyc~-}F)5kA#&0GJ0=jrNO
zJANv-e|YV^EmOC?_xzjxK6EnkUw?e(sjs<SJG*w%rJ;j=x%*k(vIXg@d<+I-&*Q6d
z8Y07e$rCn%v=WRfLDGof2?Bi#29Tid&w0~R_D@RnB}OWA5R3;QP$RNXFGa}z*36AF
zj8j~GWMF_nx()nouxi9WgGBx#nev6b*=3!+q%YlPlyQ`#Eb<<`l|2ULhsKxvp7=p8
z;<Ov2D&VR>&<O6~>3zlL<F)1aROBxPDcPMoR9TQ6{_(l`nDiwtQFcFr6yy#Xsw~Kk
z8T)QPW&EB*E8+YNQYM$#Pg#&X?BS90=Z<do`Vt>%puf@It}M(GBVHbFyL>fNq}$W)
z^SC@gS@8Rj!5!9=prUhr#7ce84VG?jXymcT^ozDJ2T!|I14I525+unBvR|zGd&9%G
zKZ>F3{syTDxF+fWm7vI^GtXx(K9C$sl~5a2SXMZc<uUuK;r|P~RKC(mc<RFqz-^#5
zs02l(T)4+oFKH;HJ5VK`g5^^xq7oFD`jP8)zoDH)R$`(eHNe$CfJRUx%_nELZAvsy
zQym7WkZ^@W(8yzv^Jgane<M9<B?1+B2;4((Y6L}c9k<Wsgt^kyT}lX7LU&OKiYzPs
zCOk3WXDbz<UTp=gmHI>_D6;msMb(AF6Wi#CsG<sDsUQxGph)r10n2~w3QwRSG;q=^
zX%1yU_BwlN-m@R)B&bR{fa{=2RDvR9eA>%T^_)1}mzZgEd<n)cVSq+h2=CZ-yXC2t
zmPIrTsG3*#SFOr|d+ZtF^VhCDgTD47E*hV4e4I^LkbR)6XU>?Pe&0Z?pqIB9xXsiG
zm7qw&ou@bYl(n9u=b$3{z_O2Co=Q;U@S5<>tQF=%zN7~=G{czTP!`!pTDxgNAKkX$
z%!r;o{!|UU-8$gvs2Y{v9?fE3M^ME(W2vE3&0gU5f=weRa(YzE%^C5H^mavh)R(xI
z9Lj?1^P$J@o?SMeuOAse*{OVLA7w%Iw`=yz-ucXbPWzIc)W`>5d;kMAA{%N^3~gOx
zGmd<_GnE=j-4*4C8mKI|$JG-D7Kg6>nntNFWji^izp@~^y>iGmJ;t}n#eil?HkU0Z
z%VT!+j-;mg^*>Ti(y!w!@eXA{cK6!Bi(CsAT%h*Tu)72N9q6MG6!|4{)P)_c#v$~w
zY1nNc{1#%>2#P%VWLtLE%fB455<j}jx4?Z%Yo|(3BsXJDUY*%s=#l&Su;`OF;+%pM
zln^uqOPL^rctb)*YITF6VId<JigZPK8A?t_9)lH<9yT!BjXt1-%93Zr`bCj5&%@Ai
z*K#jIn;~#BxG@&x01Vp<(F|9Cr3$QAtmO*rtOAc7hgU%i!@I$?8@vqHL0}!YF$m?z
z3#)@@hMU0B1Xe7Ca{Y5Qfk%(Sn;?eaQ{XxUUWQvCuoc``Hso3n)(X)KcYvh>teEj~
znmaqdqsQSL5W{d6xVpg0a3KjSByOyKa?uMbB+(3)5K9TMVr7%dnX`m=^f<hP#4ub=
zT;;@z7@uUqYY*|6_s4EMux|e#K`Y}ZV6NYL+~c#TNucq<1qtdKYcL2>6c?pK$y~Ay
zWpP<LR1RD@=<s&nw(D^%aJ71T6u6^$d>OdQdVC$Y>v~*3xB{ZvznE~v3}atGPpUQ*
z?4;_8g96xFaDvnVd<$4Hs|2YX0@}gL?)DZ0++ugdYaIxdZtOs)M!Ef9`jqOWwI>*;
zK(yWi9gdkr8A{`!$z!6@pXG)Fn}}xue~nM?3OeE^NWr#X2TBQ2ge}6sQZ}}CtPRVY
zAf?z+9H>T+7TOj%SOPDzEpxCW-3qp?;9yC)3v9c<fwl`$BiI_j!IJ7E*iM3jB}qHj
z+QFfBob47kSdgzFwl&1Tg1U^@%7_Dl6dNgeX1R;P7YENy{VR7*%)fGNMkbztek*rS
z_$@Y1$|c9Nz$kXfSHW}@L}qIzm^wj3TY0U;!b?Wq(*5Pu3f1D#>Cb9Ih0%o#6;+%W
znK=TFCP)iR3q<DrrKY7Kn$9aN7g|0^Yq{voYNLe#gN+v5xdx(ZAQ9UtK{^bk!yvNj
zX$Dg>h-f3PwPR=^en&O@IWzEnC`0YQ@L5gMS})2_`z~Z;jC;PRpk_}x4a=agpePoW
zgI$9?3?+vpN3lSh=9=bVDBF^)^Gl8^$HTm|87!MYWMNqeu1fGQTnCmq{ke~T>j-$5
zA6miEs=w-M;JOALcJ+nCQb<HRwjdP|R}t|bo?n*k9(ip{et$vo*gbA`oWbH?CqrTO
zFgHetAVrFiPKGAiC%f5Er-{>?49&OC*FC5?Vvds?Jl~%0#_$lN<>GQDLtDVUMgO2y
zf>;So=9Fr%SA(17?rsowgOlL~us7%*%p)Kk0Vi|X39z34HwKR&odWR`I2mpOdz=12
zyawVma5AU<0QMiijrk-<T_ARWli^jwzDob#6%nzBIML~XR7&in#ErE<kjjZzPMnC5
z(oj?zdkjT1fHKr}9y1u*d73wmV9+B7!)O+y)x@+~fB&0^X%i7q1+O*zr~#Y)9)gq<
zlQarDRzaF-nkwSG^6E68RdYNwHA-#&(I}*}wyC{8t8QvPhXObP$V&MnD^r>hkm6-Z
zmDxa3ierI~Ah~##4f|U9as&hiv$%~2hzMqp8tWfB2wx6f86&XNVPw%WsKWymBpe>7
zV@ttZ3K3Yk1gQc{75c>838tMOVhr-yh=Chn>7&k&=L_Zv{cF%Tr~TYW{_|^QT58Ne
zHrAkoa3OvSC7Y9VyexB;4&`#WI>~Y{mxGSG9n9PHxE9Q{dVCbjNA>tJm@n(`bueGo
z;{swXAiDjFiMg0z%s1*)bymXajI)yBRlXE#%&S`=U@HXUeF)Mn2-pR|%&Y2Pg<i!_
zp&IlJ2~JXSKn+R?K`B9+Vw)1k0!kUNw1C0{p!cJWR_uoU{2X8aQI+y|g%N}iNPpHp
z-B|AE&+<5hQP8eUQgZmZT@m=PAbCw*k%e!#DO_aXJIpjpMD4sXd7*Ngyr@50U{DK$
zuuk9#*seez77juB5o|v~Am%Ku&1aa2IG-tXgntB9R!7#5fHlO+N>~{QC}TZ1uT632
qMl^t)`_H+an=yZ;9LagPv+|U{3B<Uj|J_BNnYP~3Yk3<48vYL)TVw?Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/random_gen.hpp.B234A29FCA71839D.idx b/.cache/clangd/index/random_gen.hpp.B234A29FCA71839D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..217b12d2b98d2f086f8ad173b3f478b52edfe8ea
GIT binary patch
literal 1046
zcmWIYbaUflVPJ4h@vO*AElC7283cj2xTGkPk&%Jn8xU8_ncBb6@34W0tL<me{1X97
zI3}}{ZdhO~?OcAkH~vJh%<WBc9Ymhq=WGoMRAgMdw)(tn|K0Rr#Y;~vu98dK|7611
zr4lEWt(wBAv}ju9&D>2pg;GrSt((83f1~xr_Q07>XZ_{>B>W)e+`~oNZ_g<&wlLE(
z`tU*Wuk07~6TwFutvOC*U2m5C#mrXW*6pY&zRrzj;+)oFtQj?-o1$zs?GkP>O$jJC
z7kynhfAJ-)ODFbLA5@qcn${&!`L^o)A}fh0vTWULlh>YIF!8#qypW`q@c$drUcc~t
zZTtCs{?DZcn3gF<>|o7uSiPh{DPnrq+$3)GwU;a3>@L`tnxEIdYBED{Wo}Y3Fhr^%
z$~6l1&UnYez{$?Qp)ISe111<aS(v!E*%&z(PkV`}-#d9Z52AoemP>~NL@;nNGqQpe
zoN11E_Nwm_4@7~KtdtH1hyW?z0V`lAN=++HVPat5TJzzo&8&aV_;@(DM7U(&CUbB(
zaCsnjJS;psNIV@L69kW!g%?SVSCUr|At%Bjf+Qy;A|(TN8i$mFlm~(*%OcBz#FLa&
zMDUbZl#%4LMYLt$LByeLqwNCY@ortQUxT$^FEh|H3|t(*pacUR4k-pHB%U^dHk_v*
z@aFLf=lcmjExMe#{4kSvIP^I6ka+r>`f#3}*qK7NU7H=5SU7mhc`RU_W#N!9m$863
z+Ub{B&6CNu9hrDIWLRX7TrZ;|V**o@GO@F*O0(f4UXx%>V&np+f(`Su_owcfxUJiZ
zx#eyDe+!@mT)JFH(IKTP1-D>F+JiGLT&Es_0)&%G7%4zFrI2{qoZ4_6%sW8qKxqnQ
x9n8r<VNgnf2|syYkrQ)-c|Kkr6s6`Qf~+nsN-pLGF*pirqZ}R9GuksS003w2Z4dwe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/ranges.hpp.66B5D2199BD23CD4.idx b/.cache/clangd/index/ranges.hpp.66B5D2199BD23CD4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..737d3ebf8d5ed6da0e74839ba20633a546450491
GIT binary patch
literal 2112
zcmaJ?e^Aq97=QQO+SvAm&EXe@J2tk#HbECV+F&p;oM32PHxe0fnkeW6Qn8U*X!(Pt
z;I42HKV<4DnWUYaAQ~M0q&#eDpoTw6&8|)ckyke?wO{Y|s}DQ=XuI2Ud!PHf&+~ad
z&*y!Ymz$H57KV_uBzJ8^X;m=?Av6VlomGxklUWF<*$A!mcIElX`jW+g{eKQLay<HT
zHjd(Ve!?2%{8sy%?e5iMp2mNw_gTC>M*?EO?DH2tTcK#TXZAkLQ4|G6HvH_Ho9}K}
zu)@^oxfV9aE06!2ENhC?XgkjC9ADY|-jeQbKUsG`bJ<_D=>3-lTX#nuJXu%2zVG(#
zTMz%87T=d7AI#sAy?FSzcz;sjN$b4RL+Mk|Kn;7g^=n;6=8-#8+rlN?>Ovwmq9QIo
zqoLX@k4-C@OMS5P`0AZm1CKV2Tub-mE_2se*S&V*zQX6##!&p3)>TXQk9)0e)L34b
zp>&-Zu=TcN>aY5ryL9OIt>WzkGtG&<sx7|K17|#za1*+|viaSb?!=~D3DI}S<;RNL
z!^Vh(U*3CRwRBTpUypL{rh(1{<u9Js{<bIPf%C13WiG6nR^`7kZ^texmyifJq={#u
zn1Ha<^k<bJt>_1Te~pl!1c;P#<<U$){HO3<#3EUhSyA4G%I2Hl#1stBN9p63fcQKD
zp)d@1XI*stj6W`(AP6a+DTzo@K#I|m#nx?G`*MctwL&5a@29wwgbBe0wW;&_NBez$
za|tm9D0xZ|6A+(^8_*1P-*vrx%yVf-NKC~5DOW0GLa;*nFPkH)3umc>L?i}?xFRtV
zf`Ctd63seXS@t8{f=z088W9sfgO0O3b!J;$xR8j!`z2h7m<hrC-?6$Lnz`e*g+w$4
z$hfjdCIkUvzdCxacOBS8i;PGE&p<IDcm_x53g<a&Y2g~%6y$5qW|3s7FjWe}k(})C
zzW%`CAV^Zj*NGm>h_H|*xhWo2RA5`&)kZ%LgG5S^22PE)$aQJ}K}D8;88Zw=_f1(V
z)m}3ga<9>IBYUV|``nPAv`%I0;*u?oFi4`5#DjL2#q+xPz``s~Pjc<e+8Yv-ud*uY
zj>X*!2`Xqy=~_-{@<Vbe7rq?0@@)3zkd?scSQ*;ppbYq%5HP-mG@egPsv6XMvoONa
zV6`go*kn;M75HJzUt@<?HYMD}>FL7hQ^98ItNf!be@Zh>OBK>&@Gsm1On|-y&<Xs{
z7BIetH2K(~PZT9dy;`3HT<bR4?v-|aQH~qPr^yrHWbTdT9S1A#m*KQNS#Jb%{ZZMS
zg$oWjL#p$UQO9okIJO3Zl+j8hkRnM#tU(L*ZmE^q5Bm>Z=3)oV5TyhiC`ptUf=-}j
zQqRC)u`j0UlED{vWz*S%D^sRv4y;-;RH7Dh&e4!GK})P5{Sbgyz=R`UZa$>3{${Ba
z_Tbqi$&4^NEM1B`6QqOD;kxwm00rPoI88qbkUmwLTftInI*qCR|4qE$y2?FKU1G-|
zN=<2lf|ycFsc>jugB#G#3N-lsy7wzp_ppP_nKTW~B%5(2&1NL4ZER%x!&@OajV82t
zN;dEoh2#X&#y5g~$$%4njPRca5J!&ERmJq|;6<DGqXF&+JT#F2ktY(+$>~_`%%Cer
fOq56A!JlMfdW`ps<${qD9w8E<iRTeF4}$&!4*z+)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce.hpp.7DC5A811A95F14CB.idx b/.cache/clangd/index/reduce.hpp.7DC5A811A95F14CB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..44429614138e5dcc9164703e873718a471c31e1b
GIT binary patch
literal 1156
zcmWIYbaSg=VPJ4h@vO*AElC7283cj2xTGjEfsuhhfR%xvV$RW^+gXPVc+M_=!hESx
zI)^!!t2s|uKUcGwf7Rk`YJIsmOJo1lW$Nv1GrIBC^Hb*{W!bq28b;?hx0>r2ipfqe
zne|!3JV@dBW2<w9x-5Ljwrow`FC~fCaTiwja9i0tTd-MUj!5O_>k-dH!<JuOS1)qx
zmuq6*tZkRVEAGDMnQ57v$l4-&!L(IZ<C<n&kDH_MwJYx?OYW?_H|fYN;f%)h3MY0L
zRXD#=kdL_^wXJkc{mu_oKR>v*uPTv>D43wUQ2GvE-r^5TyI&uC_xtMJ71b{e%2jVz
zzFI5%w*FG5a}v?BWp3Vb=8DUkVH#b+nGvvb`aI6b(T@`s8l1e48RC<*uuEj-i5ol1
za#Fdkow(k1<m&7@Yu5hNwab}v^=L}^M(30KhT<wymrT0)`pvIxJjZI=9Zt3~hsz0<
zUii7n^Z7Bq8>jV8Zu<CmzW<ipPQ}kJuYRxnuXvF%a~)H+(z7qcmAOeZ!0=Z17G|Ty
z%6EyIiJ6^&*MQ4F6ihI10Of$7@qFN_$2Y<)*qB(@7<r}GrBuNLO!{()qPFJ3joZ1I
zn1H%fxm1O~1Wfu}Scbt{yVq-km{@_*)&kZtU;-vxl$utoz{J47^=rYsW^dDAMjl>Q
z9anRhE|}ApxWG<=IfaMUR>W2wrby%F%c@Np4`P9O-E`dGdi7f)BHNg`Hv)w{IXrb?
zs-06V+%&jWl+VP&E9D{;0P!OO0}rpNsH!5&Ga-IL>MthRumbhk^4bc+^v1|<5Rm7)
z^cpDaqUB-?6V48b*Sqv7{yj5LSkOfRCJgk7kgFtu=O*L^=M{*GELP$2@dTQb$dhOT
zQv~xMBNsS~dk_6yu;!t)Li2PB*)xoF^+45*+KwhL)g3QwyU&Y>W&nkq<eZdY!aa?a
zvsM>=%K|ya%*7fe40Mi_t386}X5|LwZ8B&K+pM793p6Q&BSZ(L2o@Yb=YS#>=DOQ2
zZB@Bm%Gv@|J4iS<z*IlG@Y!}_{bDDeuz`euIxOyB(m+E%(GN4^*S6^g^{h{C2C4?7
zDOf!G{=IC!{+(%GfWnRfj+!vdMX5Q7ASJ~`$;Iv<1~hSkm|P$N0ytp|1}5Id$@RIn
Vj~g<wa&vI<a4~Shr5Jb+OaMK2q3-|y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_blockwise.cpp.E231BF6C564BCCFD.idx b/.cache/clangd/index/reduce_blockwise.cpp.E231BF6C564BCCFD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cf0ddca2e40c92a0e44e6ea1f9e00f27e14d9422
GIT binary patch
literal 6846
zcmai(2Ut_d7RNI~M8FgXmtrUZ!O)}#7K%hbLPQn?;Q}fsqJXTT4^ctHhD&Dw5!b$M
zXu7&K#Da(gETF51@IYL}RZ-DrMW4E^u5Xg(<c^#7eJ|g~<o@!Xxij~inRDh2Qv(A6
zjtW_<33CI}ljE1nX0uo<J@l8lWN}iT28;Dvo5f1seuSPj#mYnX?arQE>pR#{%QW+I
z|1sBQx`m#bmHEDcnA^$cliz)?X#M!XF)hs<IuG$HxW3=MJ2Guc%2oHY_#{(X-psB<
zy8e%4tfE;nKU(cyX!EE$@?NH=q-Rd%O5Z6)D~xIiE1SJ9w7xRGKm6R&w}}b2X3pI4
z`NKBe=~dDB=Wl1)iwk>tHXHk8?sPnNFJmCp^K^sW29tXyhreHy(mo_X`b;?D2}>ty
zmF@In0nTT#dNj>+9&ha$6eXk_-WUcS`sU_tzZpqoO=<n7XLa;f4?expZ?kCRD*lCQ
z<8Lp#7T_tTPxNX2Fwalim?l{*y0&P}zT}Ph!EeG({(1cm&Dye~o&#%YstzyB3@kso
zzQeoD*Xy^#4>md9IW_P!YFFLGyN>n;LofUGjo6Y|^EA)@eDieKwAAe3)^2BP7kE#~
zvyQOkrsq`|Y087j*}WOv$KPw^nhcD5q8VuAzjN*q$y0yIOxn@Ea%=7ljyTo%V&vqA
zF)0VHE;QK$U!?4`4w%u$P2Y1-|Ml36i=s30L#CI73$#y7syAIX;d|a%L0i!G0k5{t
z=LT%J_v3cIBBzBjOyB=7tg?J=?t8bR^HYV#_y1D)-P@Cf<r$v2XM(OJTzqMKZM<$-
zL|?gh?8nKv8;()C?r}HfMqf5CD_N1Wd&jO1uQ$vwH*2fBnbP`v@5x5Gb8ckUKF3+>
zj(MEueYRtSPk!c>O9JmH!2Xu@R0~<E*E)LTg|wc`6sx!n;giij)(1qr%H6zcxJ^=?
z!=s|zYd(wX%BHtG=EZBR-P)BlaCNz-{{aVn@(#<g>ev%6!WzS6E7;Q#gV^hLy*%|U
zFXh{hi$^w8e2%V~J)>;ZkFKXiP2NAT?)5fj-~GFtTOBJ+%-VRyX+_5li8Ep4*80A{
zJ$2(9roLSMaZ2%!t-Kiq_wTo#&HKTQem)}dsq=-gPgic^jP)^}JyJq-1_$2zd|`QZ
z$Qx1n`IB8hR+~J{mm899WAfLd{&`D6MjUbc_V%P!r-iZEVNbdbHVY2u+Hw~)6`O9`
zb)&|;GAsE;Q-h)Xr5BD-%ersxJmz+ORX@io>9_m?O+ioIl-uQ5oLCq-Ib=>pUdzHk
zhssWRIu=Zs)VfPvX!`r^zu5W>@PYQyw^rhu#9vDH4wEJ?x%GC;;@&|wJHIpPk9pTA
z9X>qSDiH=qBvq>xAKF$ds~gX2dn_w9DyvzS^H4*pE#ioa=eZIekBng5t+KpF@9c}d
zbKf)cibls<VX{HekoeaXXQRB%7aiW%A2~0UJMVIaU;q9Wj~|yB`*q)5S?#p+u$Sat
zg%hXf3=s_cSoP_Y;iu<U_>V)3-iq2!XFH@WPmU8o7K{6WtG%Rp^LQ?63X4To0Dc89
z`SOJ#g7RbTF){q=zwWH!!I@K8w9r^+sVuN8joGqCJn)l@2jh@zq-8`Y3oK8(S@S^>
zZ&}A>#i9=Q27F6pf#sN{h!UaEw7Y|0D3U2nN?%!EIWOI%yXyK{His3D<n%%5LS><n
z+w6`^wB2}?!<vocYQU}rT3^0!l`X?W3$FhBDnJ(|qTyL<Tk9(eEFX2)XQR8P>^B~q
zi)0f66OppO@}*g?Lp3aH0GBlxbx;ZTl|cCA3s>pbe`8riTal0l!_ghgwatx{1(tj7
z=U?*gEPKX-2}m|IGBr~cSblVq^&q1bu(_-d)PWofl>@6UU%1Mv38_9&JE(jfoQLjU
zsclIq3za-QAwM;`w~_~EA=yI9LSI>6`CVX)V`09ZzAl`D<S|-f^pypc->?6vIb{*=
z4iAPQS)?V>Qx;feEsjq}{aK&I(pbF0%cJ?>c)+2{p=&wx$opb?&EA62#5V#LO%Cl(
z`3rG(99lL`=Eu+^Y7(IZdj|V3S_l<FXkn5tf2<8&nwlQ#|7RGRL&pw@)x$q?==kCB
z4h$s>Par69Y@#>TRMT7MZ<Crs*Mf<)ARIpihi(KD8o?BXTY!HH2w}Jb1a<(L;a(u^
z1)&UQ!N4p?=s8f91H%}-0LltTE`ib#NVG49vT{<FLz$f9YACB_SdePtmX)IChWwEa
z#^wW0ZU2^--v?T^n1+V$d^zxHi<)4TsJFelZJwT*fY<k69p$>lD-wB7S2h$Il0zQ_
z)KOr@JkMD`ohA8KK>bQ`H=w#nz6Yp#44bF@<Z}^NKSZM`1ym_8#w%UKTkDap_3K^a
za;Zo<g6Tg<6hu&{D3qW$qd03^)AGdPc~94j%VD!=sZeT$e`eD`!XSb|g`ot+3E~WJ
z&0)(#F%4_?2dNp#`uW5^+7?olNJtVS8R7ai7Y-U`>c)uF>f3G_`S98?&So{izT%@y
zQ~Z~)n&32LuFQ3BS|Mc9g}}5Bm}9Fb_?tp78S^Obb2>McPp33Ev`|-Qh+T+$HcIA+
z2gadSIIeKRP24wTF4Zb+nyuEUr}?W(^EGF`(qhwMftbPp6OpCZiXj_;4WZcy><DrY
zI1uC{a3aV>;6jj_z)eNIxldVtyjgM+u<3Qcd>t5thbkRo(S5JJUEMx7*J$V{a||<8
zCH|5S=CLD2M7U$Z`3E&oBp)2uIJ9K6#1~iM&=H~tRR{^~{CCx$sDYo<EET3LZA^~W
zEk;6wafCT;9X?rL)4zV)ph3tP+Co0#3=?e=K29(k+Em+=kFx=Cjd;Zf{Jb1GQ=F+H
zI#Ij$_$JqIweF+)ix2h`?1@toupOYzSILt@?*&?WfdO8#v(mROifuf3A9Cmp;Mjre
zjngMx0_{qm3+DJ>-^R6vYT{7E7GT!`oUxz?{`Tce#d(cV%?drgNopfX9+f<g-qt%`
zP58$zlizHXu=>!gzxFwY&IZG>fjhQIO1h-Y^wRZDdTd%aOlXCT%BD?*nOI??vT4&{
zrdHUfXc|GAb<h@%R>7X@pclir&^{MBknEicB@7or`$AGLgx-auUJ8emLMx_zDI8r2
z$1uDR+H55I-v~W7lKode`wH^@70{ysdNS>6pnVOg*Fc{d=!^B#J=af`<;AW<wkrg7
zg}@QpO2N|$!A#8Q&3F4U*zH%4^@{<w7zl83$awRnSnsTsLijps<Yr`yUKdxLuP(fl
zntGt!4eGNNIMxC;JQxn01)Z{>Y7XJPxjJ9%&=k$|Yk|rqiae{|NOB9%Zy~t@=y#CZ
z3-o(Q&Vp1Hv}1;m1C4Uv2!;!wQ31&%kY56=8ND1*<z#y~G?J6`tD#Xf!?+#L`P;s-
zXsyTAy(w?MMHb5jUfIA08=z>Q{_Kkhq6hyshU3tRB9mfUiWf15j-sMev4ulN8$^q6
zq>(QOVowUqxi=UEXIrs7z5|=K71*l6Dx0<!*c0R^a3siC;7pLKz?C334>vD-Px+@4
zzpSorU7$82DxY@Ut@t;{LxLQPkpmCBg{|5!+PBGy`$TCGqikS_EmHNyu+$*v&l;4{
z!#u)Ndo#3)Gl^F{8i!6WNgycEF>xe*>_gEvTq}zBg25bG%oXe5A#!LNt_?wUTswjs
zxDEt4ah(Wq;kppy#&uIs-T14(=4Wd~j{kav)U9^DZ(*|74JE4xp9s~?7Dpnqc~ZPQ
zFis5C4JJ0V#{hK<Wph=$Z30vi$yWe%h2$H6x<T@NK;38fc)-T={XUOwD8`~6VTjFC
z&+U<FJ*jm-W0&YjO!3brR`8Q}PM0cC?QFo#20ZMShH;j9eR*%cS2~Oo8U>bXq@^a^
zYDPF6^W94K7WctwLUa28BeOJ58445p(fj}=%j&x7PQcVUbo`nPdsergl_3+#Oz`;F
zbg)UVYP7ALGmN7=E&{bKRIg$>e|rH_gqVn0Fq^S%BM5B-(G0hMsVyLe;SLbl0pb|$
z1!27?daAsW1tYT{q31w42gWjb0i+8^E`ebskZ4~HQJy5~%ONc%xf;^d3|~4MDcN05
zN1`W=@QjdPvs}(tdt`mq5i2!AT(w$}mn8|~D^`Pxcoef5Y=ySw4A~3q333!V667p&
zCdgIjN|2kcn-t%rEqBwOtIN-Kp)v4zd<yTS+PjuUI#^s-tsI&#AK2pj(*Ce%(TVY$
zk964daUeVnOt8(^bUhH(liUo1%_N@#!gC~F1;VQ&w*g@r$vr^WL-I`^yh(B&5cZM$
z00<v2j9=Q(vQ@mXGs`wiH;f2EhXHjMsPa9BJ_V>#B)0;pmE=x9b&`AsP<I%<k=<>q
zxm4Koe>20ap2Xc%Ycwj<T#t7;JRRjB0yFHSAxXsB@qEsF0iH%2dI4tvLB|1wT!a0>
zL4Ki-YZz_@6mkv2=KzIV!|+u=A=fb61}NkjhI;^oT*L59Kq1#K+y^M+8ipSL3b_U|
z&LwCU@!zu$96J<?d|WLEz!eqyh_Q`8wKr1iRNY&EAES2w?+zeixEFZzf(Z;~LGLU`
z=sD0U2TB>e0D2XWTmn5xAkn@YdX<y99D2z~u7+OK3_qUVxi)<B%yB3{%7Km?7~>~Z
zuviX8VSdJM9zEOLp-s*A&m2R}$HaB6M?x9UD+6L2#-F7&9Nn94a~IuuzW#hwAVrSX
zU!Wp<h=eA*;slAAIf+_{`pjH80;nTE#2DcWpw5td8Bmu=?gCU7$#(&Dm*H1mj;2tj
zigPKZ(*&&qqSHD+)iI4Y^l3nyCixPeE|L6CK>d^C-vRYI!><oK+tj|IX$i8Eo{OFr
zo<R5+(L0jj5R4hiKEU6{Oh9}dhK8&-3}dRrYsC{Q>>%JDL}wPNjL`t72Ig>sLtg;Y
z1(MqV)lTv)K;2^a{i&Ml{$20?8&Ao@oKJaYVl-+GEkm#Q?jrXw_!-b^zNyTOp<uJ%
zQJBz)-2!wNivF9kcuwjp#YL2&P<}M%&kfSpz})3=gJ(>i79AB4$%@1pi#46Oj9@OK
znadf><uvAUPV(FZi3lgff1|-3q&ZkidkBZC%jZqh^Y;_$=}Z(D=~Ko--95a<c}vFo
zI2ky2T8*(4{jXmQO-#)!EQeVSx3RMy;p{TfZS+|4Q9?)82?mA%fiX}T>?;F7tp5N)
C7{&zv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_blockwise_impl.hpp.B38B56E3F9BFC70D.idx b/.cache/clangd/index/reduce_blockwise_impl.hpp.B38B56E3F9BFC70D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..080cb5f56f398ea203911d338e42ca530b3d5dcd
GIT binary patch
literal 4880
zcmb7Id011&7S9aB61b26$`X*WBND2JEFvjL0&d9eQz43A)gmCpSVU-Rk-9ulDHIh1
zq#}zgf}lW65dv+4VpS@wT4?b}Q7VfBu%bnaoxJug{rD^QyU91dduD#;%sFSyow;Ev
z0s^!QSuEk(D^e4;?217w7E2%hB)fLT1wa<7QH#Y&Eoxp9760>sS(CxHg&vXbR@yz;
zVL!`#%D^Pi@dM+-@5g36KCr9EUFQ!)IRD^J@h1D-)icT!;ugtU8;Vv3-8~D}>Jj&R
zD+U_d;*Y(uvMwwtI)632KiEp6BFm3|_*aYT0j+BJoBph}iQ6+=v1M4&J+(aiYeV7A
z&YPwCWyFKRfXdqV&!e8+i``mnz5Bk`#^jI9c0|64`g-%xnzljrG|78u7pG4@=(=c^
zy7=V2v-OU@F0)-~eN<Ed;tRjveGsHS*R57PRc&71S-*9LpCEo#EE`srpg%NCHs)k^
zz0wGenUkxytoW|^x2hEnuWR;xJ6Be?JH}WpUVrRkr?pSfhLrD#Pm*q3?C5)Pt>aAb
zwT`_WRs}DoI5{8IE#6a+ruhuXyaZ+M1x@Ejbw=3cd!j9$v<};q4wO0kwxe$L1m94;
zsBo3KdU<^#@0~fR9hu^=ZEv)>?zkMOZhEe^r@M9A_O(v?9cqf97w?kD|A&J&^JZr)
zGySOJ@!01nqc<JB)UvF4PTdPxZW9p>XJ>idpR3=v?Z|q~mfk&)(gP=I^Xf9HZF{c_
z|93FpOoD&9=JW}PPRhXJN~xqPB1v@rvj~H=0_z1w^)7VOJiH;14P6?s<VeOgl^*Uq
zWmGrVyg2TIvUw>2$(r@^>+UWa(04neuBOPo@k4RAZqL-69O-tm+NREgtY*y(UapF(
z-_Bbw*rn@PEgXuEWG5dR&*ri^mu()Z_U}Qm&Oc*bIURe!3p16ZJ@Crn>`fW#zBvsK
z>ie)HdlF-<a8^BW4v}w&{V|7N;gmyj&>YkM|FPH*ung1y0<4`|wn}6i+%;OZ^p>k5
z2N0+mT6aSyawt#)2Kly$2GtOf16Mg%L<$gU2aDQ)8;$$Hihi(?#A+@&pO@T-J)%XR
zGr;r=upk8lDh0e!@CJ>`z}zz6OyhFkP!9MsZUysN!F(Fafu|e@Xxt9G+JO&^`@zC~
z;6Y+F-|giFZR30K)cAq?VCwPC^d*bG8SqgfkjPskBp(x~AB6RT^)wy<CL^@98nw^Y
z?d8OuSJ@vLw`9X3^J=vrf$D)xJ?%&WH3EJkJs%yDp{lRyU$U{iNxn({Aq~;p#`B1)
zGH(Sq5HL^<gGd24NM@Bdl4*m9F94ehOyWAgucL{%Qj0{_<I6iWlpQ%2fHlcLpgLeu
z2dJ1UE#%ZQ<&D6kktvq~j+7~H0wztg+-P4$?&gn1qp+#v;H`2xG{)x=gr|Hm9^$Db
z`zEg?^_$#uN$WVb^hf--7xH?cEhV^Dr}e7x%Y`aC+ss|DyU;GX0Q=<M=}e&e(EdK8
zlF4HCXTDc~?EoUsQ^0)+W|681^gH1G&cx3E_ZbsU0`4S@t$pgH6Q18|t8A+6L+MrR
zMH?gV<o|6d|4@bOf3Nz#z;lSxjnk(MtB1Pvkiz^!ia*Rh8~}K{vx74gLw?t}Ra&tZ
z_%mWKKltpBQL@wd8TwoXPS==ik|)g=u}JW(_D_2vXW_iD<XbMJ4ao)ETwp}wGJwi}
zHjT>xjz0sp0&Xj_e?P#Bjgou%o%FjjWyP7H-^*hBdb~!=xCTOT260q`5Xy%}`Ot*a
zhfqB<sD})>9O}s-mzK++y&TSEVjnqNN@792W%ie!zMh30^AgN`Nf!{mq;t*|K2p_}
zyfENtPu}J6+Yq7CfP0!QA_$cLZV50WRR>(zKbkPIwFHN{60j=)hZGQ~6|!3)g;xiA
z7+>?x55+cSf_0f7itM<${rk{6&STfG^Bw`cM>KUf+rK$+3-=fv@fBo$MN>rt88|%7
z@m9w=E$u9+ydh97VCT{nMEtUG<=s<(J}T#R)HXvFfYVCFd<XtK(onH?p<$`|;~g`i
z9GzN}w5j<@KkhHK6;t+XS-rG5qpl(rmuaD`kfI@w{}O*HbOZ|V3t3L8*c!u|E|}QA
z8#4_9k72Ny6vXY6qy)dH^u=6qfNxz=K;U;7z8Uxtz<#7~d|33GT@%p?9Rj)Ox=|TM
zAa@7%H)#?Ndk<$CdFy&JdJ60Wj5nH4SICfw97GHje|vuhONdhlW5tPe#f))cfmlT9
zO9(6Pu^dNjD#t0gp?meDMP{+e19u%SUVB|p+c+bVso(THOhdPCsuRe{&WaJTQNSJr
zdZfh!QUJDsiN^qYjETnqdz{AUFOD`nN-<&KuqN6iE+J|5RR<=nDZYIb2ek@%Rlyac
zfI!u7Ni|$a;~KcM28NLMV6^w~(gPJCcwmx#l0Eqtznb)u8E6=24TCwP^x&Ak`Qpwa
zJiKb$Y~2`h^H}0RrDN7gwU3`<zxKoqaMW|0O^zkdDBz4T3!nmU6ihq@IAcsa4mjg9
zK9aHcR8{iic4ZbiyD_rDUF6P~wTH-qfxP+NF0=sxKfy8@37v%u7Eyp`6)p1@`7>BT
zR)tW0!|U@+u`AtE94U?_(UGzfy-Dv6{)}@X7ntPITZB(TQ-cmpj8Ea<<pI+?darQ&
z`KAr~IOcb-papuhzyOl!Ou{O(GW5W&2v>c-I4aK(Du6}>%&Ol24H}qLza8qe(^bDz
zG28Lskj)gEKy6@F8!#pLmDr(;xy|eXW?jINmOlVi4*(+*vY}r#WVkN}F3W*|bpKo!
zoC_mpd=J{(qhB}5y^TvRbSLF!NAC8OJ_CpF#KZi<sNz+2XkEdh_gc(xf=Pjml)0O0
z0{kX=H+O#1JNB<p-cdY0R2)i0;DY18nyB$mK7QxOAx{p?NqtJ}B8ODq>cl@SEIq<8
zz>4yrem-QRsT{K9kScx+!)Xo1hbO*Jxj;zu20IVDyIr|p|Gl%ha_!Fg>LpJfsm$1_
z=XL3q1?6vJ#v#BS0t7j^w!7u7tXnN{Dl@j1xIm}pg)cB;x_!Dg$#=^p{m{O*!Z~<b
z7`HNx+G-Ox`78<`W8>Tac@2<pv}uP%?T|;(bb8yr9~NwWLS?gVFVFri>*9B2Di6GC
z(WKt(WO-wT&wGqy(`i1IX*k@ua7`{`9B_J}M=zx6X`g4O@cDTYpt9<|jk_Lr`sL`T
zJa91n$Bc;ljo!Fi@{Rd+r1LQH<}<3dv#~Qn78)2DN%stM3!_eO!=Khn7CUv8V7gBL
zzP0IYMF15saWO!}One&PjT)nW3BVgQ3X?oQ`|ndUJ~b<^Cz7Vds#J$Q?DCVq@FbvC
z+(eO7How-&R%M4w9$0bN_WH$VGv2H(UslNG|8nAYtSS=-G68is#OL^g7^N|iotq_<
z$_7{Y_xGhuHdrh@x~V}o3A(9CH+ATyF5P6)P5i0qp9>;EY)amTKl562^)=X9n%X*q
ZE{Ch9&Qd3Puvp5E4F6op4_^OVe*ql#cX9v#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_blockwise_two_call.cpp.C62D3A32F608E374.idx b/.cache/clangd/index/reduce_blockwise_two_call.cpp.C62D3A32F608E374.idx
new file mode 100755
index 0000000000000000000000000000000000000000..22bf457f3ce3e168af39becded0e328554a0cd10
GIT binary patch
literal 7212
zcmZvg2~-rv*2m|TBir;apsd3tAmRwJC=Qz<J1Qap*Fd|WA|j}u0aPO55@i?MAc8zZ
z0Y!L$iHe9Dq7g(1h`}vE3{hNx=)(n#uZH*4RC~_HIn=4&{Hv?0ZdKiTtAnTe`i?f_
z@w}E#Pl#TzVJV--<7wjm^&4VWg<@zH@^}gP=jVk6nmcL@6h6D%WDhgqt@<xnJo8;R
z&HLN_2<L%3eJkV5&;30II^HGN@KME_1$&O9f1mle;6{96l&7SqU3+JYXHeIQ??QCm
zEmr3j9n?+wyYWBwKAcgVT9_F)z%MLpE=@Z>{Pxw>uEUSI69W6R!s5O^G`!>B-Z#BT
zj~)#!9(kcAFYv>?us9cAc)EC{;KKX3IM;$swqItoukPskO);&K@RPBVej2A@v2Bs;
zTJ)v$OLyK=GRocX-JjbQh#y(?=`f*VqP1?%`|;=_?ZL$tZ}%M2-Ez(=*F<xZW%}hL
zcH#C6Cse+{BV)lf)s4^3zxi-Z>0-wI(<=@?^vml0UHevt+1j>we&S17PNy?%VoIWq
z)~HAQdMhsM!?W85TO3?}TXTQ!uS)xV3c9iAFY6x@f4<kHr6n{lTH&LuWtL~`>+z>Y
z>h8$lZ7Q~AuLbAZEM}tUy7GmITeok1YyVul&TqF{V^5;q96cl7*)Kjj3#(oSJPFBc
z$+wO^QtxuEVxij4SAJiWe%gI?S9VrGpK^j<NB!!epS>#@j`yu-+VEEWjthV0<na!A
zKI<z_*@+!rKeKP>d3-u~(Hi>+GpEPxb{lbSOlk7-HkX+ZmaBIJpNmoXS>rx~MDR(R
z{j+DOLv5F`&&)r<@>74CCaOK4Z<OU4U6$|n&$CARjLav+_h+=emerm2j;ywNAw95W
z#~J6^TdzEHy<tOMDw-VNvCnR7u)E)SXF+*wy69eFwVCJddM9&d)j-ve7rBA{vw}_L
z-_pN@J{4S=VQpF(y3dVo9n$$c^l6O7d7JYQCzxz&rQ2KL6OYHe_P9PS)N^;DTCYo%
z%GG*X9hvGsD|0@sNf~@2Dfa)fTW#+4&JB`owHG6PpHTZ-|J$wK_(hDgYVfn`Znnq@
zIrdrNfBEU3NwLj+8}8<0WcewZN0nTExGMamqi1j93m5%|g10Jn-|tyA$wKvu!Bc(o
z!mTc2WAU)Qniw{$RNPo%%2*|Dv|Ro1MNh<rO9$>(&u|RRnJBUK(K?goXc)*W|6^i(
zU2>&lPnuQLHi=HW9`h#2aaZHu+LgYUY<-A-&8qn~qjn?zomq=@OmgzRZ!OO&D1HCP
zHS>_x$_1Y9mM53?pZ?os|D8HhL9aN->`O;|j(XxB&Fxp#T^d<(od2z-mbY`}$^hNa
z;3e%BCl37asjh8i(s=u`^_^uOCGH<S#vDFpYI`zj#O(1CXD<8Zb3)lx@u93W#UFWt
z6*sM4-l?h73D{oo+EPC-$iV1K7Ly$CU4r1_A0y2|^}UUM(tav!EF3=GzNet}#(!Kl
z+BaBR?Xk+tIv9WYc1NJ_)SCzKr{|v4*m&XIo<{3ImGwpkO;)BHTypM8UqkqVeye#i
zQUX<1B5xPB2G>hh2O>9L`sc4c!=$^%a~sNquj-(y)iuq)=yB-3?A+@Pyl`}X)OdA_
zP3KmPfy9KmlRI{AElzzjWk&3y1+7)FsqLPjIfac=3c5``j{OwETfZ?nOoXtr={mmj
zfr)RIkQc(^vGauUG=~Z-4E?En|LT68=`TA4=o>6clq7~j1(8Qa$9_oec@-w)g<^R&
zGh2J8AhOr>r|-5btnU<{U@V&{nHdfhL=I|sysAQyP$1;Z!?Ll+*kGt2^4x>(N0qd+
z<qCOguslaFM`x%Yas<1%qgXeiTYv(vY@lSIIaCmNdqQ_;z><Df0h)#7(MqF@h6*C@
zZjqdtHg^AIhPM{W-}1jz87hdJeJbME^IdM1Lf(8V8|fRF4;4htJ$G!5!yEByA#WX)
z=QHzlhYBK}TAe@8R`E7gfP%2>rsQTgR1mqNV}5v#)zU)(G#AU3N|qys3L-z;lHTY3
zpetB_7GT*_$y9r&AaYlZy>q)}si6SP#<ISW{>Y(%$bT-5T=2<1E=$P!7Rz&mb4Lsn
zME*R#k1aUy_TNI@94woNOvFP4k$JHz!q<CgVf?>+p{n-Peqh+0z~2cf3iGiAAS!^7
zl#hVu2$xGiRLbR(AUFvk+WrhEp8)~ozhL%XFrRWB^3OwogooJ+ju$qErK>V*1{h|5
zDG?Yp6AUxCoCOA1FpBEgFg6>;Q_cZN4%kqxfN>RIOSuXhs=%3YH8@mrSq2ki;7Igg
zv723;uRNIw40{puE`kXWlx|&o@p9ZxYKUQrz@iAGL}1urur7uvgq5|nO78M^YU3V^
zq(-h}2a*$r)V?}-4Eq7JKEOEgv~Z=^aaFnDCB$cCNL_|Bh#+SN8B)AfW7b8}=$G5g
z@p$sVBA+{+!(e@wjz{CHarDIf(^{0|4K?#&H0hCH4};!eu%cW9nnhf_81#y{dI@Nj
zaP>0KE2BDbB1&psIS+<ir?O6yz9dd1jzb4P`2Z*;$*?V;+`{EENcjvkf^I`<x>@hW
zjd(8C!TvgT(v4u>$mJ%mZ{qR|u)o3OX0UAr2Rf{qV0)9Y-c9>Wjb$F~coEMcjk8Eh
z8X8XTXzqEc7KzQC4K~>@nFtJ<1GYKT?8fo`a{mQJKN%XT#&yu54I4qDk;_e>(ZuB&
zpmBrC&7jfD<(r^!ld|y-RVAHHt@^-c6Acm-`_)7}**Qg}xeL!A8Ac}46Xam>a7<H(
zdOQ|PO-&t%mvWfWpmuj-h!!4ixK6ks9c`peB!{AOqBxWd{A^I8ZF7L1!{rLlr~n<J
zOW%A~S>hU3f^9bl=7ZeT*o(~f(!K3s+~2fXX=%TJ&u$TJQ8*bsn<Px)k~m2*7OYt$
zS&}#L*<R4;1wG1XXha%PtO1`rikPEFo$A#{y_)VbK3k0@RU?I0kn^l+^bKt<L*rz~
zn(&m|QCVm9e$m1<ZG*9G)SXOyb?0?u+2J@m<x?O$1={2ZhOGgO8ahHR-#7R7)%3UW
z`Rpzb?*e1ekk6)oIE6YyKAQ^SRIZ)|;xsTK?MbZWt9=!t8P-;6JCisehOGsDEhu(1
z!`=m_yD*D%;(b!^SgSW_{xF6e0Q~__n3-WupfM+qgfwK>ax|u#vTxna*HJIS3-NsS
zf$BbPfT%#K6-cpq)7s0$(eH1E;$52z>dByR2?35dYjwt_@4=?Dlv;X_7#k4isC(Tf
zHwbrZA7np=PQ*9JcMgfu0oQ&Fx--7-`Y=9Q1?p9F=i&7O^=dB5KwU;xgJdw+X!7)h
zL{WyFAf4bz-{mLu^Ca(L*!$pgAH0br=FA#l*^@naI5t_hD4es-k)lWrMTw#mL=q$X
zR(V!L-x{M5!->T4D)Ai31yL?txOwl#SH6rWJcg}hKG{5&EdGMeTf#e**?Hi>m<!F#
z$j=OGA+#_f85P4`0F?`L=a57U4@aJb(cz4fj`N^Km@1v>L5*-5oNj{`F~X|Y^>Kmk
zOWpB%tAJkxLLxAD0r=HimH}Tzd5!J^e)QVly?7q4K=&1mhcO9mErv~PpYf80TZb#c
zKA(-Wj^t33Rg~f|CS2+2-o$UciE|59#464_!!8yr=Fl?HG7en^)ytsp2Mk*es`XsH
z0;*TI+yJT#T)qmbSGjx*RIgFqG^q1I`pKjgPp=cKI>C+DWn0C}=(z>0zwjA$8*<-9
z6C{S+f!ud+`A6jbBbW2h)O@5kv1C#4;>mMlx_CZPp8&dC=YW3>49MFV_8vIhgFwQ&
z`g{zXThhh&km-x{N0AJbVZ~yxf=FXf`zq5;n=x$1tE?|hIDT9P(Petp<hB&mbNLF0
zu5h^lL=9ZN3ZkoAz6PRelu1H0T<xn=9!E1LQ%@4j@V+-UAJ5tM<`(0LiRAm<_yUY2
z6WRZ6$--?S!#jA^Cy@IIG?NH&=jL8cncNoeo8?ypYU5aySk4JeP)Xp>2{i8nTFUKm
zUODB0m@c8kI^QYy-7+MWAw>j}v%3rhkOvRfy^9N3m$(n_pkx@63<?ii6y?vK5%j}j
zB^)b9Zy7_{<9HaD6hx^Hboywt;j`&zL^?Mf9z)DAZakEsZ)7N#Jb$EDThjg9{G&3%
zR)SU~=o3NCmX)A5d-5=6aEc}y@-Syq17@`S6Yzcl(<%3WUk`9heioUWMT#8tsItuu
z!iN|B!P7SNG4&_Qdvy1LyuZG?V2I6D0v09QOri{|%V-vM^sUYwZT~@e$`P*<&xv;l
zV-pmEK6yj;q1Trrd8+gl#=WIE-M=MI=7#nMS>t598i}isA`Hsey&8FvY5dgb7N4gV
z7=e#*1=v+khbZ@hE>)mdM!6q!sphf_oMhlnI;vvtPB@Y-8i70N1<hU>p%^v|3Db}w
zOyXMy5*Bm08mU$zA?b)5d3@cF-&^Tk%|=?;NO9e`r0ybnXm@Nmp89!EJr4|d{#v0u
zvU+#uJvN-gL^6rGe>W4iNu=}ZI+iD^#x;atcb@@f8Qg7<31*pG&H~dcdR1sLNO-K8
zpgtUX^$gI>pdN{juYbCkT+RaBEE?}_nQz{`&26>_dzmfPTkOba88&HtQYfW9Fz(}S
zx^yH?M~ak+VUHoLV@Q!5%Fo&Y8Dh!vtxxLL)=q7IikBU4^-B77Ia^oK%V(QPr^eNk
z<?1+zIS&r!VG`-#H{)}|?^%s%#%8TWX0=Fh$0F$rb|Lcb1$0^Lq;_7^$lg+KFJfed
z-Gv-?Aw{BruWiWjFy(tkOcI(q)qC)}EGApbCSLBo&c%N=O&zsqxco?(o&pOpHu;e>
zuK|vyc>*(@(A<w<dmx|(6uv`#1F|?vQ@jVlt#d7&H>Km9k%i2%kOS?w44IWtCWjS=
zB>Be=I#(O1jSroxm(<IL&b0%aI>3)~Ng|N0+E?F3)L5~Sm}5MNRKhWymDGwuZc;an
zDcz;+9NGnfUDPQuIMWuSa5)tOsa#G2K^m9y!7-m^fb#pa+hKZ^<sS+hi@16*xD|8t
z5^yZx>Sf?o#?`C9xQZScx$iKp=CTZoWHeQh-#wh#z=s%*VcWr}oy!#{xB@MvdL^1%
zi4?xL>ny6<tM^5WH_`=gyg;24`4oj0QvMx^F60j2?*LWWUpLI|h6R*!k$*0|G<7FT
z=(jkOHwkZ?6r`C#ojf)a(k!C<^xDyNmq$A`;ejLy6BSPT+3NP>Ik}4_<L8akjPytz
zfN6}81ECjtD+da!S_-iP-39zz+{<VR@Kd;)3j9<qrvX2W^2e0FGhf@>55z8N0Ce#!
zh)j!*vxL$37DTxmjmEbi!h?e)7i}LNc#DT=FSQS(8{Jpx8%Q?#=YYbV4Znq4!+z`)
zOnwDZ$<y)+m(2j!Q%*q+DfA{Px0OQ?mro$)6KFDRUyhv1DU<Jnc<tmL3F+ERluqO%
z_0y!&I7$5jaC$&v82*+RwN(D#`q*XbZ^$d*^5TDgL&E>63(0Rh-f~(Eqt!*U8bzy-
zw7QH|m(c2BMRgc^N<4jrdFJ%b8ml5#N34umxMcC7rD4lfgfHi<R<z<p(CP|WT}7)a
zX?0oj@+djZSTS=Lf4GveimHGSYKYXQX?lB((Nvo@LU*K=p1zZ_>*OhJQ{8Q}tz68d
wV@>|QUv&%&jYgT8jkd6qjI*}2v!6K0(bz#eeu9^_j_>qn<P+%O2mZYO0+ASm+W-In

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_example_common.hpp.A3D8D5EB42288D30.idx b/.cache/clangd/index/reduce_example_common.hpp.A3D8D5EB42288D30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1c59e58464c13a7bf497af26145c66a39f2524e9
GIT binary patch
literal 992
zcmWIYbaT7G%)sEB;#rZKT9U}Zz`!5~#Kk2=nZJRw022d4#hkel&SxDq;5qhJ!*0qY
zr?AU%z9%xD%kE+~n6*|{?O*%b5?$5zcIv(hdEabkoY-nt`%7_RRq5U0n}zG<9@>23
z3&%QXtzEn1=AFvlYrE)`>b@*%lax)F=G)C@ue`f^O6IAq;toN*(u=Frlbx@BoVPsD
z`<3dBMf(l8KlNps)vM&5T&!}!V#$$}AMQCFI6kNN(@Nf=V-c@)H<>jyE%Dd+k>A54
zGsF8*r)Pe%|A)kY5(!zUw)71L6>iHP={=h_e+Fxy&b8gK(ue8=Z+TwUPT8cLFZHE0
zGNZg{#lh5tr?#amTUI!G(NZ42B&E~Ez2OV=yOsYln04;eU?{H4O<D*HgTNrB_G`yQ
z64;p785lT~*p-C91Oq2CBT$x+lQG!Gue;Q--A0s&je~(xMNvf!OfYZ&rGcPejo+6U
zb_`FrnRtNuHP|#{zyt#)SQe;%a$tpD`=^h;co?{W3OE@!g}?*@Ckqn;FUW#-_qjrO
z`ftn^Wnu+dpsJ{<2PR+^FchVx6~6*HM*L92w9WTs{AFa}6p#>*f|)CI`Dh`-E1Og_
zVVJ+TxFG%kJBNXRhf_mK!xZLd9!^ayO(dR{mKG9ETT2^>r=z8V#M9N%MdImc=^^p-
zwe*pA23iJi9xN;wxxirvb0f?-Kw(e-LY>M`arohrDwd09nRz%lggGQ&9_HcXl;c!J
z@Ps*q`Cz<mFXooF{r@ePcsNxgRTN<M@o=h2sw%(&8m1U%9Vj|r*3EOeZzC7z&5IV0
zt6w))Mn0U=#|R8M6%9?8TX*qsTsRi-#1~EV>7tc^zaQroqY0l8aJl}Q|BV%zFw8|j
pmw@6O79K^ZIf)=w6&EEJi-Q<kAOZ}SINdq26({$kNi%XW003*qQ~v+}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_multiblock_atomic_add.cpp.22DBB7B98E8ABAE4.idx b/.cache/clangd/index/reduce_multiblock_atomic_add.cpp.22DBB7B98E8ABAE4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87d20412c8b1e23914074619b98a5c23dcc11df1
GIT binary patch
literal 5590
zcmai&2~<-@7soR(K)^>f9tIM&0EQ(364?brmH>*fK2QM_1Ox;viwd})(uz?;a4Bjl
zAZP_e+&~)vWl=1lg%9vU#U8g>>xL~<EvU8MOWMhMZN77wbC|qe{_|$$-nnyUhKOK^
zq)(qh@k<KM%S_Hqpiw9kHvX5Bo1Hc~g+j5_p-}R^Y{M3>;k&Y)R6UFIIi1|HI<cKO
zepFJ&><E4kxXpa!daXz2rprFRx$OSKSMu{e9t#+)-ctSJx1U^*6_@06%N*waaQun5
z?p=fJtty7)4;9?P9jEU0OujkbR#h7M*1&Yxe}*RSefI3b(@*wfWw+kQ({^n6{jlMs
zikP47Z$8m=$ss0jbD^vLzR8yvA&9`c|4HIzuSt)#maFW2{6`x@f0=w#!~J&WNUZg9
z>*kV)MNXSF<_1}PCrq5N|I+R~*$F-~1a330*96o~&6PLt?p6P>Zb8v5v#yllPu5?X
zD015V@!3U<il?-ND|&XvZIK>AI*LDR`t)qD^JQDT7ie7`^Efj)WqR(xJX+-7o%}0i
zI`NNknQ{I<7hHVz%fq3-O5+#%-@3H|yLeyc9jzAo`uBU4SLeqiL^$6HlMf>^d50<U
zgq&4j63s(F4h-M>5u7}|);G_z8N-iP_)a~>j!ebI1*-#4<A4=eRZkr6bR@m{@wL21
z-W=!gv&-<M^54FmEuE|Nz1-)`ZO1>hJ(#RTtetK5+L|ZVjl`992zEI}Jj$0Ycj}i(
z(#p`^9;1olX+!Pp>runwmb0%_o!hi;sQp}aC1>PUUj3U_Mivg!58GMzeELgJP+gCW
zrE6z&*~kOyh2Pm7D30sZH{P=$tzlF3yI=RMsPycw9m(o>b*gpKnYz=l9#^wBPA}>m
zc(Zy<a`@-%+|0$k4>I@bXJ}+v2IkD3f8u?~nhTSUB-FbG&kt70sItd<Y(nQR9Aw4x
z8HTezW^~x33-%bxSyboDr;m^Pni~X8nSJ5m*HpUg*_D=pXQd1??1&xPzp}IVo<(>4
zSnuS^ogZ_`w`T^w%a>1d<~3Um46nITmL9w*c(q~E!9Iq;ZzZnv!j070!;>y$Q7ha;
zBL?FWRYqJV<3`)=8?{AjyU3H@FK)cOrsT|>yDo+e)AMQ9C&I#9n{__U>Jxoh^9Ox?
z+m`qTzQQ%(+Rq2e(|OThv(OGvZ=tdAZpKcH%Qf8+n{`<h6*G_D+Z`J>wkLIeWudd{
zc!tSF3z7W&>D-rBS_*G{>-2uxoomO^ghzKe)DJ~|mAtp+h54w4v*+&8=h1qip01&q
ziM0BW?!)~1E#V8?-`x>zj@$jm^}XN7Hia(lx)4*BJRo#B=+f2}#{`YdV;Tjw64r*>
z1l;ZYf-c>|tp91_^RJ$1k5+W$ybEW2R+CmE*r3&U;lK7s_iGdD&@;@S_q4GOOEsjd
zK<C;+2~Qh4FzLl;7l#?gMpmV?p7Ot%ZE=BP{6c$c)9vAsN}h(}y7>Dsqny#V6Gc>E
zLFu9>@z0A2_|Y}@?@qE!(yx8gjNNvmfs&J-nfL)vC?@VfQGGA&Z)Q;A!zmcgfG1EE
zcwSI`0^$#y6pmL1GN=i-%%yWtWdY^*&geZn?&619)G!JKLp4!7WdY?K{Wi@()(1Ke
z%5q#j3TQ_G{jV=r<&e2)MqmAyghfrkJFwKY)KeBvK4NX;NvRsF(WNZITVU!l1<C@-
zH7f%IS6Ml62qg)Z*J-R1DhnuI6rHwWo!onuL0y42&e!LgDhnuge_q}&+S|~gO9{o_
zt^|gaz~ZkjSY@d3_PUz>5*~vZiML>)ZDOD-pnUHKsoekO-cbfM8JCT?M#jnl%42>x
zK8uc^QU-M;E(^2;sIpMWOH!mcarbH&)TOv=MmN(_7Em64Sss-+85*ifS%?p>3NWhx
z|F17t<+tE?yIoRWJr*?)Z^4=FtfwrXJW*cVm9>U(pFxelWmCE-TUkIkJ2@ridj^G~
zd2eUg%GaEknh55H`te|+2qy8A_!1P129vZ9j}RY13q?aoT7-9mKh(6==HxB+e;-Lh
zuw_%1vEk1MmNY#{L{RedWD=#grg%Y(lVde&TNZm3K(J!qS`0jfnr5ud+EsbjC_Jj_
zm+`SfYJ%}f_LsN%Q`u?)40f81GRR@{;)b#$hOj>bI}gzFz*rT$SQ|jw2sT+??Q;WI
zKEYcr1!yTSfI(==*y=8&fBO)()Zf&92GQO;(|IHcGYunA0yn`D))ZXMPJOw>b32WO
z`SbjZ;m<T|9&a9r!gygMO5h~u!<y#nOygU&HqKWw-11e>WdCYZfD5UdR4%M<)qQr_
zMpnG3T7ByZ$0z@}gd9{8Y->IQY6?FZs0sF=NrB>1>vvIUm=qXEfeG}Af<vTWA!H}7
zYdW{{m!MM+jK|_lgHwXffs@1o4h+Ha-Sg+bSKJP4TuU$OOi=68!(>dpN;6@MPQ!#8
zAqsN|P6D9?K~@|ql4irPA(4n9B9T4Eo<t5D2NH=nVinCTeo2{pojU?(*j8Y&6*$46
z2Dq5rdL$cA_YX{@Q*@LGln6a9f3GmYY-|SR1}XgJn`0V_#;euLj?@ll%S>jK;sS=}
zBUnFbA;;=zxu=E(?pbXflkhBGo6m%)5W$SJjhHZR;eHn9yTQRDSb?}eMc7L1(yuGT
zQEJ`CJuN+ZXXnX8H31q$)%n{lfM6#8{RGg53o5lOK(mJJJ5TNXOsC9L?2CJ=)P;zU
zGtBLe8Msge*kwQrNBl{{!q*4ADUYehKjs0!ihy|$P#Kez=iP54AN+kP4da>fETC^_
z7~h<40ewTmjLeNJpl@i{4IsJ!X24-9*!u>ULvSw;^^)~o;N45s`+=yRtoH+telVMO
ze+Y<%$odeNI|O{8o^x`rb?=Vln{eBtz(xw}pwS9mA_Yq!=XE`NTtFN64);z8&@BNR
zm>>&Yzb<8W&=tbpGiPx@fpbyfajBXB`vyAy*rrRix*tB3^)?-UqX^6{0zR<wl5ttb
z&wi#q{_DDj%Mo7yid`PY3WCL;F{%(ouvq<AQy9>2^T%IN?EG-sLNG69ueq=iz6MM&
z)z)z|V?J$O`1VIycn+`@+QLq0m^H^*wR&loEytEbb{snr*?ZV~67M;397&qkT|66V
zM}NQEP$=ukP;(-5jh4>e4kHNG2=p3(6ZA2H$$*}W%-w+74crJ_4!CkM_W^DnnFj%P
zkl@;v!w+lTfF1aFDuGKSaEHg6WBZ)vc3S8@SI!-`2ngT+kG-B&ra%AvSNMh%;U1yd
zaPh5~pQM@u1WV>8lPJY5#S!}cT-<GOO$n2ug<wKmAsc!I!K`$xNMxgHLn4u`h(z|f
z_9Sx9bs&*gSFEC@IeoyY-cqrH{XLeO3LSnh<7dypQ!(a)sg5hEHLasO@)6+{!U`ds
zR4xGY0x%(5eg&Xc2>x1fAg|Hqr`w9Q^{{EsY?<zz0Q&RuB77`fY%e4D^W_c9G={yr
z7O!0dXhnbl{oLXyU_ajRcDHf@$U+-OuqB|3e@o4%tMQ+1a{I<jOHJq+IKwqw?-7V6
z1*S7oLM#cEnB@lvPCOofe;#U~pa?;~MzDaeR|W!PAe7*45ZDdE36_IkIfx{<4+Qp+
z^+6CgNN|tgoGtrK7Jh>#0|`&UhmN6PA^Z@Pn|f}>8Z2_Z0n}`hcf@$t$*>q)z&tQ-
z_;zo>*5>kU%@%6)`z$u>*ybI{R7BrQp%W2();wzyf^2!VB(me#k;vZHUP9D#<T;Ww
z@mz5L)cT7n-WTq_K8%ln$zY-|F%1l_jS-o37b=IwD+bmu^A0@eTyuHO&7bg-Tod3m
z0Y3B~4QmFxW-?y|ysKnx2fTKI;r#<Yu_`VfL>Nb~B1nt!3_#BS3)l&QT?Xi7f^UCz
z%RqB2Z}?wx^!ImG)*ah2rAE!wa8IP`DDOAKFjA+cl8*H0NIC~j8-lGuR*|R)ptviD
zAvOaPcLl*$0gAhV;C6uGu7Lb()y=I@2bX%{abF2^DuDrvaRm!2ffM9WW|wn<8@})U
zGf#}>v|K!uXY~+oAzd$B6^-~X^;W3}Uf6JVD{gCqohkGbJ;Jhc06hmx;n)zY6`-vI
zkNtJ6!y8fD>4-*>>B*!<&I0r-@e+cy0JMc*c=^NE@Sit7@){0p1hGW#G7v5U@dS5+
z$ZoKVU^&3#Ac5dM5ZOo82SMZ@!M~mvtr*zQnTtD+J(E2fb_<Qi&vuGyCSg=PVAd1!
z1aF+U?TYIrQ7ws{L<*D!z-#~p#E>ol^b&Eynz-`SXHQQ|{JE^iH@@F@w8u|5e{LOq
zcW^azb%A5T?+!+iSp<a`hd4pXPR>YB*pQRGA_xB$0)-N-DE|D!zy}H?ig;YCdYpoV
z`7A)WX(vu=Em^!MHYPfX5~~s^ONhrs#N&$0q>Pm~&PxuNLetRHqH9k@bXiPB5Zm8Z
z$kqwsaP?3FLpS%?o?hN_eC+i_9u_XvrvLBPX?!DNGl99~bSoR%84fcYXF0o?IPvVn
Te)`iS!Ew}p1#<(ze9Hd-N;d7D

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_multiblock_atomic_add_impl.hpp.014F022BC601B7F9.idx b/.cache/clangd/index/reduce_multiblock_atomic_add_impl.hpp.014F022BC601B7F9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..89b1e0fb34d1809ea356b8a4dc620509b8d22a84
GIT binary patch
literal 3762
zcmai1eLU1z7yljRC&P~!5|j5rXk<tVLqrLQB~q<khzPT4Mj@6h3Q;}fT}h8rdPkI(
zWT)+AA6sl~sdi&idJ)YmE%l_Az4O#>e)H&$xu4I>=YH?G=X=jN_nvd_FVNf5^MpEv
zc|>?8@gft#Xc&fRpr0Utzn-grVO7c)mUN&sm>2tp4Rh50R@Qg*kNg%jhfGw_`w+-_
zW%0q<dibjLExp^0u}k0md2E^brmnh4ziZLXA52LW8Ttj<avRHo>PMr6#@q|V2IU~!
z<Sz2ryn)*~CoP|UJt6Qvk-;m&y1g%Y>aA@2+bW*(A-Q`j^~LuEr-wY5Eeb7zy?eNS
zEr{aZFP+)?-zD~me&;HL$<aUjr7+mt-z$9eJnx?0DVyr;``p?e%<Ws$?&-UH!DZtH
z>)+CELKU&oOewcH%UsOvk2BRie&)m^AAXqj-FU5p_MYiuiaZbAjZ68poJ?$BdFAxP
z6Mg|<W3HKtPj27rM%2eo@i14MwQyU@;qcMML9-O!p?x2g48|D_T+GeOo$)|%lRv{~
z;`xa5^-lA5_xW1+Gx^sCEH#VbOnWw;+xx2eUw4yhztLrPBo7e7v)$DKPB$jD*(lc>
zx)-p(x_rp`a%g<)(<H@fv3>c+e-3g^zp{pDdBtJu%*e;9mcGZlZ#LvFg1)?c3NKew
z_r2G5iaa8U>EoAnshU-$HD|tNy|a9@OPsWeU0C+u4)=ZTgk7x)>77k5K`?ChsI~5<
z>fXev`pS`PWnR={L0zmpgW0vVXol8?t-uQ$n)lmG`|smBR)~-9<aHj>)6r^J*Xdk$
zH9^HD)8#cGi21d#rpGR1V5-`^tov)u&MS_3Hc~P_W42MVf#36-tfD!AhD){>Ouao6
z(Vii=(!vr6+Dg;>+1+`eM-G5ct6#(7V~VSGI;IonSc?ggdErygS>gryo@MOlQzx*A
z7>4W9^v(YFhtU-P(^Lc;Fn;7ZL98m8y1H#%hn*z@2z(T#jslxh1+lN6n{a-=0u9dw
zb~ad%0vetJRyi<(#EQ16$184x=S?8+h{+L#<S>Cp>P8w-D9R#AszTt==Fw6@0?&Zy
z8Ni|DW`j*O*i$$MX6C>g5-ajkxQ;(Pb44Z>0;>?DizwB+S(eU-Jtfy<rH;0N!Ra+B
z3Iv{Lk~oVbCh%mdWET>tvR<BSt{<bLWw>tKrR1<$l<u7MC-r{>8ZH7~5%`e;reE6>
z-e86aBITR#Wk?gO5~QTeCPSWGp;xC8n#ABwC3Oirft4T)J+g|CC?iW|CCkWC8L2X|
zG*+5aroC<JzO_GyzD1@*&y98^$#tsZJjz_R_8~`H++0G)Vb;A`<@d)f%#<4~z1iZ#
z{bqYgzd9m@iHF>-gXuFGfgb?o0nj3;3A`Mb<+Au7Fb~S&3Sd@Hc&clC)3C#@DsoLR
z?Q6QOGHX@PSEg8HA8q^d{gV?Y`1We{8k8a3pw<miY<0VOk#hK1t8o$H_#b|$y^($D
zD_YxU2hZrOxf(xi{w$X(D)*CDXyF8I<X|+HS}lYbqD`SFT9gWfqG{1GC<B-oAf4;r
zUhIDE-4w>Sb({=`+GCZ%hLmWyy{5f1yfj=0+CpHF+%&u!G`p#AN=~M_2$+;y1k*%d
zA&Xr_FptD8CB}O$<bTpa?pJ|D6_si3@zpa7U7O@5l;_3mqPW3z6)5~M^cV(7OyK=M
z?*}PfLSN9zi)8Lb3q1k)6G{-|r)mCTuQQJHjckpinNHv#^bo0b(8E<eU1h$ma!*O9
z^ZtpI_>1qxMRQ%4?#I&M?TKjh0+>?(i%B!r^97sz-yU;At7ii(8`MYvfl3`Mhr+Re
ztF3PjuNI?;90o_adk8smJV>Ruz$3pK57C5j%_J2-<#TaH#aPldHau87YM4<jS3jY!
zJfvM*bAFtgNk8@WZK2(|i4Lc+t?_Jf6#_2?dNHU|xCH1WvbYrJrLwpT=w%d6dAqCS
z>1Gy&qUUJl$R%mEo%ZDgm*2XMLUIS3?tnKbAn-2Wc7YFxGm0GyZa`W(()U{C6-Ewc
zt#&S~+;Q|9MO0-qxYNjCRH`&wWKaPp7k~jN&Gk{te6{D319H(q-C`0sEOC)hOl1{;
zmjI(g7MB8}R2G*3qm06NTOG?zZXAt~Y_B!Pj-us+a-;_5RTXu|r}qDWq8`EukyhgT
zfF!@H;h}fPz2%daOG|R$n>8!9G4$`E;pgD=96U+Vf5$DN`2_5EOhc*dZa<$Krs0=C
z`?73m_kd=PY-)=^T}-8RrFfEM-%FEs=<>}3txT9q5=pRmCUt8QcrNJVf<7hR4F<bG
zYCnNDf_oz@q~y)8uo-+QJOV}|U`pb{gI&FqBan?8WBaoG$>FN(Wu;FyUD8J(Of^ZB
zxdhX!(xjK*sWs8lKG6zxNp#KG(@EX4md}E}9SX2Q2VVrbBFa_?wiSVNLuy0s&#cU2
zXre`gpdkdA10Vvrh|1T7fh`R>JBLrmO{FIF{>~j6VkC$A|6f5Df4;l?H23vWxePz5
zJDq!Ib0h*W<O4k)2+}KI*AY9c<I;M$4A&}b71(d~zeNmg)7+d%{SK3q>}_+GFwiAy
z?_)20Z4l_77+#j$zaG%-0h#BQ7_`NpOVV^YPfHE-*DsQ5#x1WszvR~aq9=D^mthP3
zfNgx^tC;mVbb7ow#R9nuy{i%)Ik@jtl{>LN_RiLzl2y)V!`V7)Gtwu7oY~R>Gmw8~
zw4B{pfhtKbEm-EN4Wot0pm18a6ulC(oIjal-iuraofs<3jo|}LZJe4ZrgH5X-QnG6
z+Fmy@t^$vZ9c|-0EczSKdbxNlCCx=|?zk{Xg%|MG3XVxSOv&Hprb2cwOo{5E&qT5j
zWCz10P+iFj{!3FMJD37~Eoiv*ZpDJfY8py(<q0aPgc^gXPQ&QZDH!_5B%9;|^h;xq
KozI`7U(i4L^$SS=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_no_index.cpp.3EDCFAA9D727E5B6.idx b/.cache/clangd/index/reduce_no_index.cpp.3EDCFAA9D727E5B6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..173881a446cef841c9f5e9d6b348852b703cf32e
GIT binary patch
literal 4554
zcmZ{md011&7QpX>u!WnD4FU-vL`Wbk1r$LbKop3GV5D%dB8Y%sQx=tA3yOlRPm#qc
zt{8pKr7rkfse((jRa6jEL@{m@MKDz=BAV7xwf2&Ba&N!yJ^q->FXzn6nKS2{IUi3^
zDlf1QBy#SQ;{5EQ3@U;kI3k*&LiIK?1TnWpkm7B7k{%R1n81AUftz(no7YHxWa)ma
z>9Ri4@THuVRp`dr8{;!&q#^y`PPaR&1#>$4qaSbUb-JF0SRpgGmv>SO4O=%{dU4B<
zcJl0q`+vgXjtu3D{8Cz4LMbg3cWid&>DymNOe=4%;yjXEdfYpU;#@HA&chG-<~K*a
zEggBX@ZYM<Radqwmw))~6Mg%cXH6r0o8Oi$?r6=e><9`vo4F1<y!Rk&zu(h%-ASJj
zC!eiLZ;00)E=c`h<Id9fp4yS6Re85R+dVdakJ|3+`9B6XDUZm1e;?|=AEM>Aie%0m
z=JK7rQS#wmw&wM7X6~`^J2{3^`}FMVXWxsz|9WF%Hv98VO^e*_`kgJ+$8WPJbzgSq
z&R0tEU0;SZObmU{{Hl4CN+d4PIc|5kv38Q+gj3S>SL-{fg;n3U<Sp7@=izyDyjjhz
z%j++;einUmM@#GH8PiX1`WA6oHuzW-a{t)Z33o2qi+@j6{?t72Qs=^e6X6@h7Z(?g
zrB8o%>B`N<a^1b9gZGiJp@gjClKq{UiEAE}^$YWE<R+ywBsWx?%S#NlSoN8Z;_gg~
zS59m@q1=7%WmJb)$$LYmpM26fL2^Z0Y;i`q2NbD=?#EmBNdv_UK<l6Fh^UsoTUmEJ
zb6>!jUfasMi`y4^?ms1eo^s%6ad&Ry6Zsv{N4_p0F>iN9j{KPZ>erNnS!Wsxax<!W
z_C8(QWP7^Y{QLnQ+N&ED-p_V?Y8JZnzT=^Ram!luRf2T`mlw1c9&K7c*KWBik*F)Y
zo>cEqd}y32X<B<r*v_N`9U3xhe_dKJd$Y6FGvWBAub%zUvbcWQuPYj2ef^`iUf9>;
z3c7#0ROsI~-}C0tUH8J{9S>Ho&rKG-TWQxe{*B~JcG01}{<w0^jf<5db=A50l#Y}=
z@7;30QVIHBjp=)KG&}H|!}gluuQX4j^_gYXcUL$0G`GfYS|~bEyG+|#H~6ysQ2Y3O
z1wpm(u450E=(iv2?dTkN>OD)X#ZIsLxVyV=tG;{JPS-1{L%{=g941dIy}eDbHSk*3
z=H5L;izHm-)ibB}rIl>V9_qeTm-=DF{rxp-mgLM`mAEQI>z<diaR2<$7e)1=<-H49
z<0@Z=*}hM&XB@A;#W=Ttd#ZeQ?1F168o!v^-rBs*V$2a)@-*wdyt+5*g;6#4D`!8g
zGq)7WRvb<$EjZe*D^qwa<`pawuI$VDcKVo@IJ2u&^R#cK_MXr)BM|n4piWIme&#uV
zAkM)tN!NdVxQs#hBmu$f9qh%%0ncLNUugd%M@ROY87Pv4AlL%)1&+o6ilfu&4oF|^
zn9QK0;<Gtc9IkPI;<O9NU-Nl0do3x`@!3oZCdW8HaYcvhM2yGA^C&V47q<aw8?gHO
z163-$Ekdp>&i|c`%)wP!fMW}A{rdw|>c)ujuGbGMnUqZ224`z$j&XqEMz2He%mcL#
z7?gBe%(LT(i~|&36Tf~`rZKxpM^f+&Xm&KQae(5UzSV6}UA2P@N(MgLf#)DJ4p97L
zNbye7@-IL~X5+J4fo&@g{QZF{!znqdHK{$j8I+m$3IZzu$2dUon<;5NRjVg)n3ObJ
z>}%!AF%D4t+qxai3+FQ)GAJM8V!johWgMWmFgr)H#s)#m3YUZiH|rw+ifMt576ib1
zv*@W@JI}!Ovp}&hZWtf#hGGi2ViG}-+(?p48KP7WvMJmtBw5^qxNs=5T%;+U6ZJOH
z48>;J%w)lzQ7l!G>P1kRB+W!H+^95Tn85(S?O`e$i7DAO*<6AKf!81ifYBmiEDv67
zb+y+V#ll75-h^6|D2ha}qF558@KW3eO+Dc0fg4;<e7f+{=jHO1=2R@4A1;JyR4j@g
zMWR@KEQwOsDRzXW9<cSm4ld~SB4(tcfGfstBb}YjgK7a`7io7(Qn*wUvuE1dLs#Hu
z4^V_aB^1+0HQw<0P^=4Zy8s`q9C$uE?ANB2ApAN@97?<hwK7(j8$mUIRRc(@1uQL@
z>jA5t%uRsRL~w}na9gfL#;`RNb78x1VSeExc5x+0%odYmZfrLaxwG9#BxB1+<jM9V
zkpVCafC<k>#SFmL0Q@1tBbm|uQFc)59uVCF<DeRf4FJ{vkU)lI37<AvlZYq_niMpR
z@U+TXH33qhb@bW%Y~~vL&QuO6XZUl{G`9QGb#ikn{LZZHZHYVMS@UdRX+bduYX@6c
zDp0IYR49R#<M%1jkf;mLy8s6oiek3`{q{fE@iq`aIkcQY`$yqKv8#Z36|kUMdhxdT
zbIyL(M?tZAAgc#{#P0PVtR6%`&W@jWcKNZrnYf}BNVI?yL0S-~1#+mkusFQKsqN-_
z8!8@wPF;j+skp_`+h8o*nu_TFssnT~^L0Q#W{D0+$?T^iZg_Nls1C@<a-|MTCi7Gs
zh$HiK9Y`QEzC<!)&4HWW)mF?|iYHmNT{fwt<nVF?@K#YQ$3JI0iKgbLAgy{C?G)Tm
z&Bt9-17tP87rJDW<uzauWVj=s{hv1rkG{ViELg2k>_-srBM65xQS1efz5srZVQ%3Q
zN1rix`6!maO)y0yieYw`2wJ-{Zz?r5VdY~>JUcyHJmCsd%$4m*$}%ZiN+J)o2Z>}e
zWHX6*-fVA@%m5M%0E71Jey;1?^$V!LbEE|Xwt!&xI)7uZUu!q@?Oxnxse)9KaVVB1
zNHfu~)VpJMRok*HQOt$z!h&0(n3ygmksIBOMDBEV63OT?5_!@+Nn`-l1|kfPm-+U&
z2nxpmDyCAXOfMbGSPPo52_QVIWGYFITn?IJ4?T>>iZsWMM2ehPf}R7f=U_Z^&`IUS
z;=_uc?vC!tQJI!FMZXtqHPGON`z?qS<OF9ocOKyhzX+L(KcBSmG$j!wHgcPIg3z&O
zB%~{j5#{R-l(m{sF$G^?if1aO6evyWQn5H*oCDmYO;G&Ux|m^3rD6|2zyl)ss8}}$
z=mufXcq;ZBNS^~wf(L<gkl?oSGs5?u!e)$C$PksO34je09<-y~6SNdR>bRf?Xaqh*
zq%t)D*Ik!Xd>Ni-I~oaoE&)U&xbj_{36k=qB=X>UkVuv;%OW(r`Q9X%0n9Rhba;%8
zRh!?IZMcM2DU8OTaiO`L`U~?<M|Ax}p<*wA|4VW~8wUQv1jDj!Ni%kSM99Q3<4iY<
zKlC=~qy{Jh&0rH}&qOmx^b&aCJ3{kNY#4aqJ3@vXCGO6@+DXEpJs_Y5D4-&W4FTy8
zVdt~luH{KvX3OztD&iM;z$!JGs<am108@38r7a);@}TBS<A00Yd-2rG;AEI$2*olv
znI?igEN;|j155ZI%_@yUobzYk^)m>B2EINrxVdx5g(BSXbU%71JO@<ApS00#msqPG
zF#Cyj12(=^G-Ja{sHIw^k{qqJQIi@)wNNdE>!X<3S8aOUahj+m<*hnVol0m90Iva}
z!F*rwLz-E`vCr`*#9!neNT?~C6ypd|3Y7r_#W}{AibP>{UdHHgHHBFkU;Jx^$+;Q+
z2QQ>In<XbEA&Jlx2$Dqnn@s$hl|MI6jbq%rX=YS&3rj0&8<fswu$Z>79A%UThZf6q
zw6o_6LPEpBBNP)S`SE<@k_n#T|Mjbb&{;Ib)lKRy^YZ@4f9$yNLBTG8PQC%rJcpR6
NGboc)kyAiC@+XgpV^078

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduce_with_index.cpp.F84EAB2F5A59285D.idx b/.cache/clangd/index/reduce_with_index.cpp.F84EAB2F5A59285D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a165e1ee8e46bf1adab6ea5ce7ca735e2738a0b1
GIT binary patch
literal 4564
zcmZ{md010d7Qo*P0Yi8R*&vtzAqGeSi713cmKe4uiy*>7jRa)XxFJMatgYxcMbtt?
z>sGYtxU}Flxa){@se;t96{r-&3Kc4ffMl#z9p{mH^IpGiCV$+Q-#h2td)~R{oO@=)
zXtj6P2ojkSQ?xX9MJ5$N5F8QXibBI)D+IB(MUbLBhf}(ibp^2=z2jwjRI0jchSzVq
zd%WhNQ?1+Qlo{oR*2!nZ&ARgRrw?VfSN2fns8&@3EeP6R<Rg@;4xbjxT5qZr?7aD6
zbR%!M@ol2Ic3;~4gwdm&k8iSvAMXyG=d9nQZ2V6!Wn$y!Yx{YluSK>OzTKVv_RaMT
z`yPD#s;BFQVxOX~H`(U-n<pPv_$-r7&9i^-?D=gM$w~Q@^o-v(txCNpIPEoEX?6Vc
zt)<QVt+A)=c)pJYZdqSc?K;^!PgT6ELOIVjSaLM}=m_IQ{(03IQ|nOowdSJPTf&R?
z7X@*LWejOg#mS!*O7=A`^152@@#U7Np}mRk{QuA&tKSiGSKH(=IwvUO`<$p96R0Zd
zeaF{(#N|ka!_KVRGno1Bhd25Yw}d|VtemxQW}GmF;kDq^ZRxG@=r&dW53?57b$+={
zo86mqyM9qw!*m}>*Qeo+b&ZteFVAlN_VM!X)!OO`WrXSYrp67mn-2CaTok-R{M|=I
zzKoe4Kg==A+tSggAN&~!8w$^^pU~E73}^mQ{F}?M!_umln$#NEC!I$MQD!@jB6X!l
z927tL@{l+nwAEPJa#O$FYS|RugqT9wS65HO7!a@Hy?=i(M{-uZCvt5>;YoCD-P$!j
z#2@OEX9d1VXV1%@lECoaxj`T25EeFg_CVoRPwe^qESE!Xn-9Esw)t{xYFo#NV_kFZ
zPR~zXQQNgzzs36cDJ9+1Y%{kfy~XAO)!EZ$E;YSWS2it+Iwri~^5c5p7v6K~R1@Vf
zk!kV;3hOuLqmN$r)On>Ml6LRRl^;KB92I}$RlB4*szt{<@>(=-IHdDn3ii=Pc`Hx-
z!qxVdTRx7X^5W!*5{DmN7#ijq_ULLB4FxI=>k2YX%_~{7sh0~9G+gh#1->VCFAiJm
z|Ei$FW?w-4%~iL4%>4eT>Eji?J&eZf!ml^IRnGVBK8H#I=id6H&Ff8eiS*nJpZ>Je
zDFx5ZN!nEtdaD<7R|(YRjaq5v)+MP)2fw&*Wt;b&f^8+`p2dcyif<c(3wuHe1O=@s
zZbnIe>!ZWXz73a;uK9V_&bcodm+q}<&;O<@NA_%2#omNZ29|vF&!n3VuZ&#mZC`_w
z+)#(e8-EWqy|eD(eqG+RIyviC-*T^O+qpUUEjubUxHGir*~j)c-s<!nl7^P*wmQ@{
z``IxgGimyom-p(oPWvTY*Ah|?xzefJ{)e>9`*Q?2>6*&R8O)RCOJ7<CCWR^x<Eo`u
zR{(;zPMMz4_~gMFCS^_%f;l=lip>L_$L60(F3(P>IrAF>(IE&%cc6>S0~AN=cUOBp
z-#?8>NyBGzX<VLpfMWgi)XxR{xev!tX5zD1HY~1rfa207`Pu2R?blFb9xiSG)CNF%
z|9~pHCy16bR*h&`lq`G$S6f%Ed4S??y8m-&|0VWD1~MOCp$?3%1CsX-sB+HxxRh1B
zwTnr~z*pcq@I~eUio1I^Hbk{;9b{56aj_HM$;CWC@rW*OqcN@fAd@l|7Yk`Zu6cms
z-(&R3%^M@QEQ%f%`_O#2<^hUdR_wn~kk5R;q|Cy_0-Atr9-#PL&>y3d=ZGjMl8sxx
z+<LjtJXpkC_eIgm#Z3$(9T(RF`+6XJ|9~omxp~Izb_inAy=lt=6K9eQiiPpQ1kfxL
z(}Zdw2#VxIl4ROot%i`r@M1`^_@H<-l#N?yESewnI@t=v=Gx6=!`CR5=8@)2klsUY
zAsAIux;dI)9N_k_G~<aWxpuicf(C(d5Cp)W5HZ(H*+{$d&>F?mBDI20ixNeVC{`3p
zA{}4nPH6T4ejjj$3yLolE_}K+w8WZ<sRe2ms7A%21W_c46~vNA$I&?untg!N2OQvn
z?$4&bYFfq<0~E{PWbmPyqIBq$^O`<@zpeuTbs!i9Da<7+`d@Mg#RB~Tr^4ANmMBWJ
z_)!)19sTYq9gm7)jx0w<=vLg1Dorpk#V9jQv;-5@26$~i0IdwTmK*lsd|e<O8mpXE
zc@t{I>|%F<$^g3zkhl`CE6H31*i~dc57_4k4t5=BSZtFyVoSx`IBq<cwK$30Bm{{$
zVv@|A<4z(eM@k|&M@}Lyju(kcfN26O_<dB&1SXq+A7ogPa4(P5BO>$yBLWi$mnA9^
zgCHf_M%OOQWtHP6mgtn|3SXxraHPLggj&;Z=i55k6VB(`^6g<+K`|#=Cwo{$P^?f~
zC?&jV6dOs@1{iIC3k^lFcED)=E01pnA}EKJQ|NywnkaS$Q11XXRLdyZlRv-qaxVqN
zs(`!-Od@u#0%27k3Ucm@@Y;1XN3!rVt^^*HfD{UqKv@a=pyHU^zY>H%E+|qrIXASv
zwWH#p<=h55;MP?93q0F_FTouE?Enli3p#+1%pM)UlgyJkz+^IqbbwGYYdgR+GRJj*
zcrwrI07+!VIzTF9V|DB0t)=rn!LujVA(vEl^7wf|c*-c2H#yIrL~(hEkTwiQJ5Oo)
zOn|$r49LrX4|K^G`<D?e8sqRXFb&F~1%Up4?lU|($1XE?#weD=OR@wWieV0z2pVx<
zNgOpcsibEdo=q}0FCv>H90@5aJUN~ul5u1tk|)S#67v)s1xaQCcmmCa1|KpxZgGfv
zeFTrDG+~;>1{Bi^^%jB!4gZKguUvQ?q1b&8a35&kYAE&$$es}aQj>Pq_aC3xbH<^V
z8^etaO+qm-LrfxfhC7L*3@M4^3^|Fs7+xeY0b3IheP@b&dfkMDet?Q4Y7#BCAWU2v
zy15S^PC(1jk}kkJG|v&903ITqd6P(_$%`fEDNsHIf$)<rXtx)g(DdIO+m)MWS>ht2
zGnzJV${Q~+f$TtMIQvG+NxsX6!PD>nqx;jfgdgof?Pd^!`l6ALZjYb3c3X9Ey%iPH
z2sD;3rea#5*0L@Yi|5BX!Ce}JMLo7fOlv9?>jD8?#Q9LMdm!K*m<o-jVo!nWDNqnR
z2xNl<H(Z;o{^lY!d#q3fCt4bWM#<{Zzp0b$$BuQ9n~FGAi9q5?kf*?tL^6SlMDkR*
zj?h#H6eO7m=uIFK+SIgR=j-AxZsL^&qciC|c&WDZUC+NXwXL5*#fE|3Fi=5#DmDWA
zMhJ$bdK}%{=n>~1&x*Gk7k;DLWso|ZOf-`Ny@_I3XcmcvffC;lZinLADe)a4KVICn
zE@fAGD1O2#oL6|k0x{Mb(CdH`Or<gQtOF{@gT{ZIJ6a@ti05G@H`8*i_}<(s3&G|T
z&%?3Kl=u`qP0uCFeFBtEKoB%&^yJ{qmeto+;QnJwVuZl3g6jBfGuFluYyApXzY--J
zcC<9QxsxT-(r9TUe;Moyq{hzRV(^6PqnN?RV7Ym53OA5)$&hSFBQytqa)4+E-<Dq0
zTb-);0I%kgMUw*vHI1{zk07m!Rz*<!_;^d=Lf_-=8tb*;m!nuG5Oo57=pYmu0PF!E
zh5YV)8;(!=vj-<8c7uR!5CIiYYzWAPh)6>UbC+a}DH{v3jXwC(3On%jIYXCx-P$TO
zIR#0E3n54f@i&$Do4qt=i2=v>1qoJEYnyR2TRW7&VX|5Fv0QDGj7yK@jdySq2!lhy
z)KfL#5tI1JP>&!l@&9`3<l-utAaVDU%Doj6C;R&O2TpMdaQ0C}^PQ&0&8AFCjEn&@
GkpBWS<Z$W$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_common.hpp.C00F2970D4BF7FA3.idx b/.cache/clangd/index/reduction_common.hpp.C00F2970D4BF7FA3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dd6f8bfcdfd9c03fdbbbde843884067bfd387912
GIT binary patch
literal 522
zcmWIYbaP{3VqkDi@vO*AElFfyU|<jg;^LB`%pM>u!^pr;F=uPPHQylv0oGcXj2m{-
zH;O%XjZEE}{fF`3OU^}17ccEe`eJgaWs7fc#o0yYY#$j#ES9djz`p8zh2T*am(}SP
zZdE)}3)iZ1?#)^AU9ET0^GChsyLK#;nsQ`PpX3xXso0QTbG4pL+_mHA-+6cavNh{3
z9(K#i|DB%t;~jTgz|Lx0nZv^Krer-XuFOqp0J_OsT&LoLc3V3e6FVaVix9hzIGAAI
z1j;cmFtRW@$p$pvnm<ECfQf^dfkjkBR0m8jaDWvsl&nhKW;ij`m5qrFsGpOaQw&TX
z^iSd`+z|2JiUX*hiGhVjg+~WW!1NcTrWG#(8pXSH#eNOeg1t;UEI`NuF`R*cheeb{
z6wZS=mXQnW$PM$f_owcfxDDn+MkcV+V5(qF1F8nO3#J<8AfRfHYhbE&q&+y}!gcB)
s$N{`O;z$nQ6@~MPQgaePGQ~y7#b7~Jm^uaqmh9Gbt;Q8gL>O5a0K3+aHvj+t

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_enums.hpp.871AE7D90DC9E069.idx b/.cache/clangd/index/reduction_enums.hpp.871AE7D90DC9E069.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aa47ab062f577f04a9514c5e64c0a89bcaaa6072
GIT binary patch
literal 1400
zcmYL|4NOy46vxklTVG2b+!iR67jy^|G$=YmW0V<%$iOIpsU${YrWLHZE|U^vjk<Mq
zkuhbG5OoHPf*KP>_!Nzb?jt6PW5M`|q<)5hP7%>MH4?XTZ#<((|84&Nd(OT0o!|Q;
zFDK^<13>oLymEJO*=hy=R{VO(O4mQe-)aS*yw+V+xY0k&;(ukI5jYul@l<<n_N@Az
z_TZGh2OVRak-Iux4>ct0{I#<y{chTHb4Hb~Bxrs(5PK)B=7rzUlDoY2eDY9p{-#--
zM@Ra{|FB|<xA)Az#h!(!+SZ`aQ@%I-_ql05KTn@C*;`xleO;H?dpy&+tM$8p_s8}H
z1LsDnJ2%Jtv&B)gBdcoW>CfH0ErA^y>Ri=-rlj8bxanDb-@(OCQ(iZQ=lJ*Zp3MC_
zJ~cKtb@(ZtGodlVdD%Ddrq|wg{jjUK%~jEHNwY;>-*e<<&a1>n{SV)me4AarWe<Kd
z654;_o8{p7(!HiE1R!+YhGUVx41dKTEEaT)X{r=3;`6|d8vXU89`>z`Lj=YaW-&>j
zIBDg9!KV3#6FI0cpeM3Ks}zbSR}8(0+upBm2*K)b7H*V6aaL@@>=Nbd5)PKJpvSSe
z2q_dVD$@LgdvA0bz-j|MT8q}DfJvTA(UoNtd6%-A3f6Lnw1aMEcB>SsFH-IW9P8)D
zbI@@2Fczjt`9E&!&Rn$2?=*l3+sAA1b}6InSFDQyCFETd2Q${2nOT)W^=@Cr)(4gO
zDICV&-4j@XRSLyB%e%AQ-P(Vf1H;RWW8-X6D6W5U<9^*srGi5!HjiKtsuYTU`qY*D
zFRR(jLBZI>OsW)$n}^lPBj<vQ1IL@kuo#;Zid#Q?VSPKiF3kW@cr{(qKaetdwNnk7
zwzX9jgm5t8*&WPbkwW#4HeT~B-M=c+0MS^lA`?<Z>mLW|bLS3T*~h_v^+_zrDuwEw
z)U3Js*Ssc;LpYwEGj5VXacOahrwiXTVowQ^%+apI9NGdnvWYdc1!zU@$488zEdZm{
z#1`5DEP5iu6xsr)=0m9$iqBMc5J_kYjA#`Rgtowkdehcu3v8$f@q)I1g;o<MXbVWF
znD{_jz(EHjF3=WGP!r+-Z2<$>!~xm@0y;M}Lt6}=4oFi+Ta27eLe0<?qo&m~W3<J9
kX*JCjZ82CXE-n6SwRkT*rB07o{Lu{qfheaj<qv@W0PiQ2UjP6A

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_functions_accumulate.hpp.7D2C2CFC245627F0.idx b/.cache/clangd/index/reduction_functions_accumulate.hpp.7D2C2CFC245627F0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6375331dde28186131401c51dd2c43a8c0af9fc9
GIT binary patch
literal 1634
zcmWIYbaRVhV_<Mj@vO*AElFfyU|<jg;^LB`%n6JP4Eup{6?3+R`1&&$3f!H4iJ9HE
zL8~f?m-Aq^xJlcQ$7wJ4a-6s4{#q#XxPE)|li#-w<fh-7{XNa<ty=%q_qR(HOi@_x
zmMPJASnkAptCh9egUl~#EOhzQwln?5qKgfWv$*c-KHy9Cf3QIQ5yxHq1?e*%Uzf>I
zPyhaV@}1~gw<334^ql)9D9UM?%d}0}i|098XYe@Qc(Bni{$;}5A03Mh6djwtre?=7
zrMVU<95pXpEf$x*>&mz}vC?9FUSrN4w)cMa{Y#8ylsImf;iuHkGgZ;{h3$li?97)G
z!xpM096Ihj|HaLwqDLZz8G8$!beV6ARJZYR`&WACSmvz7%zBqfRBzn>pCf(s4X<pJ
z#>Iu1cTDoOJ;~etB<k(uKg+ClO}YF3m{ri{DcAVD{N6h!@a{Y?B`9izP-dbc`)T_X
zVe>PLCMc*)i#`>ya%yYift((<MGIn#9u_aj@!9*hX*cKD!*kEB3Y*n>E&Yn_LD8G;
zug|PnR#kWP@M)&KA)k^yIqN@eKbx?xrF-VeAJ&0tQh%6ARu?WVuFOrU0>+EIWNI?&
zt=k1`Oxyw-l40y&s$hbF0~mKe;HA^#+w0_;EyTnp$RO#?;?4^uVA6qYCQ&=r&$b0h
zi!w+GDhX<X37B-q<=D*<xk(e*n0N&kB>CC-<-i03Cl@n60|TQZW6WZ|-)pkt&j9rc
zF-S&gN1A{MnEokwX@RK>nUO$gVFpPxZ8Z}x0h3-X$;r6=LX|Tc6ORD9q%ym*ESP{v
zpK&}Vdi@1k07#lsQjc9v7EHjUujiNe7$nDne9k86%<ilTCScMRy}n+s=?viqNkhUy
z(t+JUm4TB57-}H5-%B{jqI5$y7w9+<21x^L0~0U-)Bpa%Jo|^+1U`anhw7JRmzHJV
zWMzWrFG@`-z71saZe6iogSB8U3y-84iy99^Edv9OqydWo63>;z6^R$g5((#Nuc~|5
zxi_qbnMYDsOjsHgCp?ly5=OExSMo^4NyN#*!W$Orj9lQ@UY(J>^1lApc35l!g+Wn$
z1{y*z6M@2@D2Bx<j0OsWq84Td%nqP1C_-UYl}?e}WTyRnAqlR8ML*CaP~5|!dz#<2
zDW2D^mXcr+Oe@eNP%?ozbw}ERGcH`G9<uUC3bF|DA%b3wS4|wjGvGBq;<@s=BJm=5
zBjG$)f&<zDN^LM(V8I9!2Bk8XFf4&FbAl5VOqd6f3Zb_1Kyo9L2a8OgNucxwvk#Wk
zfF^;G7{VlwUlHB_xe3Pm94KVB?#s(w62kDSv1quMz_Ga`XkjQy%}E4VSzMG{Yz<;a
jfCvbXEUh}#x25ErD~p)8gcO4mOq_v148eqlEdv7p9h~+k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_functions_blockwise.hpp.C958E7F57DF3365E.idx b/.cache/clangd/index/reduction_functions_blockwise.hpp.C958E7F57DF3365E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3ce2b703e9397720545620d5e22821eb6dfe0d02
GIT binary patch
literal 4528
zcmbtXYgAL$5<U|G1Uv`?Pk=}WMo`)Sf+RvU2qYvRB50~DSe5z!K@DPp7K)%%sRh(N
zrCQKF)M7381c_kfN-dRw4?v*`^)0to<Em(-R$bn^6P$apR)4YN$KILD>^(Eze3{v3
zr)V^LT`5YnXiBawEoVNLq9`I#h8%t7dJaX+aipl+iZe5(C5;Sm`OQ%HlUKM*yI1{q
zZ_U0}qmRvuinjeOwzdCkdUdX@RqXJ}jmGY~t5u(9mVEVt@4%D1E9GG~TPBMYqT7Yb
zrXLbsuXs4M?~<xF@||nfzAGO#>)Pl?*`ws*-qO@h|M<C7>bzC`#_Fvn^J=>HJ~^TP
z^Zj{#y)AaV$7&NE^>+y}-0tVVae3rH-fCq|*<Uqf*^6uE7xeW{di`AL;WeY5@I`BE
z4n5qpeDj^HKi+iOzD)ag{HL9B??v`})L!|>v%2+Ln*%kgu6;N@EV_&dnxB4S;{Et;
z{;T^OSANtston*t*Xg-@m%hO8tZrWFhQk#rI=;VnZ`!J#tK6MjN_Xe`dX3rA*E7)h
zWz(B2>+S_lG^M*;OgZq{?jNRiNZ#1kmGe=5^v=pJt3y_uY`bmZZ23mGZPOTW+2yXb
z^+By&#|&RLAGeLwd1RK?XzHTEx9y)C^SSL}?Uw6L9-MGmu_v!83MS@`b>#MS%vzlM
zdZlug&7K0&@F10;s(HmUiLm6?iGjcG*UsvHu5Ez$su_9R#bLwZW}N(~Z;9*`C;kqb
z)R5YY27y=IXxq&ra!xObo1XepWZJh?D;$12acKJNxi;hHsiQgz1H)6+yy@KTy`}2D
zH}mVvv!ChRR(L1$Y)$@tV4uje<+IntahFSP^hZoRSoB<-a{8B^tFa%S&iVA)(^qFM
z(|uL&*?|#!gEDT<Qs1o$Z4C?JHhF)&Cn4muRZXt@m+lhXowD1e#?fzked6P_G5Si~
zyU*U()+gdm-u}+R<(CVqHeC{bJ|Uyi$G5;OP8fCSY-N<^%-P|W%Qt@z(QZF(_x*((
zu}O3N;v4^MysLlv$g#*N3$&VgQ})2f!WrjMDMOxaK?32T?Rd9@;~T%2#c+7e6eAXj
z{VfZ{BYY))!xpz1e-GSr#E#=IoMMFbLSM^5IdFkbQNY0D?V_Mh;%e_O91n7StZ;0g
zWx?~aH&qEGb*lA5?MgA(!t6lHg6erE3YWW1O2}b2PDHH}>H;has^9K7MP)osJId$K
zL~R&p7-d;dy|~4-DYMqShMXq8eTBXOmIc+B-OF_P^UZh3X%~uVgprLf>cxg8%S}_W
zuJv3PAU6<`IuO=Dz>5t{^qsDZ%Hix`0uV7d48p_U|6)Uvcb;!dmvmf6ASNWt3}J@9
zWkGexu$Xo0vj14baO_E#K0+UV%Yy11u3~9?=G+iF&M*S;Li>edEeqvf%Nd0Dn-$Y1
zRJI8Z*>N0+K{eP{L*R>za<H>;v(%w>?&D4E2X9^`PDq$XKzM|NIk=&tCd1#o+M1*1
zNiB#;JqYW`J%byXoM}ydH^eTfmE1xy$rNV#TNYG5n{%`xLMC`l5G9#NM@mOo7F6rg
z(hb=(%PEedg-ggGn(-6*`D4-WTX3{6>flakV_N;kN1cK{_+rV{kRF;?yZL#dHIV}S
zLRm?!kSj47nwg-OU?pX_WqBg8XeKK#E0mQA6a`9T70nbX3aum~xEa9{DM2$v2sA<{
zE1w0$Sx_RYXyzO!&atwy#|33@=ar>4Tt=C!OgXdRO#1eE#VjsU0?8$i(khfMI>9Y}
zi_7#vaxbJz56U0?C~y5PkM2XfXO(s1$uX};uP8)?%cMo7MIkk~%mIiz08z*UNo38j
z2jS-(hjb`#(X&Y+PC*lwc=Gb{w({ha=56iC>i~G8CoXdd<d+}}vmPIJtHjPTWdfKL
zqc4FkD&#@edz0zlZ5*2M=lP4VEznE@@ESmj+za<xK+ShfT}gt7P)4X&lp-_{v8+@C
z$|6v)m=%Gh2x3uI439XKs`1UUp&8$CzFL+Qx^X%!TkAR)R|i_GA4Wt-Y3@>(KFx&E
zp<<*`jCSPi)dRd3a&xFWRE=NLOujr{&GM!V<ZYnF)MGm<CY8@t8ypBpd47C09F<y?
zggMbnq9RelUYNzp;<KTag1QuvFh80p14S8VScgX78G(<f(9Cb3{tc3lLo{Oog$Xnm
zP&|aMd%yPJ8j?k@G+4$u&6VcL*h^bL+5$2>l>BsCNMFNx*%0}Q4W7Vi?(WDVf*74I
zMqikyt)r#Avqo!S#9A1GW;9a_l46jtIFx{-#9A%`&oYprJ?^B1m}diCiv=Lw%rk+l
zbohwyF)SF-l4u!Q9i>ERB_(<$O3@MlB3=e(2>v#fHC4V+zT?qbFU4i7sh#7Cxr4XU
zOgnhA!>j0?W;(#5!&>eENe{?a67+(k*IG7#rwOEZ9_JV-)Zh?f9T`Dr1b>VKrx`Ic
zk1|A+bW&lKKUNZF8rRV>%UFyhe2EV$`N{m0SPnu)d9VsmEseL?aPPzahLb)YA&3yM
zg@|^Ec4sA}i_%I;<Rv=e1vFCxf+7&HWGM!hVsK~W5^yQ8mdk)w2F^$tn&}2XH;7pL
zdcdUz+*!F7Tzai#6YxynjChvbiRsEn*FXD5C;YhM_xtZP<nMneT2$<i-dNqVGHi&y
zMl2TLs`(*-Lv^R}s=|B5R3~y<L_kCk!pkKP1{6UMLdGSK1aw0XLVye*w++5J@V;*f
zOFwKI9;SY&6!FDEC^Yy^L3cD$3;bH}VPzx8jG#n<(o74;TR?@CsJ;7ct$WFXeM6dt
z#`J4#F^<1nCp}Wil@1?#?NSuY7`O&2sS&u1@PD~}IPZLO@~vq@#IZp!%JtR<1DlC2
z1y8|e8N2~_8-UL;vI}@!z{k9fpT3w<J?Y`?A?4LHwlE~DCVtOLl?JCZX#s=pL$(xc
zK(~PyL;Sg;y|wkupZ_8u4&~9)s{d}r)0;ZO5OKuUDPfHHX^Ld0nWg-tKCF}_%Ti*W
zqM2NIt_l%8rx>$Zo8(nCBt+ZeADvee&vXWwsREZO2tqJ^{j)-wv8`m|5R<<ieopmQ
zl?Mzdh<i`xJf2qmei)Z#0tJC0mO?89D@B+d&71<kDG*_GXr>ti&8+;#{hg85kGk}e
zH=3uw(`qm-5G)8l{2o*sYaDlQl<$zjd+72X)B55h**}`$KtXe%;OR-{e<=RGxJK6g
z;Lr*4OG^_vhfOG7lRt+|s0}o2Fb5MtE7Fk6|9s$_(Wfn&Z~g%r^a~B+%paJqe4g63
ZCXKGN86j||+|6GN{*Ay76eVCk{tvg7ZleGI

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_functions_threadwise.hpp.7A06A2856650B185.idx b/.cache/clangd/index/reduction_functions_threadwise.hpp.7A06A2856650B185.idx
new file mode 100755
index 0000000000000000000000000000000000000000..997f46ae534e4515117d2456c8ac2cca72b20213
GIT binary patch
literal 2020
zcmYjR3rtgI6u!OG0`<yYsx5N+kn$+)T3C&pnb3hiTWX<o0=DvSQYk13$)ILaH!}n1
zG`^;$Zm2Y7s8h2cD(FCgWlmAlp@@%Y<HR|)>D<`lF`bY7_nKpE(l0$dkN^9>bIzYL
z!)*2k6GVo4#?s2tB`z*O5aIaiSyElTi9-<k@gGZfc5FFkKdBR#j@;9nD9+tq>^-WD
zuH)Y4DdaU)ZNjzssSfGQe`<3S&l|2QV)v+<r&mPx`ggQ-Z;f5>>A}YSO1-6kY%i`b
z%Nm@+HwCgzUrV3&{O!^#>9ISFu19|2LRoQSo8sCfUHP~v5879&W)?oSr%u0@|IVm>
zt}5<g$E@v5`)Ui{`02;3^J+(qtnXS-|A(u7<j7DJHS0xF(lc4h#*tMA8uimNT2A&=
z&5rFHs!3|P+gtY{v)c8|n*+oKo<hIg=>IbPK&!D|vDbQGjC7hjJHGa_=e6b1__<Gh
z5oR~%J57^Y)@%sUubnb-=4^xgesb)v?3zElf_tO$*7y&XJ@tiZ`t&Kwi~jTWj%u>X
zWZ_ZH#Q0Ke(a=_{wmo~>6<O~3swV4qZF?eHpGGcj<z?*KJ-?*wlq;gAZxxZVa&wgK
z^Xz{qou)ac+^ewGu6@w0c~~W@v1J&yy?46y*%@@AyY1>H#BEuGzjFD>g^%X-yRJ5u
zMVgMON)ENw_Y8(T+;!MSc$QVpzk>TR_SfAdZ_(;GLNb^lkX>R|ut0?1-r-L~f4*S0
z%jFZ1;bRFTp(RQdz)km_?bCG}suz)A*gcV+s9^!zladba3HKoZ&c?+}v`NVVxMTV!
zq}=hlzYvi^>^9OyB@5s-y60(ESU+eKk>hdk2s$Ex1#o{n=c9wuzW%v;EXl*}D1H>h
z0=R1mhFtR(61Q+RE}llGC9nYQRc9`DdAri4h{$m`TS`lnEP#96;sMDe`;7)H3wFmy
zViYWZdwYV__ru_12hPUD&9qs`0=Q4C{;grf?{(*RWGJ6Nslll+ED)hVoG}E!L7YEx
zc2!k(C_3@Yae11il`KI1;L7VxSE;cT%ZkgV)9D%(!2Q_smF@YJ>!Um}1eX^Di)a>z
zP%inhJW*X*=83}_2i@iH*{%Ywak@3z7kMY&q$osZ(l%g2AxWmh1}&$M)FjOUohYQy
zX{LfK*qW)C9k7g<MjLdKLZ(cU4O&kjv&ozVdQm7}pPvM*C{&;?c+D!*7rtf{>5E=j
zz=`Cu;|BszNJXoZKm&#`h|dmTAV{i~CI#-H)F=%bz4nT>&dcutOI(+kPSA%!4%(pv
zns6~(mzmK(kVR;b2drFiZZxDMxSuA=$Q9p{TqK_)R{#|*vP`llAO+@|%V*~rf^bjN
zwBfa$s3iutKZVpPb+Yiu@l#=uo0>8&?@ZoNVDfM&q_ioWa48&=t#&{Vh4O59&PDIL
zPY?5Ywou5K?Q{^+wnVw(_nw6n!sEg04={k$!Ni!o0lpM6s*MIv1GiaiGJt@%uF9DV
zbwO+_E3;WpmqPMrc_LWrtU+}TH@$e9#O%dNS%7_>I8O$Ap|_YJvxJZf^MnszmM_52
z5^YHY>)W#1=Z*ZH0enCVHzw@mgGLY(vUBaPtm6hn_}w)l8?c6PA#p-D4^I}Q5(li@
zkX#`&7f%P}5Q7D$VTR12H~TbQrR|%KUa}5fG_1`-915ifQeJU_w&Gc0wjS<-l@ZG2
z0V_@%7Y&F1HRwwnI4bxXSDGN6poJiKBXA|=fdCqK-_+(L&<8hPlV<=iAr(_)9u1(1
gUyBMCqmHM##1p~%5HRragVMaCiSDeGO&kI7Kj#6(IRF3v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_operator.hpp.28F266F9F2F32A7B.idx b/.cache/clangd/index/reduction_operator.hpp.28F266F9F2F32A7B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc9e0fe29b0730ca305690f09e463ad0e1ea8651
GIT binary patch
literal 7810
zcmb_gX;@QNw?2Ct0r7-@5tIl*;3Pl;0YQj?0x1d#h@j=RT8C;KP^)c09B>vtqzd--
zLvR)WKb2Ob4t1cS2r3j5g%nXh6h|D}N=0qO_U>?4cklbhU)SgHEZ(!$+0&ZdopF&7
z5laVi-0&HZOXkl^o+jZq&W8M7keoE<crT87*_Y#%>^%KZ^p}Hl);Eu?yx?r{NuO_j
zOt}$1+)(#Znk>R9Ajvl3UeRWsEjNBwj_h;e)=lNdvXRY|QEelvKZ;HoeCyJ3zxc%G
z9T%T8<&^BJ{HgZSfxoYawpx4Sf$_ZA%LC~{O1nB{*#&sn&$GyS=;b~5L3qI7(32~+
z$~N8mr{%|GJNDeU!(W>|?0Min>+Z~ZIi}`$NkMr))bbM*{wo7NzP+yE<AU@Nwdo1>
zX2+LKPw4IHsl9eD;848TpZn#9tP0%v)U0;34>xl6Q|bddX5igky4BW&bNu(swB0yr
zY1p1uAD$dB_P3X2DYESkd;NI1I7c|kZLdrJpwQ7~XoF|<pG|F@7Juh|rD{sI+#xG;
zm5%h;W++^LTasjJnL5=z`s>U0-VT_nk^Sep+t&uf4arV!Nb|p*=Kt%a`)9Le4}CoS
zKWOL5<5?b#k=C6XHe2Ozw`TeHk974)`C}PZ^j<&Z#e?#thiu1BDcNOvtNr#F_0ihk
z_I@81W?qTdJ|yqoO*;=o#0}~<=?BH$`n^M!s}rWh=wH>mXu3IgR?&*~bAkU_^wDOo
z;-6#FcHh)^@0;|p<Ca;kWoJHJl(6^kn%xJBq7L%;E$V`<Z-Z|L<|}Ke_B?8<wJ7e|
zeSFTg`Oz__GoM<XI<V*Nr0?DvxwLJPVO7<dpjneY-xxnD=QFEg9dEKy4V$iS_h0r_
z$k~vRcQrw=HYR6kqV1w$&AraHXR~Af?sfOQ`zNg)el$dWcG2sX#eRFYT=FgUxwgJ<
zT_snxX5*SqU+xX6{p*_YV7<pZ&%}W?$8N4pxZ5y1`%;MVGx@1~)AI}cip**s3|U#e
zsLN?Wx!iB$;MA#qk3To&P`a0XwMW;kYT?m`5yzh8%Wn_=FgWSlx%2yXx-~nsHEdlv
zx@4r!x>v{U*q6`#Yjv`{!(QVMYt^tD4MNGkm-_uR%xm&pX>8x?$i{!W?ep0WR{m<R
zxY1FSXgwyY>-gK9iov~ZrsfZ|eDmH7{WiO*yH97V-P9W0Qn2`y`uM?xnRiY%#r$)p
zdW^<cRQ6HVpoNKl3~^aJUh(!s+Q}WY9j>n&9KDQ=iKVLh^_NS0R>eJY@d#e1+28V|
zv17#CuYdTu`at!8Kd&S$KH0SD*e_kbwO!NK{^OIoXmON{y7~E-hLJ(${W?C_JM?+)
z_{QV;M_#(uA6VGbyez5jy^_<<Yqq3%e6d0?DkCUqUGjICbz{eQ{Q5fRV9QH;{zBeR
z-7F8iQ%3(Tzn0cd=V!DgFP~)|iWc~mwB5eECwRs$>E%<BJ3Twjd;Q&OU$(n(LCXBk
zj`ij^tQds1ePM9!hk0jCn5dZ;)cxQrNQRSt=2;tsGmB%i*4S+@CuAdSHVWtwU?Q7m
zx#+a8GtmrN*_jKq*t`~d_Xset4tKqIz<)|qKP(%_37!&9Rrdqba+{|gw4IJgF^$k$
zD8uGuWCT(CuZiKom~tQUV6qM;PB0>^k*q@$U}9hRDkb2|r~)(G$Bq+Hv3V-)+atI>
z-J3xI@`#*p)bGc171=IXiCUu$>VCi&njh{?p67XElQ>N$(sq(*L;)tAN#hIeOZV?4
zCIdJj3v07*K#u^EfC+d0dQ}~Emuy$b2@#SARrdqbBd-NNU3H?enT$ZxG1{1b?gyx&
za+Bv3>_7O-49m$rg3W_%x*v28v9DJ*8)kl@#@OAqdtaQuTk_UD2Hd!~%jV8qXEu@9
z$x&TIl8Z>yBVdYgr*}1;FI=2VPJ!$p4QtYHP>%qU|FRE0x%>Q!>*7hYBW*i52~mK_
zc+G}ev2RaAkWt7!3>rgF_XE^Zthbk1mS>zZ!!~3eCCI!4S@sCJm)WsdH3Q~2Z!)>w
zOsK@>mBjU;fN>J9o;mSN=aI=~*qV%!Xr5@<{h(tcO-QtV`qlf1WE~_%lQqdf-4FPF
zMTyO|oTRj2<UB}(Pm@ejbw5D8x_$D?#E8}X$wtZ7MMzsjHYy4*NmU%SusIwZNK8o3
z_-XwDx*wqadaUoSwM+J2GQ<7I0>_(=x9on<1%CY~IOB?)?QC1@Vb2Mtu;LVU>JczT
z&Y0*yUV%0T$R3E(N|B}%1@#Cp*@4{O>JP*gl2Hb6LOE8HlTkzgCI#)=mAlfrc8Pmg
zkF@K_UPJ*V-$%Ab9i5_CM@-0RBxn-?x*wn}UYK@rhO429s7YkxVr?#oZc%{AVe52t
z`NYqw$T}Q3;R;q<A?pwYn4B7Ms{N0sPrfHdPDaVVk_@cs5nxiO7+t)gxj#8HD?5qM
zh|L>`<3#}`7yW*}{L|=^R<agyCap-@N_Ha(Fu8K_{Nu~+7L`0FB}ZS3`C>e@M{s>C
zFxQ_WIp6XqSY`RhG1xQ}2}EpOLZ%W0*QXaEDMXBJ`A@pD=IDKFhAoMatGTOX_rv5w
z!f%K{+j!)^Uo$bDY+Jzz*RkR{*|sQPHFpZP?VfT(`aan<38hpdNhO&l3NU%}#e&~9
z&WP?5J(P>IxnxhG0F#ahHcdHi#=IdDIdH<ySn)HNNEBegCC!|*U|WA$y-8-7N;gDe
zucT!``g0FL`|?77D!>!smlxu~;v%8^@WNcxTu+DxUdTjYnJ5xMh!^rvSU!q`1@ppH
z6m}Iw%6keiFFZzJk5MGl30}y-VL3Pw$|Lz6hvnl)2q<1K;t(Sahv?yj7OZH&N;p<%
zGzj;!7f5^QvBEO<-d~-z=LNMy&5H$(<9I<M(J;tQ;>RF^#K53%NjQUIBryz{D4EEh
zsgkJ-N{}QlC|Q!spfyOc1~GU&lB{QOHj-qsI2TEBSzLr9MJz5wl2R7eAW032jYwi-
zaVwIvvbY^d+F9I*B%Lfy!;&=2tak&JY+!K~mSnLw7fW&}mJMtO^ZDiG=U5`>t@JVo
zD2Whn6)%H;0wO?yQDgzosG$85g{k*LszKW)uf=8$_7Gzv7Q_}0tz_V}!xmlpPxz8$
z#P^Gr!^9F{zSVpgL^ALIGOS1fU|5&XA4@uh)mgstYaj^ZYZE(wwE{69OA_e-ED7SB
zC`2v*)B8-q4px@K^2~T)$-pHlXi~heLcYS8qO)k=S)_vU&kGlj`~q^OxC+UtkORd>
zvF%ap0(Ea_-KcvHbG#H5ys#O`HzOxVBVO2o<Xc!=jO4}03A!CG{D9;?u=o~|-$G7M
zb9tc+$=g``3dvs~CpZUQc#Y(*DF(_zj?Kh(fWv@?$my7P4!j0QtWEp|9266)8}M`Z
z*0JWi;H~oxgs|a-NL^$g%)krNb<+c(sq;b_(xst5D80OJ8R;&gKp;W9@D0{|g9D)>
z@gy{(Yw;K;BVb2XA~GfLD^Mh&5cv`K6$&2Nl1PW(3!p>f3lkTDJ&u2~b;yHLYlprw
zbzltgtw@yMNIqS%^UJBHOYcf~VUf!s4;nWq&M9sbWg~qyGEnDdqws7rn&Kj)FG2?D
z{2~-ygho?*2DzL;9yH3zk#jk6qqr66TakeVM=J_%MWZS1MEXu-puy3J!aLDuitBMm
zJ+1vSV?H>&=J3AH`tZV7)mRM#2`_9!s*OlPTTlT~6(9|5L5)b&h%`_hc;N|BJwY1q
z2+0Ag+Ke@jlDtrWRRvhXwjhNOJJS}_j6<672xvjTMTrwc1`1A?x9Kyt5x1ij_vQt6
znfo9p5xfv43mXJ(=Y`3#$%EiJ#S7_JmW~HO81cd(EI)*uAaS9)k-1Ea7G?sElGi34
z3sy^l=GLrIUUEb{p-9n<G<+cqh`ou}LjMQHk{A`4EDQvNn<z{)7OLCYEyElyp_Qw8
z@q$*THPFNxr5k0SiC2cB%WyoT&cF7oS{hS$p$e1Gv|Th9md*=jknI^X7=kDL_izXO
zl}rV3cAP4XadswBWwOrBN2+|**;kS3D(mdWNcEU?b`DnMu+GlMs(jYj^;l8=f1Ryp
z#?IhuI4W{-CdLj90IE)2i##33<Hm<a5<mRl$$$Sp#sHiUnM)-6Pz8YL6NSj{p;PYM
zrMa`bVxjJxwFJ_K_*kU=;CyJ1L}B9nZ~_OOcts^%*fN0l@d)-lf`>tj@+1wSE2+aF
zkI47pjQ~m_BotY%cn<(3aO}y_V*@sHeE-gldvaFWr4g1jw@n03svZF0MF^f$!yrGY
zAA<~11B1e);S7qA#xQ82bRvVMN~ba?L7KpzWN9*k)*$H`#NhQvx}L?^NSe*!TqMn9
zaS@UhvA7gTOIcilq%|xyBB_zZtw`F+;&vo$XK^QzcCt7POVcp3-VIo~fyG%^n#JN=
zEX}1D`Xh0ncwGbEmi~No!oi?F?!VK8Wl`&0p4EO;A>)M#>{5Z<Ac9Req5?-#T#a3-
zu^ZfRc%d3cRO4uh>#$25c7urGg*qHjhodR3$FB9*oyDW-aSX*q>}q87Mm);M>YK1j
z6Ly0tzza<{q6tS++>BkDu{(=LHsjG0w_tK{@c>-@ZR(ei@j`!6IG1Rb45c-2g=&R{
zqNCXDDE5N9f~zk%Yw^Mh#iz<xfAQku7iZs@-q3VO7K(Ras2W$^S1%tI>#+Hqo`N-y
zxy9Qu%zgdE&T+H1Zb*OU+)xaA6PnU}jfSsz__?kpdrhUo^ivK^X7NTqanDu1=LIw(
zvMkfp8-|5rCa=YtHXPmUVONH~DjJC1xmVzX<ZJN~4uud1g{hEr-@rlf?9HUtF+0}j
zhVg>GoxdY=OM-m;^%Mmt0+bZR+Qm9j3KQ!atEVVV5vOF3E-r+kv5K*bZjRj?M<~#|
zFvoX}o}#&mxk`#MkX;6HgeuPq8OS#S=_$@cicF+ru`UyZP`nW-HnRFWWS55=;Y@fT
z5BcUHJ;nJ*k&l!t*5#uRiVKjUfYmo3y9VS)Nm>K)Z9sa8uOh`&q-3$~Dhi>v5h)s3
z{Uc=e2su)c_XznuLVAiHBgJE+WU=ls3ZeK3QaoYx+1M@{J5rLEjeWDRp5h#=$iYe$
z>vC`i#hbBWGpo<Tc6rzlnj<+s?3;)66z5|_K31|=mybgzF2ITcR$qZFDzJ=J%nGcj
zz`hh$V~c7mqZP9nYpSs?#dX-C4$Ejotizf*>`QSywy0<I_1Lo>`%%3STNqir5qlb0
zeG|55!ZJurQw7yDVPA@yu|+ehZ^oX@*pKR4utf{i!@5b5iobCI3-8!jQs#X~DZDeq
z-?;wNq-?U^LhA2Xq&<s<0H-otq*WJ?JH=H<S%n0^Z{F-H^S-h9MSouKRC|U}#;bME
z`cvfX;LV^2bwns-!=oHV`BN0>5Xqov>S>|C&B=w;VY)v>sYsoQLZLXBF0c-1$e-ei
zNPQ87!o}2dp>?>7{3*`B>I@tT4cm0Vb@&GRQ(TPwi}4sJ^d?f*{|Ie9rb_Koh9hYu
zuEai-EUv{qwJdJLK8+Or?Q<|R$@^nB!WL8HDNeM<l_PmMa)R?EzfjJbW}0I`()0yE
kra#oChxkwUuzv6Py8}+{khr+2)SL^v=QvmTr;7ghKe$Bv$p8QV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reduction_operator_mapping.hpp.8A11EE4B065A6CAE.idx b/.cache/clangd/index/reduction_operator_mapping.hpp.8A11EE4B065A6CAE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0845ac410108dd8aea71db0b3fdb0054c50e5b7b
GIT binary patch
literal 5628
zcmY*c3tWt87k{5I%=Y%mphjkzcWR;@)QFVIehE|V&8#~qE0U<qR@;Ry32jkqE+Lns
z%Z5!Qa&7gw?6THn<+e+N#agWW?1k+9mhZgpJExhi-_LoR_nhZB=Q)?>{Aa@Gu&_Uz
z7-q!W(dh{>DYNAa!*qpz$tg*3*)oPvvka5Iy<u|XLj3^89kE$oRi?Ba-1XwNnKj(;
z9v-pl$k^6i^J>%=dbxe`y+h2qzdd}{9#Y>4I^MD8{kzGRj}F#(eslfJvzWi~rk`9?
z(A?Lz*{8Yb_@cMBOjTp-oNwt`rcIr*_smU?k4FxE`RpOH!SuZ2@w&^6s}e@fsa&#u
zkhZ1eOxf!dRfd3x>5t1R{06DsescTHF9ZJj)vk5(x4&B2rp$lam~YgDefD~@ZhP4&
zKJWE$pYl^Y|6%97IenkBhMdi-*+AyBc6?0l_#O^up1=FYF&>AXyn1lH*NC2>pY?dp
zA=!KPX|8(nT!(R$0mVt9#|M6}kUh8O(7k_?^ZMM`^M+k4?KG^zZ;X>qOA0q*&4rgv
zr>YCRe0+xO3te)>uKvD{zNm0caA=3vne*C)SII6fQ_Gu%vj;!uSMui4k%CVque-GI
zXV-?7<hqXay*DRs*)+L4=hg0#(wK95xP*Ilt9q+for-eV(bV>zwm%;~FL+m7y7w33
zPj@S}#q7V{Sis(o*iyFhtf{|msc&nlvf6kml^kBy|J>^Do~LGp7T3jWowBp-Ze-4+
z=Ti>vX47gyi<_FNrZ!w<UpIvoz0SS*#RAj(sW)5qEXlgJ_L0Mjao4^Hj}3bGt*5W{
z(9J26h8S*d`f;D#JI%<eUl-VqvvadAa-TkJYsQEfUH6q-8MnKl>|jf`&y)7W*Bu$1
z5VCTt$7;?lqT5uv>dJp2J(D8~xvAUT2CO0RLEF*?H{bo)pqi1jDd0wzu-}K;^*^in
zZRbY&vmMhOZR@(?%Uh;7iK_qZSk}CDOUGj08_91c*FRajIWu}=Qbpaov2B&_F22jV
zd~~?|IDgZWvOkVo+j6F}>ngtHVV9?sRq;%6MnZILCx)?ie)`3+gMZuE%beU88o@@m
ziiCL&sulhm#_EdaFAZn_u{%TK*f>{_pty(f_6pt@dc|JW9mEUR1#Ti~7w?WAv2|$n
zLlDE^GufGLB0;fw#{|Cjucv<iaW96Zv1x81L9zDHqS@<md|%ni3|fX#mMTTkF8(#L
z?sUWPaM<0Op|ymqC2p1lRdfr_zlf-*-_}j$sAp&%k>?S$WkD7Ft^w@Mx#!P!lJ$31
zP=AHLRwO9)eB?Ry%)s&aaDOmTL)aSPYFSWakgmn!{0xUQ;e>89ToJAnNxL}np?zsV
zhF2$<k8@|5qDXNO35o+RL?>6xe7M+2=FKxyuhzSZ1jWWbvfl*dg~tnrJJ2vim{uex
zHuq?rI-vj8GPoePS5M^ia6xH7l__oBHP7eotb`kQGPIPirEmjjL6vF$ES-0Aj?XZ`
z#jdo3C`yQnWkHqcvnKv-_=_D0xll7SfDKTJ1jTcs{>cr}FRzDNxG;2}dZ4>VP&{w8
z`fc%VIceY~)WIzIEVW2bJpZHKx#MduJrgczPdyYKT9KeQF43o<Xk~|`f_q(PHBnT9
zd(wg`i_(Uj(M*Xqg8>hQZX)tc5C>^Nl{B_z>XtKeuY(E{+hX}*wMbB$o_!>s>Dbjm
z!HAL;5=9{xkrq^0(zfRKqf}=`a8qm!hIUtU*D`i;nTlZuC0Re^<SLH_y#^D|4GHoD
zwMcM(?yUEV%0B(vESO;Ebj5TRkuY{0NDp!Us^HkG*8*c-3kK|Hup(G165PMqdGh*}
zlbj3WK*!Kg>?oy3P`u{KkmLthL$5(UK(Xf#HV67aT2N(!<D@O$hn><vntC!+tya5>
z1jQR~K6O{NbsGhx0#RH~*!55<(t;`#3d5(JTTZ+MA6yw~X3a{GptvemPELOJ;sA)j
zfHMoB6bOp<Jm}NgW#hh`AcivWVSSV$L2-3!(~{Pz?gZR}3yx8baTf`S>n1Ket~+<a
z8^ln9LFyoPk)ZfMLu==?K6-yR96HIBbybQ4#fKZKGD5a3*9pO8sX<}TiUh@t*^0;s
z`E&{F)-W`L4N-~&#a{>X`lSEJ_x~m2?mZf-h;<POiZ4#S&knqmJsn(wuuM=-a2E-R
zFXjBvI5f+6RoLB$E+>lRP+Zc2D%VC`E7&%vK1Ohlqr8IGiUh?smqsj>9rn^dY9P5Q
z3A++fBQ2=%Lzyo9LaW0Q$UijtF2e4D{7Va}+|Bwuibnr6Rxr|$suU`%NKkzLsH|~=
z^E)-vFPu<8<Q33G(t;{abMj0%k&{Z`q7bIx>f!DpLGiP(uf8CDZNj*KCQD^gT}6W8
zw(6O2wUvM8!ftrc`>Ol8iv-1Q;>O#DXT8wE4WRxu5OxFHKw40RNs5`59L_OJhmZx8
z-A=yDA{;ds3<IQogwIJ!h6J8|JXFEaS(;gTe0VsTph?hUe{yuOX0aY01P&&GCXeWG
zSa9%gXf_c&z9}4pKvO~V7#$8CSWP|A<J+rnT-N7%d5)8uqv3{dU-X27*TcJz47Bnk
z#Jj|nuO{BrRvrg&2UZ-w_*if>N)_d0y+^bz+7CG#&ag$8%vhrwjaS8cSudNYOY}ny
z2R3XGrnSu$LwqPIaWpY7(TrRV%Mgbx!UVT*%n%idu5dIuFxre<&)u&x{7;N3gX9Eg
z0(3Y$aNfg7!muA?nT#CGFl2;U(~%jNX+{n)0Aa=1kCQ@cnf-8*H285waCAR0>?fg^
z9F8`Szy@MQ9+M2`NpJvL6LEly73Y6lNx!C;W9tra5*n~15zzpcm7svFS=D~xg=Pe^
zx(4EhyjR|@Z(No>CNenB;BQkC%Zc}L(q6p?CBO**M_8Q@I4mR#8(cv^6Q~rBp7+tl
zWqr%Q#Cl>}PfYllb1?0VrNo3l01*f57vKgyJcu;Fwg5BmSWdzl7UN-qUT)BrcijPr
zU0_^b!fF~b{qw6`b~kr}u8UGeIbj(<Au8jYu#O<j$_yt=EJt&RGM6}ELO5DPltskJ
z+I8hbSx%g+UALbo_Y)_K==j0w6II>?g<vwm7-7Qh;pj}`OcRzYN8^lfCTtUqrWw;r
zScDwRMq>^!;n?Khr7><GCVYfANU^bom@wyP7{Vq2EX@5Ba<8R#+%Q*&32)%rbBW=G
zT9MEeqAZS<6ID6!vfiwU=&Fbx@~8)ikf4|~6%vnyfNcdH34jbMI3l(?Lu7B)f<GPx
zk9L2elv^9wlliCNGu2QY8w-&e93sFroRXMA_$<IRY<lcJ@J)beR?h?u{ldV2l7!N%
zY8KRH(ZpPX5ecgy2>bZ@iNkCEtd9g=eKbBi!Un7z&=>F<HUjPfdlF!?+Oq&Si#-Hz
z#7+cmsD=d!#|oGY-$cB#oE2djYID-yw;ahyV7{HR40Q{i_rbsy7t^aQz&%_wt~??X
z#ChPb0P?IS3!G&ugJ=)D5}+QoCzdOem4xl^0TUX^kV`_X4V6y<^NAUG#C~v|0Q?XN
z;41;=Nk|W)DKw<2h<IU=I9g0}#l#Q!^5-}1m;Fid;Jpbp1h<zNgvUG92QPte5sV60
z5#Kk2iSSu~hv*CD89ocx5Nome!}%WX<QvyO6jcUQyCp=6pe;Z}3{dTXq;qZCJ+{Nc
z5bYSvV<E##b4=th)f~-q%;d42IGRr!^9hfk60Yi4OnAI1N2`cq72&Z%h1cD&f$;bw
zATR|_1XPNRi0OvU0vg4K7Ow@_7f`4*`vQl7LpTb1Lja3n#b9m1X90ua-fuT0f9c(8
zS~lcj7cuT4;}HvTFdmJyWIWDG#Hp}f0I8S{tReU;6)+x*K7+3UB*o7XKt*3~c-ETp
zLJfPtda5yX9AZEY@TqYn8HdmnSEw`zxS~~|a#gv+%NqWCqRS_K$m3;TqXcL1g%n;=
zLpcex=B0`RRuMDuSXbb^fOW0j3mkgi>NoR2_JKz-fN#zkXC6zFqam6Q9@D_lQJPUa
zwj@W*8Z(czjn*J|5>`g{emvH5fY(dU6-54jc<Sec*Rm&ViF)4w%9&Srp&t-ap10vV
zcn=%Sz#D8hf8O7Q3+97uxG+A<h70GzZMax%Y(J~L6m5zPx14B~6C1veXbWxm5~3}!
z<*SLd+REea0tkZeqX6p_YXMRx{20I&8=!zH!^_%!nYv6r<eqPgzWOMnmV;^vGlbc|
z=@jpjcHo3RB*2{TD+0$7&TcR#{cNy(%nU^&)Y{R-B(Rv6kxz<=pDhF?IcZMv5aB<x
jqHbdUm#cN%<^B5h@nZT4-=x3xa57ALv>2ujiW%mA;ftG7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_avgpool_bwd.hpp.EF9EF399DB15462F.idx b/.cache/clangd/index/reference_avgpool_bwd.hpp.EF9EF399DB15462F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4b0c748f9e4b37847f9b6796b2d37d3546dfc2b2
GIT binary patch
literal 6474
zcmcgxeLR(CA3wj#A#}T)a9lVJ&LNVQ9B-K<Bw@X4m5oNlN5$H(C%vwhrMLBhrB$PP
zD#dDRH5HYl)yguxC_Y*B<k6I5l_ejw_FT^K>%MD$J%6n8;kta!?{{DK^?UzamkCi3
z5n<L6$?zFbS!qd`iBgF~(vSUSWTwwFClbkEeTgLN>(Ytwvm3oEo^~u=VOJhpW*&8W
zlue4F;mJWy-|362h7Y*DZqaIYqx&~v2R#2hFJ{Q&BiW(nDyzyrY_YxV=C&m+Xv3j}
zc}Xc23+az`$x}QEoWATHS?yW)ll4Pet611pQ!^^}h3&&d_vXvGCtg+gk9KyE#dO*<
zTpChoeN3;|eQo0=gQX8HNODHcX;>9Mm5*&`&W_)knslHrdBM7?&DZIt`q{b9LOi;1
zgSM)-Y&-W+rAx=b{RM{$LzLbvch=VJYfrm7Ze+r2%PS-OEINi&o7kFtu&a+v8?_w0
z?YZIXM+xc2($<Wq%RKtth3RQG{2u+Fbo4qsKT+8=^uzpuk2m{Y`f%BI2Mc2^f7bj?
zX}tZk;;{eDnr0l<emUnUXY#yz<DutY)}&Sk|M+yqDa`|U*Wl{wnSsjb^W(-g>^+{o
zQng`veL8isHpsrUJb(0%uj#Cu%pL03UD;n`Oub>8WxYEW6zy%6CpJd3##dH$1Y{~Y
zTRavog__IF<&!EqO#V7)bmC*b$(IuNoA2(Qv1IO^jFv+)<L2F;JN){syU!=B9hTWV
z{wIz5wyLw^S6!R*r;DrK&WIb2JGsnRixUs6KQ?GyPRnGAP0}6TT>fgo?A7(B?fczc
z|9jJ!>6~M3pV&Y3!tcr6A8%x{GjViPlstU)JooBbzn)W6f3%|YW&19d6TdosUuQ9q
zfARRrs+AL*hQW=wzx^)Bn3p!qOiv<#ovYky%I-As#I~P=`Y3${Y6poS^B?wss{IOR
zf_R=dFtSFean}w=_Mg>l;`}gSsf?JIN$7ImmxH2rLYCmk3$`1^7@RR6`Z5U(=o?_B
z9bS>IIouu<oxk0T7&3Axk(CnL-U(U8-n;dUTg$1l2Bfd4gwj6LMmtE12$-`SQW&qA
z{pE?4r$5MuF<YaCXllst-U-*36gR{pb!1W<Pwbf-UMjBu?SSMd8&mof{^b{9K=jxO
zHhpc(wF8pl=cb&udAvrY6Em0(R?waS$<uwhYd^S~XJ|(Hu%{0Yn*+qJcS4q=#ASa}
zpB<$yBOFsy5#ftSzupO1(kk5FpU@d|P4|+sz@KHh)J@2eVg4edZE3ZnFVSPG1oDBF
z+5yRVHxH)VY8)9SBL-|Ahl$Z)rc2#~EZ>!^=qzoHIc7=pm>z5mZDra4$tSyhD$taD
z@GqVy8QE9q>!ux${NtKuBflTtX=O%?*b9CpvY&~fcS4qn7xkZ!(m!AFM9x<5bn<l9
z4oI#Gb(v*nv)MqWhi!!4#`K_@kmWxIrp}GCi20Q#7HkC%rH7MtKyq8=;qLqy^CEa+
z$H?AFZ%^%j<j(dd%U4hI<77n2ws?iet`KGKge))r{YTO6v*-K>QSdr1r@S%GON?|N
zJt;XOtjDo>BZF{s06##1nlskM2RaPK2V-*Pos|MpPpBYM-^+(CbX_f9ugB3WF3TCc
ziKAt}m4Pz`5l1%?ZX<C<N3g2@-tfiVquza_bXw@NuqFDx`CU45S|+7=5Sj;Jzg-wC
zJH;6>ya&Sfzz6}?LufsO$rhFrY|WAPm(u$Xeji2%xD$qTLWqEi2`wg8=%V)g(sx!H
zUcAFZ+CdaM$N;n%M=J^YaSHAYSyg#@O52j4Y$9chmOF?wGD;bjStXGpqk7m6WlsVw
z8NX#gJZ8KYq9lu|45c*LKRE!^E~T5ne=`K2+NHDz!ipeVz>VPF2mxpYDQ$(YRtOhx
z7mVqGaRM$NngZe@NMA_23yCjc@5tsg2G+6u25cM3WIORTl$*g$cx@557QsLPH-fSe
z?1a~}f@>=b6mS>#b%8%(d=89z6EG&;K)c7)vkej-gOt`2uAcB1BFG>$)@Ec(vmA?P
zm!F9z20KS<i9;=M!7ST@U2BOKVze@Qu`Ye%LSZ{TQ$OBZ7{lCtp>m-H1?Ol5C@VmN
z8|G*}QRWj33KDT*MSE&n@^&kZPLWShpxbjaRvs%t@z(M70xc~lElhYpHpsI<fdRqM
zl^|a!#;ZZTT8!6#e2o~d2kZ4<FFaoeL4^<|Y<LIs-vMp{y#yjlV2nUN5AyS%zyRau
z1(075<7$vsi}6K}UlijSkk^RuFChO#jIV(86|fi9Z-SvsFdP-g(Prr13~mCw9U|Ld
zj6i<`nny5LpudEumoOITXg^}Ct+SXeIT~mk=#C`_t<S8kE4cVnjM0dbX-Ts6z_GMz
zbcL9ryG<Np=5jA5NZ*vBc9wRwm?}AHZ)q<=4weogq_R|rklIo$LS80b=2+J`sxi@+
zV>affw~4noRtk>#nE04uI_IdbiLW^pI*uMC>Z4=;Ix$C&5%n=KE+gtPF)k<Saxp$m
z)W-#k`(eqmKi|JU3TdqEHn1K{O57B?e`jMLQ&A%^ZzPteB4l6@+0%odN;z$fMk2F=
zXP`gYfumK#w2Dv^0E2+prKd%~7>_^tPHLR9lOu~ooBlQ$VXY$KP(<7@_j9aWa63!}
zV<&-~4O>mu*5H=#0Y=a@Gx(yU_4@snm}P`9*{C}WK>nGIRTd5={9ZebCTJ3d3W1xb
zNfe=JnrR}G;++yKaL)9eDMC4*$$_Cl2<Cz&SB&#OlPAVoz<Ub>3;bKbd#f1l0^eN_
zBG7k(?`|>P1HOC2cpn7sgJ^-j7=nw%_!?-g!B8Q@u7l>f7}tZQUW{*o&rKL6@ZSQT
zTVmV-zAZw{=IA}}y(h-^!S}uxw?S|lL<{RZgy4r_`~;$&z<UC{1EM;__$fp^72|G*
z?uJPM|8t0bF5uL>X)fV+qUW;MP8*o!k50(ZIj(d31eB@FbQRF2?w`IY${a0(fra2N
z<j@lEFA?Y51K@Q4f&_XgC`-Xrz(>IAh?w34ZcQ*4a~en6!M`2C1pE>PzZB=pB_wnS
z87<JiAj&U@tAN)L-*qGeG0x2YbuBvjKgOkjv9>kGuQ?j54jw4Tz8uucg<8+CMbu}7
z+Rm|-*lHVbz!rjI{h3=a@xiFeD7kVXcU8h7eU6?a>XXD7qrV47pCn@hTtRFrh=UkM
zRFG%^pCZarM1^>5wBF|rH~baQlcQC312BED9Idh!AqSO%2&q&m5mKwvBIIS|Wsf<L
z<!B3yLX@K|ycMDxZQ-L3<!B3Eg*Zoh93{TE$(|hTaZHTMh)0<imlKb2F+NT_jth8g
zYxwn)<n$N+a6`84-l0!oLLY1}cryxK_xm^f`b|H6(=T85Lrm1n4|NAUOt_ow46F>P
z*xkzVa2>I#BTiUjHa$y-UuJ51ja8_<#A+{bLZ|A%L3>Fg;`|Gdd#`3Mc4sTp5{p`5
zk0pO+>buhD@!3tRp==;rgV0d!9^R-7tLonNw<q>JirhJMdGU>&PEfv3A%sB%$SXi0
z1aUr*=Mx16*8!ivac2t>HnXh;I0Xz6VyT8W)sR8BAM{eDgPsE|*0$0I^GoTPCnazB
z)N#cB&{qpCw?#7T)DiPKp}3zoe=YXFkdC`=c|RCMOo^_%>#O$3SbMNXU7bq*Z+qN}
z`Q!AYbdGwC6B<ZL_k-tt@InJg=@JsSgao01&g>iakx7K%=YKOeo*kG;bhiZPmeq=^
zCZjA~UtjuKe<XeF^em+=yo)`WRr=cTTl$(5c-z2)bK;#W@eyO~odXK@i{pY-%YtJ}
z*)AK1MFX+IP^b-ZUN~W_ZT?%f2xfJr9No<aIyT;XFliilN4X$P&m{;-b0?s7__O_1
z@Rm1y*GvD8KBjanQ(_J2S3?vSeGl@!?JRJs-OkdSZ9iLzo93vJDqr1)aMZ)lgGc!|
z>dAXrpzSzzY*dyKM<Kl*A&y6gD`G4WOe<oyFxFoCpjMuIU3O~FL6y^6)=G!$e%tcZ
zV}1YOQqsBT++xe;r%nIq*Yw_<F~`(hoMF^<`FwIF^4oJE|7UtgPfATpW|wN=`};Oo
z{QlWwz|!*W0{e=kHKCsXmM=V>NYI{mjF8|KlC6@qCXaZw5-@nNI7l#hGSa7K4C%ST
lqb*tcx;{2BT_v@)v9yv{_I#)N*x(-$iIwmt`_Bn}{s*WC{Rsd7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_batched_gemm.hpp.AF38237B103D733E.idx b/.cache/clangd/index/reference_batched_gemm.hpp.AF38237B103D733E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc7a61431d9f25a8f9fea6a6469502d7cca1f187
GIT binary patch
literal 7908
zcmd5=dstM}7C(DCyoTdpW*BA$W`>6h9AHo!QKUc(umsB#Oe)Q;WJ>tR2NH>XzVlIr
zdJ&a)7>`5(2}x70sF<dvT_GhbABYlOy@-e@KJLwX*UUMKGxzJS{_(qhxEFh`#Xf7V
z=Wp+%gqL2r?9DK7?<RaWd)k5&8N)D2_@BODewqVim_}!Y`LN*1_%R#u2P)dj=l<MM
zn~}3bKEWK2=F$Gb!ArZc_O8!!aGiSao9{KddX9YMVzTtTHG1T*Jt|fBZ!L8Tr^emw
zzMkUs%f#ir0jqu+d;G6m0~Qsm(z<=!r{*U4%B{ZEsjj;$s555Hu$$p)RI3i9Ivx&b
zJU%G+o3hN7<2hvyTkhwSxheaXtX-#GlA>>U?emV@ExAvUa$RnotZZHS*5%n5bw8S-
z-#$^<-)TmRd`ZT8JHC3#4!^lj_4HQJ)1x0%uQ+h_%N>TQ`Uv+qyQcSZZGOk!wejf1
z%gLh?X8T<Uzi`9pTIl4{NtVO82LdmCT9fUQLwk+~{^GK*d+)m?*Osh|YV-{XUHipz
zW#*tWNBg~f<%c!P_mdE%g6;g^jgZ^5w>{48*nj-Hh_sbuClkERHvROXHg&|q1bw6T
z_B%uNjcazelV0PyzA3w@Fil+;^pA5tr`8t-jM{tq&Xm{EGspU8<>z}{`C#=wl3UrU
z-H}(j4;*eg+!b}gFV?xeR<>FHSJknI`0p1rZgc*2<}tS9&dHqBCPPh!s$S{7`SoDm
ziJxCBFMVY~&_{hs8xPHmW$S*(^@{rR>yhr0|1~}{;Ls(d&*RU}o+Phc46o0NT6N^t
zFKX9^Pb$4}KPu`_LeiUM>6;U?(!TiB@pk)y)q6H;CcNaM{>u-VJ2BkUR}=aKH*6eO
z7&|rOLz;Ii>J*c{aP~W8P7KpW6`2|9w(kQKb@gTVTU2q2>P5mppjiLOVqP0}wZ7tK
zSL)D@;VURtK{X;l5vN&E2kCnYhij;VAH!!;=WMDF35qBdR8DbqpL<F{StG+|STf>l
z1G1Igo|f7pPYWEW1F$_^J>6{svNgAoW7Ip71Ay(%@SBM1CgLs<6!9IA?vau+ZazoV
zCWbF3-sQwB5)=tad^G&{$MqZGu<(&|MY`TLAUk-_#;$KqC9dbF&YR%}dJi<)24shH
zR=Kq*jyfu+Bj|9SDDD%ZNKho~aOA0NF7qN(R1P0VQ=}Pe1F{D{aEklof^#LX!PqTS
z(LxO(L6PW4$6*EUe6j&P0y|_WvW&I?*~4~5h7QhcSAs7B7`~ct)kGr_6dA6ME?Ihg
z+Du2<*TC=^Cym}VFfKAkeunTQW?5p}?HwOAbJPP2RZ9FyNvKG0k1<zX(b9>jLlsov
z!|)TiiAvjm?6D6QwSGDBe3Bz|gHL1=r)=UQ66rIEf%3KQu6^MZN9qKJEvK&KR3#GJ
zXF`@{Xh%g}A8;8MJj^@HY#WfB{OZez`->kK!0E8Vdg{8Kx{CxwCOwjy@+bUroPs)o
zxl&zH4YmQ<f7#u(XZ?niI*uwKE{mymF*S<>MN-VGI?Q(ljfT(#Yvd5;9HI~jiliQW
zbY|`Qs+|zJphpL_bkI1FpvZf#j#ztnoaZr)dcrPcL|;aDk)X(o?YC71X4X0tb%F4?
zNW3nRAtFJMSzE%-?tYvU0p10(M{p4u+kotaM^&#(XxiurhXp+@Q<uvSJ_13JMGNj;
z%UR)(0H*<q>Yeop+kos9uIH9tyBMnCs5`Jj{6d0l1F~1e)##m$MvvmC3Qkc%{7OiF
zk@SjmrQ9fg!!ZaB3Krc?_1h^g5)@g}R<OY_z+!+jf?akJ@12l{0zr{azI{4p-SCPr
z5F22ND&ks2R3br<0%LP%$ckGx!RR2;MJ!z;P9!K&W;xZ=Q2wYC!UuFIq249nU4fv;
zj@5(q?7vVw0|Eea*+IN_fGz?-k@8K`-~Dpz&_IZFC~uXNtE5Vipvcay;^O<GM<jw>
z;6sJfuaNc^35rx6F4(=e$|n`_7JPG;D(+IFNKj;#eBH-y3~6-as1}TG_A>|D24sJ&
zOq{UWr6U|v1@-{%0cP8P>^(sNyvN>*WrCUwlxqMr1%e`Vcb}|Vv1U{j7#`FtrQW4r
z4}qY_ai_n{9KT!FA9jKIah`DJiBcpea`xy)KAJafpW&zmb}6EMMQ~Vwph)Aq;m>)-
ztu%uBAnOLZ4p!NQUiL4?4gZ<+o;O4~uv5HJhS-K)_N53#Tc*JRIzW_1a*;~gfb4Vi
zoAyPwaF?M-g5$3f#dTs335qnG$~+xwaoi`IqK0raa0-E-NV9y}q@v+B5;*Dw)(q8$
z`rC$HcK&O>e{-SvDx?y8q?Nd|!bb#xB5i>=i;M4!)xashqs`Q-nGO*NirlDtaqWPV
zMODyGV3$s!=p=fPpvb+vQ>k<QvpmF+a!@MQQKxm(M<k3(A0kxbrzQ7ylzn$%2&HDd
zwc|1j-&fw3V_byE{AttEKXI3OHIa)Q8aKxd-A69=Wvsm{1hTG@`eKZ;{5fhqN5e6g
zuy4bTLes{Q>-uBu_m9s1ExQlPC$UL^jiSp~{xo4vlR$I;%b%d^2^xs|W0QydtQ{U(
z5jJ<3+}6}_nYGm>WU_ojSj12)fh<4QZLSvkDYk#uK<NG0fmpFrqW}xdFua;XVV`9A
zTg2@a(PB{PF6?&xclo|SPBK0|Dn1&|CF3V+C+o25$@qE_RZpVvlrp}VXq$-+lSjrM
zpm7K2NIbQSucJ|QG+LUUquO&+hgL>pLRr3$aD_yVJx4C$6V$%nh&>1HMK;Qxsdfm8
zz-LxOgw8=Q1WE^TfqHB#avLN<pCKfIzZLKh`UM>a3@Zx4O)WQys`XR8n?tN7Og=zg
zN0T8k0(k)zVQ5FjZjJr=iJ#ja{^EJ>?$a|JQ@Z=e_@(Sp4snx=e@xiNghS0_`~%89
zfTAweXOx2vS-}-5hVcR!B!DXHP%P9HT@CTYSYRO*%*`YmV*xjyavRFROom|gjOdTS
zjED<t`5ECpBc4(K2=EI7Aola6{kpdnjfb>LHYQszJXyYh7#oO1QsN9Xo}m`B9tJKP
z%8J#nJfkx}Hvywb+Eo+JYGT1tW4VCz5^x$?AJT`y&j7q7uzV-=?W8;+BI`W3lMclB
zc$XIs&7P*PmW?czEH^YZ3(?_npSYn@Sm=x{cc~jX1#uwQCx8pEi#01G#8|`91WF1}
zP*U=LYZnVeFua0BVO3=LcIwtnwHRZ`DY<V}6<>NDlw3*NR}zh+=I_M)ccPKhd`#UR
zQw^$#U=z%2MVjdPx1#giOB`;#|A!EryzjYLZv~qCU0B{Tz|(>UVfl%^6GPEwET3vl
z9g17CP~QSJ5eo(c3(eX0JPE~cW1&fzTglL7v462(i-7gig2sVJp}q|?6iW{a_|SZr
z4n^pQfEX;jm2z9D9y@?sgvGcKMkt(Bz{IH0v}}ifWvcX8u%4#B<~cM13n6C7Cc`l1
zSZH#Vb#xdy2_ppd696<a5IKWJ`Gm_SUif9Xh@bJph@W8x0X*ZUX63xYkGwWI6U-cM
zh!2vICO#&9j5He?IyO?8)e}QK36c`3p2XCXG1B}b8FG@uNb@>ssG~uca4b~(m^wN}
znjfP>j?oyLW8y<Ov7&fPNS2S(MfyrgOg2vr$2dR_!SMu0j|RfZ3k)ls$0$N54|en8
z{4^*fw*h(-LO>5fd4$WeH?Dy0(YUKbSPz0*D7S@rNpT@ye2fbO_aInAxFY*rCO~}L
z5eq1!tAOv(T~IR!{0u<w#PZ$5x0~<?_^tEcZZZ((pWdvUH$J9g!~c<cbFYQezi_zh
zpO6EKb&E|>{pcaO9%7Q}$33dMM@?8iavx5ayu!`!lOxMd(oHg<=V2B-DPEd25nU58
zNg3Y+NQA^o^G2#`q$VlD8)<YSjmP<x@8iqQuU--gx@D>}eGrjZt*Y)OK4?{I<yLo6
zAIwW!hbR;7TIjCgSue{*C$7E@eRdgPml3;8`-HGh>^p4_WqYJfi?=g~lx>vTW?yQA
zI~s;mS==RK?6*(y|D?uvwS)Z&2v=YqiNXaBBN3Aljwamhu-)Oc4tCf~xy|-LFWm7^
z!<~2Ii{D;ZaS=33_DI%C3EV(D8i-!15@)E#8LGz*qTk`ecWoC!OliCc!iW2D{p@nr
zdMU)*-4*`)xbF(5Y=vkUY#JOU)sPHRMwsN*Dq^Z4VVG?!c+b>D!qA;8G;~ur4Z|a|
zP<>4uGz{I2mrl^c#F^|mul4?knqX+bIfM%*?zXohYp?d>dtW|Ny!uT<?Y2cf-j0WW
zT<pHMw|UF>9^&33HE$VzkGkKJns?o)pWfUR^;`Qh)fI2#VBS2+<=N+ea4*LQJ2JoR
z&RhS^8V83O<u%HH5hml05U(S|fDtC+4^pp#)PNCoZ0ET5`@ZD7>QAN)))H{2@C1P2
zbJl;++7-jcyTN|Th<X{(Nu})xQ9mI%skHS_br0MvMOclG25`_^%H`UplJInZ2W^NB
zT%0u4u;rP$0dog>k@c}c!Yt{0`T)HDK)v+M<fX*Q49*;gvqten0}N(;&cF@OO|Wcz
z#=y_uWfPW#=L@W*=tf`&uN7!_e60ZU`E&ku1;P4ofoX`C5B^$sslaWoMK;7etkRKZ
zYEyh30d0i05iD-)10C)=xoJOv7|Kv)=+T7M%B1WfdZ{vXP-O?zV<kd30Q(6KDR{b$
zoa&xJbHjE$JQry7Mg%%6Z_pa-W=5q^EAc|O5H%tsmN#=|HP&2~4|NXZ(BLee=ARaf
zw^f#(;XlJ}Hq$uMU*hEuZ4NOaJcXV?w43ecxkQ;uv=YCBa3w^I5R8Q<cCM7D@h->$
zymc-k9O7RVV#WVD3C7@I`8ML;W<URx7=I=H5<i=2v#Ak*8q05>+70&eja0djY9)Rl
z<qD}9n;zgd$`w&HW-be_xX#-the(ixi@1La4aQp&%U`AbSMBGm)YwY>k&o_&%ck(k
zC0#ZvsAmPW;C1X?!GR-THGZaN8m&F66ZV@<*FN~e9qm#0ykqgzCtmoIg_oH>ZC1)4
zc)lB1<$7J+we)p@_^o#VyJqhKR^68u`5a#U_n3JEFMYiWcn9cRz?8?EB!jh}cL708
z?*bNDyr(f(ta=wPv3eKKb<=gzKAZmH!!-n}eD4Cbf!+nI`v_7Pgao|{SlW6QusWvC
opPC+Jeb@{XUN@~{i1cfa^lPy6D{l@qeY|CH%o15}NMI22U%X-|BLDyZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_batchnorm_backward.hpp.DB836B16A0EC29D6.idx b/.cache/clangd/index/reference_batchnorm_backward.hpp.DB836B16A0EC29D6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0e02d9a8a0c0efb79972ef978861040897ac983c
GIT binary patch
literal 10216
zcmeHMc~n!^*3Ui(D)Dj+BtpRCf`A}%Ab^5eWsu^8qFA*Kb*w0k;H)SXQ3L_YpjMHh
z6%;B8bwI>gi(;)ab=DRIWw82M6+1e;gSk8R_I0hV>#f(-<sY95v-8XOoq3;q_TFb`
zzp$`7&Jsz;<bI20PMJ4BCXq<Y>HoNSvD0#iMDlJsiDXf3Nz~ZcmvpLggYO?-^q}#;
zI;ioTc15xIM4Rd6_2&-H|H0zZbnT6ZTe^RJ5vR;NIjVj5tsVYl7b-q>{yEJvopUPN
zE5E${ZgZ30qF*B_zk1x<JmIGD+AV1DA72b{SM?);nOzHe9Z=-Y>66#ieC%=0E{9fr
z{jBz$WzZ{0kNn=N{Njgh`sC+M_aDX7T(q53r2ICv_vXZLei!e}Tk~DcZ_JTXKdd{~
zv?-hHo$_*#_1K34^qh7`7ul!M@%eod*vaie9=a(?2XD6B!9DH8C3*Z(nRsqo*a&}*
zn5Jg^?}3l1Mi*Ub4%>T9K6Bn(%PEoS2i`YE)mF(X&vkP+;9O}F6CaptVL{&c;OA`D
z&3Ajfdzmbpa@N0}uh;zjmxF$s52k?;(YN9fd~cpRRa|g1KO+9>upYz8QYK~Gv_3Ov
zrR5Rd!1PyS>dD_;t#6yUB5$zscP&RofB1F4jo2Ia&y2m+ecQp`4nx%3hlATqnVuhV
z^h%zWWrXV|i|2WDJh`sy?7dSyZB4sZyq|O<zac+pctoJnz;Wuf1Cny5U*5Ipa&V5R
zPuax5ou*8S{7iOj==f*#XATb^RPs&#-TvdVvUYbJ*LmUng3O~6eF~a?JslMH$ojp(
z?^z80y1-;~QcaIj3z`ysH0>IGEN}VP$1fuzuPuy^2w2=W@2%-G)}L49-B?!;tXVoK
zJ7h)Ak9=a*O)c0Lf4$o!>lf;|<2|Ch9&L<D?|bS!Mb%W5-3^mnCA;po=$%x}*@MnU
zL<~N^^eZ1YS~lcj+qq2C^xaQKaZ^f9n!Lzb5?&Ku5K#TeySu~Nw5<0WvBKmPGp$%B
z8{X|(dEVu$vQfT?pN*QiW$N(WW4D%^xjgH^y>o?ec1v%ai@#Q4^~7qygVGes8##04
zTz#0a`S!ZA9h>47&z$(ai9}-Z>gCAO!P`%<#6v0Jwu0?e@Zg1ns&DJR_RAl81}v!F
z#}XGx*$&$6(1jOBVe<x9M(@`RV2LxO6hdGjMDPMB9R?gSIesUw+>%Hs#j~BK;`Ij>
zcMaLGZoz?=y$aIKLc%3WU6Q3ec!3me`SbVZmPgjG#EQ0%2lP+J3#0^8FML?L;JhnK
zTq$K6Xt#kAFObr^?$G&9pF9a^L*(WXPTN+idHo?VB|zs290bLljhZ-A9m^6Y+QwRN
zSqt5Gfo%jQxyE1HkXy(SFG|@0+AVZN3<4>k*(=<Z^*ue7C2o|G52}2y<^@s)X6V)z
zXkRX6i5aD=2mN{o<pokA3sc4)JCJdcC61J`1I%_nXI>a7Ws+*=i)ThDNPCrpyGYoJ
z#GDsMnK0Ks#Vhz|wt}>!YkQl>ZWAkBAZ1E~U1Y*@vvw?DX-8?GPlG|cKuTQuqVo?{
zZi{I{q_h?H_U_8pA6PuEuTJ~0@#S*`X+w*z5z}i#!waM&C3ye%*rWJMme|lvia=Ea
z8eSk}&DoJHlb3WE#u7cH?1Vl$A(9tJNxr^jewQy-_F+i}N=XNOIt23qDQhJ$cbAqo
zFJp;2rECV9&0xz5q-5{<wa41R!rmQ;DIKn>*46#>2NvfnvzTG0waHTu87)qeD$=A{
zULa-j$-16Z6|2q~7Je_Vd%>JD2&9xGnmw!aTj9wP9UaF<VEqwjd4ZHY7fO%anQ?tT
zOFB|YHh5&y*%$;;%EF?ihmf=-Ea^%qDd3R;y?J4zZ1)>7ZuU^RCjOL?1Nt1QN(O<H
zecxCQKNFrD#u7_P*$DcL(3clT*&o>=(~TKr$`U_H$pnu~2;v1&jutb$o((h&Wr;7P
zq=G(`4%Z-%a;~gs?v>I0sVuRklntQYKsz!Bq<r&n<IU`X1yq$x=<1e0`x3C=1yZiu
z^tOzi@Vta2YT8N`cx2Iz3<4?Ng?D@IcwBXuC7zU$3%0r7#tWo0%u-w*G{Sl-OMED$
z7&;V#4KI*#-?VMd;?FmH#}aQ!Nd~WE2;>D)o~&LTlKa4%8XksHGC-dJ;k-b~%k57K
zcT}JBBczMswb7Jt=5ljZVrl@fQ{IpJnUhGQu}gad9B=Fm40jY1N5N{(HlOP8i_Kfm
z3=@Xy&-T}#v6S<s7|$M>VoHl(BP<1)!73w-N$j=ddEPq<X*Z1!+6dvOt7xZb{RChV
zz)onvV6?G?4(iLnx=Xb&L`GY9rx8YA3o@>Pcvlb~#Ml7HUz>BRZ(cm8kMBqMYC3{r
z#O)aIM80bFF5Q(2t^;X7890<dcPvm>darQa7+v?L8tQ9h?Vj%}_mRuET5zoeH>@G!
zUV!fl@WVBfafzf$B5@MRvq@Ao87`D(ld;)koPdjnZxIpI-$QhJh__H)N_3@v#N$fI
zdqVjoVs(kABY6vK%dQMq+wN=~UEHWnQC)Fm7;c*FG<QUu564D7Oz63qE^L%#lr{b=
z<EGh6(;~9(y}#l~_lxp3e#&@j)7>{**?7CU*d9uXupIv@-Oc6U%ZK8`JO@vH+vn4T
zsZtpi?Go*VKg+lz&?SM77!OQ>2*kQM`$PSGq_=2mBh@2q@n>%`w?hWm-V<cpX`(qz
zw7B!=DbSr5zzz^8xxr~bCG~F1+tX=$^=uOvcN3g$f-}~aarMx(9$W<62;CbYK)_j~
zUl!>vXhH=EsUV?P?%V5x{QBZ|RSG)$VarG1co}MX+7$zfR*B*EgJM5eVlxbP5EKW+
z_z)-#iE$OEs=!9zUjX$5-~`+PsupqiPoVlqj8_xIY9i_{nJALQIGrfc#W<I!a*3$F
zJfg}I<9wpZ7vp23$1xIQT!x*=$}ved<LNe35_Kgpt}jFPi~e~av3r+^k_jE&o8MN(
zT?fzW;Ds}iao>X1x1ba73vhn{9s({R?nOj|OUa~CGFd1uBLQV3P{0-BtqRiH*m3a2
z;C;V$?(ny306L;{>Vw!pR;U4T-Wf#=4CEf4j%hRDG1cHqDU&H>(eJ=>L=RT0i;C9^
z`leEb8?72`kIgdNOt+Z<h~&Jtik+ahO2-n;hKnu@gU4#5<*%@RyfJSWJ!Iv?r<{c0
zBB3iw>kniHDur$gUM<!SJCgLnjEC=k+@`g@c{#Dd`V4o7m>&}3&q>G6NhhKF3)1lm
zF+N5*9uwmeM0<ib{Kh1Bl^0aJM8BDir!U*rXbt4N?~LO?-<i^qfJqYHbp}V;NaXD|
zU2B8;jPA)W)4re2oT5mlE0QKPO_O%O-HD2@aBlvvrwQGLLNF@?CElwH^;OMwf)d@>
zR>e236sG>pgq%^kx7`Y)E#^p-IZ{h(apbd2-i?)szJGb&j_J8Yx!t6B*59tNW9_QX
zM@bXg(4qbD#vWT9Fk_69ql1j$+`G7Y<Ek-qvFt7qcf9==>g>B+Bi%3#V7OYMsU=-8
z_&`Zl*${!Cieu<O7c-a*7G0as?$Kd_vKe9$?0rVoMCZ`D{tHzMmk3sg;DAKB;^15f
zdOYzAcLbb|i1A5Kp9Dui`A&iQlo(%yu2;cVsNV?cMsfM~p!y!{gz|@=dML)v!10-2
zu^8?pxV#jc6oyMBUa6#q(EmE(wN8vbB#s|aKZJJ-87_;sWC>0R!(|hfY%$(R9CwQO
z#l)ppTwY4_rKGpe-ag{Ck95QSXH@NU*|PPg(f_#Voi(zrX<}9XVybUL9ERv{|6=!8
zjU4%<e+&IN-ZtI^f5v?`k-rW)QFoX>IGqd)ztnrd4(G;jmx$XXq8C)Oh1j(SDv7E}
zRniczVzY~m*WFtpt2t*++t+w@#z8Qgm&QwkbQ+xqd275y$VcNNLVAr}gnTu=BIKv>
z6Cr<%zX&CRSu$8)h{90AZnhRI@D^Zb6lj(L78oE>-2$^zu)rXVnk`VKff`R3L&Icc
zI;b&%W@r$j%m6jUHw?9)%JrZ|1Iy6h&ny!xFfOFJ2xeJefw39YMKH?-3yfi@Q3JCa
zIvze6WoX!EwgD_KB4lX1X|@q8FcfXIp4NrHiLIygPT&OVd7C)g7WRwb?huDNVq8ZY
z>cqI7IMj>rUE*+8jPDVLdt%%`92&&9kvKGpaT9T967aI<cf5*@)ZG1(iNwSPy=oGH
zNdy&YjGRwzp!%-))bo5*`ofvC1+gQG6Q=c70<#i~fq<OPePDguIZ6y+Lqs$~8U$&W
z|Dd$D*|$YEwjmXQwAB96RsYI()!mliTsyjIFeO5j-rm!A@~G0=dx?<FUME7{_TD0-
zx7Ul1uf4Ab`Puu4kiWgZ2&I5B1#HA>tV{(Pu^L;aft^^5t<%8{7o$~;tuw$*tj5;s
z!A`8k%1p2ktFbZ*Y{Y7;%%*NKpNeT!V`UB<k4TM`8^A`a#>$OgBUWSELU0nRvF%Q9
z!d+$Ph39gc=<o(?Rb!VsVq8aD>cqI7xYUdBJ>qgtj2nncgBUjwmqsyeA}&n=UiYNp
z%O68*8~)3nE3Nsnxx<1UWc>dX%l&Ux?E2L;>~;U_BmZ%wefUlDqm`9Q_x$5GFP?L)
z{B_C`FFl6SyXgbbaxt8*o39A@x%r8Zzni}ZO|_lsDb$N`it!Ur0(d7t9|5lg^;&Qi
z-(Z>)a1)Mx8rY|SxA>^rr-QfnsM}|NxA>^ruLp1OQP*UGoA{_}vcOGz)HT`QCLHxb
z=voNfgrmD3g7-tD;4K}5kb^K(z=t5@kQi6NkSZ{ygcxo$>9v{+6!^)cXEGTe;0)rL
zK>`GPkNDgpL4s*$AU+LZ+(>*H#kh(1Gzl2P2kOBXa@4q)Id!hb`W)TyJ=M7ksX9aI
zh|HW{#*h7=gXRjo8fu8ThByk&Q4I;KAz_F&aFfhVtex9LTlixRal_Z)J1?zX?oL~1
zA*vQ)D-1qGs!EaCVp0g*FgiFxS{&o5+^uf)%c|yg`^#V_@4=TB_i0YY{b>x_W(<(^
zk6c~<I&SvNl<IJL+n1+ioMpT9{l^d0j*wg0;HuxZ{N<_?<eF~uT2);{=cM)2_Q4I?
z^(cJXxWuBH^yhNoT26GhdL{b6A=O*QZ=?;EfMW@G;VLpT<)%rNI^%NVV*@3Iv0)Z?
z<mCK=177;5`jj?DF{g__GdwyV4NoME`o^fnI3bPUqLtCcO08Ms-~@=mQaS%DVQhrI
zL1|wL6$_1it(<?FKpI}8l(ty0*qGIj^Uo7V!$VDJvlO!&g-#7G6-YbqV95cilqbb+
zeDd_5|7V+));tWSN}M58W(diEBd4y6D(TVE@W#&oJv0-Bo2-~@%*4>Fr((Gnxh(IF
zyVdI9hOdMXh~+e9X?;Y(-TNZ;?hj+0FBu|h&B<zSf-Rnd*6e@hT5u5X2gK$Bq7iT+
z>69qsfUBIBdw1L4zu+IY1jly`nPnE%ZWUG2|Jy!RFDdW0<iv)J^vDlY4Yk1)`RbdR
zGiSbkdh9Q)Bl@GXqeY5EqGMxtY{L~j@31uagSUotpxtCj{WGP3Xe%%DbXqoaux7&>
zuVyukRkIKGzTEUW&u?{_sBzEH6cRcAdWhDPM#fUR6u~y){ZE%_wqmx?e#!Y4MO>=u
zN!QeE=G$ETFFob!PjiPxZ%SG7mpg_pmUI~X*nY+xm-DZdIE=dP7eap7t%f&#V;XcO
zPuq5|N~QCuC8o7RDYyd3QqyFq67ME_@u0+bCSH^dH$2JW#2asirY9Y5SWP#&lDJd~
zmmp?9XpS(54Klu(G2DW#3-ow!=o{4Q*IR52okZGcC>v_b5X$+dTO%#E;tQ|R&W>;V
zYX7{t-yz4-pZ)Z=r+$Cc&pjT#=VYJ%+2;P%GjvVRs_XUtVbNlz%$N{LAHqX-x35?I
zxMBpLkBO;e3GM|tM-n{9=>16WsK>=lii>G|RM)hA82;Jg7m36{_-ZSBwG+N}GJYjo
ewN2)diwEZnXxG`!!CtGe?Ih`JER|@5um1+L<-7m@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_batchnorm_forward.hpp.0421D3C13A4F2108.idx b/.cache/clangd/index/reference_batchnorm_forward.hpp.0421D3C13A4F2108.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9959d7ff3272459e5e2fc6b62e007d228290f12a
GIT binary patch
literal 9380
zcmd^Ec~lffx1U>}0`@cmgEE87HX!>pGZ+^A1lb})jY48HqKO)eMu-VuR8&A&!V<tJ
ziyMnX1-B4c+@ldB#zaFj8jUCl2x8(AjET|YJzwGUo$l}DobUY+&+(5pr<prH>Q~iO
zw{G3Kbt^hDB4V?nL=yT|<kEz=<k^N2iNuWlB_$^=_=rd(QX`3E>E=DJ-g~*-NA_Fg
zIn&!CuAf!!3pX-8v3||q33oF0z5K(40P|hB2|rG1H|jsu=Jlwv@v0L<U%1in>*2#o
zBI|A>-|KgIzbV3aWMS&e)ftszQ!d-y`B&lcu7~M=d?&xT*|fI1Vq>Dqd%>x<PYs0+
zz9@fiMY{UP5N%aZ_~$D;KMT86+T2w>tD#bE9}}zg+nBz&Zi}~d@{AJ8>q}F2O$a#N
z*tOL4>l$t1p@ijWjd%Q;RSjl=zbHHRwi<Q3XIaK6+HG4a9?!2S*CxJO@0N4QFWLUy
zd8NHJ(bMvs!tB`g1>2f5?DuzXNKfyTrn=ad2U&g=@qNhAsu#a7TJvq0O1>gEW9{6F
zUtZ9RTo+|Nt!fohmp(d9vEh`@>hH$q?US}NDaW1I{kEbn`*EQrY1zvOHgTD4>rQNG
zJi6xG(cSVWr&DiSkDdO|HnC<`*@DX(j!w8b;adKb8B<(NeU;(jv!%FqXV}IW<B#VJ
zy6{eJ_NvV4ea+FIM(y!9{nvX(V&{y@x)ENY>Pb5EqrsWv@R(Sm!ZTTCLM+cMIc`%t
zeRZU1(w=Ko$vt8Au@l34*L9C@_B`GbuTGu#L&#}nec-npA=_7m#4k*}(pFb<E%}2d
z^J|-|XW8|ZY)czWUT~>t{KTn7d2`659nICF{jFj?DgLBn)XK|Cx`#g5QeiN+Dm81?
zA0IFwkMb<M9vYwX8t30Lv673Km43SKY=7sUE$e1Y+xA82;KnqY(vXi`9b#v#^|4e{
z$MyI{Sk~W^{<JCgc=c7^jeDA3JCn0u|IUPaa~EVJ2WISZ+x$kF=63ryi_#yzT=nPT
z>2-^)@3^+%?4Bq6_X7Jj&u@LaFX6E4wcQJ9L*L2%uO~TaaR&9F(-s*&kyzS2u1i{$
zFeh}7M8cS?oi3dh`HGAf4U=$Ji0c(Hnimpkt^*(LX5;HO0+oToh%u$OSi5*U{la4F
z{j<MTN;=9|;yzr$m4mVzoOpp0?@7N;&+)tKFqjOYEo_IO+rgX{NYU<?=|6NwfQBVL
zw3YRsTn{e1KuXYC*%r!@&#GDCLMa*GodJQoK+1$QR@pN`wT3Klqm*pWW<v-skTS9U
zSYlc6b&e$pO8EpVJ^_vwNSTyvbFIN-5$%g3rBp&-B~0Q4Ql_|Fo;l@Yt1nBOC}k}u
z*Mc=KkP>6iViJ%v#FZtklu`x3RWOAYdP++{q-W3^`XW<G$p>vdgz*9?accYM&#fj7
zVTm)Pl!CkzEO~*HrDr`uHyIg5v&4Z?GQl+y{CI(sWrfQ}A9*u=I7`eZB^R{0R4+P#
zl;tg68OyJhy0c^?rDTC;76kD^Px<Z-gSjbIor8&qg@ogVaP&jJB*p}&nE(g*D{Wrb
zFl2raOU!8->%e*)jN}Ejv0-gtT~)`$ewIinB@eWDFpd{`O6}gB-D{ccEb*X}jiB5J
z?z})sh4<-QZ&=6FvP4ZO1>jjgThR%mR85=w>rZW|3`=||r4T#|DMcrcvUkz&q~UHc
zRJWrjWesT8Q1$DCp0agzkyYb<7nW!!r3kb|l%f+zX_|ZG;t#jyEMbY3QcA$O1iW~G
zlmnA<o5$W(Q3^{b#XuhiULfU2^VIQWnoZ3zGSpJSZ6)khV#W)koNVxFjj|iNizN(g
zr5Y@%LB$KCd{w}h#b&#OutY{FIiSsfiM&9{`SvUAn$vDgEb*q44WQfr-n>A{)v+JV
zP?Q*S3?>G2xw;K?lRy2!;_JTlxd)D3eWdGZEwHs<#_0r7I-QHnFQy;4G@KaFmv|U?
z$ew;-arc=}>-ayD>saDNi_1Y;PK{b8kaA0Yw|3B^H#%ixFzv%GV!VrRyg<q?5A1KA
zwVygoMhq#XlgK)WH7}6z`|AeNQYS@jV~Hny(I!xCqJz~5J>_`%sWI*CgJfg~ZKZ)2
zHV{i*AmzWSA8g;+`sFA>RPv`DQNo#-n6VOL9grl(%}cs5LLxCpTs}IWr8fu|t^s5X
zVA(h$uWVb!s%&(!V8XlB<@`!wGXAycuZpQlXpJ@qZi6taVX8PE;`4pS!a;_#Ai51=
zu)vV(0Zk8RaS{x<K8WdqR|K3vv>C)#;1>{e0nuQ&+0Kc<z8k6(^o_;Dy_k4mfyJU)
zp9^hn6KKIsaM%g{SfDtpPIE4s-}$#1ii>-V{;JupF*4*%g40QG#u|p)4T!!0F#_%d
z>t3)Ca375BgE0cW1@^Z<CE!)WcNG~Wj3I-#We|4(*AUAZqQEtS8<eHV(~U{5T+YYc
zClF1M;nePG5z@G8L`dtd6`?rWI4_~z0;dI|1hh=OELcF9pvVMg^uP>tnDlW+U&_$6
zr^*2}E=z{`TvaZp(H%3?^QrPcjZTrF&R>-e>KJ}sV5m={k2|_MhB`3%xT9}l=&s~a
z34XX?G29*q*#lEhQ4CiHA$4MW9s<upq)^@gfgNIe2i)#JfKYx9-0le&w|c^G>A<9e
z5*1;>Z}{jrY}r)-xkjfuDKUCPv<uQ0?rqz*JrLO+N}PW?eM~kn<f4a1>)SHq-nM$%
z8cC|4#)J3$zcYF6M&|YMTQ6#MEy<@Va-QovckG`bm*$bCHx`N}r$q~8>EN9XYB7#U
zhgT6}SVT<tU`Vjs4r13qoY8Oiyw@1^g2teoF2SjaskT^OP2M*xBGqHSkgFi{vBm|Y
z?ydNA#G9>)3=O#!P_=*q78r8hfa5nXQowEC+XkYQRzN}vNEnu5R7U49l8w|)iU}W?
z;dJ}{>x9XrWv^S93>ctn4-6L!s19UxVtfEB4uF+VejY5&i*W~7bcpe7klhy7zYDUv
zV*C&+9*X&oz~Ye@7n9M&WE_siw<bp(n{|059ZwNa6cN44VyOCTDu^c%{YTD`%r<>v
zF`ez>;C38DzNZ!3TfqZIV90fWUl;fbcop$jMLY#uK*ko3AbrCjWg!iZ?M?sj)CAh}
zf4y+<?0!1=y@cr{EN+JwCL1tZ5F4alRt%TqoTNe1G2sJeoWYQ)G*y3DMGUpDe4@!G
z5!l~|BUyLjZ_le7JWybkPb{&3;XWc}ABpkDWZ1`KgiyYY3|lA0#bj8q7_TSR>xqMV
z@`Rjui6cM5(4BUBj%<$J#u+YUe9CJ=w{<Z+b{m6vI-e_;6&3<b7uO>V{UfC<W0u(o
zG+k_uGz{t~EsPB_7ihZp9%&fZQ(6j>B3cx>7++5_4*xtMPF6@&+DVK%i78GHdUsks
zlnvF7)P#@xvHmNQ%$@r_9y5q4q!Of+AjeTK^b{ej0y*|$z?RL5gqfH+yJY(AE$xBy
z#kYw378#B&p7!T@b?@N}-)A0F#E#h{uO8I*?msr_nRcz;zG9Fun0Ec&qm!BY1L9+y
z932c9&TXWd8g<0b5ZP`k@xVxcp&_FCF5-vLBg54b`ndeg2YGm!qN*Ryju*|a?qG$@
z#;=>>9M&1NL`KcZdW9<*0>iDeUMWI*!Fexeai!CDfpfhW9|Ol@;3Jf`fkhkG3HTxe
zT!e68!CZoXOJdvwiY{;z_}wt78}z=6p;4CN1~>}!Z$iLLas7U9>=(cPJ~-YN<0s(w
zM2tTqt{;+7*j{{djq9fJ-)1~>7x?b9aN~^TsGW3@Cpk>=LG3PW>AaC*c=n79!+CHX
z_V{mx^W;25$cyt5A#cuGgnT$35mIw%5z=rP5z=y65y}8*2AHGcWa!~pnhECOs7{&%
z=0a4L4bp5dM~^iS<;inEA&&Coxu6h7dGb6^h@(7tJ}A%+3`BX-0x%axdD22K7oxl(
zkQRaYUj8h~&@-xZ4aNNNb-+I;E5SzWACy&KBlw2~;?N*00)}fO4vk{mL>!vLcpq`t
zC&v4U!+tS7KpYN;@j>EnP>c@|heHBJJyJ(D-D0{*nAnAG3;hw9@EHqSf~Y*YQpd65
zL{3(hwLluq3H1xFTfFXstufp@`Mi+=S{AbG6+|X{Rzu$ee-@*Mvw>^|9`;tupX$Bi
z$l30{xuR@x3tkQWsel%6UR*H7Y#;9O`nI`nNdo=1I7Z8lGRVAB^u&Jz!?_G|;qa(J
zZCvFcawVQBPZ9D`d5MsZ%14CMDzyk{R2mV|s<a}M1@bJgLJcuA4<^qBE3wrobHGk)
zwaQ$uL&XoAO_X_HC$?H;KG=z^R$c&BVyl%Gf|b~6<wam6wp#fbibYmyTM15LtF^5H
zC$ZJKHV_|l-2+zZ+9<|N#I;F`_Y>FsVtjzO9uVV$#Py&UA0n=Y1dO>WYH;Cf_;A5m
zbZIQKmXGcqJbo|Kw&xih0}A@;-;WvhbJ73n-R9$SeZQ<ay!_K=zT1>^x_uFRbn8!a
zU3L>iH*v&hsN~_BGgq5(7wCj<BZ_UrQ3xWok-%*v0&%I$T<MA2_ixe`{yj-kdTM;l
z_t~i)w1tzz;v}&ZY_)?}bP!uSzu_99#1J-w6KqSD|3ur`z$ZQ_?IrdlQJAXBfg){l
zxqH`&=EZ)`JdDGoOy3_bi+2>dsmrY5`zwAv{#^t2b<*GTflHs(Pmm>u>g&?3Sbt0W
zONLQXvb&z1Ts(V;H_k|PYJSma)-U{-+rZY)cI%Ml$D3%!{_{lD{xW{ctc)F(>Cp0t
zTR!o@U}%pvaB^$;n`N|x?clf_yz%VL(2O2^^hb4cZ<M%LwpbrynDBXJ)B$=$N=ud{
zi)gysGSV=&OldE&FN%%|x-_$%_H#;;W#0Xr&pk=9?}hsH%g^2(Pp7t<$h(E32WG(O
zKqA<PVS;JtlHoWITyk_T)}_Ev%}2hvFn!PHUwfYW2#+Op-;Mn%Wim}#wSck(^yl4y
zB&hXCutke$c1~6MHAZcF=EmGoJNaE{gi$6{<$pHj)|CB`DJM$Hs4-8rm~4eQ$7=*y
ze<ib0|8EoiDgmeB<hiroeDl+TFP^!V!GME?gTZXDC^tCIV_pPy(uBXYz}|f8usmn=
z*cekfmR>Tdmjs~CXdC0SBKl>n^tl_x+1}<whl0-Ee4322tRM<B1iWye`i^3wL|##M
z_ktt5lyy-tc(K(FRNv1kzj*1v=IHtB^FDm$Q;2;YWb*W42t!~R)Y6+CEM(}NkfIjs
zP!9}OL)>eK7T5TI`P!WnjtR)4YhQPnglivdf)<F|@APjAeYjfz&wb|cHj7Sh8j~i9
z9CbHZsE;0>K@WzQTIN$)Ju$8)av@r{LX58vIYtY;*TM?3T6(gnB^43ZB4HE4tOiwt
z?m`R|kBfm8=&r%A;Qr==$GktYcYN+T!izXM_zCO;(IU`Y#-WaW&yGCc_;vkn|2UC^
zx}_S?UFo4l{?NYPGXB#qga1$e{b)?shvPCYb^R~znHU#8JB;4@h3y{NWpOX<H9&L5
z43z}Og6BjDz8Yf)2^w}%;@qT#19ynZfiK;M+Y#vM&s=2sjBIV}?M7HxD<yVVCXv_+
WKdpqH*1}I4;is+elip3@=YIg9X|CD;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_batchnorm_infer.hpp.B110A1DBC8883C05.idx b/.cache/clangd/index/reference_batchnorm_infer.hpp.B110A1DBC8883C05.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0f0823870227d8473b4ad01ee59a1d20fc1f5662
GIT binary patch
literal 7220
zcmcIpcT`l@)}MVT!g#q13}u)*)S(D7zzju^BGran5~J9_1d}`kBUpie3Nv;rL<N-~
zL}MfN@|5U{AhstWPr>>UF^XbIEGPkdFYhojn|qV2^||=tV=eEQU(WBIv-du`-IK9H
z2Mton1cKO4hAv1=nUicR5D2Wuzl=HQpC?iRfn_g&V8PZ?A55NcM`hD6>SAnL<G5ir
z=5F!UY%CZvzH;5FLlMi__C~vP9<CQZAKsF?|Eu=ck%^|hbj;SQf`VWAe{}Fd%a7GT
zuUAK#*pAGZx@76pvdp`s&mv?NXG^VKCc9ke?>y$bQSc*|*>O3t3%w!jYzcigc*#kV
zu<}TcDW~HrMz?1*y;yqdKxtrE`lP_f;>C$++x#QuBz$ClZ{fmkQiH3%Z(iWN=X_xL
zNj59@(3X%h>kpWRUMjx#`B~#1<2N0XncOMeUaWm`E9BD$ckSld6wZ45^XZIP8lMXr
z5?eeH(;qe;Ro~9M+_~Y#fwTL5+pT$(Jo#+mh+Y~kfFi2>>Wa%R7+;@)Q>zuWqJX`&
z6LuAazP(a^@Yt%lzH_T)p0jJ&*io?O#lxsn|CD_H!<WBFza5Zzb%}jq%B|jC>@NAe
zS9{Fb8+%TUIX&*Df`|+CA+ATK8JF1G<_E3}jK6fJ_VX92!Z26$H(zZ$8Iu@#;mmT|
z;0`BWBa=h3tH$h8eB-FB62%&6)(dM=O5_iRoqS$s_wBd+_Zgpmd&aJQw{T?fvT>Hf
zKAv)C{0fc9o$aQf7v|+9hizPNAn5Av=SM!**qknVd42roaZOQfegC#|_0D74O&;}{
zGNa|H^YAsh8g5^^eJlCS{K583u|~ZbV_ut{GMlH2d873mcj45r|CrAH(Co|f;U|ng
zyxQ`zsk6E;di<jwe(^2(LRzpydBpZ~#gMJNtuEGxrZk-XBzM-{Bu|r+wiWELf~gkr
zW0hwnwS{dsHe~<A!^wd+{%tv<(eLRxZPOQp+JB6R``W5+Sd~}Pq2li{7N$<iChkT>
zOk1E>P&%KXl(qtPB{;4GKf@p}Bc4e9n7&7HCMevGGn56PtN_^xkQ)Z1xGHmDfBAl`
zDP?3MVEgp$Bi21|+AAS!WUu)7yBW%!tgr!WHh|19AVqb#dG6Bx-aW)neuSchKrO@>
zhHlF1n<K{W3x32<aza@Tq3dC!VL(dYdF68Y+#w<=Lnw<Oa4}g$KOiNfHoGx(hJP_b
zNeE>%IIjj@!+@0j3cB-rT>NH+q6uXY1TKPT!+?~~7cCENt&;^XlpCRBdS~|2J#ad_
zuYF~S<Jt@n)!R<MZU$yESQ!SS#JfK!E6z2U#!!}Ig`B{g5xNIXk5#2ryYGG`VJL4h
zogJ7RqkG`=xQb5GtOrT=Z7FkNa8E~1rS5^#iH?CkS}Q&bW+->^c9w5er0#*!e-l<b
z3>|+@V@e6d0#@2f>ZE%J%qbw}4J@R6d34*huN(%8s9t3GH(>k*><t61v2cZJcCC8A
zK)s61z%&yT^#f8es{=fCFI_fYM46IRo`d;wa_ss6DfzWVzsmcqn#53jh@gcKRS07Y
z15yfG&&3^yh~2<YUWBsHe_^QZfz#Wdrc9j?ai5_?WI8u6H%9ls>2C|#htt<>pD>hy
zOlu%eLp(x1Am!+o$@A50-!g{sB$TxvUQ3kK4@jwe_jKE=)wOaFWlS_Fp+qH=qhUbG
zxwUt_q-Gs^?I<I%ijSF(NcX_$`ddzsh0mfdFqAKu&IjLo;#2woDGhJKmLFR&beo7W
zBbyXa-bGZHVL-~iL&pwkRyLF~lml6%0GtazZ5WX9Q|+ZG$A>$%n^GoSt3dA}ur>@x
z`PqHqrOQ#RXBo<etda-5c|;cdfRyJ6_Xe7kJryyO2cawjhh^Ys7?AS%d3N|G<J;yk
zlog>Y2gl{$VHl9|-}2UPx7VBtrYKiyT`Cc<mO@KLV6F%0Dbq7FHstrT%#>Jh)>WX{
z3J9x!$ja%#4;tUdu48hdXf}q4aYhph4G9Xf4kjo<i)Ny2IhsCKA<a^9Jw|oy#~DV(
zWSbEcFcP;hW@{j@27)+T2kJTqz$mubGa@Rec$<XC*9eM6@WTn4S-VsXbw0z%#8!}R
zg%F&ORH}2mHhtFgmp3Fg4w}8&c|3r~Pz^!V5X|ja13opN;BX7rw}1r0%$Zr-f4SM^
zCZgjoc32<Ok!F*fla+{2Qxl;s{bHEYI#0w&nj|90W^h4SX|@FvEf9!1(X5to(NcXd
z2e_V0|LobKa4lub4znBPfO{FU$<ky;L~fBsPP_`aD(tadSQ?04^XPdtxSr5Z9#Ck%
zQJ)MB*gJnE7_&*<NeZN?zBm%sH^Ar)vGXX&sBQ<rVnP<3C7R`mY@wl`;G!x+K_RpZ
zI>Uvd>1zwp)YNRtHB@;x+4~;I?tvp(PTl<VJ~pXl7O_kP1Xb`4vkqeGAP$LGmdHi}
zm;v2l3Js+RHwub4E4;Zd(Z-BsOF>i$cIXT=TLz*sKHd(Z?R>lsMEm&o0EiCo@d>aw
z0a9+gO0cQq<3<Q;glLpEXy-EVB+czFiM;h7sRu8dpos;X?tw4j5RYksWXq{@i3O@b
zSq;9pgE3nN1L`2Mn;5ex=5U9r<$Fzq<(X_9%f!mKGtpOSG!>R;LW^J`cr<;rMp~%T
z%S$s%lV6gfE2rpkioryV_1lPM$J&f_!)<6b#U~{Mkx*CDaV=B@goa9vnKouyuG`B^
zUndYJtc8GD7=+A0-)nxE`D}WrY1f2RE!g7(&0YbkD|~znY_9>!%{PE;10OenZ6hDw
z2FKeVFAh5~VRDglC-xFVn=mGfmmHFmNrO0Tx|$5s2Kx*0%0$t`ZkMFlIYZ`r#J$qh
zc;G7u<^xZqricoNo+Xrd31x}K{9vTD*W-2JCS<MkV6h&=7+5qZ<rak?#wy?CPFD4t
z-s6%tOsFWk6G~RxL5X)zcDN$8H&UUlt>F@xpTaK`rK4GYg+C9e6e=E4E7UwR)p@EP
z_uj{z9|v<NL!A-Dp-f4p7kUTHX34S?*i6xEj$4izZB4VeZn<jgK4{V(x-9}V`Xx<5
z+ifwZG4g3rUuAiqz|u>T+Aqro1-1oUmg}<~`f-*kg_u$pgYndr0b<JdcnA3Gfd1V4
ze(>4P#|OdZARo6wKs!WZQm5Gt2<YJBP6+7a<9877j>F?Si`31Pc|kp{ooGB`noY1t
z@Z{{ABu?_+P)1C~2Z)5aUKPD=Qur2exyj?S_hydCE;Uy^Fv>F}&h*DBJ^6V5nUg%-
z<wS%&9(~lfFipBumu66Mb?-SSo<l$ET4}ZwWUbI$2hj;h{C9QC*xSrlwe;+&)E5`Y
zwtsA1W)@EKiff9WPojUG{X98p*G03b02l7$4ukt)P;vMOxF6x;GvI!Pk1vAvMF_@R
zF|%r?_m+}36W^PH)5Z@p&#Q{xLj)ZxAFM)o&?AUi`aU1^Tu|NgAk(<M)`_$aGN0}u
zO0x=?f`^ncB@g+^e0j)E=Ep<+GJhUY$y7X~mZ^DYp~XULZ2M_8(<0NF7m*fO*636;
zn{AP8jj4>pr#MH#i%)T`gdd;cMIhnFr+6_)`0;6x2iE-fw8#f*E<QD2p#f|BoFGjS
zSZKi-JB_aRbyyEhT>KV-Lm@bEnePqA-*C=Evu{EEmXF&&-p0qjf&4c<ZU=cgA9sMf
zgO58w-pR-BK>m)y=;Y*-_1|b%ywKapg#OD6k3ajfasIo%umAZk;e;xY(mRKq)7=3I
z4WDfoUQ74-U8yRVpGw~6w`MEjjq;zb`8|ZCSx;L}876w-V{Sg)=##|9+!Q>dbW`$>
zpPL^K`Mde^kjhQPLuxlQ4`qq7q&&|NXG?jW<B;RR^Bjj<7ry5>ECLt4=Qu0|7ry6+
z^FYe?9C1EK`JN-zfRyh!Vl7Ddp5we8Jo%pETnL_MW}1Yh_Zv`QoOXGR_gg-01MfCI
zZU^snKJEbT4nFP#?@m5`2j1^Eyz147Grz_<Km1dtz54N?*$L54*Ze<cvvvP^{9<us
z*1q><GfWSp=Lls&yT3|kHq|RN0R2d)yD31A+|cCnL)7W*Pl!jBQj${29lMPUZAp`t
zSjrlRM?V6|BXGxP>B6CpU=ZSsQx6<`tvuD$=}A@6I1H7I7l-V;rp-~3_qCLbmU8CK
zYBOcCnR3Q2`>iF4``J}<`}MerqBD{m!kKVh%yr#&pnhdDhZ)C@(cafJyUKFP5A`e0
zS-Ivi({JE=jk-Pdj$=$!$Cmdhg3e5cX`;05%q=wBte`}@UJNOllDGRdu~#kl)Pf3&
z)^~xSqiaf%Hj(#>z`Y3kF&z<1$%-g1v>4va5aK{)pfz`}`uiE&=;+hTqxLIa?dkF7
z=dr$rin8jT4IwrwrNpIN2a2{Mc~O7egNOb7g@y^=g}r{*<E^T^a@75f#P4|TDOB$o
zon|q}EdRZyP?LFlXy)0Cn}~Ub+6<MVad0G}cU?oFjn4n8{-6JR()vS>o$rb~YuS<k
z36^A=aw@2t3PIb{g?na?9VOe?V^d&mBpQAy`qUi-78)+ZP~!ULs#zx^uRkJ(a}VtA
zfdt(Ly@P0|zb3=Oxv9A!F`X-I{8RF}*}8S?XGJTP_E>eiZzF=nF>$;b0{zt+3X0xK
z5qWeMZYU4AYoS~=bMD*}QWx7+A%PxS466$fWV&cN&tm#(I~3%h%A~cIrTt1mYX$RG
z5ObNRh%zst#AvzZpW+HM)ep7AQ1#$l&qe0rhf@ytPTVz^9QGgES-cS>OGh%1yhzkv
z4x$9Fwx0Nh{{dI`9&3o#iDdmKCd!WU3H{9?uK!j$^zZJM4!t2!^nc&Ew1>wnjn2P$
z|4+fukrun=r&iIZzvw4OPnnq<M=m4dDth0y`8D?=z^sqo4+4|{%YXon8Pk~n9U&wA
y<BWu^OHTH2+vqHf1r4T7&W<iJslcV{zv}<U|M$Z`NBkiWICDRpxSvw)=l=kk#UlRz

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_cgemm.hpp.9B0EE32433381722.idx b/.cache/clangd/index/reference_cgemm.hpp.9B0EE32433381722.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3cbf35ab101ae558208f5573ad7840183ab1047a
GIT binary patch
literal 3734
zcmd5;iCa@w7SDaMhQ}ibkpw~>1R@xTvWbL%q7Y(4K?JMdR-^=BSR58dDuT2Ia0MGj
zRO(U$Dn&uDxG`!K#TmB>Wvmso)LNZ_xImqPTW4+phs5C@7``v(zQZ~9p5>n3IX8Ni
zQdvqd7@>)?a?{m0%S{*z1{?pfbFxw>g28AvV=!{}9N!+1`p}D2{O_2Y<h<~Oliy@|
ziME}(uMNG^R<2$dvcT=cQK{vuNK5s9-dx=%yelZ-gj|SEdE6d&M9HaGa<(h1Aa!~p
z@uGL%$G<xyt(!U6lYh_c@!3nar?3Y&UGiuf4m0YH2-%?eypLy7=ZvnvWA2qsdpYad
zxDlo4WiJ{ZMoQ@o)~|be)^2mk@oy3BR(+8W@lGM@A;SH)+qUM?Vf0@+%6r@#!j7&;
zF`e124H&41Tl(c4Zr8n^7QED4yr!6&_GC#rJF=;~LsD_4=`jCgcHh6cMJ)}Jm5a-E
zM`%u7&+u+C?VBMua=g{o$K894$%n%Gn@@(+Bv($6n{T#zraoeFD0E}dn;X@!t<7E`
z&l1b87pK2EAKd6j#rvgb<PU4?>Js)Zx3=S!46-*7Q+G1GJ9|G6I3$_g_-cG?=az|H
zv*vwQrDZMjJMQ#c@?%wQY**Bq%n0khyj93LzFoJicV*9<Q{DG=U+Xzn{6VTn<f2Jz
zW@rOWBt)idXkQsGiJ6nJ_s^P=^&J@l1uYXywHB@F(?m;l|GEfmh4yf~tL&e_6<+2a
z`g}C)(4zdb>b$3mf6-oz7T;R@t2{a6M7LyW|E2W@r`+gh`)#hzyfvbtE^W`QRHw7y
zow*6V+N^>n7yq{Ur<KnyU;5M2S>0)ea;NSqDbwt+2!C*Z$5WkJROwUn2Z~EBQDlFd
zp70P?1@A`7>+rm?UKU|)&!9IU%T35}cwt!J>cM~ZTCa+?rc-5Rgc(**SSdJq0%FHL
z+j(cq&R(Vn8wUn06Un4{0%F%I--nfmFP^s}#$d6Fg$qkhKs^43NAk|UN3Wp>E*4Lx
zr+erLh$pUEQe<=C+<c1QV6j{zm+A?K{nlKlren9JQp8v+4iW^p=?REsIgj5BmWF3i
zgdG+K3Ig5q1jNDdKPq-@Z7b&xOdMDxVpSr+@B%9H>#t}Yz7u(H2n(#DK@J+^H@tw#
zlCJu(-z<7YS`lLe3|dJk?eql1Y1a;<i(ASADT2a2OrxiH=n05d&5kg?#O|dzge5kx
z1+li^{0}XlQn)^PlI4mziXvF}1b>mgR8K&>!!|&3$-jF7i?G5cR3Yyw6gs?s%E6+X
z;%;k&hDDfSm0HBAMZ)0)R2rg0+h>fw++jwLd<HEv7mn2vh9!Y;BSADe64kZ#-K$1_
z7GaIU%%}3%dII}TvznKR8af-yh%wlNvz4>0o`Cr7+qQH`*_CRFVB;8q>0l2%0rCCP
zuj#d_UsewtuA)?IT1P<qj34r<Omwc9B5d*Q08xNcPeA;~ftO!aw_f%m2p8T^$<h>)
zq8OG#3?oaOlzjphvsL@CD}x=)r36W5l9_H`l%!XYt425g1Y!=u05ZeB-DvUl;7K23
zLKh-lA>u<u03RkyJuX0>#HHk3i6%jnk#r4mu0if_Upih=gS>zT9AWRIl++$@!_=b*
z((6Qh60p(Y`U6DLk%Gwa5Q3;VE4jBYs2KYeX%}e^|8@y(I^8d8V~%ne4#z?25DBLP
z7_ke!l&=7^n0iQp5<n6wO`;|l`Kd!EC{3E1IMe6Q>dn~l9QGVra0u{<_j*xYMtgP8
zg}s1UtmH{~8Y$@j3`%mk?`o>q@;+>!9+CBkf*OZ(<Isi>4fhf9#X6V1o%>?&u{pvV
zZ*YmEbA`DhobpW3v$Wo%{bNWvoEI)MD9yK@?=+_=#q>MH=VFphv`-x2Vl2bXrmI<{
z_)2P#Wi7G+PoQ^TPlg%?c(P#mj<~~>cjb6TBeHEocCZ6l$rz@-g+O7Ew2H5CHuzGA
zJPMI7l;6^m0clIz#ID$7=W)*7kRX!Icg~lB5{Zl1V-xa+>Q<MUjw|G3hhP_k?m}Pi
zENy#&yW&CkY9>i9kSy>s1iVnP(1?qb#2RsNk~kx7nRJ=V;A)a2$w;<BvciZ<mLwZ-
zDUuW;E<>6zatd@%d~x2Cw^3E<nkj?}eH15qsCz<4&^_^P56Z)+nskj6c5mrFec^ei
z)1&_>*R8)E%Z!=vr0oAH9ndduq(jXfBHi&|!ujdvs(bLA)**f!5<^+-7>Hl`sg2_;
zoPq*=f!I(J1wjQW;L5M44t5k5Nb&9(#H~S2V5c%|wn_NB;-7JD8W7oFKw5R^U13Pe
z>(58+0K3@3p;-eS*1k|xE?fUqrw)Q+a|DJm$sL<J!qxlAqgwaI@5D+9vBJ{;f+{4g
zLZ09joJlbCg9>H<Nh|G@?f@hR4tt`TSa8<yCu@%LHucxxklt_QC#D?RmDk>@!h$+v
zTW3Hf6bkO~y4eQC$@khPwk=V5gKd&7M>ge%2bFo|AGhM_1Ao)i*!xB=)<$JmD=jpm
z^7yT5-iA-7#n@m$1L8IyJ|y;LklWhmxsK(dj{2eKHs5TO!}nNV+;(C5;*|POK1K-A
z4PpolZUg5^H+vyh4`eq(2M+VmC|i%bOkbMY@HS)bst~&h3E*a)Y;X8|ugA^(9Go4V
zix8#=NxSk~jW{t+Y{ZS@jWgoh*ls+-K6kde5$8|&a{<&OEu&;yn4BaXKm~B&u|U#t
zO3sC8Mbd#5ffS4uk`A&6qF@@4xMjF*K@zwrlHQ73w;J=M$hFj%FGH?n#{4!U+=g5Y
z`#(d%&kQ_xgwsEfnrPI5bT1i5fB&A--*_GsiAUu*OP#hn3BP0%hpm3*K3}(R<Rvyc
z&*|JJH)doaXzDNt7;q+Zc?Q&L5?@`B6gE(TpppP=Bai&-tdFv1Vrv9_bn>8g+Ed@J
cO`SzfjzWfL_%VeSq2bff@abgubT)kc7hN(B4gdfE

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_column_to_image.hpp.B1B86724AE64837E.idx b/.cache/clangd/index/reference_column_to_image.hpp.B1B86724AE64837E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1335c220bf6c822cc6c9449cf13519908e71378d
GIT binary patch
literal 9180
zcmeI2X;f547J$3T1_2*U)3ik^5NK#*4<N{9vo8V$WpNY<E~qeS6c=zAP1G0_BO^pq
zR0IbVHxwiwCMp=k1Qbw%VvrDtxDuDRMvX~yipEP-oL@O-JSRWo9O}J~Td%6R?)s{U
zjR*^Cwi1bgCq>Mko}8W}7KucL^fxVi_LQ|)B<kN=B$}UJIb!IHHW%TE_g70ZN;8hH
z{&ud7m42CZfJtG)s)O-Qmla-2cl<h69C3!o|G2l^eAW+j3ns@LI~cO4xh}&_yVtYI
zGFjZ0U(20}6@g?x9bRudcB<B%gVNhoj`I0Yh4EKrx-8i=a7XX#dc!S+nI-O94sM!x
zKTUEksraX~ksqa|>-bOpW7s{<1vXzP4aWw&>v8z;3G|`eA8wO&DI0?~E^}G0%x_v~
zWiE}Dwd*+OZ5*CFE7^ADniExyWOCrH%L!4Ic1}Yg4t4x|a$kLCRm#*SGmoy`>lPC?
zdtSBc#hqh5jXvdT6}hcqs?n{@THCAJPS}nOT{A-W#r31+x5wx#{9?!&n>cH|J9Q(1
z>wI@-?g;a{^VwL@ysP6Mn~zeeukYA@*Yxo^-=DX9_p55v<Rd57G^95>zj!8=BrSf~
zZqTpb*esnsC68u}$ncf;eV;tKGGXSbqn2Mi99f*~68AxMkV)eY|CsXcgYT?8GyBHT
zRfz+P#wa$mI4y{|VD8(xm^j=K9Fm4Mw~d>;$G)w!!)8v}xZ8(Kk0<Y0TAUu`HhcQF
z-woSld;ZLjri#nw^PSZPqOL8kYfImIKB44a=SahYu(>0q9G;>MPtG!&=y+bMyz$cz
zuR|$eF9REv%YLnmsGCrHaPBYJ1BNa4UTc4*K|H!@f`Qe<p>gHnmVvc%JNk6Y+}Q6?
zZF|DLf<F82Z)jNFckcP8NB+?|z53Vf;S*L(>g^f4cH!yGTYvTFSU2{@xf$^ngy|tq
z)l;r+I(>0qmVx?`I^)IV`zuG6_&O9^ay>fud5D44d$P5mF;z=zs`Iw?H~8UwW%b_p
z&YFjFS3NSlaPFVYhhm%je)(Z#(VBgUDT8C4rwCtMy;<cQCi-c9;RpNPH(&Fz*Qwug
zMvBt@HGRTuYMX{JHKk!QCPzuJmVt=0V`)1!dvy`%P#dN{%Kpxl+BfqQ5^O3J5kE6O
zdrbr40qrA}wTQBEq*&ijL=q*5hMES%&PQSg8E*VqTY`<~(`qxdy`};2pmo_ZaaCiE
zHrA%|$M=bsX&Mj*4jZ+rRuW?-#l5NcIMzLmO<!H0%aFCnkv7%wt7TY+cFD#4bMc^8
z7w8iAUDV==!Ly!8usL0Wx23nOrUCJ|;>_p;Y0~);tfb-~vmkd(1LE;3QnwDD+<$ua
z8g^pcopcS|7w9q(i^q*TuwzsoET(&qAW0ya2E@yIo5Y{0HXbR%dUS~`SlWV3UR|I|
zrpuXe)5SuT1e?({_?!7VXc`b_>4pw9PP6(%8w(N<QTA4vX&RAkFN{PYL{MI4!^Ga=
z!c+;ipwIi5`KUAvJio2jq*qbZ6R|cH(-o-tsEjlXh|2>M{rf$5kkYNgY%IyfhNSxf
zT`Dh>yY85Jwn&C`=?Jx0T1$7K`vP5#<d@jz_-yZx;y$!XD|T(g!LKgR<<#oAlOnTT
zj+0;$y39bcKzB_8;`-{v2MWtNa|sqxv5k(6RMUX?bpMv^LgPKUNir(-HS=}UG$1~c
zAdda=h0}Q{eus)f^+T;S4TwdvlT*@?%tWF-P35Paoj;t_OCT$ebS1I}6BfuZBt6D+
z29{>v9=U;{ImqRF`v<yW5<Mh39Ok2x5dA)*Si4wzMyW2Tt`jrTe@6Rv?h%t*<eiIr
zcn;1*A&i?)P!kH~IkpMK@toL%#sGu&BE45W3wEMt`CRAsU0)0&Vv_2R>InVCByC7q
zIHUBi^dXFLk#jC`;n_VG4Pu;+Li14=;{p^?fI=B>M-kgmB;z6!UW6hTHzBts<j%8K
z6Y}QSw+Z<%Zb6YPXei^0DEuNn-X#=u2@PZNAJKb1qA12Ukmn8L#kdoBbRtj2IoLP{
zE5P=}qy$S!umQNcn3Q2j8P6xM)d{SEkYq0T*!xqR#ttlyB$FgtW<#0SI1{UYA$n3{
z8z>oQ0v4{PiJ{<@0*Q`@9tk$1JUn~yoh9lOSWJ>gQa>0;Ofs=76ZZwS^gnRuq1U&1
zZ*}~}bPQWf{i`dnL9%k1^!s2l{85dTn2dED>ju5VWSq;mfs87!Z3VVx_H+Q-9^knW
z+g37$L`e<BMdHF4N>|Qgh7zlYwFX0Ry|Ja#Wc)WXkeI}{#JIwrVVPn*O)iC)AkdU!
z(}LfHY$T92Y|(~oAZ&NBT^n`*_OhNJN-`KXr)zYZ7+V<KD`-Pd#dD>>Ym@rC^Pv6%
z?W70v&sVW<6-yvCfxpl_?9Sh?o}j|TmJc>;iPMDPkd-JNcCIT7haBTM1BYbb9(l;x
z$DJ{=joOhwHe<KV*c(cKQ2AZMtwy~TP(dD!&%<Lt5LUJP=V^CR{-Z091>Ot%*+hF$
z;9fL@aVZKc<@qj(xQm7}`F#{|pXWy?;t|gkSY3gGK;;5CfYk?huEgp}#;`oPHQjX!
zY)zbQX!Z1DX%_t;?O3-R8^C5lB}0SGyOMY1EShu`HDDwohRJL?*Qm9m0<j}@rqD?s
z_Qake2jakyBXQ)&i8yhj)>j*Wc?!f=-`5B-u|WLv{fr>S3B+ID-v~+>fduLY8bQVv
zNRWPz5tLd2If*S#Qu}{ZRS4u1wmijiJ+`dpxdB@?@Z5+k8ySyGT0eT<);1ryVK=eS
zO-x|J;>#<)_1zeGPh%2pdDs*7JpQ+ht~ZW+;{LbE<%#OeM%%PHQ(yC|F<*F3OB|rG
zqF$o1R=I+X1zMh2WMfCjUIKM4hvV22ylu?YW$~3MGjgb*lwf%YwqQ0{hUI09K_ggC
zQ}nX&hf9asfX~Cj)KI$%UoccCe0%9>3SNjj0!dUPS}?cC!45gt4c3}0ubltRoaZ~~
zN6N#ddDt3s(#3&!IFj*ZY`U4rle4s}m&nqBs4|t#%E7RRDONRd3w_BhnLyG!(t;qt
z3nbkmoufG(b2!RJ?)fNy^({c|1w3y-?pt`?j=Z+BLQNn=$g7Cw9ms13&n?KSh3AXN
z>mtvs$g7p-OGteQ^^|4;`4Opq<hcW>J9xg10<NR?*!pjxfSWx3gaUqIJZ<d+n~;vk
zc?L89+AQeFSOWRT@uQyXDUd}Ti?}UXf*hA1FQ&!4$ayas%(xUem-1YW+{^hbt3ZAg
ztR57|LF9fA1+eGuBG0=hnDKq&d7tM;$nz1;Pf+L+6vIC9GZgj=#WMa4seeP^Fy6G9
zU3QxbetG}3^!4H5)JXeP1O2ECMVm$USooqzErXkHE}KpLr4$b<#qnVB0{I*d`<yXk
zPwEQYFLGdo5ca55yGh)7qNkqbJ&w6mOy3wQaI4J7Ud!%gzx48?fm&&2%LQtsJx2~o
z2aX(-jvP5Doj6j<)h1k^miwA;fm-fo!Ubx%zX`0E25Nbr2^Xm4K_(Cu1QH>SFo6XN
zB&vUulF=mDq@MCrAjz_1j#6YP98D!t%|J5(nWmU#&h(b5NabjTVg^UEOlMiKp7Si`
z*)YmNBnxq)W+Rf#a}HAEAa1O6$aEdzM$1K}xjb(`rW<(Pge*28ZoWdKDnt%!{$0p+
z7jkA?i%2a$e;p!qJfB45B+m^<(ZH|oG%`KS&wmzKoaN_lLKaOtpF<Yscy30jW`2F&
zBirx!^<6=ZSCAW9?=9qVi@gsCWH*-W#{HP#@4>P?JQrhGG0*$3WFPL!`j=p33ASat
zA4~T0@-nO}<K-u@(@E^kmjF(uc&^7z^*lFVrv{!Iu~Q@C)t%+D;sS4H(;Dxu1hu~!
zwEkKUg5W^2YuBO8i*1j07JXAYNFWLQ5}YA<t~u9v|KlSIOJ9pr>pSd^1spB9MI+T^
zthkI-P}8h`I)2Pj1JesMR(*yQpJ5dw)-Lw>42J>do(kV}e#Ih3Izk(kw_$TA?cubD
z#wurtb6*HNdYUsM7OSHBMw^1O=NC9!`LJe==ijDY+B$8hI5K9%6}oHJuyBpNaBN$Y
zncXPy41H}o;bUp*X^y@?GlJ%X34~NJt_j|t7=czhRy(l+GwBo9^91&XB1<4=u)!H@
z3?^I@+-4S3^L+FF?Dy{b;akUkvgdoc-z$-NB?@Ap@EB4b<2eJXGjNZ*#LXxAn}YEh
zsH)>t@y@KiYQ?Hn><sz@h0~<eeQF0CVP#a%eO3n_yYlveO5;@zi{Bm|^prnh^8Y;`
zI_OcpabeS)aB7E_vC(DrP6Aq^O6fi%1*KG-yfC_C@T2Q*mGZ`vQe(5o)$>5)j@N?2
ziQUnu`e8cD{>O9c&NVeO+<JWMZxi;QOrxR1UE*%dLU-3;Ityb_4pBi6y6a8mU(uZ}
z(W1vy{K=A-xJq2jU;y!-jwi)`s$s={nhKtv=G%2b4|77LNf*YCtEV}lhfcJBjEYKx
zg`KflxN@mqoB`eSYgm2_D<H!`2O9P~1P>L=sHZv4hq2D82YeVi+%)&Cx4bcLX}VD}
z<FJ48O=>w&K~ce=7D|axV;Dt`j2^`(#wtd|DAqc50MNz9`LUBXuFCk^I8%q$+2E)9
z6zHu9y)hHI?xOX$w}Ejc_3>70*osXc``un!`lP_&?0>Xr1!ip16Ndz1YiMf>9xo7Y
ziMJe1Jq6;U<0FCMRiGt}RV}uI8b=^?*sKm)0YA(?^p$g&QuS83Z%nyQRz!TMsxSYg
zE1Pm(K+tbmrzNG(e;o+f*XN4--r`XR3e=$s&5wS3I&c<(5*m&-MG)2Dj86nbQHmlt
zFD3M827=@OHL(ba+_c#f)8goCm;??=T_XNEWUMy*I$Pe?Sk%Ak{};M{`ofJ!Bxkq%
M*sU?URj}Lt0#Y$XiU0rr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_contraction.hpp.4808E834C3E32AEB.idx b/.cache/clangd/index/reference_contraction.hpp.4808E834C3E32AEB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..19acb98c6386220709cb3c7d012342b8c54711a2
GIT binary patch
literal 3162
zcmb_edsI_r62JEbVt~s{ct~;yfk1eQ@-V1rpn`xrL@E!XBC!jsgr)LO0_9a72q=Q!
zQshy^Rwyl6fg)-V*S1g=P!tiBwFj_>w&&C$g6N8E)$W(z7*2Z*e{Igme0TDjJKxNF
zznL##fqs5rB!>BH4a`W4O^w#VFw6jbm8mK5d;-JHGchcq`rP_0$-_e9|E?RW84urV
zGI{FEjeuZjaOmNN_v*j4I+z?|%;(%q4ITVr<@O2F)*qg6WQAKtv#I?`Cu-owAET?A
zw|uz0v?DoY;KHIWGuyJvYRdL!U=g=3?Z>x9h%Ah<Lxlg(`T3V4n}540TQ+r)p?LeQ
za>3Q?_DZk9)8?rL4W*xbvH0oTFFSgZwGRs)zP)0~C+G8upwP1JNrr!JL%bW!>tpG!
zjXlw3R&)CdE9*?o4<8)o=CMC|Ibylne(m{5mq!eV_hxL2A7d9RPt80KlbSVTyED{z
zJS#8Eq3D6zg&V`Y+x$m6vLe$rC{Gug#5kuP|1`ht)Zq_@yq{f_SN-km-H-Y^i-xY~
z4_>QJd$I6Nhj@V77&5g>)HXE|`TZVdN0d|k;NeW0mcq+Y+bfa_jlRA*9yWGi1{=%w
zJtPm@c-P(L-pM`>rx!*1gWAz`@8>Dxfzy@mI~BiVKVSE0Ph(<STJ;+7CcBm~ZEN0r
zdBq;3kNCak=8l={BO$Aba>tnB^pqR7EJ`A6j<?I3YkRRXf?e@=vMgWxy^`6NTX(BH
zv*}jk#DA)jez$)0uSxr!hw`cQ;X7G^vrF!jR9&3@cXd6bxMOg!w7hqE?=a(8-01lS
z6V~Nj(w@lfYYo41`OR_ndD|S8G#z|oJlZ9m>7?$z6UbO{=$|gd-tVM!Z)MU$0mXaW
z?an8Cchq*;q1Ll-X;L$$%uI}lL>$pt!mQlBhjUDiV3=UEl#-gL1;#==LLcU%mq8=$
zRtG79y%3|_%-tN-0^o(ID;nLd_h~T+JtK@3ED%_$1*Svb0t~}x+%XVig>Kes=MW@v
zub>nLY60)RwRnT@!7IHYLt=q3MkkskE>jDDrAZ}?5gEK@ilET$&X&#&Y60-Z-qD!O
zqo2JrBy`Z*IXEW==f7USB;jKlyz<K9Q%ph|nMm{`hH3$D5-p38vF%q<gaO*XmG3H2
z3xM-{$`ns8J#e824uV}QT^!T`-~%i4LvydE{mLX5$hlBYXrvYZSKYq&1WTFj;t*O0
zF2FejxZw2yCZCwrp6wZtX&VwOWRk(jFjEVF&*{!=vT3#bl|$$vI2ZTN#UoxXVA5LA
z+9MfR8Z}#B56Z)URtbQwO_Yx6)!dSB2qyBN)KQwL1;BR{mzK4KrBNmX10`t1vf`)(
zz@wJ=My#gwlN6zkV6nN_Q7r&|Rr~bwx~}eJ1YvGCOBb50%cd}$Sp`dpRVbgJ83Y_-
z=&HE|XF}2*JP#oxK+^G!@nYZ$@Qxs9N6L{8aq6ns2X)sQXu16S=7kcRq~)aC7CZqd
z30*ZAAxe^#($eJ+3-9)6){FYqrGyqOvyr`dSBsA4#`7QwE6=lS(~|4D^BRFMuB)at
zWCzqnKdX2RK}pRY_ikfGKgPA_Vw^0-DPRe39c64s*_miER?!|Z_CCJb>T%`A;V9#1
zL9{Dm9Imta>%>@&st^UZP65saryHWnBN}Rkr3fg+jY@G70HAeHbhCW}@oi~+D?x5+
zEn<+g&|T;c8<VtRxxyP#C28d<WiW6nwnW=zkD<~VZ6a7J*aV&u%3|z%$JV54lXQ?U
zXmK9Ly#Fe*s+OdKg~4;QFmaeC<U!Km;_x|IggD}j26ffaRd+D#LZbNXa&RX|h@|sz
zVLomLt)tQ%@^Kf?q2p1R9Vmx6WmYvmWD9+djO;1<IYz1x07kH)fK;uo2}spO03y;!
zq+UQelHSGM#f7FXTEobjIP`}^i=;hGJuNi_@-p?BqxtfDEi{;)nV$%>;-6b~gs&Pu
z^na>w`Pia%uW#ypLFZVA^9pfmAWQj7)Rw($vq4mjI9{B!h6QmRaRH!JTn?zexo_8E
z^maaOoR3>T#lZ`r+UyL3N<keEpqhN}b4@~!)|$|L_fUr*LYHVdq|Rqp;QifiNB^?J
ziK&2^cXLk;As~8TG#}X1D3*qF9gC_&VN0Zv)j%K;O;Bkrz8d0#6@@W7VIh!n<5}mp
zh0|Zm+x=fHZ>h>2co=}9EyRrqH5~>~Ah)W447*>vG#Jt7{^X~5oqn5g|B?JfRtRz`
zGnc&?k)0<(lJxypd4DOuHBolK-j(5O1eD;$B{&ai4lRm;P|beGtlz`-gRpgG74rrG
z69~OKm(g%1psFsS7R{N=s_Ox0KPKt-nQ7qR^KS5yi}uq6Y;F;XIu|#{#rbd&<7G|H
zjyd*!qC?VFTq^+tLee6xXpUxWV9nKF&Xlt;%utecp<Ilii%Hs*<w`;0lXQkC!vVNQ
z(wU;nIa-z=O9a?-b=$Yj&6d{lCh=UnqcfnX+v`t9h3>wDXa%|leH--d8Bgxs_~xHb
zP`1T4Ien{k;BYoDaWN<!%u@`GOTmjhqF6n%ABSTFW@ErAWy;@`{;FG-PMY{?C!gb<
cuFW^&n({F2>#G<oW|~Ky=FwF1$k#mn4?h7v0ssI2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_bwd_data.hpp.7B56299251FBE761.idx b/.cache/clangd/index/reference_conv_bwd_data.hpp.7B56299251FBE761.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e902493090cfac8e62d2355a2f0d09da83634b35
GIT binary patch
literal 7830
zcmd5>X;f547VfHQ723z6fd;yOwpkjG%?7oRRoOuS1<^rVB1SMSpaC6XP}G=!3kn#b
zqKJx7(YQoVG>U+N8o>n_M`w&-90*P%DshXNXfjhY7F9X(D?f6M9JtkA-&?m{y}Dan
z#*dDOC>SJ>gr<yMoSwR9hEyVvn2^7Pi!$frLh}6piDdD%>Iv!NZhOkFe*dd%dGb=(
z(G|5(@%yEU%DGmyK{K^m-AB}USNI)I_4hFUX8Dd{Z_P&&qGF1zbRWF<c4154SW}~{
zLC=d9ZP)CRemiX3kcg5y%I*!9MtR-(&?oA3^NJ4NqDrfbmIa}cCW@uas^p4ibuQB^
z>Jomr_8<S^qK}_MFM5B@nY{gzC+;g$bssu1uFSXg$c*9tta>LV+k2kll<4ZJLpgrZ
zoW`0It?hAld2zvO`%&}D=Lh?y?+70IPTsm~+rshNEJvHpcQ#dy9pKS&@zLrB#cgfJ
z3sc??6K4)>oh%mr5R}(iS!V8-vnb1WPJ8*$q(x%gvmY-^nQ(OWL5pnB+p4CtL75)7
z`s?SFuP!*QO(?VeY@<5y+x}}t{=5E(Zvecs(`jzo^f?n8pIhxs4&L+r7dJXmT$&x$
z9Uf@D@%YQR%lFQ^RGbp+cl~O8+ym=pUu?=gny6cFT)QN6>zef+6damqxz18<Kf53|
z<M{T+n!v7c`A&i1hK75FU-!LsIpeHq$iCL-BkSCTB#NroM#H&@f$_m5aCH2hDtXX#
z-<(r6Wm{ItCMC;CM|ro4|4O}ZI&15>HFvXaY@b|_mTtFXuY-}_*37$YDJ~@jPXBfG
z%8gvJEd0{-wtHPMf4h)S+M2Yr`{zduzK@o?8u0hI59hR}HNJfQZ_Vr3`(`9poq4)^
z=lj-q<0pKwHhs`1$qKhn!-N*W^6*;st5r_JbDi7cqdly?^au`LR`zwLy;Fr3?kZU-
z{-f--*kk6&cihX<y63zcuY0|&sjy_}m#wEBP3a14G(b6Dbsw4Uc}q6Us&3M!Uxo||
zj=VOxXleYi5Z6j~S?<-9_g&m&-}wDp{TGX0`z7S>zav@rVfwpyh?wO8xAm37<2Q>?
zV<N#}4q>i(A~7H=@*^9S@cG30*-u62NZ4_<aauj0?19d&QaqwfL=mb9J4O}bswb3f
zZocnDv7i4+5h@6KjB1Rlo=~<+V)j6<ONN`}(8yGRlf)zwJ)vx!;r?s3=W_3xLaAJW
zw;}m9q<J$?k)X%bJIkz(*Y#>BLShjz!My_&2?<#f_WV_1ya>(7Gem?(Owkj{Ua<JG
zJo~O|T7TG|EHZK6L{mMX?93BGS9Jw9p0t37%+?ra<a$EcSx@fBve*AK(i9qy+11FZ
z8tL8)RAl)?Z&Cd4SeFO~5e?(i<2>|)vRD2#cx?Km$~+NT5_VKX)Kop8?2Vb@$H#S~
z%@m;pVMj(pPSq32-c&Gj{ga$VM>!lw)^9^zZ7B53Kt;;l^;6g@R&|>~84<aS<hPOa
zn}Ldaap^+G$gs=#BD5!qgxG}G>j`Ba2!2mFV*3wq{h?qi!Ri5OYdw(|LSQG601S?$
zR6Bg!V0L*RMC2S3#Y77|q4RGRj_Z2m_1Ip7PDFjEL#V5sQ1(C7H%_|7zqS#f8DU4O
zqTTg`vTwcfWNq-=zm~{hKl1z?NZx@|Zw4yTg{I~fmcII%2yKamSXHc>o>2DvIi<Tc
zX}nhVhlpsfA82o;CzSn5O#FA6AM4(g!vTa{gA8ks&6}Z5WK!MM*3i5aB6J`N#M{Pe
z^@Os2UEqtv3T>7tG$a~ok-U~@=pCrYZ@ZrD-F^Oo4nmdK>seToiK4`?7i6Z+T3BI8
zKELrsDB$7ZaC2&XGQCgIeNdkyLV;qSiV5|&Cl#WeNre1FKWiq`<D`8;tyN7gub(PF
z0^SYeyMZh9NCD?Uc`oGe3M5~F`mmAb>aBA(dfhe@@N(bfqu8|7z;`v^a4B#p1zN_x
z1%}@Oehhbl;hj8w1zcY7@>gKc6&S(fZ^58jJpKu~`~>^-(|AKR?IZedo9+ocbkoXz
zR2salI5AWv#S6k0L|FWIdP02R`G-<m1Y(LnEL}^AYe9G|i17b=LV@zjym~2a0#Qw1
zG{X<Uq=z7h$(O_Mav1TEx4Z4A4)23fd<aGzf}<I(fuS`pjNwi=y%Q$0`Nhbx7%6E;
zQ;Q$@t}(dWPR?e!_4L71Ou+fbG9PIeu0)oVNJE3A-Kosk4^{-^At|12F&$GeDVb)P
zkIWgaM5dL<oKCYJb?oF5pKG$W{H4&x2Qq!8puyWC<+Z;G<}?UWJjHX$P%0(GQ@y4R
zV`!S!G!A9>Wb_%L6xRWtI%bGc+zi5+L7(AC@h<4I3v%*@Vc212h{RZ6TrC{W*7_Vd
zeU1jx*DJ*p$gKi-Fnj>H9bkB9ROj0MILW)e6ld6F*fTYCz^)E(xEXjfgCUH+3)=01
z9RD!%I1D*{E%dErk4=}C2{=cS<3rU5I1gy@0Ef2##TLNX+)dzl6NEAOcHr60<DY@z
zXI{P;1~kJsCVvwK+~o07sCWwd^vm94b((1W20d<`*MG)iAjRJ!;d>;~96)`89Di?;
zrv^ZiJaK%TsMD}$*OTw5<A>~AZvS}m_5|Wl`+?nlpkvRM5AE_HhgTuHRj3a~oX+h|
zdoXLqAOWZ1R5gtl0-l9uap*n#9*5?s<~cCQte~tuV-oOUm&IDfI|A?#pr&y}z?Fdf
za<~fMDjpvL_!y6?0j}opNuW9j9GJca5Y)g-S-|Ij%Q>KB{1y<?0{To>z^y>n3j7)W
zQ;0u>Y8tl$Tnx!Chd+b(Gahe*cq5NDL%f;CB@mbJcnic^cw7cmWzd1Gw-*NQg^}#A
z_Cc3@P|Ns7Va!qHdIGM7x@zdp_>C~O5hk$t=V9=97)kjFhT*m8sqz9KkgCG40~u3e
zn=sbY`QC<reu97pn+&$3HYH#e6BiD-nz(Yv&BTpEV~oa_(VQgUXrpK|TI>io#wf;&
zCTRi38pWE?GDg5>k<nRXM#~BTHz1=19ycPRMjoF-M&}s*%Z!al2g`5!ksaMfX7`Z=
z-O;oYzG+jPG!Dc8>>cgBs9y-gnQV)YJM}JsR6lMtXap^cQVaSyt}tB~LiDLU)V|b9
z=QuPdY>mVU;uYz-^k8Ps^L6t%v_Q9jLkqnZhA_!R-itU?19UYYnAzQFpgYauTA-`t
z@fqNK281yAv%vc-kG}>!UxP5lzX*IT^0*23H1YTf2)F{GnEX``aFxe9pl$~YrgkIX
zolv)v$K_C0&g0#1#BLbM<oCc4dw6^R`W#^OfPfD|pMyL;1bq(ixDp0b!YHP%3I<g1
z_ymkS0h1X2D;WC~k59qaQ#`JRW9#8$CVvKwJ;QKX@w?7p?PESL7I4;(tiH@gj0JcV
zfk5U?8-Uja9&ZM^&44@d1~8<7WmW;V0KXPK96bbq55Z`ruNwq*^Y{hOz2J|1FVyaZ
z{!IQT^gGHjv49()e<L3*I$>BRpL_1Yu)92d40Vs$vC)homFe$+)Yt`bWpciahS1xT
zfbSsNJII+{z<RLD9pp(djTR(R^;{aMtI_)+`MUSANcAsnZ12pLwq79xwboK=L#GSa
z)6$beUY1@QiZF{%(#$8|NV7;Ky)Ftk$}CDr@0J3-h|De`B?~A`$gGLS&B&~oVR~gH
z*)%{5=vz>xzqrz^P~}8uggAmrxjlDSDwID4ee$sIzdk)N)Ml^)+m?%s3x`~7Tsh=s
z<Hn&eiZRw)3{^y1b1_sAW6i}-MXWU+LzQQdnr*oODI0j)h?I>yK8KX&7|#D8?B?uQ
znXmukf^7Z0s*Hr-`-SA#|EOF2QMdVHZa@u*oUXU%&7E#f3xb~R#&*{;L3O)-Ac>(9
zDLavdCYp^elBcaSwz)xOZ$QcoNW)Ug1{AOXMNv%Mh4@2n^Md-rrlzQh>v_xEiRwE@
zaR;fXBy~VCqqmnqXOyQ!OUK0Lb&#U*9un@cqH%ZVZB<CaFQxydYTu)%@^?Po|Bb#L
zBqgDm#gu%A^C5><A-oFpVcIVt+Z!qlRZy#z>ANQM<<RGXGJUs%HUI>ih!Z_%$K=p3
zzqs>d$(xC$iJFO88hixuHZ*NWOE*o$kxbuMVPZ+jBwO~>#mKD~4WsY2`hIq`<=Q7-
zyyaKA<K9&zOIja9k#p-rW}U2BIeqa)(jot+ci-~ouh0Chd+O`#IC9}G0<J~Cot}pj
zKLp+n*{w>7zk?y)!M<Bm<G#50MiB;|{GVMmTQBY~TrBpJN`KeIkYZoax3AkH{asr}
ziv30ZzCeA|Vae*1BNL2?=6gtS4=Jf%wgfmW9Y4-y(_8*ZX~U3s#V~Og*9Gh89a8_I
zS%3)f??t9>9#SE?aS|<hJBQRO+kIL>UmR3|xBN}q_3Za-fU#vhImtF;(uS;QM4)+&
z?4Cb)nZEH!cTY!B%k*?T>E4Ckl_Vq}X%k1#2qWOj$oMj{q<Qmx@!{uXZY?`Z1?(s|
zn$l1wU?)>2B}2}p&K%OJv|IxBRC#j9OXbC(5Q7krMpFTYilGX6_5uzQ!xXe1D&Pnl
zq2_8G9I2)$iqtwdN=?fV0oNgeIwaB!2)G`J^+>_+86=+J@kK<+Bd$<rLb!>?%?LL$
z{AAn7dhHQA&0B6bPvu<?K3+3!_@7+AJkPnX%;MEq<3D+sm%}m|qYhsP`ol4nUSWy<
z_cr#a5B&0`qfh*2q&TuKs><|zed;&VX2|T`Za$qY$xKa~F^jb8!wwGYQ2e+&5zwR?
zy+mF2(Usr~Kv$-BZV6RNZ_^U0Y?d;0>#Xn>s{jpFv?!3!Jijn==0g9Te-tQtNbipg
c{UDLxF2kC&k6sEXQfnKfmBfm1tl7u^0lG9=p8x;=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_bwd_weight.hpp.474E92061BE55B30.idx b/.cache/clangd/index/reference_conv_bwd_weight.hpp.474E92061BE55B30.idx
new file mode 100755
index 0000000000000000000000000000000000000000..77a186319956c71237dbe6d57e91eeda692a12a1
GIT binary patch
literal 6822
zcmd^@X;c))7RRTG5n*VW5om_R0fu4U1_l_}WM3tUg8LFdMI}N|P=ZSo69cY*pcoJl
zlmr*_B$$z43>rKt5)cG;qEAI*+yfHf;r5DwOLtB_y>s$S&Y_wg_jYx6)xH0^RWXra
zVWq}AUhsm*jO2vn@q8YSH;VkGEl*8M#5~@Q>O5XXiE?uCk`5Q)tA}p|pa0Q+Vb9{9
zL@tSC$zh^%H`{Vn$f)w^yZSa-Xhu#jZhZIp;)LI>&k8D^FVy?b*p{)qzu8S{>samC
z)SXyZK541U;-}Gfo?gAMG3Nca8@JC)GT23LEkP#5>$i5hedUpQGVnRir}c^U@+s%&
zo^Oxzetd@dXm(tt!^qMTGh>h4PXFPvw%aTElfL<@;C4<|%<+Wksns5n9Xg6DrQaP{
zIqj+Dxg0ceYno45!iGbCJ$ltM!$G{SQNBu?DD-@;RA2N~FTLSaZ^y53KkKEJRBu<F
z*n4HxGn>k_F;my(Yi-=wpWL(9`a)~dGyOGks^*)ttKk!Ko8tQ{JPtMLKEC;0*>m4o
z_kBT0z4Uv5arxEP%Fx(lDRGqPNA$z1O)UnSua2>sml|7K;oYoZJzbV`tirKlyK~aU
z=**5;V~-sudQtS?iPPjOd=2I8c;R8sR%BbYc&oAe;>!3%r?;5zq&y#9bUv+>Sn;ts
z<y#s2UAf=7Huo~G;@P_d8GaAed3WkfOC6Ec6#v#LOn&-seeJ{acF*oMS;w5+J5qiZ
zt^04oh4bo1&zm~!n10!$KRUf}qj6{cwCrQ)mme*dbHDCRfv^9@o{^eD9Xj{N8b!Cf
zZ`-|J_q+8kt<T%x+i=4_VuLa$!9GwLx5BISPzdvhVfft&!HWBbpZR7j7rXTDxo<k>
zhtl$Rf9II=qUx7Bw38bBTGJM-lRshlM2D|@YS%Yiv?lsS&5qBLuOHv)^Wb1yre$=C
zSTW-0^mz@BvM!}Yo~6`+Uj+WTa)r+2g`q(Ki%w4wZO^Nju=c?e%Tp*|@A5~yw3W&8
z)YW)AenkK8FJI3+N@HCikMS4#JE|I<28l}YqmjL8f7_CE2XwK94v#6t!cshX=mIWI
zw9@sZ@3ua5Jc77HYDMa*8Zbv2D<;j{ko~m~s}pkx)+`}C3|_!x#(eotoAT{*Xsk!x
z!B6aGuWG<N^ZBK(A6;whRmWO1k1-u-DpoZ-O^n1m9%9fE<}PL4wm(LQsZl&;JUxDt
zs)75L1{l@Jq$`?ru_k!~B^D~N$<PH{(q@L;Jy+Drr?Dw{18)m&TU7(*jQ`GXNltma
zUWoamhm%-%5{ri};IigOQ|g$t*M6h18R;R^JXEG?z+CvwC#YeyWhad##2jKCB2zVB
zKH#2v&d@16kjCc3>?`%PS2bX+{qlUwiro)m46qt8OEsiIRRiXef3B|{SK2mJ9jlSo
zlW0lwR1KJ$-FZoovC=>x77%j{cCEp|Ll<yq)2Uc`ahquajV;Ine8fJMss_xr<*iwP
z1(EY;Y)s6-X2DLX2FwrNXUXT9Dfzm1BzZs~HZH^-Ll<!A`Y}9Snl=9>jZH`o0pb9A
zRWodUvd8fjv$==Hqlr0C9O$5Gz})-#n&CdtyUl|WoQvsPJc=2-fXkcc<4<1{d)H`T
zH4+0+bd-Ur0rR_kJ>OQGJ?DY3k@jF@GqixFd76Weny@e}R-erVOfjyss~E;zpvnmt
zX2=O77|H|)P_0YV7y8;y<fu{1YT;^o2uq5oMZ#Ld;Y=*d#KUZO?R$+kl{Io9pP3sz
zH$tdP7fe1cw&F8!VR7LsrG%t}LYI6dABE?m2#KL?&co))T0T>N!U|A0%S9-(2!*j+
zj6#Z0D9cqSq6$T_{WYP8CKS21J?+w#X=g(DOfw2=M&T^Cq0lxI#&SCfX-A<fUq_MG
z(KwcGpokmX{<=|kH;Q1NmybpH*bpYxj9%}xNu%j58TVL|SX*$Tm@Qbe1zQ43PNXh=
zn(3Q``An=qtP%X1Puz63U<TO2|A*Rc&sM<)JfI-Te1S@gAwnsldp1h!43bDf6Pwp!
z8<2<rwyVW1z)lO^_=xf6^=f=3+9g^J1I%Y8x=wUsX_D)tVe&TZYFwfo--Y<hFPQoT
z(;%KfW$P@uc9q7G?8rotOymLgQ_LwOImP8{EXl^h_OO#*Jx_YNaIY@K%o5L%LPk)`
z9LXFTmKJy{@L{{mMn>7l4C0+)HX)-;T;7a~Hgh=#iF1&Y_0L0+JY>UiA+jn&4lI`<
zuTtdCKCcXUm2tTNi5rj<@_}L+k)#pXu-t;IT95<FSCHKmBxCs&8hZ-`v;E&eWAAXe
z6OHX;85UV0NrOu(On9tj=+DUs!W=}A^-A+3)`8iWJVDpykL|uAD2lN*upSLcM=>@A
zHXO1wu;q}QfgOkZwEgs8>82QeZGSx|CMYIAJ3tRKpJD>F1NFLwB%Wf*u|+wy1r?>3
zeb{0jm-l0f{amiV78NW{jo&$=s=ULSe9>8~cNQDK7s0C#4RN7eOyG+Is#*fH8^#3-
zR5b)F3ltM&6eWW-WKNA&(rjx>DT!2xmBbb9rO1+Fo{Jqpv?wyx4kdUD41Pk+2&+up
zv>@_krZQ6>cvPaaUf*1st|Z|*S1~t;4cItE9Eai+@f^xWihLBrhFbwr6mWS5QtaSz
z5ppj=0jz&9axdocZsfk3%S}kpgo4;`YetG@E?-27i(GC)?ro%S99kYIrX9JrbNMoI
zzs%+9$n!c1Vf}9)&l_C6i9Bzz3}H^fcc6L$@k%O~XA^RF+-D5MERZgcffvPmBKw4s
z>dDT@-mERfE@fCbC?+K=C5E+S$TR#{Dn&M>Xjr!>rVQDXak&yXRw7T<{}7TNLY(`n
zkW&@%V(nFEToszi@^R#G90js`3pw0EV_3d}9PV(r6FGEp`7sK3j6P!hpCkY0D2nB7
z<k-!9zit%Sji#{nx5)D?3IPVmCu2NV0z-U4?Mw2;iFVRw^Jbt7hS-NBBX<v$zi>Cy
zz2t5Mt>9El10^ury=9!*hG)6&JqN@_Vr>hLAYvo2;gGGwmP2+DI}Z8j`<Xz76S2|v
zH{pnlet-!_Z1e+7xMJf}j@`Lp<Ft><`?1r0E>~cu3YNG1QN1iS@JX)DzZG-;c1-=d
zF|zZn{po<ZipNC3>abxQwuEWl*&8?K3!Twd$Smh!!#r%sYG5Ar&ck8Ag%=_!+q2d>
zk{)WYel0eGN&+@Oqz4b$gA*qMdm)IE@}wyKxJg;}NbOLMsd~0{s0i*52i5oQ{ztd4
zaU$KyXc+^MB~WdjpxXm~MwEa2<-wb(%EKrVnL$nrR31jBxSWlRvhlF}h@1DsvwPxp
zk=`RMBW0jV6xlsl)?gWYoASvjrD${ap%1v;Jg%tzdiABJ5oG4-uwEUzD#MHr;TYJr
z!jznBxH992&&vlN@Q}e!kP>}h%MOA{{cW5`Vd5I)1%(O49K<>Yu?XbsRBGq**>6@&
zB;|REb;|IziO*Ca`zquJVa{i|kaHJu0X}nJ;-}hS8XNwRO^qfQ_cSx;5q$pNN+|x{
zHp6^IPRobOD?Z~!yA2nGkP&3I2hSPct1e1cZ~Ve9R);*Y9_!a*Lr|6$U+Xn7(W47L
z;6Q-G7osxGv@=J;2F_t1O5g~FIJwjA!vii*9mIgsUC)-_-YP@%0T1Q_mk~W9bwpc8
zpw-|}HP{3~5)P?IXbzrGL1;cHIQnXjL(5)Gl6fNO@XACnmZL01AYT+?C9>k!ylHQJ
zD4Zz9N5hAPS)>?W+E*XeC5jvoNlUREWGuz(#p1nK%5oW&ma%pyW5_gfvrM2mErV%>
z>V`ORlTV;JGXtmYlG;<U?<AHVa4bMkLZ*E1un-pTKDT9OB95L5_&;?MWLQn0I$MO^
zC}zCjcyka}ipj?g`PdB}&Pz>5ieE^M8AGbH?&-f+Hw}R$YPDJ2`sTo82sV(g@8rQ8
z!j6&$vAWPOp?G0vZ#Dw?gbIZRFO!x!Kh0<0P?Om-w&%p|>50?%B13%xp5fp@=Rix&
PIpN0u{^0RM?4SPwGyTOx

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_fwd.cpp.3703F6A90F557994.idx b/.cache/clangd/index/reference_conv_fwd.cpp.3703F6A90F557994.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d1db7cf73064edbdb7114eb30b7dfc037967f21
GIT binary patch
literal 5830
zcmbtYc|26>8z1islF61`_AMhb+07_xmKj?qYgq@Awa`L!Qz9+WLWC?Wu9Or}k)<fI
zgyc$6H*2~j+`77@<#(ia&bin9<qz}u%sJ2ZJ?}jG^SsX)N4lkDwI~i}?njRb_21${
zz~OL$$TxCJM4&eU;`wp7sMA$0p$-EE0<Aga4(j^C24P|c`P<X-DmY20PR5o++Vn!&
zw$^R#6*XUF^m*cI88+qF>&nv8TeQa}b$plkw)mxMt?=(X)UVJcQ?q2^z#{G#)pJjq
zFKs$E7}RV0)MwXj)1&!@l*D}12E)W_kEH8u4m6IMd-w8LxfoaX_ceT2qg7QSrJDAr
z(Ly#+sXW^@`PGf)go$Tg>>{G^t?Of7-^mrU?XOJ{7E8+SZS>!eF7hNUxTwp%!u3>%
zR#KaYGH*#n=ag=Lw1$(UjmEmH4w+9=!6fR9-Xh~YtKE|67Tb8kw+j_ZxLDsMUP(K@
zGTS;|i+QMxUT)$Z9P+ov36CkWKUN4FeSVf1kY23J&v3M9TT(!5>$yr@TClQi=lLw1
z3ZKIcIn}!Y?p!<4!I$8a5;2;$(ZaHBP}?7W!g1KS<8W+V%7lB_ay6$^!UutTCE**r
z=51wWYZZBkeQEW^cOEM4dS9Zn(O^?>-;u3Z3m(d6IKJ&@`B<VW!F9hTLdzwf)yaFz
zd$=NCgH#pSZq!w=myyzWml3w6K;xaNTU}}|Bb&idfNKo(OrVur$lJmv;*+j)vSc-(
z@nl?o(vsqx1CK?@b{#o({L%}7o%MQ7_xY2{_3m$}yA+iXmXpPluv%P3v`Km+WUUpy
zRw%67{zdR5K~@u^b5T*e<A<puog(#THM^gvz1ttZk69Q0%!TKdOFd;=Pgask-2$)o
zJ13ZAE}*3kNEAnB`TMzzM%icOZX2hTu6P`n{Fz)#c|N|*d7;N|%U$k@g}kb>j^9-#
zzyJN-i1!Y5?zs;Vdvi-4<`$1yRv*OkKX2;&^Wp*l54p1EspMzvsW+F#Dx1X7vvg%F
zh;_&9R26q6>?mt6cweVfQm*C8vF4?IW@NF+SgnS#%_~Zhg32?=GKyk=g7ZhPf6pOW
zMo)j)x}L)~x9Kg>XXR^J(qD#uE8X5%7k)IzGmi9Cq_|;YO|tf#^A!dkMvR9G4@5E}
zEcSkG<H$)W|FTD4Ahn%#rjS0;th7FF%yNpeRjM%i!o{VQ+znj^PJDS{_)7X-WWj)$
zv)QVBQq{Sd0d1knll5_{tGr%J+Uh)1v^aBb?BYn#<of>VN8%=TF3@b-qx5V+wASYx
zhT@)W)#pOP?fLS8?7If#Q$jX9c=zu2jxWEcOaGBLsCRYq)2xPxGm@7C^?tAN^$M0g
z<@(~0`o*eeySaqP9?ZpS*F*}|M>{pBY2@YF|N3`N=htJC7YfWjWEP%3H>xgMTzPFF
zCqJ$0`jpp4r5BH+>cYj1+s`vTZEbN1D|~Me?QM8guGsqOCF!2lZG`077s5f05_!a&
zox8W6h!u|bYW!)hRBGC(7W1cs657>+EVVBO^0q&xW}STS+uFP30<O)85rb~u#M6#X
zWS)G{IA)x5W&EK0;_k19kB%_&O!$oU<QA8?efIB`m0z}*CspZS-j0|Q<#>PNS|{lz
z?qeHTJ@S3&b!Rgyzk#mL`5KYYp}r{`I2`w&Ti~MD<D8<v%^k-o1-Mc`_;JHc_ax+_
zR(AVTV5+?|5f_HTu~cv>^4~8r4{<o%)XbRn+Tc)b;DS8v0k|H7!^{nGFq1cWx}$?n
zkQX>3TO_$8g}z@P2XCH5n(FH$^F)D*>;K^px^f%+tv;vkPt**T6YOTutY7m2Cq#%i
zm$=~f3lws$SH|X#kFpv+;ZR^1v8tKw$oPpK?RVp1WK8$^aRWC*$bY(PAZUbnA^hP&
zUa%h7BEcmg^!)<$_<T9(<v{6@o}UPb@DGSwhV*~DCT~Gbi*F(zvO2-?PN17UsjbCG
zt$9df#p0L8;&rVjlSKb&TAP4h%{cVoH^bjKzz>z!YW|4stOq;>%y$x=nNjRu*B?l~
zw>f3mCm6b@jI2@ME~dtlvGcu%{;gt!!%R$JxP{HmhB;?*cF)N{bhmp*Ov=bh#ESdC
z^*-=Kt#}Ju-vUo)g_Pdd_DtoY>EF$j>I6#Y##p>mEMAGdF~Mj>{lv$TSvtvxJSj?6
zK30Y3bREcE2WpT5fz=CSdx09{AZN3en&}XbJIgOB`6Qhj1NSc+L>5_sYy(pt-ti0n
zk@`edEfB2*`cMp!H3CFOfIbwX{@PQ((aK~u!cD<k!5VUbE(Fl0FgKv{;RgMd2;u@G
zvKyYrQY0%H!QT^EjX<^$P*AaBKz0mJpjhKiPUZ38d?ygGhB}6;;Ll(cB1TM?AQVbu
znVOo~!A(S#E!ma^k%jQ6;F-nv9SA2mb2+vFmX+~uL*4}BAdzYUB27RF)npuqi~}jC
z3A_W2e`OD}99n9pe`1rUM&dEVQgN#IDp<=zmIPH|6|5*COOh(N3YuYmLD&6q8ejaJ
z3Fx@>rs~NxTGQ^(l+d(=wym=lQXM(If>VIVx&<h=u+>!wD3w@T1t?WmTn#AISX=`r
zHCTKbP;O&!6X0nA{HT+f0Z%g)_W?>DU{^Dd)ek8BSUdnI16Vu=D1%r$1SmsTJPatq
zSUe7R#sNPJ29Y%ZcqUNn{Ab8&Bi_K91On1#2%V^61lD6fc#OrdctR|m&4|E?#}ndF
z>=vM)`4_3p4RJJ8fGP|f?a^Tnt*Fr>io{lhs$vG0iAar#Tn93!TY7;=FOY$1dbXMt
zW+(Gb&RZ>byv%ABKDKl(vO@CIyd`XWd(}*5>eXL};nq}ZI<ycTTC=)Lm^cU1yX|!B
zoT&@z04dq&0}U-%V^d>xA`@8*GJ^_*hDRB-EotibB0xaWnsf_jK>re0m4H-<#Z`b*
zg~io?RE@<ofK-FUw*l!k7B>OuCZLS+Zw6w`Ko-S)fYb*xU_1#(T9f**cmR+Fuy_!V
z2C;YukcO~$7?6gscpONN17(!|1Q445vJh{5$?adSs^2~D=)$9L&L*}qvW&RtP>k`M
zVcpAf<I4IL`a2|F)Fsr}r_4UbG^2Mz+>yw#71hcCM)WWPA2)jRn*njUQPFMzQ|MtL
z5>Z+u7FPjU6&6<mS~V8e09p+e-v+eXSlkC_eZUlYaJmuE`muNb&<3!05YPs(cnHvj
zuy`2IhEbekf9!9!?N_ws9g(DNcQ*y)R3dWz%#2`YVf+|8e=HoAbAP|u_U*}eOH2>r
zOdc0XDkV&hV$(eml{nKQQ9RQlQ9RQlA%?L-l4(YDII8`7bUWRh;6Pr+oSug2fonZ>
zY<dG+-=Jd?9BgNGnY^cCzFOjaDo8rXo6EBg?+XI7)P1~6@p=0Mj?1(9YT@=*iTQ0>
zzH`>W<p)NMG7GEa=dH?Bmwi&Ze|7p3F;$7I#2#OYH$ERbOd$Eqn}6B0ltVrx-Dtyx
zQTzAAcGL1?RFTQcZZ7x|h0fJoK)DMz!}w31t%i?*7m8;x7d9ILD-N$5hj)gxNMP;9
z%k0OSL431RRq7b=)ab0Jm-D@9+rK#ohvdy;AoUnKMaSc%;?XI(;^`*kn6VLiWMczx
zZNMhjC~zG`6YSOz{_H&h$EuNyoj|V>m_VJUUqkd_@g`8Gs&6Z_-e_h_%-h{}h1&Bj
zKdbCT!bzvn8L$!7TlwD=5OUU>Gq0f2=PWW9Mr0EIXCr~(gBu=I3``EkaN!V*EUA|4
zY5AF#zqd+Pozbik=%=Rb6}s>2Ja1RP*%z4rre7st8zHhxOikEo5Lq^48+O(YF?vR*
zr>aZmE&FwH(PWl-TMp9nlyS;jFbqVNHbHwKigXA%7^F+kWuw3OPrr#wyohwnAMS;L
zfn@`;9ljLL^5f^u(_gq+x%%hi5aA!<qkwn^Rx0~%3HAaEWn=_1vU%E%e$bgXoF__d
zL`e^n+<=n4DCv)qK5UYM6;CHvEs(wEEU+;!C^%$;hnJ_1uYZ6a(;Mf_-h~T7$zYTW
zK}j!^WTIpMN(Q1NGt@6+69U(SuEi5Lxwseb@DdjZ@d*m>(^u0S>=s*Fsx4k<CoCZ%
zDk-&MrLl>rnfWSBF|v^|O<n%KR>h@dWEB*ZRF;yKX;8GZb@Zu*at2G5>spD4tLoX<
OhXTer3tJ9H-2VWM?z;W}

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_fwd.hpp.4266D54C9E6D337C.idx b/.cache/clangd/index/reference_conv_fwd.hpp.4266D54C9E6D337C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b3c0741b526e37f7b9bde09b144c2fcd269746a1
GIT binary patch
literal 8090
zcmd^Edt6Of`#)=~6P0~-bvo%BwRKLVJ6%+YitcHK!jMaiN@XIql1mswHA1|aB<ZHx
zbYol-NmCk5q(VZ*yl%OUF8nb4yyji!)Ux+{-oNH==aaS9_wzja*=s-RS>N^D%gm`h
zKKl%KJkN-!>y|~X4&(E9JT3e$Zgp(*6U5_rYw&pM4xJBLwxY|4diV5``lj1|qz4uL
zcAn=G6r3Fs9cZO@E9}l1C)L!Jiyo0z_CAjIW#wQ0F}u*b`^56gc1Amv?btYapP_ft
zB-v(bQOX$eqicWWtIXqXFRj@kvTLL=CG-3vFX)wo3*|9my~8dA1O!fAoLVKilCJ&m
z+aB4CpLDjRiCP72{ky)7D$HJQQ^RcfDEd08dP1BqE~5R1CjP%HhIRPXSoQ~76v{mg
z7R||<)|+WF%~87jd_dHMH#4?OPS=vBS9NP8s-4bAdR|yix%W=<{RQ7H`@`<suP<)h
zzdLSq$J~WCgi9Q|pNFKiyIa^<URN<Q%=GH?WIjeOTyb!BR#;E9^WIMmE&UBAlk+KC
zv$s<=&>QD6rgPMipKJYRLXzaxr5!!_c~RCCmA6+4R>TI*UZdGxe%Wx=6V>Dv_r&RS
z7epJZ@6^9O>-P2Q*cAg|AJ)G*(-1^&%DH!~VPQ>#!(WZ3e4oEukr=$$;PQuT>7%f{
z=?P&2Du1?KG##I}YopzMx0+vif<JYM!gX4}?yi}SP6W99`FN4#xC;GI##46N`ofRb
z)>mw8YK{-8jBGboNvvw$*z-`{dBWzbhG%Y5Z_=ph?oV&z=|`Uyw6yOk{`Zfwo!``E
zuktTBG;lsUw!GlbKx31h|LK`8T1UV8_Sl64mu=yh_nSt)-)Q&mtntsBp6r_V=`An$
z^8bhK@q4wZ-qiI?ihqP}L^W@X*P0t$w%*xJRNpGumi{M6=L8Guq>CRvzj6Q8#=b<C
z_s7=zt&;2?5&Jx(Y<Ad#^O*~!7SvmpHzuJmA7Af!c&9l#yyD#OC)478`$4+AqjUgi
zyNs6GKRvwb(t-s=p@E+q&ZW$0tFxbS?ac!Dc*`!w-?SWW1!dpf5EPLsm^%CH$=~)h
ztZZFrcxrt_zf7*NbHe?PKZpNO)zK&$)p_jt$%IbvclT_6dKxqT$M%G-7aG>dDUX(~
zldEiZ*EF_ElD1n#9Xz@%c=?nW$-b0}-ZN=~PWsuaQhO_;)oUg6rrcNW$-80Px*4`L
zs;X4fs)o;fqxwQG#71yoOI2>gyW5NUJ9b{b7hN+k#MgDpyxyenB`@c<QDr8No~2FG
zx_fh4=JV6{F8Urv8W$6@m(^eNw^+XK?PZr4u-#RqIIZOd^S-RA&*W-IcBz``u0O3$
z8m(}=YZPaA+{rFU`}*kgtJCuLh3a^>QfsUQ-FNP+&o8ofFPNS3q3chF@8O}HGo$u-
zzCCv8aQtv-#BG^q4?X2)=HRHNGws9kD}VVU>y7!&HMc!#(mv-?N8$cq<IQe8$v*CT
z@10KT(VLAS$4h7BEG<j9^x*!wc$bp<e=W|4$>qhZUlyK>S37Tw=zU6PVB;{z*Wxkm
zBiyx>4Nn!D@K3ESanuodWVR5hVRJefm5v;SF2qG7jaYkkJ`+sCkwP95=^eR1*$}hN
zl9o=!mw8G<YiwS`EHYL$#5_{hE7|nvUS}ax!4DXKngh^y=t5ldw}74Yx(BY(&;mzT
zZm@i;vLR*}%!ps$@Nydsb+I|hCn{Lk5VJ{T<^FpCXPRkfh0QTcjHR+6W_zuU;j~Aq
zzYwb9=jS5LT%<pAAuf~DQQ1fO)JYoJ;0P<EE3B0bF}sDLOY>!rKMW4X5#mO~36%{o
zPac?EleCK7Hw+3id5qg|H(h1JQ$rB&cz^*u*ESbr_41o&D8k{FFiWhI4Y}WUb#<2g
z2aiS?>R@v;6K$bvh<SQ}{rT>)T@5r8WAkF4#lgyknCA<wB$VCUMbmIBHisF8nJF7$
z4nFd(=XCD(2AYu8<}t4!{TgZwU5LxVz1H@bS}J#GXo({%WtLhf8)8oQePyD?vKe2|
za2z(Tl&-W^HpG0uyQX!9S*C|3q;PWTkZm3E9J&yfoGukBD?y~^;274@Yqc0fL(J#i
zk3Dm``f;Zaj=&LWk!CH@8@dpe+6MoWu~{8CBOxD;!Bf>!Q`r!6J!K&Ll$|2ggsM0~
z7E;T?Q#^PfE={3lCP@S<caDI3oQHTiex$M?W_kUzx-u#Or)e0D(1<h}5i@ilE?3l3
z`VRl7|D1-lcnsl2;TFn<m_HqUTT;~Ys{@4Qczad+7>T-s=BW*W*vP23OCxcAXf^WH
zm8+1{4mJ2bLq>{;w~QZ8hD<S4NTUkrl0qgc7N`%cQxarE`*Hb)8)wrH#Z&{$YG6SU
zPchq}=61-zzeCO6SuC*m;oy=C+b%VV$p8)+U=n*{E^x@@;v>NP2(V`D_kr_$Fqy^A
zfb%mh9suS8T>n$h^%V4E{p+D?Js01E<~QNz_=I^N#e745V@%3GeUN{KvPJj@JI$hJ
z>99@;DkAwNETm2`=|Gwe?Aak40@6cVTnR*#z>Kxu1IG7&6N|fnaW@x#0HP0E{}a&c
z1ax8j>!4X37hi{>>+o}Y^&u`KlPK;yruwn8+D?<t>;S$yfZt(x&xS>B6913SqynE*
z;LG9?;86lT=aSFJfv+6+rEkf&v9d?4l+QE+pJw38;x^#j27FlD4!qidH;a40SH0kC
zGH^bV3H>tR)ZeM-gwoKxxB1Kg=yL%2vN#8N=RhA8=R&Vs=*{9nIJFS^v$zQQ6>;Oq
zVSpS4vd`;-x_wYg7>jHQcv=_G3-sA(RqPLBS`Fx}&O21EKENfvNpI8Vq*BZasQUsM
zlh_oq2kGuX#-s|!j)G&)re|~NO0l_+*s`;+-@V%6l8pGwrjeT@Bo-gX7QTQwEZ&2J
zdyo!^Z8-U4O}|r{`WHWeWM9Tfo=?x`OrT<mCP^L}@o^FZH#H37GwW^E+p$SauuT}x
zq7AkiIA|*}*^10a!ud=RvP?o_S-b;T?qD(DP0W{ryFXcK6G!(oFC7-v2l&ioL|sNS
z*#(Fmz#|W)gBiB;6w-i<oTviK<O)eZf>L<vGOb19$gV~)^+;NeOvyf}z~=SHhG23Z
zR9C77vW?&^KoW0Gwg8X(&4zuMxwCMY*8*8Ba3GnXm{cfBg&f=sW!)_HsZDwv^C~J&
zgJO~Z{xc%%K`|*nmI6#!oCX}yxJ=&!m?mIE*nwg$0of&B%HkF<rUh8CxC1zMaGAdo
zICpY!I%LwJ5n%(o(?D4UG-dHYIOZU<WN|(;&xh75E`_e8+?}Evx|VZs6?Cm)G0|Lj
z3g^;uIa8>RUt|i2^1@C*bdVl9&k7MnoQVFyPIKrv9IjAEG2%2!&HMbaNX;}rVb!44
zh%`tkeKTEX`s`mWDip=|YWeDtM^KEPmLCUA)tbsd{#yPV6fKC>CL^VoC4wc|<mD)4
zsbHx#(QgzJBZ$!^tBzt0Bf(*$O{$+_3X!0Yi;Iw;h>MGnpqRyT!ZPMn7IaO-Z*&!D
zUqvIy8xg^Y^D%>-AtG<2J}5hf_GCihc+CxJPO@4;s-0tkO^l6jCA*Jtw<Y<bu$%NX
zv?BLWSiV_ip$SCUk$1!ook!2(%#=c|k|2?3Dr2??<2>+?-OSw_$jXW~Y7j{Ubg7WS
zJF4Rj;K9~>DsW8Y;+?>8Cl||sqa1j!HQo#yo4NQ3aJ<6BZNR<_Okw@ofqgp{UkCQr
zxi}L#W<n3P#t%Tp16-U19kaMN2ioVrDXf1kw9n<@JZPWC#f8wR5PGrxMbN2;i;JOC
zF^gl;!%e*&_^-uekp_&?xU#JTn3ZrfQVH-Y0pGAJR1O@=xryrlrXAojV^d5gFzw{x
zUSQh`T-p2kfNdWazXOi%xQQ)=mZfk4>t7Bn%elA;T2^tf96HFMJL}&J9h$kg4LY{5
ziCx$D;Bf-Ky#?#v>AKTD?^F~Mpc}wJfx3Yl6sa94Ci;|OqO_yLL{(7CV(rCZ!W$G*
zhP2C&m|ckFNV}YiE0A^ti^=yEF7*&Pgu`5lPcKsHJN?o2Klc3Pq@WdjUn8;zF|FwP
zanMx#sT}05@6SQeqG&x1t%#QBacD)fRF5n<Oe>-oJua<?4<iHiZ3>aNkc*3uxQL63
zk+_(}yPq9jHOu|w9-QibT2ueDX8u_XWIV8_pHf%!0+UA*5;q}Z!gm?(LqifZ^gHl;
z^gwYBG-f%f2TtsPJ_PTV`<32E{@w~ls7IoDWI(ofa>Zmg_<c;6iPR@HC_ab@M;63<
z#rL0){1EFBIilpOE-CWtVmulKvj67^KlS<*e7mjeu2LUJYJmZXtk4INRLH^IP}0rf
zs__#8n{q-k@yo>-$61pRP<Wumb;z2G{M5^YQ@W}B<zKv}J?(h>z=rl$et2q{kaiQh
zD-yYfr*iNFgUm5mn%JN?w;&6+_HxI(s!4C2eDN8*5c}wL@W+He{FPM<Op4hLX+F~n
zEP8<zDP_DE;5a$7A(%)5oY=vm9+L6Kp8}T)d{npp&nh8m8mTMKe%NAPD_$!lVJW5$
zTK2*5WK|Oyz>gU`79x*n)?4>bZ5^$`=l@MS^8cnc`G1ou{J(t&kpm^X_GZ#EIp-FN
zvn5g|*Nis&kmxo`1HX(MiR4I3#?$IzvT<gB{{An%fF3xtdcLm|gD*Z2Pi$1)4%g7#
ziA$vpY1JV;!rm{_kG#vVY|X<0*h}aoCX7Te-a>B<@}Yc$tZh0oog)?km;eq6WCA%T
zj1Ci#j{=Gbr^7|0vMFYf>LQw)&rwXIB+`H*nIusL9LXeEY`{vUJVfUq5i2tC5uMM)
z1xU33(X71;Ny=D}K{4e>QqIK{NK(P#{zEkvtbdjnfAJ;rVf?B_zazg+`N!nTi;av4
zi^9j*UX>#rih4E&0m7GGZ7`_Y+Ty+n5Pl|Ke4mfL$!8jmFlm%H@?ezr`xHR9ji?eH
zq3F2SMR6At2kDH{-GCDgsqwZN#$!xO%_RDIqedGT8p)(QJ%xGjXG}Ugo`mh<{}3lX
jL(<{#MzdWLwku=1W^C7#?T%u*`fOLqc8%EX7`FRg^msmv

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_fwd_bias_activation.hpp.CC5C1A67067C17BD.idx b/.cache/clangd/index/reference_conv_fwd_bias_activation.hpp.CC5C1A67067C17BD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..97991a850b0e233c6b25a4624c26b549d4894207
GIT binary patch
literal 4904
zcmcgvdsI|a7r*B+APjddj~VVDGXu;6ke34}4*_2wPrnoyOa@RM1uFynJ|z=HR1m{7
zDL`HclA0pwD3M@bD*E|Iiwup#?D1(y6q=$2XsMkU*xc)CE&g(Sti?X_Tl<{7&p!Kk
zW23^u9fbrD`f60plJqxHc?3aN;s5M6mM^v^31YyQAaZJ&HJO>$y+mEVJmKf{rdX`p
z_5Dv9e&sJ2;HPFTKHp|l5OaI`P+FKzsZ_n~$zar>+U~@Vl+<(QrVUE#a@VAne7yV5
zvThweCJnu@?*}uN6RU^ZcSQ`=OM2+Wh~wYIEZEaG?{M4NyG|Z6TKS$yHYtUny4Sn<
zb_~wBwtZ2Zt|@zY%Pk{Cn`dtQw{L&nuypQs{?~VRmprIRs#-Pv;?K%Yz8d&ud|b9q
z-0a#7^3dEwW}_c{xq10(C(Yjeb%vV1t+}Jw>%`j+9$!)vHG8f9X5WJ~x_QQT1YfkT
zk52iZ>wakY{n?_5WEEZ6?B42XQ~g%Lmr<qXKQc;hzR~uQ!}*|#)h$acS`R-ub@-?z
zpnG3+!o7a$9e$IJ{)G}xrnu%z4mI<B#p8;3O0`Wx<0r3fJk<W5orOu-BOd)*_Ufvl
z|M!C?|J8lg+P&?iHBYKbI<}PVX*|PwUsA_6S2yh%y1LZ5KD=uF(ac1jJF21NDb{tb
zgkG(i`e66|XM&IHyVhNtAGodeWsidkgu1Iy;^OLeI?8_TUhL`l$xmgPj3aLzJd*L#
z(9pxnbzfxn)>50!BBo`T=}uEO+G6|lOYe-EklK?Yk4;pb74DE9ZVWg1MLYYq%Ue&*
zdRSK6TYEBG_;|qTkBCWeIR*QBIv+f>`&aLz1XLVh^~9?4?uuJe1btriFYPcnwy1P{
z`H#n|d&+yxoV>Q%=*zqon`aj;{5#jmrzm^Xk~AkHf*}7~5lMu)sc2GUMKFP~KsN&+
zOmQUPKPn{tlyFY=MVeG#_6)}vY6AhbjU+hFr}N%pnv`R<zs%p=K!EKU_5CAy=h`5e
zw8!kJvZ?L{0_^FrZCxefE>zK^60`l3erf{&cJx3@)|v|qE3HX0OM+=YmJLWUvW#wd
zl&I|RYRjcbDV7i{3wARQV9zb$f4$?5yvT$!!V>Jv?Suvb?0NQ|^_`phG}?qT#_TZj
zFlz$=cG_R2PVEj(CJbqWg-cDPmIeas%=Qjiu}SS~Nt)o(#mKA}Nk<mgk~<KIQnvOy
zvm_~Nf{CJ|tPBL$d7r6D?Y@?})1(k92$BVP7znV}UJA*#-8wDTnl!}<-b0q}VFkkr
zY+1K|!N%@WH|6G}5mpdEM=%Bg?4q0oVg+H8Y)y{C5;h^bO~`j-fi3$hH@B@9MP`_g
z<`x7a9w&A&5W<W^4g`T1bn46!ZPk;zTAGyLOHOx~?r9)!zH6D(;&#bbuhXOqvwao5
zY6AiGwKs3}C$H*QW=Zn#4Gtjh11NN4fh|AA?Mm@ZoInphQa+;dkrgw%z?K^u7j}Kz
zl5>kD9kG&`jx*H;GRoev=H4m#*E*UMV0M5qz|BB_J=lN8*>&~2<CY{3-=Ps%HX_-`
z0$ZN$dbsz)v*&zC(vBVud`7?*(1h7AAeN_RWFrBaEhNRb)9w<85WXR2All43n=TAp
zX)8t)lSAb=Ll8hJ;j><}myA8D&nh@;Vf%H6*OebKNgflM5}Ufce{NOR#MfTpG4YGy
zlL0~2VM(sEtCW?bPe71ld-WsJzv_?r7-LD1u~BfA$7munkt~W2kB?xHHcT7NqCzAt
zM7H1`#q=Ys_XED-zc7JN9k3Y_>4_XO^o0V<2yjXxQZ15K;0Mv9YSP%tX_K`nfCBCN
zFJ_vg_9GtCjHqTrgUg}9U=<VS33lvD>5C1h2>D>0qQ7#72Ky^dmgf!r3jcKD?^(Aq
z>WnF7&iFYFpo?N+WHCw>&6m%2VNsHAQV<BBm}KANQ3TZt@keSHWM1@04a0Zzq&+z?
zr7vf2M_74$Lj{Agl%XO4lVT!*B0`}iP)rT7s6k>VBT((I2z{wzMKmCn1~d`wM=?6&
zs6#H0xq9rXLtcP2W?{#dq+9BvyAhc+A`5uyxH;A?y(PiM6va3)j-&3R7&#;7AO)k~
zASI*Zpg=*O5E>H21POwKP$DTNSP(3Pr=pmtf~i7?ABx$A6x+}QFqC4-k)oW7E0Cgs
ziz|_$lEp8lmTM30ydH=xYDL0U#DGQde^z<-9NRE?%w$}6bV^o|vr1`1F?LRN-mEY3
zkz+pcfb5}gYj7_{{?Iib6mdA{I}033FE(~sZ=D^2?;@Te4g)7GR-P6)3g~(hii!1)
z4Q1m^<FDbMIR7{fiuaA5$?nwpYB?yuH(?aPmGD)DyBKKADob;jbt8JEDc0&7>%+>@
zIBPg4&N+^Q;#Kip>`tvp%RvdM1P)44CaF2OqD&q|S*Pn=x7YrWIA+qrQG9KEXNB;F
zh{5UPL3?n@tiHd(GefB0v%?J)Oorq!;u{(%2&I@P+bBoSNHK-Ty%0?T-$I+k(uR91
zNZS%^RCwpzX94<@kULHQpW~E}%Q;9PS8$M0uH>LVQJ{pA5~3gp7>H9s6fEJSglMXS
zn-ZSe&}6VvpAw$sTwH-XE4a84c~-J`Yj4vs&CI@c|L>kJx!ij%|5)y!F+E>?!>xVB
zi4T9nNz#ODn~(|;wtO&UVS$CT3;VtR*%lxbcu0?f3s5BBiVKnTSKeCdfhFjWNQWGt
zy4BRW-&}Kgh5wi-zcXt#FM8fvH*vdZK~xLdZg#ER^6nS3f7qC}|1k8rEadd#?f=ld
z?_T82MH?Ev#f&_gJPAY_1QW?OOlB~u;9@^MVhY5oZc<>(+1iv!Y<ZX}OwGpa0i-&B
z)L>q7-|A-Z=KK3!SPyUUYpTlayd8<3q6rC`*xLE&XI)xd(1TkqOcTtZ*g`d}9&H+Y
z!!!zcecENMxBrpoJ!3lHzx!gA3c^hb|KWi1mpecI{I`egV><t$B<H4C4sFciJvRdK
zo~t&z=UO@Mx%uWxW$vbeX&MWxz6FU|kS#>-rC{eZv2&#rFYL)+HpfrqLc4G#3;mc5
zPuBggCib<e%|&Ax4e^63-0%z!K?w<k8N(wxMBa^nOQC}YZP5!e2EGBXG2XN{r>^Ti
z2f&ycBVPh|lN;@30|l3F_#S{wx4euWj<XPN!45cptPUUvSl3r}XrR{pQXMYWQn7SY
z1;xn3G7hr0vKND7ig6G-aCpA7uLz!<V*F@75saV|<8SIu!{|;i0d#-}u1n!Aq<jy#
zvgNu2DNDGx6v;}FlI52nWf{xAU(<S0-6&VRu*Zgw#BLq_UIKsPFQKf|4E$9z>!A5f
z(cN_m5JbT2T!$0+9W$397$;ybBH-O&wjn@RhHd)$8DWD(2-HCpBOr>hm%ox7r2htE
hexIm1yRdWaexB`k8!=%s@?D0P@$64A`_q>F`8U4^rhEVZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_conv_fwd_bias_activation_add.hpp.05EAEEBBFD920597.idx b/.cache/clangd/index/reference_conv_fwd_bias_activation_add.hpp.05EAEEBBFD920597.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a8d304c8b319cb7d95b82cb81222f954914d2be3
GIT binary patch
literal 5038
zcmcgwX;c(f7Oq!xBQ!<PKwFJ8ExpN(Y{6y;OW0|bpoCTeh>&2C5fj4_MHCkz7@Xi3
z5tMaA)WiirL<tHiibf4;FfP$CM&dRa6^Mx$6DRY!;8NwuIsEbr=WvU!?|t>&-M)95
z8L?4O+pGy9{O#C9^D}bN1O!1?;@|9?toOV~g77gWh(*<{@$){o;3?^Or(jhG^+|M5
z>)>zVPh(T^62@j;6&odNxP9V4PeO9w-kPGn^&Nieol}*eJC!yGTkiZTd-0=zj@X<f
z3zn9(O+_VpmjC^t?9p%USS&k!>xacn8I45#w?Dh*>f6WvQvcwozDrcL^L$q#Dhm8|
zhrZ*dyvG|l?*H<gbKobX3+)sR2lp=AQFEo{{-fhdTLLBdS+x=2EpFldi!CPZ>&P*_
z=xw{R_3Y7zxAGI2Rg>uByR$M6FZd+AVW#l()jQR)8^`mWN7o+P?YaEY=h@2|rwbnn
z-Ht5Rr8Qi9mNMarbJm%nFOKCtJY5tPA>CSXXLD_fQ_uS2EB0RXyk4t_{k6<z{&o4S
zIzgLs%}1)I>6fN>Ey^2b-9x(_JbK9^Zh5p>*+1fRJ{@Fb;>5OYU-ukg(TCl)Pt<Sy
zBDH@1l8GzwUKf5=ugh>=zRmx+uVV@`VQkY0Nu~XyyG(6H<iLUJnzQEGs`oGNfBWit
zO*gU{f=`w=uI=dmaYvnXr(7!S3(0Vm>vCkazg{~2Ah&nb7KbfoUJ1&1bF$f^{&zz^
zy;NF$DN}AwPkUamnC@7QI-iX(yXN<CoTZDg*5go}%q*^}t9hkalDd0FxpHSylyODk
z<omr-L%MGjmY&*h*uiJ;85*dY+thE8(6i<4bKlB!jtyk>wDduDhtvnB<E=jOs`+~S
z?@2~X<f)7szEA!gNzMAt<ySB4>U9(Sc|ZL|y`zQS)8Z>lkKLY`j;n}{C9?D8r$-nO
z#F*;Oa(nN_|7=VO#RL;$7GuQ<!UR_r{t@;3apUHIkf2GmCBcL%!aZ04YI)q_bw3t8
zs-j5^R{N>_JXkTJ{_@>*eaG}JR-`F@c{36>BiZl+Ocb*RC)b_cm1sqd!6v21wiNje
zKfpx&-Q8<uk<(5XlV%nKBON1kWd&hMB71^B4D#0c-w6)wy+@M@e2NG~ga<2Ny?@L+
ztM+3Tms*h~c#Co*E=P954={-i?fSEdPI^w0GW>=x`!IJ_fO^KB1$l9c>OZ4NbF2<f
z2e_~T)Ni2vCC_i(o==n3SRE`6c4GyoXKkBQ-I%F#6O)t`!Azm2Sh51tvn7s8_irxW
zDJDhucrD1Q1%(ekz+`dAtf`85=S~`vM%cvG%+{I}pkA8$L&4EnYN;7%gg4O9I))XX
zUfFoQ+<mJ!TTB|`)n&-E4EsOy0Fza;pjkUFb-I`oV3Rf^ZbS0n2bh%UR!H_-?|nj(
zD(qySI?$UHpsxR5dR(`42QCyjR%=yS4_1JB`{lz6vJZKg)1(BegXBRjtN`_ndvzsw
zvJ=KMIUcJ+<)N;u0CiL3E+(>a^&bprgw=K?c4Ah5`arwpTJo9ISv0A{YJa)ED=R>K
zNHw`xa;$$PP1<2~h&;rF6`(%sWWOobsrdV$^b{ky7+Eqy4>0*}{gZ7C-6woW(v}|T
z3&vbzP7|g>AS+{Tw%(lUC4q?TA*fplm%L*$nkXOc28wA#idN(Zi8o;9R^$m7dXA(R
zZ`xZ9&WqU2gHka2X78!rF0_{-iYcH99H7Q5Y`zWiyzu=zQotn4Nl42vb^AL1*5*C|
zlQcUi6_Cy4W|Kd*AMh4pqnLzP_*lT`qxCTyN{UK~=1_8EaukP3khBEZK<I!5@QIyi
zXIt*X21bArgZoG{OaTxdpa85G>VHtELiT+*=yzW9(#{DlCPgfRKw2qAH%=D>T?A}`
zo$;Z4c;Pn?3pnF9A)QFGNX^A@)~D&yxxgf+CeHx`#Dg6hVjb8&VEn-8dB$l2NWgR;
zssqu`e_{VzrW&ZIjmKFnMDjx91$m7+wCYjjPjl;p6f=F?bbHu=V&diTY7Wg(&T`~X
zs&8r_yntfTebYw}@E!Kii}teP5@w)2@DW%K8@bZ1V>u%O0fJGyX=KOz46&gO+mLA+
zvH(|r6!Geb^h9154g3hJ-<W3QcyHaSLW*LPjB?~06r*BPJfvpSJfvYXJQQReWDV>>
zF~R1+*03|hgqVj|L-$imsClS05DmptAypM}f)r6qHBwdcaSc+{@Nq3t)pA&$wjpuX
z=4%1i?W4&0C}O~Eh&Rr;f$8CTo%Q=huR8||Um3M8Qx5oNCOc@<MigT^!PX0cL}4~k
z6eBn2HVTusYZ=l4uLE!3Er)IoD6@=G;e_Q@*<tvQ(jaLpeEgoeOQJBRw;NMTf;J(X
zs~5di&qFh{GkGY<H|aIbGTAqohf;h~Mo{MZbjOIBx`ihA<qiqn+!DQmo`+^S%;cdY
zO_C>PnXF0Xp%hID52dP8Js`xmq}AyoD6?z3b5+eRNuzcogcLKh;TDC6Q0#RAN=yeK
zx@e#SirIoJwje1~3Vd!PVlP=pFvU!<nWEs1R)SnhkPoB`QiW4Abn(K@P&QaGbnAj*
zjn3%fhim`zGjxVh=>*M(J431BA+=J?LmH)qhk_(QGG1p$f@QqUkc7y1ogoR8@jJu4
z3cUjU8al(hnvZLcdkr7gBKKMje|m54hw-mHSo?o3fOWn1`d1vxZydD%V2Z)4<VL&k
zSjjNZKmlyH>2f5rv*AAOn|5T=jx>;^4NvE!thBJZh*z&hHmi{a@@v2Wt5Fo-@>4O}
zFBBEH;WxA*Nh`949<8o%?OW27s~wdGHfK&2=%yF-;p?FjQJviN07Dku--&jTK~ad<
zQ43CAU$AEF*R(cr)HpPRUz3M*J*)c50ij^<7Vfk=ZwtfB23vgIAG3LO;TKo&>OwTO
z5XqqKp@{HmPug>YrA6#C2dj7c1jKdM%&EXeF`5_;u7g^TrUiLGsK7hi8it7vR<}M_
z+$t>@-0{+eg8h^BZp{DYrx+aVc4XboQ8|p2_?#}Z3(pzDgbBxiCc;J@v<EL<hB<S@
z=&Or~y90l|{nE}1b%tXzG^s)myBrI=wEJ{hM@{~*&*R=Vk23lAFUddI`^}j%_n#aY
zwK`AB7WJ7fqJ;v%3l$=G;hGn`5Dmc#)o@mwzj|eGyanFA6G=Le4Yb#}R~?qjc-^l2
zrG@Frlj<4oZY(Jt)t{Sw=fa=vvOzCx*V~@LCu1z6Yp{PU$g%~=prRkFZ~V2!^<13=
zzHg*<BjiIda;cn$##@e;Lh)0Kz0{tkQ%Ds&q@}eI-~@{Cqx~c>iBpWfi9Zb$Ofdm;
zfCT0_iV36xCBTFfo=w%|$eH8&3Z$;!<BdqZ5ve)-CZyiP*H<ESC5H#AzdY>Gq}05$
zaAsv>rp?7)v?6wy^+_Htn~9+KC+FLDuIPE~Lj<=g%ufXL9SlGO+^TbJGAidrJ}pL2
xa==9+{|Cv=dMi87@U?F2(04n79LN31NWPN3`DTCG?*f}~W2J<Y)7Wr-{txY@^n?Ha

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_gemm.hpp.0A74F0F3AFBAC0AB.idx b/.cache/clangd/index/reference_gemm.hpp.0A74F0F3AFBAC0AB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9d3d50d607cbf09a6400d9afca340a0abbf470fe
GIT binary patch
literal 3096
zcmaJ?3s6&68ouWO0g}rlK!k7;0)g<5H$*^$X?awLuoeLqR|H3tP-wwoV9|z`4`ee`
zq$LI27Pnwgu?STLT0lkX6J1?k)wasE3%-gATLE>k*2<n-_zBa_keU4Vp8NgJbH4vS
z2`i$a_!L1zrmxtRqs&h;BnZL+|MT)yStk)ebQlrDwz`gl)Yvg!!L_*C#>I)dMr{NZ
z1%FJp_g=HQ$Sg~p9eG10dvw+D)z-N7lPkZk(3tEFlxzR`Bwp3?M74i+jnQ?{{%ft<
zhRnCAzyD<H)E^2rzUHu3eODDGc1?<)nzUuzN&VaDzJH!C<?S`jdU|qX$%>Z=`L&Hz
z72dr+7ep6o|C?bsrZS98Rt%iDcJo(JN&CZ(f+}+iF5JAmH1v#ve9e~wW&aA(^cUwQ
z>`QOhcbF4x{i>(qn@-|lRM&V_$`!{Kif0n_-|dze-<-Y|kX<%$JTRu`-O8Ksg~oA1
zVZ6>AhoxIvT<X4#j=lD$@+0N9_S`q{jaJAVIrU*m!%<IO{pl^5f+xLgK0{ZZkDCR%
zjlXs08C6oyEaI<MS9ASivXo4H&5uS?sueBoaV~d0eEY$nBV^Eq(7r{!rw)b3AKm>n
zum0fSPo1)|)+L`!;Et^K{82)?zccOLvd$tWU~BXBb9W!#;gxE?NdKunxMh=B*Ru_`
zla!>#=p#$x>F9v=aCh6JC9T~an?6X-u5Uefi2uvi!UxhZ`!|ouIj#>=A9scvxEJ}M
zcI5u{pfktE4@7!aoI>cbT|ANZUd~1}Nf3(~JumRmTaYDbvWQ@c_(ft?5GMGX@lX1L
z>SJTci(WoS*$_+!9b&->fE@)_-~1vm`~4X>ht9EJbOK=6@KBcWMASK&w6G=^xkT>4
z3V;JX8^5>z*V-L?(h$e4MEpu*H~RpUz(?ns3o`SM(WE){;4ks_U<JUz-M)cV$^Uky
zNj`?XrQWWr0C@GE&Au5r-G9lFG{yn)nfw4&0G#}+?y^JOS~;KOU|4Oc7PA6i<siMr
zJmU4AX_Cgyy`)~QtN^$;&233sd1WJqG{Dwkp4f^Nb8y!M$J9g%g(Yc<;c~<;N21vW
zsO&jA@@+(Q^9>$pfHyao4z^|mz#ldI@ZInw#*QW}aeN<<kCYVvf8PCJuWvwmxix8k
zVTp-^&kBHBq?661gTf-3G{dl;#Lt5j01r+}{HhMVu;h?NIK3cVkR>Yso@gsgH2Ka?
z#36YCg0VHW6|jOZC2`*qh(W)+HQgq?pWMVJx!7KV{4^+X_5nU#9ejGY@#3Wbl9X7n
zWFi=vOVfnu%tEEi$a~5m2;M;3mFG9Q_achPpfX&5aX?JOWeEuh3BNsI^76=6{-hz3
z5|R=M+J;Q7O|B5=2w$<Sj<QfI!1%L{rE{NOZ||C1+Rt6kB~Y29n8oyB5o|t}B{WdM
zV`_Wnq}B4BSV<A5NQE>gCLu0iZUwBzS<Fx#N&~KAYd6|$UI-oSAp|fUD>>870(~Aj
z@`KVc$;8!nIB65ekf}ygHKHMAM185_#J;~J;=Rta%k+Z;R|ZwTYwfk?1MD;=G$tB8
zQ_Skn)pM&v)8(gglziPeRU*?$WDZWD%Wy%u7K3t<(rVToYZzRH^=pw;EwToEXgjR0
z>p9R*{VF&+#YN_1Krv!xu^-ezF*y!7o^W&&qxMh-fd}PYBj@dwc@a2-t&6Qcc!I{o
z<;ZB+ygkzOF!;^ZZghQPIA$wWN)e~H>1$32Pgw~oP&VYUy$t~<w$lKlm|(YHUub5E
z*@et^AsgUt@$4MH)y@n7a8mYJsd{D8#mhK>Xlb-Ngrt~wY5d#@yc%%X(EyaCm=Ixz
zgWe?9BUcVvR5{YNB{6)oYQe%_wqPX}+GXAr>BbBcRx|8=Hms-dKX`kqkgy8L-~?-?
zHl*$}7vI9xn}wTY5E-vRH!EOOdqr{NMtO-Rwyr>e3S<u>qai!iFebkIE@r0|QCdAa
z8zTqpBF??|<aZWL<BEp$yN=$$K<1)M5!4Oj$6Ko#e~>mzKAfE^?K-cIbVG1XNxp2=
z#RD7mVWl8hkdJ<2HAtpGK9Fn2sJg?ZV)F9^XTz6)ZTkxc#uT`qDr8lqXZoq$TWecF
zo;+C4`i$z{Pbn|etFZMtW}O&{JKwZwvssi$+3z;vA5Mjuov$UDV?c`pS|o&1=?ir!
zN{AEJE*K38#f;VseIO%+$>~~-yn(H{GjpgwuCNcf_2LTT3+v&feUVegglNH74}6WD
z8k^g^k2BRE3k?#1^{BS>e+NAJJ~hRSWFxVIbEFu18~b@H2MY%qJ?2Gw381qn#+&vQ
zz$~N~9}^!MC`~a%PDP8M{V1l`sd(P1#ID2%FesknvX>j&SGb>TgI%EWChNM+_#BtY
z7OeESe9L)7>!q;YoFUYU2|9B_LPV7^J1qlWgKr$;-4#48U4x)$Vhinic3cnNf?&kK
zR3w17F!KoTl_6AqoUwfBF9>+hGD~d&&N@%EDKA=g=`v$sRIF1af51p=XDbvDwlg=e
RP73v}cKTP5{#C4h{Xfk0F<$@x

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_gemm_layernorm.hpp.9AA4C354D9D24BEC.idx b/.cache/clangd/index/reference_gemm_layernorm.hpp.9AA4C354D9D24BEC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9e7ec451dd5d5ec89f1735b61f425b61080520c8
GIT binary patch
literal 4930
zcmbtY3se(l7M_0y7{V|FNJ$WaOcDq*ARsCb#VQ~Ig(9LNTTwwlM4kyENVn7~Xoad<
zs#3*TR21c5rCL!GD2iI^W3jSG5nEBKMXJTd$EsWR&xA{|?Kz(1$T{5M=DWH7f8TeQ
z5gHt<v?hqaMWMNAiP`ZSf*>sMuR2?`<QXCeQxk&7EpLnJPO<aickJyC<Xs=X&)(;R
z`;wNNFwe1LWt9gKW=}ZS)>pJU!jpM9ul8Q?_}C*GCkc+{W?!_u@R?M#;<wDXt1G9P
z^S&DJ<2<#D+_0wWo0G}Y!j-2Q#a_Si<)&W!o)1cy^5s<)rjKUr3=rQhJoCUt%!${8
zX}kvNV>J7`7aizV_twW`2fwppnySn{Qu@uFzhC`wMEJiV+b2nXIoteO{^N)Xnt{cO
z`hSivEB(Fy?Z@Vu?;grZA<Dg9^t6<SOTOytlI53fx;VM*n>ABhTkhno{$l^bv-ivV
zT#WN#N{~vmvZCjivv0+nMSC8Ef4u(ui-?7_S*W^p%V#f5XUvS=RkKmCE9R$OE2ht*
z;&kno{;8wG`V`Xa>Jz8m&93sB_r(nFUyV*3D^8Db?)*5tSxkhIZl4RPZ_byCg3rnK
zZe0@m)N|dq;`W-93qKI!6RGsLx=+?EyTfZraXLPF-3LF|3ELkneWR`Q>2yJX?D>)W
z`0t(X3co(ELz=VeQNY%Q&f1N0riTfGnYi4lmf{I17yOkbBX3<ci_Cj3_h!YXXA>HO
zLn~dx4Nh-t{ccadxqI%n#-Hqp&i}OAx#CXyweG^hFT0IOx2;yDmzZC1zgzzMt_39;
z(fak%uGm^9-|R{6t8SP5k$fwN`jM$BXlOl<@zl8Y*tjTlURpwk5kVmTuQkgeUkFS{
zb31|&ju47jLYR^WCktZG2pjhZaaH9xniPo%MlDghvxMN{G{QfUoL(nNxsgMYBQcz5
zm&veX2!8pbZ$|v&f77HThSQmJIZFVR*`MyIJwK(;iZrz+m}7{44A~DpKxEv-v4xwn
zyY|zh4c3q?$#!E2z`pjfnY({1%d#SQSVITmcObjL2Z;DJO~0UA{P9v7(g<sa=0@{b
z0&r;0!$ZtR^<^|^i{U(nJOxVtj&t=QN0^<;vm&|pfO<5t9*rA(fXKXRZbn$#qaSHf
zfHh<=8FH2Yoamnv6qfnGm?ni7&XQ!gu>{~$^R~w2zUoPQl6D}NRft}NEC(MTl6CW=
zc@c-&lWB4k){rmCcVr2`YZlh$2-D`+^GRcT#cPOv4T%RIAhNc~;aco~Z@3vrVG&!N
z?I@N2T$*gTa?0M-Mkb^Qh6~IJtXKkYMXIr;a&=v~6={KC314E*5`Y_TkI8#p*c8Mk
z%`n`FygE_f-~&XyeW*yAR9Cu3XJj>^tFe)~2Z&re_|C4S_cfPkl8=qdamZ1y1mKR$
zWmLNSRs~I3V|cm4as^8O?%w8dDNQQcK$BJ&&UMW7U<trIn|)*E$>bUn(g<HM$}Gy7
zB>+F`zI;6P#Nk<das-AOkZA+XL)`;JUc7Z?+3YEAWzggs_yCne<<1g-Rf$PzS95(m
zAs$;tf9%xk+1++}oROE=z8&WiN=S-P2o?6=P>OLAx(#tqYFKCyy~q~wmd6$xSYRQ#
z65#p$^`%A}CMY~80yJ@$(2&qjJ+}o-*@6Owa_gODCar&!6j4kOaw$T7L%>moH1ACz
zmsmhhgyaB-IuCMR>^(Au!(sf!_<8FU2PuOl=($keP(M9ahlF)V1hJO9b?E3r-+%I6
zN5x=w!5)dH;{|$;Xd4Z9L^9Z8u+V&ZzJZXo-GC5u8XOm0n}Npn=8ejnfm73H>;V_9
z3$5#F1mS6W@XGAC|7aZ(#19gHjh?^$B(th^diJM}y?*|@&Tc_l1`f(Z`9v3x<}m&;
ze<wW`=ohF2Zmj<)-qpOn2{<?-xC$CR#u>qNGvK@_Z!10K!}qZP4jLmiJCRNl>3yT^
zkzh8oM{KsPH-g!|P6<T3#r$O?hY55Jl!H>7R4zfT!1)Gj3Ha(6V=)Y~;2QWqZKs8M
zC>;jQEj5WBKpf@<qHZ7>dKPqFY{e3Ki5+~M$M#_Oc7)F-88+@E8?n%+NuxrcnNiFJ
z6tn?FLe55;PCb#9$X}0xa}AkZLl$tz?D$QwH5J`T4BST6w~-A1bDo!Z{oJ@}?68Vy
z?!;}@JB{zA55p@kM6g$M69mE#cf>DsuCq*PL@`lLQC?s$g?po8HFASeqHr@(G$0@7
zQ!sI`Ll)2r3{qY@d_W1z9W0bUCm1A%b_5Lxtv7c1z)Br}T_f}mPJnxXOYwzu6A30n
zYU^ItZQCEc`^vn5d4*3-p;K)1LDh~fI9VOOWyL_wFDiU;giFL2$T<iNK3O-(;7Sm6
zlE;oN_?TkWBFD989NhRHm+$_zvN6Bzl@|nSI6QXF>2-A=c?TK|RYNgd$gvB_VZdp5
z*DmA<{H7akwoGo?-G@!Mjzrgy3@*Rv$%42KE$ln618R_{2Fdh+tU(iNPzdmwzYW>@
zebM`FSVI>QbfHn8x_rB$_r2C-KEujP@F8}Qc1Y{Zsz6i)5`l{<Qs3r;MHThpV(dj!
zuf7;}zF)lIJnb{}mA=~@*ewodeYWi{SFZI|PH)@3U<)?7Kvv)ZKBn+0AnQaPVA<h&
zD-H|SJ>35qW8v%Z^=0|({UO+Z>&W`LzMkO>oIJWU1YB_ZY)5SEgx_z!#-YusoA>9w
zSTP+(#Lv!848G<tN|AD;p4*O`w(G}etNeX0pF?4luk`52s_E(G!Q9n<@#vY$?caUZ
zr)e2h5<qHUALte=2*gEc{@M>GMO)xwdXb<Pi6C2H^TL5kqLU1IuyzT9a=kRs`MnvD
z_M2bBgRsKka~9K!4bIVSb8ya8P5F%EvUO{QWixDlSaUL+EY>$1?IH-8dyJ39ERVdT
zO<co~4zMwgT?rv$sQ;N?6mEl}Rmc%ah+^uHq7He3vnl2bvN(f;ux8xwy%hMQMuc9&
zTem%pH6tv9#rV3N$g&gJL2}<Kt$Vy(acPH`VkBaT6jG03q*Ce7=%yH%RA#_ANu3P1
z(cIAnJBA0>!+`VTdKz%i!ssD)Qp`-@OapF~<t!n%nqp=PXB%KK!WaW?j&P0v7b}c4
z;L_=I0jzKolR;+)V7gIECO4CYH%yAjqO$}r)F?(ps|2u=Q;eEc3t)eym~1**0Q(We
z6e6k+S?NP}3`vh6R|plwG$LuEA%7f+k0YrbKY^qt4Dlu;Z8GGWk+j*6Z$Z))JwH%i
z5Y4^nWryQvE3x%}G4`<hXtT$cjO;Zs<Z)4GgX~=WBU~!~@7DR$JL_af-Ra50YGSZ;
z;EtkuCxRMz?on7=<{5tAPZb4G!PU5sUTE+hqkU_FawAlUsqsm8-43cT>lHjGh(S<G
zZ|9F{T6tkg7J_qONh098iGxea(0?zIM2TCIraoDNpryeqBB1%HRSVS~+U4F+OLRX{
Z_=6ys@YZck!jg0SMiQwbAsGUQ{{kVp?yCR*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_groupnorm.hpp.368D3290BA70398F.idx b/.cache/clangd/index/reference_groupnorm.hpp.368D3290BA70398F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6fcc433bc031a43ea269a6e879a7b8bb8876fdbf
GIT binary patch
literal 4944
zcmdT{X;c$g7Oq!>)uIw01%UtwBoPtXKv01YQCV!nMo~~i#65_uxC9p#byP%@AWXX;
zve;_7wmTroC?bl8xNEDpjoYZSps4LOYLD|0yjITfoSu{8ui+f-tDAe@t^4l#-uG?|
z>EEwkgdIVIPUxSOqMkXHLlA@w{!g8$nM9BTAu}V0w9=Z`_Q^MWcvpU2!M%N;=ZJD|
zlNhVko*}`jRww1XUU{%FE+MRN@yyxwk)Pi7m*kg;#{a#1;nAx#hvu}Go!!@}eLT3I
zFvqN`^T&4-=g$474>i+YsGcowjAhyuwkx08PZE{PY<}SS{D_JDg}W|`)|Mpo{r9M=
zm8aiN@QSm0Rwv468{K_aa`G;6{;ZaTt9r=&+uk={e0aBV?Wn`;zc|cS*#zx6vSzyY
z@Sr)7e^`HL(JHgAQ=UV;8@*!oHXq%6p(1_wZ|C9%6}2lK-wR4!y>(dRr!CdyqJ-!(
zwOPzF+sm6}PipQYsE0f=3EuF%WX7YEr+KfL%x%^Q4Kpk{$Hn(cUiA7{)sGcz+<TYI
z-gip7)t0yD>UGD>3GMrA6Qgf9)va6Ake52@hE=+wIA^`nvo@z1!X(2k_O4>!b+h3=
z^7q|Z|EMV<@p0qBT#?n?>>cU(%U2wnw^YV)E<9IuGit|X^V;9%q=X*X@0mZdaoaCh
z+S$~nh={02uN%+1|GBI1)3yHm<P}e|)?KxEE0b?J>^^x`pMN+luRc(i<}&x>pGk9t
z+=8OqQr8U`rr~KlLbDXg>o%Kqw#-xB8<OSuE?F@*sC0JrEPCA2x)WScK+Ca=aP^7Y
z%HKHyV#YZ|^;Pq{>u<|eO$u8-H*?4Ru6Au($CS+6n-KjX&bH0rc}Q57)bx~bjTi@t
zrNvQaBSx&WB`tUavj_3^Am`5mVS(|2|H6+d|0Rvx$>)(A>{5?-^~mM(04_fyQ;Y93
zq)g<IX4qv5vfP5BbPV7!=zQiCQDfX~nsl=v7=N+9yIz1fT6Nm;L-VQqG%3O6Uar05
zdI9E`smic(*5^EE(ixir!~vc40?e_A?c5ty3&LzkbNv4G$Z0+5@p%B3M4@Ql8EJ{5
z8QF<TFplPqj(R~@k_fL6F=$K{*YDs(brO%X!ZB)*Z!HS_JV1<z-CKVMj_#h$BPm;g
z38%wt^a9K|#r6&UrS*4c(j6zM6e?Zx0?bAEHcu8zRCeo_y%y1Hkqy%^fXl{9JBF5A
zJd;h6t~f%lOR!8Yz`T9<&EAW5M!C|YH8ux|17&&v=3j5sIWK4*;bu!(;uX{&UJVj{
z9>Ar_&F!q9RJ(>I#W+Gxmmq~+fcdChTSWfi=}XK=6C6R@No=PVU_O~Xwqirp{>wBe
z!n0LfR6X<p%-6l&EqI#Ve94S7#j`^@h1%)`nA_z&-@i#`GVDncJloyEou?OIe(g5G
z?Lpo=cbc@r=H7zdV!Z(Ke>T3_vgt%a50Z4D^{^xuo)wQKEISm9Iw^ILH9?qYGJ^t-
zKkAJrCY~QJgdL-@dX;=!H*5|`F`6(<#Fw9$JQR|L`hm_B{1)PK`Ja<bIE*Gn6AM3c
zuv=^%8fvJI&O>8B7hKzC@p0<`1)et1ccKD*<}g~+M~nKhdMR=!MFIeVd+*BXmjM^7
zjGWeO(1h<5JLYH+huMxK+mQ@TpYNk}yM<#;Oyh7EwL<O3zBa`(r7Npx0yKfFrbUCb
zXb9*)2gvF(3Ba)lkc}h8a!1d?xlT|{2m*f&lN6MsVl^!a*Rtgd2w4+S@!pO54GAe-
zUo=fN8~Xw;_7tK{AsW~SMx1E}B|)Z84Q%LOATR&}VGMZDUOa%Km7a$H`M^Wi$(QyO
zu}(Tl0;d?u@R}4gFCT-f9$D5SE+hjqg=c%y-h6hpj<H~N>_A(OmY-CnCKTi8?CB0p
z6ccP6YzGWLF;%Eb73u~X0t|;&)j@Bts$pYS$M4*5Q;Bmqi|o!Kd&p(@ha%rchqAgG
ziRF>161$D2EmOZ?&mc9NtBz;k<R%^noHDYrRD#znmWh3#92Dk2@mk~!r-H)U=CuX+
z!#)Bp<7geE4B^%Jrm_XLsUdjfj$V%b@H2oB`|B7QNNLK-aUNm!BIk6Xm^gXd7XVYt
zaQScpZM<~64@97tY4T|Xwp2xGFl);~(kx`)n~kK|hI$T?<{0YBkaU@$o{N;Z>`|eZ
z0u)ey!XW;XV_Q8-%3dcJb>83t!tm=xcQZh^?k<9QYa>k-KP~tapo5>pv#SC8_`yO$
z_{pnKPk>+E%j&xJP|?G$IA?iF`HYyptp)!lW&P<=%kzbYGItwQ*46jA9_@2v(-RD|
z1|(=eQaEC(-;POK#C5)c>s^cl#YoCxxEQI5Q3U9PuH&uim(6&L9slb3DLfUi_41PW
z-uOjzh+l_9fPILAG2PKH0IEyNyzb9EHp}10L%y!H?m~rgQ|kZ7XTy{MoXEjT?&Do=
zLR1q=z8mLj3(nB~;YO7au!Uc$pcVE2JS%;R1L)iodebH3*!z;NN&vkNp8XyDogF({
z*9O6C=zwsYI@%u~g<XEi=o4jQ*5KKp(oi{+kHVB9twnOE^ZwS1{SLV=f5q!E<|2o#
z3P+dIasiv0uEB!bs(O|eWnOHHz{PApb`301A8x!7zbojEhep1%R`cN5$d4J(B);uE
z)m}gxIP)$Ozuflla&s$vBM0T6^qm(NDJEPHE`|UUQ;NJwksn}=VzwcA8?pyJ*6H21
zAr-4vBW^Wv0IE3VG2hp}GIFy~*f_o|da89li^X4UJG7j*v_7<tkS(pQ`-BCYymayO
z=_jv_;#V5ewlxXU?pvnOrg%pT0eMc6xpYx*3>W9pg!oNJ0LT5js`K0-gPaSE+^FX4
zYXY-bPFKI;&gNI8Lz0VfbBs(Iz{|1Hx7X~R)*z1>)D6ynjzh$C$OEvg(>?2u59s#-
z&WFC;DL_W9=An=2D|4=c7FVYh+0-H-#A{u-`%RhG`Jee1Tn_Fouu~KxagZ2jQU|Gl
zCbN+_u%4Z5IvZ#nHXa6=Kkd(l4vu01=m0)6TND#W2l9a*C`M_aq~Ru{m>`QF8qO!h
zs4P@8K#F3bs3=<&_5-Pb2HGAZ*@HX*)D*K9N%k7*`;cUxp<a$$%8`WiuRtyphI%D(
zsbqCnGbSqChY$eyWodP-e6O3-$UT8iCwxf!Xg>pH)qO=l>^DpLACxwgzy5~C?w#L^
z;}N?X`urv3Xw*~2CgG2=u$`Um^V|Oxhk!^1WQvX~Jll5$0_K2*kbo4Sqay&*Ndk3o
zQux~(1V<2foPa}+su`a;TK5TP-|-!&lS1~<fqi7=J3e6kta==0E^u@b^6ec6N7l$^
LAD!4o0sHu0Q<QBk

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_groupnorm_bwd.hpp.C24505C3C5BFB298.idx b/.cache/clangd/index/reference_groupnorm_bwd.hpp.C24505C3C5BFB298.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6720ddc84188374dddc355634ae30bbcdd64740d
GIT binary patch
literal 6272
zcmd^DX;>3i7oIaLvJ8^|sTpMv31~oN5kWvugrFcIyOnAesPWT+wSppQ)hd-Ld}?W+
zSOpXbky=2iv=$UW6tylN0)mKCh1S-J8&->TsozcDnCa6W{_rO};oNzjbLY&RbKkSv
zi42}LEtR9Ffce3xi{;DWbSa85CI89Gk`{&26xFRyQK@+~(TfwVxCq;Rd!A9}-M)KE
zTlx8Phm8w1MlLLzAWz(NBDVUB?EyPGudf`Q-cuYr;T(F_My`_<*4OFVmv=WYsk$$9
zdqYRzdb8=`DNZ)^YrOyN7;(qfdim)-xc{||Q{mV+%6i3hlMPklGh;<*n-uRf+<6u(
z$d9+&uxs`8@RnTmuMa}r3{32DOL{Uh^J&qYyQ_|_uP9h|c&~M8Z)DqU#nBg4&0!CB
zWhFa*V-{+c%ULvzT_v{*9)H`?x60_?8Lu8mhQ|ZPk9Epq0mTz!^7G>!%dDfGnVkF9
zB$2h)Q?Mo>XHopU``|l#Yf;~&@S=%rxe}LnyPC3j$xlA7afn|wvM74roan!{U-_ci
z`>n>zKl2lIwwPrX3Y#um_wua`OAqcC8dUYo&5SJf?6HO}C8s*84qZ|x6tfy`sBSHD
zIQ_QVVAp?QZJ9Q=o(ucE&mRzXZZzq2zJ4$#<?Q8oy&qog@lwvu+WS-77jCxm55BYR
z+^%z?vpF$KN{1A?hOQs_z3eSrq444c@58djx=+f}j3$_zx=dA_d^9df{zJKJ$FaQ0
ztwHh~nROexe^~piCB4=7x0Etf%-v&`p6|2Bz1FFFrMc<$^dW~1pL$UIczXQ4lFAJC
z#p#u8f4p#X-K<*lJUFz+_K{vr-UwO3^19TIQe`#YWz9aeqVwKX)$fIdQJY&N+52}J
z>|Zd)CM>^FH#GXznxvUk`jX$fdLRFs+H60YZf{NtIW7=YR!)9Cdq(q}waF2(uk?%~
zTqMJq7S10_OXN?y_WRtq)N65|IlUoy<>L6$gjdEgpHV*wZd?@7x<ZO;2Vpx{4Gfe4
z;UD>9+Pa00o>kZmrFBdx&eOnCs1c|(395?^mn`ztr}c=1SmRhTjX<?v&1P;%N!vA+
z9!l;GwGADk5vU$EwII8ovF<cWONct$F?_s6pxS9%+97T;6>LTu5aUNccmym52Atyj
zShqwmvowdLZHa{#s~CHYK=t@v3jD5h2Z)5Ui5bPsW#^h|1ghWfY|$I;+G4FwGXjcp
z8{%fE5tJbf#uNn{e3YA2;Qv7Rl%++)e56gJqekHU+h!&}`=(*(`m_#Npog)CxkjM+
z^CuZ!>%+xrmL5s&j&O`{*9cUrimi*ENAD<MX>+2E8WAPc2vnB_>31x7Yrlbz)+ffR
z!LXXVegE)s%E~iGO1>$+W=0PoQz~J2CAbd^IHmT3abMiNZIH&&cH|+mM$U532vpaz
zUoLGeJDJ+=pfX^~z?ADBa7uG+*&AEdSl6-i8^l7iWwgCUp!%1_br&w@EKC#9Mq~{w
z;MxKK0|QREZ$9mv|AvfAV(AgYLYRG+lSZKW?>!Gy%7&BfG;PE7la3P_30capAN>7w
z-iuB7llPd>hU5WFAZ!AQfdMTf$rmKYnQ8e9MguIy4Mr9dg_^>+Z^&L<BMoa`&L-wZ
zH-QsAnc-T&rWHoxtE#a>E4UyQa2NeuTCOJ0x||B4R1l4{Q{%ZRn1>j{fo8a=>{KzX
z%19FwxG3(@w+7Gl9vUyu<x<9_IN`knO-OOqA_X`oz>$w#6);|lv%w*ouP*{=5jbe|
zMc`fp9(?Rm1ipM6Py{k9R)VJzy!cqAgg`A;flT#JJW~aseEpj$n4`rNAg%xo`G=W+
zct*y`hH5>dP73H5+bNY*4?Hgz4JyxI5+u~^&w$8MOnt;<!E7+kW$G-5E_400pxolU
zh9!iWMKGfXLbMy8gb*dnM2sVOj5<wXWO&-gQ*nZKm+KPxd8K&cEVVV~B5=`SCAcg3
zbW*|;CHSG<(?8e$;FFPpS1!Zby7x?ieq1l;az6p{6R=pKq6`CT7R!ZOfgz!7*bV|z
zM9(rCeKX5PPlw_BEd5A*@2^)37h@7*jy__z6CgbSPUy_paXGQ2g;%_ZwJtF4f}v<_
z&eQF#zaCDT@XAo0=d;s%kKWA>U-?FFAM#3^H_A>z*5_gG;ffo>kSuFc24gV*7?Rq@
zRKrBf`nbzvKh$*zbEJH$p8Y3g$^K;FZcc9AI4e;RKU&r`V}*d>mPnUOK({m8Qt472
zN|q+`&~oW=9$M+Va+3DGRnDt;Xd^gp1RhQY=X5^a0@5wOv$YMR+xR#Oq*;8N2huz~
zE&yo(AMXa~Za!9lRHem<^?Mz5=0BSInw-SiPU2@8JB>r)7-=dzZd2wAoloy<d*P{O
zm&DHzClPjuVl4RVay|hb$gV9n4=jzE+Ous?>;`&D&q&kzAS0Q76k3Lx%gwdX(yIc-
zRDhe7`zJwg62#~~?3>8W1hRp{(CJ2+?g^il)qL~*`opWsUy}tnU5<5rN0qloKy?85
z*<mKg>3i?(4FbznvWRjZKRfMe%fY7{f)IcCW6<7~^tEG&g;o%?!U${&^76-Ye_X$O
z;wy7WVPc5xjEMAZlJu?ub5)!4_N>j^c8;AG_?qvi47g(DU;kw1zgCDHD0#=}>~vl~
zqHadf(U_4)ChG4#af?d_x=*6$Z#_{_=H?yVkpEsTF&k?a>x{L8A#I^u3pk^HzUx`_
zojBvczCl&azxOz>eRb3AAhOpTVBVpvJV#Eq#+FU`<Hn%!2A)m`Q~X(fUJs)_s2~Ar
zlYYKC@7bzwbrg(F;kB?UMy-%yPSNEQAW?w57H5MbTZ=JT30eD(H^|ER(Q92NR?aAX
z%@{jg68@pdG=oq76=SEGo6erS^|0=hS>TLCYPVr3t4El@@8|v}S-ok)<R}3#d=*4j
zfx`sV=rii$$S}(<2NhxPFiHp$#0K#)mik<ZglT`67rAhI#%Hf|d6#EnKwl{bgUW@N
zen}+spVl#L-*C};I7A@MBur}o(-yG6nLUc~Kl8^l?&3&M7kgS^J!UvBu@?_{i@kYh
zl6Vpi`G|dZ$k){O<<>FWWYft!G{tlZ4~4R!B0M!S+$?sM2vY{bg|T5G{ItMu;cU1F
zPmv54!A6L%jbOM)gGd(JJ%)=ih+?twF<i7kG>fca$Pr6&1RSstVYnKQ)bR09kR0XX
zT9DN8aUDqN`1lxD9RrEhUOiaV^YL-8I<Cd|HiX>)Y`{x57->G2AiE#rRn<CIjI<k6
zr2R8J_y@Z)2Ohohxb@g=>G7cQlYXyNEbzk*36cKqO&B6nk~}eP0r@T^D>d#G-COr2
zU|k7WJ>uy4^HY}swo7>Kqj2Bw=tiNj1)O}xg22Ad05dxFO%&Fq<fQkL-&KDF8`l3R
rOf9Xom)zaX$G^U)Hn=v#!qUoGJZv~MT&omoFD<o~R@zGo?dAUgcX{W&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_image_to_column.hpp.14F89CCD7031E508.idx b/.cache/clangd/index/reference_image_to_column.hpp.14F89CCD7031E508.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b59f0ac40fe8d868f2c1631e3ed055f0308fd10
GIT binary patch
literal 9402
zcmeI1`Cn8=w!r%oEwWr1gr+ajv_iA*&>$elVmG2<(h7`;hzk-yji3fFqVhBb#fU->
zqJnV&R6r9MHGqH%B5D+!XGD#T<|TtGNiY}}FwvKpqVZT)=O5q){d}z3UrycH&Z%=w
z)r}n=6?Mo#B+|?tpOu!JIZG@OiH6Ys8JX$xDu_r_HdrLe+EzC?ZT?Lkv!44e@}7V9
zv~`jFjo*sL>m`u1vhz_1-yc&-pPR2gTU2-_wfsqQcUMrs^ZL;fS`O_wayj(kcJ2M@
z!QpeXw`{x?6-XB6*&2U!AVLvY7}feikomenYM(;u$+D)Oer<`osfvhrFn(wG#O<}K
zDjOd@yzQSmkK1K+MwRPleeDrHTWK&UDXs0h?v^Umi}P@FjcfkF@@w%=n*w~ZtGWZN
zXU^O3+sNG!pIU0AK8o|V+Rc}iCr*|5WhUPHmwddFtN+ARfBd!J{O-K&qdm2yt$%Eu
zGr!_n@shgw>#JvYmQP)>!KM29eTE0-%S?I}NQyH{9b)G^Tf5fl+D)sT1qDSBl~bkH
z@BZzpf{=ZiC+uH(_;RMf50YIi7L$BO<+g9C7b#|ayJYMyn#~UW+g<OZhjluw%pz+K
zG_SlLq3^uNruBl=h<84`(b{<|;fu+QW2&s&?r&)sf9x}-Wyza6E7mo8>>uHk`2AGv
z3%-1IfNIUAeO{L@PX5cc@c+jiJoWv7si}&^_jcxA*_4*>WU+P4g=(+v&WbBBo0{JJ
zDE)&=S8c~XS{)P^_G0WGL09|&({3-n*W+Y-b-}LpN6k;~{3GST&!SHCqq$y|P8$yX
zAU&Jzt?Ia&7ynqU;HOsCZP~FW)8Ndt`xC=PxH&D}*T%&+&F(EdxW+R+*S$39afe^W
z!JP@u+veui%;UXJ#HHU`ov`Gi!^*?|2wZu>!O<yX@&8R+>T~qxs=sThxY%<CQYQQQ
zZ<sN3eB9*V(jQ0r2Ka3GX3W-M;-;Qqtt}lxGR}zKTXy(y&%I8)&<&YI&w{0=ji){1
zmsc;dFx8&@m%H~B&#cGB9Y#-;cQV8Gor;PrP8`-?Idt9JyLXg7c_v?fVX##{>-aw>
z7#00~_DWOFb5V!)PyakMBV%b=l98TBG{|rE5mo!Kf18oPW+HwEG2B6HUOq$y)S~G>
z>zkj&WNmA{-d`n$80OF_{SU0-_^x&B$B&e|CB$Ngh!0bQdFTohJEdml>}`l^G$SVF
zBEAbGT`=V31FN`-J^Vj3*g3?U7}9a<B4)d2C;bnsqAKIUEN@8<n-M$ONt8*Hqpm=4
zoJWpj+U>h#5@Jim5sC<RU4h~mPxVekoZI!jF%i=-BuElux&p;>kEXsznbuI(Kg8*h
z=|lLw0>$$>D%NN1pQx6Q;k5B^MYx-;K=Hy03-x8y?e9s5l!`+Yp>DbY#Y-psRC?)S
zjfp<dr<02}jyKm8D9&DA68+<@#Ab71M8&nltd_`MKCsG)h-sz~rg;(xQP2+53bntk
zK=GzkoU-dg5M6(JDh^SEIOqx#mxOH9YW<X;PsFqXM`K3|U4h~~f6V-_;py^XGct&Z
zj}WsXMDg;0RrY3|)l|N-##lzgv;!9d7c*Uf;(8aa&sMA|_&`D&s92-W`0EN3pYCd&
z^1x%EwHYy{9n=%wdZKyxz$!N(UcV^y=SSwmfL3`7){kMt%Li8J-qO8K^ZD8a2~pBc
zf|bE8x&p<I-@ULn*Wd=HPdHN%uN<sY=n9dc9s!XEc=)U0N#*X-XGRb*Oww=JyqS@i
zL}b_x>B%V>gNLyl0Y)$P0XA%W>Cm}nmzxUoI6eo=a=;C3hU06%tVYPU!R$6;Q>!a#
zpUXeaHx%<RQ88oDjw!{*Ol1`38t29+%_GgTR&QVSpP5}m%;$q|KB%@mTz3A7di|6>
z4#|g6j2j`K5dvqu_#~-3-XK=YH$qe+j1}^vMu<VSx^~duMb%-IwU|$HPIF=1q-)Y6
z8D*$5G>r1WIUihDm-*n451v8}$cI42MW88yNXEsWE(Q(b5*S+o;~1AhR4I&Q+z8H%
z;DST$o03N(cnaCK5mZ7R)d-<N9^D9PAx~(8cNias$m1}EaWkl!LBqHe#<s#Z#-||a
zlyFJU!I*Op#pG=e*(PkS8~ARpLf0qe^N1vmn4&YKDkPG8A(s*RGU9;#80%3h?02<j
zg~KJQMy@ls&_OspNuK1wY_JpfPH;fBYf7Kn^~tDQYCB1?q@nm{F|A|X36{vp@Y=>l
zLEjm@wS)m{2^d?c+fhkW3iQNAx(JJ&o#VCdYNw(1K%=9^(;sXx0AZA+g8tZwR^>VS
zMTmm@&&fezex}b%6`F^bPxMXnV^m073yC}0rkLMITsI22h`1Ip_Pw{GMxJ<jA&B`{
zpIBcU9X_cA2XcW9Y(jk@9VZmzm?WBII&*Q~yUwvL_)d^|i~5^C?M&ZjGI-eUZS(bK
z{Z9XUf^a8@1Vfs-Y^Cz|Ej!|LK6_jaOvAbK`RsRUAd&GMu)6~zkRywqbjGAx{0JQX
z4E&zKNDNS8qz{IaTj(vLf?k-?3)4{$b#%>dsa+{M4LN@4h@~Oe5XYCn$TA3LTmd61
zgnSV~FM@{2+aa`F$X6lss*npwXd%&{5;&@a(2YVaBB4c$;|#|fOG`E@pyNG4437{~
zTpG-m=+gL0{B4*T`cfxU1INc&#@gfR|4nP|+_`boATy44lexWmh2!02?gH|Vc?if;
z<|!aAnU{b<O+qa&cjEXklQ0Xkc8(7>3Aezsi{m3qA}nyk9IrM}TVP7W@fs721!f=|
zzlSLI5Ery`j;|!jN+DMfWtEVtiLzSAdx>%{<M(E5ivPO&W-wj8Q^evFk>T=9IaK#u
zXi0Rp&JO%~ArS3g%JUN6dq;Bo2TZh|HtI`@ib0(V-q<tK506Zp=Hy^c_3!BD=*u+T
z1-4z_fj)$zS&;h<5{Q9m`t8+I>QWXK(cPOzq<O@a8Dc(><}=2GThB<BXrl|__~};D
zZJD)qgL^mlVQ<OOx~yS~|E#2g>IIu#aKu&Z<I%kk&G;GEJY(|Yf<ca-m}jV|I+WhZ
z5cF%<4|RBb={yc%p8ZLwt%;<{oa56~>0y}UaD0X;LqM6TOaT>vstCd`LE-pfP!$XL
z3s8L_<Pr!d0X3^%3IU};-U<O*g?t=T#~}<83XX3ERkM&!fa-*hTOpv8B{<YRA>fpd
zPeZ_IA)f>Pa}dGWZv+1}A)g2T^NexpQBC+s{N&6%^c{YfvBZpvnT+lswK&N+vLkv)
zp9i#G>Wyp9@!9HZEu%7UE`wJ=h2ty0xkAX5;86)dto}uCzX*Yh+rhnE$XCJrs*rDh
z&kY!b^;1t)xqi9*X~G+U1icQOp0C7LFulIR8=GF7Rh%PE4~<7`q&vN#iE;F(=3Gm>
z@D#`K4aByAIOCzGk6jyx4|4X(1<`JKsxWG)F*Y$y*gC#WEw%pyi0z}Fqaz(58R5tj
z(RT`DBh|WV)!5dG+1Em@++Uqe_vm_}T~DU4Jz7At1&r56>#ct9`M-mFsYADObHTQ$
zL$`AmkcXXzfIRIy1>|MtC7@7gs2rD;I&^86T;R~9;c~Pp>d>VTa)Cpas^vIQ>d>Vc
zIdkaI(r7u(jOIDkT02I`=E<*eTaHgLPZ7`@^Em>VFPpDGo8b5~X_}nv`UTPj0$OCX
zXgHI7Z1b@_qjex#2ZDC;K$a)ue30gYobCAnkQNBJ5UdJeIFoM%o6TU)cpKPm182sy
zAgdL&Uk9=}A=iVfUdW9gZ3H>n<4quK67o^7Its&?yajApz@G6Luss9Lj4y!01#oA)
zhM2D*Lzy|{67yUkuO;Sdg}jbP))7lqKabeu5f{cAh-8CMo=@!Zh4MYba}V(s26N9!
zAy*O4Dj`=B&uSs>C7yd3=XW0ZC~owxpVO`Mx=VV~<-G1fHnw*@Ty<o5%^NNRFJ0&^
z^O5*i;^r~ZUA^GO+|=QABI5Y&U+8W>My!qz2h1Nf^(IdL)YRrGeft4eJpcy`fqfkO
z0HTmLpBq<oId_EzZJ~ik8^~~UB-n+%k?uAJ{ROAHnL#1PC-^3e#8^ykE{J^&%g||Z
zN&?GBI0ko)|B9G?MR;8BZQI>%WuIIeIAC&_^3)09=$PDFR240RYhj6g#ftTxUyuZj
zc|%9LH8&ONlYf5sUyFxopzgN+!V8zL`r)|p8994zP#ejCh#ZJwPNN1QYap8O9T<8C
zh9jf?X^VaLW2k`re!-JYZ%^Dp^&IaI?~Or(qxWA9^~4*MS@-L*IzI2wzPAVQ0ecX~
zh(wHZmz$_yj-O;U$r(dZeZZlT<(Im~Q9C?FERM1B!qKx=<M)R=zWY|e16IK&omcn<
z)<*An!-uq1Pg-CSWw83cd<gDcy5sxrw9$ur+kWZ%`Cp!#7%*<LO`dhjFiSFsUhfFA
zUGZx-O5)cNeer9z0^-+l&;PYw{@JHtai(-dT8OlTSfMw)ILawIcB0MZw?;l-jr`KR
zlM9=|uiU4mpjB%%=(i|M(@tj;GdboxMzQv>4vgX)<6a$Gu0PopJGUfn?SL_Lyda>P
zvi~a*Zpx0Ji<;iAt>CR~4p`f~>oe%dU{iiQ^(*ycNIj9GhyAs<rf0kR#hv;zx#eA6
z-GOtw>kwBS6EBVrmIO=jD<Q{+7=%bLN9X8$xBV{Siis`9?<R`f#2y*nqFJA}#9N9@
z8tJ|;vPd6e8T0BW`e<8Yv-bfzhqv}LVC@OFKF!$rz6aw6V0tDRn&6FRqMqFG4GwLG
zp7uq3!B+}VdUEQl6#8j6@@wN;((aY-10KjGEFXS+)jvjm1bAw~n^X}R3SLHvZ~!S*
z$y-y#^sWUAlz1u-;dvk<{ez5|v^T=v^0n1|^Y;lYgRO1kmNKiMBAfp2(|x6k{j_91
Q<?QEB_S1^}v}Qm512y$IYybcN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_layernorm.hpp.B9C72CF73E810635.idx b/.cache/clangd/index/reference_layernorm.hpp.B9C72CF73E810635.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f3f3e04ba3ab0d05067ab685d9c28ef0de1a4092
GIT binary patch
literal 6506
zcmds6X;c)~5}s-Xgt6@z)<%Y5nHhGFT~J1L34#hw^EArlHb!Le0g@0gMghgRKq7HL
zA%;aXPlF34ijioH5`qdYQ8X?QjK<@JD59d#cW0pN&dE9N{ov#r&zm#c+uu~(+tqcy
zs(Y)MFeV~m#!!JEY}S~ZH1+)HLV-YFNdB_sXD0hn0>NfIfgq=>;q6OvF9l1k#B_!A
z-kg=vmK|qEjZ4U@NX|=$dbD@dWaH$n#H074iu9j7xo>fNjr*2sr}EOhe?M_!RkiK<
zcU$w1w&~7|z3|)@#6fwF&vcfL631QB`z%lPAYpJ~)vw1Q+<JED9Gtgg#M?P5zu0<U
z<Ma82zUfnvRyju%2RtYo-?P)mbnYP|rc<={+S1c0XMPO7WwLog=gDCAw%VAuwkH!#
zdbo`kU$CHFzP_;cdBgQH0WWKB7d>xZ-nsugGd*Uu-xbfLe<V(hPsz?s__TGn%S^}m
z5bL^@2Gfy>3WK`v=$iCIv!%PO&bnpYUX*8&ocdN%_LsT4Du!4;-M8zgo2SLN%9ckB
zh9fiIeY)Jt_s4m$Rp%O=y<D<Yg(W$%aZU17#mNy*y~e2KzU%6}`a9*b`~bg}SGp?K
z4?}~i9u=HAsSH(`jgK@|28up5*y7_SO1^KS>)<?fgXvT~i&^&ys#GT~`@a)8%Qn8M
zIy7raP2{sR_x*zFDyy9)<z2toXm;&^+<nEw?PGTDnsz7sT%%fAdiM*lRqveX7gx1g
z`%YM3yL4xA*~^;suO@%2-d1(DVPnmMpKtnHe{or=ko7+L{*$RQRc_^JS%-!u7Ut`X
zEhsY!X^=0?&0Z$@Aa>6<$&`@iIpOuf^3&hn>bQ4o{}9)S(_99R&gm-3JF&_%aI$jc
zx1KMLWsDE6Y*~2UeD_HsQR}d~f%fx{mBp0)a>3d4hdHXH(aURxeiJurVZ!M;y+t|)
z8q-e-vOY+g;iMxFP#$pEJi4G+LWzt7Y$tejLRfz%&?hV-AE8Tp=5Gc~#tda^EMUWA
zVUF4c#e>DBGs8TWWiXT}5r><H+h`jU)2=o34Vxa5(KbXJDhsvKHYgT%?+@va$EPrq
zJrM^=18ubpimgg-KO6Jj*Z>Jdn+Vu?VCuoJze5$}H(opDe%3aVq2y$SV8>uTZG&Qm
z<AGky(SwpDl#o1NH%Pic*59Fui+Z;8)!4IF8A?oM7-cicLEE6%bA7?~4Hv)4HlhZR
z84iQRVeswmP$i`DlT+&AfEf&>ATtCx2Ki_k6i3yaxmbGRYcn%ShlrK>N{O~XaiY7m
zd(c}awn->Ga)U-NXavjt4pq|V>uVoTk2)mO5aNd@CdydbptvZvd0K;Le1aaOOJ;~1
z9BHI&UW=O_UOs!j^Q94`PsHaza-Q6vuS1nj^nYDkv0|qwLn(=g0X6~l+6KkzF7!Ar
zn$tB*LNVn2IZTeRwn6dH61$(;+Vd^+sKJH;Rys&JRNDv)C?Es^78*Z17;pXF#4QYE
zN2U+43USvqIKHXzU++AL-+s`DG9U)t1Iay*_IIe#y#2Q&@_n)^3}r=TP}!&)wGE2T
zyLO4c_-EZDJxYhnARjC@);1`<`0BF_o68->_VH*ZFgpp4`Z`p3S=n7veXP-!qO2ua
z-V0cXNWusV`hXx)Jv&Qi%tZ`{`r`y29^-_f+2PFap}6QGZM2|$7%fD}kMT3(lr({Z
zN*FUl$(QlvDQUvzwNn15O)oy5^gdBi2dl%8Av9YG!KDz6D<(GoIn?9q)fAzSRgYDV
zLp33rHX>~#he}~=DU8EGLbe)$t04q2Mk%?MC*x_tS*;0JG|0SdxBiQ*`@9T<Y@9OA
zj+;BlE6JNfseY;c97=OdbK_7kxE6yO-c`t!Kwt?3akvTstN6GYysN>7lOF{3Aedv)
zz@DsBU*^Ctkq7Oqf9QWo^u}0<IgI$%oAH)${?(*2^e<*I@;H5I42g&Ni;VVQJWRRK
znpB3P-DbTAkWn8V5eeDdpx6y|7!pFZ4jk*inZwPXXa+mXN<y|7BAOwR!^c5>9F!bx
z0ml|_=HuuV7{}pOaA*Z5K8|SR&vyzOPH}Qfx8yPVaxFfl=grS+rPGeh(-X3Z@<av4
znk0D=52=0BzMO0g$k%{^V@L_eOZd1Hd`iI=<^Fav1k=Sy*(6w!9FiQlX;NKNc{D8n
z?-D-ks=&L7!~UV`L+g5N#ebunBdxlcQ}m{HfslO$^fO?P8@Le>DtIwo7915cdlFIs
ze<dE)>_zzVM1#={X=+Ik(1c$G-5|zE-yUz|aO+c*t{%<0Sh~n@IL*3Ry7ExykWgdH
zTr?XtB+M8yBF$a_$15<bsDB@#*{k4qmBW*!ubr}^;?hWR=Rd&s4=}?!zw>;9=gn{P
zeFl`g;#HfBD|K(B|BdUBux~_ayuDJPL$h)_xhHy*CV9uE9$YaI(PUS5X#_tE1?<T}
z^zHgcb#;-h-D0DxP~s4&i_{N)PA!??5Z)g3K0~up?521kJ!v-CDfu<$X?CvDTpmhs
zO5ve-F7y1jaT#tIBRG`lmdQh_!EH5+z{sH4&EU3~k1Ie~0Un&Z5<DtF#o-zlQ3Inn
zybav8@yCA&ZeMaZ^~hGIP34c{1{8VtD3nNhI>K~MvwMJiEiuEbjM6E%SNzIf!)Jwy
zGmp;-R~`yA36=8rY!W8r^V$6h4CnLN{VIo7-`bNgG2~A1|KU2JPlY1w2@{z~lhc~R
zJs5@uIhySPn=Wv`@<M~1y1)Z5dX6x^?;MH~;W?Bfuwl$F-bq7qB*nP;^vtaXOTNvm
z9S~X8wmTmV{;v8qiLo9q?*SzyowZM<CFF@M&k++2fcXJXa^ZUbR0kju@j9!SLymr&
z@jFrcZzr&IKSgdmyKIpwxltFGbb&QG0b_&ERdb?7*V61>Am7(?Eh~5FSa>AcZ$L@#
z^E=<4s<Uj)`X9Ddq(%#){;{lsY%IM%_i`Ic<)W3vCm6rc19EwF*d<x$k={*z@x)i&
zBV&)1Pg_r{i&RFsbDrn~Whc0!!@j<=;A?ip!+*aq=eBQ#@7a)h`fem~R}UEXa2wab
zrgKxi3V3wmjj{Al=JosWFBZg-^_yxr)e18uIjT$JWE?62Wf9mTF@<a~D2l-j@ezkb
zo_>3ywhV{_KWvLl9}=No@D~zb3LrT_Q$b=qhNUCnh7aRojoDhHEg&%?{CMW{@#D9<
z4-Y6tFnJLJ`bto=s9BoRVUWY<Qo=vgW5Pexd03zlJKY&~UeTthRng8CMeg#vz=>k=
z@V#Kt3+9+~T2=N7C&XB;dt(TL78{7Pg*3W?{?n0>W}_^k<VYWy&9TUlBUxxtUz_g)
z8{BniQkuCO1aHiqG}{DB6PV#4NP~TvK*iy95Vr%1Tg8R$vI%oGtSB7N5~DF>iTkR5
zOf~KPtzl1gn8O?6{~I2<-GjvJO2U**Fzf^=Cc!%;wLRr7t((orQn1R`wIa>hv35LU
z&)V~lgP{Y9kwdeNhK@YsWaz|00Zf1iZo)LHVpJwbY?>X#j55JulV$_;0~vfVK(j&m
zK@3(OG#jiR%wSPVvoUmx5l50(I+lk%qCYa?D6<&o#lXYoL2(|OkODN@28uR5z5t2~
ze0&jP7eT@4{{pgK`1lgYE^+u_S^XjRUAD?MX3=|=joGhk-t+ixELEO*XZ#RZ+Za5c
zjUZE<I(;^IFBZOIaED3fN0R}AIy%?-yCo+>G616(D<c8s6D(~6m{De%t2fRb{iF~u
z72;t)fU%jCIWsFs^A^s$|5Y66WZWhB-w1x#%zPu$$%i&x9%Lc4l$n_i6-ZGe5SVe7
KmfWQUcljT@Z=;z2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_layernorm_bwd.hpp.F0026FDE37931C67.idx b/.cache/clangd/index/reference_layernorm_bwd.hpp.F0026FDE37931C67.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bc1371b4b6be270bc44bd85897b2ea92a3e00e71
GIT binary patch
literal 6248
zcmd^DX;c(f7Oq>ebyHS1(AcHW3O0zyzR4o8inPemEXM$1A`n0XF@Qu+#574n)1-})
zpcqjjK|w__LlC1xF*vR<BBGHTCvHSTG$zrgxMXx*H(a~U$q#<`Z}i};`ue`Q^_K5`
z@4iZyHFM^Cj-tZfoK?CgYe~8RMNwwtfANx{+-RDjE*eo(Y2Cr3OAD<$#21P?yUt2d
z70m%LKfWzcHK@c28@p|7aT#-_{b2e1=vjG71^*q4YB?yqwC$s_m7(YAHhSf_zFze8
zpx?=`%=mY^dLFMWHhF8p*3O5K%hy)P?yaE82i`GN3XQJD8^10em_5DI^yuOKOt=2G
z|7xq+lFn6bU3OL3`!oA^;U{}Wwf`aTzZ`%4*@BJ(TNc?g=Ui#oX}l%a(x<6m@s*lo
z;m-yBC0Tov#ziXIl;0{X4%o%7+PF==;aN-j_g}peal7o7PL(90-rT&>{fSYUov7#e
zxa_`w9mhXyTH5Im=5S?hc}Uf#{R4L_y$9L~7aAPkYR;UWdhl}A7rVsG-IHA+nr@Y!
z{;<AXaP01`>BYH4yFwNS!RLXdxA$pjPy8io$9+CEYqr`)jWJyCNoK*F+q+M0$Zf2@
z@K*ce6N~l;wq~W+-9P4$+2i8$U8}?Cn!DHBuAYhRR;TQHRMxZG|KC?9uI)JDEsxmP
zo>%tt<oSE~&+0^B8NUYasSZmmn|xU9zc<^=N;vM#{`VUToxd+VVc+s(_WL(mT{8xA
zmL$qt`ez91Of|dTy>YcLE#>Kd@)Ley(|WrPOx<51>RYdA8!L0ZI6dQW{iDc?TQ&_Q
zqhtD0etEodWAd7SnnUY)`$FbL%swNk^f;I4S8JV+6aCrVrLw_IN4|XYxtYbfUfU+A
z@{!c*^t$6wsg*xj=4~Ax?;5&cNs)I&@Rqk#C6IsZ@Ud&lg{|JM6-r6&#h?dkOs3}q
zNeb_-nK5tGDX;qLk^W=ntW;mQ{gLWcaoM7b6NFP{#|wHJ`z61#^cXXWi=G(ms56ua
z;T`#BRdcx@^4r`}BYKpW;(SbfEOZ9lHuK7MDkdq5SlW&}9b+5gpfl+9Njdx^dH$^@
zVp>R^ZUc`t2p_&+lc`5nx?eCl=qRQ~n^Rm3ur**de8DCG27gK$cEn{4HA!Vt%{Z+=
zw<0Om>d@w2ZnCs3>4S2L(pzU<xEoDAJNu2t>tpCq#O-V1E7lowr-vQrd*+!H%+eOb
zoiHK6L1)lia>Mis%M!^2bJ~QApaH}UU_E?!(d5qPpP$B_suj})q{)5|?+4lN1)Ho0
zjJ>h0yZWXPZAh9Vn<ks<47#fh4qS0{@or{mIq5^(lsFHaLAQFZ`r^j4CGVNj#^mWb
z7*_{A!xwDw;pg^E3%g7ovh)NJA=W<DMQ706GHT%XpFsty#k3KLP!EE7GK!%KHu<J8
zyRW>YzJsMDBtpDRyra&b`^1TDcX|UQhUT<@L}&$ZD_9L*u*vx!j`&Y4x2QIv84<;K
z8hcvn3?-n!l%jxxpLaZuYr9c!k)<ukD-x{}<vN4UKOAEm_m9>)KeF^h;!ct!+3O6t
zpVi;r(Qx#b4^7*!L)_%VLNQAThCoqPc5#e3MTxlcVeY+`^JoK31FvZy8QG)7Z)jjX
zVkADza3O4nrDB*SLLL2)_T0sPnf%qb&r@W;<=N&<#AgYboad-VDv+sQJRjSuV3HnJ
zf~=D7SA(+}T=af5_^Tm+k3-ZD%EysvQ0Q?rI8}o)ANy8=pB`&Kp@CU^tkgg}AJ5f5
znjY7Jv=%r_MI5}LP?zhtu8b~k(8+MAT&fL*V7OXvss%3$!Eh}gY5^&x59$P&>`*pT
zqMvN7WMIgN!EO8R`JENMQj_g*b4!cW{G*ErU)2z+hB%y31Fjl^sv#ILYAFbZ6%^M6
z=bSi(<P+7#&pv}n7rjwpK<Mg{=Zb-}FY;1@w;or+<Z9s2xf(pH!3+I<VH?7}d}1pa
zv4{8o>4;XJu^R#7@1DvtO1}#R+#z5N0b4RGm?+$Y6f@WvoSqa~4M$SouOz5eyYc5l
zLBzpDSz-+^L}~^5L4-k*V$B@})&&}lVz^-I;EDJ+!zGPQvOtMpxMrB#46aDeIq5Yi
z&D$;okZ4_C(FJ2LTJqCckDDFudXK0kX@Q$8>J0}9Mx3O1`~36MCfUhJ|9tFyJa9Y=
zSv#8=aK^=Ch&(&hgD)yKZc}8gX0kKK;IM?c%?g8LX$<Y(F)t1y!|-(V3_yb9f1Kek
z<66{G5yR!U<xIn6Vz>ov3wS8kEtiKDx-I0P0*8VZd|<f64vTrH)Tz{0AG6G<jE7dZ
zuL#upD%>l0Xr=o~9;yQOD&XPO;J%uV*MR#PKGuM<MlX5!U5$>LHb0pAl2k`N5V8$z
z^*Dvd2jUpo;L(93HxTM}bo`kid!s`uxZ<|SaP45z4i1>aTI|>k?ufAhAj=!hhVwQf
zZ83n$tBxA=-kr72{j}7U2Twr`5Nb=3toYP+Ylf*f6f7P{&xa}p>>`#7<dtD;7>`lf
z>H%N*;XrF)V(9I)|Hsr>cW&ShRmbv{m!wY3wJC=~4mI2&xz_>Y&kiY5GdO=<rO5gM
zndkLD{_ONAz8(VCg97pTlZwXPRTa)8LOV#>VFHq)ZnM+%@~#r!5k+WQ{%nJ&_*K`5
z(slyVsaM+i3iaC4tZ&3ij<F&9k}Rz2$)>;caF17j^62LI8^{=w?ULOvDHyU-*|mWi
z4g|%FaB-;kL@w^Vy=<@agL}JPshz)lYHRKCo?jJYwK~9}L%->Ked0pOo@syHc%^2e
zv#9^Kw7-@qwG_3=li{Xez~zt2x7H&SNL4Uak1Ii1smFMTAv}y=Bc!O1LfugYHLc5`
z!o#;MYWqtj!;vOs!RVPL@BJ+qj-TuK_S;+c5099xxH*t?7&=g4I-Ig9y)Gza4GA+8
ztm1_<pOG(LT^W=pB5&*jNhfe9G2MZa$`j(O*S}JP!O0@Cr#)-p@v!fHUBZIe52{9-
z{cC>R;e*W_yi${P!`<+{u}G>W6lw#rHn2iz#<MxeoJcm3SAb}b>6kgUKW=%r*{ORg
zM~;G0Um2zs!}&@5c*tMs&qD#y03MnqoyJ3fW`Qq~l;NhEP3NH?vmhRdW}_u|VrRG*
zHb#O+N`{MNV<lM4FkBoPC&6-x;gqaWg69l|i#Lg9Q6m^G!6bpjx|-n<O%hq01%{M=
z@&<6kB8uU*gM2$5?*RD@K5hhgBOf<`yorx@f@~+q_3?IrY!@Hz2H9>szE`)u)9s(O
zcCQpSv4SNO<g1=g_Z0{!_;A&%uk22>J{a+w^~kI6h@$0K$V=@fR8dxbdN%pMG^5${
zy5#qjuK^ZovzJder1~zj5U{4fbKQ$Crg#{mP^hxGtWUEe2CD#N8>=P?>!sqN%;Gn+
wpR>jeebds~@%mft`0Qd~-}5=28(Ue+#!JVJqsHl-QvI#9{#K^HwbI}I4>5E;t^fc4

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_maxpool_bwd.hpp.AA4F2921ED8F7A67.idx b/.cache/clangd/index/reference_maxpool_bwd.hpp.AA4F2921ED8F7A67.idx
new file mode 100755
index 0000000000000000000000000000000000000000..77b6d6339a46025d5e4663551892750ab187e5c7
GIT binary patch
literal 2624
zcmb_ec~BE)6yK~Q5V9sAtVuWxk`N#uARaMf1Px$fs^L%+1IihtfRtbrK~dBqqQ#<u
zQe{*`WfZMPm1+^G4k}i(RjIX(2;K#cK~$`G^xJiwarmQuH8Xj$@B8id-tV1BlrB76
zX^UZ+6kTp+>ik4248!cuH)p;%Z8HPIgiH*}-FkBU<n+7#{Gyl9=DAs!VOMXh@hW^`
zEn9Wv%F}qyE7yc&b5@SLV7=?Zk|D+Ea;I0{9cHe0(%06|yH=XL<at4Cj|%G&hA%W!
zJ{#G6JM6X6sj6XJepZPvCg{Pmtad@s#e>No#P_CTpL=t|*ylPkWY!5Q1EbnB@pXXR
zb?eGY;j6M)Z(LiN972d|ns@R$)9ZS}g}kOM%cqxe(*v5Bt!^c0d97{{DF?SHhQ4%O
zP;jDo#j`o)^I>^~V~UrtTwiuop1<w0Kexs6V#8jsf8^9_zf7H`S-r1o+k?WUWN+2`
z<S*;pE26q1*!s%yWd#fMYXw=CG_qY8wagpR%&*NHhq>&?v}*-baqElz3ejCw*9wX|
zc4_V&VWvGWZ`1wZrsr_xB=koGR&`XfJO7;5-dbJSop^mp$*p@`e%xqf-p_6OA{K<V
zR=f({k+Pt@XXd=p6n*cKs;Ps2a#;L>oLO<G>}bHlxZQ@PhntceR?m1q_CLPbGkH-;
zPh88|?Kk~GYy9LsO&fJ9HO2#ms5`u<@{anPMVU!gQGLR&bK(8Dy?HEr&|r*o<T%>X
z0<%FiL?07wQn4!gk8p857b87fJltpjaC*NWf1JGGG#_W7+XlA5nHB)&Iz^PVoiVRv
z;Y<W8I0^wR0AA$D`6BriPa(i<5FE>ob)p5p-yD%G`Deq0?Ofar`IbthUbFzXxNiQz
zzdZ%re4N9>NG(@uM+<;!#?36LY@9vuBkardwIeA3@UhenEsD+w_gJ_sfsrC>ks~cI
zHja{lF-XRV*_;D~_3Cym?ufk0WHMh`!2L$!T;8QMuCMqw8*w)V7&Wv2xMQ+0acoto
zN{A0ax1DXA`LqD|*`V}|zs|zzxwrtqau>M|EdYM04lKKu<oJ${TOs2K`~(+T0Nnr8
ztTf&HbtwYe8o`nLNM~9AjG0qSIXBSUiLW2BdAp~61jCBdi}Yeh*s$Oqk3U!bW;@_m
z<=S&0ot4;wj#G(LE^vSbVcOD-0eK_ncmO0wU$L(QQcjS1g+3765o8Q6Mg*Lr*<mL$
zQ~9L~f=pm1@W9K|DFXSUGBpz&2iOIOz*0<Nd2HQ|n<{jiW}ha6x8v)_X2f|&-3Jk*
zqs%eD!a^(2`dIuL6oxU-p|f+hW+j|#xQi?))k+OGNG)Z`mI-nupktAbND0wF+v5Z&
z87>(C)wHFXAKZayhArLna5rN^l039?(n2Ods(e&R@I;V7@}NMAruI^gvS`8H!J{o&
zsAuR%&|rqFZ0XqnF9aDYj~xyZM3C{4cscYW(!ApRL4$V?O=`rzz{;IV^Cs6epQ94u
zrt}7PFmzPpT#3CUa%%n{a!4^UNzENl{Pxk|?C9~`>;9+6%R9ZBLK<tj5W^&Kk`#(x
z-j@)+oafwz80y42DR4oWTBilAqD8ysYEgj?x*Jc%4~1Op$k1C&iY&T=dfY^qKJ_>R
zkK&<r1jJLLx#=?D^qVaMjO-6oMKo6>Y(fMSQiUJ9O^`;Z(GNTw?an(YDtW$lz>pmq
zeRy;J?~k?UILSWgQ$s^h5CLixy{~?;yQcZ<DA++%F`OjsmSRxr2#V2gzAbitQ1AT#
zo`484No^<aaLRo5Y24eq2;7Pck^~Kd$gRj=_h1=lXKEwn4Gy<iIUw*$?)k;b$3^qd
zv57Q^p&*yl9*d(4&J_a)09!<W+J8W|Q+Zj_S0`o+c>BIT^My4}v<BHW+8LdIefNd~
zuc{Q6x3ST_5jnenO@ee0xqPO%+PR7>n3O9ep^pht#+8wfNP_gR@!-OA6J(@&q!(-^
zf{b#H`b>*<iFOAJ_BB<Au@Wt~vZc2+MD~2^@l$>W-J}B|gEXN$P@5bI@j5X3jP}6U
zkpC<M_(ylCH4k^uhR;Yeq2H;{I?f&P+0y9@=zD$s(8j{n@!1SmPOvIJ{m{e4!@!tH
zocg^ftZx+qh(dp3(AhcW<eX^g-@p(`eEbOE0mJMqk7COsX=)iVBV@!*21_J%upc7C
K94wf~^7t=Q?v&yH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_pool_fwd.hpp.3BB9B8E57191E80A.idx b/.cache/clangd/index/reference_pool_fwd.hpp.3BB9B8E57191E80A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2edc9661292e493e9a0d0ae64723e1bb0cabef7f
GIT binary patch
literal 10426
zcmeHNYgALm7S0SP5RV}|^aM!=k3a%JAOtYH6(K?pAAqe@9s-v~E7}V6QMFcrf=G*K
zMNttIsaUmM<qAF!5D=@jiWGx_N{iH9TD4XM)K=>q1H+lM`ZKO|{UHnHoR2+c&zwDb
zf3wfbM20Dql@b<f^1s5CElB=2L7&B9nV`SSk2Bt%1zD_R2I#M(cGiuBH{7WHil=*4
zgdg<0f9jHieNR0aR9B^qx;Jy4=-gP_aJ9<?b>ho_&ZA}(zk2PCH8`91NvPGM86TJI
z6y6lj7ek*nBn8N}I>u#hd&$3Ju(W6N-#ZuH%ji)!-oCgeYn)?`UGNF}|MZtE^gny7
zXL;4)PmffEn=HMVI(gcGQ#~tV*L^ZGX75z^VXOXK*VsBQ*IU8WvJcmPUFB2%sC|-=
zNqyG)50B3;>D;q@e1A*+y_M4+N`9n$FAA@CvkXiPdi@vsiu4lpKHsoi`K+{rGfr~y
zeAhE~O<LlBMN-ppvs(9LFWD&HDZ3&XBQh@kxh})(*_dBWB=t39|0O**RkZ(sfq!uF
z=ni>bntt3&addB6WXiP*9`ELUo}9aQAZ-4G$Rz2e8>NRNtMuOWI$N(0_V&muUhu5G
zyb-ZJGF6q-z&n`!#UZw}&HfRYOY%yyK`q<KZPMHXy{VoV7v1-7a}N^dP5r*(MdP85
z^TJ{zS+|UL1)bS@@nq`ZUv}tyc;<4%B2UGXvX?U-nisvZ$W!7GzOUAZzb16?56A9L
zcigABnjJqnf1vPz@umltCSIYc7fvy_lp6kQ>Zg5^)7lTtzw>g|ZTs#0t-oJA=HecI
zG;&>2=x=)GntEi9+m&0?KW5bP>MM-xcG*tbFe5O_)^A%vmU?r#;daRntCn0`Gcu~O
zr)}@nDXBNkin3h&7kK=6X5!ZNgR?d7$xEkNbj6HI-RN7GmFm^pPuH6E{%TzwJNLPI
z-{)gmw%k9c4*x2lw!z<&4|eArG~)l%_r%$<H2&rJCMhfPlLd))k$1u4)9>lKvI0iJ
z5hg4;cvP^dwqO||4@5r(u0hiW==qa4(B70q`$_x)wFMSaGrw{E+a`K|3rCx==mx-P
z0D_?lt|DkFE<RGZ`X??Ni5fT!jZQ=U&;?gfP3}KYtSM<82}f~Rw7@_xPFt{yAV6IJ
zG>Goo^60>SV{UPv9qPqIYNCa<z|Y5CZ@DghHan6FDb$MtkaGZ<3|(-Q_+?Ih0{gJL
zrqB@eq7`ymq2<s8SNXD`y*wsb+{}eWs7f1@wZX|l7hI)O|JE2;^Z4dbkd3Mc*+L6#
zfyL$WdE43D{ZGe2J=9H+p@^$3uy|iISYDC&ehC+HP&btvr9fL?abrui>9~)M8grlx
zdfG?eW1}sw_);w8;gB18kOOUz*jMN))fQOX+GJ_6WX4}>M?yW+HQQ0PX4(RauVwhv
zbvpd~gabwB=>SQ9x3<9I8@yQgW79bfrqBRAT?e^!XpjakxJu9Y2j{xfkp>*-fHe3^
z{3mD&EdFg$O<}1rdhOsCR6$M^8iT<Lu5y3GM+1AkZ0>QOHPR3$3G~z!!{RYhQtE0y
z9M6G5B=)xTw$~O|JfKdQxZhyu5-!w7UAzvt*P&qOf~)*p*0-;``J5+&R>oTIWznW=
zQx3~$5U?_mQ!?k+GCr@**7`bjKMFa7xD~o!U!iCXwAMf|c8pipQ3KsD#<2k@+LhyK
zk4K-ajSxH@qgpF3JZ(Ewpij{&xGS8n4^VU!;8p<w{|>mn11B6%6uk>_cR>aZ1sY@z
zg%^%niar9-uLF(&W8T#ux3;eNY<+r;a?X@0_~T#sBQ~AVr{lunR18XtNSwx?LZB=J
zMBU9GY%?Hg7lVjmK-AZO5Die?-E_J_^>4{`eYyjLbpREE)i6X233(0-%Ynmk(@e2k
zy~sq4OcEiDaA8a`O*)N0=YjM*aA6E`0Z1<}n0M`{;q!{)?qkq<j-TVgyvtnaTxSNw
zxy8vC6z>}E#-K!xL{A1K$&)-7R0zfw0z&IXAl(QEy_<pCW<Y2z0<J}X&|M5XiUFa0
z8<1}UgnkWh)`0Ph#oB?i9XK<%1Gsen8G}24YbS7Huo^n6A@O`Rlx9QX`5fq$1BvHz
zp=&NAo<9NU6Of04mOnFk*B_4_ZqsQ*1gRz^*ILp{bW<`kTLAXZ9D##V$jS1~GyJ|A
zP;@1<tb`Km@HVG1((bMH%K`dGFt!p}V1aGmv6GJ!SJ-boL>r^Jd5G;2VQfv-;c%S>
zSA6cs*p=K^p-xI9TV=moIZhSpq{ON0`IM0kfAe%wqOeb7B^bvqMpQdjIhUZ%cvXBj
zLnTSmlBP3gBT#N6skR8HiU3i28<@5Y5cS)EvK>re0>2ZeIzc#tv!OB@67pQA%7w#n
zMbXN|hJ}p@I=!cl)x(SPM<zT(5q1kww;%^k#X-A2vwgU=Gzw|Y0ybH|lcDk&U{gcl
z$H3+>@MOrm9NLt_VI1OReeL$iR^5gJGX<iKS1BY;24-IEFCQ`c(Qchg5M>lnyC9ic
zfY|3cG`bFru?eE5nTmTi`i<aFw3sip#W?^)OZXB3Ir1F|<ivL(kd!Yake`{K1x_p|
z+TYCI0w-P+9bgt<ffG534m1n2z$q0)H$X`PbisKXMK?l8BZ-@!q>035q2w%yo1vtc
z!8qa8qi8utZpplsmz9?kBepgP$BE9YRHK3uTiv5|vhzC$MeiNjbHfMOxf_~w!?D=T
zv;QiV{Zf<d@dl?gF}(#xI@TXZ)X4y`$Gy{I97J}=@V550GHhIwE<)=n=!_?93eAqv
z({KX*xWaQD^8LZjADkZITowriKYDPKBo~aZUu~Khge)y^7AWw)(`rLTmYD>pbb5*O
zFB8SV-KwRA6rCQFKAbmF^fIqyK@3NffuJ%btEA|oAmAuaG4gI8=>{H{>L~gyN{E0z
z!yyGwQUD404(PoDhG2PmV};|6o&9s)V0K>Mu^?^Mj;{GS8GCtOUx7)3;U%4vta$oG
z&7CDY7V-w6qc9ME=jO)y6_w@en}NP?ROTv2W@ws$c{6atrp{fy+%r4zt2~{$kgw8P
zbAQ82g)?GOvDgKZ3}RBTgg}mBM*=yCod_fqO9|v><!4VYsg=Jy!K79J_5_n!1==%A
zI-5V+4$s#VJ%>MsKrz-aVn!Bc9Y>&fqIptAmMBUjP>LcYkg;BxB8@;FC_W%i0pJ$^
zqQyeMFC_6MV7&<tt!@U^n@PL{h_(Qt<zgTzCUFT+lmMc=QlKa$@ou2lP2x7dZv%Er
z5NZHlL*i?|`Wg^3@(y6#LE@Xh<0kOMo<h-Gz@v-At08|iB=oBxUrpjQ(0UCd^yfh9
z91`b2k32}|&xaoQByNE62Ixc1i}FSiH$iz5iO)j$SrRuxc{78*=&f87HR<k#|0}a?
z{Gs=cteO>v-(Y6LOC)I88C*!i^Gi{;)0v5N<-Z}K>4dyaD8dtb(UW<xYGb|zQB5`E
zRYMUo)m1~kY8Z;~mJ6X3Z8@u)kp>Mk*FYg|t7ND2_R7Y^6W-ty0rRR6Tf1(EjY82;
z-cg|pKfGi2jyuC7IbcQ(h-akvz%CyU^2@;cG6-cNtQFX`0(T}NpMn`rK|CXW0qkBd
z@;&L{`l^{Zw^4H61*tA3x!?2R)S=H#UP#x8)XP@oZ}^@wA><8qRXoukSADf&i!-kB
z$u|T+jeqX=hnwMnKr3;tmT+;RMbU>L`W;?}+y6)Cp1AC<JCXUafJGJ%;^!#323XXP
z_%X0}41~-8l|zeiIE=rM`^;$GId3Zpj9^i)E7M*Z6tzKD?ES|AF9Y${A{(8CtLpCZ
zDti5+18<j1Uc-_}wO3{Fiq?CfD6%@CSts)?Q{!i2^7FcZFOX|_Ql4Y61D-ywdG8zF
z$9LWu|1}uD*3@QoNTKC5ox-6$cFm`X^9lo<lr)N0$tE0Aef0)YAl`XG6f>SP-V{3|
zTf6^+39)JajD@3>hHL&oh~NIu`rUWG^__Wx=iLie=kjKIe+zUP$nTbCRc%yP+3BPN
z@8Tg-yKr2_F{XaCp@&U<*>1%;bwHFcGIbX;?}9v>B(?ZCtc;w_-}06sUIRt6s`{_^
z<wYB#p;!xKYr%A8hJFHMPr!6MBjaH)U~9KjnXHY>7bsdOQaUp#8i1$)IAawQT9)A*
zfYx}OhC;h9(i2dQlUj=Y78-vGY5bjnzfDo!WE`20ag$d&x41vo`${6yir449!GVY0
z2PId7yUEyiH#<`Im#@|<(<zv`6)lsW9O8ku;qV&50WXdqmviH|2^g1qwbhSZ?vC4t
zJ_BPqA2CxKG--pDIQ_(qj0Spe(;w%lcfUM5uv6Ny%ZQ>yv}jl@McbR$)3_0eb}(@u
zkT=KM{Ql5!0E+f8^x@zsilTiDeL0v#Ddbglb<h#J8AaDaK|Qo%@EIsL!(hC~gWl1d
z<37Bs&DNen!0&k9xXCj?A#KsgHF57zOU@i;6a1oV?U4rD(t|%r1G}ra4^gEk^C;ph
zz-3^5nK=vasN`h5>k%8#TQ`3VZjPr?<fwyZNH8Nkt=qS@x8mt%It_b9a(Y4vI&Bhs
zXw+@<2cNwQa9o715Y}Z~nzRTEuh<VQ^s}%t;01XW?obLZd3#F8lU#tijK2hj&unF8
zBxNprb+&~!bif5&#xS=6=GKC_9n0LBGq-g9i*ru-KguGE##!<O=3~b4ELfI|n9tmf
HV{ZQkwy`Z8

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_reduce.hpp.0C04C3B5F65B2DFD.idx b/.cache/clangd/index/reference_reduce.hpp.0C04C3B5F65B2DFD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b25d5edb8f3606ec2fbfa8752960f6fe0eb638e4
GIT binary patch
literal 7802
zcmb_BcT^Nt_xs)~1eRfF3+&<mOBYt8sK5eInh2<Xy_<+7aKHxoX+RNK1Pv{S3IRnR
zSTN{kCu%H#sEG|UQAA<^O9*z<Bm|<rx3G6-PX72fzoQ(SJC}Rsz1!b?*YV+@p@~)k
zLGZlrl?&%Dn<Eqm1V-dv(z3+3bSw~@(h&$&Zaq0^MpTQt_)668R0=glI(J6Km&L`7
zRr=8<R$N@pY?yTZ=#%>gMKAAu_S>~@&AloPhL20>S^e7|mVS}{S^8q=m8Ee32P{Ip
zJ4$T2rCS=ZzE<w~dF`>bRk_I{qE}co?LIVfH%=YLjx8?nZp<=W_V}@`xUBxj?g1UE
z!hSJcp*<+LVp8y$F{Rxu7250a>lUtDJw5*Q$l8XkG4YDywYT<fwE6X*nf3O=A<;4Z
zcYc_>-R)4e<y-~(v?MIuvEGN>ktCV?yUxZR_21|on(<e%{KUscAN(iU_O<7UgFEsM
zG<%=9`KNeq^!yw9Pj1;iy<Kth$+@v7J#N{GzxmcWf%&Ixw%)q$-C|~qG0?i*#3<Wi
zweqHYKk;<gbgR9({%KlgE4GDty>ywob%Nx<xP`Ijs;VWI7o@coHN0X26$`1RxQ18e
z_agjy^p0KqA^v3dxbM974Va>nVrZNg6t*K?wt22rQcPe@s&MpIF^Rut*v)JT_46tW
z2;PLlGup112mIyle}1#a;xA(lrHr?T5>3@%Zd)H2`)={XhOfWGMIoCaT52P2Ewh-^
zcyD(6n}%y6R$iCnW!68AcdKuyjq-3_u%&ta^ut*RYjWCeOn$k45_R&O*FP26slu1B
zE9!R`bz482emj2i;lLRA_{7wV)>BDoqwY_LITG5o&C>If-#^Ay#y$&NXFM~?->fzH
zNx`RxJu=H-{p`S}UDDQ#)f;mf@}lw@vI0JF|2y_8-O{Hzg*%SoZ+x?<pE7P8ccXr}
z<YG~II=NcUdo|h@*mA!hs9tN^;J9w7-WTJe(_IdTQg7Y=^M|rOeow!>+;Fegs1u9y
zJ_(mTU(>QiGApX;jqkikPrRMXhqotvzA#1=mIg;per@_|iy4CrO$BTb5*Hy0H4u;>
zRsHz7zkM{#CeFwZ3&jGqiZZOCEY$#5J}#E!*A~7LYh!&A0W14RX4d-<=wd_`Kr9L_
zpTxZEdU%q-gUL39$h;6assXGqO;EC&x%bavL(G^7*iOuJVk0#GmN^CkrWQ^-|AfH<
z$tqdMISYBK0kAAGZxJqBJigEne?)}s!n$3URRdsId|=6arR@9}278iKHXxS`NTCM6
zvV{3^bWZVz@Aa@2Sw$|AoA*9|oRl{=Y>{!IK7$>}3VFyi4+W_Kuq1bzb$RtzA2-B0
zWQDbqa4lu72EdZxr{`~cG5isO2a#2BkZTU|Q3GHpTw38(BA>p4!R~}56Aj8F3iknE
zDLNzWd8&2(Eraa|%O>QoiLmqmV5w5*MjyYo{IV$)5n0?V-5q-$K&}~Mcwkf~^CN=|
z$O@a0<7Q%yKA>T_X#Q*PyEUzbSeLA@opRnz`KbZ09N2Uzbk4EIP1=|uEJO5$nDjn?
ze8_K0Pve@Kb=p{)kjs!>88TJ_U^z@_&uH6~?ayEbB1ZwTC_uJq04(*<Zs*zz?#46N
zg|MtguIteVH2{{MR>hrjiCurj6w_oQH%m9i-UpB`X3ZbrRNJCtuq#<%BQo8HtkeKl
zE>{k}Q)Jvcgu!ycl8s!miC_8vurziQUn^IHJ~zV@VHu_~%&_+X<eztik6tnS%wPsP
z5pq6q$w&Tb04#TB#-!Bx+WgcP?d6CmM@DQP0G9UCm#cF>p*NXgJ)()L)OBF*LqmR{
zl@dDT_;Ch%kWEUFcPR=}17LZTwcTOQod5-coe4_`@-0D;YM^1+zvtn}@RVn!*o|!B
zVd3G}`vCIa;$`-gcalkL7?Bl<k+>L1)c{!DR=(a@b@tpajO~nj>xF<7i^PmTw+|5{
z&R>v}ZYB^=i79@H`bQ%W&B`t0HjrXMVO{Q%c{<t@%~m415*dSKXto>cbYliGnr8DU
zoqURc%rcPt66F5#op_;;y@vwtp&+x^_OY{1B$o-<9;E0&zOayxJ%a<!U>?00E1I!}
z4kZxNY&xRTkui8zq%MRI+D3cE1`PeGTtXN#uuTTGhaG6P72C9OxOnky_v;NVV~DD|
zk=<_O1B{Xr9&4P6Vq5!NLDF<Ur)O8Kr<RbtfTAv-XjnnWcA}Y`XcmXp;HVnRqc`H2
zjhIJ&gU#Py2~<~WX02j_&Lx!ukpm(Jf~qtdXBX!Vn9cFT*e7ZJnOMk1nng<C&q6lN
zD$WKl+YyJ4b^0`lKDda3qvnnB{ANWq60)(bv2L)R`;x;UUY=A7G3rEiq8<F%11~km
z$Cds>)?93#iyb)2bYS}q4)=Iewr{|vXBQL39yGcKjo}!Z@#to(!H4czZ!&Y;t<^%B
zeT!soksC15WCzDwJdDFzu&f31=pERxgQI(!4_f&`o|#1?-hs?^AV>Jn$00^A$KaFY
zy`qHnpiw{&vduWO8EbH#prW8df7=-JE3>cF%M0<^b6yggRa0~|#lWEr5Q)@BHXPLe
z2|;yULly^>ud#huR5pRcM;G$#LXjXE&ECMiH?Rf|p4Zsyej{-)QBF2cHWZeFgPwdJ
z%7j{imqhBr9#SdvyLHbJo-EkT(CkrUc@)`k+uTAnw~!Mgvq}tU7qaH)-ALNa$J4N7
z8s>eUjxE#qxDs0``FIv~$inixi%r{-*X=o_QcHRR^VYwKr8oKbZESg)|NUKTd6$ph
z!<P5>_yg?lfLnip?#Po1=Nsk`GgVQ#Rg?iZWMZVTQ+t8G7KySFq+fzepjgwSi0hXk
zlXi7|S4E-pEnnjnEN$T=oIGdqw3><*U$R3DWl}?#!46Zqi#;BlP!4_X0tvN)gi2pq
z28a1f|5B4mt@cIzhUYBV*((z%rFOq{Z&s}$VWy+K9kI#~hanyiE;QL#)`<s!BqS=Z
zybF7CXG0prrcndoY=B0Q94b>JQyd`*X?BU*l8*rssar?bB%v_IDWo-G1^Fl)nWiI6
zrK4FTGF9^Noycq_vW1vTsN3aSy6w%3e$yGmC5G(lOd)F)O&zJLtc%z~>=R)Zp{c_(
zy9?RvLhi5;bgo3gIfiqrK@plwL;h)KG~gn2Hw%ivQL8KJ?P6h5njLO5+zLLTNrvha
zB2!3Un*A?|{x5gX(ClF(K8$QY7MeYchMh)19DW^1t|JF7aUUSX0~E^PC6reQ6$~zd
z9-64URJ_y~#1N_bXpnu`p7}>T3g(Yg<qtBeLYnfvvcB~}vheC<G0hHRhw*&n#d`5D
zZ`PZK`LI4bOu;I6SSB*bL=xz@X%ZazS;!b7i6(8P{(58#9Vks~tiJ&ngG*@AA(~_(
z2{ey1>EumvkOZ21n&_pUhm1jO(#;_Kd}Iu*qN)l=OOO@63P?+l6;}n;Qg&;(2q!U0
z*`@PwC1t1N;~A7)1|MHX*{$Pna03a&zRL+j#j4oJZq?siz29%Df@Xv<m<iT22_p3s
z1{|36&EP9_c`FtY2l9{Y><B93Y5qGOW14j{bz`Axr&)JXcNQ)z#I?4=G@(JWUbbF5
z%-hzRhxyq0@UZpBWIeLrJIiDPvfw*QnvG<9XGwFA3{+A%%Onq3@SSCnk1Y7kvMxc6
zd}mpgB1gWnoYzwBU}TlEoYVQZl5$q^@eIm2gTr$=kA43<*!tnWI-uP4gG(m_KF<Gt
z7pQ`3?Jw4zNZHqKfr98KI|MR;8lTea%7H6~a4Fh%0|gInZgn{maH^_<cz7oz*-6>M
z0krwgPiLhYu-8d3FGsHBNWs;Lax}UeP2uoYSn>+nbLIUN_Iibb0N?Uc*sh;5QwNby
zE1<*$lr?AA3QAl-S=XrVr*Ii2jH8%QGEU;Y3o%H%b(_3xRo!y$_dXdb5=ILn#$~qk
zwq17Blmn<<nVOw{iSZuUuf-wMKN9uc$sr4A)?4m96y&E_AGr??%SN`@NY0O5+Z-h4
zqE|_IDk(qCNEwu828UM#w^#<%{axCxHFv)VtC+2<zC{?jkY^WnjZrmCo;R@OS_2aV
z!WhPcX=<uSJw*VsgO^Cn=^G({wtI*8MxWjGX%YFp4B3|<H)t?u!bpCd;7eFZOVm4D
zfbU72AXji8s-jI&j$~YPy}|M~cnCz3Du`qSTq%Ujut+_S;FhhRWEI@*FrXk}hA<&!
z;2n{ANCEpFd7OO2{EN;#AN*o;)azhz%GD=fB-(aTCOf%K@4FwbPuuJF>fs09@pR(-
zXES?}qe-G%K(Y&5%cHu6+IMm{8X;SQWi?pS5((KxY~P4AH=a7DR1fdN5j*<L2PoO3
zW=xe#bpadGY^-@KufC^-rcMQ#NId}p``7P|UZfwYlR@0{zpKUB<l6A$AGQ>c!Zk)b
z#sbzmf9>kUiyg1eymxhgLyg3{Y7PZq1tTaT-$-VpIczIZkD)*fNN-Y$gP0&*Esk}M
z_2WM38%)7RSL{;Orw^ZCKvb%x<kgfL6r=`!$5rE_*exGCL_iXvBtHqSM!aeC-!M@n
zNDcC+K~a#6S0B|aK0e}?-$*3nV(VP2NmC))fvr0@91dGzrq2yOw}pK~>OmT4d~;os
zM0>f#jenh$H(zfZA6xuI)_Xe<<R<#{jpv|TwE8pz|5+nJAKao3-L+nSWMH066i=gc
z(<l?JkZ-4Sx05@b`m}p=HzaRe{X^B5<}LDg%hhKg+k!n>xcdCt*~oX-URQrb3CWDj
zq!NiiDIuA$8FXVqxR?mZjLjgQGK3v9BS=y2=n*2AbZIW`3OKq57V18U!#rF(6tC1{
zMVgJ4M)Rgj^Q7~5*k^X1d4gYQHbI)eqb+h-q~K_)rK@?g)M2S194!k;vXB$MQ6~jK
z4uwS3%aLC>*UOclfhEYB>)j3@%L8Z-n1p7VP<RuX%H3O@p@3&7io>^H%PpA4UyH+Q
zF^~Te4!DFh{LmAsyou;E#2bn9&1@l%pT8&@(R$SICCQjgShfj!fkd5KkDYegZ)5+#
z0}V_-L`yYHwFire)FWaL417s4Ks6tRAowFQ{9F4A2jBFT^#3(a`)kP3`mlZH0^a)&
z>GBWB3e33K_OE&kGisuS646Rz#3|A@um(jE=O@ewAya7>T?s&ykQ%)<@$)rP5VQy|
zh7iD~(4hz*`;!vqCgrOp+)~w}??(nN0)Z{}YQen@(7a+cDR;MS*wmUsdN$U!GIOcL
e081;0z+A&1u;yMZxmO$RmHhWI{7B$MAoxE}-CBtN

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_softmax.hpp.4C7847F5A451A485.idx b/.cache/clangd/index/reference_softmax.hpp.4C7847F5A451A485.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a7ff00d725127caac16e457b88146978c3bcced2
GIT binary patch
literal 4100
zcmb7Hd0bP+6My@-!z1K+BpM=D!XYR`?z0G}T=f83rJy1T9)KuE1wj<0*oq>8P%1_Q
zQK=|Js(AeX#iCVIyi28Cc;WSHr6^XZbQ2ho(!a>(!|vNR^JZthGyB~kBw+I7K7E4l
zT^NuWADuj3ogfG{{!2<uTyhu@gsuibr0!@4IlZ*Qn{{W}Ym)}a&f9vfU$gwfExoqo
z9P2r;Y1V=zua~a;PL{INr@x{4Mty|y9&tG`S@NVIE#Yq9^X^BPA?M8t&gD(E?`LZ?
z6?dO|*7YdU&3Im2pJsz_SL}kfSBg$#_`OW-DY-nc*Sj-tPEw(JcZJv19@+8QH=DQr
zl`@-M9k%_cKuz~%`V*VgzKdSym*?IpZ@-_ApIgiOxG3W4&UU_g`TE|{$klBPdnD2{
z-#=@&F`KG8UeYJpD?ZmPNOZ2=&*`?RWo~_Ux_bY|Ys-8K)~r33H*N3l6`pC!g$)<S
z9o)?zZeL*^)`<=NQcwCDJGQ6G>Abn!<2c{uugvD`^>D4ddn|%mLK~j+;o9FgT%S{~
z-*C$Bq0`x?CtOcPKDt!?qNe3!-1{w_!M`q8d;N!({Yy66+zR{Ed(HihPv^D;I6wMu
zWy($y#&X@w(nXwgLE=S)R?D?X*VnuLj(Oi0G2JfQ;k(-OEWcHUnzsiB)bYE*Bp+H_
z1sfW(&j0o9)w?LcMaSpvXJ>n*lgXb(TF38u|8T~wSh{X;k=v!2PrHRT>Vx!CCUoT=
zeKW0U_3fslwD_o4k|0J`-D~L7xu3%%jW`5tFS55$41_k$FaC*I#S6=%;ZvEU2#ak^
zZ6%5U#N(Q09?d>cc%DV7vkAHau_};gZ~_$<uPk$|gju&(B*i7@^@zD1u?Htman)=+
z@?!4NNlbDK-o#F1XQCKD97<1f`LRFuo+hcwB51A_m#-KI9TL|bfoQZOkNmBRO8kRG
zYGN<)kxo7`9Gt+$(!8C|G%_u>aY!w!vJs8mh#UteP+1-nHNNjqqC1l`#$Gsx9BdQ=
zh&TFw=zq-4;&MnGY@rmfN|A7I0u|Y%6EQ!XIv&O%8CXSzq%!0?IDyJ8FI1Seu(fpn
z4L7D6o0bnl;sdMo)W_|5u1Tt43kJFdT*Uz5-6_pl^t{GjIHWcf7a&#v4s>7wm8!7t
z&5IMi8Dl`I;V1;!0+wO`@#RgqI-<Wl>zJe-7E5{3(TV}Yy`024f1_yxlVoD?1m1*E
ziUGtQD|+@;ww!e&NkdEr@*mA&u$Y9-z@q<5D`)O1zqOjA26oRy*M*}P;A<i=Iwt83
zgCNwTy5E;wjM{`K+Du|L7F<?K-g|T_A$*Qshr5P*090B|Er%`@ebqGVwZ7nnkKOgV
zu}GaxnVIt0v))h2$A3)hXH9J~iGR7?jHKv5R-g$Oq3B@i;31Q&<749Oi~QTLQUE_-
ztcnspcxWZ~<&bvYzM~V=@CHUH#;}1p{pFZ1jg+e##LqzjsB$<?*s{n}B!$ozibJ()
zFuQQ5<Vs8tTd5>JU!Y{7y^J<{wQ9$(uuVTXe7xJ`3S&fGS)I8%#pzx0&)Am4Ml$5p
zxP^W0B64*zhR}DcuMNNehZ->kk1`xU>64u`!pN2cIplUJQoz8$^Cw1o+7~=EoKBXR
z0fa-*9s-XsV3eXm3`0!sCN5(b8006oNTP`O`twq-rRW*aA)o{(;++SG0xHVI=Ubm2
z-<#+U+Zl*M4kTYBgq;}*`T!u^#=O1|t2zHQ{&pXt_8}(37pgO<AwHUwhf%3OIu%HN
zNH9n^*bO<`K-JHj%CYF$>Y<@Y(T0YGrm!1D+w0kLA@LOLpy$8^F2NYEJ@^nF#()kH
z0tOUaj4X@McsSHEFMCeBBH0v)-HA4d7Q;uM&o2r)P~PE&Usoe;H8KD%XMZS>J~_GG
zL5_*a0SLb*a~j*<;xx=mghwg2LZfThFa+lub&ZQK7n|YS8AuGIfEk7HH(rlyAR!c{
zQ|o->3``q2_cM##Pft5y;3a;NDM~2JP0b~+iq6xJtmP#6U{APKTt}!Puz}GXplT2R
zM`OVNL4zQ!Fv`O3LEuX4WQ2Rf6hH>!WEJhGT&*--IaI@O&HJoMWpBc9xT*yPdBJ{L
z@+{53Q-*@+A^2~;h-#;DarpmE1Y7{vy#e+I_i_W(vR`fKcahhHnc)zf;kw}23svfZ
z=OBdrDcY0esZtlAywHEug>5l%g1XpLS$roU<5>r(PJ0-8jE8vD=@4;<6pCqcSKW%x
zi7&Q{0HUqey8g&InO-vtqM`?5PkA?1%2~J)2`Z5o>aD0Za?Uz^!<!hpY$V7=Vr6k;
zBadw459{Ld{`;=yWZGa0Wr$aXjNo3}A+zqxXijk+R<R()zz*D*04#)x1@cQ0$N=)B
z!BAX|piL<*IuKH{o50N&EKqa~vd%&F;37qrApH`=2aL<(g46?N<aA=sst{GB<cNyQ
z+$|TG&Xb0L1m9uL2CiRlqSCiR<kS4V^gppfso^i=mom6{(B5r&i!Ab6462LR#TG)O
zFcM-JvIS^rQtoJb`_8GvW+eiNAs9RIV$~7;#_s(ix(w=_>PphDzVyc-RU&Ssl2HEl
z^Ubh>?teZWk+r78hyRc0TQwb9i=Pnh0Dw1JWJ;Y62kse$+nK8A%k(B|Xa5JD7IiCB
zmO6DfYn{7xwYBwm&naw1mE$hx{BZq_f#z9kMzvG8AWZGlNo8taZPm;7vRV4q>%2ns
zv33>WRUrXXG&DmPd?U(;1E?|-EfHj>&BHt+WSU{|h)`+o*%7k1WaEZmem`(&^X)w#
zKqIpC*iNHIOP`;M-IpP@3<;qqUKG~6m07oy>EO!e3q;^PMUP^S;wx9SOj{mw929M*
zZO4RblcN18e-5OKqJyX)l~oiKrLrnOW(CL!>Ykztk*E-v!TMFw$@=P4iwAOE6<dx|
zZa&Fqk|k{LsxhNKqkW~ER|6daec*NXj^j<XwI<>bf$}yd;F$PA-5WW;Dy<0guh<V`
z%2=hdlDD#u(RcPM&iL1y^~T?-mMok-b%vsAg%eF-wznun<AMI=q-I=(EFcT=wN)9i
zgEj7Q<L1ZUhS=|bZYS^eHM0<qPEfj0W7egKD~A4Gfy<Ww&)~)+VDlJ3^!AuZy&Dje
zEbs*Z-X<k3NSY^azYPX@aQR55RDAz-#w&K5mSEIqA<uwM2$e#<@{y-}9Ho35t$Y+H
GAO8Wy&#_Yg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/reference_sparse_embedding3_forward_layernorm.hpp.495E2A0ED186B796.idx b/.cache/clangd/index/reference_sparse_embedding3_forward_layernorm.hpp.495E2A0ED186B796.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8bbb9e432389b79798606abd641556242cb09d53
GIT binary patch
literal 5280
zcmbtXc~}$I7SA0PS*DYa1QHlR5|XehL_{I72BZjz6l$#-7LY|ms-U!jRavSnJgig@
zP@sae=qu$FTPUIyg(s-k;;t=iT@Y(gTcjX$Yu`=an0cRn`1!&2ab|uw=bn4+xo4Zm
z*}=hOd=@Kk`RsRB#jlOgXR%mj_&<GZ+6pegVuc&BSnpOGd41{1UN`>5c_#x$F9nC~
z>OWSjK9s6Bvr1u^Wqx_6aB_|7aFI*#)6tW&pA;20HfHGM)}}Y~_^6tL8osXzUU+5E
zg*~eW1NxVpf712v+T9gD9dloLZFFbL-4C)NOn#U6hlQ(#EIO2<4SlJeel?$oMig5_
z8|F5IzgNkwT09W*YHH@YS3}spufJWr^}*$voh7eTQ*w{48cyqBUVc&5PxlM1an%{q
zZ6ZpmV^=gzcV4^vmf-B7KEmRQ6Q?)nIp&WBFO*7+_KJ^vx>0&@Q;vD_`y(}G1K)`Y
z&#b%m>%LnFeU{~YTv75iQ`7I`Rt^pAT|=)vu9tFlMIXQX)wx4_B->CDkD6v|vn)tW
zlyz@J3p%%6-M;DW^^vu^Dx&Az=d9)im3&nk>2vdOcHOZ|X<@>0M|nc=xl<0cC0mKZ
zZ$_=0ep{HdeX*-||8I}2*aw4lPV)$LnpWEAe`>+-1((UIY7H6;x;l=g4SLqsKkZG8
zyh;7ke6=Zi3C}p|R<6BRbnl%aWr}-T_j^fmrtg0$D^MCaw(q=jb@1BIKkH%@1FG#i
zsyYn^oED{RH2YVPO~>)1IhD%n;6GINX7u$ApZV{HwZCyLn983Nm|r?^K73JjtJA&r
zRMkszru>#Pb;Hbx4A)NLN@w=mw!At?SYOaSzT+Lbq)4T(9r(P#SNwSPFLTaWDwS7+
z!iUBqHj#(Z*RP5V(qplRqei_8Cb7aO!jR9R1NngxEnykq=)r%6f*lVM>vQ*W31c3M
zE<#pC$YcBgB2)w`R>WHqNfC0a!CT_(pe4X&FK~@FyD_Nb5~f%~GvYKOdi((*+?95-
z-b~KhOA)VN4FUWBYb^n`g}d>kx389TQ-lDsRW>TQmH^wTaxgvVVPlLTp@%oHGPUAq
z39thyAC>hSj_9WdTfEv&?&qW>z<#wjveWVN23v{{V)jg%nQ|=w_R_dNpPD*^o}mZ}
z%=WVJl4}XDqitg!+Rd+X<`71B?^0x3imb;UAd+NlPVX<YFs2B7tU=9JOSJ^pDbb4t
zF5Rp;K@s+t?P=*L(Gp;9+9N8w+te)N5G3B91u0ul;P?YX^5;88{TbH7rU*-{A&4Jj
zttG(TzBzc!vioNg9D;4mqVo}zkIcp&AoB6>=z#+n#WD**4{NYCvgT+Bu&c{gL~;uA
zS91shY{L=6If6vv4-lzKj^`bkC%8GbNhlR+Mr#PLn-w2#2~SvDW=KrJ8bVA%xLN}2
zLqE_*y}ljlq6j;@zpupCK}&%Bb5}_>KYzHKOPJu*<%m;`ojmpckzW@#9e%x)h^Gh+
z*5EJmchC}GkMzfV`NQg?>kSEt%c6yogl<~GGA0m?aYUm>y_c&_pY_2$RN(bK5+4UG
zf%Sja41ZR8ywigqWSXjq|7ea0hhiCzfwcI9^gv3NO9V+zqo!GbTTQeX^(;)sGYXlE
z=MflZdiKD{zs|RALnNI{Cfk8aA-%BD$<$;Eol;Gvfl{{cg4*r;FDv!+=|tB=H&~02
zN8)rH%0b~dXr2z|B6Thb1PqyqEgMTyunh7On=qA{$}mBbs$c?SDHd|3ocTICH3<ts
z!pn_DPwKvzitWs>%dm%?_32E{OfMbEMao<>Rfh|apb!Zmk}Zt_5-gV2FT;+SFPtw1
zG1$YvM4SI$JgJRkV3KsGFjN9dNV*U?7b17iOwx6TU55k^abhal>qDc{*YQr#;%FCG
zCT&eiy1UtX3(}{ft)fNnv#sCP2Zub*m`vz!ZkCM~-dOsy2XC2!{BuwcXm`7JP0syX
z@#r~*@>W#?{nqif_4MiWcI%lwwaa9n9ORpW{6L$hB9;}yUbc?Fu6NFSb~1gusWM0B
zdQWw!`m0A$_Dktj+1x+1(lF*ec27GZ+Yto?1WGXu5jV<>5h9wZ3^%0Sv&r^;X;l<H
zXL4}zA}|>0G+ygMxiHphN;a&GGM;s0RXj(NLPwDC5yXa_--sz$SYO?%!i-MD>qHj7
zc=K_Y^46i;X%kAECEmMvH3qj+U&>3+S@40JC@02&G))E$wDcRVlqGg@nI1`7DXf(6
zGl>hhBp<ox;<gkym!cWaj^K$id8{WvGQ@8)P;BO=tFa%1YGFA13_S|l;!e5Ky4=;Y
zE7%k|6&7-&9L;q?nx^$kXh*Dl(Dl%F^hr8S7U$aiW8RWZ9>pi=<+9}r6fcWspgc4s
z52@g$Q`+kscUO)qenDtLXT)b5Ym0Cg=!e)&56Xj4tu(C>>?{b?%e(i{Kb|B>E6f$v
za2S%dGq+<PdvkjR3Lpb`Pzy;~O{#g|IFb$|19>n@k@OtS908n%q~~(xGEg`toPlbP
zv<5jq!6E6rNV=DaYmu~8hYN2XSQ8a+=c5U@7h(qcZEP5T=3I2{_F(p*tcDl(9bARi
zYbFJ7XvuY_qyFF44&caYL&7#BhYTwjeQRkB+xjBDjshesKyolZgS`t-2;iM3L+Z|N
z+312bv?6{h62lEwR65_-(7tZQgmSbxB}_l`wJkSrQR+Zshps4ro%lRsqYcDS&BlU{
zPE#{xy}%!}fxRMi`=i~@nGN@k6>DZ8aPz*wkm~5%{Xb&)WZ&dx<^2nHRmAbiw@R^_
zLgXQ?AWh=nmbV~R@I}*|%qBtp(7z|tV?epSiFOQw1CzMz+c%>rFb8PvLDGusVQ&p~
zY(;K>o2MTr%Q|~E1iQ2i@!E7n73Mj77FWuZk%5}g4g%)ziHi%rnE7Y_gig|y*7y68
zCz%nr@+LYYI)Tge>2)^io^^RVYlt$QjS21c8<jJ@4*m26*V?f!BK4JEqrB&?wd>s3
z?(Tu%qZ5k8?`2sZ<oHIhvGxwc??6I`!X9tC4UzM#cTUJFSB5Jhlgje9PRPWc4Gu1N
zXI_scuK#}!fm(&*dTa=X__^-c6F6EgL=&>T-_2ln5?fG!w{1aYEr^C_gApGGm6B5O
zpbVL42Ym>tJ3AU4R66&3VuPnXGiw0|k+eu466uhQz=nau0x<(g1QG_43Zx8VE3joC
znLx%sa)F$IJdHdlXqY7JMS1a|-;=b8Qt@G0Ch3{fOg;=ZB<)Rk^I_sAX&=gm51%F^
z?MwOcA+<=_kMiTgKt<C2ls_LlMA8vt1Q&9Fq$9~l28tr17^obH%8?ANl%#hd*)HU$
z!@H4eHxpMNSp^eUB3UI9S0Pyy6MuqapD^(rB-_Kp)ks#Y!$TDZ+g!ht$|n>-5RN!L
zHD7x0?e<{H>=yZ{10ydrr2XZ-rX!@G)Bh#v3)~lu(<#&xM)A{pRYIITc74B~%x%tp
zo=YxPT6{`O0{*NFsyDsC|6|hv1nn^_OZ;v2cL8e<40kXavOp;ecP#Kgf-rtp!mQD)
z2&xM-85VSa^t8D2Wtxvz>#;9c4VCI%MY>lj-K$vlN?ZQjW!}^!m@>&mERkASi&zpJ
I(?<9DKSRY!F8}}l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_batched_gemm_example.inc.341E18570766CD6C.idx b/.cache/clangd/index/run_batched_gemm_example.inc.341E18570766CD6C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d5c74b037668132dd3edd9b007e419f2a66f4263
GIT binary patch
literal 4858
zcmaJ_dt6ji9>2fgGR)(0=gu6S!^lgXLL#X{2!oLF5Cxk-Do`*7Fa<#bQP6f%Q#Q>^
z7ay4TLKh?pNzITHU5DIa5l9gT{AdDVrfFKMJ=WSYljq+1*Yc6W=X-zW{Lb(F`yEX)
z5)vM1C~8{njPk;~<qK$vqSW}eba}~QJy2Ahgrdr8Hf;T8s$~LG`7|i<d|RDkfph)!
z{x<pAt>tTiQqJmYHy-sH^VZ@AsrugIj`G7vUoQRQ@ol%F##2AeJE1h5DIc#rer&X=
z?Z{J?=lwp}v*~EzisgnSNh`l<ntOci?&Qrs-Ms$p&F=Q*pzX(&&rb^p3#f9+crfOV
zHFZO?6Y{%1o!z;=Gp9MIxu=4C?+3c8_{EsN)W31Cw(%dMou3a!O<$V)_b0bYmW|)_
z(u9ZuKT48!7Cc&*cXFU=L7ui<{qe$?uR5fE@LF&Gqo|68n*5k`@*`iL(#_FMZ{4=C
z$GdQQp~r=Lsh&x5+Bba|f8+FtHPPCo?_RNcXYlRTi^+F0W|i7+I#E%%FVgH99Z+y(
z$NFEV%su?pZ#5<Ub0>68UmV;!?}GoGR}LgvnratBS~6=te!A-$<$p#7PLx&@nmw@P
z+#7CQ*4(+Yk)d2E3bliMJB)euhmzA0ic<pu_m7S&9yhB-$4XR63UM6gX<H~6b_4$L
zs&p|I9_p%;N)(s}?Lcd63njNv@p+%7w#pqho+??18$Pe#6mGVKoS!fneef^qzpZ8{
z7kpl4uXDC7l-v&fyg$9{{^(af{jyyuvBw*n9h_Zki*Uh;1Z~pGM<g>@i5y?xz&WUG
z3%Ov$`nbNV#wKr8;(*CAPNudkguLCfz1aV%`7|rhVlu-qZnlMx4`)QUw14Rn%SxQ3
z6mpZfA=^U8O<VdG3`+CbS&1V)&2p@#Z6V~-nN{*#-su}zi3Xo`mAksx7DB!-*}vh3
z(WVGiBEzSpoK$UF2)T=v2H7iWPO%aOlQo>i!?qCe;7(cjlOpclti%H|cbB_6*%m?`
z4Y^pG(VRMsmAK>6N>1r!TL}67F0cOZF~zr735&@pPUUG^2$?F$TUa_<MNxL)%7*<N
z=FL(D8H0?^@$N#VAd?8C>C>VFnmNK8Eufr;91+Tm&rK#YY3I4S_r^6TBn-;%%a}|)
zGbqz9lSlG==4p{`>>%D$1xyvFgl#S0(*k}3JHD=sSbS7>9%$4CUTxq*2;6DU!ihVQ
zx@de2mob)n#@BF}B9z5tiBL9|%_Fx<;j6wp64wI^ss`t3@FFZo27!X)jV<hPYCX#u
ziNcmGBuf}{4b<1bi=5LfeN<n0V*u-gx<K9q93gnIYa=(xI&NaZ7Qp|!$&G}Qo#M|t
zdv+g#bUIxi*~Fj@pgTZCxOpqyE8gIy*(srsA=8jWw$Uig6qimANmMMdHA~4w4C)1S
zFSrw4J}o7S?yZTg#8%z{>=say^Zp*U6=hquo3Y4k5ZDHj2q8$jk2`cVynv=ryi5ET
z@|i}7E{P(P;*t^~(CR>22Yi498Gnvx+jM>_=6x8#4+9@hA!TI`Mt_>=frFz&TM|Id
zGw2`~4#EtA$G^F4-IRu3)?uN$A+Q_95pBX+U-+V0zD$o9H^6fZ5J^siJ*oJt^>NZ}
z%(z%v%*Pmm_Jd(RB$939FE*OBc0Kb3#i0A3ybm7ankn3cy`St}`MVv1z6II0z!Cz3
z?ttu$7=H({?*u%>^I>~|biqRh8W~w5N3Nrhi8YB(8k;6U7SLG0N!V+FaTW*_<Ea*i
z6L1@7+JyblA9U^eTT4{~&iF)SBF_eAi#~HIImDoBeYObY_~!7=V^9qkY9K+_+X{iL
zFoED0kLcPrhw=)r&G$gi9>JMYzrT9F@yNav*ye@mLU*#CK}G5!9wnUFaDU0Yh5Ka;
znix9qd12F}&`BaRIdrlJ=|lA*6cHLBLTAA73?PBe8Sp;?B4y5k<5_{8a<SWZ^|>Fu
zu%nTgHEYQ+8s)G#B9zPK@`!XWoU+#TCB%r%pqoV)Fz6D{mjIE#n7iVWH<Mdjf5S#9
zb1UO3(fo#nQ%6mauj4=(f`B0*kQp=#0mEWE0s$ieUNmoY-yzp`M=-yAVAuyq<l@C8
zr7LFtTpx=$nrX8m`OKgkI!A<Z>0BNanC6ea{Am6I%qxnEawSKK?VLMg((FUneap1V
zL@9Fs3<rdiA^lpWu=Q!7&M+Q~j<jP2sr*#|q)9P|^XCFcO=Zv(P+tKJF%^UQA-G>q
zl|h3LI|y+C{Tjqx6EK+w6bjoUAPAuFQ0Pb%WRTKBsUuC4L99O;K)O8^&PU}d2-gGZ
z9<eyR5Zo&krw@Yr#CQN=2Ly4f;)V^v1S0Obk7gC7@BQju*dgT;%U>kcWl+V$3K3d8
zaW#+LNw!=6{d<4BpkR<xCRLHSf<X>42Nkh5gJd$9iYU(@g-oFmX)9Bz1Z^1^qY_-f
z%2<_1cbQ5h=+4PFmEahSOrs)W%jUI_?;l<EINus2sZnV}k8N=YFt6u#GcGUj%6L(E
z$<Swr3QCqfON7k6W>H4&0mB|~c{vM#XNB@Y<i;i0I?f61?FV(g(8qQ>c`a+5N_Q1U
zYZYj#K-4>0Agl%Wac1wQT&eGWgXiGh(FN)*a3YR6xMU`sY^)r@ea$%D$d@-<o5!0(
zC@nfIhA3~HOd{HZE>ru!Ie6r|vA<$p7y1^8`gf6UkqA|TzFH87%%r%USO-#4UKo_F
zN$2~iwH~O3K}&=^{PwH6Z<V#`u)0}(S^RK?>$qPwkLv9@YMk?3o3YEv-OK%n)W?EH
z)8{@khT@uM0Z)rCj9Bqx3q+8OR-9^qmk1_7gI!^rjRjW>fO<fff|_o`tlaH3dy%yS
z8O;1?T!IWaB9uEbH=A(&cwYbWwVPEMJklBEM!qsJ$Rsz3P?|hVge>4_5$2WVgdOFL
zvA6rN`%c0OCt(iZP2<uUG6Z47bTl?)$gmiXK*)%I$&`=%VIA=Wf9wQyCx|>U1gatM
zBy7(%2M0&yv_HYdUXGUY#%`Tt%C(QLy=zUIX;CwYP^~6;e%R_)6!>#OY63sBkyMu|
zY)=Ql!mJPLOU#eUrw`{VWa(v4Ul!M^0SFrq>W_6s4Z9}bj?sBz8m=mDVDDcBugl;~
z6rgb(_PQd*J>c0RxV3Za>bKX$q^Ybv(i^7`O|XyjCJ{=Tp2jyL>*zSO4W^O3q?X_S
zvc7)=^{#{ZI_QWb-Gh2=c%?2J8|OoC`Vjb*)N^s{+HG52?!tu4;J8_c++NSm?7t52
z?!bf-pgjS6gY69uNLo~H*M*}Z!8?I}LeMD5J4u96y;DV~9{lQo_uzobx!+e!^<HI7
z>Yze??!seCP>~2#Lv%I75={qg&2W4pXw^xapQ)*-ynh)~52kv^Aln8v-0Eu`wm**L
zz6&xOB*X;dHApLLuRtLPn}KZxetKe13(RVPR|r1_jeu%IC==IbT>31s({jq5K`F`<
zQFYGs%@tMPy<pfYuE2GmsslHH={X2ICsgF?-vs7#1nm9>3!Rjj#Op@d4R)^ec`7)!
z57d1^p}w2nm>!8<{2W`xfDEFl7>D9`bhoPU@!pWG{W#X*>3D@8Ng|ynLMe2L2-ShC
z4p_36^fWAv^*JnvGYINIA*z2(j9FF@kF;2sZjg6_5And$^PK_jN0-<9Upeu+ddULo
g1Xx;<Q;PrBg_58fe@%S&7Io}raH5>ZY)Dc61MaB{qyPW_

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_batched_gemm_gemm_example.inc.B0C303C33F007446.idx b/.cache/clangd/index/run_batched_gemm_gemm_example.inc.B0C303C33F007446.idx
new file mode 100755
index 0000000000000000000000000000000000000000..904c3f9388265265e14f30b57410f678c47c880d
GIT binary patch
literal 5004
zcmaKwdsLHG62LR@yKDj<AwdXf00S`~gz&Hg0?1P&N2#vm)oKOV>eg44qT&M(d2E1g
zE6@TK=~i8}Ry_nPxV7rG_-N%}U0+8}-FDk@)>6>g8rp7rbk)6a&-dNgf0lDX&Mz}}
z?qu%Vnfa1C$7;1rmPj&xKWForm&#t`Boc{=zSftmTiH(EuhZA&cdB0h%>7J^GHWca
zZ+vg$**zydyJ(w}6n{E7`^t0sfBD8YN6A?>t-Wyk+2<EU7rjG9FW=sI@aR7CTYW3O
zoVM_;+pU%#KJ8eMe&E~FvptuOE!6LwS9@8lJ$TIbK-s80to5Yd^OxR@*+27(BPU9m
zZ@>A9Pw2(Xx3<6c_XSg~4h4VY_Oo~Op^9Qfl=j}DQ<J8;&1+d-zGg`<_3O&&SMq*}
zu*dTfH;IJMKmC(`61fy4lQ<wCS@+Vi^?&r1NG479IJEoxl5H}c*dWXXk>XNP55{_k
zX50YA2F#6MZiGZ8?}C&rn90}<>2}Cs+zsj7m=A;HFeD3d&*fUn%C3M@QjUy(_Xzk1
z0**|8Vgi%`E00fE6T5x(MLAEJK-mOS1p!Z*LDdWf#x0;~!Q2YUR;=F!sx~Zd2UR=f
z4p4Sr{T-m%f#sc`>SU~Los-P%NUbA0srINA{pU$7c+|4*!;>Be>w)Ra&0+8vhCs$Q
zA>$@w3#|Fu^orqw8CQXRU#WkSG(-^i$KToXr;wITDtJSxeM2e+BFMd^v|i)=77b<T
z!qP%PpCe^CW%&YyX?H32U7P;08%N63<>Kxfsr0MV2}&HP@~c8rt*#DdvRY8rLIh(c
zXq})JI4tL0>Zh-rdPB;S`(U~c7D2$18mXy9Dss41wq#F$ufvTel^&J8OsyF#&5*&k
z4J>V#?GSE<SSB9?pFx<ycpTElSsaZUZo50(wWn#gIw7PJQUt~5{)C|&^0o2Q(Fb69
z011MCo}AQFD-}8VZuwvPowHp5KMMXwK|Bd_%BSwv%U|%J7RDfO45!Q(#EroW!D3wW
zqql#l{%$+9a1ZqN*j7o(Gdn-%-1La!NwzFoA?Wku8QC+4=E~+GnkSnlQqq)(b1%ES
zI3efgJyzJ*b>ZkeR@7r|07V1lMo=_j9s%PBm;@I%;sAvM%N-EvfG{kd;b75p=`V1=
z3xYgZnRw>z>-SdD*o=U71f~fBp5ANq1dCP5qOX?RFY|8s4Ns~(s!$rVfTaZ=dplU#
zF?T|0C(LGMyCJ+Ar^OKX4B@o635J^xCFrO9aP7yQgS$3P;z=*~^@5n~Jm~{X9|Q@C
zJUIfIBaE#d)&IEq_Ojj7-({v{PcyaUrsarMm{uTKX<CVBm1z~CF_4de7qc@C@^Qwq
z&9wnHj(y)oQ|BZ+b&@6Dyn#zO*CHK597no9(Zw9(h#eGm2o@}Iq#G37n0r9c!}Rmx
zjTdfc+IG_&I>68Yal*pF%b%3=_gKH8$yu0Em@a(GlcLljbUhcRE=E+GQY>l~A3Aij
zD>reuKTm>mL1LZaNw6*$kzS`q6sikFG)*@RQJ5|akx^wt=@OxeKx9&x5JjpY5k+gG
zQMHKG#Uh(#of%P_E)J1JWkFUFR0)U@Rf&j_R7r@EwaI9|R849C3uT%n4N<x#9Z`lR
z15svRrg)@0nH4w-QC46UB5R;kq?PN|Z=8ScKsr6cF5tSrlWlDW&W^boxNgin!1XYG
zIk!0K%1_02XqJ>HOZ)^gtN&tX-gf4|0R8ww@c$4(1;M(84Q*5AKeAInKRngX+U2Il
z8ACN252aAS4hY!+X~K%S?zw9Uc3vK&t`0!#03Im@A#M<635q-!g18~f!w@^n^lkg9
zE9M<MJ4k1nTthCJZ1N0wi1H2jhzbk^hzgB`sKFE&ix4d~E=E*rEEehY>`AsCS{^5O
z@`RW1=t_IaJP|2m3PfHqFGNb25|K)#LgX#;Ml@MA8Ih072a&JL7m-?~M&u{+L!^;u
zMA}xH_;%M?=PK6`$CShg;eIp23WZl6EurBo_b5m6O%ud7vE$=OE5x^A?tu6X%v})H
zh3B6E@EO4K&o~&y@gV91FDGaOH@4sXT~VD^z%}aKC`631xuQcKTf@(IVeJz^;3z9>
zyJ|;6mnuH#qXR~+B{xYh#FM;)Jh3(NBtIb^QGulZX_gpEP*yl$ffE)9cJ{1(hRe=r
z_$QsDY!GFG7(w7_y)pG5mUxadKuiPXMo4O8Bd}}o{p|=m%hpdo=mcxEd++#$e|u}{
z19~xfA*>gNs1J;Ncs4%*#v_=Uq{&Uv48icbuP*2u-OwLEdtNITT3My%NgIT;LA;>I
zlXeJcXM8yPyMjd%Ii^XjF0XZfSe9Ho-NCx9i?batS1|NG_pdeQ<A(0gDd1PI{t7vQ
zfG2g*xH_rWD&GGlbyKHi{)#7cDzgioV~TdL+F>E%Zm@P^J`5>`S@3$*pPgH!dt6Dg
z;cM{xnq@<8;q|BYZu9b{*<b^?ja4?D9+$ixa|6g5FgJp{5%UOmj)01J<N&#Y$v@a8
zo0@dr`x5owG3Xz&Vef-~e?2^K=zIzde-jv*aQK@cq?v`ACoK@t!uUh?XK!EWeC=1d
zFfS?(ol=pxD4DsjW$qSn%s*z`zPTrT?7BZkf($`uY6~_5Bhnl6h(Zmah^85)Aqq2u
zAu_6sT47_3M5rSWnbamkk?Kf9(Yk1{p>ibF5Q}V@4Q51fhB!nPwFOy8P$wWtR3{=z
zQYRrw)+M9;QUg=P#>kPhz%)eZf$4}c0y7Y0>NCZ@#*tb2S%|XqS%|E9t4JRo3lC3R
zdhRaGgO9=IW7Zn`bLFd|E1!3#$C(?IE7n;$Wtj_6C924-5U<OS{dDY!hr^lF)(P-G
z!KT1--Zjf?e~{<VB2*DnfqdQ)v<1=Dpsk2%AgBh<gLM#8huH};oUA;Z-_?-w@IA#V
zbnj!}e~i_o3sWm!txL`GdUB)UZIC2ngzH8n)H4e_Er1CPm>XecBUqXK2s}5!o*i6|
z2vHNPU|#w%UU_3vz;s$g9)k8E?in_z)+QC7Vg-ZX{@>p{=uaZ9zc17$A629URXi!6
zQN7HtYu1nH1GBKsHS5P5!Q3RB*(4Pm{jz@a^Zpxj5l`mJ=JUd`f6V#wtd2dVJ@n(_
z;D4N*^sOFSVb#yp&*@3$N9Ci7yV$%KUEHm6w~80{`&B&!iIZoZrQfx5aw&RhDw|v;
z()VpW&Yck#chkt1bLH-WAM`<*t3*`ARUxY8su9%!R||-pAajEFSo!1U8%nDVcqY(@
z6+{)F6fHIvqZF-y?3(|n>5yH=_~EGwp>HQ|e&_#w75a<R>K9#yw0_;v^&a$HO7;(a
MlT~v6YKBDe9~1s9$p8QV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_batched_gemm_scale_softmax_gemm.inc.35534F29FB76E4C9.idx b/.cache/clangd/index/run_batched_gemm_scale_softmax_gemm.inc.35534F29FB76E4C9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24757b57084b4d6217be648b97b704f5b8eb258b
GIT binary patch
literal 5692
zcmaKwdstM}7Qk6>jt+1bW<UlS!9g4b7=|}8Je&a^fFhcB0kwz$wDJW43O*2ds0dz6
zDWSjwbU;KBA2&n2H#4*%GbOJynDvEGMx+**;=8-Y`<=7b{pa$1!})$^t-bbZ?X}iE
z^Jm4yg~}`}rZ1VbF8kFrnT&;ng_QhOugPC_hdd*|!eZUKTX*SJoQsjnICAMofA`iU
z%DsNtn~{#%au>^4HFMt^O}$d{Sh99|*5W7MPvXkfEc^Mx`m7lroX)Ntam={&OZwU2
zR9|^u_l6IZcXiI!0{jN-dp^1T>xgk<(0P5{wWasp`h|;U+ipJnC8FB2^igJcb^bx^
z_2NaVhF^iPI-}RuxxwSCT-T?|=EIIQ8}ZlO##;CMAHx69$+X9<`Skdk^Ti(@b9Gpr
z_*&iK2RHlo#}~~}&9ceitS{GHfANjR!#~=d{_6O*eYaomk}vGiMVVCHu?hS3xTa62
zsWv=|ZVPTapI@;sZN!Q*t}I=>K08Z6X1&9@XR%?`)eD@(SPKi*im@8|{GWxjC0LCi
zOO*fWYpZXJBZ`KPrd|s1u?LPVwUn1yPU0I{1=+SYUCgQ$acl*6SAZWc;Mi@T-Ub1b
zD?wd}xe7w7AcD$`pfkcW$}JGn0<n}^A*K~`8)(~L3NN?K2|7Dia%(mj;R$d#0m2CF
zUYBc^b-13jWY}JC=!J273x>S`;u|30S-i?=_R+b;0mMSQQ~V43Ysskz*&($vFLNBb
z6C^vSVUFDe(p{kB6*;yBq&1j#gJd_>uLWr>me+x_4s$(7>al(UNE@(x4@mc5ZUjjq
z)^7r76P7oFw3)K(%~?~Jcl6~f$CiO@8C?jD-3+#yF&_Z$0}w<fb{bTt@zz~~sn;NZ
zm%Dr!wBhrEQ?HVR8wBS;n8*uW@O1npRBBm51byJv2kyLJlEe2FJsn{jV_s9`PMFLK
z%xmho3&JSZfM*TndQjF=eQ$ZA#Ml?KVl2ZJ$%@?gW(-^GT&$o};#`8L3}j{CL#3NR
zwi$dW4}t3ts3;GEd>A}<_D*`BKfnF#E=!KR4Vv4a<pmu32sDo<`^ZIE`&^xVAIq`D
zw#80VZ5L>F!Bon%psmH+0Qv@)N#!lz(*hxsPlD<s1XAvYn0~q&>VjJ~_if}KTXAdy
zcs78JD((T#J(!!pvzc<hned)>tn<Ri#O{FR4usQ&k3sX8a=_^Nf1bHNvz~10VHkfH
zJbA@X-Fch-^)HPh7EIt~!meq8$tH;66(<KgYxuL|W-YOB3p{R7(~*+yy(j9|J|jLl
zTQpnD>vL?PC=t;d(Hul`MRO6&6U`GSa>B@k6=O0-tQiuW#1-_cGbB2Rw_&aXaV6#|
z5LaRD1$8fI_yIBO0Eh>$d;nAf;Em-`19WZ8`U?l(C0?$3`Ma!#YaDixuvubTf_$R}
zv^Cfl>OfnEc@OCKVBcs4eKY1(@M*=qatc(Zu&-PL<u&l<hY@|d?@s5z#zNvN?cm%F
z!Xe?<L*R0V9$1d;0GAHRasS$SXXUNenru0Csb*;~Z;xY_X_g^cu33&~g=Ph!m70}^
zawBr1sg=BlJVXT%1&9hG3K5yW+64BrWj|Q=Q=T)eXt7mM;5_1IcR_iVW{WwW_9pfD
z{ocbctP#XUIx&WA0dWfrVGP>};#SNDKzxAe&(+;|GwY&cD=_R47=MI%K=ND%uc4YL
zED3SBe!1xM=K19zD)1{nROnYIP}*1La?W(deMbzW>(XQB6b*U<3LY8y3`Ci_OhMDo
z-u^+y{D{}bbF4z45b`0%x+~lfO;k)o<f-sPq*AC5c`3XQ`AB_`->Ri*M7~mAL>j3E
zk-yx30-a2-A{f~WQG_50RfHl6lZGKHTB#OMxHKG5gfs$ClspRcr*qM{(%$thdPLDK
z(THMPVh~Mpn<k_+j-Bo{9nlQ88Hi%tVg<@b47I%f`bVpY|6KyROVn*LmXw;JZ49m?
zQsz45deIRTI2H)BG_A;Fbl&kJX*~q?Lm=nXmgTQ5eDOhB3~>!3Fh;QB1sw5SrUi2=
zFs+ym0CRxyiunfrD}NY%BR%9ua-4ZHE8kbXQQh74DS6!q<2z|C%v-0w?0VtKQ<h<K
zy>n5@%=6AeRN!5JsL;C*(LGpnkH*2Ze@*QvEo|43o|?e3351hfemF5ZW$#y?ktudT
zY!@W+W8>J95PuR<DW8J)Q<zUf>}jgM#aOao&cRPUBat>)nT*QQ6lDscRAnln1<D18
z($(q0spMFL+JGoSoq;G*ohi`vnUDYaNm}MeruHWU{z;SH-xI91+}`<DxE;sZh-^^o
zvK84P;zS%GJCPkCu}F-_USy9*B9b7Iilm4fL=K1?MUIHZiN+yv5;-A~iDUv*mqj#m
z<lSFx_L;z($^1U>Pq6v#4T9Vm%zryf*{vrhj(QPC?*(};2+qxs>RWb$mIyUd;~?$o
zvn*?lEwU{Viga^<7`Bu81;_4&u-%yJA*>#ABZM_#HiEYi*Roxp>cX{bKPdb0wPhIW
zhrxxPPVH#ME9LgCePp(M;M+&5yT-DgREx#uj+5#xS(_Zm>vL>Mc#2TxaBOOLDxwA2
z1xPbTorAn=2-1eYK<x~}!eMxs=Vm>;wSLJLUyyN(g6k;F)caP>W)hMrei&mear`T2
zt~ZxBf!jdHu?$-Yft8r6AhL?4MDz7%ZZFK^do*8<R5xfs+y9%B&yT~C9*r^A3f}E_
z$qs?~5MI&_P<LP+gz!OH3z@Hwk)@WxQ8#l+sijaG84J3{IvLKG$W-q^;5|CRmi?c!
zTy66Ik`((J-8TwBmSH!!Z$eZG?xi?kmxFsb<}Kj91@jO@4}p+18Fm<=hH0lqd~T+^
zJd&ibGM8}j0T6P6nS%yst!3tf0Z8Q6?<jLG`{Lvtb8LMAlb_HYj()FSTkrDX(&y28
zx)F!27Km$sMZ5)$)E;rInA<?tMtA5~T7U5VYI_H=BNbp>LBqQF?rpsdb0t_;Vy*(~
zD$KoL*9%hG@c>v4Q2B{!(WJ<S<nGL|55eOhE!Vqj=H0qdzx^?(A5;8O(5=l7YM4T;
zZA#oE#P5@Fwd*=#e(rZ=ScOu73TAhuJEDooiHJOvo`_UR6(TRC7a|{-kB}xARxMK_
z@|F1_(#SN3{1yH}t--Lt%3x$OL>Yo8R2hmWOcsW$Xk}VN;j(Z<5wZwGQHm(kpUzE(
z99ZwBM-=TAjVQ(~2GKN+Y3QCh-D5hU86Gnb#d^ddikHU=HPMGBeS9JoUl=8u`#w0n
zPj7Q)=36fhC|Y1cHaFQnSqSu`91TrJF>PbQ#w32L&+9#MN9+Fm2C>x$%0_xm?|JXz
zkkd~;pH2kFVEi%K$%TNv2j}{@`9#nKc3mKRcDdMCk@WO{coh-+Z()6DQc-2O{%`ir
zy*{i0BKgkEULU@Vo?MOuyzolQRWPjz;;4Qvr1sKU#GH}*Z_ohQdp%4txYjj@q^=Q=
zkAU!Xfn%S6{2AqMWFI}-5beH!9O!oNZ>JW_x2V8F5K4ROfWQvSgP<Rz-@LAG{mFD@
zkUdW3oFq!(_-5bEYM$}-KFvPz`Y4P)O3Sg~&Mj%hcjHWChf@7h(ecX&%|OSm6cS5m
zK{Wg@ZAq5Z6Fq6P8<e~09&v0fc-G=?Ep^~oNBP&cI`8lGz1&1HUJ+Ad!;g-9-C&9l
zl`tiU%77^Y#G8THjClw|Lm=GOITCzD!<6rs3UZ6v?83>)F7RJ~Tre{<6S-hHB$R{T
zgdDpC61Gr&diIj4VamF9|8K2LzI3k4G#4zZ^B1qSB~MHC^S3Ymk#SSE$im`3kgsU&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_batched_gemm_scale_softmax_gemm_permute.inc.DF3E33E757137FA7.idx b/.cache/clangd/index/run_batched_gemm_scale_softmax_gemm_permute.inc.DF3E33E757137FA7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..58c5a11015338b3006f4efe6c1a0b66f52e77bbb
GIT binary patch
literal 6114
zcmb7|d013O62Pa1*E$1_0cQr#0R}k~kW&N|<i0)-#DJ(o5Mf562!aYI;BpCLR4|H4
z)I?*9IW$4z5zr)NP>Ea`BT*xoxE^SttMOP9QM1u)vfq1M`zL-M^8KiK-PK)PRb8DD
z9~*ncMk0xt8(*?$UQsqDkw}KopTeT0IZx>C2#`ohwi`Bt<hF%dMt^?x^WKw&xh^}r
zCPeI5V;jL&e;BvCsei@~4S!lLKcv(D=bAHLU6ga}-EF$)mai7s_PHMKJ^xMB#eMk;
zj?3=-5&6ps??ZFak2^Hf^&e@Q|9s)<@|%Y)-giAe;w^K#pVoh}#W!>J-8bca-Mjvo
zxASz_w1=j>_eO*qT(x`G@axB)9NJ*IAY}T1Bi7n|T!X4(Tk)tpIWhA+C!};cPJBBm
za$VP%rTjHJtFkXH&E5Fn_4W;q=DuFO<oCQ$bIz@M`}Y1BuG<%6?JHZBH0tE2&5;uV
zj|Fe2pZ{u3>^i%vr8m1vl&9TwcUL>ql^5KQ6qYX1IncGQlXvRp6kO=$B|{|=V#1jW
z{rf+OTnZ*b=teD__eNp0wL}6B6O+PEMBg>x$XT#G3wFYX9O(hO9<XP80qia?HW_QS
zcjp;hy^JSW-dW=WHJ)VqW(P3R`RWkqgY;ocRvuK2s3N2ylF2GVDiKwKe>DUO(jf=>
z(<&`;n+9$3kh87#?vFWO%JC$}HD{tQ9#0m!Efh!RNv>NiqC(d~QB(GH%Cv?p(yPFe
zKA6x4qWNa$Z2s}Q$I^`o8el{Nj1tB+pRbva^O@aoDMy+>*#v4qz>!u^w1TC;iUO;I
z14-pRbOb*uzY)TpEx*(*8rK;2Jx@(m<%|~oOiflVbY+yQ&P9~3%14?7>H;KNtX_<$
zP*sRDi_}F(R;(^YRH7<Dnx*PeBr8*wiDbPoKA77Yx^gH_W}D3xm&=n(vrI%);8X=3
zYypj+Ze)jxCr2RS2ux=3Zm{Tv;fxK^7=tuP;NjnmE$iA7ae;1L8(6i0gCKB@*>!Yu
z?M|Kwu7me=2oePiQBF>p8*8ZGQ!xJ&MhXJA<Y%GZu06g%%9G3Bc^M`M0-ihs&xefN
zH8R~UJF7>gJV`f8w_<7wgBC_JDhMh-R1{H^z+|NnrHCq=E4-L&6R0<VmT?P&x8MzF
z2a9$Pj|op|rO~x)L)<@}7qYGLaQYCAoC3`$a1;jONITfJgEQj}u<gLy33i<zx|$<h
zpy>i9rr!;kZV(L}Q+(g_sj21!4XGlxqHsZtC&g~Xh)Uf`5tX@>iRAU~hPU%muI5um
z90vQt%n?4uiJe>JOU6@2TmjE3AO;~%9)ag0#y<U}UmKssZKET+3AS&t*z?Vh+9^yY
zzE1^x;MfP=f_45O-<tN8&QcSeG=qIJTm1kBG{Yo8kthA2?1$lmRJc62U2}^k&p`PM
z^9xYEU>xA{@}1u+e`%yM_JUI{vvETF_MJW3J2L2u-+}5o(6AZKf%+WSGCmLL^O$dd
z>IT+-3Jy=1e}XJe?r7e&{N)gy1j+&xY?vTf5TZ%4Nr-}F!H7a+A&5d{p&|uoA9v>t
z$$l*7NR}c?OjR7oR%9d6DRhYR3O%AG7}Ervf>|6fg2Kp>-T;p_f*Y1kGC~OEG$YIs
zI3)e5?s<`NJ&m~xvkVk-xk0%o$`%JLMpPV8jH0Y8q6|@`bESBcd9oSQn{mt?0hc4}
z)bOMOEIM$^8KfG6)I~5Z?6->#_V0ObnF&v7U|0=StOaE)_Mrik2F8&Ed%Uz^Uk<*W
zMeFM`Vh;H}r^jsrjMxC8DbZg&NIc5>*AR}(vYs_ku#h9^Ht8soGi)*t&9<J69F=LE
z=`DO0d(7}4@BSM*X(%>>ycsM60Z&>$-okkD&lAT@vRJenIGS!X*O}v|M6S@Atnht}
zTKX%v{0hOs2vge6C13P<)X8zATv0BD1xG3r6=G0uq*76ds0tKSZ2OWz9&FT|u{6+0
z*MRLBcA8R?l&*Igf{7(hBIS|D=qPy<q8L>S%A>KWSVVF1IHVabk4KcCN<f;4szgLd
z@+71=RX!C_vML#Arl?X7rOH#0=5+aVL^D-0k!G4IO{CPHGbjJ}PGZ=gFGQI|y?%ER
zvyc7#;mEVNO!R<b4^I0R!0`g(IVIZ*vc5UiN7H_-)V)^fEBJGOQ){K^j1AH;2I*ML
zsRkyWvv+UHXDNa66+Cfta1_fDPn;Z_5IH+IBXV_cMdaq-hREH)9g&C9!<tR)sq{qT
zrSwANt@K9Zqx3=KtMo<Wuk{zBh9?2q0FiX|tHSFf`Pb<wzX|G_tgg&cH<xIOUwr)c
zc3IX+Mb8avS5U1qirHw8S{kt2Af058irH*I#)?b3ZPwkQGjxJQCkuj{rG?9;KKmq`
z&QQ*ko3rsMxC%s-TqU9^;Hp66yp)fwjC{<U81yO1`@-d|`jeluQ=?Xc?P^w>mXw5k
zZ`b|9U*;T%w2$-=jN(X)B4!w)SVb(NIQuy9G;t(Bk$_~0ibO<7_DM)HS&@unDT)+C
zsrIR8_L+*ANS3BZLv$Bjzsu^)lIF`(o@#4$40=nJzl!LrTDCWYZb^x2NrbT0mHQGG
zP2bVeLA|sQyf(5l$dgUrvk5{4#erJnvl(+Ecr`NpRbM=wd#i17@Sp>CZQP%)ZV#(g
zQN!+l?;Up4shez4``h|g6K!~6CNo26hnMk)%w^_?6fy-Ol}v?5EmI?kR7Z*loF`H0
zC`2*Z7_ndCNvt*&QJgvsX~wJL5hZ97kY=Jb5mAyl329DMPeqihO-7n2+7v{o>QtmT
zT|FJqOzljhnWjxcbP!|*SsfJI2&Vt*x@ki`@UI~v7BVip)Y0q9EVEgt8_EgHi4?x!
zNq%5HqQbyJMCDH99!zs1s5j!SsRdkG*xj6`cYljhxNGVHi!RKyQcbN?OqY%QS+iGK
z*j=P&<1ly}X06WFWNEF-_N{re{x^Z93B+cPr%kMNE6a25hFIAwPVw9`=wtbQC~PX7
zd-^n;`X1QbV-4J{yaaA?@~WG(k;(GU8ZRuJUO~raiyZ++bmMi1^x=AO%Le-Hq$YTc
zec1}5TiGSygNIh`xBlh)f?l9&U{no^5fo`^2G?44h2e<-Tn(7pAg~RGXq_~qPAX2!
zlX~f-da2mx>@Gez)M}3L5?$h5@V(1o`NP#|+in*d?dZKR-!)%c22Tpy3Idrmi`^EB
z^tt;l)6*X(d(wS2g2sqnZHzG12;y_hfPAtM5`|$7aDOa16VUmP8u1tc9)oDzfoq}5
zw+)}VV4!}4mroP)c~TKoF@sTMR3)Nn2&#rDg7jcUuippt7D~F4S@JBgtk4%Bc{U=Q
zT!%<6*CT2I^Cl3pF;9(`8(D#Bu9rCnJy)Ki#y<e32duj~8rxX1Km6BTOO80YIHGRC
z$;Angvx_q#R~J`AZZ2+!++ExedDwV}S%4#+HlB#QY`hS8+jt}LvGGCVYvYT^Pva*x
z_Z;ze_7`J{BLU6<h}O{#Ln=NreRI^^Jy74>PeXA%sMoVvJMh4PZ`>z-{AirytTAWM
zrK_{)#JI2i>WZSfTnDywtermd-m2ukK2{V^!T;a0&pNMIy)tx~#en03v&17s9Uq#F
z!a*0RL!^(>i>GoRt;}d*Wp^O0c(;O|aI((*G~Tjrx!qX0=&PW)iZk95&^%$>6Sw|!
z-J#RAgI@HH+ONrFkCA5v7A@7(v7C6ty7gsQ*vK*(7d7Bf13rRH1L?u57RE9*fR_Pt
z8-%ric(eP-aQl+6j~t>igvr8q;j^FP-;dt1%X1g~^IEW7%lvsGv*`N7$Of}PFXi5*
z{ZDsze81C-78!1qsW1_JrQ*^NWpEjYW^=O<WpbH_PJ!$ch;ITseL#}6W9|T12j)(Y
qbz<%USr_JRkaaVrFVT6~1I4j$slL#R{+5!??zJnOww#<Vk^C3-1`X!`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_contraction_bilinear_example.inc.427C55016011CF17.idx b/.cache/clangd/index/run_contraction_bilinear_example.inc.427C55016011CF17.idx
new file mode 100755
index 0000000000000000000000000000000000000000..30da03eb6eb7be652c2aaefe998a676d2e2c5b35
GIT binary patch
literal 5596
zcmY+I3s_Xu7RP7l24~<nJQ{F-fstWCMg&rx@^-=#jyQsfA)?7k9%;cS59RR*idx~T
z6tV;+yIv83z|tVPQljD`AyYy^<(8USX4co!?`n6Cb<SSR_rbS+XYK#lueJ8t`z)Bl
z^Ou|?lBcugl&;QRn@%JWi39yxv(}X72NKB&`k-{*k=>u<_dUgJe&{oE)n|1__s_h$
z^ZE;AHrJ=-ulhEHFFsuJc}Y{~;<^TFKUnt8*rOYNPFUL3=)dJuO`6M{`ybL`(Z(<0
zkBsadC|%Y%H1F!-!1T6`bqlB5Is}q)QC0K)9eie+{T~f+Z?5>??zC|?qD?cNng1Zc
z@$~qL*Q3{0KbY;cWt#PS*CjJH{gJ9}D5<Kw-Sn_*(*0|@ZZ@k^$3@nEIja1@YF&xK
z>!Pakd09ie&*VD^o}Ibdwcm`oo=W!~Oz>a3$Fu(KPb611_(q&_<HgT!zueilzAo>p
zC3OprJ$WVR_m^L<|8>&l#uHyP=~RJRYvhIXdn@*T6me<tr!Jedz3<<8a73Y)xbOwr
zKcL$8T1~>AlKqmj+a8kIO-~jiAG7*p?3LuEjdA;YBx}~K&hVtOXI7;u{8pDd<RtbI
z32`6}+Q<JSvT-1>8xI5|rtFn#PCp@$jGu0MsOi&;T8<+*J~@7(kmUO0BFgj0LzEws
zKa<rg3oVOeR32K6s3NoiQRVDP1FJa%K|_GvGYnzF5XI~x5IF*~7>`2iD9jevdZ72a
zyHh)C#&e`VTM#Tf=18Ho5Rpl15@}+NYiM4p>f>>Qn8DEu3PD6jH^{reUSRo)PIEhq
z6#?TolIWD^B0T0ul2ejM_T3(<1NX!aTGI~L7kIHsh4zJrO!g*3MfOFAitURLmDrbv
z<g_-lE$FlJ1UPaWe2%kWE6-0~f2MiXC2HsahaPYf+Vw3e`r(_O=ekkDOW^tvc#DRP
zLj|v0?ysbVGr*sLWWf+%^W5M&(YrpTh6CV!049ru-`4dU`!#+q9sU+@-2z&{z>%}S
zpM^w$gQ8B@Twa%|ARHM2ZivpQ;CEyg?1#ab@dzkJz?JbRC`Z9V;NXKN%ZiOJrUM~Y
z!TTyq5e%WP&hNSPAbTb4iW!t<cv3KMbjcLm;3;sp{qcq)drKb4IFc+)mJ5$LvRJwp
z(X-NL5v52|5G|1|L6j;@MYL496j7Qq4bd{`GLgbPemwiUb^4DoLUQCeqC*jqE6+ug
zC(lEaFV9D01|Ks_6$U^^7s$JqYg*X5i&pu2rY#)N#lptIDP52%*dtp88@GJ;+94@N
zGHf$MkEdZ0QWzst;)p54gs3>A7*R<`38HdMxi8aoLQE&bGd>0UDVQ&C<i$l7e;TMw
z8gE(F@T(9Z7znur;ny&af&Ul;2=<uohwimEHx|?2-T)38KnyyLRDnYk=4x=LW~`s~
z_DsdpJMU4S-VT$tgBVb8AKpvo;(|6>6O!(n?k+S$NQQF;qD<#ZL|M*RI#x5=Ia?e#
zA-&+)3qgXF|ET8Ps=F(jXoovM)&X{c!NQ&$%%M5b39?RRf5xV2=gp)OJE(*Af?Y4D
znXn)DepoE<yi=bg4f)+aO$aHImx+-?NV&WmQH8t$QKh_6q(w%@DYth;TG?<!DN~{j
zy2xA*xyjrRsbnfd9x@L^YMC05r_2+PMy5ej2SIfZE{udD^$=9g_}Q~#8GUaacu3c5
z05k(2CYoi}H>Dg|`Qoc|Qk#KmrtvRsFD;;K!TVAhDBBo6cj(a3)&(<jR2=bB`-wil
z5r4Hmq5yROqCj;Zq9An;qF{9}qEKz97=IiI(}p1m*M=jC&_*DN(ncYQ)<z@JJLsL*
z@bun#aR=atK0=QIP@kw@#A@;mJks&rJks$IJks%rJkl8)3`l42HXxlL!hm#!L<7<p
z9gIk4^fn@$F~W#+#zdp2Tlwz2O6$4Fr>U2>f^RF#U<2p_z7OUJoSRlQ{Pv`6H>fY1
z2L3c835Gn=n&SDtAB?4`ri_$L6drS=oRlM~AQgxzNu@{y<DEX0S}%W_c3~^HZe`BG
zk#oSGV@a*BG<r~V_Vb5QLX_^xX+lecxVXC@a&vb>q;gjw@^JS+q;^*$@^tq^q;c1X
zRND8|jrHWhdFt*vLAMk9g{C=D2f8}U^`NU~T$wc-ap~43)8jZvY-XFkWws+*foCsj
zhsbuu)olr@lN&Dd(@EV9)3-w`tGENE@4&nh5_ZB;X5R(VcQN}bhVh%e-|=UdGu>Mh
zO7Tp>5f_CEA~%H_B9%gg$V1_QNUcyK@>F;t(kL{DLLEa<b_jC}Llo{9jwr%00#TG>
z6ryOyXhd}oQ^#h>vR}p2V?GGdgKP`8>`v~@*r8fL+JfyZplgAr*yOc=v<=(aK-b3X
zyJqo_cIi<DJy@pOrYo4NAfy1rS5Zh2qB2dHn4By*#;X%V$KuFI;7_u&y!*iyR~G0}
zex^aT4JK^^jnKrN-&QQEwo}ow96b^G9)M86Q1`~cfC-WR>7<5QP}H)qHo5M9zyC_I
z6*XLdCoZt`B`Cq>ubT(GlvsY|G1(OJ(@RomaK$QP#VIFbmU0%NIAt87*~;07;+64;
z<|yYNN>C;snx~v6lKJk`#J)+t9j6_w0f!ovOWs;Fml%>Ze@zbtIsQ4~BGD6@e=edt
z|2#zb;rZgV#&VfRFoRJTvE?$M?Plli{XaPAZhY(YfCkwH&}_g*fhy2cVXg*EHRc|e
z(Zd$uUt3f5-Yhz%qG4X7EeaOe<4Cc#7*UC~M5K3h-y|>nF=-k-fOUau7pMf0g{OBx
zAZ9}s9mn4pwqq>1WUq35u4TK8DoYd|)9p5@98pD71)|EimCJ>iooSb6?60+Rq)wb8
z%Mo{Bj^2TD5#`D95arAA5t(74861Qca<s>?F8b*29v+T&T=3r;Pq)zhyh2lfeE%?b
z9cI4&LEQ)aVOM6XpzD4S6c=HVpzq@Ml(xs-y6i!SpVm*z#Dw^3{SgId0}usj0}%yj
zgAfI4gAs-5Lj8oM2no}LAqv-pBZ|;PAd1pOA&S;TBhowT#nT)idS872Ye65Yk3*EK
zU&4rY=8=x~<&ln$<&ln0=8?|eY(P4LuL0={u?D0wBpZ;<=xjtfqpuO^jIl<fGbS66
zF3~knyb%(T<eDT>*U7aZOZNHv@;JkVG_#9?CBrFOK|Hf_q#cy)jJxG`1~(VvOX%|7
z1lODFKG3txu5p9?&STWUhah|i#7=Sa>KZ+a?=&NzA7PQfkx}4B*>RyS?p_x6%O^GT
zTs{Q+5In=?eHaYGFqiQN7)RJHOz4XijKV^}-f!pPee`+z`_!S^!M7cskxl@Af}N4h
zx|Mz{E0tT(&lq)(P{*<|p@}M?p7DTJ+1Bdlg?5%049N)>8lo{6l8Y!WBo9%(J|CSI
z%&^D|DNNZ7)4Exv8ZoWROoivX=(YWy-yFVl?tQdAX7c*Sd!Ek*7LJztn@<(yYVfJX
z+yh}fAa1UsH8+QkUAH<+*KfIWIVW`F+MM$}GY9&PT5%-PHWQ`DEZZzZ*|yn;Dm9h=
z*d}5QvsJz>OSzI%FnJ&Ca5Ka-<7D3onpP06ftJP5v|~OF{BgE8-@VrUYlH7#lO=hR
zbQ__=^!80M5M`1~L|G&YQ8vj&)C<yH5Wg=xy7BJ&rntT5=zKJSZ6o`FV#;2XZb`6f
zOqpvoS_)FjzrTJHg@Qz4!-_Iiv}Q#su{e&b%m2Qsp>^$T8(VIo+-ibUW@9axAlfBz
LR^(XGmKFaGANL4b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_contraction_scale_example.inc.619AD6EEB2674ADA.idx b/.cache/clangd/index/run_contraction_scale_example.inc.619AD6EEB2674ADA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cd3f3bec009a8543aecb1321a4678ad7b2a5c922
GIT binary patch
literal 4884
zcmY+I3s{s@8i4161I@rM3^c$11A{mb!>EW%Gcv${6aJwh7m<JoC|odD8&EEBsYp$s
z-Bp`RluLs~?KVV;ktoSVDH0I^wP~64)V6skGhN%tUANplW6%GeuX&!C=e*}T-?@I5
zb7u8&p6BkANbW6L-e@bYD<u+%#E*V!>uM@>Ady&UMdR~ByN_3mMR2W6yH$q11Lv&M
zrERK)Pje$K4@3JyW$8zpmq(si)xW3s(Tae{pr6NX{ObX?q2^<%kf56Km&U$5(tPan
z9P7L@>vBzB)gQRuSpUHE)}a#@S2>=4*;6wA1~GJgq%m5~H~8DOT+gkM+5h|Qq$>HF
z_~jA4Ka}45T6Q^ThiXuHYNz|V@{8*o@9&OJ>Atb>?{BtTaH_3tNqf(Is=VgW)S#I6
zp{jAc%#j`)@lD3u*D7{uE+z)OQyMkvzp$=-ZsN7?Nrh~-D=qM+%F4RC3bK-?4j<Zk
zVR!YNSEC=_(Xn?XsI}+hm+GLM!6z+k!&9fsIiH99UVf3gch`gGYJahM{jpVhblcC8
z<Ly^0e|b`pBUxDIE8$Np-jH{T{CKT4Z(r$4ov%u2n`~?6(!FzxD*m&ndEjkMB9};r
zAMu-e`$r;k1BtIY5RlZAud8*V+w30a-M8=5+U;H(DT*#yAS#mL=wd`A`jP}Dvl(qk
zjH--Ph^m)WuVAuq(2oPsoPZ@0kj&(hkTeNP8Bam#6f6_ibLOpYuf!bka^pyurtBV}
znj__!a*^gd9Ad2OSH0~<h#maxpb(l7(g!|$AQ#xDI$-6Y%qL=jBk2L@!9unCjc{9h
zd-^P)gUHL~3DtBEc{!pAc?F_Md8J4Jb;bkw_ooT%cMzfn+4Pmuam|y?rKhQ&AN=|u
zROtJzw_d$8{>l1Dx~^`}b;Dvoz>(L1e;rl{9NiEnJ(m!5h&Jp2T@P+}82H1ip{~CE
z+QlDMhSH&$LYkrkechqH4bPpOXra|_0RIN$34&O!2WMY3Kl?Tnv_n`sM2LbLO(UZ}
zrFYU18$ud1LN!N@0e=j#1=c5@^g7#QRnScv2W}kv1p!AUKt2J1j3+@c2_cN9Ksg2B
z0vmT_ja<4}zK)J(2c;eE5(FGQJw+eP71$&{;TY;{xFzLChBU)RsOHEj=_*8-(o95I
z(kw*T(riRI(i}vqrK=I;N^?arg<pSfy=UolDIrBZMWQtkQtVTVsKloPksYG#5F<>I
z5GVLJnI*b7$_dd}p6FzL;gT0P!79j;UYYG_JN8_kCr1jr3q%8_CTb{85gKu%(ol)0
zDzZw)WWAv41wG><z#oAufs;O5^Wpb1+jFQ7O@nC~5(NPvGhmuwobuh7s|TGukJ97X
z=(kbyA&%7e)re%7|6+n7=9|~4akT`sh<=iG>}tj^r+><Wko>^>Fg8R%U;(1Sz(OsP
zSp%)&DhL^XxdUt!{9m_SedNlzeeN9D53>E>E2z0xwI3oGzW}lqnEXDk&Grj9C!e9K
z7y#b^P_u?dfj<iO3%u%`_jArH_~*OsgxGv+q7f5P<x_>I+NWBi%;c-P)}E4Y1M(YB
zT$Aa4_t(!Rc~J9J%9LoP!LndPp|Vg!DwzsVxGWrzTBb(S3i?(s3Df3C8|d2@=fC&G
z+Oh6G|3VLB6e34K4D7{UZ7mpDSG|kwU<Ys=pkzICfwBvb|2wFV@q>MRNBUPMJgnr1
zR;?9}m?JUj7(_a?4pFQ+7SSU0B1CcOI7IQ9c+ojHqSxpV88ik&MvW0sf+hixNn=7}
z_A>{t<(Z?*qI+`0oM=X#W6n0OVJ$6w7Nlp1vLHQ6q6O($vMorD_v4WsAH^d*K9NUy
zd^RuYRTMq;@ryw_&eNkB27VZFgo#wv)IOT^@0U~QQQ3%Xj!?~!DpG~0npBIl!9C!V
z)U)J8YUYh0jp!^#fFEJ8d}E_|R`uTBe(@wk8K#^s^h`)_STLf{uuw#*FcqTkuy91`
zFttdHV_%+cCLc`GX>S7UCRiwR&yiNpwqkArZ5!j3vX2u_U)ox8+cdK~STJ<u%%pDM
zSw}sP)Wdl5Cnw5EbN-;D4mb@_({Q)YaLa)VTb|>C3EE>zTuZ7@&5<YLo<y`6;x>cG
ze_G+b^}A>Ow?skxR-qI#I7fmN!H7Z?p@>up6{2uOI3l$|jVRtfUJRSm(Eas@4E_d0
zMt>us1pfp?CVvy6R!C`O+u@1@DQ%edf^;v7g)Ywz>tLSEkuH#SVR;v5yI>L9E(b^*
zSndF=gUO#=%0toVTWe{q%J<G!Ftsv68M65bLj@vRq)m+S9O(t$UL1;sfFEMHYWK~L
zK3lCV_<?%J6G2aiF~0rAL&aNsRcB~*JLuZMD2%+P>*d%xlHPoc3LaNHE_#6=2CrYQ
z#`M<g`s@qoJv{MQo(FAx9PT{MQp5f)WOwhsx6n$xBuSYhZZ{#x%49^T%2Y&4l}iz&
zDbo-wQ!W!}|CN~Rv7j3#=&1F6^<or$;gOYOMNaEKXvQpBSR@_+y?QJxMpR-d5ij7b
zOK^rAGKJ;2F2R~U5O0LN*8;TXFVDM4onT|+M$tED0*|ag)DMgM*%rR^bV28Z`cV}P
zBo&&9dxXB}jYCr@(t!42-UHWj=F?p1gb*jFgk~;|b3#1k6;8+y_|O;ThE8Qx1)Y*D
z*_JI-)5w-wg{XREb%`K*t?2CH-tE5r9`w~BD-r`9M~Y>|h)QH7i0m-O4t_!x9G#WS
z38MVyo}&{>K3n_{J%TY%j6smlaF}mzJeYdvY%n2OjaE!3gv4lK5a~2JM6sG!M2j?w
z5XEWY5XEcb7YIWUqSxvX8MFpOMy(N1f;It>Nozu64m6AR5kky5b1WOdoN7)(lxHqv
zWC^q&J&VqQ^em|sq-V*qAU!^iM|!-DM|ylJkM#IF9_girq>B%M;qf{{;q%cy-j4T%
z4tC{q#d~Eph&M#nQ%Kpvcw~of&nCHj^!8K4G!EiuuBV830@8)`abyxKlg#coG6nn;
z%Y$QSSIf8`k8dM{zG?VzxKB_dw4)Uhu#)j4WKOb(L}*8=ryy65&&+%5=`H45Uzc?n
ziqNgB*iek9#8QHim>t&Gp+M-u_4=9L$D-;-;UgZO=ps&0>pTr1r$M|coeO;HR&z>3
zv&)~OH;QSO`g3#*qJCJ?4`Pt`blZiGN5AqIq`j??uHl41KVLpQk}xxNgzjXacOeSh
zR&Oh!>d5Nf1ncx(c1!v8xr0ABbhG>D((_5am(VkPw~zuvg`^OXl~@rCfOG)F*Vpx(
zdk&7}<fhU&bwEl7&STvW*-cZA@K*;li^v|#gTN267xJz11BYwUI;ZHGcY^m$_5!Ub
zf27nE`D<%_SNoK!Aa(ux`k5mXBoY}bda|MiD|(4VH!`?=!4AiUR~o&%xj8-_cSvPk
So{~F6xx|MRIac&$#s33W2DWSf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_conv2d_fwd_bias_perchannel_quantization_example.inc.453F2B26B4A84A7C.idx b/.cache/clangd/index/run_conv2d_fwd_bias_perchannel_quantization_example.inc.453F2B26B4A84A7C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..38c38a364559d5d1f432e3f0467b0b5256909c36
GIT binary patch
literal 5400
zcmchbX;f547RS5f3H9*VEDa(w(64FMrU9`#0?pEGgDCr=sDprtfuLfLML-2b2_{Ms
z6?KpVR78R#3M4Z|A%fr@Wz}c|HAq5G_Hm4w<C08~%zLkL@`*2USkABN)~%{r|GIVQ
z$N+zTePbTa_w9h}bqN_O2_BDUfPT_5(pJMMJl-~K9xuDJZ2$6<o>@kho<zI~7fZ`4
z!(X_3{^F}-Gndh0Q9Tt2r9odVNbblqY}hsKb60$+e9x&|b>NuCIpND6BuD$cz3yJV
zeQmyF;1j3!r<~ciCy5TsFW-1c^KNKG-|w+iT_;mc2O3_MFI|;Cws7&QqKmVNQv4tF
zClnoghb%S9OZ2kY;nlOGwsd@dZ?p17hFy0@*0I96-r<(B8QrVz+^SJpeADwZG^jk@
zs7W{OepT;qa-xpfy`Zhk;%>-n6RY_p&Bpd$ulO!5a((a0oZCq|jPsTpaaR<JPFG|K
zq|>i&_)#~i=uFrv>s>}k4|JC{Jls)|Tq4qZI{!iRaU(@kuvvL$Uh1p&JU%;<X!g$H
zq$5w_j%z1YAE>-7R*V)5e3%qAnpYU7ZaG|ZCnhN{?T*A{b3m~)>kH$WdD+chI2kAR
zgwa>4pU0-;_Vs+aY}_;DhvL_FC3`K)wN~?Y&E1o~W=~&e{`Z0mPjl(}52s(7|KCk+
zUPiYJdh`ZiuJxglFUwj+PM_QP_PIvE<t@W~GoQ&*wC=pvJ->FtF9kJEfBAi9!zn*S
zNq4VTbe(yT`^Uvc^RlZ0{1sosJFE#mxV}iW_On$|T?a+U!%^E+OZRC8^us#p?%uJ!
zP^i4H#c-r{`PLJGmTR^gGc<}deAhGMZ1&x;X+J6A@A(~G;r-;8pM&_x2}#A)HhJQM
z^o{G{{gEyTwTr5jz@}D;r_1BPYCu#2xBvg)ky8MV)CLgrI}Y7TS5}>+c+-#(HI*{`
z;}?%4w0Ki>kU?5PV)_<C<Zk)7x$SZXJwU+<U{V2SW+WI74pq1K4Qdfk>LQ)RG!xL#
z)p0gQF0L*tX}M0TKHM7rp)LhQmLdsbL_wOSZ8#EJiS0PDQdoI#B(af5Ig%=*Sf8Vt
zqZijFQ_55vIltws<jBRx#g`%Ny5~!`>a9I7iG66dpBo;L*69)yv=Q3K7*7<mHMe!*
z$j-vfnIn5!dpCx3mS3(8&-}+hEeZ;!2}MjZ1?}DKgE?|?cG57UUvIfieqcez#GE=d
zW=u1h6WR(nvLo%xIkGpjXX9!VI~cWiLEUzAkdFZU2sko2F*%?2&wu@J@HHf$Xr_qu
zjL2GK%lIN7ZA!EAAhM<Hxju=h1nZMHO4uk6FqfaJ&spZ$=Cb3=ZeB{|4S6NJ@ri|V
z#<M=P*<JECO<Njw*X*;|(+CJy3q-ZRfr*i2))Dm;U_F3F%=4J%&Fn|S-{!Fsw)Uas
z^@8>6)etwNC5_~H?LcnU^Vb_N*4F*?S9BF|i~csPZIx7U*gg6+>bM3>uYn+DT6?oY
z`W3fa*CK72nVMNJjtD3e3t4FqFdvxb11qLiHu&`3!NoU|khgZAZU-91o72mrubfw1
ziASn4n`UOh*i+EX#m<u>nT^bqA(ya<p&6-;C;ql%rRC#mYqeKCCXr0|BJ6-^K8=x>
zFUH7<Z-tSBFTp5>pM#Ta4l5G^=6L3?AxE?{Cz@HGYV_7v&GAE-Qz$4ml4HtOSSVQK
zt6oIa7o_rw&?X9iq5$|YqcthVoh()Llezl-XA8!XM-mUC3A4y7J;oCSv&n3hLN0ek
z+;aNQ)hPs&@nzVtIq{v?8x^_XjNO5V*%Pa}4@~caVCI+>-HQvn{YkKD5(z0v8IjtU
zGS-h)bsnhBb4O0W3qW;&<JcSS8zoLXGms(Qm5*h_T3}oYESV)F#cck*!sMd~fB8T!
zAFxS*f**i+AAoRX-$`S28pX@Or^s-j(L!scnS%R(_dXE9@anYm%&=dp)Ke)aC8X1t
zW(qnIju^=Z8AeWo6GkqC3roq7l|A%hqGb{hO3n7^Fu!)=wKf4y0rX8WTP5Jv0DWV8
z8o<*yp8@a;&h-G+<J<sX1IMYSx<j5>=9f)kZP^$4`Zp0>ZF~yOA?NV9=(v$?7`c<~
z7<rH$7%535MxLZ6MqZ>BM&6`1Mn0quMk-Q;k(yLvG?$!<k%rV@<V*UpwE68_ZoR8L
zijg8Zfw~hcVFFFTE}-t>IOpV=(y|#VZz28n0^?pFXAHNIr)!tB20TF>SAgjiE*ow?
z8N4p4tm_I=QS^*x_IgurnaMKtT2U~@G=|L`h~~vC;vyPr63bfVM;7}$`B^6&tu)do
z600hTfOipC$P8kZ1$3Dr3Kc~YUxk5M?~U&M-CYkIn_v!hMcf2#7`Y4FG4c?2V5Af%
zG4d36V&o<8!pK|TjggPQ2P2h0g^^mI#%Qi!E=C%G1|wgAFGeQ;|0FPGc6;FYirDRX
z<`bv0!EpmyxP0RJ@9NBa*E-bP3{=g)kFh_b>GiUV2%7%J-}3V0AR;hg$B!u2itWYh
zDH701X~kB9M|ZvUS>|K02`zgN(1Tp<b9CI;p?^s70!hl%(iN*nXVbGWfyu4pSVgK(
zD6p#Y2yi+A*b=w0@neIg{<ngWsG0!X1Y}HrC^!u0VUDXl+Op7iR2+j0JAk@_OVJd(
z34CvYdCb6Chk>YNk0WF#q2!q5uumK!*Bm(`Qbd6{Y~@%-JYCoN*41(3yAV0z(lrH(
z!R%tdKBjfOK3OF*!;+ApOf6$2PeEr-=TOG?moax{AN^3zP#ab4q=-$zsA?x^jKriE
zBP-GhBMB+ND2L2psh&UBv19!j9y*!7{(Dhhs;)VB!$68UT7asBQ(6Nbc3xiEAM(bP
z_RU9nhw}9IG$F0b7tY7j5g-h}D99v8%!H1DA#@0<NeYIUhq39W`TZ4@36HG%-uOOR
zZdb~A?M<sE@j_iR?5sR_|Li@alSM|0u+8oV-unTol&drBPJAi77>^>fK)3)i3ls)o
z6l@ZVMJSXG#Ud1L9?nMS+xLQWLccT2MiR)f%W`3YdOfW=;^j~4(%*RCTJ`<Df2lX3
zeB%?hOE!P7-DjcR#0`{*W-)<4H_(}8GY<vDv&0In&q`zE&yhqfVUsuobDeY9ah*2T
zU0a%T15uRQfVz!KL|y97wQcX(9Y-Ax0sRn2nT*u6VE6UHudWLxA)u~-^C=Cg?;fHo
zjY@(}u$1vdpsHX=7sodnetI^rRhx&D7GV^Dz2kd<_g)aeZ0NS-H&eSE=~lG1zow0z
z8Rvf6q@p*W@p5}P_OH1@sqkS&_cjJ}Paj&b6OFzIR2RA2(ifNUz+JO<+9Y<y_?)8q
z=LaZ%)&@KZ4uEL`K*a2Gcuh@|hd^`bk0^`UxN2v3U(K^J$Bx4~6r4%S)MeaJ(3CL6
z$c!+<NJt1VGAGP2vLGxl$_J=|XN$A38l_~sr*bihPBrkU27%1-C^!P>5$<(*-P~^b
ziC1<hQk)*Ghp9@R*2l<zHo(Y`HpECk3ovQ}_@;mRIV97|)20QjqY==J+<Q#HAwUmt
goR*Nhaw4MXX{*xF2UWZ&F#p$zmjlnELwUUa0c_Zn*#H0l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_conv2d_fwd_bias_perlayer_quantization_example.inc.41E3948246A8DC23.idx b/.cache/clangd/index/run_conv2d_fwd_bias_perlayer_quantization_example.inc.41E3948246A8DC23.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3805dff001680aa346f68e6ffc149e3c37f17353
GIT binary patch
literal 4970
zcmchbdsLH07Qi#`ffM*hfba^$JPH94Arc6NM<5Y}@=^?na;kL^#R`Sm1eZop3))h(
zT5Icj1s_NitB6!>#iHN?Xi)?a6tShOh%2JVamBV4yF<J4WqSG-|Hv7_`OVz9Gjr$8
zy*DJs#l)oZSga|F;!HUi%N9^97Rv$s<}Nd44g(g8Yt3Sr_H3IqH>oSYsX1xL^2ml+
z{wKFK`CV$N%3M_#Cw|mw{Cs9eR(k&Fwv$sv26Oimg|#exVlNBrzT9wsWxPtWoZmLf
zX5n|MGSA<iQBe7O1HCAduiBe<<WZ*TH|;}J^-x|yyna#T;_Y{&sT$vRw%qS>2wlEw
z##z1C1QvNazVl1zi~lzMBK*bfR^b=<jpMru6;bDJ)@zn^DhJZ%jBiln)igM!F0ko7
znEHcMdnCKhx;EOTYp(kA@nwpP>4D$P-e)K*KYP&W#|0}p(p+rQ5?U;FH*N0H-{i^t
z`-(a~bS``5aQ>S<EB+0uFZ!Ry1a)Svj$bI6>{iYTD38e~^R{#vSkK;GWn<SQXj#{s
zAMLzvrhjqP_zhpw90)buP06o|KEAIx=7+A&QUW!7#zQv4_XZsfUVU1T_wBpOrwuOM
zbyJk^<m#5RPZ|potuKo|nH(IO-tg0@wG|il8*eS05&2nlSI_?1O^scLH#NQ$)lsE&
zJHGnYvuV%T!ngDvI>(*aq!pHKZ*|?bz3un8U%oCVTh>>y{A8_QiJ;+HRl(1lPd0kY
zyBt4zAKz8e?WQuwZKZa5??0NIp87QEo_*4_)QFL|dY6>?ms{th{`PPV|6%e6J4Kr<
zil1)!inA;GtzoNZ_x+WvhqlchI+#%|e)|0zt|OKMSHt(@F3(AiL0+&1u9}*rzFR`G
zY*{R5XJyC#<AcSfELm3L00g&7+TGoHuJ_X{8zfK+sA3TO|356Y1z@qQk&rQCac+-2
zQumMZlbeHNTtLGez;y=@WA$J#ak{Q<>H|v(`b_i*z}*z|Rr&^+Nv@PLWT~_*-+Lu}
zBc!2FDD=caG!%(NQZtEP6Z@D+B9+L^<fHNlHIuK}S7Rnwm`rOXd8B*_ChN-Qsqb^&
zK01cIS~s76I3TaIv817=ho>LbL_;rum%o{$<E08r_Gg4SLEB=ljZPWo>5jY66zJt)
zCMjDgz~uDigOu5^m1~f7+W=}jd^fHa3buds<IWLCL7|(F(S^uM=#6JnP{bE8nh|-6
ze9dDLzJwT)Opq|!rl65yv^Pr{y^YK~_qNLhR=+0Wl`rfuY4E*&NjLBMo1+PxJ1ahy
zep&-4SPDd?K!$B3d~ctw0eJPH84*DdYJ46Mr!aIqywZcZGk6({ze=i&O9l!wMMzBs
zC&K|td-a}auu;vQ^|xic56M+~L*5uc12uqO0}}AEzV0RVrR~bMEhy;DcOQ=xQP3nX
zy@tu}!LvI%XSXduv+IDa4n*PE{x6nPDHb-QBPVgUb$7+`G?Y%1hMP$yl_@bzOe*bn
zF-$o6w=ISVhrWN?dcn}>U2%lOg^D;Lg2Ws#K@yIHAOpuhlDC0z1PU6$4H0+`h*Ax2
zVA4Kne_V06BoFOK;iO=^S^I0snGX$|&rtUY)rzTDI%?@*f1&owSYBVuw;w*v9$38B
zf`-}bY%bPB!yI-FL(}S8^zHsn8qtl)I5OrUlK=UQ&o+JF=q{Q8zZoQA)6cpyKfYsk
zqIL{XASFz1s9E~B7wI|;)Tcoh*7fG~kma8Kw_VT%xk_%Bd4r|ExfBSo*phkqH%nbN
z^&&A7*Tg(y8m@|7H3Lst^0+ce_2R)ZBoyZq=Y_jzxDG_D1Jf{P8guiKem<n5Xy`-v
z*x_y(`cl3G$tW2?a!Srnc5>Nm@eoxvhVvUtk=F#@UmvliU^zgIjSoh_BLFpmzXb40
zl8*v-l;mRo9wWH|zzQ=PPIXQj7OpKG!`@66dwN$;tygIp2CxG-W_u{vN`eB}fdmDy
zg9uWwRRjgIg9)0%o<xwEttKdh9YRnjJCvX>b{Ik7>~Mzi7i~~?XNK-T_k0oPE`mAO
zm^8cubeGJ$@>tfMVwVN&$m?za)V#4^A^Yp&bFaibMFZ7<Uu}*_`0y4sf986Ww?}_y
z*aB=@fC$fDb1X3@rMR^LxlocSiAj$%oDn>Od9F0ntMxHhBMp;-lbNx#$vYyS-W!*T
zju7h<OXTH`K*UF2CZ6<hyyXXvHvb-ChpvVfK)Ah<rz9wl7f4VLFNh!&Peo8LFPNZ7
zyh#M9d1`_}cp(IZ@<Is;<Ao6v&I>1KKj7>K&e%}fo~OOJhASAofvgExOlsM!JXxN%
zR#}O<PXhHxpuzGbQQaD$KEdvlXU+ZD2~@nk=-(&?ihae*g&f#05;-?={0g+u8vr$?
zt~fC0EbHy}{1thCpU{se!TiPkL<y#NO+l1kTY>*pz*KQ%HJ><~>kUpsK2Qb3RlpA$
zfrcFbH7Cr6Hm#oN{75nn30?rY3+5b1!+J2K9z<io<FdY#xkGwC^!N?#1}2>#QW}D=
zPa%poBx0(hp5<H#ZX86q3qW)MV1g+PH-f;8fJt1H-I4F@a!VRb#4??X@z4|VZUr9L
z$aS<rr6^m-gh*71vPA@m*<ykuYzaXIwt=BDoCnv6-pyj6RsLEfpV_0U*m>PSjs{Kv
z^(nK*RC8cUeQNKtSMD)qH*re}>_0t+JV5KACGLE($7F)0xK5E^W79BN5Y1GbwI8LG
zW&Fph=auiguA?l7)qE~<3_0SGV@uhwAHKbV+$7#9o;com5V0Oi#|Lb5@i}@zekmQ<
zMCYL+#G*W+2%73Tm9R;SAcnEYcON8-n|{O5gswT;Cwn5++hi=)zqp%|`^rsq#eV9|
z7u<ocSH5twXa2f1ku$lY@5)aYfDM7(l|sZs5E_aD#47WcBuK)9SQ;7?Mn<wl+N7*I
zXA_U?S_^cw<`B}VJ7VqK<#QMfTnDH*;#kreySeG(swR&yI7w^Yg{9Tn8~0H`kCO65
zkcxFtC@;^sWahT&yTg6&TeFZiz3%inQz_H12t*VCJw8x}@T^tm1o0KLx4#C1+b-u{
zuF#4{qq4$R!F;F0%K;nwE&R{y&FZd+>$K}nTaV<w2I{ZP!J}vXvY$etw%Cs0kPove
zQbKuAmyxboAgTpHc-4oWR}2?VxVG1thHjLb4L$@7`4pcZcgme056Xie0VN>F1WioI
z^SC0+bMSWAY-F)wpeY8icvTwS1gJ5&Vx+d&d$-242l<V?*q-nW2eAV|j$%iGcw!zw
ze?8XAd(-kX;nF&^@(Cb50VbGNZU?a4oDo^ZjO+!Y2G2Dv%tik(VOhYRL|b029$xv3
G#rhvC57{UH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_conv2d_fwd_perchannel_quantization_example.inc.4B57FDA755059955.idx b/.cache/clangd/index/run_conv2d_fwd_perchannel_quantization_example.inc.4B57FDA755059955.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9a5ce3c3d5e16cfbb98c4038734d576361803d78
GIT binary patch
literal 4974
zcmchbYgALm7J$#d5hrj6j|dtb$w?tVNB~2EA>ol2ff62K5WCu{RJ=kF0ZZaVkZQ#j
zO6A(BShXT5D@*BB#Hv*&O7THaETYyozOIzkx7^<1Z54Zm_MUUb)nELP6<~ccd-m+v
zd(Ympl5R>!$Pp04gjuGNg6w%&BtZ~v=+`>WGW%CR5Z(?1QL?2vV`|#<v7@@u`t84{
z%m_KKzEg3cvw8NyaufaNoTX-RUc|nb>!#C_et(|7b36ZlFgh-&Ykq&fwstL*mf611
zl%;46d-Tl5(RCnmd$}T{yl-9N!I;z)zrSA-plK_tQjQkpuZ*9s%8U-m(|&izZ{EMd
z+Xmh#-!@`f(?)Z$_{RH&u4f%{Z~NvpSM>Dz{`AH`S%JdvneXw?;)_KOXI?JLckp=s
zX=>&D!hyQ3zCt2n?3J6%`xc&bl3CZA(#!jjyLzNH$iKVaVSfwpr9;TJrWc2O57tha
zU0T=}mNDkJ-y(CWlltk19WTo2Ha%49ZTr^HtG`!`O|fR%66`}e1a8Yd-R!+qyUnfU
zS!Lesi3f|nNL6ees8qj@)lc1@rLLfF$-@>@$NC>V6;Yr4;msc=dW3{d3#y!cd-|i2
ze$UU7Te34hQk>Sj>8>-5JSOz0o0f1fXsz$k%lu}3#p$9}^RJ073M(qsYK(JL8e^?#
zYUj%tUBBIX;(jSjzh%#&@1AEId~vrlPq@&OrybPm^beNZX)5zPwf4uw7u3^F*`2T6
zE{teNs@^?b5S;rm|FCJ3V!e9ar-vSxvL-lHPte8qEplGBTJd1@GSk!6tey3t@vCLs
z`SHsI^K)nY_@CuY8$QTdclMOsps6!`yyvXU^)Kwf8-`tkwX|Sn0?LFXaADjG&E2gO
z;Y1M7d4#j@k3R&Dv?oT41Q0w+PS70pA+}M3BNC_uWGx8)|389f2MC@660&6HSns$Z
zb?>!~?^4NJ00lRJ(VGCxs0U9{CK}q}?%9(tFf4E^GfcuDWl%Ura;2Old!<u-?U|XY
zAq7Puksl*OK`||sazv|XEk_cmB$%T>bs#nuqzQ`V=7J-FbsWiK<r5fkXngkeM=o=A
z55rc6uGR++gBu;~Dd^|zr(iTu(BH>Dgd=HyRKw87_uA_x6~D6)E%9>l5;MaTltxNZ
zIFg0Rj10Lpi3(KJiD!p)8R_T645MA3zc)uxcPZ<iQPb~Dc`LDT2@1$HKwkrLCY+;>
zE#9#I=EgxtLeUrzs|%68D1gyLLa|WHYDN?w4&vq{LJ2k}kxST<lF-Vxx^j}%04uxB
z>%z$*;(m1YYv0&=+>`mAOV@1rOKyrht~RWbesTnm&<4adAY*(an!nBPE%5I{D`K^=
zG0b_y{BrFoc;*fo&JpCWDK2TY<UU*+y#lGp;pez9(*C_o-#V%@Gybw|z|LTGZRERy
zXrdJeTR}3jZIIVi*Sd?!Id&xUl6ZwOibz=MQ~Cx&ihIBGY<%lNE?RvY$c}S?4NI%L
z?^z_@{g?f8qW;r!9J2a{9M2bF(Gl~-7}0zhBMD!Ek%e!;DZs);m4uck3+o|7Z(H7F
zOw~WS__$%$)?#E)m~R-Hn)=_SG%YXUuSLUU>asXS+BkPth)CZwoZFYsvix!W!<<?>
z3KsARTo_FhEaVlkl-k~Hz8Lbf1BFM%m$50H{`bRy)#mUaE4qQO8>BE%$mq#Ty0ke(
zKa46cMQ1KLGWT&W(sdZ<4+A5kYx?=fQooREp2$EKr3+SeY{0_?M2uMOw6e=}qpNz6
zSgA`XdnYNlD1Ol-=2W?l8;$Dc_kKY_Nu!ednPCd90x_#VD#Noa*5b7P?lh1T3?u`c
znPCbBkwF;ANEt?QQqEF-y6qa>Pqq!?{6!_PXMIkeA9Nt$KLPsin1hjUH$WeZ_W-yD
z=LP^9aNY~xUYr{NY~;A;;MLRt(bAe>*j94<PVWNp+;<cOl{_V%^G7%@93vG^g^`-4
z#wdapfzdeLIE*wr4Mvf?NQ|^REk-(?4x=bu6h_g!XqL)mRcda{)_#WKd=kh`0t4er
z3bq4TJI5dH$=g!nnRO9m-Fe_~9;g_>a^9i6Q_q-wMiVVS*uoVU=J0lofA)IRw}*Zx
zcm+6J0b*wTl07K}Q)<qAi&AKkdJ@|nDQJ!`v)7e^>4x+)Mk56$M@(ktmZpCe`|}SY
zt>_4*Q6{V}uLLnG!DMF9$4U0f9<BL3(iw$Dpv2NXTo8_tN}$3>El^_=A&9_eoM0SA
z8i59*NI@h<T7ecLoj`|ClpqSDXhAeab%0+7JQzn+KbtXqiHpxr1m?)+upNH0@<4s@
zQe`6={tD>70&$G|R^zQ`kvZA<wYR|S`N?FGdBwj_4WxB6`xx3$IT$-N)^`Ci^g5ug
zbG>B86AxMMeZSvO284)0uqRk39m>{mRMpjC*fVPb2-yJGSDfw0KirP?Mx>w|H~{Da
zK*2bHf;RwtgX5i5iza(Kl1xK_CxGk(X9NX1fT9D0G7I<0`ld|nH!INPw|H56m@Go1
zw5TvjvarvnM)H?}GZ7t6knVC2Uk=z}O2H}+UIp03)p#p*{-!Z$LtQmEDtIEJZQryj
z;X78lxQ{?jQJ#n`k?1MP6Jtd4XpAI02}Ty4g{3C`y|XJm$Rp4$ufCF-wip^Vo_7mI
z6NiBQ5SKB{eAwOoc5muy&zP^PT(&ND{p2Xh0E4#yi@DL;h*2CJhh6Ukp9HqQep4~S
zmi?3eo!4%8+a;Td=sG%k7(L>Id$(=Rjg}siCds3cvEx;Om`cEALx*SJ?tQ^0W+I=A
z_a2XlP4J$8Q9K=w`Q%NXH(8&YT9!O=;`i<)D9i<c1z}8<bz17p&+iskU%RW$xW-+6
z?($Ig+Bfd<D_glFcCyRRy;6wAGLAs^Dpbr?5DL;pI)R&$s3dHOrJyy`%1U-iJC$um
z9g~n>j{(^+u7sR3eBlsqJ#ZJAxC-d2AegBo=MvX+e*9&p_b`&=T;H*|&HC?uM6Ywy
zluv^3j4l%OWkWm17n<)5^nK(&plnJUmB_x7DOdqwDu9_e&?V8=Bd*HnGsw1Ai@`O|
z)|U(P^iWj}(}l${^0$tdt~%eJu?h`00eur!Bkp9*yB%r#)M*%Jd63sIMJq6#M7oXv
z^zmWN@?c%VK#lxttpf$ekYgN~y(uUpg&28}UKn|k-Wd6iJ{Xn4Qubx@xIyRl<eKd*
z<d|BZtp)MSt`xip=$l+iV8($X^M%8o4;5nYo34P((Is0@-ni1PYR)`2+6^Oj+8rYS
zEx_p2^S=BcwKzIT+J;Qp59s||=1{N)&^;UzmhAkjAuFtwzgf{YCWIXfsiK0DcAf4c
Gi2nf=-h_St

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_conv2d_fwd_perlayer_quantization_example.inc.3D388FA1A8D19B72.idx b/.cache/clangd/index/run_conv2d_fwd_perlayer_quantization_example.inc.3D388FA1A8D19B72.idx
new file mode 100755
index 0000000000000000000000000000000000000000..55a32555569d9ef71d1d3e3389e85129d7ef0c52
GIT binary patch
literal 4160
zcmchaX;>3U8ppe&iQUj7K!RMM2__*x2qc922=RdR02Mrzk-<eY1_j0eJW!bN*7d++
z7*P>Wk*K)Hp^gfpj*cRz2qL2+>MXJ%Ga$;tI3BBmv&GI<r}yy_U&vEQKfkKC-m3b)
z^;TD6Cq_mFT4LC^`4h9!bW7&q7=~HFZ|0JWg<Sx{%1kjVt8n*>gsBgef?HGjO^$Dw
zp*StD8}%<;+v$s!N?o3}WaP!4Ta=W2ruFo=fxc6D8-s6ivW|WI=w-+LCx;`i<o&FQ
zHYoES?RlNeH6^ku8?^cPV_x1UeNY^E<;^EYTcXp>mI@*`DGRPAtgTyIr}#A|WVEcA
zzbL>A_g>vN{(Hr$&nx{hF8<{Mw(MXxol(F1;MUuHx@w2oUb|)0|JXjy_I+)=b=+LD
z_R_eYq~XU?yGA@7z2W(Zt<MXZ42=_A+su5uluMo6Bc3ZYDax&;RnlZ|tA5ABV_%Fl
z)IFLO9v9QI<87Uvcf%##-d~$@7o2SU{p)L*)yh~|Vr=e>Nzoj)`(nPQ&c-1>=%7{6
zn=KyNqqBqW1`DFIBR(7-h?zX=`n09h(@pz-Ofh|0wqj4noGja09+mt4>MZZPwW~O{
zVa*oPD+d!U`?}4lf2><qaph3@xy4FNY{|opLm##^JnP!lpbu_68Gb+i%Xj;We`{^`
zFMaCV7kTrSl!%+%e?052iEuBfX%C&9$o1X+MD~gH_xV}ZEGm+u^4r_Yix*P`fx9<+
zwXyl7e~;b0^V5Q^wyL|tMe|c*biIqOZ>U@KYT#<Ez~zYc=bR^>j<bIF@0>BaBHdDd
zvyX}X@`sZ%cHSNBz4?O)_mzz<b46NGB#b&&nL{@PEy<!Va||OaOfBsH`-73V31&JH
z5Ww$n?~3~J`Ti7U1`Tono(q)!dtl@UfRP+%l%ZRYS#JT|{q4IkO=@=@pomStb`uaG
z_n<FoiuUsOZWEjkdyAE*8Yd(wiI0(_Dk(!ID)X`fjY-=$6d|OAVq`=SP9i6%k!UIH
zVI;AyILJtn07<Bkq#EfsM4YNOajSSK<wKYcr|G+Ay&hF&BPc@TDDp&}D8fbPqCkY6
zBh3Dk5lS-np@?c=UJaa)n3gqkntJzy=0RU0MGmMJ^hLNhA`&EY&zv=(YAqD=G=PPV
zdfR@#I{(x^_YM#^A+!@Ry&*}2uBbOoIN3Wfz96|exf$DNdz#fo%V}l>IFZavwlrEM
zyCyUJ99rwrv7S)fun+9$-?x0fbjQAr_7!c@R*@&&c>&-=9&pYB?nong`9W<xkaWQw
zermq}v>s>fnxM@@V+X8`=f^YK)1@{exp#HwM(8G<8*hcIC7m_(X1;S~d^D`<Q4ik(
zfwKo-LpiW72b0mT-ct|vj7pc6f7AmR05g<4f9?z{nC`!k!kODY5{8c}WC`!Ybz%|C
zrCFrs>e=L~XF`M%`Vf5>S^%Ut{cJ?x?`GC)Naub9C-<4}XX8=z*u}>bba>5Bav}2k
zh8L;53l6~jn?xq@kSB_mN6up?cIi3U?iioJ;VOZBC5S>A%y=|s(!GMH@F65$!kxes
zV#R||a6q1l7i=6b57^`Z8hK8BQ8m{0efKM97A^>v7|n9Ra-tC}%*b3e_3cru2}Ou;
zu?4E82njA>krbCQlp1SzAnM024`IgXS(;|yjXMJzoY)6oA<-Ohq5#0c;{AZw&*nlv
z6tcMp5JhY*21K!u)4yqpeoe0{8p7PNu5@&+#9MBeQG^HS!8Ix%C*>^iBt2QAAQdd~
zBE4ATO?tCPNh(>SB2_H%A$?e+Ce;k(oLE#?G-~c$xLs-hEUrkfHRRdK_{NDZVM7_P
zFEb|F+7nS}GmBd4VIU)XBbc+DA|m}FncYnh6Z|GHv@Uj!=H=6onQ((>1RB<hBRed6
zI?`e5B$M^ecl;4(4&~u{Fy|{p$oX;>dGbA3q~I%9<i+=5kvHF)MM}PsMJm3EMLv8V
z7ODAa7VQGuUBCvZvinV9!djkiP_9IIB6GkMs7{wHTc@gm)uq6{6oerA;<4?abj)On
zVL!V|whU<EHBKwxjJ^S|Ft4$aJ{$MW9`QREJ{c`zz2MzMZmbvldZ1Vjm|KD2!gi~R
zoqka;fF(dw0zA?BDB>)DMS!@{y=!KCe~c$g3%!G0h++##rB}0Pl0FJi^{4lJO1AN=
zad_h(g-r6p8wcscBATRGq$l+Z)o{C;H-5edgG>GY>rYLgwqoxcD-YOU0R9GJRBE}z
z?Q3zJ(Ze2<Z*zIYt1Wk4gyIJ{2C(D~atvZoh%kh?QC!-TXwdy4=@|C8UcP5gV@($q
z4&j8mYJJ;q;-~YE;5N|;w5$bv5%$Ft6z_&nvhr$=t4UCUK*vCqQLtk$i$aB=j0V@&
zPaZkt2kR^t$Rt^kH}Z5dqb%nAleElXm!{Ej{EzQ>y}`piaFaOu?`t*Fd4sv)DO4gM
zV6J#MG0931w3=p)CW=V*N@m(xw8uHF4`s)olBWPH)+ljXChWMm_2f;*Aw;{i_Q~t6
zRU8aPJ4_J9+=VD&W!TCXG_`y5x2A1!QR7F+*?wVSn?+CJX4r59_#ZJQXUCi+j|0c<
zG9SWBp5;}%E|NDNfU?-)wq|H-im=1&SY(gev&aE=VCY3fu(<DmVHR9UE(pj4<B%*A
zaSp)3oNoh{np_J)vkKwnwG>(U8b`7cS+U4kWX;frr_swoL!_6X??XUz2*{1T&j8|#
i@nps_bg6R(@28m=^D^Q8LTm(KfAaWSnY3MoVgCW}2Pkg<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_convnd_activ_example.inc.F2FB9ED6F9BA4B8A.idx b/.cache/clangd/index/run_convnd_activ_example.inc.F2FB9ED6F9BA4B8A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2c90c4a8a9a3e4d708066bfc33b1a1aa5bf40d83
GIT binary patch
literal 1426
zcmWIYbaU%sWngel@vO*AElFfyU|<jg;^LB`%%wp30T5Tr3GLg>*KEM!`d3p@f?=Us
z>Wc~UJADc~uk4v}VT<Ua`?DLpLO#Wwv)#s%yE(S?qxRt~D{2$YnsF^P+_-Q3k~gwj
zO6v|C@QY;+<`A1QBa7=!uj)a|>jwp|OQxS$_CX?Uib~a%`8h`62airD+@vBu#U|1(
zzvQCO{@x(g_&+*<YfBAn|H@BjuP}^r^!cj1^5n6qJ5O2tzCCp0Ol#A>veiF<p|~<P
z$qDGv(pBfaRc$eu%frCQz`)MS$}0{g7&w_38CX~u*%_Z}f7+<A?dwh+1`ePCR#sL~
zFu}md#KgeG22@a#npPYK3<Q>$pQ}DRR7+swVGj`sQGyxFn=WOZxmxA`BMUnhAC~~k
zIFZ<wp<x~ynt-aE6`ZwUs>N6IHP)<(=>ZCJDRDW$RLhCpTsK*D!d{>ulJb(8Fx6_y
zwM^z)MEwP-wvx5-pB(!-f5nMUx@O;xrl{u2Wdnun0_~z;!d7x0Pi3<ueq?50XXj&w
zyM%?EgO3A=$H~VD=UI8Ks&AOHJ{o9}u9~hH%wAtbsaXH50q=prBAz0FFyWxm9qLm9
zjV6KYm9-0i*&BPwqEcD+mKac2QdCk7raFG*%5AHIEYsO}*p=j!lwra=?8@@WNIVsJ
z6(pX9p@u0!rKX`K5>Lxe3yG&~s14_(ifpZrEB&<$<X&D8F_-~q4)-?S&hDK96n5ly
zl!plyG)q)wZ}}Gm6!uW?aDWLHR@j}FId$nT8w<OXl$10~7|2tVQbXbyOPL_??4<0G
zc%D*TNW2iKP#6zJGjc(aA24NtKqKe&_0eZMKLed&CTj+d;BGJGmbd-?Er7z>LfVor
z&9DFmrjjtBFkzVR!W?fFufVoPpcLrJ<qk_zi@N^WMcc_%0S$3caDk`5rF;7X_dPN7
zVd7yoR4~+m8N$Qvrr-wWZH&2Lw5W$)2&k8xlbsi4<mRyLte3nb*nq-T0alSP_v~<+
zTCvXI`3-g!b|n=h4VaNE?8++2NIVr46(pX9nT8cYrKXuC5>Lxa3yG&~rVZzvQ?Bl4
zvJ2)1x`4}<D-h<eiw>*U)NZS<017)PIBCJmg-J7ULGm;t&>48x`C0fyV2a+Yvr>M4
zYiSrWF!6|sAf-hk1tWNX@UXipxWjp$f0^Z{-u6ldT4xqu76A)~e-|cMJUCN%6)5be
z;0O=#e@}hOY^+re0z=$D!2mAI!|tl!3g;E2<|KlgS6q}_%nhWO*e6(*DIICgI>o>M
E0Q@2^MgRZ+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_convnd_fwd_example.inc.6C6FF11BF5A73D93.idx b/.cache/clangd/index/run_convnd_fwd_example.inc.6C6FF11BF5A73D93.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6c2e11a0c444142d28dc61e986cf558749dfc478
GIT binary patch
literal 1368
zcmWIYbaM+}Wngel@vO*AElFfyU|<jg;^LB`%zhw!7>Fz8<eqZmYcddMeQ4<=uG1jk
ze)r~S6KBaK8_!G#b9COZThAo%;itd%n>R)Ho1M5Du<O@a!=qK(>K3HD`W5_aZPTaa
zy*C|M*JiQgN561R>F#Lm>v3MzS2UyWQtM>pg+ERyF1}uP$tO3Z{IQD6#Q7$E6F*iI
z#-&a>_tm@orQ+_F`<6Qk@c-w$C->L2m!Y^aH%SrbCZF_`*F^-bN%JtUF)*+Kfi#$4
z;ACQCU}j-tWn?HyO)HK9dWU7^=c*46)e@L_Sc8Oul;9d!!^FeX5xgkzC^(PTpkr3n
zRKwdqEg|9|vIs>H;t@!^81Wc5PcZaSM!^!E4U9ajG43(GFfAgnFGIsTHZ%c+?Rf0O
zVZw5vH`h&8ov@dYg_T#HR}&_zwp`0(zD3kuplTCY6aW7BiJZz_PFvO2EOg}P58Db9
zw&u51hN;$m9VZav?obC5j&P0ef(e_HhM)VNn?4^X9AXw?1rxUNTvgvNXMHqKSW!*U
z3?}R{`OAyIBH?>L;TX*reVDMT$XmgM!pv15vo$01V8Sq(kqZ*qz=!~Wpwb=cQv;1A
z0aaVbS_Ht{9kFTa1=|^3?}5T0njyL{&9Rp(DwTC_i2;RqMS105!tpCtZd)B>na<9`
zDkd)m_azUjxV$(LPeNV-i6>(yV+u2XhgH^47KtZkD2K$8H<X9-@}H{y-B;K&7vwPK
z5O<gXg%x(^WlmlC%f`aWC(8$S7?7vRrv_65<mvMnAo0xkERcBid=5xFcRmjouV!u5
zoVj9&SAljq^Ehk4gc~`xua7?C`57o|AZq}R@@_BYmbd-?Etq&%<%Q%WVXAppHN`Y#
zVc`p-nORsvghGU2ida}9gd&i5F+wqL-lDF*cF}gSpm??Cv4^L}rF;7X_dPN70a~Z3
zpsE8iiHFsZ$58^tTYCF_l99wiHK5)I^9UQ5@D8`B73&<H-(Y896;lz@fC;m(imQkt
z@g!6vka#j?GVqki!YXSfi^P*NlSAUko5{m@7adlysohpz0d#>akF6NYs>|M-MKf%Y
zo&beoEMn|n!q2roZPeKIbth1morPTlCj4%lmGb*rOT&PERZ~zy@~ab%6Wp)=p8A&A
xSgRfcDpFBU(T3^eVRhhffQMR9YEB|3`4<-@7jpw?Cf28w6;b^wc9bwM001}e#RvcZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_convnd_fwd_max_example.inc.4C55B2EACEAC1C07.idx b/.cache/clangd/index/run_convnd_fwd_max_example.inc.4C55B2EACEAC1C07.idx
new file mode 100755
index 0000000000000000000000000000000000000000..18b6d7116683392e924cb6f3eb933c2f2806d599
GIT binary patch
literal 6314
zcmai2c~}%j7N1u%BQVW5FgPHzsI<zU98o~!6a$WkfP!2mF+`LlsDM!kXyO6FMzVqi
z6%zy@pbp-<DwjzJns_6MOHhMnV7+jSh{l86>M_+_`AfgC_<mil-h1`#?wQ(<kjlX_
zS@1k<*23gZK4fGvnI-<e<de+5RfA0Sqk&A8^-amlun2n(i~fkt;9RRIAEzwV<X+qu
zSRZ*M{V3nnxyL48_7QJw&>xQ+5`Ddx$8}i`A9)PSZ%Xs23m9GUrQ!3uYmPQ+UoNgr
z`@uFpKX=cO;a)4|ZTP~`BWdTU(OaDBHP;H_=Hxx>^w0K-n*Z;~(G}OL^Z!`3AlP+d
z&H8Eio=|(pcjr^N@77xtPZ#ZLE2!vo4;>hoyKPbA##8ZD9jQ@6PQI$<XDna4{UJJL
zvDm$()wwF;m`&>O&!%g?@ETau+1m5)k9k)nMc$2HzcD<$dc^xteg~{No$cOfwwvhC
z>1KONnPK2`&8uta`gO03Tk{5zE^DWM?LT<z;jh>B7T<Yy*Y;KFuP5y+bscp3{=qpH
zt=i;0c^|h0r+!i4vcNv#=h1sF)psa*I&Rc{RTi?d_fUV;(wWP1&QI86_uy98!=}}7
zbq$|E(Wus!udU<zPdgvJ7?*k8enF8pxEQ6LNIV+i_QRS1+vpn?OaE<Hu|O5pryBpW
z>DJl9116~M<{v28)luT!=k>+hwyQ6EV|ynxR^@(~wd-+A{HgxCu@#|%UYfr%BXPfM
z$)^jGo?vf5@yHP$>}YydflOsG&Oz>AtzTrugBW|f%Ypo(r)k=x9WAWP2@{kf!*ay^
z`A24qKxS))i8M=>KI-lZvsWMmCSv8Rm41;K<G9dsaa~@T(0DuKm8r}O6R1^cC;g%)
za8zRsZqGh#{H}j$sRH3MoQd4TLcd5HA~Hq<JU$UV&s4@@0+mYTs9*F14udg=`3ugx
z+c5u^HU%=r1TJzHEBzwe<c$t#K5P1YezZQWK-QQ*DOalWizFi5z~-!x{QW8gGQk82
zxxzxfNCcm$N5sAJ3~sWrp_stU(9KrA$c&8)hv@CFVO40-uhU(RD3B5pn90p7^^3%z
za?e2I-ij-%0^yY5Oy#DQ`b8o*z&u@ecKpdl3S@x^jO0d^`b8q>;AU0?UP;PUARA0z
zDYsPV7lFV^Z|djwyNbtl&mE>fDoh}k%LnNfnXw!vrr=RoX7c<cE!YOc_sHap&+)eK
zkL6Z@^(q)fZfoG7*j;`$snC$+VuNF+lE*9;_io%Y2_^U>_(>>nQevQlnqf>ccoNK%
zKFR*?O@1!0oDL1vp;08p$X_ra5%xwSnW=lhw$u`jtN26<c(i~wVWnJ`KVXi|Zbu9k
zF*QO<9y46z)JPH8{<v!MKfAWR#zeJXRtpvqSuOCjFpA*8>Sv+HPWM@2F}71|!$`;t
zKIf5rYFAJ@V!0xOf1?RK%M~M4F&av+-K7&5vU|SCZxv+sYf9z*gCh|Z<T%VRhB(5}
zrgPsW=b~f4a61tGX^24_msbS-3?sT3mdm!twkGmeF2^E=LMtp*P-vyaN)b8PfAf{P
zJ-kWIa$1GfMq-dSJaLSKn!&so)DrFh-wp_oa4!V+Lb!wr(6j;+O|a8QrmICB<B4y@
zUj_a*Aot>Tx{NFLNqCJn{M3G7gnpFWc8jXZ9v@?KX>GNl6&bD>?3<<dBgMtQ<V{ie
zGbXT*)rkKR5g{zM1^6vel#!Byqg$vqf1bk8&0k0SJ@J@pLCyEqU)|PW-O9kU3@nLu
zq#_xbys1Qz*#7|kyLdO2(|BuwNbzC065va~oiJg!M(}I|UxFt(K0f`i{KLnF43=h9
z2BL!)P6uW>Iy7~_>!i>W_9y8;%rF7R=^%kH_SLm+&u`eYdl0@~KJZ1feoh(}Nxx7<
zs3?@>BE2K2G>z6ok0T6mdb`GrmrxB%t^v_?zL%me{eHeMO3rZY5Zw+jgplDn0RJdw
z-iLAbL1e^my)e8NMo9E|$R`g?kZ=JSS%Acp40w3Cci*n<%W&?^vzsT@B$i9IPaaMV
z1-yu<&&x2|h<)hd;UYRCq$#gA?ZNzVOO~q!^J<{Nz6Q)|XnYE6PJun)$8u*teFiuQ
zp96<;Aj)I83*dACMoRl#VADmPzX9qS^!Z!haEm^F2b}KE=X=1p2Q(7Dez57M`8@~q
zbDG}^aCkxUdj(FfXnq6WJV5hXjYh9VZiIiR$?BrtqK<D6615!o{bI08EB$`M`Lz#B
z4Oq@pZR$w670a2a%^XQ{V>wF)%VDGsg?n#YtQu2Tgn1eH7zLC575>wOsLNyidy-+e
zShHBsats$|7DqWh-YlL%iDrqS4@mLF=_@pQDt#{@{;HV1GndwlEQ#_ITpy$knoe#Q
zv$VInt?!SoaM-vzx>FUd3NckkN$TqCg%b}uem1)ipRPtm)ksA)-n;1Y>F_aERQSA`
zo0|_|%i^MJvjaI2viI-wXl3kD;k@lI+sQ;{l5JRbVGfaWKLkS#NjaS|C#R#z_RBl?
zG`|t}GSLkwy$2?bdwKUMmcC-x3bAjpTs;KTLj=+0Z<$M$MLgdbh`T~86KhHy<IH8^
zC=}1cQz(&16cL$-u%5y=B<aaFBL0kM<FrjlBd6S)wj4K)b{Nr4SGNvu>wrm;<a^+8
zPbyL@PIj+em@Mt*qVc&XNWulkuK<aSnhapL<O<UlrO#8~|0kLUm)u}YI59HIZ3ccb
ztyB|uHc2Hn!^cFImmoWin`#TVw!j!7JtND%b;#+9uW=2lL=KgbR|V``iFo2=0lQTq
zcN%+FA|DzDRH8r{hgPDg1ZQ0S?&v?B9(oV^+0xC@lVnb2e1`f#x^D(fwpQ?JrTc6*
zOzZ|~wpxY!SD_h1@nvCC&T8VkA~8SpWOX!YbXhGo?_@J==M*fLVw_?|s97%6IF&+a
z#%UB<WW0z%i;WjkDBU=nLb;&HrLDXkH0%G0z1GvAPzahrdcP373Z?y=$WgVrqz`QI
z9j2(Ki1Qp#nlX7ZS^pUVEVl<K_aJ9tQI^|_lzVBs4=MLac(w7-{fTF_PcZih>Iq`k
zTD`I^xB14+dcknB)$b9*t=T($VN6-u&p40PfbSZq%dy;A@LLOFRbjb2@XMp|IvBqW
zrVy>y&fMbv<hMafaNgz&%b}`99R$=#J+k08BgNH2u0q~h8(EJgS-5_hLGHs1|MM}z
z{!^yd5vJIG%2W!aDbpylNV$kYi<OHhl&(yt(0T}551|s*LI^CBFe&bM$P!*#sMPxg
z@jJy*z;dO)mx2djQydfmbJ_=zgt228|G6}aX;d0c%Fqx+2sLy>ct=n}$1Ke(YUqg7
z#8RjRylX(DEqONgy*1`TE@Qh@fk%~8XSPqi?&}y8X8P8-dq-LrGc9V}EnExA!Kz#u
zzeuOU*B4&OsOxV){5I<PTY=vyUB9wtn^Rf-t1<(IyA1qgsRJ|I18{jDbzp{DhsLgx
zI&kG9o6&df4S#_{Zr<2=?+|n1vBo`FoRx&v4Zk8pmCs_NDMsROS#|WbQ_qvWo!B1y
zLEsz30{Lycx9apMr$D?BW)<c@xbL2y>EM6zm+Lq@XDMg7kjE?+t&A2??fo&+GEQs}
z64P31J(C3Io~%a0$c2?5__5qft|Bj)`$u=jBtDLE#RkxUoeqW&rULfRfp{|kPt!@M
zL*@h{lQ-|TVxpVam<%POH_It(6hlc@U^&)?9ZKE?@Pxvdx)8tN357FrAwi5AM1&?m
z3cq9Ec}$9y1K0hRmky3d5!$9Ub`H@Px6QD)_a&6zpAaRX#Gu4zf)394#j~=|+zQ)5
z2gW)O`{}<L@*7e<_}m#AI6@iWMd<OEt&9}W;cYAasqUM72%qK?_zy|S)P<B}?F;Pr
z1;=m+8c~A8>d$bcXk@8W{p-!{wysN0m0=;9kliLE&Q>h971?c-aDBoi<1IsD{*DFh
z1+%>%#zBLb(S@Jl?u5TJU7GTb+P(O;{{j}p2lHV>RXl2i2Z<B^cQcJs-Ev}01R`7B
zQ6ic!eZnkaf*+O}7##omqGqg14Y<{SXpJB8=T;~GXT;67>T=xbLiOHXo33FoD~GKV
z<BH`PAfN#viJ%joe!v&hOzFhV&15p2$YYlKg!zO*%a~;p%3`uabSgJ&(2RDgEUalY
zxK&G^8qQ?yj(T}};gYv%de&k}$4m1E6W>aBet1s78vj}5_)6tqR}PMnwzV*}R(h3f
zJM6M2u%urrRN0fjcZ!KFOb(%~5GNgM2ak4o;Od63ZlGfP9{AsrCP@6L2SuzxbEWgS
zXmTzR-FB(rc7ONhav|2Thh>XSW4WUca1=x`GST4bDSVKks^@OR*HhKA1?*ZtEHzhD
z-w$L549ymb;JA5IyG-^<o+znQgo29@)wqh0M==tM<dq{fy5C=UjK>w|ueBNH&0kuT
zF#gffuWmNVdbY^o40d!C%&3BRBC$(*u5rTomLquF<&OdX6J>3Ut7g1ZR;DVZQe7?H
zHC|Loi1P3nm`zx++<q|M4}%Hrx-s&jvwZ0dA!o+$<D{JNo9-uCwI@HuaC&pka2!Xo
zd}mQ_{?WeCBD%k6Uw@gVwHzxO!i1O*0W25Fgi>fKGnGPGMoXdnV6-2oL9ZE%nk7ss
zJucM3w?xXfbq0LBfuuTr%78yj<4yzq3XQ*I_-cltKg{q)Xnca<f1>e0L;jE<p?~?Z
t=6knmoBOc0tjAi1lD34uWu|>7=(HsBqb2y;vkY-LMSYBeZDWK?_J8->2u1(^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_gemm_add_add_fastgelu_example.inc.E8AFCCD92C8D9967.idx b/.cache/clangd/index/run_gemm_add_add_fastgelu_example.inc.E8AFCCD92C8D9967.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b762d577797145913ee694d80e6a6ddeaa6af398
GIT binary patch
literal 3718
zcmaKvdsI_L9>+6yVMub5Kp-T!L5R8JWoScU!o$3LPy`92f-DyBMb&B}MT1xy>T-$)
zWV=!i^-*wrRK!Q^;^S0%s#a;cuJu7tSKY2%mv-G%&r$26wsrMbcZYRn?)<f$gE^ny
z_cy=aZ|?kN=H9ZRf`VVmMWTswik8*QT{4>yiA0I$uVG1K!FlxiHV}!HZEWrM{IxH#
z67qks`t?lw{sULn+HM|Irq8O|SW-K?iT8-Nwm0t;=O0bZ{r#_wjCoLyIq$vI!H-t=
zJ&|v$j=%>KBSTjdstb+N&s@It;O~#Q!q@v|_rEf8!{ouAKG|FB=vh^7Y?~bF=2q17
z{b^@^-cWMiwKMN9UH48snl$6>jh7{xuJyRqJ4L%LCtggHuG+D(EtH%bJ#b^bVg6Ft
zq~O7ug`cm9<<7a+&pJ}B%G<fma>DBPSan%G;l{r!FY<>SGk&{s&yCY-)%}v)@7xtN
zyir$MfM$04MSYjD&HFwp5{pD|3y9qUO#k~4iJ2HtY#e~#q5aCi-uBU}Sy4Q4iiwR$
z`|p1u@d)tDAu@MfgFz+|jc`am*fLPNO2R^8hS5sKK~si_BD2YC6Eus-LXkDg>J~JQ
z+cQ=muiHzJFW2W2G=Hw2qQJyJA=css`)}PIy-&ih&~C7g!Dbd_8L}vH@D7)tIe8}`
z*(+Smf)7(ai(z0`8BWGd3~Z7&C16T8lvX!pUE$eyltpeIh0Q37Jc}ZS!eJ0Jr@~2*
zTkfVjv*p<o<tTC}&s;?=A=NyiG6MW301Mv&y0<_U4zKFXJT9Hn2awPTRIPwSOL@1*
z%XIl$kkG7X)?k0lSB~hZUE?nzb0<)Bf^;mTx24q9Z9Cf*!@?M242jRet$^7I6u5=-
zlGTe^HOloP7^s+_cma=QVA9wmvNQ}_JaVxfYomrD8(`YwlPKj%kh~J8vB1wsu$|2x
zXhec`k~W0R7H)IGwo8t190S!UY8^H+P?MseNUPS46f{oFQKU=JnFWoH=gBzhj=8?>
z=8viY1Dm;KGJ9KN<K62&7VA(cQ@N=oJT|j^NB_4^{;?aG!?JK9Hrof*?NOaNJ`<Vi
zN7j>x+8@4g<k-)J-N;<Vm1SV>v1#{DhvR17moU&MH<Dgrph<3`$SgNgWRY7avdXO#
zh3#Pn4$i=+JW6R%Q`ACfo+z~_(P_~tta*2z-MsSfnh#=GsF7+&yQ5BaWsxB)%ywl{
z<TZGyO&keC$oBJIn0n!d!BtZj26lj~4nWn{0el_6kMj=e#T_6>^Gh9o)D`&t(w#lK
zHZMi{IG7mJVmB6+B$iNAsx0LMEu;()D)_YZ&cfUCwxEfz32chs#ExP~Paw_7<_g+k
z@nVHQ5pjf2$%%7QE<JbeWGn+cau3PEK(E|Okx%X;Eyh5<+)q(J9w43-jd!k}|MAh<
z;aye`Qz7K+0Zb1_!n<tNr&m3v6+`bMw{UtmQ*dk2HIe0=n|yFtTI1s#sNG~~GBURX
zk%pzAf9)KH9)O4O$b_+8#!Hcp@loVw{1gS40HH-AR0HvGGq#~D<y<*w>P4*=kLi|f
zwj%Si^l77UmgR?z8>W{h79w*C$Y=p}EVPszZr?xWaU!ZpE7hv7nS~CwBOp+YJBK14
z?|V+r8u$i6Eq%edicROgMB%D3t2}~Jb!Ih1HP6>j?OPM7nIU*i&zw#?SDf2@*cSR^
z_?dB1j;hY-c2f1^x^pS=^L|nf3md!*1vq%CyLRHt6}m6bQ|JY3FW~UHc8)IqH5Z@O
zqC9G;n%Z@Cmz^rq?Q&D(F?i@*$45v%to#1(n}=e-4=B$9<pLv4zwY6z=`D)XOUPVr
ztIxsa^&58Dq4(&1WUf?I67Q`Gi<!w&+O9%$JdHMz*ywl~ZKBAWX?`9@VxT3{LXkDw
zN;VU^KFh<TvC;KuYa&mn=dQ~7&zt;XH2t7HNGd>^TVFy^siBl~GYdn85JjE9)d})&
zA-yl<q?|rAY8;x9O=%lR%<a><uQeV^MTdyn;HIwsY(q9hIeZSe{#lsI=Mw6Dq|FOH
z^gSKEf?R4hj*G4!mxdy3ww63y26EXPMY?gi@xnOX#glO#Rz^K*zb^hAD!NKpm4==6
zTCZ1B-=C6+Zq_IlO~qzJjwmS_(d1|lYx|g=>IU2=@1jVppu823?%wyU|MhN7Xg+$-
z9*?IK`?JvN@lxdT`=$t*-|wd=P!y=a+WwlqID1wplyT^dPi&N7GYd^(6GdjRnIenW
zLXlN$r6?+n5;|11M*N<>;x}j#%jL_-8Xj|%&53=X=N2-{%4O9!=dm*JtM;Z!2{Nnm
z)y3F6$TzjO_{$X#Rp>NkV>7DIX`;yNGLuGSp~YpP$m+LF5XOb8!!reH8q-9U_RncY
zVlUhD1E_*7z;yv0yM2?=|1=uVM~A!H6eLF?>TXjBMWvQfvaL{eTS63dg0Y>#UHQ$m
zBH5dk=mB&K=w8%S;yf(m^LaA!zpjs8_VBfYGpL;U!uoRTG*tA}k-Wh_e2lh$PO76$
z8hyMz5j(LkU<goW%MyMGDQ4(9^UDLK?(a}sO}>VtyY=dxioB!=e?&DY)Cy{Bk~)dd
ztu=do-efzw1+9jfQA>n$8b(8rmeEqgF&srYMn_R9lS)x@e6t+K#3w6|;13$~y)*pp
zE<1MHllZuIQT^N`j=WsnyYjVk`%%f4=>4w)q;&}Izc^rw<k@G>vnV1KZU@rsAOq+5
p<+DN42Jf<s|9Ae8xeI3xk8g;)+#p52F>u>aX>v;c-_Ymge*k%-s_Fm$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_gemm_add_multiply_example.inc.3C87D5BB1B980F33.idx b/.cache/clangd/index/run_gemm_add_multiply_example.inc.3C87D5BB1B980F33.idx
new file mode 100755
index 0000000000000000000000000000000000000000..93c15d941ceba4c11199e5c3b244f706a6a417a0
GIT binary patch
literal 3652
zcmaKudr(x@9mnroFC2FFdUwHnKVesvg<Zhiy(=s)1$hQjg1AxSp_03rMF`5nRfuCp
z3u4B!NquCn1{-lSnrO7Co%jeQEr~A_Git2K3`vbR)@QXg7$-@4s6FSNzv>L$`FwxB
z^E<!uyT9MrJEyF;IESOCSwh*W=1}`$hN7rg_}A9n8omghois(Q+R^>`r;QSivge=9
zK=0f)xzy{6Ke+z!MG*7hwH1+RXMQDIZTu|K(pcB|!ef4TWPI((&b}R^DXxU^r}naY
zr`JBp_<Q)ruGwkdtbQkH@3`wo;0<%<wnN(vjlEcLZ>)dm?OoR|E}MP8oNadG?tDx5
zcX^I}!L1F(qG`9~KeY7kTQHPdbEWazuT!X7LsQ!I#^DnioC~+G(e0;(_3uA%>d(D@
ztx|U5m)xxZ`}~8m%SVk(hVQ`Dw)T?JSsy)r{>Y=VpYA_C^i%uYeb&}f8^dj@n;VMZ
zxs>z!Ufc5cH4jV0P!y0vNeut}kCINJsi<fg&?C$}T6=%Fgr%g=VLctSp0@q(K}i|O
zGpE+jlD5CG6g9~s-?wwPVUvsn#;L|+92}UYng}vG%r4PpaaahlPP6(&o5%0T63OfL
z66DMD<%%{wlPAbO%U_CY(HBnMxtnrG#;_pKmgqw=3mi5FK}q%`UbH#wPE3=R8ZyG~
znLeQz5aL2g<b<9Q*|U<tYEJxP!A4rXk&Z`&Qy#T6`?i*j#K7&j#00boZqGRgN>U`*
zM4MCLB*?|N2v0ZXCMZ>rN_eIz(lDu)B*wMUd7slPSY}){4Fy;4b`Qyg12mLd)GauY
z?!2e-%)9xQq1>hJvLbJ5kEWq{V5CRF0!f0zjbs*d(@Zz5K!=Q%{bqT$E^gZ-2CxNe
z5prXIGE0eH3j=D*HSx%nFxt8uSe~AS8?I2T&?4C`C|x6k!yQo8lxQlD>@@T%2d;RQ
zMKeI9Q5liU0JTO<kVdO9i#DxROOQ^ZOA>8*jUIPo@}5v`-`Wpr;0q3$!nh=^j*bU6
zAD0_pDtU%H2U?pxy!D{^%+Nw8SDUNxdD6#MA0Pa_v=7QThMcL$JInI$?6T;^4`mE6
zaz@-d3@~vfg3O$mAPZ+9$jVs>3dRLhC^!QIP9SW8LvRwdY=Nk~Fke`JY&rW!`ZpZg
zyf2CcYPlMRgFVSR@Ddidc{f2`o0pi=+8MQY!g4Onz4Yk9rb>nZk!cYR+67W>B#)r-
z$ZUeDBRHQK_paPO_-_A7IHDP`8G7Wzg3Q=Vf&y^?gJ{c&%feKAqWgZ+-6cEWfuu^Q
zMs$)UNbyUDHm@{OwADy!6e871YcW0l`NhgBssA_~#Q+cI!MiZP%Xta%aXzB=c#bE?
z&-t<E+$RNpkbD^znP^^Kg|9c*yu6wq&2&u$%7+2k=~{wxGj;i*O<$m&g(%o@|K=Bu
zf7miHp#_pkB(tE0W_su;XhIjAxSoAhG5UMho*`ohcZrakU;kjYYY%Mwh$Vu@KBPLb
z%F^+45A>E<WVnJa*LGbU(7t{H%3(zqpD*0nwzA^M-kGra9>z0SOu@@|3Gy*Mf_R1}
z$j|sOEuW+wj)|@xfV<=xa`7ZB@4h@}T-l!t<pOI#3fg7O(IH!1RctAgo2E9UBYEAN
zW4(ukr?D_6ja;KfG7FOYNi#%B^`{c#v-?s-Tb;cQ)4Bt*n}ctEaSDbjbQflePDSn_
zf{N3MiHRw$D6SVhOWY;cbN$8r$6OU3O;k)P*Am$|{Z1mEG=CaFyq(ASu%Irdt{4UH
z_BYI0xZe0@Sc6`g?WGOqt-qOHOxIrixdHB@lB<Y8PvjGcL|wd#Adk&M4!XS-U;A6{
z9eed?6#WS9Qx#X0h_>H4zNl`U!gK}7)vjvXwcEDuO$Rw2ABJ+iIv;z#(^SqpSJ`tN
zz#B2eh-1SWF~vlX*=<fk8!*7)wh&~UZpCxVfZ*KVi=t=96~a~8^+26`<F@@NJbp%e
z2F?K9EAg2G1#AJ_;w;FrWf9cHcLh-J-PIYIvu6@!!UHA5C77|?JHKzNW6%U|FPF_l
zd{Ve=Zh}(nsrZwE1!?v)Oz+L>%-(Xf{C6;JVO*gF#r+`pX0YaAr5paJ3I@T1WJn$X
z$A$EqfUng<%)gq4{ih$mNG(+@xcLr!lNagJRlE%A;K}y<0!3nhH`_~)FV9EZAABB9
zkiX1di`E^k{i}D^CPf^8pLtT_WF)h|BsCFamYNB&NG$|erB;H1lY_Bh6hSIr+tCG^
zrEkRtH^Kug<(A@-5At(_sG?nWpe(DB)u6S5bEJ*wUGrp6Ru!tsk$l14)w?dQN->eI
z*Ek)?Fki2UATw{qUBv<mZz0H<XDtxd1#5$gL<;#rxaZN}(K7Ur{Oqfc`qFYd>Y6Kl
z7^b$yu*Qy7f2BG3bLWis&WR3nWZ+8;cBmthpnx-gM+0`KGmD_EtS<Z``_)+4<X4@Y
ze}q2<x)*fw&@L>{&(T*S`g&W;s_|Ep=U^(;rPWnP9xXe6JaFMJe}ZkSlk14PLNC`7
z<hS{W+o9fGk5d`_HtD5dN8h(FuDVc-tAD5QU@$PH;4_#Lt7VC`O05#poy`Zof7>;(
z6TTu9qmrSVU?fIOkcQC^q-C@O=@=bBdPYxBI3~=Ym=D;jfgfX{e+$#pmT2)8=1-qo
naBR<6wd4P$NwtQW7Eip<w$_*0<nSp0`rTi@acS(kS19T~`<(L@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_gemm_example.inc.4868A822F4677354.idx b/.cache/clangd/index/run_gemm_example.inc.4868A822F4677354.idx
new file mode 100755
index 0000000000000000000000000000000000000000..96da48915bc0f269792f9a1f8fb76505afe7250e
GIT binary patch
literal 6828
zcmd6sc~q0f7Ql0d4<&pb0g1#&Br0Moi&BUf7G-gPC?Exk3w7UK1!)D9BEIU=9ziSB
zf{G;y6cj30C?Mpi1qBK!xPqeChea(yt=8=otsIv(Q71F~=lQRkn4I6aclNu@7%@38
zaKEicG-2`N#MmXP7H}ews2~2fa#j4Y0U#0$Ln2Y)-mI@{KS4vSlGpT?oH*b%^^fM!
z)fLrGF8B<*8n@d_fXmTe-Jg4*QWe{x8u#_U7nx6AfZIU3-Oguymprvy6WdsAhZ^G!
z&;HPT@G9Nu=7;NkOS-z_-k}9A>(?!c^19resQXgg^p}g9zSsNWf}MwEg(p=0YPs*o
zv(|m0lD}Hk3=3<RP<d`scXQf5Po&zcOWM1$#qvl|%u?q+qvxCp9NoU#rvAjtDDXeI
z?c3nC`R+bu&!&dw2CdbqHY^?Bx#0GUkG{-yxuBd=%6%5xn!eI%?fR*gKA0*0;pOR3
z`}Cih{V8*QS|1+w?dG1CZ!N}+u2AceSHy*SSs(o)j}NHWY18$nMOUwzy<6Jx?<%p$
z^@9(@$7VE~6RlhmyRaDVQz!&P1Rk>u7DN^z5%1>WHu}|<NZiL%WG)6CLdDzOd-=2}
zQxI8Ug$`uefo$IVh{PNc^)<sv;a4SgPL4l56byUMcO%noWcTKWD$VTM_8`lzpjQw{
zu#%aJ+2~habf{(HI@(>FX8Q;tORNxO8fEe7OC&ad*Tab7mn>cRm0To(`yo^PYbV|o
zb9@CFRDtYBQv(mFKtow}s6fLh>yfP<DOgtO(O{Ox=#hqHZ$0v1IYN)7u{>RmW)L=U
z;|p`nEKKbq@TK5T3XVh|@MYjo#&S8lUk)BrUISh=;7hp{ylPplgP=MHA##)WwLV@y
zb@?HIw{^80W87PQg6o7a<WT#RKAE}RGUoetZksIrVXIkoa8Qj2$M=A&2jpZ~9G`(?
z8HnL(B&kMLL~c=e;P`u{E2{Cb&H9>QGkN_d@NRwGj3iA?d0~mpWSepPE(qBLp`<nL
z9YQi#&Vt}9x~UwW4I$aAJQsp<S$Q4==dqj*A^EI+0fZD#wplvdIv#vafWX%xMJ;-t
zOvGltOFh517_guo4X;OTL?BC_JeK>$ccWO~ivfT76M?{&0RAvs2dX-7rScZgwm>-L
zZjg6_(Gdhb1<6yWV+gz+1?kZTRR0PZc?FH3+=fQBQI?-|OKLn2a0$=21r2OL!-(SG
zWZA;ledm+$iL`@rJ1{-mK->*vNdm|a^X)4($*nW+na7TgokZGme4KaOI7$^rUP1kH
z=r8fQzRKQHc-WJu|4B4}D2~u~H$RS6Rp4Qt!TZl(JP`=|ZN%M1G7^$pHWwW4c=l@%
zR_uX(Jzzr>Gf=+_WaLq@GjV}O4+S8O--CSips_@e<8#pP9Hb)b`ee<I#~y1lagU84
zZv=ZH5crFry$BJ6y_FA}V)`t2h~2LgB&A@qVh#rbNg2!KASq|L2S)XP(Oe8(F)D+a
z564#_Nfq@91FNdgdt_h(`&1!jSF6y>DilTJV~fsbrySi;C=vKnz#qn{(qTwC8>P}g
zlMeo*p@1DolfkkMoOLjo%CkX}&C2sYlgDxaXbM;^0_P&OeJ!YJS*vaaZ8K}tT_Epb
ztvVUWlmBkj&U$2=kKZ4y_eu_AC*Y&12b+2@<_Ups0GkHN<3BwxQt5Z(-HVhHL%vF{
zOgUDJ9pEb1UIk+&3#{67Z~5J&IXF^8TSd$02}BQ#9>I`%v?rwms{}?fVQ2!Qnc$wl
zXzoE4dytg2tU%%lWJ$Rai7P2<XL|Gfqw1dEeeM9e4%UEPg568XBr_rN*JXw=ezzfq
zHsnJ3Cou!eNvZKhG4#ZR@YazJPT?R@1Ckn0k`^3a3zAxv>p)URnS=w}<LmO|O^2cn
z+viTXtPAmb=W=A^>U&+ye?Hx5BJfcjQ4`5R1^y=DZXzk!B7y%Mala!eS>W`56Fp!2
zvRsV~CI|J)K@Ozh^wrz{n{&^X1CBoll9ONzrKEPiXJRPI$i@i#736S*o=Md7gX?d6
zHcN{;J_}>cLJ;YA?y5HT-$E8t;lQa!_IjEc4D72%enfFz;@(f^S6BUttwg4ld6RJk
z-cRl4O)O*K75VqNpgEm*n7c^0i)_il7N>1VPICNL<6GnNlKz#6%GJFE)&g%XGapQc
zktie%l&lrj49OHShHMqK3@H^#hU^vg45<_<hSW0kV6qT_*C;d&lzbFE4EZVi7@DA%
zz)+APh@oIbFhgOAFot53F^mT%*(W)Zy<QR;P`#n7U?JXXrK{3|E<4CI$b(FZD4UqS
zF0_r_*opi*k(Qd+UF3fkX~{_}i(i=#_Sbj*oPeE>s{sqr4=!_nt7W+kxH^_k0(X+~
zvh8W}_U~FU6i@OR47kR|_2-~?PJ>-c#Jo{0kLKOSitQk4r*Yem{T0t>3SJ(O|MQTC
zT<tsO%_~f7&Y1O?yAQ9R3FJ+5#sYr{w3len5LbCM^s(}btaok&pC75WpEIqW!T;-!
zLme`fp6eR!B-Q-6;RqJ&K~8&+H!+8e#UZhgS<NT=3j8hbxCKlM>jIB1mb<~Do8?W&
zV-sTZlaWU<%PGhsh2<RNor5Nmfs+GuGmDQ*?!=+92b4Wv^a@<RAY}$(xEjf-X~asd
zc(}OprzvCK8p={0Od8~MKGhZ<r?uMJi@KMO+Q*BySLy~Y-NCrW(b!Mg!K<CU%w2<t
z*T6VmgB4GD22;rV1^zi`pVNCpT0kJozVu=tuGh8#{xJ1g8VpIJ=~dv<K$FJSSGz#7
zi{(sk&SdMYEYM`J@?6m5vYZc^e3lErxsYvN1F9OfUb_I=3v9i13*@(0zutu8oBm#}
zS?cL^Z`YIgb2eHIYQ@{y3y!_)^*9eX<{{&C`I|?rUEcQ!t+CBzBUv_m0N69mX26OF
zmjiD-!QAD++?4R-n|%z$xl<_^YY#(l?o`HdIgBX>V{K%(vPSg4hvcvf#d$;qVz?Tq
zs%bHvci+ab^Y>xB*cj`<p`Jag25@L#xdnN)(5ij^ERC$Gai~9D<2JC~MuT4A(s(=H
zQ*9mCt};QI2~K1x0<Qz9j^#p-7P4Ff(jv+~ICe#T^f257S5sBUwh9d*9U0iI3aMG1
zQiY7K52Z7{P^>G6e1y*_4Wwycd_NKRbdaV~F75FBEK?r#F<yKP_|?!bXGj%(wJg^`
zXdTQaBbLo>bI(h)w8m>J1#v0O8HR@xaT&|yATDRQ3W=+bF(*}w5J!G}HupB3s=eA?
zLp?QE9jqaqT5&Yx=*RQc4>uF=i-B|(dx^~eX$H$#AkAVq8>HDR=Ylks<vft)v78Um
ze3lD9T0r@5;I70H|L!(iq+Lc1muc;Ec<J>s6V01$;ANOunc)Y|*MIoo!^+J_N1}F!
zcMpl&f-UAak{qXAQGM9a(Q8rD6MPJ9VABRn(Qpl1uR#FWc0=S*KL_LO{MhJ-#io<;
z?%-i#rp0_j)ski;%_DRyLLBY7c7{3bJq6rTz>5eB;cDVm2&Q}%v}YlVaP8%MwdmZj
zW$#?BJN137`snk<33%D3L4KOe)~%pzrDl6>*xJn-eWzKzdhc^Dr3Yy+Q};5K%fYvt
zeWvJvs2-S0w!)B^qB0P})o4^Tawqa;?e^1K%TFubxvjbwsl3VC-?)Ztt`iBJ$d*)%
zmz6)7UUeRF1`94B+e<Xb{wzEGBFS$^(yM^zR1c%b8V#3nrv{c=kZ%hzKJEM`!z}U1
zCswENsq{i%FPmoaP+%T1K7RdjeZ-<Os?6(ngw5c-nKj5%n3&2M<T=ouqXv0n?yBpa
z<J0(eZkIc+doG_{Z<6!Y%g?P_OFk}}zU3v}Bb8d^NyA;3I?R)VyY7@o(~#QkVH}4O
zh<_$^e-0-Qe;vz(Fuo8bkQ3tgA{bvp`A&36q?dGj10F4QKrB;M#SMrv(w!Y8kF%W5
z=ipGj6}YWnPI@!EK5?lm7Xw$!atUxHl<!?xz2fulEIe^XTflt_d#uIqK`|I#;SG<a
z5hXB-4s;&0=jky&KNKO^`t#gAcsJ7^C=Ikkk;CaLD4p{2nsX}M*u=g6cYXMOC6+HR
ggyWU*i&hT6zlcAxZ$$Cd*jt}~AQD~&YQy*c0EKt)`~Uy|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_gemm_test.inc.B7894CAD83BCCCF5.idx b/.cache/clangd/index/run_gemm_test.inc.B7894CAD83BCCCF5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..fc965a4f57429a883e279c1810d40af968ba75e3
GIT binary patch
literal 718
zcmWIYbaOk##K7R3;#rZKT9U}Zz`!5~#Kk2=nLB~>RUodI6WU|wf7n34&Gxf)Ud~dr
z@Iy_TLQ6VbR8=HyC?(hD`dK_bk>2#{{lc6afB&lUx!c}-efr~`jdz38EJdf)b6A%y
zzx<)>`;_U~6HkX^3U9me_PzY<y!*CqUI>J*XPoFK<h4xs+WqFY85b^GQdIr5x`&G=
z)io#f1$+6lmU%Om+cc|xzBqf$o+tTJF0S?u7k)h9x$5ly9uoF}BAwc6KCym2^VIve
zbdUMfE;jY3yY{c^7``9#jASUT%uSL9x<1mcW&Xjk2VZ#@*ccdCm|2(w!2|;*BO?<l
zGXq0WYFhCMMg|6sANn(AHN3Q9<YAFjmeq!76Px5|!!pCJ2`DV&B;@^-VZ)DGS~2Zb
zavx7+vn76HVqsC^Qxkxx2J)2ERAD@uRp;lm@0z?4s8`TQ5U$tPd{VxcZqEduFoz9?
zJ4`i<X5@l65a@Og$cU-FzF55X3CLYKvgR;V6CK?cW3K#50tyQ_3Bb)>nB&dj71;KO
z5$F<EHMrRuV{RBN>fsjxs#fDv<AoWr&m<^?)pgNppfIZ@t07GIoN{$XlU*=BNVTt8
zAWZn$^aC%ngsgjk!o23Zj<BG3&9SDR$!Y`0J8YV4CNR}7X%-e1Wf5gjm@trMsBDDb
zsfef`$*CBsz~z2zn|@Hw`t)X?xe_K4K``@*QgaeP;Zj_bT+9umnOI!B8qeR$i1%P%
F006#&7~KE>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_bwd_data_bias_relu_example.inc.1D7F06778A26066D.idx b/.cache/clangd/index/run_grouped_conv_bwd_data_bias_relu_example.inc.1D7F06778A26066D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ae56482ff0a3d73c515d6d35696b37d3bbe324cb
GIT binary patch
literal 5138
zcmZXY30PCd7JxHwfehRvKo+(@1A#0A6v7T6Y_cRQiULw4{lvN;qEtYop?yV)>$4T5
zRP85VeNre-5w#YyMS+TdD2Sj{sqJ$CYeihCrPjVTw7vJv*T?s<{Bvf`oH^&rnM3C4
z;^Hg?ELO}CUGb`=McD+4#j-))g+&D`x&VuH2n`hPtJ-xlcf`j&a&24vrC4RM?snky
z*<)W_7Kk!eWN(~E>A7roPkL;L`}GuM$&22`i%BDUCOfZvJg{@gYn8_<@AakGKh(5^
zls0G_-(B@wRA}_PJvH!Ajb>(O-1v~c&-uECk=37VIQ;13XzbaQ8DD-oBebu7Dtq61
ztcTkT@AE~8`Cj1xRm($$@&+95>>+0#c>Z=;F6;d-mX|r3_wP=X$4uPTmjwhjpZ;(`
z!w-e0*Y+Dm^96$auMAU9Js%hJB{_7YTWC$V?dx71X@BJ0I%qn>OWQ7UkBfe&Y&rR3
z$#`qsy?faQS8p9GpZDvh#_-?I+&VWEbn)Di^#vbT<?APZ$vjx&@L<TVd#L=;_w(W7
zGZOyO2mg3g@-sj7n)ii3^?zDpo0V-hHSe4X9Ifnn=idFjdoFc#jplyYn(0;lTz@so
zs^Yn5V?YsS?H<>=*K&O8=j#V^LUKDc9FFSOE!*zqcC%b6U6@uO(;f3Oa}VqGSf44c
zThXp=I%zEU(KdVA<=`6SqLOCEG!u(AKC;|3x^47onn!}@y-kO9CKU|4o~qojVUhZ1
zmDZ$dwJ5FY)YMpm8LZv>MZe9*?<~?jwKJET_>EQg_Nv9vD5~z{jjFms(FG*Sip7Ga
zW~P!qexN^rnK_&F^zOy2Q@2LmBv~X9*8!pq%>3&Y3lb(QOAA)P(q)BhC_zkptv;{4
zuy{M0gnS9#gH{X$5&=d+H=&%7iKHS0L$V;5h9P;793w@jB7%`A!<3N}5qqB*AK#KK
z0wg>E_$PprGTOZV%!6^yW=m5NiX~!SYCZ{_CC(VRxVoq*8GF2A&NsuBLrB*lz&->-
zlrHO&F01^i5(dnXE>ip&H6Q6BoiTFZxJVe8E9r{K+(<W!q#P+dQ((BCo{`X8hLm%y
zxFSY@v)VbFAvd)fMsj!gOh#6uDT-&P9O%kH8pZY(eT}^hQN!p^)j(7Y=zU#;*PALm
zD%t=68-aHt&{C&$S-(HB8@P|56%ip3^rn<T6U#r71m(~47YVNd@+y#1t7Y|`{<Xmi
zUZK^Uz@ZZ)QUZlzjrG5-_^vc3Ajh8LM5!Yn*Pe?J&w=O6$gJ(GT^O>lv!Tr=U?~ul
z0x>03&No{5_QBm~-`zm48`x0-uPvXQ`Qf)ajV2__<>cB>!z640oEBh1ndfzX^1D45
zLpf;uIS_vi=qQ2r%ba5~-t1mXkdSA^qt%ko#m^;z(I)kiVx;g;&`B2fqHBpqzROCa
zxeX+@fecD>kZxb?;J)rGB=`haf5Mm?>hk3LO7rX|mIUN;`PS4X1Qc)u7zw#Tj6_@!
zM!B}R0%o0&Ys6$mnUTJC1T=;j=@=u*G%jG)wE)`|AfW7xcz9)^si9^ax>72>ijElx
zz4%@jc?-PhsKjV)4<}!La2z?m2gLUP1Eq_Ey&%4q;kd6iPUMX*t3?wFVIhw)frLe{
zh^Ba($+&CJ(v~5IrP!s=4o*IIDfNo?;~s*5GOmo?n}Bkz93ur+fn}DGtE6pCz3!dR
zT|a7yHtz=U-AwFB*a!4|Ac?Yi?%Gy)RjPW|AE9(;V+84IC&mV!Pkn?mn95B98S_5}
z#OHudt<NZRcfOhY<6CI`T!*>t)HK88;J2mTH$=#>L@*IW4U_0%3LAk7wfgTDHD{tZ
z+W+PKNwjIAFcB-WJArm5NTJlMC@5Sv`<MDil)y4V#-WBuC@16?DF_8dN<vA~Ti>;J
zU2zgMBCF>)%%k^NQ>+~jpTGR-{|ZR)H@EH@iNT9DBveB+m)Xz{`eEb`{V@uF0T>0s
zK#YQ55Jtf;7^4svf>9_8#V8DhVWfc?jKX0!Mj4z8TgJ{TP8Lq~S#;iz(7-WZGK0N=
zmX&0Gzx0S%O9Bc<6W}!gr#}nibct*QyjF%cz?LHmFX)~je+~nOVWvcu9Zp)6UOm{2
zB9&&BCSwdrw@b$;!!IM6k<Ay)r$h4o+}+w|qZWn81ietNVkT7pZ3Xit{8MjQ{&2^C
zgSjX`e6=Gp&5!Sgkw4!bqX2#YMuGf5jDq+<7zOi#F$&>_U=+#^#VCv)hLMJ^!6=*`
zj!~9dmXfj4;AWty%2-nQ?1r5u^5z=nHT3&X^<t5sj3>S_J*-+2M6In}94c@VY@R^F
z+kpEvK)<5ZrzRsfxg+b5AcvDftJ&Z1bj!Ky@?&WD&sUNP>i%)hIp5g5=tKe^TOS8X
z|EGD0gg$lCZ3|SoWqhp63S<I|gbJZ2wS<5orHGCd0rPD0u+m+^Ex}563(&MMFZF?K
zue1xb&PLQaaI833>mXK(>6%VLSG6lfGItq$XjEQxMRcn`!i_+;kx7#SkA%v*KS_T@
z>8O&Z=p;k+UgCw3x0^TC)KvpjHSnR%cJTEe(fMB2NaQ7ZTYLJd9Ndz*cXaJ(F)Dbj
z5?8vAlh94#hLO}&N;hY9Z6Dg4z8jE)mRQOy{iqE{xEbg+gLH<E0m(7oMb*rIM=iGx
zs`+USGN26jmw^Oo;t}Oo`obrvUZw;zIvTNO!su(HwIi9{m_#jUCMH*12<Ul)Zr3KD
z+r&7Egq1*D31}xb|DavJ*M4@+bj1yq&7sz#iW@G+ND--+$H<gXN;(opv#zNRZ0FgU
zp>`Vb=|W6G0Tf^)ghGr&P=t{Y8fiM7(PDO4Z*PvA&<T<{K|1A~lM5>0Zb|x9BzTwm
zE*+k)KH$}qT34J#_D6}Mu*wtT6oUmm);Sgne7sXUEo)zIYI=6->JFs0NKr(m+gaO7
zb%$?${T*^(k}wI=y9;P{f!UNVdPMR=M?KF@cU;lpXl&Xnr&;tCB#d*8!?sUwN}y%E
z<%t$4!?wj}`$9z_os8$5if_YWt{LhRb^!GbCSC-r0O|^ycLMcJoOc2A=+p7-mzHcP
z)6U~rpzzG#(<zF=Gedxp&`0P`or#1ZUr_)<d5(E>V@tvkQ3+P-TEMIpkU+)gl1HC!
z$7xGFI!qUc?_!#h!N^8)*@(OWxyjnfTFB_Ov9iHPDOGwivLc_NP>Ke}I`dCO4Bs;$
zP;+MJ0dG)S5~w*d^fEkj>dwz&@0qhu+;l=6U3p35M(qwTm(q9D<twx6D#-<8X)SQ9
z1yV}ztzUZ3_lt`%(fSm1$}DP_gsFm5I=nYTXMQ(ENQ~$>Dw&F|j;PwmyfE@sczZDl
zs)27cpxrw1F}wK1YU@5^z?y_L^xe9ZI>0*k^7WYMenhOsa>h^WhmpV7AEN+q07il0
zK#YRKK^O&#gE0ybhhP*c4#g-;9EOobtidQ;9F9?zDhq2O4JrdhWk6TP)SkaLh135B
z%TE9`70Wwc$~TT%f89Y6Fp3=oDc2D&njMW%3_Av+S?pOD#j;~DietxN6wi*wD1n`T
zk&dmyNYB<|l*mrRD2bg!(}Ucm^bme@*B|-r(FH=NlRT(w`ng&)P&@tF5gbeEB<SB}
zf{PK4;9+D<SYu>E*kDu&*rkAO0w<b6rH^jZWuONJ$|3y(kZ=poZNVRy<3Mtpd0;5(
zfF<?O$X^}8vtI-AmyYO{qjL;{q+yUo&3u&~V#=+bxef`o1CQ;1e(+wk_sc%f6z@Y-
z-%M^6#vIUGZjO<K+yW!EoQ+Wzz&-Z+X!1Hum`4Y?cUE#MtTG)1lA}OH8H7H?$;+O$
Zy0GBQLi9-x%LFDn4ublBWbs(6{{bOVU7i2{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_bwd_data_example.inc.11E97FA73102DBD5.idx b/.cache/clangd/index/run_grouped_conv_bwd_data_example.inc.11E97FA73102DBD5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1c6903acd60c713d493c4372535238e1be39bb7b
GIT binary patch
literal 4392
zcmZXXdstIP7Qi#+0vWhTOfIj4BnBcu2(JVX<Pm~cB#1S!hN`HaS}C|-utpHkwg_|=
z5p6%MvQ!02u~KS9m-SU!AFK6kEh<&9txt4Y5vx?uTH77k&fNLB`~m#tJm$=qGiT;b
zv}R;zWdcF^9BbLUyoIwEfk1#z;99s~ZU+zuLeN0jhP9m&CR~a3cc0(%)Z45bW6w=L
zmC+N>oVg;FIrq+`GYjtR{>X7`RZ&Rj=x2Sd>}I#2mrH9N-XFO1b$|2uoAq8{TOWX{
z8q1`#DXv$}<n%ll)AFo6)cnSjg&Dn}D>IgKztI}Ha@QNzR{1TzxV5P1^^@WIPRHLp
ztn)AaR{7cdHumf5T&~-?n8S~^?r1#mB;>8gvFYMg|L&MWlB(17opq+U!wxCqjxW7u
z8gT6CvQ@x=p1bc~d$jhl>Cnd#^D^bZ$J+bX+Nv59d-go77;kIKZK&9L>V>PltCE(M
z`X0FOiNE@&_iq<f<-fM;{0D}054$t0_Wsk!cRp8qyf3~fUR~HbZq<_s2e_YZ&U)^8
zJbiiD)PLrg7VLCf8?pbC=51%u)80w@y`D_U{_(~yv*U6%M6G*iX+)uu$y~anx}!{7
zGA(e|dWG@W#i=*neRye^$IDe4K3KU^3E%souxeM9Fu1z62+X;&_tuQAmVb*Aj*2gy
zov}Em$UgOKyz-J}0~mkp;EIXDDcdIZ6*(&B%-yv<!*oKJ|H9^~+Dhw<uzvVMPJ=;b
z3wPW3aHZwVGtF0T@2ET?a4nuUdo<d^_|7%cqx(x3R^TNNK%tw^_xE4uU*P605}Zly
z98<rd^&~4`k$5#=R)gR_{|F#6MBwQmSdjOst4AUb2#sDJHJzSa=gvZzRu)VgL%CK?
z5vSuu@Up=8z+|3+;)5vCCFv|!#x(So|9D+L6R@xuaLqu619SiKz}?=k?VdwesL(2m
zSf7PTt&)(a_mJ)I1<&)yx&eq9fG@Tlx>qwVW^LwK23fO;VU#tiq)5%GDblbSLf+HA
zcjRPluSAO$dx?E12PS0-Pa2bkqM+eHQM{}mxnL|$CDxKjnEcNe%TKhXT|i;i0^eGo
z#R84b<!x(%-}xLcupUIz12bNaX3-|gF);i#vKSdNG9K&ew}izuB}{#ehTDOn9gM^J
zBb1FpKj<<RG7Mz>*#NB1K(W7=B8i_w!^@<8QW8n%%jI6iQz!}+)&TDs;D^=1Yd$=1
z^_QD%XkTWCXUOm{3tND=1;}s!BL}{i6I`q*MD}f9d>hEY0&FnE^N)H&r8@Y-!kvKK
z3Dj6EQS;zrq5G@{kbyFZY$&#3pj;xSh?8&>&5+OV<~5xXCna<0ouoV&=uC2weM2<a
zIhD6-0rD2$jRQ%&*VVVJwXqbfH<$}1xnrT8(-X@0?5)0{-dCHDQa;R=VqYvQfCYrc
z${uB0>Ywx~N=dq3I;r9DN58UnMf`YN$iP5JpbFbEFh~+akxrr`(@iYd6SCHBTK#)+
z6>T6J2kR*5yWRQI(|1uELbWg)`(oibz^nrtR?DdvuKsrXf9E6nv5K+7vD)-=35#_R
z-M&H=GLZ}^G8R>VPd(6J>%Z1!LErY#vCqDLgeJB4Sg4Lr3Cxw?MeKd<0$1sTCtEDY
zdms}i#=|TOVuC2rF*-u?zd86tSAcIDiZfG@Nj$$+W<IMrdG7iDoszP{*Kd_GooA#h
zG(dxdpCTNFQxpLsD2jxU6h*-(ilSjOMMh|($OKIk#lRSfVqq*rS%$1=J{Ft7M$<@J
z5-+nG?3Bzt(oSScW__2}?zTG<C3!uNt_J~s9Gk&|vl&P?^Sli1ZqGe!?MFH50E!O2
zB~*TrJ<qYR^BBs1W?beNeo||km7;O6<I=ILYGS>)zsJLcCQbE8HQ-?uE(7LefV8s}
zS;FdjZ~vCSqO@}cC9f9Fg;Ny4MNkyUMN$;SMNt&ZMN?$tj1-wT6Gbsx3`MbAEJZep
zjq1O4i=EKwF@1#xKKyj2Cj%2i2@nU&z)_-66eWrhDM}J0QIsr7rYJ>}LXlZyrYKdE
zN|8lmp(st1MrdvMk~RI^eqm_i^3-{x5w9JXIlWS<>KYt2<dH+Uadx72koQ0O(D0Js
zOUSLW@zY1C;=f&4godYyrx7QcS|8RNomKrWH2lZCf~#j!Z<y_{-@t#6AXFZzz?pcz
zXdIJe|N9Mh)P40bs&C5ma*DW6juZt0XUb<%-M3g$Om*KDkle!GX3e$F&ADcEIl5d~
zFP6Go6($8~%Pdrz)D#5{4<yS(jn!H}>W_t`)>1;v_kBWd{h<5V4K-|SFlid7VQcji
z)dE8;pblFdh^zxKH~}oI11WW2H0G8m@xCWd1X@s@eB?fAGF&scp{HcO3LOx&R!v$u
z3pH8|q0OJKlI_10myHe{RKrkg%fd2inS-Y$plt%-xXFH;mhT_m_=62i=!$h^Vzu_r
z+m74^_HcArIF(N7dT<(@BzPpta%N+b?aZTjr{j*_M=>t4E+T1SVKp#S1CrqFSIvtW
z{3jF)w&#>U8@5L6IVFf9U7C)#*fH&rsku%nM~^HSl##B@LOGOE#6gZCCv+0($l2|7
zF3aB?EwUYCw*v=W{obi-V7=C{90^{NyhhT!Z>6-cV(2^jQH%+y1U(;PVn8CbwaEd=
zgbpqe3Xd#z9zvc9Mih_&IwJpS^C#ba_6<r(nok<#sS22@zy$31xNpeT9brdiqX|c;
zMp0@>0ZC*E7N!KG5IRvk&f~=k@-j4G{)l{1U?&60Zoo2$8+wpc08<5@cLp_6(=wVX
zfvJ+_DqyPO8K($%;^+FGXZcIrlA22MOVYt-dZf&Rmw<L7Qce+P<l=bQOyx}S0AyjY
zZ!y(BTfmqWkcsE(w6wYFuY|OsV<qvD`0y^IUQ&v5I$Z=WD~K*g<Y}RG;Y*m#e|q!r
z?Kj;8XameXX6jH>0CNSHh}FK<>~p(p(4I!wtOLqApu>WTF^>4}W-pwK?9)x@)cuv=
zok8}g+xNh)<*C|ow0*();4o~9T7h0qQ7tgm0+N&~>qKP(&Y`E!lzEwXB;f1zvx1gq
zQ_=_D7Aivs?<-stPEmv^f}%)OBt=oGD2k$0(G(d~Mv6=-6Gbtq7>Z(5u@u>?HgX<W
zXt&xaa#>x3{`zPe{M-s@1}40{ch_utywPyB37ut@VLkcfi5W3P5=KIil#vqZ+m@)i
z-@Q2p<yxc`Q5(A0y4b<H+XS?m_%7G~^T^`7D_Qv{*Bv0c15Co<JTD$4lx&GAMS|CZ
zUni~f`N1#s?<JRQK$Ezs-N-Nt-PP_Cd8j=I{n|6WG&w2w5GoW;wI{i}Sl9@(jlh5>
gL|?l_vj$I+Yr!in^u;U~0^Q%gljZ@>-x3J^2L(I|_5c6?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_bwd_weight_example.inc.EE2E4EE0E60E4D81.idx b/.cache/clangd/index/run_grouped_conv_bwd_weight_example.inc.EE2E4EE0E60E4D81.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f62070f5bc3e738a40f04d3dae4e138b97ffdc96
GIT binary patch
literal 3260
zcmai$dr(tX8o=*?%YjR-0SN{O3B*lGkR%eqJ1-vXQl3h1=vWne0fQ8nDdA07ED(wW
z)@c#c?jjYJEoH66$Bf!m@m04XBCeHHsbyWk+KRYTD|LNz54&@2&g?&A=4R&iJ??kT
zcfapDCnG5#;Rs7nujV9G7Uh<1q$!G8j2=sA$rd@Fs0Zk&{A0)V(*+J8?oF1NRl{*D
zX9`MB?2y*yyTz;PjgN(gF7!3r$VnROYkrtrG21_Sl?|PU@jn~TbwBZ5gZm`ACwI{V
zFDSC3Kk3H!k@4eq<7>R$`o*x%G&Xtb+b`4){*iU`=gzqAb-&YG>e6rCmk*Pg^?w)(
z>O1sR`~DAKj%{vNoZm9nt7`bQvbyka_VXu$iTiBV*NaNSi<pM3*M9oGXyz{|AKg0H
zbYLdVB(8n@XrpOlL*54^qDA`?o!xk;bG(p--m8wg-sBYp#h+?ky4PpmOwGk3f!8}#
zt|L#s+xhjJ*=NbJ*`aPt;?#~4ueGducIYSb65$kRc=h7!{-L4YTz;>_>-oWW*?442
z`}>N*LWB6$;jVYiIZuYwoIV(1iys)UR26OVM>#u~-;~;EiRoe~XNrQf6D|GkAH_HT
zN7{*^N^<ioN*+ZyhVVY=9^TXnS?KHME92atSS1dzlSC?c#ZJmFWwf1Czf~pMNgb&+
zV1grq_ossfTpd^_^OtF`n1ynwoO`(|(Cm6=`=;H1g~j6HaBSp$+^;CKBk?kge0efI
zEJnVp+}}<Lo`TQ?vVok=GwX-UHxdWxoml9_^Ag%!ym{UPspP6)yUwgN8|`ETdMil4
z>~&^q=i%>my>Osm8}MrbQmmNgh&B*~+51vR^@YB;Z_tivy=s?YvFPeoZ&N=Sau@Uu
z_&6x6J8I1f<nQL=mV*6N&3_)8cz&lJWiw|{&SETPp$#xLun2QR%Lui6tm<v#9jlLx
z!D5}??9f|q<19l%U%sz97SmA77ZW7mO9;yG$nmsW+5q1M2(1kS*gznu*V_P>J{qlS
zwGFJX>)Qao&F&xb;M&ZAzCGn=JHZ~oB5cM&jX>jzDd9`&OyP81H!^AjOe1i^3Km)c
zW3_YI`BC#V^}oJ$q@j|p<os!<;;Y2iiH2&vnxl-eV~P&5_5+kCfsY^+8)cTw+`aPn
z=?)~O6?6m^XZ<azN~XRkL}I5Xr#LLmpBab`d-lW6Xo1P1<RC0&VVSWk4bzqqOZm!Q
zK8|%{p^{c|Uy6k)T1AkWR&!L)raJ1<7g~tIt?*pIRcPhe+mqGwl}oNH3}?dmH~<SH
zm<WO*nMi`7m?(mxnP`G^jE*2ZqbJC~7zi>lMuJR?iKBO!qo1uEPI`pa%JboI^;-XF
zYEfp_mA|50r>CW_v#-^>(wuFljFlN2HDv5GJ(_a1pll|Ik~DUsQe){VOxwnLiu+^s
zrlD|Tf!mTD8y{@{HHybV7Y`R7Y>g5n2q&s5LJ&bvq#%-@C_xlK(Sm4#bOIegdV!uG
zgTO$LQD7v<BrtLGzOAh%y=I5w|E;`ZZbMeRoA?^qb%CfLz`pW&psWXxnD@j_82rpB
zix!+8A590PpY@tMgTjOgLb>RB3sdN1bJIVa&;*eCa$(Q}(2B!w5E@DXBt%c=dE^lj
zpcU|2?H%0H^1`&%B&kMchY!z(n5lBDoErly4Ace^q?W6>YGz@EHiOVvL829K-?``E
zl4X-W%KnW8c(7kEaWrY98sadj4XY&%qq>MXJ>Clo>(c5nF#YMxNa4_hrLib0{vQ4k
zPTahD_f*+QF-pF|UqO^ypfr#>qk6YA`A!GlD?+}Wtfx2j?LG1P#V04jQ;~0WK(-c(
zSy-v9G-3K|+1<>wvt|{VOg1300dD@xbBGOuV#9e(wgGMs4NRCSc6+bPpPw2!Wg0d^
zQzJ#ibqovDMs+&Y9b9)^+tcdifn0nUUv8MNP|S!4k}wjEuq$x=-vt5P9p-<>UQe(M
zHOL0hj#-xV!dA!Ev@<)UncI-KkYC6p@pHb*==s%?X$zL-sR~EObK8nfBY!JUSiy4q
z;#2Q!=*_(+zrEla&UuXXetPT6Z&A)uL@7$_&%$zJ`3g)I8dIECeB)7x!ZZR!Bj76T
zV#z1wXLpJ$3x*kXf9>zj+@^I48yX7zkAcwv6v~}-=SpKlR?}Q>TIDu>uTuFEWhK^$
zxzb>vWVwVWn><mTAKn5BTY<0@aNWzIONOZp#AE%H*nSu=uINJnggoIAEJnAb5E;<{
ziVy`sN|{n^w@lZj8|-8S!B)U^($&xJJh@v3DU{|!Q6l%5SZFa?%-Ctn?+d3JO6f4#
zd!49`>(HCR!I#xB(rQ$N!SY~kn4ptAKtr66wYpj^6XTsOm2<`J7tx#QL1Mjq2;6T!
z@T5z5xf@*rg|yHWM@Dx#+KV7>+M6IB+J~c=13KBmo4s$KjjRt?Pb9laTg8QcIh9s!
w(7$pT8D#}zac$4S3T*{9pU|DTaN~S$SV}fq(AAl8fKT65#Aya@Z>6aJ0O;VDhX4Qo

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_conv_fwd_example.inc.88B5CE7F22532219.idx b/.cache/clangd/index/run_grouped_conv_conv_fwd_example.inc.88B5CE7F22532219.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4a425ad628181a269ada19cebaa5ad1b678fcaef
GIT binary patch
literal 6876
zcmd6rd011&7Qi!O;1aw9Fa{CXB3o_{*%uK}(a0ts$R?Yhq9Pyy3b+=<1*!-_u`1xs
zQ>wNWm&)2|c;bQxiip+LB4Awuf{m@#`iA!1%=GhL{*dp3e7~HTGiS~@a}F~>-m_-4
z>v1?U7I|kS$FGdzayT4q^qamiEx`?PIMHexPF7h-uwP=gjn;1gZHK-%x@XlF1^;pM
zwf2i!dv8KgT!F_fi~L-xvEF}(%|5K}?KCfX?($T6aCKk8W&1(>*UQ~go3~64u0Jy)
zce(pE+sMEza(U9CwNI#eJ&j!>UgsRxQgmwJ3IEst`t6kNdk@R>{BJ)K4erW0IQC$e
z?Vi9H4_~}j%lB`c`ir}hd)#3Dt8ep;$qY04o~}!JRx(iJp?ax#&_wae-A6eE8*JTk
zsST4fM!!>WoO5ileP^2L;GZE&XQ<@J=GvA7))m@G^GqY>th>7R<;l?%S?l!2j(Jc>
zH~W~!X9fAtH?vPp=ZUF}eGAQB-ao#l<8ps+lQec*(@n*nGlVx<leKjV>n?n8_-V=b
z7oQtS+N>*Yj8n6^Lp#iMa}GKbIZ?HzT<y}8ory)^M}wzie3g+Q2|a(ls{V)`-!<gc
z&gXlQ{4#EpbopLg5n0t$vFL*1z2imIkzopO&2m?-JKZF_egERK?$QTW^R};__}8UV
zOZrZw+}UrQx#rP=1JCOd>p#2vMzriqonX~Z^3Bml-X%R~Sw6CM%JQR0+t+F&m`;(U
z<U9ZLu03%uIahC|M&RueiVtrMzKrp0a(^GOBxCUVHC3e#A6?$g>vtA77=(|ksS8W5
ziT?Gp>z|=#YRVs<c6IB}{QRX_c%9mAw=FK8A1-M&j!4)y=f-z|6EhF02_4?IMtRO&
zT;0BMSI70y7xyn-t=`soUV7D!&tuLXa@$>N8Z@qLXQ9vAoRK<Z@#UV&*7qA3WvF_k
zuTEYt3*}Nx@?`DKaV=t+gOZ8TR?#;6`wxf59l=pi1r&H%qwcNK9rh*7;h_O}P$dtV
z|Nn!-gW#V*Y4MBGHFS}>`QJ}(v1c>^O>KgDo1h7+9`w!jpLTKPQxz`dW$R_fe$A!4
z?Y$i^`Plmq<ZJIs&>G`4B6j2m+mV$AniuR)qbUnx3p4g>ni7u}TVP_wGXzO2BonaA
z%h}5fleeq)R7^guJ_Pwr_npPcxO)fJ^mm!Zj-aU*Q1AkpDBVN#L*0HT#{3%UzQ(M2
z`c3#}nu#Ze$pNY@-#>pLIjxTDurao=WzC~0Tam5OyEJ7dvQv_VP42#BCARm`6geCt
zhX&Zx0XSxWWCa|fz^r+&a>1@b+w~}@_aJi*&SEEORU0PTmw4S&MMlszMuZU>w#Jz3
zH0%g+q#cQoP8v>x%vr-(N!lT|maaHTpGK=HV5kB{v8#@`l)d}R&wJjhb193l7US4;
z5Sd$uF^ToW`UHu^5=@L9L&z90L&zk05<(`il33$;jey?>v`DsW1Xe8T-TrSX=b2ml
zCui6=r@qWmvx2=Jok~ODy|0VS3hN=4Du6}>P{R6{OBF$*BFu)F2d6aw)5l1|!buCK
zusb&5Z=95;*4%}@J_`hAfiifanzW@avfZ`<E)ww10&RANX-{>Nx<gFxN2Zt`lQ<l3
ziFglbssS1_z<@gG`GGZWx-)622^utEwtOC=734W>1JW=6=MKPd*5--bM~<YqFV{nY
zHpsL=M^<1RaO9azs^!U#*0jsfYF$-&DiW=E9Ey&^iR=uTx&xUz(1~S-Rbee3l9wDn
z1Le?34qaKn5YLc9PnKP^E|s6#oB0<yV~3#HA;@P1G<6uN9VYn*R6By%)$rA|MAf)g
zYAB8Q7DS4P`C@_?K0}a%FCj<<_%bp%Wx!kphyx)5lVv~|IxZyx;W7}x24?Epkm@a|
z{35i)X5iQiq^y9Z+JR#`=ILiosb`Jco`VK{1w~(xPV0b72X>lA%5f{hsnx?(@n#mk
zeiE1__!&h}0T~6fW?d^C{ZjjoV8oig6M`AQ2F<G5B2Rj;cqnSMfL9AN@s{cUuMV?k
z7H4~EN{v6#a0@cGaOU_q7p3S}?I=WoD`46c;LB<VT6MvqB+zciTQX=MgL81UUWW$P
zF$X`0@#)w%d+IQi{qU-DtyA{+^dW=fkdZ?x)(O$KTvnS{{iwsGsnJfOiRLAA5)xDZ
zg$2-%okXKkE-b<v-Mu;d+~Sm>U9F|p>av@qscc}G4V+nyS=nj#WOiH~+S58_9Z@<r
z0@IB^*^IeV9q_9I{_Hd^RSyE{!5qvDAfN&B{L0GnCxRv|L1(8F8kLf*wG0}Skz5Xq
z$}umMpZ0Kg_w+Zk_Y`4@a($Xg7p4=`1Ef75gk7Dc`hc_#b3$5rM!@gKrXeL>TrUkg
z%$w^?kPp{~AYZPpl9GZdel&i?y*NyfPU$s!-;lbg{O?H5k;7m%XsKE5H}$XGdQWqy
zLTFHkqc_B6h0vTGIK<Y4(1v8^Lg+&B)IvCoWY0qAg}E3S6qEXkp;<AhzZhB<llqIH
zb1|vE7)~uF^%q0WVyr*)+&w>qVNUrlIV|Jd^!Y^teK%0@o`9kgxD3!#CuBNtUCNSm
z^{nBx-O!?`AYKq3M|v)AE<wS(V1hz;Aq0i;LJ11vg%K3a3nyqEZyrGrya<9Kd65J~
z@uCQd=0y_}!;2wkK5srjvAkG<;&^dN+OX*E<n~`Try-B5o3YLhZ*(K@*a&8`ypi|a
zH<8WWgDBH#pg|4pw3|-NP7Wz=YeGk2AF$mArm-{70R{G@zyos`urDLI9N3p*&I$U`
zeXy5GM~apUms{d#X~Hyux`DJCL}IBNNadKfo_)2bN9N-;%vC2(T5&2_bwxee{cXtH
z#<hRD&xp-03qH6Qqcjl&>0z_x3g!|NEC?njL=Zwys34S}FhLkW;ev32<_YEz6d{Np
zC{hqfP?R8wplCrfK{0|Dg60e66BH|mB`8i1M^Gj|Q;7Gsn!lQ$wFYZVv1|i>10l->
z{A}<qcFP7c@aP=C&mrYo0e>sW+W>zX$+>`^OL89I=aIYv@OO}$5BT{c?*#mvBo_dF
z0m)TRrwSUdt-9n*Z1g5g{W~b~)xfeE_qW~jt%9Jc`bH$U37MPFfvt$8Q-oty1Wh<T
z%qh6v<Dr{jps&U!cW*e{CuDk_nY=~k*3a0FXr=z+{fT>Z!1w@yJ^<?vKzVIHnj5~i
zcU6NvYU+;0j>J9HN#vxwnbMTA$XQ7h_xl#u>2{f;vNZsm2cR44{vnnQKp&RBVIGG>
zz6zW$V(5QaJsBXbZ8Bgh1NLkt4#|CFxZ<5&xFkPE)chQ+PzgnqWUIIfnY*}Ed>hqi
zcX)^9C>8X~z_U=cdi2b|6BESn7=k1`2|+TzlK~}Pm{pSX^|W6*k;$ddv=l!NaH%qA
zT1Ij?G%d$m&3}4x>xyL@l#~B{SgGNwwO$VI@%!YGhjl(w|K~voJq`{1&{PZ5Y{46=
z&DY$YtyS27ax7jy-khC9Q%m%h5R{;wKv1GVqB7&y<BVJ?hoj_Bz@`TmrOT)Rk`-{2
z0`m{sVk_dGn09?~p1aslVb5u4NEoJnj4m}+3QCy7W|Zy6py(LBGoYItWZH4ZYaG>D
zaq6eK2PhOPgewSluN1B%s250kK`0JXKalogzM^As@>@yc0;FWI{$gU<QvIa_Ez@5{
zP?ABCGM-mA`%?bxqq2~>wLn}8Z1768V0ta^X8C$rW#GH+<n&K2e$84l@4nM~;rhuL
z)?%_Yca!^E&7pEO(Lz~pQ8}{_D=RilF*XcA5=Y4-*5x$S2<#hir)={oxW2vW`nX}r
z)Z4ziv^{PMdQj*U&_w~=*`qwfo(ky4^6i>##pBOZIVj1Jg~<#yFIAXI&?6vy1m<Ds
zOCWuTxx=tl<({Q+^G7B1?(ZGt4<fsV`hP2ATJZt7zoqwXL$^v55*&qUM<LDH(-pJw
z7Z;CRBZtWktXiPXKD7T+J2}sbzgCo6-aLRFTt!@wI^NlM?s$SsxF!Uda!m;`<C+n4
z4X9lM$_vB$A6m@!xn-51{LnGeA)-FUYz#rVX1WB8H5*Hio|&GKIBD@oaYL4;r!7n$
Si9R6}c70}4^&h%I4(ERngj0F|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_fwd_bias_relu_add_example.inc.7B0A8630D499DF28.idx b/.cache/clangd/index/run_grouped_conv_fwd_bias_relu_add_example.inc.7B0A8630D499DF28.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15f5e058e020723a7feba71cd29a4dc948bdfec1
GIT binary patch
literal 6284
zcmai&d0bW1_Q3Bt9N`}B@qipWUYTVcWCjyOz!?-0cw%Isppqc-1r<@$E?g?p^aWTr
zfQYDBqGlTULL?2*dWt5R81xvFL<TL>!e{>WCHLLC-}^(9&F8!4wf5R;t$m`yX3Z*8
z%4E}*h2^Ft<|ME(nQRpL&CXFJ*MLlxuP2k`?kbIJNWbmHx2##*Fx`8iMs&EGXuOGu
zJDvcol}W39h={ar9OXK<$bK@n>Az2x-Lz~?Pr0c#?w^C`pZ(-h72auf^2L{y3QkW(
z6u;5Y@yIWxwd>twTa$!&8%G7)pLEA^Zkyqgz^|WZY8IS$cj3off0$?9xIfX{cHOP2
z)V-@7U%R!%qob<tr}l5x{U;~e<N|wR^uk7yFY87vnm?uP&dMjj!v4<6S#GX{R~8n!
zWhrkA#hv+PPmEo-fkV}C)0vyrY}%k)KjPw|Hruv7ONIS&&$^xmuiWlgwlPk;$h_fx
z%>4J22eyYr#J<<PQ}HmdzcNANz2c3aE6TOyN1FF+zFh2dX!j4jZ#~RjSNbUR?8E5F
zEqPT}eQOu}`FrNQ-h!KVCQNgj9T~p7qVbzAKJLk^Y`wfuo#I|sU~pmFJJqFPt(X2#
zjm%1KNNOqkY`5=>&;4&D6>N?^y>j%mG~+pJ+qKsI=IjJb0Bd<LC#GK=Yu>&3XLqwr
zYol89@2$HucW{T<=)u9SZn-S-ckE?VGxL`|KY#Py$sendYa=%q6+5=h`DMbhVCSvk
zMb86OL3!D0(w2OLP7PePL?_MY@6~5U$qXdXK(rVxWO8(}=x2niBmZ4>bFH>nE-4L^
z7Q+Q@<~GiA3^*{pLEEe^2?m12aDkisj6Y~K{-Ew>1v646lkD{E%!dmZ$01vh$tKmV
z?(gVo@KP{5YPQt3G#)N+b7;IY_QuAq3war`I%x|tVhiK>;*X4DKxR5xmez7KsVBLn
zRl%4T%OnFs1Lbgm*I$)uozSn0U5ge#&H9G=ro#npZW>UYiRm>ur_UIn1y}>tbhwal
ziV-5(Ufaq;i&|%$n1Cin%{KZrro#o^-jnz_CkIRN1g-3U{YVM}g#~VYen@F>QBc3a
z%f_J5aU<fKh6@>I$ZB`~yy4dL`s-EE3PwQ9^Y!PC9WL-_nJV$E>_$^G@8bHYmwjD~
zfR{9knTD~$^WuYJ4F5W?7kl9UMta_gPb$7!ve}TA(yh{UV|gjVCc^>G#7mjBnNEbV
zT(dk0sT@?Ugw!xe4Z(y9U`hc@BU}i<g)oD#24-s@oNy7$D1tD|ti!pj`6h+_NVsjx
z_-#z!3n64|*{^vIIvnFs!$%DI?}%*y+ar=%YmB%*|2LBNHAJ4fZumVK{UIa&kdg3c
zY_4du)>g@Anv73brTLbIUeV{JTt%)G-YYMy;n!Fb${U$yLnvR7PxV~Sucvx$7`Z{$
zbIjKpw;yl+_!+}WT_ALUF`h!pj$Po4S^018yo(2?UPm%KgE7xw9BwfC{+ldWZ(!m}
zXRsQ5tESjtTOF`3l#8`swR%s;1+c%1ralM(2Vn}H+FE!!;63?w9jM_3h&MpD>v5-5
zNqrjwi_lazK)3;;@Ce)R#Y2(5)y*EkOAEpl#A2i7r6fg?8SaEM=d)%iq4hrNDQbt{
zc6b%H+L`<$Yx;BouZNUxo$H0G(G)goyb0xv%k#nHY*ubu)#9CEz)H)_ms{d$R!Xu+
z(wzb;C0iux<YA>17Ax#Y&lHOkic+mob;rO;Jut2ZbhpDzZ`but+_oelfj#A(W3hwd
zrBp?#1)h+X3c;ZeCJ@$uiv}hUE&}@^@W<@&&aChcTt2ZudRNX=MqnHFl`kmB$$PB;
zRr|^PjPY9dFEV1SM<qsl8pPAYfF@Zzyp*D!@Q`DrbRnIxvkW1FqD&!^qAVecB9)+`
zs1t-v+Fm=sv6C#vN}Vvdlh~4$V>@9XmJ$1Uv}>)SC%bkwakjuB1^2fM)Eq2d#qv@y
zm=ptLLM0%UfH|IxmrB936ex{v2jzBRu)#wyr!-l@he&nvC(aMWb~)=S%|L43TNMm1
zwE@=#Mr6Kr;Mxh#Hhg>Aov8C$kr}$n-IaKMy!2YwYp>(kcw`$!J)pBOe80dDN2wQ5
z&MB+kdbyx)BwE&)cOFG%^WZ&n6h7FLdTv|xAIR+p6NCWlst|<;(+P<a!~pF2k*^OC
zrxQ|6PzGS*L8cU<oQ^4?u{o;6{Xx?RR#JnYCW>UG0uTy_B3Wd*LLpGMNCSd~cpp|O
z0-=cXj~>%^u-k025faT!?nd2*0<bS279KMxDyuGIpd0B^m@EWQ;!GBUC;=ubgYcGP
zRu0^~`sj}mG;^FsoX$t_(pHAu%7|Fman~lUv2$;g(9Da17lo1o-NuaG##rHQ|KA73
zyTm%5sy$hb48+6HBM=`+!q~9*b}zLVs$Sx-M0a76-duaLYRr~yR6SlkUc|kV298ei
zdwTCTw8fKPdJ=R#gqI$`lm`%jcayBjUN!Ic10l$3sDV|(NZi09^|J*u7Xn*Ivj$iV
z&BzBA5ytrsGC6JXgA<x8hk2W7t{zx`_Pf+!>14cGR&LN0i%Zv@{ZB;5X+HepXur4b
zqyC8wiMswdLm^-6vpu4Zs_kv<U&UgtQig^aM{jxyRSPqPx%e<~vW51d&9v}da?f>i
z=doA+E%H(saAhFi-SE;5;C9fw6S$o;?*eWY&AWlyO>;SL<uq3SS3&b0;P%j530x)3
zdx6_aa}{t^H17j$AI;UkRnxp5xcxNOgLOS^1dU+b_;+?~1S+*^0_!Hay$M{KNW1!Q
zcv{Tvs~3>b)PZvyyh0Rr23*d-M8fqjvmWLXJ_})IA(n6hxHORd1<@Y`{d&hB8=371
zFg*cOdN&AD2H_ug;QMp*3j4P{pJ=Q-I<bs)$2&mWLGw-!chbBI#9cJ+25~pd<sg>R
zTmfPQ&3i!HLvtmFl{D`KaWBnPAXd@555#>mSA$qh^L`Nb)7%Jtji57pUTOlrCc+=C
zEL;4et?R$gQN)DByhZ}Ra>eqoSX(G#@y-n<l<%ESQ9F3HljvCbBtA}UWYL1I&~0$L
zP0qG_+E0O2QFBMWG`p)zoy*RND(XZYJ>5Rt5iiS18IBoL_LAwGsXJa)%JRz6rA@4)
za#OkEaeD_xyL9*34Iv4O!KxUjr$PzXlmPWuC<WV6k`VFIc5vNJtX7+@x<3QmGW%YC
zMY?oxTnLgh$0SD=SE`eX%=<ilPD9l$au-wVTB<kqez`<Q5Ky%^Qw+zEp}OvaQD?gS
z!%<$~E%(;-=B3xeUXLNCaTUz2f=(5@^eg!M3PD)J16Qp3BhIQyP=HNVBvYxy3dIVF
zQWPl^<@x4OubU0N8z^dr$aYwar$6FyH|C9pQRDTrnU7f~36Wa%?gU@lQOk2W$)o5f
z`&-&~{!N4EIxY-bIG>2~2z(w9lR5fR(CRYfyrkh65~3!zR2xI47r+8Mzc%}vSx8@h
z4Mb=lp0pRioFa(BeC*91eD-cOGC_Jt=hCUrl)+_Cl*wgMl*MIHq~cT*bpqE(LR9Un
z(%hPmo*z)qDPnAk7+oZ<ojvRQO9giI=&T*(j=G?FYGE<=kxk5dsQL~FcStI7T97w4
z#&$=%^lJQTi&6PT<85b<rVfGmA=-971M|;lt^vy$+Ln)k<x!eH2aC^%MV(!(r+5Cn
z^aixGc5rG353Ito@5EOp{%qgz(yg60X|6up_SKK*tiJ@)FX^p3geecnJLtly{vAsK
z9<M_kUFEJ+0;&c(HL;~8$?3CWuIHB^Q9lLqPwDv%gZW{akAUS7TC`(ed5q@cU~!y?
zilaAc_}3?q7nv-I_|Aeqj`--3I*b1JlA?^$S?Q0{w~K{y4PL!&l8bcJ22O3nL42$F
zB<g8bTJ}qI^_}sIn@^4UCcV64V5)U3G#^DvImw-L-r=&{+IQ7K|1i?p_4srzd{*e&
zhew9aJ)v(OUYUNREo)*{AR(2nii)~j;L$}K+ts7<Po0QvzJONsmHX;mXWyUhdVF_-
zz6>4Q7hw7Yc_8u908ANxC_H=1)1r>`T?v53dzg6GV-MIW)&H3nY?C+ai0zB%@g?9`
zLI>+oa4v=Mcs{Ko_Sz1VEYHCA8Hp^n7k#z!{j;;jqxF_KEDONHZf}?8K1nzJ7FFMd
z(A&hV;ICik!KVH6MLpQuoCEM~d;fTI*%H0~2BGXb-#VYt@p_x})N_5qxD7gu-!HCt
zyxZ+cg#j-uXO|n|eV|VUY!XGuY%)bF*cB9|uqhO!vZ)mHfV_v~WCJym>;_w_7oh$5
z$bEFC!AsY`>>9ClUK#>^2$Xm;_|u*q_tz&qo%LH<SbF1^EHByE*!W{L{eHjwpB{JP
zZIM_Z@(>Oyh?l0yr&2UcK8>Q;^4S!H%0np%lZR0hDUa0ApLZfw1x|KoK+(`k?xk~_
zyc8Q2ONEbSFl(kAz%StQ3kfG!RCK|$-zLfNmN8-(<B#7>C`o6;9W?J`#GQn3v_e**
z{f0@Vxd#^a$fZ{$rY88JM(kE}pPh9T*l*}6u#=HxtCnUf&{#%#9M@sy$&?}ZKN0~)
A{{R30

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_conv_fwd_example.inc.C65A743D8AABBB7C.idx b/.cache/clangd/index/run_grouped_conv_fwd_example.inc.C65A743D8AABBB7C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c17461c5ad93d50763598858699fe37a66a52e5e
GIT binary patch
literal 4326
zcmai%d011&7QpX;3*3+~1Og~aBeFzR+pxGGTa^mIBm#nk7I0~i_^}8iBBE7rrz-er
z(ej=f5;VLDb*UCxtXjbp1r!(1>Z3*LtEla3rLBIgZ(^N$r|<nC-$(NO=A1J#XPGmT
zWYv@@x1<a+`CV0MezsvY%P@>R{?{9d=gkC$*=E5orCV#Nr%km~O562A(>v<U%;?*_
zY>Y#!Z__u+-%qr=wK=r*a9!`asslGRUwTwflK*;-|MuY4+ly<*dCz-bqpoFc2DR5&
zHvPV_(rIDg0;ihqyA)?DOa3L(ZTd&J-_^j~J_QGv&gnIswsOa7Yi>jrUK-qg%BJPU
zF5eXg=FYh=DX2d=bpGXR!_Pg>VylPtI{Tm6^;ube)pr|?o_oKnDtrFLyw=ERAI!{c
ztAEzLQ?onGuOh7V_L)${5&L-gUb~rxde8sd^mN39Q}2wc(Jg<=*MPQPy0)$|s6-p|
zCOCAjp1rK_>sN->)JxW9r|$8I@aqp;viGM~;y2tok(R4Vmb;f3oX*K>-~PCAmi|DZ
zt4o6+TW<{L^!eqVyH<U@^L$Hx*>^R2zPh{W@seic+w~zPzk)?sI%ECP!-m6wiE}TU
zShu0$Kgl*uvlWd27Zxt<82ZqDVD!oAZ~Wf41s)3?yvyhp<<BX{@eY+%^+~Ia-V<Bc
zFbra?S-0mOjL6DTgvAv(3zg~C?|gz72^M#NRR@fE@x_QNfUy-b_eZ!)i23cB4l%>9
zN|aTU%kvLLWX+nz&rbACKGsv8EVkgVILtcC<@tws0X#oboIO{61kY`$6gTeZn6us*
zA+=oX#ZQVf9vUBkv_4vYfil7~q6Nx~%$y`pR&17v5bJZY`aOr0VSrF41a?A%d1dn-
zTl-Vl5B-W+Sa21buY!X3Ykl$9{XzeIwiXDbj7)irOoULXN2-z#H+UlPWRLA-Y~2l9
zH#ie(=@EH;NKJevi><{ea-xr|QR+y6(!^<$EFGouvf*EqZ$5bIqu(u9)CiJBkP;&^
zD;mL<@Q4e_va_bwd+-)oVR$QyA_CXT?M2M@5!o-DLGIMsyvb+dQD9%hhBzT#xHIN7
zc;CfSmxPpzBNLC5R7|MkI{NYOR*-E44<Z=#d9iM=JYqGTXe&szf;|y<s%Jh)e0(&{
z3ZZeTacQI=5Q-PayAmZFo8Z*f1xgQ2r^pD=M$i(euj_Vc{S{>{%c3~9xL3$9i{jnm
z`4F;5<*xD;WC`*FK57A$TYDM<He_2OB;}+o<P0E`AWm>6CJ1GDXN3_8;byEbl)bqE
zyY0j|Nr?&hwy^#4*!)>m3_>fxaV7W>>u~vx=W?xQ|7gP^wM0#2QzOw(q?KqXGJ(WI
zXWj$~laP59nP8#`BFPeFPBXz<grgr^dDv{)UV<w)$03J`vreXS<CUJI9$8f+`2c$r
zHa;w#q-IL%s)u=lb9Z3ddJxrv9nnX~2qL4vakdp3?<BQVV?_xk89}0f(0JAObTXUf
zR9DiKpdZdyvM586!LQAtOi3mmZ5CxovUo}!K4=<nU2liyMI53Ml<A$xrzCa3!@J!-
z|MnqPit>nhRoLT2U@wA<)N$se@r%5I`kZmHqm|L&f?6jyb%LCz{critc2Ui`_1bb*
z9Gc1AlP3@}UB=>T^}{y|;NfB1Fd6Y!_i*2&(5L@?jQ3Iwj^)4?9zvf&<fot^_VbGM
zCDVT06N7h?!Db92!w6-vnG|KQSv(c_Onbk%d(RBKWUO~A??6#$RHys7OTYi$!8LR>
z-E<1AxR3SU@P32WH#~{?)adz%H6FI|w9=BDm!9k&C}X=X*ddfF%9RKk$`j>LlrPGs
zs6bRe(R|T-iV8)A6cvezC@K~eQ=}K^DKdx*6s?7TweXs-n<@yX`YVsEf{B8>8Um_m
zc{Pl!7UU$!U`sBq7YB8DlR7_jTlZ<4j@1yl8YT-fu7R*M5G!ycgjEW>B6&;HPj_te
zIN{42mr-}pL5MsE(}~$So8xtBzbY=rY9b`ECqYNZ4%zXvF2Qo;gX(AFak1yha;c-5
zC(ENKUzSf%fvkX{`Lg*G70L=JDv}jZR4gl|NH5b<WRMvss)Cp*h$EgLR1GoJ0)Mce
zcIHh_|1WS?AFCSsCOPZ~#fjrw1fvvX3Ux)NDbpx2!mCE1fz|vnD|5M>`xRWDbr4Vo
z{Ef4H((MR&($rxuy~^tHrm_i1tG~gGM(wRukWH{iqtNiZpGDdLZ4i;|dgSCc@V(a)
z>_v_|hkADCJap8vqZNW$A($8;)DFYjf$vsk9^DRsL~cwU=X~z$s2J=?qC+BoKN;7i
z)!td~r8};Lw2^64ZKr#r^KXiVd8^$9U%i))ZEZPQM^Z2iN2~0=90*h67li}ozz-u7
zt_n{TV!s6xTj+g$2Em^}6xr9F&XEr^$BWGkQzceWnKnYA5oQv#PyFttzWrlTAP%Dm
zTup+XW>%UYgeaOh!31i;``E|%9bso4;k{2(O-v!%+IKr@VeN=%I&;g5$>5I$Zh4b4
zXAoO_+a_dX655|}J$To8I|uBpTBPPXAwn9Fh9a#<OOXjgCg8bcN=<2V%y-vudu)d>
z?Z7_<TjHjyJGa8?7~Z7~XTx{!BX4aIZShF0#KRwh<YVB|_PK=XI+Zq{d1?85vB7S8
zxpd=~*w0e8QmT5&+{!3g;=Y70&EpF#E!!#!PT)n2Fxm)XNC1wnnbnYe&%6Jni?%s*
zH8g$QdK2fg0vs!VTIdi&9ugi+r%N9Eea?iTC0J3+iK(kP!z+W&=o#nHd-wad&A~1#
za$7{rwb*SjMen)4$GdQL<y0GOk3%VTp&mxp1Ah;nFK$eF`fa}cr7m<xW3N568w`KB
zVnp7^7U<fCm*O~%KLi)O7O!0%HQml!EP-n3TxrH=s57e#(ef<?;b-)iM&YD%?@Kvy
zD7ou2p4pLe<X?A}KmT^<?lNlz$84$NQoe;CbO0g`KoUvTRrwcIy#XE_*t!i|+kkH)
zea^@JD2w(e!{N;F&Y|pe3LSN2v_eR$(C%(*=1PAlke<i0)j@onaF>wfSld5;VDWFw
z>ma)hqsbkIP%p@OY3>7ApTPGwH4kkI=-h$B701Tel3DR{oQ<bQ#i}SuU=w(H*c|Tl
zsIOrL_DaUd_%tB&5hy+qVv0};1h)v?=hyqU{x;_BEKeJRMscG=q_ZL9$$3)b#d%TW
z%lT5|$N5p@&-qiN<di)9aYs`UG12D)cF~@*=Z`o-6I2uUwuI1TP;3Ui^CEN@f)5M#
x5oureX>b13`Xbu+Y2O6yO+uODU$A+z%|d<gJ9;txwm_ZxTOuSY_BSxhe*qGbCsP0b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_gemm_example.inc.FE653C75BFA23618.idx b/.cache/clangd/index/run_grouped_gemm_example.inc.FE653C75BFA23618.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5782d15fad14efb113f459a540a2175521b74843
GIT binary patch
literal 6854
zcmaKwd013Ow!r%|7wn6dF1BWCnhvx<plKFSwjd}V0T)CRlTn~)7dn6q2qO@1iv|Re
zmwX0fgcmgm^C@VIlTpV#sKlr-am<*Qr!fj+jH6LU5r~PBSKX6Z{u(|gzTZ8k>efAV
z&Z$$CospJy!bc`cS&~s>S^7tlAd|^l@L%O0D~d#r$)4lCn)mD9IhlDn!Z~$h?#f*u
z3vynaHt?Q$-uupqCbZAj+_I<f_{J+@ea~;1K0ZOy*0q>F`)ip;SmDtXw`0?@XKx&I
zcpi1Uv*T~u)^9Xj{pj+t>J>AuTQ03=id`1lI<Nb%`o?p4ntOfMvOVAJpL=ar+vg|$
zg__nh&%b-|9f;`Pa4Y6@FpvGQvhC#Y-u1nmy%|rt{Q}}D?ZQ6aUht2_UGKd=IpTvZ
zho}!%eBrmkaByGp@zSTN&cB?DE=%5_`L*p+Wy0XP$b!!0C9AG|kZ=6okcumdOd(J5
zKMQ*naUrrM<K)SoO}^FsU_brL;-1562mfq5lL{Fd9Vf<2=)M+yz$N|GPP=})JA;RO
zWR<U57M;M>fU7<)U)S!LCdpi7GUQ?Jp|UkH(GFyi6CiN@s&7hOO7&W$Ts2mPM2#3|
zYh(&+R{W#60C~5c-(IPdyJN9JqX@J$M67M_SY+wE*x(>{#KX(S$X#uXOtf{z!%w-t
zwe#lw@778(H>_|6?C*f)_kS``kpE{6IV&FAzVzg2tApGgE1@yS!`84-woSXLolx@n
zb)|d^o`r+PL2GNsEV}*><8l6+{}iR%#X*K#6|Qc!hKPHHe|=?cZpO<>xiju9IZJ9=
zL&S#{w1|ItA*V?xSK{6-qKlWUA>s>5FNe2&DkduB3fx=J2wGc1#MjysH6x|YpDX2_
zxVM|==5A|<_@@7=dp|GSUZ9kF;NFfJ$3R;{#82UL_!GbMca?Gxi|sY`T3bWJ6-x^%
zKla47G=4Jm%cx)k68Z+T-#`#qR`Or&_@8C9nO!E6P@{Zoqg+j_FJB&V)4ox7%T7W=
zpdErBB9PE9Xoq<o0qqE5yU4NK+b%46Ta-}I)S@)<Itdj=6i?!4VQ~SY(&*Bc7?nkr
zaa2CJJekQFz|a5_7&n5Uk>_TZ*bH%09((g2x7(il9|)*OQ=}!2oy<NlMZ0}_@#6u|
z4uC%qI9H@!sSEM0##YNq%zK$WKI~_S+LV6LK|)pDRiWgugsQ!(IjZ%p<*3fPj-%Dy
zt2tWZy@nDACp$-57?JppP}sOI1AEmlsD{Chtb(d5v~TQ^4*^f~ECingYFW=y<M#cK
zQk|rb(2M9r57vK%a)yf0Oyx|DX1UGsVX{nBrZ1y;s(BpcsB$=(ubNNE>x;<RbGxQq
z#xv>z|31)@A=IyB2ei+7$r%fJV0;h65P@&Tl?!)%bS%a$R+>;cojjIMSwa~{<tgQJ
zh>YYGY%6PGVYV`iGjO&V0@V<+l~(yiT$AELKc2uoaN7qjkkLqYJeoBR{T>N86g+%A
z#uFt8-38$;AY%C!9N#Uk_i}BM3n)H0ei~5|P-1-IWJ3N~&zyaRep`-TxdXg*fQ|}|
zRJ``n+Jpw|dHX=Q4_v6=`-+Y+e!rZ>JuZURMNkug&h>`u?Cxkuj(yKil}H|A-!oKG
z3i8?N-1&86i6Ed{-`oJAB%plXe2xnI3+Q(Vs8ClJOf>Z?Ry@4@WQH2QWgB>HgK<P4
zp<c-9g<QhnBd;GlF`V9x6}N!b7NDa`=mcb)fCWS`-ucv)1MStLcy-sw73<_mVn7M4
zmn+uu{H9#-CgXU&N9Ri%Opg>eW?hTi$#|Fois-ZiRP0*J$%><kW0__@xb_1l?T3(l
zpaCbKen{$vsZ4(WTnCtbLhSMG8|Uht3KBAg8pDX1giN6(j?AHEjx0HrSBPdp&&919
zj<j@-kx+_f3UxOL<?Hh4EU=sVmc2;2;N?uJPG?k?RL4;xxHSTeGzlGrtfTM;qMdaA
z>j!&xZLPvf(gK<m;0A34O)F!v0^msY)8Ab_?O6STWj_^e!%1V3^CT{5O!A*Z9R|xL
zOrl1X(5vFBh)h933!E0XF|w#E)PeDOgVTE8q^H2;6nGNtS!ez}=WD~i&*G5i0c8(M
z#<O?+>EX7}gLPJ?PRpgq7&~=ZK1T&J3TBeAuv5<}%p^1~({=ovjS2QhLV_qDvMo!<
zUbIJKAC-`U=zvJMAR&dQKqSpbNEAgx610RIMMp%oX9@X>{+s~<!~l+TqK>0vF`1(j
zF@>X4F_oiqF`c7qF`J_<5W8@F^m}<mDj=M^tmO}vye4~4dr0UW2=~B^to;0`XNJ~)
zzbp!8nLQxxVLr96Fl^3(qaRt5v@y|0t!wP=KDawO&WwZj83aCq5HgU2o<rbso<|{Y
zl<{l1byqq)|K}DS?>J-~hlOO?MHQ7*vwz+>#a=>L!CBl!UJo(#Fomc|XamG-;JFE6
znt1L3p$D9Zeo3}5?5oGdhxmQ5?y=PO%I#Da6b|{FI5f6`*H%7TeF9mZuxwQ!n0mLF
z-Z+F~KF=qQ3pJCEi6gU*nIlWMC6Y{_vgfN~?==>z!j@VPvVcnr>mg)4aC{2FPO*Ze
z`nRcP*H(2;#uHp^SWWFNp*4mzlxj}*->MZZeStr&3shY|_v+dQA8v6MFWtf~tq1RV
z7U@>=d2isk3A~#auR7cpkT=&Q9glbhg3dq``HXcR&a~ugx%3&f!)DNLhIk^7P&??`
zdF}vx2hV#TVGqn8`fp9Qd-MLgzs2GlD7Xu1F818*)m-ekJE*zXb62Rj*mD=vT<p0!
zs#)wwI8~7%{p~m6I(9D~F1D_Hb}Y@pZ+Ep_FI)MqYfb|4kM*Y|q<{i60W>FL3J9PH
zQ$RXT9ks52BBw>t0#-oDx@4NY1(c#o;V4y?N?k!f>AG}kQ2}N9Wz+S?6p>B698+?3
zHZhriS2wB)CKBE_H4Snvk1oO{O!iEsQ<hM!E|<O<*G|5r^aF5>kW|HO9JNWc95sMj
z0}FWx9f7PPY#V9aKJ~qVUn{I-l7E<gG?~Swk%GK+$oFgPp$EX{00a_&gswy6b)Yp*
zTkrH$)yoG%aQ^`qKfu0&tgpSJ?FdGyE(z&1dRmf8Xb@C`;K>rq5JV3_EaPE_9%lT(
zPa#>SJx5OAw`~H=CRVU)Up7OSKBr*-*JmZJC0r3O05JpXWU%9*D)`o2-7h$FS|F+g
zxQ(q9qFQ<GgDHJ5jeIsaIpVx)JvNdwETIaI3VM#PrbqniLuOB0RdBSOBbf>Obbiz~
zBy=BK?t?GsB%y~e?jeK|{wVl%&ca7?LhY;tj9)*iW34=?pB2_t&g_RQGUa_Y606(2
zX1|8NwE?0Ufci8(^g&c3&&?3u46{hbzvO;9VaMCZ6|a4fQxUCIB|N@UKdW8$N3Ptt
zaew6vn~O*X*dAqd5#4p>GKbPO)$E<}<G1(=o`L=u`!oR?OaGkbQP7VvCI@b8bn9uG
z8T}9_4}pfv*SbsLUwTS9+E)J0ryqHzVJF!OPJ5ZfjvopRj$3?w1Piu<+jf>1PfW~S
zVn1!iJuHZQId(P~_(Zl?9A5W|BYvEhkw*=CI`zR4Y3O9VgMfmy!8Aez6rv5`NUzm%
zWH1=$k_#xz5XMosA)M0bU0J10X^J;($z5L$G$O6ZUB7|nCeSxAK7IJuFH3hH{0?`_
z3(ceB;ho*b$nRdsmQ?b6=hZa>VvW)s*I>IqvkS%(0m<9g@YZ__=D5L{U@mqaf$9;9
z-7j@@jq4KUBJ2+(J|%u+ENfm!8h{kSmu{abJCbzy#|Q!CdF0W05K~YdJ(ysM&wGh<
z5s=Zt$Y~mbjGU%1-pFa1JWQOXDagcWn&M5IrrE>HX_|w~oTfS6%xM-)E}~~Hd;|$C
zriV^UuN1$^=4sJeXpR<;CDD?^$>vzNq+dFrl-n%ICX{hh9#>9xU;$OBs%UL4plVe$
zN42V2j_Oo(9IaNZ=4g#-4Mz<S+yEhLZjBJ!$a6D9Gy}J_wSZR(cr*PL2ycN1#;xGh
z%FA0Jyp@-40<TTHd@}@YX6`27DkHF+=MD($;JFWc`&hsW_^u(LkL^JMz70t_%gRIn
zor9!vJfDZu^DLPQ_;M`n8&<_wdFD4P0Sfr?D|QItND3D4<yY)5&m#~!!uYDkC!@7V
zFVxz0mADo*F$uLoTr1Cgkk$wEn&avZ8BQyMYWL%iF=d!&f{~Cp!%XQ~{SQ}8{L}tp
zoQ3M;ih8+3rbTWP@%Cc9PGsrv8Yr(ZYg^NC=5@B?OK24EFWsfS&-mhK+?S^h;({a9
zGnJ++38j0cb5syiKpjFtg&~F1VI@@VTh0|D&EVF|7aYeR>ll0e_czBcz7W`c6Mv2|
z!$`l96nQx1ST8?W$_XDAPQy(?BcK{#nf`|e`zK?N=UaHJEWfNE;*WQq<}O)e_bd^c
zwboEeLsvp|hB`|3iudNkxg~#!mEt|)>AjVN5<L?s-D}x9yhVS}dPGikPNo5bPx#I$
z9Hly^a+L0z&JlkzgYN>kn;8l5H#69#+|A7W%A*H&*NpEUE1^iYNG^Zpy5_3McjDn(
z^EoQ;EZ`Qb(7TYkrr8YAX5cRW+CggPxdWsQo*#kNBk(7#CZWgR^_b^D@ET<N=;JkW
z72igx@j@Jd<Rd`0(4RA_=PwK>`?uA@6HWANgFQUaOo<!}up?V<w%OW;XhOK{<}s)q
zv+c&}3(+_Z$rOKSdi>6Vu&cAj2)MLU<0}zjB>@Ma8efSp-UMn~Ju==5!T8w2xE+G=
zv4?R71mj~5<37;h%TB`2d%q0qNT_-LzdF+B#FdNxwlcaN8)zqZ?_{M%#nNRa>nXmn
WVsYhI{9lg39?Nv=T>m^yCi{P<$n5F>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_grouped_gemm_scale_softmax_gemm_permute.inc.0CAE0C43C07E96CE.idx b/.cache/clangd/index/run_grouped_gemm_scale_softmax_gemm_permute.inc.0CAE0C43C07E96CE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5b78a1f6b9d054f8aeaec101391eccf112d3be91
GIT binary patch
literal 7728
zcmb7}d013Ow!k@;TS}YD){S%nO+!PABHgm@0-{1x1OXk421G$)RB#DPL^7k$YJ5VX
zOjrh$xPS_Z1R-%y6WlV6afva+Eiq^^?h0zgK~N#DI`7?EXa2<RQ@<bQ)OK#2tw^36
z8#_v4W;S8|<dxYAa_8}8W@fhZZ%*!ei(b(GqruE<<+i-d5lc?_*u@0CnE9pFsL1kK
zQTeMOpk!y!`0!lC*^XAjK5yOic#FO#fkor1n(o~w{P5hy%h^{KS4E$OCnFk{#+THV
zHec7zZ~tu8vn6r8?pGf7clFNtYO$`W{-NFB#=`i6>!!bW_WajVfe+`MPrT*0rK|h!
z;zJkO*BZC%o{+r!dsT#b(H#3^S<NX&tJBA8D_oOzg?zHl;gIvAFLqYuM6QbaU~2L=
zbqh7GgJZNc0axo3Z!R-V`z>qI2T8f}L(=Lj*Y~#DF05Xm8!@CMMOPj-sxR{Gf&y2~
z$dT*s?|r&Dmfx6vcE*wqvqIMyoO{eaxv^X@%PGjtaG~e_NOmSOec2CZ1+$@MX2hH~
zANt2XW-<=Uhfrg^x8R+eb80g)crrC%+=+<~%z0AGxfFA*;y|9P<6PHqTE-hV*A0xh
z6{ByF4SWeNkWWGWDL9FW0@)6V?Vx770~9+j?*#cyrf=^1M#J{=8Ko8iSpyT+z$7-X
z2x5w03gcpkDaLGsBqO9Sc^!Dwfe+(7VAunGjO)Qrk9i-2?t@5CZhog`T!>ZdYY@l<
z7<~c4M8S~e7jLb$UeY*d8x8sP*rUgzn}<;wZ3M54FjkyVASIwLfe6N>pfAN-4qoL<
z-{NTU^s*}MXS%g2F1(7HAPo$;)p57-`E|mRXxC^r@y9$h<;XSQ$+!qc6~So6#W1QE
zvk^ubnZD&h$KXZ#)yFxWbb!1A?AXAYV09C0M7Bvdrp<nAbf+7u2Cr(iW1GX8Z2yW$
z7X{PNl^fYorlYG$z)PG#Af=!x#k>iWo3Q?7P;JKYGEkLaE(c{f)~^6n1(sKWsuFWG
zD66sl7Eo=$@~xoS%Gmy|lgILZi!2@@kYHu7<Q@Xax6W6wLllM<j%TzF6#MXv?*~Oc
z<^fO)FjiggUG+_U^bgdCJ>b>@dT}CWL(zeI>y2Zn;1XC~0%>(k_(jDVuJ_B$1yaN*
zi#UxqLLh57)mqMl@rRu1L&mzaDr39*FFyfKo`c~zj1d(_CG|#rz47>FP9Ps~J|A)6
zqCg<EoKG!dPe+T4o$BFFsk;U$1EugFkX-9r2R3L`$SP!y5{N9pj=dQoH)F1Vm<pK2
zM(+W|9&l#d2Z}z%o_iKVRIhG*dk9Z@LDLH(L`9zTflD8ZWc(6bUSfU)uCLg)=gDi(
zyao@Z-w&F8@D+J<-s53C!yHdgGgpAV0w#(AfmDLN67yEjZ)NP=9&+YyvK2wpnjdpM
zA9Kj@>o}h}#@;Uqj<)wqs-|D~5I7xTNAXGJ)K<enc2GeTH=>G@qL}YrRzL7nU5ciH
zZiwuLL~&065nm6xRxsO+2&BNaKq>y1l4pT8qt7_|&p0(d`m&Ef9eq<E&76HR<`&Mr
zg>j(wz?SE$?^e*&N;!{Gw(;P}+qPfY_Rmy$!U|4Vfp1_F=ibCgp(f<>rQsIApTD9r
zTm$E8><)+7oUHkx`uzbK9JLl&E4D*73pYgW7Ve027CMQ-G|$d188YvgoW?$@XelI7
z@{X21p(zsDuK~$ZO{`zT?!Ad8uYovmd7c!3eG%3#0z(l<1~kdziXfip7lVB<)-MJ_
zG1f1JxMHkt1bZXaH-g@X_tOERJ6H(dNher!vVd!1k4_lH*6)PyPL?@L^0zu+wkVHy
z`{#_lT=^!MK!TJ($X)ZS^N_n1stYAAFh%6ZQtX~(5Lt$~5@IT`yKaS;t(faUQP12}
zATL4jl5tem<?g2XPxGl)HG*>^yYm7$0Gb2L9R<<^nkL5Mhu0kab6kRkUZrE;aEuKc
zzpUP%i8?&^1U>Oc_tZ8az(~_)j$@8aG#XEGopL36@FdSE4^h5jzT}NO834@y1c)PJ
z<o6FxZEjW4EtYeB<(yPp#D2N1JL~Z~+v#l|EgNku>I)=F7KLcM<#+{?-3Hs+AbIkX
z8zKJT*4gjV*j>(P%4wAT{~9lTd|{?}p-&=RbSpP%D>qIOoc=E9vf=3&I^qIYT>$BT
z2@&0Y%{Xtfjtbhqsf}Hy842={_shl-7l9aL2DGs;vN4FfW!{K<WIl*|Wxj~~WPX^2
z`=L59Ko)>xfwDkEL9!r3!LndPVX`np;j(Z<V`XCzMaUu$#aqTB<0e=pAWBpwqI{F2
zOhPo>ayrtSVL1a)iZTUh&Q#7sG|O@p(wuEM8&Rq<6=|j^(-5UwrX$U{mU9thC^L{|
zrZQ8a88_xkxw2(y)S#=>l<3p{F?MG8)9Z5)WEWXoGAH+%-~FlczWifGT``C9>^g45
zIu7~725!U##_0_WhxR9jE~GVr+(qs#UPpl_Toj1xT<j33TvUi0TpSQNx;P?IYt&Mu
zD-dUmvqTxr>&KOtExkpLQpu?*IjIuJob7Pr`&TvwDlq5_-r|ljkCQx=-$he$>Vn8F
zmQ3GGExcN*{OC3nyaL55kP56t@8#rA>)kt!mV=?5p;8>6RT2c(g5-Gu*$pAPAzEBW
zAbTNXFXjdaZov8%fWH8iqJCEL?jK#B@h1lTF|$6`eO!8S?`gX6zk}1?S@2x3GV-GO
z?DzdvJc)CSlM(?>;<fSq;<7wRlqU{nlq63=G+jGg3UoY4k*6TpO!-Viv$V62W~w|D
z$<pL$h|;y`X!Q(v29jmUGZEErX*Jv&ac?WWxE|l5G1d&)K=1z^eP&I5Lj*O@7SL~j
zSaHTx-3MyyWS4H!2we*fwJf`szNO=CW=Yexbli)%0UR2be(`~+*?+3J^iNudHG@|(
z>mvnn5DW)dk0g*5FtlK91+P}De-WZCLV~!%^`AeRfBV#wv4i%<jgwyee#f>*LC?I6
z3)sd*i8Gc=F<;xY>Gu#DfyArgQP+^5N<ft8mMFE70!ea9LNr}99cj)`%|MjmmVz{A
zy3Is1OEn8=&Q{Gvl<Jm>G}GMD5T&cqk>*_0Ttpde8AvnJEmNY>y2_uIeemEXJDwPP
z4N_L-$r$G`Qdh+jZ)b1h89vTFh<ug4j-uUo;^*ckMG>9^xCcm08&3jtfn%6vkS+*O
zuzN7l4AX=mS-2(~%{f*x7Ey#I0%>Mzv!$5Flcn0Fh?Z%WAzH3oj_4G4oWkwU*WmFr
zV{y+g?El`1=rfY(Pg!&<r2a6FF-#x<$^gk)0?Dz?k?LxJtWvL%5`?MgkJ!k3OCXyd
zViV?ah$zQg4H4Ct>p)S*GLb+VJw-3(*PwXKxZ=g!v|?-ZWqQG{f!8$%7LT$uVz_d7
zvc7)MF0&&NJ}H=g?i`(QJLkHclhV)5tT=v3()wTNj7FGggjwQJrl*pG4wgcAauYmm
zveIhTli{AXAG*J!nWGVW8zD#>XexXC4?qaxCh%{<+yk*atVIz>8yDGz-he2lwsG^s
z@wIs;hYn9~ze=}W!v)l^a<Fd0Th({++SRo1-UMEoaNyYt`pv9~5J(y5%NT#=dG}9m
zKTFcnGk1bxC$qbWeLDd~@J^W02~)+bH1q$-J|B4I3Ej$GE@Ceyz3nyMjC{XZHSJv*
zZ0o?cj&;rg*#rK2@J*-(|9Z^(AZi~Ziz^&Vz2>*8)LKp_ik3x7`P1~4XuAfa^dZoX
zW7z?=;xK`3T-M3*?H8pM?qPlM@90K%aUQ!^pW7B&v9f90y=xXck?Z8B?NaCzi0pKB
zh*UZiA_tuVB1fGgBDK3(s!@32?CvblS8bl2p_yl2(2r%XF-Sfvkgve*E1Z1(yB9cd
zWw*cCx9yAmxbTB{&plP2xYRP%V(lO4&#M++&)uNk&0NZq{`7k>H-J|I(?8RA!v9GB
zHxsBSkATw=mKV=Dt^8HC(u!E}WPx@8x-;)+-$ArcyAah5?`q#gv`D)M(PHglL`$?w
z&@5TnEJVk^;~0D0<jHaHIF6rzPJqV=%qPL)B-Z~2c>IIi^7EgrPwL-ewTzzdzq_Yz
z-3vDqM^3k<-^>W<M#yB>tpidzSXVC4*f8cMdy4+<ry!euy|4DAuPJ4mri{I^3uHg1
z+0XdWq)q2aTFyBQ`tx48dve*4s2}b*@-+IW8bRv)c^Z8j4ggARP2kXkxfwi~S(4ys
z5i{x_HcktSYQfwJ9<5maE_mL>#=Q&ryI>Ga!;^d9c@N9)f&Lzr-v`h8SbiV$_p$r|
zcs{`L2cUm|<qyI0A(lS`{X;B&1fGwu{1NCMVR;vXb%FFA%F}1Jux`vxA?PV{2cA5G
z;Abp)^W--O{tffLA?V*&|2afHhbYk=S2Rb-buZWRrpv`SwQ*9<ekJ;NrasQh^y1b5
zjvdUxrpsi16Qm6Fqy4dgRZ*_1s3$i<a3kBQDba);U}4?#z7pDmxd*29unhLox;t0f
z|3!;bfsC*iA&5J?F?q+ts-60s^v8$6=`f2$zs$+K6*9i;jX__|J3i$3P-|rgoiS7s
zDuu^KpEUK<cwO9XO$!g30DEzL`tWQMh$zS=2vM+2Frplb94j_1*CH2Do<$y_e2aWU
zy&&|0oQ>=Qp%3#*5ME-Y9po#_^p{AlG53Sek9h!u0mfZ9M-S~^Ir0};8pPShp@MdS
z{Q^YIAT;BYb`UHN0y=C9Shiqp1)-JccW*e^_U*{j=hVpU;Ly$<gL<bl|LnQ8y@2*B
z6IBz@@9kn#F^Jw&zKPy$CfQCx$#t^rWQkt>&u<@hkNI)hP*bFFqAz@+6-|*wO<(vJ
zH-VbI@G))%PuejtJ_ug4V`AI_UbJIk+zOtwV-lJEPBUwssbI}{FEi&2`aeg$Th?W}
Lyeep#nc4pWQs+~c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_groupnorm_fwd_example.inc.85BD21F884D2FB4A.idx b/.cache/clangd/index/run_groupnorm_fwd_example.inc.85BD21F884D2FB4A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f885b602dd2455bb90e3a160b64ba971a21052cc
GIT binary patch
literal 3752
zcmaKvdr(x@9mn^`4_(;HuE4T97I0s<Ec<Ymr)1&6K6ne1)a_OtWd(uVDy%$I#MTtG
zydjezW=0#Ml8Ft5nAbF89vw9`Xw|6ggbb!uJKBng@zpV@lZmy5cFwuyuQIdje7@(L
zd(QdY-}(N|K3i5?d`K!1xi*ybwzu@G2a!k=i9g*vU4cpbeI9>$U*0lNA>Cz_i>}vv
z=tFxRYdYtYy(72QC10LztmP9v`D@GJVawHv7n%;4-hVH1XVi;Jf1fioUprk=K6Om?
z?l&<nuK3H)s+#_nytl_&hKHIrk2N2ecz2z3VDF@+`n!FV#Yy$tzK)mX4otR<{-BCA
z-+ayE3h#cT&YgQJ>ByC*QQuZo&u&~@{I?)l)$c0H5Or^FZ&KrRc`lw_c>cKVQ%<y0
zB!W-~P5SRYQP>hB3JpPkL|rYd-LFN6M2PD-e#3k2mxzOaI$$9K(F=NA<Kd5suxwY_
z4MYyHNv028Xd07pU`#WvA~FZ2Y?ECebB1}fK$bKM)6-gFT_d!mo6>Cp@mfBWAqQV9
zkVETW+8hoC)8;fftwK+i(Z!J4=w_(MSVSqJEg>iHM&d~X5Re9zlLy3oX6yUo>yFPu
zI2b_E0Tf3XI2cBXVWbjx1jUaaE#depU&rjY^AGbf4s;Ton#dfaNKzQGYOT3Kn{A~n
zPoNH2M~Xl}S&*SE_AND}Ej8hl<*&}!HZB3^(061J86c<+(#RW*wB+x5?!CHY04jwl
zR_LTws2S2EXjTYq+IVe}Ksv3?Bv6Vrg}z9D_JnpC3d0|+m1IBeISU-{**ud8j%)`*
zPTuJvkK@42yD1spp58n;e`Y7nx2^WA<)o8o=*WrdKioWuWo3jilE{|k2is(qe^(mH
zfmWiWd*Q&KH)N1b9GH{MX##D~ZlJ5@pxfSELfX>3{Qc37?#AL-JaHZ)m526MepXsL
ziRC_3AB{t1{^{`f?X_|&*Fjx4>HSRd!*d&#u74N`po~(=i45S4@-pO?`e~*C_*eRA
zQe!IcJIJ_#yL0!Bz5dEpoP}B8Sy4phAX}QP5GW@iCx(!FMsh=b|2Z#Sm`~@^6S?@~
zU*FsGP3u7{H-$CHh+H{)xn|D%;4ID#qtqBpWB`-YL?a4dmYV5w02Zl*>1mZ(mBKii
z)Rrtzy|kXTt*wgE-+m!4bTK<3c=`Yi91#wNoDoij+!1a{<gqCHe>R34Zt?FYj@fT_
zU)z*{ch(?ppc`&buBf`8`%OHSJ5xKKA%h#Ir{90$+1xgqG|`dKG-@2kBjpUmM8+_r
zh*U5X8yQQ<_cM?F)rC`8i<?@@d&w+#Q)?Xzxp)_SryLaVMU>hy@2uf}l(`)*a}Sd5
zK{V+8_1yVGx7?@kLe&yA%^eObN{ddIr&-%fbCQF0SvzA?J@y_q=^5zi-dg?mw}sf2
zj9`o)GWKjRF=PfaLl&?wWCbfjHn1^N5A~FSr&_Olb^GeWB^=}?=hDpPV4Hng5t+EN
zm%o&F?$1x~LCF{AhY=Z|KwQ93p}3HtVsSA;{m9&pXsz1$>7BllkNQ8rfviibqwjmy
zn`_!V2mW*h@4Hl5N==P}vZyk~j>{v<sf}?^A+2C~Rw^nP1Fnj%Vz$&!-ms3WeY8Ae
z&)w&rt`6g1nRppJVjPIYVur%RVGM<f!x@r@B@FR#JoRLpQE?82oN-Qu+;MJ7<Ly(e
zkL3A59Ee7JBP}v}|K)4gB}<&cvM<k9N)~;1Qw5aQ4BQC8hGkWz6B!#;q1P!;z!u0L
zG<845@b|B^-!H|c6qA)eWNb=F*&2azlsQaD2as}rEt<o~IE-i!=3oTbMo=~xJ>A-s
z#GjqNj(txr)2}2l2dT1Dh72+TrK49*CI0Guts3udH&X3J)WBy>MB0wV);z`O*pb~q
ziwXz9++Zbn@>_;mp1OxMCR|GW3I0SPW3u__#^1iqZ`&7F?Z+E8T8*@7;WLzHswP8l
z!O1l<cN>e<!t}IOT75##^z?LkAl|8;PoEl-Md4+7BfPZ8ao`X0)1!5~cyI5q!i9M;
zK(r#7h8{q!kTVpch+#;fP%sp$h-GLHSqBk~3BVZAk0F{I029bOf#`+f1L;?n1~&g(
zgx5HVl%t68p>d=f7x<HfH*067Kdc1+Ud&6#jsYBE2SZM=lOea*P3f#cKl4u7=_WDu
zRRd4o0HzY2E(f0pgM(>vlsK3+r_IS+I$SmvLvEX!p(0xmrStv!ubOv^pTV}PlBgKl
z)$8?4x*C;6X7f$jCdPz2?VZ$iFKTxT?c#qd!-=e1tz_<a-cm1<B!S#OJ_*w2Gp#e;
za}So|M9wqj@kIV2`J>P)M%4u@4<hBD5bG}$C;r`6pz2#Zj>;@*1Nb;<vuQQtAV;0U
zxYz(v4<Lqzk!=`dk+~L+W6lVoLAjYf=!~y=GKPbqk!Y9-ld4b66b6}An3ym$S2X`r
zXzQ|f(X?6^^n|WCwV=hGqqpm6PykX3QVYot00VC*AarN)*mnnvmkwfYRe_3Lw{UF)
zHA5QEFr)=7LpsnglmaOXwTs)OWGZYE;tqy_;vl8(hyOWo&+=t8{;}ZG_-F<^sXTbQ
z=Zo4|d<*d8dgz95B`U3}5T4tR)j+T1*oCuvl%AZtxN?79@5}$!qwue_P3sqLJl$Q*
U-4gu01UOw=#o>kBzYvN31301<cmMzZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_layernorm4d_fwd_example.inc.8870D57CA6553449.idx b/.cache/clangd/index/run_layernorm4d_fwd_example.inc.8870D57CA6553449.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c6ce91b106b3f9fe1b1fb9336b8ad3142f8debe8
GIT binary patch
literal 2926
zcmYk82~bl<7{?#*iI3zZAt8j2mypD82oGNJVmUO(p^;0D38=_fK~boJ5)eUaJJnX5
zN^7mPqmEbY<vlV^r`2|Bs}%8Qu~n*etPX9hRqH)gr&F~{JG*b037Ox2zy0>x-S7K0
zrTIBIvy}`paaMk3b3=O_FbtzbpSE^?(^~X>4t+W|FFo0#IO@{p45h!me7RWbZDqeX
z@nMN}|Mllhg)I|GtCl~$osbrCc<Y^=v%9X`=!s=VCKhklJV|Jn)zfJ%pF8-^SD#k=
zy!&!j<0w--qk7FFw(QW)nRV>w5kq-@-&^%V2Ui}lDbt7g>|JN-=A@r5Up@9u&)EUX
z)jcn!KDd}#yw~sEow;Ul;D(ywD5kBex!#VVdbcX&L0ZtJ)f^MdFc1vE*8l!7A%OwR
zh@b!nVEhfU+nN;&6C^2jZ984xt>(a#U~-a)z^CvYnapF%X);+7ER>}+*IFo#*%R%O
zOoBrgO;NUxBa`S5sWCBIq{f_r(<NKF1Q$h;AW`HN+=N)R!4X(;*BZb<t*15@I|w}#
z)jVc>-arrshG0V^mN|$}L|9}ph8WG5jNeL&PwdP-ho-csTZ~xdpw-Z7lgS_HcVLRV
zcq4qt&8tZxIS5mR*|E%lUa6<Z>adQL$852-beT#ur7<#<Y04<7^i=xrSo}u~<2Sy1
zuqp&VW6(rk86eaYYL!VFsU_P3&>3`;Wtbx@QMS}O^yESTd?p{+BS2|fX|ilt7FR|R
zA^zCq+?2=d{lI~cB2d{Pric_dGn^B#0}iANiBSBL?uNHRm)4;Bs`1ohVk6<et_ABZ
z9DfPPRi3H>EQ>SmF4CO&Bo8g8RcgszI50Ym)N<x1a{`{i!F11bk_XR??^Rsrs?;OD
zd61{TM&qLHotzt7cP{|ItMlryoOx^T_Wqq)J5UNlE24E+<{(BLV~{CU6&sGJsPBw#
zFzIgpNC1=C6ozF0Ud@wu0nBPMaR6XZTPRDb+RDrGY-(GSO!?}3GB%}H7x%~7^bx~n
zuM&s@4n&nmkyGWQNK#3JX4FoNd;i+k$-@q&2<3Pc<Y0<Okt@TMEswb~+=LqIxKrDI
z_x2(OI;D;r8wVEN5-m^hX?#@qRCp@9c&y3a)=~2K<1AD#CNQb6%mEKPMP@KlWC06B
zR<KfJ0~<y8kWZ+!Gh@Jf@`opZ0F6nFB3=rhDXEFjvVDckQ@4D79L1BYP9~MZL5ePg
zD&SOYDya?*($r~`<v7DQs$@Np9;(3=6ckk8g;y5@y?SfSvow|i1*_0tnFA%Oq=;iV
zid3wMBEcw-Qb(6&6e)5VofJt%iO`!T?_WB5`_5BT0(G`J(gD^#no+Y%V;)3uQG5|8
zcN|xM;{WFZy)46j@6DSG1-=!(3>Qq6cLKKBaVI?fw;PVXgHgeRN1L#W3dRyMN~Tyo
zmO8%{zJ;Pzp_Mdq4*Zfo1<(6n)%1;57alPW_i&Rzz%uILQ4vm=j20ug&jawV`E+vs
z9W*72%_51<LA9rvdTQn*%^@A_z?F<e8;vE6XbUEniS#LylIeVjyo!0OnR@XoE(>L8
zEwav#E$wdmBuqWk7u-9#H9B<2UX@pi<<D}~cka)+b|DC0zH5FlmI2!1+Vf=UNa`T{
z|B(9nz={PgFo7Jj@oiMGE#wyxI(ltiW#6vDl>i{Hf>QP*vLZ!J)=81XN`!jX#hf@l
zwA_T^oi=*fIP9R`5VyA{p|>98t`IMb!7`#;fkcQhGG3&{#9WaYb4pICVY?(3MUo^@
z<d)oo`ui44e^kBa3lz^>+gt%JeCm}=my(veaUAs_jZ#DPAw!~pWHP!>-b9WDZOAi?
z^r7z@OI9t*DANqzOP-fp?2l2OkGLq<P9u3fKcCFK`0Skx+qT3NB6*&E9_i{o8NPhF
zIMcRxIQjW#Qg0~v9Wik@R1RY8u~dU-vA0mvDz#Ga`cwU6=f93$IU%z6NjKVJs4|pl
zrxC7*@v?K?%9D!bpxjeV(&y&7{eNu{2DXKuGXf2H`H(vZrAP}}igci(C=9|V(u1BN
tAL~=gVU)6^6qT`MgwU_fd3D1bq|HCG4gIiY0^#1n`=hr{o|wuo{{nC7x1#_6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_layernorm_example.inc.DA0460DDD4EF632F.idx b/.cache/clangd/index/run_layernorm_example.inc.DA0460DDD4EF632F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..93845507d106fdb92e26a4753146c8ab2c94d9e3
GIT binary patch
literal 2900
zcmYk832;+I7{_1RkCG-YFHKBhlGmh10^yPOnuH!4p#cImg#u}`g;FX|u;ogop(#)f
zbvOnXK$IzrpkM_Z1(Xhqqr>6AfE_4@fIzV@paKF?ZW#`h!>!Kl+swRVe&2rk?YFz%
z_kVePXkp=S6~hdEYiL8Xy1oh+hS8!YUSBu05&fP(Ps4_pEA}c6dh|;k_TacZ4Leqp
zEN?7s%1S%4KX=`R#g6Zc4=$CJwj6l0cT3HjOXsRrr9LiC^HeMvRafQRb!YVSwr8tD
zlRHK|I=89&=FEmuAMCyw2`;Y+oEC3ikL0ZXZBpy}`cqEF0()h9u<N#27alyDkoniP
zGtbY4`yKo6=ImWNC$3q$`JLOhp5}IN$CegfVd8V56Ejg>^FN+``R#GtZnI1>!$2}5
zr~miGq$DOVU6K+YfvKymiHDU8ljK!xZ8|=2CC7rPr^!twf;r#plgW}}=_3>0lczkb
zLF;gNEL}|Z$|Tr@9uy4{3S|=QA~hxs5~(rCC3$4eY*#i#ZkL-PkIO?Sb+;{=v$p7D
z5)1ldy;%-mP#UZ<8B>gDm`p!<=N{QI=q#EN;UXq%W}((tn;}!o9J6CGpYJftyLC<K
z%7Rv<&BSIFbSfQ1R=c&YJZ9@|>n~G@rlgxpVNIB#a$k8Ej|pE?7p#Bxo3SYXYNPs9
zYzE*gT)IpevxaOBAk~;kd1~!iQTEi?b>!v%LY5HOBS49>Bv1AXJHzD0g&%$i=0B}(
z1r~&Sfhr3zU!+L#OM`I$7TkU}A=i$rvBvhp?<BA=#y4gdcFLY~f41h-w?!zfMx`N1
zvtV=^skj!Kr6-=k!YJRUAg0`c7G?XK3O$Nj2!%>)?q_>&bb4~tg9LzpE}+Ne0e9N&
zHgEZ47P{*!WtI+`S+H|<qf8F9!+<HUZ;CXw>2L~KV61QK0Bjy{=v25(y5HIrz{Hte
z!Daw6XC~VKuy7U<7l7w@%G1hOd3m0Vv)N<{azQdyI#TESV|D*7ooBBWNB|Z@wMdbq
zmMC(o-GnBLD|3E!<3L_#fKp*JUIhgx6)DR0XAhFcJbn+MDOK#TPp=PYK><=#spK?S
z;CbF5PYG#4RQrtfjUIx>rq;!0g`R#n5H*YmOloXq!3<`KEMTFC2c9A;SShlBjiMj~
z3C(EmpG-S?`gtNijZ~A3mjak9O(uk`3CjOf5ea7JXW{nY(Xx#{9Y$&9a5<!*Sjg4o
zQhlAL%_B|2LOz#IdG;~(p*p#*xi2MfMZ=25;FTBmD1LhRX7gNBR+YX=QqC)eB`v=5
z{>wheEGQL94fbY1rBG2Mm;}=N=o(ETMUqLP$Zc{HT6Of#i)ZfMdx6#*pD~_j!rCVj
z-dw0jYeVK@p_mB3rbsai3op6S*+K&Y0i4R_dj{e64*TCI!z@FV1)EU|S?yUeId}(k
zMiD+jQLU?%2sR5bZ!90r`+QmX`fD@yrFF`;Meo9977R86Q7t46RwKE}9q=f6Jg@a0
zn&MaZNzJoR;j5rtC)H9l5wjh?`e&~<g{Gh_m^>yTQK%aGn?v%NmcAD1t;TzJ%F{Z+
zIzjeK@0I=<rk$1Ty*97Z=+J2f)B!Cv?<!o|&^+))dlEp@6CH`o0Cmo~BAM!?dLsS1
zxtk}K#HKNcEX?3%P<<BX<Ae_0*j{mH>z)b#5EOz+jwC8XiX??Zkz3&=)Y91P$j_G-
zSx~;EJxcrG0If#n*E@T*OjIDv5QH3TMidlChN$!eks1?&A~hy?B}&4xz1bAGy>5y;
zUJs$xL$UHFm0x|2@|l`3)rA*6w&0_S(!BQ$BMDNg)RYA2MLj85bf3J591Yr#uarp8
z3H!Wd3;l0sI`5?@KrZ&2ZEu(Ju8iZz9Oa{A?)jH%)-`Q(4oBu``e{VaFBtc~m^&b2
zZfEiHS){R0@!RcAoGJ^BOa~<bk<181wcc7P-&pS$+4<Fir2+HE=PS_`ITc4Kslj6?
zkb|4k&7^f%DD#z(^10R6{CK15WK#+{BT$ofAqoc^MH<jhlnSX7X+cYo4s;ZS6d_Jd
nqeM|cQCJZsg#KAZsygK~UN<R@{ysB_(A#(-{_vk>1H=3asbxX@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_permute_bundle_example.inc.419BC377BBAE0EEA.idx b/.cache/clangd/index/run_permute_bundle_example.inc.419BC377BBAE0EEA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4a84c04a1c114259904713f17d38b970442de7e5
GIT binary patch
literal 2148
zcmZ8i3s4hR6y3n455fjQVgeHKp)@fetZ9&_O$>^GAmI~Gk&eYtM5t78Dp3S&9ZLbV
zP^Ab|Yw=U@Q$L8bRz~SmEu+Y2tu5M!6*@Y~Pf@h^Y1QdP=52PmGrK$I?ERc~@7?!i
z#@lQeUIbxX6knK^y&{t&2*Mw3=ZgH?@8CU~LlA{uudGPooqbzib5@7fzpL@>Z@<^_
z$zPo5RhuHWPOX}w+q>#S?5V~h+qZ2l>YmbcbH$o{EeXaSWAyyFUzIv~olW%>^K|}|
zIk_1hM^Cz$*){pZ*2U_wcfzjSaXuWNd*^2-*FP=Os4pZ*7ZE8%<teJv&*~L|7oyOB
z<gMwyc~_q}GtRL1nQV{M)5VigXNOunZ8ti+jxC>cw`uvElvh6#bZVPgb^H51ZQ<mq
zRxc!+MR^OSgLd!zhuhEB(VY|l3q)KlH{kVy@b=^o+&4IgGjQ}s>7U6Ll@#F(26nDp
z@%lk{lOCf6`Pn(n5af>J!5fVL!AC9%CHf@Fa1>EQ^Wj4+U1vBHit>sYgR_hOE?53h
z+lPRj>^GUA11}j?UpNrm0kmGCS7K}Vs$J90bKbfOv{_(gtd%cYJF3w@kq3o>ZNVwn
znnLn8c`8Qg3@LxfbMXu?3XBMh!!(70#sq0G(p>&vyZAuKmH){J->~ne?iaTY0<Ge!
z{IRv(a_hm*<?1?;LV^^*987~aty%9zR-@JIMls`KA~7;#Z7e9w(e!yzNbDyTVVXh`
zx#UeZ(ue3(7#a3v)Eq)9vp|YbW@Lm<5UsT*d@m~jnjOhl)fsy&{fjtQG`qZ@&}yE6
zLz?76?MvRs9P%WQN~mJ0N+Q~q7GNVlYFdMlsrl)Xkj;Z05SORojmI>FETNWgjHced
zI@oYvM*(P(h$S*iTWya!tDA?ekR+l4s4z^Eh#$b$xsge13UMQg*uoUV)>1mS^j^+R
zh-(lT#$uy5|6!Z!@$4L+0|NpLm`*zPYwA_{VVehuguX&Krb#676)7-s?3!EY>Rvt~
z)vmWk;-Q_{`B7wv=EDU*OM|7%>eKg3IG!a|rvc5b>VkmYV>!KzCxK@7AUEGx;CQxw
z8q`Wi3i&uCg+!!?K`VvHr*1aXj`)Vj!kA12k}ub7SLC+<ohVOKVO{GA<eJvY+je-t
z;EfZoMoo||nD0(th&M3fPa%_!i8W#7iRLJ`rNzg>`evPOwY$H?m}9IM6(32=ORKwb
z9@e5TDdI6rA*ET#xRh^lXo<zkJ_pf=MiJPk=IMfY#eTt8f!67ClQ6w|X##0aExQH%
zQ47`V>C@9p6Ti*LmqoR?y1`fACs1SGquRS^?+vAfb6`4y=wM7YlEZlyjcpG=Wr#L}
z`P*RU(MOrE(Xk9y#5WawLI_EF(@Z0pZ8e1r(|TMGGFBOzfUQq&jL+jJibu9DcFS99
zZq==GH1+`PMSC%F&prId{QJ@9F+fk0O=P<HOX!*n#Zj~TK$kj5&5R-lm_#!>#5MUT
z?sc8=BKTS*R%Svw&+YDeQNF$mwhGFJ5@THy;`{Jf$Pj5@+Ipk#Rmr7;LroAVR1g}9
zUAq3gy|&Uox52K(m2;Vm1two@r}2aVvfAx##GbyM^=aJ1Gd-bDB7s>j1;_0>{j&D`
zWzT&f9ee!yqxanJAKuahG`n&GTN<9!=`QaC6%<MFa1dAz$!8%GX<{J@X<-o7#^H?~
zOQVMl@rc)P_+;4eDoCf3>6nw_(W^qmjrQwfM}EQy!X$S(kx}Hp=s9=QuvFow1z&rj
zJq2fY8SLY#=pUQ`Ql(lc(_}bow81PS)5=&#u9Y*0$j@GyIik;*pXKy{w+EV9lS_tG
II&ul(Kg%dZGynhq

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_permute_element_example.inc.42CC4C3518B6F771.idx b/.cache/clangd/index/run_permute_element_example.inc.42CC4C3518B6F771.idx
new file mode 100755
index 0000000000000000000000000000000000000000..94fa1311cfbd102fcc0959c197dd5d7b6cca2989
GIT binary patch
literal 1764
zcmYjR4N#MH9DX<6@xu0wvB4OOeeyBb7-Iw73>u)u_%s3{y`U4~BM7kC6akG>BDKUa
z1-vXuj1<2R6z(|W%CAd8py;56gm`)heuR7k!X?xXx&P~SxA(p0d4FI3=l|R<Io4$Q
z(;matXUCQl<!sHwF%09v-?G)5-wN-IGz=^GdfS1u%&S`GC`+}vVUxep!|unOyYJA}
z38RGx!SejJ<g)E)t*nBPo}Eko&Fne1apaA>=~MSl#(W=j<Km>-hkfNK^1d2o*Tq#$
zKioO9#qgdeE#c*qYa;H=$<oQ*!1>MDeQ{nlTm6RyH@RZ3eX+jrP)B3-g!B0BB-Q!H
zAz$K0(~mXSS9NColSfPYt0KxO&AvQ-R4CV%|Hj+9cf4TXR8iF*h4RhQCp_oqqv1O)
zSW1iDUI()ehsf(iAL*Wum;;6p3<ksf`2%ybqhSnt8bO=??nuL6NLveuIYNMuVU#_8
zV2(I_CBU4MYpG{om`iZ`*)yHDuh2+B=b&>%h248~Wj{Br9ROP6rja39ws?J@ns`)7
zCkf98&sap0gg8<hhma!OlU=aDkAWaTupk=IB;n!ep+rc%@5GOz&+eUu20m;bE~0~s
zcb{C_uc(7uy+*$Z(IjCAF}#G3F6)!xk8}MW*pY-Q$JGVVB*B;Py<Q-VT;q#Sh)a1~
ze5<$;+S#UrN~XFC>@%nAaKeY@L!BNc)J|$=lnY3q_D3kR{pp<i^9edsVoI2Ph>rMs
zaH6U4n_{5(e7+db>rFGas@tdfp#do%y%CKQY<D&_98PH6wJP*2nmcQ<&b*!rl>`ET
z4$&Ka`ZZ}VaI%Yz6I>^*1kpIbbK*%6N<8-FLF?!T%TpM&#?{CL$+vc{s`P(918AX0
zNF60ZFy4|o-rNDSttk0sOL5}-$w;^ZF3x46+9bilc@)~ld-c*tQ|)rDQmmv(7P}qU
z^_A4z33P-sLV*UTIG0eATGxLAN=iZ{HolRDNh#m#|1437<!w0txrBlUKz(bTX1rCw
z5e)*Z3Q!pkU0)cF$0hB#3wrqSd~L5TNKJ`${L{tov~Y15P7-Vu+YRNCgqEeXLHaPg
z>BTogq#+)mQ@Lhwc$ak)nsYcD1)|R@@293sC8=rP4WU|u=w^JjsAqN86F8<!DYJ2}
zE=U)Of?Cq8!N;pOJSeGlR8xYsn`%m$BFC&ya*b>aHNf?nyLGz~o5vv9LE%7|q|bZD
zzKZY^4s`Mr`ciIyk_L4c>eL^3+Fmj2*9bu*i{v93j^f9mB>G))i7^x)>wyM(H$CMf
z6ijEXY!+BL%RjhQR)hpy>u||ZW$)>!3y|yO>_xfz_P_Va59$KDKy`+cL5T*Q^fPKu
zE}&P8@h_0AUB>pNG7_$gd4m=lBPSfq^5F7f?K@i*oS>bpUmqQAnyU-=<y%+>Bu+9>
zG+a5(wn184OQC7zz}XH}+aai=604~7Hoa6L9qu0T0KvFwu8k#l8eR;li$<m~S61qZ
zFxUcP`>dxlu@*`iHyD#p)S`%Gt$H{S16l}`LP~YmpGr>~BvOiOkXR|E5N6IP%v|na
WF=ts=@TL>b%Cp`*d=*>5u>S$Y>TiVr

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/run_splitK_gemm_example.inc.51FAC96E5EFEA0F0.idx b/.cache/clangd/index/run_splitK_gemm_example.inc.51FAC96E5EFEA0F0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1ddffcde2042374734d7bcea3205297065da3f91
GIT binary patch
literal 4824
zcmaKvdsLLw7QnwfeBd|3$9IOu3<!ig2N)=rafXMXhysDaNEkvS8d%9gZl;Ay-YgY-
zltrbIv_jKSfriMe?vheK;~^kmR|(9Oc5iCyrdhU1x;^ZknL6J&{d3oXGwb*5ea_iu
zpS|}vCqFwg^Q;%g&3`1jw))X0OL>mtJn{e9C)ZTC0LOic|7+i<f9a&<<9K!Y)Z)Ll
z>Wd4Co;!ABiQb|Swxc6~D|fb6t`EEDQUBKOmf}j!t6b#R<jlv##P_}|DV-b$zod9g
z`}w`k?BBC4HM8$z^OCZf<K2h%qVn6@Cm)}jIKQp4^NkkQrFmC&p51TgJoj~Q<M6)@
zr$*kHUc5au(QR%&8eIOpLRa4wr<y$Fe*9FqwSDW<s~0v!FK8|L(XUl!8{=9B%R&yV
zy0Cwht>ejq%W8hzm{#%5H!BZDd3^MeEvU9~>q*=0+Or++Ma>8*xDk8k?aH&s({FFM
zF>bC|c!{$v7%Hvo+|;ap%6NCq_lx)aqF(%_YJAoUEwKqNpT4^MWMcLE@78l`t<_~+
z*bg(dXv?Z?V_V!5Dvm>nX$miA!zr=9I1Uhm9O>P%`|@w@lETNGLxNiHb2gmdRNxAp
zx_RxBAO5?`P2r0BqiIO%Y-s<QO#iGUt%`+`LW%pks$D&u4e4J~mo`z-aU@t$Ov7R~
zwVS82A>xf!?*u8IXj>#HJh50*i@wf=h<D}hS{*T7)+s5pSgca3{G1IDAAB%TJ2V(F
zS5m0mIOHXGc{>|fT>Y8nxyG+DL=FcF(z<GWoDHYMK_YtKwRcpSe$WZee<pI?Sm7wR
z9)-X^e>f$t_(Q3)c&GBU;DYUvLc)`HDn0$24VmPWIil^mYey0#g@8@u)x4*(A>xaT
zLhaPw)Wed(7x(s7diyvVBEA;Wf8fDmmidxGgL^C0N<U{q#A_aXZ0!)@I2VI^+rBep
zJKRKM3O7X`Bnw04a5F;*#)L#F%Sq0eLn${omm$l2mK-8;8|}V+Lw`i25K&5K%1rWE
zM6*L@%jB`rKe56d*rVW))ywKblz7w*Guk19u(~s-dgkt|F>Gjxlp=q|V@b0aN|(|Z
z%8)V`%964e%9gTa(j1KKb${e-01>?c5wC!O%&L7gJB8mfXVWwh8QqPtaU!zlEOA6h
zL{&ajcTw5`d0XH?$_;=Tz>jhx_%+gniKqn@w!i~K?l)q1dayO^Jho*sOy3M)L_iYG
zMfFFb{gFnTh_nG(9Z?d|`@p{shy>t_iM!?nyd2(-w_t;~0Z}y@gxDaQaImUjb-kZg
zBi@9lc~KeUGmm0t#m*)~5?Y~hCNi-{L?fUXfxF0*A(z&?J}v0y(|7^9z-JfulKxTt
zd)2maLlw`Xcx`+D`OG7O*1(WaYm_68$DZ?VfEc11H)Z{(>vmQXUd~d#rJ>}rh_*rA
zHYg@MOMR;OgQj(NTt##agmWMf+eCC8g!3$qgD}qWryzXF@&pJIlxGFq8me+Dy(RF-
zB$?D?J|3ARGeZed0z+24HI{1H!OISEB=N`&dOQ3@o@$3Qs^1P??NmSIt+BmZKit`d
z1JdAbkcYu(V45v&l!!9YGjhogBFfd}GQnEmQ^5plJ>=E1!P)|GEs#X|rhYSiv!iuy
z4fc1TrjY4Yq$!dq^LYKuRW}~nhf_059i}CHF@=OhGGq*s#}&~b@HhnCRC)-44*?U#
zhr#18mFM=Ky?^}f?@qh$C`ZZ(BHehDE9Ej|kt{Ni^984i<8b*iTO8?<qFqP^5YY#~
ze*oSjK1w2L*N6Bv`QY`Yho;Y`Y8jy!GA(aw`-eS0d8H(x6QDZ*deU1&U7+h?xf^ud
zEcbw}hvi<-^|IUtx;~csLD$dn0O$r-9)zGl2&MB6LC_G(!w@vg@(4tXKs41Kg@{p>
z#~@;ia@oF1X~jMlE;|y|6mKTpzzJ(IFl3x(lzrn!*z^XFD@gg$^%FMV7q8%Auod#Q
z!b4<`iZyF%7X8#R*F{7*e2xeCjDwZWWyr!?WU9(ve$PL@U;ZuL9G#C&t{5U}1a%`V
zGpk)RX9TxpHmp|SSHjeY?U1*f)(CRF!13X@ThMBz3Qz@->w$=dKr;kd;xTgR;8o7c
z2JIvS2!W(Nibxrt3?v7zh=xHkOov$ad)n#eYT9OF^Rhy-Qs~07L$hUirn5J+WU*%!
zcH4^({vycfORhE8YR8?18qfop9tb3Mif8~d1C%$tol{-V{K)|JcDbQ^E}gW(P{B}@
zp-QGYC|Pn$@#C-9gtF;n@=3AvyNtl(u^s;Sb0b7H!rf%77Z$kGebexJ60S@<E|28)
zA|hA86%l8P$W3rV#ET*l1OYK_6O@R$O(m!h-3L(+5n23>jmfXupSWG=2!woHA+f+=
zQG6rFk@-qmCM-RFrwm&Z?;g+8$`n%yQ!CTc({t$%Il3I#+1Ob=<#L%5Q61#f{k2wl
zw?H82OKio*qvJkFw`L49WAu33J@s(OCKNc1jo;wEfvI>lh_k`HWVAPL=<`1IojQg6
zT&yXUC)vMhA-~}ME&q1K2aw*x92;i6nW2QFgjr-He6O-t^C-1LVml-gKJcw3`pVZ~
zKjX+K3@?-qUJ(_A7csOMk~V|vG;%%0Y2dgUlTs_9B(EgKUtNF;qCbtvK0tkxi3QkS
zjvF*B&=<k|B2^>ZK}9ame`>xHeXZc3TldG}Gr<l)b_gZ|I@oLn`NHa8iydSa9JxGa
zU6bFUa!2V&&B-U-aOp|SWyq3id6>xFT{>wzumgGF3x!E(k}Hyk%t|vu3CaY9tR7ZR
zGJuF}@UVl-$1-=<cFg^167i@BlA2%<QRMNN5_1CL$g+9t!k8|WyCJ5V<sOLXVYwG#
zdRgv+m_C;KA*P?@0f-r3c@V+|A(GBN1Ytug4@1~6%Ojv4fqSU_DCkF79s~UtW%B63
zq2qY*(9qck(ner{=LBF+1d(Ay)C>4e<`c)Fqmx$*P2uz?^(mFJsLPc9h-=1y>-e+o
zKAr5;u6$)>-1<Tl{v4kkFQ5LTM&Sr_Jg;cR4uWQoR_xR5Cue^6+u(ehvCI6HF$c=?
zkoP=2P)5VnZ`(9yF>=I+(Ud?o6vv3s%uvGIgmfyiF0snDRuMmWQrkg3OFs?z$n`=@
z@EJVs9+<HQqG-2fn9)r6Gi~>sr&EKUcI2?W@E)pGq%UG<GtAix^N93f{nd%CFY&`v
zM3W#)f=FCV9+s{u=Yx{&(0-8i!*tSLL^jB=0aJau0d>>76;VInKRNv`eG$3hjIQYm
zJf9)mAnyUG_jnH+k5sw`qo5h3Rr9N!lK*rjJtX4ZE2giAAVyrT>?la~p4W$66y+Nw
zKNCb0>l-W6^_?BJn`6fI@p$+7copf3?cfaz8F?c^4IngtM09UXJhl3n7LO!6k1^l)
zm$%&m4fIlT>*%wKg-JsIwz3Z9)zPf};l}65Q+JMjjs@#H*2$IXCtFwL@sA&$gNL>H
zTba9ZJ4CjFiOl|6&#3P8l-f7`?<&_kx~kMs8P~2^vDO{`Dp1|}%jeZUru@Qj{{vG%
B=bHck

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/sequence.hpp.25AC5091073FF3EB.idx b/.cache/clangd/index/sequence.hpp.25AC5091073FF3EB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d23d93d8d3e7852371c7e46c2a2c23184ab5e132
GIT binary patch
literal 26038
zcmd5_iC;}y``>G=y?4$z4Jx5T6Dkc#N=OqjWgf2a8ZMXXUe|P8+-nw*A)-um$yla@
zt_D)15=A02;UZ(EWbU`tS?4+S`ThQZU!V6q&%2-X-OpNkjn8`4cX|7D3ky3qP7ped
z?KW%DxIadVf*_dF|4#p7+Ak|*K{($;5N5@v4IVl9reEF1T}nF|9o=GC&}-|;@|+vZ
z5*9Z1xb*vq_Qu$$)O+;uew~w3e|VQw)iuh(wcp{}Kbyr>&8vz0@Ab=0>JD??%xINm
zJm_v%?z)t+>qc)f9NisT{{HdFlP6zp`W)>NvhA@)<`)N_g@Kbxd&eDl*5ZWInH3M`
z%-Snx?u?2pA5+ggGpc0$?tGWU6FX?+YfWPhE?Y!?9GyM2w2{RR=E*z$Jm`F6)Y14F
zSG%1z!g4nq?DxUa+3H?eO6T)S<nw>m`y=iyX_WKEr33k}(`1Xst5v4U{#uP22CsA6
zyt&-lR{A;ox8#xK%VU-f{LOyun1l8X0|z*#h4*QhK4ewz%1$qqT&rI+=CH}RQLm$J
z^&eE#hx~Xb{mveFOy7`xr-~*|znNQY^r_{O8)mMtS@B-+F%^e*?7AM7+cofXxbx)<
zX=*`c&w%kmn<R9|{&)3_lD@6p-0U+aVOLwv!i6`sySH1>z9e_psm5jvwz|dpzn?$u
zO#YE?jbg?W1%wR3*ZZ%YHZFE<m*#Odr<PA0vnO$falnB5>n^i)BrV^5CuU%BkM0+}
zR`;zleM`3EnGVa-GVkwPxGccG?4hxrUz@?>x9v4PpLjI&H`Tv|JD+*jT})4`dG_;?
zn6*B4j;#-zIp6D3#&)+GH~w9A^yHyIms;=mX5}%ecl~u4l1cKg&C@I%)oYL%)S}<(
ze}CWF<L3QAf0XUIdvN}RrvqM|tA8;)!Lv)N2d-HodyQT{?0Tq!CS<~~D}6`(Ibd;F
z@#RBnJTCQ_+IiB0-^K;cxUnnI`RZt2zy2e-7vD2YySe#Pg6pNTWiLi7ZrAGM>ce$A
zl#Ckpw~xu@4u5Rfym<M%KbIZwS-yLrb@_qx<@+5Hc8u)tPTi%^AE_Q1+t^Br5%sOz
zN*g)OAG4+OepLQcTx&aGocD+q7t?IIzw_>SE~-nbmKQ&}Z|HSz<m|sk_IbVk!_a%d
zF1aRWqc+x?T-o5%!Lz5|jj$Nj@KygD^f^@BYfZ1en-+Zh@bcKqBZ5&)#M(t+s8@{F
zFlpcZI$k#=^Ob^+(~^CKe|CtHezvWsH#efL!|;pK5@IaE)0<?R30<|MqQSkr=P!<k
zYv!`qqhe;{_8D<gEt)%a_dQ`HI=k;o-JQK`Qm4|N{nnYCIJNnO^`x!u4i8$Ua(92?
zm3%9&d{NWGZ)VRrws`jIHM>97Y%JX0ZehQLc`dq4o%Sa-xjH<ia+qIGmGQc=2@mX=
zF4_L7=W}6&-Np^h|M`ub_UPUF30IsWO7faJTh`|M`1x$ed)JbVMPpwZN6kr^84@{U
z#!rE5UwdAi)4wfQzGG2?-$ri!_QUSGg~hl_>Y5a{pr^&#zx_Jm!T@<r+l^V9q}RWX
zHOX>L_!@Ph-Jvag`%SIPb%^$Q+vI4xb?;~WFP8?+EoolR_F4U-E%SbyAK=#A?osi)
zj`m$GB2RzJ>|L8B_qFo7l^bxSy#GQ|zv@wsn>^2Gv;MR5yvn@QKUO$}-%56<{dl|E
z!PCRr)fgRHB!98BHXFJ4!R*k4u_<{cibfj6C1vf3&5VALeXelJhsa^yqOYecmLDu%
zyshfm!v<%4ei&G9im<jIYx97|Sq(S%*IVo`V_JyY<~7%61^hVP)jFl^YVUW4r~cDn
zjbom%bNgN&{TJ=)lK*-{=qVG|hTZZHKRYwlc2fSpZll}Gd@%6L_L!eXP0uWO+vN}A
z@qbPmS<y7&t#?arH@oZeqs~=3WnFInrttSO`Q0|W!3%p_`qAk5)jks{-W*=ya`l%3
z7HF1z(si4utNl*?^v&jXhc#YZZ#5Wea<txM6o1&`$G`SDZ2BQ)@8BO!giKsK;7<J5
z_Jc0`>AJVeuW@caX6)_uu>15I4_@6KwRL#8&zS#0eq6t^^Xbb@b?$6jnQNndI$_|t
z`(ZaK%l9X{dG0$feA?3ViP^1v+sR{H0&}hmcMors^`ySH=k_hlQqRllA36>@e1Ff<
zzdr4EnE!jyNSmiS6Z1`~-;C?{?W@_-ldqN~n~scpzwAn{9Fq_K#3!Fr_3LZm<8<U1
z8GLc<Jo|E+fVm5@ToxRiQ{O*xgniRt^RIvCJ#*>sm$Ca-$L+5<i=U3@x6!?B=b=`Y
z7EWx`F#LpTU&|KN<8mt=8i`}G1^2Kemu7Z&vuMPOAqfdb)_HFUICr}JYlkzTgWjy@
z{&?@`Hwmx5CjS1gRpHCFc?+I5w4OZqazNLBNvGUw63pjRnU_YFlbZNCe@Sth?hN;S
z;%C?FSLe<~dtY?_I<9cSnpOw$9)t%vCj~Y<`f{$b)UwXan@`&PUhqr#uRUvmhBW@~
z{q)(B#%!^ti_?(Q0e=)7y{N&pXbTdO&<Rm}LTmt1s6!#_@84fLX_}%+D?D{QLp{Sy
zo?%yzfFgopr`?WQk}dG4KPe~zc}5^Vka|$`i}x=+b$+42kI)Kc>ljua!<HZcCe~fQ
zY-#2E+ZOm2W-AGKCZRSU^&q=Ck4>fUx>ooAN?^94#3oUq2S|Xaa*F;-&A;cU!7H)$
zBXdM`j%bmDxCEHmHL%N0p}e047o*k8+FGn$iz$ptfN9RD?gyMqE7IUh6w6@A#59>S
z1PL&c#sm7e&aU@b#;?(gR!OJ~8<pYuAoZvTi!V6Mnf+eI&(R`gtW?yLicLWR%=ExF
zo0{4cr!8<_RvVK%CsB*+R}Xl&J9Nm4GvzYAhGqvPp*^zE9=QcbK$%@Xy1d5t*V{6F
zgJw`WN3hWmOdFd^LDk#iHWCC{*{JT)pOQ{~y3s_&zcJu5Wb_O*0V$|L?0(hZwk)WL
z9`sX`1{a~fC@@o2XUg;|xrBV{o{bpxYIx^MtjgB0uRKmnkFzS{5@6=N-(G)a*Vs~n
zBk(c?E5_<#Yzk5jd-yJU_m_BQ4ZeXkGuS??-iPag1eirezsd8q&Mnm7Q%KusUt`nP
z*e(fi2{5a9x387Ye7j1<pHU36mP$0K#2zHT8stwcb)Pxnt_FYP10_#X=ZTGz5SIXJ
zcr~@omp#W?84DM+HJmFs<w|Zzh)Y3bDGHOc72L3TQ_6?Joy9VKrUjOWMkQjSB*Z0a
z3OedoOV~8<b1r?-b~US%rL258NS%XdrEm!u+pOpka(bqErv@M9pHV>61;jQ9aS5=-
zOPB3&-yWVX<7_mW8cUar(q&pcE(KLjH(j=N*9N9{`={MMR(R-qio1v$E+Ti3V9lZT
zXgO?DRjdUb!n$QV@{C8VLFz#Tnx}1!jHJWr7;pR|#PkTE1D;Dg?Bd;ywU!kVG&qls
z<yE5TDv^EwmjHA6Z*Zd#wplAQ_&RU=y;!}MH9nUBYuccz$JRc58(QE2tYqUo$NTVK
z05q?t+q3W94v%I07R~TaLebbL8aD?CC>NpB!{W_*8w)&wnVI1^!<YX8z*WSqgM0ni
zMT0-^ZgE^x9~WusaH)s=9^d1WxQfnZqnVv~p7ZEj#(n|dG4@OQ#(3^2W8t#4J3JAc
zo`|$vxD-@P>IlD2Vco%V(UsV>zlmpzg}Yi{t>jcIg(e{`1yw^6T?g}i^v}UYqs~k=
z7Rt21NXaRZ^+YbU7I{x>)%Cw9*?Nedu~4D~rbtdH49KORYHXyl*eb2rr)GXWiFNS>
zw2k$|2<#bw*#xhZ@F~7ApDik{4UJ<<igrMi6LmSEvkaF2^KUjr7;QFkhOux-n{Afl
zl*O{;QZZ8eLpOvT`Q%^*ACfFNB{LwGit@F$A6WRp_uE7Z+>cfA8sxbK(dOb3vJJ?I
zj<>m4^BWsZb0}(?tllQm_Tmyy9qPvAzU}(7$^s8!raV18sUh|YfFPgAUMDZLejwu)
z+R~>ZqjY3N$B|YFsxHmabu($uoYW(+2h0Ajz+;)IF`i?nDfSDdLJp{xUF`Gzkg;$=
zYpO_gDw1i7aw({6%yjkN(R$$%Bky5@jfFc};1bzs2?KJeX!?#W%YF@#Hoey1C>+lk
zDnm5QU_*&Z$hhNpal&j@yiJ2AvXVKgooPL=UjTGF=)L7kc)(H(UWL~(JJm#8P3Y{-
zCBS+de(ZU0Y~TCF!c}d#Qe>wTIfQl~trS%CHM$SzkyAXlbbkDPR`*Moy-cLeWK$WJ
zU@!ce<GA~SQr>CsSw0g)i>A?{6CImc39#P%|2-T2WbQBxe#rY-5m6TrI=^$Nhpmf^
zd~(O>r3N46urow`hR~UoOMvzHVOM@(+Ld!M{z741G&7=!Ce1(stlz($aZ=6tLK#0r
z^XM0pAWaFPHOnQS2JGx~4~s9RaAS0YY$KX&%ovvd8+>p<(XVyy|6+x=^AC*_9U?_f
zI!3ip5Bt2)`%%{7*2Y4qwtlNbrz$o#aH$x{!vnThZkH;5hmlesITf&x!lhzJ4bOTT
z`oLk}TGkMM)2}!qtIx1Oz$N56{P~a%{U>{dFjO>at}R%-g)PTi0&GNGr@K><chi+I
z0&88Ed!qUtn*zB6*vO8yj?Mi_7ijPiK5-ww>I2O0<5CY>Vp~~1#<hf5)AqTKME#La
zYg__s)Tpq?Z-ZVQsEe;4)_KF6!&>rRK#UHtaL>9XL@=Ktj+K5RG2O^~4lV)q)24in
z%-nG=W!!%&LoLFZMc5c5pvGD`yg%9}d#DBO!*cWY^rt<9{Q}@OFSBkZJNB^G;Qx3p
zJ4e*#*z(Jz9(HPv>z0J-VKC_>NKOf?g}7AAWs^1*4qtnAccHORscrN7veSK;`g~jp
zDl7E_z8*}TUNofkrOiYcKhXluA)|AM)+(2Rs-weR-9$9StZtLm{i1toa3Sw)>qXP`
zqJ0wL5{fu|i(k;)$w5P`@Nlhro`xHzVaFuIC7}Ljlw<j^_azq%KF*!sB-%!7RB@??
zeRAF~e8BbHtTnYRbu3oLvM$8sci8^U=%?zd47Px^!UJUb08!7IOMuP(t4VjOpYpV>
z&?*MY!Rj1ljZ1+2{mVJu5APN<HWprKtE5bFDwF7>Yv*L=E2wN`-E1@$ZR+1{qGPcJ
zFTpdZv1+8QMrI(v*j(#(Wg*RqKFRo<*4SHO^p@01LR<=}Zq7P~Y2JiZo*8q<c^Oxu
z86A?)O=5JDSc3#}^L`mwmW?WDz3tHUTNkwpZ7lYyM@$ZPiaPVQkFl^AGuOLX5~?MN
z;Zjgp5MBA_O?xMq`sUI-Sqft1|A1^hAUelz3Fa2`pJ%q!`Ny@|h02GztH>yiEmT|r
zYT>4Xu3u&x>?Y%7+DsQ<qXn3@CYORL)IJh(cXZLBamBN~{X`dyOzu@j64OXFthfY&
zi{92t`O&0Rss>l^@o=7~&$D{t5@3s4h_k1y>SCqApU?)@qW6gU9)oeIhqe55o?Yd?
z9vb{o+X!D}`77IHa4D#|HrF-6;){FjK0Q6XPJ<U=t?!Xb)VXX;;8Jhz>Eu1*f+m$|
zXK_!O8`9ikvpAQ4ia37H^VaAM12y<0|J3)`^gXsoLR<oD>CO=s;{WM;Ok2Z&Nhli`
zWwRRQ5>Syh=AMu3GIky`>vPHJxkOh6E(Mi2)-~(0sV6$z`|a{q3p|odH3v}B1BmXl
zxddy=77V$N@waWR)_bKM_f~AQm3gmR>QQTMIyLQk@S7Dr&l}++X?&8<vCJjFmbdY>
zSsuPLhBbnAdz>w*vqjnnT<T#v23`FxMbI$!Zx$_C74ocN+j}kn6}3k4ypv#BpLu_q
zSc|?S>X*#>;}T#io^~pq@bk1U;B*#CPQ`2@<WjLzt^DQN_@Z{%6O4tM+NQrO3zua&
zGjJ)W8W;=v(PZWsuRJ&JMw^S<=mHagm$RI&VD%L?J8%g(ulm#X@h=&5J3>P|lAIo~
zhTu|A)kV67SoNPt<vgFxKgjqStrqI9?-Dh;L>G`?FM46YkjoV_sx>&3S8^dy7qWrR
zCBW7+**`R@_|J(Nyi_~eyhrNyY_{PNP-`z$zFL^|_d!@6YDA|Rw$yW}Sn6Z??w+-w
z-L@RMY_YNb0%=|#x@vI=)?%Wiom1OZJz@>39Sv8B`YIb;Tmo#JSLKZKFIjZ$ioh#b
z#pPplKC3t`0T#QJ+!?fBeY%VfYFnaEG%93HC6|gFYn-v={GmZ%X4=k42PXYXLMyS+
zO59vf^>V+ZTaV&KkB^&rb3PqjpU_5@Z6#4xveI!0*={UKG9MF_l*c-h7WSB^AF~d{
zCBQb>9q!b9$KijO&-)LvwqG>e&x*&T9#-(E<YRh7g%$oo8xwee9bRA>Fyc~=`eI&L
z@Fl>Bt!~<QMHVs5VyhdM0Nc{L^|-J-TSK{<NArAwHBT^|g1H1(e1hBkIj3f{0sk#c
za!O-;ol8Yu-)c6{Dd%z73=2Gl4gGM>aOy>~U$C?_tm5taql<oJEu~$Y)3G|8wG@|n
z*i_r!jHhn7XN8Bcy?ZHgC`I%exdc#Ro62Ob=0%yjdsEmFQL}_~Z!Yz))DcH^pV;w1
z#_zSQz8f3u#&mY$Qc#68U7%YTcAZP?nfK(@EA0Xpkc1u*qlau9a0&KyyF7k1!ExkA
z4L-xYkd>n8N|DacT<T#X7v%PG455z03ciI&7fsU{j7vRik4@F9y-m%vU9DXb+JucZ
zv3Zb7KqWbpyp?)>RB7;XzPcI73yf&jSfG_`b)#LBO(Ds%4=jwlyka;D6l&v%n?%!1
zB3)>>1bcfc*Lk(d+@ETJe`4)4*mE##C-$ocR6Uu!;?6&htnk0uRlY^z7IfBSzkt{`
z;n(E5Evq9j-bt%f`$s)fS-nxNE2#MM{#jX4pGn)pWIT@f0nLn>(dmNy3aU<yx&j~g
ziqdLEuGnFP2QOru{uXk$h3HO+OUNw6{o&5+UQyITujQ-kU826rh7gx}*zIB0-0!4P
zCxFEwgQ*o!bBo0yxdhmeMH-Lv5yjc8lC-<a!=mY7R!Lj}?C9p8(L-yl4z<G5wIi(x
zJ5*sBk>nCkY0v&jYY_EjFT~4gMW<Re;c=;$@cxbasp#+HuLKRw;fwJ*Z2At<_#2mC
z@83_a*H>6Hr6Z&WvF&}ltd3`%0+)bFe<#)N_4jhxfZ9NH6!MH>D;bvn%6R7ScTd}(
z4Q#NnIrcMBe`cFiE&-L<>~(|g@0wg;Q-$_BmlD&ZY^vZ=53Bi#CSPjW)&l>?y3BOX
z=`=s~3xKSlyV`Vpa8$c`MX;v4Bbwe}F={RWc6{3IC6zZfd}ZT8+e5dBrrX%K;1Xab
zo~q6F-D{m|g%kNGTP8XzWA)0V9`?n(<Iu^yPqHakyI5BdbrqX}xdd4Dt#P~Pu?6+J
zKO*g(>=RLcBDC(g)WaSW^jH!<af7i?rJX9SOHS9N)=9|G$+5Mds;APet0x`I=bZb)
zbu!&xXrs4Dv=!L~gG(@WvbwL2JHp`_oW<9!i$r~q^;s?fmOBQ0e(_ULXJcWXHqVu^
z(@L361Y9a&BDrsVxc9(&e^(8@&+Qcxbus$}F2UaE{nq97bvo74;Dy|QJVn%}*glR+
zfEA2t6kq!LgO=JrdRyvEW1~D4Nas?II+MQNqoe<D8Gq%D)ge)Hhy^mZ)Wa^7o;&yS
zT@I`d8ziR<Y<=KTu_-*ap@*u(=`?Mw2+V@*$7S_#wr%GU%$-Z^7ZSPZ)p8Ba=RNW>
zHvP<;c`o&^(iOjd`m(3J1{ZMHM{N2L(-=RO04qLjKhrv|8};WRFk4J6Bh$-lXT~L<
zE(A26)33K}91P<flG6_M30x{Z;X?Vow<$fJ*J0*Z{Nxc*KVpN5OE6bja^HX9<|fCr
zXD1YO4jY|gXD3_=s?JWj`R{VGQS%%pn*C*kr)jqjw~5Vd=HPM(<}P<}nV;)xOF#7x
zZ~J_r&Sw#5E&+D6iDf&R)S!E;Y>U|tEEi47*%0JX56krvhAciwd**Q-SlB>JH?R@U
zCBUvZT6gU1a<@YJ4c<xUIySn_zJW_XT_1dU#ru(sc3I$`*}UNH>E4R}0^nBQY0EAT
z<~?PF(gr9Zu{x41`CJ0*_MntKH*Qb*i@87Ac_C3WO=Rv5mjJu7%k=WH|JL`>_S^Q<
z!^cM1tlx48sJr(k7A*a=sxQlP4XcPmtWIQkatW}CE_WL&?lz>g2H)e=T0+z%gzk^H
z)Wg=*t#%&w(=lc(idnmf)i;?nE&=wycKlz>FPwU5EIihZ<X5uOD;DwKl0`hm^N7cT
zud9~dTC*<3Sjf-<D<r21i2}J4RE;FvMDyR~!8Xyc$30ob+STzkGQG|EE|-w;L(lOI
z&(B&!t1bf1rNv8@)yZsBaj8fB<&oUIx{;SIxEzN(<5+N+O8`CI-D}wt96dr8=AY#`
zt4$JOzW{jRpK#`v`eVoG;z(ZJUUW;&e)WK)s7lj*|8!!7iDZ?PkJR~WfN%-0XTizI
z3rEbRGYbtv(!kdZq`AR1uw3d<Y5gv=ZIw=!xq+;61$ze5e#m|S@M2<*Df?e7^JS+O
z+KK%=QNL%wHZB2H)es-r?l6rmI=OtRc!N#fV45A5df2MtCci$drAI3C43g%zSvK0t
z&LFuIRN*ajqqZt@$tjJw&rGeSNP!jDr~*?@kxN0<-9-m{y`@KEH6C1Sh0h`Fh%6JE
zl!?^A<`Ob~<5_Jypk~EjD}0(C+^)lo*J0-*#3jJq{n@}PtKH=c4Ss;ME6+i!K8WdR
z$0fk3o8A1^ed+W{8P{n0*EeGHjSX)u1yv7MUFoU^9d-N^a*D=sBKWcLX;FQe1wOe1
zd++Dw&#V7Am!4nn^Ob1HDVoiVTq<0g52@GgHt#lklCe;(%{WGOijjkp5SN0=+EmxE
zKUX^6_Rx6HL#H%8@bZW{kI<=`OUU@^lyxE9*3S=M2X9)}<Qq2q#v-m<>S4EkS?$-z
zCPCX8v`arHM$ehw&LyC}T@0zS_(I7?EBpgn1}-6oONee3xCBsb%b)GMDxPYXccR^d
z?GsJ+G4F&+J?!~|4nC@{k=jGZAX;Ct(P?%F$t9qKY2$vL{-Fb{Y#m1dtAUmPO+<IR
zIihEd*d9*!fu~?Zp`4=tdo&7tiH%?42Agy{K&q(19#z;M9962Q#tzlk9b8SSsKE|3
z*c}2vRPhNre8TSMbZaM7e8CQ1s6w~q0s{ik{5kyrym+FcJrJ7084qw6PmN??w+!4G
zc6U^fiQO{4D}W2Zta2g*SQT{y2xgUYA7EQYjjRzJ)`(tknnD$^qC>3c1#w`gGxjbg
zLMW>BwCH+T^oAZz-^~+U^F(hr1qZ4GdzZ5#;5?N=*I~bPI1mOU%NGa5;x2zcELg{%
zU|++j6L4GtlnIu?=@N(jaXUFKbM3z-UduL5jP<_hO0BsEx_5$QlRo-;`?rDcJber>
z3$Ri=2Db|Kf2NxTX=<oNXhX$ge~13gu<53$6D=k-go9)HIK^U0Lm*zz$9ewq0$p|I
zxb!g!wTVI<U?)KpYmna>6aa^iR1t$5Vvw^!u@O0JM9vR&D+xu!Bj0$`9!}A<ir{z@
zrcmre4m**vLU9YVxP`o6-$!40jYD7Ku0VF6iZ?ju4ektcHdU+@JFgXcDEdL17#b&b
zg?>O0mqq8xq6f^jT1C6dVz5H7K?>X;bt(W)m7Q1wCP<xrgYyjL>ILqQI_Vujc0L=J
zEOpYK-m(Mbz!a&I{<xV1Ks?eUAANX)o&9=bNk00MVMd$q$d!EbqzZPF<xwE{=nuMB
zz|f;u^3jJA*+F=pT4|8}n4BGt_E{ni(jT5u&q)r8l*1K@WwOOGxuHU_QtrA^?xo*b
zB%#%E*lIakp;#~5u9qG4Gd7DyI$oAtAX;gnqxFciCv!#*@I7edUKOpbik{$YQiX1R
zO=mjxM-}%)^ZQ~8I9M}DOFHQyO*zq)rq~H~K*?f2N5N6~=q|b&b0b<ao}#Dn(M$AF
zKKhIP%Ew?aSos(xhAAJz#c<{0V3D3e>;7ml#)xB-kK@Jh%Ew6}UDCdTr;F2-k26F{
z0r;LeOPr;AoF~$V`S;WaBt{^mA_|F728uOETw|b!Lt>nPA|8qH28tvkCK)IWAn|~K
zq6*ciLUonpz;z<9Qt?C-o`}k(s1k)L14WG})EFpgMWNO}5h)3glF~-BBt#o1Hb}w-
z14V))Bp4`mNWu;SMY1F$8z@pFA;mzECJAW<iY!UUGEn47Lau?LKoSZJ6vdKIY@jHU
zgfaugbxF8xps0|93IoL>NqA(ScrFRg4HWMr;hlk^RuXCr6iZ}biL9(b%DZ1_poo!$
z7z0I$ETkAHie#b4Kv5<OWd@4-vT)x(@k$n6eOJJm!Im+`(Sya>RL9X{dd|;-C?L23
zMye!K->7~ASoY}98ryKJ3s6Ytdz`7KNYB9L8Q8(VbuiB~Q0Vq{-v^N{NTE{z(@&z-
zIn7UDeJ{4%i+zA|MT@f!+wQ|Y<~jx*Ra6tlYSI>%B2-Z*dKHRY;IN>sjvvT;H_izJ
z1{#H)ko```p%B#ok|2X}ULa8ODD*06b(I9bfd&u-nft*xf)FvF&<l9r1v~^GTE*}S
zcoZn0>?|8j98A=OxGD4?_Be=rV50zwDuZ&;pu(G6M7$S~Q1G##s0_*pgs{n>Rxc8p
zi=-t)w7`dA-GGw^!S8}~4q+e+LQ}x%L1+iDHnMI6dTM;6`^aDjNK=cPKM3hS9%OlO
zz95`d0$-37m{SA|3hZ@}ghGL##n@uTd4kYyY4#^%^AoZ!v>{dKocQnEtwgphk%NKd
z1iV6KjFSq1QbXS@MeRya=dHTq3#zz;+FnAPVDCW{S5V6<s2v2Qz%OE-$H|8<)*N)C
zL)Jq$&k$OO8p$HgS;T#l?vRHnjuYqO#2q@IgN}*Fs)myf6;AbLvCU?&GaMAa$Y4-T
zKvZ<51K8sL_5t1tIF_uejE)E;1;?4`ISCOO56Fisg)tA|4d5KIZ{svW=s@&SPKZt?
zL}%DCQAM`slr1{L880}n%qHU>f=%Eb%Czfo&LNb7u8l7x*O!tToM2N$jpSM*xy{ot
z6{zBa<ncjjsZgZKO;TlhV1`k|dD-E->;%2Qv-j98t@h1a!B*ZVF(^t5ft?IEZtUYY
zrx6N5za>)EM9M7`j@oKjvs!Keei=|6+4$i!M|c;UPL{%XjrtTFsgb2{9wRV<fX67(
zj0=cQ0SSgNLlwoux0r-NP??sMlN%Ld=p$+Qkpzs?(V4)fW!Z3sBNPA}WR}7yjgZo=
z?x}>dO*6l-0;HfiDaaJK0aS4a)j4FKNJZYMsEq<~7}Yua9Rb@A*6^GW2~7`-NLJ&V
z1zDvF<j`z($b)yt!vR7SN94gr<l)e;z>{QF89fqg0zJ}{>NzD6K8l*iL^heo1;~I}
zMVn02S)tHTC%(@iI=0Gp#bwmwGIED`fZB+~HnG?R_Lx+Wj_uO1hhkyeK%6%a->bUl
zEk$f3&Krp@@WiMhi8v<_U!XQoMKW<tCcZ%Ur;7Vhz59}*J_5-YD2`ucuf;lk3M|yD
zIvB+hssp+`({oxUR7a?K$FEI7%#O2xFw}LZ8_eX;$Jra4-wAI3xsy#foX=UHJ7%Ee
z%M*R_#1L3vfZ$n&LK&|UibH|@8}_$V?0m<%j`e`T)70q$rw_^0omkPw8U1Jc0Fx1Y
zj6n@!kS%m5s@RAcZbY_<eeE{XY8&!{*%_PzR$|T^1yU7#_dRL-o&-Z668aVUY)&7A
z&(`-2n!@R#(8s8eOw=b64bVfF;)hH$Ora=8-O5o<gZ|Pz7WY#0mrUF(6ZbUeFMT(V
zfd>7h$41iIpuhA>B7+S2OL#KrW6)nb?n~YV{iWSkIS~2_?8#ZpaB{0+rP(63*&>F+
zN&{?H2IT}-#awfqv^Y<KfOFMb$9QE>PI6U1^GVlyG6YUUfc47in-g9E3aheAQ#iR5
zn1-~zuOO={sJWsx%aK*NfkGFE{yt!J)T-}_N66+8a)a|A2m!HzaFQ$(1W2-Mdg8oT
zMIA+n-ce#8tR1vuYsK)j;t;6U!8+zFn_D?s7P=Ee5LgPQ$Ldpb+*no;&WEklg)4y%
z%ThQ6cDIgx06TR?939zfWIq_i6v#=i;J<-0AW>i<YLke9VUSQLC&DTmh)nF4iGvhf
zGLMAhk?s)wr8drxz%!%^G&n3atg1L&Rx!{{h^{9@4~1iJRvdm-90iU#5NcV4FfuL7
z70}(;M|0*Z^ch$uSvTXXSzr;;f<<D_NZb+*uOJ|&rclP31>+PLP(Pp{B-6)GvruzI
zfUc`qS92&CP3_;LzrAACnrJrB9Na3JI>l^?IqbdYV+?8%gX|5asU{ndy<(b5MM0@3
z1eVOP<5O}5Rqy^!P2c^DY(ArAzyP3%YHU@F?F}NsRyEk(ATn(A3ELY)hONF(g)TBo
zZB!8R3StXn463+C?C%jb;6PHv8qsQvXb-)YDq=;eSkYb)sf(8#<7E$EIRNXJH326Y
z!>FO}o+lyaNq?9WAX3NLm$Qwb!vocrwJ@g_Lkf&|)}J_^7*eSDYSgG2IRlB7Dk8B>
zByJ8(4rOLV;{;>Hu#Fen#EacvY6HeGgL0NJ?1!n<W4OgJ+!CVaFxRrOa;7m%959o!
z6izM%$NG;?oLtOOIG=d2j{XUJVwS=w#Be?e6k?I?w-{|$@!iuyrw@R&k-o>t#ftP>
zlIE7wI9tbsm?O7zoPT-gaYoKND~6vH`@n#MnqyzhsmHJrr_j&H=QHw${s*I&wFBoJ
zEAl%o+8h_1VNamV#_O!u0kQ#FGW$kONrrcUip)|N0T~7*c=1fnNyqv(bc|zGWX>`+
zD01`9$Qz0bv}0z3J8xhFCJFX2oL&qmwCoSW)(^z4Fr8Ax0wb3NMgc(J0`f6?m-CGk
z8}dlv8A<#BLf<`3JdYE9*ww*6VDEAgvSO?h6W3zm1MWP1w}iNs5Fa?<h9SznhBK3)
zT45z;DV&WAUjtNRmLF#yEAq=DE%Hc^Vj?dhZbc**7J2BP+WfTKWEhGt;jkv)G-O2$
z)MD>i+yO!ou+wKoxI+y^gPj~p;rwGrfw_>SaK<t83Sb<w6wWf%zoTOrvk&FOV#SB%
zlD4@d5bA<f>?z`RinIrs4Qxx<=Ww>MVk?^?cE}OCz%kWw9sO9lmTRfUz<~wIF)JEp
z8GqFsX|Cu_B3?w-`rFcsUFy3uhAvMZTUfQQ-K(SN(MK;=FE2QpXOFGD+AAJ?Ons~r
zkB5-$A>;&gMd25b^+jY4p$)1iMb@Rr-XM@+eF@ns0vWfE^(|ztILE0%&8v_Hd?mF}
zh1yo3V1*(AH;=#`a0`zjB5>OX9IQ~R!&d9Ct>WY@iBdJOEikU=D{0s&4co$uOBFA$
z^$Tooa42eBh3yrGq9=*fNn*Q9*UI#jNYOe{v<JV8Dwc`X%S3zNoKeMDvF}-NFib&I
zQ6@JjlbgZm2vwBJV!5nQD6YwMuE}N!#Ur`SBiRfVC;G}exz0P;3}_;7;={TFCyhgY
z1RJb7aF)0}MaK|lDV!7zCyp!MWaL(Lw0rn};}Jj?XYX)|I9LOII7{KwaOk^031_{U
zv%nQ2qLg@*k|2n!0WDlRp0$*4gDJJ$XA}feDsaQu7jr&1lmrF>Gr}3*3M191b2SQs
z)dnJ{ET&LMd<sbiC@fW!5Z@9K1d|>$#(Cn3!ha$yKasY;2L$3cs{zgx2VUH|)tT+9
zyni(M-`iT~5^P<yHL|sZFQA$Ct>4#8;oFU<KSIA5X4R-An*&_WFsEiXu^dvUAs$52
zZ8X>+*;?FWE$#qE)>NTm2Y=s-=&ntCS1gq4EtGvA76nXo_MM!Z4qRs-r?V8!M~4*P
zqq7uFL5KbVV@=yyTKc(S3V)5ey~YC}@C+eJ2IbUqMHI7~v??b-5TybeER4_T=Za?g
zhU<UBUa&B*I?`TlRMe5~V&eB2I6~?lAzlzhg)>^#_c(1GDjb@O6_HWZVZET4j5HhR
z2s<Iz#ISxh)@y8Q2sY5SICov)@@3$T8Mq6)L0`zkp_w=g6fmo@966;O_No+Gj=aiI
z0PIY`<z&HTPHTrDN}-!X>rJ8`IH6R*2eG134vW@@ML*y>!^TXeozBw8(h6#TzPSxK
zZA0FQtsG~^D-g$U{bRT(xUuw=T;iBZybKmb$5X@$7Dnhstov{RJs1N5J^OslnTM~`
zep8!GVpsU4-8vFIgL2-yB2;u*^f@hd0)_|dWLbWEctC#Od$ANwn1>X|hAsS@Ag?IN
zHqmXH=mT_RTC_ybEm8CVo+9)MW|cALq1(WcY(n*nFAsWZz`e73cLndm$JWO|u_m{(
zX=kU{NB&a(mqv=M;pF<08!5JiQ>>>nQ9QnsoL)*U;1$uVUP>KaN}&oxjpS4#xqxR(
z5j9eW8Yxtv_#iodklYOvAs?hLP{4QmMLSPjqq=5Lht!mvpIspAH|SS0hQA`!xxRA~
zh~rUkcxZTc7;^NnuY2FtipTzD{TnDAXZg<x21+6Y|7G@91I6P4WVQe`fCB-V9)a3K
zpdg?yQN>bZwiGo`C{`k~m8gM&^Ady1Vo(Fc9y<>C#i2mO9(yA)+lU${6nl}`Ueo|6
zO*EHmR6iThe;SSdX%)`dsHH-YgUoY~HJoKp#BJ34Hu401CsjN^Ru52P*y&Nldt~t*
z*}&>b717u^8Y_*i#er*am||6k#jdf~8}^a(l}*@rlff%nu=y5j4PhXP*ovLE8Xyv}
zc_Ow3{wzgg<A&MTUimH9JO^9D&XXeYaKk(UL_Rjp$JVg6QbZZHE5q!XsrFB+@G8Up
z3dI#{eg#{@1vrXWOPa4Go(e<^X&z&s*hZRfGf*Uw=7|Q1W5oLyX$va=HJVPm(+w0i
ziOo&IuA*rFDB>n*eUk(z6cxm|g0xhGJSs>?1?i?x+#`PX=mtS|50u)tN4no5{S=Ca
z#O9$vE)R*@L*k>z<uS2&OdMe?rZy^xO(k)JWt%FV6PxD-HlCB_&kerzC9!!)9APU#
zU#TKCRmAax?uG<a>=IpfiQbBAcZ;sO4HQSkHb=!ESTE=+X=0l+1H}{3^NHvOhiw$`
zL=1W&c2Ov*M2{-b7Xp{89AaRV7^YCX6*X@~ONF9F^r#Vi!9Ap})QEvKVwghlS=4+M
zEftDd(W6%M)nC40_iX}e#W00pv7}inSt=Bfl1HTE3rB1;mq;lvQVLTjR!Eu^lBGfs
zEqO#szIy&E9fVR~v=pXLtdlhBBuj;2gXFP6@>TH6Hc6UIlBGhCAbBK6z7Wo$<=84|
zwn~-?#SY11hvchZqV1A2yCh46B3bfCmVDvljoR2JY4%B$3Pp<Kks|rRu^~mIN}5#3
zQlUtb{L`dhaEvJ8n4~!-St=A+l7E&I3@3jSaYE9Zpt-<5e8LX-{d1*YIQFNA(~{=2
zWT{XTNd5&<Fhr{<;;f`OD_JTO#gc!q6b#!Tint(YE=ZONMVaJRCh3<Lint<au1J=!
z=1|2WDd3UR3AQ>^Q7LIEB};|kxfJkR>ZDi^pG!TTO9K^(my*Rx$;Lq8{!(hKP`r{f
zuOv%_qFS=9mK<P-q}f(WEvltf28#C8Qb&cNMzW}pYz!3cHBxJZ;)7KGgVe-8;q^iC
zS13M9n$MD@LJ=c}#mK!BWWM#XX1#2wQ0$Rg?~#=g+a5W1kK9!;861&=kH}r&{EB|d
zY1#L*%-GA?KdmC<wA>98u=8iW8NaOuoPG+ujauGD-mt#GY{`5jeq#@!Pbs<^dn|>M
z2;rasCL!jV@q2j+*DqakOBa2>JEZw#h;A9851cH*-k%kX-`0c81%<xBVQ+A6V3)%D
z!1e|F${x5b5GrRz_!T{ck#E@h8xDs32t>0Pl;6`+96%?^zKL=O96+c4TyegWy6B#k
z3aPTQgxaJ9b5uDhAKgvd8z@peO+1y4UM613M}HH4<zui(u<|j?Bux1jZW6A19BeXJ
z`8dX8jPh~3$#~`CB$G+X$LS{1m5(z_W+)$Lnaol?&NG>(e2hRQ5lE?sLMBlLiZ#e&
zje#N#nZy|=;*m+bfg%Z+BpE0UAd>?IiYjDYg&HYpHUgVRV5Q=TsCptQo1#ioRT(I1
zL{*J}qE=MZ8Ym(qRivb}5iO~r4HO$B)dmAaf}~0?Q0$OYI}8-bk}BCiks_&53>0aS
zD$PKVC8@Fu6uFWr*FaGqsR|4f#geMnKv5>C$_y0OCDnBUMTMlQFi<>_RF4c4&n4A!
z1I0T@_0B+1E2(M?6iZ~)5?NV=D`nM614WFiiZM{6$f^_rMUkv3GEkJssxkw`eOYzi
zK=Ddez51?ztpb~%`0ZNQ2tibf#Vh!YT1CuiFAm&`LxEgH2lqZ4x(|l|b&o2lN#ANR
z5SRs2Q7Dcq6n}#F65PsV(;vTxtC)S363?ZiJwWKYS;RAow1;2|9PY3(vm3fl;;aA%
z^Z%B;Uxn!w&cw*X9A-LdAQja~MHUbchO4~H0KdHp^E7OSS+V$qU4tF7?=~?EAk=EQ
z=$kHv!88rwRW?iVTfB;Q<HT-pVt+Vu1E-Z0fZybWfH*|%m=S(?7tWSwe%YdHwio~*
zc{sshltXrdSMj+sLT2;>J_t3$@A*PH1e;kF{7NqbPG}bUaGQNN0NU|n_vMZ+il>}n
z$L4*5`gT%mT}QMX5eP8?h?TNh<F|hmp{;V_R8G7g3PufGB~Dj~7sNS%8pf)E-wB3t
z!eI|9C%*%%uzHDjT_PPJV+hKy0`Pml3g{Wq;tUCZBY0YXbHwc&=>SKw5D{es;KW-f
z04&I?0Q^F*A~=;JTIGlyaP|T2FI#5$tza+$#7kB*ehnDdnzSS*<-sTA;Q*nE5_v$0
zJQ&zWa5T*d!0!et2FXI=y^sXL5)H@N49f2ZE1(OA_W}|EP>8*;m>a(&tO(CtLBUs0
zCxFlrm7~yd6sAx-LhT-*K%j+y&(1!a-zA0*2L$_Yet8&<$f(t7)U6uzgqSx~MB-MF
z*bmO8fh^4C9)6`5tWsz(wkgK0z!s*83)trZ4pk`l)g;9dsJqgu)4-i%Rx*AM8Qe8$
ztb#PIAkDy|ql$aP{T^uz+g;%0uy^_8WJOOaAT9;OTR}oBBrb)-8v^)np@hB5Z!E(t
z6q?O7vF<go9zdw#p&0N`42F|dU^%f~%C9j4%LA@4vlRZFP*>NMTx}@Nuqpq)XC@G(
zW85KjqZwvUnyqJuXJ;S{(#N6hL%m>_)5m|1%|ECK#8Ii@5Ndh|xxmCq6{)CxDzX6u
zG${LIejgf6mnrmtIO2gg8orP!78rG1U=$AVOo&9Xcljl0#eAA5`X-7U0Ycx+5PdVm
zj^H8RzH>D3zH5&A|32OT`T(mePP>Q7f{+kP;kT&aKoyQ#*wEq^rxi_JM!d>MFw8?0
zx|`I@2){fHtplfGEQMc~hKzxG#!~nVX-I+X1xw*KqG1ODH=<bzzXc6f7a)$uQux(p
zXj=#@vJ`gX8T1gMWz-*b(HZo#PXGT<f1c=WU9-aTi`KwjfGS}roYMx=46u|~N-(`s
z4Ic(nM%J(RHEP8Y`;oYQBp$%iqIK|zxPBrYuq?m$Jp4n?Elb-lwn(99S12|IhXyuW
zte5ax*NO>vG4WnZf?xuM%@dn|`PFL$T|ZGaPn283c8X@hFQO<Wvm>(Y5jhApvd|aV
zg<5_+8@dmDH=o$$6K|NasNy_naGtmV(}F5E<3wRy_m2VTG!&t`W&B<7o>;#pZcvMG
z1(AI>zx1t`)UJvhu8N&tQiEe}2IUvOfdNFbiN&pAaR3B&;erRF;cZ84wxbZ3bLhML
zHn_rR$-sUYI2ewf=qp7ew1|WQS%WIhk)U&=8!Y*-jI*A>FO5T;!D$Oi;g`e#3n4JJ
z(D94m@D4;F*gO0tIHUl5gncl-`3<c_i<D1V<&!XI8|WLXQ}Wy2ite+NI4>pc3V%3~
zI7bq9VA6o^$O^z)3km>q237!m5nQp<mm}wL<OwS~+ze+%_+@Z74h43lwoYVr5gfYf
zv~j<UX8->wFe(HX48rtjKTYq%y38*6Z;8PlovZR4>?&=xWr@2{GTMdqAc6iz|F>K5
KzmvcJ@BabY6%(ic

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/sequence_helper.hpp.2282D53015E164C0.idx b/.cache/clangd/index/sequence_helper.hpp.2282D53015E164C0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..08824b8fba7759a5bbd3c571167333dff85fcee7
GIT binary patch
literal 896
zcmWIYbaSg<W?*nm@vO*AElFfyU|<jg;^LB`%zHrkG9v>+#hkVMXY&pj2(<jwuvB0X
zI>pr|c1m!ixZmFBW0D=B$IJdYJ9d=kbNVV~Rlm;tK6|dD?rdK@?|qufF8?hJ^wHDa
zclIt{;@-9Eip#|R3++4Lz2?*=nFrb5*j`V(BK@n&^`mT&^W)MdKkw9g&)V!OIH71m
z;fFAPkK0n}6K9+`X?V5h+_zf|!hw&i0)$g_7u!uLf6w6*m3c1EJ!su$y;E}=GX!;b
zGFSVob6mA_&X$WWeFXC*!yHWv<QC@a`G5RNN%gk#HV2eWuo#N`Q+c!_X>QJZze&|+
z<R$jZS1Yc}O=<!9Qhf2VkCK<1&+sq^@i6epE6eMF2?kCM76xtxMqWnoL*K=jb>D0A
zFbDt@NGMB~g9!#sRz^;+0`=c(H+@vyYsteP2voqO%w+&37&zIPAttPo`*D(GU($RY
z27aIdL1jS`Fu}md#>53Sp(r)2_$V;6c?~*dWlc4_&BDXWWyoa?H<(w@LeLIw3a^B{
zge%-tyz-{<)-X@-Ze6iogSB8U$N&Z|4xsnIfQMI*K@f>2!61RelV^~J^90(GSM2|F
zOOBC;S6W)y2xbe+d`2#C;B1(uy+3u=#BDHXpfD(0V8SqWGjV|f0w&DEE2Sr84b$T_
zbASK2<~O`Ro8<N7O<*?U7oUp!|4u3kC@d@@tN`=$nu^`50lACZn0R=3xOs$OHt_H&
zaw`hM^uiPaodXJFn7upF9-MLEI`t3~D4blvNP)sBh{TiNltALibIQYcJ74Qo`^wl)
z2L+0)v?I(1=dZ2)yx#W>2gs=!!a6Y98H!SK5<!*}7bO>qgBZNh+dmf0nJT@8k&l57
HBme;bfT$U<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/softmax.hpp.2F97791F40A73FC9.idx b/.cache/clangd/index/softmax.hpp.2F97791F40A73FC9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ff5a874c69a9ff2f3c61aa48cb3159e162356895
GIT binary patch
literal 1412
zcmWIYbaSg=Wngel@vO*AElFfyU|<jg;^LB`%y>oyhI2rHiaAGvee(|)@VxbZnYz7<
zu}{$HR*PAor0`?8nO82X$P`l5Tl2pzb!iRnLSLWRA}2RK{$Xpcelb97j_|C6y&L#!
zCv0l_%yZpK>G@;--)YlB4*f1jI_6-oURLd6^Az^<sEY?HTLqr&+<sNdzBnzROQmk*
zDao{{M>cA?FJ93)W7pp__oO7I#)&^zCaqb&IDLtp$fDP4uIlxF&R36Pm!J2_!FT>j
zW`0$nUt6YsoAkqbO}(O-$D`9N0m+?hN8DQ5&U9wj?<??p>ejjR(h14row^gBG8%DY
zot<-cfp~($bw^gU+?P2Gt!{w}%va1b3JRAriZpp`SN=|%Ywq^LVe69K`jok~guRMp
z*cw&w;j3YRxRtuXuJ#KD>um*3f7AS9xp=MF|E0$SzuipuW~n}Bjq<bNZIg^tvr8Sk
zeltH?s~W~ye_)@qjaj#8`fYus6U~bXoTr-Rr3=?G$hDsO#!y_Dn-mKS@6`du5wq_l
zM6oe3vodmtvx}>O2?h?J91yJKYTKuO_~|zxCKfgZE_DHQ8888pexLVELqqasG&d6y
zP+Fc#UI<LUq#24*(~1q585o$VOus}0OT1;|;W8C5b%UA8^=rYsW^dDA79K9CASsY@
zz<`HKI!GFcr=+Z;2eXNXOIcYN&J$eUDOI}e)nO(cE^PyCW0<)-Tsnq2CNQ2*`r0y`
z?=ks6%OgZ0;FhbgZkr;LEgJ$9c42oBg;}mP@9Er=Mi)YX!d4<yaMc=DzAjp9u2ldO
zwiB^~3mZSEIC0Bw={lgWv4}BTIK)p#{l!EZR-mvZuck1}C1HG1?7!7po&!2p+(kSP
zW+e}oyr{eq%!jEWTPx&Be=TDJ`bkP#8KxJ=v(<)M0Q8}h4w9U!lq)jN))h&mo0J<|
zu4!s6=Nn<GgFuJbh}ggtO$=L;BF)@&5-991;tv;wB@RX|aH8lv^n1aYht>)(Cjo^)
zDFUWy>oKEUVXG=DfQCegM8OT&)#9|$yXaOgP&h;+1TK8&aqrQ`+Wjp+VJ&toc<`T`
zlD^~$=lpj-VP6qnxazClo2*Z?IX(so8;Tghg|F$gFVUa8NDL_KE#eIqez@K~T)sHH
z2q+vT5(XE3q8{d)y06y+DC{8O02h90!1{%qU%VSAtSPAp4=PxEFmZv?G|ZPkMJk#)
zups!gZTdky>(iTodO>*y7NdU!!ZzK_Kk^Kiu(d<9!(pm<xLgffky5E^h%1tuo1q&#
z%>r{yP9n&(;-ci@U=V{FL_h!!jKRRbj$m>km|VYwjAkF*<~xIphnJm`gPVaDAqUOa
GAOQe-nfhG-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/softmax_blockwise.cpp.3EE65033A8858975.idx b/.cache/clangd/index/softmax_blockwise.cpp.3EE65033A8858975.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4653326c5413bc1dfa79c197f16103e96ffdc114
GIT binary patch
literal 6792
zcmZu#3pkZo_dn}62(RPd;B?$?xksT~4{?r5=%T6Kl2D{XO(>G4!la9CZjI*nOqZEb
zQqn~;6=~EYG3Am>DP5$SOQmK$Q~vv$@7wk}|L5~O?e_2d_FjAKwbovH?KeC)AYikQ
z#qy60PEU$Ujp49ZEKT$;B{g{kjld_0#Y&f6p0{B2(<!{(vX`w{^Q1>!u44JP<^B?o
zBDin=F6j4XL1s<<&wCEnUhxk;uWjWRHbs%z>FobaO7hm6aKCc9khe8Q9_-Bcs_6CV
zK<V!%Zq;@UKm7Aa?}5nBt?|lvryTokzKOmTyDmAi&dt+1`clq!-QLTtKWravXz^Jy
z^!|qH_uda?cBW4;RNr^C^V7?`+pDhcDHzF(m~utZ9FaA!e}8H8!nS7L*~$~w&$@MJ
zjon~gMCH_J-}<?=*3`WacCs%iMm81PJu_=wpGA-1i$|NnPh5QQO?6<6kLic)D|e07
z{IPjg4cl&RuxnX1x8y_jUw=ntJ^FRweOv9<eUsMD{PF$P7ygaMgB1F-?seR84{r?E
zF;b`WJn@EZ)RoMU>D4iB4=Qkze3SKy`XtSV8sl>f^G#OJ<<<$d2cHFha?%oNWCij=
zcn3BZ4K3>Xc47=~`P0%hYm}8&o}FZ?kZvX}6Q_z2C4Gh&OWRDt{9P_LF1j^Tw67?v
zFd}~H^&fpbA6ypOskME*=dVuZQ5fjU-_Y=*mbFe$a@3E~-Q6aK{p#vk!?zT>+*r8g
z<ly5i29bFaR|)(6()GT!H#WP^`TWXc!Lj2N*EX-av^C^nRIYDG<%i-P)6G$4`T5HW
z{ha@v$1Awdcl3x-YxYmi^SHUCE^W<Azr4I}vuk(niLT_?11D-O7q!Leb)2-1T(i15
zt@M7xg2$G>ueS{ubWvP<b>3dK@W8E-i2IUE{o8t;X5R%YZAfj>8B~nCo*9w6H={PW
zX|L*F8r$x4jx=F*hTiGr8|_x_x*_>+{N1H;XYrZ7(2<8PLe8I6p7^Rl#q^y+<|}(M
zzFMs`?9`BtRrx@1y6Dklk9#xhXPrIwee*>+I`Ythh2wPs19Lq5zNgJaSuxFv-dPtl
z?%Bc5yz(XfsA4xaLg!lR)jO>3ZI-PnbpEJ(aOJxzGn7{D)}6K{vp&o2aL=e=pR4%J
zrUkD*PFECc`^i+X!?&l@>BFb1Hy<DS&Mu<uxMo(~w>3}p`Sj*RcE3{$J)L<?x8(BH
zYo{-rjD7r@j(ySQa`k;nyKc7i|1glho>jtGWaHmI=Yh(y^mwC{1L1QElvn9J%Wd3o
zBh2~Z-1OkmGVPaw4xes|JL_!251o4<8CtPrgVi@aHmX&1kAC=K>AT*{t!0ry<o=1r
zC+M8nmA>zA+o?30TbD+sooiWfyJ>OW2(?<6=v7v3;PR~b?B_Gm)1Ql@j<`#Hocq8l
zMKf`7k9nW+%<UC%x87}baK8}pv2p1Zqk)EvSw1)OPPQIWs*3(|$6vAD8{36W&8tNV
z=D+oeUH@y0t@FUQKcAO$NGv%sB9AJFJ-n{}^m2)(;=0qdhEb1C`Igk#Y_=6`9()sA
ze^zzKd{&?S45#-n&VJRKlS^_3Gb0o=LD}_c)wIIP5FwG?yAMr~6{Qwk_jjz_v&)y7
zU!2OB<97ER$9MtTZ}}e?JTD)?;KAx856vxVtw%!hH-XgT<g=9ZNlPZPSuEIRWf67o
zl0=O?lf|Nqm5llF24x`mpOJgeSIw^<eZyrhpU0v-1RkdH2Gf$rlH(KJ{Se4yCn9?4
zxT$=3gXs|Eb<6iE+jXe35|D659Y<4ngXzfnd3%I<3wo5<a}n*L<f1KaFdZ$``MbDf
z_JBHTHlhu=hC+FR>E(+y+&iqd>zNum4ABB5fu_8{bYg*Y!^MJ&Yt&g^A=-dzpe1iG
zy{*OOT#$9qRf@F|(K*1$fpP!+4-Y9ZGfr$S9}3{HSEJ4OsrYHj8%!VC@mUlhGR#n8
zha(AkN_v{|2GgfG!%0o9zYVCd=O9|BB-E5Qn7-qCTxr4|R>#!Xvk<MTq|1{xm~O4?
zTz9A?PsnA@M{9elcnjnWrn{eRZxpxXzvZ%5A=*>mX(VqjJ>;L_OPBJubJ<@b+DFBQ
zFK@=^h0C_5Eb2bYWv@WAXq-q}-eCGuaHPY|?S9%kb`qjzjhm${Z!n!4w=AVZi$o4v
z^&b&K(Vd{*2{zck3X{~z%bJ$tsZew=j4uWwOi=WG5bg(Kf~8<21#^N+!K4%{3BCZf
z7r>F=T5za^$pqJdQysVx`~}p$0FU5Nup9+jjFp;eU%#<Et;(k83NWt#7feueCD>Pj
zH^xfcJNCwZ&~oBX^kV*EAvO<1$9To~5fsahWpMGLcroUb*QTULi$~@tP;{zVswV!K
zqSGwX><CJqn(jS@+#a{~7Js<U=FnZh?IOF7Vf!wah?P^17fxMq!k`W~bT&IQn>_;)
zylJ*iy`61xfuaY&Y7pGVh#9q}NpG^v(Y}v@`BCP+%fY^!?3;JdD9O1fs6|OOP~#|=
zVF?sn4uW#9Be()IDwupF2r8L;6=+m3`DzeU6P|C!+msi-u?1<=4!m~IC4xJE*Fmu0
z_N0uP$ELNRJ$(e;M-bwH2Eq55`d+9eAPHYU^9wVmM!{~BY#!$`WmS0|!x;$qk{2_6
z%=mOZ80UjEmQK+HU|B$HPSJ&6T1X;ErnW&lGg8~ZrJWekn7hS&`?#ha<o_t+D0{L@
zoN*ijt<_s=fjQHklb65Q?6p;aLq`dt4Drt#I!+kJKx?(u>R`^IKj^QPReV!qoD8iD
zJ^Zta_php*>lgCT;vPH?U2FifW0x2YJcyz%0Pg}Zw9DYEh7Qw5V!1~|?J$j)q61ty
zU@FEQMnjD&m12g}IkZ?y%*QDcAyctAK@u$qgA3FOWT5F<(-|mCD~y4pz?Cv%zZAHo
zOxy()T`&PF!=YaS^%A%QXS2E4?6C|WJ0cUZ*<*f{VV`XF*yeoRw)a;a``0?O>v~YD
z2W>2fq8otM00NBt1KxKP)(qW2-pL2Od>D@j4vG{(0SP}2iWEU1=wn`7q4tD7uxUP8
zVHk9ViA0JX0kaVjLil1;P?cY(Bu;5RdIs@Yf_;J;Mgf<;?N5BOtPII61C=t+z>8A!
zXHfY}Fus+MY}u8J^EyS-&a@AfM$rkz31i7AY-;{$U5C8AD7J)dLVtX>g<b!BcDuvi
zZzxOE0=E|Ucu5Xj2i!W)!FY~}<Z@CRZ#U8{L6BfX7U>6-exlhNotbXcqdOm=Wg5Y{
z5!~?nuV!dj_3iUgM5)L_#Y2Mxv8Rfs2EGg_Iu|r@$)Xg!4T)yryP$HHtVYqTpwdb(
zHZr=<WVHyf%MRe|08Ok5E=$lbSxrKQS=v~-;&KWXC&<RK%7hFuu`zMMgJio5Z-huJ
zc|~%{x|x3-^+i^y1x_ufV}e5F<kS(2Ya673tgs={Eo>|tu=F+Qq85Y3yPr86+QZDl
z9NUaTdzyKgW1Df1&Gd6gs^cJ=>2D)(m|Re^^ZCBpvB)FFddAjRcFOU)m-4qoZ$z4Q
zfngVz5>30nvkQDMPMfqXRJA^*2F-r})(^lF6IoV*a=UZ=Nr>nLK`*%svQCC3%`Lri
z1Bvbd{~id(Gbs8Ah@Ze5jB^Ji2JWatLIZULx<<I5qG&yVo)Iq2@L*8=f5qL7ppnt%
znp?Ed`nN&<Hc28orz_;V-81Tj#NGw|U2=A1;lghvIPXaD!_}ED|3JZi7Yy$*ZP*Hi
ztppc+ijK_EFla{-hrwZ(Tt>1(?=b=aSV^2OQP9Y$O&pIDoybpQoSCn6zBa}u=m!xL
ztdG5Q3vHtrc+Dhb9ZU@2OrO2=i7HCP)?!;MnnMR@1uzn9uvRbwg=&Q|P?&p|Cz&S&
z3n{S@hc1N)rDTt?vc#_o#CR(l`V~;Gn5j62&CMaH82d+6^&dquMVs1~GR{W<s1=Y~
zNp|n56*6%-h{_=tFDfhCM3qcj4Weo$t^>n5qO|P%Ng81mnLi4WQGyQ*=-51aW&UqP
zimm~E4N3DdY*hpHWMVz&)q?@S4Pe;7EO<VG@kg-0{Nc6r$~sY3?jf%=fqoNMVdBU(
zy5#xVD+VZ3s=&O8S({aZeKjd5j(pJFs}nNv6S~#o-Q!8wNT=zijRglqr#qxi!l81?
zrgv_{&}@4Yui0#2HhWB48J?8QcE(EJ?5?Q#j~5F?o7tE#d}D57&OlGW>nQ|c5fuFl
zyq*z!`jIHDL@;wXQuhKlT>yVFuoj$a$q|rcx+!(UWmSthCm+hw&{CynvAS3jha*Kx
z)FliQs2<2b)77UlP?&lc14%(eN)joB!dRu0iMv3hi;4R{y^rL63YkhJn?078&u(&j
zYNDU)jH>Sh--NNpKuthVdZq7+PT9}e3A)$=H4Uj#7D%m!k$2l+LOaQM6x{))9prRf
zRKG77_v(xtiVs?oHo)_*-8rRanC20P<~zDN29UJxG}*}?qu+F`DynRMU4jM`gK;r=
zh{@htCi_X%dUL$b=7Qs!enU#e2xIi|B8~d#PgK&mib@<bU$>Z)C>%6jcRx7<92C2L
zO%RG*$w3+3>jCjOhkgiN51F_dyt;|28?Vk69lJPpKH6l0M}i2i-)z2NTb5{!hRowm
zVi)4^0AYX;LBYac1_~91GEmsGFbNqc1$!wtVJ&4IpIFL#HFm-DE*N_@Q7CcoUy)r?
zbPiiLhduVvXvy6_vZ=7418MEZb>!n(3g12|s{eS#qg&I)*csoN6si{0@|gw6QE)v<
zE=$=%)U|?%tH8C2iEF{2mV}8cf4Mg@57QU0`$As8cY5nr|LoKD0#yJdU{t~^OAmn2
z0VYPtt%f|OI4Hl_*MI{y8vYuAGJ@<=1i6^vdGU;Fa}cP5<bI^+8$jJ);`czkC%D7t
z7sdN_CQZnae}dVc<TP~Y)_%$GG08y5q8<$EnZ-o|m^YAYDSHK4ek2i&iy9?W`PUGl
zW`<e@<JFx>WoqGQqv(y)Mg}SY0WuZdH$@)+0WuZFulcXnr(SmCB9;5Wsh_0d*9Day
zOYK_9kjnYM$p=+DL-us!6fkiia0&^2oAS%~6X{kx=(J{PXO5k8q?Y!^F@!(opt?Zz
zEr;wYeO$UW*~tjZ9f4@P^3Z=j@}NnwFFa&Y+PE~vq{F~JOakHkski&@Y^YB~8=T}c
z$rI-kyauW_WuJqHtlhxdO+5PV?eBK?PwbkBZkS=vMh_gU3kMZCCg_2K@#nCzzSNF|
z*O9q;!Mzu}Fd?guCJYmg$|@x15txGc&)e!E70y)2UTERQ;Zv|)=rL;$<BSn2IW93q
zrg}>9(v)ghXMwEw&!3$5Pd5tx#$qia-DuK{Bi%^ST}ZkSq`R1O7mall=%!(%xurLD
z7jt9dmM&SgAaY^EqNwPY#jK@cV_31IyM%O?CdDSkBRD-SRDq+Yq&!YVjZ)`naDz0(
zeiJlRgS7Ot`GWCo?w(%WA|Kz0I(8oBQ>=~u*RQ(z28Jf4W)_xKHn#STPLrHnry98k
V9VYwh=mrGGvn4a81wshx{{T8Sv{V29

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/space_filling_curve.cpp.EB322F6A66137D22.idx b/.cache/clangd/index/space_filling_curve.cpp.EB322F6A66137D22.idx
new file mode 100755
index 0000000000000000000000000000000000000000..84f616a5d324e8eb8cebc202fdd48b4e9ead6c64
GIT binary patch
literal 5936
zcmeI$dpOit9{}(<j9C5LB9us`F>~B!1~YO=#&x3bBcgQ+nWV@i%Tg+rL~iLKDu!j{
zQlg7XwRUX{+f!-ZWVdBsTeMpdVQaU|j^6iYPVe(RkH73+&GVS&e4I1C-}l^p=R4!+
z;^;_MBayhFE{RcLalvE~iKHa@i;dgOFO?*blx0bz#L{~2s9isrD?T{WUp-XE5jG|4
zTf3Uyz9@2uAhIA&!mZ4pQ{_udNE>Jz<Zf=z=#wg`+!y?6G*RJ9a=hbFr8l)r0mYT}
zeMJsauScE3Cf`kr4gEI#cIX%H7>4GV7jVF)p!#CR;`{x=@zX=k<2H_~4E-@3>-p%~
z_14!DYVv{auIh!K8U3Q`&GRH?Mg`qw%bo20o3HNNZ`;W3dLw6=iLEp;W@kKWpOzbm
zcdwESFTVUPt;&5vf4s@6zH9sx_t>7U!mbM`*PLv|atvdCX@0>=O`GYlu0MIWa7tg%
zZ@l<|fn`=(QO;<wgU)D7N3dE<h)+p&UCBi`7nSkCR3C$9z8aUEJGZLDeV6j2%0npR
zD8uvh(Fv!A^K%X#On$z9e5c^Su7URC^@lc(HCLA(TmPiu-T|Sd)tlUlnYXCv(#r8}
zRG-?5lCHw~^P7~03tPM^bl!FbO=>pT%3L2G*I%)A|5j7EuPOGw#!W7{q-&oUvHYSj
zo^|_W=<>6#HHGQVoGr#4<c>*1Mek9KT!9P7QvGmd>v*$gh!@|qZh1wa!|`s0;&xXV
zXDjcFC>Pl0xLNY1wWVE-O@qJIq<q0D>LJAx!E_W)@?`oe=ckT+wM~!8xMz9gK8w<$
zgEj-<QSC5W?xl=}hq$FwGrB@eRrZma8j&>}XWzvo@b2M1D0HZfr5?Y`b$Icp_l>lZ
z0&h$9PHuAK;@XgsJ>Fe*pSmv|Sl&vi__C~6pS^4_{FN~GmRsaKhQ<@gbJY5f<;Ra!
zu4*{Sa!#9h@sQt~&$jMXlqlJhl)O<ZlODdRwqDtOMA3Do?VL}4PbaVMSE_9#Y#%h~
zmdOpg-O%{5EZej$;&n9?T%Ah#<o+@C?c#t6<8?hH-+m+XaI_JaKe!>SzJ1s-t;Lin
zY}4i+C@mX4A%FSS&1!)_$(*Cdc3Ju-MR4<ZSdU4CaN_QM=d`K)2aRQaD)&Dv_dDe*
zx3ysBj|NUokB8#9*Yh?wI=y*xrQN*#LI0CnzX=t&6-nO@DN0OCP#zcD>HXp@>rUX6
z!Pt&(ws{?R{4A|kcFHrzjnDWC+u@!0bf$#e7EyQTS=^W74K?>X4H~~nO^CTXHfh%I
zD*b`4&cLE0rZ$fMN_l?j5B-tBBgQX<!S{Uw4aS$se7iOF;`RXW@AEO_y<x^qzA=BS
z@4kC?vmmF(<QYF~_r0o5_ts21{E)y)9S)CH&wu9|J|h`@_jq*dXHg-aiuP(q$F0Oi
zSAJn*B!3bK_m%Qx;R;d;EK$w-SJoZa;!H383L|Y6RV$H|WZu6K)%h+)7wEqmxML(g
zQS}O`6-Kzcl)Mp9P1+q69!pt5BFUL#m259&{3wIrTIyO_b05Q3s;@+(t*(v8f@NVL
zZt17)r!^NW?yv5TXp8z5L{)%Q1uTfR3hLk6ncN?Ulpy0COb=G?!?DLtTse_syaO;h
z0IPD__piGK14i_uUQ~{#P@5z$yat%n0P8WW{OM6JvJS&90kccMnrLVk`qolYBBfKE
zjN4FbsH5kq2mMU?+{pM^%34IW6kA00bbCE<O9#3GB1gI-B3FtlRUD%m#SPI0$_7N9
zG*2jQiPLaIUNkR6wSZC!s6>az_&Gp1hq?|>>QJ8t^z%SZ-2MWfUqIai=uN1v0Lm3W
zB?gF$Uj>w_s9OQ074<bhy9S`R{dGXQj`{|m-4L5Ge7Zp+sB4#K0@+-)J+UIlxBzek
zz@C_A^PdB2u)V5o0K?taxH}Ug!=Z+df)xVe7m;x;lgkoM0+++J5X*z(VL^O^j5h*i
zBVdUKpb>ByfrZ!|fYSjihz{7v3L>@EDYSoF5&zvXWPD*)gwy>}%Ts<1u{;c?YtiY%
ziVz9XSz<v55$QoaMEX!4kpVP7WC#rr89^gN#?Tm%2{b{(fDA-T$V9|~EJUGNp>$&1
zVt6PFWr-CA!w`kTa75eSHbfCH0udkb5p9Rt5k<mCL_6RPL{Tsb(N4G%(Jr_PQ8bK3
z6a!-r?S{J%#ll!baWD>1Jd8(_022`HfqM`o!bC)$!OsvS!6Zb<Fd5P3@N+~dFa=R6
zOhvR8?nSf@z<mI58h~l2_XD^e^#K46pgsuTLDcC0rlZaPFavcafSIVX0L(&t2*5+A
zvjNOTeHg&QsB-|!L7fX=F6tuy9zmT4U>@px0P|5709b&!5WqszMF19|J__Jb)WrZ6
zqdo@UG1SKaJdXMVfG1F&1n?y45&%n3p91g{>QVqpQI`Q&hWa#sr%|5)@Qm1AC#Vx#
zNbeCP02~VrSNs^M6g}wym)M}bf1Y3rw;eE@C!GkeP;F4Dah^Sm+QFK3;7lAA0scCZ
zG@XX?qT5)RSy>Y&5Qg*2c-F)T6cG22`P4PIeV&^TF-d;%LC%KZYGgGgk-5pfPwI$_
zY)5A`JXJR3zktVKM)P>+<F9#E^8`=C2V_6UC=c@_55CBo;fTh^Ox|p%cpA!qd^uQ3
zbUBCrZd>=7YZ0PmY;87E+)Mzp1%Nr%?D|xWZ4qr`0lQ1wS~Ll0>Wks3WK~0=ld|t~
zV`35tdY`^bYfElez^IBa%-bOk{v}#}4PZ?Ja3PYHq9d)AV~halkIBkEl&qYxUaA)D
zX;3rIb|@{=ANssAo;^>q3aQrU?q&M7$z<G8$C5%UJ~Hm2>p~N&8t7C53b6~wcni>N
z0W`uDsfY8vlV>|D;QFgppzOIN_l=oo{n@Rtb0Hpcb%||K?okzvi!fYAQ-?|fjNueb
z3Zm7Ts}Tig22hF6F+5N+5K)k3(41N<_6S@<O0Wg&R%@~XU*4!&FgyHYOL-!G(eRU9
z5LE+mHIOG_o*jO2i`Z>etCSvRP3Mxa4=#dza1jju@FLj%T%?nh%3wG1sur-bU%%Bg
zEM>IF%sz_E&E~#%VR$((FBgCDy4icLr{}lH8}qE(?<6#&uWai6c>x1{I9ci$aUwf&
zo=AQgXf7XuHw->L0{@{$;81F7kdnHYr)VMR%IGd5b`6Gy$b>8-c8Tb>q8-7rCuHpI
zt+9c#g-x?>|9N>51ro_aTy&o+0z6Z>w`OgU6kXcKdy|*Drw7Sru9D;}E_#WJTcbjE
z@I`dpwn36CwMbg#lf{_a5_u&>1)j2_JzZIrr?PCR>T-=Wmew|FZP(eEs2N&OIeM%9
s-+$FLwN~m-R_oHBzJU>g$ueVGXq&GvHg!-_cV6!+;k1$K0^CS{1O8>_M*si-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/span.hpp.D6316D1D55F0830C.idx b/.cache/clangd/index/span.hpp.D6316D1D55F0830C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a7494073071b4b975f238b58e16552aed6501c77
GIT binary patch
literal 3328
zcmb7Hdr(wW9KL&Z6}Wr7oPFNQdRber1zlnvc$Zxl*#h5ImX=|f3JQgS<>CXZVWw1U
zl+hvcm5cV;F{SbWY6_{K4Nlf5n5kugkE~I~3`KN&b?(8RYyRO3JHt17&$s7ye$Ve*
z#*7#`^d?JDnbSrT7v|2+R#6nC!GA@w1Nl)3iqg>(RU8ae)K5R)V9PoZeFcp#)-=8}
z|71#=yX@t{kaJd2ye;<B$y$4U^$i7caaU=~eb<KU`**4f3;GB1TQ_#LDBJ$ILJeyl
zB;L}e+?;b>+~4^^QN`<t_Rxjt$tK&{zo#!fm^N?R#?#vaH`w>fpE@|VW9Hh3U~y|f
z)!Jb-&RGjStYpjI2t6yU-*$+v`r`Tg&D%a-{n_x))SMI5c`pjpqo+=}K4_V7Y)#ba
z)V*^<r{hAx^u-DPq*_1f|9f6-rokyzor#|`_w8#9x5{71pWnK7i;><H-8X4su-}lX
zqnoEK=v~<F#JTgjdlmfHh<WFv(#b!((LtBg2P*E~Ir8Sn^S|9!Cd}Bgj4CQA%*j(y
zRF8rcmnufaXQ}C4krdK0+86><B+eTCu$Eu9RIV%UP}2;i^^Be;0JJ9Py4F}K{iUXP
zOzRk(h5*o(KAx(oFF(18XeFlIf?Fg2wEx@6z{s4Hkw)5#X$xo35&(K)kLu6bro;(|
z?uqG?*c1x^pl4)l8#bqYNR^s4V47!mEdihd`xZ3?TE<V}=-!z2bABBGpo^PU4UN6s
z^KC>MG3^mNPZ0omQF->_vCifRh*n|R&f7%-K$o?ArPuTtoYUpGm9uJ*3_vfvT{5ov
z)q{CzTEx3Gj7CQQ=<?e;@4t|%tWnc3m^Ls59RZ-1<v%m~wXJuW5gm<b7w<9=0D5_6
z$wKFVh3#Eun>n)<$pG}n)|fqpiVxOA(6JE|VtcX-0V;~d{ZA>7V&$kMD^8!;v`S4!
zVe<kb2n1mN>Y`1-a{k`|h&Ex`C-_7HK<^6OExa=6%P2L?VLF<LHWC2(ySA%e_}liY
zyPPv|CM}Wy=)*_qPaS&M@fxD_c()`-u>^qrY3#iN9Wxi;th1Q*8oee0KsSH)M{EBy
z^F@vpFrCh(BLYCTB({uddMC-Eru$$TF-T7U=*vI1dVFzZ1Jtw{(?-S^LjdUO&(*C%
zr|QbOJQq1pi(~-0vyZxcNBx*T5#0mtcEmd32>^X}gCpzQ#gZ#syKS6Ji(~*gkegS8
zoq9NJ3WiJ>)C#5yFUuna%hAFw8IEITX3#3+8sticzEC00D9`f`8@M!~y&?YCECq|S
z7Hxl+XPC@LhD;=wWDzg&@i00p5=0>$8iGZlPBcLtSY+0jP0(!kDbCz4{N&}l2B;Q`
z(pBkdD2{MqTjZUMyIe{YlI)TLDy~BAWVamz0f+FMCj&x&RUE#o5#Vc_zh-y4=P1(+
zOP;55sO;#;a}+9M&NL4Oe?e#zG!Ym$U})W<D+O(N8*u707EJ<7ZPx8I&Wp7R20StC
zAWqn4S((5jWLC)#F@j|l**UurM1@5z&Sivh7Bp<^x302d!ec&zekpNTGQ?rwviMSc
zX)vS=(KI-v_`ul<^Hohh<9MusH5Nn)(u|{*bq;<F-t=IIw81iq9O;fsXm{+Y<QfRB
zvPg1B5|orho=nehunwo;Q)DRwbCA@Y43R)sW|8S>Q!=Op4lz!&g+uhjdy>E|i+r4q
zhk#k+SNVHEZa_#?3=z<<afy@ou$ykq%|nl2r#TM~#vw3VE?H;822#>e{7_;Rxm|7v
z3_wreIAz&|EfvjwUwyy!-ESYOBG3<1c<XxGzT7<(J8}3}6(5qw#G86Qk_@anVeZjW
z@EE$xZg7G9UpM0?w`Cx-I&OhSs(_!;tr<`vmS|J>1e?<)fq;P+;Ec#uL-<Dy&LM`o
zB85v4AwN(X8aIpIl?gdz$h`puf|a&&?={Rxed17ks@HdJPt+=K10xGc0zv=0bm{V+
zxYQw!<#xyA`8)U}?VG!y4{&=$Fyz((=kRgQATJ!pBAuv91Oa7{K2Dzu3e6&eXh?*<
zW05h=m<$&SE`ZL`2MPe?#C`H;t?LdQAODT-+MnHUx*?&|nOZY4^jP?f%}s|7U%I~k
zi5Z0<z%nJ@wh%B#H(r)+Sy(=jW_@qWXmM3HoCDcc;Te`hZ+M37sdgVo42!&y*9(U-
z<kp5WPhJm>d27v*#ZL?aN<%Z`N{79;t<Ba%XzR0{Bbl9F3yN-d52eF-kY5PEZhLU|
zKKoWn-)>lfz;U_d*9Ncz6N3+s-x**TrVamEeocU7Dv(={Ex&db1*R6+<u5uJx;|R?
Z()BORgC{ic-5+mce1CkVQmfHX{{fl@A3Fd5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/sparse_embedding3_forward_layernorm.cpp.CF5214A930A81517.idx b/.cache/clangd/index/sparse_embedding3_forward_layernorm.cpp.CF5214A930A81517.idx
new file mode 100755
index 0000000000000000000000000000000000000000..63d2b0df5697604cb26960f59663bfb5dd59bd96
GIT binary patch
literal 9782
zcmchdcT^O~*1)G`$8lsD#u<?~Ok##KB$1$qBmo0S1cU`aq6`8(Fy!E(NYF(E6Nrho
zimbd<QQt#RFruukt~udb#eiX5<F0XC*ROiMt?BZd?|ko{J#vos^lxreSJ&;TdvD!#
z#?+u7M=J(nQrgs#ob-ZZ7K6dy!+-e&n#`RLI(ZC6$@=3lIqLtokGv8-&^+&v>h}Xv
z&k-l;Z<yr>HV2<?V1;gwA2sY)zH0XCUy4@^jf;5vO>mQebM~*In=jt}H0fs{U%*dU
z5nw7BMFtjav|4x3vEjt4%Ng$b4s`$FlH7dpW_S9yQkT{<@!a;`{0^m?@ZxK~mCo7w
z=16Y&#C?Cf<_{nIbIs0oS9@%G)|@?4u1b1*G5gH@*^W=YEuO8uc3eNs0JQGdF*hk_
z?8+~G6gKQzsf=6W7TfO@6K`!X+RACkUGA{#5i=KVs-N-CHPalTO#8Th%zoN2bX3!G
zV{*57z|e#ky+e702G{R$3hYO1ddpnJjm%oAe@MAAzIEz`+LQh9_BY=gyS9F~b;lNK
zgQ%+1oD<B98Rb>K_)f?y$T50$bHm2Q#@Kt8b~iFpb{sI&Pje;Tofx=fm~!XJh59AO
z8lTT_OHQ$|=RD;v^RprMP4>Auzdb-OB{`J*W74>5Ny`%N&ljEf&y14P8wN)&-p<qv
zY#jCAw>hW&k^a&CWK?r?*5<wYY%MzOKIJRi8rR2f`~K!z<Jd_vf0`x=Ua<AB?cPk&
zAI{5H1wUvkSuth$m=UvrcgQ{W<a%xL_$uBv^=gSrn_M0sY6<15BrccPBev%ZYq<O2
z+V~Z@9Yt)e!DBtklO`)ytf{>L4*uKhtm#VOsf8UiU%xJMd*qV3>(4YV`OU`MXr=P$
z`z3oP1v<Q%v98N1i!2<Mu;H~qz8BbSvyx?RYZ0D2e3ii{ea(H(9j1f3-lf&l^-kZq
z=6W~#pj;^?gC!l~V{Q={27lW7aaT1ik6v}3V-pvDJNnmg&u6wb8Gp%%nClhts;4t%
z){hRk%WF0`2XxmA{<1_?bu|3`zeapjCvyuAe#zUr#C{b&xVorx<uK#mON#a_X^tOC
zJ!Dq0nzPQ{?7p=v7i4C^<zc}YeXqnk+nWX(F5KLC@yV?|LuIkZ?zOjU-NW>7mutDF
zIS-Rg3k7%P`yHG=&LI5jO7j=ul=FL%4>uZL3Oa2#xM{+QOZ-57#HA5kK|^VmObhlG
zNZo!+lLox>vp;*rWryH?<@+sHpYSTxEkO(OPF;Ft5PrQqqWt{M^6TzP%RV&S`_OIi
zu&wy(ddKSS`x^&`cO+Cj;H~`5{9v_hqW|8__e`zpYz}Ka+C*E7O6Z%Ww<hz!c)j(C
zK=o<K^n8|It%2o&t(GSqE(-p~6X&YO7VO4jeZIF`)(+J#uFAc5^}>+^@G47kO&R2)
zsj%^SKB14hzayx4OXpY1JA)QA)QYdUdi`8}H^?O;WAb6i#?&Om=FPVrcBd+~rzVXY
z7-N6piOI|nyd$jRg+V`m%$qv9yYS+^>AQ}FTfYxy))*X*ol~UwCNttfx4?9D$+<ay
zJl^*7@cEn0mRnaJc%SZf=e?qHVBck<TB68``?gHK)BCl(^@8zFZi_qK|EDdstfOsk
z#_|QRE{|spmb3=xk00tM2TM9nk2~wHOm*e%Y2B5O@UV67s9?Y0p<S=HCEhU&tvP)2
zyv^{9FMod<=W*~+x+vx2^+U&afgwpw3*Yxz`M;g;T<_!LrQ;(nyt`WMU$wL0)NH<g
z>e;LpXD>M(y_;@ZLR>p3%U-f+fA+SyMfr|}FFoD{EQ`%A%}HTEzvY<vyqI^cYYLY+
zlffWO*`@-mfSyhJ=WKk*<nB|=mjq)zfO=P!D@QAk-ZM0Oy1ISeeJ(Q%s?FKv`dWe1
zlZ>tUH~5%*;4&*S7^H)TgTGcF^^C(eqmF+6Wdxg<&IltN1P(&2K<daXZQ(c0_U2OR
zDAHNrEYu35jxJadS$y4O9haF3$2Mjg^Rz;zKJv1Ct;OIRE^{_iE7%ILRv>lS{*$V8
z@A?Meqs?KE_8#{BT7lFmx1X~l#a|1kv1gEW0z08r=+v{yuWmM8I>cp0!<(D2&G=e@
z)a1I!$A`B)+R9}nLbZe~5orZd&)-tBdLA=PkIS3|)pE96tQAPTAUEt_!P@2%)YuWE
zlfX%+6-ZsJzWc7&Y_1EJ84r7l*dmcuAa!H!EouJcxO-e?EL2O`Qn6Mb^@`i;XU}3;
zje_&eW{^q`rN34n^~%egd}U{)kQ$poTDw|%YXwqod1*H1mesL%*gKO!+IrafYXwqw
zao&U-(m2oMGLv9$AzNsn6-d1&X!mamt9y^b-YN#^=;7$E6-a&PyR@_G5@xBOI*UQt
zdD!`D1yWy%h^as3_-7OKXwyksfvr$0ka{4io?LKbU_O_b497NN8|iBWQs13e=WiPJ
zVF{O+0@WsLlTliM)DI-H_Z3bM!53PY!yuhKoc*-|sh?e{Gg)Ee-A9cbPC5!4g<65s
z&+UO&@Znqtml+3dZppS3YXwsO>T=;KR~=<d^$sQN1@=O%K<YQv{_i$Ve(VFsp35Me
zJe>TsLZ=R0O*D>s_bt^sl2i(mLaorLH3zds_Y8wuFaeG&W{X8yfz<C4et&WPx?U-l
z`6X1#*fOzJAhjlaPX1^ig8`m}g!vu`_<1;s>;g7jz}8*-+sydn%AY9q=mO(#p3nun
zXx;;C_TcUJ0-L=!cLUpQU`My_2IIQHc*I|f;dZp1PFXU7ATwMubUzcM%2kC?rfVif
zS)N%w=mrFt?U{{Hj%N-=YELyrxgNQ?2Pa64v&IeGg}w7lY$Y#i+h?CY`^=$V9^38I
z`+^{o<VkKihf5~QlQBw>r|5_m<CM<7^Lz{-$Q?kw12~~cjNB*68M}JQSq?$YGMFW$
z@0DPXfKj4BB1TCglPu|$$s?07N*S4gQJO&-)-PQ@U4r@<AD!RyubXQ=3}cbYfN&WQ
zAqB<CWxxSv&t<@i<^~{W0ERR-0zo6qCO?iTJGs^G5}fKTV7QCEvDwArd5lM1>3{o(
z3U3wXIWZ>*j%aLTY=MG|joHSLn=g;a`hrD9nng;H#aU#OSrkUmX3;uAq18a|laNY~
zqIw`|2FQ&Qq_`f4o57pE_Ax<9>Vc#gpcq1s(t03m2FR}yq^usungQC_1Zi0hESmxH
zK?EtU2l8fsf($__>VcveME?~J335u<lsIH?f{f+Gnxp9uWct{2A6l0y%XOr6d%>8!
zU?O7F3g~^B5r|&=ki|}Z0XNbzAXrAfN*37!jGBN6>c=8`fS?E0_ke&N5Jc;HfuI-H
z_X6u)V2kVBdVvRGl!`!eQ+Wt-aM0ZCK(rk<cRLVo$IaaiB-?Rww*%>R+}!OzwjDQj
zJFwi2o4Xyzx8vq+2a4^Oxu=Cqi$)Jgka4^?BV<D8f}|Q7np(lAR$xYRH}LBQQxN++
zx%_m;){RAQcNYi>%;;Md3JNhQ5)@%nVpt-iTb3G@VpL{WhEau}0_!)AKhG5P^H*<m
z5c%x;{Js&Tk8p`mxg$y!gL-S8W<Q?O3SFkRt+x|;OoH^a^~K4{*B|LZ`_DyOc6feb
z1dGfNWa!pAi&P0z7-b4FG0GBTSs>%H$ZSD2Mmd5UjMM@()-P9(t9ywoQX|lqqPw6O
zFyVw?wo9*3oKV^e8&Nqg-3X_spAx5CSsUW>+2a**h(>y!WZQ6?<%!BLpOz;s!#rA^
zqzv<CdD1e>o8`&MFkhBuS%!JCJb4-B$MO_qv=<8s3yMLnmmni~ktWE>1R2GP!YG;-
zjZu<&66WEO-IFm&aZkZ0OOb_nzidS|MmdTcjB0?W23XVM*8)*3&UHXkM>CoOe8JSQ
z1be}O&WR`yO)R+P@-QU1mq{gHp<N&h-JM150ls^HA7XTnft!Rn#h_dh9bn)f)aivD
z1Sb_b#^`-=e1R{s1&G>kpV<P$ZMe^D0g^V{XSM)o8}2h(fUFJonJvJw4fmNXK;DM?
z%od<%!+fTfftSwx5aiUbsb8X3ovFz$3V-{pFO@HHGWf^}Fkj@TFv{d)VwA<n!YG@Q
zjZqFK2O~8{jZrQq7b6WvqoeE@Tdqr9u#SB8{g=Hm#X8{jv|=_vRsnt$(0L((TnP9J
zajpjZYMg5TzXs=8z^}!*4)E(}M*9^mAnE{1FCb;0G69_;^STB?UyBxQ`0QptUI9KG
zb$+Iw&MD9-5gn+|K^iuqPSIE+jH<SW<V0^c|2<q#3xMGQU`a2jD&SH@Z#2rlE(>w4
z1}@b&*8uk#5J>m01@5&t*8%rBoEHJ3ML<EfUkpYs26Dv6nZl>14*B%c+XmVupo4hr
zzh9&c^-c5ouWq95d7yCo*(GKKK^mDENziGTAWh6nB<OHVkfvs)5|p?Sq`8^71m#r(
zX)Cq0rk}!2YKM_hs>H}%YLAhF)Bz($sUt>CQYVa@rOrB9<mBoQ_g$g}raK}(Q3kRG
zOqcz{8CYoXlVqT!3R7i2X$BTr{A3weXz{bmz(R|kJOeEvm?HZrGLTzd92_ycA+WEF
zO7%S=v4d}vM-)cU9??2lRxfbbxLLgiu*eXj5M3IDo&>%@st!Zj&OXpS5&3Pjm7xEo
zsxfc=I_;xE4D3r)Vd%b<HkB@DiL8E;lu*SJUWSL2b-;HW2tkVGbpMCPY){ugMW{)r
zF3VdtakO#n4EwGBb+v5tAM_GM1nd9p$E$C$gIUueYVSbfHUZxzpv%x$WG@Kr1)<2#
zv&aKLaDYz#Hx1jL`-7?SN6N(3AvUPlre~vVhaOq}4tH;Xw7?4eOpt}rLX3)}MHrPR
zO6=&CrHWFF$`oZ7RY)tae)Gih<aEDXz<w7PgSc%`Y|FjEgTi6(>@Ax$7X3_+39<x?
z5@m@PrODDTN|&a)pnhn7!&O05*--84yMyU&G0dSW@CEh&s~(`U17$vOFVLOOx0Un^
zo1U{N2rhu}vhlhZZ9Bvt^H1Z5k4%_4`*=s9(FoFy=chaAciP^Mj(-th4^7<zB6@(X
z+0H@l;ue!|6%~6uGeXf~gR$3Bg;AzwCPrDlS&?+hY~O5*a(r_zQv0eg%AJ@yo$jY`
z*GxsUYt}WFO-pzJnEx*W`pW>Ho<<YkH-XVK_W=DKT93+;dU~IfC82}q1fqUC-RT74
zemvdj1d@I{-RT6<emvdj1hRfS-RT6D{dl_53FQ5Fy3+|1{aCuQ6c9`4z=O)9@ZD22
zQdE2F4q9EZ!}q~8_*!-YuiaoKn&y53&bj{Bq0qm5v+ghFWg0mCu$kI(tPnODug6Qu
zA6oPpJ~SD^2@#;5QTY}=HdVVt3Fwh4TgEcZ9L)UeDZ(jDN5uQPIr)FPSx%1(xcZ*=
z*z<3{@fpkVMOEIBJh-qMfLQ}w?INJFoYDkBkf{l{+ar5{?)-CpO2eu8-KWeyyN%C(
zv<Otyn2o05jDDzYDIj+RgBMk&r-K(xS42I-29%Q1u<vdl-c9?mf!MzK*R##HQ0vP)
zM2selQWg`vPYU~R_o_r{EIm+()mVC<lBlutKqXb9pbYD1Dw!Hf4^)<FEIm-k)mVC<
zQm9dSK#+l9fmlH&oEMH2bf)vBV-&%Qz-X51tO;l`FfO<zV3g>Zh*5@RhOS@^y{n}P
zqfE<8j4H=gdLq9_od`r#ctW`lh!)~p4Mf#AcY%N|5RCc{JZodeHT}8=28Tvq)(CW#
zCtxf#J3#YwqeFj|`AEvBple;Aq~}y<U5HVUbrD7-$`Y(NQ>rY*s7zUgQH6B{)^DC-
zp04IcRe@c00e3Xwjqg{VGW_~w3tTDR0Qon-8BOJOV!^F({>!<aeM6v38df2xZ%p*c
z7gdNB;}NDpyciEK6_Ukxe5sHw#=}d6Y%v~PDl8Y{!KFgJ7>_L#ip5xHnH)Ab9+`_E
zV|Xz_WFvx1b4~N3b!tns&VB?eGK|>+yir?Jm4LpT`Zk72RKKj;@p^;uY^xqYhI7IV
zP-}vm&Y6x;1SbNcNKPb1QJg4@qB+qR#c*OUisi&&6vv6fs1*>cK!C=7=Dupw`!mJ?
za9Icggws&iM}ZW2CF%<v{cdnVE&w*7zT;ty0v!`kWJX06*ogX?M>oRh)xFf61G$Gw
z;Zp?419gclI-EjNQlI+J85QOI@RrmkKKhn8A-@J2Q6Ku~Ms|UA2`K&|mj)YApZn-W
z4uKAdsHFPg@ENPMUM1`Q-$an1NzaBKQ{n$|(xm1eqArRl@xOnC=z_se)0esQWj=jb
zN?#VxmlgEoJl!QAe*&w$ocl^{@yaX8O6HZSbMx{G3X6(0j67W<Mj3snp)ZT*%X0d%
zguYD8Nz2ZFI4wQo3)V39aLx!Wp+AymFohrBZ^hS}A}}6hXkt2U`~<Iw-afvLqm`bP
w?l$88>sKRlp@l>$lPj!k?d+YL$BcFJ5V@K;xJ(*t6g2fKW?+opWDv~wA5dF&ng9R*

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_bf16.cpp.AB98F65366A9F1A9.idx b/.cache/clangd/index/splitK_gemm_xdl_bf16.cpp.AB98F65366A9F1A9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c28e906b8eaeef1784ff97682bc53f94e38bdcd
GIT binary patch
literal 3188
zcmZuy2~<<p7EK-)APGYVApy!P!6b%3eGnp|AVfd}46~v_0?3dc0y0!lv7pw13Nob#
z`o-3-MMOS>GKxV4@k4Py96(eMDYRlmsyO|d{B;{#t63}ioqP8__uYHXJ_+B0#}g4T
zn6(ieY2wJ_aAgbzLjvEV<ir?vFsM|=VA4*u1U`y?#L>FfKketd-21kuX^oMnD?XQ=
zoA+~f#Tq~J?`_;-J#H57{bY1A@g!$b`htJ0sM|p*k?LuzQxmEQ?fUB|dSvgS3AfhI
zdlqY_*v=OnZPPqh{5+X6dCE}LY1NNDQ7ZTo4|-l_j!w_6nHlPs{UFxa(?-+Wx3_(C
z;@D1Z)7A2`-#wUZ)^z@HZ&}~%2jinDi5W?S8dWCg8t=IJ;!K9M*sXcijJ%bdE)E*w
zhi8qXjbE(|5Vw@Xc~ZxSF-|)Q`mQYF+v~N2ed-O0MN_Yf=X9%lv{Ni=JIN0zlKmu!
z*SOaY4l%~(`a-Br)+d(wGW|MwxSYkWtQxAccjhf)+W2t0hJx+7yd>97uj^#R)eGJ~
zl}`99)sCYYx(9EN*f{rS)Zna)m)4i?J3o3Z3aQCS9C+F1#j7)CAfq{Fb-k4<-xx+4
zC}-a53$%M&T7hzK0i?ipqxyQ4Z&s={N#mKjINP24%X-s_q~!eUm=c>o5%W=6eO=SK
z+GQQgr^k=A%upSAi>@>B7;63kX6Qgia&i`Pd==84=*AV=4thpsXRC1PX`-RKcYD-R
z4jy-Pna&&<v?Y2`I?YaR4q?}bR?n4>yG)i{%d>3AvM9_fK-3$QT%Hr>B$x4%*t(K|
zz59Nczs(&xRaRCL7P9nSAimA#*<2{!;6FAgkwcywa%#g*Ez;)h@BdrsXxXK2jem&E
zRT)S-ncY)QdgAsbk9YdQAjfv|(<2=R*lR0`XlLWM*+!M5&o`Kw@64)<y_mfD+P`+O
zNAkN&I02lE#RIOwYROznf6!o`p;`#f$A<LFOnmQ>aL&8;B5vn$`uQgIWv4*B(l{1%
ze@gR%(C$Tj7SS;Q=M(l8*pyat3#v7R-4gLPzj|Fw8cCz}ewi+);Ci)xIM-5PWaSs}
zhhx<AtE`JBiiaa=b;DdgjtPH@yLi2!_Vk7MC!OW?QsbOEyMDftKC*hMr_q~pR4Fwu
za9a54Y--2E%^ChZKK(lasYC2F99GGoe{S@~6$^x=$BAc~4;}h2u=VoqE>iE^$(L*&
z#^t1p76$*S9Pli4?9TQFX4tl2UO(OZm)Dnmqx()x9w_p>O^x=>)kG$&B&So(jGkc~
z)5Fi!C)YD>j&)Y$8+&Nze@Xbvb56}ouPyYxc&PJrXZhaJ5o+V)CWB(a$8X0=X-C+%
z8t)F)idoJ7HYl$PdB)zk?SzQXoc(2gP_}Q%^Ng1deXaPCdF=+Xnzu(Cqc;<qQXBc@
zbIePx_sryK>rJ~JO}O8+o0D<f<G>xGqo(PW8XJ5M|9P-Cf6VMg!@$Z9llh?=BZps<
zWfTq`vZ~$JN6wEF=sp>?G@7aSWNt6kWTcx~wE9-g=>&hFZMku<UN}<vVJhvgu?l^_
zHhi9<YqY2Jr%mr?^!*N-H)g!WlX2U48mAsdmw8weczjmEvVX@UrHch31cTAd_i$7$
zaXLUi0x%eX8O}^o4&WC@{;~cmy{UE1T|z*-0nNZMwB!J3#|!=ChZ;|uCm<UEZGtn=
zlmnzg2c+?>RlH#W5&~!goPnAgARQgL`+fyAcZ`7e0@@U3N|FPl6O<^oRB>A?2nZk0
zOdL~74v;?F&!}@>QrJ$!Bm$bN%vF^Gq;nFqtL*DshX{xtps6@2UJj7XbLL?~?|v2#
zkPUz~#u;nM0n+*WbC>2`HV6qwD4=OLnwlISU7k5k7ivgr2*@@->*Ms*<N)bv<+tKn
zX!k1uvK7#VI75;gAl*cw5LTr;4kaL)0Bw#lC&>ZQEms%}2jR`X2*_qYTjQ*0a)5Lz
zUoF;pcMzU{_yF1hXQ3_!NZ)u7ktMZ#lB<r{0%*1_+f)va?yhWK-uS6!EddDxGz-Vl
zk^`hiMlFT6=#o(a;tyy#j;<~TNdIwvS3<@)bq)au0<<O0Qe6&^em6CeXZ0e?jDUmz
znu4RK$N|!ckx@y$;6Ds{4#9urHKG7DK^gG$(0M>4vmIOmi2Tv5Pz3!4L^AU&5K&Mg
zDB573zeAUua~w+)7;+5P!s0-4f?Sz16}faRIvL)>wzQKw!djZ2gIp?!N`t0?&IP$L
z+bVKV7HS9I3;hd3GV?AFg`;s$l(OFS7Ax4!7rcP&!uE&N%gnfv)js<4s%$!ty#-D~
z&tvib=ym0<Jx&xD>lo9Z(}@BT9TOV7F`~d+$D9V82sL1<vQ}8IRXj?DTZKKLg#W4+
zhyqg`Q!2a>qQFeYj0!6Z`vj!Onnba#g_Z>w8ks1t)UqT)YgMK-E%Ftg;enG=G$^{z
zQdO}lX6dR&9)f8ui|Ya}0u~dTT-Iia7cj^SBWNL_fJtT=!Ez7<EHcXo_5r*BaB^8A
zDi%N|l+Ztw;Mw65T)eCyp{#WUKVp&FnJKU<eG7dyJT5#LSX9=ZPz}7xL07}2u08uU
z$?oO5^&r>U!P*~A!%~A>S@SN)%|H)hp$cvvsAN67KoyUsK@l`D5Xm}ufyfzkhoULf
z%3}z(TGkx~76vx3D_}W5uB^=$<a(kTVJ<9=D*mfJ2j}d-qBs8DEEwOzqJs;^<5={?
z!f`c%UPj=uK;Mp_*B6cl5cKxKai<d6^>=X+BSqn|4JIWDliI*jNgy$4Z!0IxC28(R
zh!+bZBBP>XV&k?0RUAAB1|w30^6dW)g^Jl|MHr(96BJ>jB8*Uk+ZAEFB8*amv5Ihq
zA`~k^p*TV$Tg5ZdOG$YVPDNFXKvdV#)Nm)cy3k2@cWo+Jho;AISn0URY4sW_3X5xQ
wx5W6LW_1m|F*GqXTWn#;U@o<`v0aYZ8?Df1FI!8|<$3Hx+=89gVb^2+2cN|!ApigX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_fp16.cpp.DBCD688A2F933855.idx b/.cache/clangd/index/splitK_gemm_xdl_fp16.cpp.DBCD688A2F933855.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d41899de2ba53e0fed8019152d0447fbc2b4a1af
GIT binary patch
literal 3098
zcmZvd2{hDeAIHa^1~Y?U8e^R?nwg9>`X?jF9xalPA~iEcwvjz+_Y#pyi}jMFTeo_Z
zB(zz|v<Nw*D-~~AEakQ+-7EEm`^>!OX>`t=bLM+~&;Rp%*8lgJ@pN->Q6pe5j)86o
zLP0E74uip{f^ST0bch{jR4HLF3FWOmLlHw(#D~3ax7aUte-PemM-9Ilk;lqAdb;DJ
z-IgT>mpO;Jb(QR!rFKP^TTR>=_iQ`T$r<e~)l|+><SX(mdoPhSN$$n3oUhNNFx2AB
zzZM_5p>m*PB-U!8N>{;lb=jsMc|5D3U_@5*WFqbP#hi)HLk2VZ%&BwV(nY<rk=Xc{
z)o$_)6N$8Rbk@4!RKq>dyV0-XhPwXtQ#v2w&_QHNz+g_ogfpvU`_y1<$ElkNtslE$
zLe32)nLjEt?a+LXGHg&0CHOtUB1tq?GT9&cK4a+Z3jZp`Tle5RGoP)2h)bYF(Q=dV
z-&4!2cklSNGk++<vsB-GKtxa6do8Xg`&i-Tw8RIT_V#*CP?y|+xWJbBAfsZMZOYtj
z{fqH`q-8uTR3sI?)eGA-B?|h5E$F`Bb%c4(fRgvVKhSGczCNbMHS~p3o9(l-%l1Uy
zcj1Ctt=8N3^F7GiES1`$BI_IbpX;_%_rECr;t|Ih*`~b7RsA&grf)-a_{bBVNorpg
z@mz+Z?exu^qsD%?Gft-&9tW!<ok@T6sh%zTtYCAGJdu0H*XTF0CG~lv`Qy6k>H_VA
z14o=3-ljbsFehvvx9cDG+h$T9zGk{|%weLU?dZ}g8LXnTLPY6`tiuT5TeolKUq5f}
z4!&{2IW<Igs<+2|Tb*%ns8QzArBTVI)p^@Ank~Ya<p+m_A(P9@qIsWKy`2}^e@TzG
z8?i^JySeqk+k5*`8udNA!^(d>AL+~!71Mpz6FgsS&D9L5oD+7q>d?<^H>to|HP+to
zo6?XwgIoHi#o1s&dC2_dFI|JzmeB77D;o01!$M6aL4Hlu#=A4^jJPYz>AQyf7~vRl
z|Io@~_-k#~<V?;Rl6wxT1(AHR)V}^k;56jPma+HtXixIKR8>TKWw8_H8~66#66ADO
z+|8Y>sEeDbKbAXy<Yrf;Y)C!tJyWdi<vs8B<WuO({-~y&*s@2n4V5)-GDfT-cMI;=
zPP-?qO{pcAd7B6<t`hD4+upC#UdDTQ;Z0y9+w!tvr(IWa!d2QkigI)2$>hjJ`L(M)
ztTPTyM?S6^o@j7V{ok{s`wg8Pe~zXXuX{i_zGY;8oB4Vkt^BzEi4P~vWow2`U5>rX
z?iy{c&8N93f0~b+bBT)2OB{8$Qj-<%vAuZTu|7)KM3ZI-{#(-6F@u~P_nP_#8s9Nm
z{-a%asr02u)XswNu%^uUpS?3Z;zv?n4|<q-mVQ;E(Chy`WE1Qs*E}D4%H~~3OZ$|l
zfno9Zpw5e=!&~=!%_xvNvs%Y-%3Y`IIB`s=gxTqw^XmSpp^1F|&4Q=n6~dx{EYrrH
zderg-JoRT!S+;d2XO}D&4jClUnN+4vGOp&QTo<0>5M%P;Y@E&Q!?Fdz`5v=4C6$2i
zY9HGc7xK$&?9Nt)OcB51MNwWvsh30P=$hJ$*L!0Ug}jXj2BV(uW+V5jZ8{$D!eF@i
zSbY_VfL|KP$Lw=r^Yv-hQas`gumRS9C=rNl&h}PjH5F9jk<9?>Vf9ob0<m-6)W^`;
zZ@zeBJHQkyMNuLUuS?gw(~3sBDq;KqHdZ&LO9W#7zFQI3Yh9k=5nq6{vD%6ffjHPd
z`O!&A-Y6dN0GN)Yt4ajoNLg~X0yf|z9`OX2jb#%h0&#XP`;x0ck%)ka0+=JmQIH74
zN9|o?{QKv4cw{5MG%QU;A`s_$p1Cmn`U)S9_yMee)lifO#Fc4dEWYxsdOQ*UuohNJ
zQ6do6$^9+tUePg$M|J?Li`7+?2*j;d*=!EK>oXqN2CymCR6`;VU-wiDHA~)#!y}sj
zW?&gg5`nm*Mzp+Xw$l-hcmr&RH6%&|;^F5@`Q5D2=Xhibz)UPtNg@!xe-sy)dPd0#
zkN5!0!m^Yk0`bh7;iIPGyY%tMPJqc+vb;nf#zYH(Vzj}(6HFC?`yovPE~<yJ;r?K@
zfJq!II0!I#p#jhY69!D;cr7rIQ7SaSoXO#S$QOZ2TTG+CS4rUNE!LyJ@d#Y~#rhPu
zPMAb6i#Uc0W=Tb}W#BX$SLW|gH_1IBL*VLK={myK2y+TLi$f~y%p@{N@VsGKfk_-y
zsR=bgE#U+(v%n+{>;e-P4TGlm^-kSJzLp+f6~+$6UU2DIADneKy}4T8F@;D`herZ)
zE`$5me1n<9;a)Hk4kf{vV9J3>9C4|MzL-vd8-iJv#r=?R0+&H#kYH8`To#c<g5j!3
zXkO$YJmCVCN>(PT!%)?RqJOazat6U;juFQJt{fgSc)8-MB3*?6$$$zoMBuVXY$`lI
z0@sjaNQE~6ZU-EJ_%gu(uA`oTi=0oq0n=$QwM^hX;F7_n#diz3K`O1#HE^i5mq{^}
zuV>bSzNQ>gFZdYF2>OaI-h#dnXaY3BL;#cc`YkZoqpr~OMxo{~VqGVWnl@8=8I0Kb
zm9*~YUB_>O$HqR!0r1_!g9Ss1uk3;$*P)x?knorla6jBxu(8`R=$-E&2knP4=)l7E
zm<&3;uw93s7ZErt@QV=i=E8O#f<9Q-ZkI*x%0fT*jR@z8R~Qq`kGUZ(LdEj`&n{SC
zFk#Y?CoOkN%V24_M_LAcFJ-wThIgGM+Uukeu}2ur;|B_M2Zw|Ph3&%dznd^Y(o!fb
z1=2D^T1H6AaB0aG28N55wN9`>R&EhiUO^F0P$H@*yQ(@luvBrbY82994NWV~Dx1}|
zYwS$PM%GI#4QT)NR9#y~SC6jGU@m1F8k;RMU%p}`71c7aa3rg{xJ4q)TkY4%tjGKV
D`92cC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_fp16_fp8.cpp.60034E8F6043ABD9.idx b/.cache/clangd/index/splitK_gemm_xdl_fp16_fp8.cpp.60034E8F6043ABD9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3506d242fd6919d3e5473f6ffcf35dce21e62f49
GIT binary patch
literal 3328
zcmZuy3pkWp7are?VP?>XnIQ}_8X8RF_6;JUa*1woE4r8&cgEd6U7Sv(RMLqEf2tEo
zH!8X+3YAMriu5Z|a_JC7MM~%Fne%L;=lSP(ywBSEeb@Td-g~XZ)6Lm=LJ5OeEOZk`
z1tkRHFc^#~_=``73v&jGGGz=#oL}klQuNZA+}1p_#?fNM)5wZN^vF6<dLO@tJv#46
zt1HtZV_jxh&(!3?tX}KbGI?GOvd7~wBjj;YjJ1b`yGHru(vtezm@+%#(eGRZW?+Be
z$oq+K)7u)wwQhe4SH?^V?ejixa+?mSjN=UJw(cx8I_-OXbxfi?{pXMU)y=v+2}$t_
z-4x2-uhrdYImJ76sywaw!@J2obDaRE(b*rPEITl3HgxP4zh7!_d);V9Y5l1h#mcXZ
z@nPpXwwSeN&9B#dx~*&8vDly@k@=SUKifxI!-rG+M=S$$SwkyA)44vrLd02Uo^8SD
zD@x8^5VHQq#!UZI&phUe;a!VIhvNo~RIS*B$XHSQf%KsGrYwTeV7)l#k+X?I@?=~i
z^K#O&<0)-f1huRoy@*YtEx~E*pr(snhgc8i(b9)og<iIqOw41~@Hb2D+P~g$&5`W)
zAu_0T*MbU#x<8F-RCELMcP0Kts<sucIXh)DcJeQe2j$wwGfX^cQ!WRu|KEf1^e&x(
zY=#dZb=y$N--pU0+4-~cG8eH<*HD&wxVW70?9%^e;#*`v6_lj-ytt8Cdw)}H#pxBI
zZcV4zgv%~`|2k7sZ;iCyMZ&A~Wt}g!9wD8&j<_6uHT%C_cu%6`E2?6;#YYPP<taAY
ztZ?cl{>!PvdIwvZ;Oa8Jb0z9YC+>eZRm2Mp9BRCj9b>TN`pL@sx9j&kI%;QFH-}5?
zK++Xj#QD1$c&e|R#~Su8@-B(x^J*yj<o);fam-wgZ8Qtc-8^}Vp&y%aA^d8>{O%_e
z)aV8QIo~M%%)>JUW-+IyUiTVD^dC$yDP*15>n*+>m}RQ{V;fJWN9%BF=aCn>=MOx(
znoU~d-xWFk9nwGh%9V3o*0!#>%l&kg4WydCYvQSdo_?hneLJmpdtB3MozvY@Jzb8o
za@gkop;~+msTn@p-YHDr-rYOV(%;&?Zudx-qxY9P*_9XeILs*?&s#B8v3+NA*vjXL
zHL-JVl$Y^OY;~!`+WYRwId{ZdGe1pFB$&htmYaQRPc*+lj(C|O)a%9Cy-H6arFPb|
zV68iL6mrCDs-<Jb1ZVuSLjzOc&gY_#!x?Q2Hk|CDi+*OPy`uKbxnJ)QmMN5&Ha6cv
z@=h*^b`LpP;2|0^`j+u-(<rTfL_MiJ*RU($;g(RHk;wNsgx^k`Ev=5ddH14kRa(J|
z`U~z+%a8Weq*c6Gd(^Zma~y5B*5p57<>##!V%4?#qH;eo`z$xsiYd^o-^D{OPp$NA
zd?sq}HyNq)V(u!bR~7vbYZW(JyTx==o6Xg|calI{y&-;WW<Zgck$mac{055m^Ad*_
zK4HlNPT#+qFxDCbf2xgoi?!`=FngAikxBD>dyIdz`0ixtfHCh%@|a_%ZtCJ6)~5Ta
z2yM|%J`Skng!k6^dod0Ud~4J4(zyMHHoYtKxEm|W?WcV3;PBb_%~AYF1cRYwy4m4!
z?RSz8FARpSkJq0i1u$=<KkoSETUB3O^GL`FK(p~|vJ@a~cdq$(YQ^ERBxE(98F<Dl
zDL{H@`sk<d(jR^#WF4S2@EQauKsunMR#a8$+(|<G0IiMJCP)F&p#fXl3u)=^NXRNc
zGx1DSDL^^~OKDQXZzv=oo`5#O8<C{|>BJuTSXxN;Q)SG0Ky#=Zh7=&Zx0!v_bzXM0
z5+)ANwm4fwDL^{I(OE8_^(UW%EC;k6UT>BZAf4%1eDUkRbpZ+S2ec+$lOP31AK%ey
zAW*3-BOx0Bt%cVjNCDE9a9^UD(E1S)vH{R^JY7`^kgmMJX4?oF$4SUqKy&e24Jkmn
z%99Yz-ResuAu9pR!n2g60O|S*)fN>~4U0*LH=vF2#$+i#x~tnz&}5L;O+wZH+5m5$
zECon+AGm#jK}w4tApwA<;c1Fefb?*CVoY+evNZ|u0kk3BP+1C)o*3-PnBTXFNkTRO
znu4b&NCDDuLBa8z;6E6wIfVF2tts(QJ(LYk4=WFdBxMI<0Ff8E0g7P#VTr%g9S~7a
zIuwCb2@bu6gKJooIN~qWsKnP(*Q3FQSK{ld>(gMSAjAL`9mK*A%dIk_shk6c<dpby
zYx-jNps=7p&XQ`D<!nGUP=hzW>FwSS{A;}~;ER63qQ6h)#W#1mr50FqLqg=?&_AaM
z*a#r3qzhzWQDc+`Q-bXPM3P>ZAqqqzpePoH$vbQM{fXog`*CQ_^e14q0B=dF%<xV|
z_sT){B=;pv#(un2z@6j7@q$h}l5eC`cfYwQ+4kYAhuf(imulNIAD9ESHOLxNc<r$B
z<cPm?TzXIbO3wDRL6X#V;?R`o)ZmMNH<t9_jMRvz8q5qh72weG(@wBK!FeRzDRW|~
zGimTXVVi=;k{*?bSY(zOY%?XkA=yw37XAhCE%{YZ1<oK#iVB4a3%)em^0R^B{tmEh
zYi#QTj|l4mRGOsAWglRxvFWhDmH0+#Ms)amN_=BAV>)c&D~1H`Lyu0r2XSd+nmX)G
zcxrHklJNlJR#|!q<(C!=0<{)Pivy1bPXg{jGA*DQT$wfM07Kn<yCt4CFyRh-xi(xc
zxD68lzLHTg!&iigp$Jws5J?8g43Q)13Pmt?Ad-v~nTVmzpu#>LRJ`y9vfz@Wue4d(
zrm&yj?E_!Q%#rz`%h1*E1Mn>riN6dTaLjvh=>6#$2J1FC^!dznuN>MpbA1UxFC#E4
z&{rd9&CGQRf<B$Oet<>mrbh!NE+{fkl3{$DApVZTh?c;=zX)hxFjCFG%ShR-KxT%@
z%rKc5BQt|!rch>v$V`#U43?SU(<YX$_80MDsnQ;wS<x|~D1k62I5aFgA_U|gH7&vj
zqJ)u>pv!_husC_Vf+B&Wq)eWr;;OpD$v~CpI-91ZuAymdV{5n2-eJ*vitz${-aOs^
rw@cO1*3n}y=dujhMkZWSGYiy;ZaIfzzL-LFc8f)p`Z~JExnuqTLceS(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_fp32.cpp.A3B8F8ECA755E6B2.idx b/.cache/clangd/index/splitK_gemm_xdl_fp32.cpp.A3B8F8ECA755E6B2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6bfd4425b355fe5015feba8b0a500f4d55dacc18
GIT binary patch
literal 3098
zcmZuy3pAAL8XiA}xyWdYaUEtd!?=b&8I@azE<%!9moei$Be^T~IZBaCJ3_l6TeWMc
zNV-bVL@tv<N-0*3B0Gm&Z56t4zL~SWhPBSW);#O|zUTQa@B2URKmOjHo{NbX%xb=O
zx-cS{tAxQ|G{7$@IWfis3@X$xnDnx{fdlaa4y32u6PsM^d>#uLTxf!}co9ofc&`17
z%cdm<mwLo{OOE~WgC<EVa~NwL^}lzdgEP{3TxU@(j)&vfcVDOIkbO%3^k|&TG+vx)
z^R@JFllp;U!^sX~6$YwKt4?f;R3Wgg9vxN`zZlCNx{^OOJzzTXiw$k|`(AOk;c#+l
z(kgG2t7934d(j_`<$qj#DE@2Y>!_JQnCqnehZM9Q*|e=cKYh%Dbth!9zoz|<`>J>U
zmL$bo?B8wES8Ua;^Eh+Rw0u{@*?8OC;@M;Gdt=|_3`{Hwt1zDMi56K0Zs8-IeA^N`
z%h9u0WsXr>zi%%Z$nihU@OdVtXQb7ql;oW(_RG$A%xP_{<3vi74y5q!)J2+?8aidp
zwlJ=wKFQ8`T8twXPZ-7Rm=s4IVn=jd4miSmXi62m>*WWmC}Lo`ykcLu-*ftR_6=83
z&|iXx=DUs|`*^+-ZmxPwq1dr$-;lwbs@_*+Uwl(o!@-L-dTE{G-VeH3B^Z7Y_@34y
zAzjQ_?KE|N`-nv-{=EA+W8Z^S2_EDpJsKB^r&U)zq>L3j2r{pv*wcm*Y@XLvRUOq!
zKXAm|Z6f>mGaKT1N-N{{&|u3tf%DX<H*RC)_X<sJ=CDe#ixIV(if+Ti@0~&WKm9YL
zGrFnCBP+&Wvb)PCxYnXH)_ng<lU;kPszkxs4YmSi*}*|!%=@L*iQykv-5pn25A98L
z8Foc!QQSIVOIlCnHHPozxU%1>6FkDjrS!mcME^gx6zD{rniaNt>C-QUSe6qUuGKsU
zozzx3k2m>`?S*J!S<Kw@7fFBp68ghvoLM+!P^iNssyJ6{Xq#y@PPxgPYBS>{3Ev=(
z4lh4RsJ~aQWLa%S>H2kXBsnm5@z+Ws=c-8Y1Us!u?|t}}it@yOTz2~0wY1*mi2Tm-
zwt^q!wJDQzCkvh-1$h;j>$9pi&y;EfY@Q2!@iBI0->%zT$tU_u+ExYk;o5c$bxHEG
zCdxk@o(a?0b|FN<PoVF;C92XZSoWZ0dtb+pzd%4P$}XP%kl@^Jt)EI*Gd&coemHfg
zB<PWn^6RuaPofSorkXzg9(A<vX@|09$ysZc^Q|h`ywVmY2iJP{R~}n5?0QFpV}eb^
z9ipvF=OX>lxmVvsuP1arw(pM*R=@jNWckDXj5Ycs#m<U4^I5XoPqU#krK&n+z&qV<
z`N}xE2cfSUSudU?6Xp~y;wvXwzdgEM8Ftq8Gt<hAKW>om=!R<Q+}PTTe(Ma;n|)E<
z{8}>S{%6|d!B037oX;MI10&@Qs4HEfXSJ^AkY;yf+Kng7`qPrz|5`C%R;|7@dajIV
z{IJEMv0+2zjy2!bME2kCn{->da_yxl&7e7LQOkiF-h&svji=`ss<3*DE5A~7X?LFc
zS7)Y+%vv<m8h7u<?`u2KY_9jxwkpTA;~%?!dAl<yLm0jR!C<tCyjLpy=Cqf91Yj^+
z29}{N1MsDhJ=W714UJP?#|elJpiQx+BpE>3#^=+a7~MFM8YT?T7Frf`89;jFh3-?i
zw~tm45I;a0VU5&f0BH}=<lnJ1--8H92%xE0DozHF4(n--Z>;frNkD=Ct%uda$pF&P
zVSD<{P(>pI#23(XEL}qekWNsfbgE*vogpCpfM#RaBpE<DubX|{%d|vH#OwkzM~R~<
z14tLTdMbqV&W00^4S+Vp8mh|x(nbE~FHgO_$s-`4fY!!p<75EoQ`v7=yhY7*1Y{ea
zb+NiQ89=&L>9eqNS^IkevK7z<SOX0iK>F@2Hk-qfOcRh`KwDw0v}FM4Mt@wa^`0$w
z0<sa%##m!D89=)IlGyI{kB-#@WHX@6ux2C~KzeY<gxASBK14t^0h)<rs>uM-@A^^_
zvd*hH5RgDXv#=~R89;hwe6Y}JbO(cgYzH(2OHq*lq%nyRkx6>s-wC=3!7p$Vk&7Ck
zY?vSP77$5|1vder0CXD^L5Bg6)L!#MS!kXDR6(NwmDFYPR1}m3MbK^v_ysoGu)Jue
zmSw>a1tQnL!C*CfInaGzOR4qbTQW&ZGOQx>AP`9nC>NpTs69LvIuVGZHk>EoqH$1^
zy3W1RJjmV`+`_`mA^@hHopmcmJoKtwIvr@S6;9`V@X+UU7wCd%Dv7EEGlQlE8A=@u
zPl9F!7mzv`PJ=?@Q8K&$w6rpQft86|Lrp^}+=0k7(lnyNY7@D1O*$3UfXHQNGN|y)
z!xm7)FR(t5YfLgGLkknREE0<hJ$EU+LD^S$+7l#2Swzu-UaN^kKe1Hv`@uBFoZ|*_
zgzrfSzo2=DTvM_s4O)xHWs})7_yUMrGqM>CHV8}{6ieD;uqutHKVRA7^f)+9m#J$B
z^MF?dHI%j-RD&8ipw4isd#`sV*}t7x2i96~tODRPJQ1vw_T#*@@n|{}K}!LVv@7R{
zTv0D58dtsaE8<uy^{O6IZz=RBtRPq`ZP|Hi*P?!KEv%<1enHQI3|bV>2R~a0jGro?
z&*qQcD4?VB$F&H01%cZFy%<67&mZ?7=;Qh0Rz<W;5$eG=Lco>YFe#Cj)FcgJrSQLJ
z2MjQnIC&T@52NH^v^?A?5BWbsMJ_HsaC~WAO|5$TPN5*2$B&4Lj){$o+kxTz6k#Ie
zp->)1$io<U7%vY6@{lLw3#8Xt8?jzdNg1o6iX#x!Na~BcG~C@-8hEe8RI;YFjss`K
z%2iIzE>;wC$0hcrhW~F?OHbdxh|Vx(ny}3*te4u@En7}QbuDdIQ?xw26A+Iru4@$5
GVg3snpAeh?

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_int8.cpp.4600C416992C0160.idx b/.cache/clangd/index/splitK_gemm_xdl_int8.cpp.4600C416992C0160.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dcba2b253b5da8943b1ada169f4638e9a3195e79
GIT binary patch
literal 3070
zcmZuy4LFl)AD?Hluwgc1zDCB37J1Y*B8`ZAFJEP8j0i)c(kq?Isq}?Vl-^0FqH`pP
zC>0Tfj+3ZR`Y<hh;7#h3_pMX!vz_a1>$=W!UHk2S-@o7g{`fz?|Hkw6_Wp=Np*-V!
zvqbTkF&GpIrH6bnGQ}IY$l#PV3Y8^k4IfP%<q#h9Pp#%U1rDZMSx!yqPA!;-J#9P7
zxO?Bn(!Y3ZNyGCUe^GCXC7jot6TJ4a9%or!xslFFO`)cc)!#%mA_bNuc<uQ1V!74S
z_784m>$UR?pJ#GjS59dLR`%I6wyGa`I3PVfCd&$%3BCE@4NI%;PIqVJddIo{3MF?^
z>ug>NSZ<S^Mb*!$8#A{!UasIj<5%zW^NEJq3cO7OxBa5@99nyJMfgRJl3PM4zYIrj
zvloR$IZPZIRCo3~)^AefRF{80!2A5jxBpeF{i7m%vLbKt)UGF*=iYU_+3(-J*S}g^
zeN%Gr;OdvP2i{t@Y-W_&j?0#DtlrvRsv~YX1luhNa_Jt4WOe(Ox7QcuM+~KO&E7lX
zaE3`tW>5P@29z)2_UbfX>E@#5qr6*RR;jOP$QKX13JmaWqS-=Y`KJv6F;nl&5=}6<
z5AKAsTz@^4osO@uoSNxa0!O~QM7Xw-UHYif$-L%Hoa^sWquv{f{VB;qr6SkSS5M}z
zHHs}w@97SlJ0OYolOFCljInbh*CocfxOkM#QgahpEGsThlC3Y^xaDSQk2M)uvRRDo
zCzSZ?bt8_t<mLW$BNCHIJL^?8ed}o>pV3O@%*x8{o`#EfC#22G6D7HoMrVtt!3T_{
zKe2yxIOqBGVXq^u>?TiCRsWfp>R{dEh!aNSo!jL!a8-MH;e=OU%2S!M^H5gEU4ul|
zy|<n-{hRmAv5p0WV``ELNR4`RKFkrgD`_A1&6oX-oqhMNYa8vO#f6=%%23-g)Rg5%
z-8mL2U9rxaT#)NDoqgndZN*YzLw?x_%W&zFypS8|Gw&Z}<|p{`wnoMNR133)!|hR`
zZH+BOIEmlg!5W@)$|8mpvzhUW`RyX7QO=2Sal*BC_GRCC9anBX`tzrQ+4J1^)jf~b
z^o=~zyc!g}_0T{oiZ;Hoa`pOTa@!g8um|75KQ<VDO_V%MaPfiyvm5G2h0oh7O39xe
zFMVS+(*j=$Z93dL#>Z?jS<R*-@85U5?oh|>$Wgi#+SYtbntCycGi6XnZ6DAv6?OPs
zIH9T4mA37S1JP^hr%7%<op$!!`8V|7<kyn2Rf7?=5@T~rA)jdTa$Q(o@R3VZjR7Ha
z_LkBwKH}iHJ@2>fZ%XbXv9H9teK@eS`NL>!+o1pOz-9i}@PcPvAyfRJs^;ndw$$s%
zOzj_pw0)nC#YGJ-{4%fBZdZfJTaV#9aq;c@M=s-w&*{pz|MnPZyH`4UF>2P}_v}h~
zaYwjJBVzlt9rqF>R-EZEN{R0HUCeIA`4*<HJ#RZPr@q3zmmD^9&V3|&W6mp&&o7zt
z*U&e->lBTvEq&zfc(1vz#E>`M*_9NO`={=Je@ddsi=9`l>+G4^Bp?<LYo48u4n#!N
zl<KC1?B<;8jYtcQnc>Q2stO9rLp{o6?xrbI&8UoRqF8$fg(5hHtj)<UrfH+1P$+?&
zf!%xsAip(=$DqX56?1gi&p0R)q0O=8x(Wca!>4Um+dle~<Dfu<reo;@1pwN$rvLb!
zD~Eo=K_LijhBebw0HFB;ovCeg-cN8)BtlcL6io#HI+33}e9EvuhJ%6;nu?|BDFD#v
zXmXzhHu@9};vuvR)`p+}K=19hZSrB1cHmLz2+hW@H535oVy?Fu|Nhrl929`irdU&5
z1pxYZ?hCq5r?UYEtwU%-tf8g?09}vyQ`EPl=PeH6BeXHrSWf|fzN$yYxovsO$3bfl
z+5&5#rvN~=UbeM$7T*4ZgVrK66U(G10MKnb%_N8H2rV2GgwU2)OKk-Jx~I0o>B?Wd
zo;WBBp&3{PK>>h%I%Xs6qnD52pw$RX!_u@B0O<b=Z%)rSqs_rV;RtPowbE7qpywu^
z7TZq<EOAgYLX)xNc?tluIDSKh8S?*w7H2PY?woLk@B*qn)dajLyue)F+yr<ASPRnn
z!BThuoCEJw14RI8L?yQu*a=Ztz_vgHe1?eRCQ}iG!qGqk+=htcR#Op?VJZ*>uuIYm
z7UmbJ;RVJVV^8oRffbQlxet}OGy;tTN*W))ZOy$<9gD^bO!ZBTfran_Gkr5-usdF0
zp>JUf8~`eVtSdLLa$R@<%mQZuPDVs>8>@(7;A9}$;<uu2K_n{}Ik=sNT_`ASZqDUh
z9b?iK`E<D5Ixr2~ucq~nu19u~TV7=+Etmv$0^Ub!{or}Lz>;7|0_MaEtO!;l;HKKF
ztLnj`litWF$U0;L;G(*u#UJPz`yV0G>;-HOa41l8jMfibgBQ?AbSik&c!3Sch6+l9
z7cfW+DzHANE^-ZdbE#b8B^sVZgNuM~qzLjR1G2Woyf}5Qlat7)46O|pg6o2tA&bg;
z4yX|$a$t9`RQve9GFY$X{E%FxGcy!SgM%Tt@`hB&O@*_72-pY_$=gyz#D#r;Xi}s0
zAmman_b0`gvIzJR6dK8u_pD0pDmVn>g0gF9{m`_?4Z76eo8LPQ89z{ihg8Qe)Zhu#
zaXkcIfWWedz5{}9sE!99cu;kG3k`RFFFQ&cpAsWKVTM?kaa|s4l*8Z8dJv#cNy<<m
z|9iMWIh&viQ<b4m8OAEZ`0pWFz?Bi#Nd=2|x}wyS<d|4t+=hh2jqyqAQ7PX=s5oVq
zs0>BQFj*N2MR6(ev#g5uM`P5n^E5PZcx{5Nj*s3754xU~57CgMPch;+FLiZW=DysX
zyugLVVwnHatbvKCnZ<ldD{C8DJEp@TN2euh>SE)CuqWBT+cypJir}tP^F#d~avKDs

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitK_gemm_xdl_lds_direct_load_fp16.cpp.2D633E2EB00CCD78.idx b/.cache/clangd/index/splitK_gemm_xdl_lds_direct_load_fp16.cpp.2D633E2EB00CCD78.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b8f7f1545495dd497bf96bc990b162660676117
GIT binary patch
literal 3078
zcmZvc2~bl>9>&8XJ|G7qkbo#hI1&gD<T3(gAA*X=A#x*wKuD0oaI2t*awxKZoU$s5
zq9R5|M06Cq$m)s+4$A3pGl&->UO0~EjN*2()ds63Rr%`w-tVjKe|LZVeAl|V&Jl6A
zmBO{@@ljj&cpMH#1^?77DY2g5;SCvwORsJT+>kh8Px+;0NSw2&`sC2IY~En;*rNOV
zlk9hH-TiKco@F6=*q_sBr;XgJdBTb-Sq2rqm)ZWR^Tf))*q~K-*oMa(EtuqWHHDX?
zyI8WE2H)?nlT=@v5v4boe0(Rwet5|_#93E7i;os(<cQ6l4W#)bWm@Zg{WC8&+(|4-
z&af5K-ub*UA8C_xX^D~@uKjy#H0(kT-8~GEc(ce>;=Sx_&0_Ir$I^41sKw)|vqvh5
z34VcXGj&K4a@M&gs?@B`mq_<bzdhId#y=(Y>ewHP64t4wbXER$NrR<xhEDdnC|6o|
zPD*mzxeGJ9$=jY2&-`vbw4<=&dRGx=`>65d_pRCYlIqDDCOzZ4lXPQD_9wJX_KDAj
zyBP;24p)>Cd>dM59HCv+&PziRG~Y>XZ(+{K(wSo+OWFRK_*rE!gVsFy7#ql8MNr(L
zykbA+Iu07HS>kb1_rVC`#)XF5?&a@K9r!DCa_3N0(dCylWu03Yp^ID0y^>D$j2OAc
zF__JVO2R%at~c16!1S(L;XiR;gx9*|(jNEUbkL(hb$|EBdS;ONve%EMPAuCM<ZM1L
zrJZN2ZdlTKs({b189)_>+s%%i`ZDNHo_xkJdx6rn{Z+z(lg7P`x)Zq%r=s26J=J^*
zwRKe&_tV=V=Q@ViGkLAGAs#)if4b6gp#NCVu*hyG>1=Hp`ITxaPw^sY1Ntxg-PEb9
ziP9L0=!QPl<VL%i`nT%?3FfD$30ddchwt{O=l1_e7v$M0eMP$wkM{Nuk7-2n+_{-5
zp6ZRtt4{-{0r>q7`oCsMdUqsb7aE4R&}V4l%Zk)<_SSHo`F(CPzTC`wT_P^BsHB?j
zaU2(R+#yV>$FeuL@7{HQQc-$O>$H9)|AA*?{;A%heG?+yT(p?xz3;(qZ?`1b-Tawm
zo2F;`A<d<!V<f?5Fhz)e|HL-6vV2bI=i?S%n=c-G*u6&3<VE*8BbcHpcSWWL2a>zc
z6F;gKJas-tWq#t>YwPsX^<OHjq;yQ>WbD%NbEq=Q-<#yeto4xC-ZC*V(3?nhdZAT3
zvbRDhsW-Pm^~R+6p-0)tNAzC?wEDb_Q&VCcURW7=GlcLl@9gv8?Zr`tTO|&+AJg}0
zy;2F8(3Y6jt86bDt@3(j+fh31bYR!|Fj2R6sOTWMv3X}>T~N@?#rK=WN(<<7=$(D0
z1sl79H<`saQtzK#(dU?YX<KUm<Hhphd0(>v{&_t9Zl>vUcf+BGhMR@7t*`HVD7|sa
zVbz$^h?~-v^4Z|=jAyIRBkf32VS43H(bf2r7mpU6P)t@1i|kMHr!fLaH9<;EeXN!}
zy@O-f@$|M&+)KM3sIxroTIh|kc6Ar0cz5>4i;rEUd~CPAmu+%eAX?Qn;5F+!7JgUo
z*{2oH-RS>Cm0S{*<+=9xueF57`D%IZbj|Fr1`EmXg%K9>pyMYs%|*(uQn$xPdLcNR
zX8Bqk{t%W!MEr3$z8Pw!CKK@0k^Pp;Zoe-1;C7gZtOM8zwW7!bV&0EkNB3Q;I738y
z05(NU)no$k>izF#;u<~&6Oj;r^-+BlnLr%cb2~xO;5tS`f&n%{jZ|a;aZG5&K&{sP
zCq!gDz-*LFl?lX2iu6upG_010_yTN=T2o{Kaef#3vYTn;EfOvnU?04XvP>W@b9Pk-
z?VpPzB3=NSpeAZEfw<gPeE!2!vw(>F05Ah(sK^B3quEbc0@d3WiAWg0hNz*6OdxK=
z{}JE0wEYbc*$6NbWm07Vamy7p+d*(|mWXTs*dDcCC=-Y!zAAA`GJ*(1#2a7>)PgJ%
zh}-LL*<Sn7v66@c0BnQWP-Ft}_=Jg|lXZB4i1-1_K{;fZK>T`OYf`3|Y)?c20k%Xf
z$ufcX<Meo$<+BJwA`%X;E~={}6NtZ1l(d^a#nzE=p#XC=xdt+U7?%>YDfKT!d8!b?
zcWEN=F$Ttl%ZJ$lCTXzXAi(60g+UWc7%)lWHP1xH^q>jm3{UtjUnIT>%|r{HG>LCY
zGu49Qk@#jbGcA~YuVv+lnw+9i1rncW&s+)5FX;IB$wZArOYqv>!9EgB2XhO0l7?2^
z6Nkdlfcp>A3ry1J%1xL##)Tun3<Hxi!1GLeY%?@%^H|ks9?V@2y5PES{oxtxd%If4
zp{r;Sc&$&-*MxfjbFV=7+q{FBq+y>o69LnJGr`mYlQi;j6Pw1?f(wG#S0sFwc@p1(
zVxa-EOX6EnEHz-*>eH_;SRa4F6|}0Ws;db@)(~g+o~2wc3|{+~`?$c);qHUelwKWq
z7gicpdN4F3zO{z69^5|?-$uhm58e#89Iyq_%LNCJ*!c<<tUfUfrZZ$3a^N~(%V5#c
zy9V7LhW3~f910E%Owt=S&y;|rLlZoDV3OXuc_wGf4Voq%XRRosHjaQUxPjaVc$(A7
z^(Ba7qcnj=Od~58F}Q<ZNa<ajH>5k}1BZnBi4wlMtY8HX6|i3vpbb1fRluIje-<Oy
zX=EO6MzAaMpQ{zIBZ`p2RzWY)n?J9YOf3J~sNfqQ;!9a-iXinr(jrbO|Nlk60*Biy
zFC*opP+mrVD;4<_E%?luZ0%09gv5AJq(B%I9TOY3X>$ZF=9>wZC@%%_a+AD_l9%!F
zGFDzD$V-vD6vPWf(q7!7JQeW^P$gv*B8f~<Q+1=Ra$!*kZt7YZw1wLC4$FD}z?@cC
z(#;)>xTgAl_f~VUjxNK%(3ok$HnUh_Wo^53nV#Juj?GHCrt8`yWOb198U+vB{{S_a
B2g?8e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp16.cpp.3597E35811A1130C.idx b/.cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp16.cpp.3597E35811A1130C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d18dbd1b76943382064a67fdd22034d7d7ee4a33
GIT binary patch
literal 11556
zcmai)3tWxK`@qjU=crTX?R3AKI^A?H$|X)E5=ByqwpAgzAjG<FFKg+hmLiv^mRL%N
z5bL_DV%gP3aw(}O<?=(xr4Z^j=P<qF^ZS4Pd_Lx!Z_muU&-={GGtWG8CXE|2M%tgp
z8y!6^ZvM=r5dt2Mr-}bBS-N;mIffNt9xrakrOBZ|7BY=ok1YlL*#R2G<8{?XgzB$f
z6fxBOVDy36V`ds_kBe&W?mF@4Xj!Zl)cR-6%WFGw`)SZ-?NM5zVz=4-D~Q;jug`YA
z`%kJ>wENjTTi(xjH*-|V&F)&%+&p;2Sr?5g#|azM4Q3Z8r#-fx_4?dDSsQIG9@*~l
z{_KwHms+oG73&>eKVp#WC-*7WFKsVtc=M<;M_jb&_3*~mF}HZfPaHM8aba-TAN6_p
zt9WCxBj!HK4>)r>zaut(O;MR7@IMDP`{vzP7_1U^vFp^4?-oDIs%slkzr(U6*DiVb
z*6JB>tyGw;dFlJzKUwZP?zq8dVO^llmaz{8y7=8&HEZXrMVp)lX{g&@Jaa5Md8>2$
zZmEIO;W>9cUQ7uv9p`)~d}*n<m2g%<OaExK1@EJed20z?st;;U5dIUAA2tMhm3Q<>
z>a*~YRCM~<sojSM2=kVHUw1S}d1FjehT+0w^`Zx5ZudQ(7_Zp-(A@v0>#f_4ZM!J4
zSQ4Z;>rn07DeR3W4HmwJW;w1Ef9Jke+hJw&$WnW=Mfs*^$0Lz$VTmW+nD7GzsrY6X
z%)7GnuViJHvj(wuwq1XvwqmXMFG`ziVy9dlppk8~9?g%l(vI|an;Y$Sa6?mRV%=_Q
zyX&^U%sMR=W)6N*U}ktTao?r1&JSu&Mb8HGv^?J`7#~?!J+=O=cm4V8x5^KtuNu%2
z@2dUhyo$B~m4nO*?f)&^RDN9Qw)<5~OuyN)ytH3B=2Tk3lSvVmoGX-P%qvzoHYGA6
ztR%7`<5m0EK5L^?LPO`xnO3@I*^qFfjawR{+d@LUbTd6?KV7uTY2r<>fzyKt`};0$
zS`{^M)Y10imtqs{XXSOC_^adCzyJ7#o*%RS_^*cyrbkWN6Y4ux;<!O8QhU-`=-4}`
zBKxp(&#1HOPrPzG*;6}bT|>mH6RYnId)gGSF8f@n<U@b+jaTX>42p}6(abBbxwYMQ
z?U1XU^UiLr2-<DsGA&|j&*8SobNqV@_iP+hb?Eu7j<`^3yXpx~`_;76+#C18cG|ct
zdc`g75@x$q<A7Q54VRXxTdPJMtKWKM&FkS)Qlqcmu3GrFNn8Hd<}2C{at=qVGj16c
z(0w`m&%;@THC?xkUx^8w@yX#%=-_*m*1u1Qy5l33J*;?FdNg7BfTE&DucM#;_{{oT
z=pShXcaCq58GHDd)+oWYkd`>*ii*pZqhcRz4-UONLjPQQ#m5&RZIMaiG7paUML(|Z
zJfW>=bFp@sSHY7jKWWx6p|5>6=AV!KyKwm*3;t${OP4$?T6nU2eT!bZS*pKVvGsEA
z`pj^Xo6n-$JoZigv|7i-V|B96?DvI9Q-0VJJ9G2Q3%1T3p7Y0TDj$A3w`jQk_mvsJ
zUVe!SQRRS`$>uwvL&6`=>FQZtZIm9f%gk<i^O4CPZPq;qy6DDeJJ+nv*&b~FG$CM~
zhx@igt2Aob1MUPz?yh6?=XK|WIfpMBWTLibgWX&IVLd(<?}j}c(`7dNx%a)}ugjVT
zg<U^((5&k1{Blv_ymjj3rhfz&EmFICwpIOJ;n5M-t#5h!U=urDJ81dQ*{iPT&bj7#
z{?wsY8a_P+M}`|c{jj>=!}ezL?n!^S-0L}AIn3FeO`Bx3DsH3wQjPnLUB>oI?lp1G
z$i{};7;FDmHos4B&RfFHk6hM=$5ZY*?uTb?H;Y6tX(5lDE1s*TD6r2{{1X>2HCtRJ
zztw=*3wZ1aH1-4v{qjMo=uJ0cpK10j5Wz$qkCiA%R1`(8_{U34`OBJvwO|(Rp%`ft
zBcm@Lq>7}$r+V0FUb_gU;2s7k4Kh>|L~NFjQ*lSH$y^LS{=j3S&7z$Y1ra+eTlV77
zoAE{>2*$1LmF%??1ra;FWzU`O^vM=M02bRS*=i^XB6hZSII-vQZ!47`5~~*@iqTRO
zMC^XBGB@SymIERP!mXuBQguZ^#NO$zUd}E07$SmjEY?-hRZ$c~Jmz9jyyxwj!)m-x
zEOyd!k}3)!9;+Ik@o?f_8$~b~i%pbF^b`dVPgw2yYqnmBlLka!)y^P|GsyVM2dNUU
zRrAe#W$$xJFdz5etL$r|D2O;%sp+)Wf)9)uZw40k)9YuVD2RCStrYuQ@sAB+h{J}R
zVKl>9Q4n#2aC3G8nzmI8%dj}gB+5Zi5OI{)B+?}5)iO2S9Nf9Jv9+b5=oN=EtxbRI
z*{A`F@WhH0$7(7HB9585vSzP-dW#4Gacg5GV@*Xt#0yMM<;0Io`%R4(gT)rc7M6;F
zh!^s8D}+kndqwaa7TYP=X($RJUNSZFw7zM{Z8hFpEVeYZv{V#CoLXu3yN_*V8N*9K
zSbiA^E+gTW4^n57=NL5gR(GTbCg3VYN=E96qE~!=sQQXu?*F9(lknL1i~O|}1rcwa
zRB)v8$*Cw2OvkPDl=M^-1rcw{DR5GD|MQp#!m!v)$xTmD5OHqZ{LHTEjN@7`6L(&K
zGzyUP%Ll2l_f)g_<JWKSqMM3)FjF$qP!vSGzxbtQs<&g67DVG7b|AwYXvmijQsr>t
zSoiq!V}%+Jg{$mADtqulK7WuZM+F`8D~9~@LIkn+AqGkYnu>ymi+!{vHMWdCD1s0y
zHdQiJR}@5i<=0S?2<^;WdN2u3p8#QihN2+i8@WHcwMk#J;Pd>*MB+@O$$tJIRq9Ko
zY{>GOR3e7uxQFTH)B7n3B5o*;?#Z|}NhpRTSR82*>7Xcx_};jy=UlH{JSl={SUqzk
zb9F^Q#Lt%9OAhZkYb1uHxOKRBct1r!#Bcrc!kuikUQ^@E!>z51tt=G<5x;wOZ<FJr
z8GS{t7>l))w3QVF5x@V?-7;5W*hDp+KNdUcIZ71;5%U($oVCQ-oX0~i{3eVn^ZE}6
z*wyB%`$m0PGXgdX>1QDWVk7=Q(ScnRk!eDPT?dZqU~umYX4iw`dYaR~F^%SQa7zad
zuKr$d+e>pExaHBD4{rH17l3mCxRUltN8g65R-bq18~Q3;D|z<Z{-`RSVYh(I7H}c$
z8Fnk!Y^6C9Y%)0}yAfYiu?rFB%&@Trv1Y_N47QF#78*>fBdRLC^6uu@4TxcXLCi16
znp9MqZ77>lC@sa>mV;I~=nypoY#WH%K!Y&Z<&f%^y^Z))jR_^@^PeYL;1}5mHaj^*
z88#Pea%tWLHoG|1IAbz@aHdbCirfO)Tfmw;N^Sw|tu$wXb|%ePpq)i?HfU$loCDf9
zH17oMoiyiyb}r4kKzkR*y0-`Ke&aBt6&P$2gL1Iq6l(*WHZUhlwsmY1`JPUuuHKjJ
z98|w-;>1)8o6NUN<~x!023H5JxU%2#79N#$&};`IA|Sgx?m@oQla3kIP2$#{ct*14
zV>QFXVaCLls48~;O<$a^o$2Z133@Z$2^jVk*xjP1eJ$A4(tHo>?$O)~cFi1<5CTs-
zdHlesBr8O+Q;2BHu*G0l%-NEJ76R3;LJEeJZAP*kRI{UCcoZy&0T^}--+K)|fH0Zm
zxEp!=!5Op|*%Tv3VsVB&iOfzS8!}ntY<Ch3Ane>8c&Xa_z8Ax=A5h>26h!)G*pDdi
zBhB3?u$$%{6xc&^JOsuA)jk0N6KGC?z$BWJAuyTd6bMYAITZp^Y3_jUI$$!VUnhLm
z$+2wQj$Ktd{+^0wP8#1djo-U=+%zui{5|`}ZoC@bAmJM%=6dKr!Va1{k+73vH<Op;
z^OPc9stVZih&hkMB(xG>e=1(bIRXLpr{eWAr-3+)=5!FJ(_9X=<zP<~5wNWwZl&d|
zVBHF~w0u}A45vAy6~YMjhMTHig_#Wd35h<TUjHR`BCcCu6lun=nc$j9b2hkU)0_jY
zIW*^jYc9>Zz;zeRdC)fx9Ju};!RSZe$MGld{sa>Ud$iteE83sC91qL^&^!PVB4F5q
zpm~tyLeMPac+?O3`<r;2`PP0iYWM2CFSq9g;Te&Ej5E+c;^Dn^rc|zKzO)VtWJ1{x
zvQWLww)vfAoD-D<Y!}k&LS{rC0XC%GCz{uQ-Ws6fiJ+GV{kZyTL2oTMa{LC&-+(h=
z6922JeiipKtdYb>N<v76l}e;iGPDe9A~BJY(4S#VC8kmmNinRs#9T@O7yM4)vIz1{
zm}bgiD2kJ<q-Z^oZRqtlWBy2Pi+w(8m%7{)NIoD_qOaT)NIufsjU?SP_aI3R&G8_K
z2daGnND^pH0!b3h$skFlIRzvsG^gUhr}-hMJOp)4zlUJ{5NtSp1S*ecc?TGGfCVS-
z1mjMQ$$<;+E%MVA<;Y(m+1p<G5bI-qFF$p0)!u;m8(O<=@a~3S^0cv8-Yf6lz8i*T
z+dm-s2TX{-H@jrxwZ!LYSYT#s)|Uk5{<6(K==D#_z{|qG#b7iEry17P)z*t_9t>;e
zYUf1`H2#;bPrTi~{U3pV-3#KqU`QSzVDmtnM{_=i^J(4(;(Z(w4~Mt!&nHK+t&mUr
z?(*{@_wI8gb6rXN#jx+d@*Q}Qrw6Hc|2coAhCIYNgH+BSb<+KJ6SQoeY;y0T$FQ|X
zr4~^mc^9eNrMVud)YII6R2pb*L@JFmHzAcKnwybIGtDhXrG@4XNbLjCCK}6Uy4pvY
zyOCNq%{@r1hvs-tiwCNG0;na>oCIn~G$(^vGR-NVmO^tXsHM_e0?sAi#_3lI&ZRVW
zfMy5ibMj8m?Bw{n`f%Us96yh5MBe1Yj~9UaxX+EQtOUva?p*-iM`VT``mN3lkLLxX
zeF5o_@gyk$9Cu9?PnK|TmpnT_;;w0@Tw|wNTgszi|6%@P$)E8qgob9KQ6w&B@Nv>}
z5Ar7`&FKdYoGhH=Iva;;t;n<$QLDHOnYPjV0-3&`xgD9db4(HwSjkXvs1c{6JV!ya
zA?XPmRmpP`lqG^>VVvV^g5V~Y%31vbcz@tFI<n7TbILOqBvQl2Fuu+@e(zw-+|b8r
zjn<M;*qpurzEj8&BRLM-TAtwGROkWk9vl*VNkhz8ykvR6yMmF}E_*?+7f&0KpTKKg
zkVkVq2=Zy(2ZDVZ$4<(>YxYu5_Ki6byCu@btI79AY`zuJ6}IFZ8FrO)RWFeQ4Bpw~
z*%wZCnTyPWGv5vHya7I(!qwneO>+%+*3euBo^>?WgJ(U>jo{fxa}#(r(fkg)-*Ll3
zye1xFc@BtDau<>v=$-8hdkhSZ(E)fo-#eb)8-Op08&N4Ozwt>Zz)sA#7CDjk5wLfW
z@m-qhk#RlE4am5G=0;@PNOKc1Zlbvv88_41f{a@@Ui5pdf2T>}&TkCh;v)T!kg}n#
z@r0=YNfo#KF>EzRs%fqRNgd6NAZg^7<jI7pU**SesNpH|;Rd+|#@E0kVibn0gYk7V
z*TeXFnj2w!Bh5`PzKQ1d;QgK(#O33?DjY(F2I9rj&!gW&;=$rBHnpq}R9(eR<_y@K
z;bJg`Edjd{noGg1l;ikSr<2MWo6gAl3+fj}`b#+MH$Ql1)!*1?r(obI7|FfQX&88#
z<}=`T2BvWGvoP>1Cr?T$^Zt2LWC8B~B<P=nzNF&X&{~bY{@wmqa0>KKab`|w_m;X=
zrWs;^u7<822^vW|yhVRbL6Z5q-j|=$y0OQ_oMA0=Eu>^AGOVSprId^r!&>QDNlBWD
zVXbworNrkk>{TSZN^g$WknkGK*OBl#%~eQPMe|K0yh-ydB)mm)H4;|ST#IySQST9-
z!RtonF3t5wr=I2pq|-oiBhqQ4xe4hs(cFx5nrUu9IxRGpfJX_8Ar@uWQt&9HxeA0;
zbQpOPY;QvEI+y1UTyD{6uWArh)AjFw?H#&)Ex6Ru_3J=bN7ruv+XlM+J#e{4*KY)2
zBVGRi*gl}^H-k$vS3mtp@uDElXB+=tk4=sncrnY<)nqa;_y{EGqR)<7z;z3K*2@Cd
zESh(M>rR^YLf^e#p9L)*aYrN^Xo2@f?kq_5U+k0RX>qa_GK%Cc5i;!DVRJ`w(Pj}+
zFGBhx?@hAjI7F4F&bc=+mKaOPriz`^0@)(sq%xcAOGlp1dy1FXIb?7SS(1Pz^X-gj
z3F^|@*o|yM2HTJ&7oBWF?%R+L$Ccnv2{JN=NG1wz-tq)Jr^N4|^E<bfY`fx9aC7a?
z{qTcNBAt`Sl%!+G%i#8(Gx(e(QutOW{DI^_yARI3aI0)t6&_rPi)0wdyD_Z2tGzqP
zOznPdA9zv!?FIapBG4=X1Je4p*zp42iEAGy3)m9Glzg4qa9#&PNkw^T!+AZ;Y2chj
zb2^Mn=f+*0GxaOy##Np(wQS>(UwdBY+O<5j=u(oO`K=CcIS~1o-|8UEg<w@kb2|)a
zrz77C{_qSwmCVTCPs!lVAc~XI6gH&%o`V|**(zB|?;f8wy3Tmig|1)!XB_fUL$V((
zJ1@l(APa1=xD+kJW`j*O%{gF`!!fx-!S4C<1qvC!{p~)#g(n`kiyxJNbTg0<iJ%U}
zo$C`gf6o{!&=za!kvoFJeSB(O`pj{~1ME25aU!`JU|6T&P7}$A_3&V~z)L%4Y{!*i
z++w`BNb>}8JAu5(tRk^AR_gOz4N>Zd-J{?sFDLfn%TrK>tso`llk>n<7)-NoE4TfU
zQz>3-@*5t`ya{~o1b%Oj%CM>Y;8bo@j@}>fW41v+j3Q_qxtGrF<-y22n)6{~KFtN-
zRsjA)pJP++4BC^Xu7&5(I#5~1MY9Z^uqx|mP6L%Rn$tlgo#t{-D+f)|F+OF0N-HNP
zNtQmUUtLq-BfEuJKQhU1ZWcsqxSIlo&4<4EwD)@q7LPgacXDWP*2=2qK1hK5j`lh3
zAa#07W?a$8$9M2Sz7<)vA}R&ghAi7?et|4s(A<tJ+c`dcY3b0=9kv}h@?4Jd2h@8P
zBR}Cff26q^Id{|CgPeP4jtA#>?k<O66Tmrv<|J@VqB$9ylW9%?=M<V#aUzW7Dv(uy
zJ7@nJAiF_xHOQ)It^rvM&2=EFqq!br^)xqvtdZs>kTucV0WKX}Zk1s>!KIVqvg-W7
zyldy@d?UEN@c3}p<pXEK7y-7o@B<Q&_Yq)w3qR7_jfCAa_aI>p&G8_N2daGn2oq>d
z0$~!($skOoIR%6%G^c_vmF5m$I=CQ0fZvzt<oM$Lzy)e!gh~J75y^QO`w;o9EZLLG
zjp81t#EJW0`;>rR350VmUJ8Dt99KL(T=x9hO+VbRV~}Gg@n_d8R;DJ52vS$Z0Xott
zK?Y<CmM?7gbzA@~U)b*JX-)(8G@8@FC!O<9@|&vf${~d4!{D8$UmFY{T;;Rz-}S}+
z8hvA6$wrS~`|~w3@v+q==law-m{rLAIo~jY-<y}Xt5?#!!ozF@_HhTm;Q$?<9R!Di
zG#7$HA<gYDteuPB>XIMUoNw+^jGyop3Ev`ys7`LUvDg1z{5_R!kzaO`fg$(YxKfBX
zgi3eFue?bmk~GHNO@7tQx$#60CDQ(EEr{0Aybt>B``W*ae$3gvDST<8%cxDFZ`k(X
zfIe@O)%Dlo7d(M9Paq?*Uy)mOtb4FHn0i6^g*(yx*_NWtoeq`qbajA!K<{NA!v^XH
zQZ!L>qCT0s3>z*E*CD+#>~!&T9b!y|ogthdCYdY-hhG*2$d1Hg47(p$?5FtvGCqJT
zIC&QmbRktTMDn{A!6%y6fM5+!@<b3M(!3S~YiWK1!Z*-+SJj^BR;!X~DnIU4f^j8x
z_dt$vIB=H#+<+TFnXhajiE|mY9)j!XgjE-KcX8uIVh(%=kpC!w>f{2*E&|t$>?X)=
z(cBEO7MhFrvLk$|ehFVzN^>P&cAe%Dfvi+O>Jy)Ug9`bNA1F1dkgTRRXu~ml6Oujd
z<?e`2eo|Gm0^7zZ$SI7(MvG_0M$E!}cpVZw(0Q?XG9pWP{7TdF@mD<;Arcmm%p{LY
zEs~kzkvM6V!OWal-fvS7@lGW0;t_YSWO3Az0=dTV_{Txy-wN`D$NQQ4x|;i%OjTal
z`)58I{YTlw*81f|OcF{)i6{ZBSozbcpI5Izxc5Y=&8pruyq~zQYp5DY)b|AHd(`~s
zSo|X?p6^UQzMzkivQS0DsA;H+eKfsBS!=5LXz6R~7)V@Phq(_Q;W5%t*TKzVsIBS$
t^;bP3sj-=PUrQ?+JNtf40|pKrA~SI|>_2F<uHKk&YY}*V?>QEY|1Zh3?N9&!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp32.cpp.8A4C951A87984080.idx b/.cache/clangd/index/splitk_gemm_bias_e_permute_xdl_fp32.cpp.8A4C951A87984080.idx
new file mode 100755
index 0000000000000000000000000000000000000000..79ce2caa39e81bd6c1fa560865ad26b99a7542d7
GIT binary patch
literal 11558
zcmai)30#a@*udvJGis)JXWG-Wr_v&28K#m;6iJa?mMGcd+HbGr(jH2YC=^#E*>&yi
zRavhtlBJ}glwAoGH`I4#IK9X3`+gt4pEL9CIp@6Rea<=0d7kIY=wU;Kn04pz28Rp_
zTQGB3kbuYIY2*J(mo1rFgkhzW#|ztaWkSFhD}`3ZfaIL+>@mxb?G9$E^PCK7p|AUa
zklfirW}4~_o6*?Tdg9r!!cZM3_e+_d)o}Fwt1(-22k8t7-Ddx<AZSzP=WNTTmvJ^B
zy)I<ky|Sojrq9qn+g_sD+TJTqxof3%PgGZrznjdzx>Wzem8l_)UGJ8ZiFd93JN?en
zo6Ygk?H@k;(#!6v`?Nbxx7Sp?X??aqdNk=n|LPB6mAnrpjv3v#*t_tr%B<`)yrF49
z^Im6<JbOR8IW&9SkwTgO%L7|{vhFMzCkngNdiv<tB~Mc;8v0Bzx2?;tkDj)*bUNHB
z5T<Eg8NX+>b^4!9n~WD#_<JW0ebUp#_tBbJ>9ZDZmiN-qbhvc(cu4eCdH5a^L+8VD
zAAGqKJJNiZ{7~Ss0!tg=tcbeqA(Dmvg&g<N5xmpv)fgfCXJYoGKH#Iit5Z~`MORG3
zXKtO|bGU~vYuWgUV`J14=gdelS`@8$<Vm61V~-c6D>I*3`mMfQzwP+8OJb{~W3*=-
zDxWu*z4M~V%E!oJhwF`N8UIOk*%&{w*4<)Nv^m7-Xt3L)$P-Ow{E@vxK1qi2uW!8^
zt?qKpF!aH;+wUbSV=ULJZMF@ae6@#Gn(;=oAk0QLc)-Vu5Z?ovY6>DN_H?zsZMS~b
z8L2R(_lq10qq~v&uEe)|mb?<b?$KV?6eSoQe6Vy%<wvi|3)}A%9ZFc!qb}T4_wV^P
z8hRA>vN-7QZ^7oGKTX{Byq`12arP`v-FHqqiml<r=%6d|8|u^N=ZTI_4o;esAABR}
zedEwhF*8H~0rTfhE!exfPoQyPa&<_<!~jqI6pz`j7VmZ*bysTW{A9%bE-Px*%osK3
zSmU2hS8sTnn$>dRa`W+j|M3a9Fl7IqI}aI7n=y57fX_VHtwfz*-O(}7oZ0I}+F_Hu
zgU)?8@!siF`^#Y)s)F91SbL}6tD2w<Y5fvqpSxQoUauI@D=cJ=c9!k6d)s|t`rP!G
ze{Rc-F?(!WrUq?oKin{3u3x*+-o!yAhuqDY!UDS5myUSlSXNi|Xz^RSsl$>D^6K1W
zn-&f)2)YR8H@-9*p?^8}tW&b(^Hry`t{*#?S+c4B!{LdxcWp20Prk(T`{%ceFA2u0
zCik3hv^jTn^1B7$%}1xtz8NxLXX!$}=~k{&wlA&aU()!!@XTLR9Jj20e0XQWtarUK
zf{Fw8n4UZ$7HXfl(&bve=?nYzHJh<??c|i1Z<Zbx7gx4#dg^pFb!_gd<goR&c`@tW
z_jKx4=)G)1#*OXAZzi4G-{kpf<BGpRuN_=5#d^4VPJ_nmgMWE1<Y`7c$eJ-fc*5nW
zFR!c%l)MNY=={0eZqbSsjm@{-PY|7c<XK*1eOi=x`=0tq8!v~CV^^JPkq7i2JYKt}
z<YFiHOM9k9FWM5srf0uwuL$^6zIRPT*6M`(z<0atHjI?K?%{iTO8eZv@wxlOza0MB
zIPq$;$COJ0f+nmFU-*3Lq5qQSm4w~D9wv`-+gGWm|6(PXKR?&_Wn|N_&TR25^~8VY
zcw9RW;raNqz4N-G_P^MzZTw}y<vw#>pBfN$^kVD4o&N2Ar623==UZNw(QGpz%k}xE
zM|nkqExw|_RekS$Y#;H_Ia0fA^ktVv?PrSn$=%ub(Kc(s5*?OlJ$7m}bzm}XN!tfj
zS7prU>i6Dut&V)xQg%V`@=iRSdY56pzIMBNL=2-B@z{CNc?PNi+dS1j@!piXzqV^%
zknra7c&v@7jkT&EVrdRjmh3X&qZZ6w$YW2Sp(jwl_YYFVV44N{TDwb*7)Ie9WNI>z
zs^}2^cBLkJdF?nIn1y@DLt1&r`1=Q`BCGN)?RSRPD2B<nhhA#Ej8p{?TSV-*@xY+Q
zQVPHQ%40(;LY!3v5qDd@{Oz-*;l^SZhg&<SIq0YgB6j}Bp1;uIohF8nSZt?er==>0
zSnkm6#NMlatWtwuJiIyLIXbF>h}{nqXT+XM&K1KL+}cFVL{n7|vDfTLI-%@r2MNy)
zi=7OdOjJdOIN|-fc}KoX6hj~u>#OODR0R<a)d){|I_h$w7$#t`nVOk_svzPKYkhX6
z8N@nkK@c9=S)_FqnSTEuRYq>rZhEZlbzTh?;2wO`eQZ?)5l^@m>yRP+tx5`ESaZ{j
zr*~BqL>wgCl2(PLZk57vES_ODqnoNA;u(QVea&Bc6SZJ5UOA!CP;FH~#B-*sD$DGg
zP$vd|{A^P-Q*Bj2#Eba)H-u_|nPM1=#rA6UTB?GG<BIK1dfTNGGQ3!X<yVp5DiVJG
zAayo-3_(-wwFQe|1g>JNW~`|yI>Z<HYOY-W__7*|#`^IS`{}9*BHl7O=V;4|(=)^{
z4YxK>GZ3i?BHp$m$64L|@8e>agvD-ZZU(A?h|_EX=G$1^$(Qf~u-Mta*+f+k@s3&Y
zT3h*=M-tvVEVeeawpJBHoKdkLrL{EaPaT+vhn9o1a*)aQ4^k!bbgkv{4^4R6Ou;=^
zs99*K3L@U0_f9*`%c(>MLU0eekkKyG=lch#a=3b^dw9a}gIX{HSJ{h1d+|fQeUK{0
z1kDR>^!ewl7(($w4Al&^RRs~}dFza>t{Z$n3=^@~T+LilRS@y1W#<p<k>n;zcyqAW
z%GAnQRS@wtn>!AV=RI?g@aAH1SJST6s)C5G?+h>t(oNZI0Hg8h8Yvv9r7DQ{PR6ev
zZ4(wR{I<GMkTeBpv)?{QmCF3dn^HYT=SyJ)?qQncG)GlI#8pKh?MaVD3#G6Wi-XOA
zyQvByel+andDmN)PKjYE9-gI|rKYMN;@8U`MF+N?GnT?K+&a)Q&{0(o@!N%Sd588d
z%a!n^<JOJ_j%KQYh(Eo4wAty|^e$ppg2lROy6UQeh<Qt9&RS|`$>X87z9R+}dcFh!
zyVi1T7qX=UY%1!UiVTSo_&vqByKe-?3mJ9;IBkI59ovK52u>Smjt8fBniIe+0S0jO
zGr=vB<}7f_qB$GfvT4o%c@DUe_G-sIPF$-w|Ikmyt9C2@^&5v{8hnOL2HRwCA?+D<
zE7)$OIR$J}I3}kNUqf{Y5$nsap@yLrL^}-DPPbIln`noeQ38!0&L)Oc_$Wq^!-HWf
zVO%9l;YQL5UajCuSX`2K{nOoZn-IgUN6dQEl{`%{+ek6@ph*FK<s#520zJ|~z&3!i
z0kjB{vk+-~KLLsL)f$m+x!_Ht71l;N*rs!D%di<>n?dt#u-(nE)>*R!y;Ho4MM^d5
zCPP>9D5V;8x6+&fx+yfLf^I6!X`q`%^A6D6L329jrqi4Ox*0U@2Ho8p>)-Far>R??
zdSI|l42!^q8(0JAHGm~ya`<DNC{KQ}{tdn#`k?Xss3)po*l50WG~b!DH@w+%<@Nm@
z_wcMVf_5Vq69Ji4+=J2~ARRNTo6M~{vE9Toz{8v*on%TZj)uxV*!=x?us1&wJ;Gr6
zCjrCW1N(dQ-YEzBa+)83{Ue%d!M>Jb;%MNttMoFsAt^2)MKK~HX4pJ1%HwoN91y&E
zmEH%{>;O`npqd>6qhnxA6u_|S_+IPyBMFnOg1b?ADV##{kZm4vA{uAdQ^?{JvL%aE
z$@ZsE55n?B|0|`Ik3AWN{fzuSqcNm^hW&#4ztG%<{M%@5NB-?Jhl77OQ0*hYKZ52c
z@Q<Q78vLVajs^c%n&ZGfj^<_<+YA%9@wLF%7LFCecI__Nb!`e>Iq`h+cz(y;anrh(
zelqR1HoO~~kgy3!xgMI4u$kr-By8c>&Fo##e6^r=8UpqLVlE&lali!FC`&hRhCqOg
zvUDTO@gR+-IRT^zG#7zg5jc>62-td%*3<HO=vohUw7g$E^rv}ZJxn4@oKi;P2WJ%9
z!`X^O+%o)%#9vW|HB*`t*R3#!bi%MH;F>~n8n~v>yaQZ!(3}CT88q((*WEN{L6<D(
z#`XUU20!EMFT;KXudgtI$OqKlZ#c3)V+Ee3T+q%184)n-0nk1`^Fh!)$nl_G_jflN
zaQ0_M)Sx|UyS&?;F$S-XBxIU|dJ=o@nI2obruNDPEKmp)eTcq1&o%g-VdRl&0=5+y
zv?2>K9s&Cm8GNOA9T=<wN*)OYk>JSHj{$=iaN@WLESo@1n8Xa&j3@&J${xH#ino+K
zQ2d7!?H%m(J(j?hRvAc;W<&mpZU}udYqzFyKg&KNbMhGFewKZqxedwMXvTfF(;N=6
zaG=^pfGmRMD3C?b91XH)nqxs0OLH8aGn$`*=qYG&<9iC0Pr;VsXCQh;%bUTp8LT*Y
z3z)WWOoAtD50#M<(FdN?wTi?JtzjDgnoaa5+Q6#~#*to!rh2V<eE;Djycqrg@jqZj
z1U_l`iMJx(NU*@d)S?UV*8LP)el_SGpM+Pfp^M>Q;$JhYovWQEvGfdU?`rQ!0wBMu
zw@2OY-uRC|z-EFp6O70s1Z);avuMr+X*SLKK)R1(Vvg_%_!bP3Gk|>J6{QRTIaSP)
z&2uFl7sGx6>rdcGo<2t8_4k6AT1tQDEE1hXnxy-&BXn$EY<BNtz_8^=RF0?_dx%62
zX|6<~N}8*XsEX!lB&w#l28n8Du0^6+n(L6Lj^@uu@)_xp5i3`;<O|JhNYX}gJCd~1
z91fCjpxQ@(B!cEBkVMfO4U%Y@V?h#2a~w$GXwC<DKDcq?D*$-`&CQ_Q44pZ73uw1+
zJhn2>r*wz!fS>r<<AYKH4*gN#h9~MG(!Gcb$V8Fo9(z9%q!VPE_oEE-iT5-0v}@>;
zuGUK5)UTi4Q1WMN{ZZc(G>Ev^3=SMU_98zLQ%=jxJ#}!j>ul^a)g$wIL~X+cWZppY
zTV(#0=0;@R$T7(=;6VmR1B|&rD$@;QG{iJxXHS`CAg&G$hWL6L_#FzIxv7)QI-RA1
zuqJ#)`b;Lvo+Kjh@>gadxaF@{qF6)BJHxhvS3CAFzULk0E?K%_<foj0_ysdTkclUj
zBqZ>z6lBqy4T5Z%_kmy^$DyOMA6mQ<6#iruLz9DTJ!^b^!+Kb0veJ&cF~hDgS<^wp
zWMG@7Oqft_avCWLDWi5AjE>W;VmRL`oZsOpE)M&p*re#rSD^s!VAFEsOx{hvK18Mu
zX|6=3l{8l&(<+*)k!dx}HORDv=2~Q0OLHAEt>bv{$#TCIv&i(H=w`{0&J!mV_Wgk8
zq6B0moQ-GLQjnF>TmiBQnyW!p&GCxio;SKp?CXygh~ogqQ6z>ByHr!RQc!Xe?}4*m
zf0jF%88#p6^Jy*s`vQ)`*PMwetgbn$>~D<YB+_5RVc!Mg(o3%4@t%gBr(qz`FT<XJ
zo@Z!23%+MzGABO=J<oCSsHj4(-!}*6;QmiR=Tp#yRE!BI*XrWe=7$BRq4R0(4PzU<
zOk9iOjj%vpOW%Mv1*9E58op)H$kN&9!;h*@Y<IC_SSx)i6Oz$kSZjT26OzPYSQ~vC
z6B1=JY*+oRCM3DTus4zLCT)~&A>l2WZzJJtnoE$dgyy?Qc$emTNO+ItQY0*;xg6=0
zqmDR`!An-}A<dOYuaf2}q*q0AHPWl5xd!Rg&|Hi3YH6-RdUZ7C!+?AkLNv;-1u&q1
z<`NK=&~D0Iu)7N#TTB^by4<4^ETteUrRzTcy9adra&Rf9>sNrVg05c$c2#u!N8s{^
zu3rtpYP$XtuzNz+uLYM{u71Lcyv1WYUMK#)c8DZiMH)XOT(NI8MKP1JsI5rcN(UIp
z;F?Sa1gYSfN^?56rqi4WT{3?R7_6Rgel$Mr!RrebFp$Fn8%JfXmso#Bl~p7{hMm`M
z-eAt>If68gpw1*6OH#t>8b73h8P-^4Y(id|VNGNvCLJ*rYbrA}Ax8|un#s&ei0jF)
z<}z~=l0;=#OPQq!$vURgI1~)Lko5|0)bq&jJhCRIcFM=;QzJA@?qg%L4H<4j*4z=i
z4Y_Ya-W(T0w_;F`l}M5v_>fd)tH};$*pr}lk~<8xUH8tp8}qv(e())zcM6%4$e(1i
zaeHMtn`^IF4~j$}(@KnjP~(S;HlBMAg*U0huw^j33`Uc&F>D14ub{aShF8*D4a2Kx
zu7Tk-H2())|8c=0nJ_%oZ^?6RwqyA=vHYIo!FvwOzId;2c?n+5G8b7tl8$0n2UiDo
zlCatH#=-wm=Z_b$N{)c`5ilgJ{|Fr}@EH~JL|wq<BPRdHh+n<|`jU#uh+n>u=6I0D
z)0_YU6S%dlOeOjjacftZO0;g^azcCG>f6`7vT9XR2ER7BTxviW{MsC#`5@RFq`49L
zG}1n162E^EpNi^}_>+_P)5*X|+=3OUykejxeKS(*?C>bE23MF4y4bq&e=VA<xQS^`
zmtQc!t2GsDQ@PY7!={048qGVvb_d7g{sNy!-!3o64DN6A{v$9l_aS~%64FmX#>5pm
z6n4Io|AM_kus~O;Yd|vWhdX(fz4M;yif7oVztboZJu|Fxf9FwTO&{*<=6@x9`gUAt
zj@ukB&Vf9E+)f}bvIU5DiwE`Xo`ekQsQt5Xli!VU<SSD%M)hDq^rz&W_0XGUpL%Yg
zk^m6<p32)5PQMX+uLypJx5=<^{Bd#Iv>bc<%Wr9hBj>2R&w-hA)+-AJX3?At1G8z)
z0k<6RBjY(f<w38#@tQh#A#DKB2F@*I@QM{}q&Xf$@iZrZD1qi8kQ9M7=@|R{Agbr&
z#L9Kj_~D8ON5@tcj%1bLoF$0YaT#=m&4w=7w26NXR?j&Tf2wa@>Z+1A-bjFrr0#hx
zv>|?qhQ<#!PS^;!C<bsw=ni<?0dH=|rQlIYa~XJ)(Odx@6*O0ZM<vbG;89I;4S3Yh
z{0Y21adUrWNJ`j|fzKb{z`h<?*CQ%_*MO`WXnu>V-_qQOtQ$E#b7fiIfL(UYddeJ*
z{4?sfjZsFt@-H;EA$c3k?MU8Eb2!Mux!W9ujR1KB%~2qaqB$Dm(KN?`JeKA-oVTL6
z1QaFU&T0G(DDKc)3W`#i%Ro^^a|I|WXs!fBCC$~KsHV9F6g4zAgG)1)7-iTNaB1QA
ztn_7Z%%XWbrETFGCGk6=*22>4-n?5E=KdrQx%m9>q^r5-CNTm$5yH<%OkPxgCqnpz
z<~Ag3qq!Xk+i4C5VK`9jBS08Ia})@pXpRP9G|jOfjHNjagmE-C1JlfTeFCfqriJ56
z`~4S6h6tnnmpvzLJ$9m$7q*<eQjAed;x<+djH{uq(msRNXKvw>ARe1h<+U#Hy^4&(
zo`}Mvov@Pg!8acQIqemIZvn?Qo*yoJbL*}z9-Y$|rvPHoZ&|I1i}+=XraBJGkX}AA
zBrcQEwz+TM!X~9{bKgjFJh;cxoB-Yl+<29jSYwM|A{h^ZPh-ah=s~!|JMrI*dH)*!
zWOWe748QibdunpbJT%B}TRFgDCAQ|d&@GpCz7IgR12i9mZU<>@gno^jb6yeswCqA{
zr#!5fk4X3tF=T|~CLEt{<E7)NRFCpfoamq2hvP~Ur4y-Ci}GrmR3iSVn#K>e;n=dK
zBgMWBhk#)tK^#e2@)!`u(7X@2?EBHO4}Q+6ye4p2wacK*;-9GWX^&1#>YAN5;uk)F
zv`-*oa#EAqdHgQpq~oX;R$kPTcX^$Bq$Ryuu`>5Pvh&D}>pzC|@9a;}DD6?5$&zK*
zKxv>J>78MxNvG)%<udGa;dCj<sxjE}w8}yD#1&=O{m5!R&AG@l7g=%gRwQUe8f41w
zeFqYJrFk6))&V7t1VJRtF(8Pcxe0_#&~bZ3k`LIEQ~nx(+PP&&@wj8+8dKcLMRDfJ
z6i6|c7IW7W#Gk@`z4GS@RQrcYo6PNf#cNQ!q4_f?zR-M?ueipi>X-2q4`_bQSG=IP
zOrUrmAobgVzWSPIpp|%h&SRV>5eII`%+R1&cuby$#833zuAP9$W*ffB{6hFmkHv`i
zZX}_}BddU91$o4on`Jn2$1Japv543Y;vMqHhF!X3#?l<62Ey@gkI283<O`4YJNI=h
z_cfZTysG2Rd^Y{j?tX%odzUj&C>lkg2()t5>NUTwU59Y*kyM*C9c_54xv%S}8d22u
z2<rQc1tFpMCpbKxnZA5MCpC4UNX$sIG^O6!o`bq-Yk2E)*3~nVxw!Uo@Bhnyflm6}
z+^qWAng3sZH83_YwXp1BZDVWi;ON|=XYW1=Gr3XsUW4@wh75~A;5FW3C>s7hjx^{j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/static_buffer.hpp.E059C0D43CFA4618.idx b/.cache/clangd/index/static_buffer.hpp.E059C0D43CFA4618.idx
new file mode 100755
index 0000000000000000000000000000000000000000..679748e6a592339c3a13effe42583ff922321071
GIT binary patch
literal 5930
zcmd5=dstLu8b9Ch$YnSO21W-MKjsF*%ml+Fkc%M7MdKzA@equ<EQ&_6l2;@ag{(-3
zhnuE_WaNdoB>_pO#j;9yVYlpJ)^0tz61pg5yMp8mcfUFCcFw>4RXh*xd4AvfUElBb
zUcO0BSh`f}N>Q<S3C|a-*^<jp6y;3*H*eYW><@^dY;7p&`8Ru?T2}j2fb@3ie{-Ey
zww)Rtb19E^c=z+f{OH3IlsS8MUd)jNTk|4zbloUBvoSH$IJ=p4?7cQPerfDZ`Nyum
z)E??M7?59Jj;;Boy>0YZ*YYo`?(6HvEE#E<WwBSD^~%_B`PiYpWw$$5tiE~Fe*ZH&
zQ}(^Yj4MBFy&pMrbnnE{UiFfVrtX-lQRWFl{>!R6FE4#BI^<Z=*}KD;XND(tJq&ub
zXEfom<@P7BYx9S{%PYF;8ZUcok=FP6;GpmI%Ftqia^aa>pRF9K7;I{Y^Bc;_dTy5b
z=f;iu;+~9HaWi-O#n*S=z45f??eXUQv30)QPFqf|of^&maZg;p2zTPYKh=G=r)l(1
z?%?)GloZ_hpnTr_6XW^XvVXkxmuuq}c0?RDo_N@ICjOJ6vgCV*yxzJo(J}7R*n1>!
zsJ%`1@aVapBc8~+v_@7mHQb*4O5)jli+>0{_gC$?z9;9@bXUllEB!9}mw#Ll_`w%*
zd#cOnck{E43>>`j_TsIXS$oIYFNS=(wQA{syqK#!#=5J^FNSH#$KK$4BOQWkLkg$P
zcSjG^*zR0?@}%1+y=?aK5!u$^>a&ICmZxl~wt0}Va)t7g>(0?}kM8f!Onn<=Irl=G
zt(VynyR&v8?vTf&{?PEdXg=$&qBa*7tWuGSpR=ho&+ftP<t{WgkHT4gSqphV*%3aF
zzqx*@x4Vzty+YIq3LE{53wZ%*?-!#(dR)G;aiJZ_=uE%NNM3;2|58|d!OcsRETwd(
za3AvQL-S`ol)W88xl#zDMGMq}0fqnUn(ow(Jo`zf=?`UZJ8Qbr6Fmn$I<9RxC#5xH
zAMyZs7%yO<#LU`})X@#^OKE$uk51&#iL^5x=#sqVfOFx3P33ZWF6ol4O^@IOs8cR~
zy3p{HBPXR<qHacpW)wB^fi6$Z8O>X~y41;qb|ziSf#&(V0Cncr(5yabbC3%yA?l0;
z840`qwb}cW*TTc$UpdfjM4c>2cH{-9R~~FqzF1g&VJ@vA>QqIlCoe#~=J15%-xs4Z
zq_mo-Ejo*d7ofJhS1&tZKVCO&9%X<s3|j@Li`uoTt7E%=k<&6_J(*VK#|uzDKl=|g
zPABFMN@)i&d6B*-fES?NUHoltScgY9xnwe0@28LC1*i{CEjqa(uXSSDat+9@0mb3z
z4|I9&KU10{?nR3YJ%{ih-Zoy%3(9^r@}{gD=(^PyTUzt|Ln*B#JIL|L3E~Be|H!{~
z$>FIZa$;Jt1Kk2$0xv*){=($y&R1M3$!MZB`Iv%u0qP6oA2fB`<Jui)JyA;~Qb%5Z
z`m&4ePWI5j5@K4S4%P*mcme9+UAua-Tla35OM8*6r6^J~ya4qNwXF-@EPtcLftC|>
zk|fE27oZ;bTj(Cgebui?X(uvyxJS5}7r(0SxL$gv;f2n*v?m#zrbzST1*osZ^ePL>
zzcv!ikkQT(XGdOu`lowJ%k?iaJ#A<gVpd06M;R|DdmCEEn{~1&XyBcj#bzljCDT_T
zeI*K*`M?|xH=O?Lg?N23@f@;5lt+}77oes#ty#OdLrPIxK*c`Gp1|)o7Mns$kzjHb
z2L}X)fC*Wg;+^6P7GrU$cd9QCo5eXHIgxPDEVcw#LV(pQE($CP1sbxrSW~PQP$dei
zM4@mN7FQun719g10WE1jNkBuEc<rKQv;-WQB>^X-4@E*O;(A*L46`@%yRlft$k^#H
zM^P-6Gjb79GD;EBF**^_GkOs+GDZ<HF(wfTW`ad1oCz19C?-mT(wTG-%3v}?D3i$)
zp)4j#gv^XtgmRc15wb8A5h`MeM5q!ml}Lme5Yr&W&4_6h<4(kMig6#J`p`_0uv)Jl
zQT+mjtb!zbHUa;KY(ko?sRa1z_D#3lgO<L>XEE4a;XV&yJ%iN>^*jh|3{LY)^B<XU
z8115?$8UvwE_s9@kgLez)<gw+GUYvp@SwM!vSqNFr<+b#yu?uQ7?6;`+qB!f1XPU-
z)#x#xAcJd=wg!0#xDja^krz;s!7WJJg1mr|Kt-~X>73a0me0VN6ZIbi9-rt;%{-uH
zSnTNG7zFlVv9q(Y9)baj6SxFV*glJsos;#DTmrAX)j98vpRBiIaIPv>4c^A!b*go0
z@HPgQs!G+uVaj~Ud<9g8RCP!#9Hk!l)FWR3??<ZrNG%+#4XN6Y8jkjO{K%^|?%5#{
z2Cq`B5^bPBRUq0xiLyi`EV<2Xo4bI@)MXk0RU>6JQVGk~Ah#OiF5r5ku16XHHzK!2
z<SuNl1-Z2#ci3K7%<D1l-_tn#X7-6#U-3-y^@@!&i*-g_I3zX}XBacWp@M*li11^&
z4g!8GIr^vFuh%Z>okcj{mEbRMAW4y=gG@viVMq=FMgX4?%@<5SDJYo864TWZ%$3+M
zQ2W_SKXQl!#KrO@FdV8UGL5x#g5F?PGR<_!1PZ8?h{9SZK{K3?G+PTKXok(&N%)!w
znpvE_I6V%QW^tZVo(w3>;`L7JWk6{b??6sFkPLXk;@!w@H)4SoEN((hO-Kg3U~va>
z>OeBs|BA>)*}>VP&;4e+fGkN%__9y1Kw3arI0QX7Ex8111qhuWLXu`{(Fa!yMIULl
zR(#M5CL&j0t@L1$s>E3@|5W?GFhWeT(HsEx$Ko#J(uLfCFI6|<F6OV@^x${u-%zTO
z!&&Q8I5S*3X`U`j;neUYX17+Q;Aez*B9a5;h#{?&DU9hI`*W}BtGC)8Aq5><`>(&p
zNrtXiYY$FFUf|Pz#ML;t2q`(G2<bST2<bV!2pKt}2$?vO2nBP&A{5Sri%=97B|_<3
zx(H=(86uR)Wr|Q1mnA}G&MZPXT#g7?IEx4saYZ6jiMUE6!VQRP5aVXVHH&d4;yT5+
z583x2nP9zsWZy5~4<j|sBNda?B#A2Q6fO|KSgf(rxPY&)IKnQ%1*&rv_aM6-<N_>*
zuuS-3eQW@YK05n#_K}G3tB-Kw=OewhrKjS(k1}NDHACoTeagV%UtBDF+DCo5j5ug<
zaB&Q{v9%lyu0k;o>EN+Lwr+jtz-Xw@$)wg7&acgH{G7Y1&9sz6`+gMCj}}Afxe|6h
z_Lq0v&?D>{%6_SYe;|RIusF>-%>*Hg#krbX19%jRS7}y>&^paJ0|Yr17ibDZs6<}k
zAso6?Q)&=UnYPR*plT$qMjl`g7S|z79Wn^G9%<{5QNWEz(TFq<BUrp2Y4#(7fLoBF
z1!*8vu(%Cr+K>S-q+Mcg>+27W4b&o$#ro<4gR?kFAN3eKy1*)=llAQfV@U{(HYY*|
zhPuyI!ax3mnPfq-PzdrYc9Xj)1g<OO3K7!CH41_G{!acf;r$lk6e0t5z}zGcTVInx
zI20!k60NjyKVd0%JNG#PQfifc!cuBG^&9~mM-Io4D^$NM9zYHQ$Q3Ae>&@c>{#_cc
zN1)ue^!kr4Ri+02PP5VE*6irv5A(NLUDi+>BrvoJ8LChaFmQ7F=~DUqlTN?aQ@oOJ
z%xkFU_HR}jft`qlP5*=eegbYp6xJUufW725i*>nLk{;kWMZthy8=&jvO{+KaiO2)8
S2}NOz?bPKGNxTesQ2z%lWQlD6

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/static_tensor.hpp.55059299885EB585.idx b/.cache/clangd/index/static_tensor.hpp.55059299885EB585.idx
new file mode 100755
index 0000000000000000000000000000000000000000..27ada298945d2cb7bf610f7622a5d617a25cd92a
GIT binary patch
literal 4350
zcmY*b3sjR=68`5WyuwY)gCyii3<>f^Krj>y0THEDwutZg!s;R_XjefNTy^niePNFt
zmjxH!x@)y9i!4^E)@NO9)h)6v?!iYFsfr>{A4P3_;NJYn^#7Z4GWX0sbMKw`X70?4
z9hs5wOvQ2Oxg$##%w9ZG#&MjQ{fibC&T9vb%kkp4((PCOH0F2ziSu9mdfcnJwG+$h
zTF!5nFd{PUU9Bca6Db>3b?;`)M3cRK$HoQM%XSrP3h9`5UzPkwb?Ico<)6P_+SGNc
zJ!HqEO3!=cRenkTp1x}B>Knv-GjMF#)s^jvTUPzX1I@Zu@AosTjQSvT<@6Pg&Y%D8
zc5kWgb>mC@_I37y_m2*g@9MBksvk0@X~nZ$rsEH3UHm)CJqz^HiWhGC>%Fs;&li5#
zwdu{_<tf)%`b^tj_IBu*vNQ3U6W+N$-!ysEg;+)W%`all+K0TA)ZBMYX{&eNX?<=V
z<wH9s#}4nPUb_F=eLG*i_3^xYE5|0rRfn$He*V<u=WjMWNj>{**omJKuG@}nX*0oJ
z_Ktg0G-=J$ZId6|%X{jZU$@(H_xEi>+m3FrezH-2L^(r#?)Hlz`5PY`8d!Ay8&&ba
zes_Y;)EDTM9nki6)+)PB-DsWV)A7eYKS*6MCFp-kR{gX5N)oZ`40oK-WlgwQwC2-W
zmzpkra8wam^n?a{ntmYMk~@zEWquv?!=0_4w`-zrZ~ZJP%%rv~JAZCj?dkmN;rV-x
z$=AK7Hh;H%LwNMxbE0qPFZeBp?y)R-{BFQ?<>lp$ymfol9OQ3gla5h|Wc9N6afke2
z&2Rjw)?Ukf5%>C;Y5H34o0{VuT_eUlELyT)mOl%<_vpQosxLiK>v*f4qff!`6e8RU
z=fkps{ghviJbwL~_@OG^kEt|5SR=%_7gSLxDpOysSvFD0_hl;2!RtAg+zYA%ytLfY
zRUK+m@fN01995hk5EKXPs>uzVn{Y+X_tA0G-|8PD5ERF(ZZDbp9;Z?A2DZBgCicKo
z_kt?P=cg5ntJ@W$;(eJ)4fL;p0qzA=hFaFmp@Rp_Q}J@9(h5<nkmg=cCGGt$uHIF|
zE?4q_%#Li59U%}DXY`NVTr;)x0JDRMvm9B&1cKs`7qbV1UyL*o&g9S0az416x49S2
z*GI-_IiQfW?aPuU>QWhloSiY+FgijYc*gi3k7<qC!Xd6RdLg(MY}B=IzTO_LGschm
zM{3s}(l6?GBRfN5(!>b_&zM|O6?(!xkLr0fb2QJImnIMt=jLume-gAWnuUw)9^*B}
zBoGv@t-3sAWWU*$biA4Eo@tsHB@h%>{9LzZZPM@z9p8_Mb4)o_fuML>b&Mr>&ySD&
zcnuTh%5(h$g5upTqMxJ|b|y1`nfMuyXDs1d3#uFnS~4bm!txRwAHY<m)9Ge`p!oRU
zhg{ksugNaw3akZb)Jaf$+4sVhwC$6(G4q&nYLnV35EM5LZqGg)=;`6b6BZ(+w^Az*
z&ew}qaL(}jx8wRC%foX=+3jq<imJ>4!TopID}FrwDkhR4&y1U4nh_@u6yN96rBjBS
zd+f?%y;UDWodm@l){8qeEelVo_#n2s0zxWSlDZaDc`)h9tNbaK;$2DF2_c<er>+H6
zdKOpR>uzcD@#0m?hpFCELj}V5dV*5$;c3sc!joIC{hmEG=7TxPoFEX~|5L!i5r6C2
ze~Vec#K|Ga0|bKN!r6I6N91B{;0av~hSd;(Z&>cG5NL-YZr#@W%7Hh02#qjCnDKT(
zGmIH#yo}H?W0@I`Aan;9cYqlIA+!Nv8^DIIn9vpoZ-E#rK7`(d=(`Y)Pngih5d9e9
z5iP~x1E<107oJcO>M!#ruFA!6gz9B_31^m>C0w#BS;FPX@+4e=tU$sQ%Zepj1;{Eu
z!q<STM#?vWtWnChf~-}_cY>@_%J%}-3+`&`w7Hz;%K2Y-gaG93s*A8=)n)P71KK?h
ziD1gzwHFl;bl`=Wi=VOPvd_+fi!TFftv8F9wk!ZQJ0qsu0Rec4Gh*5X2#`jMw167T
zbw;euUGSGitj}Zc$B4<@^&73hbC{d1;*A14fC*eB8wFUaWh~|i{@Lgbp$54j3?Cw)
z35JA0n38g#nxm(~=Z;aA4p{H$MQDgM#DPkL4zmt(h<Urzy3`?}Pz}~<aEK_>gS8$U
zXg#6LU~L8mqA=p~f{HDhQV#>6xyoE47DwzO9_+<vBLc*mGqf22c(YvSEodzE76yZ>
zvDEF1IyL*)<_7F1EDd4}u@QKdT<9ve8@meI?dm1C8#@Uv7di;)6PjhsvZ0ySPgtfn
zn+XPB`qr-Qw^dK~GUsL~XIZdsV0&Q(jna;a6y0%l8oXdp<m<5D!IdAfnqhHRGSO;6
zQyeLoxP{OXM@goLeI+<5AydS@796#ZDPrFQjwZ-N><MiHM;l~f0ukB`j&8`r3@r7`
z+1Tfp?JP6D0jv!Wg$I?^PIxeW)cAo6ESZli5Q_==$DndHMMQzHM2_tN2fmvPqdY$B
z-t;C(L8w}%4#a*$s79v=M7RkJ)&v_xuZL=e8bz;jbUA_8W(l34n-Pd@htQe2nSp2t
zp-VJNj3Pc;K(hsmB0hDXsRJXrMd%gKTmhqq&u!4$2BV12L(n_~BjSS=Fc_R&9{t4}
z_K*vW9X}KHI>TEbLOiV#tep^yr(xswBy^5^P8g;Xq4VYQ!?33lx*iPcVUXz3S77`K
z643>@dsM(n4?cQr$yBV&XCdqXSr3>n2slF!xiCARJ_9phMv@3-Nn(jzylZL7(sYrl
zgv3g)i(pnmN;Ra5d@Ur_f*ma*v>sCGAzkE~Ah8MTBG}E4(hTV$-v)_oU>D))hQw~L
zBV0%FZvAVx{9>E4p(*?gn1igLDfAM~tT0QsWJR)s%Twe@xB^9igez7QOSlS9RDgu9
z0Y#0JZv;i7ly3z^tCa5qMW>YS1>auKiZ++?zUBNcJO-Op6X)QDvB8JJ(#<)wVWGg)
zQ!Wf@ShpA~lAgt2VN({W61sx-TfrN!=n?u6@AnaJz!Hiho|jx0?Zj=G1Wl6oLM@3F
zyZAywQbN*2E+sL=F22`cDZ|o{yRlu@)0O{;kF&6uEjGy(bBYa>SYb1l*rd5;t^}Jj
zx6HNDVW<hrO<<Gej=2qN(%dn3gAH>B^P3ea=YWa;vnp+!Wy9Km4U$DciXGa(KD*{p
z+>M8`@VKT?6m-Anm)~(^VlaDyS+Xn@_6!uWGU1$1#kfg32>De8AXe;d=gf+6dA2Lz
z)W<(8nZhn;fxa!!A8|l5nZ9#~Mg79r^JhA9uc+|1MSg7M>8O0gXNLIGcRB9=sOOUm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/statically_indexed_array.hpp.8A80A7A4CC47B954.idx b/.cache/clangd/index/statically_indexed_array.hpp.8A80A7A4CC47B954.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4e71ab4ae2b9950d038ff9eb3da87b0e1460212b
GIT binary patch
literal 2362
zcmYjS3s6&65WRVc2}xdJLNEw~2jLeKN)oN{Q;bnlz&I&_5XWL{iF^d{GlaHSrd3<1
zD1(6Gr;36IW9=vcQ-9!x7OPdn4~)YA;?%TSiv>XiMN75!JzWxJIK$qvd+)y8vrGJf
zsHiD?gyyF#$X%V9lgvQ~N%5~eCp#mbg-``P$lYJF_}ZFgjb!MDk&Tl}2<~<_Wt-2Q
z<=yPW#1E6wrSgyZy32NZJQ+B7@uyocyIO*`f034+ReZg>=K7~WX2;f>FYkU?HrSDt
z8#=P+$@%4L?<DujR1VK|TZ*a#iE`nU(mDN!38ju%!LREMEYtjzce#FWRG8W>ZjR*-
z#6C)vY-!I_+^p&ia-8k5&lcXhI>!<H&$CyN4%J0J!-;+FWiO+N_@@_^h2N@=9?s8N
z!m2sC^!Be=j}!@_2TRv{#@_igSyp<csV*V(s;0ra=uvE7UU7QN!J$jl>#Z9?Iv%#4
zyw}_&X<XBEr&u{!yJ1~dtoGddJwf~vH-{SsA|Hn3EP6<#O!bfHnwO#AD|d%P*AJC!
zTH@EdFE(eA$HB1Lb0s0CFJDZmTV5HqNSGL7O$)tm+F5<#kSV-n-i|4i>77X*=gm(o
z9tf7ZuYE7CH6Zb<CcYq}r}OakV;!x@`3n2O32SrxN=tP$g<gf?D`y7##7*kjzEM-%
z#OeQX-5rOvBK!K)wmk%H%PohU%YD(xC4%H`l$0lxF@W4Kg?~h=PWP3!%nBiq(gRVt
zK%IdB#21e9mmwC#QvA`Qf1Mw7K|+ud5M_)oY8Zg{5Yf|~N#8xr=MoY*qC9z?N(K<0
z&&Idocy($~#k1f|&xJ&=H=<+$nV12H*ES3tkF`|#3yJag{cvu$f&piIQh7#R$0rLn
za|tPSh~!23GvIW{yHw#w$XyZYPKa@W2CZQl1MomqiN}LcZQ^Dz;e+WEZi<irpsQo<
z&fd9WnH;C*i6}3DmzV*dk3Ba&JC$xK!1P2!DFsR~13;fEub*~(C}o#BA;Op2B5aWi
z0DYr=anx7rEE|sYLX=9N5;FjFTbuR!8&9u2mJ{wcS|6bQfB~R8k63;)HTZ6zOUt8z
zdBI8sfbL-*2$O8PA0Z(EakM4EqGJH)zpsq+c2!pNzU`Yi)I6D@0qBv;;i|9~`w0mV
zfTOLURt*DV^e=Ny9jwk!yp@GIP;H=SKzxFQ7+JhtDQ8(ezkPs~6(b}>iFslf;tS|?
zfcySNzqP2b^RR`^o=f@h{FDs911LK+&AyHHeWy>BfYIgAm8*XIT|K--q$${!xl4P8
zLsCYqaXP4wq|9MvEqohES!P=1Kp{xVI@LN8J`9A7@1<2;S02`s+E{u@bn-}3=(u&~
z`6Q)A>UTqc5J~Bf&c)Jm^yJ&mqB#S{;9?m$Mi<M>F}qk6j>W~Ya;z?vjbj_L;Cpa(
z^hklo$EHcZOc4}=q`Xz$b3v*k<)iX(vFO@5bES3Sw8AyvCKxf8MEoW_hERB**#c(r
zpim^G6>60upaW`#aE20Us-M!=&^PPPZTKdR8%G7tnf5jn-}tsqgnK6Skb1!}wgRMM
zP9!Ch%H)s>NvWhNIm`)^fe<h=?o5TA4Ip9O@#FNc!*L<+=gNZS?2RmvQc9I_7~Db=
zdvkyBn;H1nFv*~SVAu+5l8PB0;{oj;skrfR9w09Pvnk*KC_hKQYzSx-?v2?y4oKXa
zIG;qP6&@dMfdWHHIB|NBz<``dyVq(Iv^Z!*^3^hD)pUHFi=`LoJ)N9EWN@*JBBP6C
z7MWcvi^$?)Sw&VC%O<joS+L4*cJv;DuZGs(V|stVPJ-3NBAAs0-aun-T8E%tm{9CZ
z?<w$xI%98oHNi1VF!rYR(%A7kDPnJW6@fTl5wZEq0)i*8=<vTfXTiLs-oDTls64)j
zo*f9uPR&XVz!A_dR6F>ea-ry(%&~+1S5#(iyCu+`y~3`h?^eJA2vNto+gRMJy;*F9
NuOISt5`NA<{|C=T#J&Ik

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/statically_indexed_array_multi_index.hpp.E363636635FF265A.idx b/.cache/clangd/index/statically_indexed_array_multi_index.hpp.E363636635FF265A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b938c546048b9dea2700c3874a28e58ad2867ff9
GIT binary patch
literal 3500
zcmb7G4OCM{7M=-D<-zbEU`UYXD5m^`hzc~40RB)YF6^>V(7N3<RsP)NL=?oI_=g~{
zU4w|OYmq`Y7471#Eg}{bEu{_JwF1Y#`qM&XwJ3Vjbx~P`%IZ!M$M+7Np3~->+?Ox!
z&b@c;cjwMcPKb}+?*>3@MuI6jb5$w>0FaVz-l~;ZJrICX<PTHz$H_@6s%Cp!J+k>9
z{3P{+iwAU*x?2X$;V%A8(WW=N6xMdv(lzoio<pr`y*-PzdfxP4zv$esqGzLX?z#uV
zGrBJBuDy9?ig&4|vpXHg&mNms`^&paTp#tB{3EYkwr=b(eXDMCicWd*>uz~n%5~Qv
z@1eOf{hrSJtT!-Y#<54;4eJ}1fVSnGoH5t>?9jlMm8Vn7{nqJMZ66cu`k9{VeP`>3
z|4MpKr8!*G`uv91_$w7DgU7u)Qo^>k^mJ_A>DSO)zrdx*e=tQIxOu~+#w*Ok=D<{!
z$f5^@lG!(p_xp{h@XF{Ko3Zr58lT{;mlqDk?aj;ZZa7_e^Hz=LK#Qkge)7&jrXx9O
z`CoFsSD&~%@yOfRnM3_+4t(2p|CWJi&S&cUcDXG~cp8hW^PCUvgo}!j=I+(W2Nb)M
zV?MHYC#>=~J=1(*VPW*5UhTfWditEjC8>MvZn1usW_}A>6KDN->*2Udzs-tm-FMV<
zD5z(bDxswDSMx2&d+xP$CHNcezF9J*xB5`$+V+Eo1|El>dbfM(nkLS2qv~(d@8_!$
z(_bImeZ2q4-NW^ng&W2d&FGJd*PpxJIGkNv|J4Uy_x&*TDms5>Y`$;p?gQH<U#;91
zwIb(SnCv;M3bRD!t;<duA}sszNuPI@F}EZVKOaIq`QS}1o1}nsfbrMr+qxoWRwEGP
z18@cOser*R9)LxoK(GR07?<q3f8%54I0PmTC3mHJxLp92l>j*)3d<CDTtG*bud8G%
zQ9z@Shg|?x>OtHPh2;~XGK0$Y$q<+{4q!7>nqk0;2Vl7|4lB9A{|i*=b{8X{B35`M
zA8HqXb!8k1rKX{EWs)K*0=~pb0aO+cD~^XfSbHp=es$^XkE0&lL_kTb7?cLJT?lTh
zf%%~kpCo>WKmbwDEA?u-5ER<8W)J<5nKg(&AW_gLHEO#6tTS}P@y$SW@*nn?-$mdx
zq7b8uQQL)J;kx#0?30fa&=G@4X;NdG0Ib_+nLWsd)y=I_8~k2#C>WJSHMR-Bx;S~;
z6$Zm|+kR31&)Eq06BjC_N^KXm7`*HXPj8<G-{dOl5SU05WJ;OVE&%K4?g$dB%v_q6
zLueiq_Mk4O^7mgt2J1p~;dFOy#$)$t+9LxX!ctdh5armd=|F4HsGJrzgx!I=v!hA?
zKv)K3BE$n;gj7H!LK>hEp%@S&LVBPVAp<aokP#R~$OKFxQ~*H%6k#(2W-+dSphCdz
zii>Yd`?|+Swme5WM@RWmTsiDKTz4$YiNWdGbRDe}2IpvUrc+87yxwg+M;WB&1G>C8
z6%94&*31(-k4~EPN%KSPIhUVdD<oq$YP@8Vbu@$3yc$!A863e!U`h*vWBFK2BWCbC
zejcV3%izWQVl0Hc4f5L{re|v^+Ld(1c<>n^v;^u(U^G20q0IyJ`nUE@BrlK`QqjYz
z^-lw-?u%EGBrVV`&`~Fm9#%_J<=SVX5UzwwC3K-FK=^&gyf4O8kf{=53-q=?UeK?B
zOpWk<_7<<vw;$K4NEkbyZU>B}fz#4+a<R7+nqh?H!ScwLb&hc`j-)Ljv9h)NOwm0z
z4}?`yRN*w02y3QjURF55G2t;<no5M@nRr$}`f$CL?iu0v%zRcrhH!(HrWxThCXE%4
zQEQw*YX#vI%nDXOrf`#%Rt_SaDWm{~Q5Fy`hD<SJ1#E^PW*A1ZfN(iv${{P@3K&`e
zr_*%)=INf$zLQ0-kErQIwHa6A8-7?VA?b3JOGW8&#ZsCsN;#=wTXUuvUS_RRRayS~
zyAc&YBOuz2PEGF+YL2|Hb!Td(u(Dn?yt+Pt5UcQ2Or;@|`p3*I@5<Qz3g4xPPR0K&
zf377h1_h%~8XW0ED(N&@3<QNqbpo2qO_m8XYdNiqo*0DVxHy?WY%-TD6R2FmEs+W2
zWpmjwfvzo(+X7_*A!U#&gEG2(gsY&u3SvqC!WKBq0wXA<>qs^ncMwg}&I`Bu3r}vS
zACcB-i~9SG*3}Us2&~O7E$`$a-W;Lbba#D|-2CwGPCva@9bbB_GtQ;`x~+Ay?p~tS
z&B{cGXL%7)u__VLuo@AHVPix{&+0|Uz#2rz$Qnh+#F|8?0I~&8gw2pOi*W^HD+GKp
z+E04pPmjx3ghQARPM}=J=s1D$Im{eRpnL(dfD<TBXVN)=@*F0I6DTi%ObO&Db+%5#
z?0}p=c{OCJAx9~uRv3wWhM)!i(<y0P*?WJ{@cOxg+>p?amwk%HDv{Vn55cau5udtn
zzN6#Lz~>|Ch30^)wT&jiTJh!HKKSXy!CChvzjF4@Mh9$6*7NWEjIyF>CJt?*j^I#y
zwRYXUc7*LKXd{P#ac074?R_`1&1+6?{vBIuh*Hd=t)VAqTT17ARn{xv(dJ2f*#;)T
zm-tSmY&uEoLzAHN<Jb6r)h!@*+<$hw;*i!CacCQ|1Y2Ud2ET|o`oaDzo~#y+N_015
t&$juCu3eeAEY-IEyp`#BO4|?I*HX|@vi+TB5`H}3ZT~nv<LM6o{tsu3gCqa|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/stream_config.hpp.015AF40CD82DD077.idx b/.cache/clangd/index/stream_config.hpp.015AF40CD82DD077.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00718ad8b7637f281e142ec29475f54ac79a939b
GIT binary patch
literal 618
zcmWIYbaP8$VqkDi@vO*AElFfyU|<jg;^LB`%;i9ulaYa;Voq<LBi|td9+uCFSprSY
z>unm@BNPjzSNqARW(w|a4_oo*L?XZT+n#H8q7|j{XR*nh&3{?_IK)ORFGOorV`~)y
zhZLvN)_Y-b3a=+~-AZ{Hy0hu>jIFoY1^hyUI17r5s+r6k@4k6qRx4g~Z|*E<k<yO^
zFVEdAOlNv-YVqXqR__1nPS25Y|HZIKnM+qCBWR(|wuL|ECG@Ao*5)s&WM0e{cblQO
zGB;@t(7mRSrsBr#L4Wv|*qInucsO{3zyt#aP!0&nQryZFPCvHeW8wfx^KtMAfC-p%
z?X6UyT7MIuZf2k~F9)v(n1D%7{hkop#4mG|jR_bkEFA0{!eD}d6Kpmk3**+t<J{`b
z=ll4WxPbb(Ik`o_1Wf-;$%&WdZs0$`$HWPg<`UqN0uwOlqSUnFg+QZVeqiJRyB%h$
z&AT+gt;NiYj65t{EL{9B!(d(lss{N7rW)oIpfJcEFku)C6b88;>~J6fb2v~K<Z75I
wm>oc2kZWPW3`MCqi6E`TMajj=AO<UlfB-fa0~kbs@vBeOEZ?ca#Ky`10L%iXh5!Hn

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/stream_utility.hpp.7DD00DAAC010212E.idx b/.cache/clangd/index/stream_utility.hpp.7DD00DAAC010212E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e44387cf8f06b1407ad5b0f2d2e44872cca20d3c
GIT binary patch
literal 528
zcmWIYbaUfiVqkDi@vO*AElFfyU|<jg;^LB`%(FmxFCzm(#hkShHfAv!3b=gL)N9(<
zeMhau<a+Ds80qUyk?Yq!RaraXTaM3%{n3^_2STO9qj}D}Nj|nFuI^s-l#tSmSDp0u
zCh^t8=)7K|bMDhFv(@#!+0B2C$yfyNGsG!1uFrd~HDBTM%&yXtt=><$Gz`5<?>D~j
zPz_}8lUQcU*rywL)$df}SEhNJUHvVhv|PQO7uT6RihHCt-=!kr*~8^2tx6K{_PhJv
zz16?mbuRVvlqt%k_hdZx`}G~xPs~{3nK?azJ(KV7+q^BDOZPDpSLP-u0(~{>T{ZWz
zs>dZf3>*v$tXx7|dSHTqlZlamlbMl~k)bFxt$009CD)n{XKiNvd&b1W%Fn{j1J}qZ
zA}yi};{}Gg?`L?x#LURUswASM2@{U|c7T(8?W!e=EUfC%>M}6lsoxV~oA_m}0#$RF
zayi0;VKgHb#4SL}L15EuDbd0s$09&xvnUC}RNa)Ecxmni{u3Z!3neF*FwosOi6Cvo
jMajkbAO<Kxp@1F60Q#9Zq-sHka>@^8b~Y9UHiQHK5qq~k

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/synchronization.hpp.94B924A949B60CF0.idx b/.cache/clangd/index/synchronization.hpp.94B924A949B60CF0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..978a463d320b51ad5e088fbeccfc4d005141585c
GIT binary patch
literal 466
zcmWIYbaOk!$iU#7;#rZKT9U}Zz`!5~#Kk2=nQcJ&B@kE4S=(pH$E?W1Q0aH+i0%B1
z5sx{z_1$Ku-r8g5KGW)w$>kokn01N4i@WN~oX-ARJR{Y>H^kdAAf_>^=b_#yL58lk
zJV%55^tgA~RwQJ3`irw~+B<u>ucUVWF|JD<{bwA$o?ZLpU;70^*VK6*B=ru@vbijh
zeScT-X1QseeEKieh0Q9i%uPxLx<uS^8h>x0=_MWpW>yAvQ2|jcFu}md$jHFS!pP1z
zE$z^AJ4q#G9tI|$0!{%=IWU1xa4|hPmg||<Zyp8?paMAoIZ-fyP{2@>npWJ8WEUeB
z*p3bJwD+g(nz$|MiAdy5W5Ezc9(HaPZWEwFFo0PMR1LBfrg}%(gEKB%ryer%uyZnS
za=}#du!}N?BJt!H<lsD*3xOtqoCh-r=&qbZkpAML<YI9UgAGJL0Q(JX-_<?;pWS0*
GV*mh-jel4G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor.hpp.C84FDD70C53B91FD.idx b/.cache/clangd/index/tensor.hpp.C84FDD70C53B91FD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..06c01cec006740d5890094eeaa27e1043c3a5f33
GIT binary patch
literal 7706
zcmbtYcUV+c7N2|W5MkyGoktlUbZI&cpctitihwb(tymI!T?Iu=BpPM0ARsGQk!nE}
z2VH|k!4MM&*s`J@iHnMDMM1Dc#1f1IV)l|ZC-0g4W529?>hZ_@y<5&b?VfR?M~*ai
zl*uCJjQ()R+!ZtBGMUVb{Y_f2eBm-8lRcv{*@s1ypG=QEImmL;v+>EFDL2l1+<G80
z<%l!YMcmDaK3mW@*mq3CC&w?{UKcp#<nafm)F0$L>;CGlTX;#`BfnjdUMs81tl~yY
z9Qnb-U8yyt6H>KXTx{1EvqA4;<q_M=_j+wM->_uZng_>Y-3^TD3J1P$skmc*Pad}~
z|H!zW%UZjiy@E5d_B%#Z-%43GZE@z5fP#zV>2p7An)6O=a^;4#vg7^K={vi^7R6sZ
zQ;_-|Su3BNKY8Vf-3O}X`X66gpF}SQ6wg1qVP#tWeXqTDJ{lb3`&o4g@08hk?3;U6
za(-CqGOae@qtEkpZ{IdFtk_+<Ehcng`i>gc5m$$01r~T349n_yFE;T)(Xuq{nT@{0
zdeb|zt_+{Fyj2!u`ks1TqDQQ|Dt>u%(;kDw^HFDf6ho)B)}FfTs!^1_G@BBCM%Dh(
zeSBw8)SjKe6Qici%eT~7-alrzxcEr-M6Z(TKPzLJM%)M;I=s2?P{KWBi7dHQmb7}w
ztWb6h{my9o*0wiHFa;wk8J|i$Q|U0hkQuTYVn6nEQ;(l-F|RR&HyEV{JbOT^7f5lA
z{`zvyj@l!p;J_#)pe=#7^a3ef{Zn?>thPu|f{Ib{s3MQr=><|oG$a`Xmrclg1B@+Y
z{6*65B2npu%m|p7$Ouoy>kgVtK5(;BSSke^Qox-T1kMv*xiMl^&53IQr5=3h!Jiid
zQs#d5)9t2c-4<g|v)vq`j)$m5FOag}pTRj+wcA!%fhnWhByKlJkX|5VvBC4aC*3zP
zO~IZ~sz6)Cjx7kJd}z&2so>*27L?Zup1n+Yf<Q_}%0DgVpX@eOf}G8=ohr6dwO$}4
zQ~3|D{9<~nDR7LELHlRWp?ZOooRTDkN7_fHlwiOpO+?W|?DPUD1rNS3JGy4(OH(jp
zluqd13E_HylwIX7oy=pl954oRMyaG+CAHBDr0hwoGaNEFXN?jlqm)ubDK*s#r0nm1
zN*x>)JW0??HaKLnI}!v^4t(-m<eLi03zQnr)-crx0xAC(>zwmr;+omUV8IqDqslUB
zt`|u8bRqwhLAB=`VU}j_X=XPq2&5cUCrva-92+K3a={^&?L`nsslGdBahG<*1%c8A
zK5c9kK_KPCHUC34+s40Z3jNq#y2#irGFdN>a#D9b({jhgeM&H6l&eH>m8kUsDW}h@
zm|kc+_9+i0jM70|I*6xU{6@LFFmh>I2X72^jH081bTm*ekkZyMvT~T!v=~!RFiJQ0
zbi-i1K+5+s8g_l|dM;8}sFk?1vV{bJl&40+zwdV`XQ~w_*(}Y(u9<l01yY`e>IQz1
z={Z`MrG?nE5Jz4RNRcg{J1^-+YnjZz?};ukfBq>V=QF`76Rgos4aV<UcENh4hn(Xh
zTp~QG4a*+8ufDvM<3D!&xc`li_!&!A7IkoZ5w$9!wiqNheh+ooLp>0`@#SErpb_QO
zgyU<8vX*%G86`|kTz={VhA*Pf2Q7>tszGT5i*mHrl^m~i)q07`#u>&LXR9=eUnrZ?
zIR1VA_usm*zNpapY~nVKpJO=3xNf%jJl}OuhdF+w>q;+igK5MtjTpbo8(WwBF8qn(
zOG!v6i5l<pFmi6?{5c$7L!xR(4DO3#S{RrK0jLFzKMLxj;EaKi<FA4G8aQL9QdO3H
z?P-u$?ZEN2GFyG%mB~2XQRXN?J~AH(ikHPpP=YK$g0hG#i%4(|k>yBnA(0hIaWRn<
zOK}B}RY-9Qk+%@;*L=YK;g~VWTcx;z$UCICi^#jAI0fV>Ao)HU<k?c33-VkkE&+Lo
z6jy<~N{VYhUL(czAg`C=W{@{aaU00nq_`90ol@Kl@@^^a0eO!U_kz4vic_gPl}gT&
zLFE}DHnF)FF{tt8Vk+m?laTc!6zxXN>tK)$v}gx%z5xa`fEMilb1k!HA<?2eJ5;`C
z+j@O>9pzX|alJ|gVYuN~5_Qd^gV0r&Wm20`>W*O=S0bQK$T*|xt?@fQAM<x7=2Pp5
z^?Krid&HSQq0i6wMc>phrF05T?X!Bs+3k}cf$ixgvARiIF=cWrkVG|;aTsGb{uYV4
zC1P#+u5$H^(+MnC?4g79P%SR$7kwxF4cqu3Y+#1_3=j0oz|=qew54k0brwV3R=#a7
zy6;q{sanxt(}^;j*yAxczKRT}A|dFKLnCq{j{f9q@?VP%JAY)``0Hr{fjK2{zec=8
z)#g+8eCmz63o`sF_I6pB=2xQOHNs1YhWV67VsPWXm&BUG3`X2fO!gDC7$y!8lS3j7
zPd_*8QFGExb_eamyq&mX;Erlq@A~&&v>~r`V!SuGLSIF#_+Va4YSQM-oIZ*pq0gQB
zQiO<w+=O^kW6|)!DI<~$IX*}o<Syzz$~wwfMEi;Te&QiLtNkG<&IH#?7$Ankd~nJK
zZ%oiEw1ZPEcw=S2@pr-LE_h>hjxoG9vwodf#s5@lO#at@ZAg8nvSQKG*HPOWOfRbU
zwLau{pP@d%;%@z-{l<tWAUZ&TLdS%T6UX9A;w(j!5S}nnw6RsutHy{Zc}(&+5v386
zG-4^@Vlup#go*koA#NqaOT-l<qJoS<d*xVFJ1_<OMSW$%z-;ix;)-Mb98m(J(7`#r
z974(=Qp8mdQ3a#0Naa`w98?eC7-~5F5`<iWND((fP&0&MP4(`!PfbQD|Mu_KO5Ee-
zjdfb^*m>M*O@bO=AxtRr(Jm*gDBCm4JYB|2;tKJ(LPD?z?sKP@b}}8^X>w1Vw*AZ0
zfIl3MQ8{dB;Q@h5PYWuv*lTca>3>C&_iyG5X&o6m=e15e>)f%3rEMC@;=(O5@fL~4
zNILt>r<(_S`@t9g(+Ti_W5C>qDDDz9s)FM?iK3IJu@>c6;S`rjr(-3>@ms0kR%(k+
z4u$?%ga^kM&oY$oCc*)<Go~ux34{ZGwsExcIJZ#E(!>!6IReq5g<ge_s}PM!=Gap5
zL(~}ezGQEHWMJM%*Vj^I;?T}z;~Q64Ge!7m{6a)O4$uThP-sACq&T)pvnoXN+vI@c
zNJRLaW5*X>ad^rN>wjs^+;Q|UyRs)>{RFi5+S*XEE_qDZ>2KJ;K4P|y*rCI3s`k8p
zIWaMT4QvIItzd_Co;EyoNl@dH*=!(-1Z9y3qw;n|$Cx|C8U1RD(@rd3&lD1<4s;Ut
zJKB1*Br6t@;e{luV`1JDr#l=w_fDeRNlX!E#Lx{LfBj`>KaSNR<~pzyuj3%_2f;za
zhk-vV#SLKI0Jh?FH3HuV4kB&>zDbHJsl1XZ@lA>ETvk^Ik6ct;UtJMcN~5u^$aoNS
zdBMEpy}zph&?1@Yg%>b>zx8y#zbx*p`;Hmd4dQY`S{oJ9z+yTA^D@WprT%+qIO6TS
zyUn(AI_zNHR7M7ukx;xE7E8%n<s=d@#uc`8;bo0WbNmkKyMqSd*>asW78I21^J6#k
z|5qw`Ee9&>Iv<y^``-X=8^9M0BY*LC`>VrZLSM@v&@P!8g%$vwrO4`j>GgHBKd@yt
zgU@CN!)1$h`n0S$u`=)vM^C(YW@`6{;vV6}AoQ3h9ur;+LYZKg36{8GpK1E(AVf69
z2JmZu5Yg1Ks9_ehM3YeHdlYz-JyW*nj7Lv!W5%yms`-q%OB3FFlxogGfUTdK=r*B_
zp<W_dWxL7^-JIi-9h1Efp<%M?7y2Ez4zn97FQW=<RbXyrF;HlCAQ6iiHXt-PZ~(tl
z^rb9CJ24Ygv=ua4VTh=bGO#EEJ5eP^!Q!YCe+w4hf}N<MYhZCrio0P{H@uDA1CBi~
zf_oqeUmP5ZXOX=Sg8@#VZ?)h?4}SQi!8=R#j%2o%O_bTh0)rCA=MrTuvFI)ySMXWm
z0NcJW_triVj=t0PH1@3_0a({yeZf>AbZT&gLyv!Uj#d3-5wpKt)NmJdMmIg&NS)^2
zrw^Ea$2r7F3eGg*kVc&GmRLbVhLw;ITo?;Wwvf>8!G)^k-8vtosK4F!*fz4Y5FfRU
zMvfBXW8@=2@ka3ylwgz~L0QBoi%4(|G0KtRLSj@X#l^&^Sc)r%QH2z@5LF9dO%>s%
zuR2$?N^u8Kbx3g+QFTdi3aC;*@_ja_vZXi|RJl@I0;&=zt^!q+6xV>NMvCh}RWHTO
zplX)lHc+)maVMxcrMMeZ-BR2Gsvars1y!#Ur&3iam7FJosxm~34+M74LemcK`RnDa
zkEZn|zr*rK7PZQv))+3Zo5w<+(8|Lv@Ihy!(0A@|ps(0=ECG$^eO2IC1sW`Rn302H
z4QSBM`-*Lwdhimxu^DWd!3&L_V_UUp125b=#~OGx-QXp9=Wgn<oBE>*oc6vr`}<Rs
znanq4S<LdpH!sFdcJx<YY^U{YL>5-sh@wpzRvr<>BVvYMa{Q0P<VV7zO8Qcc{RVIq
z!_H=K-3;C$)`6=IyhTMcfNKMIiwAuICQpFJEW=6|YLi9ZMAa+w%}iA6H~+eD?%abP
zYX5L<sb3h@y-6PWf?e*b4adfU&Qg<Q9><ufwo}LL)CY6(CC$U@M<>qs8;f8XD~+od
z_@^nS*`h`{K8a6~q=6b@SVMS>U6&7hp7Siit((oALxOThI0n5dKNO8yke9ys4<7?B
zoSCA~$wkH7*M!`>+CF6OYvp=OCd@uwec^Z1KAZckQ}QKdA5Elx6B&RN4#(dinmc4L
ze#tSz(sYo)qP=tx&n_}hw3l??(!pG`kApDaAoz>+aTt6LgP({SVL&7Ji}un4zD?jK
zVjWfJs3{t7$4`!n?*D6&E7N!ZH7uZ(=n<XUGiIjxU4M@aG?R(V;u8s<O8pf2b~b91
z;~x<A=eHgU$3G?3Pl<<kv{c|z!CgE`DOi<)o4ED!V09kc#I2T6k5W1qtAnQxZl|}M
zzO7?BdbOGJ%=G@l)GcEJepk<8d!1d5&{W5bK3^3d-ty&wRlnA=W;cjEh9!<KrXIy&
zyR5ICb=gY?piZ8zJ@YZ&Q`gtxlD5yDZyH_YR)2)GWPjIgd*Q$AWOVtNFgEaNfb6+<
zL3U-CjgA$@cZvO7q7}`dlh}6>E$WQ*|NlI*FHI#apPi)X`wzd!-|=k9{l}}Rt&O$J
Tw(nQM&u{TtCbPlc$z=Zl>~FOb

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_adaptor.hpp.1AABBD498E3DDABA.idx b/.cache/clangd/index/tensor_adaptor.hpp.1AABBD498E3DDABA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..78e734459e123eb4e1943df784e372d615232b04
GIT binary patch
literal 7392
zcmb7I2UJu?7v7m!by?mnxJp@yG#|Yz(j|a^A|ghELaax@f=1K`VnkyoB1niqu%IHw
z8p|O_1c{1(+1ODMfq*C?3bBjCA4`nJf3{rTJMo-&<Q!)9d-qPebMKvRCTdhz*cCH@
zAavHKWea96i4_V2g8uYBX-VR|@mL_p6bS^&wpCBKG5?xEzd7lRIAB9rMo!`DW$sq%
zl4t6gKH4Gg9kc)0Uc-F$QbFOYy>C{`J{vTv<%0K>XK`IskGpFgcMtBp^(xlzbky8c
zEtVA*3w5HyZoB=N*}U(p;Jcs`CgjrO-yZ&nj~z&K><zoU<;;M_?GJ`e?~*=w?p(fQ
z^T3vmCXGH4(R=yz(3jtD=&Z?@*4Z6>^z{0bc~^dLInvU2ui{j(;lPf@$;XAehv0|f
z^=DZ8<vsg+)qxvPHN*ZFwZ+d~-;?ZH=<?&!UZ>Q!QCN9zTJD4Y8F#cVIr>F|@s@yo
zF-L!r+8hWQrt4{O)2w59L$YPy_tH}%8`8gdGG)iiu86^9BMUd2?X$1%)XsiV?+a~Z
z-=zg+{2A@IlI$Nl`CgUd;Rly2Kb;>G^4+y*I>kTU`)r=vV3MCnQ`BIm-G940LXJld
z{jonVrg)Oh9OFM?esKD7Z;<Y>0i%-~JWl(TEHLc#idXJvUvU4;rIz-(PH9`M_#a6R
zQ*&bJ#&1d!Gh0^QQapA&T%D5ZVPW3Zv8ZDOvd&7;(fiz>Yv%GH=j@bypLWFW-kP;P
zD7v{is@AP!MeWC@EGHZ1+}bcCIMORQDrk?%>1*}}1FMtEynVX|US4`YboI)Zua%0(
z*?nUVE;u?TC?xuc-Gp&p{T7-1<8|Ky_sshpUOiQyw`EOA{xIW3za0M@53S!aWYdZ(
zcZQ7F_;O#eNn@(Zh^UUGGg1%4#Tv)o9+xr1Az|Uxrehw*`}jL8UesB!^SNTPr%b2r
z+_WPdT`$(h<W6~GeBj}M>ti#HN3(Zz#@w-L79^!Ci0e-sjjRbhxW;_OE*Uno5wMo7
zmZQ{#fV!OOPu$uyUO90^kUkbuaVfGZMgH0fRCJ}|^Dpk1{)G$=pen2@8>TKm>@>tQ
zK(eoHeqU^AEnp4$84Oex0v%DmK>`6{k(b5tao_e2++=`VX!}(2RC{#+?L$`%h}rAC
zpim!6sJIx}6(cWg1uDZsYp?jNulz)c2UC?Eq~C+gwH2s@UGAK5u<6V^Lu^J>@{n~N
za?@6zGP)#X-$B<U_6FFIs-z?HbmX9|KqdB$Y*KS~p^Xensfwwq=_qvp;stRQmo}J9
z8f<`_s5lLor_mlX6{sw_J}>UIOGLa3_opiH?(rkk1&EV3F8cdc&FE43*o2B(k$Ee!
z)>fdBJI!XLRAKR#F4m<gIY^v?473%f6wchaxhX=D!w9UX$$ODaFY?t^0v#a|SgOok
z*x6HlXj<ezU2H`A$wcBz+K;A!ezxrDQ)wRcYp@Itq$=UA;UB9D5SO-X`(=mc$@z@H
zfeuqbY)VL|wi4*{!4~Ras&X5WbA0c-)EK@1Sr<^lYbxlc=33YOSIK9(4X`sECIgvg
zAX{w(DktCgUww1SVToo%ZN#RH__CTxpwpLVR&;7^_N+N6i>+l?pZ1gBo)E4sppQ#d
z_Z@TkpPDSgaw@JzuJtHHTY*YL^VOAqbXhExVLhr6=N=cXE<oIJ&%$HYuXA*z*p-SO
zBfZDSL|cK%)xq6<Q+8x6lVOIcIJ!EHQWqe;@kQ?GBYUnl>S7%#UX8@7snInRs66_0
z+l)m<ua@XybE=Ys#92tDtw5z`)gcG#Z~M6GV*{!pca{6A3lKlEei61tk}*<-WmIhB
zZWN*}K>X^^_^>|iJ#jK@OvP@lZlUS|#Dc`xbCPrnxy*pY>Wlz!%&;es?MW0oC*0<;
zZH@VlkRgy*n-@S4@&awIN(%6qVUHn)V<;HfNfc|gezwv5+CYZ&Qh0f1KQk0Wb^SAl
zVZ9aJJmjPB;h_MF0G87X2nh)1P%slLgEcbj@POf=910&B9t~d2uo2b~_8h84?$yW#
zmc_7-vD;(p^){B+9<1Mk4LSS-i=JR9hx3VBKH=#X5d8we)89Zu8wgK-JMr31{JHk$
z$?)?e;%!i|7l`Nrk#e|!jBFrd085VSueI;<c{Megk<f@yCu>?TtXwGPAyc6#4_OK=
zd59IVJme^J<RLeq8xMsG!+9u97{^2L!gwA^5GL?Y9unpu9xgz_0zNKA!eTxyMZ!`(
zu1CUpK5j+ARzB`Uf?lLOsZ?t(A%YUZ!)-*+#$i3<rVvGA`+Oo~w-b-;#1m{UWY3e3
z^CT2(D`XqUhz1e{SU+p#{@RyU2VsV-C;jV*JqQ@8*uR0pM#|?##O1LQRqVzCy0IOM
zz_9DcfOX_8HubsNT#=hK4^aoz8*Tv(I_T#;2S2-J<f6;ajBcERY#`Y)>`o%zNd|NM
z?IQABe7u{;ck}TcBHzQudx?B6A2$(s6B*1c<aZ+fosXM|yqS+%h`fc5uMqha4qMcJ
zYu^*_sDv==Yb1J&jJb(E#iFN}hYN|Qki5l^75a$Pte|YWQ|;{I`IU_JYz8r-Ju*1F
zlj!c`x?$KRqT57_VLdLhUiq8jl}fRYjd73hfFKvL8<E3C<OGM3kllp*H=zIy7b5>c
zJ}yIS8M1?1C1jr?zvom%d%z0W1K8sLmn4MjO&oR;hfltgc=YRW$}vLrF!4A{ykJ^F
zww(Bs6F<O;sn|>^_T54KB^^7aW1s$|=Lfir-A-M(2s;*W7&ua2hK(|evV!%9)l7*q
z%0}e85&6Js7`nQ^O=uK{3sGPpA74j?*O3*F4qrA-5UV#m841~H<W`LokPC<ED~pe8
z`QjfD!(Kz8Ye)($7`6?G+W5E~G3`jk$vY6!!N<1{bBm8}BhhVs`@2YVmyaJH(E|>D
zbbIZH*tGL$G;B^_lM~n+Hey)DfWCL1?Q)=o%)ur(oPatGHYwmRut|DaX{aO^FSOA9
z(}K!77ibU1kmwjPgFR>1bBH;IECB=A6smkQG!KmfYRRzck<EIf;LKf&Mi!$9Fr-*Z
zNnu22hw#6gePrV{|DwqwlISX@>rI!#W&=N^=c$IALPrc`5H_ok8<300h{c~)S_Gbc
z?^4E&9=K<Q0p9$<Q9EHX`y|xCzD||C+=lx>`kJ`Z;Ks1n`~~l9&&0nsyWg%_<@v#x
zcj~TL*Tc$QCR1ZCbXe#H43A+G?Gs%&^b>LUiFiO%G3;^Te4Gpc9Gm^y_Y3`dR#JN$
z#-4|<4|vVLy$I+%wXllYbM~!8wzVANNiZH{Vp})Qv=2l1vT-J|1}d%E2;<er8a9HV
zyv;ZZSp%hG*h(a-L^5zChCPLtQ^*kFEJ3-?p{V%jly?FUE*5lRn(G8NwM=A_iQHfY
zaJ`_4s=EcJmVulz&@f=yiydFv+4ybDq!ww$?#(#JWoKC8HvO5a=!%FUlLP{WK&IiM
zxr{)MU-osgPHoCMLAN?3GG!cW#n-*Ve9BuBU(?U&I3yj9fP|LI29L|TI_q0H;(rHd
zZUo7El=sHhit-u<BL#3-N)AoDtT>YAZQ{*Ck+R6QVZyK<i$2zapkdgtqOm*_=Mbj=
zmd~*94)F>w8N((xBq(4ZbXlLIOytn=sO6J5lxmP_1}T}LTVjxg%wUh`%8@}jGJ|ks
z=vBrb1DS#C8MYkB%8{INf(j(7;NvtbO~VG9lcr-)I@SYAsx0f9j)ORyi48Ncxd$_A
z!p)|OmJGWV8?NQ!tysF1-(L|H74iEk!oEfP{<dSo?fmw8u*n{5mB64Fscpz1>crTv
zk}p4nrKkA)pT*L%d|XDPWyAoYkzs!((x3TQNu){+1LYxN?X4DEj`}L{Y(_zFk7H;)
z74{%KNXYpOcYD&SR!|Q4uj6aucceWgZNpLOd(}i%P3*X-)DXiO;sh8>PKVZ9nmIbs
zPITLe<=bm0+fD|z6F<NZel$!pH)xpi<|Qp>hUVYTe<ve?f6_U5upT48vBl~OHniSW
z>~!l(WwOr)r_G%e+05N}jr-}a&B(MFd4SD!eU+PcQR*||9UFarB(iAQILy2=QLqhN
zn&yTNHaR+{<Kl4fg`27ih{VW<cL9;edB{{^%0reCOCDk+EDt$K9C^r1;>JVal5if1
zlf>~*yd<875+n&cl!qjFh=&W1q=1i$k))W9OOd3MkL!`7o{w9Rq?M0*kxnlfz^%Q6
z=#&s1ZX-Hv96t8IWyKh?phCLo^(3&Kgn)5X_lUp-4p)6-)xT@av$gt^?unvg;4BOq
z%|sh==mZj-;GZT9NYsF2u-6QG9f_{<$<`_?T7_lc<qVsRnQZ=zb{%Hc@o_O074!9X
zV`evBz8~xF$0l5V$1!u9FK@;At=I&}FvGTE{dPXSfd}5egE)CM(a$Dt^?^rGex!LX
zg0O&0N@+{*fM7FVA7bsB5vZw7Uz;rC45x2xM2_{NOP?#%X&|zie8O43hK#Kt69CsH
zc6LvFwR{{FvRCoYtN0_Bp^#leeAkenfFb;8zR<iyL2S^a{%5{W+I{8jQxUQEoiJ1m
zoBPW@Rmo)^%*fzuqEk^ld~r5NL(XX^lslX%kbMOTg2U%*!{xfV`@JVWIKx~pIcxvd
z%IG1~m`Y@+M4rH57`6`S)gc=g;lhg(O)r8TI#b2}KKQ|HX`<1*&A}QjHz$AUWO*i)
zrlv+@+=y&J6V7o;c2rMwpszZBozqb$9KgT3J-k*PJ!3uPknZO0PT($dqS&bp`*H63
z7k2y$`*Qdku|G$=!7JeLkFq}b82LnagHzoCJo8WxuzA%jK)fCqfcG-&4aD3)M!=-R
z+6P0J?UmkbQFAwco&CW(056TyUaDtC=&#-WoADPXILv#my@90C_L>Js=%F(#hEysh
zSbVS#cD`Pl)3$DX0kzN~#iAf^Nrp{QB=OKvcBv;PJ4&375`Qo<!&VXJDh}T-PfCc-
zOu0?>etG2bPoUYuh}DyrDM4G_>wOUWbdj2uR9GZzAN3v8`zd^8*cxP8gFL{SdzBK~
z#U_twVtkMHK4q^}I%g%_r4wD|zH9_c^l40m^~+$Ty=wl`mdC<pV6`+_RnN2B=&8=B
z0UWU`NA~5&4<hM>$KqeYOV12@Z(91>=6I`BtGtj7T7*N2I1&!7Plr`K`*OqPAeS5z
z0?bS0!tU!)7`QNf5l)CzopDKtGn3p@KWtTn=8r8(om@KBbnYHvZ)+#8htC3m9ryG9
D1x~<2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_descriptor.hpp.7E73ABD05058119F.idx b/.cache/clangd/index/tensor_descriptor.hpp.7E73ABD05058119F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..50e2c601b672f7a4d9f9b165a4badb46147bb62c
GIT binary patch
literal 11290
zcmc&(X<QUX)8FoCad(+raLsZqA~y>N3xa|OD2ND-h(^FG@kBvIgGwZc7g6v86@z%-
z5s!c<phQGPLqvm$ig@9fcpn%%5f4Oto82vE62H9p^zhRu{&Y=scU5(DRn5Tg-n|uW
zB2mcL@RY>3>7$59Bx*<hPn$k9!J2~JMj}y4ZfVT0Dc8DuasSsjZLFmpXsP(x0n;PL
zjqQ(&-V<+d^s#HtC1LH_ExORFIx^;E#nY(RN247=&oA)3ee2j$)$9khD}J3ZV$N3!
zhMM<%s@ToMO`pv;o{Afn9GTOt-KkYEj<qd`p*Lo|FARJp$t=%xZC~HrXZrS=$0wD$
zuc#<)*Eqo1recM<q)%b(Nz;V8fvJTW<)cF(vueDI(hRJVBM<tPcdVbP8d~~8VY2%}
zujwBj^=UbB@71BSo#~o+XKvhy>s9xrd2;6cwGqF!|0?47QF-<~?~8^H(w1&6bJ$K=
zawU_0J~d8q@ZPO0cjnA$+WF8`IzDtn{u+fQ);*E?Fw62rW9<9XEw0@hYVAk$&q`|N
z^I_bCrnswLDx0sabo%>6?xf+N;bn(4IjiRnQ<MAyn;WJXEY7Rd><O>DFtc2xntXCo
zV8gGo>!U~X^&gV%GsW;x^yT=n^5&EsLBAjT`(@t}tFI3oFHs+T_g%Q^?lSp{SiCRc
zY3B}=hgyt=&mU75J3aG!(7CVUL!FK`rTj8IbaO)Z`;5rb8!k?KXfFE!vG>+hY%BBs
z;&ImSMte8K!P>9BFyFOx653wd!$NcW_>GbM-7H<k@9_4BPCEJFh^$6b_D1~ZaMhnH
z)z)FA-`4dH{ob^+u5RFd^J~M@!%jxY$_9-1V&l-3Y{jsN%?DnI;=ITFxjs4gn={uV
zZWi}-m|3_tV(5wXgU*_MH!;w8)}Yu|L-W>fBNw<I586LyRl}3XPwHOASZ}}Roz?$B
z)w!1NvQ;Jq?7hVn>r8+CV_=o~c$<su`lTE&&3$5XS6q-|`19fX5pEf8QNg17HT5&o
z>??PVZ&Nb3z;NPL(R6ZQemCa(r8aHL95NbROvXPhKQSjKx_Dn>U2&TCw5Hx6ML+G0
z&%?Icd+Zyz-Y{}YaGFW!aPN_~Pd_N|U67t}qj6<nVE0>JD_2%!1Q~wcdE|knJm1w;
zYleOK`nI)smk-{nE%29zg08%{6Ztg%()8G|l1T@qELyT;$;m%1TPH4m{cH+ynV-QF
z#Y?Ogco}Z2mUMX7^TbcTGpBC9denU6YIectWUmNGy-x>^osK^gwpSFrelw%v@iO;0
z>Ggi2zCF>9n0%t(bhO>>w!3y7f7aJ2uAtAKU+27CZTao*A56H|Bjd*{OB%CyP{D%x
z{y#<++YLFgu=cO8j&@NFYt6ZmvZ>z{C$_7}Og1LFBW)JSOs^jgS?+hjEVlc1UCt4k
z8neWGjjzS~f3#oj!bwrjH)iZr^IuK3hK`Q28fsY+nj05%?9cd32UqN<tnPJneA|f9
z(VjICDHA7und4F7G~{$(i}Sc`4(<<3?N&O34*VcKJ@&Em;r#tA%?{fGIjaGgB}p;v
zS2csW9+gDC&D~Nr{=wLBd)il6EGe-XGO&73blB2>Q^$LbssCW~c-iJBqG{hHjvg!)
ziAay-s*uIeyIIkoo+9oYWA~2n(-V<IibcbFAPyOn`*iFcT6BmN4WUfGVbO2cQcpyZ
zwn!8pqD<{3g*Eo+*?yccwh9(;&B&=4dFY9VPSpCNXc)XFsc%b-636rsaS2ulF4_kS
zoBSjin(^b&hsM~N4rd{!ETqvBU@>pnR$jlT{g=krl}<7bJI%vg^aNNe(lQoz`^tNn
z0^3k2Wyrk@h3E;e*u1l~p@+fG5O=3io+0ToWTGd);<z$lws-dKbOpAjQmj0!g0&AA
zcB-?<X|7@DPNM1X1gi;Z?E{9LkBypp>++OTL)?Wb><%%wLsWVqk|>R-DyYKTCVB+t
z9bBBP!0uFrkGoHZ_5m_H-)FwJcM0EVj4kMJK61)OK6(Ny-oN)hwms>vg6fm*<1w;&
zj9l~tSh^g&Zt3f~hRAU^m2wXm-$Pb<0xZ6Mqo2&d&b4yfKSacB!R@wSTRj1m?h!7f
z!Al-bG{9bTKaWt`N61u9L=uHex1S!tF<l=_|GAej*3hZWA**x9R!=}q&tJb%zJAn^
zC~+)hiMNV();?f3cuMz{dHtHa71)IiA4cwnQ7=6KmXL1#XQIY?Tg&lK%5oPO-9;99
z0xZ4a2eeD+w5Sa$>P;70P3)>kpq_{%hKwjMh$_EVRniEuc)?XUj-YbxBjtVMtS2C+
zci05^*_YEC46rYiW6{>aRQnJ~OxktR8nE|i)?#8;qt|jAMb*Cu8!w{ALPtQx;5x;F
zqg}krsi)CtCP^omXdf^<y!h!3kzbEP3hYFOBi$n-wGS8`eS2^4NS7TB##lusNW~pe
zv74R%%S8W+DI2@Q-cjOV)OdN=B@YMd39wB0doD6f`eC*S?o6dLA=@V8r6<5L^-RD?
zYjw;{6YNLVG0S$Am-Yd}-+L5ZZ~k(}C1Y$#hYOKQA@b7`V3~8(>8oGn53n`Hj&zb1
z<kEsV>j|)=2Fd1J&zgQlw~j_++ep`;Bfzp?$3lm*aV|qlun*l$Hnz>C$6H5$C9_9W
z@~Yj@gAK4ZwMqtVn?aSXBO*zAxy}`r#db)S9S96kU<bNjYftNN?E_?FTTVaNw*8#9
zI=y~G?jKQiPDehm<ZZq5W0}`7Lu^JROqNdmLi>Q>P4Rng)(m~H)&LKnt0=~8i?K>i
zM3OIzJhlFj7ikvo(@<v*W9&wEvjMwrz&-T@<m4^Ph?#Lbn)*j4s<x*{`V=Yk1XzkB
zal=PhhNmg8ny%y^az9ApnT`O<u2Tjl&i3>#pk78NT8^d5v5B4l%kiAwirYK7zM&IQ
zDT&fVrS<{C)r-p52j46y(mB8k+ZkS5>*Le#vi<hr$xVli@!(JqSAv~Nu!o)iOU<J~
z1IPJ}RmRwvPEd;6OOd~x088!co*DC3v{_(`J5rX-$Z0e3)DvL2c%tM%Ud!ljbPli*
z+peSzpd-Li_vg&<JI_aVHN;)1>TeK(8^lac_}F^oS?Vap`bCr+_n{iuiY>QdwVr^S
zYq`po4HLGDS<yf$XBA_&ir#!WB9gE~cd1=}zw<iNwumUO9bK}$r+pvo17tjz8*aUT
zb#_x=TROZHc`il4dIBuZ3b*$5TO89zfjuZo6>_hlEII-#&&SN}GSvOIrK~8d)tQK>
znE2}nAFy8BmAM~Ftt>Ib=2T9SG)bv_fQ+}__xWw?u9=sOu_YZYKu!gyo1Os6y9-N3
z-yUlb#frY9)08uI<xHTSh$IpNf9);)zC*7M;>d4QI;Q80-E+#MBO-~ojZP~cH3psj
z2vTh|#vQ13*J1Z{xSO89WK-kDP4lwm^9v-^XBE3PxK_j+Sg{v!3Ob-mC}2RQS-5sW
zw&yi7mh;#6cNw(!&G7SQcVA(-?walb6rc$Zpb+a22c9z|IHV7c!q_kc+_fwh5fBl=
zqsab|(GY}LF3K^=jYnC?GYfTvq=e;Gp}<wB7mrsX=heua$7_)D8X;bboYxBRMr64W
z+3@@uQQ$_Qylp6A8;XKM!qVtDpasQ3U}d>dY*dPs)%v5za`ia49`^w~vRng>X~4sI
z+=RtVSkB`-rc)l%dEwl+%Xw=T&Str7WZ*V37!p8PX?}%sn{N&|vYZvMVrd-J{;-@i
zu@)eEVlO~G#7BT4Nu&VHATtDL7MUeLOA%R$1UL(kEFs>E$YvqVM<ic}3lJ#~;zC3U
zg}4-vQXwuwq)dnpB63iO4<m9|h^r8(65?}+oD<?kL>h&-36UltZbqb8h+7b85#o=C
zd=%nTOj5BxAM-GoC&Vi;St-QXm}Cp_I!x9H@divb2yqD}B|>}%iw<FZQZ5ov*Jebk
z7y&M4MCC&KoDn@2;&+Vboe&ohQ4tZmUrj{SJT@}B5Ue@>*F=VJ&vDpu-e(E-0*Af$
zCm!(vj|87JP9KwB(R{T7V(E+K&#?vupw^(lLsw`(f(jKz8<q=eAJ!2Lx>Q?dKtjiI
zlk6vXLTbixKOvW&ga>^a8n6w04F?epES*bNl7JY~%Gah$r>CxTb=#29He?5Hm|fZ}
zZpkb2LWJ9m9Csrp$e*Bk!7N>N&t?<_`;h8u7}y5XFmz5`0RwYdc70g4XGwYjl@MVR
zVGa9Zxm%3;EgGqZUb&SJ^5|%X!&?g`Tr=%tz;a1uNp^5|uv`YC%3z#8jVzbRs4|6k
z8KYVz#LF4gav@&9s8;aUx+c^4S-|6+49m@9I?iK0*@xxQ8F4ye#!s|}5ijB~6kv3J
zy6OuyZX2{`{|nz*>1V2gdJIdobr|S_<>EWV2SQl^RT(4IR%1L<qJ5$UG9Z?l>@vAa
zwm!duN{y~WSEj)=<a&L@Sy#8BE!4V~u=*1AgS~*x=_IW+8$U@>r=(zB({qv8T+|5?
zYqw?lPS4)m<D1W<bsF2^Z-r-7N(ndIb9iUCeF!%nsplhCc%N`-$R`bT<#9UlNf+W`
zWLu2v!P5x$2X^}dyThUhSA{iI*o((&7_T*qHzYxXt6;n;7;nfC2zQL}IL3GacJDtn
z%xlMtCDblwnZ9S4elP%4D%GE^L<N07ZAuNMD@?(}EH%0PGvo<A02L}>xfJshN6183
zZZ4A2!UcBBQu$%|Xb?;&)mO1FpRdDc(I}%4$#jKzjB6g_1{WQax70~<urvo-LXp98
zo3PO)+!4}rslMoi3Zt8G<@(Y!Gd|EmbODJkAUPDcEO!x!FADJ`#9l%Qe*6kzuL$uq
z#9kBPzmWJZ{(aDlfrMk*vZmmH)KSeRtHDvF`g)rm)s;^0GOI7?cwEXDmNJ%LVOShB
zl&&-f#lhjB+Z=2%*akdMs;$g{iKYJNHlO!5minV_8Vcd9la74Tg}4rx)FB&~6)Jl=
zyRNJU!BEfB*;Aw`O5XZ&q;r(K^o_`9BeH-{OmiXBuM`c2nuFyoBla>fh1sO~j=-l$
za-3tc_CBZ<v)mSjY+=kmDJ)mQh)Wn1kIysYJYxplO_MA}T+gU@oK8qOF@vOu<+2FL
z65>3<=8=y4cs^nCc?_)yss>$O0;~z`h}J{kk`lnHAr(nekrl6q0z|X=5s>Xm^(_n7
zC{QBRwkkk6t1aeWkk!`qJj!4UGZ;&V4A9h|26=^~uaFYj2dxQdXbXzs6Vi_;@*^4o
z#WYJh6TYc91ezAqMX-Gu_JB)Is&9>egNB{Y^<OtMbQ%qbaYk|G;CEa^QtV#qz(4-y
zw(8&ai>)J-c=dn2+y<n?v5S(jX3~u{;Lrvf4yqlr+Tlfr+W90Mi13K;gNp<b5$Y7J
z4I(hi;CHJ%@0ZW>{?A#{@XcfY?7gcwnOe|M?dSm|EX%2#)$Tkxfz>Cl2ds_dPGR*a
zA%26qy}>>B@priEI~)KQ%tf{1=inC#)s|G-zk)5a)=XBbPyn1;a1FXsT{8=O0lHXp
zK-a>87(B_WgN5VC1WP&1*~}-|fvQ+8(LB+PztNu}^QXuTnklU}jQKNU2R95$dm-k_
zu^qHY=(2FYDjddN1Y0rLisjHWpcex}HZgKODikr|B1Qo>3d_|nq=u36QQ;aRzQ!mZ
z`=aj?b`vrALAKm;MBb~>EO(PAZW0xAE?Dj^QQRde$UP=i=Bd}MZyNU5)dfcn;eQ2p
zhzE3_wOI%Hp7wCViwcVz$!A#xqy<#Jy6zpkG1I`qcFu)#S`2X>oJSOR$LuMb-QqtQ
zqv)Tn;`UduBbXfyKNY0w@qq=Qxkp#@&wD9#O_91KxhR~9dd+ygW<sIE0NST-=sJ90
z^95IDoyrayJNvWAmqal0A20jp+mV*2YAX$-gtSCeTWg@@$8z>+dkt)aZfk(YU><1(
zYKHL0-@@MpJeT$tEV>KOK*hjMF^S~{i3b^h%d=dxI9h<>mGLULL|HCDnV^CygykkE
zC#WDRWoc(fy%cGnX~c4?kzzG6htt7wYmj1%5EmlbLZk-cu{4oVK1M3YtXc!IZ3|L^
zO<3*@mfyijuocVQ$MX9^+<+q*@Bn^1myzc(O5V-aGxGI9T*=5Q871%LM;Q4LAwJ5;
zj|%ZiM*fmff`M7?BP0LFDEZ59E|JeA0=$KoY#~<sG1)`pdx#R+VM|}^Km0Vr`rc<b
zXU4Za4-)nphkWKB;6ZI9oKagNOSt}|`u;FrR^L$scJ^NmNaHqR*uYrB#=*5{#^2gX
zhO<L$-6<iIw{<2m%0y0J-4*8^K1(hCq3E+i4fiEILb{eS9HH!u%wI8cf7(ZDyEEA2
z40h!WRf{cZ`Pd98A(hbDn1=P_u6MmXvog8sf4(rFnN3r5T_>AYR6Vw+=O_Mo&jR*g
z!ufo<!Wv{*gE~WVldc#Uok323H_u+O?3BD~=x3XUT$@T*hv~l&&`Ol*8{=RTh~-r0
zx~@2AsWs2_YeKysHnKFWAd9dO<Sjc>%G*RGZthK;qXCN>urX*6`sh?^y6!pXt*FDB
z6Xdti8|fSMSX__I;f>;{3nr`y81#}3?7_-C*a>=XaFNi(>6-1Z>Q+w*$j0Fi+i1Rt
z$rfx3HAv~--<O&$enHz;|LP3v8+Y?mFKP8HM;h8>R;GgPo6K5(>}B=><RkMDph#Jy
z0L_ri5TIGISpu{a$(AAk&O)**A>NE+n}s+Z$?}D`0Lco3xDd$-g}4;SN`<%#$;yQI
zAd(#v;=@RGSct2TtV)Q_A=x=0ZbY(1A#OslCLwM{vSuM}L9!Mh{)l8Bg*X+<Qn5fE
z^RR555U<3tl|r12W!XZ!4$IaF@dhm0AjBnDRwBfQu;dVK&)awvBU!}=a5*C>7vkrP
z<hc;PV<hi{xQIxKh~WKdBB|!_{=Yl@&`%YVL%p~W`!w>6jPmX_?d~pom0?H=Kk;A%
zf4vQ3hneuG8i}is0!}u|)gf`6u$KLWA-@Q7rCNs63Ui~247n)8_ZjhhMgg&vo)|*1
ziJZSA_7Sp=$l(%TxpRb^BXX#m4j#Waq|m>y^-}m}Z>=I?k?rHt-&AU&XP+~;ADm*A
ztHt!+r)#fb>YXRUTfUE?5|$Iya?$~&K0F$|dpxE(R2w-=8~El)rPFNBu6t;6{@<;I
z${z7YK9nc#`p+f8v2$lno_zH3_x~KKA?2k(to0ings^IhW6p19MSIbrDGeK?VLK?g
zSZ*hy+R146)b$GEdW8w%Q`Jl&&Lk?(0n6nQhg{;v-^uHV!+Ig!KpZv*aRo81Aof*3
z14q2cOz!}ND^2QJzaM^DpZ=o-{j=`lAzV37!PEWdJ{iIrXSzGi9N1sA>N7h-la;=q
z`+^8>oUKcab?sO+hZ^G~GCPSJd6&IOtZoulVft-#m$*XuebJ-g+ON?gX@i~XYSq;h
z$}*OFf}Nk>F5pM-X_D?i_dybZKFFmj(tVAD>+^EjB{PF$i#oc8e>Tl7zswyNzk2b4
z|7?D!Q0P)~utyFKfQv?||F{Wjf}D<C>AFvvuqKwS(f$VM28+HPc%*&Uo!O75)4yh7
zUNb`>dxI>QN@)G;30kHpZKnynAf>(KdnV-dPBZkGmG1P8VfJbUS^wvgap!%;qU+07
zW>HUe(l`adDp<}%<03$A4sPE3*a_U{1datmwU)AdP6_efSp7Hl<;UM*m$!Tzq<J)U
zY2opsU#CqPKmWV?R3#ctGXSP(h)5g44hvfIKR;uTVN%QJzBu!?DMk*(s0$PfZ~&-d
z>OMFFBLvshgirgmnfH!s#PAyF4bk>R?*0moT{iZg*%$iUGy`h=6b*t~^Fr?$6a+J~
zv`I_uAtUJVHaT3|>y|$4Ayuurr+X-<wkc+hZFA2`=T?n>r5Oc(e$zSWNS~sr9(4A9
zMrvq((T(Z8bi@2`tI%cXzHjpzd&L;OVjRJ{X{|(@?~o9P0q-836<tm;%cK5U&8VvR
z#y1TnjOwHi&m)%ehz0^C2&6OWzLNicXN1I;?m_nf9ggnQxJjd0bAxG9$4v8R{e_ET
d)jn6}9>2r%i|i}v+xjQnPbB;z67}PM{Xg-^repvB

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_descriptor_helper.hpp.B97BC6F42C53AF70.idx b/.cache/clangd/index/tensor_descriptor_helper.hpp.B97BC6F42C53AF70.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5f4f3ae090bf030fbb4d66a57bf9c4d698d14451
GIT binary patch
literal 2568
zcmb7^c~BEq7{K34f+ledESMpNB~<}Q3`Zi+Ac<HILglaoK@fBhK|~-T1T-i*0s=x*
zpwtqsdQb)NtU?i`Qkk|Mc%t#Zj!3JFRZ)u+$D_4N+Q(-2!x?ubFZ;{)<-6Y(HczPp
zYyjpa%+souXT~uA066*4WM-_ahX90*13+7ASP+zABel4v+4wWz+ZY#L-rV58%W&G<
zUv{tHV4b$ecJ@`f`TW6N@sWX^mbx1SpO)z?+<J3LeVV-Mykp&a?&$B28*s|PFKe^2
z_xBCG&M>qG<wiSQI?;|@x^k&Cbsz8W>G_)qFI4tNKZ#l8v(nkN)wMo3sBiu7)<|*P
zpIXJ8BwJm~x|503i$ZlTT^(EF4g?3?_KMssZ{N=6r0pp@<Z$D~`hQ|mkJ!j*VNtT&
z($4py{>Gk*+|@4I8*bNv#alO}>PlxCB+oxz`x(CONx%PWpL4^C@-zI!!uws<yxu;r
zh?f6eoUZqr+}<&xCFs?&;j(MHLT0Ypt5UwnJ$3Tw)8J>5Pm2uneJQt7|E8OT&T0!3
z->i!8ZMz~muRWaE>u%UkzG|~YotyAvV>I^<AFb@{lpPW8CW%T4{H+aNsfKw@$EF;Z
z)hf(%wd{G6=M>YOn=@@i!TH)G^-Q77^CgmFbBB4q)V+Eikk#<u-h#)yg<Zw^?(4LK
zTIXMzmZaMl>YGahq0h3l$3;aubpv;I4f?*xd(GPE;<*rLvenDBkUdjyX8#ncw9m!>
zmjmzwkYHy#0BanLjRA<mFqg&kIlwUr0~1LFt{vCUcmS3e1V)7@tFzLjhj|Vdup$*C
zc9J>91F)Ek2}Xq#=9QYVAqOD_CXovH5adJq|GxlhJe_4!%2=MLNf{3St(+!3*w^C=
zG2Fw^LynA3lx(R^-XXoq#&BzY>!8t~aK4x?rAQDU2%$)&Q28PQ7+wH*1yC>=3|<Ta
zieU)F70|i@+EQEvg;mgn;u<KdG3A|*xf61bd}2#uhtO>G4-5?F0xpxRtntBc36Pjj
zJcu`;JP7jOD4X+c_)4zdOQADxz73y1c`$H+zd%V*HJnim#mKKkVSHo9n;rq_*NgCS
zqtaD)D1s?c6Kc27R^w?NX~Bqix4vV(YiOMbNw1##S;&u#?9t6;UExX^hNs)7yCZE3
zuY}f>&=I*8ocXpX`L4Ay*_Uu4+>vq^K}49)LSmr_MG}!F6h%atkcQAWQqD4oOcTl?
zvP@_VvBref5^G0k+MMRLLC>q~F)2AGyp!9~HgwV$_%=9o8?-}7Fz|9Xy&MV=dpaxw
zaptjEq#|F$pMgrkzzH$oga`!**~SZV^c!)<@{{G?`^%X=yADY-WZkS7Rsw1cyv@8n
zL8J^w=3E=nIFL%=6T)q%s%a)@Y>?rA>$UQlXn)N(@+J@-(<vpHNM_P!FQ%6zB}*cb
zrXvFwo-9f>>AIS!wnVbPs*cjj`A;X3V<S;Yf+z!0rPPEHW+b?xzhLB@73V_<suC&$
zoo!qQr~nu)Fc(OW5W`hWl?a&+*E<Z(6$!gWhRQE}K4n=CUG&fsSw`wKwsDhCmUqJm
zyQ$fWdaIZApUQXtsQZF^kZt>K6CvxVoeiRzy1!yR*mPYWee*!+drvxsyK&sOC{PSf
z<Ro&bbc!IS2y&6D^6p13a!;*4{6WjLb>iQ0I<sUSmCT-vy3&grng0hZ?~m;EX1As|
zlN&Y3F=;l6@}TyVS*UuClI(@dU8bN?Np??GAf=Wk>sqW!2Ajs*q!WqvFZ!`B^o=Y#
z7S{@mH91Qxi3!EC;!P+Ivhv_4w>@!QA41Hj9NAmGUP{WnB1REQk=j?SM8RNqT1;9j
zqV{u!MTg`sAC3vF({ZCz)*QI^gC#lJeXXnO@z9SSRqQK@={aFv5MPgRXmEXc_jtt8
z3bMR9Xr+TwQQL1l+_G@rbrB>5#ZXuby^-MdP;FR}zNF}*y!Yy#$9=iqd8A#cd8!3e
zh0?>)7o(c=d0v_O>L3q}330Htt;6-`RQoYMgN)^=aU;FcWW;M+M_w#Oj{hEW^a6k-
U^@^8<RvhumnG{F2v@i$%0{0^oZvX%Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_layout.hpp.83E496410F43D7E9.idx b/.cache/clangd/index/tensor_layout.hpp.83E496410F43D7E9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..413add1ed1b447b1b6824c3a58fb8a898e2b9ecc
GIT binary patch
literal 12068
zcmYjX2Ur!y8oe`DcSR8!Tu@_KrHD~1(O43V@7W9Xg8D4jiDHjYV{Zft7VI6PB6eeo
z*o`G><QX-w7sP@ZC5c#gwl{~lXT5x%^M8B(v;RLkJM+)%@wRT<pur~<9FDrZ8jp<b
zJv=gp!{I2P{tX*GG%CR9a16=saEx3&Yf-EI*DK~f_2NK_$bO;ahZnm3Lulz=CwuR@
zA6c}@=^3q}cmI7i;^c@<O&)!_>RI!shhbhHVpa}#v?xAf(2JDBloqRMuYcEaeNskD
z%GL0p58gaHdoS}*n@+FSmzbCF^0z@T_x{ZK_lm<gE5uhCmT>6Gt#>2BGrm~TX34g_
z%O1qMOZ%egxl5ON4Xd1IPt)QLsvPXT|6b(6v5GS}qH^T|1A-@eIa{2b*(@MoW<=$;
z4INt^rj|<l?cT6`DgOIs^gb|e*2t0j9^U@7*SKm2m)$>Iy<wXNV;`Q0PCXNt<F*lQ
zxJuU;6uo1_{S%QTH*{LiEXUM{lE1~=tMt5Nv8KM>U4Crd>*rY;PV|`DzEr`hNmoYp
zoS%Mj`ni$qT)k&4im3C)?0co}_c>naWZMbj{%4L_R)2KOCE>^BE`Gc8#+5^5{~1}i
z_NrbX+nSy@Ga~E!)vwcko>caas51@jru45J{N~}Y)Z~1p`feEGbX@eFI4L<`k7Hj@
zu9O|8I#0M!Fh|z#?j_Ge1|*j+*1@+*{#*lRKASPH|M~CpDy&iD(H=V<%-Q+L^AlGC
z0{yEePwAMs;pD~%KSaE}Git##XZ)R8cmDV*<;uy3g2u3lODFB{Znn;KbbIF6`$bZZ
z3=az`;rMM=zCL5iJ^Eqa^1}^V`i%JglP`CDKjB2AadP6ci4BAy+6Qhg8TlgDr&T^{
zR;}gdEqB*lb93Lh2UpH~RsPpX171Zo`}N78$9pz**|@uF)~t)0ZY@l0lht+a`=j6N
ziyrXqN{54CZ`Z##crdBUp%GouzV1@B{=xDq!v3iF>_O+t9fu}mT-y?Mb8NfnwMNg&
zw`otiy$MS;6nR{3VA<=>qyMb(dgaP?r<32lTroV@Yiy~0@5^4;@!PeAQ|qk`?c48B
z&6lN)I``MO8+0Jz+y}p!d&-n9yXW7r_0|>HKcvZ;S2urOP^4<}ZL<e$tvbH*qbs?J
zAK99>Vbt(HOTX%IZ|2xyL+YQ3n_RKx=MiSb^*KhLJrJ9D-}-ENskDfQheOh~G#J#S
z)=!TIzUa0(;YQTJkl_P1z3P!ru>Z!LotB<^*X*yOadBB*y&DhPeQ`(mVm&LheBpns
z|4-iy8x`H-A76(<mcM);{Fi>~3|S%EF<)E|<^_?5*5T!FY!VKKm@kC0tJlP*2kthL
za*=4J7MS{*FT8Vm`+$VMoJGSO=>cP6{p1QUMETB2>#Pi-b>wohZYR2m9GcSxtbU4>
zA*>9cSlW_hI4Xy)p3-6gSc4R6qSKn_R4i@D%JXc-!uGFvoiJn=`o3M5+eL0#bZeu#
z@7~|OY7zC{x9bfV4B`iAevs-RwFR-Peny|(m)f56ifN6sofL7Ww07bVv}M|S)qC7Y
ziF{rUtnrHVMp|#c(w40JZ?+Yk{bku74EZH;)-~&@jA-5Cg)uGA?n1x8YrVG{vH&<c
zq`5=prS;@xtXz}h(AQze7$+1nSz5_RsV!3$Z2f1h=+R00K^v)9SA}&IW2-G$u7MlZ
z9^R6xsv*lEX}mDw@y50FNa`BbX~6jT=?@L*2i|OH&PM07_2Bux+VgbVrB`vt+D;K`
zST$5mw9HuaxiI@Qj{g@JqZQ+pux_DW+LBfLuVzng@7P%ttgebRS6XxNjM|b_a$4QS
zrMBE3XviAq)nIe5>JwUz9$EFjdk}RwBomQlVz~a6*53%FEkgyyc}*EVet8MdW`Wkv
z>ZeF(NeP{?VN+PQRdc-^E7hb|gXHvyce-dDUQ#&LizaxLpFZiDzjebp;|%GGer=ZK
zX6a4q(YJETeZ#)q*7b8k`hz!Bno}_qwFPgLd(M{IAR}Uq416wS{KxfHICo(`cQ8l!
zv~x$pg4328(g%^so8?s;v>>c=M+vJJ7vB4?UykA#dns0|uwpUv+VU~OVj47y3r#9%
z$WjP(QJNQJK3b2?f0n0VJ(;KJVnmya5vy<2SHaLSRF!W!q<o0(yULJW2z6ALM=`s!
zMW_w#o>d$8>6A9Z&U+cMGKk5-OcrXuwDrhY<8Hq%#}wHYgq%~5^QE+2;$>>fyk91a
z+%mUewQ8Wv0IiNyN0HEyQhQ~E`$gw3-)G2@$QbK1W1VX9XzP)&&UcYM)q?$MB2;fh
zye_TlNT@BN)fxM2W{1~pUW3(2u{H^76Gm2Bvg(x$YxC9lyl)Iy7dbP9nTZmmEpjHf
zhojyXy$+eDC!R872@szO^Qp*B>(RM-@tNtZh8DVI$o$}SG&`zCK<mlb^RB3W(LpQF
zwYeAzKg&<`3@!Ur-}h;OghI~IhAf0IZOyi7?9qCJssG{ljbg=GUod0?aQc{i)Qdz5
z&N6qun>KuNI_Jw&7jk{4SfN&^N`;oGn!YWWbf@!#O<+VT#yD%7VxT1>yk4W-b5g(m
z+mO|fC`+1ISUcK!bg1RS0!yRLc8@@$9*Wq?YNaBfWu(@X*01}#?DTtvERIO&!b}%x
zB5Ugrscnv?RbISH>xoFyvnQ3XUgGV&%(e{I_FU0TgYNy22-*bDhFU`v2`wq@d}>_#
zC+%5tLl#5EG-0L*A6jJG>>j9gHCqiiIOv(`+9x1hmF88cW`edLE_4&yM@6J}TRHKV
zAsd0XP?`%t)E2}a+{6z5i>udS&R_M>yKag(Q(7}IPqk$SBYxa{zSiO<8HOy3a4pOh
z>Mf#0sKxG3of>@@^KyQzzJ~l4NbiOD9*a?15Y->5T$iy2wl1%Jpd3<mQN*p%+KQCg
zGG%w4sZCp-nB3ox2BNJJ<|@<xZ4oWWown!wIUzI4&3JCe&p^B(%o{>YXKg*w_MQ>s
z8aiNTPNZ$4h`$Ky7rZTPnYQ2E)bdSs)|?Gid&R0^RZ$$YHX|)G?p23UqYPOe`7)%L
zfuYqF`9k$$Mumo7dzMoB5+V&##0A1yfPC6AT6FW_wKrxi4>Dw4M5}C8Rzrr?qvwO}
z2faHqWYP*l76B(wnu#d)+Jd**Jrsk(axXl8^H6rt>8Y4W(n`X}YRi-(0va{S-+Fo<
zLzY3f)6zVRJy~0XTk8%tqGEK3zR?v*7*d_e=8G6%#-Q$K3*rJdaa^nSUj+G-&4owp
ztC)|4^%#$;E%Qz))&A$LjRMPn6`@#9r1b>7)0V6$lRG}%5xwmTuwt-WR!D0FXxiGe
zT_Yw(^$ss=$S=^h)zVyz<*u#A*iEfau=dJj)&PVXrHBWFbpTz{meHn#U-;LN(sSQ}
z^{ry<71my`w6$6LrWGqHQXff~D4Ob%=zL-1HvHZB9Mf*k52;(U=oUlPLZ>pNnTZ#r
ztw*O~D{kvH{$ioRNEfMysnSZtBWlZZvEPWns}l12g4I#6I$NC;2Q3*h&dv3WFTCis
zAp?<bj5$VC4YVG)=B?TM&*i)oZX-&RBF%7GGjIylmXYF9p0EBY?oBdSEfs6Mu-0P$
zv?Xi)z<MK^HBFcV)?mdNWsOoCv}7!(lW=LtxiSk4nG-$gW_I(^EqcW1vhc6=k9Y0v
zm0fc>C{_ilf(nC{VHWLMn&Ux($_5gRRg4l=3B^E5#?nf=l15Cfq)xrHk*L4fU%h>_
z9*@5C(51z9kBsh*D6=t@7E5a}##vj2TAtu@{kk#eCTJ}bD?wNZcy?`V)|HqRVdZ)*
zF{C#VRx_)qHG|e8;fk+obX)k=E5MND!C5TK#i)SVdhk~IuKni6k?Up{vLSdsN%JSn
z6K%m;<6c+^P5f4LJ$|*3Awxj?*8El_MGMk$H)(B++EL&A{iu>53xae{n)eV*TaT>k
zW<K8A!za?mkU`+3NHayM`J$~S?|Md$$^ogthRhA#NnxJEQ)uhKTOWMTd49#ly%@Kt
z7~W_rS`8ao4qBpg=IFoic`HzcE5=e`Eyd7kYqJJyznt}Gvo~1N5a_V94&yy(OWN*X
z^9KGptHE~A#wb>x6{t99$=J8N&ep{j3Y0ZuCG`BVFfXI$+IsYS-;o1NFElj98xAK9
z<hjgT{<`(xBnPHkzLQ)-9cQZ`R)`s*BBAx<TptuPJ|kW2+qu9=5@r&%9&J5%`@@Du
z7I-vjhapRYH^pgAajIzAdhiZaY27Pl{WVVv*?11d!N>Gb2M@F!q{GdVj6FL(J77o)
z!6piGB9158dhm|6zB6!1{-?tY=>l)MG^e9NYU{x}E`FS_>~icYL)HVYwpm-%c(mXw
zb<eR=trm1F^YFXuL;FC*IxnsBcthH9{LgfWj=S!)@EYDwbH$3ZA{7TM8E0!A{(V)a
zp}oNPk7Bg9+A9WHGJd^zbJ&&&<Fmi>zE+IM!kUaJqb*tgogV*u;@Qz74f#1bUCb<|
zI*k^cUgGZb`Im3EyfT+YAz5q1`dL^%Bb2rbb)mq*m<Q8_*D<7tP-}&`7L!I>k8!#X
z<htS<kk}s4CgJU@k=7cFfwl~Hv3SvT`#)dw0<>|8)x>I|IB0D~Q?H?mzwDt7ae0vN
ztT4}F)@bXI@KUVTi=!iYXMb9DR>TL=dVqY|GTNm#wGOARk1URSLltX|u;$<)wI%EK
zhCRJz&dxv0kQI@0w=j3(G^MRa&dYgDgtm#D*U6C8z-weSQbU2(gL9>6VpI;Jdv;mv
zqe%CKbsrhEWyY&HaxSUXcH0s3YlLE564oWKv?c5MjAGrr(#Ku}Ya#~qxU`PrA+@z>
zt4`c+HM>O((E2OZ@6!4mEN#iUmG|wu=ZVR`gEc_0ra7%?*y^+;>vkdE!l!QqWq)kM
zBIO%ly+O~kwP^<%U(Nftd#oY9!blvH=0PkUZ9PUK_4U9lZ|bxwV#u1{9h2rUoTjw(
z;61EdV#C^u?YRv3Z}75&nT0o?ttan`y$goS+q@4CF%R9#Z{=5=L(8tE=ezUGxqw@L
zfYMztVx$#=(a@Hx$F1`e39lEGV8}X1xJ;VMuq|lok?={b&9i0A?Q;y7BfIDnG7A-)
zFVK3D8vIr%ZD*<Ml;5LQ6Qnf(IkjcZKg&i{7?9A{&ye|$^N}<kVbHYo2=`Aw&f{t6
z;gj(MZ58pHu+HHXXv=8NFQn$aKD)IVvH(Q8EzR3FX={sUq57o#a?+wEud63MFl1Hm
zN|~k9bE5Ug_%68Vg0nTfe?i7?6zNA{{fLa(GUEqF>Z!#~tk;J86p`YD8HW?Ewur`$
zj-kE#3~N&Gqw)!smOUxCP|1WxtC&1@=X9%=@My)-{%&_GmhfnmlA<G3N_ezH$vkAh
zqqRtSC`7`e#Yako6d&Qyf+N44DLBFdosqG4CvdGW+8(#82#;10dAgyJ2#;0|xj$1q
zghxw;j7Gf>9xW2mLxm6?=z`1L=P~Mnz_k|04?=2z@M!&y(W(Ez1KE!`sQCfcGG9OU
z#~x)qc%bc3)}Xd0CH1^%?vI2V>PydQE^vC0LU9ivw6e#vl=I-x8Xi5A=iq@VN1Tcr
zaIM1eTa79lJX(U|J=ERcf!M}ZO|cEQR@it4)irpuq(%?*G<dXx#!nwgXz)Nc<LQFB
z8E`F@F&33Fc(gi3m4oUSJX#S;b|1K?h{2;}Fh=82f=BCLw#O}c!2_+!#@#{}aHw0X
zhg7$KL)0RluO2vrEaDWgfJ3_C0g}=ca0piXXycLtu2m{VqC^D`gej^o6sCY{HHvpo
zih@V0PaGktPw+r;;?$?)1YGM)jK$H1M+-~zP*j3PD@pWFMS@2wM?8J9@9;n{;vE!=
zfNO<_p8q6NDR}EYyINu6h<E0UcY1RLWdp_Iok0YsBq4}akGROF9>Jp}BZg;nfJbXY
z^iUvzM~g!1++RWzh2Vh{MA3%T0l3zG7>nW$JX-K!&EviBK<D8d6nTJaWrrTBI`C*E
zhg|L`@M!sl_pltmqxBkksMWxu^%+hp>ND_YO@`Y!H5quc48z046t@fmkJet8hr$ay
zT5MrCq1XbC7Fc*XWf6f#izbXl#RML$lJIaql>{DWBh<FJ%z#4@VI=Ayz_kX#(nJje
z9<6)u(}21MJWxCMSx&VBIHV5dpuhnfvIfgIWewmEGx!xlF#|X>491~=0bFYqT!Pdr
z;L$n-XFT^scpy(O7PX1&SlPk^J>0M1(Gmnbd`@_vIuNJg09?xq^iX4fM+*zK$Gw)|
z(NY3E)Dhr;c)-!1cmP}r27HD@w_pH|)(CF9PwUhO;DIv0qajZoz_lE}?9>A4D-VPK
z?%%4|_!RaD<tMqiEc8jCa@-u$Wij7kfqa4LvXoybJDa);afJkP2T+&g{mR?3KwVaL
zRSsq$P?y#GtA+5rsmt2FwF9|ts>`~5b?wbsT{dzx3g%~;x@_Us!rma%Wm{L<VD4P%
zvZH^;5UvPy*{x8w(wwyFvX5UM`y`_-`@8xFbMV#WVE@4(+%DARxBlOTa6+idF#%)B
zaE+<UiK6gC5lU@B0WrciMg+1BDIiYx#)&|luoVz5{NhCr>!Jb{3)f;1%$KZyRl;wT
z2;$6Ez*^y2D}uQ!6_6zSlSBxQ1q#?Me7B209+wrcTlno3L98$eNEWVS5zJzwfTP0i
zs0iXQNC78>>!b)~@mIiE;eS?yTy%f8E8v3gyC8!2vj+uS7Ou-8n6pd)H-!HU5yH(z
z0cj#2O_brbtAKP-I9-IY_$uJ3DDqU4<pD+knZh?y1ac==K$h^!5<&cuRKR=TdM|>x
zmnvYY^qVS!c%W3kbm^KdgSi_kV7Bz1Ekk$`R=`5(yHEzQuqa@e^jju_Sm710TDn%t
zV3v9XBuc+T8N^*k0h^_3vkd0utAHKSe}@d=SAzn6l72tQAeI^h9F(quGMMXD0mr2O
zF&V<oX$72?{-<RK_c#Tl$bb}Ch9yP;7iHm#GL%PN1zeTBS7jhiFbcRWeQ(P^?iC8S
zC;jfpAQmtMJd&<QGMGD^0y3m;h79D^u7FJGmnnm|Hz^=Xy0X-dK-~Y~sDKaB?}H5D
zmzV;kI9*em!Ib?J5bN}hb%t=KVPQZuf~3e=#CZ%{i-`0v6g*lsq=y9x9w>$Ui4B(`
zaIFpU4o(Jmv>Heciw-<c0Etum1FjW6z62_M@IdEdaiPu!9O@oR64gE6le3jQ_LHR)
z9te5NO%V^c7VfxGSq9+IY8~&PR0oe%=Qw$(&cOpsj)jzK8@LwY7>g<#JX(3<<e~Bg
zk5=1Q+*obk(GnY@F&{jT)reD51Fm&6RtoB9@M!IfWrx}sJX$W};<@gY%iz%x8P^CU
zGI+Es#+gM~3?8k7vCeRf!lUIc_L}+N(SjE}ycZrVaoHZXz6B3NEzVzxTEMlC#VW>1
z1dmp)7>m0xJP@on_n)}cD&Sh6Vu|EBhexYYj73Qb9xX$0POy5wqqQgA!_x&k(3^-;
zYyz$YCT>WaA@FEbiT4ccJs@(5`r#?hBleH#Y=ALrfCa|}_(^XAEWkFv-S#8k{(+Va
zaL2a+R&^WTd$9o?Z*73H%?7w+ZGhXa4X`xY0Jpu5fcwWS9|89Vy$!I)d<5J-KCuDT
zY#ZSAYXe+%9|8AIE^L51r48^C*9JJ*Y=9>i8{oOX2DshW08fWDz}?mcxG{bN-2Yu~
z11x$rz&*?cc!0A3)^8i&dC&%UPPGA^*KB~BybbWwXalUEHo)!VBjEm8umK*eY=C>{
zN5HKO*#Ik>4Y1VM0QVjn;MQmZ+!1Yn+msD(7qS7C4IAKYU<2F)Y=9Nk23S~afc4Y{
sSQKr5wa^Aw>TG~@%m!G#Y=G5i*wCKCyzv8v`J&CA?Y^&dE~SCv|G7TAdjJ3c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_partition.hpp.2F54ACF04D747753.idx b/.cache/clangd/index/tensor_partition.hpp.2F54ACF04D747753.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c0acbb0cfd90af28f89e512a04a6f3e3d0df0c78
GIT binary patch
literal 5406
zcmcJTX;c))7J$2|2b^Kr0UKl+(HVzTKvofCM;rts%Jy921|lx_SOY<l5SAGfi3Sis
zkzGIohzl+m35#(<3;_kAJP3=DpeV-mX;5DyHr4NV-t+jwK8MqM+}pQq-MaTy6+Z4_
zvFcO|n;+;N6S5(49Sg&-Db!CyWca25h+$t47#5S;p5?i?%SkCd;@Qz`+{Jn4dgtxO
zm3%x+7W(@icJ5ztW$Uo}xu>zwGMR2k%A20Qs?JLu&Ppf<zY@^%_*zW5GjuGI3=rQ<
zHkS(g8d$+W)1p_oXB*&EGe75?U(7365ud)^U3Ed)2Fpe9E2}FznyihUU*yQ?>TIbm
zJgE{$_WT$#_p^cE+rGCGjQ#KRZi(?dv3lFOW##|K%<O4UY0%I$Ua&c?G)?_nlDmHP
zRnnxPz$)0=XL-F5N9&hX*SR4#OpM}A{664zx3uPVpNd@Y4Y`B`&hyB<=f2JH_9FRx
z;SLht)X;5mCZRE>qJJ86=iRW&Nn7h(=xlwjF~bsEeY1AtMP*27Fz@;9s;b(EwDNCd
zRZ9}9=J@qo^Mk%VlV6{BQJ5=ha<FmQ(NVkgixzSJnffSAU$S#xdtqeh^S=uEr+l|E
zpj>dgRO4E!w_QnA1)f&g<*v(R8SiObvMIT@fVI*rtwrqNf4fKd>&T<0&To1ceedR%
zQI^lTm45Q%W@M#DpIWWzl>Qjk?+cXwdZU|r=Z+`F#<O%rZ&iJt7dhIa@QB4vH&Yf^
z%pNJ5_E2|Ax&F_8myGsoYrQy+h^~5e!YTH+c9zUJ%0A)X4~Z2^+N!Evu8XO+ANGE8
zGT);(ss0i7R)0!d=Ql)LisD!7-8RvGt%|vSX*K`*<e_14Zs+zc1G9h0^fnJiOnRz3
zHSR&jm5mR!r*LBpQ%~CtS%uBrc84DnEB{;;wX0W$f9%muc<iB=lDtl1y+v;tV)bv%
z&sg@t&OB+Q?t#(M>M!&4k9Qn)_TG52(bP-N=Bm7fLjIvmD-O$FZ#tg7{Oj*G*EnTZ
zT=}^z{DA6Co@|5fkrJ`UyJVGWk+GTaUCSH!`&c(ayLCL=-J-3V{=DV4qHgu<1vM7=
zo2t4Zx5ebhlb_v5p44Ib%EMs0-zLwn5X~=ap07ETUNB<%dN9PyeYj}doWa*EPR-|L
z4`=gUL%(bBwphfMA?q`#oG&eEvHbVo`J)6jm4}hdxTqO7fA_*T+)4Ok43MBaYWK-~
ztCz+Qm<pv-53zcv^6rIk6y)JsrSoqpX7$CV9vWAYKukg@jlVDsM-IGI>I%$`7g}0~
z$A{~NqHbtTj=wMt!DhWx3f~YEVL%n6oNr!OtJ*qUd4klRt#AJH=%xJMH3}{h<Yu+a
z{NhCKNq>&M4j{-#wMYg@fUN{LTC5ptE7(-Mjvyt#Qvwzv&V=Tf(290eUN%Qyx1g#P
z5abcaJ_1FwfUv%`c#FYO4o;A%P&pL}CIn<ZRPKj@iJ6cCP<a6HWYd)WoA;V^X%M8Y
zNY?;4ZOAs{X)V07Z0pIh&IIWyay7`U8_kv!HVqM^o12>#B45Inbam>RbUan%0YOHy
zqj_kuQoxo19-2-$P%8&o-7~Mr_3K~xogj|__EEq?d;zFk09pq;ymmckIsY3$UIy&T
zfQPsp7`Fq9bOlM%$-+P#f@}x&?O;CQLEtqAe3kY(8IAbg_>LflfoK>QWL6(Jd^VoF
zlpsd{dj#;1d@t16%ZMU*2(k}B9tvJLR4a#Ci0?r4J8(J*<{(rbWbz+S{SPK*;ruLI
zi1c%CA_pfCSK`7-T#x3-HRW~sa-p@82~vOw*i^p1eF)MTvu02L7QmncfF%G1X8<gN
z$wdGwVzLZiGA6eHtc}UN0PAIPGQ^T0!+ssa>X_UOv2G?e<5)93p|7ZZ!^QM^&lboc
zvw=xAFnwz8rF!p&vQQRT3>}K0BOP0{#^-9d@rWTL$V|x2gwtpNK~~^I1wN6%>Yc~f
zqh+Vf@iB{P*+3mxl;92$+>wqZi%h{CQV`E{JbCt!^{)ya{h_GW_uy%-5&>0pg;2Q=
z>d~$l2aeim!LJ(SSfof#WI&e-i@Xiy-3E4anz6`-aLz+G7df^IpRo(21&iE+Ywy8z
zX}Q^8UKM}sg?|DTc^^!_4`$GY%opC?DX+2I>aD=yKCq+(a}$@o+;JwViwd5<qQ3xD
z<1xTJ2By*VW$m)h<-|*&;(Lbec=kW95o>=*eg6WAUO+MJ3PFxS{ZZ&iv#SrkB<>m8
zl?p`?Zk~j@Ob9yTMI~>{96sQI+x1-A11lCpP-;u~OLUN<FQD5C=uKPm$f?`?OXBck
z9#sfpfhQeUf)t8{3=)wdH6#<w5}6<}R5VmU6d(@pL|rLR9LS(xmtZ#}3$qV%*|O?~
z-CZ_+2njOWKAb_(iqR@aGXdBp04LPj5`k?ZlchjW3RGyh_nt~cF*l?86YdT^L5l^!
z4~0ITaEN7pXPXVm5-mSi&a3PWC>P27x$V7F`+WcKeK)12MN|gpYw8=L4Mzg>Bw!vL
zt~EV-tp>!NfuC?LAw@a!=VpsVpK^%mtScLPJv69tiN}d}+>lQ2Q0tK}-|LZT)Y#%#
zVjHxXx=nX8Mzqjkp$n>!V3S}6lzyeav=lhfrA?4^z_boH(pzQNtXFy>o4rZYfuI=I
zEygYBkqNRIC#rEx#6RM?KO%WTV$GdX@bDvQeGWp+gUpk}5Y!lgGw2VKioc2VD!*`v
zQnZ_5r-|HjQ*qOxry!}-AL<AV{U_x{HDp&qWx91ztrRB;aW&c+L7u^hGl)|s)fc8T
zhR3rg>qd5F$hw1~g9@TW{6&mzR{_*2P`69<z?1Ipeq%j}AVW1ng{Y|efkHo+LibLB
zOo99qI1_o+2>Fe0=ENzId<A*0;8f&kGR{lJ87#$lQpBluT|0t<!bjh0(wU?AQ+AAK
zWl;eOGYYdp{)8KaPtZO=^8WqBrRJ2<|Lf#ATYf!U>*V4133E61pkaSpW7OOamU_X8
zJ!L$1KW#ZGmsP|w=%pvfg<1>s(NYC#1Zy+1SF;r8Ftb;)4(QO?OCR&0!n?zLWOR^6
zVuVymkgYhk71yK{=tZIw$Itml@dD&tfHP=?iWr&Pl8_=XRf%QLs0>=t%Uv}*xS_M>
z<a#RLX>djw)TRXl`52l#hF0{t9}Smo&awA<LQV4!R6GQQNU<DFFNdaxE1_Z~Q=W$_
z=ixfEA%d*Im1_`JU&jR-hw!_UVUa+jjW#Wjwa5;Up_ZW#QM6XH5Jgi8w4^{t_s$as
zymTD5bqhZ@rJ8%OHEKH_9r@@J(V3v$W><x#zOjQUN1)6&sB)|s6u=2!Py*m200w6O
zP6m^U0H=t_GQg2BxeajInA{6Ey-ZGqoMgzbUk5pLOzwu9ZYDS5oMwDN|J>#F<(0ON
z?|gJD=w+u;aQyoCzf$m0TkF-Ucc0XMu%KG(bmM#$Ps{pfS6VxwmKq8!enTaB0&bLm
z&!bbW-DJ4yg#X$ME<tK@wKY&0xfr_Gpb`z@25F$AN(S6ypg|`UL6!hf2{1z?TMM|g
zK!YyXvAY<z9cZAt*j>QA3p7xh&4<(S;Ve2_V|B<bgQ|#k;(9yrIkd~aJ<0V6&fY2g
zV6^BDsQ8cH%u)RRyI1sCx6FTDp$<MetV1RaJvRrfiauepy<h)$LyI7`of@JK@cV!e
zg-r^Lq|h9NsS>g)nOD>X$Zvo`y3Y}09!}&jub4GBQG@tV?%6t%YN7T=2jQ{O&~W08
z#g^}FYtO#A__gZa4da&z=B4DB#QltR)A6V89RacM4O`cZ^^Az{fC$^MUnZ&$!xYDO
f{2BWv=evijCR?s-wcB%^r94GR5mTc7j$!`+u58AW

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_space_filling_curve.hpp.058953EFBD7A2B25.idx b/.cache/clangd/index/tensor_space_filling_curve.hpp.058953EFBD7A2B25.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ad114cdc236cffbce2ed061784f424edf16c6623
GIT binary patch
literal 4256
zcmb7G2~-o;8lDLOLl_A#WFaAu1QQboVF@%~5-FPm3KBLeY8Ba`pd2C=LDS>2$5O2L
zcy$2*6|0C=sY~e-)CW)M23lLU`kvZjrS)0RR#9<neK)Z`lS9w(lsWKczWL_f`M3Ms
zn-VvDdc6yVMdid*l;kd6K*uo575$qQmlamv7?y<|R5W#@_80XAxLn-%<nP&Sn?~{?
zw-&}Zu4}#V&1Y8(jk9KLNQ;|M6#4tn4cB$gx;nZ0{m*_~?driA`84<<CuCi{GP!DC
zvvE=6FRb32ecfT2NjXbThE;Z#(%!jxRR5mZnc4Jy#CH>gvYZ#%Ujm-m4R5<IEUDk~
zadKR1OV9JZlH`O_-m4V%vL?*P!As8MpLnO$<ah5kcct%VrTfk8eeAATCEJd7^{y>E
z5^{WB)WsZg;w|s<*MB_}oHnm+!hj)f#Xy3O^ik_ikz0>v?!FLPQP+C+e)fS~P8kQz
zEhS?<=}4<ji8rN6&gvQW%HG=hpO);c5trsw_9Ygs`EY&Y%&iyN+Ahvbrz!jk3g&Ng
ze0VKRu>ISUDRy~+zPCF&UUr>~Z7II4$koN)EzD<c|3^!!^n)Y$hP8{>Mc->Kcq{vp
zGwuxCO}dT$t(e$VB>PmG98l_9G;^lUsrJhc`dnU~-NC=UY|i;jnuRY1f6JWe96MCU
z+H$V-OViSltPzxd$E>}B%dS0K=|-cwU}S_ULT5D?1Jwv!oXc5#yEd&)cca-M+MqJ%
ztp;edG~mdV9>*azO@L@GzL$^H039>OuB3d%eHoiZAetxSDXj+Rw@&Yp4V+udaHBaO
zx&c=;;8B14z$?Y3pNq2>JZWXqoY5-+zCdm@K(Frme0R>J7?Mq6B3j0md0Gw78#2kP
z?Tp*4ZZrm>)he~lYJlEUE?WD<k#UPnb49c#-_yrxfZlcRfLyU~bBGho3DF!Thh;TD
z_r8$%H3fDc?<XKy9w-mD8lW!?C#J={_wpK>#z(Y}FAT66pl_PHGF^EgDGs!WER1xU
z;Ks8W%#ntQg5f0ozxDv<n_rC(Y?=oe?=Ex~TMdjK{D}F!_jbEBH<~@7js8ZB)c}1j
zSGprwv#y3sb40Y1FZHw<pv!XeOw9y_*$3U<S=x~Q6;3A$d4>GbXP(8B>*pmn%zX0o
zzU>YIoFGLE(G&;;dIe`%djvw57*M^bXno*SE<t)kdBj1{2~s*qDg{RfQZ`8@g?bUB
zDpr*U86-%xUL6I+BFJEUunp2GwLui8Q|fF`ggzn)ydubGQS@XR6cVkakYSo38k(6P
z6GRD<ZBR&pmO{o*<21-FLCy}HZG+}|&6QDHDOu`4Ayd4`NTCKix&e=e4kyS)T-Jyy
zDcp>!oAEGPY-q-30G`;<)-AU$J?cP^a=IK9>eYo3#ZikW6v2tGK?aV&1{pa<8`OYv
z8t_rf;$4gkxbXeEG&-5D$oGM=(@C?z90xJDtjccde)eAxY66L;L=8(F5`~(;A_vd}
zyzlliUhMp5kR9qIQQ;(;DV1hFn37J`;VE@^Dqu<Y7oEeQ-!R8DNV=4(3g?EOLVjh)
zGYl}7A5wQAZTT@2laJCz1-(TleYL)NK!KeP+mmLtc%iXNh$WJtMZ!f(Pc17BytTPq
zka>I&@<X5(OasxOld3dTI*g^01;PRm1QulYvBtq8;|;WOd}Y2<;3dcs%8o@1Av>Yc
zENlU5emVMpo~!3UJJZQpTvLmOLJStHt;KbKLnAjv9(*Qdjl(ZkeKe+3YQX`5^k?~t
zz{}VRP1?rUF(x~L6f?wb;4~^ULt=xx8Q!A^Iuix<N~mGxw0h}qltSK*=JE*f2#4>Z
z6&|LFf)Ao~<TE=M!b&fy4l1qfIX*^SW<9SDymfeG@N2truC9^XW`BF(I8u`Jm^**^
ziT%6BG)%VQ;1;=4TLXGrhoiqBaIcsr;>-FAE$YW{<9Nb009}DHYuSe&ek74{(HkI<
z#m9Ck8k--@8E2u(j@h-9KfNLOlP@b8+7&O#+v5<SoTOAkK3B)veQ<Z}%TPK&y0hIq
zVXq)a9-HS$l|ZhL2fz@5ta7R13a00K{2DUn3J5VOr)AfKU{()CUoOZidoiY~KQ8<|
zzi;Zl3TUXt3=cl_0ToLnh7Z)8zkP4T;i~p=h7N1Nj%nSS!4stBtY{&n-wCV)AxIDG
z+h~Yo)24=qLPZ9sgr%x)*5RX)+|=xQbJ?+S&1<`><<SwM!kA_FQysRMSFb(etce+C
zaL_x*Q_J~*^3<p>s-WKpvKl9<aaT~IJr#%TjV0TrBPYB`Zvm{%_G?YOTZ4PvMUTDw
zyb|GY$53Slw`O=Z^6ak*@1eXKXQG*XKUnsciE=Krqo|$KYzjp<Mc5#NlfecVos2f9
z0e5P^M{(znUsa-K_(tTaBtjAct{&EjU4N_|S%bPZ-z9(av_X)SL?wsH#&VbCd<xaz
zL=8U59yxU>W3TqXPveUA*c5Z1=8O0p<18LDHl1lb*ipsPC^Q;qTY?Om5*9+CsL4@5
z(6uMef7{<b_^{`-EqD4ip?&Jxd-Nz*wr8!r!Z)9;qK7{jb22-plPyb0^4u`4_%He2
zxq2`?^AoguO-j>LD!0uzs~H!9w8P4=XIf9HRJ>xjcpB6YDLFA)LQP>-nWHIGi}P#o
z0KiusHl-BQ*L?We_^(atC1VH7gD8Gwk}?&N{zK@wsK<MF__zWEmyxPjPSpE0`*!Dc
z)qM%_$Y5HJ(C`+LcOgd%OBA)LMxv#{G$otTArN$0P%QC-^@;kGUn!!v8k}E)+u&wA
zsTog&tPXNwT4e*dToj+}aRr8>wV?IZ`ce&~^VW?bru9ru`P+yK8*wSj3`c7uE|vp6
z^;(6XLJ0FC*$=EzgX|ACb$0vj6HgvjRv)<hd}rlHNx@?_g^}ea%$(={W{rKkD(=wa
zz5|cYC;sYL>hbf>_izVG)}R!LN%0h@0alh<yuczxrm}2Ps^x~{#(4Hf7<Ty%?raX1
d=Z1|c4u-i~UU+qJ;0D7+r47S))KfR=_W#rbd;9<Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_specialization.hpp.5BC1AF8D88F729F2.idx b/.cache/clangd/index/tensor_specialization.hpp.5BC1AF8D88F729F2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d113e54a34d1008908627ed5bedaeca7ed6197e6
GIT binary patch
literal 678
zcmWIYbaR`>#K7R3;#rZKT9U}Zz`!5~#Kk2=nQwvgHbw@9iaD(l9J3D_@U;F7=F(lf
z*XPNq!yGx!jZVCC3%zpW$kyJcC3lbg`|ae#FPpgg#e(lgzJEDW_BmAX_=420YL3Zy
zlQt;7w_CYGB8>mnUgO*Lyw;zU+@)#_*B@!#l6Ep(!u;#LYnS{kN9~JQwJPrw^UCDa
zs=F`PuAP6~+&{S3|K>EGbuach+;d+h^m&V!jlwj^^yEo4iv^-Bt;LSiKUq5G^2Ljl
zjtnmzdoBpiTHcfH@iBJOQip%ebD53@cX_evmram4(#-wlcvu7bkKPu&-h!p~O!h9e
z`=&Y5?^MvDG`-_K=au8SjqEf3E!e?OT$!8H0QB`5eU94-qDxQlFmSOmaPkZD8-NK0
zP9{bMZZ<|v#<i>-*FSn>TCy^*0u^xba4LWa22P-61_q#lGtbu^d~@!y4?hD72LmS$
z3lA5VVBi2NVEC8#*GGnnv5KF887R%o!p#jPVA4gYX~l1VDuX#)_vWSw^fB^q^7HTu
z!JGx78M(mD@>D77p18}o4(2EpF0f-@!a%*|{N6A<Je<Ou!m<dS2&V{~x5I5}#X5)Q
zH<(#Ch4qCkV2W5cMf60>5Ij*4Q6m@+=1(Sw|6oqw;S`Y<QH3dj`4*@M<W+<sVR>OS
gn4+T8oJ3II6c;5Ia|3B6PVb$Xs^?BJcQP;l0Ll320{{R3

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tensor_utils.hpp.B92D33FF5A9EED5B.idx b/.cache/clangd/index/tensor_utils.hpp.B92D33FF5A9EED5B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5bdd303497dd37ecec45f79b2015afcb51f30e06
GIT binary patch
literal 4658
zcmbVPd0b6d9KYxGDz|s*RoY0ReSO;QdlkJlqReZtq>M&ID=pd(QZErTQPxJ8kP!_-
z+Qu?y9c4tNq9R+1FpZ2|=Dhbhy=VH1`}uUv{l4Gtp5O2MmhU+&aCLDpQp7Olm9Fa|
zg5m;b7=|gpzu35#Fm(tl<S}er-t8q3QJpOLw!V)$Bovhr9{9Mv`71iDG|Zb8dSROT
za9Fa#!jOYzbHa7@wi<uDU#}fB*F7(I@EMRGx9k0VsBc|VVUEPxO{;0{J<GH#dpvz3
z-rs!3y(alxX-a2!c=N4Ni*AV--<wxAxX464Yz$M=j&80$vKhNJd04_bQYsNYI;mKG
znV;*kiJ2MJ`)k?s{U@(Pg%VrWWVP{5?aJ8R<Da^8x?6vL199rWv)4aY1n!XAu{EcL
zJb%}qvd%ty@J#wDb*ZFRlYbo1FTYtf_}PCqZANOV<=c{!4tIIhuwqizKQ;X^>W&is
zO!)cZieo(y(N4*4td-pe>X>}xB%Y_-UmCkjHNz`CYj?-@NBcGM9yS!-E{JZ(t8~s$
zRr+n7gKg1{7r&fPi;z{h`cCF}c~Wz6Z0hxZWlV4P388nZPgxGV<2v=_m&Lm_SoR({
zd8TK3t`4sy;mFAH^0oFYikd7Uzu={cdm`<<N|wik>E}O`KC(QhoyqyZPhzj#r6u8I
z4;O8E-ux-^7%SsoT2tH55R+NG=(A(5Q{eW7w0{j4lCfl$*2;gvE&G=GhJ?Rp+sdH^
zLrk}8?Ug&ct+@C3C+D-9Oxl#zF3cVbAGwfFr_tP-$T?N`$D5c#fthu3ay=dOzK0bq
zs@5&onM+oCq{&7-JsS|MF3tFJiHqtXR^giB{hsGL>?&G1Dk~~x8jp0Xc4`0aug{&4
zQCg2!gQ44h9h|bsUSsyD2OknG9EVREWa>{{ocUp1Sx%{{>EBnE1eNab+33*tV^JQf
zH1BG0XibT_myIs%<I@ZyovDl4OQh%Czi(yCw!J>{(up%3D_91fqAHe6jcz=&VJLlr
zX}SJum0P>YM{WY<qVC8ET&p4-W#hBl^Vcu7Cxu;9X&QMjKe#6A%*O1!1^3Cj-G9fo
zr0^U+@1Ul5{7$cJm49wnjCbFSFHW~l2+ORs*r1c-Y}u4pB;Dzh6Q5Kyqg(Ul%KQj(
zpJRje$HF#nIv4U7Kfk^3<YfQTUv=Yl>+VmOJ$GZtz!n|%t$jlvJWamua%&t*)}v&q
zRrRA5lleZ+<Eo_RhAhAFw#)v7Y_XA5$kB5J`!ra_M=fLeTFr8$L(HC(R5w<4Httr{
z&GN9j>Q~j5>X^z=TRlQwWaYYC*)@(eN$R(p=nM}%r-#$HExAndd|z)f;}^{hn&<UG
zMjre41S~w===MEp-+`uE2769;E^CwgW#fy@pL`cT$BGi1WGd$Q`aU-4b%-=15;8s3
z9-q~fpOATB=$g(tj-2ns6+<Qa+#erKNXY!v>Bb8+s~?uWPR?!e^qy&1abmYaKud?+
z%OQQmw>z39K4>v3P3fr0X<l9!`{*+9>C>N&y$s!1D>euHq-EFVTA866S3O^AYs>nW
z<8!h<N~S%#7=y*Gj|e=2V;Bf7PzqrAt1IHu3^B@^=}id+Yy#vF_~^JE@BHIr;awG6
z8;VV}O!Wl=5*rQJ=E~2`)h00$1B_}1SUZp$oiIW}4cEm01&r6vxwgj}kD}w#p^}D_
zhLT{w2n>^f4p3mu23PaTjcGWE*+7M4z(@w#qZ3A`t4t8A;-=~sOa<+~Lt;iyX(zyT
zLQ}#CBNSzUqdr{8&FuI7M4OPKa<~I@*H^(;NiZOH7X*7u3muxifW$a(jdK7y2WX=c
z-(9-E&|<B9cRZg(zN6!8sN^T*Cnp#%LUS?<2sGle$?-jTeb`kVp9zhy_1PwZ0U24I
zRCl+vYxOXR*~9N&0mdt!G&*6#L^(-;k=P}gW-dC(N<yV%oRN$xQNjr$l;x!aO1#vv
zC*}Bwmn3ElS2_%^!$4|u!U!dXgfZWnt(Oc2+M=VQNX!;0MKPi#3kHmsqDlx>*|ya0
z7Snj=E;*b9-SwI1qb3-TyQiP^2DRTS4J0u;_<bM1`e0mz6W<6;VB~Da5cZ(%LotcX
zf=aajtA&9RPW-@;0z;(-&uTtYe%&mr6{Z$bpW+Y5P*d6Z_i6=u^o3i_#u?dg%fbmG
z6cnU&^x$p={LF9rSk=y!!nL3wV`*bjFkr;gNy6aA1O>;Q=a<cB-2xdwy`t|Vc#hUr
z>Ta9|By|VS?*MsJp^^&^t+}Ndpa%tIfL;a|s1{_+(tiryYu-75q~^_>=ZKO^Qe0!M
zxd^4>%IUcJSlXy`+#nriif}frmyMf@?GbLwJR4_=W4mnJSsc4%<DLl1mK9w#kc_@0
zLsF(RQ<C3#{D-8NG^QA0)7WCjo8~QsvH+F^M!S~(WdXnn#BmA0O2lyuz-q*C8^GGc
zaWBAnMVO)1>SWe(ZxxM3akM!)C@v5ga*RZXtIZXYaSgd+GA)(BLPNU5X&C={K(7Z3
zR8ojx04`t_lt`2}TxOI{;?@ScrN(aC_W#5pc7cV!rxv7Uts961FqX$lKo`|HQ^Z_U
z=ZYaOb+0irGb%mg*g*Ok=%R+IhCW&tu@f{Pq$Xry&W%{Yu4qxF3_%%48_1%cNh*dO
zqcGN%6c5mOKmn~nLe7%f3CJ-zqjCU|0~jKF7|0$56UTZV@)P|WP(bBM!nUK=0tFO4
z5{8Ft2Z|_OB<x+X4=AFxM(aZFg=C4`L{tgu-jp)6MrnX0=fzxvTA8F~F=i>E;E@zZ
zmt!D8&L++?M2M@)6_a_XdQC<$8375S%zEuQ#c-kT?jF~PXp1oALTW~?{=a1rL&H7s
z=^3dZbMR+ADI<G69}bxJUP!h2=F;`OH@h!@zOZRr=_00w28A?_JXwBkFYA%Z{FURj
z7*wKV6|-Vi8Wem<Pa*q?E4A0X&4)|*OZks&X?4|c!-M(%e*YC!9C0MqnSH=Tdz@{=
zRnUa+ph4cEQiB5F8G{5{j&4rm7S-0%NXkapW-5v=M3fCjglv^<#bi8Ofrn2<#!wr;
zwfRRAS{a>5(y;hUOwGlLZ#L6aRD2u_j+V%d^E_v+2yr#IVlpocuQ4PeINwnEGu$M$
zJm2}*8k({(v@sJcU^CaoU4(4ewqjdy)^;9aE|PM#cXknJ@^Eb)ZiF^SQd@B4E%+3K
zf2@_}t@?QNCJeu=j;#gyEZbK{A%$+jgg<{tf7#9bLn#+Foc@Xo=dQYW;*x`#{kRP0
zZlCINcA&TgCn-NoKRwhYB;~K^uZJouFW>m##;SPsH@Cx)q>P={<=At+!sEg>k|RHr
zzl$5Ud&E+(<%L%cR3IcZ0O$ii9w{KYgNnjyhe+|uc%D1R)wPf8o(@fYxdN&x92eY5
zP<hbcNj)ta*F`6BWpK~+*)k0e_$RTfsjBcC7M#SgOfkfkWs4ziS#L3v1&Ay#QS@a2
zAPU5B2_Q<uaSb49#Bm!S+Qe}$AbLf(N<Yb*UFvr1E5cvje5<kX>8mT>oN^Q#*u}yN
z8;X8&*S<ee@+^X2w7=xDj#qgLLJ#cN_RS^-W^nG^>17=mSL7&h(lUaZA@T&K3;uJ0
zgzyxDH%}3I4d}0dJmPD9&oZ{)N~mW^c$;ISi{meA;`x=~pr&bww<QQlTUuM}ppTv&
zt&bRvgtr5uWMGC4`j<}*uB0s+CdP#(CTMj4zc^xJR>ZF2znJp};X^sCr61s}+&X=N
i=?n`?D>J4!W`SfFW+{3#6}`?7y_$($twgU((d&P2O>8Lu

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_avg_pool3d_bwd.cpp.6DD8FAFEF9A89C97.idx b/.cache/clangd/index/test_avg_pool3d_bwd.cpp.6DD8FAFEF9A89C97.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2420f306fe7ee96210c531c3d639a22524472f62
GIT binary patch
literal 3470
zcmaJ?3p~?%8~<;)?#4{QBy$_v%yP>XF=@rzlGn9pE*)aXC?!tpyn2&Mp`4;nDqUQU
z4kv{|LM2fuZ-`EGMRmxLdjDJJhxY07?(^C6*>n3n-{<>#p5NRZIGj{01g#Bri06mI
z@}wXL!oy!dY)n|&5(vsbK~Q|bWltYLzlGw9ukDvY`^xj<ISSR8_JkGcUTQXhwt)_T
zQb-!E{Z&O*N(s#`Cad)2!K<@IQ)|Ygv=EGS#h7JHe1B}`@Jx@I>sFI59cQOc=l2`d
zA3G7S99l6r@muS4;p5@qyd^`g-b3->fk}<&O>@C}KD;%>4<;2>w~jQYXQ+I5#zdqi
z9y5M4ytAgj+G?PxdpT)bz9GHs#MYCZXK&WTY>p=Pg#}rJs`YK=pNI{Me09_Eq+u|%
zM*Trnc!|#;#}fBZikubuYpk=LY3JK18}HCdh9!3Ot0tolHxax%{zKm5e|pNkSfv_j
zHT9rZKU61vE$@y@)QeL)_ES_?;7L?axaN_Pqna8QQbPG!IxgE4KH@ItaCSw1dDeU`
zOHn8M2u?B7E&6T9z`=C=yoy|Gl|d7~q>HNPJyEpB6iGbDoYqzOl^Xd6*}4zyTttD+
z1Q5Rl;tCBp#^JiyQU0xKRgvUVZw&H%HQx{IeOjN6ex8*=JJMd4Hj#YCU3Tcn7@waa
zbg9cTIFH{Rd45yEglmn=!~Z(cgI(?yct>_B&i4A1oe}in4Rk4pD*0@xLz-+%$5Ub>
zzp~BzjML>b#C7T;v#K>(Tz={O<qxAtQ#As&X!TQ?m1FnH+q0FGb4wU+T!w@Xa?kBM
z{x_$%)w8g0m%Eo5SBUV+8^X3Ucb|GbU0O6Aw>mU%*K^1Gj8tqbOaDnoWe(F{E4bL;
zo?Z0SmJ?c*Y0ONkzEjjM<&{zX@t26}Ha$otKBs!m*11>tC#F^HZDD@aDY_eWJ3XT8
z<Kuf<`#lF<uz#touDN>t)PXJ9>~TM`?|^pK#jyiUEzXIrv>o>yJ?(w_+(+W<gTfv6
zv-0JRnarfw*VmCoX41Ch-EglyniPL@U##D&#1s1*iYAvbwJ-LrC(x`3_i;T#5mu|b
z@(jze={+AhuRq>**E&w1TK~xp+0`DFm(zXrO3g>JpNAq&9=(~1(wXhSHWd8PgQ?Y5
z59(ET-5C12^2qD#b?@vdBXneT_vjjCgl%{BZ>|s?yzX<{;e=;(`-kAYZC;kAr3K@Z
zO$lE*#+r14P<rtN(Or2vC}hFh@rFPB2zyO-%OxogdPV~J$MW8q$75JN*@fL#HrMz^
zS`5q8HLOYTmhukvqZ}B_e{D9`=-V$}Kgf=|P}@8{tT}c-xZTvD-@LRmi_H2wikM-h
z>a?Ozmj0~M{fGK}s~TVCx4Pd?E}HH7yKVBuN><R-oo#7fU)-^-m^9s9QeiSjWeZQn
z8t06Mhu-dl1iSfxIS2?+XS`0m<h$Pqjr7?9@o*@dvRJ?yFaGFK$L>irOx;2weIbZP
zlcC|o0=VsFrnev;rfx(dJ>hH&3bRx!fP1}{@3z<uN|P+oAI@e&Y?N3)qDKY(cxmiO
zw0OC?1C8{Cv(YHDidcMeLXh`Ww>PdO`ypB6W;k0;MovX6fLqv7RAy`tAd7{9esp;1
zwe*?Px)3`w6au?#AzMYU0JRGr$DxP`WlTk+A6%OtPr!%;IQ;zCIK8>)50hvp_($2b
ze_cF8ajji~L;eJ3<2CWRVga)6HVwL@hV&+(k^h3-SQJ)GEP#9Z@cMXySA8rR=>@wL
zPznUG0B$HIWShW17J{$?_79Ujbw?pEJQ9}_2q3HERvOaL`45o@F`Iy7<aZjlS0qY&
zpMXmYkH{k0gOf=4gBvE^>6jy=#N2?U!KK5wq8kUeFg#lwTRI5Rm;OChBYe3B4326?
zbpt_Edg3RZQ27>Y#w+4A08y(bWmx=BhqXkd)fZyV<NUd4i0{GM==Ua=juc11(W0RN
zHw~r@#w8lHg`U({YU@Bx>oVR?ZZH4rjl}S144NtE1=h@ImOz{5>|0#`{#!Y~FSkQ)
z^a9e=8*W83J|Jk*pDlNGd%T?brxh?~9xh7cvBOr<qQ>R%3ZQY%_YrPu&|!6Ov_3|E
z89)g}A8ti7_Y1A4vQ#-B2w34D(ewiV7#<tVCW03H@3JqQ%t4FbEH1!vatNSn=gdl&
zxBX@J>Dbc{1sJ6^SKAj<iX+4nz0Hg^lemYtQQ452<W-L(4%s=w?x7Eom&(AT+R*F)
zsj$YOaThcrwh?EMw!z4GL8IHy7v<48^!dDm^&7$ma~g|agsPgVOn@*!a2XSo_OlHP
zW${Sti)J<$c#Y)2(Yjn+KM<W9lU$tUN@|A9dR)EDzzhcIpEd!Z|JJMCn3(zgg*1kT
zCZQSNZZJFy3Bv$Ghv8vKSO(Z^3=c=bF~E{xcnTy123Tt_5lG~BbMrm${q}^MppDaK
zV1zncoq2@m(Msj3F55qXr`L#KWVS$<k%bWpXqhgDKKbmbFp;>H!&<%{-JISXkuYbw
zpR1S}s||v?qp@h#;1IZgm66SYMrY6$EeqY8KEEs&9-T#>&pQ?=oim}7Spx$X%^Tn_
zf7D{R3-qoXHdDz|b6^I{BT(Y!6-;!gBI~FS9i%D&6}&C~m${3tg}Z&86=W`f`=QW`
z7dkMa7%kH16#BeYc4tkyCaQrXQLZX<VEmvIC0N3I`M2COUiG6CxD`37oH4)*HX7bW
z(Sr}n7~WE~r7J+6b(|yd<+h`ujafXV*S=bryjgjCQle60TXW|*&Km+;NfZr8yD^f{
zIzr9*MZ_hBst(NUsfQ?i@}Jt_LqzWp7&9=EFfP$Q%>piUmbwGL)ipwP^&9Np2_H#Y
zLDL1K-E+`z->B4;37fUJTJ9ivn3<QXbj#CI0`5qjhfSoVmY+m(e>y%HWB%GuBM}YO
z0mdczi(9~@!O~a@D$V$Lbk>=by_Gm}7JqnkL{-wbg!yxC!JN!pnPCZYOh^Pz8)gZV
zfWX@bZgSouhzSxHi}(x#q(TVZp0nCJgppU&mjrALgaYOr(AI^YLHyu|C|KRd-?T(Z
zT1Hk*9*w~&;qZzI4sH&v&V&sdLjuZKS#7C`I?>A7#+Gfj)_&zO`WihJmGmR4>T=B$
o+B&-W1{9i+v5BeKD&}e}OAUtk`emwQ3$DvHq@%|=CxkonKXOU_1poj5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_avg_pool3d_fwd.cpp.5274F49320881D42.idx b/.cache/clangd/index/test_avg_pool3d_fwd.cpp.5274F49320881D42.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d9f03ea0a12a3c5bface0cd94509cd34f5805c04
GIT binary patch
literal 3298
zcmaJ?2{e@J8~<h+%WU?2ZL$-e?8bx|OJg^L=rY!_m!(JxNehvvWKSr!EX7pRUyG&1
zQl!-7L>Y>1k;+nW-EZdJSMHq7f6kfb%=3Gn<@YS_J9L_j%>V&{s6n)Z$ly4C3<N>K
z$Tv1FCbVT01SR7jD52<@$6(B$k<i#ukOwpDg(Ib$Z_61G#{eIeCd-y!wU&0=GP&J4
zZ~s0CWtBg229?9UubC=oSNK4*SA($0k79!xp{Rx!m(fB&NlMDC-fc$vrH1xPHw@C2
zzVjz6*_C!>2{?r0-H}JV@U*=fo<7k^K@09?4b~~r=D)u?+g+ld7GY8-@ExN&8%hpr
zfLp(fPEV8&r?82gpF$17h<!{0@e7r#gJw;3;SDyOjc>#LtgzR2dwzd9ZR2&T%dU(6
zv%SerKj!}4fhT3#eCgEhqb)JF76l2{44d4y3yPh{!%||c(-JPQ#KW9L!bq_$ds8~B
zpL++ghQ$YJ;7XnI_0wlm-<ld<AL*rqP24N%p!M@gRM0-AJTBYybX?-`Q1g#Z3IRcx
zeZpu%Ty1j~Ikz`%48M)4iQ5rA(IA3(OR37TVMMNDm2{!?+qLIL-;v9)?mXwdYO&_x
z8q1`btFtLZH%MRZ-z+{>Uml^ESe^Q1^X}limVYTf+?dbfDQrHnHV&%I9P&^#*`CRl
zkt0(!a)ulaE!l)K+G|EPh6HtQ8oSPa)mc7VrMNuHQpXj0)G>Q2ErwF>HaRpOnJHhZ
z5qxPgf#3J;g{&FWmhIHjQCAD_JF|;Zi~5A(ldX^JvUZ@mxa724F-sNna4SJ;zL!mt
z@Od?L!^2a?J{1|uYhqh_@42`s+J@oxyLu`n+bWI1iQ!T44$N-Gwpfa-*CC26{@yft
z_u<#-Z%saL&A2{)tK~NN&DBfdj+^ndCt?3FR&M&td;0M)rS3O=sCb*ekf7G5!PL}9
z*`!=P>`+I`=bOjHZdn~PXF3_&v@SV4t$O|_Jkyk^j@l#lErNVu_+?#GX6}K#$F8`j
z`s9R48m~>bUhzlel~dIfid&UezdEzWPlX{;wT~t#tToe%v43!gnQCAn`=Wbhvz;?d
zGIZ(E?%wiv=T~xVI)9zKt7;k@tGDNO9c@atqDf(8+Pbuss=s1d77NTyG?rCmx>|g^
zGZ&$J-;*a}ppx;;`)mD8C%R1Evx4e_!%2C~8F8d|?+$WjPVaiZkwXe+llt-pzBpH`
z&pobw>h*M1lKBZY)90{NN@4SC6E<#$vhd*b+`L=QqoVO6mjn8<o%=^HjZLO0UKlTT
zS&d_*`Gp3HcgO=WgTuawlaUWfTUYg9VO*E;`YK;qjoFv_sV8l9q%6AVj;&EP#w6qT
zfTgGU8EF;8mRhCpM_BYgW>dnLll+0vu2prPl)@?m*DcOjHWV|}a-+*lAI8QeMh3*8
zAPCdnc{g5GyGH;WvK3+oDF`WZ1aj9oA9>H@l;PgSE<Usm1Tpw|`6W0494l*$_h4?%
zKE$KFkYgQQ9bt|D$NGNG@`S?CU_5#Y5}be|h;sxuHdGQ_a6Ekh3H}@6Z2%eII09n7
zHps`w?W|p<cUcD!3i#Er$Lxl(nWt5bc=To@I1xvb=7@iMA*eWHA+A%UQ5cW*K#uV^
zyckD-;|7I=d>(aJ7Y_v?!Sx}1A&vmKnY~(i(|sy!c(f<tEr1h{<OtwBMcx^thOM&4
zLxI2Yo<><P7jrIF3!%LcZxO5rfg^x-OmIkS8Sl^Dqj8*`gI*IDV$x!YAO>;Bs(=4{
zZ#0U)P^Kz-0~i=1zU--h&;*77-ar;KmQ*;r>C>c=AqvAO4zLDOLwwmI`{}E%q^|>v
zv~>2(`=^RMMDhv=3CV#y(ZjzyXPT_dgF<s=1u}t2BO%#i2J8q7DI=*3fHN>g|Kkht
zVmOhe*9sQwY^Do<ssEb*7#VBU5QSfNtmi7lLiJgaKC3Jmf!X^5oUJoIPlwemc%cal
zcr~m93Pv;)SRH8g9V+iGp`E>u3fc1qncT+Ox)a?eX8+9u7RB93L~Z;eg@@e_tdJn+
zhsWm#7jt}QBVw&V(2xYBg5^Z|!QRZD{g5`2wgyJOwWn_vy+<u&vxH$izzflwVLH%!
zA8)8G$;GqjtW4nVK=adClly}U4w>0FpZ_XY5Mil8RiT4S@$dM8@2eTV<g!kRy+(*m
zKVX5tE4!Dx!}PYgCo2yLqDED713^+^Qp(buNv#O1PE~gYaGFZJ>%OedZ!rW0o`ly0
z_m03IkO;b9F9-}GiKq*@LSP7x1a!e2JN)2M!Nt-${YYF%1xZ5?_lT8fZlLR2B?7Bb
zRhQFcIc3jp*<Gi{Rm#cL-d|lDCQoz0IrNKlv*R}c5pu95tPPrqXqGT_g=P(JTA|s&
zjzG(g;>rCalEp+=u-gxo1XPL4m;JQ-oUf#jq$Q}J{6qhn^E0TUh_Aezyd!|Cgve)7
z@qtoYR4@OTnz-dnQ5Jg}WMD-=0MW>>1<-1`%^xY?Zj-otuWlcj%D`v_alv4BkqftS
z!~Z*NFb1R)_B;2p6d5BKD&Tyr;Z(YR{mX+$Dg0XeFfeMgIh{~z`HsC;D;KDFiK6#m
zUBNULojVU&J1^RN5JiHptwAqaSeflCtrLe3Ses8<9KeqbXKTK|&Obteu*rj3p0L!M
zeTD~jA=a=UECw3WLzAPsiag0eU?r-O9pL=}->%_yEXNyZs;q~s4}b{_IS;vIZJ0i3
zylw8~UamK7#56BO<e`TLR~4*0cZ(>K)jnJ`kJ$~3o1VSR<g%U!f0p8Y)|Z0tk|)b=
z0tCP-`+p|`j8Qnw-v?+qcq!2+P8*kB)&FSUeB3)54Qf7evHT%?(k$ZY{wXea;aSll
z?_=H(E*Odlj__ZD3<RWxKr*lo%O|lhfw3@qbberd5ahr76UET?gqW%3JEB(k`uRh?
z%Lh<kWKcvjqPj-9t-|o|^6_Kw1fqz5u#h0lh34!iy2(aE6z3==EiNG=OIdGWxnU#K
zYMrExxe8f}^ea?KUO{n<lCrA0=2~r7Pv5}U)NHkhoUWm*q?DSGy;CsS&fVGpMTh<d
Dd+_8l

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm.cpp.A022B37F45A32155.idx b/.cache/clangd/index/test_batched_gemm.cpp.A022B37F45A32155.idx
new file mode 100755
index 0000000000000000000000000000000000000000..aab8245f5887c35ba01372f50f70a3e9d77f883b
GIT binary patch
literal 6944
zcmbuDd0b6tAHdIZRku!8-P3a8wooYDluEmnBTEZKS0dUMDyKpy8e3UL-84hmL(@L(
ztDUKccq@a<&^v~*i$T^g@t$ry-6z*;{;1EV=brC*w%_?}&+i<*)4X}(R55IJkW)-}
z@Opm<48sWeKWhEDP(>QdSQr+QUGEXTu*ZV^_VDqArJ+;Ts3o3b-YPg5qEc+(@tr_z
z)&6B0db53(Z$A5}#pQ4NsaL~eb#l8mwe-cb^mQ9zQho{c;}4BF`qK%6#NRq}vK!N0
zzI*-l?#kCaX6i1xRc!9OO0v#)Yqp?pRguk|?o-TD1;=7z_WjbF*7a*gW!?J^55ujt
zY|hR%zjpGeCJ6oeSwyDmw*Oi$)HwY7T2`$6Tir8B7s|U0N){wJ-H&O{$<|5oJmPZv
zLD=K``8!KgdtwR=b8<V2xBj8k8t1-AtA3H@g)zPIN8OEU|7(!!7gDC1Z-25N#yapt
zld|_Y+3-rY*yXwH!60E(v6oW0soAH)dES3JKQYEDZ}ZQeQfj$j&2`eZv2d3xuiqh<
zTw-&@_PQ(gkEZTxMIC8un-$~VutN+V-kSewiJ)*&v|#Ms^!OEy(|H%G@66g@Vrj;a
ze3#$uHZ|h7&dlF*U5>24@Uv!HXYZ-JmzQR`e^;NGbY$6z(y3Mng{fP{XIGw_6gcKX
zv}@|6Ag{T{E_gn{yi`w@xvaUuJIbA6YTu?G947lI>BcYFD_Wy;9rYaVHL#2ps;}11
z$rF&8Zui*-CLZ4(^!v5Ny|>>7$E#Q92AwE<#P0p2aPPY@HP?4XoH)-t^suWwYS9|r
zPjXd{uC(uUdT=jgvU%_FyYZb$tK>}YCwV27XvnQ7FBKFtZBLAPA=Q0nS?>DhCuRRg
zdv?UhYx=5N(@drv>omGz7FXZ?lA$R(cLyh~XG(=~<BuKH_l~!BUH0|u%~IERI?ruM
z{T;>!)&tuH_fwSy=VG=@w){u$2*-WpyozNfLW--NIiF34=iI(@^;Kbl(o-_-cA<f9
zE9ZD|x6>SA=9_{y-vd(zqthn~{!Fio5FFUjJLh5wF@s;;zgv(vzSH1%L6Cm`8<{bW
zQ&T=>?(o0zK0!{)<5HeXYQz1jK01vzEg}rs6~m@@Yd7~bwr+2+3GTC6`BGl0vbU-=
zgI{za|F!$JJ7tc}fzG?!{^^i04}5Oetvt6yH#EIy%8fVE`%VQe(|l}T%ywy3;N3dM
zb!+}BzP8b6zf+ET4H1{reWO~&eNj)WZL`dVLtV}Go?RU_h7_gr(YZTicfhlasY{pk
z&OEtFo^tl?_wrP{lwgw~8H;DyD0UT3N#tzFpkf^}m75by;wG$qyG*&w!9%spr?mE3
za+y_|hTf**e-<{kYS|_)b$YdJNkA#JjoF!7^u|^~!KjOTS#4@&I<`UHkN0?dx5>f%
zH8PRTYZtlQRewZ&o*5v!KG8fl-8pma1+{mPT_G-KpA_<RcOAKKe(D~L-9^`ZTPmfx
z9^BeGQ~g(8d#hRTNfj@ocju&;FWDc~E6tldfqC)iq^aM9C~5fHU3h=DDMYHeFo*jY
zYZ5G5iA8M+57;#Z!&GX1d-Zx!aEct}z6_(3*h*YsLuU`+pUQt~U*0t^_@2X96@pQF
z0JaB64gFvPX@>q&zvpu*<C9l2lM#$zR0Y5)=plnYa0u^Vom1P&8zmfue*i|M04#+b
zGWY|Bc+Pz5XqNQD&2MZ(m0I1q`ImVThY`H;|7;ZgV4TiUQkTtS_|medNU10bn?XV8
zg!uXIrfSBngHqU1x?7$pPY4^>os{9XW^8twBa`7zcYjsP<dEj@bA~@}=P&|;Fe(>d
zx%86^{=g?mU153V{B|r<mf_)tQL4D=SYZP<Ntc%ArOeGum&2CR3MbeEA#7lGMzULc
zllHBfa@Y#GTY;@W2pia)B}h(;cGs|DGkoa9wPdtd!UlF9Hg+_z*0j9%4Mo><b<Oep
zG+pBxPf$1Yla1X^o!;MArQTSLC3&?dgUMJ$tLm$M(V%PBcF14eMV8@Bi^|2hV}%VA
z_2^>l7nb=OJHKJg>&6wP+hR|}GZ}&OCSO%{w@IAccUd<zzwtCrv$sa{OfJ|d$B3fk
zn4mU6OV~g;`l=24*I90U`VBeWOOx_Ct6%Sw!#ru5D6$m^VFOqAuxpFy=37<Yijj|J
zyD{aSH<x7@UbGDAxVo~ifh%C^f>%b#Dqxt5MPT1}pR5=Oi6NctwY>kjvoTIqV#pn5
zS-*;D?7sd)Vn`jDvm9pIOZ8Vv3`Jaj1+)3y=eV7ON&jQzpXbW~#!v^0QqCIL0Hb8P
zvvkLxLZOmjB4FUc1OP*YMNQ*PGl6ag(*Zqu(B{w&#`!fh-BfKm#sFV*d^r%x!GMB6
z9l$4m1K^X<D~=TOOTo1z8&cF(tGfdhRSuZtKpCzGvkoonU@XJ<o>pgR(WiFBjb3_W
z$Cm>QFPSO&FqBLsSd;~6;Rdz9qFhK9gaXI_grdplVKTpJaPzRnv|uq2?7}w1-TW(B
z|6iXD8vI{nbx>l>Sb1N6lbC5-%=brf^aVfA`cb!5cYv<wT5P(!JWX!bs9nw7RYB*`
zkJgg(`V(X($ir~+cttI;l?^=xB&awb9S0Pl_F$$L6V#`DR@tEgvTwy^_{L;M`kg-2
zL+_9cRI`B=G@$<iCtm}t10Tg^1iQ>Sy`#D?fS!>IRFZ)@^iOz((3=k)QPADOa~ic(
z^zL<uSqdI0lK6<Dghl%~a}VIa0ASIs#7qGk=wmF}Yni!#1H+3&@5ihF9JpV^I1uR9
zksUW`^+(QcSX%YbYX|+Py8&~z$o*ji(*Zqj(!r>pt;C+dg;v_|kQv;fWB$9C24nv5
zDsRoXzEMn3u~GG`$b@ZnqZWM6iAd{rx{;l9a1G`Xk%U8~!O)EXRSul9qW+=kV(7-6
z7jxAtI|N)>10_JU1ZcwGg2|JPosrq<%cPX(JZ9&!<j)qgD1<a2507XTr68>!4^Lzk
zr6{c^56=P?r6jE+50fUeI-S)<$`1239I(&zhxeY*mBwQFhJ^*K?~&>5$BD~dFIw>Y
zpxBH9WtAQ_lb1J&c`PhIWN=~O0Uek{Ig*aEVHC0`SH9~K5qXhbNM8Uy0O?yxu0{IR
zk?RnO=0_tVWB4({1hq)}-$0cE-H}D<@^tm!#u*Wk2VN-btD|i<m?fZ*q1Nfy14WQ%
zHjl>}o()%_V{@Ps5<O`$P)r6K81w^aLopk0VA>e)6U7q1fqufG8h~O0;5<B9s&vg~
z--WD)X6mlk^Hy4@1Auuz6jHFf5>t6z|MT#P>1MC#j**^L%(hC6mkk^zX1%iL$AY2`
z<A70zQF)tfs|J2XQ*4JSrII+~_E<+TkE+sFmG=JOYRag$t`52VSE2+jNNj6ZV~TAJ
zD@!p2!Zb<;(8y!}3n4h5zBVPnzvlNi+BPaY6$5B;SZorwu*ihBS(GtpjCkt|at1;s
zqzOV6e2Zb<Wl@%VON6ZWRtPze4#+$w(g`6Kz6&yPA%7u4-lR7&(uedxD3BkBj11xj
zArwVMAtN{OHx2i}BAV9c$m8tTyn`_n<ip=bZR69Qt&Mw)GGB_>_>2iZNOJB^PiaGN
zb-6~+08Kqf=X>;$cha|y0Ok=<?6f!ppJOX{7>!z@%X6eMavQ7<(#d+d_H;X#16b4?
z?Kuc}YI`CSp&c<ymrE}{|6J5}hgRCiV1;Kj?Svyw99V6`0r0Iu%)h%W({?Ca^zabV
z)_vz&k1W&A)QNS#oI{^cBQ5x)MW0HSB{Qt52Ff^9T~!m9<l#Zi#f5J*BFER^>P!^r
zW-q|^0=6h2CICDE<wSrdqAUQo0OcfrC!xF#;QLTc1$Zi8ixO-az|&Ap2Y5Ql834~f
zITPTSC}#mY3*|h3=K;1TVdn!pALRmo7oc1S@IsV}0A7T0F~Eyat^#-!U_)cDsA_;$
zqg(^<8kB1RUW;-a!0S-12Y9`RpGBQ)D2>s%JB~%|1;k#!MGZkDpqvPZM3e=95TKj{
zh$NKv0b(D@seniYT+|Rm8p`Q_NJlvX5E&?E0wNRTEI?$ToCk<Jz(ox~<fB{whys)g
z0a1u@5g>|CE(Sy~%2j}<0$kJ(L^aAafT%&a77(>4*8!pq<$6HWi}+b$m0<meC#7P6
z3%>%0B~n=L(RIkci@&H2(c|e2*CFravwNciKVKhpZiAl^#8MM{Ya5RXKOBg{+>|t(
zFA8%r(hMPU(i|aMzOA=tq#fT5A(Brb<V-pv^X8ND5n4blKxi?4F*47c?~c$C{t|?I
zNnd21AL)mXKk1K92tNdw7s?MsD2yM5(0Xz`GH(O90iliLMufKTw;=Pj^0y+ijlXS}
z{<)srFV!y9J!(V3?=b(G);~r5nYrsmovWDR(cgq9|0Q%e3~fF9p>_BMe5n||0gI&J
z8}Pt_Rg-A3sC9u+tLfDNm7`wkrZACRDjgmfv1V=P>M(5J{?OkVc!6ONqRUXxWtix4
zt>`jRbQvzX3=9vV2727W7mtyUlwwND$g;*N$PsLLr$tT+T@)SX=_=wbO5>F|Tvclu
zJ93u&>^Vj%2DVeI^fbSAcbvM$1g%MvwWso?=}$MFVPbAMbE1Wsp{aw)I32V3uIm`i
KOXe;BeC&VOq_n^Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp.52CCAAF6582C3CD0.idx b/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp.52CCAAF6582C3CD0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ab5b34b003d10dc3e54d40bbf41e8aaaba820284
GIT binary patch
literal 15106
zcmeHOXINB6+di`#*ky|>=<WhaS5U+Ps8ky$A~pm>MKrpAQVtyyEU{pW5u|sCMieDh
z?AX9As1P+46r<Q1YDm=BeP?%lIJ25B{`may=DqZ~7`*41GiT1s)8?MRXQ+qAD`$#w
z4<4En8ImxCrYK4m|HLQ6g+*Xc)0U!=O6m&6M>jf(=RKRt4~Wf~+DIkGv{~!%KqGO1
zt6n>&Y~G*VL#==K>*5pdtq&VS73~M3=Y#K@$~3v#bbDduLh-V@C+r3m-n^Zh)qk#w
zTzmJPePtK7ojIBjG1Jy~xpC*J$F(<)4*b65#*07a-*|d!z?O`{Q;$pK(-Y^!&P`eR
z`gZS$Z%Q)mhuS)slrA{?Bu8`N8~xvw-tE6Oq1WKLh(%8d3N}v3&i8i@m}yy9kP#NL
z;=t&sg#okf`$i>(^gY?`w8T;4!6S>a8}hwYpN$zKxYx7lwPz>$<5yl8Ej)24&^pHD
z$Qws9o#9i=jVtzQwXM+fs43~3KmU)>3H5tkyyit2d4(<M?8^PV>m9RQBeywv8X7F-
z<^?s5%U;=WglU6LTET_w;qUJMx@&XCnHPl4r!)Dq{lGt#)*6=laddL3Iq$}b;HFNx
zncD&)dvdI~C7QhiO-r1lF(udZBbSah)vwxl-J#9$GsjJPH}OmY>rxWGjW~3#r(rjp
zkQ*sihZ{I}zc^g$b}8-W@_v)0Yb;NV+je2+^VPi{+jtq87CUV?H}a)}<1pi*oZ+77
z{^{oNgN+_;Um1SCum8eZdby@sc-dWk(9mhtpH(qs;z5t-p24TaBrM+N-DrL1`<g@J
zP9>~gUnt(Y^RJH6Di`<RbB(W9WUn5uw&R}bI~RGoq{Q+|w?y3wuz42K5HKgqsJk$3
zrrUx}Elr`@9iNzJ&)Il+$H3UeK!ex8HErW9wJv*9>`2bnsD0qd9TNTvzrip(qu=pP
zzodocRF7z$ltJzBsazLuyKRLeSr|L)=<G<hwEOP0hc_DvkB3YRalbSuMbOy$xn&!M
zwh|m)XJ@{1{<*1RozK*?(eH6{_~OMb&sIL`Y*y*?$I)Ac)mtVl3+izE_>Z1;AustO
zCR7hS_JV)0`<}AZ`PbicHaoj{%X4S5K4xP&$C|H}8$~{<SeE(Nb>sfj<8}9kE}2l<
z?*8w=F})7lG+v)<y#B>lx8*reR#r!UeDJGz(<_UsE!!)0dW(-<onM`?JE}=5%htw_
z7clR_xklH}=U4a=$BEzXI{d!<&ifvJ?Fq?#IrjBg8<XX2lE!rEvgzHq@tbN34z@IP
zu3J5H*!bAxfg>;Rk5xv$ZLF>A=)Bv$qBhvm-R@=Dibl&<d+(GN*;Vh!9^p?HmpX1f
zJm<T4F0ReO3p(d#INl!Ya#M68sK=G2VfF278pBr2JTvn0zOh$b7o-mgOYY!tbFD^G
z_|~-OqOLRN&#GpYk4U)Dp<VB*19#2wxt1DUI$I}<TPQf=w7u_%?XnXaWGCLd&DDHp
z(Zln;*@vI(f*-V-;`Swm`gQNxRCoTuzS{hxqa!Ttx$mAft#3}*MU&f4hxK}8wBp{h
zuvepZO^INtQ?@>2Op;UI&JA3b;+LW~F?{RbYw^`hd8RFwzS|srx#)S1g1R9i&UBqI
zW9QTPeJgWj4jK7+Nd2oJ^|L(s{59@LO;!A*D>3^mzBw84=KCQn(H(xR(x{Tvq=i)?
zjdab2db5AEdQZLU>ZMD?Cr8RDO^Tw2YMLJ$-n@y+nL<&FC*{dk7L>v_;eVo#6=9>U
z?$(xa!on$LIii*$`ok}1qT6$A|IV`ucX2tBaSsEi0eocvV%bkKQg$X?+05lk!{UXA
zT8Mj4{emV2Ay&i0e<nqWs2RA431z}p79ei#a@6sdrtuaoM~=k-Q~+ODfVe}qQTr?=
zo0f^FBrLX~toX{JRb1)A)g1M3wunl`Vn@o6uPi`puDK}0d9=TufHMw@b+mNaDGTMt
z<Dz5Nd#CBD=QRS(SS*%kNpzJ3h+T4aN434)c%I9dh{e%VG+$Xj=dN3N7n~mK7b>En
zuy_NaHX#0oU(lrA>CMA-g++7zVS*=WM=W^P^7|7hXIkk0F@ux6{Vy6=PW+b33Bxb?
z+hvkJJ8#>vr@>4_CE_NZ%wp(d&$NxBtM$2@04(NeaUGNed}!nuYo9~j)lx1e1dBhJ
z<e2!9EX^aw83AVu77Mk6d}RSo$1YfZE4rw2J(n{Ti$7U|#2tYK%UmBm=5hkDSWe0L
z$^xcnQNI1z_;@Ck%lQV2y(w?LvH<atOzZePbuINGDjthJ(ZAe1KRyhv|NVe~Gaid|
zwRAfu3m95qo0O^(wjOa@P7oINpnC9?1$4g5$ydu}%XYbln*Mj<5{nhp5kX$fA}Rrk
zKe3)4AKxwRId)pAh>FAFPe#7H*YA0~em-!B%L&HfPxSuh293lr&MIvY6@$f}DAzin
zmG+vwkGqJdSS&W8jQGj|M!x>8ebu3@`@;kre=HVhiTKI_zPmyzo@&`IS@#cmx7Um@
z)NAVVy@-m&J$$k($0~n2dB^`Z(LbnVU9ThVkKRpP^bbug_tx_5a`VOmE@#r;eZ}R2
z4i&3b)wFOq6R`MiE$Gh6b^+hsoVZp%&BRUGQ|-md0_Nc*bMCm^qy0++)U3Y~$AwId
z-_rxn%9B3ESqD#UK|)5)Q!fy-hNI71b>>$5j5x~4h0%i7<Zx8p^24msN^~nE)MVhH
z6R$NFtv~f}LjF<wtU@cnPii{`{b?|1SrP}_U`$al8sa<%W0m2?^&0$B;*@Fo_y)*h
zMUOZ8IY?<TU_=}>#E<s4aI*NsVm$6FB+o(-g`rkCBZY7Ji+V0D4$re~<Oms|u22RR
zDr7{uA{i(LxOY56)w{uw=vovFbl0?1j9ulc!HvUL;4fEv6sQxeM0+)IeG*M})akzZ
z$?oOdnC*B_1xQ|i7#LJLx6&y8<(dtj{dQ^BDfP!k?tO)ONJH{86b?PWIPm*F1d-Oc
zS05+@75oJX%YuQn?>>!wuyUZoCtnOv26wD*j?kV~qSKSUZC+(Ce+I-F{ItRcLZc2J
zVhpZUI5(&^_B)xF_lH>rT&>bv&}Oq`*F1lJY|$G>LWRD9216|3@YQ0FkdX#SV<8e*
zjG?X_DDNe~ohrN)_%Dm6{@Ef+{voU^>MQy2%0fT*C5mtcV-+&(2elssBk3GbkX0n?
zyICS+Qc+|oN`mQEEJsutN@np=6t$GZ=_o3l#2F|$1I=XHXQJp#5@(_4ED~p<m~1qQ
zZJ&dpa>)L2QFJcZJ`Y9bkvJbk=aaYq#T1bJ6{46z5*MMEA`%y)*kUxBeSay6Dn-dK
zy+URcid}_fv$zb!m7zH-UW4M+khmPhm9w}<Hy6nm`MwUCLZ%earAW-q|0+bUB5@g_
z%SgNi(Q8Otj_7h0+x5Pe5)=`VjMYNj4(v0g_kH0voreVqy#_URu-+)rZ938SBNcRN
zUe;uC&?F4ERwy`xeaLgLaVX*t44PJ=MI2-Q(foxOknP}SRq77S`j&T_*!Ww901Ff`
z5aKzcKDb(4W&2y6n`M7<Wwel1I$a2Kw9=5lu4&}}!60d+F+jAYm1YQ5sq#&bCa8Qg
z1a4Y6B8bx}V~0$MR>nMtf~s$Ym`s28MEGX+)oTyf+PAaE{wfE_b69^hs5bwB-z$qg
z*bR$$VgoR79A$WV1`uVuimEf}Q<_D%k#wkZB$%j>kqwo#qG2aG-Cf5SH4eW*;WS`!
zN0#>v*jUi!yU)5*-`hv(lKt*K|FGf^7yU-+S>4NW)szYuGea}eD={a!ht6s~jgh69
z4U4)Mb}@~b{OC;ks}c8wOc%>8HY~C=v^7mkTsJn~W$-g0V{2(^!y;!xXHy7dLdMzB
z*@i`KhHj=1GK7qqrJD_lybQfeA&?3gFH0{Q7Wo+Zn6C3X(=%?twQ?cjW9ehVqCmqy
zQ}EV8CeSj_hDE`K!KRlgYR)d}BkC?>f-Qq>SQKj*YYK`iWMVC2ZCJDk8E!(R^K7@S
zufJ7gEo3$!%T34zuzy3qqx}ghJ+K-pYy(Wi{}0BypCGJKLSq+7ry?D2DKvJWbQ+15
zB6=x_(-ECc;tWJ*kT?_3nIz6abQX!T5uHuq97N}kI2X~mB+f&09*Of2oloKdL>G{_
z5YdGsE<$t>iHi|kOk$O%0~bkS7pm}dELM3s7OOlRi&dTuFl>kLe5-r9N%aFaEYA@<
z|2pFzVhhaJ|98eADa77i<*32;K*kCUCa14AI#n_EBwiXt&IS9MDZ__;)7?Dh_1A^@
z!V3eD3ePv0vrhFl+Im&osDmTLQY2r>1|!(y<Gr9_hY!mrWQ+txuCTWeGA05OS6DkC
zV<oV1g*-sW^bquLg`88!I0_tHVP7I-1_%bY!u~+WcnUmSVV@*qyanE_5cq{mfFQsX
zG7%vo7sy?K0}wLNf@oK8qCy;I1sjknB+j8FdY5kX-1r`UFq!;DZr59x;{5Iut|qe<
z7&X>gGFw+=4iLVg^ij~{UrH8yHh#hbIKw7iqW((QW@)PvAULi|qy}tUm$sLh5Xem0
zi9lA;t}K#CWW;l__A=r*nVF1uPG%)*eGayic&zF^@?)PFo7C@@4kymO?LsqV5;J{>
z_cUY9GuLKO7fBa=_<owP=2>gA$W~&j5AmL6?09zCEVAd>6UbTOtPjzjW?XnK+AMOD
zxamXmrx|yiyEcowBwqRu{b^<xZ<scVd?Y^l5dCS!m*=a^B0ruVfdVCg`j89I%tYQq
zZ59Pfg7qO6pqXI%;9e{Wu@528RQss}nr1(ZK%w@b1PZecBT%d)Rv$77nu+7ZX|pJv
z7f+xBUIKv<d5Hv?&YMo48N3+;O5!CED4CZ`pqadx1e(R0MW9VcvI*%!HbXO;5pOfn
zW^o1LRgicK;%y=ER>a%NV(<po=E$wEDkF#G28QFan^n*x+>o?)b?XP$fnTX0slcnL
z(sx+L#Jk*(V^$6gUt`c0jzQCh4La3zTpw>7Y!;x73s4`J|KyfQBMe_}eu15`f}Di#
zGAn+EFnY<{{%!u7H+L$h30@S^98lYjH-BI&SWDG;9@q#&_?kbba^b<I>$~wjFBQpC
zQ5ftA6l_3HDGG;-4SWOck=#?L?=Tk4KJ1%y;rtpr!3yRXV(p>@mM7Of)tZYJBp1nZ
zSw;}TFSZlKHXPPTwFL&R3EO1cDar4pau2Lrd?;?7Vuu<a{f3B70kT$<7NfVxVaK2b
zK5o=aYRfi~+Djb><RtA&p#D;K0*#XTvPjlVM)V@Hml3_loMc2Vvi>rn7uhITYcJ^|
zGhd9KvDxmc;eTPnPj3hqZ{8_ih*d&mE}&4tk9G+eZO$ExE;y}^RZ2Y>$ubRi>OK5%
z-q)r1!b_7YN$l-#+OFM4o&o44O#Ld(!PxbZ1<8`Z>hgbxK63o<jE(qNbylw6w>KC6
zdUI?u_A7Zvp2zwX*bZR7q6pH^t%{O^2o9bOdo*<=_*f~5qBL`U*Gu0nR`^H-Zvy=(
zsAmOl!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~!eSL~
z!eSL~!eSL~0=W2y@9+(y&iX1jcyTHk0aK`$2yq&Tmm={}5~m|^I*BuoID^EQNSsOH
zEF{h%aW)cXlQ;*7b4Z+v#JMESL*hIV=Ob}Gi3^apfW(DJTu9;~BrYOxF%lP(xD@G@
zqLC0B6&$?oDiW6=-7*rdLAq;5T#j_hSqv;O)^Ji^ffD{;G_VXM0d++_?=Hn+wa-WL
zd{*s&G{)`8a0(k|Yz^eOlZw2C9n}hh1cxd^e;@p;ddBb*f6>ah)6>plfx7jT1`P;2
z5Ip}q9!3$87qLfb<znaj6hV;wS10;~ohVTKcqzy;4>*m$)rzzh)CBS;?BB>U4)qfc
zICj-gKa0odl8&@3IAF~B>&RNs`iCjW<FA*bD-thlX&a)EwyYIxm~lUC?Jsun&vF<W
z>-yCi^Q+9hD$5sE7PwCFiQR4qz4|1TwhsO}<6n5jcWldDGG%K6%iF=t0NAYoDP8H2
zm$YfX-G=|64JS}z*x8UbOVsa~z`-R}Ns<=XTRV4K$hXE1aW(0v_P=+vYql+J`KqU1
zSWl}XFRzao(6AZ@?o1@lWb=pW<A+bK^=NJ@#PmB6NHG0QTvx&LJ8{nh)9*xrjp=vd
zvI(Z&iTgO1e#evm1Sd?t6B~X^zY`Z+F#S%DW|)2_&i57cyW-9w^mFLJ>W*dGeimW9
zBPktqg6HV$?ic2*P2PhQjN~qk?n%?6>*mXFHL0`5hJ+ju>-pP$Hgv&tGyJ^z2I5%d
zjiomd`d46qVm^R?Ja*+uRd~VJgS3z-M)G15$)?X`D5wlYvFYR*6tsrK<tV6}#WnY6
zUxAaaA|+GI6tIo8i!!Y*YQL!c=s3N$eyz>*_I(tr7kS+RZo+*s`ugwQSe-j^I{2gB
zf4ONjYQnX~=~!zdCX$X2_JOg+3qba+{;msS`YuAczPE9g<b(hdik~IFL;b6OrNh}P
zsgx?N4wfJ2EZj(4KY$nmLJZPf{7UjXQ2lhU>E^rvS-KU9i)z_Op3P>qiabAP6^ev~
zy5UqXU$4gB|Er7o!Y=A&fnWb9N7D)aqWd;CAO59(AVmjZ`{0Rq&Ty~^=R#QFNJ><l
zAieE%`JPi@)%LG?^o8~4PTkM@%Vv)lg%>3Q$urnc2iFQk9JQ-U|IH@=D#r%byY5a4
z(;u|%Z3%X%q$6uiJKwT8@!n<p!G%a($hxk_4koAKCjWE<3(}E1ojrknre*ii%cw#F
zX@7KD@<OMiIX-ROxzEPImlOO>R`%NX=wD>4aMkQzyZo<pvs{=h-#+;l&H0;iC2zIr
zwQl?m?Rwric70lv+(`o)mDp442UjWMLYOJ<afS*pp!H9hS}WjAN^1pNJ!!3g`zEaw
zaCM}$0@BLX3djRnE8uvowE}qY)(W5oTPq;jVk<uUxdUu&DcIDuR=}<n0wDX@@o_=%
z7K%(g64G^wifH|>2IH0(>RB^r%xF$TBo!1HjFWSU8Xf7YL2GJpdD;S@NUx2qSZApB
z&`~4WdU|we%OBCsKq574@7lNjfPrrAgKTBp`*m{eYW8=v`bNek<`x|-J6qdy>tSc#
e%fZRT)UiX)-a};iR(*zzoXzp_8$2BOQ2znBt5GTd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp.FB71D53BC1DE7E13.idx b/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp.FB71D53BC1DE7E13.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1208380bcf6571e2109efec2b978ad1e6309aec4
GIT binary patch
literal 15102
zcmeHOcUY9i)_-SxVHfsYSkT?2hALbT3!s94f`W>mU{FyM0T)owr6bK4j6KqO5sjiK
z#z?ViM8y!5V9m9lMx&BojY?uHXyTpS_3+M0Zv5l(k1zMp=VA8w&CEOPoH@TUXU2eW
zW5zsnp(xLYajDUfNwa8*qIB?IVp2lXTnrj|Q&ehY<J6!@?asoduge3*tCy_Zu$o&t
z$i`~ygi)O}1=V}Duhy!ww2sL5W6y(##U2+19D9(t@nK-bBA@M+%Pg0*xlJw6IBGQT
zOy;Z3mpZ;De!pPcSRNAm=fMr78*ko4?N|BfgWcN(RhVQ9535{fG$>?L)|S>2Wg$NX
zU;pvetxw}UZy%m*k#2ADg?9ePySbf}r!zwA#_kO5oo+w!^o%7YQ>ShZt1JohoStq~
zR(gNe4#T)2{mpwqx{7AZ3_9bWmZWxiar)sg6&C_NpKMS&4;erCdcJM>u#P<@9yOJx
zsyWYSdsn{`y(X2d;{H(EL<fJS<6ylya{VWkYj+O#v$iSP(`JIs=f`i<hkulCe{iPH
z_a{8uYK$%&D`0G*I8BSf4({3<`P;zILm3w5c8<Bb*wa>cl5R;cpMUlz$H+`e-YqAG
z#|69-LEDoDacsGjYJK=ORu2<TJo9MTo?~`GCiLoKL#>a_9XB1aeuZ{zN0ZB8$*d&3
z?2?s(J4Zemf2>!ZkYPdpT4OJ|%q?5KRm1V3Z1?uzQOo^x`wE1oi~{d{{=&)G_u|QM
zR+eLJ#un_a%2{w_!@Rr01DE}zTV&e6E3o-|H8&w^cw7DUu;Wg84jr~GUh5j7MZbAQ
z%(9z(;O8mhS8dt-%yLefm+OjN+A-PQ0qZB}UW;2evSw|Qv)Z|7PD%Ys(w$}1hjwWh
zF0EXb7dGyov%!EXi;B<g{c@(E<+Ik}21C!GC(HHb8a9P=1Zi1r`_0H`>+pVI18atN
z8O>Oyf6gl7i#CZX=Z?>;sk=`K>P0~-Q#VNuC7T_4V)paRIs@kT(49L*ULL(P`xk?>
z-rNC83YUjZ(Kmbg+4+vGNmUz+2ZfCB<>~m}^SuA@sK;?X{njb?G`unSlJmXeh29R4
zFEmFlJUs5$bIpa4YXJeRai10m@9v$paIJ7no^8=w-p36&dy@A}@J?BG@%p#-JDym6
zyD?&QQmTWq!HosE7u<5EymB<l9k+K);jMAMN9okRUX|k3|EvCeymq7<`SDV)(+zd=
z@M(jEyQuA3_okhF(9qX5?ts=K?(5k4Q$xF?Gpx!p5(l2zcJuDysZ$E1&$gUSnDD2b
zk#^4a<12O*iEg)CShhWE+3kymP4s6}6}t9XleKcNWqaN8{qw6rhfY1U?8SxMvK{6V
z`?LfZKkP?OyA?U7{?ij7RV}R-Z}mTVd}itqNu|xSU9&9~UN?wdXYuIbdP&tTw=d5<
zze~jp4{6?ea!6zI&d{NC<}S%vs}1Iyo%uzL@rH^U6&J>@I<3{{QT}9!!!K%)PveV*
z-M1LH>U+-+VfD&`raszfY3@hb=(u58h6d%CnyHTGDz>z7tgl*f7TG?$9CYTy_xH~v
zbPXHaWgOZuC#oyRvtG34$B~yWT$!U)_xftJQ{&d9lb7l?M(K`rclbDt=l#jto30hf
zV_uea{u*S}b!51nTcf96*UHML+QaVa@6Zoy^nALub#~V8T9;D>gl_!ya6+Hgb@sow
znm*$=L>k02Ee>xwK$|>jaZmqxbZcT-w3Je#DAdN$*uUgzGnW%aQ8I7JTT_uxa=V28
z2_nCU^1r@cL(GYqN6FS9Y8|5A{6H1S`jn-6Qh%u9a^~O`mLX~xuA=;bDtd$U{MTyE
z8q4KG;wsLRv!)_}Vso_>kuDzxPUdr_;bLud?LLY`IRc8ZV2ZT@?7ny)pwe)$F=ec&
zNZ@PBes|8&hn}9`bEe{Ak-A7nk-*pe^Pi{8uo?7>%bERGUk|+6aO#EolS}~>i;Ja{
zR8x^a<0H=2`R<L1<^0Y3M{jCE<Kq2l1XMh3VInnAQ;|RmerLUwi*=mU`BXYCHlz%N
ziUf*hYRq5qXXoQbV$Pho|7G>UCI(*8x0?AWmlK6s_)8f?I$u6uu;l6{0TqX<#8R=E
ziUc}|ORUUSYd$XHbEe>8MxD`AB(Oxu`EgUMmR}APPz!MJKx&|-B7v_LHD43PZ)(59
z<w*akIOE`|&GlDL^yP9QaIrOIt*J=h>lI$MRQ{xkv$&jaTx>#_XetsY&b%;a^=@OM
zn*u5U7kf}1nu-L9v)8n{+NQkD5Kzf~^>xvKwfE<petd+_3BtuX>N-Y>1SVEIBm7?a
zukkCnoLRVdIyGHWkwD`m|315Nmq_HrrxxPk9-`TH+wsuR9gR_ZP9QE8s0%a|2{c|O
zJ3lTr;8rx36M~C-$h}@Yk?L#|efc+<Q@^9^pow8aDVGz9Tj(+OBd7f~Y-#__Q9z~O
z;vNh8E#vMbUHelKmlKSOduZ;ry8~vG)*ks<Kqcbh9vl1EzH6-=f!73oqj%rsrf(^q
z=~*D4l5h(>R`1l{X3t;#oW0_2S~xrM(ihXjD;)$>3~u2oM16%c-~50K&igpWE<1Xr
zH<vRVSK+F2ofHWy^39{s8v=j#XSkdh|E-1gB(uZYrX`1QIWuvU9=iWvV!+atVe?D{
zR4Oj+VS5rHXD1#Qh`W0<^rQTy)&|7LbiH+hL9xI*ayUw}_-0(R0?jTI*KP9(d&Lz6
zS*K(j_2(ePzJdM1Q9&Bc;lJsJtV!j7af~akKx19JFjj8qgu$cNFPJswpl0lti>(#k
zt->S8N78&Wzih5`p&#?goqsXYIpNyn9}q_&ZIHiU(5i}g#r9V@jEvD?B(T>RnLtM%
z0UZ{E-*MVEGr9{8QLd?=yBrk+#vXT|91>U_9gFgjo@(~r;sT{%2b-_cwcg>y?YLcd
ztffd=iexa>KBKB)0@tavy>qtWezQ&=_xrjFw~&RTS!f=#0A1jfe?unSYp*&6AQxOw
zE?v+ZT7jyw%$M#z7q?rlBjvf|HTnk6K)%ml=NS%|L*Lsl(y7Oc&DQ+r{rI895KLuq
zy8uSqtTXcY*7T`;@iNL44OEY#f`HCAnkqEL$iyLH8L*XkrTypOYUu-T&vKgxRuZ@l
zXUmBWCS#G|uNLNgb_DKMo(-I)tOUs7NnnA%O7keLq$<9*#R+4;1C-Mxbi&9Cy$t<f
zs?aK4R&pe?0>O9&Bz@DBN130%!&UBX*slZoyNaer4;ra4vMNMZAtBqzMnrEUaW$f=
zNxT`+n@L=Q=o%ISU*jI-CJ&l*7;<N6`25H;oUN)Ij`}<Y)1wbEaDm){0V6osZWrj>
z_-^=H@&&rbwN;nU-Rk^8$My}e%$xbYaY<-}@PMRfg^>ocO)IuPaH~=s!5GmBEd=gU
zS`#qXv|>+!S3xU=4W^h@Od8gJR;V3tKmA4oV01k<Z8>7^*vAp?tU@F$WOtV5IiZJ(
zf0P=Yq<q7`{_$!lxi`7XVkovFLY`0`v>4k%o~}OF!_nElcFwDPHBA$1bARz5&?Jl;
z#6t+=Bz7T?o7kN|zG6QXN%~8ORwND*q7{jgglI+LCLvmp_)5B4fwjUDk{dVB37_)A
zxu=&t5a0=kM~i)kK1WNsk#B~VM$6uo;rLrQM}ocg{jTrrFFF3x@Z!n$KQuX^)FiaZ
zthE1|p1_Ry)eM=wy;N)YJFOjTE&)fS*?*2wP5Qa{RNCabPv7ht@SrC9=~~~(cT*EH
zGBX1+)5?4PuEx8^j=;#u%#KAi1~#U}VGEw7_LVp>G8-!!I~Lg+*qZ{8GBSHBdpj1n
z7`T{%C1+$VRxWld8f7rbbXoPDww)5)D~xQE)hIg_`55?^0!cG6A1fa_76ljtm;x&@
zvH+_9I~D~S1e<o4{Z_{>F3DwN!B)X`EQ&CQFx``5-*W9`{Uk;fVHIJ=qIiRNQ=kY&
z7H<`A$D&$fP>W0t)r=X?>B6;QWVOhu7TEy~{8ct_1V3t@h?Zp{IumIF4bqqv=`0dw
zBRZSJIf%|7aW0~BNt}o1JQC+4I-kS^h%O*;A)*UOT!iQ%5*H)7n8YQBE+KI#qDx6!
zhUhX9mm|8I#1)9HAhD7!fXrx_k}p`S<O>!n`GUntz5op8E4*V>i6LZU%V^6<<MnWX
z+_i>NOvvexU#%<TKm8zxH^dDK`F|Az_>$P4P*N+{RPcYH!mRP*LI;=^zWgA8_bdUa
z^UcJLm(TCKMRnjm;B2{13u2iaXuRXP@TxXuz-%PVX59xcq1eqN8B>)%WAspX*qO>D
z?_@+cu4t%+Brs)*#QH2{iw(sl1Tqu%C6KkaAB!X+3Gto8P(pksF_RGANvtK^-@#!F
zPfFEHylpn(Q%9UJ`tJKbY-pL8$V?Bojh317%r#hKBeKy0Zlh(kJX;MG*^BJ;fZJ%9
z1J6N&MUFg20=bA>^nmGTnJdp#gGHl6qx693XqhL^Q-ei5A|E{v6D{-Q`D(BzKop<{
zOh?Nm@g`}oXfkgyfr3TBdcb?MY$k7}28$v@5qiLTv@F6gVlay$9U}=e+i^C5<~YtF
z&|Js41d4KuB2c_2UJuxlmL>2KG+30#OC(SdFNr`4cnb)W%u6Ov3NM8~sk~GIrSZ}T
zl+H^h&_doq0@Wf>Ez$$#re$@ASBErMT#tD5ByK>w1`_W;yd5ltWB)r(Z3}L5Bac}-
zwSBOQ@7XRU`+A-1H*M%dZ0VLD%Oz+im_RtYVaq5FibD3n#NABn>I`?S|Ico_(u#pu
z1C9k{d;6@|9kNoLsqzR5)(KoB%;0j536?{7>IL)m*`j^Rnr__Oj}Kj$NScYFz>$>)
z0>Z1%Jn$#LT4B8>BQmP-83>EuF6B`Ywo5qzYrA5JRqK`~>KRymMMzr22FzBr>}RSH
za^yiDx%^;~ncO2ARvk@PEI$`c5l<(mV~V63!Py*tD<2;83`c#OO%5n~LuFw8qNyC5
zO~z+({PLa$rPMim*s||=%jP7mcos1sZ^m1a&1st+WZt)g!ID+#KZuQa%jeA1I^Psq
zdRFfItI~bG>GUTP$3HCoAhGu>F$f^xn4|ofh82t?M3Mws5TbK<yUT-hP)QjB1Lp%Q
z4K^I)!A<oTSMqsm^YL#|wqpUS2J#?c@sEwblwNyv>DjAeb=cY!BWW>f?ZD8!(}gXM
zQ7e3`o3l$OH_`G~1FW+=1TT*@uvi&uV6igRz+z>rfyK&L1B;ch1{N!04J=m18d$81
zHLzG2YhbZ5*1%$AtbxVKSObfdu?7|^V-0{SDi<~k^}Q>>N<v<WR8But9`NVoc)Y?T
zNLs?KFc1!2VKQc~8o^(u)ADn0<m+VIC>dQ{cVTmA+QgXkIGfbk^_nX-WF%)@=%J@o
z;U2+9B0hVK2T_is<?PK0oB;43$o5rxs}Z`CEaivmeK2G1*^GT@q-Ju?(N-VJi&O!)
zX6vPTMT{K`ZE<g~Zj4N#E$K$v?k`OXx>=bc_d+zpT0|ucNjLf`<!;uNGY-;s9;3IP
z8qxBR<{`5W0)Ni}uf3o3aJCxn02VlT;|Gy{%v9u+gYxbM4%BtcXAVh11%dcm(kq0Q
z0X_CS+)Cm`NuN`BSp@!a!`$nSGHKh050ZY*lHR?uX!WceZEVO2ZlK}t1OZBjbYKfC
zzk-B&c-33^y+5_Kzx43FhlQ+qGrDg~acZr{?Y94+(_`anH-`uI6`dbc@;$30#FVh6
zsVdRiN_9B;%C<O<woTYu$V1XR)@3<-{F~M-V;=No7?oRmM&<O%sGR2+m0$;>@G`(-
zVpI$hqw*$(QL%)KiU-fAT-GxxfoA5-+@KG*KgT<V3^R0ZGu4ZK<&8LONw`DD-^ydt
z(B!ePk<CKgK>K&nD7=co->aSrkNxn&p?RfKn`lN>fut2E8oX^rR*k}|Q4EAN7};hN
zzL~@|D7=QnEi3YDFKIk)dFuidF8FcB<jDu#SoIjsnRaXc89!c4#sU|ah%CW-fV+QO
zU#T##0`SZVzhRTVv4U)@6?|hic<O<*f^XsmF&C^Ad}}ud`e3c#4|IbN3f2m~vm3k|
zz*@oga0B+jTEX{rgNt;m75s^AK$lo6_|x6sn1{83FLeXKVy)oEy1|ngxmNJMLT>P&
z<VO2qk9?ha`75UaBrRb5VFq6ihHpgCAb^{Tn>&vEF>i}Ve!7XyM5(N#vQSJGN@H<0
zipeH%4vNVkaW0C@Md@t)JQSNp;(QdFPvQa;SAZ6>^$Sr<A=zFLiY+4R7o*r>5|^OZ
z5)zl9xKgscG89)v;&K#MPT~p_Ux7Yh`>#SVRVWP>M!rz-8__2$u0{#fXc3DyqlC>Q
zu0aVkECxwn??hF%Aeac&>ejH*&vZ`&27aJ}_e=-39x)q}SM6wiV<cc3!Ir^-AusTh
zuY=&yq`h>qdyKQ`jDOGrruKgu`c+sujKI?+UGMh$&-9Kg)kfI#@dP{<uxvm3T)8u>
z?9i?cTKAr{4&rXu2_aw5slKP{IJLF0`jZL%SSYzjn#(#M@Zt>%k$j!`AnQWsrUiGd
z;QD1qTE?35hfXFZ6T;Rv<ANL{&0(*de^+;S;bUCi3u(M{OY^Mj6*KDQsz%%c>5f0w
zu6&S~_2v0@a{h}MlUw%<-u}x!XlK0C$#<1CeER60H01f*o8zyi&rr~^FcVE+X(|f|
zvq+qcgxMs{LBbpo=OSS)iSv*!kHq;%m`~yYBrG6tArcmnxCjZ0NL-AB#Uw64!V(gf
zB4H_s%aE{)#N|j>PT~qAtRQg}(y2mzEbVSYIvYt`jdZF>ycy|iCUFhYsbTTUhVR{j
zme?J~>QCOIy*gj{hx%#tTmM7H%s-s_Jvo%tw-+`ILT}+@c%YgPDVy~gJG5|qr{LV)
zodE}Tpo;GQ^wgaJo@{ppxWwHV;Q4lEfMeX90e*9L2Jmxt26(~U8Q?N^XMiu<odHf)
zcLtC>uprf4l9&*lXd!p(qrtnUddUV((RZJ@cm8V5{Ag-^cRm#!9f3KOqCSqE)QeVA
z=khf8j6hdQN2on+;yC{ay}ifS^wylvM_(j1FmxO4?lE$dr<c8Cz=*ys{mlL<SI_t(
r6LSkotAE+r^&jZqIM`{JtEscmpdn)=de%dI{XXINOddTR1yKJ1Rc35c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_util.hpp.38E1B00A8B9BAD43.idx b/.cache/clangd/index/test_batched_gemm_bias_softmax_gemm_permute_util.hpp.38E1B00A8B9BAD43.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b57bd59bebf41fd0d06986a7d0a8ae2bf3fd4f4
GIT binary patch
literal 8154
zcmZ8l30zc1(w}Y+m^VFwz~C@|FoS~#3Jg96MMW{9;)iVFeJx&~Al?{7)OcVtDvI~@
zxhCfOJfdj45;dS`KryIb&=nT&zOo{y@xXW_Tkj3U>&4&C>iJXkx~seE>KYK+yLXwd
zO7%%%Z2FX>S!3;0Dpd>kpEhe|iZ4~EmN=+X={rsjs#;kU=6-$egIx=%C;g*m+db_b
zlxdBH>ePZLAtvfbOk!-`nbwql=MVjHLq+_wkpqPFo>xvKi<?(Ro&LkHv_s&^VRrWi
zJrBJ;!|XWmNtNCHffXU_RNUPj)qas1>h?vKQHR*?nhkn7E%J}#?^lejukpQ_+|Q?f
z*A>0LZszWHv)})<7+*g=eu~%2UZLKjmWQU#zrA?Ml}^e3{^pbJ4!&0(xNbkbE_TP3
zs)LKHd+fLWytBi;HAAnQI=-f#Lz{Pb^R0tY!xF8d96mejG=#){+uHcgghWr)ef_?S
zcDI~=DbB3gU+{d>>+G75v;T_u;r8k9=S9Jb9wkq`KOi;9`~Q<;Yu8+Q8uYmEMqTeD
z%fs_InJs<tp5ED)IkaQKz|;-Z`@%2%x*;R_Z@Y6QJe{ZMVOi;^3+p#<<D%omqrU5V
z_9x5d?Uo;pEE+y(u-k!y!N2YLwtURWvU9hS=FPME4UfONS>t56a4Jakc=Mj?zjwRw
z<ovw;W?i34?I-NNT3ohTb=1y3dQ8Z+z=3BcKQ780RkQHw#eW94JRF@iMO89u#kF@H
zcP^*4`SEe-!ir5TN9+jumw0P^@7Vjv$F8p4`(Ul`^W>K^Zkgh}1`po8s#r6FE@<!d
zSHT~rzUlX8`<^@U2fWRGeCS$2@OQKFLSs{w_g~ZDLfwu5KQGDEEUO57I6Y;grlRU@
zh4a(#qlQmh<PmYa!Y-}YkA$b4GABDPo0e_Az%g(4pg2|9+$rO-9Km2}yW~B2J|IX#
zeN-yZ#lgi<CMtC^Fd~&oTRFWt>BO5@4K;_TL|?&ICli&E#8j!8AKGc2alGb@hWZ2B
zTkzJ&1lcWCY;3p2B_&EjT?{I5k}ydp6J+bO#^Ik^oSteZ)2qZcM0i7V4FihkzwcZ#
z=aXxDG&BfiNDvY<GC{T}+q5vJ`gop(YJojg80#(*WSfIOpBkLu#WXYk7N{jcEi7Og
zP^676uYHp9gohgH3?i?I@S3<c3@8$qx@Fxq*6n*2+6qL@Q}uc3-Y}p@Xtwe8@NOpt
zfF2<7iU_Y@7TbU#;%_;^kvHe2xX?g2sy^yIOeV<gl6&~sfTYUS8tMcyyd%Opn87xn
zNOVD3KabPPA84p6h`c4jTjJI*ph(Q>-(O5R^kJBWdckeFI=HH3qEg$gE9StU$j-%?
z0UGKC$CWCiy2}K=?sK}LI_mY`1r7Cp1(r}@3H56jP^9nUZM!nUo|MB3Ao3>>{sc8`
z1BwjU8W&X`c;=FZ8eyYef|pJv$R28!SYaI8wWo%<!wmIAs3(So0Yyg4Zj-NGAAV0m
z1yD0ZNNFh(WRLo_VCtDey>Du$4mRoG;GvcYZgO;a&D`EQ?fSzp!PiMblABEM>q%LU
z4|N;z%2z{!Vf{KH)DhQ)0Y#>dB9H&~nS)M49btws!WcK1AbaNId6vx&mu}HeKiFV^
z5TKI@vcK3ex!|bFFRKj{!n6EOWUX0eRBu&;zCL4T3Al_COg>SV=q?ldI%~S?p@<)P
zRcI(gbJ5MgO^^wGopr#ie8k4us|MN*C)c*tc0zV`TD<p?{n>DX&0%xziSQl_-ZtO@
z*>i5Lojz-7$L2Hwv=1}|2FV238;;a}<-0KHBiI0_H%*x4Clh3E>?XEM*wn=U{s;@a
zAi@ixX&6vs)8O3rzDMR3HK(0m0gKw=A`@h9FB}+`^S82r4zwj~u!cC+kX{XgN^MW1
zUF}@_FkpQ1f}<MR21IlYI<-u2lR3|OFK;s_$xlNyuz*=G8)SkD9O-|3$NQt|Ebv8O
z8wI1gOpyKKvB|%#4e4I2p(bFr5?bkGg6yCCzcP5eFM^2T4ErA^jB}R>vQHmdIM;3N
zxo`+sz#cD**U1FgXWBNqzSp?-y$fv%lXp;eaFq$N3+HA%`Z{tw1RDWP?*kD&kd_St
zik#mw^zU7EFWw8*3p4l&{%$fscJcLD^ZU%dIu#-+%<z#2AK`*+1B%?9^{;Wa5@<D8
zFU;U4__@iXk$tDM+F9t}2{93NFhUrilL@l#m<q1hM`xykdjdOANHoX<*>@KeKDzL6
z(q4##z&=XUN5TJX1B%=e1I8@wbpcK=7(_-2BVA>J?E5wK`m>qN#DTVl%k^~h6lJ1P
zJJ?Bg|DZ10Zt#o~KQyPIpnaG+%vmP*^^@?@T~ppZx~rkBf$bys=wyQIryXbI4L42k
z*HAUETMMlXGC_8o$BShL<_sDL?g6SL3&}c}Ap6zznY*T6t{DP%4_})Ele<ii{n{b?
zkA6Of?F`gg(iHxQn${XCn&rD!X}c1K-8~KG4p;M%2rnT#+XkHBeX8T#xLX+`9jGUK
zJj`)eFPWg-&rF(-mJTrub21TZxmaEc$TMrJb>}V}p-enX{11}|^i?JnQ2zoN(HOs{
zXmzK?-?&YfXf;^-qfu(ZzS=%%lcfs&?9$xD9aq5=PX#%(<NYu(GCp!Rev647grqF5
z2t~w4Gzww52O%lf<8GO_n6Sm9C8~>QAH-}qA438Yvnk7_EpZ<n=|a(hH($hp!g2&{
zIETiXK#*$zWtlC6Z6R7*hKXAV+sfl@gl*&TPQrHb_-n$x=J7X#eZ%A3gze_>9>VtU
zcrRgld3=blLkh-P#ZHiG6h-YhY>tPb^c7;gLVBQ|VFjZW8tNA1tlkmBJ7UJVfkBuz
zyeDSN01Tpn;Uh6)z+jN24Sy0dCU7Rcq(LufM=Wlb_=*O-;_(|A^oGavG^k#|SU16L
zY(<lz*ec3aQ4hRCCf=piyR-)`gd-TYt(3wKv4(;fZIu%;n7ER%mHdo%sP&FABUV;0
zqf}O*ZQ{v*nc3<nWxjQktyAX1x(XJy6;+C07G+sfr)=dOwcexMaSK>&ISRHatHC~*
zIL2p;9~Kcz9P2aI4=<mI<9x>XVUl5DqEDh9hFm5l`6T&a`NG5uj|@+=3#{B=Gg8^6
z7|#;ImXJpCK}|u-R=z1)xk>dmX$Y#0W&nb=>P->cMA;_gaQbb_&o0S)s)eAF;gX?4
zeVDkKx~x`?rhmlxX&&vf)`HlGt|R)~Zb&i=thQ?{>4RoGDBiWrZ|k}?GcYkw9q5d;
z2q0acmx4|Z?Fr(BKETA2M0=9QxkQ`G<2<6x<8eOG=JWV0(VpdT0nrxlxR7WId0a%a
zMLfPpv=@0?Ng^v1A7x@SiLB;vG1V4RH#}=5UZUDdJT9Ty5+0XQZ7Gk-sJ4v9*QoXy
zk8e=z4IY<MZ8?uCskV~GRa9HW<60V7tN0HSpV7!?3dRB+Y(uKv(Vrlh#9JF<{gB@$
ze;TDpIZ5orUQWH-Fi&8058v8Kd&RE6Hr7&av{0<^!Re#|Uoq%Kge@YCcF)9nB<LQA
zKwH4tAADUZ{2M$@l~59$vvY&yO%rWKr|+0H=5*eDNFZ)4-1Jymql>}X((^!BdoG*H
zS$HY2mXhwc^4Pf8(IKV{@0%=v&k5MD?Jc2f_yaL~ASSH&Cil(0uI(Ne2Ocs#VEPcn
zkuqsOCLMz3f?)%68v_pyMYq{(HmBQVV!f>BHX}W%)VsLsqnEuH7#!FR9m!s7>(o}Q
zpdhCp4h1^}bEutDI}WvXYR{n%rw|UEAjT7<10IV#IH&O>k8_DJmxu~Kj~MfKoKKAT
zJU&Z|XL(#ej0OCBg~V9M<04`#;^i+A<3%19Q)4mhfX-(R&S$*D;}U8tp`yYsrN&Yo
zmr-LGkFQbVH6Guf#vA;6<<wZt<4S6*<mIcVu}Z=Ccm#WqUXF@$MaM@sI@jlq9g{A!
ziKu8|YqS5S*X7`0)=$Bl<-)VU`qK;UKFF~D^@;-%uaeNK%2_b6l7v?BxSE7k^Y}gu
zy|3s7(VB+V^7t7IeWqZ1*%I}C)ni|pkAAjb&F_yG6TM7c77QOu^fq~0uyACekIBb^
z<vSBwnOa%!DaJ%!ldlEqP$v4B{4Ds2VWQDwwBVbLi6)cDf^|9*{Z0NBd@(VxwW+lQ
zUjR%ro6Ht`Dl##^6kx%#U;aGCyK`xl0nX92g=<Sp<EtFc?)vKInTx<f=ef+|oNWzt
zS)-&ne89sEO0RdtL%(2OG$!Iz{p^<3X8+Ym&%_f%e}ed;KS;?;f0D<!M4!v!JfhFz
zaX!)K^Y|>$pXG4@(HHQzkmw6}TtxImJibWu7kOMsVk$|TqF*(MspfGp)fZD=^cN|a
z=`ZoPgz8IpTuSw&JT9a9G9F)}`fEJCLG?FyTu$}nJg%hrN*-5HeHD*uX-qARQ}laA
zW1cA(V?NYqQZoe;29&?V=Or=XqYcVm;`53au?S^Q)%v_9Ml5p~q%)s-V#NH$pj7mE
zON>}MO7)P(YHGkRj(r+%6}dmtu<q*cwG+BHcdMlKP4b$(_?ia2rV+S}HSufT4xjUV
zc#{Z#)jimrt*fJCoAs2fSDsE7^dV15EglRr>-<K%`ZTU?Q<G^i8Nswt$44>gVxli5
zA!r+o%^#tm<)f^=pR)Z_i@Rsy0m=^WIES(v9>XK{2#=3ac9h4*C_BdEla!t0@hQqq
z@%Sgoe&X>t%FZbmTT-w;sUL->!Ni^}JzHYwvTbU!4bNTI9EYeNRWxWRxJNiiTeAuo
z*rfu7)S|*a?{kUVoT(YFZ*s&r>AB55o$_sO2rdb(2?l(_GB`=s7sP-^ivbHXl^R=$
zraj_&B%r1ToAoV!(Kn?tXv!Pc!7EX}v=Yu3o2-nn6Q-1!Vak+U<GT*SDUT%IdbY|w
zD6z@6KOEb*a_rFxcw~s4v7Y_$TMXPVrPM&fKQpm&eCNg_hL0SmWTZBlVpQoYy)jNQ
zv4mJll-e2lYcR9aTvI-p&*nEa5KjNqHT<b%`->*Z;$GlFq|O^^bw(VJ*evGT8rXx*
ztZR?ghIJXJXKTdaMKf^&WgDocV$A>TnwW1uMX5c9%P=u2K58iHQ0$d{Po3VpnZ3PO
zMMA1bH2RRec$0+O<Z&GhsiV>8`1ay+8uDDh_<o1gY;8V8&5e|8R6PDhZP=FgnY+$|
zfsO1svKMyY7?h=5>qsxGQ!%f?d{S#r(Y{N3mqzXJ`37AFsX?gR!bHPFGv+@gCL5B?
z*ivKSB*P>#c7T|eYDhKXqm_wwNYEY95zhcx0wm}zkMEP9`#gR?f*vRsAIPAy)R0uv
zS-=)}W9-7o;r4C)NEE}9fUl)yB>tI+DTWjiK11$x-m=*B;*cRtObU+&CY3ss=;PR=
zgej#CB}zd&@$cCdP2gd_%HKL~J#)Otlvo16w^I93`L<Jhr$*nYamvZ2k%!lU&3c$U
z!tm9D#SMHb^){7nyT^BLj1G7sAS87-mHjSc3maVsODGVM`kcy4-Qv471_f+*LK5Ta
zP0X3#IkxtuVv;K;TcK1+OuS94x0QYgssg7e^*-@5{~Y@<&RZZEPE1@%OiM{yyiz89
zNlahzco{J*<MDE0TF&Dw)U<`RRXleqHErebHfq|&<L%V6UBNSxCXY3NRv7hEcq}-o
zlE?JTI`~<>K5gdsw1FT=McqGkK3aP)<h?^jYnUaxgUW)tQmHy9Kf{!t;mS{|^0TAz
F^FM-8BVGUi

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_gemm_fp16.cpp.A1B1DD17614F30D3.idx b/.cache/clangd/index/test_batched_gemm_gemm_fp16.cpp.A1B1DD17614F30D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c72336ba242faf47b44e1f0b2d35f9938c49693a
GIT binary patch
literal 13082
zcmeI2cT`kI+raN#uk5morMeJAKv{|?(xg}rCBTAG1X+qA^-=;z5m|b*L{W+w8xmtf
zV?=|(izSv|Fp3Bo3l>Bryx3c?qkgjsvv*dKcs~E~J%=1K`kQk1&df7Un>$nJ?5tvm
zVbdan;`qq<p#+9u%HaQ$`N=U{0GGKKCa(BJbl>NiE$`)rKd!9R*1l}fK&L4-w@%#N
zdTWbw-K9D~#;Z}9zrTAv=ZY{N6WxCCXo&OMSG!+qTV0Ng!s?7xb<ygZcKuN!el`D?
zYQ62zu+G<4?li4kE@<c+-gq*qC*r`CaREB-e{Qgyd^L2Ujrqn7r^eo}iEkDNm(9Bz
zUe|kY?cRM|J@Z}Jjgvjs#J%o%K-Rpn50A3EeOT}+E9Zvd_LxZP?W?;BGSk;wdZ9DC
zbJ4EaHItinSMTm#H??kRl-I)&*L}_=Wo|2XIfp1^<!&us*O~B+r4bv&+Sxj3y}z~~
zu}R(F_IDSo*H0O3HC1Dic+qK>^LF=df0!`;zL(I)`L>?zy|veVwQnBjs-<~si%E36
zX@rwdHLs^t95dtOS%<7wJ<6M-#zZe3V$Dn_IAf##O>~oCi2Eq!lb05MsxwT*TgH>b
zC2oq2tLZq=^`}#8k1AhXRD-D<sk^n<E=zYr;RRR5h}kRY_j~=E&Usq}K01H1@J5`b
zl9%{a#Ll?JSrL!hOlJ(a?!WI+-P2aZ5e9BrMDy@j_sU)?*|XBF3&%!yoUd?wbXssJ
zCVZCr<DL3muD-M|<x!&Yya~xmC$;UZPHM1Gv*<FhCp_EgR=maowA|M{5?c?uo%C>$
zTE+V4hQvGP+RF@#e>WB;)bG}{I_p(Iv%0u^_Ofj{+zC7P{IsPpuX7|r<Mz$y1N$E@
zD>~x$u#tCQPId7wHi<18zs+Bu92Z$Rj1l0Yc4yPm!zXrcdirNdTzZ1jn%?-Z?_-Dy
zwVSK&t$q2sC~?^BvZb%D7V-aZS#|R0S5|hl>sIexJ;UlkTlMPyB(>_>2G|FS&Yc)v
z+Boq$kNx|nRf>`fEtEa)#dmwt=G!jqnm@h6e@`>tFl*wsrN$ffxbA=cl||*A9cib=
z7}T<>=!&jKw_m+a#L-fBM31a-OWnVGq_tjq%vJHQlg&@P-CRzsy89$LaonaP&x;e6
z&*~kWzj$lMpAqwCW(O+Vb9ge=)BPLwg=M`D6>^IY<SYtO)vVh5Lx@3ad(MmTFAlEz
zt~q#bfVWZ!J$;U$v32sjxawOi$8X)#ZQd#z9nhVgJh^jpMXPephu9+K>wPm6#vh2?
zcsR!J7$>I0zRP{)vfPJbJ}7)YXZoJbNjb-zFT9-fRzJl1<%}tRrp2=Ks$aBjeE7qJ
zj%jI6gWC;qj~?CnHX&E|vi*JeqeJhq-Q#0v2MZeUWfxMHr%y<^+VpHa*<$+R2Zf$f
zLbJm?Z!T+#WAn0>?Af@|>fIg%cf$uq3V(c@Fo}NRLrswCRxRKCL0^r$7NufqYS3=k
zQyk*p__%ms<*|UdHxHX7o@2GGD6c%0V_r9HyW!rs#b)MD%TC-eI8mFcdUWx<if;}(
z751E|>JZE~6LjD=uLoAAq{oLn8G>O$GRn)({_s~C2TQ~-(hM_WN()S)J-~kq=G2p#
z4`XyVSS)C^!EBh)0ye9*KY!3>7xA7>oBMa8HL8Ep<9Ge>3!N4Mn$<8hrnG>gwOe}H
z6pY(v(`h2m9DoHdr3GvrHTq5~apcACEZQv4%vNA4OAG0Hfx@hLFKsmIl(q&fGz=pX
zumnO{z|oe^Z5BIi|LH&f&}>_}N&CUeZx7OG!QkOKm=04~z|ppC&pfLG1K!YSL7>?i
z^JYp5*lgz*GNED5`OS1%0B9btES~n?``&5|>EK||pn1T={K~y*M{L{PLZ{6J%>qoo
zlos&e5qkdFoSp05(rKZfdBDV?^o}oc`Zh*Pr-g&&B{;SOC;I<_Pe>fmdClWl?<zXY
z4|Jg`(1%M4xN^x=5B>}xdh1wNGH6!8RCv+?j?PcCK4gBf^fwL`51I$OUa7iLDWjsj
zmre@=%?_9YQ(C~$rQ3c^zvcT|6rDB)G;3fQOlbj|%YGfQZLP_bNDh|pcg^JtZ>#S2
z8}c|<5@;SEF{@T26&yLcsgi@mfMyrWg()rI!*^|18DFsRdm{~6c*H*}P0b&)!cFh3
z*ZfPD+TX@6bqlxM{tuE<?<E}eBhh+{g{6RZ7_fTvUNd5kKkIekU~!;%faDxF?&A|+
z<hz1S^Z&aUbTv<{>}_rf<Y3=`<^hZQVD=JKYaLg$e|h`Ym6NTeJy@T}!RCP{3=r((
zNHQhN7)V02*?(KSUe2T8Qt^cXO_L+29B^8nQs@^Ts1!i8MD(FpDcXkrr5U_IOzS&{
zmd3|Kv@&Kq6iyxe4o=PZNvoOjoG&%4KfMHJlbNE-kQ*m{E&EkL(}Q{DOmOsw4g$Q8
zWZF>5XmX_dW|PJ1?!^6IY(6f^$3x+0)%rsb<~E1Ng3*%W1~m&!j-a-p$xnxnPHs4V
zQI%R?^Fg(#n-lyhhqirU78;{N&z|%+^ai}i0?~rG@I5p(UnI4km+ez$lc~JasjwP!
z;w?ZZjhBXym?uUkotKVK1}_7lOkO5JS-dQSvU%AE<?wP4%H`!El!x>3aD?-5UOvhN
zIIjTZLY!BKauLoeLb(`MF2<+58uypai}tjaY;rlSyd32cT)710Qe3$d<#JrPoMLD+
zz%EJ7Ka?&)O0`O9rxlWRe$aMLxT>x_%nXR2L_tE8DckqHji<&+gaXP4+-V>y5_>0;
z3ORz(3Oe<vSg3YsAF<ws4%a7-Mp^?UEYSr}w?gAZW6BvVHmPEwVhuYPENX6O&-XO|
zvydEAXp_vwOi-T{c368Tn=HnOVw?x(Az4~tIm#tCQG#+QPL!ftjuYh+LzDT@s=<>A
z=87(+{L(^u`q5Y$lieKUo2MszG<`S~u!)kH!3hvj+SroZW00G>7aCd6NunpIP7=Cu
zvBfM^Bh^)=ndFrX(WiR)wU~q8LZ)Ts5A&FPQUa|AAvHbF<n;YG)Q^PJvO!zdrvsqA
z>w7p<bA(g{LCs7^Ee<q9g!Ij!&L^bm23r4VTdR(lTdG+C6_G8<mZ^xZmp^+NTixr=
zV3QmrjtUeMKx&*}2x)S(5z^u4QN&f^BICGfTx1+qlZ%Yw>TqS_gm*n`7oKzV|7Zu@
zPaAaqNb~duuPG5=xC_(0ciugnt`G@?LSn?BS*OX-bn-y=*gNS5z#~z$JH5x_iFoM3
zUQjJ@AE3W8<GkOJgEdu(pj!S+dF-$b+`Wdo@Dn;0E^Pm=!8i5kI?<f&d-p}Z2txbp
z>!D(Jr)=U-6*kG&;_Jfs021&`Dbm%_)rA7dCUph6rW6@z8R^bjWtwqS%S?++8VQU{
zDKghG*M*vdO_~eLO)0X|veShImQC6T>`W<g)N<5?wvJ6Y3LH%-5^4!`*G;r6PTr>U
zo=pk`LQ{%-w0v~onP8JX0v}U~g0zBkp>|`FL4qJtNO0-F5tHl$T(Umlhm-Mp7UwmY
z6;rQHkf}tf5;R%_xqw)JP#TejkeCo7luo20ltE-5lu2YFltpAAlucwKltbholuP6y
zl!p^}IKufjk&ki#P86VAh!ce<7vV$^%6*y@8ZDw<vr^orSt;(*tQ7ZYR>-r@H6$FV
zc5WX6j<FGEG;&mpxG0U`j)QG;nkFbb6Fn;e(kyXgp<@_)*72vxm8=(oo56#NOOXG+
z%?O<?oI6&3<=ep}{Ol!!4i(TF5|@C|8>oXo`VFtS!MmmeEJvCsO=jR^o)T2^)M$CN
zpTNh%Ls}{dKB4yKg?aHY!J(b~Y9IP|P<sIbg&G-olLI4R7E>+)W|L|xHGSx5vPlh=
zhCZ}#Y*L4%qYoVzHYs2U^r0GMlV&V4edsZ=NgI}pK6Fsnqyx)AAI2?g(uL)s4-Yk)
z^k#YMLw|=&2CxG3;XIPWlUuY;pMN);Ao=QiKD0~(smah(q)1m?mk%uyLFzH|6e%)N
zH{!$PAV_0|u_8q#3=@RR)y?_PP7$O9!$OfFJ9Rrgv{M9W&#+gd$Wh&q5A75|PGwA0
zq)4bP<U>0}kZue&MT*=R?g;s)`|zRFB1m6`uOdZ3>Op*HwFq)HW40nia~N|F3T6Z&
zG?y_Kp%6w0LZOUMgu)nM2!%7k5sF|$AQZ`nlu@#ddal#zL+d|M)5+5gngx5P_DBM~
zu+(^HGeMviz9503wD2^9#NlFu(j(I6Q9UyvG7!p)$V4bBGAohlnH`>uJS8V02kDs`
zk&93s9+8J5oR3H5<H(Z=@W=v`3-QQ8l#6h(2v4GZxfl;G#uK1gmP8}qPZGr?cvK0V
zOmQh5Rf=*s9#u{;jDlD+Imh!$a^X|E>5+|wHf@8J<g+cw0(Yk~qy5Kw0V5GNDsDWq
z1!*~TyB<yVasbCll0$+@G&7}^J#T5Q1N%#of`89v!4MM61a-0Vv+IkF-@J8z1{8!S
zH5ffa{>AuQct{HB)SqV>mj{8-sA*kPAZV)UQOE_WmLtlMCCL`8I(PAL#c(aqTM{Ef
zgR;V-^i@FG9*aRS{cJH^(H0n>Z}1%(Y!~YAuwiDsUv8m!2Y!(xO9`d9FGT=NBJ?Cc
zCrRi`b&?RY=w-TrJzmovtsdO7&)zeb*a4>jU1ym60o7;%5vFWFwItqxyRxRe_wvi$
z#y#Mq<%)7;Ck?tVz`I1FGx=C;UBb|nPgGXy=Yca00%!P0-~@ocnIcKxOpzpTrbrSv
zQzQwTDUt-v6iEVSiX?$EMUudoB1zy(QD5LpabMs}abMs}abMs}abMsJdENCz3SCRf
zoBG!riu(H6^CvRL)VG0Zc^|(buzB6g_g9j@^5)^9JY2TG(0c_7jOIS&Goh9FU4n3K
zLk-A;az=6lFmwhqh9e840a<XYsW$FNF4B{0!Iiab4xcjnSVaB@@G^zCsF1oo*uk^^
zu8<YZA3ThoeHeGB&m6H=zdIU80Gd^YS8@XcW7qw{^nhCxD}ch~C~~;S_9=2@w5#WN
z?n3dKCt$<qhID%v&Ha>ix45dsBJ2}^<L>6kr*xP{^agkJvv>99=o^oh5hf9XOZeGK
z2ydJRoB(*21vH?<PQh$MpVfj?yKmLf(3&=wBY}Atc&Q3=E%447{0Qns);q5jS5N6~
z`yX<rVfy|Hjtb#*4_IvU-U=prfZK;U7=0H;(4RQC!6|8P)tp!`7QI&a<-<ga2j=Ua
z4xY2mK4(YcukK2m(zX@YiVRVP%+EM_vZ1NU`GG1E<ouB9WRUYiTyv1~LvATS&JVdO
z0XaW#OAP%Vkn=-6c>p;-BnkjIKg6*FIX~o56Xg8B6*PQGqf6uoo{FFOCk$BQ_3KAs
z*NBc2Y><Z$72#o26j_4%m*C+rie!_exPK|i<+y)2#iw!vL(AH~Rsd3m8s}4sN=+^+
zzB={M&DoiT4wsrO9WJ0Z`YA?Br=e$1$F&7uf!QP2qoBpU;K=j;O0(oB=#A!KVP+3#
zH#Ah04_t4oI0hQIzk#u+69}gX#-iFxUbo%s%$riCy!RS#AhJYRvM)g-St<YJco<xo
z+jeX6RetyL8$4N`eX?%X1eHZ9?brx*2{lxmYkVCNY;}Xas7v=Qx2t#T)*jhExZR(<
z-SA6?{^JjI)_<FK{<1A=h9@}w>7sO*e!V-!(a`wRi?N_Ln#hA+s)1_xfbc=_#iv_S
z?_33sDZoVqR1EgWX2hA~kWD8*gIFY%-6s9T{i)#4jS$O^gs<m#d~Zh7<l_F9gn60g
z3Kp;3hVLDME6QaXUnx~QNGbmD!YA}C)EPkU$iMW*lXxBE+f;A0`{5TQ`tZvZuq3E8
zd#hYFA@1j#$6wUD_FkxXr*Kx`_WzJ$`ku3k-MjDl1hBlQLHp3|zO^9E+XPh0%S>`)
zTxdHuJTPd*;I*Nw0{Zx}3V0<btAMwVR7L-H$+8NVIG0txt4dh~JQ`F*{};Nl3h21P
zXj}HmF!;b=*deQcaR-)?9G)@~%yJ014-Z=G7PMtHEiwv=gh~U$qNsm{$4A7?1Lz&^
zGlUqbKxZhj*c=tEGEYh9CG?!3>gsH)%ABF5sji{*)x=2-QyiyGb21+`cCvx}XuZFy
u<qy{xp+9nzz{tdOjG2X{m5tp5U0dyO)-J>NhU48m=F{AKzn+fI#QqO*ez0Bu

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_gemm_util.hpp.1FF4EC29A4661CFF.idx b/.cache/clangd/index/test_batched_gemm_gemm_util.hpp.1FF4EC29A4661CFF.idx
new file mode 100755
index 0000000000000000000000000000000000000000..24635c22856b93a3961fa76b73ce65e2e756d0fa
GIT binary patch
literal 4388
zcmZ8kd0Z6t6`vWd*_q+kg_UJla5)y7U6vaZv?{0|NEVPo%2`wp5ky#QXc|rMf{LI5
zS~bQ4k5G$>0tyOh<P#H88&jhOu}M9!S`R=JjUmSTX5{T`;g1Jz-tQg1_j~XA9`mCj
zA{1f_o3%KqtT3lE+Z4kv#A5o=lBGR3h8;#9%67NK$1l6(YjY#Er~73u`^`QkfGR#y
z)2!9D?cW@})n|L9m*{`HgB?c)>oTV5@_M^3&XjyMc=qhz)W3YV`jsZ?e0atKt95r%
z-}E;&?PR_kbi8_~vT&C=;UeALlaCb-UKo;l%()lX`Y<bW&dY;QoqKQmXUl7CWajKG
zm(S-letx^l?fKw`MUL)zm&1{bHWmE^AK2BfH(QTiuH712^V7ix_7)%bovbRqoYQhM
zPo{73INh8;^!4<01>zmzZ@arkdh`!2v@A@rdotkCkvpUIrst{r%|#QJJeh9(Hm>$?
zzjud6QPwW69{>Jo{SCMJ{_={rxkaBu*G7)UI6HnfkiIk8q$di;18Ty<Lzoh)s*+{C
z;r=J%<@2Ne)*2M&P}Z}yFMi411KG>lf3%P`Hn+8u9}jBQmYM8QyOkyo(zK0#STl0@
zl6dNuttFP)v-bOvt}VQ?c%Svvj<-!m@;@8haWu<f$)|M#Wo(e`osqXbe)l>HHE%w@
zjz3%*(08<DUZCG<)vE7Kt-aD1)Tq1OH0zP-=(_gsWh=4H=-g{*Rjs0X%DZp8s;VM%
zH$QeRTNPI2M7Hk9^Ob&m;VY}oz9$Z^P3~JC3B~k(DqPsbJ#X5!$P-W7hH1je9b*-g
zLc$Y_L?Z6z>D10gL-W-%p};WK(%jO5Cz!~L;G|4Qf8z6K-E5@ofU&Vutb`|^J)z-x
zu3}_si#Z`f+R+x#Aw0oErl^)MOg;B@Ue%*D@6m(}($1hVq&xv_AOC6iilZN&q6uph
zngiuv#}g1wYxT@Le_!!EP1qyt94bf26A;g-9)6a8?3WapkR#vL=GG#fU?R>lhD9)&
zOT*k>rqP5L8F!+bBs>A_)N|CA1BJmeXu=g4&!)0zo`5*5y?$q2w<wY(C?rm&((QQy
z;`D!YFF$uI;s=_TfP72LB_f`H@2u!id9SwN3{6-d?F1^pmM5USFyVORx~$p(nh+te
zj?#&FVi51U;~TmmXcA2@NSsI|+VKR$In8N4N8Yy2vLTe{Vxgi?h9@8{nRUh|Q)m0a
zhHygKdvVcT+-~d_9NF;6aK+=vjvG;G(2+PQPQnuqSFi5Oj+)Y+ND~qyPN7n4c>>~^
z?Z=9D9P#o%mqY&TDSKO<fOzxtm8pBvf+MU5FC<pFD%Csz@s3PynUB8LjV7eX_yTHy
zoF^dO+4E`6f@@yCvm&%en-#G(JOS~Z^7|f(th!Iq#6*;LvAI~p6UckRx<hSUJ&zaC
z1dWU*Q^``Efc6pRe{Y{~adjn1EHW;mWOh6O@!wyJe6+so=OJ^#4H=hONaZ}iMCK-3
zL7nJmDBNB!b0SSxBI9XPnv^G?-SOu~v)!FY<l7dBGpS5Fo`Cqm&mlj&x#^0bq$1yP
zN^Z*&5P#uV^D0;Byoe_3kaj#3FXjn|FXdSLs-2fQ3v~(-Cs9duJOT0b*Hf-<w3B>f
zMW~SRNuo*CJOT0dT~p3lKHG2$8AoDg%2~`45I^mv*Iah#?WGB4WITq7k@5t@1EWpV
z?<aR9qiRLkj+CRACm?=aQ!)3-KN1^gf<)S>RH_|MKs;jZ|2o>K*+fn_a!Eu#EagrG
z!;8I2{q>e>c%`2W;ew27L>dWCKpQK`$<;sOI(iJn2#QB9cq)eV5A%-#BY<T%^$x>$
zcLIhXT@Jtivlw;>xx@xQ#?rTPSN=A?TR6j7T3OnGDL9Nwa_BZ@(Ex17B!_BniedGn
z-WE&(yrDxJwh1x&g!#;an5UG>hMEr@xQ+bggyk#(UHFZ3Ip7l<?8jyMac^jHz(6zV
z9rnSE*#xHqC7=+)CORc57mlF{!zMW;DWRYkHpMAL2^|lZiM(=vDMTC)7GTKody{L6
zCTKTpK@rDHi;0BlNz_R^A8wsyMlh^Wq_hSYK%7+R2wd`TLELnDX*+JpP9`VYLJ|yM
zOz|Kr+;%-quE#}?42Io+lN*fr3Y@Gk<~QNwCS$%DC##M58l0>#=Id~>&X}*q$$Ddc
z8%}OB=6B=dZh?m(fzriI2qEkevP23&!>B-qxH$n8#IUi_SQ(7h%x}V)_SDfY$K?fH
z(1haWhu|Qe%ol?LhCP4>AHZioO=QPJWqG^SjgGSgLkGpdO&q~iE}3f$wxCFmLvHQ}
zhnA7cByfmf58}ZG@tNQd1`_HKZW=+=6qilEqWJpi8w*obsZn}C37E1jq>D)4TuE0W
z&W&_4;@nAhBThxCj5rU{!@xl#D1vtb4+3BLz#`|8=j7gTx%lI}*_Smb!EvaBc`zD#
zl=Z5U@vCZv|4TBgy~17tgN|Wk3Yi9m2g5oj95m3|3@ca2H304m>!@(l0AMn#LZQ$A
zaxtuv!bt-i#<0!`XAJ-|3}o~++%$&h88(m1GayvO&>Tg`l}&OKw6%@(1ZaKeUG!w!
zs1|e!$wH$VIgAG%7HXtMx8<kbzkl3+TwdVDC=PC}!%H%(HcT50%?eM4lFdzbIJ8xn
z`XVl7VB0v8Fz=B`ZrBUq6_G_UsHmOG&FY3EHLd7s94tVOz%qbzxlI5(F>DYMWWWws
z2~4SXtAGIlbyA%Sik@NPq;WDxGsDJ9<7F^%VQD~KxQ#)Ga5=f$faQ%gq1)Hei)7<&
z2StH&xphK_&^yfA0P*k@C5&4t;C4-AXU*aY8zRu9vnAPbz`&#boZjP1opTxI=|tX-
zgH@*L#p6GZEhQ)bZX*$bFD8o(!MC;4Z_GZ~y9RlZ`bYyH9@torF1M}-x}{{Pq1n6M
zPV;}_x9izBzvlwyFE*Rg@EDmCtHfG(1ZV_gl3Q`W6T|w3`5Foc5|0jXi;m!TDY?{u
zL@=OGS##@-aHxPRFa*c2hw$J-0usR<ge-GQkl;ROVvq{z7n%+o;#MKyP+(Z#TsYLH
zUmAU3WOXUJpVn176F%P!`D?l4tK{Ty*|;Cw_b_C~xz=%?ACpa+iO^<lQc6@3Uzh|i
zF=>k5vxKXq%2QRaPBCnnJWT~VC&Q-8(^W8U7&b$mp@M&|7&cR$se;F3*j8NKiU$EC
zGHe^JZZqaj;p$Vy{ApZ$+L&+0)$IZgV1}~At$sqb{KEVU*@B?aA#MQ_ZnT12VR$Uq
z1aZ;-fi9#LbqBZpfdjPp6=b`jUNL-FG4b9l-S98<P|XI3zGRU;lItoDXe|tzB>d?q
c{HZbgNw6~><x-eV|Am=a<Ka08QyWCsZ@r|HN&o-=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_softmax_gemm_fp16.cpp.0592033B8F1CD807.idx b/.cache/clangd/index/test_batched_gemm_softmax_gemm_fp16.cpp.0592033B8F1CD807.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f43f2575278135871115a10bd4f6145177cabc82
GIT binary patch
literal 14694
zcmeHOX;f6lwm#Jy=!QNG-2yU-Ac%Iri9sAd6cmLZ4k0uFWojmmA_5LLUq|K{6qQMW
ziV7NyQ4_B?B8nQXMib`&uVWk{Y7|6$)!pnq1$u9OJb%2rC2LWvuXdets%r1rd)HUR
zL~k#zkiit?8R?xI7d11SrYK5+f0Je=P8ov1T`@%^m(<IyOf(xTE`9HPVwusZvDHmW
z)}`4!+rX1rd~Lcp#w%)abWhREx2?OJW>W<si*>aNP0n8%GJ4@x)-lT?mtWlJze@YC
z`Sx*Jo1guHP9DEgZ#2)@ZRq=Xqt1*za$8j2VALF2w<Fxiy-n_!uwd8vWbXlG*S+dr
zG(T*6<J`CI#-y_?doHDxoo?9V=|8@8uIr}f+hg{8duflh+q#OovXy6yj7q~xehcsn
znQvRX`qf-{_k$Cni-YDpNSZ!N?q+B3FU>(3Tiko}3)(!guSiz=k8XeLvo5kXPde#l
zY#Q_B=7Vk%PU$)(HW=78>F$}@AIZX&+owfn_U=CFUQNd0r-%A@_8DuqqT%Lu@+UsO
zJ7ri-iL|uvPn#<DUXt&$<mRycAwj=vZb{KAAC`ar8(v)O7>%WCEL$2QMHVg0v<dWx
zwq?veZFxSwZZXXI$1xOxPV@FgWNRL6DLT8mSaa#Esb-7L9kOmUuF@L$`eu+t&GV`H
zUMXi|f0XKXfAR2m@{+@Lz3ifd8!P86``)2X__4JccGvbQeRiU<3RSK-;^QWLk77=^
zCFN_MINLZOavLiz4X6p4mt$zPf$llRercpnj~Pd@cks<tiE3VGc^;cID&@H24{;HT
zQfo5nedqkNQv2DiKg$Z4b+H>}+&py6y1?Lm;YgROm&HLb{ryHb{VJNAvB5%j=+^DE
zM~+!GjJ65v!XJIE>BsF|TDP)$wCZ0bhCf<nrFUwLajemy%Xh!gu8KX`#e9m}*08u(
zw^sAuGLy4C7nfc%8p^xx6)v0A$PW`Hc5f<5ZwXpq{7XuH%7Nhi=2uQN9^IL)Wp<_K
z{B#YYLnj({Y|J+JeR$Aw{~>vi{LJ5;*;J*vyxjJ@w{`O1#+KJQ2X=&P9gzJxS>JoT
zeQ3p$$1YOi(6FbK!9J{spPoaHw_T=gcZ_g6x7<|EBcSHnxNg(+&z0V-os)3Yy!)Nx
zBHtpt94)?v`K!RwzXaAFaA+=mJaDk|x_`M0y~#;#c;0x~XN19x(m(DMU&$IY!GOuo
zS{hK{{Bk$5*3EXk%Y&y6^4GlEx@OhsKavlwvpu=vP*Ti~wu^Oj`TV(7V?$1^eEuln
zaM%8aw?5d@<A_K6{QC!^Lnn&&Oqu9g92(ZkFK$xgp5qDECrq-wztKAN<cpE7_pd2S
zyWG4t<=C0vef=6-p52+@edgK53lBsd4L^^#&~)_uPRIIzd;cvDI=%ejlWL<}@5RS=
zxkvu0?+~+JK4Q#`K=C=3uFGF!-O0Kp*FRV^NPCKd)4^Z%j(xh^%zR%#>OA-Sf`Wyy
zRlhxL`Z>wsMen+Ie$8zMOm~&;c%%Dp^0@h%MzkC#xvsTY%eky9arhpKMS~98y!+~y
zi_taR)?+pcjOwo2m^mi8w?&R?n>ns+TK5HOx4yacD&2kT)wh4luYYf98hXN@{IKtv
zJ+ZCU8-0&B8y^f^`1&iua?eY8^@n=*NSYfLu~Xqr7C*WWh|b*+QFADYHKL4klm(>_
z;UC_L13}k|=<Z#4F|ia|jHqHnxBpQ$DOfS^-1b$Dd|n8Cg9qiIqbxwI`EOIJzTS3|
zMbtDb-iWA;NT>Y|G|>^<3_g0N>Zy>Li<=lyhGJy_Vu?xc+_k>Leh^VJvDkz%(NPv4
zwz0eaFS_o}<3iroSgfa|Cs7uv{!^4q-PPbzbuWGRyeKSQL@lC~1w3v4{>8lAF-1S|
zd6Tf%hw{-;79bw<E`E3R<=wV?UN9C<q9*Am3lI;fYB~A*W}q9N7m39m%${F@j`<kx
zbIbX>7%cwi>;+mo>$u!)-r1EG75$O!O__bUXVaUPoBxA2u3lqkLbF{hpBIQfz}Mp2
zC<_>S($FuTmYF(RiKyux>wU6uqQ(Hz=kX#c4vXWdcpYT{gPS{a?xxZ7_kZ$vK_3$@
zZ0J{ASMK2`qT;dmgPB|Wt;=?*XUPO1m5jw7D0NB2^Lr^VUl$2^0az^55;DpHdY)=9
zCuC8==+}H+C>Hxu{yNG6x|VI(a$B#$DNRJp!eT4RN=I3MIQPJ?qdU**J{3_3Sp0!W
z)_NW-4%2+US41U#?B%=c%(^(XEE*@GW@E7f<)EW1;N?G!WriCtZ-W2Jk~pYS7W4bn
z@k{x<F#HB7CDl<D@P@-#Ze!OT{oYZ?lVLHV#W2bO#@;wbce|gZWP*sAfyG0pAv($e
zp1!4(v;XI~r@!%e5m@}e*l+EKTi0CMcgcV0UQ<fVw)B5>eb47j{@4c|@bx47?-W$=
zd2%fNVEhkk8di(T1}x(9!m;>+<@kF|4b#u-`CvZp8!Z0ldVk{YE*<jn2{y|l+~kAl
zd3MnHVr}KV%>Q8XiBZu>Wqt8vwR<kaQ+cfR=>w02t5pjQ79p?QPw{vv|J2@|R-*9}
z-&_3q;pJNJR`_XE*Fh=TgXGUmLe?eoRPN&o*X~JEjrLCfXZU5SNXMnm!v=6=BA&{H
zsXwCd4bWp+Icwn7Xr&IpqG{!PzzosK=?8<Pm39E;q3Sr8I;~uRuqJ8cY5_Z^l|Boc
z0j*r*u=eTpal*iL;=O)7xwbJCF>D5sWuP$OB-wWdil<zKzdWS$pWs$`9guLZOVHV+
z!2|L1Dbxz<KxA-f&5_5CjN4TOUK3^jS1Zg2rdo0-G1xBMh^N*8n3SZo!lC8fj9s{I
zc}SLr!l6$F9+lGr$}}(jg>nqWFOwAx^U%aYP<#&zS@m^5UG2huZt&aJia)Xv$yTCJ
z_(<2v3(@8;^S(D^*cC{;0(pQy(NMgS#3@LeLgG{;P9<>~5~q<k9f{LPoPop{B+f+Q
zOcG}yaTbZQkvN;gIY^vC;#?%oC2<}S=aD!ciStQZfW!qPE=1x&5|<!J3G!U$dQ0lQ
zzw2X$Ek%-25|<-MIf*Nfq=LkiNK(mRGugHaH(GX7cQP8YS52nD!}jL3TL<XJYTb=7
ze>|~v#(9V$s6d#pZk{`J;-;pRShtEO4RM>-0YR$3)9H{RFs$JKLsxi{+Ok!q4OXdI
z_aKDz|6f<b7mXqPoLbKp;SngjIJjNAcUrGXd=C}kYDLt948-2keZ<k>SI^*Th0g^q
zY`;zGXDx%EC!Jfj=PlTqDPF?$Qh{c5n^(U(@Y<VxeTEu-9q~_dO@=K&bP2+{Mb(c&
zWppWt%Mo2p;tE7pkhl`jl^pJ0_Hc*!t#cdj@X5Y*?72f*zaCl-movDLB7j3k=XF3X
zU(|{uHyhnLsZ*^#l>(vPDo+Nw20w!5K<QrFThU779a#U%f>{Z?q2<Wj>$#C9o<BwM
z2MQeeoyVPC{u}Ogs&=<MKi~28`glh5lD~t;_N>};aDbzMBX(F>NS4JptWjCb&)98k
zfdYmV>xsL7@nd8lHX@LT*qlHXVk-_w^d!V{5(5eGoWw*zJSVY`sGozW$BI|zMXja|
z=*z~Q_qOoH1>$Zi)-h-p0x5pBQ~2(L)nOgtJiGxyGBoh3b`Gjtt-Incg2TAU1~QMv
z^*?r6evla8{wVTDbrR75jgJm*-2AQoBiu-lk3ojU>wqR$4==x8E^ELa=s2>k?i@Wm
z*Z;wv?Mnbs2!FpWz3%6^YIbE<GgC852woW3n%Q&6+SJ+-zR$4Mw$}C>vNN@_jPBoW
zL%Cq)P=>X$wX^4tlc|#>EMtasvURfOkgKVy<^BczY&V#`ZeUnfTUUDyjWiu;36sLG
zBW*|8bI8}!*AjdW!}{9#+H)wtG{6!(0>cK_2H0~b#5BYb7B$0$*oN41D8V$ra;E*^
z1+7<&Z!m0vZGt_As*!0mvOHn)toZG|3QvZuMz+<+9&kYOwDH4)Q}!CtIAGH&kRAkV
z8V788C5clIokHSNM5mHC4bf>NPDgY)i8BzLLE=nAXOcJz(OD$UMszlba}b?F;#@@M
zk~k01c_hw9bUuj-5M4mxLPQsmSmpg8VAJj1pTjEe&taAK=djBA1BN|oC*K)-;OjT9
zHeaGTRZlv5WP05;uWLN)czTQN2<OsUqDCP<kEkgt5dQgDSM8A+=K6oDt8icM(7rti
z-h7_ePdu^UNwLo-w>lj+KG6Z)c3u}>(ZfM51<6vl*bKZ*r?oQmMe35e)`kdNP3};D
z<`L<1pm!M7NN6O5@XD|zLK7*Z$qZ{Hw30$3VpyqADuwMB!#W5Zq>x}T>=5A)DI}*1
z>ml@zLYQM%AEA#FGE|237y3&fk7L+L!bws{)EG8i7%v6Nh{3x`;YK8do!fL{Bo8}N
zvh*)vaQvnmi=DZ1Z1C{OIGT_=?2LKixwlf9H8V6bhS*25<^pqV4p|#o8$;})SsQ_k
zHizsC?TkTVG;1%g*XB?kK_3D+89Et5B&1nqfwMM;Tn$}~ArjK8o4`$*Ln94G8bc(c
z*-?T~+8pvV^fiV^NV8)FW3@TtC-5UsfMI|!#6_Bw31r$F3NZ{ZhPX(xp@L9t4uuKA
z2ox>|Cs2eSf<SVCoIsI+NCHI(q6ic%h$c{sAcjD(f>;727$z7)RHoSk#{?%1B|0V&
zD9JI2Kr<a@5@?p=ECS7ToK2u=WLS-iA+po#W+d2*v^l&53AT`UD-vuaaSamGa5&NP
zr+y)R`Y)7Q`G^%L4l)D9R$jgmP3Le5lBbY370FXcoQ5LPP(0T@9Yv;-I0HpykT?@X
zWuh5e`z$2SBHx#dBD2Z%IVduR#JMOkm&AD}Dvx|$K8ngGaRG`dAaNmzE<_33_e+qx
z1Wku%qu3Wlm!bp?m!p_+l*r)<6jMRsN)%JcVc?kX0#@w8AteWEz`jz!GJ}`kbwIEr
za3gYC+;NjU>mEPqSTC20pTpTF4rjB!9C^wvaNu|xsTSBQ7zmo1)Ao2r%<7)^aJA4t
z*bP{FU=OjL6$*vcYQ=V}RxyrX1oOHa*I1bmx&}W>?x^M`{lLVhFLu}czr6ESFbF{X
zz&;apteu(E1!je}O`V&6ZJ#5U8=z=$HA&n6`}DUZj+ZRlyZq42TR-A7e+81QKp~`$
z2`oXO;6uQMFbStP_5s^rSc!|o4R!}Bm)f4(^g?Sfw!3U3%jRfk$UZRVqac;wX%*oO
zu`;bTwoI0_r5?{QNxY@cGMRp^?A|+EO{Nz>y|-7xE3^G_Uw<=ayTfN!^NC$e_QjxY
zdh}eyU`tV~Czx-hofZKMCmUET3O@nWd6q#h#```g`mB~utd=6@-WuOlbz6>?9Am&(
zz#--8Yi+;1YCE<aSLYyE4!085bghaAc%tzF>xT^LYo==!{Lv+QmssHk6s!v<PLW$F
zSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQiefSQief
zSQiefSQo(H?{GfTk?-rw|CI!tUNiCCr35^~xk#4F&2ZV0JGOG&BRjP)b7C#F1>cBK
zA8~&IIf$J(B(atdEhT*<L`#W-MBQ@z<=g_@y4)PBK9Z!Zln<YDFmlOz+?fI-E8x!9
zs>MTAEfdO(Kd-w_yzaLCzHII{w@R`V{Ft^_m(ZlPM2%|ZJjmR1#v$S_Hqdb5i<L@7
z$J*1wPFU#FS$)>sC)VAL*y~R-=r1BaPwXe2Sh(_l4Owx(2tM_P?+eqzcU}Y@2@dlh
z6^A3AAI^kXXe4dw*1ilNXWZ*ml)P>Dy^H^Ewis{~36Ha5hyhv(#~v~P?wAhmfA8w>
zXu#TMpN;nu8}EU*%Ufp*zgUH1XgZRma{>K8!;zDlydLS}3t!}cHD+#!>o1tOrL5Ez
zkC|J1kp}D^Gq=QQz|1YaR=~-~G#N9u#I*p-+!C85%-j;>3}$YLExdxcRb0`8w*qI3
z_giE~RM(V%Lxo%sZ|M}b+2C0m&|JjoE#2W;kH=5{=Q8~djP=OE;&1iWH?erS_<i|f
zfb2dtI_jvnOTd7?2zke@UaezS%MQ^DTZm+ZD1zH}mZQLOB!@hVVJlEz1&J$BU?qov
zOvGzOH3dLk!&NZACsyWu@tgMR6P?a$;a(OVAE}m8zLTRu*Hyj6)#NRRGq-z<3%+q>
zHXgKIcRd@3`ZsU?V|bqAyDK;u$V9SCF8xw4bAhEO0!9POKGq9qE~?99z<a}8zz?SU
z)>WRsX!|}3EFd|UropS1bU7Xn@Yy4LVvhi)D_C2kZKy8vfwOMfd7)R?ulvfdQphVz
zw^#eSPak9%{5QIia>wWL{lUd`yFS~|C$^)zr#J84uwd*s>^0JmERBmSaF~Y;jEoSf
z`=fAx1!se}5gBk)*HPgd4ZIv~)G-nOzX^WwQRanbRkQA2#(l|0vV6`}J#p!NDlx3$
zC>Eq5St@tv3Ab9Yfs?TW?)&0u@`lz+Er%ALZr^Dj?GB9eYrT?3Nts*D|3z4MJ*6<L
z{;TtGofcC|u>Z-uPPLE!MvLi<ByZ@n9gCj+ow_{j_NKbCOONBp?-)wk&X>H!*BP7s
z-(=u#yZM{n`Cii=kF;Y3o)|SP{5?*DAe*7!NL^h4r|RkoI7(Mnz{$S40?rWB6|l2W
zS3tU^u7Es2T>+s;T>-9xt7yOAudV=>0J|^s|J0}}VB-a;95?KwM0t|6A|ad#Ss|st
z4v=4uy|KSqgEwU=HAP)d$>SoY;r}I}Cd7TEL2GL91=>PJ)J0b!*7F|kJ#MW27%!>5
z&R7Ez!>*=gu5RwbM|gOS9AMOUSkEDLRv)W1?q<=whqaAu?=S58^mA}@av40-a*%m{
W=g~&Sy#{{iH;?DzKgt(Pr2YqMYqWI$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_bf16.cpp.EDC4F40BC781A6DC.idx b/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_bf16.cpp.EDC4F40BC781A6DC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b455b4e79c7ec886c8cd99480cff86e5dbcb671b
GIT binary patch
literal 15052
zcmeHOXINB6+di`#U}1M<L6@Z|U|FPCC=o;yL`4JzL?jBRxPX8jI!cpRNYE&~gII!6
zY+#o}MNnT01WOVlmZ-6z1`CP3ezUtioI&%&AD=(Iyq8`V2kx0O<;?TUQ|=kPC%CvA
zv8Sjp!4ncALgHuB6h&#`pSbwga6b%^dQwzkd28`>&wj&&i(kIYD3>U;oHXie6iKA9
zVN_?>_xJ8bnvY6OrOwYfc=LWR2Zh^ym;0ph`lcO@=YBWWHP$uWV^J8gbm*}^e%`e3
z%~y?9u}Lvw>Z(H99qVpC4@=x%`Rs{XpVqzh8Ml_r+S}{cc>fjWmUliHY+6<PW&6`}
zZS@BaRQxq-r{>62*`CX4);&6w=+W+Y>b}d)Sv`}7jc!dzHx!Fk&n`dbJ7#8zq@*}I
zDQIou<T)jN$#>!+=Ld~6?{!+)PHFB_%_PlAt?z9D#-WsqD~mSdeG?S<964WSOhxqq
z>4K76)vBSI8f`Pu)Y>@xf+zGccY5C6K6J*5RC+Iex6M?oHED-Me|h;;PuJ_8$v={2
zwmrha)GJRS>`*=GKF|J(w}0MzX`AOTPc<UkRcUFd(evhDfyZ;{wsD$tr>E7!GVZ|X
zKIR#VkD^f2%Bh!TDI2tZ>7sARqYs}O_UdSp^$VR1Dh9u&t#p{yeom6nWBr*}f61uZ
zXA8oWzOnW2ZPuPR;IZ~jtK6YMt=<=JPYAs<t-!{~CT`ls%s-=|iyW*>N^=ah8jQQO
zspRgiin(`2`hH!YRbaG{n`ip1(%uf8lx?%mH@Za)4gPUz?DG9y7bQ2B9dDlYV|?|z
z65*;nPd5bDFZWef?Q^ML-Ug@Y{y+TwYr1fClb!OJX}0l$3X|<RwlwVNAyyeas&7+X
zeDp5A@~4xWuWVjmGq83QFK8`4NqTaF(?0I6!%n#wjk9o%?w{=2bU`@sMR}2fb%pyI
z`|Wkf^y$n)Q_q^NT|DS`kgLDh&3PiJo1t5uH!A||23wo^H=F4NwAJq38xz>iZu;<F
ze^J}sqkq|?J7dj91n&2+Htu`xmt`l{XI#BCu>HFEtE%L!SuqA{7xXnPxXH8G`dXrV
zofok0c2S;R`p;<@t_P1?_#>k&!{)oA-{!2Czw69wTjSrt`y6mT*4jBa%m0MW^M_&4
zLmRL6*_>3i=~<}Dnw_sbJXVZ2+dj4N@%a7MlR_dYjlMmgw8${-<r4p~P20kZ!j%1{
zM769tA7^57HJ-ok@LT;o_gvoY)*saTd*@=?Nk!KWomzdf?3JaRfx(S$Y?21)pWL}p
z&A*Ymcgq#E6};JBny3uZ*9bp;bXI5F&*>@W)(q--OI*BTQkv5gtx(HLw%q!|Keq-p
zd8IDfxjG?#M!!*y5A{xMDP4Lo^1_BS3Hf_83RaIlZ9Un4P5PNf=ar(fQuar#9UtFe
z;G_AVY*nAY6s-$h7k-_%{J`tPaix!(OP@HGrj}cUrruCC9oh90lF-;jM-R9>z{WI6
z|LVPO&Nl3d?7UXi|8#Y%r$@5Hz0R@aPI1_y5hM3|{`qR>{rZz_sj80~cBcOH+oOXI
zB`O<LzDWN0*7C1z+-(e+b*-oRtmE-f7hXPfJ#}MLw7>1d2gfJ8Js$q{`1Di9pZMPG
zeLi9E#h05Bg0FnP(A|B}iGgn(NH07XeP7usM6b4KaZu9_{8`t|J5C56nj4oCA*GZl
zik_g{w`pQWttuywq8L}oRbBp|WR?j3Q*SKOc)P;$iGYg7)qa$py8Hpvf{?Z0o|nH@
z)8>TFrI<B{T7&3!zu*<kq1BE97nbZ*<;=o8I8jdO@&{D+eB<BXmlYl<pb~Jk31y-#
ze?YZ(T|(NP#7lLmoKRf-HKM-811NsMEBYh;V|P@!&vsSL?7y3$Vc+y<&&5p}1ymHS
zmQqr6`2$+_RZb7FpX}?y<4nWV8Y&vS<PXKyQ&fNR+o$P<C&zi5FLAX<MWiWzK(%?>
z=t22~f;9pv7FU1pal=Au6P&mF`J6{3<7$0MUnqaT*!Jp|qTWnj`&vLPz}1G7p}PD5
z)uT?=x$g~+;{3znx)!GKwKh5*5>RothYzIZ(Q@~tLuynpk4pKw@n@*bTl%K+;R9_>
zXxRU<xPe~2=X51AzEb6c;~xH2z+r)|8CxeG)KTU5;a5~uRBh!CSfcQy2c5?678wbs
zXj~mdMXAdl@R1Q`%)AeI9n@Ck1mo%t<R2YZo~wN17{lXC#npTjzPkJYoyRQQd?TuK
z;7L_Z2(GrH?9}BC=sb1V_Dq#aP0Lg{GyZPu^djq@;^LSXRnAOY?L~R1%OB8sWk9)h
zZ1DzfRZak|{y=jIepvm-+>;L*d7SCET2n>SK>mQKmGnq!Xts2TRpm%=^$==^y8Hp1
zuXdviO6+@IQ{~LT)xD|S>hcFvud#VpXtS~LkSZq#SAU=y>lRENSiPvGR6s@I>Me-c
zg4Eyrf*v-vTQ?lqaUh(>@x`wQR0Qhs2TX69N*vXBK(y%}l&e06dGL6((<1>DgM0X3
zd5+fq)^f}DH^Dz>&yN{}k9EJk8YZ9;aStDer`6`jn0s&Lr2j(?r_4vz*VVT;2&g37
z!w1WBp~-gJ`t`><RXP5+`ft~vvv}hX%k3@p0%|^f<pVjyhRlg89D-Nn!LUBLO)c9I
zpV4yF3IO49rrhr|%MYkP9J$>CTcdRL=-D-wCpq9^*}8xtfidN96h`#jV$*VTBP7gl
z^r7aLD&y@hJv@?q2e(yB7(PPHF6b}anX3{xVBw-m%g}iHu^1~hUaZZyb@OM3?pN1E
zE^CLpI>15lHHA&U>4wy!Tz+jSYhI2&C>Kd{(Y%r{lY9^UD@QQt?>-3>LNIiB+1x_Q
ziyS_~*W`;~<MA1Rra%nJ1oj@cRZI!&gW#-LesyJ|C$5lLAy^2_ex;+wC|mvq7b}c5
z*mLj-c%@`k6DC2+(LSXD@A-ziTQy&-vwGehy$gRrF_IP|20j6t2JTj7F`#uX=ZZ+*
zHOg%tT|jN~*(Z;A?0bPb&Op)(G#5IC>EUU;b242b>^^_}!J+is??&5xun?djxCfcH
zgsIVTbb97j9qV<MB!ExCtz~8fUNT;F#r5I+eqUf0C-W8HJ-|hA)VsMRKBGNLI|3Y3
zzem1#0*U$|tP|ZO^S-2@0q#xa&R}dlqd!*P6NVf(uQ<1qKeA55XI7$!l_(JwNVZ^+
z87PUxnJ6-o#91gZi^SO|DjOxU@8_VX91`cEs9X}~q3Ar6!oHu6BJ;`q3Q$x5`F<ga
zDkO0ciYg*;F^VoG`zt}wB_u9I(WN9VLosD&Av=BrimX6Mu)KU`J&IY67P7b!#a5zJ
z7FVIzDiYVA*cui?z<?!B>iv*GBScfB`wf)&Oa-DVkdPJ2dPJ`$aV4TFNnC~KDiYTq
zx`xG8U))X$nirCU$Lp@ZAZY<d5t6(cke1t6u)4I|^FWwF%QqYBJz8!{!Mf6N+XMDq
zv5nzd73L2@6<TgF!M7;(1(+mSz8zq9(DFb5;s^SjwFK8^YqnLOd9``udz)o=Zqa0k
z$BIASEil*__y$}obEdGo5Mtrt?z_fO@=J>K_3xti|DFXIv!S}t(z=&5Hfs4un$H@w
zu_ub|`n>2j3|n>~SEvi-1UovemM%Ct5D#7nnE{6da&I2czKP>G4R<OtNl@%j^M&8m
z;vV08)Y3u(`ma^9DL0R_Ci&cX@=nd*dwe{!Om62oC~NZ>V?AS|OVQ1P!%{j<V<a)Q
zV3DbwsZnI$y)*ik=iT8mrV>*N7Fp_98qJ@->C2)K;~w)FONpfgi|qC6jlj$E8GDJn
z1&f^ZoQ=Rk@)>7|vjvOX^xTXf&fqg{5;qGLdFy!_ZSpxYG<NBg8b0GK@wQ-5fL?$R
zI43?6APKNwQLtXH(IV09rD^MjZs9Y*l3)uK#puNtWsRBXTQ;z|gwMoCVk}rxi}Y%d
z(PGOTn@`?oFyk||NK%U|0Q>&Qc#YzP*NbTEVd#}e11v6$Jq(>e;!H$mk~j;|StQOz
zbT)}|5S>HfTtw%RI1kZzB+f^4K8XtuT|nYOL>H2{2+>6(E=F`QiAxY&LgG?Hmy);)
z(Pbo7SbXpxH1;qui_c<(#b>d?;<H#`@d5i+TWG{8C11fCSr#rso&mWYF7ECKW}Y0q
zWlcWs$!9^lbG@*T|5rf-@;MF>hc(^bSyTv#A!5b>qQd2XO@$~QUI@$@zAxpO?pEAh
z=CU9ooIP>EtigTrUw)R?CzhAA&Gprmmx&wkddosHh#Z3I`PD9(r%Z7%xu53v_TF(>
zcu@m8#!Mv5WW8-@d3@v3g{>QKAWpvDy)|$qI4PGACZGa*rZ=y*18iA7W5_df0J-xS
z6P}3!&?S6k2yci3a14CLj%Vipc_5#0;yF3M{P~P4&(#66kI#7Vyc{53<uiUfKL<!z
z_>7b%bpY0b&qVQ}9DqUL<LHaG1vx;j3Q-we6&a-i9{|(C%1tI_ipUXMU!?kGQ~zfJ
zU2$77d5+XUL4l{rmwq&+fPU}_%LeHhvm1i^RjoT=gp($bwl3sA80l*p63AG40D(-j
z%~>QCiHUY%eKFBaY%C_)iA}^^?I4B6bM0<+-#5cCiK9*#H2?J4lxB=Y#ySu}&`e)$
zUo{q)icEFj`)S6EYo^8`OOd4xgb*}i#kEpnku}$vK=vYg9SA6BW(0SH8jGAo&N>iK
z(99U_7&R8TiQIG`pr9FduDcqGyhYwR5Kz#}6z&u?7Wr^}2oxX+(1GxRW@d0_sIe$m
z6s!Z`1<eFo2iveH#5#mPbFAkODAYQXKw;Kl1PZqfCs2$iMhAivnu+Dcs<9}J8%Lmc
zZajhJbLSIi0e1m`61WKjO5`RID2baypk!_`fl|0B1gb@%TBHL363x^hZXHr%@ixTW
zM&j*=yPd>45O)WQ!5ZMLC?norhbj^^5H~PNA7u?dlkkf&3JnYoryH8zt{QUG#J26T
zYx=~lX@UFLAI+x?^TN(%DeAuz4FhSW#O>xstytvP<NtG-sxW+D2_eh@)q6j$2akjG
zmh>zNG61$J7zM13G6D^JmV!<L1Um=6)SYO3pLku8TK{#^)!)Cz3GzxLU5UaW7m=|s
zK^15&goNo!B`wvDR2E@56d-8<%c+7x#gmq~G!T!1>i|2rQnm5B<%ePy)7b4y(Dq<m
zi?*+}Ka0c@#6+FgSKL(xABIOEX+s5T0!xw+tyOEgdu$sQk1|NxlI>U9THBUD!?i~e
z$WeO?i^PM(L_4vym}n;+E+*QE9mQSkG8;?{f=7qQ;#HZ!0s8|Y!%~pRvrD|)P?kr|
zv*`(SR-Y}$C$=0}ZGO}H4k+Sd)mCgnP)+`O*++JdUz(zdPwQambq`c>jr<;YO}t<D
zSu~$mG<i?kjqO@zw&E3+ofJS6o^Q1$^u_)MUbvX_w7}D0+Fo|l$cj-sOFSR8Zqc_<
zM~)p%kZr;4bY}5yug*Pvb+ir}u0kX&WDQrBBUg+D-73gkaQa{l@fPl`8}I8%X{2)g
zL#?b`LYX0zkvh<y?ASm?>abWr>abWr>abWr>abWr>abWr>abWr>abWr>abWr>abWr
z>abWr>abWr>abWr>abWr>abWr>Hsb~GG*cx&!48q&rgIa(Ii+x*+K|2NSuj;nIz6a
z!YmSJBVjg)bC58H#JNbAOX55v%p-9=66TY*00|37T!@5)BrZb2A`%xPVKIqIkg$Zr
zrASyx;xZ&GBXI@NtUw;%L1pJBn(Ikii8L!oT!l2NNL+(7Ygk<F*IMfR=6noR<RT<3
zVig&ZX6zefX*UQ$5f?#D4!Jabi97{R9CJXDwN5?#yG08!&VI0o;N$^UDB=kiq$0>E
zpfPyKTRf>!BrRpnDQkrGMQOYsozJ%O6WdM*sjx*NDTMBX!luHG&5x#Yu3>qRkq4Y*
zD|NpRfx{&Y?Veys`HWaY+=Vv(k(M<5_wp?4MYYwmdk`<FiM!C2ggY74r>vwO<uA6j
zjylmty}|giR(xWufIBT8?A+~PmmjR8&4NF#_$OZR-8&0b%-+$)a&~Zm0<t_Hv+G?7
z6KkE?+y09ZRXD=N+D7Jl-zQ1ZqYqt1U0wnxQ=rkX%_8~b?rie`V~OE~iKTV8HOXxr
zc<W%*VOiGsS%E*X0w0XHusPbPZ3A}QIY^qr#uEpR9d4<1>FCMFoIDXvU{0R6K7cuS
z;`$Nh<S{h{K>_CE3EB;F^29YK%*hjH?3j}$I0MYd6ZeKNCr?}u!kj!ow#ztq*~LE?
z@#y%FBSI}-%a1xBxVeXwtb06v^wOn<xy4^LVTPg%Ny|_K8@5)Wph^_UhH+IWsEWij
zD5!?TC(?7w&Z#{+@!qL9+z-J6lZylQ06&-y+(W??q!3PBH}Ukpa&Z9`n#fSpAKWpp
z$asU1303#>>Zf0q{Fu{HCEK1Pb96p+#oT>8?C0QO^04jvuI#^<=6hMVnB0!ArEwhD
zT?9a80rdniC7u%5ZTCSg#jQPd#V0=Wy(wLOy?@T3w$fF&nB-HgUSI5#tGP{fPcIKi
z^Vnch7Fq|bM-d>Ss}DNzUQ9MC#qG&N^4jp?C0fUQeLtI$Pi#uA7yCFy+8O!(i>?e^
z|Kq7nW10rS8;L96+QX?N9O_`LB`<S!zuI}z?ZWNhB@Mei>(M9Hqg%p}eWs@7Yv6rL
zrY*NxHy)^5ILQ;sDjP|&Sx?oGakg{A{M#3BMG2CYum<72t>KTcf$NUoiYz3}Vvnfd
z(%w7E0RH2#iq&(s-rJOr+Iy!A&os(A;xab<@=?d~+~23A-e&8)f6*3rshK;1+5Xjo
zf6<g*{Zjr~<)q5B|Dss>wVnB+hNJ;XSXG6tLLa!INX3RQv)^L<1k2R*-<7%w;F?2M
z0r03@1#my0s{l?Wx(eV-qN@OobGi!P<fW?sc+;)|;Gen*fNSb1fS>_#CJJ(;t^!Dw
zz$~+$9TyuE*H5-jBVe;q)V!`gm6#K=ZOhK+>;A!+7eNI@1Y^TNQIjL4DACF)s$4Z5
zU!c`PQ>Zb)Yl7#bo~|yYJ=G`m(iLgz={t;cbQ<kEW~`-n@TdXy=Ei?ls?)oVVc&lJ
nB?HYY1`V;Ywy_;P!pP2G=ojO~Iwr&1Jr;7@e8x>g-qim9CB14Y

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_fp16.cpp.0D1473C6C2BC0BFC.idx b/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_fp16.cpp.0D1473C6C2BC0BFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a1a75f4b8b18a70ac31cbe561227bb5f4660aba3
GIT binary patch
literal 15060
zcmeHOdt6Of``_!>bxxgqbVBDGxm2j$h;*Z=kfa;QBq6t)O5${viXjs}*L1&63{ftl
zn~BN1ZqqcOB#cW#BDZ2Pl>2S`);Y)1UWa+dA3uNm=Kb*bwClT`z1LosXMMlVv)1P2
z=IngRmZHW5xW$G9E}lnI6s3)Sq83L6OEK8olcHkt>$7KsU9l0ad=jte8<Adr0ClF+
z?M{8^uvtHu9-lO*%D%!rEjE7n+O(^KCVgGfPpwg$YdLO3{;F(CM+-;UjKOvEyrfGz
zr@nanWM$y3&Kqmycr=_oI`{0Ovkzk@bllFg&kKu|)N73iGf7H5(5qDY=G_pliDspz
zAC#0w{7_l&Nbh^?QQstcuHCiu@s-#~t=AetUB1`rxpc(X`UlLlVZ&^Hwka)MbLi>B
z;bwZBHKP-@pBlcu^IU7wy_%&y|JphsT>bnX13L$}8%73KZs0tX9DVC*$Uok6uixsL
z-@FEgH>~?@p7S5MRZdO*F#;n$eqxH@mwTz(>23MZcb}oA!ROC;yTxjDpLS|_@aFl!
zDOMIjtv%HT%6B-p<vEn;Nz97xo`3!8&eFNAoqPLL6dDM4^Am0@O1HMTz40f}cwzOJ
z+GC3bYz_Fb*qo!w%~!MHU*GVBC_H~|pO9)(Av5>NLq6BD%|Asqe6`lFPQT69a)#R`
zI@g8dKHqw2N|B&fv&{^XnZq;B&pcJ@RF!BwBxJks=8oSAcA$dXpWR0dd8-z5vRhP|
z#>um_Uj~$H*ga{{o~&m7@vbIgHPSZd=cd_&*Cmy6`}hd<K2aZ6Gs|%83I6^NuNCX|
zCjK%p=EpS}9ZgTSWH92hzAeXAEptfg)t2EjvZ+Q`U@6(DulxI&l0q-7wVN+oy4X^s
zf81rz^tA3tJ{un$acg_9r=MWMH+Rl8x{T|eYiV+3Si+|I?|Zmv*Y)VTe#_J{bEo9d
zI}IBW+tQA@wSU(iE#4kJHM-WtZ$hW=T$cL_-?cgC)~2nEA0SDpJ5`&tFG*c|VA1`M
zzKy&|K1tTmjxx^rI|mmRWcC|AXJ@W3b;-TutzSBaM`b+f&{;9%@WtacNuGA8y=prR
z^G2*lPv|f#nZ0;|+t8oyj&<HKci#PIXPI-}<eN^ezt~=GEBhwp53A(n!kc%V|Lb^h
zz@D#St!yOMBa$!JdrW<8H_-F96~@NDY`*ts@b#Ain>!BeJK&X4-Td{@q=Rpc4K47S
zZY8YX6z~2a@J##ep1R?-Yp%Au6<k@eJ-*$!Zhox&b&vGhl@FtW4jRP^x^ua41|D<j
z);wyLR%H#Z8v1)o!VSm93+unGuNkMiuOsMnQ$<VjfuOq8FK!pge$_vdQ9Xl+5pY5~
z7F$jq?tSfsto3TZrFEW%!(Gg5rj)f6dEAQ|8T{(c#yEOZ&;7T9B_7|7qOxy}UwUZP
zxu~Z0=vPdEf9uu_v3HvuyIt*hLH*K!`Ht=P&kU*EA+D_u*S5dO_~Q40X0CS)x_<uM
zt;T5Okd7tW+zu{S`NOp7ub%o|pE|;AtJdu;`Hv@N&56(W)$D9r%+d|t-g@D3b?N5v
zSKaR|4|=ur%Dwm&55`Nd@NrSS_VR|ngw~_gE;>hY?>1MAX)@Kb9<SF_Y_~J!){gjF
z74f(B1_dnhe3H~|mwjJw(Qe;^Sw{*BkB0Pm`;GbiXZp`LxfX&;e#grFX0IuI(r!HU
z_NZT@;zFd98bzT-j>f@+E7e?%A4SPrDOXL!g_1K8{wE0hHrTWIpoWMOypWP@LDUvR
zzx&m7OYwt)M|ay^6i{)vhJI8(O~nPw;;k{O4#YN<aydcRyc$ufaSh5}aEruB;<;IK
zo(q=~h;Ol>Y%~=YF!%3O9=mzbo_YZlj?H0In5N<a<^gJJ0&Tza@#b@;W3!gJRxiax
z*?x)|G^FheedOs0K4%&>>!|B!D=uIjn)*CurpfSUT+V!K{-E`C7Z048R(85TK*eCQ
zF=ec&xWKzdpDCU2LvR@9Z-&SHtwYG}arGDZ)KXkSAF7W~ae*3UYb;9mv*Y0d5hrND
ze;6LW$v&5Q8_iz9<pkpz{&Eh&qw-VLs*lU~oT>O0MxD`AT%eQ7tFH<pcHF$g<w&vF
zj55<yT%Zl{ht?I9T{&UM<pf~!2c2BwG?+@Aa&aD)<B!dHl%A&I0#9$6ar55g@@)?V
z)M9M*rF=CN7ceLP7-${uJvCfF#bWaZlbiA5=HC~det3+}nSssP>e_u37if9rO#gdJ
zJ0sR|IrILi^6y$Lk5uhD9L(qVV6#A7psBdP)1|WWZpmIZLb#kc*gTn<tf{y_)5_GN
zsE)%rd;Vtj%J$_P((6;6&E?F+H8@ZXnu<$T4GpV*vQGWwgn)|1<_`w8vH>NJUQ)YW
zKrO-M4@T~A!Yj`mKjuvpP?3LC`B5#6kV{qh*c^q;A1uvdm3fPAym*2u55s0DCDl}1
zplQD_W1OU$Z!P9>X8o19rt)gt9iOX$zwx_M)|;Y}=W8z!P@%Yny@=Y2G~fM#m(*KT
zkGucp{55}5%XxR3u+>!!J-HlTTnks7J5q6hDY|-<b8XS66iq(044Xe#9xYWN+kBpQ
zGF;Bg|I~(?iw%x`H+_j8moppR@&P@(u^u(F@-l;a5rNGg_&{Xf{HRPb+?od?`ltR>
zSB@B257!>Ez%79uIUFS|zH2tEK=ac@rQbREz2@r7Fpf#A^5h_e3&3jQs2~j+|3|&&
zd=tw7_6yI=LoT*X7%P`_q}BMX(er{1X@)sptjnuihg*`0q^W38&I04~NzCgpg5m4(
z7G|1X;V?2rn-RmBXJi6xff!bTz`t!|<?N7`xGnO&!V2K1AYgUGdG~Ih7{NO(#|AkS
z!X#-A%pE&U&HNW!tYmH=b?ujoNm%zyZrC8F4ImQjms>4fw1!vUE@UHVHj+UXdN~$^
z`fO2a{OBNvhRi#CeA2&P;u;c>G!ZR?8bslzAG=#U<>5EVsRsA~8ieQRowjt%lxmAX
zYw(S7E`lD?3RIXlVAa91;rsgx3!?9B8#@xeL^(GwJ|IQdAa^u)x_{yJkd^|iWA8P}
z;G1i%5C0fwjRjBc*q{Uk9f8kxFP&zHr$}B2_^+}`K%i<IO%*j`WTKg(V9>OI@t$ga
zu4hK$n&i_3TBe*lKm#B9qzQ~PtME;#oq4}=P%D0=TqdD&jI57SA5Un!p=YpdH|ulj
za98Dw1nUD{hKH4O99n^(nPx-nbf!uV^;KhJ1&A&{LiQEg5xt$ng@`UBaS@`6NW2Tt
zyI5>t-L}eqQD7Wyl<I7c99%5WzFs@(y&13x=c?V$SpGA<pBy(3dt5B{W@sIVJT6up
zz%fSWms#H27RD&M35$nTFg<WTt*8sCrWKq}_D(jGUIz+EE2ITPK}j*70kmTM0+-PW
z-v{BP6}%5}Kr5UI9F=}Y*udrEcI-T6ZrRHcud8$<O=s7YlW%b*|Huq0+)260f$lk-
z%)00Oa^MKeIYOSWH+Ur`NnVfMK$7F*3;RuPxlw_im1`Wt3*$F+b}TG;J)MW0cBp7L
zxEDqiB5MMT6xkBUUNnY9;-O;VIkAP9cuqW0OgtyH7k51ezAE4-L#=lc`Olf93AKi;
zdpMrc@hr)S102yq(lqDcj!_+WoaNjFR5odswQq5D_y74Qsd_mT9&_ov`wo7dT)R_l
zSJ%gR-+lUSk$@t5PwHXZmTIpiVq^xo2KxE;Jnj1JcNvY5k%1|TOmt23GyS5U#u|!8
zGBOh*6H^wM>zeC>I5IMGBXd(0+3MQr1K%<-TO(Uj7CGuV>aQ+5(D=Q$$7M$5Xyj<h
zB6nSPeGpzo=5FL}$|5gaFMSX|M&@PYWy+#iy0i4}7(6QFXJ#ccvROv6Oj#758=!w6
z$^7KiS7lQeS%6W1DT^X>BlLl>8Cir;gei+kkZuXmKeEetScfgwh>?{bqY`8a*r!uA
zc{D$`QiqmpMD#|a1sX<UO`{V@ya~~pNSuV|BoZehI+?^Nh)yAKDxy<KoQCK$5~m|N
zox~Z4&LD9nqBBXHh3G62XCpeB#5su0A#pCEb4i?s=sXfD=^m7bmMQ6;#Y(zov6Ajt
ztfYIu(?g#a4nDb{2liu<z9E&u#i}YW=k(aUThjSYKTpIv-GzbtzfJ^%MOgKf$_Pva
zF)iFMZ=&1WVFS`%eV&0&JOlo1v-8fZt~iW+UY>nI>M`G^-@azy6)h|vn~-!9EAzmW
zI2a@)KvfVHv@C169qK2=m6AC`s08`@HRLCJRA(SF(NNumz{0O1>dlI~sE<gGKn5a1
z0vU^jut=;UCY}@b5fjge4aCHAVq<aFbC6)+zN+%U`}`0QJG!B7?YTE5w9G)qKmyuE
z%LecUXt2mc$3z0!M#~2C25Yd$T*q7j+D6MPcorHgvgBD3$X3Tz0!l~A?09w>EOOLw
zlz`IFvT?j|8Z2_xahJgC(Xt7=2^uW&((#gj($TUhyeS$i^5%IHXqL__3Fsazo6Vc8
z!J+`2014<GEeo&=uwqf5Wgvm(Th1p?kYx~o7FaGIP_Sh%fg*GwB%q$OERq+g!J;T$
z6oD4=7859%7fqlgyd?yR;l&UrmKRH)I9?oqmhzSoXc=!Afl82036g+v)3Q>;D@7VC
zE<?OB5|<-hIf?fn-aZyXD*w@xE4q~%SYtf0@iQ-eVqRP_!RgeH=_4j%0!%=I63_@x
zrlk^*L6K+HD9mm0Ei>39TNbsS3EL1qru+ZrBtgkyKx?3};B4=wdY}bZoaNF01E9=6
zft_H-2H!-cyvn2%=<Vy}m8*Zc-f|Epfg6!@BMOE<R-VWE7odd@PXJqCzbBJO)szy_
zL;Om4w#c?ij(}~~BpB80e5@Xi=Pv_EGuTbAjof3~)>TEWq_LWe7rC<<A@UMUW082g
zn6Qbx#9cPXny~}OH;~}<Nt05Z&xk3t_`DK7@k&f;^qn!lFpI&?qx4@Gapt|4F`0jV
zNZC<pCY~75d6K)I|0yi{j06{}syJ+$U<0|6m#Xjkc=E(QGCxn`C!WX*lk%rkvHwPR
zyh-MSO(48$C_lYnXIDH^903awwrTO2Cl{S^9}6N7pc~8{<Yuz!)^gVRu<GMK#}wmM
zsBXq(UwwV)+3T88%ov$Sn#nQ-kkv=Sz}!%^^oKo?Dui;HljrMfUzGVei<S90i<S90
zi<S90i<S90i<S90i<S90i<S90i<S90i<S90i<S90i<S90i<S90i<S90i<S90;C$cu
zT(3WWi@-yfg``>RPy##Sp(Ho@RJZ-MYOdK9za;U(2QyS4-sW6XnUaVN%EKf$+$r#P
zm~|}R>$muQxk#GJ9v47ljNeCYc&ToG?GoB%t>XJjJ}>xByx`$TshjGdDi&TXQBPRe
zj7+R0?n2-FzAA1;OMVh2dy$5yJ8_eSxC`x#xtqAN!9x1cJhHfc^vV93M+`pC>?fYt
zu2Vr}f)DFndB^~%Bb<(Zu#Vwu0F${~sM)9G2tq(O;=oUn(P#OB<_8<;!2zGA^b=3%
z{_it3%-h$<ZqC4oD5MFn6$d#z=2TL~RYkqhIWxAzp|$b9C>?_HIP8#Q+VSIAU%TW%
z<<oF6xz`1!aJZQC@xyQJE!xfVIzGGaCw5=hX~ABp8XLSHE5Ly$o*4NCCXmn3kjvkN
zJ2dXVAw&w2rm!K)(c?eY?R0MM$uKGh(~OD}Fe=F}qrzB>N_N4hxC5io21emQ&>}`9
z(q$BjjkS?q4>0gi%{w!k<l-*MS13FLTNRjg$ztW{FRbbsm%wUa4<GZ7YAbyIgI}VW
zR@F2$9a)$??I)aI=OJkx3W4~Ukrg8ULKF(gJ|ina{zW9-h5UE1_{8*1n}x>x*Wx#j
z!Q06-DT6O*JU{u~X(Jp^<GYm(0NsX9j6P^+{rU19%l4g!#Rl>-iqpJ__J3+5t?#M`
zqRfVy#-1~;U0H%%L`P3&5X2_%eG>k3O#X2aycW9)zn?vP)_~oGuV)WC7ua3+#`dsv
zg58C0W)Hh9*j@NG_HZ<d-G%R94}yu^h3{$)pG;tP;ZL>)DaP)?_qB&S3A+nlY7aBc
z$inzx_V8th++Fy4kv;V5!rHj6q=g5b;SQ5QIUM8uV-LZb&Aty>ujw=Z_lrE*Xr9&@
zZ*j6g{>c$J?bmBgcuYLoYLAP_<nj8=l@6)eW%6&P(vUQbO)VIF2I;>Yg@8A=tgOCM
z^Jd{r9eMH>vJu6yZl8!k6Hy$CH=)o?Bu+x1NhD52VaaGIdw&WFOCfP83QHw%8VXNC
z%h>zVQD{0@Uj_=xAn(saVVNY(LSb1X&PL(cWPLd(Jcq=&C_I<Mc_<<eeZ{`N0EHHy
zI2aiDKt*gvU$M9lMHZsvEG|NkMI_#ZB6qPE+!WK6>cWQoN}&E5Uu1vX<AjgTXKnq&
z+6u&hXMmI)<%dFW0CO|jdrYW}{>*>Sw@f#Ge<nG$N(*5^=LvYWK;*aGFSmV>bEM+4
zAAMp!f_*w1qmiFps(y%hr+)Y0!mm6$@uVapX)+s#wI^Qa*b&`!85?qtG>2v4KSt`E
ziuBuBjSWdin#7(}KT)@M;oh&T8`6014ClF4etT5np7S3Ka$d~zuB)^vzW)z$>Q~yS
zcCzvn5B^C_Uc_yg*t~S6LOO*Tkq0X!iAb18;!Q}niNr}rm_*`aBupl83KFJ}I28#~
zNt}j+X(UcZ!gLa6AYlfHGm$Wp#92s~MdEBE%qDRT66TON7YTDooQH&YBrZVO1!xj0
zjoXp-b`lpN?Lrb4A?+d(??T$USp52Y{u}kv>Noz2vT*o<6^Ci^bQ_}WH#S2aABi*U
zhB-j!D)ffWOsU8~*}S*-CfMDeU{|B70Jb%N{k#4{Q&$1R;#~z0)OQs??A}!Xp?y~Y
zME+d`5U6(*Kpfvy06~9O0Yvj%1rQl^6@XBKuu)x>s7U{)f%3r$ff1vqMO}Y(lTD(^
z6Do$f^yMrHq5MMvu+CG|mmyQS(Q4{ko(7*0^yscF)N-5b=IPPX)!C${rbn;dIwIXZ
z_M^r)jCCC6WG)^y+R%20!C#e1`t{cvFmRC3zXqEQHM6j^8u^8tzD?iZ)-Ga+@rVhN
OzT&uhkDrLVsQ(2%mb!5O

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_util.hpp.DD433B78FEA57450.idx b/.cache/clangd/index/test_batched_gemm_softmax_gemm_permute_util.hpp.DD433B78FEA57450.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1813740dea53131d2a1f602f50e5b8899964daae
GIT binary patch
literal 7610
zcmZ8l30PD|5}t0k=1mU}U|<*-VTMD*i_zyG9w3H*7&Y+-t7sGyBB*!(9uSEdFN`Rj
zQL=s}CPA|+UWh`L(*+3_NJLFEctm3qA|mQ4cqUu(W{UTM-)HswRK4oz?&_-QniLfg
zQJ|H{-i(i0k(`hgXC;%#-hlt9X(>r=R3=+$Ba^MzS2nr*a8sa5!{@(kY`Z;urZ!u5
zdrfFZcy++B^_k32cKQr2kGgkp%j@>HV@~~aqT!a>r>*YWZ#(zzJJ@pZpNkq}xrE%r
z*nQleyN4Z(Zx7F0HTtRRX)P<Rk9}vxz>w_^z6iVEl#!HYZPgL9d*zcavu4e@8+zc@
zx?!JXI4MIF{$c;gi+Ho9SLD7X<K0c07pG}MZa2=cJstejH+f&#zxK!+bvrfjT0~9M
z#AG9F+Ut}^hyUQ^F|a<jaO+r`$QOmLW`FZ@Y;9ph)^CZu`g`xFNg6wPU_sUk&nS;x
z|Le>wcyhD+{+XLKubO`vd41LPwi4SOtIBEOh7XMYDvYiP+u(X>;kWh2Iy`=9x?UX6
zyL8{Ohox)Xv>wYJ7S!~emmH&h?~ZlG#A7XaJL3GxPs?_Fv2NmD6T)H=j{1BZdp&Dd
z^e9*RB^TVCoK7wAy>6RfeK6nS)%At#4H}=mc1?>(kL)D}##XM+b3e6m<Ef0N5tF0N
zo=Mr98*r*1t^eu;XT~`0ecSKfEb!`!M@P@U@N`KzB@chTXhzS~sqf#)FQ-F0n>Lj_
zYz|Gkljj;a?&mE<Ur)R+w5ea(?LNQE{W$RItK814ixKO`rADXwZM|vJKkWHP&Ay89
z&vKIQB?o=;FOP~TUlmpztPGyxFriP}nUUd{3wEUx7qp%E<?5fc4)^EGm^OFyr^|O;
zA*tuQ$e`uxeG_drEV8$HE!%eXok&?~dh+aSTX2-N;idN;P6|*^a3bE(#?e+FGPyN4
zk4&a)T6`^`s3S^2eFw;RPtH>%5Sg8cDU;bB-S4}ku(d-$y@Bn)d8h<}>^C;;>bKP~
zDMUdXH8Or4H%}!HWUG|AX-f@u_Z5_>W&A_JJtQj2fFkO{iRm9EugX%;z97;<xEA7I
z8M;J{1-CAL^Ws4T4S-2zax)bILAE|ezbgM)VX1;DfgQ)ixd;T=zI~T2?3>Y(DX1Sz
z&`h{yu)#c_NN-hX{{)8*?kK1Oh_n-~ow!&A6!D+GXL~LidCZaa0+9+Tub?iL0YwJp
z=zg6xvS<?60U{3w_W-n*2NdC3^0^ZoXObPMKkR0tJdz0n*~3eYw@*rF@=;Jb(C~zC
zPe6ltK#{Pj)UmE*8-G_&ClGl|xW~lVGN8y?TOL1}cl6a%1?>q3?quU67l=%5KGC-h
zPZ<_`Zmpk!I>UC&=jOWz1iy|fyK*h$>HD05y21pn2=|J3Sq2mt)0}%CBk<m3&;TOO
z2=@%kG!G~ey?0zlr+@kP3aW#J_T+l11cK~;^oYNrdvC;O1$6-pj|lgOXe<MY#4PJw
zA>T3RHwERu%p@+!O(4jg(O9*x{Ak2=1y#W!U2R<D0>MSjy4;!`vA@T7*e3WoflF`}
z2!6eI25J6hyNyagZGj!j#X1WF*(nP?G<<qz-5v$?0_*)aKb1g`{n4HURRxZ9TQt<u
zyrnXp`LJM|dvlPwGh_b`aAoX3`CM+Ui$L(}?8Qz;gHDBCQBa8Dyt9onClLHP`>^xn
zm|bn>HMAcpSNbT2B0DE_g2%`I&Vl2xhn;&$xToOF<^d<jS$=)n;<SZB>}e2K@2~d{
z5D2n&p6JZ-T$S(|4j9Z^#4Yj?2(ovL<lScO9<G7Q1{1Uru9YY(1B&c^uVliQ6Y1aC
z)1fedL2hsq2(mw`o;)u9oePs~s2eQs4zayM!YzYLZbeM1+P~(G-yHj@0tM|2A}Sk|
zTp+l}{D%=6drwL5Qcwj<;LG`H1cDQs7=LNs%K~{eTr^<oIGu|?kbU~(g2rtFLeD9v
z9@xFOUMhhg`#;`Unw~GegK*&h>z~cdb`c1&%TBIJcTPVu2tpRH=WugW0zr0pAL~np
zbcbF#(mtSkuzawSK#*OXo_+W8VLKp_aIkwFgzF$~mH|c19{k7q!}Hc00@nl$-ki6y
zK#+azQrgPMmFE{iL<J3<gzJO@HV-KBYuYcfZ_K3Ez<WW17w6?HkS_MkAFnxZgWVzE
z!3tuy7?nVfeN$g`(JE~13b=c~j_2Yv0_kGs+4Ng*{P$vr$H1P>O?MCovTv=fzFYHp
z-XVyCz%C&20=WC;0Yz@NcB(7aIuIM$9}d^u)}0rKOm5S|boReL$m#LklA=@gbRbwC
zC=YZH2!4HU(2oa_pWVHspgzFXa$1!@kbQqhTIn=>vbTcDf$hWjXas`n2d<Ac99}+U
zGF%TZFOf@B2?W_sE~Ok;T;CcEXAfWNIlYTOkp0wV(DSj{<2^Lg!(=J^;}t#%Bf8~_
zC#eVGr`{@qeTSoIBU~HAUh{w&Ue33@HSR{nbQ|gpA5XQN8ZHp@`;>$aQfE1XUQ13U
zoH-W@a{*~L_S$~t`xBJ$$BFlG5`=e^@nzJzj0RzFRt9E$HgC7#Dq*~%y`u}N!(2`|
zA&28ZFn(Cnuqo)dn9)JVl+Hz=ps1iOA<XL_WJ>C|EXJ=PYz=WkyD-0lm^r!Q-DmtJ
z$~I9qG}v_o_g!YkM^Rv~5GpMT#^m3Fv!s0S`wpV%Af9-i82<+i_=66_G_1ywfD=jq
z(Y`%|?IB9EkMVm6+bhMngyl-{e!})k@#lnnF2!FE_JtJZ5tb*#2MIeU#fJzxB*jMw
zJ1Sx<RbaiQIz=>lIa}^38eK<>bz~G?H7sD{oTY3LH|Yt{JR!bVHZVv8ny17Ua|DC9
zq3I;P7*ZG{Y0Wd@i}{=J%`~8y4#7=jd<zX|k>Yza;GPt>(tuVGW7!0&G1p9@Ve2Vd
zPhIi!8Go4?FVj&t5su*g=1K}b#1aZLnhPgnFn%3n>!ccgrpBK|jaXQLMpIRVP95}B
zh}{PhpXb1Znm%HB*iB1pgN2e2EhG%x4>u6p*j#Oi({840vp6kQUtpiP^b!R(P_}`p
z#1&tm#w#=wSBf>64QH;!EdI~<SZ%BqRwIm$)5dwBk1&3=cD5I0E5^rb<GnC;Gd@9^
z;DtpFR&-##shktN>SMw_CS87owt<+rtP_pDNYxkV05lf;0|d=gohS%9nImq;_@Et&
zT>EElgYAtO5fk~VCGmh?krB-#5*>HK$Ok)S*E@nw@YhJdH8K!AyZf!4GQp{L%k~l0
zPK@`L`#WGE14!lHQ$$5XSwx&gCB;NpEX5^6St7-yL|H1u6+~Gf#g#-^DaBPpStZ5Q
zL|OeeOqJEt8F!2EXQ}e66xUK^trXW$Wt|k)Q)RsrH&A7R6rZQc^HSVMm5n0C(i%?3
zR9B;`Gu{|w?26yfj~BcfqDU$xR(!Z!xHHCnth?b`bI~oH>?f3cLOsw&u=)l;Q|T?b
z)M~<36L$<%ScSv2GnL?&#%BHymcM_eqNCgFvsUa|6kAq$8<K$Y8_sGhgwc)RQke>M
zajtZhuELoZ|2;8&PeSq3<HkkJ8ld0#vfC63#^hX5hWd{KS6l42I9mMrHSv2*qH%e6
z<6z2IDE`GMm$7A%DQk(bR-AIlijW^Y&RuxzX2tvW_3wxNXvO!j>mwIYfL(wD^|k9O
zLH+FdNl<^g{t`66Zh!<85nT}(jH|Wci;1pSic5&DgzzH2l;}#OxPs^^q_~pkDy6uJ
z=&Gdps)?@pZ<y+;>0tD6EAVmMSt+iix?0MM{5q<ulj3@+tC!*is%wzq^Hg_Us;`mi
z8byo`FmON9>r3?Uu&A&uA79#Rn^4m`=t?&iSoU^!ec!3Zcfreq;?3fe_-G^Yzf~aX
zeG4XLZhd^0F}|n1rvWb)<302q1}v2ruhnY}SeY`um%f()Um1+|)O#AR@?yM~-pham
zI^%VEodKUKjMwY+1}s4t@2&SX;A4RCK6)Pm7PpM|)%zOo6~uTyy`KTMWaGoPJc57B
z*1+<d-f(im#Io78@<7&qQ&!tBK6HHOB)ss9A2oiI1Z6m8Nc{Ul#}6fFBX!&;&0ty7
zF-yews)ZvlJ!!=&^AY=~E5&Rc+hbi@&R;_v8DB)yMZ^;wlz{`Qi>0`Ps7s`{l&DLk
zxPquFq_~o(E2X%KsH>#7ny9P)hN-%mdZL4xg0%Xq6xUL9trXW$b)6K~Q+2%*H&Ath
z6rZQ+^HSVM)r}&?*arFD^uou$$l$r4Z6i8-NiZmLv=4|5>mCM$l(wDdu(DzhrnQfV
z4&ye1vP1ir=&%mKW&t?0&?K<TlYK06^B?)S4&Wn3jd2V*!`2Dgo=sbRY*4oggXJQe
zthtdOCWUOuvc)1EV;n?#Q&$0R<#w-_C-07Xu)CXDjJ=@N)MOAHriQ3%$N=<=u6_e&
z=`x73e@)rfREaU0@p+WxN%29-4odN1$_`6$K4tk*e1x(iQd~e;ffS#l?4%SIQdTI%
zC6twj7`q^_K2r+>cZ2bx9Y?$2mgg?C-ucjJYayJ3DG#F`;2L2k&HWK%U}FRrrmhJ7
z`5VV!pRQHRQFq(o{1qkE?<Rj40oiV*(@YIMiWuyqQ!CNn)?y$5rKVmAT6)BK%;iU8
zl7hiDsq-9Z2&wDEmFv3Ag;5NaYwEb562^~+8ZjA{dm`~h`(~?v_-@~xI=SnUxPmKC
zj`HqN?&I-W456UX)TO~cGd?&fxGTzI0Su|y)T<HQXa!r*6?Yhaju_90PbO^QfM!z%
zNBn3dTiNw^E&FT4p!<f;9(6Mo*8;m`YVV*`<@}_0>$kpag*B*rRsHdx@PP_5nY%rB
zw2a?E*%s=K?u(Hg1Wi33RLl5~sF1DzjW0D2GPit0ZCfeZDqcm*Fd%5^|A^X#M-A`N
zhV?i+98FyyaTn%l=K5khXMCb2(HHwBjGw2O=Zmj@#?RNx_r>Q0<C{o86B&Xxm+@Cg
zz*Q-}P6Dn=@eLAiL&W%a0h>)-C()ctmg#}f0-tDb+@{`>D3qj0^2DmN`#S<Hf+I7x
zqVTidg7>U(s*R5BcEq?{Aa80<;Vr|?6c}h~OQ96pXz$T|&@JwaZE6YLTVB{rB^C$p
zt*K`vemgX3XqPXy+U4iaVaK<D54rlf24eM$X&=lnHMH<sJQGl9YGmP>7{7|G>T+T%
zCqT&5&cbgQKQd}$*R{mH7NiMjNrYJl_S)RV61{CLWoyOU&-kCH@h9;$i>83VrWP0O
z-=A@>$9Zrh!wz2SM8A&o!9!*IdZJ%1#T$rzgA{Kh`i)Y&o$9yKK6rN+zk}*`Nbydp
z-zmkrsD77-Whn^@;`Cq^1|Avi3y!kH*fH5hCRV6ZQ|6>j21&}_s<HXsf#*x(ZH5>F
c4TA>D47e<rEKvM8So}Fi{Am<_4iSI;A4U*4+W-In

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_batched_gemm_softmax_gemm_util.hpp.C46CFF8584087968.idx b/.cache/clangd/index/test_batched_gemm_softmax_gemm_util.hpp.C46CFF8584087968.idx
new file mode 100755
index 0000000000000000000000000000000000000000..69bc64742ced2f6e28755a8e021a8d509905aff2
GIT binary patch
literal 4566
zcmZ8j30M?Y67Fs}obCrQFvAQ3gTpbPk6}dAA&5~7UW17c1QgIwF2M*MGb$K22Wq?_
zDyZ?;e2Q_CETD*zD66PT(5MN1nq)<D>W9We!x}L$C)w(jnx2J^ucoX2s#o>uz5c7x
ztdk~nC>bWHz*=5fxHQ*|VHi*u$5Q+JPL^T5ffwa_o6}}5_&i4SS;~dBpF4+k>?0%i
zB`sBrqoOPKzy8eDMJ39t8B?y!Z#weV*y|mU`735>XWseB(S1I?>H4>2<EkG;efrZE
z6<4p`9lUMijkCeOEj!=h+nhYTROlS2`{-vwseWcwn_<!33jb3vLz{|{@?QSxmGkEw
zZz~8MZ|Tk3=P@;<s`B)uXv+uZlD$$=4|=y>e-J)qM?u@{q8&lMcCATS{zq40O4a>>
zl#aT2Cq^!=AK%uU+n_$yabrYt#akJTs~>+?x3i-{^Q-A{dz-@^6<mEKa`=KdUyrT&
zU^?mDZ+z~0lKoQe$@!MTPhuABO-e~^oE(sOW5?g`6a_qETTyax<Ed`~SI#e4d1q_f
znLDO+iMzrx#$Wlk-T0aPleL<6h7_=w8LOh3D{36JT(6US)0XjvPqe*XJTZIzd*eM1
zH7?h7{rlCK74OeI{@%rr?u(mmxmosUv>%t*%D>aumX@_$IzDyHh`Pb_+uokc*$GEC
zwbwo#_M-7muSXaD5%bJzy@O6ns_E%m{@d!70gZyWxOltm=$>1eJFn=!zM^?A?ybx3
z9|-G?ve9Kp>m!&i!^66j8{%#jr=A*I_~Xs3pJ+c^U4K}6#AD*!U2IO8+AU}GwMz|i
zPqXE!v)zsrrFmb7hy6ps$?qSeQGp9$7)m^drz9AKI|m>`;=9s&lBZ1&IPU<4X7Cxl
zl3)gkAjA0Ael{=Yr>#dlI4yXmdZxxpf>F4^)nS;Riy^~X{(O0y4;Nx&XrdyKNCLeh
z^42;UJPsAX;s6KofxePJm`?eoc+>Z*o)b79h@H>pt0jSO>_6-N7G`hWDR2t#*6})}
zBoH2x6xMv#5or~;V1R@9V6`Ni@UckS>Dxiq1Wp5sFW?s#B!Te6%7;I`aOjVj0;hwG
zd3$&(B*7@eZ81!CD}TPb)G|)s!oa(TFA^ky-Z^Wdd%F0%Y2KU}Vw)9aA4xdj<2AdA
z+Z2-pjtB2NK2IYFglBEvQh#{Qu}cEy3Euv^zn>(W@RhcOt%oLkDRBOfn$kn5kOZbS
zC;4|>XLM}~c*EjZe3qXi&^vGX;q`0gZ0m-q0i4dKD<$EC_uY&cy>`?Pfg=EC^Vup%
zAZ$-+8D^X2_sECS1N=6tc$-!A{Wwh~tshi)vo22H20`{(UaOJ>!sTn~{a)MO{i(pI
z0M_$*r6dqu<NYSjT<iK>;CulN-~;?5f$-Wr57*s`3*0DhN{FAyXZlJ4;SHMJ?&-@8
zrg?J_;2o`q9wZ4TTzS>cOI4N(-T(*jLB5hexbo$bxz?w;(jhi1ZsTo!l0f(kn^`-|
z(HSOiYJg|+vvra{c-Q^g8-_U&>!5P5xQ5rLB!Tdo6}Kbuz1x}u&H$(8<Kd%_1fE*$
znuE=!F5Jr#H~|))#m`bp0=<u%s4d-5k~mo4ya3MSbJdbS_{7TZCxq(<g0~;Q1$==@
z5(u}w^u__d{kLB5;KCuc%2O302}a>DK)lTDfmM&^se|+3NWgmrKSL=A^gdVU*&97I
zI|&XH;2b_jB?*K-^-|8Med+KyZ!Q}4Z3s0`Ng(_~o3QGmkj_ql8w}oQe41Jk2=`R2
zn{xS5dab|>1b8k#S0xFAfAWa^HCf*{K*t4&nZh4cgexpqwO?vmYJNM|ljg&PLhKkt
zjISin+g>=&QO<)?AIGu$6K<0hlDUoJB~C*wBXl8Is2bD94_-0rcfYWG_~$I4#G9ya
zDKv&9Mb_(Ek<p!Qe{1!Y`m<mi8yY(W^Ftm4Ymo%;Jdpn&0g?RhHKB3VxaZLqSrCjw
z3dF#KzDQowV3x>*U?dWuypqLgNyU|r5y42LL)mDwb#%JZsBhn3B$A>W?p;><E^9^t
zKz?-RpYS4<lxFF(3`jwQX6v&Jd3}^ZXpTO|fVYWIo8D%?Pwu(6ZHxS)H*bNurHx6O
zjF%>Tns4OYtz+CdLJbPTAiQdzsSSasWj-uP9V^T~&bra32R`kG8*x%JFHy6+k<BdG
z%z9zv2(4sErK?`Wk}6ldnkCh)dJRizT=nfN+3u?EX31_>eGg0axaxaZve#8Vz>)*9
zj;#agFZK=jkP=9Ob8k6)4FrtD?ji4O1z8b*JHs{tMq&?<SBka9reT$4^)(YP5<3ar
zV?s038Cq=qwvO@jZ&nLG_p2}-^8jAzFme`4$PzVXL1-Oosbk0C#mRlvI>#JZ{d>Q)
zuq{CzVo#FSDkY_UxE9_d2q<<Y*~m`p{%AyKJ!`3F6EWbD^0AkLE?)i9%Z(bM3?{ra
zpqUNBWG%!T;-ZC`LtV5mbC`=3ZVq?RjAo;Y7GaKX(IU-}F4_?D5EpHzd8ms<&D2GU
zGDkTz+y+p`lXi={w)gc}sF>Jf<>MV`9XTD3*Z$X@`uo}M41@2gB%&k*YqacJ&%$$&
zah?4RVfoYvA4Scyq=L1?vVFG!TQ)=z`?VY?-Wu;D7i`>MBzA4tXq0u7^R~TaY<m8s
zs)u(Jg!=3KO?X2I)#x=QY@mc{^;#2dkI(>pfC*P8RHxUOknIQ!)CZc7MF|bk2bnhX
z^>{+{dc6r#Cp1_eY(fIWDFSj8M+mGap^L~OC-JX)Ffl0p@@5?r!`sK(9|;df4Dc1_
z3wfUjgA<Hcc1!|{#3@5Iddm8glh~2_;5Li1hMdJxvef1CImB8H$)8Wvv@O46l^u@l
zcXC?|*^g46?cUz+HI6Khhd8guA(oTnL3kN<Ep)Gb;9GSJHY)NbQUD$Y_=>}fyoV9i
z5zhOCQ;i${WSVidfpm4c7QbmiGu4?|EIOex)HAd=bKu+qQN*D~&Ttu7=In0o_>A7M
zTFBS-+g*M6zuZ$xYbQbCMZQHkB>aYt$6g5LTSNL;A1VIoznk1>LI2l%lNBT&j#Y93
z<)qx1K=aX>jk!&ot00ofq8g30z_AN_#hFX?T}f6t`RLSNV`A?Q-~D61xUD183fw2w
z-h(B5j6Ts=NL(5|cX3X`!@zPuBym2IBP}M2o%D#U6)u!GrO8GP;&7%zXaj3$km(T_
z5CV#`8^*sCzq7*k!_3TnS>H~+aX0>r)?@u%-y2x}IybseY~M<wG6tIpwsC=%W<EKg
zxw^SVocsvQ)#Vy-z$G+KmuJNOMrgh+--r_sp#{1EBNm;|lPvs?LV_gp6bt_@dMis?
zUG>u}J?*NWVd)uJ$9$l~;utC?TSm&9)x&v|Rs0`A;&%^46o*hOqP?&rHx!N#uMdL&
z{xYTQ0}_@zjB0S$^Bt4HlA}9&UjNUg2MZ>;N0}mr3}qtm1~AM}`EiK+Xp$eJ<j4O3
D&1CB`

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_bf8.cpp.16F0A495F47E95CC.idx b/.cache/clangd/index/test_bf8.cpp.16F0A495F47E95CC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0682e3b1e6370ae893ee82bbea722e8b05e262cc
GIT binary patch
literal 13388
zcmeI2cT`kI`^WEHE+VkFMd|`d@9esOG;t}46j?>eQlivd1zB)m4MMC4qJn}7q8Nfb
zR%{U)zQlmhV8N&)BJyf1fPzF~i9P=AWi!k`LeBX`|9kNq_wc#TEz_Q7zVpn41WcMV
z(FwyQhy$`xC3E8m48vIDfBD>u<O50=wo(nlvMQ>>Z>8OGW&ZR*Gx2-jCZW<R<Mi&I
zo6quFG_y6W)~y>Jx+By7=5LD@E<e?DxI!cN_}c_|yKF>9hQBYyT5{PTwH(Voz45`#
zXXhQYzuWcbPJ=XCKHp_>TF3p1IrlryR303r-@fnJOQrK|-=9hMKO89-cK+%0$G^Xw
z6}{kt^JG!%==O%ov-x@@rUA8fqK3|2+tdH(c_BD?((+2mnGKJQDr=|5Pu9~ZQf+Ut
zIPh4MP)EN$ZrAinsZmkqL+necCz`%E=yx``>gbB~cTZ_Q&3kp-v^sF@-CtdFZH|^-
zW53j7bg(@5!qr*Zx_3HNimnY`pEGk$6L;eZ?fMUs?ltOMc<^1;@BRhZ7wpRH2<z=0
zu6rD#_0_GTE@W+4c+FKJ>}r2<d*SJ&YobOv^>~>pJI4Gr;z)kc0+zI)l-4TBj#;us
zZ{O#Brs8Gm9H$;yGu=^Kd2hN(%cfIU*rMe3Q{7(&xqT5HSRfmf@~i!+QQJCqZwk>!
zmDN5jqU-!H+PpmFO9#VNrN-8#GpokdT>E)Mw5&KuVpd8|4w>KaE!$qIYh$5lSz6LC
z=3sq=`_Cb6ej{^&0_=_G-ux;n*^<!i+izx=kNbA(x;+ub##widg+FBMZjJ1m+qG{a
z^KG?!4WB)&!|Rb*+KK?xz`$voW~JIm6Fyh7N=_?{)_sxHy6wOr|GFg?)qAw->aS)!
zYTce+>)hS4<|gmPxik%_-lpKB&%BG>$^!hm>`I1LC!Tzq(=x5m`{|_>$6hZ?WMpPV
zi8h7(<bHfzRs8+>^=j`s1L@tS7aJDOPIhg$v^Rg}zWlt^8n>Ig)r)QXL|Qr9csrcm
zUR`-^QbQJNib;VibaTrplPwq5R%tuF(RkKbnO?wfVTGi=-n2z9$6qhDB-7fyDzMFR
z7R&1R_M016&BJYCfAl^ZC_g&(n)6l}`^JcRxjdV7J=`JNqgMB;)eleZN}f`u`Jvz!
zH}~$gpod?da(%hOQ})CC)|TlP<Nipm+kB-p`K-pWxJvWOd5hOS7;~Cm;#9-=ZErzB
zNb;((4L$L51@48Rl6yYk6@Jglg<oIR=(($2Q))I!Z2|p#wo*;)sBL#no4<GcA}Cg4
z)S3EM&t865_|!Pl;B|Y0pG@c59W~1f?kKN4wXpT@m%mSaax=Q|kw(Vb1M8%WXpbYT
zD)#8BQ{QfTbv9m8u=qy$yT#U3%D3)JOYck`uNQm1IwJb@e`EZ%FMaT=BTGqgXY-w{
zizYu-zOJXS^ts`GJh*T0UEy_W<k_hS4`~=SVpH>f?s;u}qe6?uFtH2fLiGM%ly{T=
zY|p$KO*`~v9)}jIxPXaO)KnOM{*BowUp?Nri1kF97X8`ZJZ*m<D5A4<bDcJA`V34=
z$LK`w4|w?)!QxlXN5+?NXcF*pRW(&6c)74eSZ_J~!4nmlh%DLyYyr{x1D=k#63(9X
z$M<`*X<^}*I0}m*dVhd9VU?Lf+}wyS1}R#KTcBOB)2vJmEe?EpxY}?X@a-I}11~bp
z)_V?8@;v#6ANS`>E!Cz)Mqy$tOpECK0pBiIXmuxc`>4=CidOnf<q1KJIbEA3iujv~
zm0dl|yQPs<Jjlzp*r-_Y_bof2LW?Au<uBjfVX#bwXPP;x(592-zqI+@w%N<Z?)>H2
zAVoW-?@H{A3)Sb)W`Sl=Q&ZCh&GJ)eqS?KPg6csU^+w2T`6IiP0fSVm+3=Q6n!&|F
z6<Qcsw7>LN+q^{0rn@tz4)XMKT_e@`dKVlAX_gOaTVjg2+?qjpEJHF=KF<!rR9zFF
z>BW4RrK}i=Jt564v2xZKMgO%kl;1{`aR`MEbfxq5r0c$HRE#SNJM$<#$&-1CLHOSm
z*{ENx#x|+npQj>k+9#pmz4ZY#r1c|>MLk_7QZjzQX}#SAIuN)Xjizw9KO0%`abRSA
zRl1?3mP5noVsw0J`fN)RRVp`sqH1UdCcK<-d%{Mtgj5IsALo42A}D%wdO1NC$Jxel
zL1pRU_^|j0NV&LeF3#nCIYXyw=VU2eoQH?!;gNt1J<PQiPuO>rY*hxq(C9MeaL=FX
z>>$KUf{74|U?IdN*nMPDVZDFmT%8EAaMYLqZ45f(ADbS;FI@F+<JN%sf_Oz<rW^;j
z5~vCJB`OMmxJ2tm+|CBqsj|l-$!F1<cuA>FkvcUe$o^wmF>OKZLFgmPg@!!uISFQb
ze()ReC6r@>k>Xq5@<SHiUQ3lO=GbtY-x;jhuq*qU8FaCsnc)~nMqDFzNTz0{h|Q8|
zX#j1uJlk=QxI8XGd>$VmC!SLu33YEey<4i{Lk<V3CqYmkYex1YB&sJNQ9TJrxU=V?
zi85X@ppE((RMB+cZiS}-<CNWFqpLQISoVoVe8`eH^mk9Y3dy3NBVP1W8Z1yxK4JB5
z2P+}w5F7?*TS9C=7$9Us7$Iarm>^_9SRiCe*doLwxCrqHK0;1}6G9$@2cw<s?|r2^
z#ETFM2mwMqgbzX!iHQjL6aEMZ2_ZtkL@+|3L?}WcLWEEZ5yJp+fe^<LaR`aYCWW?H
z#4LnTh!limgbX1$Ax9{a$V4cc$VMm!CvtEG@K!>cixashFUN`HC>P*F0m?-<QG{{{
zPL!Zrh7)Bdm*Ye^$`v?K0XZUn)>@s$u!BsxIL0u>8dRMOp3HayND@7XIiw_J5@M5a
zWERjS=g1Mt<YXe0&B;b62RF>YtwBRj0hyVH8$e!x>#e}eAs6DzLeyS@b4qXvXfMM#
zWhj^9oN|;aa83o}sLBREm66xGsW8E2j0Pc(3=?c7LM%25AvT-cM>9&hiqkH|w$tch
z(=gKr;7oL}Z@6zH7~C<9-uo*{)PEoHRK)*=Je5))g!upGsbD2R&IPEzh;w7GcW(Op
z5qTL^1Mxutv0Rgqw?hMOXyA!IZi)bF)Bz%7rX$@u?}%4DsTo;TEG`J3V982qMsz74
zElIGPcsFn!DGt%48X0_fjB(&uGWhbC2(cVk0%&79vArN!Ff9<N=+E>w05O*?PGP1X
z^HUMdEQ0ff^svO&*<zTj8L3i94FqHyy;lQ4qBIaBN&`WnG!P_813{uR5F|<iL83Ge
zBuWE8qBIaBN&`WnG!P_813{uR5F{$HLZUPfBuWE8qBIaBN&`WnG!P_813{uR5F|<i
zL83GeBuWE8qBIaBg$9DG&_IwC8VIsN13^}3AjpaV4p|XdAuBWxWQ7I-oTiH>#$Bmc
z`iVsQoEfuMKFlAnpHvnGb_P5s3=G2z5t$&c6Y$`W^$hpy6AiMt4OcJj+*33l(tqBe
zA~|R2(mWK>heG;ChV-n_b(wYgXX{AqniQ6VBuNVK@IsgrnJ=%V&nZ~!`H36@q$>wV
zmE3cgLBg?5^9Rf?AbUw(MzajddwdPtuB_sa(+--G0Ljb&KC{{^O6%B~nY+n~T6$Uf
zLthN=4hR6=xBB6c^ic1vV)B*f+*CAW&8w(+2OWm|ZOGp~(%*`jqoYjB%jn8<#n!@~
zGY`FYWP@pl-Y>yQ?^XHGMy2Zjty!e9<Hq>eNoGTN9FTbUU(e&LHHv;79Q5n@p<p!>
ztUfYWfsH?MDn<7QO4~{rOpQlAAy;r7HXbfu@}!GB%{=*#e7HX2Ao-g4BDMf#fFZOA
zc|s3J!MtFELV2MGiFl$u0{e~wWWfgY?>y<(61NN}bj6~j-jXqHPyY3`&R5emG)m|D
z^M{k0OPX=UAl#7o(ed$tkdnqHAtbeuj)k^dTr(FphIz_7T#$!PhFpjX3Q;b`t&4GI
zVBaiwxOUEB*A7ygFsvATTQgwOh1`~-)Df5&6gz~#xb@in&wf1lB9iQ~7>5`EEK<CA
zJRg#TD{+UEWS)fBq)aKYl$P;i<KUfgo*bb}UM51>yzD*#J5*#@sT36~E1Iy{&2%rz
zRZr9%3SdJ4>>~phI6W|6EwF#yqgA`EI3#?c)y)1ks@gf2r~QCHR5huqsnc!JTyjH&
zqs=jaD-J`ZA(F5%;uu*#n<>*2vDxa|T0$FFpNkM*pO27}z7s+o+8!n_pn5Vr4Iue&
zd@LaOGJO%7P+w>XZNd7%2!-m0A|%om^-<MjLBlAuTH_&y9dg)@blB=|N(<tTG=l^j
zm@oQIX8li`D2}^;Qx^R<;lc4BG6jvU=Z{yOIKImNMI`C4=)`mU$CSXBXzPtb<9KKs
zf8;nmF=>6)0k4ivb>8WmpX4l)E#iNo3G2d)V6mQ*v3$VzzjP?~Q0YDKo&mFE)6Np-
zvbH(_sXN(L>@i^D1)QxT??z9gfHQUEWpvE|{F79E)Se|&e&`P7&1d899=un)b|{;H
zoD%upK9vFvF8<?Lwzj~Qhi)5dkN)em9XQH4V9>U;Ns<nWR(>EyK6;869J?gTgyy6B
zXQzRTB5y+7qrbCs4*NP(>qX!!GAS+fl={I1RWa^WjQfL?A8<MHUFdw(fA#{$KmYNA
z;QZ)6zUgU3g;K>tcLupw%)uRUa8FPU>ST_29_|2n1<qZ8kAqx@n-`+?Vw_ow8$x>t
z&MUz^AeZ61GL*}4UOCDYIIjXSI8sY?6WZ{v7!{A@tM)d1qUiX#Th)^lb~vA0=o@Gm
zm_q$+INoqFB-8Px2w7TNx<H#iOJE8H0GSpMcm_f875E|)U=!d9Y~XpabI|Jbuk5>~
ziuc1Bhu!-`l?BI;$tFdI7M3BA#)p$p8VIt!gZF(Gz(L2p3n16fcLAukz6)Ryz~mW;
zavn7YB!Gzk!{Yn?#Kk3BoA`AFq1%<>6B4n+J_DARDwfV6aawAq5}`ayWw@$3U4uP>
z#bg8o2TT=?^q*usQcb9-r=`u&_ZaW#<?S<JB439qFn6~#{=2KXhNDKCn3`Ew+1T0h
g92}io+{PKX8jKn1r=x4>JbB7YT40QC5FUd4ADjf?Bme*a

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_block_to_ctile_map.cpp.D0968ED6B32C18D9.idx b/.cache/clangd/index/test_block_to_ctile_map.cpp.D0968ED6B32C18D9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..353428a23ac1548c5e5470639888423466ffaf52
GIT binary patch
literal 13980
zcmd^`dt6Of|Hsc-?X=r=x6@5EH7<1|l|)JEICQ@VX&S{86<dfd3^RIk(#5!?T*9dl
z8QsgJ5-R2~#%0DWJj3u9#JG&yGSW=ZZ=K0%tuoACet$fAy_|aQwa;g-+xNS^`@5Dv
zA5YH>jv|p;l+WUr84IQ}B9TZ3{YzL79}N*~&=83h=T-;DOuFi%-I8};{MTzDZ|u)D
zlBUWkwiukKTF`oL;y2ahE6Sex#P@yqyk@ViWpbF`k<(G)htexvZ3Sz@az?)LNalpT
zxn%)AHop39f_bd;)rRn?YtBf+Zwu?27W+MYwZdgfOw8s}np*qxj(bg6S*UlRG^c34
z^I@m9mJve^{j2_j5MD9xk6DhcrzdCK{kX6)vE_>e!w$TP-?G*=<eB(MoXex+=+KF4
zqJK>jhYh)q{^;DVN4U!7D~0QRORw8(a?0?Ywc^s0Z=MC4Y@9uKTxqOY+cldtTSwgt
zUSc;qgNr%m8aVZ1N$g;LU}JX8rCBPw&RKmX*GU*IU3n&J)nUH{r4yp3c;1vcJ=%PA
z*;JeStbNXMuSrj?h);!_cN9H#d*JYMn9cS2wkKx!!q*Miy^Mp7^s>369U^U3i;U3G
zZyjYCrONegP&lkPa%{U^^7ORx>_B++4+n1>0XVfQj;NUaQn_}_r4t6y{lUHBm#V}q
z_-I^t&5S48!#*||GGXzZsOVX?{z3EOviGS@KUpz*w(IXV21HbO6s6trb+tOYW58st
zFzx)hi6`@8XP7*wnps}`VyceiaAVayY-pd;=~JAJn^~3_Pkmss{M3P!n*Y|CJv(Bl
zY*V(=h58B|U)PHI<zr)Y%k?!~4_XY1{yrf3^Y3Eh=bV3xvno0;uPM}eVfOQ)T_@rq
zzkRC1EOT;lw#jR=-ah1$+G^{0*}eT189Yj@9C}W=OyiD6vfcbu(6lOS!+sEYR(d?!
zF!=e&>*g&dGvo^+_m_V8tLaKrYqz?<q2Uj-BmbP-+&X?$aC}bIqsc!t#75R?K5o3>
z5qoe!esTZrC7C}-vjZ0Hb?aBUSUl-y&X-r0Y!HQKL^#&pcszO6^FBh`V%_QUylwkj
zY1K&RXZYXRrtF4l&BE`Ps_?jn4;ag2Nl>D8WxQ$e!n$$Mixm5Q`93ZoaYOSbduJ$2
zHPThgUQ9iGVe<Ef#rwy9o|8NBdQe_=a+_A&`e>gZ<JFBhy3*ALu6;FXrS0}FwoUCd
z@6S4e)1^NzJ#yjxv%7yL&u!)B-MW}ncJu7NZ%*5Mrh44qYu;OOJyJ}JuRm$=Yq&k)
z)ZX&8->Zi@-`@1y^ra5B^X7izu6wY&?Mhy5yMf#MD?c~S&(6CQ=+|^(P)+;G{bLXG
zd3DHZ*Nr89T9qSO?3UCoAGKlr<AV#w)Qvm2)Z@A5()xoF4(CT4KDH?$S@&4`_+3_7
zKXx2Ers7-hU6jJ&LW;SxF5vv{9f~;D%Sl=#pN+qoc~Uj0m%dj1%N31nYo`Wiw9MaI
zHTO!@g3x+9-4`R{Hw~>FS(3BwlThE2XLan&RHD>ek_IhEO4Dj+kBdE)w_)hG_GNAp
zACK8{wAI;vaoNBQX#3rNd=kEhiCCy262Z@Tk<Wii>J4DfCn8}L6s>~H>t7<J*FZn2
zi&P)vo@7V7;}-Ump7^CLS7-pk!bL)~D4O}}11qSD!WtTEJgW>~=oFD~3W!bt=JhXp
zi)LbJ`*wcaaSc(VNF-?Tnwo!obk;y5GVqHPUEpdX)ZkQ9TmVD`NTKr=R_Nouc*rxA
z8HNTh_*0QEUNoNh>jM{iBuRcuJ*rcq21Aj;-`2rjf61fM-v<AwA&Nk^Xz&`Ee|_LO
z#12)9zG^$>*;^j4sA+`W4-U4w)L=0B_8ZSnoKpBiZc=bgLo^)~^E|Kl*9R6)@)%ie
z8EB_z04Ilt1Z|Nv^VbKieR4muh%Z9tw!Nh|ZG7(|iPv?`8o;Tay`_6r`S?3*elE{^
z$1O#tR8F<%uG#sPQQW1LnPl?ltF><_-gE!fBipp8>2Imxp)+}#HXOJR_KrF5ZIWpD
zf{yv;-qOKml8j?oZ!9c$M+?r~&T3wDe_)v!3__j!x3l@0hmbVOQWB>I!_X~nOx0UI
zB$p?Q>-8@+I2jecF%I{dF8{h{q0jkuwBqH>_?zbUhP-%(xrpLt%uG0BL#z+L*OkQ$
zP-(V*D8Jzv0eNAN(I6|sPlv^CYz{Z(1(~_b0TVXNH@>arFVUaLYkrDsWE*6+LG6cs
zggFf`U3doDDf))8<+v}feWJ^qhJlSm-6`^!QCQC6UY)Zs<n+tOpNmF*)M;6B{<xGl
zj%m&4wDlVo{<vtCrdEl~@_43PzVg!6-Pq74HW(UO?9+ONy^G(Rgx}IjyodDurph#W
z$h{`?Y^BX$E7Gw0N-=xm(5ouE!0Yga*m`(DLr2398!9ho>S!8vxsELrX`{x<Ly(ek
z-0(=^Vim5NN8yEPIJO$P<F4U_dT3n_Z87VO&7T*%RrT_FY&v#o5Mo#M+8f|L#5dpp
zhoD=_y0;VizaAZF>SAG#<W|~%-1>3A9tS!nj=8LzZ+$~xgcE>0LAY;CeqiDpmuhsQ
z#T1K3{7goOVxzQo{QI!)u;1p$`X0IsUE|GnqKD8EsDI+>1o7&QLp`1?*(ljD*lMo?
z@WLz5?-g+HnwCEDj@|5g)(Zt@vG>U<J%-BO<9YaD)5{OgthY3>l;A4x!c3!?5?qC0
z|J|u_WW!v$9#wGGKJs<-mPN06@q#I1%8~C*VWx2SyS%WFS;(#5nP$+m_vK(T<bbUQ
znz%=J;WA(@0}j`c7dAs?Gvx4;a5j7T!<=5zQP3o~iQV+@BM=RCv!Te}Er5`_*qypl
z0mKTRhbzDc3NTm!Y-sjZfB?cdK%7J0-vPuOKo3`t5jw!&4q!vGe+LL4EQew_r0%bR
z;wq?z8^j1zaBvm0p*ge)P9<CoEvumw;eAlNk9>dB!6xy)OzuSY7)BRO*Q4Q{=3k(p
zFflQa;=bU85R(uo8Hxmx1SxJQFO&n5av&u`@e?rl2}p5Ecp)2_WJ4*giS+v!cEdj2
zI0-$uGYY__;vQ)7O=1h{&hWxm$FZK|*;|0)7T`&q_X;?^0-i*_61r4Ef2@xkATrHP
zkBHZ4?9T8Fp6eco9yT|Bj;3r5u+0Ik<O$n>WIJ%cjGu&TxVw+VPr!;^7jj4CgnA@s
zqTU;A;zL^C?(H5($XDj;ODJS?h!54XqZ54abCvb-&H;ghw*%RB;7hn1j4lU0bkC3e
z3Hadp@`4<C%V8klbSO)QzJ#;k=xpeNdyW@sptJ^h6Fvkz55WnTebm;*ueJ<bghsA&
zea1^3UJy_RbuZ|7BliC4T(|Iy7u=?~MUXD>k@#2=Qh;d+5J4Kf4M?^DOTrysS_g<A
zjY@-(G-ye<8k$!_32p>0)IsYyvRI7k^YifbWol#5Ea-Hyc)IYfoh;^nB8SlOO`2)%
zaaEb{eOAV_dC}UaF1LWuEnr0&(GJYpi4(>v@;}bVpaAi}ds!~38oXewZmmT+)mq=$
zgpjw6w;`bvpq>J>NC&3?{S;tAI1lLL0Yk#Of%<NsMP}h0pneBvk-Faj`gec{;Z&%T
z3JnSGgz7t?78#(OP=6;hAzTY}YM~+K;O4m#q`jl}YcYZ)W7*XtMi|5lqR5Z+)5T3<
zgqh4tEo@|rki;Z)HIEVGfRU5tGD1FJ^MNjI-aoCqjL-;}M$%jc4IOif=H-xG4s~&J
zkv?SdNpl%t2V{28d>FEap)PJN4&2@uKz#P)jZsfeYrN>Ou!fx(Q?_+|@0JgwtET(=
zgxmIC-}ZqtH|oT-@rtrZeUZ1YU@QjU$<7OrOymGO*=L>m$$5z8_%Gk^*>hI+i(f5m
zoTJJJdjPu!sAKzOguQ^>OLHY)D`~C*Y!%JbfUPEs-!0IDrQZs=zazx#&ORyJ@HIl=
z<#oVZ2LrHq%<BUpr=87R_8z?~P6nVZp(_~Y#FeClijIm=Q0V36jk^<3xZ5;B-WJ|n
z(JdqRSolz+0A32<O+1GJgrmiaa2~M80}{g9fCXC22zLOl4&aTuLKz~4qs5JIDzr$2
z62fWF0xfogtKcV9Fc^Cd2E8T?K#L#EW6+xu;aWJj7FrXogM-oHNVpyjMsHS_;}%QL
ze)PlnmvM|>J=$7Ey2OUHQ72@}+EO%x9YT>EYe&&gb|^*mtUX1;*kKe6XNObdz&cRm
z$U0Iqf*nDT6YE6LNOmMe&a5*<qu5atxv(x2dGp@X!jb|;rvMolv{b;R0(HV^fK8)0
z9kA&%F9+;$n!f_<S2V8x><XG!0(K?Ms{p%-=GA~*P4gPSuAw;tuo*P31?*ayGXa}P
za~5E;XkG`{bu_OB?0TB>0G~%MpSywoZZM8`lRIGa9Uvo<^e$lU0(HXo0DF(-`+&Vq
z^KXFtjphe{eL!;)V4G-u2-t@-KLYF{njZu9G0jf^`-JAFfPG4HGhmx(eg@cQH2)6R
z-)Vjh*yl920JeqZ7l3_1b1Pt5X-<WFD(s4$7-1*$*$I8I5u*G8WDB4=%?1S|I#QOK
z7eEWT+`a$~qghq}$Iu*50LK$9glr+LUkD8fY5hWIUP$W~Li<8mzYxj_Y5hVNP)O?+
zLAHq2FM<X|w0;pZFQWB}pnVaoUj$`Ew0;o`D5CX?AzMuA7ej+$TE7^Y7t{L1(7u?~
zFNU&WTE7?u6w~@8kS(G0OQ1mstzQDoOKANPXkS9>mq1wwtzQBIN@)F3$d=OjrO=?1
z)-Q$TrL=x2v@fOgOQEck)-QztrL=w-WXov%GH6gn>z6_EGFrb3+LzJ#Wl&Z|>zBcR
zGFrbJvgNdXIW#Dz^~<4oIjvs~?aOKXawsdO^~+&EIjvs-*$P^}0vc4%`W4W;g4VBq
z_7${#1(a3L`V}ysg4VBvd@Tvb7v5Eyactm-%O6NgxoGmlJ|C~RbP>grJD|}Hl0V{w
zJ<xm)bijTBCwozrMHxilm*md$GA>+_P|eR@nKb%626H&Oj4}$!jTu}Ilu__@qmnG%
z67Oyq1#bnNMnOSYUiXZGcL$)-Bq-=gr85dB$f|;qaaZsH3ZiOBMgi{!q4<(cX5)RJ
z9+y+?p1i7?^zs<G+&#IJy3a#vn{8bPN~6nxksMgz5Q7&s0rO2HJ*7-sm>dVzc*#xu
zv0;wmt>#!XhVHiRt~fH|1z(A8*9HXMc0!Nso`u7kO>Y+Xs-!+#+u>xx3t~>JLu7S1
zU5fNLJ&N=>eTob?1BwheLyG!veJC>Gj3|<^vaSG(7bbEOb@1yAF9dRd6iwnLQ53`l
zQ54JtQ}ikKDMfR*IXa|&=W=r?is51?n#awfXg)WeqF64LqBt&&qD_F^M8YIqC<0s&
z$>8%sG2n`6E&*H#&82`VrMVVp)&hMz=DctQaA$xH4y$<KEa1-4d=7BuXl?*p1I_0F
zcb?`8fV)8RMWA^R=o9^Rz_ybt2QRFF+#0AuY$5}488okj+*+D5A(u&W7UZ&MUI)2#
zG_Qx6>!Ci;-vGG{bon00?jdeFBOtRiWYI4B55#R}Tn-KCZ;{QTw7t^h5HCpkX1;S{
z|8*hXW8;eNkuB$6r*g4YR^QOyC-^_w_<{5&b_OVmLGPAz-#g2$@yNQ;y8|T<g#$qA
z0MN$fxMiN|`X}OyDwK9r<|uI=q8(wBqa>uvQ4&(-C<!TZl!Q8Sl!QBTl!QBTkc2yP
zkc2yPkc2yPkc2yPkc2yPkeIieHu_K9<fC<OtO>bGmyMZWH}dg6?oe)@>&N&`OmAcu
z!9(mp?W*MhaW1_}mJG$o^sd)_DBe$Yt+3mFXKra{+s_-FSNy8_#)?{`?VO#`JF?FQ
zGKKM(5Y!L!6fP+_r6t^%^a_Osa%d}uuEeLOLrFU9+UYH65%&$bHR(3Gu?CvekiAA;
zI0Pk!NGw?TAmxjYo4FY%nYSD0?*_)$49XVWUtM_CKH@#bf90p8C(XD+{XUQ}#RqwM
zjOm`Km+kXcYl&GD_iDINk(0ZIJ4Fhhp#W4&k^?kyXzl<S9rT(nhZ=H7#S~Rgql#X0
z_d$()g!jz)<?$-UDoW%3Yh&j9ZoJ?_QK+lwFuuBS!^-J<FXIF?KC06F#i#RJ4L+mO
z^Grz3jc9zX25+a<oSWkt7CH!>VEBh~ANxHwF0PQ>eV+$i>N{NQn}#Z${*MeQKFapS
z+JFzRp#pkS;F#BpquS!&q3_Xy;nP-VP3!(9*7MEm_!)P8Me)o}kekhiM)_n*`}lo@
z{Ofy;4QPzqByL^rxrkicx{h!l3UUj^Z*9usD-s2H=T)+c0(2efP;!t01miGBxfdbn
z0A0K4N_OcWhdh)V)B%F=TB1CjBB_FwI9^w>dlmE~Tn)w5WMNnCRoGQSdt8oJkbmry
zA7>BH_gi<q6pdT=qX(zw^j)MDD0_?|4lgS&Ytrzoz-lY7CJjr0Rw>Y$RQw<ueh@n1
zikum+<f~=UNm}Soiifd>8J6LNTwt6F%&-hEBtzq5Xoh8YVLvq956!w{ULSx&)$BgV
z*yd3lrL+Dx=IZGb8txpKe{l50yUO-RyawWmFv3<K*$M{YiZDV7l%&9cxFU>j5ZWDt
zLvckef_V*_v>rR4Ni??K*sgTIMdLe0H|~r5drbdJ394cJev>VEquQ5#c^)AiKF|@x
z28m1})eT1hdldA?!-qW}YMn9!A#I}r5`6sdouPK)$W6)O6T!j%BYvncGqQQ(g<%b7
z7`A|ZTSzjev99HX^X?U6P(dXeSP4hrfxERnDE4VmTN5hCgG?UD%;6OX^)dbK*Lmy{
z7vBEQ3$sl1m|xKBaqsVLPG(OI=$`}3$?TEC{&HweW={<i*AO%P)%VK1k(&=yeIU^m
zJ~E7^fA`!R9uzcXyL%mM#n`N$-ul*lQ-zN5h-`O-<`-%MDz6o7MZ?($j2c1LE^ez<
zZd1bQl8f(g#c4fMkan-IH3?-FYeBzSV2;gInMrg%4!j7T0J0On4Ria&+?T!1^lE&M
zb$fZ~>9~@a71vPglNg*BMkeM_Fy$zS#5KVuwNbxzAApXiZ|of&m?a{|lw_3DwYRQ3
z%w=){SmN);&zL)XCUQqOn(O*s3SAfQnZB+IIH}!r0grLl1?=s*E@10Rh>uM0LQexi
z)1|5o^8q_bRcA#<&6p{gfjy!~6h;1wCV$Q%e=4i{`X@3bYVLgW_r#b<DvW9`HCDYh
zuOZgd(bn>r;4^-luD9nPU2dG7k-mYkiHqxK*_g3z?n4c2r54VE`~BV3KBi_L_a9&`
mv9z+b8Dc+d_=u6C`Z|4N=ip`7XP~2Bz-;I{&BGrAivABJa@g7c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_contraction.cpp.6A4280349E6B5DE6.idx b/.cache/clangd/index/test_contraction.cpp.6A4280349E6B5DE6.idx
new file mode 100755
index 0000000000000000000000000000000000000000..256ef6d681e5842ec7e34c2bab383ee99a1bb41e
GIT binary patch
literal 9750
zcmds7cU)6f`@biKAmlQ_5QTt7mV`tiLI49M!~ilx1*C$CTES{X2?XIP;=)x>aX>((
zB6Ss!I$B()#f4U_3wPZEx7I~%^>-3@$R%&z_x*f+{`d0vc<%k)=geoHCq{(^2b;PO
z#Gu4bYHCvYcqTy*9Q;p}o;Eoj0|Rq{pvv~gUW;_}Gk<&L>G|W683IYp;VugjS*dy5
zgZ<-|kGV1R$kg{&r}g^d?WpH$qaiY<D`VWkMvthCexGaf{eq~{4Oi1bcU!iGst3-w
zc6UKVe%~t}9=v(_?XRa@8rM7Aed0dT^Vyi<2(x!tOV>9#-1R<u<WzRz^|Jgkv$Kyi
zZmXOi-oLlJ?tIIUCi5Ok|E@3I-g2?!?|bi_my{<sep7zy*rW?x`>wj)ZN<Hc@aGpN
zL|t`NF14$*NoZAEuG(2~#;eZHx64e`vcy<7p=Hv7TBqc72lj7>?SJgX*M<wuU5S~j
z3}14si@az3QHB4lw~UB`mfbf;ZD}QU=2xqlkK7DV=B;pF9dyUf_i6Fs3w!2=I48xd
zbx6<sW<m?;mc%%cHE!Sb4de4X#x`Vse|AUw=40!e2Tz-CvGusE*Y2w1mCVq;R~z@<
zXc3jr(k)otB6~UaYa_YpSeM7Gv6I|?WgL0QwoABmt4B7ug}CtY(1_k6A4)jEi_cnZ
z85-R)w>go!_T|y0h-*DAt-R6rYedlQA14mKeByj?eZi!px!tah|6cv;&0B)(dy%P|
znPCR2t|yJCiO!4f^Ww+eUL)2zFAEwr+CAoJQ@?lP8mktlQ$&fsj!0ipTQ{}GAA?3r
z6inQ6yKwZ0`&Y6`(vqvxw-4rIDV&_ox-FbmA8F42Mid#!uCd$GEj)lLj;b8Bcv)b+
z>Et@EzROGUGHyDhkJ;s(lbBg%_c8I2Wu^=3s^YgM|Dp3vZ}Bn+aX-`l5wkF)a>7ZK
zW9hxj@3N;m%7%aYbjzYv=B&t`gSeZEw}}rI53Kh8Ys@^(<%zKu`%Lw^e7RPADfJf5
zx6m))0jG*x?RoNCQ%T0bAO|5&QFdYRQAc^I@41#E-49IIf7N^WpQ#ot5u^Hr41efT
zp4#u&??FHB<j444ag1p-Xm03r!6JSU*M!G!J{xPgE%bimsuO$br!6<N<T|f*ZfrdC
zJ8>x>_7A7nS03`pri{|nx2ykjVqczfL9#Njd7^Fa)@@xHYIZ-HJEiN>qs0w-Hi{Rv
zbh$wO^h@_8+ntOulJC^?c~qH`b11<x+NEiP8`E$99rh@LsLwW)(--94IB_Fs>fHv1
z8{Ye-yd~3e-G^UX?KPn1lqEN^N8Q=I(00rpei2Ib-=ogdnKbW_)Fm#M7mH5MxOLXb
z@4&XO{qhE`a#!;YH@9)~^Fx<TobM<-^ze}LZWo_>wVwAj-m;C&dE9>{=f--~`ocSJ
zCha_bH$8M_{^x$5{v4IuyghE;r;5|ZetztG$y54n(kbf)jxVy^e@MGhxM1?a#9206
z*QPn%^Ox1NI^4S!|Kz|}_0o5BSGPvoHs;>*9sfCR7r)Q5%zp28Pdz*DQn=$#Pu@$j
z3*VZ}Gku@B?D8;=7pIe69sAT{wnc%_=-=zlXUay-7_)C$iEyo8;M}=35}DPokw0HX
zy$<i%mGjB*@Fn}rStV2J`*IVD4}K^~>e(fE?dwnS4`q+{7TxW#__q5Zn@_7#r){Mg
z*&j?<FIzvYx;k`i@;GI*<&*7u%A3WDq*JySE`MQYv9ZwFJk348ZOsa@#Fg0L+-g`f
z{n{-;HDl`6YYYgYyY<3nWeq288#0vGL5*XLv(+}(b7}uPYWE66S<3lrVhO_ZDnwKv
z^Y#a!@p1UC|GVh}dDEQJ*u>(0+s9|n`$JBxr%c$4*eL|%Z0yX@Hf<9SgwOe<pP#H;
ze8Gg6gIB3WL^U#Qf50jM{dXuYy#3)Gn-Py^a4~kV&^9o`z?x;3PPkM$v57^RO=J)m
z=Gq2^%l8%8iX)>3uo+|U@IJ<Ux@#L4t|<Cy*p&X;@=Y0u;|R*$-rhmm06A#c>sQ~_
z{xy=#7>_46Wtkdj8z6_5#hab2X5iDs;mP^Nd~0n3<Y9^}rwjP6s@RMeOm;GM;%FNn
z$27@ZXNApwY(mV_Y`q2%HTd*x57_>Qd6!O84VDAh#CMuLnM9_!wgGyq_YQ-hgm4m@
zF%pwkBi3r<+WvrLzAAdKJLts6@oZw@KXPbfbk%ptiV<DeM2?2O4iW3HsM{W}OngbS
zSoG$-h|QRcyObbS336(Gz%mJ@dp@76=d5BgCg3g{7RN~2bdYQQGEw-hY&B)1V_pcj
z0=~8Za^kwfN8b&ezn8@ri%0h`?7`7C9pr~U{<OHY@~|OeDkkS6Rz7lXe?XVa6#l#q
zxttn9MhfmSgP38gZQum6duFd&BH34F&iDqe?`h;oY8x0mXAs4lZso9?O)U7gOqdg%
z7Jv5?=RBJ+8c*S7?AAluz!XcB7PUUR6qnhI5qJY;EHh(m1LRdv8@IiC^z#HZ<7-Sd
zXPFym8z8F%p_?iLC%c(4mLfd6hrNfJwgGxk<lD{c*^Ah0Mgk_g8oOF(8z8T_ZC$it
zvJKmmu}m{u#1o0M4bazLn_u;QS63qwV!q~lSmEYjU2A&)i>g1W;=9&_J~w2f;V%EE
zjyq<uJx%(A#uze^ahD`klD)QpRerkV6@JHX#7$Gi#Bl`0wdXo$8z48W+`s0{A4YrG
zjIZ$M6jn+PZ3E=rj=N{1c$SYfWhCLzcJ_7-+6KtY11>sU9W!c-A!8aI{g1xl68lQ^
zpMCOvHe{sYF1d)6i~6)bV3wOE($P5?2S2chTn!`s#h_XSemqD7mk(nzCSkIQrRt__
zVEE(w+1L67j-P4BP+{^vvhhiN!IIZG%tAxPx3~+B#k1BnFhlEW&Vx+JLo*X%u4V)O
zqNqe#(nJ-{mLSlJh)CHXMGIn5!v_x!1qEi+y&pNpY)XyZXA}1LCG`^TP|}#{#_DDY
ztC5s4L^%`~Pg1ENsWg-&%<`^x_$9?|W_eGN$`)qRPyq@lKtsX5kW?WWT!@D0uo{J^
z>GUELQbfnaD6|+2*Y)3kgd31|i=e(_UvP{mNo_*HO*(89&sW%{7))o9l$4bAhQpGS
zjFiz(AQ?zQa#G%bAf7RdJH;^)%ZNP>^#r-~`r4|usY0HF812FX0<ZMu;Fr%Gy_rmH
z_yDKL)M^NLVP?Bq1eeKdkBATnA!;%h5Gvc{z-%`saQXet<tgU|-0RLjdfU+7mZvQf
z$u^&xi1*v3@n8k8<bPethriW?bL&LF7aa`3co$=B_y!fn;&GJmb;24f87v3nvwmRL
ziNjm-5lPv2+YGp5Sh3-opLla1;&&VfG!^1J(^wndbxQ;~2D-zfBqev0(@>}+bP&)W
z6TlqNWC1WQQ)^;ibyAlLS^e*>OTpaGSRQ<4wu=GWB(t5{uv>lk1r~XHy4O&%0iUAH
zPlE51r)&~BNO$X7CuAX5J2feY?r;&3hz`q!Tm&;&lZZe_B&h&lKwmf|NyWs)!~=r-
zgk{_*FCnP{BriZgI_oP!!XhN1WB(!~)!|YkEk$w&^CVS)gcV2xIBd;<{L?v4O$@O+
z5i2D$cOp)up#mf>KoZ@gg-BdT$7&>2({T|J7twJs@+n6CFdmXC%nnUvrQ?~%O4JkL
z9Z3z53<(Bygr9JJxP>uvB=#b76(HMErgn`WA-CzkZP_g-3}kJ1T}{fSTUY8RrRjQz
zB&0*vqwgO+H6y0wUAv8eCbl^bh&bRpAk_fp0Zt<((!%x8N`C_*lCm?hGlOVIQd}dh
z8AMi+vNy6f15F_*fsw#0p*^)CDGwtLGe{&zO3V@40B1<bhvP#-z8qf~^5gh*AaGF~
zsttUJxt<W#ToNJG3q0T(vDs;YDfHI2D3B_)fVM0ZI%>q>yAH*HPmO(@JZvC)ay?!q
z&OFWz4ge1MKUz1;fjh0!rm?VyzCD9t<K46+v(S;mj@!EmMT2`CeYSC0Lv8R4oXlCV
ztc;)=lTunM1z-nED%CobhE#l2AmkBDDxIHBLj}mX011E_OdM9N3+Y&mtkrZ}gsh9`
zxENU%({U-XE=2-350ffG)@5{Dj;zb+xB^*M&~YWQuGC>j;*E?uCGR+I5T=UYxET4T
zN@+-*8l*$%^7J4$JV~X;q<^JDY9v(Cy^tFDtLa{-90|*HUI+9V3rTZd&>5navsZu6
zi-gmGH5IGb@ATAVELD~=ZvVc!IQp1z%(_EvK0YFJY~S9iKkD@!U!8fJ$K|i=2ucuT
z!3X0lXznB+er$K<0$bQ$V*BL;cWS$t=(<%Sb~O^gY)M~zwk4bICgM$n^bQ#SM@WwP
z{%WsR%tLyHl}Zgf!##2C@2*Tr;4TorY8ZJ4L^R|l7(hdV1PUDzdI)Jfg?>U>PvIaT
zt)~ay!yP<F>aHt8-Z|uy>+MJNTfCWp(=|0_k|uP*eQWwd!=$6GK9}`a3S?}Gn%=;Y
zwy+C1XO?OgnUXt4`i0RL7>RE*nu{GM3OL&T>J8vxhXts))xmg@3YUb3!Q=~_w-3)=
z{`vuqflfRp5kv`+a^^XUAU`507oLj<QW=tR<++OB=0;L|czr~W36PW<&rJj?lN6uF
z7eT~Z{4$8^cV?+A))P~XsSO1AJi~+Qmi&@77w<q7t{SE@tMu@68Y)HNQX~P7Kw^&|
zE~DdeBrd1p3M8(e<4WXHiTq)FURaj}&+<R}YNC>rogFPIS#j-XQOU~QP8XFttUT;M
zn{XWBiXDNAI1X`rXvmlAOGAEKKN?clDtIss3>YUmnwvGKljd3?FGRuMPf4m635!ub
zIu0mCfjV4<q-AIjoQ|X_k+2f=1H80aeIsR7%QdW!VrQ`sR%cQ^&OS8c>+DNIe$IX!
zXnEv{4{@2>JoT($Wv}=*!y=v+>V1ZqfQ50VS^&<ngi4ybRr)Jw?p7&R(%h{|uA;eH
zRg8+}ZVQmGfbMP!kbeQ)-4-HYA)Q``{0r%HHIl14Cz0|ZIxa?vVicm|Z7C9#>fE<R
z^Mqw|T#lsWos&#y1szu+S*1=kpg`CvHPsNERVqXZe~=6q7w)8~iRe0oNx}v}r!~_I
zR=>1Zz7yM%CZz!7fgBQ#)znLLV~2+h9|>bC=G}LhVA{AHCjgc#%m4WV04_k-$+s7W
z+PMW69PCWnE0pa7xEei5<?Xeq+ShxN`2ckedqQ_TRo`+)HkdK;4AP4pKRn5q5qG<6
zx#|CEMz!_Sz$ORd2A&=}ew_6Ci##SjQSb8(E;qjT?fT`IOP%kPP<X?-s;Rnx3X{|f
z_Ke;TYj^D^$Q{4^@_el4JU*{4C^UF!e4){l<RGUZDSx3q?Oquj7u|8MgaD1}6ZBZ$
zIr7&B7gDvdhnK%Fdhh@2@A<S}$tyipG<ZcN8U$56!H2PQwiJRL9yr!`a%1pQOMJzn
z=M|bdkEW!lV@zAQM29<T?V;xH`xB=dCgVlv#kEGo^F;FPp@rs5N;*I)0VjZwOd3c-
z!O{>KijWS`A=v;KZJbOdqm7dV%V^_d5i;7iKuKVSS7TCvfq@;jON>GS2kTY{91<9*
zLvo3nHclQWr;U?`$Z6x`L*yOf8aEA_W)j?WL06K>;$;ay-na<J%ch|m#LGbfs8*5G
zLd080#~ToD0}{}Sdb~|^T#I<MNC2*Zr0Ni_PKTQot>9lW`q1=6*JikRV9U@{lyw5n
z;AaHDt?|rV*P)|sUdzB*qcl@m!5ExKIjP)qiBthHD?nECfM!;TtROjn+y~o{rZ}y$
zBZ*K#v!kfEs17?iJ2?Mj-u9DrdTMs|vE&+qgO2~dtDP5jZF2O^Ex(P8qnmfP04V;#
zMFTtO&gu{MUyu}3<8ZauhGoOnF-;YxO3?8mJuaPwa*!|wdD4|ZxR8!Zk)jlZ0=G2b
zKd20a>2Ntxl+)=INKrw@l_;o^-oF+JYmp~id4zR347GkN_qIa6?hrMIU4w)Wv94Q~
zt4DNyeYGQ`LCF^iGniZdsd8~UA^n!Gd?i~?hIjmm=2;w%jn#8ocZ*`A(|mi4`a-98
zUu5vc(99+e3s__%vV>ISe%RGVe#H$t^%eylVRzj*Z0Ls(kU#F&xtt6N0~!Q6fVr#r
z4MfM=OxH{wDER;*F#|h$Unu^7t~}{llg}7Xw*sHW%GpW;MHL882F9I!At9-8PUHGR
zL7$|OT$98Q|41xs&s-D;(kJni+dCgA;11q8D&}Op*VboW$bP5`w`VJG{EwWKlH?y}
zzxcx0|GH52H|sR(&KFLDe<?~HzXtCOSVe$9fExse6WB-SrwRnf02oSu*#Mzq_KpuQ
zd&dWuz2ig2WU90Ys!R>{W&sNcB3t*WCAGl}-Pf5NuZS9#S(fDz&u=oCku_`P?Cj|o
znZ)!CGBI2Cnx%W4p?jS$H8C0g{{~|C)aWiu1D0V|BQ|MbVahQ#3mq0Zbcp5PU^h$S
zAyz%CyV=_L2S{as@<EDzHhl*=4{+!G)7S3y4n2E2I=S@W3xpn?{k?qrB)xpOBJU8J
S?ylnS$UMXd8x<6RqW%y5Mn$*)

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_contraction_interface.cpp.51D57FFCD74AC94A.idx b/.cache/clangd/index/test_contraction_interface.cpp.51D57FFCD74AC94A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0bd764ffd83c70f8a713c14b1b00aa3e7570ad22
GIT binary patch
literal 11242
zcmeI2cUV)&7Qi#%0tO-_hOW{f2m}*Aam9d$AczP^7tv5`6wO77K6jB&MF>R&DK->E
z7H~lY6h*Nsc5t5+D+{_-a9MY4??N2z5O#fk_%D8+!~NZJ=gyrobIzPO6E7Dhr^lug
z%8W3VgsAXUe2hY&C=>r<R^iLi2xwBEP!dXy`Ocqr*M`}*{q|l_h(kbqcv|{Omy+Am
zB8PP696jy3#pm<=@?)Oue=xUMX<}r+qT5AoW?OglZvBv_cUJGLTY<@`d1tFxqTa~f
z)j7^DUc9^d^dpY<ytJKCBvO=3nXA~>_SeG`t3C@7?QLw;mBu?<Vy~b+>Ur_;={YsC
zpMtg<_rCUP3*&Admot4<{{922xnCw7E(-qf_$(u}9kyrQuUT`4yI8<?&6Gbk@72=@
zHJ6GbPg?uUH(p@o7NBuXEi>ORV9xVS#@n)j3EAbUmF~|}diEE3<NU+-o59XHqm7v_
z6dpehJso}6K_^;p)Zk^U-qwp*Ce1pXCw~nLJrzGQ?sxfpy}#T}*pPYna+djOrToA(
zzs(Bwk2s}z%jKoUmAtqc-34a6weLT!mu}VE2~vGF;V;X({helXPU9z=mtU$bSCgJ?
z^5oF-MP8L=v{x<eF0AOo>7~DK+2h1izh{wsJbBmX$ai%HvusvGJj#xpCg0Ig;-&EC
zrN8RtX=t9e<8u8LxA^ENPK;S#S(DvSuqUo6{FROC@({y@^X;Pd8>{%H(S~!&DSnGu
zqk_|IV=Ky6DQNJE3^v#0Viz{2JY1*v;Mm){nLiv!Yt8VPm~mgfWWwq63qe|rN6LL{
z=fp>J`JeT49M_Vs-$c!K<8MmQs9*B_&hqy~b&S=?%XDu(5wymp*0Shv2`7&7`75ow
zyMJU}+grCFBJxA0SIMEp5z(<L5B{;EYsxm|`RxHt1vf=q?Uu`BOJ-b%*;?Q7TI$2Y
zl)ZIc6_o{zM$KWS+JO^J{~W_fQJm_WcyqV(#Y^X9%6nx}qqKf{BCFbQ@m^2mD5Yci
zhdRpHQGErCETNHoiu<uB^Gx+({gSI!-fH*U%odzacMp>fKh3D>h^~Javi?qLfNQ}G
z?bkEij+o;Moh~_6^++>;%4(}aXKKCgJZX%ZlvQ#nS-pC9Z^GJ%ceMET^h?g^Zb)>D
z9CIuhH|q9Q>l**Zxl$EDv|9};@=MH#CeEKfJd~l{R6qIO_V(HQG|Fh*!_KojSYei7
zOAIO}WO_E$T9<_#&*6#s9;dbobWL^3#%Q>1y4(FuXt6A_FXlLPhQjvi*4*pQMDb3h
zCQH-SpYc4AGsW+wxm>+N#23roySsS@&aH8XJ|PUsVb0-<+LK_lc4ytunpuvP@eBNX
z)K#8sjV+8_ZPdC{<icFwJb7g=NazaVww?43qJ7T#&rf5Ff*4nt{V9FZEN<L+x8M4n
z-DmYLS31{UUcjAlc*l)fFC9nWb0ZRbA5Iy!q|2iHZ3<;oO2Zk?vB&+i8ZPxEdLQ+j
zuV_>tNGMV-QTOKe24?hDirV(5H(gDZnIQOdUT)U23j2oU{h!C0=i0opk1@7-@n!nY
zr>oya{CDG=O2#9(pnaJgZF*kJ=qH~`s>*lW&udsZ{^+h`>1oTB($W@Bjn&-v$~o-%
zoTu^J-F+wiyvtJyZ;M)S{f`?>6@G0&E;CEzlhTVlMYdY6pX76!R~y_b`h52@?I%t`
zg8nrpkIo0{YRlfN6s`PgCLS;Cm?#xlwPbP3>Zp(dQWVN)d45LYRBIP`syBs#Gh`SF
z0|nu)2L6p+R&mYp!>{hL)MbSDAfOxs*w;Ule#b`qV}^L1h+I><LX{dCM!`1#$_C;Q
z@gMkz%e0n{V+2-gRciQR3Z6(w#0Cl|_IP)%x$|^Pt}1oOQVPBqP&N~95dVRX%=<eg
zrpF?=PL>)@e8;di%=<h3WxeptR(a}te+up;=fxN(#2-kZ1S=f*(%Pt;FHc=S)K-^K
zrwtTP9GdF9bNM&}T~+F`<^R&j)sK8{{B=_I);AtWKOXA%__ireiMo*ZhG1Io*nt8z
zvSD{Zh@0`-Px4gXKnkuer>{RyKykj;-u->u4WaVXAmS6%WYlN_1r!%_&$Dw3%oYBN
z2CF|ydAQD<x$a-YwH<w`Z+F#XePf6Y^12%<1KYA>sR2a$<7CF^3>2{aBg-DGda&w(
zhb(moA^yi9I@(@*yH`b-c!MADh`Wrt&Oia*aGL5<`n=}di*F3_nUo*v_ULttQKH5X
zDl(;+G6xFyhAW#aTXmP_UXrKsi3XHql;s8rD85$G<lD)Oe)NqYx^LV1=ACnwg$i{M
z(SWY5uJJ$t-_UKPb2j+5hw<O|#OEe&6yk*GFTbJ5_f2-si_AiB6>2cih>oU?-ar8x
zdH;Jy9{aaNdh*l|q8BPMDzXCw6o1%hzcKK{5;s|DC?WnwYog%ci(_o{C=|?_cK)h{
zs{^2Z9T7N6G@H;~jo^|5Ol)V@=7%e(XP+~IX34#5xoGC%iWq2Q3m4B#oA1wKK+En4
zxwcv=L}UnkgS6ZbpFqc1Wq?%);Ebl@jvU7+aPa8(eE0dj&>-n}yji>zG!5uYs8r}f
zzIGe73?E=Rltue)lfLzx9e}QhP>@)YP`f7V*`DrC95xXGu}^~yTbNH(@d-IPR*sIR
z0W1wD!EvGE834;bIS*iYDCYw#ALZ=;+m7;1fbB%N5@3}mR{^XF<tBhNp?n-*$4Q1`
zNCgsIlk6(=y+l|0-6nLTuIpw5<V=26FibVBMaOflB)=F<C5BnRiX#Wv&&$sb4y$X!
zn%;%dL7xB}_p<bw18>inoWDZVJS&&Dt!%Aq4-MA6!aA_JP$Bg@M-wg+#3zd#5S;F?
zPsFt9?;;#B@#Kd>_>EM#fu#X@D>@!V4`ad!>g~=j?#{E79!tkfHBEKMIc=tChLE|Y
zIYMkrHbNGf76?t$oQM!dlY<aflZy~<I?tJG&u+RMLcw~$Ecl6Ze4*Y#gaoz%2U3RH
z;s_-HS`tt^V(#M|{U~)59Zv!(NkEHaA)pCSxe%xbQF%6?Wdp?|ANMglntmG{7Xg|G
z<w8I!1d5+puScv)@A0MMTL5hf%0+-ygmN*U6{B1NXeB7u16n;$?Dgz=8sw{WjE*+|
zS_8_BfYwN|Z%?ELPkvdWGWOM0VaG5$9boBX$FQ%qOUlKz3p<A4V%sHIY`d^y7%sM5
zl6L`Y7y02BUJkHwl*P79J}<Uyl4}4~gFgQYz<xnlZ2M3V7=8?3$H<OjxY!qv@)m%#
zka8IA5PrKq)`9jw$8&&54lsk=>}Pfk;E<dPv~z(T$_BZ>7;^BH=~de{%J+@v8y~7~
zxPB78c&KBNbX0sdmxXM-jUPrdD4{_Zfe%}cqt!v^qN}CC?_4?r3B0RgFW&e>gg8mS
zDhXJVu7H3fFbU;pNq~p)>?A<!gT+1+!-c>~h}I`Sh}IW^X+pHV5X=^$^|OI-Hef-U
z!U&&gEJC>uunU0&=_<AW_7;?j0J{j~V!$p&xdgCFP_75;dSF4iiUz=LK)DgH8%d7s
zk_|s=Jo(nh43(InakCvSnfg!iAVM3R89F{N$WJ)+=;<2l`ehk9ZZpYdD!G93Ch-um
zn`DPj1Set=xnv7C0)%i5K1guWIm{~l6LW~ZA~VtMJFNmv!v7mX!jF}a8`#p4u~Qlt
zn+7bPanOl%S1TQ`NzMRcGf;UR(8>dB*e;z|ceV0K-ry>g@-pX(jTRl()zH;}yEP&f
zGc+|LRcmHwhLE|TIYMkhHbNGL76?r=oQM#|kb@A{kc$w{jW>^M&(6&bp<wgi!Tk&o
zi<vJ(NZ>4Rhq|KUxHFDW63|Hk#-wV5Ku3sjHehB0EmAciz!afe2$+RH3;KorSc|y@
z<s!f=Lb(_)i%~8C%o3FA0ka-xLBG%+YcU&8ZUoFmlHoRhh@SgX7cg5u$E~!ite~ny
z?z!&+;`eh$oQH(-5Zww4&@(KcWAq^#6go_keT~rQU&9A7U=wdnA?K_uW^dHk6Gr$0
zbSsyq6Zp=<?bYTHwM(cS=G4^XhUC(SjjcP&6V7OYd|3Vn1+(}FEoDU@6wg|NP$El+
zBtcl$F;+W9ij#qMGSDLpJ{cG!qXwT0Op<{aDrYAH4#_D%I|Z$u0t`~n`YFI91+AX~
z*ePiJRG^)T)=vcnsc8LFV3LZ~PX+8$w0;iI&OtW=Ilv%?+zin1Tws!mZYFX8JD1!{
zz=Wm5TrG;0-uQ+gIw;$}R`l=D;M}I;;yo3~;yo3~;yn||;yn||;yn_{;usxrh0+fN
z$?~D<BSs=&BvtXxvi4r(hJ0r%QI*gn?GaVxM8rob3M#5F%UR7+ODt|bvzNHd8pX1J
z)=_<#cVWC-v+l@fNQ?$d<V%=M?b43d8?XssBl?oi7d+V@^fffY3nLUFWbHr9fNp?}
zZw1&^FrJK~wgGG#$`t^sKzSd)_Mv<jV28nY7&jd{uql&YcMI-MVX_fs+F`yKX4+xK
zQZnCt=xx*aZFP=+NDO<U)8>R4`>yXzgEuAyl}S7JDoVT^E;U5>kA@RNgA4e@L~d1_
z+kl#s>=N7*5DM=<`+<+q@nnD{gYmF2I-UZs6qHi|mWpx?z;eKNG7!!MST4!OzSweK
zaEjlKSb~HlG@Fm@yc1y8NeuGPBosV&8zz8Z`)Ij=V>W0pbbJS3>;U6P6Db3XGL&}%
z#%`4N0LC7ae+G=7QLYAzYLxc_#(tD*0izb>qd@5>x%HqE*-!OjKp&bM9X|n-PN4D@
zpx#2t&pL;nR$};azH>plsFm=TmY^U_yoMLR3x>+VaJv9IK1l*^fj`{xVL0xM&xZt0
zKWIabDT#$}17PK#f#VjyDn;0V<6sRNpyM+*GY|@J5AcE~8FW0*JrJQ-v)GBSaXKDn
z7B@)O{G&yxSwG$*rohnSE_gEdO&0(UR)^KTKCUOBi?>RJ5Day`*MB>?;)q>)mrQ?V
z-YN}Em!$#wG%y`{gZ@mtWje4z*)AQ}lbivpGSK=Nz&?YlfAdg<zIEo-M?`qRux3ny
z2bQ<-#i4;<p+0ejHS}l_J&c6u1x#+>5SX0)EH29i1}em8O=V3-$dTm=>3&;mE9=c3
zpJCm9_+*`Zp7Jku`*THpI((QbdZf&GP{UhmFiek+rxYeJ&ind~_%sQh_V`YN{jZ%%
zy9kX9^=X~1ce<x+X+AijF$s;qpD+y582sVFu*nJ!`w15?)Hg`_3h00T?}P6ZOJBrf
zRknYp1BPZpq@4PHM1k069bg?9^p|j_F-mU8?<fev&7iW@Q=zdETtsCRqr8vGsziC6
z6e~#zc|K2ym5=gHDOMTEB^ax8aKHa|ePIiGYU=^Q=0j(OhC)9K13)5t61zwk7fD8l
zAKObm$y||n{GCdJi#9PxLoG(qg7EN2z9aE5Fo31NUqB4rfU)f04Hyaz-hiEp!9!!>
z`#UCpy`xZ~$jg=FWi)xYV(?N5r)b^h-<95RmKL)rVnt+B^hyCGVo*$3MP9~`myzUU
zXjE8a|LZ)X=1F1FGP0v+^7PRR1!W~g7Y~=YbH>baG8rQ`hpC~Ws;Ol=&CY(h!wg4`
z8k=YMgQ@O6-BlkqURzJ!z=$=$%-n)&X*GGuR2`eK6Rn-q)Qu<2c8{gH`pk3#UX=d<
DTP9E{

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_conv_tensor_rearrange.cpp.69AF0D8F1794ACAD.idx b/.cache/clangd/index/test_conv_tensor_rearrange.cpp.69AF0D8F1794ACAD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..67659fc9758e4ac2b57ccaddf224ac31b0bcc2c4
GIT binary patch
literal 5752
zcmb7I2|Scr8-L$1nKCaiMwaZ^4a0;ngXmf^*&<h2QkHp<HMDbWm7AgvA-AGcrBF(X
z>hl%Ry`@s2#VwLn@|7sncV={EMz_!RGr!-QdH?To&Uwyxp8xYa=ZqJ}$ti9Ig4l<0
z)<km?gNO)%$bfG`VtfP*fObg)SyNi;<LfbtB|WsK<B;b8PD9OXXX`kwj#AF=+$$m1
ztB;<qta}sqK;_!#C!>#Z)ur*N=P44$KLvbzf+N^vy!dngVRgA>%gIdo`qF=Oy!iNj
z^yP5RQ)`=IY`TTQo;|nnI~q4OM_G3+Q+^$5vioRJ;3KETyB#CV+fF17zx%kBeg5q3
z{Y@ud+}f@X@hO_^5+>htCjZs0>+gs2EKi_a?jHWjN~ou=>>mh-KT*ix_TM`2Fs^>E
zMYo}_;6u`J(<OUnAMT>({uz0ITfVim_KHf+>YZ)N-`4pj+u1*hI%syTHtUsXhShO>
znbnE)4zhh4oQ_~-d-e_GCf+zVoU%DaPJ%n=Y)zb}lCaZv(V*V`trCBGh9z3e=ylMS
zO?_MTuBD{VXioB{LB;a0>{fA-!H$PSZT2-K#)v~ZWiw^8UZNb?taxJOA*bMH3kK|}
z9Lmmir}ZR}I!kI0>TePM-p5W-SIaQS)k=!1oqBIXwY*QS#4Gyt-}W8FTg9k>mc@})
z&)GH1#NLDAF~@uhP7k%FM>-g0EQp_1;5R#^{A5*+`-%!;S-SJvLH50hJAMZJ6`I@i
z+bF>@)caXPW?Ys{d)1Mu?wJzB@p3BZx>5dDM^js}54fJWkmYiA&id(!hjd;lNZFjz
zj_}GW_3GL5Ds6$4{fhZ|6)`=2v=4FZ<rk}Rl*FWlG#3B9W#5y$NAlB>D?4j7YHW?S
z)-&U7{;@gjw9NaW=ZXy%ZZHxTtlu3ny=`|=j?vBRsQfP7p^ofS?tzM<e!Vt>Q)u?e
z6-6hEuN2KGxBj|tpG+I>WA1k@{BZS>4Cabf4a)+nKsirGHRtGi2l6G~Y~P~krzy9>
ztKV;%LTg;;$rr3Ijq6rMmiBaOH3axqs~`S2vp0Ztcy+T;<$SrEAqBll<()e*T805N
z^!HO<D$#HHO1$6dRn%{J)w|5L%#*Q}@lfrxWcZ_n=G%hhz6B3lUiF|<-|dLUpxqbJ
z<IP%UEbc2be+zihUGl(f#XikdAxeviR_b$GSo51c%f&3ZYxm-4tDU!;%i%q;VUE*M
zq#85oLyx<99iBgR)w*VR%QDXU*cW@_zcyUCT(httW@g{PdB+VcjOfEH_tKJM)9bSR
zgZ1`!?F(yLT(hgQ%D-kPt06k+cAXq&7Q;I7YxH^ZP6ONJ&*sHm-*%>DoBznW2ZNcJ
z$2dlAEV|8>{#QYHqf09vroZXm+kUsRclG6xTPlJ5h2FQk6sc_mR-KQMv2_`LnNc~$
z#QK>&>oF7MTS@J$uaC`9eb>~t;VY$q_=e`kRcjsS@pP*#cOOnQaOs~jI4EZD%aVxE
z)KN+|b6O-@@5{3u)}Q)Mhwm0#G^$ZO-8IAa=6#pk;e)GYeb|)We=g~&%V?JPi|YfX
zU%R#|*Ejtc9yYuCW07KPNtt}w7jnqpRQH6n(ZRO}2r|RzKNu|`U7Ji;j3Bs~xS1@!
z0a=CrF`R$>>sOV=O(eonAdW_&CHM`(OFiIYboPBgc2l4S8Ce0u?uffIzk%Y#E$`Q`
z7tb=6M52E-Iv^>!`uydgPb5MB5Zj8|s_~n#sSzZiPceTKbG=GN;(*u<ag*jZaPnw{
zS8mGJljBK*B|z+ncuMdaDBe;Wwk!U{<FjNW_GhEZzBt~stO&v+k*J@I-tTj=owD=Q
zMlxX$5Yxr!vit@vt)kmCMY`*392rRjVh_Yan%_X(N403`$9}1xlL*Uzm?O?n<2P{h
z2`{mh2+854WF#Jl|6#RO&g#O0lxsK0gkOUZoFYL{<2P{htsU<+=&rnLL`D*Rru*LG
zKf+N**)=3WAP_r<JE-v+IJ#r*C9K3T(vd_60^&>*$wVc-|G?FEzwXWN_e)nK6MR7z
zmN-k6-@qAOPJbR0x{#bIiNyS`83xrK=e56NdPxyNzz(Q1DwE&98HR$y?b9zc#*hg<
zKrAC6L*X}2{6#HNIsNr=M4I3S#PZYRF@6KZU(<T%=be}Ik_i4lzm&L?3crD3B%X^W
zyabgHGMgYVj?ozIPIG5L;?A6M=hM+mab+lhPaeoA^-cC1&B|vn1Pr%gSXn}gV7N8I
zT7WcmbvX~bon<aY<ns~|j>z{YR6yht1Nuz!$?i4B2Z&Mxbey{J%C+7)kKBx*YM>ad
zsG+C>&7m{p^bX}n6>l*NUrJgk2ltENYe{QGP##LkLj@SdFM%)?x8Fw*`N+?^?vnj)
zB^q$8pc9XbaOSa2Fox+yR&u`vUr^?)YN{d?b~5{Ts_(OG)*bL#MPJ1QE|zIfDJk=J
zi_Ii#W2@7)es*_REEvKIW*DI3S3&&FRY*dI{P%sH@v@JBqgV`<z%>}oX0Qdwk@`la
zOjc|i=w!w)V?ttJxH-dIfE>SFd8ne|bsKatWtdvP1c2cT219^c2*!pw#v87QV>sQC
z?kMbJYH2Eh%vfelLM;|73lYR(v*ro;WW}};AeR>H=SK)0i-CiA$qrf=5LUz~*jS?Z
zE?oHR%3o`}UkrUWGt8J{6ohGJjDj#_dAn0|`*N+P0EbbhQK!RD^%;y>FozsbM<DXu
z3;jsM^``2{LbSN1w$-L+em}4gok15^1;b4kCISQr144>dFW@#H*#Px<uLZ;$hhRkj
zG0Q6la4d#<l0CIxa4xxRxptOO=M3HsG^hq}2Z1-{9^0Ky9-43mOV^afE~~%L4ICJv
z8lny35g@}5h7cvICTqh5V|cP*GDC<8QPo0JJ9A3nhLJ6MreSy?YFLOeAP#cnEH&25
z|2*L;xNdXn)2a^!gH=Q*MW_q)MrcOp2~h^Bl!2-XHD{oj8K|BRm!e9gs5;b);iaf%
zDXIr?gys_kxBPQzKcqIu5fJR6EE^wm(A1z4FBYL6#yY_T#=4wdRP{L^4e&Lb6fQ_P
z7@m%j(uFAq!*`*iT|&I_If>h-%WR*}LhBQ1vDmB>+esyZjoLMy$vK0)fSxkCG8VAR
zU^qpWVgaivG!Nk6SoMPyk?;e~9B@=--Q21KoP?|5T5#6@8Q>;Dl!~gRqFO?yq@o6?
zsEH64qpHQI7SxL2#i&6sY63AV8WVaW$9?;-@9Gnb2`!eZ?)QFy>)V0}I~B)F$seHP
zoC2Z4i)-lD64O%^t1?WhfZ;3{7J`(B;Y<cofOb8%P;p1zHUoDDlY&LuokU6!K?NwO
zK<I3k#lMpfDujs>#5wP51z&^2M{{FAn;e#A7iN${6eipeOl!b^yki)g>Bl+&Ph>zL
z?>Gj#VZi6)OjHYU;OMK2wZ2bEb1;zLXmnG!OpxGcCL+j`W-5XhG=>N=qnVisbH6#w
zTm)IrEJP5K#uPy;8p~8z7uYnm2(qGCi6Cp5wE$KAHK$os;-bccHiCr!*hsXLjIXTK
z;g9>b5Dh}n@hQJLa_;{64|z$zf#bua_T-MNpqln{KqV=Llm$GM)j1WcId1p7eNus2
z|IM25+O*`|guA#@{pf~$F0v8Cqlu-7NVR2HGDMI$%Uq<|GFeO!#AdSv)fSxP*wzAc
zsp;R>4>`S-1>_b@c-XiS#$yZbybRgX(4Bp?+i=?kFo;Xyio`$)DMbVop`;>VL^M)U
z+<l!!@bBdj*1_gY+YRnd8*ZLNCBNCN>*3$^U=>hL*P!bN$!exyrX!>vOM|5Yi3O7&
z2m{`|0y+)DHC#1<V3mi5)&CPxx5Z<!o1_c6{=p1*0|82S*AHkdU`uke7Thn4IG|*_
zQDNl54B7GVSEVDYJL`Y2*HhNcG|Dv}p9v|`z0=z^#!}sXQe=d-0?X&!eT2*Rpn0&M
zZJy2V>|+&OJT;+uct{mk2+jqMEMgM;dy&u$igJqT&<z+qQ)#9)M6j5FRq-xN!Xc`z
zs@|~Eh`;ysI<v|*liKfRnOv)=>fib&HE6W0^o#fn@u!m-1eXbJoV;5d+yne2N{qY6
zVYscntsOiOVz`68gB_$6hI8~eb}*^ObEAS*1KYyB2t2;P4hUo$Oj-!k2`5A#FQGs%
zncxMSOz=W5HX%MF0sK7z{;_}#Mv&#gN3QTOT=*C&e2f%4is4Rg_MW=^-1q<~EFye)
zWGEL269|!z=+G$eX9nUKy-<ueMVvHs8X1$MNXbY~=PclOxXC&@>B~yEQ5EH8D9yCC
zv9+7~i@k%Xyzv}uwxPz)-YTf5s%g&B($O=Z8POTD&6q4JbxUOv^Lg?Lx)v_(D+%+x
J9bHi`<iBjg&bR;o

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_conv_tensor_rearrange_interface.cpp.6D91D3DCA5B6D134.idx b/.cache/clangd/index/test_conv_tensor_rearrange_interface.cpp.6D91D3DCA5B6D134.idx
new file mode 100755
index 0000000000000000000000000000000000000000..00cc6564d64404e013626c0cfb4f5a0d6708dab3
GIT binary patch
literal 15844
zcmeI3c~}!y_s4U>2pSSZ1QIGDL`@)Uvd9*YAXG(mQ2`Yf1i@C(0a2`PT|jB2R>iIE
zf?81D2sXH+F2MztE`lg-xD-Wk3spg_ZL$0Y!_6FPe|;bSR-ebod~$BSxtTfl+<VU%
zhJ_9qbUlE>2~G~po|-aa0>$BQx-tJVXJkxX$$%*X*&C0<CB*jdHQgLo9_28*-<~N0
zExL|>a%x`}_0fc8ufCspddMEFy*Od|+*fng=bl*3yRuOB<BmfY$Ne(Z*RGFUrTHRH
zPid{{;SQ(tz@~ExH!l6_hpMV&QJRe@(uP~dK5A&_{h5E_p?;fxZ7Ov$d2v~^%uaHz
z{ESavfc*Ct-X_14R;eD`J?SIg9Jw@aeDc+MKX`7K>$~}DSC@)Or^g+e9r<Lo=O>+3
z?sd;ES-s6SD{0-O#=L7wK9R*ms#jT!=rQ=-id5tD{@;{c9&#_}jK9VDAC|WWo_Swh
zm6X0Saq5ET(XpL>XbPx2Id)*_&Jb}aok%y)xBY~j3T9M&`_=Y+r%ttnrN?Q^4lb$O
zec=AuN5|to-nQc0@$RQn7L<D>n+#2hn!fB)>5}xT+7n+y^$LBaj2YFieEq%QbB_5H
z)b#sZadrM3U)uY4dH$Wlg+Vj>`Ib4S)&_6Pl<fT0=Cji7>*qU}E-YJ8k=t<KNsaNz
zqFaXrKlga_S+}jho<o%vDra9@x5VJi7N3Nn9wqLo%l*Q7cD^{MY*m+~dymgqY`(tx
z$Ih~>>eges?ZQ6kbhcCOss-Ek^S_Jixy!7oNW9&2Ok*F*plF9ewPdjN=F1Uh<(fff
z&WxDxrT_R1evMyzIjl$W!Mhhds~=Za736GP?3Z_U>9xMgLrhAj>wAtIHTrM2)JXG(
zhL@9eesN;tL)pX{(Vg*IvV(8mPE+ssvrkd=2zlD_S9H<YuO;5K4@GG`f?PBA)+gKK
z=ZrS9cAo9p*<`@!H6?!2R{kP=Q6s1x;aoGqxh3yHr-VOzhPjPP{L=07LhCly>1WSP
z8sur8=$SU<+L;YL)1Qo7wdT@@h4&4*{GpjKJ9u<tf`v1PBOhixiW%6?`s9p^yz2O}
z^sj>hBgN~&PFR{gpS#jG#rT=a^%VoV%-g%t^w!87M}~eGI;6=_#Fafv%Q?Ak%$04P
zK0Pj<JvKK#=6s6Z^<G;3o?m0mZi_!zVV1FB%6?yu@LgeH%l_(JrP}LRyvl#YG?VND
zja~-nwn^TT=tWi=>|!s8dzKls`VQ~?<7SJLwBG5)wdqG6`i0h9-Pu39Akb(|?xq<o
zGe#|5|6TI4)h;hZ9uLOm<(XAn(H`wnQhOmm`!rT#zGT^wjaEfJcCobnxMG%eu`Jb4
zC4Uj}d68rH|76US)mrSAAC!O3_xpb6q<#Z(ni3}dp*rCIK+x8F{7HU{_NBvtwq|wG
zxZTS?pLcw*>9j?e*)PMkzq*{ar?lUon!y_;r~I-dwZgtr*`VvTR#CUSTAHLju1`A0
zt$%Q|xmdQitoVfIt3=~n7W=p4eiHj>fM|66wCD*bQ@YkV8^yM~2wJ@+Ye>ta`Ypet
zs{b%4?xeYK&|}S7{?OE$$pdTD8rwOk!cixl6t>*mH0FM?Dxv(^mC=8ubZdH{tY2h4
zclEG@j~xH=?cDNBSB4Ks;`jeAU$gC??bDFdk+pP=qf5cpvz~3c3Zg5!r)W-gFZ8}~
znA){BnO~OLyXSAo>aPwgczX7yCKqMk%%~^d3_SHDBlyeFw*D_C7lbZIJlx6rVZP}_
zOVa_hE2b4yOsU9xI(O(_MwYjNw{S9NPfhA$z~NZ^xYcNW7t<0;Zqh^!{X1}e2g8mR
zPJ0|<{>hpi3U^u_5gBviIUL&3(9-1fL1*M}oFZx#h3-9MZp2Mu%>St1Dhke9@>z+I
z5jT-(vJ?!Lf>*~2QSqZ2{14BGcx=H<n8cy&?ChLgABZ_<aDgV@%fs4|o1DU-e*(@=
zOgD8eq?_SAVlNzVzTRlejbjx0@cP)kJ`jauukKkr{0**|a8ntxdoOo~*9T&r+~?!o
zGo!Tkjkw877yt1KruwUj?Dikh{>!DBE_FHeV&3@Fe`zvjZ^=swW1A8q?pWrFHUVc7
z^GkFu<d@8!X=7hmy>7EHcLdYK!qCFx^?`IX=eIM@<33%{Y{VVIm~9MgB(D#|yf8QJ
z=1=2K9Q>C{wY1Rm?uL;E4gaOdn$_Fm{QBGcY{?xv?*HlL`i$;JhdbVvn{r1nSG0eZ
z{`}Vm(hqg(<JB*No|gWLLRH|}{L^!8s*Si4m|wCM4A+9Z<Atc4&np`4v?6c7zo<M;
zG~Z)nyL^igcRZuA5)4-|-lcmXDo-QU&1=3q_mMGoB-2$HPiFi2K+G>6-7Iiv94|KJ
zj%F0B46Te_ABZ_4Wm4u~PYwqT%v-eSvDs5Dblfm<FS&o?E<O`mtj2E0HlR8T<dal~
zXAKuTZ`-Y|nLu@nmjb_aZU5<hWQA$zcyrfD;iuv6ie#8F2HglFqps&Mp0`n|vkha=
zIW2Ky8;h%@EBh`#VoK44ASwhab-;0Pzy*^~iY^9WF^E`Jf>4QB1ws{OH3-$1H6YYr
zt^{Ev<|+_Xfz^iZ<Uu25#_gc!S`gNPh~+vE)?uy(VLj$X5H@1gf>4XO8HCN4TS3$c
zR`EIVnY}yb8dG!*SD3>U5$<KOz;|)ypKk%5UJTtAgDvq}yMq~N+WG!@KrsfB#n6oy
zNWi0bZv;9%?I3jMMZyf9b``qzB96tU2S*Q%Cj#W=aLL^wo<#09Y2ftW`SrCUFQhSj
z^AY>>A)hDllFR#Bl;ku1WV8vqiQ~CC%PmGE9kw>%)8~LX2d3nU`1E<8&SSm=)FsSU
zfVzVDI#Aa!-vsI==G#Es#(W2;JDBeSbszIXpdMoW6{ue^KLP3q=D&dY3-f#~HJ_{h
zJzn`$9G6S8Y34$W@EYMqE)<zH1iaU?1&NHzo#HuIpgYBLwP5d*!e23%2=M6;{}4oB
zwqcH}E!;L7QG{&-qDb3FMA5d<h+=GG5Dl{(hA7rHR!?MBdu#9p``@mOxcaD^`MGhT
zI2Y2RFs_TE#AV@`ZhU%ze8Ld&HlHp6$0G0{BbrZ_fUE?(S>6q@-I&WkR*w05kbRH2
z0%R4KPlD_u=4x=P1|Kq_`SeAQT?B8IuYv3u=35}Uh4~)H?qO~MSrg_KkhNfb1&*)4
zhm2@Goy(Qwa=i(Y8_f)t_GzE>f{E4>-N{9Y+|hs3?r%&My>m>Cx#~RRbje)<E}!-l
z`|8IipOy>dRz#fybY^OJGYtPdERSrS5<S3pvf7rSw*s{lx{{kj(c6LAj(G=AJ23A8
zY8U2Gph_|C0csECy+G~7ydS9jm=6MV5c6T64r4wF)KSdGfjW-)6i}xyp8@I&W?e)g
zPC<1<B$jm%iDg|xVp$iFSk^@(mUR({WnDyKSr?I5)<q<ibrFeWT|{E}8BoulE6Xo{
zdVyINv4|5=v@T+?ypT&R<aT8_k4xpT9PeQrZe8uT=B*i>e1LJKH>Y|snUKjNUbXLB
zz}lgYnIK{5X4#+2T8Za<X7_fkvt-_Ow{W*1>)ynMh4K3)O<TdZaiD3SHE~=%9c&t`
zr?FK*%FQc{UwrUR`u+CKamyCATsmqvh;gDmL4D#$gq;u`I?hGB^3M<c@t*q!i7LH#
zuXoYxl<O_@wqs+9kI)B^uh18foG-UzTPjQyAF&i-8iHtu=@3M5f;b`DGF}jmC_#{b
zXpCSCqC`O=qOpRph{g%VAu0r3A()eina}K(c*U5Nz*AyY0Z)Zl4Lmhw4e&IWD}h&u
zxe9nyU`{3)KC@rq)ncv#ULEFo;MHSp1YRR%E%3CMn}OGixfOV=U{2;4KC@rq<#6?!
zV!kiiChL!4W}=wz(S+XYrOxV=Wlt#W*F6J0|9tSAi4S8%WA$&7%_p<*=}69WY)Ht)
zmoy-`l(%NTbP-G(U%fHsotJ4O_41Y?9W!qB#aVw8ylu!HP;LF=`B&rHcgXgI;6Uzp
zJ3AMG3ugC1@WAX{$ga2B+xr(n0JaAe!a$abL0Sy<n4OBj8M9k4xMTJzX4l*8J9c?7
z_+z`G7=l<<f>eq1mEfer`buz9VtpleDY3p1<Vvis1cegot3ay4`YLczVSN?2sj$8Z
zyi{0U1#%VESAjx>_0=F%V|_I^sj<Eq+|*cK4PI)juLijq>#ISb#`+qNYOuZroHSTp
z18y3uuK_O&*4KbsgY`9_&|v*akXB;-N^q*g`jy~TiS;YNs}k#1g1i#zSAwDv>sNt&
z6*!OxLD5y<Tm>$e-K)R@vv(D{KWlIAUj+f!9#jPbS*`_XE!bmrss(4vZnfZ!*{hb_
zpS5Rr<h9_B?TT6mVz~~aby&X+oa(TC9k|tD{W|cf!}@g~ufzIvps2(8^&qXs`t{&c
zkM--ptsd*wgI7J)uLpTO)~^ReJ=SjoX(QHe1gA!<-w1AvSicdx8nJ#O$Q!YKBPbfN
zz80iftgi(pE!Nk9n-=S9!ApzvwIJ7GeJv=oSic#h%~-z~oSLzIGq^Ql{bulL#`?`5
zZ^rt~plHVWtzh2@4rE55=vHuU1sBZjt>A&#yOrI~x3~9ig#c_1YK4I;=WwMtTzkw;
zIb3JVZaG|c%w9QcexN-gD9_>gW4j`U8$>uG{QBb3yf$}chIn)Ho^{VC<&@O>%7^9d
zd%oPwK5^^LXH?DEE_^yn5N1YJL3}z~5RNE95P>LC5Q!*C5QQjO5RE8C5QAu#U>KrU
zL9CwUB<FkoGTHYVW@)DAp>SurkswNNB{U~<AAef@yuf!}L6M{~vv}*MN9h-DIwlR~
zCSR7<-dmW_-i;W=X!6BOE;S}ZPjgp>JyhGro?!;no1Z4ttxmi*s#;9Wwzq>L*=RD<
z&%uW!*WRv<YzEi0pQ{g{1zir5js7{bg;|LX?m0M$bcXCmm>hn4g^bOMF9X3ca3&dc
zKD`PAtH7DmkofdE5UgWn*N)0m?yzqLwA?kyWYiNyiTXq=sj<BkF}~Ok@x_48Lj&LX
z)n6Q(=Kq_+w2KDqDg5@DAKRlhSIYd3tPim-zP2@eS71>D(*vCO&iA!WEgw*q#C*}4
zS)#9o8Z2yD*5<>M1njKrY)OWkPY0L<=+l)<4#+N0ziQ{xL1sb77Hk)cY+=GM8*)_{
zhZlw;iV#L1iWEj7iV{X4iWWvAiV?;j8YUcuC{`G&r=^k0Una~dbN>*4-}3-WD%Nie
zxMl0=qz#i52bkd@3zEf?s8C=#>REL7ud6>e&U@}Sq_X}$v*`t2eZ4h$&MJey#aZ$k
z#9IM<U3#aLPt2C?HRE>_a!A1DcNB71))jJC))jJC))jJC))jJC))jJC))jJC))jJC
z))jJC))jJC))jJC))jJC))jJC))hin))hin))hKf))hJklexo=_xkY=GBLLoEu}>u
zBS{QCy$!_MpfAfMASq#!GJJX`h<9T9ZjkK8_A(HcVS719%CUVPi1%Up_aONm+Yf>G
z5Vlu<qypQIf%q7<p9INCY(EX+)7V}O(rS>AO)#IX0dWoVW%(jVF0z|?K7AR)m$CgC
zNUmY~4G`bJ_FEvih3yR>Zou|?Ai0O_4?z3?+nYergzb+({21F?K+=Nk&q4eg+h2k7
z70AdYichzJxDEQUoXeHua`l@+KD~%5Uc}Yg$#wXj71JH2ers}#CyG9V^7lNHD`!1e
zQc~@n^v)X_W%i*?(Nnj!7qhHPtSrgXUdoDYZo7mrZOq%Dy+S=moZ5YMzz;S&jok--
zd(ZuCk8_XCiObKgdS`$8F74*yuP9ma2M>DBJ!oI~s{9G6D>+Prv60&ZlBLi7>_4xU
zti4_Q)@FvReV94%&78^Kb13RTtFucL_pZFNYpPGq@JiTd|GYhQ<Du{&F^@^zc!cPQ
zJh}VUn)lGogz`r(U$rrx4(Sns@=x)iczyoqNSnXA)@}XM58nTtdw<1uH>!dQIE$F4
z6|)ps5v;Fm1IaeFn%cgwm+Zt`29h$&`#`b}^C6HN!h8%Q$1tA;$!W|rAgN)iW9<uV
z$z{wpKym|f14tS$KLE)C%#T6x81r+GJjdJyk~X&5(Z2ANEaK{`AC+e3zTJJ}@O36A
zj1!G>AeZIT;FL;JvypD^JfzQ9N|oQqDwCK&IZ8B2O!WBlXyIu6;}#^J$vh&dd(58f
z&f3$BM!$nUd8q7M-=Ex$X?R1K-g%3?^-nvJ$MD|@2^S7Fyl9+f$YI)#5sh&o-+tTo
zn+0ZPMvnXt1m5!?aCd0kec!d3y&oLmJ$D3BVPPWBn}3_bXNs;yKY=%SG>D{dnTG9u
zs6!S_e0rAkEd7$>W!H_`%q6>O-#J>6&7T49jX&{W>P=A&(Wu@O<r+;o(ptOmm0^|P
z?RWmJS4XQP>;1DgGO-bFFgY11Qzpz}0x&5;aY$i}{8qh1{|1TQ`ZvhsBuh*Vv5=Pf
zzsn;nIphL!GBYM-mbXs;8DtjVaHg=QpR%V}?CEs&bSir~nLSNoPt)1cne6E&>}e)@
zI*mP@p+7aCed`@xq^{$BW|A=@Gi&DLsZ&1vWP19vG|t4S$y3@t5;HZ{fa+vu)VYf>
zUtng^&9rN1bZAsW_aTEEy7MB;tt>1*>fs*{s0bPu9O5Q)8PM0y(e9t#74@{~CGOov
lVlQ)Yc6Img^!An8`dIhtKUgS|dWA*K;tn1DaX1X){2%Z!SIGbX

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_elementwise_layernorm_fp16.cpp.6E7042446B0FE0D3.idx b/.cache/clangd/index/test_elementwise_layernorm_fp16.cpp.6E7042446B0FE0D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..719f0893730c85e9d9c2bec76cb2fe93f96785db
GIT binary patch
literal 3444
zcmaJ@3pkWp7ye8PW`;2{hM8Q(Jq+`U7~&9)`?%yDmD?wGxfW3=6^&A%bdi%ww-ahY
zP0o*#bU~$XipnW+Eg}_B9sf6T+BMD7^MB7{J=R+9e)rmIzkA{7WN#lU1wpp3Q(9bj
zN{A=~K?LL{IVCBodkF-MV<0H4q~7;w(o++e;YHZDF51u4qC&jyh@xxeLvDN8LhG6x
zE<|;iU=fevZCm9vH2a*N-HKia4Vw$b&$`(`gse8ha09Wra>_{GNH^6!A^D3Y)oNeO
zn_F#vRa%*PW@c)-Z*Y*;s(lxfFPojGelvzWe#1grIsTSk(8l@c4|9oz8Cyy;D*94-
zBkT$?OYax-^+#H!XEwOo&sEJUPtPwjh@D9cab{z4bXHzt(`&U(oU*S^?Yh}gbU2ag
zQKZ)v?KyQdSH9jmO0V7{)p=R(Z}I1cTFU$X5#Q5(oTYnNoPGP9MCs~8wbJ=5PEX$L
zP0=;-;HnJojI8DiEV3sa$$O4k_oj6yy(^&xh8h$EY)Pmzw2WDTJyJRQ=bWQ?RLTW8
znq}sK#P!-+V~$J6r_VLYD#0tP=39C%IqGoh-<M|<8CLl?6ncg`or*BkEPGHMVg7i>
znGroGCunK&*$SyoW8qG*1OKhw&3rVelAm9pGt9x=FYYf4$@<K68)?_`5c@!gsEJt>
z_x{R?`fNE173b)kM18+kr#|OZ?H~v2JukZtBlmD5^_<u30OwJz+Ce8zon2~<-UU(z
z67u^Mnyb%s$4DPbCM)igkM`RAIqPoy0PAW4ru%)uCTYVVtyi)~6#K|VZAB*wCie7;
zG+jCVl2ta;z=<jA`LZpqp~XKQM_HT2l#z3&7~b6MG&cA9x){05w@;hINX68x6n$cU
z>4xG4(Y%!%cWfpN&MJAB&zjX)ysC5khrziRpQcN{x$ZCB$Gcqo=M6kngVyD$Zq+ke
zsMXTv-rE=J-w?^=#AP!Rr>_Sd3#?+V;$SNu<#e6iA1i4`SRcwZINw?AK$uQ|YsZf+
z?%9$WQ!;r&qs4~Gfg?wmdLDY$Huvdkx{T*U5FD=#KPe_0mCy~Ew!Pb)x}ttK=6u}^
z?k1C}6W8tTR#d*@ZhvV0`fbB#w6raGdtNZ<<P|egski)S=_h;%IWeA9qn&GYGNF0I
zO+8-QadjlSKBZ{v(rd2Gz36o2xK+1t!VJ-u8*bZF2>Wb*K=L8qeD;iET)rzPW3;h0
zO*M7x9p~K0lx*eEGJ=E6M5kX4Dg7Dt@f^2Ovh&iX$9~rowzbxl?roXcEs3AepJ7e@
zw$RmTxl!(}S9OJ1@_xMcX7e-59UUFaG7ae7)WH$iFP!bhnL~`19~WPpm5W`RWqSpg
z|3jh4_7ChYaVecxf4x6p@DHmp{E0h-v~zi<*Ur=oxI1-GcZVOFwr}lfbe%5!1MlCZ
zA2o=#Ty!3Y%GH?aaUHB^77=~&{uGp)9v3<-0zssjBi+035B;&Iz#s^g!N|z-4f6W=
zpR8=*6~o#WrQ)dd5Cmh<SY^HeX$|_=1JS10J6P0OB&~ziA@B`IYkJ8pZ>Q{y!J>Q-
zay$l4;v0}Qy#0Lid6!rb7TSQU)`#?E_y*$5ALKJCs^~NK>oXBY1tVz&#K7<kNHeYc
zpHYS!GO#Eggq(;WlKF<e6fxEdwPT4N%Ez%##Lvj>_RqYHuKA3ddkvCCqtTLl16JoY
z*Y!-w4;5okAqY7MLn84FNV99zYw+aNi&&H&k|tmXa(n~Q1)gUw&W>FT!=eHZa&fe{
zB;SDarB&|DHkrrOrJ(Si9jyLf<KYk0202)iH$pCrktXpCSlxIO_qU(iqd^(eIwVb&
zB;)u7q$h5gIi$J<XDvlVeGS197zy4u=YL7z8<II_2tuDP4`#N0Iv|3Bm2{L`04b0I
z3d5HHxPBZ=W6{<C5Qrimo+1D>4rX8(>fqLtnLz>GGuJPOp!l~2!m1~pJ5Ik*Hx?1)
z-!+f`5E4SkQ_*)yeY(C000B86AfA>$88}#(rR)QW17H;DUom~966IF`IuEKrKs;T6
z?fwNNiI!<My?9{9)UB2xylDSLs8pakqzazszE`2jQuPB>SZVcS7`-gph^*yp0QLo<
z{J+HkX`(SdND>Fr@pK|+6ynL!<pn${T~WYOrmG5g+H^YbT-r59e@D6dBlpgGaNy)#
z!>#wyy~k(2o)y&T?M7TScE~~ug@Y+l6cRWc;!!Cy;Q7wPdRStkxX4-!oQ6CI2YSb>
zpKk3j5E8)p`_@&Zg;?HBg6I&IQxHR~@RLwDG0n15vAyeygeXigrI-P3QJ9LON(p$f
zRI-4lO#N0Ig4Chv3gnEbCIX%<^;ZGUhw2MFAOe)g5Ayv{%&DW|;kTACdxeV5NUq0i
zWN)`Z!h)B-Tl4bjt@$$}XdEm}lV*Ttj)U<uJOgMO2g}f87~mP=U;>T60MiBs6KO;S
z&>s#a(MSw1sB)(56wNyKEk*7XErXT=8UkrVth^!iU94&>wQo-b@FO7JDEkglVX63o
zhU7Uf*)@4!(G-z_;5<R3ph2MnP`=y2k83txG!$ycA(epl9;`Qqg~DJ=BKz`YB+w!v
zZFRa95bUvdvBGI$ejRc!o?ZaxoXpB2+1N0O5aor}+39NrO7amjU6L*hNUHBywkczp
zRE5ysS$>r$;O0dBJ(VY_3k#vag(2nghVu6|DY2Bk^|HRUh#hjdcPCN+)(Fc4Ag=x4
zw3=0;Jf#Rq0`yG;!=K>*-MDCBII;ZrTcHA)Z{O%RYd=9kxCN(H+n$5Ek0n30A!y!{
z25+}rA#G4HO<D{ID^L`uV5`oT%iIH>Ki!0o=+JaDK!*m+igyt2`1%TM`LfvCD&ABj
z5D5#`J?#8s%<O1Woe&!Q?GcguK_l~l-W}W()ssRbp9qJ{V>#QVgu<VDO1_})qX&h;
zP*Qkoh#GQ5ph@5UKL9QNmPt+uOIAmaBCzA1z8Ti*0<BS@VNfWzUI+>k_%#lW-H5oW
z<GhxLilN0NB(XRtqBKDU@8sd+?n-jB*CAnCk?|t8RKemGt5w!Ew!bcyXIN^kq^tfc
tN>Ng#si|vdY3u6g|Hsg11=Gx2#Z=M2*g>AM%*4el66Ne~#}e^`{s)n3^;iG^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_fp8.cpp.AF44E860BEF4215C.idx b/.cache/clangd/index/test_fp8.cpp.AF44E860BEF4215C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a927a367158e191a351b94f7a52fcf13588e2fb0
GIT binary patch
literal 13386
zcmeI2c~}!y_s26~WHU%Wh!}`00m2?bQv?kOARwV&(IAQnvJ9ZC0-{h6DJY79iiAaq
zRNP;b;sVyHMMb4WD@dg-SoJMdL9Eicv}mjTW=JphLQ9|LSNZRa&to1w_he==cg{KY
zd(WKNQg3e$A&oYjkftX@EepqKG@1_iKXqBkk^?F<+ImeIExoc~&aK2-&dl%L>P24g
zukcstH%)$g?Z$aQr(TAh)z+=*vfaylZvK+BYTb#B!<AYApZ^|_`e146_y18hp&ts~
zYd`+XmI=k8m2c9U&H9C}r){%OO6-end7PQ)60nf-^|SB(+mxC1CTqv>QwCP!tXY@7
z+A%My;&#?szjb|gqg#3(ofCZ#Y;PGk+q|Xhf=%%4+oD$E#yh+zzwOS@>UZ3J?2MU1
zy!E{%!P&~r^Ch=^{mR2X>XvW(!ru1Lyqc&ZrHgmJN!#dLTf#c*xbV71RoLnidxEST
zzkjVdeXWyhepgl352762-v2N^T=u2B`Kpv%?Rdqmq;rf;dqn?TpYLCKY!j)sO<EFB
z<+#yetoKDL?U}s=t<4_3TOB9$R0(GWU#)q0a)oi7iRU!~En$H4Y{A@9xd%<l_yzX)
zv$Jp6w2fstMm9WIyyjp)$^#A7*0{WHinl*<TOiZV{+3_+R?7R-+PltI%4&_g{H53R
z>2gk#tfrdLTDJYtA-6YHUoX5Rf4F1P(g%B0qYvIJIkx;t&Pq>PZZ+Yvz0>w>$>Q*H
z0fK7h>PpYN49=^L!=)EpdesTvSgcxO!xFDNmCBUguJzg6e5kd@uR#+#JGP<LW0l+e
z=Gl7|gj`s2<J0ZA1?|i<P5I2OymI-D9eQg&cKvUhP`)pF>nrc-wLzP|vhUU0C^T4_
zf!*;)cV^qabhYV;yZxC~{_Es~pRS!Lt=7mtHt~6R<hC{St&aHt(@HN*+RwRQ<Xttf
zopqZ3(^)P4CEu%o4Hs*PxEuY4W><%XY^csOcX|1tz1BX#(VFmwN2<>@t0q0&kvpS=
zdtifGJ8$+Us~&lBcM+%5tJa&v_sR@s;q^v0V+FhJUa5OBk@x$d{4KGUuH!F~r(f)^
zh%RgTwk9wuh4pdDwBA({KQ7$ybc~BxQ17nVqR%we79U@3Q1!{BjpactKmO7mvCL^_
zchw0OjgUWVbGm!#%igQ@PfLEX<jon=&8cyz_PPCEX{;8f?$eo~my=N{<h+@)b^0@J
zzazb|=MUIySmGO5e`<~Pu4nDRl{Vui)O3j|?i_8XyqfjJmtz`!Jd;UqU&$@6M$gLZ
zXYwsB8-%1YqB2!ZHcr}k_tYLi>hlMUv7+pCzjpV%z4gp=dHl=0-Q43x7k#^7XV;4G
z(%W~Q%=7>4=z-m<KFI8v>?77(bo!C=o71YbV{YA@m0a_ARYBdoPTSf)5+&Zda`tw&
z{i1rY(dYE^*VhhC@QVrjh~3w-=VXuC$%{=-X{i|r5j}J=cwd~o`08e|KZ_nbpGN4@
z^znfY8Wr2*Kb7m1D)IaGcdOBZX@6s8C9QHO7oId`(dW*i5i4mc@qrI;w1wKS^I3D`
z+9RCY@~-D*He2O9iyjh6BN#LWKJWpKwrX}zmmV&~)#xGQwf|zaZz!2uuYWFLglkV0
zukL?7KD?Aok5b$UjnL54VE*|x&8dC<(~jOlS6TE03u(j@+7x`?16<qX@_jc`TIbji
zP9EptYqwb#vz$#|1WvA|sWt(eJSnQQ<GOv;3l@D&5RC|?1>*xB;N+<fqJLi)lKf<Z
zX(^c7=*ckNw{?W0w~GUE_#2KEtI_9@)AE-&uNQ66^!VBJzzCCi%+MLH36mMJ=`rA!
zYieq;z%M@=g6Uc9o-=ENX}NHC^1Yvro*wCmHSZqxOKDr*CN@0`oLpT~eH=LXH+TCA
ze1+H6s?isazx*$A-jSfk$!iQFAIaa&+HDs}dR<v{rW$<?dGuc<wZG_4`M!kK<`I5*
zN>p^}3On+b{ORu)eZbFwf791|cX^w`lwxw2no%&vgJ#9I!Hp<l1Q<1(BcDwS{Um*i
z!hN3|`*q*a`b`mvFEy>q*ztVkR~8EQ@+piyM0}!Er|<`^N{nH9qNG6)H11BvSEsEy
zVaD7pFFV=Dreh2O^(j4%orytJV{VMnRagu5urC;7eKhQbK|l{1o>=|r>6iJ5tH@z0
z4nbtHk9LYT^`BaYGl($TFbAp2gUN}_1x5@aToxV#DF?I7!5nU$YMOZXh4+32k&6Z9
zVsil-E+;-+7(I769UF4)p;zLFELf|jA85yMLL1lC0>|J46K5i%gX<uqi|Y=ODI-h!
z)tx?n@+y?$0av6CA@D%RFH!ylJdQquKvE;;f=ZXb3qUd(Hbr2n=!*M+9)!*#NM*xj
zoxeCiIB&wq<D@4rc}yE{(;%~98h@M7$djOV2r}8Qa|3T9Zxwa;q3-W>uJh}3fxW&4
zgJ9dRoz^eS#03v-?P3rdbIxQ)CUz#ykhtbt#KvRtjG)byZ#xZ=1K$B50bhWSkS`o0
z5H<h!?0~!d<FoTOxGDT2N6mwDSjvsS(;(-UbR$TV8$qJn2++*^@6WF-8@G0pqA=)b
ze1LfDlH*qY)~#`ZjkC4El;VUDZiJ8tZi0{*ZibL0Zi$dBZi|ou?tqX07a%0Wg$Rjo
zkv0e|oDkz;ggkH$ggkLignV!xg#2-TgaYsYgk-o3p&&d6p-?<j8$5TM2*bk=BFHI)
zwrD&Wp;$Z?p(H#Bp%gp?p)@=Vp>#YQp==D##<W2i!-*UW&p|m4!}CyHkKyZ4F2L{t
zl#4LD2xU2j%TX@F@G_JuFuVft+`O2g39W&3Oa>9k3AF~dPUd1}xDh16n6QNu!;C>}
zN$ey`XiH(IAe6>VLnxh{j!-tn$;Pa~L{RyinTr`g&clrJFk8s^7&9NW7h&uo%o5t=
z7+a2V8OAO{xdLNXKn|{InV~lR>SHRVb+wH_mLy|Zmx+*$t`0)Fy1IiDB7d|w@tZ{t
z=nR4@;|76XVh~<|UUPwShqk&OtSTP+t1?oN_)|tIC1`Ns{}-v@e$qt_X2riNT}kOm
z#JeT^$X$Dv+Mj-|IZX3KzsGQ2bHwc&DI9fpI`$wXf?X>q9MPRcd~pQN^9SD*q{>6L
zf@C^%(4Gp8CDW+`6CoXejw`h3zNagJWXZHdG@lRC#|UIm1`*4QMOK^wj9CCzg80Kx
zUQU*(w&tYVBvlR2T#Nx#1Bp^KkSJ9HiBdI?C{+WAQZ<k$RRf7qHIOJ(1Bp^KkSJ9H
ziBdI?C{+WAQZ<mMTnUL%HIOJ(1Bp^KkSJ9HiBdI?C{+WAQZ<k$RRf7qHIOJ(1Bp^K
zkQAy0vO?8BR;U`t3RMGHp=uy292~MDS3*{(8psM&131wDiwtY4%o!!UofHzA-;+1)
zAgL9M?2PzusprTz$igqS6Z4_VN&+Q=G5}O|$bdt4h@hfF4x@X*v<rI@wd{9x4<GJU
zi)!XgxuIMDlncN+E&%DqP0O1M&o`0UHAWVL)Linhz<gL^S()0vNZPPkGD=~TmC;`Q
z_<P#$a?rY;y^LJi3fSbbgO*xBl#x4gG&qKa%0{PYr@P7$#t&P`KrM7wycL=U>mS<^
zy`Q{oo`g38T!BGI-K0JsatnKoB+J|%Z6<#Xx~^@SxurjNMV+HE`jpZ4j?uU2#{6J2
zi&DlIhT_G=pQ{#Qz|X-F#uzBaL88_TNL29+(3UK<-Pc`X7n>`qI-m^kzh2cTGMWE8
z!0$<kGFd5;)jK9D`TgP+uIczba+@cz5lsQBB!iHcO9YTS>^xi`d6|15HYroefi{1>
zzX(zQKL8;aUxrW+KWLCDmfj0GG;>!SxuT=5Kq|9x+@mJD^!@7>Jm8fb`Jx5Qsfu@N
zz_@)@S!?`CAAvgg!b&g96l58)NGb}KLW&W^AQZ=o6GB@Krk8`6!eV7E=9-K7Le9rr
z^HDCutO~LBf&Ei)Pf^lp=WbGyX!EoOU+GjIJGkJ=7cb_K&nDC{R1B{cZV@hkL<k8t
zNHG>Mh%Js8hit5q_({{?H&gg22&M7U5K8B#57K_-)Y9Cskp{|aqs%t%m~9TOcv!Hn
z%`sw>b{w_dk2-o?yqA%0H`pw;85H~+CI_iCnHZRuLmQXLMQpZ)wmfKaFmymjU?@OH
zXedNT#1fgoG$&z7j39X!c$h=-VtOGqe?xyBv;`OjAS5%CArxd7G)T3V#VwOG8%>oV
zs0_h(48eMy#+)ses~?hU4cdhR8ia-|zzs*`8^Vv=0F^e-w=lHTMbX6qnkN)pETBtb
z=t>LvjL2Fuy7~V6N}O-#eCu+hdn?`h9o_r5cS-tZZr!8R5OpG4C!^lM;OekS6KIGd
zhtaMY(AM>jH7(FBH!RYcXGEQ6>IMl$X=wyacVwiYVcz1m&FQl{$qmEq2gEKZld1aw
zrH7$`4Lb3L9ow;|*s1h>lbB2>x;$Mzc;N%IEjbq>w-Via=K4cj*XAN+Z3F2b`QPqX
z0SgD|ifB!2=&EurF!bED^*`MvT@iJz`@Wq$F6E8&|9ZRFeR8Amj6l5t^1NtV_p0ud
z>A&!eAq&fK;y7=(g(}1(h1d+>>W_ca({Vn(@;P}lTGD<pm9Np3NTXs79e$cpsZ!}V
zRh!)1Wn+%nm;{`LYL&6b#T+5$VZuDj9dbTqk&oI7F=ipgf%YPdUxbMu%Q3zj<uZ(4
zhH?eQuYe3%RmsRj_u@l$=Kb20KcyDbbc|BE0)4u}CjMpYJoj(xTH45A^u^lC#~Koy
zz{7duhTTZthzs>QN5t`k#1(N7;#u*KUPZCKm<x5d#8vVEBrjJlgrwHe!LrC79hI+D
zTUEP8>8@!hQSsqq-!cuztijIu!2zJS1_waVVsHTD?!f^NLA2DA$W%U+%OgNSr_sU(
z|HO#P?kV$R7Sd-b;SrIv$Uy@wGJ%LsB5_uNOa&jKs-~_nmZ7CPPKT*24Um55Ki<dN
zdc3B;-b8&C+fd{xadY>W?kSkyAhwulYx;Mi2AoO8W?XYiD;qm|zT<nsDK67YoQ)>G
WKVyOc&*_7i(RANHFF$NH?f(E3rTF{+

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_add.hpp.EFC09082310C599D.idx b/.cache/clangd/index/test_gemm_add.hpp.EFC09082310C599D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8007f921e36b34f6ccb99b943326eeacea406cb2
GIT binary patch
literal 3774
zcmaJ@30M<%79WO~keLZ4gfkE_D2E|Y5eNwgBHzSVJSp)+JW#1irBcNM+U~ZBTC0}J
zgRQ!@YCWprfnaT|)Jp3CDplDPtw$@xqf|s!th;#h&EQMs%YNVTeLQ&o_kZ{Mz5k@A
z+H8kC34)uMTJTo({45DU5Gwf1pFb~WtQ$cjxD!OdrtRrBhu1|ZYm<MFO`ROKY`{t;
z;g6y%s|Wly_~Px{hKu2$U(8tX@@Q&c+iyiDmU^|<cC|OOwtd%fzbmwA)-uzcj_{nd
z$4law3)i+DjGcRLbI!QeNBbYnB$Kb?uc$a1-E;7>6X&ZADrZeB>07(4b0)QFaA7&Q
zzNPkXknFvGa(g~_y05b?Dfh=YhOlV^i)J;MsK)36mmk%5ANg+g%J{qhS>Y1S&fN-E
zJk1UoQW4NOpz~1WA8Yj2+sh}?DUsJ#OkBI9qQoO)+1^s`<#XStH*ft_Zpt>PXW2F7
zXNINLr7wQmJZk909p0PLit9`NRXuQ9O!@Fn=*N0}+lbhpqth1^ubiFN<<(H!9_TY{
zvf=ZD<OTbi-9tY6DyBI4PUgL{FB->QKU$QrvGZzW?u43~FV~&?Cf3(9PosIZJH55y
z60_lDp9j|#<{hw|+uPJ3Z+Kz&kjO85Yx-T1ASCR)E2p<aHPNIrfMES8f0aN859k2=
zNcv<Q|3_9sGfnmcaUd0_5(tXfpQ@YBtJYstl71l8%k_*vP~3Y*;V)%5o?SH2n;=*l
zVN(i(P;?s)HQfDCQv2jIO?rcvlX5D75b|zucvE<zZFgx7O?rTFZ^~OO5Hucn`}EfO
z*o6d5%0TQx`S=I~#Ytsf%+^WeR}@4D7#~Uur351Q4h$lOeEjI)oBN(kph+c&87U(d
z2>MRzd1O!fm+j>~qz@R^YP3dypz*Y6S?cD0zxbFYdw^I?snr5O@r2CnH=oXam`jst
zFiuG+4}qZZ%nNmMYWLZ$(<BXI52=S-ASlkt2^iP9z_)}ZDKH*D1*imq##g`Xwa;8*
zg_b~_SXxR`0zvKa^qu>k-2HL}O#+Bmg;XII2#Pl^ZVjK|Syx4qz95!KWpaU_xGG(i
z_0gwP1x+&G+n4h776|&TPCPd&GPK_yn&<`E|I^T~g1g?<Z0xh0COtvhkMdIs1htQk
zJk?QivgZ{U=?~gUsZu2nLhcSQCy0jXMfKNHsyE0;4Twb@lfax3#Ff_Z)mieIH<hFp
z7?03LL<t0qcYa2X9h0Q=q)7#cHI&9nASj-f{YL&ZC>YMWJ0;9HhQzSF?7cFu>kMns
zo1$>I5d)yY1Hh>gW)BmoSo9W$3aYwM0v=E$WaG`(I@)`i+$2Knu$l-HlI(^DpDp}I
zgc{%+BUnhP8zkL#2!^!<+w^D&Ap@6qyr5r(4YmhoVs;1|xWuEzc}Zv2O~Fh3e9z9c
zeSIheLLC@1P>TtfBP$iEpX-xcGQc3+$Zn9m3NAUk>D`!?EqQR8ZztYD>`AC@%sEow
zqPQ5$lVQ0*T#AU}I2&@q>%;EsCPz<jgY03cywWEyc4!ojSZ87pJ`oW(J`5X?7%>Vt
z1TxfuN3l~yYu7poW^eueT;Z5|PhKUB?HV)6E8=chVh@sG4GM!AO9WhiL5JM9N4aT(
z=$s><o9vP7xP%)5Lgm*3W&lZ>I5Vcnu+})M1v%UrFjM@>K$R)yQ}%9NL-WCj9ZW6Q
zlsE0zH!V;iT1ETg(#RT-I<;S~HC-;Zapgcb{JwFPp|`!aXvE|6agGt6S@zNLck~~e
z0ZZZm$sU<<Z1KSz)4?yLpwyV>+=4-kLG>4(OC+qpVu;7;NLV9dRERi@F;K+yGxkSr
ze#fPZ@rqr|U{{-@O~+O*xFgFx8DVaA$>-ht8YXY$`^j)D9=xU7-hXH0t{+ZG0N^@P
zAGFP|ak@AWXV#fToJD7GaLdA~UjJ+HquX+ZRcqBooF0bt)_NNed<^TO^)VvK7}i(o
zYs9_Euzp%UBkoCt_1F3vaV0TqfHuI0tzcM<R%67i#;}3fKqGFV4?Bhg#W$|jKrt8<
z<A-=!HOzhK{gRg!I5vbD445AF9^l5m7C1OCT~XFpoG##!t(=G(VogP^I8Ir4^taDf
zyh1nb>EcnPPbV(gAMH~6GB)FJ+KBdbF30#T!HDsXk~0b}o=Zee5JiGD3As<^^jOp3
zRa_1E@S}|?_=G8_?x*Y3IrKEzWWuDOr_pgD&KzTobIL5y7Lkm5jT??W0f8gAG!d81
zjTLd@xeO;~9c~>Z;?k^k5jWO4LBwTPCnL8@nXtZyp5gCOpFM3wg%f|-R1DjMi{$#_
zZKz13)xlMs8+0a^sts|`J@}U!4^~McXSy8Y(7|%!Uvej^Svxx!5rj|2|9yO_8~ffk
zdhOB_(oMh<PC)mzN3Q13+Lc*V%}bzN%5Y^Qx;tW9SFmsJ9Uh`$EFuOVIQ~NcZ!>I&
zJ;bqjkG)Kc{(ZoPhc1cL#-`8gF=T5Sc#?(5`r<%*d#bKsr|ow&9P^%VUeGjn;=eiY
z7%tI6eKGhm%S$6K%b)V=jDJ#a$JYlI+>#e~_JHFk|EK_*pa#v<w0?aJ&s`9K#-Nts
z0J;;$8aB2>jGlD!r-cwDr{hAg(HCx?$oo3E>5jXEjSh+Kg`OmAOh}A~iw%i&aQKKp
zV*KY2=E<;$%0xdja4WuIvDfKw<6KJEQZXiZj)k2B$KtO1`m-fB^LzJC;aI%YTZdh}
z8~^v(Y8NAI9(zm3gaf6Jibt476uwZnv;gY-cbGG-P<!YkoI8d!>&*i&%KuCn({LoV
z^1e$rU8=Hyxd;FDz~%T^)28Rr^U_-`$MdptvtppVIEVy##=nkFOvl0H!nsRtym!l=
l$e%YOzdt`3I-C#q2K{ztVC#KF>kD^nu#O0J3bf9D{{(P<5y}7n

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_add_fastgelu.cpp.9C90F62B27922BE9.idx b/.cache/clangd/index/test_gemm_add_fastgelu.cpp.9C90F62B27922BE9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..349660da6ab613341881743ac5f2b07db2659ffa
GIT binary patch
literal 3332
zcmaJ?3pkWp7yjmRpZUfZ<Icn|!<Z%!<Ca@Q#yyvENlv*<B6pGGaVtj=(Oswf<*!H`
zis*7wiV*&uatTp}h!jr8G1BS(W@_X3&-2go%zEbC?_O)Kz1F+-xU=o-77z%s31TNk
z1jp?Vh9HQDeq-ZecD)vYAY&W^B^Fn2iHPhol<NAX-5zs-xm##1qp;`d^Lq;ZIs0YW
zhuP{*1xoB!(+5(H9#iRd>8kjMRap1QWS}9rv<RX?h8@+m5@Q=p7&D^;8>ct(;~s`}
zssRHoRju!vd6=mKuD`Ss44p<V*OQ^qP_YLVPd@ym2mje!*t*u}-rUS6^=#^aQz2Yl
zc-$ka6Dh@QH+hqD(eqP09pyp2nG-|YPOr<SN=qx~Ys?)!haK-qaMrQO*vCEYnpF`t
z5nZC~cTE3MYQMU_fkUgOVu><)Q;%xo)y_vHeG9@ili}A5>c~$TcEPS!Oo*=c?yHj$
zZSBfb40P<j&`m@A7Xk|UGTl$BV>=#3H<^uec`4);ZN`|rFYa!9luiqVr!3ZO_j&Pf
z==EK~C(6DqC9;}Z;YL!?m{-?S!JvYl96LRrT+JcBDDYyk61!R>Ev@rdrrn{m$^AEO
z;X^UT3oQrbP4|VUW|U>G*&mfF-o>lPxya=(;^`R#uhbxw3sWu4)_o%5;;K<{|Hc?}
zg){mR$BNeHZaH|ty6JdtT593VU#%~@F&_FpRKN4DpX0}UR*#r(6^^m>wPvPUSM*F-
zbySKe8?yeWeVcEx^*Louv?l%;;mRZL_P9fx^pQQ2=EF~&Ukc@!)ye8>H@j#jG(^aj
zL?7f9xkbP3srDW(EOeQ0oT$}%vw);$B}U((pZ~%riS)Xw$J}$$F|b>-Ys?9`b@-?9
z-%pX4XUw$3a+}g?avE=TRE|7<s>ADh{Bcs;v-8Z|GNJBFNAE0ScfEj=Y`qj_(tJ5Z
zqqgl@n9n??w$i^mlKaLssw~R;WWi2l`@n?K!io>#wu-JhbBwRLp4%ib6*c_(tIYG2
z+rP5fXQQQelvH?Eg-sYS-I#TIU#+QN$l5gs7Y#;DmJ5f&@pPVA$cSlg<Fw(`etN{c
z!`*E>pO%Y3IiH8abycR~J33XKF#ald5F69qc6$2K{4MeR)Hmnsal3KOn->Np0<PC_
zv;A_?%NShD+>Si+i7hQP?k6r1MBay}jhnLT0zE=)Md}GIVhwR6PQGQn*+KhrF{*)k
zb0ebz&*mS=N_?oZw|skY{2|9PGvV{8rbiEdA#W8mtXuWluiFX9z0Ej-_XWyW{f{?>
zJ(R}-lT)7E-P?*M9W5=My){|0$zmEx#oet<#I>n+9mspUHgAi=&<6Qk^X2oU9tj;A
zl0NM6TUR+ecwyK<l^r+NHGug#6wlIslJxDge*KI4_1Ouj^xpo^Zv&Ef*e53y>&vSK
zGzS_L1M0kaMw8#THObd%PvWQd3a2hinP~=KT*a_=pxC5{odg&i_iWzJ__nSw9K3BO
z#F4^Ekpu=M1Rq$9l(}Jmasvl%MQJ=1PZAhFQ?3d*KW0t%<DlIT#L<GZqyz@yza{90
zx_azC!oSTv#lZn6O%<gQLv2KDh>*m0nyGQ*_ThD<{y5kZ4US+DlE8c?A&Axf+49PL
zWfL)Y2kI>$DxoMafHqg{O&N5`$iYLqP?`?Wu>u3mw&IE?e4JAc#lik)a2c!&QDA_#
zlNmE%DBaXR7XAsP<&@+M1P0JapOaFWi`(~z!GZr2JgwO0kE()=BRJR>^(J5mG6Dm<
z-`4DZ5vKFuJ`VOnX%dzsEiiz7cRIH*|C+lM4hlnq|4-H4H8in%2hTZ3!P`-9iZ(@8
zV1V~*7Qyd+c>6F8_CjeQmPiyBK=T5mE@t{l9>BrgC@qbZCJGFo|Ee17tfHQ8$H7}r
zS`sTMD=>hDVuC|rzlosmhb~~OKo6i72uGQvYypY~)PZi|OJix10!v|W6W%1ae?X-5
z4h9zV2UM1FcF=z6qXr8jET{me2~ZRae^=6v#DoRDpyfbU=(&7Rf#ea6mZp{-hy>Jy
zZsJR8DLHwT{9<xo72U)a*is}qj=l<{YQ(6(6W_F$uSgYS^W$Q`*HZIkvOc#Y65Y)g
zCD;Y@h@Qk3BRCl0sIpYeKtXMDXQsohFL(<f9I7r=AKVi~7Ss*PkPX!iAjcNkXc_G4
zjX`gZ&L7AEL=?gbbiH&O1y9ie3PU&)TS@>x9wU$2l04tf{<tRSg!i`6!MWYj;V_y7
zi)sYUML0A=nlV68x_*~h_iQ^2WB%6UzqcIJ6YVX(G9cCObpjERW2=$KoK(vd<pUFk
zmXAMfOZQP?DOrNcX3?rPr{vANlR!8K8DRniARIgy&jeycIFe*ZCYS?+Lm(5FAUeX4
zB1<uW_7M(|Ok{$2Lpai8X(kvighL{em|)Nmjtp6b30iz4O8UA{^~N5wi!@!D0jLp*
zENM2&kS&c3P(A~C_H4X%C)!1MIe7zcp=0isZ_JL}4&>L~glYzMqmfK$<^cVYHli6Q
z*-#~pZnCF3f@KsrQ{9#!cdFMiw3X_+glLX5mu1L}=CKTU(Y%)-Uz$Ha<r5i6K5vWj
z_}MY2Yr!&_9fP)rZukFr*1?o(uwwjc?T#m2w;JtP(dPQ5hStk=A4q5&mP_P;bJJ9H
zBjFXgETJKx1G?XQFrU#aG1t68u^n73SHHo&1avo5l}ZJ9fo7o+;y=kC2*O##T4fF@
z&OPaAxyM+|Z$)0<QTe~m5*Q>5R`6hh1Au_gbNR0|SVlN1ES1H{=lOZI)*4hyt~eLS
z4fW+e<x9R~7J1S43$b8*c;$glD|{DT7Jn6O7adx$43v+C<UjvQAr)DQR-g>%|3LT-
zb#y`j4T0eWtypAYV*+E>^Jh5}3^at{_$Xv}eyJ!%goA=ZLZRSA1`1mG8M5><Fd`^C
z5`~*1JcNWrM8(7<a0p&nk|;%ByRluJN%nTyB&@THqO6>fvYEN1)dp*uje7Dr7Bth<
wlz&=PAg@wY*Pv?B8O$|n*XbJ=nXFecR?%f`l2_0&baV-a9eiw^Fz(QQ0VkH*9RL6T

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_add_relu.cpp.6A11991389EEE204.idx b/.cache/clangd/index/test_gemm_add_relu.cpp.6A11991389EEE204.idx
new file mode 100755
index 0000000000000000000000000000000000000000..81a4022ce4dfcb0bccffdff2831b07494913afea
GIT binary patch
literal 3290
zcmaJ?3p7+~8{V@?%xuP(aW@!dBSs8i+)^f!anEh!(#6nrE#Y)rPRi+Yl<<e}JEl@{
z%c4{g|5D+UTneQkb-E~VN=MG=_E*QhXF4CUTK`^a-ZjsD-uJsa-}8Rkm(SzfL?DP8
z!B2^gOx`RFK@bK1B_$_D4@p6g3<-i#L`{LCTSqyHw`Xa>k+Kz|GShpgqqU7KY9q8%
zno%N?QsrvOA9<azGpiu3%WcPCtEnPKWGU)w@y&tspu)%=b3)%txCk>ln#KBY^3BHX
zml=uQ*}FH4R~*ouo^(B8maaM8d)6uK@|9FFrz+^a(~?IScK`Wr8u!4dUHt{m#yTBJ
zS*l|c0)Hm5(WAJzu=+#iSeQkN{!_;*B`;nCHtr+&g@<ws{$2Ruf`#TexmzCbhm+Yh
zk&#;N$_6IqG;aSZ<kx7^qJ5R)Csk5@{5)CgR#iMQ<!$bNC&KrbqmP5gWx9`sJ}Et4
zQsY?fLn_S-UuLYjCN@`o_HI=3#e;U_jswj@?Ecb5l*IFOCo2T-V&ipDehm2>xv7xX
zY(6{v_)kOGBUU^|X=9p4dbfSiVDeKFAy<!>pgEaJfGq>-@_3sf^DVt&GDF$D4kImI
z(J?)+&-=>P>2^ObblO<?EK#%7yQ{5-vyWFqxs8;xe>h`rfgId57SZ6*<G(vy)Y6a{
zUC~@wOF!4MhsV$Lv2ZW4+dX^S>hzXC3x;j9?f$;;?`r(Ig593i=pE$y!_E|@UjdSy
zmi<8UQvK;ptGymAM1^*fVSk2`Zc@C_Ez>|YLm~FLu5f4#ebeBvFlr&~=^>`gORvYV
zb<4a6=<VE4Ql)loaJ@-R<<;E>lBQ)kt_+qZzbgCbqF^ex@Ipw#xsj-0^Ez&MytMsc
zp}KQ0rXE~F@7-QqL;s)^d?b3%U{}d-1zo$>ZJnU$k<1A3&dMDBYXa49R~^2>Z0)+F
z7)yQ4;^i7KaibGuoT$h3eWrEUs~Sls?quEE&pUrDh2o``ZC=}7zm_tP5Yc-jCxhG8
z8Y_C)$#g!6Y|M{&ZgV2%RL%F@rl|X5SEuZWQQxUvSr6$VvkBesO1od~-Lz&*aT*TG
z-@Se%)uZCXpDD>J)f6ULt7G!G%Ck9-P`jydk0{?TQTp#r3B_G4?3NDQzY{w?TYayw
zotrOQZ?FP#>f0Xiz_r`Ujde2Za-DH@ZHO>7MNVfawr8`*jjnp;iNh}*rI`1zvj)v^
zu`xQHoQFa|RzP*zMWKo4R%q+H8^@3TokfzOdS<1cpFCYrHg#-$*8TF0KHe_9Z#5IS
zYp4wE{Z*Mq_kC8rQ(|WJ(N*c`cqwIr()Ez?%;T3c)VnU#sRu`Xv^Vl{j9+_I@RyBI
z%^^djH%{G4%R{tESow!u9<}{r*S#5=LG9MI?$jCj&Q0u+WS2ozqsg2Odf9+WkK4PS
zAK9P3f40CbP``~hv5ZSq*(v44c~g{>8XrDQfFRk19cQ0xtJq6|gEm8$5>bgF33&HQ
z9_p=}T#UN+HxlfR(+ClvN&=v@4p{|0$WGi%f(1A&Pn4%h0-&vH<n|QnZpc)GH{dh_
zWw0ax(7rF4^S8PPI+Wp!IIW_gVlD}Q4jOdxI2-V46A2E*X+@$URT2OlMsSraX|LL_
z6b^?V%m^|fN&+Olz3^vi(VYj*sYgGNplF;ngbWoW0ni2QnHBdFykkjF%(v+3Wdn%i
zz~L4W9DvVFAyO!k@MUNSx|_QFr{8{fqP7&?gjdVS$SF$#poa(fs}d+aLJ}N|SCffk
znj`@Fp6Es|q4B^32@b(&Dv_!r34ngkKAI-D!RC_SFr20lX%tBS^qcm#)`MOdp(HpE
zrxl0_%8~%+&&zkF(WJ*~<)Nr=$&whkC24la{JFzK$$5hfAsC%a=YWGz<YovS=b2{0
z1c|gjari3)EP{_D_Q(84s%%v|FcL5do+Nh3e3BYlZ7zvX_Rw|umA(~0T5?q21eqoi
z1JW1YUc+mdGEDJ@N%A0=p`M{BNCH+P5G57^6WAEn%TpfpI4&B}uK%50Ib6jfe7zMX
zgQ0ZGv18ltY_Xq!lrFh5uVWfM1xX<o)0}Au)`O#!%+(7Jm&pUjy$5}M67~)v;H$)~
z2i6SakRnQ)4pxa^40(n=*aD7LGXwznJ^S@~s{iEMe|7+T8x&gk@O|v-I2d1r71IXH
zi(sfV$^i&;5U!lKededrWNXf;K6k2Lv;a%}A^~s!kQs)*+RxWIr6)BZ+p>2#Es}39
zdS>0uyqR%11ViWu3-|@W<mvJ(U<?FPpewL|w-Jm?C$m5sA($dvkp(t`U=%uq1sW8=
zl;}z<aEl=rl}=@W{z5PsoyG#~l#`&;VAHf_7(YeS99;o4!jTioU4UFrK0pP&)u-N0
zGz-Nd+cF)%YrMn`bp!~U$p4k*NaiyGp+e^wS$rf{raP#`ktfq<0rF)AEI<Ng@H|4@
zQLhEa2lZQk0??oZC>RX`2<U`cKzxhMTR@AgHP`2LFZQIae=N!q@6(uR25Rwr8l!Wl
z-g?Lx`Ix7@sAn$lic%U@Js)0l=$E=Wd#iabsdxj3TM4-8vS-(H(^c{a{#s5)&KNAG
z`{#qkcjex9FEV-mn4x#b!_gghHdBYm1nUA-_@2a9FBlu2iDaY!><r9<Cy6g$P=jC^
zYz=$RhrlU#lK2{)Pts;<+kzxuAUsKY8P6x_uyy7n94qm6+RD)lStJ86G+rydspo6e
z+3F5pf)647y%$#0Pc5qbNXfU4Ys~m#@oVebdDM}<eq0n)6;*RE$fqIEXPFx^PZni^
zJM5oR16o0ek#U>#a0Cuy?vn<V2mv9J5`{_j;<_z!#RN=lb3pBT*?~4`VMOE>C~}U0
zBIf@J7yONnh}(w4b@6^u(o1BP%F2-tc_jskBAM^Q_x7Z^@|IDFo-}o36%9=XN2gV*
zow;k+s>@fRb|#E(y;Y-Y>*(n-4UCLgre+qFD{O4-bvat*R%=z&46WV0Vqv%SE*=D5
F=zm~uvm^ij

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_add_relu_add_layernorm_fp16.cpp.E9706C96D2E415C1.idx b/.cache/clangd/index/test_gemm_add_relu_add_layernorm_fp16.cpp.E9706C96D2E415C1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e92621d55ca400c37aaf208add2dcfb12c02069
GIT binary patch
literal 5208
zcmaJ^2|QG58z0A@nRBwuSjI9kvYWZaHb|*2=A08*MoE?sk}gNlBGqm6U00<n;TCmU
zZrUhC=_+40mDDY%Zn;-0Dutp|Nqz53z2@Y*zwi8ho}cIaKhOJrmuGoLMoOhYTn1xu
zJeiS_urQX%U@&yRU;4tdq(M~%BN$~cGD>SgQ$jl2wBLR{`IozG;XQG~j}vucG;6U(
zm+)q3c-z*C`SwlgUf*7lFs>;t&hTPQbn?@s)dyFc;H+Yvd3^G?jY^qKbI!;2m$f-1
zk;|{_UDEMtThihQ#$`P!JED6MU!7Jx_P!-`*_oz(exL08o4b>l&#i8}dQp68;L{VA
zZTDAhOXzt2=CAD44t2k6PkgnlzoDmXD8;Qlq0_bfS^F-F${W{13yb)cnQsbjE%p^n
zUYpswB`~XEUiX5Xwh{Txd%Q(?BBNgu4s5u#agS=JR`Jbw53g0MoHV&NxzgoiSxzd>
z^4j%yqSwp*9RCM;y!~m{YRB|;-16vdAN=W#XW=F8f>$OcGyaI9qOyVsp5B>)t4mWH
z95TxN858ge_TPJkE&uTPk+aX`@SWqC^YZ7e(H);KI{N+EBaeh+pS_+BD?crH;bEz)
zmxzf8&%f5?c%<|N**2fM>h9s>)&si}kBj%^gq_v6ogGtlKd8raS;+pO@dH6`W?WEb
zd-G<_t($22aC})-5@Vxl_a!&JXIkUjgi&@2OZS@&XnCN4UEC>_C3>fPXSj$1Pv4rd
zw3)X$GOtLx;FlnK{W~3hy^Cr1Z7*Na<J4$p#PmLMJhxwEy5Z%#=SPfwy;E`eN0XD;
znP>8>-<{i97{8ml?V+2`+|pQY%W0(&H{sq)muF$o;f+};){WW@=YDV5wfse($_ioI
z#)rLgPP!~kh{?L@gOAg4D{U-|Y3cOvnCqoqZEn=;)|ii5)ICT=2HSmpIdCg-x@oN-
z@@z=9P_N_9I>Uy>ae*r(@yz0%r&ko+GZrP^epz95YE*!pX1CLUR`=^AQP1M*maN;g
zK~j~|x_?udFKd2!=c{g8@3{I8;;S#KII%k`!klWVjP-4bBI}ZiLbtFJ8<>UB>CX=`
zlaWO>?*+FmEv??z5hpu&th!&6yRD$3Frtb3v@65Wq_6f#^7iu%pYqP6r?qzeQ5&)%
zoLTAJ7qBO*CTnl)P_rJ@cqcX|LQie_^t}tNFdN@BY@OB8GJEFw1#t<v3vY;yYaBhr
zFT;GItZfa_pB>uw;N9;nF(slx(r-GNx53)5&s_iBfcov@tfsJR;o|y;H$5I|ch(r3
zsNKVRcVuO`Fp~4)L}p9@da$5r{<8b=tr22g?~&k5Q&U-2qPvd!f9x;m_o?-vt_m}2
zBHpZtZX4xYeXRRlW&YrWv0BNUmTm4uFSiRjZ(ng+lsdn>Jhf_O=ii^p7H>Hh*p;;P
zQ}myI=iN;Y;%ByfD*r*_^?FJA{)|j)O63{V)Xz0XQ=^*i73<(x?MWBR4jw9A-{Icz
z^ZG3-LmD+U-H-Jxj?Z4{Fp!#E=NRl&x8vC6k7!XjN5V*7nljr8)BtAG8qY%at&u2_
zFpoj;*?b*^098WqjkWK(#;kvP5k=-O7?iKNuZ}`M*b=!OvY{=n1x027Z#}jiS0Ny5
zH&B^Z9B^P0ip&JCA=^-2As{RcjQwD@^y1)XMk<)xpW)9|2!^~)fzP#ba6Ym8Pq4Hw
z;BCw{)=`KrAcNt%@n!$K%8y|v5(~Up>MWK*KsaEOalyJfCA}yT3B0*%E>|HS94xY|
zI}-HW3>1k0us&PgNFg8`+0sDOS4!`qNCcQ%LtTTV5D<=P*<{~ixN9-Wm=9n-hM%@V
zz?COhnHV`ZmR>^{3;qdSK2*JC?N3%iC^8*z8?lXa6ar4Z>gc+%KSHOSMUiL#o3c%H
z6#~MWQzI->6BA=lWEy~V*g71AfN;L7{P5eZ6LBaq8*sDL*(`;C@b=Y@tl~5qj-p5+
zfYsI2Sqd=>KS^4XXKP=JB2i#+6Sj%2LO^&&;+2w%-VSe3Bp$##Hjkqa5Z*cPbj}ka
zKWz;p2~4i3u4%3i5I(Trrck@*^$ir60lanDx*UaoaIL|`yBy?$rv~yp@cy50XkixT
z?wkC!4n^XCw*lM0Kp~*_!}%7et#{h1w2`?09>W@gDFlRj&U@gCrpBx?LejquQ79Xg
zcSFTD^vgjjtGdS3C=v?1Ic$!eLO?hzVNSXPQ)Y`QTahJjzZfO56*<BcVAOco_?b{M
z-*%tqdv4*X!c>F^j2xI)2>Y9a!YDtHp9eI8=>?4Bxuv9Yk-0=d0;|Yc+0C$6nF>ZZ
z+BycqVVHZsSDtuEUx`dI3mU-;R6)LE-q-!Y%mdq$r=8N*-qxPBO;4HU@HQoO6{B??
zTpB-Oo8DG8mpOOrOa;U83x#|zcg@+~<Sa%>wWPZ6Xn@9*8qhSp)SRYSNUdm^jntl|
z38juS%}MG?(<D+)rG{(aG(Mb*)A(>cPUFKZa2g+OgVXqMAx`7Nop2f-F2QMhgcd>L
zBe(>OkKhwDKEi^a@ewuzjgJr#G(N(Kpz#qBg2qQ`ku*M%OVapAK1t&vEl3(4X+zTZ
zNFhn%Bb`VZA1NV+`TQ<fJ~@B{hJjF&*F=~-uoQz(l-FTpI62Flhr<{aWMCw(#!4es
znd@vAuOW}lU0FKq>D#YS2J7s!0A1^@JTEl_qXeS`9B2(Rp1=%hp)ZqzCZI{Rz;_}x
z(Etv^O0UNLrov+sJ_+|x1`6Rtcq=v1i==Cl$YD#Ex(M>k5qul|v#w4$w7me8Z-c&+
zSbs8FBq$ghF${P4f2UlKF`?Py()A%GlM=WKJmBt_lm%unnx>7hFsErw7S2#x^h6`!
zg0-Y|1o^V`8Z0$ur8gL+CopVbphC#|m-4pkZ0%?j!b9dUTp?gv11|EOrgRb73dL|0
z8097N3WwUd-77Pv?mOSe#3(y6I~zERQTAr`G)-tGq-i2E(Xf_fc{F%M?#suT7^Nf7
z5yRGrQMv+MF>FT|#Sw7CumfNeSHKm+ki{rHfu0!FVT{rj=!;>bVw8cvKn&XwMi~kW
z#jxsP6i>ht!*+>LMgk)-JTr_k78r|Rjm0Pvfr)tg*S?EUrUFwjJSvRh3;1IAMEkMN
z*TkcFgAvFsmNrWt*3k7+RM+)x`s@yZpWe|x@Q)SP!z0TL*gP<E$zi7vw+jyhB%S;r
zp}Z4EnPUX>%^MT`I>`U|mJz=2p#c&{{>p&K4yra!YNAXPoQKmif=AFal1C0}KT~Q2
zeY&~TV0ZGIfYZVEiXQu$cM1pyCp%ZTLJ$y6_B2iCEOb+1A}0|IlX^*MsVki%9Zp@a
za;cP5u3YLT4WMa3(h#MFd*QTMa4Amn!~JlYA0C9${D?^ejgOELG(W<Rp!pF&1kI0>
zk~B-wkEB_WLFBL{j3aQN^2eL<LdVI*4W|_hOJF2_!YPg1W$web4o88J{2>SL1Ec&=
ze=GP@+o$bWv=)ss91$8Yvw*Msl?SK7D6vdDOmMJ!?b4YKOLIYTOIcD)+S#&jnnth)
znntq7VXflIgeE?_-h2dkm1}!rPUfWZU*9xvsa13Dy;#q*iyvWKlfC=Zn{-1>L`6~G
zp!Jvc5Y%dp73RbqZCwLmQ`1e;6YdAXpfJm8BRrxnl@d-|E8UV&Iq`|SuF?zW>CawH
zJ>sqXBbHO|B8yk^-PH}4)GgJ8F#7+~*l@XA+QR|Ebf%_%Z??P58fXHO$onm96qk%M
z+F2QzYGByX(2@r;p}E64FyemOVlauXneSM5;IM#!%aT8;;pQ;PP39I2eOn%%P5sUL
z<`Y#WWo>RPfWu76#@vRc*_zu9YtTl8t$2vTdD}gTSL<F34IQ!Zj^b(Fl>3+nFibzD
zM^pap@+hgV9%0=*>d+>{^T=kf6%9)bN7%#RBVUE}O(}ernZC8XTY!#;i1!iGF4wwL
z^p2n$)X5u{T=m0?5yKyEmJX?3Rc{|L4BG}cJ^8N&xJ!(3k~u}eBVeQ@B*zLsd_b51
zLkC9IFp{1Ymu>+pRj9U{9r4xpm3c^V3L^!^7K4$j{5LKoJ{kOSW(22%s4~^mHAb;e
zY_zVXj<yyk34v2Ngw&S9p33Fv=^Gh)jrX27>0gt5#0H`XmY(+J|1@f7$~O~OjJ2|{
ja}a*#<m@8x7&pe<#L*QuFtm2_3rItzg!}rd$Qb_xJ9*Ij

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_add_silu.cpp.0C3DE18B907F5A14.idx b/.cache/clangd/index/test_gemm_add_silu.cpp.0C3DE18B907F5A14.idx
new file mode 100755
index 0000000000000000000000000000000000000000..2b0530567e9416dd4f94550fbcb6e3d0ececef4c
GIT binary patch
literal 3330
zcmaJ?2{@E%8~%LunQx34!)O@mH^wr`SYi;>poV0ZwT$(%r3^*tbgI8ikt7|5{6$%l
zBqaY|4jnB{Du+}3PNzEBEKw0@InFmzFOIA0pX-|En)`j8_u1b2dB$7d;_?B3AbzAE
zEio!ZC<8$d1^<dtl4JX%A&5$ZptPa}|GxEoHp;!zv<O~dLL=$1JiF1Sv{}98fIY#I
zm#g4&D*x1Q+Lg=v(BXg^YBhcEj&@!<ugK(Q1s&KdwhB>c<=kEo_~DcAy50}JZ*?sr
z_YbwiCNHsT8I~G7^z=bVld#cdx>`PD?fCZ*nN#iN5c$Z!yL-VwTR)q-tNmL$=0+M<
z`#2>B)qZgxa-;RyE-lAL?EZ}9e|T^GPjf77%*I-_qd8(HuX(>)y>p@2t)4CcUF4J3
zQFpX0GRwd8L;8U<KLIPevbjp8IBff+)43*{!y|Nc*Ld&BJ6eWXxAoJ)GpTQ!`iMQn
z{uR4dEvr*CraQ%Fk*2M&%?`(GiEZ^wJzS5<I$5HNw;Nf>&ow?V%sID53)7BcXRQzW
zYvkeHNoD)$tSt(8hCfW?HW`-=T&H*JaiTqL>>ksUDLdnBn$_B^7gO8=H;3n^pBL41
z`{v1>cunk|OnJY`VN0M+piPp=NLK3c%#u=!@Zi`_Qzc8xk(}+%BFlcstb`_QEJCGz
zbr~i2@UT4jtX#&)uzcfGXqq1%)_x{(d2)9$wcu8jXL)`@PM4Rn&d~~)f9b6(Ll;bp
z(^{;*TN8dDo=}jis=1SX&WQir(&+zrmTBPEG=J+_cC%uCP-_XlMc2USx{+Z0{nbxQ
zs*P4l$@XT25KlYU#a~!xozIY7xTsY_fxaTEySbM7!N0-8J-jI-zP|Tdl>gqPg$lOE
zBaGUrTzxKh^74h{6`E6;LB+9m(JOAlca;t~S5{OrHXKMGX=u9`KdmNlPaQ({PWxBz
z8VU|uxoMEB8f)IAYF>ElADw?R{lrVSa{OdUNxrFh@!li*^i~b}^n1xzY7T@uoPOk4
zJHcRoG~0j9drz&PcZy`vaj&Q4X@2hR@lv5o_p6EtXBmT?-7iv<>%5unQL%+Nn`mpg
z_6$9Lt5@dL9k_K9S6%7lrSiD!<*L*BAM<QqJzs@+hZdy|I&R3nR=_Q2)A?&t+h?oI
zjE(Mxg<hy7<aldiWT=~f&01X**V<{GSs5G=pC+}P+$G<h#!4;f%pJq*g~g`)n5@zJ
zUA_(s(dBDr$<_WDqT9dOls2a%9d~yAtWB3Q_z{^*I3jeje#rk-t=ZYJyg1XL_lfC{
zNyOc)9_#m~aIVrz&NOibPSFF_UA{D^FUT2UuSoOj8Qsaqgg^RePFHqddACM&9N%5Y
zK3F@Dl;acVfevmjCNFKVS-4Km=)slbKQ;%AlkaNTUg=7IE4?j@dn6_F?mffUthQ~1
zcTV3VOlCfoZ?r$R)7D@AzsfHI@iewcFBW}R0g2KR!^toN(eejVeEBRg5%vp*FbaV}
zkqCSXNItrX8}AV1$QeY~7gr+$L`5Qi8Z~)&N9OXxbt1eDf-qDLrNr>%_!LO-ivgP2
zJ65j}XhDSHaWxxaD@z0<UNrc}$eLpJ<JwK?L^u?8P9xAL67fX|L8j_c)WPU6d3jif
z-&T}URFw#z=6yKkiUk;5CBi}YZ8CvOlL(-8is2`oaIM`Z4@cm)4Ix8<M1a_?!nf?G
z=)>|vH~_y*B~U360dD)(U;4N&QV>Uk*WhXrfutf4KpoiYRA2q(Y&;S6$JI&%B~^(4
z>T5~at!*zhoL7azaJ3pk%}gSIx`&!^co(|EjtGa~YGs15ibMeQpB4|xLVo{#oCw8z
zt7{_y)A+HF&3i~t?6(RVX$eYV>I#dM;kCGPU6w9KB7pNrlU;|FJ3hQfgoAOl3PFV`
z5kL(kM@5UKWoLN-6C`{9Z-HQ%TumELA7BS=B&No!5uHo7oiQqWwDR?czLk`W<ncju
zfI9eHF>gQu2*%c9o6Z=0B@hgMA<kDT8Oc+EMgSV&p2QrQ%|e^2J(DUp#f`*Fnhj@2
zG-Lp-EIOEYea9Bl0x4J$032Hz*N`ZWZ!Vcv0U#jmQp`RO0l^qth8?Kh(V!>8qW@$P
zepk#vkTY<N7bZRi>3lKD6Rq2__$Nz|bJp88ac4gTNFx|(hFXA*;m849HV5%h7k~h*
z@tnk*p3RBLWzO`%?I!C*73>ipg*O|YNf6xU+0|R=z5^3q%LKy<Mgeyzo`YGJ>Rk1i
zM-B)%{pfxCn*<n7&I+{#=^z+uG0O%ZFg3U*@yN`2TEJak4;q#sYCw`Qw>!>j*lyN&
zuWi|HpDQ95!bCWLc?d>gk~knm1XE%vaX=UZBQwby@Ja}#%v9!p(L*o_lfnVZ55cHR
zDhKoq!Bm(k955&dMq|=AfMW+YP|sR7INisegk{FE1eM{)k;R{boLK^Zvd!vJ6Q3(L
z&LdT>cX{o^%ccl%$9Skcu*N;vvX%e@m;g)tgOLBEXYr0_1`$f_=M`{ID^WLa7f0@>
z=N#mX`prRW(4bkwa$~KYgFIP2bC4e^U=9jmg#uLe^TuDI6^5SUJff@7&A<y*5A8}1
zd{UGxuFM!+1g!DOj9D}2qRY{=vrAvxpZCmHuC;fTy9}$~F6YigtNXRr4IShM__3n4
zqA_3&h$u~v>{oD%V7eq-eXs~|v`lvmKsWZ}7~fNTe`6lWZ}+o3f*<v5!ULe%C<<~0
zbio-bUemJyG`Jcw+y~3*|1PQiW4?~-ZFGX?)eH8)S8{++1ldXvB+DDb1)Rjw5pQmA
zj9^+^EgopbRIqR7Ma!C3^CEqu9JWiS`)+jJ@u!}m&vIRIf6hC`{|6F;dU!e@1_U-1
zc#0W~C^<r8Cr&&DWDh}6Ge4!Upqp}uIcc41Wg?@ZW1z?x4HP~5Ga@lEAqj_Gi9XUY
zvU2hYibRA&RiY@91)c&Aca@bcMk)k%+5%NIhKBtT$EC}b^PRZ#g$^uR6Wwo}sx!5;
o_4H8#Lt~EVB6AB%Yo48sji#B^3c5ObvD@kd*fr33m6SL1H$X|M0RR91

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_splitk.cpp.921AD8C68523E94F.idx b/.cache/clangd/index/test_gemm_splitk.cpp.921AD8C68523E94F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dab26e89a6e4e378d195b21d5cf184e9f29929a3
GIT binary patch
literal 3696
zcmbVMdpy(YAOBjjnQd<65;eDtY?y1Cnn`S$)C!er^>v9b_sX?SO6Vf@lITaMqtfMs
zsGLN8EG3c3aY{+<qEpiO)j7Yf)idkVUuUn^KCkcdem~Fqd7k&@^Ld<kTyCHo1ljrV
zVg&xtFdBj&MdTk99l4!>3~~t&6mzm_qi2{7V?oiWhsW50b^>|(PE`tx64JsAkr8go
zoo+QNGB;Nt#RtV}y?gv22D4JjRp)A^!b~sw)b%t-2g>#rPA<5nd?&WZI&orBrhYnQ
zEGRnS<L86ka&>1s5A!wNZM{)uqNb@SFkQ>9Rx40ooKCr=mUy%7Q|y|QQ#O@PX8sDj
zW!%gd^!72C9^cm1{G3k0?D|kM_A-xo9sT^z&l|(D7nC=ZhO-9N-p>-Yb(mGwdN7BT
zV4;x7Uf)t4XjxI?NzV4Du*mc{yX?-qG5H$TNS|8oPuTs!6He5t!Kv5#Wk|Ip_N=V9
zw4P4f4yyUOge0N#Ll57^H-z!qs?LLZW8UU1du?i7{i<PA&~QVsaOgbIDK)ES-f;2G
zheLTr?1Y(5xTJt19V%F=Q|+{~A&1+UOFvA^Jf2jrnl8J_w7<&ZcBz~Dg5GQ<r})}H
z(O9o@V2}{CZl-LiJ$%NV9p|p&&I}8FabyQgk>k*j6Ip;6ygQH!|MZEbALXy-B|WXO
ztunZ;XNa>s2zw#iXuY#;POo!WciXQW>c=B!u~ccEbY}1C9cMe0s&ekH=V)c0b(Y~|
zc=617vs}iW&2W-e*OzA)Y--SsX;`|UMZU#LLl}McT+E5e_WQQYKd0pp4#|xl&hT?w
z%ZqHRMAuiJc<EGJT9+AI-1U1XZAbFuF17GD)Uy?Ci-^Z25_2DTKdQ9MJlF49P<z0t
zW*NGNdnwDRNZV<|&<<bKZ&it|(miE&H>$9@bqi&uU8AqN!AM1(l><9DUtU-FZ56?^
z^yre5J&(>xSLJ6Cg=XqOA-}a0Fo*gsv>9DI2wj#f?H+wtMigWo(&@EVfSolxkLvvA
zMN2bd)4RnZRbKsmntF<Q!c>E_jk_E7A2N7vTwCU6@Hti=HLT+LKQq*P+W~xC>bjrP
z%d1QC9OH|+?jP6@6_@p3$uMiUsHyOHxAlfUSgupPXnqG=klj&QNB+>Yr|<l#NZ-ih
zy`P>-oqF60-y+(l`?q?Pl?HS&`fscZWQN?!Z!W1Q*&ny`GB;;3K4H<Wth6qkHch+4
z+%e|j;Hcw^r<%s;{q$cyzfF1{(D<j)`>>A-uogKJ@1<Sa3u~J7Ml_U@19q?YgAu~o
z9k{gNX1;^GU*sAUo1}&eV%XF0tNX0fL(RidmRqFpPP`Z&rnJWA)+$#&kBAr&&`cB$
zmK9!ZbA3{;$J_jfchKA*oZ^xI2h|;^i0Mm=pKiHc{7a8{!{WN*I(r@mm-e){biU(c
z;cfXpt(UWoJ{m|k9M;_)wRpF6T`7O%mGa@!<@JV!-rR!RW9j1~8>%lpOFb5IWLeJf
zzXry9Ez~ElYvaPuVcP=aCpW$s?~9LpHX=ysTzc@tn0Z{`@Rn{&M7Qy22uCEZDNB}%
z?n!)4e;E2pUEIyHvFyL4-sBxhf}&ytJ|QRw(u&1g-|mr{hlfHS2)2f-6+{7=eP@s#
znKtwQT|eD|N4X=kCB{-w6uyuUWU{4ACZ~L_5+3qLXg0)F5CyQB9ZzS)A5TujqBbM6
zE=E^f6u@e`hTNAAxcg}sROnX-R>CUH2edt#n%-2RwFi$14})M;Syeev05s>w+o_<+
zPi}bBCPW;E!Qn&!(A@p*L#B~VO|#;zurgM8z9;~iZ_rG43)|y~hXRq+Hjs^iD1eiD
zgmqR<xhGu1qP!5Aj-jiI0$A<U(HL4^$^9LV+KA9Z3=t;^fQD&j?yR`iGL46}A@p~)
zq{xiO$fg9c@K6v!vmlm&D1g1wf?~R!cz1L0(DonES<YuJP7l@i;!$o0EsK%Gi2~R=
ze=xRDLt!c(4+S9fcNZ(#JD}$)*LVew@<3<;hJX_Vu=?gMtF`})iZH~Yyb*daX0f^`
z0QxrXbj_C3{0CT+Cqf%w4Aeyd&{IBRubNKn>%&9A2>o3#Mfz`x+9f?#b5f`;wf1$n
zcER&ilM}j@Xtc-<U;|n-2}q+wc7nP?i!KPR42o4+_M0Llz$|^18IU8u*80|FK#l;j
z_1R`XjsV-}+n514fFj47O(D1>gL;9e*eA69*T)21Atn3cZoqxh1%z+PmH?CJq~#z>
z1ei=GF9&)Am`bNE2Pq=JCUleKppHQ=A$!g?5@-(4LulD=dWQfrlo@KkPGqFT(3Kd`
z7{*`(gb>}?76Trw7E%T)P<%Hu;yqH{07kq=9+<h2q^*eeND~4Mv?t=dB<wQKbT8I*
zc>1gNppm|m2zbiAlnCg-FC_x>fFH<tXPX*q2l^SQoY`guFygNhnJ?k51JzN&Uk4LM
zpoh5aWWom{u=hVICvedRcJ@M}jfg8GRT4!ai>hRb#4f5-io`CeCKR9x?u+P3+L|)S
zS!C6ISY3($;}mczAgQ3@|1lf9A4tw;ACI}aGL@MUIcI7yC34QBF~vCt{qc|7fR_dF
zaQ0c5(<M2O9Ov#zaw75PMq~%Fqr}afsC;nqbEErW-Cvwc#~}@{P?<C*r);QfIyX{P
zQPlv&jo4~S;!23ylDQJ%wp1<<N77|$!UeV>1&(9Lfs!V`OdL}@0%}OX%)W+T<9Z8c
zKWR?MC}KH*LNEe%2l~kD*~bygZat7eeT<uF{nusN-mlO1eE!=`WIIbuOENg>sLc7J
zsAZQj5SU0Int&<>j~kL8$$o<({X<|^DF_0dKpInRw<sP)95JwJCB7l5^J9$|5S-2)
zO_<!YLJTyyU0M{l>E#gc4$>c)AonMKq=|tmZJk2i4I7pisEWUFYx9SQaWU|w;wUS-
zaDWK<L3|)F0H%rokOnaToCX9$s!H(i3W`k>11!73BEjIHd@)dhwdcMvpI0gdF1*%N
z9Ca$r09ugCE?v67c;bN=NPO+VGp7}A76WHHU9_K^d=)JQ`UOu7yz(!^M@9NZIn3tU
z4<sFew$6R}h(D#^sQ~}*?9By9(tcaE`uady#1!N!@hR{N2}6eK1upZ@(ip4^4o@H|
z$tfzx^VaeBYv!-w8qAknvtXgJin<2N+Gd6AN;`W~Rbw_~8BO~~sM=zZrj9OIZwb}V
hh;C+X!LVAc#av`!>7c5nzjXCl0m^X$$4SZ=`ag9Y65#*<

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_splitk_ut_cases.inc.CD038407CD6C080D.idx b/.cache/clangd/index/test_gemm_splitk_ut_cases.inc.CD038407CD6C080D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6b4b30f5325025d1f1c1be0ed9d2a8d44c8b6e04
GIT binary patch
literal 7386
zcmd^EX;c(f7JgNhG~Hb_-7ICR2#rfb)1qd;xG~_`0gX$PpteSBQAdp+izh-L7(~RV
z#3-UN(aE5K8<L0+blecPF^H(3Be({LprDasT%%@gb@lD8>YVxIPqeK%@71mM?(*I5
z-d9ctF&Zy-B4qgNkhq9Bi_Hun#0CDvE{>V+LlN?0TSDTtWM-U*Jn7A@h~MD(nO9BG
zp3iS4=j^Gu<?GWj!QI*A;+QLiOY|MKZXC;X{-C*{wsNhx>1;(q=JDUEvUAhUWIT>y
z&w84w9u=9J&fQ(%IsE6;!oc4~eNJs7`%{Nq{eGUuyEEtA{`+T>|J&u#?TN8r^W07i
z-98~>(!$e6j5F7F%nsV_vU0hx^WDmo=L?O^JN>(?ySjNy*T06OXZd$|?`n_qkMc~J
z&12p*MPC?abPF1ocj?@cTi<F5zVzI)I^*j}#%E{DyPZu1o;?G5c`PdB?yNBU>-FxE
z89Av3(o~*@e5yZg{`b~<>4jCn4d*6syOTGM=>E!MboZq49%;woi_a$~KDo4dZtbD_
zfAkNE^`7|p>A2ao>Z(mSfo}$7J#u~W`|hP}y0~@iOk(3BW{n5usfXp|otw67BS+d1
z!iSPjmE}S>@Tr?S#Zw(7#{Of~_4-ndbOV6`iWE>P>&-6Y(Q>}z9h2`!1r4!q6rv$4
z1Xl7YN;FMd5dVdOW<e;DM5-(o!l6L-_9=_M^9(!75xrahb@93Xuv=Z~3xk4IfPk5p
zRhEl=g@EB({~6J+I!i%={`?M#>;PrOn_Vb)nEFI)<%?nktp%Y(lBlv=>;fU$hL^8?
z*QavC69f#zpt4--=8iM23;m|Y@qrxa0s=;2R9P-|fr(4H_KE2I@S=hS7xSwqvI_JN
zZ}x2_bt*1C8kDHvh&Kp1D2pafr}sB?PycE+M?64i6Gb+`O5)AFUBvc`ro=~+hA3!_
zTp#r87k5n|eunJn3c3vlNTQ4aSuRDh0=IM-$`LOR7(fQ7EEoG0+Y}>=^?q%3DCoa{
z08MC><zlyH$IKo6MQaA{<;XiApe0(B<zinU$8T)!6Yjb!j_5$Z!6#_$uF8p>uWhMc
zte~Ah$cuQXEEoGW`}%I|cQNXiH%HooKnX=kz#rnxzSX|FU0s?!%>G6}t3gO&4VqhU
zzVfG#BS8TQS_J}<2%|tz!jY2c&!QhFXe9`gQ=}ZW5pQ-~PEG!y)4WNwjSAWogd`^8
zD%HbBCzda%;W*L}1SEJxb8F};c`1$EMsfrw&2bwPx~?6cmw)d0GLCcrA&E~=pe~}i
z=d`SS2?`p>&(SBS%ddOZ24rmgg;UVZAmjif6nfe#(YK~-t(GI5K}Zs1-0F?#$)_>p
zWlD~81p!I!pg_!=xv@?_C2bz2ynp?0Sp-E}LyalnRJCG8`qtWMG-U}4hKiPrkd@$V
z0j#jN<qvm1d5s~u!z-8;=mQCbYl&Nbt0}JtriCD2SlrCZ4|M4DVhrpgI1Qmi%SQb7
ze#8F8=jFk>Vg@3Gw5$HGvg7rzZ{dNE;Fw~%v?&ipoFDHG(;^%YDae!XKqN_28o7~n
ze85-*`@M>GKdJHvM{__>unym6ETg<rN7BlG@%0I#{r^~=e*aJU>Ko>U8fITn!vg{E
z=vVJ}r^EA2nHw3#;v39T#$pum7-NYf#uoKLi;$LKWvO0B#-Of%Y`eO`sH!gLgMfuO
zhQ2e-J8#HMc7^SQ7$B4xOIR_=j1?lt5R4VWnCpIDUs;<`@t6Yzq@nQlRHj5Me{U&F
zOXG{~gs_W&um!+jjFAQdtR=!~17jVhAJ#pXcJ0~5z=S>l37b<_GAyrEYxRgPmN%-6
zdISc`C#n<mSX)@wOI<+eaqpn+RSAA|*^9J{C95#c9D^eFk-$q>ix^8(k)|EPCnSm8
zn?eAPdPMYK2pt10vasFnA}wIeK|f%jas-BBoEU3LAgc}YsWojb-FF)zA((=sOUp*X
zW}dp(I3lD5qA1chvYGLadqm^&QB&c8s1`C9`WWIS5Re^s!f0a4nWO6tMK46Z!P~;#
zC^TwkyUM;L`(DGeXgku<=$BkwUfh~`cpyj3xSnVFOkT8fDj--M54o2eJtvX_dIA#-
zRu>SA+TiA3iM2VfAy6DAB|lx!l>8zWtpZO9M%p|e8-<Q_^=Vue&{oUxjFQn*-B25@
z^?Ngx<z1DonvX*~M$VWv;uDtlR(flY8d=^~>8rswvHT$AAPo{U%LgliHP>tJ^{-Bv
zvY6$=lwq2<{EYQ6rH|@aK3o~D!6aw-C}ore8H(kTlt~)YnB~_{$~BY*En)dAN|{Az
zE<DNcDSp4h5|+=Ul)01!bB*N-DP<w0!7OI^Qc77$X^uyY-L>~#y9+E|K`ARJ&944V
zG3oKE^UyBvjFkK3l>h+@?w4n}X;Ec<%ap`}kl6;cL5~P!d9&KAM}lJkf9g$?9`VNl
z%+=+T9+@B01y&bPfZ?5f-#ur+-k+EK2~%dB_Ryu;T%nFhx!P>ld3s3sf1bV-S<Y_v
zQCnU5v)Qn=fKW67X$qhwMW9F|806@<z=}>kU49KL1%GvjAclA4U3KVbhWF;Zb%-E_
z_vL+ch#-a^#1GOTf*3xS57y!03?Iga=@3B-AI^vC5J3zd#YgE7K@6Y7C+QGD3?K-8
zIz$iy2!fvu5ySw3;HN_bF@PZW=@3B-AP9asL=XcAf}ai%6m#@tZ1v-t{lQ_<+*&+m
z!<$FF<BDNgU{~y*QGdXjbXqQy5L?79TD3H&Rk1yK2oNUi*~C(hZo?pg`{ewK42YDN
z=6v5-l{PN`azSWA?FCav7<(BM8WMYP6O$XW{i_9WRo6bbYJ7_&(&H<do-S{D0YD({
z_7H4fhrDe>hQa1y?b&e+_v&7FJEY_K?oN@RPueRTK!=9Gv|uL^NmfJrNe^vH)?3ve
zJi$SH1drm{qJJ^fw>J)L-aGmV0R;phd<WY`m=;(T-^IoOrsef3x6%1#_T^8*9hSkA
z1Fy;5>F8ctuP(NBxS!ZWBK06$z-rRdfc$KLJtSyiJtbNf!v-1LA$7IbXCZ^-PY<Fq
z1CDjHy^YNYRtk7q+Fpn~0<z%&`gxBVhNUCnZE1swbpzfNy<j&_9HX0xsuCv~w*MgX
zBvP(9Xv~buAJTBGRk6#_B?HHvjg9UFm6}%5dIToR4^R)#V=P%dQXQ$sa}*1ysop{9
zF%^%d-1()QVs<sSQ`(MVT>u!#8&E_H#{Br%lVv-N&2G?NNfQVT1(K9tEuqGULR%qe
zK0#SYQp52<(h9ON(u4*?a-;&YB<DdmXh^zV)*<tRW<~f$WY<96@dyVK;S-+yxjZd|
z=0^C~fzL}rwinM>nROBT?xpt9V`753>QFsWzMzD96{Sb=X2Fx{5=xJr!xIxAUg|6H
z^n-oH=^Hz1=9i5c0pXFR^BL32TL!;{I+&Ip1F-%>+~f}&1(Dl8P|9g?hQm$-grseg
zyotg?73d<3Pj$uMZ8u%Ye}rji4~oq{Ov`<Wr(k#>@B+FT8zoRddd!s{deNoeN_pV$
zFotlh+Y}lUIly~rt73_J{`uqX28<G#Hff4u=MGMjX9r^)BbHibHP@f|F69msS|O5H
z_8gDfLIVE3uO-;2fC^HUcri7!?4bX?2dyd`zq_m(yl{WgudT}HS5;eHJ6&+9ZB<5N
zHUR3S)hTAqLbDeB`v<?E5&Rxzo4`*Qwh26_+9t4g;8~sEk=8bW&6sThJ1pA-5&(7)
z1bYPAgsn-i_89~_44VWt8CVesRy*4SmP6YFR!iFivUhAucr5(a1bjkY@HF>!^!Q`p
Gg!~UT&K>9g

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_gemm_splitk_util.hpp.D5E2B28B24FEB85E.idx b/.cache/clangd/index/test_gemm_splitk_util.hpp.D5E2B28B24FEB85E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..6168bf6662508bf8f4890d5b3b8bbb92ea017ab2
GIT binary patch
literal 2342
zcmZ8i3s4hR6kQe~WV0X;LRdBlk&hvO3dmS~qI^w2fl_1!T0p+yh=6<sDM(dN@rNR)
z&?-gis70ZOqKpbE)Z)Jut)loD8I+Wd1w^#yNZ+n=lbP&H?!39@-23jk@7$LNWl)fW
zi(&rB%5`f}auf703^PH$oZReHn{_bk405bH*b=!c^QqMM(9hleg->@jw>2D5JdPYW
zyfEeb%HR98?IOudV~wZkr`wvW%8MHK+@w#EOIx?D?Ccy^^q!_{I3xeuFD@(p?9<UH
zDkwdk88gwoVRRy7x3@vMEV(eSR@VRD_M+(vTZBGWZy67^oH?AQ=e75R$z;_l{esIo
z?%oV@4|hI<-_Kky5K1hh?+yqD|Mq+$P3)>|x^>~N#}b$EZF{?YFS_*Z+_Jv^)Xeh8
zF!`R&&K0pG^@poZH*}voNS?c$vihCL8t05Om0#S{#Wz1Z%Sb#nZ=`X@n7Fjm@AXml
zi0(v>6|{c3D$Vejdw%2elBo)tFyD;)yylwc$1YXAm7Wt#mM!D%udn#-2+e=AZ<TBH
zhM~nRc5x1mir(FvQtfYz?U>K4|Ge|I(_rr%CwJnQ(YHxa9mB3S28(xXi_W^9`Efz~
z!`|+uERPWDkpuP$hxk{x&aaEgI77RWOZh$F+xwNC8D<lMb=PvM1nwDv2i;dI_O2HT
zKIzU`dae)4DOi)}jAIxv-sNvAJYk6Avy9Otp^q{pOpP#ffy~G69_{P8nHV9!C8ii<
zOWE471~d2qOO<)|Qg206^Y9r677-#7)&MNeGMG5JNH30u69^U(LPOR7JlA_v+3VY9
zDZs4}Y-4FNi!}hd8BKj(Gx4yMhwCC(LP+?m0eG=EY17ux;`uz>0KsBHY{D9V(?sum
zDycb9`nZt^MhRvJ%vghQail>^htf%_sS11YP2^=B&Oz~3gcYAP5MTb2N80d}W%dGm
zCW2>M&X%$U;C;twm08e+904vxu$`rygf#$H>z=MrwyI)`aUOylxsD`j0B*A9#;WR4
zb9uNvf~AC%&l-SRh|J==jfW<AI2XYdgoPn%0Pea%i`Kp^Dl)>Y5G)hOtXKo^pNf~Y
zbFVqY8sj8_-MQ|jtO2-Z?S}%}tOX~HaYF>lxpE`c0Q{z=Zs<i~&1xR5hhUnZO;`hP
zc1mha5Sm~R22QYofNqnNGC&y)6@WYtiP46Jw8U$PACN#Kh{PyFLlT>dDY#^i3L-H&
z(U24YiV(;If<YukEgF)amtO#oAR{6%n$eIHUJ7+hk)ftCRfRW4hood&*-U77)OdQh
z4{uegjvlq#U^yRB#@+QjRa?oM)N-#VDd@=~V#84qKWjfvU?(h9#?Q5`{HVo-LUagL
zr!beGQdBM@QDBjjzn8!IOnC$PDSybm>d{KM*{=EP_2uKQIV45Xv;*8MNr`BY1KbTs
zNoa`!w2q{#Xe$TUl%%Az)B(g=9=hj4bpCIyTD!oYpbD8O(Nq``5E2ZBu8_t|gfD%w
zPb(It2*olZq=^j<2v$HW%o6GhGe|(H$>3CunwFnM?PV_G%Y%erbuYNGI{)Yxw3A#W
zw+FSKF!tPE!b`HyDxl$x&kYIDK1<6DM@Drq!wFlGRG?R&y6dOYp8vZ;&mmbW7KRta
zGLx%`4e$!^fitz7soI{<+_MQ4FOUfwfz&qb?K<r8<Cs>;cDIORPT-Lt#1_hhE-*15
zOI?DUe@H@7a*=#C9Gs+F&0VF?L)afDDL2wh0B(YPAs_*SqJ}fi1GxJ<4s0~N9u=ik
z+^gUxDv#>+vs!LY8)^gdh(HpO3i1k4&sla#dO|9?Ii>$_#G^TKxeja%bj+u4o_AsO
z<}$S-Ci8ydSaeaV+R>Kp5!#ir)=BN?Q2ug0zO3dR^cVvqCp#&pgwY8NV!<$xW+~7t
t%`{6(%~Gu~^e6d%qyN&e-+4JfTZ+VdGjoB^0yEc8#hRs1v$W7G{|CAs<L&?e

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_bwd_data.cpp.BA5BCB136A7B5CF1.idx b/.cache/clangd/index/test_grouped_convnd_bwd_data.cpp.BA5BCB136A7B5CF1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b99e72e84ad670963a902ae1b3f8744be53dc049
GIT binary patch
literal 5266
zcmaJ^3pkYN7yo7&WM+sNlxwbC#;6Q)zYl|9%ow-Om1!cOqAcC<FO{NJ*ks!-n~HR!
z+g94zjU`fOyXet{NTe3qrrP>{Gxg4VZM!qi<MsW%^PTrS?>WD7-t&e^{QS09K#<P@
zNk+12c`OEkAanRHZF%Y<HyD_kKu|{6>BxS07sITDpY}s!TS()w3jwPiMVyIou&m!w
zutO;)ROWv_eBj_;cOEh%Q+FEVzm`?Rv~2CAzg$LKkNMeR14Rl|Ut~?ot6#E;oHblM
z@!qxOhQ8b4be4ld%8c6Dz4dL|a9w=~>vEQ5oVb!<?0Not!Pecwd2K(GEPpfFE2dq#
zU>`rZt>KUD6QW<Qp){Azocvc;$A$iZ?o!5O!fHOdA5&OEp_UgZvzbh}ZM9&xXLEb2
zq$D+rUUzbKEX}7<v1YDIb56K(QE8>)QrzD!s6|a4)z|IKY!+jW+&t!#;&05EA>N#^
zx1X9QpWH68+8dSMC~l2WcD=NF(K<7<WJktGF{@|aj57-rlEl6%`$*4@nsob0UStjK
zOK*La-+X!X=xfXU@l~ycMm84GRSm0l+*&ZS>y#70;M|6hi|8x^Tt%+=-Kg%S!YId?
zO%(acW8+H&(Gd1B?SrxD#0LNAi(O56lOJ_Poh2xiTLgR17cH$F46Z{B9WNsnHSF3l
zVA2$7H2?lHRdRNY>|F7bBj$+<kAyDnd6;0*w5v!OAPbebR;xCxc&=F7XBUCNZVnWX
zvlj$@`($HT++TImZ3jk6Mdh=_xsR+nygm7jLCI6v?2pHaj<(zoUOBvZ<wkLCx-1}c
zd)T)RPPk`{iwrA6Csx>=w+iCgrH9V64%&vwHF{+7>$O{{4TF{c`%rOTNIcKFQ4^gI
zHKS6wcMisVm(oDwum$absSliOobB?1^17GDPu<4}@1AU||4GpvVV{<jWSBP5pVU)1
z&a=E=O71#?_Okm;zgZnD*l5#nW{76W%5V?5UP=-jMH{c2!@AXKWqo|VTlS{Hw>L9)
zLMHy6-_O4J&C!{c>m+=cb9dt_>&=mCeQvzjy0YX!=&9ll8PidH#o_N*s6s)PJH_{K
zxId9RhuG?H=*qowDW>B>ZQSOR|B`d1bN%n!h+#(L9|ahP!WXo7zryFOtzjRtsyQCp
zrds+9@r_o;q;;UqnjNHOoe8zN5SmAxw@Esqg=aGA5ar&~X=a+y<~nu7ibRv+kN>)o
zaH6Cq^iWX}Lw@7fJ$$`)!QdJTQh)A3!%56TU26t<H`m4m5myf#_>({Mpf*Fc%{SF{
z^}7CP&${<Kn$8}nwtBzZ$+>ek`s%7#JH@v?WWF2jIen{R>;co*5!BUNilOzV3Y>df
z4=hui4tnB~xwHJ^OnLg%!Brv2i<5_i^Ss{vl9$T2WEa-h8(LC-FHZlazo+QdnIn}o
z-DM7_19mGX42JEteK>7zg!KK{zr)`C=-O4YK4JQc^{)mtH0}Jc?k=kFrc3_OPSn+0
zzsII;)3&ag@$H$sQw8x2aT2Hck0JMrSp9vORm6uE)&_P~kDl+KAFT^&5BS)z2mO{_
zaX@ZulWeo!t@Qnvk-D=IC@nKNE*K3#20U8g;R~Z8BlIi?Qkfc>I%o#Co@jmq`x{A(
z&C(<sv=kQ0A-S1mfYe0?{#p4A<H>2+ZaDNzSWGdb5H-UmAp|kT<tyS#w$<U#k@F#y
z1=fO~8Gx8|`8M_1ijST+C<WG*K{7MV04Dce;_%+CmltA+j)S$W$X0aC0JO7DOs#is
z|MQp;dN!;rHWWK*1|Z(D%V@s4QCkiU9Sv&}umqxJ0OAtna}`#!$pD9rfW_umbAo07
z;ytOY{+Z!6D~!-{;K>z`0;?H-Y1g=fxyshCViPC{7AqkoRx<$cov1-KW0DHqSTZdB
zYGe2IdllclV=lr$OJMO=PJM7cy3_Fh_Aw3}J0DUJu|%R~0H$>f^c22~$+pI!qhM{e
zA)BZffVihmv3z+i_B$i=Tv!}v80e@OfOuHAX3CsRV>w3Xd9e5^%ioocANdG1;!U8%
zu=uNee>fBHlh3j=t{FOJKBTfXw#I7)FgcW}T9|emuVd91>=$gptHdO+2m}tV3M2)J
zfW2lhc@;m$qq9(G&9Wc><`Jz4OMk?lz~EJECYuj7k5_Sc9A6#c^0{Ii;<0!FfB=Kh
z=ugc3Y%4~ycMwOwV>I>){^C_K4;c$Y({ZDZ#JTEe3Tz5r!k2>u@G5~!5UxW)xiCV9
zM1i6ZfSh(e40tu@LPcXVoB;D;G=>3ExDE!4wxn1%q3|kOC)-KD<L>=)%|iVJYvIKy
zObQPej8}Q_ynJ-XoA2$bLsS-(572ZEwq-=zMLYGpV{PI7;nR;!sxmQu)X3Fq6UbxO
zJ?bO|H3P45mO6(63qdx+j#e`sEP&C36Yz>ZseP4K*S;n2a%ysc*I!1%w{YZ{B9+*w
zj=6-5a5P88K{%Q#=OP@<3*>1XEw*@JzK=L&rP^8k6e(~4>@2C5LWjJAy+d_~8bA#O
z2$UX}zv{Z9+pw*rts~elUS)4%?*>p%qg!tUS{{XlNrTuv{tMOxPx&wscf{#?G)hB1
zkc{XrP}tL7n=`}vUVjb=@ard<06K?{3tr28I;9s*7j+VYnZjj+$<KCh4B#kR8C#75
z`-mJ$3gO`vor9-T2OrpsQIW@zi2y~{jh0{~f!?t0#Hka5!0S1!j7%4g`&Q~HovcV;
z%DFd$NB0!s)K>oYl=Cvmmdw9&u?r4_H_e+1+>2K+Xbc46(zslZBzTpOCe)%>sbt<X
z=fZb-S5XBEFJ+x9kAqjCQm72wDohF!fp`=iL$`{EBGRH2y+*3K$@IH=S6P*I8oy-A
zIzIf-e^VFKHaJt%ClW{kyh?#nxPh9s!R?e{P5w|1Y_GYkISCv^coo5xKmz9$UPZJe
zlE7JjS6SFvkU$>hz7pFpTejN36L`_QIJynfX><hQ&^R33h6OZ%7J)@zGu0=fj^B<_
z#|RMkyaZ9$pHp#%{z0*P+4eWl-yEmtb=D5Yn2`|aV3FRBz{W9H&DpE7!bR$$Eh~7H
ztJD>N+@x+=v@6A+@D(Ay2BxN($iUmBo;63(b3D~&2F;&_NPoJNj!1t-FaweP%mAh~
z{XwAMQl<_ynD<lQZ|=Rm)jgxC^I*rzILd^_lg7eDMnDnkg5V|g3e+KQKkpzNq6(>g
zIz$uD5PEbm9ihkYV<7aHLZ(*luo<g32dA8%*S8fU-)rY@m$m%kIIZl@$&7wnmTv|d
zO{RG2Oz|RnA&@uO8-b`~Dgx2SGz6lP=?KIiGZ2VLW+D)q%=Xl=kwfMn5SPqFARd{A
zKzuSEfdpg$0tv}N1QL-&T2$4{YjD8UI_Y)N$%M|Ac^J|H)R%?uboI_Hv8V24!?ZH>
zGSyc9bACk`#{|9VD)HY1;L=A9Wx%ua554PNuK(rwA-@44d{!X&h1ZmFz&6mEd`>^`
zx?$sX(jDVB4e$c$bLQVqhnB9%@>vf%(%}UtPLxTyFteT5lRyvvk>SHgZH1QfU!b-i
z<J9jsaosS>*+RBI(1rUiKaNy~xc=N=9pZ_20XoDN@)3FhKLJ8d=r2U*i9{l;9_Wtr
zwiBRf(%amD8VVOXbu+CiGfq+`ZLxcJNpjER`TWi5n!%(Y`YSq*j_9u#d<LSwVzQXp
z{_63Yl&W`v2z5>7Gx>-{RlpM<8dV`*h-g$rERnWReZnlsBK#H^sh2rjCE@%OhT|-~
z90ab1H>|!M0Or80h=-g7T=q2Z{xjxxwKaN|1I;-+llms1n@NdNx`Cbp{D*^}ZbiUH
zpL#mb@r=$2zkbfU>V;mz-<j`VC!JjTQvXMvuPC1@8(8jrd)eQHHw>TX)dfcz3ah!d
z04l)umZaDScqZ)&aE@zV08an`1kwc*3#~5r5(0jK(o&Ua@NEl%764-*Xp#02r8<J~
z`%t!aUair>g!n~?$^}roRtPC|kBPd+1l{98-J>#jK@xnUgcQkPV=xAWMq`a}coTxD
zx!E{Lh(sPt4Dg#u#0FbfTUy!Ja=3hfP~_w5Ier??jpgq2Z*M2qJ4|$Tah)`UME00Y
e@$#lK*^Ugm8C3uA6DHGwWXsTj;bJK&6#74{MpmW(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_bwd_data_interface_xdl.cpp.8747321B4790F80F.idx b/.cache/clangd/index/test_grouped_convnd_bwd_data_interface_xdl.cpp.8747321B4790F80F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..668ec02b1c6c69288101ec3366d9c6ccc4d1e678
GIT binary patch
literal 7392
zcmcIpcT`hJ7k@8=AiPBB0ff+7C>lV41*J#}y>}@pihd{p3ZV&zE{cc<f{F!oDe8h1
z6nifSiwL?#MO;Nh1;Ji%{U(7+vYy?;UvQ51&F|fLGk5O2^P4--m+j=lRmQMc^Vpol
zG0URK7=}rqe@V*{7fOm^SUv)r?Pr3*{2o}yeyq4xQ#hynNZbQOhFW-ej(v4o%qRA5
zE=TIt8m4Mns(=6Rt%+{m?(5o~tTVSVY1!E%8)<QMarNoV^)V}_UR!s``c&5OYto4(
zCwJ8ZIbYjc^Zs)j=a3!e)G?cFamp6kPN$x+v*12btDRMFaEa#Qmrvh+erG9P5%2lp
zX~ze}8RZ3&^3=S)MqBSLY5wiO(LOKL7q^_f&tCaFRjHvMA(ZP|y?VjQ)M;J@r-}+q
zF1>Fzs7MS8{PXC(vSm6uHh%4qeo-G7T6tsXY~sY~p^wF(9x6%ADw50vrd3s8cH!sP
z*Ht<uTvjRE+)l2olwxX@dsKZ)=-a;h&0zS+7`;tR-OpxP__xfSKf7s9kqJ|B%JN?u
z4$sTqVUqLLcC(0(j^Z!B2Mu-=$A)Ou+o_kzwvew_l}+>h;=-L{Ao_-o<Jhfg39hqJ
zP&d%EB%+9>X-S5*ttswdniAxJ&C{*VH>GnG{Pooz=l-60sP&T}M=aj8;e_tYO*?LI
zZ66zGCM~XUcYaHmzhu$Hf@#g|f9|Yrk#jcCWH%|BG-#~Y*jj9vlrMheLJ!wu^Ievg
z%A$Rzot<2&E9JFx>{xYbjTiC;FZ>##ZFMrIc};-X8f%RQ%&U2OJA)lgRt31PxpP0g
zm?i7uytLiTT&!4ns_O;nT~qs4)FIhQhpwnl_YX_CiJ8$V61mF#B~ff;Wtv>tIf>`T
zkIbGN_pPUeyIP%i=Dw`{!>-=cGPX(GQTJbb)61H7`lNG}-F7P8qMV!5e}Ao7v+HmB
zL!a0mQQVrx^<_ouSjWoXrhbVuqB(C3xwC6o*RDgeFU1d?*RUwH3Y3%DlTtbT_uFgd
z4{CKg-X^c%?n|k26t_#7SMo;xEGzP%LS9C{gUOw;o9AD?((${xVDIcLRRQ}hK8~>8
z6shqk#y-w|zqhTF_eSS1D`lA*D%A}s37yLQHL=<yHJ83$=)$@g_pA#-h78m{H6@pn
z?D5Mi7rU+crm*Hj!-=wlZymVm)z0MeG>3t>_4^#=1}NFp5O$(O|CORgBLCa+UR=kK
z#onx*`#$11esI5*<NBJ`x5Y!298<1a&5g>{`z^c*Eo^dH7S~?C<vhJ)!E^r~xt~QU
z|6W~9jWfBpb*`-9S1*T$=A}<PRLn&hW2PvzvhQA)B!AK3FWO+9d91j4o{Y}N;&<z^
zekf}92B{9*A9PIY3O)L(Sx|T(D_T*>SaF-%vHImM#^u3}rg!(hbSO1%^ow#MZF%SO
z?u^TUzP24Zqsp`^7L{~r9e@3Gqmt$?j$S5R=dz0Loccno_}i$vXX)Aj6}jS@vDQB-
z+h*<EX=8WS^!0=x&;6|*wgl&&xO_uv=7g@A7YARSDZ95q?8>qGX;Wujiuc|nmMoS(
zDBZWWZr-;*mv{R;9sNQU7Hqh)T&rVYQ0HRebHiVMQ~YEpsmCMwngfE$_UZ)S!IzI!
zcG4RM5AB$gm7A*KWxo7_!3VZc_p8nyRYXtE%j?GF=3RGgt=zdvqK#?jSDn=7uXiJ|
zwm0FSoP0?mA<Ed>xliKVN_pQ)FCN7@o9eW6SiKn9dN(rOK8cWD*lO8T6qv4f{q?i~
z$I0cNnsd{$PS3nLMRY^=cT!M_E$99D>d@%w0c`cqlUD%)nAD>B?-Sp?dnVyp{@6V-
zEz)wUCiX_v<SX^!r-Ri86Q1OrG_v|p8xXc8eE0U7RclMSFMLtm7<uAwY1_%r(|3!1
z;CfHvgOgGh&v_++VWKuhx%OC75J8$9i*XDk44Hg^n3Mk}E^@a&>~Bv74GTg1sVc9^
z<O@*CYJM=2IdIW|AcZ1z8YPY33s7qrJZ~eP{dk=qMIv<;B}<AgKy7-UiW{%CDN~FT
zhtx$lR)kB8JTM-AqrYhd{qaUo+06u&jiWy6acn&<Ir4x$*2jYl_Oq`nB}mbz3x~oX
z_+oef3}Y|4=98MikR(Wp(F*k>^knz~)UL<c4<G5+xP*qSKwsD3SPe2__yH66v^VJJ
zOBhWTBmEC`S%Xv7;JPCZ7{zarf@E_~;}HVO`A<7%hY>j{cDtWXBrQRPn5mlS^92kM
zaY)4Y`;Ed0lB5`<HWxRi@&%}4!_)8nMqmGwAo(LB<SFt}d;#ioy%jxyt@mwcSQb+M
zt1nlyUG3es>;Q820@Ou=qQT$`7$G~Ep?CS_)|~_?7^&qbaufLi)cJ#hs_c9IY=RVq
z)G3q{f-gW_;nmxAa`*Gg1ZfUZXHqf=z5w;f4K8Zt#?12sDG;eAQzmQi1*lIIWvybL
zo$`#pR`T3eh+~B)%)<|eLi3Eb2G`_zjpRsC$UoZZ+J<}qL$p-Zy;SpWxgthdghu$+
zIJRywZB~yhY9UDTP!~FdF2NTt!W|p@M8%$-CW15<O`t2GE5jF{=I+bcrhKmrrCB)o
zI)joy@CB$J7}gg!<ZU2{krpC#E>6kCwMQP%rT5Rg_G-uZK{PA~bz#Ue^!NhQpZmHt
z8b6EDAV~9(T8W}0#uuReN>bX+rFFRzqzGi5zJ$IsUx50%%Q0tLNzr-&%l%ImVu>;H
zlT@VzyPhP$-|dhYR8E9!gc&3v?0wXUxA9>jDrYTjv=+C6T>%^*BJq#CPvvCey4koP
zZ~zaxWaA!y87aGLui;wHa4N@VhL0Ugs=n=!%j*HnX(aN<E(Zo6kE8>niu%PA$|P6U
z7@RZ`Y)Al@0NKN19D&*Eag+7<3>c4!#xpC#ZD2f|8m3u=Md){AkBy3r7uaLa7a{5E
zY`X%z&M?j}g(R3_WF;3c=d`jYl~an7OL0lTH03zCTo_m2<O*S2i<4`G@gbaiNPvM8
zNhDyxk&($GI8Kr1aM`p&8P<V14znUyKpyc25Sux(jwbcjjTvKl^p~Lq^)u7|)Ic5`
z<`1w12o@y*k6mFRUPQgJaer{Qjn3_K)9&5zXrkfofs=l=$Wiu{)NUD(H00A1;}lb1
z79eUAI37z2l9I(>>B2dMqT_870?~mrjwS)v=5|B7cR$H1m_+@_JG`*-$OSrjm>0*(
z8x<5GTj&<_SwInx+-S+ed=LGpXl+ChZV0~dUEZi$>SI!jzPD1b@&?1;JQy=Nem$4}
zNcDg5aWp#cCFDEo>%g~h-ZP1XjTG`o#z7*GIl~zzz{43Qz?C?;5}zc%!x^~2s^Dqo
z_ds_tXE*}^!!b2Mf`2>-!eA|Ftpw|%a%?4Sg^+`!gAfX#h0p~(LusKxD2x^+gu-dz
zqiA-!MXILJeR<?cGdVK`*Z~8fW0`j-f;|zfcC|Z8h>zp$k7##xq_IZHw{f2cPI#1a
zyyIPvaV&pUB;<eWz#)6HZ!bO|Cxt15>49?~xFr0iz5DZCUfwGfb5+-zJMOFDTzdR8
zjeP45g358CIZ1<|RE{&vSqQn%T!avt#uh@ZG*=<yMspKF?lgBH<U#WgLY_2FA>>8#
z5<+vP%o#o4fS*PQ4_uy})Ig2T&$1bl$sidKW%24M_@0Um{wcY*Gt>mUBa^lk*9KOB
zQcG0gA0-x*!?I<$K%^0JW%&xBXx2g@l*U?#1I!UWelV<s{SGC?ER|WKXIHN8=9%w$
z#FEhKP=!!q2*GOYBR-jhZ=Z=!Il7Z|nNapnIeL@zm_Wi*j{anQCe$2xuN{>wZx$$_
zk>nKQ7+^mzH_BLEK`2;$Hg24ao5I1cPVD^dH5U`J5gD_LSSDaKLJ6#7K<ivZR=wKr
z!vf`twxYHgRC_3U=uQkU9c2&QSqQn%U4#&u&K5$hbXOtdMt2iJ?sRt{<U#ikLY{O_
zA>>8(8byVkMV~`b4w{Vn-T^8`yLq^fh1%xlT4$qL=zDMXfzolesHC84s;{c<4wL|f
zZiX2Uw340Lwu-thd5Nd2q3jG*{?;X;C9kCm>(T4svj;MxoK?RsIhS<^DLBR)Qz$Ip
z@__P)cXNQ=yaK~68+U>{s(LUu$3)?=4*ELHIL#Ca&%Khrl;sk06~^5OzxU;)hCJS0
zi1IiaSIWj!!7v`S%EoO0*K$sYdN1DNgbY=hs3Z@Cecc~3n-~&j)yLh?!?9h1>&VP`
z<33Tt(M|c$N4P7|LN!b^%z+LYok}?eY<sz5^5+S=o@<-Vyy_ceU?{5~lMu1;E{lSU
zU>Gsn;2uTgs2Zud!8HX6c~o-!v&8$k96j^H+N7UV{$||0bMD%u<{GDd8I&)>1+icm
zBR?>VytpjgR#o@#bSH9SxI(x-u-%o1+vfCAE(fC5svfG*u+XMoi<o!CKR1n8G6IL9
zZN<ATLv#@DQdm;3g-59cxCM-asvXD(N@KBq-kQ;9t8mtu(X$BQe4JG%j7xCVW+A+k
zxped_qH+?L2|_55nK+7WcI$eDJ$jIgmY<DJ%Eo13dv%=b{`f45f+0nSLWm)_tK&_X
zrp1xG^P`smSQ?ygD4BS-dLci^u#~|MsDAgc3Wc%Yo`+n=yXk{Zc$TPT<Jw>e9Pp^y
z@Qq*4u?8n<a3+}W=nVTF?Pb(jWOAl)rYR8K6U)sx($@n6$2}FhH~ZTxu}}*eb4T{L
z-hXbn<!t@9TLhaJxsLZsh`@2VxK}P74W0$=M7I&%FCv0&taT(-0Z9-&)^-wWhcK=s
zvFe2J7ZU5MFwPfYtrHRYzFdU0O&I5pSu4q){}Sfk+-O$!YE1t@#e{N=_Y;j!u1#mn
zgwar$<~=JIa_yVcMyt5PD+b3Ldq`@6^>riq6O=>@4~9Q@aOn2-@03=`lQF-CtK^sr
ziGC-Fiw;1(hJqCXUWWCIYDQl`z~D&0Km!U4_5zlaI5+7iFOmrmP7KQsG*bl4Btdhj
zpqVIW#tWLu1<g!BGfmK3CTJ#)HbpqGQN67zC6^>kOifSA$XuF${>Lz7d2&3KKH3LM
z6*T8Ao)^b6#b>df2w9XOHbI=A(xfL!Nls#Wv%NfIT%7b}Bs^pl<>ZwnThFkuwX>h)
zz@i(^)c(al{omdys4!GDG_`c}n1)7EOiazpEv?inl&4N}rYq=9ck^6@lUxHF-ErUl
E1N74NG5`Po

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_bwd_weight.cpp.1E0C24433031CA72.idx b/.cache/clangd/index/test_grouped_convnd_bwd_weight.cpp.1E0C24433031CA72.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ec2bf58c50f5a602adaf83570cb0798cf668195a
GIT binary patch
literal 8652
zcmb7J2Ut``*WR-rxa>uNrT3++be6uj^r9e&AP^K4K~U<YD`?ONK~!L2S$ZdcB%fV0
z(Zuo@TTr7$qhd*H5o0X)MWe<N|JjA%Ui=gD=<_i5zB^}T&OLpO5y3%0N4N}ze`0V!
zMp9n9JcGedMSr<@IVlYY)J|qF3R-tZ&&|B*u5z*Ca$j+D`qKW4PlUhZq&k<0H-2a8
zD=}T~5SUsxn7#VVr2N74=RLBoO<<Ne9sYC3(28ruJ(Aeu7^<?QFuP`HSlpx-KfBWh
ze>i#deE<C!D_nmw%5d;H*f~v7$?hJ=+J4C0<JzKqpK4$C3OM%g=$*?|*232>Rt58(
zFE?G`q0{kH?0fIh|300yHOwez`E$3YPZb0Q)jsz0&Xm8s%fY;-baAmj5N6%an9;N(
z^<K}6y_VLu2P2_v&4tcom!|jmT8Trxk5YLIZ>rL)1iMe!J@@&1^FR;to3(rNEPfGN
zh8{6pXz#mHxJJ{_bi+*Fz0>EVrN8!8{N`!tkA-IK`;LkVPgU5?jjWls;+9WR!Ns+q
z;+iX*eTr9tyY*b(KD~Exsha1eI#>4|=Am5;`!sYaEjldq%zRb8s&3Z`@HtkT8ZRHI
z>HcJj=6^N%Pj$Dam0B#)oVYel_h4}e@97UG)n*?q)Z62BG2f(OQn%$Di<zzVv6Adh
zUR(2WlQxD1K2un-RWAO@uSqGzrj=Q?;Sq9{O-DnvG<hG=bzB<l$Q6VLYEBJtFMbhd
z+nXDsS`$3G_=w5$jb>Klcl>3UuXh&LZ{sJPippCjT-;(cRQBC`y;FIIH#Mr<@6R<`
zJQ!emxL9uLG&9Symg~vESvN{9X!jYLUsPUD-uvje-WS#tM~@j#D?6g6rZw&Rz;Cvx
zz1)<lSedxIHKqS-;-Sr=HLj|8j>jF1&sQfc3eK`Ao28=vQa`pm<p;BaGai)d8J-A<
z|H$RkgzY*_uiPW}aT)9QMbiZ@+&&o^ToQC~N`~{`;FtB6vTq!UZHkT7Q2mT;_G|vV
zfSIw2_N}W|F`4z-;@jbVS?)UrUZ|~Jy#I=e=w-IrK%|bBfB0RuCP%NsKL>tQYt!p3
zn7%G=#^$0X6-o68t>lT}J46;S`z(4qgG>6ef6_PD=pNz`_Nr`uiQ(3jn}>oo-|}V^
zhIoA`u$=Xu8M#$&rnTNlAw#zs*R0mh+CAi@FyH@|i}Ke(oR{@#=6$V_2afj3o-dm^
z-*)3cjpqeX=9bTvr7GJGT;Cr*Y}^*orn&Q$P;fBbN&9Bx*I#9A-#6@V!D%2S&f$z<
zLT1$;F0suwB$@eU;=3<z30w+%l>>8Da!(HL`C&)5<eM8G%`PB!JSv%GcvigWSk;p&
zmHnj!={sMoRGs&;MNjzP!O!1j>~$Yp>zF)SA@eiht1d#X_GZhJ<HuQs(SmPtTe@TP
zHry|%vs*v;d3N((aZcwZwp?HMzwk5z?T#)B*SU`)f)9pHIAyN5p=$O8>+O=~WmRz+
z$0FwVJ~`J{s4>r}E%ZY0YyW3oProIJcec$+DgH@u<&)U-?~^CnTRgq|am4yfMMfuI
z*?)FAYk$=S=ezoG+1>gD=ZtQwbv&@jH~QYIw5+wOEl;=~Fn+Q9!J5YpIAvDrb9XDM
ze>vsAsyl`%UXe=A&UW~}_6z-ONB8jDG|e2sW0k^rx7`=h(zUkyab?984e$IbrMrH*
zx?uSFPPMH*$Kx0F-Yz;(Vo<64jPH8whrz7#+FqB|w+w8^88Xh+c%<%;JFq&zX!{{X
zZed2kkQ{@d5Prk*wD4FUn^C}Ekln!O29?nVLmJD_KgCafFP~iX(IA^SFNr}KC>dzT
zhLK4IgHX(yyXEPy2b&p-rgfBbG-Lyw)(m<=*yI+QvzamIfErM!0nX?HlUSbV{i<m2
z=y^6{B_insMla%F<bipJaGto!ME$ojip)ibqz@Q<pfvixBym@bQ_g*|{1Y~FKH@`P
zNuMJd@F8(;R?mLbx%WDoxdcsXD`{)U20R^qw9ah%?10s5#>fA}!;;RcGYKv4S{3F3
zw8H|{0zx+69dZKZEjW~~l&Z*FjCLpkg)%T1ePEKcEq8fuMSmd{qtFgoN?Pi&0Z-R>
zzD&Ls{$-3ZBk!L+vgW~*7-9ckc5LQSw1b|K9!ECd9U3HTwcS?shbA#sqiJIUV{_Sn
zr+cFN2h|$xm$I23p##c6p&aRA<bg^0R`CA&M@(IhA~PP56oW!B`ihYUCfS@H{Kw|R
zSTi;=8Xcmmq^l+y@btEdoNrFQ5}j3MEJHhd3yg1(szx4IRR<2|UlnXBy~}1SM<jo1
zbjOsH7F;}jsZ)^|he(PPicDk!KIFUXH7{+p4O~)WCZOqlVDzIyMjrT(lQCX-FMG{b
zsxVg}HceP2YO(?EaQ3D+;!mUNe<?B-q8-!~)Qx2Wp1$ymP_5s??Nnyu|I-22o;auJ
zguGtIW~@LYJ;3NeW;^o0Jp2@;z$=v;%~fP3B9gzA#r0)3Uw>QnxzQwMGGf!n#>h=J
zU_Nd?7?KRnE7oB%=b{~&K%oh#Y~+DSen0<gV8d&M4VzK;PkiJgEy-=xLJM@K?hJW2
zqzBAVW-x;|3^1A4`A2t`AN4Y9n>ZNADanEWmLfq4b%bVE<OG=<og9z-1qX-;O0vK}
zE%CSZccqpDSqITjWLV@pYL1L1$F@^;Sl66d$G@shkQLxj0Y2EO39=G`D<O=+A_x*e
zD21!Qql!*n4Z+oPdNBlv>GTrtlz=~uU<6qUp|ub};W`MegD?u$gJ(URz5zlT==6;c
z+(@Ty2A^gK!YU@n76@yBC<?biXe&fexDEW<Aeh4K;L}cD-vMDA^!1$(+DTvE1^!*s
zdex|N$yq(bc6owq2Ipq*pf+v+{}zgnAX~wwm3~kgxV3>NrpE$cD!mif3FHp;Js`vu
zYP{Az*ksEc5}nu%X6;~wcOb|PaOwa7g*(Ba6I?0W1-ve>#hBZ2efr}^rY>^wqvnbu
ziu|Yp;uGaZ^9W8cW{;--d4^Qy0zt-c<1Da=@)oG%eEHY&%S;#zA~?P<-@$kv<R_yE
zYRpeYyc<V9>-!5;BBuGQk|Rh7I7+|`3rT84zO@iU;X3fDqpZd*EODCSrsd@zPYQWL
zOY8&owo^5t687m!8J`%Zj(y)b|D)4yGnRHJNS(o|0_-SbuLMCQcv4sdt|IWDa1~fp
z(dnx}P)(;7gR7XLceC1}ta{^w=LbvN+v3NSPPe-<(ZZ3mjU$QtsAzGdl^+dBSb_nP
zr@w~haDJAVr&%-ISX#iTg>EdZ;Lu7pmNwwEQN|Le{!q1xBez1P3Nr~<QlhT~r&@5Q
za2+_*feXfgZ;xCvFo-yZIN;my9kD71vI@+qU<xJ5YH+FsHwufvK@83q&p2Uz_dm?A
z7<B$fs>8~Yx`#ArP<Xse&D_l5{lqfmA*&oQE*!1jjo#sxZKogayCa8;!_w2z8#9WF
zCS<H5qQ?heRrK_4QsYmkQ<-z;$oGYj_nwU@7pDX4PB_(Jcf!?D<SaGbuVUc@CP5mp
zj3(n;Mv%rV<H@)fiT*u(wii2P4^v)dy4dIB$%2Um9E@Ui?mp|;7&456Q~_ocbm>-t
zQzc!xB5)8<(p_-Qz0lP5il&qWu8|eRf{B%h6PCon)4rSACG2PKWg)JhHFMGaeYXjM
z6ahyBx>$iy%oEXlMGPD<eO?SaF@3%cICY?l?T8@jfL8~06mA4gBYnORc#ZV=R^YUP
zE{=B6_{(brI|_FKr;|S43A|2f9#_IBM@!3KD&g{j_+eNi1eq3{7Kg=;t7DX-rKK^o
z#LwD~mX!mo18FF7O5~V?2{I!(V=PC<hX%*lnMhu}SKio&{xy~+sqZc@%zwENDTb%Z
zGo`+Ougj++AwQ6s^X2+lQOJ+$M??Nxe;Nwl2GCF-H;{&cxIr{D$8Zjh+B4EHl8yo+
zXW_Y|nZEni?AVxq81{AZ^~66Dq@TN=7lr)Y{b^{9^BfOqF484(Dh<t#OrTI3Ol|`W
z${N~0s|^h3*r*N6Dcl8<yXf>?pw&gE?*gMPiXNK;DyXG*9Bc}icg)`$$S!V^IjBPg
z<$%(%%exZ%>E0p~QDs(kf!7)R6X)Kbgl24RY=NT!zAmAnVdSoagA;**9s670;b514
zS25`r&u#F%wzOh-TFKDH<Zx8jj@$xp0$-WCo5<X-+WUjkPpSFN@FLOgLy82cWuRq)
zuWkgXZJ=#~-Gv}^40LR8flH9O2D&!*dPI<V26{I5!bXt#2KqMGGYQhbz`zFA(xrR-
zH+>OsXzBlB9v`ZpB)cbKRUkJo<JnR6;A`#cLZJZb02-PzWezQIrAMbP!ul<nxV&|R
zs3!oOI+7i5x+r@bs4qBEp{t3`apuu%GGJ<eFLh3!Fwl=eLBb#!ssv6Y=uj3>2}YG*
zPRG`jz^8CEaH{F_)nHUjr>_R<YC3%_aB4w^il?<;R14;GY+Va{3O4|!fll85Mh$fO
z2C#0R)3*Sp1#~D6Y5}7b%0r}ux^)Zisre4zbkONLz^H>x-vQPg6n%M!+@}w#-nyfd
zWvp(jhdqKIO{STQ#gK~M_Z}MdH@fpq9VnVr{bKI&ZI18t2C?zHr?W#J)xDPp*I{G&
z#}0^17I~($!9ZDEYIN#Y!oW-;B59j}B1+IpFvTX;L{5-ARuk<)b%`|B;{%%cyS`Xf
z!rzWUNj;d=(+jc&aB83zWR2j^NEKx5>OJ{83wq0?1z9*hin6{caHyh3q-yZ4rbi?(
z2*p%H!ghyvAK(0t7HIa}AjG!`TpytVY`lxS>oUDphZ7vNZ(5)-Tmse-aG*@D7CdUf
zo5FSARtG`~*MoIEoxTA)8tC+m;MPdd<Ms?<eSA}fKyi=V4?G<A`YKn=(Rj$6FD(gq
zmSfk5&s99?OW2FG-)kA4+k}7optJr7S}T|=;9@avx?fT__hzd|cJXFaz$7e3X$D|b
z(y<6wB08=DRuvsr1FM>j#lRBNu>@EWm_!Mw7Fe}(TnDT=I<5y+JsmdytAUOifz?RI
z&A@7gNtFM#0IP+LTY=R|$8ErBqvLj9wbOA2usZ0t6Ih*e+y$&I3gZlc>N9EgjPf16
z8Q%s=n;=u7Qx;-R+Oo3#Y-Sm<4AffTCr6F_{JqzJq0#~Fd)0|ARLIds+i;_fxFh-J
z;c|JaL=<A^rTFgt+70nL2aC~J>MrU6%-)`$x`M4fcLzT>lkR^jbbP>X{@;6<#4h)*
z4XA_151;G{dcZ-!WBjE?>H*R=AdWK9Izrle#P}F{(80lWc<Wfa!}$)k8j(*)`;3@g
zx<9)Qc$f@(7UFRHLHHtzn+OAItF11vUJQKjZ3?R%F(PeOV!;#S67?kpSmkGK^TJ}U
zUtNa2zZ|&bU_p6d4RC9~g7TUs;5LB;wsO3kywbQvCqWk36a`@!5GbtJ6hjb>GXyHm
zY|0=AS8L~IuyPuO3BOE~_Ka+r!4dB+&52W6z?Z_U;NA*e6mA3iHgKVEJJ__-*LT3w
z4*L2|aPOqA?*jWSYCY~6AZ?D<#$f3sZH0W@i|^!3e1k>W8-MA>+QS>-vJ(-Gw;$Y}
zz_)UI(?nh@?b>4TqW+9VoH5oGK_(g|T4Hp4fAD4H-xjTxB|gV`dQ+$Yd>iOZp+@j(
zq?$rEUyxf~`9$f;8fVV!^j2L5_;=7-b)DeTNpID4fm;{Vs=Fm<Sfln`bo6_D@peO`
zH|cI1`@x=$m5U5`e2bbu-7@=L5aRaPoscWP3K|dWdM_Kehxe~4d)6A$7wJ>AYAnEE
z9sU^`d2)~QmWLnwA+J=$Q`2{>8vfvguYYWPqi|Z`)(0=lNlK4L{~5x-j>f>>fIX3c
z;|~tc3~X`OcNsWR;fb-$#(uz?jr}mTHaBN+ZizJXW#IV7V5E<|0vTPPdS%<2ZD*7+
zv$Jw?Q_|B?GZ;Bz^Ncj=HH&&pp<Yv|*Ieo~n|jTpUKeL1rb}UX#vD2M2?~l6S!_aC
zZIY_W<lx!CVKX_?gDg2pGu3r8G<Efap58vbe*OV`EjurB4{Pq<{i<zXXk=n)W--Oe
i#@61^$=OXX)!1F%!DX73HqUj&tRiMeWMC*nF#Zo#@^U=@

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_bwd_weight_interface_xdl.cpp.B93D201E166036D3.idx b/.cache/clangd/index/test_grouped_convnd_bwd_weight_interface_xdl.cpp.B93D201E166036D3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..453c299c2f83f82fc80e4f3fb399ae9d25ef6fe6
GIT binary patch
literal 7454
zcmcIpXIPWT7S0!9kQ9oPC`f41B1J+GqzMAjyR-<R;0j8U8kz*b^(u&hAiA(?gR6k3
zfCUvRHk6{QOI28L5wL)|tKwCcI|&>Ty?Y;i!Sgto_d93WIcMHEGkQAM*~!S`a4SL`
z*pchw*Aj3zoCNw87atogH3x?~ivYXyvM<Z4*I26KVE3`Sb?vngy$k4y;y-0t6NhU*
zmwAMCci8M}C@9_Y_5J6j#JlBzB!}Jux%sURKec|f7k3mXH@bgjP;<~#>Q3^9@uwWg
z6r=jkp-%N<zC*9#yDWQ@8~VG%?y!V^e3le_e(uWctZbKmJ$#?OX!XIL4{o_|ZQ%YB
z6S@9AN_P`G{;_aojoVQ@xivpO7K@IutlQgmC@tr0_>!;%N6Su^JCoYHMt{?)c(r=j
zZE7d^RGM;xzki#0Ky5|5-HT4AMSBBA3_R?Hmlc`b9n-1!?dIn5T|$Rrjl2RbO5bcT
zwN|LNQ>4c<6siTQx>PbNKiO<lIN$0|n`kX#Wm_LGi=t<(xMddFQDjEaJc@tsxxr7)
zNxe%`Qlr)3S&^Z(e?#)0PtAAjGqh$&1vC|w9J0AnOJsCrYoBA5Z#lVDBaERR^iou<
zsA%t2;nQB421#biTa%nwDkQ1VVk5KbO)I<Rduu8^&8o6J-Ky*VMzpVRW2#%ahG|hv
z<_%xF=RL0Vd4Fg(KYmmww@IeF+C1{W;yfR__5#z;jkaHWmpN^CtoaKAJ9%y2rG~uZ
z7w%4j=1gmS)da_oc~YIEt*o?hdW4^M?NPb<TtC~aVcBjUJqz!_+sf_Q)dS%d15(}e
z^ORrAbxj%hVY#PfuIG|Q?FEMuiCg}C#-=Ygz&YV+e{A|%c~TZ#zCAd6;X<l(%2m?v
z`Px9ah~oPdHE}D39v>JEdwIQJC|PZ!td%(u_BvUoIJc+qx!=3TWsz@+GoKhsg!+ej
zuW~ANvM8?;XEAP9#p(ShW~w_;Q7)mOUnN`EDkpVr)A%k0Kf<xtJ5|coTJjYwjD+yI
zq=Fpfjd+_^kG6LxTR&Xb&5qBDZuVVzTS>t*-}Q{N*XezUzlzYj9_ci<S9(7k=eQ5j
z?xs5&S3N_TBZWyPpE=uD<9*>-*$17xlB11;>d<#(8C|BsYagqW+|kR)5dV*TVo53G
z=NEg;4*h*RNx7-I<!o-Rg6O!x^F*O2#n=s-oK<D2+RGxCZr!UZ8`zi6wlx|?e7G9%
zV?lw=3VL@)W$KYi`>l>{rP8WjFMm9lw8y#7fwW4*ig@sox9G#8f8RQ8UiS4jwZF6$
z{QhXfEIrJ%G(lT-hfimXn48Cz;pWk6fvUD%&st*C-tN$|8Ef=b{HX9bDQuVbhdXBG
zinQX42}4C%=RtbUnCgWv?Osn4FYCM|n@j?B=Qc~P=~}kuzS{GTEy0&8yawf?Xklu&
zi}MZoonpPVkN!2^>-7~}TF{=}o;{|qyWLNW=#T8Zy}?hx(xP!Sap;nI*Fk2VtZ&#$
zE2Ue7wmUZ5{N>2U0{sa6Ih#l^r*rI@%^DvJsO~tmZMVa_I@!AQ=lX+cs?MqI-#^f~
z_<`(ap<63+5}JK#)9n`}uTl^9irC>-w6rAt{pD8@S0(z^1n1r;9qW**zSLJA=Bt*L
z$#TEHCfM`NuP%mPozp$<v)AfQ)LwWOn-Hqd!g6~ZcD*3GVUx6sNauwFhSZ>2M_o|l
zwp*N0>&&H9F_kyUg10@2`46q<Aw@X;UC^?OztZasj$V+-SwzFTIaOVzoSBfJl~Y^l
zo;Tg<@B3w-n4|vYMCPSb#S@n=9q`%z@6_c=ci06*smrr3w#SNn@{hyvf6%Jy?Dutj
zKG*7t+pculvrjoAo65#JqVBg{y;l2k_DPQ(TKi4ScT(BKN+(LYJtDje>2WuG!z_$c
zC$P1#S)y^tk?W|4C<#w+x7M%Q{*;Ul3B$1&B!(ofK<vr;o0IJHcR`C>5EZutap-Xj
zcN`;4{ouGXi~gik-Wy1sykSGe2OyOcVhWj8ObT(h`5^-@3iNKUgz+1YScRyf$SXiB
zcQPPn&Z6E-VSG3eXJN!FOnvGHR2D5Ben7bL;Vv1!2B~BaGbDHgh?g3O84kVn*es0?
zMk`QLQq$rUAT~Six46>bRty=x7PV#**<@Y;V#|Pb=d!z3)W|pvhU9q|E)Nr*`hgYN
z?t1rj<H;|6WV|m@ktRxu@CxX{BSvv!Q0$s*WV{a&%Ms<Yc?F0g%dfu?eX;El8JGEg
zN7y+wrr^-+?Lfw}P!~zWBr>mnE($!4o%%3*K7@?-Lt+`Cj0mro7U!AY2yww-WPBvD
zfF?;(l2?GZ*!b`8i%kWJ!uXA-^*{P>fA0H-=MFrs62(WM9t;%>HF*W}aL{e&!NtSF
zH_7;QsE2f7I+<61_-Llx@~*~;bSiEY64Pbq3|;}^(}eet&4zc!$oNnsrV(ipyaL3v
zlCCwIY|qlk_&_waHc4BHSAe*ouI-9Q;={#MTn1{LkKyuB1Wx{df3EuKj%kmsPaxy6
zxE9`x;dUb*PX2(-b)~yk!v081Bjbb6C>kV<dAtHfxtC${R%#$-nK*tu5}SyaP<RE1
zyDBnE7Ir^CL9`kTkVZ@+^9m3@4Jk4E>ALA+GH&btoxnij#D-&R$~sZ}HVh46ByL3G
z70}0s!tdmNaegHT<2Rupc3{LEnC8?EsC?Alo6r6nVL-+CBNYW11qQDG@pzlWhWENl
z2ZiwwNc@kM8h_cBuk$=unT-DdsmKy#g?R<^fQw!KL!8MxJ}U^3cx(WBn8IcnFwMbX
z6t=C0Z2(vfm;%ECKTP#1aqQH+fZ>$)7=^8<t!M>{0NOx9a5)2}Kw)z*bq=Nlo1Tm9
zIhYGznS*U{?Md{NXqd?r!m448s6Q^r@cYZgsM(l0@C=2$6Ql0L)WO{F9Z&j>btq^V
z<l~qY>>j0(e)~70Du#H%6gvQO5T+sxf&gZUCc@<^{zURHgFMU<no&?QqurPlG*dgq
zFgR!&@D-WKO5Vy1%)0nxkhtf{<y+9_G@Uek2t-|NQ|VRf8Wzr>u=inveV90Zm=cUo
zB8U%RghPV38Y5H-;u?%l!^cZ1i<5U2u}=SweT{au;&eKaH=~JoS$eI2`7LEsQzd@C
zXFj7@fToe5xf~7dLLoJeDGwoKTs7FSG0t;09-9gxus#G8V3UMxb4S!GTF`8?U9{t&
zJ&-*bgv<1Nqi`@c4(1K6u?W&SuK8Pc8fv{zd7%RkxTV^oRQ+d%qtRL>V+**4Ff|*9
zaquA?4=2UcNZ%{kO1+jiEGd+N21?RN(g&J{1BF2P<{&|~ux8rAS`cz#x(Xm~rY|6;
zd#Xc~csD;h<+}-Rql(LZuyU8B1<^9vxqHwAXvVY^K*g|P6r<ep3bZL~x;EV#=p5}=
zg5qM>ubv6@>O~&R0@RbOysbO<-)rz@OS1RKhiU)AaR=K1$l2#zgu4Yfyq-vyir9;r
zZB(13*nbk(lM&0ulM&0u<rtwH6XWB_2!>c8uqPuJaA3P}vWj+(44SB+v>_diwxCw4
z6Tj>wf0*_0*E&0_(NW&}dDaIC8$E4p@Ogwwtbz1E9q<S^V8-B#D!E)+7=NX3nz3PX
zB@79#zpcKsYt~o5ZhZPj3gKZ3iNdz0+RuX(P}mMs2La?rbre8OR3`!COm!AOE>sr*
z<VtlFKyFkw0pw0~7eF3Vk7<+|ca{>pGsAL57RVGie~Q)Y@F%jrxxD!_V`q65Tl_gF
z9EU!yrmr@DMQu~9^+?Zq_gsj=){xU+Kt4iYYszUd3Z^nR3Y#IvU_eg3{f+HHleS!0
z)RXjlX*#gRj@~0tm6nHH&}Ts=2692%$4wyldzypka4>z~z8puP?AN;|j42eh+5$C2
zpk)f%Xs(eMBq_)rG<!Ns9r=UiAb=ccjsnPu<|KfeY0d)3h2|oFTxqTX$c^SEfZS>B
z0?32rF^%$E_l)}|onp><H0CP?{N?UES~U9u<?K4*o?0mModY4uM-Jqs_<RR41DV0V
z{17RKxw(nHz%j~T>VpzWA>6I}!F{l;5g|+_|Ijm>R=}J<m2*kF?4RGag}8+*^hRM@
zi(3mIJ8?S!#FAvm^EFpXt`<Onl7Rv!NHS;|9f+E<_szV#Gsqy5Tn{t+N$*TWVwT=%
zBw*{X_2GDj`v|fW_c8+7++5L$gV{l-!$F4BxR(;XS~jMejje#xue^6`9drIuH8e<y
zPKrJxrpLvX?aLyE%%AmesCt=|?Ekbh4;|(lOqPQw!2G$`l!Gk?e1d&xj(g-0J2X*Q
zQCS&C=BryQn&_k}O0zyu$R31A-(28OqOy(|$&Wo|G&4m2r7^c-d>GA$Hsl9L3?oJW
zZDwqqMm1r5V>tvy=&UaariOxqd-voAeXvxp5zO>#UE$8PwOw10<y8z-OhBpLu84hd
z`B2x4M+M|AD6jqY*7E%=1~Sg+^c#Zq@`M@lmEKQpXWbr<uOYVhHkJH7DZYBA@%}Np
zSCS}fC)Wg1<=&jQ??iy%s6&*+y?2B0P}nFkHNgDn40Sj61AE_`_n^G`NyB4g?;v`R
zE_`nKaSx+YWV~s{WD^>orvmO>9)@gtrtZe-?@dJAaxgItCIvg|*Ne|SJkKEFkbp&J
zF`;=!g3A-<FVt)3Gu?+C2z&yUFcdP}rvO+dw^Kz9rV5?*Uv}uGz6jodTBqxz>jP=Q
zg>Oa!5115Z&b#a5GwUvgQy;B|`z!(0qOfBbv3g*OXT{!@QO1g_8Do9Wv1iz1|CKYd
zKJDRU;-5qq*X;i%e}%g}TF>M=3g1Kd7&#x)fj0xV+aN{mYYJa&C#Jg-TLJxmI`Y`J
zkq%n|wl^>`IzYK^EByYLPcUYm01GllG3J;c{u^UX2;wF@^B$hRH<&$mX0ITAfoHxH
z#BD;%2ST9#iRFF2-k|#RjIR1Dk#8DNvt?{X#V?OazY<%C12Zb37o>=_=oN7K*%t?=
zEJPI!tO^1g<R?I~(``W$2M)l+#fHSyam^YFw#DHR`PDRjHHlx1<5y$&)o6Y-fnQDM
zS2s^rh1jBY`<+-eZBnAisVQmcF`Lo<EG8u+M&nYZb#U?gYAU~)%&&$-hDLD5_lWco
zBFrHQ&lMq4sPjZ6#Kj!k9o$?b?d>!rNiI?gq-A8~%*-v9TUoEL(WB{Ds4iQq^iN;s
s%hMH<RTimfFtoIFnfeAxO-vPy7cMbeNt>^3<m8%-;T^qgoiWe<0xSVpe*gdg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_fwd.cpp.5B423DD33073B028.idx b/.cache/clangd/index/test_grouped_convnd_fwd.cpp.5B423DD33073B028.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c8e54471bdd88720dc4768642eb4217bda60339b
GIT binary patch
literal 6280
zcmb7I2~<<p7JUhG5(EMW$RH7k5Q37#jD*Yy2{Q=Dps0+2GAN)0#aglI*IKPoM9@}L
zaHyh!BB&s>wAQ%}(1Jrzz^M*U>bEM^{+r;<d(^*JOV{$4{mwi0-FwbH!!=5yQhn-(
zAo5gAQBGRnJYxhwY{1`w!u(7I08y3*QnalxHol;jXZ>P(_nv)ca1Gs+(f240PFeX8
z-68j`de$7PlWlGhxIa16vpGlRK6xwUWQ^V3w%DH4YJ`lGJvi@&yErsdb9VbF*TcqB
zFE2~2XxfwUYejeW-epHkN^VyNmemn=otuBy#Kkoix0QCL$;zg=wOKe^tX#c)SyFo0
z@X5$U0;lcIEzUL!^l3Q9MN1uT-|k7>!~N_@a*0@+=hSy}y8Z6jlu|xF{G%h1t&B?@
z*EALRGdj=I@B2P`RqU?U3)d7*s3Gs`m}Rnj$=QKv`}#YWC*ORw>PFQp+R=~fKW;a6
ztK6v&Ui&YhagS{lr63}E?Rn+3gp|Gkm#5b%q8v6B4Q&?OXK;@tN2;?Po!)W&(E-bl
zHJff2-IhPCx$1kdt>o1U`<)qITp5pN6Lz=uyI#;TU&*?x*Who-7AKW{x6Pz;pG{n1
z|3%wG_n8-&;rRzsswD9Up^f=A-c_`5LsG1oeEH^gt@XJ-?tJJ~6_k7ZPDy9wI-~wM
z{CF=`8twZ;0nzo-%v}}b7UL4`SM91%?cGuvRCl`V>Vz#UpUN+{HE&a=e{TJ_BeSli
zAll}vCk<C`zR6otnpm>-wsYHo`gZopkmGB|ohEj5FvRY}9IxxXu`J3j84uhuIux#R
zukQG|jOu;XTa(*m-)^?wI}B&v`NgcyYoGm<c+lsI+B2W-k`J2p_G%*<hG{!ae(O4Y
z<Nm&ott(fV91)n_P#x~&Z}te0f3s`low(&bs`>o|*~Pn+Ti?Vab90LhWGQENmJ(B0
zTZVa2j5)La!&s`|zwxQQ-;u83B%ILht<P!C>#bhV)NuH~)Lm<1+@*(tr%bppsW0uQ
zg4SCjI_BmnY2O%rrsBEl<IcJTn>Li0&q#Z@F3zWI-uaj77FXPgYS{F)-yQd8bM!C`
zS0;HH%v2=oTThKUL9GeUtT?jZwu|fT=|Lyd!>e0YxYaHB_P*vo??els=Wx#Db=1z<
zm3!VAWfUhk&hT2-zwF>(@ujDGnj<6QY_&dV^*)VfhgQB!%gK9~IrxPATi?R3YdX&C
zW4Dy{ENTBuDkvBDrd=w||D-*gw-Z;GN2?~E+S8J;iFm(To_(|aRC=&&$<VGp*vSW{
zB}J{U%<|fF@{WJyaCMnkLqTU1KkvXgFZs#u-X_i;fA_8O&m(`HpWD=5UGCE2!N29p
z&-PqBpmv)$WrNpAUdn4`RdVd{$)~pel=*sDW7dW2wk7vY{c^1575j=SN8s^7NzI(|
zJm~k9fY4{x`+a7F9;(PZ%?~hD@t<Fr81a{L4}E8h^_!;CPD`VD7bGZ-ZWryi`(tpW
zw`TUBMa(L<g4E6_(H9wSBh^(683P;E{8)PHi+@%$eEjUt>b{h}e)ag{^(Q4gCtfVO
z(6phaCCxr*hwAmR7cPTsYb-_Ca}RG{es{Gcsmbz6RM=bGyh9yaq@Xw_*~|z*aPf`@
zg>6%fh)4l~XlEibt@Q%YRT%Jx&~>6P<5-fT8GaT}lT1ky^kPJbAXfgXAN|nq>@!Ow
z52y=~LV{jE>+z~TN%VqJPa-}JsM)4$TfKnRjwa;p>RIdG5|M>K9f!nO>jktH;wE3a
z|Ji>V&G2)8I@C1OMK7RrV7&Km^Ygd_B0eDn(UJ*dTfKnlkn#aV{NHja5g!B8Rs<`u
zUO@Gtp&JhIbEtQSNIqEny>ox5sF%)aG`~W`#{z3>f;Cw$pmq7PkfE2;>jsJVRG_vY
z*x2d?RByH1un-9?F&~Fd18O_E9ak@)`kQ>w&7#F|0!t(Zd^-V2Am{}=cXdfRDZ$}x
zx*2{hP^Tej1idh*|7!iGiP~F4MCJo^ED~$27f@Xjd!&n8@pvT>KNG0^O#N*20)Bh{
zaj(A}vJ@0E{5+tJGL3T43#e{C?wiV1SVWuQlY#oZfOf_HxvjMAk40wq*+8u}RlDc~
zw7&5u(l_tPaJLyg38>#YSdU40zTn65T5EhFaEpV514%ESbzf?G=NpqBxJ0B7SieVy
zz9;v~1`<l0i1>J*7MY4{^#WS|?%CnxP-Ax063GSX_q;X~y+7kkg7_j4Spd}Uc@4==
zOD{NQh2{euKavOUmN(X~2)_Zd!k92B8%SC;RV{!%B56aZp#n%^k~WGOC4k<Cd4t1`
z<jcFSSRZ~B^DE1l3?A#82i<8Cb4@$HhIH7Nq!npI;m{EztvE~^jUq|7BnCy&AZY|d
zFc0yBQTa#G&h(ziftfJAcF;(CCg}RRJwYbGBn1x9g#^0WSif<24yQ2LV^<f)#dw_F
zPmsa*QOJFjTr;WeUM@I>j4u<z&5^Wnkz9cyg;=3Pky4<PKt$QvqaJ+b!8S712eA5G
zA{gr<1&?8@-zMbJ<Y($vKj+#}a3rl0#fb_T=<_(qIx0}|3D^odkR1a3L(+0WIFTsg
zhH|4&#8dM^A)4mvZy%G~;-XvE+{q4lW%|OIm;cGh*bdh96Z^?wE|Rni1w)M@rjn^a
z5lg~SK*XihS=!t=AMgQ3gbV}0)zwJ26_PfMm_~t-5K}vru=;l7fpvAvf)!{qtw+aH
z8uIY#uRQ9P{)>!?vgZ#Y^KF0Ak&!RLL{TgfW1=V#OE6KC3Z#Z8D(#-y?65Ujtn)pe
zErK2ez87)iDB_B_3Ka1KJTXM7U+gDSi<l<?*>nX85_Ke%AcEZictzJ5V8)WP?t$(R
zkOsl$Js*6F505h<jg-)Lz5{sy`XK5^Ebn1k0jaO+Dd@VCKuS1Vm!zE(I0=J10zC{U
z%IACE$_QIy;8tCcf&rRw?}tmp(T`rd3p^}WBix2fdW73BBcsE}Qyy*<ng9;kS=d>@
zrV~4q9U&rSHsg);)P#F7*3!&rwh+ZNG+qhV5;UMAD&H>v(hm?-#E_wgDQ03w%o4Cf
zP?vD4<zDU1({uh+EfUQe{TJG~=pJKFH~Si|*u(e6c->Yhb~L!vq$*r<rHj!x5J_yk
z8(sa9E=v!5Q$~C-=6U86ZOfb6+R|r2(lR+r%z3yRZU7_>N$bz?#~=wu5&&xpNvq%}
z3}{}6X3o?}Wy52BGc8owy1agOGWaHo$>KrJkhDA|kB5F!$P{9blquz*-&8V{29$Tt
z^Y2p&O25{HMCpgA4~!s27`iw!m>G^DmWmYw(ZUC2X=kQzyT;r!ER$pQw<QIQq`bAu
z#2-*CR(du>E?qx#AMD=B*@_0!grv20wx+>pilnu1wxPk{iKHbvlW8zfR}U&(_&=|k
z0KCcIFof_>pe%4WLe!fAjsSzi9I+7frko=;p!H#2zKs9)5OYjtz<q%tp&RkgWOfU5
zGbHRr_k`yWA<xRkTnU~DtfZR;(Um3!PBfI8%3PDOLGt>eAPaQ42n$RV<LJIcD;T==
z&gb(n1yLXpU<#tY*dJ36g#w|WAi{JqBIw6V^y5zUPVO+{Nm}O#&Ylp#2Jo)9!>zz-
z1lXZVU-)>WO~57(`!0A%H%Fl_VLJ<qA8?3n28AzyLhBvNfh*MB97tw5-ldv@Rv=+3
z(I|72oFEi&Roq|{@uWNziue*f#!jFVVC?)={un!<RA{i<Yc1Hkf|%ksW=!F*29BVc
zlhGp-5(_D?&+LD^qB#EUw)J)(Yw3Orcx*soIs=24bS4I|=qwCk)7cF4KpZ*;gSd1q
z2Jz@T4C2%I7$l$zFvy?ok3mAZ5Q9W?5d%#cF<p#761oI~q;x3;$>=f+lGEiFq@XJ>
zNJ&>3P~9cbIXA*7%9yUM*Zwx}xiKvjH@ZtT)->HoGF_ez62{frRRzi0c;u_K^BTHJ
z!DG@i5)1OZSyf(iQ1YPb1FBZ*KSJgcM%5o<Zm#X8_6vJdPi(=Ipkp}718j<rMh)_i
z9Uw(?BMLMb8L;5|a&Fa^v|lY=p8J57|8w7z@Wh^*i@;X=D1KCwZEOmgigHLq5m8};
zuLoP?ZxJT<gB%k{MS<`!xIC!DAt;grN<vX2l}Xhol1XJ4JGn}Zu~P&pFm_6r(qISg
zK_Go}7e(j;;6!2q1zv67zrayg-;PYFa6Eu<#}V}RMHug%R!z6%?{nAa>SiDxGsFu*
z1d*sS{6qaQL%dKeGz{@~U*^^gOJXK~1!Ni-=3XTalVk2xif{$yUZo6D8tzrS{EFq|
zv$3&b(g$8ez!9;7<mh3$uOeEU%X3fs0Pe#00RQSP!>E7byyH0VqxXVtJmhaVb#%-f
z!fWljp$)DCXV5Y0{Hw(13(eE(o_@f_hix|cXV-r+@Bt=oI=207dfBx115Ch;K6~EB
z;4F|X2)rm69>4-`cmOF2-$Y<Pgl{4+v0(`^tPV{P$Y`V>Kc!%X&X*a`p$L+N{!BOg
z#A!{<yIoq`=Q*2Zew3A+o`GZ=bjU~O&rI}ZN=|Av_#ZbUGG~U7v5BdfxdoAANgiio
zZKa9OgooOyRo=FQP&+64@e`bd0b+?%CRZ>VriwfTK9v9W)X~*#;v|pBR4*FccN&v5
coy+%k=ehW?105WvaDu}Y;e(=;A-E{ye?!>RjsO4v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_fwd_multi_ab_interface.cpp.D44C4377C972C730.idx b/.cache/clangd/index/test_grouped_convnd_fwd_multi_ab_interface.cpp.D44C4377C972C730.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b39ff4518cb96db4c6d75ce7ac791e138a2b78f7
GIT binary patch
literal 8800
zcmb_gc|4Wb8^7muyEoT+t#R!xV_zd%lC2_3MPy5nrNx#t7q6wH#nz&xnwFWVqzqEi
zf{LP|jWVQ+5=m5(8YX@j)9+Z0_ci9z{9$}P&ij7fbM86MInVc5&a>R!*49&rMq3bQ
z&kGBRSwpAMXd>`GIwopED+Ce{@Qz&b@N$)zEByHTr}_KLg*}ab%wWnybhOZPb2?m9
z9Ouu`3_0l~XIyu-?Xj$+tPL&Kt?c#*i;DN|3gKc7{CURp7hmm-QoDOXsdDIQHUIp%
ze;qh_OvtWn>C^RB6HBZI`@7ALh0n~&PgF3O$sCf|S|gm>_jaMF<y^UE#Ek=oe{wax
zc2)A{_L!&(ne^KgOHYM<4&%+QC^ZkLo^#Ut)?fnBYM|=1L95oGUS%M7Ptzv;oXF}w
z?rMF{zq09Z-yw!EbDdS?>TeD&$_qM`_0XkCroXNDOw;|y%PrWCn~XiATctbvcRgHL
z{81^~<tirjft}?xw~1i1<e!LsoYy;}bcOnUjfQqz)53$lww7n8sRX(GARm*{d%HpE
z*gpPQudefgmN-f%SP*|L%w3n=z({}0E#r5H(GGdo$rqGqx4Thqg|N#0-F>I$N%CLh
zyTNx}W#$7pq0_JDoyz73%&I;cTDm=LUcdfnmUA_ev#6wQKs!tD#F@sfL!X^vRbBbg
z6^>|MI2w7hR_0*(&V4c=?)$TkUAk{%r^R=nLvYST)nw+no-Olk<%!mc^E!)nXe^Tp
z&bwNZ_0VVYw}vX#kwwmtX}iubW!#NoV$u{}u2R|F&3C@&mtp4>l9?vY8iI9SeSVhs
zU9|S`oEO*AR&w|5?cQf&;!{8KkyMsz1FLUaah_Z6`%6B)-rLiyg+6U4F=Xd#%h;4s
z`Y`{z99w><x$K<df&;CGX04a8uJ1Evnq|fd=GfK?bW1%tQGcehY|n461~@~Dm8@2T
z6c&72O1H?*-0iLT%!A`!ZebojsJBJLbFI6P`YJ=uC&le1kuMx03L>uRiY-)FZ1}eE
zPS?BCOam{Q9gSx$7ujv~kM8{Cj)sM6MY(>?=QCnGPAm28+<Q%qM(WqJEV_K5z|FKp
z(d~cYT@^}=qTbs@*@~K7k8HR5FE2Zxop9jxgB!`)iUKo2<5EiIT#wxK)ZD@A&+QNA
z{eJ1rof6k1cY2}ao1?X?5CL}2;QQ2pfcIh7*$dl4Z=Siie1-I#-}4BgmSqvrW<z`G
zG_|8|4Q8#8+I&wxj?bl8x%zJ15pnV1#1fMych=qw%qvLe)~q801K!zxk#=(G*IFg%
z8I@(=VKg%e6Hs~K_|tc%U#qRN-(V`O_lmwwp~ju-`$K{B$DRMX?P1J_N<O~3%Jocc
zpC>yhA;-LIN4wjCH<g)Ya%LYj>JGmZEt;iLmbJ}ikJb$%JWts}I6OEiTg3cBRqbu!
zb$_4Fru07SoUQiGYqp<0e6a6%j~z3q)G+M$>(v%L!~14c^?Zr$FTJ$%c7}_6@<=w<
z_;Q4!SN8J3i{x0DW%o{SF8@#->?BteailNf<oq|Y?<F_a1|Rw@+PR|1TS&EQP->s)
z!b?m|ot!MmeeP@a-&xC=?W6l9Zp*S?kH_AB{4hMi%cZ8y>C^|4B9s3`tS_>2&pt?~
zK0Vs#G$@)fVDu-bF*_Ds&>7c~wrSy~w$L-?g^MnoyZZ69QvAc*&w<gaJo9b5hS$#1
z$aAlmB@$9lbcNO0knU;}-dIS_w*F#m`Q51+!Dn~7=ILP9{#dJ^{<F=@)*p*L33tD@
zYZVS%|LO8j?O&fwtCnpu-#jE^TQ;LA%~H}yOE!O{)tcbQfx=6vGeu;Q1%_vw{IZUI
zS^alQU6%E`==d=Id(g-66+Ky!QC+xL0^1ltBN_p%5irIdv=M&@|5&!Wj;lsB{~~~`
zq0tB_hLqrFGuBC?$?eS-Rma5N2w<M;X@n+AQ*^XJ-CT8z;>SB2mJ4IPusei51Rrfs
zx5;G(>j&R6GbOOlFdA_m(C)+a#~x^X$MQFYf3}oINMM0MG@=vGI^mG92O8pZ?v?x-
z4^JN^76L^q1dKwUG5$acEn7eQGJD8_#>CdbAr*j80W`-SXo&Od?Ax_UP90xa$Zf&V
zzt$GsDH6u~;coo-{RKxGG=o=Y^YHoZOgjl|LokhK1+-Qu>evH`>V2Uh-Ni1fT^#d<
zGpH!4sE#&hhOjF=t4)1$F(wuc#k>d@7vWCE9uq@eFwgu^>2aEgZG=PqQDxg`xy0M+
zoJuAZ0Eg57Mh*C8{DEfK_WpsZQ(8V<9P@y3Xh>@)jW($JyIaE=%I^Y>31F*Ww>C>#
zY_vh$N1N*w>}~3GW@3SGfq&G=$=CSlE8qMOB#y0wGpH)6s*W~jhCBh=%JSz2mBg_X
zuv<-9O=+}2-34h)hr1Q(tiQsG68axu_l;_M1+djn25E-0$Y_IRxTbboXVt(fdnUFH
zb}KRzbw?Z2UHPP{zs=G2%~w=a^{M`UO?n6W#W61^qOzj0_Gp7<Xwd1+iq9x}CV=_C
zZXK46#At)Mo3WeE%x|j631U&O+gHU`Z?u`{z8BKF((v$yuNYq2)o)W4dPKftVxe#c
z|7aWC5hce}Zr7as3T1m+&nLKx4tp`N4RDr!<mB(c4&s~CB=-wokx*1EmX_FPgGBA`
z(gbHUd1V5aJM7kD>FJF&sC&Ts<fWu2G5)WpYVhszJ-S_fN&=W4l!3$GFh?6S!w@ew
z`eV2D;8%1pY`bsnu}H<huN)UGDrjx=Q3=wg0c_kl!p1Fz-!!04YV~0roj!h=As=TE
z>K*J3`uM<b?73RPdS)Aa{G49M9_X_&<V&EBQ+y3WR)2VI$#LWoXRs!@LYyc8#wEZ4
zxg(q?2gc>V0vR<<Q~~2EV1axQPE-TqY7&dAXjmU{5-$Nbp}k1k5!s-~uPcuh+ti=M
zz_=L5`mn!YNvQ$5Yj^ua;P2HJt1m+)hCDPMYt%C%!@vmx++Y?`E>46yhkGJ#EieAG
zI9Fx&H9#kP^?YX|F3amgT7=oVeugJIX7flA`M4jfYJhSVTzLzazXe#MiO>_C)QImV
z*M0;P9s#B0?=Q;FDg9LpC!PX@r$DLa7H-<yKAR6G+JQnlP-?QdX?x}XEe|JpfI<&Y
zLPw#hi+A~nrLTg5j4>LqTs2?Q@OxA5VJJwXT;wveynb-kx2iTA&nbeV_yaq}3=A2a
z=~Dx_<8`9)Q&yOJ+Gc6VPbb2sOtmn#zuP5B>^WRobCISaI(3IbmAmTF-m&R8Q4GwA
zfhEa~QeajJEJ=1;2WHoSCCQE(!0ZNzQ5>AIFq#l!{VnF<a8JzpKojk0iOJrL5;|GC
zXW)disq_NmKyl(6Fg^#Y(L6X&28_#qH7QsvaI6LHD5~H@E%2-b0SGTWZRA~aKrnsU
zW`%YRH!>EWk%PdYmYJ<U3uzfe>M2LDJhoJAzl*_MsDl`}7-yt@L^EhXlePSJt)uu3
zSfg<r(bfI;P-hE)4W)30j{yA<5JD5uA#dqVfe_*?9rBjm4ulYI>5#Yd9w3BxixLDV
z_+;Lqv;f-ZNKSwj#)&#0TL%=;dN}bM$UX;(X#Q0Vrt#nCv`E1H`pWn!BRgGPXI*$K
zSK!kOI<XdCD}i=3UFlbMzP6{~u~*<~>Z<Ct$iE=3_^<wHjg!5vu2SyMe=Jr|8*Ak|
z(Xz+TSy6Yp{(Xs!;ZV;XP}(dYnFZ7l8<5xkS2l#cTPk#(A$bpKYcg$+Wk8}%&dqQb
zNSKX46Ga^K#84<gj7AX$4M33x$~2bIp$_yzgu^94fM5NQEz>r(NS;N#?A?Z82A%Me
z@{>l?fk@d;g@O$H=1|alKMM+4>}O9wu6}L^#eBB9a8l_k1D^eC)!EBYkcoZG47#mm
z^5j1zz$V9q_q9BYo#AR43p7@s)Rwp9XG7nXKO?6Y8?U;(GlBl-fe=o_vExJ$FK~j#
z=21{QJD!3P*a;Mr$WEjnf=y6R5<7{4lG(`=l)_G-pj5z41r)psuy;{$9pJA6f~4^`
z1O8?ph@8!cGvI#?1e+#;WE<dbBXMH%MSNpMisiJavYG$#*_HS0`{4P8%Y{z_)wJaY
zEc#wCqv7XFAejkt(LtiX0L^3U?txg3!?c0f36xRVfHVaMj@&><uV;l@O=Kq9%Qu=2
zXl!hP0d&j?$=SwCg~a1mt?>^dtBqV1QUA_|08_tnK@-Y5;wRAOOy>0;cT`2CVmM(a
zY$%SniW6qSW)mpWfp6<?yS|vRaY9*2nTxIlIH4k?!bQ#-Csd_Wx#;2p$;FA2M`r@b
z#o<v<JSU!l5;zGIl*mb>Ac8|sP!cDJf|5DO6qLeAp`cX2Nd*%)XX)-iueh@Y|FP-i
z+*Q%zV+D9X8=LI6_He!Xot@tA|Bu((_dsOFcAeKhK^+DP2XawlJIK5FlZ*M?)+u5f
z3ZIetK{V$AB+ppnMp_@qI7cEkiN_*0iN_*0!YCPs0ea*%Oj6YdH=1Ax%sO0RyphTg
z%c1k1Ohj~@Km!nC;M^nkWOD9iAlnQSNqcVtvTbDWNAVY0&&d6m{Fx<gIbl7gh0VTA
z69|x;cItv|<WquSVO&pEe!YSJwB1Z0@5&74u;aE+w;Vx^1iCa9J+QbV&#F-TkC}wB
zUj}PZ*#IYWRCH#eyZv-i;}RX2nX7g;?3?!4oLm3AH<PXzIBhjv{CQu-n)6S#z(ZDC
zplF9kQ);`PS8VnA$+QK%k}Y&BRd{y=yuxh-3R{5&vId;k1{AhY@eaV<0W?uRk_6rq
zM~XGb?r@?Au#12Zc{#ZR*q4A2O2KfV60j?Q5xP5!Tu#`vzzD^lYhTO_-syk$aoPum
z3Pw}1iHTUi8Z8f@ec*&XuFoN(NgOv$3+)vK11^t(;<@n@l)z1(phRvW1rc0=f|9sN
z6qL+Orl1sV3I(MCZYr3-H$@-j6+OG!3Q3R+B(q7S+_nq4DJ<%)KW!S+N~jf{)4b^q
zX^;*S(#aKYVmnaSPQ^O`cPB{$WO>jRj+E5M)3m@XB+1h}1vF0qV|1RsIPymwH23d>
z`_WpY<%sH8zm&iH)Rn@Z!Os~$G6QH&jIUnYZqTduY>F>KsT54ZC*RNheqVe1+w22F
zX+PeAS7|9zDJxVEnr;OM<<B!&qs2mU_Zx*bN+aoUB1$-lf?|bZDJTQ5GXRGin+e#N
zRGbCaSyY@2*x4j*9xqEy38bio0E5}&EDY7s&;V33``06U@iNtFk4VX{P&*nP8f#E4
z*fICOX3<)A_h}n!hue|5`~qhE6eYiMXz8ANu;RkBS3uQQxWeRe7>d~cdYgM&mNyV@
zuqSxRaidBHJnqR!Eh<f-0f_6+UPcPHBxB~_b0~R9q;sSf;smlHm^X|RbIG6V1(JKg
zd^8Nzb*G%)pwN+b(OCtJ({>Ul!b7%?)PzZw?}R(Ck(4;G1YbfyOYx-?<chmekQ?qs
zK`ZbT6y$^ZOdwQnhFhC#=6^OG8D`WmI;OofRIi?*SXiYFT~h$JYQNy!@@5?k4S;f3
z6ya#dl2AyaAzh=GM?=K`6y#_qlSI*-hLnn)G&K9f2Q>S{2Q>S{2Q+(hR6z9Q5#^);
zq#_zEk$jCKUt`GESn@TRe2phxH&bTYH1U~_h#l6`owxP3IKiZp<W#U3#KiG7#U{ol
zMAMQdM$vfWYXbS2LcS)EuK{6!p%D<egt_w3XE2yE`2}#csGx|jkiE10vL#}RZ8gML
zOJ+%lOGwL@&$qO)wpp-HpQCG`Vy30|PhTbF<Q0^@QC3yg)Yj26m}6)%cOKVNcD9in
TM^eq$ap@L-Ik+uy0?Ypo^AE`>

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp.D8EE1DD81FCEA622.idx b/.cache/clangd/index/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp.D8EE1DD81FCEA622.idx
new file mode 100755
index 0000000000000000000000000000000000000000..621e701f41e288592d53fb4b136e2fdeec7a91e5
GIT binary patch
literal 3526
zcmaJ>2|Scr8-KmVOlBrCjUmD?W8X7qW+Y2QWE-I<vW;aJvJ~3U;t~;(BBk$=r4W@=
zQu$OjT2yWuMV7S4Xrbcz-kH8*zUFuDH^1LJ?{m)oIp^8_=kauMaA=(mL3W`|$vjSC
z5DJ2z1;{@kF+S2CS*($PpyZM(KF?#GGl`GyeXcq#r~Ncqsg3ZP%7u>MU=7`gl^%mD
zX)bww0q>hPT($Ebcrt3$9)ne&hwn{DW|}I~sT$Pw!HaYkVps}GZ(^j*(SVxIxBmS0
z>%heK%mBB2;Y%}(*G(k0ng&@|v-;R2rDHnhsbLomu;Z3qY?<cI$5>O&yqW$h#@Tqd
z@c~h;U`JWk==Ar2xGhOL?TlkQKIz|><Tsmlc6WQ_9?Fp(9;Bb{!WwXnRGrH=8mg$y
z4k?CPpc>mNFV(x!yl90-YSaP=vop&II}IA&NfNnH@l|axYGZDsiFPu2K)Q0x@ntzS
zxMJCV(>;tbdlo6p76ty<TF{blz45|Y6ZeY^a#?PktK5}ODc^jSi`T4Lb}i+2sacEn
zr+Z~rcTn?uUpJkkn$N4aGFW;$V*Ny6qwNAzoc?Hv6Kc@ll#GK7FV*xwn8bRMJC@%s
z-i?4AQeS=5f4$+=D@E28Ux_WuSjowbmcHGWZuEM^uhF`@ruQHp`o_?}ap)E0TQH6H
zQy23-x6rTU^)xr=RHb@#23IWkqbzW~FI~SlSNXK^im)u=r$;Azx>=p`Zfok;29q>a
zlT?!4Hz&nTx9oMRi@T#xvohnd=TVD}eYch??N6+!DJC9mNYDshYhQXjP4bexhHh?w
z{i5@XsX+6w)w$SDdmjZ?@4M0S);XkM@vW!2{JrfR4<h1v84@$z=hGb;RE?T^BhLo=
z+{`||^bId($mVSBNTAck+FgA_<@g-UdrcAPFEN{=_U^0pS?hf|UMkx5kQJp$p)MxN
zWSb6aim_{WZJ5Ino!*%GdR|EE*Hp<^PLQv^L&gqu+gMNSMNi?^ob}Sw0jKR;MNJgJ
z(W=b9pD|n$_4<vfJMDHsKhEsSWv7lQIBH>}uCJ4|g>yz?ROj|WVt($kqGn=wFvYc=
z>AG^Mb9!S#Z%JuIazSFAj623rGJ5iG7rS%drF4qh*LNjk_v%%+1793Vh9ZMpPN-#G
zRT=-i>jUP64Y>#P-r3COcX_i7za^cVA-J8ajQ!GdonGYSQA?|5RpgxSWruk>U8Rh-
zxOp0UdbB7Y?<!h5{iIpGg(0Uah29#3J3Jt<X{+%r&5iReraEcW7buUt4Xx1bY`mGg
zJnZ|C)2`#UKiK;a9_*<aG@2@sW{!n*Hr%m1)SS%LnzYp>+*a@Y)^k>o5n7=iCtKwb
zwk(eyMftdyX}$9TuJ>R$BfIS7p7j04Y+CIxH~kX_<aQjZP~GI`7xF`;;K!MsbiUMI
zXw5_K`)&A|K?dgLO&7NMJ)+9FR&Cn3PVQ3GpNsn2Rn;Cm^~$e2u%19}uc@A`)E^Br
z9m*|Cq7>)Z);;5-Eo)2J%LuR=-87hu?Lu8XIPPbqGokvxa;u&z!SG>P!nWNCLv8u}
zCY!xqAH|JpqRMiN`9<Z~ucb2jnNa^@T(;imG}QyGn%7n>skW`{WZbS<(tsr6+EMLq
zg#0RtLMS1H7d$QjL8O=*OAFomU*zGyAc)I=7-*q@1k(ul$!(810>Aj2ii7<&KwLGf
znygU#1QGap@(ix0dKd@$BCrZpMNTLHOp_TQ?;N9U!@<D_tc+Eb6AA!psyL3+J+~ub
z;6UhaeCzX9T#+(R$oz-U*33>$Uc0QTje~uV%ayQ7WT60;C#L3%hNc(>%fV|ASdpq|
zAQS*TJ`z+LUT*vt2ZtcADppliC;(h})y(yUbjkz~UXQ@W(#Ck90B}9(3$N4s)<+z?
z4vCzMCNB^Q0H31_VPEQ6tK(pA<TMRSBMAk7TXOnR#-H??<KQ(2Ou<q}LNN!I_HeDF
z_y+%w_w|Oxx_xMzD+Ufge9NQdmk0$oJ@u|H&uBQ1hJ*bPn1m)_gaW|voUnxXXb6&g
zNckhH_{~EJSlDmS3TbA@C&hrH$r32xcA+fxTz0<kH+<@SQX-+kO3ZqJcX!0bf8pyT
z@LV;fniU`m;FOrPoL0?@bGug>Vr{a*Bm|TgoPx1JYJv3dTrIYiAE+cC8e$}1naD_w
zt+yT+0lFbZ0<MXSbl5s`Mj%u8XX2nPKl243pq^L;*3WFVpRp&0BN@<4X&%6@eg5Y+
z?0DHeWShdISb_Gs@Y@&9e?cgPC4MRKuLXBy9E_DW+@Fh_B{Ru(fLLy)%>7Om$?Os9
zgU>{?hZQce8QEr`SOnk!JOw0i!8;HouFck72citPYg>CPANN)4Y><o71|wSj53y~Y
z$GOyW`v7IJ3dAT#gaqCzNfV1Jp}PC>!HmG|NEW6GO_e}v!5c*45<EarTr@Uqt|6O7
zFq{r;#uKpswkDjz$7b%d7W1^PCDw9MH})~&NieKH`?4@?Co-<b?}G7MO>fOP+Tgjm
z-nw(EEN98x9}dhik+X`5ifSOb9Jkyl-=sRaSO9ryexE&D#tOx@!ISvET`i80${#1?
zG$Eb@eF9N}p@Y;zFmXin&|quKMG1<B7zyT%$cQD+QUo18MYK4Sjtfx`3$8Y4bbsYN
zdazj0YLhQXdGN{|kZ?#cBzZ6*PIp<|reIsCVwV68qOiXVH9S|Jtv}Z_JlBA2@INa<
zw&C2W-cV_te{OeySQHHreIK(>dZA+O06&p@1(Q;gFC9-OgJPU(EXWKx`$RB1<X6eN
z1EXe#!sIhHBLeOS*Z_J1ULbO%V7dZpJXf8q?hE|2{CCWXG5pG`CH=k05x!WI;11+U
z!N>(2MOcVoFu>U8aOAWTNj@fG;XGU=9e>y|cn}F*m8oh4STMjUN+sR$@e%Vh<oxUn
z>u_t6*!CCmEm3*t_3w~N=nLuUpfFQ}lJ^Pws~?HE1TB`p{xbciAC=6Y@1h69tf6>L
zbPx^636vQE>4Gl=QcQ>sNjQL5N^<?7(D1FfCULkZUUW<>Cp;o7G8Bpdg@K?*ks2;i
z!$c}aq(+O>2$32pQe#CbPozeP)DT{1w7`hXTP=x_LSyDh<M8uIG7E?VCl4ogSJ{;g
zy0TbTvZ9>)A|=aZ);6}w?d(k!8d+&t=u`i0S3!lMN>kU+($O<8G^U%HF<490n97UI
V9TzHSFInZb5q9=k;UeJ){Rfu7=LP@(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_gemm_interface.cpp.BFD580574FDC4028.idx b/.cache/clangd/index/test_grouped_gemm_interface.cpp.BFD580574FDC4028.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5394c22e180a0478b2de59174b15f40645125ebd
GIT binary patch
literal 11368
zcmdU#c~lg~*1)GGEi*kd!{Q8xC^$qwkR`0L1;HSQY>I+Hh={Up&42-XE(nUdM8yq7
z4ImLgB8u`51j8PXXhdV&Bj5@q8a0Spcr^n>7vu8%<N40fbL^g9S9dM9Zryw97Cvq+
zF7pjptgk}cQes0F1rsb5OBw!8T$B)L2BAWM#Y)Mk_Vb_BWv$e^|H{EbnPllh%GW-A
zy!j!eX`sx$QgAZ1Si87-YtH5u|6DsZHq%i#b8pFX8y~HpKVn^BWue)+ThdF9S+|x|
z(RSMDWf#O|71j<mTI2QV|E-nl=*fB&Un%+`=dVBGl=!`qxZWZcwMl95Vk_$xFZL=<
zKEi2v)}63xjJK=nGTrel-<`L}-lz2kWpYHf<?mJcCZ@}C_LXm(88~B|xWm~XC!j%S
zbY8cPTb<C_xV3eC&aOz0yAcZA&!Zd40v=fSxpdn$D&9{GoK>7HO!|jh)#~$u#Z}JQ
zGu7O<jJwX)kEN_@dP;g0nf<z8_0^2x6Cw5e>_1!8T5oqG+%rmUi*+d%IJv3csm(px
z7sCx(I;-ikGr8jR9_~AWYL>p@)t0$Lu4D^;s7ukYQ&qQmF)h1qy?#Z(lVd9c<}-Wv
zx{a-Qf_+!?ryCu9eE6QN?r%9GZasE?@;pAJ+2KFF+R@f@^n;d4%MJNm5z|jB=G;4T
zHe}Se{<fCD3m*3!Z~oNgsZh7sfBS5^h-}3%CDass!#Uk)eey=NQ|9XU=wwfPVGuKK
zjqSIE1qqoG+1XR}PMMId5^%IrZvVpPx913x7S)Rqe@y@C@mdexpA`$QD}`4U_-cD<
zojvQMTd~eIed?N|+Gz%%2L)BZzf(b9*NcCck#OVki5cC?Pv5CbH=HCl=HW^`?uPu_
zV>iOdWrwbw=;(QAJ?Xebso>Ho$7|e6_IhTk{P%`N<R)o1TjnP{OpI6@+-~)vr@>*%
z<GFsjtBQ1tXSM}*)KAM;`mJd7v%i1om~1(^CDx_L=|NscQ@O(Fbt@Hb7F}Oo#8v8X
zjCwkw^7zTG1;@BMMrnI=$M(!5?8jVLWN|CJaL#|5{3m;8w^_8<lwDe0@A0$EbK|N4
z-$Sp@dl9WA6}_Q^!0C>>RYPH*Ymw20f>zDO6J^0S#yo7cpRW7et5?-~HIFR$@$SsR
z&LHLMQ?;(;klsZb63@NioNcw%Yqeb6lgdvyeb+(my5sE~Q8%Z%Wrn_p%L(Ty=q&bd
ztC;g^Sm5{j4i`<h+q=oAMdjazfBuuynQbP{HPFxRN<5l()m86C{{!bX@y{#o&l_<k
zA<q12{{3x>7G2_Be*Vbti_SYb6+82t!vEB_47C0G{`wu~6*MBo|2VPab;@xU@20H3
z8yiD3`X9&daZadDoA)$kywl@l?X+XRM*TjAdrp<UhfHpU-PS4iJ?7}8l!T+3PLCC-
z?;q(Kw;}6X=pQRyHyNHD|3>cF^UxPUopPbWLVj0Ya_q((R~(i!x|WpG-p)y_PH%qo
z)UhLaeasYLd62bu-f!ad@;X-@d~x;1DOXOpj_&QBf1>L_*vzF(H(&kt4s9`eNx^>N
zg5h25vjw+JRNMJqK1*Kgn?>K;@x0~b!sn|}O`rCKu1i*#*iUYpdf{K?)9k!dur}~n
zcjDsM1?|vss922Jo^#B8H-{a{V$ngYAUWxUC9yE@pIU-Hhn#IXNuE7-9*f=uc)Nh+
z;BDXq7E3RBon8OjMcu0Gp!qC%4`A&9a)UQii@{E(J!`LAa+hZZhOlTwo}z;Ef@(0W
z%f6P@HFXDv9S-|o$TF0ZUeMRpA&U;1s?rwnY=8K*fG1FvUeMQ*ja1d!K-a$<b};<9
z6tGHR`vW&r!{k{Hdki~&zpKg)4SUBRC$BMCMz8pxlf#aLH8`=H<fIq$wd<iy>l&e>
znLIlHwy(+4RFhs%`_o@^YdC3-aeYta?s5u#H?LZi3fW&n31)!N89-<7hH9AcN9ASX
z(aFy8?Ah>@G{8#(%7Zub%IrPIL*%DLk?<AR2Nj-*vh;$A=P%mdx^#AylPWth>K!A9
z`ntGeXK`-{>;wE-b+oFE^n$)l-X0jmU3tY>l^q`O4*4%TIQ5GIKLw|~XMh>ANmJXR
z%hB&?dv$@+8qndEF3+9=2baoARghj#+Z)-_AADX*^Py_M;^n;M3epQI&dONYyYo=n
zxeqAUo=L2)ygzpfdr$B898R6ITRgq(12r6+xyvnn$)S^~?BE6O(V!zofWvR{85A4^
zl%Nt%NkMu+-8{N+_JvKJnuq1tzOZ;CZ=|yHf{M%ja@u~cDO-`lUI-P8lB0B_7gT(5
zj(d=qj(g_^+OOJrqx`_*jstKFKryWZyp>Qz25zX9Q{lG`p6_X&{+@YgxTN-I-ho^b
z4*MI}hb3}Lbfg#b^%=pl;2yutF;K~21DgPE6Hptxp;yjbYmI!Qs($MO-RxMhE&0L5
z7-tSU2sZLgTkxXqP@zFm;7KSd*ibs)rNfa7+)zU=e@R`dyZA!*duH^_5#r^eX!Sp#
zO@hMK=jm%oFX-!p(6GddA{L9ZUiegf-ky{Zgw&c`yVz0I-*kE|Azj<!e&BA2mnT*e
zQfqgkH!0_Cm})yQ*yS6SBYmtLeV-76b|m1;nt-u&KDo-Nb{o&`cC-ToolGPveXlv|
zNq1I`yYlabR>le;s{niOdn5*AT-odaqdM5dfksgwGO?fW1`R8kDT*!x%7sAZOtaYK
z<yw83qDz2s3B#Jt(nPPacaN3njTS%6W*xV!15k8|YRc!RlW4?jAQb?zdF){%dJuTA
zY`(LHOqHK|ZcU$kVN^OC;DBL8Hh28tEtaa!)D>mkM=s$lf#_}SrcG!3QUMDjW)^h<
z`3EQhiE}{9g`!miDr(5|QM5oHP(#jOV!hYvZ91zC!I~rv1Gx~RH750Yc^W*K4w)F)
z34Irw8Y559HiV55QWc7JARI87Ma;rz4lxI#Kq3&MI3f<CLO>J(j7tDff^!`p>Tqrb
zL^IB9fM{bFjRk6rWN4_f18R&4&AKMBa+~LmkQu0zMX%AJx~h{hkCD{aJgMts1T3-<
zIU@gNZ=m8f@hRwfYnIV$!wwDl)geV5t`5l0!Wtxz1k!Lc&G384tTO|`b->Y#FJDoT
z{!pgpj(y$z@>{83sLja}k|!c<Mtz1FBXP*gfb>QB#>hZW^we=vje^vl-U<%gAeweZ
zj$8Wg^>`WGL+)Fq&&X{<i4FK{W>xM2CcA(wTDD$=mH{O@`57P4%%FGxwmlFGFl}c7
zn@r${jP=YFwLfWfY&j((v)NuMdiS^3k1(caW71erZ18;s9U7fW(PpF>M&_hBMi!(6
zMwX-{MpmR1Ms|EV0aK?v-yS1Z(p3=&oTA-HcZ@tp4~#rXPmH`sFO0lNZ;X8SJ_4pr
zU%oF!{(OIo0{8(K1(U&wNM9&Aj10ploD9b(f{ef@l8nSCij2Z2mLDr%>J;<E7$xzO
zFxm^qy+Dz%-MN6w#W@d<c{t|-G9TvyfINV60U!%-E&}`_AYke*2K-{2%K=#q6d$p|
zqubqMY$>`Dkd-)B0kR6`Q-C~$b2T8VajpSm4bHWIUke0G{SAQMfO8|@H{#p^_$@fM
z0)8vQ$jC#Lmjt)Wde*Yh8ZuCRPhw^`$VXOCgw%>79VDbyoFR!7NBTn05-ZM-#ELT{
zvEpbp2|9sD7@BDU`d*1mNAe}0@0Hkeh6ijq8b3h~*mQ;mY&v2TK|t?2++-q?ifruH
z7fKv{FRJsCS<>^)Ivw7ZCEu&?|1oqtw)xnCnAg*MzrQ}3FEc(gCm)PXg1*h({$f=7
z9~AWB6Ku@Uj6*Lz!2zRLdb7+>eH1-MZw^L*dVv_xTC@S0e~M1jO2nuTOeh5AXuc(}
z*Mt(B>wsPzFhlc8!THx~#<>mXwK0q~Xfn<x*uyQzSUc<0CqJ!dhY8oOWP8c1UwspO
zTjY<UeS(bJp9DRUnftg=VB!38Iq4tKgCf;{>V!v!!(&Bpio>}vt?h3F1Z^xdRzu$q
zv>j<jF=S8LW8^AyRYS@}&_1LO#gH%Qi%|d>fKjk8SPiKfLC2D@6hmTCj8PJqgwb9g
z+zZr@8WMC7Ad3LSa4{f@aV`hKa-fFPlAvn=SqmtJ8vxmWb0Z)dac%)*3(l>8Y-KpH
zsQuK|wey#$!&pMYObZE#qRlnTF|rU@=rCoLB1?>{L{=ENYq)D60aCPwh6hHTB2OKr
z%uD2jk+;YjqcDvyEhJ8g4%Z0BC_)sW!<0peA~A{*MPZZ+G;)C!5-mmN0gXJI^MNRz
z*?m#;0U$bna{&+);9LnbDuEUfJw;amjVhc^0nsU-!<1J8Q8msrKvcu<xA!=qO?s0q
zfAXdx$=$ha;#{j41VyX5sLn-}gQA^vovoSq&HyeMU@i*7DY^{kmH}%dEtIl?_7q>>
zDSK@E&F2LC10N4}vx|?Hbno3F?>w5I&4gyyS~nM(V`L$;7>g8_pe==#7+DFeFme~V
zV|Ktp=z)=^&~q#r2tj)Zy)g0?dSm2I`eQaCOc;jQh;U&zMiIh@u}F;xI#L*kQIs$W
zqg)`&#Vtl25a!{W4}|$(EK+j<s<H3@&ILeNfO91fR^k?=3J9xkJ_UrQz*saf1YHe;
z)i~DxVGYAe5>HU{`qarE(bzAkR;uXgN-Sqnw3dpNHQE|dbS6;A1lDM?gwmK$H}K7A
z+OmpOwYNG>!mm{WRD&6H%m4uyAQ&ki%5us$aFpGI`%p=0li8@H0YMs2LpwPN=IR71
zff`EVt*NB#pX<}Vji82Y@u^|^V`^B^h(bB_!R<38vlmB+ROo>vSyZMc=|GSU)KJ=O
zW17wGVsUSjj1S)QsL^`Us_mQL@CLGWD20{jwYkqf`H0ykZ+5nf3wSoe<FD<X{6J7j
z{H>|VjvJBnSS9l~OqEEsGAPZ0^5gJ*+*G2x_*>QN5BSnb&ey_RMnjPCfgu&Xe1<Qk
zJeRNdhP!W#l==1m>0e&1x#L%v#{~y_xBW*=M9Dlx>3G<);YR%6eV^yO3%;+CegA=N
z@r{sW_V9hYXN5|(J66gs9Q6tA77xQD`oVK%3qGQeEcwNuMUz(}`sCFgoy0&pCpm^;
zb{ZoEBh}D3OsSCu_nXb_+diTIOMSB%9~vDM$}FR!5jP)gAPho2K`leVf|8V+STT|u
zXESan=YpORykSRH&~I@Jbx|@J`Yn!=E=u!4zr_jCMakdewmvW8`#M+4tj-f<+tvpk
zyPOW)E@ewOpk^p~iR}^x<h&|ewx|5)c<=H@oJMrk_Lgx$t{F;W;9L!Lf}%GAo6Stp
z6`ixe8u0veb-%sYL(|=VfAR~g#=<9L#MHbVI5Ud+il*pnhN742FE>R|#_5Cmdz2UT
zMaY<9Br_NeNlqA%3zEbxoIOAn1r?Img;NT2QPQi)HMBuV#n1F3&U1_U<P}*(sXr^h
z?L5bh%LMiI9D9s>I6hn^RQKihV&u>9$0&dkfKe<bmWw<hL5n$JjFLD>7!?6d5#Tc6
zb}`@-<6H|kwSdcn&kcanfO8|@G~(O>I4wB00!}N#DA<J)FmQ&8W`u%=i=th?7#&+s
z^l>omI50vDqYgs%K0E-F4G7Ow?n$+|dgl#*2Qu&s6iqQY9DZwEUHmB{EM&y9`LJMk
z#ppoxgE<;>68nF4>&E?y8TReX-@s<@C<mSFelTje)wRNLi*o5TsPWFW&dx|5BzsQV
zGT@9v)Zri5SN-kH<uW>tlp3CH4Ga*81f3?M7dmZrrm6hw=l2P^iaXP<-nZFRbNrJx
zZav(?-?42=H^Zq`(pR!T!30I8>Ze+u%^aF=s5HX^P}zX=@%gB94YEfh;|lr|a~>vc
zEV}p+%^BLhLGLv1w-U^{{tnoD$9Sig(R_!cm+of4F5$s2`jZNo<?;Gr&KtS2a-AQs
zrzqJg$CLh52IJUmlx?uYpM$akMl_cmgUl&R#B&odN>)xDO422S#snwB^@VC3`t!A+
z2Ph;NdVr!56mSfEJM;jF65UzIEe$<DE0~p-urLw+WebZHixxSH6*u(rGrE4Rzt}tR
zbqGhC5FHa28_$XvDr7}7KjWF73CvG1^K)TrNDK{eX6&rbh!JueE-6nbsPL7Q6y0XH
zd3%nW=Au86=P6JZs*cvM|8laU)0D5AP1TGYbZjSz-t~0USWPYM&&TUdFfcSSF*CQA
dWMenZddwG=u4<$7tlYg4+0*@|dH^5R{{TL@9=`wp

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_gemm_splitk.cpp.1985BB1B09859227.idx b/.cache/clangd/index/test_grouped_gemm_splitk.cpp.1985BB1B09859227.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5b58186ef44040e5ba8acaf3b5f7c629bfc3153d
GIT binary patch
literal 4558
zcmbVOc|26>8$ZTS%o*mGp&7=$g(Q-FYh;-rg%L6`Wk|NHg>HoC*0nE%v`Fi(Qi{?=
z-I7YVrIOUOCn`&#U%7Stx@SgjGd}g{^ZCvB<N2KDJm-C%?fZP+*Msipcng6bdk#Gz
zA}DS>7J?vDg<|7kHW{Kyu@nR)l-GJitWq#1^w`E8(!Gr}tFqoLou{?KgH}{IQ28=p
z<JPkYW7gL+29sy9KW5lD`^oIVDm=&<F8vrybEf^;Vic6XsFQ6@JZPP4J|OY81^HtC
z;K1ZxdHy+{Hu_~fFnIMaZsbNtQcOd$nPvBehMje#T~4<KM<yme%PL*{vXt}qS@*C0
zuAA?FzQZ2<6nyu1@8WT%trz{aME_RqpEh#+Fw)g_qsR_-+jD%dU`?>t-snG{Jf}Nl
z=;*zeW{LRicyjo|)5B9LUW{XMx<{jS{I9=^abN%P$W|3*2gf6|+pg<?#$c+R>rK&h
z2FtfurMd+<k`ATE<kxg`PS<1WKV-BgYA7`}=Oo_A(obX-tW6rQ3rgt7Y|SrtPHdR>
zoL;GLY5M)^dmCk}PZwV@JDOG$X5gS_B)`(e)5Mv^=wc+f=aGz*n7Be((DJC<0n*Na
zz1!Mj&$eM(G1n(tUL5M{L$<%afNhWJ5%15AKQO<tDTHxlC2{y(2ZyTi<!%eB)Z9D2
zF{5<EuFfRMzpT$$z00|%-Hs7i?V&J~!p8f?=q2ikF~kbGhEzLFIp`mY^+5{gt95oe
z<gw_?u7bePp6Pfo{eZwbC8u8z5^tG!wMyq-F)4YS>HW^QyZnGU<+fHRMUL93sn$X|
zQ1B{q3~S?)9ew=h&EG4+s@rFh)#Ed&JIRBiMG5u;11GP9eDtDi^|B1nsDDA*U#z7U
z84uU&K3r&iP^5I%B2kyL7v^zkeQm>?ldKafC+lxst?REiL)zli#1aX}JC>_1JM|=!
z{CHPUdWvXBYD6$6`;lXMc0^dB;yk5>;wvw`5Y^x9(t3&(pWTFGksiv>Zqm1hSTW{H
zIhiz42249n`@cA;texLljNnM|$%W<%aHIZlE-fCPrYd(jU%vHxpM;^rsb5kv39=(0
zY0Y7$YKD!?6rTS1OsCoXk@?u64>?J7<JD!u)q7Gj2Mx``V~xt0PO-T&PUW)3rdr;F
zmR_mpnjtCCD%F@UOr+*sd*AWwDaneV^TX63UCo&z+W*n5=&Wy-cQ3HrrroG+v9C<-
zf{YraYf>yL>P=TJaz4s7<83L2&6Mf=@4J`HGkq4N_+YqYJY|1yZ<GeRyfxVGX>k3y
zQZJc+eb;nmPSQ+n4E?1iRojueTP%BXVRb9Btzm)jIn2uSI%07~%}=!sFFW(#^O%@H
zH8rd(=7DdQ!rSZJ7wc`>7!$5h^Qv}LB%e_{wT^Q}sU&TP*0*#+k?5*Vx6Np`p6ZS2
zOr=l*29zC2rHj+HzA{j?c}?$58&$IP&QAB<AAVW%dgSeJt7q1W!s`a#NXCU*mQ2WW
zzFl#%YP%^*{%*`na7jzT$8j^O`<GS5kzwin9+~0$R%4WB=awW>?z!zSP5NT-X{Pau
z&Mm)&6ntA!?kd)#?f6}DD=0QGVna9vg0u?*e|Z*ivKbF;gdny&Bv0lEh|4hOqvz67
zad_zRB^cU-niU~MGEV@rMX!G2it*ixU?>tbQy>bNCxF@dqhob^oaH18g`s9LL?-hD
zFgs*Vz6-7T#|MX5gPLJ+Sb`^JJwp&v`AF2~sJgQ-6!IhAtcOjJSF0SmaTqVuEF~@_
z!4qJ0s2r((;zHJIJQVRG-=RT&=rqO8uwZC2YW^<rEt9p!sM9vCFcgHEzbi}fjSmD>
zM4E+%yiv2<0y#yV0KRu99?-ac+^hqKS&Nzp;sjZq0OtL=;g;)9PZ{B%s2}-0nEFyH
zP^#$y4#PstcyYW0Pk_~@q>WGBP$nG0VSG?CB96%N1TbF<sazN-<%!2bp{V)0j@^@_
zh)#D5sliY<YW}Vz4{MD>tj||l!cZ`3{x0wFY?G#G*IF4I#t$_U#fem&0KVTxZhBn9
z(S3u%_@ZW#IEl&=z#J149J@dgf+XYm-muDYk7KdC06?d(yf%R1V0qpFJh8k?KyIW<
zPwK+1?a~%{12xGLlehiq1BKoI#3h2^g$XhQ<ATA=j>*^WbE7wE)oo+B)Fpp)4u)hl
zeK1k8=>r(fZU6=X{DXSo0v9-+TR*TZ`V+Uy{GYfQC^?uwyyTq8gV_XE!YY7bRC6aV
z1++B;U%rO8BC7x-VK(B9Fa@+Vh%a9wdJ@+F0Ksgc57AFR^Cz<T8p)Hi24D$hlYB^i
z0-8UG4Kz;)hsKB?LXOat1Qq>hznE?^+OP7h#6)1W98L}av+u*XxbIq*p9L1Jj~c;H
zpdxCBp+MF;L^FZRbVz0(dO%>Jl5>a)Eg}tZ;BP>plU%`)_37{MZYP$d2#s6NcoesS
zr~qvEQ$$54*gT-#eKaG2<&hnzL%d?B>afnGPeL~VfE;zsB^*9dxf&R>ZLziR<~=F4
zLT3)bMXSx_Yv2WDlh=~j-&m1NSxaGqYJiqvB<IjKy2O<12$t-caG;t$$dX6V0TW<3
zFi}~l=4Gd&KZFWhmBM{%NB+oF>=b%~67SD7wXCNtUpD~p&awnZah7HP>(9~*AThwi
zC{_e%=D>10Mz1UVq0sw&JKrSdoNMS9SrhC)%~6d;a1hX(2n@c4SR;1)2|=6?245qt
zAUg2}jz}lE3TSRbcfLkiL2?3<h0=^0$sK4%3zsdwp^kqebn!trD0H|dG#_mgQ;H)f
z{^^nQMBo1M94>>H608K$GexYx4k%OF5bXtmwjtSrFwb3#xiej^RW5Xxb&kaeRd&Mw
zWPly!o+LjziLV-aaKs#rjG*YpYvT1mYr#Q{LWg?;0qHjdusR66Psm^6{r}zF!-lM0
zQk3%#2Ax57vM<0i%%-?gd_h^k*Potz0P;U2oySp_>Wp<}X+qa`Wd504jK%ppv^-Lp
zQcD2ik86&Nh0D?<F<3U0LRA5=VA;wPWfd@fpyeopxEC!@Lzt~VQ7`~Y-Z6h8WIApq
zqYx70bF?4=7Xx?>uflU$0-%I?;ojHyUX&<G1|XEbjX!V8J6dmvmYu3bT?m}RY-Kg&
zh2W=u+RJCe?~4xz-CQUpD11FPfY_M8*Z}TJAXm&jJj3kkI`9pGcz?~R{`{{S_^+IA
zucGX`8Kn}IcMLJ}0y*mf{5PzJ0wXx#=<OeJkMI=5iiyvgFM)&cGExXZn!cL8Y9-Ox
zQHLnGl0=b}lUJ~^wzXSMvtMDjV6ly+g`Vn<zLFJ{l-1NVv=-_v(l;<#YHVt5sbWT5
Wvdn1#S=+?LEgHjMI=G5>K>q^o{R^!C

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_gemm_ut_cases.inc.B622DD65303F1D04.idx b/.cache/clangd/index/test_grouped_gemm_ut_cases.inc.B622DD65303F1D04.idx
new file mode 100755
index 0000000000000000000000000000000000000000..8001fb3df3f16921de31a2ebc3fb94145acdac93
GIT binary patch
literal 9582
zcmdU#dsI}_8o<v!9%pbk;}&rSHrr9h366kD+#9d7k!vuhrBotlUP@k)rHKrp#JZQ4
z<|7p`%_jm`s0E0x^sc&^*<*!vyV*l6>-Jb)mC9S%T`NoN?wR46^Nnle{hL1=)^G29
z&e{9h-)n!{hbFpQvm+(x-olBC=M|OZDUu|Gh=1;~(m7v<zZ?2U(&B9!8cy`Dy>meN
zwc$Kz^z=zz8xjm>lN`2|y)VAd*IX4hD}V8-)g!kYxX1nEvtJBbG4QWvE+6?eDDj<5
zJ1=&>UrrZhs~2XyH8*^8YvzTY<?+qU$6GgDd2#vcEx)$ZN9Spql1w8<)!V;%Bw@t)
zipsZFEXjJVU7NMCa@>p$x_;ZAcI7ooT<i5YvCc_nKFl7wwj`|m=ahy!=fCyLnBPWQ
zo?f;0jmHX?_90W8Gm0L0EM<N3!<y~Hs*Y84TgTr_o%wwJ!s7kswzPI1iO*<X#+^sL
zn*84Wo4Z$!-2bKd`;gd1cg4K?aIr#zyLotU%M+?m4wfV~S(<F{{zzsa0ZAf^ggh8}
zbo{u?u~xaSL1K?fj~l!{lIk5Gg+`>#3tM#0DXSbZP-4TRVFvFHbPvCCQBd|?HPI^f
zGf8ZwG}GYyfo?}<^`8E_SNvd<E#m7TQi#F(qklc=g-Zvj_U@c-l!+M7N~Bg|==q}$
z8n=I4d}zY@yRCAtir3IH#f%}5amyTy9~tE^@eWy%4c;I9`6>H;*qBYX<{4!w1~?_B
z!TX~Rm^Zri(w4@Y{zh4dg>59#MllQDAAQhMiMgG1kLArV%6-KkT~K~W)}9_X_R$k}
zY_Q4}D_(5RH0w>}6IT^|nQN6p2T3eWiZghBKvaj$MAR?4F*L|13nfK@_P_q{1N`;x
zPY;-GmHQ|XTTZ0qVv@c;@Q!zz&o0(ZOMf!TM)8gcslwp>(bw?F+aGlI`S0JmjB-CQ
zK;JI^N6yxK?iiFYLToo3fJN1_hR-w3J!Clk=47i(<Xg9^XUK(%*A{2A4>@d=!@}`y
z^mJd``|?W_PaS>0DjS3F6nnbAt*MBfJvfdSWud#bEYd&8?;if_-MQa5E*a%eF-VuB
zf51;ehsLzLy0yeAoBQ9opFJNcEt>6qcBrs`E$<!u*QkaHnRs6Wua*M{j7XM!#^W;$
zN%0bN8V=;2Dwj{H;!VC?fwjoDu&M0J?mus-Z&ekqo*wInp1AGl8T%BkX}q%J4-2n#
z&?gnI3d9Y|Gb|ic{($10&YWk`@(MDeRw-Wf{o}reU$?#7+~%9bg{s>BWLzFOP4VvP
z&8Y6XerbDfgW}cGv&XB(-4|+Isd%-a+B)Ok_a<y>P<p2I-iF}CpRSaWlpc*gKV!!e
zZ+6XpMCq}^^z$=2;$lBmgO@E?bn?ieZ^imNZUFRE4j}L*MD~6eYzDlDXvWVD${ASR
za44&#mq}vKD<2<z;^;E1mrdL9TxH{?j||S+s`_x#&MWWCe6^sLtqtOkP%w{I13w4x
zM;NXr`1IJVFkBBo=crvdZ3km6Z549yxJ+=OasZ9K_(Acer<eSws;s}Mzj^9a+w_IS
z<C9fpGuh0o+a4TzaMqM~m5nrwG}jalS$3>qZlub_n8ujF!KiGKX_EQ7+|-a+85g=#
zHr+JcT={iE%*h{{ttu-v6`Py-bpCShn$z7XD>s#!A5Ff$`0&*+yH&Q7n3fXrfdN;(
z&smY4q_XwIw4RusOFA+zsjE)RX9qFuAm&*+mv=_R9xqneK4RKO%(G9tAAf92YcaSG
zAp&fAKK}wX4ZcJS@`OhSm_A2S<bLx|$*<>bK&%#ne1TdQr;lO4*l(<dwyYp0dzlVF
znjv(^lP7^+2mj|^HY^hyWbTCEhuEbdU66-}fu1BpmyU;^6GYo1X?+luJ7|~CNc^nt
zz3`^2DFd~O4P*OGP?<s%3p}H;AR2_IAMJ<8L`{f-X)q#{s)#~qs72T7FdBv^oQ5M}
zlp%_s5r{bDh=$YQ7Tr3cXcQuiYKTVA5s0E`G@=+9gD9THTXY*spb3Z)X(FN|nuI8s
zCL>CrDTq>OsztZdG@6Deou(slQ5T{Nnt^Bnoq#BlW?FQ6pGv19%A#3_rqOALvS~J=
z9GZhDm*!e@uaQUd5arW+L<O_}Q6VivR78sq&82fKy6-8WC5RT#1&B&%DIzy@BPye1
zh$?BNMfYD-v<gu*q16O&4WTudYYDBzTt{df=9Pr5Bo^Hpt|4>{=6XWwF|Q?bE#?M7
z8!)dUbRFi+gl;Al-Mcmt+K72Oq1!Px5!!^gnb2m;ErhmU-bLsxVgYZhvNl57Ft-!h
zj=6)-4$ON9-GjN4&`zDF7w+G9^xVw5J+9a&TfujT0B8)cLI4!8z&Okb37yC_jd4~;
zghdE2I<4TfM6PL^Yz5CKa!un*E5urnYZ@Q7LdGX@O=E=>VyP$BG%hDrNQphUrg0;&
z!NJSpN{y|=2Cj9+`OiN5>T2KVJrM*=8^vv32t(5rM=CV!fH;O7x-UgW(-vo5G;Ndu
zK#Gi~?KvCbX?sqrFl~qg!fyvgM2MI$5DZKV^c*DN>w!r#=$shjJBz|+<y$zf?nm>K
zoe5RFEZ#z1A=Zb3UtkIXhVg$z3Wp;RSUo2~T@gG4sqdWCmF>C5^2%>3cfXiH&U%mz
z+=m5=o$n9sI(0+HUEzuiBCx_jd8iFyr^1Hw;WmhR3XA9QHi%aWOXaCHh%^ez<e4^z
z5em!Yxi;|b3Y*L4+CXg-R>>=E;87K}lJJ$p20l<>n+e}cY@k01+eP>;VgpgSUpn>L
zy5+O}GFGGt_Mk|Zw#xe1`yn!ErdYi%SPMp^YAT{IdsrkaS7qV$a70XFv3g&G7J-Os
z98r`#DiXw`GR>|b8ljDd)%&8gXhbnu459>kLL`V+Wr_AgL`hmwtlpQbB_m4FQV^xt
z(;{J8RF-Z}N9596v3g&ImVs!3HUZI8`_xF-PL*ZZvk*<wrp4-g*;+QD94!Y?o;@!T
z)IeqV_IyMIT0yMdSEv;tD$<G&mDo!nL48!Vz`g)csa6`R_qjDUqB5-vQI)+a64Xv*
z)x=&+5Z4f`hQ#XqwM47MTt~Dz%xj2!4T%J`Rarf;*JEBww6!Ew?{6Sl1Lk!^TZg%k
z*c(YCsJ+Uz6Z>||O+;%Vv3h?q(V8*05UmAs8?m>MNH80fwG(?g<_@BDkXXHc57G8u
z?j%~L&J_pRyBiyV)5a<+hzB8W(U12-Wa1_Vn4Q9cc`zcCtBAsQ81gdVJRA|@%mHSn
zum~Q3h;xo8ibo;uq;U<=2tL9AW~Z=d9*ro5#~@1J3COD@@<c>QJjnrOr?6z6j3|Yt
zAWGwD$Q!2fbVM%ha)8+>EQ4ntn!qO@n#!jlFPg=(5KZIL9AI_|%jVgLa(E7+Jf4TV
zZ$8gQRKN=yV0H>C<b{ZecoCu!UV^;#0=@uIDKB+^*(uD;-H6I~8KNp)g}i+=;nf6j
z4dFG!0cNMLTEc5F*AZTac@5!faI~l=ydLvf!q*Z9n4Q8J2yei=j_`Gu8wqd3(Plg0
z+c7s0-b5T=b_#1Iycu%~;Vqck2yfG)l_JcJw`1-gyn{Hv>=d?#@I9D23GdX|pIiBo
zD?OIu#H-SMP@TWBYFsbNpg@yFI)fA3!1@K$dqrY|%Uix;Fx0-PZH;#vYfb*7uNrVp
zPy@~hYQQ-`4LB#L0p|p88e(F&=oDCjg8LFtZ^7lbWu*qI`p>h;;@RIXe!yjlcot`S
zptR!I+t=W?ExK-Rx=*_)2I6{QVD%78{I-L@%ZfpG6@hEoHSy)^-RFZo6@vm(NJudu
z=iD9i{PxPwcQ4=eZM*4RyWSW&ph;{NK8~pY$1yeFIHm?1$JASnW8gc)SFwr&Dh%9A
z2tC3J_0>{g-f$WdgFG9rTN_q`!>$dBJ5n#4>1Eai7jQ!70+s8vnO$uOXH{8<@Au6R
zu6cxHuv+`>fI#Qryxe1}y7=(6z=e$%6gUq!uEI^!sjE*_%yVD6=&4yd`@0~k5H)M3
z%>{ZZYSzw?E-+Y8vv!VgL9`Y%Yv&{vL|sv{c20MJ&k;3iXR!-RMAWRE<t~V3o|?6D
zDRF^bd1}_q^~43{>8V*ecMuoE1W(P{xsSNOr^DG?tk7efx)`vn!d0c%4t#us%P#RO
z_VB)&FgX9gJ*Idz@SqNtU*hRNLxtNhuw=Mz69e&{`!3@^A>b}ld<O41+<}T_ysuDY
z5U=#@tG@2QeTK8G80c|+`Y`MYVb_FGV(|nD>+<Tv<3Hbja3E2|a&bZChm^m?bXM%Y
zU#N~}=I{v-570G=hzH2VMLa;05%EA2VD#raLedHEs&G`I1futOM50rmBs|mf6kzoJ
zz?%rT_!W{3+`gZlob%y<!kBKc(7;O5Po?ji>9{-RbT9Ko*FV~J!|)Ho*S&1oa4#)1
z7n`!LA_p;9DlM9uw@Lhe5GY>y{}YS90V<dN1}Log8=z$9Z-5e}zX3w7zX38ue**+F
ze*?H@e*-uwe*@UCzX5C&Dw_Va_#0p?{svf!zX8_bZ-BM<8(=N|23U)~0oLMgfVKD=
hU@iUzSc|^_*5WQLaEt#<B*|>>813fq%=Jel>3^Ps8d?AV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_grouped_gemm_util.hpp.62D7D32AEF140979.idx b/.cache/clangd/index/test_grouped_gemm_util.hpp.62D7D32AEF140979.idx
new file mode 100755
index 0000000000000000000000000000000000000000..976f0d875565df1e3401e5ba5a90c147ebafe2d7
GIT binary patch
literal 7828
zcmbtYc|4U{_dn}#DxUKk84`}^5Q-2b9#fP|g>bv2x=n9s)<B84g-C8QX?jyPO)45`
zPOb*G;!1Ng(?}|}d3^VCwA#;K{XV~sK3dQIp0)P8_IIznB7*$=7n@2XzGH*pqoZby
zmP#ZNZSp^MX3WHXY7$8nCy~UjKQVanw9Aei-VDe*?W2r7;6BPHZQH@gYq!bQTw~1a
zua|@rxbNGSTL`c2&RAb^Db?j#cboHdx2~UiRXJ<Xh8JI^oC$t-&F{MN0Og!<J1^|f
zfAYclBXeTUt>HV&w;T4&?LTy9*_WDSZ|9oY-)~|+$7sA8d~fi+`L}F+SGM)d4NEYN
zJ*>?=c+c&h@@I;M?U`kx7Cj1E_`Au0Amc^rPpv%s<N+Hz*w#hkVBNMty;4{CUftsv
zU(*{6DoXdl<}Ag=M_nDa2dgL4sz1Hcx&C(9m!j*_+B%;&x4q57?4G?d(pQCTYS4KR
z(C}=x@=Kd$t&n${z0V(MduUzN?93alOZuPu@?_)IVOP84$E8jRU41K#e=#6;lUaE7
zxT>RXlJxD1UfG(QEv>0qF#o3WwHaP33fC^!0p|xqjW$`c)X6zz$O6+@GoS3pyBU$!
zWL?$d6+d<1;q$Sl4%gp2VOG_Y`nce)i#xLN^*=tY$nT@J$HSkUJ)^0!TBYIVGriW<
zo;2L~dyl-Y79R=*hkIY%&B-!)6)mmKSRL@hIJdU<rsD3J{L#E?bN25HjF?$CwR+Zy
z&7esBlvk})Fl<Mr|MCMXl^tIcdlwhheGOLsZ|JbXw|{!&4RZWXu69C9bN%acx5i%_
zXVdd^L$~$Oiu~7E;d?e_gkSypetpu_v(LYbDAC@3_EzoLK7AHDh5p)>FJ75=)HQO~
zFvqk<16K`-tQfd0XKN>ov5a-Q%z*Ue#=4p{54N2DeU$uEW<+M%Gwns|&!zfC7gS|D
z_1#sOy)V;S^OwTe%m3b^esS^p;p6PGOdkGbmdsbT**#tvdi>65mvQ$x4b)q^>Of-F
z71;nL+h+BKkgMI&ZWb9Vv+`te*M=-u_Hy1eH4Vf1sdKFkm6ctJ3doUlKAvp#XztPF
zlRR$BxO4n(|50)4uavzjtJ*a@cGgw9+t%AVbn*>3vMy26e&+5^{QTgxG5;M}lW%*;
z?hpIqn<4)W+tDj8`ux>S)2+t%tn0aR_LM7OHKVg1r0rANzOv|^=For}T^>#!y}IJ&
z<;}aj_Z|vAnY(J(>DoD#Q?tz*V}@C(4=?ij9P<8VN?(Wd({xwchAtnd(Ys^tnqxj?
z-Wv6;HT;I^!j5XU<1=OiU7R-a-;IH>3DN(_CC97XvqVT=WEU=DG;AdNY<9M_N=Ven
zsU!c`=5k+4{d_sY^y%F43$I`mni7o`z%7hnH=M~IBV*c;+DUAZu}Wa=_F=1s_N&+5
zB4apP3Ez$FrmGUDc6d@fe$9fkS6YlYQF~~3a4Lan=ZPP$|5@vPP{wGJ##4b!1<UUr
zG;tfAo>jF-!$QUwk`^3T2W^%3q5h}CA916P@6}>D5w)9!o2E*j+AHE&$_Lls^Ys{S
zqIOZZn5qP-ecnI&Cd=YRli`v6Yk;i*`S%Z+1f*s<`vhclk}(}f3-+wNwo0Hna_oWC
zCvOdNZ^u}6k?=NcZS+(^qS00$HsB%h$h4lBPWw`1j0x!?mW`FE1lIrUOHsEDE4n|I
zF)UFhunCqbf$EX7*5-ULxO7*GF(v9^&?tru-#=(F=HALt{$u(Ea1tLf0u#=}NF^j1
z3`lxglMzG>vA%WSJ!2+g49E<eSSMYT!1{?Nv<}zR#U5<S=#%<U?V{{eLZZ=zu(2?M
zx$SE-r~6LnE@Ldn7^bt+Wh#O7Q=d<*3%nB0AY;^tdJa3sUL{bSRxs?+zb;qa$rv-D
zp3F|xQwdaW?34AlW^m0^8Dm7$<Js|Yl|XfoR&m{(k0nYy#)^#HRpF{s2~;1MWIbWY
zAIF_#jDpmTW}`c(1gcA?r6f01R!@*IT11@;Y%*AX|DefVyKS3FKQ$kaF*?L#E<4vz
zB~X28z5hMei$jje7)_#12R0pyzkkr=+*V6@&OctpGDb;ECbARdDuL>%O9MvO=6{vS
z7&)1N9c!np5~#jpGV)^K>?$KYrWaB7H0)`m5~!|Cf3xwLsd<1N<4e@e3TJbbK=mE>
z;Emhgu62+x#-#5F>;$<=p!&fajY0ZvtjlCfN21OEHUn(Ff6(M%ssGp)!>-lH7#T5{
z#m;h62~<A|(utgQy3-~Zqes+}*hzAgKy~Aj9K~;%`;W;OT{1jJ)=^g_Q2pV?k^o;r
z&3GBpp46TX?0m5Q{y`H-Ow_p8C_Tzj469-%^h%E30^Anpf>S{{B4z>?QL}vD^1&3%
z<c<s1P0Deq1&)u^i8a7Oh-^h_3si-Nm*f3B{rr%t^vjZ09k{dR3nS$%WEOHfUj%7c
z=pa&DJ~Kn|%rokzMkdAYA1<lg_!rCZaq>6?t~JNU%i|SD%N(Czm*9px!10OlL<OEH
zj!&{oa>6-t{5<(Q1$r#Ur+_>K6zJp}Ax^tga6?Yz_%x8GfdV&yBYta{4o>LO4jn6I
z_#Zvco6O)V_<w}}Btt3x4Z3~<JH)syz^b?#7Y+L%YuQ5FbT0~kD*$WU3uJI&D9|_7
z=lK4f{e7`<x9>boIuS@5jXYES+*@(`a$hpYQs7EK6A4kuKLqX}XyPD|=t(C6%VT|x
z5Ah2L!cNebbkf50IAs6h=}l7_$8D5y{0Wes0A1{c<4=S3Y0#(m0mvVKF1Ez+PeJ=B
z=p)A4f(%2rD(Eo!dG`A;*AUmNX&vnsFDK49nV;;73>k4{v)Z#`36Vc_sb4hwPDDp(
zEmh-qlQt$YbVrW2)Ua$%k(HKJM??cdbvi#;;n~KR+`R@Krnnd!?`hyELS6=5BGk*E
zmk4<qc#DvafsY6kgR&T`=>i=D<v}q%1j<8Vd>E96#rOy)kBIS6P#zWIGLV;vH{?9X
z&x`Q|kY5nv3XoTbaU~d5f*BoO4H(vd8EzQI*MecK7~cZJTVi}03~!6^9WcBj#&^N+
zt{6WC`D3wzya4$NF@6d1mty=1<gdiI84R0gdxJ*j3@O-D>rQ;d($LZv_jFM6l(1g1
ziHD_9K2kkW7xz@kmw<W+oxPN=1NA!Ex|IJ6>Yph_|0D-O2wgA~M7or-s#pb231O<p
zgdvov5=JD9q$trX(Vg;EqIaSXMM+^v;grRa1|<!qC>0`8VF<E`z+xj(VI;+A;FbpN
zl*Q7(I}Lm&PKU5`2uD^ESZ+`{3`UIOBL`9luIN;@0=JcN)QEq27i`Fuy=~=sqL-2N
zY>CydztO=#BONTVKee)b+<wx&5a{84ja`wsV2(}lx&M>OqdFAFo6h>UO}LPhaW<k0
zNf~b=x{$mR0&(HUg``Zh5pAVwk`LXydCGY<;tNTc0yg3c$vYJS#apRN0~_(yxu$~;
zZWYF<WUGWA6@8QA13UwMT+ZVcTPJodXj(<i!$PoF2z~LEnHUorH{kttAENjQ>R&+{
zPk2ls+j~Xx^OZzlWn`tqC5+v9>3G({QNMEn<y)K(GfoNUmw*xOB}dMbV+lBm@qiNe
zh2oQ7auUqN*z+X#Q2YQa9)LBDha+cFzYdJV*s%_r#dts+{6cX(nAC&07<<-(58_#m
z+eMx2?0Tb>t`&dke*GM2btS{`m0(o~wsic>VATw^*k3|X;j`Qovc7Gl{CO}x4_0U>
z<u8Ew1u?DwqY6sBQvL;)zW^(2NXowi^Os`$3XEP+>cg``oLzX8ptIw1gkx3D5VRs5
z?H1=wJvuHhE{LLdw|IBz-|>O*K@`n%o99k_cwXSVAc|7JEd|`EN2fqw3ItJ{2V5Qr
z2)@VJ9KQp&9b&v6xcwBT76wQEnp<5?T=5(9`vyaBDe!m^*B2g8bWYoW+Ya5ZeT)=X
zR`rIWX4$}HgFc#ZJaP7wVw?k9ju=;iV>Ni-#&LW#cvnL|F%GSUAr!v_$G6~tqviOw
z;Qbc*iE-##I-7K>{o%haZGP02<L!*>Z1Dxe@%BdcwiqFCyn~U0Eyji%-_59-Er#bD
z?`Y&`i;oMAcQSIaMRv$|>1XU#wOB#caJ<2IQ+z)zYFD~xVOdNv`8nDsT8R(C#amY0
zo|g2qmN;t}kiR}SCn;YJ!Q~J_aV2!9q=BB4e+=ZW56)4_H$reDgizcJ9hzz2xA^9z
z(9oEMEYiPC7n`0)UQ0ulzaKJtPxn?H&K1g`ufu|ya$4EU%1ynqMtzF8Pr}?Xa3}*8
zoIS^%1&_1fOYviHcnmH$TaIr4j|T8XoO9o{)Td<2GcsE<4KqFD)SP#thNreSzDj;h
zW~`DKC$wKXwN3U*oux&jV<B8=$t+1E0<{X^3dJqq3dJqq3gQiY?s}Sr_m}@PLPn2h
z!>Td~^pGq8ZV9xfhh!;mOU0PP<IBaE>}3|k+b%WTi~qOlT{1VTE>>M}%K2LEtLDqb
z7_@R{3gVBo35(A2C#_G{nXW*NE3E8w&Xg@NZ)NNICfs?kNNPJ)?f(}oV)BKsl8|{p
zpXT^bztH{|avpTqzb3J|F%YD@Sw}MiTo*imYOE?FLw8{jaEqWL-Gvpvt)Td*ruxOR
zk@X=z-Qg48Ji9kKt$WqVRVi_f7~94#uYs)jM5l>vn8!hKC$nzJ7}2rs0&W*gMU-ZY
z+8Xu5^8Qc9idP9~O-L@$wsr!y6T0I%oa);&mV5J8Iw$awPPsUOF9*YNP~gcF^vZG?
zZ3x)D92~@Y_j1bZf<B-e0;#?cbQ<XkK;SjQMo{2M6|l0AMjHaQZv+Ri-o26D1%f`H
z5du+<!5o=yOSX*8cn@%UXod!tTgIx=Wz;OLN1Ti3;_eaeLa*wCh=gB7foYFK7rZ%1
zrlem|h$t}aG0z3pUdWX6NC6kzS>cTlkqW<x1JfR9;35u8`=vt&&gDYPj_^-)(Xl_>
z=nAdA*FUvybpNRlUWH_JgiIhkQ9+(TKhgsi<D^l{{E^YbFa0e1Jn*1lrjXPX(uA~b
zuwQUE8ezVW7zxQjYSiDae=r(h){qzpDMPyS%Ya)(8SjDH>J**72MucFF+&$8Owf{n
zr2SQaX%%>(YjXT+FntYP=t}kX_$j}gR@+Ec^ydkxhy7}vxUDHEYUM!D>BzDOc~Cmc
z&A@G@M-@||BrFy3p=j2iUa*XD-zy|_COJ*=L@NwpNTWh}lx~Q>r~i*f&GXn}2Y1DH
zzE5KBv!H(#lypfNK)(T$xFqkwKXv}5t2Ty^t(DA56Ui1GmvqvSWW_;o{BGcO(*TlW
zTc?b+AlGS-Z)3!Ddm9_Tyc@<Z5*%S{%$V3Z;o|kg+gBp#MvIQL*pn9RXt4(^cBMrx
zT6CdB4_b7fMQ2*<PK$0oiVXkpz?;MQT^1+nxOupHdUdt4cXH_F=+fP}XAf72{SP~d
MJ1x4>q7yCt59z>fDgXcg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_groupnorm_bwd_data_fp32.cpp.EF83656CE5FBCD24.idx b/.cache/clangd/index/test_groupnorm_bwd_data_fp32.cpp.EF83656CE5FBCD24.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87519167b0e027ee73a495bae8813a9ea4cede12
GIT binary patch
literal 3434
zcmaJ?3pmti7yr*-Ff(Q_Vi*x-#x2bnkwHl7G8lwhQVOGzJ6%RMl-3uLNE9pEms+`B
z`doH%r-no>U!_olv^FY>o|65~)SGF(eZK$m%$et0-}9dHJLh#}+SvF@LC{t<lM@<v
z&`%VCAQ|`<eK0EMu?PgUqalcsd)s4o^pLUi;^Opr9Bad)y+0);Saw!3voAWm$ukQM
z4h>GDTqSR<b&l12YPt3Dg8i)Tp2eD>!OvM5aS}~0>*!igs7Gqjr-@}nn+Vg-*^Vlu
z;pVLLoawoD5N9#z>H|E^acU|YAJ>&E^8EJMC#~ubz9u7wcEs`c2Q%iR^ZEXlN6kr6
zigxMvsu=pp!8_UJ=EItg7E_c7$BE}no7Zn(Sw~L0<YyGs8U<HM9&jgV?#jJ<%LwOY
z%}p<+{`Foda>$P2UuZBWVX|fAAdBLbeE2;^{y=b4;k`F3vpn0mlqAC2`*(Hw^0fP`
zCi8!wOdx|xDqh4|A0BroOm6y})3x?}F3HEwC4I-yi0e@|ZpU9_6mFhl^f@UFH$TbQ
zG&{Owu6VnNWj-~2sH<2(vD@=xK}a?Q5vF&-<5a#fX)s5|-KVp+%FE?O@3noKA7NOn
z-Vm;lesW)`?_*n<uD{e&XhuxHut(d&$c>q1FPlHyYNTV+Ryw#=m*uv{;X~x>5d${i
zI{n745{{iszmjj&l{s`eHp8&oN_X46lmf-7{mImU!kQh-sao|k2Unz(3|WVpBp#JI
zOlmAFyla?vzYdLS_AD>ldeTyUDB^}`v5ne7b#$7>rfqnyD7I$^6Mfjel5|*i?%bgA
z&A6mLI!e#D;uC)!$n1y;Xqph?i1<ZTYmI(nS)Jeh+r#@|+ybLuso-i8(IM;dOvZi5
zK57rwYP>wZBtLQ{dVkEoHHXhe6=hDGTL;>A4aNtzP>PQe3)9NX+tSP@m6r$89@KWH
zgp|xbj0kw@G1AeTP*Yj%mDSv4lx8#Qpt)m2c5h(qSylq`#2d4uxI~SsJ=9O6*Mnt|
z|HWrW6w;rhTl+MswTE!;bcJbqGc8P#E$gKoH6%6t=<+e0=3lhaw%^N~dVOvCxhlWm
zoxW{m_VxctR5)FQC_7d<spbB5QNs6HztO&yjMrl=(JS#)-aCnITIiYBu!x#TCxd5K
zdi@$Bw+|gkzaD%+Vt%P#zA>OElXAw8wH7xsM=t9b#Z}g)`oEcee5Lr*C~E3-m2sxY
zhQOZ#A1<k6ZJY=?I9k(Uz5K#ZgmCW3Y)8LdYjW>EP3#QU;s|?*p8f9Lh?;W`Zuv2)
ze!nMo^DlYRgZpTQ$8Or#VixvHHFejoQ4uREO>JR9PiIpuB`u7c@%&RKb)^jTI$?M=
z-|Vmdv_3k5pKq&pZI?H`h0BT<D`*_xgf?;U_I`zFzdJA0*&RJoE=?FY>Qy)1xw7y~
z#N(jtl_mWtnsnF)yLS!8naAhxj;CItqhmt-=MWHt(0i{WTesMaL3-|n*f=y!j!)nX
z!2je<h)hyj#<pRQey|#k#uNAisMS`oY$DvRGcnLU2x99)`qF#?@wN#3k#%47iZ(3V
z!ytWNwH``OhEKjIVKpsHTMDhxD+L9@YC1$m^9k^_KKJhJfznT27^D}hMxjuWd;+{<
zHFYOjebh+e$UXn$op33q!tbUt34`1YXD*ADRpb-ko$h+GY~fW+00!v|dyAvQCHVxX
z3$#v7KYPBh2!q@Sd&{6@l=%dx2fHsw<i*ZcVUQlMnt&$A@d;2rr&G1-rcBWo<PKOZ
zjh0s66QG`HXV@NM?TJ%D27MK=CD0PQt;_!wRmTIugBtRrk-K2;HIi$vd;-*;MdBJv
zNBW8|&|Wz6|Ec-s?%d>{LcdCJq%W+-pfE~&0=%OF_eTF4wc2+in%{5Wg=MQT)!aY^
zp!={9ul1`&>P&T4U?k6JimXd1{}loIYOC1+v0Bck-P{`uLxd<lN02o@3=;VTDPQvv
z<#!$c0^kNdmj@b<JC?1lsZR%yfH4C31vOu-HtCc|$vp4eEOf2~b>Nu6_%GqGm<KfA
z3^0a3e}^rHPRi6Rq_Y=%CtiU;0&Wv+-7)Ud3twY{ySO{aYQDN)CWgeaiBd!Yr~u3<
z5>)}`@h-&091~Ov`)Xme6oA<G!It8@3=A)%pxDomxqE|p#z!G>B*(l*xwCIsT$D{T
zCNe+}m}3x4Ryjq6q5!AFP!ixs43Yq+%up8K)-u+9;|@)U2X<&~dMuQ|vFO{_@T6D^
zSS(maFpO|pd16qoFGdqV=i@a`s0V_PgxkuS$5l8{m=rIN{TWK-wiBuECQ(>6j*K$^
z`w7d&lko-skXW`fS=s=M8I~<WmN5V$i)G7_WevbtST=!7FaYzKIA^WQ=t)t6`-+lA
z$%F1E+lwUpd49zhZlK^G!VM%7bwOgOPH9WtN6P3z1-Y<p$GnT<n@k}w*e~Mf@Ae9o
zO~cb<0lr~Qj;0{MDbkb$I2D?j0H;Bt3UFF9UBF!q6U&&BO|6D6#e)#2J*Wp<ZQlM`
ztu~oS_5qjrRodjzX-oi7DEp#6PRDq^$W4Vq(j;j(a8gOPIadYUpeAHpv43PXSyYED
zB)(mBG0Cs;@hLb6hK4Z!3DnwTa!M_yc)$`w(47@C`i%aSw?aW0@3wZ`w3#8mK?H$y
zx4`5#t2u<>*Wm~7<NlPU^GI@u3kTuNEf~;F5@$@3gA;?rszg;qaQU8bO=q9cr-$L#
zda8P4a7zFdxNyA94)VsbsZ8p()d;W$8}au0s*wg$!vh@f;(B{(uiS%7p)AK3S@E*>
zJUoQ1I-ccj5oWy32Nnw!X2O2x6(hT$TF4p<6%xU>1{L@+e}Dro4MOreV1xYNt;c&_
zK=L1a+`DQwmi#G{z(<)2rXdw4XNAO{26C5C4^U%5Vz8e7?}7&>!qwxwPyn)Uk1=#r
zkUSI>7~-c6>j01-;0FF7U>Y425Us(B5&%L$2=+YN$T?==;R4b>0P_DPhXO*`A>lB)
zE!0g!R1769A&J3C$>L<B@k|${GmBtnqfJ1w<P_J)E0N4LTUc(f+WMou!Ul7SiLTl|
tNr`J!RMj=eR4twLdNjJh#!U><b;im+7}_cjHH{pc_9N}xtsN1r(BC*w^UweQ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_groupnorm_bwd_gamma_beta_fp32.cpp.266D9654B5E1C67F.idx b/.cache/clangd/index/test_groupnorm_bwd_gamma_beta_fp32.cpp.266D9654B5E1C67F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..a64ab64d811c58914cfb356ecc5336fb1d64f884
GIT binary patch
literal 3444
zcmaJ@2|QG58$V8r%o*DZnXx8g$u`UoV=FB}_ARs-&V<`gSu!n(7Ao1I4JmE3NLdQq
z)culdk-l3*UkeFWWh={PyJseEO>@8BcYeQlW}fGHpZ)(nXFMG2>^PDTv;uYDhOwhq
zq7VehVxP#Uh|tb?5Tt-W5I3jHC(N_MTxN8(k3%@r<-d9S;kuA3V)RU@R@bbuvOA{f
zn&+t6mmbXc+{wr+{kpJq@_5KtOVtBkiDhs=((-kgafU@M^_@f0s#@E&&$zFDlam`;
zDHq%|ebJ!ip7vL6Wi-pmntSiDxy9>1D{te9iqvD-Nev~mwR^^gCp?@6!k=_TSukG(
zC&;L2c^Xeg?OIxO|K5W9*2W3S;QOiH#kE4&TO9RLiTAVelQ*Q1LfmNgheN~rYj!o7
z9HB5*{P!C%^rfzWCp+Rlj_tO4V;l<UiQyA!CMHb_95Xr!%B}6+iI{44#uv4s@VQ!h
z@3@*{rKy)9Up}5?R<H{)7i$cWm17!Lur?&S+f(EpW*Clh^iik9q_~0fq57=cvk!KT
z%&4qvw#YM(GNHDJo{&y6<qTW08mJC~i2N{~D&O{(ZStiW_mp^+Lxo*peD@U8ezF8o
z-W&GFU&XpO)%Ha-(WZQf^venP>Xz*$W$jn~NZe&a5!ab|hh7a!uqeJ8b@Gas$xHn$
z9lLt}ki2k0y`W|L<t2?<Kj=5MEY8#oHVGzfcvWK|>A$7hdaWpamH%@6{mTn|w;?@k
z2P7L3a8>Wqm}^_q$5X<xPpi3IzMa#T+j>4TyJXC5p;^nrZ6E5M`ehvdlo@FD8b7~~
zp`3p(tH0*vV~>G>v4Wx-T{UqRSIl3~+1Xpp8B07kbECbG6%n07xo321S#aR_SW&&k
zAltOERkl+8shWjxFH4$lcem8qOkA(FsgNxQH;M6XUN6%WpxSBwc#oud%~#{@zJZfY
zO<|k8^Vx?Jlsclh3gT+}Jug4Ha#YtRzI3yv#jBPSqoNC%KC;^1^2T&TR8s;iDn7jR
zEn+2??8#YKTlzGZP&k!97@>?A-n;Bw_^-N+a6PSJG#XwKujR!#To$1c)<2%*^@SRC
zKdJv?uTi%<on5)lyj^LivgfzlYTcPL4}*$3KFV2LwK14Crp7&$f4S`P75kYThx$$L
zW_P+AIx?Aj{dplN-Qbc=dbS-ZH~3=Ltanh%ip-`v&cEL|bo2s~YgwPI<xtV!89bsm
zLsmI9`rPSEsfTl}c2BGMMdRlDeMbxpM-k@~PDImCX}qD)yI9Ud5~J!?-pvco5~eC`
zYED(3UOg3b>fmJftM#O;BWLzWMZ|u-X<BDJS-hj0qi%V&rE#QoXlUg0Vf!(oK9{;p
zTY`od+?*@3YoB}zn6<pNqwLOf#T%o9M<Z-ctC@>OE!nI8{h-kEp4{P6UbXI5xwN6+
z)dY|2xAAXseJ3}Ca9?yQ-RL_?7#fLtYhpz|(kVHmvwT>Zo;q(fa%WiJpa=vZ)|DdL
zY_ww#c>N{_mBGs>^9i;K_>a^p6>`=2yblNz3_++dWGurc5N}~%kCJjlgjr-!syG~g
zsp$|M&nG~ws(L(TMz?BI9Nvhj5gek%Cwx<kQ1yDBiB^r)BJes)ZGbb7<rAQ$zy4s;
z6%h0Vfmv8;(s*eSp8)l$<l&*;Zch0kupefP!{H?O1gL$>^)<@=iohfATFjb=CzAOD
zs5i(DOhpkD%@AlaX8n)CPy_UN<;U7*k`NnH|D!PWnge;~OXl~=z#A~<@)Gg{J^{|-
z5_dn0ZF_$lfqgJF2~Ses6QDlskzX|0dp8Jy{W0g_IB^L+0qSdA4Dp?++8hM-#jK_9
zQgVC()LmIt`aKnS`3UTdsb%r96g~m!Pmh_aqn!id6yeZ05jr0~pSRrkPXv2&<VD=i
zxx;w=yuqXhsD`Ts3zP?@55wQ5KIcR9dqd`0v|z_%W!kL}SrM2o29N;Xm@Th;bEeRD
z+ltS{#(wxd7=*>L;GqK?BA|4HP6Zy+v)|eee$g-$5#?tACI`@i+47+C)7Dtu*bE#3
z7{U(m00iO@&_%9`d_kFJx`mRm@9M3E0xZ?~b2p<aZ!0F|17NORfIAWVFL+y1ui0}A
zcJ&n@{L(AW?&_yIz4}Jye1nd9xe&E`)2+l{0!o%7lRzdIM<J^L&S#k8W`l&5W41a3
z9R(l;$i`~o!TM)Sny#7~z#-6p|2w*kO}003GZ0>Kf&o}g4x0=YWlRc*!JU{JG-x=8
zi1mYa5kNZIxi-zp_0OlpMNzUjnF&r|9FuJElcO*x0-PdKQGiooDhY7POl1L1g{ks`
zi|!F;*Jzuy3FW&xvW&3x=)R?xSa5#ea$@b{-As@#0o8KVS`QLCq*?49n=<?nM?j^h
zQVeh<38*wxngKvXKxL>h46s27s4P{M0d6M&B~pnDFe(B{qLLUOa^m1BWoGj+MGPu9
z8QcN@s3gaE2S2CInq!y{ys;Q2sAN5mSc*&9xL-^WU8vUdh5ldNoW7h85`#@6j{jxD
z5Kx*ljR-Cf#>vqX1UL#!S%6cesS9u#G%W#6ho%R(GvQ(xgT$0ltSBCq06OxO#zKg8
ztT9ZizF6G_lu;mKaq<8XL>9^y^ckBCZ}0t_jk>G4AF#dNe_*HI`<xWaB~5}R1rjKH
zYE!3*uTmFsc{}9IK$55~DkLr|$vny`ZrhJJK}-;H5UJem1o!5$9v+kgF$4A$EW#Iq
z&GhTWyFY71|74ssBoyhvlSd8tcArRCB*8^kmyo90DluM%o#wrLz+$LZ;&w`KrNl6?
z8d;43YHse*c3t<jV+ZEMK+S*(YHnjV`tCm0_Cjd)b}~=J6Kj2Zgfi(m<+U{2eBnAw
zELa?XEY@`1+XS?NfKpwl8$klyj_tk7(^aKHF2UPIMB*3BPOS54C^sqmEEH){Hr<kQ
zd+&gdc<RrbX<QSoQ%DR&usN(n7?40lKmKlj;r-A=Mg&D_@E{iia0fwwKYog!F1WM!
zuNSWR!GS?gP#DVDis{_LJm-mu;l$@lAOuOGl&p-jgPVh^Gs)g=F$wQ1N0DEksAOfm
zbeZk)6)TMu3~e+m^wj^Cl&qqvwn&4jrK78FK%+BE%$Sx7&6SNzS1XXUm#lFKfgQb8
JIf-~ce*+c%`z-(f

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_groupnorm_fwd_fp16.cpp.E01BDD088F0473FE.idx b/.cache/clangd/index/test_groupnorm_fwd_fp16.cpp.E01BDD088F0473FE.idx
new file mode 100755
index 0000000000000000000000000000000000000000..15fac964c6ecbf12fe850b6403ac8b1dce6a96ac
GIT binary patch
literal 3460
zcmaJC3pkT||J`fGZe|-cce1frNz0net*cx@7wMvCyt<*4AyPV$(@DjmNI8A=gf5~)
z(Rt3Hr$}_6(mj!NDkVu5U61d-)xXxh^L+2~yua;tzyIEu{=U9zsR%L&^G}J7NtEFc
z1fhdhk(jXTrXGT%kic7fGAv$xZ=CVXVg8lu!jZ9h?<e)#K77@241;yYo4Di^hhFBl
zt)#E6A$_X<saR3xbor;efxRlPKUYi8-;ikq{7kgHtclka;ZrQb*Q!EV9zFc{_%-Kr
zk<sY1GoKt%r7sumy3)437gxyg>fROg?zrKZb1xn}zG!V*?-`Q!;kwH0^xih#Nrb#d
zs<^$Ed?H7e<%SHky!CRB^SyJLE27L;WhHehdF?Gp3yP?B@S-CX`7Xgh^UOB|PHD+L
z>iG7x*{c%;0gW>c9x+SY@TIpPz-83k7Y&OJ9c?KYk+v_zGq+pc)Tt-ySYe6RfBeWr
z8E3hC*5uf2)ZtrLZM{-TJ6m3Rf6VBjDlWm{aH+vmXI1sM<Cy<)+S=TjeBauYANxBu
z4H<jJZe3#-%9);4l<a)qLyLSXbqnWAd`Sq|c!#V)YB6DFsfky)ekp0xnp2A3uXPmb
zcMk-rh7#Yjdas$s2zQr1@34HH?3|NxfFBb{cyh77#dZ8Y9!`pwf<x$i)uklrp=}Qe
z%97iiX%<}3gv|wlB%~ovblg`k<HC-U{4GmgOFIrnk1ACxpkLtdar+G!JF#<n_eKw$
zmkXY)ylk(i+%Rc?8h1o#t2{lfxYw+F`fum%NG>S%9MT&pIT5=!Z^Q3-eZO|{*T##Q
z`THLIQ5Czt<D9HKDXZ!}du1x_!T7mW)Yg~jh3&OR7RH*bZr(Z0l4@Ba#ozTkeqD0O
zupp^md)8f9u*Z}ax1Y2R9-6Y}^TMdy6)DbwriC{)Se`XKx{iHVdCa9z>0My;WnO;p
z>C0Q==Kis?t}6b3yr(;}R#-dNG$V@TS3xUfZi^#cAMt0SiEL)%{M-@U4=+5fCJs~I
zKD+6DZ<GJVQekiQY>x)n=HBeI!p@l|cBChoWh91It~1Zryl+t7r}5~5sk6fC9ZRg6
zPFBl07eq9A21s)^u`;W1$I>bv^Fm+M8b%zrK5og`{AWG)6~h_T%Ze@Y1*D$!!F7>}
zgLJWpv`<_aP#v5X=a9a;FMc;6I3f0#^HX#Oc4Htx88+KQ!EC-+!u>3Dw=Hf8Jgk}+
zw#=ea*&P0+&d^I--)tI`YSu}+Q@4Bl(6uv@qm%32)i_bRzVK;}tOw)zu7(7S+_HFv
z=MCwa^P*1Xv!V!}w5o<v-3R-VM_UKZnbVxWD0w_*vLZXZEB|wqPgd@Y*;d1m>#jQA
zYsl>Bxb@NYhF-h%n$v}zMVTAQ*Y>12y_v<74IX$C-pU<~e=y{~dTd<H##?3ul}p~{
zvTTl*cwB4jiu~v5nIKF5h*ifn=l+AHmz@t$q{c@L;J}&>Z!AA*9)6LGE{Z@fBa#tQ
z0{}zw;78VeFlm13PDU0Z2<C>k8EXKdJ`&)eJ#XC7i)^zdqhUasPNJJ@z*i=Ma0E}=
z@%8U6kkR=-+m+x-*8s?ND|4N=(T<}sKqCQLgosEQ0F6)CHu!3J<>$F%bPiw>2n3=A
zK<ycS#nu<^w{ajNOa3jiu)8JZQMN9ZjL~oqnkmtgq5)7lV%GM{`{Yd$GP(e;nIxv2
z20(V)@b2YzyF@H98VcCPBx5rTfb5JD=e*V=klGB;g}^tNKxS$HWarI1c<kNN+GsL5
z7qATo21E^j>{9;Jmg|h!&SX>uj5A0KrUpQE+oeqI^YB(V6^Q}b|5KLd9c{{DpPlc?
z=sciJBhgqI0JZy?BvX@u7X4y@F8fBp3`vISgRXfJViqeJ2;lFxzM~{fzhOcYj3eTB
z!`i{_qv+SJf3tzt1O>x{GXN~92jGV#Cy~=+XvuPv-@bkY*9nJebYbW1_jLIU^gD=f
zc#Q+-56%rRsUDhdZV?LRX6Gh=LM|q0$}G;13l79=!*PS2plJ}Zx^z(YYs`Mbw^w<5
zxz;6iuAtJpTfKDr`Z#u^jaBWJ1F~8dW(Wt7K>BF_DVUHZWWZPfW+G&1Vdg?BEzCja
zsD<%_0*JwoP~1<I`MFMkt=uO^(2%(xSM?6Uq~Hqt-~P<{yYX^rXz#mkxk8=LV20Dv
z?VkE51!Gg$OjtgEnX_#n27hrR&GNwBh7HC|_a-l^111F&0SijOodRuBFBFufV0;lj
z@Vn5J2}k$llhxs9>&mA2N4kv2|D;2Pdk9Rbm+*&4N0H+MXmVAGcbiq~wJ$_G#&%~*
zAQ@m1w)6*PE-}}_EF>0Mn5D#03$v0~X<^n9Yc0%1VxxuGN^HMlNly%7&O3>3>!gsT
zI7yLbulE6D?V*CHfXvkQ1GbcciO3>09O5kQi5cs11|JhBm=VuN440OI(Reg5JOmWX
zm}e}8m7-vD9$gHVl7caK3^6=76pYDZiea0w{7dYzYMW|wc5>6e6e~&FRtqp|1Y?33
z9R4i<dK+Kl4!eVS(Y_~OUN~$4EOJh;^8LKk$Amg1*lsuf&Gc#gJ9WtLOo9@qZ{&{>
z2yKM+a6gJx=;seGa%w<0>Un`eFeRX?KA^Cg6pSb0eZOCajHLyc<Y=~z-}2{~sq-Ed
z=KzyJqR<F>s<`akVneF8*O9JT`rzdjyfdalK3TmZTXy3124IBDBaegu)cNjAsT}`A
zJp|g_{%hD%*f(bPD;*;Z&F3y2^zCB;BihHR<yWQiE#mtYAgx|^xU#L5DLup#BYi-&
zXWN^@qQf)wt&gxmZ7Qdb@UDBwz_zQcD+gxTBPq;aTnG!(i9}Tx>a)V#E*y}x<NiC~
z?o-LG`WhW+IC`M^>JJc1o`P{j+=(zQ_$9&-f0~#N;h`7n+$vt^c=|-&?h#j&{!)kh
zxuf_C;R>Nghm0h|#K}g0RH17Gt_<wUca9<<TH&BBN))UNf<%A+i^C?zq|Ms?U;&0k
zMI+JiSezX2X2j3Z!|M|a42fh4m0?6Truk3z4+&!W`SO{hAQN*_GYd;kFQ4%fCQh0>
zhUMbTl?v?tjmoyRu^r*SapXCV95q@v)=ez&uyeO^6-{NaoyG+QFGT}pPnm|BiToF~
C-`R)&

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_groupnorm_fwd_fp32.cpp.06336278490A0D06.idx b/.cache/clangd/index/test_groupnorm_fwd_fp32.cpp.06336278490A0D06.idx
new file mode 100755
index 0000000000000000000000000000000000000000..63ca52661e0649e3ca8681760cd775526ff3daa3
GIT binary patch
literal 3432
zcmaJ?2|Sc*7k}r)4D*g{MvQ&mC8;sST32>aNHih5Dj|i6+!9hZlAV&Kx+Sd&`BcA_
z`(??R7L}B2B}ziKh~z#q_0%-q{eAE6H)rOY|9Q_j&pH3|jHk1M!y0h}vBRC?B12<?
zgb@Uh2A`PN=&;8^2ogsIpW-T?$f###Qjcf#@1`9#3KyER>+3jkUuQm5?x~H~x-l*3
zQ%)HthvRPTbaehtIHPM>)O5}$r{DVeHWvB^;#Q!ag5En-r`Hx>%Lx*$?r&>+**i7R
ztabH}#Jq&R_G-pky!StNx2=0(S97|`Cb#(szQR4dTcbaGJRli(RJZZV(B`7mZR<{`
z3ptO6nRyHZmdQQ5Q#$=}ij!R}bF{wSXMQ^7<+<8iX7Hs&S<9Y?WQ9kMo#jeTo!?(m
z@+WTZlRFik9_Qz(9el=X1!nQzud%M$g|hEH1)6w2#yzvGmsxQfN0ZZNW4E8UKHzwq
zOl?`zo0EtASyjJhHoSrz=yZhC`-nYqu&||rnv&iAAVo8O%ums`LaT~&)3MX}1lxSa
z#7I9$l*4ea7B--szTsw>fB(x%X8j&}YLD52QW4RnOFqgO$C^CTIFFJ&Q)R86U3CxJ
za2MS)Su&Pju{~7vhS9#c#noHhWNfjV<9O$OUd6nW51|t{U2XPA?p0G)Z^FptfosHe
zhbUJs%k}xBjGUrlOPcd7b|vO8E;$%3Z^*0CPhbDRqWw(B+>)5p(hXYrgyW*re7uIk
z37M#mGJLbCRXwIG*=|%k;%t^`)>X6OA=z`w{-}A%YREc#TF8i15xzDj`AE+2o-Y00
zBN=z}kG;HB9)7&NCg|Lj)bi)_P4R>mi+q*Do8Ip%yjOkJKU{X(oqRJzamC9P!krG6
z9<o|Q3$_$wr*;OpTiU<tcztjDwEf{P{=u0W;|vVz{2wJN-kx)Q2mMUeMdR8mn*ybo
z<+<)xTlYu!UR!^qJn}?TUr$Q4X|?a1q+mIx3z8+Yj0mxZ8rN%Of;<CPWoq>FHVjmX
z&5D0~^Vs~^UgzB<rbB5fEN=zv8%j$k?DDM0+ZijH6w5B#A)mDG*tm#o?fKP?UhJDX
zCzb1}DucRK2h>`*SY+;%OQ|GWOeh=B^B#L58gSyF*}B`gZ~C6c%qCTa6)WZ%lKT?f
zuLQ=FN;73FhM8q9mF_tanmdbzBa4Xc(cy0l2GKnH(dX&Kt5z&;S*pa&Ef{U?`F)e@
zypjW$UrAW3ol4aGevQAMPpYh3_{uNV7VqEiwaZnAmhrT=^C4&9?-Qj~S=$;a%N=c*
z<wiyJW8QBE<)6fQoL;wmX1MNaXY$c&&u+vT`THw3J7*QCinr=dMS6@+f0eQES-JU3
z;z#>)5wj9$UV(2Dy(JfI8+;Woj!Tfs#`nEHa?fMw?#9dqQTm0~rQ*7EyIL>Zrv|Wn
zuW=sP2E_f9GvY)`O#JL9#EjI6ue+gRGlhjY)!W6yM+T1)z?#qQK6hT8-Hf4Y0uT^U
zg2od7;60LWYtx623}p=U0kAY#TAn8$tolSfV!*0fjf4gw2)+PWK;{XAdxwBW%W&|X
z@Xb#R7`h6;#zbRjp7;hLh<(QR*oLw%z8Ja^z(gWZj3*%M(|J-pSk<^i3T1=bbHwIQ
zcml!!UfHe3qUu-}x*EVVGEI#qARIAUw4t+`A%~&f0G1+4$?^n*H>=Ch%JTLc!jN?!
zw<%&O#S^gdosIUg>p$CdV945kgOe(1thYy#+eoNCfH5LQ;|Ul%{c*pdc!Ot~I1&or
z|7lB(XX(XHgViAz>I+~JktD_w@a+<P$Hs@$6$Tg@1S+SJsWhH|@Z0vbtYX{zNeuM^
zuq0Vhjwc{IT*q?UvUJTZ1vKot2^S@caxZk=6CJuX<^&P^{62S-%$qkn5Czv@XxPBc
z!|4Mj?&yCyX*0B!Kqtrn6#X{)?}>zY+R6KwGmTr<s;D3Cb{ZYgWDtaTZHDcI1OW-S
z?D+mFLcvYdObwxvu}nf%s@8-t0TidAu@L5E8e9;U?rE|X2x+U+p0J?*coZ;m8xKbT
zxke!W!ZFt-?bl~f4_^?Q%K=z`*Q?src<-T4-(y3vc)L?|2eYk2Pzp{Lr_*31K$EAd
zLJfX{OkB>wrumFVEVIGFu7HsG6n#191z8L_z@;$srr^2^UDqE@cNd@M<YL@%_}>k3
za}G4t$ek82hhLtJoqCb>hVX45a6WJ$Kql@s{LG}q&{_;L!IOev^7rG{UT%)FX;W%?
zFhdf?>E?76d>3dey2VdTo+ZzxDX<jyG)0yopQgl8;?tB_%0IL%uSucxy37uNBPPUD
zQKHfkZGoA8$dEDrw|aPR&<F0~!jV((1=s>QY*wmv#qu2o#s`QLTtZKR3AcrUOX^87
z;RaH0DLpAB>?;MA){|z!m7(BNJt`BfG6kpU(U@@7dq?e+SZ(_i0O^QQL|I79bQhuB
z?=ol20lWA&7O+c;ZV1~7`TTvju=)e;S(AR-E}9DTDO+vT-^<)a@&(NB0wR%rxp*kJ
ziKGb??j+D;Oyu}9c@rf*O~pi=Pt!Ef;nVa?454;tqe$*3^}uCtRxWj6z3^T@3A_Qq
zw*Sgl*gY^Q?kxxhN5OR%IzR5H(^3`%DOd<y;E3nmro{X879Rk4O~gzjU^y3BZ5ma`
zH`N8a%h$hnpDt{G3z(}a^U{JUI+8&UOb;`H1zd5+k1JdBnoCN4_T8AZG98{@G$s(_
z)}5N>Qiow02*UM-9cbK>W$;97vJrT5Hw`XclVV(-Se%3iFsss4<>A=d`gJ|n-Osjy
zAjYc3TJSi%NtBHJE@_$qtKP>dOrkdU_y|<eU+8VS(OiuU%=~q~a_Jmo9lUv2z#Fn0
zbf5bt0}oHZH5r<}!MfnT4MFS|&VFWlH(yzJ;hjLHY3Xdsi1MB91<YUCi)V;;iG2cQ
zBsw%ANCWf>rbZx5;9PznG0`D0n%vF?!}cOb$d8`{yzNOnHbc2kg$9QpA(41Q6u_28
zdI<@O5J{q97)6{aAuT28yv*5SDb2}2pGICPBR@x0LD9<EcG2SB?3T=zGq%yTFjW6H
tD_vPdRYOxtN6)}$?mW{43z;lSHFG5shNB!^*UZ&@J?gT;-i_di`~xYP)&&3n

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_layernorm2d_bwd_data_fp32.cpp.65C39944BBD4A16D.idx b/.cache/clangd/index/test_layernorm2d_bwd_data_fp32.cpp.65C39944BBD4A16D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d6adb6e5cc79079cb8894c1baa47271233836b42
GIT binary patch
literal 3440
zcmaJ?2{@E{7k}q9X3S!2WoDRWEMv(`Tua81k%X}eDM`YsEJY!ARFYfVo4wpph?I|%
zHX*JoEh?0juWnI9Qhc9KR4U!Rcc%WD=IQz7d7N?1`JerF&UiaJI)=+bkUioo+!z)g
zjD;Xb3H^(UkB#V)fuL?81PP03{T|0Ywje+F62Eq+uG`!+QNXZ%Yew}f@JJ{z%awCx
z_>>(e8x^+fuyd|cyD-xBs`=4XORStC)d2njax3QSg72&wO5AJPV%;it`WkCrdvDv&
zEp|f@vv>@O3A*v`4_#>^Gt=RC3!84tnu3G9Hn{$=$k-1f1=JG0$G#6&KASZhyxpZh
z7wmg~KGCG5L0Hsh`$U&|A#;0{(3bh@Rj#4C!{!mMBYB<m=A3Fmq;Hg7V9}uy<`f@?
z_TQ>>-%hG+?h)vORGL1J>#<KxcGvOANcc=rjf{$|yg|_&%h47tX`{Zr`j~UCctu(2
zO}k*14%xe_La%(;H`<f$O|=ugJHUA{pI3MJ2)lNwUodp>Z=CmmvUjbY1lAGp`1LV;
z7q^ivo%f5trxai#jm{UF6y(gfb)OZ$Ph37$VeKr<Pw=#l*Zo<#ewA1Goulh*9_Ez~
z1VEG)lO0jn#obOuoDlh!8}kxFd()l`TNJpAb@Xq4z@n?&NFG2+-O>$Ot<6cjP~@?P
z`y&|f&s+5?2?jMqm6)0QsuZ17dzP%bxsH2uK|yXsU!>0A!OX>m&ICE<+~aJ{*aV}I
zw<FLeCyH`t+`s}`a;cp1UZ(m^TFhTpBIksxw8mS;D!tvQ<CM^TplodgyD^xzJErP|
zg{*H8FJVF1$BDC{fe~CPuBdn?PZ8&o@uneY_3*>$-f&f3XR0|zo>OCm?Qu9ez<7a;
z$a{&DZeDpOEu8U*;livPHcXB?XOtj35!tr<K?>)xEH^2<a#K*!9|Ip(#!h+KTMe3+
zR@dHhx{|svpgC(7wXfrHNFYw7aIu<AnMX~v&drWZ%7t==Ki*HAgW);xOWdb<&rh-j
zt5{Y_-eJ|UKI$!+Z5uAtc1F+lcTV1+m=Yh=l{7bH&yjaJtgC*OIpyarx(l{3d0lCn
z<F+ctaYnTGi|F$*g-_zlZ`-SW$?Ui6iHX&HQfX+o=poCZur1X%^{B=_3A-K}Unm*)
zdpBh=>-@d!Ja*uorAMxp)J@9=EXdbiJ?YW6iCyQm@8NZ)V^yb*9Xx4Or+czL=kDua
z4Ud0iyZW#9H*0Q_b#Gwxz8~EblH7lM@1AR;NrNNl1J}KJRwT=Jwp{3(Ncv^yT9e~g
z>4J}?^6$^h)4#BLYk5Aa>CruftZZiY?}e9AJD0n*z$J&9mas9=EtWmY&8BH<>d#yJ
z(~@4P-S~7_!+1oa(Q9ZbL(|nhoBe+3`19CDlXrr+m(8C)ZyY)6D=0Ra+0Jfqq2&@h
z?W&$c{l~U>lEV2**L|jSas`2B$4*S+>+wG-STTna;u1H8Okg00%^ALfZGLxy1P4G6
zVt_MH5)<?eh(8WlZ(c{9o>@tPSE6bh4o46ZpcYv7RTR!UafSq~N7aUqp`w^TqFaJ~
zJg40!mzQU)BEhSIAViU<NEMT>N(c%JUMRD;F+Y<82cv2QqJoN;05w8UKkeodZ%cxg
zqiSWMGF?o7IuaAmT7P$qzC09$#x{bCh++a_$F`sGDLPRkP=o_f>$!xvWHABOahED6
zyOV3J@$eednuH@U#002!r=Q9^d-F&u3HC#+RfsBc#004Kc~_iyH+&(K1pA}bcpRP}
zCP2Oa%+`#sYtus{Xx-1+a_n&X(dl81HVIyVI;RjRbHoH#*XZR(E4b9lk>J&+T8XH{
z5EG!jzBjJvpDpgmczEs4>K<*gblT#+cBdK~@y!B}Bg%;myZ93uwl3}%4uWJ%6H}X`
z!;&y$gvn&ufNp6PJ{3%iX)nZJ#e)Qe0F=PQAJ9QY4Cfn~fk!-pQ}Rmv9X3+7wz~JX
znGTl5pkmRGz}Nt?FvOoC3WmQT>08=A+PZe=ygUJWu0)d;VeZ?TS-7EDim(PoOhz>L
z8m^#1pP{R_5`9PBeajLc*w;f@Z+Pz_S(uE_<Y_A4HOkRxEWr7_+2F+|Mbx9VI%FLx
z5Ca^eaYZ2h9+$;uIe|xOy6mco^GKq&lD5Ha&OwG+d7q@j0Q;y*5%9me)Z%No1DBwc
z82Bslzx83Wkbj%eaqlx8i_k1+mf$rOp_A$I5}Yc1wgObALD!Vfap+tL&YaGZ;Oyv2
z00+L&A&NE(qzS&!>_wXe=0iq|NJg3fzgu47!>;M^9!XWcEv}Xvlb&pgiY4I!77}$S
z+C~t8jA-+<1t5)7o$6&t*>7Ip$Owf^F$H@@MikfzreI2BM3Jp%3QiLlQDQ5Zf|Vm9
z%4}s*(0nqY!d5W_k$<0XU|8PHR6{|BQ^ctP{ARew?0%K=#R3J6<es6xVbeGuv22gr
zY5!l(m`K&SS7X(bm)qEWDKWS%c;XLNg^X|&xXNG^QEm>GD#6jY3<-|O)s*10xw;aZ
zK9>VHu)F_v-3~>|=1nMP*PxY)pbN$c+KY}vbi=-ngvn>lj^wza)q!24P?}VUCu5?z
zr<UR`e}gW-#L55eO7tHn%0?q`30w-u^3*MUJCoR~Ddpr`c;9#i)&P+bgJGh1iH_&@
zyfpcmUSJ}%=l@I(u730z8i`~~S_mH1JMI^rUi?x7N=fuJeOY8OYFzMIDpJe!E1eaN
zqbjHq(QD8cPz~BtN%2>g<Mu(uR+ml2+tYFl9ulV8QJ10{2JTRYy6_o6NRdUw>NIsa
zNcZ-jo@c<*#|da81C{|BH0GIQev0x9KR>Cw2J?Mwqb)Q7P_bkcLoPkTmIs?_q+Wwx
z3=H811drL;zLzIQPDzaj<PT%SzGrLwf%_2P+gWdR>XlRipOp4mZ#cJOTuMB1zvwfr
z3-?S)48?|R2-ZTI3WfuLU|@K&nz-1|IBik;L%|FnC}j3u45BH#+FE5t<G~@JQ0PWv
zLky~0w$VohD~rR+5lCctWr~ubg0q*ir@M;4alQ)CeGYxDs+#&D8{5TpOYD~#Qw{k#
xRvgWrNon(#EG=!euD$_xfsu)+8PC#M!-Db4LMJLs&)n4`9Cq<_aKm^*{{s~v=rsTU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_layernorm2d_bwd_gamma_beta_fp32.cpp.B3B6429EB0BBF1C4.idx b/.cache/clangd/index/test_layernorm2d_bwd_gamma_beta_fp32.cpp.B3B6429EB0BBF1C4.idx
new file mode 100755
index 0000000000000000000000000000000000000000..77f0325b797a1a14a578a6d502ba8ab62f22d97c
GIT binary patch
literal 3450
zcmaJ?3pmql8~^*8VH=yxVapsF3v0@14teS1FvW!AP=vpniA8kKS$&j5y(n)V<x9>@
zDc@H~>8&V9k)l%6>(G=#`uI}b|5i_}U0vU<YxlML_}usJ`91eMc#e(<DiE{`;U&j}
zZVJLc5Tu5F1)CBg`sE?W3=cuc1-0HWf^IJ1@yy`$=I$Eq<}V~>(oj^uwumiH?FRML
zvXjp^A1W3lUrQ}y+&<KGmt6fRn34>U-Unkya(T9EunL7O&GAR<x7#!*vCZolkGjX3
zCKlx1Q{^ijJZwJjG+2M0KV46STEdjSwSqF<y}zOEYJpq*y3tqNoe?#4*G?GqXG~3I
z(RgDi_0F1|{lxQ^Kb?3zon~z>Zx^_BQ-dwr)ggY+y)dV|jvG~@9O<RaTvw1^##Qxn
zXv(fIo=81p6;n(Ww&zD;hKsIUJCj}NaCwTT85xyOdgtBBXF5(J2hu5{cX~A<TA2@*
z{#7{7=nWiEDSWlVA*pXgX~tsPl)F0PdD{L#?%94@HWnw?Mai3ZdTd>9{m^aJbK!ZS
zp56AnFMZ1wyj<)G*XH^XX)S9~OQQ1_a<Ox_d+#o!Yj>Pb^YU+gz+uht3~xJs^6ipG
z8#Q+}Rv+BKb>fRo=V}Pk#xFKq5ys-<H_=^<_AjssoLG}4H)e0;J16H9PPD6KspZbD
z0oTx3iaK@6j<#bo{HoeBr!0Nt%`TAl-J`bTE%#qg<aV{g#pbrguC>`@Qo8;so;Up^
z{jz1Mr)O&?-J;*r3ezKW)ftnod633=J{qYWyg5lV;kEG)?RWZ94gc)je`Jf&p2PLR
z<9+Jx@z0;sx(_urswVzR_>8)73;bfS4}EU#`_BeN_T@>Z^PlPFc2z8_Xd5DhYhBkA
z<WY)UPv-2q6|}^^FkU1GPia53;tN-3&ahow+gml17j?s^VjHzItJ>~nmfacp=Z>uV
zbu9;@D?hYt4C&U4>wCH9Qq4J^-0L^FS&r{rn0{U44WV`aaNWz>F0xJExqo)xy^?9|
zp^obK$HW|^Qj4eAMU_2S!53e@2)%T)qF7IFx&HICs6dDHhG(}oX$e<XosM8-C`a6m
z?~C@aEMd++oXzfS%xSxRp!2si0(#lxRt&!_^#tb?FERerzh@erve*Yr)D^~ZwYLQb
zKG9-b2hR1>nGW0V9@aYJUZ3H4j7`Q~o{1D5b5@V_Ir(SiO^84Czsc~U9f|6Kf8KtK
zXiRf$baif)w#S>iWo58qH}W_q-T%vtT03GHxX7YS!KowpKg$mji$=|vtaGQUc0bIV
z9-pGVpm20Qn8%pvb{!KeJKMiyq9Y?r_{q_UC~BFwpWozn!Mo9+Iw&lxx`@}%<>az8
z{Y!Ga#Wr$l@8rx-y{ahH=}rA575nFl9Cti97UyHzRKDITp;eSJVJ_<8DmQhPA*ESY
zN+vl+KWqtVGrr;~OsMWkE}5L`y~*MDWK+X?y6!3GG4E8DZ6Anc!%9(*ASEVvL=J-J
zV=i~f7hd;OfkGh&F^9}?5&?<#0{YP(pAal#KN8{LwSf>q!jb3_@zn`I46}g;nEH=*
z@UTCcZHhHjlL+8uF_<>Bu~iBrcpaLpp{zlW2;k;~s}ScMb2^KMSEFt<oEl9czPmf`
zzBNz?8^*&ysGEo*QX~SnmkB!6E+4?$Q-s(5s&j`!W22E((>{3E7j<K?SY?R-*`7BN
zl;FV*1P^UM-E4?Wk_a%j|0doaQ&EGJif{nx#$)jY5&^OUHY+t09qrA<!`^6X6dXlM
zB7i&Y%h_{ExJVN`yasiXab&7Q0Jp%gY~$R@#shdL>{n-Jdz4gv9JmyMhy75uB34mZ
zB0zS51rxrsP-`3y`=G^DajNPP0o<ZLt(_8G*YBJKM|?9ON;oC)ikEy6LN^FbV<AY%
zXz;{i{_A`>SaRP0OhAWkW{?GF0OKUr56~eXdVIYgaLWcKdYnhz&0Z*nkw6I$1{^_i
z#RLH*2?(3bwgy3<h8#}95O4rs2%0OV2QU*5eZGDG$l-kB2^u6O$&VmizOD}lvNWw!
zQG40^n@n@|#_d}<eMNDoS<HPfK7a%MiVo=cYf$k;jZ%*tN2&nB(4NKg0zocHH}cD}
zl0M5IyQ0~szYOO4$f1LyYy;5%0Q4kj^Uv*gzPZ_w;x+j3Tl0WFUT$ROz!6&on1E1K
zs1z_Ns?n%=K=U4pcDKVv)S<aX1S2hA1}LGRiP;U>At2^>b7L?*pg9^O<~Yb9AclOy
zwV+bK{UsGAbMWHkqz?#M$qQr~=6sSdgQtO(6u*lfB@Orn>p;o)WV=>+^TW@I7=&s?
zwFYJkLL<;rq%=+1_jypCKFvVNXGUX5X<VA6l(vMn6leg;|BHWOzhdYursWfvS+)vl
z332H=>`}8cdG;CAtlFMA_7|3bs4`SJ;Jp$MB7?{QWF;UZ28jdS6#-FWsBr+>2?&`%
z<^ZG!2!%o6fUDR);y|}<Jva-c5tfA21RTt8k>5Lfc*Y8)g7j-gslcF`fmt)%vL<~u
zS6j%8aYWzmgZs*N$7IZ44JqP&T0aDYMP!k|qeeA#mX?%8W6`BFU6z5AHk)NErOjcP
z0WCjPA!me~c>x_tOd&u<fE>yd@tXU=6&=2gFBl5EMp*8rMfUSrJ4>?peGfi_Dl=mF
z+dV11uL?5Jb6LtPRnSgli(Qj0uHHZ<S{VM~eFnx9kulevKbjs?^JF(#1kc6WfG+DC
zk0n<v?iX`LnjBYWY%N6dkB`a}xpuqpZi%Caf))`+0|c7VvP?UbC!5e{@pb@YZPrfi
zRZdn_K+W1zZ5n{0wU5d7@99oLi<s(}GQc=Jx<vKD<=8~jtYf0%2BLedb9a&Nczerq
z(wFOHA8V!WkD8_Xe&f<LV4YWgK_(jf+sG;Z#Jeefue)^%Du-o?d{R5IC|a2MUdBA#
zUho-vAKNQqh7v-fgA7q90HhGuO91lk9zjBgU^W^lha3oV$w~;Y4TXb4pb!8f1O@;2
z84`m;$D!(~7*BbO0#;E;8Bb6ltE!QRJa-=7mE!DZLczJJ(=;?^Y1{s0zj(>gWy|Ml
zvF(g(%nW{IrRwPF8O~-H&oO1qHMig_u(V#JZ$+QC&`FER<gRcFhh4lJR?2xme*-5%
BwL<^^

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_layernorm2d_fwd_fp16.cpp.49CB42570B950D66.idx b/.cache/clangd/index/test_layernorm2d_fwd_fp16.cpp.49CB42570B950D66.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dc2d7531a9572bd8489abb20bff63cfd17d05d48
GIT binary patch
literal 3436
zcmaJ?2|Sef8-KkqV=l*xFc|kS#w;ZoglMDWK6X+R@=lJ(SxZVq$3_&Aa&6ZpgrYVf
zDiu*|rHHLI6bez24%y#K{c4*3e*W+0<Kek}&-XcbIM~}`6a@JhbBGJ)#RiBY2qFu9
zVq&91x@RHCULu0T<)7#FMEA`n-=7)2oWd!t%{y$_zb42kFI8{crCX1}UPNt@E6zM%
zH2ondGc4jjY|+a<jR;XTD#$gvbVL)$;<X$>&V@u}1m(5sDPFn#wiEjl8NY$!Qm)(R
zE<H7|tnk(*1>LH9`>gz9AIw_q>7@2tEUudFx25)3<As4J<AhE5+BsE2*Y|JNes5f$
zQ&rpgzJGkk`eI#Bzt#1emoK?<(@35kzGiZR{jBm^xnW64_a8VY6rDW1w{hn&%z>Nr
zAwD(E-BD*vvF9%&i>sgG9CZ}rM<)H|cs;;8GOXMz3+=~}Qq+2PR^*(p`h6+s@b;?`
z`U;EJrBP;<V;Ac8TS#BM;WK(D=T;XrIiurlvR3x^3T5vz+UN20OS&D7`<P9SJ$<7m
znPF;gCC;Lry5zJV`_Ay|1%Zv4gvfP6_fT=3cWJu48&BV^J=C^%QEK6cZ;^><0~-0^
zkMRVH?#R8X4_=dR>#=xyqoyFGsxNHW!<;McD`V)m<_3cC^0nJpm#x}-JGXTYhKGdW
z%U&OG_n;jJzUYA;bFIX2gLR09E9diM!&mRN*l_xf*mMK*u(po6ob9p9u&#SCqmQS;
z?HJ|eyu?x&na%h-R-JUSuX5s$e5u=SjStP6_vf9O#Wt@rSeu!YoT;+wk^bgz(*gY>
z{WVd+M>`q=N;m90`?um8i&+us>~tKjAwqTHQfO!H^+!>~w^ueQl}YXSlRg|PH807o
zfQKF3c<7#s`+{`ybC#iZtUr$i_&pl);-=J=NObOS^nO3b<LRXuxyCnV`v>>s#<qm7
z#TxJO8q>m+w`Aef>as6hFOih1T^zcFSrCjl9+JO8^W5REB}6M_+itBq=cZ4bj#8h#
zFg0CS`IZxo+>zd05+@x;S4lbE^!SjPC3&JJdZbm-Nj%%AOSUSS;T--xcmJejc>(3+
z+F1EF%pptFLW&EnZuDZV!qJol`!jped-4gcC8;s()A+WNQR|(`nZM1LKTW*NSn{{q
z;F%}YtlJh2mFK5ID&K`DnYmX#b3FZ#i5|l4N17}xWu;t93tV@Xnn=%=a_QaK5v5UT
zl2+c_N7<CWyS%o(Nyj-(>5=qQ^|6JM^XJ?#-MuQIo3^`RW=GR&`S?YXgPpp^<Esn{
zLm5_mDWin9UKNFxy8QC0s%n!f>Nh7oUXrQKX<+7hPV^MMcsH%Txw@@yYUV`-_u*%T
zT+^nPTm7YNel_bjQoOBSpCRY_N^6>N-^V%q=98Su9)|I@W1n4XY6BW|>TaBgjTruX
zB^%EkI%qP^IlOggr1Fx?&d0jrcl_erTN7g9!`F<7fo)G5MN|vYMo6gtY6O!aO3?%a
z+yKF+Z7}$kc>QEE30;XG7z@Xe6%eSl+OqQP^PvlyNGKP0%MxYj0s_@`X|Kma%RhOO
z(B;4zhr<yB1gibjR188oyi4(@A5a@1MnnNY_*V#gSWlZ->Z(%|61oCBE=`nH5RflQ
z1j+O$s+b(S5J*D3fh`_~CkO~s=gp)p4x<HG%b-5Mdp2P<SwNt=WY|H0koCxpg!%$C
zjYv}$5U4Jj&L0b1EU!mGy?|PVC?hW*P+d_J`&UX$&Sw%D0Mt|>l_nriUFXj@xMJO&
zcoMP}Kr=)PWdsB!R+ntVJ-Jb3Jsw>J)Fd2fwtztI4%$h{cmC?uBqZp+yoaxwFWKPi
zzfB1Z`Raj55+(UpTkwsTH9v9E3t34)c)<IAiW<#Z8*?1@z3(Vd&~FHdj4^B&4zP9_
zc~6#(jcS>Si3{+<wBS$xU;aeE$Yjh=)9@!~#9@_DWdCWg67hu-1AO`8@;$D)jk+t0
zyHK|yVP4<S2;j?40Hy@T2}k^4jK~;UnoWhE04K**5aQ@;6(LTGtt-Sa*#?kvY|$Ag
zL0!GYKpLQ7Fe%Sx*KWsq4NZPc1NO$tg{n80VJU%<F-3|Z4FU%oT~Q5k+}B}l)})YX
z;HyK{QGnvrZMLTivq=0Wg~f0iQ1r_dd`0l31i!!c@k(l!ln!um@H5oR`Y7THR|Qm+
zzbD_Ts%oR^4C4mfqEe&%C_x}YhCX8<6i3Bbx2v?=oyLn}ie`%DP%Mto$#jYkCr|&D
z0nn+_HH35qbhZ#TpUx5DZ0U<2w_yO!yP?PF5<#C3bDkWLvUMR43%7fxPNn;n^w-aE
zWK4=F#etWKj7c-4IWS`~Cc~8Bz;z{KvP@YHoG~&+Wl}kCSIHQSN#nr1-#uohV&1-2
z2^0e-gOi7&mAY)!j*)#cW}vUaO9lF>t!Mz#NO##kwc^(bW05p6)K^ZpIgMtEh~b?A
zz2P4zI9Om#RoNPF`VOy`$Qq-jp966CGXafYL@}b^c)`BHX+XsU1^YHRKX9jHEDk1<
z0u@Ezl-Q8AQ6`?XihUU~=nLxtQ1Xu<j6}vXZ8W(MM(NAs_!Up{(*a60fh`5Wm)*9$
zsY<NZ5Qzkv0-od_(C<&G*{C_eCt-Jh5&w*SH`1`t_;!cR20wn4D$c@00-UcpkP=YY
zl?33BSR_N}TWy~mS8g-FpEBWbf`<w0m+*p%7=1m>01p2y!#<EPBa)F4tktC%OXK0W
zx#J>m8k!ngi|mJK01iJc^u4)jKdX~4brV?g<H8Hlq8!&th?9~4Vr4~TI;=-~pPsAF
zlb(&>5tbTD8<rS){a^6C<~iG`&0MYsuD%?vh3n1KeSlcF9?zC_51JpUKPO@h|GkJ2
ze%JtbgTIaqOm5<F$iVU5b^nM=6S!8b)h1;lA|XG@W-JIhoA^pZ{HY^<8rO#F6%ixR
zys!XG5DE4ffgOgw5!jKK=)f2)U@3-Kde?a8?Pz&{t_eieK$Hj)DD-DI78U_m_i)cy
z;u1K#B!NVxP^DyLq#fKGT%Bo4?e%FyXF2+8c_n2_tA#eUi+)}_SHaL)$HGA4zod!`
rRW(g5ZCxgd{gaXLJQI%j0(CQ$Ii^b#6!qpix&))kez9{B^FaOue~iT|

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_layernorm2d_fwd_fp32.cpp.57C9823195A1B889.idx b/.cache/clangd/index/test_layernorm2d_fwd_fp32.cpp.57C9823195A1B889.idx
new file mode 100755
index 0000000000000000000000000000000000000000..bde3c4a0ed509cb6a29775964b4a2eb9296bf62c
GIT binary patch
literal 3438
zcmaJ?2{e@J8~=PX!^~ioEMx3TO=EDA%viFOEnAkPw2)8KHDm@MQAxTfS&~S)C6_`@
zB3tN0NkZGbg|rZHLx|Fq>;KKvdz<ET{@*zsp7ni}-}CTrw6ohG4MD3B$7oJqgues?
zK_v8t7s1`xD-J=&a1a!oSLNNy?PU}GT!`=+yxwiRBHkg7J!7Qcmf;$8&L~aNQQ7n2
zsf(l0P4U)_*J(Oql`s0T53$u{pngqzh_t_r7q|}UDOFbZ++HMp>=k!Pr?$?lsit<V
z{vXLj;;#mDV@%aHpYA+y0&{bu4T^SIUcVQ2vv2Uv+fRSBh}(B#?U3xJ$8X{+XljOO
z+UJ5kAo09^jT#baTsp(3F>aGFh9irw)fma$E?j-URezwT^zM>k_p-AQ;X5w(u$|>}
z@D;gF%6(gQYGoZM?DWzKmYti|%W~H1nQ8JXE^8~)tS_jsOnpU_vus(ht=wU+MY_Bp
zHfh;1g*Cq(mtM$5o;IGWlf0Eu{d{>>?NvF9)}=zMJ>%^)<FpFxDr}>Dk7I!udv4<O
zV|&T;Dm!tBUF711`>|OMN2h23&6-lXbw{p|B?g}r=i9C0EL+{R%R0v<C4X#lxsh}e
zy!)T5Nng|0*NP3LNA&WlU51WF(&U0#yS=#d#1HG=8xqnID9Cx|ql}y4oi?ofJ;RZK
zO4zg9B6kmpS8%O|*+f_c=2Wl_F6XLo0EzSPkm*cmRz#{UT&u04zS#OgMhN=>Z~Vn<
zq|2%+CV}xKL}Hxy8PglGcWr3#BQhoKNBX)=?xvq9#xhMRoOWa!Ow3R|+((ZKSowC@
z@BMY*!RMbOC!hT*>kc>I=7?0ZxPN#heQw`}-`}sVf5;3?#dExsyn`bhD%<@Ec&8~B
zUGg)N?)Y2NYzNpw52lK3Kh2&=-WS>Ixn6htvdm_ZzTZf*L(-?eJqcZ{83j#BkEqo-
z>3UA*q*>QFGaS{p32%A%cZTai$UCR6L|n+y^Y_NMIg^&m)TYp3)zmnP`u?W#Tj!O6
z=6?AtjghJ6Q}r9o<G%d!dgW6YS<?fFk{5k>V<nQ#5>*D>q-$Iimw?Z?X;I3o&cl(s
zSn>z@h^B=(RpnT8A8%#r`J+wt6XdaK6H309f<peBY+gtIt<yH{#0B21vE1GhHH8oS
zo5I)kL}t9vooq+mKd#(;VRuKi&mL{~^qZN<hxC$n^daNavNpynC&JwHi07pNkH0p`
zZ%J!)nx0);7s^e+=S~^bsZE62CX8~B3sT8_(LLnk@`aPFQ%W&5*N3x>3yPw)m>ZeK
zw&?Vc|IXQJ5qq`aV@qdfRO8{P?rx&v!RP<(4|aQ~=6l()x@h9HY0AJtiM?syr#)Wd
z?>Msu`o!k_HzZz%Gm0X-xDB7Ct2cyInJjr}aG=m;>fo#SbJKg84?9I4_aA)mY`!JS
ztaLcnXp)ta;INx!pYimCQ-47M6cE-P2JvDz+b6`(eUBf9RL>qC!^3`CAw&izLlF@4
z3J5-J-QfokjUVsg;Y|>PFfa^~fB?0rdPh=1(LYIeXa}m+hxCa80`V^i`mNfxX=Z4o
zIu;Lmqrpiy5>-IHDj~@B_|)XCD_?x@@CMWygTY7%2vEB;A2Q!mOWuHof_@ep>Fs2v
zY$=w;!yD1yvN%}<0Ri6et{W@Vk3?2VLxHHb0c3y^5Rh7iM^X95;pzZ9?1QSY7_5|l
z0PiykDQiL~K^8>V7Y!~aB}WhtpuRlns34WuXN!k7qiPC{qAnmnT{fRLv1^U|QatR1
zs);zFynq08c~QiJqjlN;;$eSOO~#Qa0s_?aKCCzFOB!`pc-zl1zmaI*{YR8?C>Gv=
zs__`SoPYrD(c32WkuHAwmEfJ<M2I9#l7H3(A2)CZuK@FX?l4X;ZxED#s932ug6icn
zY%&HrMh=R>g5H6|G|s$qm>AbuDJCJ9HUI&@3L26R6!0Pe(bv@114b-HnKbFuy;UN<
zKn=rR0rbsTLhvpaC4dw903S==b5^%fcLO<FE_t$d#lZP6^a1|RfiJ+C814s*Io;EF
zdl`1~5ks>?(F<}0(EGpXcn-F7#CVN-{8kF^@p2_E9Zol21QQTNX+;Xi3+1SaYJl^e
z3R!1?-&uqD>JW4kfOu=Cb!mPko*z<J?C<QjO<yx}PDBhg0PP1K{@?qdYNhG|LI&L@
zleycZgU||9mZ?|*acHgY*w(H9KUnOC(f&B-@Mw!J<$I6ju@Z<PThRpgN+47MRa%IX
zr~X#~s#B+G2<ddGOd-yg$`azNsWyO%e2Wddvy}Brq{O|vDni)N1WQybJWVjwZ@>g&
zOVX)u-<LWygdrd@bQu=7Z3IM?F3SSlA|OOMkp)(ifRN}U7C;sOA=Ak$u%85kLZ`66
zt|w2}Dw}j2Q$ibwA!6hKpi<U~9~wKcz(&U{yl3dRwH0+ib*Zjtvm5u68;VqyuD)q{
zo%48>h#1@^EbfP!L_nCbOfuLqlv~VH5aOszWg$+LsUgH^F_#E&bfzxgazYnnPLNY8
z(W3Z(1C0cEquav2eE?a-MAK7=_y9$bd_jj$9P&>guqGgyR+`=*RLQ%<n2oRVQqh-~
zQcM|;Kv{=Hn<}nRLnI^^4;qqxKEH=lvr=;gA;FwbBmNQnZlqzQ@!t)*8vNpYiUb1@
zd7$dr$)o-i&kmyL;2C&*;9IRUcX?Csr9zR5S7VnIeZ}f6ADzN51P2j46MQk~-d7{A
zVIVY`CI46hBLZT8H&6l*8WzkAN0;YLie%Z;+T31bH%dXX6n>!Xa2n&O)NC7S&A(dU
z^4z0E_e({~EJDRJMH&^fscT@VoA1lsC^QH|jiC)X-*1wAfPB~6Tja@sY%j}Dwz@AW
z7A|#gZO^dD>BdSCYw+JjOzH>pfA;liuQ4hc6A3a$N;eC+8vkBI{N+jBJf;)VFCvDx
zfg%2yXurU$Aka$i8v^a)aRYc-sHGS(yWxulE$~qbygdNg4tfni0YZOrkdQExb?11B
zODw`*C8h8LX|fE7DC@Y+(anY8V7H8db6HH4lUJgdud=kVwpqPqxq`lhj;XH3&!mbf
rs%n~A+DqsRrk;V}3L}<@nL1nfmzDMkic5{1T!Z1Y>usIIJfQyp*<H$7

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_layernorm4d_fwd_fp16.cpp.4CF1AE7F957B873D.idx b/.cache/clangd/index/test_layernorm4d_fwd_fp16.cpp.4CF1AE7F957B873D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..027de8bca2856845dcb60a026c4f3914337747b4
GIT binary patch
literal 3396
zcmaJ?2{e>@8-HhLFf+`UvCP=_WlST)SjLuwC=C)(!Ye5eqmbLxa!V0H;VWs=RUwk1
zTXHX#8$v?)D9OH+C0o(;y)*SM&FOsaIiBNL|IhM!o^f$lw~i$WK{l{M)ZTyyUqJ|h
z#F0NvM0ilA00f=JLQquJJ@>twP9yw_rQyHgGw6FbgM?G$+{Xje895p6(!@y0rvl?!
z?dy^jMupZFzt-{$ow=_oGI=-(BI<?<5>c6~EtnOT8|oZVDt<Ao6lUc5q~<;RkZ>Wd
z&(Fw09#ZtFy<6Q~8637)4;h<Np+fC}NM}^p!$*J2A8d##Ox{x4^Y+*DEZu=LJpqTQ
z=*RX_%t7b73vr!&i|HwI5|?sYHtHoiy0X1ZC5HNFw;Rp}AC+mVagYxD{Z?X?8@pI7
z&*AeU+bkoOQywEuhDXJUEMK|QwqAX`h&$17-kVk&6Ti?aoXwuJVSSj1FBv2UDRVYu
zriNRWyl9_@wLQ@5cr!uiQB=MBhb&ny-&Ltw5AMqex7Z_~?dTHI9g}l$^&iLX5)>^C
z{YLcMv)byi#*vQt{4LZ5MumY9JrycVx6rZPl$1CUB-oa&Kb<diR=Z^Azl}5NTPAd%
zTpw!1wQJHz&Gqr^mC>DWmq+R5umf2e)Z_GxeP$R{&_|}VR9(&uG^?8)A#fwH%kB5D
zXWDZz>W0?G({=6%Riqdet<hmWIem|suq#2M^=4o}4#!*kiMlrGiZJOMTy-RyGglM3
zdnBwu`-9i<!`)`>t#1!mNp(y(MtUT!8Ea~^PI|T0d`E4{HoY_nxwCz%#@ye^=9xI7
zBwr@l;I1nqn-%CJ_SYP4Ax~f`dv>vB7Y3}~<Q=%?`AQ+9Bi}H;dF0HljKACmS83W`
zU}@qKtB&4}uL*m4rJtSfezKIXFmfXImiw#C@?{#{m346`*zvYpN_MikU93@#L5P8>
zSGt94NO0#!Cu_dDfKGoAaJgOM`Ou-^wK`sv`S|Qi6{iu7G+U4*_tEdy>R6ldPlk=N
zz9i+%8eYY_WtlmdD<4Ff(A9N@HSNUhOW$^E5ziFf?Nh9}v)}Z5&6J_?JGGE3(b0iv
zwE^~XTb2b?J~8Sw$58W1e1+`<N&dbGIonNII(vbT)zDL!xrR+#;#9sqZ+5!u`ZM2(
zu+1I&Q~UNzw!q^pCHt=KtIF_*R6%F<&CExfZQC!wk^ScdEV*cEsvuu|vo%b6l%BwN
z(yM;(7wk4>d9`{&l+0_weA!jYStI@CvsEFicPFb}#@6^pC$f~yrX4RdyFO$n$jF=b
z#D|E?co!;^w{MH_3uMOl*nAudKB;CM63Z^h9h#GGQ+((;zwcu8<WigJ#o`4s8ezt0
z@z3{1{w*rB_CnRVn$jM034JoVQQ9GwR-t?3`kJYu-JdT_M^>G1j7s$F?`r>iOefN#
z*)XUlRDbDAw^5Q>VR^>icCo0;+WKUO6TR1O0tG>Yi0$KbJsrb1wC8pRCSVEDJc3*S
z-lwKB^i1&a>=PV%D+Iwb3{9LzfZFozjmC<Ee_L^AU&Nb)C6RdqsBIIcCU@Un^uVFF
zAl?`ZMubOz+L<&M`b_BRQ&A`YQR_mwSRR46cL@2mQVAF8n~V)`Xm2Dq8B12;kuORJ
z3OY1pZJzomObG4sGw*$r6OB$<#pXCD5K%KA29ZZV@Pm!zh2t`mIUL#@2`-Kmr}790
z9@iw?5<w2M7ea4Cym1(uB#!`fs!ML+>`<vc4()-cg)l-QJOb3|3GNZ`g(hopXg4G{
z5lfWf5um>M=D>WB-Jiuov=^e56p_U92vAonHG358X+xU215t}%#mGDY)T4JQ8?!O*
zpW>jMKPzmk+Qe@Eh8>4x&_Q24U}3B<_pI~2@PM71Xw3Jyqp`esgL?6>f|bGskf8E~
zH}(_nRSi*syxxJA0Voh(E+Rl^Jj_sGtOQ0(T7jr|cbz$(ucbz7j6v^}P(;ifIDh~E
z5)zV&qVMS_TPZt$bO4+%*dJhohv@`52_%JZ5_IWh9F;D&j8mm+EaSB3I)HPmQy;vA
zcJ@FaWguVyMS%oS=r7p(DtGPwyCK?bc=l^`0Bmj>Nt#0`tQBZHOcAA!LE;EUr6>Z<
zeJXg91um!z@m0sGO9L^0JJJF!*1xwv#Y)8;7;SH~&d*N6ao=6;>+9s;!%#_E;1dTv
zC6U5?L<5nU<hA83K@hMx=pSqPGLg=rgdJD(hL=D|h|%(X{5ZC-D2pb!y62w-1!0OY
z#RNPc2vhM?(Pf+rRdyMtMb!q}{z0LD8ck+9U*cdpkzR0D^LsCptdyLAZw%)iJ~Teo
z67gLQrQ_<w>_bUY0~kCkrX|J%+k}S+S_CFIcX*hnMP!0K#lzxS;!N-;9wup#m;jM@
zn5;!+g5iKnki@u00iXmA>*91}Kwsk~Y~@UvPRk%|#}F}60JaJ11%7*%ykv}kWBEoR
z;83IJfTEH%o|*TIEY#;~NQ%<d=}k`W&+&=D9YWIK9-;5)&=u&)0NJ2Q#E5%}KnZwQ
z(Mr(?y!)S!6=@Tsq!J`&F0eo(phjVNCkpu9AeyEh!}(MA(#ab=7VY^qD+vig7om%R
z1PU4~Y89}LmHDji?CKg%5TwC;;(K?`#`_kxA4P)TXgCIVvTWVCsM}V9T%;_YcKH%Z
ze@yTEBwvt<f2!(p*Nu^pAY5xu4gdqv3hrG6v61z#lK*yQYEPV@y%d?RMM83s4;G<L
zHmXM?N^AupmZiv2L1&tJHJ!cQbRIxnqAAkUK+Rx{&?sI402*I54sbmt8AnJ@+}-(-
z@6B+t3^7*nLd45U?O)$9WRm%~gwGlr-2c0%;8y`9@&oohbjb`vMyOu;dfh-?b5jqt
zPnS%v_{1N*-0I6#7rf)kYe784Ecni^<9y=9)~wH%MobT%7zz&v_EkX!0-FErha3$4
zn}!qa&rwAzQSeEpLl@ISM-9+^{*WISCItB}`*$xK9Ez~)y{-a+D=<RBA~?J#Nlctb
zaM<MFyn$@LPMeI~AVHOsl96SZTUuGKv01Ap&9G26)lvSLlp?R7sG_Q-p+%#w)YUgw
d%``DnGL~CqXeUk4G;-Xy3%!1`trN-x`X7Z83ON7(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_max_pool3d_bwd.cpp.AF800737CBB45892.idx b/.cache/clangd/index/test_max_pool3d_bwd.cpp.AF800737CBB45892.idx
new file mode 100755
index 0000000000000000000000000000000000000000..791cfa96998fe321a40290b2936a7908bd916827
GIT binary patch
literal 3476
zcmaJ?2{_bi7yldkEO*9Ohhgl@$jlJMWT&x?okG&z7BUekDV3DE?xoO*Pj#!WZniek
z?WSBsRAjj*6<MN>8cRtb_dioF?mRu;JkOkG&N=UU&hMPxdCzz|v)RAmAjm%4Ihh-g
z1dBlsB!hetlM<q9MIk5)4ME9;75-eGZZpEv{M%TpWPQcs0Z~G0WT1L&%j<KI+)^$l
zyg{!)`&LNLv3CTgs!atWsnWNbV|qf|P)DGy-<~SyoOx8dC7{HL45RXjoBO^@^-Vmg
zs47%D{SM-VjJJJmN`E^$a8tE+G)Xd<8@g?Vkk;Msaj^5S+pf9Ez{fn^fBQnMR4B{U
z?_Yej?(yAq7H+)varYnKwBn?#f`dQq3kZCc-`aT2p}eTIj@}?`a;bCuXOo*-!z?57
z@9!(rn446N?w}d^L?pa)VcE>bb4ra5Zl0qoGig$A$?MLqTIJj;W2)GfbGkDe<r99M
zW)sGK?$}{I($}MxFycw|VMof7JsyX_!C79;B>8)J^ph(TjNcrUzQ=o*9Kpd4@m{4S
z<BhJlM@yz<Xy>1qunouE-gP$il#cZ7-|g0?oks;(p3G@CYc#sD*1N0)-e74h#eL)l
zx!<jP{ft~{R(B>Q6h~QCAUQJ2zW;2?ii+ohH3w7TNs9WT;sG%QXS2t$pS0fCTbI9;
z8-%|TZ_v_6XdkZf2^xz_P$+szJxbdc-L&kFYgwtvB^`}B-%H75-q_zIb|CO-Uv^<4
zenp9ja(pc{u{6|X3{}Y~!Cz`!Tb6X#RDbaJlp1<KLm6@`dLbL+oty4W&gho18^I1y
z&%UYN5_R?O`uaQVYdXd^+<Qh9!;Hr#>Ce107!G>58FyVJ0Um9=NBqyvKTnZj!__eR
zN(|8nPg9mfYeiYslO5HvMt`|W9W7Q&`yYGK_hWBvyoiTJnQuwwk<ew7#G<b;4FeXD
z{#*SsVw0}~>+vQwXqx%4VlPL9<~ePuEX`BpNlnsHLS?u3P-YGhdk6L3T;KFHYfibz
zY^0vE->rLB&2O9+N8sw-XM<O6T=r9<=$x6&-x;E}K{m?nL3S7IQ%hy`Y3ISZ=#gNH
zl^ZWo%X6u1{cSBRhng+7Cz5YZJEDpQt~{b9`=ka+6j_haJzO`ra&qPdB~CZLfHPLf
zW>tK-`0=#qDwp;uV$P}@&7elFS0>JlVfn0r&as){@T8-WjpvWWOjz9dqeh&wqh_|C
z@kP-dW7f5X@JL;yF%9>Dfihh4W6Z8Pt(z%tthUQoCn@$`_$zf&uH;l=dhw2-!3653
zy_1ucP7XE>mFIG&@0feE%@<ePxiYWncQB=Uj=ni9t?Tzn&p(6DWzHm@spa7xmK%|&
zUh((FTLan-y!hK?*0qOxEE6x*v&YCHThV=tRQsG~@bzIqP{~^@QeV08yg^NR(FaCe
z+)n?;6Hwv~ZfKqe1mVLE{JGoVzB~rCCK!V8XuO<2Ag^8UA-oRVt5#+7NfNalf?#{d
z9xV_Me<dIv<&G2|l_y^&B~j}Tx0HmGf<Sz8LJ$Q@ur<wRd16q0NN_9~D=!eht!*&;
zRIGaHAqEwI+@?#=Wdvfuy(cn1ne)6Dg9=1$6VL>CfdKB5mP14ux5*X*g(J7^AUlFU
z0BFYzIHX&X7w=+F{>W_%8lxZ(;P%Z^y#Di5>{|p>5aO1Tmct4Ja6cd>KP{YpQiwsV
zMcg<vPEjC$`}z3(fa1u}Z8#|GM{Vg@XQ65{cVid>Mf?cb;OgR?MV)x=PpFMZaH1Mf
zTOdGiC?O&;(MS@4aIYLXc2Bp&iC|$ZN3Hb$GN25F7ElE^z``UJ$pIV!#Gz2%2>Tvf
zOrXW9WH|d#Q?kKTBod!H000ZyW9(Hx1VExFT0kDSgN5-*crAbh3#+nJSA*ml1wBq<
z@3l-t!~~Q8rU2haIzHKp>Fl)Z=nIFwvyL3%Q@)r^nWemtKv%r-D0!2rFVY!9J3~(p
z$$*|Z4X4XJM9f42kqr90l2(e7>20(Wf@7`QzH@p1**L_^A8f$s>a2<Jn6lZmC@f58
z(w76!h-OMR1Db22&d_C)$66#4fARnTpTYZ2cKE)X`ko2Qfv-DJZ#c(7oR3SIhzGd*
zCt^LVFws@WX??7|BIsVAe1DT)ftP~FH|hWn!4EQ{WNYkCiy}x^bW4B)3o|Sj){7dG
z#a!^MZ?Y{Z%ER#CEKLBrIfu|S^Jaw1kss}*PrZ!5BPg{UwfzAcFf1q$L7M=ae}hZ<
zpkLUm%JCDzymd%2qE2sRmr$l{pKaUdJ$mvINGaBIhehyoHr;hmW7sm-OEf12XHjEX
zGnd?BvY89_z#t)$zm%GPle{yr0vngP#~J~)l>5E$=GFRZ`cEJ&&~?;Z16q)lkajlH
zL!%Zk>pAMJ1?B@f*Su1*C*B~8U?><m*f>}iOTp5?v%$hR3XTr;$w3F%yfCj%7m!Fg
zjyemGevK!Vt*mh9LS~M^WSA_L$ymW)0WI78$aKK=a$})p=BNjL@N|2BOvs$;eX)FM
z_+}UqlF4K)8Dpj?b771DWaMr0H{{}DBPbEHwZFeaEKFvR*`VlCo5l0Sh}k!g3ivYx
zCcj{4g{nLBz5y{C>KU2>^Cbf7cseFbNeEo=NcxU|*M-^qS*Fn$phBd28e@sZq%jw?
z(kE68YUpYWp-5MwUh#erqlF8ZE3O~Sgs*k&L*PgmNHIWRpa>DPVA}(875i9nnavP?
zOqWi>op}r6QKS5Eq1bopYad-?kIEvk`KLjtfAwcQ=r=gLpWgwjp)xQ7=8=8I{~G}*
zv9JnD#S!3X$T&jll%8%tGVx~??2YDrJ&&NC?rlg&I)$zc+Vs?!?CnKt&PL3dj+*Wu
z2v}g>-yi@0g8kq3&}V)kjrh>tU#Qgn6Ta4)&D4X08Uq#>QWO8*Yq2H;k^%+58FYC*
zY<{x(hET|l*H_0G+8ruIz?0NS=HT?_Am2w-CZ+F$GR?^3tctz5n<r$RZ7-aYXp-m?
zG9&+-v9KmmI6wq}Z3X6G!IPK}mdN0@C=~Pwf;KGv3|aUo0!ylI(bN&Ox*!^|AruPZ
zhR4PsPA{&HsF=8fq?9xUizEIdL%=(GI(u+ro!C@aG)Imkub`}AVQFn+YiI9Zq)4~Y
yVHs%r$f~5OM$y#L*43viGh`SqH(6=EO5KdiG<8%|(p%x`9*J`CUF{~~4gCY>Kum-H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_max_pool3d_fwd.cpp.BE8A2BFD19420616.idx b/.cache/clangd/index/test_max_pool3d_fwd.cpp.BE8A2BFD19420616.idx
new file mode 100755
index 0000000000000000000000000000000000000000..abc44f4fb7ab6fc7d188046377edab21fa0cf200
GIT binary patch
literal 3690
zcmaJ?2{@GN7yoALGvAmo6tZVu$G(oO;k(AZ%(Yjam}E;Ciu$KqN$G|NEtHB%NDD>X
z6e_e(@rY81vR@^I?yda4neLnN^qA*y%<nztyl4BpXS|sRvPKkw7-7uB7<POJKLkNy
z*dHf8E~1eSf_4xfD6z1{H-_`f6z-ZEX>iWhiQ){A_k>=lH{C<y&=eA~g_(4p62v>X
zr;4W0x@k>#)L6o5**@GPC3Ytj=K!g-=h@@dU5fL6Jw^_fYIr`(F-)$l`(o(6UVX?<
z=;ptE|AcMYfEr$#{Tr&iN{3eYxGD_d5^H5%CVy#Zd}7<&IzF;7r}zG;CXQgv>yG&R
z`}Mu|)w4WiS|3;|$4cAmLyAI{6`VRz8?W)Ol^uL!Ne{ot*?qZsZLYt|q?ertv-i#Y
zCn?K3ri6N-^H3yzQQ!8fr*q02uYDrJH70eeuo;OgNiq7#{uT78$&BjoJ=NJ38m`Q`
z;=rir*=s7-CQCaVJ-c4KzeOLq<11BIk}oe?R>+P~WV=O@|Cd>vi)`L7J@Gt!hO#`W
zNR)DhSv+kx=7vn=(F#fBDmN=j+MOu}LQig4W?xeZTOSc%6LhysbFM!!4i@P9pgpiT
zzIssVdc|(*!Ygir`x1i0!<&0p?>&Wv%D>)yx8t){cABB)Sx8~S(yO>mBvz}#eE83;
zdvgvKueewAG<AEf$$1BDua-Y6rOVgtRr^pDURlEF6>Cw|#2pu*=7lx$Sq)9>d)e7J
zaM))sFl|SlK=i1mLUx*I@Q7%+*Y4}%vaeehoqYMSJ*q~`279Jx(|S_<g~L-4o*Q49
zS3ew`nra($4o%YU8>Rl0a^5`fz*}pr;h^X1jmv{J9)IB6M5q+qvq?HSMwIhMbt~uM
zLvx*VQR`%zWfA$B;fx!)<K^N<&jwDi68bVW;9L}rvd;J&Tce>6mH9Q&DV)9*xvaCz
zl-_;0=1S7}zDgaPE_QLZ+T-DEV-Dp(jTNw&fn#D>`$`9~cLOcq9R_o|l0QZjera;q
z$vPC>a4M!Zd(Gv(o;T+YRFo(jbyR%5YwK!9borTWt&#@XGmnh(vqU1&Hx5MmoB0(O
z*s89)HB^v#30JjsVo*PRGE2DYQN$6WR|TW}jZ>dfE(HWj`DhSEY3jX|bsf8dGM-9(
z?D%Tc?(3`-tZsje!mOvJR-8L|u-W*sT6y10!|gsD_tH$4fban0tIYy#vnh3EMRnG5
z4wm|lb~7UA&yP1J9lTqXtV9TWTiqU0JT?8w)-*ZY*2MC#(HHA;;kuu~R5N#sjVWu0
z*ik!uGaADOEi>n;GB@3s*z9?9%d=NQ8^|9T_u|!a6grDvcNHb}%!Kzpp$L_c%yrC8
zKFfY)sXF)A$R+J{(M@gJ5psHI{=31(yK+`vTMh|T3@b<UlMUu(ZeEg0P(K-~+t$*=
zNs3t$kAomdua857#EB^)J|GxEVFJ8_C$P89`^dV^k)~<u1t|DHOidM`!aM=gD%!&j
z`Rk`!h<JYpLLCJi#drd!O@dnGplH31M7$s7ElLoT<O!hmDkE_?RF^g)-WO972}B8=
z0P28oeUY#@+EXEX5O%q+pzu<j0P3Rtxjn;q^(RE}YcX$%JVk{kfckVgI_AHn?WPcZ
zCFbo4xe|B+;;u66Q-M1p6Kt)znTY=#^Cl5U(me4^2|?W}JMQ%&Ucp2t4D)t@TqryN
zyq|S9B@H(a?h&ETU#W+XnLCs8@<<Z2=2x-f*lRhO!Vtt~7g<VenB0zsQF&*1H_&Ss
zHS#p_0f_X%p(}azQ7mSp%1{ji$1rN?WVsq3^3}gqyiD^?#Nl~Gg4)4w`3StBf`-DV
zw5hZ`@P$!5y54U9f&VxHZ_vO+>Bvi`cN5AcIDTFqK>c9;c*0L}hfx=07d>F4t+#D9
zBvqmX2cr}UMHb8rkONDO%M1_+MrBN87$6N8)u-z-0W#C7B#HIiwO){gty<?+qk-eG
z*fDp`p!<&d-cLrJ`|OW(4pBsuL9Q55MKl0%xvM&K7VqJYCC8mIm_MK<mK>Lt^T|n>
zO4)&QU{sf`ivZ;F=GKEG*2wgK$pHqj+^Oor`{@GQ>P0AIP`&T_XipnrL@k~l)q=Go
z!6p2tEo}Q8fv{M6x%3Cg!Kf?IRR#=s<?W1h;th;r+%jp1At($(9*8$UtJ>@<j^+}l
z7Y!HU%3Q6tJon#4jo7<1(Ng0>-FVzjRQs8DYd@}-J@lMS#&S|)sCk3xVARUV$`7D~
zAtCl1b+hh8!)@i%z_A(IZ5PaTTkkPFBYTqO0>bauU{SB(H&B4Z)nI6Re?u^8?PTo_
z;-<u<9MAAjYQl^(8Ja9`45KzqHUR*osb2Kjmic~=A4Z8vM18;q7=@K!eekYel%zz`
z2de}|MU_PL0j+m6pFMo~*o`hMij=&R3Ah*IjEb%mK<N(RxuIS&rI?XAL;Z(T=uY(S
zsbsom&jxI+GF-Gz`xOH}c)CvHEozkGeY$FT_(~{N6rzRb0&-x;4q^O&91*7<kTc>2
z(4qALx$mf%)mXk<*3J`>%UXZ{#WB`$w=y6xj7msLD1gcVd-30h1AM=|V8F366wCZn
zVyM)jMU{`GCI!4I%*1j+mLUotRSa1owg7<{Vegr{>*sS?X1dG~<OHKMI?V;3$_F;B
z@`QS&MT1{h_u}ndeyy-YjcP6($OyUE{Rg%JgtdifAb2g3mw3)@sCxmmug}p9#|;io
zE_&sS=B9h6k#TYCO76-7?Q0y^cY8oPYbSQ>Amp$F9JlVu)_Ni`+qz&iwGXI!1igH=
z3A0AXhy>`>L#C{^7qy++P!)!XGw275S~yvF0Q7`-ThnJxP9R1lSzp;8a15hzzH&bx
zgJ6S@d7WXf;dcb~BSwH97w^w29Q=eVYSeF0kV0+o^<A`K0|hMG_2!B}3!44L`RTC5
z!TRb&eZddPf6E4RbtI}Yg>`Th9qW_4rM%O8nMQU-?f`uXV%@7XuKasJ+<(RPTSZ?;
ze!JlKb7$cf!8XDE1;<bvJ33@J78g(l0?!9f>$`>%7s{bwmN>Mk_uON4p;$NHs?{OT
zYLGMph0K3lHUBj<CM-G@gI+N{eEb4}Lc$_Mm_!v7qmY@NOb<74Cqz@6;3grxWa%;)
zy0xvH{R)PIfux>|s-?EluTG`p<Q0}HtEg#c>FCl7jf~AKtQ5^<^-X@0lu|c!acAS5
KS&ptaZ|MKXV^XUC

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_permute_scale.cpp.0607B34D8846C932.idx b/.cache/clangd/index/test_permute_scale.cpp.0607B34D8846C932.idx
new file mode 100755
index 0000000000000000000000000000000000000000..efaa9296eaed4d18bd03183f36c04a935ab7afe4
GIT binary patch
literal 5080
zcmcgvc|6ox8~^!XhMBR=bY((zV>glF4(SrIlcgjfewDIZMn&9KQCe=+%9Spq^|e$;
zD=C#FZMsRhLQ*$di>2kg-ZR6aq5k#ye4L!~{LXWp^L@VG=Q%C1x3#^`z%c7z`<O5y
zI#2||Fd6jAjgAU=B#dE}R1AyBF7*!MHk(Qhj=a&66mM&;@85dmzEK!m(~hNNvHMDy
z$(>Z!RQq!O){HPK<v@eZ-fNY;=M9@T?8H>DG@>?uvf5@=?fEW4mfZD$$3rh)g~xyD
z(7c%aF)6OHOL3FYGmnGy4K71l+5)7S*3G*52l++s=U4B$7He+$H#V()W-qHIW{%<j
zb75-4BOjwLvvOQ}mBQ2B>8tgLU-R=T-4?R!U}}3+V6sbLd#!njv)j?7oHr{1YimuH
zX*pyDo1ALBuycB;XNX>@+d2okCZ&i|FRSwWMyOI5CWkzOFBtURQ;k_ry(1=gAhl#w
za<q=ITe)&4zP8|OTb#}M9_pzNmF97Ej1LEyzAM<NOE>;{%;<ROR#EPvxB!cXE;CxH
z+8>=TJfgDTFS#BSH}~Xb#|OLiykhw?<yUHp4yj+&QqHHXeSFdW)H(J0=S?+o@8_-J
zJP~)k>xH>g%>K7p(So~YdFD;ohG$D-f?8~EK8>7{|Gew^r*m;ediyssf(z?$uPsCI
zuP%`qZHpW|IIDN~?n=%2b3xtV=53qPOp2{^XJu^5-{w;itN6JwRzIKnT;`7YscBXW
zb)A~TxTuB}`HI4mH%v~{mr{hT>o$4j&Qs=wl|R(?4$KM<=~A|ynI~EN@=jq>nv<DU
zO<>dB$|v639PRYyb0zglF11GF7qT<S!imEDq^h>nZkvXje`$TYzTi;He#fw!+2*;A
zRUXXQzPCQmw)o`vJ$j9j+-9*gw`V1#*c~pbZk@K??uLhtcFt+pc-3W9wMiM2fqxd!
zI)fS>_?v`W4_X=Akzbccem3X5rj+KDjEmK$#SG@`YTWqjQc0VBu2z*}w`%A=LGtb*
zR}`LxZoF|M*T03i>PzviVmDhopORhjvRawumlZNo^FJO<6FcV1?W+@W63H+eRvE1Q
zZFVV9>e=kRj3eK0_gCo;I&aK-r8crG<_$f+>;Af3NygmnwqHZS+{e`yyFLl8vdNZD
z)1`E8?(OfaSL`Z}bWhP=>$;0^a#QOA-{}(R-SaKnz1-`$^1f;grhSJ8T-tXuMylwl
z<h?W5n4QGD9G_bl9%;|`+iEDsGQ4f;s*ReIinomA<$cZxA65=1tx>%DG~lm)Dvlc(
zEa%$)`Q^#oC-PrEVw=`uN(z<tq+Z;Rl(Gz*9{F{%v!iaPWnA^yqMnh)@U8or!dz5i
z{o_QXs#5xXi+`APGj&7O8tnt|a>KnWa=MFUUi6?${&l<b2FtwJ@vhq0FBAF<GQ)4b
zP`bHc$-~7DZ*HDBtRz+Bul{MCC2L-kqmKrzXj6fS6frm$H#TfV4US>t&x0E~4^cBD
zun-I*95F`<f53P(0R6~@>Q6iDV6P&9twClF%!9%oz${-awU8xSLm}h7$V??sW%$GB
zUJO&$eSKf#+Ta~3?t{#15?h8pfSDCw)}MHTnn}gImSKbxMT)^6z|0XIK3?w`nL@@_
zpxwThFNHtA**1IMzgtuIbqN(;jLakwNt{2xZr=pXOKa1fA~Noe%+{DSg+IXVK;?)g
zthmKO0$YX59E?NZ4`5EJHCTKudCPP%zWj&g`|UPsd;6m$u(iniU1+I`P81Ekz7Ryk
zmmo8lL>A`{5L((JPKL_0f_y3#hRjBoku-mRvky#*-c&}~epVV^ieOO_SEKO<*nQNf
zKCi-*xtWXypxxhP?DBD~hmPCLtf*M%kDOPx<OUVoa?qpVp2)c*MUufEz<F)rk)PcP
zI*g?8Uy)gcA)~+_z+8XzuM<0zoWvxs)yVmGpf+4j`S<gPS1%d&LuO;ln8F|6Vo#5l
z$G`oYg)m?FBlD}>N-yl3&UKUVKxF=oYu(jmcI%u2w#>jI#=;@QC}O-mHTn}ptmJm0
zP6dn>!bh1rW~LldNj$N$vJMyGF9V9BAHj^S{>dyjS!F)@h(`%v0X|^{nm`Bef`s$O
zwLqDm0UBXUHI@ZrXatAKkp;gr!kTI=3rwXEj#Nik2!%#?P(5S;6B^-5^_2xhD4b;-
zt8l<*V-dhdDu1*t?C(`G5fb6^377yX#3>WhM<a}88fn8vnxbXt3TqWTg=hp*j42D3
z0*Ru%Jbl6)X@s_`_7WiKT%7_5nHRMdlg@=`aeP#_V?IUrvv)_&3_Y=&|He|YG0wOn
zH-aR>rwUL+gpW9gRD@48AfE_7fskeqei9%jqt27Y<OwB0m@=U>&<Hl0Z49uWnF-qz
zW{$NQU8ix^B?t$eFaZaT&YO2*J>LwDT@e)Q=|X3{PB9l1MiQl~$pIiVLY=NIFw>xG
zjL*E^hq}$9hLj^ux-{MCuw}7{xsY?Puo+4zPgcMQ_!nWwTZV8(0V$sGXh-^GPT=x}
zU;dMmM&$)3=+b}ggXc+QkM5wWXlrX*!$|?djjPH|Rpw114`7X!dDMrK0-fyr4OG*<
zO!7@!UXqF9`6^>y1EC`e-&}g_U+aUGc|{Qzu&#?t++xjbn&c#oTS|*a-fTG~x!<92
z&i%K2?_c3GLQ-B*2kIV;kdl|u0a`=c2wuTUkFpPn0kCQ8Qh-cSzm0;~y*)D!yd-Io
zEO2eNgK$F6?<1xZ+Jy3i2%@N_4mm&sQB)V0X((z4%uMJe%rjdq4R6!Qe@7!M=oU-}
zg+|!WZJ3aE8sSWLW<p`35ngmJ=8(%_e551}5z>$D#{`w_*-r9LQ;W<(R}f^^K9~K&
zJ{yXRCgmkXdFe+tr{0W7%h1)3ns_Y@a6tO8W!r;((9DtT0yEGF8K0o0sn9w=D@5UU
zOQ=+b!bQ{i>HA7hmU!BLpby>37wZX)nuY-3wJ9jiP<#;|c#Rdl9<4Y4^sy}mwl0$(
zLeO5pCXs*P=uY*{kc()Uw-;QUl>Rw$J2gmd5}1%*)M*3*m>ABPx7%ZV+Oqc{4zqRG
zhJY%X&1c)dtl;jvnwgYqDw7TXlAr?wQ$CRZXIKB-w_8Lfc+%d|^XbWf=bE-5(5RYJ
z9l)~OHY28Repl0!=oK2#|A8u3N6d}7XJx&cbnA-SH|kE<b~8w*Wg|{N@=?ncsE4R!
z3)ErMvIS^@TDHJ@0JUsX0>MjF$KL7=-QN=sqzo~}cz=bmg!(J*Ee1%VioHPF87O_r
zVOw~n;pqg#EWxT;lek~!iTH30x(2URf!5#&hg|aBZ{J+9Xsn;WCG=9{Qt&>Y1H^wp
z2uAe^l)cx%a`kO}wjN!Dt;ALb9zyFzE(Q0tn;qCnzvJya3G9voo|fUJ%Dz*=rA~+D
z*JcN=l}tJl-bv_of{7Y^K|tYzXQq(&g!&DY6v;_22NN?gp!Ya7Ls3U-3<5wCT%*0b
zbH~G&F#F`|Syw*0pl3<|`683T2|c^1Z)%)dtuGijT$F)sEm-)zHfH~q)Ie&Vk~6U=
zA~bLV5*N@LgSW=`0z93@7ogdXFF=VNUx0*-FF>lGAbvZJ8x_P=L+1+-XQFe2%Kkj~
zHy#v>1%cc#Ecn}>L1Dq6y#KSfhj|E#h?2-+;#8UhT~bC`%HGZ1)tRx-R+~X_p2qx1
zc80vUh2{KTtgIKznr<*p<7Zu!AFawM&QwxWQ`gke(bG3HGM+uxjH7I-z&5d)F2|bV
P=n{@Qc-uG$EyDf>R!HAA

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_pool_fwd_common.hpp.CD53AEA46B3FF1EB.idx b/.cache/clangd/index/test_pool_fwd_common.hpp.CD53AEA46B3FF1EB.idx
new file mode 100755
index 0000000000000000000000000000000000000000..1212340dedd8626ed85049bfa0ddd06ad68a6532
GIT binary patch
literal 1562
zcmY*ZeQXnD7{9)|U3c$&_1bRTb=UR7;9%+?7-0<~IyTBiB4dRFm=QKhQ4EBoozOyD
zhKvv<GT0K-tOz49{2|%mjM+azV#)|1nUfe1XQoq<WkIHb;)VnDx!WFSFX?lA@A>@w
zp5I-#HWadR3{$neHnBU}(!??hqrxxV65BZj_W=vTBvPkORMw@Jh=*s7e2`nC?`PY`
zmrh-|e$Bo7(}^qfsqoD7_UEdaPdiImw{6{euBCJFpfc7ly>IN5PaggG>+FVo>+{$B
zJl@gs@6k)YhYD_6e7CD@!^zhlTDG`KC%d{AM-Ek<zc6upRpMM%&3O4|iI?iW_iY^5
zABnb|b+><IFPD}c7+p4)boF1HvM)=G-W*ExOAX%cACwE8k5=fe&f(zBhVEeUpUa~+
zDmFRYJ;TYgcYXO~L0#J$x^rr8$o1E=eFIO=E;v=x{7%nZrT_5C7amF9R`g0tJh}Td
z17V(g@?LtRwXcA1R)#99Vlxes2MoX^{aAgYb>{jWnb^QxtQ9XY4Z5ExdwZtokn<`>
zfJiD?C5LGkelC0Ud-v|X+4pd>KnS=wj^j;(?pKF8-f(<bbw(h-J>@x`Hx0UX++Q{G
z;aF6Wi3+{Fns=dTu=m~vmEV;8RFW@{T*wwUK`;$ww>1YZE!=eLeVK@m?brN^OoQ16
zV+VUX>ilV$NZ<}=fl||;`$)B;J5oRWu}l<j`!rvvY0!PF+1|JO^Xi)d0Sc*=vr^Mw
z@2>Fq{=X*&BLc|*cOI9=n+Dxq*XrSV`^a&b$WYv)d3>h9;&+CEYu<XfvCU0{IT{%X
zmx|f3=#KaTsGAczup*eAc+nzKUgDidCN9)M%|+sdaFIl$1+>7AQPf-~2H~C1xi!wW
z7qcqiQlWvd(fJr=<2rlE<jIO$5u$Y^fDJ@y(`^|iRaY}kcHN$Fa_EkXlT&wQoIJco
zLEc1K%oi(&NTgogtDsm!>f?P1u1}<X-mjn(L>k}&3T_?=1006*5XZJeYER?EFKz*?
zR@GXFo5uRUpNqOQghx$DB5x>3C~4>sOJaY>)u`rvl540Eb8##Pa#MHFlORKz7=*YW
zsE`(B1Q`l_9Q4O(JvegA4G1`ubK$<;U7OQB)%jmJJT~PtKOF*unGGG|%qU@aZ0Hvs
zUn@%d!YAZhj-~lDU&SfeEK#vp6tP*OvRM?cSy*wufG)E*HQx?I_nk1bXv5L227n=y
v3C|<t_|jPTEObmPx~s_;J|2t2pECZ?nw8-PYks@hNC4GiT<{?@oO9Rz{9LzK

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_softmax_interface.cpp.9F0A118BAF3CB91E.idx b/.cache/clangd/index/test_softmax_interface.cpp.9F0A118BAF3CB91E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..ae6630a48e228f48ebf7cb38599549f2655445da
GIT binary patch
literal 13524
zcmeI23s4hR7suDQ5Cf4%2uVl?5eOl~AQ4DH1SA1XNk|YOf`}rpprW-@qG;6?u~vMG
z%0m&+ib%DB)VHnGDy^@g_(1FHtEg>(Dr&_Kv}*Y_*vnpM>3rib)5(m?u#?|CH@o+7
z?*E>>Inf%my4D{-!csMJrl-wJMiB(D$N!BpGt+PNN08mt2r{Q+e{6il%>cWro3B@-
z><TzD=_c~g>Ew)+=aPR?x>yuDYZiG$M?`1cns_XX`$d_c@?KWXH+Saywa!><yk@UN
zlP(#Y1PkIzSMPcH{QG{(8=l<VoI2l_^^r94yl2fim&eZwf;Rk@D%e|A<MMvyt^SEI
zx(pZ6`$om^?3WvDR=3|v&6*vkIX~(0#k>Xl=fgI6?EQRi?wOYV%*xEtpH5h|o5@_6
z!fxFDO~7oO#ItBddh8U3?Hk4gHmAI|>{5U%ZoT-zR0FSlNn?S*t+4oWZG+1nFXo99
z2WzXcb3WZCD6nc-e78-w&~tl}Yw77D%9iFu<4)7*8_P0P`>tMloV_%w#kgbdprCmX
z+}3<{LUjXw|I>!3@|jmHf8lFdavnqlrqAr(I%@OL`HSu1Tk6BMj5xlp=*So+ecXEr
z4sG;;1WldMZv7tZet%NL@{$jtt$$h;|I3-W{o0s|gOT!ytd7v!I(k`=#oFA~19yKM
z`@;@l;MwhYA3Bxz+!$PgJb$#XHh+VkV}r}&?SqqUG|fsDkymDxjgEGw*G8Ai8sz&d
z3v0CkzNfe6n(Mo)W=)N^Zgw7nMlaPWg$q-)Yq3$I4$r&ygU5kG+w+?%LzUGQ8-0)1
zp4M^ZHd_wfV7GCvG0%7ERz`Kxg&|8dR#w_*)`6I3XQ!TxQL1K@T?pNgTcf_8iR~EL
z@DuAuPIbW4{D-d|s&4PhD%U(%Wj#HX9b5Zr-jsE!3Zr=B`ndFE4ACjy2lRrn^<m%p
zc^990<kD8~$*HY|`Y{6!g&tT?ZtL#&<LCOJK|GC3M`e_8s-OF`LK$c3y9xL22Q@WR
z)`|zWDld9OH|v{6cqgO=*~;xluCK2UY#lHq;L`P(y3L;^wy-@rBugc6i+#ecd^gZ|
zl3D*1y>RU(Ps=_Yuq`2`W%&RdT3*-emQp^f`0UyZd0F(h^NVL0Go3FBo5e0Jwl<dg
z9_}bEi~8=E+niNxH%ln`oy+}3y&|7~65uC@VSRf$^4XqW+T5$WN;DrHA71+4fYF;@
znHOc;qTk`xKe2GM?iD+s#`>A@egA+ZjnStLixsPfR9z=|`lmk(&K6y<{fsR5GUKsV
zrJH(_Q*hpI)>n&<Qc6pYQaoeMtt=0|zdLI3@Y(J4RhsITTb_@4{wT`e?2~pU>(fW^
z9r>?H)p=QeR9$R8cl%y;eU<3@>T4s~e+t=>T{oB9)~1;4ep@DZ__%+A|EiAN@ze6p
zH4eJneqcgH<H5^ui91zCzZ<`LvdiR>%gln&Ho;We^rV%$i&BfE9rqrW{E>3$;;L^a
zuPX1bSmm*pXJhZZuqkBY(BOy052mN=Bq0c8!(k08q<IgSl!PEyFcNIpc|lBCfd5%P
zKXK8+?Xi(e8jG9V5I4)t3otXr%f(~=pjVJd<8ZS*VsF`b0cNgb=9i7l)4dp^#EA$N
zg+y6)UVzzk%8S`2?))B2CdK04rdU$gotM{7MiBShnx7*ij^zx}*!aJBw&%TNsqVv)
zq8Oy{6A(;-NGv-q;O0uL_1T$cW+T0Pdu;gqrR%oJb}>jv$$!)D@j<NLTiT1Z^wRPL
zo}<cyb{`xilM?Zk|D|IKzuG$2>9==I^z!W`Z3AW8eo3o(X?gX!8Efya@Ygd);}ZU6
zP!GR|Ipvq6XZK>RU$Q;T`_}l3UOIM<8oI(Nepz-e<|bBaBDHDvxn6Gm>->8&HhJZa
z?4{*T_fQ<u*Xh<-lg7j$*g&g+PMsGpp)Z!)yiNb`N-LQ(9v`*8%-G9^X}{gs_TD$W
z%vfgHBx9=>K?Z1d+df~s{6Rlb=PkfZDFIlkQ{u)s7HFpno}Y9ud6MU~3bfN-eU8=S
zIF=lrh<1LZ{dv=xhYkm~qMfdO$isGeRpX5pXy-fzU3s~w_2>J6U|o<%uh+w$eR%y9
zU<9iJuVh*ppr5aw35@pHfh*1zD+33j7z1V4-L+U-R!TithGI;V2@wlrLBvMc5OGir
zL|l{$5f9}-Bt(UFVFSnwUmrfHK`{|3f=G;tA(Em}JI{A+o{1FH2vJOq${|vqiY@|6
z9PguPeG`2&5il}uzbq;mvF;W=CngC2;{(<|UTj(;VAZ%ZQ^FWCE*LDRm<`oN2)u@h
zg}8<U&r?^V6dtP)QL&M(BO%hbY9NYnjR}6>$eiiFGI^29(;U|q8;ZGRI>0=e9L3TB
zwRZ+mF~%513h)mq#xyXK336gMxf8^WVJ8vf%y5QmE({llcm^JH7cYjFJMa%GHkdIO
zB5#H_L^6iV9ry<o3t$95G>kE<i-3q*T6Gg|DrQf%=K%kpVs2zN4p<RXOhT4$z`~_s
z!Q@~La2YBVMULVCPjPJ?U%xkLgw9N-7VKcS|C{sy^q;?Ljg;}!$<|DJ1-ydPubTjO
zmt)TvA0M@QG}#O9IxG}m)5eQU`VV}-JJ+XXlNHCxOh?4WJJ4}V$`NM@6;q;0JD`zN
zOogf-3PnR9Qln~!w5S%M2s8pB9jb#!kLn>ZpoT62yBt0-u)B44KHqzC<kM407G^rs
zGXFO39VfmzY^Kcz;RQZL@Q&Dhe^mO~SRblf_`pp22qNzus;AP;Jx(tXS>eTS*THV|
zk4_u;nRHxbk<BN46x@woaQqwIoi#_ybbt+!G|#3rK^9~x=0I_<CE@@&g$@y)!nXyZ
zOT|1W9uNsA0*HJmzP3QsshA(duZv=zyc;##I{gp@?Nl}poS~h{MvzI_2r?-fK_+D*
z$fRrpnUswnld=(HQZ^9cfwwi>ci{Lk;d15hA9kLIHPd+nFPS}7MB;hXuF`eZZ4C2#
zVfc4FCGq~LrjD9vHfHJ)$vVx%A#Uqlyq$YU3I1P|z)lx>H~5a)-a{qQ$lGWV{q-O4
zVmK6?7Cmvur6&^~nW?9|$LfbA)s-ylqb+g9oZ)rOXD_`HqnHENfe-c+6r*$L5HZ3T
z8p6g5XF|jZXF<dX=Rm{_=R(8_=Rw5h^7&w&LNO1n2Sfs{KuFkxGNGIxu}ln+FV~k3
z_AnIl<N65+l7>i?1Svuk5Gf_fp~R-6R4P>v#VN-Uq>`whnkuCVs;P=oK{Z1qp-|0G
zWhhiLG%ggXsg|grnrfvQs;Q1sLp8M$EmTvh)Iv42aayQmgd_s08KI1TYDUCGKs9v|
z9aK}N)Il|MaXP4`UZRI;>Xmw^ran#&)ig*9L&3I<Vg{u_1yP(~ETCD9<g}BX@@sFE
z?hTY~j_;sN1=VM+;gSZwMgw2PV`(@{>K-)BGw!EmEXxk*BjEr?HTe0UDXa(YzOPSH
zfe{>t_lR_Izsv|~_sY0}9|Xc@i}+~~bef8v7U3ra;J)~25q`i(5d6&CV~&o|7nket
zr@-&OOJu*FGo;`BwZ~wAQ}`am%eDO%wXE(4Fa_6i2M!QBDn_T%A>z~d93Vw_a837s
zNI(}r<V*ME0P(?tYr0<-t;kUne?Fsas+sOUE4CH>Cu4r|&HgCH4rfDxE|Q5LVU~x;
zyX1CNWZtv**}Hw;F8ViE^lJu<dm5#CSnzi6-eB;GPF_0ODhSSJ;HT^&hLl*=Oc4_z
zmWTxrTf~NlBjP~B6>%ZriFgnRMM5bd4kD2VBC$vekyIp=65=2i$stmR6c8!-N*|!_
z_`xDy1yLwJ6e2ZW4Uv|wg(!j_0g;ZcgGkTULuB9^e8674d0eo}a*V&9nF`fbb(oE~
zxWU0po4<_`f8c;i;RQ2oetAV({#>Rj)l8dT(G=a@ma^q5Gj0B7cEbukw#Hch@&Vqi
z+082|>U>lBDCn#_H~fr?)lt{CTj3jQg;mp*d~hqRWDy?t!sj8?-kbpw@-F4~(G-C!
zmL5;}V#<$?mZPH7f7u`RO4$?K8gAL+t%HAz*uOV_S@NE13;R(qiVwvP{I(5JME>!;
z>X5o%PL;CpS|5$y;qtm)z9_AA##;!#%z*5l9>e-|N09G>p~b(wT@Y`uAdXdRzHdMC
z*%ULC_>=w?ORD>SoYccpPQH`&tsO0PXdn6GS<alhlsPsQZ2VUV86y}3&^U!K)C4hQ
zOvuIzVL~>Rj0M?PAuPzoma!olJA@6{I5G}o<AiV^8&}4KY}^noWaG(rkc}6@gKWYu
zp%#SbC??bi;|LNB6R98y6Gajv)`+7ZiWW~GNGg*CgJ2%Tq+011g5)weWRr)<BMF;A
zrhse;y<#k|fh;y$cwnjSF5GorZp>EP`1zG7fe249BhE%l31$SD63hrPC72OpN-!hH
zlwd}XDZz{&Q-T>mrUWyBObKQLnG(zhG9{Q1WJ)jt1QHf;_k|CwuJV2leX`)7phxkY
z`1EIA&-(ErGhM4-q5R{7-c=Ok&`!t3{>ytLNEPid13znTG%3~}-S>8@d4sJ6<k#Y3
z3NL7oSBn?J=egkh8}}ACO6eZ6fb>ZltIlLw5VPQ7Kx4Q9f5Wp0X-tS%G!{f`8XF=G
zjRO&v#)XJS<3S{(30;BwQ85us1d*5~hDb`2x&l$7Vse@sA_Yy+MeUh2<JX=j*@h1U
z++&cjiw^`mKHW#_nQh4f<BFa2eYD|$+%)`P8$N)3ok9$ZFY8NQS)RAN+egm?DbRTL
z;h+!6OnWzZ5-ylva=X&@y9z*VVOIf2z3eIgXYpMHpbsGE?Yg@$6Eg<lPwR)p8BZT#
zcx?4V(<Y^)BB?;K5hR89XlmD=e`(+Yp58q@_1z2{M^BIGk6Kug2T-i318Fw)cD9-*
z%_zM?xZ2CXN>6ukWUx6w!68ajXxNCM%pt=)h7ET8yHXa{h3DqZ7YIdSZy#Sjf4L%X
ZP=K>U`VN!jDT|1lOwt-g>iR_^{{yXuOK<=H

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_softmax_rank3.cpp.F74F60A3C98E1A4C.idx b/.cache/clangd/index/test_softmax_rank3.cpp.F74F60A3C98E1A4C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3f4241702e553a456058dc94f98734dec75bba0b
GIT binary patch
literal 2138
zcmWIYbaM;hU|?`g@vO*AElFfyU|<jg;^LB`%p=SU3{$xn7%Jv$4ZWRr+d|;#`8UCa
zX>(FGzh$cP2-<CWOT}%rQvab2DbE(oAn%~EJ@NaUJ^N(MySL?Wv+$+!Jui6Y#UR8G
zEpjV~??sBKX7_P(^%)0lzvh4W<=dxMk4^T@>ec5HSwH{e=A#E&Z$5hS#i6jUZb5~g
z;r_k-UAMPiUakNB$A_n%O=4?y&0ada`)A6TE|XgO_$A%_)4y-t?Qj0_DQ|Wh)4llk
z?+0Q+wy6bHp1HVsYQ~#cF<oq5k7?^KSu6e8@bVPBlaiP0^AtCQy^PUNE}eZ@@Z0g|
ztf`hIv*#yd%+Z~qwtevqYwrh)rs|Ix*I!Rznt$TSDu!OkOnKMmYr6l3KQUVvw5;b)
za={%#{?<}cH|5k22LD|THgBA;afuJZKKEnV*;QL>zP>t}-7&A!ja@23!(Diaj6(L0
zleWpHm7YGhu62T|NNdAYW?#Qv<<7>3bC$TZH@^MDaBJFT2AR6t!@{z*oBAEE9uc$K
z5ytlVkJzIZ7PF#cPd#w9wUK1ksCk;QGPliY6~Be|WhSX?v#q(dUmoyXJSA}D#f~8B
zr~9Aithu0hW@<{vrjB(v$#dDRMOB^`ZFHJ=PEpKpVOUN@L-Ct~ZNARU_X<wTS^Yxb
z&B|Qi&&;b#rWxOTEp_Je-J8`K6J}=wSbTaCc4kRYnb$_?Jnoijl6}VQvtA!tcQmN~
zxa}+L*E<X!Jd$Z@-C<#+yZ?8nQ~av3?1(#Z(~qs=U2}1!M27r9vCk0#DY=0Yny)2$
zMlbmKrZ}h4q;meB+@$$;IYP{5?OqkWj&u43uEHfr``QGOe?@%y;_vNNG&}O@MxKiW
z$-DX9KYHL?u&U%n$iyURkL<Fgr>$=ap3>TPM5HiCzN%m7!9xwP&zk>_Rq<xedHYz!
zY1!Hsqb*;hjiz|Ni!JkB9o2f$lj&6cj&G$*W{qkge*`uc$>zTP5w-4)%a(gbON$Hj
zDyH~o9(%rcvf%pab<b47KCWkxoBY@;{nX@Fu97?cAF$T+%J%jR{PVQMKl`1<qrOA8
zm9G4FeDIxgxsw&gO6Scs(kGXBvV31v5O(du^A|ZKN=B(JOQkd?zm|@<)U*7|{*=?I
zzvNF%^nFxbeT>=E->zido$|xA=dauIFFfLZ;J(wJbvfS?{#&o=t9@Jj>HF!Gw<mag
zdEzc}>U~0ZRmj_)cc<<bS~_J9%i3=u2{os96P?W)Zr7Z$_gnwN-{9D1_x?$<J_yTb
zcgZhRZ<_92n||ZjjCGZ<rh6u{zPe;3y2|S1soBS_PXAPQ_}21j-D!2q`;N}KH&OAQ
z$7c0ApH9|mUO&0JxH32CCNNJkytUr*>O+tgH&ZwRLo_!VH#?XB$$)^0_KOQF`@Wsv
zW{Lnxi?fLff(eKu1A}Ytr;l0d{zPyyg#x8{*?2j@1WdZ=&DNV$bJRq*84`ig3JeMY
zV1fZupa8+7pylhoz1W`2%@hWd=4a#Q1QQTR28P?GIF2(evs7hgiUdjvu?Y!+37GVQ
z3%XOJ>I2SjGbCX%oS`T+t=NJIs2b`+W-b(;@kC2mNLj-C#}h4MAp_^}8g$Icnre8P
zktdp)k6Q$0g5vD^9v|K*n=-P1y#ezb%nT+jl(6B6R*+NBgefxDTFWc=;FK+zURVGE
z9R~_MnBRbo6O)pJ84TnZNEyR;-ltTbuVV6#06Gp7qA)GNPmi3f40-ttY5^>CaxV3;
z1Se+&0#ysi3u(c^6y_{OCKN;3Ke)<Rp6`(Ys^%5o6@$4B7E!ERC=mp7m$H-^%w51B
z6q7-6vy8F~vYdeo+y|>)H&;eJoYThyG=Niz56J*d8949I<KCl>wfkFuig?(0L}A|K
ziI!xSg!4|eP0+f|`Qs#-r_VoD_m94RyA0?Jad~kSm<2G;Gjd@~NN8bEl$w)R4zv~)
zg7EYRGloH^Furi&{np=b0i&Go$%)y9l5l~*r?V@X=EnQL1<L>K*>-k;+I_gdZ0WgC
zvFT~PaDhv^OG0B3{)I3U7bO=5gS?#pBEWz#x_-+I2N_3^e{2bH@lmlc3^6b%28LJ!
zGYY|sM=;|M%;emZoO~cRFgKW)g_Vt+gOi(wSBOtgfZsE~)8ALv!%b6|%U48FR7^_R
z#@61!(aG7xP+ZSW)k<3to2Z1WoV=2<ikgO&j;_9uv5C2*wSt9=fvLN=gu0oxUow+d
Kh^r4{5CZ@UjVz!5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_softmax_rank4.cpp.05FB4DA05EDB8025.idx b/.cache/clangd/index/test_softmax_rank4.cpp.05FB4DA05EDB8025.idx
new file mode 100755
index 0000000000000000000000000000000000000000..87c8bd25709b13e712d33e72a84b696b17ee7717
GIT binary patch
literal 2036
zcmZuy3pkT)82&cP%nqWmtT{E3InHq-5ixT%iM0?zVG651`zs}fq(7oi5-C?lAq$=8
zM5QhjF2~7f&W&`SlJtMnr*!pR*Y|wS^WN|Cy!Ugz-+{@nvlAl$z$TCp9UipP9{~V>
zEO_yDMuk*M06>N`07U1Mc?!1+sf(LliajVBbD{;K@ECjM9piR2yVEOEAD$#I;`#OI
zC-lBfn4jotjTCHqQ*PBAF%W?pEJCQjN^JAB2M)Mw5S6yxr9DD)cJ6(;YPVO5vPRp)
z&S-u?^f#y67pb~VgZM#u^7Hz%URsTKOqD#==~8s>uIN$L%j4Xc;h1cnx~(Erty1wf
zZf>2WuZ7cF%h$nP6@`3M^nP9=OmL$#yWtrp#kZ)Ty{I;5n_GD0$IB(R%a!oSrRp@B
zktaG0wj|G72hxV)ji||sZ1Sy$p>oRyZsu%z+wHEA`_t}|$S&0+zm`CF-qXHPAOX!>
zoBV=>|6cqw+(zA5A;CGKgOZxCi*P>8$l!p6=Xjg9cZr%aP(^?27-m)$Juy-jmXup&
zM<-m;q%SELQVsiDE5UibvcJ2<AW1rkT;guw>{3RGjqS;aw>}g*x>u$>Q2-FD%6kr{
zn20&YvRfY_Tavb%=~_orq@VA<j#)c2Q%kk#(b8}X!e2T!l_W62p1mG?HC&W0@hK@i
z(dxX&VKn1=x}d~bgOYX7{lUuHrH6Nou*yE>YQYoKGHkRDu2*jxml|x1!wR$it?W5Q
z^Qu>+x(Uq&kbxoETW>cuogDls_~fQy>RuetS9|RwV?cZ&(6j?5b<~i!6_@?yaZwFZ
z=8)ou)|;D@?kL0q>&?V#nO)z=`FfsrDB5Ewj+_?B<{>A{jU!tQ@wKJoo!yVcJ?1dX
z&%P7$xO^ji*0fJR-fO9Xuo|Bx&6apa?*uB7;<qn}vQL{oby1acv#yZKn%Z{ry3dh{
z$Z=IF%dGIK{dIH!<!eL#HL;ZKGQWUi>gCikbkumn(hqI%`#F_U#xEu#<^1yty$x=B
zSRubUv*V69GksIyI~KfQ$E`^-*l8<8a!eBIR!Vu~UY)G%IVPPoyLdDQ{t2g3IC`7_
zR=lTZfV%NrF~X!a<g}rPF%hD0^ZR}Cw9Mwr<a3{&qw<<NLaKEVDr{eU7ClfRR9H8o
zPuU2F*!4adnx8^P3SW&6b_Yac>J~7xSm?Blj^6YOYl3Ff*}C=L-EdsFdqNMTORV?#
z6MgxLW5}wRTQce~@EMoJu8dKfI0o5txp<%JObl(;5`);>=>s-j6va&J6^VPo_g*2I
z2)%Vz*DP9|%d=^Hh3nSM>rMR-Hsh!u`f~p0S+i-xljL6d@Zs)$_pTKtwBRsbFXOj;
zOFcu*U@~0V-v>ud1l{d&`zk(LQ@Z2+?n}dp3caaDdmifKO(Tt*YVYy*;Q?*nI}MDQ
z-yIoelF@K40AQmf(a1S7i_8u+t)W+ls>!ElxHkynCGoOz<_8D>wnyK;3oV-VM#DWp
z7%PcI%^3*hZt7@XTa(^^1_D4>1yET$XMouc3OLf47mFu{DmZBPMi7>fltImzA0Pl!
zmdWV7ZuB%n1O9)@7Cw`D0=qy{N5WYkEGH=^J7=J5ASx)BM}>po&kiM|7j%YU6Q~3l
z)E|aTq!Q;NtbS5<Xa+?H!?4jf^irsT>gg`~@iCGy3^CUR#1E<gmtLR^hOMHcq6sZB
zC11nJc0aTDBNq}7BsZ%bL>?q3Pf&mggONVL5Q-e1sSjR(yLyA<W)+1}Jo>Ba_#1{N
ze{KLt58Yw#1*^FYDF_cvs5@9hPFao&X?bYemPmW~9|4@kF2>43?4W~MB)#B(Kwc!m
zDu@>-p*-<7&O{RNcTAr+-@(;4aeP)!P6`|>fFj_2D}W--N04xT5Pv8-S|id*jR$iv
zNX#;*dkk9vsW2baCLSZVpg!0Bp~Xv)hAX>E_z&0}UKy_j?F%&zlV0!uMFnyFXa9PT
zQ~<ON;6-tG9<w`dnKeJgX2=0xY*s*KFxK0>S#pc7-zJ~UfX_SzZ2IN;{c?SOxt#Do
z?hY_?5BHEjERsY@q0ktt98Pwz48x7#>b%6>PIHO0^HPOn@`M#@EG(_oS=+3q;C0uk
xn`x>3ZFae$lJZKD+N#xLZ5=%WLnBj~xeAr2Z%oH8*D!H(;lLd>*gC<O!2czr0agG2

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_softmax_ut_cases.inc.A86C258A9B038466.idx b/.cache/clangd/index/test_softmax_ut_cases.inc.A86C258A9B038466.idx
new file mode 100755
index 0000000000000000000000000000000000000000..eae4e8b7f8a8cb131694c496394cc386da5a437e
GIT binary patch
literal 3360
zcmcImYfuwc6uwzj^4MJ=<RK(cv6C(+1||3aE5ufT(ZTuvv?x|l31UHOpo)UvBTPqS
zXc4KcK&!O)Y7t)u44qm_6#*5omLk?iu@*WbbsXzdrIy~^c#>WJCo|m4IXQdpcfapD
zXPG}UI@*^(X!^>T*%>KW$s|HZiT^UQY-tbicP)WXc6n*>n{%4R(*=*?*58f&v34Ij
zs$Zt-X8Tt)?D!=*Y^LVW_x7g`Pp>E(AsZ9ZyLnvurUh+f|1^&IVv3?Ktz+eGdLTLX
z2m6S~1*;R)3s+mFKIjR_$=z<car<h-<L-@H=RJ$c%jz^UGn9SdjeiwhJa>QnPpPva
ziyHFv$1KB6P40}GyUMnh{jTX4)fsZqoKRg<bmZEix{i`HLrgn!I+o;Ck-i_#>6ua1
z*BJOt=(*6UY2P$oIGtNiF0H4prj8FZ)&={1kaniaT6v-6xIQuPq~+%;mAfJan<GYV
zF3PZNO?>*dA+9Ol{YIt&Wp2z^UMNAxXWi0u#dit@q=XD1ia>;dUl2p!ACZKT3~R3R
z*!%gi9SovzK@bX3@C(9RBDn$=3?J&fXZHMCeGHePWJspq7gxcES$@wef4}{lK|^tY
z5-An@;wmtwFDyxZFT_tuc!-Yx4D;@nWpArk8Y?9{ae)r$6#U`>x#7jPXD;t&??{Ou
z;v-;(&C>^~%YxFZ4D!YWW@J|Iiwi@UzGBgmRO@Uhft?||1?*6>%71F=#7WN>&Nf1#
zfQOnRqZ3;<#-%XG3l|7LLcz7htc0Xx|HVrQsTczsVN0=QK9b{dJ6qN2y}m|5@F3wi
zc}%27c*`LloaS@_<=~rer19hYhvbWXO*l~(b@<5~P!T`NNeD%t9Zqw?K^hdpX-+Ok
zgHHG);U**>oo^2u=sA4-;KbeCQJo8mV<a>ctP1vp^rYfJ-MRJy=`R7mR+AoAoJpN*
z)e{n)F%X&%MHcpYPUyV@Ydsxz0wqWuGhBh>p#pM}-vA_WAOsFbo>g!}Oy1?nq?lpe
zFYgLYaw2g+0SX-Z0_ym;eFFN5rw)<Q6e%aYfNYwg8Cvz>P=8xXZAK|gX&8;__(6Nu
zlZuOZG-Z$*ykG;GGU|<^ToJ2h1(8W^@&f)6UYkC0XPy5oj@Gfk<DBUuc^9gis~=(X
zLQf>gzfv3Rsa%>fB3vfP{O&2$CpFw%IRQb0b=i^ahSqB@NRpyQ(PIHPNg0M40s$OJ
z8MQ_~SHx;rX9T2SV+j4?kOXd?lUYlz-I;H3i;AEoz8TltAlJ4BwKq1Evud}T2|Qu%
za?Jx1#L-F+?V9Cb@Be5pjI)TE)C(-)OOJ#dutj_|BY`Mv5w$@NOk<0%Mw2UISym94
z)FwT6q44aJ!f7P|z4)@^1acxAMN$e@5dcS#l!{daz)>WniO@v4qHtZfzP`b<$Xb=s
zO;S^IQ=HN6-+D&pU*5mr6-Rw)OZzi4#U~1<Ipx6!aC>YB;aqWaEI_F)e5<!1_NW9u
zE94K_8>cx-gOt$AF9ARruwxN~0K{-=!#8{xyL=z*7G~<BOS6wn>+N!j$3P5xJ<iwg
zDiFUI&Ner-9g2RS@}xzK(IW0?QESrtTmTYj(LAR`Ed=hu2!u03JV9FgEL%#B=3gZ6
zlu#Qr;8v1i)vPnRMChgU3tI2GMFl4_;sM~8^}u4+fPDsP*#Gq_j=R|^YnB`gf1Trx
zBDC;A#r#n+Em#@zM=7;naLga2)52WE{82$#7^j#&%B%%I(1w^l;e-~?>$_J?dvy78
zFS*4-_x8qH)`a;lcFX2}d=_HVl$Nw%N`(e})G=>M!DM9^_qvh?Ghm`;i{snsCsg;l
z1<HW3_>J@vc@N#P`M=l8UwB^e>~YKH$d;0xY{lIOPIUfT?o5EWoC)xrGXd6gCcwGQ
Z1XwcDwjz__dY@X8d;VsA{kS-U{s*3{2Oa<b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_softmax_util.hpp.A85F4CCFF59B9D2C.idx b/.cache/clangd/index/test_softmax_util.hpp.A85F4CCFF59B9D2C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d2126fc80a7f12e7208f32092d4c2b9c84734378
GIT binary patch
literal 4926
zcmaKv30PCd7J%<uK*-Gn0)!<HAY34rAXI|LQbdIpiago0)c{syQ-gpCQBbt;xwO8B
zC<OtP($-pATtKlvWp}BH8;iawV6C;bS_^2U(pqYLlaS%g_wjw^`<S^u=VZ>DGv}N;
znKvgqe7Y4ugr&~O%S=-x84?6x0snGT*~*P1L7Xrqh`frrcamcIr}Cc1R2d$aMK+i2
z%Iw}2^k~l!<<z3A(G{n!cPrLrTxl&?x8qgFo0;!M9^ItaSJHPqKezLtZNIzmsQKVr
zyTOxduRJ$NsCjq5$-Lu?ctb?UH><0-<+bnI&vdFAaSq}%Rv%5AoDOEh``uI?^e(aZ
zanV2TPf{Ovx@v0=c*FahL5oFHUK!b%Wgd`Uy~k4`>VER^!<rk;mfzhvZ8dZDCu^=8
zoSEf1KYcXn;Hvb2^&8KA-066BpV2kX6{hJ&?SHRY{?_r#b;r9N+@$)ZDZk15>r_$M
zx2;8W+vELHJzFcE-r0WQ&RT=(9m`!d2w#ZW*ZeJ?ZDhY{$}eZ!Y?cn4xPRll+p8CG
zZx{ZPeJ5e{gCm>5tE1mMR8yR~q=r8lQghtT>{HRn+O(^=Lo@z6Qrw?#af$cy2+#r4
z3K_9`=Zj0cuO<sDdf)q_@z;|kUMa~znGH?rb1${<uYShueVAXCf2F+HGjhl-$h*77
zYSDJ{%Uc71IJqJ%YPiAAJG7E-z30s;(Pt&4hslHui?>CJZ|=`uU9hrx+2Re-xoNu|
zh8bnh(OVCmOE|KnmOS=_E3LVBmlGrtPYG$ASGD~Jx4d|G<j~)@dP|d&g_XrUeYU*N
zuDb!tw)hA??^dO*Oqezj7aw0|=R31OBOg#ln+~_s5A?SDeu7&6>AsaEk@pL1l3vy3
zRjli-yMBIdnK+_7`DyZ9zx)+0-!{Ab^z|oxX(>gG9~N#42u~?74$jK`(sz<_aN*~g
zg&*yF+SUGI^}yzZO_QfRTxSdl{^&J(QT%hM+dN;R4ck6nex6$`8oB4T@P~g=uYK&r
zDSNBbi>fpn%DQmLdFi^2H*(fzCOa7r1XxsX@yWHHOSq)5J;5{rsu5W0C&C2gD*TWK
zj8k^?RGs6J9H`O;s5aoNpO6Z7@bSg;dxG;^k_T0q0M!In8Jm!b(=VJbn>Af;Pa-J+
z!JGodr+}+|A~;5HCj?-CsPRux&zB<|X5?fCf{{w4KDq&|A^k{g<Z%0gUM|UnX6gV{
z2Tb)7Qt{d`<Lc@*We|_#!xj8F{+7A{i38hL6t9>zcQ==$;R^MDs)s9#O-LmyBGC0{
z`hjI!(iEzk2GnW5*H1_#D*S)moezGC<&q{)r3Fwez(zkImAHz>!tl*stm2YpP^B4A
z%`hTk6H>|ODKwm)k{)MF+Q5j2CW$O`1HqZ3Z4@vX>iXVueH$&hd88Her5td|0j;0V
z8oU2Gw$k!@bB-Bl3spQhp60p%iA!1(FD&Ow&*YL8a2E}LY5->X38|D_=(#*Bo>9Uh
z&7n#s;B-QF$0nq*ue$Y%dy`bI@N}UOcgo#bH;jwVES*>L@-WN~Qz&+&T=}{IiTAJR
z*gD5Ive}IEgyJcjDLmbP#1)};in|q8AHW2JVmHdoS~nnZ<NdzA2i2vYn2{b(ED=k5
zbORE11UNtY?MkJJM_NO%H^<vdHz4toUr+O^IFHWLgan3|PxBpg1HmByA+>`c9u5h5
z885GUW=&ec?~5J9F1i7IKRaz%&QY4J5|Yx_2FhldX$R5mARg7o_)O_5gjKu&>}~)r
z6mC>8aK*7YVHMAS=rin!wr4k&p1NE9k~CxlT!AU_2og_ViOBxOH*1K;!D+8mLKXm}
zue%eKo^5H3MfIrapg>!4C@<ZHB3^W6@m*lj?FY8|fe_^*DnkQGUwmj~8rF&6G!UW^
zqM>hsCLl!4qT+;xw3Uerp`wJb7%N3~y+$Bt1VYrL(7I4@tS(u_HW1wg6e#>r{vWQg
z@haE@x84AJ8bB~A1XQ_jg|V__S2zcv&w&KA0;*rAI99c+;#m-VmOY<&H$O9Y+_YZt
z+R)GE$Xhbgc!dTu!#Cn{(Xr8t1INLfB_U5}gXryOOOHoI!A7PuBjHHQQ5`_CmN>8^
z<w(uZ`O%EjTI#@(%vt7+3WR24o-!Yn{O$am&~eaAfL#DaA*LZ#tgIYZmIFI<_B8Ab
zqD~-2c}z21K-7hEHxPBRydY_B{IO5_gWyRx2pojSy9LiOVnevfuYn=67ntk?<MtL`
zpM9*dm^(5-kTiGB5^tt>bb<h75lkv=4@F76CNHE@c)91Li6JAANqkYnAo)qAV<eY^
zWAwJ<9hPMFGDj3KL&lfqYk`rguNy{wzS9w9=Tozbo(+~jkA)6G5%M^@`0JfN6<te(
z|84`pZ6FF|TF#N4<_~u)-DFHNVGd!=Y~Z6s(QYioi{dd_?ywxw%ob&1vQ45*80`R}
z9blZ-J~U3d;NgFJLKM;Y!YGctXi!+UyVu5TDtz{{A<ei-Tm$y$n?23AN!$WZ^QIYh
ziF*KgPtc55A`U>U^@FE!VPI#0J#@SdSkwVq)IN4b8f+Wf{VITh`qpA0(uWx?l(?c(
zfO#vE1)|u}jK8zLm<>vRa{xvmb|F|$%7LI9h>%~}y!Ge=-gw@6bm80$Ji1vfI8^%G
zs*S(&!?F%>37L#c8p7y!mxDLZvW3xm{*2G9kxhsvdNG?&x)(BvCe1`BA{L|7XlA}*
zK1Pcbi!sVjWMGu5$i-+2P;3Dh?*fWlEEkCqevh3yw0FX;P#ZL&^tS-Et2F~bGjK$0
zkB040&;lG$zoeOd;MEU&*znx}UUyjD`-8N7R_me1@Q`gdHr6QVkvhz(u`30#w=c`6
z_^D;#S5SYf{<RlP?FZ!$jGOkN!NyG<Dvx6078@FiQDSH!M#@knMygO1M(R*CMj8;R
z0ppBP8=BT$VldMZxi4l~5{pryBoQN}M2V3~qQXcmQDdY55)Bw<bVq`wwHGAJv_ifF
zGp&flC{dA!ky4?=NTpC=q*kag(g1}9j58`>E~UG7v8JU$sR#Qm3A74yVmA_G6@-!8
zPd<y4#rnl!l<1d;k<w3zk;+enk=jp<kp}o_z&KZiejn@<_ixkJzGM&g4Bg8c_WT~^
zlyI`p6XoESyr9yp+!VVB#c=#uNQV>E>M}uaa>d;zI}N3&69jdqN;FBex3@qGBC$w{
zVpTuAwmiRkAOh;!*x1-1!;NMpT^AG&qF%dc)aT$mLwmVJOTxBC%WP5GYk3tc8}Qlx
z+k_3BT^9uw{yOC@e0RwC^6@T3GqL<wj1u{Y7%BNmj8uFTMrytqBMsndz&LjZW^OJ%
zqP_%oA+wVS(B{w#Yg~K(Mjl+ru8JG|F*E12-e2YY>(;3GK+tQI?gy6r$MVvG1rxmg
zDYEZTV0r6T69j`fP4!3fJbr-ZQ~~TNfD3BZ=xE?xwLb>XU19vyL+Nh+$OW32YCUz_
zPnwx+JzK!$xjaTbpQYHCSd0>55;01%Ps6H38KcBxsu&eU>KHXf8W5ua7*_!3ju`p!
zN8G5#D=UK(=)I6DG()Zt*=ZR`zOa+d2^*Mq%S$2w=zx%0ME_-zi{NmeJ4$C_T#}QW
zlCxFYhTXKo*vAPy5QIDX=)yiSH*AK_sh=-xq1;^E#X?6XXORox%F2c8qa*w1%sz_Q
IM-lt@f9T5J%>V!Z

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_transpose.cpp.EA23A250A71A4476.idx b/.cache/clangd/index/test_transpose.cpp.EA23A250A71A4476.idx
new file mode 100755
index 0000000000000000000000000000000000000000..26fb935e79494c451449f75da4cc4af07a193773
GIT binary patch
literal 3372
zcmb7G2{@GN7yss)MP|Xwlr>|<PBfM<M0AD5zGjclmPBGGqPlI0N-9gDx)rwt*$elE
zo+v}*y6q|{i6WAlxT*f%Ox~9H_jsN=&*MCE&Uw#!&iVb$d51gK)>f4QLDpg1q?qso
z1P4J775(ES#ILTvLXZU!f|3ery<&I|jmdXEzr37LptFWIAn6=(vC%l4&0{Mi>fyQO
z9!0k9n;(?3HTmny!iMfVY+R0GKrM?^ex1`CmTzzv_H*0&Hu}RDtWwwX;h-bq+QlFZ
z{_BLK$Dyg56Q{?IH79JIYP?q0_)6c|={iGZ>sXRy{l|%P$tRx<tBwzxUTtu!)7Cmd
zl|S&nJb#NK`*)X4N&e(k(^$;$(GkC3X_?=P>ed}Td1$E4K%q)J%5$+sppK)j+%5+j
zenzSKyCM0ps$A`T#nR`TUy?tL8|4J})%>vqIZ>s4C}X&*<yJV3eTGA2>#RIDk*a*O
zoji9Tx8BNcP4r|re<YQa$k<(W@sU0|zfz5$cD7i|PP?z#xFXCxN>azJn|sRtLel$@
zKCwB+jO;BWJhM+uxF=+FrS?m<r*Z8s#}vuni?We2Gea$}T-?Sq4;oQ-<6WL&WYkS~
z-DL5Y7_TN=>5@*mQQ^hfH&+`hTDsyE6%XI)O??n3M`&w`H}H-~*RDu0<^^J-PHiYY
zt60-Zb#td@N7lRBy$A}Uly=!W76*Ikr5l%9A2ll}PRLvU=d0%YqPcCEns4PU-f+)k
zpp9Y>udT48S8BT1*sD#lEkNOBLPv?7)BaThOzX8LGZFrQz-qPpDvyJYD^i;8t>6`E
zWe1q7*UjF@iFWgr?bwI8=iN)$(b~0cs>fS=Q%QGiMEsV3)zw`UOG~VOi@~K_M{Fxf
ztM}^Mmf&?`YeSD)cbsy5_<WR}qWtu1>D&!o9gn2W2fx0ud}nGREA1cs{=v>8HND1$
zRh56+T-g_+@MrpV$)5H{AznE82dk(}I)UAP<+{#&s%}hGe^>lY6J}&aFDZCqa=XO8
zuT)8$n%lm6o^?tSk{;4}sbj5{uf1i%ZmR~xdyRXWm$^=DV~3Ok*m?L%I}|In)>b2r
z1A<!3?alV@^4`7&Q=W4Eg{J4&`#Bf7k_!x<auef5>OVL8QU|U+&w!_nE0~wokCRQ*
zIUBn5r-y8v?<g3UW?|jD+^%1e_gAr>j9F(|w>!`M+RoA)yRW_<jHCYhYGPeOO7lFY
z<hKD&r1OR;<5>DmlhpPr?~2DFyY^1;%1+@)Y>BN$T}@0*C(zEksEl=v-JUn<uV=o{
zByM`kol5+}(cd`lFMLu-3QO1(Do^KoEYhs=T-Eb#oDsQ=`)Of+Zalxby@WHR!M*u;
z4Zk9n)??w*aO|!No>5ukQk1Ubwt1>vO@*HlRWLcY*vGs=mX{n8QiXvatXGSKe(rKB
z3KRlC2n%8ng#r@nH~OLK)$7mIz42NM4*Z#0b;00YxW<V#66}k*83YDZD89r(kXFi(
z6S-wg5hU0v074{)5_F*eZev`7(etbVG6f1nqm?0LqELX`W|_&6B^J_NBq$tpGa;sw
zPyn~hzPDph=cj#0us7<)6Y%0f0WObkJ{o<z^~EJIc;(M7-_h{O?v!oZQ8CybbrT6h
zx=?`VG~*^5Cs_5m6zqpqE+sBS77E}#=zh9<V(?-p3HCvw2?T<;Pylz{y*qAUy5fJ5
zpfJ?^f0ft1%HicygkwptCwe)BNTCY_h#qb=wM%pg{#71cHDf}=h+=}rR``hzkKi>B
zzA=Y^zc4oA28_W!nsLHI%X~c4RsYSFWEDJF{$@nQ2!lfe)q#2eoiOnS8p#Ng$y@*q
z0aDRJ0!9IlWJJb6#uMNMs6~qt@H$f)WTIV0q4u_0h~x!KV_{)A;PjW8!Gz+3H-U>S
z)o-U5_MM1@XFFuATOj8=)ckCw3NR7?1sEd)<tt)8Xkfgs#2?0rj4<>Vh5#s3Gh%Rn
z=FqM-a1M6y!JvSlsDd_J-@A{knj$Em=>%ASD!w2CG>`t48_Aw8CuWcVh<G~FGzJft
z<6$yFqtNI8UR0B#DFJQ8+i2G%q}6q3tQ=WR2AEggTv~eU5J_<5_h!VA8?;dQ&<7E7
zT+))x`R#vy#NrSdhh_>c!69@qo${T=q_cnqYDY)ydy>Bp1-j;aycLa|j=DcHrU2~3
z0kQC_8YcHesz~A0l4CQrSk6i-7&(+0!Quj($q0+Yk_8FKh%!l87L>SCt;%gv=Gzx2
zjuZuoHrPNiBB3Cm4JI1&g&_);9i&Nf;1`}a*Th6=mC=&11gCbrv6*k$xmO-F<E8M@
zU|#mvV|TpX|Cxhk`hIIs##CtwKy0RS)}-%-as!d>4>0}yadjNY5ix_;2JKhy)O_og
zVawox;X_B!f#D1^u&n=Kqw?3{551vf)}Z|e#sy?KBW0GmofIl7Qd7y$_GI6ug_&rZ
z3~dG*<V7_rh7HilZY<$5h>a|f*viPB(LFfrFcEWYb#?|)b$=UL08x{u1K<Z-q3@00
zsRU+}3@aHf0JP3FC+WP^!2Mb0=F(?#gJJz6V<LH*u3hOkZ97ayTVbKUf1q)6fWCXV
z_-R?%_rxdpXslqj!G5+YBs~*PlE9<pd9w55KslX#8ZQ1%9<E1QVJI<F0Upmx^S093
zR;&;ypfAtUaxI7HKWp**_T7V~hZ}1|V*6u5TdMtSvP9Mk{Dxq}f0)84f6tD(MP;vN
z%`|m+y{w7V4p)&^^8@w3;^@lFqqD|N-!A+}xJh_6>!JAYXv7eO8BqE4zl^U2u(n_%
zz8&Sohw@a=SPZh0cl_Lq2l+#AP%spXT4p{$;0pwW#)L)3qN-br2Ns7Xh|Livkts9@
zs+1(xmFwa}Uv8^KCpyiQn<p)=U~XY)we%P3W%@D;m#CR7VExQ0tH@ANR#8>g)Yj2u
g8yGG!F<s2$%-1urlabXhc5se>?N``1V%(ws0m8F_d;kCd

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_wrapper_copy.cpp.732CF6ED569867C1.idx b/.cache/clangd/index/test_wrapper_copy.cpp.732CF6ED569867C1.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b0822c47c858bac582e075891ea93ecbc03f3b88
GIT binary patch
literal 5522
zcmcInc~}$I7EeMD(P1YLM7FT+1j4=}gdGXNgiW>y7=(a=D2mnyf>u;q*o{0(D}n-6
z1+*$!!J>d5ZiphHl|?F4v{iXl<qg$46HT9d?~ga%m&yIzIrpA(&wfLE>GT~W9Bv86
zH;b3BCXR^1;pC8C#+vjb4Fu#$;c!`kCe}defSv4}>k~{WOTanB<2wzH()y!z+6+1^
zpjq#m&-|g}2f>+`i`(`U7buui*oi~sR&DkI$8uR+&t`*0lD(w<vD}kqW=`tLJJ^44
zbZl{OI&0UxTklDWb^F+gKRIWc<17d7MZ*)54b;jdf|x<i_B$QpHwvvxzS-#7T|LyG
zfA8_1?_wc($tLH2CEhA%JJ9j!bl1d`yx=Y|rm!&k()x&d6}LJIqOzKNRI{&`DbSxD
zX1f$Or`9Q3b>9;ljC;R!*P1?~gGV}j21nmXzkidw(mJEn)YgqFFfB@aGk<l<{@(iD
z`Ik+b@1_t~De24hC|UKn?CS|vo>87C`7We-=jBbSYRU#$&qjFaCWLBUKl}6ju!=1q
zda|!iyy&c`G<Dtd_GNih!oi!1wk_((lk{uw?0a-ozf(zPZ1{QT9@0qEcGmIO^wh=v
za+*EiD<6CED`O;i1qQMYj`OnkPqO>Tq20YxmzpfY#w6s<=h`+!9O>A$A(NAGXsY#U
z$Uu4T-j@9Eiju7-62k}jnC2UcxCvz|AMdI49$mV&M5)zqud!q;(J<;>yHkB|@`$sg
zM_&9vW<ZX7g!}41>7iVyu6*`Ju~NBXTv_M10loHZp-Co`!9N!59V<+JUMp)*MQS^-
z;xukwXIxQ^N~-MEK9%Wm=C1RX8q0orai;Wbr-S0W$;pr(6(zi{{PL4I;cY~7q*!`b
z&#o4$HTr^YcfVGLEX@~Ghes+-)V!)1cJJNLzt}&z?7Yke+5Q!dguA=w%U)e`RLY3h
z+8;uVHdcnG-PV6NouJJ~WMvu{rv>*n>|ggyten8;Aou8es@=_sE4N3~Cc~~8RT8)N
z*Pd?ilKjqo$N0YuFHmESQKjGae@nZ?>d{&>QC+?#<F?tgn@3rXV~w>&O0RA@xF@hi
zk|a$qddZF4#``H<BI()k#kSe!3?n*<#<SKnTDP9@wvmZfn(=$;vVLgzT5qk3sn~CY
z=l3Ti@oMe&X-H37j6Rfj;Qm01DYf!#z3kFrM@pnzd1H6F&H+7N#?jhSC#zk(9lx(%
zQymadOYa<>4!O2INAHj0H5B*bV!Kl;6O_9ikp-txr3d>z%xLJE1zpg$?Ja0kh&Jeu
z8BWeVWo0u_wDYA&>c)ZJ-@l(3+PurA9zHQc`Ox*=PAfX`ucWu<eydh~J;ZqR@YXyF
z`)ztR`{o^97FOM2cdN?!<f3H;+q-4RKR$^)rFX3N{)Ef?Y=cD8l4dClXyuFXw24OJ
zu|luHrLyN3i~Al%%WaHxzI;C3{`yqd+wQU9<<AqX@AmN&YhuU@v2@$oqm4!Kcbo21
zitRd0J(u#dVTVNWOutL3d0xP;NBJYFnojGxI1zC!DY(1~ZPx5GT-=n+&l?ADahG0D
z%VJygXYh`XANAuh*7M?7cpQ$P_4Azo>VxGFZW#^-2Z;yi{QW{Q$={EP;NSZU@;V<V
z5yB&Ia6TTFk0*Zm!_B4~<O$w+f5+x-o99Xdb{G!M!{hQ0l8=9Y<Wztl<6_;7F$n@2
zVe*e8=M;>2qxvD*Utx5Ky!9PCt<n1xlGdi&#Bpl+(XYtT&dtc#UY@r2D<oIox_FtP
zb~On?I3n&p9{RqzJG0d9R0D)#AtYp3vijezj~c<@(i0Lh%B3;wiY@p|s}Q0C%RD^8
ztJ7E|7k~FPgfOAy<C)31cb+K*=BN<P$bUl%x}-$Fe^M4uAmIxv&a7O0Lpe?l55a0w
zHEST%{KPi0)We!X_%G=t?chIDq7P&dfQ0}jU{p@snge+q)+tgJ+bdQUQwJajc5ri`
zgN+fi)L|Kh7)A^W)PrD#Bg0(?1u~czVj3}pD4C8-AxdT-Q;3pF<2qpLxL#a;3?-Q-
zSqmYzBsxG+lL4{K+6!U{5WG%xod%eO;4IZF6v;m=9j%4N*C2ZNXjNcoMUSdfR;76V
zoO&U5Q#W+fxZg$;z6B3$!7BkqAW%fMR<(J)?tpG#IkF-LubP9`0V_1d^_wNvK5|7q
zka6S%z;(34GNVREh6m?CurtY71=F<~$xR4(kUWHtH_2NF`H*~skT1y>MaJvBI*M3?
zG$I76d#DFvo6~G*UKon^j}Hf8LNJHU2?fabwed-%A>ND_1Z!w$7-37SRjhR|6sjAF
zTAMZ%{7awKkSB7^Fw-z}+R>+_>E-b^C;u<!be+8OdqYj-T7+lx!e|8`y;rs`^<q~Z
z0hy+eXof&u5QcMxoA)h&<ADgT3<(IPKoljg9)cMV1N8(1sWNl{O9*B{OrdW~Ri@B4
zt}fRcqfF9FG6Kl^uWy2j`E7NGKT|9yb{H!PixLfxFMYLEcp2d&avt>%bu=#h$}bl7
z<W5K-(+)Ncj(|F_y_a#vbm}VN<dM>mihvUlj*^Zd0TfhY9Z_E@oi}GBfV3}?V;Fnt
zW#|j{4<g)!b5cm+QcdoZRtVxb!qX6(LP}8q+#xuXl!~IzAbI_#d!57)9k3<YG8hpb
zFUiNFnDsIx*kxhT8Kmsw;E6fdhvgL$lisqgs*v@`X30VchiAq^QFM=cUFH6TuSHhJ
zJS?_(Ligu<agMkE-xxE>{=Y=%<GY$YQb)|i5PF(7%^zqTWK_htKF>K{=AfkXjI~Rr
zkGLVyMiL{Xu+v5nqfmrah`Dy-Rh2KXEu(8Ak^Z^5BPo%3LVIgR>0>C%G>ZyQDj5&Q
zUKixd=?Ix)5((}4w*4Xs$<Ank(spciopXomEu4AsDw7=|u9Q<Bz*~0k#uY>*usQ4k
zM1$Z2qXc_^@>>kX`&O+=LmW%pP@M+!4#Dx3@y>t<$zrA)JAexO#I{Zr&e%yK%27~T
zge{@yQ_O@=ASDz-GzKjULyk(0Y8Y}-auPypMYbwng&5N*atVf_oTEH31gpbB*&u<O
zfR-x6Wu=qRsDxmiG*1Ze@rryr%42uXp6T#)$1O!Hu)9Fsvhw@1W(gw9HejQ<1&M!B
zq%x2KiGQUiGKP|Ml2HdC!aN<G5X#cbqF|JKydfWN4Djy$h;WK-34|EyQ}_3&q|J6W
zvwb6|aC<c){tOQIy;wxt;y1C&b7U6ZK_VzqH52vA^5ovy7B@dvk<7uiW}{vJ!F;?8
zACGnp<*DNPr{s4xB71#I3V{0|n`bkF`969gh5FEhQE;2`n^t5+3>1UbjOwcEwCT~d
z<H)h}NP0pw&VXcqq8eF;lH7$bS>#AHnT3s|(d;Zlb`z2bT5%xQjASN+EJzk8sx1+p
zbV%dYi3nJ?ZhYsF>z46z##4Q_ONaV`3tA%5H1zN25&Ph)gGlqW>dhr_HGx%6a34M|
zAR2K*_Ci}y&|?AxgVqudG@UBi5qGj5OC;a}=s7_XcKYIyt~!gd%OYvu481t3!EJ2r
z?(EDjYiz;=2ConkhN*<9gEN4;g@{I5W}GZ1s1+|n4hOCpA}hcJLqx{5nL4A#{DQ;(
zgQ45K6E4fjv#4|06I@wDm;%iaKrO*+@OiE2Wv?VpuRW-NC`OH_hFnZOKKGR_e%Rou
zxdHL*XpLxXps(59R(lH`4M33R2n#;D<UsH$@+ver3~<W$S~M&ASL5VY`_e7)S6M6(
zg%6Vto9r2tF+*1W_x63n@BUN!vI~tOt4D`CcjvW@<RhA7tFckJiAYXR<Kqp$xyDR>
zs<%=*R)uIGlbE>(OplwpJiIq*QzepLQruF6a&@X(DvF-@OfM}<Du0cL8y^`T2X;cL
z42u&5(6a=Wo!UiI`m9U%n+R1BGcU6spi>C82(Umqw-+}U{5)GZyIBX83`48_w2WXP
zW9QM^b9&0V+k*G=dglE)Cn0h}O^Kr*UI3yK2ci<40M3<>4rd&h^%u@Zr<r}nV#8=`
z7>Nzzupt*4#$&_f*l-0ljKYR7*f0SbCZa<;tks>f`9+X}htvvAVgffZDn4#`bZiU`
z=5bQe5HbdjDMl2ZCox}gAtWU)Ehj7E8|=F@P{D_8sIVYVQH7+mNY%;N)y>^wi6>Rr
z%thD1SnHo=$!h8v6m1<n10xetb4x3lt-YhB-C_%CZ)LK+jek%Q!7tP+03U+;53qUK
AWB>pF

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_wrapper_gemm.cpp.74B7113BB3F74220.idx b/.cache/clangd/index/test_wrapper_gemm.cpp.74B7113BB3F74220.idx
new file mode 100755
index 0000000000000000000000000000000000000000..888a4a2a8a4b91e9cf554bc0e7b58e529ec9d7c2
GIT binary patch
literal 13228
zcmeI2d00)``@qjyooLvaXws;pS%Xv3Nb{sLCn`sSBz7w4Ufh&)sYDVYBEt=tQr%mm
zOEie+s)S5wFjJZC^}}y>S!b>4%YD8-kLUUQaXpXD`t0}JYp=ETyWaJ#cP~#Td;7hU
zSS*`Rr-X%Jaa=Zw#S&-!$Hv9XPh>zwjKxaGEA?CGe&0l_F~8hhd-5rTU(?Txyj`(t
z)~*i!#dmbB<&O7NFq>zlTN#|VK70S_m2Nux2O=%#%NG-BLMC%{xEH(2%Tz+n*7yBZ
zr&RTHP0Eh^_w{;<+}sjpzqs1_?B2@WSF#p6sWCO#?<>Lm7ju435Lvt=R!MepkMu8A
zRUy6oM};lRPb&0h*SGXel*tPYn*O%4w*LLw*Uz@)&tBEP?QY+bqS3SW?K`ZYTUnRl
z*I!a595Y*H_m-ePy>A>W3vxB;m7QYq*EQWg=I@EytfjgnyUFay9r{2^v*54oaCf}1
z-?sJ<9r0FNQR@&@#a%&Jg8o_c-i;+&H@?n_Ypi%{rv59n&(LnX;wqb3^UhWum9d(8
zVE^9OqIF?jGcK@4SQrM)yn4KAi{72n#WFJ&aPA)4X2?z6sGVgL{pY#1<h>bxt`;f|
zmF+5M(%za8>EOv;GihUhiTIozKkJ)wq-Fj%8s%Tzq&k;6sN(<j?pWEP{z`%C<F-4G
zn%`|x8Y3$@BSAcL_fzi&b&7MO_H)nrpD2pDKU2NkvBdD4*1Z`AM)e2BY`(PGOT|mV
zQ_%Cbg0)KmcKcpf=AF`&7FRP@eA5-7RK6)OqQ*K40(wVz)&!kj**3cD@cj?^J~HV$
zo+Xsl6<@lsY_l+@DeYce_I5s}FahUr)HPx2`&8dW|J4=aO8g()y5;uB<?$7nCofj=
zH+0;MD1Y&8u1)Zv_>ifu8*+~eS4*AE-5RlFnqI$(y79ukJoyEuGzz6TiUzftr6)W#
zuDa7UZ*0=RfIWvyyOYmlT(U9uTwLX=S@V>y>tTLshVAy5Ed@oz(?U-A-tJDai#`}#
zFj3Vu-!a5^Uqir+tPNo=*=rI!l`7Uqq*iCSADQ?<;c@BM`^VO+XH757nY8Aw->!St
z$^IzKo$qrX`c$@l*OtN+Ir`g=nQn?^dp9>3i$vze)Xu*jC_nk<k(!wsHUGG^s;N=n
zzG91crqBPVa@W_to1k9n^lHBNmBfo?xl*lC+f@F%xAL*U8_T;=bCaA~9y|<Qoe<eB
zvE1%O^1;aQk4>GL4jwpBJhy87YX`H3-nk{-rJL%V0v~s=YbQ3|II+NBkAYiWBLB#i
z9~D-FW}n#D`6P3>xls?tZ{um3t*&+Fedk|$qrXnrGV?^na|^{6t7<JiEZx8Rm2&sF
zJvDMInVG3+M`cBg^j0^HOk2?ERjEDtw8+a=od@TaJ=K_Qb!fFI_l<?toyUhqW>4}j
zh?qVjV0GAywdGRlvbqXw>8R(*K_^=mr|*wAdbXr+`zFqL@oOLZR`->7{Hc6-h412u
zfNGbE(lJUYxAHY~H<t-2Q5Cn>6b9>AiB?+dRZ72|^KR9&5VaL|cIX~ox9ZG0BhyDS
zGoz=~He^>t%vxJ^Yt65>%0I5<-YQ!h6=VLR)phsQ%+Mf@uv@;BuSWz=ZOwmDpb&9$
z={>n;mlVq@EgMQUNYIXlEcnY#{w&;*`ucrw=Br;Ty_S2{jlZq8W|?q`x@go9$%%V(
zcOBy<Ec2N)RmR+3;e&3a64$L)$vJtRL!S~S{N^H7><<g)9)pZe8XL9G^mu*#NH~+l
zqMt$5Gsyn@!y52e%s<()mVyr9Q3olOAEV-|>pc1IKc9pwmfGp&n>9Aw)JPc2h_67_
z6{dzyKd1)vaSz+YZXEi7V)-&E^8E5*fB!)>n80AuvaCe56r44iMW2DJGfWMieozf2
z$$)P`{6sA&I5Xf2&9vf+KhT_i$6N{q&-p?-9HEgxSozd7BjGHj9slt`ZhuZHZI=|U
zrdZyLiiwzs{NI02Q$4mE89PC4ipE!(8YZT{_U6Jp{S?cesUZ#Wr$MpLKd6?lam7vB
z9!)8ef`LIS`YvSMWjg%n2X%PqP4g#(d$Xo}Mde6lv~_EgR_;g`$UNvj)>3n$%zvI&
zBr~#K9JX4|rJlbWde02o7gesjJ>_m!TYmB@oxanrWO-O}&hxJ{x2JU#$GSwx=qvB%
z!E|m)73*32S5#uc!ecjRF#{P~HLY=yrYN9jjvQw^vM})6)otn-+lM3A^m3!+#>jNp
zbfQrrK}A5n2pF7cu8(_bULVe;i-3I*aKL;PxSa*=`0fk9?E+?gubqpVq`4C$DB4Tc
z%K)iSw70G|K|Z=Z1SJ9OB%p_NlYw?J$$7vv56r;w-N1b}@WOl+=$r-mn5%()HQ-=A
z2XxMn@(V!c0?7?Pzk$@h2y`xz@?K!m3mkBLAAs2hV1+p!8s|e(o^jK=q492#3!!Zx
zbVTwI$KLoSiY_SnuJ1wMn)Pl+{}r_meD*iU6~Q4aXcE#d8?-%vjogZZ`H}*^vFYoP
zqe3;IS(wszRer0p;7`aEHZJtmm|(Oqca;8`ZIL+!H!UeTN-k;~9-Sz&sA&YbMtNdd
zuCaU~K{m^sFePdv60$^_L_)R>7_9?lSa%(;T?d>=_F4yKV$K1a9AJuhGvI6{xeQ2`
z0R=2C0~%$3czPL_Q3iamyaGs9kopzCssh-N@);GR{&}E!9%y3uC7^nV<PIRy0Ti*k
z188=TGoS-_bO1jr?*TGBq<#;W+5_xKxknGFUjsQc(6sut*E~x4&=rcVfp#^}g=D`P
z7>KzRa%xHaT4+~G>eoWwS~wf)*Fn8HXhgDg9dsnwrw-1-+zdI*&{WDvXwLXNl|YJa
zhIY-+g=D{G7>KzAa#~3J7HHQ(>bF4O7C0O0w?e&EXhgDgD|95;rxnh^+zC0I&~(+_
zJGSrc<||TkC$#H?E+qSP!a&Sjkkdu#cR{-@QojrOcEQ<LzZ>dxLnD%{yP+e=KHYE@
z=03>jgQh2c-@on}y;+o^`=DJPbRpTV4+diHhn#*=zaQH5lluM8w;#^N`Y)m0OK3#0
z^-JhTvd>F63-PF@u5-#%h2}5|$}iCH7if%J8H$h~L~+sQWgIyNq%g+1V3F@m{<d%S
zGU%24qhFIsPTCl<^4ToFrv0@1w9(W+zffpUphnSqp}<}!fjTB<{Cw$lwe`6}mQbaK
z>ScoQcdeMq&Mapk6tk%MFPYul)p<*hqAvsS%Rmx6g`%5)coWH2f%sLDTYz{A$&Y~S
zBhs-v1G3LZeg$M-k^Bb8z9BgYiYGzd^T&6V*9vihyoYpyw-%tq0CfcUe58Ov2WGSe
zB2%PB(WfE%G!#Xnr8&~`rmI4FCi9qFAejp+QUA4nu&dwX3+a&*9WyP47c@|G+_X4?
z;-|$ElmaYMfGetkqEmroD#=w~auw-oYr*7N%-XNDi>2-?{3uM(fs%pBxSC+eV1l@k
zT!Mbo|B+ZsR_L!F=yxFdI~b4asROcgB-cR68mLUVHpynFjO=v4l}L6%WpZju_CaMd
zwRHX{bP<?)RA-2p>bdWEB(+$(Xvm90y8!?{N9ZVOM|%K9@M)idCZ4Hc?}Y^;xV_B8
zD~BTGPz=4Z0iIkAr;zMg4&5+UK=}%&h}iVQagVkI5-G!;VCGm^+IRF#@&7yHw`hM=
z-_#H+#<cj;mJ3-m+6@g8Skr~NidN;{H!`z+K5?-yU%kSxr`x>mnSEH3^LUuRZkdxo
zRl{lsDY^<sR*`W-Es(6mY=3I?v#2NG`Fs>z3<Znv1RY@IVmJZ4Oo~1M1y5i(iuo9?
z`f2Bf>}sf;Ib6*%+a6#SJAtADv;v3_J5Vi<2;c+F0*PQKSS^^41)BvEGOikzka5kp
zge=r3lnAoJjlv0vFp3~3ii(oLGc-ykinpJk=tbs>h`=b?Jer_o)H33}<<xS5exiON
zD3MAeXf4oQ3ykm+i-1@W7>o8u6kP@c%76rxR{((ulFx!MXUXvG0vL0F<a<Eu9vF-7
z?*?MsB-cQ-8mNbTPAyccCAkhN*Fh~TZ-#2kP!Id07O2)jaw}ABg<4qN3Dr8G9`;#X
zP_2vPZm8T1wXnPos`WuV<kKj+AFB0}{1Pg^gjz`M@b0G>IV-D+7<*0wwrQkqPY1T?
zn0Z!+%n&U=Xs?14oa`5mo4FMhGc~3Fw-n%khb|SkrDArL4h%O6+j5=pt9yX>9>9C2
z&bd|V8&^FU#R%2`?R9wfH+}L(T1q!#10ztiSM@+E;`Ev;9xb{r>iF1n78J>XqtVo2
z(;K14Mv}9kNH%7-LZg7gTSQiWb0mbaF>IkhSs02A6$l-J`yVb4PEdqE1dlv692T_~
z&c5~UR-M?o58kSylosQX2m3|OaCha17vntt9lI}%omZ!@O{qf5MrQ0j<=;?F&f85w
z{Xv#F{|JhfmzP&TGm4@W<P``~lv0$(vK%OwgV!sH-V6mdlY9yap2F<i9^~A*&DH9=
zrdWLYqGnhM&;N}H-8gKHC~9lqwkJ;=#9kG!DcWDl-vZCX>%ibTu*IAYCGw#f@*5PB
zHImqkIVi5p@PRA$6yqnBt1MSVw=vl|l|+J)fK?KhilxcGDj9Q7<Fulk*}|`fG%@(6
znj6jSQT<GuunAaf0@mn$irxY&wvb!|G>U*OTG@hc=NR9&cM1L1?7wLCTvw;yi7MIe
zh79@KhBE6#U3z?f$8Zi+zy%L^EpVvCjO;-`Xh>q5qU#~M9*W{s{UT&vB>5U-UxT7}
z6>o>^c9MG_y9bKmRr?WSKO*@RWWR!<Xq85Z_+j%mZ2Gjyq~rVNGArUAQ1TBbhn^Ub
z$$yKIouoX>Th5P(jdy?h%T&hRvLHJPjzXQJn8lL4k>qU1&c+<&x$lN@FZ;}QoyQ9c
za&0DWwO9SusK02`i)28^oXdI1-yF0oF1<eewc@HB-}QyY`YGw#k6$^4$vXUZ?rk7&
zw=N?y{T0BV0vO}D+5v1jfFtG}VAumDBaX}2;dkNrv9ckX`6WFg!)nwgw+(R#@qe9u
zEx7*I3ye$Xf$BZbkDQt6k1(U}UBhPLve;54gif^j=GT&Z*ma@Lj;2~>f*yn>2cZL=
zpogHzA(D%rNfGAccKvn6rLw1(rVY+JpuR9W%E7Du&}aaCPWVT~)RNxN&T41#Z(0yj
zD>WXrMn}{9OhQT{W)gaFOhQT{j-Ysrc!E-ZS_<9}3?!t~QZfJh!cKn5<&3e6JM4h6
z9Z(5%J>8iv<>~s5CgK#WB&S3~Hp+6!1dWp$N6>WZ>BMIaH)}V7+^yXSN&{wT<Vu|m
z%+g8D1QMA*7WeQVa5xCuFt-7bHZt?r4ybl88q3q5TpCothL8^B(n-#Mav3D=gd#hk
zB-Y;rMRsA%bldQD=F+2vL*_tcTfj^ewapY0EPei{{dtQ=q0i(z{`?J?$DiyC?Fh4!
z176N>eU>w%{+u^+0+U7MTbP%#4;t=+mZ*K3!_W3u%Bt`&7XK88Jta5aFM-%glHUV?
z_jse2yWq-!%Bjv)|E?BaTrF}J@n^q~$Sh$7eIPJKuPFC~%ca?p*F+e9w@7l4JhrrG
z$!H#Jw~&;M_S7l-rdgi9#$2>F@~>E?0p(Do9Ou1uJ<*y`E&b*!Bglh7c~Bgg!|o^-
zwzEfS2V*`JP_qJRBSFD)NzKkjs(&$dqAjG&3-T#iM@WYtjw^?k8rWZOne(>0`Y6-S
zEFhUh=8BI3^`pQTMbm|1rkhs@hsX@E*}^63OST00B`}ZN4<z^F^d3bQ0?9&>j|2JR
z*b@~keVFmfWy87Oyc`r&i5nc0_V~#3KwVB<5AAFyT2o!q4AJ3Bmi3B4r^gLx{SoTj
z*809P!OXpFP@xUFq3DpJ+o5_p^hbPTne(n{->5L=T{w$5>mVyZyF})b(%>17zCR;{
zBW-Dod>@>~Gp^(Rp1wb-+3ho{*F%R%<CH_~a;S^m`v6;)LtD%hP^1E<&yUY;Gc3pu
zmH6Al1j}*Adi;E8X3WdTXh!e~s=tDMD4s5<__-n?I8lumm@QCr3p7MkQ}k9Sx)pQj
ziA`zT;x-&?p)A+XuQNkOWavk?-|TY|j!f<O_K$AAh2OnsD|F!=(<c*_3GcI4xyZxI
zsnL-v<}!KCq@s89L!nz!qlaw+&`2;#{or6>Xz)<^(cJKf!M5HZrd6eCd7Dm{RW~fy
zI5X!;qpsSDRl}Zmrgh&WR^#dU-!<Gs8?t{)o>zY?(z>@YecyjQvCj!qUp>5MlnP@y
z^-#SY`k_Zs^hKzC5p#`0*f}u?UxRP<0O)Yku%SjrqJ{|?I3+)>IjLvC7=yXIxf04w
zHmWXLlWgfVa2BZo8dN|dWcBFS$Uhpx<<cMilG(LMor&hx0-LpD^qm1rG60cb`ER5B
z!$9LOPHs^25ukAd^A*2HnYZyfl$cRZ0_-Fpf?O0uCj)jeW^@K@nE5q1u5U;$Zjk$~
z=GUrtlK;B4^2H&u@Mh@NmE$;*?3vboeHB8-)0ox|j@0i3#oaZr&JzBvt?!t!B}t;r
z*LTQN@08i6v!PpbBGdZ8Ck~i35E>TZ69`?-Hy%vcdb(uD{XIJ6hbJi<Dqsd>AC%h%
z4UpOOxOh5&m^%YGiov}_NAMWI;ID<~;N5URsl}lpRgY|@jlDp&7x2y)Q1k~N{sHi^
zbPx3`Z(r{=F8DWO{9?%X*zv9V*7>=w8P^!&5knkwit~sgDBdHUM~}lSGgLPj*#GTK
z9rlyw0}FveA<4&q>v7<N0?x;gM|`G?wm#3;hK8I5FS>c0edP5H?aKonPu1D#LMY5(
zl-L>sX|gqW)LWeBCD3Ld$JC@IpeBiK>6g3Wvb#ROjk&xNsO|*3({Hb)6jiNXD)nxd
z@x2wzdlH**u=$()_S<lF7t1WMfte2FaBMl0M|S@|&3W%fG>+)_<|n-Wy>Y@e^Mt&C
zVHy-0l4ChIQj7_U;@UFK69qQBgVnqX=*v9s0t$I}7f{p}8xs=yV!#p4L*AIhionG%
zUXhPp|KpRI))@iKV`8IY7RPZHg)EAk8?_*GVOT`?y!k9HPm{F(7Z>7U94<!TVkj=g
z;$kE&&c(%WTwIKcA-Fgn7Z>4T3@%3ViXjU_qiBXa7JBiq`A3Wt5EQ0HNr;Gxi8{@2
za-Tlh(Oz@3(DX5~l2T*k%q^^@PP4YLHISZYsb;D@?mu0XQBYJ;Q5`=)LrX_j&u|iF
c@)R>=6M20j2Wc5~V;8q2(Ame%6?n4#7pcac761SM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_wrapper_layout.cpp.6A8EFC4867A12799.idx b/.cache/clangd/index/test_wrapper_layout.cpp.6A8EFC4867A12799.idx
new file mode 100755
index 0000000000000000000000000000000000000000..49228f9cc3e489494c5c5c8ff3449b22f19f167e
GIT binary patch
literal 25600
zcmeI43tUZE`}ogV>yWY=IbD>UE)puG%ZU(?q*5qJE@P04TWn=qehgD|Q@NB<Ni!N#
zlS}T3gfyj+TZ593QF+})qZ#Dxzd4rm9A?aC{_p?&_`Uzh=c9AJ=UKPC_gc?d&$FJr
zeB51K9|cGxgQvPL3JRP*St^l8+Kc}~<_FJk6`|)R63L>p0>81N>K!y5etjv+tXS&5
zAi~@r&vr|aNh{Cm=e~2>vtw0{#io5uJbUuIVZ}gU!L+X~nVaYB+?jtNQmehJ{gg^y
zy>iw4J)FuaFKp=Fab3*sM-xt7{&DY}lMS=wjx|%3E*rF_>v*+q%L-rU)ZY|hnq*U!
zKGoID?XcnVt4HeUFBb*c9z4pMW+gP0X~cFp*l|RB*@Ij4HydAc?NflVMhIUlt8G&<
zZ;o2~8>@qFH*~f-<x{<K=GKeD4<)B$$XB0U@Aj|bGdC~0(m8#!Pm}-5OmnN+NiP=+
zGyOcN=0QIVk6Aw1E3XXvsgui85A9E;s6B9=zNY8qgPzeDcAF;{?Xd6l?6Wrc4-cLn
zYvfm(5s;a^yOp}D&#)0`Bh8NNH?|w;vjq3wyh(m&^4<|&A6eSfcu&prn86-}*$?CQ
zcmDR;Zz;QxnZvwcx5f_7bDQzqI9)3b3*84AA%$A|&ljEE?>y?7R-4M6vCG3WEq=He
zXgulK;T_4nFMiQ^YUFISgI=09VlD>i8oe|x^-R$l9a<2ca%<Jdp!<ic6HVd-6K|ik
zJ|km1f|9R>+fN;|W6K0vH$AtRd(F>h53$@5@_ifk??$Gr?CiF+3zyqadWAn~HRNQK
zo8ITzxbvCnOzy_ushR4-4_!-K|NE)jkS^a%&q$gauDJkpDQmypMfZT_qw<*zW$vMe
zjQ5SaeQnpeLqGY;TleYPrFQKuhnZH-j`TfyWagNTQ=9u#P0KEeAJ<71WP5Y!>ddf;
zt#a)o-__GHG<`p@iA>5$?Hzfw(=Tn+d#>`?yshKEe9rb6`^eho+Tj|lqNdQ`y01N^
zNA$AW<2qySpn%KcCf*2bH#czFj1-k9t92f8Jx<K3Q7ixLXH|{KK3QHa?F)r12~DQU
z1{c)WfAZ8idTJ)Oaq#(KuW9Q~oy%`CtjzgILP3;UuIt^HsN|^i10qen$ylAaXJmb4
z`VWbFCtuO;RHAFTsko@9uvT)%;DV#-=w}w4x7`fdXji!9ki+jco$7kT*ag^{;>Xv%
z8ZoOe=$QQvvPAQ9KOR+U`YEq#e3F~iryI-L`HywANwq%^q1|w%uy9h(boWj-_Dt>3
z_vq#jo8!p`CkANE=y3CxOMSs~$Jj=TPW_#49MbtzwN?KO6WXkQ=KK4StvB2?UdEno
znz!bnwrl^x23@|L*`d;6)5I;2X-W8nfwi80asII<=btoLyPSP${%rX78#-E+@9R_H
z`oKPGgvF&BykSB7r+%mKi+1&&s}<;~IL>I?Z#eM5?Yqv;jL#2?y_f2+y8fztUfiz7
zof_soxEQ*7v3_mmlH<Dr{4RDFGT~>Rb<sZ<Of>ahUlZOxwc(3zxnS<r+@?dtwlA}`
zt?|nF&By)dCm!MFj%7Y8+LizO(wy9dU%fcrw{h2<9cd@@0v4s|USHX*=*wSb)Zd9|
zGCSt<@XWmqgG&UPx+6<fj_=(tKBG(KuboD=UNdY^%))^^To&cI=EY<zN4?d|f<hJt
zO*t%T6Rua-x?P*KJL}+SpG$<>NOBvse)A!DrFq1E+}=Lnt;UnCI(Xt|65$M%oWasJ
zA2f-!j<MdfO0_y|>^E5=Xm`}sfBi?1Un1!^YP4LtUG~p9c-(l2P(UOFMEd4~me6&k
zrRTXmZB=!!|5%Ce97~>yl_)-F61$F{>P_?C@lpp*36Kab5|`Gm|4^~pgcKFumD8+s
z@bnoHAy5+7`t=_w_B_(-(dit?79Bj{Q;CpFB*|hM6dyE+?=j=leNpX}s^F;-iLez(
zwu<*c@j;VJiuS7KR}|T);7MYV_O07@c>RZ7=#&{|`KKagJo!K+Q|m8S>0Te&K?hF_
z{F6$iyFEO&G=BVrKe>{HiTyjhkR(=qpps<<HwtohoC*EF4G-JZtF}P$X~UmX5*cqd
zu}eeKSQR`@ywE><WRWj&78mZB5~_kHib?*kUqoNj{k8kUi(UTYk*%LAsjllbe$xl;
zMdC?)zt)FVXnmmbW$*nyahu-wmnwL=*pBxe{En|3)wVbdHB!O;qWBMc)}F>0liQh=
z{P=;+x98DW!*SyetWd#I#8Tete0vX=u1V+{7Wsi;bmZ5EGdeF^yZZwdd$eub%Z3$a
zKKno^XG>M*#|Iwyvpf3ZX2bipm%ND8z+=RLsIRKez5YY{=fx8tc1z4;Csgn>@hbjs
zv|nm&=;$1~Xx9hs#Vt)|w{wr$5BfmwYBmZeb?!D;9BqH9GoJ6dGGeZ=W1co1EjCrF
zqn7^bKeVaO8=5*++IDGD!4t%${^58H4xAR!eXN*a<J2BA|F!Q3!f$#Z?U^WHcaeVV
z#uFI5?iRG8U?n19gB*^PoIZp(A%^fV#Hg}O^WrbB@MWA3OZZsEtwyEKJJ(@y*Vde1
z(#xbTz15sxYG`T#$l27Dk)4~Jr`f$1^|w{i(l|lhO>V)+zMK79YMpOy<Txa!pbaOC
z&>JD2K4pX9jBN%boG@Q+zMNhMCoI-m4Co-zJBZ|4eGM+;hd&i_9zl9XActYSFf8x4
z-rv4iySbhdBC%d1<keVjHDgtWfZz2dZd}wx`nrF6g+Gm&eqc|P^!44c?&UisbN;yv
z(l-h*^lML|(a)`w^i417#5rHbeXrZmnc?iXm76S8Q#Gj)56-%;H_+W>4%NBCJ<nDx
zZv7Ra^T6>1FJ?SC74O*g=y#f--#5$=pQU1`(&0-dDp8>5x3n+Q>FK?vudex1yDx&j
z9V7}~DFyoU)vcpaXC~TQ_U8mG6D>2^4LL#EMB9u`WSn4UX=leCm8YYpGkYeU!##Z&
zc{zFknb&YHAoG*^^<^pjT>U&5jggN5vN5h>fNUN=&zL>+d6x6+06EPY0?2#b7)HT-
zFvu5d84U6TI|YM$!QR0jUpO)hN2aWtaAXsXdP8;zM;?&<!qEiA5y(6OSwnV=K+cRK
z5g&;_%c2k;1^EEt4<KXKlMf)v1IP}t(*ZODviAWrhH)<9b7B5mWSI-|=OU+Em_HYJ
z=feD@$h;I;v+_!jV<~cGT!zfcpu7w@mO*(rGB1bna^zSJ<tLH(Nhm*w98W@dGaA*5
z#<G$62zfk0BN;zN9*-eEK^{*aKSdr-AwNSN&mccX9?v1aKprn3zeFA{A%BHEzQUmX
z6l|5kGzBN5;DIUFm9;Mgk4nLy{i)b66`L}7Dz-_*y&0!rt29_%8XlMi%S*$f(qMU;
zu;C_H-X?6b36@ub&1<lA-Ky)O_Fi=mHFphmtijHZN7Ufaj4xyUGB$>6c^TU=jw0q!
z#G1_nQN%HdI71#0MMg7@CVVt8hHM#4>=<Vf^Gsshczx{Gr<Sx`&k32tF_Sn$9+62#
zGtMG>7BPlwnMLdvXA|>mSY9@9%!cJ<lM&glysd=a3d`F{EVsh)%7}RxvG&+iY;|-%
zY%V915yvv(40%Ku8O^wy@a4o9vSm53qpV*1KerEdI($gfyo%*BU0G`;c@1hC8E_SG
z!WzP_VRNgNbH*IM^j6i<e|P?%3LL4uwm|55omwDbixg{Fy1u3JmbgSvo>tzSW;!{V
z>MS3;?VOcVFll4bp3d-6!L*GjpdMTgZ6>qiECJbR*y+(JMJmWO<bVcn13*eA&IwR0
zmeyi*+Bv1dRV=*<xeiO~AlGAQJ>&)~ZGe0oORqz|iKRCo-@?*ckZ)t@ZOC`9^bX{^
zSb7(7BbGKozK5mvAU9!Y6Xg3?dLQxwEPVjE8B3cXKf=;SkRM~|W5`di^a<n`B8?&H
zbTCTAWt}t@avYJyL0&_oYap*9(shvIi8LN^0+A*_P9)Mq$Q49dLDcEskP4@W^c3XN
zM0y(XSt30P`5cj+gIr0Zm5|R9>3PT(i1Y&Fi$r=6ay5}wGw!T2Ws6x`l0!J5g6LNe
zBic|-Y^eSz$ft?^X~tbfHD8+_7IIj|2{*CzO{hL^Ve4CvZ)5A*kndpYJCN^U>$}SA
zau*M!J)RR9v2`PqH)5AYD8GlT??L%J>~asvo3M2gls93QCMds;t?xtm18n^Oax=DW
zhWrRyKZ5)iTR(>U1Y190Onn=o@=;#XT0FGaqjNRrr79=*Y5N(mrC})24n;<6`LqJr
ztU!*8i;;FQGGfb<6UgQSa%8+7Tdv0rY`WcmZ8l&>#yg4SPU66(lifsnH!-42*ZLnc
zo2WNNyr@^!Ftv~S*hTo%7HxYV<%FgD(q62|U-Dl9T7~&lxEE{CI?S(Q-1pR8!)C3*
zf_KWM$^TMU5B)K%-^Ou6LL$X^#%!e7II562!Q9t;A|s1o7NZ$?T6*?okHE{)3y?SG
zt;1yAa_<3*d^jH<8^w(R<lEnOC`;*Q>erLe7}GI;=5h0M*xd@|f&ney7C@3O7{F34
z<Q4+iB5o0&CEOB7@+BbOGHw}=E$5a43Pq-&s3)r<6v;yo$Qy=C!=OA2$-|(0CE`{h
zP+mCF2uB@Rc@c<<fbs~$M?mW%16gK3>*E084j>)Yy>k(l3ppQg`N}NMM*~>;!-zW!
z<pqc<fP4&b$CO!q4CXIH@<PbR5qBJ>FG5@q<YHu6jC!*Ai;=t-<}XH`#mJAzOOR;^
zOkaZJB`|#n@+^VrOA%KJ>nlZkDKcf{l_9<iayjD4Vfu2!mqYnU#Giy*hb-$*Z&qF%
z8eE6G7&jxAW;C1*M^1Q#`aMHK7_Y?KN~{A}z7h{$9FDngD38Eg1mtMUMJuyB8s?9|
z@)*dmn2Uw!<1iNoc`fGFDzkho%pZ@rcqmW6Tms}o%q1!_p9rmyL~Nc2t(inTG!c8V
z>rKL363m~3`6QS>37aRu{7HCd63m~Bxn!6>1#6_>j;#C?Y?y-07^h;5R4Ctwxs5P?
z8rDdI@-%Fi2J>&i8k?YeH#Xai?O6H6*rpiwgFLVpdoV7+HYHGAjk#*9!>03U%vWPm
z$mZ2p4tZ!b_GWwu54wbhF|NT}4J@w)^EI%%%b2?i<q?F7AUbpb$O#dIk07Ry%_E2$
z^3Vw4%{Y<_iX_7rM-eUxmKR0%C|F)J;i92Dj+n-go-BVHk;jn%j1x$g1Y!l*E`juC
z{2l4`9kHR)^DFC-%OpCG`AlNUIE!#uQ2sqJ`yN_**@Vl6@@&Fq!~9zbw-w5_6K=aQ
z^V?zi?ZkXLk+b@@lcC#*H)C;anFI6Z5IzUy&mrbHFn<mingjEf5Uzyiuytt(;Y)}q
zWb+athdi`|crz{~gG$LT#$|*ngXNVGz6_REPPlR=ryEP+Y6SWb-8Yq-$>?WbF8-|8
z`=DNDkCdZpPOtbyy&Wg`Yx(Q32|pTXMI${n4JIS4WMwu;MkY*NfwU@+9#gwjNUKVj
z4XTg{ln<^#Zj95hRyx*W!!jFdWh=8mHa21MT&$G~<+<1(S6SXY7hAIQdDtM2$?5GC
zS7*v5_PZuG-H8)#oWiY7UHx9Vm*mgFRGh6u2fJuu)13g(!7iHEj9!`8j9!`8j9!`8
zj9!`8j9!`8j9!`8j1(p|V}*&$SYcu_R+!j~6(%-gg^A5rVPZ2@nAnUJCN^V*iOpDH
zVl!5l*o+k>He-c}%~)Y#Ggg?`j1?v}V}*&$SYcu_R+!j~6(%-gg^A5rVPZ2@nAnUJ
zCN^V*iOpDHVl!5l*o+k>He-c}%~)Y#Ggg?`j1?v}V}*&$SYcu_R+!j~6(%-gg^5j>
zZl#Dr1KLW;mJ{6(5fdp7i+6{~@bB%%?bz^1_($ER@m{;5oom~KE28Q8|9&%!6V4O;
z^UTx02^Wa|1;`hP{zb;V!)4uWu5)Z9HpWQT$bgx2{;K}%p#D{jM(vn>O-8E8kSmaC
z1!~9iXBASdf}D<3({Ved|FW@aHsoBan#-8_JTY~G(%wL~&}ag`PfRk?S6}QX>O*D4
z>G#|0kptS!$g^&7IkvbQ|DMb7Adwy<?dS?jDjXuxLy-S>yClcx-%{(pWOaZzBA~t;
z({uL;?@ghd#GjRY_5PYCY$oOCnu<;*PtLPD8}WjP;KauJeq^#AIkEBGj7*x56CKZE
zYh7+kZaeCX7zeP9@au>nTYtwBKA!Q!(}T9BC$xR|Zw^bUr4%~>j9PXAsG%nn{(svE
z`0pN;pY_*1WZLUwH*x$a`>{W(P3%?gI&7-$e`YaIhxs~eNP7h*)MLILas%cYAYaG)
zb;vg{e-rX8%-@218}qjz-@*JH$agV+7jh%!8zJAr{5{A`m~Vo7AM^JiKfwG0$jz8<
zhWrTgk03wB{A0*ZF#m)x-QsMerWA_E2`(xwy3{St34tnsx-<la6Srkmwjy1+0^vk`
zs&X6Y(vX_Tp6;JnbxwHNqTz&=g~vZ>;X&hfMGdEHgVCs6QJ@^r5U90(k--;FGR0NH
zKBB%4dbx6mdM@L@5v^kv#|~9{Z}Aeir?Wgbz%g9htBNCh9E*vdVZdSz=y#&4n-3b3
zK^ZB7dndb^h(Uq>3G=KL^R-L^abiRQf0gkZT@*0sOj_7SHBf25p?I5=C#!cSuP<(>
z1eSE`UEQMWT9n=24D9&tUe3|?2;Nh6|I@f*la}ha22puTY(KFRRXCzugml-TF3h0H
zK>8V|JL5{ERSClq>X3dN>dp+VEm&&{HlhX>C+x<$yKxuB#n`AA_n^EaWIs1|)pDng
zsz#O$Id12kAKd>PCH&>-%eA3b7Dz;;T}||>i6QlV(>E`~^vcd08Za&DBjp+7-5G_x
znDI`Hk-;Iwn&}TGh!Juhia-1F7&992Ci<fkEo1&D<(5Vsn(NqXHur*PO~10fX)<vM
zDMnE<Qd|XNg-K4CZia}pDRxJwfCl){1koO)Uv!u?a?N`;jJ`V}3ZN4p#)GW9yXK{X
zzzN+nx^<v-vl!^1W7dh0xsEv?3mpqUeRcW*>Zj8WkfV+xprNv%VE@KV<_5@J<_^e1
z<^jmxz~7V&8-ELbTSnu!ahi-KXiVt9Xu51V*dLl9n*nI1Y$l*tvRQy;%Vq-#k_7>p
zBbx(gu52!#Xk-u#123YHMKrQy13DTxL?hrqi9uWp(q!^Dq!EWYFkXYS)*zm7GU}cT
z14WXNeKL$=N=Ac|VOV)G8kvkhSb8erQek~*NFxo_myWd3VSV2q-EWW~tA7WQ?SMOG
zIY^cRc_)(XguDyMc0t~aWV<2nL9#uN_afO|$or6NALI&TPyqugDv(75vSkYKG~!Mp
zO~z-D##z*XaTV%b1p`B>kbM=5^QuB#Rmc~nA6JDYGp<2g4J`jBr129hzZPlLBA%uH
z8R`Cv3>n`=vbzZEAT}adBjjI^>{rP5knA4hCM0Wu{2P+}2Kjd+`yKLqB)bp!0g^p{
z{1C|=LQcm9=`fHY9b2ShTRMSo!Z(=v25U0Tz#1921LJJmJsbCgY@ZF|<g)SLY#7+E
z4RhOIc{{Mi4p`n!thE#Kth`)on~VEEcFKie?zz}27yB}K9`2q8+mnaw^I&`OF_#bP
zFTffFu>PZ1>nN<h5bG9VLsotf)-Pf^wwzFjWtCWuc~*YJvL7Lz$FlQ~FJRdP$QQBf
zBIK`#>?@+j9H*f~7796x$ig76AhH#ZR}$Gu$Qy}lBhjO7Y)(icvNXt>h-?$&bRtWK
z{56q%4fz`)`v&r6BHIi(gUB)<e@kTFLf%4TTOc1JvSUP#&i$NFNMwbOj}zH($VEg}
z1o;Od`+;%n><}aAPW5Htq9T#-iOl01yCvz$+@-fJiz^|2d;dOkzK{yh$T=FhQ++4C
zTZP6YqsfdbkbMQy7gC`bwW(%1T2dh$+o!YsB^9!<S2pX9QsE$ObCB(5tscJi(b$DM
zZCfm*7E9?LWht$$9yhkLX%Z){h?N(!+#BBlUEFeS+(wL*m)h%$#@!$3*_hZef?7t<
zKWYT6|NLca-I9>wqA{oJSfxh*hyrD^n7;Tg3Y5d^Qf8gaE*R+H^dET*J7rF*gnP2p
z`$TuH!nwmnvSJIJa@w-6L4nqNer<UjtYsws&5``SJN9BrbE_Ar2OiHC7oiD+PhgAC
z?=|`+t!Nvd+cI~x^!UHo<Nv$;<aF0ubY(&hsN!4<Dx^M>zjW_zUwA7v>x}J`kLuOl
zz9_%d$e;{Y@pu>?-p7O*l$<b%9|dS3zYv6@FX9&gTE;JfWV5UfH9I+BnaeUyM$7r-
zAmwta<-LGx(DJwDD<>RA{9zc-UV!)l$j1?XT$yc-qdxT3b3zf~i=ey+SrwsPOdgK;
zaOnPsz<dPcSj@*NvrQ~?-^5~<SnSE%DRG#OgX!b2RUCVZQQXhPgW|9|OCOK<c$hvO
z+r-226EL3u<q6m-ft5#3LJ|9+Vn>s%Md%qTG{KHj1Ii54ju^LS>K0A?kJ8jR)vHdE
z7w%nDIKkJ_*MX_EDr{MW9hjf<61KU79T|rc%W&er{Faf#CXzTZK0$0w5J%?cD<w9i
zjCWb`Z9ZLFxnA6wd$m(Vqy7Kh!AZNP*ZvwKHJ>UTd-vh#k{(tj8b->24K&tE)DLjq
zPSGeT+`G}s)M?HYW6{9bPaO7>{>-_WM;!7P(-Y0aY7{$rG&pQu{<WT7`#ntFv)K0U
zUB7Dbp6akxoY1~Udutk(_U?w?yV{DL2u0rlq9;Pp{mR@)Bi8-;JR#z}hUA`|<uP%b
zS@OSWg3&$0mN;BG9;8Cci5wsEM2<Yy#6>v+Z&ZKOC1Sp#ar=g-hjF4EXscn{k&cr5
z;>ORHNG6PaPx}<SbhQ1&CRKd#>@W1!k(+HN_-^~WyI6v<D~KMD_?}{+$0EL`0gmfE
zZJu;F-{t>0CW4+iC$8F+V<2dZ1XV)CODD9snoT}?tv2=d6Dz1T_462)3=RB2L*{4m
zQJ)<>d{4YAW#8GmzBGE8q<Cjxj0K(e#Qe%r%)3)g<xfdRLep#m#j|7168*FAM5uE_
z{~Y8>qF>3Fb|+P}*XL8xxq=g>s7+C4^L8RqOGN6-T*yLdS&%OvwF^j{E%0w3wHuJL
zuv!*YXAAKCSZzP#3anPan9gwG1gad&LJv8l3FsM_;%!sDKXrXJgvv_&{QhzfP*wYv
zPw2Ww&CNc^NA|U><y+SB|ERS*J!DoKW6G-emf7r|DxV&;69vjU2Df?+sJ-;BUK#(9
zje%RK57QTp*U%JeRaU8uZTEiC_oU_%Q2?W~>HHuLaOGL<-C6HJKxuaD)ccX|T_-(4
z_q`gMpYGh=;dW4*8Yg)3-rxwK1^fbVXwMRU37~w$=fg;%V~9Tnxe)P%kXK@UC5%Lg
z#(Xs77|h2&UW@s)j32d2`z0iH>-mp53qEaf++LG2V25}E;pF!0eA@HYr(4&3)V4ie
zX7XE7X6<aTZOYw)9!E>mUY~}{CPtGkCSB>QDTY4Uck4}w8cyQ0r6`N}g_O$<4xBwX
zMHJFPCG`EDw?9yC=Gz~r+4%McIt5Z=?`>gkf1pby_QRWV!r%Tt{Zwy%pjRLX2@VKZ
z@v8Z==`~1@Fn5XWlZIjLC)DPJ1kay2XYTBvS(2c)a>*?Abq@PFmwgRpUuUwf^V!!B
z_H`cn8W1#f_N!97gGLc)tJW%QRNHduvQOG;v~wTnKEg}W&DBg(%}Yy9Tc;!M*xza3
zAm_m@HXW=6nD(_W{-a-YI~f>uHtEu}yLk`Gp0;-K-hKKRIp|y45AC4at=F*OUtka4
KA)d%b@_zvBNDDv!

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_wrapper_partition.cpp.EDFCA4C2C7218226.idx b/.cache/clangd/index/test_wrapper_partition.cpp.EDFCA4C2C7218226.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0e37ec3b0c085d59e48334ee502e8dcfe9065bad
GIT binary patch
literal 5914
zcmcInd0Z3M77lR)qQf2n5knve2|Fb0VuA?-Lxd!(N>m(3Si_PUTx-Q@!77MPL1YsZ
zidYmEa6zf^s-ofo?t+R7MbT<is#u;@eS^NqjJ(Q!@cRMxyJxxQp83w1GZ7)d!KM~C
zT%a;!aYkZxJb=UDOyFNl_WabLaX8!s5Qke_c`#;f<ZWN#;MVTCj&+26>E}VCY3(Jk
z_GPXVYwp@*87kV0xDd(N4{y0^Gw=>QJ<RMp!qLATTlYxY3E=`|zF_u-B6Z^wr~ETU
zdbI89H|Hp4ZtZl-+|!ufKBSy6@F90<S!PCQ=z7D$MAn(vYc~^bZhcYH`|8ZD{JpJ;
zfx*sQ>2DuqZrmc>+j&~1?5=y7>V3?wb?cq)j$G`>+P=hKD0@k_&dt<BmBh(y{)6kZ
zjAt9?G7By_?Oo=W(KE{}sl3V0|K!^<E>~iAWmh@U7gW45eA6;J%}aBl;SH2h=~A3@
zczd(I_%T0)(W`S?<*no7eZ*0ePgYTd{k!u&T0D!{9<ig~ghKteboGg&@*sL*luhT+
zf8Ux_&W~Ukzu7-<u3X}B@7TMSRn^I=8W#hv%86#h&umrdJ`csj4Vph^m@9>T7S-3e
zVIAuf4*2_(wIm79^&jg4E{)UlTMs>0Q%HTFka?Xvp?C8fw1sLDv^pj2pi7a_f*j@K
zZ7<gq)gJQLUb8g6$<u|PxLALvq0aD*jhwtg?i-a>^LNFT+FZ4`PmJGWoa=OZR$bM>
zn7cK!DEk|JZ8ZDs=HOxVDeLrEUU!}v2!i)i2fhx?s3@gI)b2l<U8f4F%DQnk()Lo&
zu@n1~595l?Jt$sjm5CEP1$!_29NcqX?mW2o&i3PthU@sOGeqaMd__ask$a5S2h$Gg
z_U@~+YVkZk>^L>8MPK(q{>s!Fr1Ab~$LyEj<VIT+=O-2|YNwjMOPGHD=;g=tk|RoX
zo$yDx`K#-zEwAdeznbS_wZ^+C?c$Z`t4=ReEu-j>e*O8%njLG2fi?}@#jf+t7v}En
z4cYc%(c4Q)ijDJXO5^4|@VT+8^>jRywIS)&IlpbcW|b5CroTAh(vW7}VQ3n$Jona3
zLiZ2lvIjQ%stXS-=qhnqRC4HU+tWvVzHRnp913;4VQ};F@QY2e+U}Vy4PDsc=Mh>j
zJ-^}u{e@C4pPg3NTVckVt7$RW>rk>YYE{*T@U{J?Rklxa!9R?no?YIMUd|A<?&vsI
zzMMi?XLzNIlUh5;q3HZ<c0Yl6F{jhI_N~S58^WIJNc83|RMcLwVz2yo?6=pi4|4xf
zKWzb|HtD}<H03)2b~RGE-WS%-m0F$cn_MDFwknQ^-?UJ@L#R@E)yo>*8(B90c74?k
z(U!c0S*N%MDo9SZ^L&`=I-7gWUu}4?S;X5})flnAGI{9!P<wyhFS3nI<*KJK?<xm#
z&MezDPx|eql8D}(rdyhA1GbOu<qd9m*Y~H}YKx&6--p~eqRuhOOQw!5iz)5q?sT_a
z?Djj?9<+{K?e<F#nKZez-t^AK_JIDJEiIZQ83{vhA0Vpb)t2rrnn?KQ7#w7RGXXyR
z!43Bb_-FEjo3Wz$k8Bct<}4iKkMjpU{XxW>pW=3#aU?Z*_-NceSPNOd%7F{}xnHra
z`)1JPg<-^vubh2!fb%X)S)lrg_0{B~iV(Je?pLg@^?&!oxb=Is9)2bq+W$p9KQT$u
zWQ@ZRd=;<AaTSa4pYQGb>Vmh#l)dNkHT%AKSpCx2I>5mF$MUpI?AtaVM06y&hx;6N
zOrvO;U^JB|z#s`R3PmJG5=N8cPvT=_5)wv}%BK3DGOB<o!k|cMEQ)BJv?&-Qq|Lyf
z7}^{Z(LL!HO}dbd(WJ-FF`7&c(-*a41~NlX#PVYWpor_t<*X@dDBL&2KNf_zJZ=yM
zNx9J|;yLp$nmjHq5F?ZFFq$GOku53{Ig2oMBCZHyCz6PwP))Iw7-J`P7GvzhTrtKj
z&Mz(iu_FqqGh?d8ca8eAk&@8Ce<}rp<K$qn!LvKn??!AR6#{P$lMJC9WvU%1Cvewg
zT>$dZ_cB3}2teNY-WcSg?}I^^`k5w3r~s6upM^nceYF-*pKpHDs=SsxA{3`3!uH}A
zd=?&rD6W*L$eAD{bQMlT{5rjHY9!su_&5%PEbJ^W45gXVY*3Q`W&j&8L5c#e^F>0w
zOb?rM;P7P0t%2dgc@!b)Nyrll5l=`i@t<u!_ZS@RbL}BQ&zG{c8@Bg?U{8~5Coxb@
z0$TwC@r1ON5ow=(wbH*rF4P#nwFweQs4dZoXon&aNW#cSRwRs!N~1cXnlw9_8;a<5
zbPQsGOpF&Uh08*s10f!The0AxL_%d^P>exgV3-z(%UaiWta@gk141&QjEtCsS$2E=
zy;4u03w|OKk~Z)YnNT!LCKL^m2}Q$XLeVgp5G2{@C2!blm^*4sASoD?0uf14o0ONf
z5iqsbWHtq{frF-c(NUC1&7>nYi;|eK9&eaB9yTEp$VR9VpTIXll;C6#WLgySXS$3x
zF|~P2Dl4vozq_hv;+rvX9v>Pq&y`m6ZcLnKRF?(%lm^oQIFoK<4>XgGwvHHfb0%@D
zQ5oNj?}21F(^5wPbylt%l?R%cRQF)U3L;#aOa{{d^(3Ato`ZN=(65(xn&s0qCih(W
zykQHGaJdgIk(pLAwKevw<{+50HYZ@z=<|-*Ke?e+e`=8C6il}hnhR}^=#l?0)v*oz
z%l4ix9{0l^>p%Ju;A|7gR%jkrrYtgw96$#%6!D3CE5tj@X=|As206(vZ!np!O#9A)
zHHETJR5OkgM?p}*PbEEB%bwiU0U()F7K?1Jt8v)qV&26X(HsjU{~C&2w3jm`&!sTm
zlv7ocb;uNe$TYIHYXcC=i=~}z0LYQ(sO^~m#0J@hNPYl_W5CgN008m`@X$_70K}*9
zwY?vJ!Y#tJ%@Keits=D<0ibAVw01H8pg6BMA0#aRlwhCWh$1CR>42g{Rw4!^)04GN
z2Y^!ODHxPSPs1P;U8O~ZUGi9(T?q)6$M7hPv{>--(ELBRet>bLpx_*D#M^s=jr_GN
z^!cVSc{)XBx4y)0VHA&GM49T+&&Q^T1K?sx@JjGQl2N&QB|p*K@5Pvef=vA*uD{cG
zO3A8;3d=D$vwq{HK<b2XBmkbjXx`d)5ditnd@v}Bl4XtL0D#mKwHEDO^rWD!lanxN
zduh1HYqy?o+;&W^aImR(ZG2-F(w>oF>GOvQGIbulq!P!;9ZJU}g67~&i}y_HAANXk
zZp-$Lt+XE;9^T9*St7*^LIN{^C6Zg@%Wec4Pk!9(_+UkTEzCw&va1VXdwP1}DPz+Z
z_Nci>z6M9I3P}t+a1J+fq#(b{+AUJ+hou2FOW~L)p%fW%E<87f3S~&lx0R(UY&|`K
z;ou-LWRGkkpNONrEgp(*J@zX-eqc-<%cFZig}Emm=E3P9Q^>YR^l;0h@U@Nbi5u^3
z_apCWGZ69!^}t*)j1i{23kZb?!h(^TfslkD!N?>62}UMmNHH?0K#GxtGr}>lNc%`<
z)MvDFv^$CvUJ7l~0U@QQ(ifE_dM08}l6#UTDob`x#-J4U6bwprPsN}#_cRPrxvR8j
z;BxqiTpttPVWtu$86bIsnMx2CAV<B)3u$JYYaAR0Kw%zXT3Y~;ct|iv>LJylA*0G?
znu7Xkqtf+l#wu&G_b&Jm74Y}9%J;e*y1k>?;^rr&$J1acBP{`kGz0AhLVL}8g{ESd
z^m7sB<8X7(Wh}akLzju@avr*zjV_hwQjRVYv`Zam`I{?eCv#eU0hRHJgv40+ytvtO
z=He7u4O~3BRAeaAvtSsW5jhSRudAm|m<Sq}8k!gzg~&prVP>LW2eS!bM6x+)lGRjh
zz%;%fFo<p8%Cno|WcyFMmK19$jm}`&J370#aXdVI{QPZvC%b!1x3FY+&yXbH#Zkgg
Ioe1220q)BnDgXcg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/test_wrapper_tensor.cpp.A4AB6B2A22D4689E.idx b/.cache/clangd/index/test_wrapper_tensor.cpp.A4AB6B2A22D4689E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..4577e6c7301da64d80ab6af2962478a043d4ab47
GIT binary patch
literal 14482
zcmeI33p7?)`@qlhI@GJ<lFJ*t$vyNU>4Hc_h)F~#w<2^>D3_$mmo8I6sfbc8r3n!s
z*Kw;vMaWFesHwj0^38Ocnbg1Q|LASM`_N3^zg25}-&%jI)p36IK6~%;>}NmEv!8vq
z*xTC1nG%GxzkO72K!h(z5X4aCe|SV#pb7&ID1wMeD|7R7x?`%)mU$&-H(yfUeDfKx
z3hB$Iok4S?=FgV)$0kMo6e)RqVSQ%!Mwj`mR_@&7w8t;jSw7eD)%U${zpC8J_|f#$
zbMEf^XJpjrwps5kv_+4|^va-hlAfNDc=BwEMQZSfeHVHnO*U+PBXv`|@z1l~<odMq
z%-63IVlpebuKad?O`=MBww)j09yvBW;6%%v>hsT9=SHp*ziZIg*d3U&FZ)P=uEFHJ
z*YX2Aei;;|Y~ppi=J%sTbAu;mE^Fr9rMjPce)orMVCrbi2+eM>9@UCVOCMEw2A^^^
zKOk^?lzMy37~kh!nnRzPF4)bD($JjSHa;_{XLdx}_b;ODLlhkX;&d$}ViKy_3-Y7J
zyV#1~x9^IYn9{1QC$H4D`)TujY2L!Rb+4MU=z<Q3(@PVtk=kpLCsc;Dc3Nqjf4x~U
z%Re#f#)Ax7mxnW+5M^42s_sB!-Mp(3?mmO-OK!y`jJ)ONta$45pr5z-q^r)^CfzY-
zsq13PId76fH=Gve3<>V?vWwb$!bs{))Bd>u{Ps!rPS^^pkKL9|UL)~t6n!Ppyzz+K
zMagZ~6djMbwtMJ3o;@^0@IzvwhHjxEW%B2?$(==cN6$)G*%oG52T7H$yddMcrTAP#
z%w@0UF8Z;1TgUXtH_s(`ZnC#GJ#up^U*S-wM5u$?p^y~@DQheqmDZI??>Hsh<6fk1
zn80hSJudgst&A@g8s(+*%BNFo)hLOpHbxID+GAS}Sm>M`-acfw!j72s;(JxQy?>86
zQ)fCfH*doCd19*D#&jB=*&X*k|FYiRz4>nU@vYmZ9&o>0@@L1vV(<LrcTQTb$Xiik
zsM#nx!!P3-9nUL6hXvGd6+&Geo9~p{dg~nhWBBW)HP4of=n0j(;^kSqVd;|ttyS|6
zuDX@tmMm#!&UIbC<?1}{32wf1Z2ZZ%NmrY0?VIyRH@GJL*ZV$G8g)}OR37h)w4MJv
z_x|i-w^g}oXIhq<jhmV06u;rE-;LPPqVkIhjZQl2Ji=8qL-a;EA00mV&4Rd<cihgc
zQICIh@%!DmziQc>$Zm^^dt-T(|M0u;fFE2gTWRIYTQ^O9dusHLN;WS|e%S4pz2rzW
zZ$^k%&%r41Yp2S^W_=?UUp#Yt>~95{CJ$2A=5W$GT(}-T`a9)4eyCBi=j~R>EnCCi
z#*O~f&E4Yl`uK(hlVHPRGWEBwoBom5?Ize*`?PtDN{r3A<s-AZ-(Aj1FL|YstXdna
zCO>3;mihjJrIPujuJnEr+dYTn(ktGGi><nL=qG>eM8`)-l`&Nv>jDluIse_8+@NXS
zKUL?QAGG9LTu_%r{?)=??iao*wUc-hnZ74boxG9fwpKEzCSKgny1FCt^_$@5WVBuc
zhp$q%{8UlT`Q3ptA5NU}z`bTio`G(PLVcy2e#c$a9^<?&B7Al5qFEe*P<yvN@R^Zr
z8IR-XMbO^@@hy-aeh|GDh54tsGcE40LCQWJXQ2;4?+0Q(Q$_d#Rnbb^QQ_Spm_iXA
z1VQ`D`s=;F2%jei-9uL6PFxMt9>iJ56#r!lV^Zfj9R2f|IFIAKfS~6Q^T_uX)Pc>&
zXLf0R+<;Gv!N%4%FP@4_8^rNsI{C{taTu&|`AAbXpCa6uDm-Q0==T>i%zzEE2VFYX
z*-a61nPNL-JI(hOR2(K2@z(X;6Xj2cYW?@6*BjUSe@hYbnHB`fg3<3UsD-W9<sB_z
zIEfVD#uPg%JL|o_pyGtfx9ZnQoQ(WL2T6^yb^f)osEs1#F)f%Wn;E^opcazuZ10i`
z&R9K&<INO*B*n}l=62EEzcv1ikk6+6yv3rZ+wE_Pe+>~=vP_@+`V(^y7O*(H+=w8=
zT{A*EhWU;Xr)Xt;Wh0T?X}%GEJSLU#DyEnW^$m?g0;KKr?Tyfca83RGP+o8#N=&E}
zi(`EqCsYesk)j4I<u=K<7u`i#sjxQxW$+*QE2NCJ?O&Qa_7{?(EqRtBM6#n@EnKZc
zb3um~g-jGFFdgC&G9QstS#G_q_=;NQHFyMXFp@Gw3kU&=%m_0U*%5XuaweQbB=vBn
zPr0U~4~L?60)Ho{PB~%wt$p*`lN7xR_`5I<)ssAsd-0ep2R<wow17DMSiveQ#4SU-
z^D_m51oUs)S85%*=6vq7OF+>nAe#ci(TG(nw0Y6i`5j6modaYJNTDH<bS{v&Y(5I)
zQ8woRnaAcrAPd=i9LVErE&{TM&BZ_#v$+h&GB%e3S<dDPAS>8h31lUkYk;g_b1jgy
zY_0>cj?MKz*0Z?*$Obl_0rCu+8-Z+Oa}$tFY;FOvh0Pa$yujvGAY0ko24ov%)oz_~
zZsg4=%nbPx{*p*Vky!`M2TO>oBQoYpfzS>k)zWf(ek)>AcnzbR%F~pm<7LUW;m<&X
z76wPMkA0`;1mGvYc%&VuO%Cb=4GXPMRAF@Yq*U4IshScLeH8ddK@F)9MHd0Th|Lus
zT><jAybkzv?D9t7H?p}Eq+4-)!D^fHd)zq7nU(Bq>TT<NJ>WZ))T-MQ9cmbAf!>Ou
z!z{uaFh!U|*kH;6(=4z>0}%8GN_hVae#eL_L@`7SS33lw4#8Md&Cto;&NzQnJhMR6
zRMpgx97)<p%V-obb|h`7XDLAD33)<hhI$P+GU;P^r4=J)U46=Q=sm;R3BN%W%*X;K
z)FGMzW@!suAe#98iZJEj<wu3N8Ig?Q#DP>CD558=xh18$>G51<$lh+=^H4`5Q%x-n
z)KN8(&H}eAn1?z>9{MB25)A;204qb^5!$S^A6wtDg~N;}LNP)QiPY}LH1h)<HsK>F
zT2)(h3=$zj<5ev&85tW{Ab~RW-o$bm5-3G`P4Jq6-h-j(Ub8R-O$eHTx}xZy=|QtF
zg{X%JSY#dIh$$0PGC>EuJw<1NekM%Bd=OL)vda&G{y}zmHppj#3a+06dO0wT&89hE
zg}DIK3SczmLeMG1dPdQOFtHG(;_^~ZD`nR&1D!H<{W6$X#;#utBdb9dKVJjdH86(F
z<7>bYa|7r$fHCG~P-|xQ*8(~%?EYF{Vhiri-ZoU-eIKWUnd`8bVNOV_XgW#BK7koU
z`^x&N<2hT$U#E{Aq3CFSw1^y)u8e=ZX^@o?^}#P8Wurd$CCoy<gjwj9Fbn+>W}#og
zEc8p5g?<UM&@W*Y`X$Ulzl2%nmoN+c5@w-a!YuSln1y}`v(PVL7WyU3LcfGr=$9}H
z{Ss!OU&1W(OPGaz39)0-#o3n)o}A;5bSy~60uS4bc#w_<WyG^`CwLU>mEQ81A<XPu
z1v0inlCFQH2-9V;liHFe>MldRPQXghdBD#DRlGckfnUt#N|3GuIb2>3{CakI6Y!hZ
z+y>HZxW3!(LC%&#0xQ)?+K2R!K`Vo#eMw&y?FDi#$RJ67d8n2TWIo8CL6eNl`LfWh
z1jrJ0dtVyLRRLMWZtqJ&yOThkWViRFq25^_&$8S5ve1vD+ktGy?Rl(D=<!TnWd2di
zhtz|K3D^-u-}qud+q7dDNke)DRI1PJ;QzNuEnMGp_0qMVKNu@p3g)HQfA_LmDR?4V
zK+)AOxtblXoCbr_ed7c3X8ioZ8_5&z*nZ>xmCU#AO#6j~8cDAPly^XRMgD0~OSx6Z
zy)k0F`w6><Z&1L9l1k!|eX@HLy?|UGjZ7s)uOe4Tqemz@ii~2>VIU8KG}3*Pj{2l|
zU1G2+LDSgj)JNN+z_9R0rx!U>nSEU>kg*_z+%-kV0~wEb=}XzJ13Iy$AGLteU;lQ?
z7o3%9ZL-f+^D8a75GC;QX;0hqnUNbxqkiP3!sn?P*=lW2V1}0sf^0BC0`6t2Y_Maq
zM>Z_PTma?;*tk%10k{^xd{n+tX+v<dOyQ%CB)76_P(YK8N#}ss7%&?mvk}#B>+Tw|
zqg9NeJ-MDNix30c7?8%xGXuC8*b1x-FQb;nMNj$6)$otDUv6x?9!W6E|L@aoD6b(Y
z+2=TcqJ3n2Sli;Y!`HIX73=xyMcZPE&Ij3iP)93*qDw%wgw0hTTgB#+AbXO{XF>KX
zo7+LQ9rLEmW%tTCM}KEB?MkDR^wBn+q76nF=%dk7wD~A=eVmoCx3r&*jPa&|nZa|@
zJI*p<NCd4!7>C821UgAD5%Jbq-R_@OtXR(U<gMyG8r$no#ZXojE=)7b5L@poP|d=*
z36%L~l9;_mFlfh0(W}YT3h1GD2eEG-?tEt|$IJ_*ERWxaFUuc>Nku`0k4a!HaKLYC
zNEymwvZw6j(c4nAyPUf+9=oT!=WtAm<rXWWVNf(JPY=fwI5<#-MdJb|VTza?Av(FB
z=&c~P74NbsItJupaQ?p6*$8667>Sjl(?Bi_l(B+jfP4lFN4&lM+4iYPI$gkIE)|V<
z=(V>0UUllRwdy@a>Hn|k&4gLIUbwG1ZaiRf2W+m$=Ax6&k79y^P7ikt?G%bJ#=Tot
zyjF#&W6U4Uqd)v&x0=VF&NCk+4=CAyl8KZIc^<~@_a2;M-Ps8IjbMxnDLQar4zJiM
zCp>yag~_z;%N;8t>LhxTYFas<jSOXPQcR-|G%=TfMj4xHK&=LQPLgf_%?8lI^;<xr
z1=ml0?&G~lX2fM?6SD*OJ3t3n?9^o!kJe6gnDUi^`P2~cq$hF-rCcVcU^0qWIK3G~
zOu~#JW?@DUvoNEFS(rY-EX=537G}gS3o}}X)7EdA7GPxdc;J+o(USk<DKq+Tgt5=;
z3}UYAq?FA={FDaF?0}gSnb~8jy0_=F8!!6IjZi_W<@u44b?O7UG@wf&T`KI1YZH5G
zuf=Q&y|lfI@E$2hJIDw*?BXRi9><fT{kfla7}ozVtN&qC|IB~?!=V0$IeiafiaVbW
zx0+t%Ghepk0za31ly@BX$JtyC(&hMQkD_aVU&}5(1N<{=z5vn}aQzef@VPQ8oTf67
zNH$1k<KrNb$%63<U<8gVN`YU>=4w!`##=m+J`E#I<4+Pux*7P*xPF=0s>3rE?Wdsc
z7O?EZo;cr<PLu^i0TTK)4}Ja7|I43>i_07qc<AjLaP9wL*IsFx991~A>vKOLtXzEa
z!gR^zTTJ-L<MQND#KwFY$rZ53jBCasJFXpzoVm^-s<u!cdUspT4rZZG9z0nNh3XXT
zI@nbXC2VT%|D(?5O^{W;D5SQ_f7!FK5gqSiapfGgMi;`cLQujhpbUnUvAG5mYk-H#
z8$hW6CqYQM1%|cY`nB&YjDH)u|LK2~p=q)UI3p|PHuf{!E&A5CpDZtM_=76lns4@7
z<n3jvUn(8kqr>b%nqgiucp;HpoVQfzMP#}vQ$UlnG*W{fqT3g{3~}t~*ZL0?uFB!j
zYD+(v?EJA{g8AMtt@X^aeZLRu-14)M#LD46eMYkBy1|`AUGFYZAGLMsaBzUs_SG*x
zIg;CEdpF9Rb94Kzn(w<xNA!2ymoa2U{^^g1?>St3V79NWbReJ_2&hB>RhNEM!8n!t
ztoP{zatBTbcaO2WapnHRtgn<#_;tn$r+tB`Jxo8LmZ8&8sD}Q8fLV`yzdnAhlmh7~
z8>{*mj2AL#zlCqlyY!jH?e(R!Hxg|Uw?Ff-uz(=naf}H=A&uzoRz$x*(TwO9DBK~!
z!~DWG_nN=O$OaO`Qhe)&Zx`X)0DS9@Zv*kIFTPzSx&?YCF=<SIUPGYdlBI#md>1eB
z^A8}Fh{_4SVE>@y49*UA6(hwZ1`Xy8p`_%bhsw&>JKH-s%FnRXlb3W<P*&s(=UYsk
zGIg4j_4KjB3@x=M3Dp1Uw@NA_RW&rVM(K{$HyC3)ZoG+^xti$+qX~AylyoM3Gb@PW
LFvn&lxDfvf<~}y=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group.hpp.568EEF4CF01C2640.idx b/.cache/clangd/index/thread_group.hpp.568EEF4CF01C2640.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f86d2531a94fae2458e103b3687e2e54c5fba651
GIT binary patch
literal 682
zcmWIYbaPw8#K7R3;#rZKT9U}Zz`!5~#Kk2=nOlLhGa~~-#hkVMd-)C-@U(o^;5ieq
zCiLn;?j**xcYz*vj(N>;`N^ntWOwi_Uc*fh-^xF|?7MpV`97uP9}a$Ac3H4>S3=I(
z>Cdii__h7g2IgnMlIQ=PvVXoyR@8c#*xGXD*CtD+>iPT3S`~F{y20gnI$;b8{&p;w
z9H4ZOw`qU!lS7==s$L3zC2EAiyYD2FWHip%;;QJutU7tA&B2D5lW)Bg_?Izj`i7d0
zeJw|Coq2klZO&dsC5|#xhT_WHq`5%XD{QazzNYKGR)~p(gMmv(PDm3>FmM3nfI!!%
zduyW*hXNZD8#@CR7dw|2m|);!0{WVPk&Dqgp>p}TgVK&dOsqit0;&SKU;?JUiDA#3
zJ-M6XM47mO(mb*}8ejq@J$a_q7a{K{?m|q=KxsY=J~J=@lV&JNO)LHfv<T)(MlP`H
zV6KAEKw*&6Hq6uBpSo+}Hkg}$!XO91RKeWB#07Q(Oqhp@&xy|$=8|qN=9ahp|1Fq#
zxOhZ)WMF!LiWK<NVTyRTgv5koVUB`{0<8o20A}uvv<GKgxK2F;n#9M##|H{c5a8hw
zU=cv#39$&lc|Wh#n_T>~aUUbl0#hMxm<2_tIf)>x#YM@*;vfc>vkm`u_WdUhF>*3+
If&?G{0L#M5&;S4c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_direct_load.hpp.5F79EA1E492991C0.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_direct_load.hpp.5F79EA1E492991C0.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7a45d33ad545cb5fa7cc44d5356da86b8a7c7577
GIT binary patch
literal 6644
zcmcIpc|4U_AHU~x<#s)|WB1_N%aWTVTUVBB7ons@H4~-Kh)~8v)EkpfsL)uNHes@4
zElVn1Q>5`~v`SNrA<L8zvetWC&imZY{P%vExgXBCU%%gTp5OlaJr55j2Zup<8qL<%
zDPqsgeLi#=jrIll4ciwQ5P^WOD2*18ePe@bu!1S8t30|%F*F`DDR7HBR|Fke44=_B
z9(E3)Ti4cJl=k6NYgTXQQ2BJD@l|w-;Y-^GiF$-z&;3ZFa$mzETXZWr&k5a>W|K$7
zm;nYK&LwGC@9y1mDl(vAL9mMHD4Y38D6p{LoaZh*PqXV@OV2m7JelcBDwi>&xBJ}v
z&0D?bee%&QOLO1*cPYkl^5r_(bFKVs-2AS}^<UgB>i?UKx{Q?5=*i%?I4Snw*Ztqd
zp5l1BgeDAjd}JJ+Ai^I6yZ&YiYzTb?mc4uKvc-F+e;FKN|B-#ramDQoW^2k$eswi)
zqfrFV8rC!4tl@g)@;Z2v!LdJds>YL%*rsgBxjO?NMJB)9&p5vJgv+S|LtV!rHN>9X
ziRJ1Z3w1km#`LQf*P7F9PM%P@^*;Z*xADopt6wwr8!8K2ZCiB4Q}15UlAzO@4hZkc
zbILw$8`B_ay5<=7WL>G8qmx!s%Xs*$28MrM)h4^94P4cO#WTAvD6R7<9}_Y<d%M5B
zziX9h$H_FEcTeQb1uraUd@3gys<}1r*6x7!LLu8mS-0DdF%Lhh+-#=ozAc5lMxnQR
zvrgZKkT=H0E5w8bdOz&<HxKk=#nnm~MqRjV7_NFnoAK~vz@SdYiG--G_3t%;Q~n+P
zy@=-Eul_sk>Z}fPa{V)8k8yUTllY89Zb<fe(}$ULPRi`1GNrxGZ~tcH`aJH)x6`7W
z7sEq=Nuowu3an$Y4>O;2l>FQq5~3{gKsEb@rNlSk%@-#kIwlsqkIa+`srgQu{dK|Z
zr`cBeaqqvJJk-_JP~qCu|L3!nRlzI8+#RMREPoYJ`i`;WQ`4IM-<#^<I+M${ew8Fy
z`sS_1zL9j1h`h|9ht)P;P8KQbE6{t?u>JPG$;}Vfbe{#p>Eq}0hU6ZnT4ptmmY4W#
z^x3!Hn_TzF$8}vyR%3ZdZ)5wGox`=cC(Wx;28a6f!#B%JZ6z%Ks4tz!yZyZC?EQ>&
z!TZd!V$L4;re|C6CC-pbk58RN)X3K{%U`UrToGv=>+09rnbkdch760?8sGlZK-A%m
z-b`xl&ghr6gJMxnLX-NvmC6+-+$L@~MPG|3Y;j-Jtae`UT~cXdvSwF$^HfIjR0?O}
z-gEWCMiXtZ-r>!Tmr7QC>mGRAE&a&ok+fW27uim)Kz4hLa*~P1O!}jrH;MI~)$po&
z6a40EOil0TrBjY%L|yQ~+DXpn{k}-UKbk6p#VfW6cee_?S#5t><7REmYb*C%wjW1w
zTZYeV`pIgCL*tLTnjLiNoPXK(wa)5mZc+`+Qx%spj@_r<*fi!=v)okrOaJXf`n9%K
z*5-Qj{-~{g|D8wMxR(u4RM@DO*|Rq5#je4=s^S>p(c!+BxRxrO?9*=?mY+^lRuAwp
z)poZbOHvDonV&b`d~)aQ`IAi>17u!QSe3Coc^|Ugee~n@?cF=!c6VfwSeEzFQ{Gi4
z@xuCQJEe2_BFn-u{5R*s26YG3H@5XfU28eMWcYAtXYnJ&ZJa`1nc}V&mR?Oi-`zhg
zJSkeUChXfi+dUW1Xex@v`mOpM;u4T9MI)Uxoo)CFje%+u`WQsk4wdxmIW7&^h-|E6
ztjb?7dHDk|-8Z*vOGTg<BAYSIB=`#^TV_?LY7Oh`5Q8#^Tm_j`P+@-iA4`K-z%LKG
zm8GF7Vkw2_hcmxn7Mr0O$w{f!A4Q-LVlii$OYj#=UiG+rsg(t9R2m8+vL$J$$X_to
zWoKQF{?(~?5x5YMEtnQE`~{QO$J&dqHJ+--z%LLvN<L~8f5Bwm`!s9A4QT_?P!o|I
zNyo+f1(Wy6hj|A|Bs5AxHAJ={ZS?sICP(Q+{&eDprMFm66p;&oR3Xru-!RL;fi2Va
zHxoaIL1n~J4n@nM?EHpVk_Oft)a_o9!-522DFaev=#;q)vm}SVOMOr~zJvvth~+wv
zx(=4iZ<r<HlP$OZ;I<?dT!dIMp;RVZJilR<Vr!9sHSNKtMIZyQm@rMm`3okO^2n^j
zsXuGQpb8>aLS`lM=G=x^%A_0jYOACkV?iavavn;Zhida1W~r8oakD!m(Z_-!h@~D#
z)dSi24YM@3%DS%q*U^<CP#m#vnOqtEg30$)J%T3!IVZ%RG$L0((F$a;xec?t(kNV@
zYku)}X{do%cqEU*Uod&Vt?{Z*zE*`alt5%_(przdVDi+KQC~Up%q`MT9g%HGTLb=r
z$sYsXR=7Ltsa^<Kk~Gpt*hrth&=xI#fJOr(2;I3W>>O(Hbn6sF+l2L{bb?HPatV-w
zbtlL~D3>URlc8L)AWngDDHImoZ+@&b++P_Gq#esn3(F^}AD_89&7gAuL0T$XYT`sd
zOfGh7F%v-+L81ssVAqJtb*$3=<;B-RbTS`s^MM7<06JL!xCMf^52*G5HSw9(yN$$C
zed**hSTPN_6wZd6Y`E-yf%Vh$e@-TM10-aSZjc!+EOL)a8Pj>!4C!>zk>}`y6N*k=
z29}qB6?OM%pf(NEsnheJWj?gRV&mf=lV6{(4+!!HK>Pr-@G)vu71hIL|6zXS-Sm8?
zxk~V0gD{;;0}5$C37=0V(}8k2P^EANP|gs<jlif87-LV<$tGabB#8TgeLrv%R8uz}
z>R~l;y_ID0>nx?F$ubiooS(?S9(Efj2m8wSlGr{3>1XX{N1*_X04)lI%7hAV#skH8
zpn_E)kg76ifTVCbK-#h?oB_xTLEH`Wx`Bab*W`fPGkIBp>;dc^DqaY(4=nBjI#^YL
z907VGg8buvJub)}2gx|d#-=04Jjl+2YS=Whfh<`BWid9W9w~O;a9#xxWH(@SgT<5}
zBY-tRVS}Et#R}d{!GKQk?0F8jHsIccB5JmAQKl}3^m0hx!fecW^?Vbde+GqhDKsdB
zJj?lxWv02E<|Dps|3W(HXYc2Ljf{|!pEHF5v;%YmkX`_XLcL&9FYv+!qLcmLtA4Nz
zW3yEmtI8);#s4BVH#puew=X^f>D>rS8i504uqI&AL}A>0QF*8Yj7lJn8wo+y0FxTv
zh`R=Jz9(Z7U=7iQnZPs?xKJT*2@sd4kYUbuYJ4H?)X1c>Z5scXK-HOE0SR2VaiLnk
z<acU{Z44-n0adI&K~4bk3Bc9Pl^nS;*5qO5p>Qc`jA4g7aS51Q0uER#Cx;MKuQRX=
z5t;z63AkX3;NB0I{GN|*v^MAwPu+-Eg7_Z;#$zCf`SF1O@`nJ-ziH}>K~L?G<$rt0
z+-i{}c|qt+&|f@^Yb3JI?8Je^AV{7gZygR_g3O23`EWH(hd?<tx%ggy0#eFE!-Rut
zMPNgoc2>^lmVYct@d$*3ot=LuVI2(|9kC{;Drx&!;lRgZ5sKh^z{m&k*zU~v;Rqjs
zb5NKdivgn;NZ>08@(Iv*0ytEtbb$?Bz?;H}&^!^^V#g9>3N%j<#5s_W10`_pV$P3T
z_{OmpV-C3HH{C%Qa0H+qbu1x4HUm~Okf#oA1*}#ICps^P{UiC48M@F|+E@{nbJP+g
zg5<IOvw<x&0Tggx<Ge)nC;*HCL3_-Nc~~r*uqeOgW;{$F$PCEHfC^Zw)CX^d4%bE&
z|6^XqhC`V(H%ekjvU7A_M>Xs>{o8ZpLUdwstC*?%Kc-SKd%T%mza7mJr0sHBd)!`e
zB8f8j!!Xv9AkD1J>~RkvNONm*dtCkqlIz5E!&59lS~yv_;Zj79(PGh(R5%?1VuyeL
zE(G?4z?q8kGGJc@oN<Jpf(#upp$j%YLFPhME|gZg<uK%S>;4{8=^?!mF2bWafhs$_
z3NFH}pd|F@?(bK6jvyV4HH=Mgcw%-Tf-C`y5+G0AKN~W#p*(fJY^azG^(dSJSLQ%B
zjG6P#1K3KqIH53^dl<lk%7FuwvI)aKe&(EZeb>uYJ7&u(20=QDI7`nxv(jh;8AXp0
zqfjBB7XkrX2Iysi_&T6p7sT~|UN49<Aw3fcTz?+Y&r=vD9a3`caRw_{YoYo@cl_9K
zB%LqASAt6aP)4W#>H(}CKvJcv4@mU^HigFlYg~{%1*E0~>2Z*VgECl(*|CY02T5%D
zy1Ng(in#B8|4d6aEQvNSt#Ho$$I=g{3$o(u6CbwWod@mr?)8t{`rGRX9!3_T=>#w;
zfCQCeHGokAByh1M$frQ#sbH^3faVE;y(Af$CsUX?|163%XtP)x<>9WF{+SMFf3$z2
zuF6j_s28RIb{bep)sS?cmJakO+ze2c(xL*Q8Q3)g4+^&evQ?1Y3hY`1>AwQKUjY~M
zKU_B0S-#%;6sj2)q4-6pgd^+8NVbRnnS`T%yK6eLrAG#L7ch`(%-Ckj1&S=&d^uHQ
zi-BS>Fr-RV1yHO2hJtRCt^vA~8*`vy4m8B23eQn=CjZp{_7*_~hzF=ssTw98CV<+&
z(l%g?E%VB(&Gti~9QaJn4)Dg^&jfrwh78;Yc#XgfpO0Ttpmsm|j)D@AXTY<=^8dLb
z+GI1QJ_%jj1z26c1s9It?CP7#uBvK$rgUT9IbGLiDI5LEjEdVM8vABnkWjKYIy?GM
zW8Y<9eHm=Tp*hu&xc#*+xdu%~Ts7`Wobsqcs#&b0I^-27dj%TeFq>i9ePp=pRs6p$
z!tx-~&Ay<ydRF)5&YILJ)ELkkK1>YV8R#?HB*H>>ggMTBk(*ukA4pov^jWJ6PahFd
Zl#`cLpvhtijix|-RiwV6|MB3@e*w3;CanMf

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v4r1.hpp.CC30E09EE6FAA0EA.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v4r1.hpp.CC30E09EE6FAA0EA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f4738a483ea5463f95e028212cb037ab111f043f
GIT binary patch
literal 3890
zcmb7Hc{r478-JfMw((;0j-@eTEFH|GA?q>XGK+9U#Alnzk|lLY#@dHNt1}@al=YNE
z$x&Y^l}m9d;Uq08TPTjU)A{6R(U<QvdCXk>XRd3W=l#unzxQ+Bzk7M#@pk3&O;r(u
zv)MK2r|`HCB7z`l_*WPgyTwHYL99^(Nh)nF+~C>IqU;wAZW>WCvsP4d`;<kd%2?8~
z<Bx9>@au*51LNPcC-nJ7&88;gkDs|aVfcg>^s@8=tz|lS;CJmx&-Jc!e#KF}gX+HN
z`gP{_mk8DO(9nAiN=+{iYiov$msGA}YGI6y$Ti)6TyhT7yGF5e7OknfLo>=%XyW_S
zqH<A5>1oBcuxu-%;&cBv92(Z-KK2eZ7<H4$$<6FhZcR4UxPPJ0`o}*)x60_~Xc<_C
zI;;kM&K_DufBR6Ci%zuCG^r~PZ^IjVGq8r~xQ!Y5ER`*%_r0{;)f`4<M|^3!eZ^eu
zR$W&npZ2ekYOcakuN<r52a{$WHtx^Zc<TC2!Wd~&Rr~#CGg0rxYCP5&HihLPk9)7%
z$_nGgGlfrU3I|OG8TKlFed44&yc7}T`&UlnFK0UsKMx%_d?jr4tCTaFGP<V(W4j;J
zztk=~dZWF1yv`*|{!S#TP_12%G}KpZRdirwVE^gXwzkFrv^G<<|Fh}P>YG8w1Q@~X
z(QDOy`3Y~D{6lZ8=>E|k$ZdP`)SC{QMvCm49z+cOKAbg{n7glLU8&WXGSU0D%diz&
zh##(*Tz#_rMXrvH+<LmdU7m7t^?wbF@7buAK0bG}V5G^IdPUS_k;?NMI;r$3b&A@p
z%v0@{iP98lr#2$nUkKhzpDcOOXcL-!Av;mus9I;O{rP<nE`xWPV(Lz;B?tAX1fLZR
z2~-nuM(c+w)(Jai=x2&%3ypX^RY{eC<WB}||0cUP^uBphk?mhMnLFrVusQH)_lep|
zE#3Lie=o1JIe)HdRY$Yjt-CF=VRi#N;nOQB&6g@Qzb+qDIX~z$yYD&anW&G{EAo;*
zrlQ|(*r*tn_ose+(S?nz2?w?&);A_cvp;w4I>g+qH`^C7!C$&wF5}&dFfZNF&$d-$
z_CKGxETXWhS7cg|X{b)}T)I~&msad)7ZN~R9q{qy4CivPiMi#+s^TdBmKs{ugQ?Ws
zS9`y{_8y@XHxYJRJ^n;s`g38#*M~z!ElJsEWI}OB?fQh$U$@_V?~qN==+Jc+R2=Fl
z`P6P1?(^GOW3>UAbxudy`1tg$v0p7ilx%LaQW#&Ou2sLx_I#&$V?$OpB24}%v<o1J
z?9xa6GCL+6$OHvd1dE}>Xp0v_0q;5dL%LOWWYcx;73u^{TznExPXe>~4XRK)&TgCJ
zV;W?F9IlcNDEVOd{03DF<p=#&?z7mSPSC<tSX7poc!A;-74ohY13z1n39`7j1W-zF
z3*R=V!fH6!I@+WAhD?yhRWbo36KKwFP{p~^ai`p1-$Nw=5m&JwS*VH^DE4`ol3|$C
zlS?Kj;^Gua3QfE~aoTWsBsEm`nF?VkE^Y>z&44k#L6yTw<@vj_p1Ua#lyQ|BK&k<>
z`3<TZchh{HaHl_0gMi{HmP;+!;suJU#_wgl>j*JYCurbeGaWO!c!A=M7x_b+kFE}K
z1SJ&&W6Cp4#S5Y+L%=fyFfh0|*tPq9zABkO!tW2J1RIGLcz<kogs@Oil2bqu=MtKQ
zxg?{QhmD6j^cBUtm|jj2<ZbNDmLMOdkCZHw8R`U`K(R36Ft!ARGsC51am+X;SX(F-
zZye8-pme}Y2Tm{{D3$??GXPtH3jnhKIKjF?u_M6v2w+Qa8DN$HCs;KoRt}8I0UKh{
ze0{>Ol@;?kvnPCh0Vq}k=tW=^G^EVRJQjV#yk8c@VrjAZ68~b?#Ck|jnogR5#J4op
zG*1c20$N#M87xj!t%uI62SZzBh*+Ro-~z!!OyDMvLR;Oox<lHq$Br@Sk+=PE=k@^l
z9$*4(LnrYZerqWhj@ILI3SUlLvcL~<=ehG?rihrIvmZx-qMV~R@DP2+3r&-@_ehJl
z1$L|~N!+RMa1?WAxpUzuDCWoY^Ov9~Zj?Vf6UAzQWi40(o5u8F@YH$Fx9NDGiS$Hc
zNbicxU+i@VsjVmx<7jhqpe-WC)#cJ9h^NbwLVR7m6mrpZkw#n>DNejhyciNIp(ddW
z=pPY_Gl^qBH;7m|Fi8gt$UYG(044>10kc5F%7953V89xKJD#0ya|#j!O92;r8&XhU
zK3-n09ei^KAYv>Yix2s=G8-B2M_I)<E!9Aw8tB3%wRJ9X{?BK+>Ua7-9x4<$e%+~u
z>(>H>T3`zKh0{-tB%XTkSQN8XvBn?;C}yW(CxwD}!ETTc6t6RWK5&O2qj+Te65tL)
z<$7yZrA#Yu2`J`E_SKV^j37toK_ieqJp8!{3%&54DK;O!y2o-6jUix{V8>-AwtpA@
zHmzmU<(*r!@%Sy4SkPe9aC$4Em_5mUnZ&sOG(gJLAT&q{MWfLRj^kB{ChEa*m}^FK
z572|*%uNoo2<S<;sRU>x(1YBN=ARTW-h*y3d*2?KwZ*+OQ#Yf*u@7@aAk9BEpgxM<
zPizJnFpSbN#%M}wyv-u`Dlg0V)BaHd4L`CH&?~_z=o{n$5X6rncqH_RL=r!O;LkX%
zIu>+DD~hqL*mlqYiaA<2N+FI0hbobAsaz@Ec~qVh;#2uj$c5@6jT~JT#@27K2cnNU
z_)N|5W@u&&eZ2U3*F{ay<mpAsugej`U0hPfh&+l}5-nB4?=tv;V!^~<DHKDDkwPiN
z6e*Mmh?zhN=L2HCG%f+e5(&Q6Jm%Y7(2L;h5kL>Hknj>skG6pAajpGPz*+nEV~Ysn
z_WhpD&R0{{7f}i@mw1`Yy+bsVGg`v0<Tc_<;O`?)*v(#^;_XQ*T!bq<>ot+s=V_C8
zc7o_Z3nBLO#``h!ZvBi$&RND`N+Rd#b6F5QnJV>;Di$4BL>Q1Xe3s0;=|IOvY`Zxh
z>a@Tj9Dh3___(%YzYXrVJL<kL|0YYDT2?mD4Hsc)dcW&s!;bU+d^eS5CVG%3JGD;X
z6tH#KmN00%VRRiAP;B_NkhwY(#)b)9<}T*`*3N$@@B$%V+gF-YWnGT1msM9=LZ%?(
N1reeyxm+T-{696PtOEc5

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r1.hpp.F8F9EEB58869763C.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r1.hpp.F8F9EEB58869763C.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5ef0ff2ddb61953e2549ad82ae694bef0dfa8ef9
GIT binary patch
literal 2992
zcmaJ@YgAKL7QVTJN5X|0UV;gb1Y!geVhRZa2@ey95JEdfG01Dc01=H30ysjWl*dA-
ztsS9Q6oj>^70OH<sirFE@`06#BE!=M6~(ERDIydsQb9YH)RScVk*vke{mysp-rwH)
zoE_vb5=pZy1c@_Zc5lnfOUFYH<beI<=4rAo!VpwSfS}#=7i-pjFw7-&&fIF6vOw$B
zAFP3gy~@g<xM;EL@^_s-<a0ia?>wMxO5snmPi{I}Ns$l5C|A;6Pd2vCyuQ6XZ|_`g
z{oQ!WszVVCL_6O6@%WV2<bxfJohf6I2tps{eDOSyr%R6c`7q<`w7v8z#hrn^z+K+s
zRmqg`l9yo*V!n93cBfV2>f7q{<Y#9deqYo4sijBZ?OnS$x)U0!m#Zu$KA9{P9pUc@
z{$0FWsrqIiXV|VhxL{%X@W?M$E}twYxZ&p3Kj=wao}4x1t^0|>{Rgj5o7QLLxzuHS
z5pKdGOj{+ZaU8D>9j%nvlPQIe#jMo#Xikh@^3&aq_uWT}w5wh?>}nrBwtq}pF<WuL
zrZ93g-JL=B)oPk_FJ+XprS{$o@;Se&!8-dlVQA+5A1j2)p_tG_?`cK7)kHfXCGu?C
zy|MTGpp@-@f7t4z%X1teEL)%vxnadTX-LNjd7GgBKYN5fCyjG2SklM0UF*pBD|BC@
z{#7@Sn6O8Jt5eBlJ6ppElj^KfIcwAnjORU`PEo4cmW&*C*Osz#T;8$rm5-%<y`p6<
ztd6)_>o@v}o8Z3S%>=KfRh6qsYrB=>RZUCnBWGH!kOIPcHmrNq+}v{K>%r<bTfaU7
z6}P|K!D{Ia>^$9aph?81clsaYbNh8YJi$83w6<HKwH+_&zJ75pRrjQ1I9|D;=308c
zI#gY$V1H{Fl!3dSwx6*sE&62HUt&2byi$`@5nsk}?y+$^cF5&;-X(`J%DRp#a}`?t
zoLiD`*~o@g4NA%A4*AHlGDH^erA8CgT&?-4k_sIj3T~KvG*^ND+s?5gaoVl&jsDin
z6;_|G_TPMD#+&g#;vsue&yAUSeyL#iNhUFAkbgSzs;0VH*AsPSq+x2q2}*Cj)0@yw
zL~V5?y*v79K6?Y@7Hr#;go7Zs`Wa)SD@|bw6Ko+gkC^AKH;@f>(y<SoSiw$KsCSTH
zON>r+PG#y1pgo>v(e`dgJ!J)36CsplO<SQi5CO(HK@bkbd5+YTM>e^B<N~{3?=#2@
zj^2Rx1;tI*ip0KSF0eC3=a6$adIRVX_<B;^eA`YJ*b$?>JiO?71L%#*1L|J&c0aO*
zmtb@W&b0)`diw>oY(DJQcEkHajxB6~ZDA6bsNMj&>c-kE)>!jA3AVy$xwG6=Zvfp`
zQ+hItTzilN<1t$0tYYd7pdUI+vNtU)3|j)*V6@LtAEDj=8q#ED<=R;rYYY?hwFN~(
zFs=9!+!Bxms5OqLuQAw$x($f9{UcjpJQ^#FjRisRXq+_8fZXo1+AK7k^~W{|BZRAs
zo1#V0Vt{~0OvhrYPBpMDWwsss4$-RyB<L8PsP7b5Mo=M37!EukXmn(Bf)N$+3ZubE
z@(Mq5ep>KvqFE_^84DpE@Sdv#Jj$c+Jb))W%BS+_MifX5G(kbsAQL2@3QUm5SHv>D
zr|?y<KyUGAnr|8l6c>+X`)0F1W8Zt)aFUYt{SF)+WkfJmgWUO?(H|2KZY>T$H8vV5
zIHdy77oyXz=;%AU0d0$*;X&bsmLRBzEs6k|5LE0V78+3$JIaJpuoV$NFM=lcBnXWt
zjhz+&)FWuRPrA^Evf0@YAQuGP;<H5v5K(`TK=;5!A`tZ#2rMHg(}PI|cLYH>d`_@Y
zBA3rKLA)hAvXSGH`JP4;NDeeXLF6D4Bp?e6C{j*2uAQ}s#Of*`Noc@<<o<!UX^qYe
zTUOK60U(97#kU?_KyEHBE9uGrkafa*?R!t#OL=DR1GrPx2<k`lGYCLXDIzt*MNk<c
zGeL<+BF$KLC8DH(n3ym`O#{M+Z-)#hW3sT$=}XJ8?eClmP*+Q$e&&GP2<j{LjRv_R
zC`-&TL8+3|SYuR`L=_97s@*)v%$wO1tQ<^7raSPyUusb{@!1~&Y}w1j%N=wY<be&7
z#jys0*B@m))?UpIGF!Z(&VNjN$X-){xs=kS{y-`ag1KBAg}~*%H2bz34qg9ZQ2@g$
z5D3tX7wTL;o^CNa2F_(Yx)n+68pUHX2=8P6=Wh-KA*dWLH$kcRR1>7as|@JE4}%-O
z37L6fHcDq-@6DUPytr&O>8?WcJh{Vh081m^CVvH(W<5ftR{nweA+rH`hw>FH+VvtB
zL3y4$HYgZ^^1b+6fCgN(t-lYhOvVDp=rTi9@4l#)Z?64#pILFh2M+$fSv@M~5x@MW
z8_uk_iAMhW?azL?j>S!+Co;gnn3zwVi>oUS!j^@I&<RACu0MZi<taMDtjpQ*m~)J-
zZ-0B|z|W8MLU(iAny@l(s2m<}iJ)96*MP8JMRodOqjNRN+%=02c8liE4+%VAXTb0X
dqhTldvmgAoX3EmR&K@D@{|kT18V&*k{SR=T1w{Y=

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r1r2.hpp.34BED592A25E7D4A.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r1r2.hpp.34BED592A25E7D4A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..97ce4e83639138f50e714f50e60dbdcbb7badffc
GIT binary patch
literal 3768
zcmbtWc~}!!67Scv9I=Hw2v-0}#0Z#hCj=#qF&ttd<BB>Qq7V+{iVg-5Bw{{L1l(C*
zL<J8-j$zhaTn|+6h7nw^K}K916<r($yk}LskWaUXJUZjQ^!KH@fB99{tE#T5_rm4`
z2OluSuz-X)S*pYgC4pfWq<-p*RmuBt3|nA=VOhIQl`l!X?a3-w*AugM^U^ft<7tnM
zUn!R*pSMZNXzIyKX1%{@RmO9=a6$I#`g<>q-Jib1Fm<JQR%hz&lKzmp?_UHkZ>hld
zebVPX6NDe@7DWzKmHoEqJ#)~7W0Wkub2b$0h}xg<ERwA9Z?)Bc%3GJa4_<rG6<#B>
zYI>4=;=KCrqOGoX9@=E6Bx>{I%uu(+-`C`8RB0Y(?yb&O9vM>J=(*5pH@p8mFFy8i
z>2zb6d;B}g$d?8ci`X+lWA7S0x3OHf-k`^C$NwfpPvza}cz%$(Mcc7<MXUBynEZTn
z*!b-xzM93p1{1z+_^K`C?YfNK-SgnCGUwiD%I!Oy=5cRZXC}8ei>vMH6NN81o645g
zJ@u?Cn=Ld<KBo-+Sah~--KVD>Cp#BK{qc_|gX529eLN@c{$ZPE3BT9quEWirvud+1
zrq;hYQf%;IV%_}XrkU3RpZGXDe6QU9YWK#9yZK9ou66Ix_V_&SdJ~va8lP)=Z1zdh
z*{l8PJ0f5Aw9Gil_-U}edRmRR7nj|8>$5d~k1cz_Pd9S+`X<$UU)<Q+v3#mW`CbL+
zVl)xG>I;Fk`JLaKZY(K@voGok5^45`ES`z#=N+oas)}}c6tp17*7#2Ir=9ax`#)_6
zdgQTjpG<VcrSaR~#?FOOrxS{SX^Q0Hq)+=AbBg5hOXH5+zF5vLGH6Km|F+2b%3)h^
z!h~C~Ws|>&4XN4i7d!U_d0v(5ILp1M?eTftKPK8<;D)bNG<)&v$2mVZ(b|72Zco<o
z_M^Fjoc&7;214${vn{rk@@1*6^LGfHIG4Ozb4&W1&$SdS*skg-pR$$5*XAWjZ71jN
zYhG5EaC2X6`#m!<Km(SZn)q`fLvfvVII!&p|1Gae3T~_@&GD)II_+Ii-=cx&(@SK3
zn6;MVey7e>#nl50V{c(azIhx@SmM?bFzCc~a@PsWn5t~*^7?DagS4e>#fG>Ei9tsr
zM}baYCOELfFaW{)-Gd>{w|n1O;w<WYt^=2=6X^YguQ=?%UrZ{QxFJP{vBIo%0?~`(
zR{pI+6>e*VGbq~G*7+-)7^SaX!D`L@pN_{3D4NIMndt<gbG5(v)W@(i<MHtny#+9~
z0Q->}Qd!?&-0wY9AIZcCN+pSvG({&6z18HwnhoS{6-?ZiqSIOFJe@#v-Qdm@9{xAZ
zGVyT~9m9&5q7#U|=g=G)8>=X1;zks$1uQM#jogq*|C6e1O+U8hS>jfdN-1EM0{4*{
zQhCOxJ7`two@t8PP%72Hx*G6DZb$`NmAFiOo0>1CZi4CEd)tIK3FmU>dZYA_uo!TQ
zfj4r<%%)rBdF^g0K|qCEp+IT`43&rKA&p$4Kp7z5W+2}T6v%|ljf2KR4adAG6B4Op
z7HuL#77{?CY*}^yZK4p!3PAueflAAep(`m=N+gu<B{Pu;5{3kZM9?TZFgpSnBw-;4
zECdnAAPFmhZzYhU?>=c}Q_D-gtBf)D=$0_#?|`ps8&5!?ozMaKK|qnCh)W}1M_)Z8
zb`<L&KSw`3By*MV>G#52!}+KM2&i;b@=-|<FvT^6kIIvPYF9NM#ZACG;F<^cC~lNF
zoS|#J16BMpiO0*1lDb|&q*q_ne!JNAm4JXwQm0uc`6Bn;I}xO31t4J=FfId*s3QGj
z+hvCbxn^JJ2Zq_zY2FwI8jvuU8O%YUkZ=ui4F@?$!UDi70375X8UiK^UGpQCNXT>G
zaTkv?F$vv8?$c<4o+3{@B(xH;X->ozO{9@8+gA^X*<wB9$MzdVDARzU8#2g^hz}KB
zeJApTW9SCt3&8)Ba}=5pLpPVuf+Tb`b+th6kdSZ6*F!PEF$z=#BuolUQlKgzVR}%y
zoKCV91ZjaBB{^EM)1qeF!xdkAur!;A>?xO$?WhmTGw0bOmr##THS>8e(#}-_q8eDC
zv4a+;k~TaZky^pkqpuFtKJlT<Moo=!L#f>?H`ws3<dYxuH`kJDk3Kx|nKCjwM3IqQ
zyZQGuZL7s&$b1#Pm?g~c;!z8#xGEpyElMD@;P6;R3;t`>rj*JV3kQbXo6Pk^?j3ou
z_K0=y%Y9=Q6wS;2dpVGCvDc95I>MPc27bOs7)FHYp%@}Y4<!*vdMKSp*F#!BXu&8x
z`S0tIjsCBGAA^VOZ%(htYnt9b#a<4K%Yguu!1-UgE?s&&aCQt|;iCM=oXW1WwUq4y
zZh{ZWey3!LCTyPlwlR3!^`F(@d``!DYDW`J6bk635ls?#B6M?LS6EU-@uo3EfVzk(
zn9mzEDw$rt&VaZ7v;kv?0gY5h!kOZkqZ=#<WdfNL#X!OUmjDTkf&@W&T(}@yih7KM
z5iSuD8Yu-zDatqr<6Pn-G)fVqNKsFaFx4ehLL;?6Ek%t<!VH%T35_xZnR;9v5aa<V
zN)ZY3flEG+(0C&dY@~U{$N>gr5KR&j61p+mMyD|e<)mD11DQ+C)k9Ha6o(EmflT1g
zL8(X;hYll$%+cF;iUC<nZ#;~VgVLxgAEMud?ko^fbt)kxqeJ)kcUR*N-#6~2icHRx
z>m8eNxH;Y^DKB@mobfooojis@cvCQ^dGf_WufG`4A0BpLH`8qzsO8-q-Gy`wcshEH
zBI=BnrW_tB>Q(XT*~1TJ!{YO0jvlBZ;OHX`TdcRGE@XNu$61&$P0cZjQ3_+ykN*e3
C`7Qzg

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r2.hpp.B8541658FFD968DD.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r2.hpp.B8541658FFD968DD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..b6bea96007c45be989b2d6ba1029f20f7fbc4211
GIT binary patch
literal 3398
zcmb7HeK?d^8-Je1Q%%!@ldq<U82K7AVa$xN8O)AZ%t%Q->NTa1BC>0+QzkCyVw0<{
zy+W2$iaxLkMQLqBy%j2b*IVq{s<tnYYE$j|n4Fn;`@?nhT-Tgue&=_e``jPrKKD$p
z)YsQbA3=O#q{_q<Da&yLL8g))MM`p9FNPpXsR*JhIUTxW{#_T7PHxK|UIVq#P#>yE
z7{dl%C8kGV$31;(6g`VlpPf&;6T0f3^<93(_WKX|M>|)iylArg?a{4QP6m-1cHNF~
z@Xs44aj>}(Zaeg>n_>`Ym-x6KI<K{AS4S>k%9)$DLm@n6{I*Ee(8S2$9gWyt`F(ld
zbk)HNrS*UAb}xUY-`;9-GQadqWn+Jn+d+9(%A`j!rnT9-J=P(%uP(*yf4qRRXj9pb
z1AET2)Hhl*Wo4HywYXMK#ZQX^mtC#C)_cyN>ncmO;Az_Z-k1)_#ugtv@9p$M>`H2h
zTb8}deZhkTq4|!&(#``9huwBRH%rMsOdFocFJ8?IJ&SXcrg!<C{X3)1+}={v+gUZV
zmz{ooiL~&3`?jOL0~M9NwOdO|tZRNYSDu%gYOTq<ccCHmWYFd8__o(WRf8X!okH%m
zeK)6rc7_#DAJk3EQkLnbz4~$bvTZF9?*m@e6!n}cWYt#%?~P1LX>!^dRg1E}=R~wC
z<__hH1450CKX1Dp68NT1o^na>?3<g5m-y(9Jk}qnuZiots@&h;`}C&{%KDEM|A~C^
z(csX7y@Q{y@JmsfBaZtO)-C?KF4wMLI(G~A!CT{9`^r`sP3;WJu)A3x#`|V;*3Vy2
zEDh{8Z?kpGy0>z}pio@qw0@zP;jB}__G4oIvamc}QOJL7E_zjj8@~Uwx)pCqecAHd
z;l$fBkugo)kB(7x{wrs*_)=~E6LHt&#&sDFt*%psBl{Y+mdz66bL*J?hfLCL+Z-;O
z{VetM+*V#yP@IRg#E5@(hoUJq3U#wHqvzwtcCfJ{G!8GT{zC=YGv{3PGQzw5qa!U{
zmR|RFh02-ZxYX$4)EjL!*z`93Hlw@)8{$!y?&$P8+5e`!n0sCBuNQEZFu6ma=6yF3
zFTK33zgJ1Td6Dt{UhIOnh(jfG(0D)ga{tNnhW9V3sV9~-tY3Qir`<=o!oW-C&el6`
z6>AftHlPS%c6xYY%F%F%DP~L{z}CW+p&k$lSqtQ2eImwi_{hH5rkE*7=b-c))M@Mm
zMIxGGCCxv(1e#)Iq=-G;-bp<`dhHD)flX%(O~fcu5HM@fEJyW#OvX?X1VI6Mq<nzm
z{6N3}*hF%@!bD-O9^m@MstM(LDmH93!zPn7YdTA)9w6P~urhq-2mTrZY#K=yqtl8}
z*4PV*v@s`_e!5yP2(SsHNW4kBy?TIjN7UWmsOmW@0fv*b%0y+S9w6OgF5imZk}9TP
zRFd9=Qa7RIV=s+J`3DK@Br^?QlSmP{iF}%Rfb^43=`p=~_{LK(J(6ZpnTF~C(#b1g
z6>>7%hO}#9m!A8p^>DC|yO0O{#z81IR0oC2!UG@&lQ)H`P8tj~0hks{2HchaTt4?3
zxP$<FK3@k3_yQf|Vdg>CN($-1nOfvY_tZgNbT1v`P50J8BDzR}Xg7~iK2)A?#c&`I
zNxp^Za1bO8@`1?OyS{1A{W5iYp|PW=PXIrGpC$$bkg;VVsEGgp4go?f3S<ZBaMA2&
zQO{Vi2oU2CBh;b<c7h0=6apkVBnq`Cm7OYr$CLnR4rxLy%3x=R;1MD~rbDI>BI;P7
zz_iSc`QTxJt#6%z1CJRV7SKo>2rY#SE%LPV)InaBUOLFz(pv|KEJZp<#+0$NTB4cJ
zEa)5#5|{}r=nf82nW-%301h&k87ycrr+;_NjAiXfC=TpJ_TR$NcI5{xLF>#IC;{?N
zN*?M6i-PaecW)8lQbZEDC?yxQg!hHJxKLd8o?-ZfB1!O!{i-(<2`N7h%(H^#E*dG~
z^q<P+eZ|p-rr}ieZ3AOW00~Rt4fPTrND{P2i?*Y@?WjLAUM~E_c;AG_Yre2Cfi{zF
zcQI}TX(Q8!X%5S1O{U<TZ81+wPRBB0ZD4HS#u!z7t<~;kwX8U~{Wzh%q#&8S`jLI|
zL~%E{Hkc92g$8Yw>1}+z^&fX~+Sb(899Hud;p(+>4pja{PAeHo9?VopDeDhaYpUQY
zf&`bMRP~h#&G|JsH-49A$m`L4ZEW4(z70>;HB8GJJUng{Hpcc|l29-7O&~qQSq7iG
z1cDGij>~mWJRYxu6u3eMsc@AB!BC*s{{^i1{Po4v9z%bQYfAH{^FO#sYbr=Lk{C%m
zc)nprm`ngys;lNAB12CEYfc;iLWmF@6h?&UpjaZ-3MQCTP9#}D<zz|-r4@`T0kTjc
z3$=pfLmlfpFwm_J2YZvU3OS^OStuh5<w8NIj!evGKZ5d*CZ|XH5S%`D^~%MIf4%<w
zD>C0Lv{MDoH!m1BfLCs$hO?|XGW7`HG1G$$b3p*1wa`V2Jgq%72-XgH+&{NMXd^+o
zju)e5F#0fjq${Hh5sKY<T@t*iC@=RbT<LS~_IX=HL&tRmR_P?Fy0^mZ2*CDX`)gfs
z@^R8ZE<P?gC|(*r59Wmc3aMhAHq9zuRRB!R^AY*Gz+Z9%WKt!B#26}n$vx1(k-757
z=@3SUrp5<Lnp&J|88+iu|0YM;XmjD%@E14H$nsZ8xo}_~sWOZa<_dQjd3*)A^#iSv
zOdHpdtNCd_c6Lj?4iZ=jG$?sR!t&8lRU}6%{6@cM3h?v4p3y1tA^#7JJs7BSwsYCs
VB~P4SWMHU26)}Pg`9nqf^?&j1^Hcx;

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r3.hpp.F9C1F18AD045D6D7.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v6r3.hpp.F9C1F18AD045D6D7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..19e6e5742fe619282de6fc791725986d9ec06a88
GIT binary patch
literal 3876
zcmbVPdpuNWA3x7HiE&(xTN>js>Z4(X2<tZNG&8P=4NY#j6_ragYQ?fii+5;kL{kjO
zmM+SYl4!ea(Hm{bCa>sCyU8lOyH+W@+a8n0oKJuFqxpP1&-p&T-}5}b-}iSphrgG*
zJ6#<?_#3@;CdO=!LJ<VPiNBQXqAjWl2=appg6x$3Sh_y`k(G9C=0L=eg75^5k-pz2
zqr8uti_sQ3^=)mn&Cj40+$t{mZDZ8SD}I0f%fk7~16Nzh(Z_#{3=c~^&ZcBecc_dz
zDHH~9lf)=J^MJ|Sul>6ITz*B>#maN{3!SxpNO#{TqlTx~TwAg$JMh+Pm!4*^%O4!K
z32(`@=jH>!<*!QDP$IJK?x?=qd;DviRQ5XB(NRaEKl{DK@932==vmwFAbl@Uh_-M<
zVlWUtUR(Iez9qF!u`@Vi;6);m9|*2o;A<Ux)3lkgX~2Ru5O>REb&@hT)~A*J96ZCP
zF}q1Mfyx%n+9%b@moz);r>SQ}frA~9rfBWdMC@cu^uC*m!@0XJYw_GmHmF``*z$5!
z_?_NXTJW>ckY|SrPmiV6jpYgt_3X81*;bc_R902I^zCx(4e?5dwn$UzX{f(*vf`rq
zIab=E;{WcCTRL?T<NZG{bmn<H;h4eznmE|w?b2mBY;~h;dPDNJ@7n8W<KHRl_?4sk
zTvT~Ob8oG5N<Ag8!Ti`$V}&yNs&`GXE1Y#JldhKTI~En--^`J8XL24IbmsluwD?uE
zv(HPlFG5{I*`YLjrO-DazlGeVU7ZXI<|J<0j_?+Cr63MdHUW+WjD&u*V07iX#&*wd
zZE7&e!yQ^h+2u@nl$&I$sjw~EF*e?<bd{h?IcF1geSxQ1=P_wLhazEztv6L)(!I&6
zF<U_rsE}yO@GEA2b@<2h`@06reX<t6eLq?lol?=CcA(Y9H2(|7j^uF6=5*#My-g=Z
z(=KmftW|y;v9f>aVNv|;6XA<$0+YKF=nw7;4GlRt9idvRG!HPzN156xZ8!4A99LMo
zuJcs+Mw0D*%Lj`ut~xjU*5p*6n#W`1$(qhQukwI0kE}Bt%k-+BGFBLF)oa5wty}ob
zCDDvE4-ngq4Jmy~-lM^($-`x<YwtIAZ3g=o%I=E3Rm{3LMx5BG>~EJSjf1W42RD>U
zlmbq5irwkLa_P}DOKF+mlrUGnq0F$>aASMrt62VXb)SA=9g>ok7<~pHh=Ox}!?w<v
z2@T2uRRmAfPF*G&5GA5=h@aw}h|$o~&cA9<ln8OGcI+bA0L3c9>R-lH(YZPlZ9>d6
zU@~L_6sxcKxo;Jfo1sBL32`ydE(WHv6I7wEj^Dpux9^S$MS)OZQ`zdW0g6qctC8`~
zntsxuEF#3_G;>SY0L9xymd{Po!>TnXii9`|XlH?CvlCRwdqd6Pi%Wzm6cs|{FrXd=
z^w|lj6t?=*ZU*?zY7{*}r4cM@1gzN!svNFAk??(_d4&!|mr#)cniN>hPEh5-`H7a?
z|K0nKE`>^{ST45Y$Ob6B`hu(X)pi|c4a!17EC$+Q!qg8FRB1<Fo~`w3@>QfLV+d}#
z(9~QuAj%Y=gCGD0h$tpD<%zO9YJe(>*4)O~Xc5aZp@PdWN~%;@bl@_Evklk`7yt}+
zuyb&PPGGo`of8Rh>^LOkYUfHqJY615zK>7i8_JP@CLkd<ni~nZ)7(kOgXTd(o;1%n
zq^|c1&#L=zJc{C>Hleoga1>u}vz~+^Y$8Y~$~$Tm)P!kB(DES%Ap=^k_IHRG-V*@B
za{wa;*us3$d)S~fef+`-3X1!1eOAJQP&|ki#D@qnr%aWlEIbs$m-Cl<LM05h=3A4H
z4c~@@?D%#hl<Jl03+oXh(%35t_(F>?B3Hb`z!#<wM!4c$3|2wkOg=w%am=BMI<LjB
znoz2&isTmiuzWt3TO1?^3X-Ff!1g5Ygk{9;-A{QWM!N{gs(?}zFoyz*jnmEsl+5d@
zD9+R88NmH0&Nt#S<Vav7AR#v+HxhC;awj1VBM%btH1Z@Ne=~oU{LCb?Bo=H`D4u4P
z#)2%Ocs4N01}w-TiWdO00>FYSqId}~D*-IXqV-T^oqpuaB!J@PT=SKXFFVV@{vgci
zAmMZcP^th%upKz?B)rB+hWbbP0)PFZ8E=#XguYNiXbde4nmTAb)Vjy^lM^|NQ?6CL
zV0i0C*Cpr~(o8$H=o8nmHQ98Sm)pMAyc8R7a~H83!U(Z}RzNO@YWUDFVH#w<9_ZdF
z9a%!?7Xn5hu!Sta5L2kLjT7#`h%SjX0$tcZ3%nJ+8U60P191XVhe?NhXRq&2k@umT
z<-~FZBf}O32kuhAaC@r#T)7k0Vj**!VR#r8Mnd6OI0;2!kt8I-M8+@=h^w#+V|dj}
zePcPm82UcrAyx>Cp@$eQ0hk0B!#tqQwtwho*`eO<?)0C2Av_Q<L`&f%(hAnojO#n>
zKRKW<kJK^T&)#pYVPUvG+n);qhT#Fs0FE53Wv?a4lGsUHXb*-bGm|-Tl*Ufu$|*=^
zrgP*d8?dthS5C+tz}y2ka$Eq|1%N9jXfI&y1spjp0qhdMg@j>vDPWcY4#d>imm?&w
zN=K(h)z&qiATW{<xdJ+Ks%KoFY~21iu^h$-n@iCjM0s(CTmr|5W#&@mXE2Y>zeqp7
zD1YqKygd#c4JfkLFEpSz_DQFJdV5e2!!1!uwGUr62!i3UXe<e(qNya51<))&!eW4m
z$+#Gx#d3V<)`Q?P&VN0d*E~odQGFj?{?I@;6Gb-j=C7}0U5RNaY9h=fz9gWq{r<g<
zj*;=JpPX=aIi?x@e)I$L+Sv7^Z{*)~Zl(nN;f(OPsNCz>v5sZjwhP5@o*|D7iN<ig
z3ExVN1SSF!ax-z8L%rit|IG*U^FG=7L5I8iiPL&uUh7Av!rrbf&$TD4_rZL$pkZUu
zi&vJ{GnUS~cdU(?T68+CfY>Vm3<<D-`TQo=t9j||v#&quyzd_Dvb{IfaGB5-GKBW<
zuBqXBNbfei;{;O<Mh<gdCB#RSkrN4Vj5s9ZYUDbHL@^0bGbx%P+K}Qg^FaJx3xAsM
h0YT<c6QSUj(g*Iozb0*7prx*%rip0DwKU|9{{YZ>Glu{G

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v7.hpp.85ADAEC27150419B.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v7.hpp.85ADAEC27150419B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0fc4218a34ce48a27f06718c75074e84069c18ee
GIT binary patch
literal 3830
zcmaJ^dpK0<8lP{BOk)|#WrhYZC-<2&W*Adr9y8P+mxL`P<B}w|+AftsZ4nVl5h?rV
zBD>PAT;kN;mCA1Gf+9V-cjTn|5w#P$vrN8O>-2|t9^YEO^}g%<-s|sM{5|FJw{#N8
zEy{C4!pf8g3W-G0!Typ{l2)_TNTd%m5^2N!-vWIW-It;x6=C)obTirV-xnIO59{|G
z^VgyJZE7eRF;|>^)VQwKb5{4)N2i^NdVQ%2Y$f4)DxAXJ-Skaq8c*S+d>}Kf>%0|g
z<&MXuEFF8a_07@CcT-pWWs!X2?#DxSGgpVkxx9VWm9)lX@Y&vvpX$czuNaqBPA}+H
zXANWrISw@Feens6$PE4Ky|!$wjVROI@yCi!8Lt<mvb|n)W|wbW-?2sB#&f(BxBK9K
zuWdQ(wX0a<9;f!{sb8)BW(#_z+4?@TO<Yt~B)(-XaEftmJy-li>p?(-Snf9b@Zf^=
z)s3Dxk#zSJv9Fr~!^|tWeJ*Ks`||n>GiDyByOXXR)KB*=ye$c|h}P~A)a`CE>Z)bc
z&?#CsUNxlOGC5{=_tlHD8l3jVnyP3{&forZtNyL%9CRKHa?}kSw9##->9-oGS=G4x
z*fz`j{PfXr&m)!|`wwQH-{bo_B$QROsUXZ|*^<&(?2FZB`nJ^ZWedq}zAc<fP4c^C
zbLnSE^l7{j-E^j=_Yt(;Y%5#wABXBEKeyZa$nA60nvcsBL(6i1*c0hxvtzJ0!NX^J
z(X^*U-HkDY;i=Ttnc51goBYaOA6w2FwAv^;?vs!Z8oMw4#F1`6wdE83=(4q}dw~V|
z+@tQ71lR0)qv=^O?u$mX>YEnl=_b!wUR{`zJJ;}Vjo;!9aevB6w$5;~<NeTIWuYOf
zHHz~VpbL7$+clSTWa!Ud`|NQ=iP%pR*O@-o<pw|SB4;Eo=fFd2k;Ny+W<|%gi%}a#
z`Oi)y7xi(vXa4+gpy6orc+Q}4*;~od&2r)L;=jA7MOWK*ZFrur;koDg_p#lU=Ff^R
zM~tWKQnVK4T!}FLqvDqREv{Du<4$v@fXutv|A1zce88aZjK=mqXR&hl%>3*8tK}|j
znK~uW33;!|H{B{(S2}I!;`2UrRf8AG7V`Hrgnw;1m>49i(qB1dmiqiysrO5_D(>2m
z?Y_&84}>&5H3}X(-t?yKob`d7tv5M-CmA|7Z>x9ZSIx*cQJ^oWiIMp<CwZ<IzUle(
zpNCF;u6g`IyT(dy9Vt02A@UbMBB|$Ivn?>a_mx4`LrAEti7i{Xkf_)+VLv^Kg>~oZ
zK5s$E(=c(0eu{x|fnu*2$E$kW9i$AhJ|?y^v12P2C=Ly{Qtw^T!eo${n0PN>?FDv|
z7pSsD@aN|k&J81!td6N<0ZJBNPF|qOj*;5t)^*fgO)?Er*#~I*fbrx7s+5*-FEba!
z(ivm}Ooh+l+bI_)F5i>;Qllv-L6fY8iFq`hwsL{ux=JI3{E+tgX=JJ<3ANU+wpK2r
z>13cmA^{Y%<*%6ax!&LoootA`UkeOs0cY|8ZQQD;=o4JIna3bAFqPRRv)Rf8ihH$u
zf><UkOeT3cCU%(NFi*KaasSwpIX3LlhbWnXi3<Ux5X-{13sfN`t&C2to{m>cgQlz$
zbRI!@COkHb6hS>u5B=Y?+PU%X60RVqFX~G`{!D)s^aw!%SOGI}6vzrxAuV3szSvUx
z`xFF?L1UP>Qal<@Kv{sF1w^ot2<AAy5Qtz*+J@b(ygQHL$P{#ybd?Mmp`dZnI2D3@
z258Fuf<fsut{bgTjClYA%?0dSAb^_~_MWDW9X-p(j0Q`CRc0t?lxvh5+`>fI>A5$Y
zRfWBs&Q3Rnd;G(tp=C(aNt^QB*->L+6LLnYHU*U$Nll<h3^_^X;z(vHW8=ux)Rl<D
zt`c14X6i<ec`!T-V09>{FT<CB{Ehqx`vsT=5Z((k4J06ije>*kkY<y{fiY6hY+#cO
zI51xnv>4bF0}d>TearC4MM2eOSY?HDp*b{ypdz}6fW)R^M_eW`m8g*YK-Ed3@T+S8
z1?5ZmGB|(&airvELYp6!@)E!<0Rp(8qutX7!3eJ$6O;pLIWUFXAadR7T0hFx{?5>R
zf1|_cAF1;({YW&@3_1Y`he}ft4eS%lZ?1?3l@L@Sk;ri5%5qh;G=jPcUFYF4H<lYg
z=FRjb3`V>`yb;t4c3?cr-IW}I`IE>_6sR75DYqXux7A_F=3eexsdrHQMdEiM!VG94
zXoypY7c_~W3a&zmqcH0*XB<UvBM7oIZkiO996>i)Z*;~{HsEFhDQo})%>mXqz!}HI
zfLjctus#U76IkyA&Jfckc^YOow0_@?EklA1EdEMBtpwJPmE#R(vMjH!Ho^ovU7j%{
zptYiRKL*@MF;GKLjseFK9y1s(S$POy=VQEN<*P!_HH-tkoda<1f`w|?Z;HQ)Fk7|^
zTVveTY{qP3NW2JY$FMVo{Q<or(<V;}XbZu3n^FsykkVn}QfSH(16K>?1iRpAkD4D9
zjH7D6uLg3s2Mie_<cXsNHzBkT+Tt&XEJOq(7K+8VOd^yJP!{mY0)8+Z6MX7b2>f7W
z_8;IpOK(kenzH4f6IlEc=h>+$&V`l#+SP<8PaP<1e6_LBAb;rCltkDZ{p7}cTKiLs
zDI6#c*f;mLAA+V(QV8fLK=}y}@D4!Pf#Y+|Y@PnRkA-9el~d)qc;MbtZxuR!t!HVS
z_`~xl-GOz}K+t)#d4$nTL=sgjfuL)UH3XE7q?_S`kqeMqV1}o+1Ry2A45pVhd9cGI
z{c*eN`t_H?7pF`&Y(k8dCXRmiE=I0gA%T{}5;1|6q)aJ+Y@~dt5YLN@B_qTpXUYkh
zF$OUzTIz7NO!r@8TsCFSy6&bf;h42;R>x+M%3%`b$9y&)x{l3{xg8(Q4>qTU9+reT
zHS|@Xo;k1Y*Z&Z{8!KisKbpWsvHVySf}sHN|Fzt`;r;$G<@vc&X7hnDKjW9N3lxk}
z*09&8xb~^2!|l@vLoj8mu#6aqeEWWawRrPo`CmSj+ng|?!AP(k^lxZ?%Y`|o*yd9j
z{kYAu*}VPC$ae((xo5LO*U;(TF_V$(NGE6#`%;UKn3$5}q!r2T6E~{4`1W7w#3K@k
fj3&zcHYHIq53B3awY79eTHl^eth)HEHh%j*-&p=W

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/thread_group_tensor_slice_transfer_v7r2.hpp.E55C206DD2AD1B19.idx b/.cache/clangd/index/thread_group_tensor_slice_transfer_v7r2.hpp.E55C206DD2AD1B19.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5176faff8e49be827be63f8871d18c6fbe45eb70
GIT binary patch
literal 4554
zcmaJ_d0bP+7S3EQ62b+-g@m0D1qE4zumy|~Ab@Ozf&#6yXb@x-3@%i_wTcVz2|kM`
zRHYWF3W8YK6f9L)+$tcpb)zgQ&m95l_IyKNxcBuB?jJMv%Q^SVnKS2nXMPFw^YLlc
zVlcdx`DLWUu3CyS7z_v*>8oT3OOzOlRD!|CC_2Araqt~Go$C%w&6=W(__f+<7JqeR
z(B5t0jhADV6?I;2u9|TtA>zo#H^wbNec%533Z5B%{QK9%?+Z@dY<t&t{kQ9CdmL3{
z2NMntt}Soxy}|ur!`+l9mjlIN0YR}3lAb=g>^%C>*@SpKy2W{Yf_w7QXyty>XSs&8
z*5a@$`%c>VmjBKh!iKIi=kQbp@Z~SH`!B}be^WR2R-uJl*;u+}emO&arc{O%xG6u&
zoBOOSW4@HgRb@9k)6ZmQdfA5kT^w1^t7~_{vH4u;hQVKZcE2#qZCJW6BwP8<Euvn#
z+~?j#t7o0Rq<T%yvGm+DrO`mkVws;!?&aD|PeRvguR8sE=d0uK)^C!B6IONjR@e-E
z8?ou9(uFDmk)0l^1G9}1eY=KL!&I)SUd=xG#el|OM?$^#me1Z8NAp|_k{0aZo=txq
zeo^~DQ{Ix;yK0Zz<?Cy|E>Uw*zAe8qY<%`u+`cv^-hR9G><as)_l=!*J^P=e1REu7
z-)PCqFK8Z0`N99%lB8Ew$}wMMi-c9C@{WtOU1r|X^Ne=6N>9pjEz|WXv`#yi&fssq
zA!Y8fRFxg}=#xJb3YJH=+wb0K-=HMxTT^+?i}P>|acy^-p8IfN*Z%lawyf&-)>Y=g
zIpU}P?MbQmGVb-rL5a#e$sD5>b!WS;b3Stp_bk&d(3U!68Av%t!Y+kL^%}FRXSnwJ
z-%xG&c}w&=j#FtPduCilzh{Ei^G6aTXJs<;zWA7p8j#2htN$2wDN789ZBr6!NtTX;
z_cn*8zfjg9-lVQ>v(B?SaWmP~y)f2PHU0LlHp$68o0i=R@7Z~31lv29cRySG;ZgH!
zd9Rb_?Z=B}Bz;WmCacb_842ii@V_-9WL|w&_YY?^Mh1fqJwNyPiu+{(ljqz;Ig+PK
zg(po;2{os;L|w4k-4{5(UcRlg>t#YznZ9S)QR#}>9fj>`v5(Jh><4FXt@JiYq32Z-
zX>jS$Qj`0$<JLd&^J?)3*Wrboh<SJF$9w`OhqJAY9~gRd)S@r*NYRP<*wByqW@=I9
zuXm`wc$+8xo6C*fy^{C9H=(X)&2w>Ger@uB#yh8<+`Zc~&!+L@p76p~Yt1j@UA%3?
zyX-?g_%nXxragc7r3@Rgzg}CC<~0&p(zN5|AtL>o6lo*|3(hdNaP*rU${LuF27~MX
zx*fo3Vq!3%7(RrMr@B1Y;UODSfhrj~8HS1hiS^1mF8O~VUvV&fD7H7Y=PL#zHY%FE
zSNc_vR1M=o@n*o<49JNIsqhz{|1o%TV$d{<rOqI&Rjg+#28JpIR2d9F0`oMlPlI3X
z(_&-F&_*%9ivf3HLN+AVUY*%7m>9~z455m%fwQe*K;i)6UaeF8h88x4LvanzsR1)5
zCZv*Sue>)PLpO$l=|L5Nv4F1_kT`2v+N>X4+0|+o0mZ)n)-S+lVnQn0UfTw~U4%8W
zF(y>m3UszYmGKFwRO-|Wzx_CCPz_UpDvm5iO~ru3r(Z_LtW#R7KkjA)z$@TJ$0wxn
z%e6tRpIje&7>`9M&?$vG8K00!`-acdHyz}4aWOTxh)`cBQVdAkvvfg1()-W#IhZaK
z+Zx;Q6$29Yx(!K;{foD<v1w484|MW@?!<&t26}wXb^gi9;9v$&r4bl30^5lRsmNmE
z($}lO?3u_fRf3$)ns0(!BS;w`Go?^EkxoO|M7Al)I)dB<*t>uk3NAs)0b5STC4eXa
zrYJ`UvKkQ8z!Zg&HIdCI$&g*ZQe-juS>0~?fy90GcW{DCW+!tg&(hdwG*k=(#Xv}{
zRRIJQK#11T)a&xJ>%5<Y;bf9H$paa|$y9ObBtrGVWGU(dIY5wZ)^2v_6M}S~=RWxf
zK}xt114<?lNSrC;ZQxCdt3TJDrkSjpY=|`VZdWls9IFvP@1n#}lg4o}-ZS0{DWP(O
z;Tf-0<R6+bmZCT-zqe|Nt<O74C{havr@No28xsD;nsNygJ-EbpDWk7#mVKCOjAb`#
z;iSY+VvJmZ$VK8xA#a{HpF%!7A3B=rGmnz_@_cEsKu)0Uy?FJL$M=3+kCUOCP#Ow1
z45zIZ!Hb}M7s-pHBT1wWwSv@AYDJ-Smg}rge&J*eu*?BgC~tA{5U@N1tkCXpvJzNU
z0xOg|wqL#MemuOe3J_#EKivTtwXYkl2o67D3iHxU!_5r+OpqcC5e>QX+?^<y2hU>?
z*+2Zg!f?r@RDhGtVrLJ;lTNmO-Hjyd_QU0x0J913P)!Rx3p}ge@iqUUF9|g~nl;Ll
zK>av&++?jHb2v*ux9B7ZGRi8-opLB%5+6ySR^ZqQ79c`J)C|*MoR^W2$oKp0A2w`p
zoU)ptL<hA?lpnA7hr7voCI&OFBp5<>9Z5$cM7wFi7c0X$SLrGd<Q&~OGZ7OKWZ<ko
zXOtTR89FO;5}~#L%XO?LAm`u#Sc_=o>M_s5gW92wfe)=DeVl#VD4DN;FHIK04WX55
znr_-;xuW>Mf*bEP)Y2UQ?*Lpxtti7_0LPmRQp{PdInVv0KMsQ+2d;7;LYlk$mA-xU
zuTMf423wA;5fuh|j=d4$CxR4k1V(84g*!|ls|HVWLR*``jLl#+B8sA7tKO{U)h<(>
z(%~=n9MkYyXa>LD3G{bT?2Qgam8Ix|$a9$N=5Dj8n2F3qH018=K9`btID61gF$gFI
zA*jd**clg;f*{m13D`^nDnJM-Is&HLf*KHnauEFvLpk1ckwZT}S@*N<&Y|N|5?gIt
z&s7Pl{VygYMO>(%xIB&BpU4b6gNC*Od@G>ge1PZEaS_0a=(rT%rF2{l@Nzn?0eB6C
zYrEYRZ~TtY07EU~%O-iGUO(<saZBzcY^<M}?*cX`7shhKTu$Kz5nro+`=}fejz80%
zq~aXH456Viz$^nK71uhztOF#9V?)<KfrTV74LVT*_$9yr9mu)XTVIsVeLXN`94}n3
z+q5=hWh?wTo1blg7!@rB4`=+ogAS*q<K~qs{~5k8WeT=lXuUnXHme32tp@yR;DF8=
z?GRx9*UD~)W@d#38+|vWr#HJ-M_8F&-k=QmggJ*xD<m^}0j-eC1Oik@kZbe;XVSPv
zFLV;!n)CEd^`a$v;UOgo62-_ff?Oe3F^PU3E((p`zcXjb-6BsQYmHs-P^2Im8awh)
z(Lq<(fjTr-*r7BO&JCwAQ-nbTjkO{TBI(FIa`OK0m+*$yo0EEA%CMm;COpaU`z7+P
zud~D^YkHX~Jfvp8YzDIt38L<#PLR<q(E(Isq}I~O$Pi?VMU04&#ahSGWb3TgiBT#M
zWTr)?h(bBQItPfUlGqL`wo`OTz}9Vj2#8S<6XaoFaTtgwTnVfzff!{JK~@2aDj-6b
zHE}mayL|RO>hGX~+g#u-qlhSN<nT~YUDKROzS)$a8{OttYu0{x<RAU+?X3-tH;1Yk
zp~*OYoC`83i%niSMnmbcnDnr*7xvgN{%E5Y219f56(hY`PkUn@UM`uY!`9N&<}f&u
OGKLQII`&u1#PdI@mpZ8c

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer.hpp.CE773F216A04AFAA.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer.hpp.CE773F216A04AFAA.idx
new file mode 100755
index 0000000000000000000000000000000000000000..501c836879fcedd805e71a58bbc3b0ef695e0b15
GIT binary patch
literal 24630
zcmeHPd0dUz|L@tC`{1hEcB{)Rm9$YQrc_!UAzPXxyRpm|`!>wjLeVOsEQKgprbt;P
z)F@k7nozP!4M~PbGFg77?mf;kfBpWL@Avx`ubJm~`+V;6oada+c0T8Go}huE=rB=7
z$9KxWC9|e3oJ8vA=yV4EFIYJLhyJvVPN}($&XSZDVYBAkb(L36|IFjI`33<s7v?>T
zHQ(X>)1jPws>yTSckg+<ua$WsJ@D?pm{0G_4j2x0db`PCR?gHe%7L}THTM@>tl5}-
zOYJowfL`D<CSZ4Ul1*22#pdq^{<eKh!@^4e37ry+CoQws(?MQ3XFZ+0wp+N|X-dkB
z;ZrU*tq$y6VYeiA=S7P%w>o*6buj8wedkuKBb%3Cas0zFZ`HE;sV{D<5`#ZsPV_R*
z1Ina@bvyLu=95lwp*k<aC)7s_niBus#xCNtaaxB&{ea*BuTCsG)6F7d*3&6d(=PVy
z<<xWT-Fpvv*R1;W{pijedK|f07E_n^-oNwJq~fkR8CFi@513<2O*3|QaqXqum+qF|
z__fo;q`%`|*OZ<P^K<mN6V>NIa<7U_5j~$Q4P7}QLDoETuG(*kUuro1aj-$J;@{|{
z$Dj0SSSt9A`1VM@r%z8;F8Jj|)r(^VcYo~Lr7&(HH+Dxw0yW}x^)F|W-{jx;bIbT&
zD(W_6zNq=e5I?r!*~ei^?5lsg@OEA&^~>|k_NT_yIp@T6UpH?_>g$yL4<@aQUb^dI
zw{9EyP8-><WA^Qf8#+H48$WijSJlU)X)m`movutbtmyl0>d3I=6%NxqX9Y&SOX}jX
zY)-LPSAEOXve;QeFLw;sac`-2>E@?ly2l3_P03qpbG%{JN6Uo=%KZPz?p;{QSLr%O
zbl$RDf9PDnflxkn!-8M-pd9aaF3)aCG91B1MzeD!-#I_+)X5nGe(3#O-z|+V{YlRq
zg+sqL7B1BP<Zyi1dLr4@;B2gYb$Lx_UCHm+Pe1Gac_MSQ$Fi;UD_>R}JTPf~YJ9yV
zH)z29*K-z~weP#wx;lU1m~BCKr(Wsx@Q?Ko1Ji{H=I^ru-+uhFam=~<!w<aGrOkZe
zJ4S5^Uz+=}Byg?G{Hf>U2M5Y!Kkpf_>HJ$ehk_41FO_$w-*G?E=v>a0`oUXU?M^!c
zB>q^Cc+)G^Vw$I(UWtQomtM!Bzn{2f@A#<c>!xI_`tZu{htNg82Blr!rslfDMRuJ%
zJ)k<oI)}h=X4Gs_|8o8B98sNEH_)ee_~aCM)4U_`GkY|IhMPQ}?la4%^Vtp~e`=5?
zn^`>w3>-Ek*S=!x{Ioe|Go5<|;*PeDbqB{*g`9KgQQ;7_rM%Rz^7ao^l~JF<I<NCd
z&F_=(SLfPwB3J5bYyBzVk13bp)z0%X_ZWv+UFko{bwW<g><4S&$MstIIJ&%fP}0~c
z&LidCU#i=e6^19@nbbAioo%}~dEhU@e;HL)l^*socxKnQ-xgdcnEd08>u(cWsyj~o
zl$vJ#WU0e8dd>{%A$uyeP6@kSvgw`AOCC4K*&yH4*P-#vjT@7u2PkTijjlaiNL=h!
z6X3gMIGcY`ofi7tChMS??$eUZ=bG7dFC=~T2F&b0HyNz57&vkN9e&NXskd>;G9tDy
zdD4cNqZQRUP8T!F`$uP(h0oajsiWsQLT_r}cGc9p)ZE>dyhg9TG$Ujyb|d)jS?Q<W
zGeNJyj|beSdvP*w{I3tEC6-U|eh_6}j?el{ED!wXci~R$sE8bKbJCAja$>)GeX;NP
z6u*T_I|sRV?zZAC?n)iL>ao|1quzgQT&&ulcm3{1zGtr6hsv1T+Dsj<@mribd(EL%
zr0RDaeShN?lhIqMvlC-12mE%udq{D1kCxhy=)XgoCQphS+OlWF!}6h{HopAzMe5Y$
zdV_aZEWL-nHhWRlCEvEw-9=9uME>WvRUuDqnS|MG8<#PEYSh@DO6RJ2D(z=4?EB5>
z?}uFDvQAbNc3W^SPQR~B<mwpps3pNiLyT<~Ev&phxA*YMIpng3-M?vke5Q2Ru6?1U
zo{HHu1%_ig(9LDDHq{ymC9h{3%CW+Z5zTkbh7EddeMIgub;L7^2lJOt{jIsz{r8QX
zGYzJ7oa32vvN4TK+qXWi@OZ20+~6uh+lR%`G<|>ePwJ#U#&uT)op-&j9}>`~LuPs0
z<bx)+zHR<v8MVpx=2VaMmFk6S4$b~BxZrKh6x+n8%R_ur4k3r;xNIq0#rFNX-ua;Z
z&V9>A$|vjBGe7<oekyfjy2+S#^N;;y<vDm)0heJ``{Om!<Xt7jLw7h$+|cdW$~Px1
z^edImmo5L!d+^!vZ7zXRM%fdG9bWZ|e|9;hgYN6{@5)S%Y?_)dKE2znmBb9YwY<)P
zrL!h~;&gQI#-9&A7%^|V95->-5wA0f>x`AQ(P4qd0e^^wh26#(St;bW**7}k23)ZL
zH`F$eg;~@$>dNBpyU=(*KOIrYD7&;@bT}Tj($T?0tmF7Sls@-hlN|5bS4Rw0gj%*=
z;O{%uxH)eNcNx|ZcLu$9!hrvVXdB2PA7bNY@#w`^IX>7^M>MrIb!@*tvWZh;s-F=r
zqj4Ko9q}^HT;@HrjSkm|wrKB#>Xfd&u8n%X3|IEi5zBbKjPImv;C<B0o7jZm7V~v+
zBhb+t%rXb_)HaZ%`|>`!HWc(6MdSA1QH>bWh*@hJ9Zt?F+8@>Zp;tkO?^s(I-VMA_
z5y2M`y4nWb$7aNl)zhaf_5fo7Snd#tJA|dSfh@LX)@^qWEk3G`4+2B~4l{d)skIGc
z=~)<-G3($cr7ms^Sf1jRPjOFe16g|S*x@-{ZhTsfn*x^4xZ*QzrEMUKYoDs>eX4tx
zz#O}Q2XTsboT0XXEPW3Y-`)CM@1-5_Kya5dqC*;CsBIw2Hy_U>P50!_(s(a0GHY38
zEo-f9bhyr(Zu`i%D_8csrfO8la04*H0epaN`vu;^!@}QSz$Vo>8Ey#3gZM!P?H5Sy
z-_2Y1;tcsSIc^Sm5rr$Fa7%6Tg{5jv2JN(94R{xDhw+N>hV2(f4w;qX7T9qQm_A$3
zNgu|?rv0MBb)fXxJ2=Vg^#avK%PT<AymUlMYfH!W3;g}$sWq7f!Nr*dcn=^%<-Bn@
zKUCX5mT7;<jCU;QHMAq{4QBQk)8QFosBIw24}ovDXO5mXQjT{4EQe^tA=*^id|^4i
z??q(Z>4S1y1z1vPMJjEgZ6M2>?BXhq*_G>Qd?3&}3&z5t{i4Ik2%xw?@8)*R$#y0?
zuL4u=p(7S>iUQ6`+rWFwb@yMFqrYpI<}uX-Sxv}9P4k6i^5C0^VcTL2@SdP2mk8ra
zL|<(KSr)!=Gv2y(omz%lf_eFf@gFe*Z39^r7X;>XHZdv(6A$iEOz_1(B{U6WS<%JL
zxJ&n;hCpk;U0N}|71Px=kY%MqMTz&U!#OnW2&V50!JHwiwT%uZkoxVfy=Ki4E>?a0
ztpRQaL?F;O(7pWv?-5baOBWZTvVbUqp42fN>cISH8pskIwV_3Rc%eHm6`&{YG5$T)
zN!vh{n0X&=R&uv{cf|XHhup@x+{Sun8_2TRZ6~jn@9|O}cL5KoW|Y;yx@sE8vMp1w
z=(~r?H$dJ2%LRhJKp1En$dcjt>9l|ErY8ot1L(_V%=k0bSKC0AeGM~G#*Q^kH^93C
z{jKAS>o|XH^M$4P&4dLXDPMiOH+WDjqpSse(KL`Hcg~hyN_Hq$fR_euB=TaX_6sEE
z)#;Bu8PRJ2utwk^wVYWkr`9%*<>YWz>*(IE41qNQ4>?I&o}@js4P-gJ<fqFWA71$$
zJOqg3dX`_$cG5PGrQrO~PvJj*dlCo}xXXEhKM%%4(?FJ@vmf#@Ue5y7s6XgS9&en-
zchfeIrKGs4BCoh(5?}#}dP-(|O4d!=d|`Rn=h6Q9g;k(0;4V)%{t4Gf+d!7mi@RPL
zJc-%e5g!7&dyVUGjWf|UkmahwgmwKa{2zdE0mJ5J=BIAId?ByA>fn0X@@I{ze<I0G
zq#$Y<$Z}24DdI0UuBCBnAk>dJ<}v52ZFIPf4Aw5x*GInPF6%stkmG95lS`c964yoB
z!28@}e_H?Q+M-UNC*YN@@@7|gwYGsQw`O+qr5>FITBrnL`#Zt^PIS^XkfmzGw%7wz
z*`t9Lf(J#jd^D@8Z6M2?SiOTL8<)QUHWobSA+LDI8)_TKQf)MIqE9gI4onYtP!U5G
zF$UTOvi#Y{X=#AqSqtVFJSdjsV_5@j16iI_8eGAzMCO4v0(U7S%nAv$wt+0KJiLyt
z=%hCv^ab1{g;k`mhT7%}OVtW@n^K=uz#xMes=_){VTRfUvNSDOuuwT;*%9FQ!GoG;
z%O)VXn&t~ju2qzm`>)G@?*S|ogrWjCY)u1M{+|2HcbDVz-3GW5=u13n9MAUEHeXow
z${)P`C97C7E;)=M2k4=ufh?^+KMLP0n+1Y0(A^SDQG!`%n=dSKeN#`1Z$E-@0S}5~
z%_3Q~wt+04?hYZmuez3N+<pQ<CIBtcG>~Qf)M*Q54MR~N#%p6gXk}sS2MC(L4+cLZ
zSrpeWg{3h;tm4@!z9V#tq<4a;Z}hi0x`Ma|lkdT-`fJ^>AZB4g7G{p{O^m&Xso=|k
z8{zUQ+zLKY5NmLu1~*4InqZ>|6^v#CF_mCb2^D-zL3~0QJs};SKNG~~q|tLJZXtyh
z(j3tvX`@Km;i^_W1u=#;ijm^2w6K*nNA%MSdzw-Gt(`wXEMwR*Mg`?u5brbWeMSW%
zYC(L>3a?pnD4v3NfnzUlD(I5wkd|vl(rFV)5L*eMm9Rw}+e!&rDO>d53$$>7wuQD;
z5LYw8YQ`2ytRUWHg}baRd=L!AWqf-Oe(3uj4?J{yyG{@nC>GeE(a*yad6*rHcLi}D
zuGoj$q0z3#74^6s3@-(-kx(=ecBqr_lp>z8L!B(76@|1N>SQaeXr=8?Clh!@0&fQc
zO+jqo6)n6SG$b8*uHQamORu}Wf*5QaY!4kc=ycG6P}J$g!HdHXO2DiWFnfekF!vO!
zKf-xfP#zYFUa<&sD#H3mabOV^gm46I9f8|JciT23LHqDfG$aS`-~)IV!e?-|Gk8CQ
zFX5(_qyk%po0ds&Jswn#hoT{E!JS(0K2jXmf(Id7Mp%~-_H(t-mmroC*5y+CoG^V(
zNCdipFl~_HMk1(@2t`xYL<BbxVF*W&LKLZsa55<*OYu3<{T$gJ(XWt3S4b7Y>nYRq
zlte+IDAOn@j;Dg+sZb;$8>!%pR2ahhC}AI^i*Oz#<VkTg<y1}eLG%}t*9*!Y;dI(J
zogVmuHcl4A4B9tCijUI1M*-G`+Jcx%`{qh<K5d#$n<J4eq=O3SP=uRl(`MQniEt|&
z)JlgUyqUG$%-TbnFNlS#Ng=C-dMJp+tVyvHUtom`tS+KoWlgS1>E*0RxfI`Ig_~0T
zPpsD`)*t<TJvVqgHv-|koatUpVsK7yrYEHMC1?GTvxl*wASUoZ34AD0$wWRlkq<-o
zSKj(p-X7t@ywPD^g>WTrUCG-+%ipHbK`neJRB2$>_~2GP4B=#1&t#b^!dWuYESWjN
z*)r2?DbACb=1K8!ndxzaI~?0z*qxet1{XwA(o|^ARe*~iT9TGu5KKG(2iBH#fQvtv
za3BQn3Ms!r8X}>trg~IUy`Yfs+B5`oh@FgX`t-c<;72tn#$YxvSdSy?wjA#s9($t=
zJH=qm2$x~LWtbmyZ=_g^2NmNZ_eON<V=l`tZ^M&{@yQ5B5WW$FA9Qb|m`wO46MoRW
zkzxkvlR^4UrSEdd0}2$RSW0@9l3qj9C3ilTC%-4fN>W`(nj?A<<x@oYR=#Z*@Oy{b
z#-#Xw@_ayfT}$*2`)Rq_j1-$Gbu(p-=+|k_>$De?eo~BQJmVQJDD9+pg7H1U_(Ata
ziY1J13F8OdA1PKbzEzALbbqAyoV9(<+NIh}{xGWKSurU_aLy5&OS-`|#WVLMEp2#6
z1UD4nG|oDW>n_Fp(>Nc5k8#$=r1WE4|6@{mI&YKC_xSyni`I`;2X!IEblxeQcSbmq
zx5<>!GkK>>DLqbR6DRBOy6PVDEthUVig7ZhIGHoT2{M}mDLp~vlz`|kKZW7#c`4|G
z4U66x680}`fsmeJ9Z#{oNLx!-=Mr`R#74Jr*bkW}oWR7!V6HKk8xo{*xZ^pzClZtl
z(maE-KofnE3_D4VLAaFkDJ6Z8U=&g2MU(}a=n5*Vf*OPH1Ip(C<%<O4I<2}+cR_+s
zONZ3bqY#c~RPjs~BpADykloBEgrBklpR$9{#7A<<NX`UJWE$t6#)$|Y<NS|FaXRms
z&bvt`&NWkt<7BRJGB@eOxF#SBGl4j-&Ho+QX<Ilof8%lBKiiWvkO5``0iz}r2nlVB
z-<I=X(6Fh_rXQSf*U+98#GhS$_CoQ+NzCOW=7r*iEZijv_d@Z&U%1O(xEJ#0$-GN4
z?}hB=XWr#A?*;8&SH-(i+|@bnK}fxg>$8mufT95(0F2ck%=Hj94EBK6W+tIk10gBq
zUxtl`Hpkp6@3+@|{@}mL6fg$=MyBAZ2Z)L$--~2Qla7H5FzpK%H924CHw2JhBI_6d
zrk-H(7x*Pj<`*);lrP0=6TR@(0!TlZy~orrGu0NJj^2llL*jM-A9DbofN(uNx*i_~
z#n0~an^{$(68ABJc$ySWlU<;}5X9S*aGUA^4T2yxu|gBu1uC#0ZsO#dI17Z6IGZHS
z1L0)OCRvIPa`J<S-+symcQbr>kr62d2M341(4G{R+bp+*qD+dXu&$>tb13V;J>bC+
zcnA!}NwE>{+K8LOw;{zcBDjnQfx$5eg5%(FDXt-$Ylz;^uajae;Z{rZMR*PAyoQv}
zH<B(JNmoR#B3-M<K5&poaTVpfis}vH0#Xd8+`_582=AtxcT4HnluI_{is&trYYWu}
z4l*fjW`j4gAuxC(#WQS|Gbs2Y#mj7$%ToN33x3Ikz~GG(xA8XHcw1;pfcxizEBO#;
zvq1l3HsLZ`Xj~kIO&Q>H=%*N*6nA1)JF)I);BVqqH}URJY<iBxO*`^^b^|MP8CPA#
zZNCt-8x0_NRgJq=<HI0<*KRi;L5(2%BZ%=Zk`X{8<DX27hmlOL-<q<Ag`{-_<}{U*
zr;=8`|31&(*e>a3V0TG*xfJJ9@_fpwr&*`9PY<3*0&7r8-YCVTv`|W$kIh@q=}e+Q
zp&&*wLL_5O*Jm_bv>dMN^<-zZ$;9(>C=lxq?#+Y;n%qRnJ(2Q2lXI4KKTCTc7x$5N
z|44fvmsZKTRkD4cL3RqLji!x9yD~tcy@I`wL`UFW5x6&!*)qbbjPRDqs8_iZZ)UwV
zv))JsUUFV9Id3$_mAqFa?+xb^4k)l0noS%y;!p?xL9=xO^ND>dC+ke=G+~iJ+pN3n
z#CoGyFT(taFn^?OH*uGncyBnTq}YP{wc!3p-8Paw8%bX@f_Y>}9ytoBM;;lLCk;Yw
zk`r%|lMua%992bzBAiY6WK+Iq2Fj_Ba%vP(k8&!kTpFY_Q4^b}Nr>J;jcTDn5iX<q
zmC>FfZp{tX3AnVI6tC0cuhSFIj8@Y@)$~Y&8|Wbo^iYIjn0_&g=WSuYcQz}|#kNJi
z<KvkLNCHxspj2ig!uy#a`<bB-_wh;g$^Bd1=|3)NaMz6!#BiJs#}!bxK)MQNV{l!#
zAts3Vct}1zar@5`a<_Th$QHx`+_^xSow$vwZX-Q!ORw42;BIin1u>eii6)$(khP`P
zY*GnlsI|OyuMf%`-0K6LPP5U6T%soJ2^rviA7E_T>_f3i6-HKJ!YgfR)_u<U=^4+~
z&jabO8p5@P5TU3Ez;b_6OAJDI73I2$5~1jTfuX($rv@SXgz5Z*QNtykN5_8WCWzgZ
z0)zILVIMQLa8m~EGy<(@+h&9vg98fwS+l(e-x%r+Ahc~NLjD0aQrvfh`7ZzwemX6r
z(*`IXag-L00<6tH3>4?7M{mWC0>)Y*u$CBx-YA?345x;{HyAOyGQh%Fum<toc9Pvr
z>LJDx6nlcwgNz{M$g=lYJ!qG{ow43&Zv5}(|Klk|_^x0M+cr$mRJU!dL1Jj@)?QCe
zt#SFfH2%6Y{>RdIjP>+s^KRAr05tFo*7FVK4mE15#}2(cRNd_VxYdMP_+T;Kwto+|
z$)O_!|NPIT;N;84eP;jZNPr$bVJx07*3fu-XOp|DyKMh95cHfR4NsD;(18hJ1!Y)4
zxk3j7beh(yrJbNd5yah$-fqSTItW-20POfxg1<^AU{D~4_XzeLp$iuTV4*;T)KL@B
zjh|EU=adrZ{5sld9o-Y*SlTL9inr4CTWL2$zt7m*XPjYPhSwG%!1V^SPQa>Zst~>q
zv}sj9(9|A4!b~HT((3XLCLn$jbwyLSY23Cx(S47ZqV>t|M7Q4w2e{!VfK6WA^{gq}
z{{=(M=|161OSF5iVDZZ9mhq-=Sw1syvbA6Jpr63dm*T2Yya)1Bf8a)c;N1|erWDmw
zSM<h*Iitf|H)s`S7G>J(NPRozYY-&<cY~lg7H;RvEzcZm_q2A8nE}w#Lj7a-_VN_8
zBemFSu@7{sf_MP8I)MA2X>B5`ng}1{{5Dcn8z~<&&568KBJTr(-?ntKRV(iUS3Gc^
z4al#i?ga)U@P>e(DRY5>4OJ4D2ThF&B+Tpa(b}T#mbE~6XFJUHKs_$T?Tc}Dn3)sA
z6RgV#)&~k6uo<k&Nhv<ZIi2IYVb4Ga$2-LE9_S5Xd52gje#1Mw;XU9RfrUBm(1h^(
zB4Yod_m9l~wQ+*QJ77dKl{?Tt!;&4~_cet&a6|-PMaXB20L>aK+yRW5svUR-0fdcY
z8X>?q7Ty3bYHD{-Eb-4V58sw;^I=i;KoBjZ(cRN%4-{4%rn?^oSQ}Qs)fbpGX*2>8
zEj-BZ6Z{pa!M#PWw+P)P?QXsxo+U%hk`s|-KTkTJC;LG$5yYpY)l;%3!VRQVgA})r
z_AR6vqMv4LPBYFhQQo#sY*WTK!$di+tu%rovF??>Y2UI~CD58U?wdHy8;TD!q(Eyl
z^+#w-TQToe%pVTA0A_9AM{FpZMnPPKd#}R%(NZA__m9E{!N?ZS3Hw!q6AJY=5@s6-
za~SH!3?kOP*zmcJ5^StEnf8U5ZLqQ8WZ4(GRzdXB_cMj1FJSe6DWfqpOj-*dd8Pjj
zGlkhKK`g?RMYtNNN+F?NNSMO%4p1yWD)x{jT>p>n-6wf37Q_-#SwgB|5*zHTQHny!
z<f>lN2$z}u4?yHgDeqHin6wteU9|2l+87B*7Ok5F@Rz;6GFo0n8^ZM=P(4N-!x+N#
z9!MRq`jM<@tG28{5Z7|5wVWkPQVZf{PPv&=qny`nPPJP~&*M~iQe4aF*K($?WJ(bC
z^2)uu8YZ<t@{U*Tlj5tq{#D)-7VrpStV|gzQ^O>;Aa0N;Hy{ja&%hF4I6qt;>Hyf%
z!{iB=3A9kWwjK?J=rQ$vmETXB|KWd}xQKoI+u4`CUH<%klTUGVN9T>~|9r#0_W2vH
z*S!iq^-K0wpZ{O>`Ow9;%^0aDCA%VxeLyK5P+g&h0%zM^{s_$;2$&h0TBi3tZCPT{
zM7_!BY~4O!?~5>tBCHooO@VD?+@b~V1yfP5U=lbhO|>MvV@j&+<7GvQ+`g_qzpg+3
zvHncUTPM{0aAQ9!h|>+GJ3!+oh+A=kt+)dUII~!TEY<<~UqQ@fmD#Kl!r{C@IPZYM
zwbi_GHSdJ*Gv44C?|{O!dR|%2J3*ZG_O!5M_~d|p9j@(J5wqb+M>qek#_hjs++Y9&
zLd$<{^T7-~AZTi>VZt3+GZ5ZtN~@u9YD+9QJz+dyV&OpU(e5_0clHM6?hNL526I8?
z?hWqv26utx?qG6k{c@dkHb8h&WvWz}JrpNF+$mG<l=Vb-ugqny%opMPGH}#ILeG-9
zWJ&3VWk!c()`<U@%;=aDAD6itNBr5UmJ+gH@-Dd`&Uc*ejkIX7u-FWtT+A^S^G0uz
zfjeg4-blv|;KBji4B=<E<1^eFzO5j>A{<|VEfVda1hC{GgeJla;W)}Mj`Bu}z>So!
zkupQLfOagPy-^zGZ`$#1+8d@}Kpe$8Zs5I<&L;9gB5#K9JKpgf?~Mkfl^0rhGl*fb
zgXY^S<6&GShymUK{xH)mh=T?W8Vb|Pf;hxy$N&_fh5CmMf>lYtX=3)Jm=oNv&$H&f
zO}Lu&46FrWxo=`QPiT8!97zJXYOlM80@s#BF?+(PQ5GeEADzIDLobxbk4fYwKz!`^
z&&Ld-UmyHh67Vle0-%HeYu-${Hj`hn`MmbL1Je8xjQ<H{JX#NyF#aXXc(@)sIqlIE
zKfbhD4OUwnObsPvY#l5mD3A>7h-jf?s058C$4gKSM&@7=T!N7$Qv4YsKTB~GPDbGp
z{tY;}L5iQ^<WnjBjFX?GcmqLhASA!PM39%HxPl-nr1%a&-jU)*1o=pcKa%7}QgXjk
znoOl7_z+DVlH!vzc~Xj-XtGI)a~Lv*k^KHFL!Oo5>kN5aisM-_o|W*Yuw;r97jR?&
zC!t^B$V*aO$B}hX{Ej2vNpU$(mh+O|Kjg`W2%qv$cdlLE9A^!dpcuOr(?zB=3KOC*
z1B8#_>``17nbNbka27W}cr__sO)8P;Od#b6Qk+K0(?})Cd!&<!bW(-zQBr<XO20%3
zmq-J2zZ<0dhLm1ODk`PFuO@|RDgQ1?zDt_O&Z6X5Qd~gE3n(Q@e4nKhXDJoJWt6;3
zO20=5_b3Bo`rlCUH&S{trD&G^KAaZ9X$k)|TDV3VKt%@C2eiDLRwDd}79L6I4Ybf8
z#nFrq%@`op6358n7$w3fjF2Lw?`MSlQoM$fui=!?#Dj2@lW&#cU7UQE6lZbrEGf?A
z<hfG3lV^AGy3o>tlnO8G;SCUul(CUAUFga{%0?zc%M2ifd3jKHq&buY^QC~1^^Ru!
z;jtj#o7jP|>`-`mM*s(0yyIDaSjHfLLrGSbI9qsL>GY+F(OK?q{`~54{-w)-$Kt@s
zS92H+x|kyC6}{Z@24#F*5d4=H1n|@$NW0Z@d_AWR7XpI#D=+_*H-XB=YtJIWFs=B5
zm3>ODV=M^w^D(!4RCogdUfk^!Dzv$Hz5Md!`ljE%CNaJyG5(RnfTvl31SQMl$ubjY
zw|VWER#?3ZR|3FFXb!YO87%o|MvWZxYrp?^bqG8n3vRrD<u|ZK$P>H*=ik_#C~JO!
z>v4hWgEqiQIkuEDfS!Wa9;}6)2ktHam(+Hu7Ol2KGf@qfV&I6_=mq0YRJ3@^;t2@l
zVWab~ad0^bV+Alangh7dg2FKZ>vqlQTeL0p&jo&&ht;y6VWvr6o!P&1X62QOM%i|~
zu+kJH<s7H?Lo+tZVU}cvbt~?;Ra(TBf-6#_N9s~=&lEfm#k{+5huydb!gYkAPI@G+
zj_|A_0#OVbK`A1nN7f=J&j>0I%84LmvWiUU8O}`BGm{NOG24FDXg_Nu#XkGl;Rt83
zj#+F!gtJ-aY_>nbxvY6E+Y{kD);v#&V>xv!=ZNTWoH|a5!+FPWz8`cu;AklCyqfQi
za1`$lCEZbu<{hG?_!;l`jPD0+32<w?b3NZ5;YQw}5$!mGGp)SCYlNZV0)5h)M~9jV
z_pv~}LURfoTBo)u53@2}4XZrb4vmd&;m4sTw(?_I`3VqLI5ynP{dQs^NPm4ItUsZG
zZV-i%-D6R~wjl1HtaqRSY*0=>bw7X#u-mHh4AW^#v_qId8)iuHQQGh*z+ZL@b7{j|
zge!meV_&P?nAc#vzq0d6$tmVlnAIxG4asjVuAhrrp)$<lxbbnk2gI-%#(>uzx`!4B
zY<m#iMT9>*K>$t<65hpxKiq;BK+HGrJTVjof#5hXG4KL06mExr9aF-)gz$$&44{0Q
z@V+d?6@-5UF$isY-y!_(5Q9+N(JE5DiquETzj#s}FU3bl`4Lhd6>Fzb@>EJ6Ezb^7
z@<UR5m6Bhj^ilEl16p`McSgn9^|Vkg#oHNSJJT5!Z-X6DM%X9CM;ZI0j1x+u6f<VU
zj5$oB2;vRK{s!X&>zv?f1z2>=7Yv|i3c$&b*Ej)st*VGyM)V?O<^Q`mS=E#h^!?W8
zbzeiXuc6sLLNjQqfECny2?H9tyD9oj4Kog~oj`tA72tyM!^8B4*`kban0}Zo$_TfD
zPa<G0sJ0^<*AK^S;Z{Uj*|&ZSZi{L;iU?H^;UYB~s$#+gnT;!i{uS_XfiH*1uM+xK
zrT7-1e~Yk1X5k*8e~++5W+9qYMYAqYo!UY?;|;7M3h_!g<5JEMdV+hN#d-s3R@8$q
zHH{vTMt_U!?O}SvVSu&v73wV*G|jg`pdqa8Q|9~bS0n6eg!L~Ytj8<zCw4aK9R%t$
zmzXbcMjq-oW_}!VhA}ZXREV1&!ktmf)PS2e;Ld1EW(#k=g?C0V#Yf)!Bkv5Q9;Si7
zL~6d^0u3#!?did5KiC3iSOBii@pr-*`ogwvON?$L#-XUJi5SyFOh9-wsah@laz_-Y
z7bQKjoj|G*r1WG`FIh@YCGAtmUdVN&k*YLwR9X;^lB%OpTu$1TOZjh*svA;zC23Vj
z_CT((np9RxKa3DhjgF_rp(t`AHD)6<0pVSgYL_(X+(+r{qYM!}i&AAt>3Ni1o|K+X
z+2>QekZUcVR0ZfLH;Az)RhbkwQud8f{x_8Bjg;O@Sv6BVkSh<TmEp7s^(UPklui#t
zs+>U&%8=rt^q`{vOG2kXxl&w6k1nLgq0DM2ZBt6yqfp=)t-MC75d9IYd?dvUw6Z~p
zTj|lQ^f;7hjbv;h8G95uL^H~0Muq4pj50-v_cO}<2tN;K8MgI@ZOy=D)er$S#84!7
ztEhlg)KEAOFZ8zuJ<T~k5iqXoztSIRR4(qHiw}e`-iyCgGoD=!u>iYxPZ;tEBZr&5
zFIva84A~Ls2998+;6e(nhjgxv5b6lYL5~PZh>#xi$Yg~~bkIW(_p|c-tRY$##B#>5
zTvvqSIO8~k;lBeN)qEcb>L@&Y1H$WJ{4j;YC$V9=$R{OXAxYT8`a9x?pW<fZgJUU}
zd^@H@@4pvQ?L|il+rIQ;pCyfnZ(=q#F=x2T$!kBo1lRM8->q;Bb0~-f`GX`(o`elX
z56Z{o`S@Tsux+K9^0VY%q<D3dyp9@-1h<8?YGEhAe8cOM-wGX%sJnf2CI8Zuypca^
zSQ9nO<$qi^2o)HlTa#$lB>Kw<DPH?!F62?F82>6}JPiI}Gy~*a^L;MJ*u4D03WL_O
zo&Lu!(i}PPWVh0zZ~x;5UEWu$bSZzGQvgiOL7Cn`nGN*HFg^nnD*rr=2iH$vXfz)k
zgF^$g6A(0C9E0i!YhuA#Uh|zXXgu0ZbeOGY)o|wsEJ<0t3sS3elhwP4aO=&O$&aTf
zhZ}+d3G+eyP=Ul=tjAtVVoncX@<Z5{qv3*h9kaNOdBP61l{=e%#QLFf=i7LX+qge6
zrq#H-8Xa5)-6Sm56P~Etc{^dTo$!R^&TT7e^YcVM>B`#t0?`kxtRqRoNYV*;h0Ubl
zW+`qa<;|qz`(1k}!@ZOf^csSgLmB2saS|<0q9xz=%4aO{8BaJfpkSS`xWahC2pN3e
zi?O)NctYWB+kZ2UW&6RvOA!BJ<$tlZ(CP`|IZimoS;H*rN0r|vcIaHwufzU79d;;Q
zK%X=p-Gk-(I`gN_oYYpzv0(mp3kJ0P#i#9}`6KS>I3xVn&^be?0e(Kd{yKh;prhlD
N{@VU44fI9-`hOo((is2%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v3r1.hpp.2D0C9495DC723ACC.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v3r1.hpp.2D0C9495DC723ACC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..c76e2c27095379a743a1cdfb17f25937f988483b
GIT binary patch
literal 13920
zcmeHNcTiQw{=a9-5w9y=>0aOxsUE6;Qlw}^3>Hv90X0T6k=Tu*KJ=LcjG`i_SU|BL
zibO%MMIs^)QLtb^u@E(4jaY~TEEvM$x0H3yZ{Ey%e`MxOG7|@P?`OW>bI$I!?RUSs
z!9xZQ4tG;1{5~5JJ!{JRNl2kkbf^EL=Fgi}02PXE+6qPVwu@r|=iK&CUC%FGx&jp7
zJ9~q!I_<78FPL2*-a5YWr#HPSu6bCSCjrm+PxYFry)r&4-Du^t;=8J0_LolFC!gOv
zF=&pa#b`~&(EDy5y170u*18sH-_Rny_;qdNgqy?b61F-;*d}z^);H!*Y40~nZeCyW
zrSZGd$D{$jnpD|u0KQ9)mu$*$t$vpfo${f!^5)f&?-Zv-O&{}QPQ=DNh6{aD_l{UM
z<4BLFM~N@D+eR5a*<M-Vm)2>O>7A_;(JFOAxvFbubVIhPuy3*U6_;Wa*uOnhFKeN8
z+VWV_m9SF1<IRXqqZiKHc<LM2u<}XMKQ2_9)w7;%_-1m;F2}0!D+8Zjn6RL;L7iDn
zpsoH#$HL7wmmgODR5u3i&i*Z8P-gB965g`<<&uK!%KbT7L$98@A$SeGW;&?bN1a~y
zz9O!kpG8NN2R9sA-r3=w_A}GR7oD2$%P)_T?!NtPvC)n0cSMs(W4hKnshkjhd5Z7E
z9`~LOuNZXt*{u1KW^FlD)BV#^x{-+i!P9#^w4JiA^tXr@De=X)r$;B7Y^hk((=|HY
zrIXUzXKrcp1jm$!C6;3j`;4lI_WLR2XxhEc$7r7SN||bLAY*#{@-n|go#t1*>)bRw
zHvQ}F;~P<TtGA|8eIvD&d!cS?eHF8XqOhU`nsJN3cNSZcaMx3do*wd<`{?QFt=_Ax
zXKP)t`4Y~)`pcf`X^x9)1J{^(S*PBz`sv|U#m1{t7nh9DxV)~Nn>?%HMzG<HC5^{#
z`TI9KjqW$O<yl|9i^=uL>Rz=IXD;3UUBonl3%ew&GsFJ)rDs;<(M18C<74ijthH-{
zyM{K$&Mci1les&~WAUs_6+=o)HeN3a^K}1sta{YEsK~gtF(*2Y*%y|uxbM`kY3kT%
z^Y<ie80_Eg^lm@bfjL@zcMUT>-X}e*%i_NGDt<G&^?P$-^_!{%_Y1Eq+G-Y6_`Usu
z#eM3(Uehf{%#8UsGRLnp_4IF@towYstf|a;*esoIJ6poe=M>95omIbhu4-8}y=awV
zR@Bk@`5O)#sQfv@ZdG!I;ckyLVf|J_Pg!(6=ftWmg$vK=Uv>+<RrNGJZ%NXJ^STcH
z&|sL=VVC9MAD^i)Q;)X{UX{^r`~LU`Cv)Fy>bL2GOL>9Cea9SLnh;liu*d6#rVq1s
zB|P^m+w&l@a`^RMpK3SSHH4Y$`+CVOb!hX$ld0!gzR~&J#b>0c)%oEM%3|l`OugX!
z<w5t^FO!ouFS86?_uJN&1<#Yr?<Zs#Olx?u;q1F-7ypq~u=VAo?c?v>|Kae?#k<qu
z6dw7Cs713T+f#>Dy2jS$7ms%C0t`(R(iTy<MO1fm3N7mE^k3J`N#%((qgSbbin1hQ
zy<}|N(J_{8M)?8d6%JphfTNj0GO{spYHJu-C&Xxd#q3G3+MqXm!+z0xzv$f2F&5pM
z{lAF${7kqC=unmxpw|LSIy%OpcXnue#iZ|>wZR9JC0p#8EgE)oj79(HXgdqTS+lf(
z3uVa^`)7*b9UWt_JsR=q@QB(N73fY`=IYHgZfh9Xaq2(ZZLbzK>jMW$4m1ty+19)#
z*KHoY>2!{zHZZ5;y`uhJu~$dOSbEP1yt^=>-w7Sig|eh0t90bs(J>a+$quy}+^pga
zfSR&|_XziDYZ$rT+T{5D|C;-O2zr<(q+y+gNo|dyGXhwl5G3J4t(WIJmHw{i0<`J#
zLCT<RZ4G;Vpj+KUn**k*E<mK@QOZ#|Z4Dzw`p52lxo38~KIlrxq3TeRwuX_X4*m6@
zb7ZWyKCqzV6j7Zbx^#4mW!g^nvH>sNK2ZT<%CaAz{lKlGV=OaZ$>l-LJ?eA-p)5zR
z)e$_fqhl;zEqe-oxU|Mw2Xv+^w~^ItDya63u_W}eo^2q_bLj%~Da%OZNS(HZk++bt
z&K51hd*}maN=_El$#kgg9b-uyJg;!+;gT0!KqtyFS~=RFtzqOH+t)ntx_;0^1-eo4
z9;~;A+ID-#SPrcSjxIAw+1&+bQI?&eawqkW_KvZfD11ET^2Qqq70{zBk$RC<Z4D!r
z__u7leIBK1gPxR}Bbw(>^Kb7M%assA#k_}K)vJImWtpKj!>X-e<QoN>-X_EhJ0b#m
zjY2Z$WZ=}+C^|#IsLdko4_GjC`0BEoU4Q|7K13N}(AKc$YqLh_Z)zy%s|}neIZN!C
zMO~-8V=TW9yYaQd+!jY2prkC1u+<}~u=b9z%$pJ!b<;tiKm*p9_{EPoq(!8Sz-S|I
zWIjctEx>4t9A^TO3CuZtH8NU_9GSBdX&o|JC&$|n*^bOPJsoSMV-1TyL^_DI4q^?{
zni$%6>0pLvjEEF3ka)q4e`pINTc90(Q4S(G$d36Yk)9y(1lchY6bt%(;O(DV2#9na
zce{^mnfVfFz1VHNXv?J2X@Tde3k#>3iA0JK^kM{i{>AlBZ#}f<U(7^$naG}h@gCB<
zhwRxGiIj)+@~}PsPz~0r!S?J!EG`3OM>J+ZfzWtnoGIF|P{(8=C_7>@W9-^{W$N@z
z&bMud6lxY~#k`(KQ6^C`jYI`R1@TG52*EKzIL9dhNfC57&Jak39G@4g&I@)ibsO}v
z(#8HnIxqO07lv@WAGq%aJ}h1ksR%d~fxa9U1E*p+E(3$hz;I4K2L_*$;|idv00x|1
z1*9tAWStVQR4*WM457fU3Jl@+G0=Mq)Exf;^nQ`!SZEpxEx6KR;gDE3g5&itcs&f~
zI1`dgsKfCQNRG(ykI?EzX!reQFl7B)O%aiPgnmE5AslC-;7k-QH;UkUC|qt7!Ff2G
z8$|*3DZqXlpTz-ZaS+FsF}Y0BrjB4sq&t}0k>hF{P%Y=L!ND~+oEaIBYOzl(_T%^+
z4tR%yIF1oXjHttLqDT_uc%2xqPJGYb_2}N>Uigbbfk;Nkh_t5<bf3_X9SEQ;D`3H_
zo6)s_y({m|P$KEDFz~S8$IJ#v%RtX%z?|cUKzaylnHrJw0?<DhJ_!e&gnmrSNNR#T
zo1i(z`N%gP^=D#7QUmgBK>eB6k+co_ZNr0@D<LTx`(@)nEX=U<r8AKxDJS({t2dEu
z3jw!<Fg8pg{Uq4_BzSUsPYAdtr#}#EAIRyCg@DI$`V+zSiJbmQP`wh&nD;3=QYPl$
z%nGUJx93ZYz!D`o-Uwxc9v|;4tywy3rJ=S0L+`){CL|)&K~)`8v$ql{0a+)Y-W(?(
z>m)hejvTflcTPWuX?S*JLQ{66W^7O_UeQ?nP9X0DD>f}6Jpg78fCtkqks5(rBN)V-
zT-lMpu@@PZyfuUlFYO?a&Iyj^gn?Y%H-Y0#Fp$5f3OZK7foz6GHxFt3{_XcJU0UtU
z{IFob9z;^9Fs@V>&+&E8_d0N6FGAG#93BEkj$Z(W7jk?OnxBLg>_bSp2FG24<2lYp
zntWu&?XMh#mZLC^8<3^}nQ?nujzgE@Fow<hf0tkHQK9_H4V<Z%T4-yc%`|9#qG485
z*oI6@q|1WkWx<^Zo=7(Y%Nufh3s~Kfo5pQmbz6?_Ld(0*oeA3hh5niz;PGM_e|`n3
zUqMf1Lk^$)?qv$TEa`@%1+EMFvT-44q5DElj+O~SmkGl;UN5+=7i8RNLZ39jjnm_R
zTO9CUqes#T=(+;-WgdZ~mC$`9^yD}lx~9wNxv)<zbmR19=++E9*ldt=8k?QQJy~=|
zQYAL4l;hXf>^1JmqB9MfqS-R>J!V^(KFYQ&43j>Qf}Da}*%qQTj~N>yOkk5DQj(xa
z5@ZQNilCb!7;^d!L9;_n&k%Go<n%1TAxm&hDXIMD_vIf`C(RW!xx)KcPNZT%Q!K|V
zf<ud(KNe_WdF&^2EwEk&Y?+XWv=*q>0u3Ke83-%`qnM^?Pym7F<hTMDRsds8uL47>
zKoIj~n&82)>)`~Z1S0K(nw?OVf@VV9OlZjIIZ%@$ryqg3N96Q;=#USc9cEu?(aH%i
zB2ppL6!Nu*NSC1Ik{rK*4sYcAjZo7lr#C_CCTPoyf=F>l9fvf0Jeg>0CYr!v36TyV
z`sYw~;J4;#nDQZ!jw1C@q~Y{4NPR|*E0MZVj_;wd_s|3udx-QL(LaZIMN@nq{^&e{
zNH3B4CDL$uELO*28BWCNL^;mGWApF?X2e7~i$|QrAsk=Ex|gvb$JepLb?nUX9jtpt
zPOruzs^!nu;ITD$0&@boy5JG-a0tgSqHc_6$Z?$L5GOiwoG9uh%IWLG5$nYF&pUQ(
z`029s9GE=^F5GY$fvOQ$aKm~CRS)?hpzJ6MFt=k0pdOI{+%mujX2Hsi;s9eLG#l~X
zhCgB1h;#m{+!3MaA9Ny8mY~WKthsBo2&xu2&WEaeXw6OU4OG36<3orXLgw5wo+I)c
znRC@<iYQZ5vz6NU=N-Am5v6lFA!(`LwN&tCBSz9y;B^&vv#BF#BlOw`y_roSX)p5H
zi@e#Ck@Nt0JwV>fW{`9YdmY2xOvgxihP|F)Z>AHMfJgCSkFjn7jc7vOWjx}tOok?s
ztwj?iF%PTBib>rjF3>?znlLa;@Z$qLDTJQn85WXG3gb@77y6|jVkwx!f(w%3Kv*0I
z=QtM*%!Ph@Vihp70){bXL{bGDS0P{M-@%A?a1y6C!?0!;&T$^{$wR)AO)WxB=yv%D
zN%`oLd^C{@>MROAi$XZAM1hqkh~qlsQ-^%#8(6|MKDJG*iR>p0Xd)NvQXIS#hj5&T
z0~2u&!yW@u1|IpXXSe^kMm=5Qg72Qneogn+7qIaK?8a12q-Jc~jNO>(nM&!lx3%td
zaED#-l;cTx<p1g0YtG8g_dH!Wmu|urxGfmWv`wUi{T2@8Xa#gz0S7bP5xN@pTM379
zoGSK86+PHSn+AT-XQvp*%01rNKBp%Rwp>JqT#QLEcH%>>!=w&7u_1exeLN%NV45RM
zIvzrYhtQ2_n=P^W%C_npn;MZu4jAdrDqBPvHDHuKtJo1KWJpL5E9MYs^pMd(tb#?P
zaoXc_nF|nUg7yR%iu8{h%F0kgn&CfVC{Og)3aYh&J~L_>Uj)@QIW81bg@Qg0D?6ZS
z2h?ZNAW}9|Wy|pesJa04d6+3hq!e}MVdV-USLC<_krvdQhndxwtj02&f*n$@3yXKO
zz=0hyu?yQsFxgO(Xf0)OF=mS>TO@3=<Nw`^4Jm8;AEo+wDLB0p`tkYJ2~Kq!e^j~6
zYwI}wD9U&S8=b-CT(OPVs1chp#SR%fx98Z+;4o!eIDFyzc-?A9!&mYm!N|hUTI<pG
zJt*68?0X!K;_rKbeP7^F?0utV-w3d9B{p;^`Bu>SR?y|9dJT|kz>vjLBE1FVtsLh;
ztvsm9On1!ml`dbb`GJc450<^cn>IWDT<}ewKm88;WAClHyKv3y|9Rm0<nWBTpz^Z&
z)VWQ1n3%9i#-tS$nX|p>zNCx!UBT+E(2E<}L&55y9M=d|H5^Ya&l@=Vu@j`mk}CE~
z6@BI5XT(l1goU3^r6Xy*y6oLTbzCYKmI`kC3)i6GH6CS%RE~7ZkqcWRh_oE*F2^n`
z=1&RyurT&e=43j9W5Dbfuwqum3W8mfZRLONSNjFC{X$Rvy#s>h0l}-g<$05XkGs$r
zHwj~#gz?;X3!qs6?8(DL5%esAUS~gjcdaPRZyk~9p?^L6h#4=Djv(VB$dr#a4|(Px
zFSaJqniMjwMW$^0v@U@>>yQ^)^VsnO4F~OK6MW#i#lgG9PnZPRJn81M{e*%sGXJD+
zXg`-=1VYaae3QlT>{ww&nvT^k_ha@^WtCy$GTfg%NgWg$zs3ETXU{a!sIAMV8B-Yt
zSq3?>z(2F(TWRa|*W>@rOQiptm&~znKW%PRm_-F$fJp(i;exKkq!!ySLC?uORpmYV
z#>!TW0HZR{hbg&r+h=r+<9Q`;?}B&t%>S_aL>uUvuJNw>nSSv<M|Wng)Iqi)x2@=X
z5n0)Brp2u+8~bPDPuTjlqR!vM^Kyck8c!ZD%i|8ZdMI30ySBxHmhp^Sj67NFAyS}z
zpb6XD6X}4UcR(=a?pZ4s)C#8DeT#rW5isR~Dh39{a$E%Uil8xf*F2=kLx$WP>yWAr
z88Y`<{cO;Usgd)3|F_(nB{bBlTF>eDY_mj^B@%W5m2|bHF1FyGxqth_fAI+xgs42*
zPc*qiw~Nl(#UX6vVEG}9bFIgljBVR?y9Y~37P$ZE*LLOybo}iHuWSZ8$F2KTd&p=@
z>QT1^)7yLzCDJ{?^qw3)7EB*=oc{U*i5@vQ;14_huCG?Fy`b&x|F^OJ7h~NYU4Tc<
z+BKN&$gX4K>)3(^@E@`9k8)g%dskx*PLC6f<3tM<;AuuH8pq4=I<faU(Sy-*HI3&`
z(d0B56Tf!)dH_#CN`ZSR2;fofEO0+7$K}AO91P&}B<Pd`2k@k14|3Xr2Job$8aY*?
z0W7G~DjIe=iU;svKfz8<@BlU}mUB>NZNC}7Y?#n05ZWv1u`<%p=ZlXT#@A>4<x2U-
zBJW)oq&$DWD_xAb8X2;qJ-RVR;q)3{`#&N@qDUE<fo90iTr^jPvH{8lGTZ`CiyWsT
zl#XQd+X&s3<7AAIv5dY4qdjtb1fwHz{0O5*92a=^?EYxwn?!YMy<3$esJSKW5L7$l
zI9E{R3TnQd9uo8p2^x-z1y!+}UN4Y(!GOOn7N}x@?D=ItZyAt1zZQ_SK*qlls&>kc
z;c}oVM~(}jst~Gqy8Q#x`vGb=z64d5<n+goJcb6`(i@?wQBH4ydQI}@;}D5MGXA58
z97P7)GmatEF{I}B3?gUb^h!i3<@hBcFOdQFq<W;PM{15^F^R=8dLkx?9G{qDFz<=F
zzZ1>Bs)0i_@Z|N^8emfcTp4D?54w!BUcg}vc;f8!u{mB1kN<KGS7P&(v-@{L_x!!~
z{Fk+d-L9f(!er%S+3EaKsCo*`nBglsZd@_LICJ$%dHJ(fKm6(QqjSxsI3#Sp&mXS#
zT=uqJ5<J2r`7h@RtddDxyY*%q)8LgG3&QL^DEpcw9gFoA%T6p8`!0Teh}*iUvdNQQ
zF**ipj>&NX)Ju?`=OsYj1UQ_T&2^_2w~vg8*q|g*qF|HAPlIUzN9dKx&n9SD720g$
zr@KVj2Yc<~r@KTd!ImZ1g{Kf#u=N$(hvRBYs<E2m25i}YU08}iq@|+uQn3%m>qN3n
zR5Q$O)zB2A^?qRX)EWD#onEh?1$SU{foDEdueJqG-5kH-PjB|Ox9PunF@2vF)$T~m
z$Uba)N$cgBQGHlVm`D@#C))9vc!Yk03{BOaYR8rZBF)mDWydV(zQ<+1xBHERznm+u
zqXFs)t(QEREBxF%v*sVGk}lF3LN36$z=LIE&jT7qY@W8|4P8DL1T2tU^jiS~R=^-;
z<!pCO)9Q)JiLx6*Q?;h*a1mY<LN5vtJoULIxLy-{INk)TH-X+f^+^TRsdAhN95R7B
zr>A2n9lNr^zp~>-Dw|Yo)5nd0TjShm;9E=<%c|vhz%UPJcrJeo7#@@31V|Eiwfxsl
zzj7PrSd>6Ta8XcQ6o&Ef-hrw+a2QLS>bK<=J00w4`Ij?_2G!FS->e$Z=ih2FujVH|
zsja%NMW>!BYNd)ctY*-9*k!s?^k(~(H(!>0Wzcl8+rPcxP27-Umgf(@{+}1#o35>%
z{NOY7NE&xT%tL&b-P0p6^U*%6<UyoOg5f5?g<I4v!El!xX8=_OkX;MTfQA{+g_%9Q
zECCG<$?;aC+KOb?hSRV`8un!tO1%qP<YHe|vS6{7#@zOs(#&)Ttxfx;h~rtYx>+;m
zUs^$5nEa_{qnI~k=A_oZ5jF4AsEMsVtXiA)f7a}WLeY!=wB$c+_)oj{KY^5Uc67Ao
Vi|L`AY^|(qdRf{jY~K?V{|5#k*2Mq-

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v4r1.hpp.42FD0DF31A3BFDC2.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v4r1.hpp.42FD0DF31A3BFDC2.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3c798531443192eae8e39f1707e83a3a1509a29a
GIT binary patch
literal 3270
zcmb7Fd011|62IpX1BT0TLH0#MNC*KFj6sk^WR*xCC}06=Wk*1)tdB(&aRWgG6a)mJ
z*0Ko76NI9u*vC_(qHJ2~_s}Zhf`YB!LaQL}B=s0x{^<9S@5|gf%egaW<~K8*LKl~I
z0>d0bgh|^%;{z!ehLP|e7atqpj$>Fo6~mIsFBh)gI%KAN{cX7i-Pc&>`-y?-Av1bD
zlNYG7mlc)gQq<F!y;SRXxu7s@>ZhEMyT3KPJ2CF8{;YDgZT5LjNa}=uc)3dU8vk3j
z|2(_d^dK!|B0i1Q+dnv5wnCvRE$L2m-tZ9h_*%sW_Gh;3`gp&l(}-UdtC*eWvU6g)
zU2D&u3-3jq!}tDIVBTI{6}Hju0yjR>eR_#CeWc;`s+Y&2oH!YJ)p`+ak-L^UZo73Q
zX{C})Yk$D~_xTQnBQZshn%q}q1J>p(%a1?d?=8Cda7H#IIr-*{?m&&w#+oC28u1R*
z+0N9!yn*4J-wi+M`Mv%~ty7zEDEnOD;yn$gy(aIqJZo2vA*v0Z?I`}&X}se6L|s9q
z<>5z-RbjLQPwnKY9oatbd%9Y@<Mt=tunhTj{K}E@dVP3i`$Y%&<X$6Y-=Irf%BAqD
zF-8ox?x<g~WE4*=wQ{LAbx=`jVZ=QJFV>zIJ%^QU$puzlpGnyqJ@$Jg|G{eS`X<ki
zi2~DoAtxv{hZm>FH*ngg(uK@7hwf?QobUT$E91qj0WG8YtC*_w{^(Km?9VsM_Vze7
z>y-|h<j2_Z-j>|Cy<W+Ahk6?$y5Y5d|0~rFi@?$1Vy}dN@j88~ef}u1i&7fEZnbYn
zR1D0i+EJ)k*m|;GsgHVKmmoi<X7A3O+wrk$jNw$R(qf<63K7d)C#UW7)R_5CUv(Nq
z){I`{`aW|GNDHiA<M*Jp<Lv3^kTPcaNgJApSHu*vFye^~e|>ViS4F|dY-aHp-2yvL
zuaHx5$=ibb0fteVzZ%J?c)Up&msP~bO0cLBus%I7Ify>^BkY6suf;EnE8}v|#CDOb
zhL|8V=f0n?-Lkh`4POA&hT4X@VuI9G3*R5&rsoZ+;S{K5Xft%h1gV`rj?cKcz57xg
zr$RNG##Ry&q+S#F$|ds&`)hf8Ayji{92GG^>Qd^1#7l0fp^Eq-sICVJ^+4y-1DUiP
zd%m(b(V|%mSA`}lJr-R|kovApiG@L#p}9OxgKATnsk)dT^{=UpTKs=5dZ>;QP|eli
zT8jx%|2CExxQ|_Rm5R$NU?g3ZZY(C)LLA6o7$5-_8yXf@3lm9Mov-Vdv!M|u$V@<F
z0t1v)f~)~@H9#K~g8JnjJ!ikC>OjK)<p7?<ESiWmK^!s$v`@7{fgl5@5ETlE70j}a
zMHVd^q7WowF@uR}LXd`w4VkEb1X%(!OMpK520@krt5UEU@k*L^9L#S2x>bfkrkJFd
zx;0s)C$-AvQOKi!aTFM?PYCZi5Rg44;%$$DRfyO8rBQCWrj-vUWP&hZEjngk*L-2p
zvX3SuW+rQJ?KJGV073c_{-($`iV>%Y!G-Qckip8qD#*z6((7$^eU4Pav%3MY8!Se?
zmWE7Q>)~k~(3iQp`8uVpThE7vE%TN=t!U5*xV9aNw_0qqM4#ciKF*<P`K=IHbIp0k
z+R~xe;oN(=l9V;dC5%i2Zb@Xhz#<o{L?1;sr_jXsmY8u?bDR;%3DQ((>W&~ukX#4u
z8iYV(1O15FK=OflK&b}=>V);*47xIw6Gjj|v2wAh$UQ;k0U{4*Al|!#Iq7K3?uMxf
zQV&{!#2X^B`O$}dy!N-?^HsgRETBP?3D0gJwy@Dz-w!3ro?hH7m?u0o*_5e%w_x-e
zsk8=7Y52QzHr7euAlUd^C0o9>6pA-5-|UHeM``M7v$`X6pqQ<~)<wyVIuc~$G`uzu
zic5f539v!-+Zh&D4Vahf;61k(Z*f9mf($nf|IEd3W{emsi}t{#@fG+wNR0dhesj2d
zV3sdIJ3&?feic}aIyAOJrZ(ZvhXzss=dP96^D=MNQidQ|YAjteqzLjb&^`?4sCNjm
z2xu2c_}mvxgW+MZ(^AkpINB8BWj_W>EOk$W3U~z`ji-sQM35N<8D^+EL=%Wp11wR<
zv`-ThL9G0^$)nVc1dDkV_C#|QF&c5Z7UtMp-JL5*pNF;wA8{w%ULum#1-VM4r)4aq
zMI&dj4&HO_%s3zR@TL>(#xTqpm%>&2YdygQMcPtq=WvyPQVBluD7P{+@!&_df=JP0
z(-Be#QlKu-M)|pDLnj|+zdHz{637WukoZ#9&wgjDUq6wG>iR%`clV1oKh85hKh5YZ
z@bWor`?tN;Yr{(Q^w2#j<+6V;!N+7tXNDX>x-N2^vrAl8xqg9iOOQfMA!`ojAY3Qm
z+<ES{63%0}hrNXJHt;r+a2pIZeCE(>g8N0ZD^W7<zA5(%FUiT8X9|YO)!vLpo|1zp
z;OKJ%Xv9IVYVmj|>*xt^@#;k`hb?;@r%r1Esm#1kcdllQglfRj&D|d_KWzHrA7%S2
zpeI+tb<S86DhodkMf7PnVnwU`Y=y~{(6)jfAlTsEN6Lxd^c5Sq?FXQPEWpYFfvAlp
z%P(BvouwN}1@v0^`S|Yr9s($yyG`Dub*8I-__t`A{GWKA@Vw!TrngW$cO`welN<CX
zM0*_szNvw!Evf=RrYWbHpa4Xpho}NZsE=p#oj%IBN9)e(P;6*aps26nVuRy&q8B9*
d`Rk{QUNB7M^D9oC^tkMO&!uEcMny>(`yZ3ulsf<b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r1.hpp.7AB8A56ED737D765.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r1.hpp.7AB8A56ED737D765.idx
new file mode 100755
index 0000000000000000000000000000000000000000..01fe10d83a1853daaab2ad56bcf35519660f1114
GIT binary patch
literal 3766
zcmcImd010d7JugvA%;f?nD8LPpaLNzMiv7^jVQ7NEP}v@s3?o5fI$1P6pO+LxFcEx
zM$m}Vf=XH)aYKZvXiJNsg335nq!zVK$6|F<tTRyUC31cF{D;3L-<NyedFQ;d{LXT3
zWJplZF)l&)Cxxs@Uy+@_A_&3?{j;(&Q<jqivDl0t)@Z-_i$0^ro%f(AuGN2sPB1O~
zHYbQ^<Z?_bV!sYr%ePHUw>lJDSlci3do$B%;CQW57&pbuTPgaj|LTqnl2yM)q;bcv
zcbhDeO?TWhWbyQeKVGprPo?yg_}VK6o^JMSE}SLqTg{!uKDkP0Ge(+N$*<1(q5eC&
zjSdsMB{yF8M;)H5KU`jP)4ZbS+2I4rC6TYnY=_K`?79-EeqHy?p7M_l78L2&ua(nx
z{_CD;W<57UccyA}#KU8s{PnC<mchRp(%yTdv8HuhS!7)(&*a8w$XOEY7;t~v$`5C0
zn&*^#_C1KF-|<V$GSm7#E3fOd9`DnflNV(AYS(5>a(vwSxaG0y<BoqRQGa+-us3z-
zL*9kUDq+IW<lVR1J*-^APS-o|<0o4D$Z4W0Z9`WqbLT#>(6=rV?Tj^#(+>#CFQg3Y
zby=J9@$r?%|99uQn{51J^#`?y_L9Y4`L3DkGdubl%MIsqU7i<DvEU~c?09%}X|`i-
ze0xF2g`<6$PbT(c(I+x$oGUk-eL7QB+vm0AU`4+yYIY>ORlrtMxE$qM-M(7YKV`uS
z&XwAE$@U%IJJYEHVR5svFMpZi`BJ*1m$jeMS%0T1T{L)L&#F@ajY}TsoY=Yq>qSn<
z>5X=q6Dr$9mqN?Rlh2r&H#PcfoiJ2W#qGUWC&?A<R!BjeRP@Yq<EI(^_r3DN1ahs1
zf3Pkhn*Sz5*1o>S;*YGrrM$uFm@!wsmV8!syR|3viey#aA*uiU+}Jlw3tZ+oyw3OP
zO}`M~P?w;sDt7!iGA1r$@t`d2$gaVn?!r9c?{^!!h^$Z26FI2XOlMksH$O|V-JCSD
zB4{Z`YH3&qHmXPTv)N&F)IkPdPO=bN&XHRi7R+|*4p;wbyUC7A3K6>#EINT6*|12^
z?H%);WuCU>k!&P#5%?EDGO}TjkkU!riSm=%xTG}_=>m%`q~+~~MIs72?zT{Rq-87;
zX#jo$((-o0B3n;zJynHG-`kLuNW{&~ZH8gN>^<Mc)Oaow?XxDwAa)fvSHX;t4U6pO
zzj?fNvStg9q>zZ3uXZ*pn0+NLZBTgGy4#fG@(9|=%*oZT5MxO&Aqb$Mn_k!O-+#Ep
z@<?;^y&vDt*|6a6f9kDWR6eg|GnX8P*w?_}8puXAEb{n(S6FlZgEk(?K_VKyhBhpi
z{Z!aF5S^DAYff4qwu~dQHY}L^`xh@mcUyIm0@4(*6+(r`uweF((APd!v|1z}?GRfs
zUNX_JV0Pw;m02BQ34%5IQ<49+1r4SYy&ib$L4uv4=px`18PlaemBJ*3UjbBwG2H^}
z77*jyQM4Y|dJyA^z@;~qV<<TcFGb5{%K~wz6fK`E55#3i(e88HH8|50?K#KOh*Aft
z=i(evv|pg#T$~Y#J^_jo;K{hs3W`?n#IBg-OxxP=ak2wR(c%2?$xK9A;A>$rt{n^a
zEqhZdJbDl`3I~NV#+#y5E-D4%Rd{H4q_K)nEkt4!6s-eK9jJ`yP#wd!wC$ebe&fg0
zB#UkYO(O(jmsxZhXxbncyNn?rIfjm8(CY$L7f`sT#I3cu)E-)aB8yT+&0^fD0c8!$
z!fs4%d7gf2Vc8)hw#aFbPn&7Y%T>SqI-R0ponqfnilm_#<H`I>$TwA(YCaB_MSl(x
zK8HzI*Oc%ijprA+CBUNdCHYcp63+@YMJqW<KK`1bec8U&Oxoh3<Kr3gNmG+%GE^am
z3*ntYiY^63DfnOm6x{-@Eie@$#u=GJ7#h#u2ir{T7I`DTRBF{s?10KSkw~zN%Ry1)
zEAzduHWo@$fDQt2R2X;Qj7%@A79)+~^5S?I#pTcTw`K}3J38Aat~I<hLTrhmi$Pp$
z99IP>DvaaOgR35<Vh=bYvk^Oh<3*mmorstbGg#^Iu(?<{YAvPNiiHw=iak)Y)LJUS
zjhS}ou<O4414~9h1_KGCVc=rxC38(SJl^?6g_KGJ5)qz0C65AbCa=sKdM^vxpEb2U
z@pt(7y`tH5{g<bO&1;X1;=<qJ$0gRXe>sF~>A+D39yn`wA4Ab4;9UZt4CrOxUB=Ly
zk+T4MPoee;X@ocoE`{1Jqz&S5614l}_tss?@fj6cT-hi)!zT*Xf3TorX9w3OaFoF6
zPo9vvW>kBj!0N!Q4wN{cW8zostn-NnP{W4F!_-Vz`#`=A)L887%AYy{I9K}2DB6wX
zhVGnie-y1|sf{Q<mY)%&VQGx0DqvNC5#0c+24ngnur4xmvwEV{gDo#hkzX<F7;DC_
z6e`6AQ<q!{JbX^gts5oy%R>@{*5foA1%D8H4l=#??e%MV{i7EbM+G0FjfUj#d40TE
zeqqJAU1ty&=z4&exJOZjhR<yL`qsvO+M4DFZ@yQNw_a!?lPkAv9@Tb_R`8-fwVsU*
z3z=A^G&<14awVSLXch{Ta@;{^f(X2fC|`lE9M=y;`wRT#*jId)FrjE~j<<kOkj_iD
z$1rn7ZWlPBdj;p>t!%v_QEnn8MNGwF6kQFIs$nXw8T=mMF?<8TU*iA}mcus>{F*{j
zmemOq9`FBIUf{iW%K7ccNIWn817>(<^RnMj&w2muFWwrid~XQ6%x^2O8EmzD?=?PW
zk>%k_4P!YkFi(TM9qcYiyq6@LkKir@Ss?^s#_)KR6&uqfz$*bc!!H9~nK6A19M6F-
z)A{Y-*bct9^YI{KaSYcd#vAE)=|ns_(LEw6d?8~(3sG8Fh!Orn9PpYQ_L1oS?rY|X
tw1nY~&&ph$HDmaJKfD-zBt5Qxo8jo#7R<8c3v37*%peF`=25^r{t53KQQH6j

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r1r2.hpp.AD9623F82BA99EFC.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r1r2.hpp.AD9623F82BA99EFC.idx
new file mode 100755
index 0000000000000000000000000000000000000000..952c3bdd4237a7e1e4312fc747c3b796f50c3197
GIT binary patch
literal 3784
zcmcImdpMM78=w32#uzUKGre*ck@M*@45l^7X$)#9+KTGRxspSn^3}3QvXRtEOGyWz
zlv)Z~Y@u&;Nn4VR%C>zKwOSqM>#MfD$K>gq%Rm15-s_s@dGGhW=Xrkj?>;<Fu)nWw
ziw2A39qpf%5R)1~vREuF^h-%iUg1ZuSfmPzl~(%8yd_E39e5*7=i4+l*#6ety79=K
zd6r&<9^NY_73}Q4*D&AE?d;mH>aCx)KmS(L(NLGuI_sE$(eEikXDp1bpL#E@%^`T5
znf5!bmc*Qw4G(4{Z|-kzcp=@dzD4Ej*Zc3cBo(|^raIlSiB?YJc*JJU&rG><)N)eJ
z$pbgnMz!5(3Jl$ORjd+~=%7SZ4e1<od|MRuPU+_x1D5cq>!I2e(+fiz`NpN~{T=&m
zb~k>fraWj-Bh<+aEv~VBkPuqWzsUbB|26M<?(CeK`2$_`8o~N54~BPMjDNO!K-^Pv
z-D#PR(c9$B%DcA>EWLb0HK=!$G`4-$bkmcT`4@Yp206Lv#8{V?)Q%_--`wr4>E0kM
zsaV{fz2n8*T!VYRcHC}!QM<T?b>Zu(Taq)!UT5it^z?7fnEO<B$hUg#@$9&fcMq)Y
zYpu49nEI;1EmSu(;^5R<{g3Xx_dWO2P@MeJr>axILrRa<|K)W>y^ht(P4dzdtVmPe
z)cnJeMDD^TKmJ0-h?KjvOB@rgq^{d43@nrGY54ETs9Tja>UaKq=b0?7_;BDRfocD|
z<$;g0lL}2gNR~OuMixCj7Z`9r&-jKkV18nOKToQ^&-=8{Hp+MKkk`=tC6&bz-3XKO
zNeQ?2ZBA7?;^FJHIy9F1G9fxE&d)Nptyj2NwB1$ua{u9Ob-%_RzdNy<^Q4bkH8ne|
zs`+1r&BrSfyZJSFj&%{8Bm0-Gy*sScH?Q~i|ER4A4l~Z4krsTSC}np{{i#dYC2vp4
z43Ex~1{O!9{4*hvgYrq`2ffT2ud>V42vUnhTXU?n<pYb2at=KjcD7&j9BdPD2{okl
z<awIO2dvgk5|JjlZqXWq9#UTg^{YUSP1uCic%v6rTmu?}HZth~^)6)cdBP?_;heDc
zT}=*ZgbFgT;8<wN2dwrj+ObQirO!f}n1s|NU{(TdV-q$>s@g3rH#)DcLFgcpYhZK@
ztj8v7vW5N3?(dI-_LB)sWMV(XeujMbqV8&VU04366PMs1wTvgD<pWmd%sQSr$K3S<
zpHN0>TRmF?`GD1B#3G(3RcOW~*ht+1ycRGYo3P2RZ_cgTbo#3*WkQ3?qNk}$vyl%h
zRRWY)ETG|1K)rv-wfOa1LKUr13%pvi#^(vI(bpRlTJa=sh)+yGCT50a#_|EH|H|)*
zI`xp<!zC!BmhdEI@&T(qKC!It%i(ya5$Z^7#j(<s4_KWX6Pxl*6?GVo9R}XH-_$Bo
z^hV%r1oJIp)KK&$;BHcs^MT3-6GmSIRFR_G4D4nQV8<xB6WE;~z+|bcnZEV%+T}(B
zMF;VMY?)<BfmaH)IP5q>364B7F^?2VBC-o8oL}k=-|vYl680k4wnnyQIA9blwh-Il
z;G>;p2L&tcG`kdn@lF(74UW|yR+MK~GkRPMz>yaLMxv?P<K=L@ZzVy}jUa0TKkP6`
zw}Y%5{IJ8gTG$jV;)r+*D0jBIHVz&|htCfWHw`k98JG7}QgpOa^bAJI1VJWPG7jW}
zT|T%nax+*pgA)e88LLH%6mx|DICGp80C=;#wXt^;ojN~N0YDlzO^?|i3j|q;0E)n_
zND)9MSapIE_LwtP=@@{Z^>B4uFKvfH9wG{v$t-&SL<e9dCf>IBb;9ihg_X!G)HKwU
zaeuMt;xCe2^anRXBE1dKE)EpW!EBOrIZQ1F6TFsvP_)eP*t$F*>2&k-FVx|lqso!@
z8Vr!4tvsxxxKb$E+QV9k6Of`EG#qsBJ{0Y!;iw=<q>=z8hMrQ-0EW;yu&V<{92JUg
z1-n*o#3VWTJ<3&{66%5wiY4M1cmuInBr8I5=^Er;x+vWlb48*g^{ECa_7BgKC^`>Z
z@?bVLr07C$DP&~MSkGi!Z3O>DSjxa^2mf|hieW8$Uo3pmlqH%lt>bkCdgh%g^ieHZ
zXj>RyX5$MZEqwk6Scg8ZlCGMIVa2yqS?%37vS<Rga7S07XcvwPACp4S3ETvIycTDy
z$77eU6%g{?{^j#n14DH?+Su`D+#*n88ExqJGexhLu9x90@&c4LKQH(sMqZfl%?xk{
z&U@x_b9rp?hjB66_UcUQkhjs}amkiHP`5h7^Us2DRP5{?c#+w(;lu<~><-t7Jj@=b
zME|aaY1JUaAuru$^YGiYHLepz2!|9A@cC@Qz8=cRE4ZxTDxCnq(NlGnR*7ZK2tft>
zqXI-2!3n|J4)R>&epIM|Hh~gM8b$8~o4p{xW<SS1xa`BZ^oI&X+mrU_cKi9EXbCA%
zkUU9G1xZHA6r>U$OF%)c1+rFAZUM4|k((sOT93YelaFvMWG~caaIK(LOvX~1MTXGz
zbU?)f&R_UoZdWR-XCvnif$Jfr@-T@ge@2h<pUP~`$DG~SfP6%k3#`1@eXX<e*>KB*
zF~VhtV)XgY$1B{<zNM?YM(^ghoVq<+8oazXH)q0ydrSGlgK<aM=wP%GScw#Ro4{Jc
z^fnP+WP|Gm9S?kG1<9T7Zo~98Z@#w;?rpdpNRIp_f-$2`U}#~CI{`X%LxL7y+9|pe
zLP}u)PL78eXT!C0or6)V=9<iPV$90Gqzs&}8Qvf9GkUYZJH3p`a0#<Nw+;O}oEtui
zIVw`Qsb8em+^2(CTZ2U7s`X8Jd&cDVt(xOLzwgP5d>E}a2N9MD)|nt>AZ3AdmZF>o
z+&r*h^o77JRFu!cw6oxjLrc*eFs%dJaXJ1i@cGCNTxIZo-w4MXW$Ng)4|7MM>j8Nc
zD10wTj){*Lt<03<s1&!+7vAU~e+c?1|Dwu0vax~G<?$!8CSwJQrOUkXnb-dT3&(dm

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r2.hpp.D28171A7A44ECC24.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r2.hpp.D28171A7A44ECC24.idx
new file mode 100755
index 0000000000000000000000000000000000000000..e9dc97dc49a19e27e83e50d29cc91cd93f90e4d2
GIT binary patch
literal 4606
zcmd5;eLR(S8$Z|W9;f5R!9nhDMj?(v2M3`v<mIG9nN}+=AxbLZiJ~>_DMV@|ncfV&
ztdBL>B5RdxZC2Z|-m{JAW%48~QaxUJeon64_tW!7pZ<G%K7PORy}#Fe-PiAW{e=ey
z1oZ0*1b%VBYnI2aj-dqt0fB<F)u~A(lt8dXNdC)D&0V<S<|N6Fy*VF9|6Qt2Ek7~+
zkiobs8|(59@9$Qm?$Q1%=>GQ^(T8rX|FnDErmg)4TD2$qxp&}?wMDl=YW5CIGk1>m
zc%x1d=kc}Zd3wv0-b1+q%Bd+2ch6O>`%R{?ecoVoWOdVu#(4Xub2Qp)i>Wzl?PIoW
zIFp~6FnQksYeQAL*0I9;s>1A~omKaCh>KG#t;B9!dKy1iJvFrN)85?So0k7xlvW9C
zJ8zH<o0R`>V(D?Mpazdpjcd1lJ?^c%)!J?)tH0Hr*ZWP%ANmsxttc<p?0hu(<lApr
z^ma&w(|_!BdXysCo03+P@n=Yp>*|(m>*R;#*377_c-(t(wq1nw<f`z-fEoKU_7u4}
zdWKhR$jYkkl&k7Hl{)OXhVETkde@##s}8?wM@z+zw1T=4Vk%GMPAZr^A>x^pKv8qC
z<gY1y^DLV#ZLxV6)*@MbH!=F3Ps={N_GWT*%DhweF5h^|X-mnfnbOvP1Fmaz8upL-
zp(|(M!|<qUEer4GmG^XieB(n$qqv{ebzUf>e$r^$V6(t-{j#v9zxyV9Zae98;XU)o
zGiFEhH|Zu6dp2l1GY^`R4>K#*%>R~J8F%yx%BJDi;LbB4UB7K?T%gF(i2vo+;=8ez
zpB!k=EF4&NX<vWVwT|p1C(N>M-~J-XY5wDi8BV4<l5@`tfBBWkOjir5YE|d7RhNsG
z95{ZpW{mgG1(~8n4y!g=|CVXm_OFl&D^rW-W}RF3ZpT5oNiNjP3p5RWB=oL*yrVm*
zCwKqucNR!HV#P~ttY2z$t~haMI5Ti}Otv(A)5hAFL8+hRKIv|jnPh$B@MOo8`azfQ
zO;^nN8~5K_dis)CL&cZ~EkD)dOKZOkYly6t_goK+y*F{_e&o+ii5J<-w9MtPMI<}l
zxH*_{;@GA6x|F#@zz%`#5SWf`0+P5Rf0|Pdx2(w=$}*-5h=_}+i<f#qdd!vM=fCrs
zvPqlLBD72-(^D@<i^rex@DLh0>r%#qevgvAN7;{VDB@HwFK1b5-ZpJYkBG=cazpik
zw6k*dLO5=0sY&S@3Rqh$TRZh47(;=EKmaUw)+{{azkJIg1L_U(emz*%gV*SWIz0RA
zI%-ztdFoOYL}VCrhsiNtY$)OrYnW76Wof8QiHV4#$k9-}ARS82`Q+mbX8PI`O=u^P
zlYx3cIx4qXywS`|&xkT3w4J$~rFua+Wxe}T>oQkgF{MrD2uXyAdik6F&TnXm!K}B9
zC{02uScRE-K{|KYpSG!*zCVj89YXs_{A|<<(na3OGCivfc8Mtwq1!;x2DYOciX2=J
zze#2)zp6_a5s~Mh^Bhb@Hx&7D`1@7P!8O%lY77y%2$G9L%Zm*~nr=U8U$J4ALQDyX
z$ODi(0GrVbMS5JSK3AxmUB#3x5eb$A+o%_$2V(wO^RD;93NghHI#d#BtzM8$jZa8h
z#0UiRv|=;A{D{xSFzikc?*tq4AH(hf@h(2z4dUH=TnS7iSaI@4fjP>@b->i|aVrR0
z!2%D@uw5YR0t<90!#)GyGqAvuW4Q<!)>GtZgr&i-0m1-73=_kqMW)57%w-cl+I^#z
zVY8H3(>PQD(h{)e46FplN|?;yR<LgcB^o3eEyll#V6+&4QWqgU)=Hn8cOfHDN-=Dd
zB+7x))(DbDaKH+|TGbJ$YZTvRSh=q}5X+EZ9eo`Gu^brI=`E)K%zlP-^>q#8j_u~_
z#zP*y9z5hP_jl%8qLQneF*z7Ebb6=?bAe%7z_A5f@W>3?4vy{M!o{Em<UQbwG0>T`
zxoqi9MXM&mDx?bQl+oH@SRY#-IcHi_NK`oA4<U^Zj(%X+W^ir>4?YfQ=6JoamwhK)
z`hA%O&0d9|s}PK~(d-=vx&y&z8_o7XU>^jbZP@KJ7<P(i3NJr$L7WTY&>qof2gCqa
zoUJ^vwADmT!nfFUu@8zd>}ha44L)e5!<W6wuSb4ZO|&OjCwXu=m292NLn|h%a7Rua
zH|S+fWj9eY>odz|HoAvqo8ZkRutLp_p>cuEpZuc$Xm&f;Y=?<x0wx6|Qukyum?3FN
z4}lKW13o8WBMti4UfJpRliR*v{wqnHgKp<*n!e*kKH}r;GX-t&u#OeP=tZRy6Ssrw
zc2MHCX%Z#>W|)pf;;zN872sY0Gf|LXkAV9TKCT6~S_na2G&&8?N`^#Lbrqs9stk#$
z>JCI>R2dT0;68}LsN(b>RqrM8;zcz?7-GmJbw*?cuj;dv*}SSR1!*b2>W_lsQGV5T
zfqfS!(E!otY(j&v7>RpF1`|GJh!{N#I%pqODG?hPP@LFvpq~SSLtx%Osp4*9uIp=O
zHx4;kq#kaZewn$<5_=2jKy#uZ(e(`52y+@C5@VM<)`GP;pEM2_R%ReG!<Ukq4%&UR
z=kbD9<^bvik$Qll3D^Y4$w#IqK4#dhFnuco;#=5`$lD`R6vYZu8XLQ}KKCF-*%{fH
z;V}!k{I4b^qz=B6uUj59obB<G-ua(e!RQg~FNQGc-}X!A!wrpVYhHP}l4!%&N?~U;
zF{ByBH-iG*R=!8pv92lI{k20@kyEhsg0B1YSAH6kgXFlk$XkNN$*^I<Faxet=fT@~
z5RKi1VGBWA2rT-SVavdxjO!i@TL<zwaOe1EVD}7Mu&_j<cMx>c;p4Rv?NatC$@#wl
zr*FU&&tGHlw3+@Yb{}bwi)4$uIejNVb`rc%tUjUrg1_j8Kghkog?7<>aeoyE7}k&W
z<Dp<Wn1@2?P#%h)BY3DD=z8Gci$GuG<2In%`1k?P4>;WDZK;2IbAKVRI#HNtz*+r1
zFz<sgVr)YsZ(iKD&>c;-ISThWRn;qlhyElQUfjx1!`Fj0@^Z!JLgLV~;C`0trQiM1
za{hc*-<j7=sOukG8>^X48tBWD;(A|scw%Y(rq|ZkT5i;Lf9YW%37T9gSK`DW*AS_r
z5<4HchDe>1*x<<R%h=O_%TX_5FCOwX@|JO2fN_8W);+_97>77;9Xre@OoqALk#jOy
zf2=Z`1S`fW#s!1Ju&2Q46u4lD;mgF;Bez?0^k0i}+~zxewS|o4G;!Jtu0S%x89bB+
z;ym~p_lob`&&>;UeW}6pXP?V49yqJ_(qlA~e0k*Jj9(tOToBt4$Id41D*=ZRuGukc
zDL9n!aT(Z`!F29q6(Ftv8HbO6_y`}@g1DBCZ@~B)Fcs64VLM@bCrrh*$FNUe{1cdp
z$v9-;KP;TJ((M1vu+;eEn32|ymby62YvhN?$fEwESmng3@QBzGMf5n6v8Do3Bm{zS
K+|RMx&;I~iBP87b

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r3.hpp.C25C8A66C4CF26F7.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v6r3.hpp.C25C8A66C4CF26F7.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9300b0337c3440ca2864219478eea864a2e595db
GIT binary patch
literal 5430
zcmdT{dt6gT7S4=U3E=_(5+2GWA_4*-LNErEBJW3(htPs3ixh>TRaB_eYNd;o0tQ8V
z^RQT}MNxT4txqZ{ScS4Gh%8D~e2cCMsJMvk5IKAAwtx7Ce{Ft0X7c5nJ9o}G-+X5X
z3-t4oiFrKlIf2U-#DDQA#pCe|$Y07A$#YL)9xqdy$6Ho-;^V(9Y;zL*-kq8*`S&(O
zP~u#ZA5`soep2n*CTO#;x^m?2zH$DChIn6ovtq%$Ngiz(Nn>Nz<vh7^MC-HU7abY%
ziqL7h9Y;o*gztYk%Vhfrin&(yY?_w-yw{aA#bKA9elpfAAnw`Xv0tapdU-0cH%R&Q
z&~N76=%d17=lNSV$c59VHA>I^npit^Oq%3m#5_lJeZ<@F{67qESfdZU^RS-R`M6#q
zyB?j)4zH9o>S-&TZCz_RY%^C)i=OqJ{v(4h*^30x63>^?kf0$ilDCMj?R%QneWvlA
z_UJte3)S1?`)?%e^>DirFS(U}Prs}5^BAY!j@d0cS{Pn6^Rwh~MmBMJwoKHsBbhlm
z`j~cqs8;BO)v5<Ge{;Tk)p^G9Uv7P|>C200b-K;rSA3sbyzINsuFWZU?mwM^_fD4w
zcP{O`7MNEoO!fM0vVYiCquS-E{DY>oVVfsYBkRO%PSvyd4nE8BFH-If`KwaR3JXS6
z->tXi_l4Q+Pkm6_bi%^U@rh#P93$ULQ{6jOm5$lhy`<)c<E1|yY+YCQ<Z;XV;*_Cz
z-HUFznXWDLuO1q7DC})Z<+JVS4Nr4o-xx>}XD?4`H2+w<I6<ulV9JNQ$_TxX*0XP?
zzISPe=i(Xrhu&zN{{(NleX2T|->}Ofp^8yooHS-*UH)3D5mqg4KA(T(@B!zj#|Lh0
zc{p+J!W~-{>(NbzYIj%-``zjOYD=B<L8fgvzi92vN|=(BWjG<U!LlRjw)A<-ci-PT
z@Ri_?86Bomjy{)HH;r{ZeJ&>KL0<nI!^yV71s+ZTZs!9woV}5|ws?MUr>m!nOOyMV
z+E)|n--MSW>Sh_gEc$GdS<AXBW`^e@49?XWyN|ami)fx1f24SS-JPwu6H6YJzjZ4)
z@Ivcyu0ww4&&pk!Tt134-<lHNroXPg$x*2)>Tj8_r?D(pXL(l2R}10{hwykt|Jqoe
zWvO~8!~zo@lZHiU*miK@@d@9_$GEF$wZx@6PzM_j7bk&}k!Aq1S;#Q!p(l@?)xn0u
z>@09L(hOj>RK7|T$6Q~jgDGNm5x5v?hX0!DXN|tos+}mt5@MDzQcKMMX6v%PnlUvM
zvkdW2Vs1m$ZD`Wq1TNCnjreL7bzvA58S@x>Eqgo7z|+Qv#D^Fp%UQXyNB!!$F&2|G
zlxE6FngO1VN>B8g*y`P(hb@V@4C|F)v%v{m7XN5?JR!XH8zI&sE-I1AMl*nU<@5HY
zWrYu|gjk1|*JIIolEZfsxP0^0v-p0}!7F-r6mcoXdgbIL?<R1`4!)bK%1a+B#M;EA
z3yHdr?cfA1dD>Z<EREZT=wL14A`{5;H3OLQ^IDslmY;4GVw#xaL~$mX0n9tZ;uhU;
z?rVftmzbkP(bk#)%zM@y-VpL_T9qCiNzA2Kuas=@-2^TrEj>QkBhOA2VtwKgDhjpH
z3}F79n%;`Hn%L`MGh(j9l1gkpIDyMQc)RXj3Q399!A8VIE|42(1~6Z^(f06MsjrO?
z3yC>G6k($oz}#f1dVMfsj-DPiCFTmOTY*gmCvbUm!ZO~zw^1m@!-$JwxWZgBfVrzq
zurzo7S85^V6LS$36=9pf30#un6H@N$@pzQSdUNkJpH$E^a{vhsAR9<8&6FWw85dU|
zVFec-M8bnyT#e{zG>Uz`4$*a7+=%E#E^a||3m5kzem}ARg3wGh=4WFIIJz`bfcXX3
z0%%AxC754=Eg)8{B@?n5zD$&0nn@BRN!e|*BT+k&!V%MT%FLg)+4<@)no&p;){sb=
z@w4@Fgfm2zn3@#EU1DlG3WFtRrW46Kkt-KZ?PTrxrp;bXO@Dk&QA{ri=tY4L62)ZT
zfD9Z6AyLd09IyokLP!*|8~g9Z0T2@SA;F+lfe3(BVSz@Y;Aff;nJUG)z;Vzl;x-^y
ztXmnT%CIg-f<?`)l6hB697*Pq$0bi@!@h~e-9(ciOprLT+JL^Xt5smC0@J`+%R#B5
z1R9BB-DHf+5{@=WaClNA`=)jj-i{*SO)zWI%yFbVj)GXnGe~)ciyM%#fs32b_+~T}
z`~`y|%5I_;sd`Z?6oe)PQDxv*NHk3fqS}IEp&&FVNZ@Wf3ko7V-o2nTa^p`F%`CHC
z=K45flFtZl@lu*uVZDNbQms=tC=HEHLrU<4t&l7Zh<N$;4!_19%{@eG#2So_b6!tk
z@XAw|V*G;rLV=bPa~qAkjYdI!WFd3><@=VZ5yg}sn-Vkz;)Vwh{-0&#0k_u-Gm9Ee
z{QcbgJ_0XS>p0%0`m>ghx0RrAB}f6wP=x3{oyZs7H@$D0^W9^s$9*udfq8tI@f3K9
zVIwpX&JQ;N`K6gm6q$+QAagXc0Vy^h4;GgrNjb8I-O@}QlGULJEY8OE*;oO^6%4jJ
zNb-DBi{Uo)b4MVWkr~R&ftvGA726eTe?IfQ*?@LQPUk?cgk^y{<mY$I5e(2ANzh&b
zFHX=={3s)62{fZdk!o(xn~`EOH|R<vspJOTh-8i2pbM~l0am~+1cU7tHlbDx`Sx+n
zTUQdaR4g@z!>WGh(~_8w{Q6JKfOA4(dDp37ne473HBY=HcmBe>+)MYLrX5>Bbm%=?
z+7fFNw~K%OC$f=FWYvkBfn-nz1A(UN!y9SF$;-(fgpX#Ny`23)-)KheCHIG>OEWH>
zE`Fd^G~?>!>d!K4qSr(Ya`STIAa^f!4w~XPMGiHlnQ4yG<Sa8I9V6vz$)bIuRdDT~
znQkQOMlO&#npuTqtFQ~OkY;kREEl^#foNt2cHDvGAThvRlJbG84TN8~-J#>_Q%jT|
zeDq+SWc`7=5v;#6Rh`qIr}TYq-JrJ+7JCRhI4ln02N|*HN=K3DTo&gbMGlw6rASiB
zWpOo<RdZR~kL>%A0+J{gyrIEv_a3YqW0&;3f{^Df8haOw1M)zHNOS}DHt;XC=sibO
z#}$zdx4>b68|0W~t|5nO$PJ!8oN%MTM{xe84pC!DNiIk4J~X4ER2&pag>p~?6~RH#
zR5S<0QE?p9hNw2g!Ci>z;^H(+rD2YJJ*L)kaS^79SbWUQ(%^RHi?t-CrTnFaY)tux
z&PT?8PuZqAE2~ty-&@$zuWcL)oqy(&oZdppTdeAzy>zj@{$bC_52hYmoQKeikHCjR
zodP5*KqFwsg2Br)6c5fd>6imIY518Y%0`uAU655cj->PhH)?Q$YYE}2x4>J(3PB7%
z#t;qz&Fn-GJ5em#T6QDhZZsUWL^D;$qKfUWG}D3{Tac2qmteaR>;i{aFnC>uO|+(8
zF&(yKc*|epvbCo$EOFbKj1SK3)k1O4qj`Jz<QnfJaZ+$}QsS)O=%hrhfP502l(;BZ
zo%AsFkh1CVH1^~mFJmtb@)mnLu%1E2K~nbMhZ~1W;S|zLlyQ_43PUq7;ur@=c}MEe
zSOZf<7@_Z?QHxyIOkPK$t|J#Xl5jsK#UJ=D2NFs%#mKi9`NLz+W~YvyA*;?N7qfIE
zOh;4LN@gNqCKs!bP|d|Vk#HxAp@M`(@BVFqH+BmjJX)O<GVV_u^2N%!RO7xH{XhNO
z*Q}1~m@_<-l%o_$OWBPQ)*@*+7gr)_B^OsA`zquMfeqY%g~yQti_ak887^)>!UisG
zM#5$;?n736$Q_iFX4YV<HQ1ePM|s#P54*GNXb-m9gWVwy{T4oN_`!?J|L+4yj-UVO
tfSji!&rX>%@MZYGp!r~CTXpTC=>~^ADU0DF%z5Tu;PEWjuOry6{{bQb0B`^R

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v7.hpp.5C98C34294C833A3.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v7.hpp.5C98C34294C833A3.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cc94facf536b0622d50f7f28234be291faf64ff4
GIT binary patch
literal 5142
zcmcInd011|63;nA2;mw6guo4(h-?ND_C<&&ix5;6mE8&gF4!VG_WFRTND&ZG5sHGS
zeBizVDoaJFD^f+kt!Q0}3yOk>6b(M#A#!^2+CTW~<@=bM$#2fNGjnEsb66mrGl#{Y
z(cGfNX|c<ameOf7nhN?$OiGAJm7~!@nKW8j`Og9K<GLN#w;h`!S)$stQMR93kEw9t
zJKsz_FK_y)@2!8tfs-DkQzx;O9rqt<PW1{2Pgxq?AaY;l8TX@gzi9KevK8USmuCj;
zRWWR7xmMVs^vQ1Bitz9c?+SY)Bd;xW#dh!PqtZ2g=)Ur$J-G5+n8%3UUbQ75YL@+T
z>QkTVUe7rsZr0V_UwpB-{Yd()CQ0m;eXaG|y@nnqB)xmnZ@nyGRQP^C**i|hEpqhL
zvOgpBv$=1r8{}^_>8cG*Qv7_hP7pLFtz0N-5Y~OHx^yV!>8g2M;!#P@^NoFq+V7kh
zbsW7tELWGQDdo2kW|)cgBr<d@9oHGCO4>ojAE$j)EZ@H23}&>f9of9iCFekC?Le?$
zs<5cB-7i6S=boHN{TZ#EK8F|04{9D>X8&nNyc1u88A|E#tE!1?Kczm`;w>ZMn4@j4
zPN!(i(>&S#=48Cd;M@YOF26g!9iDrtuE^p3m)5Q6FY}8*;g3h;n3wcdT-w`nZvA?V
z#L9+?_a~jd&3PW`wlI5EskXeW!{gE>ox}q@Icw_OuQ(pxcY;xs5_+k2irTQ%g>7Fh
z=5{Q8`5|fL&NQxsxj84Iuguh@(`ClejarTdnL$S#n}tKeC)&xl^3|_(3M0Os5hc>R
zsNIxZcQR~>agFohJ<H<<-w2gEN{TzudING7KmD{Q_rQ|nT5(C$tja&O73Jp#$Hmm0
z4oguPS-+d}Dz6~GaZAs2zLLqw+SLkq57!oG{Kx2AK*fq^%MSvc?)&=0NCmSOM$^l(
z!!DgEaxP+|D4v<;KZ|}Tql#N0zvQFY#!zc{R{3CYGTqlH{;G+M+0FM`!h(2i7X<T!
z!-ANPAqIh2W)&+JogF+H)IeMde1EKDU3YAL$*!&w=IygexxoW3gu_8*eP^$1EJ&@X
z^FCcaQ|WAP(CX$G$)xW7+j_@4Z*g^mH-FjXc(2&gX|QjVbEC?$HFo)hiE`FvX;ul|
zA?tq)-f{F)^tFoe>X^`bs-5fI1zxcot*agVHTR}gnN0q@Zd!%S@^6#mlMij`O5S{M
z@&1B{5Wa9>lIE$^p^I({XH4h&)dYT#<Z@=h`=>rKqpi~AGvDqwoxHf=(5u?|I>`^S
zdyl*^5bz%5e=ce`@N-9NKRi(TV~G04ihy4-pDTR4cp^tF@A8d9evWHg6vIw__$~F&
z*>(Xf@%z}wYDlB0)unvev1HVX4HeaBWESLPL8I{pje+_UeRS;F#Xm*xD%g;TOyW85
z8d8B_zQcVJ%hYSnxKIVbo}{OvRAAWJDN?jQ_Oc%vG7!8Sa<-!ve0^XOPfoAmEq4_~
zb*P0*_&R)jslaf+gW&WBg^%J?AsxX3z;FOKjX$tSu#Nh|oUMsrY)BxJKu(~MRA4yl
zY3okb`MBSdpge*tm=?-Xf#Gjzc5YfbXnj*2vQ=rMv4XL&RL~S5P^QrU38MFHy>U0N
zvP=n1M(f>~?pjiT*T?>xO1C+(&{P#FAlQy)XDSsK&I+V0P0;%Y)S(W7&2`N6r2@m7
zjl6eeWF6+SpdNy&Agc<J;}2}IEtr$-m8<@f5@aD0XQs27RA6{}$U?P7$Gth~a1w$|
zd8XP@f#KSm6)o!CkIh+79l`sc-aa^e{DDpGAE;Yd#@S)71m%#4h$&K&3JgEcw#qm-
z+DvCdB?O0a!i}T?!*6r%D%yqYic^QW2%gQG&6f%cj|@dFKJna}REOFK7U&4{r2@mB
z3wkpSSLKAWAP2$KkX4Q1|N6itw1j2L6VoTtX!KcIwB7OoPRbMHuYmI_n2u|cAUgr4
zQ&xWh*iV2Fg}(;u*RpycWE3KQ$IG7}_dv!TsEeyrA=xpvEoG%1AV^oXs|k*9vc<-o
ztBNhVl?c*Q&(shno**4{9jD<kKuf$_eee>3^f&f5$4x?zne0q1<<)M$*$sHpx#cUq
z-|6H^kOu+hpsX$dRubTVJAfc1V4eglp!5yUcmp)2q9}&O#n2o_p{&{EYSH=ow~$We
zLC-wsg$?Os8T2fJUf2--4w&P;k6QwJih!}k$3N>{s5-Q&Tf-cs$HT$H4fh9*5<n>h
zU~99c_p$eaiuNJvbggs&-qSpK)JX?sv>@;Nef<6KVmdhlrVN3pI7oqSw5QeK)O<iE
zuL9$%z!a~ytRFh&8@P80ogkf^oc(c=5TwXS<c}+wAjK^4G`u%K&SlO0OSAIo$&)8T
z?9k5s&i*27gp-0|{>oE04hO?X+EV54WaRKwV0#tVVNdAjT-+tV1G|g~O^zugGQ6H3
zEp#mOF&UvU@pWB@%LFf-M39fb%tv6+qio}(&xebB2(lN<?4@+(IQ3DR=0R~D497=E
zpraL+!EhV}rbZM^AmT;5ELc)%`Mfd9X2R0I?kX^)oQc<?oG)+=lF9i3QIJf|4aL)O
zI#JFI=gMf=&^#MjQSO#N^Ac!<-NpMurj&zERPdXZHEcai{7*sQdXr~Lc^I!D$aP@a
zI$(>Bg&<pjpcS}KsT>5;27xWkH9-yo?O|X}=@P(}0DY`~r>V;`x*ns2>efQlLK|n|
zyO#Z?WxGCvO;`q)5GFCD^MG>GS<hJ@lR7<-fJ)tb?tDWkbv|65zqELV{(mj(Owxep
zh=F|>nDe^x;oVZE*pVQsfqgZYv&jFc#4EbxFdjGzrgYHAyf?)88|uAcGqDBrrVwrj
z57!Do<^o<W(4`W&8))qYdK6E!0o^t*mBPCLw;S-Vp6@NU;Z5Ne2h^E-b-p%k2ZA(L
zH`m7A5u`v}ppAzEL8fq1bnr>#zjD8`a(Tk&-#T;CVEu)DH@#>7paMKY83Z|>IbWTM
zDwGJ-prYCg{5FFyDvGVZX)Bn6y(Y*TfPaJPYJz+MY@dKx6kZ4|3ZWA&UFP_d#`)da
z@p52Y<G+thSQq#RP>qkxa=e~E`_Zf5WE>2hE{H|OW;$iI1-NVhvnjVO0=J961M4Lb
zYLQipmt$-z0R|FaiOacs7yoH`OS1ihEnk_LU)aX7_n0t{$_}F&<!^Uep;y%ab`8+L
zJ{+t$Wn#Lz!WIF;KsXF+un)&{N3PM&MeatxY{6_7%zpTaB1e$EOkbJvs{xudzyNQ{
z9KW{URGeM@xXqo}`a}WI1KpGL^}0%<5o91eP(}-<hs$X3^mrL<0MG}3jGhJQSx`pb
z4(Zz|z0p};rDx-C0h1s@7$K@uWkoU~Rq-u}AU6QU2B3<)9_uC|7syQg5<ry5>Sq9P
zMpkbF>^7>K2(kmPJ7o24!049U{~2IBqx7aYZo&(mixBZdDd3j^d#d0ofng=E#CnTy
zrlozYclCtl{L843dAVgB1D)C+upgub-Ib18*RQ`CIzM4a-MW(;#Gi8c2Q=Q+7_PC#
z9uQ<65aa=8N-qN@Wx!6B*-WZ|9cDIyJPQP8Df`>PzV4cUh$0lKgR+ATCR~C{Q%lpw
z`M;NaS5rQjcjq6hac`)6!OGISoC&?h?2H<A>~4(95AiOfSHsEp2HRb(KGe7BAOlsg
zgSCU0@+MR#)S9Z5D4i&4Tr2pArt{df8t?oxvmru7({=$$xCg_7#Pbff3-V*^mW{2E
z(R`03*yvqUrd_Du+$NNh0>CK%zEpA>0jH7D-?MxBvT}WG{zkIH8J9D;k_*az`xz58
z8Y*K~c<SAv5Yaa{L1@;S&6OD{k=#fg4sq;qt94Lzh@1huGqOXb1L$_h4wYws`;01k
zNmqX4(`cR_y8GV*{F}18eH-v^%j%B+=MfO%VKR1u;`GYu1<;@X&cr-UkQLCN0?x!&
z5Q$`8v+zO55K7=Vz&i(Is<jzRX$Fosp`Uf#|HJT!*Z#kIgbNIfHy@^t3_V5>Q5sZ!
zn}A=F?09PdoED&m+khau0KZFiyuAjtuVu&E9%!)#I^i@h$Nve#C!MhDyQO0co0t%l
n=rs0)J0`w<H1HQ1O_Tb{`|B$tKPeZxhJ{`lq4T&Jnza7{K65DM

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_tensor_slice_transfer_v7r2.hpp.891A886C76B49727.idx b/.cache/clangd/index/threadwise_tensor_slice_transfer_v7r2.hpp.891A886C76B49727.idx
new file mode 100755
index 0000000000000000000000000000000000000000..20fa3b763d819e1a54e7aa9ef94762674a7b5eff
GIT binary patch
literal 6980
zcmcIod010d)6W^A#&CfU!eSr^7ug}~hzcUOuqh%aB2X09f(sVcS}Rnrbr+R_#ifWv
zEY%jRXi==-0$QLh2)LlqTC~)cYEfAVg6~OSdh^fs`bVEH&%>PiW9HsFGiSDQBZ310
zd`x+~@$-W}T)JT896pc7)2DwcRxVHcU5m$ipW*R7Jn&;==(5I<qWV!~X`AFR<|p?R
z6)fftI{(LN{rGfA@|iJ0#`Db3Y5C=Ot<s9VPT8J@qdKLVTIG(DE_+?hS(|aWx#Q;>
zuTD08erPROvZ&(y=Og=8R6TlgKH<;YTc<m!edSXp^h7F}?iR)ND1^eBNlJbBq!SN<
zztVa-{zm7lv3_kc13LF#UmYUs(tiCw`S||T6#MVL5RlCoBZipr%Gy`lA9SEhG3nst
zjGDg%O@1Ti4K=!ReZS{&`SG}wy#;9wNo8p}4s9-J{piKS_u6YyfxqVc?LQxRs%z~&
zq@%EQwj}rJq8ml)>dmgCh9s?-Ic9<P#?_NO_iELZ7iV0&5EwekYvraNF76x?FgRAP
zp`)!Xrj%(vk$%0?XLUnLRz^y}rd_#W*y#Cs+3h{rT^3n;K1zwYXJ{GvxN9<-Ird_`
zE<1Dcm{n7|4XYM4RrU5XKQhkp`OB)9H(6dId{c1YZBt>=DP8NPtesmT+RJQyIs5i)
z(vkG}(wBl0feyb%eC1XD<&MPtQ$K9{_S-!BhK}<mim&FR8QRX6UiFi^N0#9S=X$Hi
z!LWl9D$Ei*vh$jDKCzrJW9qY6AKMsQ`R0?hq4RzpJ|cW;h3&PH>(MrYhdJsOvtQ<S
z{kZSv@;x^yHiwQ$-tk`A{=HBf5%1pJrj=)R^U2r8I!_qOhMtM#&nXIUobh(a*?^Gd
z&E@+NlHZQF+3s>vzgITpKU!JwUyojDcAGb9#b?UX(eA|-)xGyJKK<%gnVCU>?U6Kh
z8|_0PuTA6l9c~8owHu`~d)TnBDdW!PgjXF84yjSLZ1HuPJuJsv;<4sRn&DKF(m8wT
zc;VWq!w<d}=qb)wWN@9|Y_Rz6&-4!-D_`<7%t?o=tedcKVe8JAX8kXdjbo~3jgI-d
zN`J9yxBTWGw!t@B7hHA}2|aGTYUr791QaJ+e=Geq<XN91soH5AS*QKpxfq8Xo6HMG
z#SU$1T(!eUr+&$gyq`Zly<L#nleF&2n-5cT?|u-oxN!b4({*jyyBa$x6P&kpq$Cu5
z+WUk53vtVhzA?j7Z{)uS>*|}|<aMhv^yODtZ(1H0Rkw}w+3B_YlUa8*7}s3#{%oV$
zvyjU1OD{^THb>tK$@XiVy|BtL=N2(Jnt0A-P`-BasKB!!lTLN$dqkOj{zZ5C++Ah6
zUWtzn&I(aZ&l+!^oj5&k#^Nm{Ci}L_lvOWN@7u?Ia`@%jSF&5Pz3(|b^3M6o_{M*p
zCoW1n`s(|T-!|XvTN^%qm3+=OXTNirWtUMEZ1u1*UHDaDkx%OB3k@f&B9_U1GI%Lk
zt9w66=k$Do(5o5p7IEnN4!0u1wHx)iW4h<Gya;+bdvH?bbKZ)zOXv2`eAhaDb<2_8
z9^KOh@h~3iC-gH=6P|$91o{({tnO$FmtGQsC1tNMUE`o8$QIXB{9SfJ|2;7bq3jZ{
zDgobt1w|y={VF9ocXW-wn2JmrHqBd2kgW*c*k_g=<ERUEl>G*D-@vee1w|a^wk^$B
z(zHz*BvhmwgzaE5u%L+Yq@i=h&21LCU`9o{!K)jB2No1@iu<df<@NNFA`nrLBuSFB
znjqV~$2@;k<Dp7J5Ky*Ksx(s*WKaKLvsLA?*@1>&PuUZt6P?ur*-@VuG(6v+yrvC<
zD7yoM9aP2s1x4bL|8$sn!9vFntf+{y%$ZdaWT&+cv5}4Y{FE4kls#HH+D=W7y>lvC
zk-MvLtq61|y9gvjU_7v($nH9xOI@<vOKs4iBK|^u12sYRckvCc<`q}2)dmB~4iW~*
z)dblkyk2n;6Kt#vLn+%s=%J@3|F9cI$9|H{D;0r0WiOB{uvQafpBuKY=-}1u6(Z21
z>`o%-BxVB(ij>wp*=svt(FY=6s7NhHYQcJ7L6Nd}<wuPVOvmYhl#29%ZZD06{sl$K
z%Uw3CnVNh}1iDmYE0Js^rUMI#Tyz^gX~zDliDDQ`MXrF=6&jQM3yNGSiZ~$naY?u#
z*iw-$kamIPz=9&z%CFtADEasY5olA9W{@;fb^8|-xh}ao#pu5H8$+<6BCax5R!xw7
zXGz-6_nzlXHUv}3o-Cd0tR~2QyuR>ibmBZ357w0JCUaxe1lhmcKCJxM;_IP<!N`Qi
zIt+2JQxo1$Acj01Vae|^q>2}VBolN&M(^(fMIYD?EclFWxr5@??x@w;z^5W%!Z5j-
zAe*;*!NL`H#9R#kp*r{R)eOx;{u|K26y^KvG8?}&^5kHK{e?(=A@;Z#4Evmjo)c>h
ze-4t*!5*W6Ve>$g2i6>J0znhVF(eqa4Fqi<N1qQ_HEK)E>I4NaY=9`h7ITkbLq(ys
z=vJ&ugM{je#3{X#D0UL(ukSsIm0NWNGi(M?WN6}cuxtkpEGi65Q_BwUzzo)L*}QL2
zmTRLP!zvX@D=Z5PJ4Qan5|fnbH6>uWhF()br=wmB8)p-z#Hz=z@iy@qD4j6rM2xyK
zY#EW15ou8P4L!?to1qL_O(fNtxEw~7gD;j@hAoGPa+t~C8W>pvz8LZhTLTd_Fq6Z#
z!R9t7IY&PMn<t<|N9&o?jdQvEXfg2F``~z=6Xdh4;MmFu;tRDHww3s{k|=bqaG=_A
z?!OOC_aPA9h<+fTuKFk@znK5cqDEsS-Qf`T5MQiK3>)qq?u%iIaVKEdXkoM=K8j%z
znMC7~fr7-ad&$(jB$jjGKH|NP1aPtVD{=UhjNtHd;`yBTakvRwn!p<s77jEY)D+`{
z2nSluKaBIQmfnrZEFcWKff#Qf_PD9a`ER^su>KOC&&ExUn}KTc+4$-48t5FDp92eY
zET63e^GdM5ox_Gk9e$TEcZr<Sw}&Wti0dYoQdjuKDqlFz=1?uCu=#;*$JT74TiQr$
z8i^g=$Y(RaCIjp+39%ni!&J=>?{_`>`n#|hdFF!{*3a879LEEO_4oD<#}>n|k-CwV
zXc@ys=|*XwZs2!=6fJNIYTi0{So8?0P~6x!KfH;$V{H6C2-8eUsP4L`48!J-u{p#a
zZD81KFn$|ML>oq0&E?I}o4ty<x{-J`a=WZD)-wY}pdg<nPDnXS;@n*WAvG`w-952*
zapvBPv8QP~mJ+X0GR`>K?d8F7g_{`mD)G8X#sy|RO+B^jkPhZNEtRSvfKxOiC?pQU
z7;6a?>Yo8Pv2HN!20Qf4%$_e?pO&V3yfgoBc;E}wLj(6b519vxt(#V*$be|nkXFmc
z&}ghmb2i;9Nb>$`J&mWs#Og4y#RO*9qr~c{CO%HAj&pd4iCk)3ooGtK+(YbPhVNQ(
z>8RcQgYB{Z+VdZbe33lu_DW;w`Lo39ED7K|UqQSoNC0{sYcH+$s_BSxfuF+9Nh3@Z
z{!Uz&MjJ(2VZ<?Pj8V)#Xj_QZ#y>OOj-+w!Z0Kx;4rW+aLsv8G><sH>=w^nUnqk+8
z*A2roXV{a(=p>P$Bey-9(2%fjIc=H$BhI(KJbUrQc*T=<W+%>jG&}o;J+yasZOdO9
zO4oh&uN{w0qCu*f74i8D9k}?tpo6nmcD%vdd_hYA-Jf!>D+e!3Ayu!js{t=ep#ujU
z9)EmkmDf9C`B2QRqq@P<Eoms`5?L-6Y7CoCWcix7kjM%-d^lxWMvbo5#D8u2;acnK
z2cGA;(U`6$qIzPC^`-FC_qKLPUwTqw&x6}}O=GWt+cnPE@8n%S@h{BHrGgXPCXVNF
zZnE3t@tAX14rs)x#%<2S4~f%5G7bYr^)9pX;KDh!91P0AhQnQ8(*@o*v|<Vl5vnJ2
z?y-Imeo-2U5aAz%dra%KRir(a2vJs18mJDGb>M~#kYS&K@+q};U;r*z_)E<MVb!nN
z3_FcKO;>#?q=aD?@E2&HBz}?xTEkzXfwmI<R-%DB3BOYl7XiNrH25XJFVVzTfPY03
z*8;y*6E_3DnZr04(9}_#Q_w%y3h2N4&nL(z^<nj!H+Sry1||p+#CSWyCJU0q*d-Zu
z2NCQbVvH1q-9wl?#E8Sigelg<O+?g0q#VD62wF7n?<Rt7B1WwkwvRA<#E8S0Aj$+O
zhxdbEKl~$)yG5<(KT4rB=a(5R|6MxPnMURr;&6s|af8t%Vs(kQBE|%#(^ZTxMq|2K
z&MY_P%ua`?=@8o`stX_AZSWJrZUM;_P;hExfkPJ8XH}EFXCC-*{5EiD18)o_;lObY
zwMF+(Gy2bR$iS+iDiYwE3sKy>(|?phK{}9vQ!51G8!p+TxOx?b9QoHp0@F>X+IpiM
z#2j^H=&WJiNIW>)O^mvUIcIYQ*k^zThtGlGIZ$x<Tn=N)A(+FJU|0zXG+9+x$JRhF
zV&TBal(V=S9J;|1hltB>td$2`PYY-@>w+;|FbOrkT6?Ry`q}I9cUENd9o>cg6E40A
z`=CImKH1{G>1Y`37sr*FNXJMQuDB!^B}lm>+C!v!G^f`xVp>KN7=#SlMC47xg7aky
zF>2w`=2pY1C<pV(RNAFe*;MDgTK}citG++Zu$jaullX9tIZJHMk`bD-gKY&FfwKd{
zz9mj?i4TTxy>n~hsmS<Dnj}7YJ~A9F7<RqhdKor&EaS96st)g*@!>(?ad<|#`{^B%
z!K<VV?`-IIUmu7__+sm(cWy@G0mIj=iwXra@_pQWf;rt1Wr^;bUzf|4yJME(fI_!P
z^&NrJ)IZukj_UzYj!_!hlx&nN#RsU)ou<{A+tNp*eMHWMa3;tz!Ggp4!Dv5MogSy0
zi`O^tjG+DuVuReU@iA-|8|H?tLba$rRi86XeCbfO;N0#rw5`x@g2=?L8uTa}D~#2s
zO1UH`mrUb&W*)K6Bc2@2Cqem|`~qTMpvf;JL4}(9W5oWLCcmDT)RW;Dhr)p`DX3Q4
zyp$0!P8YUQ1G9-Fn}l(Zmj;qF2*b#FCHkW^eM^{XqW`~7J00sQQ;d2p>ith4bQ>M{
zRbOsU6NcRlPP<_o)<cH924>g51*?UsdKlDz4dR}<ZF3*bmrkW0Z5|Vc$C?M@Rw8ND
z{4kOY4%raE?Qt$ha>1E<(L)&i5PY%oFzgE${sMflc=YrXlsP}heobBUDM&vBcXSEE
zeg@{B!6?MN@(F!{DIb{qziULlZ!MXlnpajVpSL1Vbs<*~^$*x^%jzFm8%<^pvXCpx
Thx5z_7<3uVU0QIL3hwg1o3QyH

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/threadwise_welford.hpp.CCCD2666E9AE4B4E.idx b/.cache/clangd/index/threadwise_welford.hpp.CCCD2666E9AE4B4E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..9c1f309e8ae278e861c2623a5f70e78b10f954e4
GIT binary patch
literal 3816
zcmb_edr(y86~BAA%CcVB<->iltbnk{F7lL@Adls_5anfATtQwUT7gzr5D}S56M`b5
zl8{6oQpQImfQq0}No--F5<x90HKik>k)TX6RwX2AW52!exRgISrNf=sGyA*0@0|Oc
z^PTUU6Q>RfyJt!e!5h>?IhndNiXaFx{L9znZT^Hs5GnXYQGIJ{-|Ai$-sC4IpX`b{
zS{Ax1>b?Dmwx)j+4&7}z;hxh{w7Xkpc7J5+!CJu}VeE8zY@lvx<N7OKjt=*^v_vj%
zJmCNN)T7bi)CGPkTvaa4_j0qBk2-8)f4!{uFX6AgkRQ(NJLlihccFVK>S<-)k6B-v
z4z#X6{BzET<=vlu^-`DK+O$p77E;j|x30|W&O^`dZ;dP|Rkw9lOSt3e>V3++k<kk>
zgypKnp2u+tKTdlSU2L&CrQ*1HV#Vs?<BEe^-jT){iFadGRb<xd?(mx6@VdP>2KM`f
z1Rru23_egcpD0*s(G}6vasN)8YD>pIw>!DBBSKuILx%;e3r=4CY@;q|O+Z6Uzj51*
ze$};t<adiJvTkz@y%S#&d?_*4qx~Pj^VeTnyIDByEs9Y(uly|itW3_?zUx|J`XHO{
zzM(0VRdsIPFiMW>mv}0dY<XuWy3;CUxrOZ8%%8Z&TqEpy_iT>XHelS@T>fxuQ$x>K
z{%=S2O|H&vUs_ms?Ad#D?1>lg6E)qrj_0%^W~*wKb#bphuFWFyx96nCk_2Ht@?Fd+
ztD>a&B%3A>&zL7-1Ytsv^9X{4SgP|*-UaqH!+cV*kU#>7K*<P#jgJ{W(YJ1rH{5%>
z`J}*<KpKfg&Iq8_NQnVwjsGB@6k<9-k|1XU(4~7{DC*a~SZ_)$!p0-nk%$pMA3YHC
z;pkD52~(1bX*FAo7y)$i=&?Jgg;gp($-~AIC5duI0DVTWkXub{I?W>&U~QR&%$gBz
z>At^b=W|c!5IG-fCsWBLi~zbz<GJ6rcUUr?<lw->8pm2Pf-o_{d94pj&yk?rvM++2
z5NV9{b1ZUXi~#*Bn>Igc%D8IEC&l>dLW$6g5kOy!E${TH^=jdhBCH)RiC@MDpl`2>
ziyjVpkj*E}Fug{yM#%`EhhlybcR4P4$tU^PxVgmKg%NM*Z#q1^#J^MXNhzk01SuH-
z^hDXgn&d#4JC8KRv>WAS%m|>L<UPAm{;#$?0cnD1saUFD1kjTnCojJ~l@i7$EilcO
z@Rf`Jn#jxCn16&r5Jq`Be0_Sy0!SJqFHV-9c@u7oLXx8ZEp(|k^y?2@hiMw|WBD4e
zO(Q|9Km(~qBVmFt2?9!^AbF4@_&_6-ol0py!FIvZP(paZDo8;ZX~H$DAYW;e7@8Ob
z8APKUOLjQGiPEUVrevuB?X=yg0K^sDQn}o`yM;xe3}J={{1E-^H2c-@b8hn}<S2HO
zK<ZH_U6d|{MObBVhSfA=pRsfra?Nl9WE0TQ_0;<YXO7IlB_}z|<v#VP@?z$n4xFW#
za|Ke_T>5DD=LvgRG_n#|S%M48HoWq+BzN|Z6cXEt<*=GU+QnLHxFoJ)pSD?~e)$_C
z3MnKC3(%!dsy1~UI7Oi{mRlLi9YX1CmrkT{Qwp$>gUCT_P*RGO(;}EAPqPOT-l~sO
zXI@xyXZ0OOBkaAte1Qiv@)UU5L*|6vtoN<chUOd5Xl1}kwR2x-Wyw&v7mWf}1cn$;
zVN9V`clh2#qfT)sq!PBPuX_*+8Wl(i<lrP^pAm;Ce1=6@q?H+vPN1^^mBg1HyN+G>
z{qh+pmgHAw-ZXDKcUHb<*QDErj7Dz5moU{a)f?i7QBdkCK-oRV9qJk;*UjP}3O=xG
z_f$^(VMek*J;za$`b#6AlF-OY>LrIzbC?<r$50t*^zL`nuf4G)KbfB$gt}_HeQSKs
zMD>5&_<_OkC%Z1}Xq&~xYgeEBKs|K3V8+Jl8<h7;dJ4Q|#Js7rvbx{YD|8kgfSxrE
zhrn6tV+esls*ppD<A89z#1Ie~`5+%DAP&>xAUsf^aFXd8oFQYhHd-e@U=}aZw>i+o
zm2y##1nk0T<3;*r2f8#W;uY~B_c|wvJB53mHqTMvIlu8<e+cKRQGGBdDIOGmd1xe~
zWYdr<<vI-|QAyKK4wds3!Jb(hrWZpUyWnSS_D#hPRTRF`Y)=FZ)5w?S`__FL>DW31
ze&S0?7YG1#U%9=u)%V%MIr92q-h=OxUKPhsxEh)%%;9u!Yv4H|14`$no57*r)?k{!
z1Jq+*>gv@Id2|+O2d?1I=tmKV)aAjyUcLJG`JZRxHS|)P3;M~#kT7C<aR49~Q78^G
znt(|fB{(GbL0!XL)j{J2rI1F2j)ehGAg}w$OX8xXwR03!n2GSF^#hS%KZy>BE8w2M
zWQ6Ojek=lC@Ce}{i~_fDjTfzSf>JZ|W?I62>PIDm3g3etHf<vD5QIT1!EHel=l}-p
zJg<-2^zokeX9RiF`+D$Hi-<Kx?n7GR6V{JxuuUj1Eb>3P`P2HYi_WL4ZRhCY&tj2(
z+w(_NGs@<qJ6|K<U22&5`pt-l4?#a4!d7V{M&cy~<SX-a2Csf7SD&~2s(lit{Qo{d
u{OG>5H=^xQ;QxKm@-nm2^d%`jFC$;5|AQzSxBP%_=|Gk_VXlAi%kytoq6QrR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/transform_contraction_to_gemm.hpp.51AB58D6126FEA84.idx b/.cache/clangd/index/transform_contraction_to_gemm.hpp.51AB58D6126FEA84.idx
new file mode 100755
index 0000000000000000000000000000000000000000..08c0cb41becfdca291b34c20d3c9034bfb363813
GIT binary patch
literal 6932
zcmb7Id03NI*3Y>HTk&cj7+^36MgzgusE`mKVc+V4pw_BLRn%g^O50CrUBJpx+10{O
z5LDVaQ>{)##X7~}T8j=;ow_Sd{ao6LR+(`rj+WNBA)MU%^gn$b-;*Eb-20w;?pfa{
z<KpA#V3{oDwQ<Wb)83gYm&s(#?0?ofi!+jmOm^Z$nQYm{15;0z-WW6J<km+~)ycYJ
zcTS8P<a{nqbH=>upnk$Xwk$PXGHFXXn*yF4%*gzC=FBD4-!K1e_^plq+K`<z#due(
zuJIkZUzzD}<-aAzuWra(GVnK#JGVMMF6<k!;@^e4Zqiqm_A607c*C*$Z1J<rb!Fd}
z^WOSP^Qr4c(lnFXr<mXQ-gAuphh4?y#_%13N>t+pm3SxL-m8hP@Lj$?bHka+qPPFm
z^`{?yyuH$~-<nT?_h^RP9~b%4Ym=wW>{lO^s5xSKz5C<srwW1_&QFW}^3mK^`+1G(
z?|WeUlVx$Yr?!o*`S9%gXQy{9_<QF0oQ6JmncsJGc61&%)_x&7EU^0TS29v|B-WQU
zrQY~({PCdlwHIc&gaj2YJ{q5XA;$2+$*1Y2$UWJ{(fwn7&xj8i)0I$GAL-xv^4XTx
z+tOSTk3Fo<XkFH6Y)tmOwI-n?X@$Pzj=Xk#Lf(_5bs3`{ep<Wz#jX09tdQxOx@N{q
z`{oPn`7b_+i5m9ujRAT2ZQF00R^8n1{@XlS`hR(@=huGv*OC>^0ritkUT{i`+pJBy
zta>v4t3eK%X8m}#=(8y)4+n1f|NZru%hzTuX`XfE!O5?K`X8Dj%LCF!zH(3Z+Ky@c
zzcDrZ?V&U-)-mAF+J-CS*^f^W_PY(qEAyXS=Ki0Vr#~2R@aVSx_N)Bv{rKz_)dEf8
z>dg@v@4+ARi(T(Kvb`|(a^H(d1=DZ4dU>g4*A9=}6t=C?Jn>o9h4pnw=cREO?)lxh
zIS!BgmkfEm@0m|;HPt!48?Yfgs<G?YU$Pbju4}sb`sPdWIeqlm*%vY&4W3kEaFk^&
z&zu*_+>IP{E8o)kL$ZqW^OZ>}JyxoF9x~>5)_(&E*A&eNX}O~!dd4mVk5cgMB`D(b
z>!0^_PL0b^5)~794D!dI=_M#Kyy@x^v$5DgMP6hgIUYIco(Hn^N!QmU8)p5kB3_K0
z3!b?!yqBPeW!S@#4Kv=Hry?3AQVgEO;L}U&BKKWqC!W7ot0H}w$O?}Y>YfL(Uka(5
z^q^C=t}hwpBa@cCur$2qA?qh+aV7%^CYHUD&^>KJoQe!#>+``gpJ`zuxJQzHN6FN^
zt+SOx!9?o8sh(NGMo?t(o>e|>5L2xp{h7#W@LbIlwGk9ai9b-4^it#zC2?gU9pKaf
zBYFvnyebdA9=>iynv%FNk$d2D4|Kf*MN(U%%_9cCp;M9pOk^v_x5Cg~f+A^EFJDQD
z$jerdp-iLzJPX)~Yy?H#N{XJdv~A=jCHWH*sUS`j#IKj2$lICkK0eft^^=k)naE*q
zI?U9y5fsVI3EE%2^P-=UcrcM_;#5t1dx>3ScgN9@kGh+bWFQl%1*cl@>m_!P=HGM<
zuHNq{$si_j4V<pAL)!?76c~qU=Kp)hn=0bTL{@pO8s76j_WP~5-E(}?4ycGXW9NZq
z9y_p&ph)G@gUvI(to2e6H4`ZW&q8)THi9A_CZ!FVu;#1VO5)B$N{CYlJB^K?NX@{n
zPJVaUuvST&naDpu{!ixfHi9BwzB3`sZQkBiC2?XRP2kkTPGch|a^!<m(;)qak5$Bx
zi4=in5!1y+P^5K1M1IJfIU`lXfr%{lSg!7QAp2Z<$d;#tJ7bl^g|S<}sfE47Mo{E@
zaG=u>UCl@;8_fJ(NtK$Ohpb;87_PPY{Z%^k*_By+yH(@`wnw%{wz}tm>mMm*Cp0fE
zFH{oB*cISZ!F=9EP-Jo1f~==5JeLs1o*aS^N14xs6+i^KW8daL@6fo00}fQ0MyL7O
z6_ghEEr{i)0Tc}|0^OWS7eH|VMxf=X^gAeihY^^428=1LTlk^>6?ZDB<JAd#8*g84
zy?~6~W{xzz8iA)VY6PA+(j3nvEs+)hg+_)7C_)!8ic3c7A_WwsixN<rE>1x4x_AN2
z*38!NU1kT)ws4fHNfmghfvEy7V`zq8!i?|?!Gz1bmTS?2sI<ash1SxWB&n3`m953A
zr&5krjux{tl~#JK)M8=43gF<_oA>#(egcY5K#fU}O0A%11vUC4mF|P$KBzI5QK^<F
zYKa<?#elt^9nv|xeV`AOwt##K(4K<B9#lF7@>3%20(lq53ior-W6oZE1LTrAOdXD9
zaw?qv*`cTHL)d;=A8i=^nM!`neuDM%&U!mS=K;swf`w-r9Mo~)i?YHDwwKx2Je2Qc
zakdC3)H&3Syawill)QN2I#9{o#oZg14E<_x`rX{nLZ+<yKzC>SGupb3V^1Z-_fl4H
z!tx;kYY}BNF1#2bvHW4BWXo)2k}s@+(N$o=Xr)p)8C_0Hn9KZss&5#a+VWN(x%4><
z{v6a8S90kb3_b^Hj4Qd+4eD<2MAHOo=gH<O<}6_g)sxjzP=C3U;gw-OBtxMYc8do`
z*F^7mI#lu8?+P&#PPuyYZ@)0B?*h{<n2btN=@OVO!DLjDN_oUBj|8BSXdae$Y^?z|
zOF8Jv!GunX?%&t3r#IkYxe%5M5olh@DnfW6L}Ia^(suCQ4hF8z8Sp;?2CmOj2zUxc
z)Cc!ver0VmVmnvv|K-}^2QQu7XxT6`F3Xil;WXS8ZAqmF8X=%q8Y`eU8YiIXy6FLY
z-Avs~0j20t1e8wGg>@M;LqH4ZLILdrX(t5ov+ab5J7GG<yFuD5@*6>F6!ArnE<zx;
z#YLES5vFr|8Klc1za6A@5$6(`OI*=*RLUiza!D}9`Gn?+`~pG?IL1(6Az^FQ(D+oU
zg5WBM;33fr!OakXg&Fe}yCho+hYPWVW8P$I+>pV73TQc~(3_?vDt$XDLLEG)q#UI*
z@_TU^<szU8FjPPYSECXPl_IW%z*?~6_BKdVs)N8f5!ZvE9zsxEDt!cok3`%A#wIZ1
zv8c2UjQd1<7)*yD6x%o|9Rbr35x0P`1<X8c9|Pkt5nlu2H8AJ)c4t(&4#w*u?f_E<
zgko1qrQ2Y-E#iA%ya#4p20FpmDdG}hC?O%3si;&+45cEjAb}Ocf}JXrDoJ3ah^vXA
znuK6kp;8So)Nnj?z4k$jpYB_0Di1J+@=LM=SOgRr5Nby=T#oyvHn?l$R0@;AUgAq)
z{9@ubN~DSQN<yXAw68@WZ^p0d{eNmMF#e%Jp^2Es*}8soEVK<{*m`|CBXonyW`~Y9
zpF5Re^s*goqXF)SO&ALo+a!PtD&2(9H^Id3xts)-lL)+*MPG0A+py`El;<b}cx6lp
z+hl?7LSJMI+aQ6AtnFv^R~DzPBveujQEAYgRB|2SYDaj}Y_)CFz||N+jA0uv?2KOD
zWrw!S6Ud;_&!G4jeDP+e^Z@7s7>*Z$^9E~vtiuLhRs}v)pyx5&3_i`E=P`H_d~SkX
zw53lu(W5PKPGS1krW3w#GmLHqlekgvO^Cpa%EtG3|K7T1!M-fx1S;(gOC^^;7XhtO
ztx}`vET4iZ7u3AA<$)>>)To2C<xu5=8uO^NTUQl;8XbkQB&b>qYRrn(+Ndf7HLs0D
zpeh124l~x4MO6%HtjSa=1HUpDgGtkh1Ii#o#FjFM7IADDOyIZz%oPyDZCwfGN)gvW
zXf4Fx<y-S-Xq|}b!CVhfV*WILB;qEpG(j{s`984h6Y*gPI}EYtv)23>c0|N2U}=G9
zF@IW)iTD~=u0b?jp*4S6u8X(>!a5+9yWnjIyDj2-V7Uj;=#o_G1WTugONhCIM4@k4
z^QXB~#1$m8g2Z4jTJvXUrHHGExtc^_R9W+<xrXDl2fr!JOn7$}3(2vA#(MIQTr_Br
z9aWs^_)lJQcEfWd2lO9SWo_d(kAZS<D+h0M06Z8A7TXZc1*?dniumx8)e%}phT%!D
z17YdFHk#wYjhmz!IbSac`a^o&{9f6*69a?dp5qYPrk@FS{PGPwD=s_1v=b6A?^4zh
znJz*C>R|0^+;WK?yPD6JH9Jho+#K&rr4&PoFwUh!r;O(|PMMsN%+W%(g<b;kStyK;
zyTP;@!ujiWL)>nd$Z;c>8by91jBSKToPQZimm!?1dl}*`!$gkT!PGAD+hJ@wOyc}}
z;+9Xm@OUh#5TAT9isJ&}Rv_{Vh+hHmNB$0nu@~=;%?M$`Q?PHa-5eN`?CNb_PVfe)
zbQBav!IuYR8_+fwjzQS8;M$2;$F}S2Ubeu1E#M~3mIF?SxC;h!alG$mefC6kL=9W0
zP1M@Q9X3Rc^P9+Rub-u#iwF}Rb7b4M8?IzIXv#rCNB-LPJ^!FD#((-8{&oD!>2JTi
z^XTv&JQZhpHpW;#3Go-M3R@My6aQ)mUk#Bwac=|vZQ>Af8vIX-L(CHhc*2L6Hjm@F
zCo5kGV5S)3F-CATgNH#tCJ&Pxow&8*$lX(C6Rg)5794{X!dqd1X!|yW3ZEZ+_sX7E
zQvb~4)e5yeo8UTDb*<m4aD1{h^olCb@=DMQie}KFmsmR(#ZA!iOj}MA<wT1K<kF*!
zQ_|O$7X87P-&f6VFY$PF5mTZJoXTJb-WPryVcuu`Duzn4;S<90Ac7arA_y;nNc4ld
z<5wrsl!)5rup6F;y~*}vjlXLnC^mu*FK(NGZiZo4%(@E4H4SfT{DryH^Nsb7f8F`H
z+nvlc>>J#4!*rp=one?Epkza`fYRO4CA=Fdr3a-?K(sh*(OheYWi6hcWnq*^bGFAQ
dKbv}Uw|u0ZZiLJa<4-2j@t-63&yoD+{{TPoDHs3%

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/transform_conv_bwd_data_to_gemm_v1.hpp.341C45E4BB62B32F.idx b/.cache/clangd/index/transform_conv_bwd_data_to_gemm_v1.hpp.341C45E4BB62B32F.idx
new file mode 100755
index 0000000000000000000000000000000000000000..d41b4c9578ccadc88eb8f77de8edd4c91d75fc12
GIT binary patch
literal 22458
zcmbu{d0bUx{s(ZM^_-mBP2<YBfO*toU6}+0T#zs~s4O#68_k)qZJL>ymig68Gi@^@
zn+leuf}ntcnp<WFE|@DS;FhQaxEmOV;4UJX{4TFNJRj!2d5y1EUEhA6=ecKj&T_xc
zbI!O?Pdt%3u|tQEe;D=8w_pF;E4)L84xPRKz5loO-q=ezbm%WTbol4OEngmgv#MXG
zgr9qFt$lscfRCp<-|_$M-`r!DYjQ>5%(1KA2<r23+>(Xsub+Iha?M{7>!;*z@>_B#
zwSU;QM+(9}o<HuzUsZ${e_FBno&Nv$a#hi&PZ}?0{`Bxt=bK3hi)Gcrs{^~NojrK$
zNAWjgS;Y3ZnsX@uv)?Ovw}18B`)+L*+%0i@N#nG`)lG-~lohn>{#zAumRMOIU43Zf
zn;+i)$n*14XEuJg<aMiI-&4O$&im$X@yma7eY5G$Q{JxKz|x0IdBS+IK5f^#b^CjD
zn}0QTRKkMUU0z%GP*IZc$)Vdlii30h7W_f!7oC2pDgDGh=)?Qgx!!tb$p>4{SN^y#
z=+n36{b5nBDJ7r0HE-2txqo<Z?kguhtUZ}mSuo?B`@d;;DE-Zsi-HnDEB@&Jv*Y1A
z9nbzl|M!USi$4#&{O_!r8!j%PPwRne>9|S38|MXN|KG0?uB5NpGB2fi(i?f7ERC7>
z@|$C$e%LhmSKU7xRj@GoFGv40B<jznx=haM-R03?w;Ep>^xC_<H<kb8@aLPhRXjKN
zw|{@_VBF2r%F^(xMRR97X8mLI`#r`S{^+Z~tws4Cull56uXQuBaLG&GzWn%a-+rv{
zq1@XaEO}{#yJgkW1)qjbxOFKgW@x}|e(~dHzk2_}w_mOI?$Hj9ezoS<MCV>ff7Pjj
zv4UtTNUwVz9qyy81&BfDx|)tdFTA?63l*I^7^{f?D$@1dM+e<o#QTS3%O_qbh=_cU
z>fR#fS*P<X@ZN`AL|!cD{My$qo$o>&-XiOW|2op`-iKYpI&&ko*RnB}y3qT*MK%%t
zP2M8EeAq>T_tf28u>Hfz2kCv@BISaY3*+90UF7kT>szjR-jDA>AMh4gPyE+=TlwX~
zE;6pmoPLiE>wUHh<=!IS68~?#MSl6Pi@d&Z?X2H_@RZescJvlmL;Tl}NAG>uMdCbT
zLbkko^QQ-?<}Ff4Iu(+@dmna@8Dsw5FYOOipZ&793w6FwH;i9C>>`Wj-<{C$cXJPS
zq0HOLTH?RfyDYza*hTiA-Z?R%=H`qp)afm<f%tFm7Ww7FE>b;E=Cy2&c66cATVyry
zU+vvTzkJw5-h2Iz@4q|VyR+`?bJFqNo{qZs>328Iymyt(rXj1(9vk6u7)8`wMElq)
z=rBsCy97O(S+m(76`#wjx#(18rJ@VDyO8&>dx^u?&E31vF`{dX=%?aSMc1k5IMFo@
zog&>SvX9*v97d{ir=s(vJ0HDXy4RyONcRTxmzwpZHb}KMSF`4#OEs$$eMGa4pvyF?
zOxZqQ0z2A=$-PbPFcwq)#k8y1a5>aJ2fdW~FGX*m1Gms{RelSN*g_+fx6*-IF}{^X
zY{mEy`uGt#OqDOAkC&mZ(&w(y=T$t84UA*(eQ_)zj=}fEvw`s#k7p6_7*AjW6EL2@
zA`&pZij7#sqSW_&%|?8U{)UbC23^XYD`l{JS^R}8K2dd)gFNCOk5vAiM|_WN;LkPi
z=T$sKJf0$kDW{6ZQ_=Iq<MYv*#dDhl)X$UyG9`4BEIA+xT_^_>qIbywyU@GkfZgaT
z(*KI=syf~^>3<D<L;Bx9@6;#k)PG~2Rvbo|?q8;NRW8^4%h6T(s4D##J8rMH_t5|4
z9BrQqSWdQa?Zd_qzj7GKG&GqGP<v=54V{VpoQ8gmUd2OK@d0+r4r4VBU5);lhklJd
zEkjSs0cwrU$j~$BvoiE7x>ye_)(6;Y;5`-Tp(V<4^MLM=(QCGJcNn+Hu-oLfYQYQW
zpaMEhxsX0tNMBVxL|unyU*&T&;v9WJ`8pkbo&HYwGdAcm2JL;$T%R+jznr<3voKY@
zkVO<SXzvIceuP1L>AX`qe^hO_ZT!h?{8i<BeB?epMfnVmIKy91zQl)L;=fbA!(DfH
zU*$Q%HAjHgh=?@;>TeOlwus-V@?XmEFXec(V;0KDg>s_uw{pn0a+2~68My=F@!G(6
z?RgddLUVng^;KS=J+VN0MR~0jxmJVrcW4niv=>zTpyoP=^{X{kHM&KMY*F>?+lvP}
z+BX$#=iExtuafjvJLV(_IZ2?0pCe)C2wcq0lkoEd&Up<avVp)kG=_)7V2_^0!>8dn
zCz(ekb2!J;i+=T@zgn@&BIL4wBcMryH3>MU+z|urh{39Sj0}&Fk;-v0GENS&eSg1c
z>2J;pIqq~A0e%7IA5!-}_d5G~wZnVqF;|g??Q4<uz$E5s^00lQ;PBqe1g|0e?fV>u
z_nxE2TH>*9M;t~j3CJZq>|3Q)56&e+mGekI9_e9kcZZQjhUAfv%Iisw^~7V}jX8{s
zBw!=yVc*NNdhkXvRJniz6p$YFJ)pxVAVUhsNZSv%kB;bf<kVjnH(ne1+Aw=Hx$(i!
z4~G3Q?duQU{%+k-ZlsZ+X=K=##IKT`ZHd^>`Gv!O{_7IQTH-LqM~{ETu4Fnt8$4a#
ze^zrCulc;z-Tr5X@qy0=-R-s<-nM+wNO!v}hjE$(oMydMKPnLcC8D=FfR4$4W3sos
zUmV6PJz$pJ=I#k2NA>*r+&kW7`J6<5PM)zF?osgGpX*~H6FB#NgYRkPXMY1XN`!BT
z@Uy>x8^@&YG3jT21NVM|?<~E|)|v4c8=hU#-FyFbgjz=^bm^niI*Kl%RvG#SYW;v-
z#XPInWA=i0`wsK0Mt{vbU!&JB&l>bM%<~O;E%U5JALQmi9%iqY!}y+?-=kB6l_CbI
zc&e~c(es5hADyL{Sz4IAVZ66lnz<NVt68<`{>)+2X;vM&UbE_z?LMb<w10iHol6dr
zo`=aJYQvl*ea{j&={K>SP3#f1;BmZf9In_o(es>mL@nrb(f7J&TkuCOOzAZ4$zlKU
z_L1o5=%?+DX&3iiwtxA>V_&vIe#%@E@bd0=y<O)j9e<ThuosLQrEGjD12^#T4Sa&V
zVBFX&#%~sF4j+*^VqJ5O2ma@V28Sh#J5{*qMn8w~GYR;a478WXVSGo;?`VYbHfnA|
z@1gEJ6gK5P>fVPwOs&K8aaF&ZTIJ}Q)Vzu9#WFM2d-r<pN1DS(WM(3I7PDqCXfK^v
z>FAZrvyzQa^}l7FZ_z(6^9O8iAs@7mkF&S1!`R8)J8{<+ad#2=05=ct2vz<tHxHw$
zxLJk1%FV0j7T&dm!!eQ|x+dT;F;7_Y#N(>{g~D2h{#KaZ3i$r*!rYFYC*AX8Ti0|L
z^QC(}da*PY%Luz~de8sTorm#_(!CK~B+Vi$pQxFM8q}YmnKRI{HEXu^xN1K|vr^C*
znwf#+vo$jty-qXNp$j#$5M891Md(ACbqL!p*Q|2%Pnz`;x<RuV(6PEZR&U#0k<KBz
zUV5U}hu&+<XrIvo?L*mNjP)4{n(Q+fG{t8M=$!}N2~}64_a1l;l+OF4^XKfHYhPr&
zC#-vkj<yrlA=2{@p0G}n5vR!$s#iC%o{f0In#xB^<xkiRJB%vPvkFgGSH*~{;t6}N
zM9sNg`}})8r@SYunIw9qx~N6F8aglbVIO$|v1C{*iMFqHPaawL*n-K??|<iwc=~xF
z)Q0KZuQ#YqzdoR#enFte`aK4U>=y|d-f6hMYI8)V5unk7Mh{o9v4h5fCJ&knnlfk#
z=$){4hQN*pdk>UI`XrKp_S!j&X{66I^eob67J4@6GaLOg>GU)4SNr2G3B8LqpDm<Q
z3&vw;rx*(5<Ebm2c2o7gqpt7J+o@|idIxpwK<}om-RL43RD^c|#WbiGT|$FO(0ggn
zUi5zI+K=rWpsoYx!_<A4_EgJTLERPTQ`CJ5T|-?pSigq$ufel^E$v^6K2Kfev3wnM
z)uAs@_a$t<k-8c&-be>FVuIurb=|^vGj%niW0@<K!S<iVT+`4qm}>@l7IV*Hu>M~%
z&zCGjZI3z3GY377dFG*0m?s52pLyn^)0ig>y@0tF;P)+J?nUU8tm{e!H#E7dYc9Hg
zbuB>eX0F}1K6_Zu9?bL=v7jP!F>@7T`4Z+TK_6u9gSb3p%vFZ*GB&V`!H=&B=BmJW
zC397x)A`tRK3Sc=7Vtg`_(1fS1$>fn2Je%B@l4()6P?8av+$~&#h=bn_fZaG33n~w
z-Bf!!xN8TxkOvjwReBc>+J)ZDgLb3$@Sr{DV(u!&_DZ;`1bu+J4xo>4*Aa9jcU7W)
z<gOpvyZ?_I?h>lGs~Y3yxa%Ccj=Sr4Pqn^Rxa$hWuke9a@W=0U?z)cg8{BmReV4oM
zVtX+nBt|@@+W$<1e1@JXLZ+f)MMx|<PK3mv<3&h3I#IY21uTEEa3`bZ3HLm7hHz(~
zmkRe%^jhIsD`5Yx75&!=*q*t<m5cFp!nF>)QMfl^d)tI-8^*VZf!i=4xl_1yV!Tke
z3ej_=Yp(2OU!NUDiVRA@gky>vo+6)7@l+X<it$uAJQd?<(v^norAt>jI#aqb(TimN
zMH24XvZX5<<2llmgI+G(%VkfsJnN)u9mexzP(ChSz8s#9%eP(zt;hI!Ieb0FH^`t3
z7~dd=Z@~B_>Dq+fw^_P2qYI_05WPqC--GK{EM3JIKOkKP&_|^E2(DkE=1SCH|0HRy
zBy_UoN=DDt+_N><K3{6#Uuq-N_L`%G&q2@C!snvrY2owGshT?#+ncYs=c6+;R|bB6
zj^@fiuh3j8(7Bo`x4rx4YOw$FG*=$R^EFpKx<GRm;P)45u0o6#Yp!DSUd^=^eMoa3
z!uq9Jcqy*$VJ-YH`iK^O1btKsKZ-8X!pqP<XyHGgD>Qcne&0#WeG>hX=KcwNQFC8J
zH)`%i^k=&3GadTFR6S^_4#~hc-4%!Nc-<9`o}s&E=+HlQ>Onj8K`LFoQy;ZcAFo`c
zcdF9;(T`W@Pbt^vooZCv&h~k;s`rLR+kRM7kbV`IRXssMPGDBGnuJwjR<)Ld*W!(B
zJ&CNxtZEAjX<;yqkjTRmF{_%yBa`sLazXUFfLYZ`BIFWgRd0&0o0wI-Ee71itZIu0
zZxN7Hjg^tHDyurlrzqs*HSUM+^ZrWl@$^-{8v6M30X^pP7-*Q!Fi;|SAdx((ei_Un
z56nVmaYq)1^x;mu(@wpQ+7mzO;XmpVY)`s9Kjf#)vHkz6{O9i?7x-k0+IRlvw9fC-
z>l>czbR)Z)!)PLpG?BsThIKCOH5Y#X%%i>Lp;KtD6m%L5NTb23{sJ1X0G&YtGSJIu
zx8)RWkXO=fE77ZIx7Fx%G;ke`facM_Jaj${%txp2M^gA;b;Gfa_gaTHV0pY(9y*`*
z%13YHfg7=(ZsLKP&;>lO09`8sYGtrGf!E1^I&{4Zs7E)+ZcQ>om2Z~an$dS;w>#+X
z^?>j7V7uEmj8Z+I6n#wZc1#aZ@e_Kt6Uwh79$NO!W0ybG9mZF*?^iTb?Wh&B?+WxP
z+IJQDrtEuDhN=a<CHvk&-<ExEqbv2km3pXJkmGva<I489MBk3~tc$iY&sx%ZEg6gh
z6+LrFn0=~h9jF+VOP*KGBR%s-m>QVKBT;$e1?Bam_j)qe&Y`ppO!VAH!qmXTMlx(8
zd0x4I^eiA@c9OVtpdzY(yr6uD1Rf%gcq}D>rRc*X@G!cHL{yPy)P}Am5!L8(B;p*p
zh77GCV^#TDGPD+bo(w&YzCcD^AZ@8&hfz;P)}tH9$Od#H3v6VN7`({>Z=#!6U=uol
z4^7}>)e)4)hbE$@@uAbuGx*3E_!BCLk4!=*^O4EuDiK&EAW>K?0;|#IMBq7eqljn}
zaEp6WMBGF-iHIh2vxsaK&#I1lTSVSQ-w}~_(09eqyJD;=-y(*#pkw6F7zy#Ia^zGw
zL6wh{BV*BVa%7zHKMT)Rf5b}<1Un4BAiq#`<%uDEVn~R(p6nt%yNFwLgWbeuH+m28
z*@G@3K1Jwa;!}(+AwDJOy~JlPdLQxGhu%+o_M;CFp9AQF#OENoj``Fvx9TPrn9l`t
zJ@ctYH!z<D^hM@#5q*jITtZ)FK9|u~n9mjTRpxUQeU15CLtke;*U{y|Di>|TcMhXM
zSQY3>VO64!3F{d8xUi0+PYCM-`lPT<qE89y6#BHVPNUBV>kRs=u+E|v>efQNw>qC@
z=vD?gQ@1kFS-O>lUZh)#(2I3zF*;kfve7xZm4jZQTT9SOb!#blnQkppw)?GjOzW4K
zbM}~4T+Y8gdUpBlBJWg`hdMt5+1q};{a|!E+BX_B-gi9ct<G=xt9<QWegA4I#gfjk
z#9tjnaint`I-YcnN6#d_Gl{9ne@=WqSGKpdcP;MC>brN9N_v{bHGI($zQCL2cJ+3_
zZDE(*T|mCQeL+TV1LW@Q2Kj~gL7LlinqAbXp}Vs?sE4x$DAXAW3Uh{m`aAoB1~><R
z-st#7XSLbh2za9>sLvZ=pdoKWfyTV?g3`yHkAv05`^57J=u^+9pnrM(1^Tz=-=G*0
z8bb!Fc6JfJU8K9+g}l=ViQjJY9^$tLT}1qf(8a{B7+pgAO3-_W-(K`S;<pdIpZM)Z
zA0U1Q&<Ba%LG&S_A0iK^<t-(8Df%$c52KF~=TXu{l`kXCGITj{mZK|(vjSa7oR#R~
z#CaTjf;dm0PZH-z^l9QejXp!1XV7Pf^DMfGIIGar#956#N1W%-HN;tit|iV|^m*bu
zkG?>h7tr;@S&wcY&IWW7d9aDJ&7*L52giM{(udUcEM>l>3|z<j>R5NYH2GaX*E7F*
zbOZBiKwo5j7txoP-zD^A=64x=h521UUuAw*(bt&YHS~4ncOBix^hWl8THc#Xzlm;Q
zdK0>tIh$D*RsJ?}-bUYH&O7M4%y}2x!kjJW819VW(B4$;oQjU+&RBFDcgCR;xHAEr
z$eoGkY1}ytJ%c-Epp&>W37yQH$><dRU<z-WH{&oGxNif0NUi^7;k#LY%SB+h=%?<`
zDnwufx>5vIqK}EdW9Z`|@HqN}2t0v4DFRQTPl>=&=+h$bH2RDPJcB+f0?(qWgkB{c
zP|I5_^lJ1up`Sz72xpDxqRQ6_XDzx;IP1_Cg!2NrUO4N~7lrd8`jT*7LSGin%jm1Z
zc@=$4IIp3v3+Huoqi{B&Zwlv4bdzv4p__%X8GTzgZ=>%B=N<H2;k=7(5zZELjC96G
z*j`hmb1FJkI%Cms(iw+ds0S|8``I@K4kJSk%s^-AftlzmJunNsNDo|uUaSW$MrZ4R
z+2|ZSFbBOv4_ty?ss}DbFVh2;p_l8<<$4#jyeo9)3iK-7xeC2ncdkaS(Vc71-{{V7
z(BJCLZ_&BBGZ&qwJM+-_x-%cWL3eIIZ`7R|(FMA*0KHjvZdOh@Fxxf9FZHzdRxF+j
ziziRoHxi#GJs6sq^X~emxp(p-<`npUOnz)N<``OylObD;IS1~Y42hE?Ta7u0R^w#J
zR%1?rd#7<=PJ;i(<jGcJj-u5#nX=WGv*6ww1x~JPHRdo{jgu`~jX4ePbh36i3~sc~
zVQ`~;&Vn25a~9lapQGSL`y2)TkDLTI+UF#^^Ezo%8f}}l!i@zqY5_WfMrEKE)9A(Y
zN&AY=jT{=CgI-Fbm!g-`sO4CHIi0wizNqT2q){s|zLHK{iSgAmYBk1J(}}Aw-ad!I
zjrKVdZnV#taHD<Bgd6R1B;079Bk^wkb~zF6_HUOH;YRx$i1+*3<v_U6KIg%W_Bju3
zw9j#HqkWEp8|`x%+-RTE@a}&ZRV&-34RNDRM%AI~WmG--ij2M@pR})7+_)yAuc2?q
z=o{!J8P$aKo8-hM`J$@dETft+-Yh3JWBiVcx`XjMa^f9~f3HV<uRmd*A-%_g9#x7i
z)1%AuCsn*$k1j_a)1!`I`D6OTWBQA#{0Tkk1jbM36Hlo4oXm%Z`FZNDd9$)sm^DfL
z@Ea612sA2e6lh}DM9>Wqc7sf^JG8@?!^7s_Jcwi&mW;Fex9VYA^-1;uribLM_#if<
zbs}|~chE*n=USNY=#QW5{QIE)YFJ|NJMTYFCzdVr{=|=^!Lf9>TC4^d+CZa~_p{Lb
z44ln_v-xm!Cs)fuYk9QtdJ(!_fS1Y8Wis0C2i|PB3@t@()Iv9E5TB-pPSe5X^s(pk
z*X+$;=bXL6VfUWUYfDO(lab}5?S$hnj+3b4<SDfxRV1tmhr?<}cnuDRT_BMca5(HP
zi@M96QY)6g!xL~AVFr(!fx}^SVq~3YJ765fMG<vTJf&8=QG_+(a9Fb#(2T=jcSZPJ
z91fc*Bd4n2u<U;&eX-v;DB8h|L&S54w9SI&MimLKB5mVA+^8W#YDn9(T5en*!!MAw
zX|&vEWS&OWHtfZX1U@8zw@sMf#tc4u1|MOsxVL@bsZwb!Z+#Kos1}?X%_5>%JZ7&b
zH|~ldcSYL(G&iQo;ZtSXfcetnmtGBZpMAvJk>48jR<xa{STXVN;~h8rB}h7qyCmQ)
zf$^qBYBu8YPG!t2WAOCI0zPN~f7V`JhjE3QSMceOZNl6p+Qv0k|8uh#`*!vdGpWO9
zAptGKQeC8l{JMp}fJ+Pwh@lXVp}&s70ha{oO2B6Yw^MUF^{Do?)8OqCCSY%;L$>3w
zwVjUMuBMDTj2+b6f%SLL;2l_h2OY8l>+hhWcVPYf)ZCBt_tW70SbskqvLEa3r=#~{
z{R7lIfb|d1-~(9y03CLK!Xb5ljypgntM#p*Rt0@jxsqCy=wsA6hCWWM<LDFAI)Oe(
zt&`|e)H;PeO|8@DGt@eRK1;2$=pU)|Bf5%ORp|57Jdf*tp7uYFBb9a3tiyO64X(rO
zbD6p?Q^<<lqUJ3ue~Sj+!t%}3Y{qyq4QZxrN3z41&fL=(v^Sl3rsD%|)7il3tnEl|
zEgwCdJ*&3=bT(l+KJb>r+)3D867wWsdr53y61JDbqLZ+_BsL)l+gre_1q}B0LS`*Q
zXD}-Royn|BbQZI+(2JP02)&qDi_zK4%0}lfD+j%VSxeBl%#(}jw~l$%q4St051r3E
z`RMh`vmU*Hc{ZRoGS5czCg$0ME?}Ml^k(MSjQ);!zC&+eo-OFD%(E4}jd`}Aw=>Un
z^bY3Pfi7ldF@y6_F&k9OhN}Ki!pstkm$2Xx@8id94{UtT+}~q=sbFRWmakyJ6<EHK
znUxr?WFeId`f~;!l);CpGj#@kB7=`n&g6qKF`mgsW#Y5LOSx+)@2bidbF-Lx(80wx
zb+nidDaLz~Vm`VU$2Cj1S%URTcyI~UFX6*VINSo1@Np$Lwe<)$k6`_y+&qdd<7OH9
z2X6j=F6U-Bx`LY(=t^!@qK|R&82UIjkE2g;^91@NH&3EZaq|?qnw!<_-M<=dan5n`
z9LCS_;B&ZrE^zk+4q4ml+`NwEuk+yRSpEh#Z(#fe54nLeKN5wNDBywIX~LR@o-VBE
z=o!MAfld-u5;|E}$>^EFnu-2gSf8W65Y`vyS;Cry&Jb1xI#XDg=qzDnp%)2j5qhz(
z7NfI;m5t63Rt|cJu$G{g3Tr8PnXs0jzY^A0=;gv%j$R?G73h`1T8UmItX1gM!di{a
z6=tr0<1JSV$`x=tuM_4vjIR^H>jX^w-Xz?c1oW4k!rY1FcZ%ShSiVr0g%~dsA%y~F
z3Z+RiO?vD*Pw(@hGB`~RL=Q>B-x_IhbebCHcNpo?Ovn1^GB_RUr^{jK5^mws<+yY?
zS+$=n&1|fnErYYMezqKvjo+Uw$7N&v9BJlY{TvyTgHPV)$j5TzNcDX=^64BoPI-lN
zuaMo93#7ZCy*&jI9sn<p0}Eu^ow;}3jEpXjFz=#3PAHJSQ|)b*=4M>J@1*%1dW$r-
zptnkMD|(wWx1o!rS&S}`W(j()H20$SNpl~1zclxw4@&nzT;GH8k%KsM>!9p^5TDyV
zC`TNWPpkDiD4##5atsdRsB|C2_L4L+NrV2Cqy;Bw(7%$jAxZe#HAx$tq`^FsWX(**
z`pH^wGS*Mlh9qPCWNmaZ)=$-}RP9l<ys27fDn2lmstr%oMqzwxss=9zNYlEeX}wfD
zO?xyAAIeSBhNR*5r)kfnsR#NUMvi9Y;PNlg%q8fhnz<CcOf#3EztYUF(91P*IeLX=
zu0XHU%$4X>nz;(SS~FLpzt+sJ(Q7nw4LVOV^V++Ao(3~c^EERc<M~=}KJK5*ntQVb
z$5)BwE@^L13BC}aL>pLwQ?E<3=n@T%uM%xSi3Z2lKFz%k+uNsk_F;Sbw1NAuy?t8r
zK5TEFHenyOSD{%I*k39&s}g-ovyP#UYu0h}3C%izKB-wJ(Wf-)6#BGgokpM0tTX7d
znspX^QL`?hFKN~#^kvPujJ~2-SI}2A>ni%1W?e&H*R1R48=7?k-KbfO=$o2#6Wyd)
zP3T*ibqn3BS<UF%nspm}N3-srf7Yy@(Q&#Nr^ERqP9GGf4^_u!yl%#0JYElu*YSL#
zyOVV2kF|QITJP-hd#||4`+M<%e~#Wh$D5;EL_-$g5B+Q!l8s(MLzbW~%aF@>p}HzV
zuA;BYkn8B9ddN{d+WumP@q-@ngYxcQAARk@uKeT&9L6DT9^(Dg9eOD@OVMe<N)zEK
zo-V9(^a5cmKxaxbQ^K1(vZR@XE|g{=dY3eJp?6DjH#$=@Gc{;$k!CJJ*JxIa7OwWt
zdCfYHzMxqbly{$8@M?|R^&9UV=;hZdNM(t`{lY;{`#lYs<TnY_Nc<W}kopNWm;24-
zL23idlzua1kXoQ^y5BZE$leo0hfDYEtGmAaze)!0ZTjwL)6n`p-lbkleHT+d_3b&-
zHwV3x`YuIZk-k^d>o6SNK?2`v=o`}a2D(i5Ez|w%rF9tPx^KC%J;B%e5cR!xtF<L}
zOG(dC0uM=^A$`x_TOV$+o;UF!>R8@479Sd{7Coy4y!GLl=zC4Q_2I}f_q{hi^3`f@
zPV%qOe|^TzNtQM3tLl7V&A^ThV+%F6;2ig@)ZB`WV`dya^cm00cyt0Y6VTH%bD9Qo
z%BO4Qbo4IG+=bq)nY+<@G;@z~#r1{b{ycw1;(xW<Pj4Ajd4Ju@)4ePG(6EQ1?aq3(
z>7B1{4|~Di`_KKN`?bBX#eQo|WJmj@c5VBn{m7u>KSl<<WAE)o23tEBjtqM5GTMy{
zI*j%sgASwp$Y85+WYA%>9~o>ljtqKVoziY(u+=y+=rD@u?~Cb6s?+bKzu${)qGOsU
z%=VtcCeC3ms`4ppVhTEqO-w_VvoYli>Sys6vpBr=X&0Zo3tuC(hfm&vzQ)H~!`a)P
zi7B6nm(?jDR!oUS$BQZP=pACr4gs^nQ)GAwz6K&ij!uzd)F~xZhNogYRgO-@c)kqJ
z$9TRRosaSLGJHM8*UQoCF}^{DZ@~BlIeG)e+mBRwUlr4Cq|!TkzTHTrH{aiGq_XuS
zgd>#>qy0#w_xYK2BbBYjkxGZrex$P1I8x~_+K*JW8b>O<uN~1R?bIi$lU)0e%hu;X
zaOAS}4RSbg*?LmMkxPfse&n*%ICAMQ+K*he8b>a@FOblmt<s;fm!~md)W%+Wb8r05
zPbA&E{mp8l?(Lc0vttc4Yw$%ewbZOd?_use3|?kg#N0*bAG!G>o~_pkbFF}v6D^YF
zB7Bj_9%=4D=W1py{;rv-yQk`HFQ@GA-s^9@()y0n_uqT%{npoSc6hM$f4}_cV*jN>
l2bcP*Z`)s~vF7N)-%N3R)FRyv`uMuKba1!DI{2u+{y(D47OwyR

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/transform_conv_fwd_to_gemm.hpp.14B4756DD59EEE91.idx b/.cache/clangd/index/transform_conv_fwd_to_gemm.hpp.14B4756DD59EEE91.idx
new file mode 100755
index 0000000000000000000000000000000000000000..73d34c5cdd7da53a4f0109995086b80757a5f944
GIT binary patch
literal 10592
zcmc)Qi(eJhz5sB3zcrfKG&7EyFh@VA!)rDmTOMvnyoeGa5`l2L=((q|_o+M<^uVNZ
zfDL&{Mc(gDS$V_c0A;0JH6arM1$kH|m5L8cGdqG_e1P4pG_9Y1fS>$)+?((I{nlgF
zthLw7-h1)FsHkMa#by3u3)7ZA{=zcB#l>Zu^MC3KDSwFt7nhfvFQirNEVysQ7dG*=
ztBJ21{YS}%k7no1?C=V19Ca=s`oAZJJQUOQ(&S^i0{llcei>)`>o-5y|2FBnmb7|p
z*@n>c{JiN;&ij|&oms8Jw#jE7w5^V@HKaZL!sCzKo!$9M)7J;i7mvE@)nk>PKGHL>
z_KE*1d@OzAHQUae7d*<ty$2FLdVAE4_J$zuXKlHDUp@K$&Nu(E_rb~sFO{_PB>Wt7
zJZZtE%RT8Ci*NU}NpHXU!rgb&?p`*uI(=u<$4}0~s4*v_pmom418c@ENp$E*iHFj9
z#|5=cyx*=yL!2XWUeJ=1xixd=tR1{>!S^d#w!Qy-&utr`rNC=fKmW4Zqw|}RAD_GH
zeD7&%^h?SAao+8_zt0{2YFt-SQSYv_)jl!3X;D8~@?R1?HizE++Ox+Ke%SQ-tp76R
zbyp-!dQK`j^67sXGjsp9uj=`tT)#hCpUZl&$o<5sPaa8Kx%`n-=a^db^0`G(m6KE)
zJ=R5TU`hkie|vHn;~abEzd4~LVG)+jmq${nnNZE#<dG*cW&Sul;dH#~l+ifmb{Bcv
zZQ~~2cyPHLolUra3{le~q|jBn23=8bHe;D+nRMg9EWh-djg#E2MSiB@C}%higgkKj
z?a55})8X#^g!(<(T#-3b#!nnS>BfT@e*DL#jz9Gj7OHrgGn^|Zxq|-NlbKRh{_wSw
zLyPB+q#R+w5$AD?JeesQUafBX^z4ZrMp6cuGUyzQktZ`{*O3e3%WtpWGSbz26!Otc
z9(giTuFOe(bowtJc8#RC3+`_6^#?QTlJfWysdsp}xL`icsOjyQCyH`{mJnbT6Xlhb
zl|Ecepfy2&IYgq|3|cewWw2ZZA0EE~mMhfzP}_$AX1_(b4YfAvcGTLb*D%W(=EL*X
zGRs=(b<DDkd-V2@DPi~4?o>oM&L?g@@Ad=!4@7e<H<m|oy=Z;$Uar3@fA{6u4Zgb}
ziXToJC~e@&eHQ%BLNxb92)GC_+!?53AjxmR=`FaBAI{h4|20N)Uq%0`)T>$WYIZ-5
zmoooSM)EtDwS(PjuA4h;|EPGnc*b`q%H!R~dz!OUlr8QSPjlAG6AsL?bsv6aj3Cbn
zo)uyqt031va1Df*eHY|<46erzvu#1HW8QVl$6SDd+|0b2nUC4FAfIBkQ*4Ggu?0Cx
zuw@A|%vmhRrGl+gm|?CBvmA{5ZH1f5cEZ6A#|*u{HE^^b+g1B`-oMSDZiex^e|4zV
z;dtJ^J|^`snYW)QNST7n+b<EM5<xcGpBDb{XZ^D}#@^y)LW5pge0cZfi+71~ycR#x
zJQ`76tSu%=j!b@t$5v=7NbGs-d7?B+8d;pLSY9F81KJ*#X^xO6e+AlC)ce7{9~SfY
zInd70_<7LIQ(pqhCAzq;g5@gpUexyDO!Ewg@_y9zQ-6!vx75R^4O3?_Et8S*Z!pUn
zjJ!GIGD|LX7jx@kf#%E!ljeN+P?X0?=WPFoi65~5b83q6X=Xjm?lY%?`MPtejqDWW
zR5PEQf$Q5u<hfUOTs)f6Z>|(^?2UcF%o61=&oFO(mKJ(0Bzns8DWZDttOsvC(w~Fp
z=hUab^Az>R==m{v^O4+*p1Y|pqUS~Gy==-}Hq9IfQ66Sfhna59^#$Ekvn!Y0oBFIM
zhx>=yd6&cIhu_B)tHe$u@`|-_#Vhe7mS9V`hsP4;Cd6<hDv2aBNl79~R+6b~$)uJg
z$`TS=sw^dXMtO$H>lxCs70L<{OHopYQk7Jqmz9@^RvIgVc&}H6t&HMYrK}>E=}J0L
z2KZ$FS%4W3n*k5-vY8N@Nu33LSv21Peh%s_;J*dzd}?fg`CH&V?nY1=VIsA)5rVn*
zfU<|io50ot_waZV%x$7uP&3$?X}lTcHq&?uC@r+UR!~~0+rh7$zA3arY&$%_+v@<o
z4jS(SzfS5ipq!!godN3^dOYVqIY;CDp!8D@fHFWm2+AP!WiT#75bw`r2)j(T>?>eg
zq46sac7?`=KpCR-4TCaFor!*#NZwR4F*Xy)@n>Od7PSNY95g>0{j#aIqW@O3^QFEO
z@7;=Ii*G`u2`BP+Gb+v0dr{d-y$@~s@E)GO59jW~816Q-wb6JR&TXUdcC@wAcstH*
zr|}L{I%s=csB}>uMCIVkwjHGP9YW<0jrX9^LwyOAOVpQ9xlBEb$}n{nQ?eKtpKPXN
zQ@_rX*QwVqV+{-9<FST?tzl8zYnic@#@Dj2wKTqt8S7|#9Sd7W<LjBSp0=0Ilzi%P
zrj*}oTR9`=tAg1os4JOLN%O0iQbqj{vwg(wG50i4Zf9aUyNhoH?QB{*^W*MfVi%2f
zvDhy50FQSwv75#ZF!2EOX%=^yC7XMxDCY=9ju2#SHP;=UBP`^u5R3{T$h^!F<q9FZ
zLRe@zd3^JXhifNzI&U+=m2f})HoC}X5xFLR+UIGaEud@xKfc5pz_$SwaGwU>)6`p0
z*@}Mr6z{?5d+3xN#OZ_7ZOpTcMerfX5yTw9haQqxA^7}$2v^Hy^lf(p|G^i~=HWXp
zUPkUw&Fd8NvLDCZxG6PXzJ4ZRDuh_`F#dczJNc6*Ruqg8W%~rX2R{>=VZvtcFsm2k
zI-F349(*)wnR_ks<rA=pxi>LiK6-u3(#NLo36m*UGKDF8G)n|ai7>_N#)BtLEp!oI
z+&)^AgC_^i;oXRw7CDcr8YWl69NwvIFl`&m<F3KUH8_WNatBV^f%8n8jXB3^<Zj#?
zl(gNwCteXc+k8aXU3Vw<xl{BhL{s&tMAP(XM7pjMdFUQQp1LQIm+nR6t$P#s=srYw
zpy$!M=6ulesS7|ape_Wxkh%!;BI;t$i>XUMFQG04y_C8P^fKyl(95awQO`$mS6+a6
z0d*njh15l;7f~0ZUQAtrdI@zY>ZR0WsFzWfqh3y3fqDgXHPfs4t*R(*V)`cP1{To3
zg87tfWC4xTN0@$u`S2-xl<7yQ2U)-%3+7Y$5(~IQoht<73c=>A6Xo?nz<Tb#J$7*R
zfm7Kj&XqSSZdNj%<TVgi1Ic_+*JE5gCYzJkxe{4K9b3d#UNehmW{db?oMJJjSQ77c
zmJpLAB=K&R3NfWZlG&{_F)aJ%^}o(q1J22BxRXolDaI6{sm4^IX~r}n-O!0V3=bkt
z!;{F%@FMayyor1aA0n${b>(B?EBO-nOa4RwQUK8mX$H|ub>`i?7Q3gN>@~C1*(A0=
zT|gAAMiV`(K1{SsT}JeX`UugZ>Z3$Ys80|*sXj@R2Sy&)_*u>eBcHkei~{OHFbb)Q
zz$l_F2BVm|1dI~uQZP!X%fKk3E(fEWx)P*HaOM500;!7npCJ8{`Yn*&qTT?~2I_Y}
zdWZTwklv%-2+~IC4?z6@?&j@p2G7lO_pbxBj>flxx}ACls5_{4g1VFXAgBkazXkPM
z>K;&gsJ{dCJL(@m{ek*0sE4VKf_jv?7t~(r6QG`;J_+hc>ON5WsPoatM;m_^C_tlt
zx)6;*>LN6XsEg4krY=FFgt`=sQtC1^%BaiHD5tJKqk_5$r7Cpg{d*IoH>ux3=`HHF
zQF@zt14<jH-$m(N>i1B3k9s3Y8>!z%>3!;Il&YyWp|pv*7S&q3oA<X4J?rSz;uci5
z(D*J?cTqQ>+CbfiY9sX*sD43x6xE~Dy{Pt5A4ByR^$Ao@P@hEgBy}IEebhgq`Xlve
zR8Lc%LG=vvSyazb|BUL-)aOt=M|~dE^VI#Q_EQg_IzW8^)eF?s%&2BIzH4n_#wO~`
zOxny`dAx>6HPp3Cs=e8MwJeC|H?WWf7RGnMMi$aYy_-q9nJbTf$)qo-o0!yev;CS_
z5YKO6QVWf@GO3mN2s4f_8{c)0GUF)qF(w^jt~`F6Nyn*AGU+6BADh~D^TkHUAPeKW
z_$3x{iTVnYt}s^~zsjVm)Yq7FjXFb+G6a&JDNN0zcRRU4NUjiOUL{<=cnMigoi9lF
zf-8>~2vPxcp&%7f7YkA`b%`*wgx^h><KnzDANl&ryfm+R)Av~V7cbiX;M**7T%G+J
z`R>i^-#e2VmG8fQVTg%xWMpIlzt-3Wk=r1F?_WDGat9`u?>9AR+egJOe<$iOQBKqn
zbzUsdnixnFni#>Aq$QEeBx@4MObSgRnM<@KBy)-Xk`P|YQf(=TEwwHswWMe%B$i@L
zp|Q{u(n6}1N-|TesU$NsG?iqo(pHhoRn}D`b5-anl9{fhlgxB$I>}5AOD7)%WPv>k
z!uc^}!NM#cgYN*lgT@^Y<$ySz-vU|-==|igfVBkzsY6>Jf;zed;<;NvYo+B|!P-j8
zw?b$uE#C^!t+adx*gGJc*Vh5D9gxJ`3HDAJ?}VsMh~x2o(E4fn{b22<?e{}yKW)Ds
zqWfw41E3Ai@&jNUpydZ3bbyv0fan2Qeh9Q7T7C$uL$v%5gbva2Ll8Yg%MXJ#Ov?|0
zb(oePhOl9X<i|S<al`b}W(V3GNXE~Bu?{5TmyPyp8qdb4Y$W5?iIz^BWG+Tg?nJ#4
zeW?RGk$gJWi4mRj3)@bN@1(b6-Dv5i^>w4(P3!B%z;0S!H%4^R`noZ`o7Q&-wL_@$
z{v5)fLpYne2elp=??G!12J*OrSsaWUpM&WRMvl+H0v+_zMhA;<F!F`8gT*@-x#i4Z
zmK<7N4%2gJeK{;Jht`+FB64VbIV?Vh)|baDd9=PfrsvW6@>pOVtuK#7<k9-_SbQF>
zFP~}ojEq+S(+a2ynN~<$#Iz#nVx|>SS2C^gW(QR=GX7OetD^BLX02jm{JWW3H*@C~
z58ceGn@#6Fz}ybd_yOj1fX6@i>6K61-su0}7Jn(d<EyWi{A>2W$y?lvIV+u8=g4nT
z%&oJzt8?$(GZz~E;H`asm{<DZ`zMv$;$i4Ib$Cbp51+(1FBF2qgA@5)R1M+PkjUTb
zYB0P86U|Zk`qJLx<9@20>HJ{yIqf-rUhEgpegS`8?BAgMoBA4R*U;arMwEX=?N{my
zre(0-<KJGWTKrVm8>|1|bJ06%|NOz432=+gMSs<2y8`N`dfeh>2CvrDxSd(PZmcLL
zYsusn8W~{EfJOXVW`aGFx*hE8u!x`AOtfbraTeOMs1Krckp5h%oN49s7ZVjstDx>;
z_AdU3q1g}T0=@p5Fmri2e-rlXvg<##OHFw+HSzi%cU*rM`FE%JhYQL}-wpoj`tbjG
bd-OQv4pF+(MfyE<<EN6pN#U=G{LlXf25;CU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/transpose_3d.hpp.F3212DEAAECABFBD.idx b/.cache/clangd/index/transpose_3d.hpp.F3212DEAAECABFBD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..0178c3a083d5a044d20e88a9a878debda8dc37a3
GIT binary patch
literal 1528
zcmWIYbaVT_%D~{9;#rZKT9U}Zz`!5~#Kk2=nHv}x7@XJ`7%JxMof4Qo+d$y>PYs)#
zt6B$RI&QIRKT7n9{+fEz`JV3em?qKW>Tff-LVv$g7MNwS@>cFlrN~2e|99TMD?j_i
z>E+LRs&q7S9TpT`u9MsHN3tQZ<g%nm(1Ep0QYLnu{J|TSl?bkBdKp_5tn%^UFZS<s
zftJEyp+fQEu7dF%ZzpCm{=1%9B2_qPzuLFg{T<I=zZba?zr9TI>w4vkd>Lag8TU++
zrK$Z3zes#vkS_Cc(bfyjlPcssC~Q)n->*6ONA-&IO?q7!CoH$?PhR^ZG5V_P{gQib
z-y4!oxbM-q86e_(eCzgO%+YD)K6MH+pHvGu&VKQ?&0TkCy-v;kUh{ntWnoLFS;sLq
zd9S~;TE(s_HUG$`x!tj~-YXk~oV0?OSK8?BoX)%P8K=zSS*x#k`?svLmjAJsX&c*e
zN5Pe?T&e;a-yOSt>el}WFD|HU+SJJ7mn+Xx#x{p>*WKVfri}g%HGjU?k^OyV@{0-o
zi_RUJ)?a?3$U-E2eMa)xcR@a<!dWG{kA$YEx^x}W;GdCZS$Ir=^WIyB`JeOeD%_3Z
z{rar!Uv<;N_hP1Q(=O&`*n0F9SLP<o2gYB$|K|$*P@CmE3|u@wE)WPwu}Y~kaI!El
zh;lOuF;07@n9Ln5lP|=?!OI||DWE9>CKxy%sz2X3;a{uKv6P2_6Q*B`RSc+~nGvG@
zhl_m86`RB7xtZ92`qjA9gunz$|DQPr`OClW?P6nM=V296VpmcH6EJCpqSUlv5oQJk
zu3ro8HG7)|Gjj=vu!#uv9{Rmt%|mMi9wG4{@mRPcgrtL{V_{wrT;C~Gy6)9sRxTky
zHbG7#EjHrz2rc41;{FJpp}nCK%*#ANMh-^KFrL@U{r%^f-|%t(?KBg&MraXt6Zb~&
zq|KzQVJ_eil6I5!M(}h@bSz=U@d)Xf>RKUqhL(nQ2%eFZkv)v(Yd$GoOt)tOBae`X
zg@`jOBtraz)L%@rVPyuohf|y%$pxI!{0KX>d9{UMCRHe|7tc7C;>H5hq9AUFq(wp6
z5TQkdLq!r2y9^9GLh2mqlCVgGsbS&-rxHYv>htL9!E9++Twmq8;^=XpEjpq)axh__
zhedUfc!pAj$_P0lDI++q{e!EF<@p{dpoawo1jS(S2vf|+1x|sml=IYp^$R<{csD2<
zM8)NiJS8eE4|6b&khY{Y+ya;ypf^B?72ypN9uv4XV18!e0w-9QbwHC;v~^%^`L%8O
zK|Sl!n}JRRrCnI4!4w0X3KE7{@K+#g)7|_d&wwdXh)oD7MM~R9+asJR?IZ1v;OSWD
z*uty>hM~1CJed@w<|Kk_EG|ke4hAuVb_99$+_~t;%_btu$1fnrAPf?P009_-fq@Uf
Q1Qp>>AOd4BFz_Ro0DaX6ZvX%Q

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/transpose_vectors.hpp.7945D9ABB9DC55BD.idx b/.cache/clangd/index/transpose_vectors.hpp.7945D9ABB9DC55BD.idx
new file mode 100755
index 0000000000000000000000000000000000000000..41e7f1503951e0cecdd0f11384bcf22f620febf9
GIT binary patch
literal 4920
zcmdUzi(eC07RTp?2qcUo%oqd$8z7+uCaa`{kQkN}<Pn4lSPed)@(}HkgaiWWTBX1S
zMO{T+E3d`+s92$bZ>#uGQ9o^cw6?Zu)z7-xReZEvt-4yp>fR(==KcqM%;&?MeDgaq
zbI-Zwo;kzx)a2wynjogkPhDazD$nN<1mOe!Y~^KtJ^%<}2>!67^V7z6EL$RlD}H}>
z*>=NP``<a2OtqQo{`B9l>ECzuexh6>@{nH1kDvKs__B9M^40_TWlOsJk9G;R-0JR=
zb?^Sh{bKNrX20iKKGdZAc8AHmfAXPX;}_OdIUX`iUuN)W@x#k~_qMyQ>-l-GVclJ?
zl%b%b55xg)U9#WaJgiL|f9j&JXyWxGkI(^m++h5_HZ`Pk^7m!e_Mg`(j-1%-xw7S4
z@B9THzs1bmG`R07`A6fmFYcF^G?uCVys};=DeRfhud`&u3qIQztZW#&>)zFbZmnC)
z%*D57M$Wn?jXm|lz3x1F_Lz%_?+QPvv-uu2cP$Q{rWo8fop00(BxxVLsJ{B7ZYZei
z>2LK}=WiQ|lP4#yopC7Gl2DVXx;PM;Q+?hv{EPd!m#W^Ad)+G*W}F&x=0ED4Hycjv
zTR}zL`_b^CaP`}61?O`<pYexJxAxsSHseI%S#iX)ur+s%$LW6Bzr5=Fm)7Iwgd6gX
z9k_e^(B-P@i-ZNil?A=BkH(RLxCKhkH)-y+V{^V{wyw%;s(dbZ|8iB2>e98n8E!Ww
zhP?OWCvoV+)n_-?9o_M8`>)pTQs3MB(jD|nAZ&jz7ld&L0`MyHJEKnC5phHUf?gi9
zT;?1IPq@bLqxf00&YyF0sXIqJ*7XZ5@DTVBJkK!@0S#hK*cbNj8V0;Ld~bqQ^VOtt
zK<DT47G|t_vZU9G;{|06Uqd<vBrpDT$xPoJPh=$FKbD{=Ar<Nz2%Z}lA4tN>e-N}Q
zXS2Mw!f}6hTIr$ma}MbI_NjUMLNcl%9I^+!)MKfib3k(E5#NiCyN{TlbMWp_K~Xa2
zfaGnqft&eL)L+5LLEmctzXp&a2V}C}{f#X?<g;uq4iB1C0e%(m8aW`7gEy?t7FKy(
z5pjH>No7!_%sC)=u*#}dmyUl1WjKfEplF$MK=N6NyQf>frclTk2W9OzZJ={N@`aCQ
z`!5TY--j~1xE+kMgTRpkGP$J?Ux`*!H^C{wE5z_)yqp7)?>_CTUsKROmLvjThz&wR
zsB<8AT#lbJ#1Foee)&l1SqfJW-l7`tt6>ye2XqahtZ0Glt~Z-49KJJ4Q1~78(rOTa
zF6CDCfmU;3NEb-bGH;n^bd~8a?=Xy*$xJ+}F<HzkmK2i}n5M{5VA?$HJg-r==sa#7
zM)}_PB33K#F2JbJyAY%K-t#dkmKRTA&nwmyC$nUf8!^qOGh&)aZo)K^&V*@Zxf#>U
zIy0t~$V)J-L|1}o7P$q}EIJFOS>;wtv+AsvRw^&Wv{GH^s5UmTuD!TLdBvTi<D_xn
zDCtOAC)HuJ5=d8qaFlZ-T?eFfIIjZIRXDE(($y>r{Ju+wyl~a%$))v5{hP-x=hpXz
z4%c((bS7QFQWH=%fj9Tg%iVK<8NSM;n}M<!=N6!B!MP1EZ9u`=w*#gf=MKPhu#5_p
z8_!v&`n#B;MTZ*pldiJiDq#fasuCll{Or-r48x{iE=g;{v|2Vev69$mmh2$R4zz4=
zYJsE{L?c!Xb?p@AowATzIwdnDi*;8&Q*U4?ZE9K?OX(TunJhJetVS@Mb+ic>nm{(o
z&A`x%a|<xE;M@w*T0uH%-v%<<Ko-mGAhR9k4v^WwGAiV7YdfkqD!ecz3ED&y1$DxP
zggwtf1h1`^Nqdb=8|roY;99E%>Hy?4d@2bv0q9}nC})Cd5P-^=<Dnp^0RiZFF>AP0
zIdu51QLi-7b7DK_&0}ttytYe`m-%ncO+De6e3GWZsYtf;Q?b+(j8dsgj0{vROAN*M
zv!tLD7^Nhnq@uk=()twrG?vn0)26VL9-pp9gi-+R0!OpL?t)lVtQzftqIcqtRPWkW
z0F!8`|5C|F8G%)$3J9uz2=xq(G!j>X5Il{<H6R3~5t=!?m!tDxOI3z2BNW{W%>eM7
z6$}}zyq^@P`TB1Ga0*I)B{qc^L5v7>1Te?N>JTi)#hMW0;F_s!_Fw8>s^JBqL{WjL
zQzGeTQFI{cSx9=hXn7!N5lOmIR2hic9rWyn6La)kXkwlJV18USSP}C|_XhF?M%K~v
zp%}t=NYxQ29%vb$k*oDX9+7k{P}Kr8D((G`2OflN)9V6An&B}%XxShs84X5y21`^0
zPk~t~lnTsJp;us*#ggI)Y_CzQDvo2xC^2H1QDww5lf;B+CY1@(%n~!EnN?;?E0L67
zT8XLz(<~ATrdd=LOtVU?m}XU3F|AZmifN^)QcSb)Y(A_<<-Br?DtHwbE#fU2CA5iP
zdbzr8v^obDTsxb{KYbl0E=5p$cJpbuTG*kwJ|t}*447sITss(5Q3}JAceLy1%8oKG
ztpO1zF%HLu_}r*_ycXUm(KC@kJCdZ6$RuBu>VaoHpir`q^m;(92fm2WR)Y6(^!{ur
zYNVP7v>8AD_M0<j?mjy9N;eJi2ymQlAZVsv$s3;sJsbRe9E@qWLCLOUc6c^Mv%+U#
zG&_7YMsuWd81}rm(zzJrNOLgCmF8k(qig}}dF50&Mio>AMvJIL7&U@~MlcmMJ|x`)
zQkp;p%grFA8Rr&|(t>j<h;0RltbH4ZZv&|;w}be0oI5~#2g_*XVH{oWYbd{dc+@%l
z{nphhU+F5NX9L_<j#ovr$|%e*O&za@sG5^>7Z7%VP~;POltDAs8zS<JguW)&K_W7I
zydw3W?DX#cy|&#+S&=E<QA=!Pg|<nKFZ4s|TEx@uKkRoCQa%Lb_{sGVqA#w0jE&I$
E0os5n;{X5v

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tuple.hpp.3819E4E0F87690B8.idx b/.cache/clangd/index/tuple.hpp.3819E4E0F87690B8.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cfed3c0a380571aed5bd84b4b68a86bb821326ee
GIT binary patch
literal 5602
zcmcgvdstM}7C-B70B1Ny1{6>l5C#Qx<T3Ja!yvB_2}D$ctAUSDASK9T6!5{-k`fCg
z6*Z0Q8d-pnFEk-)EXllb4MVP`_E4{wAc|Jn^?mo5Gs`*salh~O&*l4An;(0fwfA0Y
zueE+VHF5m-^DYcy$WAON$XS%dF%08Ge~pU@=O?fXGr^u=N~#;urYtxgq`Eulw&Arm
zxZBMMRo(}CLUxS(4Lslaa75v|Hlv&`90;u0-LvAo8y~#Fow>bo;JM<T`@Z?<hkJ1^
zEtu8k`+L!iYbzgwCT{q|d(E-<9Ob_=ViZc#(Z}n3COwLJKXzNk7Z<xyvmItMUo2X_
z=F|!A&xXgW{^0DB_C9v<5%u&XU2$=S4R-fayKZgpe)qGr*G^vlS#eVsK8u~%+Hb+>
z{=G8`rUY-A03Y|YEb@DH#gZW|wbd8%rad`%X_{?POoYRli3yK~_6{u=Q8};j#-&pe
zO4lV^+o_ML%;|nBB1-sf+0ZZ4ZV~eiEuCGWzGwGcRm9t#ccV00o5BmLF7L=_II=W!
z^ibnH`{S3pZ_kh0G`?q~!@w~yTKlv;wYHA^pPSlUd+N-qF{#Urg*N;&dDOUqdG#IF
zi)T-oq|)uEoSj%2shzm+t^Ngd?>P5_gnfTM)%&~KAKjk$!BK7Jm92WOIr^=w!(Jzw
zPtULES~<gY?9PN6|2iG>yZUnBv9)9D_V7JNzHnV#*Zm;rJ~wFYGc(x>WV_LCMQ}~=
z^T8jN>?`ij?U+#GvEpvlq32x&CXbww79O-O$K^!a<?Z)Aa$yQL<ySP#nNg=Yv)S#r
zD<@rMg@68H(vq&d`+Cl<>%K5!b?4@!uCJ5o7w7-=?E@t~1^3n7t|ci=x1yEBo_h~C
zUwW@R^!YPOBA@fL>um{t@$2e*)!JVNuSzP@HH^I-zTe~54_{h8D9J0Q`iBh*cC4G&
z2yIg?T*%V?b^gvF(TOkj?7fz^_3Y6Fy@7Y$eB%UTTv9M+sV&3UcCUH#?Y37^UD*+%
z8Iqz&(V7L*kKQ@@qpaafV>@Tp^k-vq4Dl6woy`KLh1egw9Lt5nuI#WNhB#@QbY_9m
z&XqoU79MI!bY?vQ7}5#>tq^G$P-Rehdq2P6OkX>8c%TzWwog`>1x~9!G)>)@epGN}
zef$|x2AVSPvka&bxUp|xa%^#higlwZRWQ5?Mp_0`3EQ#nz^->U3J$E6sw`G6?r#=2
zZHTd3`D4eDJZX_qO{pId1x_cNuGYVk5pU#~*ij7W0G|%<w+xIz!7(ElAds}xRQb)S
zUcrv_rb<S8BX1Ur!qt(wkt(E~XgBaza=KhuKk7-1MiXilxXARf^MAYi&mNU4>rGwN
z44P(oP-#GwjGLoZJz28H&zT)URa!vT0-=@xRc0%D?fROI>P2sw<HwLh`$QF^7$``)
zn3p-LerVLyM?4c1$dFOYC@-^M6iQoth`7!C1Ho;i_R>CQ)|0xb5ds?NBGQ11yiEEQ
zPCdQI&4EpzM{NhYc5tx_jKbMLT4*6r*1wj)w5r%~)J`=7R703$Ks&`tXU8Omhq89;
zcxvZ3*dK?1mceRg^$U(;qaP)4JQGiq&Vl<m@URSw!VaP};-=rUKeD6z%fgkS^Ht<9
z*dGSoGT>tCO&327%Z^j4SOeWp9r)Ejpk+XnjSn)<oLw0;%#IyOedB2F=xP>>VgN6l
zsjAh{?UmzoH>nL802BC`s10dA8{2Qa>Hx8>dmY$kXq4TQZq8<b)Ab%z3vUgJz39L`
z8_bY4P_{vT%YZ6-PuqCC&~4L0qf9qi4!Uw00BJy#=5KCnpO$pfR>i7m&SnQ>$C?FB
zx6SqF9yHeYM4HYUo~I|G!0A(IFJv~KYi*Y@K24LRC8EITj+B=c%-y}Si{>m{e?J85
zr#UMPsB-43M4Knfkap=3MKY0IL==p|p7nAU<8f}8Nw;B5QLUs9!i0DcQ80>rHj#rx
zh4Y78>VJ8t3zagiL{s8NM1c!@@4lt&7+=4OXU5V<7<>%jW+C3vh(Y4_-w#{*dG6!H
zAEjhcYt*4c6!`tst5xf-jXrSBj`gMAyV$$9n+2nA=`Tg{>Ye0=2Y2@!l4?c;1XMr}
zkp{HU8!~u${TlC+G!E3sS-PxHv%u-!{^GHUUG#OaijAXgs)6A()Pgji%J0V}9De<d
z;<V)G3{8fXhytezb8?NZ(zL`146M2GVnM{pYtIvPggP2e%o9ztW*in1rMVj8ad{Hp
z8=!rf;v`H6bAB2M66GD``!sSS-aUQ@x>*P+-<G#YcY))n4|EkU2HnHcI;yP#9hO+0
zW^i~lL}POE^u)n65RT5kc0i9J_5w^O;b5(4xb4fIdGN%Ea}txu{6i1MsbxsRX=F&t
zXe~X6VR#bE1j~?~(aVs5G00E~mm;&0#-+(n2A3g2SzMM3Wpmjwl*i@CPzhHeL#14)
z43z;_1~Oa@T)7-q09PT$RlrrraW!z&a$EykjU3kjS0~5&f!i;~jlea^aWinua@+!3
ziyWJPGs$r)aIJFO0Za!x^{49fxwWSI=RH~4&PEzzvCHwK$fqa*+bXtsP+FQh)?>WO
zoYLIXcg=gT?x(;ztNA}%blzNkP?b052%<q}_w{@D(I=yAIFdUqHy#TgNAkzz$77x2
z$R2Rq11c=>$QkG@ln4XvS`*#a^b0wq{8QHavmfd`*wQb7BfG$B7ie%9ObU8xiFDvH
z=oWjWnPuR=c~Yt?4MncO)4he2K?H7=C+om{9e7$(uM!+9VSp8H2KUY2iL3Ht3pj3p
z0f^BqJ+#POa8cwgbdwTs*}Gk9%4uGI2o>y<W}bro=FN2&C+U`Dw!|%q{L4Dq2@aj0
z!Zmo>mzC$h5nDDEDY_+*?qHQdrbD+R@f*A?$ZzO`L~d{bSq-N!vl;6(66vC~ni$=T
zUe>)8sfq?;B(NmfD9vogY9QA!HxMrsPx2k}1MyNJ52Blq*biDp_CqHm>VvMsqDYS`
z(H;B(c@C?zup9h>rxv{1AplJ-ce5$oYV-ImE%}-x%@p(>PvTYaZdj~&k{+3Ez$VU<
z%*adwb{ZNleQ6wC1u<SRWsqRKVx8dM2_YCSj1xVH#H(;$=y&QaiA&+cQ&RP`g8wms
zZzkE4-Kcy*Y<QA2A}a`6C{Ovzh;j%*YQxj`j%a}(3^-;K^_)b^&~w;8c%_+%S>q<}
zE<8z&PENs!LEYn-?k{tXXQsc^J=t;DQ>^jM4a$wSqP)nw2^f8P7lQJmt*8=ol@Mdi
z#Vw%Q0x_6-Jf(24Z7><x9r8qa9*G^ILy;ZQn;=p{%m|+Na=tt^e6&K9B!Y<dN4zfj
zc8JF7vb*F{+sOs*kEdKW*(rH2CIWhyMo*%R=uMs$FrQuEhwkP{1GqMTJK}@ammYL`
z<CpqBRUQxLUi`;6<%vsTZx@{0a1*q*3u+nC2pSpEDzpQvn!$=-8PY5CGGtH~WGF>Q
zky%L-(qt$@$dI8dAxnm`g=`th6Y^xJL@1G=QlV6a%0MUs87>E*T#hS1sF33-5US+3
z8iZ;&t^uJ&j_W|EljHp$?3d$45E|vU8H8pzZULc1j!ht#<hT`tRypnfMF%|fXVZ26
zWfRrWTg2Wz-8jQ~O^T)zrDN&D{Gdfbd}v^reH!WQbfx_6T6*^ypkD(x<0blP{I{*A
zE=->Df2v`aqI7%W8wa=79_hP0b&`7XpSsFeG-%q24<fYn?U}{Xv|h(n(4%&OQzs0;
z`iU<ky0-K_LJuO#rWq){i>zhqzqcEt>GTuvrS!jif^R6gUGeRNi({Xl<wJZpSuM4J
zrVR*I7kp39uf>-WmLeRLO7n}!`n6n%wkKIX=gBC)Q4!c;c@p9s5`xF!Nu+mV2>OF3
zh8V+C+#s!5+T<woGOgRnbPo&xZEMO*53Ji*cd4VLHyWN58w{NgA82R`pCj~a(u)lB
z(IA}=Ut*|Vn3JESq4fxP27~(O4F)xhg>#L1v2ulC`C^zc)<4O^Z4dsQwtRFi7d9q5
MlnJxWgj@go4`|ZA$N&HU

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/tuple_helper.hpp.73EC15D672B6492A.idx b/.cache/clangd/index/tuple_helper.hpp.73EC15D672B6492A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..cecd253d159c4566b62e702fed43f561c2d3b553
GIT binary patch
literal 4482
zcmb7G30M=?7QUAtFk}cq46;ao4#7~C5FuppFf0<Hf+#ya6l(;lxS`OfRa~kT5LeJw
zRJ0;d5p9du#%f)tSnE=ruPhc7wP>w%tv;1%ZPj-t!)2x)-}l<)%Xe=6$(ei3J<ES?
z%1FI_ffE3+xg+P#ocxZ#8UTPGD14_NZ<Q4Q&IADSx0=%u^E*Sx?lISQggASbwtRfV
zy4dk#Ms=n1JJHyA%CxFkE6!BZt_z<~{yK8xLGkI4S6m(h+jNhJthljZ`PoZPe!bz`
zwEtpr*Zp7HZd~5@O#Z{Y+)#(CDL<XF&U9ZKb4c2)>3VX*=C$ihx34p&MQ&gBsmje^
zwyoji!HKrxkJe9!Z2^rwzyEfAhHa7NLb7<C&}T?s`8`1q*i>fNy4C*KwQ+BOVGl;<
zz|%!@cZV-Lrk?)8vCakeOlLdWlYer$*3lsK-8(m4y*{D4b=!#yv-H51v)(E>L)z@q
zxE$zwc<9-h+-(oudHmu`PTau-cRD{FG56bOtzdZVD)IitJMYyGuKeVrU>BMC#fsGD
zOK;1z9D8)!dw0zJow|3|-b*`~evQ}@d;Ce!hEZ3ZTz&aCSiikhYVe8EOl+@_6_x8c
zBn2sNT$xqj{c|_He8R$!wYSR`ef4tj+)G)$IpbE(bMQUU;a(6}YH%kjmno|p7OYGP
zsZl-diqAH+Zd$mv%l4LE?4wbSMr>;67_Hf3ADZ<o>3iMLPM`b=R2EO!VlTb2X|HpN
z;JRw|>&Tb^4kZOY&Zu{53Go?4(A#Z}x2r#TdO%rKv+ZSRin+u?AM~p2+0};?u0hLY
zn8!XJvvy0;Ez$AOcZCl<Q<8T&3pf63=+|%Zq0*YE{_(3nD7&d^Ilaeuj6Ai8*e6L(
z_B$6ZFUa4JDc2k*eY|?}Akz$qTcX<qW7=ox=!>U{{+aG|W>osL(@o85e?QUuYS=5=
zDPQO^c06r&Jd^Carf&Dei07v+wz$WRQ`QYTb3n6C;Js_|DXU`#o6f~_MCaCy3aEb&
zd7=68<q@LM&(3%2wl&+n7w+sge(L%O-?eplT}*hM4GQ0#Iq5N~Ns_)iX1aeSK?1oF
z&?ZQkU`Wpb2oDm307ygcABN@6%Y46x1OZ5?4pMb6sAmC$E5RxSTq$n}2>irEf}u#s
zKp8?U0uU}XtU}<rL;qgf5Ivh!NTpIkX+{9T*@{(AFVB3j@{qBU1PZi6K9wJ45rA;B
zV->V*S7#<TTTLWEC{j>R3WY@gLMnRG;;61;DYf&pGLqfFNEyRuMgU@vkli8P?zZ9F
z5~t7D9b}YDK{EmnUXH9n@^NwHk8%FbgjNyAL^ZUjhN7MY5PlTuW)zRK?;BS{gp}5h
zKsyA`2_QjY5r7zI%kGl%^~WVa|Cx7~HKU?b3Yrms@b1ql6nvU=DLb=fI|+i2nPNy4
zLq*TRIa_%9!^e&8l_)n%M6;-@P>TS>0Bd%ajlXYqTKUlP0}=!y1v8|~C@<^+2-zE~
zQl<C%Teg^X4`E}HN9Bdli~xkY{a{Pb%(Z9XovCA2kRS~0;z4-~wFp3XIz(C&j^?`*
z{Nx^`Ab|!ch$*qcA^_n8SrZo&_d7pLpSXbp|3C^PMJg?VQ@8{&tvmi}_mVB3dPvnH
zN7;p|vTI|0(=B`3ri6{E4(KE_BLE@k$C~N#+oZU3#I=wFfyhh+q$*%=&jJWvAAgIp
zH>Z`(YAq|xVt3I}TAF4AAjAT8mx9S}7j6`Cl@INE>V5x*0gaQtC_m^BQo3c@zm%OK
zk|s4|RPW(vv0AL{jRhU$8AbIPqxGJ8s`mi2Ua41eDA6;K>Ma>MQJJXbp}52Z4wXUI
zGAOl6Yg}G1wS6o}uYq=Jps2T4=t}5X38lSppsS#J74+utVd#1oN-OWq85*HU_af<|
z(C#P{;WBYB?`f3xn|(}z4hBd&RLhnRNsED)hdh7>56OUxhZI1;Ln@%+AuZ7IP%Mb$
zp?DC_L#ZH@hq6Ex4;g@ghw?xk59Nb=9x?(W4;4dD40*T$f(kx1L15zJItc3cxE_Lf
zJ~l&O=CGq{`>>F<Ytx}MeHi*4hB6%E{>3?a8ed$N_9JOW562-~vAB4+4B?6;V_-&b
zFaOcm?%B%T=B9H!bHgxsVy3mS@8MTmlqB5%B@NIEcM^U9RWC!eSPk1ij~Q;ok1?~N
zZ5dj{k4buo?-CUzO<Fby7#!a+T3OR&Fsux!a2+xjeKi<k8n(8#&@Z=}TBsh_OMSF@
zkXp{6B?FgG*i!T|vCZvYYWyK{DyIfG0G|pgAU_%2=62@$IlOIaL{(N?Aqu%DSVVIn
zj}k^ncpWL!OL#pfOq6haHb*!|f@6Y50y+&d3OJj8^)&ID8`qs*e&3-#NqfY4jOe9P
zdQj{j9`d5R<lKL<a9ISlOwzJMSqg{bK5{<}DZ&*IIN~I&3RgvNK{ZAgqjAhgG(W<N
zVFbq(Wzee(`r@oHY~r{E4&eT;fI}-_496ZOs5Zf&y*C0x&|&pZgL6Zo$r5IU8q6km
zQle60$0Uv%4l@cUGaC0zonH67qPv8o)5Ym>T%aU9UOZmTH9?LzM~-WH>~EWtw~j0i
z>(gxGYjUsZ_rEX-Q09dqCoc06p~Q)6wndJMBseNKR>&D<0%IRZ`mbhpMpLbS)wb?&
zf7%r|t|%VtJjb^1A<-Yq49Aa)1_~B!7QFh!iln`wy!2dY$hC4EPJhW1M{U&0cx%)t
zl3<AvpK+D#mpe+E3YMbBc@cThc*L)g#U!n}oV&J9UzNAse)Qgvg`fQCA#ie$uk1Yn
zdy7q=&SY*6`0*bZWb;bt*oHg%+K~Nf=v58<aE)x;7I34e$*As=(dw-AwW5(}zUYlz
zklY0a;_9neylP#uNFDd59@n(`o!|QN_8};YQlZqD3*%tnU}qf0kLDlgmo&3Rj|xW{
zuT9}dLKmuw;7B4NDq%RT-u-O??<sc#chSCOFsKYF`F-TM2KwN9V~Z$c%-sn`=Fr<+
z%{u#Y*O_5UkYe7jCB!`BL3r?xjF9n=f>7{~ics;8meBH0ED_5?@kBfir4p$;ltpCm
zkbyApP#%%TL-|BL4;cv~4;4eA81irhBr5pW1PK!#*FmC=kLw{(&&Os+m^pm-UeLnC
zK~a^=Ow~HZaBc3Qb&0_(K++j88F3t)Wb3j=;7%jy+~K(im`?B;C_&8K6({JFzU8ZL
zJCi2;-~Kz@+Vagecb|UQr*O|bKid90>V9CKR>R&3>@9a-Y>lK#p|BLX;i@6&&!O;h
z=!P2(vn=W__Fl}r_{YJ2O};v{In6(>k6~QKi0PPX^WUdkJP42r?A@Akp$fiP1yxu9
zw-2(!T&@4UrKe2}eskO(uJ3b(g2^)sOs5nU<P<8HA1X$$AKGo@ck_)UL6ktH4hL%d
K7=Uo@=l=jUgtc)1

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/type.hpp.3093CAE9B6EE9081.idx b/.cache/clangd/index/type.hpp.3093CAE9B6EE9081.idx
new file mode 100755
index 0000000000000000000000000000000000000000..43c0f57d6eaf1658027b3d9b32902ebf5a2bdbed
GIT binary patch
literal 1350
zcmWIYbaS&~Wngel@vO*AElFfyU|<jg;^LB`%nv|1m5G6&V$Rx$2eX(J1=?yw_!p`O
zXuS<|nzyL^?&KYlZ@$~Gu4`8A^bPOhc$X~uvP<E`|Fib{jW%6=_f75cql>RNe;%(&
ztM}Y2_w(JiILF`TC!O3Eap2~Rio@~yt}|Pky5CDl@2wNt8F70)_uI<&WWF`2vZ}0B
zYHXX54ZKcP8ELc%G|vv;35+jU5?S%R)iHDRSs_iXg30B9<!o0w4lWW*IT)j_k#Od9
z>4&=J1=&i@nSwvx9(DM8Y-fsa9b4{|TD4otdOx4O@qVJn)bgj>7Cci*`qzDYp67v+
zXICVIt(?dFE!1vTz2mwh2ksdc`J+$Yi!ZLsO;TrMV30ewWL4H}ab0dEVQvO~88#U~
zFu}md3Y27E<Y%;;`yoa@ars>~CP8i%enECYJ}|++0h9)UrURQfANpn%ax;kn^((O{
ziGc}({+7k{RlX~Z9_M8e0xA#_6O#uM44iDh5a$9~Fe~hu&x;J#PF^MfpaNAfRV6S1
zF#)7teNlVXOxutz+)QFX1(Ix%5?}&h#fEQtKB+GAu;*dm1S-%_(9i%A44mxDKxYC?
zxOlBJTwCE-JvWmGP=P#~ya<>;nDF*@%n#q$ZRQ|PGVt@T^YDQQh$BHJFchVx6?Xw$
z%)52Peht=wy^K8k8VnlHKtKc|6Bju6Hq6uBpSo+}HXeR?ad{<}^(t#7nu>jq<Ok}N
zF_1BV1-)*&{^7Y?(p*4cNjpg=n0NKW&J?=s+U&^0!mlf;D+co<kSD1t1>?bDhmi{$
zD=>RqPELCj@=Mu_iHDy@f=3hPJsy5RIYE7xp3u2XNBw>Eq=EJ-sVkYnVlw*jTD$A3
z4uyf7qbX?#b56>{&bBJehLd=m12Y5Y98gk#IR};ofWn{z027AAJWv=E?J(g*Zq9r<
zvmOY8f=Nk68x~AUj>WB7BJc4JD6GYzr431b3=A;$0Sy5q4VWRY00Ihw5(G?mN7{ok
zE?lP`0u9mN)DVUlvfryy>FwEqLqK6MLoqX$dym${aNJ+X91j$h*O0e@#pAIPdYsE0
z#U22KRkc)2VBUa91FZxlF_@K49$4hW9ATc1H{f6~#LNXwcrd-d2m@v~m@9#fQxemJ
zIj$%*C(#%v1&eD2SiHgoV1~kIhN<(J?}ar?*}+g;lw7P2GL{!+HUk3>f(f^gfq{R0
QrMP;>)=Q<#Jbb(i0LQ1mkpKVy

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/type_convert.hpp.5842DA15997DD61D.idx b/.cache/clangd/index/type_convert.hpp.5842DA15997DD61D.idx
new file mode 100755
index 0000000000000000000000000000000000000000..769821a378a182754738589f19e214d14171b189
GIT binary patch
literal 7482
zcmY+J4Oo=NwSafd`OL!diQlmC2Nq!w0SSL13N8qASrLCQf`(|CNTZEW6I0^Pz17;@
zr-Tw>V+aW`{serK5Pni<HmOP1px9D3F{wtVxi=e0ZV)xebsui&hN!gMa%cINohi?g
zZ+!R6oS8W@bI#1k+7-)|EtqXEI38QE_o*#=HYo;!fwKR0@7c9Io)`?lXfW(;=zRH)
zPeA&;h2JGsTq`)a$W!=|{fkpyzVF<)_%~WcMZ^4~hwC2OQ8;!d{?(>;r<`>C`RujG
zp12*Eo;AOA?Oy(lW8Spyu01+l(lKy-{^sAU%a3k(;{5UCr}~OE*iV&|pU>IQ?`gKS
zKNHqmWL<UV<F(I~d^9iSA3b+xKAWBS?YdKWJF|cJtB`FMkCz_ZQ5>GL^|7UG``4B}
z{7m*v@6UWk+UJIz_Exn_P5x@e`j4)z9yiQP8^<3V?s#l8eb9TKDf&SBd-FEr=KkBS
zzDeKv;FDETzB%*uY+w9^mG_*z`APTdGhc50e9COsJB`1y|90xHXVm}c>D3>POyBjN
zzPu-1zh3)I-}#-VJ~(ap*_Hj@rGK%(d!^#-MvwK&4Gjr*%kQ@3rTvFOBmbe;-U+|C
zd*4%=f6FX|soUR=IbOI#g|H9<El(-8O?(VN%%bcMcjKHSOWCLjCMIrEQreW<$%0qW
zwq3p9+>r8j70j$kMPkL`iBCY>_{!m~?g&r~Q$uu}^aQb<ASs4mAVI+fLP`8z?%P!q
zZYWVfWg6<hS_i3<1=o@g=4lGU)%WV1)RW|Tl0)@^S6SG*{KmJ3QX=%3Bqk+hQU8aF
zZU0sDe_uSk&!ieyCli14AXpDVs(&2DiE$ME>FR~O=_%874I^M3fn=%|T+8}rx4ty<
zlLy|?tGIBE3sb3H@G3j%HzzE;w(_T{!EaHs)to%>@pq$4bW`=>zM6;+HvL4O%l@SO
z+0_5x?O*uzot^9K-WT=uQP!wr>i_uLPnl@{%FX(ZVq?sIRe@QQHWO<zNtrCTj+5_o
zJluQw`wjXqJJH&S$y6_Rl{XI`y!8u1O`>Y>FWXk@*0hO_zZ;zBrXxEkKLLOLqOPL}
zW1BFI>IK))@yi-`j?~_-S1C>@wo(6wi#wNjKfPsJwnrC_L(Dj&P`%()-ntkW_Q_ZK
z8uco*IJXw(Q@!9-PJOcEOYh#Q5A-T!$z_YF|HH+n6YXD5+qUmXeTR-=${6NSy$DpP
z$}Ecf@Utq_z&5eI^moyE7w7vYd(XsV&py;;`Ydkz5Be5#g4M~kpk8n-zTCafcGoBi
z^l><n9gC^|6A<rez=oyS)AUtw6RVr8ieB(4x1GnWeb=M+=?h;&5^Km}su#S<Kj(Mm
z?@C*rpm&m)l4+y<4;Sy+^7!uCGYtl)j=q-uw=Zu_QJJwUwHO<)Bjfa1v<Ys+*hWke
z+=R9!v<DWKt*E03S8?X<fS4xSfb)3Lxh?O%^d5Hu>}TBvaDk~~_|lSN|NL52X-ZIv
z7C3`6EhNn>C?_aK3rvl+5~Y=>fo)7L6Xh~decXVwpD6vZyb6>mkmRip+6prQUW1;7
z(9<v@u!Qt;2>o1^dm+>-%PW<jN=4%vXX^atjy-WGz2B(P!x(p1w$)KgI*RGScE>R8
zm@IF=qy|~uh-r<uNYrn^H7!`iZL88&T+=Fe>bq^{=1qC}>|B+a4dzLQYcQxZ$`B<{
zqA_urzH9xmOk<`*4x>Y&Vq>vHWyUgz%8lg`Z8dC_dZ;i~NV5IL{Sws>V-1nGo*3(8
zK0yp8h@`8T7@K9en;6|Pdx+5^^C&Tn%6t$E2SL(T2gW*(`gelSDRU1Pdt@F2<DkqV
zU>uQo9E{^K*P^i&CH+lkY?8SXjh!;P(CCucjYhZ3y=d%}c>s+AGJDbJmH93j?xHl`
zF*J_J@;1fTrUb^vV=X&}{y@u5p*9&~N$_NB<pO?2z%&_8xxhqwAf^Y>IP>_*6zfry
z3zA~PYis&%EChuf!h}P(fbVXFzJb;^FoEyxq=p~g7&<Wc6jA7_81X8`@QU0EplKo=
zbHyqxG#5q(oGM*!ULVa}J&)CFK6;GijSBvr`STm~{dQ1kUPzu<xRlqGkk=J6Khivq
zvuW$$o3}SnH^#Uzh4ae%P5-`a)8GDAQR!_^Zi7V}w6{=s3nl&#l@DbeL}gHLY3-gZ
z@iEulyho+eXHAa}+yUtHjOlhk)>+p0Kzygx469wx_F3EG13{T?&)9C~#4jCeI{H<E
z58p(+ILC{roPV^pq2kfAXRe!7nwX!sj9Z@3^29ZQGUGCnh3PW0Gwp&LcE@tDD2|nm
zQbEOW#mRw?Pm8mQ?Se{Xl|&0FvzIOBi>cDGm1U)Z%HztDd6-dYd3L#7(B@g2qXms&
z+!!YFMO9f;&l|%+!EH)ho0809lgdI<UYk-VxI>BTP+~bhTJlxU$=@WFpvq#Nr3vTo
zo$KfMO_(LP4exD33+FAX?Pn|BZrRL&ZCPAdl9)-Ett?;ALo*+mEvVd4zCz5Td{ucV
zryUEP{;ul#V9LgE7^fdb3qL3-J%-bd$=ryMjW~zP`8HzPz<;aci>}hS5px5(YR9|I
zhaYdaxSR##GsJd=I7FE3CTqLNBZ4mx+a=-<VYiR0>mwTl4-?z4+}=ml`N&4mz7lMe
zAhoZC2dbf5v~P#C?eGY9K`QNnbzQJggz0`**AE+od<Y&Gf^sh3x!Ck(^$!i_rzrF&
z-p3*_H-kbO@IDrK1vg?Ui@btc@O~C~`SK|&?B35JFK2%JX0FbE?dE2z87_>pT|O8Q
zr&33pBT4Lfhs}{MsAy)<Y(d42;uT`o7q2QV6;wm=YRJ;SosV7gmer7A!OxSR=SisG
zdJ<PplDH#QnIB$KPYMJdAwfq-s9-l)=qB01+-|bOO$r2GCb5@EqToJa?jzxXM@i-=
zSt4`ZC|N4F3Sz53YF`cJYLMDHVUZJZ_;RW2ww&t(hu|(ScR{$|LC6|}T$%F+p-}J;
zn1|&0otV*yw!lk<JzsJ<(Jt7978k~F->5RznC8Mog3n>xIZP5ffQtq&N3<Wn+yQh5
z_M*ir>+@ooSJvmlcpoMUeQiono3fgF7r*ye)25Vhe%^c~_x$gV?VII4hoNUklsGqM
zNWvMCCAgdXpqrEkzC=PVktnecmq@}Tk|ns0*!xJS;9(LvEbAX83B$5}AF=yLsnA~u
zp_L%%uY`n3kn~qWK{c!u`dc8h1){_`ZGnUq$P(NRKWK*%!7d1OL6lfn7bLhKOK=z1
zyP#CC2SPoveh(ygWc~eM?}t*M{}zPalJ(z$gj=%yAt)GvwOs!TZsRjgj{UKhT^Z(B
z<|K-E8*7OT5dW3AR{vX@`-FdWt*`9MpJ9mqtdD7`#hhBSbC2U^M=h@4%uTJBCZFY8
zpNV=%jE4l`wMy?m$Q_u)?<*?3fFTz!z()$`+2EMdj|Zu=n*?=}Fs@mpeI%%lgo#D+
zkszNeuZEy%kmT(U)DB_7h+PoW1!3GsD(#1$epx;QK|?~$PYJV+o<$Hd8o-zVaiQW@
z7Uo?2DFt4Q`(F05o<Rr*CNl_3;LjZJpLynh^{3|xxWGhvF{W4a$4$=K>sbR{(Rurf
zFRo>IK4AwFL)y^Hhj1z(xUBBd$^V(Hvc<79lW1YnW|G=W76|Sp5#92^aubW2L<^g{
zNvfMH5Zp&BeX_g`EOihqY+eUxb&w&r9U|K0jqe1D6QaeMINALiG6Z*lrAwBNfMo=t
z#pW4-lo3c5d<Vksz-+<euxK1?+_ow+ZO%C43qF9+2QXIfA<Q_0S=_`bJ%W)(&?@*S
z#va84!LMQDYqGonV;f|7Ct5l&n(rl*ZKae>Oc(6La3{_d?7}P;<_M4L!dw^H1)sy{
za~R8+-@VvY@aKp4MsPhoaB->W<7@s`cCgtjF8EQwQ4%F;i4j7Usbxy!&>RvKYsC_k
zX=M_XYvmGc4c;pCP@z>wvi;hAiE4;eLnN*zTD{CCNbm_F>1rlgvn+QL%`LNsXdanI
zi8d<pK?ptwlD<07>Okt>37S*p9?*JZ9t3Sr<`K|FWF7}?T;^KTYEjbPgj$o#ov3xn
z>_W{Yvl}(H%)O}f$~=JDfXrUhyfWX#;JYZzcMP>LS>C2-ZAxH#{Dr~1kDf^6dr74o
zN=S!d;X(fNWMYxMQ}u)*ugK3V)1W62xxhpRA!bm#yFYlj{kn1M*FR?u(4!c6R9yB|
zcFT=ykhul(Td<J3F@G{Czg6&`SI$WI?~kthgaxNo65dMY2{+bH!uthZta{|@TeUux
z=wx{}^JOwe_^&E3S3!V#zkhi6-95jEVNcov7<K@ogkv~@(~ijRl-F?DYcik1uyYv2
z52i}5V%Sx=y${2DLjGy*GuFRuNS?-Q&`OG0<*U$TQgE595+m#<Mg6k83JR(qApc^@
zn>VAbgoHCiEnsT_yKuEGu(`m_mqn!>uz6(pEwJ4Z@|!oeY<v5`1LN%S<{=gju?k~)
zNScQ%6zn4rJ~@Vr63ZyDiV=^Jv{AB9a5Y$}L6Y}?r3b8Hpgl0Z2NntLhlqYTh75vb
z5UgS%gD`&(76~2#%aAO0qs5I@5d++q>c$Mg7cl$+MhWi4MZK6Umbw?0^rBtxRgAuh
za|I8eWdN<hXAfZN0A>ijj^WoaO0XBRytqVc4KL<-(IMD}(LS8Zd3bly;uk)BrGUMa
z|26xhvLuu8Dozu@=M6Nyfz$c%;7>Fb4E#xYe*E}4u$!5ln&$!&9VIcN^3BYNAx@kn
z27Vnwt_vP{aCQ~!`S7PKEp`~^vb-@Lh{7K6b6K94Gk@Q)f#{ij-Xl-Bu(X&z!_R+K
zX-P=QO!11`9I|;P&m8R9vSX7!M(y79<K1cge`x-X{uj#sVK4-;h3x-}ZqSaXtNQNU
KiJ{Xp!~X*;p7QGe

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/type_convert_const.cpp.46FAD4095BF1793B.idx b/.cache/clangd/index/type_convert_const.cpp.46FAD4095BF1793B.idx
new file mode 100755
index 0000000000000000000000000000000000000000..21ac067ba2ea0b23507b3290add6854925735aaa
GIT binary patch
literal 4342
zcmWIYbaQ(rz`)>~;#rZKT9U}Zz`!5~#Kk2=neEIB4ANW-3>9;>hQ{XKw&MBvUsA#}
z`+6?xrweH(#qMn7Ei?Xdae?YhF(vPPljfNG{yP2B@kwd27gBF`D}U;6SK(i(!ob;Z
zwm3JC!79$L>(9r(l4>6kfB9SO-o5Xqjd%XD$ETH#?^gd-Q@JH^pG|%3lO203ifoof
z+*!Nm@w>xsp1pRDs`*<!O>h65^Os*gO)^q*pH&y7bNT7>*H{1iJDpoq>pkyzga74B
zcAbb!?UP>pOnYx_xRtlm`{IIol}{zNhvwR*MJ>yI*>h#T+`&z5nFpsl+$H#_<;P{^
zt*2ym-OV?+a_*Uz_x9yG?EN1wnwmap+!lU`CCXDdgip$NWv$8F&3gaxmfUU)nyDn!
zfAHcFn<v*oC7P!MGVZxzxwf~pM<<o(f?ejGP0NZuhp)f4zj;Qabjg#(C@vGnk8+#Z
z`fND1rFb#ct!C19JRu`}_cBMM^+`hk9hK`Z3<Bo7cf1`S-f(Jt&JRCzdClVsv^)jv
zcC@j-{=*Ph8sfXOFZaX3SN%QW8h<9FEX-}Q+SPBteVIWjyKHK%P00hfi>EwRT$mVC
z{oMSePR<2QkuKp=Ctbo#0+-L*8gy^6(gf|5tS*8gS+0%l7FXqRZd~2;U0Q2qm8ftN
z^Y*pL^$w!Pf|gsy_d30~e$Tnj#AWYcPyM@XGo~z9BC_-$n}W%?fVoUsmwJP`gZhul
zzVd#ZVesG+Q&a1Yf-P}Y-`m^lmX{@;keQa9yq#^^3m=^=s}l~`JQLcac7AE|v1CQ|
z1s|sr%T$^y%KmrZhPrjrnk?VotCrh^u5aM{=ykO2tLm|D6Q4e<?=_io?98fZOV_ut
z-QHRe&92)N+<JIdkBGBk?@e#@sN0@TQf)dl9{JY3F!E~9bKe(p=7sdvSMv+I`ppH6
za(~WRpqPJR!urfIP2IIxqH5`!8vB*su4c($D%E(YkfbV``#58EK;9v#_m7H;D-XWV
zn%UH=@0PAy_0IFGXUNA~rv8(e`srTgS2rtN_|sTENA#Lll-uWi;kx=~e#&{8ENni}
zt@Tr%WLk93-!l8{gl+2&cU&($lo2;`^PVR2#h))qo}QApbd{z5vh6#j74OR2x72IC
zqQcs`L)M#4b~R)^pYK+FOS-s<)%v)W+27RL>(<rnDol+(P`0CU$0niDd2K5Hq+<=g
zZEtyh;GMOWziqC<<LqlJOG9tGO_iOx<l+3swj!U)SHzoDo1~vqobreNN9(c*hT_WH
zBr`?^2F;!$%q9YJpNKJqM=->4F>tYf2?kKU1p?#y<%jNzO3AV_g)%V2sxhdsfC-qi
z<H3$Z^~Wl{Voaf7SfnEx6b_$AtX{~@6b{slVR+`;_^X@W-%rG$dxG(7E`!&FXQY@S
zA&y~)W#?iC6((RX>w9kN&m#NTICO85wqRM65G*al6c5!9%fZDV24yoGXZSt)VrA}p
z9RB{7|9iXioVq?7(nYCh#d_Qf44f9ppC#jFS1>_B?c*_K`A?M{){sDYzRtb*bosF!
z5vN6YVe?pT;8i4gb?N@q#!sV}7{LJy3wI_Io~EMU>YcpIW{haU&^QDNgTfdV>Tok4
zs$o8dI1VOhvLHqMmqEA%nqHF&9#39O)s<r7iB(Wi(5pRuh@)$lML18anUa|v%&k1J
zHZnFE2%eLS6B5ry$p=YgkW3I<4wh=ryy3UPG-~a1zE-^96yhhO{$ipHD-%yFKR>@H
z!eQF{+M+P0gnh~hu;b3$!NC$M!6P9KQ^XRh$fJnFGv_fEhiPGnwdJuz;yLp;Bk_EB
ze35v;Ji$o3NS;VI4;J)HkeEOOIdZ(jopW3>Z!-5cy1OpEVorX--aU^Qc~CM^qK{|1
zp3>ysbazkcIsFwz^J47L(jKf>KuaDulath49(t7nU5}iRb2g^BnyImAbMS!U8W!X{
z;J8NcKyi)Wf#Mp$1I0Ch2a0P14;0r39w@G1ygUiU<b?Bcn(!vq!V0_dGN&&61uEiK
z<yS+bL~T`VHCWm!ea@bGOx^qj-6Np7cq>nSSG_%6FTmUkESZo~F=|=U`(jgJu;c4~
zPSnWZi4|88M=CvJ<Ycs9=JCWD$r)+E!hk2%LCyh*=OO2T#0!)OgqI#Xv0-vyNOBP}
z5pcOazX?Ag%hwyxJtX^XL`5p8PU8U<OQ1l4r4b%TZ3*SVDl|q;l>9fLj(y2>+pG+_
zJ9?r*)RzFCcT?$ZF|6uF%i${otfn<{Cri-XJ!{He_AEbdl!P}QtUYUYQjTk%BHdlL
z!SPaj)XAVmv|@7W<|$o?n{L3$A6V6cTD`+sLBJ%3Txf37*FLHI=E7OD2-toh+bb?g
za|YeLxWjE~#X5)Q_?t)vd3Vm>46GGLbM!Iy)FT2yVTO1k57yj63&Zo)>Lx}pw|juu
zR#;gWUOe-}>MQHRi|3o4Z6B_G)IX8#p#p0_;%yVc+MIZc=O6rY-W2z&x`H>1{#=>!
zm+c(eBfN_K8LYK0(hpY$hM1^?s0yMmx0bPn7sIgDGg@dDrRF3m0n0*I%7ynX-~zBj
zTwIh~tPfPl7`ytyPTd#Q8>TTwMaM8i!^%|#hA0Fx2Ek0uP07gzaszXNnORub*f}`4
zd3c5R1O@m#13dkGg+1Ifg}Ho1Bt^xfq-|{N9UPsUT@1zb>{P9^6|sp*$jZqpDXXYy
kXzA$c8yTCJTUskv$QYQqi%Y1RdHcmNd4;(8Fa|LI0Hd@XXaE2J

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/unary_element_wise_operation.hpp.DD0C457D046F7875.idx b/.cache/clangd/index/unary_element_wise_operation.hpp.DD0C457D046F7875.idx
new file mode 100755
index 0000000000000000000000000000000000000000..5e11cde6f128ebd1826506d86dc3d3da12f4bd50
GIT binary patch
literal 19700
zcmds8d0bW1_P=|deL(SCB!??pJtE>|ltDxWMN~X0I8T|G6>9nF<rMmgtjqxo6~XCg
z1Vrl7L@`uEB%E+UMMXsn6(ungOVfN3IAp(dId{2tuRr?Zy-$Dq`g}Ue@7;?%uD$l!
zYpp#prhorCgAInhZ^Xn;eRI|XW-u6Dfd6OCihKJvYB1b%Fc{*~^Txa~>`Gv(UuJ%?
zaN#d!qZWEknzE&$)3K@1*?p(x?>ilIB4%p09m@hXwyr)<I`q(Gx8*Ar@9VVxqpJ7M
z>_|yjmVB+U`J!nX`|PCmH(t1w5mdZs!uRV^sw(g0cpmtme%3E{haYdX=EC(Zr@nqU
z$n{aR-KoQ#vmWl9?0TTXx5cGCvRlBvp7egz)$`xim*++-y}MwoNB_*TG5)(2%~A>m
zcn<VGxTo~ig2O-0dRWo(x4`xp+j^|((!E*5%*pEhxm)k4Kc(;P8x>p@dnaQ?#0c*@
z9$PxdVPB67YF3@m`dMQAdsR~d@(UNJ#hvb5e&xnDPqd}i-h00wH^1=x8TaQn84g_g
zW%fSDmx_9iK6H)VoOrA#c1@>i+uvB2ADfjQxXSz5z@rnlb}gG@JX3!4^|1q`?pv>W
zG(B>v<VaIXNud9i$)4FAomaGLdw9%0_QnnmfAyz9;q$(pd-|oC*)OJzT=rVmcfP9h
z>UAUS?w;@`wcWcX3_R0mLip&aNiJ)y&TqTp+~|*6pT9CFwW)Wd!;mA{!#f&(iyo1>
zb<g{U;)36N(Co2)>dR_X#g-90Uirt+qf!4Fp68jLvSHOLrOKDCfAf5JtLFB@W)ll;
zCarHf<9gdw>z{@#ALFv+(3qWjdIvc6b(EtAEEsgwY1fLs4{`a#@<xx5sk^6ouX?xd
zv6pW+x9^%QKfM$Y@!<TD)Pz=^>8X*MZWT<NHp4lk{<pIB-O8uEP@O4Va60gA`sJ=C
zrHa+2l(+j%Y4%&m;cIqJ;wHScW#m6oj@8~=cVXjZ?+8=&qKdH>%~7sPUvC?q@#9pl
z7u}{FPORCjMw)7lZ*CSp?(^VQQ|FKHh<fG}wZ~`fjFOR)kIt|CY}@LXH96VaJ5}b7
zd-=UpQpSQ6%Uimws{E?=%J2_<FgC5XWKDH)eRW-$>Dz;^eqHBVQ9ty1w~u4Y?E`}&
z|Nh#``LW5r*~ffS@yVOz&ONrhU%2{<9qXI4xql#H{lPb`TC!dnas6n*mLa1yUFjRL
z=7lBM-;Q(nG<SHzI~&Hn85Sw;7!_8wpzp)rZfB?8eE8JUF!O_{6I(hN4D`F6%QCBn
zj(4Jix)`)LHO|ubF*Jd_0{=<N_di?x=%p2Uag-Wm(X1aVwm)&JxOjZXbiFuUjkjpl
z4;HulVRD}0)cTksjp=I8?lPyl%w!|T;(Ygufr<S;v9=#!(CVmKM=dsjEN#D<ej-<P
znrBB_dKk1z%>EKnY{byCv6n<6s`<v}Ee5~(E7a57pzUVrZf3C&tjBeb&zP>q6I}Gp
z#;UOv&HBOO4z217o1UKj$d0yvW?I;{Xw&%NO?2q)_>`r+*G&KW9HQnBi>4E-!#&OU
z{J4|Hb9I&sqGkY#PM)(=)Gl57!?al@Dtj8V2aG*n&20o(Jo~r*>9r4Ul-f}<^kfIK
z-vI;G3Ga<>{oFkdTK1c7N9};An%Gy9);7YK{OtNZ*qIWU4E=!xsA6gr%!*F1F~8(D
zHjSQidA}WP4)rXM?H9;SHp1%(^z0jQ<k+r}cGLt+Mby5Cde{hO3ikO|#XnE)+2cq%
z1{$;#veOFLWFy#E@V@X3<)8b0Y)4h7=Mu5ML|WPiuP3zq;u+t_wO`v&8JO-f`}?eg
zjR;H|BAZMe@hHWPwg#rP)P5~hY=kp~wOzE@ZTD(l7=BQrpcu5)YU}4K9u}WCZS`EA
z^>(xa)KN|Ct7!`x;dO-X`85CNIG?Fd2aGCFRug55jo_$yMW;V}IDODfU<o&9r<i&Q
zW?3i5(tFl}7W==S<7%Q!eGS?gnXQrSZ3J1gpFZAO_g&lrN{4o`b+(y9GqbHH2JKnm
zg{!M*J7}bft=LU<dp?$aFZC(Rn)UFkBXtL{nkzfam0Q~gHqbx#=T#jyypUl>+rx0r
zQTuZ+T%GXY4zO%7m#s?Ax1-L$w1(NQVa;rWGsQf5-~M0w2i<Dyi<Ust1Q2tbU}FQd
z^x4G?RqKGIn?Wm}Y60wPogm8)H}k8*@`m5Dqb|@?F|#j*ouCum)R54X(@lO!%kAh3
zz|`KpeT&AAz>vSA_m^GU^waw?P90~_tRHM-XwRQkPCVtb&_r89KT@f4DphR+S%#hQ
zn$)jNh(WI*kEnSt8l51^@EL(pZm{=bVDUF-wQ}29+1*B-v#h?iHh*jr>wrO#W)U?D
zYS9U@j2=5O|KPyg+gs2+P|HQR<3-uqMv!Iv{>NJ;JGR~gdkExYz3j9e_K;4Vvt0M_
z@7?|B8zyQ7O_$49E^BHd$PznWnfB59OSVH_LJitdrY;46)Csb@xqIGV%lQTVz!G54
zo=EBw$zmhVS)P9W^wNMAnwY2yi1i7^PJkNH39|gHLuvehVe>K_X;dWW7phdzRyKkx
zZ<)S$`q`4QGCSH77I3v}zgl*%5x$7;?0Dr%_m2%@pf}xY!ktdlbozW5r+m6MwZ{^_
zWJfwU#Gsv_4ri#7jbKAlH^v`n<I%qw>ga6HN~9JglG#Q$%d(Z;YwpwoOE3t7q^824
zb%HF@==UBS&ux3X8FlnCXgRV&j@-&dkY&2V)xRg~caMNtLTsHqM%81yv(qCp+&2Yg
zM*)MELAy!Rn;;50!8&GqlNZ)v>_1=Cdy*olDbN$0Aj^BhJIaw2H^S{`E10+>$v#PH
zW+Qy!-k&sWX5zxYub>`SfQwYU2-{32SWmpGSF4ks558`q%|U@Jl-WYrXd}q-;nV@q
z^VWC`FwvH9Ok82?3W$_WkY(=8;_vJ>7yk?mfncV}YAUEfogmA+ONZW=Fk-PYu)tQ!
zq-rMgMJLEI-?QQLon@n10ZR{qwuz~mKrnTJEDIv8ol-B4n_@>552l@w?a#;^Yy?@B
z9$0vOTG0JbPz!9eq3Tde;|GhEeeCk-qSsE;Kyhz_RzuYq=!Z^_CE2~jP{)trj=@fZ
zRo*Fg+zC5TC(l{BbZoDAzjGdH=>_Maq%MFJ(h0Jx_B<XR+s82xSYW^#sk#vw(h0Jx
z@vI+t-h&o`IKh@LlO4)HoOJS>#qZJglk&$^S_f><Hp}*#VZb^uG^J#S!2ssX+5ugU
znO^K+2YLb2`)|zYH?X91g7tj<<+Z$BJr>yOvob}UV$rN0EKb`{k^Siq!(}Lj4L^^m
z^I+<Af-LFpywKs5qpu9Hqivxl8<_nD(EU2$5|=Uj=!V(DLw|#MdK$D!qE>=%>jdk`
z^!fR%E1fr{={;FO)Fr^86J*)c;-keK&g29EOOQcp<Jl&l@$<ZRXS!poZ(>U*hMCQ!
zYA(#IPLO5GkV&lY!O>N~5(Wav)KnOcPLO5mfo^>V#a1t|qfKBu|0eeTCM|4)k0-lL
zvd6&jDThte1e)bAbv{f(Z3OGdo;bhQr`?I(aJItsxWk<8u+}z$EIIbw=PY8TAAtpw
z@)72A1Xy%}EZaQ_a@$vAeBGM5fhe4nUCzoKZR9yi&+Dz`ynU@)pSWl>+M-!MSe#qD
z$+h|0hNgP)2BL0&dDRKB>}$xLr~dlXx;0?pwo!E(Oq@=T<$#}^`r(r{%R!W2G(R$@
zA3>CK@|-1jdhUC!@AYX$onU)xmmRi)F`<*^EUzCgyqebasos}BHPE72KUjRQ!_tO<
z-kzzTuc06Nsl$HI*E&I#qMX9eBmY#gp&xL(Tx3ocp&vRymZPf+4*lG_Wp}+1AJxaA
zSwC2O{OsnF%IfGcD2AmwLDds5Se+otiR?kCTbFIIP}<jG)76TDVzMp2p`_=~f1R{%
z_L3d72MJk7?HAIPHo_&OWX;Tdex1g`{lgWQ=1TT+B_|u<OsB5Y{Q8YozV$o>ozzL~
zWNG|hKh9)d-7JUZCRsPRq-~S!x4|aY39_6USo+%Wy)CRa7^opajj%L+u=xC#UxwfK
z_ud&$>}}8%Q*|*6Rwu}E;duG_XK(mfFF2qi>X=#wTS6zuQr@(F*gD6KGl2yT$sDOg
z4jhs?L6(c2Il~kC`ERkK?O?bMsQm*97nDZAhx_g7vuh8(^T{|n+6I^snSCO2v=Pp9
z`N&%n2DDo035^BVCci;a8>Ht8cqOp!jY0KA*&t`IB`(Mg7eLN*g1!09m|{HSTN-9Z
zn?XaHB>PQLa~rWX1YR7_j@QSY*i!C1^B6P)Q#VJQV`=<g9XDS7Y2-HVR!MsCXmzwj
zvwofzf8)}n^Ut0OOtd|mp=H#$jJn$hvfOrib$<Hl)>|5<t#K$?mee9kG9$yC9aZDs
ze#>%BZz4nuv1ry07T=!}x8dfrl$MQuuT@Dcs$kl6f-Db~cPPyJ&bK45_!zW$qSnLa
z*2#00yh(oxpU`TD-pFKivPH9gu(+;bNz)aNC%s@tjWAVpl6@T*3_7t+)xgH7depw^
zx2mTvT5p7~wNJ`zPl6?@6RgJ&_vYl8lfuC_u+a5-m!GfJ5Jhuo@8W?v8(jv3Ho6MX
zLYMsC{>JT#ei6VuTbb>Qt^q0T!ye&>Z}w`XXhXe*2H@EZsLRkW0d*hRM?iyzj^ZeR
z1SXJBymTmV(-@dQMsQq4eadK;8@6Qx-vG5-qMgw+QotZ(RkRt}s)|-iJ!`2ynhA>5
zK%*LH40<9IZKV{oQi?$nP0>=AC53fHqgv6@S+8`~9}gf!+rlhcSZB0t6s?p+m9iMT
z6)IXai>hWZco&EESX2Xx!QHB8OXaAgatvNHp?*0kU5-JE3tccU-BI9r!Rf~Ng`cf1
z7yP+N;mA=d*3zr63MDk6glecH9kTDAj`;l9%_fSrTW+#jc0wD-ii38`-8s&eo8*h-
z`EpRcSY9Bz6v%Fz{~Oum8xa@EE`=gKAiEq8@PeLcPrn%RKJ+kAG9^mpWSgaB%-ejx
zOP;Y$Q?%BswW4bbgF(^Sur>m6V{QWK#5xJchxrI7kOc}TgoOwwf<*`@ibV-1nnep}
zC>tuE(QLGU#<6h%ie<3^n#?8(XbPJmpg0yMpm-K9pgC-gfD#BxAOc=O*b)(E5SAh0
z4TNnFaTZ}&BF-T!N5pxA<%zhEutE_R5mqGPV#10=e44P+A}%AWOvIIhRf_l~VK+rw
zPguQ(7gM&F3gb<sELFtml%<O}ld?<^Z=`Iah__L;O~kpB<%+m~vH}qwrtGkYk5P6^
z#3v{_A>uO1%0zsTvWp_Fp{z#4b(GbKc!9(gNWyqiB$guLREedEI7?z#BF>Rmj)+Sn
zRwCjmiB*ZXL1GOeej>3aBA&<CJSOyKDPv1ToXS|Lh&M5|NyNJu+b!Z#jGYp36=PK*
zu4AlD#ECLXl!f+FWtPfubLR_Dfo0d;VN6@eEGwA@DrKf+GM`N5i%OYk7g*;D%ny|?
z(`uPdE%QYs+&uZvXQO8i^8_XIl$xGW7u5bv8LQV8?l)5?m`9a))CJFF^kqwmR>?dn
zS!cX7E7~pQcZ+q!svT?|Zgg8af4if})35S_?3wlxb^M7o$Aib(=XRM)<$XTL%m-Op
zbY)w|5P5<162}l$%fhiCrlrVUDY7>n98CLC4*F8=iU$V{2~G&zua1@-_KZ5%+~{Z}
zE4E8NGGI@j2Hm5Mo@sQYLxI(Wj)xDr(1}B%>l~{^$2p}ocRBJ??KDwOljncmvDdDG
z2VrL&{_?3pzdiI&_tx_s7813P_t4$uy@wI%u6uM=<IKRE>YjQWykhgsAp`pDp$6S)
zkBpsrpIJ3^s`t1*h$OBzOt$X2$K^rSJ+!I&?XfDw=HEdEf77XV`+t+$hcc~79#$m}
zM^7ZvzLQ6PC%=d|Kv^~9%sX43fsn6cT~@MS++PaFYDgvvM_E-sHp4Ek9w?g_Ab_su
z(E;8S{Ys4tf#YdAM~AR4VU#7*ql9)v`=rawFTQRYdvv<Yv`Xq&Nt>hnX5H%TS29=Z
zignw#XR=P%6{Zz2OA+%xVYP0XVHenN>>7q8U`z{T%R<=~9b$@>EPE%*0l0r0Y{3eg
zH4Ik3O6idb98>Vw=<jR)>1T!-$1CHVhuA#OiuM)l@D=ThZXrcWlgw$7$JC9w*AYbp
z=131}V13Ge$Wo{+wgHTKV1qO24s2w=_y!@3aShylVU~%83C&;s2Q2qW)@>yV5vQ(u
zChLJy7Z$ZPD)+Hl^Z)srV`RlD<c(n!90ATX$g&<v!O0H4IN;aH86P-+T&$wXDr(02
zZp6)H{U*#kF&7GcW$wQ+544+ceW5Bn)`IZ_jI}`HdY}be(7orkw5$DPy?0|(k-xG&
zSQW-rU>>b872Nz7Q-O7U!1_O6gK+o5p~1YLGJpJg2V0~CSGeD<cUG-U>6Hf|sWREQ
zOm;^%AZ#1i`GV|@7vFw|FCUNZ;@27&w-RM5aUO1qB7mcvC?`0^*bRt;9<jj*g%%nw
ztjK`z8|&a3gEz>C*Cm)PJyL_C#7GSk=<ykpE}TzjQ;*7ERTu?82CNC%v<7I94fh39
zY>mhu7e-`&`}fj}aTrj-dh`YVf@)<%E#pcUqcKpi9*RMqG0q;;phsSK4aGz)<~59%
z^+S<goq7f&;R5S)fq9}Zp(2Da>+u<^8g7kpbgeuL+kR=$-&V9aGdc%WZU<3z5NAB(
z;QB|DQjW*I*Sf0rgcVc&AiraueAeY=;ez1*+Ljp~k=Am(Ts!T5zKby!1@gZ}R@cbx
zxU=9iki9eI0JJn5Y#}P#Xya`mDp+DYK81(h>+$K+WB2X7;jC!UUeN*DZT?O`U82JT
z)IGY7fCfj8;%Kzj=m2!+DB9?-(Y-ht<1(fLN8<v<g`f;7+PLU(!#IldiVZ*`M$ux!
zV|#Nn(Pd%>j^e%I1Mse`Xz}6ky*ZlgGP?suw~5<r;)TH<1tI}n8)*PP4K~s)8)=w`
zyKkg@L_ByS9mR1L^~<8c9H&ZNsZs!*CJF?ux~EEgIL?qfGbDe+vBRsHtlR2690asY
zNSmJMyiqiBd$R=*9t^OP^zb11^(K6JcGI*T_douF)V{G|@uJz@bKx5FC#J-kL;CES
zw`0@9_KG&eYl=TcCjbRb2@z1(lt=-^Oc};eTwq)XI-3+NE+%dmNAX?aL;2Ll4~ieb
zQ3CNwApWTEtp1|#1k#J+MZ{$hapiak@mfOsF>0Yes3LR;>B(^taY-Vs92Zcp0_u-?
zQvtOZSU^KW99BRhMI2K=hjCm;)k5mdaV_m#OZ(%sK+zsi%OmQG<~tmSl4X_T%l%Qw
ztbZ~ajB980M|q|*KRnbGEsOQfVuQIq>MXOIWxlB56zvAH+~8LsXpdPQF<(5)6fH@%
zB+0&bR4ZDBY{`&)akFFO7-U8d8>8RxO<TMecD)`Y#`6I^Z=grYXrD3~!(DZiv_mEJ
zLQ~Jd7CXkn4D|_UV?A(;M~hV<F-W0628<uW0j|f5@eISbF*JRc4meDQV7SL>e#Vs0
zVQ79j*rLg3r%qWl(NnuV@O@WB>#THk#z=&s1t~$!2W-I)MFU@zGdk!L?Jeak=La??
zt)k6RW;t_#T||^c#F-0S5>b+fGZ(6BM7c(sF}kK`>!{N@+74r?idIOS3TZnoG=(&%
zP!zIC>QqVF;Y`9QBy}#8!Z0YVXy+v5oaD^$C8_fzvHXgpToKD_B&9}jM#r_H-IJ7i
zB3{AFEBO6e(Y|2jFPH}x!h@{cLFR=<xdKLd%Od8=@p)!G&pfzLmN9dgh-;Wy!@TkC
zqi8podQ-%=n0iaZPngRSX5m8ol({?=akA`^EL*rxXUZ;_vIT{D%Av2k(=zH_`BRe_
zClE{wJ;eY66jUphT9r%f5ThQ3ZK7uwU_2i)3?v6zUIE626o^3k?39DKWA&u$aZ(OM
zm;JOi=g*%1k1Y#9CvRh>ZOn{291{{i`t?Kv6k<G!pbvWb0bV{)mq3A@bby_UJJ~b-
z_3Anm+W5U+WM-d4|2ZR${{+X$uTuM8B{x(9pk7&I0_%hEWvmIbfS$8}dlp3#+SC&i
zaF^qK4E{pTPQW|Chht;115T}7`G?M+>jnNo&r-k}F)pWd(i>A0P(k26f&x7w0Tl$!
zAr$C&2p9xM!w=fflMrwm2^&6{J!JFIVh3>iP}42y$`$u5>UWEVh`8S^I+WvDYOba3
zBJNsCBREcDrZnct_fQ)1OJgA-?w7`fa-7c0>C9ckLFxQ<_v!t9?ZQiw%mWoI!Yjfb
z<BNa-BSHie77;0+n22E<MR`T}bM+P#8r72{%|&zND7H&%C^wp72gQ!yXmY^hZd@%+
z9x!<bM{!<p{#;4Mg~dg3G|Od{D@QrRD~I@VGdqWb<d6uCcMz8y#FgVb;-5!?@lv6{
zouXeJ8N~5v;&ocA?=%TLO?vXbFC{LeVtr*Kpp10me_uuhl#wAE-yuEkkiNL(6zv)5
z`Hb{M_o1RKr#+X`zId$#6PkKuQhyXoD-O)0AtDaTq>&<y$)v+LUPsk+)Scs7wEHdE
zn+sC|wKRxU#7fDsQnVsUB(+3xM|}q8p=2qMd`0Y6B6SsUWQnA4d`@yXC%JN5F1eOV
z77=@vOMW5_D3`j4xOchKpW`c%%N4P{DoL%9-1z}iC3UWnf;g^`Txuj&j&Dn@w<U{+
z{clUb9N&{%?uq>KSm${x2-Oz2ds)~#7Aa!*X&}e**}(a1IL8UhDFODYEv?3?20asC
zzX~`sf%W7#nRzF(0B*IVu`y|EJm0)o%sYz(pgsbtl6e=f0DfTZW8V8jT*@q^qE%PT
zEY+e_*T5_dqE)w4wk#E`x^&r+&aFB)Q)K^baxjm&ZI`{a%l-#!sUQkyLH{#i`32eQ
zg6xl;3`MJu11jWhc!5^5t8(B~IRx>i4M%Q|9#NkFT62!qoB)j21L`s-OhDb|^byeD
zIion*CcABuy?6(6<W4!VFCM4X4!WHY%gf|WWwI}Ky;R5@D`ao%5V-N>&Xsak#7ixG
zhNoP9Bf0U$BRYpL4h<)SHPsAV3b@hWXFazJeO}9IcfLES&yOk1741K5y8pD%{(sp}
zXt#j1uBXy*!{($MeNrBVh71OLVe0gRI?S+HalZEcyuuH^{!=RqFU3%Uo^OZ8`RWgP
znG7D^o`Y3vGq}x2yyvcIKgoDx@w>O+=ikSiQ98gF)3fsM*jK>+<8x6CM)9!ve>`gC
zK=l8tv1RSSVD(%*G!LwDXw6wfImOJqMkiwE1J;u}5uY&cCoBL>RtH-?A36~~oAvFn
z?kj#+(I}9j;}b8)FnDW%9||OD8u(dH>%)5)S~OtE=*fK8)$|PC>W>cnBlN%aK;f+g
zn$|M`(RKdCk40y5+nt>Ghk3_)5B!Cm7>J`-w>$flvjufy!OqMlN<MMs2W&ZUDktp_
zub;Ct<&<M+zdw}~lmTeUnrDbk@=dlBL-<)wEJP97ls7o-RdAd9Vcs|2$t#z8w0j8(
zida|?>w&r%bvRU|Cm^B|cFVJ-+28h>f15CEm)vTXY{o5unTfC@dRihb30_a3;P-Kp
z?C!q3Jf-{;Xx#jQS;awNY9dHiC3C4{?a?&0UYPuDvLL)bIoJ{w(NfNy`zcK%$wQ&O
ziONK0G}gBD|M9{P;r(a9^lv1}M&gV)hj4!;%3+Q%;Sn0nWv;ob6Z-rV?I3G)kh$RG
zIM~u4(FK@us$GjSRUbgog68sq%kzO?b|h>>JvEYV#B;LqIoXZlYS~;ZyW_@0&mc^`
zHCGbl2>T89G(AHS3#e8^)FSQ*-sRc!#pS2h-F5;$2l2`voiP?@#r_#2NW{SzBuvD8
zGspmrH;{l0q${?pK*RkukQk1$h*uWrj81Cs(~xdiB%I@|#APc{InE)0IV6}{q&Z|j
z4jII89`VW}opD20E!>bi(w*ac;*w8Pjvo`Z$Hc-d(j@AZL@j8M=KVN3ug#JNTmJIk
z>z>KqozfgmUXM|<=FOV7M#TwPK*rW>IBH{TBOo`Un}9kQI|<0g=p&#&W1xUSj3EMw
zFh&R{$`~b}Xk)a1h8l+oXtZ&(fW{fe2`JVWE1=27$pV^UoFbq&W1N8Ejqw7SW1J(P
z1Y%4e0$xIlOGKPOj2R-{K#UtioJEXTBF-Vk91-UcW1ff$iLp?`MZ{Pn;$mVf7V&9f
zJT2lfVk{GJB{5ct_$D#l6mdN<){A&CH7=&YcvGn{RmADkm@eW>YRnYzMrzzB;%(Hp
zO~kp>m@DD}YAg`(VQM@q;$zf!OvERs@q~!WsIg4M7pd{0h-;{^M#Od0SSR8Il5v3~
zj5kFxrieIIGNy_+TVmOgBi|c^5-SvOxx~svd|P6-Mcg168^rpaNX923p2v*yn9!f4
z%(zs<smz!v;!VuBNyNLEakq$1G2<x_S21Iii0hcKPQ;0_F;N!UPnC_SBHk&pow6|g
zlQKIg;)^o7DB@a~)pA_0AY?$h;%^tg5p<6z_eeXuv=>IWz3_eFFUz2yT5eJ;x4=VY
z@A2}NxApjU&0nq?F`pk!H$96VtrX1SZ#?)NDwz{M_-mM7jhK$4=l0{lj~6=TV9V&|
zR{Boav{T5_1e>MFF1aPzEDEFvH$N$x(TYbWICRaL`H$x-X8zklfQtI1(ykcqwYsz|
z>!>gHSZt>4Hq(x1%PSfLQ*&qw95S3|RLP?)&``6wz%7N;7hT{Ew)X@$Blw;GX5<9x
zc!GIhJVMdVGoSM;5Pi)M+HHJnfZ>Tlo^!f{?ivW0gMX4Pab(WrjZJTqxq^6nDMx-O
z_eFOjMh9RjtnU+WDvHmaD=GQm;lcmfzQ!O0)UQ8pKnL&1hP@XWdfn*^J^Gc2msWh(
z0CuGQssRm5ylO##{*nO;@Bss~VST-TV?gNuyG(zsfc#d6f%O#wS3(tXlM2}lw@hh-
z$44WFwOjL-ZyEGn!a%H#C~zQX=D^STs|lPUynjJ~{#F7nVE9(T4${^dA4;HWpK66t
zt3vMTXAo?0ER{^?%VOFk$?=kiZ%B?eBoq3gn08lkyvs4><-wZh?<{bRF&Ya6`V$K*
zsCbw@@~zL7eEMHY7CzB{(1HFm122u}*8w%5Kgi(r_eI%sk*~9&U6z|%mRsVwTQADZ
zYh^QDl&^UujXE{j`|Ur7Xw}n}9)o{Q>;{iYb7bcn*`0fGw#&}jWp~t6R!>cbO4*mU
za8>Sbm1A5Mm?iyv2+q>=u+qNu+0NuI7eVw{fk;^&qo4?4Y6ARheSh-&XWPRC=>P9?
zdA|QHY|8i}5~Mi)Nvyr5=FPWuj9L5}++hldS0VA^calO9P)LGB98ySnh`4_t8N_iB
z2`D1L{LWHD`WKNw92XO>V&cc|D8(eCnDpTI1aUb*+H-uG1fC|{_?_i6i9St+a9l>b
z%7`Doqm+@*G7`aYIZ?|=M~)wph{vRlc!!BdqJ404<2xNnZ0~cP%Z8RMwaVsO_J_~h
z9=gn1{urb(Pcr98DsFas$OCI+eZYfjboa^lh9PN-Li;M(ZDP7jR5TeB?G7>B`8^K0
zL&ACaV`6$tRNmvqq|0Lx#_=;^dM1`XBSFtdI4@74rX;H32q8L8yCl&tj+ax@a<P0l
z4O&jad3gmjRZtb9Z{Vk;K@~Kd<4e?hiMnxIPfhhy<=F@IG^n12bKF494b)A<p$#;W
z<44r|NGx9{nHEYa`urdXQwmxrg>$?@GOdtQj7Tcl3MpuX6wdKV$-Gi>6LIKDDU#z=
zl6jR_zKofcF_rt-m$9H_ES%$HW=>{qBJPpQqBu@r<`l7fGc#>wD)-B8W<i@-ILBGc
zoW<Nk+#`!calD0@w}|C?m}w7F(LJJQdsxsO7S3@2GZ!#7j`uRtUXgz<3)(C4?_=hD
zBL5j?I>S^ncEBCUg3hpTj!T)jl(~sGw3J11e3qHdisjdt={i%<kECeVS<rQ!gl>J^
zWUgjzyu6x)R<lTsZ!q%>vAmv{>Y0kCuA<empn4Y0aRW0qFgFp0Hb5S{?QNqqTgCiH
zEMFp<mdGj(>@SgnmdN28FO|(pWj7IrE|nuWPLj<@VtJZuN|RMI?I4L)4oZ{5IZl_&
z>9U)ML(}C*jx%I)1~31~*lLky$$`7zP;+FC?YXsdSdtG*u87ez0C$!CED#kJ)ox3z
zZga(jdI9v4{x}d_i4W2X4g_p>>-3i&b{NZr9$8-_qQ-o9$msE|`QGwBocncMTfby-
zI%ahMmQYg(ZHv||njUZt>aP~@9RFoe%$`msch>#S6CCj>16|jjK%%w#>y@5&N9Vq8
zgg*X04dPkgwq9`oFL}ZxV~jG!8BcWZ<<SmVV(NE>WX|BJ-_JH5J(Dz1vHFy6Gtb-1
z2e${jwVX1+Dg-m*Ce7?&{fCz1Gxg&&PhYNQ{rV4xHuSUprvKL;|6wo;;Qx%~|NK9t
CFy<%#

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/welford_helper.hpp.DC75228A25D5F78E.idx b/.cache/clangd/index/welford_helper.hpp.DC75228A25D5F78E.idx
new file mode 100755
index 0000000000000000000000000000000000000000..7c9d29fa4e245b5aa5f97b11ae376bc43f661065
GIT binary patch
literal 1736
zcmY*Z4@?_%6h4kyDR-||wr<5DQ0QnG{RebNN3pn3LSW$9wIe!6NLYcmq+q0DmJtnE
zh$5&AXN(Cm7z_d2z{D^_v#8r#2opD8PNhQ1GLspG#{Y4P27kTe^@mORUA}wwd+)pN
zy?5Tq^77AR4C89747D_!Xy6!z*-rk!6K%)!aSS8IGfe2*$l<E0`23_B?q!#fzxM8|
z=hWDO&e&rKr7ksw6@$&C=@S?Kj;y}CXXef6P-0djXb2SSTjB2(eCjf$Hyo?15&X}!
zmi!WY8pP*6`*9fd{JD~H`?9~#F`TNNE%O&%{iD}mx-`16B)s%`{>Vb@M)eo=DHH#Y
zT_1aD?Txj2fwM;oAKsdC{#>+JRreL!6j|vSKiJc4^)kOb(rEjkdtqkcLFu`!n~}LE
zE}U`|)+-WUjW_vEA8`wTnxWId`XyP1X5_8Kz{jI$vNMOmFBA-z!lB-_qk<7ci~Z2@
zg@N<En=P~7mi#`M5ourc3{Tt{T<ZL$`J-zYQ@1Y<*NyIY`^&+hXG0Z{*0cZmEYC#x
z^Sas>cTP?&C%ZL<aVFT&ax6sLO7Pu#?_%fsCq-7|8SsJ6A^{^Oj*_3U|HIBJ{?=kO
zD<feuXU>%X!)0sQP3vLyH5t22&VVt&Xp?|RWJzi!4&v&^nW^5o`=cVOB<DTg(Mf>k
zo4*@gfy7s2No-Og1FB?|Rssz7JT`x8y|(nK$R?1}Zg6WP!0<cg)29QwYx_i2LBd|}
z>LtMNc>7-sUypS-L{=cl4PekpfZ@r1GU9X(x)>Q7Pqgz1e7*!!`=y?_`Ju7!eKpIG
z^T2^s0z7}?`rRYp>0iGmuSUXp(CZ|?@cPZw`R!*us3osP#MzVWS_v@R*3=waC6kK&
z%K-Mu_vWDA0Q<A|S6~i+a#MLVWfkU%Cn>8kSH&1Fbr0Ku+2UBKYj_*4EH8DdGV5gD
zz$dO3;eQWd=e7alX6HIF6@blZE2FH)QRJaginEKI$atyKSPMH%q(s|{<{#gEWoL#$
z9AyC%xr$=69Kf1qeH^Rc7;r=ipPP)giIfUk4P}m0M-D9>ofa(KcQC&Dr~b`-JAh<>
zLRv+oRH;!JfJ&#zp-clB6=f-qvW4k6y@34(FmMI|hmDw+XSP#OmORTAW>?zPC=tM}
zv&RGz=CB{2F{j+Aq|5~_6=iPDEubd=9?m0RvjDuDSHQgh@Nqr?Ck*eBhUUxjGpWP^
zMS+2`BB3aYmU8NyCCGSbyU<g7VUjLMp7bd7beFIzo%*IRYMhj%7*k@5m$nfH0k32x
zqpL_~GP-vdoM~R0FXOM@A=9GbRHF2SB0#+3>@kK)$fRvOi|rzdCh}qiJ_Vr8R>w2}
zq?ytVP`z5Swv@6=b7qY3(z2s4w20J<ZaR9zv^519qAG)J{-8jvEUX2?y6TNg8^inu
D0Wv?$

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/wmma_gemm.hpp.260BC1EC1DC7109A.idx b/.cache/clangd/index/wmma_gemm.hpp.260BC1EC1DC7109A.idx
new file mode 100755
index 0000000000000000000000000000000000000000..dc97c586c708b18f6f8b7f6d6e8acb40a17da600
GIT binary patch
literal 15720
zcmcgzcUToy)4%7ORoWsYE>~d}6hROv0<K;Z8$#3=<BN@`7!|CsU`t|&T~Sa`EZDKd
zUJw-(d$5Z_P_P%G#@I=0U_|+5@7`fIU;cRic%LVEo_ELL%spk!oS8YlId7+SpMTz`
znZeMyU%Rox`j6?$84Ly|_}}O;qXstz=2CToVeG0)-Fl9`X{!Ei$m{xVerfZ)G&Eu_
z-?!?yuM!8Zm2w)*ckOHUOP^f#_SF^~s+j2gzV3<I3%oC_pV9W-`;@%Yyr|$n!&tG}
z;bNy|eH+wvui7hNW+R@zw&88e2=??+)?cZoUd}&0to@gl^F1;q+YS$T*}HCWahca3
zhd<v0+TM4JeY)apOvCSXeMhf-Ro7}(-Y~DF6E^I!Up>M5YPUF_c@O;t*X_OV*~6gI
zdBLl1h5vjn=<KMD$#ef|);#2G=Z5l2PMmmaZM8i;v-}bR63+~qmh5&l;p4nd#~ezR
zor~?a@#@ecM+>X<e!H$y)8?0Ues;P1ai^vM-}Ua`_eZCvlYUHI{r;z(!C4JExNUj8
zctd7r!(COXo$4G{F>BHA(}h<Xy0v_t-D1U!Uk+X?ZW+GwMcJC{BUKV^8-t$J$hUc5
zqI+x6xmHf=Mo$ra-ld++wDY_eH^29fD-5mD{XOT#HaIBXTr>G<&-usZS6%$4^_G&}
z@*nRv7kL(*9&)ZI{!`2JKQq3%@MxuD=ZEnpJgqKNiCli>Sifyv-`5*iGsbylV_UbA
zC8mYn_?B&((x%C6|Kf~$se8-Y{xIXo>ZaSL-g&ld>h0vb;aT0{Y&Y*6|JkDoerwmA
z5z(g#hGm@^^R#c<0SgX{u68h+?B9~XJ#M?VVC(Lz{!b?c+_;+1bn)<~y#k!>I45P~
z58UcG`fb3RpNdX|_s$)Y+c_)p^9i%JEp=U1b*fcer}h4Pjh6#kE)9O(@o0yRBN}{n
z?_Neqfk#{45ApY0Cq4|zoMn2sY2K5bk&kBF^-Kwg3oPlHThgE;>0O)SmpYBh@7b+y
z&cc=-P9C{4zf;nP`fU<>oqZqdP^a7D+wo0)x2r3ZIGtt(R(-59)VRXgCE)9-rly-B
zM-ATgZntNCzWeu`-+t)h{d0IgVeB^hz<I6i)cE?$iAGiJ<45%83!F<6f=3tc{=G`x
zsAgwse=b(r80eqE+Vs4fy71*1Czn}24tp@M=nK1z`>u}_JGXX!RPDp0;K9$H{4_9n
zd3gWJi`%X8&PYEnp<d6z9V?pkacFwK%IzLU&W}Iq800%9YJID^@m?|O8rTkg-f?6<
zQ*O8R6Mq|6nAY8)-qq|n!hcq#h6xG#_O`rV`$iAP3qxA&KQMmpqQTpCc{LkevaHE(
z{p*b{+P->Q)pUB__CcpolSj(<k$106-|)@)d%x$*8atve_pL|xGAD-vUtWpjnm01?
z*|Sr9Liui{X|p{OPFjbqK4I*-eW?G)$jpJxPse7?=`e8cs4u%a{#@&n(|xbCkHY6(
z5|Sf2OML?Up5{l@NdLLh#iTZ6XCjZ;{dRZY-EIrYr{v6hywud>_1)(8?XrGzPM*-|
zuG^%rjW-*<TQ@Yi;~tyMyJxSR*8gnlAe&7~rj2c~blB>xLQKNcxW&`j?i;she2uRX
zirQXUY#(~(Xl(ACy~}?({!8ZE3xRQ6V!p2uySAzD%a+KD&@M+kqM~Y6&|3amvPO>^
z7T3_hV4z8CaLJ31Xn{tB806VRnoa6jh`}1X4gAM-NglR+iqXcAwrXgQlWFy2YP1j(
zv5Su)Yc@)r0<9Npkhc(N3-Pm%N|6u8>jZJ9Y6-L%43tEqBqCb~itrgP?{AEMy$gDQ
zf!-169rU6S6cNssxwJU_s#u`mp$2(?G$7Rc!RDGrPEK1>Jo&6Mb@4aI7n$!x7G@!p
zA}{af-)kHj=uB&XNCk7NU_lmwBI3MP3uaFXO&93rAd=3xrgQEVf+DrMe&H?gN%x)U
zXCShH<2G>N7E&qF|Em=ladu|}8ri@g=MpKG_*)2yc!vCv@18xVo(uI2G|2hXGoJ=p
z2#R<$@cVe=c57FGhJj``5@{pxu@DsTb#%6SJvwWTgQ0zZL3ZYxYnngUT>tdsE5}-g
zZg!?lFx3=JO5qw<2#N%xjW9jBm2;~q{i?o=+|jC|qxpl)L4W0Bl^kfZRG<>{@RYhf
zrJ{wPNQ21>58ByeoB$hv#oi^=?-HYhph$3cr^d9r15pOCm7f!<=fu%MP^6*5n{WS;
zuiFaL4-BA`NTr}Mm7qwA_H7JjcC0(*OkINv@(JpCg4VVW6loc`Vu8){^*aUX0n>d>
zU7y2rRe~a|zG*WqZ1N!|*c;I5W+H8dfmBi{av=0$iQ(3t0(FC49#hxH(2Gh?B;wdr
zQ&7$*u-7Os#y(P?y5<iyw|%tg&h-aXe{-ge&_fdEmBh8S5ES_$a>S17BNy!xXiJ#K
zHX?0<L!lBB`KsOM$qCk<Ew!W7pqD?W?H|<HLQv#CsSlb~Z4hT8(AIFYrV?o?tVSg$
z(zWh}-i=!n9T#X*80aS={RC#D5)_HCuFYH<{@^aq=CGQ0DPA^zu(|J#izn84d0`Ip
z06Tn+Navsjm7qwR<Kl`P(Y6;s1h(%LkzRp_N>HTV4;96G@-7^6FnkHN>clzKG=H$U
ze~3?&$Ko9;fkwcAxJX?uQXdOJkpce^PB*(GoQ7$_N-~I)0V-2TrN|U6?)Li$Hj1TL
z$?3!@9V|^HC^D$Q?B~lDeR<xQI)hfzIFB^0p@pExkS~1f$=*lD1lj`hF-RKZXZ}<+
z_pwhb{M`ieXaV}TN~EjctW<&`L$B^>I<#+4(<+9#^$oHeZzq^P1{<?mB{CT{E2m@O
zgRT$Ya6=#GiF6(mrV{KUfxr1OdqAIB&QySjeqe4NSg?hl$jC19N1fSsd4xclfZG2g
z(x1?aN-9N`d=oj~-Pp!(JV9w$q<R({PnA@Ptl1O2aYEQe7wQcIy`rA4sGo(P$e6yq
z6T=pN=LG{nFB^!o0R~bDihQ?VlF@SuZR%iX3;XN9In*?NDw~&0ePG+9+zI9Zmb!pc
zUjXw^35tB*D|T!23wj*}0g*}4Bp>q!n}0|<=IU@f;4g>~&^(<;(?RVjL6LEjt}dx#
z8u~dLOVDZokqRKBsRTvF=O1prxa81U2qU1fPel4eLM;SECS*M?@osl(BlH3yABgk;
zdQl0A{21ufYnpGTSO{rgA#qZiZ2n;L#3AxGHp}e)P|VOEx3g;JX#N;%s!(Du!1hm?
zaqMZwk1y84?_r{siF6rEStZ!VRQve{AI^<`4-NuE@`#iNA}T?V>2@z4FHTxe7jz1<
z7*c(R(fq;Y88<&BuJk&v1I`O5ESX5j(1S`)Byqs-gmsf1c*7vD)G{KKfrv^fMJBsW
zsJG54A3QVkGJ{Anz{RKpMP~0G9(${O|2_hB2LpOWU7vx&R0)bCb$Qh7N~2lv;2>bC
z-K1_l<_|W{pD=hygxwSZQ-#G|A<`9?he}Xn(V>31Zuu)NI@8)PNCo98Xsm^x$l}n;
zliPk)|5s;f1d(^l?H%*A5ENN<vPYBD<o>C4)E-1isBH;#v=9_oosf8M&de5@9BCa8
zDP+|PnbAT}B;9ef&kxrd^m8zL0k&?>+1E6GDw};frJZ{<VLn(YEVhT#!_WMwY|a?d
zZJ=x;fYXNNucfd3%^z%Dmo+lCpPlDj&;{(tM<RU$U8n>_HpcJU&@;B>I<QbUjys98
z6D(9ED6(a=OK?RjndeL$U==esuNe@6RDvSgCbjatd2MhS>?|n#5|J*!&Z-1OcC=~#
za_yv9C*ib!NFk95;k2lvQlzX^czuVPlLbgF4e|@>`T{mzB`C7{-G=WI{8F0;R0NSH
z)b$C7s02l_F5Kd_SKZl5pdCPDE$6zH^Rf^W+3%g-$Ceyf=S-`E$Q$PNhIv^?rAWs^
zjiV=26=6l7*`88QfAa^M4-~j%miJ76jJGvt_6?EVz=BnRA_tpi&$-T8ZLdmw;N+%L
zt91IAg&1sj<suHJ_r&hi7jnPzSqaV?j>HrqO#yGH66_}@=v(*c)zctGgoDT&BFzEC
zsRTvN$Hv!q-ER70$V6Z^Z;A94W}}iykxA{RRKFQ95C($v#7Hqd<_|XKM)*csf3|o6
zco8szT}0XimZFkMk@2@$k7-^9e1IQ_Y$eiG7(^u~l9$};#rNMVNq}$$^O!=VDd7H9
zQYlj0W~yK7I*(vA(93Qj?S|E;1V!@G{MLVLIR+v`3lP~uq&?uBRe~bF1a|6na>S7m
z@ByHc$<kyW^9P$R7YZ4s*H>mLPP3|<L#%QjcB=$Mu5{v-cmK5nGP5RNj%SH<78I`%
z6e+Zmk|qs*U*DN}f|YD$+-4SQAt-XMw6y;hJxU!wtFYkLM0yRqs02lR?KL$pYjD#V
zU@0)r1tMJlwW$O}9zL$Q*Yo6OXCNI7G)HyF(W&c3S{ns^%Pb%8<H|cT9Sm*2NjP$j
zZsrd*KkBvArPw}eJ8VDfZFi}=zxjjBPtI?*dYXIaDa-)$xrs=dU<N8dkylTS76+Ez
zDt9ong8`~>)m+UVY<`{LeCSAx8-d{4LESTnG!v{tB`ESHbh&px8V5TP2_n;oGz|t(
z35vWe+EZ@y-48dSCNRyeQdb}I2b)W;o|x%n6VFwpUT|I>QL9JP!9ol+c1%4lrMDM1
zZ+@pYq{B5~PYRfC0SmQ|%6@)5dOqCWP6ktkJt-$rIZRe1C{nh=b4}^DS+`+5Fq2+V
zFWLOT=5puWZQsvp0U@;>Y}rL3T?7ME35tB`mU8fSmu9(&g;~qvtj0N-Kb6fFwg-<%
znZkhygEr<7X&&^T5)>KLf57O(dfMH&imm0wj9v5cToO5#binh^Lk9u#NTW}dzIk~e
z5f&2n?=0yMFQ*aLG*Sn7HfloCu#=Bk5MC~3b&8oEMkBn66I;s#yz)|YSgsWwwd*K_
zp_#*6=7)ijm-jJ;easI}gw;{wtnb<$n*d#B6F!^xkF=zQc!j67mW%rUOBT(`nZz@b
z1m#(dzn$eG4|7N|-{LD*ZMnWf59*~IC&$Yb#Iu3~Vb*2eVD^e{mjvIc&dYu3^$Egj
z6c1W7%_hwelHehjnRb%q80UCM;!FjkIp%jfq+_NJq&a44ynKmvzC^p@C4z_Rn#W`6
z-O6&4<00C3JfYqgh<LcedOW4xc)jHzTk&{Cy)lvD;nL{woO<JZl81oi@q&8qu*5oE
zPGupf%%rJ$Aq!ck=a(@5B`id1&u0GFECjcLmyfWJBg~}nM_I^GJ-?ZA-OSZOtbv!e
zaIRZ4-ac!~ao;K<j#}}68F2ZOR~?YSz(d~7z0&iaDECRvZ(`ggrjwt+aWgm_pTu!V
zdVW2}t=D)2VZZ_u1P0NA8rL$d%QqhgvEt-wH1HY?LU=>ryI!N+H2ykma-BBS_!~6z
z25qS4zq&y?YJ46G&SN2n?r?G{S3i~WM`Q#44)akVAM{5xEm#k;wIDp$z{`nT<3z4K
z=106dhZE;;4e-j22oRXE3jb(7T+8{a<?3UmB5pd~c~RGs_8ccK;>1OqC(cRX!xnKZ
zG=3!~uH?M%;>pP?xfUzAHpt^9!x~gH2}i7LfstS;Du#p+l84kd>I&(CSu<*jPz!{F
zS&V0kec7)}zu5lk`-0dNtVso~(4`<^1q}*Zh4*CNU%kTLy3h*nP4@|Zp9CVRRPWZ!
z<eERbb%Tb*gkMYo&{5Vu@onzNnsXaf;pJ$LXg}=$Oz@bX<JJ<7wM56CA|9vo{7d5T
zQsZ&UV0J26hFcNPue_;5C+1Y)<N_L0K!Z``oP31_UD5MJG^j|=U#C5;(^yTulr}7-
zjYeAT+<1auL<M}?onHTV|3j(!16?kT{QS_~R)mwMa(+{}`e@JSu|Qr$<8X$kFlbOQ
zI5ZY?#n7N4Z#YKqqx4;FeNHD(PJYb9$4sgWHF6mf%b0}Mso)n$xlYxCy22i$P(FqF
zV=M}3`!Km`&F+oh2$WK>luEcN4)XGLZ`c=@vUAX5>oOLJJ#z9#=Jt_^2yr0#2bQj4
zemH3*euPwzRv14TUpd?VwO7}-X1#!F8lhfH5lbWX{#D!Izb+;$G_lj&liT0g^74G<
zIG@$meDnh5xIoV@VvdXS{9@*~SkEtEj!X3X0p@ss)z^IbLFRZ+&u25oY(0O7IUdsU
zhneGHjYqf>_D#h)S6}?nHB}<_F(`p+Q=v^X0bWjKG0CjEwxT^OW)JI*GeYPSw5$T2
zcqn;5(V{X)7X;WaY(jRU!k*d)Gg!<F{Rlf)%noe?9$Z`4O0FroHieH}$$f)7#(U^k
z!A#Nkd3j*uzz!H&czIC3phk$n@$xWXm^&gJygbr%WF6$%Tv^q0Wsj)QHei^HPhl?F
z%2Sz7Dht;5G{&dt+gCAum7d?v`2EZUHKoL7pCc?-<Bu`^n7;iK<4@`Njhw?qt|q#2
z3?{I5D)@^EMAR3yV4E58)wW;{3EM-W@fwBLFK25(e6a^e7&(_T&K*^%s8}Ce6kZ=)
z5Fv&URb03-XM-FrDv5N#qz-Xn7*R!swW)0*rfnnw=dM_bDT73ywJ2GlkV!l+OYGX)
z_jzW$u7}NMG<qu68P6!f%%BwoW5(T&4PW`$`YMly|JS42d+e%&K1Yu{0GEG*dEa3E
z=&%(2%Nwk-#@}Jycl7Od*q3+o?RQz7yIO9I05>dMg|*Rl@&LG*@<;??aZ2DdT_F(|
zczL;q2t~vLVG@MFp`%x33|#YLWyDlQA`oxm0njmhBoT=9D1J_;ARg%F5G;o?t{~+Y
zCLu2+;&T$LMY|m8nnQ!oSM*CwN*rHr%K3kU4NQz}Ax|<x>zE-xW=&lRXdR4CutMr`
zMb8&emm)oXjk;X>JKx|MZKnPHI(5HJ>+AXO>$IK5-=MBHsE?kPZ_o(j`*&z{bZX|N
zIF1Le=Xrwq;?fk}?*wh2@%NbbJyu`a^LwoQJr<2TS|=ErS|UJmMhXpTVF2SHo+xNg
zs{(lB&=o<0QVT%49bzd1z@Js{AFXC+4R_bNxr5!{Mpi<F66%4VBM)a;_=9@j;l^#X
zvNcx+a3Z|i(W#@Ww#Va~#_70p;*?Hwd=7ES(euxV({nwaLmhHx4OAP}9bnRGwE-uM
zp%~_(mKt#NJcL=Nshkg<Vud$N<-XGRL{3cPLi7nW<kaoa2lDa)&TRn~jJXvrFXP;n
zalxo=tYLtutCb7w%#;&TIf*E>Km*Ddv92J_ct-I03QJMz8Mw2Ug+My27Bg@uSjB*$
z)Y=99f(O*hw2MR_pr*Jq>rCQ|E)8oN;6#<e2JRaV79^%?#}RW=n2}o9KwpcM4H#G{
zYv7-GaKWY|5`ivQ0S~O_5@!Stu;Ky3s&x+ShjWSNToQzWcrS*ys#ZI2R}}{pSwuRZ
zgW|zd1Fn)rXuE(i5TTg3WA@0)4~gqTQU`g|HLOT2mEh*!%^OBkD<jyT>{4U}>43lF
z0ZI&*LK`8<2y3LmT<VTFBwVGbYbvdSJXTg<M5U~PhwaDWeGLoV&pHe_=p({E(k=*C
zY5}`d>MXi51nIQy3<21qah8gp<qtffjFr5aC~NYUsP!f4giaKTH!wxDWP{qqFai6e
z7HY6TQG3``5{+ud1O|Fgsy4VtEZD%G)iMq4H--)9SS`+I+MYtgrqF2AZla}F1A=Oa
z276FUD`YNdg%&gGk0>!Lf37RU(rc{tHRgkUQ{mfOV;wa97W2NP_p`TH`&*iy<>lL~
z&TUo?<q<*wQ&%fMc=+Jr&0-Sux#KPtvrEezu|fo~Qz;LjPI$SOsaFJIf4m%TijTlV
z0`f*-pa<rSbG~3RUd*pBwc+KiLRS%iOJ43KbQ7_6UXBrB#Pb$hotM8Bz82Fh^%!36
zE_4?Wm*wRiLJu+BQp(}wo<dJ?zootalM;G~+NAmjeMHP{c{xsq6VZ%e!a}^LO?ZGX
zK*R|{qALs%F}m>bBw><>o`RPr3zJ1e%;69bVG0qk@BluT2vdoO$sG@%tuT#<i0AU)
zFNEntL<hoyH3&0^h~~iq4lK+hB0|Ip{wB;OqK3aE5h00)nqx{PLNXE2Kmjx%!W<%^
zH7O{kFpr2D%DI6E8;FPwmxq*6_=$*^WAcCs2pfrrFaZz9wy=qa2zv8yK@m0+5pOJj
ze-U8|%*XQBh=&q}u$73pmg)u%*|e~Yh+2xelL$MBsNsjZh_DOfEd>)Lg%vW0h)Dtu
z#SCFL5z*iCP`MEH5D~L+9*PXYSt4TnfrmU%I7dWG9C-+)!g(TMbmgH|AY33KmJxUW
zaD-eUBBH|sFfCjpB4XBnUlQRG5wRkt<efqu5z(P2_@_`nMD+7KI3D3L5z!MVd8u%P
zh+1A+NQ6Qn>fOF@m5AtncnGz^pG3si3$P&(UJ()VS02hC!fPU8rG$qnnec{)SU}^U
zP9?l0B9^dtfPIB`M8s!YJY3_2QX=BLo(H!jl!5(O%FPOXD|{d#MiCyMexaO*TB!d>
zgpWkj(&JA=_(Vi~P!Xn35#1i>lhwM&d=L%<eX&{v%m)h(N)BSZfOF7tkW|h(RnITu
zte0^Pn3M4Ga?X0W#?N^gd3Df$QRV+L(2>%3WJchm%iAD~jpj%DV;o4C;~bK{Az?q{
z?SBWdaYG>@sMT<cD0s7l2DKE9ZWbXH@V{yS92+3ea&D=ducm|LoZE8F7uSqQD)?Wu
zEUx`9g>y;aMD2%v$2<A|hj%VdpY%)|+_pNbF@aC0T;qx+t$$#-HvaJEMU-Epp-31x
zU@2;`9S2UYx+yzl`0!Q#xJY989ip64d)Lyjj$+5kyAgVB5K=1eT09x(IWP!aaC*eO
zaQEo;_U0jirgFiU)nZJ5Un>QA?bmTpab3{4qj|#2)%rY!W%N6c9qlw1=e0yLlSF2c
z4thruu!l6#9L)hD93bwP+w$^Z;(D0WK^~KA7*Va|qpr5bTyAcAU`PPmIZ~NpDyxeI
zqCgprM_65iGSJ1q!qiefT9lG1h7^%j=rfQi+hVT!qu`FQhpYT&*!s_F02INDY0N_#
zYAG`=)$_}kahaZ9&Wy|T{0e4Vq32gJ<4QfhiWyhw`PIz0TF<95W4fMS!;EY6{90yQ
ztLN7-<2pTmj2VwH4^#~=A7{qndj146p3w6@Gvm*CK8G1|^!!O?JgMhTG2<ycf0`Lj
z>-jUxct+2kWyZ65{v0!&)AQ$<@w~=kH41h}eQ$xrfMqT)KlPObHb9(bF|+kK*lrfH
zTg$;9E_1FMIT3Lzh|8SoCXL71hX4+qx|dR4ylN>{7naIeYo@S}g)P+cOIYX<)=X>9
zW}(@vnWp+9EbIttt?@@$*ik*7#~kulP0hjQvCur$Oydh!ivre8bNG*$`(x&dz6A<=
z%<&m>)%X(TSfc0Ou=a0QR|KpfU@`YH=8N7`;iJk}M~(l;q>s!8Z_SEx4M^o0qmQ(;
zyj{^0znQDOnX9iUb_-W~i^gX?8=qxNezomiFVx_IAzY|0YP1=bF~>4iQ@>(|kE{l+
z7*k>JBg#7)%|Q+f8Q2QlE9MIzq&~Aj2dM=7sB+Q;13to0Fb(z1jdqVy-$tVVFR$UO
z*KpObOm}cV@x@5n3%`Oz4&jDW)2?g_2)BUf_$<O@>G`{ayQ}AuDVI!j^7+(0pBlCP
zUQzp3dj1o&|D@+PG5bwSCx4M~7n!cV0>%~S`9j7O>iIW}d!y&yG47q7|G>BpdcJ~j
z6?%RKXFr3}ji1EXC+YbVj!WTm^GV~lG(EqbvtO^tAF5{k%bBkqcKAnJJRDxz>EY~m
z^B{^pX4a3HGoGl!vuD|yIo<ckKZ5m<b4A^EHhuluzi!%iSpxT{zRkkBFy30AL4A>h
z4Vcb?xvKB5G;{q(qCS!?81V7_2t6o|ve5YPdI5h{UtwX`&mFZj`on`^qyM2!yuiSS
z>Z2_Dg@QnZWRg~xC}5_mEXMpo3#Wzx(4aoZ!XEO&>;BlOqnP%u>m73$n34Ma3ulD4
zS;#Pz$6u&*<uYoTNg}k%XfiP+lL+myIhP1?i3e^Q<{{9(`Z^4K1>VlUiKuVFP*S-r
znJ$wE?YdM%tc!><rZ^aNphxw=7&<|`|3HKKJ`A<~t0}*A>5dxYUso>PD8XRV=V<yh
zMqMUdaE-qv?KZ`BUp5pD>N3JF(+cMgx)fifU+g?<wncpj^<lATXzVGr#@wK<Ipa$=
zYGZlfra9wF)cF#viNF}Xc!S9)&)u}VJO`ix5{k|g5f%_rp1tAe0jMM@z6(Mn2nB$U
z`sxj>NWp5Oc9AZ4*j{{yE$gr{v9W`KT-je^)zKp<eCRdSMB}eB`|JAl8_ekjtD&{u
zV5S?ag~s1xPB-=Kx0utdzw?c6vF4imZRT`a-+qTV-T6Cjx})#^E_1r8Z@<T!?){x_
zd{5thF>@-`+Fz~8J|48Yw%5OIgZS<cwnlk$sBQB&aa;&)@@vN@&!;DKZ1_K4nb9r4
zXzJrd&2P*jVe?2df=+L$|FJahQH@Bj_@jhBs=aSRR1A7h-!^JJB$KFQ(gi&YLWbZ#
zl!uO*VaA!_BJk1&priFbXUuTkO*}Kn<-^Yo|MN26CAK^0bz#rj|5^n_-vp~spIf5)
zQ{H|l|JMx^z+4!dm-^(gvcXbNfd=)3C3-akG@)zdd8M}gL1BZU5vP1#G`V5HtIT7d
z<~OXt8}>P-kO*eMuhnOlnrkW|o<$@G-EH|ncfWO3b`k&ZK)5e3G39Y4E(|vq{;a;n
vM7dG@hxS$MZS<)4(Lu^T8Yv(3KNJ4LU}#+VPbw4Xc|~5eES@zsH8T7kpG)wV

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/workgroup_barrier.hpp.E88FB4AF7CFEB4E5.idx b/.cache/clangd/index/workgroup_barrier.hpp.E88FB4AF7CFEB4E5.idx
new file mode 100755
index 0000000000000000000000000000000000000000..f5902d559933b65554d7ef33138152e81ae58dc9
GIT binary patch
literal 1398
zcmYjQd29?(9DX~U*V*HBcWp+G-LWn0XxFCG+ET;X9*Yg)h^uKd2(@*jEp`zVe@LWA
z+(&{&i7N<^(o}1U#+f*a6jdQBR3hStBPDUY9sTXjB>TPn-Z%4`_kF+LOdaR*Rbzxk
z%^$a_%D;RrixBEVf5GK}MQwC-Xb@Vp-@ol)^|f5>nBSxF>Yse?Us&E}!{e;$E3Z^{
zv>kjtzNXV2KJ2{TSH8HSB#>vV?drI(qHcCoPH1-1H*3Xo|D#ub8y>LeWF>oRY4K}!
z$GKhcce708>q^hhBnu8T?5wbag=0DFvf7uWEuQe4(Bvlj&tv%yEX~`N{Oz919c(FZ
zE^Oa2?8`e8a%kK4<c+?vrnbGY@aG-P%rd$8Y$3_qG|gf)wf4J!LziZBm2J8Cxo2dL
zp{8T;2^6fUnkQ2aSv`|hUJm{!H8TblVSyKH3Lq`@NUyP-dnY6fI(5{{=%}nq)>#z*
zxuj{+(Ygi0*P9taWsw(c3Pk13A<p7(YqFV1pt8hEf&w5f%pF|6-=v?C$QY^2nYc6s
zKwjJUA(%aIK!KUjP+8_>s{$ZzP5ra4_05Y)Goz=nn|Iq3h|1fqJ{{56JkiK7RQ4Ht
zi3)(+olv$h;iHc=Gg>Npcu$f7ApZ)^>ORu7&Kb`b;}FIiPF4WvlvFfW6Yy6AQ{oWP
zpZ<JrWZj>)MuNHioC7QotTXA7AU(=#y7U+m1R(>0Cs@c9hQ^pEh#3$o!D6;JG{&TU
zl33E4DZ!E?<*QhcG&aU$o16~$5$qP+84wr2o@7rN*eBS>`pDMatl$(Hy)pz4g16>V
zLEFpv2^xZ37MBg)0)FE3%1prB1gBZjcqm1%!{Xq<0<81}kzs*Hpdzgl84}ngSeE2`
zm5m(tFt7nZ(@K%Sfk)tGx*Zu9*oLR*c4SbZ+r7RqZJvu80W}FuNlnQFd(^bx%77yh
z>`nCAAwIn_Q&6s@B7AqW{`M;^Eyr?3WmG88?z1`-6IemTL{?NWiIr4LW@Qy~vu+jh
zu%0M;HvGib)+MV`X@k7P+ZRp;D|H<ia5(0Bf5GPT-GhBB!FGdP)u~KFX0%fT=NfYP
znCl`#5fAmC5TjQ<iiQKuHh*ZxdFoo#2^8>G&W%J643r1c>B`_~i_au%Kkhoj#B(S<
LBK^NO_=C_tsX>*(

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/workgroup_synchronization.hpp.EE3FBA63C405EAA9.idx b/.cache/clangd/index/workgroup_synchronization.hpp.EE3FBA63C405EAA9.idx
new file mode 100755
index 0000000000000000000000000000000000000000..39f30da4f402c88a90ac2eaf4bf7bec5f92a3023
GIT binary patch
literal 1004
zcmWIYbaQ*c%)sEB;#rZKT9U}Zz`!5~#Kk2=nI{+-7#0HsD(0LGvdxz^<hlAL{-9#>
z`hp@4xBRyO2fVGyY+jV?v3{BoxKw4g--Q2hul_Wu6*ThbUT|!1nfTt;E=V^&?%nlk
z>OVxJob3#EF59ZPFYMK)uQRsaH*Np1t7-Q1cKg(L&iRX;{}PQkyL0`6+2!mW%1-%~
zYC8>Iz1vorqPB8o{iB1kC!V~WKKI1o15=o9a2U2b2wHovxLK>86neYa{N$wl(KaC;
zcKV-qdj1z<TW*eo^0L&evl=hY>(Y~%khnbdScJy?u9m%`PsAQd&Ro*(@gVb{ghAxO
zlS^InZi$vUabMZ*o>6?CyE!BI&qUqiaH-|>*DP`_Io)G*yv6%uaqim9F9Wtr&FE8d
z&$@X0%#T8ClleIjTW(CYsVm=hW5b$%+1K}Hq;>C^cxk;Kf7WNQ#nU4gt(G!w_0PEE
zwd;#(*SGe?aR=tJIWuicxVF4c+2Mw;|K|yhwsPL;?+=*y)$N?$%*?%Em!|1-L@cX%
zT@ZWyt4H*j<14IX->&1^;-0tV`a6N#HN3g3QQP@ubKYFMdab!$dBki_t>mntS=SfJ
zt*tOyG*c=?ZJ)IB4{nk4FWu5-t;N&kZ4!HXF{3n2zVxQ(R^{6(<3Dq3nN(bvo74h~
zS1a+A<(u2nRCpMKco_Jd#GMqt1Oq2CBdahsBR^x!6Pr1ndzY~AGOz;`aEfxOfC&ap
z1_lmZE}(+>whOblFK<}S!yo`updhZG1SUWxaDq)Z%q;tU%0j70JPd+B1p?v%@?e62
zlZA;5te_}0t=I#YwqR~y<N~{4!#wT%sk<g_bI2^YJKMx1n~{g#jn7RSrpkL+r!mj$
z<+4CwLk>ehm@{DRW#Iz56J`hxzl4Z{GTbTr+M?PjFfBlHb=<6B=E6jQt^j!kX26cL
z2WMQkPCaDi;TK>R0A&d<;Ne$bS3u%9u{*(eFx!A8fqV@!>2<1IkMa~lH=s$<?9u`-
zRXqGE94c^c@bJ5`y9&VE^y9rreEofXVW3}?B$X6lPW<C(?ksb4`yQaM42O&$%%q~!
hoJ5f0ii?tq#X$@{5CH~E{G0#ID`6J?zL=4Z0RW+Mas~hZ

literal 0
HcmV?d00001

diff --git a/.cache/clangd/index/xdlops_gemm.hpp.C8FD7C44D6227313.idx b/.cache/clangd/index/xdlops_gemm.hpp.C8FD7C44D6227313.idx
new file mode 100755
index 0000000000000000000000000000000000000000..3a523c7a8a98de000851f428d21121b14eb79107
GIT binary patch
literal 39200
zcmd6QWqcJ^^Z#Ap?gC3AWOE_x;slBWFQiyo+^x73yDe5IQnX00011I0#odCtyL)g6
zuEAY`|G#_BnZ5hG`9CjyZ=QZW-Ra59T{&lF&N*}DJFQo>Ql$)2U9R%&s`ly9zGrLM
z<#Gk#-yS`?cUUU9T+M=Au0AtXG`co6sdV7m`Q1(*YfwIFOkmR$=h~ldu)b!~4U>{e
z9n0xn{`u(NTer13-}_r<<<-dYRkqf1eL6n%<)w^?+Rb;@vh;uYH2BG)=BqO{xH0JY
z;1LHp)v7mR+~v?7QRkjKKk?*Rmfvfh9k)Ge;nzEZ-{p&0H>1;@dBG`NgynMWR(}QV
zNcgA2vO5FQH(8ZFvck6EL!aCmGjL<6{-d7!F*sM%e=`>BK4AQsZX35v8Ze^j&>5Yc
zXw!z2Zx%JQS<^0`y5u(=4l8>-?a3di_&C$ItX>>{ODl5o=E{Obm#U3he77}6!i_W&
z3nrgwIB|R8s9ej%-%_WOlcM`|{IO<I^xBqtAI55d4Vw-e__C`vx%bY41EfVokDORu
zWca>uX{IeKSvKsdw@K+wd%7Kc9W&uXxri&9PMViCj~@7WNWgDH@63r_lQqjABg39s
zo=WeF{j*!lGAj3n9ZkMg*<WV-;eT6vzbh_B`tU~e{xG)8y>PitkHmMzuIvNCAGS0$
z$CtT$zSf)d#XrT=zdfSkf|k=iz8zEH(d#>_y#sT6dj6)|nKNHU4N9L%?EPKyOq;*!
z=Gi@b_ZL_5uP<k{FR2ys-0aYDS@CNfQf}Xh?wnZ-$Z(?K!BaV2-9J1#M`ER&v89#1
zY484gx76>)N0n|>HtpT=bq*}c{7>$}sdMfqzNO3l8bupUE8IOa?8fMjiF+%@=Zw8o
z{K8maVCcmY<G!2sw&|eae>N{(Zp)v8|2)08<i5QvuOB~}c&+`Jy>(}&Pq=<4=+wW>
z`}V%_a^a?wu??@DFWYj(g7o*REL_^L_UY;Ke)wy{_ha^?9;q%i+FZ`ve8ssA=As8R
z>rUUCwBW$OnQhh;&6prBHok1${a}5uwr?w~Xi%c;{9h||UUDsSi-l(&PO0?hLHMJN
zGb^<!Uhnq?OP~Bv<HF?j=GtM6m9gcIKkS+>aaQW`jT=sCRBMagZq)KTeKMsF37Q;#
zX<mmwPqsBthl6w8TieGQ*}8Fq?44$NE)?}{c{?wykaVoti?8F7yRL4S?T;D<mJOc}
zIO)QTORIC{YBKt}wUx6!iS3fRv9bP0nPa1kzs`r&9h-fBmB#X;2R{wF`~AEXogc>c
zTadJJXSKc&<5Ds%`Ll|rbb}%}{*3B>DZEIrl0wZ|5AxR?`m{#Z{WZG<T&ePB|I>#T
zv<RHme5ku~uXlML)o9qZ!J>}k(+yowL(--$7?V&#(5B8Avm$Sm_p2L3xjN`=JB-h2
zPTHR8bg_cNru|jsM4HpnimvVYdid8Kb=%hXVRQ3MV{%sbaqR4S--Z-<^-WP<<oL5k
z-!5&^M#6OSj2b<oZ`OES>{Qd`9S}AkkMH7A!Lt4-lxo_Rvw?yxcC794K9$X<a&As+
zk;WzS)L7fAa<CX)&E-u}%_OxDCnO>S&rRR2(276Y;uxXgcit1CazYH`#FklqWWxGQ
z*@uRT6>ytqC401zpA!-hE6r-&J*1>7z}35w%NryIW$;~yE$2Pn=39r*8M-)DXyo#)
zmd(|2K2B_z_CxM(h}WM7it$2Ym-nU=a#QkfV#~yyxY;7=RYP5zD71EY_eka*DV!5q
zW=DmM_qNsQY(1;F%ez1@7YO+{AsP4bs|l5>W~rr%F+x3;_n}}u6v8<n8Kq69u|E`^
zGfo#H@zw9uu=i>%PDn(R6L+4f9bY_HjDW~sCC^}`2q(6PdZg~yj>ZmM94FLpc~1!D
z384TdwoJ2Tdi#=|KXoxyXyWn~HVYT@U5KAr&y-_I$)r}g*tx#T+um#+?z`CRQg?S=
zIr-!u^rWfFTPLhe9^b`g&-$=2@15rr16<uPG&$rPnS2*wr&)LM__M#>RIy&Lfy+BY
zHiyUsI3by|l@k^O9q1P7Q<8Mvrb^R5-^Jz+Ei+>Gji+(e(4oth3n9w|4<{s}4Oy7=
z?a6TstdFdYVG+#*VnI$wCf&9pc@hVAnrz*tt;@S!3Rw>oV8oVrxw%lp%6$=*3bb>1
zCrRccr~o4*^W9%VYM&S>E{AFShM^P8_d;$?Y?0M9|MKM7*F+cp#L&G|!(OVHIU$jN
zRJ&HZc{{JEF2)JJy1eH_^Sqd!6I<rz7K3(XPr_Gq!IQoU=2z$uBP0^^-K0IS?;Ay0
zFASMag82zDjMy>(3A0x<+gt}j*v#eKCx-14^KfE|oH<qam$&a>?p^A*yoY7;upG{b
zE%Itha_f}xp_W!-C`ZcXNXRfkGQk<d<$o;g_uR4$SlK1Xyd>r0#FqKVJ9V3y_qZ;O
zx8C@pV15(|aY8a7F<Xki%5o&XCDR69E}CD(LY$CHXy*})n|*g>yY(z6Zfmo3KHtUW
zubrny<!||m)xmIKuVwSKoSzet@pRc+C$_@2d>Bd?>Mg;%g`s4`7FqIncJ#W#F6&kB
z*-^3?B^TnvmO1`g_&*)Kt-x*2!CORgi<p}e63I|6;(F%a{$6SI9)|m(XkHY<IU$*_
z0-g3=DbY8bHI`7~#iF?wGK|<V5ql?n{Gzw>8B3t|ysVs;iP$o~Up3A?oA;yj%6Qfp
z(L96RV}xWf)$UyQ>6BHFi51}gdzd{6`YtxV-i1m18$KLw^)9?dakF?n-^J$7xN)`1
z#G#jb@8R)|5tT9U?2Onl$47NIR4MaO>zmOpdqi^&zK0Q#G4?K5tM~i4hh>z|ypwXs
zN!h~*$z<M9JTg!JMkCO@c$GqCq1?U;v9qKbx1-gLUhAw^t?TkWlgwvOIz~t)TXey#
zbqfS3)?-?_ynBU^z3?H7*fL}Oh)<4M1BK~=Ub`Tg7x1=>kVy8cy{nHclF}c(5^oY|
zMuz(?#Ln4w`tJ0vwoJ1I5FTckWG;jEV8oVL5Y_X7{!dn`<KfNb2<9B2ASbrWhg-+`
zPVPCx>Nre92PE?V?!ySl<a!!+BJlf)AFXFWckc=2Js3VCBoqGkkcdfP8Q>hcz<1X)
zYlizSHv6AtZ4$Sn3HFVl+j~S%jtGIAkWB6Yhi+Y&_r4_hz83tUY)-;DfDsb;enF0p
zV|Q2mKEO3hz#I@Fhv>cw@$)V(kZtz03eBvqY31^c7tHZ^Q$|Q8-?+nJYd0N8vSjd_
znX)+(&V><M=J@5)Q&aK`z}LVKN66*~h%iDT`9sTB>vAP=j-~k+prU5cLcWX5A2gs;
z!RD<uTaSTHcq^N4aSujFra-!x{nt-;-40`nF<dQ~tI_w2*dpfEyzxal^?(TUdar2i
zg{ClKi!@JvKjW1?Q>=-vq04(&HZNo1W5kwu9QNmlE{_{oN{Ra%6hjV*9!^LmBJxYp
zous%D@ZFe?rb}Vdr97P2BHdQiD*fwZECM>g=U$V-u0ezmTO=jbpml*I_FDY{nG2G6
zK`O)v$rP;e^VUm=8)pWKc`#?qS2N95b8<o=h5GL4)uC(cFFu6{_MVoM(^4QOBvW`y
zr9Ff9=Ie#Gf$AQT!VckW7$K1&X<lwlxzl--<(Xju3k7qbke?G<ri?hE<}6`0ZiB%d
zC5MfY^KfE|bjq0eMA!7qEepVH<_qS0=oKTj%$@>ELLPc6;Wn5@hD%|?@ivT*NYRuT
zH`7QX5BWYy^-hqL3Gk?lkW8^(b`Dsa{e3iU11EJ%4m*aoVT44A54$<-&UZ1dpq=pD
zOGR@jbb=8Q`QfoN^|aaf8ALFaQzUZ=G>Z`uDbamQ^#O{vrsd0^ps!@}6_&D$*fLT5
zp7%YHJ*_pESnjTr&6Svu7_ntW{q<vsR*|ummuc_v-WSdLSPwB`%Z!VVJtd}MA|7k4
zbtX&ZWcW=+NTy`NQnN!3oxKv^in455mvxWtV)MHn?V9!C>>L5EK46E*VHtcEVn^mG
z(ewK23LPygz#QGuY+1l}vH5G#-JPB){|0y$*xF*rTrB10#1`3=bz`F<r88S^2~8L#
zo5Ro}jF8Na-YJ)p4!Q!Y9>EOoL^7X1y%`~yQguQThW{}UlSn7n<1IPt7KV@!TcrD)
zU*=9+F9(ZxE4#dp-R5I=K2B_r{tW}Y%fBSzg<H71v7#A^nVu0_WLlNuWe0A+f_b7f
z)oc;WEl^8FNTzH;ja~KY4#g6(GemX@=1zz(LL%h`)Nh@&{N0&Q7CdR9XikJaF+w6g
zB?c{+m|^KMh~Sg1%jR`BGe$_{m&ym%ZdiIJ*?Jiaz;EVn;l2y;%P(&6W^v}D!z_mg
zU0Nrb>u?W7NTx!*JL@VX9lc`d5=LOKYz~$Sa$?I=>XbG@SGHOv4IdvTm~k*^Mr@fD
zyDOZE*wx7zC78$}(OiUO2_qy^ad-a4TLupJV7)I!@~LD#g^e&mGL>>S+WTl~>dV1m
zE;y>gYR<!Ic}_^AO8MR=c3pom)OsCwp?#9MPs-1UE%UTp|E{6wy9bMfahu_4p5bay
zPHd4EmDBC`cmGe8gT|8Jsu*$=4w?~?soK;KdIqfBW~n$HwM#O0Nd-8uWsbdD8q>1a
zLQ99S<ozs|pRs9Vgk-82Gmht8TF4E}LMMNd%x^FSMr@H%Q+uVWIVpRvSfDc7q<crF
z3LBj&n=KIkJblADx8tDW=-?G%*a|H486lDCfibSFXKul1cEWgXki#~>3>dLRdbay|
zAoG>!5W&k15yOVyE{xbBpO>y^B9+YH^O>o=k7eaCd?q6#Q!{6^BHLDG>Th)nT-kHk
ze2y8A5t6Bu;rRYP)Bcd+7Ry(3dEdB|H|`)#T&c5)_(;Jk)Ebp||E1^4V)P8u{Gx1L
z#1e!N`#vZ3c8=*A86PYb!dMPg^9)u?aY7<><l)OwhF6w?#oQ1{R<k6l#W=A=BHu)}
zPn#N+8)NOJ&kE*Qm?a~&%(i<YMh+8vi%mGzN22*i%*P4I)GZixVOPIlyYa$Uc2zPf
z<@Q}{_WsS&(s%#uj%80+x!o4b+gJfIV$0lbTK+|ur;jiOnA#GAumm9wC$>n+m1=p1
zuL*((bn~(pb{WIO2#M6YUM<TaPtKi|{z4&Eiy^C_zl_*2F9$ttdp&1CxFV?ULCHJ_
zgJXn5>i<0X!gsxU=dg?tO0!BfSHVdzV#~B@@aN6A4{h+xm?8Jb<{k_ZBP8;xx&L#G
zI;l@t-;FiiWZ9eyKf(yfGzcluq3n#xn3LkHzDX9$WcU$AY?-b@79Sc?y*M6)IsKDp
zeu58X#1>iEf4ukJzMm}J!_+oGGAF<&8L?$vRnHjFX<<)mwBcbViRL7DSVl;uq5Jf}
z-~VcJ43C2MY;HEs?Yr3Qa*xBxj^0?=X9i*3w}SE(3us1cnNF?cQZbuaTVsHGbO`C-
z@m+}D=%@FUe%bK{M!E~WCQ34+(Ep6sBL5U^{C;u17^~wjO)inlB~n37NT%^`akD1B
zJAU8#8tmu#nf=0j7veV=_;mJ-Jc-w=*T8e4rI2U@KNzuPW_;_sakY1?^&B|y(SkV|
ziphvAvt9`=GiDQ3&=ajeoGP1Bp$m-IGJ6UH430k8#JUfh_#?@Dg!g5HWSWND3hG$$
z!9Hs;z<s93<`gVq7$KQve?=NyHhdTuEarsCZd7w_grjG~7CHN1UHI*HSU_~b82&4U
z{R<;tghZMLSD#sC_xTIflQ7wC7R=3vH!xz$)K1^EZpNq3msqQJK8xmOESDIuWp>q@
zwz%$NEX`uA$>5Y|o`MV`B-5hw<7^La<+>E$>WAe}xEvnlyV&f(AD)eGSah^C55U2k
z63kPW2N<zsCYpQR%vf+2_6W_LCYjTq4~&pV%Rgssh!~b|-ZB#SpvR*57@;FZY?)ro
zN?)yZ1-FT__7=NDa~EV7v1R&hXmI^o?h4je;yw>W^C2vN5nJX%t?-_cX8wTJfjvGF
z%tvsEjF3pH3Oh!Pd7lEcnP{yKhRWtps01UnOsfG0*M{!;#Zo``#QGuiJ-!R^e>*WD
zs!#kcyDasEm9#V4h5IhVZ+%^;IDP)sBbGs9V^_p1Qowg1e%qfu9&hH3PJtD{L#>s~
zwRi<aY?1Bh<JWiGS0KRk7slp0`MV6h3$fd^m{q5L@yefq#e7)&-E*7w+<7@65pVLh
z-K7)zezG(GPHwMk?u7$m#FqK=d;e`2I`;{14Y3x?9@&%0cOm}o6;Eb3n0_L(D8}mZ
zr=s~3ea;BUbU0lrC7`;BkBPO$ZmVc+g{Cn=G981zMR%J$dX%L~m=&+e=2a|<86lZY
zgYq@nmvSBxO`O$fKbSw{_g(z_x1M_^dcB45zyxQ@VY9IoV8j-QNj+fC+&b&5al>L@
zm>4n)GYBIj(`C;0neWfh7oqRromNWbN@y%2B=Tpf7iGRH{S^DpI7{(Q3+8FKbw)_0
zYvB@CCytV3%T6&9?Uu~lSj00zGTj!o?zN}MoQcpz48T#@Jc?mvghYCT&VMjIWpXB~
zV=(6}5zQr-pc%1cI@C_z@#g&8mNr4F?+E4{gv}YTWn#+Sne(9lR?Zz^hvOx4Jl123
zkVwyF-)8sAcxyJ+V^v+=BdNlUq{_yLEz+fVv*U|T*0Y8iZ+t{HkKm0NA(>u3X^WpH
zcAsQ@IGorx(Hw_$F(bCj{l}*o<^RiPKUhqCltVruOvVVw{H+%}-=<!r=f2q}m3OYF
z%tc6!5t8XWIKzm3Bdh56D)_KO(M-hL$%rkoIaQy@19J@2#p3X~AKYOd+y*DM$hVs@
z@$tXEvn(FgT+OT&?z<4bPwI~~)AXD^(f1~y-kE|j6Z0Y?woGuR$zA%Ti^7{=t{Ekm
zqY!#!ghcvIypZeR;Hk}gw@L4fP$B|-7vlFjI%Z^+Hr47`%84<WC56nwoW_VPbG%K9
z=o8nnSu)V7qf*FG$S^`O{fjM%?R)AS=Av=dT=cJO{)?3=BP25*v|IAI3{e>@8QA1D
z(cC5$;KY_$KPzA3-Uqt|iv`hhQA(aDr5Go+$kUrO&Yh`(Ao>IWZf37LY_B^vCnPhl
zOw#j5#}C)CdKWr#KsFCRXBZ)wL8%V}nlbkfp6!A+o*|nvAi{_(GNr2=)S~8Z)@?A>
zXGQZYJPjkZ%&sNp>n}Sq#Zo=28hV+%3i&QJ|K-<RkwP7Wmd08ukL8lN93cWmNM^_n
zd-e^zlC`PTGZ@J+f;k4p#0bem>20n@9o=3wz|{**F`Jw%qwiv~N98(nxk3uOf2?Hz
zV?}c;93CSiGjyr`dExDMS77=OnJk)<@imN)$nYPs-2b`9C7g3~fs;8Sm}f9oF=C4(
zlwUrn#in%D2jgX91TzMU0Y+?@WvN~?ZLtTD)mY25UXsmAaPy3i%!mwYlTVgWwg$Mm
z;e$fu&<wtd&A$25!Cx!pgmO=``t@(~?{MFR_|d(8?RmF;QcJukX1lg#+i>57*rR?9
zsW@+?cR1M4+5M8aA0CwvTjbwGSFbcU_mAaKvBr8Xna{DqWrSo#pPzg`xn4&EiaMh|
z&I#r@^amp(GN$H?yBYGH!T`rvsy|&er{j=>5nJZo4W-e+n)3o&BdsYdvz*!EyAVG%
zs}M4#T<hQQ7~Eo(Y|er%GD0F_Yu>N?Ec8Z3t3R;5`X-v+FnWyGGE;|qN^B%_vMd!N
zJWULlhN*@TTjts4UYY(bRmYM+zYYo+<ndjIKdx%uyb}w}L$D~un(<1SCBuCeKYvE{
z{6B@}vvd`c!$Q$q2rFiUWX3mYc;ir&PhBk;#7fUg=6UQL7_nt;wqG%=T?w4t##ttM
zNH!0l+ZeHBl0!x;8u<H8pZ5y#Cdf(x_Q;HoOnm9SHS5(bjle)>Xk)x!#$$vSu|*oy
zn}2#pgKd_^LK~B0GYJc2Mr@gOciSFmaw0LnHPjlR068GkcOm|SF=Og?T3-%Y-vyJ$
z8`*q=5n{v^`Ff;6)WWIOHV_@#-|QdmyAXR~t8VqrWO(Wd7E8mmT~#8kDvdZHkx5x{
zHm`dzS2NrJ4|yk>@9+>tY>`#Ea8OL0V0A6j;EQN}LH9F4GLr|+9-8l;FTdzwQ9S9r
zJM2BCA4W*z-_fU<=Ku8-7M$a(HB_=_CgV1Ykj#_?C#0B+b=O#$fKj|6nO9&#jF8Oq
z2?3cq^{9<~Rc9<3*URR5yfGsrGNa1R<&3qN5RZzpX1W??jr_g~@n?=J6k7Sw!5P-*
zKoR!K=6;M0Beu+%b4&I`kI5V?es4{vYQzV%Dkmf|Ywp35m4cRC(#28`d8>xKMeLms
z5}C7X)aIT2AH@Z@24IM@%2_k|E;f64s<D0U-O7*Y6aj<Tf;k)0DI>PX+BGe-svXl?
z--31eaM>IVrDepHd0jAh)X%T?Ti*kpaZ)f(!ap)%%Y6DXam@Yd9ej$ddpC*7Cag0U
zA(^?OHXc6zzC4aCx?q9zKrkObu^Azec|Cd>Wg0gh3lW%V3A04F??UYPT6~#}Wn&S`
z8fOI^R><ZGOv;Rq%mPW>H)2UDoCtM<4}2(@53xpLghUqhl1?gnwnGnMt@(eaWbVXX
zlo66y)ca`K2gim_upWg$+$Wp+@F+%XnFU*a9apqSaqB)%?`N|443h#Qw#;U&SMwn!
zTjPuHIxi&i1yqg^Tja{&8WB@(#0I$j!6>Ga(`E2oh`qRL*{}P{DHx*8P}-}Kc@=kH
zghZBPiM}}G+S#+gVp*ulc17N<<m1E^DHeIZz?szFTizCKB~~zF(ff?pGPSNY&r?5q
zSFl(T>yR7nup91hPHd6Kzb0H2XFRp^1(W{@$y|Xch7nul^@>Yx7KDGo3q$Re2<8$z
zi4hV>Nb~;AvPl_nc2W#4oZ=2kfk883i(K0=B4S5rxZEz-p>G$=?YIjgw#c_1k2UF?
z&%{u}<nM{*JuF-qA(3TnkN7t0rEGW?ILe!{c@yT$h%KU(x)t(!M#Ox^Tb|>AXg<JF
zn-N<k^P7u*WNPq-)q9wS#*5~7%vFroGA(W{ANyq_99c&w)egzr0mWm4M3#5`J;Rss
z_tRi=1$}uhhrP$<iV<7nbdlfYg#P<Hz%|Np+nMA{p}q_8SCmf7Td+{GX92De7T=H!
zkMCmhdzah$KE5J?W#cX9RMo6n$af+B%FHdx{n2>;;9xO7dVa2&X)e49BeqDXLhHwz
ztN73wV|4vC!Q6(886zaKDlz0?@UiqGtVJM<`MzM@N7RB5l35+GJWKZ7zZ|k;;MoUD
z=3p!o8L?%0op_Y3(Sf$sv(T{<L~{Z{5{%d~uV<e;l`~%|>poa!?-b3Q*u61AGHWsw
zY+5#9PiCM0$>iN8Dcht#PHdUq=1$HP+%Ojg2yX0{U><`zWW*MUUY`(HW=k4tEdd8q
zCZvqVcd_|<Dr{eqP@p{q5IwU?2-}4aEF&bcw)53(k#pDJDV?!79xR%Jp(%`z$hy#q
z9eak>6?`v?eUzwtfU{(TWY)ip-?w<jzKfQ|L36Lj<`rlzBP6q-X3X8qMcW48ZSb-`
zhW(hwcd^;cf6g*F%_6K>$62<TAejkRv@l}JY?<4_OxzFUkF`ANa@kxC7siM!bFg@y
z&)@%w@I+?}<1on_1_#6liEPT26fv#k)!%(zl-V06DRJn2Mr@gXe~c}+boN`TZ((Pz
z#E@55%P~SSn=72^*CA-b%mCNlQ10|{dXMiy>@D%rJN?)urIGJ3X}!skl8nbNLNZ&k
zSNkz!_prACuKu`1PB~|o?_#raJ?NBFqpQznW0~E;Y*Em6vH1`C-no?HRz<5rU_X`3
z$_0HF;%`%W&d8OeenV>w0q?e0Fc-rwF=ERczBz1W#q3emm!eZQ%ORVwF=vEiw)gvY
zd7a-rltHIr-i?;cXbdkSB(fuBVC<od6Yp5JflQ1X5`)_?V#}Nxxb9J*$ojq>(Y!YV
z<py@ujMy??wqCB8e*byPC16WDQ8FhYh{6cT?99~opCe7~r?%`6gFHkshrq8hLNdEr
z?in_y`kDmav$A;ah{_$PEhDx}|7U8ap<z~V9Fx^q$vg`u$cQbnEk51c`|+`sW<%zp
zWL|_mF+wuC<-#SphW)a{QeE`&T)~`+{Ual`%#@M`H#b?*7upZ)UL%FA!7)A~w#e*N
z?mwcGg>LbuiqPCt%CJ=FIdP@Vh7@({C~5c8dzDg^y&Gb^5Bx%+U?yTj8KE2PiRqiJ
zX_^HEtnR^5xR2Q<zwbi)y-S`~zWKBZg19kO_YIZIq4*|7Y?(ESr{662UAQHKVVEkJ
zQz65MEpy@A=N=<=|KRK50Phz``2q!J#Flx~x<JpuJv(D!ME500W)g102#M?~)9gXO
zq*AchPMGT_2w@Z8rWmnBHV>(`d{tb#0M{UxNk%zi2H%C)`@<`ybo_A`R#BZ1C%7b<
zm*B-1u|*oq-&<o=L?b)|=KNd?dyX(WBeux2vb8rP7RT&6)-tB*X7zl&3-J#a<1Y%k
z1Fl*cj#=cH5b_KeSd5U&p_QlJ4(z|_IX)MUc_Nxm@FtAdBD+?%ZfoYp&iMz-X&>ET
zAF-NXghUS4JlH6tbmfQElQ77agpf;k5+k-ulL|AR-w%>38EjhT$szM#Z;X)4k;2Jy
z1`W@(6a9j>c_y0A@HULtBFzg&ue+JKg*DFj=vrp2aNot|PmFGRIx!ZJrHR%YQ_L(D
z?z`Ch*NrYO?+{f?7fax4KDom_;cFNnk)umbUal>aTMaA2G_*l5H((}X#1>h-ulV@Z
zY{i1L#!K0J2?xUn$sEhKZs~`<hf`TvjXCC;Y+l3SpAlPTaQ;+z%a5#VbqZwe${}~r
zDU6WJ@eT<>){#B$S;GwL=xz4S=eyYa*TpZsFIZqQVxm=GV5!0q5ff#EL{2Qd<$f-1
z8EoAG3U^L4&tZ;dgk%zby&Lv(XYZ*%@fUPVv>Fnv;<U;~Y>|KdxjyAs|3J%cpl3D-
z<|g<JMr@gNx4%ChmlvlOG1h*6nP@J<7&2nZ9Q|&`?Fs_Ue&Vd7hc;%L{JsnE&(yow
zeoGtmg~f-5Un84q&~uEC%-KaP+q@o@A=r8stOPy?<_EkBBP4UK*!7y{>c6ZV;OYx2
z%r0lo<h$7HdIQ(LFI@tW!5FLaZ^-5i3?d^Wb0K{6%*bc?5YQQC1$`b%=3~tAjMy?y
z%MO_nQM#x#1HjJi3g%tx;20s9OZ)GPAKCQUH){l-{nKP~8V(N_A(_irzl&&J>^k;e
zW3AO%ylBQ_lo+vP#%?HbQ4UURDGc7{ofPs8!Bs{`CaG<nuE|yYTpr*WY%OAg<zSER
zLi{Tw%RL?by@ptPjCGHflKB!ojuBht-J7tRcYneDAjVqVtdh)CkYR*mu7>BB-MV4)
zk8n}A&1)&_HKqtgY?1ukh1xqVA|l!aYs=lDxf{yB2#MUf6mj*SCw^F3F$eVSt2@nC
zcQ7X;ayzPTv=Z<|z=NPJ-Y{<---XzBYV^MLy5Fk}uvgeece8tL-^FIvk6SkE{loHB
zx58hK70j_%OfzE3Tz^0Gmux?tv4$B2a#}P`Bfi0iE%WyK2Cpi1fxC^hw11;uZp2wF
zBP4UT^zTuf)^^3rS_B{S&K>p+-h~lcWO)7GVl#wI#m7Kd9t&ZQF?%p#i`;R?tO^)1
z5g!EgUnZN&aI((`iQMb!`BE}`bU}11zG<9bj>D|V2#MTp|8z(~t;-$pAo%3-LfCmc
zh!I=lAEW8Qs1TM^lE%ABR(2s^#t6wImoU3^9ysHir3o;!Z-V&^k7C4@DgXX@uClS2
ztnPzTu3}ax;JeuT76%(mJuo%QvJi}5E3;K0-^J#4tDf{`?xafAGtjj=1ak+T!H6x>
zb62;jDH9Nt8*9mI70j&|Mn*{HLC7B$F1}sa*EhlQe4TtszCc%MSzybN2a9TMFVX5v
zM6g%}12k2Ur(!S9h<%ftCC2^!<ulYN&RRiTmdwjotT93|kB!@{qVr_-O?c@4$wJ6v
zOn8jgGMA=iURyMM`&6!SSRSR-((1m8E4A;%urzv-zy5?-PuCgN7hzAkS2Fiv2geBA
z<k`pz$4#SUKda|pLEA-hyI6=5TV}!JvL9Q-6-g^*$2Wa(r}^T>s>Vlbkww*izV>la
zDNL+z%F9J_IqZ-TTjceSiYs3H1rvz17Etk$8IQ0HBP8?u@a*}=T8=Is;OYU#l3UIl
z=DQI4Mbd-&^U@7j>g)axZv~}7pzlKbmzSO{4l0qOr;ne}J5o?aVmZwS$-J6Xvt5Nb
z$Ld(Gf!;qSndk5tjF8OhZn-}9%#&uJF8*j)nHu&ENFGLPk%UaWD(`HEWpJD|4JC+X
z0(QQP*fO_94jLYKr<0|%a2!3&o&|gtoByC@p}nPRhX=TZTYfZ)oW<k25dTeFjvgWR
z)82=f;t}&jb3OtjjF8CN3Zdnq-K`MpkF|^^MKDv~GZ?XD#x9R(TVTnF09Um27*jSq
zzKhM@w{TFu`-8Bjin9jhU&;Iz`p5{$yt~@z{i$|gwR}CRc)Kax0(}?azhAIzT}Y=(
z82|BBEbWDCzCh5F5t2!1JiJfoDC1}72yD2vSv%Z!vDvM{tFP|7w-r1N_6&;za}i9G
z5nH5d!4fGq4&K5YFd@8>%vYE*7_mj3U2HaM<E(O_;?D?z9dKtq;Ev#gL_X*@PEGH+
z5do6X0*3ds8ul7N6h>^Bj&ojw4$8aV_r0cfp`<LtN}dst`Ix`+*CzF99SCrBMYm><
zGZ?;$%^vx+*_8v8u324z51J>K^Dt{OLNcGGt@w~G%Q=L2$5}ISoM^^jT401^K4&Vl
z;cb}@{R3PB@t8n4FoW-6vp=;u8a}GnVc(l(@U9b-b<lZ6NaoA9UlP126%g_sZ%wvK
zC37jZkc`+e2Qr-d?&~|uK5>>NeG$wr2pchC%OviYP<Uz(_G}X^XR$*xcfiaUA(^iS
zeh8Wt(s@vTt0!JJmz*oicd^+?8#-2+az(X#5(2XaW%D4`DU6WJx3FacMy>WBcpGc2
z_KyqZaVRw-w#>)kYj!tXSsU5{)9P#X4fkD$?dsm%+v8(1ctVbe6w>&xkfg_ig~Xp3
z5VAT?e{^vKBp3|-NcoI}dlb`fu;c1k`F6G6I)0-@;72wexC9P<6hsbrq&pDg5qGfg
zj*q$f*1S!PriJ*AT?Bl@^BL%IJU?@iIoc5nhIFLNeVIcX!C>%43Z}^%*9Zo~G}>qQ
z9Lxv?12XQQFvPavSxZ-3xkJ*tbCiHNN?w}HH19qoV4sqgBGsDro)U0R$xCNQ1bxIE
z8Sat3nE;PqFtnr1;4nup7{oE0BUe<8Z3M%@HB!O=K^pOA3(rUZ!_*wgh(B96Mw%c}
za|k24+QKjTyLy6bOpuFGXA_hWzh*GTy?g^tx5w`-?4p0Dtrq<{EeA2;*9^c&{Y%J2
zFc@l)(m2u;ik{YSQsFCs8rE!C-nnzpa!$i>SeEzhTy#>VA)1~g#hr@+?HaN*vV3so
z>dg~5H0+PFe01kJ#JA2Gc9>Z{xpN)in@`QVNQqpel=F|@VkL61v%N%#T;gmeC`A*L
zNWc6MrRWie@RJG6dsK-$s+9A$k13JIobBUE<Z);Fv#NbogJ<${GR^y0jr^>ZqjtIt
zOLpZ+)oq8W!8gkX+#2<%2Ah$i6-RrDB2RI)w<+>AXZxxmUv;)Os`5tF@%zK7eAwCk
zpvoWoZ32*Ds2FPW19$xCIBFCO1{I|kDM3ZSU^vl+JY)1b4iJ^oa9Aj*x`w$W_iHsD
z%{v;70rR|7^V10-ltIn&PR&oN4uX;52@FC?t|uVFxnByMz1~Tby|>hox70{-R<ifD
zTI#l1n%V&zWQsd5XeqrTK}*44cv73e@uXldFzFnAA{fL`q+l=z>3zPW58~KSFc@Hz
zY=;1&U@)*KIjUfe5(R@HMCsuK4F!Y2LMg&dcu+7bEGSuJD2IXK&kO=e&YQ5GU@*id
zIW`YRe1gHyp7gRF4)X+qVLa)(3ET+=19s9NWVpCS{k(Wnku<XRkh|a^ccB!%;*h;3
z-NjG3fAF_2xQkzKwlBJiTyz)p^ON0mlil?wOHuX?S8@$k!Uyn_Bdc9{xKhU7zNKcn
zrDmr=wgx@tZ8ev(UE#J`k=g`3#TPKVDS4qVjyHvKV^~wadt+@f#xM>##T^*Zl=^^>
zreIhoQ<_vWaeyiQY=KM3b!XzZQaq1=N&U~`;8Om;OdMQ_-!nuh4TiC3SN#bM_N9a3
z-&Kvf&OW%SM&4D+(Ki$H6n9{-Q_@z#n}T71P3hGLXNo^Fa48*85V#Z!hA92ams#03
zq7)2<Bc(pcUT*!CS0zuS`JdMd0=(i2EVL_i1p!_0XNGPiPneUVTftz+R+{E=a!4x}
z2Yhf=T6pA&`j+F`!`-<AOk<LoBIUfygGvN&#V;7DmF$yHtzcM?R+`3XO#sbdNLIh*
zur_HbL0a*924p1!AOTr%2MfgNUkJPvjJHAo`hDJt12*k_oaY|ic#fpbk+M@BA^}}-
zV}@>}EheE`!C;V9dSX5fX$6A;S;^H9kQEFI#7h01pW|2YX9ij&XGWk^Fc?&o2BZLo
zs)E4)s$?bwI6xH)hDs$%FVOCDDO3DsuWI?MdWYP<`hCzbcgbV!AN}o<?q5&38~WR4
z+{Ml~eb`xd$+J%R3+~z%+;#o(SKKA9IQj8P(Riget%(b8cr4zKL1XFF2@?wj1H{sF
z8^MvVU@-hE%_b2X{|W}fz7m5#uc^{-sH=Y--lCdY)NsNeXx?M0c}xu_@R{bls+w2T
zaB?P^_n~S&RKp2oL=S*2F_^4h8(EvQsc^Zgc}tDB)DhZrTT*XJIce=qKw7*rgU^!1
z6uv)eWv1T_bd_cAFeUdeC6C|fN2$4^)I7B~8cOy~P>l(yNx^g(>+PcN)nc>&Cmb%G
zz%aOEtVKBZ77T{GCABP4<?d*A#wI_&x5g_*ykgQ=$;d0HI#;bmDF;P3HWzo8@5ANV
zx8{IcFc=J%rf)I{bQr_l`a6tcae0ps78iG52wZa8#UqbYTR3}AD0*_Ntd5n#=}?uR
zx%dSG<kI9ufLt&b43`wCL}c&TG0mPF#jW0m>KifJCVrqz_+0#gfpW<JOK?yw7z~L^
zTd@)xi3<jU-_p@)2@byngJEvz;U$0h_m_1avjzO0$CtpnQW_s-*YD5P3C231AZZ{}
zJ5|`RR0SxVy(Gu<Vk9g;um5K!MB{{LZ$1gzi}$y1y|mtnEIOjz^?l3UV%suL@yt_-
z__qo370-NUdx7Fv;A}5cJPV!e1jUo!Y#&fO2b3cIZN)*wbI{p7q<9WF+lLj;VQ2e@
z;yL1MCo7(0r3h_HHSYt(^T64DsCXVa+m95_BWL@u;(6?BKUX}@{q0DOZASkvU^BVo
zNRDU*gTa`|x(LP$2E#7X0)w#2U@*io`RY;}u?z--EAQr~;H5ZD84QL_rgc>*j!p)H
zL6T{$LXc!I7!a8l1Vjdd!H?-y1V09YfsW~|2y_gF1vn<nAgnR|%mBt@A_Oo7gMo{w
zLn?4oF&GR=OcQE_dYKjsc&Ao?eXdl?u2gH0AFseM#`pyT7L%zHuo#S|K2S0BEje^B
z7?hZ1fXZiQl$zeOVvkTuk+peBs9%v6DG`g5lK%EmMO*6RuU53x&h`OCJD`O675AtT
zaa1YkZ=X=K6HfkFMLX+kCo5XA66#mfCrZQ<rKG?8LeXA0`R^6&y|cYU)t0EC)JK{(
zK@Ch$!~E@~s<zb0U#4oyob7dLz&bUfUw*fmez$7+r^fwi`u)!K88!VG6|h%^)7HEf
z)btme?N4gJCp9DO?+L7p5!mMgYSTnlg(I}VVDM}jnkpQZ4F&^a)4H0#*kCXqHr=W!
zM__}&VAr&rB-k|=3~WuhSB-<M#nc>RO{RtLmOFT#JL3Qj<*3F%*SG_NT9e8V)EW#1
zuqJ)4&H>h7FjzIc3{58RybMe2p949zn)__RR^tu~P))z29>N_Mn%aK{R#}Y%2vm(b
zFeo({FD)LxU^r^}1+8qrVCZR^!C}*2FswAK#OrXZG#CsgO;aBMq`{~JHK7%JT@H-K
zpBWmOR{V828X61+g{IX+U5<eUgCU<u5eWGV27^1(pw#1VXD}GZnKXnz&R{TPGmTb#
zj%)^l;hAZ@L3m~`7>JpcFa%--gCUpcI~uf@Fd}sJuX*6M7Rc%XIs7nBLutSv&G-ca
zGSjN40S7XJ!O+XJNNB*(%V03bGI?c!EQ4X;mC5;%ia{|LKH0Ar9Iwn(t0BiL<M$R;
znSM_&W&D|elj#i`ad0vi446#Q93hgyu+Ye~?{EBRO5yw4h8Fv;aj7N+$4rb`y*Cv)
z>y)gXl5^31ouJIPF~cv@t2E*GWiS|MnWl>-9JCAuLn_lqHQ`8QFc>_UW(9&LgTXM#
z^b2x&U@#;ySwS<7L<WQ5k7+;%e+&i#9n-|xy!7<+^HM$d^Z&fgYtGTjcmjhgQ-?I?
zkYz9!PMOSuaLUjI3#9B{;+>a`^Kubd8MfefW$X7otTI`43l35SgF%x?EeV<o2E!wh
z$+zTqWH1;AnKpARIS3gHhCH@^!I8&cFuXBYQ7eu&27`f&$z`|VAY(8rfHAq1R()o4
z{cXqgJ5Zz;wMvXyoAyHlI)=OjI3`smtTFy<;f!hTO%P-JnE{K*LA2$-Vsxg3D)x6~
zH=~=qTBg7qe$BAOG@Z1s`SXq;YZkS^+r_FmV%6{I?Fem*UoeO<`5iBZ7~^>syx9M|
zF`_X>w9hMk518C|!0vXI{oI$;`*N;hJiUt`$ha}X9aE1J?idUPHl{HmurU}6VN8<+
zA&kMW(8bhModz*rG5*Zp#AH34IGh*^h9M@m)Hy@uu<&CYGKm^;1@tfOEYElvfrjiG
z{i{0*CFT&=7<XV0V{$hHF$RO-ib+=qR}2ON6O+d$FfkYmLQEY@5MnSGZkQ&I9vn9e
z#vLDKm{hU{2MdG2Ai`u1w5I}t;e)T9S-(U0KHb@#92jgde4t=@IH7^TU=U#PBfWNc
zZbcRz`U}SMg`mC=vQtK7FOChyFBl-076bI!P!xs&_WOK}3g(7Rs9@ZIVS&j{&=C@j
zDHsr#yq)HqrIebbl<}{D_b8?IC}sR>+S^K*+e$fFqxR;AVLZV?3zKQmJdQVH=wSaF
za<nji!#*4>jNdbiFu8V8Io!d*2m60-waAb4O?jGKUtOgdx^9S|4iU1C<v2o$+~Kwi
zHcaEzkHdz+U|3;N8o~;L!2rUf+5`{=gMov|cqlpp27?4sAJPKS0|WQrfXO9j2)u<J
za_68xnbj_Q$X(XoKI{%V?9T0PA9IHubIKocmpbOI;^&`qhn{ruPr6H-bXV~66Ww8n
z&ikElhn{iDpK+Hs<CH(^4m<0Vzu*qN;FQ1ME^)yrf6*Ow(J6n?UF@R!XaDn)+@VR%
z-(PWOy5i2`=U;J`y5jtvWOrb)+a#Z=d6V4*lij8L?Fa6F2X4dPe(Wys*j<5k%LK5-
zh%i(&*&d;)!C+Ww8r(q~OAQ7CP?LcS;s9zeEbuffat3kSH2%!M(&VKGEDZ)jNRu8%
z1zfD#cu@Z(P|@R(dK^Jzu7L!dhJstTX}`BmmW*VnFsXPHhfw3j3^z?~gK*PeFtju=
z2rUf;!$_0E9L_P)U@%}bz2k5Wj0S_Dp=lZ*G&BND78Kh5jz<LJh+xlwn)h_7u+ym`
zj`JMh5ga9r`&kHS>b#LJDjJcAF=0q7IFc&+kyLreiE2nB&YqYm5B=Vn&vV51XG}s>
z<9-&Tn&xvtQ{&GJLru0kiesq3c<#eb)9R0a(_k>TG<D6GYSOT$ao<W#!#q_;FJ#b3
zk>)L|7d9vxLGu>Tix{+w(7Z+Uq6P)@HE%J!m~n*Xg=^m8dU1n}!Zhy>`VR)pshYQh
zUcw-hq~<NDmoy09ta&5#NP`Y%G}Je&WYDHc^H$a?8(TTHMe|nCs~E{VB~tTN)vFpS
zIBrbyR@18)<eD{ab-lW=pQrI_-WqxhgW#_kDiYQ-=-@^3*3xSkM|keL=B=&QHYi_9
z^ETI;8?*`0ye;$=##WBQ(Y!77mIlQQHE%1um66PoF*WaR`fmnBjx=v;y|qCnx|+9*
z-o{AedD@z{t=`t4vDUop^mYcV<TP)4y}d!{ADXwj-rb;tT+Q1<?_tnsq~`6Z_cUJf
zq&v;qOYdcnF>Btx^}h|$PR-j}?`=@tyXNhq_b~_~qIvu3eT{ECy-4%+)B72;@6o*d
z_5Q{Lj;q!%Z|S21gC;T!TNHh?V9=gLgFn#62nIO_=(eE82nHD=R)>NfD;P^Swi7XK
zK_4p^cX&Rrh6tG+Cm0km(~y6wj}r_^&DC%urH>a3I-}A6GuPt<gAnK%z{2`$!I;Bw
z#5mg$^f`h-2YMP1Ir?0|AhZsGVuC(TFh26cH4SloeZF8&;2%3&L0=#klvkwTh(TWn
zKCk_Wq!U43Bp4|?kxv7zN?$A(gt^jyBGs4RdpW>V1Li^BAQ%LP(2(n_Z^ZZWJZKG>
z|N17uNaUyh4Vx`}vtZEfOGCI*-y#?k|J0DIt8W#IWS+>Y0raSE6O5G{5UBxws&5yJ
z1P)ErkN~XjfILqQ)(|DtcM1lDO%X&F^kagth64}b#s&SjU{HD*62AofgkY#Vl}Q7r
zT0e>Y=74MXHbGBB|8j&9Hoby=N-$^(2Hz>@rv-zSxf-0deg^u%YhGxGz368JqYe+o
zXz*P6Il&-TrXk5eza<!y^?=hRLBA~+6m!wAqSfyR1}$+lKnC=?f^mVv36R7k==TJJ
zPGzyF5cK<kLEuo#!GfMF7!><JG+odi2nGScaXu#K4+Vp2<7?Q|>5l}XCEw*CsY%e^
z3kGG(;51m!Q*iFeb5@bmB<LRmgOC<TY7+F1f<bvANNN)FPl7>7D@Y;|^v{Apfj{gk
z1pSL(eBqgk*jEVpSHYm{7Y#||`ZtU(PcYYj<J1R>20=Vwf1(~M8WdQE{fYWm(V%%3
z_9yCbqCrWTus=~BCmOVzhy98Ac+ptK(Z#SoQI8i5f>~$)!s-)5V>JiKY6!OL6Gelf
z?;659`XteyxQ~X}T%Rl&G`++AM18(!P#Q4oPt+HP2Bnw5{zQGDh#7=cz<~XU`XbSw
zfH~|>)EA2eC4#{IM16^9&|U~ZV^L2Kjn_PtS;JA9zEm{mY)u2IMqefxi5zjGA)#Ae
zE*j5xt~cyY)VGKR<<P+XM18Ah(4G_aC+gcogO<^-KT+R~-}97n*q^BH5RET9s}%Mp
z>O0Y295AdQtf21_4T>{p0CwoRMdK>Rd}zoL(f5c3rHyDf4%hdJ2Az+?{zN@dG-xXX
z`xEt3qCp`P*q^AM7L7X`n*sY1^)u)%p7sL!6ZNyAK`TYrpQxY1ff)xoX-Ln~&x;1-
zYau8u>K8<VQbaUl4(S&~gHnk!An^1{qCwz1*q^B16O9CpgNOZz`hC%$<3-q?s3(gC
z)y9DRiTVTR8%Mmu{zUyD`iF;GV1J_iNHi!$t07*aKL(%ImeYW{)1Qb2LGm=j3iPL<
zLBRtJNI3nOXb?ya_9yC}pr1UF3HuZE&*(oMn}Yp``WMld#7ls{{zUz&Xiz#A>`&Cc
zi3U}Qf&EGPV96k0Id*Q6K14DIys9A#sYgi$#U(W$F!iAlVuTExsR6aD50i`y9MBB=
zll1YDK|6KWpQOi2#xo8yfc;7O1j(R)E9_6wCrZW+4yc9wN%|zopj1EDpQKNg3_8Wt
zkOQy(D;aNje!PYgWqpcd&=Io+^rt=*@5hm$8UknfG|8Yy8thNf7fS}68p8f0eF?sg
zhgV^LlAeI?;XrTLpQJC949eDq{Ym;V$)L;}*q@{?mkip8XeivEuaJx-yr=^X&Ln-M
zWKcXs1Dr}<B^g&Z)Jj7hiN0DgD6s_gC+RyRV;Qe}0Q-~losvQIHDG^|zDqLb1P1me
z>ANL^qA##NN#7$Glz<HTlk~lkL9i2?sz~}i$)Lj^4Qq9Mzhuzj9l4*9egNOc^N2Kr
z@AZR{LGgUppQN9aj9na*4EvMxbCN*_4%nZhpO*}(?gRUi^b3+fnIW)0Nxz8SbHp<2
zPtq?*1|=wJz})DUCF23d<7j~4>PZ++j?mRmTSmVk8B})$`5}^iRWhie2kcMMA4&$5
zEP?$=`XkAp2sG?Z(jQ9(LHuBUlKw<8=u8;)C+SZmgU&Hwf0F)8GAJoc!xlz=j{f3%
z8U%?X{e@&K;x$P$K<f0D7+(&z(*T9gUr9y^M@hi`B>kIY(0Lf_Pu2&^2A$2q{$zcK
zY)}?7>`&ICP?3TsOTzwSeW+|ua2EC_>%(M&4!boJn$d^L#vERRMnl4rK0-Dq14=_;
zs6J9QC{q-<y|NxH8-$&Q{mJ?y*`VYt*q^LVmJK?Df&Iz)zp_E`8Q7n!Pmv7@<iq}C
zeX4BGQ2^{u)~Cq^K_N91z0#-41{K885CPC<$OeT3G-Ph+Gx5Ise^h_~n5@r|4Jv*G
z`;+x$vT>KA<zRoZzFamo@S+Q_KUrTP8&uW?_9yErWrOlYV1Kf{N;c>i3ic=Kt7U^y
z{52p>^)<3VXjBbBXnn10P*hsOnYzACHt3LDLm?x5y=+j8BiNs;@4@}~X%_5H*7wQ=
zMVVoLvc69?s1^?FPuBO#1{FAh{mJ?P*`U}7>`&GY%Eng?(Ljt!)(^?XcwQMxLk6&Z
z7~{!Pg0YvE^&_%DaZe<P%lc8-pdc*lPu4HW2Axa7{$%}<Y*5Sr_9yF?WrNbzV1Kfn
zBpdO(JQnOv*00D0)zX0d$@*2<pwP7jth0ViHV6l;A!k9qE*q4@fO-nDenU3SaTu<K
z$}##)*`T5_us>OUDjS3af&Iz)GxQg)3k3U<_2;rdC$q3WS$`oLgmr@b$@)v#ILfiw
zus>OUh4JSF|B?4C>#t>l(v^{TB<pVgK;WgjkOwa7Z)Jlr0FehS>+fWPs%~L@;MVW`
zXUt1~>^3Oo1$$Pq4pwp##sc=MWQ|gCQ*;u1HS1tCHy!`NUe&C_)!cLjkN#E*C#fa<
z`EE&S`6RWPzkN$pZ>a(P_H9+Y?QAEj#go<2bc&4iidr*S{gv8ecL=$#>T?u&8^f#5
zA^v4mp7Y0+7Gm;|>?Ec!k1AaF!`uDyf6YuIgT;ebjXHV|tvtYBh3V)NZ7eTL2L`J~
zM-eoNB7<R7qNDg#+`SRi`Yp>9j!<3$wLx||m7>@Uq^-Ji6u71ZIsUxRSB{RN_B2z0
z!HUmON+FpT7_8zPg)iwq3SGs@%=x3?yzCs07Sg64cVNZlD8fh91O}@xM-jewUSSTm
zwyMkdM`5I3j1(d$3OS)-&#6T__n(JPFQ=BvzJ{0Bnb7v@=7$Hj=fI!Wit1W1HyzKD
zcH?=h93359P}l?vs}>z~EgdJ|&#VR=#nQ<Pz+ffk=ti{O1cTL?qfsUFOj<xQd`0Hy
zR`C2vsTt0rja^EqUCyJ8n@X9R{-X_eYBlF-HQb*bvt7-(-I+OhQq6hNnK^nx&3VI_
zIr>D+`NW?&ij}ik{Ignu(qpW4)6Z&iYSRiIU&Kn@(eWKEJ-}df?nq&#hyLjK`_S{;
z=>Ef^dRWXsnYA=I;1{gy9cd87{lKtF-qCT*jG6gv57^tQ#D6`zr&$BPXSMK1Md>5J
zV1@9g*Jr7hVhbMrJsR)0QBpTb*$Hezz!zv9D~LzGK*T`J^Ik1N*{oC`4}w;8Jim&b
zmW<Q>Ksg1t@CK|r9-Tqa$^;Bn6OWFSDMSMXD}+bUU^?N!6It~;zrFElc-$^12!cDX
z8hCVWNYPy|tO9uC6KIiuKeK{&G%mDy1cQ~iqlh&v!@*#+?I^}g^Enu-pdF0_O<Z8G
zDt2Tsv|<E<)vlvwEaJ86@VV(wxX$0f5k`a^<EQm5e$DFEkrL9$AQ-G@9gQPJG{Imc
z>c|-_;U((8VD;(fd~C_QK@qpBWZnr~*($4B<y_l%5hmK);1{e+9UZ$+6cG$ooQ?*a
zFmuTNW+mz9t+3uxO3ie7f!#`}-A*rXODS{9=>^uPIoJ5TKtiK4Wz%fxQ~-w<ixkfy
zrGWpW^r+%FsuZC9z(QZ~JW&deYomZ8UX@kAqjVA4`Gdjg-jUtY5(Nxa?T(x_?ft=E
zCGPBB@E{Zztgao+XtcWngH^L5HKJ;EV6Yl?q%F()Y)e(A%e+c?upw1EOO-7C?_8#M
zmO0zY70+^Kdxhdz;cTx|JS&~;Rf=bov%OmJtai57D4sRW_FBcW*4bXCc-A@F>lM#>
zXM2O<+2CxSP&_A;EdFshsd!F0+lh)N(b+zwcuqOnrxnj>XZwueIpb`fRXk^%?Q@Ff
zoU?si@tk+IFDRZ1&h|ycbJ5wpq<Ahl+b<N)3ndHb5f+?^=cTj#O7XmMwqGlr*Ut7E
z#q-A5eye!iI@|9Q&pT)Pz2bTAY^Nxm6leQ`;`!ige^fjlo$XJG=aaMjS@C>!w!bKz
zFV6N-)w5L1;-6-hsh(y2_HtfK5#Grvr0DlfcM!A_?G*~uRD_sSPth;-K{P&y_HvSr
z$e{05Sw)%v3EqHw2v$eY?=BXq#zHlh|Lr!b=4R(2;HYXI^)CYGm;v`^#TLoRS6-d5
zZ%(n@o#5LRt3NDO8`2#q(2ifQ(u*{QP=pK&R&SA3I21|-gB4jME2o$#7_71)O=M*4
z=p0s6(LbTFHm&pMlmmBQMHOjEq!k|+RzXFY(AMx;iug0Dph%xf;dL-r-9$<VqL09<
zuwse+SK-AI`Ma#oow~ue6LTNJ(Z;G(V$~XyKZS_Anlnz#<8O~sYmHOuQG5Nuo2_RQ
zJsOFg9xbV(rR<bfNLxs}6RW{U{*-Dkf?*Y4q>rPJ9RAD-F4C++dw(!kl|`DyDJTgB
ztF1_1yWw%a4W7}jmj2J<0x3@h!xQH#0O=nd)+SBaR3NHkjf{2wL@y@^YLbwHs<RQI
z27SQlLek{4Y3tPnX<K%CgJ0|y)crzsf>%<(NIcW37U?(L8-j7e4`<ozy?MMsnxnV!
zYu++)neXUDG;e*mzN6Ja?%-$*k_S0jlLdLQ;Ak%w<mJxxUP0dLY+n-OOV0K)L4M|J
z4-@5KqT_zkM0uLCy;_u4JKG0E`Jl6XRg|wf+pk3Vm9rf!$<dPI{<9=`mb1NHlGi)i
zM<w~Fvwc&NZ#vuWB>A1Q9V5#zvg7{qWO<&my;+twJKHB^`J}UbSC;QO+aG25qqBX$
zo$i3!;~%NPiac0xJTFR-qnzz%C0(@Qke{NYo8oM5Q_^j7wy!Gbt~%R;Re7-Lxc_if
z9`0;Mt8%onJy(_II@|MAdA_r~QBAi|b-d4EHQixnJ4uz3obt)4oa}6WP}6<z%hP%r
zdd|w7`fZDsKjl3|<xe4K)jstbM1pK2__KH^{Q{O|l|ZGm7|IaApRLNLG{w;YJpRlo
zozhZyTlke~6>fY8fnAOe)e&M=O6J^FI$O})=y$PTOi<McYA|_NT9)HhtRO1+KMLxC
z!OEc0R%`p=^G#NidVTMI-v1G_5MRKGw9*&Qtc)>Xm0A6R$1An+!J~sE+<}#6C6`Jk
zTwt)8tTZ{0gT$>_Ay)sbd0|$5YdQnL9auS5+Jn*I1{kacD={dm0}NJpl?Ht`FT4r{
zE4NB>8|}rA^2KVc`d^0CQ6d4XL)C0A)dE!Uf)X5X$3$QGRa!bwP7N4V%~cBd(19NQ
z%xbJschE^17_6`=bs!A|`iWIj^}iaguFCs~&TeoAR#TO_gKDaR!78cJmUJJlq>5Wx
zWmNqZw@EZM`Js_?Xp7&oI;xZ|xS!Wi1%p*jr7@+nD=@6esbq|l;)_4CN~xq?2bv7Y
zR<CqQJZx^Ipsp0MQ+*jauf;D|;Z%A8Re!;&vvR5arRyD~%pLzu_RyKr$BsREv-`gW
zd1xlbqgZiO8c9O6fWa!NQizofi>=S`l~bjcqg*mDtQxBH`IIJyKU<|!X~9K>Yw%}Q
zE0sQll7qlt6;f%LNl|w&SXorszS4037_1g5t@S8b9t>6hl^Dl^c8*%JKj<@heu|(@
z5wcSCJ<1TnFRZGlq`#D`h(EJ3s1=K+1dI-Jm88r7Fj(zVQj)~d;REW`HYWeiJ3;dE
z_ySg%mBxXRu)tvTSZN$6I|K|?gq4Ebl-vh~Re+V|C_3%JpIO0Gnn5TH2n<$bmG%u(
zWff{}m00y#=og%oOEx%yj(PBVR$`Tg_uRUQOZp$1a}mS)Syn&GxhT7y4rcKSR&bT3
zY08iSgH>528>OQ~Fj#3-GJ{KRX4Ly*%IJ~*)xsz_2zOwGS7`@FH8Q|p)mG_csM;z%
z->S6g|NNPPG1I?Brc+@2o>f|<-_!9w7^{3$R>|K}c~vl2QC0c{CD?$$N~zLj^5&lU
z8J{*g`vN_AT~x1&IS93Rt7KyN_scT~|MU7ym=o;wStVG~Z|m?XG+E6|hgUcZRE=k9
zPG`H^Gqp0csbnnr$f_4h+D)0M__I|fmOKWfbmGsfE-cMHlzRjQD+Ws*gOVV@U=?7=
z-t3ZL&=X(zSN|7GS4vHHhBtRArFJ^Qn>Um)H~f<V(zDc@+nn)?6Kc*A{&)t)Sj~Ce
zd8qMN&H31QsPRlK@Jua8J!17(<TJIDvt8wxTFu%1^_kj`+IM-KT6`g^PD@5c=j~vy
znzZC0DDf2xR*9B+@!q}mf2?UZ>Kg(EW7QmEodJS4wPu_@K!6lJHRm{gdO4M_#T{AU
zTAJ@Dqz<EF)oT3<4qm;MFE}W<5O-iDYsqU-el{3Zxmue4C^;H`wyM;U-cSxI{=CCi
zoR)^0l3BrErD%ykrDy?Wwkpv2-*2B=-RBM>BnriD@q1Q<mehr^Ucg|bXKAHGlOGtY
z-YoSpd7jK^e0661PmBnP2;<%1(!3RdDmYqAgPJ;8-GaI~S|f#^k%FT=QwW;rY_Ai7
z);ZfpgrFnN_6;HEhO_-v2zu*mj}e2$h>rWu6@%tF+ndCoP0scSG3bP|eMbzs<7|Hr
zgFZOhaZ*s6<hcJrDQKayy-f<*=4_vqf=)Zz$x=|Vv;9R1`r>R)kb@@3j{7IbK?%<G
zE;(qIvwdC;I`3>hmV+KU+s9G`97~nK54|{@D&Tag478e|v@MJ~D-leg!>4_tmhE^t
zrzyOCC$&>{LK8pz)I6o;^bsY}BF|mb#;ECO?O?SFk5Nnd+c9cjjOuZ=i^r&?{q0ya
zFxJVBRg1?u`D4|<u}*%R8Wg8y@c%weEf%Mi^0&vSLF1hKacc2#{`*n>DL#~yJErN4
z0E(C#SaoBv6)RIBG)Xo5nGpY3h76h2|51jFGGy@tRs)&VM9=nzKi>4GX9c=)hN#XE
z!zmGgau)FmRuP$YK2#AIj>sy9Og4zzGTC_S4+xNZftXbl*>9Uxi)@n&8A7ZkvR{nX
zNam8SW^CD2aCNqv|MSy1r2?YcSOsTtVy^=qJ{dZywgpN5|5r7lqKoJgR^Zt`=I_<w
z@715^ESj>U@hDd1nZ}d=Ctz3=o@vT|&5JwZ##Uix|G;gNjBWlT7fSBNuUTbhx-DhN
zgTcx<)5P+YmvhFAty0eZ8^=jTod3p@wTj=fQqH7N@BTSBF!a-&fdASx??zPJob%|~
zxBq7_J9P=(*eV?D*E*|3%1K#OcwbgFn)Y+=KdoBsdDO*<MJ-W1OVoUHd_wpn{K6_G
zO(sC_6a1M~kEXx}<wb(QYDUv~mq3ZA*!!P>4yxFh%}TW@J6NgWPqiAZ1P)ii{Ha#M
zmE6OX!p?Tt;Yt;MJ6_2gujHZHTGk2jPm7dlG&|#DNcm~8v%N(5X^FF)p!}GiRPf6m
zQGPt4RPdi{9#wuis#Np0k10PLbGDBwKOJ|rpD2+}l=A+Q#ivT-Q)l~`68X&8ey$XK
zu9Wf1N2%XOsYRWsW#xveRsE@D6V%KToa?a3YT#rw%<rP7tAW#<?YV08xoQJCjI-+5
zWKK}?&`?|L$ON^#zr9n{cdA*abOTcRRsEEj#b2N1mYU<1TEOqHZ>!<A)k4m8rQ2#9
zYGViFs{-yf{?BUq&#FneEmp$LPoLFl)TS*IG@8`}_v<vT5zf{6!}&qQPQO}X<u%Mz
z3(Zw4(HBzjaNNPF9PYQ`xuP*wEJjvASS;u)s~k?sOrR7n;(S%ZX&*`D!f|7(R=EGh
z6^sgYN(U_>A;c<$``@2e4(BgGx#)NdD;G{*PM{kwzW7RolY1gG2N<j{IL)q97#z>E
zs)GApY?o~8^3P<HeTv_+s^GK~C)gDjtQ0t{)d{NyhE)fgyiND^om*R9(W86Y9@ZhS
g=zY8FP}6&_HusSJ=<vI%1OGyh-~IpjhyOqS9{|%P&j0`b

literal 0
HcmV?d00001

diff --git a/.clang-format b/.clang-format
old mode 100644
new mode 100755
diff --git a/.clang-tidy b/.clang-tidy
old mode 100644
new mode 100755
diff --git a/.github/CODEOWNERS b/.github/CODEOWNERS
old mode 100644
new mode 100755
diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml
old mode 100644
new mode 100755
diff --git a/.github/ISSUE_TEMPLATE/issue_report.yml b/.github/ISSUE_TEMPLATE/issue_report.yml
old mode 100644
new mode 100755
diff --git a/.github/dependabot.yml b/.github/dependabot.yml
old mode 100644
new mode 100755
diff --git a/.gitignore b/.gitignore
old mode 100644
new mode 100755
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
old mode 100644
new mode 100755
diff --git a/.readthedocs.yaml b/.readthedocs.yaml
old mode 100644
new mode 100755
diff --git a/CHANGELOG.md b/CHANGELOG.md
old mode 100644
new mode 100755
diff --git a/CITATION.cff b/CITATION.cff
old mode 100644
new mode 100755
diff --git a/CMakeLists.txt b/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
old mode 100644
new mode 100755
diff --git a/Config.cmake.in b/Config.cmake.in
old mode 100644
new mode 100755
diff --git a/Dockerfile b/Dockerfile
old mode 100644
new mode 100755
diff --git a/Jenkinsfile b/Jenkinsfile
old mode 100644
new mode 100755
diff --git a/LICENSE b/LICENSE
old mode 100644
new mode 100755
diff --git a/README.md b/README.md
old mode 100644
new mode 100755
diff --git a/client_example/01_gemm/CMakeLists.txt b/client_example/01_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/01_gemm/gemm.cpp b/client_example/01_gemm/gemm.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
old mode 100644
new mode 100755
diff --git a/client_example/03_gemm_layernorm/CMakeLists.txt b/client_example/03_gemm_layernorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp b/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
old mode 100644
new mode 100755
diff --git a/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp b/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/CMakeLists.txt b/client_example/04_contraction/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/contraction_bilinear_fp32.cpp b/client_example/04_contraction/contraction_bilinear_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/contraction_bilinear_fp64.cpp b/client_example/04_contraction/contraction_bilinear_fp64.cpp
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp b/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/contraction_scale_fp32.cpp b/client_example/04_contraction/contraction_scale_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/04_contraction/contraction_scale_fp64.cpp b/client_example/04_contraction/contraction_scale_fp64.cpp
old mode 100644
new mode 100755
diff --git a/client_example/05_layernorm/CMakeLists.txt b/client_example/05_layernorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/05_layernorm/layernorm2d_bwd_data.cpp b/client_example/05_layernorm/layernorm2d_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/client_example/05_layernorm/layernorm2d_bwd_gamma_beta.cpp b/client_example/05_layernorm/layernorm2d_bwd_gamma_beta.cpp
old mode 100644
new mode 100755
diff --git a/client_example/05_layernorm/layernorm2d_fwd.cpp b/client_example/05_layernorm/layernorm2d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/05_layernorm/layernorm4d_fwd.cpp b/client_example/05_layernorm/layernorm4d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/06_softmax/CMakeLists.txt b/client_example/06_softmax/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/06_softmax/softmax4d.cpp b/client_example/06_softmax/softmax4d.cpp
old mode 100644
new mode 100755
diff --git a/client_example/07_grouped_convnd_fwd/CMakeLists.txt b/client_example/07_grouped_convnd_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/08_fused_attention/CMakeLists.txt b/client_example/08_fused_attention/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/08_fused_attention/fused_attention.cpp b/client_example/08_fused_attention/fused_attention.cpp
old mode 100644
new mode 100755
diff --git a/client_example/08_fused_attention/fused_attention_bias.cpp b/client_example/08_fused_attention/fused_attention_bias.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/CMakeLists.txt b/client_example/09_quantization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/09_quantization/gemm_quantization.cpp b/client_example/09_quantization/gemm_quantization.cpp
old mode 100644
new mode 100755
diff --git a/client_example/10_grouped_convnd_bwd_data/CMakeLists.txt b/client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt b/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/common.hpp b/client_example/11_grouped_conv_bwd_weight/common.hpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/12_elementwise_normalization/CMakeLists.txt b/client_example/12_elementwise_normalization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp b/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
old mode 100644
new mode 100755
diff --git a/client_example/13_batchnorm/CMakeLists.txt b/client_example/13_batchnorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp b/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/client_example/14_instance_id/CMakeLists.txt b/client_example/14_instance_id/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
old mode 100644
new mode 100755
diff --git a/client_example/15_convnd_bwd_data/CMakeLists.txt b/client_example/15_convnd_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/15_convnd_bwd_data/common.hpp b/client_example/15_convnd_bwd_data/common.hpp
old mode 100644
new mode 100755
diff --git a/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/15_gemm_add_multiply/CMakeLists.txt b/client_example/15_gemm_add_multiply/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp b/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
old mode 100644
new mode 100755
diff --git a/client_example/15_reduce/CMakeLists.txt b/client_example/15_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/15_reduce/reduce_nhwc_c.cpp b/client_example/15_reduce/reduce_nhwc_c.cpp
old mode 100644
new mode 100755
diff --git a/client_example/16_convnd_fwd/CMakeLists.txt b/client_example/16_convnd_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/16_convnd_fwd/common.hpp b/client_example/16_convnd_fwd/common.hpp
old mode 100644
new mode 100755
diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt b/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp b/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/client_example/18_groupnorm/CMakeLists.txt b/client_example/18_groupnorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/18_groupnorm/groupnorm_bwd_data.cpp b/client_example/18_groupnorm/groupnorm_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/client_example/18_groupnorm/groupnorm_bwd_gamma_beta.cpp b/client_example/18_groupnorm/groupnorm_bwd_gamma_beta.cpp
old mode 100644
new mode 100755
diff --git a/client_example/18_groupnorm/groupnorm_swish_fwd.cpp b/client_example/18_groupnorm/groupnorm_swish_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/19_pool/CMakeLists.txt b/client_example/19_pool/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/19_pool/avg_pool3d_bwd.cpp b/client_example/19_pool/avg_pool3d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/19_pool/avg_pool3d_fwd.cpp b/client_example/19_pool/avg_pool3d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/19_pool/max_pool2d_bwd.cpp b/client_example/19_pool/max_pool2d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/19_pool/max_pool2d_fwd.cpp b/client_example/19_pool/max_pool2d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/client_example/20_splitk_gemm/CMakeLists.txt b/client_example/20_splitk_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp b/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/21_grouped_gemm_bias/CMakeLists.txt b/client_example/21_grouped_gemm_bias/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp b/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_grouped_gemm/CMakeLists.txt b/client_example/22_grouped_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_bf16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_im2col_col2im/CMakeLists.txt b/client_example/22_im2col_col2im/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/22_im2col_col2im/column_to_image.cpp b/client_example/22_im2col_col2im/column_to_image.cpp
old mode 100644
new mode 100755
diff --git a/client_example/22_im2col_col2im/image_to_column.cpp b/client_example/22_im2col_col2im/image_to_column.cpp
old mode 100644
new mode 100755
diff --git a/client_example/23_elementwise_transpose/CMakeLists.txt b/client_example/23_elementwise_transpose/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp b/client_example/23_elementwise_transpose/elementwise_transpose_3d.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/CMakeLists.txt b/client_example/24_grouped_conv_activation/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_bilinear/grouped_conv_bwd_data_bilinear_residual_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_bwd_data_scale/grouped_conv_bwd_data_scale_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_bilinear/grouped_conv_fwd_bilinear_residual_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scale/grouped_conv_fwd_scale_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab.inc b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab.inc
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_bf16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_ab/grouped_conv_fwd_scaleadd_ab_int8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu.inc b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu.inc
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_bf16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_fp32.cpp
old mode 100644
new mode 100755
diff --git a/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp b/client_example/24_grouped_conv_activation/grouped_convnd_fwd_scaleadd_scaleadd_relu/grouped_conv_fwd_scaleadd_scaleadd_relu_int8.cpp
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/CMakeLists.txt b/client_example/25_wrapper/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/README.md b/client_example/25_wrapper/README.md
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/tensor_transform_using_wrapper.cpp b/client_example/25_wrapper/tensor_transform_using_wrapper.cpp
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/wrapper_basic_gemm.cpp b/client_example/25_wrapper/wrapper_basic_gemm.cpp
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/wrapper_img2col.cpp b/client_example/25_wrapper/wrapper_img2col.cpp
old mode 100644
new mode 100755
diff --git a/client_example/25_wrapper/wrapper_optimized_gemm.cpp b/client_example/25_wrapper/wrapper_optimized_gemm.cpp
old mode 100644
new mode 100755
diff --git a/client_example/CMakeLists.txt b/client_example/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/client_example/README.md b/client_example/README.md
old mode 100644
new mode 100755
diff --git a/cmake/Analyzers.cmake b/cmake/Analyzers.cmake
old mode 100644
new mode 100755
diff --git a/cmake/ClangTidy.cmake b/cmake/ClangTidy.cmake
old mode 100644
new mode 100755
diff --git a/cmake/CppCheck.cmake b/cmake/CppCheck.cmake
old mode 100644
new mode 100755
diff --git a/cmake/DoxygenDoc.cmake b/cmake/DoxygenDoc.cmake
old mode 100644
new mode 100755
diff --git a/cmake/Embed.cmake b/cmake/Embed.cmake
old mode 100644
new mode 100755
diff --git a/cmake/EnableCompilerWarnings.cmake b/cmake/EnableCompilerWarnings.cmake
old mode 100644
new mode 100755
diff --git a/cmake/TargetFlags.cmake b/cmake/TargetFlags.cmake
old mode 100644
new mode 100755
diff --git a/cmake/getopt.cmake b/cmake/getopt.cmake
old mode 100644
new mode 100755
diff --git a/cmake/gtest.cmake b/cmake/gtest.cmake
old mode 100644
new mode 100755
diff --git a/codegen/CMakeLists.txt b/codegen/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/codegen/driver/main.cpp b/codegen/driver/main.cpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/device_gemm_multiple_d.hpp b/codegen/include/ck/host/device_gemm_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/device_gemm_multiple_d/operation.hpp b/codegen/include/ck/host/device_gemm_multiple_d/operation.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/device_gemm_multiple_d/problem.hpp b/codegen/include/ck/host/device_gemm_multiple_d/problem.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/headers.hpp b/codegen/include/ck/host/headers.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/operation/gemm.hpp b/codegen/include/ck/host/operation/gemm.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/stringutils.hpp b/codegen/include/ck/host/stringutils.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/types.hpp b/codegen/include/ck/host/types.hpp
old mode 100644
new mode 100755
diff --git a/codegen/include/ck/host/utils.hpp b/codegen/include/ck/host/utils.hpp
old mode 100644
new mode 100755
diff --git a/codegen/src/device_gemm_multiple_d.cpp b/codegen/src/device_gemm_multiple_d.cpp
old mode 100644
new mode 100755
diff --git a/codegen/src/device_gemm_multiple_d_operation_xdl_cshuffle.cpp b/codegen/src/device_gemm_multiple_d_operation_xdl_cshuffle.cpp
old mode 100644
new mode 100755
diff --git a/codegen/src/headers.cpp b/codegen/src/headers.cpp
old mode 100644
new mode 100755
diff --git a/codegen/src/types.cpp b/codegen/src/types.cpp
old mode 100644
new mode 100755
diff --git a/codegen/src/utils.cpp b/codegen/src/utils.cpp
old mode 100644
new mode 100755
diff --git a/codegen/test/CMakeLists.txt b/codegen/test/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/codegen/test/gemm_multiple_d.cpp b/codegen/test/gemm_multiple_d.cpp
old mode 100644
new mode 100755
diff --git a/codegen/test/include/test.hpp b/codegen/test/include/test.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/CMakeLists.txt b/codegen/test/rtc/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/include/rtc/compile_kernel.hpp b/codegen/test/rtc/include/rtc/compile_kernel.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/include/rtc/hip.hpp b/codegen/test/rtc/include/rtc/hip.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/include/rtc/kernel.hpp b/codegen/test/rtc/include/rtc/kernel.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/include/rtc/manage_ptr.hpp b/codegen/test/rtc/include/rtc/manage_ptr.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/include/rtc/tmp_dir.hpp b/codegen/test/rtc/include/rtc/tmp_dir.hpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/src/compile_kernel.cpp b/codegen/test/rtc/src/compile_kernel.cpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/src/hip.cpp b/codegen/test/rtc/src/hip.cpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/src/kernel.cpp b/codegen/test/rtc/src/kernel.cpp
old mode 100644
new mode 100755
diff --git a/codegen/test/rtc/src/tmp_dir.cpp b/codegen/test/rtc/src/tmp_dir.cpp
old mode 100644
new mode 100755
diff --git a/dev-requirements.txt b/dev-requirements.txt
old mode 100644
new mode 100755
diff --git a/docs/Contributors_Guide.rst b/docs/Contributors_Guide.rst
old mode 100644
new mode 100755
diff --git a/docs/conceptual/what-is-ck.rst b/docs/conceptual/what-is-ck.rst
old mode 100644
new mode 100755
diff --git a/docs/conf.py b/docs/conf.py
old mode 100644
new mode 100755
diff --git a/docs/data/ck_component.png b/docs/data/ck_component.png
old mode 100644
new mode 100755
diff --git a/docs/data/ck_layer.png b/docs/data/ck_layer.png
old mode 100644
new mode 100755
diff --git a/docs/doxygen/Doxyfile b/docs/doxygen/Doxyfile
old mode 100644
new mode 100755
diff --git a/docs/index.rst b/docs/index.rst
old mode 100644
new mode 100755
diff --git a/docs/install/dockerhub.rst b/docs/install/dockerhub.rst
old mode 100644
new mode 100755
diff --git a/docs/license.rst b/docs/license.rst
old mode 100644
new mode 100755
diff --git a/docs/reference/API_Reference_Guide.rst b/docs/reference/API_Reference_Guide.rst
old mode 100644
new mode 100755
diff --git a/docs/reference/Supported_Primitives_Guide.rst b/docs/reference/Supported_Primitives_Guide.rst
old mode 100644
new mode 100755
diff --git a/docs/reference/wrapper.rst b/docs/reference/wrapper.rst
old mode 100644
new mode 100755
diff --git a/docs/refs.bib b/docs/refs.bib
old mode 100644
new mode 100755
diff --git a/docs/sphinx/_toc.yml.in b/docs/sphinx/_toc.yml.in
old mode 100644
new mode 100755
diff --git a/docs/sphinx/requirements.in b/docs/sphinx/requirements.in
old mode 100644
new mode 100755
diff --git a/docs/sphinx/requirements.txt b/docs/sphinx/requirements.txt
old mode 100644
new mode 100755
diff --git a/docs/tutorial/tutorial_hello_world.rst b/docs/tutorial/tutorial_hello_world.rst
old mode 100644
new mode 100755
diff --git a/example/01_gemm/CMakeLists.txt b/example/01_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/01_gemm/README.md b/example/01_gemm/README.md
old mode 100644
new mode 100755
diff --git a/example/01_gemm/common.hpp b/example/01_gemm/common.hpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_dl_fp16.cpp b/example/01_gemm/gemm_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_dl_fp32.cpp b/example/01_gemm/gemm_dl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_dl_int4.cpp b/example/01_gemm/gemm_dl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_dl_int8.cpp b/example/01_gemm/gemm_dl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_dpp_fp16.cpp b/example/01_gemm/gemm_dpp_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_wmma_fp16.cpp b/example/01_gemm/gemm_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_bf16.cpp b/example/01_gemm/gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_bf16_rtn.cpp b/example/01_gemm/gemm_xdl_bf16_rtn.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp16.cpp b/example/01_gemm/gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp16_fp8.cpp b/example/01_gemm/gemm_xdl_fp16_fp8.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp16_v2.cpp b/example/01_gemm/gemm_xdl_fp16_v2.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp64.cpp b/example/01_gemm/gemm_xdl_fp64.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp8.cpp b/example/01_gemm/gemm_xdl_fp8.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_fp8_bf8.cpp b/example/01_gemm/gemm_xdl_fp8_bf8.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_int4.cpp b/example/01_gemm/gemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_int8.cpp b/example/01_gemm/gemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp b/example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_lds_direct_load_fp32.cpp b/example/01_gemm/gemm_xdl_lds_direct_load_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp b/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_streamk.cpp b/example/01_gemm/gemm_xdl_streamk.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/gemm_xdl_wavelet_fp16.cpp b/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/01_gemm/run_gemm_example.inc b/example/01_gemm/run_gemm_example.inc
old mode 100644
new mode 100755
diff --git a/example/02_gemm_bilinear/CMakeLists.txt b/example/02_gemm_bilinear/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/02_gemm_bilinear/README.md b/example/02_gemm_bilinear/README.md
old mode 100644
new mode 100755
diff --git a/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp b/example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/03_gemm_bias_relu/CMakeLists.txt b/example/03_gemm_bias_relu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/03_gemm_bias_relu/README.md b/example/03_gemm_bias_relu/README.md
old mode 100644
new mode 100755
diff --git a/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp b/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/CMakeLists.txt b/example/04_gemm_add_add_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/README.md b/example/04_gemm_add_add_fastgelu/README.md
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/common.hpp b/example/04_gemm_add_add_fastgelu/common.hpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_lds_direct_load_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc b/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/CMakeLists.txt b/example/09_convnd_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/README.md b/example/09_convnd_fwd/README.md
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_common.hpp b/example/09_convnd_fwd/convnd_fwd_common.hpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_common.hpp b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
old mode 100644
new mode 100755
diff --git a/example/09_convnd_fwd/run_convnd_fwd_example.inc b/example/09_convnd_fwd/run_convnd_fwd_example.inc
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc b/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
old mode 100644
new mode 100755
diff --git a/example/12_reduce/CMakeLists.txt b/example/12_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/12_reduce/README.md b/example/12_reduce/README.md
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_blockwise.cpp b/example/12_reduce/reduce_blockwise.cpp
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_blockwise_impl.hpp b/example/12_reduce/reduce_blockwise_impl.hpp
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_blockwise_two_call.cpp b/example/12_reduce/reduce_blockwise_two_call.cpp
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_example_common.hpp b/example/12_reduce/reduce_example_common.hpp
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_multiblock_atomic_add.cpp b/example/12_reduce/reduce_multiblock_atomic_add.cpp
old mode 100644
new mode 100755
diff --git a/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp b/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
old mode 100644
new mode 100755
diff --git a/example/13_pool2d_fwd/CMakeLists.txt b/example/13_pool2d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/13_pool2d_fwd/README.md b/example/13_pool2d_fwd/README.md
old mode 100644
new mode 100755
diff --git a/example/13_pool2d_fwd/pool2d_fwd_common.hpp b/example/13_pool2d_fwd/pool2d_fwd_common.hpp
old mode 100644
new mode 100755
diff --git a/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/14_gemm_quantization/CMakeLists.txt b/example/14_gemm_quantization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/CMakeLists.txt b/example/15_grouped_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/README.md b/example/15_grouped_gemm/README.md
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/15_grouped_gemm/run_grouped_gemm_example.inc b/example/15_grouped_gemm/run_grouped_gemm_example.inc
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp b/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
old mode 100644
new mode 100755
diff --git a/example/17_convnd_bwd_data/CMakeLists.txt b/example/17_convnd_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/17_convnd_bwd_data/README.md b/example/17_convnd_bwd_data/README.md
old mode 100644
new mode 100755
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp b/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
old mode 100644
new mode 100755
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/18_batched_gemm_reduce/CMakeLists.txt b/example/18_batched_gemm_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/19_binary_elementwise/CMakeLists.txt b/example/19_binary_elementwise/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp b/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
old mode 100644
new mode 100755
diff --git a/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp b/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
old mode 100644
new mode 100755
diff --git a/example/19_binary_elementwise/elementwise_add_1d.cpp b/example/19_binary_elementwise/elementwise_add_1d.cpp
old mode 100644
new mode 100755
diff --git a/example/19_binary_elementwise/elementwise_add_4d.cpp b/example/19_binary_elementwise/elementwise_add_4d.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/CMakeLists.txt b/example/20_grouped_conv_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/common.hpp b/example/20_grouped_conv_bwd_weight/common.hpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
old mode 100644
new mode 100755
diff --git a/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc b/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
old mode 100644
new mode 100755
diff --git a/example/21_gemm_layernorm/CMakeLists.txt b/example/21_gemm_layernorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp b/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/CMakeLists.txt b/example/22_cgemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_bf16.cpp b/example/22_cgemm/cgemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_common.hpp b/example/22_cgemm/cgemm_xdl_common.hpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_fp16.cpp b/example/22_cgemm/cgemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_fp32.cpp b/example/22_cgemm/cgemm_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_int4.cpp b/example/22_cgemm/cgemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/22_cgemm/cgemm_xdl_int8.cpp b/example/22_cgemm/cgemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/23_softmax/CMakeLists.txt b/example/23_softmax/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/23_softmax/README.md b/example/23_softmax/README.md
old mode 100644
new mode 100755
diff --git a/example/23_softmax/softmax_blockwise.cpp b/example/23_softmax/softmax_blockwise.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/CMakeLists.txt b/example/24_batched_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/batched_gemm_xdl_bf16.cpp b/example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/batched_gemm_xdl_int4.cpp b/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/batched_gemm_xdl_int8.cpp b/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/24_batched_gemm/run_batched_gemm_example.inc b/example/24_batched_gemm/run_batched_gemm_example.inc
old mode 100644
new mode 100755
diff --git a/example/25_gemm_bias_e_permute/CMakeLists.txt b/example/25_gemm_bias_e_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/CMakeLists.txt b/example/26_contraction/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/26_contraction/README.md b/example/26_contraction/README.md
old mode 100644
new mode 100755
diff --git a/example/26_contraction/common_instances.hpp b/example/26_contraction/common_instances.hpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_bf16_compute_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_bf16_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp16_compute_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_fp16_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp32_compute_bf16.cpp b/example/26_contraction/contraction_bilinear_xdl_fp32_compute_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp32_compute_fp16.cpp b/example/26_contraction/contraction_bilinear_xdl_fp32_compute_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp64.cpp b/example/26_contraction/contraction_bilinear_xdl_fp64.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_bilinear_xdl_fp64_compute_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_fp64_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_bf16_compute_fp32.cpp b/example/26_contraction/contraction_scale_xdl_bf16_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp16_compute_fp32.cpp b/example/26_contraction/contraction_scale_xdl_fp16_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp32.cpp b/example/26_contraction/contraction_scale_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp32_compute_bf16.cpp b/example/26_contraction/contraction_scale_xdl_fp32_compute_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp32_compute_fp16.cpp b/example/26_contraction/contraction_scale_xdl_fp32_compute_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp64.cpp b/example/26_contraction/contraction_scale_xdl_fp64.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/contraction_scale_xdl_fp64_compute_fp32.cpp b/example/26_contraction/contraction_scale_xdl_fp64_compute_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/26_contraction/run_contraction_bilinear_example.inc b/example/26_contraction/run_contraction_bilinear_example.inc
old mode 100644
new mode 100755
diff --git a/example/26_contraction/run_contraction_scale_example.inc b/example/26_contraction/run_contraction_scale_example.inc
old mode 100644
new mode 100755
diff --git a/example/27_layernorm2d_fwd/CMakeLists.txt b/example/27_layernorm2d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/27_layernorm2d_fwd/common.hpp b/example/27_layernorm2d_fwd/common.hpp
old mode 100644
new mode 100755
diff --git a/example/27_layernorm2d_fwd/layernorm2d_fwd_fp16.cpp b/example/27_layernorm2d_fwd/layernorm2d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/27_layernorm2d_fwd/layernorm2d_fwd_splitk_fp16.cpp b/example/27_layernorm2d_fwd/layernorm2d_fwd_splitk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/27_layernorm2d_fwd/run_layernorm_example.inc b/example/27_layernorm2d_fwd/run_layernorm_example.inc
old mode 100644
new mode 100755
diff --git a/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt b/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp b/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/29_batched_gemm_bias_e_permute/CMakeLists.txt b/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/README.md b/example/30_grouped_conv_fwd_multiple_d/README.md
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/common.hpp b/example/30_grouped_conv_fwd_multiple_d/common.hpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp b/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
old mode 100644
new mode 100755
diff --git a/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/CMakeLists.txt b/example/31_batched_gemm_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc b/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt b/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/cross_attention_forward_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_query_attention_forward_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/multi_query_attention_forward_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute_wmma.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc b/example/32_batched_gemm_scale_softmax_gemm/run_cross_attention_wmma.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc b/example/32_batched_gemm_scale_softmax_gemm/run_grouped_query_attention_forward_wmma.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc b/example/32_batched_gemm_scale_softmax_gemm/run_multi_query_attention_forward_wmma.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc b/example/32_batched_gemm_scale_softmax_gemm/run_self_attention_wmma.inc
old mode 100644
new mode 100755
diff --git a/example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/self_attention_forward_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/33_multiple_reduce/CMakeLists.txt b/example/33_multiple_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/33_multiple_reduce/README.md b/example/33_multiple_reduce/README.md
old mode 100644
new mode 100755
diff --git a/example/33_multiple_reduce/dual_reduce_common.hpp b/example/33_multiple_reduce/dual_reduce_common.hpp
old mode 100644
new mode 100755
diff --git a/example/33_multiple_reduce/dual_reduce_multiblock.cpp b/example/33_multiple_reduce/dual_reduce_multiblock.cpp
old mode 100644
new mode 100755
diff --git a/example/33_multiple_reduce/dual_reduce_threadwise.cpp b/example/33_multiple_reduce/dual_reduce_threadwise.cpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/CMakeLists.txt b/example/34_batchnorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/README.md b/example/34_batchnorm/README.md
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_backward_nhwc.cpp b/example/34_batchnorm/batchnorm_backward_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_common.hpp b/example/34_batchnorm/batchnorm_common.hpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
old mode 100644
new mode 100755
diff --git a/example/34_batchnorm/batchnorm_infer_impl.hpp b/example/34_batchnorm/batchnorm_infer_impl.hpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/CMakeLists.txt b/example/35_splitK_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/run_splitK_gemm_example.inc b/example/35_splitK_gemm/run_splitK_gemm_example.inc
old mode 100644
new mode 100755
index e3690984a..944377a4c
--- a/example/35_splitK_gemm/run_splitK_gemm_example.inc
+++ b/example/35_splitK_gemm/run_splitK_gemm_example.inc
@@ -60,10 +60,14 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
     case 1:
         a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
         b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        //a_m_k.GenerateTensorValue_control_entropy(GeneratorTensor_2<ADataType>);
+        //b_k_n.GenerateTensorValue_control_entropy(GeneratorTensor_2<BDataType>{-5, 5});
         break;
     case 2:
         a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
         b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        //a_m_k.GenerateTensorValue(GeneratorTensor_3_control_entropy<ADataType>{});
+        //b_k_n.GenerateTensorValue(GeneratorTensor_3_control_entropy<BDataType>{});
         break;
     default:
         a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_fp16_fp8.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp16_fp8.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_lds_direct_load_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/36_sparse_embedding/CMakeLists.txt b/example/36_sparse_embedding/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp b/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
old mode 100644
new mode 100755
diff --git a/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt b/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp b/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/common.hpp b/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
old mode 100644
new mode 100755
diff --git a/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
old mode 100644
new mode 100755
diff --git a/example/39_permute/CMakeLists.txt b/example/39_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/39_permute/common.hpp b/example/39_permute/common.hpp
old mode 100644
new mode 100755
diff --git a/example/39_permute/permute_1xHxW_fp16.cpp b/example/39_permute/permute_1xHxW_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/39_permute/permute_HxWx4_fp16.cpp b/example/39_permute/permute_HxWx4_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/39_permute/permute_NxHxW_fp16.cpp b/example/39_permute/permute_NxHxW_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/39_permute/run_permute_bundle_example.inc b/example/39_permute/run_permute_bundle_example.inc
old mode 100644
new mode 100755
diff --git a/example/39_permute/run_permute_element_example.inc b/example/39_permute/run_permute_element_example.inc
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/CMakeLists.txt b/example/40_conv2d_fwd_quantization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/common.hpp b/example/40_conv2d_fwd_quantization/common.hpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
old mode 100644
new mode 100755
diff --git a/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/CMakeLists.txt b/example/41_grouped_conv_conv_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc b/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/CMakeLists.txt b/example/42_groupnorm_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/common.hpp b/example/42_groupnorm_fwd/common.hpp
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/groupnorm_fwd_sigmoid_mul_fp16.cpp b/example/42_groupnorm_fwd/groupnorm_fwd_sigmoid_mul_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/groupnorm_fwd_splitk_fp16.cpp b/example/42_groupnorm_fwd/groupnorm_fwd_splitk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/groupnorm_fwd_swish_fp16.cpp b/example/42_groupnorm_fwd/groupnorm_fwd_swish_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/42_groupnorm_fwd/run_groupnorm_fwd_example.inc b/example/42_groupnorm_fwd/run_groupnorm_fwd_example.inc
old mode 100644
new mode 100755
diff --git a/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt b/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/CMakeLists.txt b/example/44_elementwise_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute.cpp b/example/44_elementwise_permute/elementwise_permute.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_3d.cpp b/example/44_elementwise_permute/elementwise_permute_3d.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp
old mode 100644
new mode 100755
diff --git a/example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp
old mode 100644
new mode 100755
diff --git a/example/45_elementwise_normalization/CMakeLists.txt b/example/45_elementwise_normalization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp b/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/CMakeLists.txt b/example/46_gemm_add_multiply/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/README.md b/example/46_gemm_add_multiply/README.md
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/common.hpp b/example/46_gemm_add_multiply/common.hpp
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc b/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
old mode 100644
new mode 100755
diff --git a/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
old mode 100644
new mode 100755
diff --git a/example/48_pool3d_fwd/CMakeLists.txt b/example/48_pool3d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/48_pool3d_fwd/pool3d_fwd_common.hpp b/example/48_pool3d_fwd/pool3d_fwd_common.hpp
old mode 100644
new mode 100755
diff --git a/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp b/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/49_maxpool2d_bwd/CMakeLists.txt b/example/49_maxpool2d_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
old mode 100644
new mode 100755
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/50_put_element/CMakeLists.txt b/example/50_put_element/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/50_put_element/put_element_fp16.cpp b/example/50_put_element/put_element_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/51_avgpool3d_bwd/CMakeLists.txt b/example/51_avgpool3d_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
old mode 100644
new mode 100755
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/52_im2col_col2im/CMakeLists.txt b/example/52_im2col_col2im/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/52_im2col_col2im/column_to_image_f32.cpp b/example/52_im2col_col2im/column_to_image_f32.cpp
old mode 100644
new mode 100755
diff --git a/example/52_im2col_col2im/common.hpp b/example/52_im2col_col2im/common.hpp
old mode 100644
new mode 100755
diff --git a/example/52_im2col_col2im/image_to_column_f32.cpp b/example/52_im2col_col2im/image_to_column_f32.cpp
old mode 100644
new mode 100755
diff --git a/example/53_layernorm2d_bwd/CMakeLists.txt b/example/53_layernorm2d_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/53_layernorm2d_bwd/layernorm2d_bwd_fp32.cpp b/example/53_layernorm2d_bwd/layernorm2d_bwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/54_groupnorm_bwd/CMakeLists.txt b/example/54_groupnorm_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/54_groupnorm_bwd/groupnorm_bwd_fp32.cpp b/example/54_groupnorm_bwd/groupnorm_bwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/60_gemm_multi_ABD/CMakeLists.txt b/example/60_gemm_multi_ABD/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp b/example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/61_contraction_multi_ABD/CMakeLists.txt b/example/61_contraction_multi_ABD/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp b/example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/CMakeLists.txt b/example/62_convnd_activ/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/binary/CMakeLists.txt b/example/62_convnd_activ/binary/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/binary/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp b/example/62_convnd_activ/binary/convnd_bwd_data_xdl_bilinear_residual_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/binary/convnd_fwd_xdl_bilinear_residual_fp16.cpp b/example/62_convnd_activ/binary/convnd_fwd_xdl_bilinear_residual_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp b/example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp b/example/62_convnd_activ/convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/CMakeLists.txt b/example/62_convnd_activ/multi_AB/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp b/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_bf16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp b/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp b/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_fp32.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp b/example/62_convnd_activ/multi_AB/conv_fwd_xdl_scaleadd_ab_int8.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp b/example/62_convnd_activ/multi_AB/convnd_fwd_activ_multi_ab_common.hpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/run_convnd_activ_example.inc b/example/62_convnd_activ/run_convnd_activ_example.inc
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/CMakeLists.txt b/example/62_convnd_activ/unary/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_activ_unary_common.hpp b/example/62_convnd_activ/unary/convnd_fwd_activ_unary_common.hpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_abs_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_abs_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_clippedrelu_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_clippedrelu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_elu_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_elu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_leakyrelu_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_leakyrelu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_pow_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_pow_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_relu_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_relu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_sigmoid_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_sigmoid_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_softrelu_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_softrelu_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/62_convnd_activ/unary/convnd_fwd_xdl_tanh_fp16.cpp b/example/62_convnd_activ/unary/convnd_fwd_xdl_tanh_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/63_layernorm4d_fwd/CMakeLists.txt b/example/63_layernorm4d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/63_layernorm4d_fwd/common.hpp b/example/63_layernorm4d_fwd/common.hpp
old mode 100644
new mode 100755
diff --git a/example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp b/example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/63_layernorm4d_fwd/layernorm4d_fwd_splitk_fp16.cpp b/example/63_layernorm4d_fwd/layernorm4d_fwd_splitk_fp16.cpp
old mode 100644
new mode 100755
diff --git a/example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc b/example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc
old mode 100644
new mode 100755
diff --git a/example/64_fpAintB_gemm/CMakeLists.txt b/example/64_fpAintB_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/example/64_fpAintB_gemm/common.hpp b/example/64_fpAintB_gemm/common.hpp
old mode 100644
new mode 100755
diff --git a/example/64_fpAintB_gemm/fp16int8_gemm_wmma.cpp b/example/64_fpAintB_gemm/fp16int8_gemm_wmma.cpp
old mode 100644
new mode 100755
diff --git a/example/64_fpAintB_gemm/run_gemm_example.inc b/example/64_fpAintB_gemm/run_gemm_example.inc
old mode 100644
new mode 100755
diff --git a/example/CMakeLists.txt b/example/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/include/ck/ck.hpp b/include/ck/ck.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/config.h.in b/include/ck/config.h.in
old mode 100644
new mode 100755
diff --git a/include/ck/host_utility/device_prop.hpp b/include/ck/host_utility/device_prop.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/host_utility/hip_check_error.hpp b/include/ck/host_utility/hip_check_error.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/host_utility/io.hpp b/include/ck/host_utility/io.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/host_utility/kernel_launch.hpp b/include/ck/host_utility/kernel_launch.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/host_utility/stream_utility.hpp b/include/ck/host_utility/stream_utility.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp b/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/stream_config.hpp b/include/ck/stream_config.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor/static_tensor.hpp b/include/ck/tensor/static_tensor.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/cluster_descriptor.hpp b/include/ck/tensor_description/cluster_descriptor.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/multi_index_transform.hpp b/include/ck/tensor_description/multi_index_transform.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/multi_index_transform_helper.hpp b/include/ck/tensor_description/multi_index_transform_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/tensor_adaptor.hpp b/include/ck/tensor_description/tensor_adaptor.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/tensor_descriptor.hpp b/include/ck/tensor_description/tensor_descriptor.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/tensor_descriptor_helper.hpp b/include/ck/tensor_description/tensor_descriptor_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_description/tensor_space_filling_curve.hpp b/include/ck/tensor_description/tensor_space_filling_curve.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp b/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp b/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp b/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_dequant.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_dequant.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp b/include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_base.hpp b/include/ck/tensor_operation/gpu/device/device_base.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_cgemm.hpp b/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp b/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp b/include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_elementwise.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_elementwise_scale.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise_scale.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_dequantB.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_dequantB.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp b/include/ck/tensor_operation/gpu/device/device_normalization_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp b/include/ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_normalization_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_normalization_fwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_permute.hpp b/include/ck/tensor_operation/gpu/device/device_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_put_element.hpp b/include/ck/tensor_operation/gpu/device/device_put_element.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_reduce.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_softmax.hpp b/include/ck/tensor_operation/gpu/device/device_softmax.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp b/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_fpAintB_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_fpAintB_gemm_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_query_attention_forward_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_query_attention_forward_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_multi_query_attention_forward_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multi_query_attention_forward_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/masking_specialization.hpp b/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/matrix_padder.hpp b/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp b/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/tensor_layout.hpp b/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp b/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/device/welford_helper.hpp b/include/ck/tensor_operation/gpu/device/welford_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/element/quantization_operation.hpp b/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp b/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d_scale.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d_scale.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_3d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_3d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_fpAintB_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_fpAintB_gemm_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle_lds_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v4_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_splitk_lds_direct_load.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
old mode 100644
new mode 100755
index b52f5c51b..fb89b45f1
--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
@@ -27,13 +27,13 @@ template <typename GridwiseGemm,
           typename CElementwiseOperation>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
-    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
-        kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
-                                             const Block2CTileMap& b2c_map,
-                                             const AElementwiseOperation a_element_op,
-                                             const BElementwiseOperation b_element_op,
-                                             const CElementwiseOperation c_element_op)
+    kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
+                                         const Block2CTileMap& b2c_map,
+                                         const AElementwiseOperation a_element_op,
+                                         const BElementwiseOperation b_element_op,
+                                         const CElementwiseOperation c_element_op)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
     defined(__gfx94__))
@@ -669,433 +669,450 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
                                const BElementwiseOperation b_element_op = BElementwiseOperation{},
                                const CElementwiseOperation c_element_op = CElementwiseOperation{})
     {
-        const FloatA* p_a_grid           = karg.p_a_grid;
-        const FloatB* p_b_grid           = karg.p_b_grid;
-        FloatC* p_c_grid                 = karg.p_c_grid;
-        const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
-            karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0Padded, karg.KPadded);
-        const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
-            karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0Padded, karg.KPadded);
-        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
-
-        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
-            MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
-
-        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
-        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
-        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
-            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
-
-        // divide block work by [KBatch, M, N]
-        const auto block_work_idx =
-            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
-
-        if(!block_2_ctile_map.ValidCTileIndex(
-               block_work_idx,
-               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
-                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+        for(auto i = 0; i < 1500; i++)
         {
-            return;
-        }
+            const FloatA* p_a_grid           = karg.p_a_grid;
+            const FloatB* p_b_grid           = karg.p_b_grid;
+            FloatC* p_c_grid                 = karg.p_c_grid;
+            const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(karg.M,
+                                                                                  karg.MPadded,
+                                                                                  karg.K,
+                                                                                  karg.StrideA,
+                                                                                  karg.k_batch,
+                                                                                  karg.K0Padded,
+                                                                                  karg.KPadded);
+            const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(karg.K,
+                                                                                  karg.NPadded,
+                                                                                  karg.N,
+                                                                                  karg.StrideB,
+                                                                                  karg.k_batch,
+                                                                                  karg.K0Padded,
+                                                                                  karg.KPadded);
+            const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
+
+            const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
+                MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
+
+            const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
+            const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
+            auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+                p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            // divide block work by [KBatch, M, N]
+            const auto block_work_idx =
+                block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+            if(!block_2_ctile_map.ValidCTileIndex(
+                   block_work_idx,
+                   make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
+                              c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
+            {
+                return;
+            }
 
-        const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
-        const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
-        const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+            const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
+            const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
+            const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
 
-        // HACK: this force m/n_block_data_idx_on_grid into SGPR
-        const index_t m_block_data_idx_on_grid =
-            __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
+            // HACK: this force m/n_block_data_idx_on_grid into SGPR
+            const index_t m_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
 
-        const index_t n_block_data_idx_on_grid =
-            __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
+            const index_t n_block_data_idx_on_grid =
+                __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
 
-        // lds max alignment
-        constexpr auto max_lds_align = K1;
+            // lds max alignment
+            constexpr auto max_lds_align = K1;
 
-        // A matrix in LDS memory, dst of blockwise copy
-        constexpr auto a_k0_m_k1_block_desc = [&]() {
-            if constexpr(ABlockLdsExtraM)
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
-                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor_aligned(
-                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
-            }
-        }();
+            // A matrix in LDS memory, dst of blockwise copy
+            constexpr auto a_k0_m_k1_block_desc = [&]() {
+                if constexpr(ABlockLdsExtraM)
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor_aligned(
+                        make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
+                }
+            }();
 
-        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
-            if constexpr(ABlockLdsExtraM)
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
-                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
-                               Number<MPerBlock + 1>{} * K1,
-                               K1,
-                               I1));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor_aligned(
-                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
-                    max_lds_align);
-            }
-        }();
-        // B matrix in LDS memory, dst of blockwise copy
-        constexpr auto b_k0_n_k1_block_desc = [&]() {
-            if constexpr(BBlockLdsExtraN)
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
-                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor_aligned(
-                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
-            }
-        }();
+            constexpr auto a_b_k0_m_k1_block_desc = [&]() {
+                if constexpr(ABlockLdsExtraM)
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
+                                   Number<MPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor_aligned(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
+                        max_lds_align);
+                }
+            }();
+            // B matrix in LDS memory, dst of blockwise copy
+            constexpr auto b_k0_n_k1_block_desc = [&]() {
+                if constexpr(BBlockLdsExtraN)
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor_aligned(
+                        make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
+                }
+            }();
 
-        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
-            if constexpr(BBlockLdsExtraN)
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
-                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
-                               Number<NPerBlock + 1>{} * K1,
-                               K1,
-                               I1));
-            }
-            else
+            constexpr auto b_b_k0_n_k1_block_desc = [&]() {
+                if constexpr(BBlockLdsExtraN)
+                {
+                    return make_naive_tensor_descriptor(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
+                                   Number<NPerBlock + 1>{} * K1,
+                                   K1,
+                                   I1));
+                }
+                else
+                {
+                    return make_naive_tensor_descriptor_aligned(
+                        make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
+                        max_lds_align);
+                }
+            }();
+            // A matrix blockwise copy
+            auto a_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    AElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<1, K0PerBlock, MPerBlock, K1>,
+                                                    ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                    ABlockTransferThreadClusterArrangeOrder,
+                                                    FloatA,
+                                                    LDSTypeA,
+                                                    decltype(a_b_k0_m_k1_grid_desc),
+                                                    decltype(a_b_k0_m_k1_block_desc),
+                                                    ABlockTransferSrcAccessOrder,
+                                                    Sequence<0, 2, 1, 3>,
+                                                    ABlockTransferSrcVectorDim,
+                                                    3,
+                                                    ABlockTransferSrcScalarPerVector,
+                                                    ABlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    AThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    a_b_k0_m_k1_grid_desc,
+                    make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
+                    a_element_op,
+                    a_b_k0_m_k1_block_desc,
+                    make_multi_index(0, 0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // B matrix blockwise copy
+            auto b_blockwise_copy =
+                ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                    BElementwiseOperation,
+                                                    ck::tensor_operation::element_wise::PassThrough,
+                                                    InMemoryDataOperationEnum::Set,
+                                                    Sequence<1, K0PerBlock, NPerBlock, K1>,
+                                                    BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                    BBlockTransferThreadClusterArrangeOrder,
+                                                    FloatB,
+                                                    LDSTypeB,
+                                                    decltype(b_b_k0_n_k1_grid_desc),
+                                                    decltype(b_b_k0_n_k1_block_desc),
+                                                    BBlockTransferSrcAccessOrder,
+                                                    Sequence<0, 2, 1, 3>,
+                                                    BBlockTransferSrcVectorDim,
+                                                    3,
+                                                    BBlockTransferSrcScalarPerVector,
+                                                    BBlockTransferDstScalarPerVector_K1,
+                                                    1,
+                                                    1,
+                                                    BThreadTransferSrcResetCoordinateAfterRun,
+                                                    true>(
+                    b_b_k0_n_k1_grid_desc,
+                    make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
+                    b_element_op,
+                    b_b_k0_n_k1_block_desc,
+                    make_multi_index(0, 0, 0, 0),
+                    ck::tensor_operation::element_wise::PassThrough{});
+
+            // GEMM definition
+            //   c_mtx += transpose(a_mtx) * b_mtx
+            //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+            //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+            //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+            //       register
+            // sanity check
+
+            auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+                BlockSize,
+                LDSTypeA,
+                LDSTypeB,
+                FloatAcc,
+                decltype(a_k0_m_k1_block_desc),
+                decltype(b_k0_n_k1_block_desc),
+                MPerXDL,
+                NPerXDL,
+                MRepeat,
+                NRepeat,
+                K1,
+                LoopSched,
+                ComputeTypeA,
+                ComputeTypeB>();
+
+            auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+            // LDS allocation for A and B: be careful of alignment
+            constexpr auto a_block_space_size = math::integer_least_multiple(
+                a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
+
+            auto p_a_block = reinterpret_cast<LDSTypeA*>(p_shared_block);
+            auto p_b_block = reinterpret_cast<LDSTypeB*>(p_a_block + a_block_space_size);
+
+            constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+            constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
+
+            auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
+            auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
+
+            // gridwise GEMM pipeline
+            const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
+                (a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
+                (K0PerBlock * K1));
+
+            const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
+
+            gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
+                                                                   a_b_k0_m_k1_block_desc,
+                                                                   a_blockwise_copy,
+                                                                   a_grid_buf,
+                                                                   a_block_buf,
+                                                                   a_block_slice_copy_step,
+                                                                   b_b_k0_n_k1_grid_desc,
+                                                                   b_b_k0_n_k1_block_desc,
+                                                                   b_blockwise_copy,
+                                                                   b_grid_buf,
+                                                                   b_block_buf,
+                                                                   b_block_slice_copy_step,
+                                                                   blockwise_gemm,
+                                                                   c_thread_buf,
+                                                                   num_k_block_main_loop);
+
+            // output: register to global memory
             {
-                return make_naive_tensor_descriptor_aligned(
-                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
-                    max_lds_align);
-            }
-        }();
-        // A matrix blockwise copy
-        auto a_blockwise_copy =
-            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
-                                                AElementwiseOperation,
-                                                ck::tensor_operation::element_wise::PassThrough,
-                                                InMemoryDataOperationEnum::Set,
-                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
-                                                ABlockTransferThreadClusterLengths_K0_M_K1,
-                                                ABlockTransferThreadClusterArrangeOrder,
-                                                FloatA,
-                                                LDSTypeA,
-                                                decltype(a_b_k0_m_k1_grid_desc),
-                                                decltype(a_b_k0_m_k1_block_desc),
-                                                ABlockTransferSrcAccessOrder,
-                                                Sequence<0, 2, 1, 3>,
-                                                ABlockTransferSrcVectorDim,
-                                                3,
-                                                ABlockTransferSrcScalarPerVector,
-                                                ABlockTransferDstScalarPerVector_K1,
-                                                1,
-                                                1,
-                                                AThreadTransferSrcResetCoordinateAfterRun,
-                                                true>(
-                a_b_k0_m_k1_grid_desc,
-                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
-                a_element_op,
-                a_b_k0_m_k1_block_desc,
-                make_multi_index(0, 0, 0, 0),
-                ck::tensor_operation::element_wise::PassThrough{});
-
-        // B matrix blockwise copy
-        auto b_blockwise_copy =
-            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
-                                                BElementwiseOperation,
-                                                ck::tensor_operation::element_wise::PassThrough,
-                                                InMemoryDataOperationEnum::Set,
-                                                Sequence<1, K0PerBlock, NPerBlock, K1>,
-                                                BBlockTransferThreadClusterLengths_K0_N_K1,
-                                                BBlockTransferThreadClusterArrangeOrder,
-                                                FloatB,
-                                                LDSTypeB,
-                                                decltype(b_b_k0_n_k1_grid_desc),
-                                                decltype(b_b_k0_n_k1_block_desc),
-                                                BBlockTransferSrcAccessOrder,
-                                                Sequence<0, 2, 1, 3>,
-                                                BBlockTransferSrcVectorDim,
-                                                3,
-                                                BBlockTransferSrcScalarPerVector,
-                                                BBlockTransferDstScalarPerVector_K1,
-                                                1,
-                                                1,
-                                                BThreadTransferSrcResetCoordinateAfterRun,
-                                                true>(
-                b_b_k0_n_k1_grid_desc,
-                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
-                b_element_op,
-                b_b_k0_n_k1_block_desc,
-                make_multi_index(0, 0, 0, 0),
-                ck::tensor_operation::element_wise::PassThrough{});
-
-        // GEMM definition
-        //   c_mtx += transpose(a_mtx) * b_mtx
-        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
-        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
-        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
-        //       register
-        // sanity check
-
-        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
-            BlockSize,
-            LDSTypeA,
-            LDSTypeB,
-            FloatAcc,
-            decltype(a_k0_m_k1_block_desc),
-            decltype(b_k0_n_k1_block_desc),
-            MPerXDL,
-            NPerXDL,
-            MRepeat,
-            NRepeat,
-            K1,
-            LoopSched,
-            ComputeTypeA,
-            ComputeTypeB>();
-
-        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
-
-        // LDS allocation for A and B: be careful of alignment
-        constexpr auto a_block_space_size =
-            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
-
-        auto p_a_block = reinterpret_cast<LDSTypeA*>(p_shared_block);
-        auto p_b_block = reinterpret_cast<LDSTypeB*>(p_a_block + a_block_space_size);
-
-        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
-        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
-
-        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
-        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
-
-        // gridwise GEMM pipeline
-        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
-            (a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
-            (K0PerBlock * K1));
-
-        const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
-
-        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
-                                                               a_b_k0_m_k1_block_desc,
-                                                               a_blockwise_copy,
-                                                               a_grid_buf,
-                                                               a_block_buf,
-                                                               a_block_slice_copy_step,
-                                                               b_b_k0_n_k1_grid_desc,
-                                                               b_b_k0_n_k1_block_desc,
-                                                               b_blockwise_copy,
-                                                               b_grid_buf,
-                                                               b_block_buf,
-                                                               b_block_slice_copy_step,
-                                                               blockwise_gemm,
-                                                               c_thread_buf,
-                                                               num_k_block_main_loop);
-
-        // output: register to global memory
-        {
-            constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
-            constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
-
-            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
-                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
-
-            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
-                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
-
-            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
-            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
-            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
-            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
-            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
-            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
-            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
-            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
-
-            constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
-                GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
-
-            auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-                static_cast<FloatC*>(p_shared_block),
-                c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
-
-            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
-                c_block_desc_mblock_mperblock_nblock_nperblock,
-                make_tuple(
-                    make_freeze_transform(I0), // freeze mblock
-                    make_unmerge_transform(make_tuple(CShuffleMRepeatPerShuffle,
-                                                      M1,
-                                                      M2,
-                                                      M3,
-                                                      M4)), // M1 = MWave, M2 * M3 * M4 = MPerXDL
-                    make_freeze_transform(I0),              // freeze nblock
-                    make_unmerge_transform(make_tuple(CShuffleNRepeatPerShuffle,
-                                                      N1,
-                                                      N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-                make_tuple(
-                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
-
-            // calculate origin of thread output tensor on global memory
-            //     blockwise GEMM c matrix starting index
-            const auto c_thread_mtx_on_block =
-                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
-
-            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
-            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
-
-            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
-                make_single_stage_tensor_adaptor(
-                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
-                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
-                    make_tuple(Sequence<0>{}));
-
-            const auto m_thread_data_on_block_idx =
-                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
-                    make_multi_index(m_thread_data_on_block));
-
-            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
-                make_single_stage_tensor_adaptor(
-                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
-                    make_tuple(Sequence<0, 1, 2>{}),
-                    make_tuple(Sequence<0>{}));
-
-            const auto n_thread_data_on_block_idx =
-                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
-                    make_multi_index(n_thread_data_on_block));
-
-            // VGPR to LDS
-            auto c_thread_copy_vgpr_to_lds =
-                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
-                                                   FloatC,
-                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
-                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
-                                                   ck::tensor_operation::element_wise::PassThrough,
-                                                   Sequence<CShuffleMRepeatPerShuffle,
-                                                            CShuffleNRepeatPerShuffle,
-                                                            I1,
-                                                            I1,
-                                                            M2,
-                                                            I1,
-                                                            M4,
-                                                            I1>,
-                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
-                                                   7,
-                                                   1,
-                                                   InMemoryDataOperationEnum::Set,
-                                                   1,
-                                                   true>{
-                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
-                    make_multi_index(0,
-                                     0,
-                                     m_thread_data_on_block_idx[I1],
-                                     n_thread_data_on_block_idx[I1],
-                                     m_thread_data_on_block_idx[I2],
-                                     m_thread_data_on_block_idx[I3],
-                                     m_thread_data_on_block_idx[I4],
-                                     n_thread_data_on_block_idx[I2]),
-                    ck::tensor_operation::element_wise::PassThrough{}};
-
-            // LDS to global
-            auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
-                ThisThreadBlock,            // index_t BlockSize,
-                CElementwiseOperation,      // ElementwiseOperation,
-                CGlobalMemoryDataOperation, // DstInMemOp,
-                Sequence<1,
-                         CShuffleMRepeatPerShuffle * MWave * MPerXDL,
-                         1,
-                         CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
-                CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
-                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
-                FloatC,               // typename SrcData,
-                FloatC,               // typename DstData,
-                decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
-                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
-                Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
-                3,                                          // index_t VectorDim,
-                CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
-                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
-                false> // bool ThreadTransferDstResetCoordinateAfterRun
-                {c_block_desc_mblock_mperblock_nblock_nperblock,
-                 make_multi_index(0, 0, 0, 0),
-                 c_grid_desc_mblock_mperblock_nblock_nperblock,
-                 make_multi_index(block_m_id, 0, block_n_id, 0),
-                 c_element_op};
-
-            constexpr auto mxdlperwave_forward_step =
-                make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
-            constexpr auto nxdlperwave_forward_step =
-                make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
-            constexpr auto nxdlperwave_backward_step =
-                make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
-
-            static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
-                constexpr auto mxdlperwave = mxdlperwave_iter;
-
-                static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
-                    constexpr bool nxdlperwave_forward_sweep =
-                        (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
-
-                    constexpr index_t nxdlperwave_value =
-                        nxdlperwave_forward_sweep
-                            ? nxdlperwave_iter
-                            : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
-
-                    constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
-
-                    // make sure it's safe to do ds_write
-                    block_sync_lds();
-
-                    // VGPR to LDS
-                    c_thread_copy_vgpr_to_lds.Run(
-                        c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
-                        make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
-                        c_thread_buf,
-                        c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
-                        c_block_buf);
-
-                    // make sure it's safe to do ds_read
-                    block_sync_lds();
-
-                    // LDS to global
-                    c_block_copy_lds_to_global.Run(c_block_desc_mblock_mperblock_nblock_nperblock,
-                                                   c_block_buf,
-                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
-                                                   c_grid_buf);
-
-                    // move on nxdlperwave dimension
-                    if constexpr(nxdlperwave_forward_sweep &&
-                                 (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
-                    {
-                        c_block_copy_lds_to_global.MoveDstSliceWindow(
+                constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
+                constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
+                    blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
+                    blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+                constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
+                constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
+                constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
+                constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
+                constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
+                constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
+                constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
+                constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
+
+                constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
+                    GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+                auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                    static_cast<FloatC*>(p_shared_block),
+                    c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+                constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                    c_block_desc_mblock_mperblock_nblock_nperblock,
+                    make_tuple(make_freeze_transform(I0), // freeze mblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleMRepeatPerShuffle,
+                                              M1,
+                                              M2,
+                                              M3,
+                                              M4)),       // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                               make_freeze_transform(I0), // freeze nblock
+                               make_unmerge_transform(
+                                   make_tuple(CShuffleNRepeatPerShuffle,
+                                              N1,
+                                              N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<>{},
+                               Sequence<0, 2, 4, 5, 6>{},
+                               Sequence<>{},
+                               Sequence<1, 3, 7>{}));
+
+                // calculate origin of thread output tensor on global memory
+                //     blockwise GEMM c matrix starting index
+                const auto c_thread_mtx_on_block =
+                    blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+                const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+                const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+                const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                        make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto m_thread_data_on_block_idx =
+                    m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                        make_multi_index(m_thread_data_on_block));
+
+                const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                    make_single_stage_tensor_adaptor(
+                        make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                        make_tuple(Sequence<0, 1, 2>{}),
+                        make_tuple(Sequence<0>{}));
+
+                const auto n_thread_data_on_block_idx =
+                    n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                        make_multi_index(n_thread_data_on_block));
+
+                // VGPR to LDS
+                auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
+                    FloatAcc,
+                    FloatC,
+                    decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
+                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                    ck::tensor_operation::element_wise::PassThrough,
+                    Sequence<CShuffleMRepeatPerShuffle,
+                             CShuffleNRepeatPerShuffle,
+                             I1,
+                             I1,
+                             M2,
+                             I1,
+                             M4,
+                             I1>,
+                    Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                    7,
+                    1,
+                    InMemoryDataOperationEnum::Set,
+                    1,
+                    true>{c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                          make_multi_index(0,
+                                           0,
+                                           m_thread_data_on_block_idx[I1],
+                                           n_thread_data_on_block_idx[I1],
+                                           m_thread_data_on_block_idx[I2],
+                                           m_thread_data_on_block_idx[I3],
+                                           m_thread_data_on_block_idx[I4],
+                                           n_thread_data_on_block_idx[I2]),
+                          ck::tensor_operation::element_wise::PassThrough{}};
+
+                // LDS to global
+                auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
+                    ThisThreadBlock,            // index_t BlockSize,
+                    CElementwiseOperation,      // ElementwiseOperation,
+                    CGlobalMemoryDataOperation, // DstInMemOp,
+                    Sequence<1,
+                             CShuffleMRepeatPerShuffle * MWave * MPerXDL,
+                             1,
+                             CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
+                    CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                    Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                    FloatC,               // typename SrcData,
+                    FloatC,               // typename DstData,
+                    decltype(c_block_desc_mblock_mperblock_nblock_nperblock),
+                    decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
+                    Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
+                    3,                                          // index_t VectorDim,
+                    CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
+                    true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
+                    false> // bool ThreadTransferDstResetCoordinateAfterRun
+                    {c_block_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(0, 0, 0, 0),
+                     c_grid_desc_mblock_mperblock_nblock_nperblock,
+                     make_multi_index(block_m_id, 0, block_n_id, 0),
+                     c_element_op};
+
+                constexpr auto mxdlperwave_forward_step =
+                    make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
+                constexpr auto nxdlperwave_forward_step =
+                    make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+                constexpr auto nxdlperwave_backward_step =
+                    make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
+
+                static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
+                    constexpr auto mxdlperwave = mxdlperwave_iter;
+
+                    static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
+                        constexpr bool nxdlperwave_forward_sweep =
+                            (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
+
+                        constexpr index_t nxdlperwave_value =
+                            nxdlperwave_forward_sweep
+                                ? nxdlperwave_iter
+                                : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
+
+                        constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
+
+                        // make sure it's safe to do ds_write
+                        block_sync_lds();
+
+                        // VGPR to LDS
+                        c_thread_copy_vgpr_to_lds.Run(
+                            c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
+                            make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
+                            c_thread_buf,
+                            c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                            c_block_buf);
+
+                        // make sure it's safe to do ds_read
+                        block_sync_lds();
+
+                        // LDS to global
+                        c_block_copy_lds_to_global.Run(
+                            c_block_desc_mblock_mperblock_nblock_nperblock,
+                            c_block_buf,
                             c_grid_desc_mblock_mperblock_nblock_nperblock,
-                            nxdlperwave_forward_step);
-                    }
-                    else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                            c_grid_buf);
+
+                        // move on nxdlperwave dimension
+                        if constexpr(nxdlperwave_forward_sweep &&
+                                     (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_forward_step);
+                        }
+                        else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
+                        {
+                            c_block_copy_lds_to_global.MoveDstSliceWindow(
+                                c_grid_desc_mblock_mperblock_nblock_nperblock,
+                                nxdlperwave_backward_step);
+                        }
+                    });
+
+                    // move on mxdlperwave dimension
+                    if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
                     {
                         c_block_copy_lds_to_global.MoveDstSliceWindow(
                             c_grid_desc_mblock_mperblock_nblock_nperblock,
-                            nxdlperwave_backward_step);
+                            mxdlperwave_forward_step);
                     }
                 });
-
-                // move on mxdlperwave dimension
-                if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
-                {
-                    c_block_copy_lds_to_global.MoveDstSliceWindow(
-                        c_grid_desc_mblock_mperblock_nblock_nperblock, mxdlperwave_forward_step);
-                }
-            });
+            }
         }
     }
 
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_data.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_gamma_beta.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_gamma_beta.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp b/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_dequant.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1_dequant.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp b/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp b/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm_arraybase.hpp b/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm_arraybase.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_address_space.hpp b/include/ck/utility/amd_address_space.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_buffer_addressing.hpp b/include/ck/utility/amd_buffer_addressing.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_gemm_dpp.hpp b/include/ck/utility/amd_gemm_dpp.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_inline_asm.hpp b/include/ck/utility/amd_inline_asm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_lds.hpp b/include/ck/utility/amd_lds.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_wave_read_first_lane.hpp b/include/ck/utility/amd_wave_read_first_lane.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_wmma.hpp b/include/ck/utility/amd_wmma.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/amd_xdlops.hpp b/include/ck/utility/amd_xdlops.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/array.hpp b/include/ck/utility/array.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/array_multi_index.hpp b/include/ck/utility/array_multi_index.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/c_style_pointer_cast.hpp b/include/ck/utility/c_style_pointer_cast.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/common_header.hpp b/include/ck/utility/common_header.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/container_element_picker.hpp b/include/ck/utility/container_element_picker.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/container_helper.hpp b/include/ck/utility/container_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/data_type.hpp b/include/ck/utility/data_type.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/debug.hpp b/include/ck/utility/debug.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/dynamic_buffer.hpp b/include/ck/utility/dynamic_buffer.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/enable_if.hpp b/include/ck/utility/enable_if.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/f8_utils.hpp b/include/ck/utility/f8_utils.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/functional.hpp b/include/ck/utility/functional.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/functional2.hpp b/include/ck/utility/functional2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/functional3.hpp b/include/ck/utility/functional3.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/functional4.hpp b/include/ck/utility/functional4.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/generic_memory_space_atomic.hpp b/include/ck/utility/generic_memory_space_atomic.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/get_id.hpp b/include/ck/utility/get_id.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/get_shift.hpp b/include/ck/utility/get_shift.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/ignore.hpp b/include/ck/utility/ignore.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/inner_product.hpp b/include/ck/utility/inner_product.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/inner_product_dpp8.hpp b/include/ck/utility/inner_product_dpp8.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/integral_constant.hpp b/include/ck/utility/integral_constant.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/is_detected.hpp b/include/ck/utility/is_detected.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/is_known_at_compile_time.hpp b/include/ck/utility/is_known_at_compile_time.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/loop_scheduler.hpp b/include/ck/utility/loop_scheduler.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/magic_division.hpp b/include/ck/utility/magic_division.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/math.hpp b/include/ck/utility/math.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/math_v2.hpp b/include/ck/utility/math_v2.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/multi_index.hpp b/include/ck/utility/multi_index.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/number.hpp b/include/ck/utility/number.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/random_gen.hpp b/include/ck/utility/random_gen.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/reduction_common.hpp b/include/ck/utility/reduction_common.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/reduction_enums.hpp b/include/ck/utility/reduction_enums.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/reduction_functions_accumulate.hpp b/include/ck/utility/reduction_functions_accumulate.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/reduction_operator.hpp b/include/ck/utility/reduction_operator.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/sequence.hpp b/include/ck/utility/sequence.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/sequence_helper.hpp b/include/ck/utility/sequence_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/span.hpp b/include/ck/utility/span.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/static_buffer.hpp b/include/ck/utility/static_buffer.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/statically_indexed_array.hpp b/include/ck/utility/statically_indexed_array.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/statically_indexed_array_multi_index.hpp b/include/ck/utility/statically_indexed_array_multi_index.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/synchronization.hpp b/include/ck/utility/synchronization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/thread_group.hpp b/include/ck/utility/thread_group.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/transpose_vectors.hpp b/include/ck/utility/transpose_vectors.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/tuple.hpp b/include/ck/utility/tuple.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/tuple_helper.hpp b/include/ck/utility/tuple_helper.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/type.hpp b/include/ck/utility/type.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/type_convert.hpp b/include/ck/utility/type_convert.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/workgroup_barrier.hpp b/include/ck/utility/workgroup_barrier.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/utility/workgroup_synchronization.hpp b/include/ck/utility/workgroup_synchronization.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/version.h.in b/include/ck/version.h.in
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/layout.hpp b/include/ck/wrapper/layout.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/operations/copy.hpp b/include/ck/wrapper/operations/copy.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/operations/gemm.hpp b/include/ck/wrapper/operations/gemm.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/tensor.hpp b/include/ck/wrapper/tensor.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/traits/blockwise_gemm_xdl_traits.hpp b/include/ck/wrapper/traits/blockwise_gemm_xdl_traits.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/utils/kernel_utils.hpp b/include/ck/wrapper/utils/kernel_utils.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/utils/layout_utils.hpp b/include/ck/wrapper/utils/layout_utils.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/utils/tensor_partition.hpp b/include/ck/wrapper/utils/tensor_partition.hpp
old mode 100644
new mode 100755
diff --git a/include/ck/wrapper/utils/tensor_utils.hpp b/include/ck/wrapper/utils/tensor_utils.hpp
old mode 100644
new mode 100755
diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_fpAintB_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_fpAintB_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp b/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp b/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction/device_contraction_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_f16_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_f16_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_i8_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_i8_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_scale_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_scale_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_bilinear_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_bilinear_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scale_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scale_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_ab_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_scaleadd_scaleadd_relu_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_bilinear.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_scale.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bilinear.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scale.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_ab.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_scaleadd_relu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_scaleadd_scaleadd_relu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp b/library/include/ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp b/library/include/ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization_fwd_swish.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp b/library/include/ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/transpose/device_transpose_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/transpose/device_transpose_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/transpose_3d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/transpose_3d.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/algorithm.hpp b/library/include/ck/library/utility/algorithm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/check_err.hpp b/library/include/ck/library/utility/check_err.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/conv_common.hpp b/library/include/ck/library/utility/conv_common.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp b/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/convolution_parameter.hpp b/library/include/ck/library/utility/convolution_parameter.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/device_memory.hpp b/library/include/ck/library/utility/device_memory.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/fill.hpp b/library/include/ck/library/utility/fill.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/host_common_util.hpp b/library/include/ck/library/utility/host_common_util.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/host_gemm.hpp b/library/include/ck/library/utility/host_gemm.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/host_tensor.hpp b/library/include/ck/library/utility/host_tensor.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/host_tensor_generator.hpp b/library/include/ck/library/utility/host_tensor_generator.hpp
old mode 100644
new mode 100755
index 6fd7ed8aa..b55cfb655
--- a/library/include/ck/library/utility/host_tensor_generator.hpp
+++ b/library/include/ck/library/utility/host_tensor_generator.hpp
@@ -142,6 +142,45 @@ struct GeneratorTensor_3
     }
 };
 
+// template <typename T>
+// struct GeneratorTensor_3_control_entropy
+// {
+
+//     // constexpr static int fp32_exponent_size = 8;
+//     // constexpr static int fp32_mantissa_size = 23;
+//     constexpr static int fp16_exponent_size = 5;
+//     constexpr static int fp16_mantissa_size = 10;
+
+//     mutable std::mt19937 gen{std::random_device{}()};
+//     mutable std::uniform_int_distribution<uint16_t> dis{0, 0xFFFF};
+//     float MAX_FP32_NUM = 0xFFFFFFFF;
+//     float MAX_FP16_NUM = 0xFFFF;
+
+//     template <typename... Is>
+//     T operator()(Is...) const
+//     {
+//         uint16_t bits = dis(gen); // Generate 32 random bits
+
+//         // Combine the bits into a floating-point number according to IEEE 754 format
+//         uint16_t sign     = bits >> 15;           // Get the sign bit
+//         uint16_t exponent = (bits >> 10) & 0x1F; // Get the exponent bits
+//         uint16_t mantissa = bits & 0x3FF;      // Get the mantissa bits
+
+//         if(exponent == 0x1F || exponent == 0)
+//         {
+//             // Avoid NaN and denormalized numbers
+//             return static_cast<T>(0.0);
+//         }
+
+//         // Shift the exponent to the correct position and set the sign bit
+
+//         int16_t result = (sign << 15) | ((exponent - 15 + 0x1F) << 10) |
+//                           (mantissa >> (10 - fp16_mantissa_size));
+//         float v= static_cast<float>(*reinterpret_cast<uint16_t*>(std::bitset<16>(result).to_string().c_str())) / MAX_FP16_NUM;
+//         printf("%0f,\t", v / MAX_FP16_NUM);
+//         return static_cast<T>(vector_type_maker / MAX_FP16_NUM);
+//     }
+// };
 template <>
 struct GeneratorTensor_3<ck::bhalf_t>
 {
diff --git a/library/include/ck/library/utility/iterator.hpp b/library/include/ck/library/utility/iterator.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/literals.hpp b/library/include/ck/library/utility/literals.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/numeric.hpp b/library/include/ck/library/utility/numeric.hpp
old mode 100644
new mode 100755
diff --git a/library/include/ck/library/utility/ranges.hpp b/library/include/ck/library/utility/ranges.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gndhwc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gndhwc_3d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnhwc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnhwc_2d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnwc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_gnwc_1d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_ndhwgc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_ndhwgc_3d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwgc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwgc_2d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nwgc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nwgc_1d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_bf16_compute_f32_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_f16_compute_f32_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_bf16_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_compute_f16_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_compute_f32_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_bf16_bf16_bf16_compute_f32_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f16_f16_f16_compute_f32_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_bf16_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_compute_f16_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_compute_f32_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_wmma_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f8_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_default_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_default_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_interwave_padded_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v1_padded_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_default_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_v2_padded_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_lds_direct_load_f32_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add/device_gemm_add_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_relu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu/device_gemm_add_relu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_silu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_silu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_silu/device_gemm_add_silu_xdl_c_shuffle_f16_i8_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_interwave_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v1_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_v2_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_interwave_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v1_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_v2_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v1_interwave_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_kpb128_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v1_interwave_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v1_interwave_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_v2_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_lds_direct_load_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_scale/xdl/device_grouped_conv3d_bwd_data_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scale/xdl/device_grouped_conv3d_fwd_xdl_scale_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_ab/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_ab_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_scaleadd_scaleadd_relu/xdl/device_grouped_conv3d_fwd_xdl_scaleadd_scaleadd_relu_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f8_f16_f16_mk_kn_mn_irregular_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gndhwc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gndhwc_3d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnhwc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnhwc_2d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnwc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_gnwc_1d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_ndhwgc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_ndhwgc_3d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwgc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwgc_2d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nwgc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nwgc_1d_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/normalization_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_groupnorm_bwd_data_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_groupnorm_bwd_data_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_data/device_layernorm2d_bwd_data_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_data/normalization_bwd_data_instance_common.hpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_data/normalization_bwd_data_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_groupnorm_bwd_gamma_beta_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_groupnorm_bwd_gamma_beta_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/device_layernorm2d_bwd_gamma_beta_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/normalization_bwd_gamma_beta_instance_common.hpp b/library/src/tensor_operation_instance/gpu/normalization_bwd_gamma_beta/normalization_bwd_gamma_beta_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/normalization_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_f32_f32_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_1d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_1d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_2d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_2d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_3d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_3d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_4d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_4d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_5d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_5d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_instances.cpp b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_instances.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/transpose/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/transpose/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp b/library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp
old mode 100644
new mode 100755
diff --git a/library/src/utility/CMakeLists.txt b/library/src/utility/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/library/src/utility/convolution_parameter.cpp b/library/src/utility/convolution_parameter.cpp
old mode 100644
new mode 100755
diff --git a/library/src/utility/device_memory.cpp b/library/src/utility/device_memory.cpp
old mode 100644
new mode 100755
diff --git a/library/src/utility/host_tensor.cpp b/library/src/utility/host_tensor.cpp
old mode 100644
new mode 100755
diff --git a/profiler/CMakeLists.txt b/profiler/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/profiler/README.md b/profiler/README.md
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/data_type_enum.hpp b/profiler/include/profiler/data_type_enum.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp b/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batchnorm_backward_impl.hpp b/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batchnorm_forward_impl.hpp b/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_batchnorm_infer_impl.hpp b/profiler/include/profiler/profile_batchnorm_infer_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_contraction_impl.hpp b/profiler/include/profiler/profile_contraction_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_contraction_utils.hpp b/profiler/include/profiler/profile_contraction_utils.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_conv_fwd_impl.hpp b/profiler/include/profiler/profile_conv_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp b/profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp b/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_impl.hpp b/profiler/include/profiler/profile_gemm_add_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp b/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_relu_impl.hpp b/profiler/include/profiler/profile_gemm_add_relu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_add_silu_impl.hpp b/profiler/include/profiler/profile_gemm_add_silu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_bilinear_impl.hpp b/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_impl.hpp b/profiler/include/profiler/profile_gemm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_reduce_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_splitk_impl.hpp b/profiler/include/profiler/profile_gemm_splitk_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_gemm_streamk_impl.hpp b/profiler/include/profiler/profile_gemm_streamk_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_grouped_gemm_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_groupnorm_bwd_data_impl.hpp b/profiler/include/profiler/profile_groupnorm_bwd_data_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp b/profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_groupnorm_fwd_impl.hpp b/profiler/include/profiler/profile_groupnorm_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_layernorm_bwd_data_impl.hpp b/profiler/include/profiler/profile_layernorm_bwd_data_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp b/profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_layernorm_fwd_impl.hpp b/profiler/include/profiler/profile_layernorm_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_permute_scale_impl.hpp b/profiler/include/profiler/profile_permute_scale_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_pool3d_fwd_impl.hpp b/profiler/include/profiler/profile_pool3d_fwd_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_reduce_impl.hpp b/profiler/include/profiler/profile_reduce_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_softmax_impl.hpp b/profiler/include/profiler/profile_softmax_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/include/profiler/profile_transpose_impl.hpp b/profiler/include/profiler/profile_transpose_impl.hpp
old mode 100644
new mode 100755
diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_avg_pool3d_bwd.cpp b/profiler/src/profile_avg_pool3d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batched_gemm.cpp b/profiler/src/profile_batched_gemm.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp b/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batched_gemm_gemm.cpp b/profiler/src/profile_batched_gemm_gemm.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batched_gemm_multi_d.cpp b/profiler/src/profile_batched_gemm_multi_d.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batched_gemm_reduce.cpp b/profiler/src/profile_batched_gemm_reduce.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batchnorm_bwd.cpp b/profiler/src/profile_batchnorm_bwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batchnorm_fwd.cpp b/profiler/src/profile_batchnorm_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_batchnorm_infer.cpp b/profiler/src/profile_batchnorm_infer.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_contraction_bilinear.cpp b/profiler/src/profile_contraction_bilinear.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_contraction_scale.cpp b/profiler/src/profile_contraction_scale.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_conv_bwd_data.cpp b/profiler/src/profile_conv_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_conv_fwd.cpp b/profiler/src/profile_conv_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_conv_fwd_bias_relu.cpp b/profiler/src/profile_conv_fwd_bias_relu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_conv_fwd_bias_relu_add.cpp b/profiler/src/profile_conv_fwd_bias_relu_add.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_conv_tensor_rearrange.cpp b/profiler/src/profile_conv_tensor_rearrange.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm.cpp b/profiler/src/profile_gemm.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add.cpp b/profiler/src/profile_gemm_add.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_add_fastgelu.cpp b/profiler/src/profile_gemm_add_add_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_fastgelu.cpp b/profiler/src/profile_gemm_add_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_multiply.cpp b/profiler/src/profile_gemm_add_multiply.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_relu.cpp b/profiler/src/profile_gemm_add_relu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_relu_add_layernorm.cpp b/profiler/src/profile_gemm_add_relu_add_layernorm.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_add_silu.cpp b/profiler/src/profile_gemm_add_silu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_bias_add_reduce.cpp b/profiler/src/profile_gemm_bias_add_reduce.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_bilinear.cpp b/profiler/src/profile_gemm_bilinear.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_fastgelu.cpp b/profiler/src/profile_gemm_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_multiply_add.cpp b/profiler/src/profile_gemm_multiply_add.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_reduce.cpp b/profiler/src/profile_gemm_reduce.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_splitk.cpp b/profiler/src/profile_gemm_splitk.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_gemm_streamk.cpp b/profiler/src/profile_gemm_streamk.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_conv_bwd_data.cpp b/profiler/src/profile_grouped_conv_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_conv_bwd_weight.cpp b/profiler/src/profile_grouped_conv_bwd_weight.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_conv_fwd.cpp b/profiler/src/profile_grouped_conv_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_gemm.cpp b/profiler/src/profile_grouped_gemm.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_gemm_fastgelu.cpp b/profiler/src/profile_grouped_gemm_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_grouped_gemm_fixed_nk.cpp b/profiler/src/profile_grouped_gemm_fixed_nk.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_groupnorm_bwd_data.cpp b/profiler/src/profile_groupnorm_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_groupnorm_bwd_gamma_beta.cpp b/profiler/src/profile_groupnorm_bwd_gamma_beta.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_groupnorm_fwd.cpp b/profiler/src/profile_groupnorm_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_layernorm_bwd_data.cpp b/profiler/src/profile_layernorm_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_layernorm_bwd_gamma_beta.cpp b/profiler/src/profile_layernorm_bwd_gamma_beta.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_layernorm_fwd.cpp b/profiler/src/profile_layernorm_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_max_pool3d_bwd.cpp b/profiler/src/profile_max_pool3d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_max_pool3d_fwd.cpp b/profiler/src/profile_max_pool3d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_permute_scale.cpp b/profiler/src/profile_permute_scale.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_reduce.cpp b/profiler/src/profile_reduce.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_softmax.cpp b/profiler/src/profile_softmax.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profile_transpose.cpp b/profiler/src/profile_transpose.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profiler.cpp b/profiler/src/profiler.cpp
old mode 100644
new mode 100755
diff --git a/profiler/src/profiler_operation_registry.hpp b/profiler/src/profiler_operation_registry.hpp
old mode 100644
new mode 100755
diff --git a/rbuild.ini b/rbuild.ini
old mode 100644
new mode 100755
diff --git a/requirements.txt b/requirements.txt
old mode 100644
new mode 100755
diff --git a/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh
index 51d6f7a30..13f5f0867 100755
--- a/script/cmake-ck-dev.sh
+++ b/script/cmake-ck-dev.sh
@@ -11,7 +11,7 @@ cmake
 -D CMAKE_CXX_FLAGS="-std=c++17 -O3 -ftemplate-backtrace-limit=0  -fPIE  -Wno-gnu-line-marker"     \
 -D CMAKE_BUILD_TYPE=Release                                                                       \
 -D BUILD_DEV=ON                                                                                   \
--D GPU_TARGETS="gfx908;gfx90a;gfx940"                                                             \
+-D GPU_TARGETS="gfx942"                                                             \
 -D CMAKE_VERBOSE_MAKEFILE:BOOL=ON                                                                 \
 -D USE_BITINT_EXTENSION_INT4=OFF                                                                  \
 ${MY_PROJECT_SOURCE}
diff --git a/script/hip_fatbin_insert b/script/hip_fatbin_insert
old mode 100644
new mode 100755
diff --git a/script/process_perf_data.py b/script/process_perf_data.py
old mode 100644
new mode 100755
diff --git a/script/redis-cli.conf b/script/redis-cli.conf
old mode 100644
new mode 100755
diff --git a/script/test_convnd_fwd.sh b/script/test_convnd_fwd.sh
old mode 100644
new mode 100755
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm/CMakeLists.txt b/test/batched_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm/test_batched_gemm.cpp b/test/batched_gemm/test_batched_gemm.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_gemm/CMakeLists.txt b/test/batched_gemm_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_multi_d/CMakeLists.txt b/test/batched_gemm_multi_d/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp b/test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_reduce/CMakeLists.txt b/test/batched_gemm_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp b/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm/CMakeLists.txt b/test/batched_gemm_softmax_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
old mode 100644
new mode 100755
diff --git a/test/batchnorm/CMakeLists.txt b/test/batchnorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/batchnorm/batchnorm_bwd_rank_4.cpp b/test/batchnorm/batchnorm_bwd_rank_4.cpp
old mode 100644
new mode 100755
diff --git a/test/batchnorm/batchnorm_fwd_rank_4.cpp b/test/batchnorm/batchnorm_fwd_rank_4.cpp
old mode 100644
new mode 100755
diff --git a/test/batchnorm/batchnorm_infer_rank_4.cpp b/test/batchnorm/batchnorm_infer_rank_4.cpp
old mode 100644
new mode 100755
diff --git a/test/block_swizzle_test/block_swizzle_test.cpp b/test/block_swizzle_test/block_swizzle_test.cpp
old mode 100644
new mode 100755
diff --git a/test/block_swizzle_test/rebuild.sh b/test/block_swizzle_test/rebuild.sh
old mode 100644
new mode 100755
diff --git a/test/block_swizzle_test/simple_args.h b/test/block_swizzle_test/simple_args.h
old mode 100644
new mode 100755
diff --git a/test/block_to_ctile_map/CMakeLists.txt b/test/block_to_ctile_map/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/block_to_ctile_map/test_block_to_ctile_map.cpp b/test/block_to_ctile_map/test_block_to_ctile_map.cpp
old mode 100644
new mode 100755
diff --git a/test/contraction/CMakeLists.txt b/test/contraction/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/contraction/test_contraction.cpp b/test/contraction/test_contraction.cpp
old mode 100644
new mode 100755
diff --git a/test/contraction/test_contraction_interface.cpp b/test/contraction/test_contraction_interface.cpp
old mode 100644
new mode 100755
diff --git a/test/conv_tensor_rearrange/CMakeLists.txt b/test/conv_tensor_rearrange/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp b/test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
old mode 100644
new mode 100755
diff --git a/test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp b/test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
old mode 100644
new mode 100755
diff --git a/test/conv_util/CMakeLists.txt b/test/conv_util/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/conv_util/conv_util.cpp b/test/conv_util/conv_util.cpp
old mode 100644
new mode 100755
diff --git a/test/convnd_bwd_data/CMakeLists.txt b/test/convnd_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/convnd_bwd_data/convnd_bwd_data.cpp b/test/convnd_bwd_data/convnd_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/test/convnd_fwd/CMakeLists.txt b/test/convnd_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/convnd_fwd/convnd_fwd.cpp b/test/convnd_fwd/convnd_fwd.cpp
old mode 100644
new mode 100755
diff --git a/test/data_type/CMakeLists.txt b/test/data_type/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/data_type/test_bf8.cpp b/test/data_type/test_bf8.cpp
old mode 100644
new mode 100755
diff --git a/test/data_type/test_fp8.cpp b/test/data_type/test_fp8.cpp
old mode 100644
new mode 100755
diff --git a/test/data_type/test_int4.cpp b/test/data_type/test_int4.cpp
old mode 100644
new mode 100755
diff --git a/test/data_type/type_convert_const.cpp b/test/data_type/type_convert_const.cpp
old mode 100644
new mode 100755
diff --git a/test/elementwise_normalization/CMakeLists.txt b/test/elementwise_normalization/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp b/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/CMakeLists.txt b/test/gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_bf16.cpp b/test/gemm/gemm_bf16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_fp16.cpp b/test/gemm/gemm_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_fp32.cpp b/test/gemm/gemm_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_fp64.cpp b/test/gemm/gemm_fp64.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_int8.cpp b/test/gemm/gemm_int8.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_standalone_xdl_fp16.cpp b/test/gemm/gemm_standalone_xdl_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/gemm_util.hpp b/test/gemm/gemm_util.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_nn_instance.cpp b/test/gemm/instance/gemm_f16_nn_instance.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_nn_instance.hpp b/test/gemm/instance/gemm_f16_nn_instance.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_nt_instance.cpp b/test/gemm/instance/gemm_f16_nt_instance.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_nt_instance.hpp b/test/gemm/instance/gemm_f16_nt_instance.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_tn_instance.cpp b/test/gemm/instance/gemm_f16_tn_instance.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_tn_instance.hpp b/test/gemm/instance/gemm_f16_tn_instance.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_tt_instance.cpp b/test/gemm/instance/gemm_f16_tt_instance.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_f16_tt_instance.hpp b/test/gemm/instance/gemm_f16_tt_instance.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm/run_gemm_test.inc b/test/gemm/run_gemm_test.inc
old mode 100644
new mode 100755
diff --git a/test/gemm_add/CMakeLists.txt b/test/gemm_add/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/gemm_add/test_gemm_add.hpp b/test/gemm_add/test_gemm_add.hpp
old mode 100644
new mode 100755
diff --git a/test/gemm_add/test_gemm_add_fastgelu.cpp b/test/gemm_add/test_gemm_add_fastgelu.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_add/test_gemm_add_relu.cpp b/test/gemm_add/test_gemm_add_relu.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_add/test_gemm_add_silu.cpp b/test/gemm_add/test_gemm_add_silu.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_layernorm/CMakeLists.txt b/test/gemm_layernorm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp b/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_reduce/CMakeLists.txt b/test/gemm_reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/gemm_reduce/gemm_reduce_fp16.cpp b/test/gemm_reduce/gemm_reduce_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_split_k/CMakeLists.txt b/test/gemm_split_k/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/gemm_split_k/test_gemm_splitk.cpp b/test/gemm_split_k/test_gemm_splitk.cpp
old mode 100644
new mode 100755
diff --git a/test/gemm_split_k/test_gemm_splitk_ut_cases.inc b/test/gemm_split_k/test_gemm_splitk_ut_cases.inc
old mode 100644
new mode 100755
diff --git a/test/gemm_split_k/test_gemm_splitk_util.hpp b/test/gemm_split_k/test_gemm_splitk_util.hpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_data/CMakeLists.txt b/test/grouped_convnd_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_weight/CMakeLists.txt b/test/grouped_convnd_bwd_weight/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_fwd/CMakeLists.txt b/test/grouped_convnd_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp b/test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_ab_interface.cpp b/test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_ab_interface.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp b/test/grouped_convnd_fwd/test_grouped_convnd_fwd_multi_d_interface_compatibility.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_gemm/CMakeLists.txt b/test/grouped_gemm/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/grouped_gemm/test_grouped_gemm_interface.cpp b/test/grouped_gemm/test_grouped_gemm_interface.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_gemm/test_grouped_gemm_splitk.cpp b/test/grouped_gemm/test_grouped_gemm_splitk.cpp
old mode 100644
new mode 100755
diff --git a/test/grouped_gemm/test_grouped_gemm_ut_cases.inc b/test/grouped_gemm/test_grouped_gemm_ut_cases.inc
old mode 100644
new mode 100755
diff --git a/test/grouped_gemm/test_grouped_gemm_util.hpp b/test/grouped_gemm/test_grouped_gemm_util.hpp
old mode 100644
new mode 100755
diff --git a/test/magic_number_division/CMakeLists.txt b/test/magic_number_division/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/magic_number_division/magic_number_division.cpp b/test/magic_number_division/magic_number_division.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_data/CMakeLists.txt b/test/normalization_bwd_data/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_data/test_groupnorm_bwd_data_fp32.cpp b/test/normalization_bwd_data/test_groupnorm_bwd_data_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_data/test_layernorm2d_bwd_data_fp32.cpp b/test/normalization_bwd_data/test_layernorm2d_bwd_data_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_gamma_beta/CMakeLists.txt b/test/normalization_bwd_gamma_beta/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_gamma_beta/test_groupnorm_bwd_gamma_beta_fp32.cpp b/test/normalization_bwd_gamma_beta/test_groupnorm_bwd_gamma_beta_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_bwd_gamma_beta/test_layernorm2d_bwd_gamma_beta_fp32.cpp b/test/normalization_bwd_gamma_beta/test_layernorm2d_bwd_gamma_beta_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/CMakeLists.txt b/test/normalization_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/test_groupnorm_fwd_fp16.cpp b/test/normalization_fwd/test_groupnorm_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/test_groupnorm_fwd_fp32.cpp b/test/normalization_fwd/test_groupnorm_fwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/test_layernorm2d_fwd_fp16.cpp b/test/normalization_fwd/test_layernorm2d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/test_layernorm2d_fwd_fp32.cpp b/test/normalization_fwd/test_layernorm2d_fwd_fp32.cpp
old mode 100644
new mode 100755
diff --git a/test/normalization_fwd/test_layernorm4d_fwd_fp16.cpp b/test/normalization_fwd/test_layernorm4d_fwd_fp16.cpp
old mode 100644
new mode 100755
diff --git a/test/permute_scale/CMakeLists.txt b/test/permute_scale/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/permute_scale/test_permute_scale.cpp b/test/permute_scale/test_permute_scale.cpp
old mode 100644
new mode 100755
diff --git a/test/pool/CMakeLists.txt b/test/pool/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/pool/test_avg_pool3d_bwd.cpp b/test/pool/test_avg_pool3d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/test/pool/test_avg_pool3d_fwd.cpp b/test/pool/test_avg_pool3d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/test/pool/test_max_pool3d_bwd.cpp b/test/pool/test_max_pool3d_bwd.cpp
old mode 100644
new mode 100755
diff --git a/test/pool/test_max_pool3d_fwd.cpp b/test/pool/test_max_pool3d_fwd.cpp
old mode 100644
new mode 100755
diff --git a/test/pool/test_pool_fwd_common.hpp b/test/pool/test_pool_fwd_common.hpp
old mode 100644
new mode 100755
diff --git a/test/reduce/CMakeLists.txt b/test/reduce/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/reduce/reduce_no_index.cpp b/test/reduce/reduce_no_index.cpp
old mode 100644
new mode 100755
diff --git a/test/reduce/reduce_with_index.cpp b/test/reduce/reduce_with_index.cpp
old mode 100644
new mode 100755
diff --git a/test/reference_conv_fwd/CMakeLists.txt b/test/reference_conv_fwd/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/reference_conv_fwd/reference_conv_fwd.cpp b/test/reference_conv_fwd/reference_conv_fwd.cpp
old mode 100644
new mode 100755
diff --git a/test/softmax/CMakeLists.txt b/test/softmax/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/softmax/test_softmax_interface.cpp b/test/softmax/test_softmax_interface.cpp
old mode 100644
new mode 100755
diff --git a/test/softmax/test_softmax_rank3.cpp b/test/softmax/test_softmax_rank3.cpp
old mode 100644
new mode 100755
diff --git a/test/softmax/test_softmax_rank4.cpp b/test/softmax/test_softmax_rank4.cpp
old mode 100644
new mode 100755
diff --git a/test/softmax/test_softmax_ut_cases.inc b/test/softmax/test_softmax_ut_cases.inc
old mode 100644
new mode 100755
diff --git a/test/softmax/test_softmax_util.hpp b/test/softmax/test_softmax_util.hpp
old mode 100644
new mode 100755
diff --git a/test/space_filling_curve/CMakeLists.txt b/test/space_filling_curve/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/space_filling_curve/space_filling_curve.cpp b/test/space_filling_curve/space_filling_curve.cpp
old mode 100644
new mode 100755
diff --git a/test/transpose/CMakeLists.txt b/test/transpose/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/transpose/test_transpose.cpp b/test/transpose/test_transpose.cpp
old mode 100644
new mode 100755
diff --git a/test/wmma_op/CMakeLists.txt b/test/wmma_op/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/wmma_op/wmma_op.cpp b/test/wmma_op/wmma_op.cpp
old mode 100644
new mode 100755
diff --git a/test/wmma_op/wmma_op_util.hpp b/test/wmma_op/wmma_op_util.hpp
old mode 100644
new mode 100755
diff --git a/test/wrapper/CMakeLists.txt b/test/wrapper/CMakeLists.txt
old mode 100644
new mode 100755
diff --git a/test/wrapper/test_wrapper_copy.cpp b/test/wrapper/test_wrapper_copy.cpp
old mode 100644
new mode 100755
diff --git a/test/wrapper/test_wrapper_gemm.cpp b/test/wrapper/test_wrapper_gemm.cpp
old mode 100644
new mode 100755
diff --git a/test/wrapper/test_wrapper_layout.cpp b/test/wrapper/test_wrapper_layout.cpp
old mode 100644
new mode 100755
diff --git a/test/wrapper/test_wrapper_partition.cpp b/test/wrapper/test_wrapper_partition.cpp
old mode 100644
new mode 100755
diff --git a/test/wrapper/test_wrapper_tensor.cpp b/test/wrapper/test_wrapper_tensor.cpp
old mode 100644
new mode 100755
-- 
GitLab